diff --git a/.gitattributes b/.gitattributes index d43b8ce152afbe5249aaa59dfb15faefbba90001..93fce726a35124122f48e426448d871829454e33 100644 --- a/.gitattributes +++ b/.gitattributes @@ -276,3 +276,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2024/Blind[[:space:]]Image[[:space:]]Deconvolution[[:space:]]by[[:space:]]Generative-based[[:space:]]Kernel[[:space:]]Prior[[:space:]]and[[:space:]]Initializer[[:space:]]via[[:space:]]Latent[[:space:]]Encoding/43dfd375-9b0b-43fc-bd91-b16f42fac7e2_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/Blind[[:space:]]image[[:space:]]deblurring[[:space:]]with[[:space:]]noise-robust[[:space:]]kernel[[:space:]]estimation/d41417e8-8761-43ae-8b44-cd29d9b6b480_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/BlinkVision_[[:space:]]A[[:space:]]Benchmark[[:space:]]for[[:space:]]Optical[[:space:]]Flow,[[:space:]]Scene[[:space:]]Flow[[:space:]]and[[:space:]]Point[[:space:]]Tracking[[:space:]]Estimation[[:space:]]using[[:space:]]RGB[[:space:]]Frames[[:space:]]and[[:space:]]Events/6b3549ce-8cc9-43f1-a37e-d09d28b58d56_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Bones[[:space:]]Can't[[:space:]]Be[[:space:]]Triangles_[[:space:]]Accurate[[:space:]]and[[:space:]]Efficient[[:space:]]Vertebrae[[:space:]]Keypoint[[:space:]]Estimation[[:space:]]through[[:space:]]Collaborative[[:space:]]Error[[:space:]]Revision/79ad0f6a-e5ca-4e70-bd32-205034f5cde0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Boost[[:space:]]Your[[:space:]]NeRF_[[:space:]]A[[:space:]]Model-Agnostic[[:space:]]Mixture[[:space:]]of[[:space:]]Experts[[:space:]]Framework[[:space:]]for[[:space:]]High[[:space:]]Quality[[:space:]]and[[:space:]]Efficient[[:space:]]Rendering/cdbb2d54-c057-481d-a45f-f2c36fd59358_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Boosting[[:space:]]3D[[:space:]]Single[[:space:]]Object[[:space:]]Tracking[[:space:]]with[[:space:]]2D[[:space:]]Matching[[:space:]]Distillation[[:space:]]and[[:space:]]3D[[:space:]]Pre-training/515a618d-9747-46e7-9c9a-8a152456fd1b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Boosting[[:space:]]Gaze[[:space:]]Object[[:space:]]Prediction[[:space:]]via[[:space:]]Pixel-level[[:space:]]Supervision[[:space:]]from[[:space:]]Vision[[:space:]]Foundation[[:space:]]Model/93449abc-6022-4367-830a-8808cd9b5ba3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Boosting[[:space:]]Transferability[[:space:]]in[[:space:]]Vision-Language[[:space:]]Attacks[[:space:]]via[[:space:]]Diversification[[:space:]]along[[:space:]]the[[:space:]]Intersection[[:space:]]Region[[:space:]]of[[:space:]]Adversarial[[:space:]]Trajectory/d6848265-b938-496f-9e0e-c16ce7da086b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Boosting[[:space:]]the[[:space:]]Power[[:space:]]of[[:space:]]Small[[:space:]]Multimodal[[:space:]]Reasoning[[:space:]]Models[[:space:]]to[[:space:]]Match[[:space:]]Larger[[:space:]]Models[[:space:]]with[[:space:]]Self-Consistency[[:space:]]Training/cda53ae2-a211-4097-a2a7-2df98969f4ef_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Bottom-Up[[:space:]]Domain[[:space:]]Prompt[[:space:]]Tuning[[:space:]]for[[:space:]]Generalized[[:space:]]Face[[:space:]]Anti-Spoofing/3e12da5b-7b20-4ba8-b65d-4a8790f34fde_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Brain[[:space:]]Netflix_[[:space:]]Scaling[[:space:]]Data[[:space:]]to[[:space:]]Reconstruct[[:space:]]Videos[[:space:]]from[[:space:]]Brain[[:space:]]Signals/9af6f00c-d99e-4633-82c3-bc4c9c80afb6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Brain-ID_[[:space:]]Learning[[:space:]]Contrast-agnostic[[:space:]]Anatomical[[:space:]]Representations[[:space:]]for[[:space:]]Brain[[:space:]]Imaging/b53a82d6-784c-4409-ad49-9bb017717031_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Bridge[[:space:]]Past[[:space:]]and[[:space:]]Future_[[:space:]]Overcoming[[:space:]]Information[[:space:]]Asymmetry[[:space:]]in[[:space:]]Incremental[[:space:]]Object[[:space:]]Detection/1c449442-a7f3-4df8-9bc6-973b26235638_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Bridging[[:space:]]Different[[:space:]]Language[[:space:]]Models[[:space:]]and[[:space:]]Generative[[:space:]]Vision[[:space:]]Models[[:space:]]for[[:space:]]Text-to-Image[[:space:]]Generation/1764d397-410b-451c-a099-8966f6b5772c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Bridging[[:space:]]Synthetic[[:space:]]and[[:space:]]Real[[:space:]]Worlds[[:space:]]for[[:space:]]Pre-training[[:space:]]Scene[[:space:]]Text[[:space:]]Detectors/80dda151-f601-4dfe-a0ed-8536e8b997b9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Bridging[[:space:]]the[[:space:]]Gap[[:space:]]Between[[:space:]]Human[[:space:]]Motion[[:space:]]and[[:space:]]Action[[:space:]]Semantics[[:space:]]via[[:space:]]Kinematics[[:space:]]Phrases/ceda5829-f6a0-4ea2-94dd-2ddc6da8e6ed_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Bridging[[:space:]]the[[:space:]]Gap_[[:space:]]Studio-like[[:space:]]Avatar[[:space:]]Creation[[:space:]]from[[:space:]]a[[:space:]]Monocular[[:space:]]Phone[[:space:]]Capture/3e2c48bd-a83b-478a-9094-3e6d6e7a5043_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Bridging[[:space:]]the[[:space:]]Pathology[[:space:]]Domain[[:space:]]Gap_[[:space:]]Efficiently[[:space:]]Adapting[[:space:]]CLIP[[:space:]]for[[:space:]]Pathology[[:space:]]Image[[:space:]]Analysis[[:space:]]with[[:space:]]Limited[[:space:]]Labeled[[:space:]]Data/3b6fb56d-1649-4e06-bdb2-706d8b03ddb5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/BrushNet_[[:space:]]A[[:space:]]Plug-and-Play[[:space:]]Image[[:space:]]Inpainting[[:space:]]Model[[:space:]]with[[:space:]]Decomposed[[:space:]]Dual-Branch[[:space:]]Diffusion/8cdf68e6-b7ca-45f7-9a98-9bdc1df71997_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Bucketed[[:space:]]Ranking-based[[:space:]]Losses[[:space:]]for[[:space:]]Efficient[[:space:]]Training[[:space:]]of[[:space:]]Object[[:space:]]Detectors/7957777c-4d0a-4feb-89c6-b57e75289486_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/BugNIST[[:space:]]-[[:space:]]a[[:space:]]Large[[:space:]]Volumetric[[:space:]]Dataset[[:space:]]for[[:space:]]Detection[[:space:]]under[[:space:]]Domain[[:space:]]Shift/fceefb33-6fd3-4a63-9cb1-f0d439dc4d49_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/BurstM_[[:space:]]Deep[[:space:]]Burst[[:space:]]Multi-scale[[:space:]]SR[[:space:]]using[[:space:]]Fourier[[:space:]]Space[[:space:]]with[[:space:]]Optical[[:space:]]Flow/608ca9aa-ec95-4beb-807e-424a93d0d485_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/ByteEdit_[[:space:]]Boost,[[:space:]]Comply[[:space:]]and[[:space:]]Accelerate[[:space:]]Generative[[:space:]]Image[[:space:]]Editing/f2b7631d-9e11-4a5b-9762-122914fe2a69_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/C2C_[[:space:]]Component-to-Composition[[:space:]]Learning[[:space:]]for[[:space:]]Zero-Shot[[:space:]]Compositional[[:space:]]Action[[:space:]]Recognition/c79e54f7-761f-47c3-bbd9-d80fc5c9ad18_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CARB-Net_[[:space:]]Camera-Assisted[[:space:]]Radar-Based[[:space:]]Network[[:space:]]for[[:space:]]Vulnerable[[:space:]]Road[[:space:]]User[[:space:]]Detection/ed9a2da4-1d84-4538-8283-1656bf597855_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CARFF_[[:space:]]Conditional[[:space:]]Auto-encoded[[:space:]]Radiance[[:space:]]Field[[:space:]]for[[:space:]]3D[[:space:]]Scene[[:space:]]Forecasting/4d2110cb-f049-4580-b1d7-c4b728d00449_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CAT-SAM_[[:space:]]Conditional[[:space:]]Tuning[[:space:]]for[[:space:]]Few-Shot[[:space:]]Adaptation[[:space:]]of[[:space:]]Segment[[:space:]]Anything[[:space:]]Model/bd649bc1-acc8-4027-a33f-08ccf2414aa5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CAT_[[:space:]]Enhancing[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]to[[:space:]]Answer[[:space:]]Questions[[:space:]]in[[:space:]]Dynamic[[:space:]]Audio-Visual[[:space:]]Scenarios/825b02cb-2d21-47e3-b7b8-2d59e5594aa4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CC-SAM_[[:space:]]Enhancing[[:space:]]SAM[[:space:]]with[[:space:]]Cross-feature[[:space:]]Attention[[:space:]]and[[:space:]]Context[[:space:]]for[[:space:]]Ultrasound[[:space:]]Image[[:space:]]Segmentation/4f25c9c8-80da-4658-ab8a-cd035c32acd6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CG-SLAM_[[:space:]]Efficient[[:space:]]Dense[[:space:]]RGB-D[[:space:]]SLAM[[:space:]]in[[:space:]]a[[:space:]]Consistent[[:space:]]Uncertainty-aware[[:space:]]3D[[:space:]]Gaussian[[:space:]]Field/4bc60e08-fd5b-49a8-aee4-fd600735bd69_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CIC-BART-SSA_[[:space:]]_[[:space:]]Controllable[[:space:]]Image[[:space:]]Captioning[[:space:]]with[[:space:]]Structured[[:space:]]Semantic[[:space:]]Augmentation/87b30ae0-4f07-48fc-b46b-fbcdc143e18d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CLAMP-ViT_[[:space:]]Contrastive[[:space:]]Data-Free[[:space:]]Learning[[:space:]]for[[:space:]]Adaptive[[:space:]]Post-Training[[:space:]]Quantization[[:space:]]of[[:space:]]ViTs/9ac2383c-dda4-4497-a269-a04b9fe79777_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CLAP_[[:space:]]Isolating[[:space:]]Content[[:space:]]from[[:space:]]Style[[:space:]]through[[:space:]]Contrastive[[:space:]]Learning[[:space:]]with[[:space:]]Augmented[[:space:]]Prompts/f676b4a9-3170-48e9-b0be-002172d2221e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CLEO_[[:space:]]Continual[[:space:]]Learning[[:space:]]of[[:space:]]Evolving[[:space:]]Ontologies/7399fe90-4063-4ee4-8949-4afe680e9068_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CLIFF_[[:space:]]Continual[[:space:]]Latent[[:space:]]Diffusion[[:space:]]for[[:space:]]Open-Vocabulary[[:space:]]Object[[:space:]]Detection/51812c3e-19bd-4e17-9855-dbf65e0d5287_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CLIP-DINOiser_[[:space:]]Teaching[[:space:]]CLIP[[:space:]]a[[:space:]]few[[:space:]]DINO[[:space:]]tricks[[:space:]]for[[:space:]]open-vocabulary[[:space:]]semantic[[:space:]]segmentation/ab181842-dc49-431d-ab8a-9cea7505e7b5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CLIP-DPO_[[:space:]]Vision-Language[[:space:]]Models[[:space:]]as[[:space:]]a[[:space:]]Source[[:space:]]of[[:space:]]Preference[[:space:]]for[[:space:]]Fixing[[:space:]]Hallucinations[[:space:]]in[[:space:]]LVLMs/174c66bb-6859-4a7e-835b-0b1709492a92_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CLIP-Guided[[:space:]]Generative[[:space:]]Networks[[:space:]]for[[:space:]]Transferable[[:space:]]Targeted[[:space:]]Adversarial[[:space:]]Attacks/ddf47f95-f1c1-405a-8323-f8a66b1071ea_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CLOSER_[[:space:]]Towards[[:space:]]Better[[:space:]]Representation[[:space:]]Learning[[:space:]]for[[:space:]]Few-Shot[[:space:]]Class-Incremental[[:space:]]Learning/10463618-6ae0-4309-91d1-f1ab71be0dbc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CLR-GAN_[[:space:]]Improving[[:space:]]GANs[[:space:]]Stability[[:space:]]and[[:space:]]Quality[[:space:]]via[[:space:]]Consistent[[:space:]]Latent[[:space:]]Representation[[:space:]]and[[:space:]]Reconstruction/c724f22f-d6fb-423d-a760-902dfb6a8a1d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CMD_[[:space:]]A[[:space:]]Cross[[:space:]]Mechanism[[:space:]]Domain[[:space:]]Adaptation[[:space:]]Dataset[[:space:]]for[[:space:]]3D[[:space:]]Object[[:space:]]Detection/55286f6e-e84b-4ee2-994b-347fc5f78848_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CMTA_[[:space:]]Cross-Modal[[:space:]]Temporal[[:space:]]Alignment[[:space:]]for[[:space:]]Event-guided[[:space:]]Video[[:space:]]Deblurring/7cf7582c-c0da-4c34-8883-80e1b8a2e3ef_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/COD_[[:space:]]Learning[[:space:]]Conditional[[:space:]]Invariant[[:space:]]Representation[[:space:]]for[[:space:]]Domain[[:space:]]Adaptation[[:space:]]Regression/5f47c561-6103-4c2b-b6da-ac9368ec9383_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/COHO_[[:space:]]Context-Sensitive[[:space:]]City-Scale[[:space:]]Hierarchical[[:space:]]Urban[[:space:]]Layout[[:space:]]Generation/ad29cc50-ea60-4c58-8c17-9b4b4ddc225f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/COIN-Matting_[[:space:]]Confounder[[:space:]]Intervention[[:space:]]for[[:space:]]Image[[:space:]]Matting/7e767e14-ba96-40b5-847c-1a51f71239aa_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/COIN_[[:space:]]Control-Inpainting[[:space:]]Diffusion[[:space:]]Prior[[:space:]]for[[:space:]]Human[[:space:]]and[[:space:]]Camera[[:space:]]Motion[[:space:]]Estimation/213c3e91-2355-4081-9e3c-28b7d1b0600c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/COM[[:space:]]Kitchens_[[:space:]]An[[:space:]]Unedited[[:space:]]Overhead-view[[:space:]]Procedural[[:space:]]Videos[[:space:]]Dataset[[:space:]]a[[:space:]]Vision-Language[[:space:]]Benchmark/82c10f54-7bc3-400b-b41c-eee014f604fb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/COMO_[[:space:]]Compact[[:space:]]Mapping[[:space:]]and[[:space:]]Odometry/005f4f87-73cd-41e4-ad7b-b179ff03449c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/COMPOSE_[[:space:]]Comprehensive[[:space:]]Portrait[[:space:]]Shadow[[:space:]]Editing/2be02136-d81b-443f-a244-263becec430b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CONDA_[[:space:]]Condensed[[:space:]]Deep[[:space:]]Association[[:space:]]Learning[[:space:]]for[[:space:]]Co-Salient[[:space:]]Object[[:space:]]Detection./4b7d9a37-9ffb-4ebc-9e75-7e4ada571495_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/COSMU_[[:space:]]Complete[[:space:]]3D[[:space:]]human[[:space:]]shape[[:space:]]from[[:space:]]monocular[[:space:]]unconstrained[[:space:]]images/5adea58c-9270-4078-881c-e504ff34f51f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CPM_[[:space:]]Class-conditional[[:space:]]Prompting[[:space:]]Machine[[:space:]]for[[:space:]]Audio-visual[[:space:]]Segmentation/462b4fbe-3dda-4385-88f4-05811a4623a9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CPT-VR_[[:space:]]Improving[[:space:]]Surface[[:space:]]Rendering[[:space:]]via[[:space:]]Closest[[:space:]]Point[[:space:]]Transform[[:space:]]with[[:space:]]View-Reflection[[:space:]]Appearance/255265d5-d4ab-4d8b-8210-adf9209a31be_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CRM_[[:space:]]Single[[:space:]]Image[[:space:]]to[[:space:]]3D[[:space:]]Textured[[:space:]]Mesh[[:space:]]with[[:space:]]Convolutional[[:space:]]Reconstruction[[:space:]]Model/9e4b61b2-e1dd-4f99-9ab9-4fe8376f780a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CSOT_[[:space:]]Cross-Scan[[:space:]]Object[[:space:]]Transfer[[:space:]]for[[:space:]]Semi-Supervised[[:space:]]LiDAR[[:space:]]Object[[:space:]]Detection/4b43039d-68da-4e38-a8cf-39348329a3bc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CTRLorALTer_[[:space:]]Conditional[[:space:]]LoRAdapter[[:space:]]for[[:space:]]Efficient[[:space:]]0-Shot[[:space:]]Control[[:space:]]&[[:space:]]Altering[[:space:]]of[[:space:]]T2I[[:space:]]Models/d556ef6f-c8cf-4800-af84-90cf37eef09f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CVT-Occ_[[:space:]]Cost[[:space:]]Volume[[:space:]]Temporal[[:space:]]Fusion[[:space:]]for[[:space:]]3D[[:space:]]Occupancy[[:space:]]Prediction/d0af9ebe-697b-4a4d-bc71-225f5f0fd15a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CadVLM_[[:space:]]Bridging[[:space:]]Language[[:space:]]and[[:space:]]Vision[[:space:]]in[[:space:]]the[[:space:]]Generation[[:space:]]of[[:space:]]Parametric[[:space:]]CAD[[:space:]]Sketches/7ca1dca8-97a9-4690-a292-e4d8a7eacc0d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CaesarNeRF_[[:space:]]Calibrated[[:space:]]Semantic[[:space:]]Representation[[:space:]]for[[:space:]]Few-Shot[[:space:]]Generalizable[[:space:]]Neural[[:space:]]Rendering/df1f2576-4362-497d-ae9b-cdaf433e83ec_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Caltech[[:space:]]Aerial[[:space:]]RGB-Thermal[[:space:]]Dataset[[:space:]]in[[:space:]]the[[:space:]]Wild/377537ff-2360-4d17-b986-8337083e74f6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Camera[[:space:]]Calibration[[:space:]]using[[:space:]]a[[:space:]]Collimator[[:space:]]System/26f230b2-04ba-4e24-b664-739a6c3feb85_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Camera[[:space:]]Height[[:space:]]Doesn't[[:space:]]Change_[[:space:]]Unsupervised[[:space:]]Training[[:space:]]for[[:space:]]Metric[[:space:]]Monocular[[:space:]]Road-Scene[[:space:]]Depth[[:space:]]Estimation/b1560f9a-1722-4ac2-a48d-d70d8859d7c8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Camera-LiDAR[[:space:]]Cross-modality[[:space:]]Gait[[:space:]]Recognition/ce94d361-28a2-4f35-a27c-eca979dcca0e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/CamoTeacher_[[:space:]]Dual-Rotation[[:space:]]Consistency[[:space:]]Learning[[:space:]]for[[:space:]]Semi-Supervised[[:space:]]Camouflaged[[:space:]]Object[[:space:]]Detection/e7b0e9c5-9645-43e1-9e04-1cc9804f36e6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Can[[:space:]]OOD[[:space:]]Object[[:space:]]Detectors[[:space:]]Learn[[:space:]]from[[:space:]]Foundation[[:space:]]Models_/6d5daec6-da83-48aa-809b-d1c5b6b09141_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Can[[:space:]]Textual[[:space:]]Semantics[[:space:]]Mitigate[[:space:]]Sounding[[:space:]]Object[[:space:]]Segmentation[[:space:]]Preference_/88d8a449-254f-4c3c-bdf6-605afa985fb1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Canonical[[:space:]]Shape[[:space:]]Projection[[:space:]]is[[:space:]]All[[:space:]]You[[:space:]]Need[[:space:]]for[[:space:]]3D[[:space:]]Few-shot[[:space:]]Class[[:space:]]Incremental[[:space:]]Learning/354b1d2e-f84c-4d0d-a2a6-9954d66577ef_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2024/Bones Can't Be Triangles_ Accurate and Efficient Vertebrae Keypoint Estimation through Collaborative Error Revision/79ad0f6a-e5ca-4e70-bd32-205034f5cde0_content_list.json b/2024/Bones Can't Be Triangles_ Accurate and Efficient Vertebrae Keypoint Estimation through Collaborative Error Revision/79ad0f6a-e5ca-4e70-bd32-205034f5cde0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..27400d6e0995715e55acb96bea6a3f372ef948a0 --- /dev/null +++ b/2024/Bones Can't Be Triangles_ Accurate and Efficient Vertebrae Keypoint Estimation through Collaborative Error Revision/79ad0f6a-e5ca-4e70-bd32-205034f5cde0_content_list.json @@ -0,0 +1,1729 @@ +[ + { + "type": "text", + "text": "Bones Can't Be Triangles: Accurate and Efficient Vertebrae Keypoint Estimation through Collaborative Error Revision", + "text_level": 1, + "bbox": [ + 220, + 141, + 781, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinhee Kim\\*1, Taesung Kim\\*1, and Jaegul Choo", + "bbox": [ + 303, + 234, + 697, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Korea Advanced Institute of Science and Technology {seharanul17, zkm1989, jchoo}@kaist.ac.kr * indicates equal contributions.", + "bbox": [ + 323, + 262, + 676, + 303 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ca6445e836a61c2d145b6f13cfaf09dea4fef77f9263271ae1fd7975ee058164.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 306, + 352, + 457 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/132500992809c6c3958818e7a42e4c2489ba2eadd87c6a10cb23401559e95777.jpg", + "image_caption": [ + "(a) Image" + ], + "image_footnote": [], + "bbox": [ + 217, + 459, + 352, + 659 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/847cf46633eefd9c685d4c73531b621ad8a2e6f72b7db6627f1282407879e2cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 306, + 496, + 457 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9685d7166195f6848cb71988cfdf95e39ea470e09feba1f1b3dc9713fceb971d.jpg", + "image_caption": [ + "(b) Initial prediction" + ], + "image_footnote": [], + "bbox": [ + 361, + 459, + 496, + 659 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3cd1e50afc61a263e3f4f2abc30cc90802eb3743068ba91f9beb550ee72f7eb8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 306, + 640, + 457 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0fedc50436e78b8a35f4b85c7791dab4b8cbe58fbb911a6aa16df3a7898fb8af.jpg", + "image_caption": [ + "(c) KeyBot" + ], + "image_footnote": [], + "bbox": [ + 504, + 459, + 640, + 659 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/998da9e1a72fe3c716e25d0022c4dba31e48eece06842be5e6ef5a0d7f740960.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 306, + 785, + 457 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aef0418b5efadb9b952b22f025ff400f2320b2c6109aba5707dda6a9a68b6109.jpg", + "image_caption": [ + "(d) Groundtruth", + "Fig. 1: Vertebrae keypoint refinement results on BUU-LA [11] and AASCE [24]. Initial predictions exhibit significant errors in vertebra shapes due to difficulty identifying individual vertebrae. KeyBot effectively reduces these errors without user input." + ], + "image_footnote": [], + "bbox": [ + 651, + 459, + 785, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recent advances in interactive keypoint estimation methods have enhanced accuracy while minimizing user intervention. However, these methods require user input for error correction, which can be costly in vertebrae keypoint estimation where inaccurate keypoints are densely clustered or overlap. We introduce a novel approach, KeyBot, specifically designed to identify and correct significant and typical errors in existing models, akin to user revision. By characterizing typical error types", + "bbox": [ + 259, + 742, + 740, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and using simulated errors for training, KeyBot effectively corrects these errors and significantly reduces user workload. Comprehensive quantitative and qualitative evaluations on three public datasets confirm that KeyBot significantly outperforms existing methods, achieving state-of-the-art performance in interactive vertebrae keypoint estimation. The source code and demo video are available on our project page.", + "bbox": [ + 259, + 147, + 743, + 231 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keywords: Interactive vertebrae keypoint estimation $\\cdot$ X-ray image analysis $\\cdot$ Automatic error correction $\\cdot$ User interaction", + "bbox": [ + 261, + 243, + 743, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 295, + 375, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Accurate vertebrae keypoint estimation from X-ray images is crucial for effective medical diagnosis and treatment planning [20, 23], with errors having significant impacts on clinical decisions, such as in spinal surgeries and pain management [6, 13, 15]. This task presents a distinct set of challenges due to the complex anatomy of the spine [17, 27]. The similar and repetitive shapes of vertebrae complicate distinguishing between individual vertebrae and accurately estimating their keypoints, as shown in Fig. 1. This requires a deep understanding of sequential context and anatomy, unlike more distinct keypoints found in human pose [5, 9, 14, 19] or facial keypoint [1, 3] estimation tasks.", + "bbox": [ + 212, + 325, + 785, + 462 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Such challenges often result in significant errors during automated vertebrae keypoint estimation, including vertex misidentification, bone misidentification, and left-right inversion, as illustrated in Fig. 2. Manual revision of these errors is time-consuming and labor-intensive, especially as the number of vertebrae increases. Missing even a single vertebra can lead to extensive adjustments, thereby significantly increasing the time and effort required for accurate annotation. To mitigate these, interactive keypoint estimation approaches have been developed, significantly improving performance with minimal user input [10,26]. These methods aim to automatically refine inaccurate keypoints with minimal user feedback, using the feedback as a hint to adjust the remaining errors, as shown in Fig. 3(a). However, they heavily rely on users to identify and correct errors, which can be time-consuming and labor-intensive, especially when erroneous keypoints are densely clustered or overlapped.", + "bbox": [ + 212, + 462, + 787, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In response, we introduce a novel method named KeyBot, designed to autonomously detect and correct specific errors in vertebrae keypoint estimation before user evaluation, building upon existing interactive keypoint estimation models, as illustrated in Fig. 3(b). Initially, the interaction model predicts keypoints based on the input image. During refinement, KeyBot autonomously preevaluates these predictions, identifies specific errors, and provides corrective feedback similar to user input. In contrast to previous models requiring at least one user input, KeyBot autonomously refines model predictions before user input.", + "bbox": [ + 212, + 659, + 787, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We generate synthetic data characterizing typical error types, vertex misidentification, bone misidentification, and left-right inversion, and train KeyBot using this data to focus specifically on predefined error types. Independently trained from the interaction model, KeyBot effectively identifies errors that might be", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4713e0e8374727d2eb2389713f0aac795f3b598075edb2a36d4b2b870ccc9894.jpg", + "image_caption": [ + "(a) Vertex misidentification" + ], + "image_footnote": [], + "bbox": [ + 267, + 143, + 406, + 345 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b48c1d1d2c6796b0c4f9d01233f3f962c4c7ed249243c2388fd6210c1f811e7f.jpg", + "image_caption": [ + "(b) Bone misidentification" + ], + "image_footnote": [], + "bbox": [ + 433, + 143, + 571, + 345 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/16ef5afecc64764dd1aace2df0b321480e920542e5dfa64e9c730b76afef6c9b.jpg", + "image_caption": [ + "(c) Left-right inversion", + "Fig. 2: Error types in vertebrae keypoint estimation, illustrated on the AASCE dataset." + ], + "image_footnote": [], + "bbox": [ + 601, + 143, + 740, + 345 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "overlooked by the interaction model, significantly reducing the user's revision efforts. Rather than correcting errors from scratch, users can focus on verifying and fine-tuning results, with KeyBot having already eliminated major errors.", + "bbox": [ + 212, + 415, + 784, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "KeyBot delivers quick, cost-effective, and iterative feedback without user input, which is crucial for complex cases requiring multiple refinements. This automated feedback is typically much faster than user review, significantly enhancing workflow efficiency. The collaborative feedback loop between the user, KeyBot, and the interaction model enriches the process by combining expert medical insights with KeyBot's focus on typical error types. This synergy addresses a broad range of potential errors and enhancements, leading to an overall improvement in keypoint estimation accuracy.", + "bbox": [ + 212, + 460, + 785, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Through extensive experiments on three public datasets [11, 24], KeyBot demonstrates significant improvements in vertebrae keypoint estimation accuracy. The model reduces the mean radial error (MRE) by $19\\%$ and decreases the number of user clicks (NoC) required for target performance by $17\\%$ on the AASCE dataset, achieving state-of-the-art performance. These results highlight the significance of KeyBot in advancing vertebrae keypoint estimation, offering an advanced, user-centric solution to this unique and complex challenge.", + "bbox": [ + 212, + 582, + 785, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our contributions are: (1) We introduce a novel method, KeyBot, designed to provide automated corrections for specific errors in vertebrae keypoint estimation. This enhances existing interactive keypoint estimation frameworks and improves overall keypoint estimation accuracy and efficiency. (2) We train KeyBot on synthetic data representing common error types in vertebrae, enabling it to effectively identify and address these errors. (3) KeyBot significantly reduces the need for user intervention during the keypoint annotation process. (4) Extensive evaluation across three public datasets confirms the efficacy of our proposed method in improving keypoint estimation accuracy and significantly reducing user intervention.", + "bbox": [ + 212, + 688, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Bones Can't Be Triangles", + "bbox": [ + 558, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/123c58c08b8eb78e2db6d319ad6fb374a38f2f4682ac3ad59224e3551e70f5a5.jpg", + "image_caption": [ + "Fig. 3: Overview of (a) the interactive keypoint estimation and (b) the KeyBot framework. The interaction model generates initial keypoint predictions from an image, which are then revised by users, leading to updated results. This approach requires user input for error correction. In contrast, KeyBot offers a cost-efficient feedback mechanism via automated, rapid, and iterative refinement without user input. It independently identifies and corrects major errors, allowing users to focus on the final adjustments." + ], + "image_footnote": [], + "bbox": [ + 215, + 146, + 774, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 560, + 382, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Automated error refinement. Despite the advancements in deep learning, errors in model predictions remain a formidable challenge. To mitigate these inaccuracies, a specialized area of research dedicated to refining prediction results has emerged, leading to the development of advanced methods for identifying and correcting errors in model outputs [4,14,16,25]. For example, Qu et al. [16] introduced a method leveraging attention maps to detect label errors in CT data, markedly improving annotation efficiency and facilitating the collection of extensive datasets. Similarly, efforts to refine keypoint estimation have been noteworthy. For instance, Ronchi et al. [18] analyzed human pose estimation results to categorize three main types of errors, while Moon et al. [14] developed a method that specifically addresses and corrects these categorized errors, achieving state-of-the-art human pose estimation performances.", + "bbox": [ + 212, + 597, + 787, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite these advancements, estimating vertebrae keypoints from X-ray images introduces additional challenges due to the complex sequential structure of the spine. Error patterns, such as the mislocalization of an entire vertebra, often require extensive user inputs for correction. In response, we propose KeyBot,", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a method tailored for revising errors specific to vertebrae keypoint estimation, effectively addressing the unique challenges of this task.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "User interaction. To maximize the accuracy of predictions while minimizing the need for user input, the development of interaction-based models is paramount. Interactive segmentation is a particularly active area within this domain. Backpropagating refinement schemes such as BRS [8] and f-BRS [21] have improved interactive image segmentation by incorporating user hints through backpropagation. Similarly, RITM [22] has shown superior performance by integrating an iterative refinement process into its training process. In the field of keypoint estimation, interactive keypoint estimation methods have been researched to reduce user input through interactive refinement frameworks. Kim et al. [10] pioneered an interactive framework for keypoint estimation, significantly improving the accuracy and efficiency of the keypoint annotation process in medical datasets. ClickPose [26] developed an interactive keypoint estimation model for multi-person 2D keypoints, significantly reducing annotation costs.", + "bbox": [ + 212, + 176, + 785, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recently, Liu et al. [12] introduced a method that generates pseudo clicks to mimic human interactions for refining segmentation predictions autonomously, thereby minimizing the required number of user clicks. However, their error decoder is trained end-to-end and can thus be affected by inherent biases or limitations of the segmentation module. To address this challenge, our KeyBot is trained independently to correct simulated errors, which allows KeyBot to provide corrective feedback free from any potential biases present in the keypoint estimation model. Unlike traditional interactive keypoint estimation models that rely on users to identify and correct errors, KeyBot autonomously generates corrective feedback. This significant innovation not only enhances the keypoint estimation accuracy but also remarkably reduces the need for user interaction, thereby streamlining the error correction process.", + "bbox": [ + 212, + 373, + 787, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 KeyBot", + "text_level": 1, + "bbox": [ + 215, + 577, + 328, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce KeyBot, a novel method designed to enhance vertebrae keypoint estimation by identifying and correcting errors prior to user intervention. We employ an existing interactive model that can incorporate user feedback for interactive modifications. Given model predictions, KeyBot identifies erroneous keypoints and generates their pseudo-corrections, which are then fed to the interaction model for updates. KeyBot consists of two components: The detector discerns whether each input keypoint is accurate, and the corrector refines input keypoints accurately. Both components are trained with synthetic errors to focus specifically on the target error types.", + "bbox": [ + 212, + 609, + 787, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Overview of the propose approach", + "text_level": 1, + "bbox": [ + 214, + 768, + 545, + 784 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Interactive keypoint estimation framework. Consider an interactive keypoint estimation model $\\mathcal{F}_{\\theta}$ with parameters $\\theta$ , operating on an image $\\pmb{x}$ . The model aims to precisely estimate $K$ target keypoints through user-interactive", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Bones Can't Be Triangles", + "bbox": [ + 558, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e9134607dfc9aa107aa066d901484079bff86056ec607e6a9a6c70af4c985650.jpg", + "image_caption": [ + "(a)", + "Fig. 4: Training overview of KeyBot. KeyBot consists of two main components, the detector and the corrector. (a) Synthetic errors are introduced to the groundtruth keypoints, creating inaccurate keypoints and corresponding anomaly labels. (b) The detector is trained to discern whether each input keypoint is accurate or not. (c) The corrector is trained to refine these inaccurate keypoints accurately." + ], + "image_footnote": [], + "bbox": [ + 251, + 142, + 748, + 376 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "refinement. In each iteration $t$ , the model generates the $t$ -th prediction, $\\pmb{y}_t$ , using three inputs: the image $\\pmb{x}$ , user revisions $\\pmb{u}_t$ , and its previous predictions $\\pmb{e}_t$ :", + "bbox": [ + 214, + 491, + 787, + 522 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} _ {t} = \\mathcal {F} _ {\\theta} (\\boldsymbol {x}, \\boldsymbol {u} _ {t}, \\boldsymbol {e} _ {t}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 536, + 784, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\pmb{u}_0\\coloneqq 0$ and $\\pmb {e}_0\\coloneqq 0$ . Subsequently, a user reviews the results, and the adjustment made to a keypoint position is encoded as Gaussian heatmaps $\\pmb{u}_t^i$ , effectively capturing and representing the refined locations:", + "bbox": [ + 214, + 558, + 787, + 604 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\rho} _ {t + 1}, \\boldsymbol {u} _ {t + 1} = \\Psi (\\boldsymbol {x}, \\boldsymbol {y} _ {t}, \\boldsymbol {u} _ {t}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 614, + 784, + 631 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the set $\\rho_{t}$ represents keypoint indices revised by the user over $t$ iterations. Next, we selectively gather the known-to-be-false predictions from $\\pmb{y}$ to $\\pmb{e}$ and reintroduce them as inputs in the next step to differentiate mispredictions [10]:", + "bbox": [ + 214, + 642, + 787, + 688 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {e} _ {t + 1} ^ {i} = \\boldsymbol {y} _ {t} ^ {i} \\text {f o r} i \\in \\boldsymbol {\\rho} _ {t}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 696, + 784, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $\\pmb{e}$ denotes predictions modified by the user. These updates on $\\pmb{u}_t$ and $\\pmb{e}_t$ are critical for refining the model predictions, leading to the updated keypoint predictions reflecting the user feedback:", + "bbox": [ + 214, + 726, + 787, + 771 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} _ {t + 1} = \\mathcal {F} _ {\\theta} (\\boldsymbol {x}, \\boldsymbol {u} _ {t + 1}, \\boldsymbol {e} _ {t + 1}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 782, + 784, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This process repeats until sufficient accuracy is achieved or the predefined maximum user interaction $T$ is reached.", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The KeyBot framework. KeyBot offers an automated keypoint refinement step that precedes and complements the existing user-based refinement, effectively reducing the need for user interaction in the process. KeyBot consists of the detector $\\mathcal{G}_{\\phi_r}$ and corrector $\\mathcal{G}_{\\phi_s}$ , which together identifies and corrects inaccurate keypoints predicted by $\\mathcal{F}_{\\theta}$ . Following each prediction iteration of $\\mathcal{F}_{\\theta}$ , KeyBot evaluates $\\pmb{y}_t$ and suggests necessary corrections, identifying erroneous keypoints $\\pmb{\\nu}_{t,n+1}$ and their pseudo corrections $\\pmb{z}_{t,n+1}$ :", + "bbox": [ + 214, + 146, + 787, + 253 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\nu} _ {t, n + 1} = \\mathcal {G} _ {\\phi_ {r}} (\\boldsymbol {x}, \\boldsymbol {y} _ {t, n}) \\text {a n d} \\boldsymbol {z} _ {t, n + 1} = \\mathcal {G} _ {\\phi_ {s}} (\\boldsymbol {x}, \\boldsymbol {y} _ {t, n}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 260, + 785, + 277 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with $\\pmb{y}_{t,0} \\coloneqq \\pmb{y}_t$ . This step is similar to the user interaction step outlined in Eq. (2). Here, KeyBot can detect multiple inaccurate keypoints at once and simultaneously offer corrective feedback for each identified error. If no inaccuracies are detected by the detector, KeyBot concludes and $\\pmb{y}_{t,n}$ becomes the final prediction. Revision information, $\\pmb{c}_{t,n}$ , now includes both user $(\\pmb{u}_{t,n})$ and model $(\\pmb{z}_{t,n})$ corrections. Here, user modifications are prioritized as the definitive ground truth. Thus, keypoints already revised by the user are exempt from further updates by the pseudo-corrections from KeyBot.", + "bbox": [ + 214, + 282, + 787, + 404 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {c} _ {t, n + 1} ^ {i} = \\boldsymbol {z} _ {t, n + 1} ^ {i} \\text {f o r} i \\in \\boldsymbol {\\nu} _ {t, n + 1} \\backslash \\boldsymbol {\\rho} _ {t}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 382, + 410, + 785, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Previous work [10,22] relies solely on the prediction at the most recent step as the previous prediction, as described in Eq. (3). However, this approach can result in losing track of erroneous predictions once they are corrected in the model. Hence, our approach accumulates predictions identified as erroneous by either KeyBot or the user across all iterations. These false predictions retain the history of incorrect predictions and serve to inform the model about specific prediction errors to avoid, thereby guiding it toward more accurate future predictions:", + "bbox": [ + 214, + 433, + 787, + 540 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {e} _ {t, n + 1} ^ {i} = \\boldsymbol {y} _ {t, n} ^ {i} \\text {f o r} i \\in \\boldsymbol {\\nu} _ {t, n + 1} \\backslash \\boldsymbol {\\nu} _ {t,: n} \\backslash \\boldsymbol {\\rho} _ {t}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 545, + 785, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, $e$ denotes predictions modified by either the user or KeyBot. Finally, these are then fed back into the backbone interaction network for a new, corrected prediction akin to user feedback:", + "bbox": [ + 214, + 568, + 787, + 614 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} _ {t, n + 1} = \\mathcal {F} _ {\\theta} (\\boldsymbol {x}, \\boldsymbol {c} _ {t, n + 1}, \\boldsymbol {e} _ {t, n + 1}). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 400, + 622, + 785, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Each iteration of KeyBot contributes progressively towards enhancing the overall accuracy and reliability of the keypoint predictions. This cycle repeats until the detector detects no errors or the maximum KeyBot iteration count $N$ is reached. Upon the KeyBot process concludes, say, after $n + 1$ iterations, the prediction $\\pmb{y}_{t,n + 1}$ is established as the final output for the $t$ -th iteration of the interactive keypoint estimation process. For the subsequent $(t + 1)$ -th user-side phase, given new user feedback as Eq. (2), $\\pmb{c}_{t,n + 1}$ and $\\pmb{e}_{t,n + 1}$ are updated to $\\pmb{c}_{t + 1,0}$ and $\\pmb{e}_{t + 1,0}$ ,", + "bbox": [ + 214, + 643, + 787, + 751 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {c} _ {t + 1, 0} = \\boldsymbol {u} _ {t + 1} \\text {a n d} \\boldsymbol {e} _ {t + 1, 0} ^ {i} = \\boldsymbol {y} _ {t, n + 1} ^ {i} \\text {f o r} i \\in \\boldsymbol {\\rho} _ {t + 1}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 757, + 785, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and the keypoints are revised accordingly:", + "bbox": [ + 214, + 779, + 522, + 795 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} _ {t + 1, 0} = \\mathcal {F} _ {\\theta} (\\boldsymbol {x}, \\boldsymbol {c} _ {t + 1, 0}, \\boldsymbol {e} _ {t + 1, 0}). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 803, + 785, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The complete procedure is summarized in Algorithm 1 in Appendix B.", + "bbox": [ + 214, + 824, + 725, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Bones Can't Be Triangles", + "bbox": [ + 558, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/465d142995c3e4d69f7368bba681efcd8c913cc9152991d1b7a4a46a67f9945e.jpg", + "image_caption": [ + "Fig. 5: Implementing realistic estimation inaccuracy simulations for training KeyBot." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 784, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.2 Model design of KeyBot", + "text_level": 1, + "bbox": [ + 215, + 407, + 464, + 424 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "KeyBot integrates the detector and corrector to identify potential errors and generates corrective feedback for their revision.", + "bbox": [ + 212, + 435, + 782, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Error detector. The error detector specializes in identifying up to $K$ erroneous keypoints, focusing on prevalent error types in vertebrae keypoint estimation. Our empirical analysis indicates that evaluating the entire bone structure at once is less effective for error detection. Therefore, the detector evaluates keypoints in smaller groups, examining $k \\leq K$ keypoints simultaneously. This approach simplifies the task and improves its performance. During training, we randomly select a set of consecutive groundtruth keypoints $\\mathbf{y}^*$ and either maintain their accuracy or introduce errors. The cropped image area around these keypoints, along with their corresponding Gaussian keypoint heatmaps, are input to the detector. It then predicts the likelihood of error presence in each keypoint.", + "bbox": [ + 212, + 465, + 784, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Let $\\varepsilon$ denote the function that induces synthetic errors in groundtruth keypoints, and $v$ represents the anomaly labels for the keypoints. The detector performs multi-label binary classification to differentiate between normal and abnormal keypoints using a sigmoid function in its output layer. It is supervised to minimize the Binary Cross-Entropy (BCE) loss between the anomaly labels $v$ and the predictions:", + "bbox": [ + 212, + 617, + 784, + 705 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {r} \\leftarrow \\phi_ {r} - \\eta_ {r} \\nabla_ {\\phi_ {r}} \\mathcal {L} _ {b c e} (v, \\mathcal {G} _ {\\phi_ {r}} (\\boldsymbol {x}, \\varepsilon (\\boldsymbol {y} ^ {*})), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 720, + 785, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with $\\eta_r$ representing the learning rate for the detector.", + "bbox": [ + 215, + 750, + 609, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During inference, the detector iteratively detects errors across all keypoints, examining $k$ keypoints at a time with a stride of $s$ . To facilitate a comprehensive assessment of each keypoint in different contexts, we set the stride $s$ to be smaller than $k$ . This results in multiple predictions for a single keypoint, and if any prediction indicates an anomaly, the keypoint is flagged as erroneous.", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Error corrector. The error corrector aims to generate pseudo-corrections for keypoints inaccurately identified by the detector, thus providing feedback to the model $\\mathcal{F}_{\\theta}$ . During training, we randomly transform the groundtruth keypoints to erroneous ones using $\\varepsilon$ . Taking these keypoints and the image $\\pmb{x}$ as inputs, the corrector is trained to revert these to their original, accurate state, focusing particularly on the typical error types implemented by $\\varepsilon$ . It generates $K$ -channel heatmaps indicating reconstructed keypoint locations, employing a sigmoid function in its output. Its learning objective is the BCE loss between the ground truth keypoint heatmaps and the reconstructed outputs:", + "bbox": [ + 214, + 146, + 787, + 282 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {s} \\leftarrow \\phi_ {s} - \\eta_ {s} \\nabla_ {\\phi_ {s}} \\mathcal {L} _ {b c e} (\\boldsymbol {y}, \\mathcal {G} _ {\\phi_ {s}} (\\boldsymbol {x}, \\varepsilon (\\boldsymbol {y} ^ {*})), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 295, + 785, + 311 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\eta_{s}$ is the learning rate for the corrector.", + "bbox": [ + 214, + 323, + 550, + 338 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "During inference, the corrector refines predictions from the backbone model $\\mathcal{F}_{\\theta}$ , and the output corresponding to inaccurately assessed keypoints by the detector is fed to $\\mathcal{F}_{\\theta}$ to serve as pseudo-corrections.", + "bbox": [ + 214, + 338, + 785, + 383 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.3 Training KeyBot via realistic error simulation", + "text_level": 1, + "bbox": [ + 214, + 406, + 640, + 422 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "KeyBot aims to address significant and typical errors in vertebrae keypoint estimation. The key idea is to train KeyBot to identify and correct three error types commonly observed in vertebrae keypoint estimation: vertex misidentification, bone misidentification, and left-right inversion. During training, KeyBot is provided with both accurate keypoints and keypoints with synthetic errors. We randomly simulate the predefined error types, as shown in Fig. 5. Rather than relying on highly variable and not clearly defined real errors, we use characterized and consistent synthetic errors.", + "bbox": [ + 212, + 431, + 787, + 551 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "By employing this training approach, we expose KeyBot to a wide range of error scenarios, equipping it with the proficiency to detect and correct these errors effectively. As a result, KeyBot provides a targeted evaluation of the keypoints predicted by the interaction model, focusing on specific error types.", + "bbox": [ + 212, + 551, + 787, + 613 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Vertex misidentification (misvertex). Misvertex is a type of error where a keypoint is correctly identified on a vertex of a vertebrae but is incorrectly attributed to the wrong vertex. For instance, a keypoint intended for the first vertex is inaccurately predicted at the third vertex. To simulate misvertex, we randomly displace a keypoint to its neighbors within a $[-r, +r]$ index range for a predefined number $r$ , excluding its original position. Keypoints moving beyond their index range wrap around cyclically.", + "bbox": [ + 214, + 613, + 787, + 719 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Bone Misidentification (misbone). Misbone errors represent errors where the entire vertebra is incorrectly classified. Our simulation randomly shifts keypoints up or down by one vertebra. This is achieved by selecting start and end keypoint indices and moving all-encompassed keypoints to an adjacent vertebra.", + "bbox": [ + 214, + 719, + 787, + 780 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Left-right inversion (lr-inversion). LR-inversion occurs when keypoint pairs are incorrectly positioned on the opposite lateral side. To simulate LR-inversion in our model, we randomly interchange the positions of keypoint pairs that delineate one side of each vertebra.", + "bbox": [ + 214, + 779, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Bones Can't Be Triangles", + "bbox": [ + 558, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/a55ed82ebc087172044aaadf5902ed130e81367a5c31fbbe144b62f26621aca9.jpg", + "table_caption": [ + "Table 1: Performance comparison of mean radial error in keypoint estimation across three datasets. UC denotes the count of user clicks provided to the model." + ], + "table_footnote": [], + "table_body": "
MethodAASCEBUU-APBUU-LA
-UC1UC2UC3UC4-UC1UC2-UC1UC2
BRS [8]45.6535.8829.9325.4021.9751.2230.9136.1740.2043.1935.07
f-BRS [21]64.0657.3052.5347.8343.7944.0524.9917.5236.0327.9319.29
RITM [22]56.0337.9631.0326.7623.7936.4325.7617.9623.2717.799.89
Kim et al. [10]51.5830.6025.7821.6019.0842.3123.2616.4623.4314.2910.29
Click-Pose [26]54.6546.5044.0841.7340.0432.7229.3026.2033.7021.6217.38
KeyBot-i144.1825.9320.6618.0316.3732.0121.7915.9718.7713.399.12
KeyBot-i242.5227.0322.9518.7216.2331.8820.6515.8119.1113.479.03
KeyBot-i341.7026.5925.0221.2316.7631.8720.6615.8418.7413.368.97
", + "bbox": [ + 228, + 183, + 767, + 347 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 386, + 374, + 402 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 419, + 426, + 435 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. We rigorously validate our method across three publicly available real-world datasets. The AASCE dataset [24] consists of 608 spinal anterior-posterior radiographs, each annotated with 68 keypoints, marking the vertices of 17 vertebrae. The BUU dataset [11] features two distinct subsets from 400 unique patients: BUU-AP with 400 anterior-posterior (AP) view radiographs, each annotated with 20 keypoints, and BUU-LA with 400 left lateral (LA) view radiographs, each annotated with 22 spinal keypoints. Images that exhibit critical keypoint annotation errors, such as those with incorrectly ordered keypoint indices similar to typical errors in keypoint estimation models, are excluded from the datasets. Comprehensive details about the dataset, including the data split and examples of the excluded images, are available in Appendix C.", + "bbox": [ + 212, + 446, + 785, + 611 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Metrics. Following the evaluation protocol of previous work [10, 26], we evaluate the model performance using the mean radial error (MRE) and the number of clicks (NoC). MRE computes the mean Euclidean distance between the groundtruth and predicted keypoints. No $C_a$ @ $b$ measures the average number of user clicks required to achieve a target MRE of $b$ , given a maximum of $a$ clicks.", + "bbox": [ + 212, + 612, + 785, + 686 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baseline models. We compare our method with state-of-the-art interactive keypoint estimation models, including the model proposed by Kim et al. [10] and Click-Pose [26]. Additionally, we use interactive image segmentation models such as BRS [8], f-BRS [21], and RITM [22] as baselines.", + "bbox": [ + 212, + 688, + 785, + 750 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation details. We utilize ResNet [7]-18 and DeepLab-V3 [2] as the backbone for the detector and corrector, respectively. We build on the pretrained model of Kim et al. [10], which is designed to estimate keypoints based on user feedback. All of the models are trained independently. We apply the misvertex error type to train the detector and all three error types for the corrector. KeyBot-i1, KeyBot-i2, and KeyBot-i3 refer to KeyBot with one, two, and three", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/5f48cc8ce8f498422611c98de846b0e31b969d6bb1c5c12f3d05eca6beff6188.jpg", + "table_caption": [ + "Table 2: Comparison of required number of user clicks to reach target performance on the AASCE, BUU-AP, and BUU-LA datasets." + ], + "table_footnote": [], + "table_body": "
MethodAASCEBUU-APBUU-LA
NoC10@20NoC10@30NoC10@40NoC10@50NoC5@15NoC5@20NoC5@25NoC5@30NoC5@15NoC5@20NoC5@25NoC5@30
BRS [8]2.531.971.511.232.801.510.920.613.892.261.340.86
f-BRS [21]7.165.304.213.482.231.180.680.481.240.610.440.40
RITM [22]3.202.371.871.381.050.580.420.300.570.240.170.14
Kim et al. [10]2.101.541.231.030.940.540.370.320.380.150.140.12
Click-Pose [26]4.153.563.202.971.200.890.760.711.450.700.560.45
KeyBot-i11.841.260.920.530.870.510.330.280.340.100.060.05
KeyBot-i21.781.210.910.610.870.480.300.240.350.110.060.05
KeyBot-i31.741.320.940.680.870.490.320.240.340.100.060.05
", + "bbox": [ + 230, + 183, + 769, + 349 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "iterations, respectively. Unless explicitly stated, the experimental results are conducted without user interaction. The training and inference are conducted on a single NVIDIA GeForce RTX 3090 GPU with 24GB memory. More details on experimental settings, including training details, are provided in Appendix C.", + "bbox": [ + 212, + 388, + 787, + 450 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Experimental results and analysis", + "text_level": 1, + "bbox": [ + 214, + 468, + 540, + 484 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison with state-of-the-art methods. In this section, we present a comparative analysis of interactive keypoint estimation performance between KeyBot and established baseline models by increasing the maximum KeyBot iteration count $N$ from one to three. As shown in Table 1, KeyBot achieves state-of-the-art performance in mean radial error (MRE) across the AASCE, BUU-AP, and BUU-LA datasets. KeyBot consistently outperforms baseline models both in initial predictions and during subsequent user interactions. Remarkably, KeyBot with a single user click (UC1) shows superior performance compared to BRS, f-BRS, RITM, and Click-Pose with two user clicks (UC2) on the AASCE dataset. Furthermore, the performance of KeyBot improves with additional iterations, demonstrating its ability to refine predictions through multiple correction cycles. These results highlight the effectiveness of KeyBot in enhancing keypoint estimation accuracy and reducing user intervention.", + "bbox": [ + 212, + 492, + 787, + 688 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In terms of the number of clicks (NoC), as shown in Table 2, KeyBot exhibits superior performance compared to the baseline models across all three datasets. For instance, on the AASCE dataset, our proposed model nearly halves the NoC required to achieve a target MRE of 50 ( $\\mathrm{NoC}_{10} @ 50$ ) compared to that of the baseline models. This substantial reduction in NoC underscores KeyBot's proficiency in interactive frameworks, significantly lowering annotation costs.", + "bbox": [ + 212, + 688, + 789, + 780 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Analysis on the detector. We investigate the effectiveness of the detector in identifying erroneous keypoints within images. In Fig. 6, we provide (a) the number of images where the detector identifies no errors versus those where errors are detected, as well as (b) the distribution of their actual mean radial error.", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Bones Can't Be Triangles", + "bbox": [ + 558, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/68e2aeac048d4f26decab5ca04079d817be705d917ef8f32c2bb855b0cdeab9e.jpg", + "image_caption": [ + "Fig.6: Effectiveness of the detector in identifying errors on the AASCE dataset." + ], + "image_footnote": [], + "bbox": [ + 253, + 143, + 532, + 281 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/ef18d7f108047369bb8757460a3d680b56ac32e75ee3613df7d466cfc6d0fab3.jpg", + "table_caption": [ + "Table 3: Keypoint prediction errors on the AASCE dataset, comparing MRE for examples where the detector identifies errors versus where it does not." + ], + "table_footnote": [], + "table_body": "
MethodNo error detectedError detected
Initial prediction28.9964.88
KeyBot-i1-53.12
KeyBot-i2-50.49
KeyBot-i3-49.18
", + "bbox": [ + 545, + 228, + 743, + 319 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Out of the total images analyzed, 43 images have no errors detected, while 73 images have errors identified. The images classified as \"No error detected\" exhibit significantly lower errors compared to those classified as \"Error detected.\" These findings indicate the precision and effectiveness of the detector in distinguishing between erroneous predictions and accurate ones.", + "bbox": [ + 212, + 352, + 782, + 428 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We further examine error reduction through successive KeyBot iterations for images where errors are initially detected. As demonstrated in Table 3, KeyBot effectively enhances the accuracy of these error-detected images, significantly reducing errors in subsequent iterations. Notably, after three iterations of KeyBot, the MRE significantly decreases from 64.88 to 49.18. This reduction highlights KeyBot's ability to iteratively identify inaccuracies and improve predictions.", + "bbox": [ + 212, + 429, + 784, + 520 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation study. This section validates each component of KeyBot. First, we study the role of the detector under two scenarios: (i) employing corrector outputs directly as the final results and (ii) utilizing all corrector outputs as pseudocorrections for the interaction model. As detailed in Table 4, both scenarios exhibit significant performance degradation compared to KeyBot, underscoring the importance of using the detector to selectively apply corrector outputs.", + "bbox": [ + 212, + 520, + 784, + 611 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Additionally, utilizing only the interactive keypoint estimation network, without the support of KeyBot, results in a significant performance degradation compared to the refined predictions with KeyBot. When ablating the error types simulated during the training of the corrector, our findings show that training the corrector to recognize and rectify three distinct types of errors significantly improves its prediction accuracy compared to when it is trained on only one error type. This finding highlights the necessity of incorporating all three typical errors into the training process.", + "bbox": [ + 212, + 612, + 784, + 733 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative results. In Fig. 7, we analyze the qualitative results of KeyBot with three iterations. The results demonstrate that KeyBot effectively identifies and reduces all three error types: vertex misidentification (Misvertex), bone misidentification (Misbone), and left-right inversion (LR-inversion). For these errors, KeyBot provides corrective feedback to the interactive model, enabling necessary adjustments. For instance, in Fig. 7(a), the initial prediction skips a bone, resulting in a Misbone error (red), and also exhibits a Misvertex error", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/112aa05676ba1c1718c6bcbaed33d21ac397a67c22272a57e8607e7fa6a6c272.jpg", + "table_caption": [ + "Table 4: Ablation study of our keypoint refinement method on the AASCE dataset." + ], + "table_footnote": [], + "table_body": "
Interaction modelKeyBotMean radial error
DetectorCorrectorError type-UC1UC2UC3
✓✓✓all44.1825.9320.6618.03
XX✓all50.9939.1732.9830.36
✓X✓all57.6145.9439.1136.01
✓XX-51.5830.6025.5721.07
✓✓✓misvertex52.6731.1425.9221.65
✓✓✓misbone49.0729.8523.4320.75
✓✓✓lr-inversion52.3830.1625.1320.90
", + "bbox": [ + 272, + 170, + 725, + 325 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f8a01f1b7f685cbdefd8c45e6fc67b2d0c9e11cbe1977c9c63b91dde52df3074.jpg", + "image_caption": [ + "(a) misvertex+misbone (MRE: $71.7\\rightarrow 11.9$" + ], + "image_footnote": [], + "bbox": [ + 220, + 352, + 496, + 492 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f48d389d76dda02e382386fb0e9c14c0299ed4d9349fb1bdb3da9820b02fc4f5.jpg", + "image_caption": [ + "(b) misvertex+misbone (MRE: 91.9 → 20.2)" + ], + "image_footnote": [], + "bbox": [ + 504, + 352, + 782, + 492 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/24533e71014362e7a8f3f8e268de202f96af687e578696373c94e40e411c1a7d.jpg", + "image_caption": [ + "(c) misvertex+misbone (MRE: 193.3 → 15.3)", + "Image Initial prediction KeyBot" + ], + "image_footnote": [], + "bbox": [ + 220, + 502, + 496, + 642 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ca6b4f126fbb2f2d2c6876befd69939cd9175e3c631c7f82631f59aec355b680.jpg", + "image_caption": [ + "(d) misvertex+Ir-inversion (MRE: 443.5 → 40.3)", + "Image Initial prediction KeyBot", + "Fig. 7: Prediction results on the (a) AASCE, (b) BUU-LA, and (c,d) BUU-AP (bottom) datasets. KeyBot effectively reduces all three types of errors, even in challenging cases where error types are mixed. The keypoints classified as inaccurate by our detector are indicated with red dots, while their corresponding keypoints in the revised results are visualized as blue dots. A diagonal line connecting the upper-right vertex to the lower-left vertex is indicated in each vertebra. More results are in Appendix F." + ], + "image_footnote": [], + "bbox": [ + 504, + 502, + 782, + 643 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "(purple). KeyBot accurately refines these errors, significantly reducing the MRE from 71.7 to 11.9. The results highlight KeyBot's capability to significantly enhance keypoint estimation accuracy without user clicks.", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Bones Can't Be Triangles", + "bbox": [ + 558, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7d487f580d16bbb59390c9e49d8798c928bcd715e8ca7486605c8e0152f011c3.jpg", + "image_caption": [ + "Fig. 8: Collaborative keypoint refinement process among the interaction model, KeyBot, and the user. Faced with substantial errors in initial predictions, KeyBot effectively removes all major errors, allowing users to easily identify and correct remaining inaccuracies with minimal effort. Here, refinement is achieved with just two user clicks." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 282 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Additionally, KeyBot excels in interpretability. It includes a separate detector that enables users to easily comprehend the keypoints identified as errors and the corrections made, leading to efficient evaluation of the results. In the given examples, identified errors are indicated by red dots, and the corrections are shown as blue dots, illustrating how the keypoints are adjusted. This transparency contributes to improving the efficiency of the overall annotation process. Collaborative refinement process. In Fig. 8, we showcase the qualitative results of the collaborative refinement process involving both KeyBot and the user. Our approach establishes a feedback loop among the interaction model, KeyBot, and the user, combining expert medical knowledge and a targeted focus on specific error types. This significantly enhances the overall accuracy and efficiency of the annotation process.", + "bbox": [ + 212, + 378, + 787, + 561 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 585, + 359, + 599 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduce a novel approach, KeyBot, specifically designed to identify and correct significant errors in vertebrae keypoint estimation. Our proposed error simulation training procedure enables KeyBot to accurately and independently address specific error types, thereby significantly reducing the need for user intervention in the annotation process. Comprehensive evaluations on public real-world datasets confirm the superiority of KeyBot over existing methods.", + "bbox": [ + 212, + 616, + 784, + 708 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 215, + 732, + 392, + 750 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was supported by Institute for Information & communications Technology Promotion (IITP) grant funded by the Korea government (MSIT) (No.RS2019-II190075 Artificial Intelligence Graduate School Program (KAIST)) and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. NRF-2022R1A2B5B02001913).", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Albiero, V., Chen, X., Yin, X., Pang, G., Hassner, T.: img2pose: Face alignment and detection via 6dof, face pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7617-7627 (2021)", + "2. Chen, L.C., Papandreou, G., Schroff, F., Adam, H.: Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587 (2017)", + "3. Colaco, S.J., Han, D.S.: Deep learning-based facial landmarks localization using compound scaling. IEEE Access 10, 7653-7663 (2022)", + "4. Daniele, A., van Krieken, E., Serafini, L., van Harmelen, F.: Refining neural network predictions using background knowledge. Machine Learning pp. 1-39 (2023)", + "5. Guo, W., Du, Y., Shen, X., Lepetit, V., Alameda-Pineda, X., Moreno-Noguer, F.: Back to mlp: A simple baseline for human motion prediction. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4809-4819 (2023)", + "6. Guo, Y., Li, Y., Zhou, X., He, W.: A keypoint transformer to discover spine structure for cobb angle estimation. In: 2021 IEEE International Conference on Multimedia and Expo (ICME). pp. 1-6. IEEE (2021)", + "7. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)", + "8. Jang, W.D., Kim, C.S.: Interactive image segmentation via backpropagating refinement scheme. CVPR pp. 5297-5306 (2019)", + "9. Jin, S., Xu, L., Xu, J., Wang, C., Liu, W., Qian, C., Ouyang, W., Luo, P.: Whole-body human pose estimation in the wild. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16. pp. 196-214. Springer (2020)", + "10. Kim, J., Kim, T., Kim, T., Choo, J., Kim, D.W., Ahn, B., Song, I.S., Kim, Y.J.: Morphology-aware interactive keypoint estimation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 675-685. Springer (2022)", + "1. Klinwichit, P., Yookwan, W., Limchareon, S., Chinnasarn, K., Jang, J.S., Onuean, A.: Buu-lspine: A thai open lumbar spine dataset for spondylolisthesis detection. Applied Sciences 13(15), 8646 (2023)", + "2. Liu, Q., Zheng, M., Planche, B., Karanam, S., Chen, T., Niethammer, M., Wu, Z.: Pseudoclick: Interactive image segmentation with click imitation. In: European Conference on Computer Vision. pp. 728-745. Springer (2022)", + "3. Liu, T., Wang, Y., Yang, Y., Sun, M., Fan, W., Bunger, C., Wu, C.: A multi-scale keypoint estimation network with self-supervision for spinal curvature assessment of idiopathic scoliosis from the imperfect dataset. Artificial Intelligence in Medicine 125, 102235 (2022)", + "4. Moon, G., Chang, J.Y., Lee, K.M.: Posefix: Model-agnostic general human pose refinement network. CVPR pp. 7773-7781 (2019)", + "5. Pisov, M., Kondratenko, V., Zakharov, A., Petraikin, A., Gombolevskiy, V., Morozov, S., Belyaev, M.: Keypoints localization for joint vertebra detection and fracture severity quantification. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part VI 23. pp. 723-732. Springer (2020)", + "6. Qu, C., Zhang, T., Qiao, H., Tang, Y., Yuille, A.L., Zhou, Z., et al.: Abdomenatlas-8k: Annotating 8,000 ct volumes for multi-organ segmentation in three weeks. Advances in Neural Information Processing Systems 36 (2024)" + ], + "bbox": [ + 225, + 175, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Bones Can't Be Triangles", + "bbox": [ + 558, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Rahmaniar, W., Suzuki, K., Lin, T.L.: Auto-ca: automated cobb angle measurement based on vertebrae detection for assessment of spinal curvature deformity. IEEE Transactions on Biomedical Engineering (2023)", + "18. Ruggero Ronchi, M., Perona, P.: Benchmarking and error diagnosis in multi-instance pose estimation. In: Proceedings of the IEEE international conference on computer vision. pp. 369-378 (2017)", + "19. Shi, D., Wei, X., Li, L., Ren, Y., Tan, W.: End-to-end multi-person pose estimation with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11069-11078 (2022)", + "20. Singh, S., Sandhu, N., Puri, T., Gulati, R., Kashyap, R.: A study of correlation of various growth indicators with chronological age. International Journal of Clinical Pediatric Dentistry 8(3), 190 (2015)", + "21. Sofiuk, K., Petrov, I., Barinova, O., Konushin, A.: f-brs: Rethinking backpropagating refinement for interactive segmentation. CVPR pp. 8623-8632 (2020)", + "22. Sofiuk, K., Petrov, I.A., Konushin, A.: Reviving iterative training with mask guidance for interactive segmentation. In: 2022 IEEE International Conference on Image Processing (ICIP). pp. 3141-3145. IEEE (2022)", + "23. Türkoz, C., Kaygirisiz, E., Ulusoy, C., Ates, C.: A practical formula for determining growth. Diagnostic and Interventional Radiology 23(3), 194 (2017)", + "24. Wu, H., Bailey, C., Rasoulinejad, P., Li, S.: Automatic landmark estimation for adolescent idiopathic scoliosis assessment using boostnet. In: MICCAI. pp. 127-135. Springer (2017)", + "25. Xu, Y., Xu, D., Hong, X., Ouyang, W., Ji, R., Xu, M., Zhao, G.: Structured modeling of joint deep feature and prediction refinement for salient object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3789-3798 (2019)", + "26. Yang, J., Zeng, A., Li, F., Liu, S., Zhang, R., Zhang, L.: Neural interactive keypoint detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15122-15132 (2023)", + "27. Zhang, H., Chung, A.C.: A dual coordinate system vertebral landmark detection network with sparse-to-dense vertebral line interpolation. Bioengineering 11(1), 101 (2024)" + ], + "bbox": [ + 215, + 147, + 784, + 590 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/2024/Bones Can't Be Triangles_ Accurate and Efficient Vertebrae Keypoint Estimation through Collaborative Error Revision/79ad0f6a-e5ca-4e70-bd32-205034f5cde0_model.json b/2024/Bones Can't Be Triangles_ Accurate and Efficient Vertebrae Keypoint Estimation through Collaborative Error Revision/79ad0f6a-e5ca-4e70-bd32-205034f5cde0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..34a64b7163718ad0b9d86040ed2cc2a5e4cbd4f2 --- /dev/null +++ b/2024/Bones Can't Be Triangles_ Accurate and Efficient Vertebrae Keypoint Estimation through Collaborative Error Revision/79ad0f6a-e5ca-4e70-bd32-205034f5cde0_model.json @@ -0,0 +1,2190 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.221, + 0.142, + 0.782, + 0.21 + ], + "angle": 0, + "content": "Bones Can't Be Triangles: Accurate and Efficient Vertebrae Keypoint Estimation through Collaborative Error Revision" + }, + { + "type": "text", + "bbox": [ + 0.305, + 0.235, + 0.699, + 0.253 + ], + "angle": 0, + "content": "Jinhee Kim\\*1, Taesung Kim\\*1, and Jaegul Choo" + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.263, + 0.678, + 0.304 + ], + "angle": 0, + "content": "Korea Advanced Institute of Science and Technology {seharanul17, zkm1989, jchoo}@kaist.ac.kr * indicates equal contributions." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.307, + 0.353, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.46, + 0.353, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.256, + 0.662, + 0.315, + 0.675 + ], + "angle": 0, + "content": "(a) Image" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.307, + 0.497, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.46, + 0.497, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.374, + 0.661, + 0.486, + 0.675 + ], + "angle": 0, + "content": "(b) Initial prediction" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.307, + 0.642, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.46, + 0.642, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.661, + 0.616, + 0.675 + ], + "angle": 0, + "content": "(c) KeyBot" + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.307, + 0.787, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.46, + 0.787, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.673, + 0.661, + 0.764, + 0.675 + ], + "angle": 0, + "content": "(d) Groundtruth" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.685, + 0.788, + 0.729 + ], + "angle": 0, + "content": "Fig. 1: Vertebrae keypoint refinement results on BUU-LA [11] and AASCE [24]. Initial predictions exhibit significant errors in vertebra shapes due to difficulty identifying individual vertebrae. KeyBot effectively reduces these errors without user input." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.743, + 0.741, + 0.841 + ], + "angle": 0, + "content": "Abstract. Recent advances in interactive keypoint estimation methods have enhanced accuracy while minimizing user intervention. However, these methods require user input for error correction, which can be costly in vertebrae keypoint estimation where inaccurate keypoints are densely clustered or overlap. We introduce a novel approach, KeyBot, specifically designed to identify and correct significant and typical errors in existing models, akin to user revision. By characterizing typical error types" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.148, + 0.744, + 0.232 + ], + "angle": 0, + "content": "and using simulated errors for training, KeyBot effectively corrects these errors and significantly reduces user workload. Comprehensive quantitative and qualitative evaluations on three public datasets confirm that KeyBot significantly outperforms existing methods, achieving state-of-the-art performance in interactive vertebrae keypoint estimation. The source code and demo video are available on our project page." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.244, + 0.744, + 0.273 + ], + "angle": 0, + "content": "Keywords: Interactive vertebrae keypoint estimation \\(\\cdot\\) X-ray image analysis \\(\\cdot\\) Automatic error correction \\(\\cdot\\) User interaction" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.296, + 0.376, + 0.312 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.787, + 0.463 + ], + "angle": 0, + "content": "Accurate vertebrae keypoint estimation from X-ray images is crucial for effective medical diagnosis and treatment planning [20, 23], with errors having significant impacts on clinical decisions, such as in spinal surgeries and pain management [6, 13, 15]. This task presents a distinct set of challenges due to the complex anatomy of the spine [17, 27]. The similar and repetitive shapes of vertebrae complicate distinguishing between individual vertebrae and accurately estimating their keypoints, as shown in Fig. 1. This requires a deep understanding of sequential context and anatomy, unlike more distinct keypoints found in human pose [5, 9, 14, 19] or facial keypoint [1, 3] estimation tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.659 + ], + "angle": 0, + "content": "Such challenges often result in significant errors during automated vertebrae keypoint estimation, including vertex misidentification, bone misidentification, and left-right inversion, as illustrated in Fig. 2. Manual revision of these errors is time-consuming and labor-intensive, especially as the number of vertebrae increases. Missing even a single vertebra can lead to extensive adjustments, thereby significantly increasing the time and effort required for accurate annotation. To mitigate these, interactive keypoint estimation approaches have been developed, significantly improving performance with minimal user input [10,26]. These methods aim to automatically refine inaccurate keypoints with minimal user feedback, using the feedback as a hint to adjust the remaining errors, as shown in Fig. 3(a). However, they heavily rely on users to identify and correct errors, which can be time-consuming and labor-intensive, especially when erroneous keypoints are densely clustered or overlapped." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.78 + ], + "angle": 0, + "content": "In response, we introduce a novel method named KeyBot, designed to autonomously detect and correct specific errors in vertebrae keypoint estimation before user evaluation, building upon existing interactive keypoint estimation models, as illustrated in Fig. 3(b). Initially, the interaction model predicts keypoints based on the input image. During refinement, KeyBot autonomously preevaluates these predictions, identifies specific errors, and provides corrective feedback similar to user input. In contrast to previous models requiring at least one user input, KeyBot autonomously refines model predictions before user input." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We generate synthetic data characterizing typical error types, vertex misidentification, bone misidentification, and left-right inversion, and train KeyBot using this data to focus specifically on predefined error types. Independently trained from the interaction model, KeyBot effectively identifies errors that might be" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.56, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bones Can't Be Triangles" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.145, + 0.408, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.348, + 0.409, + 0.358 + ], + "angle": 0, + "content": "(a) Vertex misidentification" + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.144, + 0.573, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.432, + 0.348, + 0.571, + 0.358 + ], + "angle": 0, + "content": "(b) Bone misidentification" + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.144, + 0.741, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.348, + 0.733, + 0.359 + ], + "angle": 0, + "content": "(c) Left-right inversion" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.369, + 0.785, + 0.384 + ], + "angle": 0, + "content": "Fig. 2: Error types in vertebrae keypoint estimation, illustrated on the AASCE dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.785, + 0.461 + ], + "angle": 0, + "content": "overlooked by the interaction model, significantly reducing the user's revision efforts. Rather than correcting errors from scratch, users can focus on verifying and fine-tuning results, with KeyBot having already eliminated major errors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.787, + 0.583 + ], + "angle": 0, + "content": "KeyBot delivers quick, cost-effective, and iterative feedback without user input, which is crucial for complex cases requiring multiple refinements. This automated feedback is typically much faster than user review, significantly enhancing workflow efficiency. The collaborative feedback loop between the user, KeyBot, and the interaction model enriches the process by combining expert medical insights with KeyBot's focus on typical error types. This synergy addresses a broad range of potential errors and enhancements, leading to an overall improvement in keypoint estimation accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.787, + 0.689 + ], + "angle": 0, + "content": "Through extensive experiments on three public datasets [11, 24], KeyBot demonstrates significant improvements in vertebrae keypoint estimation accuracy. The model reduces the mean radial error (MRE) by \\(19\\%\\) and decreases the number of user clicks (NoC) required for target performance by \\(17\\%\\) on the AASCE dataset, achieving state-of-the-art performance. These results highlight the significance of KeyBot in advancing vertebrae keypoint estimation, offering an advanced, user-centric solution to this unique and complex challenge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In summary, our contributions are: (1) We introduce a novel method, KeyBot, designed to provide automated corrections for specific errors in vertebrae keypoint estimation. This enhances existing interactive keypoint estimation frameworks and improves overall keypoint estimation accuracy and efficiency. (2) We train KeyBot on synthetic data representing common error types in vertebrae, enabling it to effectively identify and address these errors. (3) KeyBot significantly reduces the need for user intervention during the keypoint annotation process. (4) Extensive evaluation across three public datasets confirms the efficacy of our proposed method in improving keypoint estimation accuracy and significantly reducing user intervention." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.128 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.147, + 0.775, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.446, + 0.788, + 0.531 + ], + "angle": 0, + "content": "Fig. 3: Overview of (a) the interactive keypoint estimation and (b) the KeyBot framework. The interaction model generates initial keypoint predictions from an image, which are then revised by users, leading to updated results. This approach requires user input for error correction. In contrast, KeyBot offers a cost-efficient feedback mechanism via automated, rapid, and iterative refinement without user input. It independently identifies and corrects major errors, allowing users to focus on the final adjustments." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.561, + 0.383, + 0.576 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Automated error refinement. Despite the advancements in deep learning, errors in model predictions remain a formidable challenge. To mitigate these inaccuracies, a specialized area of research dedicated to refining prediction results has emerged, leading to the development of advanced methods for identifying and correcting errors in model outputs [4,14,16,25]. For example, Qu et al. [16] introduced a method leveraging attention maps to detect label errors in CT data, markedly improving annotation efficiency and facilitating the collection of extensive datasets. Similarly, efforts to refine keypoint estimation have been noteworthy. For instance, Ronchi et al. [18] analyzed human pose estimation results to categorize three main types of errors, while Moon et al. [14] developed a method that specifically addresses and corrects these categorized errors, achieving state-of-the-art human pose estimation performances." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Despite these advancements, estimating vertebrae keypoints from X-ray images introduces additional challenges due to the complex sequential structure of the spine. Error patterns, such as the mislocalization of an entire vertebra, often require extensive user inputs for correction. In response, we propose KeyBot," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.56, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bones Can't Be Triangles" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "a method tailored for revising errors specific to vertebrae keypoint estimation, effectively addressing the unique challenges of this task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.373 + ], + "angle": 0, + "content": "User interaction. To maximize the accuracy of predictions while minimizing the need for user input, the development of interaction-based models is paramount. Interactive segmentation is a particularly active area within this domain. Backpropagating refinement schemes such as BRS [8] and f-BRS [21] have improved interactive image segmentation by incorporating user hints through backpropagation. Similarly, RITM [22] has shown superior performance by integrating an iterative refinement process into its training process. In the field of keypoint estimation, interactive keypoint estimation methods have been researched to reduce user input through interactive refinement frameworks. Kim et al. [10] pioneered an interactive framework for keypoint estimation, significantly improving the accuracy and efficiency of the keypoint annotation process in medical datasets. ClickPose [26] developed an interactive keypoint estimation model for multi-person 2D keypoints, significantly reducing annotation costs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.555 + ], + "angle": 0, + "content": "Recently, Liu et al. [12] introduced a method that generates pseudo clicks to mimic human interactions for refining segmentation predictions autonomously, thereby minimizing the required number of user clicks. However, their error decoder is trained end-to-end and can thus be affected by inherent biases or limitations of the segmentation module. To address this challenge, our KeyBot is trained independently to correct simulated errors, which allows KeyBot to provide corrective feedback free from any potential biases present in the keypoint estimation model. Unlike traditional interactive keypoint estimation models that rely on users to identify and correct errors, KeyBot autonomously generates corrective feedback. This significant innovation not only enhances the keypoint estimation accuracy but also remarkably reduces the need for user interaction, thereby streamlining the error correction process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.578, + 0.33, + 0.596 + ], + "angle": 0, + "content": "3 KeyBot" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.788, + 0.747 + ], + "angle": 0, + "content": "We introduce KeyBot, a novel method designed to enhance vertebrae keypoint estimation by identifying and correcting errors prior to user intervention. We employ an existing interactive model that can incorporate user feedback for interactive modifications. Given model predictions, KeyBot identifies erroneous keypoints and generates their pseudo-corrections, which are then fed to the interaction model for updates. KeyBot consists of two components: The detector discerns whether each input keypoint is accurate, and the corrector refines input keypoints accurately. Both components are trained with synthetic errors to focus specifically on the target error types." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.77, + 0.547, + 0.785 + ], + "angle": 0, + "content": "3.1 Overview of the propose approach" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Interactive keypoint estimation framework. Consider an interactive keypoint estimation model \\(\\mathcal{F}_{\\theta}\\) with parameters \\(\\theta\\), operating on an image \\(\\pmb{x}\\). The model aims to precisely estimate \\(K\\) target keypoints through user-interactive" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.219, + 0.169, + 0.241, + 0.182 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.143, + 0.749, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.392, + 0.789, + 0.463 + ], + "angle": 0, + "content": "Fig. 4: Training overview of KeyBot. KeyBot consists of two main components, the detector and the corrector. (a) Synthetic errors are introduced to the groundtruth keypoints, creating inaccurate keypoints and corresponding anomaly labels. (b) The detector is trained to discern whether each input keypoint is accurate or not. (c) The corrector is trained to refine these inaccurate keypoints accurately." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.492, + 0.789, + 0.523 + ], + "angle": 0, + "content": "refinement. In each iteration \\( t \\), the model generates the \\( t \\)-th prediction, \\( \\pmb{y}_t \\), using three inputs: the image \\( \\pmb{x} \\), user revisions \\( \\pmb{u}_t \\), and its previous predictions \\( \\pmb{e}_t \\):" + }, + { + "type": "equation", + "bbox": [ + 0.432, + 0.537, + 0.785, + 0.553 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} _ {t} = \\mathcal {F} _ {\\theta} (\\boldsymbol {x}, \\boldsymbol {u} _ {t}, \\boldsymbol {e} _ {t}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.559, + 0.788, + 0.605 + ], + "angle": 0, + "content": "where \\(\\pmb{u}_0\\coloneqq 0\\) and \\(\\pmb {e}_0\\coloneqq 0\\). Subsequently, a user reviews the results, and the adjustment made to a keypoint position is encoded as Gaussian heatmaps \\(\\pmb{u}_t^i\\), effectively capturing and representing the refined locations:" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.616, + 0.785, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\rho} _ {t + 1}, \\boldsymbol {u} _ {t + 1} = \\Psi (\\boldsymbol {x}, \\boldsymbol {y} _ {t}, \\boldsymbol {u} _ {t}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.643, + 0.788, + 0.689 + ], + "angle": 0, + "content": "where the set \\(\\rho_{t}\\) represents keypoint indices revised by the user over \\(t\\) iterations. Next, we selectively gather the known-to-be-false predictions from \\(\\pmb{y}\\) to \\(\\pmb{e}\\) and reintroduce them as inputs in the next step to differentiate mispredictions [10]:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.698, + 0.785, + 0.716 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {e} _ {t + 1} ^ {i} = \\boldsymbol {y} _ {t} ^ {i} \\text {f o r} i \\in \\boldsymbol {\\rho} _ {t}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.727, + 0.788, + 0.772 + ], + "angle": 0, + "content": "Here, \\( \\pmb{e} \\) denotes predictions modified by the user. These updates on \\( \\pmb{u}_t \\) and \\( \\pmb{e}_t \\) are critical for refining the model predictions, leading to the updated keypoint predictions reflecting the user feedback:" + }, + { + "type": "equation", + "bbox": [ + 0.414, + 0.784, + 0.785, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} _ {t + 1} = \\mathcal {F} _ {\\theta} (\\boldsymbol {x}, \\boldsymbol {u} _ {t + 1}, \\boldsymbol {e} _ {t + 1}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "This process repeats until sufficient accuracy is achieved or the predefined maximum user interaction \\( T \\) is reached." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.56, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Bones Can't Be Triangles" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.255 + ], + "angle": 0, + "content": "The KeyBot framework. KeyBot offers an automated keypoint refinement step that precedes and complements the existing user-based refinement, effectively reducing the need for user interaction in the process. KeyBot consists of the detector \\(\\mathcal{G}_{\\phi_r}\\) and corrector \\(\\mathcal{G}_{\\phi_s}\\), which together identifies and corrects inaccurate keypoints predicted by \\(\\mathcal{F}_{\\theta}\\). Following each prediction iteration of \\(\\mathcal{F}_{\\theta}\\), KeyBot evaluates \\(\\pmb{y}_t\\) and suggests necessary corrections, identifying erroneous keypoints \\(\\pmb{\\nu}_{t,n+1}\\) and their pseudo corrections \\(\\pmb{z}_{t,n+1}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.261, + 0.786, + 0.278 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\nu} _ {t, n + 1} = \\mathcal {G} _ {\\phi_ {r}} (\\boldsymbol {x}, \\boldsymbol {y} _ {t, n}) \\text {a n d} \\boldsymbol {z} _ {t, n + 1} = \\mathcal {G} _ {\\phi_ {s}} (\\boldsymbol {x}, \\boldsymbol {y} _ {t, n}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.283, + 0.789, + 0.405 + ], + "angle": 0, + "content": "with \\(\\pmb{y}_{t,0} \\coloneqq \\pmb{y}_t\\). This step is similar to the user interaction step outlined in Eq. (2). Here, KeyBot can detect multiple inaccurate keypoints at once and simultaneously offer corrective feedback for each identified error. If no inaccuracies are detected by the detector, KeyBot concludes and \\(\\pmb{y}_{t,n}\\) becomes the final prediction. Revision information, \\(\\pmb{c}_{t,n}\\), now includes both user \\((\\pmb{u}_{t,n})\\) and model \\((\\pmb{z}_{t,n})\\) corrections. Here, user modifications are prioritized as the definitive ground truth. Thus, keypoints already revised by the user are exempt from further updates by the pseudo-corrections from KeyBot." + }, + { + "type": "equation", + "bbox": [ + 0.383, + 0.411, + 0.786, + 0.429 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {c} _ {t, n + 1} ^ {i} = \\boldsymbol {z} _ {t, n + 1} ^ {i} \\text {f o r} i \\in \\boldsymbol {\\nu} _ {t, n + 1} \\backslash \\boldsymbol {\\rho} _ {t}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.434, + 0.788, + 0.541 + ], + "angle": 0, + "content": "Previous work [10,22] relies solely on the prediction at the most recent step as the previous prediction, as described in Eq. (3). However, this approach can result in losing track of erroneous predictions once they are corrected in the model. Hence, our approach accumulates predictions identified as erroneous by either KeyBot or the user across all iterations. These false predictions retain the history of incorrect predictions and serve to inform the model about specific prediction errors to avoid, thereby guiding it toward more accurate future predictions:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.546, + 0.786, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {e} _ {t, n + 1} ^ {i} = \\boldsymbol {y} _ {t, n} ^ {i} \\text {f o r} i \\in \\boldsymbol {\\nu} _ {t, n + 1} \\backslash \\boldsymbol {\\nu} _ {t,: n} \\backslash \\boldsymbol {\\rho} _ {t}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.569, + 0.788, + 0.616 + ], + "angle": 0, + "content": "Here, \\( e \\) denotes predictions modified by either the user or KeyBot. Finally, these are then fed back into the backbone interaction network for a new, corrected prediction akin to user feedback:" + }, + { + "type": "equation", + "bbox": [ + 0.401, + 0.623, + 0.786, + 0.64 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} _ {t, n + 1} = \\mathcal {F} _ {\\theta} (\\boldsymbol {x}, \\boldsymbol {c} _ {t, n + 1}, \\boldsymbol {e} _ {t, n + 1}). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.645, + 0.788, + 0.752 + ], + "angle": 0, + "content": "Each iteration of KeyBot contributes progressively towards enhancing the overall accuracy and reliability of the keypoint predictions. This cycle repeats until the detector detects no errors or the maximum KeyBot iteration count \\(N\\) is reached. Upon the KeyBot process concludes, say, after \\(n + 1\\) iterations, the prediction \\(\\pmb{y}_{t,n + 1}\\) is established as the final output for the \\(t\\)-th iteration of the interactive keypoint estimation process. For the subsequent \\((t + 1)\\)-th user-side phase, given new user feedback as Eq. (2), \\(\\pmb{c}_{t,n + 1}\\) and \\(\\pmb{e}_{t,n + 1}\\) are updated to \\(\\pmb{c}_{t + 1,0}\\) and \\(\\pmb{e}_{t + 1,0}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.758, + 0.786, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {c} _ {t + 1, 0} = \\boldsymbol {u} _ {t + 1} \\text {a n d} \\boldsymbol {e} _ {t + 1, 0} ^ {i} = \\boldsymbol {y} _ {t, n + 1} ^ {i} \\text {f o r} i \\in \\boldsymbol {\\rho} _ {t + 1}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.524, + 0.796 + ], + "angle": 0, + "content": "and the keypoints are revised accordingly:" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.804, + 0.786, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} _ {t + 1, 0} = \\mathcal {F} _ {\\theta} (\\boldsymbol {x}, \\boldsymbol {c} _ {t + 1, 0}, \\boldsymbol {e} _ {t + 1, 0}). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.825, + 0.726, + 0.842 + ], + "angle": 0, + "content": "The complete procedure is summarized in Algorithm 1 in Appendix B." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.145, + 0.785, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.364, + 0.785, + 0.379 + ], + "angle": 0, + "content": "Fig. 5: Implementing realistic estimation inaccuracy simulations for training KeyBot." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.409, + 0.465, + 0.425 + ], + "angle": 0, + "content": "3.2 Model design of KeyBot" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.784, + 0.464 + ], + "angle": 0, + "content": "KeyBot integrates the detector and corrector to identify potential errors and generates corrective feedback for their revision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.785, + 0.617 + ], + "angle": 0, + "content": "Error detector. The error detector specializes in identifying up to \\( K \\) erroneous keypoints, focusing on prevalent error types in vertebrae keypoint estimation. Our empirical analysis indicates that evaluating the entire bone structure at once is less effective for error detection. Therefore, the detector evaluates keypoints in smaller groups, examining \\( k \\leq K \\) keypoints simultaneously. This approach simplifies the task and improves its performance. During training, we randomly select a set of consecutive groundtruth keypoints \\( \\mathbf{y}^* \\) and either maintain their accuracy or introduce errors. The cropped image area around these keypoints, along with their corresponding Gaussian keypoint heatmaps, are input to the detector. It then predicts the likelihood of error presence in each keypoint." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.618, + 0.785, + 0.707 + ], + "angle": 0, + "content": "Let \\(\\varepsilon\\) denote the function that induces synthetic errors in groundtruth keypoints, and \\(v\\) represents the anomaly labels for the keypoints. The detector performs multi-label binary classification to differentiate between normal and abnormal keypoints using a sigmoid function in its output layer. It is supervised to minimize the Binary Cross-Entropy (BCE) loss between the anomaly labels \\(v\\) and the predictions:" + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.721, + 0.786, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\phi_ {r} \\leftarrow \\phi_ {r} - \\eta_ {r} \\nabla_ {\\phi_ {r}} \\mathcal {L} _ {b c e} (v, \\mathcal {G} _ {\\phi_ {r}} (\\boldsymbol {x}, \\varepsilon (\\boldsymbol {y} ^ {*})), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.75, + 0.61, + 0.764 + ], + "angle": 0, + "content": "with \\(\\eta_r\\) representing the learning rate for the detector." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "During inference, the detector iteratively detects errors across all keypoints, examining \\( k \\) keypoints at a time with a stride of \\( s \\). To facilitate a comprehensive assessment of each keypoint in different contexts, we set the stride \\( s \\) to be smaller than \\( k \\). This results in multiple predictions for a single keypoint, and if any prediction indicates an anomaly, the keypoint is flagged as erroneous." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.56, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bones Can't Be Triangles" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "Error corrector. The error corrector aims to generate pseudo-corrections for keypoints inaccurately identified by the detector, thus providing feedback to the model \\(\\mathcal{F}_{\\theta}\\). During training, we randomly transform the groundtruth keypoints to erroneous ones using \\(\\varepsilon\\). Taking these keypoints and the image \\(\\pmb{x}\\) as inputs, the corrector is trained to revert these to their original, accurate state, focusing particularly on the typical error types implemented by \\(\\varepsilon\\). It generates \\(K\\)-channel heatmaps indicating reconstructed keypoint locations, employing a sigmoid function in its output. Its learning objective is the BCE loss between the ground truth keypoint heatmaps and the reconstructed outputs:" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.296, + 0.786, + 0.312 + ], + "angle": 0, + "content": "\\[\n\\phi_ {s} \\leftarrow \\phi_ {s} - \\eta_ {s} \\nabla_ {\\phi_ {s}} \\mathcal {L} _ {b c e} (\\boldsymbol {y}, \\mathcal {G} _ {\\phi_ {s}} (\\boldsymbol {x}, \\varepsilon (\\boldsymbol {y} ^ {*})), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.324, + 0.551, + 0.339 + ], + "angle": 0, + "content": "where \\(\\eta_{s}\\) is the learning rate for the corrector." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.339, + 0.787, + 0.385 + ], + "angle": 0, + "content": "During inference, the corrector refines predictions from the backbone model \\(\\mathcal{F}_{\\theta}\\), and the output corresponding to inaccurately assessed keypoints by the detector is fed to \\(\\mathcal{F}_{\\theta}\\) to serve as pseudo-corrections." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.407, + 0.642, + 0.423 + ], + "angle": 0, + "content": "3.3 Training KeyBot via realistic error simulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.553 + ], + "angle": 0, + "content": "KeyBot aims to address significant and typical errors in vertebrae keypoint estimation. The key idea is to train KeyBot to identify and correct three error types commonly observed in vertebrae keypoint estimation: vertex misidentification, bone misidentification, and left-right inversion. During training, KeyBot is provided with both accurate keypoints and keypoints with synthetic errors. We randomly simulate the predefined error types, as shown in Fig. 5. Rather than relying on highly variable and not clearly defined real errors, we use characterized and consistent synthetic errors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.614 + ], + "angle": 0, + "content": "By employing this training approach, we expose KeyBot to a wide range of error scenarios, equipping it with the proficiency to detect and correct these errors effectively. As a result, KeyBot provides a targeted evaluation of the keypoints predicted by the interaction model, focusing on specific error types." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.614, + 0.788, + 0.72 + ], + "angle": 0, + "content": "Vertex misidentification (misvertex). Misvertex is a type of error where a keypoint is correctly identified on a vertex of a vertebrae but is incorrectly attributed to the wrong vertex. For instance, a keypoint intended for the first vertex is inaccurately predicted at the third vertex. To simulate misvertex, we randomly displace a keypoint to its neighbors within a \\([-r, +r]\\) index range for a predefined number \\( r \\), excluding its original position. Keypoints moving beyond their index range wrap around cyclically." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.72, + 0.788, + 0.781 + ], + "angle": 0, + "content": "Bone Misidentification (misbone). Misbone errors represent errors where the entire vertebra is incorrectly classified. Our simulation randomly shifts keypoints up or down by one vertebra. This is achieved by selecting start and end keypoint indices and moving all-encompassed keypoints to an adjacent vertebra." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Left-right inversion (lr-inversion). LR-inversion occurs when keypoint pairs are incorrectly positioned on the opposite lateral side. To simulate LR-inversion in our model, we randomly interchange the positions of keypoint pairs that delineate one side of each vertebra." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 1: Performance comparison of mean radial error in keypoint estimation across three datasets. UC denotes the count of user clicks provided to the model." + }, + { + "type": "table", + "bbox": [ + 0.23, + 0.184, + 0.769, + 0.348 + ], + "angle": 0, + "content": "
MethodAASCEBUU-APBUU-LA
-UC1UC2UC3UC4-UC1UC2-UC1UC2
BRS [8]45.6535.8829.9325.4021.9751.2230.9136.1740.2043.1935.07
f-BRS [21]64.0657.3052.5347.8343.7944.0524.9917.5236.0327.9319.29
RITM [22]56.0337.9631.0326.7623.7936.4325.7617.9623.2717.799.89
Kim et al. [10]51.5830.6025.7821.6019.0842.3123.2616.4623.4314.2910.29
Click-Pose [26]54.6546.5044.0841.7340.0432.7229.3026.2033.7021.6217.38
KeyBot-i144.1825.9320.6618.0316.3732.0121.7915.9718.7713.399.12
KeyBot-i242.5227.0322.9518.7216.2331.8820.6515.8119.1113.479.03
KeyBot-i341.7026.5925.0221.2316.7631.8720.6615.8418.7413.368.97
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.387, + 0.375, + 0.404 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.42, + 0.427, + 0.436 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.787, + 0.612 + ], + "angle": 0, + "content": "Datasets. We rigorously validate our method across three publicly available real-world datasets. The AASCE dataset [24] consists of 608 spinal anterior-posterior radiographs, each annotated with 68 keypoints, marking the vertices of 17 vertebrae. The BUU dataset [11] features two distinct subsets from 400 unique patients: BUU-AP with 400 anterior-posterior (AP) view radiographs, each annotated with 20 keypoints, and BUU-LA with 400 left lateral (LA) view radiographs, each annotated with 22 spinal keypoints. Images that exhibit critical keypoint annotation errors, such as those with incorrectly ordered keypoint indices similar to typical errors in keypoint estimation models, are excluded from the datasets. Comprehensive details about the dataset, including the data split and examples of the excluded images, are available in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.787, + 0.688 + ], + "angle": 0, + "content": "Metrics. Following the evaluation protocol of previous work [10, 26], we evaluate the model performance using the mean radial error (MRE) and the number of clicks (NoC). MRE computes the mean Euclidean distance between the groundtruth and predicted keypoints. No \\( C_a \\) @ \\( b \\) measures the average number of user clicks required to achieve a target MRE of \\( b \\), given a maximum of \\( a \\) clicks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.75 + ], + "angle": 0, + "content": "Baseline models. We compare our method with state-of-the-art interactive keypoint estimation models, including the model proposed by Kim et al. [10] and Click-Pose [26]. Additionally, we use interactive image segmentation models such as BRS [8], f-BRS [21], and RITM [22] as baselines." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Implementation details. We utilize ResNet [7]-18 and DeepLab-V3 [2] as the backbone for the detector and corrector, respectively. We build on the pretrained model of Kim et al. [10], which is designed to estimate keypoints based on user feedback. All of the models are trained independently. We apply the misvertex error type to train the detector and all three error types for the corrector. KeyBot-i1, KeyBot-i2, and KeyBot-i3 refer to KeyBot with one, two, and three" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.56, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bones Can't Be Triangles" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.173 + ], + "angle": 0, + "content": "Table 2: Comparison of required number of user clicks to reach target performance on the AASCE, BUU-AP, and BUU-LA datasets." + }, + { + "type": "table", + "bbox": [ + 0.231, + 0.184, + 0.77, + 0.35 + ], + "angle": 0, + "content": "
MethodAASCEBUU-APBUU-LA
NoC10@20NoC10@30NoC10@40NoC10@50NoC5@15NoC5@20NoC5@25NoC5@30NoC5@15NoC5@20NoC5@25NoC5@30
BRS [8]2.531.971.511.232.801.510.920.613.892.261.340.86
f-BRS [21]7.165.304.213.482.231.180.680.481.240.610.440.40
RITM [22]3.202.371.871.381.050.580.420.300.570.240.170.14
Kim et al. [10]2.101.541.231.030.940.540.370.320.380.150.140.12
Click-Pose [26]4.153.563.202.971.200.890.760.711.450.700.560.45
KeyBot-i11.841.260.920.530.870.510.330.280.340.100.060.05
KeyBot-i21.781.210.910.610.870.480.300.240.350.110.060.05
KeyBot-i31.741.320.940.680.870.490.320.240.340.100.060.05
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.788, + 0.451 + ], + "angle": 0, + "content": "iterations, respectively. Unless explicitly stated, the experimental results are conducted without user interaction. The training and inference are conducted on a single NVIDIA GeForce RTX 3090 GPU with 24GB memory. More details on experimental settings, including training details, are provided in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.469, + 0.542, + 0.485 + ], + "angle": 0, + "content": "4.2 Experimental results and analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.789, + 0.689 + ], + "angle": 0, + "content": "Comparison with state-of-the-art methods. In this section, we present a comparative analysis of interactive keypoint estimation performance between KeyBot and established baseline models by increasing the maximum KeyBot iteration count \\(N\\) from one to three. As shown in Table 1, KeyBot achieves state-of-the-art performance in mean radial error (MRE) across the AASCE, BUU-AP, and BUU-LA datasets. KeyBot consistently outperforms baseline models both in initial predictions and during subsequent user interactions. Remarkably, KeyBot with a single user click (UC1) shows superior performance compared to BRS, f-BRS, RITM, and Click-Pose with two user clicks (UC2) on the AASCE dataset. Furthermore, the performance of KeyBot improves with additional iterations, demonstrating its ability to refine predictions through multiple correction cycles. These results highlight the effectiveness of KeyBot in enhancing keypoint estimation accuracy and reducing user intervention." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.79, + 0.781 + ], + "angle": 0, + "content": "In terms of the number of clicks (NoC), as shown in Table 2, KeyBot exhibits superior performance compared to the baseline models across all three datasets. For instance, on the AASCE dataset, our proposed model nearly halves the NoC required to achieve a target MRE of 50 (\\(\\mathrm{NoC}_{10} @ 50\\)) compared to that of the baseline models. This substantial reduction in NoC underscores KeyBot's proficiency in interactive frameworks, significantly lowering annotation costs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Analysis on the detector. We investigate the effectiveness of the detector in identifying erroneous keypoints within images. In Fig. 6, we provide (a) the number of images where the detector identifies no errors versus those where errors are detected, as well as (b) the distribution of their actual mean radial error." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.145, + 0.533, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.294, + 0.528, + 0.322 + ], + "angle": 0, + "content": "Fig.6: Effectiveness of the detector in identifying errors on the AASCE dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.54, + 0.149, + 0.754, + 0.217 + ], + "angle": 0, + "content": "Table 3: Keypoint prediction errors on the AASCE dataset, comparing MRE for examples where the detector identifies errors versus where it does not." + }, + { + "type": "table", + "bbox": [ + 0.546, + 0.229, + 0.745, + 0.32 + ], + "angle": 0, + "content": "
MethodNo error detectedError detected
Initial prediction28.9964.88
KeyBot-i1-53.12
KeyBot-i2-50.49
KeyBot-i3-49.18
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.784, + 0.429 + ], + "angle": 0, + "content": "Out of the total images analyzed, 43 images have no errors detected, while 73 images have errors identified. The images classified as \"No error detected\" exhibit significantly lower errors compared to those classified as \"Error detected.\" These findings indicate the precision and effectiveness of the detector in distinguishing between erroneous predictions and accurate ones." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.785, + 0.521 + ], + "angle": 0, + "content": "We further examine error reduction through successive KeyBot iterations for images where errors are initially detected. As demonstrated in Table 3, KeyBot effectively enhances the accuracy of these error-detected images, significantly reducing errors in subsequent iterations. Notably, after three iterations of KeyBot, the MRE significantly decreases from 64.88 to 49.18. This reduction highlights KeyBot's ability to iteratively identify inaccuracies and improve predictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.785, + 0.612 + ], + "angle": 0, + "content": "Ablation study. This section validates each component of KeyBot. First, we study the role of the detector under two scenarios: (i) employing corrector outputs directly as the final results and (ii) utilizing all corrector outputs as pseudocorrections for the interaction model. As detailed in Table 4, both scenarios exhibit significant performance degradation compared to KeyBot, underscoring the importance of using the detector to selectively apply corrector outputs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.785, + 0.734 + ], + "angle": 0, + "content": "Additionally, utilizing only the interactive keypoint estimation network, without the support of KeyBot, results in a significant performance degradation compared to the refined predictions with KeyBot. When ablating the error types simulated during the training of the corrector, our findings show that training the corrector to recognize and rectify three distinct types of errors significantly improves its prediction accuracy compared to when it is trained on only one error type. This finding highlights the necessity of incorporating all three typical errors into the training process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Qualitative results. In Fig. 7, we analyze the qualitative results of KeyBot with three iterations. The results demonstrate that KeyBot effectively identifies and reduces all three error types: vertex misidentification (Misvertex), bone misidentification (Misbone), and left-right inversion (LR-inversion). For these errors, KeyBot provides corrective feedback to the interactive model, enabling necessary adjustments. For instance, in Fig. 7(a), the initial prediction skips a bone, resulting in a Misbone error (red), and also exhibits a Misvertex error" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.56, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bones Can't Be Triangles" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.145, + 0.782, + 0.159 + ], + "angle": 0, + "content": "Table 4: Ablation study of our keypoint refinement method on the AASCE dataset." + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.171, + 0.727, + 0.326 + ], + "angle": 0, + "content": "
Interaction modelKeyBotMean radial error
DetectorCorrectorError type-UC1UC2UC3
✓✓✓all44.1825.9320.6618.03
XX✓all50.9939.1732.9830.36
✓X✓all57.6145.9439.1136.01
✓XX-51.5830.6025.5721.07
✓✓✓misvertex52.6731.1425.9221.65
✓✓✓misbone49.0729.8523.4320.75
✓✓✓lr-inversion52.3830.1625.1320.90
" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.353, + 0.497, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.493, + 0.472, + 0.502 + ], + "angle": 0, + "content": "(a) misvertex+misbone (MRE: \\(71.7\\rightarrow 11.9\\)" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.353, + 0.783, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.493, + 0.757, + 0.503 + ], + "angle": 0, + "content": "(b) misvertex+misbone (MRE: 91.9 → 20.2)" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.503, + 0.498, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.643, + 0.475, + 0.653 + ], + "angle": 0, + "content": "(c) misvertex+misbone (MRE: 193.3 → 15.3)" + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.657, + 0.482, + 0.67 + ], + "angle": 0, + "content": "Image Initial prediction KeyBot" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.503, + 0.783, + 0.644 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.643, + 0.766, + 0.654 + ], + "angle": 0, + "content": "(d) misvertex+Ir-inversion (MRE: 443.5 → 40.3)" + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.657, + 0.766, + 0.67 + ], + "angle": 0, + "content": "Image Initial prediction KeyBot" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.68, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Fig. 7: Prediction results on the (a) AASCE, (b) BUU-LA, and (c,d) BUU-AP (bottom) datasets. KeyBot effectively reduces all three types of errors, even in challenging cases where error types are mixed. The keypoints classified as inaccurate by our detector are indicated with red dots, while their corresponding keypoints in the revised results are visualized as blue dots. A diagonal line connecting the upper-right vertex to the lower-left vertex is indicated in each vertebra. More results are in Appendix F." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "(purple). KeyBot accurately refines these errors, significantly reducing the MRE from 71.7 to 11.9. The results highlight KeyBot's capability to significantly enhance keypoint estimation accuracy without user clicks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.294, + 0.785, + 0.351 + ], + "angle": 0, + "content": "Fig. 8: Collaborative keypoint refinement process among the interaction model, KeyBot, and the user. Faced with substantial errors in initial predictions, KeyBot effectively removes all major errors, allowing users to easily identify and correct remaining inaccuracies with minimal effort. Here, refinement is achieved with just two user clicks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.788, + 0.562 + ], + "angle": 0, + "content": "Additionally, KeyBot excels in interpretability. It includes a separate detector that enables users to easily comprehend the keypoints identified as errors and the corrections made, leading to efficient evaluation of the results. In the given examples, identified errors are indicated by red dots, and the corrections are shown as blue dots, illustrating how the keypoints are adjusted. This transparency contributes to improving the efficiency of the overall annotation process. Collaborative refinement process. In Fig. 8, we showcase the qualitative results of the collaborative refinement process involving both KeyBot and the user. Our approach establishes a feedback loop among the interaction model, KeyBot, and the user, combining expert medical knowledge and a targeted focus on specific error types. This significantly enhances the overall accuracy and efficiency of the annotation process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.586, + 0.36, + 0.601 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.617, + 0.785, + 0.709 + ], + "angle": 0, + "content": "We introduce a novel approach, KeyBot, specifically designed to identify and correct significant errors in vertebrae keypoint estimation. Our proposed error simulation training procedure enables KeyBot to accurately and independently address specific error types, thereby significantly reducing the need for user intervention in the annotation process. Comprehensive evaluations on public real-world datasets confirm the superiority of KeyBot over existing methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.733, + 0.393, + 0.75 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.841 + ], + "angle": 0, + "content": "This work was supported by Institute for Information & communications Technology Promotion (IITP) grant funded by the Korea government (MSIT) (No.RS2019-II190075 Artificial Intelligence Graduate School Program (KAIST)) and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. NRF-2022R1A2B5B02001913)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.56, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bones Can't Be Triangles" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.785, + 0.218 + ], + "angle": 0, + "content": "1. Albiero, V., Chen, X., Yin, X., Pang, G., Hassner, T.: img2pose: Face alignment and detection via 6dof, face pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7617-7627 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.245 + ], + "angle": 0, + "content": "2. Chen, L.C., Papandreou, G., Schroff, F., Adam, H.: Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.246, + 0.785, + 0.273 + ], + "angle": 0, + "content": "3. Colaco, S.J., Han, D.S.: Deep learning-based facial landmarks localization using compound scaling. IEEE Access 10, 7653-7663 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.785, + 0.3 + ], + "angle": 0, + "content": "4. Daniele, A., van Krieken, E., Serafini, L., van Harmelen, F.: Refining neural network predictions using background knowledge. Machine Learning pp. 1-39 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.301, + 0.785, + 0.355 + ], + "angle": 0, + "content": "5. Guo, W., Du, Y., Shen, X., Lepetit, V., Alameda-Pineda, X., Moreno-Noguer, F.: Back to mlp: A simple baseline for human motion prediction. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4809-4819 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.357, + 0.785, + 0.398 + ], + "angle": 0, + "content": "6. Guo, Y., Li, Y., Zhou, X., He, W.: A keypoint transformer to discover spine structure for cobb angle estimation. In: 2021 IEEE International Conference on Multimedia and Expo (ICME). pp. 1-6. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.785, + 0.425 + ], + "angle": 0, + "content": "7. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.785, + 0.452 + ], + "angle": 0, + "content": "8. Jang, W.D., Kim, C.S.: Interactive image segmentation via backpropagating refinement scheme. CVPR pp. 5297-5306 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.453, + 0.785, + 0.508 + ], + "angle": 0, + "content": "9. Jin, S., Xu, L., Xu, J., Wang, C., Liu, W., Qian, C., Ouyang, W., Luo, P.: Whole-body human pose estimation in the wild. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16. pp. 196-214. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.785, + 0.563 + ], + "angle": 0, + "content": "10. Kim, J., Kim, T., Kim, T., Choo, J., Kim, D.W., Ahn, B., Song, I.S., Kim, Y.J.: Morphology-aware interactive keypoint estimation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 675-685. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "1. Klinwichit, P., Yookwan, W., Limchareon, S., Chinnasarn, K., Jang, J.S., Onuean, A.: Buu-lspine: A thai open lumbar spine dataset for spondylolisthesis detection. Applied Sciences 13(15), 8646 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "2. Liu, Q., Zheng, M., Planche, B., Karanam, S., Chen, T., Niethammer, M., Wu, Z.: Pseudoclick: Interactive image segmentation with click imitation. In: European Conference on Computer Vision. pp. 728-745. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.701 + ], + "angle": 0, + "content": "3. Liu, T., Wang, Y., Yang, Y., Sun, M., Fan, W., Bunger, C., Wu, C.: A multi-scale keypoint estimation network with self-supervision for spinal curvature assessment of idiopathic scoliosis from the imperfect dataset. Artificial Intelligence in Medicine 125, 102235 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.729 + ], + "angle": 0, + "content": "4. Moon, G., Chang, J.Y., Lee, K.M.: Posefix: Model-agnostic general human pose refinement network. CVPR pp. 7773-7781 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.785, + 0.798 + ], + "angle": 0, + "content": "5. Pisov, M., Kondratenko, V., Zakharov, A., Petraikin, A., Gombolevskiy, V., Morozov, S., Belyaev, M.: Keypoints localization for joint vertebra detection and fracture severity quantification. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part VI 23. pp. 723-732. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. Qu, C., Zhang, T., Qiao, H., Tang, Y., Yuille, A.L., Zhou, Z., et al.: Abdomenatlas-8k: Annotating 8,000 ct volumes for multi-organ segmentation in three weeks. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.176, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "17. Rahmaniar, W., Suzuki, K., Lin, T.L.: Auto-ca: automated cobb angle measurement based on vertebrae detection for assessment of spinal curvature deformity. IEEE Transactions on Biomedical Engineering (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "18. Ruggero Ronchi, M., Perona, P.: Benchmarking and error diagnosis in multi-instance pose estimation. In: Proceedings of the IEEE international conference on computer vision. pp. 369-378 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "19. Shi, D., Wei, X., Li, L., Ren, Y., Tan, W.: End-to-end multi-person pose estimation with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11069-11078 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "20. Singh, S., Sandhu, N., Puri, T., Gulati, R., Kashyap, R.: A study of correlation of various growth indicators with chronological age. International Journal of Clinical Pediatric Dentistry 8(3), 190 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "21. Sofiuk, K., Petrov, I., Barinova, O., Konushin, A.: f-brs: Rethinking backpropagating refinement for interactive segmentation. CVPR pp. 8623-8632 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "22. Sofiuk, K., Petrov, I.A., Konushin, A.: Reviving iterative training with mask guidance for interactive segmentation. In: 2022 IEEE International Conference on Image Processing (ICIP). pp. 3141-3145. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "23. Türkoz, C., Kaygirisiz, E., Ulusoy, C., Ates, C.: A practical formula for determining growth. Diagnostic and Interventional Radiology 23(3), 194 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.453 + ], + "angle": 0, + "content": "24. Wu, H., Bailey, C., Rasoulinejad, P., Li, S.: Automatic landmark estimation for adolescent idiopathic scoliosis assessment using boostnet. In: MICCAI. pp. 127-135. Springer (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.508 + ], + "angle": 0, + "content": "25. Xu, Y., Xu, D., Hong, X., Ouyang, W., Ji, R., Xu, M., Zhao, G.: Structured modeling of joint deep feature and prediction refinement for salient object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3789-3798 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.785, + 0.55 + ], + "angle": 0, + "content": "26. Yang, J., Zeng, A., Li, F., Liu, S., Zhang, R., Zhang, L.: Neural interactive keypoint detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15122-15132 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "27. Zhang, H., Chung, A.C.: A dual coordinate system vertebral landmark detection network with sparse-to-dense vertebral line interpolation. Bioengineering 11(1), 101 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.591 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Bones Can't Be Triangles_ Accurate and Efficient Vertebrae Keypoint Estimation through Collaborative Error Revision/79ad0f6a-e5ca-4e70-bd32-205034f5cde0_origin.pdf b/2024/Bones Can't Be Triangles_ Accurate and Efficient Vertebrae Keypoint Estimation through Collaborative Error Revision/79ad0f6a-e5ca-4e70-bd32-205034f5cde0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a697cf4e570f5107590df1aa7dca2442e1d661ac --- /dev/null +++ b/2024/Bones Can't Be Triangles_ Accurate and Efficient Vertebrae Keypoint Estimation through Collaborative Error Revision/79ad0f6a-e5ca-4e70-bd32-205034f5cde0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae20ef6570cc5400fbc426d8e18c22f3d0f7912ff76b2ab73458fb3042830e93 +size 3236543 diff --git a/2024/Bones Can't Be Triangles_ Accurate and Efficient Vertebrae Keypoint Estimation through Collaborative Error Revision/full.md b/2024/Bones Can't Be Triangles_ Accurate and Efficient Vertebrae Keypoint Estimation through Collaborative Error Revision/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0368a80ebf66432a4a6ee8fdd01b7764ede9149b --- /dev/null +++ b/2024/Bones Can't Be Triangles_ Accurate and Efficient Vertebrae Keypoint Estimation through Collaborative Error Revision/full.md @@ -0,0 +1,304 @@ +# Bones Can't Be Triangles: Accurate and Efficient Vertebrae Keypoint Estimation through Collaborative Error Revision + +Jinhee Kim\*1, Taesung Kim\*1, and Jaegul Choo + +Korea Advanced Institute of Science and Technology {seharanul17, zkm1989, jchoo}@kaist.ac.kr * indicates equal contributions. + +![](images/ca6445e836a61c2d145b6f13cfaf09dea4fef77f9263271ae1fd7975ee058164.jpg) + +![](images/132500992809c6c3958818e7a42e4c2489ba2eadd87c6a10cb23401559e95777.jpg) +(a) Image + +![](images/847cf46633eefd9c685d4c73531b621ad8a2e6f72b7db6627f1282407879e2cf.jpg) + +![](images/9685d7166195f6848cb71988cfdf95e39ea470e09feba1f1b3dc9713fceb971d.jpg) +(b) Initial prediction + +![](images/3cd1e50afc61a263e3f4f2abc30cc90802eb3743068ba91f9beb550ee72f7eb8.jpg) + +![](images/0fedc50436e78b8a35f4b85c7791dab4b8cbe58fbb911a6aa16df3a7898fb8af.jpg) +(c) KeyBot + +![](images/998da9e1a72fe3c716e25d0022c4dba31e48eece06842be5e6ef5a0d7f740960.jpg) + +![](images/aef0418b5efadb9b952b22f025ff400f2320b2c6109aba5707dda6a9a68b6109.jpg) +(d) Groundtruth +Fig. 1: Vertebrae keypoint refinement results on BUU-LA [11] and AASCE [24]. Initial predictions exhibit significant errors in vertebra shapes due to difficulty identifying individual vertebrae. KeyBot effectively reduces these errors without user input. + +Abstract. Recent advances in interactive keypoint estimation methods have enhanced accuracy while minimizing user intervention. However, these methods require user input for error correction, which can be costly in vertebrae keypoint estimation where inaccurate keypoints are densely clustered or overlap. We introduce a novel approach, KeyBot, specifically designed to identify and correct significant and typical errors in existing models, akin to user revision. By characterizing typical error types + +and using simulated errors for training, KeyBot effectively corrects these errors and significantly reduces user workload. Comprehensive quantitative and qualitative evaluations on three public datasets confirm that KeyBot significantly outperforms existing methods, achieving state-of-the-art performance in interactive vertebrae keypoint estimation. The source code and demo video are available on our project page. + +Keywords: Interactive vertebrae keypoint estimation $\cdot$ X-ray image analysis $\cdot$ Automatic error correction $\cdot$ User interaction + +# 1 Introduction + +Accurate vertebrae keypoint estimation from X-ray images is crucial for effective medical diagnosis and treatment planning [20, 23], with errors having significant impacts on clinical decisions, such as in spinal surgeries and pain management [6, 13, 15]. This task presents a distinct set of challenges due to the complex anatomy of the spine [17, 27]. The similar and repetitive shapes of vertebrae complicate distinguishing between individual vertebrae and accurately estimating their keypoints, as shown in Fig. 1. This requires a deep understanding of sequential context and anatomy, unlike more distinct keypoints found in human pose [5, 9, 14, 19] or facial keypoint [1, 3] estimation tasks. + +Such challenges often result in significant errors during automated vertebrae keypoint estimation, including vertex misidentification, bone misidentification, and left-right inversion, as illustrated in Fig. 2. Manual revision of these errors is time-consuming and labor-intensive, especially as the number of vertebrae increases. Missing even a single vertebra can lead to extensive adjustments, thereby significantly increasing the time and effort required for accurate annotation. To mitigate these, interactive keypoint estimation approaches have been developed, significantly improving performance with minimal user input [10,26]. These methods aim to automatically refine inaccurate keypoints with minimal user feedback, using the feedback as a hint to adjust the remaining errors, as shown in Fig. 3(a). However, they heavily rely on users to identify and correct errors, which can be time-consuming and labor-intensive, especially when erroneous keypoints are densely clustered or overlapped. + +In response, we introduce a novel method named KeyBot, designed to autonomously detect and correct specific errors in vertebrae keypoint estimation before user evaluation, building upon existing interactive keypoint estimation models, as illustrated in Fig. 3(b). Initially, the interaction model predicts keypoints based on the input image. During refinement, KeyBot autonomously preevaluates these predictions, identifies specific errors, and provides corrective feedback similar to user input. In contrast to previous models requiring at least one user input, KeyBot autonomously refines model predictions before user input. + +We generate synthetic data characterizing typical error types, vertex misidentification, bone misidentification, and left-right inversion, and train KeyBot using this data to focus specifically on predefined error types. Independently trained from the interaction model, KeyBot effectively identifies errors that might be + +![](images/4713e0e8374727d2eb2389713f0aac795f3b598075edb2a36d4b2b870ccc9894.jpg) +(a) Vertex misidentification + +![](images/b48c1d1d2c6796b0c4f9d01233f3f962c4c7ed249243c2388fd6210c1f811e7f.jpg) +(b) Bone misidentification + +![](images/16ef5afecc64764dd1aace2df0b321480e920542e5dfa64e9c730b76afef6c9b.jpg) +(c) Left-right inversion +Fig. 2: Error types in vertebrae keypoint estimation, illustrated on the AASCE dataset. + +overlooked by the interaction model, significantly reducing the user's revision efforts. Rather than correcting errors from scratch, users can focus on verifying and fine-tuning results, with KeyBot having already eliminated major errors. + +KeyBot delivers quick, cost-effective, and iterative feedback without user input, which is crucial for complex cases requiring multiple refinements. This automated feedback is typically much faster than user review, significantly enhancing workflow efficiency. The collaborative feedback loop between the user, KeyBot, and the interaction model enriches the process by combining expert medical insights with KeyBot's focus on typical error types. This synergy addresses a broad range of potential errors and enhancements, leading to an overall improvement in keypoint estimation accuracy. + +Through extensive experiments on three public datasets [11, 24], KeyBot demonstrates significant improvements in vertebrae keypoint estimation accuracy. The model reduces the mean radial error (MRE) by $19\%$ and decreases the number of user clicks (NoC) required for target performance by $17\%$ on the AASCE dataset, achieving state-of-the-art performance. These results highlight the significance of KeyBot in advancing vertebrae keypoint estimation, offering an advanced, user-centric solution to this unique and complex challenge. + +In summary, our contributions are: (1) We introduce a novel method, KeyBot, designed to provide automated corrections for specific errors in vertebrae keypoint estimation. This enhances existing interactive keypoint estimation frameworks and improves overall keypoint estimation accuracy and efficiency. (2) We train KeyBot on synthetic data representing common error types in vertebrae, enabling it to effectively identify and address these errors. (3) KeyBot significantly reduces the need for user intervention during the keypoint annotation process. (4) Extensive evaluation across three public datasets confirms the efficacy of our proposed method in improving keypoint estimation accuracy and significantly reducing user intervention. + +![](images/123c58c08b8eb78e2db6d319ad6fb374a38f2f4682ac3ad59224e3551e70f5a5.jpg) +Fig. 3: Overview of (a) the interactive keypoint estimation and (b) the KeyBot framework. The interaction model generates initial keypoint predictions from an image, which are then revised by users, leading to updated results. This approach requires user input for error correction. In contrast, KeyBot offers a cost-efficient feedback mechanism via automated, rapid, and iterative refinement without user input. It independently identifies and corrects major errors, allowing users to focus on the final adjustments. + +# 2 Related work + +Automated error refinement. Despite the advancements in deep learning, errors in model predictions remain a formidable challenge. To mitigate these inaccuracies, a specialized area of research dedicated to refining prediction results has emerged, leading to the development of advanced methods for identifying and correcting errors in model outputs [4,14,16,25]. For example, Qu et al. [16] introduced a method leveraging attention maps to detect label errors in CT data, markedly improving annotation efficiency and facilitating the collection of extensive datasets. Similarly, efforts to refine keypoint estimation have been noteworthy. For instance, Ronchi et al. [18] analyzed human pose estimation results to categorize three main types of errors, while Moon et al. [14] developed a method that specifically addresses and corrects these categorized errors, achieving state-of-the-art human pose estimation performances. + +Despite these advancements, estimating vertebrae keypoints from X-ray images introduces additional challenges due to the complex sequential structure of the spine. Error patterns, such as the mislocalization of an entire vertebra, often require extensive user inputs for correction. In response, we propose KeyBot, + +a method tailored for revising errors specific to vertebrae keypoint estimation, effectively addressing the unique challenges of this task. + +User interaction. To maximize the accuracy of predictions while minimizing the need for user input, the development of interaction-based models is paramount. Interactive segmentation is a particularly active area within this domain. Backpropagating refinement schemes such as BRS [8] and f-BRS [21] have improved interactive image segmentation by incorporating user hints through backpropagation. Similarly, RITM [22] has shown superior performance by integrating an iterative refinement process into its training process. In the field of keypoint estimation, interactive keypoint estimation methods have been researched to reduce user input through interactive refinement frameworks. Kim et al. [10] pioneered an interactive framework for keypoint estimation, significantly improving the accuracy and efficiency of the keypoint annotation process in medical datasets. ClickPose [26] developed an interactive keypoint estimation model for multi-person 2D keypoints, significantly reducing annotation costs. + +Recently, Liu et al. [12] introduced a method that generates pseudo clicks to mimic human interactions for refining segmentation predictions autonomously, thereby minimizing the required number of user clicks. However, their error decoder is trained end-to-end and can thus be affected by inherent biases or limitations of the segmentation module. To address this challenge, our KeyBot is trained independently to correct simulated errors, which allows KeyBot to provide corrective feedback free from any potential biases present in the keypoint estimation model. Unlike traditional interactive keypoint estimation models that rely on users to identify and correct errors, KeyBot autonomously generates corrective feedback. This significant innovation not only enhances the keypoint estimation accuracy but also remarkably reduces the need for user interaction, thereby streamlining the error correction process. + +# 3 KeyBot + +We introduce KeyBot, a novel method designed to enhance vertebrae keypoint estimation by identifying and correcting errors prior to user intervention. We employ an existing interactive model that can incorporate user feedback for interactive modifications. Given model predictions, KeyBot identifies erroneous keypoints and generates their pseudo-corrections, which are then fed to the interaction model for updates. KeyBot consists of two components: The detector discerns whether each input keypoint is accurate, and the corrector refines input keypoints accurately. Both components are trained with synthetic errors to focus specifically on the target error types. + +# 3.1 Overview of the propose approach + +Interactive keypoint estimation framework. Consider an interactive keypoint estimation model $\mathcal{F}_{\theta}$ with parameters $\theta$ , operating on an image $\pmb{x}$ . The model aims to precisely estimate $K$ target keypoints through user-interactive + +![](images/e9134607dfc9aa107aa066d901484079bff86056ec607e6a9a6c70af4c985650.jpg) +(a) +Fig. 4: Training overview of KeyBot. KeyBot consists of two main components, the detector and the corrector. (a) Synthetic errors are introduced to the groundtruth keypoints, creating inaccurate keypoints and corresponding anomaly labels. (b) The detector is trained to discern whether each input keypoint is accurate or not. (c) The corrector is trained to refine these inaccurate keypoints accurately. + +refinement. In each iteration $t$ , the model generates the $t$ -th prediction, $\pmb{y}_t$ , using three inputs: the image $\pmb{x}$ , user revisions $\pmb{u}_t$ , and its previous predictions $\pmb{e}_t$ : + +$$ +\boldsymbol {y} _ {t} = \mathcal {F} _ {\theta} (\boldsymbol {x}, \boldsymbol {u} _ {t}, \boldsymbol {e} _ {t}), \tag {1} +$$ + +where $\pmb{u}_0\coloneqq 0$ and $\pmb {e}_0\coloneqq 0$ . Subsequently, a user reviews the results, and the adjustment made to a keypoint position is encoded as Gaussian heatmaps $\pmb{u}_t^i$ , effectively capturing and representing the refined locations: + +$$ +\boldsymbol {\rho} _ {t + 1}, \boldsymbol {u} _ {t + 1} = \Psi (\boldsymbol {x}, \boldsymbol {y} _ {t}, \boldsymbol {u} _ {t}), \tag {2} +$$ + +where the set $\rho_{t}$ represents keypoint indices revised by the user over $t$ iterations. Next, we selectively gather the known-to-be-false predictions from $\pmb{y}$ to $\pmb{e}$ and reintroduce them as inputs in the next step to differentiate mispredictions [10]: + +$$ +\boldsymbol {e} _ {t + 1} ^ {i} = \boldsymbol {y} _ {t} ^ {i} \text {f o r} i \in \boldsymbol {\rho} _ {t}. \tag {3} +$$ + +Here, $\pmb{e}$ denotes predictions modified by the user. These updates on $\pmb{u}_t$ and $\pmb{e}_t$ are critical for refining the model predictions, leading to the updated keypoint predictions reflecting the user feedback: + +$$ +\boldsymbol {y} _ {t + 1} = \mathcal {F} _ {\theta} (\boldsymbol {x}, \boldsymbol {u} _ {t + 1}, \boldsymbol {e} _ {t + 1}). \tag {4} +$$ + +This process repeats until sufficient accuracy is achieved or the predefined maximum user interaction $T$ is reached. + +The KeyBot framework. KeyBot offers an automated keypoint refinement step that precedes and complements the existing user-based refinement, effectively reducing the need for user interaction in the process. KeyBot consists of the detector $\mathcal{G}_{\phi_r}$ and corrector $\mathcal{G}_{\phi_s}$ , which together identifies and corrects inaccurate keypoints predicted by $\mathcal{F}_{\theta}$ . Following each prediction iteration of $\mathcal{F}_{\theta}$ , KeyBot evaluates $\pmb{y}_t$ and suggests necessary corrections, identifying erroneous keypoints $\pmb{\nu}_{t,n+1}$ and their pseudo corrections $\pmb{z}_{t,n+1}$ : + +$$ +\boldsymbol {\nu} _ {t, n + 1} = \mathcal {G} _ {\phi_ {r}} (\boldsymbol {x}, \boldsymbol {y} _ {t, n}) \text {a n d} \boldsymbol {z} _ {t, n + 1} = \mathcal {G} _ {\phi_ {s}} (\boldsymbol {x}, \boldsymbol {y} _ {t, n}), \tag {5} +$$ + +with $\pmb{y}_{t,0} \coloneqq \pmb{y}_t$ . This step is similar to the user interaction step outlined in Eq. (2). Here, KeyBot can detect multiple inaccurate keypoints at once and simultaneously offer corrective feedback for each identified error. If no inaccuracies are detected by the detector, KeyBot concludes and $\pmb{y}_{t,n}$ becomes the final prediction. Revision information, $\pmb{c}_{t,n}$ , now includes both user $(\pmb{u}_{t,n})$ and model $(\pmb{z}_{t,n})$ corrections. Here, user modifications are prioritized as the definitive ground truth. Thus, keypoints already revised by the user are exempt from further updates by the pseudo-corrections from KeyBot. + +$$ +\boldsymbol {c} _ {t, n + 1} ^ {i} = \boldsymbol {z} _ {t, n + 1} ^ {i} \text {f o r} i \in \boldsymbol {\nu} _ {t, n + 1} \backslash \boldsymbol {\rho} _ {t}. \tag {6} +$$ + +Previous work [10,22] relies solely on the prediction at the most recent step as the previous prediction, as described in Eq. (3). However, this approach can result in losing track of erroneous predictions once they are corrected in the model. Hence, our approach accumulates predictions identified as erroneous by either KeyBot or the user across all iterations. These false predictions retain the history of incorrect predictions and serve to inform the model about specific prediction errors to avoid, thereby guiding it toward more accurate future predictions: + +$$ +\boldsymbol {e} _ {t, n + 1} ^ {i} = \boldsymbol {y} _ {t, n} ^ {i} \text {f o r} i \in \boldsymbol {\nu} _ {t, n + 1} \backslash \boldsymbol {\nu} _ {t,: n} \backslash \boldsymbol {\rho} _ {t}. \tag {7} +$$ + +Here, $e$ denotes predictions modified by either the user or KeyBot. Finally, these are then fed back into the backbone interaction network for a new, corrected prediction akin to user feedback: + +$$ +\boldsymbol {y} _ {t, n + 1} = \mathcal {F} _ {\theta} (\boldsymbol {x}, \boldsymbol {c} _ {t, n + 1}, \boldsymbol {e} _ {t, n + 1}). \tag {8} +$$ + +Each iteration of KeyBot contributes progressively towards enhancing the overall accuracy and reliability of the keypoint predictions. This cycle repeats until the detector detects no errors or the maximum KeyBot iteration count $N$ is reached. Upon the KeyBot process concludes, say, after $n + 1$ iterations, the prediction $\pmb{y}_{t,n + 1}$ is established as the final output for the $t$ -th iteration of the interactive keypoint estimation process. For the subsequent $(t + 1)$ -th user-side phase, given new user feedback as Eq. (2), $\pmb{c}_{t,n + 1}$ and $\pmb{e}_{t,n + 1}$ are updated to $\pmb{c}_{t + 1,0}$ and $\pmb{e}_{t + 1,0}$ , + +$$ +\boldsymbol {c} _ {t + 1, 0} = \boldsymbol {u} _ {t + 1} \text {a n d} \boldsymbol {e} _ {t + 1, 0} ^ {i} = \boldsymbol {y} _ {t, n + 1} ^ {i} \text {f o r} i \in \boldsymbol {\rho} _ {t + 1}, \tag {9} +$$ + +and the keypoints are revised accordingly: + +$$ +\boldsymbol {y} _ {t + 1, 0} = \mathcal {F} _ {\theta} (\boldsymbol {x}, \boldsymbol {c} _ {t + 1, 0}, \boldsymbol {e} _ {t + 1, 0}). \tag {10} +$$ + +The complete procedure is summarized in Algorithm 1 in Appendix B. + +![](images/465d142995c3e4d69f7368bba681efcd8c913cc9152991d1b7a4a46a67f9945e.jpg) +Fig. 5: Implementing realistic estimation inaccuracy simulations for training KeyBot. + +# 3.2 Model design of KeyBot + +KeyBot integrates the detector and corrector to identify potential errors and generates corrective feedback for their revision. + +Error detector. The error detector specializes in identifying up to $K$ erroneous keypoints, focusing on prevalent error types in vertebrae keypoint estimation. Our empirical analysis indicates that evaluating the entire bone structure at once is less effective for error detection. Therefore, the detector evaluates keypoints in smaller groups, examining $k \leq K$ keypoints simultaneously. This approach simplifies the task and improves its performance. During training, we randomly select a set of consecutive groundtruth keypoints $\mathbf{y}^*$ and either maintain their accuracy or introduce errors. The cropped image area around these keypoints, along with their corresponding Gaussian keypoint heatmaps, are input to the detector. It then predicts the likelihood of error presence in each keypoint. + +Let $\varepsilon$ denote the function that induces synthetic errors in groundtruth keypoints, and $v$ represents the anomaly labels for the keypoints. The detector performs multi-label binary classification to differentiate between normal and abnormal keypoints using a sigmoid function in its output layer. It is supervised to minimize the Binary Cross-Entropy (BCE) loss between the anomaly labels $v$ and the predictions: + +$$ +\phi_ {r} \leftarrow \phi_ {r} - \eta_ {r} \nabla_ {\phi_ {r}} \mathcal {L} _ {b c e} (v, \mathcal {G} _ {\phi_ {r}} (\boldsymbol {x}, \varepsilon (\boldsymbol {y} ^ {*})), \tag {11} +$$ + +with $\eta_r$ representing the learning rate for the detector. + +During inference, the detector iteratively detects errors across all keypoints, examining $k$ keypoints at a time with a stride of $s$ . To facilitate a comprehensive assessment of each keypoint in different contexts, we set the stride $s$ to be smaller than $k$ . This results in multiple predictions for a single keypoint, and if any prediction indicates an anomaly, the keypoint is flagged as erroneous. + +Error corrector. The error corrector aims to generate pseudo-corrections for keypoints inaccurately identified by the detector, thus providing feedback to the model $\mathcal{F}_{\theta}$ . During training, we randomly transform the groundtruth keypoints to erroneous ones using $\varepsilon$ . Taking these keypoints and the image $\pmb{x}$ as inputs, the corrector is trained to revert these to their original, accurate state, focusing particularly on the typical error types implemented by $\varepsilon$ . It generates $K$ -channel heatmaps indicating reconstructed keypoint locations, employing a sigmoid function in its output. Its learning objective is the BCE loss between the ground truth keypoint heatmaps and the reconstructed outputs: + +$$ +\phi_ {s} \leftarrow \phi_ {s} - \eta_ {s} \nabla_ {\phi_ {s}} \mathcal {L} _ {b c e} (\boldsymbol {y}, \mathcal {G} _ {\phi_ {s}} (\boldsymbol {x}, \varepsilon (\boldsymbol {y} ^ {*})), \tag {12} +$$ + +where $\eta_{s}$ is the learning rate for the corrector. + +During inference, the corrector refines predictions from the backbone model $\mathcal{F}_{\theta}$ , and the output corresponding to inaccurately assessed keypoints by the detector is fed to $\mathcal{F}_{\theta}$ to serve as pseudo-corrections. + +# 3.3 Training KeyBot via realistic error simulation + +KeyBot aims to address significant and typical errors in vertebrae keypoint estimation. The key idea is to train KeyBot to identify and correct three error types commonly observed in vertebrae keypoint estimation: vertex misidentification, bone misidentification, and left-right inversion. During training, KeyBot is provided with both accurate keypoints and keypoints with synthetic errors. We randomly simulate the predefined error types, as shown in Fig. 5. Rather than relying on highly variable and not clearly defined real errors, we use characterized and consistent synthetic errors. + +By employing this training approach, we expose KeyBot to a wide range of error scenarios, equipping it with the proficiency to detect and correct these errors effectively. As a result, KeyBot provides a targeted evaluation of the keypoints predicted by the interaction model, focusing on specific error types. + +Vertex misidentification (misvertex). Misvertex is a type of error where a keypoint is correctly identified on a vertex of a vertebrae but is incorrectly attributed to the wrong vertex. For instance, a keypoint intended for the first vertex is inaccurately predicted at the third vertex. To simulate misvertex, we randomly displace a keypoint to its neighbors within a $[-r, +r]$ index range for a predefined number $r$ , excluding its original position. Keypoints moving beyond their index range wrap around cyclically. + +Bone Misidentification (misbone). Misbone errors represent errors where the entire vertebra is incorrectly classified. Our simulation randomly shifts keypoints up or down by one vertebra. This is achieved by selecting start and end keypoint indices and moving all-encompassed keypoints to an adjacent vertebra. + +Left-right inversion (lr-inversion). LR-inversion occurs when keypoint pairs are incorrectly positioned on the opposite lateral side. To simulate LR-inversion in our model, we randomly interchange the positions of keypoint pairs that delineate one side of each vertebra. + +Table 1: Performance comparison of mean radial error in keypoint estimation across three datasets. UC denotes the count of user clicks provided to the model. + +
MethodAASCEBUU-APBUU-LA
-UC1UC2UC3UC4-UC1UC2-UC1UC2
BRS [8]45.6535.8829.9325.4021.9751.2230.9136.1740.2043.1935.07
f-BRS [21]64.0657.3052.5347.8343.7944.0524.9917.5236.0327.9319.29
RITM [22]56.0337.9631.0326.7623.7936.4325.7617.9623.2717.799.89
Kim et al. [10]51.5830.6025.7821.6019.0842.3123.2616.4623.4314.2910.29
Click-Pose [26]54.6546.5044.0841.7340.0432.7229.3026.2033.7021.6217.38
KeyBot-i144.1825.9320.6618.0316.3732.0121.7915.9718.7713.399.12
KeyBot-i242.5227.0322.9518.7216.2331.8820.6515.8119.1113.479.03
KeyBot-i341.7026.5925.0221.2316.7631.8720.6615.8418.7413.368.97
+ +# 4 Experiments + +# 4.1 Experimental Setup + +Datasets. We rigorously validate our method across three publicly available real-world datasets. The AASCE dataset [24] consists of 608 spinal anterior-posterior radiographs, each annotated with 68 keypoints, marking the vertices of 17 vertebrae. The BUU dataset [11] features two distinct subsets from 400 unique patients: BUU-AP with 400 anterior-posterior (AP) view radiographs, each annotated with 20 keypoints, and BUU-LA with 400 left lateral (LA) view radiographs, each annotated with 22 spinal keypoints. Images that exhibit critical keypoint annotation errors, such as those with incorrectly ordered keypoint indices similar to typical errors in keypoint estimation models, are excluded from the datasets. Comprehensive details about the dataset, including the data split and examples of the excluded images, are available in Appendix C. + +Metrics. Following the evaluation protocol of previous work [10, 26], we evaluate the model performance using the mean radial error (MRE) and the number of clicks (NoC). MRE computes the mean Euclidean distance between the groundtruth and predicted keypoints. No $C_a$ @ $b$ measures the average number of user clicks required to achieve a target MRE of $b$ , given a maximum of $a$ clicks. + +Baseline models. We compare our method with state-of-the-art interactive keypoint estimation models, including the model proposed by Kim et al. [10] and Click-Pose [26]. Additionally, we use interactive image segmentation models such as BRS [8], f-BRS [21], and RITM [22] as baselines. + +Implementation details. We utilize ResNet [7]-18 and DeepLab-V3 [2] as the backbone for the detector and corrector, respectively. We build on the pretrained model of Kim et al. [10], which is designed to estimate keypoints based on user feedback. All of the models are trained independently. We apply the misvertex error type to train the detector and all three error types for the corrector. KeyBot-i1, KeyBot-i2, and KeyBot-i3 refer to KeyBot with one, two, and three + +Table 2: Comparison of required number of user clicks to reach target performance on the AASCE, BUU-AP, and BUU-LA datasets. + +
MethodAASCEBUU-APBUU-LA
NoC10@20NoC10@30NoC10@40NoC10@50NoC5@15NoC5@20NoC5@25NoC5@30NoC5@15NoC5@20NoC5@25NoC5@30
BRS [8]2.531.971.511.232.801.510.920.613.892.261.340.86
f-BRS [21]7.165.304.213.482.231.180.680.481.240.610.440.40
RITM [22]3.202.371.871.381.050.580.420.300.570.240.170.14
Kim et al. [10]2.101.541.231.030.940.540.370.320.380.150.140.12
Click-Pose [26]4.153.563.202.971.200.890.760.711.450.700.560.45
KeyBot-i11.841.260.920.530.870.510.330.280.340.100.060.05
KeyBot-i21.781.210.910.610.870.480.300.240.350.110.060.05
KeyBot-i31.741.320.940.680.870.490.320.240.340.100.060.05
+ +iterations, respectively. Unless explicitly stated, the experimental results are conducted without user interaction. The training and inference are conducted on a single NVIDIA GeForce RTX 3090 GPU with 24GB memory. More details on experimental settings, including training details, are provided in Appendix C. + +# 4.2 Experimental results and analysis + +Comparison with state-of-the-art methods. In this section, we present a comparative analysis of interactive keypoint estimation performance between KeyBot and established baseline models by increasing the maximum KeyBot iteration count $N$ from one to three. As shown in Table 1, KeyBot achieves state-of-the-art performance in mean radial error (MRE) across the AASCE, BUU-AP, and BUU-LA datasets. KeyBot consistently outperforms baseline models both in initial predictions and during subsequent user interactions. Remarkably, KeyBot with a single user click (UC1) shows superior performance compared to BRS, f-BRS, RITM, and Click-Pose with two user clicks (UC2) on the AASCE dataset. Furthermore, the performance of KeyBot improves with additional iterations, demonstrating its ability to refine predictions through multiple correction cycles. These results highlight the effectiveness of KeyBot in enhancing keypoint estimation accuracy and reducing user intervention. + +In terms of the number of clicks (NoC), as shown in Table 2, KeyBot exhibits superior performance compared to the baseline models across all three datasets. For instance, on the AASCE dataset, our proposed model nearly halves the NoC required to achieve a target MRE of 50 ( $\mathrm{NoC}_{10} @ 50$ ) compared to that of the baseline models. This substantial reduction in NoC underscores KeyBot's proficiency in interactive frameworks, significantly lowering annotation costs. + +Analysis on the detector. We investigate the effectiveness of the detector in identifying erroneous keypoints within images. In Fig. 6, we provide (a) the number of images where the detector identifies no errors versus those where errors are detected, as well as (b) the distribution of their actual mean radial error. + +![](images/68e2aeac048d4f26decab5ca04079d817be705d917ef8f32c2bb855b0cdeab9e.jpg) +Fig.6: Effectiveness of the detector in identifying errors on the AASCE dataset. + +Table 3: Keypoint prediction errors on the AASCE dataset, comparing MRE for examples where the detector identifies errors versus where it does not. + +
MethodNo error detectedError detected
Initial prediction28.9964.88
KeyBot-i1-53.12
KeyBot-i2-50.49
KeyBot-i3-49.18
+ +Out of the total images analyzed, 43 images have no errors detected, while 73 images have errors identified. The images classified as "No error detected" exhibit significantly lower errors compared to those classified as "Error detected." These findings indicate the precision and effectiveness of the detector in distinguishing between erroneous predictions and accurate ones. + +We further examine error reduction through successive KeyBot iterations for images where errors are initially detected. As demonstrated in Table 3, KeyBot effectively enhances the accuracy of these error-detected images, significantly reducing errors in subsequent iterations. Notably, after three iterations of KeyBot, the MRE significantly decreases from 64.88 to 49.18. This reduction highlights KeyBot's ability to iteratively identify inaccuracies and improve predictions. + +Ablation study. This section validates each component of KeyBot. First, we study the role of the detector under two scenarios: (i) employing corrector outputs directly as the final results and (ii) utilizing all corrector outputs as pseudocorrections for the interaction model. As detailed in Table 4, both scenarios exhibit significant performance degradation compared to KeyBot, underscoring the importance of using the detector to selectively apply corrector outputs. + +Additionally, utilizing only the interactive keypoint estimation network, without the support of KeyBot, results in a significant performance degradation compared to the refined predictions with KeyBot. When ablating the error types simulated during the training of the corrector, our findings show that training the corrector to recognize and rectify three distinct types of errors significantly improves its prediction accuracy compared to when it is trained on only one error type. This finding highlights the necessity of incorporating all three typical errors into the training process. + +Qualitative results. In Fig. 7, we analyze the qualitative results of KeyBot with three iterations. The results demonstrate that KeyBot effectively identifies and reduces all three error types: vertex misidentification (Misvertex), bone misidentification (Misbone), and left-right inversion (LR-inversion). For these errors, KeyBot provides corrective feedback to the interactive model, enabling necessary adjustments. For instance, in Fig. 7(a), the initial prediction skips a bone, resulting in a Misbone error (red), and also exhibits a Misvertex error + +Table 4: Ablation study of our keypoint refinement method on the AASCE dataset. + +
Interaction modelKeyBotMean radial error
DetectorCorrectorError type-UC1UC2UC3
✓✓✓all44.1825.9320.6618.03
XX✓all50.9939.1732.9830.36
✓X✓all57.6145.9439.1136.01
✓XX-51.5830.6025.5721.07
✓✓✓misvertex52.6731.1425.9221.65
✓✓✓misbone49.0729.8523.4320.75
✓✓✓lr-inversion52.3830.1625.1320.90
+ +![](images/f8a01f1b7f685cbdefd8c45e6fc67b2d0c9e11cbe1977c9c63b91dde52df3074.jpg) +(a) misvertex+misbone (MRE: $71.7\rightarrow 11.9$ + +![](images/f48d389d76dda02e382386fb0e9c14c0299ed4d9349fb1bdb3da9820b02fc4f5.jpg) +(b) misvertex+misbone (MRE: 91.9 → 20.2) + +![](images/24533e71014362e7a8f3f8e268de202f96af687e578696373c94e40e411c1a7d.jpg) +(c) misvertex+misbone (MRE: 193.3 → 15.3) +Image Initial prediction KeyBot + +![](images/ca6b4f126fbb2f2d2c6876befd69939cd9175e3c631c7f82631f59aec355b680.jpg) +(d) misvertex+Ir-inversion (MRE: 443.5 → 40.3) +Image Initial prediction KeyBot +Fig. 7: Prediction results on the (a) AASCE, (b) BUU-LA, and (c,d) BUU-AP (bottom) datasets. KeyBot effectively reduces all three types of errors, even in challenging cases where error types are mixed. The keypoints classified as inaccurate by our detector are indicated with red dots, while their corresponding keypoints in the revised results are visualized as blue dots. A diagonal line connecting the upper-right vertex to the lower-left vertex is indicated in each vertebra. More results are in Appendix F. + +(purple). KeyBot accurately refines these errors, significantly reducing the MRE from 71.7 to 11.9. The results highlight KeyBot's capability to significantly enhance keypoint estimation accuracy without user clicks. + +![](images/7d487f580d16bbb59390c9e49d8798c928bcd715e8ca7486605c8e0152f011c3.jpg) +Fig. 8: Collaborative keypoint refinement process among the interaction model, KeyBot, and the user. Faced with substantial errors in initial predictions, KeyBot effectively removes all major errors, allowing users to easily identify and correct remaining inaccuracies with minimal effort. Here, refinement is achieved with just two user clicks. + +Additionally, KeyBot excels in interpretability. It includes a separate detector that enables users to easily comprehend the keypoints identified as errors and the corrections made, leading to efficient evaluation of the results. In the given examples, identified errors are indicated by red dots, and the corrections are shown as blue dots, illustrating how the keypoints are adjusted. This transparency contributes to improving the efficiency of the overall annotation process. Collaborative refinement process. In Fig. 8, we showcase the qualitative results of the collaborative refinement process involving both KeyBot and the user. Our approach establishes a feedback loop among the interaction model, KeyBot, and the user, combining expert medical knowledge and a targeted focus on specific error types. This significantly enhances the overall accuracy and efficiency of the annotation process. + +# 5 Conclusion + +We introduce a novel approach, KeyBot, specifically designed to identify and correct significant errors in vertebrae keypoint estimation. Our proposed error simulation training procedure enables KeyBot to accurately and independently address specific error types, thereby significantly reducing the need for user intervention in the annotation process. Comprehensive evaluations on public real-world datasets confirm the superiority of KeyBot over existing methods. + +# Acknowledgments + +This work was supported by Institute for Information & communications Technology Promotion (IITP) grant funded by the Korea government (MSIT) (No.RS2019-II190075 Artificial Intelligence Graduate School Program (KAIST)) and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. NRF-2022R1A2B5B02001913). + +# References + +1. Albiero, V., Chen, X., Yin, X., Pang, G., Hassner, T.: img2pose: Face alignment and detection via 6dof, face pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7617-7627 (2021) +2. Chen, L.C., Papandreou, G., Schroff, F., Adam, H.: Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587 (2017) +3. Colaco, S.J., Han, D.S.: Deep learning-based facial landmarks localization using compound scaling. IEEE Access 10, 7653-7663 (2022) +4. Daniele, A., van Krieken, E., Serafini, L., van Harmelen, F.: Refining neural network predictions using background knowledge. Machine Learning pp. 1-39 (2023) +5. Guo, W., Du, Y., Shen, X., Lepetit, V., Alameda-Pineda, X., Moreno-Noguer, F.: Back to mlp: A simple baseline for human motion prediction. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4809-4819 (2023) +6. Guo, Y., Li, Y., Zhou, X., He, W.: A keypoint transformer to discover spine structure for cobb angle estimation. In: 2021 IEEE International Conference on Multimedia and Expo (ICME). pp. 1-6. IEEE (2021) +7. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016) +8. Jang, W.D., Kim, C.S.: Interactive image segmentation via backpropagating refinement scheme. CVPR pp. 5297-5306 (2019) +9. Jin, S., Xu, L., Xu, J., Wang, C., Liu, W., Qian, C., Ouyang, W., Luo, P.: Whole-body human pose estimation in the wild. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16. pp. 196-214. Springer (2020) +10. Kim, J., Kim, T., Kim, T., Choo, J., Kim, D.W., Ahn, B., Song, I.S., Kim, Y.J.: Morphology-aware interactive keypoint estimation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 675-685. Springer (2022) +1. Klinwichit, P., Yookwan, W., Limchareon, S., Chinnasarn, K., Jang, J.S., Onuean, A.: Buu-lspine: A thai open lumbar spine dataset for spondylolisthesis detection. Applied Sciences 13(15), 8646 (2023) +2. Liu, Q., Zheng, M., Planche, B., Karanam, S., Chen, T., Niethammer, M., Wu, Z.: Pseudoclick: Interactive image segmentation with click imitation. In: European Conference on Computer Vision. pp. 728-745. Springer (2022) +3. Liu, T., Wang, Y., Yang, Y., Sun, M., Fan, W., Bunger, C., Wu, C.: A multi-scale keypoint estimation network with self-supervision for spinal curvature assessment of idiopathic scoliosis from the imperfect dataset. Artificial Intelligence in Medicine 125, 102235 (2022) +4. Moon, G., Chang, J.Y., Lee, K.M.: Posefix: Model-agnostic general human pose refinement network. CVPR pp. 7773-7781 (2019) +5. Pisov, M., Kondratenko, V., Zakharov, A., Petraikin, A., Gombolevskiy, V., Morozov, S., Belyaev, M.: Keypoints localization for joint vertebra detection and fracture severity quantification. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part VI 23. pp. 723-732. Springer (2020) +6. Qu, C., Zhang, T., Qiao, H., Tang, Y., Yuille, A.L., Zhou, Z., et al.: Abdomenatlas-8k: Annotating 8,000 ct volumes for multi-organ segmentation in three weeks. Advances in Neural Information Processing Systems 36 (2024) + +17. Rahmaniar, W., Suzuki, K., Lin, T.L.: Auto-ca: automated cobb angle measurement based on vertebrae detection for assessment of spinal curvature deformity. IEEE Transactions on Biomedical Engineering (2023) +18. Ruggero Ronchi, M., Perona, P.: Benchmarking and error diagnosis in multi-instance pose estimation. In: Proceedings of the IEEE international conference on computer vision. pp. 369-378 (2017) +19. Shi, D., Wei, X., Li, L., Ren, Y., Tan, W.: End-to-end multi-person pose estimation with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11069-11078 (2022) +20. Singh, S., Sandhu, N., Puri, T., Gulati, R., Kashyap, R.: A study of correlation of various growth indicators with chronological age. International Journal of Clinical Pediatric Dentistry 8(3), 190 (2015) +21. Sofiuk, K., Petrov, I., Barinova, O., Konushin, A.: f-brs: Rethinking backpropagating refinement for interactive segmentation. CVPR pp. 8623-8632 (2020) +22. Sofiuk, K., Petrov, I.A., Konushin, A.: Reviving iterative training with mask guidance for interactive segmentation. In: 2022 IEEE International Conference on Image Processing (ICIP). pp. 3141-3145. IEEE (2022) +23. Türkoz, C., Kaygirisiz, E., Ulusoy, C., Ates, C.: A practical formula for determining growth. Diagnostic and Interventional Radiology 23(3), 194 (2017) +24. Wu, H., Bailey, C., Rasoulinejad, P., Li, S.: Automatic landmark estimation for adolescent idiopathic scoliosis assessment using boostnet. In: MICCAI. pp. 127-135. Springer (2017) +25. Xu, Y., Xu, D., Hong, X., Ouyang, W., Ji, R., Xu, M., Zhao, G.: Structured modeling of joint deep feature and prediction refinement for salient object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3789-3798 (2019) +26. Yang, J., Zeng, A., Li, F., Liu, S., Zhang, R., Zhang, L.: Neural interactive keypoint detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15122-15132 (2023) +27. Zhang, H., Chung, A.C.: A dual coordinate system vertebral landmark detection network with sparse-to-dense vertebral line interpolation. Bioengineering 11(1), 101 (2024) \ No newline at end of file diff --git a/2024/Bones Can't Be Triangles_ Accurate and Efficient Vertebrae Keypoint Estimation through Collaborative Error Revision/images.zip b/2024/Bones Can't Be Triangles_ Accurate and Efficient Vertebrae Keypoint Estimation through Collaborative Error Revision/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..039dd3f07a2235448b7dd1499862272e3ec92b39 --- /dev/null +++ b/2024/Bones Can't Be Triangles_ Accurate and Efficient Vertebrae Keypoint Estimation through Collaborative Error Revision/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8afbb263aab2f728466940bdbe09954b0fc56dd49d4c8fa4be9c934e37b63288 +size 759217 diff --git a/2024/Bones Can't Be Triangles_ Accurate and Efficient Vertebrae Keypoint Estimation through Collaborative Error Revision/layout.json b/2024/Bones Can't Be Triangles_ Accurate and Efficient Vertebrae Keypoint Estimation through Collaborative Error Revision/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a520aa46791067f4af4a6c1eb9090841e5201ba7 --- /dev/null +++ b/2024/Bones Can't Be Triangles_ Accurate and Efficient Vertebrae Keypoint Estimation through Collaborative Error Revision/layout.json @@ -0,0 +1,8425 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 112, + 478, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 112, + 478, + 166 + ], + "spans": [ + { + "bbox": [ + 135, + 112, + 478, + 166 + ], + "type": "text", + "content": "Bones Can't Be Triangles: Accurate and Efficient Vertebrae Keypoint Estimation through Collaborative Error Revision" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 186, + 186, + 427, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 186, + 427, + 200 + ], + "spans": [ + { + "bbox": [ + 186, + 186, + 427, + 200 + ], + "type": "text", + "content": "Jinhee Kim\\*1, Taesung Kim\\*1, and Jaegul Choo" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 208, + 414, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 208, + 414, + 240 + ], + "spans": [ + { + "bbox": [ + 198, + 208, + 414, + 240 + ], + "type": "text", + "content": "Korea Advanced Institute of Science and Technology {seharanul17, zkm1989, jchoo}@kaist.ac.kr * indicates equal contributions." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 243, + 216, + 362 + ], + "blocks": [ + { + "bbox": [ + 133, + 243, + 216, + 362 + ], + "lines": [ + { + "bbox": [ + 133, + 243, + 216, + 362 + ], + "spans": [ + { + "bbox": [ + 133, + 243, + 216, + 362 + ], + "type": "image", + "image_path": "ca6445e836a61c2d145b6f13cfaf09dea4fef77f9263271ae1fd7975ee058164.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 364, + 216, + 522 + ], + "blocks": [ + { + "bbox": [ + 133, + 364, + 216, + 522 + ], + "lines": [ + { + "bbox": [ + 133, + 364, + 216, + 522 + ], + "spans": [ + { + "bbox": [ + 133, + 364, + 216, + 522 + ], + "type": "image", + "image_path": "132500992809c6c3958818e7a42e4c2489ba2eadd87c6a10cb23401559e95777.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 524, + 192, + 534 + ], + "lines": [ + { + "bbox": [ + 156, + 524, + 192, + 534 + ], + "spans": [ + { + "bbox": [ + 156, + 524, + 192, + 534 + ], + "type": "text", + "content": "(a) Image" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 220, + 243, + 304, + 362 + ], + "blocks": [ + { + "bbox": [ + 220, + 243, + 304, + 362 + ], + "lines": [ + { + "bbox": [ + 220, + 243, + 304, + 362 + ], + "spans": [ + { + "bbox": [ + 220, + 243, + 304, + 362 + ], + "type": "image", + "image_path": "847cf46633eefd9c685d4c73531b621ad8a2e6f72b7db6627f1282407879e2cf.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 221, + 364, + 304, + 522 + ], + "blocks": [ + { + "bbox": [ + 221, + 364, + 304, + 522 + ], + "lines": [ + { + "bbox": [ + 221, + 364, + 304, + 522 + ], + "spans": [ + { + "bbox": [ + 221, + 364, + 304, + 522 + ], + "type": "image", + "image_path": "9685d7166195f6848cb71988cfdf95e39ea470e09feba1f1b3dc9713fceb971d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 523, + 297, + 534 + ], + "lines": [ + { + "bbox": [ + 228, + 523, + 297, + 534 + ], + "spans": [ + { + "bbox": [ + 228, + 523, + 297, + 534 + ], + "type": "text", + "content": "(b) Initial prediction" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 243, + 392, + 362 + ], + "blocks": [ + { + "bbox": [ + 309, + 243, + 392, + 362 + ], + "lines": [ + { + "bbox": [ + 309, + 243, + 392, + 362 + ], + "spans": [ + { + "bbox": [ + 309, + 243, + 392, + 362 + ], + "type": "image", + "image_path": "3cd1e50afc61a263e3f4f2abc30cc90802eb3743068ba91f9beb550ee72f7eb8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 364, + 392, + 522 + ], + "blocks": [ + { + "bbox": [ + 309, + 364, + 392, + 522 + ], + "lines": [ + { + "bbox": [ + 309, + 364, + 392, + 522 + ], + "spans": [ + { + "bbox": [ + 309, + 364, + 392, + 522 + ], + "type": "image", + "image_path": "0fedc50436e78b8a35f4b85c7791dab4b8cbe58fbb911a6aa16df3a7898fb8af.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 523, + 376, + 534 + ], + "lines": [ + { + "bbox": [ + 326, + 523, + 376, + 534 + ], + "spans": [ + { + "bbox": [ + 326, + 523, + 376, + 534 + ], + "type": "text", + "content": "(c) KeyBot" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 399, + 243, + 481, + 362 + ], + "blocks": [ + { + "bbox": [ + 399, + 243, + 481, + 362 + ], + "lines": [ + { + "bbox": [ + 399, + 243, + 481, + 362 + ], + "spans": [ + { + "bbox": [ + 399, + 243, + 481, + 362 + ], + "type": "image", + "image_path": "998da9e1a72fe3c716e25d0022c4dba31e48eece06842be5e6ef5a0d7f740960.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 399, + 364, + 481, + 522 + ], + "blocks": [ + { + "bbox": [ + 399, + 364, + 481, + 522 + ], + "lines": [ + { + "bbox": [ + 399, + 364, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 399, + 364, + 481, + 522 + ], + "type": "image", + "image_path": "aef0418b5efadb9b952b22f025ff400f2320b2c6109aba5707dda6a9a68b6109.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 523, + 467, + 534 + ], + "lines": [ + { + "bbox": [ + 411, + 523, + 467, + 534 + ], + "spans": [ + { + "bbox": [ + 411, + 523, + 467, + 534 + ], + "type": "text", + "content": "(d) Groundtruth" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 542, + 482, + 577 + ], + "lines": [ + { + "bbox": [ + 130, + 542, + 482, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 542, + 482, + 577 + ], + "type": "text", + "content": "Fig. 1: Vertebrae keypoint refinement results on BUU-LA [11] and AASCE [24]. Initial predictions exhibit significant errors in vertebra shapes due to difficulty identifying individual vertebrae. KeyBot effectively reduces these errors without user input." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 159, + 588, + 453, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 588, + 453, + 666 + ], + "spans": [ + { + "bbox": [ + 159, + 588, + 453, + 666 + ], + "type": "text", + "content": "Abstract. Recent advances in interactive keypoint estimation methods have enhanced accuracy while minimizing user intervention. However, these methods require user input for error correction, which can be costly in vertebrae keypoint estimation where inaccurate keypoints are densely clustered or overlap. We introduce a novel approach, KeyBot, specifically designed to identify and correct significant and typical errors in existing models, akin to user revision. By characterizing typical error types" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 117, + 455, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 117, + 455, + 183 + ], + "spans": [ + { + "bbox": [ + 159, + 117, + 455, + 183 + ], + "type": "text", + "content": "and using simulated errors for training, KeyBot effectively corrects these errors and significantly reduces user workload. Comprehensive quantitative and qualitative evaluations on three public datasets confirm that KeyBot significantly outperforms existing methods, achieving state-of-the-art performance in interactive vertebrae keypoint estimation. The source code and demo video are available on our project page." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 193, + 455, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 193, + 455, + 216 + ], + "spans": [ + { + "bbox": [ + 160, + 193, + 455, + 216 + ], + "type": "text", + "content": "Keywords: Interactive vertebrae keypoint estimation " + }, + { + "bbox": [ + 160, + 193, + 455, + 216 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 193, + 455, + 216 + ], + "type": "text", + "content": " X-ray image analysis " + }, + { + "bbox": [ + 160, + 193, + 455, + 216 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 193, + 455, + 216 + ], + "type": "text", + "content": " Automatic error correction " + }, + { + "bbox": [ + 160, + 193, + 455, + 216 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 193, + 455, + 216 + ], + "type": "text", + "content": " User interaction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 234, + 230, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 234, + 230, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 234, + 230, + 247 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 258, + 481, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 481, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 481, + 366 + ], + "type": "text", + "content": "Accurate vertebrae keypoint estimation from X-ray images is crucial for effective medical diagnosis and treatment planning [20, 23], with errors having significant impacts on clinical decisions, such as in spinal surgeries and pain management [6, 13, 15]. This task presents a distinct set of challenges due to the complex anatomy of the spine [17, 27]. The similar and repetitive shapes of vertebrae complicate distinguishing between individual vertebrae and accurately estimating their keypoints, as shown in Fig. 1. This requires a deep understanding of sequential context and anatomy, unlike more distinct keypoints found in human pose [5, 9, 14, 19] or facial keypoint [1, 3] estimation tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 366, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 521 + ], + "type": "text", + "content": "Such challenges often result in significant errors during automated vertebrae keypoint estimation, including vertex misidentification, bone misidentification, and left-right inversion, as illustrated in Fig. 2. Manual revision of these errors is time-consuming and labor-intensive, especially as the number of vertebrae increases. Missing even a single vertebra can lead to extensive adjustments, thereby significantly increasing the time and effort required for accurate annotation. To mitigate these, interactive keypoint estimation approaches have been developed, significantly improving performance with minimal user input [10,26]. These methods aim to automatically refine inaccurate keypoints with minimal user feedback, using the feedback as a hint to adjust the remaining errors, as shown in Fig. 3(a). However, they heavily rely on users to identify and correct errors, which can be time-consuming and labor-intensive, especially when erroneous keypoints are densely clustered or overlapped." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 522, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 617 + ], + "type": "text", + "content": "In response, we introduce a novel method named KeyBot, designed to autonomously detect and correct specific errors in vertebrae keypoint estimation before user evaluation, building upon existing interactive keypoint estimation models, as illustrated in Fig. 3(b). Initially, the interaction model predicts keypoints based on the input image. During refinement, KeyBot autonomously preevaluates these predictions, identifies specific errors, and provides corrective feedback similar to user input. In contrast to previous models requiring at least one user input, KeyBot autonomously refines model predictions before user input." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "We generate synthetic data characterizing typical error types, vertex misidentification, bone misidentification, and left-right inversion, and train KeyBot using this data to focus specifically on predefined error types. Independently trained from the interaction model, KeyBot effectively identifies errors that might be" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 164, + 114, + 249, + 274 + ], + "blocks": [ + { + "bbox": [ + 164, + 114, + 249, + 274 + ], + "lines": [ + { + "bbox": [ + 164, + 114, + 249, + 274 + ], + "spans": [ + { + "bbox": [ + 164, + 114, + 249, + 274 + ], + "type": "image", + "image_path": "4713e0e8374727d2eb2389713f0aac795f3b598075edb2a36d4b2b870ccc9894.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 275, + 250, + 283 + ], + "lines": [ + { + "bbox": [ + 161, + 275, + 250, + 283 + ], + "spans": [ + { + "bbox": [ + 161, + 275, + 250, + 283 + ], + "type": "text", + "content": "(a) Vertex misidentification" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 265, + 114, + 350, + 274 + ], + "blocks": [ + { + "bbox": [ + 265, + 114, + 350, + 274 + ], + "lines": [ + { + "bbox": [ + 265, + 114, + 350, + 274 + ], + "spans": [ + { + "bbox": [ + 265, + 114, + 350, + 274 + ], + "type": "image", + "image_path": "b48c1d1d2c6796b0c4f9d01233f3f962c4c7ed249243c2388fd6210c1f811e7f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 264, + 275, + 349, + 283 + ], + "lines": [ + { + "bbox": [ + 264, + 275, + 349, + 283 + ], + "spans": [ + { + "bbox": [ + 264, + 275, + 349, + 283 + ], + "type": "text", + "content": "(b) Bone misidentification" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 368, + 114, + 453, + 274 + ], + "blocks": [ + { + "bbox": [ + 368, + 114, + 453, + 274 + ], + "lines": [ + { + "bbox": [ + 368, + 114, + 453, + 274 + ], + "spans": [ + { + "bbox": [ + 368, + 114, + 453, + 274 + ], + "type": "image", + "image_path": "16ef5afecc64764dd1aace2df0b321480e920542e5dfa64e9c730b76afef6c9b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 275, + 448, + 284 + ], + "lines": [ + { + "bbox": [ + 374, + 275, + 448, + 284 + ], + "spans": [ + { + "bbox": [ + 374, + 275, + 448, + 284 + ], + "type": "text", + "content": "(c) Left-right inversion" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 292, + 480, + 304 + ], + "lines": [ + { + "bbox": [ + 132, + 292, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 292, + 480, + 304 + ], + "type": "text", + "content": "Fig. 2: Error types in vertebrae keypoint estimation, illustrated on the AASCE dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 480, + 365 + ], + "type": "text", + "content": "overlooked by the interaction model, significantly reducing the user's revision efforts. Rather than correcting errors from scratch, users can focus on verifying and fine-tuning results, with KeyBot having already eliminated major errors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 365, + 481, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 481, + 461 + ], + "type": "text", + "content": "KeyBot delivers quick, cost-effective, and iterative feedback without user input, which is crucial for complex cases requiring multiple refinements. This automated feedback is typically much faster than user review, significantly enhancing workflow efficiency. The collaborative feedback loop between the user, KeyBot, and the interaction model enriches the process by combining expert medical insights with KeyBot's focus on typical error types. This synergy addresses a broad range of potential errors and enhancements, leading to an overall improvement in keypoint estimation accuracy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 461, + 481, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 481, + 545 + ], + "type": "text", + "content": "Through extensive experiments on three public datasets [11, 24], KeyBot demonstrates significant improvements in vertebrae keypoint estimation accuracy. The model reduces the mean radial error (MRE) by " + }, + { + "bbox": [ + 130, + 461, + 481, + 545 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 130, + 461, + 481, + 545 + ], + "type": "text", + "content": " and decreases the number of user clicks (NoC) required for target performance by " + }, + { + "bbox": [ + 130, + 461, + 481, + 545 + ], + "type": "inline_equation", + "content": "17\\%" + }, + { + "bbox": [ + 130, + 461, + 481, + 545 + ], + "type": "text", + "content": " on the AASCE dataset, achieving state-of-the-art performance. These results highlight the significance of KeyBot in advancing vertebrae keypoint estimation, offering an advanced, user-centric solution to this unique and complex challenge." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 665 + ], + "type": "text", + "content": "In summary, our contributions are: (1) We introduce a novel method, KeyBot, designed to provide automated corrections for specific errors in vertebrae keypoint estimation. This enhances existing interactive keypoint estimation frameworks and improves overall keypoint estimation accuracy and efficiency. (2) We train KeyBot on synthetic data representing common error types in vertebrae, enabling it to effectively identify and address these errors. (3) KeyBot significantly reduces the need for user intervention during the keypoint annotation process. (4) Extensive evaluation across three public datasets confirms the efficacy of our proposed method in improving keypoint estimation accuracy and significantly reducing user intervention." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 342, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 342, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bones Can't Be Triangles" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 116, + 474, + 345 + ], + "blocks": [ + { + "bbox": [ + 132, + 116, + 474, + 345 + ], + "lines": [ + { + "bbox": [ + 132, + 116, + 474, + 345 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 474, + 345 + ], + "type": "image", + "image_path": "123c58c08b8eb78e2db6d319ad6fb374a38f2f4682ac3ad59224e3551e70f5a5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 353, + 482, + 420 + ], + "lines": [ + { + "bbox": [ + 130, + 353, + 482, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 482, + 420 + ], + "type": "text", + "content": "Fig. 3: Overview of (a) the interactive keypoint estimation and (b) the KeyBot framework. The interaction model generates initial keypoint predictions from an image, which are then revised by users, leading to updated results. This approach requires user input for error correction. In contrast, KeyBot offers a cost-efficient feedback mechanism via automated, rapid, and iterative refinement without user input. It independently identifies and corrects major errors, allowing users to focus on the final adjustments." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 444, + 234, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 444, + 234, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 444, + 234, + 456 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 473, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 616 + ], + "type": "text", + "content": "Automated error refinement. Despite the advancements in deep learning, errors in model predictions remain a formidable challenge. To mitigate these inaccuracies, a specialized area of research dedicated to refining prediction results has emerged, leading to the development of advanced methods for identifying and correcting errors in model outputs [4,14,16,25]. For example, Qu et al. [16] introduced a method leveraging attention maps to detect label errors in CT data, markedly improving annotation efficiency and facilitating the collection of extensive datasets. Similarly, efforts to refine keypoint estimation have been noteworthy. For instance, Ronchi et al. [18] analyzed human pose estimation results to categorize three main types of errors, while Moon et al. [14] developed a method that specifically addresses and corrects these categorized errors, achieving state-of-the-art human pose estimation performances." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Despite these advancements, estimating vertebrae keypoints from X-ray images introduces additional challenges due to the complex sequential structure of the spine. Error patterns, such as the mislocalization of an entire vertebra, often require extensive user inputs for correction. In response, we propose KeyBot," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "a method tailored for revising errors specific to vertebrae keypoint estimation, effectively addressing the unique challenges of this task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 295 + ], + "type": "text", + "content": "User interaction. To maximize the accuracy of predictions while minimizing the need for user input, the development of interaction-based models is paramount. Interactive segmentation is a particularly active area within this domain. Backpropagating refinement schemes such as BRS [8] and f-BRS [21] have improved interactive image segmentation by incorporating user hints through backpropagation. Similarly, RITM [22] has shown superior performance by integrating an iterative refinement process into its training process. In the field of keypoint estimation, interactive keypoint estimation methods have been researched to reduce user input through interactive refinement frameworks. Kim et al. [10] pioneered an interactive framework for keypoint estimation, significantly improving the accuracy and efficiency of the keypoint annotation process in medical datasets. ClickPose [26] developed an interactive keypoint estimation model for multi-person 2D keypoints, significantly reducing annotation costs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 296, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 439 + ], + "type": "text", + "content": "Recently, Liu et al. [12] introduced a method that generates pseudo clicks to mimic human interactions for refining segmentation predictions autonomously, thereby minimizing the required number of user clicks. However, their error decoder is trained end-to-end and can thus be affected by inherent biases or limitations of the segmentation module. To address this challenge, our KeyBot is trained independently to correct simulated errors, which allows KeyBot to provide corrective feedback free from any potential biases present in the keypoint estimation model. Unlike traditional interactive keypoint estimation models that rely on users to identify and correct errors, KeyBot autonomously generates corrective feedback. This significant innovation not only enhances the keypoint estimation accuracy but also remarkably reduces the need for user interaction, thereby streamlining the error correction process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 457, + 201, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 201, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 201, + 472 + ], + "type": "text", + "content": "3 KeyBot" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 483, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 591 + ], + "type": "text", + "content": "We introduce KeyBot, a novel method designed to enhance vertebrae keypoint estimation by identifying and correcting errors prior to user intervention. We employ an existing interactive model that can incorporate user feedback for interactive modifications. Given model predictions, KeyBot identifies erroneous keypoints and generates their pseudo-corrections, which are then fed to the interaction model for updates. KeyBot consists of two components: The detector discerns whether each input keypoint is accurate, and the corrector refines input keypoints accurately. Both components are trained with synthetic errors to focus specifically on the target error types." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 609, + 334, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 334, + 621 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 334, + 621 + ], + "type": "text", + "content": "3.1 Overview of the propose approach" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Interactive keypoint estimation framework. Consider an interactive keypoint estimation model " + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": ", operating on an image " + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": ". The model aims to precisely estimate " + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": " target keypoints through user-interactive" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 342, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 342, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bones Can't Be Triangles" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 154, + 113, + 458, + 298 + ], + "blocks": [ + { + "bbox": [ + 134, + 133, + 147, + 144 + ], + "lines": [ + { + "bbox": [ + 134, + 133, + 147, + 144 + ], + "spans": [ + { + "bbox": [ + 134, + 133, + 147, + 144 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 154, + 113, + 458, + 298 + ], + "lines": [ + { + "bbox": [ + 154, + 113, + 458, + 298 + ], + "spans": [ + { + "bbox": [ + 154, + 113, + 458, + 298 + ], + "type": "image", + "image_path": "e9134607dfc9aa107aa066d901484079bff86056ec607e6a9a6c70af4c985650.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 310, + 482, + 366 + ], + "lines": [ + { + "bbox": [ + 131, + 310, + 482, + 366 + ], + "spans": [ + { + "bbox": [ + 131, + 310, + 482, + 366 + ], + "type": "text", + "content": "Fig. 4: Training overview of KeyBot. KeyBot consists of two main components, the detector and the corrector. (a) Synthetic errors are introduced to the groundtruth keypoints, creating inaccurate keypoints and corresponding anomaly labels. (b) The detector is trained to discern whether each input keypoint is accurate or not. (c) The corrector is trained to refine these inaccurate keypoints accurately." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 389, + 482, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 389, + 482, + 414 + ], + "spans": [ + { + "bbox": [ + 131, + 389, + 482, + 414 + ], + "type": "text", + "content": "refinement. In each iteration " + }, + { + "bbox": [ + 131, + 389, + 482, + 414 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 389, + 482, + 414 + ], + "type": "text", + "content": ", the model generates the " + }, + { + "bbox": [ + 131, + 389, + 482, + 414 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 389, + 482, + 414 + ], + "type": "text", + "content": "-th prediction, " + }, + { + "bbox": [ + 131, + 389, + 482, + 414 + ], + "type": "inline_equation", + "content": "\\pmb{y}_t" + }, + { + "bbox": [ + 131, + 389, + 482, + 414 + ], + "type": "text", + "content": ", using three inputs: the image " + }, + { + "bbox": [ + 131, + 389, + 482, + 414 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 131, + 389, + 482, + 414 + ], + "type": "text", + "content": ", user revisions " + }, + { + "bbox": [ + 131, + 389, + 482, + 414 + ], + "type": "inline_equation", + "content": "\\pmb{u}_t" + }, + { + "bbox": [ + 131, + 389, + 482, + 414 + ], + "type": "text", + "content": ", and its previous predictions " + }, + { + "bbox": [ + 131, + 389, + 482, + 414 + ], + "type": "inline_equation", + "content": "\\pmb{e}_t" + }, + { + "bbox": [ + 131, + 389, + 482, + 414 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 264, + 425, + 480, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 425, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 264, + 425, + 480, + 437 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} _ {t} = \\mathcal {F} _ {\\theta} (\\boldsymbol {x}, \\boldsymbol {u} _ {t}, \\boldsymbol {e} _ {t}), \\tag {1}", + "image_path": "6a8a7583ab56f2cc8890ba0c91696bfc07d95c8f973264ece978bd864bd51294.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 442, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 442, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 131, + 442, + 482, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 442, + 482, + 479 + ], + "type": "inline_equation", + "content": "\\pmb{u}_0\\coloneqq 0" + }, + { + "bbox": [ + 131, + 442, + 482, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 442, + 482, + 479 + ], + "type": "inline_equation", + "content": "\\pmb {e}_0\\coloneqq 0" + }, + { + "bbox": [ + 131, + 442, + 482, + 479 + ], + "type": "text", + "content": ". Subsequently, a user reviews the results, and the adjustment made to a keypoint position is encoded as Gaussian heatmaps " + }, + { + "bbox": [ + 131, + 442, + 482, + 479 + ], + "type": "inline_equation", + "content": "\\pmb{u}_t^i" + }, + { + "bbox": [ + 131, + 442, + 482, + 479 + ], + "type": "text", + "content": ", effectively capturing and representing the refined locations:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 250, + 487, + 480, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 487, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 250, + 487, + 480, + 500 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\rho} _ {t + 1}, \\boldsymbol {u} _ {t + 1} = \\Psi (\\boldsymbol {x}, \\boldsymbol {y} _ {t}, \\boldsymbol {u} _ {t}), \\tag {2}", + "image_path": "08c081d17b2a7ed8aef0c52150e3814f26a3c4237e16fae3ab256759b0b7be8f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 509, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 509, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 131, + 509, + 482, + 545 + ], + "type": "text", + "content": "where the set " + }, + { + "bbox": [ + 131, + 509, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\rho_{t}" + }, + { + "bbox": [ + 131, + 509, + 482, + 545 + ], + "type": "text", + "content": " represents keypoint indices revised by the user over " + }, + { + "bbox": [ + 131, + 509, + 482, + 545 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 509, + 482, + 545 + ], + "type": "text", + "content": " iterations. Next, we selectively gather the known-to-be-false predictions from " + }, + { + "bbox": [ + 131, + 509, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 131, + 509, + 482, + 545 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 131, + 509, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\pmb{e}" + }, + { + "bbox": [ + 131, + 509, + 482, + 545 + ], + "type": "text", + "content": " and reintroduce them as inputs in the next step to differentiate mispredictions [10]:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 261, + 552, + 480, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 552, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 261, + 552, + 480, + 567 + ], + "type": "interline_equation", + "content": "\\boldsymbol {e} _ {t + 1} ^ {i} = \\boldsymbol {y} _ {t} ^ {i} \\text {f o r} i \\in \\boldsymbol {\\rho} _ {t}. \\tag {3}", + "image_path": "e4857777dbb60eaa3a518c52fccb3de0a69ddc5c0d6ca46149eb9c457fd565cf.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 575, + 482, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 575, + 482, + 611 + ], + "spans": [ + { + "bbox": [ + 131, + 575, + 482, + 611 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 575, + 482, + 611 + ], + "type": "inline_equation", + "content": "\\pmb{e}" + }, + { + "bbox": [ + 131, + 575, + 482, + 611 + ], + "type": "text", + "content": " denotes predictions modified by the user. These updates on " + }, + { + "bbox": [ + 131, + 575, + 482, + 611 + ], + "type": "inline_equation", + "content": "\\pmb{u}_t" + }, + { + "bbox": [ + 131, + 575, + 482, + 611 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 575, + 482, + 611 + ], + "type": "inline_equation", + "content": "\\pmb{e}_t" + }, + { + "bbox": [ + 131, + 575, + 482, + 611 + ], + "type": "text", + "content": " are critical for refining the model predictions, leading to the updated keypoint predictions reflecting the user feedback:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 253, + 620, + 480, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 620, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 253, + 620, + 480, + 633 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} _ {t + 1} = \\mathcal {F} _ {\\theta} (\\boldsymbol {x}, \\boldsymbol {u} _ {t + 1}, \\boldsymbol {e} _ {t + 1}). \\tag {4}", + "image_path": "e37ba2a5a80cfcc49a922d07fb5dd545cda2f49925237c511ca61feff4fac6b1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "This process repeats until sufficient accuracy is achieved or the predefined maximum user interaction " + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": " is reached." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 201 + ], + "type": "text", + "content": "The KeyBot framework. KeyBot offers an automated keypoint refinement step that precedes and complements the existing user-based refinement, effectively reducing the need for user interaction in the process. KeyBot consists of the detector " + }, + { + "bbox": [ + 131, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\phi_r}" + }, + { + "bbox": [ + 131, + 116, + 482, + 201 + ], + "type": "text", + "content": " and corrector " + }, + { + "bbox": [ + 131, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\phi_s}" + }, + { + "bbox": [ + 131, + 116, + 482, + 201 + ], + "type": "text", + "content": ", which together identifies and corrects inaccurate keypoints predicted by " + }, + { + "bbox": [ + 131, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 131, + 116, + 482, + 201 + ], + "type": "text", + "content": ". Following each prediction iteration of " + }, + { + "bbox": [ + 131, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 131, + 116, + 482, + 201 + ], + "type": "text", + "content": ", KeyBot evaluates " + }, + { + "bbox": [ + 131, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\pmb{y}_t" + }, + { + "bbox": [ + 131, + 116, + 482, + 201 + ], + "type": "text", + "content": " and suggests necessary corrections, identifying erroneous keypoints " + }, + { + "bbox": [ + 131, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\pmb{\\nu}_{t,n+1}" + }, + { + "bbox": [ + 131, + 116, + 482, + 201 + ], + "type": "text", + "content": " and their pseudo corrections " + }, + { + "bbox": [ + 131, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\pmb{z}_{t,n+1}" + }, + { + "bbox": [ + 131, + 116, + 482, + 201 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 205, + 206, + 481, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 206, + 481, + 220 + ], + "spans": [ + { + "bbox": [ + 205, + 206, + 481, + 220 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\nu} _ {t, n + 1} = \\mathcal {G} _ {\\phi_ {r}} (\\boldsymbol {x}, \\boldsymbol {y} _ {t, n}) \\text {a n d} \\boldsymbol {z} _ {t, n + 1} = \\mathcal {G} _ {\\phi_ {s}} (\\boldsymbol {x}, \\boldsymbol {y} _ {t, n}), \\tag {5}", + "image_path": "29dc06ef578091e0f67dd063905035c847d3911f8310f08a4aa1b140f427ba2d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 224, + 482, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 224, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 131, + 224, + 482, + 320 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 131, + 224, + 482, + 320 + ], + "type": "inline_equation", + "content": "\\pmb{y}_{t,0} \\coloneqq \\pmb{y}_t" + }, + { + "bbox": [ + 131, + 224, + 482, + 320 + ], + "type": "text", + "content": ". This step is similar to the user interaction step outlined in Eq. (2). Here, KeyBot can detect multiple inaccurate keypoints at once and simultaneously offer corrective feedback for each identified error. If no inaccuracies are detected by the detector, KeyBot concludes and " + }, + { + "bbox": [ + 131, + 224, + 482, + 320 + ], + "type": "inline_equation", + "content": "\\pmb{y}_{t,n}" + }, + { + "bbox": [ + 131, + 224, + 482, + 320 + ], + "type": "text", + "content": " becomes the final prediction. Revision information, " + }, + { + "bbox": [ + 131, + 224, + 482, + 320 + ], + "type": "inline_equation", + "content": "\\pmb{c}_{t,n}" + }, + { + "bbox": [ + 131, + 224, + 482, + 320 + ], + "type": "text", + "content": ", now includes both user " + }, + { + "bbox": [ + 131, + 224, + 482, + 320 + ], + "type": "inline_equation", + "content": "(\\pmb{u}_{t,n})" + }, + { + "bbox": [ + 131, + 224, + 482, + 320 + ], + "type": "text", + "content": " and model " + }, + { + "bbox": [ + 131, + 224, + 482, + 320 + ], + "type": "inline_equation", + "content": "(\\pmb{z}_{t,n})" + }, + { + "bbox": [ + 131, + 224, + 482, + 320 + ], + "type": "text", + "content": " corrections. Here, user modifications are prioritized as the definitive ground truth. Thus, keypoints already revised by the user are exempt from further updates by the pseudo-corrections from KeyBot." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 234, + 325, + 481, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 325, + 481, + 339 + ], + "spans": [ + { + "bbox": [ + 234, + 325, + 481, + 339 + ], + "type": "interline_equation", + "content": "\\boldsymbol {c} _ {t, n + 1} ^ {i} = \\boldsymbol {z} _ {t, n + 1} ^ {i} \\text {f o r} i \\in \\boldsymbol {\\nu} _ {t, n + 1} \\backslash \\boldsymbol {\\rho} _ {t}. \\tag {6}", + "image_path": "fbcec041e744f692ad01087a369da833a79ba002c6a45f0d5abf705c665b50f0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 343, + 482, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 343, + 482, + 428 + ], + "spans": [ + { + "bbox": [ + 131, + 343, + 482, + 428 + ], + "type": "text", + "content": "Previous work [10,22] relies solely on the prediction at the most recent step as the previous prediction, as described in Eq. (3). However, this approach can result in losing track of erroneous predictions once they are corrected in the model. Hence, our approach accumulates predictions identified as erroneous by either KeyBot or the user across all iterations. These false predictions retain the history of incorrect predictions and serve to inform the model about specific prediction errors to avoid, thereby guiding it toward more accurate future predictions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 225, + 432, + 481, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 432, + 481, + 447 + ], + "spans": [ + { + "bbox": [ + 225, + 432, + 481, + 447 + ], + "type": "interline_equation", + "content": "\\boldsymbol {e} _ {t, n + 1} ^ {i} = \\boldsymbol {y} _ {t, n} ^ {i} \\text {f o r} i \\in \\boldsymbol {\\nu} _ {t, n + 1} \\backslash \\boldsymbol {\\nu} _ {t,: n} \\backslash \\boldsymbol {\\rho} _ {t}. \\tag {7}", + "image_path": "a784a5bb415b4064be5d86f4e68506e1290a10524b8423ce6c86493bfc7f6a0b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 450, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 450, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 131, + 450, + 482, + 487 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 450, + 482, + 487 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 131, + 450, + 482, + 487 + ], + "type": "text", + "content": " denotes predictions modified by either the user or KeyBot. Finally, these are then fed back into the backbone interaction network for a new, corrected prediction akin to user feedback:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 245, + 493, + 481, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 493, + 481, + 506 + ], + "spans": [ + { + "bbox": [ + 245, + 493, + 481, + 506 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} _ {t, n + 1} = \\mathcal {F} _ {\\theta} (\\boldsymbol {x}, \\boldsymbol {c} _ {t, n + 1}, \\boldsymbol {e} _ {t, n + 1}). \\tag {8}", + "image_path": "b27b90430867622b1ae649970f5c430b47f53e5f544cd51455ff53ef90f438b3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 510, + 482, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 510, + 482, + 595 + ], + "spans": [ + { + "bbox": [ + 131, + 510, + 482, + 595 + ], + "type": "text", + "content": "Each iteration of KeyBot contributes progressively towards enhancing the overall accuracy and reliability of the keypoint predictions. This cycle repeats until the detector detects no errors or the maximum KeyBot iteration count " + }, + { + "bbox": [ + 131, + 510, + 482, + 595 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 510, + 482, + 595 + ], + "type": "text", + "content": " is reached. Upon the KeyBot process concludes, say, after " + }, + { + "bbox": [ + 131, + 510, + 482, + 595 + ], + "type": "inline_equation", + "content": "n + 1" + }, + { + "bbox": [ + 131, + 510, + 482, + 595 + ], + "type": "text", + "content": " iterations, the prediction " + }, + { + "bbox": [ + 131, + 510, + 482, + 595 + ], + "type": "inline_equation", + "content": "\\pmb{y}_{t,n + 1}" + }, + { + "bbox": [ + 131, + 510, + 482, + 595 + ], + "type": "text", + "content": " is established as the final output for the " + }, + { + "bbox": [ + 131, + 510, + 482, + 595 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 510, + 482, + 595 + ], + "type": "text", + "content": "-th iteration of the interactive keypoint estimation process. For the subsequent " + }, + { + "bbox": [ + 131, + 510, + 482, + 595 + ], + "type": "inline_equation", + "content": "(t + 1)" + }, + { + "bbox": [ + 131, + 510, + 482, + 595 + ], + "type": "text", + "content": "-th user-side phase, given new user feedback as Eq. (2), " + }, + { + "bbox": [ + 131, + 510, + 482, + 595 + ], + "type": "inline_equation", + "content": "\\pmb{c}_{t,n + 1}" + }, + { + "bbox": [ + 131, + 510, + 482, + 595 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 510, + 482, + 595 + ], + "type": "inline_equation", + "content": "\\pmb{e}_{t,n + 1}" + }, + { + "bbox": [ + 131, + 510, + 482, + 595 + ], + "type": "text", + "content": " are updated to " + }, + { + "bbox": [ + 131, + 510, + 482, + 595 + ], + "type": "inline_equation", + "content": "\\pmb{c}_{t + 1,0}" + }, + { + "bbox": [ + 131, + 510, + 482, + 595 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 510, + 482, + 595 + ], + "type": "inline_equation", + "content": "\\pmb{e}_{t + 1,0}" + }, + { + "bbox": [ + 131, + 510, + 482, + 595 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 211, + 600, + 481, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 600, + 481, + 613 + ], + "spans": [ + { + "bbox": [ + 211, + 600, + 481, + 613 + ], + "type": "interline_equation", + "content": "\\boldsymbol {c} _ {t + 1, 0} = \\boldsymbol {u} _ {t + 1} \\text {a n d} \\boldsymbol {e} _ {t + 1, 0} ^ {i} = \\boldsymbol {y} _ {t, n + 1} ^ {i} \\text {f o r} i \\in \\boldsymbol {\\rho} _ {t + 1}, \\tag {9}", + "image_path": "5d9baf08912673764447d7760e854ac96cd66ebd0122d2dd7eb3d059ac1bd7e4.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 617, + 320, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 320, + 630 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 320, + 630 + ], + "type": "text", + "content": "and the keypoints are revised accordingly:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 246, + 636, + 481, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 636, + 481, + 649 + ], + "spans": [ + { + "bbox": [ + 246, + 636, + 481, + 649 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} _ {t + 1, 0} = \\mathcal {F} _ {\\theta} (\\boldsymbol {x}, \\boldsymbol {c} _ {t + 1, 0}, \\boldsymbol {e} _ {t + 1, 0}). \\tag {10}", + "image_path": "88fd18af725d0ea07812a0d095b8189f8c83db2796856b119bd3f35e52bcf260.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 653, + 444, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 653, + 444, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 653, + 444, + 666 + ], + "type": "text", + "content": "The complete procedure is summarized in Algorithm 1 in Appendix B." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 342, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 342, + 91, + 448, + 102 + ], + "type": "text", + "content": "Bones Can't Be Triangles" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 480, + 279 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 480, + 279 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 480, + 279 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 480, + 279 + ], + "type": "image", + "image_path": "465d142995c3e4d69f7368bba681efcd8c913cc9152991d1b7a4a46a67f9945e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 288, + 480, + 300 + ], + "lines": [ + { + "bbox": [ + 132, + 288, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 288, + 480, + 300 + ], + "type": "text", + "content": "Fig. 5: Implementing realistic estimation inaccuracy simulations for training KeyBot." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 323, + 284, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 323, + 284, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 323, + 284, + 336 + ], + "type": "text", + "content": "3.2 Model design of KeyBot" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 345, + 479, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 479, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 479, + 367 + ], + "type": "text", + "content": "KeyBot integrates the detector and corrector to identify potential errors and generates corrective feedback for their revision." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 369, + 480, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 480, + 488 + ], + "type": "text", + "content": "Error detector. The error detector specializes in identifying up to " + }, + { + "bbox": [ + 130, + 369, + 480, + 488 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 369, + 480, + 488 + ], + "type": "text", + "content": " erroneous keypoints, focusing on prevalent error types in vertebrae keypoint estimation. Our empirical analysis indicates that evaluating the entire bone structure at once is less effective for error detection. Therefore, the detector evaluates keypoints in smaller groups, examining " + }, + { + "bbox": [ + 130, + 369, + 480, + 488 + ], + "type": "inline_equation", + "content": "k \\leq K" + }, + { + "bbox": [ + 130, + 369, + 480, + 488 + ], + "type": "text", + "content": " keypoints simultaneously. This approach simplifies the task and improves its performance. During training, we randomly select a set of consecutive groundtruth keypoints " + }, + { + "bbox": [ + 130, + 369, + 480, + 488 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^*" + }, + { + "bbox": [ + 130, + 369, + 480, + 488 + ], + "type": "text", + "content": " and either maintain their accuracy or introduce errors. The cropped image area around these keypoints, along with their corresponding Gaussian keypoint heatmaps, are input to the detector. It then predicts the likelihood of error presence in each keypoint." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 489, + 480, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 489, + 480, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 489, + 480, + 559 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 489, + 480, + 559 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 130, + 489, + 480, + 559 + ], + "type": "text", + "content": " denote the function that induces synthetic errors in groundtruth keypoints, and " + }, + { + "bbox": [ + 130, + 489, + 480, + 559 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 489, + 480, + 559 + ], + "type": "text", + "content": " represents the anomaly labels for the keypoints. The detector performs multi-label binary classification to differentiate between normal and abnormal keypoints using a sigmoid function in its output layer. It is supervised to minimize the Binary Cross-Entropy (BCE) loss between the anomaly labels " + }, + { + "bbox": [ + 130, + 489, + 480, + 559 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 489, + 480, + 559 + ], + "type": "text", + "content": " and the predictions:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 220, + 571, + 481, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 571, + 481, + 585 + ], + "spans": [ + { + "bbox": [ + 220, + 571, + 481, + 585 + ], + "type": "interline_equation", + "content": "\\phi_ {r} \\leftarrow \\phi_ {r} - \\eta_ {r} \\nabla_ {\\phi_ {r}} \\mathcal {L} _ {b c e} (v, \\mathcal {G} _ {\\phi_ {r}} (\\boldsymbol {x}, \\varepsilon (\\boldsymbol {y} ^ {*})), \\tag {11}", + "image_path": "366a8705b062a459c32433b8f6daaf9a8cc1e1b90a93f5baff283bc6b3562303.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 594, + 373, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 373, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 373, + 605 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 132, + 594, + 373, + 605 + ], + "type": "inline_equation", + "content": "\\eta_r" + }, + { + "bbox": [ + 132, + 594, + 373, + 605 + ], + "type": "text", + "content": " representing the learning rate for the detector." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "During inference, the detector iteratively detects errors across all keypoints, examining " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": " keypoints at a time with a stride of " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": ". To facilitate a comprehensive assessment of each keypoint in different contexts, we set the stride " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": " to be smaller than " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": ". This results in multiple predictions for a single keypoint, and if any prediction indicates an anomaly, the keypoint is flagged as erroneous." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 224 + ], + "type": "text", + "content": "Error corrector. The error corrector aims to generate pseudo-corrections for keypoints inaccurately identified by the detector, thus providing feedback to the model " + }, + { + "bbox": [ + 131, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 131, + 116, + 482, + 224 + ], + "type": "text", + "content": ". During training, we randomly transform the groundtruth keypoints to erroneous ones using " + }, + { + "bbox": [ + 131, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 131, + 116, + 482, + 224 + ], + "type": "text", + "content": ". Taking these keypoints and the image " + }, + { + "bbox": [ + 131, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 131, + 116, + 482, + 224 + ], + "type": "text", + "content": " as inputs, the corrector is trained to revert these to their original, accurate state, focusing particularly on the typical error types implemented by " + }, + { + "bbox": [ + 131, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 131, + 116, + 482, + 224 + ], + "type": "text", + "content": ". It generates " + }, + { + "bbox": [ + 131, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 116, + 482, + 224 + ], + "type": "text", + "content": "-channel heatmaps indicating reconstructed keypoint locations, employing a sigmoid function in its output. Its learning objective is the BCE loss between the ground truth keypoint heatmaps and the reconstructed outputs:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 223, + 234, + 481, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 234, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 223, + 234, + 481, + 247 + ], + "type": "interline_equation", + "content": "\\phi_ {s} \\leftarrow \\phi_ {s} - \\eta_ {s} \\nabla_ {\\phi_ {s}} \\mathcal {L} _ {b c e} (\\boldsymbol {y}, \\mathcal {G} _ {\\phi_ {s}} (\\boldsymbol {x}, \\varepsilon (\\boldsymbol {y} ^ {*})), \\tag {12}", + "image_path": "5cc74d712ff1bb8d2c481511937622d8e4910db025e1e01939346fa81061e38c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 256, + 337, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 256, + 337, + 268 + ], + "spans": [ + { + "bbox": [ + 131, + 256, + 337, + 268 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 256, + 337, + 268 + ], + "type": "inline_equation", + "content": "\\eta_{s}" + }, + { + "bbox": [ + 131, + 256, + 337, + 268 + ], + "type": "text", + "content": " is the learning rate for the corrector." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 268, + 481, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 268, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 131, + 268, + 481, + 304 + ], + "type": "text", + "content": "During inference, the corrector refines predictions from the backbone model " + }, + { + "bbox": [ + 131, + 268, + 481, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 131, + 268, + 481, + 304 + ], + "type": "text", + "content": ", and the output corresponding to inaccurately assessed keypoints by the detector is fed to " + }, + { + "bbox": [ + 131, + 268, + 481, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 131, + 268, + 481, + 304 + ], + "type": "text", + "content": " to serve as pseudo-corrections." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 322, + 392, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 322, + 392, + 335 + ], + "spans": [ + { + "bbox": [ + 131, + 322, + 392, + 335 + ], + "type": "text", + "content": "3.3 Training KeyBot via realistic error simulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 342, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 437 + ], + "type": "text", + "content": "KeyBot aims to address significant and typical errors in vertebrae keypoint estimation. The key idea is to train KeyBot to identify and correct three error types commonly observed in vertebrae keypoint estimation: vertex misidentification, bone misidentification, and left-right inversion. During training, KeyBot is provided with both accurate keypoints and keypoints with synthetic errors. We randomly simulate the predefined error types, as shown in Fig. 5. Rather than relying on highly variable and not clearly defined real errors, we use characterized and consistent synthetic errors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 437, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 486 + ], + "type": "text", + "content": "By employing this training approach, we expose KeyBot to a wide range of error scenarios, equipping it with the proficiency to detect and correct these errors effectively. As a result, KeyBot provides a targeted evaluation of the keypoints predicted by the interaction model, focusing on specific error types." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 486, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 486, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 131, + 486, + 482, + 570 + ], + "type": "text", + "content": "Vertex misidentification (misvertex). Misvertex is a type of error where a keypoint is correctly identified on a vertex of a vertebrae but is incorrectly attributed to the wrong vertex. For instance, a keypoint intended for the first vertex is inaccurately predicted at the third vertex. To simulate misvertex, we randomly displace a keypoint to its neighbors within a " + }, + { + "bbox": [ + 131, + 486, + 482, + 570 + ], + "type": "inline_equation", + "content": "[-r, +r]" + }, + { + "bbox": [ + 131, + 486, + 482, + 570 + ], + "type": "text", + "content": " index range for a predefined number " + }, + { + "bbox": [ + 131, + 486, + 482, + 570 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 131, + 486, + 482, + 570 + ], + "type": "text", + "content": ", excluding its original position. Keypoints moving beyond their index range wrap around cyclically." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 570, + 482, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 570, + 482, + 618 + ], + "spans": [ + { + "bbox": [ + 131, + 570, + 482, + 618 + ], + "type": "text", + "content": "Bone Misidentification (misbone). Misbone errors represent errors where the entire vertebra is incorrectly classified. Our simulation randomly shifts keypoints up or down by one vertebra. This is achieved by selecting start and end keypoint indices and moving all-encompassed keypoints to an adjacent vertebra." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 482, + 665 + ], + "type": "text", + "content": "Left-right inversion (lr-inversion). LR-inversion occurs when keypoint pairs are incorrectly positioned on the opposite lateral side. To simulate LR-inversion in our model, we randomly interchange the positions of keypoint pairs that delineate one side of each vertebra." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 342, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 342, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bones Can't Be Triangles" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 145, + 470, + 275 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 1: Performance comparison of mean radial error in keypoint estimation across three datasets. UC denotes the count of user clicks provided to the model." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 145, + 470, + 275 + ], + "lines": [ + { + "bbox": [ + 140, + 145, + 470, + 275 + ], + "spans": [ + { + "bbox": [ + 140, + 145, + 470, + 275 + ], + "type": "table", + "html": "
MethodAASCEBUU-APBUU-LA
-UC1UC2UC3UC4-UC1UC2-UC1UC2
BRS [8]45.6535.8829.9325.4021.9751.2230.9136.1740.2043.1935.07
f-BRS [21]64.0657.3052.5347.8343.7944.0524.9917.5236.0327.9319.29
RITM [22]56.0337.9631.0326.7623.7936.4325.7617.9623.2717.799.89
Kim et al. [10]51.5830.6025.7821.6019.0842.3123.2616.4623.4314.2910.29
Click-Pose [26]54.6546.5044.0841.7340.0432.7229.3026.2033.7021.6217.38
KeyBot-i144.1825.9320.6618.0316.3732.0121.7915.9718.7713.399.12
KeyBot-i242.5227.0322.9518.7216.2331.8820.6515.8119.1113.479.03
KeyBot-i341.7026.5925.0221.2316.7631.8720.6615.8418.7413.368.97
", + "image_path": "a55ed82ebc087172044aaadf5902ed130e81367a5c31fbbe144b62f26621aca9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 306, + 229, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 306, + 229, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 306, + 229, + 319 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 332, + 261, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 261, + 345 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 261, + 345 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 354, + 481, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 481, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 481, + 484 + ], + "type": "text", + "content": "Datasets. We rigorously validate our method across three publicly available real-world datasets. The AASCE dataset [24] consists of 608 spinal anterior-posterior radiographs, each annotated with 68 keypoints, marking the vertices of 17 vertebrae. The BUU dataset [11] features two distinct subsets from 400 unique patients: BUU-AP with 400 anterior-posterior (AP) view radiographs, each annotated with 20 keypoints, and BUU-LA with 400 left lateral (LA) view radiographs, each annotated with 22 spinal keypoints. Images that exhibit critical keypoint annotation errors, such as those with incorrectly ordered keypoint indices similar to typical errors in keypoint estimation models, are excluded from the datasets. Comprehensive details about the dataset, including the data split and examples of the excluded images, are available in Appendix C." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 485, + 481, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 481, + 544 + ], + "type": "text", + "content": "Metrics. Following the evaluation protocol of previous work [10, 26], we evaluate the model performance using the mean radial error (MRE) and the number of clicks (NoC). MRE computes the mean Euclidean distance between the groundtruth and predicted keypoints. No " + }, + { + "bbox": [ + 130, + 485, + 481, + 544 + ], + "type": "inline_equation", + "content": "C_a" + }, + { + "bbox": [ + 130, + 485, + 481, + 544 + ], + "type": "text", + "content": " @ " + }, + { + "bbox": [ + 130, + 485, + 481, + 544 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 485, + 481, + 544 + ], + "type": "text", + "content": " measures the average number of user clicks required to achieve a target MRE of " + }, + { + "bbox": [ + 130, + 485, + 481, + 544 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 485, + 481, + 544 + ], + "type": "text", + "content": ", given a maximum of " + }, + { + "bbox": [ + 130, + 485, + 481, + 544 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 485, + 481, + 544 + ], + "type": "text", + "content": " clicks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 481, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 594 + ], + "type": "text", + "content": "Baseline models. We compare our method with state-of-the-art interactive keypoint estimation models, including the model proposed by Kim et al. [10] and Click-Pose [26]. Additionally, we use interactive image segmentation models such as BRS [8], f-BRS [21], and RITM [22] as baselines." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "Implementation details. We utilize ResNet [7]-18 and DeepLab-V3 [2] as the backbone for the detector and corrector, respectively. We build on the pretrained model of Kim et al. [10], which is designed to estimate keypoints based on user feedback. All of the models are trained independently. We apply the misvertex error type to train the detector and all three error types for the corrector. KeyBot-i1, KeyBot-i2, and KeyBot-i3 refer to KeyBot with one, two, and three" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 141, + 145, + 471, + 277 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 2: Comparison of required number of user clicks to reach target performance on the AASCE, BUU-AP, and BUU-LA datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 145, + 471, + 277 + ], + "lines": [ + { + "bbox": [ + 141, + 145, + 471, + 277 + ], + "spans": [ + { + "bbox": [ + 141, + 145, + 471, + 277 + ], + "type": "table", + "html": "
MethodAASCEBUU-APBUU-LA
NoC10@20NoC10@30NoC10@40NoC10@50NoC5@15NoC5@20NoC5@25NoC5@30NoC5@15NoC5@20NoC5@25NoC5@30
BRS [8]2.531.971.511.232.801.510.920.613.892.261.340.86
f-BRS [21]7.165.304.213.482.231.180.680.481.240.610.440.40
RITM [22]3.202.371.871.381.050.580.420.300.570.240.170.14
Kim et al. [10]2.101.541.231.030.940.540.370.320.380.150.140.12
Click-Pose [26]4.153.563.202.971.200.890.760.711.450.700.560.45
KeyBot-i11.841.260.920.530.870.510.330.280.340.100.060.05
KeyBot-i21.781.210.910.610.870.480.300.240.350.110.060.05
KeyBot-i31.741.320.940.680.870.490.320.240.340.100.060.05
", + "image_path": "5f48cc8ce8f498422611c98de846b0e31b969d6bb1c5c12f3d05eca6beff6188.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 482, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 357 + ], + "type": "text", + "content": "iterations, respectively. Unless explicitly stated, the experimental results are conducted without user interaction. The training and inference are conducted on a single NVIDIA GeForce RTX 3090 GPU with 24GB memory. More details on experimental settings, including training details, are provided in Appendix C." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 371, + 331, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 371, + 331, + 384 + ], + "spans": [ + { + "bbox": [ + 131, + 371, + 331, + 384 + ], + "type": "text", + "content": "4.2 Experimental results and analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 390, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 545 + ], + "type": "text", + "content": "Comparison with state-of-the-art methods. In this section, we present a comparative analysis of interactive keypoint estimation performance between KeyBot and established baseline models by increasing the maximum KeyBot iteration count " + }, + { + "bbox": [ + 130, + 390, + 482, + 545 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 390, + 482, + 545 + ], + "type": "text", + "content": " from one to three. As shown in Table 1, KeyBot achieves state-of-the-art performance in mean radial error (MRE) across the AASCE, BUU-AP, and BUU-LA datasets. KeyBot consistently outperforms baseline models both in initial predictions and during subsequent user interactions. Remarkably, KeyBot with a single user click (UC1) shows superior performance compared to BRS, f-BRS, RITM, and Click-Pose with two user clicks (UC2) on the AASCE dataset. Furthermore, the performance of KeyBot improves with additional iterations, demonstrating its ability to refine predictions through multiple correction cycles. These results highlight the effectiveness of KeyBot in enhancing keypoint estimation accuracy and reducing user intervention." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 483, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 483, + 618 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 483, + 618 + ], + "type": "text", + "content": "In terms of the number of clicks (NoC), as shown in Table 2, KeyBot exhibits superior performance compared to the baseline models across all three datasets. For instance, on the AASCE dataset, our proposed model nearly halves the NoC required to achieve a target MRE of 50 (" + }, + { + "bbox": [ + 130, + 545, + 483, + 618 + ], + "type": "inline_equation", + "content": "\\mathrm{NoC}_{10} @ 50" + }, + { + "bbox": [ + 130, + 545, + 483, + 618 + ], + "type": "text", + "content": ") compared to that of the baseline models. This substantial reduction in NoC underscores KeyBot's proficiency in interactive frameworks, significantly lowering annotation costs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "Analysis on the detector. We investigate the effectiveness of the detector in identifying erroneous keypoints within images. In Fig. 6, we provide (a) the number of images where the detector identifies no errors versus those where errors are detected, as well as (b) the distribution of their actual mean radial error." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 342, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 342, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bones Can't Be Triangles" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 155, + 114, + 326, + 223 + ], + "blocks": [ + { + "bbox": [ + 155, + 114, + 326, + 223 + ], + "lines": [ + { + "bbox": [ + 155, + 114, + 326, + 223 + ], + "spans": [ + { + "bbox": [ + 155, + 114, + 326, + 223 + ], + "type": "image", + "image_path": "68e2aeac048d4f26decab5ca04079d817be705d917ef8f32c2bb855b0cdeab9e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 152, + 232, + 323, + 255 + ], + "lines": [ + { + "bbox": [ + 152, + 232, + 323, + 255 + ], + "spans": [ + { + "bbox": [ + 152, + 232, + 323, + 255 + ], + "type": "text", + "content": "Fig.6: Effectiveness of the detector in identifying errors on the AASCE dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 334, + 181, + 455, + 253 + ], + "blocks": [ + { + "bbox": [ + 330, + 118, + 461, + 171 + ], + "lines": [ + { + "bbox": [ + 330, + 118, + 461, + 171 + ], + "spans": [ + { + "bbox": [ + 330, + 118, + 461, + 171 + ], + "type": "text", + "content": "Table 3: Keypoint prediction errors on the AASCE dataset, comparing MRE for examples where the detector identifies errors versus where it does not." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 181, + 455, + 253 + ], + "lines": [ + { + "bbox": [ + 334, + 181, + 455, + 253 + ], + "spans": [ + { + "bbox": [ + 334, + 181, + 455, + 253 + ], + "type": "table", + "html": "
MethodNo error detectedError detected
Initial prediction28.9964.88
KeyBot-i1-53.12
KeyBot-i2-50.49
KeyBot-i3-49.18
", + "image_path": "ef18d7f108047369bb8757460a3d680b56ac32e75ee3613df7d466cfc6d0fab3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 279, + 479, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 479, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 479, + 339 + ], + "type": "text", + "content": "Out of the total images analyzed, 43 images have no errors detected, while 73 images have errors identified. The images classified as \"No error detected\" exhibit significantly lower errors compared to those classified as \"Error detected.\" These findings indicate the precision and effectiveness of the detector in distinguishing between erroneous predictions and accurate ones." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 340, + 480, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 480, + 412 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 480, + 412 + ], + "type": "text", + "content": "We further examine error reduction through successive KeyBot iterations for images where errors are initially detected. As demonstrated in Table 3, KeyBot effectively enhances the accuracy of these error-detected images, significantly reducing errors in subsequent iterations. Notably, after three iterations of KeyBot, the MRE significantly decreases from 64.88 to 49.18. This reduction highlights KeyBot's ability to iteratively identify inaccuracies and improve predictions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 412, + 480, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 480, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 480, + 484 + ], + "type": "text", + "content": "Ablation study. This section validates each component of KeyBot. First, we study the role of the detector under two scenarios: (i) employing corrector outputs directly as the final results and (ii) utilizing all corrector outputs as pseudocorrections for the interaction model. As detailed in Table 4, both scenarios exhibit significant performance degradation compared to KeyBot, underscoring the importance of using the detector to selectively apply corrector outputs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 485, + 480, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 480, + 581 + ], + "type": "text", + "content": "Additionally, utilizing only the interactive keypoint estimation network, without the support of KeyBot, results in a significant performance degradation compared to the refined predictions with KeyBot. When ablating the error types simulated during the training of the corrector, our findings show that training the corrector to recognize and rectify three distinct types of errors significantly improves its prediction accuracy compared to when it is trained on only one error type. This finding highlights the necessity of incorporating all three typical errors into the training process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "Qualitative results. In Fig. 7, we analyze the qualitative results of KeyBot with three iterations. The results demonstrate that KeyBot effectively identifies and reduces all three error types: vertex misidentification (Misvertex), bone misidentification (Misbone), and left-right inversion (LR-inversion). For these errors, KeyBot provides corrective feedback to the interactive model, enabling necessary adjustments. For instance, in Fig. 7(a), the initial prediction skips a bone, resulting in a Misbone error (red), and also exhibits a Misvertex error" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 135, + 444, + 258 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 478, + 125 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 478, + 125 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 478, + 125 + ], + "type": "text", + "content": "Table 4: Ablation study of our keypoint refinement method on the AASCE dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 135, + 444, + 258 + ], + "lines": [ + { + "bbox": [ + 167, + 135, + 444, + 258 + ], + "spans": [ + { + "bbox": [ + 167, + 135, + 444, + 258 + ], + "type": "table", + "html": "
Interaction modelKeyBotMean radial error
DetectorCorrectorError type-UC1UC2UC3
✓✓✓all44.1825.9320.6618.03
XX✓all50.9939.1732.9830.36
✓X✓all57.6145.9439.1136.01
✓XX-51.5830.6025.5721.07
✓✓✓misvertex52.6731.1425.9221.65
✓✓✓misbone49.0729.8523.4320.75
✓✓✓lr-inversion52.3830.1625.1320.90
", + "image_path": "112aa05676ba1c1718c6bcbaed33d21ac397a67c22272a57e8607e7fa6a6c272.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 135, + 279, + 304, + 390 + ], + "blocks": [ + { + "bbox": [ + 135, + 279, + 304, + 390 + ], + "lines": [ + { + "bbox": [ + 135, + 279, + 304, + 390 + ], + "spans": [ + { + "bbox": [ + 135, + 279, + 304, + 390 + ], + "type": "image", + "image_path": "f8a01f1b7f685cbdefd8c45e6fc67b2d0c9e11cbe1977c9c63b91dde52df3074.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 390, + 288, + 397 + ], + "lines": [ + { + "bbox": [ + 153, + 390, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 153, + 390, + 288, + 397 + ], + "type": "text", + "content": "(a) misvertex+misbone (MRE: " + }, + { + "bbox": [ + 153, + 390, + 288, + 397 + ], + "type": "inline_equation", + "content": "71.7\\rightarrow 11.9" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 279, + 479, + 390 + ], + "blocks": [ + { + "bbox": [ + 309, + 279, + 479, + 390 + ], + "lines": [ + { + "bbox": [ + 309, + 279, + 479, + 390 + ], + "spans": [ + { + "bbox": [ + 309, + 279, + 479, + 390 + ], + "type": "image", + "image_path": "f48d389d76dda02e382386fb0e9c14c0299ed4d9349fb1bdb3da9820b02fc4f5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 390, + 463, + 398 + ], + "lines": [ + { + "bbox": [ + 327, + 390, + 463, + 398 + ], + "spans": [ + { + "bbox": [ + 327, + 390, + 463, + 398 + ], + "type": "text", + "content": "(b) misvertex+misbone (MRE: 91.9 → 20.2)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 135, + 398, + 304, + 509 + ], + "blocks": [ + { + "bbox": [ + 135, + 398, + 304, + 509 + ], + "lines": [ + { + "bbox": [ + 135, + 398, + 304, + 509 + ], + "spans": [ + { + "bbox": [ + 135, + 398, + 304, + 509 + ], + "type": "image", + "image_path": "24533e71014362e7a8f3f8e268de202f96af687e578696373c94e40e411c1a7d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 509, + 290, + 517 + ], + "lines": [ + { + "bbox": [ + 151, + 509, + 290, + 517 + ], + "spans": [ + { + "bbox": [ + 151, + 509, + 290, + 517 + ], + "type": "text", + "content": "(c) misvertex+misbone (MRE: 193.3 → 15.3)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 153, + 520, + 294, + 530 + ], + "lines": [ + { + "bbox": [ + 153, + 520, + 294, + 530 + ], + "spans": [ + { + "bbox": [ + 153, + 520, + 294, + 530 + ], + "type": "text", + "content": "Image Initial prediction KeyBot" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 398, + 479, + 510 + ], + "blocks": [ + { + "bbox": [ + 309, + 398, + 479, + 510 + ], + "lines": [ + { + "bbox": [ + 309, + 398, + 479, + 510 + ], + "spans": [ + { + "bbox": [ + 309, + 398, + 479, + 510 + ], + "type": "image", + "image_path": "ca6b4f126fbb2f2d2c6876befd69939cd9175e3c631c7f82631f59aec355b680.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 509, + 468, + 517 + ], + "lines": [ + { + "bbox": [ + 320, + 509, + 468, + 517 + ], + "spans": [ + { + "bbox": [ + 320, + 509, + 468, + 517 + ], + "type": "text", + "content": "(d) misvertex+Ir-inversion (MRE: 443.5 → 40.3)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 520, + 468, + 530 + ], + "lines": [ + { + "bbox": [ + 328, + 520, + 468, + 530 + ], + "spans": [ + { + "bbox": [ + 328, + 520, + 468, + 530 + ], + "type": "text", + "content": "Image Initial prediction KeyBot" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 538, + 482, + 605 + ], + "lines": [ + { + "bbox": [ + 130, + 538, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 538, + 482, + 605 + ], + "type": "text", + "content": "Fig. 7: Prediction results on the (a) AASCE, (b) BUU-LA, and (c,d) BUU-AP (bottom) datasets. KeyBot effectively reduces all three types of errors, even in challenging cases where error types are mixed. The keypoints classified as inaccurate by our detector are indicated with red dots, while their corresponding keypoints in the revised results are visualized as blue dots. A diagonal line connecting the upper-right vertex to the lower-left vertex is indicated in each vertebra. More results are in Appendix F." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "(purple). KeyBot accurately refines these errors, significantly reducing the MRE from 71.7 to 11.9. The results highlight KeyBot's capability to significantly enhance keypoint estimation accuracy without user clicks." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 342, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 342, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bones Can't Be Triangles" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 224 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 224 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 224 + ], + "type": "image", + "image_path": "7d487f580d16bbb59390c9e49d8798c928bcd715e8ca7486605c8e0152f011c3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 232, + 480, + 277 + ], + "lines": [ + { + "bbox": [ + 130, + 232, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 480, + 277 + ], + "type": "text", + "content": "Fig. 8: Collaborative keypoint refinement process among the interaction model, KeyBot, and the user. Faced with substantial errors in initial predictions, KeyBot effectively removes all major errors, allowing users to easily identify and correct remaining inaccuracies with minimal effort. Here, refinement is achieved with just two user clicks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 300, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 445 + ], + "type": "text", + "content": "Additionally, KeyBot excels in interpretability. It includes a separate detector that enables users to easily comprehend the keypoints identified as errors and the corrections made, leading to efficient evaluation of the results. In the given examples, identified errors are indicated by red dots, and the corrections are shown as blue dots, illustrating how the keypoints are adjusted. This transparency contributes to improving the efficiency of the overall annotation process. Collaborative refinement process. In Fig. 8, we showcase the qualitative results of the collaborative refinement process involving both KeyBot and the user. Our approach establishes a feedback loop among the interaction model, KeyBot, and the user, combining expert medical knowledge and a targeted focus on specific error types. This significantly enhances the overall accuracy and efficiency of the annotation process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 464, + 220, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 220, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 220, + 475 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 488, + 480, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 488, + 480, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 488, + 480, + 561 + ], + "type": "text", + "content": "We introduce a novel approach, KeyBot, specifically designed to identify and correct significant errors in vertebrae keypoint estimation. Our proposed error simulation training procedure enables KeyBot to accurately and independently address specific error types, thereby significantly reducing the need for user intervention in the annotation process. Comprehensive evaluations on public real-world datasets confirm the superiority of KeyBot over existing methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 580, + 240, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 580, + 240, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 240, + 594 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "This work was supported by Institute for Information & communications Technology Promotion (IITP) grant funded by the Korea government (MSIT) (No.RS2019-II190075 Artificial Intelligence Graduate School Program (KAIST)) and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. NRF-2022R1A2B5B02001913)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 139, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 139, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 480, + 172 + ], + "type": "text", + "content": "1. Albiero, V., Chen, X., Yin, X., Pang, G., Hassner, T.: img2pose: Face alignment and detection via 6dof, face pose estimation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7617-7627 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 194 + ], + "type": "text", + "content": "2. Chen, L.C., Papandreou, G., Schroff, F., Adam, H.: Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 194, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 480, + 216 + ], + "type": "text", + "content": "3. Colaco, S.J., Han, D.S.: Deep learning-based facial landmarks localization using compound scaling. IEEE Access 10, 7653-7663 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 217, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 480, + 237 + ], + "type": "text", + "content": "4. Daniele, A., van Krieken, E., Serafini, L., van Harmelen, F.: Refining neural network predictions using background knowledge. Machine Learning pp. 1-39 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 238, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 238, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 238, + 480, + 281 + ], + "type": "text", + "content": "5. Guo, W., Du, Y., Shen, X., Lepetit, V., Alameda-Pineda, X., Moreno-Noguer, F.: Back to mlp: A simple baseline for human motion prediction. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4809-4819 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 282, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 282, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 282, + 480, + 315 + ], + "type": "text", + "content": "6. Guo, Y., Li, Y., Zhou, X., He, W.: A keypoint transformer to discover spine structure for cobb angle estimation. In: 2021 IEEE International Conference on Multimedia and Expo (ICME). pp. 1-6. IEEE (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 316, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 480, + 336 + ], + "type": "text", + "content": "7. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 337, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 480, + 357 + ], + "type": "text", + "content": "8. Jang, W.D., Kim, C.S.: Interactive image segmentation via backpropagating refinement scheme. CVPR pp. 5297-5306 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 358, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 358, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 138, + 358, + 480, + 402 + ], + "type": "text", + "content": "9. Jin, S., Xu, L., Xu, J., Wang, C., Liu, W., Qian, C., Ouyang, W., Luo, P.: Whole-body human pose estimation in the wild. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16. pp. 196-214. Springer (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 403, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 445 + ], + "type": "text", + "content": "10. Kim, J., Kim, T., Kim, T., Choo, J., Kim, D.W., Ahn, B., Song, I.S., Kim, Y.J.: Morphology-aware interactive keypoint estimation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 675-685. Springer (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "text", + "content": "1. Klinwichit, P., Yookwan, W., Limchareon, S., Chinnasarn, K., Jang, J.S., Onuean, A.: Buu-lspine: A thai open lumbar spine dataset for spondylolisthesis detection. Applied Sciences 13(15), 8646 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 511 + ], + "type": "text", + "content": "2. Liu, Q., Zheng, M., Planche, B., Karanam, S., Chen, T., Niethammer, M., Wu, Z.: Pseudoclick: Interactive image segmentation with click imitation. In: European Conference on Computer Vision. pp. 728-745. Springer (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 512, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 555 + ], + "type": "text", + "content": "3. Liu, T., Wang, Y., Yang, Y., Sun, M., Fan, W., Bunger, C., Wu, C.: A multi-scale keypoint estimation network with self-supervision for spinal curvature assessment of idiopathic scoliosis from the imperfect dataset. Artificial Intelligence in Medicine 125, 102235 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 555, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 577 + ], + "type": "text", + "content": "4. Moon, G., Chang, J.Y., Lee, K.M.: Posefix: Model-agnostic general human pose refinement network. CVPR pp. 7773-7781 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 578, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 632 + ], + "type": "text", + "content": "5. Pisov, M., Kondratenko, V., Zakharov, A., Petraikin, A., Gombolevskiy, V., Morozov, S., Belyaev, M.: Keypoints localization for joint vertebra detection and fracture severity quantification. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part VI 23. pp. 723-732. Springer (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "6. Qu, C., Zhang, T., Qiao, H., Tang, Y., Yuille, A.L., Zhou, Z., et al.: Abdomenatlas-8k: Annotating 8,000 ct volumes for multi-organ segmentation in three weeks. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 342, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 342, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bones Can't Be Triangles" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 468 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "17. Rahmaniar, W., Suzuki, K., Lin, T.L.: Auto-ca: automated cobb angle measurement based on vertebrae detection for assessment of spinal curvature deformity. IEEE Transactions on Biomedical Engineering (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "text", + "content": "18. Ruggero Ronchi, M., Perona, P.: Benchmarking and error diagnosis in multi-instance pose estimation. In: Proceedings of the IEEE international conference on computer vision. pp. 369-378 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "type": "text", + "content": "19. Shi, D., Wei, X., Li, L., Ren, Y., Tan, W.: End-to-end multi-person pose estimation with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11069-11078 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "text", + "content": "20. Singh, S., Sandhu, N., Puri, T., Gulati, R., Kashyap, R.: A study of correlation of various growth indicators with chronological age. International Journal of Clinical Pediatric Dentistry 8(3), 190 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "21. Sofiuk, K., Petrov, I., Barinova, O., Konushin, A.: f-brs: Rethinking backpropagating refinement for interactive segmentation. CVPR pp. 8623-8632 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "text", + "content": "22. Sofiuk, K., Petrov, I.A., Konushin, A.: Reviving iterative training with mask guidance for interactive segmentation. In: 2022 IEEE International Conference on Image Processing (ICIP). pp. 3141-3145. IEEE (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "23. Türkoz, C., Kaygirisiz, E., Ulusoy, C., Ates, C.: A practical formula for determining growth. Diagnostic and Interventional Radiology 23(3), 194 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 358 + ], + "type": "text", + "content": "24. Wu, H., Bailey, C., Rasoulinejad, P., Li, S.: Automatic landmark estimation for adolescent idiopathic scoliosis assessment using boostnet. In: MICCAI. pp. 127-135. Springer (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "type": "text", + "content": "25. Xu, Y., Xu, D., Hong, X., Ouyang, W., Ji, R., Xu, M., Zhao, G.: Structured modeling of joint deep feature and prediction refinement for salient object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3789-3798 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 402, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 480, + 435 + ], + "type": "text", + "content": "26. Yang, J., Zeng, A., Li, F., Liu, S., Zhang, R., Zhang, L.: Neural interactive keypoint detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15122-15132 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "27. Zhang, H., Chung, A.C.: A dual coordinate system vertebral landmark detection network with sparse-to-dense vertebral line interpolation. Bioengineering 11(1), 101 (2024)" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Boost Your NeRF_ A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering/cdbb2d54-c057-481d-a45f-f2c36fd59358_content_list.json b/2024/Boost Your NeRF_ A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering/cdbb2d54-c057-481d-a45f-f2c36fd59358_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e08ba152368d76d1fa6c19a617e52597fa7f3789 --- /dev/null +++ b/2024/Boost Your NeRF_ A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering/cdbb2d54-c057-481d-a45f-f2c36fd59358_content_list.json @@ -0,0 +1,1718 @@ +[ + { + "type": "text", + "text": "Boost Your NeRF: A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering", + "text_level": 1, + "bbox": [ + 222, + 140, + 782, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Francesco Di Sario $^{1}$ [0009-0005-6969-1246], Riccardo Renzulli $^{1}$ [0000-0003-0532-5966], Enzo Tartaglione $^{2}$ [0000-0003-4274-8298], and Marco Grangetto $^{1}$ [0000-0002-2709-7864]", + "bbox": [ + 344, + 233, + 656, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Turin, Italy \n2 LTCI, Télécom Paris, Institut Polytechnique de Paris, France \nfrancesco.disario@unito.it", + "bbox": [ + 289, + 306, + 710, + 349 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Since the introduction of NeRFs, considerable attention has been focused on improving their training and inference times, leading to the development of Fast-NeRFs models. Despite demonstrating impressive rendering speed and quality, the rapid convergence of such models poses challenges for further improving reconstruction quality. Common strategies to improve rendering quality involves augmenting model parameters or increasing the number of sampled points. However, these computationally intensive approaches encounter limitations in achieving significant quality enhancements. This study introduces a model-agnostic framework inspired by Sparsely-Gated Mixture of Experts to enhance rendering quality without escalating computational complexity. Our approach enables specialization in rendering different scene components by employing a mixture of experts with varying resolutions. We present a novel gate formulation designed to maximize expert capabilities and propose a resolution-based routing technique to effectively induce sparsity and decompose scenes. Our work significantly improves reconstruction quality while maintaining competitive performance.", + "bbox": [ + 261, + 383, + 743, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 643, + 375, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRFs) [24] have recently shown impressive results in synthesizing photo-realistic 3D scenes from a set of 2D images. However, NeRFs suffer from limited scene diversity, long training time, and sensitivity to training data [21]. Since the introduction of NeRFs, significant attention has been directed toward improving their training and inference times, resulting in the development of Fast-NeRFs. By using auxiliary data structures such as voxel grids to store scene geometry and skip empty spaces, the training and inference process can be accelerated by several orders of magnitude. As a result, the neural component of these models, which is usually used to transform learned features into view-dependent color representations, becomes much smaller and can sometimes be replaced entirely by spherical harmonics [10]. Despite Fast", + "bbox": [ + 212, + 672, + 787, + 843 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/af8de6ff0478d385dfb6d0336d7364361ddfeed921b911b58cb796a1bdd2990d.jpg", + "image_caption": [ + "Fig. 1: Different strategies for improving reconstruction quality with Fast-NeRFs (DVGO [29]). Increasing the resolution of data structures like voxel grids can improve render quality, but only up to a certain point, after which quality declines (left). The MLP component's impact on rendering quality is analyzed by varying its depth and width while keeping other variables constant (center). The effect of the number of sampled points along each ray is examined by decreasing the step size (right). In gray, our method's performance, which significantly improves PSNR at a low computational cost." + ], + "image_footnote": [], + "bbox": [ + 228, + 146, + 403, + 242 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5065657261c1d40643a7eb8e77731e04eac0762360d4e7e017cc81711548c4b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 146, + 588, + 242 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e28add08a4a50158e527829cf420b945c72423bc1371faf9d592672e183b246d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 146, + 776, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NeRFs being an improvement from the original NeRF framework, achieving high-quality and efficient reconstructed geometry is still an open problem.", + "bbox": [ + 212, + 396, + 782, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Typical \"naive\" approaches to enhance the reconstruction quality of these models include:", + "bbox": [ + 212, + 426, + 782, + 455 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Increasing the parameters and resolution of the used data structures (e.g., voxel grid, hash grid, etc.).", + "2. Increasing the number of sampled points per ray.", + "3. Increasing the number of parameters in the neural network or the order of spherical harmonics." + ], + "bbox": [ + 220, + 467, + 784, + 541 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, as evident from Figure 1, the increase in reconstruction quality results in a significant increase in computational costs. In the first approach, increasing the resolution can lead to a significant improvement in the reconstruction quality, but a plateau is reached beyond which overfitting occurs, and reconstruction quality degrades. This comes at the expense of a considerable increase in spatial complexity (especially with dense voxel grids) and training times. Increasing the number of sampled points per ray can marginally improve reconstruction quality further but at a significant increase in computational complexity. In fact, the higher the number of sampled rays, the higher forward passes through a neural network are needed. Augmenting the number of parameters in the neural network is another solution. However, as the neural network grows in depth and width, it tends towards a fully implicit model, deviating from the principles of fast models (limiting the neural part as much as possible - or even removing it altogether).", + "bbox": [ + 212, + 551, + 787, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose a technique capable of significantly enhancing the reconstruction quality of such models while maintaining competitive training and rendering times. Inspired by the Sparse Mixture of Experts (MoE) paradigm [28], we have developed a model-agnostic framework capable of improving the reconstruction quality of various state-of-the-art models. Intuitively, our mixture of experts consists of different-capacity (resolution) models. During training, each model", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "F. Di Sario et al.", + "bbox": [ + 271, + 114, + 387, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "specializes in rendering the most suitable parts of the scene, i.e. low-resolution models render low-frequency parts of the scene, while high-resolution models render high-frequency parts. Our formulation of the gate allows our method to be model-agnostic, enabling, on the one hand, the insertion of the gate function at the early stages of the MoE mechanism and, on the other hand, the multiplication of the gate's output with the experts' output as late as possible. This maximizes the capabilities of the mixture of experts and allows for working directly with the output of each expert, which is considered a black box. This is why our method is inherently model-agnostic. In summary, our contributions are as follows:", + "bbox": [ + 212, + 146, + 782, + 282 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose the first model-agnostic framework based on Sparse MoE of models at different resolutions, which significantly improves the rendering quality of such models while maintaining competitive training and inference times (Sec. 3).", + "2. We provide a novel gate formulation inspired by Fast-NeRF models, which maximizes the capability of the mixture of experts (Sec. 3.3).", + "3. We introduce a new routing technique based on resolution, favoring the assignment of tokens to low-resolution models and discouraging the assignment of tokens to high-resolution models, inducing increasing sparsity in high-resolution models and decomposing the scene based on frequency (Sec. 3.5).", + "4. We conduct extensive experiments to test our method, including different NeRFs architectures and datasets, showing higher rendering quality and efficiency (Sec. 4)." + ], + "bbox": [ + 222, + 287, + 784, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related works", + "text_level": 1, + "bbox": [ + 215, + 503, + 390, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Neural Radiance Field", + "text_level": 1, + "bbox": [ + 215, + 531, + 446, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "NeRF [24] (Neural Radiance Field) has emerged as a prominent method for synthesizing novel views, showing significant progress. This approach requires a moderate number of input images along with their known camera poses. Unlike traditional methods that rely on explicit and discretized volumetric representations such as voxel grids and multiplane images, NeRF employs a coordinate-based multilayer perceptron (MLP) to create an implicit and continuous volumetric representation. NeRFs can represent a 3D scene as a MLP $F_{\\theta}$ , with $\\theta$ being the set of its trainable parameters, such that $F_{\\theta}:(\\mathbf{x},\\mathbf{d})\\to (\\mathbf{c},\\sigma)$ which maps $(\\mathbf{x},\\mathbf{d})$ , a 3D position $\\mathbf{x}$ and a viewing direction $\\mathbf{d}$ , to a view-dependent color emission $\\mathbf{c}$ and density value $\\sigma$ . To render the color of a pixel $\\hat{C}(\\mathbf{r})$ , a ray $\\mathbf{r}$ traverses the center of the camera through the pixel of the image plane from an origin point $\\mathbf{o}$ to a ray having position $\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}$ . Then, $N$ points are sampled on $\\mathbf{r}$ , and the MLP is queried for each point, obtaining a density and a color value. Finally, these results are accumulated into a single color with the volume rendering equation [22]:", + "bbox": [ + 212, + 553, + 784, + 781 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {C} (\\mathbf {r}) = \\sum_ {i = 1} ^ {K} T _ {i} \\alpha_ {i} c _ {i}, \\quad T _ {i} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\quad \\alpha_ {i} = 1 - \\exp \\left(\\sigma_ {i} \\delta_ {i}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 272, + 801, + 784, + 842 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Boost Your NeRF", + "bbox": [ + 609, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{i}$ is the probability of termination at point $i$ and $T_{i}$ is the accumulated transmittance from the near plane to point $i$ . NeRFs are trained by minimizing a photometric loss, which is an L2 loss between the rendered and the ground truth pixels. In more detail, given a batch $B$ of randomly sampled rays, the loss is defined as", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {n e r f}} = \\frac {1}{| B |} \\sum_ {r \\in B} \\left\\| \\hat {C} (\\mathbf {r}) - C (\\mathbf {r}) \\right\\| _ {2} ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 233, + 785, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{C}(\\mathbf{r})$ is the predicted color and $C(\\mathbf{r})$ the ground truth color for the ray $\\mathbf{r}$ .", + "bbox": [ + 212, + 277, + 785, + 309 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Fast Neural Radiance Fields", + "text_level": 1, + "bbox": [ + 214, + 330, + 495, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the introduction of NeRF, significant effort has been directed towards the development of faster models. Several studies focus on speeding up rendering times, for example by working on ray sampling efficiency [3, 6, 12, 37], by integrating explicit volumetric representations [5, 11, 11, 15, 31-33] or by utilizing thousands of tiny MLPs [26]. However, all these methods still require large training times. One noteworthy development is represented by the introduction of explicit volumetric representations in the training pipeline, directly optimizing such representations [4, 9, 10, 13, 16, 25, 29]. These models enable fast training and inference, with render quality only slightly inferior to full-implicit models [1, 2]. We refer to these as Fast-NeRFs. Plenoxels [10] is the first important work in this context, as it represents a scene as a sparse 3D grid where each voxel stores spherical harmonic coefficients and density. Spherical harmonics serve as an orthogonal basis for functions defined over the sphere and thus can be used for computing view-dependent color emission, without the need for a multi-layer perceptron. Additionally, Plenoxels demonstrated the advantages of linearly interpolating voxels, facilitating the learning of a continuous plenoptic function throughout the volume, akin to NeRF, albeit with discrete data. Another notable work is DVGO [29]. Each scene is there represented as two dense voxel grids (one for density and one for feature colors) alongside a MLP, for learning view-dependent color. Similar to Plenoxels, the value of each voxel is linearly interpolated with the 8 nearest voxels, but after applying the activation functions. DVGO also incorporates a preliminary coarse geometry stage to learn the scene's general structure, facilitating adjustments to the bounding box and enabling a more intelligent ray sampling strategy. Despite utilizing lower-resolution models, DVGO achieves high reconstruction quality. The major drawback of these models is their substantial memory storage requirements. TensoRF [4] addresses this challenge by replacing the dense voxel grid with a planes and vectors decomposition, significantly reducing storage demands while maintaining comparable performance and rendering quality. Alternatively, Instant-NGP [25] proposed a multi-resolution voxel grid encoded via a hash function. They define $L$ hash grids of increased resolutions: each entry of each hash grid has $2^T$ parameters and $F$ features. This enables even faster training times and real-time rendering", + "bbox": [ + 212, + 356, + 787, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "F. Di Sario et al.", + "bbox": [ + 271, + 114, + 385, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "performances while maintaining a compact model; moreover, it demonstrates the efficacy of a multi-resolution approach in enhancing render quality. Inspired by this methodology, K-Planes [9] introduces a multi-resolution planar factorization of 3D space, offering reconstruction quality and model compactness similar to the previous methods, but with slightly larger training times. A similar multi-resolution planar factorization has also been proposed in Tri-MipRF [13] for mitigating the aliasing in distant or low-resolution views and blurriness in close-up shots. Given their properties, we decided to evaluate our paradigm using three different models: a 3D dense voxel grid-based model (DVGO), a decomposed grid-based model (TensoRF), and a multi-resolution hash grid-based model (Instant-NGP).", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.3 Mixture of Experts and Sparse MoE", + "text_level": 1, + "bbox": [ + 214, + 333, + 565, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Mixture of Experts (MoE) paradigm [14] has gained prominence in various machine learning applications. MoE consists of multiple expert networks, each specializing in different regions of the input space, with a meta-network determining the contribution of each expert to the final prediction. Building upon the MoE framework, Sparse Mixture of Experts [19, 27, 28, 36] constitutes a scalable and efficient variant. At the core of all Sparse MoE algorithms lies an assignment problem between tokens and experts. One approach to tackle this is by approximating the solution with a gating function, which learns to assign input tokens to the most suitable experts. It typically comprises a linear layer, a softmax activation, and a Top- $K$ (where $1 \\leq K \\leq 2$ ) operation, aimed at routing the input token to only a subset of the experts. To balance the assignment across all the experts, auxiliary loss functions penalizing unevenly distributed routing are often employed. This sparsity-inducing technique significantly reduces computational overhead while preserving the expressive power of the MoE architecture.", + "bbox": [ + 212, + 357, + 787, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.4 MoE and NeRF", + "text_level": 1, + "bbox": [ + 214, + 589, + 395, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Combining the sparse mixture of experts' paradigms with neural radiance fields presents a non-trivial challenge. In recent research, Switch-NeRF [35] has been introduced as a novel end-to-end large-scale NeRF with learning-based scene decomposition. Inspired by Fedus et al. [8], they propose a full-implicit model with an MLP-based gate comprising 4 layers with 128 neurons and a Top-1 function. However, their architecture is ad-hoc, and suffers from extensive training times. This poses a challenge for us, as we aim to consider each expert as a black box, taking a point in space and a direction as input and outputting radiance and density values. Moreover, as the mixture comprises experts at different resolutions, we also aim to prioritize routing input tokens toward lower-resolution models and minimize the usage of high-resolution models.", + "bbox": [ + 212, + 613, + 787, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our work intends to overcome these limitations. We design a novel gate formulation inspired by Fast-NeRFs, that guarantees fast convergence and better performances. We position the gate at the beginning of our MoE pipeline and postpone the multiplication of the gate output with expert predictions as much as", + "bbox": [ + 212, + 780, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Boost Your NeRF", + "bbox": [ + 609, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/235ba8e61c477ad6888196f5ebadba5868808d3226d41c575b74082eb7587470.jpg", + "image_caption": [ + "Fig. 2: A density field is used to compute density values for sampled points along a ray. A filtering step discards low-density points, routed through gating network $G$ for expert assignment. Top- $K$ experts compute radiance and density, aggregating and weighting these values by the corresponding gating probability to get the final values $\\mathbf{c}_f$ and $\\sigma_f$ . Volume rendering equation yields pixel colors, and joint optimization refines our resolution-weighted auxiliary loss, allowing for high-quality and efficient rendering." + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 784, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "possible. This allows our framework to be entirely model-agnostic. Furthermore, our mixture of experts maintains competitive training times, enabling, in the worst case, training models on complex scenes in approximately 1 hour while achieving state-of-the-art accuracy.", + "bbox": [ + 212, + 421, + 787, + 482 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 503, + 330, + 520 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we introduce our model-agnostic assembling approach. We present in Sec. 3.1 an overview of our method. We also describe the ray sampling and filtering mechanism in Sec. 3.2. Then, in Sec. 3.3, we discuss the effective construction of a mixture of experts with a gating mechanism in the context of NeRF. Finally, in Sec. 3.5, we present a novel formulation for decomposing the scene into resolution-based parts with an end-to-end training procedure.", + "bbox": [ + 212, + 536, + 787, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 648, + 341, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 2 shows a high-level scheme of our method. After independently training a set of $M$ NeRF models $\\{E_i\\}_{i=1}^M$ at different resolutions, we sample $N$ points along a ray $\\mathbf{r}$ . For each point $\\mathbf{x}_p$ , a density value $\\sigma$ is calculated, and based on this, a filtering step is performed based on the density volume $\\mathbf{V}_D$ , discarding points in regions with a density below a certain threshold $T$ . Initializing the density volume $\\mathbf{V}_D$ with the one learned from the lowest resolution model can be helpful (though it can also be learned). Subsequently, the filtered point $\\mathbf{x}_f$ is fed to the gate $G$ , so we denote with $G(\\mathbf{x}_f)$ the probabilities with which the point is assigned to the $M$ experts. Based on these values, each point is routed to the Top- $K$ experts. Each selected expert then computes the radiance $\\mathbf{c}_{i,f}$ and density values $\\sigma_{i,f}$ for the point, both of which are multiplied by their respective", + "bbox": [ + 212, + 672, + 787, + 843 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "F. Di Sario et al.", + "bbox": [ + 271, + 114, + 387, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "probabilities and summed together to get the final radiance $\\mathbf{c}_f$ and density $\\sigma_f$ . Subsequently, the volume rendering Equation 1 is used to aggregate all colors and compute the pixel color for the given ray. Finally, we compute the total loss and jointly optimize the gate and the experts.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the next sections, we will dive into the gating and sparse mixture of expert modules.", + "bbox": [ + 212, + 208, + 782, + 238 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Ray Sampling and Filtering", + "text_level": 1, + "bbox": [ + 214, + 268, + 491, + 285 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The first phase of our method involves learning a coarse and explicit density field starting from the density volume $\\mathbf{V}_D$ , that we can leverage for skipping empty spaces. Given a ray $\\mathbf{r}$ as explained in Sec. 2.1, we sample $N$ points along it:", + "bbox": [ + 212, + 303, + 782, + 349 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {p, r} = \\text {s a m p l i n g} (\\mathbf {r}) \\in \\mathbb {R} ^ {N \\times 3}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 363, + 785, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with $p \\in [0, \\dots, N - 1]$ . We suppress the index $r$ for abuse of notation. Next, we compute the density for each point $\\mathbf{x}_p \\in \\mathbb{R}^3$ and discard those with negligible density. The density value $\\sigma_{D,p}$ for each point is computed by linearly interpolating $v$ neighboring voxels:", + "bbox": [ + 212, + 395, + 782, + 455 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {D, p} = \\operatorname {a c t} (\\operatorname {i n t e r p o l a t e} \\left(\\mathbf {x} _ {p}, V _ {D}\\right)) \\in \\mathbb {R} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 479, + 784, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where act represents a density activation function (such as softmax). We denote $\\mathbf{x}_{\\mathbf{f}}$ with $\\mathbf{x}_{\\mathbf{f}} \\in R^3$ a remaining point after the filtering operation.", + "bbox": [ + 212, + 505, + 782, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Trainable Gating Model", + "text_level": 1, + "bbox": [ + 214, + 566, + 460, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our gating mechanism incorporates a hybrid architecture: an explicit feature grid $\\mathbf{V}_G$ and a shallow MLP. The gating mechanism can also be seen as a probability field. First, we compute per-point features:", + "bbox": [ + 212, + 599, + 782, + 646 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f e a t} = \\operatorname {i n t e r p o l a t e} \\left(\\mathbf {x} _ {f}, \\mathbf {V} _ {G}\\right) \\in \\mathbb {R} ^ {C} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 661, + 784, + 678 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $C$ represents the number of channels of $\\mathbf{V}_G$ . Subsequently, we transform each feature into a probability. We compute logits as", + "bbox": [ + 212, + 691, + 782, + 723 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {l o g i t s} = \\operatorname {M L P} (\\mathbf {f e a t}) \\in \\mathbb {R} ^ {M} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 405, + 744, + 784, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $M$ is the number of experts in the mixture of experts. Finally, we apply a per-row softmax. Our gating function can be summarized as:", + "bbox": [ + 212, + 772, + 782, + 801 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nG \\left(\\mathbf {x} _ {\\mathbf {f}}\\right) = \\operatorname {s o f t m a x} (\\operatorname {M L P} (\\text {i n t e r p o l a t e} \\left(\\mathbf {x} _ {f}, \\mathbf {V} _ {G}\\right))) \\in \\mathbb {R} ^ {M} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 824, + 784, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Boost Your NeRF", + "bbox": [ + 609, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 NeRFs Experts", + "text_level": 1, + "bbox": [ + 215, + 146, + 388, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As mentioned in Sec. 3.1, we employ a set of $M$ pretrained NeRFs models as experts. After feeding as input each point $\\mathbf{x}_f$ to the gate, we route them to the $k$ experts with the $k$ highest probabilities $G(\\mathbf{x}_f)$ . We define this set of indexes of the selected experts as", + "bbox": [ + 212, + 169, + 787, + 231 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {K} = \\underset {T o p - k} {\\operatorname {a r g m a x}} (G (\\mathbf {x} _ {f})). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 241, + 785, + 266 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Each expert is a Fast-NeRF model that receives the dispatched points and the direction of the ray they lie on as input and outputs a density and a radiance value. We treat each expert as a black box, ensuring our method is inherently model-agnostic. The radiance $\\mathbf{c}_f$ and the density $\\sigma_f$ for the point $\\mathbf{x}_f$ laying on a ray with direction $\\mathbf{d}$ can be expressed as the sum of each expert predictions weighted by their probability:", + "bbox": [ + 214, + 273, + 787, + 364 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {f}, \\sigma_ {f} = \\sum_ {i \\in \\mathcal {K}} E _ {i} \\left(\\mathbf {x} _ {\\mathbf {f}}, \\mathbf {d}\\right) \\cdot G \\left(\\mathbf {x} _ {\\mathbf {f}}\\right) _ {i} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 378, + 785, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $E_{i}$ denotes the $i$ -th expert and $G(\\mathbf{x_f})_i$ the probability for the point to be dispatched to that expert $i$ .", + "bbox": [ + 212, + 414, + 785, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Once we obtain radiance and density values for all the points on the ray, we can compute the pixel color with Equation 1.", + "bbox": [ + 212, + 444, + 785, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Resolution-based Routing", + "text_level": 1, + "bbox": [ + 214, + 494, + 475, + 510 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To balance the load and prevent the gate from focusing on assigning points to a single expert (typically the one with the highest resolution), we employ a resolution-weighted auxiliary loss. Given $M$ experts, and a batch of points $B$ , the auxiliary loss is defined as:", + "bbox": [ + 212, + 518, + 785, + 579 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {a u x}} = \\frac {M}{| B | ^ {2}} \\sum_ {i = 1} ^ {M} c _ {i} m _ {i}, \\quad m _ {i} = \\sum_ {\\mathbf {x} _ {f} \\in B} G (\\mathbf {x} _ {f}) _ {i}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 592, + 785, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $c_{i}$ represents the number of inputs dispatched to the expert $i$ , and $m_{i}$ is the sum of all the probabilities for each point in the batch for the expert $i$ . This loss helps balance the workload, ensuring that each expert processes a similar number of points. Ideally, in a perfectly balanced scenario, $L_{aux}$ is expected to be 1, as both $c_{i}$ and $m_{i}$ would be $\\frac{M}{N}$ , resulting in $\\sum_{i=1}^{M} c_{i} m_{i}$ being $\\frac{M^2}{N}$ . However, in the case of total imbalance, it tends to the number $M$ . We aim to take a step further by assigning as many points as possible to low-resolution experts while discouraging point assignment to high-resolution models. Hence, we introduce some penalty terms $w_{i}$ associated with each expert. The higher the resolution of the model, the higher the penalty. We define a novel resolution-based auxiliary loss:", + "bbox": [ + 212, + 638, + 787, + 803 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {r w - a u x}} = \\frac {M}{| B | ^ {2}} \\sum_ {i = 1} ^ {M} c _ {i} m _ {i} w _ {i}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 803, + 785, + 844 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "F. Di Sario et al.", + "bbox": [ + 271, + 114, + 387, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As for the weighting strategy, we opted for the following geometric progression. The weight $w_{i}$ for the $i$ expert is computed as:", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i} = \\exp \\left(\\frac {\\ln M}{M - 1}\\right) ^ {i}, \\quad i \\in [ 0, \\dots , M - 1 ]. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 189, + 785, + 223 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The total loss is then defined as:", + "bbox": [ + 238, + 224, + 478, + 239 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {t o t}} = L _ {\\mathrm {n e r f}} + \\lambda L _ {\\mathrm {r w - a u x}}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 247, + 785, + 263 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As we will see in Sec. 4, our proposed loss not only further improves the reconstruction quality but also encourages sparsity in higher-resolution models. A comparison with other weighting strategies is proposed in the Supplementary ??", + "bbox": [ + 214, + 267, + 782, + 314 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 332, + 375, + 349 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Setting", + "text_level": 1, + "bbox": [ + 215, + 358, + 323, + 373 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The code was written in Python 3.8, using PyTorch 1.12, and executed on a single NVIDIA A40 GPU. We tested our technique on the DVGO, TensoRF, and Instant-NGP architecture. The code from official repositories was used as the starting point. For Instant-NGP, the native implementation in PyTorch ngp_p1 was employed, which exhibits comparable performance and reconstruction quality to the official NVIDIA implementation. We experimented with various configurations, ranging from a minimum of 3 models to 5 models of different resolutions. The experts are ordered by resolution, such that the $i + 1$ -th expert has approximately double the parameters of the $i$ -th expert. For all configurations and models, $\\lambda$ was set to $10^{-3}$ . The Gate consists of a grid (dense voxel grid for DVGO, factorized grid for TensoRF, and hash grid for Instant-NGP) and a shallow MLP (2 layers with 64 neurons each) with ReLU activation function. The resolution of the gate is low: $128^3$ for both DVGO and TensoRF, while for Instant-NGP we used $L = 6$ . The number of iterations is set to $20k$ for all the architectures. All other hyperparameters are left as the original implementations. We draw experiments with our method using Top- $k$ experts, with $k = 1$ and $k = 2$ . We compare our results versus baselines with comparable resolutions, as well as a Fast-NeRF ensemble (Ens) obtained by jointly fine-tuning all models and averaging their predictions. It can be noted that this ensemble can be interpreted as a limit case for our method with $k = M$ and $G(\\mathbf{x}_f)_i = 1 / M$ , $\\forall i$ similar to the method proposed by [7].", + "bbox": [ + 212, + 378, + 787, + 696 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Metrics and Datasets", + "text_level": 1, + "bbox": [ + 215, + 714, + 439, + 728 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For each test, image-quality metrics such as PSNR, SSIM [30], and LPIPS [34] (computed on AlexNet [18]) are presented. Additionally, we report the number of non-zero parameters as $\\| w_0\\|$ , the average GFLOPs required by each model to render the images of the test set, and total training times. We present results obtained across four major datasets, namely: Synthetic-NeRF [24], Neural Sparse Voxel Field Dataset (NSVF) [20], TanksAndTemple [17], and Local Light Field Fusion Dataset (LLFF) [23].", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Boost Your NeRF", + "bbox": [ + 609, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/f738e21099d61812d7fbdc8b44896c72d49c65e3e1d0319804c87d6cfa7cdce2.jpg", + "table_caption": [ + "Table 1: Results on DVGO, TensoRF and Instant-NGP with $M = 5$ . $\\| w\\| _0$ is expressed as multiple of $10^{6}$ , while for Instant-NGP as a multiple of $10^{5}$ ." + ], + "table_footnote": [], + "table_body": "
DatasetMetricsDVGOTensorRFInstant-NGP
baselineTop-1Top-2EnsbaselineTop-1Top-2EnsbaselineTop-1Top-2Ens
BlenderPSNR ↑33.0433.4333.7433.7932.9833.6834.0934.0033.3533.5633.8334.01
SSIM ↑0.9610.9640.9650.9660.9580.9650.9680.9680.9630.9630.9650.966
LPIPS ↓0.0260.0240.0220.0220.0290.0230.0210.0220.0250.0450.0430.042
||w||0 ↓992639974024334931172126
GFLOPs ↓6354999402206886732134422144671123208
Time ↓26'21'25'32'44'69'76'70'11'32'34'37'
NSVFPSNR ↑35.2137.1237.5937.6836.7037.4037.9838.0836.4433.5937.0437.31
SSIM ↑0.9770.9840.9860.9860.9810.9840.9860.9870.9830.9830.9840.985
LPIPS ↓0.0150.0090.0080.0070.0130.0090.0080.0080.0100.0230.0220.020
||w||0 ↓9527431004228385430172027
GFLOPs ↓564430811190370657510531763285272123
Time ↓12'22'25'30'46'75'75'75'10'31'33'34'
TaTPSNR ↑28.9329.1429.2729.3728.4428.7829.1429.1129.0729.1629.3229.38
SSIM ↑0.9270.9250.9290.9320.9050.9240.9290.9280.9240.9270.9290.930
LPIPS ↓0.1070.1080.1050.10300.1060.0990.1090.1010.1250.1240.121
||w||0 ↓7416266579158088374770
GFLOPs ↓266616263066719835672791512691462112295311003
Time ↓22'25'28'38'72'70'78'101'14'37'39'45'
LLFFPSNR ↑26.2426.4326.6226.6526.7126.7327.0927.1024.9724.9025.1725.19
SSIM ↑0.8310.8320.8390.8430.8350.8360.8620.8640.7640.7630.7770.778
LPIPS ↓0.1360.1150.1110.1070.1140.1110.1010.1010.1280.2390.2370.234
||w||0 ↓622640113191116231526875148
GFLOPs ↓16781514250849724542322659211352257388713912712
Time ↓24'28'32'36'58'49'57'68'24'25'46'42'
", + "bbox": [ + 215, + 183, + 785, + 477 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Quantitative Results", + "text_level": 1, + "bbox": [ + 215, + 512, + 431, + 527 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The main experimental results in terms of the metrics defined above are shown in Table 1. Here, we present the results obtained using a mixture of $M = 5$ experts; further configurations can be found in the supplementary Sec. ?? Our experiments reveal that our MoE provides a significant rendering quality improvement with respect to the baseline with no or limited impact in terms of computational cost. The increase in rendering quality is notable in Synthetic NeRF (up to 1 dB) and NSVF (up to 1.3 dB). Although more moderate, improvements are still evident, even in more challenging datasets such as TanksAndTemple and LLFF, with about 0.5 dB gain. While Top-1 can already achieve state-of-the-art render quality, the Top-2 strategy further enhances image quality, reaching levels comparable to the ensemble but with much greater efficiency (about half the average FLOPs per rendering) and using significantly fewer parameters. This observation is also illustrated in Figure 3 , where FLOPS/PSNR plots are shown (each marker in every curve refers to the cases with $M = 3,4,5$ experts respectively). Training times are on average are longer but still acceptable (in the worst-case scenario, around 1h is required to train our MoE). However, in the case of Top1, they can be faster than baselines. This is because of our resolution methods, which tend to favor low-resolution models. Based on these analyses, the Top-2 strategy strikes an excellent quality/cost trade-off. The ensembling configura", + "bbox": [ + 212, + 551, + 787, + 843 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "F. Di Sario et al.", + "bbox": [ + 271, + 114, + 385, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/f9550c86b308a27fc0fb872ad9021cf67ac989f7a485604387820576fd5030fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 142, + 395, + 253 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e568527cb91f343b5ef103030064e80d5f334f78c6b4f9455a9202897cb0d82a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 142, + 584, + 253 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7cd19a8e4c18091a0f7e3b41fca58af0f5707f5d7c0c8a4f760f6dd06a21bb2e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 142, + 776, + 253 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/69dafdc302de52af7289909dc5a53fa52dace5473cbb51e58fccb4ac78aa80a7.jpg", + "image_caption": [ + "(a) DVGO" + ], + "image_footnote": [], + "bbox": [ + 220, + 263, + 395, + 375 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/fd742c4c39cf2372498587d65a3a7167952c51b97931c2178c4a43167cc806dc.jpg", + "image_caption": [ + "(b) TensorRF" + ], + "image_footnote": [], + "bbox": [ + 411, + 263, + 584, + 375 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/892b5dcd660df7c1bf65c52e7c0813c3841a8918b524ed6ce7a22262a941f1db.jpg", + "image_caption": [ + "(c) Instant-NGP", + "Fig. 3: PSNR/GFLOPs plots for Synthetic-NeRF and TanksAndTemple. The remaining datasets show similar results." + ], + "image_footnote": [], + "bbox": [ + 602, + 263, + 774, + 375 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "tion, using all experts, represents an upper bound in terms of image quality while being the worst case in computational terms.", + "bbox": [ + 214, + 462, + 784, + 492 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Qualitative Results", + "text_level": 1, + "bbox": [ + 215, + 513, + 419, + 527 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Figure 4 shows visual comparisons among the baseline, ensemble, Top-1, and Top-2. From the selected image crops, one can appreciate that our method ensures superior reconstruction quality compared to the baselines, effectively reproducing sharper details while reducing noise on texture-less spots. This is particularly notable when examining elements such as the window decorations in the Palace scene or the text on the box above the desk in the Room scene. Additionally, surfaces such as the semi-transparent glass in the Wineholder or the Caterpillar scene appear less noisy and more faithful to the original. This improvement can be attributed to the synergy among different resolution models: low-resolution models excel at representing lower frequencies, thereby introducing less noise, while high-resolution models can focus solely on high-frequency components. Additionally, it is important to notice that there is no significant difference between the ensemble and Top-2.", + "bbox": [ + 212, + 537, + 787, + 734 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.5 Comparison with Naive Methods", + "text_level": 1, + "bbox": [ + 215, + 770, + 534, + 785 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As shown in Figure 1, our method leads to significantly higher quality reconstructions at a greatly reduced computational cost. Decreasing the step size yields the least noticeable improvement while incurring a substantial computational", + "bbox": [ + 214, + 794, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Boost Your NeRF", + "bbox": [ + 609, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2d38a8a6f81d2ff182f38c17d027390221ce77d0fa37588021c595da6f3655b2.jpg", + "image_caption": [ + "Fig. 4: Qualitative results on some of the scenes of each dataset. From left to right: ground truths, baselines, Top-1, Top-2 and Ensemble. Each model has the same parameters and has been trained for the same number of iterations (DVGO)." + ], + "image_footnote": [], + "bbox": [ + 272, + 142, + 733, + 428 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "expense, as each sampled point requires an evaluation by the color decoder. Our method achieves high-quality reconstructions with the same step size as the baseline models. Increasing the MLP parameters can boost accuracy, but again, the computational costs become substantial compared to our method. Similarly, increasing resolution results in way inferior PSNR with respect to our method at a comparable computational cost. Our method also allows scaling to higher resolutions, while baseline models tend to introduce noise and artifacts, leading to a decrease in reconstruction quality as the resolution increases.", + "bbox": [ + 212, + 507, + 787, + 628 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.6 Gate Visualization and Scene Decomposition", + "text_level": 1, + "bbox": [ + 214, + 648, + 632, + 666 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In Figure 5, we visualize the gate (probability field in grayscale) and each expert output in the case $M = 3$ . On the right side of the figure, one can appreciate per-model renders and experts' specialization. It is worth noting how higher-resolution experts render high-frequency details. This is particularly evident in the Mic scene.", + "bbox": [ + 212, + 672, + 787, + 750 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.7 Ablation", + "text_level": 1, + "bbox": [ + 214, + 770, + 336, + 784 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Here we investigate how different design choices of our model affect the performance and rendering quality. These choices include the gate resolution, different gate formulations, and the number of experts in the Top- $k$ .", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "F. Di Sario et al.", + "bbox": [ + 271, + 114, + 387, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/08a346b4472e12ae5271af5ff9628e37dfb6f56de18ed3137174d89ffa77872c.jpg", + "image_caption": [ + "Fig. 5: Gate (gray-scale images) and per-model output visualizations with 3 experts and resolution-based routing (DVGO). Images are ordered by increasing resolutions." + ], + "image_footnote": [], + "bbox": [ + 272, + 143, + 730, + 390 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Here we investigate how different design choices of our model affect the performance and rendering quality. These choices include the gate resolution, different gate formulations, and the number of experts in the Top- $k$ .", + "bbox": [ + 212, + 467, + 784, + 513 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Gate Resolution and Gate Formulation. We show in Table 2 that a low-resolution gate is sufficient to achieve high-quality rendering. Interestingly, as the gate resolution increases, there is a slight decrease in rendering quality during testing, coupled with an increase in required FLOPs per image rendering and number of parameters. We also compare various gating strategies, including linear gating, the configuration proposed by Switch-NeRF, and our gate formulation. Through experimentation, it becomes evident that our approach outperforms others in terms of both render quality and performance, achieving performances comparable to a linear gating function.", + "bbox": [ + 212, + 555, + 787, + 691 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Why Top-2? Here we investigate the performance and rendering quality trends with varying values of $k$ . In Figure 6, we can see that while $k = 1$ can lead to a significant increase in quality with comparable performance to baseline models, the Top-2 further enhances quality at the expense of increased (but still acceptable) computation. The Top-3, Top-4, and Top-5 provide marginal quality improvements at a significantly higher computational cost. Hence, we consider the Top-2 the optimal balance between performance and quality.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Boost Your NeRF", + "bbox": [ + 609, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1f26f236a43fd9b466762daa74876b4e314eca3af1742b4d02c64143be1cb780.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ResGate TypePSNR||w||0GFLOPs
1283Ours36.826476
2563Ours36.7733601
3003Ours36.7942677
-Linear36.2524444
-Switch-NeRF36.44251095
", + "bbox": [ + 217, + 196, + 446, + 265 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 2: Comparison of different gate resolutions and formulations on Lego scene and DVGO.", + "bbox": [ + 215, + 276, + 446, + 316 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/044d1bb3a35c35583d4c3b24f456b2581b24df7d67eeed7e84f9e961634af261.jpg", + "image_caption": [ + "Fig. 6: Comparison of Gating Functions (Top-1 to Top-5) on Synthetic-NeRF with DVGO." + ], + "image_footnote": [], + "bbox": [ + 480, + 150, + 776, + 270 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Limitations", + "text_level": 1, + "bbox": [ + 215, + 349, + 362, + 366 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our study present some limitations. First of all, a pre-training phase, on which each model is trained independently is required for achieving good results. Training end-to-end, without a pre-training phase, can lead to reconstruction quality that is noticeably inferior to baselines. Pre-training models at different resolutions allows for diversified architectures, making it easier for the gate to learn the decomposition. The training is also strongly dependent on the auxiliary loss. Different values of $\\lambda$ can significantly influence performances and load balancing. We conducted a sweep to identify a value that works well across many scenarios, but it may not be suitable for different datasets. Another limitation is the overhead introduced by the MoE. Each input token is first interpolated with the gate's grid, decoded into probabilities and routed to the chosen expert. These operations can significantly slow-down the rendering process, leading to higher training and inference times. Although considerable effort was devoted to developing the most efficient gate possible, our MoE is still slower than the respective baselines.", + "bbox": [ + 212, + 390, + 787, + 616 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 648, + 369, + 666 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduced a model-agnostic framework for enhancing the rendering of Fast-NeRF models. Our formulation of the Gate reduces computational costs in both training and inference phases while ensuring better quality compared to a traditional gate. Additionally, the introduction of an auxiliary loss with res penalty allows for increased utilization of low-resolution models, reducing the number of active parameters and promoting sparsity in higher-resolution models. Our results demonstrate how this approach can significantly improve reconstruction quality while considering performance metrics. Specifically, we show that a Top-2 strategy strikes a good balance between performance and quality.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "F. Di Sario et al.", + "bbox": [ + 271, + 114, + 387, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was partially funded by Hi!PARIS Center on Data Analytics and Artificial Intelligence.", + "bbox": [ + 215, + 178, + 784, + 208 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 234, + 321, + 250 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)", + "2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706 (2023)", + "3. Chan, E.R., Monteiro, M., Kellnhofer, P., Wu, J., Wetzstein, G.: pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5799-5809 (2021)", + "4. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. ECCV (2022)", + "5. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16569-16578 (2023)", + "6. Deng, Y., Yang, J., Xiang, J., Tong, X.: Gram: Generative radiance manifolds for 3d-aware image generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10673-10683 (2022)", + "7. Di Sario, F., Renzulli, R., Tartaglione, E., Grangetto, M.: Two is better than one: Achieving high-quality 3d scene modeling with a nerf ensemble. In: International Conference on Image Analysis and Processing. pp. 320-331. Springer (2023)", + "8. Fedus, W., Zoph, B., Shazeer, N.: Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. The Journal of Machine Learning Research 23(1), 5232-5270 (2022)", + "9. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes for radiance fields in space, time, and appearance (2023)", + "0. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)", + "1. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High-fidelity neural rendering at 200fps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14346-14355 (2021)", + "2. Hou, A., Liu, F., Ren, Z., Sarkis, M., Bi, N., Tong, Y., Liu, X.: Infamous-nerf: Improving face modeling using semantically-aligned hypernetworks with neural radiance fields. arXiv preprint arXiv:2312.16197 (2023)", + "3. Hu, W., Wang, Y., Ma, L., Yang, B., Gao, L., Liu, X., Ma, Y.: Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19774-19783 (2023)" + ], + "bbox": [ + 225, + 267, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Boost Your NeRF", + "bbox": [ + 609, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Jacobs, R.A., Jordan, M.I., Nowlan, S.J., Hinton, G.E.: Adaptive mixtures of local experts. Neural computation 3(1), 79-87 (1991)", + "15. Jeong, Y., Ahn, S., Choy, C., Anandkumar, A., Cho, M., Park, J.: Self-calibrating neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5846-5854 (2021)", + "16. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023)", + "17. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG) 36(4), 1-13 (2017)", + "18. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25 (2012)", + "19. Lepikhin, D., Lee, H., Xu, Y., Chen, D., First, O., Huang, Y., Krikun, M., Shazeer, N., Gshard, Z.: Scaling giant models with conditional computation and automatic sharding. arXiv preprint arXiv:2006.16668 (2020)", + "20. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020)", + "21. Liu, Y., Li, X., Yu, F., Zhou, Q.: Probabilistic neural scene representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021)", + "22. Max, N.: Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics 1(2), 99-108 (1995)", + "23. Mildenhall, B., Srinivasan, P.P., Ortiz-Cayon, R., Kalantari, N.K., Ramamoorthi, R., Ng, R., Kar, A.: Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG) 38(4), 1-14 (2019)", + "24. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European conference on computer vision. pp. 405-421. Springer (2020)", + "25. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022)", + "26. Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14335-14345 (2021)", + "27. Riquelme, C., Puigcerver, J., Mustafa, B., Neumann, M., Jenatton, R., Susano Pinto, A., Keysers, D., Houlsby, N.: Scaling vision with sparse mixture of experts. Advances in Neural Information Processing Systems 34, 8583-8595 (2021)", + "28. Shazeer, N., Mirhoseini, A., Maziarz, K., Davis, A., Le, Q., Hinton, G., Dean, J.: Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538 (2017)", + "29. Sun, C., Sun, M., Chen, H.T.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5459-5469 (2022)", + "30. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "F. Di Sario et al.", + "bbox": [ + 271, + 114, + 385, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Wizadwongsa, S., Phongthawee, P., Yenphraphai, J., Suwajanakorn, S.: Nex: Real-time view synthesis with neural basis expansion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8534-8543 (2021)", + "32. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time neural irradiance fields for free-viewpoint video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9421-9431 (2021)", + "33. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoctrees for real-time rendering of neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5752-5761 (2021)", + "34. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)", + "35. Zhenxing, M., Xu, D.: Switch-nerf: Learning scene decomposition with mixture of experts for large-scale neural radiance fields. In: The Eleventh International Conference on Learning Representations (2022)", + "36. Zhou, Y., Lei, T., Liu, H., Du, N., Huang, Y., Zhao, V., Dai, A.M., Le, Q.V., Laudon, J., et al.: Mixture-of-experts with expert choice routing. Advances in Neural Information Processing Systems 35, 7103-7114 (2022)", + "37. Zhuang, Y., Zhu, H., Sun, X., Cao, X.: Mofanerf: Morphable facial neural radiance field. In: European conference on computer vision. pp. 268-285. Springer (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 424 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Boost Your NeRF", + "bbox": [ + 609, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Boost Your NeRF_ A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering/cdbb2d54-c057-481d-a45f-f2c36fd59358_model.json b/2024/Boost Your NeRF_ A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering/cdbb2d54-c057-481d-a45f-f2c36fd59358_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ab17f2a8c1ee3ee5903d1d53fe07893c4137b579 --- /dev/null +++ b/2024/Boost Your NeRF_ A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering/cdbb2d54-c057-481d-a45f-f2c36fd59358_model.json @@ -0,0 +1,2203 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.223, + 0.141, + 0.784, + 0.21 + ], + "angle": 0, + "content": "Boost Your NeRF: A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering" + }, + { + "type": "text", + "bbox": [ + 0.346, + 0.234, + 0.658, + 0.296 + ], + "angle": 0, + "content": "Francesco Di Sario\\(^{1}\\)[0009-0005-6969-1246], Riccardo Renzulli\\(^{1}\\)[0000-0003-0532-5966], Enzo Tartaglione\\(^{2}\\)[0000-0003-4274-8298], and Marco Grangetto\\(^{1}\\)[0000-0002-2709-7864]" + }, + { + "type": "text", + "bbox": [ + 0.29, + 0.307, + 0.712, + 0.35 + ], + "angle": 0, + "content": "1 University of Turin, Italy \n2 LTCI, Télécom Paris, Institut Polytechnique de Paris, France \nfrancesco.disario@unito.it" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.384, + 0.744, + 0.621 + ], + "angle": 0, + "content": "Abstract. Since the introduction of NeRFs, considerable attention has been focused on improving their training and inference times, leading to the development of Fast-NeRFs models. Despite demonstrating impressive rendering speed and quality, the rapid convergence of such models poses challenges for further improving reconstruction quality. Common strategies to improve rendering quality involves augmenting model parameters or increasing the number of sampled points. However, these computationally intensive approaches encounter limitations in achieving significant quality enhancements. This study introduces a model-agnostic framework inspired by Sparsely-Gated Mixture of Experts to enhance rendering quality without escalating computational complexity. Our approach enables specialization in rendering different scene components by employing a mixture of experts with varying resolutions. We present a novel gate formulation designed to maximize expert capabilities and propose a resolution-based routing technique to effectively induce sparsity and decompose scenes. Our work significantly improves reconstruction quality while maintaining competitive performance." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.644, + 0.377, + 0.66 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRFs) [24] have recently shown impressive results in synthesizing photo-realistic 3D scenes from a set of 2D images. However, NeRFs suffer from limited scene diversity, long training time, and sensitivity to training data [21]. Since the introduction of NeRFs, significant attention has been directed toward improving their training and inference times, resulting in the development of Fast-NeRFs. By using auxiliary data structures such as voxel grids to store scene geometry and skip empty spaces, the training and inference process can be accelerated by several orders of magnitude. As a result, the neural component of these models, which is usually used to transform learned features into view-dependent color representations, becomes much smaller and can sometimes be replaced entirely by spherical harmonics [10]. Despite Fast" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.128 + ], + "angle": 0, + "content": "F. Di Sario et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.147, + 0.404, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.147, + 0.589, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.147, + 0.777, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.257, + 0.788, + 0.367 + ], + "angle": 0, + "content": "Fig. 1: Different strategies for improving reconstruction quality with Fast-NeRFs (DVGO [29]). Increasing the resolution of data structures like voxel grids can improve render quality, but only up to a certain point, after which quality declines (left). The MLP component's impact on rendering quality is analyzed by varying its depth and width while keeping other variables constant (center). The effect of the number of sampled points along each ray is examined by decreasing the step size (right). In gray, our method's performance, which significantly improves PSNR at a low computational cost." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.784, + 0.427 + ], + "angle": 0, + "content": "NeRFs being an improvement from the original NeRF framework, achieving high-quality and efficient reconstructed geometry is still an open problem." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.784, + 0.456 + ], + "angle": 0, + "content": "Typical \"naive\" approaches to enhance the reconstruction quality of these models include:" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.468, + 0.784, + 0.498 + ], + "angle": 0, + "content": "1. Increasing the parameters and resolution of the used data structures (e.g., voxel grid, hash grid, etc.)." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.498, + 0.598, + 0.512 + ], + "angle": 0, + "content": "2. Increasing the number of sampled points per ray." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.512, + 0.785, + 0.542 + ], + "angle": 0, + "content": "3. Increasing the number of parameters in the neural network or the order of spherical harmonics." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.468, + 0.785, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.75 + ], + "angle": 0, + "content": "However, as evident from Figure 1, the increase in reconstruction quality results in a significant increase in computational costs. In the first approach, increasing the resolution can lead to a significant improvement in the reconstruction quality, but a plateau is reached beyond which overfitting occurs, and reconstruction quality degrades. This comes at the expense of a considerable increase in spatial complexity (especially with dense voxel grids) and training times. Increasing the number of sampled points per ray can marginally improve reconstruction quality further but at a significant increase in computational complexity. In fact, the higher the number of sampled rays, the higher forward passes through a neural network are needed. Augmenting the number of parameters in the neural network is another solution. However, as the neural network grows in depth and width, it tends towards a fully implicit model, deviating from the principles of fast models (limiting the neural part as much as possible - or even removing it altogether)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We propose a technique capable of significantly enhancing the reconstruction quality of such models while maintaining competitive training and rendering times. Inspired by the Sparse Mixture of Experts (MoE) paradigm [28], we have developed a model-agnostic framework capable of improving the reconstruction quality of various state-of-the-art models. Intuitively, our mixture of experts consists of different-capacity (resolution) models. During training, each model" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.61, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Boost Your NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.283 + ], + "angle": 0, + "content": "specializes in rendering the most suitable parts of the scene, i.e. low-resolution models render low-frequency parts of the scene, while high-resolution models render high-frequency parts. Our formulation of the gate allows our method to be model-agnostic, enabling, on the one hand, the insertion of the gate function at the early stages of the MoE mechanism and, on the other hand, the multiplication of the gate's output with the experts' output as late as possible. This maximizes the capabilities of the mixture of experts and allows for working directly with the output of each expert, which is considered a black box. This is why our method is inherently model-agnostic. In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.289, + 0.785, + 0.349 + ], + "angle": 0, + "content": "1. We propose the first model-agnostic framework based on Sparse MoE of models at different resolutions, which significantly improves the rendering quality of such models while maintaining competitive training and inference times (Sec. 3)." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.349, + 0.785, + 0.379 + ], + "angle": 0, + "content": "2. We provide a novel gate formulation inspired by Fast-NeRF models, which maximizes the capability of the mixture of experts (Sec. 3.3)." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.379, + 0.785, + 0.438 + ], + "angle": 0, + "content": "3. We introduce a new routing technique based on resolution, favoring the assignment of tokens to low-resolution models and discouraging the assignment of tokens to high-resolution models, inducing increasing sparsity in high-resolution models and decomposing the scene based on frequency (Sec. 3.5)." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.438, + 0.785, + 0.484 + ], + "angle": 0, + "content": "4. We conduct extensive experiments to test our method, including different NeRFs architectures and datasets, showing higher rendering quality and efficiency (Sec. 4)." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.289, + 0.785, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.504, + 0.391, + 0.518 + ], + "angle": 0, + "content": "2 Related works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.532, + 0.447, + 0.546 + ], + "angle": 0, + "content": "2.1 Neural Radiance Field" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.785, + 0.782 + ], + "angle": 0, + "content": "NeRF [24] (Neural Radiance Field) has emerged as a prominent method for synthesizing novel views, showing significant progress. This approach requires a moderate number of input images along with their known camera poses. Unlike traditional methods that rely on explicit and discretized volumetric representations such as voxel grids and multiplane images, NeRF employs a coordinate-based multilayer perceptron (MLP) to create an implicit and continuous volumetric representation. NeRFs can represent a 3D scene as a MLP \\( F_{\\theta} \\), with \\( \\theta \\) being the set of its trainable parameters, such that \\( F_{\\theta}:(\\mathbf{x},\\mathbf{d})\\to (\\mathbf{c},\\sigma) \\) which maps \\( (\\mathbf{x},\\mathbf{d}) \\), a 3D position \\( \\mathbf{x} \\) and a viewing direction \\( \\mathbf{d} \\), to a view-dependent color emission \\( \\mathbf{c} \\) and density value \\( \\sigma \\). To render the color of a pixel \\( \\hat{C}(\\mathbf{r}) \\), a ray \\( \\mathbf{r} \\) traverses the center of the camera through the pixel of the image plane from an origin point \\( \\mathbf{o} \\) to a ray having position \\( \\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d} \\). Then, \\( N \\) points are sampled on \\( \\mathbf{r} \\), and the MLP is queried for each point, obtaining a density and a color value. Finally, these results are accumulated into a single color with the volume rendering equation [22]:" + }, + { + "type": "equation", + "bbox": [ + 0.273, + 0.802, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\hat {C} (\\mathbf {r}) = \\sum_ {i = 1} ^ {K} T _ {i} \\alpha_ {i} c _ {i}, \\quad T _ {i} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\quad \\alpha_ {i} = 1 - \\exp \\left(\\sigma_ {i} \\delta_ {i}\\right), \\tag {1}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.387, + 0.128 + ], + "angle": 0, + "content": "F. Di Sario et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "where \\(\\alpha_{i}\\) is the probability of termination at point \\(i\\) and \\(T_{i}\\) is the accumulated transmittance from the near plane to point \\(i\\). NeRFs are trained by minimizing a photometric loss, which is an L2 loss between the rendered and the ground truth pixels. In more detail, given a batch \\(B\\) of randomly sampled rays, the loss is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.234, + 0.786, + 0.27 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {n e r f}} = \\frac {1}{| B |} \\sum_ {r \\in B} \\left\\| \\hat {C} (\\mathbf {r}) - C (\\mathbf {r}) \\right\\| _ {2} ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.278, + 0.787, + 0.31 + ], + "angle": 0, + "content": "where \\(\\hat{C}(\\mathbf{r})\\) is the predicted color and \\(C(\\mathbf{r})\\) the ground truth color for the ray \\(\\mathbf{r}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.332, + 0.496, + 0.346 + ], + "angle": 0, + "content": "2.2 Fast Neural Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Since the introduction of NeRF, significant effort has been directed towards the development of faster models. Several studies focus on speeding up rendering times, for example by working on ray sampling efficiency [3, 6, 12, 37], by integrating explicit volumetric representations [5, 11, 11, 15, 31-33] or by utilizing thousands of tiny MLPs [26]. However, all these methods still require large training times. One noteworthy development is represented by the introduction of explicit volumetric representations in the training pipeline, directly optimizing such representations [4, 9, 10, 13, 16, 25, 29]. These models enable fast training and inference, with render quality only slightly inferior to full-implicit models [1, 2]. We refer to these as Fast-NeRFs. Plenoxels [10] is the first important work in this context, as it represents a scene as a sparse 3D grid where each voxel stores spherical harmonic coefficients and density. Spherical harmonics serve as an orthogonal basis for functions defined over the sphere and thus can be used for computing view-dependent color emission, without the need for a multi-layer perceptron. Additionally, Plenoxels demonstrated the advantages of linearly interpolating voxels, facilitating the learning of a continuous plenoptic function throughout the volume, akin to NeRF, albeit with discrete data. Another notable work is DVGO [29]. Each scene is there represented as two dense voxel grids (one for density and one for feature colors) alongside a MLP, for learning view-dependent color. Similar to Plenoxels, the value of each voxel is linearly interpolated with the 8 nearest voxels, but after applying the activation functions. DVGO also incorporates a preliminary coarse geometry stage to learn the scene's general structure, facilitating adjustments to the bounding box and enabling a more intelligent ray sampling strategy. Despite utilizing lower-resolution models, DVGO achieves high reconstruction quality. The major drawback of these models is their substantial memory storage requirements. TensoRF [4] addresses this challenge by replacing the dense voxel grid with a planes and vectors decomposition, significantly reducing storage demands while maintaining comparable performance and rendering quality. Alternatively, Instant-NGP [25] proposed a multi-resolution voxel grid encoded via a hash function. They define \\( L \\) hash grids of increased resolutions: each entry of each hash grid has \\( 2^T \\) parameters and \\( F \\) features. This enables even faster training times and real-time rendering" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.61, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Boost Your NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "performances while maintaining a compact model; moreover, it demonstrates the efficacy of a multi-resolution approach in enhancing render quality. Inspired by this methodology, K-Planes [9] introduces a multi-resolution planar factorization of 3D space, offering reconstruction quality and model compactness similar to the previous methods, but with slightly larger training times. A similar multi-resolution planar factorization has also been proposed in Tri-MipRF [13] for mitigating the aliasing in distant or low-resolution views and blurriness in close-up shots. Given their properties, we decided to evaluate our paradigm using three different models: a 3D dense voxel grid-based model (DVGO), a decomposed grid-based model (TensoRF), and a multi-resolution hash grid-based model (Instant-NGP)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.334, + 0.566, + 0.35 + ], + "angle": 0, + "content": "2.3 Mixture of Experts and Sparse MoE" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.789, + 0.57 + ], + "angle": 0, + "content": "The Mixture of Experts (MoE) paradigm [14] has gained prominence in various machine learning applications. MoE consists of multiple expert networks, each specializing in different regions of the input space, with a meta-network determining the contribution of each expert to the final prediction. Building upon the MoE framework, Sparse Mixture of Experts [19, 27, 28, 36] constitutes a scalable and efficient variant. At the core of all Sparse MoE algorithms lies an assignment problem between tokens and experts. One approach to tackle this is by approximating the solution with a gating function, which learns to assign input tokens to the most suitable experts. It typically comprises a linear layer, a softmax activation, and a Top- \\( K \\) (where \\( 1 \\leq K \\leq 2 \\)) operation, aimed at routing the input token to only a subset of the experts. To balance the assignment across all the experts, auxiliary loss functions penalizing unevenly distributed routing are often employed. This sparsity-inducing technique significantly reduces computational overhead while preserving the expressive power of the MoE architecture." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.59, + 0.396, + 0.605 + ], + "angle": 0, + "content": "2.4 MoE and NeRF" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Combining the sparse mixture of experts' paradigms with neural radiance fields presents a non-trivial challenge. In recent research, Switch-NeRF [35] has been introduced as a novel end-to-end large-scale NeRF with learning-based scene decomposition. Inspired by Fedus et al. [8], they propose a full-implicit model with an MLP-based gate comprising 4 layers with 128 neurons and a Top-1 function. However, their architecture is ad-hoc, and suffers from extensive training times. This poses a challenge for us, as we aim to consider each expert as a black box, taking a point in space and a direction as input and outputting radiance and density values. Moreover, as the mixture comprises experts at different resolutions, we also aim to prioritize routing input tokens toward lower-resolution models and minimize the usage of high-resolution models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.781, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Our work intends to overcome these limitations. We design a novel gate formulation inspired by Fast-NeRFs, that guarantees fast convergence and better performances. We position the gate at the beginning of our MoE pipeline and postpone the multiplication of the gate output with expert predictions as much as" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.128 + ], + "angle": 0, + "content": "F. Di Sario et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.143, + 0.785, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.308, + 0.789, + 0.395 + ], + "angle": 0, + "content": "Fig. 2: A density field is used to compute density values for sampled points along a ray. A filtering step discards low-density points, routed through gating network \\( G \\) for expert assignment. Top- \\( K \\) experts compute radiance and density, aggregating and weighting these values by the corresponding gating probability to get the final values \\( \\mathbf{c}_f \\) and \\( \\sigma_f \\). Volume rendering equation yields pixel colors, and joint optimization refines our resolution-weighted auxiliary loss, allowing for high-quality and efficient rendering." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.789, + 0.483 + ], + "angle": 0, + "content": "possible. This allows our framework to be entirely model-agnostic. Furthermore, our mixture of experts maintains competitive training times, enabling, in the worst case, training models on complex scenes in approximately 1 hour while achieving state-of-the-art accuracy." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.505, + 0.331, + 0.521 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.789, + 0.629 + ], + "angle": 0, + "content": "In this section, we introduce our model-agnostic assembling approach. We present in Sec. 3.1 an overview of our method. We also describe the ray sampling and filtering mechanism in Sec. 3.2. Then, in Sec. 3.3, we discuss the effective construction of a mixture of experts with a gating mechanism in the context of NeRF. Finally, in Sec. 3.5, we present a novel formulation for decomposing the scene into resolution-based parts with an end-to-end training procedure." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.65, + 0.342, + 0.663 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Figure 2 shows a high-level scheme of our method. After independently training a set of \\(M\\) NeRF models \\(\\{E_i\\}_{i=1}^M\\) at different resolutions, we sample \\(N\\) points along a ray \\(\\mathbf{r}\\). For each point \\(\\mathbf{x}_p\\), a density value \\(\\sigma\\) is calculated, and based on this, a filtering step is performed based on the density volume \\(\\mathbf{V}_D\\), discarding points in regions with a density below a certain threshold \\(T\\). Initializing the density volume \\(\\mathbf{V}_D\\) with the one learned from the lowest resolution model can be helpful (though it can also be learned). Subsequently, the filtered point \\(\\mathbf{x}_f\\) is fed to the gate \\(G\\), so we denote with \\(G(\\mathbf{x}_f)\\) the probabilities with which the point is assigned to the \\(M\\) experts. Based on these values, each point is routed to the Top-\\(K\\) experts. Each selected expert then computes the radiance \\(\\mathbf{c}_{i,f}\\) and density values \\(\\sigma_{i,f}\\) for the point, both of which are multiplied by their respective" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.61, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Boost Your NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "probabilities and summed together to get the final radiance \\(\\mathbf{c}_f\\) and density \\(\\sigma_f\\). Subsequently, the volume rendering Equation 1 is used to aggregate all colors and compute the pixel color for the given ray. Finally, we compute the total loss and jointly optimize the gate and the experts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.209, + 0.784, + 0.239 + ], + "angle": 0, + "content": "In the next sections, we will dive into the gating and sparse mixture of expert modules." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.27, + 0.492, + 0.286 + ], + "angle": 0, + "content": "3.2 Ray Sampling and Filtering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.304, + 0.784, + 0.351 + ], + "angle": 0, + "content": "The first phase of our method involves learning a coarse and explicit density field starting from the density volume \\(\\mathbf{V}_D\\), that we can leverage for skipping empty spaces. Given a ray \\(\\mathbf{r}\\) as explained in Sec. 2.1, we sample \\(N\\) points along it:" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.364, + 0.786, + 0.382 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {p, r} = \\text {s a m p l i n g} (\\mathbf {r}) \\in \\mathbb {R} ^ {N \\times 3}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.784, + 0.457 + ], + "angle": 0, + "content": "with \\( p \\in [0, \\dots, N - 1] \\). We suppress the index \\( r \\) for abuse of notation. Next, we compute the density for each point \\( \\mathbf{x}_p \\in \\mathbb{R}^3 \\) and discard those with negligible density. The density value \\( \\sigma_{D,p} \\) for each point is computed by linearly interpolating \\( v \\) neighboring voxels:" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.48, + 0.785, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {D, p} = \\operatorname {a c t} (\\operatorname {i n t e r p o l a t e} \\left(\\mathbf {x} _ {p}, V _ {D}\\right)) \\in \\mathbb {R} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.784, + 0.537 + ], + "angle": 0, + "content": "where act represents a density activation function (such as softmax). We denote \\( \\mathbf{x}_{\\mathbf{f}} \\) with \\( \\mathbf{x}_{\\mathbf{f}} \\in R^3 \\) a remaining point after the filtering operation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.568, + 0.462, + 0.584 + ], + "angle": 0, + "content": "3.3 Trainable Gating Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.784, + 0.647 + ], + "angle": 0, + "content": "Our gating mechanism incorporates a hybrid architecture: an explicit feature grid \\(\\mathbf{V}_G\\) and a shallow MLP. The gating mechanism can also be seen as a probability field. First, we compute per-point features:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.662, + 0.785, + 0.679 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f e a t} = \\operatorname {i n t e r p o l a t e} \\left(\\mathbf {x} _ {f}, \\mathbf {V} _ {G}\\right) \\in \\mathbb {R} ^ {C} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.693, + 0.784, + 0.724 + ], + "angle": 0, + "content": "where \\( C \\) represents the number of channels of \\( \\mathbf{V}_G \\). Subsequently, we transform each feature into a probability. We compute logits as" + }, + { + "type": "equation", + "bbox": [ + 0.406, + 0.745, + 0.785, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\operatorname {l o g i t s} = \\operatorname {M L P} (\\mathbf {f e a t}) \\in \\mathbb {R} ^ {M} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.773, + 0.784, + 0.803 + ], + "angle": 0, + "content": "where \\( M \\) is the number of experts in the mixture of experts. Finally, we apply a per-row softmax. Our gating function can be summarized as:" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.825, + 0.785, + 0.842 + ], + "angle": 0, + "content": "\\[\nG \\left(\\mathbf {x} _ {\\mathbf {f}}\\right) = \\operatorname {s o f t m a x} (\\operatorname {M L P} (\\text {i n t e r p o l a t e} \\left(\\mathbf {x} _ {f}, \\mathbf {V} _ {G}\\right))) \\in \\mathbb {R} ^ {M} \\tag {7}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.128 + ], + "angle": 0, + "content": "F. Di Sario et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.39, + 0.163 + ], + "angle": 0, + "content": "3.4 NeRFs Experts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.232 + ], + "angle": 0, + "content": "As mentioned in Sec. 3.1, we employ a set of \\(M\\) pretrained NeRFs models as experts. After feeding as input each point \\(\\mathbf{x}_f\\) to the gate, we route them to the \\(k\\) experts with the \\(k\\) highest probabilities \\(G(\\mathbf{x}_f)\\). We define this set of indexes of the selected experts as" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.242, + 0.786, + 0.267 + ], + "angle": 0, + "content": "\\[\n\\mathcal {K} = \\underset {T o p - k} {\\operatorname {a r g m a x}} (G (\\mathbf {x} _ {f})). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.275, + 0.789, + 0.366 + ], + "angle": 0, + "content": "Each expert is a Fast-NeRF model that receives the dispatched points and the direction of the ray they lie on as input and outputs a density and a radiance value. We treat each expert as a black box, ensuring our method is inherently model-agnostic. The radiance \\(\\mathbf{c}_f\\) and the density \\(\\sigma_f\\) for the point \\(\\mathbf{x}_f\\) laying on a ray with direction \\(\\mathbf{d}\\) can be expressed as the sum of each expert predictions weighted by their probability:" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.379, + 0.786, + 0.41 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {f}, \\sigma_ {f} = \\sum_ {i \\in \\mathcal {K}} E _ {i} \\left(\\mathbf {x} _ {\\mathbf {f}}, \\mathbf {d}\\right) \\cdot G \\left(\\mathbf {x} _ {\\mathbf {f}}\\right) _ {i} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.787, + 0.445 + ], + "angle": 0, + "content": "where \\( E_{i} \\) denotes the \\( i \\)-th expert and \\( G(\\mathbf{x_f})_i \\) the probability for the point to be dispatched to that expert \\( i \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.787, + 0.475 + ], + "angle": 0, + "content": "Once we obtain radiance and density values for all the points on the ray, we can compute the pixel color with Equation 1." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.496, + 0.476, + 0.511 + ], + "angle": 0, + "content": "3.5 Resolution-based Routing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.787, + 0.58 + ], + "angle": 0, + "content": "To balance the load and prevent the gate from focusing on assigning points to a single expert (typically the one with the highest resolution), we employ a resolution-weighted auxiliary loss. Given \\( M \\) experts, and a batch of points \\( B \\), the auxiliary loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.593, + 0.786, + 0.634 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {a u x}} = \\frac {M}{| B | ^ {2}} \\sum_ {i = 1} ^ {M} c _ {i} m _ {i}, \\quad m _ {i} = \\sum_ {\\mathbf {x} _ {f} \\in B} G (\\mathbf {x} _ {f}) _ {i}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.639, + 0.788, + 0.804 + ], + "angle": 0, + "content": "where \\( c_{i} \\) represents the number of inputs dispatched to the expert \\( i \\), and \\( m_{i} \\) is the sum of all the probabilities for each point in the batch for the expert \\( i \\). This loss helps balance the workload, ensuring that each expert processes a similar number of points. Ideally, in a perfectly balanced scenario, \\( L_{aux} \\) is expected to be 1, as both \\( c_{i} \\) and \\( m_{i} \\) would be \\( \\frac{M}{N} \\), resulting in \\( \\sum_{i=1}^{M} c_{i} m_{i} \\) being \\( \\frac{M^2}{N} \\). However, in the case of total imbalance, it tends to the number \\( M \\). We aim to take a step further by assigning as many points as possible to low-resolution experts while discouraging point assignment to high-resolution models. Hence, we introduce some penalty terms \\( w_{i} \\) associated with each expert. The higher the resolution of the model, the higher the penalty. We define a novel resolution-based auxiliary loss:" + }, + { + "type": "equation", + "bbox": [ + 0.391, + 0.804, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {r w - a u x}} = \\frac {M}{| B | ^ {2}} \\sum_ {i = 1} ^ {M} c _ {i} m _ {i} w _ {i}. \\tag {11}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.61, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Boost Your NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "As for the weighting strategy, we opted for the following geometric progression. The weight \\( w_{i} \\) for the \\( i \\) expert is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.19, + 0.786, + 0.224 + ], + "angle": 0, + "content": "\\[\nw _ {i} = \\exp \\left(\\frac {\\ln M}{M - 1}\\right) ^ {i}, \\quad i \\in [ 0, \\dots , M - 1 ]. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.225, + 0.479, + 0.24 + ], + "angle": 0, + "content": "The total loss is then defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.418, + 0.248, + 0.786, + 0.264 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {t o t}} = L _ {\\mathrm {n e r f}} + \\lambda L _ {\\mathrm {r w - a u x}}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.268, + 0.784, + 0.315 + ], + "angle": 0, + "content": "As we will see in Sec. 4, our proposed loss not only further improves the reconstruction quality but also encourages sparsity in higher-resolution models. A comparison with other weighting strategies is proposed in the Supplementary ??" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.333, + 0.376, + 0.35 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.359, + 0.324, + 0.374 + ], + "angle": 0, + "content": "4.1 Setting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.789, + 0.697 + ], + "angle": 0, + "content": "The code was written in Python 3.8, using PyTorch 1.12, and executed on a single NVIDIA A40 GPU. We tested our technique on the DVGO, TensoRF, and Instant-NGP architecture. The code from official repositories was used as the starting point. For Instant-NGP, the native implementation in PyTorch ngp_p1 was employed, which exhibits comparable performance and reconstruction quality to the official NVIDIA implementation. We experimented with various configurations, ranging from a minimum of 3 models to 5 models of different resolutions. The experts are ordered by resolution, such that the \\(i + 1\\)-th expert has approximately double the parameters of the \\(i\\)-th expert. For all configurations and models, \\(\\lambda\\) was set to \\(10^{-3}\\). The Gate consists of a grid (dense voxel grid for DVGO, factorized grid for TensoRF, and hash grid for Instant-NGP) and a shallow MLP (2 layers with 64 neurons each) with ReLU activation function. The resolution of the gate is low: \\(128^3\\) for both DVGO and TensoRF, while for Instant-NGP we used \\(L = 6\\). The number of iterations is set to \\(20k\\) for all the architectures. All other hyperparameters are left as the original implementations. We draw experiments with our method using Top-\\(k\\) experts, with \\(k = 1\\) and \\(k = 2\\). We compare our results versus baselines with comparable resolutions, as well as a Fast-NeRF ensemble (Ens) obtained by jointly fine-tuning all models and averaging their predictions. It can be noted that this ensemble can be interpreted as a limit case for our method with \\(k = M\\) and \\(G(\\mathbf{x}_f)_i = 1 / M\\), \\(\\forall i\\) similar to the method proposed by [7]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.715, + 0.44, + 0.729 + ], + "angle": 0, + "content": "4.2 Metrics and Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.842 + ], + "angle": 0, + "content": "For each test, image-quality metrics such as PSNR, SSIM [30], and LPIPS [34] (computed on AlexNet [18]) are presented. Additionally, we report the number of non-zero parameters as \\( \\| w_0\\| \\), the average GFLOPs required by each model to render the images of the test set, and total training times. We present results obtained across four major datasets, namely: Synthetic-NeRF [24], Neural Sparse Voxel Field Dataset (NSVF) [20], TanksAndTemple [17], and Local Light Field Fusion Dataset (LLFF) [23]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.387, + 0.128 + ], + "angle": 0, + "content": "F. Di Sario et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 1: Results on DVGO, TensoRF and Instant-NGP with \\( M = 5 \\). \\( \\| w\\| _0 \\) is expressed as multiple of \\( 10^{6} \\), while for Instant-NGP as a multiple of \\( 10^{5} \\)." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.184, + 0.787, + 0.478 + ], + "angle": 0, + "content": "
DatasetMetricsDVGOTensorRFInstant-NGP
baselineTop-1Top-2EnsbaselineTop-1Top-2EnsbaselineTop-1Top-2Ens
BlenderPSNR ↑33.0433.4333.7433.7932.9833.6834.0934.0033.3533.5633.8334.01
SSIM ↑0.9610.9640.9650.9660.9580.9650.9680.9680.9630.9630.9650.966
LPIPS ↓0.0260.0240.0220.0220.0290.0230.0210.0220.0250.0450.0430.042
||w||0 ↓992639974024334931172126
GFLOPs ↓6354999402206886732134422144671123208
Time ↓26'21'25'32'44'69'76'70'11'32'34'37'
NSVFPSNR ↑35.2137.1237.5937.6836.7037.4037.9838.0836.4433.5937.0437.31
SSIM ↑0.9770.9840.9860.9860.9810.9840.9860.9870.9830.9830.9840.985
LPIPS ↓0.0150.0090.0080.0070.0130.0090.0080.0080.0100.0230.0220.020
||w||0 ↓9527431004228385430172027
GFLOPs ↓564430811190370657510531763285272123
Time ↓12'22'25'30'46'75'75'75'10'31'33'34'
TaTPSNR ↑28.9329.1429.2729.3728.4428.7829.1429.1129.0729.1629.3229.38
SSIM ↑0.9270.9250.9290.9320.9050.9240.9290.9280.9240.9270.9290.930
LPIPS ↓0.1070.1080.1050.10300.1060.0990.1090.1010.1250.1240.121
||w||0 ↓7416266579158088374770
GFLOPs ↓266616263066719835672791512691462112295311003
Time ↓22'25'28'38'72'70'78'101'14'37'39'45'
LLFFPSNR ↑26.2426.4326.6226.6526.7126.7327.0927.1024.9724.9025.1725.19
SSIM ↑0.8310.8320.8390.8430.8350.8360.8620.8640.7640.7630.7770.778
LPIPS ↓0.1360.1150.1110.1070.1140.1110.1010.1010.1280.2390.2370.234
||w||0 ↓622640113191116231526875148
GFLOPs ↓16781514250849724542322659211352257388713912712
Time ↓24'28'32'36'58'49'57'68'24'25'46'42'
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.513, + 0.433, + 0.529 + ], + "angle": 0, + "content": "4.3 Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.844 + ], + "angle": 0, + "content": "The main experimental results in terms of the metrics defined above are shown in Table 1. Here, we present the results obtained using a mixture of \\( M = 5 \\) experts; further configurations can be found in the supplementary Sec. ?? Our experiments reveal that our MoE provides a significant rendering quality improvement with respect to the baseline with no or limited impact in terms of computational cost. The increase in rendering quality is notable in Synthetic NeRF (up to 1 dB) and NSVF (up to 1.3 dB). Although more moderate, improvements are still evident, even in more challenging datasets such as TanksAndTemple and LLFF, with about 0.5 dB gain. While Top-1 can already achieve state-of-the-art render quality, the Top-2 strategy further enhances image quality, reaching levels comparable to the ensemble but with much greater efficiency (about half the average FLOPs per rendering) and using significantly fewer parameters. This observation is also illustrated in Figure 3 , where FLOPS/PSNR plots are shown (each marker in every curve refers to the cases with \\( M = 3,4,5 \\) experts respectively). Training times are on average are longer but still acceptable (in the worst-case scenario, around 1h is required to train our MoE). However, in the case of Top1, they can be faster than baselines. This is because of our resolution methods, which tend to favor low-resolution models. Based on these analyses, the Top-2 strategy strikes an excellent quality/cost trade-off. The ensembling configura" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.61, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Boost Your NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.143, + 0.396, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.413, + 0.143, + 0.586, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.143, + 0.777, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.264, + 0.396, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.278, + 0.382, + 0.342, + 0.394 + ], + "angle": 0, + "content": "(a) DVGO" + }, + { + "type": "image", + "bbox": [ + 0.413, + 0.264, + 0.586, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.461, + 0.382, + 0.537, + 0.394 + ], + "angle": 0, + "content": "(b) TensorRF" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.264, + 0.776, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.382, + 0.738, + 0.394 + ], + "angle": 0, + "content": "(c) Instant-NGP" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.406, + 0.785, + 0.434 + ], + "angle": 0, + "content": "Fig. 3: PSNR/GFLOPs plots for Synthetic-NeRF and TanksAndTemple. The remaining datasets show similar results." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.463, + 0.785, + 0.493 + ], + "angle": 0, + "content": "tion, using all experts, represents an upper bound in terms of image quality while being the worst case in computational terms." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.515, + 0.421, + 0.529 + ], + "angle": 0, + "content": "4.4 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Figure 4 shows visual comparisons among the baseline, ensemble, Top-1, and Top-2. From the selected image crops, one can appreciate that our method ensures superior reconstruction quality compared to the baselines, effectively reproducing sharper details while reducing noise on texture-less spots. This is particularly notable when examining elements such as the window decorations in the Palace scene or the text on the box above the desk in the Room scene. Additionally, surfaces such as the semi-transparent glass in the Wineholder or the Caterpillar scene appear less noisy and more faithful to the original. This improvement can be attributed to the synergy among different resolution models: low-resolution models excel at representing lower frequencies, thereby introducing less noise, while high-resolution models can focus solely on high-frequency components. Additionally, it is important to notice that there is no significant difference between the ensemble and Top-2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.771, + 0.535, + 0.786 + ], + "angle": 0, + "content": "4.5 Comparison with Naive Methods" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "As shown in Figure 1, our method leads to significantly higher quality reconstructions at a greatly reduced computational cost. Decreasing the step size yields the least noticeable improvement while incurring a substantial computational" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.128 + ], + "angle": 0, + "content": "F. Di Sario et al." + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.143, + 0.734, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.437, + 0.788, + 0.48 + ], + "angle": 0, + "content": "Fig. 4: Qualitative results on some of the scenes of each dataset. From left to right: ground truths, baselines, Top-1, Top-2 and Ensemble. Each model has the same parameters and has been trained for the same number of iterations (DVGO)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.63 + ], + "angle": 0, + "content": "expense, as each sampled point requires an evaluation by the color decoder. Our method achieves high-quality reconstructions with the same step size as the baseline models. Increasing the MLP parameters can boost accuracy, but again, the computational costs become substantial compared to our method. Similarly, increasing resolution results in way inferior PSNR with respect to our method at a comparable computational cost. Our method also allows scaling to higher resolutions, while baseline models tend to introduce noise and artifacts, leading to a decrease in reconstruction quality as the resolution increases." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.65, + 0.633, + 0.667 + ], + "angle": 0, + "content": "4.6 Gate Visualization and Scene Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.75 + ], + "angle": 0, + "content": "In Figure 5, we visualize the gate (probability field in grayscale) and each expert output in the case \\( M = 3 \\). On the right side of the figure, one can appreciate per-model renders and experts' specialization. It is worth noting how higher-resolution experts render high-frequency details. This is particularly evident in the Mic scene." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.771, + 0.337, + 0.785 + ], + "angle": 0, + "content": "4.7 Ablation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Here we investigate how different design choices of our model affect the performance and rendering quality. These choices include the gate resolution, different gate formulations, and the number of experts in the Top-\\(k\\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.61, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Boost Your NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.144, + 0.731, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.402, + 0.785, + 0.43 + ], + "angle": 0, + "content": "Fig. 5: Gate (gray-scale images) and per-model output visualizations with 3 experts and resolution-based routing (DVGO). Images are ordered by increasing resolutions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.468, + 0.785, + 0.514 + ], + "angle": 0, + "content": "Here we investigate how different design choices of our model affect the performance and rendering quality. These choices include the gate resolution, different gate formulations, and the number of experts in the Top-\\(k\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.788, + 0.693 + ], + "angle": 0, + "content": "Gate Resolution and Gate Formulation. We show in Table 2 that a low-resolution gate is sufficient to achieve high-quality rendering. Interestingly, as the gate resolution increases, there is a slight decrease in rendering quality during testing, coupled with an increase in required FLOPs per image rendering and number of parameters. We also compare various gating strategies, including linear gating, the configuration proposed by Switch-NeRF, and our gate formulation. Through experimentation, it becomes evident that our approach outperforms others in terms of both render quality and performance, achieving performances comparable to a linear gating function." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Why Top-2? Here we investigate the performance and rendering quality trends with varying values of \\( k \\). In Figure 6, we can see that while \\( k = 1 \\) can lead to a significant increase in quality with comparable performance to baseline models, the Top-2 further enhances quality at the expense of increased (but still acceptable) computation. The Top-3, Top-4, and Top-5 provide marginal quality improvements at a significantly higher computational cost. Hence, we consider the Top-2 the optimal balance between performance and quality." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.388, + 0.128 + ], + "angle": 0, + "content": "F. Di Sario et al." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.197, + 0.447, + 0.266 + ], + "angle": 0, + "content": "
ResGate TypePSNR||w||0GFLOPs
1283Ours36.826476
2563Ours36.7733601
3003Ours36.7942677
-Linear36.2524444
-Switch-NeRF36.44251095
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.277, + 0.447, + 0.318 + ], + "angle": 0, + "content": "Table 2: Comparison of different gate resolutions and formulations on Lego scene and DVGO." + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.151, + 0.777, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.47, + 0.291, + 0.785, + 0.319 + ], + "angle": 0, + "content": "Fig. 6: Comparison of Gating Functions (Top-1 to Top-5) on Synthetic-NeRF with DVGO." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.351, + 0.364, + 0.367 + ], + "angle": 0, + "content": "5 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.788, + 0.617 + ], + "angle": 0, + "content": "Our study present some limitations. First of all, a pre-training phase, on which each model is trained independently is required for achieving good results. Training end-to-end, without a pre-training phase, can lead to reconstruction quality that is noticeably inferior to baselines. Pre-training models at different resolutions allows for diversified architectures, making it easier for the gate to learn the decomposition. The training is also strongly dependent on the auxiliary loss. Different values of \\(\\lambda\\) can significantly influence performances and load balancing. We conducted a sweep to identify a value that works well across many scenarios, but it may not be suitable for different datasets. Another limitation is the overhead introduced by the MoE. Each input token is first interpolated with the gate's grid, decoded into probabilities and routed to the chosen expert. These operations can significantly slow-down the rendering process, leading to higher training and inference times. Although considerable effort was devoted to developing the most efficient gate possible, our MoE is still slower than the respective baselines." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.65, + 0.37, + 0.667 + ], + "angle": 0, + "content": "6 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we introduced a model-agnostic framework for enhancing the rendering of Fast-NeRF models. Our formulation of the Gate reduces computational costs in both training and inference phases while ensuring better quality compared to a traditional gate. Additionally, the introduction of an auxiliary loss with res penalty allows for increased utilization of low-resolution models, reducing the number of active parameters and promoting sparsity in higher-resolution models. Our results demonstrate how this approach can significantly improve reconstruction quality while considering performance metrics. Specifically, we show that a Top-2 strategy strikes a good balance between performance and quality." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.61, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Boost Your NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.179, + 0.785, + 0.209 + ], + "angle": 0, + "content": "This work was partially funded by Hi!PARIS Center on Data Analytics and Artificial Intelligence." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.235, + 0.323, + 0.25 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.268, + 0.786, + 0.323 + ], + "angle": 0, + "content": "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.786, + 0.365 + ], + "angle": 0, + "content": "2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.367, + 0.786, + 0.42 + ], + "angle": 0, + "content": "3. Chan, E.R., Monteiro, M., Kellnhofer, P., Wu, J., Wetzstein, G.: pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5799-5809 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.422, + 0.786, + 0.449 + ], + "angle": 0, + "content": "4. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.45, + 0.786, + 0.505 + ], + "angle": 0, + "content": "5. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16569-16578 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.507, + 0.786, + 0.546 + ], + "angle": 0, + "content": "6. Deng, Y., Yang, J., Xiang, J., Tong, X.: Gram: Generative radiance manifolds for 3d-aware image generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10673-10683 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.548, + 0.786, + 0.589 + ], + "angle": 0, + "content": "7. Di Sario, F., Renzulli, R., Tartaglione, E., Grangetto, M.: Two is better than one: Achieving high-quality 3d scene modeling with a nerf ensemble. In: International Conference on Image Analysis and Processing. pp. 320-331. Springer (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.591, + 0.786, + 0.631 + ], + "angle": 0, + "content": "8. Fedus, W., Zoph, B., Shazeer, N.: Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. The Journal of Machine Learning Research 23(1), 5232-5270 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.632, + 0.786, + 0.659 + ], + "angle": 0, + "content": "9. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes for radiance fields in space, time, and appearance (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.786, + 0.701 + ], + "angle": 0, + "content": "0. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.786, + 0.742 + ], + "angle": 0, + "content": "1. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High-fidelity neural rendering at 200fps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14346-14355 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.743, + 0.786, + 0.784 + ], + "angle": 0, + "content": "2. Hou, A., Liu, F., Ren, Z., Sarkis, M., Bi, N., Tong, Y., Liu, X.: Infamous-nerf: Improving face modeling using semantically-aligned hypernetworks with neural radiance fields. arXiv preprint arXiv:2312.16197 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.786, + 0.84 + ], + "angle": 0, + "content": "3. Hu, W., Wang, Y., Ma, L., Yang, B., Gao, L., Liu, X., Ma, Y.: Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19774-19783 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.268, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.386, + 0.127 + ], + "angle": 0, + "content": "F. Di Sario et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "14. Jacobs, R.A., Jordan, M.I., Nowlan, S.J., Hinton, G.E.: Adaptive mixtures of local experts. Neural computation 3(1), 79-87 (1991)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "15. Jeong, Y., Ahn, S., Choy, C., Anandkumar, A., Cho, M., Park, J.: Self-calibrating neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5846-5854 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.219, + 0.785, + 0.26 + ], + "angle": 0, + "content": "16. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.261, + 0.785, + 0.303 + ], + "angle": 0, + "content": "17. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG) 36(4), 1-13 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.304, + 0.785, + 0.346 + ], + "angle": 0, + "content": "18. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.346, + 0.785, + 0.388 + ], + "angle": 0, + "content": "19. Lepikhin, D., Lee, H., Xu, Y., Chen, D., First, O., Huang, Y., Krikun, M., Shazeer, N., Gshard, Z.: Scaling giant models with conditional computation and automatic sharding. arXiv preprint arXiv:2006.16668 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.785, + 0.417 + ], + "angle": 0, + "content": "20. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.418, + 0.785, + 0.459 + ], + "angle": 0, + "content": "21. Liu, Y., Li, X., Yu, F., Zhou, Q.: Probabilistic neural scene representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.487 + ], + "angle": 0, + "content": "22. Max, N.: Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics 1(2), 99-108 (1995)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.488, + 0.785, + 0.543 + ], + "angle": 0, + "content": "23. Mildenhall, B., Srinivasan, P.P., Ortiz-Cayon, R., Kalantari, N.K., Ramamoorthi, R., Ng, R., Kar, A.: Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG) 38(4), 1-14 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.545, + 0.785, + 0.586 + ], + "angle": 0, + "content": "24. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European conference on computer vision. pp. 405-421. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.587, + 0.785, + 0.628 + ], + "angle": 0, + "content": "25. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.63, + 0.785, + 0.671 + ], + "angle": 0, + "content": "26. Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14335-14345 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.672, + 0.785, + 0.713 + ], + "angle": 0, + "content": "27. Riquelme, C., Puigcerver, J., Mustafa, B., Neumann, M., Jenatton, R., Susano Pinto, A., Keysers, D., Houlsby, N.: Scaling vision with sparse mixture of experts. Advances in Neural Information Processing Systems 34, 8583-8595 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.714, + 0.785, + 0.755 + ], + "angle": 0, + "content": "28. Shazeer, N., Mirhoseini, A., Maziarz, K., Davis, A., Le, Q., Hinton, G., Dean, J.: Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "29. Sun, C., Sun, M., Chen, H.T.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5459-5469 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "30. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.61, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Boost Your NeRF" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "31. Wizadwongsa, S., Phongthawee, P., Yenphraphai, J., Suwajanakorn, S.: Nex: Real-time view synthesis with neural basis expansion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8534-8543 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.231 + ], + "angle": 0, + "content": "32. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time neural irradiance fields for free-viewpoint video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9421-9431 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.272 + ], + "angle": 0, + "content": "33. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoctrees for real-time rendering of neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5752-5761 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.314 + ], + "angle": 0, + "content": "34. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.355 + ], + "angle": 0, + "content": "35. Zhenxing, M., Xu, D.: Switch-nerf: Learning scene decomposition with mixture of experts for large-scale neural radiance fields. In: The Eleventh International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.397 + ], + "angle": 0, + "content": "36. Zhou, Y., Lei, T., Liu, H., Du, N., Huang, Y., Zhao, V., Dai, A.M., Le, Q.V., Laudon, J., et al.: Mixture-of-experts with expert choice routing. Advances in Neural Information Processing Systems 35, 7103-7114 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.425 + ], + "angle": 0, + "content": "37. Zhuang, Y., Zhu, H., Sun, X., Cao, X.: Mofanerf: Morphable facial neural radiance field. In: European conference on computer vision. pp. 268-285. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.425 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Boost Your NeRF_ A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering/cdbb2d54-c057-481d-a45f-f2c36fd59358_origin.pdf b/2024/Boost Your NeRF_ A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering/cdbb2d54-c057-481d-a45f-f2c36fd59358_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9b6d263892b0af542001918fc4ab6a017f721224 --- /dev/null +++ b/2024/Boost Your NeRF_ A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering/cdbb2d54-c057-481d-a45f-f2c36fd59358_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa350c955bd51e735ee729b8c03c210280dd3bfe056d56deee9e3dd90078398a +size 10148820 diff --git a/2024/Boost Your NeRF_ A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering/full.md b/2024/Boost Your NeRF_ A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b6be32337dd230067dfd5e96d7c0d98b8159fc63 --- /dev/null +++ b/2024/Boost Your NeRF_ A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering/full.md @@ -0,0 +1,299 @@ +# Boost Your NeRF: A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering + +Francesco Di Sario $^{1}$ [0009-0005-6969-1246], Riccardo Renzulli $^{1}$ [0000-0003-0532-5966], Enzo Tartaglione $^{2}$ [0000-0003-4274-8298], and Marco Grangetto $^{1}$ [0000-0002-2709-7864] + +1 University of Turin, Italy +2 LTCI, Télécom Paris, Institut Polytechnique de Paris, France +francesco.disario@unito.it + +Abstract. Since the introduction of NeRFs, considerable attention has been focused on improving their training and inference times, leading to the development of Fast-NeRFs models. Despite demonstrating impressive rendering speed and quality, the rapid convergence of such models poses challenges for further improving reconstruction quality. Common strategies to improve rendering quality involves augmenting model parameters or increasing the number of sampled points. However, these computationally intensive approaches encounter limitations in achieving significant quality enhancements. This study introduces a model-agnostic framework inspired by Sparsely-Gated Mixture of Experts to enhance rendering quality without escalating computational complexity. Our approach enables specialization in rendering different scene components by employing a mixture of experts with varying resolutions. We present a novel gate formulation designed to maximize expert capabilities and propose a resolution-based routing technique to effectively induce sparsity and decompose scenes. Our work significantly improves reconstruction quality while maintaining competitive performance. + +# 1 Introduction + +Neural Radiance Fields (NeRFs) [24] have recently shown impressive results in synthesizing photo-realistic 3D scenes from a set of 2D images. However, NeRFs suffer from limited scene diversity, long training time, and sensitivity to training data [21]. Since the introduction of NeRFs, significant attention has been directed toward improving their training and inference times, resulting in the development of Fast-NeRFs. By using auxiliary data structures such as voxel grids to store scene geometry and skip empty spaces, the training and inference process can be accelerated by several orders of magnitude. As a result, the neural component of these models, which is usually used to transform learned features into view-dependent color representations, becomes much smaller and can sometimes be replaced entirely by spherical harmonics [10]. Despite Fast + +![](images/af8de6ff0478d385dfb6d0336d7364361ddfeed921b911b58cb796a1bdd2990d.jpg) +Fig. 1: Different strategies for improving reconstruction quality with Fast-NeRFs (DVGO [29]). Increasing the resolution of data structures like voxel grids can improve render quality, but only up to a certain point, after which quality declines (left). The MLP component's impact on rendering quality is analyzed by varying its depth and width while keeping other variables constant (center). The effect of the number of sampled points along each ray is examined by decreasing the step size (right). In gray, our method's performance, which significantly improves PSNR at a low computational cost. + +![](images/5065657261c1d40643a7eb8e77731e04eac0762360d4e7e017cc81711548c4b4.jpg) + +![](images/e28add08a4a50158e527829cf420b945c72423bc1371faf9d592672e183b246d.jpg) + +NeRFs being an improvement from the original NeRF framework, achieving high-quality and efficient reconstructed geometry is still an open problem. + +Typical "naive" approaches to enhance the reconstruction quality of these models include: + +1. Increasing the parameters and resolution of the used data structures (e.g., voxel grid, hash grid, etc.). +2. Increasing the number of sampled points per ray. +3. Increasing the number of parameters in the neural network or the order of spherical harmonics. + +However, as evident from Figure 1, the increase in reconstruction quality results in a significant increase in computational costs. In the first approach, increasing the resolution can lead to a significant improvement in the reconstruction quality, but a plateau is reached beyond which overfitting occurs, and reconstruction quality degrades. This comes at the expense of a considerable increase in spatial complexity (especially with dense voxel grids) and training times. Increasing the number of sampled points per ray can marginally improve reconstruction quality further but at a significant increase in computational complexity. In fact, the higher the number of sampled rays, the higher forward passes through a neural network are needed. Augmenting the number of parameters in the neural network is another solution. However, as the neural network grows in depth and width, it tends towards a fully implicit model, deviating from the principles of fast models (limiting the neural part as much as possible - or even removing it altogether). + +We propose a technique capable of significantly enhancing the reconstruction quality of such models while maintaining competitive training and rendering times. Inspired by the Sparse Mixture of Experts (MoE) paradigm [28], we have developed a model-agnostic framework capable of improving the reconstruction quality of various state-of-the-art models. Intuitively, our mixture of experts consists of different-capacity (resolution) models. During training, each model + +specializes in rendering the most suitable parts of the scene, i.e. low-resolution models render low-frequency parts of the scene, while high-resolution models render high-frequency parts. Our formulation of the gate allows our method to be model-agnostic, enabling, on the one hand, the insertion of the gate function at the early stages of the MoE mechanism and, on the other hand, the multiplication of the gate's output with the experts' output as late as possible. This maximizes the capabilities of the mixture of experts and allows for working directly with the output of each expert, which is considered a black box. This is why our method is inherently model-agnostic. In summary, our contributions are as follows: + +1. We propose the first model-agnostic framework based on Sparse MoE of models at different resolutions, which significantly improves the rendering quality of such models while maintaining competitive training and inference times (Sec. 3). +2. We provide a novel gate formulation inspired by Fast-NeRF models, which maximizes the capability of the mixture of experts (Sec. 3.3). +3. We introduce a new routing technique based on resolution, favoring the assignment of tokens to low-resolution models and discouraging the assignment of tokens to high-resolution models, inducing increasing sparsity in high-resolution models and decomposing the scene based on frequency (Sec. 3.5). +4. We conduct extensive experiments to test our method, including different NeRFs architectures and datasets, showing higher rendering quality and efficiency (Sec. 4). + +# 2 Related works + +# 2.1 Neural Radiance Field + +NeRF [24] (Neural Radiance Field) has emerged as a prominent method for synthesizing novel views, showing significant progress. This approach requires a moderate number of input images along with their known camera poses. Unlike traditional methods that rely on explicit and discretized volumetric representations such as voxel grids and multiplane images, NeRF employs a coordinate-based multilayer perceptron (MLP) to create an implicit and continuous volumetric representation. NeRFs can represent a 3D scene as a MLP $F_{\theta}$ , with $\theta$ being the set of its trainable parameters, such that $F_{\theta}:(\mathbf{x},\mathbf{d})\to (\mathbf{c},\sigma)$ which maps $(\mathbf{x},\mathbf{d})$ , a 3D position $\mathbf{x}$ and a viewing direction $\mathbf{d}$ , to a view-dependent color emission $\mathbf{c}$ and density value $\sigma$ . To render the color of a pixel $\hat{C}(\mathbf{r})$ , a ray $\mathbf{r}$ traverses the center of the camera through the pixel of the image plane from an origin point $\mathbf{o}$ to a ray having position $\mathbf{r}(t) = \mathbf{o} + t\mathbf{d}$ . Then, $N$ points are sampled on $\mathbf{r}$ , and the MLP is queried for each point, obtaining a density and a color value. Finally, these results are accumulated into a single color with the volume rendering equation [22]: + +$$ +\hat {C} (\mathbf {r}) = \sum_ {i = 1} ^ {K} T _ {i} \alpha_ {i} c _ {i}, \quad T _ {i} = \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), \quad \alpha_ {i} = 1 - \exp \left(\sigma_ {i} \delta_ {i}\right), \tag {1} +$$ + +where $\alpha_{i}$ is the probability of termination at point $i$ and $T_{i}$ is the accumulated transmittance from the near plane to point $i$ . NeRFs are trained by minimizing a photometric loss, which is an L2 loss between the rendered and the ground truth pixels. In more detail, given a batch $B$ of randomly sampled rays, the loss is defined as + +$$ +L _ {\text {n e r f}} = \frac {1}{| B |} \sum_ {r \in B} \left\| \hat {C} (\mathbf {r}) - C (\mathbf {r}) \right\| _ {2} ^ {2}, \tag {2} +$$ + +where $\hat{C}(\mathbf{r})$ is the predicted color and $C(\mathbf{r})$ the ground truth color for the ray $\mathbf{r}$ . + +# 2.2 Fast Neural Radiance Fields + +Since the introduction of NeRF, significant effort has been directed towards the development of faster models. Several studies focus on speeding up rendering times, for example by working on ray sampling efficiency [3, 6, 12, 37], by integrating explicit volumetric representations [5, 11, 11, 15, 31-33] or by utilizing thousands of tiny MLPs [26]. However, all these methods still require large training times. One noteworthy development is represented by the introduction of explicit volumetric representations in the training pipeline, directly optimizing such representations [4, 9, 10, 13, 16, 25, 29]. These models enable fast training and inference, with render quality only slightly inferior to full-implicit models [1, 2]. We refer to these as Fast-NeRFs. Plenoxels [10] is the first important work in this context, as it represents a scene as a sparse 3D grid where each voxel stores spherical harmonic coefficients and density. Spherical harmonics serve as an orthogonal basis for functions defined over the sphere and thus can be used for computing view-dependent color emission, without the need for a multi-layer perceptron. Additionally, Plenoxels demonstrated the advantages of linearly interpolating voxels, facilitating the learning of a continuous plenoptic function throughout the volume, akin to NeRF, albeit with discrete data. Another notable work is DVGO [29]. Each scene is there represented as two dense voxel grids (one for density and one for feature colors) alongside a MLP, for learning view-dependent color. Similar to Plenoxels, the value of each voxel is linearly interpolated with the 8 nearest voxels, but after applying the activation functions. DVGO also incorporates a preliminary coarse geometry stage to learn the scene's general structure, facilitating adjustments to the bounding box and enabling a more intelligent ray sampling strategy. Despite utilizing lower-resolution models, DVGO achieves high reconstruction quality. The major drawback of these models is their substantial memory storage requirements. TensoRF [4] addresses this challenge by replacing the dense voxel grid with a planes and vectors decomposition, significantly reducing storage demands while maintaining comparable performance and rendering quality. Alternatively, Instant-NGP [25] proposed a multi-resolution voxel grid encoded via a hash function. They define $L$ hash grids of increased resolutions: each entry of each hash grid has $2^T$ parameters and $F$ features. This enables even faster training times and real-time rendering + +performances while maintaining a compact model; moreover, it demonstrates the efficacy of a multi-resolution approach in enhancing render quality. Inspired by this methodology, K-Planes [9] introduces a multi-resolution planar factorization of 3D space, offering reconstruction quality and model compactness similar to the previous methods, but with slightly larger training times. A similar multi-resolution planar factorization has also been proposed in Tri-MipRF [13] for mitigating the aliasing in distant or low-resolution views and blurriness in close-up shots. Given their properties, we decided to evaluate our paradigm using three different models: a 3D dense voxel grid-based model (DVGO), a decomposed grid-based model (TensoRF), and a multi-resolution hash grid-based model (Instant-NGP). + +# 2.3 Mixture of Experts and Sparse MoE + +The Mixture of Experts (MoE) paradigm [14] has gained prominence in various machine learning applications. MoE consists of multiple expert networks, each specializing in different regions of the input space, with a meta-network determining the contribution of each expert to the final prediction. Building upon the MoE framework, Sparse Mixture of Experts [19, 27, 28, 36] constitutes a scalable and efficient variant. At the core of all Sparse MoE algorithms lies an assignment problem between tokens and experts. One approach to tackle this is by approximating the solution with a gating function, which learns to assign input tokens to the most suitable experts. It typically comprises a linear layer, a softmax activation, and a Top- $K$ (where $1 \leq K \leq 2$ ) operation, aimed at routing the input token to only a subset of the experts. To balance the assignment across all the experts, auxiliary loss functions penalizing unevenly distributed routing are often employed. This sparsity-inducing technique significantly reduces computational overhead while preserving the expressive power of the MoE architecture. + +# 2.4 MoE and NeRF + +Combining the sparse mixture of experts' paradigms with neural radiance fields presents a non-trivial challenge. In recent research, Switch-NeRF [35] has been introduced as a novel end-to-end large-scale NeRF with learning-based scene decomposition. Inspired by Fedus et al. [8], they propose a full-implicit model with an MLP-based gate comprising 4 layers with 128 neurons and a Top-1 function. However, their architecture is ad-hoc, and suffers from extensive training times. This poses a challenge for us, as we aim to consider each expert as a black box, taking a point in space and a direction as input and outputting radiance and density values. Moreover, as the mixture comprises experts at different resolutions, we also aim to prioritize routing input tokens toward lower-resolution models and minimize the usage of high-resolution models. + +Our work intends to overcome these limitations. We design a novel gate formulation inspired by Fast-NeRFs, that guarantees fast convergence and better performances. We position the gate at the beginning of our MoE pipeline and postpone the multiplication of the gate output with expert predictions as much as + +![](images/235ba8e61c477ad6888196f5ebadba5868808d3226d41c575b74082eb7587470.jpg) +Fig. 2: A density field is used to compute density values for sampled points along a ray. A filtering step discards low-density points, routed through gating network $G$ for expert assignment. Top- $K$ experts compute radiance and density, aggregating and weighting these values by the corresponding gating probability to get the final values $\mathbf{c}_f$ and $\sigma_f$ . Volume rendering equation yields pixel colors, and joint optimization refines our resolution-weighted auxiliary loss, allowing for high-quality and efficient rendering. + +possible. This allows our framework to be entirely model-agnostic. Furthermore, our mixture of experts maintains competitive training times, enabling, in the worst case, training models on complex scenes in approximately 1 hour while achieving state-of-the-art accuracy. + +# 3 Method + +In this section, we introduce our model-agnostic assembling approach. We present in Sec. 3.1 an overview of our method. We also describe the ray sampling and filtering mechanism in Sec. 3.2. Then, in Sec. 3.3, we discuss the effective construction of a mixture of experts with a gating mechanism in the context of NeRF. Finally, in Sec. 3.5, we present a novel formulation for decomposing the scene into resolution-based parts with an end-to-end training procedure. + +# 3.1 Overview + +Figure 2 shows a high-level scheme of our method. After independently training a set of $M$ NeRF models $\{E_i\}_{i=1}^M$ at different resolutions, we sample $N$ points along a ray $\mathbf{r}$ . For each point $\mathbf{x}_p$ , a density value $\sigma$ is calculated, and based on this, a filtering step is performed based on the density volume $\mathbf{V}_D$ , discarding points in regions with a density below a certain threshold $T$ . Initializing the density volume $\mathbf{V}_D$ with the one learned from the lowest resolution model can be helpful (though it can also be learned). Subsequently, the filtered point $\mathbf{x}_f$ is fed to the gate $G$ , so we denote with $G(\mathbf{x}_f)$ the probabilities with which the point is assigned to the $M$ experts. Based on these values, each point is routed to the Top- $K$ experts. Each selected expert then computes the radiance $\mathbf{c}_{i,f}$ and density values $\sigma_{i,f}$ for the point, both of which are multiplied by their respective + +probabilities and summed together to get the final radiance $\mathbf{c}_f$ and density $\sigma_f$ . Subsequently, the volume rendering Equation 1 is used to aggregate all colors and compute the pixel color for the given ray. Finally, we compute the total loss and jointly optimize the gate and the experts. + +In the next sections, we will dive into the gating and sparse mixture of expert modules. + +# 3.2 Ray Sampling and Filtering + +The first phase of our method involves learning a coarse and explicit density field starting from the density volume $\mathbf{V}_D$ , that we can leverage for skipping empty spaces. Given a ray $\mathbf{r}$ as explained in Sec. 2.1, we sample $N$ points along it: + +$$ +\mathbf {X} _ {p, r} = \text {s a m p l i n g} (\mathbf {r}) \in \mathbb {R} ^ {N \times 3}, \tag {3} +$$ + +with $p \in [0, \dots, N - 1]$ . We suppress the index $r$ for abuse of notation. Next, we compute the density for each point $\mathbf{x}_p \in \mathbb{R}^3$ and discard those with negligible density. The density value $\sigma_{D,p}$ for each point is computed by linearly interpolating $v$ neighboring voxels: + +$$ +\sigma_ {D, p} = \operatorname {a c t} (\operatorname {i n t e r p o l a t e} \left(\mathbf {x} _ {p}, V _ {D}\right)) \in \mathbb {R} \tag {4} +$$ + +where act represents a density activation function (such as softmax). We denote $\mathbf{x}_{\mathbf{f}}$ with $\mathbf{x}_{\mathbf{f}} \in R^3$ a remaining point after the filtering operation. + +# 3.3 Trainable Gating Model + +Our gating mechanism incorporates a hybrid architecture: an explicit feature grid $\mathbf{V}_G$ and a shallow MLP. The gating mechanism can also be seen as a probability field. First, we compute per-point features: + +$$ +\mathbf {f e a t} = \operatorname {i n t e r p o l a t e} \left(\mathbf {x} _ {f}, \mathbf {V} _ {G}\right) \in \mathbb {R} ^ {C} \tag {5} +$$ + +where $C$ represents the number of channels of $\mathbf{V}_G$ . Subsequently, we transform each feature into a probability. We compute logits as + +$$ +\operatorname {l o g i t s} = \operatorname {M L P} (\mathbf {f e a t}) \in \mathbb {R} ^ {M} \tag {6} +$$ + +where $M$ is the number of experts in the mixture of experts. Finally, we apply a per-row softmax. Our gating function can be summarized as: + +$$ +G \left(\mathbf {x} _ {\mathbf {f}}\right) = \operatorname {s o f t m a x} (\operatorname {M L P} (\text {i n t e r p o l a t e} \left(\mathbf {x} _ {f}, \mathbf {V} _ {G}\right))) \in \mathbb {R} ^ {M} \tag {7} +$$ + +# 3.4 NeRFs Experts + +As mentioned in Sec. 3.1, we employ a set of $M$ pretrained NeRFs models as experts. After feeding as input each point $\mathbf{x}_f$ to the gate, we route them to the $k$ experts with the $k$ highest probabilities $G(\mathbf{x}_f)$ . We define this set of indexes of the selected experts as + +$$ +\mathcal {K} = \underset {T o p - k} {\operatorname {a r g m a x}} (G (\mathbf {x} _ {f})). \tag {8} +$$ + +Each expert is a Fast-NeRF model that receives the dispatched points and the direction of the ray they lie on as input and outputs a density and a radiance value. We treat each expert as a black box, ensuring our method is inherently model-agnostic. The radiance $\mathbf{c}_f$ and the density $\sigma_f$ for the point $\mathbf{x}_f$ laying on a ray with direction $\mathbf{d}$ can be expressed as the sum of each expert predictions weighted by their probability: + +$$ +\mathbf {c} _ {f}, \sigma_ {f} = \sum_ {i \in \mathcal {K}} E _ {i} \left(\mathbf {x} _ {\mathbf {f}}, \mathbf {d}\right) \cdot G \left(\mathbf {x} _ {\mathbf {f}}\right) _ {i} \tag {9} +$$ + +where $E_{i}$ denotes the $i$ -th expert and $G(\mathbf{x_f})_i$ the probability for the point to be dispatched to that expert $i$ . + +Once we obtain radiance and density values for all the points on the ray, we can compute the pixel color with Equation 1. + +# 3.5 Resolution-based Routing + +To balance the load and prevent the gate from focusing on assigning points to a single expert (typically the one with the highest resolution), we employ a resolution-weighted auxiliary loss. Given $M$ experts, and a batch of points $B$ , the auxiliary loss is defined as: + +$$ +L _ {\text {a u x}} = \frac {M}{| B | ^ {2}} \sum_ {i = 1} ^ {M} c _ {i} m _ {i}, \quad m _ {i} = \sum_ {\mathbf {x} _ {f} \in B} G (\mathbf {x} _ {f}) _ {i}, \tag {10} +$$ + +where $c_{i}$ represents the number of inputs dispatched to the expert $i$ , and $m_{i}$ is the sum of all the probabilities for each point in the batch for the expert $i$ . This loss helps balance the workload, ensuring that each expert processes a similar number of points. Ideally, in a perfectly balanced scenario, $L_{aux}$ is expected to be 1, as both $c_{i}$ and $m_{i}$ would be $\frac{M}{N}$ , resulting in $\sum_{i=1}^{M} c_{i} m_{i}$ being $\frac{M^2}{N}$ . However, in the case of total imbalance, it tends to the number $M$ . We aim to take a step further by assigning as many points as possible to low-resolution experts while discouraging point assignment to high-resolution models. Hence, we introduce some penalty terms $w_{i}$ associated with each expert. The higher the resolution of the model, the higher the penalty. We define a novel resolution-based auxiliary loss: + +$$ +L _ {\mathrm {r w - a u x}} = \frac {M}{| B | ^ {2}} \sum_ {i = 1} ^ {M} c _ {i} m _ {i} w _ {i}. \tag {11} +$$ + +As for the weighting strategy, we opted for the following geometric progression. The weight $w_{i}$ for the $i$ expert is computed as: + +$$ +w _ {i} = \exp \left(\frac {\ln M}{M - 1}\right) ^ {i}, \quad i \in [ 0, \dots , M - 1 ]. \tag {12} +$$ + +The total loss is then defined as: + +$$ +L _ {\mathrm {t o t}} = L _ {\mathrm {n e r f}} + \lambda L _ {\mathrm {r w - a u x}}. \tag {13} +$$ + +As we will see in Sec. 4, our proposed loss not only further improves the reconstruction quality but also encourages sparsity in higher-resolution models. A comparison with other weighting strategies is proposed in the Supplementary ?? + +# 4 Experiments + +# 4.1 Setting + +The code was written in Python 3.8, using PyTorch 1.12, and executed on a single NVIDIA A40 GPU. We tested our technique on the DVGO, TensoRF, and Instant-NGP architecture. The code from official repositories was used as the starting point. For Instant-NGP, the native implementation in PyTorch ngp_p1 was employed, which exhibits comparable performance and reconstruction quality to the official NVIDIA implementation. We experimented with various configurations, ranging from a minimum of 3 models to 5 models of different resolutions. The experts are ordered by resolution, such that the $i + 1$ -th expert has approximately double the parameters of the $i$ -th expert. For all configurations and models, $\lambda$ was set to $10^{-3}$ . The Gate consists of a grid (dense voxel grid for DVGO, factorized grid for TensoRF, and hash grid for Instant-NGP) and a shallow MLP (2 layers with 64 neurons each) with ReLU activation function. The resolution of the gate is low: $128^3$ for both DVGO and TensoRF, while for Instant-NGP we used $L = 6$ . The number of iterations is set to $20k$ for all the architectures. All other hyperparameters are left as the original implementations. We draw experiments with our method using Top- $k$ experts, with $k = 1$ and $k = 2$ . We compare our results versus baselines with comparable resolutions, as well as a Fast-NeRF ensemble (Ens) obtained by jointly fine-tuning all models and averaging their predictions. It can be noted that this ensemble can be interpreted as a limit case for our method with $k = M$ and $G(\mathbf{x}_f)_i = 1 / M$ , $\forall i$ similar to the method proposed by [7]. + +# 4.2 Metrics and Datasets + +For each test, image-quality metrics such as PSNR, SSIM [30], and LPIPS [34] (computed on AlexNet [18]) are presented. Additionally, we report the number of non-zero parameters as $\| w_0\|$ , the average GFLOPs required by each model to render the images of the test set, and total training times. We present results obtained across four major datasets, namely: Synthetic-NeRF [24], Neural Sparse Voxel Field Dataset (NSVF) [20], TanksAndTemple [17], and Local Light Field Fusion Dataset (LLFF) [23]. + +Table 1: Results on DVGO, TensoRF and Instant-NGP with $M = 5$ . $\| w\| _0$ is expressed as multiple of $10^{6}$ , while for Instant-NGP as a multiple of $10^{5}$ . + +
DatasetMetricsDVGOTensorRFInstant-NGP
baselineTop-1Top-2EnsbaselineTop-1Top-2EnsbaselineTop-1Top-2Ens
BlenderPSNR ↑33.0433.4333.7433.7932.9833.6834.0934.0033.3533.5633.8334.01
SSIM ↑0.9610.9640.9650.9660.9580.9650.9680.9680.9630.9630.9650.966
LPIPS ↓0.0260.0240.0220.0220.0290.0230.0210.0220.0250.0450.0430.042
||w||0 ↓992639974024334931172126
GFLOPs ↓6354999402206886732134422144671123208
Time ↓26'21'25'32'44'69'76'70'11'32'34'37'
NSVFPSNR ↑35.2137.1237.5937.6836.7037.4037.9838.0836.4433.5937.0437.31
SSIM ↑0.9770.9840.9860.9860.9810.9840.9860.9870.9830.9830.9840.985
LPIPS ↓0.0150.0090.0080.0070.0130.0090.0080.0080.0100.0230.0220.020
||w||0 ↓9527431004228385430172027
GFLOPs ↓564430811190370657510531763285272123
Time ↓12'22'25'30'46'75'75'75'10'31'33'34'
TaTPSNR ↑28.9329.1429.2729.3728.4428.7829.1429.1129.0729.1629.3229.38
SSIM ↑0.9270.9250.9290.9320.9050.9240.9290.9280.9240.9270.9290.930
LPIPS ↓0.1070.1080.1050.10300.1060.0990.1090.1010.1250.1240.121
||w||0 ↓7416266579158088374770
GFLOPs ↓266616263066719835672791512691462112295311003
Time ↓22'25'28'38'72'70'78'101'14'37'39'45'
LLFFPSNR ↑26.2426.4326.6226.6526.7126.7327.0927.1024.9724.9025.1725.19
SSIM ↑0.8310.8320.8390.8430.8350.8360.8620.8640.7640.7630.7770.778
LPIPS ↓0.1360.1150.1110.1070.1140.1110.1010.1010.1280.2390.2370.234
||w||0 ↓622640113191116231526875148
GFLOPs ↓16781514250849724542322659211352257388713912712
Time ↓24'28'32'36'58'49'57'68'24'25'46'42'
+ +# 4.3 Quantitative Results + +The main experimental results in terms of the metrics defined above are shown in Table 1. Here, we present the results obtained using a mixture of $M = 5$ experts; further configurations can be found in the supplementary Sec. ?? Our experiments reveal that our MoE provides a significant rendering quality improvement with respect to the baseline with no or limited impact in terms of computational cost. The increase in rendering quality is notable in Synthetic NeRF (up to 1 dB) and NSVF (up to 1.3 dB). Although more moderate, improvements are still evident, even in more challenging datasets such as TanksAndTemple and LLFF, with about 0.5 dB gain. While Top-1 can already achieve state-of-the-art render quality, the Top-2 strategy further enhances image quality, reaching levels comparable to the ensemble but with much greater efficiency (about half the average FLOPs per rendering) and using significantly fewer parameters. This observation is also illustrated in Figure 3 , where FLOPS/PSNR plots are shown (each marker in every curve refers to the cases with $M = 3,4,5$ experts respectively). Training times are on average are longer but still acceptable (in the worst-case scenario, around 1h is required to train our MoE). However, in the case of Top1, they can be faster than baselines. This is because of our resolution methods, which tend to favor low-resolution models. Based on these analyses, the Top-2 strategy strikes an excellent quality/cost trade-off. The ensembling configura + +![](images/f9550c86b308a27fc0fb872ad9021cf67ac989f7a485604387820576fd5030fb.jpg) + +![](images/e568527cb91f343b5ef103030064e80d5f334f78c6b4f9455a9202897cb0d82a.jpg) + +![](images/7cd19a8e4c18091a0f7e3b41fca58af0f5707f5d7c0c8a4f760f6dd06a21bb2e.jpg) + +![](images/69dafdc302de52af7289909dc5a53fa52dace5473cbb51e58fccb4ac78aa80a7.jpg) +(a) DVGO + +![](images/fd742c4c39cf2372498587d65a3a7167952c51b97931c2178c4a43167cc806dc.jpg) +(b) TensorRF + +![](images/892b5dcd660df7c1bf65c52e7c0813c3841a8918b524ed6ce7a22262a941f1db.jpg) +(c) Instant-NGP +Fig. 3: PSNR/GFLOPs plots for Synthetic-NeRF and TanksAndTemple. The remaining datasets show similar results. + +tion, using all experts, represents an upper bound in terms of image quality while being the worst case in computational terms. + +# 4.4 Qualitative Results + +Figure 4 shows visual comparisons among the baseline, ensemble, Top-1, and Top-2. From the selected image crops, one can appreciate that our method ensures superior reconstruction quality compared to the baselines, effectively reproducing sharper details while reducing noise on texture-less spots. This is particularly notable when examining elements such as the window decorations in the Palace scene or the text on the box above the desk in the Room scene. Additionally, surfaces such as the semi-transparent glass in the Wineholder or the Caterpillar scene appear less noisy and more faithful to the original. This improvement can be attributed to the synergy among different resolution models: low-resolution models excel at representing lower frequencies, thereby introducing less noise, while high-resolution models can focus solely on high-frequency components. Additionally, it is important to notice that there is no significant difference between the ensemble and Top-2. + +# 4.5 Comparison with Naive Methods + +As shown in Figure 1, our method leads to significantly higher quality reconstructions at a greatly reduced computational cost. Decreasing the step size yields the least noticeable improvement while incurring a substantial computational + +![](images/2d38a8a6f81d2ff182f38c17d027390221ce77d0fa37588021c595da6f3655b2.jpg) +Fig. 4: Qualitative results on some of the scenes of each dataset. From left to right: ground truths, baselines, Top-1, Top-2 and Ensemble. Each model has the same parameters and has been trained for the same number of iterations (DVGO). + +expense, as each sampled point requires an evaluation by the color decoder. Our method achieves high-quality reconstructions with the same step size as the baseline models. Increasing the MLP parameters can boost accuracy, but again, the computational costs become substantial compared to our method. Similarly, increasing resolution results in way inferior PSNR with respect to our method at a comparable computational cost. Our method also allows scaling to higher resolutions, while baseline models tend to introduce noise and artifacts, leading to a decrease in reconstruction quality as the resolution increases. + +# 4.6 Gate Visualization and Scene Decomposition + +In Figure 5, we visualize the gate (probability field in grayscale) and each expert output in the case $M = 3$ . On the right side of the figure, one can appreciate per-model renders and experts' specialization. It is worth noting how higher-resolution experts render high-frequency details. This is particularly evident in the Mic scene. + +# 4.7 Ablation + +Here we investigate how different design choices of our model affect the performance and rendering quality. These choices include the gate resolution, different gate formulations, and the number of experts in the Top- $k$ . + +![](images/08a346b4472e12ae5271af5ff9628e37dfb6f56de18ed3137174d89ffa77872c.jpg) +Fig. 5: Gate (gray-scale images) and per-model output visualizations with 3 experts and resolution-based routing (DVGO). Images are ordered by increasing resolutions. + +Here we investigate how different design choices of our model affect the performance and rendering quality. These choices include the gate resolution, different gate formulations, and the number of experts in the Top- $k$ . + +Gate Resolution and Gate Formulation. We show in Table 2 that a low-resolution gate is sufficient to achieve high-quality rendering. Interestingly, as the gate resolution increases, there is a slight decrease in rendering quality during testing, coupled with an increase in required FLOPs per image rendering and number of parameters. We also compare various gating strategies, including linear gating, the configuration proposed by Switch-NeRF, and our gate formulation. Through experimentation, it becomes evident that our approach outperforms others in terms of both render quality and performance, achieving performances comparable to a linear gating function. + +Why Top-2? Here we investigate the performance and rendering quality trends with varying values of $k$ . In Figure 6, we can see that while $k = 1$ can lead to a significant increase in quality with comparable performance to baseline models, the Top-2 further enhances quality at the expense of increased (but still acceptable) computation. The Top-3, Top-4, and Top-5 provide marginal quality improvements at a significantly higher computational cost. Hence, we consider the Top-2 the optimal balance between performance and quality. + +
ResGate TypePSNR||w||0GFLOPs
1283Ours36.826476
2563Ours36.7733601
3003Ours36.7942677
-Linear36.2524444
-Switch-NeRF36.44251095
+ +Table 2: Comparison of different gate resolutions and formulations on Lego scene and DVGO. + +![](images/044d1bb3a35c35583d4c3b24f456b2581b24df7d67eeed7e84f9e961634af261.jpg) +Fig. 6: Comparison of Gating Functions (Top-1 to Top-5) on Synthetic-NeRF with DVGO. + +# 5 Limitations + +Our study present some limitations. First of all, a pre-training phase, on which each model is trained independently is required for achieving good results. Training end-to-end, without a pre-training phase, can lead to reconstruction quality that is noticeably inferior to baselines. Pre-training models at different resolutions allows for diversified architectures, making it easier for the gate to learn the decomposition. The training is also strongly dependent on the auxiliary loss. Different values of $\lambda$ can significantly influence performances and load balancing. We conducted a sweep to identify a value that works well across many scenarios, but it may not be suitable for different datasets. Another limitation is the overhead introduced by the MoE. Each input token is first interpolated with the gate's grid, decoded into probabilities and routed to the chosen expert. These operations can significantly slow-down the rendering process, leading to higher training and inference times. Although considerable effort was devoted to developing the most efficient gate possible, our MoE is still slower than the respective baselines. + +# 6 Conclusions + +In this paper, we introduced a model-agnostic framework for enhancing the rendering of Fast-NeRF models. Our formulation of the Gate reduces computational costs in both training and inference phases while ensuring better quality compared to a traditional gate. Additionally, the introduction of an auxiliary loss with res penalty allows for increased utilization of low-resolution models, reducing the number of active parameters and promoting sparsity in higher-resolution models. Our results demonstrate how this approach can significantly improve reconstruction quality while considering performance metrics. Specifically, we show that a Top-2 strategy strikes a good balance between performance and quality. + +# Acknowledgements + +This work was partially funded by Hi!PARIS Center on Data Analytics and Artificial Intelligence. + +# References + +1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021) +2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706 (2023) +3. Chan, E.R., Monteiro, M., Kellnhofer, P., Wu, J., Wetzstein, G.: pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5799-5809 (2021) +4. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. ECCV (2022) +5. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16569-16578 (2023) +6. Deng, Y., Yang, J., Xiang, J., Tong, X.: Gram: Generative radiance manifolds for 3d-aware image generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10673-10683 (2022) +7. Di Sario, F., Renzulli, R., Tartaglione, E., Grangetto, M.: Two is better than one: Achieving high-quality 3d scene modeling with a nerf ensemble. In: International Conference on Image Analysis and Processing. pp. 320-331. Springer (2023) +8. Fedus, W., Zoph, B., Shazeer, N.: Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. The Journal of Machine Learning Research 23(1), 5232-5270 (2022) +9. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes for radiance fields in space, time, and appearance (2023) +0. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022) +1. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High-fidelity neural rendering at 200fps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14346-14355 (2021) +2. Hou, A., Liu, F., Ren, Z., Sarkis, M., Bi, N., Tong, Y., Liu, X.: Infamous-nerf: Improving face modeling using semantically-aligned hypernetworks with neural radiance fields. arXiv preprint arXiv:2312.16197 (2023) +3. Hu, W., Wang, Y., Ma, L., Yang, B., Gao, L., Liu, X., Ma, Y.: Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19774-19783 (2023) + +14. Jacobs, R.A., Jordan, M.I., Nowlan, S.J., Hinton, G.E.: Adaptive mixtures of local experts. Neural computation 3(1), 79-87 (1991) +15. Jeong, Y., Ahn, S., Choy, C., Anandkumar, A., Cho, M., Park, J.: Self-calibrating neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5846-5854 (2021) +16. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023) +17. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG) 36(4), 1-13 (2017) +18. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25 (2012) +19. Lepikhin, D., Lee, H., Xu, Y., Chen, D., First, O., Huang, Y., Krikun, M., Shazeer, N., Gshard, Z.: Scaling giant models with conditional computation and automatic sharding. arXiv preprint arXiv:2006.16668 (2020) +20. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020) +21. Liu, Y., Li, X., Yu, F., Zhou, Q.: Probabilistic neural scene representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021) +22. Max, N.: Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics 1(2), 99-108 (1995) +23. Mildenhall, B., Srinivasan, P.P., Ortiz-Cayon, R., Kalantari, N.K., Ramamoorthi, R., Ng, R., Kar, A.: Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG) 38(4), 1-14 (2019) +24. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European conference on computer vision. pp. 405-421. Springer (2020) +25. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022) +26. Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14335-14345 (2021) +27. Riquelme, C., Puigcerver, J., Mustafa, B., Neumann, M., Jenatton, R., Susano Pinto, A., Keysers, D., Houlsby, N.: Scaling vision with sparse mixture of experts. Advances in Neural Information Processing Systems 34, 8583-8595 (2021) +28. Shazeer, N., Mirhoseini, A., Maziarz, K., Davis, A., Le, Q., Hinton, G., Dean, J.: Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538 (2017) +29. Sun, C., Sun, M., Chen, H.T.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5459-5469 (2022) +30. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004) + +31. Wizadwongsa, S., Phongthawee, P., Yenphraphai, J., Suwajanakorn, S.: Nex: Real-time view synthesis with neural basis expansion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8534-8543 (2021) +32. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time neural irradiance fields for free-viewpoint video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9421-9431 (2021) +33. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoctrees for real-time rendering of neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5752-5761 (2021) +34. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018) +35. Zhenxing, M., Xu, D.: Switch-nerf: Learning scene decomposition with mixture of experts for large-scale neural radiance fields. In: The Eleventh International Conference on Learning Representations (2022) +36. Zhou, Y., Lei, T., Liu, H., Du, N., Huang, Y., Zhao, V., Dai, A.M., Le, Q.V., Laudon, J., et al.: Mixture-of-experts with expert choice routing. Advances in Neural Information Processing Systems 35, 7103-7114 (2022) +37. Zhuang, Y., Zhu, H., Sun, X., Cao, X.: Mofanerf: Morphable facial neural radiance field. In: European conference on computer vision. pp. 268-285. Springer (2022) \ No newline at end of file diff --git a/2024/Boost Your NeRF_ A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering/images.zip b/2024/Boost Your NeRF_ A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5d16d8029447d22098b3f1b9ee58a1183ffe9b7d --- /dev/null +++ b/2024/Boost Your NeRF_ A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:325d209acfbe43288fe524a264135c19cb60b1f5ea3e114e8b11833fd35ba2da +size 547037 diff --git a/2024/Boost Your NeRF_ A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering/layout.json b/2024/Boost Your NeRF_ A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..25f4123d7e66c78ac04fe2ca32b54a1d601df62d --- /dev/null +++ b/2024/Boost Your NeRF_ A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering/layout.json @@ -0,0 +1,9233 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 111, + 479, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 111, + 479, + 166 + ], + "spans": [ + { + "bbox": [ + 136, + 111, + 479, + 166 + ], + "type": "text", + "content": "Boost Your NeRF: A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 211, + 185, + 402, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 185, + 402, + 234 + ], + "spans": [ + { + "bbox": [ + 211, + 185, + 402, + 234 + ], + "type": "text", + "content": "Francesco Di Sario" + }, + { + "bbox": [ + 211, + 185, + 402, + 234 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 211, + 185, + 402, + 234 + ], + "type": "text", + "content": "[0009-0005-6969-1246], Riccardo Renzulli" + }, + { + "bbox": [ + 211, + 185, + 402, + 234 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 211, + 185, + 402, + 234 + ], + "type": "text", + "content": "[0000-0003-0532-5966], Enzo Tartaglione" + }, + { + "bbox": [ + 211, + 185, + 402, + 234 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 211, + 185, + 402, + 234 + ], + "type": "text", + "content": "[0000-0003-4274-8298], and Marco Grangetto" + }, + { + "bbox": [ + 211, + 185, + 402, + 234 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 211, + 185, + 402, + 234 + ], + "type": "text", + "content": "[0000-0002-2709-7864]" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 177, + 243, + 435, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 243, + 435, + 277 + ], + "spans": [ + { + "bbox": [ + 177, + 243, + 435, + 277 + ], + "type": "text", + "content": "1 University of Turin, Italy \n2 LTCI, Télécom Paris, Institut Polytechnique de Paris, France \nfrancesco.disario@unito.it" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 304, + 455, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 304, + 455, + 491 + ], + "spans": [ + { + "bbox": [ + 160, + 304, + 455, + 491 + ], + "type": "text", + "content": "Abstract. Since the introduction of NeRFs, considerable attention has been focused on improving their training and inference times, leading to the development of Fast-NeRFs models. Despite demonstrating impressive rendering speed and quality, the rapid convergence of such models poses challenges for further improving reconstruction quality. Common strategies to improve rendering quality involves augmenting model parameters or increasing the number of sampled points. However, these computationally intensive approaches encounter limitations in achieving significant quality enhancements. This study introduces a model-agnostic framework inspired by Sparsely-Gated Mixture of Experts to enhance rendering quality without escalating computational complexity. Our approach enables specialization in rendering different scene components by employing a mixture of experts with varying resolutions. We present a novel gate formulation designed to maximize expert capabilities and propose a resolution-based routing technique to effectively induce sparsity and decompose scenes. Our work significantly improves reconstruction quality while maintaining competitive performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 510, + 230, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 510, + 230, + 522 + ], + "spans": [ + { + "bbox": [ + 133, + 510, + 230, + 522 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRFs) [24] have recently shown impressive results in synthesizing photo-realistic 3D scenes from a set of 2D images. However, NeRFs suffer from limited scene diversity, long training time, and sensitivity to training data [21]. Since the introduction of NeRFs, significant attention has been directed toward improving their training and inference times, resulting in the development of Fast-NeRFs. By using auxiliary data structures such as voxel grids to store scene geometry and skip empty spaces, the training and inference process can be accelerated by several orders of magnitude. As a result, the neural component of these models, which is usually used to transform learned features into view-dependent color representations, becomes much smaller and can sometimes be replaced entirely by spherical harmonics [10]. Despite Fast" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 116, + 247, + 192 + ], + "blocks": [ + { + "bbox": [ + 140, + 116, + 247, + 192 + ], + "lines": [ + { + "bbox": [ + 140, + 116, + 247, + 192 + ], + "spans": [ + { + "bbox": [ + 140, + 116, + 247, + 192 + ], + "type": "image", + "image_path": "af8de6ff0478d385dfb6d0336d7364361ddfeed921b911b58cb796a1bdd2990d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 203, + 482, + 290 + ], + "lines": [ + { + "bbox": [ + 130, + 203, + 482, + 290 + ], + "spans": [ + { + "bbox": [ + 130, + 203, + 482, + 290 + ], + "type": "text", + "content": "Fig. 1: Different strategies for improving reconstruction quality with Fast-NeRFs (DVGO [29]). Increasing the resolution of data structures like voxel grids can improve render quality, but only up to a certain point, after which quality declines (left). The MLP component's impact on rendering quality is analyzed by varying its depth and width while keeping other variables constant (center). The effect of the number of sampled points along each ray is examined by decreasing the step size (right). In gray, our method's performance, which significantly improves PSNR at a low computational cost." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 253, + 116, + 360, + 192 + ], + "blocks": [ + { + "bbox": [ + 253, + 116, + 360, + 192 + ], + "lines": [ + { + "bbox": [ + 253, + 116, + 360, + 192 + ], + "spans": [ + { + "bbox": [ + 253, + 116, + 360, + 192 + ], + "type": "image", + "image_path": "5065657261c1d40643a7eb8e77731e04eac0762360d4e7e017cc81711548c4b4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 366, + 116, + 475, + 192 + ], + "blocks": [ + { + "bbox": [ + 366, + 116, + 475, + 192 + ], + "lines": [ + { + "bbox": [ + 366, + 116, + 475, + 192 + ], + "spans": [ + { + "bbox": [ + 366, + 116, + 475, + 192 + ], + "type": "image", + "image_path": "e28add08a4a50158e527829cf420b945c72423bc1371faf9d592672e183b246d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 314, + 479, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 479, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 479, + 338 + ], + "type": "text", + "content": "NeRFs being an improvement from the original NeRF framework, achieving high-quality and efficient reconstructed geometry is still an open problem." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 338, + 479, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 479, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 479, + 361 + ], + "type": "text", + "content": "Typical \"naive\" approaches to enhance the reconstruction quality of these models include:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 135, + 370, + 480, + 429 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 135, + 370, + 479, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 370, + 479, + 394 + ], + "spans": [ + { + "bbox": [ + 135, + 370, + 479, + 394 + ], + "type": "text", + "content": "1. Increasing the parameters and resolution of the used data structures (e.g., voxel grid, hash grid, etc.)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 135, + 394, + 365, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 394, + 365, + 405 + ], + "spans": [ + { + "bbox": [ + 135, + 394, + 365, + 405 + ], + "type": "text", + "content": "2. Increasing the number of sampled points per ray." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 135, + 405, + 480, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 405, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 135, + 405, + 480, + 429 + ], + "type": "text", + "content": "3. Increasing the number of parameters in the neural network or the order of spherical harmonics." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 437, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 594 + ], + "type": "text", + "content": "However, as evident from Figure 1, the increase in reconstruction quality results in a significant increase in computational costs. In the first approach, increasing the resolution can lead to a significant improvement in the reconstruction quality, but a plateau is reached beyond which overfitting occurs, and reconstruction quality degrades. This comes at the expense of a considerable increase in spatial complexity (especially with dense voxel grids) and training times. Increasing the number of sampled points per ray can marginally improve reconstruction quality further but at a significant increase in computational complexity. In fact, the higher the number of sampled rays, the higher forward passes through a neural network are needed. Augmenting the number of parameters in the neural network is another solution. However, as the neural network grows in depth and width, it tends towards a fully implicit model, deviating from the principles of fast models (limiting the neural part as much as possible - or even removing it altogether)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "We propose a technique capable of significantly enhancing the reconstruction quality of such models while maintaining competitive training and rendering times. Inspired by the Sparse Mixture of Experts (MoE) paradigm [28], we have developed a model-agnostic framework capable of improving the reconstruction quality of various state-of-the-art models. Intuitively, our mixture of experts consists of different-capacity (resolution) models. During training, each model" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "type": "text", + "content": "F. Di Sario et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": "specializes in rendering the most suitable parts of the scene, i.e. low-resolution models render low-frequency parts of the scene, while high-resolution models render high-frequency parts. Our formulation of the gate allows our method to be model-agnostic, enabling, on the one hand, the insertion of the gate function at the early stages of the MoE mechanism and, on the other hand, the multiplication of the gate's output with the experts' output as late as possible. This maximizes the capabilities of the mixture of experts and allows for working directly with the output of each expert, which is considered a black box. This is why our method is inherently model-agnostic. In summary, our contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 228, + 480, + 383 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 136, + 228, + 480, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 228, + 480, + 276 + ], + "spans": [ + { + "bbox": [ + 136, + 228, + 480, + 276 + ], + "type": "text", + "content": "1. We propose the first model-agnostic framework based on Sparse MoE of models at different resolutions, which significantly improves the rendering quality of such models while maintaining competitive training and inference times (Sec. 3)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 276, + 480, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 276, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 136, + 276, + 480, + 300 + ], + "type": "text", + "content": "2. We provide a novel gate formulation inspired by Fast-NeRF models, which maximizes the capability of the mixture of experts (Sec. 3.3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 300, + 480, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 300, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 136, + 300, + 480, + 346 + ], + "type": "text", + "content": "3. We introduce a new routing technique based on resolution, favoring the assignment of tokens to low-resolution models and discouraging the assignment of tokens to high-resolution models, inducing increasing sparsity in high-resolution models and decomposing the scene based on frequency (Sec. 3.5)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 346, + 480, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 346, + 480, + 383 + ], + "spans": [ + { + "bbox": [ + 136, + 346, + 480, + 383 + ], + "type": "text", + "content": "4. We conduct extensive experiments to test our method, including different NeRFs architectures and datasets, showing higher rendering quality and efficiency (Sec. 4)." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 399, + 239, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 399, + 239, + 410 + ], + "spans": [ + { + "bbox": [ + 132, + 399, + 239, + 410 + ], + "type": "text", + "content": "2 Related works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 421, + 273, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 421, + 273, + 432 + ], + "spans": [ + { + "bbox": [ + 132, + 421, + 273, + 432 + ], + "type": "text", + "content": "2.1 Neural Radiance Field" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "text", + "content": "NeRF [24] (Neural Radiance Field) has emerged as a prominent method for synthesizing novel views, showing significant progress. This approach requires a moderate number of input images along with their known camera poses. Unlike traditional methods that rely on explicit and discretized volumetric representations such as voxel grids and multiplane images, NeRF employs a coordinate-based multilayer perceptron (MLP) to create an implicit and continuous volumetric representation. NeRFs can represent a 3D scene as a MLP " + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "text", + "content": " being the set of its trainable parameters, such that " + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "inline_equation", + "content": "F_{\\theta}:(\\mathbf{x},\\mathbf{d})\\to (\\mathbf{c},\\sigma)" + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "text", + "content": " which maps " + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "inline_equation", + "content": "(\\mathbf{x},\\mathbf{d})" + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "text", + "content": ", a 3D position " + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "text", + "content": " and a viewing direction " + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "text", + "content": ", to a view-dependent color emission " + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "text", + "content": " and density value " + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "text", + "content": ". To render the color of a pixel " + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "inline_equation", + "content": "\\hat{C}(\\mathbf{r})" + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "text", + "content": ", a ray " + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "text", + "content": " traverses the center of the camera through the pixel of the image plane from an origin point " + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "text", + "content": " to a ray having position " + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}" + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "text", + "content": " points are sampled on " + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 130, + 438, + 480, + 619 + ], + "type": "text", + "content": ", and the MLP is queried for each point, obtaining a density and a color value. Finally, these results are accumulated into a single color with the volume rendering equation [22]:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 167, + 635, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 635, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 167, + 635, + 480, + 667 + ], + "type": "interline_equation", + "content": "\\hat {C} (\\mathbf {r}) = \\sum_ {i = 1} ^ {K} T _ {i} \\alpha_ {i} c _ {i}, \\quad T _ {i} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\quad \\alpha_ {i} = 1 - \\exp \\left(\\sigma_ {i} \\delta_ {i}\\right), \\tag {1}", + "image_path": "f8bf80318dae299866953c4760ebc1f3cb72812cea00ccc2da4d96e8fb8973a9.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 373, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 373, + 91, + 447, + 100 + ], + "type": "text", + "content": "Boost Your NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " is the probability of termination at point " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " is the accumulated transmittance from the near plane to point " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": ". NeRFs are trained by minimizing a photometric loss, which is an L2 loss between the rendered and the ground truth pixels. In more detail, given a batch " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " of randomly sampled rays, the loss is defined as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 236, + 185, + 481, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 185, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 236, + 185, + 481, + 213 + ], + "type": "interline_equation", + "content": "L _ {\\text {n e r f}} = \\frac {1}{| B |} \\sum_ {r \\in B} \\left\\| \\hat {C} (\\mathbf {r}) - C (\\mathbf {r}) \\right\\| _ {2} ^ {2}, \\tag {2}", + "image_path": "b1a8ccb91779304d788d1465923a821e72afd0179c86ee09bd814a8ba679aaab.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 220, + 481, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 481, + 245 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 220, + 481, + 245 + ], + "type": "inline_equation", + "content": "\\hat{C}(\\mathbf{r})" + }, + { + "bbox": [ + 130, + 220, + 481, + 245 + ], + "type": "text", + "content": " is the predicted color and " + }, + { + "bbox": [ + 130, + 220, + 481, + 245 + ], + "type": "inline_equation", + "content": "C(\\mathbf{r})" + }, + { + "bbox": [ + 130, + 220, + 481, + 245 + ], + "type": "text", + "content": " the ground truth color for the ray " + }, + { + "bbox": [ + 130, + 220, + 481, + 245 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 130, + 220, + 481, + 245 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 262, + 303, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 262, + 303, + 274 + ], + "spans": [ + { + "bbox": [ + 131, + 262, + 303, + 274 + ], + "type": "text", + "content": "2.2 Fast Neural Radiance Fields" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 282, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 668 + ], + "type": "text", + "content": "Since the introduction of NeRF, significant effort has been directed towards the development of faster models. Several studies focus on speeding up rendering times, for example by working on ray sampling efficiency [3, 6, 12, 37], by integrating explicit volumetric representations [5, 11, 11, 15, 31-33] or by utilizing thousands of tiny MLPs [26]. However, all these methods still require large training times. One noteworthy development is represented by the introduction of explicit volumetric representations in the training pipeline, directly optimizing such representations [4, 9, 10, 13, 16, 25, 29]. These models enable fast training and inference, with render quality only slightly inferior to full-implicit models [1, 2]. We refer to these as Fast-NeRFs. Plenoxels [10] is the first important work in this context, as it represents a scene as a sparse 3D grid where each voxel stores spherical harmonic coefficients and density. Spherical harmonics serve as an orthogonal basis for functions defined over the sphere and thus can be used for computing view-dependent color emission, without the need for a multi-layer perceptron. Additionally, Plenoxels demonstrated the advantages of linearly interpolating voxels, facilitating the learning of a continuous plenoptic function throughout the volume, akin to NeRF, albeit with discrete data. Another notable work is DVGO [29]. Each scene is there represented as two dense voxel grids (one for density and one for feature colors) alongside a MLP, for learning view-dependent color. Similar to Plenoxels, the value of each voxel is linearly interpolated with the 8 nearest voxels, but after applying the activation functions. DVGO also incorporates a preliminary coarse geometry stage to learn the scene's general structure, facilitating adjustments to the bounding box and enabling a more intelligent ray sampling strategy. Despite utilizing lower-resolution models, DVGO achieves high reconstruction quality. The major drawback of these models is their substantial memory storage requirements. TensoRF [4] addresses this challenge by replacing the dense voxel grid with a planes and vectors decomposition, significantly reducing storage demands while maintaining comparable performance and rendering quality. Alternatively, Instant-NGP [25] proposed a multi-resolution voxel grid encoded via a hash function. They define " + }, + { + "bbox": [ + 130, + 282, + 482, + 668 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 282, + 482, + 668 + ], + "type": "text", + "content": " hash grids of increased resolutions: each entry of each hash grid has " + }, + { + "bbox": [ + 130, + 282, + 482, + 668 + ], + "type": "inline_equation", + "content": "2^T" + }, + { + "bbox": [ + 130, + 282, + 482, + 668 + ], + "type": "text", + "content": " parameters and " + }, + { + "bbox": [ + 130, + 282, + 482, + 668 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 282, + 482, + 668 + ], + "type": "text", + "content": " features. This enables even faster training times and real-time rendering" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 236, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 236, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 236, + 101 + ], + "type": "text", + "content": "F. Di Sario et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "performances while maintaining a compact model; moreover, it demonstrates the efficacy of a multi-resolution approach in enhancing render quality. Inspired by this methodology, K-Planes [9] introduces a multi-resolution planar factorization of 3D space, offering reconstruction quality and model compactness similar to the previous methods, but with slightly larger training times. A similar multi-resolution planar factorization has also been proposed in Tri-MipRF [13] for mitigating the aliasing in distant or low-resolution views and blurriness in close-up shots. Given their properties, we decided to evaluate our paradigm using three different models: a 3D dense voxel grid-based model (DVGO), a decomposed grid-based model (TensoRF), and a multi-resolution hash grid-based model (Instant-NGP)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 264, + 346, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 264, + 346, + 277 + ], + "spans": [ + { + "bbox": [ + 131, + 264, + 346, + 277 + ], + "type": "text", + "content": "2.3 Mixture of Experts and Sparse MoE" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 283, + 482, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 451 + ], + "type": "text", + "content": "The Mixture of Experts (MoE) paradigm [14] has gained prominence in various machine learning applications. MoE consists of multiple expert networks, each specializing in different regions of the input space, with a meta-network determining the contribution of each expert to the final prediction. Building upon the MoE framework, Sparse Mixture of Experts [19, 27, 28, 36] constitutes a scalable and efficient variant. At the core of all Sparse MoE algorithms lies an assignment problem between tokens and experts. One approach to tackle this is by approximating the solution with a gating function, which learns to assign input tokens to the most suitable experts. It typically comprises a linear layer, a softmax activation, and a Top- " + }, + { + "bbox": [ + 130, + 283, + 482, + 451 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 283, + 482, + 451 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 130, + 283, + 482, + 451 + ], + "type": "inline_equation", + "content": "1 \\leq K \\leq 2" + }, + { + "bbox": [ + 130, + 283, + 482, + 451 + ], + "type": "text", + "content": ") operation, aimed at routing the input token to only a subset of the experts. To balance the assignment across all the experts, auxiliary loss functions penalizing unevenly distributed routing are often employed. This sparsity-inducing technique significantly reduces computational overhead while preserving the expressive power of the MoE architecture." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 467, + 242, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 467, + 242, + 479 + ], + "spans": [ + { + "bbox": [ + 131, + 467, + 242, + 479 + ], + "type": "text", + "content": "2.4 MoE and NeRF" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 617 + ], + "type": "text", + "content": "Combining the sparse mixture of experts' paradigms with neural radiance fields presents a non-trivial challenge. In recent research, Switch-NeRF [35] has been introduced as a novel end-to-end large-scale NeRF with learning-based scene decomposition. Inspired by Fedus et al. [8], they propose a full-implicit model with an MLP-based gate comprising 4 layers with 128 neurons and a Top-1 function. However, their architecture is ad-hoc, and suffers from extensive training times. This poses a challenge for us, as we aim to consider each expert as a black box, taking a point in space and a direction as input and outputting radiance and density values. Moreover, as the mixture comprises experts at different resolutions, we also aim to prioritize routing input tokens toward lower-resolution models and minimize the usage of high-resolution models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "text", + "content": "Our work intends to overcome these limitations. We design a novel gate formulation inspired by Fast-NeRFs, that guarantees fast convergence and better performances. We position the gate at the beginning of our MoE pipeline and postpone the multiplication of the gate output with expert predictions as much as" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 373, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 373, + 91, + 447, + 101 + ], + "type": "text", + "content": "Boost Your NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 480, + 236 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 480, + 236 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 480, + 236 + ], + "type": "image", + "image_path": "235ba8e61c477ad6888196f5ebadba5868808d3226d41c575b74082eb7587470.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 243, + 482, + 312 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 312 + ], + "type": "text", + "content": "Fig. 2: A density field is used to compute density values for sampled points along a ray. A filtering step discards low-density points, routed through gating network " + }, + { + "bbox": [ + 130, + 243, + 482, + 312 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 243, + 482, + 312 + ], + "type": "text", + "content": " for expert assignment. Top- " + }, + { + "bbox": [ + 130, + 243, + 482, + 312 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 243, + 482, + 312 + ], + "type": "text", + "content": " experts compute radiance and density, aggregating and weighting these values by the corresponding gating probability to get the final values " + }, + { + "bbox": [ + 130, + 243, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_f" + }, + { + "bbox": [ + 130, + 243, + 482, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 243, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\sigma_f" + }, + { + "bbox": [ + 130, + 243, + 482, + 312 + ], + "type": "text", + "content": ". Volume rendering equation yields pixel colors, and joint optimization refines our resolution-weighted auxiliary loss, allowing for high-quality and efficient rendering." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 334, + 482, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 482, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 482, + 382 + ], + "type": "text", + "content": "possible. This allows our framework to be entirely model-agnostic. Furthermore, our mixture of experts maintains competitive training times, enabling, in the worst case, training models on complex scenes in approximately 1 hour while achieving state-of-the-art accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 399, + 202, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 399, + 202, + 412 + ], + "spans": [ + { + "bbox": [ + 132, + 399, + 202, + 412 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 425, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 498 + ], + "type": "text", + "content": "In this section, we introduce our model-agnostic assembling approach. We present in Sec. 3.1 an overview of our method. We also describe the ray sampling and filtering mechanism in Sec. 3.2. Then, in Sec. 3.3, we discuss the effective construction of a mixture of experts with a gating mechanism in the context of NeRF. Finally, in Sec. 3.5, we present a novel formulation for decomposing the scene into resolution-based parts with an end-to-end training procedure." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 514, + 209, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 209, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 209, + 525 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": "Figure 2 shows a high-level scheme of our method. After independently training a set of " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " NeRF models " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{E_i\\}_{i=1}^M" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " at different resolutions, we sample " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " points along a ray " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": ". For each point " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_p" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": ", a density value " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " is calculated, and based on this, a filtering step is performed based on the density volume " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_D" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": ", discarding points in regions with a density below a certain threshold " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": ". Initializing the density volume " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_D" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " with the one learned from the lowest resolution model can be helpful (though it can also be learned). Subsequently, the filtered point " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_f" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " is fed to the gate " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": ", so we denote with " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "G(\\mathbf{x}_f)" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " the probabilities with which the point is assigned to the " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " experts. Based on these values, each point is routed to the Top-" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " experts. Each selected expert then computes the radiance " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{i,f}" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " and density values " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\sigma_{i,f}" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " for the point, both of which are multiplied by their respective" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "type": "text", + "content": "F. Di Sario et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "probabilities and summed together to get the final radiance " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_f" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " and density " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\sigma_f" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": ". Subsequently, the volume rendering Equation 1 is used to aggregate all colors and compute the pixel color for the given ray. Finally, we compute the total loss and jointly optimize the gate and the experts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 165, + 479, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 165, + 479, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 165, + 479, + 189 + ], + "type": "text", + "content": "In the next sections, we will dive into the gating and sparse mixture of expert modules." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 213, + 301, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 213, + 301, + 226 + ], + "spans": [ + { + "bbox": [ + 131, + 213, + 301, + 226 + ], + "type": "text", + "content": "3.2 Ray Sampling and Filtering" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 240, + 479, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 240, + 479, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 479, + 277 + ], + "type": "text", + "content": "The first phase of our method involves learning a coarse and explicit density field starting from the density volume " + }, + { + "bbox": [ + 130, + 240, + 479, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_D" + }, + { + "bbox": [ + 130, + 240, + 479, + 277 + ], + "type": "text", + "content": ", that we can leverage for skipping empty spaces. Given a ray " + }, + { + "bbox": [ + 130, + 240, + 479, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 130, + 240, + 479, + 277 + ], + "type": "text", + "content": " as explained in Sec. 2.1, we sample " + }, + { + "bbox": [ + 130, + 240, + 479, + 277 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 240, + 479, + 277 + ], + "type": "text", + "content": " points along it:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 244, + 288, + 481, + 302 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 288, + 481, + 302 + ], + "spans": [ + { + "bbox": [ + 244, + 288, + 481, + 302 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {p, r} = \\text {s a m p l i n g} (\\mathbf {r}) \\in \\mathbb {R} ^ {N \\times 3}, \\tag {3}", + "image_path": "7b8977824980eaaff610322b760ea4f9a318c8702b35d280ff120ae3dfd02bbb.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 313, + 479, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 479, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 479, + 361 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 130, + 313, + 479, + 361 + ], + "type": "inline_equation", + "content": "p \\in [0, \\dots, N - 1]" + }, + { + "bbox": [ + 130, + 313, + 479, + 361 + ], + "type": "text", + "content": ". We suppress the index " + }, + { + "bbox": [ + 130, + 313, + 479, + 361 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 313, + 479, + 361 + ], + "type": "text", + "content": " for abuse of notation. Next, we compute the density for each point " + }, + { + "bbox": [ + 130, + 313, + 479, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_p \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 313, + 479, + 361 + ], + "type": "text", + "content": " and discard those with negligible density. The density value " + }, + { + "bbox": [ + 130, + 313, + 479, + 361 + ], + "type": "inline_equation", + "content": "\\sigma_{D,p}" + }, + { + "bbox": [ + 130, + 313, + 479, + 361 + ], + "type": "text", + "content": " for each point is computed by linearly interpolating " + }, + { + "bbox": [ + 130, + 313, + 479, + 361 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 313, + 479, + 361 + ], + "type": "text", + "content": " neighboring voxels:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 228, + 380, + 480, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 380, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 228, + 380, + 480, + 392 + ], + "type": "interline_equation", + "content": "\\sigma_ {D, p} = \\operatorname {a c t} (\\operatorname {i n t e r p o l a t e} \\left(\\mathbf {x} _ {p}, V _ {D}\\right)) \\in \\mathbb {R} \\tag {4}", + "image_path": "cead5546794ff76d097c686b3e3eef78625269f27ed255eff5da7ef07b1c15e1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 400, + 479, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 479, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 479, + 425 + ], + "type": "text", + "content": "where act represents a density activation function (such as softmax). We denote " + }, + { + "bbox": [ + 130, + 400, + 479, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathbf{f}}" + }, + { + "bbox": [ + 130, + 400, + 479, + 425 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 400, + 479, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathbf{f}} \\in R^3" + }, + { + "bbox": [ + 130, + 400, + 479, + 425 + ], + "type": "text", + "content": " a remaining point after the filtering operation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 449, + 282, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 449, + 282, + 462 + ], + "spans": [ + { + "bbox": [ + 131, + 449, + 282, + 462 + ], + "type": "text", + "content": "3.3 Trainable Gating Model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 475, + 479, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 479, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 479, + 512 + ], + "type": "text", + "content": "Our gating mechanism incorporates a hybrid architecture: an explicit feature grid " + }, + { + "bbox": [ + 130, + 475, + 479, + 512 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_G" + }, + { + "bbox": [ + 130, + 475, + 479, + 512 + ], + "type": "text", + "content": " and a shallow MLP. The gating mechanism can also be seen as a probability field. First, we compute per-point features:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 235, + 524, + 480, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 524, + 480, + 537 + ], + "spans": [ + { + "bbox": [ + 235, + 524, + 480, + 537 + ], + "type": "interline_equation", + "content": "\\mathbf {f e a t} = \\operatorname {i n t e r p o l a t e} \\left(\\mathbf {x} _ {f}, \\mathbf {V} _ {G}\\right) \\in \\mathbb {R} ^ {C} \\tag {5}", + "image_path": "2b0db990191a7e600327238f69d675220f2186e4050f013702c26d8f22588787.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 548, + 479, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 479, + 573 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 479, + 573 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 548, + 479, + 573 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 548, + 479, + 573 + ], + "type": "text", + "content": " represents the number of channels of " + }, + { + "bbox": [ + 130, + 548, + 479, + 573 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_G" + }, + { + "bbox": [ + 130, + 548, + 479, + 573 + ], + "type": "text", + "content": ". Subsequently, we transform each feature into a probability. We compute logits as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 248, + 590, + 480, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 590, + 480, + 602 + ], + "spans": [ + { + "bbox": [ + 248, + 590, + 480, + 602 + ], + "type": "interline_equation", + "content": "\\operatorname {l o g i t s} = \\operatorname {M L P} (\\mathbf {f e a t}) \\in \\mathbb {R} ^ {M} \\tag {6}", + "image_path": "6bb31a749a3c9704bf05a859abb62aac90cb04f0579b3cbc2c6ffadf2fa0dd5e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 612, + 479, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 612, + 479, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 612, + 479, + 635 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 612, + 479, + 635 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 612, + 479, + 635 + ], + "type": "text", + "content": " is the number of experts in the mixture of experts. Finally, we apply a per-row softmax. Our gating function can be summarized as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 194, + 653, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 653, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 194, + 653, + 480, + 666 + ], + "type": "interline_equation", + "content": "G \\left(\\mathbf {x} _ {\\mathbf {f}}\\right) = \\operatorname {s o f t m a x} (\\operatorname {M L P} (\\text {i n t e r p o l a t e} \\left(\\mathbf {x} _ {f}, \\mathbf {V} _ {G}\\right))) \\in \\mathbb {R} ^ {M} \\tag {7}", + "image_path": "6bf406d00046948ce4ab2f1b65922d1f86f038772293737f0fd8ba859ac347fa.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 373, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 373, + 91, + 447, + 100 + ], + "type": "text", + "content": "Boost Your NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 238, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 238, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 238, + 129 + ], + "type": "text", + "content": "3.4 NeRFs Experts" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "text", + "content": "As mentioned in Sec. 3.1, we employ a set of " + }, + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "text", + "content": " pretrained NeRFs models as experts. After feeding as input each point " + }, + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_f" + }, + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "text", + "content": " to the gate, we route them to the " + }, + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "text", + "content": " experts with the " + }, + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "text", + "content": " highest probabilities " + }, + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "inline_equation", + "content": "G(\\mathbf{x}_f)" + }, + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "text", + "content": ". We define this set of indexes of the selected experts as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 261, + 191, + 481, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 191, + 481, + 211 + ], + "spans": [ + { + "bbox": [ + 261, + 191, + 481, + 211 + ], + "type": "interline_equation", + "content": "\\mathcal {K} = \\underset {T o p - k} {\\operatorname {a r g m a x}} (G (\\mathbf {x} _ {f})). \\tag {8}", + "image_path": "897845bd87afd916c00249a92e392074a22b8ae3de2662a61eb33533b66062b4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 217, + 482, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 217, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 131, + 217, + 482, + 289 + ], + "type": "text", + "content": "Each expert is a Fast-NeRF model that receives the dispatched points and the direction of the ray they lie on as input and outputs a density and a radiance value. We treat each expert as a black box, ensuring our method is inherently model-agnostic. The radiance " + }, + { + "bbox": [ + 131, + 217, + 482, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_f" + }, + { + "bbox": [ + 131, + 217, + 482, + 289 + ], + "type": "text", + "content": " and the density " + }, + { + "bbox": [ + 131, + 217, + 482, + 289 + ], + "type": "inline_equation", + "content": "\\sigma_f" + }, + { + "bbox": [ + 131, + 217, + 482, + 289 + ], + "type": "text", + "content": " for the point " + }, + { + "bbox": [ + 131, + 217, + 482, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_f" + }, + { + "bbox": [ + 131, + 217, + 482, + 289 + ], + "type": "text", + "content": " laying on a ray with direction " + }, + { + "bbox": [ + 131, + 217, + 482, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 131, + 217, + 482, + 289 + ], + "type": "text", + "content": " can be expressed as the sum of each expert predictions weighted by their probability:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 241, + 300, + 481, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 300, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 241, + 300, + 481, + 324 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {f}, \\sigma_ {f} = \\sum_ {i \\in \\mathcal {K}} E _ {i} \\left(\\mathbf {x} _ {\\mathbf {f}}, \\mathbf {d}\\right) \\cdot G \\left(\\mathbf {x} _ {\\mathbf {f}}\\right) _ {i} \\tag {9}", + "image_path": "3d9bc839caaeaaddd05af4952f078de3d0a62d49b783c7392c863540576dbf7a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 328, + 481, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 481, + 352 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 328, + 481, + 352 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 130, + 328, + 481, + 352 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 130, + 328, + 481, + 352 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 328, + 481, + 352 + ], + "type": "text", + "content": "-th expert and " + }, + { + "bbox": [ + 130, + 328, + 481, + 352 + ], + "type": "inline_equation", + "content": "G(\\mathbf{x_f})_i" + }, + { + "bbox": [ + 130, + 328, + 481, + 352 + ], + "type": "text", + "content": " the probability for the point to be dispatched to that expert " + }, + { + "bbox": [ + 130, + 328, + 481, + 352 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 328, + 481, + 352 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 352, + 481, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 481, + 376 + ], + "type": "text", + "content": "Once we obtain radiance and density values for all the points on the ray, we can compute the pixel color with Equation 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 392, + 291, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 392, + 291, + 404 + ], + "spans": [ + { + "bbox": [ + 131, + 392, + 291, + 404 + ], + "type": "text", + "content": "3.5 Resolution-based Routing" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 411, + 481, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 481, + 459 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 481, + 459 + ], + "type": "text", + "content": "To balance the load and prevent the gate from focusing on assigning points to a single expert (typically the one with the highest resolution), we employ a resolution-weighted auxiliary loss. Given " + }, + { + "bbox": [ + 130, + 411, + 481, + 459 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 411, + 481, + 459 + ], + "type": "text", + "content": " experts, and a batch of points " + }, + { + "bbox": [ + 130, + 411, + 481, + 459 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 411, + 481, + 459 + ], + "type": "text", + "content": ", the auxiliary loss is defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 209, + 469, + 481, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 469, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 209, + 469, + 481, + 502 + ], + "type": "interline_equation", + "content": "L _ {\\text {a u x}} = \\frac {M}{| B | ^ {2}} \\sum_ {i = 1} ^ {M} c _ {i} m _ {i}, \\quad m _ {i} = \\sum_ {\\mathbf {x} _ {f} \\in B} G (\\mathbf {x} _ {f}) _ {i}, \\tag {10}", + "image_path": "ac07dd2c5d2a725b609b966df959964a37088a2206d114d3aff81c6b9da22d68.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "text", + "content": " represents the number of inputs dispatched to the expert " + }, + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "text", + "content": " is the sum of all the probabilities for each point in the batch for the expert " + }, + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "text", + "content": ". This loss helps balance the workload, ensuring that each expert processes a similar number of points. Ideally, in a perfectly balanced scenario, " + }, + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "inline_equation", + "content": "L_{aux}" + }, + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "text", + "content": " is expected to be 1, as both " + }, + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "text", + "content": " would be " + }, + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "inline_equation", + "content": "\\frac{M}{N}" + }, + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{M} c_{i} m_{i}" + }, + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "text", + "content": " being " + }, + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "inline_equation", + "content": "\\frac{M^2}{N}" + }, + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "text", + "content": ". However, in the case of total imbalance, it tends to the number " + }, + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "text", + "content": ". We aim to take a step further by assigning as many points as possible to low-resolution experts while discouraging point assignment to high-resolution models. Hence, we introduce some penalty terms " + }, + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 130, + 506, + 482, + 636 + ], + "type": "text", + "content": " associated with each expert. The higher the resolution of the model, the higher the penalty. We define a novel resolution-based auxiliary loss:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 239, + 636, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 636, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 239, + 636, + 481, + 669 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {r w - a u x}} = \\frac {M}{| B | ^ {2}} \\sum_ {i = 1} ^ {M} c _ {i} m _ {i} w _ {i}. \\tag {11}", + "image_path": "21812fcc730253f0099132b94466eba1142e834ebf30ac8243ca35dc54836a59.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "type": "text", + "content": "F. Di Sario et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "As for the weighting strategy, we opted for the following geometric progression. The weight " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " expert is computed as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 217, + 150, + 481, + 177 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 150, + 481, + 177 + ], + "spans": [ + { + "bbox": [ + 217, + 150, + 481, + 177 + ], + "type": "interline_equation", + "content": "w _ {i} = \\exp \\left(\\frac {\\ln M}{M - 1}\\right) ^ {i}, \\quad i \\in [ 0, \\dots , M - 1 ]. \\tag {12}", + "image_path": "d54ff91910505fa2556421c22a4304e117e43fe34572f3eea1a03bf6bf98c3be.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 178, + 293, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 178, + 293, + 190 + ], + "spans": [ + { + "bbox": [ + 146, + 178, + 293, + 190 + ], + "type": "text", + "content": "The total loss is then defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 255, + 196, + 481, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 196, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 255, + 196, + 481, + 209 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {t o t}} = L _ {\\mathrm {n e r f}} + \\lambda L _ {\\mathrm {r w - a u x}}. \\tag {13}", + "image_path": "e7f45a668f391c3d5f4535abd11bce6a78fd41af52b2c74d29d0cc9919236ff5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 212, + 479, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 212, + 479, + 249 + ], + "spans": [ + { + "bbox": [ + 131, + 212, + 479, + 249 + ], + "type": "text", + "content": "As we will see in Sec. 4, our proposed loss not only further improves the reconstruction quality but also encourages sparsity in higher-resolution models. A comparison with other weighting strategies is proposed in the Supplementary ??" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 263, + 230, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 230, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 230, + 277 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 284, + 198, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 284, + 198, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 284, + 198, + 296 + ], + "type": "text", + "content": "4.1 Setting" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "text", + "content": "The code was written in Python 3.8, using PyTorch 1.12, and executed on a single NVIDIA A40 GPU. We tested our technique on the DVGO, TensoRF, and Instant-NGP architecture. The code from official repositories was used as the starting point. For Instant-NGP, the native implementation in PyTorch ngp_p1 was employed, which exhibits comparable performance and reconstruction quality to the official NVIDIA implementation. We experimented with various configurations, ranging from a minimum of 3 models to 5 models of different resolutions. The experts are ordered by resolution, such that the " + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "inline_equation", + "content": "i + 1" + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "text", + "content": "-th expert has approximately double the parameters of the " + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "text", + "content": "-th expert. For all configurations and models, " + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "text", + "content": " was set to " + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "text", + "content": ". The Gate consists of a grid (dense voxel grid for DVGO, factorized grid for TensoRF, and hash grid for Instant-NGP) and a shallow MLP (2 layers with 64 neurons each) with ReLU activation function. The resolution of the gate is low: " + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "inline_equation", + "content": "128^3" + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "text", + "content": " for both DVGO and TensoRF, while for Instant-NGP we used " + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "inline_equation", + "content": "L = 6" + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "text", + "content": ". The number of iterations is set to " + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "inline_equation", + "content": "20k" + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "text", + "content": " for all the architectures. All other hyperparameters are left as the original implementations. We draw experiments with our method using Top-" + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "text", + "content": " experts, with " + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "inline_equation", + "content": "k = 2" + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "text", + "content": ". We compare our results versus baselines with comparable resolutions, as well as a Fast-NeRF ensemble (Ens) obtained by jointly fine-tuning all models and averaging their predictions. It can be noted that this ensemble can be interpreted as a limit case for our method with " + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "inline_equation", + "content": "k = M" + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "inline_equation", + "content": "G(\\mathbf{x}_f)_i = 1 / M" + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\forall i" + }, + { + "bbox": [ + 130, + 300, + 482, + 552 + ], + "type": "text", + "content": " similar to the method proposed by [7]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 566, + 269, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 566, + 269, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 269, + 577 + ], + "type": "text", + "content": "4.2 Metrics and Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "For each test, image-quality metrics such as PSNR, SSIM [30], and LPIPS [34] (computed on AlexNet [18]) are presented. Additionally, we report the number of non-zero parameters as " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\| w_0\\|" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ", the average GFLOPs required by each model to render the images of the test set, and total training times. We present results obtained across four major datasets, namely: Synthetic-NeRF [24], Neural Sparse Voxel Field Dataset (NSVF) [20], TanksAndTemple [17], and Local Light Field Fusion Dataset (LLFF) [23]." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 373, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 373, + 91, + 447, + 101 + ], + "type": "text", + "content": "Boost Your NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 145, + 481, + 378 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 1: Results on DVGO, TensoRF and Instant-NGP with " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "M = 5" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "\\| w\\| _0" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": " is expressed as multiple of " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "10^{6}" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": ", while for Instant-NGP as a multiple of " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "10^{5}" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 145, + 481, + 378 + ], + "lines": [ + { + "bbox": [ + 132, + 145, + 481, + 378 + ], + "spans": [ + { + "bbox": [ + 132, + 145, + 481, + 378 + ], + "type": "table", + "html": "
DatasetMetricsDVGOTensorRFInstant-NGP
baselineTop-1Top-2EnsbaselineTop-1Top-2EnsbaselineTop-1Top-2Ens
BlenderPSNR ↑33.0433.4333.7433.7932.9833.6834.0934.0033.3533.5633.8334.01
SSIM ↑0.9610.9640.9650.9660.9580.9650.9680.9680.9630.9630.9650.966
LPIPS ↓0.0260.0240.0220.0220.0290.0230.0210.0220.0250.0450.0430.042
||w||0 ↓992639974024334931172126
GFLOPs ↓6354999402206886732134422144671123208
Time ↓26'21'25'32'44'69'76'70'11'32'34'37'
NSVFPSNR ↑35.2137.1237.5937.6836.7037.4037.9838.0836.4433.5937.0437.31
SSIM ↑0.9770.9840.9860.9860.9810.9840.9860.9870.9830.9830.9840.985
LPIPS ↓0.0150.0090.0080.0070.0130.0090.0080.0080.0100.0230.0220.020
||w||0 ↓9527431004228385430172027
GFLOPs ↓564430811190370657510531763285272123
Time ↓12'22'25'30'46'75'75'75'10'31'33'34'
TaTPSNR ↑28.9329.1429.2729.3728.4428.7829.1429.1129.0729.1629.3229.38
SSIM ↑0.9270.9250.9290.9320.9050.9240.9290.9280.9240.9270.9290.930
LPIPS ↓0.1070.1080.1050.10300.1060.0990.1090.1010.1250.1240.121
||w||0 ↓7416266579158088374770
GFLOPs ↓266616263066719835672791512691462112295311003
Time ↓22'25'28'38'72'70'78'101'14'37'39'45'
LLFFPSNR ↑26.2426.4326.6226.6526.7126.7327.0927.1024.9724.9025.1725.19
SSIM ↑0.8310.8320.8390.8430.8350.8360.8620.8640.7640.7630.7770.778
LPIPS ↓0.1360.1150.1110.1070.1140.1110.1010.1010.1280.2390.2370.234
||w||0 ↓622640113191116231526875148
GFLOPs ↓16781514250849724542322659211352257388713912712
Time ↓24'28'32'36'58'49'57'68'24'25'46'42'
", + "image_path": "f738e21099d61812d7fbdc8b44896c72d49c65e3e1d0319804c87d6cfa7cdce2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 406, + 264, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 406, + 264, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 406, + 264, + 418 + ], + "type": "text", + "content": "4.3 Quantitative Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 437, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 668 + ], + "type": "text", + "content": "The main experimental results in terms of the metrics defined above are shown in Table 1. Here, we present the results obtained using a mixture of " + }, + { + "bbox": [ + 130, + 437, + 482, + 668 + ], + "type": "inline_equation", + "content": "M = 5" + }, + { + "bbox": [ + 130, + 437, + 482, + 668 + ], + "type": "text", + "content": " experts; further configurations can be found in the supplementary Sec. ?? Our experiments reveal that our MoE provides a significant rendering quality improvement with respect to the baseline with no or limited impact in terms of computational cost. The increase in rendering quality is notable in Synthetic NeRF (up to 1 dB) and NSVF (up to 1.3 dB). Although more moderate, improvements are still evident, even in more challenging datasets such as TanksAndTemple and LLFF, with about 0.5 dB gain. While Top-1 can already achieve state-of-the-art render quality, the Top-2 strategy further enhances image quality, reaching levels comparable to the ensemble but with much greater efficiency (about half the average FLOPs per rendering) and using significantly fewer parameters. This observation is also illustrated in Figure 3 , where FLOPS/PSNR plots are shown (each marker in every curve refers to the cases with " + }, + { + "bbox": [ + 130, + 437, + 482, + 668 + ], + "type": "inline_equation", + "content": "M = 3,4,5" + }, + { + "bbox": [ + 130, + 437, + 482, + 668 + ], + "type": "text", + "content": " experts respectively). Training times are on average are longer but still acceptable (in the worst-case scenario, around 1h is required to train our MoE). However, in the case of Top1, they can be faster than baselines. This is because of our resolution methods, which tend to favor low-resolution models. Based on these analyses, the Top-2 strategy strikes an excellent quality/cost trade-off. The ensembling configura" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 236, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 236, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 236, + 101 + ], + "type": "text", + "content": "F. Di Sario et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 113, + 242, + 201 + ], + "blocks": [ + { + "bbox": [ + 136, + 113, + 242, + 201 + ], + "lines": [ + { + "bbox": [ + 136, + 113, + 242, + 201 + ], + "spans": [ + { + "bbox": [ + 136, + 113, + 242, + 201 + ], + "type": "image", + "image_path": "f9550c86b308a27fc0fb872ad9021cf67ac989f7a485604387820576fd5030fb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 252, + 113, + 358, + 201 + ], + "blocks": [ + { + "bbox": [ + 252, + 113, + 358, + 201 + ], + "lines": [ + { + "bbox": [ + 252, + 113, + 358, + 201 + ], + "spans": [ + { + "bbox": [ + 252, + 113, + 358, + 201 + ], + "type": "image", + "image_path": "e568527cb91f343b5ef103030064e80d5f334f78c6b4f9455a9202897cb0d82a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 369, + 113, + 475, + 201 + ], + "blocks": [ + { + "bbox": [ + 369, + 113, + 475, + 201 + ], + "lines": [ + { + "bbox": [ + 369, + 113, + 475, + 201 + ], + "spans": [ + { + "bbox": [ + 369, + 113, + 475, + 201 + ], + "type": "image", + "image_path": "7cd19a8e4c18091a0f7e3b41fca58af0f5707f5d7c0c8a4f760f6dd06a21bb2e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 135, + 209, + 242, + 297 + ], + "blocks": [ + { + "bbox": [ + 135, + 209, + 242, + 297 + ], + "lines": [ + { + "bbox": [ + 135, + 209, + 242, + 297 + ], + "spans": [ + { + "bbox": [ + 135, + 209, + 242, + 297 + ], + "type": "image", + "image_path": "69dafdc302de52af7289909dc5a53fa52dace5473cbb51e58fccb4ac78aa80a7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 302, + 209, + 312 + ], + "lines": [ + { + "bbox": [ + 170, + 302, + 209, + 312 + ], + "spans": [ + { + "bbox": [ + 170, + 302, + 209, + 312 + ], + "type": "text", + "content": "(a) DVGO" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 252, + 209, + 358, + 297 + ], + "blocks": [ + { + "bbox": [ + 252, + 209, + 358, + 297 + ], + "lines": [ + { + "bbox": [ + 252, + 209, + 358, + 297 + ], + "spans": [ + { + "bbox": [ + 252, + 209, + 358, + 297 + ], + "type": "image", + "image_path": "fd742c4c39cf2372498587d65a3a7167952c51b97931c2178c4a43167cc806dc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 302, + 328, + 312 + ], + "lines": [ + { + "bbox": [ + 282, + 302, + 328, + 312 + ], + "spans": [ + { + "bbox": [ + 282, + 302, + 328, + 312 + ], + "type": "text", + "content": "(b) TensorRF" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 369, + 209, + 474, + 297 + ], + "blocks": [ + { + "bbox": [ + 369, + 209, + 474, + 297 + ], + "lines": [ + { + "bbox": [ + 369, + 209, + 474, + 297 + ], + "spans": [ + { + "bbox": [ + 369, + 209, + 474, + 297 + ], + "type": "image", + "image_path": "892b5dcd660df7c1bf65c52e7c0813c3841a8918b524ed6ce7a22262a941f1db.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 302, + 451, + 312 + ], + "lines": [ + { + "bbox": [ + 392, + 302, + 451, + 312 + ], + "spans": [ + { + "bbox": [ + 392, + 302, + 451, + 312 + ], + "type": "text", + "content": "(c) Instant-NGP" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 321, + 480, + 343 + ], + "lines": [ + { + "bbox": [ + 131, + 321, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 131, + 321, + 480, + 343 + ], + "type": "text", + "content": "Fig. 3: PSNR/GFLOPs plots for Synthetic-NeRF and TanksAndTemple. The remaining datasets show similar results." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 366, + 480, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 366, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 131, + 366, + 480, + 390 + ], + "type": "text", + "content": "tion, using all experts, represents an upper bound in terms of image quality while being the worst case in computational terms." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 407, + 257, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 257, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 257, + 418 + ], + "type": "text", + "content": "4.4 Qualitative Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 426, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 582 + ], + "type": "text", + "content": "Figure 4 shows visual comparisons among the baseline, ensemble, Top-1, and Top-2. From the selected image crops, one can appreciate that our method ensures superior reconstruction quality compared to the baselines, effectively reproducing sharper details while reducing noise on texture-less spots. This is particularly notable when examining elements such as the window decorations in the Palace scene or the text on the box above the desk in the Room scene. Additionally, surfaces such as the semi-transparent glass in the Wineholder or the Caterpillar scene appear less noisy and more faithful to the original. This improvement can be attributed to the synergy among different resolution models: low-resolution models excel at representing lower frequencies, thereby introducing less noise, while high-resolution models can focus solely on high-frequency components. Additionally, it is important to notice that there is no significant difference between the ensemble and Top-2." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 610, + 327, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 327, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 327, + 622 + ], + "type": "text", + "content": "4.5 Comparison with Naive Methods" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 480, + 665 + ], + "type": "text", + "content": "As shown in Figure 1, our method leads to significantly higher quality reconstructions at a greatly reduced computational cost. Decreasing the step size yields the least noticeable improvement while incurring a substantial computational" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 373, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 373, + 91, + 447, + 100 + ], + "type": "text", + "content": "Boost Your NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 113, + 449, + 339 + ], + "blocks": [ + { + "bbox": [ + 167, + 113, + 449, + 339 + ], + "lines": [ + { + "bbox": [ + 167, + 113, + 449, + 339 + ], + "spans": [ + { + "bbox": [ + 167, + 113, + 449, + 339 + ], + "type": "image", + "image_path": "2d38a8a6f81d2ff182f38c17d027390221ce77d0fa37588021c595da6f3655b2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 346, + 482, + 380 + ], + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 380 + ], + "type": "text", + "content": "Fig. 4: Qualitative results on some of the scenes of each dataset. From left to right: ground truths, baselines, Top-1, Top-2 and Ensemble. Each model has the same parameters and has been trained for the same number of iterations (DVGO)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "type": "text", + "content": "expense, as each sampled point requires an evaluation by the color decoder. Our method achieves high-quality reconstructions with the same step size as the baseline models. Increasing the MLP parameters can boost accuracy, but again, the computational costs become substantial compared to our method. Similarly, increasing resolution results in way inferior PSNR with respect to our method at a comparable computational cost. Our method also allows scaling to higher resolutions, while baseline models tend to introduce noise and artifacts, leading to a decrease in reconstruction quality as the resolution increases." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 514, + 387, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 514, + 387, + 528 + ], + "spans": [ + { + "bbox": [ + 131, + 514, + 387, + 528 + ], + "type": "text", + "content": "4.6 Gate Visualization and Scene Decomposition" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "text", + "content": "In Figure 5, we visualize the gate (probability field in grayscale) and each expert output in the case " + }, + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "inline_equation", + "content": "M = 3" + }, + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "text", + "content": ". On the right side of the figure, one can appreciate per-model renders and experts' specialization. It is worth noting how higher-resolution experts render high-frequency details. This is particularly evident in the Mic scene." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 610, + 206, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 610, + 206, + 621 + ], + "spans": [ + { + "bbox": [ + 131, + 610, + 206, + 621 + ], + "type": "text", + "content": "4.7 Ablation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Here we investigate how different design choices of our model affect the performance and rendering quality. These choices include the gate resolution, different gate formulations, and the number of experts in the Top-" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "type": "text", + "content": "F. Di Sario et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 114, + 447, + 309 + ], + "blocks": [ + { + "bbox": [ + 167, + 114, + 447, + 309 + ], + "lines": [ + { + "bbox": [ + 167, + 114, + 447, + 309 + ], + "spans": [ + { + "bbox": [ + 167, + 114, + 447, + 309 + ], + "type": "image", + "image_path": "08a346b4472e12ae5271af5ff9628e37dfb6f56de18ed3137174d89ffa77872c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 318, + 480, + 340 + ], + "lines": [ + { + "bbox": [ + 132, + 318, + 480, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 480, + 340 + ], + "type": "text", + "content": "Fig. 5: Gate (gray-scale images) and per-model output visualizations with 3 experts and resolution-based routing (DVGO). Images are ordered by increasing resolutions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 370, + 480, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 370, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 480, + 407 + ], + "type": "text", + "content": "Here we investigate how different design choices of our model affect the performance and rendering quality. These choices include the gate resolution, different gate formulations, and the number of experts in the Top-" + }, + { + "bbox": [ + 130, + 370, + 480, + 407 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 370, + 480, + 407 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 440, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 482, + 548 + ], + "type": "text", + "content": "Gate Resolution and Gate Formulation. We show in Table 2 that a low-resolution gate is sufficient to achieve high-quality rendering. Interestingly, as the gate resolution increases, there is a slight decrease in rendering quality during testing, coupled with an increase in required FLOPs per image rendering and number of parameters. We also compare various gating strategies, including linear gating, the configuration proposed by Switch-NeRF, and our gate formulation. Through experimentation, it becomes evident that our approach outperforms others in terms of both render quality and performance, achieving performances comparable to a linear gating function." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Why Top-2? Here we investigate the performance and rendering quality trends with varying values of " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ". In Figure 6, we can see that while " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " can lead to a significant increase in quality with comparable performance to baseline models, the Top-2 further enhances quality at the expense of increased (but still acceptable) computation. The Top-3, Top-4, and Top-5 provide marginal quality improvements at a significantly higher computational cost. Hence, we consider the Top-2 the optimal balance between performance and quality." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 373, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 373, + 91, + 447, + 100 + ], + "type": "text", + "content": "Boost Your NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 156, + 273, + 210 + ], + "blocks": [ + { + "bbox": [ + 133, + 156, + 273, + 210 + ], + "lines": [ + { + "bbox": [ + 133, + 156, + 273, + 210 + ], + "spans": [ + { + "bbox": [ + 133, + 156, + 273, + 210 + ], + "type": "table", + "html": "
ResGate TypePSNR||w||0GFLOPs
1283Ours36.826476
2563Ours36.7733601
3003Ours36.7942677
-Linear36.2524444
-Switch-NeRF36.44251095
", + "image_path": "1f26f236a43fd9b466762daa74876b4e314eca3af1742b4d02c64143be1cb780.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 219, + 273, + 251 + ], + "lines": [ + { + "bbox": [ + 132, + 219, + 273, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 219, + 273, + 251 + ], + "type": "text", + "content": "Table 2: Comparison of different gate resolutions and formulations on Lego scene and DVGO." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 294, + 119, + 475, + 214 + ], + "blocks": [ + { + "bbox": [ + 294, + 119, + 475, + 214 + ], + "lines": [ + { + "bbox": [ + 294, + 119, + 475, + 214 + ], + "spans": [ + { + "bbox": [ + 294, + 119, + 475, + 214 + ], + "type": "image", + "image_path": "044d1bb3a35c35583d4c3b24f456b2581b24df7d67eeed7e84f9e961634af261.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 287, + 230, + 480, + 252 + ], + "lines": [ + { + "bbox": [ + 287, + 230, + 480, + 252 + ], + "spans": [ + { + "bbox": [ + 287, + 230, + 480, + 252 + ], + "type": "text", + "content": "Fig. 6: Comparison of Gating Functions (Top-1 to Top-5) on Synthetic-NeRF with DVGO." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 277, + 222, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 222, + 290 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 222, + 290 + ], + "type": "text", + "content": "5 Limitations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 309, + 482, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 482, + 488 + ], + "type": "text", + "content": "Our study present some limitations. First of all, a pre-training phase, on which each model is trained independently is required for achieving good results. Training end-to-end, without a pre-training phase, can lead to reconstruction quality that is noticeably inferior to baselines. Pre-training models at different resolutions allows for diversified architectures, making it easier for the gate to learn the decomposition. The training is also strongly dependent on the auxiliary loss. Different values of " + }, + { + "bbox": [ + 130, + 309, + 482, + 488 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 309, + 482, + 488 + ], + "type": "text", + "content": " can significantly influence performances and load balancing. We conducted a sweep to identify a value that works well across many scenarios, but it may not be suitable for different datasets. Another limitation is the overhead introduced by the MoE. Each input token is first interpolated with the gate's grid, decoded into probabilities and routed to the chosen expert. These operations can significantly slow-down the rendering process, leading to higher training and inference times. Although considerable effort was devoted to developing the most efficient gate possible, our MoE is still slower than the respective baselines." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 514, + 226, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 226, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 226, + 528 + ], + "type": "text", + "content": "6 Conclusions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we introduced a model-agnostic framework for enhancing the rendering of Fast-NeRF models. Our formulation of the Gate reduces computational costs in both training and inference phases while ensuring better quality compared to a traditional gate. Additionally, the introduction of an auxiliary loss with res penalty allows for increased utilization of low-resolution models, reducing the number of active parameters and promoting sparsity in higher-resolution models. Our results demonstrate how this approach can significantly improve reconstruction quality while considering performance metrics. Specifically, we show that a Top-2 strategy strikes a good balance between performance and quality." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 237, + 101 + ], + "type": "text", + "content": "F. Di Sario et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 480, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 480, + 165 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 480, + 165 + ], + "type": "text", + "content": "This work was partially funded by Hi!PARIS Center on Data Analytics and Artificial Intelligence." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 186, + 197, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 186, + 197, + 198 + ], + "spans": [ + { + "bbox": [ + 133, + 186, + 197, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 212, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 212, + 481, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 212, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 212, + 481, + 255 + ], + "type": "text", + "content": "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 256, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 481, + 289 + ], + "type": "text", + "content": "2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 290, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 290, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 138, + 290, + 481, + 332 + ], + "type": "text", + "content": "3. Chan, E.R., Monteiro, M., Kellnhofer, P., Wu, J., Wetzstein, G.: pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5799-5809 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 334, + 481, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 334, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 138, + 334, + 481, + 355 + ], + "type": "text", + "content": "4. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. ECCV (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 356, + 481, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 356, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 356, + 481, + 399 + ], + "type": "text", + "content": "5. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16569-16578 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 401, + 481, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 401, + 481, + 432 + ], + "spans": [ + { + "bbox": [ + 138, + 401, + 481, + 432 + ], + "type": "text", + "content": "6. Deng, Y., Yang, J., Xiang, J., Tong, X.: Gram: Generative radiance manifolds for 3d-aware image generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10673-10683 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 434, + 481, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 481, + 466 + ], + "type": "text", + "content": "7. Di Sario, F., Renzulli, R., Tartaglione, E., Grangetto, M.: Two is better than one: Achieving high-quality 3d scene modeling with a nerf ensemble. In: International Conference on Image Analysis and Processing. pp. 320-331. Springer (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 468, + 481, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 481, + 499 + ], + "type": "text", + "content": "8. Fedus, W., Zoph, B., Shazeer, N.: Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. The Journal of Machine Learning Research 23(1), 5232-5270 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 500, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 500, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 500, + 481, + 521 + ], + "type": "text", + "content": "9. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes for radiance fields in space, time, and appearance (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 522, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 481, + 555 + ], + "type": "text", + "content": "0. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "type": "text", + "content": "1. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High-fidelity neural rendering at 200fps. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14346-14355 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 588, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 481, + 620 + ], + "type": "text", + "content": "2. Hou, A., Liu, F., Ren, Z., Sarkis, M., Bi, N., Tong, Y., Liu, X.: Infamous-nerf: Improving face modeling using semantically-aligned hypernetworks with neural radiance fields. arXiv preprint arXiv:2312.16197 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 481, + 665 + ], + "type": "text", + "content": "3. Hu, W., Wang, Y., Ma, L., Yang, B., Gao, L., Liu, X., Ma, Y.: Tri-miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19774-19783 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 373, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 373, + 91, + 447, + 100 + ], + "type": "text", + "content": "Boost Your NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "14. Jacobs, R.A., Jordan, M.I., Nowlan, S.J., Hinton, G.E.: Adaptive mixtures of local experts. Neural computation 3(1), 79-87 (1991)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "type": "text", + "content": "15. Jeong, Y., Ahn, S., Choy, C., Anandkumar, A., Cho, M., Park, J.: Self-calibrating neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5846-5854 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 173, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 173, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 173, + 480, + 205 + ], + "type": "text", + "content": "16. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 206, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 206, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 134, + 206, + 480, + 239 + ], + "type": "text", + "content": "17. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG) 36(4), 1-13 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 240, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 240, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 133, + 240, + 480, + 274 + ], + "type": "text", + "content": "18. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25 (2012)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 274, + 480, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 274, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 133, + 274, + 480, + 307 + ], + "type": "text", + "content": "19. Lepikhin, D., Lee, H., Xu, Y., Chen, D., First, O., Huang, Y., Krikun, M., Shazeer, N., Gshard, Z.: Scaling giant models with conditional computation and automatic sharding. arXiv preprint arXiv:2006.16668 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 480, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 480, + 330 + ], + "type": "text", + "content": "20. Liu, L., Gu, J., Zaw Lin, K., Chua, T.S., Theobalt, C.: Neural sparse voxel fields. Advances in Neural Information Processing Systems 33, 15651-15663 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "type": "text", + "content": "21. Liu, Y., Li, X., Yu, F., Zhou, Q.: Probabilistic neural scene representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "type": "text", + "content": "22. Max, N.: Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics 1(2), 99-108 (1995)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 386, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 386, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 133, + 386, + 480, + 430 + ], + "type": "text", + "content": "23. Mildenhall, B., Srinivasan, P.P., Ortiz-Cayon, R., Kalantari, N.K., Ramamoorthi, R., Ng, R., Kar, A.: Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG) 38(4), 1-14 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 431, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 431, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 133, + 431, + 480, + 464 + ], + "type": "text", + "content": "24. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European conference on computer vision. pp. 405-421. Springer (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 464, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 464, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 133, + 464, + 480, + 497 + ], + "type": "text", + "content": "25. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 498, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 498, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 133, + 498, + 480, + 531 + ], + "type": "text", + "content": "26. Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14335-14345 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 532, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 532, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 133, + 532, + 480, + 564 + ], + "type": "text", + "content": "27. Riquelme, C., Puigcerver, J., Mustafa, B., Neumann, M., Jenatton, R., Susano Pinto, A., Keysers, D., Houlsby, N.: Scaling vision with sparse mixture of experts. Advances in Neural Information Processing Systems 34, 8583-8595 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 565, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 565, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 133, + 565, + 480, + 597 + ], + "type": "text", + "content": "28. Shazeer, N., Mirhoseini, A., Maziarz, K., Davis, A., Le, Q., Hinton, G., Dean, J.: Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 599, + 480, + 632 + ], + "type": "text", + "content": "29. Sun, C., Sun, M., Chen, H.T.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5459-5469 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "text", + "content": "30. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 236, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 236, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 236, + 100 + ], + "type": "text", + "content": "F. Di Sario et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 336 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "31. Wizadwongsa, S., Phongthawee, P., Yenphraphai, J., Suwajanakorn, S.: Nex: Real-time view synthesis with neural basis expansion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8534-8543 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "text", + "content": "32. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time neural irradiance fields for free-viewpoint video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9421-9431 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "text", + "content": "33. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: Plenoctrees for real-time rendering of neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5752-5761 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "type": "text", + "content": "34. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "type": "text", + "content": "35. Zhenxing, M., Xu, D.: Switch-nerf: Learning scene decomposition with mixture of experts for large-scale neural radiance fields. In: The Eleventh International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "type": "text", + "content": "36. Zhou, Y., Lei, T., Liu, H., Du, N., Huang, Y., Zhao, V., Dai, A.M., Le, Q.V., Laudon, J., et al.: Mixture-of-experts with expert choice routing. Advances in Neural Information Processing Systems 35, 7103-7114 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 314, + 482, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 336 + ], + "type": "text", + "content": "37. Zhuang, Y., Zhu, H., Sun, X., Cao, X.: Mofanerf: Morphable facial neural radiance field. In: European conference on computer vision. pp. 268-285. Springer (2022)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 373, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 373, + 91, + 447, + 100 + ], + "type": "text", + "content": "Boost Your NeRF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Boosting 3D Single Object Tracking with 2D Matching Distillation and 3D Pre-training/515a618d-9747-46e7-9c9a-8a152456fd1b_content_list.json b/2024/Boosting 3D Single Object Tracking with 2D Matching Distillation and 3D Pre-training/515a618d-9747-46e7-9c9a-8a152456fd1b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..53db67122d9b6d399c91d4e0c60a563ee2625c21 --- /dev/null +++ b/2024/Boosting 3D Single Object Tracking with 2D Matching Distillation and 3D Pre-training/515a618d-9747-46e7-9c9a-8a152456fd1b_content_list.json @@ -0,0 +1,1666 @@ +[ + { + "type": "text", + "text": "Boosting 3D Single Object Tracking with 2D Matching Distillation and 3D Pre-training", + "text_level": 1, + "bbox": [ + 245, + 140, + 754, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qiangqiang $\\mathsf{W u}^{1}$ , Yan Xia\\*2,3, Jia Wan4, and Antoni B. Chan1", + "bbox": [ + 250, + 210, + 751, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Computer Science, City University of Hong Kong", + "bbox": [ + 277, + 238, + 723, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Technical University of Munich", + "bbox": [ + 388, + 253, + 612, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Munich Center for Machine Learning (MCML)", + "bbox": [ + 338, + 267, + 663, + 280 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ School of Computer Science and Technology, Harbin Institute of Technology, Shenzhen", + "bbox": [ + 238, + 281, + 761, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "qiangqwu2-c@my.cityu.edu.hk, yan.xia@tum.de, jiawan1998@gmail.com, abchan@cityu.edu.hk", + "bbox": [ + 243, + 310, + 754, + 335 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. 3D single object tracking (SOT) is an essential task in autonomous driving and robotics. However, learning robust 3D SOT trackers remains challenging due to the limited category-specific point cloud data and the inherent sparsity and incompleteness of LiDAR scans. To tackle these issues, we propose a unified 3D SOT framework that leverages 3D generative pre-training and learns robust 3D matching abilities from 2D pre-trained foundation trackers. Our framework features a consistent target-matching architecture with the widely used 2D trackers, facilitating the transfer of 2D matching knowledge. Specifically, we first propose a lightweight Target-Aware Projection (TAP) module, allowing the pre-trained 2D tracker to work well on the projected point clouds without further fine-tuning. We then propose a novel IoU-guided matching-distillation framework that utilizes the powerful 2D pre-trained trackers to guide 3D matching learning in the 3D tracker, i.e., the 3D template-to-search matching should be consistent with its corresponding 2D template-to-search matching obtained from 2D pre-trained trackers. Our designs are applied to two mainstream 3D SOT frameworks: memory-less Siamese and contextual memory-based approaches, which are respectively named SiamDisst and MemDisst. Extensive experiments show that SiamDisst and MemDisst achieve state-of-the-art performance on KITTI, Waymo Open Dataset and nuScenes benchmarks, while running at above real-time speed of 25 and 90 FPS on a RTX3090 GPU.", + "bbox": [ + 259, + 368, + 738, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 696, + 374, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Single Object Tracking (SOT) has emerged as a basic 3D task for numerous practical applications in autonomous driving [30,67], robotics [41,66] and virtual reality [1]. Given a target described as a 3D bounding box (bbox) in the first frame, 3D SOT aims to track the target across all the frames in the 3D scene by predicting thebbox pose and position, which can be considered as a 3D extension of the 2D SOT task [62] but using a different input modality of 3D point clouds.", + "bbox": [ + 212, + 726, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding Author", + "bbox": [ + 217, + 824, + 385, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6b91c85c43066eab893cbc213a499c5a74a5f698278d387ac8aa7e361e91653d.jpg", + "image_caption": [ + "Fig. 1: Comparison of our SiamDisst and MemDisst to state-of-the-art 3D SOT on the Van category of the KITTI dataset [18]. Compared with Siamese-based trackers (plotted as squares), our SiamDisst achieves the best performance while running at the real-time speed of 25 FPS. Our lightweight MemDisst outperforms the other contextual memory and motion-based approaches (plotted as circles), setting new SOTA performance on KITTI-Van, and runs efficiently at 90 FPS. This demonstrates that our designs effectively boost 3D matching with limited category-specific training data (e.g., the KITTI-Van training set only contains 1994 examples)." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 500, + 267 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7588f4b20e70ff6f1c065388725e3bf7c0be719bf7a912100a5f9df01681eafe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 143, + 756, + 266 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, existing 3D SOT methods [38,62,73] mainly focus on learning robust 3D tracking models from sparse point cloud datasets [5,18]. For example, 3D siamese trackers [27,33,47,77] follow a siamese matching paradigm for accurate localization. To leverage more historical cues, MBPTrack [70] further extends this framework to a contextual memory-matching mechanism. Although the 3D matching models proposed by these methods play a key role in obtaining favorable performance [5,18], the sparsity and incompleteness of the point cloud training limits their abilities. Moreover, a paucity of large-scale category-specific 3D tracking datasets leads to insufficient training data for effectively learning the 3D matching modules in existing 3D trackers.", + "bbox": [ + 212, + 439, + 787, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast to the 3D SOT task, various large-scale annotated video datasets [29] are available for training 2D SOT methods. Existing 2D deep trackers can fully benefit from the feature backbones learned from 2D image [22] or video pre-training methods [62], enabling the learning of robust matching modules for 2D SOT on videos. The state-of-the-art trackers [38,62,73] with strong matching abilities show excellent tracking results even without using online memory mechanisms. With the remarkable success of 2D SOT, our work is motivated by the following questions: 1) Inspired by 2D SOT, can we design a general and effective 3D SOT framework that benefits from 3D pre-training? 2) Can we transfer the well-learned matching ability of state-of-the-art 2D trackers to 3D matching in 3D SOT trackers, so as to facilitate training from limited category-specific point cloud data?", + "bbox": [ + 212, + 594, + 787, + 775 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main challenge of achieving the transfer of the 2D matching ability lies in two aspects. First, there is a significant misalignment between the input modalities for 2D and 3D trackers, dense 2d images and sparse 3d point clouds, respectively. Note that in the 3D SOT setting, only the single modality of point", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Q. Wu et al.", + "bbox": [ + 271, + 114, + 357, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2b1c7e2963833f85e0e21dfcca44039fc3b73ae116c6d54e3f0ac5b7d3f188ad.jpg", + "image_caption": [ + "(a) Previous Siamese 3D SOT" + ], + "image_footnote": [], + "bbox": [ + 246, + 143, + 401, + 284 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5f064ab3210a8271c795d2d5bbd54b93213e202ad044b6e7d9f055ccdc9c929f.jpg", + "image_caption": [ + "(b) Our Proposed 3D SOT (i.e., SiamDisst and MemDisst)", + "Fig. 2: The comparison of 3D SOT frameworks (a) the previous Siamese 3D SOT trackers [27,47,77] focus on learning the 3D matching module from limited category-specific point cloud (PC) training data. (b) The proposed 3D SOT framework bridges the gap between 2D and 3D SOT by using our designed target-aware projection (TAP) module, which enables the 2D tracker [73] trained on large-scale 2D videos to operate well on the point clouds. The well-learned 2D matching pattern is further distilled to facilitate the 3D matching learning. Note that only the 3D-related modules (denoted as blue components) are optimized during the distillation training stage. The 2D tracker and TAP module are not used during online inference to increase speed." + ], + "image_footnote": [], + "bbox": [ + 421, + 143, + 756, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cloud data is used as input, which makes it difficult to directly apply the current 2D SOT trackers. Second, how to effectively transfer the 2D matching knowledge in pre-trained 2D trackers to guide the learning of the 3D matching in 3D trackers is unclear given the differences in internal representations between 2D and 3D SOT, i.e., 2d dense feature maps and unordered 3D point embeddings.", + "bbox": [ + 212, + 487, + 787, + 563 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To tackle the aforementioned challenges, we propose a general 3D SOT framework that can effectively utilize knowledge learned in 2D SOT, which is illustrated in Fig. 2(b). We first propose to align the different input modalities of 2D and 3D trackers via a learnable target-aware projection (TAP) module, which takes the raw point cloud as the input and projects them into the 2D image space. These projected images are suitable for 2D target matching and 2D bounding box regression by 2D pre-trained trackers, i.e., no fine-tuning of 2D trackers is required. Secondly, we propose a matching distillation approach that uses the knowledge in the high-quality 2D matching patterns from the 2D tracker as guidance for learning the matching patterns for 3D SOT. The proposed framework employs a standard Vision Transformer (ViT) as the matching module, which can naturally benefit from 3D pre-training approaches [42,49] and also creates a design that is consistent with the 2D SOT trackers (e.g., OSTrack [73]) for more natural distillation of matching knowledge.", + "bbox": [ + 212, + 565, + 787, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We implement our framework with both Siamese [26, 27] and contextual memory-based [70] matching paradigms, which are respectively named SiamDisst and MemDisst. Extensive experiments demonstrate that our SiamDisst achieves significant improvements over its baseline, and MemDisst sets new state-of-the", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "2D Matching Distillation and 3D Pre-training", + "bbox": [ + 424, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "art performance on KITTI, while respectively running at a fast speed of 90 FPS on a single RTX3090 GPU.", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, the main contributions of our work are:", + "bbox": [ + 238, + 176, + 620, + 190 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- We build a general SOT framework that employs a standard ViT as the target matching module, which is consistent with dominant 2D trackers [62, 73] for better knowledge transfer of matching patterns, and meanwhile can naturally benefit from 3D pre-training.", + "bbox": [ + 225, + 191, + 782, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- We propose a lightweight target-aware projection (TAP) module to bridge the gap between the 2D and 3D tracking domains, which enables 2D pretrained trackers to be operated on point cloud data without additional 2D tracker fine-tuning.", + "bbox": [ + 223, + 253, + 782, + 310 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- We propose an IoU-guided matching distillation approach to guide 3D SOT learning, which facilitates transfer of knowledge from powerful 2D pre-trained foundation trackers to the 3D tracker. We show that the matching distillation facilitates the 3D model learning in the low-data regime.", + "bbox": [ + 225, + 313, + 782, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 398, + 385, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D Single Object Tracking. Inspired by the great success in the 2D tracking community, much progress has been made in 3D SOT [43]. The pioneering work SC3D [19] proposes the first 3D Siamese network to compute the similarity between the template and 3D candidate proposals. However, the time-consuming matching pipeline in SC3D is not end-to-end trainable, thus degrading its performance on 3D tracking benchmarks [5,18,53]. Inspired by the 2D region proposal network (RPN) [20,36], P2B [48] proposes a 3D RPN to generate high-quality 3D proposals, which achieves better tracking results while running at the real-time speed. BAT [77] extends P2B with a box-aware feature design, which provides a strong prior of the target shape. In contrast to previous approaches that directly regress 3D bboxes from a sparse point cloud, V2B [26] proposes a voxel-based localization head to localize the target in the Bird's Eye View (BEV).", + "bbox": [ + 212, + 431, + 784, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the development of transformers, 3D transformer SOT trackers (e.g., LTTR [8], PTTR [79] and CMT [21]) are designed for target-aware feature learning. Further extensions include iterative transformer blocks [27], historical cue modeling [33], contextual memory [69, 70], motion prediction [68, 78] and advanced architecture design [16, 64]. Despite these successes, learning robust 3D tracking models is still limited by the sparse and insufficient category-specific point cloud data. In this work, our aim is to leverage off-the-shelf 2D trackers trained on large-scale video data for enhancing 3D trackers, thus alleviating the current data bottleneck in 3D SOT.", + "bbox": [ + 212, + 613, + 784, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2D Single Object Tracking. Traditional 2D SOT [10,17,23,39,59,60] follow a correlation filter tracking framework due to its fast running speed and favorable tracking performance. With the development of the deep learning, much progress has been made in designing convolutional neural networks for video object tracking. Specifically, SiamFC [3] and SINT [50] employ a siamese neural network and treat tracking as a template matching problem. Based on SiamFC, many", + "bbox": [ + 212, + 750, + 784, + 838 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Q. Wu et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "improvements have been made, including architecture design [14,35,36,76], unsupervised representation learning [56,61,62] and template updating [4,72,75]. Recent advances using transformers [7,9,38,63,71,73] achieve state-of-the-art performance on popular tracking benchmarks [13,25,32,65]. The great success in 2D tracking significantly inspires the development in the 3D tracking community. However, there is no previous work trying to bridge the gap between the 2D and 3D tracking areas. In this paper, we propose a lightweight target-aware projection (TAP) module to bridge the gap between 2D and 3D SOT, and use the 2D tracker to effectively guide the 3D learning.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Knowledge Distillation. Knowledge distillation (KD) aims to transfer the knowledge from a teacher model to a student model. Previous response-based KD approaches [24,54,55,80] focus on using the final output of the teacher model to supervise the training of the student model. To supervise the intermediate feature learning, various feature-based KD approaches [2,12,28,37,45,51] have been proposed. FitNet [51] first uses the intermediate representations learned by the teacher as hints to distill a student model. However, the intermediate supervision is unsuitable in our cross-modality distillation case, since the 2D and 3D representations have a large gap. Unlike the above approaches, relation-based KD [34,44] aims to leverage the relationships between various samples or feature maps, e.g., FPS [74] uses the inner products of feature maps for distillation. However, naively applying feature correlation as supervision is not helpful in 2D-to-3D tracking, which is due to the transfer noise and causes performance degradation (see Table 2). In this paper, we propose a novel IoU-guided matching distillation approach for better 2D knowledge transfer, which is also the first attempt to perform cross-modality transfer for 3D SOT.", + "bbox": [ + 212, + 282, + 787, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 547, + 380, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first present our general 3D SOT framework with a standard ViT as the target relation modeling module, which could naturally benefit from 3D self-supervised pre-training and be consistent with 2D pre-trained trackers for better distillation. To use the 2D pre-trained model (frozen) for guiding 3D matching, we introduce a lightweight target-aware projection (TAP) module to bridge the domain gap between 2D and 3D tracking. Finally, we explore our IoU-guided matching distillation to boost the 3D tracking performance.", + "bbox": [ + 212, + 579, + 787, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Unified 3D SOT Framework", + "text_level": 1, + "bbox": [ + 215, + 705, + 495, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Revisit of 3D SOT. Given the 3D bbox in the first frame, the goal of 3D SOT is to accurately predict 3D bboxes in the following frames. Existing Siamese matching approaches [26,47,68,77] mainly use the point cloud (PC) data $\\hat{\\mathbf{p}}_1\\in$ $\\mathbb{R}^{\\hat{N}\\times 3}$ cropped in the first frame as the template to predict the 3D bbox $\\mathcal{B}_t\\in$ $\\mathbb{R}^7$ from the PC $\\mathbf{p}_t\\in \\mathbb{R}^{N_s\\times 3}$ in the $t$ -th search frame. $\\hat{N}$ is the number of sampled target points within the annotated 3D bounding box, $N_{s}$ is the number of points in each search frame, and $\\mathcal{B}_t$ is parameterized by its center coordinates", + "bbox": [ + 212, + 731, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "2D Matching Distillation and 3D Pre-training", + "bbox": [ + 424, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c2fc1c2d824f0e586e6b0dd457511e02af10688ec65dbe30f1eb28460669cb71.jpg", + "image_caption": [ + "Fig. 3: The pipeline converting 3D point clouds (PC) to 2D images using our target-aware projection (TAP) module, followed by the 2D pre-trained tracker." + ], + "image_footnote": [], + "bbox": [ + 246, + 145, + 751, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(x, y, and z), orientation $\\theta$ and box size (width, height and length). Recent approaches [69, 70] fully leverage contextual information and use more memory frames $\\{\\mathbf{p}_i\\}_{i = t - k}^t$ to predict $\\mathcal{B}_t$ , where $k$ is the number of memory frames. Instead of explicitly removing template points, a target mask set $\\{\\mathcal{U}_i\\}_{i = t - k - 1}^{t - 1}$ is used to identify the points within the 3D bbox, where $\\mathcal{U}_i \\in \\mathbb{R}^{N_s \\times 1}$ .", + "bbox": [ + 212, + 358, + 784, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Target matching module. For simplicity and consistency, we use $\\mathbf{m} \\in \\mathbb{R}^{N_m \\times 3}$ to represent the unified memory used in the above approaches, where $N_m = \\hat{N}$ in memory-less Siamese approaches and $N_m = N_s \\times k$ in memory approaches. We then obtain the search point features $\\mathbf{f}_t = \\phi(\\mathbf{p}_t) \\in \\mathbb{R}^{N_s \\times C}$ and memory point features $\\mathbf{f}_m = \\phi(\\mathbf{m}) \\in \\mathbb{R}^{N_m \\times C}$ by applying a point feature extractor $\\phi(\\cdot)$ on $\\mathbf{p}_t$ and $\\mathbf{m}$ respectively. The current 3D SOT trackers mainly focus on designing a robust 3D matching module $\\mathcal{M}(\\mathbf{f}_s, \\mathbf{f}_m)$ to perform the target matching between the search and memory point features. However, learning this matching module is challenging due to the sparsity of LIDAR point clouds and the lack of sufficient category-specific training data.", + "bbox": [ + 212, + 436, + 785, + 587 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To tackle these issues, inspired by 2D SOT [62, 73], we propose to use a standard ViT [11] for 3D matching since: 1) the standard ViT can benefit from 3D self-supervised pre-training (e.g., [42, 49]), which eases the need for large-scale 3D annotated data for 3D SOT training; 2) it has the consistent target relation modelling architecture (i.e., ViT) with dominant 2D trackers, which could more naturally transfer its 2D matching knowledge. The 3D matching process with ViT is formulated as:", + "bbox": [ + 212, + 588, + 785, + 691 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {f}} _ {t} = \\mathcal {M} \\left(\\mathbf {f} _ {t}, \\mathbf {f} _ {m} + \\theta (\\mathcal {U})\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 703, + 785, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the output $\\hat{\\mathbf{f}}_t\\in \\mathbb{R}^{N_s\\times C}$ is the updated search features output from the last layer of $\\mathcal{M}$ , $\\theta (\\cdot)$ is a learnable linear mapping that converts the 1-dim masks into the $C$ -dim feature maps. For siamese tracking paradigms, we drop this operation since the template is well cropped. Finally, the matching output $\\hat{\\mathbf{f}}_t$ is fed to the prediction head [26, 27, 70] for predicting the 3D bbox $\\mathcal{B}_t$ . Briefly, we call the memory-less Siamese-based and memory-based approaches with our standard ViT matching module as SiamBase and MemBase (see Supp. for more details).", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Q. Wu et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Target-Aware Projection Module", + "text_level": 1, + "bbox": [ + 215, + 146, + 535, + 161 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Although SiamBase and MemBase can naturally benefit from 3D self-supervised pre-training [42, 49], their performance is still limited due to the absence of large-scale 3D datasets (e.g., ShapeNet [6] only contains 55 object categories with 50k point clouds) for obtaining effective pre-training weights. To further improve the 3D tracking performance, we propose to employ a powerful 2D pre-trained tracker (denoted as $\\mathcal{T}_{2D}$ ) as the distillation teacher to guide the 3D matching learning in SiamBase and MemBase. However, in 3D SOT, only the single modality of point cloud data is available, making it hard to directly utilize current 2D SOT trackers. To bridge the modality gap between 2D and 3D tracking, we thus propose a learnable target-aware projection (TAP) module in this subsection, illustrated in Fig. 3.", + "bbox": [ + 212, + 171, + 787, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3D-to-2D Projection. Using the domain knowledge that the objects-of-interest for 3D tracking are typically on the ground plane (at similar fixed z-level heights), we project the search point cloud $\\mathbf{p}_t$ from the orthogonal view along the $z$ axis. For each point, we round down the $(x,y)$ coordinates to obtain its location on the 2D $xy$ plane, i.e., from the bird-eye view (BEV). We denote this 3D-to-2D projection as $\\psi (\\mathbf{p}_t)$ .", + "bbox": [ + 212, + 337, + 787, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Target-Aware Projection. Based on [58], the point features $\\mathbf{f}_t$ can be mapped to the 2D plane based on the corresponding projection $\\psi(\\mathbf{p}_t)$ . However, simply projecting search features to the 2D plane is not optimal since it does not consider any target information, which may not well support the subsequent 2D matching (see Table 2). Therefore, we use a cross-attention module $\\mathcal{C}$ here to obtain more effective target-aware search features:", + "bbox": [ + 212, + 429, + 787, + 518 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {t} ^ {*} = \\mathcal {C} \\left(Q (\\mathbf {f} _ {t}), K \\left(\\mathbf {f} _ {m} + \\theta (\\mathcal {U}) + E (\\mathbf {m})\\right), V \\left(\\mathbf {f} _ {m} + \\theta (\\mathcal {U}) + E (\\mathbf {m})\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 282, + 530, + 784, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $Q$ , $K$ and $V$ are query, key and value embedding modules, respectively. $E$ is the 3D positional embedding module. We project the target-aware search features $\\mathbf{f}_t^* \\in \\mathbb{R}^{N_s \\times C}$ into the 2D plane via $\\psi(\\mathbf{p}_t)$ and obtain the projected feature maps with $C$ channels, which are further input to a lightweight CNN $g$ [58] to reduce its channel dimension, yielding the projected search image:", + "bbox": [ + 214, + 558, + 787, + 635 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nI _ {t} = g \\left(\\psi \\left(\\mathbf {p} _ {t}\\right)\\right) \\in \\mathbb {R} ^ {H _ {t} \\times W _ {t} \\times 3}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 643, + 784, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $H_{t}$ and $W_{t}$ are height and width of the projected search image. The same operations are made on each memory frame and its corresponding features, thus we obtain $k$ projected memory images $\\{I_m^i\\}_{i = t - k - 1}^{t - 1}$ , where $I_{m}^{i}\\in \\mathbb{R}^{H_{m}\\times W_{m}\\times 3}$ .", + "bbox": [ + 212, + 672, + 787, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TAP training. The TAP module is trained via a simulated 2D tracking process on the projected search and target images, so that the projected images can obtain good 2D tracking performance. As illustrated in Fig. 3, the projected 2D search and memory images are input to a ViT-based pre-trained 2D tracker (e.g., OSTrack [73]) for 2D bounding box regression. Specifically, the input images are flattened into a sequence of 2D patches, projected into image embeddings, and then concatenated along the spatial dimension to input to the 2D ViT model for joint feature extraction and matching. The updated search features output", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "2D Matching Distillation and 3D Pre-training", + "bbox": [ + 424, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/88a89314ba2d9f4f2c421e04db7229384f3cc4c7cbf930f2882d196a77f85c9a.jpg", + "image_caption": [ + "Fig. 4: Visualization of the projected search image $I_{t}$ obtained by our TAP module on (a) KITTI-Pedestrian and (b) KITTI-Cycler. (c): The attention maps are generated via template-to-search matching from the 2D tracker on (b) KITTI-Cycler. The red box indicates the ground-truth (GT) 2D box projected from the GT 3D box. The blue box denotes the prediction of the 2D pre-trained OSTrack [73]. Based on the target-aware search features, the TAP module can effectively encode the target region with unique color and texture, enabling subsequent 2D matching." + ], + "image_footnote": [], + "bbox": [ + 303, + 143, + 702, + 295 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "from the last layer is input to the prediction head for 2D box regression, which can be generally formulated as:", + "bbox": [ + 212, + 445, + 787, + 477 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} _ {t} ^ {2 D}, \\hat {\\mathbf {f}} _ {t} ^ {2 D}, \\hat {\\mathbf {f}} _ {m} ^ {2 D} = \\mathcal {T} _ {2 D} (\\{I _ {m} ^ {i} \\} _ {i = t - k - 1} ^ {t - 1}, I _ {t}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 486, + 785, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{B}_t^{2D} \\in \\mathbb{R}^4$ is the predicted 2D bounding box in the projected search frame, and $(\\hat{\\mathbf{f}}_t^{2D}, \\hat{\\mathbf{f}}_m^{2D})$ are the updated 2D search and memory feature maps from the last layer of the 2D tracker. Here we use OSTrack as the 2D pre-trained tracker, and use its same 2D loss $\\mathcal{L}_{2D}$ [62, 73] to evaluate the simulated 2D tracking result, thus providing supervision on the TAP module. The ground-truth 2D bounding box $\\mathcal{G}_t^{2D}$ is obtained by projecting the corners of the ground-truth 3D box to the projected search image via the 3D-to-2D projection $\\psi$ .", + "bbox": [ + 212, + 516, + 787, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "After training the TAP module, the projection process is aligned with the 2D tracker. Note that we only update the parameters of the TAP module while keeping the parameters in the 2D pre-trained tracker $\\mathcal{T}_{2D}$ fixed, which makes the TAP training more efficient, and less likely to overfit. We also visualize the projected 2D search images in Fig. 4. The learned TAP module can support accurate 2D matching in the projected search image space well via the pretrained OSTrack model, which shows the effectiveness of the proposed TAP module.", + "bbox": [ + 212, + 625, + 789, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 IoU-guided Matching Distillation", + "text_level": 1, + "bbox": [ + 214, + 767, + 535, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce how to distill the 2D matching knowledge in $\\mathcal{T}_{2D}$ to guide the 3D matching learning in our 3D SOT trackers, SiamBase and MemBase. In (1), we obtain the updated 3D search features $\\hat{\\mathbf{f}}_t\\in \\mathbb{R}^{N_s\\times C}$ and updated 3D memory", + "bbox": [ + 212, + 792, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Q. Wu et al.", + "bbox": [ + 271, + 114, + 357, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2c13193de7125ae7c3a33e1a672f1fd4544ca0e92a9003c528ad14b4484b5258.jpg", + "image_caption": [ + "Fig. 5: Training pipeline of our SiamDisst or MemDisst via IoU-guided matching distillation. During the training stage, only the SiamDisst or MemDisst model is updated." + ], + "image_footnote": [], + "bbox": [ + 250, + 147, + 754, + 299 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "features $\\hat{\\mathbf{f}}_m\\in \\mathbb{R}^{N_m\\times C}$ output from the last layer of $\\mathcal{M}$ . Similarly, the updated 2D search and memory feature maps, $\\hat{\\mathbf{f}}_t^{2D}\\in \\mathbb{R}^{\\frac{H_t}{s}\\times \\frac{W_t}{s}\\times C}$ and $\\hat{\\mathbf{f}}_m^{2D}\\in \\mathbb{R}^{\\frac{H_m}{s}\\times \\frac{W_m}{s}\\times k\\times C}$ , are obtained via (4) in the last layer of 2D ViT, where $s$ denotes the patch size for patch embedding. The 3D and 2D features belong to different embedding spaces, and thus directly guiding the feature embeddings of the 3D tracker using the 2D tracker is not straightforward (see Supplementary). Instead, we use the matching correspondences (between feature maps) used by the 2D tracker to guide the matching correspondences (between feature vectors) of the 3D tracker. 3D-to-2D feature indexing. Given a search point coordinate in the search point cloud $\\mathbf{p}_t$ , we index its corresponding 2D feature vector via the 3D-to-2D projection $\\psi$ . Specifically, we first obtain its 2D location in the projected search image. Then, we interpolate its 2D features $\\hat{\\mathbf{f}}_t^{2D}$ to its original spatial size $H_{t}\\times W_{t}$ via bilinear interpolation, and finally index its feature vector via the projected 2D location. Based on the above 3D-to-2D feature indexing, we obtain the 2D search and memory features, denoted as $\\tilde{\\mathbf{f}}_t\\in \\mathbb{R}^{N_s\\times C}$ and $\\tilde{\\mathbf{f}}_m\\in \\mathbb{R}^{N_m\\times C}$ , that correspond to the 3D search and memory point clouds.", + "bbox": [ + 212, + 377, + 787, + 621 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Matching distillation. Note that $\\tilde{\\mathbf{f}}_t$ and $\\tilde{\\mathbf{f}}_m$ are obtained from the 2D ViT in the 2D tracker, which is well learned on large-scale 2D annotated videos, thus providing more robust matching results than our 3D counterparts (SimBase and MemBase) trained on limited category-specific point cloud data. Thus, we use the 2D memory-to-search matching pattern to guided the 3D matching process:", + "bbox": [ + 212, + 621, + 787, + 698 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d i s t} = \\mathrm {I o U} (\\mathcal {B} _ {t} ^ {2 D}, \\mathcal {G} _ {t} ^ {2 D}) \\cdot \\mathcal {L} _ {1} (\\hat {\\mathbf {f}} _ {t} \\hat {\\mathbf {f}} _ {m} ^ {T}, \\tilde {\\mathbf {f}} _ {t} \\tilde {\\mathbf {f}} _ {m} ^ {T}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 719, + 785, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{f}}_t\\hat{\\mathbf{f}}_m^T$ is the inner-product (matching) matrix between the 3D search and memory features, and likewise for $\\tilde{\\mathbf{f}}_t\\tilde{\\mathbf{f}}_m^T$ , and $\\mathcal{L}_1$ is the L1 loss that calculates the mean absolute error between two matrices. IoU $(\\cdot ,\\cdot)$ calculates the intersection-over-union between the predicted 2D box $\\mathcal{B}_t^{2D}$ and projected 2D ground-truth box $\\mathcal{G}_t^{2D}$ , which measures the reliability of our 2D tracker, i.e. if the 2D matching is reliable enough and predicts the 2D box accurately (i.e., with high IoU),", + "bbox": [ + 212, + 748, + 787, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "2D Matching Distillation and 3D Pre-training", + "bbox": [ + 424, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "this 2D matching pattern will be used more for guiding the 3D matching. We also use the typical 3D box regression losses $\\mathcal{L}_{3D}$ in the siamese paradigm (i.e., STNet [27]) and memory-based paradigm (i.e., MBPTrack [70]) to supervise our SiamBase and MemBase, respectively. The overall training loss is:", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {o v e r a l l}} = \\mathcal {L} _ {3 D} + \\lambda \\mathcal {L} _ {\\text {d i s t}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 220, + 782, + 234 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\lambda$ is a hyperparameter. After matching distillation, we denote the updated 3D SOT trackers as SiamDisst and MemDisst. Note that the TAP module, the 2D pre-trained tracker and the matching distillation are only used during the offline training stage to improve the 3D matching ability. During online inference, these modules are removed, and only our 3D trackers SiamDisst and MemDisst are required, which run at the real-time speed of 25 and 90 FPS, respectively.", + "bbox": [ + 212, + 248, + 782, + 338 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 363, + 374, + 380 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation details. For training the TAP module, as illustrated in Fig. 3, we sample memory and search PCs as the input, and project them into the 2D images for further matching in the 2D tracker. We use OSTrack [73] trained on large-scale 2D video datasets [13, 25, 40], as the 2D tracker, and freeze its weights, thus leading to efficient TAP training. The original 2D loss in OSTrack is used to evaluate the 2D predictions without modification. For TAP training, we use a batch size of 8 with a learning rate of 1e-4 and a total epoch of 40. For SiamDisst, we use the same hierarchical backbone network and 3D prediction head in STNet [27] for initial PC feature extraction and 3D prediction. For MemDisst, we use the same lightweight PointNet [46] with Point-MAE [42] as the backbone, and use the prediction head and memory mechanism in MBPTrack [70] for a fair comparison. We employ ViT-384-6 and ViT-96-3 as matching modules for SiamDisst and MemDisst respectively, which is illustrated in Table 3. The hyper-parameter $\\lambda$ is set to 1.0. The training and testing settings of SiamDisst and MemDisst are following their baselines STNet and MBPTrack. More details on implementation can be found in the Supplementary.", + "bbox": [ + 212, + 396, + 785, + 637 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets and Evaluation. We use three 3D datasets, KITTI [18], nuScenes [5] and Waymo Open Dataset (WOD) [53], for training and testing our proposed approaches. The KITTI dataset contains 21 training video sequences and 29 test sequences. Following previous work [19], the training set is split into three parts: 0-16 for training, 17-18 for validation and 19-20 for testing. Following [26,27,69], we use nuScenes and WOD to test the generalization ability of our approaches. We follow the one one pass evaluation [31], which is widely used in previous 3D SOT works [19,26,27,70], to report Precision and Success metrics for evaluation.", + "bbox": [ + 212, + 638, + 782, + 758 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 782, + 388, + 797 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Component analysis. We evaluate the effect of the two main components used in our proposed framework: 3D pre-training and 2D matching distillation.", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Q. Wu et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/6ab9b6b44367c9ed927f73de24060e7e916a15ccceadef6dde54e937f3a05149.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VariantBaseline3D Pre-training2D DistillationSuccessPrecision
SiamBase✓55.664.3
SiamBase✓60.369.3
SiamBase✓✓63.575.0
MemBase✓61.372.7
MemBase✓63.576.5
MemBase✓✓66.679.3
", + "bbox": [ + 248, + 143, + 751, + 243 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/98005893eebd8dfc6bdacf0fe13d44bd98a00d4289d61008835088d2b9dc056c.jpg", + "table_caption": [ + "Table 1: Component analysis of SiamDisst and MemDisst. 'Baseline' indicates the variant initialized with the random weights and trained from scratch." + ], + "table_footnote": [], + "table_body": "
VariantSucc./Prec.Pre-trainingSucc./Prec.
SiamDisst63.5/75.0Point-MAE [42]63.5/75.0
unlearnable TAP61.1/72.3ReCon [49]62.9/73.6
TAP w/o target-aware interact.61.2/73.12D MAE [22]60.5/69.9
2D distill. w/o IoU weight.56.4/66.3DropMAE [62]61.9/74.0
", + "bbox": [ + 238, + 299, + 751, + 373 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 2: Ablation studies of SiamDisst on KITTI-VAN: (left) variants of the TAP module and distillation; (right) using different 3D pre-trained models.", + "bbox": [ + 215, + 373, + 782, + 398 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The 3D pre-training means the usage of the default pre-trained Point-MAE [42] weights for 3D initialization. As shown in Table 1, SiamDisst and MemDisst with 3D pre-training achieve $5\\%$ and $3.8\\%$ improvements over the naive baselines regarding the precision metric. By applying the 2D matching distillation, SiamDisst and MemDisst can be significantly improved, leading to $5.7\\%$ and $2.8\\%$ precision gains, which shows the effectiveness of our approach.", + "bbox": [ + 212, + 444, + 784, + 535 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "TAP module. We study several variants of the proposed TAP module: including 1) unlearnable TAP: we remove the learnable projection module in the TAP, and use the z-axis value as the projected feature in the 2D image plane (repeated for each of the 3 image channels); 2) TAP module w/o target-aware design: we remove the cross attention module $\\mathcal{C}$ and separately project the template and search images; Table 2 (left) presents the results. The unlearnable TAP obtains inferior performance, which is due to the low-quality projected images and more 2D tracking failures. Without the IoU weighting, the unlearnable TAP also has severely degraded performance (see Supplementary for more details). For the target-aware design, we can observe $2.1\\% / 1.9\\%$ gains by using it, which shows that learning target-aware features is beneficial in the 2D projection stage.", + "bbox": [ + 212, + 536, + 785, + 700 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "IoU-guided matching distillation. In the IoU-guided matching distillation, we use the IoU to measure the reliability of the 2D prediction. To test its effectiveness, we remove this IoU weighting in (5) and show the comparison in Table 2 (left). The variant w/o the IoU weighting degrades the performance with large margins of $7.1\\% / 8.7\\%$ , which is mainly because the 2D matching distillation may introduce some low-quality and inaccurate 2D matching patterns for the 3D matching learning, thus severely degrading the learning.", + "bbox": [ + 212, + 703, + 784, + 808 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison of 3D pre-trained weights. We use various pre-trained weights to initialize the ViT in our 3D tracker and test them in Table 2 (right). The", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "2D Matching Distillation and 3D Pre-training", + "bbox": [ + 424, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/de469e7b2be04432a9667be91202ec219dc4bd50ff8ad012e09d34d25ab978c6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
methodmatching modulesucc./prec.FPS
SiamDisstViT-384-663.5/75.025
SiamDisstViT-96-360.5/71.730
MemDisstViT-384-666.7/79.178
MemDisstViT-96-366.6/79.390
", + "bbox": [ + 338, + 143, + 663, + 215 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 3: Comparisons of our approaches with various matching modules.", + "bbox": [ + 254, + 215, + 745, + 229 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "generative 3D Point-MAE pre-training provides the best result. Note that ReCon [49] adds the contrastive learning in the multi-modality generative pretraining, which may not be suitable in the pure point cloud tracking setting. We also test two typical 2D pre-trained models, i.e., MAE [22] and DropMAE [62] (see Supp. for details). Interestingly, our SiamDisst with DropMAE initialization achieves competitive results, which indicates that the temporal matching ability of DropMAE learned in 2D videos is also helpful for the 3D tracking task.", + "bbox": [ + 212, + 273, + 782, + 380 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Lightweight or heavy matching module. We explore various matching modules used in SiamDisst and MemDisst in Table 3. Note that ViT-384-6 is built from the first 6 layers of ViT-Base [11] with 384 hidden dimensions and 6 heads. We also design a lightweight ViT-96-3 model with 4 heads for comparison. We initialize ViT-384-6 with the official Point-MAE weights and rerun Point-MAE pre-training to initialize ViT-96-35. In Table 3, SiamDisst benefits more from the heavy ViT-384-6 while MemDisst achieves favorable performance even using the lightweight ViT-96-3. This is because SiamDisst has no effective online memory mechanism and heavily relies on its matching module. For MemDisst, the online memory provides more historical cues for matching, which eases the need for a heavy online matching module. Considering the speed-accuracy trade-off, we use ViT-384-6 and ViT-96-3 for SiamDisst and MemDisst, respectively.", + "bbox": [ + 212, + 380, + 784, + 561 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 State-of-the-art Comparison", + "text_level": 1, + "bbox": [ + 215, + 588, + 495, + 604 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We compare the proposed SiamDisst and MemDisst with state-of-the-art 3D trackers on the KITTI, nuScenes and WOD datasets. Specifically, for KITTI, comprehensive comparisons are made with two groups of trackers: 1) memoryless Siamese-based 3D trackers, SC3D [19], 3DSiamRPN [15], P2B [48], LTTR [8], MLVSNet [57], BAT [77], PTT [52], V2B [26], CMT [21], PTTR [79], and STNet [27]; and 2) contextual memory or motion prediction based 3D trackers, TAT [33], DMT [68], M2Track [78], CXTrack [69], and MBPTrack [70].", + "bbox": [ + 212, + 617, + 782, + 723 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "As shown in Table 4, our memory-less tracker SiamDisst outperforms the other Siamese 3D trackers by large margins in terms of both success and precision metrics. Specifically, the overall performance (Mean column) achieved by SiamDisst is $66.2\\% /83.9\\%$ , which has significant improvements $(4.9\\% /2.8\\%)$ over the baseline tracker STNet and demonstrates the effectiveness of our proposed", + "bbox": [ + 212, + 724, + 784, + 800 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Q. Wu et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 11 + }, + { + "type": "page_footnote", + "text": "5 We follow the same pre-training hyper-parameters, steps and dataset (ShapeNet [6]) with Point-MAE [42] for pre-training.", + "bbox": [ + 217, + 810, + 782, + 839 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/8cacb846a2267500f0734957ca36788daa5e9c2f657e086d1dd9b655e0a4c567.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4: 3D tracking results on KITTI [18] for (top) memory-less 3D trackers; (bottom) memory/motion-based 3D trackers. CM and MP represent the contextual memory and motion prediction. Trackers are evaluated by per-category Success/Precision metrics. 'Mean' indicates the overall result averaged over frames. The best result is shown in bold. For each category, the number in brackets indicates the test frame number." + ], + "table_body": "
MethodCM MPFPSCar (6424)Pedestrian (6088)Van (1248)Cyclist (308)Mean (14068)
SC3D [19]241.3/57.918.2/37.840.4/47.041.5/70.431.2/48.5
3DSiamRPN [15]2158.2/76.235.2/56.245.7/52.936.2/49.046.7/64.9
P2B [48]4656.2/72.828.7/49.640.8/48.432.1/44.742.4/60.0
LTTR [8]2365.0/77.133.2/56.835.8/45.666.2/89.948.7/65.8
MLVSNet [57]7056.0/74.034.1/61.152.0/61.434.3/44.545.7/66.7
BAT [77]6560.5/77.742.1/70.152.4/67.033.7/45.451.2/72.8
PTT [52]4067.8/81.844.9/72.043.6/52.537.2/47.355.1/74.2
V2B [26]3770.5/81.348.3/73.550.1/58.040.8/49.758.4/75.2
CMT [21]3270.5/81.949.1/75.554.1/64.155.1/82.459.4/77.6
PTTR [79]5365.2/77.450.9/81.652.5/61.865.1/90.557.9/78.1
STNet [27]3572.1/84.049.9/77.258.0/70.673.5/93.761.3/80.1
SiamDisst2573.7/85.158.4/83.963.5/75.076.9/94.566.2/83.9
TAT [33]✓-72.2/83.357.4/84.458.9/69.274.2/93.964.7/82.8
DMT [68]✓7266.4/79.448.1/77.953.3/65.670.4/92.655.1/75.8
M2Track [78]✓5765.5/80.861.5/88.253.8/70.773.2/93.562.9/83.4
CXTrack [69]✓3469.1/81.667.0/91.560.0/71.874.2/94.367.5/85.3
MBPTrack [70]✓5673.4/84.868.6/93.961.3/72.776.7/94.370.3/87.9
MemDisst✓9074.1/85.669.1/94.166.6/79.377.2/94.771.3/88.9
", + "bbox": [ + 217, + 143, + 802, + 422 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "IoU-guided matching distillation. Moreover, the proposed SiamDisst is more effective in identifying distractors during online tracking, e.g., obtaining best overall performance on the Pedestrian category. For MemDisst, it sets new state-of-the-art performance on all the categories while running at a fast speed of 90 FPS. The lightweight model used in MemDisst is well learned via the proposed learning framework, thus achieving promising tracking performance. Note that both SiamDisst and MemDisst performs well in the low data regime (e.g., the Van category only has 1994 training examples), which is mainly due to the effectiveness of transferring the matching knowledge from the 2D pre-trained tracker.", + "bbox": [ + 217, + 546, + 785, + 681 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To test the generalization performance of our proposed trackers, we use our SiamDisst and MemDisst trained on KITTI to evaluate on the WOD and nuScenes datasets. As seen in Table 5, SiamDisst performs best on both WOD and nuScenes among all the Siamese-based 3D trackers, e.g., STNet, V2B and BAT. This is mainly because our SiamDisst learns a more effective target matching module via the 2D matching distillation, thus leading to more accurate tracking results. Compared with MBPTrack, our MemDisst obtains comparable performance on WOD and outperforms it on nuScenes. Meanwhile, MemDisst runs efficiently at the above real-time speed of 89 FPS, which is $53\\%$ and $162\\%$ faster than MBPTrack and CXTrack, respectively. This lightweight MemDisst tracker", + "bbox": [ + 217, + 689, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "2D Matching Distillation and 3D Pre-training", + "bbox": [ + 424, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/5219c8ed0b4d9d6882b245966c56b7ef6ffd2a1511333e11780f06bb3f129c67.jpg", + "table_caption": [], + "table_footnote": [ + "Table 5: 3D tracking results on Waymo Open Dataset (WOD) [53] and nuScenes [5]. For a fair comparison, we compare our SiamDisst and MemDisst with memory-less Siamese-based trackers, and more complicated trackers with contexture memory or motion prediction, respectively. The best result in each comparison is shown in bold. * indicates the speed is re-evaluated on WOD for a fair comparison." + ], + "table_body": "
MethodWOD [53]nuScenes [5]FPS
Category (No. Frames)Vehicle (185731)Pedestrian (241752)Mean (427483)Car (15578)Pedestrian (8019)Truck (3710)Bicycle (501)Mean (27808)- -
SuccessP2B52.617.933.027.015.921.520.022.946
SC3D---25.014.225.717.021.82
BAT54.718.234.122.517.319.317.020.565*
V2B57.623.738.431.317.321.722.225.837
STNet59.725.540.432.219.122.321.226.935
SiamDisst61.228.542.733.419.825.017.728.125
CXTrack57.130.742.229.620.427.618.526.534
M2Track61.132.044.6-----57
MBPTrack61.933.746.033.619.823.920.028.160*
MemDisst61.933.645.934.020.028.118.128.989*
CategoryVehiclePedestrianMeanCarPedestrianTruckBicycleMeanFPS
PrecisionP2B61.730.143.829.222.016.226.425.346
SC3D---27.116.221.918.223.12
BAT62.730.344.424.124.515.818.823.065*
V2B65.937.950.135.123.416.719.129.037
STNet68.039.952.136.127.216.829.230.835
SiamDisst70.544.555.837.330.420.123.732.825
CXTrack66.149.456.733.432.920.826.831.534
M2Track69.349.758.2-----57
MBPTrack71.952.761.037.632.720.729.233.860*
MemDisst71.852.861.138.032.223.924.134.289*
", + "bbox": [ + 217, + 143, + 802, + 479 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "shows its potential to serve as a simple and strong baseline in the 3D tracking community.", + "bbox": [ + 215, + 590, + 787, + 621 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 217, + 643, + 359, + 659 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper improves the power of 3D point cloud tracking from two aspects. First, we design a novel 3D SOT framework that could fully benefit from the 3D self-supervised pre-training, which shows that the self-supervised pre-training weights (e.g., Point-MAE [42] and Recon [49]) are effective to improve the 3D tracking performance, especially in the low data regime (e.g., KITTI-Van and KITTI-Cyclist). Second, we bridge the gap between 2D and 3D tracking domains by leveraging a target-aware projection module, which enables to use a powerful 2D pre-trained tracker to facilitate the learning of 3D matching in existing 3D SOT trackers [27, 70] via matching distillation. We show that our proposed SiamDisst and MemDisst trackers can achieve state-of-the-art performance on KITTI and Waymo Open datasets, while running at real-time speed.", + "bbox": [ + 215, + 672, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Q. Wu et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 217, + 143, + 382, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by a Strategic Research Grant from City University of Hong Kong (Project No. 7005665).", + "bbox": [ + 215, + 178, + 787, + 209 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 233, + 321, + 250 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Anthes, C., García-Hernández, R.J., Wiedemann, M., Kranzlüller, D.: State of the art of virtual reality technology. In: 2016 IEEE aerospace conference. pp. 1-19. IEEE (2016) 1", + "2. Ben-Baruch, E., Karklinsky, M., Biton, Y., Ben-Cohen, A., Lawen, H., Zamir, N.: It's all in the head: Representation knowledge distillation through classifier sharing. In: arXiv:2201.06945 (2022) 5", + "3. Bertinetto, L., Valmadre, J., Henriques, J.F., Vedaldi, A., Torr, P.H.: Fully convolutional siamese networks for object tracking. In: Proceedings of the European Conference on Computer Vision. pp. 850-865. Springer (2016) 4", + "4. Bhat, G., Danelljan, M., Gool, L.V., Timofte, R.: Leanring discriminative model prediction for tracking. In: ICCV. pp. 6182-6191 (2019) 5", + "5. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11621-11631 (2020) 2, 4, 10, 14", + "6. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015) 7, 12", + "7. Chen, X., Yan, B., Zhu, J., Wang, D., Yang, X., Lu, H.: Transformer tracking. In: CVPR. pp. 8126-8135 (2021) 5", + "8. Cui, Y., Fang, Z., Shan, J., Gu, Z., Zhou, S.: 3d object tracking with transformer. arXiv preprint arXiv:2110.14921 (2021) 4, 12, 13", + "9. Cui, Y., Jiang, C., Wang, L., Wu, G.: Mixformer: End-to-end tracking with iterative mixed attention. In: CVPR (2022) 5", + "0. Danelljan, M., Bhat, G., Khan, F.S., Felsberg, M.: Eco: Efficient convolution operators for tracking. In: CVPR. pp. 21-26 (2017) 4", + "1. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arxiv 2020. arXiv preprint arXiv:2010.11929 (2010) 6, 12", + "2. Duong, C.N., Luu, K., Quach, K.G., Shrinkteanet, N.L.: Shrinkteanet: Millionscale lightweight face recognition via shrinking teacher-student networks. In: arXiv:1905.10620 (2019) 5", + "3. Fan, H., Lin, L., Yang, F.: Lasot: A high-quality benchmark for large-scale single object tracking. In: CVPR. pp. 5374-5383 (2019) 5, 10", + "4. Fan, H., Ling, H.: Siamese cascaded region proposal networks for real-time visual tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7952-7961 (2019) 5", + "5. Fang, Z., Zhou, S., Cui, Y., Scherer, S.: 3d-siamrpn: An end-to-end learning method for real-time 3d single object tracking using raw point cloud. IEEE Sensors Journal 21(4), 4995-5011 (2020) 12, 13" + ], + "bbox": [ + 225, + 266, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "2D Matching Distillation and 3D Pre-training", + "bbox": [ + 424, + 114, + 732, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Feng, S., Liang, P., Gao, J., Cheng, E.: Multi-correlation siamese transformer network with dense connection for 3d single object tracking. IEEE Robotics and Automation Letters (2023) 4", + "17. Galoogahi, H., Fagg, A., Lucey, S.: Learning background-aware correlation filters for visual tracking. In: ICCV (2017) 4", + "18. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. The International Journal of Robotics Research 32(11), 1231-1237 (2013) 2, 4, 10, 13", + "19. Giancola, S., Zarzar, J., Ghanem, B.: Leveraging shape completion for 3d siamese tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1359-1368 (2019) 4, 10, 12, 13", + "20. Girshick, R.: Fast r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 1440-1448 (2015) 4", + "21. Guo, Z., Mao, Y., Zhou, W., Wang, M., Li, H.: Cmt: Context-matching-guided transformer for 3d tracking in point clouds. In: European Conference on Computer Vision. pp. 95-111. Springer (2022) 4, 12, 13", + "22. He, K., Chen, X., Xie, S., Li, Y., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022) 2, 11, 12", + "23. Henriques, J.F., Caseiro, R., Martins, P., Batista, J.: High-speed tracking with kernelized correlation filters. In: IEEE Transactions on Pattern Analysis and Machine Intelligence. pp. 583-596 (2015) 4", + "24. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. In: arXiv:1503.02531 (2015) 5", + "25. Huang, L., Zhao, X., Huang, K.: Got-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence (2019) 5, 10", + "26. Hui, L., Wang, L., Cheng, M., Xie, J., Yang, J.: 3d siamese voxel-to-bev tracker for sparse point clouds. Advances in Neural Information Processing Systems 34, 28714-28727 (2021) 3, 4, 5, 6, 10, 12, 13", + "27. Hui, L., Wang, L., Tang, L., Lan, K., Xie, J., Yang, J.: 3d siamese transformer network for single object tracking on point clouds. In: European Conference on Computer Vision. pp. 293-310. Springer (2022) 2, 3, 4, 6, 10, 12, 13, 14", + "28. Jin, X., Peng, B., Wu, Y., Liu, Y., Liu, J.: Knowledge distillation via route constrained optimization. In: IEEE/CVF International Conference on Computer Vision (2019) 5", + "29. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017) 2", + "30. Kiran, B.R., Sobh, I., Talpaert, V., Mannion, P., Al Sallab, A.A., Yogamani, S., Pérez, P.: Deep reinforcement learning for autonomous driving: A survey. IEEE Transactions on Intelligent Transportation Systems (2021) 1", + "31. Kristan, M., Matas, J., Leonardis, A., Vojir, T., Pflugfelder, R., Fernandez, G., Nebehay, G., Porikli, F., Cehovin, L.: A novel performance evaluation methodology for single-target trackers. IEEE Transactions on Pattern Analysis and Machine Intelligence (2016) 10", + "32. Kristan, M., Matas, J., Danelljan, M.: The first visual object tracking segmentation votes2023 challenge results. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023) 5" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Q. Wu et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Lan, K., Jiang, H., Xie, J.: Temporal-aware siamese tracker: Integrate temporal context for 3d object tracking. In: Proceedings of the Asian Conference on Computer Vision. pp. 399-414 (2022) 2, 4, 12, 13", + "34. Lee, S.H., Kim, D.H., Song, B.C.: Self-supervised knowledge distillation using singular value decomposition. In: In Proceedings of the European Conference on Computer Vision (ECCV) (2018) 5", + "35. Li, B., Wu, W., Wang, Q., Zhang, F., Xing, J., Yan, J.: Siamrpn++: Evolution of siamese visual tracking with very deep networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4282-4291 (2019) 5", + "36. Li, B., Yan, J., Wu, W., Zhu, Z., Hu, X.: High performance visual tracking with siamese region proposal network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8971-8980 (2018) 4, 5", + "37. Li, J., Guo, Z., Li, H., Han, S., Baek, J.w., Yang, M., Yang, R., Suh, S.: Rethinking feature-based knowledge distillation for face recognition. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) 5", + "38. Lin, L., Fan, H., Zhang, Z., Xu, Y., Ling, H.: Swintrack: A simple and strong baseline for transformer tracking pp. 16743-16754 (2022) 2, 5", + "39. Liu, Y., Liang, Y., Wu, Q., Zhang, L., Wang, H.: A new framework for multiple deep correlation filters based object tracking. In: ICASSP (2022) 4", + "40. Muller, M., Bibi, A., S, G.: Trackingnet: A large-scale dataset and benchmark for object tracking in the wild. In: ECCV. pp. 300-317 (2018) 10", + "41. Osep, A., Mehner, W., Mathias, M., Leibe, B.: Combined image-and world-space tracking in traffic scenes. In: IEEE International Conference on Robotics and Automation. pp. 1988-1995. IEEE (2017) 1", + "42. Pang, Y., Wang, W., Tay, F.E., Liu, W., Tian, Y., Yuan, L.: Masked autoencoders for point cloud self-supervised learning. In: European conference on computer vision. pp. 604-621 (2022) 3, 6, 7, 10, 11, 12, 14", + "43. Pang, Z., Li, Z., Wang, N.: Model-free vehicle tracking and state estimation in point cloud sequences. In: IROS (2021) 4", + "44. Peng, B., Jin, X., Liu, J., Li, D., Wu, Y., Liu, Y., Zhou, S., Zhang, Z.: Self-supervised knowledge distillation using singular value decomposition. In: IEEE/CVF International Conference on Computer Vision (2019) 5", + "45. Peng, B., Jin, X., Liu, J., Li, D., Wu, Y., Liu, Y., Zhou, S., Zhang, Z.: Shrinkteanet: Million-scale lightweight face recognition via shrinking teacher-student networks. In: IEEE/CVF International Conference on Computer Vision (2019) 5", + "46. Qi, C., Su, H., Mo, K., Guibas, L.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: IEEE conference on computer vision and pattern recognition (2017) 10", + "47. Qi, H., Feng, C., Cao, Z., Zhao, F., Xiao, Y.: P2b: Point-to-box network for 3d object tracking in point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6329-6338 (2020) 2, 3, 5", + "48. Qi, H., Feng, C., Cao, Z., Zhao, F., Xiao, Y.: P2b: Point-to-box network for 3d object tracking in point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6329-6338 (2020) 4, 12, 13", + "49. Qi, Z., Dong, R., Fan, G., Ge, Z., Zhang, X., Ma, K., Yi, L.: Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining (2023) 3, 6, 7, 11, 12, 14", + "50. Ran, T., Gavves, E., Smeulders, A.W.: Siamese instance search for tracking. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1420-1429 (2016) 4" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "2D Matching Distillation and 3D Pre-training", + "bbox": [ + 424, + 114, + 733, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "51. Romero, A., Ballas, N.: Fitnets: Hints for thin deep nets. In: arXiv:1412.6550 (2014) 5", + "52. Shan, J., Zhou, S., Fang, Z., Cui, Y.: Ptt: Point-track-transformer module for 3d single object tracking in point clouds. arXiv preprint arXiv:2108.06455 (2021) 12, 13", + "53. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. p. 2446-2454 (2020) 4, 10, 14", + "54. Tao, F., Wang, M.: Response-based distillation for incremental object detection. In: arXiv:2110.13471 (2021) 5", + "55. Tao, F., Wang, M., Yuan, H.: Overcoming catastrophic forgetting in incremental object detection via elastic response distillation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022) 5", + "56. Wang, N., Song, Y., Ma, C.: Unsupervised deep tracking. In: CVPR. pp. 3708-1317 (2019) 5", + "57. Wang, Z., Xie, Q., Lai, Y.K., Wu, J., Long, K., Wang, J.: Mlvsnet: Multi-level voting siamese network for 3d visual tracking. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3101-3110 (2021) 12, 13", + "58. Wang, Z., Yu, X., Rao, Y., Zhou, J., Lu, J.: P2p: Tuning pre-trained image models for point cloud analysis with point-to-pixel prompting. Advances in neural information processing systems 35, 14388-14402 (2022) 7", + "59. Wu, Q., Chan, A.: Meta-graph adaptation for visual object tracking. In: ICME (2021) 4", + "60. Wu, Q., Yan, Y., Liang, Y., Liu, Y., Wang, H.: Dsnet: Deep and shallow feature learning for efficient visual tracking. In: ACCV. pp. 119-134 (2018) 4", + "61. Wu, Q., Wan, J., Chan, A.B.: Progressive unsupervised learning for visual object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2993-3002 (2021) 5", + "62. Wu, Q., Yang, T., Liu, Z., Wu, B., Shan, Y., Chan, A.B.: Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14561-14571 (2023) 1, 2, 4, 5, 6, 8, 11, 12", + "63. Wu, Q., Yang, T., Wu, W., Chan, A.B.: Scalable video object segmentation with simplified framework. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) 5", + "64. Wu, Q., Sun, C., Wang, J.: Multi-level structure-enhanced network for 3d single object tracking in sparse point clouds. IEEE Robotics and Automation Letters 8(1), 9-16 (2022) 4", + "65. Wu, Y., Lim, J., Yang, M.H.: Online object tracking: A benchmark. In: CVPR. pp. 2411-2418 (2013) 5", + "66. Xia, Y., Gladkova, M., Wang, R., Li, Q., Stilla, U., Henriques, J.F., Cremers, D.: Casspr: Cross attention single scan place recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8461-8472 (2023) 1", + "67. Xia, Y., Shi, L., Ding, Z., Henriques, J.F., Cremers, D.: Text2loc: 3d point cloud localization from natural language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14958-14967 (2024) 1", + "68. Xia, Y., Wu, Q., Li, W., Chan, A.B., Stilla, U.: A lightweight and detector-free 3d single object tracker on point clouds. IEEE Transactions on Intelligent Transportation Systems (2023) 4, 5, 12, 13" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Q. Wu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "69. Xu, T.X., Guo, Y.C., Lai, Y.K., Zhang, S.H.: Cxtrack: Improving 3d point cloud tracking with contextual information. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1084-1093 (2023) 4, 6, 10, 12, 13", + "70. Xu, T.X., Guo, Y.C., Lai, Y.K., Zhang, S.H.: Mbptrack: Improving 3d point cloud tracking with memory networks and box priors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9911-9920 (October 2023) 2, 3, 4, 6, 10, 12, 13, 14", + "71. Yan, B., Peng, H., Fu, J., Wang, D., Lu, H.: Learning spatio-temporal transformer for visual tracking. In: ICCV. pp. 10448-10457 (2021) 5", + "72. Yang, T., Chan, A.B.: Learning dynamic memory networks for object tracking. In: Proceedings of the European Conference on Computer Vision. pp. 152-167 (2018) 5", + "73. Ye, B., Chang, H., Ma, B., Shan, S., Chen, X.: Joint feature learning and relation modeling for tracking: A one-stream framework. In: European Conference on Computer Vision. pp. 341-357 (2022) 2, 3, 4, 5, 6, 7, 8, 10", + "74. Yim, J., Joo, D., Bae, J., Kim, J.: A gift from knowledge distillation: Fast optimization, network minimization and transfer learning. In: IEEE conference on computer vision and pattern recognition (2017) 5", + "75. Zhang, L., Gonzalez-Garcia, A., Weijer, J.v.d., Danelljan, M., Khan, F.S.: Learning the model update for siamese trackers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4010-4019 (2019) 5", + "76. Zhang, Z., Peng, H.: Deeper and wider siamese networks for real-time visual tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4591-4600 (2019) 5", + "77. Zheng, C., Yan, X., Gao, J., Zhao, W., Zhang, W., Li, Z., Cui, S.: Box-aware feature enhancement for single object tracking on point clouds. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13199-13208 (2021) 2, 3, 4, 5, 12, 13", + "78. Zheng, C., Yan, X., Zhang, H., Wang, B., Cheng, S., Cui, S., Li, Z.: Beyond 3d siamese tracking: A motion-centric paradigm for 3d single object tracking in point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8111-8120 (2022) 4, 12, 13", + "79. Zhou, C., Luo, Z., Luo, Y., Liu, T., Pan, L., Cai, Z., Zhao, H., Lu, S.: Pttr: Relational 3d point cloud object tracking with transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8531-8540 (2022) 4, 12, 13", + "80. Zhou, H., Song, L., Chen, J., Zhou, Y., Guoli: Rethinking soft labels for knowledge distillation: A bias-variance tradeoff perspective. In: arXiv:2102.00650 (2021) 5" + ], + "bbox": [ + 212, + 146, + 787, + 686 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "2D Matching Distillation and 3D Pre-training", + "bbox": [ + 424, + 114, + 733, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Boosting 3D Single Object Tracking with 2D Matching Distillation and 3D Pre-training/515a618d-9747-46e7-9c9a-8a152456fd1b_model.json b/2024/Boosting 3D Single Object Tracking with 2D Matching Distillation and 3D Pre-training/515a618d-9747-46e7-9c9a-8a152456fd1b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a698d8bd970154acb7a78115cfcb99b5b1626c71 --- /dev/null +++ b/2024/Boosting 3D Single Object Tracking with 2D Matching Distillation and 3D Pre-training/515a618d-9747-46e7-9c9a-8a152456fd1b_model.json @@ -0,0 +1,2537 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.246, + 0.141, + 0.756, + 0.187 + ], + "angle": 0, + "content": "Boosting 3D Single Object Tracking with 2D Matching Distillation and 3D Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.212, + 0.752, + 0.229 + ], + "angle": 0, + "content": "Qiangqiang \\(\\mathsf{W u}^{1}\\) , Yan Xia\\*2,3, Jia Wan4, and Antoni B. Chan1" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.239, + 0.725, + 0.253 + ], + "angle": 0, + "content": "\\(^{1}\\) Department of Computer Science, City University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.39, + 0.254, + 0.613, + 0.267 + ], + "angle": 0, + "content": "2 Technical University of Munich" + }, + { + "type": "text", + "bbox": [ + 0.339, + 0.268, + 0.664, + 0.281 + ], + "angle": 0, + "content": "3 Munich Center for Machine Learning (MCML)" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.282, + 0.762, + 0.308 + ], + "angle": 0, + "content": "\\(^{4}\\) School of Computer Science and Technology, Harbin Institute of Technology, Shenzhen" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.311, + 0.756, + 0.337 + ], + "angle": 0, + "content": "qiangqwu2-c@my.cityu.edu.hk, yan.xia@tum.de, jiawan1998@gmail.com, abchan@cityu.edu.hk" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.369, + 0.74, + 0.673 + ], + "angle": 0, + "content": "Abstract. 3D single object tracking (SOT) is an essential task in autonomous driving and robotics. However, learning robust 3D SOT trackers remains challenging due to the limited category-specific point cloud data and the inherent sparsity and incompleteness of LiDAR scans. To tackle these issues, we propose a unified 3D SOT framework that leverages 3D generative pre-training and learns robust 3D matching abilities from 2D pre-trained foundation trackers. Our framework features a consistent target-matching architecture with the widely used 2D trackers, facilitating the transfer of 2D matching knowledge. Specifically, we first propose a lightweight Target-Aware Projection (TAP) module, allowing the pre-trained 2D tracker to work well on the projected point clouds without further fine-tuning. We then propose a novel IoU-guided matching-distillation framework that utilizes the powerful 2D pre-trained trackers to guide 3D matching learning in the 3D tracker, i.e., the 3D template-to-search matching should be consistent with its corresponding 2D template-to-search matching obtained from 2D pre-trained trackers. Our designs are applied to two mainstream 3D SOT frameworks: memory-less Siamese and contextual memory-based approaches, which are respectively named SiamDisst and MemDisst. Extensive experiments show that SiamDisst and MemDisst achieve state-of-the-art performance on KITTI, Waymo Open Dataset and nuScenes benchmarks, while running at above real-time speed of 25 and 90 FPS on a RTX3090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.697, + 0.375, + 0.712 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.727, + 0.788, + 0.819 + ], + "angle": 0, + "content": "3D Single Object Tracking (SOT) has emerged as a basic 3D task for numerous practical applications in autonomous driving [30,67], robotics [41,66] and virtual reality [1]. Given a target described as a 3D bounding box (bbox) in the first frame, 3D SOT aims to track the target across all the frames in the 3D scene by predicting thebbox pose and position, which can be considered as a 3D extension of the 2D SOT task [62] but using a different input modality of 3D point clouds." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.386, + 0.84 + ], + "angle": 0, + "content": "* Corresponding Author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.129 + ], + "angle": 0, + "content": "Q. Wu et al." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.144, + 0.501, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.144, + 0.758, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.291, + 0.789, + 0.404 + ], + "angle": 0, + "content": "Fig. 1: Comparison of our SiamDisst and MemDisst to state-of-the-art 3D SOT on the Van category of the KITTI dataset [18]. Compared with Siamese-based trackers (plotted as squares), our SiamDisst achieves the best performance while running at the real-time speed of 25 FPS. Our lightweight MemDisst outperforms the other contextual memory and motion-based approaches (plotted as circles), setting new SOTA performance on KITTI-Van, and runs efficiently at 90 FPS. This demonstrates that our designs effectively boost 3D matching with limited category-specific training data (e.g., the KITTI-Van training set only contains 1994 examples)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.789, + 0.592 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, existing 3D SOT methods [38,62,73] mainly focus on learning robust 3D tracking models from sparse point cloud datasets [5,18]. For example, 3D siamese trackers [27,33,47,77] follow a siamese matching paradigm for accurate localization. To leverage more historical cues, MBPTrack [70] further extends this framework to a contextual memory-matching mechanism. Although the 3D matching models proposed by these methods play a key role in obtaining favorable performance [5,18], the sparsity and incompleteness of the point cloud training limits their abilities. Moreover, a paucity of large-scale category-specific 3D tracking datasets leads to insufficient training data for effectively learning the 3D matching modules in existing 3D trackers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.789, + 0.776 + ], + "angle": 0, + "content": "In contrast to the 3D SOT task, various large-scale annotated video datasets [29] are available for training 2D SOT methods. Existing 2D deep trackers can fully benefit from the feature backbones learned from 2D image [22] or video pre-training methods [62], enabling the learning of robust matching modules for 2D SOT on videos. The state-of-the-art trackers [38,62,73] with strong matching abilities show excellent tracking results even without using online memory mechanisms. With the remarkable success of 2D SOT, our work is motivated by the following questions: 1) Inspired by 2D SOT, can we design a general and effective 3D SOT framework that benefits from 3D pre-training? 2) Can we transfer the well-learned matching ability of state-of-the-art 2D trackers to 3D matching in 3D SOT trackers, so as to facilitate training from limited category-specific point cloud data?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.842 + ], + "angle": 0, + "content": "The main challenge of achieving the transfer of the 2D matching ability lies in two aspects. First, there is a significant misalignment between the input modalities for 2D and 3D trackers, dense 2d images and sparse 3d point clouds, respectively. Note that in the 3D SOT setting, only the single modality of point" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.425, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "2D Matching Distillation and 3D Pre-training" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.145, + 0.403, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.293, + 0.392, + 0.304 + ], + "angle": 0, + "content": "(a) Previous Siamese 3D SOT" + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.145, + 0.758, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.467, + 0.293, + 0.712, + 0.304 + ], + "angle": 0, + "content": "(b) Our Proposed 3D SOT (i.e., SiamDisst and MemDisst)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.327, + 0.788, + 0.453 + ], + "angle": 0, + "content": "Fig. 2: The comparison of 3D SOT frameworks (a) the previous Siamese 3D SOT trackers [27,47,77] focus on learning the 3D matching module from limited category-specific point cloud (PC) training data. (b) The proposed 3D SOT framework bridges the gap between 2D and 3D SOT by using our designed target-aware projection (TAP) module, which enables the 2D tracker [73] trained on large-scale 2D videos to operate well on the point clouds. The well-learned 2D matching pattern is further distilled to facilitate the 3D matching learning. Note that only the 3D-related modules (denoted as blue components) are optimized during the distillation training stage. The 2D tracker and TAP module are not used during online inference to increase speed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.564 + ], + "angle": 0, + "content": "cloud data is used as input, which makes it difficult to directly apply the current 2D SOT trackers. Second, how to effectively transfer the 2D matching knowledge in pre-trained 2D trackers to guide the learning of the 3D matching in 3D trackers is unclear given the differences in internal representations between 2D and 3D SOT, i.e., 2d dense feature maps and unordered 3D point embeddings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.778 + ], + "angle": 0, + "content": "To tackle the aforementioned challenges, we propose a general 3D SOT framework that can effectively utilize knowledge learned in 2D SOT, which is illustrated in Fig. 2(b). We first propose to align the different input modalities of 2D and 3D trackers via a learnable target-aware projection (TAP) module, which takes the raw point cloud as the input and projects them into the 2D image space. These projected images are suitable for 2D target matching and 2D bounding box regression by 2D pre-trained trackers, i.e., no fine-tuning of 2D trackers is required. Secondly, we propose a matching distillation approach that uses the knowledge in the high-quality 2D matching patterns from the 2D tracker as guidance for learning the matching patterns for 3D SOT. The proposed framework employs a standard Vision Transformer (ViT) as the matching module, which can naturally benefit from 3D pre-training approaches [42,49] and also creates a design that is consistent with the 2D SOT trackers (e.g., OSTrack [73]) for more natural distillation of matching knowledge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We implement our framework with both Siamese [26, 27] and contextual memory-based [70] matching paradigms, which are respectively named SiamDisst and MemDisst. Extensive experiments demonstrate that our SiamDisst achieves significant improvements over its baseline, and MemDisst sets new state-of-the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Q. Wu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "art performance on KITTI, while respectively running at a fast speed of 90 FPS on a single RTX3090 GPU." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.178, + 0.621, + 0.191 + ], + "angle": 0, + "content": "In summary, the main contributions of our work are:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.193, + 0.784, + 0.25 + ], + "angle": 0, + "content": "- We build a general SOT framework that employs a standard ViT as the target matching module, which is consistent with dominant 2D trackers [62, 73] for better knowledge transfer of matching patterns, and meanwhile can naturally benefit from 3D pre-training." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.254, + 0.784, + 0.311 + ], + "angle": 0, + "content": "- We propose a lightweight target-aware projection (TAP) module to bridge the gap between the 2D and 3D tracking domains, which enables 2D pretrained trackers to be operated on point cloud data without additional 2D tracker fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.314, + 0.784, + 0.372 + ], + "angle": 0, + "content": "- We propose an IoU-guided matching distillation approach to guide 3D SOT learning, which facilitates transfer of knowledge from powerful 2D pre-trained foundation trackers to the 3D tracker. We show that the matching distillation facilitates the 3D model learning in the low-data regime." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.399, + 0.387, + 0.415 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.785, + 0.612 + ], + "angle": 0, + "content": "3D Single Object Tracking. Inspired by the great success in the 2D tracking community, much progress has been made in 3D SOT [43]. The pioneering work SC3D [19] proposes the first 3D Siamese network to compute the similarity between the template and 3D candidate proposals. However, the time-consuming matching pipeline in SC3D is not end-to-end trainable, thus degrading its performance on 3D tracking benchmarks [5,18,53]. Inspired by the 2D region proposal network (RPN) [20,36], P2B [48] proposes a 3D RPN to generate high-quality 3D proposals, which achieves better tracking results while running at the real-time speed. BAT [77] extends P2B with a box-aware feature design, which provides a strong prior of the target shape. In contrast to previous approaches that directly regress 3D bboxes from a sparse point cloud, V2B [26] proposes a voxel-based localization head to localize the target in the Bird's Eye View (BEV)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.785, + 0.748 + ], + "angle": 0, + "content": "With the development of transformers, 3D transformer SOT trackers (e.g., LTTR [8], PTTR [79] and CMT [21]) are designed for target-aware feature learning. Further extensions include iterative transformer blocks [27], historical cue modeling [33], contextual memory [69, 70], motion prediction [68, 78] and advanced architecture design [16, 64]. Despite these successes, learning robust 3D tracking models is still limited by the sparse and insufficient category-specific point cloud data. In this work, our aim is to leverage off-the-shelf 2D trackers trained on large-scale video data for enhancing 3D trackers, thus alleviating the current data bottleneck in 3D SOT." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.839 + ], + "angle": 0, + "content": "2D Single Object Tracking. Traditional 2D SOT [10,17,23,39,59,60] follow a correlation filter tracking framework due to its fast running speed and favorable tracking performance. With the development of the deep learning, much progress has been made in designing convolutional neural networks for video object tracking. Specifically, SiamFC [3] and SINT [50] employ a siamese neural network and treat tracking as a template matching problem. Based on SiamFC, many" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.425, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "2D Matching Distillation and 3D Pre-training" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "improvements have been made, including architecture design [14,35,36,76], unsupervised representation learning [56,61,62] and template updating [4,72,75]. Recent advances using transformers [7,9,38,63,71,73] achieve state-of-the-art performance on popular tracking benchmarks [13,25,32,65]. The great success in 2D tracking significantly inspires the development in the 3D tracking community. However, there is no previous work trying to bridge the gap between the 2D and 3D tracking areas. In this paper, we propose a lightweight target-aware projection (TAP) module to bridge the gap between 2D and 3D SOT, and use the 2D tracker to effectively guide the 3D learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.525 + ], + "angle": 0, + "content": "Knowledge Distillation. Knowledge distillation (KD) aims to transfer the knowledge from a teacher model to a student model. Previous response-based KD approaches [24,54,55,80] focus on using the final output of the teacher model to supervise the training of the student model. To supervise the intermediate feature learning, various feature-based KD approaches [2,12,28,37,45,51] have been proposed. FitNet [51] first uses the intermediate representations learned by the teacher as hints to distill a student model. However, the intermediate supervision is unsuitable in our cross-modality distillation case, since the 2D and 3D representations have a large gap. Unlike the above approaches, relation-based KD [34,44] aims to leverage the relationships between various samples or feature maps, e.g., FPS [74] uses the inner products of feature maps for distillation. However, naively applying feature correlation as supervision is not helpful in 2D-to-3D tracking, which is due to the transfer noise and causes performance degradation (see Table 2). In this paper, we propose a novel IoU-guided matching distillation approach for better 2D knowledge transfer, which is also the first attempt to perform cross-modality transfer for 3D SOT." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.548, + 0.381, + 0.566 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.788, + 0.687 + ], + "angle": 0, + "content": "In this section, we first present our general 3D SOT framework with a standard ViT as the target relation modeling module, which could naturally benefit from 3D self-supervised pre-training and be consistent with 2D pre-trained trackers for better distillation. To use the 2D pre-trained model (frozen) for guiding 3D matching, we introduce a lightweight target-aware projection (TAP) module to bridge the domain gap between 2D and 3D tracking. Finally, we explore our IoU-guided matching distillation to boost the 3D tracking performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.707, + 0.496, + 0.723 + ], + "angle": 0, + "content": "3.1 Unified 3D SOT Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.732, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Revisit of 3D SOT. Given the 3D bbox in the first frame, the goal of 3D SOT is to accurately predict 3D bboxes in the following frames. Existing Siamese matching approaches [26,47,68,77] mainly use the point cloud (PC) data \\(\\hat{\\mathbf{p}}_1\\in\\) \\(\\mathbb{R}^{\\hat{N}\\times 3}\\) cropped in the first frame as the template to predict the 3D bbox \\(\\mathcal{B}_t\\in\\) \\(\\mathbb{R}^7\\) from the PC \\(\\mathbf{p}_t\\in \\mathbb{R}^{N_s\\times 3}\\) in the \\(t\\) -th search frame. \\(\\hat{N}\\) is the number of sampled target points within the annotated 3D bounding box, \\(N_{s}\\) is the number of points in each search frame, and \\(\\mathcal{B}_t\\) is parameterized by its center coordinates" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Q. Wu et al." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.146, + 0.753, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.301, + 0.785, + 0.331 + ], + "angle": 0, + "content": "Fig. 3: The pipeline converting 3D point clouds (PC) to 2D images using our target-aware projection (TAP) module, followed by the 2D pre-trained tracker." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.785, + 0.437 + ], + "angle": 0, + "content": "(x, y, and z), orientation \\(\\theta\\) and box size (width, height and length). Recent approaches [69, 70] fully leverage contextual information and use more memory frames \\(\\{\\mathbf{p}_i\\}_{i = t - k}^t\\) to predict \\(\\mathcal{B}_t\\), where \\(k\\) is the number of memory frames. Instead of explicitly removing template points, a target mask set \\(\\{\\mathcal{U}_i\\}_{i = t - k - 1}^{t - 1}\\) is used to identify the points within the 3D bbox, where \\(\\mathcal{U}_i \\in \\mathbb{R}^{N_s \\times 1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.437, + 0.787, + 0.588 + ], + "angle": 0, + "content": "Target matching module. For simplicity and consistency, we use \\(\\mathbf{m} \\in \\mathbb{R}^{N_m \\times 3}\\) to represent the unified memory used in the above approaches, where \\(N_m = \\hat{N}\\) in memory-less Siamese approaches and \\(N_m = N_s \\times k\\) in memory approaches. We then obtain the search point features \\(\\mathbf{f}_t = \\phi(\\mathbf{p}_t) \\in \\mathbb{R}^{N_s \\times C}\\) and memory point features \\(\\mathbf{f}_m = \\phi(\\mathbf{m}) \\in \\mathbb{R}^{N_m \\times C}\\) by applying a point feature extractor \\(\\phi(\\cdot)\\) on \\(\\mathbf{p}_t\\) and \\(\\mathbf{m}\\) respectively. The current 3D SOT trackers mainly focus on designing a robust 3D matching module \\(\\mathcal{M}(\\mathbf{f}_s, \\mathbf{f}_m)\\) to perform the target matching between the search and memory point features. However, learning this matching module is challenging due to the sparsity of LIDAR point clouds and the lack of sufficient category-specific training data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.589, + 0.787, + 0.693 + ], + "angle": 0, + "content": "To tackle these issues, inspired by 2D SOT [62, 73], we propose to use a standard ViT [11] for 3D matching since: 1) the standard ViT can benefit from 3D self-supervised pre-training (e.g., [42, 49]), which eases the need for large-scale 3D annotated data for 3D SOT training; 2) it has the consistent target relation modelling architecture (i.e., ViT) with dominant 2D trackers, which could more naturally transfer its 2D matching knowledge. The 3D matching process with ViT is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.418, + 0.704, + 0.786, + 0.722 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {f}} _ {t} = \\mathcal {M} \\left(\\mathbf {f} _ {t}, \\mathbf {f} _ {m} + \\theta (\\mathcal {U})\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.842 + ], + "angle": 0, + "content": "where the output \\(\\hat{\\mathbf{f}}_t\\in \\mathbb{R}^{N_s\\times C}\\) is the updated search features output from the last layer of \\(\\mathcal{M}\\), \\(\\theta (\\cdot)\\) is a learnable linear mapping that converts the 1-dim masks into the \\(C\\) -dim feature maps. For siamese tracking paradigms, we drop this operation since the template is well cropped. Finally, the matching output \\(\\hat{\\mathbf{f}}_t\\) is fed to the prediction head [26, 27, 70] for predicting the 3D bbox \\(\\mathcal{B}_t\\). Briefly, we call the memory-less Siamese-based and memory-based approaches with our standard ViT matching module as SiamBase and MemBase (see Supp. for more details)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.425, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "2D Matching Distillation and 3D Pre-training" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.536, + 0.162 + ], + "angle": 0, + "content": "3.2 Target-Aware Projection Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.337 + ], + "angle": 0, + "content": "Although SiamBase and MemBase can naturally benefit from 3D self-supervised pre-training [42, 49], their performance is still limited due to the absence of large-scale 3D datasets (e.g., ShapeNet [6] only contains 55 object categories with 50k point clouds) for obtaining effective pre-training weights. To further improve the 3D tracking performance, we propose to employ a powerful 2D pre-trained tracker (denoted as \\(\\mathcal{T}_{2D}\\)) as the distillation teacher to guide the 3D matching learning in SiamBase and MemBase. However, in 3D SOT, only the single modality of point cloud data is available, making it hard to directly utilize current 2D SOT trackers. To bridge the modality gap between 2D and 3D tracking, we thus propose a learnable target-aware projection (TAP) module in this subsection, illustrated in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.338, + 0.788, + 0.429 + ], + "angle": 0, + "content": "3D-to-2D Projection. Using the domain knowledge that the objects-of-interest for 3D tracking are typically on the ground plane (at similar fixed z-level heights), we project the search point cloud \\(\\mathbf{p}_t\\) from the orthogonal view along the \\(z\\) axis. For each point, we round down the \\((x,y)\\) coordinates to obtain its location on the 2D \\(xy\\) plane, i.e., from the bird-eye view (BEV). We denote this 3D-to-2D projection as \\(\\psi (\\mathbf{p}_t)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.519 + ], + "angle": 0, + "content": "Target-Aware Projection. Based on [58], the point features \\(\\mathbf{f}_t\\) can be mapped to the 2D plane based on the corresponding projection \\(\\psi(\\mathbf{p}_t)\\). However, simply projecting search features to the 2D plane is not optimal since it does not consider any target information, which may not well support the subsequent 2D matching (see Table 2). Therefore, we use a cross-attention module \\(\\mathcal{C}\\) here to obtain more effective target-aware search features:" + }, + { + "type": "equation", + "bbox": [ + 0.283, + 0.531, + 0.785, + 0.549 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {t} ^ {*} = \\mathcal {C} \\left(Q (\\mathbf {f} _ {t}), K \\left(\\mathbf {f} _ {m} + \\theta (\\mathcal {U}) + E (\\mathbf {m})\\right), V \\left(\\mathbf {f} _ {m} + \\theta (\\mathcal {U}) + E (\\mathbf {m})\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.559, + 0.788, + 0.636 + ], + "angle": 0, + "content": "where \\(Q\\), \\(K\\) and \\(V\\) are query, key and value embedding modules, respectively. \\(E\\) is the 3D positional embedding module. We project the target-aware search features \\(\\mathbf{f}_t^* \\in \\mathbb{R}^{N_s \\times C}\\) into the 2D plane via \\(\\psi(\\mathbf{p}_t)\\) and obtain the projected feature maps with \\(C\\) channels, which are further input to a lightweight CNN \\(g\\) [58] to reduce its channel dimension, yielding the projected search image:" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.645, + 0.785, + 0.663 + ], + "angle": 0, + "content": "\\[\nI _ {t} = g \\left(\\psi \\left(\\mathbf {p} _ {t}\\right)\\right) \\in \\mathbb {R} ^ {H _ {t} \\times W _ {t} \\times 3}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.72 + ], + "angle": 0, + "content": "where \\(H_{t}\\) and \\(W_{t}\\) are height and width of the projected search image. The same operations are made on each memory frame and its corresponding features, thus we obtain \\(k\\) projected memory images \\(\\{I_m^i\\}_{i = t - k - 1}^{t - 1}\\), where \\(I_{m}^{i}\\in \\mathbb{R}^{H_{m}\\times W_{m}\\times 3}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "TAP training. The TAP module is trained via a simulated 2D tracking process on the projected search and target images, so that the projected images can obtain good 2D tracking performance. As illustrated in Fig. 3, the projected 2D search and memory images are input to a ViT-based pre-trained 2D tracker (e.g., OSTrack [73]) for 2D bounding box regression. Specifically, the input images are flattened into a sequence of 2D patches, projected into image embeddings, and then concatenated along the spatial dimension to input to the 2D ViT model for joint feature extraction and matching. The updated search features output" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.129 + ], + "angle": 0, + "content": "Q. Wu et al." + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.145, + 0.703, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.319, + 0.789, + 0.419 + ], + "angle": 0, + "content": "Fig. 4: Visualization of the projected search image \\( I_{t} \\) obtained by our TAP module on (a) KITTI-Pedestrian and (b) KITTI-Cycler. (c): The attention maps are generated via template-to-search matching from the 2D tracker on (b) KITTI-Cycler. The red box indicates the ground-truth (GT) 2D box projected from the GT 3D box. The blue box denotes the prediction of the 2D pre-trained OSTrack [73]. Based on the target-aware search features, the TAP module can effectively encode the target region with unique color and texture, enabling subsequent 2D matching." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.788, + 0.478 + ], + "angle": 0, + "content": "from the last layer is input to the prediction head for 2D box regression, which can be generally formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.487, + 0.787, + 0.508 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} _ {t} ^ {2 D}, \\hat {\\mathbf {f}} _ {t} ^ {2 D}, \\hat {\\mathbf {f}} _ {m} ^ {2 D} = \\mathcal {T} _ {2 D} (\\{I _ {m} ^ {i} \\} _ {i = t - k - 1} ^ {t - 1}, I _ {t}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.517, + 0.789, + 0.625 + ], + "angle": 0, + "content": "where \\(\\mathcal{B}_t^{2D} \\in \\mathbb{R}^4\\) is the predicted 2D bounding box in the projected search frame, and \\((\\hat{\\mathbf{f}}_t^{2D}, \\hat{\\mathbf{f}}_m^{2D})\\) are the updated 2D search and memory feature maps from the last layer of the 2D tracker. Here we use OSTrack as the 2D pre-trained tracker, and use its same 2D loss \\(\\mathcal{L}_{2D}\\) [62, 73] to evaluate the simulated 2D tracking result, thus providing supervision on the TAP module. The ground-truth 2D bounding box \\(\\mathcal{G}_t^{2D}\\) is obtained by projecting the corners of the ground-truth 3D box to the projected search image via the 3D-to-2D projection \\(\\psi\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.79, + 0.748 + ], + "angle": 0, + "content": "After training the TAP module, the projection process is aligned with the 2D tracker. Note that we only update the parameters of the TAP module while keeping the parameters in the 2D pre-trained tracker \\(\\mathcal{T}_{2D}\\) fixed, which makes the TAP training more efficient, and less likely to overfit. We also visualize the projected 2D search images in Fig. 4. The learned TAP module can support accurate 2D matching in the projected search image space well via the pretrained OSTrack model, which shows the effectiveness of the proposed TAP module." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.768, + 0.537, + 0.785 + ], + "angle": 0, + "content": "3.3 IoU-guided Matching Distillation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.793, + 0.789, + 0.843 + ], + "angle": 0, + "content": "We introduce how to distill the 2D matching knowledge in \\(\\mathcal{T}_{2D}\\) to guide the 3D matching learning in our 3D SOT trackers, SiamBase and MemBase. In (1), we obtain the updated 3D search features \\(\\hat{\\mathbf{f}}_t\\in \\mathbb{R}^{N_s\\times C}\\) and updated 3D memory" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.425, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "2D Matching Distillation and 3D Pre-training" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.148, + 0.756, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.323, + 0.785, + 0.352 + ], + "angle": 0, + "content": "Fig. 5: Training pipeline of our SiamDisst or MemDisst via IoU-guided matching distillation. During the training stage, only the SiamDisst or MemDisst model is updated." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.378, + 0.788, + 0.622 + ], + "angle": 0, + "content": "features \\(\\hat{\\mathbf{f}}_m\\in \\mathbb{R}^{N_m\\times C}\\) output from the last layer of \\(\\mathcal{M}\\). Similarly, the updated 2D search and memory feature maps, \\(\\hat{\\mathbf{f}}_t^{2D}\\in \\mathbb{R}^{\\frac{H_t}{s}\\times \\frac{W_t}{s}\\times C}\\) and \\(\\hat{\\mathbf{f}}_m^{2D}\\in \\mathbb{R}^{\\frac{H_m}{s}\\times \\frac{W_m}{s}\\times k\\times C}\\), are obtained via (4) in the last layer of 2D ViT, where \\(s\\) denotes the patch size for patch embedding. The 3D and 2D features belong to different embedding spaces, and thus directly guiding the feature embeddings of the 3D tracker using the 2D tracker is not straightforward (see Supplementary). Instead, we use the matching correspondences (between feature maps) used by the 2D tracker to guide the matching correspondences (between feature vectors) of the 3D tracker. 3D-to-2D feature indexing. Given a search point coordinate in the search point cloud \\(\\mathbf{p}_t\\), we index its corresponding 2D feature vector via the 3D-to-2D projection \\(\\psi\\). Specifically, we first obtain its 2D location in the projected search image. Then, we interpolate its 2D features \\(\\hat{\\mathbf{f}}_t^{2D}\\) to its original spatial size \\(H_{t}\\times W_{t}\\) via bilinear interpolation, and finally index its feature vector via the projected 2D location. Based on the above 3D-to-2D feature indexing, we obtain the 2D search and memory features, denoted as \\(\\tilde{\\mathbf{f}}_t\\in \\mathbb{R}^{N_s\\times C}\\) and \\(\\tilde{\\mathbf{f}}_m\\in \\mathbb{R}^{N_m\\times C}\\), that correspond to the 3D search and memory point clouds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.622, + 0.788, + 0.699 + ], + "angle": 0, + "content": "Matching distillation. Note that \\(\\tilde{\\mathbf{f}}_t\\) and \\(\\tilde{\\mathbf{f}}_m\\) are obtained from the 2D ViT in the 2D tracker, which is well learned on large-scale 2D annotated videos, thus providing more robust matching results than our 3D counterparts (SimBase and MemBase) trained on limited category-specific point cloud data. Thus, we use the 2D memory-to-search matching pattern to guided the 3D matching process:" + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.72, + 0.787, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d i s t} = \\mathrm {I o U} (\\mathcal {B} _ {t} ^ {2 D}, \\mathcal {G} _ {t} ^ {2 D}) \\cdot \\mathcal {L} _ {1} (\\hat {\\mathbf {f}} _ {t} \\hat {\\mathbf {f}} _ {m} ^ {T}, \\tilde {\\mathbf {f}} _ {t} \\tilde {\\mathbf {f}} _ {m} ^ {T}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{f}}_t\\hat{\\mathbf{f}}_m^T\\) is the inner-product (matching) matrix between the 3D search and memory features, and likewise for \\(\\tilde{\\mathbf{f}}_t\\tilde{\\mathbf{f}}_m^T\\), and \\(\\mathcal{L}_1\\) is the L1 loss that calculates the mean absolute error between two matrices. IoU \\((\\cdot ,\\cdot)\\) calculates the intersection-over-union between the predicted 2D box \\(\\mathcal{B}_t^{2D}\\) and projected 2D ground-truth box \\(\\mathcal{G}_t^{2D}\\), which measures the reliability of our 2D tracker, i.e. if the 2D matching is reliable enough and predicts the 2D box accurately (i.e., with high IoU)," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Q. Wu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.207 + ], + "angle": 0, + "content": "this 2D matching pattern will be used more for guiding the 3D matching. We also use the typical 3D box regression losses \\(\\mathcal{L}_{3D}\\) in the siamese paradigm (i.e., STNet [27]) and memory-based paradigm (i.e., MBPTrack [70]) to supervise our SiamBase and MemBase, respectively. The overall training loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.222, + 0.784, + 0.235 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {o v e r a l l}} = \\mathcal {L} _ {3 D} + \\lambda \\mathcal {L} _ {\\text {d i s t}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.249, + 0.784, + 0.339 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a hyperparameter. After matching distillation, we denote the updated 3D SOT trackers as SiamDisst and MemDisst. Note that the TAP module, the 2D pre-trained tracker and the matching distillation are only used during the offline training stage to improve the 3D matching ability. During online inference, these modules are removed, and only our 3D trackers SiamDisst and MemDisst are required, which run at the real-time speed of 25 and 90 FPS, respectively." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.364, + 0.375, + 0.381 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.787, + 0.638 + ], + "angle": 0, + "content": "Implementation details. For training the TAP module, as illustrated in Fig. 3, we sample memory and search PCs as the input, and project them into the 2D images for further matching in the 2D tracker. We use OSTrack [73] trained on large-scale 2D video datasets [13, 25, 40], as the 2D tracker, and freeze its weights, thus leading to efficient TAP training. The original 2D loss in OSTrack is used to evaluate the 2D predictions without modification. For TAP training, we use a batch size of 8 with a learning rate of 1e-4 and a total epoch of 40. For SiamDisst, we use the same hierarchical backbone network and 3D prediction head in STNet [27] for initial PC feature extraction and 3D prediction. For MemDisst, we use the same lightweight PointNet [46] with Point-MAE [42] as the backbone, and use the prediction head and memory mechanism in MBPTrack [70] for a fair comparison. We employ ViT-384-6 and ViT-96-3 as matching modules for SiamDisst and MemDisst respectively, which is illustrated in Table 3. The hyper-parameter \\(\\lambda\\) is set to 1.0. The training and testing settings of SiamDisst and MemDisst are following their baselines STNet and MBPTrack. More details on implementation can be found in the Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.784, + 0.759 + ], + "angle": 0, + "content": "Datasets and Evaluation. We use three 3D datasets, KITTI [18], nuScenes [5] and Waymo Open Dataset (WOD) [53], for training and testing our proposed approaches. The KITTI dataset contains 21 training video sequences and 29 test sequences. Following previous work [19], the training set is split into three parts: 0-16 for training, 17-18 for validation and 19-20 for testing. Following [26,27,69], we use nuScenes and WOD to test the generalization ability of our approaches. We follow the one one pass evaluation [31], which is widely used in previous 3D SOT works [19,26,27,70], to report Precision and Success metrics for evaluation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.784, + 0.389, + 0.799 + ], + "angle": 0, + "content": "4.1 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Component analysis. We evaluate the effect of the two main components used in our proposed framework: 3D pre-training and 2D matching distillation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.425, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "2D Matching Distillation and 3D Pre-training" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.249, + 0.144, + 0.753, + 0.244 + ], + "angle": 0, + "content": "
VariantBaseline3D Pre-training2D DistillationSuccessPrecision
SiamBase✓55.664.3
SiamBase✓60.369.3
SiamBase✓✓63.575.0
MemBase✓61.372.7
MemBase✓63.576.5
MemBase✓✓66.679.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.245, + 0.785, + 0.271 + ], + "angle": 0, + "content": "Table 1: Component analysis of SiamDisst and MemDisst. 'Baseline' indicates the variant initialized with the random weights and trained from scratch." + }, + { + "type": "table", + "bbox": [ + 0.24, + 0.3, + 0.753, + 0.374 + ], + "angle": 0, + "content": "
VariantSucc./Prec.Pre-trainingSucc./Prec.
SiamDisst63.5/75.0Point-MAE [42]63.5/75.0
unlearnable TAP61.1/72.3ReCon [49]62.9/73.6
TAP w/o target-aware interact.61.2/73.12D MAE [22]60.5/69.9
2D distill. w/o IoU weight.56.4/66.3DropMAE [62]61.9/74.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.374, + 0.784, + 0.4 + ], + "angle": 0, + "content": "Table 2: Ablation studies of SiamDisst on KITTI-VAN: (left) variants of the TAP module and distillation; (right) using different 3D pre-trained models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.785, + 0.536 + ], + "angle": 0, + "content": "The 3D pre-training means the usage of the default pre-trained Point-MAE [42] weights for 3D initialization. As shown in Table 1, SiamDisst and MemDisst with 3D pre-training achieve \\(5\\%\\) and \\(3.8\\%\\) improvements over the naive baselines regarding the precision metric. By applying the 2D matching distillation, SiamDisst and MemDisst can be significantly improved, leading to \\(5.7\\%\\) and \\(2.8\\%\\) precision gains, which shows the effectiveness of our approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.787, + 0.702 + ], + "angle": 0, + "content": "TAP module. We study several variants of the proposed TAP module: including 1) unlearnable TAP: we remove the learnable projection module in the TAP, and use the z-axis value as the projected feature in the 2D image plane (repeated for each of the 3 image channels); 2) TAP module w/o target-aware design: we remove the cross attention module \\(\\mathcal{C}\\) and separately project the template and search images; Table 2 (left) presents the results. The unlearnable TAP obtains inferior performance, which is due to the low-quality projected images and more 2D tracking failures. Without the IoU weighting, the unlearnable TAP also has severely degraded performance (see Supplementary for more details). For the target-aware design, we can observe \\(2.1\\% / 1.9\\%\\) gains by using it, which shows that learning target-aware features is beneficial in the 2D projection stage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.785, + 0.809 + ], + "angle": 0, + "content": "IoU-guided matching distillation. In the IoU-guided matching distillation, we use the IoU to measure the reliability of the 2D prediction. To test its effectiveness, we remove this IoU weighting in (5) and show the comparison in Table 2 (left). The variant w/o the IoU weighting degrades the performance with large margins of \\(7.1\\% / 8.7\\%\\), which is mainly because the 2D matching distillation may introduce some low-quality and inaccurate 2D matching patterns for the 3D matching learning, thus severely degrading the learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Comparison of 3D pre-trained weights. We use various pre-trained weights to initialize the ViT in our 3D tracker and test them in Table 2 (right). The" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Q. Wu et al." + }, + { + "type": "table", + "bbox": [ + 0.339, + 0.144, + 0.665, + 0.216 + ], + "angle": 0, + "content": "
methodmatching modulesucc./prec.FPS
SiamDisstViT-384-663.5/75.025
SiamDisstViT-96-360.5/71.730
MemDisstViT-384-666.7/79.178
MemDisstViT-96-366.6/79.390
" + }, + { + "type": "table_caption", + "bbox": [ + 0.255, + 0.217, + 0.746, + 0.23 + ], + "angle": 0, + "content": "Table 3: Comparisons of our approaches with various matching modules." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.275, + 0.784, + 0.381 + ], + "angle": 0, + "content": "generative 3D Point-MAE pre-training provides the best result. Note that ReCon [49] adds the contrastive learning in the multi-modality generative pretraining, which may not be suitable in the pure point cloud tracking setting. We also test two typical 2D pre-trained models, i.e., MAE [22] and DropMAE [62] (see Supp. for details). Interestingly, our SiamDisst with DropMAE initialization achieves competitive results, which indicates that the temporal matching ability of DropMAE learned in 2D videos is also helpful for the 3D tracking task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.785, + 0.563 + ], + "angle": 0, + "content": "Lightweight or heavy matching module. We explore various matching modules used in SiamDisst and MemDisst in Table 3. Note that ViT-384-6 is built from the first 6 layers of ViT-Base [11] with 384 hidden dimensions and 6 heads. We also design a lightweight ViT-96-3 model with 4 heads for comparison. We initialize ViT-384-6 with the official Point-MAE weights and rerun Point-MAE pre-training to initialize ViT-96-35. In Table 3, SiamDisst benefits more from the heavy ViT-384-6 while MemDisst achieves favorable performance even using the lightweight ViT-96-3. This is because SiamDisst has no effective online memory mechanism and heavily relies on its matching module. For MemDisst, the online memory provides more historical cues for matching, which eases the need for a heavy online matching module. Considering the speed-accuracy trade-off, we use ViT-384-6 and ViT-96-3 for SiamDisst and MemDisst, respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.589, + 0.496, + 0.605 + ], + "angle": 0, + "content": "4.2 State-of-the-art Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.618, + 0.784, + 0.724 + ], + "angle": 0, + "content": "We compare the proposed SiamDisst and MemDisst with state-of-the-art 3D trackers on the KITTI, nuScenes and WOD datasets. Specifically, for KITTI, comprehensive comparisons are made with two groups of trackers: 1) memoryless Siamese-based 3D trackers, SC3D [19], 3DSiamRPN [15], P2B [48], LTTR [8], MLVSNet [57], BAT [77], PTT [52], V2B [26], CMT [21], PTTR [79], and STNet [27]; and 2) contextual memory or motion prediction based 3D trackers, TAT [33], DMT [68], M2Track [78], CXTrack [69], and MBPTrack [70]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.725, + 0.785, + 0.801 + ], + "angle": 0, + "content": "As shown in Table 4, our memory-less tracker SiamDisst outperforms the other Siamese 3D trackers by large margins in terms of both success and precision metrics. Specifically, the overall performance (Mean column) achieved by SiamDisst is \\(66.2\\% /83.9\\%\\), which has significant improvements \\((4.9\\% /2.8\\%)\\) over the baseline tracker STNet and demonstrates the effectiveness of our proposed" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.783, + 0.84 + ], + "angle": 0, + "content": "5 We follow the same pre-training hyper-parameters, steps and dataset (ShapeNet [6]) with Point-MAE [42] for pre-training." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.425, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "2D Matching Distillation and 3D Pre-training" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.144, + 0.803, + 0.424 + ], + "angle": 0, + "content": "
MethodCM MPFPSCar (6424)Pedestrian (6088)Van (1248)Cyclist (308)Mean (14068)
SC3D [19]241.3/57.918.2/37.840.4/47.041.5/70.431.2/48.5
3DSiamRPN [15]2158.2/76.235.2/56.245.7/52.936.2/49.046.7/64.9
P2B [48]4656.2/72.828.7/49.640.8/48.432.1/44.742.4/60.0
LTTR [8]2365.0/77.133.2/56.835.8/45.666.2/89.948.7/65.8
MLVSNet [57]7056.0/74.034.1/61.152.0/61.434.3/44.545.7/66.7
BAT [77]6560.5/77.742.1/70.152.4/67.033.7/45.451.2/72.8
PTT [52]4067.8/81.844.9/72.043.6/52.537.2/47.355.1/74.2
V2B [26]3770.5/81.348.3/73.550.1/58.040.8/49.758.4/75.2
CMT [21]3270.5/81.949.1/75.554.1/64.155.1/82.459.4/77.6
PTTR [79]5365.2/77.450.9/81.652.5/61.865.1/90.557.9/78.1
STNet [27]3572.1/84.049.9/77.258.0/70.673.5/93.761.3/80.1
SiamDisst2573.7/85.158.4/83.963.5/75.076.9/94.566.2/83.9
TAT [33]✓-72.2/83.357.4/84.458.9/69.274.2/93.964.7/82.8
DMT [68]✓7266.4/79.448.1/77.953.3/65.670.4/92.655.1/75.8
M2Track [78]✓5765.5/80.861.5/88.253.8/70.773.2/93.562.9/83.4
CXTrack [69]✓3469.1/81.667.0/91.560.0/71.874.2/94.367.5/85.3
MBPTrack [70]✓5673.4/84.868.6/93.961.3/72.776.7/94.370.3/87.9
MemDisst✓9074.1/85.669.1/94.166.6/79.377.2/94.771.3/88.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.493 + ], + "angle": 0, + "content": "Table 4: 3D tracking results on KITTI [18] for (top) memory-less 3D trackers; (bottom) memory/motion-based 3D trackers. CM and MP represent the contextual memory and motion prediction. Trackers are evaluated by per-category Success/Precision metrics. 'Mean' indicates the overall result averaged over frames. The best result is shown in bold. For each category, the number in brackets indicates the test frame number." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.547, + 0.786, + 0.683 + ], + "angle": 0, + "content": "IoU-guided matching distillation. Moreover, the proposed SiamDisst is more effective in identifying distractors during online tracking, e.g., obtaining best overall performance on the Pedestrian category. For MemDisst, it sets new state-of-the-art performance on all the categories while running at a fast speed of 90 FPS. The lightweight model used in MemDisst is well learned via the proposed learning framework, thus achieving promising tracking performance. Note that both SiamDisst and MemDisst performs well in the low data regime (e.g., the Van category only has 1994 training examples), which is mainly due to the effectiveness of transferring the matching knowledge from the 2D pre-trained tracker." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.69, + 0.786, + 0.84 + ], + "angle": 0, + "content": "To test the generalization performance of our proposed trackers, we use our SiamDisst and MemDisst trained on KITTI to evaluate on the WOD and nuScenes datasets. As seen in Table 5, SiamDisst performs best on both WOD and nuScenes among all the Siamese-based 3D trackers, e.g., STNet, V2B and BAT. This is mainly because our SiamDisst learns a more effective target matching module via the 2D matching distillation, thus leading to more accurate tracking results. Compared with MBPTrack, our MemDisst obtains comparable performance on WOD and outperforms it on nuScenes. Meanwhile, MemDisst runs efficiently at the above real-time speed of 89 FPS, which is \\(53\\%\\) and \\(162\\%\\) faster than MBPTrack and CXTrack, respectively. This lightweight MemDisst tracker" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Q. Wu et al." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.144, + 0.803, + 0.481 + ], + "angle": 0, + "content": "
MethodWOD [53]nuScenes [5]FPS
Category (No. Frames)Vehicle (185731)Pedestrian (241752)Mean (427483)Car (15578)Pedestrian (8019)Truck (3710)Bicycle (501)Mean (27808)- -
SuccessP2B52.617.933.027.015.921.520.022.946
SC3D---25.014.225.717.021.82
BAT54.718.234.122.517.319.317.020.565*
V2B57.623.738.431.317.321.722.225.837
STNet59.725.540.432.219.122.321.226.935
SiamDisst61.228.542.733.419.825.017.728.125
CXTrack57.130.742.229.620.427.618.526.534
M2Track61.132.044.6-----57
MBPTrack61.933.746.033.619.823.920.028.160*
MemDisst61.933.645.934.020.028.118.128.989*
CategoryVehiclePedestrianMeanCarPedestrianTruckBicycleMeanFPS
PrecisionP2B61.730.143.829.222.016.226.425.346
SC3D---27.116.221.918.223.12
BAT62.730.344.424.124.515.818.823.065*
V2B65.937.950.135.123.416.719.129.037
STNet68.039.952.136.127.216.829.230.835
SiamDisst70.544.555.837.330.420.123.732.825
CXTrack66.149.456.733.432.920.826.831.534
M2Track69.349.758.2-----57
MBPTrack71.952.761.037.632.720.729.233.860*
MemDisst71.852.861.138.032.223.924.134.289*
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.217, + 0.482, + 0.788, + 0.551 + ], + "angle": 0, + "content": "Table 5: 3D tracking results on Waymo Open Dataset (WOD) [53] and nuScenes [5]. For a fair comparison, we compare our SiamDisst and MemDisst with memory-less Siamese-based trackers, and more complicated trackers with contexture memory or motion prediction, respectively. The best result in each comparison is shown in bold. * indicates the speed is re-evaluated on WOD for a fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.591, + 0.788, + 0.622 + ], + "angle": 0, + "content": "shows its potential to serve as a simple and strong baseline in the 3D tracking community." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.644, + 0.36, + 0.66 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This paper improves the power of 3D point cloud tracking from two aspects. First, we design a novel 3D SOT framework that could fully benefit from the 3D self-supervised pre-training, which shows that the self-supervised pre-training weights (e.g., Point-MAE [42] and Recon [49]) are effective to improve the 3D tracking performance, especially in the low data regime (e.g., KITTI-Van and KITTI-Cyclist). Second, we bridge the gap between 2D and 3D tracking domains by leveraging a target-aware projection module, which enables to use a powerful 2D pre-trained tracker to facilitate the learning of 3D matching in existing 3D SOT trackers [27, 70] via matching distillation. We show that our proposed SiamDisst and MemDisst trackers can achieve state-of-the-art performance on KITTI and Waymo Open datasets, while running at real-time speed." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.425, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "2D Matching Distillation and 3D Pre-training" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.383, + 0.163 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.179, + 0.788, + 0.21 + ], + "angle": 0, + "content": "This work was supported by a Strategic Research Grant from City University of Hong Kong (Project No. 7005665)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.234, + 0.323, + 0.25 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.267, + 0.787, + 0.308 + ], + "angle": 0, + "content": "1. Anthes, C., García-Hernández, R.J., Wiedemann, M., Kranzlüller, D.: State of the art of virtual reality technology. In: 2016 IEEE aerospace conference. pp. 1-19. IEEE (2016) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.787, + 0.351 + ], + "angle": 0, + "content": "2. Ben-Baruch, E., Karklinsky, M., Biton, Y., Ben-Cohen, A., Lawen, H., Zamir, N.: It's all in the head: Representation knowledge distillation through classifier sharing. In: arXiv:2201.06945 (2022) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.352, + 0.787, + 0.393 + ], + "angle": 0, + "content": "3. Bertinetto, L., Valmadre, J., Henriques, J.F., Vedaldi, A., Torr, P.H.: Fully convolutional siamese networks for object tracking. In: Proceedings of the European Conference on Computer Vision. pp. 850-865. Springer (2016) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.394, + 0.787, + 0.422 + ], + "angle": 0, + "content": "4. Bhat, G., Danelljan, M., Gool, L.V., Timofte, R.: Leanring discriminative model prediction for tracking. In: ICCV. pp. 6182-6191 (2019) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.423, + 0.787, + 0.476 + ], + "angle": 0, + "content": "5. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11621-11631 (2020) 2, 4, 10, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.478, + 0.787, + 0.518 + ], + "angle": 0, + "content": "6. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015) 7, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.519, + 0.787, + 0.546 + ], + "angle": 0, + "content": "7. Chen, X., Yan, B., Zhu, J., Wang, D., Yang, X., Lu, H.: Transformer tracking. In: CVPR. pp. 8126-8135 (2021) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.548, + 0.787, + 0.574 + ], + "angle": 0, + "content": "8. Cui, Y., Fang, Z., Shan, J., Gu, Z., Zhou, S.: 3d object tracking with transformer. arXiv preprint arXiv:2110.14921 (2021) 4, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.576, + 0.787, + 0.602 + ], + "angle": 0, + "content": "9. Cui, Y., Jiang, C., Wang, L., Wu, G.: Mixformer: End-to-end tracking with iterative mixed attention. In: CVPR (2022) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.604, + 0.787, + 0.631 + ], + "angle": 0, + "content": "0. Danelljan, M., Bhat, G., Khan, F.S., Felsberg, M.: Eco: Efficient convolution operators for tracking. In: CVPR. pp. 21-26 (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.632, + 0.787, + 0.687 + ], + "angle": 0, + "content": "1. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arxiv 2020. arXiv preprint arXiv:2010.11929 (2010) 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.688, + 0.787, + 0.728 + ], + "angle": 0, + "content": "2. Duong, C.N., Luu, K., Quach, K.G., Shrinkteanet, N.L.: Shrinkteanet: Millionscale lightweight face recognition via shrinking teacher-student networks. In: arXiv:1905.10620 (2019) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.787, + 0.756 + ], + "angle": 0, + "content": "3. Fan, H., Lin, L., Yang, F.: Lasot: A high-quality benchmark for large-scale single object tracking. In: CVPR. pp. 5374-5383 (2019) 5, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.787, + 0.798 + ], + "angle": 0, + "content": "4. Fan, H., Ling, H.: Siamese cascaded region proposal networks for real-time visual tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7952-7961 (2019) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "5. Fang, Z., Zhou, S., Cui, Y., Scherer, S.: 3d-siamrpn: An end-to-end learning method for real-time 3d single object tracking using raw point cloud. IEEE Sensors Journal 21(4), 4995-5011 (2020) 12, 13" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.267, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Q. Wu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "16. Feng, S., Liang, P., Gao, J., Cheng, E.: Multi-correlation siamese transformer network with dense connection for 3d single object tracking. IEEE Robotics and Automation Letters (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.191, + 0.785, + 0.218 + ], + "angle": 0, + "content": "17. Galoogahi, H., Fagg, A., Lucey, S.: Learning background-aware correlation filters for visual tracking. In: ICCV (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.219, + 0.785, + 0.26 + ], + "angle": 0, + "content": "18. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. The International Journal of Robotics Research 32(11), 1231-1237 (2013) 2, 4, 10, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.261, + 0.785, + 0.303 + ], + "angle": 0, + "content": "19. Giancola, S., Zarzar, J., Ghanem, B.: Leveraging shape completion for 3d siamese tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1359-1368 (2019) 4, 10, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.304, + 0.785, + 0.332 + ], + "angle": 0, + "content": "20. Girshick, R.: Fast r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 1440-1448 (2015) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.333, + 0.785, + 0.374 + ], + "angle": 0, + "content": "21. Guo, Z., Mao, Y., Zhou, W., Wang, M., Li, H.: Cmt: Context-matching-guided transformer for 3d tracking in point clouds. In: European Conference on Computer Vision. pp. 95-111. Springer (2022) 4, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.785, + 0.416 + ], + "angle": 0, + "content": "22. He, K., Chen, X., Xie, S., Li, Y., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022) 2, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.418, + 0.785, + 0.459 + ], + "angle": 0, + "content": "23. Henriques, J.F., Caseiro, R., Martins, P., Batista, J.: High-speed tracking with kernelized correlation filters. In: IEEE Transactions on Pattern Analysis and Machine Intelligence. pp. 583-596 (2015) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.487 + ], + "angle": 0, + "content": "24. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. In: arXiv:1503.02531 (2015) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.53 + ], + "angle": 0, + "content": "25. Huang, L., Zhao, X., Huang, K.: Got-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence (2019) 5, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.531, + 0.785, + 0.572 + ], + "angle": 0, + "content": "26. Hui, L., Wang, L., Cheng, M., Xie, J., Yang, J.: 3d siamese voxel-to-bev tracker for sparse point clouds. Advances in Neural Information Processing Systems 34, 28714-28727 (2021) 3, 4, 5, 6, 10, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.785, + 0.614 + ], + "angle": 0, + "content": "27. Hui, L., Wang, L., Tang, L., Lan, K., Xie, J., Yang, J.: 3d siamese transformer network for single object tracking on point clouds. In: European Conference on Computer Vision. pp. 293-310. Springer (2022) 2, 3, 4, 6, 10, 12, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.616, + 0.785, + 0.656 + ], + "angle": 0, + "content": "28. Jin, X., Peng, B., Wu, Y., Liu, Y., Liu, J.: Knowledge distillation via route constrained optimization. In: IEEE/CVF International Conference on Computer Vision (2019) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.785, + 0.699 + ], + "angle": 0, + "content": "29. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.785, + 0.742 + ], + "angle": 0, + "content": "30. Kiran, B.R., Sobh, I., Talpaert, V., Mannion, P., Al Sallab, A.A., Yogamani, S., Pérez, P.: Deep reinforcement learning for autonomous driving: A survey. IEEE Transactions on Intelligent Transportation Systems (2021) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.798 + ], + "angle": 0, + "content": "31. Kristan, M., Matas, J., Leonardis, A., Vojir, T., Pflugfelder, R., Fernandez, G., Nebehay, G., Porikli, F., Cehovin, L.: A novel performance evaluation methodology for single-target trackers. IEEE Transactions on Pattern Analysis and Machine Intelligence (2016) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "32. Kristan, M., Matas, J., Danelljan, M.: The first visual object tracking segmentation votes2023 challenge results. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023) 5" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.425, + 0.115, + 0.734, + 0.13 + ], + "angle": 0, + "content": "2D Matching Distillation and 3D Pre-training" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "33. Lan, K., Jiang, H., Xie, J.: Temporal-aware siamese tracker: Integrate temporal context for 3d object tracking. In: Proceedings of the Asian Conference on Computer Vision. pp. 399-414 (2022) 2, 4, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "34. Lee, S.H., Kim, D.H., Song, B.C.: Self-supervised knowledge distillation using singular value decomposition. In: In Proceedings of the European Conference on Computer Vision (ECCV) (2018) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.272 + ], + "angle": 0, + "content": "35. Li, B., Wu, W., Wang, Q., Zhang, F., Xing, J., Yan, J.: Siamrpn++: Evolution of siamese visual tracking with very deep networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4282-4291 (2019) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.314 + ], + "angle": 0, + "content": "36. Li, B., Yan, J., Wu, W., Zhu, Z., Hu, X.: High performance visual tracking with siamese region proposal network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8971-8980 (2018) 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.315, + 0.788, + 0.355 + ], + "angle": 0, + "content": "37. Li, J., Guo, Z., Li, H., Han, S., Baek, J.w., Yang, M., Yang, R., Suh, S.: Rethinking feature-based knowledge distillation for face recognition. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.384 + ], + "angle": 0, + "content": "38. Lin, L., Fan, H., Zhang, Z., Xu, Y., Ling, H.: Swintrack: A simple and strong baseline for transformer tracking pp. 16743-16754 (2022) 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.412 + ], + "angle": 0, + "content": "39. Liu, Y., Liang, Y., Wu, Q., Zhang, L., Wang, H.: A new framework for multiple deep correlation filters based object tracking. In: ICASSP (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.412, + 0.788, + 0.439 + ], + "angle": 0, + "content": "40. Muller, M., Bibi, A., S, G.: Trackingnet: A large-scale dataset and benchmark for object tracking in the wild. In: ECCV. pp. 300-317 (2018) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.481 + ], + "angle": 0, + "content": "41. Osep, A., Mehner, W., Mathias, M., Leibe, B.: Combined image-and world-space tracking in traffic scenes. In: IEEE International Conference on Robotics and Automation. pp. 1988-1995. IEEE (2017) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.522 + ], + "angle": 0, + "content": "42. Pang, Y., Wang, W., Tay, F.E., Liu, W., Tian, Y., Yuan, L.: Masked autoencoders for point cloud self-supervised learning. In: European conference on computer vision. pp. 604-621 (2022) 3, 6, 7, 10, 11, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.55 + ], + "angle": 0, + "content": "43. Pang, Z., Li, Z., Wang, N.: Model-free vehicle tracking and state estimation in point cloud sequences. In: IROS (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.591 + ], + "angle": 0, + "content": "44. Peng, B., Jin, X., Liu, J., Li, D., Wu, Y., Liu, Y., Zhou, S., Zhang, Z.: Self-supervised knowledge distillation using singular value decomposition. In: IEEE/CVF International Conference on Computer Vision (2019) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.591, + 0.788, + 0.633 + ], + "angle": 0, + "content": "45. Peng, B., Jin, X., Liu, J., Li, D., Wu, Y., Liu, Y., Zhou, S., Zhang, Z.: Shrinkteanet: Million-scale lightweight face recognition via shrinking teacher-student networks. In: IEEE/CVF International Conference on Computer Vision (2019) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.675 + ], + "angle": 0, + "content": "46. Qi, C., Su, H., Mo, K., Guibas, L.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: IEEE conference on computer vision and pattern recognition (2017) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.715 + ], + "angle": 0, + "content": "47. Qi, H., Feng, C., Cao, Z., Zhao, F., Xiao, Y.: P2b: Point-to-box network for 3d object tracking in point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6329-6338 (2020) 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.716, + 0.788, + 0.758 + ], + "angle": 0, + "content": "48. Qi, H., Feng, C., Cao, Z., Zhao, F., Xiao, Y.: P2b: Point-to-box network for 3d object tracking in point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6329-6338 (2020) 4, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.758, + 0.788, + 0.799 + ], + "angle": 0, + "content": "49. Qi, Z., Dong, R., Fan, G., Ge, Z., Zhang, X., Ma, K., Yi, L.: Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining (2023) 3, 6, 7, 11, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.841 + ], + "angle": 0, + "content": "50. Ran, T., Gavves, E., Smeulders, A.W.: Siamese instance search for tracking. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1420-1429 (2016) 4" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "Q. Wu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.173 + ], + "angle": 0, + "content": "51. Romero, A., Ballas, N.: Fitnets: Hints for thin deep nets. In: arXiv:1412.6550 (2014) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.214 + ], + "angle": 0, + "content": "52. Shan, J., Zhou, S., Fang, Z., Cui, Y.: Ptt: Point-track-transformer module for 3d single object tracking in point clouds. arXiv preprint arXiv:2108.06455 (2021) 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.216, + 0.785, + 0.271 + ], + "angle": 0, + "content": "53. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. p. 2446-2454 (2020) 4, 10, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.272, + 0.785, + 0.297 + ], + "angle": 0, + "content": "54. Tao, F., Wang, M.: Response-based distillation for incremental object detection. In: arXiv:2110.13471 (2021) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.338 + ], + "angle": 0, + "content": "55. Tao, F., Wang, M., Yuan, H.: Overcoming catastrophic forgetting in incremental object detection via elastic response distillation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.785, + 0.365 + ], + "angle": 0, + "content": "56. Wang, N., Song, Y., Ma, C.: Unsupervised deep tracking. In: CVPR. pp. 3708-1317 (2019) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.785, + 0.406 + ], + "angle": 0, + "content": "57. Wang, Z., Xie, Q., Lai, Y.K., Wu, J., Long, K., Wang, J.: Mlvsnet: Multi-level voting siamese network for 3d visual tracking. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3101-3110 (2021) 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.785, + 0.446 + ], + "angle": 0, + "content": "58. Wang, Z., Yu, X., Rao, Y., Zhou, J., Lu, J.: P2p: Tuning pre-trained image models for point cloud analysis with point-to-pixel prompting. Advances in neural information processing systems 35, 14388-14402 (2022) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.473 + ], + "angle": 0, + "content": "59. Wu, Q., Chan, A.: Meta-graph adaptation for visual object tracking. In: ICME (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.785, + 0.5 + ], + "angle": 0, + "content": "60. Wu, Q., Yan, Y., Liang, Y., Liu, Y., Wang, H.: Dsnet: Deep and shallow feature learning for efficient visual tracking. In: ACCV. pp. 119-134 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.541 + ], + "angle": 0, + "content": "61. Wu, Q., Wan, J., Chan, A.B.: Progressive unsupervised learning for visual object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2993-3002 (2021) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.595 + ], + "angle": 0, + "content": "62. Wu, Q., Yang, T., Liu, Z., Wu, B., Shan, Y., Chan, A.B.: Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14561-14571 (2023) 1, 2, 4, 5, 6, 8, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.785, + 0.637 + ], + "angle": 0, + "content": "63. Wu, Q., Yang, T., Wu, W., Chan, A.B.: Scalable video object segmentation with simplified framework. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.677 + ], + "angle": 0, + "content": "64. Wu, Q., Sun, C., Wang, J.: Multi-level structure-enhanced network for 3d single object tracking in sparse point clouds. IEEE Robotics and Automation Letters 8(1), 9-16 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.704 + ], + "angle": 0, + "content": "65. Wu, Y., Lim, J., Yang, M.H.: Online object tracking: A benchmark. In: CVPR. pp. 2411-2418 (2013) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.785, + 0.757 + ], + "angle": 0, + "content": "66. Xia, Y., Gladkova, M., Wang, R., Li, Q., Stilla, U., Henriques, J.F., Cremers, D.: Casspr: Cross attention single scan place recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8461-8472 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.8 + ], + "angle": 0, + "content": "67. Xia, Y., Shi, L., Ding, Z., Henriques, J.F., Cremers, D.: Text2loc: 3d point cloud localization from natural language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14958-14967 (2024) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "68. Xia, Y., Wu, Q., Li, W., Chan, A.B., Stilla, U.: A lightweight and detector-free 3d single object tracker on point clouds. IEEE Transactions on Intelligent Transportation Systems (2023) 4, 5, 12, 13" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.425, + 0.115, + 0.734, + 0.13 + ], + "angle": 0, + "content": "2D Matching Distillation and 3D Pre-training" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.202 + ], + "angle": 0, + "content": "69. Xu, T.X., Guo, Y.C., Lai, Y.K., Zhang, S.H.: Cxtrack: Improving 3d point cloud tracking with contextual information. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1084-1093 (2023) 4, 6, 10, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.788, + 0.258 + ], + "angle": 0, + "content": "70. Xu, T.X., Guo, Y.C., Lai, Y.K., Zhang, S.H.: Mbptrack: Improving 3d point cloud tracking with memory networks and box priors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9911-9920 (October 2023) 2, 3, 4, 6, 10, 12, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.259, + 0.786, + 0.286 + ], + "angle": 0, + "content": "71. Yan, B., Peng, H., Fu, J., Wang, D., Lu, H.: Learning spatio-temporal transformer for visual tracking. In: ICCV. pp. 10448-10457 (2021) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.287, + 0.787, + 0.326 + ], + "angle": 0, + "content": "72. Yang, T., Chan, A.B.: Learning dynamic memory networks for object tracking. In: Proceedings of the European Conference on Computer Vision. pp. 152-167 (2018) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.328, + 0.787, + 0.369 + ], + "angle": 0, + "content": "73. Ye, B., Chang, H., Ma, B., Shan, S., Chen, X.: Joint feature learning and relation modeling for tracking: A one-stream framework. In: European Conference on Computer Vision. pp. 341-357 (2022) 2, 3, 4, 5, 6, 7, 8, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.37, + 0.787, + 0.41 + ], + "angle": 0, + "content": "74. Yim, J., Joo, D., Bae, J., Kim, J.: A gift from knowledge distillation: Fast optimization, network minimization and transfer learning. In: IEEE conference on computer vision and pattern recognition (2017) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.411, + 0.787, + 0.452 + ], + "angle": 0, + "content": "75. Zhang, L., Gonzalez-Garcia, A., Weijer, J.v.d., Danelljan, M., Khan, F.S.: Learning the model update for siamese trackers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4010-4019 (2019) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.453, + 0.787, + 0.493 + ], + "angle": 0, + "content": "76. Zhang, Z., Peng, H.: Deeper and wider siamese networks for real-time visual tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4591-4600 (2019) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.494, + 0.787, + 0.549 + ], + "angle": 0, + "content": "77. Zheng, C., Yan, X., Gao, J., Zhao, W., Zhang, W., Li, Z., Cui, S.: Box-aware feature enhancement for single object tracking on point clouds. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13199-13208 (2021) 2, 3, 4, 5, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.549, + 0.787, + 0.605 + ], + "angle": 0, + "content": "78. Zheng, C., Yan, X., Zhang, H., Wang, B., Cheng, S., Cui, S., Li, Z.: Beyond 3d siamese tracking: A motion-centric paradigm for 3d single object tracking in point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8111-8120 (2022) 4, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.606, + 0.787, + 0.659 + ], + "angle": 0, + "content": "79. Zhou, C., Luo, Z., Luo, Y., Liu, T., Pan, L., Cai, Z., Zhao, H., Lu, S.: Pttr: Relational 3d point cloud object tracking with transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8531-8540 (2022) 4, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.66, + 0.787, + 0.688 + ], + "angle": 0, + "content": "80. Zhou, H., Song, L., Chen, J., Zhou, Y., Guoli: Rethinking soft labels for knowledge distillation: A bias-variance tradeoff perspective. In: arXiv:2102.00650 (2021) 5" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.688 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Boosting 3D Single Object Tracking with 2D Matching Distillation and 3D Pre-training/515a618d-9747-46e7-9c9a-8a152456fd1b_origin.pdf b/2024/Boosting 3D Single Object Tracking with 2D Matching Distillation and 3D Pre-training/515a618d-9747-46e7-9c9a-8a152456fd1b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d46d4c3a92887cc43364de8b1ee2707f878fc359 --- /dev/null +++ b/2024/Boosting 3D Single Object Tracking with 2D Matching Distillation and 3D Pre-training/515a618d-9747-46e7-9c9a-8a152456fd1b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:035dd5a59e0d21c42dec6402cb237ac3345c288931536f248b7097e5775f2451 +size 4785662 diff --git a/2024/Boosting 3D Single Object Tracking with 2D Matching Distillation and 3D Pre-training/full.md b/2024/Boosting 3D Single Object Tracking with 2D Matching Distillation and 3D Pre-training/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ace26c8046448e98f3560ee1db2a323bb9c3f55d --- /dev/null +++ b/2024/Boosting 3D Single Object Tracking with 2D Matching Distillation and 3D Pre-training/full.md @@ -0,0 +1,297 @@ +# Boosting 3D Single Object Tracking with 2D Matching Distillation and 3D Pre-training + +Qiangqiang $\mathsf{W u}^{1}$ , Yan Xia\*2,3, Jia Wan4, and Antoni B. Chan1 + +$^{1}$ Department of Computer Science, City University of Hong Kong + +2 Technical University of Munich + +3 Munich Center for Machine Learning (MCML) + +$^{4}$ School of Computer Science and Technology, Harbin Institute of Technology, Shenzhen + +qiangqwu2-c@my.cityu.edu.hk, yan.xia@tum.de, jiawan1998@gmail.com, abchan@cityu.edu.hk + +Abstract. 3D single object tracking (SOT) is an essential task in autonomous driving and robotics. However, learning robust 3D SOT trackers remains challenging due to the limited category-specific point cloud data and the inherent sparsity and incompleteness of LiDAR scans. To tackle these issues, we propose a unified 3D SOT framework that leverages 3D generative pre-training and learns robust 3D matching abilities from 2D pre-trained foundation trackers. Our framework features a consistent target-matching architecture with the widely used 2D trackers, facilitating the transfer of 2D matching knowledge. Specifically, we first propose a lightweight Target-Aware Projection (TAP) module, allowing the pre-trained 2D tracker to work well on the projected point clouds without further fine-tuning. We then propose a novel IoU-guided matching-distillation framework that utilizes the powerful 2D pre-trained trackers to guide 3D matching learning in the 3D tracker, i.e., the 3D template-to-search matching should be consistent with its corresponding 2D template-to-search matching obtained from 2D pre-trained trackers. Our designs are applied to two mainstream 3D SOT frameworks: memory-less Siamese and contextual memory-based approaches, which are respectively named SiamDisst and MemDisst. Extensive experiments show that SiamDisst and MemDisst achieve state-of-the-art performance on KITTI, Waymo Open Dataset and nuScenes benchmarks, while running at above real-time speed of 25 and 90 FPS on a RTX3090 GPU. + +# 1 Introduction + +3D Single Object Tracking (SOT) has emerged as a basic 3D task for numerous practical applications in autonomous driving [30,67], robotics [41,66] and virtual reality [1]. Given a target described as a 3D bounding box (bbox) in the first frame, 3D SOT aims to track the target across all the frames in the 3D scene by predicting thebbox pose and position, which can be considered as a 3D extension of the 2D SOT task [62] but using a different input modality of 3D point clouds. + +![](images/6b91c85c43066eab893cbc213a499c5a74a5f698278d387ac8aa7e361e91653d.jpg) +Fig. 1: Comparison of our SiamDisst and MemDisst to state-of-the-art 3D SOT on the Van category of the KITTI dataset [18]. Compared with Siamese-based trackers (plotted as squares), our SiamDisst achieves the best performance while running at the real-time speed of 25 FPS. Our lightweight MemDisst outperforms the other contextual memory and motion-based approaches (plotted as circles), setting new SOTA performance on KITTI-Van, and runs efficiently at 90 FPS. This demonstrates that our designs effectively boost 3D matching with limited category-specific training data (e.g., the KITTI-Van training set only contains 1994 examples). + +![](images/7588f4b20e70ff6f1c065388725e3bf7c0be719bf7a912100a5f9df01681eafe.jpg) + +As illustrated in Fig. 2, existing 3D SOT methods [38,62,73] mainly focus on learning robust 3D tracking models from sparse point cloud datasets [5,18]. For example, 3D siamese trackers [27,33,47,77] follow a siamese matching paradigm for accurate localization. To leverage more historical cues, MBPTrack [70] further extends this framework to a contextual memory-matching mechanism. Although the 3D matching models proposed by these methods play a key role in obtaining favorable performance [5,18], the sparsity and incompleteness of the point cloud training limits their abilities. Moreover, a paucity of large-scale category-specific 3D tracking datasets leads to insufficient training data for effectively learning the 3D matching modules in existing 3D trackers. + +In contrast to the 3D SOT task, various large-scale annotated video datasets [29] are available for training 2D SOT methods. Existing 2D deep trackers can fully benefit from the feature backbones learned from 2D image [22] or video pre-training methods [62], enabling the learning of robust matching modules for 2D SOT on videos. The state-of-the-art trackers [38,62,73] with strong matching abilities show excellent tracking results even without using online memory mechanisms. With the remarkable success of 2D SOT, our work is motivated by the following questions: 1) Inspired by 2D SOT, can we design a general and effective 3D SOT framework that benefits from 3D pre-training? 2) Can we transfer the well-learned matching ability of state-of-the-art 2D trackers to 3D matching in 3D SOT trackers, so as to facilitate training from limited category-specific point cloud data? + +The main challenge of achieving the transfer of the 2D matching ability lies in two aspects. First, there is a significant misalignment between the input modalities for 2D and 3D trackers, dense 2d images and sparse 3d point clouds, respectively. Note that in the 3D SOT setting, only the single modality of point + +![](images/2b1c7e2963833f85e0e21dfcca44039fc3b73ae116c6d54e3f0ac5b7d3f188ad.jpg) +(a) Previous Siamese 3D SOT + +![](images/5f064ab3210a8271c795d2d5bbd54b93213e202ad044b6e7d9f055ccdc9c929f.jpg) +(b) Our Proposed 3D SOT (i.e., SiamDisst and MemDisst) +Fig. 2: The comparison of 3D SOT frameworks (a) the previous Siamese 3D SOT trackers [27,47,77] focus on learning the 3D matching module from limited category-specific point cloud (PC) training data. (b) The proposed 3D SOT framework bridges the gap between 2D and 3D SOT by using our designed target-aware projection (TAP) module, which enables the 2D tracker [73] trained on large-scale 2D videos to operate well on the point clouds. The well-learned 2D matching pattern is further distilled to facilitate the 3D matching learning. Note that only the 3D-related modules (denoted as blue components) are optimized during the distillation training stage. The 2D tracker and TAP module are not used during online inference to increase speed. + +cloud data is used as input, which makes it difficult to directly apply the current 2D SOT trackers. Second, how to effectively transfer the 2D matching knowledge in pre-trained 2D trackers to guide the learning of the 3D matching in 3D trackers is unclear given the differences in internal representations between 2D and 3D SOT, i.e., 2d dense feature maps and unordered 3D point embeddings. + +To tackle the aforementioned challenges, we propose a general 3D SOT framework that can effectively utilize knowledge learned in 2D SOT, which is illustrated in Fig. 2(b). We first propose to align the different input modalities of 2D and 3D trackers via a learnable target-aware projection (TAP) module, which takes the raw point cloud as the input and projects them into the 2D image space. These projected images are suitable for 2D target matching and 2D bounding box regression by 2D pre-trained trackers, i.e., no fine-tuning of 2D trackers is required. Secondly, we propose a matching distillation approach that uses the knowledge in the high-quality 2D matching patterns from the 2D tracker as guidance for learning the matching patterns for 3D SOT. The proposed framework employs a standard Vision Transformer (ViT) as the matching module, which can naturally benefit from 3D pre-training approaches [42,49] and also creates a design that is consistent with the 2D SOT trackers (e.g., OSTrack [73]) for more natural distillation of matching knowledge. + +We implement our framework with both Siamese [26, 27] and contextual memory-based [70] matching paradigms, which are respectively named SiamDisst and MemDisst. Extensive experiments demonstrate that our SiamDisst achieves significant improvements over its baseline, and MemDisst sets new state-of-the + +art performance on KITTI, while respectively running at a fast speed of 90 FPS on a single RTX3090 GPU. + +In summary, the main contributions of our work are: + +- We build a general SOT framework that employs a standard ViT as the target matching module, which is consistent with dominant 2D trackers [62, 73] for better knowledge transfer of matching patterns, and meanwhile can naturally benefit from 3D pre-training. + +- We propose a lightweight target-aware projection (TAP) module to bridge the gap between the 2D and 3D tracking domains, which enables 2D pretrained trackers to be operated on point cloud data without additional 2D tracker fine-tuning. + +- We propose an IoU-guided matching distillation approach to guide 3D SOT learning, which facilitates transfer of knowledge from powerful 2D pre-trained foundation trackers to the 3D tracker. We show that the matching distillation facilitates the 3D model learning in the low-data regime. + +# 2 Related Work + +3D Single Object Tracking. Inspired by the great success in the 2D tracking community, much progress has been made in 3D SOT [43]. The pioneering work SC3D [19] proposes the first 3D Siamese network to compute the similarity between the template and 3D candidate proposals. However, the time-consuming matching pipeline in SC3D is not end-to-end trainable, thus degrading its performance on 3D tracking benchmarks [5,18,53]. Inspired by the 2D region proposal network (RPN) [20,36], P2B [48] proposes a 3D RPN to generate high-quality 3D proposals, which achieves better tracking results while running at the real-time speed. BAT [77] extends P2B with a box-aware feature design, which provides a strong prior of the target shape. In contrast to previous approaches that directly regress 3D bboxes from a sparse point cloud, V2B [26] proposes a voxel-based localization head to localize the target in the Bird's Eye View (BEV). + +With the development of transformers, 3D transformer SOT trackers (e.g., LTTR [8], PTTR [79] and CMT [21]) are designed for target-aware feature learning. Further extensions include iterative transformer blocks [27], historical cue modeling [33], contextual memory [69, 70], motion prediction [68, 78] and advanced architecture design [16, 64]. Despite these successes, learning robust 3D tracking models is still limited by the sparse and insufficient category-specific point cloud data. In this work, our aim is to leverage off-the-shelf 2D trackers trained on large-scale video data for enhancing 3D trackers, thus alleviating the current data bottleneck in 3D SOT. + +2D Single Object Tracking. Traditional 2D SOT [10,17,23,39,59,60] follow a correlation filter tracking framework due to its fast running speed and favorable tracking performance. With the development of the deep learning, much progress has been made in designing convolutional neural networks for video object tracking. Specifically, SiamFC [3] and SINT [50] employ a siamese neural network and treat tracking as a template matching problem. Based on SiamFC, many + +improvements have been made, including architecture design [14,35,36,76], unsupervised representation learning [56,61,62] and template updating [4,72,75]. Recent advances using transformers [7,9,38,63,71,73] achieve state-of-the-art performance on popular tracking benchmarks [13,25,32,65]. The great success in 2D tracking significantly inspires the development in the 3D tracking community. However, there is no previous work trying to bridge the gap between the 2D and 3D tracking areas. In this paper, we propose a lightweight target-aware projection (TAP) module to bridge the gap between 2D and 3D SOT, and use the 2D tracker to effectively guide the 3D learning. + +Knowledge Distillation. Knowledge distillation (KD) aims to transfer the knowledge from a teacher model to a student model. Previous response-based KD approaches [24,54,55,80] focus on using the final output of the teacher model to supervise the training of the student model. To supervise the intermediate feature learning, various feature-based KD approaches [2,12,28,37,45,51] have been proposed. FitNet [51] first uses the intermediate representations learned by the teacher as hints to distill a student model. However, the intermediate supervision is unsuitable in our cross-modality distillation case, since the 2D and 3D representations have a large gap. Unlike the above approaches, relation-based KD [34,44] aims to leverage the relationships between various samples or feature maps, e.g., FPS [74] uses the inner products of feature maps for distillation. However, naively applying feature correlation as supervision is not helpful in 2D-to-3D tracking, which is due to the transfer noise and causes performance degradation (see Table 2). In this paper, we propose a novel IoU-guided matching distillation approach for better 2D knowledge transfer, which is also the first attempt to perform cross-modality transfer for 3D SOT. + +# 3 Methodology + +In this section, we first present our general 3D SOT framework with a standard ViT as the target relation modeling module, which could naturally benefit from 3D self-supervised pre-training and be consistent with 2D pre-trained trackers for better distillation. To use the 2D pre-trained model (frozen) for guiding 3D matching, we introduce a lightweight target-aware projection (TAP) module to bridge the domain gap between 2D and 3D tracking. Finally, we explore our IoU-guided matching distillation to boost the 3D tracking performance. + +# 3.1 Unified 3D SOT Framework + +Revisit of 3D SOT. Given the 3D bbox in the first frame, the goal of 3D SOT is to accurately predict 3D bboxes in the following frames. Existing Siamese matching approaches [26,47,68,77] mainly use the point cloud (PC) data $\hat{\mathbf{p}}_1\in$ $\mathbb{R}^{\hat{N}\times 3}$ cropped in the first frame as the template to predict the 3D bbox $\mathcal{B}_t\in$ $\mathbb{R}^7$ from the PC $\mathbf{p}_t\in \mathbb{R}^{N_s\times 3}$ in the $t$ -th search frame. $\hat{N}$ is the number of sampled target points within the annotated 3D bounding box, $N_{s}$ is the number of points in each search frame, and $\mathcal{B}_t$ is parameterized by its center coordinates + +![](images/c2fc1c2d824f0e586e6b0dd457511e02af10688ec65dbe30f1eb28460669cb71.jpg) +Fig. 3: The pipeline converting 3D point clouds (PC) to 2D images using our target-aware projection (TAP) module, followed by the 2D pre-trained tracker. + +(x, y, and z), orientation $\theta$ and box size (width, height and length). Recent approaches [69, 70] fully leverage contextual information and use more memory frames $\{\mathbf{p}_i\}_{i = t - k}^t$ to predict $\mathcal{B}_t$ , where $k$ is the number of memory frames. Instead of explicitly removing template points, a target mask set $\{\mathcal{U}_i\}_{i = t - k - 1}^{t - 1}$ is used to identify the points within the 3D bbox, where $\mathcal{U}_i \in \mathbb{R}^{N_s \times 1}$ . + +Target matching module. For simplicity and consistency, we use $\mathbf{m} \in \mathbb{R}^{N_m \times 3}$ to represent the unified memory used in the above approaches, where $N_m = \hat{N}$ in memory-less Siamese approaches and $N_m = N_s \times k$ in memory approaches. We then obtain the search point features $\mathbf{f}_t = \phi(\mathbf{p}_t) \in \mathbb{R}^{N_s \times C}$ and memory point features $\mathbf{f}_m = \phi(\mathbf{m}) \in \mathbb{R}^{N_m \times C}$ by applying a point feature extractor $\phi(\cdot)$ on $\mathbf{p}_t$ and $\mathbf{m}$ respectively. The current 3D SOT trackers mainly focus on designing a robust 3D matching module $\mathcal{M}(\mathbf{f}_s, \mathbf{f}_m)$ to perform the target matching between the search and memory point features. However, learning this matching module is challenging due to the sparsity of LIDAR point clouds and the lack of sufficient category-specific training data. + +To tackle these issues, inspired by 2D SOT [62, 73], we propose to use a standard ViT [11] for 3D matching since: 1) the standard ViT can benefit from 3D self-supervised pre-training (e.g., [42, 49]), which eases the need for large-scale 3D annotated data for 3D SOT training; 2) it has the consistent target relation modelling architecture (i.e., ViT) with dominant 2D trackers, which could more naturally transfer its 2D matching knowledge. The 3D matching process with ViT is formulated as: + +$$ +\hat {\mathbf {f}} _ {t} = \mathcal {M} \left(\mathbf {f} _ {t}, \mathbf {f} _ {m} + \theta (\mathcal {U})\right), \tag {1} +$$ + +where the output $\hat{\mathbf{f}}_t\in \mathbb{R}^{N_s\times C}$ is the updated search features output from the last layer of $\mathcal{M}$ , $\theta (\cdot)$ is a learnable linear mapping that converts the 1-dim masks into the $C$ -dim feature maps. For siamese tracking paradigms, we drop this operation since the template is well cropped. Finally, the matching output $\hat{\mathbf{f}}_t$ is fed to the prediction head [26, 27, 70] for predicting the 3D bbox $\mathcal{B}_t$ . Briefly, we call the memory-less Siamese-based and memory-based approaches with our standard ViT matching module as SiamBase and MemBase (see Supp. for more details). + +# 3.2 Target-Aware Projection Module + +Although SiamBase and MemBase can naturally benefit from 3D self-supervised pre-training [42, 49], their performance is still limited due to the absence of large-scale 3D datasets (e.g., ShapeNet [6] only contains 55 object categories with 50k point clouds) for obtaining effective pre-training weights. To further improve the 3D tracking performance, we propose to employ a powerful 2D pre-trained tracker (denoted as $\mathcal{T}_{2D}$ ) as the distillation teacher to guide the 3D matching learning in SiamBase and MemBase. However, in 3D SOT, only the single modality of point cloud data is available, making it hard to directly utilize current 2D SOT trackers. To bridge the modality gap between 2D and 3D tracking, we thus propose a learnable target-aware projection (TAP) module in this subsection, illustrated in Fig. 3. + +3D-to-2D Projection. Using the domain knowledge that the objects-of-interest for 3D tracking are typically on the ground plane (at similar fixed z-level heights), we project the search point cloud $\mathbf{p}_t$ from the orthogonal view along the $z$ axis. For each point, we round down the $(x,y)$ coordinates to obtain its location on the 2D $xy$ plane, i.e., from the bird-eye view (BEV). We denote this 3D-to-2D projection as $\psi (\mathbf{p}_t)$ . + +Target-Aware Projection. Based on [58], the point features $\mathbf{f}_t$ can be mapped to the 2D plane based on the corresponding projection $\psi(\mathbf{p}_t)$ . However, simply projecting search features to the 2D plane is not optimal since it does not consider any target information, which may not well support the subsequent 2D matching (see Table 2). Therefore, we use a cross-attention module $\mathcal{C}$ here to obtain more effective target-aware search features: + +$$ +\mathbf {f} _ {t} ^ {*} = \mathcal {C} \left(Q (\mathbf {f} _ {t}), K \left(\mathbf {f} _ {m} + \theta (\mathcal {U}) + E (\mathbf {m})\right), V \left(\mathbf {f} _ {m} + \theta (\mathcal {U}) + E (\mathbf {m})\right)\right), \tag {2} +$$ + +where $Q$ , $K$ and $V$ are query, key and value embedding modules, respectively. $E$ is the 3D positional embedding module. We project the target-aware search features $\mathbf{f}_t^* \in \mathbb{R}^{N_s \times C}$ into the 2D plane via $\psi(\mathbf{p}_t)$ and obtain the projected feature maps with $C$ channels, which are further input to a lightweight CNN $g$ [58] to reduce its channel dimension, yielding the projected search image: + +$$ +I _ {t} = g \left(\psi \left(\mathbf {p} _ {t}\right)\right) \in \mathbb {R} ^ {H _ {t} \times W _ {t} \times 3}, \tag {3} +$$ + +where $H_{t}$ and $W_{t}$ are height and width of the projected search image. The same operations are made on each memory frame and its corresponding features, thus we obtain $k$ projected memory images $\{I_m^i\}_{i = t - k - 1}^{t - 1}$ , where $I_{m}^{i}\in \mathbb{R}^{H_{m}\times W_{m}\times 3}$ . + +TAP training. The TAP module is trained via a simulated 2D tracking process on the projected search and target images, so that the projected images can obtain good 2D tracking performance. As illustrated in Fig. 3, the projected 2D search and memory images are input to a ViT-based pre-trained 2D tracker (e.g., OSTrack [73]) for 2D bounding box regression. Specifically, the input images are flattened into a sequence of 2D patches, projected into image embeddings, and then concatenated along the spatial dimension to input to the 2D ViT model for joint feature extraction and matching. The updated search features output + +![](images/88a89314ba2d9f4f2c421e04db7229384f3cc4c7cbf930f2882d196a77f85c9a.jpg) +Fig. 4: Visualization of the projected search image $I_{t}$ obtained by our TAP module on (a) KITTI-Pedestrian and (b) KITTI-Cycler. (c): The attention maps are generated via template-to-search matching from the 2D tracker on (b) KITTI-Cycler. The red box indicates the ground-truth (GT) 2D box projected from the GT 3D box. The blue box denotes the prediction of the 2D pre-trained OSTrack [73]. Based on the target-aware search features, the TAP module can effectively encode the target region with unique color and texture, enabling subsequent 2D matching. + +from the last layer is input to the prediction head for 2D box regression, which can be generally formulated as: + +$$ +\mathcal {B} _ {t} ^ {2 D}, \hat {\mathbf {f}} _ {t} ^ {2 D}, \hat {\mathbf {f}} _ {m} ^ {2 D} = \mathcal {T} _ {2 D} (\{I _ {m} ^ {i} \} _ {i = t - k - 1} ^ {t - 1}, I _ {t}), \tag {4} +$$ + +where $\mathcal{B}_t^{2D} \in \mathbb{R}^4$ is the predicted 2D bounding box in the projected search frame, and $(\hat{\mathbf{f}}_t^{2D}, \hat{\mathbf{f}}_m^{2D})$ are the updated 2D search and memory feature maps from the last layer of the 2D tracker. Here we use OSTrack as the 2D pre-trained tracker, and use its same 2D loss $\mathcal{L}_{2D}$ [62, 73] to evaluate the simulated 2D tracking result, thus providing supervision on the TAP module. The ground-truth 2D bounding box $\mathcal{G}_t^{2D}$ is obtained by projecting the corners of the ground-truth 3D box to the projected search image via the 3D-to-2D projection $\psi$ . + +After training the TAP module, the projection process is aligned with the 2D tracker. Note that we only update the parameters of the TAP module while keeping the parameters in the 2D pre-trained tracker $\mathcal{T}_{2D}$ fixed, which makes the TAP training more efficient, and less likely to overfit. We also visualize the projected 2D search images in Fig. 4. The learned TAP module can support accurate 2D matching in the projected search image space well via the pretrained OSTrack model, which shows the effectiveness of the proposed TAP module. + +# 3.3 IoU-guided Matching Distillation + +We introduce how to distill the 2D matching knowledge in $\mathcal{T}_{2D}$ to guide the 3D matching learning in our 3D SOT trackers, SiamBase and MemBase. In (1), we obtain the updated 3D search features $\hat{\mathbf{f}}_t\in \mathbb{R}^{N_s\times C}$ and updated 3D memory + +![](images/2c13193de7125ae7c3a33e1a672f1fd4544ca0e92a9003c528ad14b4484b5258.jpg) +Fig. 5: Training pipeline of our SiamDisst or MemDisst via IoU-guided matching distillation. During the training stage, only the SiamDisst or MemDisst model is updated. + +features $\hat{\mathbf{f}}_m\in \mathbb{R}^{N_m\times C}$ output from the last layer of $\mathcal{M}$ . Similarly, the updated 2D search and memory feature maps, $\hat{\mathbf{f}}_t^{2D}\in \mathbb{R}^{\frac{H_t}{s}\times \frac{W_t}{s}\times C}$ and $\hat{\mathbf{f}}_m^{2D}\in \mathbb{R}^{\frac{H_m}{s}\times \frac{W_m}{s}\times k\times C}$ , are obtained via (4) in the last layer of 2D ViT, where $s$ denotes the patch size for patch embedding. The 3D and 2D features belong to different embedding spaces, and thus directly guiding the feature embeddings of the 3D tracker using the 2D tracker is not straightforward (see Supplementary). Instead, we use the matching correspondences (between feature maps) used by the 2D tracker to guide the matching correspondences (between feature vectors) of the 3D tracker. 3D-to-2D feature indexing. Given a search point coordinate in the search point cloud $\mathbf{p}_t$ , we index its corresponding 2D feature vector via the 3D-to-2D projection $\psi$ . Specifically, we first obtain its 2D location in the projected search image. Then, we interpolate its 2D features $\hat{\mathbf{f}}_t^{2D}$ to its original spatial size $H_{t}\times W_{t}$ via bilinear interpolation, and finally index its feature vector via the projected 2D location. Based on the above 3D-to-2D feature indexing, we obtain the 2D search and memory features, denoted as $\tilde{\mathbf{f}}_t\in \mathbb{R}^{N_s\times C}$ and $\tilde{\mathbf{f}}_m\in \mathbb{R}^{N_m\times C}$ , that correspond to the 3D search and memory point clouds. + +Matching distillation. Note that $\tilde{\mathbf{f}}_t$ and $\tilde{\mathbf{f}}_m$ are obtained from the 2D ViT in the 2D tracker, which is well learned on large-scale 2D annotated videos, thus providing more robust matching results than our 3D counterparts (SimBase and MemBase) trained on limited category-specific point cloud data. Thus, we use the 2D memory-to-search matching pattern to guided the 3D matching process: + +$$ +\mathcal {L} _ {d i s t} = \mathrm {I o U} (\mathcal {B} _ {t} ^ {2 D}, \mathcal {G} _ {t} ^ {2 D}) \cdot \mathcal {L} _ {1} (\hat {\mathbf {f}} _ {t} \hat {\mathbf {f}} _ {m} ^ {T}, \tilde {\mathbf {f}} _ {t} \tilde {\mathbf {f}} _ {m} ^ {T}), \tag {5} +$$ + +where $\hat{\mathbf{f}}_t\hat{\mathbf{f}}_m^T$ is the inner-product (matching) matrix between the 3D search and memory features, and likewise for $\tilde{\mathbf{f}}_t\tilde{\mathbf{f}}_m^T$ , and $\mathcal{L}_1$ is the L1 loss that calculates the mean absolute error between two matrices. IoU $(\cdot ,\cdot)$ calculates the intersection-over-union between the predicted 2D box $\mathcal{B}_t^{2D}$ and projected 2D ground-truth box $\mathcal{G}_t^{2D}$ , which measures the reliability of our 2D tracker, i.e. if the 2D matching is reliable enough and predicts the 2D box accurately (i.e., with high IoU), + +this 2D matching pattern will be used more for guiding the 3D matching. We also use the typical 3D box regression losses $\mathcal{L}_{3D}$ in the siamese paradigm (i.e., STNet [27]) and memory-based paradigm (i.e., MBPTrack [70]) to supervise our SiamBase and MemBase, respectively. The overall training loss is: + +$$ +\mathcal {L} _ {\text {o v e r a l l}} = \mathcal {L} _ {3 D} + \lambda \mathcal {L} _ {\text {d i s t}}, \tag {6} +$$ + +where $\lambda$ is a hyperparameter. After matching distillation, we denote the updated 3D SOT trackers as SiamDisst and MemDisst. Note that the TAP module, the 2D pre-trained tracker and the matching distillation are only used during the offline training stage to improve the 3D matching ability. During online inference, these modules are removed, and only our 3D trackers SiamDisst and MemDisst are required, which run at the real-time speed of 25 and 90 FPS, respectively. + +# 4 Experiments + +Implementation details. For training the TAP module, as illustrated in Fig. 3, we sample memory and search PCs as the input, and project them into the 2D images for further matching in the 2D tracker. We use OSTrack [73] trained on large-scale 2D video datasets [13, 25, 40], as the 2D tracker, and freeze its weights, thus leading to efficient TAP training. The original 2D loss in OSTrack is used to evaluate the 2D predictions without modification. For TAP training, we use a batch size of 8 with a learning rate of 1e-4 and a total epoch of 40. For SiamDisst, we use the same hierarchical backbone network and 3D prediction head in STNet [27] for initial PC feature extraction and 3D prediction. For MemDisst, we use the same lightweight PointNet [46] with Point-MAE [42] as the backbone, and use the prediction head and memory mechanism in MBPTrack [70] for a fair comparison. We employ ViT-384-6 and ViT-96-3 as matching modules for SiamDisst and MemDisst respectively, which is illustrated in Table 3. The hyper-parameter $\lambda$ is set to 1.0. The training and testing settings of SiamDisst and MemDisst are following their baselines STNet and MBPTrack. More details on implementation can be found in the Supplementary. + +Datasets and Evaluation. We use three 3D datasets, KITTI [18], nuScenes [5] and Waymo Open Dataset (WOD) [53], for training and testing our proposed approaches. The KITTI dataset contains 21 training video sequences and 29 test sequences. Following previous work [19], the training set is split into three parts: 0-16 for training, 17-18 for validation and 19-20 for testing. Following [26,27,69], we use nuScenes and WOD to test the generalization ability of our approaches. We follow the one one pass evaluation [31], which is widely used in previous 3D SOT works [19,26,27,70], to report Precision and Success metrics for evaluation. + +# 4.1 Ablation Study + +Component analysis. We evaluate the effect of the two main components used in our proposed framework: 3D pre-training and 2D matching distillation. + +
VariantBaseline3D Pre-training2D DistillationSuccessPrecision
SiamBase✓55.664.3
SiamBase✓60.369.3
SiamBase✓✓63.575.0
MemBase✓61.372.7
MemBase✓63.576.5
MemBase✓✓66.679.3
+ +Table 1: Component analysis of SiamDisst and MemDisst. 'Baseline' indicates the variant initialized with the random weights and trained from scratch. + +
VariantSucc./Prec.Pre-trainingSucc./Prec.
SiamDisst63.5/75.0Point-MAE [42]63.5/75.0
unlearnable TAP61.1/72.3ReCon [49]62.9/73.6
TAP w/o target-aware interact.61.2/73.12D MAE [22]60.5/69.9
2D distill. w/o IoU weight.56.4/66.3DropMAE [62]61.9/74.0
+ +Table 2: Ablation studies of SiamDisst on KITTI-VAN: (left) variants of the TAP module and distillation; (right) using different 3D pre-trained models. + +The 3D pre-training means the usage of the default pre-trained Point-MAE [42] weights for 3D initialization. As shown in Table 1, SiamDisst and MemDisst with 3D pre-training achieve $5\%$ and $3.8\%$ improvements over the naive baselines regarding the precision metric. By applying the 2D matching distillation, SiamDisst and MemDisst can be significantly improved, leading to $5.7\%$ and $2.8\%$ precision gains, which shows the effectiveness of our approach. + +TAP module. We study several variants of the proposed TAP module: including 1) unlearnable TAP: we remove the learnable projection module in the TAP, and use the z-axis value as the projected feature in the 2D image plane (repeated for each of the 3 image channels); 2) TAP module w/o target-aware design: we remove the cross attention module $\mathcal{C}$ and separately project the template and search images; Table 2 (left) presents the results. The unlearnable TAP obtains inferior performance, which is due to the low-quality projected images and more 2D tracking failures. Without the IoU weighting, the unlearnable TAP also has severely degraded performance (see Supplementary for more details). For the target-aware design, we can observe $2.1\% / 1.9\%$ gains by using it, which shows that learning target-aware features is beneficial in the 2D projection stage. + +IoU-guided matching distillation. In the IoU-guided matching distillation, we use the IoU to measure the reliability of the 2D prediction. To test its effectiveness, we remove this IoU weighting in (5) and show the comparison in Table 2 (left). The variant w/o the IoU weighting degrades the performance with large margins of $7.1\% / 8.7\%$ , which is mainly because the 2D matching distillation may introduce some low-quality and inaccurate 2D matching patterns for the 3D matching learning, thus severely degrading the learning. + +Comparison of 3D pre-trained weights. We use various pre-trained weights to initialize the ViT in our 3D tracker and test them in Table 2 (right). The + +
methodmatching modulesucc./prec.FPS
SiamDisstViT-384-663.5/75.025
SiamDisstViT-96-360.5/71.730
MemDisstViT-384-666.7/79.178
MemDisstViT-96-366.6/79.390
+ +Table 3: Comparisons of our approaches with various matching modules. + +generative 3D Point-MAE pre-training provides the best result. Note that ReCon [49] adds the contrastive learning in the multi-modality generative pretraining, which may not be suitable in the pure point cloud tracking setting. We also test two typical 2D pre-trained models, i.e., MAE [22] and DropMAE [62] (see Supp. for details). Interestingly, our SiamDisst with DropMAE initialization achieves competitive results, which indicates that the temporal matching ability of DropMAE learned in 2D videos is also helpful for the 3D tracking task. + +Lightweight or heavy matching module. We explore various matching modules used in SiamDisst and MemDisst in Table 3. Note that ViT-384-6 is built from the first 6 layers of ViT-Base [11] with 384 hidden dimensions and 6 heads. We also design a lightweight ViT-96-3 model with 4 heads for comparison. We initialize ViT-384-6 with the official Point-MAE weights and rerun Point-MAE pre-training to initialize ViT-96-35. In Table 3, SiamDisst benefits more from the heavy ViT-384-6 while MemDisst achieves favorable performance even using the lightweight ViT-96-3. This is because SiamDisst has no effective online memory mechanism and heavily relies on its matching module. For MemDisst, the online memory provides more historical cues for matching, which eases the need for a heavy online matching module. Considering the speed-accuracy trade-off, we use ViT-384-6 and ViT-96-3 for SiamDisst and MemDisst, respectively. + +# 4.2 State-of-the-art Comparison + +We compare the proposed SiamDisst and MemDisst with state-of-the-art 3D trackers on the KITTI, nuScenes and WOD datasets. Specifically, for KITTI, comprehensive comparisons are made with two groups of trackers: 1) memoryless Siamese-based 3D trackers, SC3D [19], 3DSiamRPN [15], P2B [48], LTTR [8], MLVSNet [57], BAT [77], PTT [52], V2B [26], CMT [21], PTTR [79], and STNet [27]; and 2) contextual memory or motion prediction based 3D trackers, TAT [33], DMT [68], M2Track [78], CXTrack [69], and MBPTrack [70]. + +As shown in Table 4, our memory-less tracker SiamDisst outperforms the other Siamese 3D trackers by large margins in terms of both success and precision metrics. Specifically, the overall performance (Mean column) achieved by SiamDisst is $66.2\% /83.9\%$ , which has significant improvements $(4.9\% /2.8\%)$ over the baseline tracker STNet and demonstrates the effectiveness of our proposed + +
MethodCM MPFPSCar (6424)Pedestrian (6088)Van (1248)Cyclist (308)Mean (14068)
SC3D [19]241.3/57.918.2/37.840.4/47.041.5/70.431.2/48.5
3DSiamRPN [15]2158.2/76.235.2/56.245.7/52.936.2/49.046.7/64.9
P2B [48]4656.2/72.828.7/49.640.8/48.432.1/44.742.4/60.0
LTTR [8]2365.0/77.133.2/56.835.8/45.666.2/89.948.7/65.8
MLVSNet [57]7056.0/74.034.1/61.152.0/61.434.3/44.545.7/66.7
BAT [77]6560.5/77.742.1/70.152.4/67.033.7/45.451.2/72.8
PTT [52]4067.8/81.844.9/72.043.6/52.537.2/47.355.1/74.2
V2B [26]3770.5/81.348.3/73.550.1/58.040.8/49.758.4/75.2
CMT [21]3270.5/81.949.1/75.554.1/64.155.1/82.459.4/77.6
PTTR [79]5365.2/77.450.9/81.652.5/61.865.1/90.557.9/78.1
STNet [27]3572.1/84.049.9/77.258.0/70.673.5/93.761.3/80.1
SiamDisst2573.7/85.158.4/83.963.5/75.076.9/94.566.2/83.9
TAT [33]✓-72.2/83.357.4/84.458.9/69.274.2/93.964.7/82.8
DMT [68]✓7266.4/79.448.1/77.953.3/65.670.4/92.655.1/75.8
M2Track [78]✓5765.5/80.861.5/88.253.8/70.773.2/93.562.9/83.4
CXTrack [69]✓3469.1/81.667.0/91.560.0/71.874.2/94.367.5/85.3
MBPTrack [70]✓5673.4/84.868.6/93.961.3/72.776.7/94.370.3/87.9
MemDisst✓9074.1/85.669.1/94.166.6/79.377.2/94.771.3/88.9
+ +Table 4: 3D tracking results on KITTI [18] for (top) memory-less 3D trackers; (bottom) memory/motion-based 3D trackers. CM and MP represent the contextual memory and motion prediction. Trackers are evaluated by per-category Success/Precision metrics. 'Mean' indicates the overall result averaged over frames. The best result is shown in bold. For each category, the number in brackets indicates the test frame number. + +IoU-guided matching distillation. Moreover, the proposed SiamDisst is more effective in identifying distractors during online tracking, e.g., obtaining best overall performance on the Pedestrian category. For MemDisst, it sets new state-of-the-art performance on all the categories while running at a fast speed of 90 FPS. The lightweight model used in MemDisst is well learned via the proposed learning framework, thus achieving promising tracking performance. Note that both SiamDisst and MemDisst performs well in the low data regime (e.g., the Van category only has 1994 training examples), which is mainly due to the effectiveness of transferring the matching knowledge from the 2D pre-trained tracker. + +To test the generalization performance of our proposed trackers, we use our SiamDisst and MemDisst trained on KITTI to evaluate on the WOD and nuScenes datasets. As seen in Table 5, SiamDisst performs best on both WOD and nuScenes among all the Siamese-based 3D trackers, e.g., STNet, V2B and BAT. This is mainly because our SiamDisst learns a more effective target matching module via the 2D matching distillation, thus leading to more accurate tracking results. Compared with MBPTrack, our MemDisst obtains comparable performance on WOD and outperforms it on nuScenes. Meanwhile, MemDisst runs efficiently at the above real-time speed of 89 FPS, which is $53\%$ and $162\%$ faster than MBPTrack and CXTrack, respectively. This lightweight MemDisst tracker + +
MethodWOD [53]nuScenes [5]FPS
Category (No. Frames)Vehicle (185731)Pedestrian (241752)Mean (427483)Car (15578)Pedestrian (8019)Truck (3710)Bicycle (501)Mean (27808)- -
SuccessP2B52.617.933.027.015.921.520.022.946
SC3D---25.014.225.717.021.82
BAT54.718.234.122.517.319.317.020.565*
V2B57.623.738.431.317.321.722.225.837
STNet59.725.540.432.219.122.321.226.935
SiamDisst61.228.542.733.419.825.017.728.125
CXTrack57.130.742.229.620.427.618.526.534
M2Track61.132.044.6-----57
MBPTrack61.933.746.033.619.823.920.028.160*
MemDisst61.933.645.934.020.028.118.128.989*
CategoryVehiclePedestrianMeanCarPedestrianTruckBicycleMeanFPS
PrecisionP2B61.730.143.829.222.016.226.425.346
SC3D---27.116.221.918.223.12
BAT62.730.344.424.124.515.818.823.065*
V2B65.937.950.135.123.416.719.129.037
STNet68.039.952.136.127.216.829.230.835
SiamDisst70.544.555.837.330.420.123.732.825
CXTrack66.149.456.733.432.920.826.831.534
M2Track69.349.758.2-----57
MBPTrack71.952.761.037.632.720.729.233.860*
MemDisst71.852.861.138.032.223.924.134.289*
+ +Table 5: 3D tracking results on Waymo Open Dataset (WOD) [53] and nuScenes [5]. For a fair comparison, we compare our SiamDisst and MemDisst with memory-less Siamese-based trackers, and more complicated trackers with contexture memory or motion prediction, respectively. The best result in each comparison is shown in bold. * indicates the speed is re-evaluated on WOD for a fair comparison. + +shows its potential to serve as a simple and strong baseline in the 3D tracking community. + +# 5 Conclusion + +This paper improves the power of 3D point cloud tracking from two aspects. First, we design a novel 3D SOT framework that could fully benefit from the 3D self-supervised pre-training, which shows that the self-supervised pre-training weights (e.g., Point-MAE [42] and Recon [49]) are effective to improve the 3D tracking performance, especially in the low data regime (e.g., KITTI-Van and KITTI-Cyclist). Second, we bridge the gap between 2D and 3D tracking domains by leveraging a target-aware projection module, which enables to use a powerful 2D pre-trained tracker to facilitate the learning of 3D matching in existing 3D SOT trackers [27, 70] via matching distillation. We show that our proposed SiamDisst and MemDisst trackers can achieve state-of-the-art performance on KITTI and Waymo Open datasets, while running at real-time speed. + +# Acknowledgment + +This work was supported by a Strategic Research Grant from City University of Hong Kong (Project No. 7005665). + +# References + +1. Anthes, C., García-Hernández, R.J., Wiedemann, M., Kranzlüller, D.: State of the art of virtual reality technology. In: 2016 IEEE aerospace conference. pp. 1-19. IEEE (2016) 1 +2. Ben-Baruch, E., Karklinsky, M., Biton, Y., Ben-Cohen, A., Lawen, H., Zamir, N.: It's all in the head: Representation knowledge distillation through classifier sharing. In: arXiv:2201.06945 (2022) 5 +3. Bertinetto, L., Valmadre, J., Henriques, J.F., Vedaldi, A., Torr, P.H.: Fully convolutional siamese networks for object tracking. In: Proceedings of the European Conference on Computer Vision. pp. 850-865. Springer (2016) 4 +4. Bhat, G., Danelljan, M., Gool, L.V., Timofte, R.: Leanring discriminative model prediction for tracking. In: ICCV. pp. 6182-6191 (2019) 5 +5. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11621-11631 (2020) 2, 4, 10, 14 +6. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015) 7, 12 +7. Chen, X., Yan, B., Zhu, J., Wang, D., Yang, X., Lu, H.: Transformer tracking. In: CVPR. pp. 8126-8135 (2021) 5 +8. Cui, Y., Fang, Z., Shan, J., Gu, Z., Zhou, S.: 3d object tracking with transformer. arXiv preprint arXiv:2110.14921 (2021) 4, 12, 13 +9. Cui, Y., Jiang, C., Wang, L., Wu, G.: Mixformer: End-to-end tracking with iterative mixed attention. In: CVPR (2022) 5 +0. Danelljan, M., Bhat, G., Khan, F.S., Felsberg, M.: Eco: Efficient convolution operators for tracking. In: CVPR. pp. 21-26 (2017) 4 +1. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arxiv 2020. arXiv preprint arXiv:2010.11929 (2010) 6, 12 +2. Duong, C.N., Luu, K., Quach, K.G., Shrinkteanet, N.L.: Shrinkteanet: Millionscale lightweight face recognition via shrinking teacher-student networks. In: arXiv:1905.10620 (2019) 5 +3. Fan, H., Lin, L., Yang, F.: Lasot: A high-quality benchmark for large-scale single object tracking. In: CVPR. pp. 5374-5383 (2019) 5, 10 +4. Fan, H., Ling, H.: Siamese cascaded region proposal networks for real-time visual tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7952-7961 (2019) 5 +5. Fang, Z., Zhou, S., Cui, Y., Scherer, S.: 3d-siamrpn: An end-to-end learning method for real-time 3d single object tracking using raw point cloud. IEEE Sensors Journal 21(4), 4995-5011 (2020) 12, 13 + +16. Feng, S., Liang, P., Gao, J., Cheng, E.: Multi-correlation siamese transformer network with dense connection for 3d single object tracking. IEEE Robotics and Automation Letters (2023) 4 +17. Galoogahi, H., Fagg, A., Lucey, S.: Learning background-aware correlation filters for visual tracking. In: ICCV (2017) 4 +18. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. The International Journal of Robotics Research 32(11), 1231-1237 (2013) 2, 4, 10, 13 +19. Giancola, S., Zarzar, J., Ghanem, B.: Leveraging shape completion for 3d siamese tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1359-1368 (2019) 4, 10, 12, 13 +20. Girshick, R.: Fast r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 1440-1448 (2015) 4 +21. Guo, Z., Mao, Y., Zhou, W., Wang, M., Li, H.: Cmt: Context-matching-guided transformer for 3d tracking in point clouds. In: European Conference on Computer Vision. pp. 95-111. Springer (2022) 4, 12, 13 +22. He, K., Chen, X., Xie, S., Li, Y., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022) 2, 11, 12 +23. Henriques, J.F., Caseiro, R., Martins, P., Batista, J.: High-speed tracking with kernelized correlation filters. In: IEEE Transactions on Pattern Analysis and Machine Intelligence. pp. 583-596 (2015) 4 +24. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. In: arXiv:1503.02531 (2015) 5 +25. Huang, L., Zhao, X., Huang, K.: Got-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence (2019) 5, 10 +26. Hui, L., Wang, L., Cheng, M., Xie, J., Yang, J.: 3d siamese voxel-to-bev tracker for sparse point clouds. Advances in Neural Information Processing Systems 34, 28714-28727 (2021) 3, 4, 5, 6, 10, 12, 13 +27. Hui, L., Wang, L., Tang, L., Lan, K., Xie, J., Yang, J.: 3d siamese transformer network for single object tracking on point clouds. In: European Conference on Computer Vision. pp. 293-310. Springer (2022) 2, 3, 4, 6, 10, 12, 13, 14 +28. Jin, X., Peng, B., Wu, Y., Liu, Y., Liu, J.: Knowledge distillation via route constrained optimization. In: IEEE/CVF International Conference on Computer Vision (2019) 5 +29. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017) 2 +30. Kiran, B.R., Sobh, I., Talpaert, V., Mannion, P., Al Sallab, A.A., Yogamani, S., Pérez, P.: Deep reinforcement learning for autonomous driving: A survey. IEEE Transactions on Intelligent Transportation Systems (2021) 1 +31. Kristan, M., Matas, J., Leonardis, A., Vojir, T., Pflugfelder, R., Fernandez, G., Nebehay, G., Porikli, F., Cehovin, L.: A novel performance evaluation methodology for single-target trackers. IEEE Transactions on Pattern Analysis and Machine Intelligence (2016) 10 +32. Kristan, M., Matas, J., Danelljan, M.: The first visual object tracking segmentation votes2023 challenge results. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023) 5 + +33. Lan, K., Jiang, H., Xie, J.: Temporal-aware siamese tracker: Integrate temporal context for 3d object tracking. In: Proceedings of the Asian Conference on Computer Vision. pp. 399-414 (2022) 2, 4, 12, 13 +34. Lee, S.H., Kim, D.H., Song, B.C.: Self-supervised knowledge distillation using singular value decomposition. In: In Proceedings of the European Conference on Computer Vision (ECCV) (2018) 5 +35. Li, B., Wu, W., Wang, Q., Zhang, F., Xing, J., Yan, J.: Siamrpn++: Evolution of siamese visual tracking with very deep networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4282-4291 (2019) 5 +36. Li, B., Yan, J., Wu, W., Zhu, Z., Hu, X.: High performance visual tracking with siamese region proposal network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8971-8980 (2018) 4, 5 +37. Li, J., Guo, Z., Li, H., Han, S., Baek, J.w., Yang, M., Yang, R., Suh, S.: Rethinking feature-based knowledge distillation for face recognition. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) 5 +38. Lin, L., Fan, H., Zhang, Z., Xu, Y., Ling, H.: Swintrack: A simple and strong baseline for transformer tracking pp. 16743-16754 (2022) 2, 5 +39. Liu, Y., Liang, Y., Wu, Q., Zhang, L., Wang, H.: A new framework for multiple deep correlation filters based object tracking. In: ICASSP (2022) 4 +40. Muller, M., Bibi, A., S, G.: Trackingnet: A large-scale dataset and benchmark for object tracking in the wild. In: ECCV. pp. 300-317 (2018) 10 +41. Osep, A., Mehner, W., Mathias, M., Leibe, B.: Combined image-and world-space tracking in traffic scenes. In: IEEE International Conference on Robotics and Automation. pp. 1988-1995. IEEE (2017) 1 +42. Pang, Y., Wang, W., Tay, F.E., Liu, W., Tian, Y., Yuan, L.: Masked autoencoders for point cloud self-supervised learning. In: European conference on computer vision. pp. 604-621 (2022) 3, 6, 7, 10, 11, 12, 14 +43. Pang, Z., Li, Z., Wang, N.: Model-free vehicle tracking and state estimation in point cloud sequences. In: IROS (2021) 4 +44. Peng, B., Jin, X., Liu, J., Li, D., Wu, Y., Liu, Y., Zhou, S., Zhang, Z.: Self-supervised knowledge distillation using singular value decomposition. In: IEEE/CVF International Conference on Computer Vision (2019) 5 +45. Peng, B., Jin, X., Liu, J., Li, D., Wu, Y., Liu, Y., Zhou, S., Zhang, Z.: Shrinkteanet: Million-scale lightweight face recognition via shrinking teacher-student networks. In: IEEE/CVF International Conference on Computer Vision (2019) 5 +46. Qi, C., Su, H., Mo, K., Guibas, L.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: IEEE conference on computer vision and pattern recognition (2017) 10 +47. Qi, H., Feng, C., Cao, Z., Zhao, F., Xiao, Y.: P2b: Point-to-box network for 3d object tracking in point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6329-6338 (2020) 2, 3, 5 +48. Qi, H., Feng, C., Cao, Z., Zhao, F., Xiao, Y.: P2b: Point-to-box network for 3d object tracking in point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6329-6338 (2020) 4, 12, 13 +49. Qi, Z., Dong, R., Fan, G., Ge, Z., Zhang, X., Ma, K., Yi, L.: Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining (2023) 3, 6, 7, 11, 12, 14 +50. Ran, T., Gavves, E., Smeulders, A.W.: Siamese instance search for tracking. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1420-1429 (2016) 4 + +51. Romero, A., Ballas, N.: Fitnets: Hints for thin deep nets. In: arXiv:1412.6550 (2014) 5 +52. Shan, J., Zhou, S., Fang, Z., Cui, Y.: Ptt: Point-track-transformer module for 3d single object tracking in point clouds. arXiv preprint arXiv:2108.06455 (2021) 12, 13 +53. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. p. 2446-2454 (2020) 4, 10, 14 +54. Tao, F., Wang, M.: Response-based distillation for incremental object detection. In: arXiv:2110.13471 (2021) 5 +55. Tao, F., Wang, M., Yuan, H.: Overcoming catastrophic forgetting in incremental object detection via elastic response distillation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022) 5 +56. Wang, N., Song, Y., Ma, C.: Unsupervised deep tracking. In: CVPR. pp. 3708-1317 (2019) 5 +57. Wang, Z., Xie, Q., Lai, Y.K., Wu, J., Long, K., Wang, J.: Mlvsnet: Multi-level voting siamese network for 3d visual tracking. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3101-3110 (2021) 12, 13 +58. Wang, Z., Yu, X., Rao, Y., Zhou, J., Lu, J.: P2p: Tuning pre-trained image models for point cloud analysis with point-to-pixel prompting. Advances in neural information processing systems 35, 14388-14402 (2022) 7 +59. Wu, Q., Chan, A.: Meta-graph adaptation for visual object tracking. In: ICME (2021) 4 +60. Wu, Q., Yan, Y., Liang, Y., Liu, Y., Wang, H.: Dsnet: Deep and shallow feature learning for efficient visual tracking. In: ACCV. pp. 119-134 (2018) 4 +61. Wu, Q., Wan, J., Chan, A.B.: Progressive unsupervised learning for visual object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2993-3002 (2021) 5 +62. Wu, Q., Yang, T., Liu, Z., Wu, B., Shan, Y., Chan, A.B.: Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14561-14571 (2023) 1, 2, 4, 5, 6, 8, 11, 12 +63. Wu, Q., Yang, T., Wu, W., Chan, A.B.: Scalable video object segmentation with simplified framework. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) 5 +64. Wu, Q., Sun, C., Wang, J.: Multi-level structure-enhanced network for 3d single object tracking in sparse point clouds. IEEE Robotics and Automation Letters 8(1), 9-16 (2022) 4 +65. Wu, Y., Lim, J., Yang, M.H.: Online object tracking: A benchmark. In: CVPR. pp. 2411-2418 (2013) 5 +66. Xia, Y., Gladkova, M., Wang, R., Li, Q., Stilla, U., Henriques, J.F., Cremers, D.: Casspr: Cross attention single scan place recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8461-8472 (2023) 1 +67. Xia, Y., Shi, L., Ding, Z., Henriques, J.F., Cremers, D.: Text2loc: 3d point cloud localization from natural language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14958-14967 (2024) 1 +68. Xia, Y., Wu, Q., Li, W., Chan, A.B., Stilla, U.: A lightweight and detector-free 3d single object tracker on point clouds. IEEE Transactions on Intelligent Transportation Systems (2023) 4, 5, 12, 13 + +69. Xu, T.X., Guo, Y.C., Lai, Y.K., Zhang, S.H.: Cxtrack: Improving 3d point cloud tracking with contextual information. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1084-1093 (2023) 4, 6, 10, 12, 13 +70. Xu, T.X., Guo, Y.C., Lai, Y.K., Zhang, S.H.: Mbptrack: Improving 3d point cloud tracking with memory networks and box priors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9911-9920 (October 2023) 2, 3, 4, 6, 10, 12, 13, 14 +71. Yan, B., Peng, H., Fu, J., Wang, D., Lu, H.: Learning spatio-temporal transformer for visual tracking. In: ICCV. pp. 10448-10457 (2021) 5 +72. Yang, T., Chan, A.B.: Learning dynamic memory networks for object tracking. In: Proceedings of the European Conference on Computer Vision. pp. 152-167 (2018) 5 +73. Ye, B., Chang, H., Ma, B., Shan, S., Chen, X.: Joint feature learning and relation modeling for tracking: A one-stream framework. In: European Conference on Computer Vision. pp. 341-357 (2022) 2, 3, 4, 5, 6, 7, 8, 10 +74. Yim, J., Joo, D., Bae, J., Kim, J.: A gift from knowledge distillation: Fast optimization, network minimization and transfer learning. In: IEEE conference on computer vision and pattern recognition (2017) 5 +75. Zhang, L., Gonzalez-Garcia, A., Weijer, J.v.d., Danelljan, M., Khan, F.S.: Learning the model update for siamese trackers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4010-4019 (2019) 5 +76. Zhang, Z., Peng, H.: Deeper and wider siamese networks for real-time visual tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4591-4600 (2019) 5 +77. Zheng, C., Yan, X., Gao, J., Zhao, W., Zhang, W., Li, Z., Cui, S.: Box-aware feature enhancement for single object tracking on point clouds. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13199-13208 (2021) 2, 3, 4, 5, 12, 13 +78. Zheng, C., Yan, X., Zhang, H., Wang, B., Cheng, S., Cui, S., Li, Z.: Beyond 3d siamese tracking: A motion-centric paradigm for 3d single object tracking in point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8111-8120 (2022) 4, 12, 13 +79. Zhou, C., Luo, Z., Luo, Y., Liu, T., Pan, L., Cai, Z., Zhao, H., Lu, S.: Pttr: Relational 3d point cloud object tracking with transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8531-8540 (2022) 4, 12, 13 +80. Zhou, H., Song, L., Chen, J., Zhou, Y., Guoli: Rethinking soft labels for knowledge distillation: A bias-variance tradeoff perspective. In: arXiv:2102.00650 (2021) 5 \ No newline at end of file diff --git a/2024/Boosting 3D Single Object Tracking with 2D Matching Distillation and 3D Pre-training/images.zip b/2024/Boosting 3D Single Object Tracking with 2D Matching Distillation and 3D Pre-training/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..30a91a1ea792b97e2279d23b9728ad8dc1af2bbc --- /dev/null +++ b/2024/Boosting 3D Single Object Tracking with 2D Matching Distillation and 3D Pre-training/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:caac9247184522bed376e98a16dfd16a849c8c16b79ed868c24260bb0dcbadd9 +size 561558 diff --git a/2024/Boosting 3D Single Object Tracking with 2D Matching Distillation and 3D Pre-training/layout.json b/2024/Boosting 3D Single Object Tracking with 2D Matching Distillation and 3D Pre-training/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..04e0c7cb655ad91279ff489785b5b62ca0eb77ff --- /dev/null +++ b/2024/Boosting 3D Single Object Tracking with 2D Matching Distillation and 3D Pre-training/layout.json @@ -0,0 +1,9666 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 150, + 111, + 462, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 111, + 462, + 148 + ], + "spans": [ + { + "bbox": [ + 150, + 111, + 462, + 148 + ], + "type": "text", + "content": "Boosting 3D Single Object Tracking with 2D Matching Distillation and 3D Pre-training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 153, + 167, + 460, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 167, + 460, + 181 + ], + "spans": [ + { + "bbox": [ + 153, + 167, + 460, + 181 + ], + "type": "text", + "content": "Qiangqiang " + }, + { + "bbox": [ + 153, + 167, + 460, + 181 + ], + "type": "inline_equation", + "content": "\\mathsf{W u}^{1}" + }, + { + "bbox": [ + 153, + 167, + 460, + 181 + ], + "type": "text", + "content": " , Yan Xia\\*2,3, Jia Wan4, and Antoni B. Chan1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 170, + 189, + 443, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 189, + 443, + 200 + ], + "spans": [ + { + "bbox": [ + 170, + 189, + 443, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 170, + 189, + 443, + 200 + ], + "type": "text", + "content": " Department of Computer Science, City University of Hong Kong" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 238, + 201, + 375, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 201, + 375, + 211 + ], + "spans": [ + { + "bbox": [ + 238, + 201, + 375, + 211 + ], + "type": "text", + "content": "2 Technical University of Munich" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 207, + 212, + 406, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 212, + 406, + 222 + ], + "spans": [ + { + "bbox": [ + 207, + 212, + 406, + 222 + ], + "type": "text", + "content": "3 Munich Center for Machine Learning (MCML)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 223, + 466, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 223, + 466, + 243 + ], + "spans": [ + { + "bbox": [ + 146, + 223, + 466, + 243 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 146, + 223, + 466, + 243 + ], + "type": "text", + "content": " School of Computer Science and Technology, Harbin Institute of Technology, Shenzhen" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 149, + 246, + 462, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 246, + 462, + 266 + ], + "spans": [ + { + "bbox": [ + 149, + 246, + 462, + 266 + ], + "type": "text", + "content": "qiangqwu2-c@my.cityu.edu.hk, yan.xia@tum.de, jiawan1998@gmail.com, abchan@cityu.edu.hk" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 292, + 452, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 292, + 452, + 533 + ], + "spans": [ + { + "bbox": [ + 159, + 292, + 452, + 533 + ], + "type": "text", + "content": "Abstract. 3D single object tracking (SOT) is an essential task in autonomous driving and robotics. However, learning robust 3D SOT trackers remains challenging due to the limited category-specific point cloud data and the inherent sparsity and incompleteness of LiDAR scans. To tackle these issues, we propose a unified 3D SOT framework that leverages 3D generative pre-training and learns robust 3D matching abilities from 2D pre-trained foundation trackers. Our framework features a consistent target-matching architecture with the widely used 2D trackers, facilitating the transfer of 2D matching knowledge. Specifically, we first propose a lightweight Target-Aware Projection (TAP) module, allowing the pre-trained 2D tracker to work well on the projected point clouds without further fine-tuning. We then propose a novel IoU-guided matching-distillation framework that utilizes the powerful 2D pre-trained trackers to guide 3D matching learning in the 3D tracker, i.e., the 3D template-to-search matching should be consistent with its corresponding 2D template-to-search matching obtained from 2D pre-trained trackers. Our designs are applied to two mainstream 3D SOT frameworks: memory-less Siamese and contextual memory-based approaches, which are respectively named SiamDisst and MemDisst. Extensive experiments show that SiamDisst and MemDisst achieve state-of-the-art performance on KITTI, Waymo Open Dataset and nuScenes benchmarks, while running at above real-time speed of 25 and 90 FPS on a RTX3090 GPU." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 552, + 229, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 552, + 229, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 552, + 229, + 563 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 575, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 575, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 575, + 482, + 648 + ], + "type": "text", + "content": "3D Single Object Tracking (SOT) has emerged as a basic 3D task for numerous practical applications in autonomous driving [30,67], robotics [41,66] and virtual reality [1]. Given a target described as a 3D bounding box (bbox) in the first frame, 3D SOT aims to track the target across all the frames in the 3D scene by predicting thebbox pose and position, which can be considered as a 3D extension of the 2D SOT task [62] but using a different input modality of 3D point clouds." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "text", + "content": "* Corresponding Author" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 306, + 212 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 306, + 212 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 306, + 212 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 306, + 212 + ], + "type": "image", + "image_path": "6b91c85c43066eab893cbc213a499c5a74a5f698278d387ac8aa7e361e91653d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 230, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 130, + 230, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 230, + 482, + 319 + ], + "type": "text", + "content": "Fig. 1: Comparison of our SiamDisst and MemDisst to state-of-the-art 3D SOT on the Van category of the KITTI dataset [18]. Compared with Siamese-based trackers (plotted as squares), our SiamDisst achieves the best performance while running at the real-time speed of 25 FPS. Our lightweight MemDisst outperforms the other contextual memory and motion-based approaches (plotted as circles), setting new SOTA performance on KITTI-Van, and runs efficiently at 90 FPS. This demonstrates that our designs effectively boost 3D matching with limited category-specific training data (e.g., the KITTI-Van training set only contains 1994 examples)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 463, + 211 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 463, + 211 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 463, + 211 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 463, + 211 + ], + "type": "image", + "image_path": "7588f4b20e70ff6f1c065388725e3bf7c0be719bf7a912100a5f9df01681eafe.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 348, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 482, + 468 + ], + "type": "text", + "content": "As illustrated in Fig. 2, existing 3D SOT methods [38,62,73] mainly focus on learning robust 3D tracking models from sparse point cloud datasets [5,18]. For example, 3D siamese trackers [27,33,47,77] follow a siamese matching paradigm for accurate localization. To leverage more historical cues, MBPTrack [70] further extends this framework to a contextual memory-matching mechanism. Although the 3D matching models proposed by these methods play a key role in obtaining favorable performance [5,18], the sparsity and incompleteness of the point cloud training limits their abilities. Moreover, a paucity of large-scale category-specific 3D tracking datasets leads to insufficient training data for effectively learning the 3D matching modules in existing 3D trackers." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 471, + 482, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 614 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 614 + ], + "type": "text", + "content": "In contrast to the 3D SOT task, various large-scale annotated video datasets [29] are available for training 2D SOT methods. Existing 2D deep trackers can fully benefit from the feature backbones learned from 2D image [22] or video pre-training methods [62], enabling the learning of robust matching modules for 2D SOT on videos. The state-of-the-art trackers [38,62,73] with strong matching abilities show excellent tracking results even without using online memory mechanisms. With the remarkable success of 2D SOT, our work is motivated by the following questions: 1) Inspired by 2D SOT, can we design a general and effective 3D SOT framework that benefits from 3D pre-training? 2) Can we transfer the well-learned matching ability of state-of-the-art 2D trackers to 3D matching in 3D SOT trackers, so as to facilitate training from limited category-specific point cloud data?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "The main challenge of achieving the transfer of the 2D matching ability lies in two aspects. First, there is a significant misalignment between the input modalities for 2D and 3D trackers, dense 2d images and sparse 3d point clouds, respectively. Note that in the 3D SOT setting, only the single modality of point" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "text", + "content": "Q. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 114, + 246, + 225 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 246, + 225 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 246, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 246, + 225 + ], + "type": "image", + "image_path": "2b1c7e2963833f85e0e21dfcca44039fc3b73ae116c6d54e3f0ac5b7d3f188ad.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 232, + 239, + 240 + ], + "lines": [ + { + "bbox": [ + 163, + 232, + 239, + 240 + ], + "spans": [ + { + "bbox": [ + 163, + 232, + 239, + 240 + ], + "type": "text", + "content": "(a) Previous Siamese 3D SOT" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 258, + 114, + 463, + 227 + ], + "blocks": [ + { + "bbox": [ + 258, + 114, + 463, + 227 + ], + "lines": [ + { + "bbox": [ + 258, + 114, + 463, + 227 + ], + "spans": [ + { + "bbox": [ + 258, + 114, + 463, + 227 + ], + "type": "image", + "image_path": "5f064ab3210a8271c795d2d5bbd54b93213e202ad044b6e7d9f055ccdc9c929f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 232, + 435, + 240 + ], + "lines": [ + { + "bbox": [ + 285, + 232, + 435, + 240 + ], + "spans": [ + { + "bbox": [ + 285, + 232, + 435, + 240 + ], + "type": "text", + "content": "(b) Our Proposed 3D SOT (i.e., SiamDisst and MemDisst)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 258, + 482, + 358 + ], + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 358 + ], + "type": "text", + "content": "Fig. 2: The comparison of 3D SOT frameworks (a) the previous Siamese 3D SOT trackers [27,47,77] focus on learning the 3D matching module from limited category-specific point cloud (PC) training data. (b) The proposed 3D SOT framework bridges the gap between 2D and 3D SOT by using our designed target-aware projection (TAP) module, which enables the 2D tracker [73] trained on large-scale 2D videos to operate well on the point clouds. The well-learned 2D matching pattern is further distilled to facilitate the 3D matching learning. Note that only the 3D-related modules (denoted as blue components) are optimized during the distillation training stage. The 2D tracker and TAP module are not used during online inference to increase speed." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 386, + 482, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 446 + ], + "type": "text", + "content": "cloud data is used as input, which makes it difficult to directly apply the current 2D SOT trackers. Second, how to effectively transfer the 2D matching knowledge in pre-trained 2D trackers to guide the learning of the 3D matching in 3D trackers is unclear given the differences in internal representations between 2D and 3D SOT, i.e., 2d dense feature maps and unordered 3D point embeddings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 448, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 616 + ], + "type": "text", + "content": "To tackle the aforementioned challenges, we propose a general 3D SOT framework that can effectively utilize knowledge learned in 2D SOT, which is illustrated in Fig. 2(b). We first propose to align the different input modalities of 2D and 3D trackers via a learnable target-aware projection (TAP) module, which takes the raw point cloud as the input and projects them into the 2D image space. These projected images are suitable for 2D target matching and 2D bounding box regression by 2D pre-trained trackers, i.e., no fine-tuning of 2D trackers is required. Secondly, we propose a matching distillation approach that uses the knowledge in the high-quality 2D matching patterns from the 2D tracker as guidance for learning the matching patterns for 3D SOT. The proposed framework employs a standard Vision Transformer (ViT) as the matching module, which can naturally benefit from 3D pre-training approaches [42,49] and also creates a design that is consistent with the 2D SOT trackers (e.g., OSTrack [73]) for more natural distillation of matching knowledge." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "We implement our framework with both Siamese [26, 27] and contextual memory-based [70] matching paradigms, which are respectively named SiamDisst and MemDisst. Extensive experiments demonstrate that our SiamDisst achieves significant improvements over its baseline, and MemDisst sets new state-of-the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 260, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 260, + 91, + 448, + 102 + ], + "type": "text", + "content": "2D Matching Distillation and 3D Pre-training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "art performance on KITTI, while respectively running at a fast speed of 90 FPS on a single RTX3090 GPU." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 140, + 380, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 140, + 380, + 151 + ], + "spans": [ + { + "bbox": [ + 146, + 140, + 380, + 151 + ], + "type": "text", + "content": "In summary, the main contributions of our work are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 152, + 479, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 152, + 479, + 198 + ], + "spans": [ + { + "bbox": [ + 138, + 152, + 479, + 198 + ], + "type": "text", + "content": "- We build a general SOT framework that employs a standard ViT as the target matching module, which is consistent with dominant 2D trackers [62, 73] for better knowledge transfer of matching patterns, and meanwhile can naturally benefit from 3D pre-training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 201, + 479, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 201, + 479, + 246 + ], + "spans": [ + { + "bbox": [ + 137, + 201, + 479, + 246 + ], + "type": "text", + "content": "- We propose a lightweight target-aware projection (TAP) module to bridge the gap between the 2D and 3D tracking domains, which enables 2D pretrained trackers to be operated on point cloud data without additional 2D tracker fine-tuning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 248, + 479, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 248, + 479, + 294 + ], + "spans": [ + { + "bbox": [ + 138, + 248, + 479, + 294 + ], + "type": "text", + "content": "- We propose an IoU-guided matching distillation approach to guide 3D SOT learning, which facilitates transfer of knowledge from powerful 2D pre-trained foundation trackers to the 3D tracker. We show that the matching distillation facilitates the 3D model learning in the low-data regime." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 316, + 236, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 316, + 236, + 328 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 236, + 328 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 342, + 480, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 480, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 480, + 484 + ], + "type": "text", + "content": "3D Single Object Tracking. Inspired by the great success in the 2D tracking community, much progress has been made in 3D SOT [43]. The pioneering work SC3D [19] proposes the first 3D Siamese network to compute the similarity between the template and 3D candidate proposals. However, the time-consuming matching pipeline in SC3D is not end-to-end trainable, thus degrading its performance on 3D tracking benchmarks [5,18,53]. Inspired by the 2D region proposal network (RPN) [20,36], P2B [48] proposes a 3D RPN to generate high-quality 3D proposals, which achieves better tracking results while running at the real-time speed. BAT [77] extends P2B with a box-aware feature design, which provides a strong prior of the target shape. In contrast to previous approaches that directly regress 3D bboxes from a sparse point cloud, V2B [26] proposes a voxel-based localization head to localize the target in the Bird's Eye View (BEV)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 486, + 480, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 480, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 480, + 592 + ], + "type": "text", + "content": "With the development of transformers, 3D transformer SOT trackers (e.g., LTTR [8], PTTR [79] and CMT [21]) are designed for target-aware feature learning. Further extensions include iterative transformer blocks [27], historical cue modeling [33], contextual memory [69, 70], motion prediction [68, 78] and advanced architecture design [16, 64]. Despite these successes, learning robust 3D tracking models is still limited by the sparse and insufficient category-specific point cloud data. In this work, our aim is to leverage off-the-shelf 2D trackers trained on large-scale video data for enhancing 3D trackers, thus alleviating the current data bottleneck in 3D SOT." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 480, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 664 + ], + "type": "text", + "content": "2D Single Object Tracking. Traditional 2D SOT [10,17,23,39,59,60] follow a correlation filter tracking framework due to its fast running speed and favorable tracking performance. With the development of the deep learning, much progress has been made in designing convolutional neural networks for video object tracking. Specifically, SiamFC [3] and SINT [50] employ a siamese neural network and treat tracking as a template matching problem. Based on SiamFC, many" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "improvements have been made, including architecture design [14,35,36,76], unsupervised representation learning [56,61,62] and template updating [4,72,75]. Recent advances using transformers [7,9,38,63,71,73] achieve state-of-the-art performance on popular tracking benchmarks [13,25,32,65]. The great success in 2D tracking significantly inspires the development in the 3D tracking community. However, there is no previous work trying to bridge the gap between the 2D and 3D tracking areas. In this paper, we propose a lightweight target-aware projection (TAP) module to bridge the gap between 2D and 3D SOT, and use the 2D tracker to effectively guide the 3D learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 415 + ], + "type": "text", + "content": "Knowledge Distillation. Knowledge distillation (KD) aims to transfer the knowledge from a teacher model to a student model. Previous response-based KD approaches [24,54,55,80] focus on using the final output of the teacher model to supervise the training of the student model. To supervise the intermediate feature learning, various feature-based KD approaches [2,12,28,37,45,51] have been proposed. FitNet [51] first uses the intermediate representations learned by the teacher as hints to distill a student model. However, the intermediate supervision is unsuitable in our cross-modality distillation case, since the 2D and 3D representations have a large gap. Unlike the above approaches, relation-based KD [34,44] aims to leverage the relationships between various samples or feature maps, e.g., FPS [74] uses the inner products of feature maps for distillation. However, naively applying feature correlation as supervision is not helpful in 2D-to-3D tracking, which is due to the transfer noise and causes performance degradation (see Table 2). In this paper, we propose a novel IoU-guided matching distillation approach for better 2D knowledge transfer, which is also the first attempt to perform cross-modality transfer for 3D SOT." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 434, + 233, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 233, + 448 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 233, + 448 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 459, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 482, + 544 + ], + "type": "text", + "content": "In this section, we first present our general 3D SOT framework with a standard ViT as the target relation modeling module, which could naturally benefit from 3D self-supervised pre-training and be consistent with 2D pre-trained trackers for better distillation. To use the 2D pre-trained model (frozen) for guiding 3D matching, we introduce a lightweight target-aware projection (TAP) module to bridge the domain gap between 2D and 3D tracking. Finally, we explore our IoU-guided matching distillation to boost the 3D tracking performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 559, + 303, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 559, + 303, + 572 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 303, + 572 + ], + "type": "text", + "content": "3.1 Unified 3D SOT Framework" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 579, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 579, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 579, + 482, + 666 + ], + "type": "text", + "content": "Revisit of 3D SOT. Given the 3D bbox in the first frame, the goal of 3D SOT is to accurately predict 3D bboxes in the following frames. Existing Siamese matching approaches [26,47,68,77] mainly use the point cloud (PC) data " + }, + { + "bbox": [ + 130, + 579, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}_1\\in" + }, + { + "bbox": [ + 130, + 579, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{\\hat{N}\\times 3}" + }, + { + "bbox": [ + 130, + 579, + 482, + 666 + ], + "type": "text", + "content": " cropped in the first frame as the template to predict the 3D bbox " + }, + { + "bbox": [ + 130, + 579, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_t\\in" + }, + { + "bbox": [ + 130, + 579, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^7" + }, + { + "bbox": [ + 130, + 579, + 482, + 666 + ], + "type": "text", + "content": " from the PC " + }, + { + "bbox": [ + 130, + 579, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_t\\in \\mathbb{R}^{N_s\\times 3}" + }, + { + "bbox": [ + 130, + 579, + 482, + 666 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 130, + 579, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 579, + 482, + 666 + ], + "type": "text", + "content": " -th search frame. " + }, + { + "bbox": [ + 130, + 579, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{N}" + }, + { + "bbox": [ + 130, + 579, + 482, + 666 + ], + "type": "text", + "content": " is the number of sampled target points within the annotated 3D bounding box, " + }, + { + "bbox": [ + 130, + 579, + 482, + 666 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 130, + 579, + 482, + 666 + ], + "type": "text", + "content": " is the number of points in each search frame, and " + }, + { + "bbox": [ + 130, + 579, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_t" + }, + { + "bbox": [ + 130, + 579, + 482, + 666 + ], + "type": "text", + "content": " is parameterized by its center coordinates" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 260, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 260, + 91, + 447, + 102 + ], + "type": "text", + "content": "2D Matching Distillation and 3D Pre-training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 115, + 460, + 220 + ], + "blocks": [ + { + "bbox": [ + 151, + 115, + 460, + 220 + ], + "lines": [ + { + "bbox": [ + 151, + 115, + 460, + 220 + ], + "spans": [ + { + "bbox": [ + 151, + 115, + 460, + 220 + ], + "type": "image", + "image_path": "c2fc1c2d824f0e586e6b0dd457511e02af10688ec65dbe30f1eb28460669cb71.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 238, + 480, + 262 + ], + "lines": [ + { + "bbox": [ + 131, + 238, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 131, + 238, + 480, + 262 + ], + "type": "text", + "content": "Fig. 3: The pipeline converting 3D point clouds (PC) to 2D images using our target-aware projection (TAP) module, followed by the 2D pre-trained tracker." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 284, + 480, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 480, + 346 + ], + "type": "text", + "content": "(x, y, and z), orientation " + }, + { + "bbox": [ + 130, + 284, + 480, + 346 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 284, + 480, + 346 + ], + "type": "text", + "content": " and box size (width, height and length). Recent approaches [69, 70] fully leverage contextual information and use more memory frames " + }, + { + "bbox": [ + 130, + 284, + 480, + 346 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}_i\\}_{i = t - k}^t" + }, + { + "bbox": [ + 130, + 284, + 480, + 346 + ], + "type": "text", + "content": " to predict " + }, + { + "bbox": [ + 130, + 284, + 480, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_t" + }, + { + "bbox": [ + 130, + 284, + 480, + 346 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 284, + 480, + 346 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 284, + 480, + 346 + ], + "type": "text", + "content": " is the number of memory frames. Instead of explicitly removing template points, a target mask set " + }, + { + "bbox": [ + 130, + 284, + 480, + 346 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{U}_i\\}_{i = t - k - 1}^{t - 1}" + }, + { + "bbox": [ + 130, + 284, + 480, + 346 + ], + "type": "text", + "content": " is used to identify the points within the 3D bbox, where " + }, + { + "bbox": [ + 130, + 284, + 480, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_i \\in \\mathbb{R}^{N_s \\times 1}" + }, + { + "bbox": [ + 130, + 284, + 480, + 346 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 346, + 481, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 481, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 481, + 465 + ], + "type": "text", + "content": "Target matching module. For simplicity and consistency, we use " + }, + { + "bbox": [ + 130, + 346, + 481, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{m} \\in \\mathbb{R}^{N_m \\times 3}" + }, + { + "bbox": [ + 130, + 346, + 481, + 465 + ], + "type": "text", + "content": " to represent the unified memory used in the above approaches, where " + }, + { + "bbox": [ + 130, + 346, + 481, + 465 + ], + "type": "inline_equation", + "content": "N_m = \\hat{N}" + }, + { + "bbox": [ + 130, + 346, + 481, + 465 + ], + "type": "text", + "content": " in memory-less Siamese approaches and " + }, + { + "bbox": [ + 130, + 346, + 481, + 465 + ], + "type": "inline_equation", + "content": "N_m = N_s \\times k" + }, + { + "bbox": [ + 130, + 346, + 481, + 465 + ], + "type": "text", + "content": " in memory approaches. We then obtain the search point features " + }, + { + "bbox": [ + 130, + 346, + 481, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t = \\phi(\\mathbf{p}_t) \\in \\mathbb{R}^{N_s \\times C}" + }, + { + "bbox": [ + 130, + 346, + 481, + 465 + ], + "type": "text", + "content": " and memory point features " + }, + { + "bbox": [ + 130, + 346, + 481, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_m = \\phi(\\mathbf{m}) \\in \\mathbb{R}^{N_m \\times C}" + }, + { + "bbox": [ + 130, + 346, + 481, + 465 + ], + "type": "text", + "content": " by applying a point feature extractor " + }, + { + "bbox": [ + 130, + 346, + 481, + 465 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot)" + }, + { + "bbox": [ + 130, + 346, + 481, + 465 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 130, + 346, + 481, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_t" + }, + { + "bbox": [ + 130, + 346, + 481, + 465 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 346, + 481, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 130, + 346, + 481, + 465 + ], + "type": "text", + "content": " respectively. The current 3D SOT trackers mainly focus on designing a robust 3D matching module " + }, + { + "bbox": [ + 130, + 346, + 481, + 465 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\mathbf{f}_s, \\mathbf{f}_m)" + }, + { + "bbox": [ + 130, + 346, + 481, + 465 + ], + "type": "text", + "content": " to perform the target matching between the search and memory point features. However, learning this matching module is challenging due to the sparsity of LIDAR point clouds and the lack of sufficient category-specific training data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 466, + 481, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 466, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 466, + 481, + 548 + ], + "type": "text", + "content": "To tackle these issues, inspired by 2D SOT [62, 73], we propose to use a standard ViT [11] for 3D matching since: 1) the standard ViT can benefit from 3D self-supervised pre-training (e.g., [42, 49]), which eases the need for large-scale 3D annotated data for 3D SOT training; 2) it has the consistent target relation modelling architecture (i.e., ViT) with dominant 2D trackers, which could more naturally transfer its 2D matching knowledge. The 3D matching process with ViT is formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 255, + 557, + 481, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 557, + 481, + 571 + ], + "spans": [ + { + "bbox": [ + 255, + 557, + 481, + 571 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {f}} _ {t} = \\mathcal {M} \\left(\\mathbf {f} _ {t}, \\mathbf {f} _ {m} + \\theta (\\mathcal {U})\\right), \\tag {1}", + "image_path": "fff44ac31d4096a2eed61ca6181448d46ef8164815595467b93a7dc8a30196c9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "where the output " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}_t\\in \\mathbb{R}^{N_s\\times C}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " is the updated search features output from the last layer of " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\theta (\\cdot)" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " is a learnable linear mapping that converts the 1-dim masks into the " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " -dim feature maps. For siamese tracking paradigms, we drop this operation since the template is well cropped. Finally, the matching output " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}_t" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " is fed to the prediction head [26, 27, 70] for predicting the 3D bbox " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_t" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ". Briefly, we call the memory-less Siamese-based and memory-based approaches with our standard ViT matching module as SiamBase and MemBase (see Supp. for more details)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 328, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 328, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 328, + 128 + ], + "type": "text", + "content": "3.2 Target-Aware Projection Module" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 266 + ], + "type": "text", + "content": "Although SiamBase and MemBase can naturally benefit from 3D self-supervised pre-training [42, 49], their performance is still limited due to the absence of large-scale 3D datasets (e.g., ShapeNet [6] only contains 55 object categories with 50k point clouds) for obtaining effective pre-training weights. To further improve the 3D tracking performance, we propose to employ a powerful 2D pre-trained tracker (denoted as " + }, + { + "bbox": [ + 130, + 136, + 482, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{2D}" + }, + { + "bbox": [ + 130, + 136, + 482, + 266 + ], + "type": "text", + "content": ") as the distillation teacher to guide the 3D matching learning in SiamBase and MemBase. However, in 3D SOT, only the single modality of point cloud data is available, making it hard to directly utilize current 2D SOT trackers. To bridge the modality gap between 2D and 3D tracking, we thus propose a learnable target-aware projection (TAP) module in this subsection, illustrated in Fig. 3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 267, + 482, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 482, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 482, + 339 + ], + "type": "text", + "content": "3D-to-2D Projection. Using the domain knowledge that the objects-of-interest for 3D tracking are typically on the ground plane (at similar fixed z-level heights), we project the search point cloud " + }, + { + "bbox": [ + 130, + 267, + 482, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_t" + }, + { + "bbox": [ + 130, + 267, + 482, + 339 + ], + "type": "text", + "content": " from the orthogonal view along the " + }, + { + "bbox": [ + 130, + 267, + 482, + 339 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 267, + 482, + 339 + ], + "type": "text", + "content": " axis. For each point, we round down the " + }, + { + "bbox": [ + 130, + 267, + 482, + 339 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 130, + 267, + 482, + 339 + ], + "type": "text", + "content": " coordinates to obtain its location on the 2D " + }, + { + "bbox": [ + 130, + 267, + 482, + 339 + ], + "type": "inline_equation", + "content": "xy" + }, + { + "bbox": [ + 130, + 267, + 482, + 339 + ], + "type": "text", + "content": " plane, i.e., from the bird-eye view (BEV). We denote this 3D-to-2D projection as " + }, + { + "bbox": [ + 130, + 267, + 482, + 339 + ], + "type": "inline_equation", + "content": "\\psi (\\mathbf{p}_t)" + }, + { + "bbox": [ + 130, + 267, + 482, + 339 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 340, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 411 + ], + "type": "text", + "content": "Target-Aware Projection. Based on [58], the point features " + }, + { + "bbox": [ + 130, + 340, + 482, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t" + }, + { + "bbox": [ + 130, + 340, + 482, + 411 + ], + "type": "text", + "content": " can be mapped to the 2D plane based on the corresponding projection " + }, + { + "bbox": [ + 130, + 340, + 482, + 411 + ], + "type": "inline_equation", + "content": "\\psi(\\mathbf{p}_t)" + }, + { + "bbox": [ + 130, + 340, + 482, + 411 + ], + "type": "text", + "content": ". However, simply projecting search features to the 2D plane is not optimal since it does not consider any target information, which may not well support the subsequent 2D matching (see Table 2). Therefore, we use a cross-attention module " + }, + { + "bbox": [ + 130, + 340, + 482, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 340, + 482, + 411 + ], + "type": "text", + "content": " here to obtain more effective target-aware search features:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 173, + 420, + 480, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 420, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 173, + 420, + 480, + 434 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {t} ^ {*} = \\mathcal {C} \\left(Q (\\mathbf {f} _ {t}), K \\left(\\mathbf {f} _ {m} + \\theta (\\mathcal {U}) + E (\\mathbf {m})\\right), V \\left(\\mathbf {f} _ {m} + \\theta (\\mathcal {U}) + E (\\mathbf {m})\\right)\\right), \\tag {2}", + "image_path": "a9c8402d2ec60e507f2dcd576c62b7cca36d9a3e2dea8ca30296509f2dbd16f9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 442, + 482, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 442, + 482, + 503 + ], + "spans": [ + { + "bbox": [ + 131, + 442, + 482, + 503 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 442, + 482, + 503 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 131, + 442, + 482, + 503 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 442, + 482, + 503 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 442, + 482, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 442, + 482, + 503 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 131, + 442, + 482, + 503 + ], + "type": "text", + "content": " are query, key and value embedding modules, respectively. " + }, + { + "bbox": [ + 131, + 442, + 482, + 503 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 131, + 442, + 482, + 503 + ], + "type": "text", + "content": " is the 3D positional embedding module. We project the target-aware search features " + }, + { + "bbox": [ + 131, + 442, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t^* \\in \\mathbb{R}^{N_s \\times C}" + }, + { + "bbox": [ + 131, + 442, + 482, + 503 + ], + "type": "text", + "content": " into the 2D plane via " + }, + { + "bbox": [ + 131, + 442, + 482, + 503 + ], + "type": "inline_equation", + "content": "\\psi(\\mathbf{p}_t)" + }, + { + "bbox": [ + 131, + 442, + 482, + 503 + ], + "type": "text", + "content": " and obtain the projected feature maps with " + }, + { + "bbox": [ + 131, + 442, + 482, + 503 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 131, + 442, + 482, + 503 + ], + "type": "text", + "content": " channels, which are further input to a lightweight CNN " + }, + { + "bbox": [ + 131, + 442, + 482, + 503 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 131, + 442, + 482, + 503 + ], + "type": "text", + "content": " [58] to reduce its channel dimension, yielding the projected search image:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 246, + 510, + 480, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 510, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 246, + 510, + 480, + 525 + ], + "type": "interline_equation", + "content": "I _ {t} = g \\left(\\psi \\left(\\mathbf {p} _ {t}\\right)\\right) \\in \\mathbb {R} ^ {H _ {t} \\times W _ {t} \\times 3}, \\tag {3}", + "image_path": "1f1898fea52e493d73e7d52dd715a33cf024cee8643dbad1380bc3c7d3a5dccd.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 533, + 482, + 570 + ], + "type": "inline_equation", + "content": "H_{t}" + }, + { + "bbox": [ + 130, + 533, + 482, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 482, + 570 + ], + "type": "inline_equation", + "content": "W_{t}" + }, + { + "bbox": [ + 130, + 533, + 482, + 570 + ], + "type": "text", + "content": " are height and width of the projected search image. The same operations are made on each memory frame and its corresponding features, thus we obtain " + }, + { + "bbox": [ + 130, + 533, + 482, + 570 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 533, + 482, + 570 + ], + "type": "text", + "content": " projected memory images " + }, + { + "bbox": [ + 130, + 533, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\{I_m^i\\}_{i = t - k - 1}^{t - 1}" + }, + { + "bbox": [ + 130, + 533, + 482, + 570 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 533, + 482, + 570 + ], + "type": "inline_equation", + "content": "I_{m}^{i}\\in \\mathbb{R}^{H_{m}\\times W_{m}\\times 3}" + }, + { + "bbox": [ + 130, + 533, + 482, + 570 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "TAP training. The TAP module is trained via a simulated 2D tracking process on the projected search and target images, so that the projected images can obtain good 2D tracking performance. As illustrated in Fig. 3, the projected 2D search and memory images are input to a ViT-based pre-trained 2D tracker (e.g., OSTrack [73]) for 2D bounding box regression. Specifically, the input images are flattened into a sequence of 2D patches, projected into image embeddings, and then concatenated along the spatial dimension to input to the 2D ViT model for joint feature extraction and matching. The updated search features output" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 260, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 260, + 91, + 448, + 102 + ], + "type": "text", + "content": "2D Matching Distillation and 3D Pre-training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 186, + 114, + 430, + 234 + ], + "blocks": [ + { + "bbox": [ + 186, + 114, + 430, + 234 + ], + "lines": [ + { + "bbox": [ + 186, + 114, + 430, + 234 + ], + "spans": [ + { + "bbox": [ + 186, + 114, + 430, + 234 + ], + "type": "image", + "image_path": "88a89314ba2d9f4f2c421e04db7229384f3cc4c7cbf930f2882d196a77f85c9a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 252, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 131, + 252, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 131, + 252, + 482, + 331 + ], + "type": "text", + "content": "Fig. 4: Visualization of the projected search image " + }, + { + "bbox": [ + 131, + 252, + 482, + 331 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 131, + 252, + 482, + 331 + ], + "type": "text", + "content": " obtained by our TAP module on (a) KITTI-Pedestrian and (b) KITTI-Cycler. (c): The attention maps are generated via template-to-search matching from the 2D tracker on (b) KITTI-Cycler. The red box indicates the ground-truth (GT) 2D box projected from the GT 3D box. The blue box denotes the prediction of the 2D pre-trained OSTrack [73]. Based on the target-aware search features, the TAP module can effectively encode the target region with unique color and texture, enabling subsequent 2D matching." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 353, + 482, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 482, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 482, + 378 + ], + "type": "text", + "content": "from the last layer is input to the prediction head for 2D box regression, which can be generally formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 220, + 385, + 481, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 385, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 220, + 385, + 481, + 402 + ], + "type": "interline_equation", + "content": "\\mathcal {B} _ {t} ^ {2 D}, \\hat {\\mathbf {f}} _ {t} ^ {2 D}, \\hat {\\mathbf {f}} _ {m} ^ {2 D} = \\mathcal {T} _ {2 D} (\\{I _ {m} ^ {i} \\} _ {i = t - k - 1} ^ {t - 1}, I _ {t}), \\tag {4}", + "image_path": "0ef0697260f844680f0373abd63f17423a050b5d1c89efcd2279406efa5d0ddb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 409, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 409, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 482, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 409, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_t^{2D} \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 130, + 409, + 482, + 495 + ], + "type": "text", + "content": " is the predicted 2D bounding box in the projected search frame, and " + }, + { + "bbox": [ + 130, + 409, + 482, + 495 + ], + "type": "inline_equation", + "content": "(\\hat{\\mathbf{f}}_t^{2D}, \\hat{\\mathbf{f}}_m^{2D})" + }, + { + "bbox": [ + 130, + 409, + 482, + 495 + ], + "type": "text", + "content": " are the updated 2D search and memory feature maps from the last layer of the 2D tracker. Here we use OSTrack as the 2D pre-trained tracker, and use its same 2D loss " + }, + { + "bbox": [ + 130, + 409, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{2D}" + }, + { + "bbox": [ + 130, + 409, + 482, + 495 + ], + "type": "text", + "content": " [62, 73] to evaluate the simulated 2D tracking result, thus providing supervision on the TAP module. The ground-truth 2D bounding box " + }, + { + "bbox": [ + 130, + 409, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t^{2D}" + }, + { + "bbox": [ + 130, + 409, + 482, + 495 + ], + "type": "text", + "content": " is obtained by projecting the corners of the ground-truth 3D box to the projected search image via the 3D-to-2D projection " + }, + { + "bbox": [ + 130, + 409, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 130, + 409, + 482, + 495 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 495, + 483, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 483, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 483, + 592 + ], + "type": "text", + "content": "After training the TAP module, the projection process is aligned with the 2D tracker. Note that we only update the parameters of the TAP module while keeping the parameters in the 2D pre-trained tracker " + }, + { + "bbox": [ + 130, + 495, + 483, + 592 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{2D}" + }, + { + "bbox": [ + 130, + 495, + 483, + 592 + ], + "type": "text", + "content": " fixed, which makes the TAP training more efficient, and less likely to overfit. We also visualize the projected 2D search images in Fig. 4. The learned TAP module can support accurate 2D matching in the projected search image space well via the pretrained OSTrack model, which shows the effectiveness of the proposed TAP module." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 608, + 328, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 608, + 328, + 621 + ], + "spans": [ + { + "bbox": [ + 131, + 608, + 328, + 621 + ], + "type": "text", + "content": "3.3 IoU-guided Matching Distillation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 628, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 628, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 482, + 667 + ], + "type": "text", + "content": "We introduce how to distill the 2D matching knowledge in " + }, + { + "bbox": [ + 130, + 628, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{2D}" + }, + { + "bbox": [ + 130, + 628, + 482, + 667 + ], + "type": "text", + "content": " to guide the 3D matching learning in our 3D SOT trackers, SiamBase and MemBase. In (1), we obtain the updated 3D search features " + }, + { + "bbox": [ + 130, + 628, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}_t\\in \\mathbb{R}^{N_s\\times C}" + }, + { + "bbox": [ + 130, + 628, + 482, + 667 + ], + "type": "text", + "content": " and updated 3D memory" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "text", + "content": "Q. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 117, + 462, + 237 + ], + "blocks": [ + { + "bbox": [ + 153, + 117, + 462, + 237 + ], + "lines": [ + { + "bbox": [ + 153, + 117, + 462, + 237 + ], + "spans": [ + { + "bbox": [ + 153, + 117, + 462, + 237 + ], + "type": "image", + "image_path": "2c13193de7125ae7c3a33e1a672f1fd4544ca0e92a9003c528ad14b4484b5258.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 255, + 480, + 278 + ], + "lines": [ + { + "bbox": [ + 131, + 255, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 131, + 255, + 480, + 278 + ], + "type": "text", + "content": "Fig. 5: Training pipeline of our SiamDisst or MemDisst via IoU-guided matching distillation. During the training stage, only the SiamDisst or MemDisst model is updated." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 299, + 482, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 299, + 482, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 482, + 492 + ], + "type": "text", + "content": "features " + }, + { + "bbox": [ + 130, + 299, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}_m\\in \\mathbb{R}^{N_m\\times C}" + }, + { + "bbox": [ + 130, + 299, + 482, + 492 + ], + "type": "text", + "content": " output from the last layer of " + }, + { + "bbox": [ + 130, + 299, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 299, + 482, + 492 + ], + "type": "text", + "content": ". Similarly, the updated 2D search and memory feature maps, " + }, + { + "bbox": [ + 130, + 299, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}_t^{2D}\\in \\mathbb{R}^{\\frac{H_t}{s}\\times \\frac{W_t}{s}\\times C}" + }, + { + "bbox": [ + 130, + 299, + 482, + 492 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 299, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}_m^{2D}\\in \\mathbb{R}^{\\frac{H_m}{s}\\times \\frac{W_m}{s}\\times k\\times C}" + }, + { + "bbox": [ + 130, + 299, + 482, + 492 + ], + "type": "text", + "content": ", are obtained via (4) in the last layer of 2D ViT, where " + }, + { + "bbox": [ + 130, + 299, + 482, + 492 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 299, + 482, + 492 + ], + "type": "text", + "content": " denotes the patch size for patch embedding. The 3D and 2D features belong to different embedding spaces, and thus directly guiding the feature embeddings of the 3D tracker using the 2D tracker is not straightforward (see Supplementary). Instead, we use the matching correspondences (between feature maps) used by the 2D tracker to guide the matching correspondences (between feature vectors) of the 3D tracker. 3D-to-2D feature indexing. Given a search point coordinate in the search point cloud " + }, + { + "bbox": [ + 130, + 299, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_t" + }, + { + "bbox": [ + 130, + 299, + 482, + 492 + ], + "type": "text", + "content": ", we index its corresponding 2D feature vector via the 3D-to-2D projection " + }, + { + "bbox": [ + 130, + 299, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 130, + 299, + 482, + 492 + ], + "type": "text", + "content": ". Specifically, we first obtain its 2D location in the projected search image. Then, we interpolate its 2D features " + }, + { + "bbox": [ + 130, + 299, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}_t^{2D}" + }, + { + "bbox": [ + 130, + 299, + 482, + 492 + ], + "type": "text", + "content": " to its original spatial size " + }, + { + "bbox": [ + 130, + 299, + 482, + 492 + ], + "type": "inline_equation", + "content": "H_{t}\\times W_{t}" + }, + { + "bbox": [ + 130, + 299, + 482, + 492 + ], + "type": "text", + "content": " via bilinear interpolation, and finally index its feature vector via the projected 2D location. Based on the above 3D-to-2D feature indexing, we obtain the 2D search and memory features, denoted as " + }, + { + "bbox": [ + 130, + 299, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{f}}_t\\in \\mathbb{R}^{N_s\\times C}" + }, + { + "bbox": [ + 130, + 299, + 482, + 492 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 299, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{f}}_m\\in \\mathbb{R}^{N_m\\times C}" + }, + { + "bbox": [ + 130, + 299, + 482, + 492 + ], + "type": "text", + "content": ", that correspond to the 3D search and memory point clouds." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 492, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 482, + 553 + ], + "type": "text", + "content": "Matching distillation. Note that " + }, + { + "bbox": [ + 130, + 492, + 482, + 553 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{f}}_t" + }, + { + "bbox": [ + 130, + 492, + 482, + 553 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 492, + 482, + 553 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{f}}_m" + }, + { + "bbox": [ + 130, + 492, + 482, + 553 + ], + "type": "text", + "content": " are obtained from the 2D ViT in the 2D tracker, which is well learned on large-scale 2D annotated videos, thus providing more robust matching results than our 3D counterparts (SimBase and MemBase) trained on limited category-specific point cloud data. Thus, we use the 2D memory-to-search matching pattern to guided the 3D matching process:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 220, + 570, + 481, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 570, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 220, + 570, + 481, + 586 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d i s t} = \\mathrm {I o U} (\\mathcal {B} _ {t} ^ {2 D}, \\mathcal {G} _ {t} ^ {2 D}) \\cdot \\mathcal {L} _ {1} (\\hat {\\mathbf {f}} _ {t} \\hat {\\mathbf {f}} _ {m} ^ {T}, \\tilde {\\mathbf {f}} _ {t} \\tilde {\\mathbf {f}} _ {m} ^ {T}), \\tag {5}", + "image_path": "82acbf7d5507e1bf7a00c062b0cbefae8d07a545800f3f8f777071cf371bead5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}_t\\hat{\\mathbf{f}}_m^T" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": " is the inner-product (matching) matrix between the 3D search and memory features, and likewise for " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{f}}_t\\tilde{\\mathbf{f}}_m^T" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": " is the L1 loss that calculates the mean absolute error between two matrices. IoU " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": " calculates the intersection-over-union between the predicted 2D box " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_t^{2D}" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": " and projected 2D ground-truth box " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t^{2D}" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": ", which measures the reliability of our 2D tracker, i.e. if the 2D matching is reliable enough and predicts the 2D box accurately (i.e., with high IoU)," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 260, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 260, + 91, + 448, + 102 + ], + "type": "text", + "content": "2D Matching Distillation and 3D Pre-training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "this 2D matching pattern will be used more for guiding the 3D matching. We also use the typical 3D box regression losses " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{3D}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " in the siamese paradigm (i.e., STNet [27]) and memory-based paradigm (i.e., MBPTrack [70]) to supervise our SiamBase and MemBase, respectively. The overall training loss is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 255, + 175, + 479, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 175, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 255, + 175, + 479, + 186 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {o v e r a l l}} = \\mathcal {L} _ {3 D} + \\lambda \\mathcal {L} _ {\\text {d i s t}}, \\tag {6}", + "image_path": "c51781eab2099a8d05ce6584db72480dee7da128926f1090709699c37d6c7501.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 197, + 479, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 197, + 479, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 197, + 479, + 268 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 197, + 479, + 268 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 197, + 479, + 268 + ], + "type": "text", + "content": " is a hyperparameter. After matching distillation, we denote the updated 3D SOT trackers as SiamDisst and MemDisst. Note that the TAP module, the 2D pre-trained tracker and the matching distillation are only used during the offline training stage to improve the 3D matching ability. During online inference, these modules are removed, and only our 3D trackers SiamDisst and MemDisst are required, which run at the real-time speed of 25 and 90 FPS, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 288, + 229, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 288, + 229, + 301 + ], + "spans": [ + { + "bbox": [ + 131, + 288, + 229, + 301 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 314, + 481, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 481, + 505 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 481, + 505 + ], + "type": "text", + "content": "Implementation details. For training the TAP module, as illustrated in Fig. 3, we sample memory and search PCs as the input, and project them into the 2D images for further matching in the 2D tracker. We use OSTrack [73] trained on large-scale 2D video datasets [13, 25, 40], as the 2D tracker, and freeze its weights, thus leading to efficient TAP training. The original 2D loss in OSTrack is used to evaluate the 2D predictions without modification. For TAP training, we use a batch size of 8 with a learning rate of 1e-4 and a total epoch of 40. For SiamDisst, we use the same hierarchical backbone network and 3D prediction head in STNet [27] for initial PC feature extraction and 3D prediction. For MemDisst, we use the same lightweight PointNet [46] with Point-MAE [42] as the backbone, and use the prediction head and memory mechanism in MBPTrack [70] for a fair comparison. We employ ViT-384-6 and ViT-96-3 as matching modules for SiamDisst and MemDisst respectively, which is illustrated in Table 3. The hyper-parameter " + }, + { + "bbox": [ + 130, + 314, + 481, + 505 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 314, + 481, + 505 + ], + "type": "text", + "content": " is set to 1.0. The training and testing settings of SiamDisst and MemDisst are following their baselines STNet and MBPTrack. More details on implementation can be found in the Supplementary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 506, + 479, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 479, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 479, + 601 + ], + "type": "text", + "content": "Datasets and Evaluation. We use three 3D datasets, KITTI [18], nuScenes [5] and Waymo Open Dataset (WOD) [53], for training and testing our proposed approaches. The KITTI dataset contains 21 training video sequences and 29 test sequences. Following previous work [19], the training set is split into three parts: 0-16 for training, 17-18 for validation and 19-20 for testing. Following [26,27,69], we use nuScenes and WOD to test the generalization ability of our approaches. We follow the one one pass evaluation [31], which is widely used in previous 3D SOT works [19,26,27,70], to report Precision and Success metrics for evaluation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 620, + 238, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 620, + 238, + 632 + ], + "spans": [ + { + "bbox": [ + 131, + 620, + 238, + 632 + ], + "type": "text", + "content": "4.1 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Component analysis. We evaluate the effect of the two main components used in our proposed framework: 3D pre-training and 2D matching distillation." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 152, + 114, + 460, + 193 + ], + "blocks": [ + { + "bbox": [ + 152, + 114, + 460, + 193 + ], + "lines": [ + { + "bbox": [ + 152, + 114, + 460, + 193 + ], + "spans": [ + { + "bbox": [ + 152, + 114, + 460, + 193 + ], + "type": "table", + "html": "
VariantBaseline3D Pre-training2D DistillationSuccessPrecision
SiamBase✓55.664.3
SiamBase✓60.369.3
SiamBase✓✓63.575.0
MemBase✓61.372.7
MemBase✓63.576.5
MemBase✓✓66.679.3
", + "image_path": "6ab9b6b44367c9ed927f73de24060e7e916a15ccceadef6dde54e937f3a05149.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 146, + 237, + 460, + 296 + ], + "blocks": [ + { + "bbox": [ + 132, + 194, + 480, + 214 + ], + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 214 + ], + "type": "text", + "content": "Table 1: Component analysis of SiamDisst and MemDisst. 'Baseline' indicates the variant initialized with the random weights and trained from scratch." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 146, + 237, + 460, + 296 + ], + "lines": [ + { + "bbox": [ + 146, + 237, + 460, + 296 + ], + "spans": [ + { + "bbox": [ + 146, + 237, + 460, + 296 + ], + "type": "table", + "html": "
VariantSucc./Prec.Pre-trainingSucc./Prec.
SiamDisst63.5/75.0Point-MAE [42]63.5/75.0
unlearnable TAP61.1/72.3ReCon [49]62.9/73.6
TAP w/o target-aware interact.61.2/73.12D MAE [22]60.5/69.9
2D distill. w/o IoU weight.56.4/66.3DropMAE [62]61.9/74.0
", + "image_path": "98005893eebd8dfc6bdacf0fe13d44bd98a00d4289d61008835088d2b9dc056c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 296, + 479, + 316 + ], + "lines": [ + { + "bbox": [ + 132, + 296, + 479, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 296, + 479, + 316 + ], + "type": "text", + "content": "Table 2: Ablation studies of SiamDisst on KITTI-VAN: (left) variants of the TAP module and distillation; (right) using different 3D pre-trained models." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 352, + 480, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 480, + 424 + ], + "type": "text", + "content": "The 3D pre-training means the usage of the default pre-trained Point-MAE [42] weights for 3D initialization. As shown in Table 1, SiamDisst and MemDisst with 3D pre-training achieve " + }, + { + "bbox": [ + 130, + 352, + 480, + 424 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 352, + 480, + 424 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 352, + 480, + 424 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 130, + 352, + 480, + 424 + ], + "type": "text", + "content": " improvements over the naive baselines regarding the precision metric. By applying the 2D matching distillation, SiamDisst and MemDisst can be significantly improved, leading to " + }, + { + "bbox": [ + 130, + 352, + 480, + 424 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 130, + 352, + 480, + 424 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 352, + 480, + 424 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 130, + 352, + 480, + 424 + ], + "type": "text", + "content": " precision gains, which shows the effectiveness of our approach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 425, + 481, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 481, + 555 + ], + "type": "text", + "content": "TAP module. We study several variants of the proposed TAP module: including 1) unlearnable TAP: we remove the learnable projection module in the TAP, and use the z-axis value as the projected feature in the 2D image plane (repeated for each of the 3 image channels); 2) TAP module w/o target-aware design: we remove the cross attention module " + }, + { + "bbox": [ + 130, + 425, + 481, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 425, + 481, + 555 + ], + "type": "text", + "content": " and separately project the template and search images; Table 2 (left) presents the results. The unlearnable TAP obtains inferior performance, which is due to the low-quality projected images and more 2D tracking failures. Without the IoU weighting, the unlearnable TAP also has severely degraded performance (see Supplementary for more details). For the target-aware design, we can observe " + }, + { + "bbox": [ + 130, + 425, + 481, + 555 + ], + "type": "inline_equation", + "content": "2.1\\% / 1.9\\%" + }, + { + "bbox": [ + 130, + 425, + 481, + 555 + ], + "type": "text", + "content": " gains by using it, which shows that learning target-aware features is beneficial in the 2D projection stage." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "text", + "content": "IoU-guided matching distillation. In the IoU-guided matching distillation, we use the IoU to measure the reliability of the 2D prediction. To test its effectiveness, we remove this IoU weighting in (5) and show the comparison in Table 2 (left). The variant w/o the IoU weighting degrades the performance with large margins of " + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "inline_equation", + "content": "7.1\\% / 8.7\\%" + }, + { + "bbox": [ + 130, + 557, + 480, + 640 + ], + "type": "text", + "content": ", which is mainly because the 2D matching distillation may introduce some low-quality and inaccurate 2D matching patterns for the 3D matching learning, thus severely degrading the learning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "Comparison of 3D pre-trained weights. We use various pre-trained weights to initialize the ViT in our 3D tracker and test them in Table 2 (right). The" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 260, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 260, + 91, + 448, + 102 + ], + "type": "text", + "content": "2D Matching Distillation and 3D Pre-training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 207, + 114, + 406, + 171 + ], + "blocks": [ + { + "bbox": [ + 207, + 114, + 406, + 171 + ], + "lines": [ + { + "bbox": [ + 207, + 114, + 406, + 171 + ], + "spans": [ + { + "bbox": [ + 207, + 114, + 406, + 171 + ], + "type": "table", + "html": "
methodmatching modulesucc./prec.FPS
SiamDisstViT-384-663.5/75.025
SiamDisstViT-96-360.5/71.730
MemDisstViT-384-666.7/79.178
MemDisstViT-96-366.6/79.390
", + "image_path": "de469e7b2be04432a9667be91202ec219dc4bd50ff8ad012e09d34d25ab978c6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 156, + 171, + 456, + 182 + ], + "lines": [ + { + "bbox": [ + 156, + 171, + 456, + 182 + ], + "spans": [ + { + "bbox": [ + 156, + 171, + 456, + 182 + ], + "type": "text", + "content": "Table 3: Comparisons of our approaches with various matching modules." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 217, + 479, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 479, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 479, + 301 + ], + "type": "text", + "content": "generative 3D Point-MAE pre-training provides the best result. Note that ReCon [49] adds the contrastive learning in the multi-modality generative pretraining, which may not be suitable in the pure point cloud tracking setting. We also test two typical 2D pre-trained models, i.e., MAE [22] and DropMAE [62] (see Supp. for details). Interestingly, our SiamDisst with DropMAE initialization achieves competitive results, which indicates that the temporal matching ability of DropMAE learned in 2D videos is also helpful for the 3D tracking task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 301, + 480, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 480, + 445 + ], + "type": "text", + "content": "Lightweight or heavy matching module. We explore various matching modules used in SiamDisst and MemDisst in Table 3. Note that ViT-384-6 is built from the first 6 layers of ViT-Base [11] with 384 hidden dimensions and 6 heads. We also design a lightweight ViT-96-3 model with 4 heads for comparison. We initialize ViT-384-6 with the official Point-MAE weights and rerun Point-MAE pre-training to initialize ViT-96-35. In Table 3, SiamDisst benefits more from the heavy ViT-384-6 while MemDisst achieves favorable performance even using the lightweight ViT-96-3. This is because SiamDisst has no effective online memory mechanism and heavily relies on its matching module. For MemDisst, the online memory provides more historical cues for matching, which eases the need for a heavy online matching module. Considering the speed-accuracy trade-off, we use ViT-384-6 and ViT-96-3 for SiamDisst and MemDisst, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 466, + 303, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 466, + 303, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 466, + 303, + 479 + ], + "type": "text", + "content": "4.2 State-of-the-art Comparison" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 489, + 479, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 489, + 479, + 573 + ], + "spans": [ + { + "bbox": [ + 130, + 489, + 479, + 573 + ], + "type": "text", + "content": "We compare the proposed SiamDisst and MemDisst with state-of-the-art 3D trackers on the KITTI, nuScenes and WOD datasets. Specifically, for KITTI, comprehensive comparisons are made with two groups of trackers: 1) memoryless Siamese-based 3D trackers, SC3D [19], 3DSiamRPN [15], P2B [48], LTTR [8], MLVSNet [57], BAT [77], PTT [52], V2B [26], CMT [21], PTTR [79], and STNet [27]; and 2) contextual memory or motion prediction based 3D trackers, TAT [33], DMT [68], M2Track [78], CXTrack [69], and MBPTrack [70]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 574, + 480, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 480, + 634 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 480, + 634 + ], + "type": "text", + "content": "As shown in Table 4, our memory-less tracker SiamDisst outperforms the other Siamese 3D trackers by large margins in terms of both success and precision metrics. Specifically, the overall performance (Mean column) achieved by SiamDisst is " + }, + { + "bbox": [ + 130, + 574, + 480, + 634 + ], + "type": "inline_equation", + "content": "66.2\\% /83.9\\%" + }, + { + "bbox": [ + 130, + 574, + 480, + 634 + ], + "type": "text", + "content": ", which has significant improvements " + }, + { + "bbox": [ + 130, + 574, + 480, + 634 + ], + "type": "inline_equation", + "content": "(4.9\\% /2.8\\%)" + }, + { + "bbox": [ + 130, + 574, + 480, + 634 + ], + "type": "text", + "content": " over the baseline tracker STNet and demonstrates the effectiveness of our proposed" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Wu et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 479, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 479, + 665 + ], + "type": "text", + "content": "5 We follow the same pre-training hyper-parameters, steps and dataset (ShapeNet [6]) with Point-MAE [42] for pre-training." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 114, + 491, + 335 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 491, + 335 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 491, + 335 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 491, + 335 + ], + "type": "table", + "html": "
MethodCM MPFPSCar (6424)Pedestrian (6088)Van (1248)Cyclist (308)Mean (14068)
SC3D [19]241.3/57.918.2/37.840.4/47.041.5/70.431.2/48.5
3DSiamRPN [15]2158.2/76.235.2/56.245.7/52.936.2/49.046.7/64.9
P2B [48]4656.2/72.828.7/49.640.8/48.432.1/44.742.4/60.0
LTTR [8]2365.0/77.133.2/56.835.8/45.666.2/89.948.7/65.8
MLVSNet [57]7056.0/74.034.1/61.152.0/61.434.3/44.545.7/66.7
BAT [77]6560.5/77.742.1/70.152.4/67.033.7/45.451.2/72.8
PTT [52]4067.8/81.844.9/72.043.6/52.537.2/47.355.1/74.2
V2B [26]3770.5/81.348.3/73.550.1/58.040.8/49.758.4/75.2
CMT [21]3270.5/81.949.1/75.554.1/64.155.1/82.459.4/77.6
PTTR [79]5365.2/77.450.9/81.652.5/61.865.1/90.557.9/78.1
STNet [27]3572.1/84.049.9/77.258.0/70.673.5/93.761.3/80.1
SiamDisst2573.7/85.158.4/83.963.5/75.076.9/94.566.2/83.9
TAT [33]✓-72.2/83.357.4/84.458.9/69.274.2/93.964.7/82.8
DMT [68]✓7266.4/79.448.1/77.953.3/65.670.4/92.655.1/75.8
M2Track [78]✓5765.5/80.861.5/88.253.8/70.773.2/93.562.9/83.4
CXTrack [69]✓3469.1/81.667.0/91.560.0/71.874.2/94.367.5/85.3
MBPTrack [70]✓5673.4/84.868.6/93.961.3/72.776.7/94.370.3/87.9
MemDisst✓9074.1/85.669.1/94.166.6/79.377.2/94.771.3/88.9
", + "image_path": "8cacb846a2267500f0734957ca36788daa5e9c2f657e086d1dd9b655e0a4c567.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 132, + 336, + 480, + 390 + ], + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 390 + ], + "type": "text", + "content": "Table 4: 3D tracking results on KITTI [18] for (top) memory-less 3D trackers; (bottom) memory/motion-based 3D trackers. CM and MP represent the contextual memory and motion prediction. Trackers are evaluated by per-category Success/Precision metrics. 'Mean' indicates the overall result averaged over frames. The best result is shown in bold. For each category, the number in brackets indicates the test frame number." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 433, + 481, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 433, + 481, + 540 + ], + "spans": [ + { + "bbox": [ + 133, + 433, + 481, + 540 + ], + "type": "text", + "content": "IoU-guided matching distillation. Moreover, the proposed SiamDisst is more effective in identifying distractors during online tracking, e.g., obtaining best overall performance on the Pedestrian category. For MemDisst, it sets new state-of-the-art performance on all the categories while running at a fast speed of 90 FPS. The lightweight model used in MemDisst is well learned via the proposed learning framework, thus achieving promising tracking performance. Note that both SiamDisst and MemDisst performs well in the low data regime (e.g., the Van category only has 1994 training examples), which is mainly due to the effectiveness of transferring the matching knowledge from the 2D pre-trained tracker." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 546, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 546, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 546, + 481, + 665 + ], + "type": "text", + "content": "To test the generalization performance of our proposed trackers, we use our SiamDisst and MemDisst trained on KITTI to evaluate on the WOD and nuScenes datasets. As seen in Table 5, SiamDisst performs best on both WOD and nuScenes among all the Siamese-based 3D trackers, e.g., STNet, V2B and BAT. This is mainly because our SiamDisst learns a more effective target matching module via the 2D matching distillation, thus leading to more accurate tracking results. Compared with MBPTrack, our MemDisst obtains comparable performance on WOD and outperforms it on nuScenes. Meanwhile, MemDisst runs efficiently at the above real-time speed of 89 FPS, which is " + }, + { + "bbox": [ + 133, + 546, + 481, + 665 + ], + "type": "inline_equation", + "content": "53\\%" + }, + { + "bbox": [ + 133, + 546, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 133, + 546, + 481, + 665 + ], + "type": "inline_equation", + "content": "162\\%" + }, + { + "bbox": [ + 133, + 546, + 481, + 665 + ], + "type": "text", + "content": " faster than MBPTrack and CXTrack, respectively. This lightweight MemDisst tracker" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 260, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 260, + 91, + 447, + 102 + ], + "type": "text", + "content": "2D Matching Distillation and 3D Pre-training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 114, + 491, + 380 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 491, + 380 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 491, + 380 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 491, + 380 + ], + "type": "table", + "html": "
MethodWOD [53]nuScenes [5]FPS
Category (No. Frames)Vehicle (185731)Pedestrian (241752)Mean (427483)Car (15578)Pedestrian (8019)Truck (3710)Bicycle (501)Mean (27808)- -
SuccessP2B52.617.933.027.015.921.520.022.946
SC3D---25.014.225.717.021.82
BAT54.718.234.122.517.319.317.020.565*
V2B57.623.738.431.317.321.722.225.837
STNet59.725.540.432.219.122.321.226.935
SiamDisst61.228.542.733.419.825.017.728.125
CXTrack57.130.742.229.620.427.618.526.534
M2Track61.132.044.6-----57
MBPTrack61.933.746.033.619.823.920.028.160*
MemDisst61.933.645.934.020.028.118.128.989*
CategoryVehiclePedestrianMeanCarPedestrianTruckBicycleMeanFPS
PrecisionP2B61.730.143.829.222.016.226.425.346
SC3D---27.116.221.918.223.12
BAT62.730.344.424.124.515.818.823.065*
V2B65.937.950.135.123.416.719.129.037
STNet68.039.952.136.127.216.829.230.835
SiamDisst70.544.555.837.330.420.123.732.825
CXTrack66.149.456.733.432.920.826.831.534
M2Track69.349.758.2-----57
MBPTrack71.952.761.037.632.720.729.233.860*
MemDisst71.852.861.138.032.223.924.134.289*
", + "image_path": "5219c8ed0b4d9d6882b245966c56b7ef6ffd2a1511333e11780f06bb3f129c67.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 132, + 381, + 482, + 436 + ], + "lines": [ + { + "bbox": [ + 132, + 381, + 482, + 436 + ], + "spans": [ + { + "bbox": [ + 132, + 381, + 482, + 436 + ], + "type": "text", + "content": "Table 5: 3D tracking results on Waymo Open Dataset (WOD) [53] and nuScenes [5]. For a fair comparison, we compare our SiamDisst and MemDisst with memory-less Siamese-based trackers, and more complicated trackers with contexture memory or motion prediction, respectively. The best result in each comparison is shown in bold. * indicates the speed is re-evaluated on WOD for a fair comparison." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 468, + 482, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 482, + 492 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 482, + 492 + ], + "type": "text", + "content": "shows its potential to serve as a simple and strong baseline in the 3D tracking community." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 510, + 220, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 510, + 220, + 522 + ], + "spans": [ + { + "bbox": [ + 133, + 510, + 220, + 522 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 482, + 666 + ], + "type": "text", + "content": "This paper improves the power of 3D point cloud tracking from two aspects. First, we design a novel 3D SOT framework that could fully benefit from the 3D self-supervised pre-training, which shows that the self-supervised pre-training weights (e.g., Point-MAE [42] and Recon [49]) are effective to improve the 3D tracking performance, especially in the low data regime (e.g., KITTI-Van and KITTI-Cyclist). Second, we bridge the gap between 2D and 3D tracking domains by leveraging a target-aware projection module, which enables to use a powerful 2D pre-trained tracker to facilitate the learning of 3D matching in existing 3D SOT trackers [27, 70] via matching distillation. We show that our proposed SiamDisst and MemDisst trackers can achieve state-of-the-art performance on KITTI and Waymo Open datasets, while running at real-time speed." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 234, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 234, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 234, + 129 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 482, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 482, + 166 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 482, + 166 + ], + "type": "text", + "content": "This work was supported by a Strategic Research Grant from City University of Hong Kong (Project No. 7005665)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "spans": [ + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 211, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 211, + 481, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 211, + 481, + 243 + ], + "spans": [ + { + "bbox": [ + 138, + 211, + 481, + 243 + ], + "type": "text", + "content": "1. Anthes, C., García-Hernández, R.J., Wiedemann, M., Kranzlüller, D.: State of the art of virtual reality technology. In: 2016 IEEE aerospace conference. pp. 1-19. IEEE (2016) 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 481, + 277 + ], + "type": "text", + "content": "2. Ben-Baruch, E., Karklinsky, M., Biton, Y., Ben-Cohen, A., Lawen, H., Zamir, N.: It's all in the head: Representation knowledge distillation through classifier sharing. In: arXiv:2201.06945 (2022) 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 278, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 278, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 278, + 481, + 311 + ], + "type": "text", + "content": "3. Bertinetto, L., Valmadre, J., Henriques, J.F., Vedaldi, A., Torr, P.H.: Fully convolutional siamese networks for object tracking. In: Proceedings of the European Conference on Computer Vision. pp. 850-865. Springer (2016) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 312, + 481, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 481, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 481, + 334 + ], + "type": "text", + "content": "4. Bhat, G., Danelljan, M., Gool, L.V., Timofte, R.: Leanring discriminative model prediction for tracking. In: ICCV. pp. 6182-6191 (2019) 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 335, + 481, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 481, + 376 + ], + "type": "text", + "content": "5. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11621-11631 (2020) 2, 4, 10, 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 378, + 481, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 378, + 481, + 410 + ], + "spans": [ + { + "bbox": [ + 138, + 378, + 481, + 410 + ], + "type": "text", + "content": "6. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015) 7, 12" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 411, + 481, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 481, + 432 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 481, + 432 + ], + "type": "text", + "content": "7. Chen, X., Yan, B., Zhu, J., Wang, D., Yang, X., Lu, H.: Transformer tracking. In: CVPR. pp. 8126-8135 (2021) 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 434, + 481, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 481, + 454 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 481, + 454 + ], + "type": "text", + "content": "8. Cui, Y., Fang, Z., Shan, J., Gu, Z., Zhou, S.: 3d object tracking with transformer. arXiv preprint arXiv:2110.14921 (2021) 4, 12, 13" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 456, + 481, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 481, + 476 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 481, + 476 + ], + "type": "text", + "content": "9. Cui, Y., Jiang, C., Wang, L., Wu, G.: Mixformer: End-to-end tracking with iterative mixed attention. In: CVPR (2022) 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 478, + 481, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 478, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 478, + 481, + 499 + ], + "type": "text", + "content": "0. Danelljan, M., Bhat, G., Khan, F.S., Felsberg, M.: Eco: Efficient convolution operators for tracking. In: CVPR. pp. 21-26 (2017) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 500, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 500, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 500, + 481, + 544 + ], + "type": "text", + "content": "1. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arxiv 2020. arXiv preprint arXiv:2010.11929 (2010) 6, 12" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 544, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 481, + 576 + ], + "type": "text", + "content": "2. Duong, C.N., Luu, K., Quach, K.G., Shrinkteanet, N.L.: Shrinkteanet: Millionscale lightweight face recognition via shrinking teacher-student networks. In: arXiv:1905.10620 (2019) 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 577, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 481, + 598 + ], + "type": "text", + "content": "3. Fan, H., Lin, L., Yang, F.: Lasot: A high-quality benchmark for large-scale single object tracking. In: CVPR. pp. 5374-5383 (2019) 5, 10" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 481, + 632 + ], + "type": "text", + "content": "4. Fan, H., Ling, H.: Siamese cascaded region proposal networks for real-time visual tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7952-7961 (2019) 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "5. Fang, Z., Zhou, S., Cui, Y., Scherer, S.: 3d-siamrpn: An end-to-end learning method for real-time 3d single object tracking using raw point cloud. IEEE Sensors Journal 21(4), 4995-5011 (2020) 12, 13" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 260, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 260, + 91, + 448, + 102 + ], + "type": "text", + "content": "2D Matching Distillation and 3D Pre-training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "16. Feng, S., Liang, P., Gao, J., Cheng, E.: Multi-correlation siamese transformer network with dense connection for 3d single object tracking. IEEE Robotics and Automation Letters (2023) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 151, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 151, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 151, + 480, + 172 + ], + "type": "text", + "content": "17. Galoogahi, H., Fagg, A., Lucey, S.: Learning background-aware correlation filters for visual tracking. In: ICCV (2017) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 173, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 173, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 173, + 480, + 205 + ], + "type": "text", + "content": "18. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. The International Journal of Robotics Research 32(11), 1231-1237 (2013) 2, 4, 10, 13" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 206, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 206, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 134, + 206, + 480, + 239 + ], + "type": "text", + "content": "19. Giancola, S., Zarzar, J., Ghanem, B.: Leveraging shape completion for 3d siamese tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1359-1368 (2019) 4, 10, 12, 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 240, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 240, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 133, + 240, + 480, + 262 + ], + "type": "text", + "content": "20. Girshick, R.: Fast r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 1440-1448 (2015) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 263, + 480, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 480, + 296 + ], + "type": "text", + "content": "21. Guo, Z., Mao, Y., Zhou, W., Wang, M., Li, H.: Cmt: Context-matching-guided transformer for 3d tracking in point clouds. In: European Conference on Computer Vision. pp. 95-111. Springer (2022) 4, 12, 13" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 297, + 480, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 480, + 329 + ], + "type": "text", + "content": "22. He, K., Chen, X., Xie, S., Li, Y., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022) 2, 11, 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "type": "text", + "content": "23. Henriques, J.F., Caseiro, R., Martins, P., Batista, J.: High-speed tracking with kernelized correlation filters. In: IEEE Transactions on Pattern Analysis and Machine Intelligence. pp. 583-596 (2015) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 385 + ], + "type": "text", + "content": "24. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. In: arXiv:1503.02531 (2015) 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 480, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 419 + ], + "type": "text", + "content": "25. Huang, L., Zhao, X., Huang, K.: Got-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence (2019) 5, 10" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 420, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 480, + 453 + ], + "type": "text", + "content": "26. Hui, L., Wang, L., Cheng, M., Xie, J., Yang, J.: 3d siamese voxel-to-bev tracker for sparse point clouds. Advances in Neural Information Processing Systems 34, 28714-28727 (2021) 3, 4, 5, 6, 10, 12, 13" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 453, + 480, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 480, + 486 + ], + "type": "text", + "content": "27. Hui, L., Wang, L., Tang, L., Lan, K., Xie, J., Yang, J.: 3d siamese transformer network for single object tracking on point clouds. In: European Conference on Computer Vision. pp. 293-310. Springer (2022) 2, 3, 4, 6, 10, 12, 13, 14" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 487, + 480, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 480, + 519 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 480, + 519 + ], + "type": "text", + "content": "28. Jin, X., Peng, B., Wu, Y., Liu, Y., Liu, J.: Knowledge distillation via route constrained optimization. In: IEEE/CVF International Conference on Computer Vision (2019) 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 521, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 480, + 553 + ], + "type": "text", + "content": "29. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017) 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 554, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 480, + 587 + ], + "type": "text", + "content": "30. Kiran, B.R., Sobh, I., Talpaert, V., Mannion, P., Al Sallab, A.A., Yogamani, S., Pérez, P.: Deep reinforcement learning for autonomous driving: A survey. IEEE Transactions on Intelligent Transportation Systems (2021) 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 588, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 632 + ], + "type": "text", + "content": "31. Kristan, M., Matas, J., Leonardis, A., Vojir, T., Pflugfelder, R., Fernandez, G., Nebehay, G., Porikli, F., Cehovin, L.: A novel performance evaluation methodology for single-target trackers. IEEE Transactions on Pattern Analysis and Machine Intelligence (2016) 10" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "32. Kristan, M., Matas, J., Danelljan, M.: The first visual object tracking segmentation votes2023 challenge results. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023) 5" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "33. Lan, K., Jiang, H., Xie, J.: Temporal-aware siamese tracker: Integrate temporal context for 3d object tracking. In: Proceedings of the Asian Conference on Computer Vision. pp. 399-414 (2022) 2, 4, 12, 13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "34. Lee, S.H., Kim, D.H., Song, B.C.: Self-supervised knowledge distillation using singular value decomposition. In: In Proceedings of the European Conference on Computer Vision (ECCV) (2018) 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "text", + "content": "35. Li, B., Wu, W., Wang, Q., Zhang, F., Xing, J., Yan, J.: Siamrpn++: Evolution of siamese visual tracking with very deep networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4282-4291 (2019) 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "type": "text", + "content": "36. Li, B., Yan, J., Wu, W., Zhu, Z., Hu, X.: High performance visual tracking with siamese region proposal network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8971-8980 (2018) 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 249, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 281 + ], + "type": "text", + "content": "37. Li, J., Guo, Z., Li, H., Han, S., Baek, J.w., Yang, M., Yang, R., Suh, S.: Rethinking feature-based knowledge distillation for face recognition. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 281, + 482, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 304 + ], + "type": "text", + "content": "38. Lin, L., Fan, H., Zhang, Z., Xu, Y., Ling, H.: Swintrack: A simple and strong baseline for transformer tracking pp. 16743-16754 (2022) 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 304, + 482, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 326 + ], + "type": "text", + "content": "39. Liu, Y., Liang, Y., Wu, Q., Zhang, L., Wang, H.: A new framework for multiple deep correlation filters based object tracking. In: ICASSP (2022) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 326, + 482, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 482, + 347 + ], + "type": "text", + "content": "40. Muller, M., Bibi, A., S, G.: Trackingnet: A large-scale dataset and benchmark for object tracking in the wild. In: ECCV. pp. 300-317 (2018) 10" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 347, + 482, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 380 + ], + "type": "text", + "content": "41. Osep, A., Mehner, W., Mathias, M., Leibe, B.: Combined image-and world-space tracking in traffic scenes. In: IEEE International Conference on Robotics and Automation. pp. 1988-1995. IEEE (2017) 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "type": "text", + "content": "42. Pang, Y., Wang, W., Tay, F.E., Liu, W., Tian, Y., Yuan, L.: Masked autoencoders for point cloud self-supervised learning. In: European conference on computer vision. pp. 604-621 (2022) 3, 6, 7, 10, 11, 12, 14" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 413, + 482, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 435 + ], + "type": "text", + "content": "43. Pang, Z., Li, Z., Wang, N.: Model-free vehicle tracking and state estimation in point cloud sequences. In: IROS (2021) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "type": "text", + "content": "44. Peng, B., Jin, X., Liu, J., Li, D., Wu, Y., Liu, Y., Zhou, S., Zhang, Z.: Self-supervised knowledge distillation using singular value decomposition. In: IEEE/CVF International Conference on Computer Vision (2019) 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "type": "text", + "content": "45. Peng, B., Jin, X., Liu, J., Li, D., Wu, Y., Liu, Y., Zhou, S., Zhang, Z.: Shrinkteanet: Million-scale lightweight face recognition via shrinking teacher-student networks. In: IEEE/CVF International Conference on Computer Vision (2019) 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 501, + 482, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 534 + ], + "type": "text", + "content": "46. Qi, C., Su, H., Mo, K., Guibas, L.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: IEEE conference on computer vision and pattern recognition (2017) 10" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 534, + 482, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 566 + ], + "type": "text", + "content": "47. Qi, H., Feng, C., Cao, Z., Zhao, F., Xiao, Y.: P2b: Point-to-box network for 3d object tracking in point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6329-6338 (2020) 2, 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 567, + 482, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 600 + ], + "type": "text", + "content": "48. Qi, H., Feng, C., Cao, Z., Zhao, F., Xiao, Y.: P2b: Point-to-box network for 3d object tracking in point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6329-6338 (2020) 4, 12, 13" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 600, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 600, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 600, + 482, + 632 + ], + "type": "text", + "content": "49. Qi, Z., Dong, R., Fan, G., Ge, Z., Zhang, X., Ma, K., Yi, L.: Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining (2023) 3, 6, 7, 11, 12, 14" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "text", + "content": "50. Ran, T., Gavves, E., Smeulders, A.W.: Siamese instance search for tracking. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1420-1429 (2016) 4" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 260, + 91, + 449, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 91, + 449, + 102 + ], + "spans": [ + { + "bbox": [ + 260, + 91, + 449, + 102 + ], + "type": "text", + "content": "2D Matching Distillation and 3D Pre-training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 137 + ], + "type": "text", + "content": "51. Romero, A., Ballas, N.: Fitnets: Hints for thin deep nets. In: arXiv:1412.6550 (2014) 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 169 + ], + "type": "text", + "content": "52. Shan, J., Zhou, S., Fang, Z., Cui, Y.: Ptt: Point-track-transformer module for 3d single object tracking in point clouds. arXiv preprint arXiv:2108.06455 (2021) 12, 13" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 214 + ], + "type": "text", + "content": "53. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. p. 2446-2454 (2020) 4, 10, 14" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 215, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 215, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 215, + 480, + 235 + ], + "type": "text", + "content": "54. Tao, F., Wang, M.: Response-based distillation for incremental object detection. In: arXiv:2110.13471 (2021) 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "type": "text", + "content": "55. Tao, F., Wang, M., Yuan, H.: Overcoming catastrophic forgetting in incremental object detection via elastic response distillation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022) 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "type": "text", + "content": "56. Wang, N., Song, Y., Ma, C.: Unsupervised deep tracking. In: CVPR. pp. 3708-1317 (2019) 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 289, + 480, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 321 + ], + "type": "text", + "content": "57. Wang, Z., Xie, Q., Lai, Y.K., Wu, J., Long, K., Wang, J.: Mlvsnet: Multi-level voting siamese network for 3d visual tracking. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3101-3110 (2021) 12, 13" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 322, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 480, + 353 + ], + "type": "text", + "content": "58. Wang, Z., Yu, X., Rao, Y., Zhou, J., Lu, J.: P2p: Tuning pre-trained image models for point cloud analysis with point-to-pixel prompting. Advances in neural information processing systems 35, 14388-14402 (2022) 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "text", + "content": "59. Wu, Q., Chan, A.: Meta-graph adaptation for visual object tracking. In: ICME (2021) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 375, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 480, + 396 + ], + "type": "text", + "content": "60. Wu, Q., Yan, Y., Liang, Y., Liu, Y., Wang, H.: Dsnet: Deep and shallow feature learning for efficient visual tracking. In: ACCV. pp. 119-134 (2018) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 396, + 480, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 428 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 428 + ], + "type": "text", + "content": "61. Wu, Q., Wan, J., Chan, A.B.: Progressive unsupervised learning for visual object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2993-3002 (2021) 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 429, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 471 + ], + "type": "text", + "content": "62. Wu, Q., Yang, T., Liu, Z., Wu, B., Shan, Y., Chan, A.B.: Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14561-14571 (2023) 1, 2, 4, 5, 6, 8, 11, 12" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "text", + "content": "63. Wu, Q., Yang, T., Wu, W., Chan, A.B.: Scalable video object segmentation with simplified framework. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "text", + "content": "64. Wu, Q., Sun, C., Wang, J.: Multi-level structure-enhanced network for 3d single object tracking in sparse point clouds. IEEE Robotics and Automation Letters 8(1), 9-16 (2022) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "type": "text", + "content": "65. Wu, Y., Lim, J., Yang, M.H.: Online object tracking: A benchmark. In: CVPR. pp. 2411-2418 (2013) 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 558, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 480, + 599 + ], + "type": "text", + "content": "66. Xia, Y., Gladkova, M., Wang, R., Li, Q., Stilla, U., Henriques, J.F., Cremers, D.: Casspr: Cross attention single scan place recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8461-8472 (2023) 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 600, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 633 + ], + "type": "text", + "content": "67. Xia, Y., Shi, L., Ding, Z., Henriques, J.F., Cremers, D.: Text2loc: 3d point cloud localization from natural language. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14958-14967 (2024) 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "68. Xia, Y., Wu, Q., Li, W., Chan, A.B., Stilla, U.: A lightweight and detector-free 3d single object tracker on point clouds. IEEE Transactions on Intelligent Transportation Systems (2023) 4, 5, 12, 13" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Q. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 544 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 159 + ], + "type": "text", + "content": "69. Xu, T.X., Guo, Y.C., Lai, Y.K., Zhang, S.H.: Cxtrack: Improving 3d point cloud tracking with contextual information. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1084-1093 (2023) 4, 6, 10, 12, 13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 160, + 482, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 482, + 204 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 482, + 204 + ], + "type": "text", + "content": "70. Xu, T.X., Guo, Y.C., Lai, Y.K., Zhang, S.H.: Mbptrack: Improving 3d point cloud tracking with memory networks and box priors. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9911-9920 (October 2023) 2, 3, 4, 6, 10, 12, 13, 14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "text", + "content": "71. Yan, B., Peng, H., Fu, J., Wang, D., Lu, H.: Learning spatio-temporal transformer for visual tracking. In: ICCV. pp. 10448-10457 (2021) 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "72. Yang, T., Chan, A.B.: Learning dynamic memory networks for object tracking. In: Proceedings of the European Conference on Computer Vision. pp. 152-167 (2018) 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 292 + ], + "type": "text", + "content": "73. Ye, B., Chang, H., Ma, B., Shan, S., Chen, X.: Joint feature learning and relation modeling for tracking: A one-stream framework. In: European Conference on Computer Vision. pp. 341-357 (2022) 2, 3, 4, 5, 6, 7, 8, 10" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 324 + ], + "type": "text", + "content": "74. Yim, J., Joo, D., Bae, J., Kim, J.: A gift from knowledge distillation: Fast optimization, network minimization and transfer learning. In: IEEE conference on computer vision and pattern recognition (2017) 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "type": "text", + "content": "75. Zhang, L., Gonzalez-Garcia, A., Weijer, J.v.d., Danelljan, M., Khan, F.S.: Learning the model update for siamese trackers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4010-4019 (2019) 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "text", + "content": "76. Zhang, Z., Peng, H.: Deeper and wider siamese networks for real-time visual tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4591-4600 (2019) 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 391, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 481, + 434 + ], + "type": "text", + "content": "77. Zheng, C., Yan, X., Gao, J., Zhao, W., Zhang, W., Li, Z., Cui, S.: Box-aware feature enhancement for single object tracking on point clouds. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13199-13208 (2021) 2, 3, 4, 5, 12, 13" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 434, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 481, + 479 + ], + "type": "text", + "content": "78. Zheng, C., Yan, X., Zhang, H., Wang, B., Cheng, S., Cui, S., Li, Z.: Beyond 3d siamese tracking: A motion-centric paradigm for 3d single object tracking in point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8111-8120 (2022) 4, 12, 13" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 479, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 521 + ], + "type": "text", + "content": "79. Zhou, C., Luo, Z., Luo, Y., Liu, T., Pan, L., Cai, Z., Zhao, H., Lu, S.: Pttr: Relational 3d point cloud object tracking with transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8531-8540 (2022) 4, 12, 13" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 522, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 544 + ], + "type": "text", + "content": "80. Zhou, H., Song, L., Chen, J., Zhou, Y., Guoli: Rethinking soft labels for knowledge distillation: A bias-variance tradeoff perspective. In: arXiv:2102.00650 (2021) 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 260, + 91, + 449, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 91, + 449, + 102 + ], + "spans": [ + { + "bbox": [ + 260, + 91, + 449, + 102 + ], + "type": "text", + "content": "2D Matching Distillation and 3D Pre-training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Boosting Gaze Object Prediction via Pixel-level Supervision from Vision Foundation Model/93449abc-6022-4367-830a-8808cd9b5ba3_content_list.json b/2024/Boosting Gaze Object Prediction via Pixel-level Supervision from Vision Foundation Model/93449abc-6022-4367-830a-8808cd9b5ba3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3db2a4044d0a04f719c99101909b1ee4af6e8546 --- /dev/null +++ b/2024/Boosting Gaze Object Prediction via Pixel-level Supervision from Vision Foundation Model/93449abc-6022-4367-830a-8808cd9b5ba3_content_list.json @@ -0,0 +1,1740 @@ +[ + { + "type": "text", + "text": "Boosting Gaze Object Prediction via Pixel-level Supervision from Vision Foundation Model", + "text_level": 1, + "bbox": [ + 228, + 140, + 774, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yang Jin1, Lei Zhang2, Shi Yan2, Bin Fan3, and Binglu Wang1", + "bbox": [ + 240, + 215, + 759, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Xi'an University of Architecture and Technology, Xi'an, China {jin91999, wbl921129}@gmail.com † Corresponding author", + "bbox": [ + 285, + 243, + 715, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Automation, Northwestern Polytechnical University, Xi'an, China \nzl_hnly@163.com, yanshi@mail.nwpu.edu.cn", + "bbox": [ + 228, + 273, + 759, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ National Key Lab of General AI, School of Intelligence Science and Technology, Peking University, Beijing, China binfan@pku.edu.cn", + "bbox": [ + 228, + 300, + 771, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Gaze object prediction (GOP) aims to predict the category and location of the object that a human is looking at. Previous methods utilized box-level supervision to identify the object that a person is looking at, but struggled with semantic ambiguity, i.e., a single box may contain several items since objects are close together. The Vision foundation model (VFM) has improved in object segmentation using box prompts, which can reduce confusion by more precisely locating objects, offering advantages for fine-grained prediction of gaze objects. This paper presents a more challenging gaze object segmentation (GOS) task, which involves inferring the pixel-level mask corresponding to the object captured by human gaze behavior. In particular, we propose that the pixel-level supervision provided by VFM can be integrated into gaze object prediction to mitigate semantic ambiguity. This leads to our gaze object detection and segmentation framework that enables accurate pixel-level predictions. Different from previous methods that require additional head input or ignore head features, we propose to automatically obtain head features from scene features to ensure the model's inference efficiency and flexibility in the real world. Moreover, rather than directly fuse features to predict gaze heatmap as in existing methods, which may overlook spatial location and subtle details of the object, we develop a space-to-object gaze regression method to facilitate human-object gaze interaction. Specifically, it first constructs an initial human-object spatial connection, then refines this connection by interacting with semantically clear features in the segmentation branch, ultimately predicting a gaze heatmap for precise localization. Extensive experiments on GOO-Synth and GOO-Real datasets demonstrate the effectiveness of our method. The code will be available at https://github.com/jinyang06/SamGOP.", + "bbox": [ + 259, + 419, + 738, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Gaze object prediction $\\cdot$ Vision foundation model $\\cdot$ Object segmentation $\\cdot$ Space-to-object gaze regression", + "bbox": [ + 259, + 811, + 740, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/494d3ca1c7ce9f1f6a9dcf3a12974465e1826e94bdf96377c54067106ffe7cb6.jpg", + "image_caption": [ + "(a) Illustration of semantic ambiguity and semantic clarity", + "Fig. 1: (a) Box-level supervision often fails to localize objects in dense settings precisely and leads to semantic ambiguity problems, whereas pixel-level supervision excels by providing clear semantic distinction through pixel-by-pixel predictions. (b) Vision foundation models can produce instance masks, thereby segmentation features can be used to improve the gaze regression branch's spatial perception, and the gaze object mask can help the heatmap focus on the gaze object." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 500, + 255 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/97b8cf9b4a6ee2cbbb92b1087e24d231e9bdf2d10ffc3f2dbfb1f1020fd5c394.jpg", + "image_caption": [ + "(b) The main idea of our pipeline" + ], + "image_footnote": [], + "bbox": [ + 506, + 146, + 787, + 253 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 387, + 375, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The objects stared at by humans contain important semantic information, which can reveal a person's behavior and state of mind. Therefore, identifying the objects that people are looking at has a wide range of applications. For example, in medical diagnosis, whether a child focuses on an object can potentially reveal whether he/she has autism or visual impairment [4,26,32,34]. In assisted driving scenarios, determining what object the driver is looking at can assess their attention and provide corresponding driving guidance [5,15,23]. In general, detecting the object gazed at by humans realizes the interaction between human sight and specific objects, fostering a richer semantic understanding of the scene.", + "bbox": [ + 212, + 417, + 787, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Tomas et al. [35] first proposed the gaze object detection task and the benchmark datasets GOO-Synth and GOO-Real, aiming to predict the object boxes and categories stared at by humans. Recently, Wang et al. [40, 41] conducted an in-depth exploration of this task. Despite achieving promising results, they still suffer from the following limitations: 1) They typically rely on box-level supervision, which is prone to suffer semantic ambiguity in dense object scenes. Especially when multiple objects are densely stacked or placed adjacent to each other, the box-level supervision algorithm may struggle to accurately separate them, as illustrated in Fig. 1(a). 2) The GOO [35] dataset only contains pixel-level annotations of synthetic images, leading to difficulties in generalizing to fine-grained predictions in real scenes. 3) They usually require additional head information as input, affecting the efficiency and flexibility of the model. While global modeling using the Transformer [38, 39] can alleviate this problem, it lacks special attention to head features, which are crucial for gaze direction perception. 4) They directly fuse features to regress the gaze heatmap, lacking a full understanding of the spatial location and subtle detailed information of objects.", + "bbox": [ + 212, + 553, + 787, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, vision foundational models (VFM) have gained significant attention due to their powerful generalization capabilities, which can achieve accurate pixel-level segmentation using a variety of prompts, i.e., boxes, points, or masks.", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Jin et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In real scenes with small and densely packed objects, VFM can generate accurate boundaries and pixel-level masks (See Fig 3) by utilizing boxes as prompt, which reduces semantic ambiguity and thus opens opportunities for gaze object segmentation. Motivated by this, we propose a more challenging task, gaze object segmentation (GOS). This task involves inferring the pixel-level mask of objects captured by human gaze behavior, helping to further enhance the understanding of human gaze behavior. Furthermore, we establish a GOS framework specifically for this task, which integrates the pixel-level supervision provided by VFM into gaze object prediction to alleviate the aforementioned limitations of previous methods [40, 41].", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the proposed model, we introduce the unified detection and segmentation Transformer, i.e. MaskDINO [20], to simultaneously obtain candidate gaze object boxes, segmentation masks, and head boxes of the subject, which are indispensable for enabling the model to infer in the real world and achieve accurate detection. Instead of requiring additional head input [40, 41] or ignoring head features by global modeling [38, 39], we propose a RoI reconstruction module to automatically obtain head region features based on holistic features and head box provided by the unified Transformer. This helps capture gaze direction information during gaze regression to improve model efficiency and flexibility. Subsequently, a space-to-object gaze regression strategy is developed to enhance the perception of the spatial location of objects and the understanding of detailed information. Specifically, inspired by the ability of humans to infer others' gaze objects, a dual attention fusion module is presented to establish an initial human-object gaze spatial connection. Then, we propose an object feature interaction module to refine the gaze spatial connection through semantically clear feature interaction between gaze regression features and object mask features, thus strengthening the modeling of the human-object gaze relationship. In the end, an energy aggregation loss is used to further guide the regression of the gaze heatmap. By focusing on pixel-level details and progressively improving the gaze representation, it further reduces semantic ambiguity, leading to the generation of more precise heatmaps for gazed objects.", + "bbox": [ + 212, + 297, + 787, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our main contributions could be summarized as follows:", + "bbox": [ + 238, + 616, + 643, + 628 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the vision foundation model into gaze object prediction to provide pixel-level supervision and propose a more challenging gaze object segmentation task, which focuses on identifying the pixel-level masks of objects captured by human gaze behavior.", + "- We present an all-in-one end-to-end framework for gaze object detection and segmentation without relying on additional head-related inputs, capable of simultaneously addressing gaze estimation, object detection, and segmentation. To the best of our knowledge, our work is the first to tackle gaze object prediction from the image segmentation perspective.", + "- We propose a space-to-object gaze regression approach that enhances the interaction between object and gaze branches, incrementally improving the modeling of the human-object gaze relationship and alleviating semantic ambiguity." + ], + "bbox": [ + 225, + 642, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Boosting Gaze Object Prediction via Pixel-level Supervision", + "bbox": [ + 330, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 143, + 397, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vision Foundation Model. The vision foundation models (VFMs) [19, 24, 48] are typically built on large-scale datasets using self-supervised or semi-supervised methods, which have strong generalizability and zero-shot transfer capabilities. Segment Anything Model (SAM) [19] received attention as a foundation model for image segmentation. Its portability enables it to be flexibly applied to various downstream tasks. Then, the medical foundation model MedSAM [24] extends the foundation model to medical diagnosis. In this article, we choose SAM [19] to obtain high-quality masks to boost gaze estimation and gaze object prediction.", + "bbox": [ + 212, + 176, + 785, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Gaze Object Detection. The gaze object detection is first proposed in [35], which aims to predict the box and categories of the object stared at by humans in an RGB image. Tomas et al. [35] released a benchmark dataset for GOP. Wang et al. [41] first proposed a GOP method named GaTector. However, GaTector requires additional head-related prior information, which hinders its application in the real world. Recently, the Transformer-based method [40] achieves comparable performance. Although previous methods are effective to some extent, they can only provide instance-level box outputs, leading to potential ambiguities or misjudgments in scenarios with small and densely packed objects.", + "bbox": [ + 212, + 297, + 785, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Gaze Target Detection. Early research on gaze focused on gaze direction estimation [2, 5-7, 16, 18, 27, 28, 31, 42, 44, 46, 47, 49], which only predict the gaze vector and ignores the impact of the surrounding environment on the sight, resulting in limited application scenarios. Therefore, to capture higher-level gaze-related semantic information, researchers propose gaze target detection [1, 8-10, 12-14, 17, 21, 22, 25, 30, 34, 36, 37, 43], which involves to infer the specific gaze point. Recasens et al. [29] introduced the GazeFollow dataset and first proposed gaze target detection. Recently, Tu et al. [38,39] achieved end-to-end gaze target detection using Transformer, with a relatively low detection efficiency. Despite advancements in gaze target estimation, previous methods struggle with gaze regression due to spatial perception ambiguity. In contrast, our space-to-object regression strategy helps the model gradually converge to a certain gaze point, simplifying learning and enhancing detailed interaction with the object.", + "bbox": [ + 212, + 434, + 785, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 652, + 330, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an RGB image, our model aims to precisely predict the pixel-level mask of the object humans are looking at. We first present the overall structure of our method, then introduce the VFM-based mask supervision generation method, and finally describe each module of the proposed model in detail.", + "bbox": [ + 212, + 685, + 785, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 768, + 339, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2, our method consists of four components: a feature extractor for specific feature extraction, a detection and segmentation module predicting candidate gaze objects boxes, masks, and human head location, a gaze regressor", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Jin et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8f1529fbd145da8de17a7eb09b2daee56dd8fcd3712c58baa4cc8dd28a46ada4.jpg", + "image_caption": [ + "Fig. 2: Overview of the proposed model. (a) The feature extraction module extracts features for detection and regression branches. (b) The detection and segmentation branch identifies object and human head positions. (c) After obtaining head features, the gaze regression branch progressively refines its output: 1) employing a dual attention fusion module for initial human-object correlations; 2) leveraging a feature interaction module to incorporate semantically clear object-aware insights from the segmentation branch; 3) ultimately predicting the gaze heatmap. (d) Supervision signals are applied to both branches only during training." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 782, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "producing gaze object heatmap with a space-to-object strategy, and a supervision information generation module creating pixel-level masks and ground truth heatmap during training. At training, the overall loss function is defined as:", + "bbox": [ + 212, + 441, + 782, + 488 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {d e t}} + \\alpha \\mathcal {L} _ {\\mathrm {d i r}} + \\beta \\mathcal {L} _ {\\mathrm {g a z e}} + \\gamma \\mathcal {L} _ {\\mathrm {e n g}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 496, + 785, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ , $\\beta$ , and $\\gamma$ are the weights of the gaze direction loss, gaze heatmap loss, and energy aggregation loss respectively. Before model training, we choose SAM to generate pixel-level masks for the objects. See Sec. 3.2 for details.", + "bbox": [ + 212, + 521, + 782, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature extraction. As shown in 2(a), given an RGB image $\\mathcal{I} \\in \\mathbb{R}^{3 \\times H \\times W}$ , we first use the backbone $\\mathcal{B}$ to extract multi-scale features $\\{\\mathbf{f}_1, \\mathbf{f}_2, \\mathbf{f}_3, \\mathbf{f}_4\\}$ . In previous work [41], pixel shuffle [33] significantly improved small object detection by enlarging feature maps for detailed feature capture. Therefore, we integrate it into dense prediction tasks, i.e., image segmentation. In our model, the multi-scale features $\\{\\mathbf{f}_1, \\mathbf{f}_2, \\mathbf{f}_3, \\mathbf{f}_4\\}$ are fed into the pixel shuffle module to generate object-specific features $\\mathbf{f}_i^{obj} = \\phi^{obj}(\\mathbf{f}_i, \\eta)$ , where $\\mathbf{f}_i$ denotes the $i$ -th feature layer, $\\eta$ denotes the scale factor, $\\phi^{obj}(\\cdot)$ denotes the pixel shuffle module, which consists of a pixel shuffle operation and a convolution layer that transforms the feature channels to be consistent with the original features of the unified Transformer [20]. Finally, the scene saliency features $\\mathbf{f}_{\\mathrm{scene}} \\in \\mathbb{R}^{1024 \\times 7 \\times 7}$ are extracted by a scene residual block using high dimensional features $\\mathbf{f}_4 \\in \\mathbb{R}^{2048 \\times 7 \\times 7}$ .", + "bbox": [ + 212, + 566, + 784, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unified detection and segmentation Transformer. In this paper, we employ a unified Transformer for candidate gaze object detection, segmentation, and human head positioning, integrating MaskDINO [20] for candidate gaze object detection, segmentation, and the generation of pixel-level mask feature embeddings and a head decoder for head box regression, which facilitate accurate head-related feature extraction in head feature reconstruction module and", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Boosting Gaze Object Prediction via Pixel-level Supervision", + "bbox": [ + 330, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1ccb143e786d159e727b559a8131322e7e1b5183edccaf11d6264df7973d0178.jpg", + "image_caption": [ + "Fig. 3: Illustration of instance mask generated by VFM." + ], + "image_footnote": [], + "bbox": [ + 218, + 146, + 782, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "accurate prediction in real-world settings. The training loss is defined as:", + "bbox": [ + 215, + 258, + 740, + 273 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {d e t}} = \\mathcal {L} _ {\\mathrm {d e t}} ^ {\\mathrm {o b j}} + \\mathcal {L} _ {\\mathrm {m a s k}} ^ {\\mathrm {o b j}} + \\mathcal {L} _ {\\mathrm {d e t}} ^ {\\mathrm {h e a d}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 282, + 785, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{\\mathrm{det}}^{\\mathrm{obj}}$ and $\\mathcal{L}_{\\mathrm{mask}}^{\\mathrm{obj}}$ is the object detection and segmentation loss respectively. $\\mathcal{L}_{\\mathrm{det}}^{\\mathrm{head}}$ is the head decoder loss, which is consistent with the $\\mathcal{L}_{\\mathrm{det}}^{\\mathrm{obj}}$ . See Sec. 3.3 for details. Space-to-object gaze regression. We employ a space-to-object regression strategy for gaze heatmap prediction. The module begins by representing human gaze behavior with a gaze vector. Subsequently, a dual attention fusion module establishes an initial human-object spatial connection. Finally, pixel-level object location knowledge from the mask branch is introduced to achieve semantically clear feature interaction, refining the gaze heatmap regression while alleviating semantic ambiguity. Optimization involves $\\mathcal{L}_{\\mathrm{dir}}$ and $\\mathcal{L}_{\\mathrm{gaze}}$ for gaze direction and gaze heatmap, respectively. The energy aggregation loss $\\mathcal{L}_{\\mathrm{eng}}$ guides the gaze heatmap to focus on the gaze object mask. See Sec. 3.4 for details.", + "bbox": [ + 214, + 310, + 787, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Instance mask generation via VFM.", + "text_level": 1, + "bbox": [ + 215, + 502, + 555, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given an RGB image $\\mathcal{I}$ , we choose the vision foundation model SAM [19] to obtain instance masks for model training. Specifically, image $\\mathcal{I}$ and prompt $\\mathcal{P}_{box}$ are first fed into the image encoder $\\varPsi_{Image}$ and prompt encoder $\\varPsi_{prompt}$ , respectively. Then, the mask $\\mathcal{P}_{mask}$ and confidence $\\mathcal{S}$ are output by the mask decoder $\\varPhi_{mask}$ , which can be formulated as:", + "bbox": [ + 214, + 525, + 787, + 602 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} _ {\\text {m a s k}}, \\mathcal {S} = \\Phi_ {\\text {m a s k}} (\\Psi_ {\\text {I m a g e}} (\\mathcal {I}), \\Psi_ {\\text {p r o m p t}} (\\mathcal {P} _ {\\text {b o x}})) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 612, + 785, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Instance segmentation by using boxes produces accurate masks for most objects, but accurate segmentation of objects with unclear texture or edge features remains challenging. Therefore, we segment again using $\\mathcal{P}_{mask}$ as a dense prompt to produce final mask $\\mathcal{M}_{\\mathrm{mask}}$ , which can be formulated as:", + "bbox": [ + 214, + 637, + 787, + 699 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {\\text {m a s k}}, \\mathcal {S} = \\Phi_ {\\text {m a s k}} (\\Psi_ {\\text {I m a g e}} (\\mathcal {I}), \\Psi_ {\\text {p r o m p t}} (\\mathcal {P} _ {\\text {m a s k}})) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 709, + 785, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, $\\mathcal{M}_{mask}$ is used to train the gaze object segmentation model. The generated masks are shown in Fig. 3.", + "bbox": [ + 214, + 734, + 785, + 765 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Unified detection and segmentation Transformer", + "text_level": 1, + "bbox": [ + 215, + 785, + 663, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The unified detection and segmentation Transformer aims to detect candidate gaze objects and locate human heads while extracting high-level pixel mask", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Jin et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "features, which comprises a detection and segmentation Transformer, a head decoder, and a RoI feature reconstruction module. As shown in Fig. 2(b), the object-specific features $\\{\\mathbf{f}_1^{obj},\\mathbf{f}_2^{obj},\\mathbf{f}_3^{obj},\\mathbf{f}_4^{obj}\\}$ are fed it to detect object boxes, categories, masks, and human head boxes.", + "bbox": [ + 212, + 146, + 782, + 207 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Object detection and segmentation The global attention of Transformer ensures powerful long-distance modeling and global perception, achieving excellent detection performance through layer-by-layer self-attention refinement. We use MaskDINO [20] as a foundational detection and segmentation structure due to its robust representation capabilities and fast convergence speed.", + "bbox": [ + 212, + 209, + 784, + 284 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Head decoder Our model is designed for real-world gaze object detection and segmentation without relying on head priors. We introduce a head decoder based on the unified Transformer, using a 3-layer self-attention decoder to generate proposals for head-related feature reconstruction. During training, we filter head queries with IoU $>0.5$ as positives. During inference, boxes are selected based on confidence. Similar to other DETR-like detectors [45, 50], the head decoder uses cross-entropy, $L_{1}$ , and GIoU loss for classification and box regression.", + "bbox": [ + 212, + 285, + 784, + 390 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Head feature reconstruction Previous methods typically extract head features from manually cropped head images, limiting model inference efficiency and flexibility in the real world. To this end, we design a RoI feature reconstruction module to reconstruct head-related features based on holistic scene features and the head box from the unified Transformer. In this module, we first use the head bounding box $R = (x_{1},y_{1},x_{2},y_{2})$ , which represents the upper left and lower right corner coordinates, as the region proposal to crop the regions of interest (ROIs). Then, those candidate ROIs are refined by ROAlign [11], which maps the head region to the corresponding pixels in the holistic image features $\\mathbf{f}_4 \\in \\mathbb{R}^{2048 \\times 7 \\times 7}$ and produces a fixed-size feature map $\\mathbf{f}_{\\mathrm{head}} \\in \\mathbb{R}^{2048 \\times 7 \\times 7}$ . For head location map generation, following previous works [9,10,29,41] to generate a binary image, setting the head area to 1 and the rest to 0. Additionally, a specific residual block extracts gaze saliency features $\\mathbf{f}_{\\mathrm{gaze}} \\in \\mathbb{R}^{1024 \\times 7 \\times 7}$ , further enhancing the model's attention to the gaze direction.", + "bbox": [ + 212, + 391, + 785, + 602 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Space-to-object gaze regression", + "text_level": 1, + "bbox": [ + 214, + 628, + 519, + 645 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We propose a space-to-object gaze regression approach to gradually refine the human-object gaze relationship. Fig. 2(c) shows our gaze regression branch with a dual attention fusion module establishing an initial human-object spatial connection and a feature interaction module strengthening this connection using semantically clear feature interaction.", + "bbox": [ + 212, + 657, + 782, + 732 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Spatial perception for gaze objects. When humans infer what another person is staring at, they usually first observe the gaze direction of the others and search for the gazed object along this direction. Inspired by this, we propose the dual attention fusion module for the spatial perception of gaze objects, establishing an initial human-object spatial connection.", + "bbox": [ + 212, + 733, + 782, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Fig. 4(a), we first use global average pooling to obtain the global features of $\\mathbf{f}_{\\mathrm{head}}$ , followed by fully connected layers to transform the features into", + "bbox": [ + 214, + 809, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Boosting Gaze Object Prediction via Pixel-level Supervision", + "bbox": [ + 330, + 114, + 730, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aee492adf4c07b0548b42361e9eb28b1f882d8fd7d163065ef5b5ad4e38dcb3a.jpg", + "image_caption": [ + "Fig. 4: Illustration of spatial perception for gaze objects." + ], + "image_footnote": [], + "bbox": [ + 256, + 142, + 486, + 243 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cdd85fdf43e9d108402b9f524e92bb64f9a24f8526c299371dbb66ebc2b77697.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 145, + 743, + 243 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a 2D gaze vector $\\mathbf{v}_g$ . To achieve spatial perception, we project it into a gaze cone map $\\mathbf{M}_{\\mathrm{cone}} \\in \\mathbb{R}^{1 \\times 7 \\times 7}$ , activating the area around the gaze direction:", + "bbox": [ + 214, + 295, + 785, + 325 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {\\mathrm {i}, \\mathrm {j}} ^ {\\text {c o n e}} = M a x (0, \\cos (\\mathbf {v} _ {\\mathrm {g}}, \\mathbf {v} _ {\\mathrm {i}, \\mathrm {j}})) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 337, + 785, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{v}_{\\mathrm{g}}$ and $\\mathbf{v}_{\\mathrm{i,j}}$ respectively denotes the predicted gaze vector and the vector composed of the eye point and each point on the gaze field. During training, $\\mathbf{v}_{\\mathrm{g}}$ is optimized by $\\mathcal{L}_{\\mathrm{dir}}$ to obtain a more accurate gaze direction and activation map. $\\mathcal{L}_{\\mathrm{dir}}$ is defined as follows:", + "bbox": [ + 214, + 364, + 787, + 425 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {d i r}} = 1 - \\sum_ {i = 1} ^ {n} \\left(\\mathbf {v} _ {\\mathrm {g}} ^ {i} \\cdot \\mathbf {v} _ {\\mathrm {t}} ^ {i}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 435, + 785, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where vector $\\mathbf{v}_{\\mathrm{g}}$ and $\\mathbf{v}_{\\mathrm{t}}$ denote the predicted gaze direction and ground truth, respectively. During backpropagation, $\\mathcal{L}_{\\mathrm{dir}}$ adjusts parameters of head feature $\\mathbf{f}_{\\mathrm{head}}$ , providing more direction information for gaze regression. To boost scene attention, we adopt the method from GaTector [41] to generate an original attention map $\\mathbf{M}_{\\mathrm{ori}} \\in \\mathbb{R}^{1 \\times 7 \\times 7}$ , activating scene salience area based on head properties. By multiplying $\\mathbf{M}_{\\mathrm{cone}}$ with $\\mathbf{M}_{\\mathrm{ori}}$ via element-wise dot product, we generate the final spatial perception map $\\mathbf{M}$ :", + "bbox": [ + 214, + 484, + 787, + 590 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} = \\mathbf {M} _ {\\text {o r i}} \\odot \\mathbf {M} _ {\\text {c o n e}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 428, + 601, + 785, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This compels the model to search spaces more likely gazed at by humans along gaze direction, establishing an initial human-object gaze spatial relationship.", + "bbox": [ + 212, + 628, + 782, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Semantically clear feature interaction. Previous methods [40, 41] usually directly fuse features to regress the gaze heatmap based on spatial saliency, lacking interaction with object details and facing challenges in accurately predicting heatmap for the gazed object. In our method, we introduce a feature interaction module, which leverages semantically clear mask knowledge from VFM to achieve feature interaction, refining human-object gaze relationship modeling.", + "bbox": [ + 212, + 657, + 784, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig. 2(c), we initially extract detailed mask features from the mask branch. Subsequently, the fused feature $\\mathbf{f}_{\\mathrm{fuse}}$ is utilized for pixel-level attention perception. Given that candidate gaze objects can be located anywhere in the image, the Transformer's global modeling excels in accomplishing this task. Therefore, utilize a transformer layer, akin to DETR [3], to capture the global relationship between human gaze behavior and object masks. For fine-grained", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Jin et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "object mask features, a three-layer MLP is applied to process the decoder query $\\mathcal{Q}_{\\mathrm{obj}}$ of the unified Transformer, generating mask embeddings:", + "bbox": [ + 215, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Q} _ {\\text {m a s k}} = \\mathbf {M L P} \\left(\\mathcal {Q} _ {\\text {o b j}}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 185, + 785, + 200 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where the $\\mathcal{Q}_{\\mathrm{mask}}$ denote the generated mask embedding. We first apply a layer of self-attention for $\\mathbf{f}_{\\mathrm{fuse}}$ :", + "bbox": [ + 212, + 207, + 782, + 238 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {\\mathrm {e}} = \\mathbf {F F N} (\\mathbf {S e l f A t t n} (\\mathbf {f} _ {\\mathrm {f u s e}})) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 246, + 785, + 262 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbf{f}_{\\mathrm{fuse}}$ , SelfAttn( $\\cdot$ ), and $\\mathbf{f}_{\\mathrm{e}}$ denote the fused feature $\\mathbf{f}_{\\mathrm{fuse}}$ , self-attention operation, and encoded $\\mathbf{f}_{\\mathrm{fuse}}$ , respectively. Then, the $\\mathcal{Q}_{\\mathrm{mask}}$ is perceived by the $\\mathbf{f}_{\\mathrm{e}}$ as key-value pairs, i.e., $\\{\\mathbf{K}_{\\mathrm{mask}}, \\mathbf{V}_{\\mathrm{mask}}\\}$ :", + "bbox": [ + 212, + 268, + 782, + 315 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {\\text {r e g}} = \\mathbf {F F N} (\\mathbf {C r o s s A t t n} (\\mathbf {f} _ {\\mathrm {e}}, \\{\\mathbf {K} _ {\\text {m a s k}}, \\mathbf {V} _ {\\text {m a s k}} \\})) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 321, + 785, + 338 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbf{f}_{\\mathrm{reg}}$ and CrossAttn $(q, kv)$ denote the regression feature and cross-attention operation, respectively. This operation refines the perception of the object by providing a pixel-level understanding of the gaze object in each token of the regression feature. It effectively alleviates semantic ambiguity and strengthens the ability of the model to represent the human-object gaze relationship.", + "bbox": [ + 212, + 344, + 782, + 419 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ultimate gaze regression. Finally, following previous work [41] to generate the ground truth gaze heatmap and use $L_{2}$ loss to supervise gaze heatmap regression. $\\mathcal{L}_{\\mathrm{eng}}$ proposed by [40] guides gaze heatmap energy aggregation towards the gaze object. We use the gaze object mask for more accurate error measurements. Space-to-object gaze regression approach gradually refines the human-object gaze behavior representation, regressing a more accurate gaze heatmap.", + "bbox": [ + 212, + 419, + 782, + 510 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 530, + 375, + 547 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experiments Settings", + "text_level": 1, + "bbox": [ + 215, + 558, + 437, + 573 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets and implementation details are given in supplementary material. Metrics: Following previous work [40,41], we use AUC (Area Under the Curve), L2 distance, and Angle error for gaze estimation evaluation, and mSoC (min Single over Closure)4 under different thresholds for gaze object prediction evaluation. Following MaskDINO [20], AP (Average Precision) is used for instance segmentation and object detection evaluation. To evaluate the gaze object segmentation, we use masks to replace boxes when calculating mSoC (details refer to the supplementary material). For a fair comparison, we reported results for both real-world and non-real-world settings. The non-real-world settings involve inputting additional head images and head location maps while real-world settings only use scene images. Furthermore, to compare with GTR [39], we also adopt the evaluation method proposed by [39]5. This method computes gaze estimation metrics only for results with L2 distance $< 0.15$ and head box IoU $> 0.5$ , which is an incomplete evaluation for AUC, L2 distance, and Angle error.", + "bbox": [ + 212, + 579, + 782, + 790 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Boosting Gaze Object Prediction via Pixel-level Supervision", + "bbox": [ + 328, + 114, + 730, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "4 mSoC metric proposed by Wang et al. URL: https://arxiv.org/abs/2112.03549.", + "bbox": [ + 217, + 796, + 784, + 811 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "$^{5}$ Evaluation method proposed by Tu et al., URL: https://ieeexplore.ieee.org/document/10262016.", + "bbox": [ + 222, + 811, + 784, + 838 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/6939b4e3815121082d710f7ae7faf4f9f8226fc9995f3c520679c1963ba2e42f.jpg", + "table_caption": [ + "Table 1: Gaze object detection results on GOO-Synth and GOO-Real datasets. * indicates using additional head detector to achieve real-world settings." + ], + "table_footnote": [], + "table_body": "
MethodGOO-SynthGOO-Real No PretrainGOO-Real PretrainParamsFPS
mSoC507595mSoC507595mSoC507595
Non-RealGaTector [41]67.9498.1486.250.1262.4095.1073.500.2071.2097.5088.802.7060.78M14.1
TransGOP [40]92.8099.0098.5051.9082.6097.8089.406.5089.0098.9097.5033.2094.03M13.7
Ours91.1993.9793.4673.2781.7598.7896.437.1381.8998.9996.637.2880.99M14.5
RealGaTector [41]*56.1988.0866.18-40.8782.0533.540.3066.4096.2280.490.4075.73M9.3
TransGOP [40]*91.6099.0098.1048.7079.7097.8091.4011.2083.2098.8095.3014.40106.37M11.6
Ours90.7793.9293.4569.5881.3598.9596.022.6181.6099.0396.546.0384.26M14.1
", + "bbox": [ + 230, + 184, + 769, + 297 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/7cbb47521c2f3b69e511518bd9ff3188aa0095d05b7730f6960ad3e64aa1897c.jpg", + "table_caption": [ + "Table 2: Ours gaze object segmentation results on GOO-Real and GOO-Synth." + ], + "table_footnote": [], + "table_body": "
SettingsGOO-SynthGOO-Real No PretrainGOO-Real PretrainParams|FPS
mSoC507595mSoC507595mSoC507595
Non-Real82.6893.7091.6826.8386.3998.9597.4420.0789.3998.8098.0233.0380.99M
Real82.7393.6791.7926.8585.9498.9997.2916.7086.9698.9497.7522.2184.26M
", + "bbox": [ + 232, + 338, + 769, + 402 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/6793ddd6a6bbc7d9cf83f6932f142bd4de3c417a0116edb813388a1f0217cd64.jpg", + "image_caption": [ + "Fig. 5: Comparison of TransGOP [40] and our method inference in the real world." + ], + "image_footnote": [], + "bbox": [ + 246, + 416, + 514, + 474 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/639656f757280972a12e0fd517dee9c05df1976bbc2c10c95b56bcb3c05cbe5e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 417, + 756, + 474 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Comparison with SOTA", + "text_level": 1, + "bbox": [ + 215, + 529, + 460, + 545 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Gaze object detection and segmentation. Table 1 shows the performance of gaze object detection on GOO-Synth and GOO-Real. In real-world settings, previous methods [40, 41] significantly increase parameters and slow down the inference speed due to the additional head detector (As shown in Fig. 5 (a)) and they can only achieve box-level prediction. In contrast, although our method is slightly inferior to TransGOP [40] in detection performance, our model runs significantly more efficiently because it integrates the acquisition of head features into the overall framework (As shown in Fig. 5 (b)) and can output bounding boxes and segmentation masks simultaneously. The above results also show that our space-to-object gaze regression method can effectively model the human-object gaze relationship. Table 2 shows our gaze object segmentation performance in real-world and non-real-world scenarios. On GOO-Synth, our model achieves $82.68\\%$ and $82.73\\%$ mSoC in real-world and non-real-world settings. On GOO-Real, without pre-training, our model performs well in both scenarios $(86.39\\% \\mathrm{mSoC}$ and $85.94\\%$ mSoC). After pre-training, it reaches $89.39\\%$ mSoC and $86.96\\%$ mSoC. The above experimental results show that the pixel-level mask knowledge provided by VFMs can effectively alleviate semantic ambiguity while achieving pixel-level prediction. Even under the multi-task learning framework, we still achieve an FPS of 14.1. This is mainly attributed to the automatic", + "bbox": [ + 212, + 551, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Jin et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/dcce279da873f4d0f497e8d0a51ffbcd8506684679294d1c2ec6bbad076f5654.jpg", + "table_caption": [ + "Table 3: Instance segmentation results on GOO-Synth and GOO-Real datasets" + ], + "table_footnote": [], + "table_body": "
MethodGOO-Synth DatasetGOO-Real DatasetParams
AP507595AP507595
Non-RealOurs79.3393.5191.0911.7282.9198.8595.0911.2280.99M
RealMaskDINO [20]70.8595.5483.311.8177.0598.3890.454.2652.03M
Ours78.5696.0090.568.4382.8699.0295.5710.0984.26M
", + "bbox": [ + 259, + 171, + 743, + 262 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/e2a0309ef89fe08b463bb6b863e67a860be4c5015c5f48dea4e3b6d451e90e34.jpg", + "table_caption": [ + "Table 4: Object detection results on GOO-Synth and GOO-Real datasets. * indicates using additional head detector to achieve real-world settings." + ], + "table_footnote": [], + "table_body": "
MethodGOO-Synth DatasetGOO-Real DatasetParams
AP507595AP507595
Non-RealGaTector [41]56.8095.3062.500.1052.2591.9255.340.1060.78M
TransGOP [40]87.6099.0097.3025.5077.2097.8089.406.5094.03M
Ours88.3393.8893.0850.7077.3298.6593.911.9380.99M
RealGaTector [41]*43.5086.2337.52-29.2174.0216.13-75.73M
TransGOP [40]*86.0098.8096.2023.8073.3096.7086.004.10106.37M
MaskDINO [20]72.9396.2688.132.0474.4498.7291.190.1952.03M
Ours87.6093.7273.0745.5375.2098.9493.200.7684.26M
", + "bbox": [ + 259, + 316, + 743, + 459 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "incorporation of head features, avoiding the need for an additional head detector and thus enhancing the efficiency of the model.", + "bbox": [ + 212, + 489, + 784, + 518 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Object detection and segmentation. Tables 3 and 4 show our model's instance segmentation and object detection performance. For a fair comparison, we adjusted the input size of MaskDINO [20] to match ours $(224 \\times 224)$ . In Table 3, our model outperforms MaskDINO by 8.66 AP on GOO-Real (70.85 vs. 79.51) and by 5.81 AP on GOO-Synth (77.05 vs. 82.86) in real-world scenarios, with similar advantages in non-real-world settings of our model. These results show the challenge faced by the original MaskDINO [20] in capturing detailed features in scenes with dense objects and the effectiveness of the pixel shuffle of our model in capturing detailed features. Table 4 further demonstrates the superiority of our model in object detection over previous methods and MaskDINO [20].", + "bbox": [ + 212, + 521, + 787, + 672 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Gaze estimation. Tables 5 and 6 show the gaze estimation performance on GOO-Synth and GOO-Real, respectively. In Table 5, our method outperforms GTR [39] and TransGOP [40] by $23.9\\mathrm{mAP}$ and $10.6\\mathrm{mAP}$ , respectively. Even directly evaluating all prediction results in the real world, our model reduces L2 distance and angle errors compared to TransGOP [40] by 0.015 and 1.8, respectively. Applying non-real-world methods to the real world significantly degrades performance, highlighting our real-world model is more efficient. Table 6 verifies our method on GOO-Real, achieving state-of-the-art results without pretraining. Even with direct inference in a real-world setting, our method outperforms previous non-real-world methods. After pre-training, our model exhibits more pronounced performance advantages. These results show the effectiveness", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Boosting Gaze Object Prediction via Pixel-level Supervision", + "bbox": [ + 330, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/054fc8fb59f66f9311576d9087d90375446ed37787bc8b8af8a7d7abdc6804ef.jpg", + "table_caption": [ + "Table 5: Gaze estimation results on GOO-Synth dataset. $\\dagger$ indicates the evaluation method proposed by [39]." + ], + "table_footnote": [], + "table_body": "
MethodAUCDist.↓Ang.↓mAP
Non-RealGazeFollow [29]0.9290.16233.0-
Lian [22]0.9540.10719.7-
VideoAttention [9]0.9520.07515.1-
GaTector [41]0.9570.07314.9-
GTR [39]0.9600.07114.5-
TransGOP [40]0.9630.07913.3-
Ours0.9470.07214.8-
RealGazeFollow [29]†0.8320.31742.60.468
Lian [22]†0.9030.15328.10.454
VideoAttention [9]†0.9120.14324.50.489
GaTector [41]†0.9180.13924.50.510
GTR [39]†0.9620.06814.20.597
TransGOP [40]†0.9770.07011.50.730
TransGOP [40]0.9450.10620.7-
Ours†0.9780.05710.20.836
Ours0.9380.09118.9-
", + "bbox": [ + 220, + 198, + 488, + 414 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/c9a9871646f06f4ccfe74e5e1696bf7f0f4953b040b76652a08e2f1ab96c984d.jpg", + "table_caption": [ + "Table 6: Gaze estimation results on GOO-Real dataset. $\\dagger$ indicates the evaluation method proposed by [39]." + ], + "table_footnote": [], + "table_body": "
MethodAUCDist.↓Ang.↓mAP
No PretrainNon-RealGazeFollow [29]0.8500.22044.4-
Lian [22]0.8400.32143.5-
VideoAttention [9]0.7960.25251.4-
GaTector [41]0.9270.19639.5-
Tonini [36]0.9180.164--
TransGOP [40]0.9470.09716.7-
Ours0.9430.07812.9-
RealOurs†0.9710.0599.80.810
Ours0.9440.08814.7-
PretrainNon-RealGazeFollow [29]0.9030.19539.8-
Lian [22]0.8900.16832.6-
VideoAttention [9]0.8890.15029.1-
GaTector [41]0.9400.08714.8-
TransGOP [40]0.9570.08114.7-
Ours0.9630.07312.4-
RealOurs†0.9780.0518.80.824
Ours0.9490.08213.8-
", + "bbox": [ + 504, + 196, + 779, + 414 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/d9267b53a3af2e8e67046d817965d7bf97846dcd5b03f065675fb0bfbbdb1f2a.jpg", + "table_caption": [ + "Table 7: Ablation study about each component on GOO-Real" + ], + "table_footnote": [], + "table_body": "
PS RoI-Rec. DAF FIM EALGaze Object SegmentationInstance SegmentationGaze Estimation
mSoCmSoC50mSoC75APAP50AP75AUCDist.↓Ang.↓
79.0098.4892.5074.2498.0087.620.8540.30454.2
✓84.5098.7596.5680.5398.6893.670.8390.33459.8
✓✓84.6398.6796.7281.7098.8394.660.9100.11420.5
✓✓✓85.4498.9296.9681.5898.7694.290.9190.09516.8
✓✓✓✓85.8998.8097.1482.4598.9294.820.9320.08914.9
✓✓✓✓✓85.9498.9997.2082.8699.0295.570.9440.08814.7
", + "bbox": [ + 246, + 454, + 754, + 556 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "of our space-to-object gaze regression method in modeling human-object gaze relationships while achieving more accurate gaze heatmap regression.", + "bbox": [ + 212, + 584, + 787, + 617 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation study and model analysis", + "text_level": 1, + "bbox": [ + 214, + 635, + 549, + 651 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation study about the effect of each component. In Table 7, we conduct various ablation studies on the GOO-Real dataset. We first establish a baseline using MaskDINO [20] and a gaze regression branch similar to TransGOP [40], and then gradually add the proposed modules and demonstrate their contribution to overall model performance. (i) Pixel shuffle (PS). After adding the pixel shuffle module, instance segmentation performance improved by $6.29\\%$ AP $(80.53\\%)$ vs $74.24\\%$ compared to baseline, indicating that large feature maps significantly enhance detailed features in dense prediction tasks. (ii) RoI Reconstruction (RoI-Rec.). The RoI reconstruction module reduces angle error by 39.3 and L2 distance by 0.22, demonstrating the importance of head-related features for gaze direction perception and the ability of the module to learn accurate gaze features in real-world settings. (iii) Dual attention fusion (DAF).", + "bbox": [ + 212, + 657, + 787, + 843 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Jin et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/12ee9406870d81717b4dc856df00e3b3393f8df47eac8cd740ae0d5017fdb234.jpg", + "table_caption": [ + "Table 8: Analysis about dual attention fusion module." + ], + "table_footnote": [], + "table_body": "
Original Attention Gaze ConeGaze Object SegmentationInstance SegmentationGaze Estimation
mSoCmSoC50mSoC75APAP50AP75AUCDist.↓Ang.↓
✓85.5198.8697.1782.0898.8794.780.9380.08915.6
✓84.1398.7896.3581.3098.7893.690.9260.09716.3
✓✓85.9498.9997.2082.7599.0195.170.9440.08814.7
", + "bbox": [ + 245, + 170, + 754, + 242 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/6c0bb8202cce1d96ff63a2136fc34ab68a9535bf035acb8166021be070f36abb.jpg", + "table_caption": [ + "Table 9: Analysis about feature interaction module" + ], + "table_footnote": [], + "table_body": "
Encoder QueryDecoder QueryMask EmbeddingGaze Object SegmentationInstance SegmentationGaze Estimation
mSoCmSoC50mSoC75APAP50AP75AUC Dist.↓Ang.↓
✓85.7998.3896.4481.2898.8794.870.9280.095 16.0
✓85.2098.2197.0181.6598.8694.650.9490.093 15.1
✓85.9498.9997.2082.8699.0295.570.9440.088 14.7
", + "bbox": [ + 245, + 281, + 754, + 354 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Introducing the dual attention fusion improves gaze heatmap quality (AUC: 0.919 vs 0.910), demonstrating that perceiving gaze object spatial location beforehand produces more accurate heatmaps and eases regression. (iv) Feature interaction module (FIM). The feature interaction module improves gaze estimation, reducing angle error by 1.9 and L2 distance by 0.006. Gaze object segmentation achieves an $85.89\\mathrm{mSoC}$ , showing that pixel-level object information enhances feature interaction and human-object gaze relationship modeling. (v) Energy aggregation loss (EAL). Adding energy aggregation loss improves the performance of all three sub-tasks, showing that the gaze object mask effectively guides the regression of the gaze heatmap towards the gaze object.", + "bbox": [ + 212, + 383, + 784, + 532 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Analysis about dual attention fusion module. Table 8 analyzed the effect of different spatial perception methods. Compared to using original attention $\\mathbf{M}_{\\mathrm{ori}}$ or gaze cone $\\mathbf{M}_{\\mathrm{cone}}$ alone, our dual attention fusion performs better. This indicates that combining spatial search along the gaze direction with human head-based scene saliency is more effective. Fig. 4(b) shows that element-wise fusion of $\\mathbf{M}_{\\mathrm{ori}}$ and $\\mathbf{M}_{\\mathrm{cone}}$ creates an initial human-object spatial connection.", + "bbox": [ + 212, + 534, + 784, + 625 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Analysis about feature interaction module. In Table 9, we evaluate the efficacy of different object information for gaze regression. Injecting the global encoder query of MaskDINO [20] results in $85.79\\%$ mSoC for gaze object segmentation but poor gaze estimation performance. Refining the decoder query with mask supervision reduces the angle error by 0.9, indicating better object representation. Using mask embedding with mask features during gaze regression achieves optimal performance, demonstrating that semantically clear mask features enhance gaze object positioning and the human-object gaze relationship.", + "bbox": [ + 212, + 625, + 785, + 746 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Qualitative Results.", + "text_level": 1, + "bbox": [ + 215, + 768, + 426, + 782 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Fig. 6(a) presents qualitative results in the real-world setting. In object-dense scenes, our method accurately predicts the gaze object mask without relying on head-related cues. In Fig. 6(b), compared to previous methods, our method ac", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Boosting Gaze Object Prediction via Pixel-level Supervision", + "bbox": [ + 330, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/96de6f78ea012101fb96efc55275baa794bc376c136ab1332a48dacc373341f3.jpg", + "image_caption": [ + "Fig. 6: Qualitative comparison of ground truth gaze object (GT), gaze object detection (GOD), instance segmentation (IS), gaze object segmentation (GOS)" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 390 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "curately locates adjacent objects using a pixel-level mask, effectively alleviating semantic ambiguity. See supplementary material for more visualization.", + "bbox": [ + 212, + 460, + 787, + 493 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 517, + 359, + 534 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce the gaze object segmentation task, which aims to infer pixel-level masks for objects captured by human eyesight, and establish a framework based on the pixel-level supervision from VFM to solve this task while alleviating the semantic ambiguity existing in previous box-level supervision methods. In our method, we choose SAM to generate pixel-level supervision information for the model training. The RoI reconstruction module is proposed to reconstruct head features directly from holistic features using the head box provided by the unified Transformer, which improves the inference efficiency and flexibility of our model in the real world. To boost the representation of the human-object gaze relationship, we propose a space-to-object gaze regression strategy, which first uses the dual attention fusion module to establish a human-object spatial connection, and then a semantically clear mask feature is injected into the gaze regression to strengthen the human-object gaze relationship modeling. Extensive experiments on GOO-Real and GOO-Synth demonstrate the effectiveness of our method in various settings.", + "bbox": [ + 212, + 551, + 787, + 777 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "**Limitation:** The proposed method requires box supervision as prompts for the SAM when generating masks, which limits its application in a wider range of scenarios without strong location priors, more flexible approaches for generating masks via the latest VFM models need to be explored in future work.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Jin et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported by the National Natural Science Foundation of China. (No. 82272020).", + "bbox": [ + 215, + 172, + 782, + 203 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Bin Fan is sponsored by China National Postdoctoral Program for Innovative Talents (No. BX20230013).", + "bbox": [ + 215, + 204, + 784, + 233 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 253, + 321, + 271 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bao, J., Liu, B., Yu, J.: ESCNet: Gaze target detection with the understanding of 3D scenes. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 14126-14135 (2022)", + "2. Cai, X., Zeng, J., Shan, S., Chen, X.: Source-free adaptive gaze estimation by uncertainty reduction. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 22035-22045 (2023)", + "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Eur. Conf. Comput. Vis. pp. 213-229 (2020)", + "4. Chen, W., Li, R., Yu, Q., Xu, A., Feng, Y., Wang, R., Zhao, L., Lin, Z., Yang, Y., Lin, D., et al.: Early detection of visual impairment in young children using a smartphone-based deep learning system. Nature Medicine 29(2), 493-503 (2023)", + "5. Chen, Y., Nan, Z., Xiang, T.: FBLNet: Feedback loop network for driver attention prediction. In: Int. Conf. Comput. Vis. pp. 13371-13380 (2023)", + "6. Cheng, Y., Lu, F.: Gaze estimation using transformer. In: Int. Conf. Pattern Recog. pp. 3341-3347 (2022)", + "7. Cheng, Y., Lu, F., Zhang, X.: Appearance-based gaze estimation via evaluation-guided asymmetric regression. In: Eur. Conf. Comput. Vis. pp. 100-115 (2018)", + "8. Chong, E., Ruiz, N., Wang, Y., Zhang, Y., Rozga, A., Rehg, J.M.: Connecting gaze, scene, and attention: Generalized attention estimation via joint modeling of gaze and scene saliency. In: Eur. Conf. Comput. Vis. pp. 383-398 (2018)", + "9. Chong, E., Wang, Y., Ruiz, N., Rehg, J.M.: Detecting attended visual targets in video. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5396-5406 (2020)", + "0. Gupta, A., Tafasca, S., Odobez, J.M.: A modular multimodal architecture for gaze target prediction: Application to privacy-sensitive settings. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5041-5050 (2022)", + "1. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask R-CNN. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2961-2969 (2017)", + "2. Hu, Z., Yang, D., Cheng, S., Zhou, L., Wu, S., Liu, J.: We know where they are looking at from the RGB-D camera: Gaze following in 3D. IEEE Trans. Instrum. Meas. 71, 1-14 (2022)", + "3. Hu, Z., Yang, Y., Zhai, X., Yang, D., Zhou, B., Liu, J.: GFIE: A dataset and baseline for gaze-following from 2D to 3D in indoor environments. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8907-8916 (2023)", + "4. Hu, Z., Zhao, K., Zhou, B., Guo, H., Wu, S., Yang, Y., Liu, J.: Gaze target estimation inspired by interactive attention. IEEE Trans. Circuit Syst. Video Technol. 32(12), 8524-8536 (2022)", + "5. Huang, T., Fu, R.: Driver distraction detection based on the true driver's focus of attention. IEEE Trans. Intell. Transp. Syst. 23(10), 19374-19386 (2022)", + "6. Jin, S., Wang, Z., Wang, L., Bi, N., Nguyen, T.: ReDirTrans: Latent-to-latent translation for gaze and head redirection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5547-5556 (2023)" + ], + "bbox": [ + 225, + 282, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Boosting Gaze Object Prediction via Pixel-level Supervision", + "bbox": [ + 330, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Jin, T., Yu, Q., Zhu, S., Lin, Z., Ren, J., Zhou, Y., Song, W.: Depth-aware gaze-following via auxiliary networks for robotics. Eng. Appl. Artif. Intell. 113, 104924 (2022)", + "18. Kellnhofer, P., Recasens, A., Stent, S., Matusik, W., Torralba, A.: Gaze360: Physically unconstrained gaze estimation in the wild. In: Int. Conf. Comput. Vis. pp. 6912-6921 (2019)", + "19. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Int. Conf. Comput. Vis. pp. 4015-4026 (2023)", + "20. Li, F., Zhang, H., Xu, H., Liu, S., Zhang, L., Ni, L.M., Shum, H.Y.: Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3041-3050 (2023)", + "21. Li, Y., Shen, W., Gao, Z., Zhu, Y., Zhai, G., Guo, G.: Looking here or there? gaze following in 360-degree images. In: Int. Conf. Comput. Vis. pp. 3742-3751 (2021)", + "22. Lian, D., Yu, Z., Gao, S.: Believe it or not, we know what you are looking at! In: Asian Conf. Comput. Vis. pp. 35-50 (2018)", + "23. Lv, K., Sheng, H., Xiong, Z., Li, W., Zheng, L.: Improving driver gaze prediction with reinforced attention. IEEE Trans. Multimedia 23, 4198-4207 (2020)", + "24. Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature Communications 15(1), 654 (2024)", + "25. Miao, Q., Hoai, M., Samaras, D.: Patch-level gaze distribution prediction for gaze following. In: IEEE Winter Conf. Appl. Comput. Vis. pp. 880-889 (2023)", + "26. Mundy, P., Sigman, M., Kasari, C.: A longitudinal study of joint attention and language development in autistic children. J. Autism Dev. Disord. 20(1), 115-128 (1990)", + "27. Park, S., Spurr, A., Hilliges, O.: Deep pictorial gaze estimation. In: Eur. Conf. Comput. Vis. pp. 721-738 (2018)", + "28. Park, S., Zhang, X., Bulling, A., Hilliges, O.: Learning to find eye region landmarks for remote gaze estimation in unconstrained settings. In: Proc. ACM Symp. Eye Track. Res. Appl. pp. 1-10 (2018)", + "29. Recasens, A., Khosla, A., Vondrick, C., Torralba, A.: Where are they looking? Adv. Neural Inform. Process. Syst. 28 (2015)", + "30. Recasens, A., Vondrick, C., Khosla, A., Torralba, A.: Following gaze in video. In: Int. Conf. Comput. Vis. pp. 1435-1443 (2017)", + "31. Ruzzi, A., Shi, X., Wang, X., Li, G., De Mello, S., Chang, H.J., Zhang, X., Hilliges, O.: GazeNeRF: 3d-aware gaze redirection with neural radiance fields. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 9676-9685 (2023)", + "32. Senju, A., Johnson, M.H.: Atypical eye contact in autism: Models, mechanisms and development. Neurosci. Biobehav. Rev. 33(8), 1204-1214 (2009)", + "33. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient subpixel convolutional neural network. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1874-1883 (2016)", + "34. Tafasca, S., Gupta, A., Odobez, J.M.: ChildPlay: A new benchmark for understanding children's gaze behaviour. In: Int. Conf. Comput. Vis. pp. 20935-20946 (2023)", + "35. Tomas, H., Reyes, M., Dionido, R., Ty, M., Mirando, J., Casimiro, J., Atienza, R., Quinto, R.: Goo: A dataset for gaze object prediction in retail environments. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. pp. 3125-3133 (2021)", + "36. Tonini, F., Beyer, C., Ricci, E.: Multimodal across domains gaze target detection. In: Int. Conf. Multimodal Interact. pp. 420-431 (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Jin et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Tonini, F., Dall'Asen, N., Bayan, C., Ricci, E.: Object-aware gaze target detection. In: Int. Conf. Comput. Vis. pp. 21860-21869 (2023)", + "38. Tu, D., Min, X., Duan, H., Guo, G., Zhai, G., Shen, W.: End-to-end human-gazetarget detection with transformers. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2192-2200 (2022)", + "39. Tu, D., Shen, W., Sun, W., Min, X., Zhai, G., Chen, C.: Un-Gaze: A unified transformer for joint gaze-location and gaze-object detection. IEEE Trans. Circuit Syst. Video Technol. (2023)", + "40. Wang, B., Guo, C., Jin, Y., Xia, H., Liu, N.: TransGOP: Transformer-based gaze object prediction. AAAI Conf. Artif. Intell. (2024)", + "41. Wang, B., Hu, T., Li, B., Chen, X., Zhang, Z.: GaTector: A unified framework for gaze object prediction. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19588-19597 (2022)", + "42. Wang, K., Ji, Q.: Real time eye gaze tracking with 3D deformable eye-face model. In: Int. Conf. Comput. Vis. pp. 1003-1011 (2017)", + "43. Wang, X., Zhang, H., Wang, Z., Nie, W., Yang, Z., Ren, W., Xu, Q., Xu, X., Liu, H.: Dual regression-enhanced gaze target detection in the wild. IEEE Trans. Cybern. 54(1), 219-229 (2024)", + "44. Wang, Z., Zhao, J., Lu, C., Yang, F., Huang, H., Guo, Y., et al.: Learning to detect head movement in unconstrained remote gaze estimation in the wild. In: IEEE Winter Conf. Appl. Comput. Vis. pp. 3443-3452 (2020)", + "45. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. Int. Conf. Learn. Represent. (2022)", + "46. Zhang, X., Sugano, Y., Fritz, M., Bulling, A.: Appearance-based gaze estimation in the wild. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4511-4520 (2015)", + "47. Zhang, X., Sugano, Y., Fritz, M., Bulling, A.: Mpiigaze: Real-world dataset and deep appearance-based gaze estimation. IEEE Trans. Pattern Anal. Mach. Intell. 41(1), 162-175 (2017)", + "48. Zhao, X., Ding, W., An, Y., Du, Y., Yu, T., Li, M., Tang, M., Wang, J.: Fast segment anything. arXiv preprint arXiv:2306.12156 (2023)", + "49. Zhu, W., Deng, H.: Monocular free-head 3D gaze tracking with deep learning and geometry constraints. In: Int. Conf. Comput. Vis. pp. 3143-3152 (2017)", + "50. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: Int. Conf. Learn. Represent. (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 646 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Boosting Gaze Object Prediction via Pixel-level Supervision", + "bbox": [ + 330, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Boosting Gaze Object Prediction via Pixel-level Supervision from Vision Foundation Model/93449abc-6022-4367-830a-8808cd9b5ba3_model.json b/2024/Boosting Gaze Object Prediction via Pixel-level Supervision from Vision Foundation Model/93449abc-6022-4367-830a-8808cd9b5ba3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3db9711b797bcf4fbc62d9b6b5b4079248335141 --- /dev/null +++ b/2024/Boosting Gaze Object Prediction via Pixel-level Supervision from Vision Foundation Model/93449abc-6022-4367-830a-8808cd9b5ba3_model.json @@ -0,0 +1,2390 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.229, + 0.141, + 0.776, + 0.187 + ], + "angle": 0, + "content": "Boosting Gaze Object Prediction via Pixel-level Supervision from Vision Foundation Model" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.217, + 0.761, + 0.234 + ], + "angle": 0, + "content": "Yang Jin1, Lei Zhang2, Shi Yan2, Bin Fan3, and Binglu Wang1" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.244, + 0.717, + 0.274 + ], + "angle": 0, + "content": "\\(^{1}\\) Xi'an University of Architecture and Technology, Xi'an, China {jin91999, wbl921129}@gmail.com † Corresponding author" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.274, + 0.76, + 0.3 + ], + "angle": 0, + "content": "\\(^{2}\\) School of Automation, Northwestern Polytechnical University, Xi'an, China \nzl_hnly@163.com, yanshi@mail.nwpu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.301, + 0.772, + 0.343 + ], + "angle": 0, + "content": "\\(^{3}\\) National Key Lab of General AI, School of Intelligence Science and Technology, Peking University, Beijing, China binfan@pku.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.42, + 0.74, + 0.794 + ], + "angle": 0, + "content": "Abstract. Gaze object prediction (GOP) aims to predict the category and location of the object that a human is looking at. Previous methods utilized box-level supervision to identify the object that a person is looking at, but struggled with semantic ambiguity, i.e., a single box may contain several items since objects are close together. The Vision foundation model (VFM) has improved in object segmentation using box prompts, which can reduce confusion by more precisely locating objects, offering advantages for fine-grained prediction of gaze objects. This paper presents a more challenging gaze object segmentation (GOS) task, which involves inferring the pixel-level mask corresponding to the object captured by human gaze behavior. In particular, we propose that the pixel-level supervision provided by VFM can be integrated into gaze object prediction to mitigate semantic ambiguity. This leads to our gaze object detection and segmentation framework that enables accurate pixel-level predictions. Different from previous methods that require additional head input or ignore head features, we propose to automatically obtain head features from scene features to ensure the model's inference efficiency and flexibility in the real world. Moreover, rather than directly fuse features to predict gaze heatmap as in existing methods, which may overlook spatial location and subtle details of the object, we develop a space-to-object gaze regression method to facilitate human-object gaze interaction. Specifically, it first constructs an initial human-object spatial connection, then refines this connection by interacting with semantically clear features in the segmentation branch, ultimately predicting a gaze heatmap for precise localization. Extensive experiments on GOO-Synth and GOO-Real datasets demonstrate the effectiveness of our method. The code will be available at https://github.com/jinyang06/SamGOP." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.812, + 0.741, + 0.84 + ], + "angle": 0, + "content": "Keywords: Gaze object prediction \\(\\cdot\\) Vision foundation model \\(\\cdot\\) Object segmentation \\(\\cdot\\) Space-to-object gaze regression" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "Y. Jin et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.501, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.256, + 0.501, + 0.267 + ], + "angle": 0, + "content": "(a) Illustration of semantic ambiguity and semantic clarity" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.147, + 0.788, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.256, + 0.757, + 0.267 + ], + "angle": 0, + "content": "(b) The main idea of our pipeline" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.278, + 0.788, + 0.362 + ], + "angle": 0, + "content": "Fig. 1: (a) Box-level supervision often fails to localize objects in dense settings precisely and leads to semantic ambiguity problems, whereas pixel-level supervision excels by providing clear semantic distinction through pixel-by-pixel predictions. (b) Vision foundation models can produce instance masks, thereby segmentation features can be used to improve the gaze regression branch's spatial perception, and the gaze object mask can help the heatmap focus on the gaze object." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.388, + 0.376, + 0.403 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.788, + 0.553 + ], + "angle": 0, + "content": "The objects stared at by humans contain important semantic information, which can reveal a person's behavior and state of mind. Therefore, identifying the objects that people are looking at has a wide range of applications. For example, in medical diagnosis, whether a child focuses on an object can potentially reveal whether he/she has autism or visual impairment [4,26,32,34]. In assisted driving scenarios, determining what object the driver is looking at can assess their attention and provide corresponding driving guidance [5,15,23]. In general, detecting the object gazed at by humans realizes the interaction between human sight and specific objects, fostering a richer semantic understanding of the scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Tomas et al. [35] first proposed the gaze object detection task and the benchmark datasets GOO-Synth and GOO-Real, aiming to predict the object boxes and categories stared at by humans. Recently, Wang et al. [40, 41] conducted an in-depth exploration of this task. Despite achieving promising results, they still suffer from the following limitations: 1) They typically rely on box-level supervision, which is prone to suffer semantic ambiguity in dense object scenes. Especially when multiple objects are densely stacked or placed adjacent to each other, the box-level supervision algorithm may struggle to accurately separate them, as illustrated in Fig. 1(a). 2) The GOO [35] dataset only contains pixel-level annotations of synthetic images, leading to difficulties in generalizing to fine-grained predictions in real scenes. 3) They usually require additional head information as input, affecting the efficiency and flexibility of the model. While global modeling using the Transformer [38, 39] can alleviate this problem, it lacks special attention to head features, which are crucial for gaze direction perception. 4) They directly fuse features to regress the gaze heatmap, lacking a full understanding of the spatial location and subtle detailed information of objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Recently, vision foundational models (VFM) have gained significant attention due to their powerful generalization capabilities, which can achieve accurate pixel-level segmentation using a variety of prompts, i.e., boxes, points, or masks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Boosting Gaze Object Prediction via Pixel-level Supervision" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "In real scenes with small and densely packed objects, VFM can generate accurate boundaries and pixel-level masks (See Fig 3) by utilizing boxes as prompt, which reduces semantic ambiguity and thus opens opportunities for gaze object segmentation. Motivated by this, we propose a more challenging task, gaze object segmentation (GOS). This task involves inferring the pixel-level mask of objects captured by human gaze behavior, helping to further enhance the understanding of human gaze behavior. Furthermore, we establish a GOS framework specifically for this task, which integrates the pixel-level supervision provided by VFM into gaze object prediction to alleviate the aforementioned limitations of previous methods [40, 41]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.788, + 0.617 + ], + "angle": 0, + "content": "In the proposed model, we introduce the unified detection and segmentation Transformer, i.e. MaskDINO [20], to simultaneously obtain candidate gaze object boxes, segmentation masks, and head boxes of the subject, which are indispensable for enabling the model to infer in the real world and achieve accurate detection. Instead of requiring additional head input [40, 41] or ignoring head features by global modeling [38, 39], we propose a RoI reconstruction module to automatically obtain head region features based on holistic features and head box provided by the unified Transformer. This helps capture gaze direction information during gaze regression to improve model efficiency and flexibility. Subsequently, a space-to-object gaze regression strategy is developed to enhance the perception of the spatial location of objects and the understanding of detailed information. Specifically, inspired by the ability of humans to infer others' gaze objects, a dual attention fusion module is presented to establish an initial human-object gaze spatial connection. Then, we propose an object feature interaction module to refine the gaze spatial connection through semantically clear feature interaction between gaze regression features and object mask features, thus strengthening the modeling of the human-object gaze relationship. In the end, an energy aggregation loss is used to further guide the regression of the gaze heatmap. By focusing on pixel-level details and progressively improving the gaze representation, it further reduces semantic ambiguity, leading to the generation of more precise heatmaps for gazed objects." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.617, + 0.645, + 0.63 + ], + "angle": 0, + "content": "Our main contributions could be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.643, + 0.787, + 0.704 + ], + "angle": 0, + "content": "- We introduce the vision foundation model into gaze object prediction to provide pixel-level supervision and propose a more challenging gaze object segmentation task, which focuses on identifying the pixel-level masks of objects captured by human gaze behavior." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.705, + 0.788, + 0.78 + ], + "angle": 0, + "content": "- We present an all-in-one end-to-end framework for gaze object detection and segmentation without relying on additional head-related inputs, capable of simultaneously addressing gaze estimation, object detection, and segmentation. To the best of our knowledge, our work is the first to tackle gaze object prediction from the image segmentation perspective." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.781, + 0.788, + 0.84 + ], + "angle": 0, + "content": "- We propose a space-to-object gaze regression approach that enhances the interaction between object and gaze branches, incrementally improving the modeling of the human-object gaze relationship and alleviating semantic ambiguity." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.643, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "Y. Jin et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.398, + 0.161 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.786, + 0.298 + ], + "angle": 0, + "content": "Vision Foundation Model. The vision foundation models (VFMs) [19, 24, 48] are typically built on large-scale datasets using self-supervised or semi-supervised methods, which have strong generalizability and zero-shot transfer capabilities. Segment Anything Model (SAM) [19] received attention as a foundation model for image segmentation. Its portability enables it to be flexibly applied to various downstream tasks. Then, the medical foundation model MedSAM [24] extends the foundation model to medical diagnosis. In this article, we choose SAM [19] to obtain high-quality masks to boost gaze estimation and gaze object prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.787, + 0.435 + ], + "angle": 0, + "content": "Gaze Object Detection. The gaze object detection is first proposed in [35], which aims to predict the box and categories of the object stared at by humans in an RGB image. Tomas et al. [35] released a benchmark dataset for GOP. Wang et al. [41] first proposed a GOP method named GaTector. However, GaTector requires additional head-related prior information, which hinders its application in the real world. Recently, the Transformer-based method [40] achieves comparable performance. Although previous methods are effective to some extent, they can only provide instance-level box outputs, leading to potential ambiguities or misjudgments in scenarios with small and densely packed objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.787, + 0.631 + ], + "angle": 0, + "content": "Gaze Target Detection. Early research on gaze focused on gaze direction estimation [2, 5-7, 16, 18, 27, 28, 31, 42, 44, 46, 47, 49], which only predict the gaze vector and ignores the impact of the surrounding environment on the sight, resulting in limited application scenarios. Therefore, to capture higher-level gaze-related semantic information, researchers propose gaze target detection [1, 8-10, 12-14, 17, 21, 22, 25, 30, 34, 36, 37, 43], which involves to infer the specific gaze point. Recasens et al. [29] introduced the GazeFollow dataset and first proposed gaze target detection. Recently, Tu et al. [38,39] achieved end-to-end gaze target detection using Transformer, with a relatively low detection efficiency. Despite advancements in gaze target estimation, previous methods struggle with gaze regression due to spatial perception ambiguity. In contrast, our space-to-object regression strategy helps the model gradually converge to a certain gaze point, simplifying learning and enhancing detailed interaction with the object." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.654, + 0.331, + 0.67 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.686, + 0.787, + 0.747 + ], + "angle": 0, + "content": "Given an RGB image, our model aims to precisely predict the pixel-level mask of the object humans are looking at. We first present the overall structure of our method, then introduce the VFM-based mask supervision generation method, and finally describe each module of the proposed model in detail." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.77, + 0.341, + 0.784 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.841 + ], + "angle": 0, + "content": "As shown in Fig. 2, our method consists of four components: a feature extractor for specific feature extraction, a detection and segmentation module predicting candidate gaze objects boxes, masks, and human head location, a gaze regressor" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Boosting Gaze Object Prediction via Pixel-level Supervision" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.145, + 0.784, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.303, + 0.785, + 0.415 + ], + "angle": 0, + "content": "Fig. 2: Overview of the proposed model. (a) The feature extraction module extracts features for detection and regression branches. (b) The detection and segmentation branch identifies object and human head positions. (c) After obtaining head features, the gaze regression branch progressively refines its output: 1) employing a dual attention fusion module for initial human-object correlations; 2) leveraging a feature interaction module to incorporate semantically clear object-aware insights from the segmentation branch; 3) ultimately predicting the gaze heatmap. (d) Supervision signals are applied to both branches only during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.784, + 0.489 + ], + "angle": 0, + "content": "producing gaze object heatmap with a space-to-object strategy, and a supervision information generation module creating pixel-level masks and ground truth heatmap during training. At training, the overall loss function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.497, + 0.786, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {d e t}} + \\alpha \\mathcal {L} _ {\\mathrm {d i r}} + \\beta \\mathcal {L} _ {\\mathrm {g a z e}} + \\gamma \\mathcal {L} _ {\\mathrm {e n g}} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.784, + 0.566 + ], + "angle": 0, + "content": "where \\(\\alpha\\), \\(\\beta\\), and \\(\\gamma\\) are the weights of the gaze direction loss, gaze heatmap loss, and energy aggregation loss respectively. Before model training, we choose SAM to generate pixel-level masks for the objects. See Sec. 3.2 for details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.785, + 0.749 + ], + "angle": 0, + "content": "Feature extraction. As shown in 2(a), given an RGB image \\(\\mathcal{I} \\in \\mathbb{R}^{3 \\times H \\times W}\\), we first use the backbone \\(\\mathcal{B}\\) to extract multi-scale features \\(\\{\\mathbf{f}_1, \\mathbf{f}_2, \\mathbf{f}_3, \\mathbf{f}_4\\}\\). In previous work [41], pixel shuffle [33] significantly improved small object detection by enlarging feature maps for detailed feature capture. Therefore, we integrate it into dense prediction tasks, i.e., image segmentation. In our model, the multi-scale features \\(\\{\\mathbf{f}_1, \\mathbf{f}_2, \\mathbf{f}_3, \\mathbf{f}_4\\}\\) are fed into the pixel shuffle module to generate object-specific features \\(\\mathbf{f}_i^{obj} = \\phi^{obj}(\\mathbf{f}_i, \\eta)\\), where \\(\\mathbf{f}_i\\) denotes the \\(i\\)-th feature layer, \\(\\eta\\) denotes the scale factor, \\(\\phi^{obj}(\\cdot)\\) denotes the pixel shuffle module, which consists of a pixel shuffle operation and a convolution layer that transforms the feature channels to be consistent with the original features of the unified Transformer [20]. Finally, the scene saliency features \\(\\mathbf{f}_{\\mathrm{scene}} \\in \\mathbb{R}^{1024 \\times 7 \\times 7}\\) are extracted by a scene residual block using high dimensional features \\(\\mathbf{f}_4 \\in \\mathbb{R}^{2048 \\times 7 \\times 7}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Unified detection and segmentation Transformer. In this paper, we employ a unified Transformer for candidate gaze object detection, segmentation, and human head positioning, integrating MaskDINO [20] for candidate gaze object detection, segmentation, and the generation of pixel-level mask feature embeddings and a head decoder for head box regression, which facilitate accurate head-related feature extraction in head feature reconstruction module and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "Y. Jin et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.217, + 0.692, + 0.232 + ], + "angle": 0, + "content": "Fig. 3: Illustration of instance mask generated by VFM." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.26, + 0.741, + 0.275 + ], + "angle": 0, + "content": "accurate prediction in real-world settings. The training loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.283, + 0.786, + 0.304 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {d e t}} = \\mathcal {L} _ {\\mathrm {d e t}} ^ {\\mathrm {o b j}} + \\mathcal {L} _ {\\mathrm {m a s k}} ^ {\\mathrm {o b j}} + \\mathcal {L} _ {\\mathrm {d e t}} ^ {\\mathrm {h e a d}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.311, + 0.789, + 0.482 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{\\mathrm{det}}^{\\mathrm{obj}}\\) and \\(\\mathcal{L}_{\\mathrm{mask}}^{\\mathrm{obj}}\\) is the object detection and segmentation loss respectively. \\(\\mathcal{L}_{\\mathrm{det}}^{\\mathrm{head}}\\) is the head decoder loss, which is consistent with the \\(\\mathcal{L}_{\\mathrm{det}}^{\\mathrm{obj}}\\). See Sec. 3.3 for details. Space-to-object gaze regression. We employ a space-to-object regression strategy for gaze heatmap prediction. The module begins by representing human gaze behavior with a gaze vector. Subsequently, a dual attention fusion module establishes an initial human-object spatial connection. Finally, pixel-level object location knowledge from the mask branch is introduced to achieve semantically clear feature interaction, refining the gaze heatmap regression while alleviating semantic ambiguity. Optimization involves \\(\\mathcal{L}_{\\mathrm{dir}}\\) and \\(\\mathcal{L}_{\\mathrm{gaze}}\\) for gaze direction and gaze heatmap, respectively. The energy aggregation loss \\(\\mathcal{L}_{\\mathrm{eng}}\\) guides the gaze heatmap to focus on the gaze object mask. See Sec. 3.4 for details." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.503, + 0.557, + 0.518 + ], + "angle": 0, + "content": "3.2 Instance mask generation via VFM." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.526, + 0.788, + 0.603 + ], + "angle": 0, + "content": "Given an RGB image \\(\\mathcal{I}\\), we choose the vision foundation model SAM [19] to obtain instance masks for model training. Specifically, image \\(\\mathcal{I}\\) and prompt \\(\\mathcal{P}_{box}\\) are first fed into the image encoder \\(\\varPsi_{Image}\\) and prompt encoder \\(\\varPsi_{prompt}\\), respectively. Then, the mask \\(\\mathcal{P}_{mask}\\) and confidence \\(\\mathcal{S}\\) are output by the mask decoder \\(\\varPhi_{mask}\\), which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.34, + 0.613, + 0.787, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} _ {\\text {m a s k}}, \\mathcal {S} = \\Phi_ {\\text {m a s k}} (\\Psi_ {\\text {I m a g e}} (\\mathcal {I}), \\Psi_ {\\text {p r o m p t}} (\\mathcal {P} _ {\\text {b o x}})) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.638, + 0.788, + 0.7 + ], + "angle": 0, + "content": "Instance segmentation by using boxes produces accurate masks for most objects, but accurate segmentation of objects with unclear texture or edge features remains challenging. Therefore, we segment again using \\(\\mathcal{P}_{mask}\\) as a dense prompt to produce final mask \\(\\mathcal{M}_{\\mathrm{mask}}\\), which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.33, + 0.71, + 0.787, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {\\text {m a s k}}, \\mathcal {S} = \\Phi_ {\\text {m a s k}} (\\Psi_ {\\text {I m a g e}} (\\mathcal {I}), \\Psi_ {\\text {p r o m p t}} (\\mathcal {P} _ {\\text {m a s k}})) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.735, + 0.787, + 0.766 + ], + "angle": 0, + "content": "Finally, \\(\\mathcal{M}_{mask}\\) is used to train the gaze object segmentation model. The generated masks are shown in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.786, + 0.664, + 0.802 + ], + "angle": 0, + "content": "3.3 Unified detection and segmentation Transformer" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "The unified detection and segmentation Transformer aims to detect candidate gaze objects and locate human heads while extracting high-level pixel mask" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Boosting Gaze Object Prediction via Pixel-level Supervision" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.208 + ], + "angle": 0, + "content": "features, which comprises a detection and segmentation Transformer, a head decoder, and a RoI feature reconstruction module. As shown in Fig. 2(b), the object-specific features \\(\\{\\mathbf{f}_1^{obj},\\mathbf{f}_2^{obj},\\mathbf{f}_3^{obj},\\mathbf{f}_4^{obj}\\}\\) are fed it to detect object boxes, categories, masks, and human head boxes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.21, + 0.785, + 0.285 + ], + "angle": 0, + "content": "Object detection and segmentation The global attention of Transformer ensures powerful long-distance modeling and global perception, achieving excellent detection performance through layer-by-layer self-attention refinement. We use MaskDINO [20] as a foundational detection and segmentation structure due to its robust representation capabilities and fast convergence speed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.286, + 0.785, + 0.391 + ], + "angle": 0, + "content": "Head decoder Our model is designed for real-world gaze object detection and segmentation without relying on head priors. We introduce a head decoder based on the unified Transformer, using a 3-layer self-attention decoder to generate proposals for head-related feature reconstruction. During training, we filter head queries with IoU \\(>0.5\\) as positives. During inference, boxes are selected based on confidence. Similar to other DETR-like detectors [45, 50], the head decoder uses cross-entropy, \\(L_{1}\\), and GIoU loss for classification and box regression." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.392, + 0.787, + 0.603 + ], + "angle": 0, + "content": "Head feature reconstruction Previous methods typically extract head features from manually cropped head images, limiting model inference efficiency and flexibility in the real world. To this end, we design a RoI feature reconstruction module to reconstruct head-related features based on holistic scene features and the head box from the unified Transformer. In this module, we first use the head bounding box \\( R = (x_{1},y_{1},x_{2},y_{2}) \\), which represents the upper left and lower right corner coordinates, as the region proposal to crop the regions of interest (ROIs). Then, those candidate ROIs are refined by ROAlign [11], which maps the head region to the corresponding pixels in the holistic image features \\( \\mathbf{f}_4 \\in \\mathbb{R}^{2048 \\times 7 \\times 7} \\) and produces a fixed-size feature map \\( \\mathbf{f}_{\\mathrm{head}} \\in \\mathbb{R}^{2048 \\times 7 \\times 7} \\). For head location map generation, following previous works [9,10,29,41] to generate a binary image, setting the head area to 1 and the rest to 0. Additionally, a specific residual block extracts gaze saliency features \\( \\mathbf{f}_{\\mathrm{gaze}} \\in \\mathbb{R}^{1024 \\times 7 \\times 7} \\), further enhancing the model's attention to the gaze direction." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.63, + 0.52, + 0.646 + ], + "angle": 0, + "content": "3.4 Space-to-object gaze regression" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.784, + 0.733 + ], + "angle": 0, + "content": "We propose a space-to-object gaze regression approach to gradually refine the human-object gaze relationship. Fig. 2(c) shows our gaze regression branch with a dual attention fusion module establishing an initial human-object spatial connection and a feature interaction module strengthening this connection using semantically clear feature interaction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.784, + 0.809 + ], + "angle": 0, + "content": "Spatial perception for gaze objects. When humans infer what another person is staring at, they usually first observe the gaze direction of the others and search for the gazed object along this direction. Inspired by this, we propose the dual attention fusion module for the spatial perception of gaze objects, establishing an initial human-object spatial connection." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "As shown in Fig. 4(a), we first use global average pooling to obtain the global features of \\(\\mathbf{f}_{\\mathrm{head}}\\), followed by fully connected layers to transform the features into" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "Y. Jin et al." + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.143, + 0.487, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.146, + 0.745, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.253, + 0.692, + 0.267 + ], + "angle": 0, + "content": "Fig. 4: Illustration of spatial perception for gaze objects." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.296, + 0.787, + 0.327 + ], + "angle": 0, + "content": "a 2D gaze vector \\(\\mathbf{v}_g\\). To achieve spatial perception, we project it into a gaze cone map \\(\\mathbf{M}_{\\mathrm{cone}} \\in \\mathbb{R}^{1 \\times 7 \\times 7}\\), activating the area around the gaze direction:" + }, + { + "type": "equation", + "bbox": [ + 0.392, + 0.338, + 0.786, + 0.356 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {\\mathrm {i}, \\mathrm {j}} ^ {\\text {c o n e}} = M a x (0, \\cos (\\mathbf {v} _ {\\mathrm {g}}, \\mathbf {v} _ {\\mathrm {i}, \\mathrm {j}})) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.366, + 0.788, + 0.426 + ], + "angle": 0, + "content": "where \\(\\mathbf{v}_{\\mathrm{g}}\\) and \\(\\mathbf{v}_{\\mathrm{i,j}}\\) respectively denotes the predicted gaze vector and the vector composed of the eye point and each point on the gaze field. During training, \\(\\mathbf{v}_{\\mathrm{g}}\\) is optimized by \\(\\mathcal{L}_{\\mathrm{dir}}\\) to obtain a more accurate gaze direction and activation map. \\(\\mathcal{L}_{\\mathrm{dir}}\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.418, + 0.436, + 0.786, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {d i r}} = 1 - \\sum_ {i = 1} ^ {n} \\left(\\mathbf {v} _ {\\mathrm {g}} ^ {i} \\cdot \\mathbf {v} _ {\\mathrm {t}} ^ {i}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.485, + 0.788, + 0.592 + ], + "angle": 0, + "content": "where vector \\(\\mathbf{v}_{\\mathrm{g}}\\) and \\(\\mathbf{v}_{\\mathrm{t}}\\) denote the predicted gaze direction and ground truth, respectively. During backpropagation, \\(\\mathcal{L}_{\\mathrm{dir}}\\) adjusts parameters of head feature \\(\\mathbf{f}_{\\mathrm{head}}\\), providing more direction information for gaze regression. To boost scene attention, we adopt the method from GaTector [41] to generate an original attention map \\(\\mathbf{M}_{\\mathrm{ori}} \\in \\mathbb{R}^{1 \\times 7 \\times 7}\\), activating scene salience area based on head properties. By multiplying \\(\\mathbf{M}_{\\mathrm{cone}}\\) with \\(\\mathbf{M}_{\\mathrm{ori}}\\) via element-wise dot product, we generate the final spatial perception map \\(\\mathbf{M}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.429, + 0.602, + 0.786, + 0.619 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} = \\mathbf {M} _ {\\text {o r i}} \\odot \\mathbf {M} _ {\\text {c o n e}} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.784, + 0.659 + ], + "angle": 0, + "content": "This compels the model to search spaces more likely gazed at by humans along gaze direction, establishing an initial human-object gaze spatial relationship." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.785, + 0.749 + ], + "angle": 0, + "content": "Semantically clear feature interaction. Previous methods [40, 41] usually directly fuse features to regress the gaze heatmap based on spatial saliency, lacking interaction with object details and facing challenges in accurately predicting heatmap for the gazed object. In our method, we introduce a feature interaction module, which leverages semantically clear mask knowledge from VFM to achieve feature interaction, refining human-object gaze relationship modeling." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As shown in Fig. 2(c), we initially extract detailed mask features from the mask branch. Subsequently, the fused feature \\(\\mathbf{f}_{\\mathrm{fuse}}\\) is utilized for pixel-level attention perception. Given that candidate gaze objects can be located anywhere in the image, the Transformer's global modeling excels in accomplishing this task. Therefore, utilize a transformer layer, akin to DETR [3], to capture the global relationship between human gaze behavior and object masks. For fine-grained" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.33, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Boosting Gaze Object Prediction via Pixel-level Supervision" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "object mask features, a three-layer MLP is applied to process the decoder query \\(\\mathcal{Q}_{\\mathrm{obj}}\\) of the unified Transformer, generating mask embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.186, + 0.786, + 0.202 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Q} _ {\\text {m a s k}} = \\mathbf {M L P} \\left(\\mathcal {Q} _ {\\text {o b j}}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.784, + 0.239 + ], + "angle": 0, + "content": "where the \\(\\mathcal{Q}_{\\mathrm{mask}}\\) denote the generated mask embedding. We first apply a layer of self-attention for \\(\\mathbf{f}_{\\mathrm{fuse}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.247, + 0.786, + 0.263 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {\\mathrm {e}} = \\mathbf {F F N} (\\mathbf {S e l f A t t n} (\\mathbf {f} _ {\\mathrm {f u s e}})) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.784, + 0.316 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}_{\\mathrm{fuse}}\\), SelfAttn(\\(\\cdot\\)), and \\(\\mathbf{f}_{\\mathrm{e}}\\) denote the fused feature \\(\\mathbf{f}_{\\mathrm{fuse}}\\), self-attention operation, and encoded \\(\\mathbf{f}_{\\mathrm{fuse}}\\), respectively. Then, the \\(\\mathcal{Q}_{\\mathrm{mask}}\\) is perceived by the \\(\\mathbf{f}_{\\mathrm{e}}\\) as key-value pairs, i.e., \\(\\{\\mathbf{K}_{\\mathrm{mask}}, \\mathbf{V}_{\\mathrm{mask}}\\}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.323, + 0.786, + 0.339 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {\\text {r e g}} = \\mathbf {F F N} (\\mathbf {C r o s s A t t n} (\\mathbf {f} _ {\\mathrm {e}}, \\{\\mathbf {K} _ {\\text {m a s k}}, \\mathbf {V} _ {\\text {m a s k}} \\})) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.345, + 0.784, + 0.42 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}_{\\mathrm{reg}}\\) and CrossAttn \\((q, kv)\\) denote the regression feature and cross-attention operation, respectively. This operation refines the perception of the object by providing a pixel-level understanding of the gaze object in each token of the regression feature. It effectively alleviates semantic ambiguity and strengthens the ability of the model to represent the human-object gaze relationship." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.784, + 0.511 + ], + "angle": 0, + "content": "Ultimate gaze regression. Finally, following previous work [41] to generate the ground truth gaze heatmap and use \\( L_{2} \\) loss to supervise gaze heatmap regression. \\( \\mathcal{L}_{\\mathrm{eng}} \\) proposed by [40] guides gaze heatmap energy aggregation towards the gaze object. We use the gaze object mask for more accurate error measurements. Space-to-object gaze regression approach gradually refines the human-object gaze behavior representation, regressing a more accurate gaze heatmap." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.531, + 0.376, + 0.548 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.559, + 0.439, + 0.574 + ], + "angle": 0, + "content": "4.1 Experiments Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.784, + 0.791 + ], + "angle": 0, + "content": "Datasets and implementation details are given in supplementary material. Metrics: Following previous work [40,41], we use AUC (Area Under the Curve), L2 distance, and Angle error for gaze estimation evaluation, and mSoC (min Single over Closure)4 under different thresholds for gaze object prediction evaluation. Following MaskDINO [20], AP (Average Precision) is used for instance segmentation and object detection evaluation. To evaluate the gaze object segmentation, we use masks to replace boxes when calculating mSoC (details refer to the supplementary material). For a fair comparison, we reported results for both real-world and non-real-world settings. The non-real-world settings involve inputting additional head images and head location maps while real-world settings only use scene images. Furthermore, to compare with GTR [39], we also adopt the evaluation method proposed by [39]5. This method computes gaze estimation metrics only for results with L2 distance \\(< 0.15\\) and head box IoU \\(> 0.5\\), which is an incomplete evaluation for AUC, L2 distance, and Angle error." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.797, + 0.785, + 0.812 + ], + "angle": 0, + "content": "4 mSoC metric proposed by Wang et al. URL: https://arxiv.org/abs/2112.03549." + }, + { + "type": "page_footnote", + "bbox": [ + 0.223, + 0.812, + 0.785, + 0.839 + ], + "angle": 0, + "content": "\\(^{5}\\) Evaluation method proposed by Tu et al., URL: https://ieeexplore.ieee.org/document/10262016." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.797, + 0.785, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "Y. Jin et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.174 + ], + "angle": 0, + "content": "Table 1: Gaze object detection results on GOO-Synth and GOO-Real datasets. * indicates using additional head detector to achieve real-world settings." + }, + { + "type": "table", + "bbox": [ + 0.232, + 0.185, + 0.77, + 0.299 + ], + "angle": 0, + "content": "
MethodGOO-SynthGOO-Real No PretrainGOO-Real PretrainParamsFPS
mSoC507595mSoC507595mSoC507595
Non-RealGaTector [41]67.9498.1486.250.1262.4095.1073.500.2071.2097.5088.802.7060.78M14.1
TransGOP [40]92.8099.0098.5051.9082.6097.8089.406.5089.0098.9097.5033.2094.03M13.7
Ours91.1993.9793.4673.2781.7598.7896.437.1381.8998.9996.637.2880.99M14.5
RealGaTector [41]*56.1988.0866.18-40.8782.0533.540.3066.4096.2280.490.4075.73M9.3
TransGOP [40]*91.6099.0098.1048.7079.7097.8091.4011.2083.2098.8095.3014.40106.37M11.6
Ours90.7793.9293.4569.5881.3598.9596.022.6181.6099.0396.546.0384.26M14.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.234, + 0.311, + 0.768, + 0.326 + ], + "angle": 0, + "content": "Table 2: Ours gaze object segmentation results on GOO-Real and GOO-Synth." + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.339, + 0.77, + 0.404 + ], + "angle": 0, + "content": "
SettingsGOO-SynthGOO-Real No PretrainGOO-Real PretrainParams|FPS
mSoC507595mSoC507595mSoC507595
Non-Real82.6893.7091.6826.8386.3998.9597.4420.0789.3998.8098.0233.0380.99M
Real82.7393.6791.7926.8585.9498.9997.2916.7086.9698.9497.7522.2184.26M
" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.417, + 0.516, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.418, + 0.758, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.487, + 0.776, + 0.502 + ], + "angle": 0, + "content": "Fig. 5: Comparison of TransGOP [40] and our method inference in the real world." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.53, + 0.461, + 0.546 + ], + "angle": 0, + "content": "4.2 Comparison with SOTA" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Gaze object detection and segmentation. Table 1 shows the performance of gaze object detection on GOO-Synth and GOO-Real. In real-world settings, previous methods [40, 41] significantly increase parameters and slow down the inference speed due to the additional head detector (As shown in Fig. 5 (a)) and they can only achieve box-level prediction. In contrast, although our method is slightly inferior to TransGOP [40] in detection performance, our model runs significantly more efficiently because it integrates the acquisition of head features into the overall framework (As shown in Fig. 5 (b)) and can output bounding boxes and segmentation masks simultaneously. The above results also show that our space-to-object gaze regression method can effectively model the human-object gaze relationship. Table 2 shows our gaze object segmentation performance in real-world and non-real-world scenarios. On GOO-Synth, our model achieves \\(82.68\\%\\) and \\(82.73\\%\\) mSoC in real-world and non-real-world settings. On GOO-Real, without pre-training, our model performs well in both scenarios \\((86.39\\% \\mathrm{mSoC}\\) and \\(85.94\\%\\) mSoC). After pre-training, it reaches \\(89.39\\%\\) mSoC and \\(86.96\\%\\) mSoC. The above experimental results show that the pixel-level mask knowledge provided by VFMs can effectively alleviate semantic ambiguity while achieving pixel-level prediction. Even under the multi-task learning framework, we still achieve an FPS of 14.1. This is mainly attributed to the automatic" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Boosting Gaze Object Prediction via Pixel-level Supervision" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.235, + 0.145, + 0.768, + 0.16 + ], + "angle": 0, + "content": "Table 3: Instance segmentation results on GOO-Synth and GOO-Real datasets" + }, + { + "type": "table", + "bbox": [ + 0.261, + 0.172, + 0.744, + 0.263 + ], + "angle": 0, + "content": "
MethodGOO-Synth DatasetGOO-Real DatasetParams
AP507595AP507595
Non-RealOurs79.3393.5191.0911.7282.9198.8595.0911.2280.99M
RealMaskDINO [20]70.8595.5483.311.8177.0598.3890.454.2652.03M
Ours78.5696.0090.568.4382.8699.0295.5710.0984.26M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.277, + 0.787, + 0.307 + ], + "angle": 0, + "content": "Table 4: Object detection results on GOO-Synth and GOO-Real datasets. * indicates using additional head detector to achieve real-world settings." + }, + { + "type": "table", + "bbox": [ + 0.261, + 0.318, + 0.744, + 0.46 + ], + "angle": 0, + "content": "
MethodGOO-Synth DatasetGOO-Real DatasetParams
AP507595AP507595
Non-RealGaTector [41]56.8095.3062.500.1052.2591.9255.340.1060.78M
TransGOP [40]87.6099.0097.3025.5077.2097.8089.406.5094.03M
Ours88.3393.8893.0850.7077.3298.6593.911.9380.99M
RealGaTector [41]*43.5086.2337.52-29.2174.0216.13-75.73M
TransGOP [40]*86.0098.8096.2023.8073.3096.7086.004.10106.37M
MaskDINO [20]72.9396.2688.132.0474.4498.7291.190.1952.03M
Ours87.6093.7273.0745.5375.2098.9493.200.7684.26M
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.785, + 0.52 + ], + "angle": 0, + "content": "incorporation of head features, avoiding the need for an additional head detector and thus enhancing the efficiency of the model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Object detection and segmentation. Tables 3 and 4 show our model's instance segmentation and object detection performance. For a fair comparison, we adjusted the input size of MaskDINO [20] to match ours \\((224 \\times 224)\\). In Table 3, our model outperforms MaskDINO by 8.66 AP on GOO-Real (70.85 vs. 79.51) and by 5.81 AP on GOO-Synth (77.05 vs. 82.86) in real-world scenarios, with similar advantages in non-real-world settings of our model. These results show the challenge faced by the original MaskDINO [20] in capturing detailed features in scenes with dense objects and the effectiveness of the pixel shuffle of our model in capturing detailed features. Table 4 further demonstrates the superiority of our model in object detection over previous methods and MaskDINO [20]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Gaze estimation. Tables 5 and 6 show the gaze estimation performance on GOO-Synth and GOO-Real, respectively. In Table 5, our method outperforms GTR [39] and TransGOP [40] by \\(23.9\\mathrm{mAP}\\) and \\(10.6\\mathrm{mAP}\\), respectively. Even directly evaluating all prediction results in the real world, our model reduces L2 distance and angle errors compared to TransGOP [40] by 0.015 and 1.8, respectively. Applying non-real-world methods to the real world significantly degrades performance, highlighting our real-world model is more efficient. Table 6 verifies our method on GOO-Real, achieving state-of-the-art results without pretraining. Even with direct inference in a real-world setting, our method outperforms previous non-real-world methods. After pre-training, our model exhibits more pronounced performance advantages. These results show the effectiveness" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "Y. Jin et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.488, + 0.189 + ], + "angle": 0, + "content": "Table 5: Gaze estimation results on GOO-Synth dataset. \\(\\dagger\\) indicates the evaluation method proposed by [39]." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.199, + 0.489, + 0.415 + ], + "angle": 0, + "content": "
MethodAUCDist.↓Ang.↓mAP
Non-RealGazeFollow [29]0.9290.16233.0-
Lian [22]0.9540.10719.7-
VideoAttention [9]0.9520.07515.1-
GaTector [41]0.9570.07314.9-
GTR [39]0.9600.07114.5-
TransGOP [40]0.9630.07913.3-
Ours0.9470.07214.8-
RealGazeFollow [29]†0.8320.31742.60.468
Lian [22]†0.9030.15328.10.454
VideoAttention [9]†0.9120.14324.50.489
GaTector [41]†0.9180.13924.50.510
GTR [39]†0.9620.06814.20.597
TransGOP [40]†0.9770.07011.50.730
TransGOP [40]0.9450.10620.7-
Ours†0.9780.05710.20.836
Ours0.9380.09118.9-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.501, + 0.145, + 0.779, + 0.188 + ], + "angle": 0, + "content": "Table 6: Gaze estimation results on GOO-Real dataset. \\(\\dagger\\) indicates the evaluation method proposed by [39]." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.198, + 0.78, + 0.415 + ], + "angle": 0, + "content": "
MethodAUCDist.↓Ang.↓mAP
No PretrainNon-RealGazeFollow [29]0.8500.22044.4-
Lian [22]0.8400.32143.5-
VideoAttention [9]0.7960.25251.4-
GaTector [41]0.9270.19639.5-
Tonini [36]0.9180.164--
TransGOP [40]0.9470.09716.7-
Ours0.9430.07812.9-
RealOurs†0.9710.0599.80.810
Ours0.9440.08814.7-
PretrainNon-RealGazeFollow [29]0.9030.19539.8-
Lian [22]0.8900.16832.6-
VideoAttention [9]0.8890.15029.1-
GaTector [41]0.9400.08714.8-
TransGOP [40]0.9570.08114.7-
Ours0.9630.07312.4-
RealOurs†0.9780.0518.80.824
Ours0.9490.08213.8-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.291, + 0.428, + 0.713, + 0.442 + ], + "angle": 0, + "content": "Table 7: Ablation study about each component on GOO-Real" + }, + { + "type": "table", + "bbox": [ + 0.247, + 0.455, + 0.756, + 0.558 + ], + "angle": 0, + "content": "
PS RoI-Rec. DAF FIM EALGaze Object SegmentationInstance SegmentationGaze Estimation
mSoCmSoC50mSoC75APAP50AP75AUCDist.↓Ang.↓
79.0098.4892.5074.2498.0087.620.8540.30454.2
✓84.5098.7596.5680.5398.6893.670.8390.33459.8
✓✓84.6398.6796.7281.7098.8394.660.9100.11420.5
✓✓✓85.4498.9296.9681.5898.7694.290.9190.09516.8
✓✓✓✓85.8998.8097.1482.4598.9294.820.9320.08914.9
✓✓✓✓✓85.9498.9997.2082.8699.0295.570.9440.08814.7
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.788, + 0.618 + ], + "angle": 0, + "content": "of our space-to-object gaze regression method in modeling human-object gaze relationships while achieving more accurate gaze heatmap regression." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.636, + 0.55, + 0.652 + ], + "angle": 0, + "content": "4.3 Ablation study and model analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Ablation study about the effect of each component. In Table 7, we conduct various ablation studies on the GOO-Real dataset. We first establish a baseline using MaskDINO [20] and a gaze regression branch similar to TransGOP [40], and then gradually add the proposed modules and demonstrate their contribution to overall model performance. (i) Pixel shuffle (PS). After adding the pixel shuffle module, instance segmentation performance improved by \\(6.29\\%\\) AP \\((80.53\\%)\\) vs \\(74.24\\%\\) compared to baseline, indicating that large feature maps significantly enhance detailed features in dense prediction tasks. (ii) RoI Reconstruction (RoI-Rec.). The RoI reconstruction module reduces angle error by 39.3 and L2 distance by 0.22, demonstrating the importance of head-related features for gaze direction perception and the ability of the module to learn accurate gaze features in real-world settings. (iii) Dual attention fusion (DAF)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Boosting Gaze Object Prediction via Pixel-level Supervision" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.317, + 0.145, + 0.684, + 0.159 + ], + "angle": 0, + "content": "Table 8: Analysis about dual attention fusion module." + }, + { + "type": "table", + "bbox": [ + 0.246, + 0.171, + 0.756, + 0.243 + ], + "angle": 0, + "content": "
Original Attention Gaze ConeGaze Object SegmentationInstance SegmentationGaze Estimation
mSoCmSoC50mSoC75APAP50AP75AUCDist.↓Ang.↓
✓85.5198.8697.1782.0898.8794.780.9380.08915.6
✓84.1398.7896.3581.3098.7893.690.9260.09716.3
✓✓85.9498.9997.2082.7599.0195.170.9440.08814.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.327, + 0.256, + 0.677, + 0.27 + ], + "angle": 0, + "content": "Table 9: Analysis about feature interaction module" + }, + { + "type": "table", + "bbox": [ + 0.246, + 0.282, + 0.756, + 0.356 + ], + "angle": 0, + "content": "
Encoder QueryDecoder QueryMask EmbeddingGaze Object SegmentationInstance SegmentationGaze Estimation
mSoCmSoC50mSoC75APAP50AP75AUC Dist.↓Ang.↓
✓85.7998.3896.4481.2898.8794.870.9280.095 16.0
✓85.2098.2197.0181.6598.8694.650.9490.093 15.1
✓85.9498.9997.2082.8699.0295.570.9440.088 14.7
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.785, + 0.534 + ], + "angle": 0, + "content": "Introducing the dual attention fusion improves gaze heatmap quality (AUC: 0.919 vs 0.910), demonstrating that perceiving gaze object spatial location beforehand produces more accurate heatmaps and eases regression. (iv) Feature interaction module (FIM). The feature interaction module improves gaze estimation, reducing angle error by 1.9 and L2 distance by 0.006. Gaze object segmentation achieves an \\(85.89\\mathrm{mSoC}\\), showing that pixel-level object information enhances feature interaction and human-object gaze relationship modeling. (v) Energy aggregation loss (EAL). Adding energy aggregation loss improves the performance of all three sub-tasks, showing that the gaze object mask effectively guides the regression of the gaze heatmap towards the gaze object." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.785, + 0.625 + ], + "angle": 0, + "content": "Analysis about dual attention fusion module. Table 8 analyzed the effect of different spatial perception methods. Compared to using original attention \\(\\mathbf{M}_{\\mathrm{ori}}\\) or gaze cone \\(\\mathbf{M}_{\\mathrm{cone}}\\) alone, our dual attention fusion performs better. This indicates that combining spatial search along the gaze direction with human head-based scene saliency is more effective. Fig. 4(b) shows that element-wise fusion of \\(\\mathbf{M}_{\\mathrm{ori}}\\) and \\(\\mathbf{M}_{\\mathrm{cone}}\\) creates an initial human-object spatial connection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.787, + 0.747 + ], + "angle": 0, + "content": "Analysis about feature interaction module. In Table 9, we evaluate the efficacy of different object information for gaze regression. Injecting the global encoder query of MaskDINO [20] results in \\(85.79\\%\\) mSoC for gaze object segmentation but poor gaze estimation performance. Refining the decoder query with mask supervision reduces the angle error by 0.9, indicating better object representation. Using mask embedding with mask features during gaze regression achieves optimal performance, demonstrating that semantically clear mask features enhance gaze object positioning and the human-object gaze relationship." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.769, + 0.427, + 0.784 + ], + "angle": 0, + "content": "4.4 Qualitative Results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Fig. 6(a) presents qualitative results in the real-world setting. In object-dense scenes, our method accurately predicts the gaze object mask without relying on head-related cues. In Fig. 6(b), compared to previous methods, our method ac" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "Y. Jin et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.402, + 0.788, + 0.434 + ], + "angle": 0, + "content": "Fig. 6: Qualitative comparison of ground truth gaze object (GT), gaze object detection (GOD), instance segmentation (IS), gaze object segmentation (GOS)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.494 + ], + "angle": 0, + "content": "curately locates adjacent objects using a pixel-level mask, effectively alleviating semantic ambiguity. See supplementary material for more visualization." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.518, + 0.36, + 0.535 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.779 + ], + "angle": 0, + "content": "In this paper, we introduce the gaze object segmentation task, which aims to infer pixel-level masks for objects captured by human eyesight, and establish a framework based on the pixel-level supervision from VFM to solve this task while alleviating the semantic ambiguity existing in previous box-level supervision methods. In our method, we choose SAM to generate pixel-level supervision information for the model training. The RoI reconstruction module is proposed to reconstruct head features directly from holistic features using the head box provided by the unified Transformer, which improves the inference efficiency and flexibility of our model in the real world. To boost the representation of the human-object gaze relationship, we propose a space-to-object gaze regression strategy, which first uses the dual attention fusion module to establish a human-object spatial connection, and then a semantically clear mask feature is injected into the gaze regression to strengthen the human-object gaze relationship modeling. Extensive experiments on GOO-Real and GOO-Synth demonstrate the effectiveness of our method in various settings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.841 + ], + "angle": 0, + "content": "**Limitation:** The proposed method requires box supervision as prompts for the SAM when generating masks, which limits its application in a wider range of scenarios without strong location priors, more flexible approaches for generating masks via the latest VFM models need to be explored in future work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Boosting Gaze Object Prediction via Pixel-level Supervision" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.174, + 0.784, + 0.204 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China. (No. 82272020)." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.205, + 0.785, + 0.234 + ], + "angle": 0, + "content": "Bin Fan is sponsored by China National Postdoctoral Program for Innovative Talents (No. BX20230013)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.255, + 0.323, + 0.272 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.283, + 0.787, + 0.311 + ], + "angle": 0, + "content": "1. Bao, J., Liu, B., Yu, J.: ESCNet: Gaze target detection with the understanding of 3D scenes. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 14126-14135 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.787, + 0.352 + ], + "angle": 0, + "content": "2. Cai, X., Zeng, J., Shan, S., Chen, X.: Source-free adaptive gaze estimation by uncertainty reduction. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 22035-22045 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.352, + 0.787, + 0.392 + ], + "angle": 0, + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Eur. Conf. Comput. Vis. pp. 213-229 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.392, + 0.787, + 0.433 + ], + "angle": 0, + "content": "4. Chen, W., Li, R., Yu, Q., Xu, A., Feng, Y., Wang, R., Zhao, L., Lin, Z., Yang, Y., Lin, D., et al.: Early detection of visual impairment in young children using a smartphone-based deep learning system. Nature Medicine 29(2), 493-503 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.787, + 0.461 + ], + "angle": 0, + "content": "5. Chen, Y., Nan, Z., Xiang, T.: FBLNet: Feedback loop network for driver attention prediction. In: Int. Conf. Comput. Vis. pp. 13371-13380 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.787, + 0.488 + ], + "angle": 0, + "content": "6. Cheng, Y., Lu, F.: Gaze estimation using transformer. In: Int. Conf. Pattern Recog. pp. 3341-3347 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.488, + 0.787, + 0.514 + ], + "angle": 0, + "content": "7. Cheng, Y., Lu, F., Zhang, X.: Appearance-based gaze estimation via evaluation-guided asymmetric regression. In: Eur. Conf. Comput. Vis. pp. 100-115 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.514, + 0.787, + 0.556 + ], + "angle": 0, + "content": "8. Chong, E., Ruiz, N., Wang, Y., Zhang, Y., Rozga, A., Rehg, J.M.: Connecting gaze, scene, and attention: Generalized attention estimation via joint modeling of gaze and scene saliency. In: Eur. Conf. Comput. Vis. pp. 383-398 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.556, + 0.787, + 0.582 + ], + "angle": 0, + "content": "9. Chong, E., Wang, Y., Ruiz, N., Rehg, J.M.: Detecting attended visual targets in video. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5396-5406 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.787, + 0.623 + ], + "angle": 0, + "content": "0. Gupta, A., Tafasca, S., Odobez, J.M.: A modular multimodal architecture for gaze target prediction: Application to privacy-sensitive settings. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5041-5050 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.623, + 0.787, + 0.65 + ], + "angle": 0, + "content": "1. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask R-CNN. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2961-2969 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.65, + 0.787, + 0.69 + ], + "angle": 0, + "content": "2. Hu, Z., Yang, D., Cheng, S., Zhou, L., Wu, S., Liu, J.: We know where they are looking at from the RGB-D camera: Gaze following in 3D. IEEE Trans. Instrum. Meas. 71, 1-14 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.787, + 0.732 + ], + "angle": 0, + "content": "3. Hu, Z., Yang, Y., Zhai, X., Yang, D., Zhou, B., Liu, J.: GFIE: A dataset and baseline for gaze-following from 2D to 3D in indoor environments. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8907-8916 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.787, + 0.772 + ], + "angle": 0, + "content": "4. Hu, Z., Zhao, K., Zhou, B., Guo, H., Wu, S., Yang, Y., Liu, J.: Gaze target estimation inspired by interactive attention. IEEE Trans. Circuit Syst. Video Technol. 32(12), 8524-8536 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "5. Huang, T., Fu, R.: Driver distraction detection based on the true driver's focus of attention. IEEE Trans. Intell. Transp. Syst. 23(10), 19374-19386 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "6. Jin, S., Wang, Z., Wang, L., Bi, N., Nguyen, T.: ReDirTrans: Latent-to-latent translation for gaze and head redirection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5547-5556 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.283, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.354, + 0.127 + ], + "angle": 0, + "content": "Y. Jin et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "17. Jin, T., Yu, Q., Zhu, S., Lin, Z., Ren, J., Zhou, Y., Song, W.: Depth-aware gaze-following via auxiliary networks for robotics. Eng. Appl. Artif. Intell. 113, 104924 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.23 + ], + "angle": 0, + "content": "18. Kellnhofer, P., Recasens, A., Stent, S., Matusik, W., Torralba, A.: Gaze360: Physically unconstrained gaze estimation in the wild. In: Int. Conf. Comput. Vis. pp. 6912-6921 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.787, + 0.272 + ], + "angle": 0, + "content": "19. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Int. Conf. Comput. Vis. pp. 4015-4026 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.787, + 0.313 + ], + "angle": 0, + "content": "20. Li, F., Zhang, H., Xu, H., Liu, S., Zhang, L., Ni, L.M., Shum, H.Y.: Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3041-3050 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.312, + 0.787, + 0.34 + ], + "angle": 0, + "content": "21. Li, Y., Shen, W., Gao, Z., Zhu, Y., Zhai, G., Guo, G.: Looking here or there? gaze following in 360-degree images. In: Int. Conf. Comput. Vis. pp. 3742-3751 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.787, + 0.367 + ], + "angle": 0, + "content": "22. Lian, D., Yu, Z., Gao, S.: Believe it or not, we know what you are looking at! In: Asian Conf. Comput. Vis. pp. 35-50 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.787, + 0.393 + ], + "angle": 0, + "content": "23. Lv, K., Sheng, H., Xiong, Z., Li, W., Zheng, L.: Improving driver gaze prediction with reinforced attention. IEEE Trans. Multimedia 23, 4198-4207 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.392, + 0.787, + 0.42 + ], + "angle": 0, + "content": "24. Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature Communications 15(1), 654 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.787, + 0.447 + ], + "angle": 0, + "content": "25. Miao, Q., Hoai, M., Samaras, D.: Patch-level gaze distribution prediction for gaze following. In: IEEE Winter Conf. Appl. Comput. Vis. pp. 880-889 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.787, + 0.488 + ], + "angle": 0, + "content": "26. Mundy, P., Sigman, M., Kasari, C.: A longitudinal study of joint attention and language development in autistic children. J. Autism Dev. Disord. 20(1), 115-128 (1990)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.487, + 0.787, + 0.515 + ], + "angle": 0, + "content": "27. Park, S., Spurr, A., Hilliges, O.: Deep pictorial gaze estimation. In: Eur. Conf. Comput. Vis. pp. 721-738 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.787, + 0.556 + ], + "angle": 0, + "content": "28. Park, S., Zhang, X., Bulling, A., Hilliges, O.: Learning to find eye region landmarks for remote gaze estimation in unconstrained settings. In: Proc. ACM Symp. Eye Track. Res. Appl. pp. 1-10 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.787, + 0.582 + ], + "angle": 0, + "content": "29. Recasens, A., Khosla, A., Vondrick, C., Torralba, A.: Where are they looking? Adv. Neural Inform. Process. Syst. 28 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.787, + 0.609 + ], + "angle": 0, + "content": "30. Recasens, A., Vondrick, C., Khosla, A., Torralba, A.: Following gaze in video. In: Int. Conf. Comput. Vis. pp. 1435-1443 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.608, + 0.787, + 0.651 + ], + "angle": 0, + "content": "31. Ruzzi, A., Shi, X., Wang, X., Li, G., De Mello, S., Chang, H.J., Zhang, X., Hilliges, O.: GazeNeRF: 3d-aware gaze redirection with neural radiance fields. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 9676-9685 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.787, + 0.677 + ], + "angle": 0, + "content": "32. Senju, A., Johnson, M.H.: Atypical eye contact in autism: Models, mechanisms and development. Neurosci. Biobehav. Rev. 33(8), 1204-1214 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.676, + 0.787, + 0.733 + ], + "angle": 0, + "content": "33. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient subpixel convolutional neural network. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1874-1883 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.787, + 0.772 + ], + "angle": 0, + "content": "34. Tafasca, S., Gupta, A., Odobez, J.M.: ChildPlay: A new benchmark for understanding children's gaze behaviour. In: Int. Conf. Comput. Vis. pp. 20935-20946 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.813 + ], + "angle": 0, + "content": "35. Tomas, H., Reyes, M., Dionido, R., Ty, M., Mirando, J., Casimiro, J., Atienza, R., Quinto, R.: Goo: A dataset for gaze object prediction in retail environments. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. pp. 3125-3133 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.787, + 0.841 + ], + "angle": 0, + "content": "36. Tonini, F., Beyer, C., Ricci, E.: Multimodal across domains gaze target detection. In: Int. Conf. Multimodal Interact. pp. 420-431 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Boosting Gaze Object Prediction via Pixel-level Supervision" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "37. Tonini, F., Dall'Asen, N., Bayan, C., Ricci, E.: Object-aware gaze target detection. In: Int. Conf. Comput. Vis. pp. 21860-21869 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.217 + ], + "angle": 0, + "content": "38. Tu, D., Min, X., Duan, H., Guo, G., Zhai, G., Shen, W.: End-to-end human-gazetarget detection with transformers. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2192-2200 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.217, + 0.788, + 0.259 + ], + "angle": 0, + "content": "39. Tu, D., Shen, W., Sun, W., Min, X., Zhai, G., Chen, C.: Un-Gaze: A unified transformer for joint gaze-location and gaze-object detection. IEEE Trans. Circuit Syst. Video Technol. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.286 + ], + "angle": 0, + "content": "40. Wang, B., Guo, C., Jin, Y., Xia, H., Liu, N.: TransGOP: Transformer-based gaze object prediction. AAAI Conf. Artif. Intell. (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.327 + ], + "angle": 0, + "content": "41. Wang, B., Hu, T., Li, B., Chen, X., Zhang, Z.: GaTector: A unified framework for gaze object prediction. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19588-19597 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.327, + 0.788, + 0.356 + ], + "angle": 0, + "content": "42. Wang, K., Ji, Q.: Real time eye gaze tracking with 3D deformable eye-face model. In: Int. Conf. Comput. Vis. pp. 1003-1011 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.398 + ], + "angle": 0, + "content": "43. Wang, X., Zhang, H., Wang, Z., Nie, W., Yang, Z., Ren, W., Xu, Q., Xu, X., Liu, H.: Dual regression-enhanced gaze target detection in the wild. IEEE Trans. Cybern. 54(1), 219-229 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.439 + ], + "angle": 0, + "content": "44. Wang, Z., Zhao, J., Lu, C., Yang, F., Huang, H., Guo, Y., et al.: Learning to detect head movement in unconstrained remote gaze estimation in the wild. In: IEEE Winter Conf. Appl. Comput. Vis. pp. 3443-3452 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.481 + ], + "angle": 0, + "content": "45. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. Int. Conf. Learn. Represent. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.508 + ], + "angle": 0, + "content": "46. Zhang, X., Sugano, Y., Fritz, M., Bulling, A.: Appearance-based gaze estimation in the wild. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4511-4520 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.55 + ], + "angle": 0, + "content": "47. Zhang, X., Sugano, Y., Fritz, M., Bulling, A.: Mpiigaze: Real-world dataset and deep appearance-based gaze estimation. IEEE Trans. Pattern Anal. Mach. Intell. 41(1), 162-175 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.577 + ], + "angle": 0, + "content": "48. Zhao, X., Ding, W., An, Y., Du, Y., Yu, T., Li, M., Tang, M., Wang, J.: Fast segment anything. arXiv preprint arXiv:2306.12156 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.577, + 0.788, + 0.605 + ], + "angle": 0, + "content": "49. Zhu, W., Deng, H.: Monocular free-head 3D gaze tracking with deep learning and geometry constraints. In: Int. Conf. Comput. Vis. pp. 3143-3152 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.605, + 0.788, + 0.647 + ], + "angle": 0, + "content": "50. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: Int. Conf. Learn. Represent. (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.647 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Boosting Gaze Object Prediction via Pixel-level Supervision from Vision Foundation Model/93449abc-6022-4367-830a-8808cd9b5ba3_origin.pdf b/2024/Boosting Gaze Object Prediction via Pixel-level Supervision from Vision Foundation Model/93449abc-6022-4367-830a-8808cd9b5ba3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..834d44fa1cacd881e4928a931c6479da6b854a88 --- /dev/null +++ b/2024/Boosting Gaze Object Prediction via Pixel-level Supervision from Vision Foundation Model/93449abc-6022-4367-830a-8808cd9b5ba3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e727f7d7b4b88ad99c9d99e68fc59117a0b189b9a2c8c64122f4857520d83687 +size 10529872 diff --git a/2024/Boosting Gaze Object Prediction via Pixel-level Supervision from Vision Foundation Model/full.md b/2024/Boosting Gaze Object Prediction via Pixel-level Supervision from Vision Foundation Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..aea6b6e48affcc836c2ae4837a373c814eca463b --- /dev/null +++ b/2024/Boosting Gaze Object Prediction via Pixel-level Supervision from Vision Foundation Model/full.md @@ -0,0 +1,313 @@ +# Boosting Gaze Object Prediction via Pixel-level Supervision from Vision Foundation Model + +Yang Jin1, Lei Zhang2, Shi Yan2, Bin Fan3, and Binglu Wang1 + +$^{1}$ Xi'an University of Architecture and Technology, Xi'an, China {jin91999, wbl921129}@gmail.com † Corresponding author + +$^{2}$ School of Automation, Northwestern Polytechnical University, Xi'an, China +zl_hnly@163.com, yanshi@mail.nwpu.edu.cn + +$^{3}$ National Key Lab of General AI, School of Intelligence Science and Technology, Peking University, Beijing, China binfan@pku.edu.cn + +Abstract. Gaze object prediction (GOP) aims to predict the category and location of the object that a human is looking at. Previous methods utilized box-level supervision to identify the object that a person is looking at, but struggled with semantic ambiguity, i.e., a single box may contain several items since objects are close together. The Vision foundation model (VFM) has improved in object segmentation using box prompts, which can reduce confusion by more precisely locating objects, offering advantages for fine-grained prediction of gaze objects. This paper presents a more challenging gaze object segmentation (GOS) task, which involves inferring the pixel-level mask corresponding to the object captured by human gaze behavior. In particular, we propose that the pixel-level supervision provided by VFM can be integrated into gaze object prediction to mitigate semantic ambiguity. This leads to our gaze object detection and segmentation framework that enables accurate pixel-level predictions. Different from previous methods that require additional head input or ignore head features, we propose to automatically obtain head features from scene features to ensure the model's inference efficiency and flexibility in the real world. Moreover, rather than directly fuse features to predict gaze heatmap as in existing methods, which may overlook spatial location and subtle details of the object, we develop a space-to-object gaze regression method to facilitate human-object gaze interaction. Specifically, it first constructs an initial human-object spatial connection, then refines this connection by interacting with semantically clear features in the segmentation branch, ultimately predicting a gaze heatmap for precise localization. Extensive experiments on GOO-Synth and GOO-Real datasets demonstrate the effectiveness of our method. The code will be available at https://github.com/jinyang06/SamGOP. + +Keywords: Gaze object prediction $\cdot$ Vision foundation model $\cdot$ Object segmentation $\cdot$ Space-to-object gaze regression + +![](images/494d3ca1c7ce9f1f6a9dcf3a12974465e1826e94bdf96377c54067106ffe7cb6.jpg) +(a) Illustration of semantic ambiguity and semantic clarity +Fig. 1: (a) Box-level supervision often fails to localize objects in dense settings precisely and leads to semantic ambiguity problems, whereas pixel-level supervision excels by providing clear semantic distinction through pixel-by-pixel predictions. (b) Vision foundation models can produce instance masks, thereby segmentation features can be used to improve the gaze regression branch's spatial perception, and the gaze object mask can help the heatmap focus on the gaze object. + +![](images/97b8cf9b4a6ee2cbbb92b1087e24d231e9bdf2d10ffc3f2dbfb1f1020fd5c394.jpg) +(b) The main idea of our pipeline + +# 1 Introduction + +The objects stared at by humans contain important semantic information, which can reveal a person's behavior and state of mind. Therefore, identifying the objects that people are looking at has a wide range of applications. For example, in medical diagnosis, whether a child focuses on an object can potentially reveal whether he/she has autism or visual impairment [4,26,32,34]. In assisted driving scenarios, determining what object the driver is looking at can assess their attention and provide corresponding driving guidance [5,15,23]. In general, detecting the object gazed at by humans realizes the interaction between human sight and specific objects, fostering a richer semantic understanding of the scene. + +Tomas et al. [35] first proposed the gaze object detection task and the benchmark datasets GOO-Synth and GOO-Real, aiming to predict the object boxes and categories stared at by humans. Recently, Wang et al. [40, 41] conducted an in-depth exploration of this task. Despite achieving promising results, they still suffer from the following limitations: 1) They typically rely on box-level supervision, which is prone to suffer semantic ambiguity in dense object scenes. Especially when multiple objects are densely stacked or placed adjacent to each other, the box-level supervision algorithm may struggle to accurately separate them, as illustrated in Fig. 1(a). 2) The GOO [35] dataset only contains pixel-level annotations of synthetic images, leading to difficulties in generalizing to fine-grained predictions in real scenes. 3) They usually require additional head information as input, affecting the efficiency and flexibility of the model. While global modeling using the Transformer [38, 39] can alleviate this problem, it lacks special attention to head features, which are crucial for gaze direction perception. 4) They directly fuse features to regress the gaze heatmap, lacking a full understanding of the spatial location and subtle detailed information of objects. + +Recently, vision foundational models (VFM) have gained significant attention due to their powerful generalization capabilities, which can achieve accurate pixel-level segmentation using a variety of prompts, i.e., boxes, points, or masks. + +In real scenes with small and densely packed objects, VFM can generate accurate boundaries and pixel-level masks (See Fig 3) by utilizing boxes as prompt, which reduces semantic ambiguity and thus opens opportunities for gaze object segmentation. Motivated by this, we propose a more challenging task, gaze object segmentation (GOS). This task involves inferring the pixel-level mask of objects captured by human gaze behavior, helping to further enhance the understanding of human gaze behavior. Furthermore, we establish a GOS framework specifically for this task, which integrates the pixel-level supervision provided by VFM into gaze object prediction to alleviate the aforementioned limitations of previous methods [40, 41]. + +In the proposed model, we introduce the unified detection and segmentation Transformer, i.e. MaskDINO [20], to simultaneously obtain candidate gaze object boxes, segmentation masks, and head boxes of the subject, which are indispensable for enabling the model to infer in the real world and achieve accurate detection. Instead of requiring additional head input [40, 41] or ignoring head features by global modeling [38, 39], we propose a RoI reconstruction module to automatically obtain head region features based on holistic features and head box provided by the unified Transformer. This helps capture gaze direction information during gaze regression to improve model efficiency and flexibility. Subsequently, a space-to-object gaze regression strategy is developed to enhance the perception of the spatial location of objects and the understanding of detailed information. Specifically, inspired by the ability of humans to infer others' gaze objects, a dual attention fusion module is presented to establish an initial human-object gaze spatial connection. Then, we propose an object feature interaction module to refine the gaze spatial connection through semantically clear feature interaction between gaze regression features and object mask features, thus strengthening the modeling of the human-object gaze relationship. In the end, an energy aggregation loss is used to further guide the regression of the gaze heatmap. By focusing on pixel-level details and progressively improving the gaze representation, it further reduces semantic ambiguity, leading to the generation of more precise heatmaps for gazed objects. + +Our main contributions could be summarized as follows: + +- We introduce the vision foundation model into gaze object prediction to provide pixel-level supervision and propose a more challenging gaze object segmentation task, which focuses on identifying the pixel-level masks of objects captured by human gaze behavior. +- We present an all-in-one end-to-end framework for gaze object detection and segmentation without relying on additional head-related inputs, capable of simultaneously addressing gaze estimation, object detection, and segmentation. To the best of our knowledge, our work is the first to tackle gaze object prediction from the image segmentation perspective. +- We propose a space-to-object gaze regression approach that enhances the interaction between object and gaze branches, incrementally improving the modeling of the human-object gaze relationship and alleviating semantic ambiguity. + +# 2 Related Works + +Vision Foundation Model. The vision foundation models (VFMs) [19, 24, 48] are typically built on large-scale datasets using self-supervised or semi-supervised methods, which have strong generalizability and zero-shot transfer capabilities. Segment Anything Model (SAM) [19] received attention as a foundation model for image segmentation. Its portability enables it to be flexibly applied to various downstream tasks. Then, the medical foundation model MedSAM [24] extends the foundation model to medical diagnosis. In this article, we choose SAM [19] to obtain high-quality masks to boost gaze estimation and gaze object prediction. + +Gaze Object Detection. The gaze object detection is first proposed in [35], which aims to predict the box and categories of the object stared at by humans in an RGB image. Tomas et al. [35] released a benchmark dataset for GOP. Wang et al. [41] first proposed a GOP method named GaTector. However, GaTector requires additional head-related prior information, which hinders its application in the real world. Recently, the Transformer-based method [40] achieves comparable performance. Although previous methods are effective to some extent, they can only provide instance-level box outputs, leading to potential ambiguities or misjudgments in scenarios with small and densely packed objects. + +Gaze Target Detection. Early research on gaze focused on gaze direction estimation [2, 5-7, 16, 18, 27, 28, 31, 42, 44, 46, 47, 49], which only predict the gaze vector and ignores the impact of the surrounding environment on the sight, resulting in limited application scenarios. Therefore, to capture higher-level gaze-related semantic information, researchers propose gaze target detection [1, 8-10, 12-14, 17, 21, 22, 25, 30, 34, 36, 37, 43], which involves to infer the specific gaze point. Recasens et al. [29] introduced the GazeFollow dataset and first proposed gaze target detection. Recently, Tu et al. [38,39] achieved end-to-end gaze target detection using Transformer, with a relatively low detection efficiency. Despite advancements in gaze target estimation, previous methods struggle with gaze regression due to spatial perception ambiguity. In contrast, our space-to-object regression strategy helps the model gradually converge to a certain gaze point, simplifying learning and enhancing detailed interaction with the object. + +# 3 Method + +Given an RGB image, our model aims to precisely predict the pixel-level mask of the object humans are looking at. We first present the overall structure of our method, then introduce the VFM-based mask supervision generation method, and finally describe each module of the proposed model in detail. + +# 3.1 Overview + +As shown in Fig. 2, our method consists of four components: a feature extractor for specific feature extraction, a detection and segmentation module predicting candidate gaze objects boxes, masks, and human head location, a gaze regressor + +![](images/8f1529fbd145da8de17a7eb09b2daee56dd8fcd3712c58baa4cc8dd28a46ada4.jpg) +Fig. 2: Overview of the proposed model. (a) The feature extraction module extracts features for detection and regression branches. (b) The detection and segmentation branch identifies object and human head positions. (c) After obtaining head features, the gaze regression branch progressively refines its output: 1) employing a dual attention fusion module for initial human-object correlations; 2) leveraging a feature interaction module to incorporate semantically clear object-aware insights from the segmentation branch; 3) ultimately predicting the gaze heatmap. (d) Supervision signals are applied to both branches only during training. + +producing gaze object heatmap with a space-to-object strategy, and a supervision information generation module creating pixel-level masks and ground truth heatmap during training. At training, the overall loss function is defined as: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {d e t}} + \alpha \mathcal {L} _ {\mathrm {d i r}} + \beta \mathcal {L} _ {\mathrm {g a z e}} + \gamma \mathcal {L} _ {\mathrm {e n g}} \tag {1} +$$ + +where $\alpha$ , $\beta$ , and $\gamma$ are the weights of the gaze direction loss, gaze heatmap loss, and energy aggregation loss respectively. Before model training, we choose SAM to generate pixel-level masks for the objects. See Sec. 3.2 for details. + +Feature extraction. As shown in 2(a), given an RGB image $\mathcal{I} \in \mathbb{R}^{3 \times H \times W}$ , we first use the backbone $\mathcal{B}$ to extract multi-scale features $\{\mathbf{f}_1, \mathbf{f}_2, \mathbf{f}_3, \mathbf{f}_4\}$ . In previous work [41], pixel shuffle [33] significantly improved small object detection by enlarging feature maps for detailed feature capture. Therefore, we integrate it into dense prediction tasks, i.e., image segmentation. In our model, the multi-scale features $\{\mathbf{f}_1, \mathbf{f}_2, \mathbf{f}_3, \mathbf{f}_4\}$ are fed into the pixel shuffle module to generate object-specific features $\mathbf{f}_i^{obj} = \phi^{obj}(\mathbf{f}_i, \eta)$ , where $\mathbf{f}_i$ denotes the $i$ -th feature layer, $\eta$ denotes the scale factor, $\phi^{obj}(\cdot)$ denotes the pixel shuffle module, which consists of a pixel shuffle operation and a convolution layer that transforms the feature channels to be consistent with the original features of the unified Transformer [20]. Finally, the scene saliency features $\mathbf{f}_{\mathrm{scene}} \in \mathbb{R}^{1024 \times 7 \times 7}$ are extracted by a scene residual block using high dimensional features $\mathbf{f}_4 \in \mathbb{R}^{2048 \times 7 \times 7}$ . + +Unified detection and segmentation Transformer. In this paper, we employ a unified Transformer for candidate gaze object detection, segmentation, and human head positioning, integrating MaskDINO [20] for candidate gaze object detection, segmentation, and the generation of pixel-level mask feature embeddings and a head decoder for head box regression, which facilitate accurate head-related feature extraction in head feature reconstruction module and + +![](images/1ccb143e786d159e727b559a8131322e7e1b5183edccaf11d6264df7973d0178.jpg) +Fig. 3: Illustration of instance mask generated by VFM. + +accurate prediction in real-world settings. The training loss is defined as: + +$$ +\mathcal {L} _ {\mathrm {d e t}} = \mathcal {L} _ {\mathrm {d e t}} ^ {\mathrm {o b j}} + \mathcal {L} _ {\mathrm {m a s k}} ^ {\mathrm {o b j}} + \mathcal {L} _ {\mathrm {d e t}} ^ {\mathrm {h e a d}} \tag {2} +$$ + +$\mathcal{L}_{\mathrm{det}}^{\mathrm{obj}}$ and $\mathcal{L}_{\mathrm{mask}}^{\mathrm{obj}}$ is the object detection and segmentation loss respectively. $\mathcal{L}_{\mathrm{det}}^{\mathrm{head}}$ is the head decoder loss, which is consistent with the $\mathcal{L}_{\mathrm{det}}^{\mathrm{obj}}$ . See Sec. 3.3 for details. Space-to-object gaze regression. We employ a space-to-object regression strategy for gaze heatmap prediction. The module begins by representing human gaze behavior with a gaze vector. Subsequently, a dual attention fusion module establishes an initial human-object spatial connection. Finally, pixel-level object location knowledge from the mask branch is introduced to achieve semantically clear feature interaction, refining the gaze heatmap regression while alleviating semantic ambiguity. Optimization involves $\mathcal{L}_{\mathrm{dir}}$ and $\mathcal{L}_{\mathrm{gaze}}$ for gaze direction and gaze heatmap, respectively. The energy aggregation loss $\mathcal{L}_{\mathrm{eng}}$ guides the gaze heatmap to focus on the gaze object mask. See Sec. 3.4 for details. + +# 3.2 Instance mask generation via VFM. + +Given an RGB image $\mathcal{I}$ , we choose the vision foundation model SAM [19] to obtain instance masks for model training. Specifically, image $\mathcal{I}$ and prompt $\mathcal{P}_{box}$ are first fed into the image encoder $\varPsi_{Image}$ and prompt encoder $\varPsi_{prompt}$ , respectively. Then, the mask $\mathcal{P}_{mask}$ and confidence $\mathcal{S}$ are output by the mask decoder $\varPhi_{mask}$ , which can be formulated as: + +$$ +\mathcal {P} _ {\text {m a s k}}, \mathcal {S} = \Phi_ {\text {m a s k}} (\Psi_ {\text {I m a g e}} (\mathcal {I}), \Psi_ {\text {p r o m p t}} (\mathcal {P} _ {\text {b o x}})) \tag {3} +$$ + +Instance segmentation by using boxes produces accurate masks for most objects, but accurate segmentation of objects with unclear texture or edge features remains challenging. Therefore, we segment again using $\mathcal{P}_{mask}$ as a dense prompt to produce final mask $\mathcal{M}_{\mathrm{mask}}$ , which can be formulated as: + +$$ +\mathcal {M} _ {\text {m a s k}}, \mathcal {S} = \Phi_ {\text {m a s k}} (\Psi_ {\text {I m a g e}} (\mathcal {I}), \Psi_ {\text {p r o m p t}} (\mathcal {P} _ {\text {m a s k}})) \tag {4} +$$ + +Finally, $\mathcal{M}_{mask}$ is used to train the gaze object segmentation model. The generated masks are shown in Fig. 3. + +# 3.3 Unified detection and segmentation Transformer + +The unified detection and segmentation Transformer aims to detect candidate gaze objects and locate human heads while extracting high-level pixel mask + +features, which comprises a detection and segmentation Transformer, a head decoder, and a RoI feature reconstruction module. As shown in Fig. 2(b), the object-specific features $\{\mathbf{f}_1^{obj},\mathbf{f}_2^{obj},\mathbf{f}_3^{obj},\mathbf{f}_4^{obj}\}$ are fed it to detect object boxes, categories, masks, and human head boxes. + +Object detection and segmentation The global attention of Transformer ensures powerful long-distance modeling and global perception, achieving excellent detection performance through layer-by-layer self-attention refinement. We use MaskDINO [20] as a foundational detection and segmentation structure due to its robust representation capabilities and fast convergence speed. + +Head decoder Our model is designed for real-world gaze object detection and segmentation without relying on head priors. We introduce a head decoder based on the unified Transformer, using a 3-layer self-attention decoder to generate proposals for head-related feature reconstruction. During training, we filter head queries with IoU $>0.5$ as positives. During inference, boxes are selected based on confidence. Similar to other DETR-like detectors [45, 50], the head decoder uses cross-entropy, $L_{1}$ , and GIoU loss for classification and box regression. + +Head feature reconstruction Previous methods typically extract head features from manually cropped head images, limiting model inference efficiency and flexibility in the real world. To this end, we design a RoI feature reconstruction module to reconstruct head-related features based on holistic scene features and the head box from the unified Transformer. In this module, we first use the head bounding box $R = (x_{1},y_{1},x_{2},y_{2})$ , which represents the upper left and lower right corner coordinates, as the region proposal to crop the regions of interest (ROIs). Then, those candidate ROIs are refined by ROAlign [11], which maps the head region to the corresponding pixels in the holistic image features $\mathbf{f}_4 \in \mathbb{R}^{2048 \times 7 \times 7}$ and produces a fixed-size feature map $\mathbf{f}_{\mathrm{head}} \in \mathbb{R}^{2048 \times 7 \times 7}$ . For head location map generation, following previous works [9,10,29,41] to generate a binary image, setting the head area to 1 and the rest to 0. Additionally, a specific residual block extracts gaze saliency features $\mathbf{f}_{\mathrm{gaze}} \in \mathbb{R}^{1024 \times 7 \times 7}$ , further enhancing the model's attention to the gaze direction. + +# 3.4 Space-to-object gaze regression + +We propose a space-to-object gaze regression approach to gradually refine the human-object gaze relationship. Fig. 2(c) shows our gaze regression branch with a dual attention fusion module establishing an initial human-object spatial connection and a feature interaction module strengthening this connection using semantically clear feature interaction. + +Spatial perception for gaze objects. When humans infer what another person is staring at, they usually first observe the gaze direction of the others and search for the gazed object along this direction. Inspired by this, we propose the dual attention fusion module for the spatial perception of gaze objects, establishing an initial human-object spatial connection. + +As shown in Fig. 4(a), we first use global average pooling to obtain the global features of $\mathbf{f}_{\mathrm{head}}$ , followed by fully connected layers to transform the features into + +![](images/aee492adf4c07b0548b42361e9eb28b1f882d8fd7d163065ef5b5ad4e38dcb3a.jpg) +Fig. 4: Illustration of spatial perception for gaze objects. + +![](images/cdd85fdf43e9d108402b9f524e92bb64f9a24f8526c299371dbb66ebc2b77697.jpg) + +a 2D gaze vector $\mathbf{v}_g$ . To achieve spatial perception, we project it into a gaze cone map $\mathbf{M}_{\mathrm{cone}} \in \mathbb{R}^{1 \times 7 \times 7}$ , activating the area around the gaze direction: + +$$ +\mathbf {M} _ {\mathrm {i}, \mathrm {j}} ^ {\text {c o n e}} = M a x (0, \cos (\mathbf {v} _ {\mathrm {g}}, \mathbf {v} _ {\mathrm {i}, \mathrm {j}})) \tag {5} +$$ + +where $\mathbf{v}_{\mathrm{g}}$ and $\mathbf{v}_{\mathrm{i,j}}$ respectively denotes the predicted gaze vector and the vector composed of the eye point and each point on the gaze field. During training, $\mathbf{v}_{\mathrm{g}}$ is optimized by $\mathcal{L}_{\mathrm{dir}}$ to obtain a more accurate gaze direction and activation map. $\mathcal{L}_{\mathrm{dir}}$ is defined as follows: + +$$ +\mathcal {L} _ {\mathrm {d i r}} = 1 - \sum_ {i = 1} ^ {n} \left(\mathbf {v} _ {\mathrm {g}} ^ {i} \cdot \mathbf {v} _ {\mathrm {t}} ^ {i}\right) \tag {6} +$$ + +where vector $\mathbf{v}_{\mathrm{g}}$ and $\mathbf{v}_{\mathrm{t}}$ denote the predicted gaze direction and ground truth, respectively. During backpropagation, $\mathcal{L}_{\mathrm{dir}}$ adjusts parameters of head feature $\mathbf{f}_{\mathrm{head}}$ , providing more direction information for gaze regression. To boost scene attention, we adopt the method from GaTector [41] to generate an original attention map $\mathbf{M}_{\mathrm{ori}} \in \mathbb{R}^{1 \times 7 \times 7}$ , activating scene salience area based on head properties. By multiplying $\mathbf{M}_{\mathrm{cone}}$ with $\mathbf{M}_{\mathrm{ori}}$ via element-wise dot product, we generate the final spatial perception map $\mathbf{M}$ : + +$$ +\mathbf {M} = \mathbf {M} _ {\text {o r i}} \odot \mathbf {M} _ {\text {c o n e}} \tag {7} +$$ + +This compels the model to search spaces more likely gazed at by humans along gaze direction, establishing an initial human-object gaze spatial relationship. + +Semantically clear feature interaction. Previous methods [40, 41] usually directly fuse features to regress the gaze heatmap based on spatial saliency, lacking interaction with object details and facing challenges in accurately predicting heatmap for the gazed object. In our method, we introduce a feature interaction module, which leverages semantically clear mask knowledge from VFM to achieve feature interaction, refining human-object gaze relationship modeling. + +As shown in Fig. 2(c), we initially extract detailed mask features from the mask branch. Subsequently, the fused feature $\mathbf{f}_{\mathrm{fuse}}$ is utilized for pixel-level attention perception. Given that candidate gaze objects can be located anywhere in the image, the Transformer's global modeling excels in accomplishing this task. Therefore, utilize a transformer layer, akin to DETR [3], to capture the global relationship between human gaze behavior and object masks. For fine-grained + +object mask features, a three-layer MLP is applied to process the decoder query $\mathcal{Q}_{\mathrm{obj}}$ of the unified Transformer, generating mask embeddings: + +$$ +\mathcal {Q} _ {\text {m a s k}} = \mathbf {M L P} \left(\mathcal {Q} _ {\text {o b j}}\right) \tag {8} +$$ + +where the $\mathcal{Q}_{\mathrm{mask}}$ denote the generated mask embedding. We first apply a layer of self-attention for $\mathbf{f}_{\mathrm{fuse}}$ : + +$$ +\mathbf {f} _ {\mathrm {e}} = \mathbf {F F N} (\mathbf {S e l f A t t n} (\mathbf {f} _ {\mathrm {f u s e}})) \tag {9} +$$ + +where $\mathbf{f}_{\mathrm{fuse}}$ , SelfAttn( $\cdot$ ), and $\mathbf{f}_{\mathrm{e}}$ denote the fused feature $\mathbf{f}_{\mathrm{fuse}}$ , self-attention operation, and encoded $\mathbf{f}_{\mathrm{fuse}}$ , respectively. Then, the $\mathcal{Q}_{\mathrm{mask}}$ is perceived by the $\mathbf{f}_{\mathrm{e}}$ as key-value pairs, i.e., $\{\mathbf{K}_{\mathrm{mask}}, \mathbf{V}_{\mathrm{mask}}\}$ : + +$$ +\mathbf {f} _ {\text {r e g}} = \mathbf {F F N} (\mathbf {C r o s s A t t n} (\mathbf {f} _ {\mathrm {e}}, \{\mathbf {K} _ {\text {m a s k}}, \mathbf {V} _ {\text {m a s k}} \})) \tag {10} +$$ + +where $\mathbf{f}_{\mathrm{reg}}$ and CrossAttn $(q, kv)$ denote the regression feature and cross-attention operation, respectively. This operation refines the perception of the object by providing a pixel-level understanding of the gaze object in each token of the regression feature. It effectively alleviates semantic ambiguity and strengthens the ability of the model to represent the human-object gaze relationship. + +Ultimate gaze regression. Finally, following previous work [41] to generate the ground truth gaze heatmap and use $L_{2}$ loss to supervise gaze heatmap regression. $\mathcal{L}_{\mathrm{eng}}$ proposed by [40] guides gaze heatmap energy aggregation towards the gaze object. We use the gaze object mask for more accurate error measurements. Space-to-object gaze regression approach gradually refines the human-object gaze behavior representation, regressing a more accurate gaze heatmap. + +# 4 Experiments + +# 4.1 Experiments Settings + +Datasets and implementation details are given in supplementary material. Metrics: Following previous work [40,41], we use AUC (Area Under the Curve), L2 distance, and Angle error for gaze estimation evaluation, and mSoC (min Single over Closure)4 under different thresholds for gaze object prediction evaluation. Following MaskDINO [20], AP (Average Precision) is used for instance segmentation and object detection evaluation. To evaluate the gaze object segmentation, we use masks to replace boxes when calculating mSoC (details refer to the supplementary material). For a fair comparison, we reported results for both real-world and non-real-world settings. The non-real-world settings involve inputting additional head images and head location maps while real-world settings only use scene images. Furthermore, to compare with GTR [39], we also adopt the evaluation method proposed by [39]5. This method computes gaze estimation metrics only for results with L2 distance $< 0.15$ and head box IoU $> 0.5$ , which is an incomplete evaluation for AUC, L2 distance, and Angle error. + +Table 1: Gaze object detection results on GOO-Synth and GOO-Real datasets. * indicates using additional head detector to achieve real-world settings. + +
MethodGOO-SynthGOO-Real No PretrainGOO-Real PretrainParamsFPS
mSoC507595mSoC507595mSoC507595
Non-RealGaTector [41]67.9498.1486.250.1262.4095.1073.500.2071.2097.5088.802.7060.78M14.1
TransGOP [40]92.8099.0098.5051.9082.6097.8089.406.5089.0098.9097.5033.2094.03M13.7
Ours91.1993.9793.4673.2781.7598.7896.437.1381.8998.9996.637.2880.99M14.5
RealGaTector [41]*56.1988.0866.18-40.8782.0533.540.3066.4096.2280.490.4075.73M9.3
TransGOP [40]*91.6099.0098.1048.7079.7097.8091.4011.2083.2098.8095.3014.40106.37M11.6
Ours90.7793.9293.4569.5881.3598.9596.022.6181.6099.0396.546.0384.26M14.1
+ +Table 2: Ours gaze object segmentation results on GOO-Real and GOO-Synth. + +
SettingsGOO-SynthGOO-Real No PretrainGOO-Real PretrainParams|FPS
mSoC507595mSoC507595mSoC507595
Non-Real82.6893.7091.6826.8386.3998.9597.4420.0789.3998.8098.0233.0380.99M
Real82.7393.6791.7926.8585.9498.9997.2916.7086.9698.9497.7522.2184.26M
+ +![](images/6793ddd6a6bbc7d9cf83f6932f142bd4de3c417a0116edb813388a1f0217cd64.jpg) +Fig. 5: Comparison of TransGOP [40] and our method inference in the real world. + +![](images/639656f757280972a12e0fd517dee9c05df1976bbc2c10c95b56bcb3c05cbe5e.jpg) + +# 4.2 Comparison with SOTA + +Gaze object detection and segmentation. Table 1 shows the performance of gaze object detection on GOO-Synth and GOO-Real. In real-world settings, previous methods [40, 41] significantly increase parameters and slow down the inference speed due to the additional head detector (As shown in Fig. 5 (a)) and they can only achieve box-level prediction. In contrast, although our method is slightly inferior to TransGOP [40] in detection performance, our model runs significantly more efficiently because it integrates the acquisition of head features into the overall framework (As shown in Fig. 5 (b)) and can output bounding boxes and segmentation masks simultaneously. The above results also show that our space-to-object gaze regression method can effectively model the human-object gaze relationship. Table 2 shows our gaze object segmentation performance in real-world and non-real-world scenarios. On GOO-Synth, our model achieves $82.68\%$ and $82.73\%$ mSoC in real-world and non-real-world settings. On GOO-Real, without pre-training, our model performs well in both scenarios $(86.39\% \mathrm{mSoC}$ and $85.94\%$ mSoC). After pre-training, it reaches $89.39\%$ mSoC and $86.96\%$ mSoC. The above experimental results show that the pixel-level mask knowledge provided by VFMs can effectively alleviate semantic ambiguity while achieving pixel-level prediction. Even under the multi-task learning framework, we still achieve an FPS of 14.1. This is mainly attributed to the automatic + +Table 3: Instance segmentation results on GOO-Synth and GOO-Real datasets + +
MethodGOO-Synth DatasetGOO-Real DatasetParams
AP507595AP507595
Non-RealOurs79.3393.5191.0911.7282.9198.8595.0911.2280.99M
RealMaskDINO [20]70.8595.5483.311.8177.0598.3890.454.2652.03M
Ours78.5696.0090.568.4382.8699.0295.5710.0984.26M
+ +Table 4: Object detection results on GOO-Synth and GOO-Real datasets. * indicates using additional head detector to achieve real-world settings. + +
MethodGOO-Synth DatasetGOO-Real DatasetParams
AP507595AP507595
Non-RealGaTector [41]56.8095.3062.500.1052.2591.9255.340.1060.78M
TransGOP [40]87.6099.0097.3025.5077.2097.8089.406.5094.03M
Ours88.3393.8893.0850.7077.3298.6593.911.9380.99M
RealGaTector [41]*43.5086.2337.52-29.2174.0216.13-75.73M
TransGOP [40]*86.0098.8096.2023.8073.3096.7086.004.10106.37M
MaskDINO [20]72.9396.2688.132.0474.4498.7291.190.1952.03M
Ours87.6093.7273.0745.5375.2098.9493.200.7684.26M
+ +incorporation of head features, avoiding the need for an additional head detector and thus enhancing the efficiency of the model. + +Object detection and segmentation. Tables 3 and 4 show our model's instance segmentation and object detection performance. For a fair comparison, we adjusted the input size of MaskDINO [20] to match ours $(224 \times 224)$ . In Table 3, our model outperforms MaskDINO by 8.66 AP on GOO-Real (70.85 vs. 79.51) and by 5.81 AP on GOO-Synth (77.05 vs. 82.86) in real-world scenarios, with similar advantages in non-real-world settings of our model. These results show the challenge faced by the original MaskDINO [20] in capturing detailed features in scenes with dense objects and the effectiveness of the pixel shuffle of our model in capturing detailed features. Table 4 further demonstrates the superiority of our model in object detection over previous methods and MaskDINO [20]. + +Gaze estimation. Tables 5 and 6 show the gaze estimation performance on GOO-Synth and GOO-Real, respectively. In Table 5, our method outperforms GTR [39] and TransGOP [40] by $23.9\mathrm{mAP}$ and $10.6\mathrm{mAP}$ , respectively. Even directly evaluating all prediction results in the real world, our model reduces L2 distance and angle errors compared to TransGOP [40] by 0.015 and 1.8, respectively. Applying non-real-world methods to the real world significantly degrades performance, highlighting our real-world model is more efficient. Table 6 verifies our method on GOO-Real, achieving state-of-the-art results without pretraining. Even with direct inference in a real-world setting, our method outperforms previous non-real-world methods. After pre-training, our model exhibits more pronounced performance advantages. These results show the effectiveness + +Table 5: Gaze estimation results on GOO-Synth dataset. $\dagger$ indicates the evaluation method proposed by [39]. + +
MethodAUCDist.↓Ang.↓mAP
Non-RealGazeFollow [29]0.9290.16233.0-
Lian [22]0.9540.10719.7-
VideoAttention [9]0.9520.07515.1-
GaTector [41]0.9570.07314.9-
GTR [39]0.9600.07114.5-
TransGOP [40]0.9630.07913.3-
Ours0.9470.07214.8-
RealGazeFollow [29]†0.8320.31742.60.468
Lian [22]†0.9030.15328.10.454
VideoAttention [9]†0.9120.14324.50.489
GaTector [41]†0.9180.13924.50.510
GTR [39]†0.9620.06814.20.597
TransGOP [40]†0.9770.07011.50.730
TransGOP [40]0.9450.10620.7-
Ours†0.9780.05710.20.836
Ours0.9380.09118.9-
+ +Table 6: Gaze estimation results on GOO-Real dataset. $\dagger$ indicates the evaluation method proposed by [39]. + +
MethodAUCDist.↓Ang.↓mAP
No PretrainNon-RealGazeFollow [29]0.8500.22044.4-
Lian [22]0.8400.32143.5-
VideoAttention [9]0.7960.25251.4-
GaTector [41]0.9270.19639.5-
Tonini [36]0.9180.164--
TransGOP [40]0.9470.09716.7-
Ours0.9430.07812.9-
RealOurs†0.9710.0599.80.810
Ours0.9440.08814.7-
PretrainNon-RealGazeFollow [29]0.9030.19539.8-
Lian [22]0.8900.16832.6-
VideoAttention [9]0.8890.15029.1-
GaTector [41]0.9400.08714.8-
TransGOP [40]0.9570.08114.7-
Ours0.9630.07312.4-
RealOurs†0.9780.0518.80.824
Ours0.9490.08213.8-
+ +Table 7: Ablation study about each component on GOO-Real + +
PS RoI-Rec. DAF FIM EALGaze Object SegmentationInstance SegmentationGaze Estimation
mSoCmSoC50mSoC75APAP50AP75AUCDist.↓Ang.↓
79.0098.4892.5074.2498.0087.620.8540.30454.2
✓84.5098.7596.5680.5398.6893.670.8390.33459.8
✓✓84.6398.6796.7281.7098.8394.660.9100.11420.5
✓✓✓85.4498.9296.9681.5898.7694.290.9190.09516.8
✓✓✓✓85.8998.8097.1482.4598.9294.820.9320.08914.9
✓✓✓✓✓85.9498.9997.2082.8699.0295.570.9440.08814.7
+ +of our space-to-object gaze regression method in modeling human-object gaze relationships while achieving more accurate gaze heatmap regression. + +# 4.3 Ablation study and model analysis + +Ablation study about the effect of each component. In Table 7, we conduct various ablation studies on the GOO-Real dataset. We first establish a baseline using MaskDINO [20] and a gaze regression branch similar to TransGOP [40], and then gradually add the proposed modules and demonstrate their contribution to overall model performance. (i) Pixel shuffle (PS). After adding the pixel shuffle module, instance segmentation performance improved by $6.29\%$ AP $(80.53\%)$ vs $74.24\%$ compared to baseline, indicating that large feature maps significantly enhance detailed features in dense prediction tasks. (ii) RoI Reconstruction (RoI-Rec.). The RoI reconstruction module reduces angle error by 39.3 and L2 distance by 0.22, demonstrating the importance of head-related features for gaze direction perception and the ability of the module to learn accurate gaze features in real-world settings. (iii) Dual attention fusion (DAF). + +Table 8: Analysis about dual attention fusion module. + +
Original Attention Gaze ConeGaze Object SegmentationInstance SegmentationGaze Estimation
mSoCmSoC50mSoC75APAP50AP75AUCDist.↓Ang.↓
✓85.5198.8697.1782.0898.8794.780.9380.08915.6
✓84.1398.7896.3581.3098.7893.690.9260.09716.3
✓✓85.9498.9997.2082.7599.0195.170.9440.08814.7
+ +Table 9: Analysis about feature interaction module + +
Encoder QueryDecoder QueryMask EmbeddingGaze Object SegmentationInstance SegmentationGaze Estimation
mSoCmSoC50mSoC75APAP50AP75AUC Dist.↓Ang.↓
✓85.7998.3896.4481.2898.8794.870.9280.095 16.0
✓85.2098.2197.0181.6598.8694.650.9490.093 15.1
✓85.9498.9997.2082.8699.0295.570.9440.088 14.7
+ +Introducing the dual attention fusion improves gaze heatmap quality (AUC: 0.919 vs 0.910), demonstrating that perceiving gaze object spatial location beforehand produces more accurate heatmaps and eases regression. (iv) Feature interaction module (FIM). The feature interaction module improves gaze estimation, reducing angle error by 1.9 and L2 distance by 0.006. Gaze object segmentation achieves an $85.89\mathrm{mSoC}$ , showing that pixel-level object information enhances feature interaction and human-object gaze relationship modeling. (v) Energy aggregation loss (EAL). Adding energy aggregation loss improves the performance of all three sub-tasks, showing that the gaze object mask effectively guides the regression of the gaze heatmap towards the gaze object. + +Analysis about dual attention fusion module. Table 8 analyzed the effect of different spatial perception methods. Compared to using original attention $\mathbf{M}_{\mathrm{ori}}$ or gaze cone $\mathbf{M}_{\mathrm{cone}}$ alone, our dual attention fusion performs better. This indicates that combining spatial search along the gaze direction with human head-based scene saliency is more effective. Fig. 4(b) shows that element-wise fusion of $\mathbf{M}_{\mathrm{ori}}$ and $\mathbf{M}_{\mathrm{cone}}$ creates an initial human-object spatial connection. + +Analysis about feature interaction module. In Table 9, we evaluate the efficacy of different object information for gaze regression. Injecting the global encoder query of MaskDINO [20] results in $85.79\%$ mSoC for gaze object segmentation but poor gaze estimation performance. Refining the decoder query with mask supervision reduces the angle error by 0.9, indicating better object representation. Using mask embedding with mask features during gaze regression achieves optimal performance, demonstrating that semantically clear mask features enhance gaze object positioning and the human-object gaze relationship. + +# 4.4 Qualitative Results. + +Fig. 6(a) presents qualitative results in the real-world setting. In object-dense scenes, our method accurately predicts the gaze object mask without relying on head-related cues. In Fig. 6(b), compared to previous methods, our method ac + +![](images/96de6f78ea012101fb96efc55275baa794bc376c136ab1332a48dacc373341f3.jpg) +Fig. 6: Qualitative comparison of ground truth gaze object (GT), gaze object detection (GOD), instance segmentation (IS), gaze object segmentation (GOS) + +curately locates adjacent objects using a pixel-level mask, effectively alleviating semantic ambiguity. See supplementary material for more visualization. + +# 5 Conclusion + +In this paper, we introduce the gaze object segmentation task, which aims to infer pixel-level masks for objects captured by human eyesight, and establish a framework based on the pixel-level supervision from VFM to solve this task while alleviating the semantic ambiguity existing in previous box-level supervision methods. In our method, we choose SAM to generate pixel-level supervision information for the model training. The RoI reconstruction module is proposed to reconstruct head features directly from holistic features using the head box provided by the unified Transformer, which improves the inference efficiency and flexibility of our model in the real world. To boost the representation of the human-object gaze relationship, we propose a space-to-object gaze regression strategy, which first uses the dual attention fusion module to establish a human-object spatial connection, and then a semantically clear mask feature is injected into the gaze regression to strengthen the human-object gaze relationship modeling. Extensive experiments on GOO-Real and GOO-Synth demonstrate the effectiveness of our method in various settings. + +**Limitation:** The proposed method requires box supervision as prompts for the SAM when generating masks, which limits its application in a wider range of scenarios without strong location priors, more flexible approaches for generating masks via the latest VFM models need to be explored in future work. + +# Acknowledgements + +This work is supported by the National Natural Science Foundation of China. (No. 82272020). + +Bin Fan is sponsored by China National Postdoctoral Program for Innovative Talents (No. BX20230013). + +# References + +1. Bao, J., Liu, B., Yu, J.: ESCNet: Gaze target detection with the understanding of 3D scenes. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 14126-14135 (2022) +2. Cai, X., Zeng, J., Shan, S., Chen, X.: Source-free adaptive gaze estimation by uncertainty reduction. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 22035-22045 (2023) +3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Eur. Conf. Comput. Vis. pp. 213-229 (2020) +4. Chen, W., Li, R., Yu, Q., Xu, A., Feng, Y., Wang, R., Zhao, L., Lin, Z., Yang, Y., Lin, D., et al.: Early detection of visual impairment in young children using a smartphone-based deep learning system. Nature Medicine 29(2), 493-503 (2023) +5. Chen, Y., Nan, Z., Xiang, T.: FBLNet: Feedback loop network for driver attention prediction. In: Int. Conf. Comput. Vis. pp. 13371-13380 (2023) +6. Cheng, Y., Lu, F.: Gaze estimation using transformer. In: Int. Conf. Pattern Recog. pp. 3341-3347 (2022) +7. Cheng, Y., Lu, F., Zhang, X.: Appearance-based gaze estimation via evaluation-guided asymmetric regression. In: Eur. Conf. Comput. Vis. pp. 100-115 (2018) +8. Chong, E., Ruiz, N., Wang, Y., Zhang, Y., Rozga, A., Rehg, J.M.: Connecting gaze, scene, and attention: Generalized attention estimation via joint modeling of gaze and scene saliency. In: Eur. Conf. Comput. Vis. pp. 383-398 (2018) +9. Chong, E., Wang, Y., Ruiz, N., Rehg, J.M.: Detecting attended visual targets in video. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5396-5406 (2020) +0. Gupta, A., Tafasca, S., Odobez, J.M.: A modular multimodal architecture for gaze target prediction: Application to privacy-sensitive settings. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5041-5050 (2022) +1. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask R-CNN. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2961-2969 (2017) +2. Hu, Z., Yang, D., Cheng, S., Zhou, L., Wu, S., Liu, J.: We know where they are looking at from the RGB-D camera: Gaze following in 3D. IEEE Trans. Instrum. Meas. 71, 1-14 (2022) +3. Hu, Z., Yang, Y., Zhai, X., Yang, D., Zhou, B., Liu, J.: GFIE: A dataset and baseline for gaze-following from 2D to 3D in indoor environments. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8907-8916 (2023) +4. Hu, Z., Zhao, K., Zhou, B., Guo, H., Wu, S., Yang, Y., Liu, J.: Gaze target estimation inspired by interactive attention. IEEE Trans. Circuit Syst. Video Technol. 32(12), 8524-8536 (2022) +5. Huang, T., Fu, R.: Driver distraction detection based on the true driver's focus of attention. IEEE Trans. Intell. Transp. Syst. 23(10), 19374-19386 (2022) +6. Jin, S., Wang, Z., Wang, L., Bi, N., Nguyen, T.: ReDirTrans: Latent-to-latent translation for gaze and head redirection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5547-5556 (2023) + +17. Jin, T., Yu, Q., Zhu, S., Lin, Z., Ren, J., Zhou, Y., Song, W.: Depth-aware gaze-following via auxiliary networks for robotics. Eng. Appl. Artif. Intell. 113, 104924 (2022) +18. Kellnhofer, P., Recasens, A., Stent, S., Matusik, W., Torralba, A.: Gaze360: Physically unconstrained gaze estimation in the wild. In: Int. Conf. Comput. Vis. pp. 6912-6921 (2019) +19. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Int. Conf. Comput. Vis. pp. 4015-4026 (2023) +20. Li, F., Zhang, H., Xu, H., Liu, S., Zhang, L., Ni, L.M., Shum, H.Y.: Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3041-3050 (2023) +21. Li, Y., Shen, W., Gao, Z., Zhu, Y., Zhai, G., Guo, G.: Looking here or there? gaze following in 360-degree images. In: Int. Conf. Comput. Vis. pp. 3742-3751 (2021) +22. Lian, D., Yu, Z., Gao, S.: Believe it or not, we know what you are looking at! In: Asian Conf. Comput. Vis. pp. 35-50 (2018) +23. Lv, K., Sheng, H., Xiong, Z., Li, W., Zheng, L.: Improving driver gaze prediction with reinforced attention. IEEE Trans. Multimedia 23, 4198-4207 (2020) +24. Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature Communications 15(1), 654 (2024) +25. Miao, Q., Hoai, M., Samaras, D.: Patch-level gaze distribution prediction for gaze following. In: IEEE Winter Conf. Appl. Comput. Vis. pp. 880-889 (2023) +26. Mundy, P., Sigman, M., Kasari, C.: A longitudinal study of joint attention and language development in autistic children. J. Autism Dev. Disord. 20(1), 115-128 (1990) +27. Park, S., Spurr, A., Hilliges, O.: Deep pictorial gaze estimation. In: Eur. Conf. Comput. Vis. pp. 721-738 (2018) +28. Park, S., Zhang, X., Bulling, A., Hilliges, O.: Learning to find eye region landmarks for remote gaze estimation in unconstrained settings. In: Proc. ACM Symp. Eye Track. Res. Appl. pp. 1-10 (2018) +29. Recasens, A., Khosla, A., Vondrick, C., Torralba, A.: Where are they looking? Adv. Neural Inform. Process. Syst. 28 (2015) +30. Recasens, A., Vondrick, C., Khosla, A., Torralba, A.: Following gaze in video. In: Int. Conf. Comput. Vis. pp. 1435-1443 (2017) +31. Ruzzi, A., Shi, X., Wang, X., Li, G., De Mello, S., Chang, H.J., Zhang, X., Hilliges, O.: GazeNeRF: 3d-aware gaze redirection with neural radiance fields. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 9676-9685 (2023) +32. Senju, A., Johnson, M.H.: Atypical eye contact in autism: Models, mechanisms and development. Neurosci. Biobehav. Rev. 33(8), 1204-1214 (2009) +33. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient subpixel convolutional neural network. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1874-1883 (2016) +34. Tafasca, S., Gupta, A., Odobez, J.M.: ChildPlay: A new benchmark for understanding children's gaze behaviour. In: Int. Conf. Comput. Vis. pp. 20935-20946 (2023) +35. Tomas, H., Reyes, M., Dionido, R., Ty, M., Mirando, J., Casimiro, J., Atienza, R., Quinto, R.: Goo: A dataset for gaze object prediction in retail environments. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. pp. 3125-3133 (2021) +36. Tonini, F., Beyer, C., Ricci, E.: Multimodal across domains gaze target detection. In: Int. Conf. Multimodal Interact. pp. 420-431 (2022) + +37. Tonini, F., Dall'Asen, N., Bayan, C., Ricci, E.: Object-aware gaze target detection. In: Int. Conf. Comput. Vis. pp. 21860-21869 (2023) +38. Tu, D., Min, X., Duan, H., Guo, G., Zhai, G., Shen, W.: End-to-end human-gazetarget detection with transformers. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2192-2200 (2022) +39. Tu, D., Shen, W., Sun, W., Min, X., Zhai, G., Chen, C.: Un-Gaze: A unified transformer for joint gaze-location and gaze-object detection. IEEE Trans. Circuit Syst. Video Technol. (2023) +40. Wang, B., Guo, C., Jin, Y., Xia, H., Liu, N.: TransGOP: Transformer-based gaze object prediction. AAAI Conf. Artif. Intell. (2024) +41. Wang, B., Hu, T., Li, B., Chen, X., Zhang, Z.: GaTector: A unified framework for gaze object prediction. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19588-19597 (2022) +42. Wang, K., Ji, Q.: Real time eye gaze tracking with 3D deformable eye-face model. In: Int. Conf. Comput. Vis. pp. 1003-1011 (2017) +43. Wang, X., Zhang, H., Wang, Z., Nie, W., Yang, Z., Ren, W., Xu, Q., Xu, X., Liu, H.: Dual regression-enhanced gaze target detection in the wild. IEEE Trans. Cybern. 54(1), 219-229 (2024) +44. Wang, Z., Zhao, J., Lu, C., Yang, F., Huang, H., Guo, Y., et al.: Learning to detect head movement in unconstrained remote gaze estimation in the wild. In: IEEE Winter Conf. Appl. Comput. Vis. pp. 3443-3452 (2020) +45. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. Int. Conf. Learn. Represent. (2022) +46. Zhang, X., Sugano, Y., Fritz, M., Bulling, A.: Appearance-based gaze estimation in the wild. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4511-4520 (2015) +47. Zhang, X., Sugano, Y., Fritz, M., Bulling, A.: Mpiigaze: Real-world dataset and deep appearance-based gaze estimation. IEEE Trans. Pattern Anal. Mach. Intell. 41(1), 162-175 (2017) +48. Zhao, X., Ding, W., An, Y., Du, Y., Yu, T., Li, M., Tang, M., Wang, J.: Fast segment anything. arXiv preprint arXiv:2306.12156 (2023) +49. Zhu, W., Deng, H.: Monocular free-head 3D gaze tracking with deep learning and geometry constraints. In: Int. Conf. Comput. Vis. pp. 3143-3152 (2017) +50. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: Int. Conf. Learn. Represent. (2021) \ No newline at end of file diff --git a/2024/Boosting Gaze Object Prediction via Pixel-level Supervision from Vision Foundation Model/images.zip b/2024/Boosting Gaze Object Prediction via Pixel-level Supervision from Vision Foundation Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..16ca174510c89671deeb38ca45a1afbb4e09b2a1 --- /dev/null +++ b/2024/Boosting Gaze Object Prediction via Pixel-level Supervision from Vision Foundation Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ecedb1d061889cd9790afc98afaa50e1707df3ff1dac0dc6744732c0d6b7fb1 +size 767770 diff --git a/2024/Boosting Gaze Object Prediction via Pixel-level Supervision from Vision Foundation Model/layout.json b/2024/Boosting Gaze Object Prediction via Pixel-level Supervision from Vision Foundation Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a2625d61b81330d74fc037ec3678a849a50b5211 --- /dev/null +++ b/2024/Boosting Gaze Object Prediction via Pixel-level Supervision from Vision Foundation Model/layout.json @@ -0,0 +1,9330 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "text", + "content": "Boosting Gaze Object Prediction via Pixel-level Supervision from Vision Foundation Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 147, + 171, + 465, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 171, + 465, + 185 + ], + "spans": [ + { + "bbox": [ + 147, + 171, + 465, + 185 + ], + "type": "text", + "content": "Yang Jin1, Lei Zhang2, Shi Yan2, Bin Fan3, and Binglu Wang1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 175, + 193, + 438, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 193, + 438, + 217 + ], + "spans": [ + { + "bbox": [ + 175, + 193, + 438, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 175, + 193, + 438, + 217 + ], + "type": "text", + "content": " Xi'an University of Architecture and Technology, Xi'an, China {jin91999, wbl921129}@gmail.com † Corresponding author" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 217, + 465, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 217, + 465, + 237 + ], + "spans": [ + { + "bbox": [ + 140, + 217, + 465, + 237 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 217, + 465, + 237 + ], + "type": "text", + "content": " School of Automation, Northwestern Polytechnical University, Xi'an, China \nzl_hnly@163.com, yanshi@mail.nwpu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 238, + 472, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 238, + 472, + 271 + ], + "spans": [ + { + "bbox": [ + 140, + 238, + 472, + 271 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 140, + 238, + 472, + 271 + ], + "type": "text", + "content": " National Key Lab of General AI, School of Intelligence Science and Technology, Peking University, Beijing, China binfan@pku.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 332, + 452, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 332, + 452, + 628 + ], + "spans": [ + { + "bbox": [ + 159, + 332, + 452, + 628 + ], + "type": "text", + "content": "Abstract. Gaze object prediction (GOP) aims to predict the category and location of the object that a human is looking at. Previous methods utilized box-level supervision to identify the object that a person is looking at, but struggled with semantic ambiguity, i.e., a single box may contain several items since objects are close together. The Vision foundation model (VFM) has improved in object segmentation using box prompts, which can reduce confusion by more precisely locating objects, offering advantages for fine-grained prediction of gaze objects. This paper presents a more challenging gaze object segmentation (GOS) task, which involves inferring the pixel-level mask corresponding to the object captured by human gaze behavior. In particular, we propose that the pixel-level supervision provided by VFM can be integrated into gaze object prediction to mitigate semantic ambiguity. This leads to our gaze object detection and segmentation framework that enables accurate pixel-level predictions. Different from previous methods that require additional head input or ignore head features, we propose to automatically obtain head features from scene features to ensure the model's inference efficiency and flexibility in the real world. Moreover, rather than directly fuse features to predict gaze heatmap as in existing methods, which may overlook spatial location and subtle details of the object, we develop a space-to-object gaze regression method to facilitate human-object gaze interaction. Specifically, it first constructs an initial human-object spatial connection, then refines this connection by interacting with semantically clear features in the segmentation branch, ultimately predicting a gaze heatmap for precise localization. Extensive experiments on GOO-Synth and GOO-Real datasets demonstrate the effectiveness of our method. The code will be available at https://github.com/jinyang06/SamGOP." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 643, + 453, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 643, + 453, + 665 + ], + "spans": [ + { + "bbox": [ + 159, + 643, + 453, + 665 + ], + "type": "text", + "content": "Keywords: Gaze object prediction " + }, + { + "bbox": [ + 159, + 643, + 453, + 665 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 643, + 453, + 665 + ], + "type": "text", + "content": " Vision foundation model " + }, + { + "bbox": [ + 159, + 643, + 453, + 665 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 643, + 453, + 665 + ], + "type": "text", + "content": " Object segmentation " + }, + { + "bbox": [ + 159, + 643, + 453, + 665 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 643, + 453, + 665 + ], + "type": "text", + "content": " Space-to-object gaze regression" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 306, + 202 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 306, + 202 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 306, + 202 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 306, + 202 + ], + "type": "image", + "image_path": "494d3ca1c7ce9f1f6a9dcf3a12974465e1826e94bdf96377c54067106ffe7cb6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 202, + 306, + 211 + ], + "lines": [ + { + "bbox": [ + 133, + 202, + 306, + 211 + ], + "spans": [ + { + "bbox": [ + 133, + 202, + 306, + 211 + ], + "type": "text", + "content": "(a) Illustration of semantic ambiguity and semantic clarity" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 220, + 482, + 286 + ], + "lines": [ + { + "bbox": [ + 130, + 220, + 482, + 286 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 482, + 286 + ], + "type": "text", + "content": "Fig. 1: (a) Box-level supervision often fails to localize objects in dense settings precisely and leads to semantic ambiguity problems, whereas pixel-level supervision excels by providing clear semantic distinction through pixel-by-pixel predictions. (b) Vision foundation models can produce instance masks, thereby segmentation features can be used to improve the gaze regression branch's spatial perception, and the gaze object mask can help the heatmap focus on the gaze object." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 116, + 482, + 201 + ], + "blocks": [ + { + "bbox": [ + 310, + 116, + 482, + 201 + ], + "lines": [ + { + "bbox": [ + 310, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 310, + 116, + 482, + 201 + ], + "type": "image", + "image_path": "97b8cf9b4a6ee2cbbb92b1087e24d231e9bdf2d10ffc3f2dbfb1f1020fd5c394.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 202, + 463, + 211 + ], + "lines": [ + { + "bbox": [ + 364, + 202, + 463, + 211 + ], + "spans": [ + { + "bbox": [ + 364, + 202, + 463, + 211 + ], + "type": "text", + "content": "(b) The main idea of our pipeline" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 307, + 230, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 307, + 230, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 230, + 319 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 331, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 437 + ], + "type": "text", + "content": "The objects stared at by humans contain important semantic information, which can reveal a person's behavior and state of mind. Therefore, identifying the objects that people are looking at has a wide range of applications. For example, in medical diagnosis, whether a child focuses on an object can potentially reveal whether he/she has autism or visual impairment [4,26,32,34]. In assisted driving scenarios, determining what object the driver is looking at can assess their attention and provide corresponding driving guidance [5,15,23]. In general, detecting the object gazed at by humans realizes the interaction between human sight and specific objects, fostering a richer semantic understanding of the scene." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 438, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 629 + ], + "type": "text", + "content": "Tomas et al. [35] first proposed the gaze object detection task and the benchmark datasets GOO-Synth and GOO-Real, aiming to predict the object boxes and categories stared at by humans. Recently, Wang et al. [40, 41] conducted an in-depth exploration of this task. Despite achieving promising results, they still suffer from the following limitations: 1) They typically rely on box-level supervision, which is prone to suffer semantic ambiguity in dense object scenes. Especially when multiple objects are densely stacked or placed adjacent to each other, the box-level supervision algorithm may struggle to accurately separate them, as illustrated in Fig. 1(a). 2) The GOO [35] dataset only contains pixel-level annotations of synthetic images, leading to difficulties in generalizing to fine-grained predictions in real scenes. 3) They usually require additional head information as input, affecting the efficiency and flexibility of the model. While global modeling using the Transformer [38, 39] can alleviate this problem, it lacks special attention to head features, which are crucial for gaze direction perception. 4) They directly fuse features to regress the gaze heatmap, lacking a full understanding of the spatial location and subtle detailed information of objects." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Recently, vision foundational models (VFM) have gained significant attention due to their powerful generalization capabilities, which can achieve accurate pixel-level segmentation using a variety of prompts, i.e., boxes, points, or masks." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Jin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "In real scenes with small and densely packed objects, VFM can generate accurate boundaries and pixel-level masks (See Fig 3) by utilizing boxes as prompt, which reduces semantic ambiguity and thus opens opportunities for gaze object segmentation. Motivated by this, we propose a more challenging task, gaze object segmentation (GOS). This task involves inferring the pixel-level mask of objects captured by human gaze behavior, helping to further enhance the understanding of human gaze behavior. Furthermore, we establish a GOS framework specifically for this task, which integrates the pixel-level supervision provided by VFM into gaze object prediction to alleviate the aforementioned limitations of previous methods [40, 41]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 488 + ], + "type": "text", + "content": "In the proposed model, we introduce the unified detection and segmentation Transformer, i.e. MaskDINO [20], to simultaneously obtain candidate gaze object boxes, segmentation masks, and head boxes of the subject, which are indispensable for enabling the model to infer in the real world and achieve accurate detection. Instead of requiring additional head input [40, 41] or ignoring head features by global modeling [38, 39], we propose a RoI reconstruction module to automatically obtain head region features based on holistic features and head box provided by the unified Transformer. This helps capture gaze direction information during gaze regression to improve model efficiency and flexibility. Subsequently, a space-to-object gaze regression strategy is developed to enhance the perception of the spatial location of objects and the understanding of detailed information. Specifically, inspired by the ability of humans to infer others' gaze objects, a dual attention fusion module is presented to establish an initial human-object gaze spatial connection. Then, we propose an object feature interaction module to refine the gaze spatial connection through semantically clear feature interaction between gaze regression features and object mask features, thus strengthening the modeling of the human-object gaze relationship. In the end, an energy aggregation loss is used to further guide the regression of the gaze heatmap. By focusing on pixel-level details and progressively improving the gaze representation, it further reduces semantic ambiguity, leading to the generation of more precise heatmaps for gazed objects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 488, + 394, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 488, + 394, + 498 + ], + "spans": [ + { + "bbox": [ + 146, + 488, + 394, + 498 + ], + "type": "text", + "content": "Our main contributions could be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 509, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 509, + 481, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 509, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 138, + 509, + 481, + 557 + ], + "type": "text", + "content": "- We introduce the vision foundation model into gaze object prediction to provide pixel-level supervision and propose a more challenging gaze object segmentation task, which focuses on identifying the pixel-level masks of objects captured by human gaze behavior." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 558, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 558, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 138, + 558, + 482, + 617 + ], + "type": "text", + "content": "- We present an all-in-one end-to-end framework for gaze object detection and segmentation without relying on additional head-related inputs, capable of simultaneously addressing gaze estimation, object detection, and segmentation. To the best of our knowledge, our work is the first to tackle gaze object prediction from the image segmentation perspective." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 618, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 618, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 618, + 482, + 665 + ], + "type": "text", + "content": "- We propose a space-to-object gaze regression approach that enhances the interaction between object and gaze branches, incrementally improving the modeling of the human-object gaze relationship and alleviating semantic ambiguity." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "text", + "content": "Boosting Gaze Object Prediction via Pixel-level Supervision" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 243, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 243, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 243, + 127 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 236 + ], + "type": "text", + "content": "Vision Foundation Model. The vision foundation models (VFMs) [19, 24, 48] are typically built on large-scale datasets using self-supervised or semi-supervised methods, which have strong generalizability and zero-shot transfer capabilities. Segment Anything Model (SAM) [19] received attention as a foundation model for image segmentation. Its portability enables it to be flexibly applied to various downstream tasks. Then, the medical foundation model MedSAM [24] extends the foundation model to medical diagnosis. In this article, we choose SAM [19] to obtain high-quality masks to boost gaze estimation and gaze object prediction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 481, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 481, + 344 + ], + "type": "text", + "content": "Gaze Object Detection. The gaze object detection is first proposed in [35], which aims to predict the box and categories of the object stared at by humans in an RGB image. Tomas et al. [35] released a benchmark dataset for GOP. Wang et al. [41] first proposed a GOP method named GaTector. However, GaTector requires additional head-related prior information, which hinders its application in the real world. Recently, the Transformer-based method [40] achieves comparable performance. Although previous methods are effective to some extent, they can only provide instance-level box outputs, leading to potential ambiguities or misjudgments in scenarios with small and densely packed objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 344, + 481, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 481, + 499 + ], + "type": "text", + "content": "Gaze Target Detection. Early research on gaze focused on gaze direction estimation [2, 5-7, 16, 18, 27, 28, 31, 42, 44, 46, 47, 49], which only predict the gaze vector and ignores the impact of the surrounding environment on the sight, resulting in limited application scenarios. Therefore, to capture higher-level gaze-related semantic information, researchers propose gaze target detection [1, 8-10, 12-14, 17, 21, 22, 25, 30, 34, 36, 37, 43], which involves to infer the specific gaze point. Recasens et al. [29] introduced the GazeFollow dataset and first proposed gaze target detection. Recently, Tu et al. [38,39] achieved end-to-end gaze target detection using Transformer, with a relatively low detection efficiency. Despite advancements in gaze target estimation, previous methods struggle with gaze regression due to spatial perception ambiguity. In contrast, our space-to-object regression strategy helps the model gradually converge to a certain gaze point, simplifying learning and enhancing detailed interaction with the object." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 517, + 202, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 517, + 202, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 202, + 530 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 543, + 481, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 543, + 481, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 543, + 481, + 591 + ], + "type": "text", + "content": "Given an RGB image, our model aims to precisely predict the pixel-level mask of the object humans are looking at. We first present the overall structure of our method, then introduce the VFM-based mask supervision generation method, and finally describe each module of the proposed model in detail." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 609, + 208, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 208, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 208, + 620 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "As shown in Fig. 2, our method consists of four components: a feature extractor for specific feature extraction, a detection and segmentation module predicting candidate gaze objects boxes, masks, and human head location, a gaze regressor" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Jin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 479, + 232 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 479, + 232 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 479, + 232 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 479, + 232 + ], + "type": "image", + "image_path": "8f1529fbd145da8de17a7eb09b2daee56dd8fcd3712c58baa4cc8dd28a46ada4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 239, + 480, + 328 + ], + "lines": [ + { + "bbox": [ + 130, + 239, + 480, + 328 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 480, + 328 + ], + "type": "text", + "content": "Fig. 2: Overview of the proposed model. (a) The feature extraction module extracts features for detection and regression branches. (b) The detection and segmentation branch identifies object and human head positions. (c) After obtaining head features, the gaze regression branch progressively refines its output: 1) employing a dual attention fusion module for initial human-object correlations; 2) leveraging a feature interaction module to incorporate semantically clear object-aware insights from the segmentation branch; 3) ultimately predicting the gaze heatmap. (d) Supervision signals are applied to both branches only during training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 350, + 479, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 479, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 479, + 387 + ], + "type": "text", + "content": "producing gaze object heatmap with a space-to-object strategy, and a supervision information generation module creating pixel-level masks and ground truth heatmap during training. At training, the overall loss function is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 228, + 393, + 481, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 393, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 228, + 393, + 481, + 407 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {d e t}} + \\alpha \\mathcal {L} _ {\\mathrm {d i r}} + \\beta \\mathcal {L} _ {\\mathrm {g a z e}} + \\gamma \\mathcal {L} _ {\\mathrm {e n g}} \\tag {1}", + "image_path": "e8f1ee5554d8ffb1ee66be238b83aebee4d51b6fcecee0fec8f42ca202d294a9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 413, + 479, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 479, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 479, + 448 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 413, + 479, + 448 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 413, + 479, + 448 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 413, + 479, + 448 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 413, + 479, + 448 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 413, + 479, + 448 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 413, + 479, + 448 + ], + "type": "text", + "content": " are the weights of the gaze direction loss, gaze heatmap loss, and energy aggregation loss respectively. Before model training, we choose SAM to generate pixel-level masks for the objects. See Sec. 3.2 for details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "text", + "content": "Feature extraction. As shown in 2(a), given an RGB image " + }, + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{I} \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "text", + "content": ", we first use the backbone " + }, + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "text", + "content": " to extract multi-scale features " + }, + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{f}_1, \\mathbf{f}_2, \\mathbf{f}_3, \\mathbf{f}_4\\}" + }, + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "text", + "content": ". In previous work [41], pixel shuffle [33] significantly improved small object detection by enlarging feature maps for detailed feature capture. Therefore, we integrate it into dense prediction tasks, i.e., image segmentation. In our model, the multi-scale features " + }, + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{f}_1, \\mathbf{f}_2, \\mathbf{f}_3, \\mathbf{f}_4\\}" + }, + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "text", + "content": " are fed into the pixel shuffle module to generate object-specific features " + }, + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i^{obj} = \\phi^{obj}(\\mathbf{f}_i, \\eta)" + }, + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "text", + "content": "-th feature layer, " + }, + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "text", + "content": " denotes the scale factor, " + }, + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "inline_equation", + "content": "\\phi^{obj}(\\cdot)" + }, + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "text", + "content": " denotes the pixel shuffle module, which consists of a pixel shuffle operation and a convolution layer that transforms the feature channels to be consistent with the original features of the unified Transformer [20]. Finally, the scene saliency features " + }, + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{scene}} \\in \\mathbb{R}^{1024 \\times 7 \\times 7}" + }, + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "text", + "content": " are extracted by a scene residual block using high dimensional features " + }, + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_4 \\in \\mathbb{R}^{2048 \\times 7 \\times 7}" + }, + { + "bbox": [ + 130, + 449, + 480, + 593 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "Unified detection and segmentation Transformer. In this paper, we employ a unified Transformer for candidate gaze object detection, segmentation, and human head positioning, integrating MaskDINO [20] for candidate gaze object detection, segmentation, and the generation of pixel-level mask feature embeddings and a head decoder for head box regression, which facilitate accurate head-related feature extraction in head feature reconstruction module and" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "text", + "content": "Boosting Gaze Object Prediction via Pixel-level Supervision" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 116, + 479, + 163 + ], + "blocks": [ + { + "bbox": [ + 134, + 116, + 479, + 163 + ], + "lines": [ + { + "bbox": [ + 134, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 479, + 163 + ], + "type": "image", + "image_path": "1ccb143e786d159e727b559a8131322e7e1b5183edccaf11d6264df7973d0178.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 171, + 423, + 183 + ], + "lines": [ + { + "bbox": [ + 190, + 171, + 423, + 183 + ], + "spans": [ + { + "bbox": [ + 190, + 171, + 423, + 183 + ], + "type": "text", + "content": "Fig. 3: Illustration of instance mask generated by VFM." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 205, + 453, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 453, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 453, + 217 + ], + "type": "text", + "content": "accurate prediction in real-world settings. The training loss is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 242, + 224, + 481, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 224, + 481, + 240 + ], + "spans": [ + { + "bbox": [ + 242, + 224, + 481, + 240 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {d e t}} = \\mathcal {L} _ {\\mathrm {d e t}} ^ {\\mathrm {o b j}} + \\mathcal {L} _ {\\mathrm {m a s k}} ^ {\\mathrm {o b j}} + \\mathcal {L} _ {\\mathrm {d e t}} ^ {\\mathrm {h e a d}} \\tag {2}", + "image_path": "fd43483a1bbcbf068342ab50d42b1cfb5cce56adb25962460f5aea79ded70e78.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 246, + 482, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 246, + 482, + 381 + ], + "spans": [ + { + "bbox": [ + 131, + 246, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{det}}^{\\mathrm{obj}}" + }, + { + "bbox": [ + 131, + 246, + 482, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 246, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mask}}^{\\mathrm{obj}}" + }, + { + "bbox": [ + 131, + 246, + 482, + 381 + ], + "type": "text", + "content": " is the object detection and segmentation loss respectively. " + }, + { + "bbox": [ + 131, + 246, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{det}}^{\\mathrm{head}}" + }, + { + "bbox": [ + 131, + 246, + 482, + 381 + ], + "type": "text", + "content": " is the head decoder loss, which is consistent with the " + }, + { + "bbox": [ + 131, + 246, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{det}}^{\\mathrm{obj}}" + }, + { + "bbox": [ + 131, + 246, + 482, + 381 + ], + "type": "text", + "content": ". See Sec. 3.3 for details. Space-to-object gaze regression. We employ a space-to-object regression strategy for gaze heatmap prediction. The module begins by representing human gaze behavior with a gaze vector. Subsequently, a dual attention fusion module establishes an initial human-object spatial connection. Finally, pixel-level object location knowledge from the mask branch is introduced to achieve semantically clear feature interaction, refining the gaze heatmap regression while alleviating semantic ambiguity. Optimization involves " + }, + { + "bbox": [ + 131, + 246, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dir}}" + }, + { + "bbox": [ + 131, + 246, + 482, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 246, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{gaze}}" + }, + { + "bbox": [ + 131, + 246, + 482, + 381 + ], + "type": "text", + "content": " for gaze direction and gaze heatmap, respectively. The energy aggregation loss " + }, + { + "bbox": [ + 131, + 246, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{eng}}" + }, + { + "bbox": [ + 131, + 246, + 482, + 381 + ], + "type": "text", + "content": " guides the gaze heatmap to focus on the gaze object mask. See Sec. 3.4 for details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 398, + 340, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 398, + 340, + 410 + ], + "spans": [ + { + "bbox": [ + 132, + 398, + 340, + 410 + ], + "type": "text", + "content": "3.2 Instance mask generation via VFM." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 416, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 416, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 131, + 416, + 482, + 477 + ], + "type": "text", + "content": "Given an RGB image " + }, + { + "bbox": [ + 131, + 416, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 131, + 416, + 482, + 477 + ], + "type": "text", + "content": ", we choose the vision foundation model SAM [19] to obtain instance masks for model training. Specifically, image " + }, + { + "bbox": [ + 131, + 416, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 131, + 416, + 482, + 477 + ], + "type": "text", + "content": " and prompt " + }, + { + "bbox": [ + 131, + 416, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{box}" + }, + { + "bbox": [ + 131, + 416, + 482, + 477 + ], + "type": "text", + "content": " are first fed into the image encoder " + }, + { + "bbox": [ + 131, + 416, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\varPsi_{Image}" + }, + { + "bbox": [ + 131, + 416, + 482, + 477 + ], + "type": "text", + "content": " and prompt encoder " + }, + { + "bbox": [ + 131, + 416, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\varPsi_{prompt}" + }, + { + "bbox": [ + 131, + 416, + 482, + 477 + ], + "type": "text", + "content": ", respectively. Then, the mask " + }, + { + "bbox": [ + 131, + 416, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{mask}" + }, + { + "bbox": [ + 131, + 416, + 482, + 477 + ], + "type": "text", + "content": " and confidence " + }, + { + "bbox": [ + 131, + 416, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 131, + 416, + 482, + 477 + ], + "type": "text", + "content": " are output by the mask decoder " + }, + { + "bbox": [ + 131, + 416, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\varPhi_{mask}" + }, + { + "bbox": [ + 131, + 416, + 482, + 477 + ], + "type": "text", + "content": ", which can be formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 208, + 485, + 481, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 485, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 208, + 485, + 481, + 499 + ], + "type": "interline_equation", + "content": "\\mathcal {P} _ {\\text {m a s k}}, \\mathcal {S} = \\Phi_ {\\text {m a s k}} (\\Psi_ {\\text {I m a g e}} (\\mathcal {I}), \\Psi_ {\\text {p r o m p t}} (\\mathcal {P} _ {\\text {b o x}})) \\tag {3}", + "image_path": "2895436b37f2b101f339830f4a05c81ea48dcf627c421837fe033a81119a45a3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 505, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 505, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 131, + 505, + 482, + 554 + ], + "type": "text", + "content": "Instance segmentation by using boxes produces accurate masks for most objects, but accurate segmentation of objects with unclear texture or edge features remains challenging. Therefore, we segment again using " + }, + { + "bbox": [ + 131, + 505, + 482, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{mask}" + }, + { + "bbox": [ + 131, + 505, + 482, + 554 + ], + "type": "text", + "content": " as a dense prompt to produce final mask " + }, + { + "bbox": [ + 131, + 505, + 482, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathrm{mask}}" + }, + { + "bbox": [ + 131, + 505, + 482, + 554 + ], + "type": "text", + "content": ", which can be formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 201, + 562, + 481, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 562, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 201, + 562, + 481, + 576 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {\\text {m a s k}}, \\mathcal {S} = \\Phi_ {\\text {m a s k}} (\\Psi_ {\\text {I m a g e}} (\\mathcal {I}), \\Psi_ {\\text {p r o m p t}} (\\mathcal {P} _ {\\text {m a s k}})) \\tag {4}", + "image_path": "b4318580cf55a9da04fdeab23890bad7d7172b994dc179ca5d9377ab0c4330c4.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 582, + 481, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 582, + 481, + 606 + ], + "spans": [ + { + "bbox": [ + 131, + 582, + 481, + 606 + ], + "type": "text", + "content": "Finally, " + }, + { + "bbox": [ + 131, + 582, + 481, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{mask}" + }, + { + "bbox": [ + 131, + 582, + 481, + 606 + ], + "type": "text", + "content": " is used to train the gaze object segmentation model. The generated masks are shown in Fig. 3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 622, + 406, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 406, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 406, + 635 + ], + "type": "text", + "content": "3.3 Unified detection and segmentation Transformer" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "The unified detection and segmentation Transformer aims to detect candidate gaze objects and locate human heads while extracting high-level pixel mask" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Jin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "content": "features, which comprises a detection and segmentation Transformer, a head decoder, and a RoI feature reconstruction module. As shown in Fig. 2(b), the object-specific features " + }, + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{f}_1^{obj},\\mathbf{f}_2^{obj},\\mathbf{f}_3^{obj},\\mathbf{f}_4^{obj}\\}" + }, + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "content": " are fed it to detect object boxes, categories, masks, and human head boxes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 166, + 480, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 166, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 130, + 166, + 480, + 225 + ], + "type": "text", + "content": "Object detection and segmentation The global attention of Transformer ensures powerful long-distance modeling and global perception, achieving excellent detection performance through layer-by-layer self-attention refinement. We use MaskDINO [20] as a foundational detection and segmentation structure due to its robust representation capabilities and fast convergence speed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 226, + 480, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 480, + 309 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 480, + 309 + ], + "type": "text", + "content": "Head decoder Our model is designed for real-world gaze object detection and segmentation without relying on head priors. We introduce a head decoder based on the unified Transformer, using a 3-layer self-attention decoder to generate proposals for head-related feature reconstruction. During training, we filter head queries with IoU " + }, + { + "bbox": [ + 130, + 226, + 480, + 309 + ], + "type": "inline_equation", + "content": ">0.5" + }, + { + "bbox": [ + 130, + 226, + 480, + 309 + ], + "type": "text", + "content": " as positives. During inference, boxes are selected based on confidence. Similar to other DETR-like detectors [45, 50], the head decoder uses cross-entropy, " + }, + { + "bbox": [ + 130, + 226, + 480, + 309 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 130, + 226, + 480, + 309 + ], + "type": "text", + "content": ", and GIoU loss for classification and box regression." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 310, + 481, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 481, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 481, + 477 + ], + "type": "text", + "content": "Head feature reconstruction Previous methods typically extract head features from manually cropped head images, limiting model inference efficiency and flexibility in the real world. To this end, we design a RoI feature reconstruction module to reconstruct head-related features based on holistic scene features and the head box from the unified Transformer. In this module, we first use the head bounding box " + }, + { + "bbox": [ + 130, + 310, + 481, + 477 + ], + "type": "inline_equation", + "content": "R = (x_{1},y_{1},x_{2},y_{2})" + }, + { + "bbox": [ + 130, + 310, + 481, + 477 + ], + "type": "text", + "content": ", which represents the upper left and lower right corner coordinates, as the region proposal to crop the regions of interest (ROIs). Then, those candidate ROIs are refined by ROAlign [11], which maps the head region to the corresponding pixels in the holistic image features " + }, + { + "bbox": [ + 130, + 310, + 481, + 477 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_4 \\in \\mathbb{R}^{2048 \\times 7 \\times 7}" + }, + { + "bbox": [ + 130, + 310, + 481, + 477 + ], + "type": "text", + "content": " and produces a fixed-size feature map " + }, + { + "bbox": [ + 130, + 310, + 481, + 477 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{head}} \\in \\mathbb{R}^{2048 \\times 7 \\times 7}" + }, + { + "bbox": [ + 130, + 310, + 481, + 477 + ], + "type": "text", + "content": ". For head location map generation, following previous works [9,10,29,41] to generate a binary image, setting the head area to 1 and the rest to 0. Additionally, a specific residual block extracts gaze saliency features " + }, + { + "bbox": [ + 130, + 310, + 481, + 477 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{gaze}} \\in \\mathbb{R}^{1024 \\times 7 \\times 7}" + }, + { + "bbox": [ + 130, + 310, + 481, + 477 + ], + "type": "text", + "content": ", further enhancing the model's attention to the gaze direction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 498, + 318, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 318, + 511 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 318, + 511 + ], + "type": "text", + "content": "3.4 Space-to-object gaze regression" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 479, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 479, + 580 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 479, + 580 + ], + "type": "text", + "content": "We propose a space-to-object gaze regression approach to gradually refine the human-object gaze relationship. Fig. 2(c) shows our gaze regression branch with a dual attention fusion module establishing an initial human-object spatial connection and a feature interaction module strengthening this connection using semantically clear feature interaction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 479, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 479, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 479, + 640 + ], + "type": "text", + "content": "Spatial perception for gaze objects. When humans infer what another person is staring at, they usually first observe the gaze direction of the others and search for the gazed object along this direction. Inspired by this, we propose the dual attention fusion module for the spatial perception of gaze objects, establishing an initial human-object spatial connection." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "As shown in Fig. 4(a), we first use global average pooling to obtain the global features of " + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{head}}" + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": ", followed by fully connected layers to transform the features into" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 447, + 103 + ], + "type": "text", + "content": "Boosting Gaze Object Prediction via Pixel-level Supervision" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 113, + 298, + 193 + ], + "blocks": [ + { + "bbox": [ + 157, + 113, + 298, + 193 + ], + "lines": [ + { + "bbox": [ + 157, + 113, + 298, + 193 + ], + "spans": [ + { + "bbox": [ + 157, + 113, + 298, + 193 + ], + "type": "image", + "image_path": "aee492adf4c07b0548b42361e9eb28b1f882d8fd7d163065ef5b5ad4e38dcb3a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 200, + 423, + 211 + ], + "lines": [ + { + "bbox": [ + 189, + 200, + 423, + 211 + ], + "spans": [ + { + "bbox": [ + 189, + 200, + 423, + 211 + ], + "type": "text", + "content": "Fig. 4: Illustration of spatial perception for gaze objects." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 308, + 115, + 455, + 193 + ], + "blocks": [ + { + "bbox": [ + 308, + 115, + 455, + 193 + ], + "lines": [ + { + "bbox": [ + 308, + 115, + 455, + 193 + ], + "spans": [ + { + "bbox": [ + 308, + 115, + 455, + 193 + ], + "type": "image", + "image_path": "cdd85fdf43e9d108402b9f524e92bb64f9a24f8526c299371dbb66ebc2b77697.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 234, + 481, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 234, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 131, + 234, + 481, + 258 + ], + "type": "text", + "content": "a 2D gaze vector " + }, + { + "bbox": [ + 131, + 234, + 481, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_g" + }, + { + "bbox": [ + 131, + 234, + 481, + 258 + ], + "type": "text", + "content": ". To achieve spatial perception, we project it into a gaze cone map " + }, + { + "bbox": [ + 131, + 234, + 481, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{cone}} \\in \\mathbb{R}^{1 \\times 7 \\times 7}" + }, + { + "bbox": [ + 131, + 234, + 481, + 258 + ], + "type": "text", + "content": ", activating the area around the gaze direction:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 239, + 267, + 481, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 267, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 239, + 267, + 481, + 281 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {\\mathrm {i}, \\mathrm {j}} ^ {\\text {c o n e}} = M a x (0, \\cos (\\mathbf {v} _ {\\mathrm {g}}, \\mathbf {v} _ {\\mathrm {i}, \\mathrm {j}})) \\tag {5}", + "image_path": "e7d0cad0276cfd2937cb895b37ff17a87d0ea9abc865eb3120767c2fcf98de9b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 289, + 482, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 289, + 482, + 337 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 482, + 337 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 289, + 482, + 337 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{g}}" + }, + { + "bbox": [ + 131, + 289, + 482, + 337 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 289, + 482, + 337 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{i,j}}" + }, + { + "bbox": [ + 131, + 289, + 482, + 337 + ], + "type": "text", + "content": " respectively denotes the predicted gaze vector and the vector composed of the eye point and each point on the gaze field. During training, " + }, + { + "bbox": [ + 131, + 289, + 482, + 337 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{g}}" + }, + { + "bbox": [ + 131, + 289, + 482, + 337 + ], + "type": "text", + "content": " is optimized by " + }, + { + "bbox": [ + 131, + 289, + 482, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dir}}" + }, + { + "bbox": [ + 131, + 289, + 482, + 337 + ], + "type": "text", + "content": " to obtain a more accurate gaze direction and activation map. " + }, + { + "bbox": [ + 131, + 289, + 482, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dir}}" + }, + { + "bbox": [ + 131, + 289, + 482, + 337 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 255, + 345, + 481, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 345, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 255, + 345, + 481, + 376 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {d i r}} = 1 - \\sum_ {i = 1} ^ {n} \\left(\\mathbf {v} _ {\\mathrm {g}} ^ {i} \\cdot \\mathbf {v} _ {\\mathrm {t}} ^ {i}\\right) \\tag {6}", + "image_path": "8320a09b6d7daaef32f8ff80c09c7111f2075ba781a0578fb2d084ee0ce9683a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 384, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 384, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 131, + 384, + 482, + 468 + ], + "type": "text", + "content": "where vector " + }, + { + "bbox": [ + 131, + 384, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{g}}" + }, + { + "bbox": [ + 131, + 384, + 482, + 468 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 384, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{t}}" + }, + { + "bbox": [ + 131, + 384, + 482, + 468 + ], + "type": "text", + "content": " denote the predicted gaze direction and ground truth, respectively. During backpropagation, " + }, + { + "bbox": [ + 131, + 384, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dir}}" + }, + { + "bbox": [ + 131, + 384, + 482, + 468 + ], + "type": "text", + "content": " adjusts parameters of head feature " + }, + { + "bbox": [ + 131, + 384, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{head}}" + }, + { + "bbox": [ + 131, + 384, + 482, + 468 + ], + "type": "text", + "content": ", providing more direction information for gaze regression. To boost scene attention, we adopt the method from GaTector [41] to generate an original attention map " + }, + { + "bbox": [ + 131, + 384, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{ori}} \\in \\mathbb{R}^{1 \\times 7 \\times 7}" + }, + { + "bbox": [ + 131, + 384, + 482, + 468 + ], + "type": "text", + "content": ", activating scene salience area based on head properties. By multiplying " + }, + { + "bbox": [ + 131, + 384, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{cone}}" + }, + { + "bbox": [ + 131, + 384, + 482, + 468 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 131, + 384, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{ori}}" + }, + { + "bbox": [ + 131, + 384, + 482, + 468 + ], + "type": "text", + "content": " via element-wise dot product, we generate the final spatial perception map " + }, + { + "bbox": [ + 131, + 384, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 131, + 384, + 482, + 468 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 262, + 476, + 481, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 476, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 262, + 476, + 481, + 490 + ], + "type": "interline_equation", + "content": "\\mathbf {M} = \\mathbf {M} _ {\\text {o r i}} \\odot \\mathbf {M} _ {\\text {c o n e}} \\tag {7}", + "image_path": "7230f77f9d303ba1585cabba69aaeca78ec17b4f3a5250ef9fa2d16fa7d5cdfb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 498, + 479, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 479, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 479, + 521 + ], + "type": "text", + "content": "This compels the model to search spaces more likely gazed at by humans along gaze direction, establishing an initial human-object gaze spatial relationship." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 521, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 480, + 593 + ], + "type": "text", + "content": "Semantically clear feature interaction. Previous methods [40, 41] usually directly fuse features to regress the gaze heatmap based on spatial saliency, lacking interaction with object details and facing challenges in accurately predicting heatmap for the gazed object. In our method, we introduce a feature interaction module, which leverages semantically clear mask knowledge from VFM to achieve feature interaction, refining human-object gaze relationship modeling." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "As shown in Fig. 2(c), we initially extract detailed mask features from the mask branch. Subsequently, the fused feature " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{fuse}}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " is utilized for pixel-level attention perception. Given that candidate gaze objects can be located anywhere in the image, the Transformer's global modeling excels in accomplishing this task. Therefore, utilize a transformer layer, akin to DETR [3], to capture the global relationship between human gaze behavior and object masks. For fine-grained" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Jin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "content": "object mask features, a three-layer MLP is applied to process the decoder query " + }, + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{\\mathrm{obj}}" + }, + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "content": " of the unified Transformer, generating mask embeddings:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 261, + 147, + 481, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 147, + 481, + 159 + ], + "spans": [ + { + "bbox": [ + 261, + 147, + 481, + 159 + ], + "type": "interline_equation", + "content": "\\mathcal {Q} _ {\\text {m a s k}} = \\mathbf {M L P} \\left(\\mathcal {Q} _ {\\text {o b j}}\\right) \\tag {8}", + "image_path": "a6d90fd4b2ef575933de295efe423555718c8d16e83519befeb633121e7eba5a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 164, + 479, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 479, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 479, + 189 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 130, + 164, + 479, + 189 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{\\mathrm{mask}}" + }, + { + "bbox": [ + 130, + 164, + 479, + 189 + ], + "type": "text", + "content": " denote the generated mask embedding. We first apply a layer of self-attention for " + }, + { + "bbox": [ + 130, + 164, + 479, + 189 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{fuse}}" + }, + { + "bbox": [ + 130, + 164, + 479, + 189 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 247, + 195, + 481, + 208 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 195, + 481, + 208 + ], + "spans": [ + { + "bbox": [ + 247, + 195, + 481, + 208 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {\\mathrm {e}} = \\mathbf {F F N} (\\mathbf {S e l f A t t n} (\\mathbf {f} _ {\\mathrm {f u s e}})) \\tag {9}", + "image_path": "9438d38fe9eadad2632fb6bc53eb4b08432c582b6ea526f230e95a67245fd9b0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 213, + 479, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 479, + 250 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 479, + 250 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 213, + 479, + 250 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{fuse}}" + }, + { + "bbox": [ + 130, + 213, + 479, + 250 + ], + "type": "text", + "content": ", SelfAttn(" + }, + { + "bbox": [ + 130, + 213, + 479, + 250 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 130, + 213, + 479, + 250 + ], + "type": "text", + "content": "), and " + }, + { + "bbox": [ + 130, + 213, + 479, + 250 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{e}}" + }, + { + "bbox": [ + 130, + 213, + 479, + 250 + ], + "type": "text", + "content": " denote the fused feature " + }, + { + "bbox": [ + 130, + 213, + 479, + 250 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{fuse}}" + }, + { + "bbox": [ + 130, + 213, + 479, + 250 + ], + "type": "text", + "content": ", self-attention operation, and encoded " + }, + { + "bbox": [ + 130, + 213, + 479, + 250 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{fuse}}" + }, + { + "bbox": [ + 130, + 213, + 479, + 250 + ], + "type": "text", + "content": ", respectively. Then, the " + }, + { + "bbox": [ + 130, + 213, + 479, + 250 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{\\mathrm{mask}}" + }, + { + "bbox": [ + 130, + 213, + 479, + 250 + ], + "type": "text", + "content": " is perceived by the " + }, + { + "bbox": [ + 130, + 213, + 479, + 250 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{e}}" + }, + { + "bbox": [ + 130, + 213, + 479, + 250 + ], + "type": "text", + "content": " as key-value pairs, i.e., " + }, + { + "bbox": [ + 130, + 213, + 479, + 250 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{K}_{\\mathrm{mask}}, \\mathbf{V}_{\\mathrm{mask}}\\}" + }, + { + "bbox": [ + 130, + 213, + 479, + 250 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 207, + 255, + 481, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 255, + 481, + 268 + ], + "spans": [ + { + "bbox": [ + 207, + 255, + 481, + 268 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {\\text {r e g}} = \\mathbf {F F N} (\\mathbf {C r o s s A t t n} (\\mathbf {f} _ {\\mathrm {e}}, \\{\\mathbf {K} _ {\\text {m a s k}}, \\mathbf {V} _ {\\text {m a s k}} \\})) \\tag {10}", + "image_path": "bfc0c62e11eb39cdb590a6c1d04d3ee52d762a5bec1f45446cde95ef7eeb6e6e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 273, + 479, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 273, + 479, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 273, + 479, + 332 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 273, + 479, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 130, + 273, + 479, + 332 + ], + "type": "text", + "content": " and CrossAttn " + }, + { + "bbox": [ + 130, + 273, + 479, + 332 + ], + "type": "inline_equation", + "content": "(q, kv)" + }, + { + "bbox": [ + 130, + 273, + 479, + 332 + ], + "type": "text", + "content": " denote the regression feature and cross-attention operation, respectively. This operation refines the perception of the object by providing a pixel-level understanding of the gaze object in each token of the regression feature. It effectively alleviates semantic ambiguity and strengthens the ability of the model to represent the human-object gaze relationship." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 332, + 479, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 479, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 479, + 404 + ], + "type": "text", + "content": "Ultimate gaze regression. Finally, following previous work [41] to generate the ground truth gaze heatmap and use " + }, + { + "bbox": [ + 130, + 332, + 479, + 404 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 332, + 479, + 404 + ], + "type": "text", + "content": " loss to supervise gaze heatmap regression. " + }, + { + "bbox": [ + 130, + 332, + 479, + 404 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{eng}}" + }, + { + "bbox": [ + 130, + 332, + 479, + 404 + ], + "type": "text", + "content": " proposed by [40] guides gaze heatmap energy aggregation towards the gaze object. We use the gaze object mask for more accurate error measurements. Space-to-object gaze regression approach gradually refines the human-object gaze behavior representation, regressing a more accurate gaze heatmap." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 420, + 230, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 230, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 230, + 434 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 442, + 268, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 268, + 454 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 268, + 454 + ], + "type": "text", + "content": "4.1 Experiments Settings" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 459, + 479, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 479, + 626 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 479, + 626 + ], + "type": "text", + "content": "Datasets and implementation details are given in supplementary material. Metrics: Following previous work [40,41], we use AUC (Area Under the Curve), L2 distance, and Angle error for gaze estimation evaluation, and mSoC (min Single over Closure)4 under different thresholds for gaze object prediction evaluation. Following MaskDINO [20], AP (Average Precision) is used for instance segmentation and object detection evaluation. To evaluate the gaze object segmentation, we use masks to replace boxes when calculating mSoC (details refer to the supplementary material). For a fair comparison, we reported results for both real-world and non-real-world settings. The non-real-world settings involve inputting additional head images and head location maps while real-world settings only use scene images. Furthermore, to compare with GTR [39], we also adopt the evaluation method proposed by [39]5. This method computes gaze estimation metrics only for results with L2 distance " + }, + { + "bbox": [ + 130, + 459, + 479, + 626 + ], + "type": "inline_equation", + "content": "< 0.15" + }, + { + "bbox": [ + 130, + 459, + 479, + 626 + ], + "type": "text", + "content": " and head box IoU " + }, + { + "bbox": [ + 130, + 459, + 479, + 626 + ], + "type": "inline_equation", + "content": "> 0.5" + }, + { + "bbox": [ + 130, + 459, + 479, + 626 + ], + "type": "text", + "content": ", which is an incomplete evaluation for AUC, L2 distance, and Angle error." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 447, + 103 + ], + "type": "text", + "content": "Boosting Gaze Object Prediction via Pixel-level Supervision" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 631, + 480, + 643 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 631, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 631, + 480, + 643 + ], + "type": "text", + "content": "4 mSoC metric proposed by Wang et al. URL: https://arxiv.org/abs/2112.03549." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 136, + 643, + 480, + 664 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 643, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 136, + 643, + 480, + 664 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 136, + 643, + 480, + 664 + ], + "type": "text", + "content": " Evaluation method proposed by Tu et al., URL: https://ieeexplore.ieee.org/document/10262016." + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 141, + 146, + 471, + 236 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 1: Gaze object detection results on GOO-Synth and GOO-Real datasets. * indicates using additional head detector to achieve real-world settings." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 146, + 471, + 236 + ], + "lines": [ + { + "bbox": [ + 141, + 146, + 471, + 236 + ], + "spans": [ + { + "bbox": [ + 141, + 146, + 471, + 236 + ], + "type": "table", + "html": "
MethodGOO-SynthGOO-Real No PretrainGOO-Real PretrainParamsFPS
mSoC507595mSoC507595mSoC507595
Non-RealGaTector [41]67.9498.1486.250.1262.4095.1073.500.2071.2097.5088.802.7060.78M14.1
TransGOP [40]92.8099.0098.5051.9082.6097.8089.406.5089.0098.9097.5033.2094.03M13.7
Ours91.1993.9793.4673.2781.7598.7896.437.1381.8998.9996.637.2880.99M14.5
RealGaTector [41]*56.1988.0866.18-40.8782.0533.540.3066.4096.2280.490.4075.73M9.3
TransGOP [40]*91.6099.0098.1048.7079.7097.8091.4011.2083.2098.8095.3014.40106.37M11.6
Ours90.7793.9293.4569.5881.3598.9596.022.6181.6099.0396.546.0384.26M14.1
", + "image_path": "6939b4e3815121082d710f7ae7faf4f9f8226fc9995f3c520679c1963ba2e42f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 142, + 268, + 471, + 319 + ], + "blocks": [ + { + "bbox": [ + 143, + 246, + 470, + 258 + ], + "lines": [ + { + "bbox": [ + 143, + 246, + 470, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 246, + 470, + 258 + ], + "type": "text", + "content": "Table 2: Ours gaze object segmentation results on GOO-Real and GOO-Synth." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 142, + 268, + 471, + 319 + ], + "lines": [ + { + "bbox": [ + 142, + 268, + 471, + 319 + ], + "spans": [ + { + "bbox": [ + 142, + 268, + 471, + 319 + ], + "type": "table", + "html": "
SettingsGOO-SynthGOO-Real No PretrainGOO-Real PretrainParams|FPS
mSoC507595mSoC507595mSoC507595
Non-Real82.6893.7091.6826.8386.3998.9597.4420.0789.3998.8098.0233.0380.99M
Real82.7393.6791.7926.8585.9498.9997.2916.7086.9698.9497.7522.2184.26M
", + "image_path": "7cbb47521c2f3b69e511518bd9ff3188aa0095d05b7730f6960ad3e64aa1897c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 151, + 330, + 315, + 376 + ], + "blocks": [ + { + "bbox": [ + 151, + 330, + 315, + 376 + ], + "lines": [ + { + "bbox": [ + 151, + 330, + 315, + 376 + ], + "spans": [ + { + "bbox": [ + 151, + 330, + 315, + 376 + ], + "type": "image", + "image_path": "6793ddd6a6bbc7d9cf83f6932f142bd4de3c417a0116edb813388a1f0217cd64.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 385, + 474, + 397 + ], + "lines": [ + { + "bbox": [ + 138, + 385, + 474, + 397 + ], + "spans": [ + { + "bbox": [ + 138, + 385, + 474, + 397 + ], + "type": "text", + "content": "Fig. 5: Comparison of TransGOP [40] and our method inference in the real world." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 320, + 331, + 463, + 376 + ], + "blocks": [ + { + "bbox": [ + 320, + 331, + 463, + 376 + ], + "lines": [ + { + "bbox": [ + 320, + 331, + 463, + 376 + ], + "spans": [ + { + "bbox": [ + 320, + 331, + 463, + 376 + ], + "type": "image", + "image_path": "639656f757280972a12e0fd517dee9c05df1976bbc2c10c95b56bcb3c05cbe5e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 419, + 282, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 282, + 432 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 282, + 432 + ], + "type": "text", + "content": "4.2 Comparison with SOTA" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": "Gaze object detection and segmentation. Table 1 shows the performance of gaze object detection on GOO-Synth and GOO-Real. In real-world settings, previous methods [40, 41] significantly increase parameters and slow down the inference speed due to the additional head detector (As shown in Fig. 5 (a)) and they can only achieve box-level prediction. In contrast, although our method is slightly inferior to TransGOP [40] in detection performance, our model runs significantly more efficiently because it integrates the acquisition of head features into the overall framework (As shown in Fig. 5 (b)) and can output bounding boxes and segmentation masks simultaneously. The above results also show that our space-to-object gaze regression method can effectively model the human-object gaze relationship. Table 2 shows our gaze object segmentation performance in real-world and non-real-world scenarios. On GOO-Synth, our model achieves " + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "82.68\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "82.73\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": " mSoC in real-world and non-real-world settings. On GOO-Real, without pre-training, our model performs well in both scenarios " + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "(86.39\\% \\mathrm{mSoC}" + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "85.94\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": " mSoC). After pre-training, it reaches " + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "89.39\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": " mSoC and " + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "86.96\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": " mSoC. The above experimental results show that the pixel-level mask knowledge provided by VFMs can effectively alleviate semantic ambiguity while achieving pixel-level prediction. Even under the multi-task learning framework, we still achieve an FPS of 14.1. This is mainly attributed to the automatic" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Jin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 159, + 136, + 455, + 208 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 470, + 126 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 470, + 126 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 470, + 126 + ], + "type": "text", + "content": "Table 3: Instance segmentation results on GOO-Synth and GOO-Real datasets" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 136, + 455, + 208 + ], + "lines": [ + { + "bbox": [ + 159, + 136, + 455, + 208 + ], + "spans": [ + { + "bbox": [ + 159, + 136, + 455, + 208 + ], + "type": "table", + "html": "
MethodGOO-Synth DatasetGOO-Real DatasetParams
AP507595AP507595
Non-RealOurs79.3393.5191.0911.7282.9198.8595.0911.2280.99M
RealMaskDINO [20]70.8595.5483.311.8177.0598.3890.454.2652.03M
Ours78.5696.0090.568.4382.8699.0295.5710.0984.26M
", + "image_path": "dcce279da873f4d0f497e8d0a51ffbcd8506684679294d1c2ec6bbad076f5654.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 159, + 251, + 455, + 364 + ], + "blocks": [ + { + "bbox": [ + 132, + 219, + 481, + 243 + ], + "lines": [ + { + "bbox": [ + 132, + 219, + 481, + 243 + ], + "spans": [ + { + "bbox": [ + 132, + 219, + 481, + 243 + ], + "type": "text", + "content": "Table 4: Object detection results on GOO-Synth and GOO-Real datasets. * indicates using additional head detector to achieve real-world settings." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 251, + 455, + 364 + ], + "lines": [ + { + "bbox": [ + 159, + 251, + 455, + 364 + ], + "spans": [ + { + "bbox": [ + 159, + 251, + 455, + 364 + ], + "type": "table", + "html": "
MethodGOO-Synth DatasetGOO-Real DatasetParams
AP507595AP507595
Non-RealGaTector [41]56.8095.3062.500.1052.2591.9255.340.1060.78M
TransGOP [40]87.6099.0097.3025.5077.2097.8089.406.5094.03M
Ours88.3393.8893.0850.7077.3298.6593.911.9380.99M
RealGaTector [41]*43.5086.2337.52-29.2174.0216.13-75.73M
TransGOP [40]*86.0098.8096.2023.8073.3096.7086.004.10106.37M
MaskDINO [20]72.9396.2688.132.0474.4498.7291.190.1952.03M
Ours87.6093.7273.0745.5375.2098.9493.200.7684.26M
", + "image_path": "e2a0309ef89fe08b463bb6b863e67a860be4c5015c5f48dea4e3b6d451e90e34.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 388, + 480, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 480, + 411 + ], + "type": "text", + "content": "incorporation of head features, avoiding the need for an additional head detector and thus enhancing the efficiency of the model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 413, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 533 + ], + "type": "text", + "content": "Object detection and segmentation. Tables 3 and 4 show our model's instance segmentation and object detection performance. For a fair comparison, we adjusted the input size of MaskDINO [20] to match ours " + }, + { + "bbox": [ + 130, + 413, + 482, + 533 + ], + "type": "inline_equation", + "content": "(224 \\times 224)" + }, + { + "bbox": [ + 130, + 413, + 482, + 533 + ], + "type": "text", + "content": ". In Table 3, our model outperforms MaskDINO by 8.66 AP on GOO-Real (70.85 vs. 79.51) and by 5.81 AP on GOO-Synth (77.05 vs. 82.86) in real-world scenarios, with similar advantages in non-real-world settings of our model. These results show the challenge faced by the original MaskDINO [20] in capturing detailed features in scenes with dense objects and the effectiveness of the pixel shuffle of our model in capturing detailed features. Table 4 further demonstrates the superiority of our model in object detection over previous methods and MaskDINO [20]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "Gaze estimation. Tables 5 and 6 show the gaze estimation performance on GOO-Synth and GOO-Real, respectively. In Table 5, our method outperforms GTR [39] and TransGOP [40] by " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "23.9\\mathrm{mAP}" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "10.6\\mathrm{mAP}" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": ", respectively. Even directly evaluating all prediction results in the real world, our model reduces L2 distance and angle errors compared to TransGOP [40] by 0.015 and 1.8, respectively. Applying non-real-world methods to the real world significantly degrades performance, highlighting our real-world model is more efficient. Table 6 verifies our method on GOO-Real, achieving state-of-the-art results without pretraining. Even with direct inference in a real-world setting, our method outperforms previous non-real-world methods. After pre-training, our model exhibits more pronounced performance advantages. These results show the effectiveness" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "text", + "content": "Boosting Gaze Object Prediction via Pixel-level Supervision" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 157, + 299, + 328 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 298, + 149 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 298, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 298, + 149 + ], + "type": "text", + "content": "Table 5: Gaze estimation results on GOO-Synth dataset. " + }, + { + "bbox": [ + 132, + 114, + 298, + 149 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 132, + 114, + 298, + 149 + ], + "type": "text", + "content": " indicates the evaluation method proposed by [39]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 157, + 299, + 328 + ], + "lines": [ + { + "bbox": [ + 135, + 157, + 299, + 328 + ], + "spans": [ + { + "bbox": [ + 135, + 157, + 299, + 328 + ], + "type": "table", + "html": "
MethodAUCDist.↓Ang.↓mAP
Non-RealGazeFollow [29]0.9290.16233.0-
Lian [22]0.9540.10719.7-
VideoAttention [9]0.9520.07515.1-
GaTector [41]0.9570.07314.9-
GTR [39]0.9600.07114.5-
TransGOP [40]0.9630.07913.3-
Ours0.9470.07214.8-
RealGazeFollow [29]†0.8320.31742.60.468
Lian [22]†0.9030.15328.10.454
VideoAttention [9]†0.9120.14324.50.489
GaTector [41]†0.9180.13924.50.510
GTR [39]†0.9620.06814.20.597
TransGOP [40]†0.9770.07011.50.730
TransGOP [40]0.9450.10620.7-
Ours†0.9780.05710.20.836
Ours0.9380.09118.9-
", + "image_path": "054fc8fb59f66f9311576d9087d90375446ed37787bc8b8af8a7d7abdc6804ef.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 309, + 156, + 477, + 328 + ], + "blocks": [ + { + "bbox": [ + 306, + 114, + 476, + 148 + ], + "lines": [ + { + "bbox": [ + 306, + 114, + 476, + 148 + ], + "spans": [ + { + "bbox": [ + 306, + 114, + 476, + 148 + ], + "type": "text", + "content": "Table 6: Gaze estimation results on GOO-Real dataset. " + }, + { + "bbox": [ + 306, + 114, + 476, + 148 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 306, + 114, + 476, + 148 + ], + "type": "text", + "content": " indicates the evaluation method proposed by [39]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 156, + 477, + 328 + ], + "lines": [ + { + "bbox": [ + 309, + 156, + 477, + 328 + ], + "spans": [ + { + "bbox": [ + 309, + 156, + 477, + 328 + ], + "type": "table", + "html": "
MethodAUCDist.↓Ang.↓mAP
No PretrainNon-RealGazeFollow [29]0.8500.22044.4-
Lian [22]0.8400.32143.5-
VideoAttention [9]0.7960.25251.4-
GaTector [41]0.9270.19639.5-
Tonini [36]0.9180.164--
TransGOP [40]0.9470.09716.7-
Ours0.9430.07812.9-
RealOurs†0.9710.0599.80.810
Ours0.9440.08814.7-
PretrainNon-RealGazeFollow [29]0.9030.19539.8-
Lian [22]0.8900.16832.6-
VideoAttention [9]0.8890.15029.1-
GaTector [41]0.9400.08714.8-
TransGOP [40]0.9570.08114.7-
Ours0.9630.07312.4-
RealOurs†0.9780.0518.80.824
Ours0.9490.08213.8-
", + "image_path": "c9a9871646f06f4ccfe74e5e1696bf7f0f4953b040b76652a08e2f1ab96c984d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 151, + 360, + 462, + 441 + ], + "blocks": [ + { + "bbox": [ + 178, + 338, + 436, + 350 + ], + "lines": [ + { + "bbox": [ + 178, + 338, + 436, + 350 + ], + "spans": [ + { + "bbox": [ + 178, + 338, + 436, + 350 + ], + "type": "text", + "content": "Table 7: Ablation study about each component on GOO-Real" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 360, + 462, + 441 + ], + "lines": [ + { + "bbox": [ + 151, + 360, + 462, + 441 + ], + "spans": [ + { + "bbox": [ + 151, + 360, + 462, + 441 + ], + "type": "table", + "html": "
PS RoI-Rec. DAF FIM EALGaze Object SegmentationInstance SegmentationGaze Estimation
mSoCmSoC50mSoC75APAP50AP75AUCDist.↓Ang.↓
79.0098.4892.5074.2498.0087.620.8540.30454.2
✓84.5098.7596.5680.5398.6893.670.8390.33459.8
✓✓84.6398.6796.7281.7098.8394.660.9100.11420.5
✓✓✓85.4498.9296.9681.5898.7694.290.9190.09516.8
✓✓✓✓85.8998.8097.1482.4598.9294.820.9320.08914.9
✓✓✓✓✓85.9498.9997.2082.8699.0295.570.9440.08814.7
", + "image_path": "d9267b53a3af2e8e67046d817965d7bf97846dcd5b03f065675fb0bfbbdb1f2a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 463, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 489 + ], + "type": "text", + "content": "of our space-to-object gaze regression method in modeling human-object gaze relationships while achieving more accurate gaze heatmap regression." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 503, + 336, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 503, + 336, + 516 + ], + "spans": [ + { + "bbox": [ + 131, + 503, + 336, + 516 + ], + "type": "text", + "content": "4.3 Ablation study and model analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "content": "Ablation study about the effect of each component. In Table 7, we conduct various ablation studies on the GOO-Real dataset. We first establish a baseline using MaskDINO [20] and a gaze regression branch similar to TransGOP [40], and then gradually add the proposed modules and demonstrate their contribution to overall model performance. (i) Pixel shuffle (PS). After adding the pixel shuffle module, instance segmentation performance improved by " + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "inline_equation", + "content": "6.29\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "content": " AP " + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "inline_equation", + "content": "(80.53\\%)" + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "content": " vs " + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "inline_equation", + "content": "74.24\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "content": " compared to baseline, indicating that large feature maps significantly enhance detailed features in dense prediction tasks. (ii) RoI Reconstruction (RoI-Rec.). The RoI reconstruction module reduces angle error by 39.3 and L2 distance by 0.22, demonstrating the importance of head-related features for gaze direction perception and the ability of the module to learn accurate gaze features in real-world settings. (iii) Dual attention fusion (DAF)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Jin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 150, + 135, + 462, + 192 + ], + "blocks": [ + { + "bbox": [ + 194, + 114, + 418, + 125 + ], + "lines": [ + { + "bbox": [ + 194, + 114, + 418, + 125 + ], + "spans": [ + { + "bbox": [ + 194, + 114, + 418, + 125 + ], + "type": "text", + "content": "Table 8: Analysis about dual attention fusion module." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 150, + 135, + 462, + 192 + ], + "lines": [ + { + "bbox": [ + 150, + 135, + 462, + 192 + ], + "spans": [ + { + "bbox": [ + 150, + 135, + 462, + 192 + ], + "type": "table", + "html": "
Original Attention Gaze ConeGaze Object SegmentationInstance SegmentationGaze Estimation
mSoCmSoC50mSoC75APAP50AP75AUCDist.↓Ang.↓
✓85.5198.8697.1782.0898.8794.780.9380.08915.6
✓84.1398.7896.3581.3098.7893.690.9260.09716.3
✓✓85.9498.9997.2082.7599.0195.170.9440.08814.7
", + "image_path": "12ee9406870d81717b4dc856df00e3b3393f8df47eac8cd740ae0d5017fdb234.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 150, + 223, + 462, + 281 + ], + "blocks": [ + { + "bbox": [ + 200, + 202, + 414, + 213 + ], + "lines": [ + { + "bbox": [ + 200, + 202, + 414, + 213 + ], + "spans": [ + { + "bbox": [ + 200, + 202, + 414, + 213 + ], + "type": "text", + "content": "Table 9: Analysis about feature interaction module" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 150, + 223, + 462, + 281 + ], + "lines": [ + { + "bbox": [ + 150, + 223, + 462, + 281 + ], + "spans": [ + { + "bbox": [ + 150, + 223, + 462, + 281 + ], + "type": "table", + "html": "
Encoder QueryDecoder QueryMask EmbeddingGaze Object SegmentationInstance SegmentationGaze Estimation
mSoCmSoC50mSoC75APAP50AP75AUC Dist.↓Ang.↓
✓85.7998.3896.4481.2898.8794.870.9280.095 16.0
✓85.2098.2197.0181.6598.8694.650.9490.093 15.1
✓85.9498.9997.2082.8699.0295.570.9440.088 14.7
", + "image_path": "6c0bb8202cce1d96ff63a2136fc34ab68a9535bf035acb8166021be070f36abb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 304, + 480, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 480, + 422 + ], + "type": "text", + "content": "Introducing the dual attention fusion improves gaze heatmap quality (AUC: 0.919 vs 0.910), demonstrating that perceiving gaze object spatial location beforehand produces more accurate heatmaps and eases regression. (iv) Feature interaction module (FIM). The feature interaction module improves gaze estimation, reducing angle error by 1.9 and L2 distance by 0.006. Gaze object segmentation achieves an " + }, + { + "bbox": [ + 130, + 304, + 480, + 422 + ], + "type": "inline_equation", + "content": "85.89\\mathrm{mSoC}" + }, + { + "bbox": [ + 130, + 304, + 480, + 422 + ], + "type": "text", + "content": ", showing that pixel-level object information enhances feature interaction and human-object gaze relationship modeling. (v) Energy aggregation loss (EAL). Adding energy aggregation loss improves the performance of all three sub-tasks, showing that the gaze object mask effectively guides the regression of the gaze heatmap towards the gaze object." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 423, + 480, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 480, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 480, + 495 + ], + "type": "text", + "content": "Analysis about dual attention fusion module. Table 8 analyzed the effect of different spatial perception methods. Compared to using original attention " + }, + { + "bbox": [ + 130, + 423, + 480, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{ori}}" + }, + { + "bbox": [ + 130, + 423, + 480, + 495 + ], + "type": "text", + "content": " or gaze cone " + }, + { + "bbox": [ + 130, + 423, + 480, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{cone}}" + }, + { + "bbox": [ + 130, + 423, + 480, + 495 + ], + "type": "text", + "content": " alone, our dual attention fusion performs better. This indicates that combining spatial search along the gaze direction with human head-based scene saliency is more effective. Fig. 4(b) shows that element-wise fusion of " + }, + { + "bbox": [ + 130, + 423, + 480, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{ori}}" + }, + { + "bbox": [ + 130, + 423, + 480, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 423, + 480, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{cone}}" + }, + { + "bbox": [ + 130, + 423, + 480, + 495 + ], + "type": "text", + "content": " creates an initial human-object spatial connection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 495, + 481, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 481, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 481, + 591 + ], + "type": "text", + "content": "Analysis about feature interaction module. In Table 9, we evaluate the efficacy of different object information for gaze regression. Injecting the global encoder query of MaskDINO [20] results in " + }, + { + "bbox": [ + 130, + 495, + 481, + 591 + ], + "type": "inline_equation", + "content": "85.79\\%" + }, + { + "bbox": [ + 130, + 495, + 481, + 591 + ], + "type": "text", + "content": " mSoC for gaze object segmentation but poor gaze estimation performance. Refining the decoder query with mask supervision reduces the angle error by 0.9, indicating better object representation. Using mask embedding with mask features during gaze regression achieves optimal performance, demonstrating that semantically clear mask features enhance gaze object positioning and the human-object gaze relationship." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 609, + 261, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 261, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 261, + 620 + ], + "type": "text", + "content": "4.4 Qualitative Results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Fig. 6(a) presents qualitative results in the real-world setting. In object-dense scenes, our method accurately predicts the gaze object mask without relying on head-related cues. In Fig. 6(b), compared to previous methods, our method ac" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "text", + "content": "Boosting Gaze Object Prediction via Pixel-level Supervision" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 309 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 309 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 309 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 309 + ], + "type": "image", + "image_path": "96de6f78ea012101fb96efc55275baa794bc376c136ab1332a48dacc373341f3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 318, + 482, + 343 + ], + "lines": [ + { + "bbox": [ + 132, + 318, + 482, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 482, + 343 + ], + "type": "text", + "content": "Fig. 6: Qualitative comparison of ground truth gaze object (GT), gaze object detection (GOD), instance segmentation (IS), gaze object segmentation (GOS)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 365, + 482, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 391 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 391 + ], + "type": "text", + "content": "curately locates adjacent objects using a pixel-level mask, effectively alleviating semantic ambiguity. See supplementary material for more visualization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 410, + 220, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 410, + 220, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 410, + 220, + 423 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": "In this paper, we introduce the gaze object segmentation task, which aims to infer pixel-level masks for objects captured by human eyesight, and establish a framework based on the pixel-level supervision from VFM to solve this task while alleviating the semantic ambiguity existing in previous box-level supervision methods. In our method, we choose SAM to generate pixel-level supervision information for the model training. The RoI reconstruction module is proposed to reconstruct head features directly from holistic features using the head box provided by the unified Transformer, which improves the inference efficiency and flexibility of our model in the real world. To boost the representation of the human-object gaze relationship, we propose a space-to-object gaze regression strategy, which first uses the dual attention fusion module to establish a human-object spatial connection, and then a semantically clear mask feature is injected into the gaze regression to strengthen the human-object gaze relationship modeling. Extensive experiments on GOO-Real and GOO-Synth demonstrate the effectiveness of our method in various settings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "**Limitation:** The proposed method requires box supervision as prompts for the SAM when generating masks, which limits its application in a wider range of scenarios without strong location priors, more flexible approaches for generating masks via the latest VFM models need to be explored in future work." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Jin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 479, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 479, + 161 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 479, + 161 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China. (No. 82272020)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 162, + 480, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 480, + 185 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 480, + 185 + ], + "type": "text", + "content": "Bin Fan is sponsored by China National Postdoctoral Program for Innovative Talents (No. BX20230013)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 201, + 197, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 201, + 197, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 201, + 197, + 215 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 224, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 224, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 481, + 246 + ], + "type": "text", + "content": "1. Bao, J., Liu, B., Yu, J.: ESCNet: Gaze target detection with the understanding of 3D scenes. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 14126-14135 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 246, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 481, + 278 + ], + "type": "text", + "content": "2. Cai, X., Zeng, J., Shan, S., Chen, X.: Source-free adaptive gaze estimation by uncertainty reduction. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 22035-22045 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 278, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 278, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 138, + 278, + 481, + 310 + ], + "type": "text", + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Eur. Conf. Comput. Vis. pp. 213-229 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 310, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 310, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 310, + 481, + 342 + ], + "type": "text", + "content": "4. Chen, W., Li, R., Yu, Q., Xu, A., Feng, Y., Wang, R., Zhao, L., Lin, Z., Yang, Y., Lin, D., et al.: Early detection of visual impairment in young children using a smartphone-based deep learning system. Nature Medicine 29(2), 493-503 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 342, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 481, + 365 + ], + "type": "text", + "content": "5. Chen, Y., Nan, Z., Xiang, T.: FBLNet: Feedback loop network for driver attention prediction. In: Int. Conf. Comput. Vis. pp. 13371-13380 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 365, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 481, + 386 + ], + "type": "text", + "content": "6. Cheng, Y., Lu, F.: Gaze estimation using transformer. In: Int. Conf. Pattern Recog. pp. 3341-3347 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 386, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 386, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 386, + 481, + 407 + ], + "type": "text", + "content": "7. Cheng, Y., Lu, F., Zhang, X.: Appearance-based gaze estimation via evaluation-guided asymmetric regression. In: Eur. Conf. Comput. Vis. pp. 100-115 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 407, + 481, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 481, + 440 + ], + "type": "text", + "content": "8. Chong, E., Ruiz, N., Wang, Y., Zhang, Y., Rozga, A., Rehg, J.M.: Connecting gaze, scene, and attention: Generalized attention estimation via joint modeling of gaze and scene saliency. In: Eur. Conf. Comput. Vis. pp. 383-398 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 440, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 440, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 440, + 481, + 460 + ], + "type": "text", + "content": "9. Chong, E., Wang, Y., Ruiz, N., Rehg, J.M.: Detecting attended visual targets in video. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5396-5406 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 460, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 481, + 493 + ], + "type": "text", + "content": "0. Gupta, A., Tafasca, S., Odobez, J.M.: A modular multimodal architecture for gaze target prediction: Application to privacy-sensitive settings. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5041-5050 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 493, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 493, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 493, + 481, + 514 + ], + "type": "text", + "content": "1. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask R-CNN. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2961-2969 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 514, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 481, + 546 + ], + "type": "text", + "content": "2. Hu, Z., Yang, D., Cheng, S., Zhou, L., Wu, S., Liu, J.: We know where they are looking at from the RGB-D camera: Gaze following in 3D. IEEE Trans. Instrum. Meas. 71, 1-14 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 547, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 481, + 579 + ], + "type": "text", + "content": "3. Hu, Z., Yang, Y., Zhai, X., Yang, D., Zhou, B., Liu, J.: GFIE: A dataset and baseline for gaze-following from 2D to 3D in indoor environments. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8907-8916 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 579, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 481, + 611 + ], + "type": "text", + "content": "4. Hu, Z., Zhao, K., Zhou, B., Guo, H., Wu, S., Yang, Y., Liu, J.: Gaze target estimation inspired by interactive attention. IEEE Trans. Circuit Syst. Video Technol. 32(12), 8524-8536 (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "text", + "content": "5. Huang, T., Fu, R.: Driver distraction detection based on the true driver's focus of attention. IEEE Trans. Intell. Transp. Syst. 23(10), 19374-19386 (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "6. Jin, S., Wang, Z., Wang, L., Bi, N., Nguyen, T.: ReDirTrans: Latent-to-latent translation for gaze and head redirection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5547-5556 (2023)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "text", + "content": "Boosting Gaze Object Prediction via Pixel-level Supervision" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "17. Jin, T., Yu, Q., Zhu, S., Lin, Z., Ren, J., Zhou, Y., Song, W.: Depth-aware gaze-following via auxiliary networks for robotics. Eng. Appl. Artif. Intell. 113, 104924 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "text", + "content": "18. Kellnhofer, P., Recasens, A., Stent, S., Matusik, W., Torralba, A.: Gaze360: Physically unconstrained gaze estimation in the wild. In: Int. Conf. Comput. Vis. pp. 6912-6921 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 215 + ], + "type": "text", + "content": "19. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Int. Conf. Comput. Vis. pp. 4015-4026 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 481, + 247 + ], + "type": "text", + "content": "20. Li, F., Zhang, H., Xu, H., Liu, S., Zhang, L., Ni, L.M., Shum, H.Y.: Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3041-3050 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 247, + 481, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 481, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 481, + 269 + ], + "type": "text", + "content": "21. Li, Y., Shen, W., Gao, Z., Zhu, Y., Zhai, G., Guo, G.: Looking here or there? gaze following in 360-degree images. In: Int. Conf. Comput. Vis. pp. 3742-3751 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 268, + 481, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 481, + 290 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 481, + 290 + ], + "type": "text", + "content": "22. Lian, D., Yu, Z., Gao, S.: Believe it or not, we know what you are looking at! In: Asian Conf. Comput. Vis. pp. 35-50 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 289, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 481, + 311 + ], + "type": "text", + "content": "23. Lv, K., Sheng, H., Xiong, Z., Li, W., Zheng, L.: Improving driver gaze prediction with reinforced attention. IEEE Trans. Multimedia 23, 4198-4207 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 310, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 481, + 332 + ], + "type": "text", + "content": "24. Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature Communications 15(1), 654 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 332, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 481, + 354 + ], + "type": "text", + "content": "25. Miao, Q., Hoai, M., Samaras, D.: Patch-level gaze distribution prediction for gaze following. In: IEEE Winter Conf. Appl. Comput. Vis. pp. 880-889 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "text", + "content": "26. Mundy, P., Sigman, M., Kasari, C.: A longitudinal study of joint attention and language development in autistic children. J. Autism Dev. Disord. 20(1), 115-128 (1990)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 385, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 481, + 407 + ], + "type": "text", + "content": "27. Park, S., Spurr, A., Hilliges, O.: Deep pictorial gaze estimation. In: Eur. Conf. Comput. Vis. pp. 721-738 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 407, + 481, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 440 + ], + "type": "text", + "content": "28. Park, S., Zhang, X., Bulling, A., Hilliges, O.: Learning to find eye region landmarks for remote gaze estimation in unconstrained settings. In: Proc. ACM Symp. Eye Track. Res. Appl. pp. 1-10 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 440, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 481, + 460 + ], + "type": "text", + "content": "29. Recasens, A., Khosla, A., Vondrick, C., Torralba, A.: Where are they looking? Adv. Neural Inform. Process. Syst. 28 (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 460, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 481, + 482 + ], + "type": "text", + "content": "30. Recasens, A., Vondrick, C., Khosla, A., Torralba, A.: Following gaze in video. In: Int. Conf. Comput. Vis. pp. 1435-1443 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 481, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 481, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 481, + 481, + 515 + ], + "type": "text", + "content": "31. Ruzzi, A., Shi, X., Wang, X., Li, G., De Mello, S., Chang, H.J., Zhang, X., Hilliges, O.: GazeNeRF: 3d-aware gaze redirection with neural radiance fields. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 9676-9685 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 514, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 481, + 536 + ], + "type": "text", + "content": "32. Senju, A., Johnson, M.H.: Atypical eye contact in autism: Models, mechanisms and development. Neurosci. Biobehav. Rev. 33(8), 1204-1214 (2009)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 535, + 481, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 535, + 481, + 580 + ], + "spans": [ + { + "bbox": [ + 132, + 535, + 481, + 580 + ], + "type": "text", + "content": "33. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient subpixel convolutional neural network. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1874-1883 (2016)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "text", + "content": "34. Tafasca, S., Gupta, A., Odobez, J.M.: ChildPlay: A new benchmark for understanding children's gaze behaviour. In: Int. Conf. Comput. Vis. pp. 20935-20946 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "35. Tomas, H., Reyes, M., Dionido, R., Ty, M., Mirando, J., Casimiro, J., Atienza, R., Quinto, R.: Goo: A dataset for gaze object prediction in retail environments. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. pp. 3125-3133 (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "36. Tonini, F., Beyer, C., Ricci, E.: Multimodal across domains gaze target detection. In: Int. Conf. Multimodal Interact. pp. 420-431 (2022)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Y. Jin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 512 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "37. Tonini, F., Dall'Asen, N., Bayan, C., Ricci, E.: Object-aware gaze target detection. In: Int. Conf. Comput. Vis. pp. 21860-21869 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 171 + ], + "type": "text", + "content": "38. Tu, D., Min, X., Duan, H., Guo, G., Zhai, G., Shen, W.: End-to-end human-gazetarget detection with transformers. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2192-2200 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 171, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 171, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 171, + 482, + 205 + ], + "type": "text", + "content": "39. Tu, D., Shen, W., Sun, W., Min, X., Zhai, G., Chen, C.: Un-Gaze: A unified transformer for joint gaze-location and gaze-object detection. IEEE Trans. Circuit Syst. Video Technol. (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 226 + ], + "type": "text", + "content": "40. Wang, B., Guo, C., Jin, Y., Xia, H., Liu, N.: TransGOP: Transformer-based gaze object prediction. AAAI Conf. Artif. Intell. (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 226, + 482, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 258 + ], + "type": "text", + "content": "41. Wang, B., Hu, T., Li, B., Chen, X., Zhang, Z.: GaTector: A unified framework for gaze object prediction. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19588-19597 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 258, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 281 + ], + "type": "text", + "content": "42. Wang, K., Ji, Q.: Real time eye gaze tracking with 3D deformable eye-face model. In: Int. Conf. Comput. Vis. pp. 1003-1011 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 281, + 482, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 315 + ], + "type": "text", + "content": "43. Wang, X., Zhang, H., Wang, Z., Nie, W., Yang, Z., Ren, W., Xu, Q., Xu, X., Liu, H.: Dual regression-enhanced gaze target detection in the wild. IEEE Trans. Cybern. 54(1), 219-229 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "type": "text", + "content": "44. Wang, Z., Zhao, J., Lu, C., Yang, F., Huang, H., Guo, Y., et al.: Learning to detect head movement in unconstrained remote gaze estimation in the wild. In: IEEE Winter Conf. Appl. Comput. Vis. pp. 3443-3452 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 347, + 482, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 380 + ], + "type": "text", + "content": "45. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. Int. Conf. Learn. Represent. (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 380, + 482, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 402 + ], + "type": "text", + "content": "46. Zhang, X., Sugano, Y., Fritz, M., Bulling, A.: Appearance-based gaze estimation in the wild. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4511-4520 (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 402, + 482, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 435 + ], + "type": "text", + "content": "47. Zhang, X., Sugano, Y., Fritz, M., Bulling, A.: Mpiigaze: Real-world dataset and deep appearance-based gaze estimation. IEEE Trans. Pattern Anal. Mach. Intell. 41(1), 162-175 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 435, + 482, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 456 + ], + "type": "text", + "content": "48. Zhao, X., Ding, W., An, Y., Du, Y., Yu, T., Li, M., Tang, M., Wang, J.: Fast segment anything. arXiv preprint arXiv:2306.12156 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 456, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 479 + ], + "type": "text", + "content": "49. Zhu, W., Deng, H.: Monocular free-head 3D gaze tracking with deep learning and geometry constraints. In: Int. Conf. Comput. Vis. pp. 3143-3152 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 479, + 482, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 512 + ], + "type": "text", + "content": "50. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: Int. Conf. Learn. Represent. (2021)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 202, + 91, + 447, + 102 + ], + "type": "text", + "content": "Boosting Gaze Object Prediction via Pixel-level Supervision" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory/d6848265-b938-496f-9e0e-c16ce7da086b_content_list.json b/2024/Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory/d6848265-b938-496f-9e0e-c16ce7da086b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9c72cd5877c27691ed42f18c618288f58069dd61 --- /dev/null +++ b/2024/Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory/d6848265-b938-496f-9e0e-c16ce7da086b_content_list.json @@ -0,0 +1,1947 @@ +[ + { + "type": "text", + "text": "Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory", + "text_level": 1, + "bbox": [ + 222, + 140, + 781, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sensen Gao $^{1,3}$ , Xiaojun Jia $^{2*}$ , Xuhong Ren $^{2}$ , Ivor Tsang $^{2,3}$ , and Qing Guo $^{3\\dagger}$", + "bbox": [ + 320, + 233, + 679, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Nankai University, Tianjin, China", + "bbox": [ + 344, + 276, + 622, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Nanyang Technological University, Singapore", + "bbox": [ + 344, + 290, + 658, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 CFAR and IHPC, Agency for Science, Technology and Research (A*STAR), Singapore", + "bbox": [ + 240, + 305, + 759, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Vision-language pre-training (VLP) models exhibit remarkable capabilities in comprehending both images and text, yet they remain susceptible to multimodal adversarial examples (AEs). Strengthening attacks and uncovering vulnerabilities, especially common issues in VLP models (e.g., high transferable AEs), can advance reliable and practical VLP models. A recent work (i.e., Set-level guidance attack) indicates that augmenting image-text pairs to increase AE diversity along the optimization path enhances the transferability of adversarial examples significantly. However, this approach predominantly emphasizes diversity around the online adversarial examples (i.e., AEs in the optimization period), leading to the risk of overfitting the victim model and affecting the transferability. In this study, we posit that the diversity of adversarial examples towards the clean input and online AEs are both pivotal for enhancing transferability across VLP models. Consequently, we propose using diversification along the intersection region of adversarial trajectory to expand the diversity of AEs. To fully leverage the interaction between modalities, we introduce text-guided adversarial example selection during optimization. Furthermore, to further mitigate the potential overfitting, we direct the adversarial text deviating from the last intersection region along the optimization path, rather than adversarial images as in existing methods. Extensive experiments affirm the effectiveness of our method in improving transferability across various VLP models and downstream vision-and-language tasks. Code is available at https://github.com/SensenGao/VLTPredictAttack.", + "bbox": [ + 259, + 362, + 743, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Vision-Language Attack $\\cdot$ Adversarial Transferability $\\cdot$ Diversification $\\cdot$ Intersection Region of Adversarial Trajectory", + "bbox": [ + 259, + 708, + 740, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 758, + 375, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-language pre-training (VLP) models utilize multimodal learning, leveraging large-scale image-text pairs to bridge the gap between visual and lan", + "bbox": [ + 212, + 787, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* co-first authors; † co-corresponding authors", + "bbox": [ + 217, + 824, + 524, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f6a4ec0e69e781aa3dff3ed1823ef6f5ab5b39259ef55a4b14a441af60845903.jpg", + "image_caption": [ + "(a) SGA" + ], + "image_footnote": [], + "bbox": [ + 274, + 143, + 442, + 253 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1433f9d0752d691d5c98279c201a9d8e7f6b4247bd6c9ca74ec0679249d814a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 143, + 751, + 253 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d4cfe77792677fc82598e233a866afc7b3630619180647168f842c8a85f8716c.jpg", + "image_caption": [ + "(c) ALBEF to $\\mathrm{CLIP}_{\\mathrm{CNN}}$" + ], + "image_footnote": [], + "bbox": [ + 248, + 267, + 493, + 380 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a1b3abab916e308cf1df4006807d87b09cc0a3194075a0ad8300b6fae7cc8653.jpg", + "image_caption": [ + "(b) Ours", + "(d) $\\mathrm{CLIP}_{\\mathrm{ViT}}$ to $\\mathrm{CLIP}_{\\mathrm{CNN}}$", + "Fig. 1: Our method vs. set-level guided attack (SGA) [36]. (a) shows the main idea of SGA, i.e., conducting augmentation around the online adversarial examples. (b) shows the main idea of our method, that is, we perform augmentation in the intersection region of adversarial trajectory. The red and blue dots both depict images sampled from the intersection region, with red dots indicating the best samples selected using the text-guided adversarial example selection strategy. The surrounding light red dots represent applying the same resizing data augmentation to the best samples as utilized in SGA. (c) and (d) compare the transferability of our method and SGA by using the adversarial examples of ALBEF [30] and $\\mathrm{CLIP}_{\\mathrm{ViT}}$ to attack $\\mathrm{CLIP}_{\\mathrm{CNN}}$ , respectively." + ], + "image_footnote": [], + "bbox": [ + 504, + 267, + 753, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "guage understanding. These models showcase revolutionary performance across various downstream Vision-and-Language tasks, including image-text retrieval, image captioning, visual grounding, and visual entailment, as demonstrated in [20, 27, 28, 44]. Despite their success, recent research underscores the significant vulnerability of VLP models, particularly when confronted with multimodal adversarial examples [7, 12, 16, 17, 36, 37, 57]. Uncovering the vulnerabilities, especially common issues, can drive further research aimed at building more reliable and practical VLP models.", + "bbox": [ + 212, + 565, + 787, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Current research predominantly concentrates on attacking VLP models via a white-box setting, where the model's structural information can be exploited. However, exploring the transferability of multimodal adversarial examples is pivotal, especially given the limited access to detailed model structures in real-world scenarios. There have been some efforts made to enhance the transferability of attacks on VLP models by introducing input diversity, as seen in the work SGA [36]. While they have achieved some effectiveness, how to enhance the transferability of multimodal adversarial examples is still an open question.", + "bbox": [ + 212, + 686, + 787, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we undertake a comprehensive examination of the factors contributing to the limited transferability of the cutting-edge multimodal attack", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. Gao et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "method, i.e., SGA [36]. As shown in Figure 1 (a), throughout the iterative generation of subsequent adversarial images, SGA conducts data augmentation around the online adversarial image (i.e., adversarial examples generated during optimization). This strategy enhances the diversity of adversarial examples along the optimization path, leading to a certain improvement in transferability. However, such an approach still carries the potential risk of overfitting to the victim model with the high reliance on the examples along the adversarial trajectory (See Figure 1 (b)), which leads to low attack success rates when we migrate the adversarial example to other VLP models (See Figure 1 (c) and (d)). To mitigate this overfitting risk, one potential solution is to further enhance the diversity of augmented adversarial examples in a judicious manner.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building upon the aforementioned analysis, SGA overfits local adversarial examples, while the clean image is the only accessible example that is far from local adversarial examples. Therefore, we embark on a pioneering endeavor to enhance the transferability of multimodal adversarial attacks by considering the diversity of adversarial examples (AEs) around clean inputs and online AEs throughout the optimization process. To achieve this, we consider the intersection region of adversarial trajectory, which encompasses the original image, the adversarial image from the previous step, and the current adversarial image during the iterative attack process (depicted in Figure 1 (b)). This innovative approach aims to circumvent overfitting by strategic sampling within this region, thereby avoiding an undue focus on adversarial example diversity solely around adversarial images. After obtaining multiple samples, we calculate gradients for each to determine perturbation directions away from the text. Subsequently, we individually incorporate these perturbations into the current adversarial image and select the one that deviates the most from the text.", + "bbox": [ + 212, + 316, + 787, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, in the text modality, SGA only considers deviating the text from the last adversarial image in the optimization period, but the adversarial image is solely generated by the surrogate model, still posing the risk of overfitting the surrogate model. For this reason, we propose to have the text deviate simultaneously from the last intersection region along the optimization path.", + "bbox": [ + 212, + 547, + 787, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed method is evaluated on two widely recognized multimodal datasets, Flickr30K [41] and MSCOCO [35]. We conduct experiments on three vision-and-language downstream tasks (i.e., image-text retrieval (ITR), visual grounding(VG), and image captioning (IC)), and all results indicate the high effectiveness of our method in generating more transferable multimodal adversarial examples. Moreover, when adversarial examples generated from image-text retrieval are transferred to other vision-and-language downstream tasks (i.e., VG and IC), there is a substantial improvement in attack performance.", + "bbox": [ + 212, + 627, + 787, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as follows:", + "bbox": [ + 238, + 752, + 632, + 766 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Boosting Transferability in Vision-Language Attacks", + "bbox": [ + 380, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "- We propose using the intersection region of adversarial trajectory to expand the diversity of adversarial examples during optimization, based on which we develop a high-transferability attack against VLP models.", + "bbox": [ + 225, + 794, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We extend the generation of adversarial text to deviate from the last intersection region along the optimization path, aiming to reduce overfitting the surrogate model, thereby achieving enhanced transferability.", + "bbox": [ + 223, + 146, + 782, + 190 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Extensive experiments robustly demonstrate the efficacy of our proposed method in elevating the transferability of multimodal adversarial examples across diverse models and three downstream tasks.", + "bbox": [ + 225, + 191, + 784, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 261, + 387, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Vision-Language Pre-training Models", + "text_level": 1, + "bbox": [ + 215, + 294, + 571, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VLP models leverage multimodal learning from extensive image-text pairs to improve the performance of various Vision-and-Language $(\\mathrm{V} + \\mathrm{L})$ tasks [29]. Early VLP models predominantly depend on pre-trained object detectors for acquiring multimodal representations [6, 32, 45, 49, 59]. Recently, with the advent of end-to-end image encoders like the Vision Transformer (ViT) [10, 46, 56] offering faster inference speeds, some work propose to use them as substitution for computationally expensive object detectors [11, 29, 30, 48, 53].", + "bbox": [ + 212, + 321, + 784, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There are two popular approaches for VLP models in learning vision-language representations: the fused architecture and the aligned architecture. Fused VLP models (e.g., ALBEF [30], TCL [53]), initially employ two separate unimodal encoders to learn features for text and images. Subsequently, a multimodal encoder is utilized to fuse the embeddings of text and images. In contrast, aligned VLP models, exemplified by CLIP [42], focus on aligning the feature spaces of distinct unimodal encoders and benefit downstream tasks significantly [1]. This paper concentrates on evaluating our proposed method using multiple popular fused and aligned VLP models.", + "bbox": [ + 212, + 428, + 785, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Downstream Vision-and-Language Tasks", + "text_level": 1, + "bbox": [ + 215, + 585, + 594, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image-Text Retrieval (ITR) involves retrieving pertinent information, textual or visual, in response to queries from another modality [5, 8, 51, 60]. This undertaking usually encompasses two sub-tasks: image-to-text retrieval (retrieving text based on an image query) and text-to-image retrieval (retrieving images given a text query).", + "bbox": [ + 212, + 613, + 784, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In aligned VLP models, both the Text Retrieval (TR) and Image Retrieval (IR) tasks leverage ranking results determined by the similarity between text and image embeddings. However, in fused VLP models, where internal embedding spaces lack alignment across unimodal encoders, the similarity scores between image and text modalities are computed for all image-text pairs to retrieve the Top-N candidates. Subsequently, these Top-N candidates serve as input for the multimodal encoder, which computes the image-text matching score to establish the final ranking.", + "bbox": [ + 212, + 688, + 784, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual Grounding (VG) refers to the task of localizing the region within a visual scene with corresponding entities or concepts in natural language. Among", + "bbox": [ + 214, + 809, + 784, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. Gao et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/688911c98308805d427f5d6efa63fa10c5bd25e52e789e82eee9266838fba66d.jpg", + "table_caption": [ + "Table 1: Attack Success Rate (%) of SGA with and without image augmentation. The SGA w.o. Aug doesn't consider image augmentation. We use ALBEF to generate multimodal adversarial examples on the ITR task to evaluate transferability." + ], + "table_footnote": [], + "table_body": "
SourceAttackALBEFTCLCLIPvITCLIPCNN
TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1
ALBEFSGA w.o. Aug99.999.9570.0771.6730.5539.8832.3142.54
SGA w. Aug99.999.9887.8888.0536.6946.7839.5949.78
", + "bbox": [ + 246, + 210, + 759, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "VLP models, ALBEF expands Grad-CAM [43] and utilizes the acquired attention map to rank the detected proposals [54].", + "bbox": [ + 215, + 300, + 784, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Captioning (IC) is to generate a textual description that logically describes or implies the content of a given visual input, typically involving the creation of captions for images. The evaluation of Image Captioning models often employs metrics such as BLEU [39], METEOR [4], ROUGE [33], CIDEr [47] and SPICE [2], which serve to assess the quality and relevance of the generated captions in comparison to reference captions.", + "bbox": [ + 215, + 330, + 785, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.3 Transferability of Adversarial Examples", + "text_level": 1, + "bbox": [ + 215, + 444, + 588, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adversarial attacks [13-15, 18, 25] are typically categorized as white-box and black-box attacks. In a white-box setting [23, 26], the attacker has full access to the model, whereas black-box attacks [3, 40], more realistic in practical applications, occur when information about the model is limited. In the realm of image attacks [21, 22, 24], prevalent methods for crafting transferable adversarial examples often leverage data augmentation techniques (e.g., DIM [52], TIM [9], SIM [34], ADMIX [50], PAM [58]). Zhang et al. [57] introduced a white-box attack targeting popular VLP models for downstream tasks in the multimodal domain. Building upon this work, Lu et al. [36] proposed SGA, considering the diversity of adversarial examples by expanding single image-text pairs to sets of images and texts to conduct black-box attacks on VLP models.", + "bbox": [ + 215, + 470, + 787, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, SGA primarily emphasizes diversity in the vicinity of adversarial examples during the optimization process, potentially increasing the risk of overfitting the victim model and impacting transferability. Therefore, our primary focus in this study is to further enhance the diversity of adversarial examples in a thoughtful manner and avoid an undue focus on adversarial example diversity solely around adversarial images.", + "bbox": [ + 215, + 637, + 785, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 750, + 380, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Background and Motivation", + "text_level": 1, + "bbox": [ + 215, + 782, + 493, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adversarial attacks on VLP models involve inducing a mismatch between adversarial images and corresponding adversarial text while adhering to specified", + "bbox": [ + 215, + 809, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Boosting Transferability in Vision-Language Attacks", + "bbox": [ + 380, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "constraints on image and text perturbations. Here, $(v,t)$ denotes an original image-text pair from a multimodal dataset, with $v^{\\prime}$ representing an adversarial image and $t^\\prime$ denoting adversarial text. The allowable perturbations are restricted within the ranges $B[v,\\xi_v]$ for images and $B[t,\\xi_t]$ for text. The image and text encoders of the multimodal model are denoted as $F_{I}$ and $F_{T}$ , respectively. To generate valid multimodal adversarial examples, the objective is to maximize the loss function $J$ specific to VLP models:", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\max J \\left(F _ {I} \\left(v ^ {\\prime}\\right), F _ {T} \\left(t ^ {\\prime}\\right)\\right) \\\\ s. t. v ^ {\\prime} \\in B [ v, \\xi_ {v} ], t ^ {\\prime} \\in B [ t, \\xi_ {t} ]. \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 263, + 785, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The state-of-the-art approach for exploring the transferability of multimodal adversarial examples (i.e., SGA [36]) involves augmenting image-text pairs to enhance the diversity of adversarial examples along the optimization path. Specifically, during the iterative generation of adversarial images, let $v_{i}^{\\prime}$ represent the adversarial image generated at the $i$ -th step. In the subsequent step $(i + 1)$ , SGA initiates the process by applying a resizing operation for data augmentation to $v_{i}^{\\prime}$ , resulting in $V_{i}^{\\prime} = \\{v_{i1}^{\\prime}, v_{i2}^{\\prime}, \\dots, v_{iM}^{\\prime}\\}$ (See Figure 1 (a)). The iterative formula can be expressed as follows:", + "bbox": [ + 214, + 306, + 787, + 429 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nv _ {i + 1} ^ {\\prime} = v _ {i} ^ {\\prime} + \\alpha \\cdot \\operatorname {s i g n} \\left(\\frac {\\nabla_ {v} \\sum_ {j = 1} ^ {M} J \\left(F _ {I} \\left(v _ {i j} ^ {\\prime}\\right) , F _ {T} (t)\\right)}{\\left\\| \\nabla_ {v} \\sum_ {j = 1} ^ {M} J \\left(F _ {I} \\left(v _ {i j} ^ {\\prime}\\right) , F _ {T} (t)\\right) \\right\\|}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 320, + 440, + 785, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further examine the impact of image augmentation along the optimization path in the SGA method, we utilize ALBEF as a surrogate model to generate multimodal adversarial examples. These examples are then employed to target VLP models such as TCL and CLIP, assessing the transferability of the attacks. Detailed results are presented in Table 1. Our observations reveal that SGA enhances the transferability of adversarial attacks, showing an increase ranging from $6.14\\%$ to $17.81\\%$ . However, it is noteworthy that the success rate of attacks on the target models remains notably lower than that on the source model. This discrepancy is primarily attributed to the fact that SGA predominantly emphasizes diversity around AE $v_{i}^{\\prime}$ during the optimization period, without adequately considering the diversity of adversarial examples toward the clean image, leading to the risk of overfitting the victim model and affecting the transferability.", + "bbox": [ + 212, + 492, + 787, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For this purpose, we propose to consider diversification along the intersection region of adversarial trajectory, which encompasses the original image $v$ , the adversarial image from the previous step $v_{i-1}'$ , and the current adversarial image $v_i'$ during the iterative attack process. This region is established to sample images within it to broaden the diversity of adversarial examples (See Figure 1 (b)). Moreover, to fully leverage the interplay between modalities, we aim for perturbations guided by textual information that induce $v_i'$ to deviate significantly from the associated text $t$ . Additionally, in the text modality, our objective is to identify adversarial perturbations that simultaneously deviate from the intersection region rather than only adversarial images, thereby reducing overfitting the surrogate model and enhancing the effectiveness of black-box attacks.", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. Gao et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Diversification along the Intersection Region", + "text_level": 1, + "bbox": [ + 215, + 146, + 630, + 162 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As outlined in Section 3.1, we enhance the diversity of adversarial examples by introducing diversification along the intersection region. Specifically, at the $i$ th iteration during optimization, we have the $v_{i}^{\\prime}$ , $v_{i-1}^{\\prime}$ , and the clean $v$ , and these variables form a triangle region denoted as $\\triangle vv_{i-1}^{\\prime}v_{i}^{\\prime}$ , i.e., the intersection region of adversarial trajectory in Figure 1. Then, we initially sample multiple instances within the region $\\triangle vv_{i-1}^{\\prime}v_{i}^{\\prime}$ , representing the set of samples as $e = \\{e_1,e_2,\\dots,e_N\\}$ . Each sample can be expressed as $e_k = \\beta \\cdot v + \\gamma \\cdot v_{i-1}^{\\prime} + \\eta \\cdot v_i^{\\prime}$ where $\\beta + \\gamma + \\eta = 1.0$ . Consequently, we can compute the gradient perturbation for each sample. For the k-th sample, denoted as $e_k$ , its gradient perturbation $p_k$ is calculated as follows. In this way, we can get a perturbation set $P = \\{p_1,p_2,\\dots,p_N\\}$ by", + "bbox": [ + 212, + 171, + 787, + 338 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np _ {k} = \\alpha \\cdot \\operatorname {s i g n} \\left(\\frac {\\nabla_ {e} J \\left(F _ {I} \\left(e _ {k}\\right) , F _ {T} (t)\\right)}{\\| \\nabla_ {e} J \\left(F _ {I} \\left(e _ {k}\\right) , F _ {T} (t)\\right) \\|}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 348, + 785, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Text-guided Augmentation Selection", + "text_level": 1, + "bbox": [ + 215, + 400, + 563, + 416 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Section 3.2, a diverse perturbation set $P$ is derived from the intersection region. To harness the full potential of modality interactions, we introduce text-guided augmentation selection to obtain the optimal sample. Specifically, we individually incorporate each element from the perturbation set $P$ into the adversarial image $v_{i}^{\\prime}$ . The selection process aims to identify the sample that maximally distances $v_{i}^{\\prime}$ from $t$ . This procedure can be represented as:", + "bbox": [ + 212, + 424, + 784, + 516 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nm = \\underset {p _ {m} \\in P} {\\arg \\max } J \\left(F _ {I} \\left(v _ {i} ^ {\\prime} + p _ {m}\\right), F _ {T} (t)\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 530, + 784, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "At this juncture, $e_m$ represents the selected sample. We employ SGA as our baseline and incorporate the image augmentation methods considered along its optimization path. The chosen optimal sample $e_m$ is resized and expanded into the set $E_m = \\{e_{m1}, e_{m2}, \\dots, e_{mM}\\}$ . Subsequently, we utilize the expanded set $E_m$ to generate the final adversarial perturbation, yielding $v_{i+1}'$ :", + "bbox": [ + 212, + 561, + 787, + 638 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nv _ {i + 1} ^ {\\prime} = v _ {i} ^ {\\prime} + \\alpha \\cdot \\operatorname {s i g n} \\left(\\frac {\\nabla_ {e} \\sum_ {j = 1} ^ {M} J \\left(F _ {I} \\left(e _ {m j}\\right) , F _ {T} (t)\\right)}{\\| \\nabla_ {e} \\sum_ {j = 1} ^ {M} J \\left(F _ {I} \\left(e _ {m j}, F _ {T} (t)\\right) \\right\\|}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 647, + 784, + 689 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Adversarial Text deviating from the Intersection Region", + "text_level": 1, + "bbox": [ + 215, + 709, + 725, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the text modality, SGA only considers deviating adversarial text from the ultimate adversarial image generated during the iterative optimization process. If there are a total of T iterations, The generation of $t'$ only considers deviating from the adversarial image $v_T'$ . The adversarial image $v_T'$ is exclusively created by the surrogate model, thereby still presenting the risk of overfitting to the surrogate model. However, during the optimization process of the adversarial image, the clean image is entirely independent of the surrogate model. For this", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Boosting Transferability in Vision-Language Attacks", + "bbox": [ + 380, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "reason, we propose to have the text deviate simultaneously from the last intersection region along the optimization path. Specifically, the adversarial text deviates from the triangle region constituted by $v$ , $v_{T-1}'$ and $v_T'$ .", + "bbox": [ + 212, + 146, + 782, + 194 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nt ^ {\\prime} = \\underset {t ^ {\\prime} \\in B [ t, \\epsilon_ {t} ]} {\\arg \\max } (\\lambda \\cdot J (F _ {I} (v), F _ {T} (t ^ {\\prime})) + \\mu \\cdot J (F _ {I} (v _ {T} ^ {\\prime}), F _ {T} (t ^ {\\prime}))) + \\nu \\cdot J (F _ {I} (v _ {T - 1} ^ {\\prime}), F _ {T} (t ^ {\\prime}))). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 204, + 785, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also set adjustable scaling factors, among which $\\lambda +\\mu +\\nu = 1.0$", + "bbox": [ + 215, + 244, + 710, + 260 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 282, + 455, + 299 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the specific process of our attack, we employ an iterative approach. In each iteration, we sample within the intersection region of adversarial trajectory, guided by textual information, to select a sample that maximally deviates the current adversarial image $v_{i}^{\\prime}$ from the text $t$ . Subsequently, we subject this sample to image augmentation processing, calculate gradients to determine the perturbation direction, and overlay it onto the current adversarial image, resulting in $v_{i + 1}^{\\prime}$ . Through multiple iterative steps, we obtain an adversarial image $v_{T}^{\\prime}$ . For the text modality, in contrast to previous methods that solely focus on deviating from the adversarial image $v_{T}^{\\prime}$ , our goal is to derive an adversarial text $t^{\\prime}$ that simultaneously deviates from the last intersection region $\\triangle vv_{T - 1}^{\\prime}v_{T}^{\\prime}$ along the optimization path.", + "bbox": [ + 212, + 310, + 787, + 476 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 500, + 375, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we present experimental evidence demonstrating the enhanced transferability of multimodal examples generated from our proposed method across VLP models and various Vision-and-Language tasks. First, in Section 4.1, we introduce the experimental settings, including the popular image-text pair datasets and VLP models we use, as well as restrictions for adversarial attacks. Subsequently, the process of searching for optimal parameters is shown in Section 4.2. After that, we evaluate cross-model transferability in the context of the image-text retrieval task, as detailed in Section 4.3. Following this, in Section 4.4, we extend our investigation to the transfer of multimodal adversarial examples generated within the image-text retrieval task to other tasks, aiming to gauge cross-task transferability. Lastly, Section 4.5 outlines ablation studies.", + "bbox": [ + 212, + 532, + 787, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Setups", + "text_level": 1, + "bbox": [ + 215, + 723, + 318, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "VLP Models. In our transferability evaluation experiments across various VLP models, we explore two typical architectures: fused and aligned VLP models. We select CLIP [42] for the aligned VLP model. CLIP offers a choice between two distinct image encoders, namely ViT-B/16 [10] and ResNet-101 [19], denoted as $\\mathrm{CLIP}_{\\mathrm{ViT}}$ and $\\mathrm{CLIP}_{\\mathrm{CNN}}$ , respectively. In the case of fused VLP models, we opt for ALBEF [30] and TCL [53]. ALBEF uses a 12-layer ViT-B/16 image encoder", + "bbox": [ + 212, + 748, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. Gao et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and two 6-layer transformers for text and multimodal encoding. TCL shares this architecture but has different pre-training objectives.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. In this study, we leverage two widely recognized multimodal datasets, namely Flickr30K [41] and MSCOCO [35], for evaluating the image-text retrieval task. The Flickr30K dataset comprises 31,783 images, each accompanied by five captions for annotation. Similarly, the MSCOCO dataset consists of 123,287 images, and approximately five captions are provided for each image.", + "bbox": [ + 212, + 176, + 785, + 252 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Additionally, we employ the RefCOCO+ [55] dataset to assess the Visual Grounding task. RefCOCO+ is a dataset containing 141,564 referring expressions for 50,000 objects within 19,992 MSCOCO images. This dataset serves the purpose of evaluating grounding models by focusing on the localization of objects described through natural language. For another Vision-and-Language task, Image Captioning, we leverage the MSCOCO dataset as well.", + "bbox": [ + 212, + 252, + 787, + 343 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Adversarial Attack Settings. In our study, we adopt adversarial attack settings of SGA [36] to ensure a fair comparison. Specifically, we leverage BERTAttack [31] to craft adversarial texts. The perturbation bound $\\xi_{t}$ is set as 1 and length of word list $W = 10$ . PGD [38] is employed to get adversarial images and the perturbation bound, denoted as $\\xi_v$ , is set as 8/255. Additionally, iteration steps $T$ is set as 10 and each step size $\\alpha = 2 / 255$ . Furthermore, when randomly sampling from the intersection region of adversarial trajectory, we set the number of samples to 5. When generating adversarial texts, we set the three parameters $\\lambda, \\mu, \\nu$ in Equation 6 to 0.6, 0.2, and 0.2 respectively. The values chosen for these adjustable parameters can be found in Section 4.2.", + "bbox": [ + 212, + 343, + 787, + 492 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Metrics. The key metric for adversarial transferability is the Attack Success Rate (ASR), which measures the percentage of successful attacks among all generated adversarial examples. A higher ASR indicates more effective and transferable attacks.", + "bbox": [ + 212, + 493, + 785, + 554 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Optimal Parameters", + "text_level": 1, + "bbox": [ + 214, + 575, + 429, + 589 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In our proposed method, the number of samples $N$ taken from the intersection region of adversarial trajectory and scaling factors in Equation 6 are adjustable. We conduct specific experiments to explore optimal parameter settings and examine their influence on the efficacy of our approach. To be more specific, we utilize ALBEF for generating multimodal adversarial examples on the Flickr30K dataset and assess the transferability on the other three VLP models.", + "bbox": [ + 212, + 598, + 785, + 686 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Number of Samples taken from Intersection Region of Adversarial trajectory $N$ . The bottom of Table 2 illustrates when the sample size reaches 5, transfer effects are observed for both the Image Retrieval task and the Text Retrieval task. As the sample size continues to increase, the transferability only fluctuates and does not exhibit further improvement. Taking into account both transfer effects and computational costs, a sample size of 5 is identified as the optimal configuration.", + "bbox": [ + 212, + 688, + 785, + 794 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Scaling factors $\\lambda, \\mu, \\nu$ in Adversarial Text Generation. In Equation 6, $\\lambda$ represents the weight of clean images, while $\\mu$ and $\\nu$ represent the weights of adversarial images. Therefore, we stipulate that $\\lambda$ cannot be zero, and $\\mu$ and $\\nu$", + "bbox": [ + 212, + 795, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Boosting Transferability in Vision-Language Attacks", + "bbox": [ + 380, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/e51f203b1430437788955279349076032a2ee6e76575b5a6360d5cf54822cb1d.jpg", + "table_caption": [ + "Table 2: Optimal Parameters: Attack Success Rate(%) on different settings, Top for different values of $\\lambda ,\\mu ,\\nu$ and Bottom for different numbers of samples $N$" + ], + "table_footnote": [], + "table_body": "
SourceAttackALBEFTCLCLIPvITCLIPCNN
TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1
ALBEF[λ,μ,v] = [0.2,0.0,0.8]99.999.9389.6790.542.2152.046.2355.44
[λ,μ,v] = [0.2,0.2,0.6]99.999.9390.4190.4341.9651.8746.4955.3
[λ,μ,v] = [0.2,0.4,0.4]99.999.9390.3190.4341.9651.8745.3454.82
[λ,μ,v] = [0.2,0.6,0.2]99.999.9590.5290.5742.2151.8445.0854.92
[λ,μ,v] = [0.2,0.8,0.0]99.999.9390.3190.5741.7251.7446.154.68
[λ,μ,v] = [0.4,0.0,0.6]99.999.9391.2590.8845.5255.3248.9157.63
[λ,μ,v] = [0.4,0.2,0.4]99.999.9391.2590.8345.0355.2248.2857.77
[λ,μ,v] = [0.4,0.4,0.2]99.999.9391.1590.7144.9155.2848.1557.87
[λ,μ,v] = [0.4,0.6,0.0]99.999.9391.1590.8844.9154.7749.0457.56
[λ,μ,v] = [0.6,0.0,0.4]99.999.9391.4690.9546.3856.3849.0459.11
[λ,μ,v] = [0.6,0.2,0.2]99.999.9391.5791.1746.2656.849.5559.01
[λ,μ,v] = [0.6,0.4,0.0]99.999.9390.9490.9846.0156.7249.5558.87
[λ,μ,v] = [0.8,0.0,0.2]99.999.9390.7390.9846.1356.7149.4658.74
[λ,μ,v] = [0.8,0.2,0.0]99.999.9390.3190.9545.7756.6549.3458.87
ALBEFN=399.999.9590.6290.7945.6456.5448.8358.73
N=499.7999.9191.3691.1745.8356.7850.4559.01
N=599.999.9391.5791.1746.2656.849.5559.01
N=699.999.9390.9490.3845.7956.9650.758.52
N=799.999.9189.8890.9545.456.3550.759.07
", + "bbox": [ + 246, + 196, + 759, + 375 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "can have at most one zero value. Initially, as the value of $\\lambda$ gradually increases, indicating the gradual introduction of clean images, the transferability of multimodal adversarial examples increases accordingly. However, when the value of $\\lambda$ becomes too large, it also leads to a disproportionately low proportion of adversarial images, resulting in a decrease in adversarial transferability. According to the experiments, the optimal parameters we select are [0.6, 0.2, 0.2].", + "bbox": [ + 212, + 415, + 787, + 506 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Cross-Model Transferability", + "text_level": 1, + "bbox": [ + 215, + 527, + 493, + 544 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As outlined in Section 4.1, our experimental design focuses on assessing the transferability of adversarial examples across two widely adopted VLP model architectures: fused and aligned. There are four VLP models selected, namely: ALBEF, TCL, $\\mathrm{CLIP}_{\\mathrm{ViT}}$ , $\\mathrm{CLIP}_{\\mathrm{CNN}}$ . The selected downstream $\\mathrm{V} + \\mathrm{L}$ task for evaluation is image-text retrieval. Our approach involves employing one of four models to generate multimodal adversarial examples within the specified parameters of our proposed method. Subsequently, we validate the effectiveness of the generated adversarial examples through a comprehensive set of experiments, encompassing both self-attacks and attacks on the other three models. This evaluation encompasses one white-box attack and three distinct black-box attacks.", + "bbox": [ + 212, + 553, + 787, + 703 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We adopt the methodology proposed by SGA [36] as our baseline, Therefore, the effectiveness of our method is compared against it. Moreover, we also present the effectiveness and transferability results of various other attack methods on VLP models. Table 3 provides a comprehensive comparison of these methods on the dataset Flickr30K, More experiments on the MSCOCO dataset are provided in the Appendix. In the comparison, PGD [38] is an image-only attack, while Bert-Attack [31] focuses solely on text-based attacks. Sep-Attack involves perturbing text and image separately. Furthermore, Co-Attack takes into account cross-modal interactions, generating an adversarial example for one modality", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. Gao et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/fa112191c91e91d2277120ecde637d54a2308149461c8aeb4606ed354312286e.jpg", + "image_caption": [ + "Cleannnnnne", + "a man riding on the back of a motorcycle down a street" + ], + "image_footnote": [], + "bbox": [ + 282, + 148, + 344, + 195 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b7ca1681e86010cf84c61bc96873c5e2ac7db0d058d23c0f6eb5c61370b5a36f.jpg", + "image_caption": [ + "a man wearing a black shirt and a green tie" + ], + "image_footnote": [], + "bbox": [ + 351, + 147, + 413, + 195 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d759c7dc32f0c0e468c00a5596e8b6f92738d77d463285e551d7e74b8348f12e.jpg", + "image_caption": [ + "a parking meter sitting on the side of a road" + ], + "image_footnote": [], + "bbox": [ + 418, + 147, + 480, + 195 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/20ec060bcb801515f1977f1ffba3f0ae7943db9ad671f0e3a1bce39029325548.jpg", + "image_caption": [ + "a cat standing on a toilet seat in a bathroom" + ], + "image_footnote": [], + "bbox": [ + 486, + 147, + 547, + 195 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7306439b9c64d067525644f744d093202dc47c844c13cda18ce367a99fcb14d3.jpg", + "image_caption": [ + "am an riding a bike next to a train" + ], + "image_footnote": [], + "bbox": [ + 552, + 147, + 614, + 195 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f8244d470fc556955750045aa8029aa5e6317db3fe6466b5ec4fe42bd51bff50.jpg", + "image_caption": [ + "a man sitting on a bench next to a white dog" + ], + "image_footnote": [], + "bbox": [ + 620, + 147, + 683, + 195 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ec9796b597d6cea774128ee16fa8c3987566cdc05c892c51f399dbd46011bc52.jpg", + "image_caption": [ + "a siamese cat eating a banana from a person's hand" + ], + "image_footnote": [], + "bbox": [ + 687, + 147, + 750, + 195 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6a87880aeea6b07ebd99c6a6394751353bc2c97486a063c3b55c7ef4717f70b9.jpg", + "image_caption": [ + "Caaee", + "Adversarial", + "Caaee", + "a woman sitting on a motorcycle in the street", + "Fig. 2: Visualization on Image Captioning. We use the ALBEF model, pre-trained on Image Text Retrieval(ITR) task, to generate adversarial images on the MSCOCO dataset and use the BLIP [29] model for Image Captioning on both clean images and adversarial images, respectively." + ], + "image_footnote": [], + "bbox": [ + 284, + 232, + 346, + 281 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3d2fdd59f275d91b97284a1af2a57e302901040857a8043d9fc9415d4c50bb93.jpg", + "image_caption": [ + "a young girl wearing glasses and a tie" + ], + "image_footnote": [], + "bbox": [ + 351, + 232, + 413, + 281 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1bd6db1a76ba15eaa41e39e3c73392fe25888edf259b42e1eac659314dcd1c64.jpg", + "image_caption": [ + "a group of people riding motorcycles down a street" + ], + "image_footnote": [], + "bbox": [ + 418, + 232, + 480, + 281 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/94ab0d8f2de445360357fa605bc658cca27c1d0d3c6a926a6dc2612bdac7d41b.jpg", + "image_caption": [ + "a black and white dog laying on top of a toilet" + ], + "image_footnote": [], + "bbox": [ + 486, + 232, + 547, + 281 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2907a68e8f28b9598049757ab815de30dda8d019a8beace6e512755557aabd53.jpg", + "image_caption": [ + "a train on a train track with a dog standing next to it" + ], + "image_footnote": [], + "bbox": [ + 553, + 232, + 614, + 281 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3ddd831a1f68db9f81c9b47b7a79996bd2b973fab15e046a381c94e80f2f1162.jpg", + "image_caption": [ + "a boy riding a bike next to a white dog" + ], + "image_footnote": [], + "bbox": [ + 620, + 232, + 683, + 281 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5c58d90dfdffac9a457e7743617d5af84eec7c9581ab90d1c023a6876fbfb8a1.jpg", + "image_caption": [ + "a small dog is eating a banana on the floor" + ], + "image_footnote": [], + "bbox": [ + 687, + 232, + 750, + 281 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "under the guidance of the other modality. SGA, our baseline, expands a single image-text pair into a set of images and a set of texts to enhance diversity.", + "bbox": [ + 212, + 417, + 784, + 446 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "First, we compare the performance of various methods under white-box attacks, wherein we can leverage the model architecture and interact with the model. It is evident that our method, along with SGA, performs the best in the four white-box attack experiments compared to other methods. Whether TR or IR, the attack success rate at the top-1 rank $(\\mathbb{R}@\\mathbb{1})$ consistently exceeds $99.9\\%$ . Given the already high white-box attack success rate achieved by the SGA method, there is limited room for improvement in our approach within this context. Subsequently, we shift our focus to elucidating the enhancements our method brings to transferability, specifically in the realm of black-box attack performance. We delineate this exploration into two segments, contingent on whether the source model and the target model share the same architecture.", + "bbox": [ + 212, + 446, + 784, + 612 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Cross-Model Transferability in Same Architecture. ALBEF and TCL both utilize a similar model architecture but differ in their pre-training objectives while maintaining a common fundamental model structure. Therefore, when ALBEF and TCL are employed as target models for each other, the success rate of attacks using multimodal adversarial examples is remarkably high. Notably, the black-box attack success rate of IR reaches $95.58\\%$ when TCL is the target model in our proposed method. Additionally, the transferability of SGA can reach approximately $90\\%$ , but the room for improvement is very limited, our method has improved compared to SGA, ranging from $2.71\\%$ to $3.69\\%$ . In contrast, when considering $\\mathrm{CLIP}_{\\mathrm{ViT}}$ and $\\mathrm{CLIP}_{\\mathrm{CNN}}$ —both being aligned VLP models—their image encoders vary significantly, with one utilizing the Vision Transformer and the other employing ResNet-101. Given the significant structural differences between traditional CNNs and Transformers, existing methods show low cross-model transferability, highlighting room for improvement. Our method outperforms SGA by about $7.67\\%$ to $11.62\\%$ .", + "bbox": [ + 212, + 613, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Boosting Transferability in Vision-Language Attacks", + "bbox": [ + 380, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/e6fd463d4d74e9b0c0118626ed42fe1033e8e01a2ec61f2e6aad6d4de79afd66.jpg", + "table_caption": [ + "Table 3: Comparison with state-of-the-art methods on image-text retrieval. The source column shows VLP models we use to generate multimodal adversarial examples. The gray area represents adversarial attacks under a white-box setting, the rest are black-box attacks. For both Image Retrieval and Text Retrieval, we provide R@1 attack success rate(%)" + ], + "table_footnote": [], + "table_body": "
SourceAttackALBEFTCLCLIP_VITCLIP_CNN
TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1
ALBEFPGD93.7494.4324.0327.910.6715.8214.0519.11
BERT-Attack11.5727.4612.6428.0729.3343.1732.6946.11
Sep-Attack95.7296.1439.351.7934.1145.7235.7647.92
Co-Attack97.0898.3639.5251.2429.8238.9231.2941.99
SGA99.999.9887.8888.0536.6946.7839.5949.78
Ours99.999.9391.5791.1746.2656.849.5559.01
TCLPGD35.7741.6799.3799.3310.1816.314.8121.1
BERT-Attack11.8926.8214.5429.1729.6944.4933.4646.07
Sep-Attack52.4561.4499.5899.4537.0645.8137.4249.91
Co-Attack49.8460.3691.6895.4832.6442.6932.0647.82
SGA92.4992.77100.0100.036.8146.9741.8951.53
Ours95.295.58100.099.9847.2457.2852.2362.23
CLIP_VITPGD3.136.484.438.8369.3384.7913.0317.43
BERT-Attack9.5922.6411.8025.0728.3439.0830.4037.43
Sep-Attack7.6120.5810.1220.7476.9387.4429.8938.32
Co-Attack8.5520.1810.0121.2978.5387.529.538.49
SGA22.4234.5925.0836.45100.0100.053.2661.1
Ours27.8442.8427.8244.6100.0100.064.8869.5
CLIP_CNNPGD2.296.154.538.885.412.0889.7893.04
BERT-Attack8.8623.2712.3325.4827.1237.4430.4040.10
Sep-Attack9.3822.9911.2825.4526.1339.2493.6195.3
Co-Attack10.5323.6212.5426.0527.2440.6295.9196.5
SGA15.6428.618.0233.0739.0251.4599.8799.9
Ours19.534.5921.637.8848.4759.1299.8799.9
", + "bbox": [ + 246, + 238, + 759, + 479 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Cross-Model Transferability in Different Architectures. When fused VLP models (i.e., ALBEF and TCL) are used for generating adversarial examples, our method achieves a significant improvement compared to existing methods, outperforming the current state-of-the-art (SOTA) methods by $9.23\\%$ to $10.7\\%$ . When the fused VLP models are targeted for attack, our method still outperforms all existing methods when generating adversarial examples against the aligned VLP models (i.e., CLIP). In this scenario, the transferability of all methods is relatively low. Compared to our method's baseline SGA, we attain an improvement ranging from $2.74\\%$ to $8.25\\%$ .", + "bbox": [ + 212, + 523, + 787, + 659 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Cross-Task Transferability", + "text_level": 1, + "bbox": [ + 215, + 686, + 478, + 704 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We not only assess the transferability of multimodal adversarial examples generated by our proposed method across different VLP models but also conduct experiments to evaluate its effectiveness in transferring across diverse $\\mathrm{V} + \\mathrm{L}$ tasks. Specifically, we craft adversarial examples for the Image-Text Retrieval (ITR) task and evaluate them on Visual Grounding (VG) and Image Captioning (IC) tasks. As evident from Table 4 and visual results in Figure 2 and 3, the adversarial examples generated for ITR demonstrate transferability, successfully impacting both VG and IC tasks. This highlights the efficacy of cross-task trans", + "bbox": [ + 212, + 718, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. Gao et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/dd50538e90d8f02ac790684bf3ff72dc8c7f817cc7fca1353f6606f5da6073d9.jpg", + "image_caption": [ + "Fig. 3: Visualization on Visual Grounding. We use the ALBEF model, pre-trained on the ITR task, to generate adversarial images on the RefCOCO+ dataset and use the same model, pre-trained on Visual Grouding(VG) task, to localize the regions corresponding to red words on both clean images and adversarial images, respectively." + ], + "image_footnote": [], + "bbox": [ + 251, + 150, + 750, + 417 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "ferability in our proposed method. Furthermore, our transferability consistently outperforms that of SGA.", + "bbox": [ + 212, + 518, + 784, + 550 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 571, + 388, + 588 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Our proposed method builds upon the SGA [36] as a baseline, introducing two key improvements. Firstly, we utilize diversification along the intersection region of adversarial trajectory to expand the diversity of adversarial examples. Secondly, we generate adversarial text while simultaneously distancing it from the last intersection region along the optimization path. To investigate the impact of each improvement on the effectiveness of our method, we conduct ablation studies on the ITR task and employ transferability from ALBEF to the other three VLP models as the evaluation metric.", + "bbox": [ + 212, + 597, + 787, + 717 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In the ablation study, we systematically eliminate each enhancement from our approach and compare their transferability with both the baseline SGA and our final method, which incorporates a combination of two improvements. The results are depicted in Figure 4. It is evident that when attacking models with the same architecture versus different architectures, the impacts of two distinct improvements are also different. When attacking TCL with the same structure as ALBEF, under Setting1, our method exhibits a greater decrease in transferability, indicating that diversification along the intersection region is", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Boosting Transferability in Vision-Language Attacks", + "bbox": [ + 380, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/dec395851dece59e370bb92a6943a7bf357d431c53e687bc67191bcdfbad9574.jpg", + "table_caption": [ + "Table 4: Cross-Task Transferability. We utilize ALBEF to generate multimodal adversarial examples for attacking both Visual Grounding(VG) on the RefCOCO+ dataset and Image Captioning(IC) on the MSCOCO dataset. The baseline represents the performance of each task without any attack, where a lower value indicates better effectiveness of the adversarial attack for both tasks." + ], + "table_footnote": [], + "table_body": "
AttackITR → VGITR → IC
ValTestATestBB@4METEORROUGE-LCIDErSPICE
Baseline58.4665.8946.2539.731.060.0133.323.8
SGA50.5657.4240.6628.024.651.291.417.7
Ours49.7056.3240.5427.224.250.788.317.2
", + "bbox": [ + 246, + 238, + 759, + 309 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/8268410413513309c42bce47d08364b6f9b93b3b4de585e0fb11a9230cc8d25c.jpg", + "image_caption": [ + "Fig. 4: Ablation Study: Attack Success Rate(%) on other three target models. The baseline is SGA. Setting 1 removes diversification along the intersection region of adversarial trajectory. Setting 2 removes the text deviating from the last intersection region along the optimization path." + ], + "image_footnote": [], + "bbox": [ + 246, + 335, + 413, + 445 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/ba2b94f80a5264bb12ca3d19cac0ff96a4d84d6598e8a24ad2156afe5adc77aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 337, + 584, + 445 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3d1524d5e2bdd08b7b423cf627dcc13674fa14fc8a8f732e289db634576f69b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 337, + 754, + 445 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "most crucial in our approach under this scenario. However, under Setting2, the decrease in transferability is more pronounced when attacking the aligned VLP model, indicating that at this point, adversarial texts deviating from the last intersection region play a larger role in the effectiveness of our method.", + "bbox": [ + 212, + 542, + 787, + 604 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 627, + 359, + 643 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we conduct a systematic evaluation of existing multimodal attacks regarding transferability. We found that these methods predominantly prioritize diversity around adversarial examples (AEs) during the optimization process, potentially leading to overfitting to the victim model and hindering transferability. To address this issue, we propose diversification along the intersection region of adversarial trajectory to broaden diversity not only around AEs but also towards clean inputs. Moreover, we pioneer an exploration into extending adversarial texts deviating from the intersection region. Through extensive experiments, we demonstrate the effectiveness of our method in enhancing transferability across VLP models and $\\mathrm{V} + \\mathrm{L}$ tasks. This work could act as a catalyst for more profound research on the transferability of multimodal AEs, alongside fortifying the adversarial robustness of VLP models.", + "bbox": [ + 212, + 657, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. Gao et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research is supported by the National Research Foundation, Singapore, and DSO National Laboratories under the AI Singapore Programme (AISG Award No: AISG2-GC-2023-008), Career Development Fund (CDF) of Agency for Science, Technology and Research (A*STAR) (No.: C233312028), and National Research Foundation, Singapore and Infocomm Media Development Authority under its Trust Tech Funding Initiative (No. DTC-RGC-04).", + "bbox": [ + 212, + 180, + 787, + 272 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 297, + 321, + 314 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abdelfattah, R., Guo, Q., Li, X., Wang, X., Wang, S.: Cdul: Clip-driven unsupervised learning for multi-label image classification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1348-1357 (2023)", + "2. Anderson, P., Fernando, B., Johnson, M., Gould, S.: Spice: Semantic propositional image caption evaluation. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 382-398. Springer (2016)", + "3. Bai, Y., Zeng, Y., Jiang, Y., Wang, Y., Xia, S.T., Guo, W.: Improving query efficiency of black-box adversarial attack. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16. pp. 101-116. Springer (2020)", + "4. Banerjee, S., Lavie, A.: Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In: Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization. pp. 65-72 (2005)", + "5. Chen, H., Ding, G., Liu, X., Lin, Z., Liu, J., Han, J.: Imram: Iterative matching with recurrent attention memory for cross-modal image-text retrieval. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12655-12663 (2020)", + "6. Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: Uniter: Universal image-text representation learning. In: European conference on computer vision. pp. 104-120. Springer (2020)", + "7. Cheng, H., Xiao, E., Cao, J., Yang, L., Xu, K., Gu, J., Xu, R.: Typography leads semantic diversifying: Amplifying adversarial transferability across multimodal large language models. arXiv preprint arXiv:2405.20090 (2024)", + "8. Cheng, M., Sun, Y., Wang, L., Zhu, X., Yao, K., Chen, J., Song, G., Han, J., Liu, J., Ding, E., et al.: Vista: vision and scene text aggregation for cross-modal retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5184-5193 (2022)", + "9. Dong, Y., Pang, T., Su, H., Zhu, J.: Evading defenses to transferable adversarial examples by translation-invariant attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4312-4321 (2019)", + "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arxiv 2020. arXiv preprint arXiv:2010.11929 (2010)" + ], + "bbox": [ + 218, + 333, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Boosting Transferability in Vision-Language Attacks", + "bbox": [ + 380, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "11. Dou, Z.Y., Xu, Y., Gan, Z., Wang, J., Wang, S., Wang, L., Zhu, C., Zhang, P., Yuan, L., Peng, N., et al.: An empirical study of training end-to-end vision-and-language transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18166-18176 (2022)", + "12. Gao, K., Bai, Y., Bai, J., Yang, Y., Xia, S.T.: Adversarial robustness for visual grounding of multimodal large language models. arXiv preprint arXiv:2405.09981 (2024)", + "13. Gu, J., Jia, X., de Jorge, P., Yu, W., Liu, X., Ma, A., Xun, Y., Hu, A., Khakzar, A., Li, Z., et al.: A survey on transferability of adversarial examples across deep neural networks. arXiv preprint arXiv:2310.17626 (2023)", + "14. Guo, Q., Cheng, Z., Juefei-Xu, F., Ma, L., Xie, X., Liu, Y., Zhao, J.: Learning to adversarially blur visual object tracking. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10839-10848 (2021)", + "15. Guo, Q., Juefei-Xu, F., Xie, X., Ma, L., Wang, J., Yu, B., Feng, W., Liu, Y.: Watch out! motion is blurring the vision of your deep neural networks. In: Advances in Neural Information Processing Systems. vol. 33, pp. 975-985 (2020)", + "16. Han, D., Jia, X., Bai, Y., Gu, J., Liu, Y., Cao, X.: Ot-attack: Enhancing adversarial transferability of vision-language models via optimal transport optimization. arXiv preprint arXiv:2312.04403 (2023)", + "17. He, B., Jia, X., Liang, S., Lou, T., Liu, Y., Cao, X.: Sa-attack: Improving adversarial transferability of vision-language pre-training models via self-augmentation. arXiv preprint arXiv:2312.04913 (2023)", + "18. He, B., Liu, J., Li, Y., Liang, S., Li, J., Jia, X., Cao, X.: Generating transferable 3d adversarial point cloud via random perturbation factorization. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 764-772 (2023)", + "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2016)", + "20. Hu, X., Gan, Z., Wang, J., Yang, Z., Liu, Z., Lu, Y., Wang, L.: Scaling up vision-language pre-training for image captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17980-17989 (June 2022)", + "21. Huang, Y., Guo, Q., Juefei-Xu, F., Ma, L., Miao, W., Liu, Y., Pu, G.: Advfilter: predictive perturbation-aware filtering against adversarial attack via multi-domain learning. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 395-403 (2021)", + "22. Huang, Y., Juefei-Xu, F., Guo, Q., Zhang, J., Wu, Y., Hu, M., Li, T., Pu, G., Liu, Y.: Personalization as a shortcut for few-shot backdoor attack against text-to-image diffusion models. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 21169-21178 (2024)", + "23. Huang, Y., Sun, L., Guo, Q., Juefei-Xu, F., Zhu, J., Feng, J., Liu, Y., Pu, G.: Ala: Naturalness-aware adversarial lightness attack. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 2418-2426 (2023)", + "24. Huang, Y., Sun, L., Guo, Q., Juefei-Xu, F., Zhu, J., Feng, J., Liu, Y., Pu, G.: Ala: Naturalness-aware adversarial lightness attack. In: Proceedings of the 31st ACM International Conference on Multimedia. p. 2418-2426 (2023)", + "25. Jia, X., Wei, X., Cao, X., Han, X.: Adv-watermark: A novel watermark perturbation for adversarial examples. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 1579-1587 (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. Gao et al.", + "bbox": [ + 271, + 114, + 356, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "26. Jia, X., Zhang, Y., Wei, X., Wu, B., Ma, K., Wang, J., Cao Sr, X.: Improving fast adversarial training with prior-guided knowledge. arXiv preprint arXiv:2304.00202 (2023)", + "27. Khan, Z., Fu, Y.: Exploiting bert for multimodal target sentiment classification through input space translation. In: Proceedings of the 29th ACM international conference on multimedia. pp. 3034-3042 (2021)", + "28. Lei, C., Luo, S., Liu, Y., He, W., Wang, J., Wang, G., Tang, H., Miao, C., Li, H.: Understanding chinese video and language via contrastive multimodal pre-training. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 2567-2576 (2021)", + "29. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)", + "30. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems 34, 9694-9705 (2021)", + "31. Li, L., Ma, R., Guo, Q., Xue, X., Qiu, X.: Bert-attack: Adversarial attack against bert using bert (2020)", + "32. Li, X., Yin, X., Li, C., Zhang, P., Hu, X., Zhang, L., Wang, L., Hu, H., Dong, L., Wei, F., et al.: Oscar: Object-semantics aligned pre-training for vision-language tasks. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16. pp. 121-137. Springer (2020)", + "33. Lin, C.Y.: Rouge: A package for automatic evaluation of summaries. In: Text summarization branches out. pp. 74-81 (2004)", + "34. Lin, J., Song, C., He, K., Wang, L., Hopcroft, J.E.: Nesterov accelerated gradient and scale invariance for adversarial attacks. arXiv preprint arXiv:1908.06281 (2019)", + "35. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "36. Lu, D., Wang, Z., Wang, T., Guan, W., Gao, H., Zheng, F.: Set-level guidance attack: Boosting adversarial transferability of vision-language pre-training models (2023)", + "37. Luo, H., Gu, J., Liu, F., Torr, P.: An image is worth 1000 lies: Adversarial transferability across prompts on vision-language models. arXiv preprint arXiv:2403.09766 (2024)", + "38. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2017)", + "39. Naseer, M.M., Ranasinghe, K., Khan, S.H., Hayat, M., Shahbaz Khan, F., Yang, M.H.: Intriguing properties of vision transformers. Advances in Neural Information Processing Systems 34, 23296-23308 (2021)", + "40. Park, J., Miller, P., McLaughlin, N.: Hard-label based small query black-box adversarial attack. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3986-3995 (2024)", + "41. Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In: Proceedings of the IEEE international conference on computer vision. pp. 2641-2649 (2015)", + "42. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Boosting Transferability in Vision-Language Attacks", + "bbox": [ + 380, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "43. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)", + "44. Shi, L., Shuang, K., Geng, S., Gao, P., Fu, Z., de Melo, G., Chen, Y., Su, S.: Dense contrastive visual-linguistic pretraining. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 5203-5212 (2021)", + "45. Tan, H., Bansal, M.: Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490 (2019)", + "46. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: International conference on machine learning. pp. 10347-10357. PMLR (2021)", + "47. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4566-4575 (2015)", + "48. Wang, T., Ge, Y., Zheng, F., Cheng, R., Shan, Y., Qie, X., Luo, P.: Accelerating vision-language pretraining with free language modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23161-23170 (2023)", + "49. Wang, T., Jiang, W., Lu, Z., Zheng, F., Cheng, R., Yin, C., Luo, P.: Vlmixer: Unpaired vision-language pre-training via cross-modal cutmix. In: International Conference on Machine Learning. pp. 22680-22690. PMLR (2022)", + "50. Wang, X., He, X., Wang, J., He, K.: Admix: Enhancing the transferability of adversarial attacks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16158-16167 (2021)", + "51. Wang, Z., Liu, X., Li, H., Sheng, L., Yan, J., Wang, X., Shao, J.: Camp: Cross-modal adaptive message passing for text-image retrieval. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5764-5773 (2019)", + "52. Xie, C., Zhang, Z., Zhou, Y., Bai, S., Wang, J., Ren, Z., Yuille, A.L.: Improving transferability of adversarial examples with input diversity. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2730-2739 (2019)", + "53. Yang, J., Duan, J., Tran, S., Xu, Y., Chanda, S., Chen, L., Zeng, B., Chilimbi, T., Huang, J.: Vision-language pre-training with triple contrastive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15671-15680 (2022)", + "54. Yu, L., Lin, Z., Shen, X., Yang, J., Lu, X., Bansal, M., Berg, T.L.: Mattnet: Modular attention network for referring expression comprehension. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1307-1315 (2018)", + "55. Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in referring expressions. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 69-85. Springer (2016)", + "56. Yuan, L., Chen, Y., Wang, T., Yu, W., Shi, Y., Jiang, Z.H., Tay, F.E., Feng, J., Yan, S.: Tokens-to-token vit: Training vision transformers from scratch on imagenet. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 558-567 (2021)" + ], + "bbox": [ + 217, + 147, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "S. Gao et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "57. Zhang, J., Yi, Q., Sang, J.: Towards adversarial attack on vision-language pretraining models. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 5005-5013 (2022)", + "58. Zhang, J., Huang, J.t., Wang, W., Li, Y., Wu, W., Wang, X., Su, Y., Lyu, M.R.: Improving the transferability of adversarial samples by path-augmented method. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8173-8182 (2023)", + "59. Zhang, P., Li, X., Hu, X., Yang, J., Zhang, L., Wang, L., Choi, Y., Gao, J.: Vinyl: Revisiting visual representations in vision-language models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5579-5588 (2021)", + "60. Zhang, Q., Lei, Z., Zhang, Z., Li, S.Z.: Context-aware attention network for image-text retrieval. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3536-3545 (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 340 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Boosting Transferability in Vision-Language Attacks", + "bbox": [ + 380, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory/d6848265-b938-496f-9e0e-c16ce7da086b_model.json b/2024/Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory/d6848265-b938-496f-9e0e-c16ce7da086b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8fbb278474ddd202ebad7c8241c2553827e7ce80 --- /dev/null +++ b/2024/Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory/d6848265-b938-496f-9e0e-c16ce7da086b_model.json @@ -0,0 +1,2779 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.223, + 0.141, + 0.782, + 0.21 + ], + "angle": 0, + "content": "Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.234, + 0.681, + 0.266 + ], + "angle": 0, + "content": "Sensen Gao \\(^{1,3}\\), Xiaojun Jia \\(^{2*}\\), Xuhong Ren \\(^{2}\\), Ivor Tsang \\(^{2,3}\\), and Qing Guo \\(^{3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.345, + 0.277, + 0.623, + 0.291 + ], + "angle": 0, + "content": "1 Nankai University, Tianjin, China" + }, + { + "type": "text", + "bbox": [ + 0.345, + 0.291, + 0.659, + 0.306 + ], + "angle": 0, + "content": "\\(^{2}\\) Nanyang Technological University, Singapore" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.306, + 0.761, + 0.334 + ], + "angle": 0, + "content": "3 CFAR and IHPC, Agency for Science, Technology and Research (A*STAR), Singapore" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.363, + 0.744, + 0.697 + ], + "angle": 0, + "content": "Abstract. Vision-language pre-training (VLP) models exhibit remarkable capabilities in comprehending both images and text, yet they remain susceptible to multimodal adversarial examples (AEs). Strengthening attacks and uncovering vulnerabilities, especially common issues in VLP models (e.g., high transferable AEs), can advance reliable and practical VLP models. A recent work (i.e., Set-level guidance attack) indicates that augmenting image-text pairs to increase AE diversity along the optimization path enhances the transferability of adversarial examples significantly. However, this approach predominantly emphasizes diversity around the online adversarial examples (i.e., AEs in the optimization period), leading to the risk of overfitting the victim model and affecting the transferability. In this study, we posit that the diversity of adversarial examples towards the clean input and online AEs are both pivotal for enhancing transferability across VLP models. Consequently, we propose using diversification along the intersection region of adversarial trajectory to expand the diversity of AEs. To fully leverage the interaction between modalities, we introduce text-guided adversarial example selection during optimization. Furthermore, to further mitigate the potential overfitting, we direct the adversarial text deviating from the last intersection region along the optimization path, rather than adversarial images as in existing methods. Extensive experiments affirm the effectiveness of our method in improving transferability across various VLP models and downstream vision-and-language tasks. Code is available at https://github.com/SensenGao/VLTPredictAttack." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.709, + 0.741, + 0.737 + ], + "angle": 0, + "content": "Keywords: Vision-Language Attack \\(\\cdot\\) Adversarial Transferability \\(\\cdot\\) Diversification \\(\\cdot\\) Intersection Region of Adversarial Trajectory" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.759, + 0.377, + 0.775 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.788, + 0.788, + 0.819 + ], + "angle": 0, + "content": "Vision-language pre-training (VLP) models utilize multimodal learning, leveraging large-scale image-text pairs to bridge the gap between visual and lan" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.525, + 0.841 + ], + "angle": 0, + "content": "* co-first authors; † co-corresponding authors" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "S. Gao et al." + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.145, + 0.443, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.256, + 0.378, + 0.267 + ], + "angle": 0, + "content": "(a) SGA" + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.145, + 0.753, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.257, + 0.617, + 0.268 + ], + "angle": 0, + "content": "(b) Ours" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.268, + 0.495, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.383, + 0.453, + 0.395 + ], + "angle": 0, + "content": "(c) ALBEF to \\(\\mathrm{CLIP}_{\\mathrm{CNN}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.268, + 0.754, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.585, + 0.383, + 0.713, + 0.395 + ], + "angle": 0, + "content": "(d) \\(\\mathrm{CLIP}_{\\mathrm{ViT}}\\) to \\(\\mathrm{CLIP}_{\\mathrm{CNN}}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.408, + 0.789, + 0.534 + ], + "angle": 0, + "content": "Fig. 1: Our method vs. set-level guided attack (SGA) [36]. (a) shows the main idea of SGA, i.e., conducting augmentation around the online adversarial examples. (b) shows the main idea of our method, that is, we perform augmentation in the intersection region of adversarial trajectory. The red and blue dots both depict images sampled from the intersection region, with red dots indicating the best samples selected using the text-guided adversarial example selection strategy. The surrounding light red dots represent applying the same resizing data augmentation to the best samples as utilized in SGA. (c) and (d) compare the transferability of our method and SGA by using the adversarial examples of ALBEF [30] and \\(\\mathrm{CLIP}_{\\mathrm{ViT}}\\) to attack \\(\\mathrm{CLIP}_{\\mathrm{CNN}}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.789, + 0.687 + ], + "angle": 0, + "content": "guage understanding. These models showcase revolutionary performance across various downstream Vision-and-Language tasks, including image-text retrieval, image captioning, visual grounding, and visual entailment, as demonstrated in [20, 27, 28, 44]. Despite their success, recent research underscores the significant vulnerability of VLP models, particularly when confronted with multimodal adversarial examples [7, 12, 16, 17, 36, 37, 57]. Uncovering the vulnerabilities, especially common issues, can drive further research aimed at building more reliable and practical VLP models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.789, + 0.809 + ], + "angle": 0, + "content": "Current research predominantly concentrates on attacking VLP models via a white-box setting, where the model's structural information can be exploited. However, exploring the transferability of multimodal adversarial examples is pivotal, especially given the limited access to detailed model structures in real-world scenarios. There have been some efforts made to enhance the transferability of attacks on VLP models by introducing input diversity, as seen in the work SGA [36]. While they have achieved some effectiveness, how to enhance the transferability of multimodal adversarial examples is still an open question." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.84 + ], + "angle": 0, + "content": "In this paper, we undertake a comprehensive examination of the factors contributing to the limited transferability of the cutting-edge multimodal attack" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.381, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Boosting Transferability in Vision-Language Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "method, i.e., SGA [36]. As shown in Figure 1 (a), throughout the iterative generation of subsequent adversarial images, SGA conducts data augmentation around the online adversarial image (i.e., adversarial examples generated during optimization). This strategy enhances the diversity of adversarial examples along the optimization path, leading to a certain improvement in transferability. However, such an approach still carries the potential risk of overfitting to the victim model with the high reliance on the examples along the adversarial trajectory (See Figure 1 (b)), which leads to low attack success rates when we migrate the adversarial example to other VLP models (See Figure 1 (c) and (d)). To mitigate this overfitting risk, one potential solution is to further enhance the diversity of augmented adversarial examples in a judicious manner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.317, + 0.788, + 0.543 + ], + "angle": 0, + "content": "Building upon the aforementioned analysis, SGA overfits local adversarial examples, while the clean image is the only accessible example that is far from local adversarial examples. Therefore, we embark on a pioneering endeavor to enhance the transferability of multimodal adversarial attacks by considering the diversity of adversarial examples (AEs) around clean inputs and online AEs throughout the optimization process. To achieve this, we consider the intersection region of adversarial trajectory, which encompasses the original image, the adversarial image from the previous step, and the current adversarial image during the iterative attack process (depicted in Figure 1 (b)). This innovative approach aims to circumvent overfitting by strategic sampling within this region, thereby avoiding an undue focus on adversarial example diversity solely around adversarial images. After obtaining multiple samples, we calculate gradients for each to determine perturbation directions away from the text. Subsequently, we individually incorporate these perturbations into the current adversarial image and select the one that deviates the most from the text." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.625 + ], + "angle": 0, + "content": "Additionally, in the text modality, SGA only considers deviating the text from the last adversarial image in the optimization period, but the adversarial image is solely generated by the surrogate model, still posing the risk of overfitting the surrogate model. For this reason, we propose to have the text deviate simultaneously from the last intersection region along the optimization path." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Our proposed method is evaluated on two widely recognized multimodal datasets, Flickr30K [41] and MSCOCO [35]. We conduct experiments on three vision-and-language downstream tasks (i.e., image-text retrieval (ITR), visual grounding(VG), and image captioning (IC)), and all results indicate the high effectiveness of our method in generating more transferable multimodal adversarial examples. Moreover, when adversarial examples generated from image-text retrieval are transferred to other vision-and-language downstream tasks (i.e., VG and IC), there is a substantial improvement in attack performance." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.753, + 0.633, + 0.767 + ], + "angle": 0, + "content": "Our main contributions can be summarized as follows:" + }, + { + "type": "page_footnote", + "bbox": [ + 0.226, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- We propose using the intersection region of adversarial trajectory to expand the diversity of adversarial examples during optimization, based on which we develop a high-transferability attack against VLP models." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "S. Gao et al." + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "- We extend the generation of adversarial text to deviate from the last intersection region along the optimization path, aiming to reduce overfitting the surrogate model, thereby achieving enhanced transferability." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.193, + 0.785, + 0.237 + ], + "angle": 0, + "content": "- Extensive experiments robustly demonstrate the efficacy of our proposed method in elevating the transferability of multimodal adversarial examples across diverse models and three downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.262, + 0.388, + 0.279 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.295, + 0.572, + 0.312 + ], + "angle": 0, + "content": "2.1 Vision-Language Pre-training Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.322, + 0.785, + 0.427 + ], + "angle": 0, + "content": "VLP models leverage multimodal learning from extensive image-text pairs to improve the performance of various Vision-and-Language \\((\\mathrm{V} + \\mathrm{L})\\) tasks [29]. Early VLP models predominantly depend on pre-trained object detectors for acquiring multimodal representations [6, 32, 45, 49, 59]. Recently, with the advent of end-to-end image encoders like the Vision Transformer (ViT) [10, 46, 56] offering faster inference speeds, some work propose to use them as substitution for computationally expensive object detectors [11, 29, 30, 48, 53]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.787, + 0.563 + ], + "angle": 0, + "content": "There are two popular approaches for VLP models in learning vision-language representations: the fused architecture and the aligned architecture. Fused VLP models (e.g., ALBEF [30], TCL [53]), initially employ two separate unimodal encoders to learn features for text and images. Subsequently, a multimodal encoder is utilized to fuse the embeddings of text and images. In contrast, aligned VLP models, exemplified by CLIP [42], focus on aligning the feature spaces of distinct unimodal encoders and benefit downstream tasks significantly [1]. This paper concentrates on evaluating our proposed method using multiple popular fused and aligned VLP models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.587, + 0.596, + 0.603 + ], + "angle": 0, + "content": "2.2 Downstream Vision-and-Language Tasks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.785, + 0.688 + ], + "angle": 0, + "content": "Image-Text Retrieval (ITR) involves retrieving pertinent information, textual or visual, in response to queries from another modality [5, 8, 51, 60]. This undertaking usually encompasses two sub-tasks: image-to-text retrieval (retrieving text based on an image query) and text-to-image retrieval (retrieving images given a text query)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.785, + 0.809 + ], + "angle": 0, + "content": "In aligned VLP models, both the Text Retrieval (TR) and Image Retrieval (IR) tasks leverage ranking results determined by the similarity between text and image embeddings. However, in fused VLP models, where internal embedding spaces lack alignment across unimodal encoders, the similarity scores between image and text modalities are computed for all image-text pairs to retrieve the Top-N candidates. Subsequently, these Top-N candidates serve as input for the multimodal encoder, which computes the image-text matching score to establish the final ranking." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Visual Grounding (VG) refers to the task of localizing the region within a visual scene with corresponding entities or concepts in natural language. Among" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.381, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Boosting Transferability in Vision-Language Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 1: Attack Success Rate (%) of SGA with and without image augmentation. The SGA w.o. Aug doesn't consider image augmentation. We use ALBEF to generate multimodal adversarial examples on the ITR task to evaluate transferability." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.212, + 0.761, + 0.261 + ], + "angle": 0, + "content": "
SourceAttackALBEFTCLCLIPvITCLIPCNN
TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1
ALBEFSGA w.o. Aug99.999.9570.0771.6730.5539.8832.3142.54
SGA w. Aug99.999.9887.8888.0536.6946.7839.5949.78
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.331 + ], + "angle": 0, + "content": "VLP models, ALBEF expands Grad-CAM [43] and utilizes the acquired attention map to rank the detected proposals [54]." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.332, + 0.787, + 0.424 + ], + "angle": 0, + "content": "Image Captioning (IC) is to generate a textual description that logically describes or implies the content of a given visual input, typically involving the creation of captions for images. The evaluation of Image Captioning models often employs metrics such as BLEU [39], METEOR [4], ROUGE [33], CIDEr [47] and SPICE [2], which serve to assess the quality and relevance of the generated captions in comparison to reference captions." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.445, + 0.589, + 0.463 + ], + "angle": 0, + "content": "2.3 Transferability of Adversarial Examples" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.471, + 0.788, + 0.636 + ], + "angle": 0, + "content": "Adversarial attacks [13-15, 18, 25] are typically categorized as white-box and black-box attacks. In a white-box setting [23, 26], the attacker has full access to the model, whereas black-box attacks [3, 40], more realistic in practical applications, occur when information about the model is limited. In the realm of image attacks [21, 22, 24], prevalent methods for crafting transferable adversarial examples often leverage data augmentation techniques (e.g., DIM [52], TIM [9], SIM [34], ADMIX [50], PAM [58]). Zhang et al. [57] introduced a white-box attack targeting popular VLP models for downstream tasks in the multimodal domain. Building upon this work, Lu et al. [36] proposed SGA, considering the diversity of adversarial examples by expanding single image-text pairs to sets of images and texts to conduct black-box attacks on VLP models." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.638, + 0.787, + 0.729 + ], + "angle": 0, + "content": "However, SGA primarily emphasizes diversity in the vicinity of adversarial examples during the optimization process, potentially increasing the risk of overfitting the victim model and impacting transferability. Therefore, our primary focus in this study is to further enhance the diversity of adversarial examples in a thoughtful manner and avoid an undue focus on adversarial example diversity solely around adversarial images." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.751, + 0.381, + 0.77 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.784, + 0.495, + 0.8 + ], + "angle": 0, + "content": "3.1 Background and Motivation" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Adversarial attacks on VLP models involve inducing a mismatch between adversarial images and corresponding adversarial text while adhering to specified" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "S. Gao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "constraints on image and text perturbations. Here, \\((v,t)\\) denotes an original image-text pair from a multimodal dataset, with \\(v^{\\prime}\\) representing an adversarial image and \\(t^\\prime\\) denoting adversarial text. The allowable perturbations are restricted within the ranges \\(B[v,\\xi_v]\\) for images and \\(B[t,\\xi_t]\\) for text. The image and text encoders of the multimodal model are denoted as \\(F_{I}\\) and \\(F_{T}\\), respectively. To generate valid multimodal adversarial examples, the objective is to maximize the loss function \\(J\\) specific to VLP models:" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.265, + 0.786, + 0.298 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\max J \\left(F _ {I} \\left(v ^ {\\prime}\\right), F _ {T} \\left(t ^ {\\prime}\\right)\\right) \\\\ s. t. v ^ {\\prime} \\in B [ v, \\xi_ {v} ], t ^ {\\prime} \\in B [ t, \\xi_ {t} ]. \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.308, + 0.789, + 0.43 + ], + "angle": 0, + "content": "The state-of-the-art approach for exploring the transferability of multimodal adversarial examples (i.e., SGA [36]) involves augmenting image-text pairs to enhance the diversity of adversarial examples along the optimization path. Specifically, during the iterative generation of adversarial images, let \\( v_{i}^{\\prime} \\) represent the adversarial image generated at the \\( i \\)-th step. In the subsequent step \\( (i + 1) \\), SGA initiates the process by applying a resizing operation for data augmentation to \\( v_{i}^{\\prime} \\), resulting in \\( V_{i}^{\\prime} = \\{v_{i1}^{\\prime}, v_{i2}^{\\prime}, \\dots, v_{iM}^{\\prime}\\} \\) (See Figure 1 (a)). The iterative formula can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.321, + 0.441, + 0.786, + 0.484 + ], + "angle": 0, + "content": "\\[\nv _ {i + 1} ^ {\\prime} = v _ {i} ^ {\\prime} + \\alpha \\cdot \\operatorname {s i g n} \\left(\\frac {\\nabla_ {v} \\sum_ {j = 1} ^ {M} J \\left(F _ {I} \\left(v _ {i j} ^ {\\prime}\\right) , F _ {T} (t)\\right)}{\\left\\| \\nabla_ {v} \\sum_ {j = 1} ^ {M} J \\left(F _ {I} \\left(v _ {i j} ^ {\\prime}\\right) , F _ {T} (t)\\right) \\right\\|}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.789, + 0.674 + ], + "angle": 0, + "content": "To further examine the impact of image augmentation along the optimization path in the SGA method, we utilize ALBEF as a surrogate model to generate multimodal adversarial examples. These examples are then employed to target VLP models such as TCL and CLIP, assessing the transferability of the attacks. Detailed results are presented in Table 1. Our observations reveal that SGA enhances the transferability of adversarial attacks, showing an increase ranging from \\(6.14\\%\\) to \\(17.81\\%\\). However, it is noteworthy that the success rate of attacks on the target models remains notably lower than that on the source model. This discrepancy is primarily attributed to the fact that SGA predominantly emphasizes diversity around AE \\(v_{i}^{\\prime}\\) during the optimization period, without adequately considering the diversity of adversarial examples toward the clean image, leading to the risk of overfitting the victim model and affecting the transferability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.789, + 0.841 + ], + "angle": 0, + "content": "For this purpose, we propose to consider diversification along the intersection region of adversarial trajectory, which encompasses the original image \\( v \\), the adversarial image from the previous step \\( v_{i-1}' \\), and the current adversarial image \\( v_i' \\) during the iterative attack process. This region is established to sample images within it to broaden the diversity of adversarial examples (See Figure 1 (b)). Moreover, to fully leverage the interplay between modalities, we aim for perturbations guided by textual information that induce \\( v_i' \\) to deviate significantly from the associated text \\( t \\). Additionally, in the text modality, our objective is to identify adversarial perturbations that simultaneously deviate from the intersection region rather than only adversarial images, thereby reducing overfitting the surrogate model and enhancing the effectiveness of black-box attacks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.381, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Boosting Transferability in Vision-Language Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.631, + 0.163 + ], + "angle": 0, + "content": "3.2 Diversification along the Intersection Region" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.339 + ], + "angle": 0, + "content": "As outlined in Section 3.1, we enhance the diversity of adversarial examples by introducing diversification along the intersection region. Specifically, at the \\(i\\)th iteration during optimization, we have the \\(v_{i}^{\\prime}\\), \\(v_{i-1}^{\\prime}\\), and the clean \\(v\\), and these variables form a triangle region denoted as \\(\\triangle vv_{i-1}^{\\prime}v_{i}^{\\prime}\\), i.e., the intersection region of adversarial trajectory in Figure 1. Then, we initially sample multiple instances within the region \\(\\triangle vv_{i-1}^{\\prime}v_{i}^{\\prime}\\), representing the set of samples as \\(e = \\{e_1,e_2,\\dots,e_N\\}\\). Each sample can be expressed as \\(e_k = \\beta \\cdot v + \\gamma \\cdot v_{i-1}^{\\prime} + \\eta \\cdot v_i^{\\prime}\\) where \\(\\beta + \\gamma + \\eta = 1.0\\). Consequently, we can compute the gradient perturbation for each sample. For the k-th sample, denoted as \\(e_k\\), its gradient perturbation \\(p_k\\) is calculated as follows. In this way, we can get a perturbation set \\(P = \\{p_1,p_2,\\dots,p_N\\}\\) by" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.349, + 0.786, + 0.381 + ], + "angle": 0, + "content": "\\[\np _ {k} = \\alpha \\cdot \\operatorname {s i g n} \\left(\\frac {\\nabla_ {e} J \\left(F _ {I} \\left(e _ {k}\\right) , F _ {T} (t)\\right)}{\\| \\nabla_ {e} J \\left(F _ {I} \\left(e _ {k}\\right) , F _ {T} (t)\\right) \\|}\\right). \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.401, + 0.564, + 0.417 + ], + "angle": 0, + "content": "3.3 Text-guided Augmentation Selection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.425, + 0.785, + 0.517 + ], + "angle": 0, + "content": "In Section 3.2, a diverse perturbation set \\( P \\) is derived from the intersection region. To harness the full potential of modality interactions, we introduce text-guided augmentation selection to obtain the optimal sample. Specifically, we individually incorporate each element from the perturbation set \\( P \\) into the adversarial image \\( v_{i}^{\\prime} \\). The selection process aims to identify the sample that maximally distances \\( v_{i}^{\\prime} \\) from \\( t \\). This procedure can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.376, + 0.531, + 0.785, + 0.556 + ], + "angle": 0, + "content": "\\[\nm = \\underset {p _ {m} \\in P} {\\arg \\max } J \\left(F _ {I} \\left(v _ {i} ^ {\\prime} + p _ {m}\\right), F _ {T} (t)\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.562, + 0.788, + 0.64 + ], + "angle": 0, + "content": "At this juncture, \\( e_m \\) represents the selected sample. We employ SGA as our baseline and incorporate the image augmentation methods considered along its optimization path. The chosen optimal sample \\( e_m \\) is resized and expanded into the set \\( E_m = \\{e_{m1}, e_{m2}, \\dots, e_{mM}\\} \\). Subsequently, we utilize the expanded set \\( E_m \\) to generate the final adversarial perturbation, yielding \\( v_{i+1}' \\):" + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.648, + 0.785, + 0.69 + ], + "angle": 0, + "content": "\\[\nv _ {i + 1} ^ {\\prime} = v _ {i} ^ {\\prime} + \\alpha \\cdot \\operatorname {s i g n} \\left(\\frac {\\nabla_ {e} \\sum_ {j = 1} ^ {M} J \\left(F _ {I} \\left(e _ {m j}\\right) , F _ {T} (t)\\right)}{\\| \\nabla_ {e} \\sum_ {j = 1} ^ {M} J \\left(F _ {I} \\left(e _ {m j}, F _ {T} (t)\\right) \\right\\|}\\right). \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.71, + 0.726, + 0.726 + ], + "angle": 0, + "content": "3.4 Adversarial Text deviating from the Intersection Region" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In the text modality, SGA only considers deviating adversarial text from the ultimate adversarial image generated during the iterative optimization process. If there are a total of T iterations, The generation of \\( t' \\) only considers deviating from the adversarial image \\( v_T' \\). The adversarial image \\( v_T' \\) is exclusively created by the surrogate model, thereby still presenting the risk of overfitting to the surrogate model. However, during the optimization process of the adversarial image, the clean image is entirely independent of the surrogate model. For this" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "S. Gao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.195 + ], + "angle": 0, + "content": "reason, we propose to have the text deviate simultaneously from the last intersection region along the optimization path. Specifically, the adversarial text deviates from the triangle region constituted by \\( v \\), \\( v_{T-1}' \\) and \\( v_T' \\)." + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.205, + 0.786, + 0.246 + ], + "angle": 0, + "content": "\\[\nt ^ {\\prime} = \\underset {t ^ {\\prime} \\in B [ t, \\epsilon_ {t} ]} {\\arg \\max } (\\lambda \\cdot J (F _ {I} (v), F _ {T} (t ^ {\\prime})) + \\mu \\cdot J (F _ {I} (v _ {T} ^ {\\prime}), F _ {T} (t ^ {\\prime}))) + \\nu \\cdot J (F _ {I} (v _ {T - 1} ^ {\\prime}), F _ {T} (t ^ {\\prime}))). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.246, + 0.712, + 0.261 + ], + "angle": 0, + "content": "We also set adjustable scaling factors, among which \\(\\lambda +\\mu +\\nu = 1.0\\)" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.284, + 0.457, + 0.3 + ], + "angle": 0, + "content": "3.5 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.788, + 0.477 + ], + "angle": 0, + "content": "In the specific process of our attack, we employ an iterative approach. In each iteration, we sample within the intersection region of adversarial trajectory, guided by textual information, to select a sample that maximally deviates the current adversarial image \\( v_{i}^{\\prime} \\) from the text \\( t \\). Subsequently, we subject this sample to image augmentation processing, calculate gradients to determine the perturbation direction, and overlay it onto the current adversarial image, resulting in \\( v_{i + 1}^{\\prime} \\). Through multiple iterative steps, we obtain an adversarial image \\( v_{T}^{\\prime} \\). For the text modality, in contrast to previous methods that solely focus on deviating from the adversarial image \\( v_{T}^{\\prime} \\), our goal is to derive an adversarial text \\( t^{\\prime} \\) that simultaneously deviates from the last intersection region \\( \\triangle vv_{T - 1}^{\\prime}v_{T}^{\\prime} \\) along the optimization path." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.501, + 0.376, + 0.518 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.788, + 0.701 + ], + "angle": 0, + "content": "In this section, we present experimental evidence demonstrating the enhanced transferability of multimodal examples generated from our proposed method across VLP models and various Vision-and-Language tasks. First, in Section 4.1, we introduce the experimental settings, including the popular image-text pair datasets and VLP models we use, as well as restrictions for adversarial attacks. Subsequently, the process of searching for optimal parameters is shown in Section 4.2. After that, we evaluate cross-model transferability in the context of the image-text retrieval task, as detailed in Section 4.3. Following this, in Section 4.4, we extend our investigation to the transfer of multimodal adversarial examples generated within the image-text retrieval task to other tasks, aiming to gauge cross-task transferability. Lastly, Section 4.5 outlines ablation studies." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.724, + 0.32, + 0.74 + ], + "angle": 0, + "content": "4.1 Setups" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.843 + ], + "angle": 0, + "content": "VLP Models. In our transferability evaluation experiments across various VLP models, we explore two typical architectures: fused and aligned VLP models. We select CLIP [42] for the aligned VLP model. CLIP offers a choice between two distinct image encoders, namely ViT-B/16 [10] and ResNet-101 [19], denoted as \\(\\mathrm{CLIP}_{\\mathrm{ViT}}\\) and \\(\\mathrm{CLIP}_{\\mathrm{CNN}}\\), respectively. In the case of fused VLP models, we opt for ALBEF [30] and TCL [53]. ALBEF uses a 12-layer ViT-B/16 image encoder" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.381, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Boosting Transferability in Vision-Language Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "and two 6-layer transformers for text and multimodal encoding. TCL shares this architecture but has different pre-training objectives." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.253 + ], + "angle": 0, + "content": "Datasets. In this study, we leverage two widely recognized multimodal datasets, namely Flickr30K [41] and MSCOCO [35], for evaluating the image-text retrieval task. The Flickr30K dataset comprises 31,783 images, each accompanied by five captions for annotation. Similarly, the MSCOCO dataset consists of 123,287 images, and approximately five captions are provided for each image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.344 + ], + "angle": 0, + "content": "Additionally, we employ the RefCOCO+ [55] dataset to assess the Visual Grounding task. RefCOCO+ is a dataset containing 141,564 referring expressions for 50,000 objects within 19,992 MSCOCO images. This dataset serves the purpose of evaluating grounding models by focusing on the localization of objects described through natural language. For another Vision-and-Language task, Image Captioning, we leverage the MSCOCO dataset as well." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Adversarial Attack Settings. In our study, we adopt adversarial attack settings of SGA [36] to ensure a fair comparison. Specifically, we leverage BERTAttack [31] to craft adversarial texts. The perturbation bound \\(\\xi_{t}\\) is set as 1 and length of word list \\(W = 10\\). PGD [38] is employed to get adversarial images and the perturbation bound, denoted as \\(\\xi_v\\), is set as 8/255. Additionally, iteration steps \\(T\\) is set as 10 and each step size \\(\\alpha = 2 / 255\\). Furthermore, when randomly sampling from the intersection region of adversarial trajectory, we set the number of samples to 5. When generating adversarial texts, we set the three parameters \\(\\lambda, \\mu, \\nu\\) in Equation 6 to 0.6, 0.2, and 0.2 respectively. The values chosen for these adjustable parameters can be found in Section 4.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.787, + 0.555 + ], + "angle": 0, + "content": "Evaluation Metrics. The key metric for adversarial transferability is the Attack Success Rate (ASR), which measures the percentage of successful attacks among all generated adversarial examples. A higher ASR indicates more effective and transferable attacks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.576, + 0.431, + 0.59 + ], + "angle": 0, + "content": "4.2 Optimal Parameters" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.688 + ], + "angle": 0, + "content": "In our proposed method, the number of samples \\(N\\) taken from the intersection region of adversarial trajectory and scaling factors in Equation 6 are adjustable. We conduct specific experiments to explore optimal parameter settings and examine their influence on the efficacy of our approach. To be more specific, we utilize ALBEF for generating multimodal adversarial examples on the Flickr30K dataset and assess the transferability on the other three VLP models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.795 + ], + "angle": 0, + "content": "Number of Samples taken from Intersection Region of Adversarial trajectory \\( N \\). The bottom of Table 2 illustrates when the sample size reaches 5, transfer effects are observed for both the Image Retrieval task and the Text Retrieval task. As the sample size continues to increase, the transferability only fluctuates and does not exhibit further improvement. Taking into account both transfer effects and computational costs, a sample size of 5 is identified as the optimal configuration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Scaling factors \\(\\lambda, \\mu, \\nu\\) in Adversarial Text Generation. In Equation 6, \\(\\lambda\\) represents the weight of clean images, while \\(\\mu\\) and \\(\\nu\\) represent the weights of adversarial images. Therefore, we stipulate that \\(\\lambda\\) cannot be zero, and \\(\\mu\\) and \\(\\nu\\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "S. Gao et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 2: Optimal Parameters: Attack Success Rate(%) on different settings, Top for different values of \\(\\lambda ,\\mu ,\\nu\\) and Bottom for different numbers of samples \\(N\\)" + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.198, + 0.761, + 0.376 + ], + "angle": 0, + "content": "
SourceAttackALBEFTCLCLIPvITCLIPCNN
TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1
ALBEF[λ,μ,v] = [0.2,0.0,0.8]99.999.9389.6790.542.2152.046.2355.44
[λ,μ,v] = [0.2,0.2,0.6]99.999.9390.4190.4341.9651.8746.4955.3
[λ,μ,v] = [0.2,0.4,0.4]99.999.9390.3190.4341.9651.8745.3454.82
[λ,μ,v] = [0.2,0.6,0.2]99.999.9590.5290.5742.2151.8445.0854.92
[λ,μ,v] = [0.2,0.8,0.0]99.999.9390.3190.5741.7251.7446.154.68
[λ,μ,v] = [0.4,0.0,0.6]99.999.9391.2590.8845.5255.3248.9157.63
[λ,μ,v] = [0.4,0.2,0.4]99.999.9391.2590.8345.0355.2248.2857.77
[λ,μ,v] = [0.4,0.4,0.2]99.999.9391.1590.7144.9155.2848.1557.87
[λ,μ,v] = [0.4,0.6,0.0]99.999.9391.1590.8844.9154.7749.0457.56
[λ,μ,v] = [0.6,0.0,0.4]99.999.9391.4690.9546.3856.3849.0459.11
[λ,μ,v] = [0.6,0.2,0.2]99.999.9391.5791.1746.2656.849.5559.01
[λ,μ,v] = [0.6,0.4,0.0]99.999.9390.9490.9846.0156.7249.5558.87
[λ,μ,v] = [0.8,0.0,0.2]99.999.9390.7390.9846.1356.7149.4658.74
[λ,μ,v] = [0.8,0.2,0.0]99.999.9390.3190.9545.7756.6549.3458.87
ALBEFN=399.999.9590.6290.7945.6456.5448.8358.73
N=499.7999.9191.3691.1745.8356.7850.4559.01
N=599.999.9391.5791.1746.2656.849.5559.01
N=699.999.9390.9490.3845.7956.9650.758.52
N=799.999.9189.8890.9545.456.3550.759.07
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.788, + 0.507 + ], + "angle": 0, + "content": "can have at most one zero value. Initially, as the value of \\(\\lambda\\) gradually increases, indicating the gradual introduction of clean images, the transferability of multimodal adversarial examples increases accordingly. However, when the value of \\(\\lambda\\) becomes too large, it also leads to a disproportionately low proportion of adversarial images, resulting in a decrease in adversarial transferability. According to the experiments, the optimal parameters we select are [0.6, 0.2, 0.2]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.529, + 0.494, + 0.545 + ], + "angle": 0, + "content": "4.3 Cross-Model Transferability" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.704 + ], + "angle": 0, + "content": "As outlined in Section 4.1, our experimental design focuses on assessing the transferability of adversarial examples across two widely adopted VLP model architectures: fused and aligned. There are four VLP models selected, namely: ALBEF, TCL, \\(\\mathrm{CLIP}_{\\mathrm{ViT}}\\), \\(\\mathrm{CLIP}_{\\mathrm{CNN}}\\). The selected downstream \\(\\mathrm{V} + \\mathrm{L}\\) task for evaluation is image-text retrieval. Our approach involves employing one of four models to generate multimodal adversarial examples within the specified parameters of our proposed method. Subsequently, we validate the effectiveness of the generated adversarial examples through a comprehensive set of experiments, encompassing both self-attacks and attacks on the other three models. This evaluation encompasses one white-box attack and three distinct black-box attacks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We adopt the methodology proposed by SGA [36] as our baseline, Therefore, the effectiveness of our method is compared against it. Moreover, we also present the effectiveness and transferability results of various other attack methods on VLP models. Table 3 provides a comprehensive comparison of these methods on the dataset Flickr30K, More experiments on the MSCOCO dataset are provided in the Appendix. In the comparison, PGD [38] is an image-only attack, while Bert-Attack [31] focuses solely on text-based attacks. Sep-Attack involves perturbing text and image separately. Furthermore, Co-Attack takes into account cross-modal interactions, generating an adversarial example for one modality" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.382, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Boosting Transferability in Vision-Language Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.162, + 0.274, + 0.185 + ], + "angle": 0, + "content": "Cleannnnnne" + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.195, + 0.272, + 0.226 + ], + "angle": 0, + "content": "Caaee" + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.149, + 0.346, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.284, + 0.197, + 0.346, + 0.23 + ], + "angle": 0, + "content": "a man riding on the back of a motorcycle down a street" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.148, + 0.415, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.197, + 0.415, + 0.222 + ], + "angle": 0, + "content": "a man wearing a black shirt and a green tie" + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.148, + 0.481, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.42, + 0.197, + 0.485, + 0.222 + ], + "angle": 0, + "content": "a parking meter sitting on the side of a road" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.148, + 0.549, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.489, + 0.197, + 0.549, + 0.222 + ], + "angle": 0, + "content": "a cat standing on a toilet seat in a bathroom" + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.148, + 0.616, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.197, + 0.611, + 0.222 + ], + "angle": 0, + "content": "am an riding a bike next to a train" + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.148, + 0.684, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.197, + 0.686, + 0.222 + ], + "angle": 0, + "content": "a man sitting on a bench next to a white dog" + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.148, + 0.751, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.197, + 0.748, + 0.229 + ], + "angle": 0, + "content": "a siamese cat eating a banana from a person's hand" + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.233, + 0.275, + 0.276 + ], + "angle": 0, + "content": "Adversarial" + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.284, + 0.272, + 0.315 + ], + "angle": 0, + "content": "Caaee" + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.233, + 0.347, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.286, + 0.286, + 0.345, + 0.31 + ], + "angle": 0, + "content": "a woman sitting on a motorcycle in the street" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.233, + 0.415, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.286, + 0.411, + 0.31 + ], + "angle": 0, + "content": "a young girl wearing glasses and a tie" + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.233, + 0.481, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.421, + 0.286, + 0.489, + 0.31 + ], + "angle": 0, + "content": "a group of people riding motorcycles down a street" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.233, + 0.549, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.489, + 0.286, + 0.551, + 0.31 + ], + "angle": 0, + "content": "a black and white dog laying on top of a toilet" + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.233, + 0.616, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.286, + 0.622, + 0.31 + ], + "angle": 0, + "content": "a train on a train track with a dog standing next to it" + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.233, + 0.684, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.628, + 0.285, + 0.678, + 0.31 + ], + "angle": 0, + "content": "a boy riding a bike next to a white dog" + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.233, + 0.751, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.689, + 0.285, + 0.757, + 0.31 + ], + "angle": 0, + "content": "a small dog is eating a banana on the floor" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.33, + 0.785, + 0.387 + ], + "angle": 0, + "content": "Fig. 2: Visualization on Image Captioning. We use the ALBEF model, pre-trained on Image Text Retrieval(ITR) task, to generate adversarial images on the MSCOCO dataset and use the BLIP [29] model for Image Captioning on both clean images and adversarial images, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.785, + 0.447 + ], + "angle": 0, + "content": "under the guidance of the other modality. SGA, our baseline, expands a single image-text pair into a set of images and a set of texts to enhance diversity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.785, + 0.613 + ], + "angle": 0, + "content": "First, we compare the performance of various methods under white-box attacks, wherein we can leverage the model architecture and interact with the model. It is evident that our method, along with SGA, performs the best in the four white-box attack experiments compared to other methods. Whether TR or IR, the attack success rate at the top-1 rank \\((\\mathbb{R}@\\mathbb{1})\\) consistently exceeds \\(99.9\\%\\). Given the already high white-box attack success rate achieved by the SGA method, there is limited room for improvement in our approach within this context. Subsequently, we shift our focus to elucidating the enhancements our method brings to transferability, specifically in the realm of black-box attack performance. We delineate this exploration into two segments, contingent on whether the source model and the target model share the same architecture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Cross-Model Transferability in Same Architecture. ALBEF and TCL both utilize a similar model architecture but differ in their pre-training objectives while maintaining a common fundamental model structure. Therefore, when ALBEF and TCL are employed as target models for each other, the success rate of attacks using multimodal adversarial examples is remarkably high. Notably, the black-box attack success rate of IR reaches \\(95.58\\%\\) when TCL is the target model in our proposed method. Additionally, the transferability of SGA can reach approximately \\(90\\%\\), but the room for improvement is very limited, our method has improved compared to SGA, ranging from \\(2.71\\%\\) to \\(3.69\\%\\). In contrast, when considering \\(\\mathrm{CLIP}_{\\mathrm{ViT}}\\) and \\(\\mathrm{CLIP}_{\\mathrm{CNN}}\\)—both being aligned VLP models—their image encoders vary significantly, with one utilizing the Vision Transformer and the other employing ResNet-101. Given the significant structural differences between traditional CNNs and Transformers, existing methods show low cross-model transferability, highlighting room for improvement. Our method outperforms SGA by about \\(7.67\\%\\) to \\(11.62\\%\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "S. Gao et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 3: Comparison with state-of-the-art methods on image-text retrieval. The source column shows VLP models we use to generate multimodal adversarial examples. The gray area represents adversarial attacks under a white-box setting, the rest are black-box attacks. For both Image Retrieval and Text Retrieval, we provide R@1 attack success rate(%)" + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.239, + 0.761, + 0.481 + ], + "angle": 0, + "content": "
SourceAttackALBEFTCLCLIP_VITCLIP_CNN
TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1
ALBEFPGD93.7494.4324.0327.910.6715.8214.0519.11
BERT-Attack11.5727.4612.6428.0729.3343.1732.6946.11
Sep-Attack95.7296.1439.351.7934.1145.7235.7647.92
Co-Attack97.0898.3639.5251.2429.8238.9231.2941.99
SGA99.999.9887.8888.0536.6946.7839.5949.78
Ours99.999.9391.5791.1746.2656.849.5559.01
TCLPGD35.7741.6799.3799.3310.1816.314.8121.1
BERT-Attack11.8926.8214.5429.1729.6944.4933.4646.07
Sep-Attack52.4561.4499.5899.4537.0645.8137.4249.91
Co-Attack49.8460.3691.6895.4832.6442.6932.0647.82
SGA92.4992.77100.0100.036.8146.9741.8951.53
Ours95.295.58100.099.9847.2457.2852.2362.23
CLIP_VITPGD3.136.484.438.8369.3384.7913.0317.43
BERT-Attack9.5922.6411.8025.0728.3439.0830.4037.43
Sep-Attack7.6120.5810.1220.7476.9387.4429.8938.32
Co-Attack8.5520.1810.0121.2978.5387.529.538.49
SGA22.4234.5925.0836.45100.0100.053.2661.1
Ours27.8442.8427.8244.6100.0100.064.8869.5
CLIP_CNNPGD2.296.154.538.885.412.0889.7893.04
BERT-Attack8.8623.2712.3325.4827.1237.4430.4040.10
Sep-Attack9.3822.9911.2825.4526.1339.2493.6195.3
Co-Attack10.5323.6212.5426.0527.2440.6295.9196.5
SGA15.6428.618.0233.0739.0251.4599.8799.9
Ours19.534.5921.637.8848.4759.1299.8799.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.66 + ], + "angle": 0, + "content": "Cross-Model Transferability in Different Architectures. When fused VLP models (i.e., ALBEF and TCL) are used for generating adversarial examples, our method achieves a significant improvement compared to existing methods, outperforming the current state-of-the-art (SOTA) methods by \\(9.23\\%\\) to \\(10.7\\%\\). When the fused VLP models are targeted for attack, our method still outperforms all existing methods when generating adversarial examples against the aligned VLP models (i.e., CLIP). In this scenario, the transferability of all methods is relatively low. Compared to our method's baseline SGA, we attain an improvement ranging from \\(2.74\\%\\) to \\(8.25\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.688, + 0.48, + 0.705 + ], + "angle": 0, + "content": "4.4 Cross-Task Transferability" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.787, + 0.841 + ], + "angle": 0, + "content": "We not only assess the transferability of multimodal adversarial examples generated by our proposed method across different VLP models but also conduct experiments to evaluate its effectiveness in transferring across diverse \\(\\mathrm{V} + \\mathrm{L}\\) tasks. Specifically, we craft adversarial examples for the Image-Text Retrieval (ITR) task and evaluate them on Visual Grounding (VG) and Image Captioning (IC) tasks. As evident from Table 4 and visual results in Figure 2 and 3, the adversarial examples generated for ITR demonstrate transferability, successfully impacting both VG and IC tasks. This highlights the efficacy of cross-task trans" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.382, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Boosting Transferability in Vision-Language Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.151, + 0.75, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.491 + ], + "angle": 0, + "content": "Fig. 3: Visualization on Visual Grounding. We use the ALBEF model, pre-trained on the ITR task, to generate adversarial images on the RefCOCO+ dataset and use the same model, pre-trained on Visual Grouding(VG) task, to localize the regions corresponding to red words on both clean images and adversarial images, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.785, + 0.551 + ], + "angle": 0, + "content": "ferability in our proposed method. Furthermore, our transferability consistently outperforms that of SGA." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.573, + 0.39, + 0.589 + ], + "angle": 0, + "content": "4.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Our proposed method builds upon the SGA [36] as a baseline, introducing two key improvements. Firstly, we utilize diversification along the intersection region of adversarial trajectory to expand the diversity of adversarial examples. Secondly, we generate adversarial text while simultaneously distancing it from the last intersection region along the optimization path. To investigate the impact of each improvement on the effectiveness of our method, we conduct ablation studies on the ITR task and employ transferability from ALBEF to the other three VLP models as the evaluation metric." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In the ablation study, we systematically eliminate each enhancement from our approach and compare their transferability with both the baseline SGA and our final method, which incorporates a combination of two improvements. The results are depicted in Figure 4. It is evident that when attacking models with the same architecture versus different architectures, the impacts of two distinct improvements are also different. When attacking TCL with the same structure as ALBEF, under Setting1, our method exhibits a greater decrease in transferability, indicating that diversification along the intersection region is" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "S. Gao et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.214 + ], + "angle": 0, + "content": "Table 4: Cross-Task Transferability. We utilize ALBEF to generate multimodal adversarial examples for attacking both Visual Grounding(VG) on the RefCOCO+ dataset and Image Captioning(IC) on the MSCOCO dataset. The baseline represents the performance of each task without any attack, where a lower value indicates better effectiveness of the adversarial attack for both tasks." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.239, + 0.761, + 0.31 + ], + "angle": 0, + "content": "
AttackITR → VGITR → IC
ValTestATestBB@4METEORROUGE-LCIDErSPICE
Baseline58.4665.8946.2539.731.060.0133.323.8
SGA50.5657.4240.6628.024.651.291.417.7
Ours49.7056.3240.5427.224.250.788.317.2
" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.336, + 0.415, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.338, + 0.585, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.338, + 0.756, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.515 + ], + "angle": 0, + "content": "Fig. 4: Ablation Study: Attack Success Rate(%) on other three target models. The baseline is SGA. Setting 1 removes diversification along the intersection region of adversarial trajectory. Setting 2 removes the text deviating from the last intersection region along the optimization path." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.788, + 0.605 + ], + "angle": 0, + "content": "most crucial in our approach under this scenario. However, under Setting2, the decrease in transferability is more pronounced when attacking the aligned VLP model, indicating that at this point, adversarial texts deviating from the last intersection region play a larger role in the effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.628, + 0.36, + 0.644 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this paper, we conduct a systematic evaluation of existing multimodal attacks regarding transferability. We found that these methods predominantly prioritize diversity around adversarial examples (AEs) during the optimization process, potentially leading to overfitting to the victim model and hindering transferability. To address this issue, we propose diversification along the intersection region of adversarial trajectory to broaden diversity not only around AEs but also towards clean inputs. Moreover, we pioneer an exploration into extending adversarial texts deviating from the intersection region. Through extensive experiments, we demonstrate the effectiveness of our method in enhancing transferability across VLP models and \\(\\mathrm{V} + \\mathrm{L}\\) tasks. This work could act as a catalyst for more profound research on the transferability of multimodal AEs, alongside fortifying the adversarial robustness of VLP models." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.381, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Boosting Transferability in Vision-Language Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.788, + 0.273 + ], + "angle": 0, + "content": "This research is supported by the National Research Foundation, Singapore, and DSO National Laboratories under the AI Singapore Programme (AISG Award No: AISG2-GC-2023-008), Career Development Fund (CDF) of Agency for Science, Technology and Research (A*STAR) (No.: C233312028), and National Research Foundation, Singapore and Infocomm Media Development Authority under its Trust Tech Funding Initiative (No. DTC-RGC-04)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.299, + 0.323, + 0.315 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.334, + 0.785, + 0.375 + ], + "angle": 0, + "content": "1. Abdelfattah, R., Guo, Q., Li, X., Wang, X., Wang, S.: Cdul: Clip-driven unsupervised learning for multi-label image classification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1348-1357 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.432 + ], + "angle": 0, + "content": "2. Anderson, P., Fernando, B., Johnson, M., Gould, S.: Spice: Semantic propositional image caption evaluation. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 382-398. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.488 + ], + "angle": 0, + "content": "3. Bai, Y., Zeng, Y., Jiang, Y., Wang, Y., Xia, S.T., Guo, W.: Improving query efficiency of black-box adversarial attack. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16. pp. 101-116. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.49, + 0.785, + 0.543 + ], + "angle": 0, + "content": "4. Banerjee, S., Lavie, A.: Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In: Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization. pp. 65-72 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.546, + 0.785, + 0.6 + ], + "angle": 0, + "content": "5. Chen, H., Ding, G., Liu, X., Lin, Z., Liu, J., Han, J.: Imram: Iterative matching with recurrent attention memory for cross-modal image-text retrieval. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12655-12663 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.602, + 0.785, + 0.643 + ], + "angle": 0, + "content": "6. Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: Uniter: Universal image-text representation learning. In: European conference on computer vision. pp. 104-120. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.645, + 0.785, + 0.685 + ], + "angle": 0, + "content": "7. Cheng, H., Xiao, E., Cao, J., Yang, L., Xu, K., Gu, J., Xu, R.: Typography leads semantic diversifying: Amplifying adversarial transferability across multimodal large language models. arXiv preprint arXiv:2405.20090 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.687, + 0.785, + 0.741 + ], + "angle": 0, + "content": "8. Cheng, M., Sun, Y., Wang, L., Zhu, X., Yao, K., Chen, J., Song, G., Han, J., Liu, J., Ding, E., et al.: Vista: vision and scene text aggregation for cross-modal retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5184-5193 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.743, + 0.785, + 0.784 + ], + "angle": 0, + "content": "9. Dong, Y., Pang, T., Su, H., Zhu, J.: Evading defenses to transferable adversarial examples by translation-invariant attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4312-4321 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arxiv 2020. arXiv preprint arXiv:2010.11929 (2010)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.334, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.357, + 0.127 + ], + "angle": 0, + "content": "S. Gao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "11. Dou, Z.Y., Xu, Y., Gan, Z., Wang, J., Wang, S., Wang, L., Zhu, C., Zhang, P., Yuan, L., Peng, N., et al.: An empirical study of training end-to-end vision-and-language transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18166-18176 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.787, + 0.247 + ], + "angle": 0, + "content": "12. Gao, K., Bai, Y., Bai, J., Yang, Y., Xia, S.T.: Adversarial robustness for visual grounding of multimodal large language models. arXiv preprint arXiv:2405.09981 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.248, + 0.787, + 0.289 + ], + "angle": 0, + "content": "13. Gu, J., Jia, X., de Jorge, P., Yu, W., Liu, X., Ma, A., Xun, Y., Hu, A., Khakzar, A., Li, Z., et al.: A survey on transferability of adversarial examples across deep neural networks. arXiv preprint arXiv:2310.17626 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.787, + 0.331 + ], + "angle": 0, + "content": "14. Guo, Q., Cheng, Z., Juefei-Xu, F., Ma, L., Xie, X., Liu, Y., Zhao, J.: Learning to adversarially blur visual object tracking. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10839-10848 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.787, + 0.374 + ], + "angle": 0, + "content": "15. Guo, Q., Juefei-Xu, F., Xie, X., Ma, L., Wang, J., Yu, B., Feng, W., Liu, Y.: Watch out! motion is blurring the vision of your deep neural networks. In: Advances in Neural Information Processing Systems. vol. 33, pp. 975-985 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.787, + 0.416 + ], + "angle": 0, + "content": "16. Han, D., Jia, X., Bai, Y., Gu, J., Liu, Y., Cao, X.: Ot-attack: Enhancing adversarial transferability of vision-language models via optimal transport optimization. arXiv preprint arXiv:2312.04403 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.418, + 0.787, + 0.459 + ], + "angle": 0, + "content": "17. He, B., Jia, X., Liang, S., Lou, T., Liu, Y., Cao, X.: Sa-attack: Improving adversarial transferability of vision-language pre-training models via self-augmentation. arXiv preprint arXiv:2312.04913 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.787, + 0.502 + ], + "angle": 0, + "content": "18. He, B., Liu, J., Li, Y., Liang, S., Li, J., Jia, X., Cao, X.: Generating transferable 3d adversarial point cloud via random perturbation factorization. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 764-772 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.503, + 0.787, + 0.544 + ], + "angle": 0, + "content": "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.545, + 0.787, + 0.599 + ], + "angle": 0, + "content": "20. Hu, X., Gan, Z., Wang, J., Yang, Z., Liu, Z., Lu, Y., Wang, L.: Scaling up vision-language pre-training for image captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17980-17989 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.787, + 0.657 + ], + "angle": 0, + "content": "21. Huang, Y., Guo, Q., Juefei-Xu, F., Ma, L., Miao, W., Liu, Y., Pu, G.: Advfilter: predictive perturbation-aware filtering against adversarial attack via multi-domain learning. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 395-403 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.787, + 0.713 + ], + "angle": 0, + "content": "22. Huang, Y., Juefei-Xu, F., Guo, Q., Zhang, J., Wu, Y., Hu, M., Li, T., Pu, G., Liu, Y.: Personalization as a shortcut for few-shot backdoor attack against text-to-image diffusion models. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 21169-21178 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.787, + 0.755 + ], + "angle": 0, + "content": "23. Huang, Y., Sun, L., Guo, Q., Juefei-Xu, F., Zhu, J., Feng, J., Liu, Y., Pu, G.: Ala: Naturalness-aware adversarial lightness attack. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 2418-2426 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.798 + ], + "angle": 0, + "content": "24. Huang, Y., Sun, L., Guo, Q., Juefei-Xu, F., Zhu, J., Feng, J., Liu, Y., Pu, G.: Ala: Naturalness-aware adversarial lightness attack. In: Proceedings of the 31st ACM International Conference on Multimedia. p. 2418-2426 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "25. Jia, X., Wei, X., Cao, X., Han, X.: Adv-watermark: A novel watermark perturbation for adversarial examples. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 1579-1587 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.382, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Boosting Transferability in Vision-Language Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.189 + ], + "angle": 0, + "content": "26. Jia, X., Zhang, Y., Wei, X., Wu, B., Ma, K., Wang, J., Cao Sr, X.: Improving fast adversarial training with prior-guided knowledge. arXiv preprint arXiv:2304.00202 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "27. Khan, Z., Fu, Y.: Exploiting bert for multimodal target sentiment classification through input space translation. In: Proceedings of the 29th ACM international conference on multimedia. pp. 3034-3042 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.786, + 0.286 + ], + "angle": 0, + "content": "28. Lei, C., Luo, S., Liu, Y., He, W., Wang, J., Wang, G., Tang, H., Miao, C., Li, H.: Understanding chinese video and language via contrastive multimodal pre-training. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 2567-2576 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.786, + 0.328 + ], + "angle": 0, + "content": "29. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.786, + 0.37 + ], + "angle": 0, + "content": "30. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems 34, 9694-9705 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.786, + 0.397 + ], + "angle": 0, + "content": "31. Li, L., Ma, R., Guo, Q., Xue, X., Qiu, X.: Bert-attack: Adversarial attack against bert using bert (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.786, + 0.453 + ], + "angle": 0, + "content": "32. Li, X., Yin, X., Li, C., Zhang, P., Hu, X., Zhang, L., Wang, L., Hu, H., Dong, L., Wei, F., et al.: Oscar: Object-semantics aligned pre-training for vision-language tasks. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16. pp. 121-137. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.786, + 0.48 + ], + "angle": 0, + "content": "33. Lin, C.Y.: Rouge: A package for automatic evaluation of summaries. In: Text summarization branches out. pp. 74-81 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.786, + 0.508 + ], + "angle": 0, + "content": "34. Lin, J., Song, C., He, K., Wang, L., Hopcroft, J.E.: Nesterov accelerated gradient and scale invariance for adversarial attacks. arXiv preprint arXiv:1908.06281 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.786, + 0.563 + ], + "angle": 0, + "content": "35. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.786, + 0.605 + ], + "angle": 0, + "content": "36. Lu, D., Wang, Z., Wang, T., Guan, W., Gao, H., Zheng, F.: Set-level guidance attack: Boosting adversarial transferability of vision-language pre-training models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.786, + 0.646 + ], + "angle": 0, + "content": "37. Luo, H., Gu, J., Liu, F., Torr, P.: An image is worth 1000 lies: Adversarial transferability across prompts on vision-language models. arXiv preprint arXiv:2403.09766 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.786, + 0.674 + ], + "angle": 0, + "content": "38. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.786, + 0.715 + ], + "angle": 0, + "content": "39. Naseer, M.M., Ranasinghe, K., Khan, S.H., Hayat, M., Shahbaz Khan, F., Yang, M.H.: Intriguing properties of vision transformers. Advances in Neural Information Processing Systems 34, 23296-23308 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.786, + 0.757 + ], + "angle": 0, + "content": "40. Park, J., Miller, P., McLaughlin, N.: Hard-label based small query black-box adversarial attack. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3986-3995 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.786, + 0.813 + ], + "angle": 0, + "content": "41. Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In: Proceedings of the IEEE international conference on computer vision. pp. 2641-2649 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.786, + 0.84 + ], + "angle": 0, + "content": "42. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "S. Gao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.786, + 0.232 + ], + "angle": 0, + "content": "43. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.233, + 0.786, + 0.273 + ], + "angle": 0, + "content": "44. Shi, L., Shuang, K., Geng, S., Gao, P., Fu, Z., de Melo, G., Chen, Y., Su, S.: Dense contrastive visual-linguistic pretraining. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 5203-5212 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.274, + 0.786, + 0.3 + ], + "angle": 0, + "content": "45. Tan, H., Bansal, M.: Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.301, + 0.786, + 0.342 + ], + "angle": 0, + "content": "46. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: International conference on machine learning. pp. 10347-10357. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.343, + 0.786, + 0.384 + ], + "angle": 0, + "content": "47. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4566-4575 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.385, + 0.786, + 0.438 + ], + "angle": 0, + "content": "48. Wang, T., Ge, Y., Zheng, F., Cheng, R., Shan, Y., Qie, X., Luo, P.: Accelerating vision-language pretraining with free language modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23161-23170 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.439, + 0.786, + 0.481 + ], + "angle": 0, + "content": "49. Wang, T., Jiang, W., Lu, Z., Zheng, F., Cheng, R., Yin, C., Luo, P.: Vlmixer: Unpaired vision-language pre-training via cross-modal cutmix. In: International Conference on Machine Learning. pp. 22680-22690. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.482, + 0.786, + 0.522 + ], + "angle": 0, + "content": "50. Wang, X., He, X., Wang, J., He, K.: Admix: Enhancing the transferability of adversarial attacks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16158-16167 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.523, + 0.786, + 0.564 + ], + "angle": 0, + "content": "51. Wang, Z., Liu, X., Li, H., Sheng, L., Yan, J., Wang, X., Shao, J.: Camp: Cross-modal adaptive message passing for text-image retrieval. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5764-5773 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.565, + 0.786, + 0.618 + ], + "angle": 0, + "content": "52. Xie, C., Zhang, Z., Zhou, Y., Bai, S., Wang, J., Ren, Z., Yuille, A.L.: Improving transferability of adversarial examples with input diversity. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2730-2739 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.62, + 0.786, + 0.674 + ], + "angle": 0, + "content": "53. Yang, J., Duan, J., Tran, S., Xu, Y., Chanda, S., Chen, L., Zeng, B., Chilimbi, T., Huang, J.: Vision-language pre-training with triple contrastive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15671-15680 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.675, + 0.786, + 0.729 + ], + "angle": 0, + "content": "54. Yu, L., Lin, Z., Shen, X., Yang, J., Lu, X., Bansal, M., Berg, T.L.: Mattnet: Modular attention network for referring expression comprehension. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1307-1315 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.73, + 0.786, + 0.785 + ], + "angle": 0, + "content": "55. Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in referring expressions. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 69-85. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.786, + 0.786, + 0.84 + ], + "angle": 0, + "content": "56. Yuan, L., Chen, Y., Wang, T., Yu, W., Shi, Y., Jiang, Z.H., Tay, F.E., Feng, J., Yan, S.: Tokens-to-token vit: Training vision transformers from scratch on imagenet. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 558-567 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.381, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Boosting Transferability in Vision-Language Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "57. Zhang, J., Yi, Q., Sang, J.: Towards adversarial attack on vision-language pretraining models. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 5005-5013 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.245 + ], + "angle": 0, + "content": "58. Zhang, J., Huang, J.t., Wang, W., Li, Y., Wu, W., Wang, X., Su, Y., Lyu, M.R.: Improving the transferability of adversarial samples by path-augmented method. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8173-8182 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.246, + 0.788, + 0.3 + ], + "angle": 0, + "content": "59. Zhang, P., Li, X., Hu, X., Yang, J., Zhang, L., Wang, L., Choi, Y., Gao, J.: Vinyl: Revisiting visual representations in vision-language models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5579-5588 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.342 + ], + "angle": 0, + "content": "60. Zhang, Q., Lei, Z., Zhang, Z., Li, S.Z.: Context-aware attention network for image-text retrieval. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3536-3545 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.342 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory/d6848265-b938-496f-9e0e-c16ce7da086b_origin.pdf b/2024/Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory/d6848265-b938-496f-9e0e-c16ce7da086b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a9df25c3892e7b452e24dc23dbbffdc48663a249 --- /dev/null +++ b/2024/Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory/d6848265-b938-496f-9e0e-c16ce7da086b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dec23c1eb6fc28f27ce8e761cd188d507950d57033fbc3e567aa9ad90cbf9aa6 +size 3968822 diff --git a/2024/Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory/full.md b/2024/Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5b7476a9ce654f3dcb5bad94d582f8db3174f3f5 --- /dev/null +++ b/2024/Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory/full.md @@ -0,0 +1,341 @@ +# Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory + +Sensen Gao $^{1,3}$ , Xiaojun Jia $^{2*}$ , Xuhong Ren $^{2}$ , Ivor Tsang $^{2,3}$ , and Qing Guo $^{3\dagger}$ + +1 Nankai University, Tianjin, China + +$^{2}$ Nanyang Technological University, Singapore + +3 CFAR and IHPC, Agency for Science, Technology and Research (A*STAR), Singapore + +Abstract. Vision-language pre-training (VLP) models exhibit remarkable capabilities in comprehending both images and text, yet they remain susceptible to multimodal adversarial examples (AEs). Strengthening attacks and uncovering vulnerabilities, especially common issues in VLP models (e.g., high transferable AEs), can advance reliable and practical VLP models. A recent work (i.e., Set-level guidance attack) indicates that augmenting image-text pairs to increase AE diversity along the optimization path enhances the transferability of adversarial examples significantly. However, this approach predominantly emphasizes diversity around the online adversarial examples (i.e., AEs in the optimization period), leading to the risk of overfitting the victim model and affecting the transferability. In this study, we posit that the diversity of adversarial examples towards the clean input and online AEs are both pivotal for enhancing transferability across VLP models. Consequently, we propose using diversification along the intersection region of adversarial trajectory to expand the diversity of AEs. To fully leverage the interaction between modalities, we introduce text-guided adversarial example selection during optimization. Furthermore, to further mitigate the potential overfitting, we direct the adversarial text deviating from the last intersection region along the optimization path, rather than adversarial images as in existing methods. Extensive experiments affirm the effectiveness of our method in improving transferability across various VLP models and downstream vision-and-language tasks. Code is available at https://github.com/SensenGao/VLTPredictAttack. + +Keywords: Vision-Language Attack $\cdot$ Adversarial Transferability $\cdot$ Diversification $\cdot$ Intersection Region of Adversarial Trajectory + +# 1 Introduction + +Vision-language pre-training (VLP) models utilize multimodal learning, leveraging large-scale image-text pairs to bridge the gap between visual and lan + +![](images/f6a4ec0e69e781aa3dff3ed1823ef6f5ab5b39259ef55a4b14a441af60845903.jpg) +(a) SGA + +![](images/1433f9d0752d691d5c98279c201a9d8e7f6b4247bd6c9ca74ec0679249d814a4.jpg) + +![](images/d4cfe77792677fc82598e233a866afc7b3630619180647168f842c8a85f8716c.jpg) +(c) ALBEF to $\mathrm{CLIP}_{\mathrm{CNN}}$ + +![](images/a1b3abab916e308cf1df4006807d87b09cc0a3194075a0ad8300b6fae7cc8653.jpg) +(b) Ours +(d) $\mathrm{CLIP}_{\mathrm{ViT}}$ to $\mathrm{CLIP}_{\mathrm{CNN}}$ +Fig. 1: Our method vs. set-level guided attack (SGA) [36]. (a) shows the main idea of SGA, i.e., conducting augmentation around the online adversarial examples. (b) shows the main idea of our method, that is, we perform augmentation in the intersection region of adversarial trajectory. The red and blue dots both depict images sampled from the intersection region, with red dots indicating the best samples selected using the text-guided adversarial example selection strategy. The surrounding light red dots represent applying the same resizing data augmentation to the best samples as utilized in SGA. (c) and (d) compare the transferability of our method and SGA by using the adversarial examples of ALBEF [30] and $\mathrm{CLIP}_{\mathrm{ViT}}$ to attack $\mathrm{CLIP}_{\mathrm{CNN}}$ , respectively. + +guage understanding. These models showcase revolutionary performance across various downstream Vision-and-Language tasks, including image-text retrieval, image captioning, visual grounding, and visual entailment, as demonstrated in [20, 27, 28, 44]. Despite their success, recent research underscores the significant vulnerability of VLP models, particularly when confronted with multimodal adversarial examples [7, 12, 16, 17, 36, 37, 57]. Uncovering the vulnerabilities, especially common issues, can drive further research aimed at building more reliable and practical VLP models. + +Current research predominantly concentrates on attacking VLP models via a white-box setting, where the model's structural information can be exploited. However, exploring the transferability of multimodal adversarial examples is pivotal, especially given the limited access to detailed model structures in real-world scenarios. There have been some efforts made to enhance the transferability of attacks on VLP models by introducing input diversity, as seen in the work SGA [36]. While they have achieved some effectiveness, how to enhance the transferability of multimodal adversarial examples is still an open question. + +In this paper, we undertake a comprehensive examination of the factors contributing to the limited transferability of the cutting-edge multimodal attack + +method, i.e., SGA [36]. As shown in Figure 1 (a), throughout the iterative generation of subsequent adversarial images, SGA conducts data augmentation around the online adversarial image (i.e., adversarial examples generated during optimization). This strategy enhances the diversity of adversarial examples along the optimization path, leading to a certain improvement in transferability. However, such an approach still carries the potential risk of overfitting to the victim model with the high reliance on the examples along the adversarial trajectory (See Figure 1 (b)), which leads to low attack success rates when we migrate the adversarial example to other VLP models (See Figure 1 (c) and (d)). To mitigate this overfitting risk, one potential solution is to further enhance the diversity of augmented adversarial examples in a judicious manner. + +Building upon the aforementioned analysis, SGA overfits local adversarial examples, while the clean image is the only accessible example that is far from local adversarial examples. Therefore, we embark on a pioneering endeavor to enhance the transferability of multimodal adversarial attacks by considering the diversity of adversarial examples (AEs) around clean inputs and online AEs throughout the optimization process. To achieve this, we consider the intersection region of adversarial trajectory, which encompasses the original image, the adversarial image from the previous step, and the current adversarial image during the iterative attack process (depicted in Figure 1 (b)). This innovative approach aims to circumvent overfitting by strategic sampling within this region, thereby avoiding an undue focus on adversarial example diversity solely around adversarial images. After obtaining multiple samples, we calculate gradients for each to determine perturbation directions away from the text. Subsequently, we individually incorporate these perturbations into the current adversarial image and select the one that deviates the most from the text. + +Additionally, in the text modality, SGA only considers deviating the text from the last adversarial image in the optimization period, but the adversarial image is solely generated by the surrogate model, still posing the risk of overfitting the surrogate model. For this reason, we propose to have the text deviate simultaneously from the last intersection region along the optimization path. + +Our proposed method is evaluated on two widely recognized multimodal datasets, Flickr30K [41] and MSCOCO [35]. We conduct experiments on three vision-and-language downstream tasks (i.e., image-text retrieval (ITR), visual grounding(VG), and image captioning (IC)), and all results indicate the high effectiveness of our method in generating more transferable multimodal adversarial examples. Moreover, when adversarial examples generated from image-text retrieval are transferred to other vision-and-language downstream tasks (i.e., VG and IC), there is a substantial improvement in attack performance. + +Our main contributions can be summarized as follows: + +- We extend the generation of adversarial text to deviate from the last intersection region along the optimization path, aiming to reduce overfitting the surrogate model, thereby achieving enhanced transferability. + +- Extensive experiments robustly demonstrate the efficacy of our proposed method in elevating the transferability of multimodal adversarial examples across diverse models and three downstream tasks. + +# 2 Related Work + +# 2.1 Vision-Language Pre-training Models + +VLP models leverage multimodal learning from extensive image-text pairs to improve the performance of various Vision-and-Language $(\mathrm{V} + \mathrm{L})$ tasks [29]. Early VLP models predominantly depend on pre-trained object detectors for acquiring multimodal representations [6, 32, 45, 49, 59]. Recently, with the advent of end-to-end image encoders like the Vision Transformer (ViT) [10, 46, 56] offering faster inference speeds, some work propose to use them as substitution for computationally expensive object detectors [11, 29, 30, 48, 53]. + +There are two popular approaches for VLP models in learning vision-language representations: the fused architecture and the aligned architecture. Fused VLP models (e.g., ALBEF [30], TCL [53]), initially employ two separate unimodal encoders to learn features for text and images. Subsequently, a multimodal encoder is utilized to fuse the embeddings of text and images. In contrast, aligned VLP models, exemplified by CLIP [42], focus on aligning the feature spaces of distinct unimodal encoders and benefit downstream tasks significantly [1]. This paper concentrates on evaluating our proposed method using multiple popular fused and aligned VLP models. + +# 2.2 Downstream Vision-and-Language Tasks + +Image-Text Retrieval (ITR) involves retrieving pertinent information, textual or visual, in response to queries from another modality [5, 8, 51, 60]. This undertaking usually encompasses two sub-tasks: image-to-text retrieval (retrieving text based on an image query) and text-to-image retrieval (retrieving images given a text query). + +In aligned VLP models, both the Text Retrieval (TR) and Image Retrieval (IR) tasks leverage ranking results determined by the similarity between text and image embeddings. However, in fused VLP models, where internal embedding spaces lack alignment across unimodal encoders, the similarity scores between image and text modalities are computed for all image-text pairs to retrieve the Top-N candidates. Subsequently, these Top-N candidates serve as input for the multimodal encoder, which computes the image-text matching score to establish the final ranking. + +Visual Grounding (VG) refers to the task of localizing the region within a visual scene with corresponding entities or concepts in natural language. Among + +Table 1: Attack Success Rate (%) of SGA with and without image augmentation. The SGA w.o. Aug doesn't consider image augmentation. We use ALBEF to generate multimodal adversarial examples on the ITR task to evaluate transferability. + +
SourceAttackALBEFTCLCLIPvITCLIPCNN
TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1
ALBEFSGA w.o. Aug99.999.9570.0771.6730.5539.8832.3142.54
SGA w. Aug99.999.9887.8888.0536.6946.7839.5949.78
+ +VLP models, ALBEF expands Grad-CAM [43] and utilizes the acquired attention map to rank the detected proposals [54]. + +Image Captioning (IC) is to generate a textual description that logically describes or implies the content of a given visual input, typically involving the creation of captions for images. The evaluation of Image Captioning models often employs metrics such as BLEU [39], METEOR [4], ROUGE [33], CIDEr [47] and SPICE [2], which serve to assess the quality and relevance of the generated captions in comparison to reference captions. + +# 2.3 Transferability of Adversarial Examples + +Adversarial attacks [13-15, 18, 25] are typically categorized as white-box and black-box attacks. In a white-box setting [23, 26], the attacker has full access to the model, whereas black-box attacks [3, 40], more realistic in practical applications, occur when information about the model is limited. In the realm of image attacks [21, 22, 24], prevalent methods for crafting transferable adversarial examples often leverage data augmentation techniques (e.g., DIM [52], TIM [9], SIM [34], ADMIX [50], PAM [58]). Zhang et al. [57] introduced a white-box attack targeting popular VLP models for downstream tasks in the multimodal domain. Building upon this work, Lu et al. [36] proposed SGA, considering the diversity of adversarial examples by expanding single image-text pairs to sets of images and texts to conduct black-box attacks on VLP models. + +However, SGA primarily emphasizes diversity in the vicinity of adversarial examples during the optimization process, potentially increasing the risk of overfitting the victim model and impacting transferability. Therefore, our primary focus in this study is to further enhance the diversity of adversarial examples in a thoughtful manner and avoid an undue focus on adversarial example diversity solely around adversarial images. + +# 3 Methodology + +# 3.1 Background and Motivation + +Adversarial attacks on VLP models involve inducing a mismatch between adversarial images and corresponding adversarial text while adhering to specified + +constraints on image and text perturbations. Here, $(v,t)$ denotes an original image-text pair from a multimodal dataset, with $v^{\prime}$ representing an adversarial image and $t^\prime$ denoting adversarial text. The allowable perturbations are restricted within the ranges $B[v,\xi_v]$ for images and $B[t,\xi_t]$ for text. The image and text encoders of the multimodal model are denoted as $F_{I}$ and $F_{T}$ , respectively. To generate valid multimodal adversarial examples, the objective is to maximize the loss function $J$ specific to VLP models: + +$$ +\left\{ \begin{array}{l} \max J \left(F _ {I} \left(v ^ {\prime}\right), F _ {T} \left(t ^ {\prime}\right)\right) \\ s. t. v ^ {\prime} \in B [ v, \xi_ {v} ], t ^ {\prime} \in B [ t, \xi_ {t} ]. \end{array} \right. \tag {1} +$$ + +The state-of-the-art approach for exploring the transferability of multimodal adversarial examples (i.e., SGA [36]) involves augmenting image-text pairs to enhance the diversity of adversarial examples along the optimization path. Specifically, during the iterative generation of adversarial images, let $v_{i}^{\prime}$ represent the adversarial image generated at the $i$ -th step. In the subsequent step $(i + 1)$ , SGA initiates the process by applying a resizing operation for data augmentation to $v_{i}^{\prime}$ , resulting in $V_{i}^{\prime} = \{v_{i1}^{\prime}, v_{i2}^{\prime}, \dots, v_{iM}^{\prime}\}$ (See Figure 1 (a)). The iterative formula can be expressed as follows: + +$$ +v _ {i + 1} ^ {\prime} = v _ {i} ^ {\prime} + \alpha \cdot \operatorname {s i g n} \left(\frac {\nabla_ {v} \sum_ {j = 1} ^ {M} J \left(F _ {I} \left(v _ {i j} ^ {\prime}\right) , F _ {T} (t)\right)}{\left\| \nabla_ {v} \sum_ {j = 1} ^ {M} J \left(F _ {I} \left(v _ {i j} ^ {\prime}\right) , F _ {T} (t)\right) \right\|}\right). \tag {2} +$$ + +To further examine the impact of image augmentation along the optimization path in the SGA method, we utilize ALBEF as a surrogate model to generate multimodal adversarial examples. These examples are then employed to target VLP models such as TCL and CLIP, assessing the transferability of the attacks. Detailed results are presented in Table 1. Our observations reveal that SGA enhances the transferability of adversarial attacks, showing an increase ranging from $6.14\%$ to $17.81\%$ . However, it is noteworthy that the success rate of attacks on the target models remains notably lower than that on the source model. This discrepancy is primarily attributed to the fact that SGA predominantly emphasizes diversity around AE $v_{i}^{\prime}$ during the optimization period, without adequately considering the diversity of adversarial examples toward the clean image, leading to the risk of overfitting the victim model and affecting the transferability. + +For this purpose, we propose to consider diversification along the intersection region of adversarial trajectory, which encompasses the original image $v$ , the adversarial image from the previous step $v_{i-1}'$ , and the current adversarial image $v_i'$ during the iterative attack process. This region is established to sample images within it to broaden the diversity of adversarial examples (See Figure 1 (b)). Moreover, to fully leverage the interplay between modalities, we aim for perturbations guided by textual information that induce $v_i'$ to deviate significantly from the associated text $t$ . Additionally, in the text modality, our objective is to identify adversarial perturbations that simultaneously deviate from the intersection region rather than only adversarial images, thereby reducing overfitting the surrogate model and enhancing the effectiveness of black-box attacks. + +# 3.2 Diversification along the Intersection Region + +As outlined in Section 3.1, we enhance the diversity of adversarial examples by introducing diversification along the intersection region. Specifically, at the $i$ th iteration during optimization, we have the $v_{i}^{\prime}$ , $v_{i-1}^{\prime}$ , and the clean $v$ , and these variables form a triangle region denoted as $\triangle vv_{i-1}^{\prime}v_{i}^{\prime}$ , i.e., the intersection region of adversarial trajectory in Figure 1. Then, we initially sample multiple instances within the region $\triangle vv_{i-1}^{\prime}v_{i}^{\prime}$ , representing the set of samples as $e = \{e_1,e_2,\dots,e_N\}$ . Each sample can be expressed as $e_k = \beta \cdot v + \gamma \cdot v_{i-1}^{\prime} + \eta \cdot v_i^{\prime}$ where $\beta + \gamma + \eta = 1.0$ . Consequently, we can compute the gradient perturbation for each sample. For the k-th sample, denoted as $e_k$ , its gradient perturbation $p_k$ is calculated as follows. In this way, we can get a perturbation set $P = \{p_1,p_2,\dots,p_N\}$ by + +$$ +p _ {k} = \alpha \cdot \operatorname {s i g n} \left(\frac {\nabla_ {e} J \left(F _ {I} \left(e _ {k}\right) , F _ {T} (t)\right)}{\| \nabla_ {e} J \left(F _ {I} \left(e _ {k}\right) , F _ {T} (t)\right) \|}\right). \tag {3} +$$ + +# 3.3 Text-guided Augmentation Selection + +In Section 3.2, a diverse perturbation set $P$ is derived from the intersection region. To harness the full potential of modality interactions, we introduce text-guided augmentation selection to obtain the optimal sample. Specifically, we individually incorporate each element from the perturbation set $P$ into the adversarial image $v_{i}^{\prime}$ . The selection process aims to identify the sample that maximally distances $v_{i}^{\prime}$ from $t$ . This procedure can be represented as: + +$$ +m = \underset {p _ {m} \in P} {\arg \max } J \left(F _ {I} \left(v _ {i} ^ {\prime} + p _ {m}\right), F _ {T} (t)\right). \tag {4} +$$ + +At this juncture, $e_m$ represents the selected sample. We employ SGA as our baseline and incorporate the image augmentation methods considered along its optimization path. The chosen optimal sample $e_m$ is resized and expanded into the set $E_m = \{e_{m1}, e_{m2}, \dots, e_{mM}\}$ . Subsequently, we utilize the expanded set $E_m$ to generate the final adversarial perturbation, yielding $v_{i+1}'$ : + +$$ +v _ {i + 1} ^ {\prime} = v _ {i} ^ {\prime} + \alpha \cdot \operatorname {s i g n} \left(\frac {\nabla_ {e} \sum_ {j = 1} ^ {M} J \left(F _ {I} \left(e _ {m j}\right) , F _ {T} (t)\right)}{\| \nabla_ {e} \sum_ {j = 1} ^ {M} J \left(F _ {I} \left(e _ {m j}, F _ {T} (t)\right) \right\|}\right). \tag {5} +$$ + +# 3.4 Adversarial Text deviating from the Intersection Region + +In the text modality, SGA only considers deviating adversarial text from the ultimate adversarial image generated during the iterative optimization process. If there are a total of T iterations, The generation of $t'$ only considers deviating from the adversarial image $v_T'$ . The adversarial image $v_T'$ is exclusively created by the surrogate model, thereby still presenting the risk of overfitting to the surrogate model. However, during the optimization process of the adversarial image, the clean image is entirely independent of the surrogate model. For this + +reason, we propose to have the text deviate simultaneously from the last intersection region along the optimization path. Specifically, the adversarial text deviates from the triangle region constituted by $v$ , $v_{T-1}'$ and $v_T'$ . + +$$ +t ^ {\prime} = \underset {t ^ {\prime} \in B [ t, \epsilon_ {t} ]} {\arg \max } (\lambda \cdot J (F _ {I} (v), F _ {T} (t ^ {\prime})) + \mu \cdot J (F _ {I} (v _ {T} ^ {\prime}), F _ {T} (t ^ {\prime}))) + \nu \cdot J (F _ {I} (v _ {T - 1} ^ {\prime}), F _ {T} (t ^ {\prime}))). \tag {6} +$$ + +We also set adjustable scaling factors, among which $\lambda +\mu +\nu = 1.0$ + +# 3.5 Implementation Details + +In the specific process of our attack, we employ an iterative approach. In each iteration, we sample within the intersection region of adversarial trajectory, guided by textual information, to select a sample that maximally deviates the current adversarial image $v_{i}^{\prime}$ from the text $t$ . Subsequently, we subject this sample to image augmentation processing, calculate gradients to determine the perturbation direction, and overlay it onto the current adversarial image, resulting in $v_{i + 1}^{\prime}$ . Through multiple iterative steps, we obtain an adversarial image $v_{T}^{\prime}$ . For the text modality, in contrast to previous methods that solely focus on deviating from the adversarial image $v_{T}^{\prime}$ , our goal is to derive an adversarial text $t^{\prime}$ that simultaneously deviates from the last intersection region $\triangle vv_{T - 1}^{\prime}v_{T}^{\prime}$ along the optimization path. + +# 4 Experiments + +In this section, we present experimental evidence demonstrating the enhanced transferability of multimodal examples generated from our proposed method across VLP models and various Vision-and-Language tasks. First, in Section 4.1, we introduce the experimental settings, including the popular image-text pair datasets and VLP models we use, as well as restrictions for adversarial attacks. Subsequently, the process of searching for optimal parameters is shown in Section 4.2. After that, we evaluate cross-model transferability in the context of the image-text retrieval task, as detailed in Section 4.3. Following this, in Section 4.4, we extend our investigation to the transfer of multimodal adversarial examples generated within the image-text retrieval task to other tasks, aiming to gauge cross-task transferability. Lastly, Section 4.5 outlines ablation studies. + +# 4.1 Setups + +VLP Models. In our transferability evaluation experiments across various VLP models, we explore two typical architectures: fused and aligned VLP models. We select CLIP [42] for the aligned VLP model. CLIP offers a choice between two distinct image encoders, namely ViT-B/16 [10] and ResNet-101 [19], denoted as $\mathrm{CLIP}_{\mathrm{ViT}}$ and $\mathrm{CLIP}_{\mathrm{CNN}}$ , respectively. In the case of fused VLP models, we opt for ALBEF [30] and TCL [53]. ALBEF uses a 12-layer ViT-B/16 image encoder + +and two 6-layer transformers for text and multimodal encoding. TCL shares this architecture but has different pre-training objectives. + +Datasets. In this study, we leverage two widely recognized multimodal datasets, namely Flickr30K [41] and MSCOCO [35], for evaluating the image-text retrieval task. The Flickr30K dataset comprises 31,783 images, each accompanied by five captions for annotation. Similarly, the MSCOCO dataset consists of 123,287 images, and approximately five captions are provided for each image. + +Additionally, we employ the RefCOCO+ [55] dataset to assess the Visual Grounding task. RefCOCO+ is a dataset containing 141,564 referring expressions for 50,000 objects within 19,992 MSCOCO images. This dataset serves the purpose of evaluating grounding models by focusing on the localization of objects described through natural language. For another Vision-and-Language task, Image Captioning, we leverage the MSCOCO dataset as well. + +Adversarial Attack Settings. In our study, we adopt adversarial attack settings of SGA [36] to ensure a fair comparison. Specifically, we leverage BERTAttack [31] to craft adversarial texts. The perturbation bound $\xi_{t}$ is set as 1 and length of word list $W = 10$ . PGD [38] is employed to get adversarial images and the perturbation bound, denoted as $\xi_v$ , is set as 8/255. Additionally, iteration steps $T$ is set as 10 and each step size $\alpha = 2 / 255$ . Furthermore, when randomly sampling from the intersection region of adversarial trajectory, we set the number of samples to 5. When generating adversarial texts, we set the three parameters $\lambda, \mu, \nu$ in Equation 6 to 0.6, 0.2, and 0.2 respectively. The values chosen for these adjustable parameters can be found in Section 4.2. + +Evaluation Metrics. The key metric for adversarial transferability is the Attack Success Rate (ASR), which measures the percentage of successful attacks among all generated adversarial examples. A higher ASR indicates more effective and transferable attacks. + +# 4.2 Optimal Parameters + +In our proposed method, the number of samples $N$ taken from the intersection region of adversarial trajectory and scaling factors in Equation 6 are adjustable. We conduct specific experiments to explore optimal parameter settings and examine their influence on the efficacy of our approach. To be more specific, we utilize ALBEF for generating multimodal adversarial examples on the Flickr30K dataset and assess the transferability on the other three VLP models. + +Number of Samples taken from Intersection Region of Adversarial trajectory $N$ . The bottom of Table 2 illustrates when the sample size reaches 5, transfer effects are observed for both the Image Retrieval task and the Text Retrieval task. As the sample size continues to increase, the transferability only fluctuates and does not exhibit further improvement. Taking into account both transfer effects and computational costs, a sample size of 5 is identified as the optimal configuration. + +Scaling factors $\lambda, \mu, \nu$ in Adversarial Text Generation. In Equation 6, $\lambda$ represents the weight of clean images, while $\mu$ and $\nu$ represent the weights of adversarial images. Therefore, we stipulate that $\lambda$ cannot be zero, and $\mu$ and $\nu$ + +Table 2: Optimal Parameters: Attack Success Rate(%) on different settings, Top for different values of $\lambda ,\mu ,\nu$ and Bottom for different numbers of samples $N$ + +
SourceAttackALBEFTCLCLIPvITCLIPCNN
TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1
ALBEF[λ,μ,v] = [0.2,0.0,0.8]99.999.9389.6790.542.2152.046.2355.44
[λ,μ,v] = [0.2,0.2,0.6]99.999.9390.4190.4341.9651.8746.4955.3
[λ,μ,v] = [0.2,0.4,0.4]99.999.9390.3190.4341.9651.8745.3454.82
[λ,μ,v] = [0.2,0.6,0.2]99.999.9590.5290.5742.2151.8445.0854.92
[λ,μ,v] = [0.2,0.8,0.0]99.999.9390.3190.5741.7251.7446.154.68
[λ,μ,v] = [0.4,0.0,0.6]99.999.9391.2590.8845.5255.3248.9157.63
[λ,μ,v] = [0.4,0.2,0.4]99.999.9391.2590.8345.0355.2248.2857.77
[λ,μ,v] = [0.4,0.4,0.2]99.999.9391.1590.7144.9155.2848.1557.87
[λ,μ,v] = [0.4,0.6,0.0]99.999.9391.1590.8844.9154.7749.0457.56
[λ,μ,v] = [0.6,0.0,0.4]99.999.9391.4690.9546.3856.3849.0459.11
[λ,μ,v] = [0.6,0.2,0.2]99.999.9391.5791.1746.2656.849.5559.01
[λ,μ,v] = [0.6,0.4,0.0]99.999.9390.9490.9846.0156.7249.5558.87
[λ,μ,v] = [0.8,0.0,0.2]99.999.9390.7390.9846.1356.7149.4658.74
[λ,μ,v] = [0.8,0.2,0.0]99.999.9390.3190.9545.7756.6549.3458.87
ALBEFN=399.999.9590.6290.7945.6456.5448.8358.73
N=499.7999.9191.3691.1745.8356.7850.4559.01
N=599.999.9391.5791.1746.2656.849.5559.01
N=699.999.9390.9490.3845.7956.9650.758.52
N=799.999.9189.8890.9545.456.3550.759.07
+ +can have at most one zero value. Initially, as the value of $\lambda$ gradually increases, indicating the gradual introduction of clean images, the transferability of multimodal adversarial examples increases accordingly. However, when the value of $\lambda$ becomes too large, it also leads to a disproportionately low proportion of adversarial images, resulting in a decrease in adversarial transferability. According to the experiments, the optimal parameters we select are [0.6, 0.2, 0.2]. + +# 4.3 Cross-Model Transferability + +As outlined in Section 4.1, our experimental design focuses on assessing the transferability of adversarial examples across two widely adopted VLP model architectures: fused and aligned. There are four VLP models selected, namely: ALBEF, TCL, $\mathrm{CLIP}_{\mathrm{ViT}}$ , $\mathrm{CLIP}_{\mathrm{CNN}}$ . The selected downstream $\mathrm{V} + \mathrm{L}$ task for evaluation is image-text retrieval. Our approach involves employing one of four models to generate multimodal adversarial examples within the specified parameters of our proposed method. Subsequently, we validate the effectiveness of the generated adversarial examples through a comprehensive set of experiments, encompassing both self-attacks and attacks on the other three models. This evaluation encompasses one white-box attack and three distinct black-box attacks. + +We adopt the methodology proposed by SGA [36] as our baseline, Therefore, the effectiveness of our method is compared against it. Moreover, we also present the effectiveness and transferability results of various other attack methods on VLP models. Table 3 provides a comprehensive comparison of these methods on the dataset Flickr30K, More experiments on the MSCOCO dataset are provided in the Appendix. In the comparison, PGD [38] is an image-only attack, while Bert-Attack [31] focuses solely on text-based attacks. Sep-Attack involves perturbing text and image separately. Furthermore, Co-Attack takes into account cross-modal interactions, generating an adversarial example for one modality + +![](images/fa112191c91e91d2277120ecde637d54a2308149461c8aeb4606ed354312286e.jpg) +Cleannnnnne +a man riding on the back of a motorcycle down a street + +![](images/b7ca1681e86010cf84c61bc96873c5e2ac7db0d058d23c0f6eb5c61370b5a36f.jpg) +a man wearing a black shirt and a green tie + +![](images/d759c7dc32f0c0e468c00a5596e8b6f92738d77d463285e551d7e74b8348f12e.jpg) +a parking meter sitting on the side of a road + +![](images/20ec060bcb801515f1977f1ffba3f0ae7943db9ad671f0e3a1bce39029325548.jpg) +a cat standing on a toilet seat in a bathroom + +![](images/7306439b9c64d067525644f744d093202dc47c844c13cda18ce367a99fcb14d3.jpg) +am an riding a bike next to a train + +![](images/f8244d470fc556955750045aa8029aa5e6317db3fe6466b5ec4fe42bd51bff50.jpg) +a man sitting on a bench next to a white dog + +![](images/ec9796b597d6cea774128ee16fa8c3987566cdc05c892c51f399dbd46011bc52.jpg) +a siamese cat eating a banana from a person's hand + +![](images/6a87880aeea6b07ebd99c6a6394751353bc2c97486a063c3b55c7ef4717f70b9.jpg) +Caaee +Adversarial +Caaee +a woman sitting on a motorcycle in the street +Fig. 2: Visualization on Image Captioning. We use the ALBEF model, pre-trained on Image Text Retrieval(ITR) task, to generate adversarial images on the MSCOCO dataset and use the BLIP [29] model for Image Captioning on both clean images and adversarial images, respectively. + +![](images/3d2fdd59f275d91b97284a1af2a57e302901040857a8043d9fc9415d4c50bb93.jpg) +a young girl wearing glasses and a tie + +![](images/1bd6db1a76ba15eaa41e39e3c73392fe25888edf259b42e1eac659314dcd1c64.jpg) +a group of people riding motorcycles down a street + +![](images/94ab0d8f2de445360357fa605bc658cca27c1d0d3c6a926a6dc2612bdac7d41b.jpg) +a black and white dog laying on top of a toilet + +![](images/2907a68e8f28b9598049757ab815de30dda8d019a8beace6e512755557aabd53.jpg) +a train on a train track with a dog standing next to it + +![](images/3ddd831a1f68db9f81c9b47b7a79996bd2b973fab15e046a381c94e80f2f1162.jpg) +a boy riding a bike next to a white dog + +![](images/5c58d90dfdffac9a457e7743617d5af84eec7c9581ab90d1c023a6876fbfb8a1.jpg) +a small dog is eating a banana on the floor + +under the guidance of the other modality. SGA, our baseline, expands a single image-text pair into a set of images and a set of texts to enhance diversity. + +First, we compare the performance of various methods under white-box attacks, wherein we can leverage the model architecture and interact with the model. It is evident that our method, along with SGA, performs the best in the four white-box attack experiments compared to other methods. Whether TR or IR, the attack success rate at the top-1 rank $(\mathbb{R}@\mathbb{1})$ consistently exceeds $99.9\%$ . Given the already high white-box attack success rate achieved by the SGA method, there is limited room for improvement in our approach within this context. Subsequently, we shift our focus to elucidating the enhancements our method brings to transferability, specifically in the realm of black-box attack performance. We delineate this exploration into two segments, contingent on whether the source model and the target model share the same architecture. + +Cross-Model Transferability in Same Architecture. ALBEF and TCL both utilize a similar model architecture but differ in their pre-training objectives while maintaining a common fundamental model structure. Therefore, when ALBEF and TCL are employed as target models for each other, the success rate of attacks using multimodal adversarial examples is remarkably high. Notably, the black-box attack success rate of IR reaches $95.58\%$ when TCL is the target model in our proposed method. Additionally, the transferability of SGA can reach approximately $90\%$ , but the room for improvement is very limited, our method has improved compared to SGA, ranging from $2.71\%$ to $3.69\%$ . In contrast, when considering $\mathrm{CLIP}_{\mathrm{ViT}}$ and $\mathrm{CLIP}_{\mathrm{CNN}}$ —both being aligned VLP models—their image encoders vary significantly, with one utilizing the Vision Transformer and the other employing ResNet-101. Given the significant structural differences between traditional CNNs and Transformers, existing methods show low cross-model transferability, highlighting room for improvement. Our method outperforms SGA by about $7.67\%$ to $11.62\%$ . + +Table 3: Comparison with state-of-the-art methods on image-text retrieval. The source column shows VLP models we use to generate multimodal adversarial examples. The gray area represents adversarial attacks under a white-box setting, the rest are black-box attacks. For both Image Retrieval and Text Retrieval, we provide R@1 attack success rate(%) + +
SourceAttackALBEFTCLCLIP_VITCLIP_CNN
TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1
ALBEFPGD93.7494.4324.0327.910.6715.8214.0519.11
BERT-Attack11.5727.4612.6428.0729.3343.1732.6946.11
Sep-Attack95.7296.1439.351.7934.1145.7235.7647.92
Co-Attack97.0898.3639.5251.2429.8238.9231.2941.99
SGA99.999.9887.8888.0536.6946.7839.5949.78
Ours99.999.9391.5791.1746.2656.849.5559.01
TCLPGD35.7741.6799.3799.3310.1816.314.8121.1
BERT-Attack11.8926.8214.5429.1729.6944.4933.4646.07
Sep-Attack52.4561.4499.5899.4537.0645.8137.4249.91
Co-Attack49.8460.3691.6895.4832.6442.6932.0647.82
SGA92.4992.77100.0100.036.8146.9741.8951.53
Ours95.295.58100.099.9847.2457.2852.2362.23
CLIP_VITPGD3.136.484.438.8369.3384.7913.0317.43
BERT-Attack9.5922.6411.8025.0728.3439.0830.4037.43
Sep-Attack7.6120.5810.1220.7476.9387.4429.8938.32
Co-Attack8.5520.1810.0121.2978.5387.529.538.49
SGA22.4234.5925.0836.45100.0100.053.2661.1
Ours27.8442.8427.8244.6100.0100.064.8869.5
CLIP_CNNPGD2.296.154.538.885.412.0889.7893.04
BERT-Attack8.8623.2712.3325.4827.1237.4430.4040.10
Sep-Attack9.3822.9911.2825.4526.1339.2493.6195.3
Co-Attack10.5323.6212.5426.0527.2440.6295.9196.5
SGA15.6428.618.0233.0739.0251.4599.8799.9
Ours19.534.5921.637.8848.4759.1299.8799.9
+ +Cross-Model Transferability in Different Architectures. When fused VLP models (i.e., ALBEF and TCL) are used for generating adversarial examples, our method achieves a significant improvement compared to existing methods, outperforming the current state-of-the-art (SOTA) methods by $9.23\%$ to $10.7\%$ . When the fused VLP models are targeted for attack, our method still outperforms all existing methods when generating adversarial examples against the aligned VLP models (i.e., CLIP). In this scenario, the transferability of all methods is relatively low. Compared to our method's baseline SGA, we attain an improvement ranging from $2.74\%$ to $8.25\%$ . + +# 4.4 Cross-Task Transferability + +We not only assess the transferability of multimodal adversarial examples generated by our proposed method across different VLP models but also conduct experiments to evaluate its effectiveness in transferring across diverse $\mathrm{V} + \mathrm{L}$ tasks. Specifically, we craft adversarial examples for the Image-Text Retrieval (ITR) task and evaluate them on Visual Grounding (VG) and Image Captioning (IC) tasks. As evident from Table 4 and visual results in Figure 2 and 3, the adversarial examples generated for ITR demonstrate transferability, successfully impacting both VG and IC tasks. This highlights the efficacy of cross-task trans + +![](images/dd50538e90d8f02ac790684bf3ff72dc8c7f817cc7fca1353f6606f5da6073d9.jpg) +Fig. 3: Visualization on Visual Grounding. We use the ALBEF model, pre-trained on the ITR task, to generate adversarial images on the RefCOCO+ dataset and use the same model, pre-trained on Visual Grouding(VG) task, to localize the regions corresponding to red words on both clean images and adversarial images, respectively. + +ferability in our proposed method. Furthermore, our transferability consistently outperforms that of SGA. + +# 4.5 Ablation Study + +Our proposed method builds upon the SGA [36] as a baseline, introducing two key improvements. Firstly, we utilize diversification along the intersection region of adversarial trajectory to expand the diversity of adversarial examples. Secondly, we generate adversarial text while simultaneously distancing it from the last intersection region along the optimization path. To investigate the impact of each improvement on the effectiveness of our method, we conduct ablation studies on the ITR task and employ transferability from ALBEF to the other three VLP models as the evaluation metric. + +In the ablation study, we systematically eliminate each enhancement from our approach and compare their transferability with both the baseline SGA and our final method, which incorporates a combination of two improvements. The results are depicted in Figure 4. It is evident that when attacking models with the same architecture versus different architectures, the impacts of two distinct improvements are also different. When attacking TCL with the same structure as ALBEF, under Setting1, our method exhibits a greater decrease in transferability, indicating that diversification along the intersection region is + +Table 4: Cross-Task Transferability. We utilize ALBEF to generate multimodal adversarial examples for attacking both Visual Grounding(VG) on the RefCOCO+ dataset and Image Captioning(IC) on the MSCOCO dataset. The baseline represents the performance of each task without any attack, where a lower value indicates better effectiveness of the adversarial attack for both tasks. + +
AttackITR → VGITR → IC
ValTestATestBB@4METEORROUGE-LCIDErSPICE
Baseline58.4665.8946.2539.731.060.0133.323.8
SGA50.5657.4240.6628.024.651.291.417.7
Ours49.7056.3240.5427.224.250.788.317.2
+ +![](images/8268410413513309c42bce47d08364b6f9b93b3b4de585e0fb11a9230cc8d25c.jpg) +Fig. 4: Ablation Study: Attack Success Rate(%) on other three target models. The baseline is SGA. Setting 1 removes diversification along the intersection region of adversarial trajectory. Setting 2 removes the text deviating from the last intersection region along the optimization path. + +![](images/ba2b94f80a5264bb12ca3d19cac0ff96a4d84d6598e8a24ad2156afe5adc77aa.jpg) + +![](images/3d1524d5e2bdd08b7b423cf627dcc13674fa14fc8a8f732e289db634576f69b9.jpg) + +most crucial in our approach under this scenario. However, under Setting2, the decrease in transferability is more pronounced when attacking the aligned VLP model, indicating that at this point, adversarial texts deviating from the last intersection region play a larger role in the effectiveness of our method. + +# 5 Conclusion + +In this paper, we conduct a systematic evaluation of existing multimodal attacks regarding transferability. We found that these methods predominantly prioritize diversity around adversarial examples (AEs) during the optimization process, potentially leading to overfitting to the victim model and hindering transferability. To address this issue, we propose diversification along the intersection region of adversarial trajectory to broaden diversity not only around AEs but also towards clean inputs. Moreover, we pioneer an exploration into extending adversarial texts deviating from the intersection region. Through extensive experiments, we demonstrate the effectiveness of our method in enhancing transferability across VLP models and $\mathrm{V} + \mathrm{L}$ tasks. This work could act as a catalyst for more profound research on the transferability of multimodal AEs, alongside fortifying the adversarial robustness of VLP models. + +# Acknowledgments + +This research is supported by the National Research Foundation, Singapore, and DSO National Laboratories under the AI Singapore Programme (AISG Award No: AISG2-GC-2023-008), Career Development Fund (CDF) of Agency for Science, Technology and Research (A*STAR) (No.: C233312028), and National Research Foundation, Singapore and Infocomm Media Development Authority under its Trust Tech Funding Initiative (No. DTC-RGC-04). + +# References + +1. Abdelfattah, R., Guo, Q., Li, X., Wang, X., Wang, S.: Cdul: Clip-driven unsupervised learning for multi-label image classification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1348-1357 (2023) +2. Anderson, P., Fernando, B., Johnson, M., Gould, S.: Spice: Semantic propositional image caption evaluation. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 382-398. Springer (2016) +3. Bai, Y., Zeng, Y., Jiang, Y., Wang, Y., Xia, S.T., Guo, W.: Improving query efficiency of black-box adversarial attack. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16. pp. 101-116. Springer (2020) +4. Banerjee, S., Lavie, A.: Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In: Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization. pp. 65-72 (2005) +5. Chen, H., Ding, G., Liu, X., Lin, Z., Liu, J., Han, J.: Imram: Iterative matching with recurrent attention memory for cross-modal image-text retrieval. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12655-12663 (2020) +6. Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: Uniter: Universal image-text representation learning. In: European conference on computer vision. pp. 104-120. Springer (2020) +7. Cheng, H., Xiao, E., Cao, J., Yang, L., Xu, K., Gu, J., Xu, R.: Typography leads semantic diversifying: Amplifying adversarial transferability across multimodal large language models. arXiv preprint arXiv:2405.20090 (2024) +8. Cheng, M., Sun, Y., Wang, L., Zhu, X., Yao, K., Chen, J., Song, G., Han, J., Liu, J., Ding, E., et al.: Vista: vision and scene text aggregation for cross-modal retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5184-5193 (2022) +9. Dong, Y., Pang, T., Su, H., Zhu, J.: Evading defenses to transferable adversarial examples by translation-invariant attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4312-4321 (2019) +10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arxiv 2020. arXiv preprint arXiv:2010.11929 (2010) + +11. Dou, Z.Y., Xu, Y., Gan, Z., Wang, J., Wang, S., Wang, L., Zhu, C., Zhang, P., Yuan, L., Peng, N., et al.: An empirical study of training end-to-end vision-and-language transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18166-18176 (2022) +12. Gao, K., Bai, Y., Bai, J., Yang, Y., Xia, S.T.: Adversarial robustness for visual grounding of multimodal large language models. arXiv preprint arXiv:2405.09981 (2024) +13. Gu, J., Jia, X., de Jorge, P., Yu, W., Liu, X., Ma, A., Xun, Y., Hu, A., Khakzar, A., Li, Z., et al.: A survey on transferability of adversarial examples across deep neural networks. arXiv preprint arXiv:2310.17626 (2023) +14. Guo, Q., Cheng, Z., Juefei-Xu, F., Ma, L., Xie, X., Liu, Y., Zhao, J.: Learning to adversarially blur visual object tracking. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10839-10848 (2021) +15. Guo, Q., Juefei-Xu, F., Xie, X., Ma, L., Wang, J., Yu, B., Feng, W., Liu, Y.: Watch out! motion is blurring the vision of your deep neural networks. In: Advances in Neural Information Processing Systems. vol. 33, pp. 975-985 (2020) +16. Han, D., Jia, X., Bai, Y., Gu, J., Liu, Y., Cao, X.: Ot-attack: Enhancing adversarial transferability of vision-language models via optimal transport optimization. arXiv preprint arXiv:2312.04403 (2023) +17. He, B., Jia, X., Liang, S., Lou, T., Liu, Y., Cao, X.: Sa-attack: Improving adversarial transferability of vision-language pre-training models via self-augmentation. arXiv preprint arXiv:2312.04913 (2023) +18. He, B., Liu, J., Li, Y., Liang, S., Li, J., Jia, X., Cao, X.: Generating transferable 3d adversarial point cloud via random perturbation factorization. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 764-772 (2023) +19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2016) +20. Hu, X., Gan, Z., Wang, J., Yang, Z., Liu, Z., Lu, Y., Wang, L.: Scaling up vision-language pre-training for image captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17980-17989 (June 2022) +21. Huang, Y., Guo, Q., Juefei-Xu, F., Ma, L., Miao, W., Liu, Y., Pu, G.: Advfilter: predictive perturbation-aware filtering against adversarial attack via multi-domain learning. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 395-403 (2021) +22. Huang, Y., Juefei-Xu, F., Guo, Q., Zhang, J., Wu, Y., Hu, M., Li, T., Pu, G., Liu, Y.: Personalization as a shortcut for few-shot backdoor attack against text-to-image diffusion models. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 21169-21178 (2024) +23. Huang, Y., Sun, L., Guo, Q., Juefei-Xu, F., Zhu, J., Feng, J., Liu, Y., Pu, G.: Ala: Naturalness-aware adversarial lightness attack. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 2418-2426 (2023) +24. Huang, Y., Sun, L., Guo, Q., Juefei-Xu, F., Zhu, J., Feng, J., Liu, Y., Pu, G.: Ala: Naturalness-aware adversarial lightness attack. In: Proceedings of the 31st ACM International Conference on Multimedia. p. 2418-2426 (2023) +25. Jia, X., Wei, X., Cao, X., Han, X.: Adv-watermark: A novel watermark perturbation for adversarial examples. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 1579-1587 (2020) + +26. Jia, X., Zhang, Y., Wei, X., Wu, B., Ma, K., Wang, J., Cao Sr, X.: Improving fast adversarial training with prior-guided knowledge. arXiv preprint arXiv:2304.00202 (2023) +27. Khan, Z., Fu, Y.: Exploiting bert for multimodal target sentiment classification through input space translation. In: Proceedings of the 29th ACM international conference on multimedia. pp. 3034-3042 (2021) +28. Lei, C., Luo, S., Liu, Y., He, W., Wang, J., Wang, G., Tang, H., Miao, C., Li, H.: Understanding chinese video and language via contrastive multimodal pre-training. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 2567-2576 (2021) +29. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022) +30. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems 34, 9694-9705 (2021) +31. Li, L., Ma, R., Guo, Q., Xue, X., Qiu, X.: Bert-attack: Adversarial attack against bert using bert (2020) +32. Li, X., Yin, X., Li, C., Zhang, P., Hu, X., Zhang, L., Wang, L., Hu, H., Dong, L., Wei, F., et al.: Oscar: Object-semantics aligned pre-training for vision-language tasks. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16. pp. 121-137. Springer (2020) +33. Lin, C.Y.: Rouge: A package for automatic evaluation of summaries. In: Text summarization branches out. pp. 74-81 (2004) +34. Lin, J., Song, C., He, K., Wang, L., Hopcroft, J.E.: Nesterov accelerated gradient and scale invariance for adversarial attacks. arXiv preprint arXiv:1908.06281 (2019) +35. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +36. Lu, D., Wang, Z., Wang, T., Guan, W., Gao, H., Zheng, F.: Set-level guidance attack: Boosting adversarial transferability of vision-language pre-training models (2023) +37. Luo, H., Gu, J., Liu, F., Torr, P.: An image is worth 1000 lies: Adversarial transferability across prompts on vision-language models. arXiv preprint arXiv:2403.09766 (2024) +38. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2017) +39. Naseer, M.M., Ranasinghe, K., Khan, S.H., Hayat, M., Shahbaz Khan, F., Yang, M.H.: Intriguing properties of vision transformers. Advances in Neural Information Processing Systems 34, 23296-23308 (2021) +40. Park, J., Miller, P., McLaughlin, N.: Hard-label based small query black-box adversarial attack. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3986-3995 (2024) +41. Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In: Proceedings of the IEEE international conference on computer vision. pp. 2641-2649 (2015) +42. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from + +natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +43. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017) +44. Shi, L., Shuang, K., Geng, S., Gao, P., Fu, Z., de Melo, G., Chen, Y., Su, S.: Dense contrastive visual-linguistic pretraining. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 5203-5212 (2021) +45. Tan, H., Bansal, M.: Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490 (2019) +46. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: International conference on machine learning. pp. 10347-10357. PMLR (2021) +47. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4566-4575 (2015) +48. Wang, T., Ge, Y., Zheng, F., Cheng, R., Shan, Y., Qie, X., Luo, P.: Accelerating vision-language pretraining with free language modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23161-23170 (2023) +49. Wang, T., Jiang, W., Lu, Z., Zheng, F., Cheng, R., Yin, C., Luo, P.: Vlmixer: Unpaired vision-language pre-training via cross-modal cutmix. In: International Conference on Machine Learning. pp. 22680-22690. PMLR (2022) +50. Wang, X., He, X., Wang, J., He, K.: Admix: Enhancing the transferability of adversarial attacks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16158-16167 (2021) +51. Wang, Z., Liu, X., Li, H., Sheng, L., Yan, J., Wang, X., Shao, J.: Camp: Cross-modal adaptive message passing for text-image retrieval. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5764-5773 (2019) +52. Xie, C., Zhang, Z., Zhou, Y., Bai, S., Wang, J., Ren, Z., Yuille, A.L.: Improving transferability of adversarial examples with input diversity. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2730-2739 (2019) +53. Yang, J., Duan, J., Tran, S., Xu, Y., Chanda, S., Chen, L., Zeng, B., Chilimbi, T., Huang, J.: Vision-language pre-training with triple contrastive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15671-15680 (2022) +54. Yu, L., Lin, Z., Shen, X., Yang, J., Lu, X., Bansal, M., Berg, T.L.: Mattnet: Modular attention network for referring expression comprehension. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1307-1315 (2018) +55. Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in referring expressions. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 69-85. Springer (2016) +56. Yuan, L., Chen, Y., Wang, T., Yu, W., Shi, Y., Jiang, Z.H., Tay, F.E., Feng, J., Yan, S.: Tokens-to-token vit: Training vision transformers from scratch on imagenet. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 558-567 (2021) + +57. Zhang, J., Yi, Q., Sang, J.: Towards adversarial attack on vision-language pretraining models. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 5005-5013 (2022) +58. Zhang, J., Huang, J.t., Wang, W., Li, Y., Wu, W., Wang, X., Su, Y., Lyu, M.R.: Improving the transferability of adversarial samples by path-augmented method. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8173-8182 (2023) +59. Zhang, P., Li, X., Hu, X., Yang, J., Zhang, L., Wang, L., Choi, Y., Gao, J.: Vinyl: Revisiting visual representations in vision-language models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5579-5588 (2021) +60. Zhang, Q., Lei, Z., Zhang, Z., Li, S.Z.: Context-aware attention network for image-text retrieval. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3536-3545 (2020) \ No newline at end of file diff --git a/2024/Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory/images.zip b/2024/Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..98522c3a4167f3e054ce250f1341b14eb63cca50 --- /dev/null +++ b/2024/Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9aef47cc1ec940f0c6830a3a36579d6c70e2ff9f28eb4555d010aad329ff48ca +size 484380 diff --git a/2024/Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory/layout.json b/2024/Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2a5ab34ee7d636e0fc9e454cf84cd49615b6ea99 --- /dev/null +++ b/2024/Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory/layout.json @@ -0,0 +1,10829 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 111, + 478, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 111, + 478, + 166 + ], + "spans": [ + { + "bbox": [ + 136, + 111, + 478, + 166 + ], + "type": "text", + "content": "Boosting Transferability in Vision-Language Attacks via Diversification along the Intersection Region of Adversarial Trajectory" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 196, + 185, + 416, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 185, + 416, + 210 + ], + "spans": [ + { + "bbox": [ + 196, + 185, + 416, + 210 + ], + "type": "text", + "content": "Sensen Gao " + }, + { + "bbox": [ + 196, + 185, + 416, + 210 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 196, + 185, + 416, + 210 + ], + "type": "text", + "content": ", Xiaojun Jia " + }, + { + "bbox": [ + 196, + 185, + 416, + 210 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 196, + 185, + 416, + 210 + ], + "type": "text", + "content": ", Xuhong Ren " + }, + { + "bbox": [ + 196, + 185, + 416, + 210 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 196, + 185, + 416, + 210 + ], + "type": "text", + "content": ", Ivor Tsang " + }, + { + "bbox": [ + 196, + 185, + 416, + 210 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 196, + 185, + 416, + 210 + ], + "type": "text", + "content": ", and Qing Guo " + }, + { + "bbox": [ + 196, + 185, + 416, + 210 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 211, + 219, + 381, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 219, + 381, + 230 + ], + "spans": [ + { + "bbox": [ + 211, + 219, + 381, + 230 + ], + "type": "text", + "content": "1 Nankai University, Tianjin, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 211, + 230, + 403, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 230, + 403, + 242 + ], + "spans": [ + { + "bbox": [ + 211, + 230, + 403, + 242 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 211, + 230, + 403, + 242 + ], + "type": "text", + "content": " Nanyang Technological University, Singapore" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 147, + 242, + 465, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 242, + 465, + 264 + ], + "spans": [ + { + "bbox": [ + 147, + 242, + 465, + 264 + ], + "type": "text", + "content": "3 CFAR and IHPC, Agency for Science, Technology and Research (A*STAR), Singapore" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 287, + 455, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 287, + 455, + 552 + ], + "spans": [ + { + "bbox": [ + 159, + 287, + 455, + 552 + ], + "type": "text", + "content": "Abstract. Vision-language pre-training (VLP) models exhibit remarkable capabilities in comprehending both images and text, yet they remain susceptible to multimodal adversarial examples (AEs). Strengthening attacks and uncovering vulnerabilities, especially common issues in VLP models (e.g., high transferable AEs), can advance reliable and practical VLP models. A recent work (i.e., Set-level guidance attack) indicates that augmenting image-text pairs to increase AE diversity along the optimization path enhances the transferability of adversarial examples significantly. However, this approach predominantly emphasizes diversity around the online adversarial examples (i.e., AEs in the optimization period), leading to the risk of overfitting the victim model and affecting the transferability. In this study, we posit that the diversity of adversarial examples towards the clean input and online AEs are both pivotal for enhancing transferability across VLP models. Consequently, we propose using diversification along the intersection region of adversarial trajectory to expand the diversity of AEs. To fully leverage the interaction between modalities, we introduce text-guided adversarial example selection during optimization. Furthermore, to further mitigate the potential overfitting, we direct the adversarial text deviating from the last intersection region along the optimization path, rather than adversarial images as in existing methods. Extensive experiments affirm the effectiveness of our method in improving transferability across various VLP models and downstream vision-and-language tasks. Code is available at https://github.com/SensenGao/VLTPredictAttack." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 561, + 453, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 561, + 453, + 583 + ], + "spans": [ + { + "bbox": [ + 159, + 561, + 453, + 583 + ], + "type": "text", + "content": "Keywords: Vision-Language Attack " + }, + { + "bbox": [ + 159, + 561, + 453, + 583 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 561, + 453, + 583 + ], + "type": "text", + "content": " Adversarial Transferability " + }, + { + "bbox": [ + 159, + 561, + 453, + 583 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 561, + 453, + 583 + ], + "type": "text", + "content": " Diversification " + }, + { + "bbox": [ + 159, + 561, + 453, + 583 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 561, + 453, + 583 + ], + "type": "text", + "content": " Intersection Region of Adversarial Trajectory" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 601, + 230, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 230, + 613 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 230, + 613 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 624, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 624, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 624, + 482, + 648 + ], + "type": "text", + "content": "Vision-language pre-training (VLP) models utilize multimodal learning, leveraging large-scale image-text pairs to bridge the gap between visual and lan" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 321, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 321, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 321, + 666 + ], + "type": "text", + "content": "* co-first authors; † co-corresponding authors" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 168, + 114, + 271, + 201 + ], + "blocks": [ + { + "bbox": [ + 168, + 114, + 271, + 201 + ], + "lines": [ + { + "bbox": [ + 168, + 114, + 271, + 201 + ], + "spans": [ + { + "bbox": [ + 168, + 114, + 271, + 201 + ], + "type": "image", + "image_path": "f6a4ec0e69e781aa3dff3ed1823ef6f5ab5b39259ef55a4b14a441af60845903.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 202, + 231, + 211 + ], + "lines": [ + { + "bbox": [ + 203, + 202, + 231, + 211 + ], + "spans": [ + { + "bbox": [ + 203, + 202, + 231, + 211 + ], + "type": "text", + "content": "(a) SGA" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 272, + 114, + 460, + 201 + ], + "blocks": [ + { + "bbox": [ + 272, + 114, + 460, + 201 + ], + "lines": [ + { + "bbox": [ + 272, + 114, + 460, + 201 + ], + "spans": [ + { + "bbox": [ + 272, + 114, + 460, + 201 + ], + "type": "image", + "image_path": "1433f9d0752d691d5c98279c201a9d8e7f6b4247bd6c9ca74ec0679249d814a4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 152, + 212, + 302, + 301 + ], + "blocks": [ + { + "bbox": [ + 152, + 212, + 302, + 301 + ], + "lines": [ + { + "bbox": [ + 152, + 212, + 302, + 301 + ], + "spans": [ + { + "bbox": [ + 152, + 212, + 302, + 301 + ], + "type": "image", + "image_path": "d4cfe77792677fc82598e233a866afc7b3630619180647168f842c8a85f8716c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 303, + 277, + 312 + ], + "lines": [ + { + "bbox": [ + 200, + 303, + 277, + 312 + ], + "spans": [ + { + "bbox": [ + 200, + 303, + 277, + 312 + ], + "type": "text", + "content": "(c) ALBEF to " + }, + { + "bbox": [ + 200, + 303, + 277, + 312 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{CNN}}" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 212, + 461, + 301 + ], + "blocks": [ + { + "bbox": [ + 348, + 203, + 377, + 212 + ], + "lines": [ + { + "bbox": [ + 348, + 203, + 377, + 212 + ], + "spans": [ + { + "bbox": [ + 348, + 203, + 377, + 212 + ], + "type": "text", + "content": "(b) Ours" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 212, + 461, + 301 + ], + "lines": [ + { + "bbox": [ + 309, + 212, + 461, + 301 + ], + "spans": [ + { + "bbox": [ + 309, + 212, + 461, + 301 + ], + "type": "image", + "image_path": "a1b3abab916e308cf1df4006807d87b09cc0a3194075a0ad8300b6fae7cc8653.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 303, + 436, + 312 + ], + "lines": [ + { + "bbox": [ + 358, + 303, + 436, + 312 + ], + "spans": [ + { + "bbox": [ + 358, + 303, + 436, + 312 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 358, + 303, + 436, + 312 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{ViT}}" + }, + { + "bbox": [ + 358, + 303, + 436, + 312 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 358, + 303, + 436, + 312 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{CNN}}" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 323, + 482, + 422 + ], + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 422 + ], + "type": "text", + "content": "Fig. 1: Our method vs. set-level guided attack (SGA) [36]. (a) shows the main idea of SGA, i.e., conducting augmentation around the online adversarial examples. (b) shows the main idea of our method, that is, we perform augmentation in the intersection region of adversarial trajectory. The red and blue dots both depict images sampled from the intersection region, with red dots indicating the best samples selected using the text-guided adversarial example selection strategy. The surrounding light red dots represent applying the same resizing data augmentation to the best samples as utilized in SGA. (c) and (d) compare the transferability of our method and SGA by using the adversarial examples of ALBEF [30] and " + }, + { + "bbox": [ + 130, + 323, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{ViT}}" + }, + { + "bbox": [ + 130, + 323, + 482, + 422 + ], + "type": "text", + "content": " to attack " + }, + { + "bbox": [ + 130, + 323, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{CNN}}" + }, + { + "bbox": [ + 130, + 323, + 482, + 422 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "text", + "content": "guage understanding. These models showcase revolutionary performance across various downstream Vision-and-Language tasks, including image-text retrieval, image captioning, visual grounding, and visual entailment, as demonstrated in [20, 27, 28, 44]. Despite their success, recent research underscores the significant vulnerability of VLP models, particularly when confronted with multimodal adversarial examples [7, 12, 16, 17, 36, 37, 57]. Uncovering the vulnerabilities, especially common issues, can drive further research aimed at building more reliable and practical VLP models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "content": "Current research predominantly concentrates on attacking VLP models via a white-box setting, where the model's structural information can be exploited. However, exploring the transferability of multimodal adversarial examples is pivotal, especially given the limited access to detailed model structures in real-world scenarios. There have been some efforts made to enhance the transferability of attacks on VLP models by introducing input diversity, as seen in the work SGA [36]. While they have achieved some effectiveness, how to enhance the transferability of multimodal adversarial examples is still an open question." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "In this paper, we undertake a comprehensive examination of the factors contributing to the limited transferability of the cutting-edge multimodal attack" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "S. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "method, i.e., SGA [36]. As shown in Figure 1 (a), throughout the iterative generation of subsequent adversarial images, SGA conducts data augmentation around the online adversarial image (i.e., adversarial examples generated during optimization). This strategy enhances the diversity of adversarial examples along the optimization path, leading to a certain improvement in transferability. However, such an approach still carries the potential risk of overfitting to the victim model with the high reliance on the examples along the adversarial trajectory (See Figure 1 (b)), which leads to low attack success rates when we migrate the adversarial example to other VLP models (See Figure 1 (c) and (d)). To mitigate this overfitting risk, one potential solution is to further enhance the diversity of augmented adversarial examples in a judicious manner." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 251, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 430 + ], + "type": "text", + "content": "Building upon the aforementioned analysis, SGA overfits local adversarial examples, while the clean image is the only accessible example that is far from local adversarial examples. Therefore, we embark on a pioneering endeavor to enhance the transferability of multimodal adversarial attacks by considering the diversity of adversarial examples (AEs) around clean inputs and online AEs throughout the optimization process. To achieve this, we consider the intersection region of adversarial trajectory, which encompasses the original image, the adversarial image from the previous step, and the current adversarial image during the iterative attack process (depicted in Figure 1 (b)). This innovative approach aims to circumvent overfitting by strategic sampling within this region, thereby avoiding an undue focus on adversarial example diversity solely around adversarial images. After obtaining multiple samples, we calculate gradients for each to determine perturbation directions away from the text. Subsequently, we individually incorporate these perturbations into the current adversarial image and select the one that deviates the most from the text." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 434, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 495 + ], + "type": "text", + "content": "Additionally, in the text modality, SGA only considers deviating the text from the last adversarial image in the optimization period, but the adversarial image is solely generated by the surrogate model, still posing the risk of overfitting the surrogate model. For this reason, we propose to have the text deviate simultaneously from the last intersection region along the optimization path." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 497, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 594 + ], + "type": "text", + "content": "Our proposed method is evaluated on two widely recognized multimodal datasets, Flickr30K [41] and MSCOCO [35]. We conduct experiments on three vision-and-language downstream tasks (i.e., image-text retrieval (ITR), visual grounding(VG), and image captioning (IC)), and all results indicate the high effectiveness of our method in generating more transferable multimodal adversarial examples. Moreover, when adversarial examples generated from image-text retrieval are transferred to other vision-and-language downstream tasks (i.e., VG and IC), there is a substantial improvement in attack performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 596, + 387, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 596, + 387, + 607 + ], + "spans": [ + { + "bbox": [ + 146, + 596, + 387, + 607 + ], + "type": "text", + "content": "Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "type": "text", + "content": "Boosting Transferability in Vision-Language Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 138, + 629, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 629, + 481, + 665 + ], + "type": "text", + "content": "- We propose using the intersection region of adversarial trajectory to expand the diversity of adversarial examples during optimization, based on which we develop a high-transferability attack against VLP models." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 479, + 151 + ], + "type": "text", + "content": "- We extend the generation of adversarial text to deviate from the last intersection region along the optimization path, aiming to reduce overfitting the surrogate model, thereby achieving enhanced transferability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 152, + 480, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 152, + 480, + 187 + ], + "spans": [ + { + "bbox": [ + 138, + 152, + 480, + 187 + ], + "type": "text", + "content": "- Extensive experiments robustly demonstrate the efficacy of our proposed method in elevating the transferability of multimodal adversarial examples across diverse models and three downstream tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 207, + 237, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 237, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 237, + 220 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 233, + 350, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 233, + 350, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 233, + 350, + 247 + ], + "type": "text", + "content": "2.1 Vision-Language Pre-training Models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 255, + 480, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 480, + 338 + ], + "type": "text", + "content": "VLP models leverage multimodal learning from extensive image-text pairs to improve the performance of various Vision-and-Language " + }, + { + "bbox": [ + 130, + 255, + 480, + 338 + ], + "type": "inline_equation", + "content": "(\\mathrm{V} + \\mathrm{L})" + }, + { + "bbox": [ + 130, + 255, + 480, + 338 + ], + "type": "text", + "content": " tasks [29]. Early VLP models predominantly depend on pre-trained object detectors for acquiring multimodal representations [6, 32, 45, 49, 59]. Recently, with the advent of end-to-end image encoders like the Vision Transformer (ViT) [10, 46, 56] offering faster inference speeds, some work propose to use them as substitution for computationally expensive object detectors [11, 29, 30, 48, 53]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 339, + 481, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 481, + 445 + ], + "type": "text", + "content": "There are two popular approaches for VLP models in learning vision-language representations: the fused architecture and the aligned architecture. Fused VLP models (e.g., ALBEF [30], TCL [53]), initially employ two separate unimodal encoders to learn features for text and images. Subsequently, a multimodal encoder is utilized to fuse the embeddings of text and images. In contrast, aligned VLP models, exemplified by CLIP [42], focus on aligning the feature spaces of distinct unimodal encoders and benefit downstream tasks significantly [1]. This paper concentrates on evaluating our proposed method using multiple popular fused and aligned VLP models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 464, + 364, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 364, + 477 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 364, + 477 + ], + "type": "text", + "content": "2.2 Downstream Vision-and-Language Tasks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 486, + 480, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 480, + 544 + ], + "type": "text", + "content": "Image-Text Retrieval (ITR) involves retrieving pertinent information, textual or visual, in response to queries from another modality [5, 8, 51, 60]. This undertaking usually encompasses two sub-tasks: image-to-text retrieval (retrieving text based on an image query) and text-to-image retrieval (retrieving images given a text query)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 480, + 640 + ], + "type": "text", + "content": "In aligned VLP models, both the Text Retrieval (TR) and Image Retrieval (IR) tasks leverage ranking results determined by the similarity between text and image embeddings. However, in fused VLP models, where internal embedding spaces lack alignment across unimodal encoders, the similarity scores between image and text modalities are computed for all image-text pairs to retrieve the Top-N candidates. Subsequently, these Top-N candidates serve as input for the multimodal encoder, which computes the image-text matching score to establish the final ranking." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "type": "text", + "content": "Visual Grounding (VG) refers to the task of localizing the region within a visual scene with corresponding entities or concepts in natural language. Among" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "S. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 167, + 465, + 206 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 1: Attack Success Rate (%) of SGA with and without image augmentation. The SGA w.o. Aug doesn't consider image augmentation. We use ALBEF to generate multimodal adversarial examples on the ITR task to evaluate transferability." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 167, + 465, + 206 + ], + "lines": [ + { + "bbox": [ + 151, + 167, + 465, + 206 + ], + "spans": [ + { + "bbox": [ + 151, + 167, + 465, + 206 + ], + "type": "table", + "html": "
SourceAttackALBEFTCLCLIPvITCLIPCNN
TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1
ALBEFSGA w.o. Aug99.999.9570.0771.6730.5539.8832.3142.54
SGA w. Aug99.999.9887.8888.0536.6946.7839.5949.78
", + "image_path": "688911c98308805d427f5d6efa63fa10c5bd25e52e789e82eee9266838fba66d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 238, + 480, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 262 + ], + "type": "text", + "content": "VLP models, ALBEF expands Grad-CAM [43] and utilizes the acquired attention map to rank the detected proposals [54]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 262, + 481, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 481, + 335 + ], + "type": "text", + "content": "Image Captioning (IC) is to generate a textual description that logically describes or implies the content of a given visual input, typically involving the creation of captions for images. The evaluation of Image Captioning models often employs metrics such as BLEU [39], METEOR [4], ROUGE [33], CIDEr [47] and SPICE [2], which serve to assess the quality and relevance of the generated captions in comparison to reference captions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 352, + 360, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 360, + 366 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 360, + 366 + ], + "type": "text", + "content": "2.3 Transferability of Adversarial Examples" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 373, + 482, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 373, + 482, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 373, + 482, + 503 + ], + "type": "text", + "content": "Adversarial attacks [13-15, 18, 25] are typically categorized as white-box and black-box attacks. In a white-box setting [23, 26], the attacker has full access to the model, whereas black-box attacks [3, 40], more realistic in practical applications, occur when information about the model is limited. In the realm of image attacks [21, 22, 24], prevalent methods for crafting transferable adversarial examples often leverage data augmentation techniques (e.g., DIM [52], TIM [9], SIM [34], ADMIX [50], PAM [58]). Zhang et al. [57] introduced a white-box attack targeting popular VLP models for downstream tasks in the multimodal domain. Building upon this work, Lu et al. [36] proposed SGA, considering the diversity of adversarial examples by expanding single image-text pairs to sets of images and texts to conduct black-box attacks on VLP models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 505, + 481, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 505, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 505, + 481, + 577 + ], + "type": "text", + "content": "However, SGA primarily emphasizes diversity in the vicinity of adversarial examples during the optimization process, potentially increasing the risk of overfitting the victim model and impacting transferability. Therefore, our primary focus in this study is to further enhance the diversity of adversarial examples in a thoughtful manner and avoid an undue focus on adversarial example diversity solely around adversarial images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 594, + 233, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 233, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 233, + 609 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 620, + 302, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 302, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 302, + 633 + ], + "type": "text", + "content": "3.1 Background and Motivation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "content": "Adversarial attacks on VLP models involve inducing a mismatch between adversarial images and corresponding adversarial text while adhering to specified" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "type": "text", + "content": "Boosting Transferability in Vision-Language Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "constraints on image and text perturbations. Here, " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "(v,t)" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " denotes an original image-text pair from a multimodal dataset, with " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "v^{\\prime}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " representing an adversarial image and " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "t^\\prime" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " denoting adversarial text. The allowable perturbations are restricted within the ranges " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "B[v,\\xi_v]" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " for images and " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "B[t,\\xi_t]" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " for text. The image and text encoders of the multimodal model are denoted as " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "F_{I}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "F_{T}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ", respectively. To generate valid multimodal adversarial examples, the objective is to maximize the loss function " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " specific to VLP models:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 241, + 209, + 481, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 209, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 241, + 209, + 481, + 236 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\max J \\left(F _ {I} \\left(v ^ {\\prime}\\right), F _ {T} \\left(t ^ {\\prime}\\right)\\right) \\\\ s. t. v ^ {\\prime} \\in B [ v, \\xi_ {v} ], t ^ {\\prime} \\in B [ t, \\xi_ {t} ]. \\end{array} \\right. \\tag {1}", + "image_path": "9dcbcf1cf0a901c8af21b4311431351662f123bc82cd0fa7c28c15445ba8a872.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 243, + 482, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 243, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 131, + 243, + 482, + 340 + ], + "type": "text", + "content": "The state-of-the-art approach for exploring the transferability of multimodal adversarial examples (i.e., SGA [36]) involves augmenting image-text pairs to enhance the diversity of adversarial examples along the optimization path. Specifically, during the iterative generation of adversarial images, let " + }, + { + "bbox": [ + 131, + 243, + 482, + 340 + ], + "type": "inline_equation", + "content": "v_{i}^{\\prime}" + }, + { + "bbox": [ + 131, + 243, + 482, + 340 + ], + "type": "text", + "content": " represent the adversarial image generated at the " + }, + { + "bbox": [ + 131, + 243, + 482, + 340 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 243, + 482, + 340 + ], + "type": "text", + "content": "-th step. In the subsequent step " + }, + { + "bbox": [ + 131, + 243, + 482, + 340 + ], + "type": "inline_equation", + "content": "(i + 1)" + }, + { + "bbox": [ + 131, + 243, + 482, + 340 + ], + "type": "text", + "content": ", SGA initiates the process by applying a resizing operation for data augmentation to " + }, + { + "bbox": [ + 131, + 243, + 482, + 340 + ], + "type": "inline_equation", + "content": "v_{i}^{\\prime}" + }, + { + "bbox": [ + 131, + 243, + 482, + 340 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 131, + 243, + 482, + 340 + ], + "type": "inline_equation", + "content": "V_{i}^{\\prime} = \\{v_{i1}^{\\prime}, v_{i2}^{\\prime}, \\dots, v_{iM}^{\\prime}\\}" + }, + { + "bbox": [ + 131, + 243, + 482, + 340 + ], + "type": "text", + "content": " (See Figure 1 (a)). The iterative formula can be expressed as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 196, + 349, + 481, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 349, + 481, + 383 + ], + "spans": [ + { + "bbox": [ + 196, + 349, + 481, + 383 + ], + "type": "interline_equation", + "content": "v _ {i + 1} ^ {\\prime} = v _ {i} ^ {\\prime} + \\alpha \\cdot \\operatorname {s i g n} \\left(\\frac {\\nabla_ {v} \\sum_ {j = 1} ^ {M} J \\left(F _ {I} \\left(v _ {i j} ^ {\\prime}\\right) , F _ {T} (t)\\right)}{\\left\\| \\nabla_ {v} \\sum_ {j = 1} ^ {M} J \\left(F _ {I} \\left(v _ {i j} ^ {\\prime}\\right) , F _ {T} (t)\\right) \\right\\|}\\right). \\tag {2}", + "image_path": "b741edc8d0092bb79682eb5f1e17724d4c53bfb510d01f8fd0645db45138a44d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 390, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 533 + ], + "type": "text", + "content": "To further examine the impact of image augmentation along the optimization path in the SGA method, we utilize ALBEF as a surrogate model to generate multimodal adversarial examples. These examples are then employed to target VLP models such as TCL and CLIP, assessing the transferability of the attacks. Detailed results are presented in Table 1. Our observations reveal that SGA enhances the transferability of adversarial attacks, showing an increase ranging from " + }, + { + "bbox": [ + 130, + 390, + 482, + 533 + ], + "type": "inline_equation", + "content": "6.14\\%" + }, + { + "bbox": [ + 130, + 390, + 482, + 533 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 390, + 482, + 533 + ], + "type": "inline_equation", + "content": "17.81\\%" + }, + { + "bbox": [ + 130, + 390, + 482, + 533 + ], + "type": "text", + "content": ". However, it is noteworthy that the success rate of attacks on the target models remains notably lower than that on the source model. This discrepancy is primarily attributed to the fact that SGA predominantly emphasizes diversity around AE " + }, + { + "bbox": [ + 130, + 390, + 482, + 533 + ], + "type": "inline_equation", + "content": "v_{i}^{\\prime}" + }, + { + "bbox": [ + 130, + 390, + 482, + 533 + ], + "type": "text", + "content": " during the optimization period, without adequately considering the diversity of adversarial examples toward the clean image, leading to the risk of overfitting the victim model and affecting the transferability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "For this purpose, we propose to consider diversification along the intersection region of adversarial trajectory, which encompasses the original image " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": ", the adversarial image from the previous step " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "v_{i-1}'" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": ", and the current adversarial image " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "v_i'" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": " during the iterative attack process. This region is established to sample images within it to broaden the diversity of adversarial examples (See Figure 1 (b)). Moreover, to fully leverage the interplay between modalities, we aim for perturbations guided by textual information that induce " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "v_i'" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": " to deviate significantly from the associated text " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": ". Additionally, in the text modality, our objective is to identify adversarial perturbations that simultaneously deviate from the intersection region rather than only adversarial images, thereby reducing overfitting the surrogate model and enhancing the effectiveness of black-box attacks." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "S. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 386, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 386, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 386, + 129 + ], + "type": "text", + "content": "3.2 Diversification along the Intersection Region" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "text", + "content": "As outlined in Section 3.1, we enhance the diversity of adversarial examples by introducing diversification along the intersection region. Specifically, at the " + }, + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "text", + "content": "th iteration during optimization, we have the " + }, + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "inline_equation", + "content": "v_{i}^{\\prime}" + }, + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "inline_equation", + "content": "v_{i-1}^{\\prime}" + }, + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "text", + "content": ", and the clean " + }, + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "text", + "content": ", and these variables form a triangle region denoted as " + }, + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "inline_equation", + "content": "\\triangle vv_{i-1}^{\\prime}v_{i}^{\\prime}" + }, + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "text", + "content": ", i.e., the intersection region of adversarial trajectory in Figure 1. Then, we initially sample multiple instances within the region " + }, + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "inline_equation", + "content": "\\triangle vv_{i-1}^{\\prime}v_{i}^{\\prime}" + }, + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "text", + "content": ", representing the set of samples as " + }, + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "inline_equation", + "content": "e = \\{e_1,e_2,\\dots,e_N\\}" + }, + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "text", + "content": ". Each sample can be expressed as " + }, + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "inline_equation", + "content": "e_k = \\beta \\cdot v + \\gamma \\cdot v_{i-1}^{\\prime} + \\eta \\cdot v_i^{\\prime}" + }, + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "inline_equation", + "content": "\\beta + \\gamma + \\eta = 1.0" + }, + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "text", + "content": ". Consequently, we can compute the gradient perturbation for each sample. For the k-th sample, denoted as " + }, + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "inline_equation", + "content": "e_k" + }, + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "text", + "content": ", its gradient perturbation " + }, + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "text", + "content": " is calculated as follows. In this way, we can get a perturbation set " + }, + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "inline_equation", + "content": "P = \\{p_1,p_2,\\dots,p_N\\}" + }, + { + "bbox": [ + 130, + 136, + 482, + 268 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 225, + 276, + 481, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 276, + 481, + 301 + ], + "spans": [ + { + "bbox": [ + 225, + 276, + 481, + 301 + ], + "type": "interline_equation", + "content": "p _ {k} = \\alpha \\cdot \\operatorname {s i g n} \\left(\\frac {\\nabla_ {e} J \\left(F _ {I} \\left(e _ {k}\\right) , F _ {T} (t)\\right)}{\\| \\nabla_ {e} J \\left(F _ {I} \\left(e _ {k}\\right) , F _ {T} (t)\\right) \\|}\\right). \\tag {3}", + "image_path": "ffcf312c5f33c9675d3c59ac6a32f6be0f5a2e6b53f8d11d2b3cbd93a0e4b428.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 317, + 345, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 317, + 345, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 317, + 345, + 330 + ], + "type": "text", + "content": "3.3 Text-guided Augmentation Selection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 336, + 480, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 480, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 480, + 409 + ], + "type": "text", + "content": "In Section 3.2, a diverse perturbation set " + }, + { + "bbox": [ + 130, + 336, + 480, + 409 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 336, + 480, + 409 + ], + "type": "text", + "content": " is derived from the intersection region. To harness the full potential of modality interactions, we introduce text-guided augmentation selection to obtain the optimal sample. Specifically, we individually incorporate each element from the perturbation set " + }, + { + "bbox": [ + 130, + 336, + 480, + 409 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 336, + 480, + 409 + ], + "type": "text", + "content": " into the adversarial image " + }, + { + "bbox": [ + 130, + 336, + 480, + 409 + ], + "type": "inline_equation", + "content": "v_{i}^{\\prime}" + }, + { + "bbox": [ + 130, + 336, + 480, + 409 + ], + "type": "text", + "content": ". The selection process aims to identify the sample that maximally distances " + }, + { + "bbox": [ + 130, + 336, + 480, + 409 + ], + "type": "inline_equation", + "content": "v_{i}^{\\prime}" + }, + { + "bbox": [ + 130, + 336, + 480, + 409 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 336, + 480, + 409 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 336, + 480, + 409 + ], + "type": "text", + "content": ". This procedure can be represented as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 230, + 420, + 480, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 420, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 230, + 420, + 480, + 440 + ], + "type": "interline_equation", + "content": "m = \\underset {p _ {m} \\in P} {\\arg \\max } J \\left(F _ {I} \\left(v _ {i} ^ {\\prime} + p _ {m}\\right), F _ {T} (t)\\right). \\tag {4}", + "image_path": "3e50a96d9c87a1c6033ccccc9d00645194e4187f6b7c28849eed93d182f54f7e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "text", + "content": "At this juncture, " + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "inline_equation", + "content": "e_m" + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "text", + "content": " represents the selected sample. We employ SGA as our baseline and incorporate the image augmentation methods considered along its optimization path. The chosen optimal sample " + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "inline_equation", + "content": "e_m" + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "text", + "content": " is resized and expanded into the set " + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "inline_equation", + "content": "E_m = \\{e_{m1}, e_{m2}, \\dots, e_{mM}\\}" + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "text", + "content": ". Subsequently, we utilize the expanded set " + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "inline_equation", + "content": "E_m" + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "text", + "content": " to generate the final adversarial perturbation, yielding " + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "inline_equation", + "content": "v_{i+1}'" + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 195, + 513, + 480, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 513, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 195, + 513, + 480, + 546 + ], + "type": "interline_equation", + "content": "v _ {i + 1} ^ {\\prime} = v _ {i} ^ {\\prime} + \\alpha \\cdot \\operatorname {s i g n} \\left(\\frac {\\nabla_ {e} \\sum_ {j = 1} ^ {M} J \\left(F _ {I} \\left(e _ {m j}\\right) , F _ {T} (t)\\right)}{\\| \\nabla_ {e} \\sum_ {j = 1} ^ {M} J \\left(F _ {I} \\left(e _ {m j}, F _ {T} (t)\\right) \\right\\|}\\right). \\tag {5}", + "image_path": "dba6a273c277b320adb6e4be366a6a7c0128283fa96b5be811a00c8734cfa4e8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 562, + 444, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 562, + 444, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 562, + 444, + 574 + ], + "type": "text", + "content": "3.4 Adversarial Text deviating from the Intersection Region" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "In the text modality, SGA only considers deviating adversarial text from the ultimate adversarial image generated during the iterative optimization process. If there are a total of T iterations, The generation of " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " only considers deviating from the adversarial image " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "v_T'" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ". The adversarial image " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "v_T'" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " is exclusively created by the surrogate model, thereby still presenting the risk of overfitting to the surrogate model. However, during the optimization process of the adversarial image, the clean image is entirely independent of the surrogate model. For this" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 233, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 233, + 91, + 448, + 103 + ], + "type": "text", + "content": "Boosting Transferability in Vision-Language Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 154 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 154 + ], + "type": "text", + "content": "reason, we propose to have the text deviate simultaneously from the last intersection region along the optimization path. Specifically, the adversarial text deviates from the triangle region constituted by " + }, + { + "bbox": [ + 130, + 116, + 479, + 154 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 116, + 479, + 154 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 479, + 154 + ], + "type": "inline_equation", + "content": "v_{T-1}'" + }, + { + "bbox": [ + 130, + 116, + 479, + 154 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 154 + ], + "type": "inline_equation", + "content": "v_T'" + }, + { + "bbox": [ + 130, + 116, + 479, + 154 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 162, + 481, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 481, + 194 + ], + "type": "interline_equation", + "content": "t ^ {\\prime} = \\underset {t ^ {\\prime} \\in B [ t, \\epsilon_ {t} ]} {\\arg \\max } (\\lambda \\cdot J (F _ {I} (v), F _ {T} (t ^ {\\prime})) + \\mu \\cdot J (F _ {I} (v _ {T} ^ {\\prime}), F _ {T} (t ^ {\\prime}))) + \\nu \\cdot J (F _ {I} (v _ {T - 1} ^ {\\prime}), F _ {T} (t ^ {\\prime}))). \\tag {6}", + "image_path": "3631caeeecfbe9a466ac2c17512137251cb836f0e53f9f35ac7af9b736721015.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 435, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 435, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 435, + 206 + ], + "type": "text", + "content": "We also set adjustable scaling factors, among which " + }, + { + "bbox": [ + 132, + 194, + 435, + 206 + ], + "type": "inline_equation", + "content": "\\lambda +\\mu +\\nu = 1.0" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 224, + 279, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 279, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 279, + 237 + ], + "type": "text", + "content": "3.5 Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 246, + 482, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 377 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 377 + ], + "type": "text", + "content": "In the specific process of our attack, we employ an iterative approach. In each iteration, we sample within the intersection region of adversarial trajectory, guided by textual information, to select a sample that maximally deviates the current adversarial image " + }, + { + "bbox": [ + 130, + 246, + 482, + 377 + ], + "type": "inline_equation", + "content": "v_{i}^{\\prime}" + }, + { + "bbox": [ + 130, + 246, + 482, + 377 + ], + "type": "text", + "content": " from the text " + }, + { + "bbox": [ + 130, + 246, + 482, + 377 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 246, + 482, + 377 + ], + "type": "text", + "content": ". Subsequently, we subject this sample to image augmentation processing, calculate gradients to determine the perturbation direction, and overlay it onto the current adversarial image, resulting in " + }, + { + "bbox": [ + 130, + 246, + 482, + 377 + ], + "type": "inline_equation", + "content": "v_{i + 1}^{\\prime}" + }, + { + "bbox": [ + 130, + 246, + 482, + 377 + ], + "type": "text", + "content": ". Through multiple iterative steps, we obtain an adversarial image " + }, + { + "bbox": [ + 130, + 246, + 482, + 377 + ], + "type": "inline_equation", + "content": "v_{T}^{\\prime}" + }, + { + "bbox": [ + 130, + 246, + 482, + 377 + ], + "type": "text", + "content": ". For the text modality, in contrast to previous methods that solely focus on deviating from the adversarial image " + }, + { + "bbox": [ + 130, + 246, + 482, + 377 + ], + "type": "inline_equation", + "content": "v_{T}^{\\prime}" + }, + { + "bbox": [ + 130, + 246, + 482, + 377 + ], + "type": "text", + "content": ", our goal is to derive an adversarial text " + }, + { + "bbox": [ + 130, + 246, + 482, + 377 + ], + "type": "inline_equation", + "content": "t^{\\prime}" + }, + { + "bbox": [ + 130, + 246, + 482, + 377 + ], + "type": "text", + "content": " that simultaneously deviates from the last intersection region " + }, + { + "bbox": [ + 130, + 246, + 482, + 377 + ], + "type": "inline_equation", + "content": "\\triangle vv_{T - 1}^{\\prime}v_{T}^{\\prime}" + }, + { + "bbox": [ + 130, + 246, + 482, + 377 + ], + "type": "text", + "content": " along the optimization path." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 396, + 230, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 230, + 410 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 230, + 410 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 422, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 555 + ], + "type": "text", + "content": "In this section, we present experimental evidence demonstrating the enhanced transferability of multimodal examples generated from our proposed method across VLP models and various Vision-and-Language tasks. First, in Section 4.1, we introduce the experimental settings, including the popular image-text pair datasets and VLP models we use, as well as restrictions for adversarial attacks. Subsequently, the process of searching for optimal parameters is shown in Section 4.2. After that, we evaluate cross-model transferability in the context of the image-text retrieval task, as detailed in Section 4.3. Following this, in Section 4.4, we extend our investigation to the transfer of multimodal adversarial examples generated within the image-text retrieval task to other tasks, aiming to gauge cross-task transferability. Lastly, Section 4.5 outlines ablation studies." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 573, + 195, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 573, + 195, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 573, + 195, + 586 + ], + "type": "text", + "content": "4.1 Setups" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": "VLP Models. In our transferability evaluation experiments across various VLP models, we explore two typical architectures: fused and aligned VLP models. We select CLIP [42] for the aligned VLP model. CLIP offers a choice between two distinct image encoders, namely ViT-B/16 [10] and ResNet-101 [19], denoted as " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{ViT}}" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{CNN}}" + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": ", respectively. In the case of fused VLP models, we opt for ALBEF [30] and TCL [53]. ALBEF uses a 12-layer ViT-B/16 image encoder" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "S. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "and two 6-layer transformers for text and multimodal encoding. TCL shares this architecture but has different pre-training objectives." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 200 + ], + "type": "text", + "content": "Datasets. In this study, we leverage two widely recognized multimodal datasets, namely Flickr30K [41] and MSCOCO [35], for evaluating the image-text retrieval task. The Flickr30K dataset comprises 31,783 images, each accompanied by five captions for annotation. Similarly, the MSCOCO dataset consists of 123,287 images, and approximately five captions are provided for each image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 272 + ], + "type": "text", + "content": "Additionally, we employ the RefCOCO+ [55] dataset to assess the Visual Grounding task. RefCOCO+ is a dataset containing 141,564 referring expressions for 50,000 objects within 19,992 MSCOCO images. This dataset serves the purpose of evaluating grounding models by focusing on the localization of objects described through natural language. For another Vision-and-Language task, Image Captioning, we leverage the MSCOCO dataset as well." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "text", + "content": "Adversarial Attack Settings. In our study, we adopt adversarial attack settings of SGA [36] to ensure a fair comparison. Specifically, we leverage BERTAttack [31] to craft adversarial texts. The perturbation bound " + }, + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\xi_{t}" + }, + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "text", + "content": " is set as 1 and length of word list " + }, + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "inline_equation", + "content": "W = 10" + }, + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "text", + "content": ". PGD [38] is employed to get adversarial images and the perturbation bound, denoted as " + }, + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\xi_v" + }, + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "text", + "content": ", is set as 8/255. Additionally, iteration steps " + }, + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "text", + "content": " is set as 10 and each step size " + }, + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\alpha = 2 / 255" + }, + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "text", + "content": ". Furthermore, when randomly sampling from the intersection region of adversarial trajectory, we set the number of samples to 5. When generating adversarial texts, we set the three parameters " + }, + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\lambda, \\mu, \\nu" + }, + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "text", + "content": " in Equation 6 to 0.6, 0.2, and 0.2 respectively. The values chosen for these adjustable parameters can be found in Section 4.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 391, + 481, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 481, + 439 + ], + "type": "text", + "content": "Evaluation Metrics. The key metric for adversarial transferability is the Attack Success Rate (ASR), which measures the percentage of successful attacks among all generated adversarial examples. A higher ASR indicates more effective and transferable attacks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 456, + 263, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 456, + 263, + 467 + ], + "spans": [ + { + "bbox": [ + 131, + 456, + 263, + 467 + ], + "type": "text", + "content": "4.2 Optimal Parameters" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 474, + 481, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 544 + ], + "type": "text", + "content": "In our proposed method, the number of samples " + }, + { + "bbox": [ + 130, + 474, + 481, + 544 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 474, + 481, + 544 + ], + "type": "text", + "content": " taken from the intersection region of adversarial trajectory and scaling factors in Equation 6 are adjustable. We conduct specific experiments to explore optimal parameter settings and examine their influence on the efficacy of our approach. To be more specific, we utilize ALBEF for generating multimodal adversarial examples on the Flickr30K dataset and assess the transferability on the other three VLP models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 629 + ], + "type": "text", + "content": "Number of Samples taken from Intersection Region of Adversarial trajectory " + }, + { + "bbox": [ + 130, + 545, + 481, + 629 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 545, + 481, + 629 + ], + "type": "text", + "content": ". The bottom of Table 2 illustrates when the sample size reaches 5, transfer effects are observed for both the Image Retrieval task and the Text Retrieval task. As the sample size continues to increase, the transferability only fluctuates and does not exhibit further improvement. Taking into account both transfer effects and computational costs, a sample size of 5 is identified as the optimal configuration." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": "Scaling factors " + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\lambda, \\mu, \\nu" + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": " in Adversarial Text Generation. In Equation 6, " + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": " represents the weight of clean images, while " + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": " represent the weights of adversarial images. Therefore, we stipulate that " + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": " cannot be zero, and " + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\nu" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "type": "text", + "content": "Boosting Transferability in Vision-Language Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 156, + 465, + 297 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 2: Optimal Parameters: Attack Success Rate(%) on different settings, Top for different values of " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "\\lambda ,\\mu ,\\nu" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": " and Bottom for different numbers of samples " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "N" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 156, + 465, + 297 + ], + "lines": [ + { + "bbox": [ + 151, + 156, + 465, + 297 + ], + "spans": [ + { + "bbox": [ + 151, + 156, + 465, + 297 + ], + "type": "table", + "html": "
SourceAttackALBEFTCLCLIPvITCLIPCNN
TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1
ALBEF[λ,μ,v] = [0.2,0.0,0.8]99.999.9389.6790.542.2152.046.2355.44
[λ,μ,v] = [0.2,0.2,0.6]99.999.9390.4190.4341.9651.8746.4955.3
[λ,μ,v] = [0.2,0.4,0.4]99.999.9390.3190.4341.9651.8745.3454.82
[λ,μ,v] = [0.2,0.6,0.2]99.999.9590.5290.5742.2151.8445.0854.92
[λ,μ,v] = [0.2,0.8,0.0]99.999.9390.3190.5741.7251.7446.154.68
[λ,μ,v] = [0.4,0.0,0.6]99.999.9391.2590.8845.5255.3248.9157.63
[λ,μ,v] = [0.4,0.2,0.4]99.999.9391.2590.8345.0355.2248.2857.77
[λ,μ,v] = [0.4,0.4,0.2]99.999.9391.1590.7144.9155.2848.1557.87
[λ,μ,v] = [0.4,0.6,0.0]99.999.9391.1590.8844.9154.7749.0457.56
[λ,μ,v] = [0.6,0.0,0.4]99.999.9391.4690.9546.3856.3849.0459.11
[λ,μ,v] = [0.6,0.2,0.2]99.999.9391.5791.1746.2656.849.5559.01
[λ,μ,v] = [0.6,0.4,0.0]99.999.9390.9490.9846.0156.7249.5558.87
[λ,μ,v] = [0.8,0.0,0.2]99.999.9390.7390.9846.1356.7149.4658.74
[λ,μ,v] = [0.8,0.2,0.0]99.999.9390.3190.9545.7756.6549.3458.87
ALBEFN=399.999.9590.6290.7945.6456.5448.8358.73
N=499.7999.9191.3691.1745.8356.7850.4559.01
N=599.999.9391.5791.1746.2656.849.5559.01
N=699.999.9390.9490.3845.7956.9650.758.52
N=799.999.9189.8890.9545.456.3550.759.07
", + "image_path": "e51f203b1430437788955279349076032a2ee6e76575b5a6360d5cf54822cb1d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 329, + 482, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 401 + ], + "type": "text", + "content": "can have at most one zero value. Initially, as the value of " + }, + { + "bbox": [ + 130, + 329, + 482, + 401 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 329, + 482, + 401 + ], + "type": "text", + "content": " gradually increases, indicating the gradual introduction of clean images, the transferability of multimodal adversarial examples increases accordingly. However, when the value of " + }, + { + "bbox": [ + 130, + 329, + 482, + 401 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 329, + 482, + 401 + ], + "type": "text", + "content": " becomes too large, it also leads to a disproportionately low proportion of adversarial images, resulting in a decrease in adversarial transferability. According to the experiments, the optimal parameters we select are [0.6, 0.2, 0.2]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 418, + 302, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 302, + 431 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 302, + 431 + ], + "type": "text", + "content": "4.3 Cross-Model Transferability" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 438, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 557 + ], + "type": "text", + "content": "As outlined in Section 4.1, our experimental design focuses on assessing the transferability of adversarial examples across two widely adopted VLP model architectures: fused and aligned. There are four VLP models selected, namely: ALBEF, TCL, " + }, + { + "bbox": [ + 130, + 438, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{ViT}}" + }, + { + "bbox": [ + 130, + 438, + 482, + 557 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 438, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{CNN}}" + }, + { + "bbox": [ + 130, + 438, + 482, + 557 + ], + "type": "text", + "content": ". The selected downstream " + }, + { + "bbox": [ + 130, + 438, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\mathrm{V} + \\mathrm{L}" + }, + { + "bbox": [ + 130, + 438, + 482, + 557 + ], + "type": "text", + "content": " task for evaluation is image-text retrieval. Our approach involves employing one of four models to generate multimodal adversarial examples within the specified parameters of our proposed method. Subsequently, we validate the effectiveness of the generated adversarial examples through a comprehensive set of experiments, encompassing both self-attacks and attacks on the other three models. This evaluation encompasses one white-box attack and three distinct black-box attacks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "We adopt the methodology proposed by SGA [36] as our baseline, Therefore, the effectiveness of our method is compared against it. Moreover, we also present the effectiveness and transferability results of various other attack methods on VLP models. Table 3 provides a comprehensive comparison of these methods on the dataset Flickr30K, More experiments on the MSCOCO dataset are provided in the Appendix. In the comparison, PGD [38] is an image-only attack, while Bert-Attack [31] focuses solely on text-based attacks. Sep-Attack involves perturbing text and image separately. Furthermore, Co-Attack takes into account cross-modal interactions, generating an adversarial example for one modality" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "S. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 173, + 118, + 211, + 155 + ], + "blocks": [ + { + "bbox": [ + 151, + 128, + 167, + 146 + ], + "lines": [ + { + "bbox": [ + 151, + 128, + 167, + 146 + ], + "spans": [ + { + "bbox": [ + 151, + 128, + 167, + 146 + ], + "type": "text", + "content": "Cleannnnnne" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 173, + 118, + 211, + 155 + ], + "lines": [ + { + "bbox": [ + 173, + 118, + 211, + 155 + ], + "spans": [ + { + "bbox": [ + 173, + 118, + 211, + 155 + ], + "type": "image", + "image_path": "fa112191c91e91d2277120ecde637d54a2308149461c8aeb4606ed354312286e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 156, + 211, + 182 + ], + "lines": [ + { + "bbox": [ + 173, + 156, + 211, + 182 + ], + "spans": [ + { + "bbox": [ + 173, + 156, + 211, + 182 + ], + "type": "text", + "content": "a man riding on the back of a motorcycle down a street" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 215, + 117, + 253, + 155 + ], + "blocks": [ + { + "bbox": [ + 215, + 117, + 253, + 155 + ], + "lines": [ + { + "bbox": [ + 215, + 117, + 253, + 155 + ], + "spans": [ + { + "bbox": [ + 215, + 117, + 253, + 155 + ], + "type": "image", + "image_path": "b7ca1681e86010cf84c61bc96873c5e2ac7db0d058d23c0f6eb5c61370b5a36f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 156, + 253, + 175 + ], + "lines": [ + { + "bbox": [ + 216, + 156, + 253, + 175 + ], + "spans": [ + { + "bbox": [ + 216, + 156, + 253, + 175 + ], + "type": "text", + "content": "a man wearing a black shirt and a green tie" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 256, + 117, + 294, + 155 + ], + "blocks": [ + { + "bbox": [ + 256, + 117, + 294, + 155 + ], + "lines": [ + { + "bbox": [ + 256, + 117, + 294, + 155 + ], + "spans": [ + { + "bbox": [ + 256, + 117, + 294, + 155 + ], + "type": "image", + "image_path": "d759c7dc32f0c0e468c00a5596e8b6f92738d77d463285e551d7e74b8348f12e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 156, + 296, + 175 + ], + "lines": [ + { + "bbox": [ + 257, + 156, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 257, + 156, + 296, + 175 + ], + "type": "text", + "content": "a parking meter sitting on the side of a road" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 298, + 117, + 335, + 155 + ], + "blocks": [ + { + "bbox": [ + 298, + 117, + 335, + 155 + ], + "lines": [ + { + "bbox": [ + 298, + 117, + 335, + 155 + ], + "spans": [ + { + "bbox": [ + 298, + 117, + 335, + 155 + ], + "type": "image", + "image_path": "20ec060bcb801515f1977f1ffba3f0ae7943db9ad671f0e3a1bce39029325548.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 156, + 335, + 175 + ], + "lines": [ + { + "bbox": [ + 299, + 156, + 335, + 175 + ], + "spans": [ + { + "bbox": [ + 299, + 156, + 335, + 175 + ], + "type": "text", + "content": "a cat standing on a toilet seat in a bathroom" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 338, + 117, + 376, + 155 + ], + "blocks": [ + { + "bbox": [ + 338, + 117, + 376, + 155 + ], + "lines": [ + { + "bbox": [ + 338, + 117, + 376, + 155 + ], + "spans": [ + { + "bbox": [ + 338, + 117, + 376, + 155 + ], + "type": "image", + "image_path": "7306439b9c64d067525644f744d093202dc47c844c13cda18ce367a99fcb14d3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 156, + 373, + 175 + ], + "lines": [ + { + "bbox": [ + 339, + 156, + 373, + 175 + ], + "spans": [ + { + "bbox": [ + 339, + 156, + 373, + 175 + ], + "type": "text", + "content": "am an riding a bike next to a train" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 380, + 117, + 418, + 155 + ], + "blocks": [ + { + "bbox": [ + 380, + 117, + 418, + 155 + ], + "lines": [ + { + "bbox": [ + 380, + 117, + 418, + 155 + ], + "spans": [ + { + "bbox": [ + 380, + 117, + 418, + 155 + ], + "type": "image", + "image_path": "f8244d470fc556955750045aa8029aa5e6317db3fe6466b5ec4fe42bd51bff50.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 382, + 156, + 419, + 175 + ], + "lines": [ + { + "bbox": [ + 382, + 156, + 419, + 175 + ], + "spans": [ + { + "bbox": [ + 382, + 156, + 419, + 175 + ], + "type": "text", + "content": "a man sitting on a bench next to a white dog" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 421, + 117, + 459, + 155 + ], + "blocks": [ + { + "bbox": [ + 421, + 117, + 459, + 155 + ], + "lines": [ + { + "bbox": [ + 421, + 117, + 459, + 155 + ], + "spans": [ + { + "bbox": [ + 421, + 117, + 459, + 155 + ], + "type": "image", + "image_path": "ec9796b597d6cea774128ee16fa8c3987566cdc05c892c51f399dbd46011bc52.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 156, + 457, + 181 + ], + "lines": [ + { + "bbox": [ + 421, + 156, + 457, + 181 + ], + "spans": [ + { + "bbox": [ + 421, + 156, + 457, + 181 + ], + "type": "text", + "content": "a siamese cat eating a banana from a person's hand" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 174, + 184, + 212, + 223 + ], + "blocks": [ + { + "bbox": [ + 157, + 154, + 166, + 178 + ], + "lines": [ + { + "bbox": [ + 157, + 154, + 166, + 178 + ], + "spans": [ + { + "bbox": [ + 157, + 154, + 166, + 178 + ], + "type": "text", + "content": "Caaee" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 152, + 184, + 168, + 218 + ], + "lines": [ + { + "bbox": [ + 152, + 184, + 168, + 218 + ], + "spans": [ + { + "bbox": [ + 152, + 184, + 168, + 218 + ], + "type": "text", + "content": "Adversarial" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 157, + 224, + 166, + 249 + ], + "lines": [ + { + "bbox": [ + 157, + 224, + 166, + 249 + ], + "spans": [ + { + "bbox": [ + 157, + 224, + 166, + 249 + ], + "type": "text", + "content": "Caaee" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 174, + 184, + 212, + 223 + ], + "lines": [ + { + "bbox": [ + 174, + 184, + 212, + 223 + ], + "spans": [ + { + "bbox": [ + 174, + 184, + 212, + 223 + ], + "type": "image", + "image_path": "6a87880aeea6b07ebd99c6a6394751353bc2c97486a063c3b55c7ef4717f70b9.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 226, + 211, + 245 + ], + "lines": [ + { + "bbox": [ + 175, + 226, + 211, + 245 + ], + "spans": [ + { + "bbox": [ + 175, + 226, + 211, + 245 + ], + "type": "text", + "content": "a woman sitting on a motorcycle in the street" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 261, + 480, + 306 + ], + "lines": [ + { + "bbox": [ + 131, + 261, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 131, + 261, + 480, + 306 + ], + "type": "text", + "content": "Fig. 2: Visualization on Image Captioning. We use the ALBEF model, pre-trained on Image Text Retrieval(ITR) task, to generate adversarial images on the MSCOCO dataset and use the BLIP [29] model for Image Captioning on both clean images and adversarial images, respectively." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 215, + 184, + 253, + 223 + ], + "blocks": [ + { + "bbox": [ + 215, + 184, + 253, + 223 + ], + "lines": [ + { + "bbox": [ + 215, + 184, + 253, + 223 + ], + "spans": [ + { + "bbox": [ + 215, + 184, + 253, + 223 + ], + "type": "image", + "image_path": "3d2fdd59f275d91b97284a1af2a57e302901040857a8043d9fc9415d4c50bb93.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 226, + 251, + 245 + ], + "lines": [ + { + "bbox": [ + 216, + 226, + 251, + 245 + ], + "spans": [ + { + "bbox": [ + 216, + 226, + 251, + 245 + ], + "type": "text", + "content": "a young girl wearing glasses and a tie" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 256, + 184, + 294, + 223 + ], + "blocks": [ + { + "bbox": [ + 256, + 184, + 294, + 223 + ], + "lines": [ + { + "bbox": [ + 256, + 184, + 294, + 223 + ], + "spans": [ + { + "bbox": [ + 256, + 184, + 294, + 223 + ], + "type": "image", + "image_path": "1bd6db1a76ba15eaa41e39e3c73392fe25888edf259b42e1eac659314dcd1c64.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 226, + 299, + 245 + ], + "lines": [ + { + "bbox": [ + 257, + 226, + 299, + 245 + ], + "spans": [ + { + "bbox": [ + 257, + 226, + 299, + 245 + ], + "type": "text", + "content": "a group of people riding motorcycles down a street" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 298, + 184, + 335, + 223 + ], + "blocks": [ + { + "bbox": [ + 298, + 184, + 335, + 223 + ], + "lines": [ + { + "bbox": [ + 298, + 184, + 335, + 223 + ], + "spans": [ + { + "bbox": [ + 298, + 184, + 335, + 223 + ], + "type": "image", + "image_path": "94ab0d8f2de445360357fa605bc658cca27c1d0d3c6a926a6dc2612bdac7d41b.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 226, + 337, + 245 + ], + "lines": [ + { + "bbox": [ + 299, + 226, + 337, + 245 + ], + "spans": [ + { + "bbox": [ + 299, + 226, + 337, + 245 + ], + "type": "text", + "content": "a black and white dog laying on top of a toilet" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 339, + 184, + 376, + 223 + ], + "blocks": [ + { + "bbox": [ + 339, + 184, + 376, + 223 + ], + "lines": [ + { + "bbox": [ + 339, + 184, + 376, + 223 + ], + "spans": [ + { + "bbox": [ + 339, + 184, + 376, + 223 + ], + "type": "image", + "image_path": "2907a68e8f28b9598049757ab815de30dda8d019a8beace6e512755557aabd53.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 226, + 380, + 245 + ], + "lines": [ + { + "bbox": [ + 339, + 226, + 380, + 245 + ], + "spans": [ + { + "bbox": [ + 339, + 226, + 380, + 245 + ], + "type": "text", + "content": "a train on a train track with a dog standing next to it" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 380, + 184, + 418, + 223 + ], + "blocks": [ + { + "bbox": [ + 380, + 184, + 418, + 223 + ], + "lines": [ + { + "bbox": [ + 380, + 184, + 418, + 223 + ], + "spans": [ + { + "bbox": [ + 380, + 184, + 418, + 223 + ], + "type": "image", + "image_path": "3ddd831a1f68db9f81c9b47b7a79996bd2b973fab15e046a381c94e80f2f1162.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 225, + 414, + 245 + ], + "lines": [ + { + "bbox": [ + 384, + 225, + 414, + 245 + ], + "spans": [ + { + "bbox": [ + 384, + 225, + 414, + 245 + ], + "type": "text", + "content": "a boy riding a bike next to a white dog" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 421, + 184, + 459, + 223 + ], + "blocks": [ + { + "bbox": [ + 421, + 184, + 459, + 223 + ], + "lines": [ + { + "bbox": [ + 421, + 184, + 459, + 223 + ], + "spans": [ + { + "bbox": [ + 421, + 184, + 459, + 223 + ], + "type": "image", + "image_path": "5c58d90dfdffac9a457e7743617d5af84eec7c9581ab90d1c023a6876fbfb8a1.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 225, + 463, + 245 + ], + "lines": [ + { + "bbox": [ + 421, + 225, + 463, + 245 + ], + "spans": [ + { + "bbox": [ + 421, + 225, + 463, + 245 + ], + "type": "text", + "content": "a small dog is eating a banana on the floor" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "bbox": [ + 130, + 331, + 480, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 480, + 354 + ], + "type": "text", + "content": "under the guidance of the other modality. SGA, our baseline, expands a single image-text pair into a set of images and a set of texts to enhance diversity." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 130, + 354, + 480, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 480, + 485 + ], + "type": "text", + "content": "First, we compare the performance of various methods under white-box attacks, wherein we can leverage the model architecture and interact with the model. It is evident that our method, along with SGA, performs the best in the four white-box attack experiments compared to other methods. Whether TR or IR, the attack success rate at the top-1 rank " + }, + { + "bbox": [ + 130, + 354, + 480, + 485 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}@\\mathbb{1})" + }, + { + "bbox": [ + 130, + 354, + 480, + 485 + ], + "type": "text", + "content": " consistently exceeds " + }, + { + "bbox": [ + 130, + 354, + 480, + 485 + ], + "type": "inline_equation", + "content": "99.9\\%" + }, + { + "bbox": [ + 130, + 354, + 480, + 485 + ], + "type": "text", + "content": ". Given the already high white-box attack success rate achieved by the SGA method, there is limited room for improvement in our approach within this context. Subsequently, we shift our focus to elucidating the enhancements our method brings to transferability, specifically in the realm of black-box attack performance. We delineate this exploration into two segments, contingent on whether the source model and the target model share the same architecture." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "text", + "content": "Cross-Model Transferability in Same Architecture. ALBEF and TCL both utilize a similar model architecture but differ in their pre-training objectives while maintaining a common fundamental model structure. Therefore, when ALBEF and TCL are employed as target models for each other, the success rate of attacks using multimodal adversarial examples is remarkably high. Notably, the black-box attack success rate of IR reaches " + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "inline_equation", + "content": "95.58\\%" + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "text", + "content": " when TCL is the target model in our proposed method. Additionally, the transferability of SGA can reach approximately " + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "text", + "content": ", but the room for improvement is very limited, our method has improved compared to SGA, ranging from " + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "inline_equation", + "content": "2.71\\%" + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "inline_equation", + "content": "3.69\\%" + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "text", + "content": ". In contrast, when considering " + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{ViT}}" + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{CNN}}" + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "text", + "content": "—both being aligned VLP models—their image encoders vary significantly, with one utilizing the Vision Transformer and the other employing ResNet-101. Given the significant structural differences between traditional CNNs and Transformers, existing methods show low cross-model transferability, highlighting room for improvement. Our method outperforms SGA by about " + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "inline_equation", + "content": "7.67\\%" + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "inline_equation", + "content": "11.62\\%" + }, + { + "bbox": [ + 130, + 486, + 481, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "type": "text", + "content": "Boosting Transferability in Vision-Language Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 189, + 465, + 380 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 3: Comparison with state-of-the-art methods on image-text retrieval. The source column shows VLP models we use to generate multimodal adversarial examples. The gray area represents adversarial attacks under a white-box setting, the rest are black-box attacks. For both Image Retrieval and Text Retrieval, we provide R@1 attack success rate(%)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 189, + 465, + 380 + ], + "lines": [ + { + "bbox": [ + 151, + 189, + 465, + 380 + ], + "spans": [ + { + "bbox": [ + 151, + 189, + 465, + 380 + ], + "type": "table", + "html": "
SourceAttackALBEFTCLCLIP_VITCLIP_CNN
TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1TR R@1IR R@1
ALBEFPGD93.7494.4324.0327.910.6715.8214.0519.11
BERT-Attack11.5727.4612.6428.0729.3343.1732.6946.11
Sep-Attack95.7296.1439.351.7934.1145.7235.7647.92
Co-Attack97.0898.3639.5251.2429.8238.9231.2941.99
SGA99.999.9887.8888.0536.6946.7839.5949.78
Ours99.999.9391.5791.1746.2656.849.5559.01
TCLPGD35.7741.6799.3799.3310.1816.314.8121.1
BERT-Attack11.8926.8214.5429.1729.6944.4933.4646.07
Sep-Attack52.4561.4499.5899.4537.0645.8137.4249.91
Co-Attack49.8460.3691.6895.4832.6442.6932.0647.82
SGA92.4992.77100.0100.036.8146.9741.8951.53
Ours95.295.58100.099.9847.2457.2852.2362.23
CLIP_VITPGD3.136.484.438.8369.3384.7913.0317.43
BERT-Attack9.5922.6411.8025.0728.3439.0830.4037.43
Sep-Attack7.6120.5810.1220.7476.9387.4429.8938.32
Co-Attack8.5520.1810.0121.2978.5387.529.538.49
SGA22.4234.5925.0836.45100.0100.053.2661.1
Ours27.8442.8427.8244.6100.0100.064.8869.5
CLIP_CNNPGD2.296.154.538.885.412.0889.7893.04
BERT-Attack8.8623.2712.3325.4827.1237.4430.4040.10
Sep-Attack9.3822.9911.2825.4526.1339.2493.6195.3
Co-Attack10.5323.6212.5426.0527.2440.6295.9196.5
SGA15.6428.618.0233.0739.0251.4599.8799.9
Ours19.534.5921.637.8848.4759.1299.8799.9
", + "image_path": "e6fd463d4d74e9b0c0118626ed42fe1033e8e01a2ec61f2e6aad6d4de79afd66.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 415, + 482, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 522 + ], + "type": "text", + "content": "Cross-Model Transferability in Different Architectures. When fused VLP models (i.e., ALBEF and TCL) are used for generating adversarial examples, our method achieves a significant improvement compared to existing methods, outperforming the current state-of-the-art (SOTA) methods by " + }, + { + "bbox": [ + 130, + 415, + 482, + 522 + ], + "type": "inline_equation", + "content": "9.23\\%" + }, + { + "bbox": [ + 130, + 415, + 482, + 522 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 415, + 482, + 522 + ], + "type": "inline_equation", + "content": "10.7\\%" + }, + { + "bbox": [ + 130, + 415, + 482, + 522 + ], + "type": "text", + "content": ". When the fused VLP models are targeted for attack, our method still outperforms all existing methods when generating adversarial examples against the aligned VLP models (i.e., CLIP). In this scenario, the transferability of all methods is relatively low. Compared to our method's baseline SGA, we attain an improvement ranging from " + }, + { + "bbox": [ + 130, + 415, + 482, + 522 + ], + "type": "inline_equation", + "content": "2.74\\%" + }, + { + "bbox": [ + 130, + 415, + 482, + 522 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 415, + 482, + 522 + ], + "type": "inline_equation", + "content": "8.25\\%" + }, + { + "bbox": [ + 130, + 415, + 482, + 522 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 544, + 293, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 293, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 293, + 558 + ], + "type": "text", + "content": "4.4 Cross-Task Transferability" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "text", + "content": "We not only assess the transferability of multimodal adversarial examples generated by our proposed method across different VLP models but also conduct experiments to evaluate its effectiveness in transferring across diverse " + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{V} + \\mathrm{L}" + }, + { + "bbox": [ + 130, + 569, + 481, + 666 + ], + "type": "text", + "content": " tasks. Specifically, we craft adversarial examples for the Image-Text Retrieval (ITR) task and evaluate them on Visual Grounding (VG) and Image Captioning (IC) tasks. As evident from Table 4 and visual results in Figure 2 and 3, the adversarial examples generated for ITR demonstrate transferability, successfully impacting both VG and IC tasks. This highlights the efficacy of cross-task trans" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "S. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 154, + 119, + 459, + 331 + ], + "blocks": [ + { + "bbox": [ + 154, + 119, + 459, + 331 + ], + "lines": [ + { + "bbox": [ + 154, + 119, + 459, + 331 + ], + "spans": [ + { + "bbox": [ + 154, + 119, + 459, + 331 + ], + "type": "image", + "image_path": "dd50538e90d8f02ac790684bf3ff72dc8c7f817cc7fca1353f6606f5da6073d9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 343, + 482, + 388 + ], + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 388 + ], + "type": "text", + "content": "Fig. 3: Visualization on Visual Grounding. We use the ALBEF model, pre-trained on the ITR task, to generate adversarial images on the RefCOCO+ dataset and use the same model, pre-trained on Visual Grouding(VG) task, to localize the regions corresponding to red words on both clean images and adversarial images, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 411, + 480, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 480, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 480, + 436 + ], + "type": "text", + "content": "ferability in our proposed method. Furthermore, our transferability consistently outperforms that of SGA." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 453, + 238, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 238, + 466 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 238, + 466 + ], + "type": "text", + "content": "4.5 Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 473, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 568 + ], + "type": "text", + "content": "Our proposed method builds upon the SGA [36] as a baseline, introducing two key improvements. Firstly, we utilize diversification along the intersection region of adversarial trajectory to expand the diversity of adversarial examples. Secondly, we generate adversarial text while simultaneously distancing it from the last intersection region along the optimization path. To investigate the impact of each improvement on the effectiveness of our method, we conduct ablation studies on the ITR task and employ transferability from ALBEF to the other three VLP models as the evaluation metric." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "In the ablation study, we systematically eliminate each enhancement from our approach and compare their transferability with both the baseline SGA and our final method, which incorporates a combination of two improvements. The results are depicted in Figure 4. It is evident that when attacking models with the same architecture versus different architectures, the impacts of two distinct improvements are also different. When attacking TCL with the same structure as ALBEF, under Setting1, our method exhibits a greater decrease in transferability, indicating that diversification along the intersection region is" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "type": "text", + "content": "Boosting Transferability in Vision-Language Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 189, + 465, + 245 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": "Table 4: Cross-Task Transferability. We utilize ALBEF to generate multimodal adversarial examples for attacking both Visual Grounding(VG) on the RefCOCO+ dataset and Image Captioning(IC) on the MSCOCO dataset. The baseline represents the performance of each task without any attack, where a lower value indicates better effectiveness of the adversarial attack for both tasks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 189, + 465, + 245 + ], + "lines": [ + { + "bbox": [ + 151, + 189, + 465, + 245 + ], + "spans": [ + { + "bbox": [ + 151, + 189, + 465, + 245 + ], + "type": "table", + "html": "
AttackITR → VGITR → IC
ValTestATestBB@4METEORROUGE-LCIDErSPICE
Baseline58.4665.8946.2539.731.060.0133.323.8
SGA50.5657.4240.6628.024.651.291.417.7
Ours49.7056.3240.5427.224.250.788.317.2
", + "image_path": "dec395851dece59e370bb92a6943a7bf357d431c53e687bc67191bcdfbad9574.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 151, + 266, + 253, + 353 + ], + "blocks": [ + { + "bbox": [ + 151, + 266, + 253, + 353 + ], + "lines": [ + { + "bbox": [ + 151, + 266, + 253, + 353 + ], + "spans": [ + { + "bbox": [ + 151, + 266, + 253, + 353 + ], + "type": "image", + "image_path": "8268410413513309c42bce47d08364b6f9b93b3b4de585e0fb11a9230cc8d25c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 363, + 482, + 407 + ], + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 407 + ], + "type": "text", + "content": "Fig. 4: Ablation Study: Attack Success Rate(%) on other three target models. The baseline is SGA. Setting 1 removes diversification along the intersection region of adversarial trajectory. Setting 2 removes the text deviating from the last intersection region along the optimization path." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 261, + 267, + 358, + 353 + ], + "blocks": [ + { + "bbox": [ + 261, + 267, + 358, + 353 + ], + "lines": [ + { + "bbox": [ + 261, + 267, + 358, + 353 + ], + "spans": [ + { + "bbox": [ + 261, + 267, + 358, + 353 + ], + "type": "image", + "image_path": "ba2b94f80a5264bb12ca3d19cac0ff96a4d84d6598e8a24ad2156afe5adc77aa.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 365, + 267, + 462, + 353 + ], + "blocks": [ + { + "bbox": [ + 365, + 267, + 462, + 353 + ], + "lines": [ + { + "bbox": [ + 365, + 267, + 462, + 353 + ], + "spans": [ + { + "bbox": [ + 365, + 267, + 462, + 353 + ], + "type": "image", + "image_path": "3d1524d5e2bdd08b7b423cf627dcc13674fa14fc8a8f732e289db634576f69b9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 430, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 479 + ], + "type": "text", + "content": "most crucial in our approach under this scenario. However, under Setting2, the decrease in transferability is more pronounced when attacking the aligned VLP model, indicating that at this point, adversarial texts deviating from the last intersection region play a larger role in the effectiveness of our method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 497, + 220, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 497, + 220, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 497, + 220, + 510 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": "In this paper, we conduct a systematic evaluation of existing multimodal attacks regarding transferability. We found that these methods predominantly prioritize diversity around adversarial examples (AEs) during the optimization process, potentially leading to overfitting to the victim model and hindering transferability. To address this issue, we propose diversification along the intersection region of adversarial trajectory to broaden diversity not only around AEs but also towards clean inputs. Moreover, we pioneer an exploration into extending adversarial texts deviating from the intersection region. Through extensive experiments, we demonstrate the effectiveness of our method in enhancing transferability across VLP models and " + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{V} + \\mathrm{L}" + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": " tasks. This work could act as a catalyst for more profound research on the transferability of multimodal AEs, alongside fortifying the adversarial robustness of VLP models." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "S. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 482, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 482, + 216 + ], + "type": "text", + "content": "This research is supported by the National Research Foundation, Singapore, and DSO National Laboratories under the AI Singapore Programme (AISG Award No: AISG2-GC-2023-008), Career Development Fund (CDF) of Agency for Science, Technology and Research (A*STAR) (No.: C233312028), and National Research Foundation, Singapore and Infocomm Media Development Authority under its Trust Tech Funding Initiative (No. DTC-RGC-04)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 236, + 197, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 197, + 249 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 197, + 249 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 264, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 138, + 264, + 480, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 264, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 264, + 480, + 297 + ], + "type": "text", + "content": "1. Abdelfattah, R., Guo, Q., Li, X., Wang, X., Wang, S.: Cdul: Clip-driven unsupervised learning for multi-label image classification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1348-1357 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 299, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 342 + ], + "type": "text", + "content": "2. Anderson, P., Fernando, B., Johnson, M., Gould, S.: Spice: Semantic propositional image caption evaluation. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 382-398. Springer (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 343, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 386 + ], + "type": "text", + "content": "3. Bai, Y., Zeng, Y., Jiang, Y., Wang, Y., Xia, S.T., Guo, W.: Improving query efficiency of black-box adversarial attack. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16. pp. 101-116. Springer (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 388, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 480, + 430 + ], + "type": "text", + "content": "4. Banerjee, S., Lavie, A.: Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In: Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization. pp. 65-72 (2005)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 432, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 432, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 432, + 480, + 475 + ], + "type": "text", + "content": "5. Chen, H., Ding, G., Liu, X., Lin, Z., Liu, J., Han, J.: Imram: Iterative matching with recurrent attention memory for cross-modal image-text retrieval. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12655-12663 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 476, + 480, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 476, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 138, + 476, + 480, + 509 + ], + "type": "text", + "content": "6. Chen, Y.C., Li, L., Yu, L., El Kholy, A., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: Uniter: Universal image-text representation learning. In: European conference on computer vision. pp. 104-120. Springer (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 510, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 480, + 542 + ], + "type": "text", + "content": "7. Cheng, H., Xiao, E., Cao, J., Yang, L., Xu, K., Gu, J., Xu, R.: Typography leads semantic diversifying: Amplifying adversarial transferability across multimodal large language models. arXiv preprint arXiv:2405.20090 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 544, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 586 + ], + "type": "text", + "content": "8. Cheng, M., Sun, Y., Wang, L., Zhu, X., Yao, K., Chen, J., Song, G., Han, J., Liu, J., Ding, E., et al.: Vista: vision and scene text aggregation for cross-modal retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5184-5193 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 588, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 480, + 620 + ], + "type": "text", + "content": "9. Dong, Y., Pang, T., Su, H., Zhu, J.: Evading defenses to transferable adversarial examples by translation-invariant attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4312-4321 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 621, + 480, + 665 + ], + "type": "text", + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arxiv 2020. arXiv preprint arXiv:2010.11929 (2010)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "type": "text", + "content": "Boosting Transferability in Vision-Language Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "11. Dou, Z.Y., Xu, Y., Gan, Z., Wang, J., Wang, S., Wang, L., Zhu, C., Zhang, P., Yuan, L., Peng, N., et al.: An empirical study of training end-to-end vision-and-language transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18166-18176 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 195 + ], + "type": "text", + "content": "12. Gao, K., Bai, Y., Bai, J., Yang, Y., Xia, S.T.: Adversarial robustness for visual grounding of multimodal large language models. arXiv preprint arXiv:2405.09981 (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 196, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 481, + 228 + ], + "type": "text", + "content": "13. Gu, J., Jia, X., de Jorge, P., Yu, W., Liu, X., Ma, A., Xun, Y., Hu, A., Khakzar, A., Li, Z., et al.: A survey on transferability of adversarial examples across deep neural networks. arXiv preprint arXiv:2310.17626 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "text", + "content": "14. Guo, Q., Cheng, Z., Juefei-Xu, F., Ma, L., Xie, X., Liu, Y., Zhao, J.: Learning to adversarially blur visual object tracking. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10839-10848 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "type": "text", + "content": "15. Guo, Q., Juefei-Xu, F., Xie, X., Ma, L., Wang, J., Yu, B., Feng, W., Liu, Y.: Watch out! motion is blurring the vision of your deep neural networks. In: Advances in Neural Information Processing Systems. vol. 33, pp. 975-985 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "type": "text", + "content": "16. Han, D., Jia, X., Bai, Y., Gu, J., Liu, Y., Cao, X.: Ot-attack: Enhancing adversarial transferability of vision-language models via optimal transport optimization. arXiv preprint arXiv:2312.04403 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 331, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 481, + 363 + ], + "type": "text", + "content": "17. He, B., Jia, X., Liang, S., Lou, T., Liu, Y., Cao, X.: Sa-attack: Improving adversarial transferability of vision-language pre-training models via self-augmentation. arXiv preprint arXiv:2312.04913 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 364, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 481, + 397 + ], + "type": "text", + "content": "18. He, B., Liu, J., Li, Y., Liang, S., Li, J., Jia, X., Cao, X.: Generating transferable 3d adversarial point cloud via random perturbation factorization. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 764-772 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 398, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 398, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 398, + 481, + 430 + ], + "type": "text", + "content": "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 431, + 481, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 481, + 474 + ], + "type": "text", + "content": "20. Hu, X., Gan, Z., Wang, J., Yang, Z., Liu, Z., Lu, Y., Wang, L.: Scaling up vision-language pre-training for image captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17980-17989 (June 2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 475, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 520 + ], + "type": "text", + "content": "21. Huang, Y., Guo, Q., Juefei-Xu, F., Ma, L., Miao, W., Liu, Y., Pu, G.: Advfilter: predictive perturbation-aware filtering against adversarial attack via multi-domain learning. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 395-403 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 521, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 481, + 564 + ], + "type": "text", + "content": "22. Huang, Y., Juefei-Xu, F., Guo, Q., Zhang, J., Wu, Y., Hu, M., Li, T., Pu, G., Liu, Y.: Personalization as a shortcut for few-shot backdoor attack against text-to-image diffusion models. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 21169-21178 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "type": "text", + "content": "23. Huang, Y., Sun, L., Guo, Q., Juefei-Xu, F., Zhu, J., Feng, J., Liu, Y., Pu, G.: Ala: Naturalness-aware adversarial lightness attack. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 2418-2426 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "text", + "content": "24. Huang, Y., Sun, L., Guo, Q., Juefei-Xu, F., Zhu, J., Feng, J., Liu, Y., Pu, G.: Ala: Naturalness-aware adversarial lightness attack. In: Proceedings of the 31st ACM International Conference on Multimedia. p. 2418-2426 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "25. Jia, X., Wei, X., Cao, X., Han, X.: Adv-watermark: A novel watermark perturbation for adversarial examples. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 1579-1587 (2020)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 218, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 218, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 218, + 100 + ], + "type": "text", + "content": "S. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "text", + "content": "26. Jia, X., Zhang, Y., Wei, X., Wu, B., Ma, K., Wang, J., Cao Sr, X.: Improving fast adversarial training with prior-guided knowledge. arXiv preprint arXiv:2304.00202 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "27. Khan, Z., Fu, Y.: Exploiting bert for multimodal target sentiment classification through input space translation. In: Proceedings of the 29th ACM international conference on multimedia. pp. 3034-3042 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "type": "text", + "content": "28. Lei, C., Luo, S., Liu, Y., He, W., Wang, J., Wang, G., Tang, H., Miao, C., Li, H.: Understanding chinese video and language via contrastive multimodal pre-training. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 2567-2576 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "text", + "content": "29. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "30. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems 34, 9694-9705 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "text", + "content": "31. Li, L., Ma, R., Guo, Q., Xue, X., Qiu, X.: Bert-attack: Adversarial attack against bert using bert (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 358 + ], + "type": "text", + "content": "32. Li, X., Yin, X., Li, C., Zhang, P., Hu, X., Zhang, L., Wang, L., Hu, H., Dong, L., Wei, F., et al.: Oscar: Object-semantics aligned pre-training for vision-language tasks. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16. pp. 121-137. Springer (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "type": "text", + "content": "33. Lin, C.Y.: Rouge: A package for automatic evaluation of summaries. In: Text summarization branches out. pp. 74-81 (2004)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "text", + "content": "34. Lin, J., Song, C., He, K., Wang, L., Hopcroft, J.E.: Nesterov accelerated gradient and scale invariance for adversarial attacks. arXiv preprint arXiv:1908.06281 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 445 + ], + "type": "text", + "content": "35. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "text", + "content": "36. Lu, D., Wang, Z., Wang, T., Guan, W., Gao, H., Zheng, F.: Set-level guidance attack: Boosting adversarial transferability of vision-language pre-training models (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "37. Luo, H., Gu, J., Liu, F., Torr, P.: An image is worth 1000 lies: Adversarial transferability across prompts on vision-language models. arXiv preprint arXiv:2403.09766 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "text", + "content": "38. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "text", + "content": "39. Naseer, M.M., Ranasinghe, K., Khan, S.H., Hayat, M., Shahbaz Khan, F., Yang, M.H.: Intriguing properties of vision transformers. Advances in Neural Information Processing Systems 34, 23296-23308 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "text", + "content": "40. Park, J., Miller, P., McLaughlin, N.: Hard-label based small query black-box adversarial attack. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3986-3995 (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "type": "text", + "content": "41. Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In: Proceedings of the IEEE international conference on computer vision. pp. 2641-2649 (2015)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "text", + "content": "42. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "type": "text", + "content": "Boosting Transferability in Vision-Language Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 147, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 117, + 480, + 138 + ], + "type": "text", + "content": "natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 183 + ], + "type": "text", + "content": "43. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 184, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 184, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 133, + 184, + 481, + 216 + ], + "type": "text", + "content": "44. Shi, L., Shuang, K., Geng, S., Gao, P., Fu, Z., de Melo, G., Chen, Y., Su, S.: Dense contrastive visual-linguistic pretraining. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 5203-5212 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 217, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 217, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 217, + 481, + 237 + ], + "type": "text", + "content": "45. Tan, H., Bansal, M.: Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 481, + 270 + ], + "type": "text", + "content": "46. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: International conference on machine learning. pp. 10347-10357. PMLR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 271, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 271, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 133, + 271, + 481, + 304 + ], + "type": "text", + "content": "47. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4566-4575 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 304, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 304, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 133, + 304, + 481, + 346 + ], + "type": "text", + "content": "48. Wang, T., Ge, Y., Zheng, F., Cheng, R., Shan, Y., Qie, X., Luo, P.: Accelerating vision-language pretraining with free language modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23161-23170 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 133, + 347, + 481, + 380 + ], + "type": "text", + "content": "49. Wang, T., Jiang, W., Lu, Z., Zheng, F., Cheng, R., Yin, C., Luo, P.: Vlmixer: Unpaired vision-language pre-training via cross-modal cutmix. In: International Conference on Machine Learning. pp. 22680-22690. PMLR (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 381, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 381, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 133, + 381, + 481, + 413 + ], + "type": "text", + "content": "50. Wang, X., He, X., Wang, J., He, K.: Admix: Enhancing the transferability of adversarial attacks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16158-16167 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 414, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 414, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 133, + 414, + 481, + 446 + ], + "type": "text", + "content": "51. Wang, Z., Liu, X., Li, H., Sheng, L., Yan, J., Wang, X., Shao, J.: Camp: Cross-modal adaptive message passing for text-image retrieval. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5764-5773 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 447, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 447, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 133, + 447, + 481, + 489 + ], + "type": "text", + "content": "52. Xie, C., Zhang, Z., Zhou, Y., Bai, S., Wang, J., Ren, Z., Yuille, A.L.: Improving transferability of adversarial examples with input diversity. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2730-2739 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 491, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 491, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 133, + 491, + 481, + 533 + ], + "type": "text", + "content": "53. Yang, J., Duan, J., Tran, S., Xu, Y., Chanda, S., Chen, L., Zeng, B., Chilimbi, T., Huang, J.: Vision-language pre-training with triple contrastive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15671-15680 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 534, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 534, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 133, + 534, + 481, + 577 + ], + "type": "text", + "content": "54. Yu, L., Lin, Z., Shen, X., Yang, J., Lu, X., Bansal, M., Berg, T.L.: Mattnet: Modular attention network for referring expression comprehension. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1307-1315 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 578, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 578, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 133, + 578, + 481, + 621 + ], + "type": "text", + "content": "55. Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in referring expressions. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 69-85. Springer (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 622, + 481, + 665 + ], + "type": "text", + "content": "56. Yuan, L., Chen, Y., Wang, T., Yu, W., Shi, Y., Jiang, Z.H., Tay, F.E., Feng, J., Yan, S.: Tokens-to-token vit: Training vision transformers from scratch on imagenet. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 558-567 (2021)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "S. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 270 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "57. Zhang, J., Yi, Q., Sang, J.: Towards adversarial attack on vision-language pretraining models. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 5005-5013 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "type": "text", + "content": "58. Zhang, J., Huang, J.t., Wang, W., Li, Y., Wu, W., Wang, X., Su, Y., Lyu, M.R.: Improving the transferability of adversarial samples by path-augmented method. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8173-8182 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 237 + ], + "type": "text", + "content": "59. Zhang, P., Li, X., Hu, X., Yang, J., Zhang, L., Wang, L., Choi, Y., Gao, J.: Vinyl: Revisiting visual representations in vision-language models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5579-5588 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "text", + "content": "60. Zhang, Q., Lei, Z., Zhang, Z., Li, S.Z.: Context-aware attention network for image-text retrieval. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3536-3545 (2020)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 233, + 91, + 447, + 102 + ], + "type": "text", + "content": "Boosting Transferability in Vision-Language Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Boosting the Power of Small Multimodal Reasoning Models to Match Larger Models with Self-Consistency Training/cda53ae2-a211-4097-a2a7-2df98969f4ef_content_list.json b/2024/Boosting the Power of Small Multimodal Reasoning Models to Match Larger Models with Self-Consistency Training/cda53ae2-a211-4097-a2a7-2df98969f4ef_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bdfa91202992ae0d21356b14fee10a9e610ba49c --- /dev/null +++ b/2024/Boosting the Power of Small Multimodal Reasoning Models to Match Larger Models with Self-Consistency Training/cda53ae2-a211-4097-a2a7-2df98969f4ef_content_list.json @@ -0,0 +1,1922 @@ +[ + { + "type": "text", + "text": "Boosting the Power of Small Multimodal Reasoning Models to Match Larger Models with Self-Consistency Training", + "text_level": 1, + "bbox": [ + 225, + 140, + 779, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cheng Tan $^{1,4,5\\star}$ , Jingxuan Wei $^{2,3\\star\\boxtimes}$ , Zhangyang Gao $^{1,4,5\\star}$ , Linzhuang Sun $^{2,3}$ , Siyuan Li $^{1,4,5}$ , Ruifeng Guo $^{2,3}$ , Bihui Yu $^{2,3}$ , and Stan Z. Li $^{4\\boxtimes}$", + "bbox": [ + 223, + 234, + 781, + 267 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ Zhejiang University \n $^{2}$ Shenyang Institute of Computing Technology, Chinese Academy of Sciences \n $^{3}$ University of Chinese Academy of Sciences", + "$^{4}$ AI Lab, Research Center for Industries of the Future, Westlake University", + "$^{5}$ Institute of Advanced Technology, Westlake Institute for Advanced Study tancheng@westlake.edu.cn, weijingxuan20@mails.ucas.edu.cn $\\star$ Equal contribution, $\\boxtimes$ Corresponding author" + ], + "bbox": [ + 240, + 276, + 759, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Multimodal reasoning is a challenging task that requires models to reason across multiple modalities to answer questions. Existing approaches have made progress by incorporating language and visual modalities into a two-stage reasoning framework, separating rationale generation from answer inference. However, these approaches often fall short due to the inadequate quality of the generated rationales. In this work, we delve into the importance of rationales in model reasoning. We observe that when rationales are completely accurate, the model's accuracy significantly improves, highlighting the need for high-quality rationale generation. Motivated by this, we propose MC-CoT, a self-consistency training strategy that generates multiple rationales and answers, subsequently selecting the most accurate through a voting process. This approach not only enhances the quality of generated rationales but also leads to more accurate and robust answers. Through extensive experiments, we demonstrate that our approach significantly improves model performance across various benchmarks. Remarkably, we show that even smaller base models, when equipped with our proposed approach, can achieve results comparable to those of larger models, illustrating the potential of our approach in harnessing the power of rationales for improved multimodal reasoning. The code is available at github.com/chengtan9907/mc-cot.", + "bbox": [ + 259, + 414, + 743, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Multimodal reasoning $\\cdot$ visual question answering", + "bbox": [ + 261, + 705, + 683, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 746, + 375, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in large language models [2, 4, 8, 13, 14, 39, 40, 42, 47, 48, 62] have led to the exploration of Chain-of-Thought (CoT) prompting [6, 50, 51, 56]. This approach, which directs the model to systematically unravel rationales before providing answers, rather than responding directly, has showcased the model's", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a89124767fb747bb36370ebf0917cd03d167d845edf0d2729bf1601e7253aac0.jpg", + "image_caption": [ + "Fig. 1: An example of multimodal reasoning that answers the question by reasoning across both vision and language modalities." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 547, + 250 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ba5d01f034521d1a765437c9b1c8a328c8b2f2b979ba3ba88f8c0f26e35988e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 143, + 784, + 250 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "impressive efficacy across a variety of natural language processing (NLP) tasks. Moreover, the advent of CoT prompting has catalyzed a plethora of research endeavors delving into the reasoning prowess of large language models. A diverse range of Chain-of-Thought strategies has been investigated, including the voting-facilitated CoT-SC [50], the transition from chain-like to tree-like thinking with Tree-of-Thoughts [56], and further expansion into graph-structured thinking [6].", + "bbox": [ + 212, + 328, + 787, + 420 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While CoT reasoning has been thoroughly established in the realm of language models, its foray into the vast and intricate landscape of multimodal reasoning is still in its infancy. As shown in Fig. 1, Multimodal reasoning [5, 16, 17, 21, 29, 30, 32, 33, 36, 53, 57, 60], which inherently involves the seamless fusion of information from disparate modalities such as text and images, presents unique challenges. The process of extracting, correlating, and generating rationales across multiple modalities is decidedly more complex than the tasks encountered in a solely text-based modality. A recent seminal work, Multimodal-CoT [63], has pioneered the application of the Chain-of-Thought prompting strategy to multimodal reasoning tasks. This approach encompasses a two-stage framework that distinctively separates rationale generation from answer inference. By obliging the model to generate rationales prior to answering questions, Multimodal-CoT mirrors the language-only CoT prompting, thus enabling reasoning across multiple modalities.", + "bbox": [ + 212, + 422, + 789, + 621 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite Multimodal-CoT has made promising strides in the realm of multimodal reasoning, as evidenced in Fig. 2, its improvements over the no-rationale baseline are still limited. Moreover, compared to the ground-truth rationale, the predicted rationale falls short, often yielding results that lack relevance to the posed question. This discrepancy primarily stems from the quality of the generated rationales, highlighting the crucial role of rationale quality in the success of the chain-of-thought reasoning process.", + "bbox": [ + 212, + 623, + 787, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In light of the above observations, it becomes evident that the potency of the Chain-of-Thought reasoning in a multimodal scenario is intrinsically tethered to the accuracy of the rationales generated. A high-quality rationale not only sheds light on the model's thought process but also paves the way for more precise answers. This leads to a critical question: how can we develop a strategy to enhance the quality of these rationales, and consequently, improve the overall performance of multimodal CoT?", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Tan. et al.", + "bbox": [ + 271, + 114, + 341, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/35433dbd48d7b8159bd6cdb1801e57e751418a77099e9d857d7c8f6c54d21bfd.jpg", + "image_caption": [ + "Fig. 2: The comparison of answer accuracy on ScienceQA using the Multimodal-CoT framework with no rationale, predicted rationales, and ground-truth rationales." + ], + "image_footnote": [], + "bbox": [ + 305, + 138, + 697, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our study is driven by the hypothesis that enhancing the quality of rationale generation can significantly improve the model's reasoning capabilities and overall performance. To this end, we introduce a simple yet effective strategy that capitalizes on the inherent variability of deep neural models during training, particularly stemming from mechanisms like dropout. Our approach involves having the model generate multiple rationales and then voting for the most consistent words across these rationales to yield a more refined and accurate rationale. The same voting mechanism is then applied to answer generation, further boosting the model's confidence in its predictions. It is important to note that the inference phase remains completely unaffected by this voting mechanism and continues to operate in the same manner as it does in the original Multimodal CoT framework. Through this approach, we aim to facilitate a more robust and accurate multimodal reasoning ability.", + "bbox": [ + 212, + 396, + 787, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Extensive experiments across ScienceQA [32] and A-OKVQA [45] benchmark datasets demonstrate the efficacy of our proposed approach. Notably, by improving the rationale quality, even smaller models equipped with our strategy manifest performance metrics that rival, and at times surpass, those of considerably larger models. This not only confirms the efficacy of our rationale refinement strategy but also opens up a promising avenue where smaller, more efficient models can be rendered competitive in the multimodal reasoning landscape.", + "bbox": [ + 212, + 593, + 787, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 722, + 387, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Chain-of-Thought", + "text_level": 1, + "bbox": [ + 214, + 753, + 408, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Chain-of-Thought (CoT) paradigm has emerged as a transformative approach, aiming to elucidate the reasoning processes of large language models and boost their ability in a variety of NLP tasks. The vanilla CoT [51] prompting strategy models to generate intermediate reasoning steps, leading to significant", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "MC-CoT", + "bbox": [ + 668, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "performance improvements across a spectrum of tasks. Its methodology laid the groundwork for subsequent advancements in the CoT paradigm. CoT-SC [50] introduced a self-consistency decoding strategy by sampling multiple reasoning paths and selecting the most consistent path to enhance the reliability of the generated rationales.", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The versatility and adaptability of the CoT paradigm were further demonstrated with the advent of advanced reasoning structures. Tree-of-Thoughts (ToT) [56] transitioned from linear chains to more intricate tree-like structures, aiming to capture more complex reasoning patterns. In a parallel development, [6] ventured into graph-based reasoning, presenting the Graph-of-Thought (GoT) approach, which allowed for a more interconnected reasoning process. Skeleton-of-Thought (SoT) [38] took a different route by emphasizing the benefits of efficiency by underscoring the potential of parallel decoding, enabling models to explore multiple reasoning paths simultaneously.", + "bbox": [ + 212, + 220, + 789, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the aforementioned studies have made significant strides in the CoT paradigm, we seek to extend these advancements into the realm of multimodal reasoning. Multimodal-CoT [63] is a seminal work, but the unsatisfying performance of its generated rationales has been a major bottleneck. Our work aims to fully explore the potential of the CoT paradigm in multimodal reasoning.", + "bbox": [ + 212, + 357, + 789, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Multimodal Visual Question Answering", + "text_level": 1, + "bbox": [ + 214, + 454, + 589, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multimodal Visual Question Answering (VQA) has emerged as a key research area bridging the domains of vision and language. Pioneering works like MCAN [58], BAN [25], and Top-Down [1] have established a foundation by introducing diverse attention mechanisms. Building on this, DFAF [18] introduced innovative dynamic fusion strategies to facilitate both intra- and inter-modality attention flows. The introduction of Transformer architectures has significantly impacted the VQA field, as demonstrated by state-of-the-art models such as ViLT [26] and VisualBERT [28]. In the same vein, Patch-TRM [35] leveraged a pyramid cross-modal Transformer, paired with pre-trained input embeddings from the icon dataset, advancing the understanding of abstract diagrams. Recently, incorporating large language models into the multimodal framework has seen a surge, exemplified by models like LLaMA-Adapter [60] and BLIP-2 [27]. LaVIN [36] introduced a novel routing algorithm that helps the model autonomously switch between unimodal and multimodal instruction reasoning paths. These models employ frozen visual encoders paired with fine-tuned language models, thereby setting new standards in VQA tasks.", + "bbox": [ + 212, + 477, + 789, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ScienceQA [32] introduced a reasoning process into the VQA task, setting a benchmark for multimodal chain-of-thought reasoning. MM-REACT [55], with its innovative prompting design, enables language models to process multimodal information, facilitating the integration of ChatGPT with visual experts. A-OKVQA [45] requires commonsense reasoning about the depicted scene in an image. Multimodal-CoT [63] integrates language (text) and vision (image) modalities into a two-stage framework, distinctly separating rationale generation from answer inference.", + "bbox": [ + 212, + 719, + 789, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Tan. et al.", + "bbox": [ + 271, + 114, + 343, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we typically focus on multimodal reasoning tasks, a subtype of multimodal visual question answering (VQA) that incorporates rationales. The significance of this task lies in its ability to more accurately simulate human cognitive processes. When humans answer questions, they often rely on observed visual information and existing textual knowledge, synthesizing the two to form a reasoned conclusion. These multimodal reasoning tasks comprehensively assess a model's capacity to understand and fuse visual and textual information, thereby propelling the development of multimodal learning. Moreover, these tasks enhance the interpretability of artificial intelligence systems. Because models must supply not only answers but also supporting rationales, their decision-making processes become more interpretable and comprehensible.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 332, + 330, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 359, + 372, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this work, we focus on multimodal reasoning, a challenging task that involves reasoning across multiple modalities to answer questions. Specifically, we consider visual question answering [3, 52], where the model is tasked with answering questions based on the information provided in the question itself and the accompanying image.", + "bbox": [ + 212, + 381, + 787, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Formally, given a dataset $\\mathcal{D} = \\{\\mathcal{X},\\mathcal{Y}\\}$ , where $X\\in \\mathcal{X}$ represents a multimodal input consisting of text and images, and $Y\\in \\mathcal{V}$ represents the corresponding output, the task of multimodal reasoning involves learning a mapping function parameterized by $\\Theta$ : $\\mathcal{F}_{\\Theta}:\\mathcal{X}\\to \\mathcal{V}$ that accurately predicts the output $Y$ for a given input $X$ . The multimodal input $X$ can be represented as $X = (T,I)$ , where $T$ is the text component, and $I$ is the image component of the input. The function $\\mathcal{F}$ should be able to effectively leverage the information from both the text and image modalities to make accurate predictions. Specifically, the basic visual question answering task is defined as:", + "bbox": [ + 212, + 457, + 787, + 592 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nY = \\arg \\max _ {Y ^ {\\prime}} p \\left(Y ^ {\\prime} \\mid T, I\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 599, + 784, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p(Y' \\mid T, I)$ is the probability of answer $Y'$ given the text $T$ and image $I$ . The multimodal reasoning task builds upon the basic visual question answering task, extending its requirements to necessitate the generation of a rationale $R'$ that elucidates the reasoning process underpinning the answer $Y$ . This task can be mathematically represented as follows:", + "bbox": [ + 212, + 627, + 787, + 703 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nY, R = \\arg \\max _ {Y ^ {\\prime}, R ^ {\\prime}} p \\left(Y ^ {\\prime}, R ^ {\\prime} \\mid T, I\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 710, + 784, + 732 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multimodal-CoT [63] decomposed the multimodal reasoning task into a two-stage framework, where the model first generates a rationale $R$ and then utilizes this rationale to predict the answer $Y$ . The process is delineated as:", + "bbox": [ + 212, + 744, + 787, + 789 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR = \\arg \\max _ {R ^ {\\prime}} p \\left(R ^ {\\prime} \\mid T, I\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 797, + 784, + 825 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nY = \\arg \\max _ {Y ^ {\\prime}} p (Y ^ {\\prime} | R, T, I)\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 821, + 599, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "MC-CoT", + "bbox": [ + 668, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8aa890f28b192b53fa4f8f9d2164f950d793374a17edcc096e4a0f6f3c80ec72.jpg", + "image_caption": [ + "(a) Vanilla Prompt" + ], + "image_footnote": [], + "bbox": [ + 222, + 142, + 302, + 273 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6217550fc74f3491f094618f9361a7e9db0a7b81e3b730ab9512799130caf1e2.jpg", + "image_caption": [ + "(b) CoT" + ], + "image_footnote": [], + "bbox": [ + 320, + 142, + 400, + 273 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/874a7e6b31956ba881aab82b289a377849e49776a704aefda476d44a02339608.jpg", + "image_caption": [ + "(c) CoT-SC" + ], + "image_footnote": [], + "bbox": [ + 411, + 142, + 506, + 273 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2f903064a751c15bfbf661e3b7f428015888d35a55209d180173ccf9b8946d61.jpg", + "image_caption": [ + "(d) Multimodal-CoT", + "Fig. 3: A comparison schematic diagram of different Chain-of-Thought (CoT) prompt-based reasoning methods. (a) The basic input-output prompt, (b) Chain-of-Thought with intermediate chain-like reasoning, (c) Chain-of-Thought Self-Consistency (CoT-SC) that utilizes the consistency of multiple independent chains of thoughts for reasoning, (d) Multimodal-CoT, which infers the rationale using the input text and image, and then predicts the answer using the rationale as part of the input, and (e) MC-CoT that infers a high-quality rationale through word-level voting, and then obtains a high-quality answer using majority vote. It is worth noting that our approach leverages multiple chain consistency only during the training phase, in contrast to CoT-SC, which employs it during the inference stage." + ], + "image_footnote": [], + "bbox": [ + 545, + 143, + 638, + 273 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cdda3e35ff217268fa2b311aa2590b3e37e0ebab713a38208cf1fb47e2cde73a.jpg", + "image_caption": [ + "(e) MC-CoT" + ], + "image_footnote": [], + "bbox": [ + 661, + 143, + 782, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this work, we follow the same two-stage framework as Multimodal-CoT [63] and explore the potential of such a chain-of-thought reasoning paradigm.", + "bbox": [ + 212, + 474, + 787, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Multimodal Consistent Chain-of-Thought", + "text_level": 1, + "bbox": [ + 214, + 530, + 602, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Though the existing multimodal CoT framework has been proven to be beneficial in multimodal reasoning, there remains a large room to fully explore its potential. In this work, we introduce a simple yet effective training strategy, Multimodal Consistent Chain-of-Thought (MC-CoT), to learn high-quality rationales and thus boost the reasoning performance of the multimodal CoT framework.", + "bbox": [ + 212, + 556, + 787, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Rationale Generation Leveraging the inherent randomness introduced by the dropout operations, the model is able to generate multiple diverse rationales for a given input text-image pair $(T,I)$ by sampling from the model $N_{r}$ times:", + "bbox": [ + 212, + 651, + 787, + 696 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nR ^ {i} \\sim p (R \\mid T, I), i = 1, 2, \\dots , N _ {r} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 379, + 709, + 785, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We perform a voting process on the generated rationales to select the most consistent words across the rationales. Specifically, the best rationale $R^{*}$ is selected as follows:", + "bbox": [ + 212, + 738, + 787, + 782 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nR _ {j} ^ {*} = \\operatorname {V o t e} \\left(\\left\\{R _ {j} ^ {i} \\right\\} _ {1} ^ {N _ {r}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 782, + 785, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where, for each position $j$ in the rationale, we choose the majority of the words at position $j$ across the $N_r$ generated rationales to form the best rationale $R_j^*$ .", + "bbox": [ + 212, + 809, + 785, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Tan. et al.", + "bbox": [ + 271, + 114, + 343, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Answer Inference In the answer inference stage, we also generate multiple answers based on the best rationale $R^{*}$ and the input text-image pair $(T, I)$ by sampling from the model $N_{a}$ times:", + "bbox": [ + 214, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nY ^ {i} \\sim p (Y \\mid R ^ {*}, T, I), i = 1, 2, \\dots , N _ {a}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 199, + 784, + 215 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The final answer is selected by the majority voting:", + "bbox": [ + 214, + 220, + 584, + 237 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nY ^ {*} = \\operatorname {V o t e} \\left(\\left\\{Y ^ {i} \\right\\} _ {1} ^ {N _ {a}}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 428, + 243, + 784, + 260 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the majority of the answers across the $N_{a}$ generated answers are regarded as the most convincing answers $Y^{*}$ . The pseudocode of the pipeline is shown in Appendix.", + "bbox": [ + 212, + 267, + 782, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Voting Strategy Consider the answer inference stage as an illustration. The vanilla Multimodal-CoT model computes a cross-entropy loss between a $d$ -dimensional logit $L \\in \\mathbb{R}^d$ and the ground truth label $Y \\in \\mathbb{R}^d$ , which can be represented as:", + "bbox": [ + 214, + 321, + 782, + 381 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\operatorname {C r o s s E n t r o p y} (L, Y), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 383, + 784, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our averaging-based voting strategy initially calculates the mean logit $\\bar{L} \\in \\mathbb{R}^d$ and weighted mean logit $\\hat{L} \\in \\mathbb{R}^d$ :", + "bbox": [ + 214, + 402, + 782, + 433 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {L} = \\frac {1}{N _ {a}} \\sum_ {i = 1} ^ {N _ {a}} L ^ {i}, \\hat {L} = \\frac {1}{N _ {a}} \\frac {\\sum_ {i = 1} ^ {N _ {a}} w L ^ {i}}{\\sum_ {j} w _ {j}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 441, + 784, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $w = 1 / (1 + \\sigma) \\in \\mathbb{R}^d$ is a weight vector derived from the standard deviation of the logits, calculated across different predictions for each dimension. The weight for each logit's dimension is inversely proportional to its variability, assigning higher confidence to predictions with lower variability and lower confidence to those with higher variability. The final prediction is then a linear combination of $\\bar{L}$ and $\\hat{L}$ :", + "bbox": [ + 214, + 487, + 782, + 575 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL ^ {*} = \\alpha \\bar {L} + (1 - \\alpha) \\hat {L} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 429, + 579, + 782, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with $\\alpha$ being a hyperparameter set empirically to 0.5. The cross-entropy loss is subsequently computed between $L^{*}$ and the ground truth label $Y$ :", + "bbox": [ + 214, + 599, + 782, + 628 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\operatorname {C r o s s E n t r o p y} \\left(L ^ {*}, Y\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 637, + 782, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present a comparison of our approach with existing CoT-based reasoning methods in Fig. 3. Specifically, subfigures (a-c) illustrate the prompt learning methods commonly employed for text modality in large language models, while subfigures (d-e) depict the CoT reasoning frameworks in the multimodal field. It can be seen that our approach shares similarities with CoT-SC, such as the utilization of a voting mechanism. However, there are also significant differences between the two approaches: (i) the objective differs. CoT-SC aims to facilitate prompt learning in large language models during the inference stage, while our method seeks to enhance the robustness of model reasoning during the training phase; (ii) the voting process differs. CoT-SC only votes on the final answers inferred from each thought, whereas our method incorporates voting during the rationale generation stage.", + "bbox": [ + 214, + 657, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "MC-CoT", + "bbox": [ + 668, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Theoretical Insights", + "text_level": 1, + "bbox": [ + 215, + 143, + 444, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Aggregation Minimizes Expected Loss", + "text_level": 1, + "bbox": [ + 215, + 184, + 578, + 199 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Theorem 1. (Jensen's inequality) Let $X$ be a random variable and let $\\phi$ be a convex function. Then the following inequality holds: $\\phi(\\mathbb{E}[X]) \\leq \\mathbb{E}[\\phi(X)]$ .", + "bbox": [ + 212, + 215, + 782, + 246 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our proposed method, MC-CoT, leverages the variability introduced by dropout techniques to generate a diverse set of rationales. By aggregating these rationales, we aim to produce explanations and corresponding answers of higher quality. In the context of our framework, the convex function $\\phi$ is represented by the cross-entropy loss function, and the random variable in question is the set of logits $L^i$ . By applying Jensen's inequality to our model, we deduce that the aggregation process leads to a lower expected loss: $\\mathcal{L}(\\mathbb{E}(L^{*}),Y)\\leq \\mathbb{E}(\\mathcal{L}(L^{*},Y))$ .", + "bbox": [ + 212, + 248, + 787, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Bias-Variance Trade-off", + "text_level": 1, + "bbox": [ + 215, + 383, + 457, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce a linear combination of two voting strategies to derive the final prediction, as delineated in Equation 10. We can analytically decompose the expected value of the squared difference between the predicted outcomes $Y^{*}$ and the ground truth $Y$ into three distinct components:", + "bbox": [ + 212, + 412, + 784, + 474 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} \\left[ \\left(Y ^ {*} - Y\\right) ^ {2} \\right] = \\operatorname {B i a s} ^ {2} \\left(Y ^ {*}\\right) + \\operatorname {V a r} \\left(Y ^ {*}\\right) + \\epsilon , \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 488, + 784, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathrm{Bias}^2 (Y^*) = (\\mathbb{E}[Y^*] - Y)^2$ quantifies the systematic deviation of the expected prediction from the ground truth, indicative of the error inherently introduced by the predictive model itself. Meanwhile, $\\mathrm{Var}(Y^{*}) = \\mathbb{E}[(Y^{*} - \\mathbb{E}[Y^{*}])^{2}]$ captures the variability of the predictions, reflecting the extent to which these predictions will fluctuate around their expected value. Lastly, $\\epsilon$ denotes the irreducible error, encapsulating the intrinsic noise within the data.", + "bbox": [ + 212, + 518, + 782, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The mean logits, denoted by $\\bar{L}$ , may exhibit lower bias since they represent the composite prediction averaged across multiple models. Employing solely the mean logits promotes exploration by treating all predictions as equally informative. Nevertheless, its measure of central tendency does not account for the individual variability within the predictions. The weighted mean logits, denoted by $\\hat{L}$ , aim to reduce variability by assigning greater significance to consistent (low-variance) predictions. This reflects a confidence-based strategy, allocating higher weight to predictions that are less variable. However, it could introduce bias if predictions are consistently erroneous.", + "bbox": [ + 212, + 611, + 787, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The final logits, which include both mean logits and weighted mean logits, are designed to achieve an optimal balance between bias and variance. While the mean logits promote exploration by treating all predictions equally, the weighted mean logits perform exploitation by assigning higher weights to consistent predictions. The combination allows balancing exploration with exploitation, thus improving the performance towards an optimal standard.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Tan. et al.", + "bbox": [ + 271, + 114, + 343, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Experiment", + "text_level": 1, + "bbox": [ + 215, + 143, + 364, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets We evaluated MC-CoT on two prominent multimodal reasoning benchmarks. The first, ScienceQA [32], is a comprehensive multimodal science question dataset that contains annotated answers with detailed reasoning and explanations. It encompasses more than 21,000 multiple-choice questions, showcasing a vast range of domains that span three subjects, include 26 topics, cover 127 categories, and incorporate 379 distinct skills. The benchmark dataset is divided into training, validation, and test sets, consisting of 12,726, 4,241, and 4,241 examples, respectively. The second dataset is A-OKVQA [45], a knowledge-based visual question answering dataset that contains 25,000 questions requiring a broad base of commonsense and world knowledge for accurate responses, divided into 17,000 training, 1,000 validation, and 7,000 test examples.", + "bbox": [ + 212, + 176, + 787, + 342 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baselines In our assessment of the ScienceQA dataset, we have compared MC-CoT against recent strong baselines, covering five categories of methodologies: (i) heuristic and expert-guided choices, such as random choice and human evaluation [32]; (ii) standard multimodal visual question answering approaches, which include MCAN [58], Top-Down [1], BAN [25], DFAF [18], ViLT [26], PatchTRM [35], and VisualBERT [28]; (iii) Instruction-tuned large language models like GPT-3.5 [9] and its CoT-enhanced variants [32], in addition to ChatGPT, GPT-4 [39], and Chameleon [34]; (iv) Specifically finetuned large language models, notably LLaMA-Adapter [61], LaVIN [36], LLaMA-SciTune [20], and LLaVa [29]; (v) multimodal reasoning models based on chain-of-thought, Multimodal-CoT [63] and our MC-CoT both belong to this category. Regarding the A-OKVQA dataset, the baselines encompass key methods that have propelled the field of visual question answering forward, such as Pythia [7], ViLBERT [31], LXMERT [46], KRISP [37], GPV-2 [22], BLIP-2 [27], PICa [54], and IPVR [12]. Additionally, Multimodal-CoT [63] is incorporated as the chain-of-thought multimodal benchmark for this dataset.", + "bbox": [ + 212, + 356, + 789, + 597 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details Following the experimental setup detailed in [63], we implemented the T5 encoder-decoder framework [43]. We utilize the UnifiedQA [24] models, which have 223M and 738M parameters, as our default Base and Large models, respectively. Additionally, we employ the FLAN-T5 [15] models with 248M and 783M parameters as our F-Base and F-Large models. Batch sizes were set to 16 for the base model and 8 for the large model. The learning rate was set to $5 \\times 10^{-5}$ . Regarding sequence length, the initial output was constrained to 512 tokens, while the subsequent output was limited to 64 tokens. The experiments utilized four NVIDIA Tesla A100 GPUs, each equipped with 80GB of memory.", + "bbox": [ + 212, + 612, + 787, + 750 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Main Results", + "text_level": 1, + "bbox": [ + 215, + 768, + 372, + 784 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conducted a comparison of MC-CoT with state-of-the-art approaches on the ScienceQA dataset, as depicted in Table 1. We present the overall model size rather than the tunable model size, as it more accurately reflects the model's", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "MC-CoT", + "bbox": [ + 668, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/e11b0a29994f6b00d637c253843cc2dd63ba3284bbd3e8d48f1801968f147eea.jpg", + "table_caption": [ + "Table 1: The comparison on ScienceQA dataset. Question classes: NAT = natural science, SOC = social science, LAN = language science, TXT = text context, IMG = image context, NO = no context, G1-6 = grades 1-6, G7-12 = grades 7-12." + ], + "table_footnote": [], + "table_body": "
ModelSizeNATSOCLANTXTIMGNOG1-6G7-12AVG
Random Choice [32]-40.2846.1329.2547.7540.0833.6639.3540.6739.83
Human [32]-90.2384.9787.4889.6087.5088.1091.5982.4288.40
MCAN [58]95M56.0846.2358.0959.4351.1755.4051.6559.7254.54
Top-Down [1]70M59.5054.3361.8262.9054.8859.7957.2762.1659.02
BAN [25]112M60.8846.5766.6462.6152.6065.5156.8363.9459.37
DFAF [18]74M64.0348.8263.5565.8854.4964.1157.1267.1760.72
ViLT [26]113M60.4863.8960.2763.2061.3857.0060.7261.9061.14
Patch-TRM [35]90M65.1946.7965.5566.9655.2864.9558.0467.5061.42
VisualBERT [28]111M59.3369.1861.1862.7162.1758.5462.9659.9261.87
UnifiedQABase [24]223M68.1669.1874.9163.7861.3877.8472.9865.0070.12
UnifiedQABase w/ CoT [32]223M71.0076.0478.9166.4266.5381.8177.0668.8274.11
GPT-3.5 [32]173B74.6469.7476.0074.4467.2877.4276.8068.8973.97
GPT-3.5 w/ CoT [32]173B75.4470.8778.0974.6867.4379.9378.2369.6875.17
ChatGPT w/ CoT [39]-78.8270.9883.1877.3767.9286.1380.7274.0378.31
GPT-4 w/ CoT [39]-85.4872.4490.2782.6571.4992.8986.6679.0483.99
Chameleon + ChatGPT [34]-81.6270.6484.0079.7770.8086.6281.8676.5379.93
Chameleon + GPT-4 [34]-89.8374.1389.8288.2777.6492.1388.0383.7286.54
LLaMA-Adapter (T) [61]6B79.0073.7980.5578.3070.3583.1479.7775.6878.31
LLaMA-Adapter [61]6B84.3788.3084.3683.7280.3286.9085.8384.0585.19
LaVIN-7B [36]7B89.2594.9485.2488.5187.4688.0890.1688.0789.41
LLaMA-SciTuneBase [20]7B84.5094.1582.9188.3583.6488.7485.0585.6086.11
LaVIN-13B [36]13B90.3294.3887.7389.4487.6590.3191.1989.2690.50
LLaVa [29]13B90.3695.9588.0089.4988.0090.6690.9390.9090.92
LLaVa + GPT-4 [29]13B91.5696.7491.0990.6288.9993.5292.7392.1692.53
LLaMA-SciTuneLarge [20]13B89.3095.6187.0093.0886.6791.7584.3791.3090.03
Mutimodal-CoTBase [63]223M87.5277.1785.8287.8882.9086.8384.6585.3784.91
Mutimodal-CoTLarge [63]738M95.9182.0090.8295.2688.8092.8992.4490.3191.68
MC-CoTBase223M91.8784.5993.0092.2888.3092.7590.6490.6490.64
MC-CoTLarge738M95.4789.9991.8295.1192.6693.2494.2791.7693.37
MC-CoTf-Base248M93.5683.5890.7394.1389.2490.9490.9390.3890.73
MC-CoTf-Large783M97.4790.4493.1896.9793.7594.4995.3094.1394.88
", + "bbox": [ + 220, + 198, + 784, + 646 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "capacity. The results indicate that our MC-CoTBase model, with a model size of only 223 million parameters, achieved an average accuracy of $90.64\\%$ , approaching the performance of Lavin-13B and LLaVa-13B models that were fine-tuned based on larger language models. Furthermore, our MC-CoTF-Large model, with 783 million parameters, achieved state-of-the-art results, surpassing the strongest fine-tuned large language model baseline LLaVa+GPT-4 by $2.35\\%$ .", + "bbox": [ + 212, + 683, + 787, + 775 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Furthermore, in comparison to multimodal chain-of-thought baselines, our MC-CoTBase model showcased an average accuracy improvement of $5.70\\%$ compared to Multimodal-CoTBase. Similarly, our MC-CoTLarge model exhibited an improvement of $1.69\\%$ compared to the Multimodal-CoT Large model. These", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Tan. et al.", + "bbox": [ + 271, + 114, + 341, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "observations demonstrate that our approach has the potential to significantly enhance the model's reasoning capabilities.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We report the experimental results on the A-OKVQA dataset within Table 2. Our MC-CoTBase model demonstrates exceptional improvements over the current state-of-the-art models, both in terms of direct-answer accuracy and multi-choice task performance. Specifically, MC-CoTBase achieves a direct-answer accuracy of $68.7\\%$ , substantially outperforming the strongest competitor, BLIP-2, which possesses an extensive parameter count of over 11 billion. This represents an improvement of over $15.5\\%$ , a margin that underscores the efficiency and effectiveness of our model's reasoning capabilities. In the multi-choice task, MC-CoTBase reaches an accuracy of $71.0\\%$ , which not only exceeds BLIP-2's performance by $0.8\\%$ but also does so with significantly fewer parameters. This comparison highlights the fact that our model's parameter utilization is exceptionally efficient, leading to enhanced performance even against models with far greater complexity.", + "bbox": [ + 212, + 180, + 787, + 363 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/4940c69a24bcaeaf95c52ddfe9054209073352de05d1ca45a0b55698a8c6d7df.jpg", + "table_caption": [ + "Table 2: The comparison on A-OKVQA dataset. We conduct the evaluation on both direct-answer and multi-choice tasks." + ], + "table_footnote": [], + "table_body": "
ModelVision ModelText ModelParametersDirect-answerMulti-choice
Pythia [7]ResNet [19]BERT [23]70M25.249.0
ViLBERT [31]Faster R-CNN [44]BERT [23]300M30.649.1
LXMERT [46]Transformer [49]Transformer [49]220M30.751.4
KRISP [37]Faster R-CNN [44]BERT [23]200M33.751.9
GPV-2 [22]VinVL [59]T5-Base [42]300M48.660.3
BLIP-2 [27]CLIP-VIT-LARGE [41]FlanT5XXL [15]11B53.270.2
PaLM-CoT [51]-PaLM [13]540B41.548.1
PICa [54]VinVL [59]GPT-3 [10]175B42.446.1
IPVR [12]Faster-RCNN [44]OPT [62]66B46.448.6
Multimodal-CoTBase [63]DETR [11]UnifiedQABase [24]223M-50.6
MC-CoTBaseDETR [11]UnifiedQABase [24]223M68.771.0
", + "bbox": [ + 220, + 435, + 787, + 594 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "These results are particularly noteworthy given the complexity and diversity of the A-OKVQA dataset, which is designed to challenge models with questions that require advanced understanding and reasoning across both textual and visual modalities. The performance leap made by $\\mathrm{MC - CoT_{Base}}$ reinforces the potential of our improved training strategy. By introducing a voting mechanism among multiple rationales, our approach refines the quality of the generated rationales, which in turn significantly enhances the accuracy of the final answer. This innovative training paradigm proves effective in a multimodal context, where the interplay between textual and visual elements is critical for achieving high performance. Moreover, the robustness of $\\mathrm{MC - CoT_{Base}}$ suggests that it is not only the size of the model that dictates performance but also the methodological advancements that come with thoughtful design and algorithmic improvements. This is evident from the fact that our base-sized model can outstrip the performance of models with parameter counts orders of magnitude larger.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "MC-CoT", + "bbox": [ + 668, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Quantative Analysis", + "text_level": 1, + "bbox": [ + 215, + 146, + 429, + 161 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation Study We present the experimental results on the ScienceQA dataset illustrated in Table 3. It can be seen that both using $\\hat{Y}$ only and $\\bar{Y}$ only can achieve comparable performance but the combined model outperforms them. This indicates that the two voting strategies are complementary to each other. Moreover, we can see that the model without voting R (rationales) obtains a significant drop in performance across all categories, with the average accuracy plummeting to $84.70\\%$ . This highlights the importance of the rationale voting in improving the model's reasoning capabilities and accuracy. The absence of the answer voting mechanism results in a decrease in performance in the SOC, LAN, NO, G1-6, and G7-12 categories, with a slight improvement in NAT and TXT. The overall average is $90.19\\%$ , indicating that answer voting contributes to the overall robustness, albeit not as critically as rationale voting. Inference voting means that voting is only conducted during the inference stage and not during the training stage. The significant performance drop of this strategy indicates that inference phase voting is not as effective as training phase voting.", + "bbox": [ + 212, + 166, + 787, + 393 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/a6b15c60f0050700b2a4d6a916e2b667d8359ac3d7ecc54f24a0144ca24986aa.jpg", + "table_caption": [ + "Table 3: Ablation study on ScienceQA dataset." + ], + "table_footnote": [], + "table_body": "
ModelNATSOCLANTXTIMGNOG1-6G7-12AVG
MC-CoTBase91.8784.5993.0092.2888.3092.7590.6490.6490.64
w/ mean only92.1083.8090.8292.4788.8490.6690.1689.7890.03
w/ weighted only91.6183.6991.0092.3888.4590.8089.7289.9189.79
w/o voting R87.3477.3985.1887.6883.2486.3484.5884.9084.70
w/o voting A92.3684.1490.6492.4288.3091.5090.3589.9190.19
inference voting87.4377.2884.6487.3982.8586.4184.9183.9884.58
", + "bbox": [ + 220, + 429, + 782, + 545 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effect of Rationale Generation We categorized the predictions from the Multimodal-CoTBase and MC-CoTBase models into four types: (i) Good Rationale and Answer; (ii) Bad Rationale and Answer; (iii) Good Rationale but Bad Answer; (iv) Bad Rationale but Good Answer. A 'Good Rationale' indicates the high-quality generated rationale. The results are depicted in Fig. 4. It can be seen that MC-CoT exhibits a higher frequency of Good Rationale and Answer, and a lower frequency of Bad Rationale and Answer. This suggests the self-consistency training approach improves the rationale's quality and the accuracy of the answers.", + "bbox": [ + 212, + 566, + 787, + 688 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We present the evaluation of the generated rationales and answers on the ScienceQA dataset in Table 4. It can be seen that our proposed MC-CoTBase model generated rationales that improved RougeL scores by approximately $1\\%$ compared to the Multimodal-CoTBase model. However, this led to an approximately $6\\%$ increase in the average accuracy of predicted answers. Similarly, the rationales generated by the MC-CoTF-Large model showed only about a $0.5\\%$ improvement in RougeL scores compared to the MC-CoT base model. Yet, this resulted in an approximate $4\\%$ increase in the average accuracy of predicted answers. This demonstrates that even slight improvements in the quality of rationales can impact the answer inference of the multimodal reasoning models.", + "bbox": [ + 212, + 688, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Tan. et al.", + "bbox": [ + 271, + 114, + 341, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4b8205ac181f05a28e1cc6195644aa15562e4761b872b5a4c97aeda8037e4d24.jpg", + "image_caption": [ + "Fig. 4: The relationship between rationale and answer." + ], + "image_footnote": [], + "bbox": [ + 220, + 146, + 488, + 300 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/88b86336f32b88749e0ab7713d1b8eaaa13bb617dc47e5bb60981716ad81d4dd.jpg", + "table_caption": [ + "Table 4: The RougeL score of the generated rationales and the average accuracy of predicted answers on the ScienceQA dataset." + ], + "table_footnote": [], + "table_body": "
MethodRougeLAvg Acc
Multimodal-CoTBase96.9784.91
MC-CoTBase97.9890.64
MC-CoTF-Large98.4794.88
", + "bbox": [ + 511, + 242, + 777, + 314 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Qualitative Analysis", + "text_level": 1, + "bbox": [ + 215, + 373, + 429, + 388 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We present two illustrative examples of predictions, as depicted in Figure 5, to demonstrate the efficacy of our proposed model. These examples have been carefully selected to showcase the model's capabilities in handling complex patterns and making accurate choices. For readers interested in exploring further, a more detailed collection of examples is provided in Appendix.", + "bbox": [ + 212, + 400, + 787, + 476 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b5926755da59c6b8a81ca1d1475e3866da2c900f1fe3e77b06f99dd2dc80d2c4.jpg", + "table_caption": [], + "table_footnote": [ + "Context: Use the data to answer the question below.", + "Question: Is the following statement about our solar system true or false? Of the four smallest planets, two are made mainly of gas.", + "Options:", + "(A) false", + "(B) true" + ], + "table_body": "
Planet Volume (cm3) Primary composition
Mercury 6.08 x 1015 rock Venus 9.28 x 1015 rock Earth 1.08 x 1012 rock Mars 1.63 x 1014 rock Jupiter 1.43 x 1015 gas Saturn 8.27 x 1014 gas Uranus 6.83 x 1013 ice Neptune 6.25 x 1013 ice
", + "bbox": [ + 279, + 505, + 406, + 580 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Multimodal-CoT", + "text_level": 1, + "bbox": [ + 475, + 508, + 553, + 518 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A planet's volume tells you the size of the planet .... The volumes of Jupiter, Venus, Earth, and Mars have the smallest exponents. So, these four planets are the smallest. The four planets are made mainly of gas. So, of the four smallest planets, two are made mainly of gas.. The answer is (B) true.", + "bbox": [ + 475, + 535, + 604, + 628 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "MC-CoT", + "text_level": 1, + "bbox": [ + 629, + 508, + 671, + 518 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A planet's volume tells you the size of the planet .... The volumes of Mercury, Venus, Earth, and Mars have the smallest exponents. So, these four planets are the smallest. These four planets are made mainly of rock. So, of the four smallest planets, none are made mainly of gas.. The answer is (A) false.", + "bbox": [ + 633, + 535, + 764, + 628 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/751a83fdb3f6dcad53dce7a5eb5e0422ebcfda47bfb2720d6ad89a6148659da4.jpg", + "image_caption": [ + "Fig. 5: Comparison on predicted examples." + ], + "image_footnote": [ + "Context: Use the graph to answer the question below.", + "Question: Which month is the hottest on average in Adelaide?" + ], + "bbox": [ + 269, + 666, + 410, + 742 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/8205a5b75cc22a8789b2d2e198bcdb96c341cfcfc3e9e36ca23ec89a0c34a173.jpg", + "table_caption": [ + "Options:" + ], + "table_footnote": [ + "(B) April and May", + "(C) January and February" + ], + "table_body": "
(A) June, July, and August
", + "bbox": [ + 272, + 768, + 375, + 780 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Multimodal-CoT", + "text_level": 1, + "bbox": [ + 475, + 670, + 553, + 680 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Scientists record climate data from places around the world .... The average temperatures in April and February are around over $25^{\\circ}\\mathrm{C}$ . These months have the highest average temperatures of all of the months. So, they are hotter than the other months..", + "bbox": [ + 473, + 689, + 604, + 755 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The answer is (A). June, July, and August.", + "bbox": [ + 475, + 762, + 602, + 780 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "MC-CoT", + "text_level": 1, + "bbox": [ + 629, + 670, + 669, + 680 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Scientists record climate data from places around the world .... The average temperatures in January and February are between above $20^{\\circ}\\mathrm{F}$ . These months have the highest average temperatures of all of the months. So, they are hotter than the other months.", + "bbox": [ + 630, + 689, + 766, + 755 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The answer is (C) January and February.", + "bbox": [ + 630, + 762, + 764, + 780 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "MC-CoT", + "bbox": [ + 668, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The first example highlights a scenario where the Multimodal-CoT model encountered difficulties in generating a coherent and accurate rationale, ultimately leading to an erroneous conclusion. This misstep underscores the challenges inherent in integrating multiple sources of information and extracting a logical reasoning path. Conversely, the MC-CoT model demonstrated its proficiency by navigating these complexities successfully, thereby arriving at the correct answer. This contrast exemplifies the significance of robust reasoning capabilities in achieving high accuracy in complex question-answering tasks.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The second example delves into a different set of challenges faced by both models. Here, the Multimodal-CoT model stumbled over the interpretation of data presented in a chart, a mistake that significantly derailed its reasoning process and resulted in an incorrect answer. This error highlights the critical importance of accurate data interpretation skills in models designed for multimodal contexts. On the other hand, the MC-CoT model, while generally more adept in this instance, exhibited a lapse in understanding units of measurement—mistaking degrees Celsius $(^{\\circ}\\mathrm{C})$ for degrees Fahrenheit $(^{\\circ}\\mathrm{F})$ . Interestingly, this error did not compromise the integrity of its final answer, suggesting a certain resilience in the model's reasoning process despite minor inaccuracies.", + "bbox": [ + 212, + 268, + 789, + 417 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 443, + 359, + 459 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this study, we conducted an exhaustive analysis of the Multimodal-CoT framework and identified that the quality of rationales plays a crucial role in enhancing the performance of multimodal reasoning models. Recognizing the critical importance of this aspect, we developed an innovative solution aimed at improving the quality of rationales. Our solution employs a self-consistency training strategy that leverages the natural variability introduced by dropout techniques during training to foster a consensus mechanism through voting. This approach is supported by both theoretical underpinnings and empirical evidence, highlighting its effectiveness and wide applicability. We meticulously designed experiments to evaluate the impact of our self-consistency training strategy on the performance of multimodal reasoning models across diverse scenarios and datasets. The results from these experiments were revealing, clearly showing that our method significantly boosts the capabilities of multimodal models. A particularly significant finding was our approach's ability to enhance the performance of smaller models, enabling them to match or even exceed the capabilities of much larger models. We believe our strategy offers a promising avenue for advancing multimodal reasoning.", + "bbox": [ + 212, + 477, + 787, + 734 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "A limitation of our approach is that we do not explicitly impose grammatical constraints, meaning the generated rationales are not guaranteed to be grammatically correct. It is worth noting that the Multimodal-CoT framework also does not guarantee the grammatical validity of the generated rationales. Despite this, empirical evidence has shown that our approach can yield superior results across various datasets, indicating that the rationales produced by our approach can significantly enhance the model's reasoning ability towards the correct answers.", + "bbox": [ + 212, + 734, + 789, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Tan. et al.", + "bbox": [ + 271, + 114, + 341, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Acknowledgement", + "text_level": 1, + "bbox": [ + 215, + 143, + 426, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the Science & Technology Innovation 2030 Major Program Project No. 2021ZD0150100, National Natural Science Foundation of China Project No. U21A20427, Project No. WU2022A009 from the Center of Synthetic Biology and Integrated Bioengineering of Westlake University, Project No. WU2023C019 from the Westlake University Industries of the Future Research. Finally, we thank the Westlake University HPC Center for providing part of the computational resources, and Project No. 23-407-3-29 from the Shenyang Science and Technology Program.", + "bbox": [ + 212, + 174, + 787, + 295 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 316, + 321, + 332 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6077-6086 (2018) 4, 9, 10", + "2. Anil, R., Dai, A.M., First, O., Johnson, M., Lepikhin, D., Passos, A., Shakeri, S., Taropa, E., Bailey, P., Chen, Z., et al.: Palm 2 technical report. arXiv preprint arXiv:2305.10403 (2023) 1", + "3. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: Vqa: Visual question answering. In: Proceedings of the IEEE international conference on computer vision. pp. 2425-2433 (2015) 5", + "4. Askell, A., Bai, Y., Chen, A., Drain, D., Ganguli, D., Henighan, T., Jones, A., Joseph, N., Mann, B., DasSarma, N., et al.: A general language assistant as a laboratory for alignment. arXiv preprint arXiv:2112.00861 (2021) 1", + "5. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966 (2023) 2", + "6. Besta, M., Blach, N., Kubicek, A., Gerstenberger, R., Gianinazzi, L., Gajda, J., Lehmann, T., Podstawski, M., Niewiadomski, H., Nczyk, P., et al.: Graph of thoughts: Solving elaborate problems with large language models. arXiv preprint arXiv:2308.09687 (2023) 1, 2, 4", + "7. Biderman, S., Schoelkopf, H., Anthony, Q.G., Bradley, H., O'Brien, K., Hallahan, E., Khan, M.A., Purohit, S., Prashanth, U.S., Raff, E., et al.: Pythia: A suite for analyzing large language models across training and scaling. In: International Conference on Machine Learning. pp. 2397-2430. PMLR (2023) 9, 11", + "8. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) 1", + "9. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) 9", + "10. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) 11" + ], + "bbox": [ + 218, + 345, + 785, + 837 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "MC-CoT", + "bbox": [ + 668, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "11. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020) 11", + "12. Chen, Z., Zhou, Q., Shen, Y., Hong, Y., Zhang, H., Gan, C.: See, think, confirm: Interactive prompting between vision and language models for knowledge-based visual reasoning. arXiv preprint arXiv:2301.05226 (2023) 9, 11", + "13. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311 (2022) 1, 11", + "14. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, E., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022) 1", + "15. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022) 9, 11", + "16. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Qiu, Z., Lin, W., Yang, J., Zheng, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023) 2", + "17. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023) 2", + "18. Gao, P., Jiang, Z., You, H., Lu, P., Hoi, S.C., Wang, X., Li, H.: Dynamic fusion with intra-and inter-modality attention flow for visual question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6639-6648 (2019) 4, 9, 10", + "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 11", + "20. Horawalavithana, S., Munikoti, S., Stewart, I., Kvinge, H.: Scitune: Aligning large language models with scientific multimodal instructions. arXiv preprint arXiv:2307.01139 (2023) 9, 10", + "21. Huang, J., Chang, K.C.C.: Towards reasoning in large language models: A survey. arXiv preprint arXiv:2212.10403 (2022) 2", + "22. Kamath, A., Clark, C., Gupta, T., Kolve, E., Hoiem, D., Kembhavi, A.: Webly supervised concept expansion for general purpose vision models. In: European Conference on Computer Vision. pp. 662-681. Springer (2022) 9, 11", + "23. Kenton, J.D.M.W.C., Toutanova, L.K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of naacL-HLT. vol. 1, p. 2 (2019) 11", + "24. Khashabi, D., Min, S., Khot, T., Sabharwal, A., Tafjord, O., Clark, P., Hajishirzi, H.: UNIFIEDQA: Crossing format boundaries with a single QA system. In: Cohn, T., He, Y., Liu, Y. (eds.) Findings of the Association for Computational Linguistics: EMNLP 2020. pp. 1896-1907. Association for Computational Linguistics, Online (Nov 2020). https://doi.org/10.18653/v1/2020-findings-emnlp.171, https://aclanthology.org/2020_findings-emnlp.171_9, 10, 11", + "25. Kim, J.H., Jun, J., Zhang, B.T.: Bilinear attention networks. Advances in neural information processing systems 31 (2018) 4, 9, 10", + "26. Kim, W., Son, B., Kim, I.: Vilt: Vision-and-language transformer without convolution or region supervision. In: International Conference on Machine Learning. pp. 5583-5594. PMLR (2021) 4, 9, 10" + ], + "bbox": [ + 217, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Tan. et al.", + "bbox": [ + 271, + 114, + 341, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "27. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) 4, 9, 11", + "28. Li, L.H., Yatskar, M., Yin, D., Hsieh, C.J., Chang, K.W.: What does bert with vision look at? In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. pp. 5265-5275 (2020) 4, 9, 10", + "29. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023) 2, 9, 10", + "30. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281 (2023) 2", + "31. Lu, J., Batra, D., Parikh, D., Lee, S.: Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems 32 (2019) 9, 11", + "32. Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems 35, 2507-2521 (2022) 2, 3, 4, 9, 10", + "33. Lu, P., Peng, B., Cheng, H., Galley, M., Chang, K.W., Wu, Y.N., Zhu, S.C., Gao, J.: Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842 (2023) 2", + "34. Lu, P., Peng, B., Cheng, H., Galley, M., Chang, K.W., Wu, Y.N., Zhu, S.C., Gao, J.: Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842 (2023) 9, 10", + "35. Lu, P., Qiu, L., Chen, J., Xia, T., Zhao, Y., Zhang, W., Yu, Z., Liang, X., Zhu, S.C.: Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. arXiv preprint arXiv:2110.13214 (2021) 4, 9, 10", + "36. Luo, G., Zhou, Y., Ren, T., Chen, S., Sun, X., Ji, R.: Cheap and quick: Efficient vision-language instruction tuning for large language models. arXiv preprint arXiv:2305.15023 (2023) 2, 4, 9, 10", + "37. Marino, K., Chen, X., Parikh, D., Gupta, A., Rohrbach, M.: Krisp: Integrating implicit and symbolic knowledge for open-domain knowledge-based vqa. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14111-14121 (2021) 9, 11", + "38. Ning, X., Lin, Z., Zhou, Z., Yang, H., Wang, Y.: Skeleton-of-thought: Large language models can do parallel decoding. arXiv preprint arXiv:2307.15337 (2023) 4", + "39. OpenAI: Gpt-4 technical report (2023) 1, 9, 10", + "40. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022) 1", + "41. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 11", + "42. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020) 1, 11" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "MC-CoT", + "bbox": [ + 668, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020) 9", + "44. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems 28 (2015) 11", + "45. Schwenk, D., Khandelwal, A., Clark, C., Marino, K., Mottaghi, R.: A-okvqa: A benchmark for visual question answering using world knowledge. In: European Conference on Computer Vision. pp. 146-162. Springer (2022) 3, 4, 9", + "46. Tan, H., Bansal, M.: Lxmert: Learning cross-modality encoder representations from transformers. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (2019) 9, 11", + "47. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) 1", + "48. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) 1", + "49. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) 11", + "50. Wang, X., Wei, J., Schuurmans, D., Le, Q.V., Chi, E.H., Narang, S., Chowdhery, A., Zhou, D.: Self-consistency improves chain of thought reasoning in language models. In: The Eleventh International Conference on Learning Representations (2022) 1, 2, 4", + "51. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022) 1, 3, 11", + "52. Wu, C., Liu, J., Wang, X., Dong, X.: Chain of reasoning for visual question answering. Advances in Neural Information Processing Systems 31 (2018) 5", + "53. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023) 2", + "54. Yang, Z., Gan, Z., Wang, J., Hu, X., Lu, Y., Liu, Z., Wang, L.: An empirical study of gpt-3 for few-shot knowledge-based vqa. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 3081-3089 (2022) 9, 11", + "55. Yang, Z., Li, L., Wang, J., Lin, K., Azarnasab, E., Ahmed, F., Liu, Z., Liu, C., Zeng, M., Wang, L.: Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381 (2023) 4", + "56. Yao, S., Yu, D., Zhao, J., Shafran, I., Griffiths, T.L., Cao, Y., Narasimhan, K.: Tree of thoughts: Deliberate problem solving with large language models. arXiv preprint arXiv:2305.10601 (2023) 1, 2, 4", + "57. Yin, S., Fu, C., Zhao, S., Li, K., Sun, X., Xu, T., Chen, E.: A survey on multimodal large language models. arXiv preprint arXiv:2306.13549 (2023) 2", + "58. Yu, Z., Yu, J., Cui, Y., Tao, D., Tian, Q.: Deep modular co-attention networks for visual question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6281-6290 (2019) 4, 9, 10", + "59. Zhang, P., Li, X., Hu, X., Yang, J., Zhang, L., Wang, L., Choi, Y., Gao, J.: Vinvl: Revisiting visual representations in vision-language models. In: Proceedings of the" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Tan. et al.", + "bbox": [ + 271, + 114, + 341, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF conference on computer vision and pattern recognition. pp. 5579-5588 (2021) 11", + "60. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199 (2023) 2, 4", + "61. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199 (2023) 9, 10", + "62. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022) 1, 11", + "63. Zhang, Z., Zhang, A., Li, M., Zhao, H., Karypis, G., Smola, A.: Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923 (2023) 2, 4, 5, 6, 9, 10, 11" + ], + "bbox": [ + 215, + 146, + 787, + 340 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "MC-CoT", + "bbox": [ + 668, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Boosting the Power of Small Multimodal Reasoning Models to Match Larger Models with Self-Consistency Training/cda53ae2-a211-4097-a2a7-2df98969f4ef_model.json b/2024/Boosting the Power of Small Multimodal Reasoning Models to Match Larger Models with Self-Consistency Training/cda53ae2-a211-4097-a2a7-2df98969f4ef_model.json new file mode 100644 index 0000000000000000000000000000000000000000..812671f7f702dd8494d06f479d2fdb2b29cdc210 --- /dev/null +++ b/2024/Boosting the Power of Small Multimodal Reasoning Models to Match Larger Models with Self-Consistency Training/cda53ae2-a211-4097-a2a7-2df98969f4ef_model.json @@ -0,0 +1,2768 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.226, + 0.141, + 0.78, + 0.21 + ], + "angle": 0, + "content": "Boosting the Power of Small Multimodal Reasoning Models to Match Larger Models with Self-Consistency Training" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.235, + 0.782, + 0.268 + ], + "angle": 0, + "content": "Cheng Tan\\(^{1,4,5\\star}\\), Jingxuan Wei\\(^{2,3\\star\\boxtimes}\\), Zhangyang Gao\\(^{1,4,5\\star}\\), Linzhuang Sun\\(^{2,3}\\), Siyuan Li\\(^{1,4,5}\\), Ruifeng Guo\\(^{2,3}\\), Bihui Yu\\(^{2,3}\\), and Stan Z. Li\\(^{4\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.277, + 0.761, + 0.32 + ], + "angle": 0, + "content": "\\(^{1}\\) Zhejiang University \n\\(^{2}\\) Shenyang Institute of Computing Technology, Chinese Academy of Sciences \n\\(^{3}\\) University of Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.247, + 0.32, + 0.756, + 0.334 + ], + "angle": 0, + "content": "\\(^{4}\\) AI Lab, Research Center for Industries of the Future, Westlake University" + }, + { + "type": "text", + "bbox": [ + 0.249, + 0.334, + 0.753, + 0.376 + ], + "angle": 0, + "content": "\\(^{5}\\) Institute of Advanced Technology, Westlake Institute for Advanced Study tancheng@westlake.edu.cn, weijingxuan20@mails.ucas.edu.cn \\(\\star\\) Equal contribution, \\(\\boxtimes\\) Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.241, + 0.277, + 0.761, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.415, + 0.744, + 0.693 + ], + "angle": 0, + "content": "Abstract. Multimodal reasoning is a challenging task that requires models to reason across multiple modalities to answer questions. Existing approaches have made progress by incorporating language and visual modalities into a two-stage reasoning framework, separating rationale generation from answer inference. However, these approaches often fall short due to the inadequate quality of the generated rationales. In this work, we delve into the importance of rationales in model reasoning. We observe that when rationales are completely accurate, the model's accuracy significantly improves, highlighting the need for high-quality rationale generation. Motivated by this, we propose MC-CoT, a self-consistency training strategy that generates multiple rationales and answers, subsequently selecting the most accurate through a voting process. This approach not only enhances the quality of generated rationales but also leads to more accurate and robust answers. Through extensive experiments, we demonstrate that our approach significantly improves model performance across various benchmarks. Remarkably, we show that even smaller base models, when equipped with our proposed approach, can achieve results comparable to those of larger models, illustrating the potential of our approach in harnessing the power of rationales for improved multimodal reasoning. The code is available at github.com/chengtan9907/mc-cot." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.706, + 0.684, + 0.721 + ], + "angle": 0, + "content": "Keywords: Multimodal reasoning \\(\\cdot\\) visual question answering" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.747, + 0.377, + 0.763 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Recent advances in large language models [2, 4, 8, 13, 14, 39, 40, 42, 47, 48, 62] have led to the exploration of Chain-of-Thought (CoT) prompting [6, 50, 51, 56]. This approach, which directs the model to systematically unravel rationales before providing answers, rather than responding directly, has showcased the model's" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.343, + 0.127 + ], + "angle": 0, + "content": "Tan. et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.145, + 0.549, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.145, + 0.785, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.262, + 0.788, + 0.292 + ], + "angle": 0, + "content": "Fig. 1: An example of multimodal reasoning that answers the question by reasoning across both vision and language modalities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.789, + 0.421 + ], + "angle": 0, + "content": "impressive efficacy across a variety of natural language processing (NLP) tasks. Moreover, the advent of CoT prompting has catalyzed a plethora of research endeavors delving into the reasoning prowess of large language models. A diverse range of Chain-of-Thought strategies has been investigated, including the voting-facilitated CoT-SC [50], the transition from chain-like to tree-like thinking with Tree-of-Thoughts [56], and further expansion into graph-structured thinking [6]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.79, + 0.622 + ], + "angle": 0, + "content": "While CoT reasoning has been thoroughly established in the realm of language models, its foray into the vast and intricate landscape of multimodal reasoning is still in its infancy. As shown in Fig. 1, Multimodal reasoning [5, 16, 17, 21, 29, 30, 32, 33, 36, 53, 57, 60], which inherently involves the seamless fusion of information from disparate modalities such as text and images, presents unique challenges. The process of extracting, correlating, and generating rationales across multiple modalities is decidedly more complex than the tasks encountered in a solely text-based modality. A recent seminal work, Multimodal-CoT [63], has pioneered the application of the Chain-of-Thought prompting strategy to multimodal reasoning tasks. This approach encompasses a two-stage framework that distinctively separates rationale generation from answer inference. By obliging the model to generate rationales prior to answering questions, Multimodal-CoT mirrors the language-only CoT prompting, thus enabling reasoning across multiple modalities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.789, + 0.732 + ], + "angle": 0, + "content": "Despite Multimodal-CoT has made promising strides in the realm of multimodal reasoning, as evidenced in Fig. 2, its improvements over the no-rationale baseline are still limited. Moreover, compared to the ground-truth rationale, the predicted rationale falls short, often yielding results that lack relevance to the posed question. This discrepancy primarily stems from the quality of the generated rationales, highlighting the crucial role of rationale quality in the success of the chain-of-thought reasoning process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In light of the above observations, it becomes evident that the potency of the Chain-of-Thought reasoning in a multimodal scenario is intrinsically tethered to the accuracy of the rationales generated. A high-quality rationale not only sheds light on the model's thought process but also paves the way for more precise answers. This leads to a critical question: how can we develop a strategy to enhance the quality of these rationales, and consequently, improve the overall performance of multimodal CoT?" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.669, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MC-CoT" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.139, + 0.699, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.353, + 0.785, + 0.381 + ], + "angle": 0, + "content": "Fig. 2: The comparison of answer accuracy on ScienceQA using the Multimodal-CoT framework with no rationale, predicted rationales, and ground-truth rationales." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.593 + ], + "angle": 0, + "content": "Our study is driven by the hypothesis that enhancing the quality of rationale generation can significantly improve the model's reasoning capabilities and overall performance. To this end, we introduce a simple yet effective strategy that capitalizes on the inherent variability of deep neural models during training, particularly stemming from mechanisms like dropout. Our approach involves having the model generate multiple rationales and then voting for the most consistent words across these rationales to yield a more refined and accurate rationale. The same voting mechanism is then applied to answer generation, further boosting the model's confidence in its predictions. It is important to note that the inference phase remains completely unaffected by this voting mechanism and continues to operate in the same manner as it does in the original Multimodal CoT framework. Through this approach, we aim to facilitate a more robust and accurate multimodal reasoning ability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.788, + 0.701 + ], + "angle": 0, + "content": "Extensive experiments across ScienceQA [32] and A-OKVQA [45] benchmark datasets demonstrate the efficacy of our proposed approach. Notably, by improving the rationale quality, even smaller models equipped with our strategy manifest performance metrics that rival, and at times surpass, those of considerably larger models. This not only confirms the efficacy of our rationale refinement strategy but also opens up a promising avenue where smaller, more efficient models can be rendered competitive in the multimodal reasoning landscape." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.723, + 0.388, + 0.739 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.755, + 0.41, + 0.77 + ], + "angle": 0, + "content": "2.1 Chain-of-Thought" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The Chain-of-Thought (CoT) paradigm has emerged as a transformative approach, aiming to elucidate the reasoning processes of large language models and boost their ability in a variety of NLP tasks. The vanilla CoT [51] prompting strategy models to generate intermediate reasoning steps, leading to significant" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.344, + 0.128 + ], + "angle": 0, + "content": "Tan. et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "performance improvements across a spectrum of tasks. Its methodology laid the groundwork for subsequent advancements in the CoT paradigm. CoT-SC [50] introduced a self-consistency decoding strategy by sampling multiple reasoning paths and selecting the most consistent path to enhance the reliability of the generated rationales." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.222, + 0.79, + 0.358 + ], + "angle": 0, + "content": "The versatility and adaptability of the CoT paradigm were further demonstrated with the advent of advanced reasoning structures. Tree-of-Thoughts (ToT) [56] transitioned from linear chains to more intricate tree-like structures, aiming to capture more complex reasoning patterns. In a parallel development, [6] ventured into graph-based reasoning, presenting the Graph-of-Thought (GoT) approach, which allowed for a more interconnected reasoning process. Skeleton-of-Thought (SoT) [38] took a different route by emphasizing the benefits of efficiency by underscoring the potential of parallel decoding, enabling models to explore multiple reasoning paths simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.79, + 0.435 + ], + "angle": 0, + "content": "While the aforementioned studies have made significant strides in the CoT paradigm, we seek to extend these advancements into the realm of multimodal reasoning. Multimodal-CoT [63] is a seminal work, but the unsatisfying performance of its generated rationales has been a major bottleneck. Our work aims to fully explore the potential of the CoT paradigm in multimodal reasoning." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.455, + 0.59, + 0.47 + ], + "angle": 0, + "content": "2.2 Multimodal Visual Question Answering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.79, + 0.719 + ], + "angle": 0, + "content": "Multimodal Visual Question Answering (VQA) has emerged as a key research area bridging the domains of vision and language. Pioneering works like MCAN [58], BAN [25], and Top-Down [1] have established a foundation by introducing diverse attention mechanisms. Building on this, DFAF [18] introduced innovative dynamic fusion strategies to facilitate both intra- and inter-modality attention flows. The introduction of Transformer architectures has significantly impacted the VQA field, as demonstrated by state-of-the-art models such as ViLT [26] and VisualBERT [28]. In the same vein, Patch-TRM [35] leveraged a pyramid cross-modal Transformer, paired with pre-trained input embeddings from the icon dataset, advancing the understanding of abstract diagrams. Recently, incorporating large language models into the multimodal framework has seen a surge, exemplified by models like LLaMA-Adapter [60] and BLIP-2 [27]. LaVIN [36] introduced a novel routing algorithm that helps the model autonomously switch between unimodal and multimodal instruction reasoning paths. These models employ frozen visual encoders paired with fine-tuned language models, thereby setting new standards in VQA tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.79, + 0.84 + ], + "angle": 0, + "content": "ScienceQA [32] introduced a reasoning process into the VQA task, setting a benchmark for multimodal chain-of-thought reasoning. MM-REACT [55], with its innovative prompting design, enables language models to process multimodal information, facilitating the integration of ChatGPT with visual experts. A-OKVQA [45] requires commonsense reasoning about the depicted scene in an image. Multimodal-CoT [63] integrates language (text) and vision (image) modalities into a two-stage framework, distinctly separating rationale generation from answer inference." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.669, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MC-CoT" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "In this work, we typically focus on multimodal reasoning tasks, a subtype of multimodal visual question answering (VQA) that incorporates rationales. The significance of this task lies in its ability to more accurately simulate human cognitive processes. When humans answer questions, they often rely on observed visual information and existing textual knowledge, synthesizing the two to form a reasoned conclusion. These multimodal reasoning tasks comprehensively assess a model's capacity to understand and fuse visual and textual information, thereby propelling the development of multimodal learning. Moreover, these tasks enhance the interpretability of artificial intelligence systems. Because models must supply not only answers but also supporting rationales, their decision-making processes become more interpretable and comprehensible." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.333, + 0.331, + 0.348 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.361, + 0.373, + 0.374 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.788, + 0.457 + ], + "angle": 0, + "content": "In this work, we focus on multimodal reasoning, a challenging task that involves reasoning across multiple modalities to answer questions. Specifically, we consider visual question answering [3, 52], where the model is tasked with answering questions based on the information provided in the question itself and the accompanying image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.458, + 0.788, + 0.593 + ], + "angle": 0, + "content": "Formally, given a dataset \\(\\mathcal{D} = \\{\\mathcal{X},\\mathcal{Y}\\}\\), where \\(X\\in \\mathcal{X}\\) represents a multimodal input consisting of text and images, and \\(Y\\in \\mathcal{V}\\) represents the corresponding output, the task of multimodal reasoning involves learning a mapping function parameterized by \\(\\Theta\\): \\(\\mathcal{F}_{\\Theta}:\\mathcal{X}\\to \\mathcal{V}\\) that accurately predicts the output \\(Y\\) for a given input \\(X\\). The multimodal input \\(X\\) can be represented as \\(X = (T,I)\\), where \\(T\\) is the text component, and \\(I\\) is the image component of the input. The function \\(\\mathcal{F}\\) should be able to effectively leverage the information from both the text and image modalities to make accurate predictions. Specifically, the basic visual question answering task is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.601, + 0.785, + 0.621 + ], + "angle": 0, + "content": "\\[\nY = \\arg \\max _ {Y ^ {\\prime}} p \\left(Y ^ {\\prime} \\mid T, I\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.704 + ], + "angle": 0, + "content": "where \\( p(Y' \\mid T, I) \\) is the probability of answer \\( Y' \\) given the text \\( T \\) and image \\( I \\). The multimodal reasoning task builds upon the basic visual question answering task, extending its requirements to necessitate the generation of a rationale \\( R' \\) that elucidates the reasoning process underpinning the answer \\( Y \\). This task can be mathematically represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.391, + 0.712, + 0.785, + 0.733 + ], + "angle": 0, + "content": "\\[\nY, R = \\arg \\max _ {Y ^ {\\prime}, R ^ {\\prime}} p \\left(Y ^ {\\prime}, R ^ {\\prime} \\mid T, I\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.745, + 0.788, + 0.79 + ], + "angle": 0, + "content": "Multimodal-CoT [63] decomposed the multimodal reasoning task into a two-stage framework, where the model first generates a rationale \\( R \\) and then utilizes this rationale to predict the answer \\( Y \\). The process is delineated as:" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.798, + 0.785, + 0.827 + ], + "angle": 0, + "content": "\\[\nR = \\arg \\max _ {R ^ {\\prime}} p \\left(R ^ {\\prime} \\mid T, I\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.823, + 0.6, + 0.843 + ], + "angle": 0, + "content": "\\[\nY = \\arg \\max _ {Y ^ {\\prime}} p (Y ^ {\\prime} | R, T, I)\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.344, + 0.128 + ], + "angle": 0, + "content": "Tan. et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.143, + 0.303, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.28, + 0.3, + 0.291 + ], + "angle": 0, + "content": "(a) Vanilla Prompt" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.143, + 0.401, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.28, + 0.378, + 0.291 + ], + "angle": 0, + "content": "(b) CoT" + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.143, + 0.508, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.434, + 0.28, + 0.483, + 0.291 + ], + "angle": 0, + "content": "(c) CoT-SC" + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.145, + 0.64, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.28, + 0.64, + 0.291 + ], + "angle": 0, + "content": "(d) Multimodal-CoT" + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.145, + 0.783, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.28, + 0.744, + 0.291 + ], + "angle": 0, + "content": "(e) MC-CoT" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.306, + 0.79, + 0.446 + ], + "angle": 0, + "content": "Fig. 3: A comparison schematic diagram of different Chain-of-Thought (CoT) prompt-based reasoning methods. (a) The basic input-output prompt, (b) Chain-of-Thought with intermediate chain-like reasoning, (c) Chain-of-Thought Self-Consistency (CoT-SC) that utilizes the consistency of multiple independent chains of thoughts for reasoning, (d) Multimodal-CoT, which infers the rationale using the input text and image, and then predicts the answer using the rationale as part of the input, and (e) MC-CoT that infers a high-quality rationale through word-level voting, and then obtains a high-quality answer using majority vote. It is worth noting that our approach leverages multiple chain consistency only during the training phase, in contrast to CoT-SC, which employs it during the inference stage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.788, + 0.507 + ], + "angle": 0, + "content": "In this work, we follow the same two-stage framework as Multimodal-CoT [63] and explore the potential of such a chain-of-thought reasoning paradigm." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.531, + 0.604, + 0.547 + ], + "angle": 0, + "content": "3.2 Multimodal Consistent Chain-of-Thought" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.558, + 0.789, + 0.635 + ], + "angle": 0, + "content": "Though the existing multimodal CoT framework has been proven to be beneficial in multimodal reasoning, there remains a large room to fully explore its potential. In this work, we introduce a simple yet effective training strategy, Multimodal Consistent Chain-of-Thought (MC-CoT), to learn high-quality rationales and thus boost the reasoning performance of the multimodal CoT framework." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.788, + 0.698 + ], + "angle": 0, + "content": "Rationale Generation Leveraging the inherent randomness introduced by the dropout operations, the model is able to generate multiple diverse rationales for a given input text-image pair \\((T,I)\\) by sampling from the model \\(N_{r}\\) times:" + }, + { + "type": "equation", + "bbox": [ + 0.38, + 0.71, + 0.787, + 0.728 + ], + "angle": 0, + "content": "\\[\nR ^ {i} \\sim p (R \\mid T, I), i = 1, 2, \\dots , N _ {r} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.739, + 0.788, + 0.784 + ], + "angle": 0, + "content": "We perform a voting process on the generated rationales to select the most consistent words across the rationales. Specifically, the best rationale \\( R^{*} \\) is selected as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.426, + 0.784, + 0.787, + 0.805 + ], + "angle": 0, + "content": "\\[\nR _ {j} ^ {*} = \\operatorname {V o t e} \\left(\\left\\{R _ {j} ^ {i} \\right\\} _ {1} ^ {N _ {r}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.843 + ], + "angle": 0, + "content": "where, for each position \\( j \\) in the rationale, we choose the majority of the words at position \\( j \\) across the \\( N_r \\) generated rationales to form the best rationale \\( R_j^* \\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.669, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MC-CoT" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "Answer Inference In the answer inference stage, we also generate multiple answers based on the best rationale \\( R^{*} \\) and the input text-image pair \\( (T, I) \\) by sampling from the model \\( N_{a} \\) times:" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.2, + 0.785, + 0.216 + ], + "angle": 0, + "content": "\\[\nY ^ {i} \\sim p (Y \\mid R ^ {*}, T, I), i = 1, 2, \\dots , N _ {a}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.222, + 0.586, + 0.238 + ], + "angle": 0, + "content": "The final answer is selected by the majority voting:" + }, + { + "type": "equation", + "bbox": [ + 0.429, + 0.244, + 0.785, + 0.261 + ], + "angle": 0, + "content": "\\[\nY ^ {*} = \\operatorname {V o t e} \\left(\\left\\{Y ^ {i} \\right\\} _ {1} ^ {N _ {a}}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.784, + 0.313 + ], + "angle": 0, + "content": "where the majority of the answers across the \\(N_{a}\\) generated answers are regarded as the most convincing answers \\(Y^{*}\\). The pseudocode of the pipeline is shown in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.323, + 0.784, + 0.382 + ], + "angle": 0, + "content": "Voting Strategy Consider the answer inference stage as an illustration. The vanilla Multimodal-CoT model computes a cross-entropy loss between a \\(d\\)-dimensional logit \\(L \\in \\mathbb{R}^d\\) and the ground truth label \\(Y \\in \\mathbb{R}^d\\), which can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.385, + 0.785, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\operatorname {C r o s s E n t r o p y} (L, Y), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.403, + 0.784, + 0.434 + ], + "angle": 0, + "content": "Our averaging-based voting strategy initially calculates the mean logit \\(\\bar{L} \\in \\mathbb{R}^d\\) and weighted mean logit \\(\\hat{L} \\in \\mathbb{R}^d\\):" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.442, + 0.785, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\bar {L} = \\frac {1}{N _ {a}} \\sum_ {i = 1} ^ {N _ {a}} L ^ {i}, \\hat {L} = \\frac {1}{N _ {a}} \\frac {\\sum_ {i = 1} ^ {N _ {a}} w L ^ {i}}{\\sum_ {j} w _ {j}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.488, + 0.784, + 0.577 + ], + "angle": 0, + "content": "where \\( w = 1 / (1 + \\sigma) \\in \\mathbb{R}^d \\) is a weight vector derived from the standard deviation of the logits, calculated across different predictions for each dimension. The weight for each logit's dimension is inversely proportional to its variability, assigning higher confidence to predictions with lower variability and lower confidence to those with higher variability. The final prediction is then a linear combination of \\( \\bar{L} \\) and \\( \\hat{L} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.431, + 0.58, + 0.784, + 0.595 + ], + "angle": 0, + "content": "\\[\nL ^ {*} = \\alpha \\bar {L} + (1 - \\alpha) \\hat {L} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.6, + 0.784, + 0.629 + ], + "angle": 0, + "content": "with \\(\\alpha\\) being a hyperparameter set empirically to 0.5. The cross-entropy loss is subsequently computed between \\(L^{*}\\) and the ground truth label \\(Y\\):" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.638, + 0.784, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\operatorname {C r o s s E n t r o p y} \\left(L ^ {*}, Y\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.659, + 0.786, + 0.84 + ], + "angle": 0, + "content": "We present a comparison of our approach with existing CoT-based reasoning methods in Fig. 3. Specifically, subfigures (a-c) illustrate the prompt learning methods commonly employed for text modality in large language models, while subfigures (d-e) depict the CoT reasoning frameworks in the multimodal field. It can be seen that our approach shares similarities with CoT-SC, such as the utilization of a voting mechanism. However, there are also significant differences between the two approaches: (i) the objective differs. CoT-SC aims to facilitate prompt learning in large language models during the inference stage, while our method seeks to enhance the robustness of model reasoning during the training phase; (ii) the voting process differs. CoT-SC only votes on the final answers inferred from each thought, whereas our method incorporates voting during the rationale generation stage." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.344, + 0.128 + ], + "angle": 0, + "content": "Tan. et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.446, + 0.163 + ], + "angle": 0, + "content": "4 Theoretical Insights" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.185, + 0.579, + 0.2 + ], + "angle": 0, + "content": "4.1 Aggregation Minimizes Expected Loss" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.217, + 0.784, + 0.247 + ], + "angle": 0, + "content": "Theorem 1. (Jensen's inequality) Let \\( X \\) be a random variable and let \\( \\phi \\) be a convex function. Then the following inequality holds: \\( \\phi(\\mathbb{E}[X]) \\leq \\mathbb{E}[\\phi(X)] \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.249, + 0.788, + 0.356 + ], + "angle": 0, + "content": "Our proposed method, MC-CoT, leverages the variability introduced by dropout techniques to generate a diverse set of rationales. By aggregating these rationales, we aim to produce explanations and corresponding answers of higher quality. In the context of our framework, the convex function \\(\\phi\\) is represented by the cross-entropy loss function, and the random variable in question is the set of logits \\(L^i\\). By applying Jensen's inequality to our model, we deduce that the aggregation process leads to a lower expected loss: \\(\\mathcal{L}(\\mathbb{E}(L^{*}),Y)\\leq \\mathbb{E}(\\mathcal{L}(L^{*},Y))\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.385, + 0.459, + 0.399 + ], + "angle": 0, + "content": "4.2 Bias-Variance Trade-off" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.785, + 0.475 + ], + "angle": 0, + "content": "We introduce a linear combination of two voting strategies to derive the final prediction, as delineated in Equation 10. We can analytically decompose the expected value of the squared difference between the predicted outcomes \\( Y^{*} \\) and the ground truth \\( Y \\) into three distinct components:" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.489, + 0.785, + 0.506 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} \\left[ \\left(Y ^ {*} - Y\\right) ^ {2} \\right] = \\operatorname {B i a s} ^ {2} \\left(Y ^ {*}\\right) + \\operatorname {V a r} \\left(Y ^ {*}\\right) + \\epsilon , \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.784, + 0.61 + ], + "angle": 0, + "content": "where \\(\\mathrm{Bias}^2 (Y^*) = (\\mathbb{E}[Y^*] - Y)^2\\) quantifies the systematic deviation of the expected prediction from the ground truth, indicative of the error inherently introduced by the predictive model itself. Meanwhile, \\(\\mathrm{Var}(Y^{*}) = \\mathbb{E}[(Y^{*} - \\mathbb{E}[Y^{*}])^{2}]\\) captures the variability of the predictions, reflecting the extent to which these predictions will fluctuate around their expected value. Lastly, \\(\\epsilon\\) denotes the irreducible error, encapsulating the intrinsic noise within the data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.788, + 0.748 + ], + "angle": 0, + "content": "The mean logits, denoted by \\(\\bar{L}\\), may exhibit lower bias since they represent the composite prediction averaged across multiple models. Employing solely the mean logits promotes exploration by treating all predictions as equally informative. Nevertheless, its measure of central tendency does not account for the individual variability within the predictions. The weighted mean logits, denoted by \\(\\hat{L}\\), aim to reduce variability by assigning greater significance to consistent (low-variance) predictions. This reflects a confidence-based strategy, allocating higher weight to predictions that are less variable. However, it could introduce bias if predictions are consistently erroneous." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The final logits, which include both mean logits and weighted mean logits, are designed to achieve an optimal balance between bias and variance. While the mean logits promote exploration by treating all predictions equally, the weighted mean logits perform exploitation by assigning higher weights to consistent predictions. The combination allows balancing exploration with exploitation, thus improving the performance towards an optimal standard." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.669, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MC-CoT" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.366, + 0.163 + ], + "angle": 0, + "content": "5 Experiment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.343 + ], + "angle": 0, + "content": "Datasets We evaluated MC-CoT on two prominent multimodal reasoning benchmarks. The first, ScienceQA [32], is a comprehensive multimodal science question dataset that contains annotated answers with detailed reasoning and explanations. It encompasses more than 21,000 multiple-choice questions, showcasing a vast range of domains that span three subjects, include 26 topics, cover 127 categories, and incorporate 379 distinct skills. The benchmark dataset is divided into training, validation, and test sets, consisting of 12,726, 4,241, and 4,241 examples, respectively. The second dataset is A-OKVQA [45], a knowledge-based visual question answering dataset that contains 25,000 questions requiring a broad base of commonsense and world knowledge for accurate responses, divided into 17,000 training, 1,000 validation, and 7,000 test examples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.79, + 0.598 + ], + "angle": 0, + "content": "Baselines In our assessment of the ScienceQA dataset, we have compared MC-CoT against recent strong baselines, covering five categories of methodologies: (i) heuristic and expert-guided choices, such as random choice and human evaluation [32]; (ii) standard multimodal visual question answering approaches, which include MCAN [58], Top-Down [1], BAN [25], DFAF [18], ViLT [26], PatchTRM [35], and VisualBERT [28]; (iii) Instruction-tuned large language models like GPT-3.5 [9] and its CoT-enhanced variants [32], in addition to ChatGPT, GPT-4 [39], and Chameleon [34]; (iv) Specifically finetuned large language models, notably LLaMA-Adapter [61], LaVIN [36], LLaMA-SciTune [20], and LLaVa [29]; (v) multimodal reasoning models based on chain-of-thought, Multimodal-CoT [63] and our MC-CoT both belong to this category. Regarding the A-OKVQA dataset, the baselines encompass key methods that have propelled the field of visual question answering forward, such as Pythia [7], ViLBERT [31], LXMERT [46], KRISP [37], GPV-2 [22], BLIP-2 [27], PICa [54], and IPVR [12]. Additionally, Multimodal-CoT [63] is incorporated as the chain-of-thought multimodal benchmark for this dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Implementation Details Following the experimental setup detailed in [63], we implemented the T5 encoder-decoder framework [43]. We utilize the UnifiedQA [24] models, which have 223M and 738M parameters, as our default Base and Large models, respectively. Additionally, we employ the FLAN-T5 [15] models with 248M and 783M parameters as our F-Base and F-Large models. Batch sizes were set to 16 for the base model and 8 for the large model. The learning rate was set to \\(5 \\times 10^{-5}\\). Regarding sequence length, the initial output was constrained to 512 tokens, while the subsequent output was limited to 64 tokens. The experiments utilized four NVIDIA Tesla A100 GPUs, each equipped with 80GB of memory." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.77, + 0.373, + 0.785 + ], + "angle": 0, + "content": "5.1 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We conducted a comparison of MC-CoT with state-of-the-art approaches on the ScienceQA dataset, as depicted in Table 1. We present the overall model size rather than the tunable model size, as it more accurately reflects the model's" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.342, + 0.127 + ], + "angle": 0, + "content": "Tan. et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: The comparison on ScienceQA dataset. Question classes: NAT = natural science, SOC = social science, LAN = language science, TXT = text context, IMG = image context, NO = no context, G1-6 = grades 1-6, G7-12 = grades 7-12." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.199, + 0.785, + 0.647 + ], + "angle": 0, + "content": "
ModelSizeNATSOCLANTXTIMGNOG1-6G7-12AVG
Random Choice [32]-40.2846.1329.2547.7540.0833.6639.3540.6739.83
Human [32]-90.2384.9787.4889.6087.5088.1091.5982.4288.40
MCAN [58]95M56.0846.2358.0959.4351.1755.4051.6559.7254.54
Top-Down [1]70M59.5054.3361.8262.9054.8859.7957.2762.1659.02
BAN [25]112M60.8846.5766.6462.6152.6065.5156.8363.9459.37
DFAF [18]74M64.0348.8263.5565.8854.4964.1157.1267.1760.72
ViLT [26]113M60.4863.8960.2763.2061.3857.0060.7261.9061.14
Patch-TRM [35]90M65.1946.7965.5566.9655.2864.9558.0467.5061.42
VisualBERT [28]111M59.3369.1861.1862.7162.1758.5462.9659.9261.87
UnifiedQABase [24]223M68.1669.1874.9163.7861.3877.8472.9865.0070.12
UnifiedQABase w/ CoT [32]223M71.0076.0478.9166.4266.5381.8177.0668.8274.11
GPT-3.5 [32]173B74.6469.7476.0074.4467.2877.4276.8068.8973.97
GPT-3.5 w/ CoT [32]173B75.4470.8778.0974.6867.4379.9378.2369.6875.17
ChatGPT w/ CoT [39]-78.8270.9883.1877.3767.9286.1380.7274.0378.31
GPT-4 w/ CoT [39]-85.4872.4490.2782.6571.4992.8986.6679.0483.99
Chameleon + ChatGPT [34]-81.6270.6484.0079.7770.8086.6281.8676.5379.93
Chameleon + GPT-4 [34]-89.8374.1389.8288.2777.6492.1388.0383.7286.54
LLaMA-Adapter (T) [61]6B79.0073.7980.5578.3070.3583.1479.7775.6878.31
LLaMA-Adapter [61]6B84.3788.3084.3683.7280.3286.9085.8384.0585.19
LaVIN-7B [36]7B89.2594.9485.2488.5187.4688.0890.1688.0789.41
LLaMA-SciTuneBase [20]7B84.5094.1582.9188.3583.6488.7485.0585.6086.11
LaVIN-13B [36]13B90.3294.3887.7389.4487.6590.3191.1989.2690.50
LLaVa [29]13B90.3695.9588.0089.4988.0090.6690.9390.9090.92
LLaVa + GPT-4 [29]13B91.5696.7491.0990.6288.9993.5292.7392.1692.53
LLaMA-SciTuneLarge [20]13B89.3095.6187.0093.0886.6791.7584.3791.3090.03
Mutimodal-CoTBase [63]223M87.5277.1785.8287.8882.9086.8384.6585.3784.91
Mutimodal-CoTLarge [63]738M95.9182.0090.8295.2688.8092.8992.4490.3191.68
MC-CoTBase223M91.8784.5993.0092.2888.3092.7590.6490.6490.64
MC-CoTLarge738M95.4789.9991.8295.1192.6693.2494.2791.7693.37
MC-CoTf-Base248M93.5683.5890.7394.1389.2490.9490.9390.3890.73
MC-CoTf-Large783M97.4790.4493.1896.9793.7594.4995.3094.1394.88
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.684, + 0.788, + 0.776 + ], + "angle": 0, + "content": "capacity. The results indicate that our MC-CoTBase model, with a model size of only 223 million parameters, achieved an average accuracy of \\(90.64\\%\\), approaching the performance of Lavin-13B and LLaVa-13B models that were fine-tuned based on larger language models. Furthermore, our MC-CoTF-Large model, with 783 million parameters, achieved state-of-the-art results, surpassing the strongest fine-tuned large language model baseline LLaVa+GPT-4 by \\(2.35\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Furthermore, in comparison to multimodal chain-of-thought baselines, our MC-CoTBase model showcased an average accuracy improvement of \\(5.70\\%\\) compared to Multimodal-CoTBase. Similarly, our MC-CoTLarge model exhibited an improvement of \\(1.69\\%\\) compared to the Multimodal-CoT Large model. These" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.669, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MC-CoT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "observations demonstrate that our approach has the potential to significantly enhance the model's reasoning capabilities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.788, + 0.364 + ], + "angle": 0, + "content": "We report the experimental results on the A-OKVQA dataset within Table 2. Our MC-CoTBase model demonstrates exceptional improvements over the current state-of-the-art models, both in terms of direct-answer accuracy and multi-choice task performance. Specifically, MC-CoTBase achieves a direct-answer accuracy of \\(68.7\\%\\), substantially outperforming the strongest competitor, BLIP-2, which possesses an extensive parameter count of over 11 billion. This represents an improvement of over \\(15.5\\%\\), a margin that underscores the efficiency and effectiveness of our model's reasoning capabilities. In the multi-choice task, MC-CoTBase reaches an accuracy of \\(71.0\\%\\), which not only exceeds BLIP-2's performance by \\(0.8\\%\\) but also does so with significantly fewer parameters. This comparison highlights the fact that our model's parameter utilization is exceptionally efficient, leading to enhanced performance even against models with far greater complexity." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.395, + 0.784, + 0.424 + ], + "angle": 0, + "content": "Table 2: The comparison on A-OKVQA dataset. We conduct the evaluation on both direct-answer and multi-choice tasks." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.436, + 0.789, + 0.595 + ], + "angle": 0, + "content": "
ModelVision ModelText ModelParametersDirect-answerMulti-choice
Pythia [7]ResNet [19]BERT [23]70M25.249.0
ViLBERT [31]Faster R-CNN [44]BERT [23]300M30.649.1
LXMERT [46]Transformer [49]Transformer [49]220M30.751.4
KRISP [37]Faster R-CNN [44]BERT [23]200M33.751.9
GPV-2 [22]VinVL [59]T5-Base [42]300M48.660.3
BLIP-2 [27]CLIP-VIT-LARGE [41]FlanT5XXL [15]11B53.270.2
PaLM-CoT [51]-PaLM [13]540B41.548.1
PICa [54]VinVL [59]GPT-3 [10]175B42.446.1
IPVR [12]Faster-RCNN [44]OPT [62]66B46.448.6
Multimodal-CoTBase [63]DETR [11]UnifiedQABase [24]223M-50.6
MC-CoTBaseDETR [11]UnifiedQABase [24]223M68.771.0
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.842 + ], + "angle": 0, + "content": "These results are particularly noteworthy given the complexity and diversity of the A-OKVQA dataset, which is designed to challenge models with questions that require advanced understanding and reasoning across both textual and visual modalities. The performance leap made by \\(\\mathrm{MC - CoT_{Base}}\\) reinforces the potential of our improved training strategy. By introducing a voting mechanism among multiple rationales, our approach refines the quality of the generated rationales, which in turn significantly enhances the accuracy of the final answer. This innovative training paradigm proves effective in a multimodal context, where the interplay between textual and visual elements is critical for achieving high performance. Moreover, the robustness of \\(\\mathrm{MC - CoT_{Base}}\\) suggests that it is not only the size of the model that dictates performance but also the methodological advancements that come with thoughtful design and algorithmic improvements. This is evident from the fact that our base-sized model can outstrip the performance of models with parameter counts orders of magnitude larger." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.342, + 0.127 + ], + "angle": 0, + "content": "Tan. et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.43, + 0.162 + ], + "angle": 0, + "content": "5.2 Quantative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.167, + 0.788, + 0.394 + ], + "angle": 0, + "content": "Ablation Study We present the experimental results on the ScienceQA dataset illustrated in Table 3. It can be seen that both using \\(\\hat{Y}\\) only and \\(\\bar{Y}\\) only can achieve comparable performance but the combined model outperforms them. This indicates that the two voting strategies are complementary to each other. Moreover, we can see that the model without voting R (rationales) obtains a significant drop in performance across all categories, with the average accuracy plummeting to \\(84.70\\%\\). This highlights the importance of the rationale voting in improving the model's reasoning capabilities and accuracy. The absence of the answer voting mechanism results in a decrease in performance in the SOC, LAN, NO, G1-6, and G7-12 categories, with a slight improvement in NAT and TXT. The overall average is \\(90.19\\%\\), indicating that answer voting contributes to the overall robustness, albeit not as critically as rationale voting. Inference voting means that voting is only conducted during the inference stage and not during the training stage. The significant performance drop of this strategy indicates that inference phase voting is not as effective as training phase voting." + }, + { + "type": "table_caption", + "bbox": [ + 0.339, + 0.403, + 0.662, + 0.418 + ], + "angle": 0, + "content": "Table 3: Ablation study on ScienceQA dataset." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.43, + 0.784, + 0.546 + ], + "angle": 0, + "content": "
ModelNATSOCLANTXTIMGNOG1-6G7-12AVG
MC-CoTBase91.8784.5993.0092.2888.3092.7590.6490.6490.64
w/ mean only92.1083.8090.8292.4788.8490.6690.1689.7890.03
w/ weighted only91.6183.6991.0092.3888.4590.8089.7289.9189.79
w/o voting R87.3477.3985.1887.6883.2486.3484.5884.9084.70
w/o voting A92.3684.1490.6492.4288.3091.5090.3589.9190.19
inference voting87.4377.2884.6487.3982.8586.4184.9183.9884.58
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Effect of Rationale Generation We categorized the predictions from the Multimodal-CoTBase and MC-CoTBase models into four types: (i) Good Rationale and Answer; (ii) Bad Rationale and Answer; (iii) Good Rationale but Bad Answer; (iv) Bad Rationale but Good Answer. A 'Good Rationale' indicates the high-quality generated rationale. The results are depicted in Fig. 4. It can be seen that MC-CoT exhibits a higher frequency of Good Rationale and Answer, and a lower frequency of Bad Rationale and Answer. This suggests the self-consistency training approach improves the rationale's quality and the accuracy of the answers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We present the evaluation of the generated rationales and answers on the ScienceQA dataset in Table 4. It can be seen that our proposed MC-CoTBase model generated rationales that improved RougeL scores by approximately \\(1\\%\\) compared to the Multimodal-CoTBase model. However, this led to an approximately \\(6\\%\\) increase in the average accuracy of predicted answers. Similarly, the rationales generated by the MC-CoTF-Large model showed only about a \\(0.5\\%\\) improvement in RougeL scores compared to the MC-CoT base model. Yet, this resulted in an approximate \\(4\\%\\) increase in the average accuracy of predicted answers. This demonstrates that even slight improvements in the quality of rationales can impact the answer inference of the multimodal reasoning models." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.669, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MC-CoT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.147, + 0.489, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.315, + 0.493, + 0.342 + ], + "angle": 0, + "content": "Fig. 4: The relationship between rationale and answer." + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.187, + 0.789, + 0.241 + ], + "angle": 0, + "content": "Table 4: The RougeL score of the generated rationales and the average accuracy of predicted answers on the ScienceQA dataset." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.243, + 0.779, + 0.315 + ], + "angle": 0, + "content": "
MethodRougeLAvg Acc
Multimodal-CoTBase96.9784.91
MC-CoTBase97.9890.64
MC-CoTF-Large98.4794.88
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.374, + 0.43, + 0.389 + ], + "angle": 0, + "content": "5.3 Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.401, + 0.789, + 0.477 + ], + "angle": 0, + "content": "We present two illustrative examples of predictions, as depicted in Figure 5, to demonstrate the efficacy of our proposed model. These examples have been carefully selected to showcase the model's capabilities in handling complex patterns and making accurate choices. For readers interested in exploring further, a more detailed collection of examples is provided in Appendix." + }, + { + "type": "table", + "bbox": [ + 0.28, + 0.506, + 0.408, + 0.582 + ], + "angle": 0, + "content": "
Planet Volume (cm3) Primary composition
Mercury 6.08 x 1015 rock Venus 9.28 x 1015 rock Earth 1.08 x 1012 rock Mars 1.63 x 1014 rock Jupiter 1.43 x 1015 gas Saturn 8.27 x 1014 gas Uranus 6.83 x 1013 ice Neptune 6.25 x 1013 ice
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.228, + 0.585, + 0.419, + 0.594 + ], + "angle": 0, + "content": "Context: Use the data to answer the question below." + }, + { + "type": "table_footnote", + "bbox": [ + 0.228, + 0.598, + 0.459, + 0.632 + ], + "angle": 0, + "content": "Question: Is the following statement about our solar system true or false? Of the four smallest planets, two are made mainly of gas." + }, + { + "type": "table_footnote", + "bbox": [ + 0.228, + 0.633, + 0.266, + 0.644 + ], + "angle": 0, + "content": "Options:" + }, + { + "type": "table_footnote", + "bbox": [ + 0.274, + 0.633, + 0.307, + 0.644 + ], + "angle": 0, + "content": "(A) false" + }, + { + "type": "table_footnote", + "bbox": [ + 0.33, + 0.633, + 0.361, + 0.644 + ], + "angle": 0, + "content": "(B) true" + }, + { + "type": "title", + "bbox": [ + 0.476, + 0.51, + 0.555, + 0.52 + ], + "angle": 0, + "content": "Multimodal-CoT" + }, + { + "type": "text", + "bbox": [ + 0.477, + 0.536, + 0.605, + 0.629 + ], + "angle": 0, + "content": "A planet's volume tells you the size of the planet .... The volumes of Jupiter, Venus, Earth, and Mars have the smallest exponents. So, these four planets are the smallest. The four planets are made mainly of gas. So, of the four smallest planets, two are made mainly of gas.. The answer is (B) true." + }, + { + "type": "title", + "bbox": [ + 0.63, + 0.51, + 0.672, + 0.52 + ], + "angle": 0, + "content": "MC-CoT" + }, + { + "type": "text", + "bbox": [ + 0.634, + 0.536, + 0.766, + 0.629 + ], + "angle": 0, + "content": "A planet's volume tells you the size of the planet .... The volumes of Mercury, Venus, Earth, and Mars have the smallest exponents. So, these four planets are the smallest. These four planets are made mainly of rock. So, of the four smallest planets, none are made mainly of gas.. The answer is (A) false." + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.667, + 0.411, + 0.743 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.228, + 0.745, + 0.425, + 0.754 + ], + "angle": 0, + "content": "Context: Use the graph to answer the question below." + }, + { + "type": "image_footnote", + "bbox": [ + 0.228, + 0.756, + 0.457, + 0.766 + ], + "angle": 0, + "content": "Question: Which month is the hottest on average in Adelaide?" + }, + { + "type": "table_caption", + "bbox": [ + 0.228, + 0.769, + 0.264, + 0.779 + ], + "angle": 0, + "content": "Options:" + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.769, + 0.376, + 0.781 + ], + "angle": 0, + "content": "
(A) June, July, and August
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.276, + 0.784, + 0.345, + 0.793 + ], + "angle": 0, + "content": "(B) April and May" + }, + { + "type": "table_footnote", + "bbox": [ + 0.276, + 0.798, + 0.369, + 0.807 + ], + "angle": 0, + "content": "(C) January and February" + }, + { + "type": "title", + "bbox": [ + 0.476, + 0.671, + 0.555, + 0.681 + ], + "angle": 0, + "content": "Multimodal-CoT" + }, + { + "type": "text", + "bbox": [ + 0.475, + 0.69, + 0.605, + 0.756 + ], + "angle": 0, + "content": "Scientists record climate data from places around the world .... The average temperatures in April and February are around over \\(25^{\\circ}\\mathrm{C}\\) . These months have the highest average temperatures of all of the months. So, they are hotter than the other months.." + }, + { + "type": "text", + "bbox": [ + 0.476, + 0.763, + 0.604, + 0.781 + ], + "angle": 0, + "content": "The answer is (A). June, July, and August." + }, + { + "type": "title", + "bbox": [ + 0.63, + 0.671, + 0.671, + 0.681 + ], + "angle": 0, + "content": "MC-CoT" + }, + { + "type": "text", + "bbox": [ + 0.631, + 0.69, + 0.767, + 0.756 + ], + "angle": 0, + "content": "Scientists record climate data from places around the world .... The average temperatures in January and February are between above \\(20^{\\circ}\\mathrm{F}\\) . These months have the highest average temperatures of all of the months. So, they are hotter than the other months." + }, + { + "type": "text", + "bbox": [ + 0.632, + 0.763, + 0.766, + 0.781 + ], + "angle": 0, + "content": "The answer is (C) January and February." + }, + { + "type": "image_caption", + "bbox": [ + 0.356, + 0.822, + 0.645, + 0.837 + ], + "angle": 0, + "content": "Fig. 5: Comparison on predicted examples." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.342, + 0.127 + ], + "angle": 0, + "content": "Tan. et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "The first example highlights a scenario where the Multimodal-CoT model encountered difficulties in generating a coherent and accurate rationale, ultimately leading to an erroneous conclusion. This misstep underscores the challenges inherent in integrating multiple sources of information and extracting a logical reasoning path. Conversely, the MC-CoT model demonstrated its proficiency by navigating these complexities successfully, thereby arriving at the correct answer. This contrast exemplifies the significance of robust reasoning capabilities in achieving high accuracy in complex question-answering tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.79, + 0.419 + ], + "angle": 0, + "content": "The second example delves into a different set of challenges faced by both models. Here, the Multimodal-CoT model stumbled over the interpretation of data presented in a chart, a mistake that significantly derailed its reasoning process and resulted in an incorrect answer. This error highlights the critical importance of accurate data interpretation skills in models designed for multimodal contexts. On the other hand, the MC-CoT model, while generally more adept in this instance, exhibited a lapse in understanding units of measurement—mistaking degrees Celsius \\((^{\\circ}\\mathrm{C})\\) for degrees Fahrenheit \\((^{\\circ}\\mathrm{F})\\). Interestingly, this error did not compromise the integrity of its final answer, suggesting a certain resilience in the model's reasoning process despite minor inaccuracies." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.444, + 0.36, + 0.46 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.789, + 0.735 + ], + "angle": 0, + "content": "In this study, we conducted an exhaustive analysis of the Multimodal-CoT framework and identified that the quality of rationales plays a crucial role in enhancing the performance of multimodal reasoning models. Recognizing the critical importance of this aspect, we developed an innovative solution aimed at improving the quality of rationales. Our solution employs a self-consistency training strategy that leverages the natural variability introduced by dropout techniques during training to foster a consensus mechanism through voting. This approach is supported by both theoretical underpinnings and empirical evidence, highlighting its effectiveness and wide applicability. We meticulously designed experiments to evaluate the impact of our self-consistency training strategy on the performance of multimodal reasoning models across diverse scenarios and datasets. The results from these experiments were revealing, clearly showing that our method significantly boosts the capabilities of multimodal models. A particularly significant finding was our approach's ability to enhance the performance of smaller models, enabling them to match or even exceed the capabilities of much larger models. We believe our strategy offers a promising avenue for advancing multimodal reasoning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.79, + 0.84 + ], + "angle": 0, + "content": "A limitation of our approach is that we do not explicitly impose grammatical constraints, meaning the generated rationales are not guaranteed to be grammatically correct. It is worth noting that the Multimodal-CoT framework also does not guarantee the grammatical validity of the generated rationales. Despite this, empirical evidence has shown that our approach can yield superior results across various datasets, indicating that the rationales produced by our approach can significantly enhance the model's reasoning ability towards the correct answers." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.669, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MC-CoT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.427, + 0.163 + ], + "angle": 0, + "content": "7 Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.296 + ], + "angle": 0, + "content": "This work was supported by the Science & Technology Innovation 2030 Major Program Project No. 2021ZD0150100, National Natural Science Foundation of China Project No. U21A20427, Project No. WU2022A009 from the Center of Synthetic Biology and Integrated Bioengineering of Westlake University, Project No. WU2023C019 from the Westlake University Industries of the Future Research. Finally, we thank the Westlake University HPC Center for providing part of the computational resources, and Project No. 23-407-3-29 from the Shenyang Science and Technology Program." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.317, + 0.323, + 0.333 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.346, + 0.787, + 0.401 + ], + "angle": 0, + "content": "1. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6077-6086 (2018) 4, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.402, + 0.787, + 0.442 + ], + "angle": 0, + "content": "2. Anil, R., Dai, A.M., First, O., Johnson, M., Lepikhin, D., Passos, A., Shakeri, S., Taropa, E., Bailey, P., Chen, Z., et al.: Palm 2 technical report. arXiv preprint arXiv:2305.10403 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.443, + 0.787, + 0.484 + ], + "angle": 0, + "content": "3. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: Vqa: Visual question answering. In: Proceedings of the IEEE international conference on computer vision. pp. 2425-2433 (2015) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.485, + 0.787, + 0.525 + ], + "angle": 0, + "content": "4. Askell, A., Bai, Y., Chen, A., Drain, D., Ganguli, D., Henighan, T., Jones, A., Joseph, N., Mann, B., DasSarma, N., et al.: A general language assistant as a laboratory for alignment. arXiv preprint arXiv:2112.00861 (2021) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.526, + 0.787, + 0.566 + ], + "angle": 0, + "content": "5. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.567, + 0.787, + 0.62 + ], + "angle": 0, + "content": "6. Besta, M., Blach, N., Kubicek, A., Gerstenberger, R., Gianinazzi, L., Gajda, J., Lehmann, T., Podstawski, M., Niewiadomski, H., Nczyk, P., et al.: Graph of thoughts: Solving elaborate problems with large language models. arXiv preprint arXiv:2308.09687 (2023) 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.621, + 0.787, + 0.675 + ], + "angle": 0, + "content": "7. Biderman, S., Schoelkopf, H., Anthony, Q.G., Bradley, H., O'Brien, K., Hallahan, E., Khan, M.A., Purohit, S., Prashanth, U.S., Raff, E., et al.: Pythia: A suite for analyzing large language models across training and scaling. In: International Conference on Machine Learning. pp. 2397-2430. PMLR (2023) 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.676, + 0.787, + 0.729 + ], + "angle": 0, + "content": "8. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.787, + 0.784 + ], + "angle": 0, + "content": "9. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.785, + 0.787, + 0.838 + ], + "angle": 0, + "content": "10. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) 11" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.346, + 0.787, + 0.838 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.342, + 0.127 + ], + "angle": 0, + "content": "Tan. et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "11. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.231 + ], + "angle": 0, + "content": "12. Chen, Z., Zhou, Q., Shen, Y., Hong, Y., Zhang, H., Gan, C.: See, think, confirm: Interactive prompting between vision and language models for knowledge-based visual reasoning. arXiv preprint arXiv:2301.05226 (2023) 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.787, + 0.272 + ], + "angle": 0, + "content": "13. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311 (2022) 1, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "14. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, E., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "15. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022) 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.357, + 0.787, + 0.397 + ], + "angle": 0, + "content": "16. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Qiu, Z., Lin, W., Yang, J., Zheng, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.398, + 0.787, + 0.438 + ], + "angle": 0, + "content": "17. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.439, + 0.787, + 0.493 + ], + "angle": 0, + "content": "18. Gao, P., Jiang, Z., You, H., Lu, P., Hoi, S.C., Wang, X., Li, H.: Dynamic fusion with intra-and inter-modality attention flow for visual question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6639-6648 (2019) 4, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.495, + 0.787, + 0.535 + ], + "angle": 0, + "content": "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.536, + 0.787, + 0.577 + ], + "angle": 0, + "content": "20. Horawalavithana, S., Munikoti, S., Stewart, I., Kvinge, H.: Scitune: Aligning large language models with scientific multimodal instructions. arXiv preprint arXiv:2307.01139 (2023) 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.578, + 0.787, + 0.604 + ], + "angle": 0, + "content": "21. Huang, J., Chang, K.C.C.: Towards reasoning in large language models: A survey. arXiv preprint arXiv:2212.10403 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "22. Kamath, A., Clark, C., Gupta, T., Kolve, E., Hoiem, D., Kembhavi, A.: Webly supervised concept expansion for general purpose vision models. In: European Conference on Computer Vision. pp. 662-681. Springer (2022) 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "23. Kenton, J.D.M.W.C., Toutanova, L.K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of naacL-HLT. vol. 1, p. 2 (2019) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.689, + 0.787, + 0.77 + ], + "angle": 0, + "content": "24. Khashabi, D., Min, S., Khot, T., Sabharwal, A., Tafjord, O., Clark, P., Hajishirzi, H.: UNIFIEDQA: Crossing format boundaries with a single QA system. In: Cohn, T., He, Y., Liu, Y. (eds.) Findings of the Association for Computational Linguistics: EMNLP 2020. pp. 1896-1907. Association for Computational Linguistics, Online (Nov 2020). https://doi.org/10.18653/v1/2020-findings-emnlp.171, https://aclanthology.org/2020_findings-emnlp.171_9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "25. Kim, J.H., Jun, J., Zhang, B.T.: Bilinear attention networks. Advances in neural information processing systems 31 (2018) 4, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "26. Kim, W., Son, B., Kim, I.: Vilt: Vision-and-language transformer without convolution or region supervision. In: International Conference on Machine Learning. pp. 5583-5594. PMLR (2021) 4, 9, 10" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.669, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MC-CoT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "27. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) 4, 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.191, + 0.787, + 0.233 + ], + "angle": 0, + "content": "28. Li, L.H., Yatskar, M., Yin, D., Hsieh, C.J., Chang, K.W.: What does bert with vision look at? In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. pp. 5265-5275 (2020) 4, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.786, + 0.261 + ], + "angle": 0, + "content": "29. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023) 2, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.787, + 0.303 + ], + "angle": 0, + "content": "30. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.787, + 0.346 + ], + "angle": 0, + "content": "31. Lu, J., Batra, D., Parikh, D., Lee, S.: Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems 32 (2019) 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.787, + 0.402 + ], + "angle": 0, + "content": "32. Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems 35, 2507-2521 (2022) 2, 3, 4, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.787, + 0.445 + ], + "angle": 0, + "content": "33. Lu, P., Peng, B., Cheng, H., Galley, M., Chang, K.W., Wu, Y.N., Zhu, S.C., Gao, J.: Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.787, + 0.487 + ], + "angle": 0, + "content": "34. Lu, P., Peng, B., Cheng, H., Galley, M., Chang, K.W., Wu, Y.N., Zhu, S.C., Gao, J.: Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842 (2023) 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.53 + ], + "angle": 0, + "content": "35. Lu, P., Qiu, L., Chen, J., Xia, T., Zhao, Y., Zhang, W., Yu, Z., Liang, X., Zhu, S.C.: Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. arXiv preprint arXiv:2110.13214 (2021) 4, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.787, + 0.572 + ], + "angle": 0, + "content": "36. Luo, G., Zhou, Y., Ren, T., Chen, S., Sun, X., Ji, R.: Cheap and quick: Efficient vision-language instruction tuning for large language models. arXiv preprint arXiv:2305.15023 (2023) 2, 4, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.787, + 0.628 + ], + "angle": 0, + "content": "37. Marino, K., Chen, X., Parikh, D., Gupta, A., Rohrbach, M.: Krisp: Integrating implicit and symbolic knowledge for open-domain knowledge-based vqa. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14111-14121 (2021) 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.787, + 0.657 + ], + "angle": 0, + "content": "38. Ning, X., Lin, Z., Zhou, Z., Yang, H., Wang, Y.: Skeleton-of-thought: Large language models can do parallel decoding. arXiv preprint arXiv:2307.15337 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.561, + 0.671 + ], + "angle": 0, + "content": "39. OpenAI: Gpt-4 technical report (2023) 1, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.787, + 0.728 + ], + "angle": 0, + "content": "40. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.787, + 0.784 + ], + "angle": 0, + "content": "41. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "42. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020) 1, 11" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.342, + 0.127 + ], + "angle": 0, + "content": "Tan. et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "43. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.231 + ], + "angle": 0, + "content": "44. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems 28 (2015) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.787, + 0.273 + ], + "angle": 0, + "content": "45. Schwenk, D., Khandelwal, A., Clark, C., Marino, K., Mottaghi, R.: A-okvqa: A benchmark for visual question answering using world knowledge. In: European Conference on Computer Vision. pp. 146-162. Springer (2022) 3, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "46. Tan, H., Bansal, M.: Lxmert: Learning cross-modality encoder representations from transformers. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (2019) 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "47. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.787, + 0.397 + ], + "angle": 0, + "content": "48. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.438 + ], + "angle": 0, + "content": "49. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.493 + ], + "angle": 0, + "content": "50. Wang, X., Wei, J., Schuurmans, D., Le, Q.V., Chi, E.H., Narang, S., Chowdhery, A., Zhou, D.: Self-consistency improves chain of thought reasoning in language models. In: The Eleventh International Conference on Learning Representations (2022) 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.787, + 0.548 + ], + "angle": 0, + "content": "51. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022) 1, 3, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.549, + 0.787, + 0.577 + ], + "angle": 0, + "content": "52. Wu, C., Liu, J., Wang, X., Dong, X.: Chain of reasoning for visual question answering. Advances in Neural Information Processing Systems 31 (2018) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.577, + 0.787, + 0.618 + ], + "angle": 0, + "content": "53. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.787, + 0.66 + ], + "angle": 0, + "content": "54. Yang, Z., Gan, Z., Wang, J., Hu, X., Lu, Y., Liu, Z., Wang, L.: An empirical study of gpt-3 for few-shot knowledge-based vqa. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 3081-3089 (2022) 9, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.787, + 0.702 + ], + "angle": 0, + "content": "55. Yang, Z., Li, L., Wang, J., Lin, K., Azarnasab, E., Ahmed, F., Liu, Z., Liu, C., Zeng, M., Wang, L.: Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.787, + 0.742 + ], + "angle": 0, + "content": "56. Yao, S., Yu, D., Zhao, J., Shafran, I., Griffiths, T.L., Cao, Y., Narasimhan, K.: Tree of thoughts: Deliberate problem solving with large language models. arXiv preprint arXiv:2305.10601 (2023) 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.787, + 0.771 + ], + "angle": 0, + "content": "57. Yin, S., Fu, C., Zhao, S., Li, K., Sun, X., Xu, T., Chen, E.: A survey on multimodal large language models. arXiv preprint arXiv:2306.13549 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.787, + 0.813 + ], + "angle": 0, + "content": "58. Yu, Z., Yu, J., Cui, Y., Tao, D., Tian, Q.: Deep modular co-attention networks for visual question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6281-6290 (2019) 4, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "59. Zhang, P., Li, X., Hu, X., Yang, J., Zhang, L., Wang, L., Choi, Y., Gao, J.: Vinvl: Revisiting visual representations in vision-language models. In: Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.669, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "MC-CoT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "IEEE/CVF conference on computer vision and pattern recognition. pp. 5579-5588 (2021) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.788, + 0.217 + ], + "angle": 0, + "content": "60. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199 (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "61. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199 (2023) 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.3 + ], + "angle": 0, + "content": "62. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022) 1, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.787, + 0.342 + ], + "angle": 0, + "content": "63. Zhang, Z., Zhang, A., Li, M., Zhao, H., Karypis, G., Smola, A.: Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923 (2023) 2, 4, 5, 6, 9, 10, 11" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.342 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Boosting the Power of Small Multimodal Reasoning Models to Match Larger Models with Self-Consistency Training/cda53ae2-a211-4097-a2a7-2df98969f4ef_origin.pdf b/2024/Boosting the Power of Small Multimodal Reasoning Models to Match Larger Models with Self-Consistency Training/cda53ae2-a211-4097-a2a7-2df98969f4ef_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..261a6990fe1e88ba451ef8c0b1076b90f0d125d6 --- /dev/null +++ b/2024/Boosting the Power of Small Multimodal Reasoning Models to Match Larger Models with Self-Consistency Training/cda53ae2-a211-4097-a2a7-2df98969f4ef_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8fdb95b0b5998ccd50b58e256500ab8984695d4a5592fd69de28f703be2f7e61 +size 777515 diff --git a/2024/Boosting the Power of Small Multimodal Reasoning Models to Match Larger Models with Self-Consistency Training/full.md b/2024/Boosting the Power of Small Multimodal Reasoning Models to Match Larger Models with Self-Consistency Training/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ba0ec61028c18c02fac4162f4628dbbba75a18f9 --- /dev/null +++ b/2024/Boosting the Power of Small Multimodal Reasoning Models to Match Larger Models with Self-Consistency Training/full.md @@ -0,0 +1,361 @@ +# Boosting the Power of Small Multimodal Reasoning Models to Match Larger Models with Self-Consistency Training + +Cheng Tan $^{1,4,5\star}$ , Jingxuan Wei $^{2,3\star\boxtimes}$ , Zhangyang Gao $^{1,4,5\star}$ , Linzhuang Sun $^{2,3}$ , Siyuan Li $^{1,4,5}$ , Ruifeng Guo $^{2,3}$ , Bihui Yu $^{2,3}$ , and Stan Z. Li $^{4\boxtimes}$ + +$^{1}$ Zhejiang University + $^{2}$ Shenyang Institute of Computing Technology, Chinese Academy of Sciences + $^{3}$ University of Chinese Academy of Sciences +$^{4}$ AI Lab, Research Center for Industries of the Future, Westlake University +$^{5}$ Institute of Advanced Technology, Westlake Institute for Advanced Study tancheng@westlake.edu.cn, weijingxuan20@mails.ucas.edu.cn $\star$ Equal contribution, $\boxtimes$ Corresponding author + +Abstract. Multimodal reasoning is a challenging task that requires models to reason across multiple modalities to answer questions. Existing approaches have made progress by incorporating language and visual modalities into a two-stage reasoning framework, separating rationale generation from answer inference. However, these approaches often fall short due to the inadequate quality of the generated rationales. In this work, we delve into the importance of rationales in model reasoning. We observe that when rationales are completely accurate, the model's accuracy significantly improves, highlighting the need for high-quality rationale generation. Motivated by this, we propose MC-CoT, a self-consistency training strategy that generates multiple rationales and answers, subsequently selecting the most accurate through a voting process. This approach not only enhances the quality of generated rationales but also leads to more accurate and robust answers. Through extensive experiments, we demonstrate that our approach significantly improves model performance across various benchmarks. Remarkably, we show that even smaller base models, when equipped with our proposed approach, can achieve results comparable to those of larger models, illustrating the potential of our approach in harnessing the power of rationales for improved multimodal reasoning. The code is available at github.com/chengtan9907/mc-cot. + +Keywords: Multimodal reasoning $\cdot$ visual question answering + +# 1 Introduction + +Recent advances in large language models [2, 4, 8, 13, 14, 39, 40, 42, 47, 48, 62] have led to the exploration of Chain-of-Thought (CoT) prompting [6, 50, 51, 56]. This approach, which directs the model to systematically unravel rationales before providing answers, rather than responding directly, has showcased the model's + +![](images/a89124767fb747bb36370ebf0917cd03d167d845edf0d2729bf1601e7253aac0.jpg) +Fig. 1: An example of multimodal reasoning that answers the question by reasoning across both vision and language modalities. + +![](images/ba5d01f034521d1a765437c9b1c8a328c8b2f2b979ba3ba88f8c0f26e35988e6.jpg) + +impressive efficacy across a variety of natural language processing (NLP) tasks. Moreover, the advent of CoT prompting has catalyzed a plethora of research endeavors delving into the reasoning prowess of large language models. A diverse range of Chain-of-Thought strategies has been investigated, including the voting-facilitated CoT-SC [50], the transition from chain-like to tree-like thinking with Tree-of-Thoughts [56], and further expansion into graph-structured thinking [6]. + +While CoT reasoning has been thoroughly established in the realm of language models, its foray into the vast and intricate landscape of multimodal reasoning is still in its infancy. As shown in Fig. 1, Multimodal reasoning [5, 16, 17, 21, 29, 30, 32, 33, 36, 53, 57, 60], which inherently involves the seamless fusion of information from disparate modalities such as text and images, presents unique challenges. The process of extracting, correlating, and generating rationales across multiple modalities is decidedly more complex than the tasks encountered in a solely text-based modality. A recent seminal work, Multimodal-CoT [63], has pioneered the application of the Chain-of-Thought prompting strategy to multimodal reasoning tasks. This approach encompasses a two-stage framework that distinctively separates rationale generation from answer inference. By obliging the model to generate rationales prior to answering questions, Multimodal-CoT mirrors the language-only CoT prompting, thus enabling reasoning across multiple modalities. + +Despite Multimodal-CoT has made promising strides in the realm of multimodal reasoning, as evidenced in Fig. 2, its improvements over the no-rationale baseline are still limited. Moreover, compared to the ground-truth rationale, the predicted rationale falls short, often yielding results that lack relevance to the posed question. This discrepancy primarily stems from the quality of the generated rationales, highlighting the crucial role of rationale quality in the success of the chain-of-thought reasoning process. + +In light of the above observations, it becomes evident that the potency of the Chain-of-Thought reasoning in a multimodal scenario is intrinsically tethered to the accuracy of the rationales generated. A high-quality rationale not only sheds light on the model's thought process but also paves the way for more precise answers. This leads to a critical question: how can we develop a strategy to enhance the quality of these rationales, and consequently, improve the overall performance of multimodal CoT? + +![](images/35433dbd48d7b8159bd6cdb1801e57e751418a77099e9d857d7c8f6c54d21bfd.jpg) +Fig. 2: The comparison of answer accuracy on ScienceQA using the Multimodal-CoT framework with no rationale, predicted rationales, and ground-truth rationales. + +Our study is driven by the hypothesis that enhancing the quality of rationale generation can significantly improve the model's reasoning capabilities and overall performance. To this end, we introduce a simple yet effective strategy that capitalizes on the inherent variability of deep neural models during training, particularly stemming from mechanisms like dropout. Our approach involves having the model generate multiple rationales and then voting for the most consistent words across these rationales to yield a more refined and accurate rationale. The same voting mechanism is then applied to answer generation, further boosting the model's confidence in its predictions. It is important to note that the inference phase remains completely unaffected by this voting mechanism and continues to operate in the same manner as it does in the original Multimodal CoT framework. Through this approach, we aim to facilitate a more robust and accurate multimodal reasoning ability. + +Extensive experiments across ScienceQA [32] and A-OKVQA [45] benchmark datasets demonstrate the efficacy of our proposed approach. Notably, by improving the rationale quality, even smaller models equipped with our strategy manifest performance metrics that rival, and at times surpass, those of considerably larger models. This not only confirms the efficacy of our rationale refinement strategy but also opens up a promising avenue where smaller, more efficient models can be rendered competitive in the multimodal reasoning landscape. + +# 2 Related Work + +# 2.1 Chain-of-Thought + +The Chain-of-Thought (CoT) paradigm has emerged as a transformative approach, aiming to elucidate the reasoning processes of large language models and boost their ability in a variety of NLP tasks. The vanilla CoT [51] prompting strategy models to generate intermediate reasoning steps, leading to significant + +performance improvements across a spectrum of tasks. Its methodology laid the groundwork for subsequent advancements in the CoT paradigm. CoT-SC [50] introduced a self-consistency decoding strategy by sampling multiple reasoning paths and selecting the most consistent path to enhance the reliability of the generated rationales. + +The versatility and adaptability of the CoT paradigm were further demonstrated with the advent of advanced reasoning structures. Tree-of-Thoughts (ToT) [56] transitioned from linear chains to more intricate tree-like structures, aiming to capture more complex reasoning patterns. In a parallel development, [6] ventured into graph-based reasoning, presenting the Graph-of-Thought (GoT) approach, which allowed for a more interconnected reasoning process. Skeleton-of-Thought (SoT) [38] took a different route by emphasizing the benefits of efficiency by underscoring the potential of parallel decoding, enabling models to explore multiple reasoning paths simultaneously. + +While the aforementioned studies have made significant strides in the CoT paradigm, we seek to extend these advancements into the realm of multimodal reasoning. Multimodal-CoT [63] is a seminal work, but the unsatisfying performance of its generated rationales has been a major bottleneck. Our work aims to fully explore the potential of the CoT paradigm in multimodal reasoning. + +# 2.2 Multimodal Visual Question Answering + +Multimodal Visual Question Answering (VQA) has emerged as a key research area bridging the domains of vision and language. Pioneering works like MCAN [58], BAN [25], and Top-Down [1] have established a foundation by introducing diverse attention mechanisms. Building on this, DFAF [18] introduced innovative dynamic fusion strategies to facilitate both intra- and inter-modality attention flows. The introduction of Transformer architectures has significantly impacted the VQA field, as demonstrated by state-of-the-art models such as ViLT [26] and VisualBERT [28]. In the same vein, Patch-TRM [35] leveraged a pyramid cross-modal Transformer, paired with pre-trained input embeddings from the icon dataset, advancing the understanding of abstract diagrams. Recently, incorporating large language models into the multimodal framework has seen a surge, exemplified by models like LLaMA-Adapter [60] and BLIP-2 [27]. LaVIN [36] introduced a novel routing algorithm that helps the model autonomously switch between unimodal and multimodal instruction reasoning paths. These models employ frozen visual encoders paired with fine-tuned language models, thereby setting new standards in VQA tasks. + +ScienceQA [32] introduced a reasoning process into the VQA task, setting a benchmark for multimodal chain-of-thought reasoning. MM-REACT [55], with its innovative prompting design, enables language models to process multimodal information, facilitating the integration of ChatGPT with visual experts. A-OKVQA [45] requires commonsense reasoning about the depicted scene in an image. Multimodal-CoT [63] integrates language (text) and vision (image) modalities into a two-stage framework, distinctly separating rationale generation from answer inference. + +In this work, we typically focus on multimodal reasoning tasks, a subtype of multimodal visual question answering (VQA) that incorporates rationales. The significance of this task lies in its ability to more accurately simulate human cognitive processes. When humans answer questions, they often rely on observed visual information and existing textual knowledge, synthesizing the two to form a reasoned conclusion. These multimodal reasoning tasks comprehensively assess a model's capacity to understand and fuse visual and textual information, thereby propelling the development of multimodal learning. Moreover, these tasks enhance the interpretability of artificial intelligence systems. Because models must supply not only answers but also supporting rationales, their decision-making processes become more interpretable and comprehensible. + +# 3 Method + +# 3.1 Preliminaries + +In this work, we focus on multimodal reasoning, a challenging task that involves reasoning across multiple modalities to answer questions. Specifically, we consider visual question answering [3, 52], where the model is tasked with answering questions based on the information provided in the question itself and the accompanying image. + +Formally, given a dataset $\mathcal{D} = \{\mathcal{X},\mathcal{Y}\}$ , where $X\in \mathcal{X}$ represents a multimodal input consisting of text and images, and $Y\in \mathcal{V}$ represents the corresponding output, the task of multimodal reasoning involves learning a mapping function parameterized by $\Theta$ : $\mathcal{F}_{\Theta}:\mathcal{X}\to \mathcal{V}$ that accurately predicts the output $Y$ for a given input $X$ . The multimodal input $X$ can be represented as $X = (T,I)$ , where $T$ is the text component, and $I$ is the image component of the input. The function $\mathcal{F}$ should be able to effectively leverage the information from both the text and image modalities to make accurate predictions. Specifically, the basic visual question answering task is defined as: + +$$ +Y = \arg \max _ {Y ^ {\prime}} p \left(Y ^ {\prime} \mid T, I\right) \tag {1} +$$ + +where $p(Y' \mid T, I)$ is the probability of answer $Y'$ given the text $T$ and image $I$ . The multimodal reasoning task builds upon the basic visual question answering task, extending its requirements to necessitate the generation of a rationale $R'$ that elucidates the reasoning process underpinning the answer $Y$ . This task can be mathematically represented as follows: + +$$ +Y, R = \arg \max _ {Y ^ {\prime}, R ^ {\prime}} p \left(Y ^ {\prime}, R ^ {\prime} \mid T, I\right) \tag {2} +$$ + +Multimodal-CoT [63] decomposed the multimodal reasoning task into a two-stage framework, where the model first generates a rationale $R$ and then utilizes this rationale to predict the answer $Y$ . The process is delineated as: + +$$ +R = \arg \max _ {R ^ {\prime}} p \left(R ^ {\prime} \mid T, I\right) \tag {3} +$$ + +$$ +Y = \arg \max _ {Y ^ {\prime}} p (Y ^ {\prime} | R, T, I) +$$ + +![](images/8aa890f28b192b53fa4f8f9d2164f950d793374a17edcc096e4a0f6f3c80ec72.jpg) +(a) Vanilla Prompt + +![](images/6217550fc74f3491f094618f9361a7e9db0a7b81e3b730ab9512799130caf1e2.jpg) +(b) CoT + +![](images/874a7e6b31956ba881aab82b289a377849e49776a704aefda476d44a02339608.jpg) +(c) CoT-SC + +![](images/2f903064a751c15bfbf661e3b7f428015888d35a55209d180173ccf9b8946d61.jpg) +(d) Multimodal-CoT +Fig. 3: A comparison schematic diagram of different Chain-of-Thought (CoT) prompt-based reasoning methods. (a) The basic input-output prompt, (b) Chain-of-Thought with intermediate chain-like reasoning, (c) Chain-of-Thought Self-Consistency (CoT-SC) that utilizes the consistency of multiple independent chains of thoughts for reasoning, (d) Multimodal-CoT, which infers the rationale using the input text and image, and then predicts the answer using the rationale as part of the input, and (e) MC-CoT that infers a high-quality rationale through word-level voting, and then obtains a high-quality answer using majority vote. It is worth noting that our approach leverages multiple chain consistency only during the training phase, in contrast to CoT-SC, which employs it during the inference stage. + +![](images/cdda3e35ff217268fa2b311aa2590b3e37e0ebab713a38208cf1fb47e2cde73a.jpg) +(e) MC-CoT + +In this work, we follow the same two-stage framework as Multimodal-CoT [63] and explore the potential of such a chain-of-thought reasoning paradigm. + +# 3.2 Multimodal Consistent Chain-of-Thought + +Though the existing multimodal CoT framework has been proven to be beneficial in multimodal reasoning, there remains a large room to fully explore its potential. In this work, we introduce a simple yet effective training strategy, Multimodal Consistent Chain-of-Thought (MC-CoT), to learn high-quality rationales and thus boost the reasoning performance of the multimodal CoT framework. + +Rationale Generation Leveraging the inherent randomness introduced by the dropout operations, the model is able to generate multiple diverse rationales for a given input text-image pair $(T,I)$ by sampling from the model $N_{r}$ times: + +$$ +R ^ {i} \sim p (R \mid T, I), i = 1, 2, \dots , N _ {r} \tag {4} +$$ + +We perform a voting process on the generated rationales to select the most consistent words across the rationales. Specifically, the best rationale $R^{*}$ is selected as follows: + +$$ +R _ {j} ^ {*} = \operatorname {V o t e} \left(\left\{R _ {j} ^ {i} \right\} _ {1} ^ {N _ {r}}\right), \tag {5} +$$ + +where, for each position $j$ in the rationale, we choose the majority of the words at position $j$ across the $N_r$ generated rationales to form the best rationale $R_j^*$ . + +Answer Inference In the answer inference stage, we also generate multiple answers based on the best rationale $R^{*}$ and the input text-image pair $(T, I)$ by sampling from the model $N_{a}$ times: + +$$ +Y ^ {i} \sim p (Y \mid R ^ {*}, T, I), i = 1, 2, \dots , N _ {a}. \tag {6} +$$ + +The final answer is selected by the majority voting: + +$$ +Y ^ {*} = \operatorname {V o t e} \left(\left\{Y ^ {i} \right\} _ {1} ^ {N _ {a}}\right). \tag {7} +$$ + +where the majority of the answers across the $N_{a}$ generated answers are regarded as the most convincing answers $Y^{*}$ . The pseudocode of the pipeline is shown in Appendix. + +Voting Strategy Consider the answer inference stage as an illustration. The vanilla Multimodal-CoT model computes a cross-entropy loss between a $d$ -dimensional logit $L \in \mathbb{R}^d$ and the ground truth label $Y \in \mathbb{R}^d$ , which can be represented as: + +$$ +\mathcal {L} = \operatorname {C r o s s E n t r o p y} (L, Y), \tag {8} +$$ + +Our averaging-based voting strategy initially calculates the mean logit $\bar{L} \in \mathbb{R}^d$ and weighted mean logit $\hat{L} \in \mathbb{R}^d$ : + +$$ +\bar {L} = \frac {1}{N _ {a}} \sum_ {i = 1} ^ {N _ {a}} L ^ {i}, \hat {L} = \frac {1}{N _ {a}} \frac {\sum_ {i = 1} ^ {N _ {a}} w L ^ {i}}{\sum_ {j} w _ {j}}, \tag {9} +$$ + +where $w = 1 / (1 + \sigma) \in \mathbb{R}^d$ is a weight vector derived from the standard deviation of the logits, calculated across different predictions for each dimension. The weight for each logit's dimension is inversely proportional to its variability, assigning higher confidence to predictions with lower variability and lower confidence to those with higher variability. The final prediction is then a linear combination of $\bar{L}$ and $\hat{L}$ : + +$$ +L ^ {*} = \alpha \bar {L} + (1 - \alpha) \hat {L} \tag {10} +$$ + +with $\alpha$ being a hyperparameter set empirically to 0.5. The cross-entropy loss is subsequently computed between $L^{*}$ and the ground truth label $Y$ : + +$$ +\mathcal {L} = \operatorname {C r o s s E n t r o p y} \left(L ^ {*}, Y\right). \tag {11} +$$ + +We present a comparison of our approach with existing CoT-based reasoning methods in Fig. 3. Specifically, subfigures (a-c) illustrate the prompt learning methods commonly employed for text modality in large language models, while subfigures (d-e) depict the CoT reasoning frameworks in the multimodal field. It can be seen that our approach shares similarities with CoT-SC, such as the utilization of a voting mechanism. However, there are also significant differences between the two approaches: (i) the objective differs. CoT-SC aims to facilitate prompt learning in large language models during the inference stage, while our method seeks to enhance the robustness of model reasoning during the training phase; (ii) the voting process differs. CoT-SC only votes on the final answers inferred from each thought, whereas our method incorporates voting during the rationale generation stage. + +# 4 Theoretical Insights + +# 4.1 Aggregation Minimizes Expected Loss + +Theorem 1. (Jensen's inequality) Let $X$ be a random variable and let $\phi$ be a convex function. Then the following inequality holds: $\phi(\mathbb{E}[X]) \leq \mathbb{E}[\phi(X)]$ . + +Our proposed method, MC-CoT, leverages the variability introduced by dropout techniques to generate a diverse set of rationales. By aggregating these rationales, we aim to produce explanations and corresponding answers of higher quality. In the context of our framework, the convex function $\phi$ is represented by the cross-entropy loss function, and the random variable in question is the set of logits $L^i$ . By applying Jensen's inequality to our model, we deduce that the aggregation process leads to a lower expected loss: $\mathcal{L}(\mathbb{E}(L^{*}),Y)\leq \mathbb{E}(\mathcal{L}(L^{*},Y))$ . + +# 4.2 Bias-Variance Trade-off + +We introduce a linear combination of two voting strategies to derive the final prediction, as delineated in Equation 10. We can analytically decompose the expected value of the squared difference between the predicted outcomes $Y^{*}$ and the ground truth $Y$ into three distinct components: + +$$ +\mathbb {E} \left[ \left(Y ^ {*} - Y\right) ^ {2} \right] = \operatorname {B i a s} ^ {2} \left(Y ^ {*}\right) + \operatorname {V a r} \left(Y ^ {*}\right) + \epsilon , \tag {12} +$$ + +where $\mathrm{Bias}^2 (Y^*) = (\mathbb{E}[Y^*] - Y)^2$ quantifies the systematic deviation of the expected prediction from the ground truth, indicative of the error inherently introduced by the predictive model itself. Meanwhile, $\mathrm{Var}(Y^{*}) = \mathbb{E}[(Y^{*} - \mathbb{E}[Y^{*}])^{2}]$ captures the variability of the predictions, reflecting the extent to which these predictions will fluctuate around their expected value. Lastly, $\epsilon$ denotes the irreducible error, encapsulating the intrinsic noise within the data. + +The mean logits, denoted by $\bar{L}$ , may exhibit lower bias since they represent the composite prediction averaged across multiple models. Employing solely the mean logits promotes exploration by treating all predictions as equally informative. Nevertheless, its measure of central tendency does not account for the individual variability within the predictions. The weighted mean logits, denoted by $\hat{L}$ , aim to reduce variability by assigning greater significance to consistent (low-variance) predictions. This reflects a confidence-based strategy, allocating higher weight to predictions that are less variable. However, it could introduce bias if predictions are consistently erroneous. + +The final logits, which include both mean logits and weighted mean logits, are designed to achieve an optimal balance between bias and variance. While the mean logits promote exploration by treating all predictions equally, the weighted mean logits perform exploitation by assigning higher weights to consistent predictions. The combination allows balancing exploration with exploitation, thus improving the performance towards an optimal standard. + +# 5 Experiment + +Datasets We evaluated MC-CoT on two prominent multimodal reasoning benchmarks. The first, ScienceQA [32], is a comprehensive multimodal science question dataset that contains annotated answers with detailed reasoning and explanations. It encompasses more than 21,000 multiple-choice questions, showcasing a vast range of domains that span three subjects, include 26 topics, cover 127 categories, and incorporate 379 distinct skills. The benchmark dataset is divided into training, validation, and test sets, consisting of 12,726, 4,241, and 4,241 examples, respectively. The second dataset is A-OKVQA [45], a knowledge-based visual question answering dataset that contains 25,000 questions requiring a broad base of commonsense and world knowledge for accurate responses, divided into 17,000 training, 1,000 validation, and 7,000 test examples. + +Baselines In our assessment of the ScienceQA dataset, we have compared MC-CoT against recent strong baselines, covering five categories of methodologies: (i) heuristic and expert-guided choices, such as random choice and human evaluation [32]; (ii) standard multimodal visual question answering approaches, which include MCAN [58], Top-Down [1], BAN [25], DFAF [18], ViLT [26], PatchTRM [35], and VisualBERT [28]; (iii) Instruction-tuned large language models like GPT-3.5 [9] and its CoT-enhanced variants [32], in addition to ChatGPT, GPT-4 [39], and Chameleon [34]; (iv) Specifically finetuned large language models, notably LLaMA-Adapter [61], LaVIN [36], LLaMA-SciTune [20], and LLaVa [29]; (v) multimodal reasoning models based on chain-of-thought, Multimodal-CoT [63] and our MC-CoT both belong to this category. Regarding the A-OKVQA dataset, the baselines encompass key methods that have propelled the field of visual question answering forward, such as Pythia [7], ViLBERT [31], LXMERT [46], KRISP [37], GPV-2 [22], BLIP-2 [27], PICa [54], and IPVR [12]. Additionally, Multimodal-CoT [63] is incorporated as the chain-of-thought multimodal benchmark for this dataset. + +Implementation Details Following the experimental setup detailed in [63], we implemented the T5 encoder-decoder framework [43]. We utilize the UnifiedQA [24] models, which have 223M and 738M parameters, as our default Base and Large models, respectively. Additionally, we employ the FLAN-T5 [15] models with 248M and 783M parameters as our F-Base and F-Large models. Batch sizes were set to 16 for the base model and 8 for the large model. The learning rate was set to $5 \times 10^{-5}$ . Regarding sequence length, the initial output was constrained to 512 tokens, while the subsequent output was limited to 64 tokens. The experiments utilized four NVIDIA Tesla A100 GPUs, each equipped with 80GB of memory. + +# 5.1 Main Results + +We conducted a comparison of MC-CoT with state-of-the-art approaches on the ScienceQA dataset, as depicted in Table 1. We present the overall model size rather than the tunable model size, as it more accurately reflects the model's + +Table 1: The comparison on ScienceQA dataset. Question classes: NAT = natural science, SOC = social science, LAN = language science, TXT = text context, IMG = image context, NO = no context, G1-6 = grades 1-6, G7-12 = grades 7-12. + +
ModelSizeNATSOCLANTXTIMGNOG1-6G7-12AVG
Random Choice [32]-40.2846.1329.2547.7540.0833.6639.3540.6739.83
Human [32]-90.2384.9787.4889.6087.5088.1091.5982.4288.40
MCAN [58]95M56.0846.2358.0959.4351.1755.4051.6559.7254.54
Top-Down [1]70M59.5054.3361.8262.9054.8859.7957.2762.1659.02
BAN [25]112M60.8846.5766.6462.6152.6065.5156.8363.9459.37
DFAF [18]74M64.0348.8263.5565.8854.4964.1157.1267.1760.72
ViLT [26]113M60.4863.8960.2763.2061.3857.0060.7261.9061.14
Patch-TRM [35]90M65.1946.7965.5566.9655.2864.9558.0467.5061.42
VisualBERT [28]111M59.3369.1861.1862.7162.1758.5462.9659.9261.87
UnifiedQABase [24]223M68.1669.1874.9163.7861.3877.8472.9865.0070.12
UnifiedQABase w/ CoT [32]223M71.0076.0478.9166.4266.5381.8177.0668.8274.11
GPT-3.5 [32]173B74.6469.7476.0074.4467.2877.4276.8068.8973.97
GPT-3.5 w/ CoT [32]173B75.4470.8778.0974.6867.4379.9378.2369.6875.17
ChatGPT w/ CoT [39]-78.8270.9883.1877.3767.9286.1380.7274.0378.31
GPT-4 w/ CoT [39]-85.4872.4490.2782.6571.4992.8986.6679.0483.99
Chameleon + ChatGPT [34]-81.6270.6484.0079.7770.8086.6281.8676.5379.93
Chameleon + GPT-4 [34]-89.8374.1389.8288.2777.6492.1388.0383.7286.54
LLaMA-Adapter (T) [61]6B79.0073.7980.5578.3070.3583.1479.7775.6878.31
LLaMA-Adapter [61]6B84.3788.3084.3683.7280.3286.9085.8384.0585.19
LaVIN-7B [36]7B89.2594.9485.2488.5187.4688.0890.1688.0789.41
LLaMA-SciTuneBase [20]7B84.5094.1582.9188.3583.6488.7485.0585.6086.11
LaVIN-13B [36]13B90.3294.3887.7389.4487.6590.3191.1989.2690.50
LLaVa [29]13B90.3695.9588.0089.4988.0090.6690.9390.9090.92
LLaVa + GPT-4 [29]13B91.5696.7491.0990.6288.9993.5292.7392.1692.53
LLaMA-SciTuneLarge [20]13B89.3095.6187.0093.0886.6791.7584.3791.3090.03
Mutimodal-CoTBase [63]223M87.5277.1785.8287.8882.9086.8384.6585.3784.91
Mutimodal-CoTLarge [63]738M95.9182.0090.8295.2688.8092.8992.4490.3191.68
MC-CoTBase223M91.8784.5993.0092.2888.3092.7590.6490.6490.64
MC-CoTLarge738M95.4789.9991.8295.1192.6693.2494.2791.7693.37
MC-CoTf-Base248M93.5683.5890.7394.1389.2490.9490.9390.3890.73
MC-CoTf-Large783M97.4790.4493.1896.9793.7594.4995.3094.1394.88
+ +capacity. The results indicate that our MC-CoTBase model, with a model size of only 223 million parameters, achieved an average accuracy of $90.64\%$ , approaching the performance of Lavin-13B and LLaVa-13B models that were fine-tuned based on larger language models. Furthermore, our MC-CoTF-Large model, with 783 million parameters, achieved state-of-the-art results, surpassing the strongest fine-tuned large language model baseline LLaVa+GPT-4 by $2.35\%$ . + +Furthermore, in comparison to multimodal chain-of-thought baselines, our MC-CoTBase model showcased an average accuracy improvement of $5.70\%$ compared to Multimodal-CoTBase. Similarly, our MC-CoTLarge model exhibited an improvement of $1.69\%$ compared to the Multimodal-CoT Large model. These + +observations demonstrate that our approach has the potential to significantly enhance the model's reasoning capabilities. + +We report the experimental results on the A-OKVQA dataset within Table 2. Our MC-CoTBase model demonstrates exceptional improvements over the current state-of-the-art models, both in terms of direct-answer accuracy and multi-choice task performance. Specifically, MC-CoTBase achieves a direct-answer accuracy of $68.7\%$ , substantially outperforming the strongest competitor, BLIP-2, which possesses an extensive parameter count of over 11 billion. This represents an improvement of over $15.5\%$ , a margin that underscores the efficiency and effectiveness of our model's reasoning capabilities. In the multi-choice task, MC-CoTBase reaches an accuracy of $71.0\%$ , which not only exceeds BLIP-2's performance by $0.8\%$ but also does so with significantly fewer parameters. This comparison highlights the fact that our model's parameter utilization is exceptionally efficient, leading to enhanced performance even against models with far greater complexity. + +Table 2: The comparison on A-OKVQA dataset. We conduct the evaluation on both direct-answer and multi-choice tasks. + +
ModelVision ModelText ModelParametersDirect-answerMulti-choice
Pythia [7]ResNet [19]BERT [23]70M25.249.0
ViLBERT [31]Faster R-CNN [44]BERT [23]300M30.649.1
LXMERT [46]Transformer [49]Transformer [49]220M30.751.4
KRISP [37]Faster R-CNN [44]BERT [23]200M33.751.9
GPV-2 [22]VinVL [59]T5-Base [42]300M48.660.3
BLIP-2 [27]CLIP-VIT-LARGE [41]FlanT5XXL [15]11B53.270.2
PaLM-CoT [51]-PaLM [13]540B41.548.1
PICa [54]VinVL [59]GPT-3 [10]175B42.446.1
IPVR [12]Faster-RCNN [44]OPT [62]66B46.448.6
Multimodal-CoTBase [63]DETR [11]UnifiedQABase [24]223M-50.6
MC-CoTBaseDETR [11]UnifiedQABase [24]223M68.771.0
+ +These results are particularly noteworthy given the complexity and diversity of the A-OKVQA dataset, which is designed to challenge models with questions that require advanced understanding and reasoning across both textual and visual modalities. The performance leap made by $\mathrm{MC - CoT_{Base}}$ reinforces the potential of our improved training strategy. By introducing a voting mechanism among multiple rationales, our approach refines the quality of the generated rationales, which in turn significantly enhances the accuracy of the final answer. This innovative training paradigm proves effective in a multimodal context, where the interplay between textual and visual elements is critical for achieving high performance. Moreover, the robustness of $\mathrm{MC - CoT_{Base}}$ suggests that it is not only the size of the model that dictates performance but also the methodological advancements that come with thoughtful design and algorithmic improvements. This is evident from the fact that our base-sized model can outstrip the performance of models with parameter counts orders of magnitude larger. + +# 5.2 Quantative Analysis + +Ablation Study We present the experimental results on the ScienceQA dataset illustrated in Table 3. It can be seen that both using $\hat{Y}$ only and $\bar{Y}$ only can achieve comparable performance but the combined model outperforms them. This indicates that the two voting strategies are complementary to each other. Moreover, we can see that the model without voting R (rationales) obtains a significant drop in performance across all categories, with the average accuracy plummeting to $84.70\%$ . This highlights the importance of the rationale voting in improving the model's reasoning capabilities and accuracy. The absence of the answer voting mechanism results in a decrease in performance in the SOC, LAN, NO, G1-6, and G7-12 categories, with a slight improvement in NAT and TXT. The overall average is $90.19\%$ , indicating that answer voting contributes to the overall robustness, albeit not as critically as rationale voting. Inference voting means that voting is only conducted during the inference stage and not during the training stage. The significant performance drop of this strategy indicates that inference phase voting is not as effective as training phase voting. + +Table 3: Ablation study on ScienceQA dataset. + +
ModelNATSOCLANTXTIMGNOG1-6G7-12AVG
MC-CoTBase91.8784.5993.0092.2888.3092.7590.6490.6490.64
w/ mean only92.1083.8090.8292.4788.8490.6690.1689.7890.03
w/ weighted only91.6183.6991.0092.3888.4590.8089.7289.9189.79
w/o voting R87.3477.3985.1887.6883.2486.3484.5884.9084.70
w/o voting A92.3684.1490.6492.4288.3091.5090.3589.9190.19
inference voting87.4377.2884.6487.3982.8586.4184.9183.9884.58
+ +Effect of Rationale Generation We categorized the predictions from the Multimodal-CoTBase and MC-CoTBase models into four types: (i) Good Rationale and Answer; (ii) Bad Rationale and Answer; (iii) Good Rationale but Bad Answer; (iv) Bad Rationale but Good Answer. A 'Good Rationale' indicates the high-quality generated rationale. The results are depicted in Fig. 4. It can be seen that MC-CoT exhibits a higher frequency of Good Rationale and Answer, and a lower frequency of Bad Rationale and Answer. This suggests the self-consistency training approach improves the rationale's quality and the accuracy of the answers. + +We present the evaluation of the generated rationales and answers on the ScienceQA dataset in Table 4. It can be seen that our proposed MC-CoTBase model generated rationales that improved RougeL scores by approximately $1\%$ compared to the Multimodal-CoTBase model. However, this led to an approximately $6\%$ increase in the average accuracy of predicted answers. Similarly, the rationales generated by the MC-CoTF-Large model showed only about a $0.5\%$ improvement in RougeL scores compared to the MC-CoT base model. Yet, this resulted in an approximate $4\%$ increase in the average accuracy of predicted answers. This demonstrates that even slight improvements in the quality of rationales can impact the answer inference of the multimodal reasoning models. + +![](images/4b8205ac181f05a28e1cc6195644aa15562e4761b872b5a4c97aeda8037e4d24.jpg) +Fig. 4: The relationship between rationale and answer. + +Table 4: The RougeL score of the generated rationales and the average accuracy of predicted answers on the ScienceQA dataset. + +
MethodRougeLAvg Acc
Multimodal-CoTBase96.9784.91
MC-CoTBase97.9890.64
MC-CoTF-Large98.4794.88
+ +# 5.3 Qualitative Analysis + +We present two illustrative examples of predictions, as depicted in Figure 5, to demonstrate the efficacy of our proposed model. These examples have been carefully selected to showcase the model's capabilities in handling complex patterns and making accurate choices. For readers interested in exploring further, a more detailed collection of examples is provided in Appendix. + +
Planet Volume (cm3) Primary composition
Mercury 6.08 x 1015 rock Venus 9.28 x 1015 rock Earth 1.08 x 1012 rock Mars 1.63 x 1014 rock Jupiter 1.43 x 1015 gas Saturn 8.27 x 1014 gas Uranus 6.83 x 1013 ice Neptune 6.25 x 1013 ice
+ +Context: Use the data to answer the question below. +Question: Is the following statement about our solar system true or false? Of the four smallest planets, two are made mainly of gas. +Options: +(A) false +(B) true + +# Multimodal-CoT + +A planet's volume tells you the size of the planet .... The volumes of Jupiter, Venus, Earth, and Mars have the smallest exponents. So, these four planets are the smallest. The four planets are made mainly of gas. So, of the four smallest planets, two are made mainly of gas.. The answer is (B) true. + +# MC-CoT + +A planet's volume tells you the size of the planet .... The volumes of Mercury, Venus, Earth, and Mars have the smallest exponents. So, these four planets are the smallest. These four planets are made mainly of rock. So, of the four smallest planets, none are made mainly of gas.. The answer is (A) false. + +Fig. 5: Comparison on predicted examples. +![](images/751a83fdb3f6dcad53dce7a5eb5e0422ebcfda47bfb2720d6ad89a6148659da4.jpg) +Context: Use the graph to answer the question below. +Question: Which month is the hottest on average in Adelaide? + +Options: + +
(A) June, July, and August
+ +(B) April and May +(C) January and February + +# Multimodal-CoT + +Scientists record climate data from places around the world .... The average temperatures in April and February are around over $25^{\circ}\mathrm{C}$ . These months have the highest average temperatures of all of the months. So, they are hotter than the other months.. + +The answer is (A). June, July, and August. + +# MC-CoT + +Scientists record climate data from places around the world .... The average temperatures in January and February are between above $20^{\circ}\mathrm{F}$ . These months have the highest average temperatures of all of the months. So, they are hotter than the other months. + +The answer is (C) January and February. + +The first example highlights a scenario where the Multimodal-CoT model encountered difficulties in generating a coherent and accurate rationale, ultimately leading to an erroneous conclusion. This misstep underscores the challenges inherent in integrating multiple sources of information and extracting a logical reasoning path. Conversely, the MC-CoT model demonstrated its proficiency by navigating these complexities successfully, thereby arriving at the correct answer. This contrast exemplifies the significance of robust reasoning capabilities in achieving high accuracy in complex question-answering tasks. + +The second example delves into a different set of challenges faced by both models. Here, the Multimodal-CoT model stumbled over the interpretation of data presented in a chart, a mistake that significantly derailed its reasoning process and resulted in an incorrect answer. This error highlights the critical importance of accurate data interpretation skills in models designed for multimodal contexts. On the other hand, the MC-CoT model, while generally more adept in this instance, exhibited a lapse in understanding units of measurement—mistaking degrees Celsius $(^{\circ}\mathrm{C})$ for degrees Fahrenheit $(^{\circ}\mathrm{F})$ . Interestingly, this error did not compromise the integrity of its final answer, suggesting a certain resilience in the model's reasoning process despite minor inaccuracies. + +# 6 Conclusion + +In this study, we conducted an exhaustive analysis of the Multimodal-CoT framework and identified that the quality of rationales plays a crucial role in enhancing the performance of multimodal reasoning models. Recognizing the critical importance of this aspect, we developed an innovative solution aimed at improving the quality of rationales. Our solution employs a self-consistency training strategy that leverages the natural variability introduced by dropout techniques during training to foster a consensus mechanism through voting. This approach is supported by both theoretical underpinnings and empirical evidence, highlighting its effectiveness and wide applicability. We meticulously designed experiments to evaluate the impact of our self-consistency training strategy on the performance of multimodal reasoning models across diverse scenarios and datasets. The results from these experiments were revealing, clearly showing that our method significantly boosts the capabilities of multimodal models. A particularly significant finding was our approach's ability to enhance the performance of smaller models, enabling them to match or even exceed the capabilities of much larger models. We believe our strategy offers a promising avenue for advancing multimodal reasoning. + +A limitation of our approach is that we do not explicitly impose grammatical constraints, meaning the generated rationales are not guaranteed to be grammatically correct. It is worth noting that the Multimodal-CoT framework also does not guarantee the grammatical validity of the generated rationales. Despite this, empirical evidence has shown that our approach can yield superior results across various datasets, indicating that the rationales produced by our approach can significantly enhance the model's reasoning ability towards the correct answers. + +# 7 Acknowledgement + +This work was supported by the Science & Technology Innovation 2030 Major Program Project No. 2021ZD0150100, National Natural Science Foundation of China Project No. U21A20427, Project No. WU2022A009 from the Center of Synthetic Biology and Integrated Bioengineering of Westlake University, Project No. WU2023C019 from the Westlake University Industries of the Future Research. Finally, we thank the Westlake University HPC Center for providing part of the computational resources, and Project No. 23-407-3-29 from the Shenyang Science and Technology Program. + +# References + +1. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6077-6086 (2018) 4, 9, 10 +2. Anil, R., Dai, A.M., First, O., Johnson, M., Lepikhin, D., Passos, A., Shakeri, S., Taropa, E., Bailey, P., Chen, Z., et al.: Palm 2 technical report. arXiv preprint arXiv:2305.10403 (2023) 1 +3. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: Vqa: Visual question answering. In: Proceedings of the IEEE international conference on computer vision. pp. 2425-2433 (2015) 5 +4. Askell, A., Bai, Y., Chen, A., Drain, D., Ganguli, D., Henighan, T., Jones, A., Joseph, N., Mann, B., DasSarma, N., et al.: A general language assistant as a laboratory for alignment. arXiv preprint arXiv:2112.00861 (2021) 1 +5. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966 (2023) 2 +6. Besta, M., Blach, N., Kubicek, A., Gerstenberger, R., Gianinazzi, L., Gajda, J., Lehmann, T., Podstawski, M., Niewiadomski, H., Nczyk, P., et al.: Graph of thoughts: Solving elaborate problems with large language models. arXiv preprint arXiv:2308.09687 (2023) 1, 2, 4 +7. Biderman, S., Schoelkopf, H., Anthony, Q.G., Bradley, H., O'Brien, K., Hallahan, E., Khan, M.A., Purohit, S., Prashanth, U.S., Raff, E., et al.: Pythia: A suite for analyzing large language models across training and scaling. In: International Conference on Machine Learning. pp. 2397-2430. PMLR (2023) 9, 11 +8. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) 1 +9. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) 9 +10. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) 11 + +11. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020) 11 +12. Chen, Z., Zhou, Q., Shen, Y., Hong, Y., Zhang, H., Gan, C.: See, think, confirm: Interactive prompting between vision and language models for knowledge-based visual reasoning. arXiv preprint arXiv:2301.05226 (2023) 9, 11 +13. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311 (2022) 1, 11 +14. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, E., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022) 1 +15. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022) 9, 11 +16. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Qiu, Z., Lin, W., Yang, J., Zheng, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023) 2 +17. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023) 2 +18. Gao, P., Jiang, Z., You, H., Lu, P., Hoi, S.C., Wang, X., Li, H.: Dynamic fusion with intra-and inter-modality attention flow for visual question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6639-6648 (2019) 4, 9, 10 +19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 11 +20. Horawalavithana, S., Munikoti, S., Stewart, I., Kvinge, H.: Scitune: Aligning large language models with scientific multimodal instructions. arXiv preprint arXiv:2307.01139 (2023) 9, 10 +21. Huang, J., Chang, K.C.C.: Towards reasoning in large language models: A survey. arXiv preprint arXiv:2212.10403 (2022) 2 +22. Kamath, A., Clark, C., Gupta, T., Kolve, E., Hoiem, D., Kembhavi, A.: Webly supervised concept expansion for general purpose vision models. In: European Conference on Computer Vision. pp. 662-681. Springer (2022) 9, 11 +23. Kenton, J.D.M.W.C., Toutanova, L.K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of naacL-HLT. vol. 1, p. 2 (2019) 11 +24. Khashabi, D., Min, S., Khot, T., Sabharwal, A., Tafjord, O., Clark, P., Hajishirzi, H.: UNIFIEDQA: Crossing format boundaries with a single QA system. In: Cohn, T., He, Y., Liu, Y. (eds.) Findings of the Association for Computational Linguistics: EMNLP 2020. pp. 1896-1907. Association for Computational Linguistics, Online (Nov 2020). https://doi.org/10.18653/v1/2020-findings-emnlp.171, https://aclanthology.org/2020_findings-emnlp.171_9, 10, 11 +25. Kim, J.H., Jun, J., Zhang, B.T.: Bilinear attention networks. Advances in neural information processing systems 31 (2018) 4, 9, 10 +26. Kim, W., Son, B., Kim, I.: Vilt: Vision-and-language transformer without convolution or region supervision. In: International Conference on Machine Learning. pp. 5583-5594. PMLR (2021) 4, 9, 10 + +27. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) 4, 9, 11 +28. Li, L.H., Yatskar, M., Yin, D., Hsieh, C.J., Chang, K.W.: What does bert with vision look at? In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. pp. 5265-5275 (2020) 4, 9, 10 +29. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023) 2, 9, 10 +30. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281 (2023) 2 +31. Lu, J., Batra, D., Parikh, D., Lee, S.: Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems 32 (2019) 9, 11 +32. Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems 35, 2507-2521 (2022) 2, 3, 4, 9, 10 +33. Lu, P., Peng, B., Cheng, H., Galley, M., Chang, K.W., Wu, Y.N., Zhu, S.C., Gao, J.: Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842 (2023) 2 +34. Lu, P., Peng, B., Cheng, H., Galley, M., Chang, K.W., Wu, Y.N., Zhu, S.C., Gao, J.: Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842 (2023) 9, 10 +35. Lu, P., Qiu, L., Chen, J., Xia, T., Zhao, Y., Zhang, W., Yu, Z., Liang, X., Zhu, S.C.: Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. arXiv preprint arXiv:2110.13214 (2021) 4, 9, 10 +36. Luo, G., Zhou, Y., Ren, T., Chen, S., Sun, X., Ji, R.: Cheap and quick: Efficient vision-language instruction tuning for large language models. arXiv preprint arXiv:2305.15023 (2023) 2, 4, 9, 10 +37. Marino, K., Chen, X., Parikh, D., Gupta, A., Rohrbach, M.: Krisp: Integrating implicit and symbolic knowledge for open-domain knowledge-based vqa. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14111-14121 (2021) 9, 11 +38. Ning, X., Lin, Z., Zhou, Z., Yang, H., Wang, Y.: Skeleton-of-thought: Large language models can do parallel decoding. arXiv preprint arXiv:2307.15337 (2023) 4 +39. OpenAI: Gpt-4 technical report (2023) 1, 9, 10 +40. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022) 1 +41. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 11 +42. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020) 1, 11 + +43. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020) 9 +44. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems 28 (2015) 11 +45. Schwenk, D., Khandelwal, A., Clark, C., Marino, K., Mottaghi, R.: A-okvqa: A benchmark for visual question answering using world knowledge. In: European Conference on Computer Vision. pp. 146-162. Springer (2022) 3, 4, 9 +46. Tan, H., Bansal, M.: Lxmert: Learning cross-modality encoder representations from transformers. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (2019) 9, 11 +47. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) 1 +48. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) 1 +49. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) 11 +50. Wang, X., Wei, J., Schuurmans, D., Le, Q.V., Chi, E.H., Narang, S., Chowdhery, A., Zhou, D.: Self-consistency improves chain of thought reasoning in language models. In: The Eleventh International Conference on Learning Representations (2022) 1, 2, 4 +51. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022) 1, 3, 11 +52. Wu, C., Liu, J., Wang, X., Dong, X.: Chain of reasoning for visual question answering. Advances in Neural Information Processing Systems 31 (2018) 5 +53. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023) 2 +54. Yang, Z., Gan, Z., Wang, J., Hu, X., Lu, Y., Liu, Z., Wang, L.: An empirical study of gpt-3 for few-shot knowledge-based vqa. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 3081-3089 (2022) 9, 11 +55. Yang, Z., Li, L., Wang, J., Lin, K., Azarnasab, E., Ahmed, F., Liu, Z., Liu, C., Zeng, M., Wang, L.: Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381 (2023) 4 +56. Yao, S., Yu, D., Zhao, J., Shafran, I., Griffiths, T.L., Cao, Y., Narasimhan, K.: Tree of thoughts: Deliberate problem solving with large language models. arXiv preprint arXiv:2305.10601 (2023) 1, 2, 4 +57. Yin, S., Fu, C., Zhao, S., Li, K., Sun, X., Xu, T., Chen, E.: A survey on multimodal large language models. arXiv preprint arXiv:2306.13549 (2023) 2 +58. Yu, Z., Yu, J., Cui, Y., Tao, D., Tian, Q.: Deep modular co-attention networks for visual question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6281-6290 (2019) 4, 9, 10 +59. Zhang, P., Li, X., Hu, X., Yang, J., Zhang, L., Wang, L., Choi, Y., Gao, J.: Vinvl: Revisiting visual representations in vision-language models. In: Proceedings of the + +IEEE/CVF conference on computer vision and pattern recognition. pp. 5579-5588 (2021) 11 +60. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199 (2023) 2, 4 +61. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199 (2023) 9, 10 +62. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022) 1, 11 +63. Zhang, Z., Zhang, A., Li, M., Zhao, H., Karypis, G., Smola, A.: Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923 (2023) 2, 4, 5, 6, 9, 10, 11 \ No newline at end of file diff --git a/2024/Boosting the Power of Small Multimodal Reasoning Models to Match Larger Models with Self-Consistency Training/images.zip b/2024/Boosting the Power of Small Multimodal Reasoning Models to Match Larger Models with Self-Consistency Training/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3b7011f91eb1aec5b0490a07364e097b4ab3b310 --- /dev/null +++ b/2024/Boosting the Power of Small Multimodal Reasoning Models to Match Larger Models with Self-Consistency Training/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85297cb0dff0caf2701de68646a8daf98eea033ac60824503f04b53977c97c85 +size 528390 diff --git a/2024/Boosting the Power of Small Multimodal Reasoning Models to Match Larger Models with Self-Consistency Training/layout.json b/2024/Boosting the Power of Small Multimodal Reasoning Models to Match Larger Models with Self-Consistency Training/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..df30d8fb36236abe8302aef5ee70d05064de301e --- /dev/null +++ b/2024/Boosting the Power of Small Multimodal Reasoning Models to Match Larger Models with Self-Consistency Training/layout.json @@ -0,0 +1,10355 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 111, + 477, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 111, + 477, + 166 + ], + "spans": [ + { + "bbox": [ + 138, + 111, + 477, + 166 + ], + "type": "text", + "content": "Boosting the Power of Small Multimodal Reasoning Models to Match Larger Models with Self-Consistency Training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 137, + 186, + 478, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 186, + 478, + 212 + ], + "spans": [ + { + "bbox": [ + 137, + 186, + 478, + 212 + ], + "type": "text", + "content": "Cheng Tan" + }, + { + "bbox": [ + 137, + 186, + 478, + 212 + ], + "type": "inline_equation", + "content": "^{1,4,5\\star}" + }, + { + "bbox": [ + 137, + 186, + 478, + 212 + ], + "type": "text", + "content": ", Jingxuan Wei" + }, + { + "bbox": [ + 137, + 186, + 478, + 212 + ], + "type": "inline_equation", + "content": "^{2,3\\star\\boxtimes}" + }, + { + "bbox": [ + 137, + 186, + 478, + 212 + ], + "type": "text", + "content": ", Zhangyang Gao" + }, + { + "bbox": [ + 137, + 186, + 478, + 212 + ], + "type": "inline_equation", + "content": "^{1,4,5\\star}" + }, + { + "bbox": [ + 137, + 186, + 478, + 212 + ], + "type": "text", + "content": ", Linzhuang Sun" + }, + { + "bbox": [ + 137, + 186, + 478, + 212 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 137, + 186, + 478, + 212 + ], + "type": "text", + "content": ", Siyuan Li" + }, + { + "bbox": [ + 137, + 186, + 478, + 212 + ], + "type": "inline_equation", + "content": "^{1,4,5}" + }, + { + "bbox": [ + 137, + 186, + 478, + 212 + ], + "type": "text", + "content": ", Ruifeng Guo" + }, + { + "bbox": [ + 137, + 186, + 478, + 212 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 137, + 186, + 478, + 212 + ], + "type": "text", + "content": ", Bihui Yu" + }, + { + "bbox": [ + 137, + 186, + 478, + 212 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 137, + 186, + 478, + 212 + ], + "type": "text", + "content": ", and Stan Z. Li" + }, + { + "bbox": [ + 137, + 186, + 478, + 212 + ], + "type": "inline_equation", + "content": "^{4\\boxtimes}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 147, + 219, + 465, + 297 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 147, + 219, + 465, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 219, + 465, + 253 + ], + "spans": [ + { + "bbox": [ + 147, + 219, + 465, + 253 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 219, + 465, + 253 + ], + "type": "text", + "content": " Zhejiang University \n" + }, + { + "bbox": [ + 147, + 219, + 465, + 253 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 147, + 219, + 465, + 253 + ], + "type": "text", + "content": " Shenyang Institute of Computing Technology, Chinese Academy of Sciences \n" + }, + { + "bbox": [ + 147, + 219, + 465, + 253 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 147, + 219, + 465, + 253 + ], + "type": "text", + "content": " University of Chinese Academy of Sciences" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 151, + 253, + 462, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 253, + 462, + 264 + ], + "spans": [ + { + "bbox": [ + 151, + 253, + 462, + 264 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 151, + 253, + 462, + 264 + ], + "type": "text", + "content": " AI Lab, Research Center for Industries of the Future, Westlake University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 264, + 460, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 264, + 460, + 297 + ], + "spans": [ + { + "bbox": [ + 152, + 264, + 460, + 297 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 152, + 264, + 460, + 297 + ], + "type": "text", + "content": " Institute of Advanced Technology, Westlake Institute for Advanced Study tancheng@westlake.edu.cn, weijingxuan20@mails.ucas.edu.cn " + }, + { + "bbox": [ + 152, + 264, + 460, + 297 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 152, + 264, + 460, + 297 + ], + "type": "text", + "content": " Equal contribution, " + }, + { + "bbox": [ + 152, + 264, + 460, + 297 + ], + "type": "inline_equation", + "content": "\\boxtimes" + }, + { + "bbox": [ + 152, + 264, + 460, + 297 + ], + "type": "text", + "content": " Corresponding author" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 159, + 328, + 455, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 328, + 455, + 548 + ], + "spans": [ + { + "bbox": [ + 159, + 328, + 455, + 548 + ], + "type": "text", + "content": "Abstract. Multimodal reasoning is a challenging task that requires models to reason across multiple modalities to answer questions. Existing approaches have made progress by incorporating language and visual modalities into a two-stage reasoning framework, separating rationale generation from answer inference. However, these approaches often fall short due to the inadequate quality of the generated rationales. In this work, we delve into the importance of rationales in model reasoning. We observe that when rationales are completely accurate, the model's accuracy significantly improves, highlighting the need for high-quality rationale generation. Motivated by this, we propose MC-CoT, a self-consistency training strategy that generates multiple rationales and answers, subsequently selecting the most accurate through a voting process. This approach not only enhances the quality of generated rationales but also leads to more accurate and robust answers. Through extensive experiments, we demonstrate that our approach significantly improves model performance across various benchmarks. Remarkably, we show that even smaller base models, when equipped with our proposed approach, can achieve results comparable to those of larger models, illustrating the potential of our approach in harnessing the power of rationales for improved multimodal reasoning. The code is available at github.com/chengtan9907/mc-cot." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 559, + 418, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 559, + 418, + 571 + ], + "spans": [ + { + "bbox": [ + 160, + 559, + 418, + 571 + ], + "type": "text", + "content": "Keywords: Multimodal reasoning " + }, + { + "bbox": [ + 160, + 559, + 418, + 571 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 559, + 418, + 571 + ], + "type": "text", + "content": " visual question answering" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 591, + 230, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 591, + 230, + 604 + ], + "spans": [ + { + "bbox": [ + 132, + 591, + 230, + 604 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Recent advances in large language models [2, 4, 8, 13, 14, 39, 40, 42, 47, 48, 62] have led to the exploration of Chain-of-Thought (CoT) prompting [6, 50, 51, 56]. This approach, which directs the model to systematically unravel rationales before providing answers, rather than responding directly, has showcased the model's" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 335, + 198 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 335, + 198 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 335, + 198 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 335, + 198 + ], + "type": "image", + "image_path": "a89124767fb747bb36370ebf0917cd03d167d845edf0d2729bf1601e7253aac0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 207, + 482, + 231 + ], + "lines": [ + { + "bbox": [ + 131, + 207, + 482, + 231 + ], + "spans": [ + { + "bbox": [ + 131, + 207, + 482, + 231 + ], + "type": "text", + "content": "Fig. 1: An example of multimodal reasoning that answers the question by reasoning across both vision and language modalities." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 337, + 114, + 480, + 198 + ], + "blocks": [ + { + "bbox": [ + 337, + 114, + 480, + 198 + ], + "lines": [ + { + "bbox": [ + 337, + 114, + 480, + 198 + ], + "spans": [ + { + "bbox": [ + 337, + 114, + 480, + 198 + ], + "type": "image", + "image_path": "ba5d01f034521d1a765437c9b1c8a328c8b2f2b979ba3ba88f8c0f26e35988e6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 260, + 482, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 333 + ], + "type": "text", + "content": "impressive efficacy across a variety of natural language processing (NLP) tasks. Moreover, the advent of CoT prompting has catalyzed a plethora of research endeavors delving into the reasoning prowess of large language models. A diverse range of Chain-of-Thought strategies has been investigated, including the voting-facilitated CoT-SC [50], the transition from chain-like to tree-like thinking with Tree-of-Thoughts [56], and further expansion into graph-structured thinking [6]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 335, + 483, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 483, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 483, + 492 + ], + "type": "text", + "content": "While CoT reasoning has been thoroughly established in the realm of language models, its foray into the vast and intricate landscape of multimodal reasoning is still in its infancy. As shown in Fig. 1, Multimodal reasoning [5, 16, 17, 21, 29, 30, 32, 33, 36, 53, 57, 60], which inherently involves the seamless fusion of information from disparate modalities such as text and images, presents unique challenges. The process of extracting, correlating, and generating rationales across multiple modalities is decidedly more complex than the tasks encountered in a solely text-based modality. A recent seminal work, Multimodal-CoT [63], has pioneered the application of the Chain-of-Thought prompting strategy to multimodal reasoning tasks. This approach encompasses a two-stage framework that distinctively separates rationale generation from answer inference. By obliging the model to generate rationales prior to answering questions, Multimodal-CoT mirrors the language-only CoT prompting, thus enabling reasoning across multiple modalities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 494, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 482, + 579 + ], + "type": "text", + "content": "Despite Multimodal-CoT has made promising strides in the realm of multimodal reasoning, as evidenced in Fig. 2, its improvements over the no-rationale baseline are still limited. Moreover, compared to the ground-truth rationale, the predicted rationale falls short, often yielding results that lack relevance to the posed question. This discrepancy primarily stems from the quality of the generated rationales, highlighting the crucial role of rationale quality in the success of the chain-of-thought reasoning process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "In light of the above observations, it becomes evident that the potency of the Chain-of-Thought reasoning in a multimodal scenario is intrinsically tethered to the accuracy of the rationales generated. A high-quality rationale not only sheds light on the model's thought process but also paves the way for more precise answers. This leads to a critical question: how can we develop a strategy to enhance the quality of these rationales, and consequently, improve the overall performance of multimodal CoT?" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "type": "text", + "content": "Tan. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 187, + 110, + 427, + 269 + ], + "blocks": [ + { + "bbox": [ + 187, + 110, + 427, + 269 + ], + "lines": [ + { + "bbox": [ + 187, + 110, + 427, + 269 + ], + "spans": [ + { + "bbox": [ + 187, + 110, + 427, + 269 + ], + "type": "image", + "image_path": "35433dbd48d7b8159bd6cdb1801e57e751418a77099e9d857d7c8f6c54d21bfd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 279, + 480, + 301 + ], + "lines": [ + { + "bbox": [ + 131, + 279, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 131, + 279, + 480, + 301 + ], + "type": "text", + "content": "Fig. 2: The comparison of answer accuracy on ScienceQA using the Multimodal-CoT framework with no rationale, predicted rationales, and ground-truth rationales." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 314, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 469 + ], + "type": "text", + "content": "Our study is driven by the hypothesis that enhancing the quality of rationale generation can significantly improve the model's reasoning capabilities and overall performance. To this end, we introduce a simple yet effective strategy that capitalizes on the inherent variability of deep neural models during training, particularly stemming from mechanisms like dropout. Our approach involves having the model generate multiple rationales and then voting for the most consistent words across these rationales to yield a more refined and accurate rationale. The same voting mechanism is then applied to answer generation, further boosting the model's confidence in its predictions. It is important to note that the inference phase remains completely unaffected by this voting mechanism and continues to operate in the same manner as it does in the original Multimodal CoT framework. Through this approach, we aim to facilitate a more robust and accurate multimodal reasoning ability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 470, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 482, + 555 + ], + "type": "text", + "content": "Extensive experiments across ScienceQA [32] and A-OKVQA [45] benchmark datasets demonstrate the efficacy of our proposed approach. Notably, by improving the rationale quality, even smaller models equipped with our strategy manifest performance metrics that rival, and at times surpass, those of considerably larger models. This not only confirms the efficacy of our rationale refinement strategy but also opens up a promising avenue where smaller, more efficient models can be rendered competitive in the multimodal reasoning landscape." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 572, + 237, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 572, + 237, + 585 + ], + "spans": [ + { + "bbox": [ + 131, + 572, + 237, + 585 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 597, + 250, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 597, + 250, + 609 + ], + "spans": [ + { + "bbox": [ + 131, + 597, + 250, + 609 + ], + "type": "text", + "content": "2.1 Chain-of-Thought" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "The Chain-of-Thought (CoT) paradigm has emerged as a transformative approach, aiming to elucidate the reasoning processes of large language models and boost their ability in a variety of NLP tasks. The vanilla CoT [51] prompting strategy models to generate intermediate reasoning steps, leading to significant" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "text", + "content": "MC-CoT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "performance improvements across a spectrum of tasks. Its methodology laid the groundwork for subsequent advancements in the CoT paradigm. CoT-SC [50] introduced a self-consistency decoding strategy by sampling multiple reasoning paths and selecting the most consistent path to enhance the reliability of the generated rationales." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 175, + 483, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 483, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 483, + 283 + ], + "type": "text", + "content": "The versatility and adaptability of the CoT paradigm were further demonstrated with the advent of advanced reasoning structures. Tree-of-Thoughts (ToT) [56] transitioned from linear chains to more intricate tree-like structures, aiming to capture more complex reasoning patterns. In a parallel development, [6] ventured into graph-based reasoning, presenting the Graph-of-Thought (GoT) approach, which allowed for a more interconnected reasoning process. Skeleton-of-Thought (SoT) [38] took a different route by emphasizing the benefits of efficiency by underscoring the potential of parallel decoding, enabling models to explore multiple reasoning paths simultaneously." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 283, + 483, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 483, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 483, + 344 + ], + "type": "text", + "content": "While the aforementioned studies have made significant strides in the CoT paradigm, we seek to extend these advancements into the realm of multimodal reasoning. Multimodal-CoT [63] is a seminal work, but the unsatisfying performance of its generated rationales has been a major bottleneck. Our work aims to fully explore the potential of the CoT paradigm in multimodal reasoning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 360, + 361, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 360, + 361, + 372 + ], + "spans": [ + { + "bbox": [ + 131, + 360, + 361, + 372 + ], + "type": "text", + "content": "2.2 Multimodal Visual Question Answering" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 378, + 483, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 483, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 483, + 569 + ], + "type": "text", + "content": "Multimodal Visual Question Answering (VQA) has emerged as a key research area bridging the domains of vision and language. Pioneering works like MCAN [58], BAN [25], and Top-Down [1] have established a foundation by introducing diverse attention mechanisms. Building on this, DFAF [18] introduced innovative dynamic fusion strategies to facilitate both intra- and inter-modality attention flows. The introduction of Transformer architectures has significantly impacted the VQA field, as demonstrated by state-of-the-art models such as ViLT [26] and VisualBERT [28]. In the same vein, Patch-TRM [35] leveraged a pyramid cross-modal Transformer, paired with pre-trained input embeddings from the icon dataset, advancing the understanding of abstract diagrams. Recently, incorporating large language models into the multimodal framework has seen a surge, exemplified by models like LLaMA-Adapter [60] and BLIP-2 [27]. LaVIN [36] introduced a novel routing algorithm that helps the model autonomously switch between unimodal and multimodal instruction reasoning paths. These models employ frozen visual encoders paired with fine-tuned language models, thereby setting new standards in VQA tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 483, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 483, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 483, + 665 + ], + "type": "text", + "content": "ScienceQA [32] introduced a reasoning process into the VQA task, setting a benchmark for multimodal chain-of-thought reasoning. MM-REACT [55], with its innovative prompting design, enables language models to process multimodal information, facilitating the integration of ChatGPT with visual experts. A-OKVQA [45] requires commonsense reasoning about the depicted scene in an image. Multimodal-CoT [63] integrates language (text) and vision (image) modalities into a two-stage framework, distinctly separating rationale generation from answer inference." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "type": "text", + "content": "Tan. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "In this work, we typically focus on multimodal reasoning tasks, a subtype of multimodal visual question answering (VQA) that incorporates rationales. The significance of this task lies in its ability to more accurately simulate human cognitive processes. When humans answer questions, they often rely on observed visual information and existing textual knowledge, synthesizing the two to form a reasoned conclusion. These multimodal reasoning tasks comprehensively assess a model's capacity to understand and fuse visual and textual information, thereby propelling the development of multimodal learning. Moreover, these tasks enhance the interpretability of artificial intelligence systems. Because models must supply not only answers but also supporting rationales, their decision-making processes become more interpretable and comprehensible." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 263, + 202, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 202, + 275 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 202, + 275 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 285, + 228, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 228, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 228, + 296 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 302, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 482, + 361 + ], + "type": "text", + "content": "In this work, we focus on multimodal reasoning, a challenging task that involves reasoning across multiple modalities to answer questions. Specifically, we consider visual question answering [3, 52], where the model is tasked with answering questions based on the information provided in the question itself and the accompanying image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "text", + "content": "Formally, given a dataset " + }, + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{\\mathcal{X},\\mathcal{Y}\\}" + }, + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "inline_equation", + "content": "X\\in \\mathcal{X}" + }, + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "text", + "content": " represents a multimodal input consisting of text and images, and " + }, + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "inline_equation", + "content": "Y\\in \\mathcal{V}" + }, + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "text", + "content": " represents the corresponding output, the task of multimodal reasoning involves learning a mapping function parameterized by " + }, + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\Theta}:\\mathcal{X}\\to \\mathcal{V}" + }, + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "text", + "content": " that accurately predicts the output " + }, + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "text", + "content": " for a given input " + }, + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "text", + "content": ". The multimodal input " + }, + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "text", + "content": " can be represented as " + }, + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "inline_equation", + "content": "X = (T,I)" + }, + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "text", + "content": " is the text component, and " + }, + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "text", + "content": " is the image component of the input. The function " + }, + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 362, + 482, + 469 + ], + "type": "text", + "content": " should be able to effectively leverage the information from both the text and image modalities to make accurate predictions. Specifically, the basic visual question answering task is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 252, + 475, + 480, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 475, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 252, + 475, + 480, + 491 + ], + "type": "interline_equation", + "content": "Y = \\arg \\max _ {Y ^ {\\prime}} p \\left(Y ^ {\\prime} \\mid T, I\\right) \\tag {1}", + "image_path": "d6ed5f55884b27eabc740fa1a257ce1de2487fd23c4d9eb73b536cfb1f76883b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 497, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 557 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 497, + 482, + 557 + ], + "type": "inline_equation", + "content": "p(Y' \\mid T, I)" + }, + { + "bbox": [ + 130, + 497, + 482, + 557 + ], + "type": "text", + "content": " is the probability of answer " + }, + { + "bbox": [ + 130, + 497, + 482, + 557 + ], + "type": "inline_equation", + "content": "Y'" + }, + { + "bbox": [ + 130, + 497, + 482, + 557 + ], + "type": "text", + "content": " given the text " + }, + { + "bbox": [ + 130, + 497, + 482, + 557 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 497, + 482, + 557 + ], + "type": "text", + "content": " and image " + }, + { + "bbox": [ + 130, + 497, + 482, + 557 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 497, + 482, + 557 + ], + "type": "text", + "content": ". The multimodal reasoning task builds upon the basic visual question answering task, extending its requirements to necessitate the generation of a rationale " + }, + { + "bbox": [ + 130, + 497, + 482, + 557 + ], + "type": "inline_equation", + "content": "R'" + }, + { + "bbox": [ + 130, + 497, + 482, + 557 + ], + "type": "text", + "content": " that elucidates the reasoning process underpinning the answer " + }, + { + "bbox": [ + 130, + 497, + 482, + 557 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 497, + 482, + 557 + ], + "type": "text", + "content": ". This task can be mathematically represented as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 239, + 563, + 480, + 580 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 563, + 480, + 580 + ], + "spans": [ + { + "bbox": [ + 239, + 563, + 480, + 580 + ], + "type": "interline_equation", + "content": "Y, R = \\arg \\max _ {Y ^ {\\prime}, R ^ {\\prime}} p \\left(Y ^ {\\prime}, R ^ {\\prime} \\mid T, I\\right) \\tag {2}", + "image_path": "f409747fbfbc6ca069625a47b0f3e75e11ce3db593e020a577a6fb332579338b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 590, + 482, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 482, + 625 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 482, + 625 + ], + "type": "text", + "content": "Multimodal-CoT [63] decomposed the multimodal reasoning task into a two-stage framework, where the model first generates a rationale " + }, + { + "bbox": [ + 130, + 590, + 482, + 625 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 590, + 482, + 625 + ], + "type": "text", + "content": " and then utilizes this rationale to predict the answer " + }, + { + "bbox": [ + 130, + 590, + 482, + 625 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 590, + 482, + 625 + ], + "type": "text", + "content": ". The process is delineated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 247, + 632, + 480, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 632, + 480, + 654 + ], + "spans": [ + { + "bbox": [ + 247, + 632, + 480, + 654 + ], + "type": "interline_equation", + "content": "R = \\arg \\max _ {R ^ {\\prime}} p \\left(R ^ {\\prime} \\mid T, I\\right) \\tag {3}", + "image_path": "5205b5e9684eeaeed571241ec97e134a62037bce0821a13840096d020f2cb006.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 247, + 651, + 367, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 651, + 367, + 667 + ], + "spans": [ + { + "bbox": [ + 247, + 651, + 367, + 667 + ], + "type": "interline_equation", + "content": "Y = \\arg \\max _ {Y ^ {\\prime}} p (Y ^ {\\prime} | R, T, I)", + "image_path": "bd41bdf1f295b5a7000ba0a451bd63bbef684901b83cd5b0e6a335e7299be944.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "text", + "content": "MC-CoT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 113, + 185, + 217 + ], + "blocks": [ + { + "bbox": [ + 136, + 113, + 185, + 217 + ], + "lines": [ + { + "bbox": [ + 136, + 113, + 185, + 217 + ], + "spans": [ + { + "bbox": [ + 136, + 113, + 185, + 217 + ], + "type": "image", + "image_path": "8aa890f28b192b53fa4f8f9d2164f950d793374a17edcc096e4a0f6f3c80ec72.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 221, + 183, + 230 + ], + "lines": [ + { + "bbox": [ + 136, + 221, + 183, + 230 + ], + "spans": [ + { + "bbox": [ + 136, + 221, + 183, + 230 + ], + "type": "text", + "content": "(a) Vanilla Prompt" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 196, + 113, + 245, + 217 + ], + "blocks": [ + { + "bbox": [ + 196, + 113, + 245, + 217 + ], + "lines": [ + { + "bbox": [ + 196, + 113, + 245, + 217 + ], + "spans": [ + { + "bbox": [ + 196, + 113, + 245, + 217 + ], + "type": "image", + "image_path": "6217550fc74f3491f094618f9361a7e9db0a7b81e3b730ab9512799130caf1e2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 221, + 231, + 230 + ], + "lines": [ + { + "bbox": [ + 209, + 221, + 231, + 230 + ], + "spans": [ + { + "bbox": [ + 209, + 221, + 231, + 230 + ], + "type": "text", + "content": "(b) CoT" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 252, + 113, + 310, + 217 + ], + "blocks": [ + { + "bbox": [ + 252, + 113, + 310, + 217 + ], + "lines": [ + { + "bbox": [ + 252, + 113, + 310, + 217 + ], + "spans": [ + { + "bbox": [ + 252, + 113, + 310, + 217 + ], + "type": "image", + "image_path": "874a7e6b31956ba881aab82b289a377849e49776a704aefda476d44a02339608.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 265, + 221, + 295, + 230 + ], + "lines": [ + { + "bbox": [ + 265, + 221, + 295, + 230 + ], + "spans": [ + { + "bbox": [ + 265, + 221, + 295, + 230 + ], + "type": "text", + "content": "(c) CoT-SC" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 334, + 114, + 391, + 217 + ], + "blocks": [ + { + "bbox": [ + 334, + 114, + 391, + 217 + ], + "lines": [ + { + "bbox": [ + 334, + 114, + 391, + 217 + ], + "spans": [ + { + "bbox": [ + 334, + 114, + 391, + 217 + ], + "type": "image", + "image_path": "2f903064a751c15bfbf661e3b7f428015888d35a55209d180173ccf9b8946d61.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 221, + 391, + 230 + ], + "lines": [ + { + "bbox": [ + 341, + 221, + 391, + 230 + ], + "spans": [ + { + "bbox": [ + 341, + 221, + 391, + 230 + ], + "type": "text", + "content": "(d) Multimodal-CoT" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 242, + 483, + 353 + ], + "lines": [ + { + "bbox": [ + 130, + 242, + 483, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 483, + 353 + ], + "type": "text", + "content": "Fig. 3: A comparison schematic diagram of different Chain-of-Thought (CoT) prompt-based reasoning methods. (a) The basic input-output prompt, (b) Chain-of-Thought with intermediate chain-like reasoning, (c) Chain-of-Thought Self-Consistency (CoT-SC) that utilizes the consistency of multiple independent chains of thoughts for reasoning, (d) Multimodal-CoT, which infers the rationale using the input text and image, and then predicts the answer using the rationale as part of the input, and (e) MC-CoT that infers a high-quality rationale through word-level voting, and then obtains a high-quality answer using majority vote. It is worth noting that our approach leverages multiple chain consistency only during the training phase, in contrast to CoT-SC, which employs it during the inference stage." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 405, + 114, + 479, + 217 + ], + "blocks": [ + { + "bbox": [ + 405, + 114, + 479, + 217 + ], + "lines": [ + { + "bbox": [ + 405, + 114, + 479, + 217 + ], + "spans": [ + { + "bbox": [ + 405, + 114, + 479, + 217 + ], + "type": "image", + "image_path": "cdda3e35ff217268fa2b311aa2590b3e37e0ebab713a38208cf1fb47e2cde73a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 221, + 455, + 230 + ], + "lines": [ + { + "bbox": [ + 424, + 221, + 455, + 230 + ], + "spans": [ + { + "bbox": [ + 424, + 221, + 455, + 230 + ], + "type": "text", + "content": "(e) MC-CoT" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 376, + 482, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 401 + ], + "type": "text", + "content": "In this work, we follow the same two-stage framework as Multimodal-CoT [63] and explore the potential of such a chain-of-thought reasoning paradigm." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 420, + 369, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 420, + 369, + 433 + ], + "spans": [ + { + "bbox": [ + 131, + 420, + 369, + 433 + ], + "type": "text", + "content": "3.2 Multimodal Consistent Chain-of-Thought" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 441, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 502 + ], + "type": "text", + "content": "Though the existing multimodal CoT framework has been proven to be beneficial in multimodal reasoning, there remains a large room to fully explore its potential. In this work, we introduce a simple yet effective training strategy, Multimodal Consistent Chain-of-Thought (MC-CoT), to learn high-quality rationales and thus boost the reasoning performance of the multimodal CoT framework." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 516, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 482, + 552 + ], + "type": "text", + "content": "Rationale Generation Leveraging the inherent randomness introduced by the dropout operations, the model is able to generate multiple diverse rationales for a given input text-image pair " + }, + { + "bbox": [ + 130, + 516, + 482, + 552 + ], + "type": "inline_equation", + "content": "(T,I)" + }, + { + "bbox": [ + 130, + 516, + 482, + 552 + ], + "type": "text", + "content": " by sampling from the model " + }, + { + "bbox": [ + 130, + 516, + 482, + 552 + ], + "type": "inline_equation", + "content": "N_{r}" + }, + { + "bbox": [ + 130, + 516, + 482, + 552 + ], + "type": "text", + "content": " times:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 232, + 562, + 481, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 562, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 232, + 562, + 481, + 576 + ], + "type": "interline_equation", + "content": "R ^ {i} \\sim p (R \\mid T, I), i = 1, 2, \\dots , N _ {r} \\tag {4}", + "image_path": "b31ac56a29020d3abedac7bde991b24ec18a4460d3e698aba2b722be8b126645.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 585, + 482, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 585, + 482, + 620 + ], + "spans": [ + { + "bbox": [ + 130, + 585, + 482, + 620 + ], + "type": "text", + "content": "We perform a voting process on the generated rationales to select the most consistent words across the rationales. Specifically, the best rationale " + }, + { + "bbox": [ + 130, + 585, + 482, + 620 + ], + "type": "inline_equation", + "content": "R^{*}" + }, + { + "bbox": [ + 130, + 585, + 482, + 620 + ], + "type": "text", + "content": " is selected as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 260, + 620, + 481, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 620, + 481, + 637 + ], + "spans": [ + { + "bbox": [ + 260, + 620, + 481, + 637 + ], + "type": "interline_equation", + "content": "R _ {j} ^ {*} = \\operatorname {V o t e} \\left(\\left\\{R _ {j} ^ {i} \\right\\} _ {1} ^ {N _ {r}}\\right), \\tag {5}", + "image_path": "fc87211ef5dcb4e6da7703266ff10f3d358641e59405e732538623eb0ef3ea7a.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "text", + "content": "where, for each position " + }, + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "text", + "content": " in the rationale, we choose the majority of the words at position " + }, + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "text", + "content": " across the " + }, + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "inline_equation", + "content": "N_r" + }, + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "text", + "content": " generated rationales to form the best rationale " + }, + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "inline_equation", + "content": "R_j^*" + }, + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "type": "text", + "content": "Tan. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "type": "text", + "content": "Answer Inference In the answer inference stage, we also generate multiple answers based on the best rationale " + }, + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "R^{*}" + }, + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "type": "text", + "content": " and the input text-image pair " + }, + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "(T, I)" + }, + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "type": "text", + "content": " by sampling from the model " + }, + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "N_{a}" + }, + { + "bbox": [ + 131, + 116, + 479, + 152 + ], + "type": "text", + "content": " times:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 223, + 158, + 480, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 158, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 223, + 158, + 480, + 171 + ], + "type": "interline_equation", + "content": "Y ^ {i} \\sim p (Y \\mid R ^ {*}, T, I), i = 1, 2, \\dots , N _ {a}. \\tag {6}", + "image_path": "86dc98c03aa6443333c52304b1dc06a2cc9fb268316f7d0af3e3463db9d1775e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 175, + 358, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 175, + 358, + 188 + ], + "spans": [ + { + "bbox": [ + 131, + 175, + 358, + 188 + ], + "type": "text", + "content": "The final answer is selected by the majority voting:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 262, + 193, + 480, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 193, + 480, + 206 + ], + "spans": [ + { + "bbox": [ + 262, + 193, + 480, + 206 + ], + "type": "interline_equation", + "content": "Y ^ {*} = \\operatorname {V o t e} \\left(\\left\\{Y ^ {i} \\right\\} _ {1} ^ {N _ {a}}\\right). \\tag {7}", + "image_path": "97611d0b78bc3b08433914664712d7df86e2c8499049133e70a8d61952eea286.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 212, + 479, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 479, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 479, + 247 + ], + "type": "text", + "content": "where the majority of the answers across the " + }, + { + "bbox": [ + 130, + 212, + 479, + 247 + ], + "type": "inline_equation", + "content": "N_{a}" + }, + { + "bbox": [ + 130, + 212, + 479, + 247 + ], + "type": "text", + "content": " generated answers are regarded as the most convincing answers " + }, + { + "bbox": [ + 130, + 212, + 479, + 247 + ], + "type": "inline_equation", + "content": "Y^{*}" + }, + { + "bbox": [ + 130, + 212, + 479, + 247 + ], + "type": "text", + "content": ". The pseudocode of the pipeline is shown in Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 255, + 479, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 255, + 479, + 302 + ], + "spans": [ + { + "bbox": [ + 131, + 255, + 479, + 302 + ], + "type": "text", + "content": "Voting Strategy Consider the answer inference stage as an illustration. The vanilla Multimodal-CoT model computes a cross-entropy loss between a " + }, + { + "bbox": [ + 131, + 255, + 479, + 302 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 131, + 255, + 479, + 302 + ], + "type": "text", + "content": "-dimensional logit " + }, + { + "bbox": [ + 131, + 255, + 479, + 302 + ], + "type": "inline_equation", + "content": "L \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 131, + 255, + 479, + 302 + ], + "type": "text", + "content": " and the ground truth label " + }, + { + "bbox": [ + 131, + 255, + 479, + 302 + ], + "type": "inline_equation", + "content": "Y \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 131, + 255, + 479, + 302 + ], + "type": "text", + "content": ", which can be represented as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 252, + 304, + 480, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 304, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 252, + 304, + 480, + 316 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\operatorname {C r o s s E n t r o p y} (L, Y), \\tag {8}", + "image_path": "4f9d5df4f9de6b9544f7aaa9b7728356d5b8f177a03754e97a92f6824a5e8271.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 319, + 479, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 319, + 479, + 343 + ], + "spans": [ + { + "bbox": [ + 131, + 319, + 479, + 343 + ], + "type": "text", + "content": "Our averaging-based voting strategy initially calculates the mean logit " + }, + { + "bbox": [ + 131, + 319, + 479, + 343 + ], + "type": "inline_equation", + "content": "\\bar{L} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 131, + 319, + 479, + 343 + ], + "type": "text", + "content": " and weighted mean logit " + }, + { + "bbox": [ + 131, + 319, + 479, + 343 + ], + "type": "inline_equation", + "content": "\\hat{L} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 131, + 319, + 479, + 343 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 228, + 350, + 480, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 350, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 228, + 350, + 480, + 380 + ], + "type": "interline_equation", + "content": "\\bar {L} = \\frac {1}{N _ {a}} \\sum_ {i = 1} ^ {N _ {a}} L ^ {i}, \\hat {L} = \\frac {1}{N _ {a}} \\frac {\\sum_ {i = 1} ^ {N _ {a}} w L ^ {i}}{\\sum_ {j} w _ {j}}, \\tag {9}", + "image_path": "5e0685ae3619644fcccebd571ef25754a3d23b6fd29ca1cd09dba5497ce962c2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 386, + 479, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 386, + 479, + 456 + ], + "spans": [ + { + "bbox": [ + 131, + 386, + 479, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 386, + 479, + 456 + ], + "type": "inline_equation", + "content": "w = 1 / (1 + \\sigma) \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 131, + 386, + 479, + 456 + ], + "type": "text", + "content": " is a weight vector derived from the standard deviation of the logits, calculated across different predictions for each dimension. The weight for each logit's dimension is inversely proportional to its variability, assigning higher confidence to predictions with lower variability and lower confidence to those with higher variability. The final prediction is then a linear combination of " + }, + { + "bbox": [ + 131, + 386, + 479, + 456 + ], + "type": "inline_equation", + "content": "\\bar{L}" + }, + { + "bbox": [ + 131, + 386, + 479, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 386, + 479, + 456 + ], + "type": "inline_equation", + "content": "\\hat{L}" + }, + { + "bbox": [ + 131, + 386, + 479, + 456 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 263, + 459, + 479, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 459, + 479, + 471 + ], + "spans": [ + { + "bbox": [ + 263, + 459, + 479, + 471 + ], + "type": "interline_equation", + "content": "L ^ {*} = \\alpha \\bar {L} + (1 - \\alpha) \\hat {L} \\tag {10}", + "image_path": "ee71e46bbf219174ad5c0f038501ca286588ca3119b27ff8d73c447cdeb15270.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 475, + 479, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 475, + 479, + 498 + ], + "spans": [ + { + "bbox": [ + 131, + 475, + 479, + 498 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 131, + 475, + 479, + 498 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 475, + 479, + 498 + ], + "type": "text", + "content": " being a hyperparameter set empirically to 0.5. The cross-entropy loss is subsequently computed between " + }, + { + "bbox": [ + 131, + 475, + 479, + 498 + ], + "type": "inline_equation", + "content": "L^{*}" + }, + { + "bbox": [ + 131, + 475, + 479, + 498 + ], + "type": "text", + "content": " and the ground truth label " + }, + { + "bbox": [ + 131, + 475, + 479, + 498 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 131, + 475, + 479, + 498 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 250, + 505, + 479, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 505, + 479, + 516 + ], + "spans": [ + { + "bbox": [ + 250, + 505, + 479, + 516 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\operatorname {C r o s s E n t r o p y} \\left(L ^ {*}, Y\\right). \\tag {11}", + "image_path": "f1a9ccd79d314d150a79d9443b4e4f9072040f1499bc786bd13bc6e6ee07bef2.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 521, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 521, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 521, + 481, + 665 + ], + "type": "text", + "content": "We present a comparison of our approach with existing CoT-based reasoning methods in Fig. 3. Specifically, subfigures (a-c) illustrate the prompt learning methods commonly employed for text modality in large language models, while subfigures (d-e) depict the CoT reasoning frameworks in the multimodal field. It can be seen that our approach shares similarities with CoT-SC, such as the utilization of a voting mechanism. However, there are also significant differences between the two approaches: (i) the objective differs. CoT-SC aims to facilitate prompt learning in large language models during the inference stage, while our method seeks to enhance the robustness of model reasoning during the training phase; (ii) the voting process differs. CoT-SC only votes on the final answers inferred from each thought, whereas our method incorporates voting during the rationale generation stage." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "text", + "content": "MC-CoT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 272, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 272, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 272, + 129 + ], + "type": "text", + "content": "4 Theoretical Insights" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 146, + 354, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 146, + 354, + 158 + ], + "spans": [ + { + "bbox": [ + 132, + 146, + 354, + 158 + ], + "type": "text", + "content": "4.1 Aggregation Minimizes Expected Loss" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 171, + 479, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 171, + 479, + 195 + ], + "spans": [ + { + "bbox": [ + 130, + 171, + 479, + 195 + ], + "type": "text", + "content": "Theorem 1. (Jensen's inequality) Let " + }, + { + "bbox": [ + 130, + 171, + 479, + 195 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 171, + 479, + 195 + ], + "type": "text", + "content": " be a random variable and let " + }, + { + "bbox": [ + 130, + 171, + 479, + 195 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 171, + 479, + 195 + ], + "type": "text", + "content": " be a convex function. Then the following inequality holds: " + }, + { + "bbox": [ + 130, + 171, + 479, + 195 + ], + "type": "inline_equation", + "content": "\\phi(\\mathbb{E}[X]) \\leq \\mathbb{E}[\\phi(X)]" + }, + { + "bbox": [ + 130, + 171, + 479, + 195 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 197, + 482, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 197, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 197, + 482, + 281 + ], + "type": "text", + "content": "Our proposed method, MC-CoT, leverages the variability introduced by dropout techniques to generate a diverse set of rationales. By aggregating these rationales, we aim to produce explanations and corresponding answers of higher quality. In the context of our framework, the convex function " + }, + { + "bbox": [ + 130, + 197, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 197, + 482, + 281 + ], + "type": "text", + "content": " is represented by the cross-entropy loss function, and the random variable in question is the set of logits " + }, + { + "bbox": [ + 130, + 197, + 482, + 281 + ], + "type": "inline_equation", + "content": "L^i" + }, + { + "bbox": [ + 130, + 197, + 482, + 281 + ], + "type": "text", + "content": ". By applying Jensen's inequality to our model, we deduce that the aggregation process leads to a lower expected loss: " + }, + { + "bbox": [ + 130, + 197, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\mathbb{E}(L^{*}),Y)\\leq \\mathbb{E}(\\mathcal{L}(L^{*},Y))" + }, + { + "bbox": [ + 130, + 197, + 482, + 281 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 304, + 280, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 280, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 280, + 316 + ], + "type": "text", + "content": "4.2 Bias-Variance Trade-off" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 327, + 480, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 480, + 376 + ], + "type": "text", + "content": "We introduce a linear combination of two voting strategies to derive the final prediction, as delineated in Equation 10. We can analytically decompose the expected value of the squared difference between the predicted outcomes " + }, + { + "bbox": [ + 130, + 327, + 480, + 376 + ], + "type": "inline_equation", + "content": "Y^{*}" + }, + { + "bbox": [ + 130, + 327, + 480, + 376 + ], + "type": "text", + "content": " and the ground truth " + }, + { + "bbox": [ + 130, + 327, + 480, + 376 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 327, + 480, + 376 + ], + "type": "text", + "content": " into three distinct components:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 216, + 387, + 480, + 400 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 387, + 480, + 400 + ], + "spans": [ + { + "bbox": [ + 216, + 387, + 480, + 400 + ], + "type": "interline_equation", + "content": "\\mathbb {E} \\left[ \\left(Y ^ {*} - Y\\right) ^ {2} \\right] = \\operatorname {B i a s} ^ {2} \\left(Y ^ {*}\\right) + \\operatorname {V a r} \\left(Y ^ {*}\\right) + \\epsilon , \\tag {12}", + "image_path": "f0b352a16c32cbabc6804e796f656d89168a785e4fc3176c3b5507f58d2d77aa.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 411, + 479, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 479, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 479, + 483 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 411, + 479, + 483 + ], + "type": "inline_equation", + "content": "\\mathrm{Bias}^2 (Y^*) = (\\mathbb{E}[Y^*] - Y)^2" + }, + { + "bbox": [ + 130, + 411, + 479, + 483 + ], + "type": "text", + "content": " quantifies the systematic deviation of the expected prediction from the ground truth, indicative of the error inherently introduced by the predictive model itself. Meanwhile, " + }, + { + "bbox": [ + 130, + 411, + 479, + 483 + ], + "type": "inline_equation", + "content": "\\mathrm{Var}(Y^{*}) = \\mathbb{E}[(Y^{*} - \\mathbb{E}[Y^{*}])^{2}]" + }, + { + "bbox": [ + 130, + 411, + 479, + 483 + ], + "type": "text", + "content": " captures the variability of the predictions, reflecting the extent to which these predictions will fluctuate around their expected value. Lastly, " + }, + { + "bbox": [ + 130, + 411, + 479, + 483 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 411, + 479, + 483 + ], + "type": "text", + "content": " denotes the irreducible error, encapsulating the intrinsic noise within the data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 484, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 592 + ], + "type": "text", + "content": "The mean logits, denoted by " + }, + { + "bbox": [ + 130, + 484, + 482, + 592 + ], + "type": "inline_equation", + "content": "\\bar{L}" + }, + { + "bbox": [ + 130, + 484, + 482, + 592 + ], + "type": "text", + "content": ", may exhibit lower bias since they represent the composite prediction averaged across multiple models. Employing solely the mean logits promotes exploration by treating all predictions as equally informative. Nevertheless, its measure of central tendency does not account for the individual variability within the predictions. The weighted mean logits, denoted by " + }, + { + "bbox": [ + 130, + 484, + 482, + 592 + ], + "type": "inline_equation", + "content": "\\hat{L}" + }, + { + "bbox": [ + 130, + 484, + 482, + 592 + ], + "type": "text", + "content": ", aim to reduce variability by assigning greater significance to consistent (low-variance) predictions. This reflects a confidence-based strategy, allocating higher weight to predictions that are less variable. However, it could introduce bias if predictions are consistently erroneous." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "The final logits, which include both mean logits and weighted mean logits, are designed to achieve an optimal balance between bias and variance. While the mean logits promote exploration by treating all predictions equally, the weighted mean logits perform exploitation by assigning higher weights to consistent predictions. The combination allows balancing exploration with exploitation, thus improving the performance towards an optimal standard." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 210, + 101 + ], + "type": "text", + "content": "Tan. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 223, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 223, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 223, + 129 + ], + "type": "text", + "content": "5 Experiment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 271 + ], + "type": "text", + "content": "Datasets We evaluated MC-CoT on two prominent multimodal reasoning benchmarks. The first, ScienceQA [32], is a comprehensive multimodal science question dataset that contains annotated answers with detailed reasoning and explanations. It encompasses more than 21,000 multiple-choice questions, showcasing a vast range of domains that span three subjects, include 26 topics, cover 127 categories, and incorporate 379 distinct skills. The benchmark dataset is divided into training, validation, and test sets, consisting of 12,726, 4,241, and 4,241 examples, respectively. The second dataset is A-OKVQA [45], a knowledge-based visual question answering dataset that contains 25,000 questions requiring a broad base of commonsense and world knowledge for accurate responses, divided into 17,000 training, 1,000 validation, and 7,000 test examples." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 282, + 483, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 483, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 483, + 473 + ], + "type": "text", + "content": "Baselines In our assessment of the ScienceQA dataset, we have compared MC-CoT against recent strong baselines, covering five categories of methodologies: (i) heuristic and expert-guided choices, such as random choice and human evaluation [32]; (ii) standard multimodal visual question answering approaches, which include MCAN [58], Top-Down [1], BAN [25], DFAF [18], ViLT [26], PatchTRM [35], and VisualBERT [28]; (iii) Instruction-tuned large language models like GPT-3.5 [9] and its CoT-enhanced variants [32], in addition to ChatGPT, GPT-4 [39], and Chameleon [34]; (iv) Specifically finetuned large language models, notably LLaMA-Adapter [61], LaVIN [36], LLaMA-SciTune [20], and LLaVa [29]; (v) multimodal reasoning models based on chain-of-thought, Multimodal-CoT [63] and our MC-CoT both belong to this category. Regarding the A-OKVQA dataset, the baselines encompass key methods that have propelled the field of visual question answering forward, such as Pythia [7], ViLBERT [31], LXMERT [46], KRISP [37], GPV-2 [22], BLIP-2 [27], PICa [54], and IPVR [12]. Additionally, Multimodal-CoT [63] is incorporated as the chain-of-thought multimodal benchmark for this dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "text", + "content": "Implementation Details Following the experimental setup detailed in [63], we implemented the T5 encoder-decoder framework [43]. We utilize the UnifiedQA [24] models, which have 223M and 738M parameters, as our default Base and Large models, respectively. Additionally, we employ the FLAN-T5 [15] models with 248M and 783M parameters as our F-Base and F-Large models. Batch sizes were set to 16 for the base model and 8 for the large model. The learning rate was set to " + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "text", + "content": ". Regarding sequence length, the initial output was constrained to 512 tokens, while the subsequent output was limited to 64 tokens. The experiments utilized four NVIDIA Tesla A100 GPUs, each equipped with 80GB of memory." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 609, + 228, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 228, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 228, + 621 + ], + "type": "text", + "content": "5.1 Main Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "We conducted a comparison of MC-CoT with state-of-the-art approaches on the ScienceQA dataset, as depicted in Table 1. We present the overall model size rather than the tunable model size, as it more accurately reflects the model's" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "text", + "content": "MC-CoT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 157, + 480, + 512 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: The comparison on ScienceQA dataset. Question classes: NAT = natural science, SOC = social science, LAN = language science, TXT = text context, IMG = image context, NO = no context, G1-6 = grades 1-6, G7-12 = grades 7-12." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 157, + 480, + 512 + ], + "lines": [ + { + "bbox": [ + 135, + 157, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 135, + 157, + 480, + 512 + ], + "type": "table", + "html": "
ModelSizeNATSOCLANTXTIMGNOG1-6G7-12AVG
Random Choice [32]-40.2846.1329.2547.7540.0833.6639.3540.6739.83
Human [32]-90.2384.9787.4889.6087.5088.1091.5982.4288.40
MCAN [58]95M56.0846.2358.0959.4351.1755.4051.6559.7254.54
Top-Down [1]70M59.5054.3361.8262.9054.8859.7957.2762.1659.02
BAN [25]112M60.8846.5766.6462.6152.6065.5156.8363.9459.37
DFAF [18]74M64.0348.8263.5565.8854.4964.1157.1267.1760.72
ViLT [26]113M60.4863.8960.2763.2061.3857.0060.7261.9061.14
Patch-TRM [35]90M65.1946.7965.5566.9655.2864.9558.0467.5061.42
VisualBERT [28]111M59.3369.1861.1862.7162.1758.5462.9659.9261.87
UnifiedQABase [24]223M68.1669.1874.9163.7861.3877.8472.9865.0070.12
UnifiedQABase w/ CoT [32]223M71.0076.0478.9166.4266.5381.8177.0668.8274.11
GPT-3.5 [32]173B74.6469.7476.0074.4467.2877.4276.8068.8973.97
GPT-3.5 w/ CoT [32]173B75.4470.8778.0974.6867.4379.9378.2369.6875.17
ChatGPT w/ CoT [39]-78.8270.9883.1877.3767.9286.1380.7274.0378.31
GPT-4 w/ CoT [39]-85.4872.4490.2782.6571.4992.8986.6679.0483.99
Chameleon + ChatGPT [34]-81.6270.6484.0079.7770.8086.6281.8676.5379.93
Chameleon + GPT-4 [34]-89.8374.1389.8288.2777.6492.1388.0383.7286.54
LLaMA-Adapter (T) [61]6B79.0073.7980.5578.3070.3583.1479.7775.6878.31
LLaMA-Adapter [61]6B84.3788.3084.3683.7280.3286.9085.8384.0585.19
LaVIN-7B [36]7B89.2594.9485.2488.5187.4688.0890.1688.0789.41
LLaMA-SciTuneBase [20]7B84.5094.1582.9188.3583.6488.7485.0585.6086.11
LaVIN-13B [36]13B90.3294.3887.7389.4487.6590.3191.1989.2690.50
LLaVa [29]13B90.3695.9588.0089.4988.0090.6690.9390.9090.92
LLaVa + GPT-4 [29]13B91.5696.7491.0990.6288.9993.5292.7392.1692.53
LLaMA-SciTuneLarge [20]13B89.3095.6187.0093.0886.6791.7584.3791.3090.03
Mutimodal-CoTBase [63]223M87.5277.1785.8287.8882.9086.8384.6585.3784.91
Mutimodal-CoTLarge [63]738M95.9182.0090.8295.2688.8092.8992.4490.3191.68
MC-CoTBase223M91.8784.5993.0092.2888.3092.7590.6490.6490.64
MC-CoTLarge738M95.4789.9991.8295.1192.6693.2494.2791.7693.37
MC-CoTf-Base248M93.5683.5890.7394.1389.2490.9490.9390.3890.73
MC-CoTf-Large783M97.4790.4493.1896.9793.7594.4995.3094.1394.88
", + "image_path": "e11b0a29994f6b00d637c253843cc2dd63ba3284bbd3e8d48f1801968f147eea.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 541, + 482, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 541, + 482, + 614 + ], + "spans": [ + { + "bbox": [ + 130, + 541, + 482, + 614 + ], + "type": "text", + "content": "capacity. The results indicate that our MC-CoTBase model, with a model size of only 223 million parameters, achieved an average accuracy of " + }, + { + "bbox": [ + 130, + 541, + 482, + 614 + ], + "type": "inline_equation", + "content": "90.64\\%" + }, + { + "bbox": [ + 130, + 541, + 482, + 614 + ], + "type": "text", + "content": ", approaching the performance of Lavin-13B and LLaVa-13B models that were fine-tuned based on larger language models. Furthermore, our MC-CoTF-Large model, with 783 million parameters, achieved state-of-the-art results, surpassing the strongest fine-tuned large language model baseline LLaVa+GPT-4 by " + }, + { + "bbox": [ + 130, + 541, + 482, + 614 + ], + "type": "inline_equation", + "content": "2.35\\%" + }, + { + "bbox": [ + 130, + 541, + 482, + 614 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Furthermore, in comparison to multimodal chain-of-thought baselines, our MC-CoTBase model showcased an average accuracy improvement of " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "5.70\\%" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " compared to Multimodal-CoTBase. Similarly, our MC-CoTLarge model exhibited an improvement of " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "1.69\\%" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " compared to the Multimodal-CoT Large model. These" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "type": "text", + "content": "Tan. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "observations demonstrate that our approach has the potential to significantly enhance the model's reasoning capabilities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 482, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 482, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 482, + 288 + ], + "type": "text", + "content": "We report the experimental results on the A-OKVQA dataset within Table 2. Our MC-CoTBase model demonstrates exceptional improvements over the current state-of-the-art models, both in terms of direct-answer accuracy and multi-choice task performance. Specifically, MC-CoTBase achieves a direct-answer accuracy of " + }, + { + "bbox": [ + 130, + 143, + 482, + 288 + ], + "type": "inline_equation", + "content": "68.7\\%" + }, + { + "bbox": [ + 130, + 143, + 482, + 288 + ], + "type": "text", + "content": ", substantially outperforming the strongest competitor, BLIP-2, which possesses an extensive parameter count of over 11 billion. This represents an improvement of over " + }, + { + "bbox": [ + 130, + 143, + 482, + 288 + ], + "type": "inline_equation", + "content": "15.5\\%" + }, + { + "bbox": [ + 130, + 143, + 482, + 288 + ], + "type": "text", + "content": ", a margin that underscores the efficiency and effectiveness of our model's reasoning capabilities. In the multi-choice task, MC-CoTBase reaches an accuracy of " + }, + { + "bbox": [ + 130, + 143, + 482, + 288 + ], + "type": "inline_equation", + "content": "71.0\\%" + }, + { + "bbox": [ + 130, + 143, + 482, + 288 + ], + "type": "text", + "content": ", which not only exceeds BLIP-2's performance by " + }, + { + "bbox": [ + 130, + 143, + 482, + 288 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 130, + 143, + 482, + 288 + ], + "type": "text", + "content": " but also does so with significantly fewer parameters. This comparison highlights the fact that our model's parameter utilization is exceptionally efficient, leading to enhanced performance even against models with far greater complexity." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 345, + 482, + 471 + ], + "blocks": [ + { + "bbox": [ + 130, + 312, + 479, + 335 + ], + "lines": [ + { + "bbox": [ + 130, + 312, + 479, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 479, + 335 + ], + "type": "text", + "content": "Table 2: The comparison on A-OKVQA dataset. We conduct the evaluation on both direct-answer and multi-choice tasks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 345, + 482, + 471 + ], + "lines": [ + { + "bbox": [ + 135, + 345, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 135, + 345, + 482, + 471 + ], + "type": "table", + "html": "
ModelVision ModelText ModelParametersDirect-answerMulti-choice
Pythia [7]ResNet [19]BERT [23]70M25.249.0
ViLBERT [31]Faster R-CNN [44]BERT [23]300M30.649.1
LXMERT [46]Transformer [49]Transformer [49]220M30.751.4
KRISP [37]Faster R-CNN [44]BERT [23]200M33.751.9
GPV-2 [22]VinVL [59]T5-Base [42]300M48.660.3
BLIP-2 [27]CLIP-VIT-LARGE [41]FlanT5XXL [15]11B53.270.2
PaLM-CoT [51]-PaLM [13]540B41.548.1
PICa [54]VinVL [59]GPT-3 [10]175B42.446.1
IPVR [12]Faster-RCNN [44]OPT [62]66B46.448.6
Multimodal-CoTBase [63]DETR [11]UnifiedQABase [24]223M-50.6
MC-CoTBaseDETR [11]UnifiedQABase [24]223M68.771.0
", + "image_path": "4940c69a24bcaeaf95c52ddfe9054209073352de05d1ca45a0b55698a8c6d7df.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "These results are particularly noteworthy given the complexity and diversity of the A-OKVQA dataset, which is designed to challenge models with questions that require advanced understanding and reasoning across both textual and visual modalities. The performance leap made by " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{MC - CoT_{Base}}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " reinforces the potential of our improved training strategy. By introducing a voting mechanism among multiple rationales, our approach refines the quality of the generated rationales, which in turn significantly enhances the accuracy of the final answer. This innovative training paradigm proves effective in a multimodal context, where the interplay between textual and visual elements is critical for achieving high performance. Moreover, the robustness of " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{MC - CoT_{Base}}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " suggests that it is not only the size of the model that dictates performance but also the methodological advancements that come with thoughtful design and algorithmic improvements. This is evident from the fact that our base-sized model can outstrip the performance of models with parameter counts orders of magnitude larger." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "text", + "content": "MC-CoT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 263, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 263, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 263, + 128 + ], + "type": "text", + "content": "5.2 Quantative Analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 132, + 482, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 132, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 132, + 482, + 312 + ], + "type": "text", + "content": "Ablation Study We present the experimental results on the ScienceQA dataset illustrated in Table 3. It can be seen that both using " + }, + { + "bbox": [ + 130, + 132, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 130, + 132, + 482, + 312 + ], + "type": "text", + "content": " only and " + }, + { + "bbox": [ + 130, + 132, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\bar{Y}" + }, + { + "bbox": [ + 130, + 132, + 482, + 312 + ], + "type": "text", + "content": " only can achieve comparable performance but the combined model outperforms them. This indicates that the two voting strategies are complementary to each other. Moreover, we can see that the model without voting R (rationales) obtains a significant drop in performance across all categories, with the average accuracy plummeting to " + }, + { + "bbox": [ + 130, + 132, + 482, + 312 + ], + "type": "inline_equation", + "content": "84.70\\%" + }, + { + "bbox": [ + 130, + 132, + 482, + 312 + ], + "type": "text", + "content": ". This highlights the importance of the rationale voting in improving the model's reasoning capabilities and accuracy. The absence of the answer voting mechanism results in a decrease in performance in the SOC, LAN, NO, G1-6, and G7-12 categories, with a slight improvement in NAT and TXT. The overall average is " + }, + { + "bbox": [ + 130, + 132, + 482, + 312 + ], + "type": "inline_equation", + "content": "90.19\\%" + }, + { + "bbox": [ + 130, + 132, + 482, + 312 + ], + "type": "text", + "content": ", indicating that answer voting contributes to the overall robustness, albeit not as critically as rationale voting. Inference voting means that voting is only conducted during the inference stage and not during the training stage. The significant performance drop of this strategy indicates that inference phase voting is not as effective as training phase voting." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 340, + 479, + 432 + ], + "blocks": [ + { + "bbox": [ + 207, + 319, + 405, + 331 + ], + "lines": [ + { + "bbox": [ + 207, + 319, + 405, + 331 + ], + "spans": [ + { + "bbox": [ + 207, + 319, + 405, + 331 + ], + "type": "text", + "content": "Table 3: Ablation study on ScienceQA dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 340, + 479, + 432 + ], + "lines": [ + { + "bbox": [ + 135, + 340, + 479, + 432 + ], + "spans": [ + { + "bbox": [ + 135, + 340, + 479, + 432 + ], + "type": "table", + "html": "
ModelNATSOCLANTXTIMGNOG1-6G7-12AVG
MC-CoTBase91.8784.5993.0092.2888.3092.7590.6490.6490.64
w/ mean only92.1083.8090.8292.4788.8490.6690.1689.7890.03
w/ weighted only91.6183.6991.0092.3888.4590.8089.7289.9189.79
w/o voting R87.3477.3985.1887.6883.2486.3484.5884.9084.70
w/o voting A92.3684.1490.6492.4288.3091.5090.3589.9190.19
inference voting87.4377.2884.6487.3982.8586.4184.9183.9884.58
", + "image_path": "a6b15c60f0050700b2a4d6a916e2b667d8359ac3d7ecc54f24a0144ca24986aa.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 545 + ], + "type": "text", + "content": "Effect of Rationale Generation We categorized the predictions from the Multimodal-CoTBase and MC-CoTBase models into four types: (i) Good Rationale and Answer; (ii) Bad Rationale and Answer; (iii) Good Rationale but Bad Answer; (iv) Bad Rationale but Good Answer. A 'Good Rationale' indicates the high-quality generated rationale. The results are depicted in Fig. 4. It can be seen that MC-CoT exhibits a higher frequency of Good Rationale and Answer, and a lower frequency of Bad Rationale and Answer. This suggests the self-consistency training approach improves the rationale's quality and the accuracy of the answers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": "We present the evaluation of the generated rationales and answers on the ScienceQA dataset in Table 4. It can be seen that our proposed MC-CoTBase model generated rationales that improved RougeL scores by approximately " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": " compared to the Multimodal-CoTBase model. However, this led to an approximately " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": " increase in the average accuracy of predicted answers. Similarly, the rationales generated by the MC-CoTF-Large model showed only about a " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": " improvement in RougeL scores compared to the MC-CoT base model. Yet, this resulted in an approximate " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": " increase in the average accuracy of predicted answers. This demonstrates that even slight improvements in the quality of rationales can impact the answer inference of the multimodal reasoning models." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "type": "text", + "content": "Tan. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 116, + 299, + 238 + ], + "blocks": [ + { + "bbox": [ + 135, + 116, + 299, + 238 + ], + "lines": [ + { + "bbox": [ + 135, + 116, + 299, + 238 + ], + "spans": [ + { + "bbox": [ + 135, + 116, + 299, + 238 + ], + "type": "image", + "image_path": "4b8205ac181f05a28e1cc6195644aa15562e4761b872b5a4c97aeda8037e4d24.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 249, + 301, + 270 + ], + "lines": [ + { + "bbox": [ + 132, + 249, + 301, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 301, + 270 + ], + "type": "text", + "content": "Fig. 4: The relationship between rationale and answer." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 313, + 192, + 476, + 249 + ], + "blocks": [ + { + "bbox": [ + 311, + 148, + 482, + 190 + ], + "lines": [ + { + "bbox": [ + 311, + 148, + 482, + 190 + ], + "spans": [ + { + "bbox": [ + 311, + 148, + 482, + 190 + ], + "type": "text", + "content": "Table 4: The RougeL score of the generated rationales and the average accuracy of predicted answers on the ScienceQA dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 192, + 476, + 249 + ], + "lines": [ + { + "bbox": [ + 313, + 192, + 476, + 249 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 476, + 249 + ], + "type": "table", + "html": "
MethodRougeLAvg Acc
Multimodal-CoTBase96.9784.91
MC-CoTBase97.9890.64
MC-CoTF-Large98.4794.88
", + "image_path": "88b86336f32b88749e0ab7713d1b8eaaa13bb617dc47e5bb60981716ad81d4dd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 296, + 263, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 296, + 263, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 296, + 263, + 308 + ], + "type": "text", + "content": "5.3 Qualitative Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 317, + 482, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 317, + 482, + 377 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 482, + 377 + ], + "type": "text", + "content": "We present two illustrative examples of predictions, as depicted in Figure 5, to demonstrate the efficacy of our proposed model. These examples have been carefully selected to showcase the model's capabilities in handling complex patterns and making accurate choices. For readers interested in exploring further, a more detailed collection of examples is provided in Appendix." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 171, + 400, + 249, + 460 + ], + "blocks": [ + { + "bbox": [ + 171, + 400, + 249, + 460 + ], + "lines": [ + { + "bbox": [ + 171, + 400, + 249, + 460 + ], + "spans": [ + { + "bbox": [ + 171, + 400, + 249, + 460 + ], + "type": "table", + "html": "
Planet Volume (cm3) Primary composition
Mercury 6.08 x 1015 rock Venus 9.28 x 1015 rock Earth 1.08 x 1012 rock Mars 1.63 x 1014 rock Jupiter 1.43 x 1015 gas Saturn 8.27 x 1014 gas Uranus 6.83 x 1013 ice Neptune 6.25 x 1013 ice
", + "image_path": "b5926755da59c6b8a81ca1d1475e3866da2c900f1fe3e77b06f99dd2dc80d2c4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 139, + 463, + 256, + 470 + ], + "lines": [ + { + "bbox": [ + 139, + 463, + 256, + 470 + ], + "spans": [ + { + "bbox": [ + 139, + 463, + 256, + 470 + ], + "type": "text", + "content": "Context: Use the data to answer the question below." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 139, + 473, + 280, + 500 + ], + "lines": [ + { + "bbox": [ + 139, + 473, + 280, + 500 + ], + "spans": [ + { + "bbox": [ + 139, + 473, + 280, + 500 + ], + "type": "text", + "content": "Question: Is the following statement about our solar system true or false? Of the four smallest planets, two are made mainly of gas." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 139, + 501, + 162, + 510 + ], + "lines": [ + { + "bbox": [ + 139, + 501, + 162, + 510 + ], + "spans": [ + { + "bbox": [ + 139, + 501, + 162, + 510 + ], + "type": "text", + "content": "Options:" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 167, + 501, + 187, + 510 + ], + "lines": [ + { + "bbox": [ + 167, + 501, + 187, + 510 + ], + "spans": [ + { + "bbox": [ + 167, + 501, + 187, + 510 + ], + "type": "text", + "content": "(A) false" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 201, + 501, + 220, + 510 + ], + "lines": [ + { + "bbox": [ + 201, + 501, + 220, + 510 + ], + "spans": [ + { + "bbox": [ + 201, + 501, + 220, + 510 + ], + "type": "text", + "content": "(B) true" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "bbox": [ + 291, + 403, + 339, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 403, + 339, + 411 + ], + "spans": [ + { + "bbox": [ + 291, + 403, + 339, + 411 + ], + "type": "text", + "content": "Multimodal-CoT" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 291, + 424, + 370, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 424, + 370, + 498 + ], + "spans": [ + { + "bbox": [ + 291, + 424, + 370, + 498 + ], + "type": "text", + "content": "A planet's volume tells you the size of the planet .... The volumes of Jupiter, Venus, Earth, and Mars have the smallest exponents. So, these four planets are the smallest. The four planets are made mainly of gas. So, of the four smallest planets, two are made mainly of gas.. The answer is (B) true." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 385, + 403, + 411, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 403, + 411, + 411 + ], + "spans": [ + { + "bbox": [ + 385, + 403, + 411, + 411 + ], + "type": "text", + "content": "MC-CoT" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 388, + 424, + 468, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 424, + 468, + 498 + ], + "spans": [ + { + "bbox": [ + 388, + 424, + 468, + 498 + ], + "type": "text", + "content": "A planet's volume tells you the size of the planet .... The volumes of Mercury, Venus, Earth, and Mars have the smallest exponents. So, these four planets are the smallest. These four planets are made mainly of rock. So, of the four smallest planets, none are made mainly of gas.. The answer is (A) false." + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 165, + 528, + 251, + 588 + ], + "blocks": [ + { + "bbox": [ + 165, + 528, + 251, + 588 + ], + "lines": [ + { + "bbox": [ + 165, + 528, + 251, + 588 + ], + "spans": [ + { + "bbox": [ + 165, + 528, + 251, + 588 + ], + "type": "image", + "image_path": "751a83fdb3f6dcad53dce7a5eb5e0422ebcfda47bfb2720d6ad89a6148659da4.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 139, + 590, + 260, + 597 + ], + "lines": [ + { + "bbox": [ + 139, + 590, + 260, + 597 + ], + "spans": [ + { + "bbox": [ + 139, + 590, + 260, + 597 + ], + "type": "text", + "content": "Context: Use the graph to answer the question below." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 139, + 598, + 279, + 606 + ], + "lines": [ + { + "bbox": [ + 139, + 598, + 279, + 606 + ], + "spans": [ + { + "bbox": [ + 139, + 598, + 279, + 606 + ], + "type": "text", + "content": "Question: Which month is the hottest on average in Adelaide?" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 217, + 651, + 394, + 662 + ], + "lines": [ + { + "bbox": [ + 217, + 651, + 394, + 662 + ], + "spans": [ + { + "bbox": [ + 217, + 651, + 394, + 662 + ], + "type": "text", + "content": "Fig. 5: Comparison on predicted examples." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 167, + 609, + 230, + 618 + ], + "blocks": [ + { + "bbox": [ + 139, + 609, + 161, + 616 + ], + "lines": [ + { + "bbox": [ + 139, + 609, + 161, + 616 + ], + "spans": [ + { + "bbox": [ + 139, + 609, + 161, + 616 + ], + "type": "text", + "content": "Options:" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 609, + 230, + 618 + ], + "lines": [ + { + "bbox": [ + 167, + 609, + 230, + 618 + ], + "spans": [ + { + "bbox": [ + 167, + 609, + 230, + 618 + ], + "type": "table", + "html": "
(A) June, July, and August
", + "image_path": "8205a5b75cc22a8789b2d2e198bcdb96c341cfcfc3e9e36ca23ec89a0c34a173.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 168, + 620, + 211, + 628 + ], + "lines": [ + { + "bbox": [ + 168, + 620, + 211, + 628 + ], + "spans": [ + { + "bbox": [ + 168, + 620, + 211, + 628 + ], + "type": "text", + "content": "(B) April and May" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 168, + 632, + 225, + 639 + ], + "lines": [ + { + "bbox": [ + 168, + 632, + 225, + 639 + ], + "spans": [ + { + "bbox": [ + 168, + 632, + 225, + 639 + ], + "type": "text", + "content": "(C) January and February" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 22 + }, + { + "bbox": [ + 291, + 531, + 339, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 531, + 339, + 539 + ], + "spans": [ + { + "bbox": [ + 291, + 531, + 339, + 539 + ], + "type": "text", + "content": "Multimodal-CoT" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 290, + 546, + 370, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 546, + 370, + 598 + ], + "spans": [ + { + "bbox": [ + 290, + 546, + 370, + 598 + ], + "type": "text", + "content": "Scientists record climate data from places around the world .... The average temperatures in April and February are around over " + }, + { + "bbox": [ + 290, + 546, + 370, + 598 + ], + "type": "inline_equation", + "content": "25^{\\circ}\\mathrm{C}" + }, + { + "bbox": [ + 290, + 546, + 370, + 598 + ], + "type": "text", + "content": " . These months have the highest average temperatures of all of the months. So, they are hotter than the other months.." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 291, + 604, + 369, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 291, + 604, + 369, + 618 + ], + "spans": [ + { + "bbox": [ + 291, + 604, + 369, + 618 + ], + "type": "text", + "content": "The answer is (A). June, July, and August." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 385, + 531, + 410, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 531, + 410, + 539 + ], + "spans": [ + { + "bbox": [ + 385, + 531, + 410, + 539 + ], + "type": "text", + "content": "MC-CoT" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 386, + 546, + 469, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 546, + 469, + 598 + ], + "spans": [ + { + "bbox": [ + 386, + 546, + 469, + 598 + ], + "type": "text", + "content": "Scientists record climate data from places around the world .... The average temperatures in January and February are between above " + }, + { + "bbox": [ + 386, + 546, + 469, + 598 + ], + "type": "inline_equation", + "content": "20^{\\circ}\\mathrm{F}" + }, + { + "bbox": [ + 386, + 546, + 469, + 598 + ], + "type": "text", + "content": " . These months have the highest average temperatures of all of the months. So, they are hotter than the other months." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 386, + 604, + 468, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 604, + 468, + 618 + ], + "spans": [ + { + "bbox": [ + 386, + 604, + 468, + 618 + ], + "type": "text", + "content": "The answer is (C) January and February." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "text", + "content": "MC-CoT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "The first example highlights a scenario where the Multimodal-CoT model encountered difficulties in generating a coherent and accurate rationale, ultimately leading to an erroneous conclusion. This misstep underscores the challenges inherent in integrating multiple sources of information and extracting a logical reasoning path. Conversely, the MC-CoT model demonstrated its proficiency by navigating these complexities successfully, thereby arriving at the correct answer. This contrast exemplifies the significance of robust reasoning capabilities in achieving high accuracy in complex question-answering tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 483, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 483, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 483, + 331 + ], + "type": "text", + "content": "The second example delves into a different set of challenges faced by both models. Here, the Multimodal-CoT model stumbled over the interpretation of data presented in a chart, a mistake that significantly derailed its reasoning process and resulted in an incorrect answer. This error highlights the critical importance of accurate data interpretation skills in models designed for multimodal contexts. On the other hand, the MC-CoT model, while generally more adept in this instance, exhibited a lapse in understanding units of measurement—mistaking degrees Celsius " + }, + { + "bbox": [ + 130, + 213, + 483, + 331 + ], + "type": "inline_equation", + "content": "(^{\\circ}\\mathrm{C})" + }, + { + "bbox": [ + 130, + 213, + 483, + 331 + ], + "type": "text", + "content": " for degrees Fahrenheit " + }, + { + "bbox": [ + 130, + 213, + 483, + 331 + ], + "type": "inline_equation", + "content": "(^{\\circ}\\mathrm{F})" + }, + { + "bbox": [ + 130, + 213, + 483, + 331 + ], + "type": "text", + "content": ". Interestingly, this error did not compromise the integrity of its final answer, suggesting a certain resilience in the model's reasoning process despite minor inaccuracies." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 351, + 220, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 351, + 220, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 351, + 220, + 364 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 378, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 582 + ], + "type": "text", + "content": "In this study, we conducted an exhaustive analysis of the Multimodal-CoT framework and identified that the quality of rationales plays a crucial role in enhancing the performance of multimodal reasoning models. Recognizing the critical importance of this aspect, we developed an innovative solution aimed at improving the quality of rationales. Our solution employs a self-consistency training strategy that leverages the natural variability introduced by dropout techniques during training to foster a consensus mechanism through voting. This approach is supported by both theoretical underpinnings and empirical evidence, highlighting its effectiveness and wide applicability. We meticulously designed experiments to evaluate the impact of our self-consistency training strategy on the performance of multimodal reasoning models across diverse scenarios and datasets. The results from these experiments were revealing, clearly showing that our method significantly boosts the capabilities of multimodal models. A particularly significant finding was our approach's ability to enhance the performance of smaller models, enabling them to match or even exceed the capabilities of much larger models. We believe our strategy offers a promising avenue for advancing multimodal reasoning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 582, + 483, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 483, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 483, + 665 + ], + "type": "text", + "content": "A limitation of our approach is that we do not explicitly impose grammatical constraints, meaning the generated rationales are not guaranteed to be grammatically correct. It is worth noting that the Multimodal-CoT framework also does not guarantee the grammatical validity of the generated rationales. Despite this, empirical evidence has shown that our approach can yield superior results across various datasets, indicating that the rationales produced by our approach can significantly enhance the model's reasoning ability towards the correct answers." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "type": "text", + "content": "Tan. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 261, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 261, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 261, + 129 + ], + "type": "text", + "content": "7 Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 234 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 234 + ], + "type": "text", + "content": "This work was supported by the Science & Technology Innovation 2030 Major Program Project No. 2021ZD0150100, National Natural Science Foundation of China Project No. U21A20427, Project No. WU2022A009 from the Center of Synthetic Biology and Integrated Bioengineering of Westlake University, Project No. WU2023C019 from the Westlake University Industries of the Future Research. Finally, we thank the Westlake University HPC Center for providing part of the computational resources, and Project No. 23-407-3-29 from the Shenyang Science and Technology Program." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 251, + 197, + 263 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 197, + 263 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 197, + 263 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 274, + 481, + 663 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 138, + 274, + 481, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 274, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 481, + 317 + ], + "type": "text", + "content": "1. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6077-6086 (2018) 4, 9, 10" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 318, + 481, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 318, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 138, + 318, + 481, + 350 + ], + "type": "text", + "content": "2. Anil, R., Dai, A.M., First, O., Johnson, M., Lepikhin, D., Passos, A., Shakeri, S., Taropa, E., Bailey, P., Chen, Z., et al.: Palm 2 technical report. arXiv preprint arXiv:2305.10403 (2023) 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 350, + 481, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 350, + 481, + 383 + ], + "spans": [ + { + "bbox": [ + 138, + 350, + 481, + 383 + ], + "type": "text", + "content": "3. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: Vqa: Visual question answering. In: Proceedings of the IEEE international conference on computer vision. pp. 2425-2433 (2015) 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 384, + 481, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 384, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 138, + 384, + 481, + 415 + ], + "type": "text", + "content": "4. Askell, A., Bai, Y., Chen, A., Drain, D., Ganguli, D., Henighan, T., Jones, A., Joseph, N., Mann, B., DasSarma, N., et al.: A general language assistant as a laboratory for alignment. arXiv preprint arXiv:2112.00861 (2021) 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 416, + 481, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 416, + 481, + 448 + ], + "spans": [ + { + "bbox": [ + 138, + 416, + 481, + 448 + ], + "type": "text", + "content": "5. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966 (2023) 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 449, + 481, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 481, + 491 + ], + "type": "text", + "content": "6. Besta, M., Blach, N., Kubicek, A., Gerstenberger, R., Gianinazzi, L., Gajda, J., Lehmann, T., Podstawski, M., Niewiadomski, H., Nczyk, P., et al.: Graph of thoughts: Solving elaborate problems with large language models. arXiv preprint arXiv:2308.09687 (2023) 1, 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 491, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 481, + 534 + ], + "type": "text", + "content": "7. Biderman, S., Schoelkopf, H., Anthony, Q.G., Bradley, H., O'Brien, K., Hallahan, E., Khan, M.A., Purohit, S., Prashanth, U.S., Raff, E., et al.: Pythia: A suite for analyzing large language models across training and scaling. In: International Conference on Machine Learning. pp. 2397-2430. PMLR (2023) 9, 11" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 535, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 535, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 138, + 535, + 481, + 577 + ], + "type": "text", + "content": "8. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 578, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 481, + 620 + ], + "type": "text", + "content": "9. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) 9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 621, + 481, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 621, + 481, + 663 + ], + "spans": [ + { + "bbox": [ + 134, + 621, + 481, + 663 + ], + "type": "text", + "content": "10. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) 11" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "text", + "content": "MC-CoT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "11. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020) 11" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "type": "text", + "content": "12. Chen, Z., Zhou, Q., Shen, Y., Hong, Y., Zhang, H., Gan, C.: See, think, confirm: Interactive prompting between vision and language models for knowledge-based visual reasoning. arXiv preprint arXiv:2301.05226 (2023) 9, 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 481, + 215 + ], + "type": "text", + "content": "13. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311 (2022) 1, 11" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 481, + 248 + ], + "type": "text", + "content": "14. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, E., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022) 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 133, + 249, + 481, + 281 + ], + "type": "text", + "content": "15. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022) 9, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 282, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 282, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 133, + 282, + 481, + 314 + ], + "type": "text", + "content": "16. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Qiu, Z., Lin, W., Yang, J., Zheng, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023) 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 133, + 315, + 481, + 346 + ], + "type": "text", + "content": "17. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023) 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 347, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 347, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 133, + 347, + 481, + 390 + ], + "type": "text", + "content": "18. Gao, P., Jiang, Z., You, H., Lu, P., Hoi, S.C., Wang, X., Li, H.: Dynamic fusion with intra-and inter-modality attention flow for visual question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6639-6648 (2019) 4, 9, 10" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 392, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 392, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 133, + 392, + 481, + 423 + ], + "type": "text", + "content": "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 424, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 424, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 133, + 424, + 481, + 456 + ], + "type": "text", + "content": "20. Horawalavithana, S., Munikoti, S., Stewart, I., Kvinge, H.: Scitune: Aligning large language models with scientific multimodal instructions. arXiv preprint arXiv:2307.01139 (2023) 9, 10" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 457, + 481, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 457, + 481, + 478 + ], + "spans": [ + { + "bbox": [ + 133, + 457, + 481, + 478 + ], + "type": "text", + "content": "21. Huang, J., Chang, K.C.C.: Towards reasoning in large language models: A survey. arXiv preprint arXiv:2212.10403 (2022) 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 133, + 479, + 481, + 511 + ], + "type": "text", + "content": "22. Kamath, A., Clark, C., Gupta, T., Kolve, E., Hoiem, D., Kembhavi, A.: Webly supervised concept expansion for general purpose vision models. In: European Conference on Computer Vision. pp. 662-681. Springer (2022) 9, 11" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 133, + 512, + 481, + 544 + ], + "type": "text", + "content": "23. Kenton, J.D.M.W.C., Toutanova, L.K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of naacL-HLT. vol. 1, p. 2 (2019) 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 545, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 545, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 545, + 481, + 609 + ], + "type": "text", + "content": "24. Khashabi, D., Min, S., Khot, T., Sabharwal, A., Tafjord, O., Clark, P., Hajishirzi, H.: UNIFIEDQA: Crossing format boundaries with a single QA system. In: Cohn, T., He, Y., Liu, Y. (eds.) Findings of the Association for Computational Linguistics: EMNLP 2020. pp. 1896-1907. Association for Computational Linguistics, Online (Nov 2020). https://doi.org/10.18653/v1/2020-findings-emnlp.171, https://aclanthology.org/2020_findings-emnlp.171_9, 10, 11" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 611, + 481, + 632 + ], + "type": "text", + "content": "25. Kim, J.H., Jun, J., Zhang, B.T.: Bilinear attention networks. Advances in neural information processing systems 31 (2018) 4, 9, 10" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 481, + 665 + ], + "type": "text", + "content": "26. Kim, W., Son, B., Kim, I.: Vilt: Vision-and-language transformer without convolution or region supervision. In: International Conference on Machine Learning. pp. 5583-5594. PMLR (2021) 4, 9, 10" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "type": "text", + "content": "Tan. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "27. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) 4, 9, 11" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 151, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 151, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 151, + 481, + 184 + ], + "type": "text", + "content": "28. Li, L.H., Yatskar, M., Yin, D., Hsieh, C.J., Chang, K.W.: What does bert with vision look at? In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. pp. 5265-5275 (2020) 4, 9, 10" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 481, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 481, + 206 + ], + "type": "text", + "content": "29. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023) 2, 9, 10" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 207, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 481, + 239 + ], + "type": "text", + "content": "30. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281 (2023) 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "text", + "content": "31. Lu, J., Batra, D., Parikh, D., Lee, S.: Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems 32 (2019) 9, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 481, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 318 + ], + "type": "text", + "content": "32. Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems 35, 2507-2521 (2022) 2, 3, 4, 9, 10" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "type": "text", + "content": "33. Lu, P., Peng, B., Cheng, H., Galley, M., Chang, K.W., Wu, Y.N., Zhu, S.C., Gao, J.: Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842 (2023) 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "type": "text", + "content": "34. Lu, P., Peng, B., Cheng, H., Galley, M., Chang, K.W., Wu, Y.N., Zhu, S.C., Gao, J.: Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842 (2023) 9, 10" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 481, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 419 + ], + "type": "text", + "content": "35. Lu, P., Qiu, L., Chen, J., Xia, T., Zhao, Y., Zhang, W., Yu, Z., Liang, X., Zhu, S.C.: Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. arXiv preprint arXiv:2110.13214 (2021) 4, 9, 10" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "type": "text", + "content": "36. Luo, G., Zhou, Y., Ren, T., Chen, S., Sun, X., Ji, R.: Cheap and quick: Efficient vision-language instruction tuning for large language models. arXiv preprint arXiv:2305.15023 (2023) 2, 4, 9, 10" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 453, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 497 + ], + "type": "text", + "content": "37. Marino, K., Chen, X., Parikh, D., Gupta, A., Rohrbach, M.: Krisp: Integrating implicit and symbolic knowledge for open-domain knowledge-based vqa. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14111-14121 (2021) 9, 11" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 498, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 520 + ], + "type": "text", + "content": "38. Ning, X., Lin, Z., Zhou, Z., Yang, H., Wang, Y.: Skeleton-of-thought: Large language models can do parallel decoding. arXiv preprint arXiv:2307.15337 (2023) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 521, + 343, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 343, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 343, + 531 + ], + "type": "text", + "content": "39. OpenAI: Gpt-4 technical report (2023) 1, 9, 10" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 532, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 481, + 576 + ], + "type": "text", + "content": "40. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022) 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 577, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 481, + 620 + ], + "type": "text", + "content": "41. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 11" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "42. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020) 1, 11" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "text", + "content": "MC-CoT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "43. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020) 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "type": "text", + "content": "44. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems 28 (2015) 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 216 + ], + "type": "text", + "content": "45. Schwenk, D., Khandelwal, A., Clark, C., Marino, K., Mottaghi, R.: A-okvqa: A benchmark for visual question answering using world knowledge. In: European Conference on Computer Vision. pp. 146-162. Springer (2022) 3, 4, 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "46. Tan, H., Bansal, M.: Lxmert: Learning cross-modality encoder representations from transformers. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (2019) 9, 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "47. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "text", + "content": "48. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "text", + "content": "49. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 390 + ], + "type": "text", + "content": "50. Wang, X., Wei, J., Schuurmans, D., Le, Q.V., Chi, E.H., Narang, S., Chowdhery, A., Zhou, D.: Self-consistency improves chain of thought reasoning in language models. In: The Eleventh International Conference on Learning Representations (2022) 1, 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 391, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 481, + 434 + ], + "type": "text", + "content": "51. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022) 1, 3, 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 434, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 481, + 456 + ], + "type": "text", + "content": "52. Wu, C., Liu, J., Wang, X., Dong, X.: Chain of reasoning for visual question answering. Advances in Neural Information Processing Systems 31 (2018) 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 456, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 481, + 489 + ], + "type": "text", + "content": "53. Wu, C., Yin, S., Qi, W., Wang, X., Tang, Z., Duan, N.: Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671 (2023) 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "type": "text", + "content": "54. Yang, Z., Gan, Z., Wang, J., Hu, X., Lu, Y., Liu, Z., Wang, L.: An empirical study of gpt-3 for few-shot knowledge-based vqa. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 3081-3089 (2022) 9, 11" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 522, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 555 + ], + "type": "text", + "content": "55. Yang, Z., Li, L., Wang, J., Lin, K., Azarnasab, E., Ahmed, F., Liu, Z., Liu, C., Zeng, M., Wang, L.: Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381 (2023) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 555, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 587 + ], + "type": "text", + "content": "56. Yao, S., Yu, D., Zhao, J., Shafran, I., Griffiths, T.L., Cao, Y., Narasimhan, K.: Tree of thoughts: Deliberate problem solving with large language models. arXiv preprint arXiv:2305.10601 (2023) 1, 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 588, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 481, + 610 + ], + "type": "text", + "content": "57. Yin, S., Fu, C., Zhao, S., Li, K., Sun, X., Xu, T., Chen, E.: A survey on multimodal large language models. arXiv preprint arXiv:2306.13549 (2023) 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "text", + "content": "58. Yu, Z., Yu, J., Cui, Y., Tao, D., Tian, Q.: Deep modular co-attention networks for visual question answering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6281-6290 (2019) 4, 9, 10" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "59. Zhang, P., Li, X., Hu, X., Yang, J., Zhang, L., Wang, L., Choi, Y., Gao, J.: Vinvl: Revisiting visual representations in vision-language models. In: Proceedings of the" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "type": "text", + "content": "Tan. et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 270 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "text", + "content": "IEEE/CVF conference on computer vision and pattern recognition. pp. 5579-5588 (2021) 11" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 482, + 171 + ], + "type": "text", + "content": "60. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199 (2023) 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "61. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199 (2023) 9, 10" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "62. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022) 1, 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "text", + "content": "63. Zhang, Z., Zhang, A., Li, M., Zhao, H., Karypis, G., Smola, A.: Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923 (2023) 2, 4, 5, 6, 9, 10, 11" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 409, + 91, + 447, + 100 + ], + "type": "text", + "content": "MC-CoT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing/3e12da5b-7b20-4ba8-b65d-4a8790f34fde_content_list.json b/2024/Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing/3e12da5b-7b20-4ba8-b65d-4a8790f34fde_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..32c507a28bb257076d8024aa2d74cf533953396e --- /dev/null +++ b/2024/Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing/3e12da5b-7b20-4ba8-b65d-4a8790f34fde_content_list.json @@ -0,0 +1,1804 @@ +[ + { + "type": "text", + "text": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing", + "text_level": 1, + "bbox": [ + 274, + 140, + 727, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Si-Qi Liu $^{1,\\ast,\\boxtimes}$ , Qirui Wang $^{1,\\ast}$ , and Pong C. Yuen $^{2}$", + "bbox": [ + 320, + 210, + 679, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Shenzhen Research Institute of Big Data", + "bbox": [ + 357, + 239, + 645, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Department of Computer Science, Hong Kong Baptist University siqiliu@sribd.cn, qwangcw@connect.ust.hk, pcyuen@comp.hkbu.edu.hk", + "bbox": [ + 254, + 253, + 746, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Face anti-spoofing (FAS) which plays an important role in securing face recognition systems has been attracting increasing attention. Recently, vision-language model CLIP has been proven to be effective for FAS, where outstanding performance can be achieved by simply transferring the class label into textual prompt. In this work, we aim to improve the generalization ability of CLIP-based FAS from a prompt learning perspective. Specifically, a Bottom-Up Domain Prompt Tuning method (BUDoPT) that covers the different levels of domain variance, including the domain of recording settings and domain of attack types is proposed. To handle domain discrepancies of recording settings, we design a context-aware adversarial domain-generalized prompt learning strategy that can learn domain-invariant prompt. For spoofing domain with different attack types, we construct a fine-grained textual prompt that guides CLIP to look through the subtle details of different attack instruments. Extensive experiments are conducted on five FAS datasets with variations of camera types, resolutions, image qualities, lighting conditions, and recording environments. The effectiveness of our proposed method is evaluated with different amounts of source domains from multiple angles, where we boost the generalizability compared with the state of the arts with multiple or limited numbers of training datasets.", + "bbox": [ + 259, + 311, + 743, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Face Anti-spoofing $\\cdot$ Face presentation attack detection $\\cdot$ Face liveness detection", + "bbox": [ + 259, + 601, + 740, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 652, + 375, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the wide deployment of face recognition in real-world scenarios, the security concerns of FR systems attract increasing attention. Face images can be easily acquired from social networks, and the costs of conducting a face spoofing attack can be very low. Print or screens of mobile phones or tablets are the two common low-cost spoofing instruments. 3D mask attack could be more challenging since both the 3D shape and skin texture can be imitated. To counter these presentation attacks, a series of face anti-spoofing (FAS) methods based on different liveness cues are proposed, various from the hand-crafted features [1,19,40] to the deep-learning-based representations [34,62,64].", + "bbox": [ + 212, + 681, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution, ⌒ Corresponding author", + "bbox": [ + 230, + 825, + 542, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Early works are mainly developed for intra-domain testing where the recording environment, devices, and attack types are identical. Since the main difference between genuine and attack samples is located on small vestiges such as the texture and color tone, the subtle variations introduced by recording cameras, lighting conditions, and attack mediums form the domain gaps in FAS.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To improve the robustness and generalizability, the generic domain adaptation and domain generalization approaches are applied or customized in the context of FAS [52,53,57]. Few-shot domain adaptation and unsupervised domain adaptation methods tend to have better results since they can learn from the testing data with or without the labels [21,23,54]. The zero-shot domain generalization setting is more practical as Since the testing data may not be accessible in real scenarios. Popular techniques such as adversarial learning, contrastive learning, or meta-learning are adopted to align representations from different domains and keep on boosting the performances [5,44]. Directly constructing generalized liveness cues can also be effective. The facial heartbeat signal (rPPG) based methods exhibit strong generalizability to different types of attacks but existing methods still require long observation time [29,30].", + "bbox": [ + 212, + 226, + 784, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, vision-language models like CLIP exhibited astonishing generalizability and zero-shot performance with the joint embedding of image-text pairs. The first trial of CLIP for FAS (FLIP) that transforms binary class labels into text and measures their similarity to image embeddings also achieves high performance [49]. However, since CLIP is trained for generic tasks with image pairs from the internet, the domain discrepancy of lighting conditions, camera qualities, or attack types in FAS can hardly be well addressed by simply finetuning it with the image-to-class-text contrastive loss. Intuitively, we may enrich the class text prompt with domain labels like the camera types, lighting conditions, or recording scene and construct hand-crafted prompts such as {a photo of [CLS] captured under [Recording Setting]} to learn a fine-grained representation. However, this approach may not work since 1) the detailed domain information in FAS can hardly be defined with text since it is related to multiple factors; 2) learning fine-grained representation that can tell domain discrepancies may contradict the final classification of genuine and spoofing samples.", + "bbox": [ + 212, + 412, + 785, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we aim to improve the generalization ability of CLIP-based FAS from a prompt learning perspective. Specifically, we proposed to learn a domain-invariant prompt from the visual embedding that can guide CLIP to avoid interference from the specific domain. Inspired by the CLIP style contrastive learning objective which pulls images to their textual descriptions while pushes away unmatched pairs, we achieve domain-generalized prompt learning by pulling together image embeddings with a domain-invariant prompt and pushing away the image embeddings with their domain-specific prompt. Considering the discrepancies of recording domain are low-level features that scatter in the entire image, we extract domain-specific information from both the facial content and background context regions as the domain-specific negative prompt (push away). The domain-generalized prompt embedding extract from multiple domains is regarded as the positive prompt (pull together). An adversarial", + "bbox": [ + 212, + 643, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S.Q. Liu et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4747e110a4a9b784448f68ce50222b72f696944a1937a2f8c6d7f0779f9d8c8e.jpg", + "image_caption": [ + "(a) Intuitive solution of CLIP for Face Anti-Spoofing" + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 444, + 220 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6dfca787fa20bd353a89eaca2669cb1e90e7bd5210b8a12d28402ea5b90d5fd8.jpg", + "image_caption": [ + "(b) Bottom-Up Domain-Generalized Prompt Tuning for FAS (ours)", + "Fig. 1: The comparison between (a) the intuitive solution of CLIP for face anti-spoofing and (b) the proposed bottom-up domain-generalized prompt tuning. The Domain Token prompt guides the model to learn domain-generalized FAS features for different recording domains. The Attack Type prompt boosts the discriminability by encouraging the model to differentiate fine-grained attack instruments. The Domain Token is learned from the visual embedding through the Domain Prompt Learner." + ], + "image_footnote": [], + "bbox": [ + 467, + 143, + 771, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "learning strategy is designed to further ensure the prompt learner can learn the domain-generalized information. The various attack instruments for spoofing samples can be regarded as another perspective of domains. Different from the treatment of recording domain prompt learning, we want the visual encoder to distinguish the attack types so that it can learn more discriminative FAS features [20, 24]. In this case, we form a textual prompt that guides the generic visual encoder to adapt to the FAS task and look through the fine-grained difference for various attack types. We refer to this newly proposed method that learns the low-level domain-generalized visual prompt and tunes high-level attack-type textual prompt as the Bottom-Up Domain Prompt Tuning (BUDoPT) for generalized face anti-spoofing.", + "bbox": [ + 212, + 364, + 787, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Extensive experiments are conducted on five FAS datasets with a large number of variations (camera types, resolutions, image qualities, lighting conditions, and recording environments). The effectiveness of our proposed method is evaluated with different amounts of source domains from multiple angles, where we boost the generalizability compared with the state of the arts with multiple training datasets or with only one dataset.", + "bbox": [ + 212, + 534, + 787, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In sum, the main contributions of this work are:", + "bbox": [ + 238, + 626, + 589, + 640 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We improve the domain generalizability of FAS from a prompt tuning perspective within the context of vision-language model CLIP. A Bottom-Up Domain Prompt Tuning method (BUDoPT) that covers the different levels of domain variance, including the domain of recording settings and domain of attack types is proposed.", + "- We design 1) a context-aware adversarial domain-generalized prompt learning strategy that can learn domain-invariant prompt for different recording settings, 2) an effective textual prompt that guides CLIP to adapt to FAS task and look through the subtle details of different attack types.", + "- We improve the generalizability compared with the state of the arts with different scales of training domains, and give in-depth analysis with extensive experiments on five FAS datasets that cover typical variations in FAS." + ], + "bbox": [ + 223, + 655, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing", + "bbox": [ + 253, + 114, + 732, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 387, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Face Anti-Spoofing. Face anti-spoofing has attracted increasing attention in the last decades, developed from the early handcrafted liveness feature construction [37, 60] to the latest deep learning-based design [64]. Early works focus on searching the discriminative liveness cues such as the appearance artifacts [31, 42, 58], and the motion difference between genuine faces and static attack instruments. Accordingly, hand-crafted features, such as SIFT [41], HoG [60], and LBP [3] are used to characterize spoof patterns. More discriminative features can be learned with Deep learning-based [38, 59]. Auxiliary tasks, such as the regression of facial depth map [32], 3D structure [12], and material [61], encourage the network to learn fine-grained appearance details. There are also classification-based methods [34, 51, 59] and Generative models [22, 33, 57]. Spatiotemporal information can also be learned through CNN and LSTM framework [9, 39].", + "bbox": [ + 212, + 176, + 787, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generalized FAS. Generalized FAS that can perform robustly on different domains are necessary for real-world application scenarios. Recent works proposed to learn the FAS model on multiple source domains and evaluated a target domain. Under the domain adaptation settings where the target domain is accessible, aligning the source and target feature distributions is the typical solution [21,52,53,65,71]. Adversarial learning strategies or the triplet loss encourage the model to learn shared feature for generalization [20,47,57]. Meta-learning and continue learning approaches are also proposed by [5,6,27,44,48] to simulate the domain shift during training.", + "bbox": [ + 212, + 369, + 787, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prompt Learning. Prompt learning originated from the field of Natural Language Processing (NLP) motives to shape the output of pre-trained language models by providing contextual information or constraints, thereby enhancing their ability to adapt to downstream tasks [43]. Prompt engineering was extended from NLP to Visual-Language Models [45], such as CLIP which classifies images by matching them with class prompts. CoOp [69] is designed to turn the handcrafted textual prompt into a learnable visual prompt. CoCoOp [68] incorporates the image as the condition for prompts and improves the model's generalization capabilities. Within the visual-textual contrastive learning framework, generalizability can be improved with a smaller data scale. CFPL1 [25] and this work aims to improve the generalizability for FAS with prompt learning. Different from CFPL which regards style (in AdaIN [18]) as the domain-specific cue, in this work we learn the domain-invariant prompt from an intrinsic CLIP-style contrastive learning perspective.", + "bbox": [ + 212, + 518, + 787, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Preliminary on CLIP for FAS", + "text_level": 1, + "bbox": [ + 215, + 753, + 540, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CLIP [45] comprises an image encoder denoted $f_{v}$ and a text encoder denoted $f_{t}$ . In the context of the FAS, the two class text prompt $p_{k}$ encompasses the", + "bbox": [ + 212, + 785, + 785, + 816 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S.Q. Liu et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1 CFPL was published after this manuscript was submitted for review.", + "bbox": [ + 217, + 824, + 687, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "descriptions of real and spoof faces, representing two classes. CLIP determines the similarity between the input image $\\pmb{x}$ and $\\pmb{p}$ instead of directly training the classification head. The model predicts the image's category $\\hat{y}$ by selecting the class prompt with the highest response:", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} = \\underset {k \\in \\{\\mathrm {r}, \\mathrm {s} \\}} {\\operatorname {a r g m a x}} \\left\\langle f _ {v} (\\boldsymbol {x}), f _ {t} \\left(\\boldsymbol {p} _ {k}\\right) \\right\\rangle \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 220, + 785, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\langle ,\\rangle$ denotes the cosine similarity. $p_r$ and $p_s$ represent the prompt with real and spoof face descriptions.", + "bbox": [ + 212, + 256, + 784, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With such learning framework, pretrained CLIP exhibits great performance in FAS with the standard finetuning process [49]. However, since CLIP is trained for generic tasks with image pairs from the internet, when handling the low-level feature in cross-domain settings in FAS, the domain discrepancy of lighting conditions, camera qualities, or attack types can hardly be well addressed.", + "bbox": [ + 212, + 287, + 785, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Proposed method", + "text_level": 1, + "bbox": [ + 215, + 386, + 423, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overview. The Bottom-Up Domain Prompt Tuning method (BUDoPT) is proposed to improve the generalization ability of FAS in the context of the vision language model CLIP. On top of the intuitive CLIP-based face anti-spoofing which turns the class label into textual prompt, we incorporate the domain token for the domain discrepancies from different recording settings and the attack type token that describes the fine-grained types of attacks. First, the context-aware adversarial domain-generalized prompt learning is to train the domain prompt learner given the input of facial visual feature. A CLIP score style domain prompt alignment mechanism and an adversarial learning strategy are designed to learn a domain-generalized prompt. Second, the textual prompt of fine-grained attack types aims to boost discriminability by looking through the subtle details of different attack instruments.", + "bbox": [ + 212, + 417, + 787, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Context-aware Domain Generalized Prompt Learning", + "text_level": 1, + "bbox": [ + 215, + 621, + 702, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recording Domain Generalized Prompt Learning. Although text prompt can effectively assist in classifying concrete instances in upper-stream tasks, it encounters difficulties in describing the recording domain with various complex factors, including camera qualities, lighting conditions, and blurriness in FAS. Therefore we proposed the to learn domain-generalized prompt from the visual embedding. Inspired by [68], we employ vision domain features as conditional inputs to generate learnable prompts. The primary objective of the component is learning domain-invariant prompts to mitigate the impact of domain gaps.", + "bbox": [ + 212, + 646, + 787, + 766 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we propose a Prompt learner $PL$ to obtain the domain prompt from the image $\\mathbf{x}$ . $PL$ utilizes the features from $i$ -th layer of $f_{v}$ as the input. The domain feature $D$ is defined as:", + "bbox": [ + 212, + 767, + 787, + 811 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD (\\boldsymbol {x}) = P L \\left(f _ {v, i} (\\boldsymbol {x})\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 428, + 825, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing", + "bbox": [ + 253, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9b496ce7dbef0d00662e55612e172142ea398b9c5e9f237694769ceff3d568a9.jpg", + "image_caption": [ + "Fig. 2: The proposed Bottom-Up Domain Prompt Tuning method (BUDoPT) covers the different levels of domain variance for generalized FAS, including the domain of recording settings and domain of attack types. The context-aware adversarial domain-generalized prompt learning can learn domain-invariant prompt for different recording settings. The textual prompt of fine-grained attack types can boost the discriminability by looking through the subtle details of different attack instruments." + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 769, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $f_{v,i}$ denotes $i$ -layer features of $f_v$", + "bbox": [ + 212, + 518, + 504, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To accurately represent the domain feature while minimizing influences from the content feature of a single image, we utilize all images' domain features in a mini-batch to generate the domain prompt. Given the training mini-batch $\\{\\pmb{x}_n\\}_{n=1}^N$ for domain $m$ , the domain prompt $\\pmb{p}_{D_m}$ is defined as:", + "bbox": [ + 212, + 532, + 787, + 595 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} _ {D _ {m}} = f _ {t} \\left(\\frac {1}{N} \\sum_ {n = 1} ^ {N} D \\left(\\boldsymbol {x} _ {n} ^ {m}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 604, + 785, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $N$ is the mini-batch size and $x_{n}^{m}$ denotes the $n$ -th image from $m$ .", + "bbox": [ + 212, + 656, + 730, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inspired by the CLIP style contrastive learning objective, we design a CLIP score-based domain generalization learning strategy that pulls together image embeddings with a domain-invariant prompt and pushes away the image embeddings with their domain-specific prompt. Given a batch of training data with $M$ domains, the generalized domain prompt $\\pmb{p}_{GD}$ can be defined as:", + "bbox": [ + 212, + 671, + 787, + 748 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} _ {G D} = f _ {t} \\left(\\frac {1}{N * M} \\sum_ {m = 1} ^ {M} \\sum_ {n = 1} ^ {N} D \\left(\\boldsymbol {x} _ {n} ^ {m}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 758, + 785, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As the Figure 3 shows, only the labels belonging to the generalized domain can be marked as positive while other domain-specific prompts are always labeled", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S.Q. Liu et al.", + "bbox": [ + 269, + 114, + 367, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/90a415da0d07628e0f3f4cde1ba8e925ba73509daa935f6f5ad413bffdb4275a.jpg", + "image_caption": [ + "Fig. 3: Illustration of the CLIP score-based domain generalization learning strategy, where we pull together image embeddings with a domain-generalized prompt (▲) and push away the image embeddings with their domain-specific prompt" + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 754, + 236 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "as negative. We denote $\\pmb{p}$ as the set of prompts and then calculate the image cross-entropy loss $\\mathcal{L}_{ce}$ by:", + "bbox": [ + 212, + 315, + 782, + 345 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c e} = - \\sum_ {z = 1} ^ {2 \\times (M + 1)} y \\log \\left(\\frac {\\exp \\left(\\left\\langle f _ {v} (\\boldsymbol {x}) , f _ {t} \\left(\\boldsymbol {p} _ {z}\\right) \\right\\rangle\\right)}{\\sum_ {k = 1} ^ {2 \\times (M + 1)} \\exp \\left(\\left\\langle f _ {v} (\\boldsymbol {x}), f _ {t} \\left(\\boldsymbol {p} _ {k}\\right) \\right\\rangle\\right)}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 300, + 353, + 785, + 396 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $y$ is the ground truth label, $k$ and $z$ denote the indices in $\\pmb{p}$ . Note that we have $2 \\times (M + 1)$ classes in total, i.e., $M$ domains plus $\\pmb{p}_{GD}$ for each class since the prompt set $\\pmb{p} = \\{\\pmb{p}_{D_1}^{real}, \\dots, \\pmb{p}_{D_M}^{real}, \\pmb{p}_{GD}^{real}, \\pmb{p}_{D_1}^{fake}, \\dots, \\pmb{p}_{D_M}^{fake}, \\pmb{p}_{GD}^{fake}\\}$ ,", + "bbox": [ + 212, + 402, + 782, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "During inference, we assume that all samples in the batch are sourced from the same domain and possess the identical prompt. The model selects the class prompt with the highest similarity score as the final prediction.", + "bbox": [ + 212, + 450, + 782, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Context-aware Domain Clue. To better learn the domain-generalized prompt, we try to extract more domain-specific information from the visual embedding as it is treated as the negative label to be pushed away. Considering the discrepancies of recording domain are low-level features that scatter in the entire image, we extract domain-specific information from both the facial content and background context regions. Specifically, we want to highlight the context information from the surrounding background regions with purer domain information without disturbance of facial content. Consequently, we classify the output patches of $f_{v}$ into two types: context patches $c^{ctx}$ and content patches $c^{ctn}$ . Then the Equation (2) can be rewritten as follows:", + "bbox": [ + 212, + 503, + 787, + 654 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nD (x) = P L ([ c _ {i} ^ {c t x}, c _ {i} ^ {c t n} ]), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 661, + 785, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $c_{i}^{ctx}$ denotes the context patch obtained from $f_{v,i}$ . To enhance the representation of domain features while reducing the influence of facial content, we extract context patches $c_{s}^{ctx}$ from a shallower layer ( $s < i$ ) of $f_{v}$ and replace the original $c_{i}^{ctx}$ , as shown in Figure 2(green block). For the VIT visual encoder in CLIP, the context patches in shallower layers preserve low-level recording domain information as they have less interaction with content patches. Hence, the domain feature is formed as:", + "bbox": [ + 212, + 686, + 787, + 791 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nD (x) = P L \\left(\\left[ c _ {s} ^ {c t x}, c _ {i} ^ {c t n} \\right]\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 799, + 785, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to emphasize the domain-specific information in the prompts.", + "bbox": [ + 214, + 824, + 658, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing", + "bbox": [ + 253, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adversarial Domain Prompt Learning. We propose the Adversarial Domain Prompt Learning strategy to further ensure that the prompt learner can learn the domain-generalized information. A Discriminator $DC$ is constructed to differentiate the domain features and determine their belonging domain. $PL$ , as the generator, is required to minimize the domain-specific trace in the output to fool $DC$ :", + "bbox": [ + 212, + 146, + 787, + 234 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a d v} = - \\frac {1}{N * M} \\sum_ {m = 1} ^ {M} \\sum_ {n = 1} ^ {N} y _ {m} \\log \\left(D C \\left(D \\left(\\boldsymbol {x} _ {n} ^ {m}\\right)\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 246, + 785, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$y_{m}$ is the domain label. $\\mathcal{L}_{adv}$ encourages discriminator $DC$ to try its best to find the domain-specific vestige so that domain prompt learner can learn the domain-invariant information. Empirically, the adversarial learning strategy aggravates the unstable loss convergence. To ensure a more consistent learning direction, we employ a distribution constraint between domain features themselves. The Jensen-Shannon divergence is adopted to pull every two features from $M$ domains together in a training batch:", + "bbox": [ + 214, + 296, + 787, + 404 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d i s} = \\sum_ {j, t = 1, \\dots , N * M, j < t} \\operatorname {J S} \\left(D \\left(\\boldsymbol {x} _ {j}\\right), D \\left(\\boldsymbol {x} _ {t}\\right)\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 412, + 785, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here $N * M$ is the total number of samples in a batch.", + "bbox": [ + 215, + 457, + 611, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Spoofing Domain Discriminative Prompt Tuning.", + "text_level": 1, + "bbox": [ + 214, + 494, + 666, + 510 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In contrast to the visual recording domain generalized prompt that aims to avoid bias, the attack types of spoofing domain can be utilized to boost the discriminability. Specific artifacts are highly correlated with the attack instruments, which can be regarded as spoofing clues, such as the saturation change in printing attacks, the moiré pattern in video replay attacks, and the edge defects of 3D masks. Different from the low-level recording domain prompt, the attack types can be clearly described with the (high-level) textual prompt. Regularize the visual embedding with finer grained attack types text helps the visual encoder learn more discriminative FAS features. This will also set clearer boundaries between genuine faces and spoofing attacks thereby enhancing model generalizability to unseen attack types. Therefore, we propose the Spoof Domain Prompt tuning strategy to encourage the model to differentiate finer-grained attack types.", + "bbox": [ + 212, + 518, + 787, + 699 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specifically, we design a textual prompt template, a structured sentence that describes spoofing face photos with a placeholder of the attack types, such as \"a {fake} face photo {and it is the printing attack}\" Given $N * M$ samples in a training batch, we generate the same amount of text prompts $\\boldsymbol{p}\\boldsymbol{a}$ . Spoofing Domain Discriminative Prompt Tuning updates the parameters of $f_{v}$ with the text cross-entropy loss $\\mathcal{L}_t$ , which is calculated by:", + "bbox": [ + 214, + 700, + 787, + 791 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t} = - \\sum_ {o = 1} ^ {N * M} o \\log \\left(\\frac {\\exp \\left(\\left\\langle f _ {v} (\\boldsymbol {x}) , f _ {t} \\left(\\boldsymbol {p} a _ {o}\\right) \\right\\rangle\\right)}{\\sum_ {w = 1} ^ {N * M} \\exp \\left(\\left\\langle f _ {v} (\\boldsymbol {x}) , f _ {t} \\left(\\boldsymbol {p} a _ {w}\\right) \\right\\rangle\\right)}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 803, + 785, + 844 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S.Q. Liu et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $o$ and $w$ are the indices in $\\pmb{pa}$ .", + "bbox": [ + 215, + 146, + 482, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The final loss function of our method is formed as:", + "bbox": [ + 238, + 162, + 604, + 176 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\alpha \\mathcal {L} _ {c e} + \\beta \\mathcal {L} _ {t} - \\gamma \\mathcal {L} _ {a d v} + \\delta \\mathcal {L} _ {d i s}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 382, + 194, + 784, + 209 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\alpha, \\beta, \\gamma, \\delta$ are hyperparameters that balance each loss.", + "bbox": [ + 215, + 224, + 643, + 239 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 270, + 375, + 286 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Experimental Setups", + "text_level": 1, + "bbox": [ + 215, + 309, + 434, + 324 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Databases and Protocols. We evaluate our method with six widely used FAS databases: MSU-MFSD(M) [58], CASIA-FASD(C) [67], Idiap Replay attack (I) [8], OULU-NPU(O) [2], WMCA(W) [10] and CelebA-Sproof [66]. Five protocols that cover various domain generalization settings with different amounts of source domains are involved to comprehensively evaluate our method's performance. They are leave-one-out(P1) and one-to-one(P2) cross-domain generalization evaluation, different domain partition evaluation(P3), unseen attack type evaluation(P4) and various devices domain evaluation(P5).", + "bbox": [ + 212, + 340, + 787, + 460 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Metrics. Following [49], we evaluate the model performance using three metrics: Half Total Error Rate (HTER), Area Under the Receiver Operating Characteristic Curve (AUC) and True Positive Rate (TPR) at a False Positive Rate (FPR) $1\\%$ (TPR@FPR=1%). HTER and AUC reflect the theoretical performance. TPR@FPR=1% evaluates the model performance in practice.", + "bbox": [ + 212, + 462, + 787, + 537 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. We utilize FAN [4] to obtain facial landmarks, aligning the images based on nose and eye landmarks before cropping them to a size of 224x224. We employ the pre-trained CLIP-ViT-B-16 [45] as the foundation model. An eleven-layer VIT following a two-layer MLP is adopted as the prompt learner $PL$ . We extract content patches from the 5th layer of $f_{v}$ and context patches from the 1st layer. For adversarial learning an eight-layer 1D ResNet is adopted as the discriminator $DC$ .", + "bbox": [ + 212, + 537, + 787, + 643 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training strategy. We divide the training process into two stages for the Bottom-Up domain prompt tuning. In the first training stage, we conduct the Spoofing Domain Discriminative Prompt Tuning to boost the discriminability of $f_{v}$ . Adam optimizer is adopted with a learning rate of $10^{-6}$ . In the second training stage, we freeze $f_{v}$ and train the $PL$ for the Recording Domain Generalization Prompt Learning. The optimizer for both $PL$ and $DC$ are Adam with learning rate $10^{-3}$ and $3 * 10^{-4}$ . Before the adversarial learning, we first train at least one epoch for each domain source to warm up(exclude supplementary training data). Then, we take turns freezing the $PL$ and $DC$ for 50 iterations. For each stage we train 4000 iterations to ensure the largest dataset is traversed. The mini-batch size is $4(N = 4)$ for each domain. $M$ varies according to the number of datasets used for training under different evaluation protocols. The hyperparameter $\\alpha$ , $\\beta$ , $\\gamma$ , and $\\delta$ are set as 1, 1, 1, and 0.5, respectively.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing", + "bbox": [ + 253, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/37af7debb983da7c646e5a954cf72f3989ff97f547882201bd367ced3fbb18d1.jpg", + "table_caption": [ + "Table 1: Evaluation of cross-domain performance under leave-one-out protocol (P1)" + ], + "table_footnote": [ + "$\\dagger$ indicates CelebA-Sproof [66] is employed as supplementary training dataset" + ], + "table_body": "
MethodsOCI→MOMI→COCM→IICM→OAvg.
HTERAUCTPR@ FPR=1%HTERAUCTPR@ FPR=1%HTERAUCTPR@ FPR=1%HTERAUCTPR@ FPR=1%
MADDG [47]17.6988.06-24.5084.51-22.1984.99-27.9880.02-23.09
MDDR [53]17.0290.10-19.6887.43-20.8786.72-25.02581.47-20.64
NAS-FAS [64]16.8590.42-15.2192.64-11.6396.98-13.1694.18-14.21
RFMeta [48]13.8993.98-20.2788.16-17.3090.48-16.4591.16-16.97
D²AM [7]12.7095.6-20.9885.58-15.4391.22-15.2790.87-16.09
DRDG [28]12.4395.81-19.0588.79-15.5691.79-15.6391.75-15.66
Self-DA [55]15.4091.80-24.5084.40-15.6090.10-23.1084.30-19.65
ANRL [27]10.8396.75-17.8589.26-16.0391.04-15.6791.90-15.09
FGHV [26]9.1796.92-12.4793.47-16.2990.11-13.5893.55-12.87
SSDG-R [20]7.3897.17-10.4495.94-11.7196.59-15.6191.54-11.28
SSAB-R [57]6.6798.75-10.0096.67-8.8896.79-13.7293.65-9.80
PatchNet [51]7.1098.46-11.3394.58-13.4095.67-11.8295.07-10.90
GDA [71]9.2098.00-12.2093.00-10.0096.00-14.4092.60-11.45
LDCformer [16]6.4398.39-8.1196.67-8.5797.09-11.1795.58-8.57
TTDG-V [70]4.1698.48-7.5998.18-9.6298.18-1096.15-7.84
BUDoPT(ours)0.9599.7087.122.8598.0355.004.498.5486.462.2698.7855.252.62
ViT†[15]1.5899.6896.675.7098.9788.579.2597.1551.547.4798.4269.306.00
FLIP-V†[49]3.7999.3187.991.2799.7595.854.7198.8075.844.1598.7666.473.48
FLIP-IT†[49]5.2798.4279.330.4499.9899.862.9499.4284.623.6199.1584.763.06
FLIP-MCL†[49]4.9598.1174.670.5499.98100.004.2599.0784.622.3199.6392.283.01
BUDoPT-CoOp†1.9899.5480.631.9399.5888.063.1899.390.43.3398.9189.292.61
BUDoPT-CoCoOp†1.5199.4375.551.5599.5684.743.5399.292.533.5198.9887.982.53
BUDoPT(ours)†0.4099.9999.670.2699.9699.921.3899.6998.11.6099.5197.180.91
", + "bbox": [ + 215, + 180, + 785, + 434 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Evaluation", + "text_level": 1, + "bbox": [ + 215, + 500, + 351, + 513 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Cross-domain generalization evaluation. Following [49], the leave-one-out (P1) and one-to-one (P2) on M, C, I and O are conducted to evaluate cross-dataset performance. $\\mathrm{OCM} \\rightarrow \\mathrm{I}$ refers to the model trains on O, C, M and tests on I. For a fair comparison with [49] and other models, we evaluate the performance of our method with and without the supplementary training data CelebA-Sproof [66]. For simplicity, we regard dataset as the domain partition in our proposed method. As shown in Table 1, without supplementary training data, the proposed BUDoPT outperforms SOTA in all settings. With supplementary training data, BUDoPT also surpasses the FLIP family model in HTER for all settings and achieves more than $69\\%$ improvement on average. In terms of TPR@FPR $= 1\\%$ , our method performs better in three out of four settings. Although the performance slightly drops in TPR@FPR $= 1\\% (\\mathrm{C})$ and AUC(C,O), the margin is relatively small as they already exceed $99\\%$ .", + "bbox": [ + 212, + 522, + 785, + 718 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In P2 of low-data regime [49] the model is trained separately on M, C, I and O on the remaining datasets, resulting in a total of 12 different scenario combinations. Since the model is trained on one dataset only, domain partition is based on capturing devices(C,M), environments(I), and sessions(O) for different datasets. We also test our method with and without supplementary CelebA-Sproof. CelebA-Sproof is regarded as a domain if it is involved. As illustrated in Table 2, without help of CelebA-Sproof, BUDoPT performs better than SOTA in nine out of twelve settings. We also outperform domain adap", + "bbox": [ + 212, + 719, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S.Q. Liu et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/f68470aed5e821ad9dbed19fa7d16356d9567a83bb1ad998cd0941639254541e.jpg", + "table_caption": [ + "Table 2: Evaluation of cross-domain performance under one-to-one protocol (P2)" + ], + "table_footnote": [ + "$\\dagger$ indicates CelebA-Sproof is employed as supplementary training data." + ], + "table_body": "
MethodsUnseenC → IC → MC → OI → CI → MOM → CM → IM → OO → CO → IO → MAvg.
ADDA [50]No41.836.6-49.835.1-39.035.2----39.6
DRCN [11]No44.427.4-48.942.0-28.936.8----38.1
DupGAN [14]No42.133.4-46.536.2-27.135.4----36.8
KSA [23]No39.315.1-12.333.3-9.134.9----24.0
DR-UDA [54]No15.69.028.734.229.03.516.83.030.219.525.427.423.1
ADA [52]No17.59.329.141.530.539.617.75.131.219.826.831.525.0
USDAN-Un [21]No16.09.2-30.225.8-13.33.4----16.3
GDA [71]No15.105.8-29.720.8-12.22.5----14.4
CDFTN-L [65]No1.78.129.911.99.629.98.81.325.619.15.86.313.2
BUDoPT(ours)Yes9.654.376.768.265.647.736.072.96.164.812.972.945.69
FLIP-V†[49]Yes15.0813.7312.344.309.687.870.563.964.792.095.016.007.12
FLIP-IT†[49]Yes12.3315.187.981.128.376.980.195.214.960.164.275.636.03
FLIP-MCL†[49]Yes10.577.153.910.687.224.220.195.883.950.195.698.404.84
BUDoPT(ours)†Yes4.332.624.980.481.834.1402.451.870.442.531.432.46
", + "bbox": [ + 215, + 181, + 785, + 330 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "tive models (Unseen=No) even without any information of target domain. With CelebA-Sproof, BUDoPT surpasses FLIP family models in ten settings.", + "bbox": [ + 212, + 398, + 784, + 429 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Effect of Domain Partition. We set $\\mathrm{MI} \\rightarrow \\mathrm{O}$ and $\\mathrm{MI} \\rightarrow \\mathrm{C}$ to evaluate our model with different amounts of training data sources. Here we treat dataset as domain. Results in Table 3 show that with less training data, BUDoPT outperforms competitors by a larger margin. However, we also note that comparing with the $\\mathrm{M} \\rightarrow \\mathrm{O}$ results in P2 the performance increase is limited. We believe this is due to the difference in domain partition strategies. For $\\mathrm{M} \\rightarrow \\mathrm{O}$ the domain is defined based on the device while for $\\mathrm{MI} \\rightarrow \\mathrm{O}$ it is based on dataset. The results indicate that a fine-grained partition strategy may help achieve better performance for our method even with a limited data scale. More analysis of partition strategy is in supplementary Sec 2.1.", + "bbox": [ + 212, + 445, + 787, + 595 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Generalize to Unseen attack type. We employ $\\mathbf{W}$ with high-quality 3D silicon mask attack samples to evaluate the model generalizability to unseen attack types (P4). Different from the baselines trained on IMCO, we only utilize $\\mathbf{C}$ as the source dataset. Same domain partition strategy in P2 is adopted. As shown in Table 4, with less training data BUDoPT still outperforms the others. This result not only demonstrates our generalizability to the unseen attack types but also shows the learning efficiency of our method.", + "bbox": [ + 212, + 612, + 787, + 719 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Device Variance Only. In P5, we utilize the O's intra protocol 3 and protocol 4 [2] to evaluate the performance with device variance only. The domain gaps in P5 are located on the unseen device settings and parameters and other underlying factors are similar. Same domain partition strategy in P2 is adopted. As shown in Table 5, under constraint setting BUDoPT can handle the domain gap of devices well. This indicates the usability in real applications where some local samples can be available for fine-tuning.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing", + "bbox": [ + 253, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/acf5ac2bd119ed2b0e4456b6169b258d48e0bd8456656f1d0e1f4b2c69791e08.jpg", + "table_caption": [ + "Table 3: Evaluation of different domain partition (P3)" + ], + "table_footnote": [], + "table_body": "
MethodsMI → CMI → O
HTERAUCHTERAUC
SSAN-R [57]25.5683.8924.4482.86
HFN+MP [6]30.8972.4820.9486.71
CIFAS [35]22.6783.8924.6381.48
DiVT-M [24]20.1186.7123.6185.73
SSDG-R [20]19.8686.4627.9278.72
DGUA-FAS [13]19.2286.8120.0588.75
BUDoPT(ours)5.3398.925.9498.37
", + "bbox": [ + 215, + 184, + 475, + 296 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/8167bbbea841be6070670685a00a75d7d6a391bf94f9739635433567664bc1b7.jpg", + "table_caption": [ + "Table 4: Evaluation of unseen attack type performance (P4)" + ], + "table_footnote": [], + "table_body": "
MethodsIMCO → WMCA
HTERAUC
DiVT-M [24]22.3686.82
DGUA-FAS [13]20.6288.07
C → WMCA
BUDoPT(ours)9.2095.93
", + "bbox": [ + 535, + 184, + 777, + 273 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/39a98a64a92a85b1326296e46b760e2517fee4432959f0f513d9cce7c6be3d81.jpg", + "table_caption": [ + "Table 5: Evaluation of Device Variance Only (P5) on O. P refers protocol" + ], + "table_footnote": [], + "table_body": "
MethodsPACPERBCPERACER
DC-CDN [63]32.2±2.81.6±2.11.9±1.1
RAEDFL [17]31.38±1.780.28±0.680.83±0.86
TranFAS [56]30.6±0.71.1±2.50.9±1.1
LDCFormer [16]32.35±2.050.28±0.681.31±1.03
BUDoPT(ours)30±00±00±0
DC-CDN [63]45.4±3.32.5±4.24.0±3.1
RAEDFL [17]45.41±6.402.50±2.743.96±3.90
TranFAS [56]42.1±2.23.8±3.52.9±2.4
LDCFormer [16]41.08±1.281.17±1.941.13±1.02
BUDoPT(ours)40±00±00±0
", + "bbox": [ + 215, + 383, + 475, + 503 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/720cb935fe98a72f8b5e1cc04745018e9f24ac591c6ad9b155450f81bbf60022.jpg", + "table_caption": [ + "Table 6: Ablation study of major components of the proposed BUDoPT" + ], + "table_footnote": [], + "table_body": "
SDSDDRDGADCDCICM → O
✓----3.51
-✓---5.78
✓-✓--3.01
✓-✓✓-2.50
✓-✓✓✓2.26
", + "bbox": [ + 532, + 383, + 779, + 470 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 558, + 401, + 571 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We analyze the effectiveness of major components(Table 6), test the intuitive textual recording domain prompt tuning(Table 6), and compared with generic visual prompt tuning CoOp [69], CoCoOp [68](Table 1).", + "bbox": [ + 212, + 595, + 785, + 641 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 6 reports the results of the ablation studies of our major components. We set Spoofing Domain Discriminative Prompt Tuning(SD) as the baseline and evaluate performance with and without Recording Domain Generalization Prompt Learning (RDG), Adversarial Domain Learning(AD) and Context-aware Domain Clue(CDC). With Recording Domain Generalization Prompt Learning, the HTER is improved by $14\\%$ . With Adversarial Domain Learning, the HTER decreases by around $17\\%$ . The Context-aware Domain Clue contributes around $10\\%$ improvement. The ablation studies indicate that each component within our method is crucial for improving the domain generalization ability.", + "bbox": [ + 212, + 643, + 787, + 779 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Intuitive textual recording domain prompt tuning. Considering the success of employing context information as a prompt in upper-stream tasks, we evaluate an intuitive domain prompt tuning strategy, i.e., using the textual domain description as the prompt. We mark the textual domain prompts with the", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S.Q. Liu et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f81e81fb30622b2a61c02f1227aee00a55c968ebf485b4387670203becc4f325.jpg", + "image_caption": [ + "Fig. 4: The t-SNE visualizations of original visual embedding (first row) and the learned domain-invariant prompt embedding (second row) under P1." + ], + "image_footnote": [], + "bbox": [ + 220, + 146, + 784, + 318 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Spoofing Domain Discriminative Prompt Tuning as SDD2. As shown in Table 6, adding the intuitive textual recording domain prompt (SDD) harms the model's performance. This aligns with our analysis that the textual domain prompt of context encourages the model to extract more domain-specific information and contradicts the goal of domain-generalized FAS.", + "bbox": [ + 212, + 388, + 787, + 463 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Generic visual prompt tuning. Since the proposed method follows the idea of visual prompt tuning, the classical visual prompt learning methods: CoOp and CoCoOp are compared. For a fair comparison, the CLIP visual encoder of the two models is finetuned with the Spoofing Domain Discriminative Prompt Learning module in our framework so we regard them as BUDoPT-CoOp and BUDoPT-CoCoOp. As the results shown in Table 1, by comparing with the FLIP-family models, BUDoPT-CoOp, and BUDoPT-CoCoOp, we comprehensively verify the effectiveness of our proposed method that adapts CLIP to FAS with better generalization ability.", + "bbox": [ + 212, + 464, + 787, + 599 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We also conduct hyperparameter analysis of loss function. Results are summarized in Eq. (11) in the supplementary Section 2.2.", + "bbox": [ + 212, + 599, + 785, + 631 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Visualizations", + "text_level": 1, + "bbox": [ + 214, + 652, + 377, + 667 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Domain Prompt Visualization. Figure 4 shows the t-SNE [36] visualization of original visual embedding (first row) and the learned domain-invariant prompt embedding (second row) under P1. The original visual embedding naturally forms multiple clusters of domain-specific information. Each cluster shares some low-level underlying factors, such as the camera qualities (see supplementary material for details). Our BUDoPT learns a domain-generalized prompt embedding where samples from different domains are grouped together in a single cluster, thereby guiding the model to minimize the influences of domain gaps. Attention Visualization. We utilize the GradCam [46] to visualize attention", + "bbox": [ + 212, + 679, + 787, + 816 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing", + "bbox": [ + 253, + 114, + 732, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_footnote", + "text": "2 Please refer to supplementary material for details", + "bbox": [ + 217, + 824, + 563, + 839 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ea9e1286e03706a17b62f75e42ccd2ae27fa58fe1e7e780a2492e397a300f9ae.jpg", + "image_caption": [ + "Fig. 5: Typical attention maps of real and fake samples under P1." + ], + "image_footnote": [], + "bbox": [ + 218, + 141, + 362, + 256 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/13948c45e91ce012c3b2fc4aae866a1b3f37457b3e5801b836cde11a51cc8343.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 142, + 504, + 257 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/46d5899b36ace9e81d326b04129d87f6a4f6c228bf197a47658c5e88ad66c5d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 513, + 142, + 645, + 256 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/7ec57ffabc1d59b0bb99eb0b60c9e903e1c609f9b2da4654cc4bbb9a5de02cda.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 142, + 785, + 257 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "maps for genuine and spoofing samples in P1. Typical results are visualized in Figure 5. We can observe that BUDoPT takes different attention for live and spoof faces. Specifically, the model focuses on face regions for real samples while highlights spoof clues of fake faces. With the help of Spoofing Domain Discriminative Prompt Tuning, the model is able to detect the intrinsic spoofing clues for different types of attack, including moiré patterns(M, O), paper mask edges(C) and paper texture(I).", + "bbox": [ + 212, + 311, + 787, + 417 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 446, + 359, + 463 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we aim to improve the generalization ability of CLIP-based FAS from a prompt learning perspective. A Bottom-Up Domain Prompt Tuning method (BUDoPT) that covers the different levels of domain variance, including the domain of recording settings and domain of attack types is proposed. It contains a context-aware adversarial domain-generalized prompt learning strategy that can learn domain-invariant prompt for different recording settings and a fine-grained textual prompt that guides CLIP to look through the subtle details of different attack instruments. Extensive experiments are conducted on five FAS datasets with a large number of variations (camera types, resolutions, image qualities, lighting conditions, and recording environments). The effectiveness of our proposed method is evaluated with different amounts of source domains from multiple angles, where we boost the generalizability compared with the state of the arts with multiple training datasets or with only one dataset. We also observe the limitations of our methods with different domain partitioning strategies. Designing an adaptive domain partition approach is a potential way to further improve the generalizability of our method.", + "bbox": [ + 212, + 484, + 787, + 727 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Acknowledgement", + "text_level": 1, + "bbox": [ + 215, + 756, + 426, + 773 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work is supported by GuangDong Basic and Applied Basic Research Foundation 2023A1515110706, RGC-NSFC joint project N_HKBU212/23, and NSFC Young Scientist Fund Application No. 6240075068.", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S.Q. Liu et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bao, W., Li, H., Li, N., Jiang, W.: A liveness detection method for face recognition based on optical flow field. In: IASP (2009)", + "2. Boulkenafet, Z., Komulainen, J., Li, L., Feng, X., Hadid, A.: OULU-NPU: A mobile face presentation attack database with real-world variations. In: FG (2017)", + "3. Boulkenafet, Z., Komulainen, J., Hadid, A.: Face spoofing detection using colour texture analysis. IEEE Trans. on Information Forensics and Security 11(8), 1818-1830 (2016)", + "4. Bulat, A., Tzimiropoulos, G.: How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In: International Conference on Computer Vision (2017)", + "5. Cai, R., Cui, Y., Li, Z., Yu, Z., Li, H., Hu, Y., Kot, A.: Rehearsal-free domain continual face anti-spoofing: Generalize more and forget less. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8037-8048 (2023)", + "6. Cai, R., Li, Z., Wan, R., Li, H., Hu, Y., Kot, A.C.: Learning meta pattern for face anti-spoofing. IEEE Transactions on Information Forensics and Security 17, 1201-1213 (2022)", + "7. Chen, Z., Yao, T., Sheng, K., Ding, S., Tai, Y., Li, J., Huang, F., Jin, X.: Generalizable representation learning for mixture domain face anti-spoofing. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 1132-1139 (2021)", + "8. Chingovska, I., Anjos, A., Marcel, S.: On the effectiveness of local binary patterns in face anti-spoofing. In: BIOSIG. pp. 1-7 (2012)", + "9. Ge, H., Tu, X., Ai, W., Luo, Y., Ma, Z., Xie, M.: Face anti-spoofing by the enhancement of temporal motion. In: CTISC. pp. 106-111. IEEE (2020)", + "0. George, A., Mostaani, Z., Geissenbuhler, D., Nikisins, O., Anjos, A., Marcel, S.: Biometric face presentation attack detection with multi-channel convolutional neural network. IEEE Trans. on Information Forensics and Security (2019)", + "1. Ghifary, M., Kleijn, W.B., Zhang, M., Balduzzi, D., Li, W.: Deep reconstruction-classification networks for unsupervised domain adaptation. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14. pp. 597-613. Springer (2016)", + "2. Guo, J., Zhu, X., Xiao, J., Lei, Z., Wan, G., Li, S.Z.: Improving face anti-spoofing by 3d virtual synthesis. arXiv (2019)", + "3. Hong, Z.W., Lin, Y.C., Liu, H.T., Yeh, Y.R., Chen, C.S.: Domain-generalized face anti-spoofing with unknown attacks. In: 2023 IEEE International Conference on Image Processing (ICIP). pp. 820-824. IEEE (2023)", + "4. Hu, L., Kan, M., Shan, S., Chen, X.: Duplex generative adversarial network for unsupervised domain adaptation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1498-1507 (2018)", + "5. Huang, H.P., Sun, D., Liu, Y., Chu, W.S., Xiao, T., Yuan, J., Adam, H., Yang, M.H.: Adaptive transformers for robust few-shot cross-domain face anti-spoofing. In: European Conference on Computer Vision. pp. 37-54. Springer (2022)", + "6. Huang, P.K., Chiang, C.H., Chong, J.X., Chen, T.H., Ni, H.Y., Hsu, C.T.: Ldcformer: Incorporating learnable descriptive convolution to vision transformer for face anti-spoofing. In: 2023 IEEE International Conference on Image Processing (ICIP). pp. 121-125. IEEE (2023)", + "7. Huang, P.K., Chin, M.C., Hsu, C.T.: Face anti-spoofing via robust auxiliary estimation and discriminative feature learning. In: Asian Conference on Pattern Recognition. pp. 443-458. Springer (2021)" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing", + "bbox": [ + 253, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: Proceedings of the IEEE international conference on computer vision. pp. 1501-1510 (2017)", + "19. Jee, H.K., Jung, S.U., Yoo, J.H.: Liveness detection for embedded face recognition system. International Journal of Biological and Medical Sciences 1(4), 235-238 (2006)", + "20. Jia, Y., Zhang, J., Shan, S., Chen, X.: Single-side domain generalization for face anti-spoofing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8484-8493 (2020)", + "21. Jia, Y., Zhang, J., Shan, S., Chen, X.: Unified unsupervised and semi-supervised domain adaptation network for cross-scenario face anti-spoofing. Pattern Recognition 115, 107888 (2021)", + "22. Jourabloo, A., Liu, Y., Liu, X.: Face de-spoofing: Anti-spoofing via noise modeling. In: ECCV. pp. 290-306 (2018)", + "23. Li, H., Li, W., Cao, H., Wang, S., Huang, F., Kot, A.C.: Unsupervised domain adaptation for face anti-spoofing. IEEE Transactions on Information Forensics and Security 13(7), 1794-1809 (2018)", + "24. Liao, C.H., Chen, W.C., Liu, H.T., Yeh, Y.R., Hu, M.C., Chen, C.S.: Domain invariant vision transformer learning for face anti-spoofing. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6098-6107 (2023)", + "25. Liu, A., Xue, S., Gan, J., Wan, J., Liang, Y., Deng, J., Escalera, S., Lei, Z.: Cfplfas: Class free prompt learning for generalizable face anti-spoofing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 222-232 (2024)", + "26. Liu, S., Lu, S., Xu, H., Yang, J., Ding, S., Ma, L.: Feature generation and hypothesis verification for reliable face anti-spoofing. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 1782-1791 (2022)", + "27. Liu, S., Zhang, K.Y., Yao, T., Bi, M., Ding, S., Li, J., Huang, F., Ma, L.: Adaptive normalized representation learning for generalizable face anti-spoofing. In: Proceedings of the 29th ACM international conference on multimedia. pp. 1469-1477 (2021)", + "28. Liu, S., Zhang, K.Y., Yao, T., Sheng, K., Ding, S., Tai, Y., Li, J., Xie, Y., Ma, L.: Dual reweighting domain generalization for face presentation attack detection. arXiv preprint arXiv:2106.16128 (2021)", + "29. Liu, S.Q., Lan, X., Yuen, P.C.: Multi-channel remote photoplethysmography correspondence feature for 3d mask face presentation attack detection. IEEE Trans. on Information Forensics and Security 16, 2683-2696 (2021)", + "30. Liu, S.Q., Lan, X., Yuen, P.C.: Learning temporal similarity of remote photoplethysmography for fast 3d mask face presentation attack detection. IEEE Transactions on Information Forensics and Security 17, 3195-3210 (2022)", + "31. Liu, Y., Tai, Y., Li, J., Ding, S., Wang, C., Huang, F., Li, D., Qi, W., Ji, R.: Aurora guard: Real-time face anti-spoofing via light reflection. arXiv (2019)", + "32. Liu, Y., Jourabloo, A., Liu, X.: Learning deep models for face anti-spoofing: Binary or auxiliary supervision. In: CVPR. pp. 389-398 (2018)", + "33. Liu, Y., Liu, X.: Spoof trace disentanglement for generic face anti-spoofing. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(3), 3813-3830 (2022)", + "34. Liu, Y., Stehouwer, J., Jourabloo, A., Liu, X.: Deep tree learning for zero-shot face anti-spoofing. In: CVPR (2019)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S.Q. Liu et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Liu, Y., Chen, Y., Dai, W., Li, C., Zou, J., Xiong, H.: Causal intervention for generalizable face anti-spoofing. In: 2022 IEEE International Conference on Multimedia and Expo (ICME). pp. 01-06. IEEE (2022)", + "36. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008)", + "37. Määttra, J., Hadid, A., Pietikainen, M.: Face spoofing detection from single images using micro-texture analysis. In: IJCB. pp. 1-7 (2011)", + "38. Menotti, D., Chiachia, G., Pinto, A., Schwartz, W.R., Pedrini, H., Falcao, A.X., Rocha, A.: Deep representations for iris, face, and fingerprint spoofing detection. IEEE Trans. on Information Forensics and Security 10(4), 864-879 (2015)", + "39. Muhammad, U., Holmberg, T., de Melo, W.C., Hadid, A.: Face anti-spoofing via sample learning based recurrent neural network (rnn). In: BMVC. p. 113 (2019)", + "40. Pan, G., Sun, L., Wu, Z., Lao, S.: Eyeblink-based anti-spoofing in face recognition from a generic webcam camera. In: ICCV. pp. 1-8 (2007)", + "41. Patel, K., Han, H., Jain, A.K.: Secure face unlock: Spoof detection on smartphones. IEEE transactions on information forensics and security 11(10), 2268-2283 (2016)", + "42. Patel, K., Han, H., Jain, A.K., Ott, G.: Live face video vs. spoof face video: Use of moiré patterns to detect replay video attacks. In: ICB. pp. 98-105 (2015)", + "43. Petroni, F., Rocktäschel, T., Lewis, P., Bakhtin, A., Wu, Y., Miller, A.H., Riedel, S.: Language models as knowledge bases? arXiv preprint arXiv:1909.01066 (2019)", + "44. Qin, Y., Yu, Z., Yan, L., Wang, Z., Zhao, C., Lei, Z.: Meta-teacher for face antispoofing. IEEE transactions on pattern analysis and machine intelligence 44(10), 6311-6326 (2021)", + "45. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "46. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)", + "47. Shao, R., Lan, X., Li, J., Yuen, P.C.: Multi-adversarial discriminative deep domain generalization for face presentation attack detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10023-10031 (2019)", + "48. Shao, R., Lan, X., Yuen, P.C.: Regularized fine-grained meta face anti-spoofing. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 11974-11981 (2020)", + "49. Srivatsan, K., Naseer, M., Nandakumar, K.: Flip: Cross-domain face anti-spoofing with language guidance. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19685-19696 (2023)", + "50. Tzeng, E., Hoffman, J., Saenko, K., Darrell, T.: Adversarial discriminative domain adaptation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7167-7176 (2017)", + "51. Wang, C.Y., Lu, Y.D., Yang, S.T., Lai, S.H.: Patchnet: A simple face anti-spoofing framework via fine-grained patch recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20281-20290 (2022)", + "52. Wang, G., Han, H., Shan, S., Chen, X.: Improving cross-database face presentation attack detection via adversarial domain adaptation. In: 2019 International Conference on Biometrics (ICB). pp. 1-8. IEEE (2019)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing", + "bbox": [ + 253, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "53. Wang, G., Han, H., Shan, S., Chen, X.: Cross-domain face presentation attack detection via multi-domain disentangled representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6678-6687 (2020)", + "54. Wang, G., Han, H., Shan, S., Chen, X.: Unsupervised adversarial domain adaptation for cross-domain face presentation attack detection. IEEE Transactions on Information Forensics and Security 16, 56-69 (2020)", + "55. Wang, J., Zhang, J., Bian, Y., Cai, Y., Wang, C., Pu, S.: Self-domain adaptation for face anti-spoofing. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 2746-2754 (2021)", + "56. Wang, Z., Wang, Q., Deng, W., Guo, G.: Face anti-spoofing using transformers with relation-aware mechanism. IEEE Transactions on Biometrics, Behavior, and Identity Science 4(3), 439-450 (2022)", + "57. Wang, Z., Wang, Z., Yu, Z., Deng, W., Li, J., Gao, T., Wang, Z.: Domain generalization via shuffled style assembly for face anti-spoofing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4123-4133 (2022)", + "58. Wen, D., Han, H., Jain, A.K.: Face spoof detection with image distortion analysis. IEEE Trans. on Information Forensics and Security 10(4), 746-761 (2015)", + "59. Yang, J., Lei, Z., Li, S.Z.: Learn convolutional neural network for face anti-spoofing. arXiv preprint arXiv:1408.5601 (2014)", + "60. Yang, J., Lei, Z., Liao, S., Li, S.Z.: Face liveness detection with component dependent descriptor. In: 2013 International Conference on Biometrics (ICB). pp. 1-6. IEEE (2013)", + "61. Yu, Z., Li, X., Niu, X., Shi, J., Zhao, G.: Face anti-spoofing with human material perception. In: ECCV. pp. 557-575. Springer (2020)", + "62. Yu, Z., Qin, Y., Li, X., Zhao, C., Lei, Z., Zhao, G.: Deep learning for face antispoofing: A survey. IEEE transactions on pattern analysis and machine intelligence 45(5), 5609-5631 (2022)", + "63. Yu, Z., Qin, Y., Zhao, H., Li, X., Zhao, G.: Dual-cross central difference network for face anti-spoofing. arXiv preprint arXiv:2105.01290 (2021)", + "64. Yu, Z., Wan, J., Qin, Y., Li, X., Li, S.Z., Zhao, G.: Nas-fas: Static-dynamic central difference network search for face anti-spoofing. IEEE Trans. on Pattern Analysis and Machine Intelligence (2020)", + "65. Yue, H., Wang, K., Zhang, G., Feng, H., Han, J., Ding, E., Wang, J.: Cyclically disentangled feature translation for face anti-spoofing. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 3358-3366 (2023)", + "66. Zhang, Y., Yin, Z., Li, Y., Yin, G., Yan, J., Shao, J., Liu, Z.: Celeba-spoof: Large-scale face anti-spoofing dataset with rich annotations. In: European Conference on Computer Vision (ECCV) (2020)", + "67. Zhang, Z., Yan, J., Liu, S., Lei, Z., Yi, D., Li, S.Z.: A face antispoofing database with diverse attacks. In: ICB (2012)", + "68. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022)", + "69. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)", + "70. Zhou, Q., Zhang, K.Y., Yao, T., Lu, X., Ding, S., Ma, L.: Test-time domain generalization for face anti-spoofing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 175-187 (2024)" + ], + "bbox": [ + 215, + 147, + 785, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "S.Q. Liu et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 17 + }, + { + "type": "ref_text", + "text": "71. Zhou, Q., Zhang, K.Y., Yao, T., Yi, R., Sheng, K., Ding, S., Ma, L.: Generative domain adaptation for face anti-spoofing. In: European Conference on Computer Vision. pp. 335-356. Springer (2022)", + "bbox": [ + 217, + 146, + 787, + 189 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing", + "bbox": [ + 253, + 114, + 730, + 130 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 127 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing/3e12da5b-7b20-4ba8-b65d-4a8790f34fde_model.json b/2024/Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing/3e12da5b-7b20-4ba8-b65d-4a8790f34fde_model.json new file mode 100644 index 0000000000000000000000000000000000000000..429f35b49f8d7caca8d5e9d28a16ae472cd3749b --- /dev/null +++ b/2024/Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing/3e12da5b-7b20-4ba8-b65d-4a8790f34fde_model.json @@ -0,0 +1,2636 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.276, + 0.141, + 0.728, + 0.187 + ], + "angle": 0, + "content": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.212, + 0.68, + 0.228 + ], + "angle": 0, + "content": "Si-Qi Liu\\(^{1,\\ast,\\boxtimes}\\), Qirui Wang\\(^{1,\\ast}\\), and Pong C. Yuen\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.358, + 0.24, + 0.646, + 0.254 + ], + "angle": 0, + "content": "1 Shenzhen Research Institute of Big Data" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.255, + 0.747, + 0.282 + ], + "angle": 0, + "content": "\\(^{2}\\) Department of Computer Science, Hong Kong Baptist University siqiliu@sribd.cn, qwangcw@connect.ust.hk, pcyuen@comp.hkbu.edu.hk" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.312, + 0.744, + 0.59 + ], + "angle": 0, + "content": "Abstract. Face anti-spoofing (FAS) which plays an important role in securing face recognition systems has been attracting increasing attention. Recently, vision-language model CLIP has been proven to be effective for FAS, where outstanding performance can be achieved by simply transferring the class label into textual prompt. In this work, we aim to improve the generalization ability of CLIP-based FAS from a prompt learning perspective. Specifically, a Bottom-Up Domain Prompt Tuning method (BUDoPT) that covers the different levels of domain variance, including the domain of recording settings and domain of attack types is proposed. To handle domain discrepancies of recording settings, we design a context-aware adversarial domain-generalized prompt learning strategy that can learn domain-invariant prompt. For spoofing domain with different attack types, we construct a fine-grained textual prompt that guides CLIP to look through the subtle details of different attack instruments. Extensive experiments are conducted on five FAS datasets with variations of camera types, resolutions, image qualities, lighting conditions, and recording environments. The effectiveness of our proposed method is evaluated with different amounts of source domains from multiple angles, where we boost the generalizability compared with the state of the arts with multiple or limited numbers of training datasets." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.602, + 0.741, + 0.629 + ], + "angle": 0, + "content": "Keywords: Face Anti-spoofing \\(\\cdot\\) Face presentation attack detection \\(\\cdot\\) Face liveness detection" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.653, + 0.377, + 0.669 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.682, + 0.788, + 0.818 + ], + "angle": 0, + "content": "With the wide deployment of face recognition in real-world scenarios, the security concerns of FR systems attract increasing attention. Face images can be easily acquired from social networks, and the costs of conducting a face spoofing attack can be very low. Print or screens of mobile phones or tablets are the two common low-cost spoofing instruments. 3D mask attack could be more challenging since both the 3D shape and skin texture can be imitated. To counter these presentation attacks, a series of face anti-spoofing (FAS) methods based on different liveness cues are proposed, various from the hand-crafted features [1,19,40] to the deep-learning-based representations [34,62,64]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.544, + 0.84 + ], + "angle": 0, + "content": "* Equal contribution, ⌒ Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "S.Q. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.222 + ], + "angle": 0, + "content": "Early works are mainly developed for intra-domain testing where the recording environment, devices, and attack types are identical. Since the main difference between genuine and attack samples is located on small vestiges such as the texture and color tone, the subtle variations introduced by recording cameras, lighting conditions, and attack mediums form the domain gaps in FAS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.227, + 0.785, + 0.409 + ], + "angle": 0, + "content": "To improve the robustness and generalizability, the generic domain adaptation and domain generalization approaches are applied or customized in the context of FAS [52,53,57]. Few-shot domain adaptation and unsupervised domain adaptation methods tend to have better results since they can learn from the testing data with or without the labels [21,23,54]. The zero-shot domain generalization setting is more practical as Since the testing data may not be accessible in real scenarios. Popular techniques such as adversarial learning, contrastive learning, or meta-learning are adopted to align representations from different domains and keep on boosting the performances [5,44]. Directly constructing generalized liveness cues can also be effective. The facial heartbeat signal (rPPG) based methods exhibit strong generalizability to different types of attacks but existing methods still require long observation time [29,30]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.787, + 0.639 + ], + "angle": 0, + "content": "Recently, vision-language models like CLIP exhibited astonishing generalizability and zero-shot performance with the joint embedding of image-text pairs. The first trial of CLIP for FAS (FLIP) that transforms binary class labels into text and measures their similarity to image embeddings also achieves high performance [49]. However, since CLIP is trained for generic tasks with image pairs from the internet, the domain discrepancy of lighting conditions, camera qualities, or attack types in FAS can hardly be well addressed by simply finetuning it with the image-to-class-text contrastive loss. Intuitively, we may enrich the class text prompt with domain labels like the camera types, lighting conditions, or recording scene and construct hand-crafted prompts such as {a photo of [CLS] captured under [Recording Setting]} to learn a fine-grained representation. However, this approach may not work since 1) the detailed domain information in FAS can hardly be defined with text since it is related to multiple factors; 2) learning fine-grained representation that can tell domain discrepancies may contradict the final classification of genuine and spoofing samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In this work, we aim to improve the generalization ability of CLIP-based FAS from a prompt learning perspective. Specifically, we proposed to learn a domain-invariant prompt from the visual embedding that can guide CLIP to avoid interference from the specific domain. Inspired by the CLIP style contrastive learning objective which pulls images to their textual descriptions while pushes away unmatched pairs, we achieve domain-generalized prompt learning by pulling together image embeddings with a domain-invariant prompt and pushing away the image embeddings with their domain-specific prompt. Considering the discrepancies of recording domain are low-level features that scatter in the entire image, we extract domain-specific information from both the facial content and background context regions as the domain-specific negative prompt (push away). The domain-generalized prompt embedding extract from multiple domains is regarded as the positive prompt (pull together). An adversarial" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.254, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.145, + 0.446, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.227, + 0.456, + 0.237 + ], + "angle": 0, + "content": "(a) Intuitive solution of CLIP for Face Anti-Spoofing" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.144, + 0.772, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.486, + 0.227, + 0.768, + 0.238 + ], + "angle": 0, + "content": "(b) Bottom-Up Domain-Generalized Prompt Tuning for FAS (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.25, + 0.788, + 0.334 + ], + "angle": 0, + "content": "Fig. 1: The comparison between (a) the intuitive solution of CLIP for face anti-spoofing and (b) the proposed bottom-up domain-generalized prompt tuning. The Domain Token prompt guides the model to learn domain-generalized FAS features for different recording domains. The Attack Type prompt boosts the discriminability by encouraging the model to differentiate fine-grained attack instruments. The Domain Token is learned from the visual embedding through the Domain Prompt Learner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.366, + 0.788, + 0.534 + ], + "angle": 0, + "content": "learning strategy is designed to further ensure the prompt learner can learn the domain-generalized information. The various attack instruments for spoofing samples can be regarded as another perspective of domains. Different from the treatment of recording domain prompt learning, we want the visual encoder to distinguish the attack types so that it can learn more discriminative FAS features [20, 24]. In this case, we form a textual prompt that guides the generic visual encoder to adapt to the FAS task and look through the fine-grained difference for various attack types. We refer to this newly proposed method that learns the low-level domain-generalized visual prompt and tunes high-level attack-type textual prompt as the Bottom-Up Domain Prompt Tuning (BUDoPT) for generalized face anti-spoofing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.625 + ], + "angle": 0, + "content": "Extensive experiments are conducted on five FAS datasets with a large number of variations (camera types, resolutions, image qualities, lighting conditions, and recording environments). The effectiveness of our proposed method is evaluated with different amounts of source domains from multiple angles, where we boost the generalizability compared with the state of the arts with multiple training datasets or with only one dataset." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.627, + 0.59, + 0.641 + ], + "angle": 0, + "content": "In sum, the main contributions of this work are:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.656, + 0.787, + 0.732 + ], + "angle": 0, + "content": "- We improve the domain generalizability of FAS from a prompt tuning perspective within the context of vision-language model CLIP. A Bottom-Up Domain Prompt Tuning method (BUDoPT) that covers the different levels of domain variance, including the domain of recording settings and domain of attack types is proposed." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.733, + 0.787, + 0.794 + ], + "angle": 0, + "content": "- We design 1) a context-aware adversarial domain-generalized prompt learning strategy that can learn domain-invariant prompt for different recording settings, 2) an effective textual prompt that guides CLIP to adapt to FAS task and look through the subtle details of different attack types." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- We improve the generalizability compared with the state of the arts with different scales of training domains, and give in-depth analysis with extensive experiments on five FAS datasets that cover typical variations in FAS." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.656, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "S.Q. Liu et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.388, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.36 + ], + "angle": 0, + "content": "Face Anti-Spoofing. Face anti-spoofing has attracted increasing attention in the last decades, developed from the early handcrafted liveness feature construction [37, 60] to the latest deep learning-based design [64]. Early works focus on searching the discriminative liveness cues such as the appearance artifacts [31, 42, 58], and the motion difference between genuine faces and static attack instruments. Accordingly, hand-crafted features, such as SIFT [41], HoG [60], and LBP [3] are used to characterize spoof patterns. More discriminative features can be learned with Deep learning-based [38, 59]. Auxiliary tasks, such as the regression of facial depth map [32], 3D structure [12], and material [61], encourage the network to learn fine-grained appearance details. There are also classification-based methods [34, 51, 59] and Generative models [22, 33, 57]. Spatiotemporal information can also be learned through CNN and LSTM framework [9, 39]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.508 + ], + "angle": 0, + "content": "Generalized FAS. Generalized FAS that can perform robustly on different domains are necessary for real-world application scenarios. Recent works proposed to learn the FAS model on multiple source domains and evaluated a target domain. Under the domain adaptation settings where the target domain is accessible, aligning the source and target feature distributions is the typical solution [21,52,53,65,71]. Adversarial learning strategies or the triplet loss encourage the model to learn shared feature for generalization [20,47,57]. Meta-learning and continue learning approaches are also proposed by [5,6,27,44,48] to simulate the domain shift during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.788, + 0.732 + ], + "angle": 0, + "content": "Prompt Learning. Prompt learning originated from the field of Natural Language Processing (NLP) motives to shape the output of pre-trained language models by providing contextual information or constraints, thereby enhancing their ability to adapt to downstream tasks [43]. Prompt engineering was extended from NLP to Visual-Language Models [45], such as CLIP which classifies images by matching them with class prompts. CoOp [69] is designed to turn the handcrafted textual prompt into a learnable visual prompt. CoCoOp [68] incorporates the image as the condition for prompts and improves the model's generalization capabilities. Within the visual-textual contrastive learning framework, generalizability can be improved with a smaller data scale. CFPL1 [25] and this work aims to improve the generalizability for FAS with prompt learning. Different from CFPL which regards style (in AdaIN [18]) as the domain-specific cue, in this work we learn the domain-invariant prompt from an intrinsic CLIP-style contrastive learning perspective." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.755, + 0.541, + 0.772 + ], + "angle": 0, + "content": "3 Preliminary on CLIP for FAS" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.786, + 0.787, + 0.817 + ], + "angle": 0, + "content": "CLIP [45] comprises an image encoder denoted \\( f_{v} \\) and a text encoder denoted \\( f_{t} \\). In the context of the FAS, the two class text prompt \\( p_{k} \\) encompasses the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.689, + 0.84 + ], + "angle": 0, + "content": "1 CFPL was published after this manuscript was submitted for review." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.254, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "descriptions of real and spoof faces, representing two classes. CLIP determines the similarity between the input image \\(\\pmb{x}\\) and \\(\\pmb{p}\\) instead of directly training the classification head. The model predicts the image's category \\(\\hat{y}\\) by selecting the class prompt with the highest response:" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.221, + 0.786, + 0.246 + ], + "angle": 0, + "content": "\\[\n\\hat {y} = \\underset {k \\in \\{\\mathrm {r}, \\mathrm {s} \\}} {\\operatorname {a r g m a x}} \\left\\langle f _ {v} (\\boldsymbol {x}), f _ {t} \\left(\\boldsymbol {p} _ {k}\\right) \\right\\rangle \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.257, + 0.785, + 0.287 + ], + "angle": 0, + "content": "where \\(\\langle ,\\rangle\\) denotes the cosine similarity. \\(p_r\\) and \\(p_s\\) represent the prompt with real and spoof face descriptions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.288, + 0.787, + 0.364 + ], + "angle": 0, + "content": "With such learning framework, pretrained CLIP exhibits great performance in FAS with the standard finetuning process [49]. However, since CLIP is trained for generic tasks with image pairs from the internet, when handling the low-level feature in cross-domain settings in FAS, the domain discrepancy of lighting conditions, camera qualities, or attack types can hardly be well addressed." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.387, + 0.424, + 0.403 + ], + "angle": 0, + "content": "4 Proposed method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.789, + 0.6 + ], + "angle": 0, + "content": "Overview. The Bottom-Up Domain Prompt Tuning method (BUDoPT) is proposed to improve the generalization ability of FAS in the context of the vision language model CLIP. On top of the intuitive CLIP-based face anti-spoofing which turns the class label into textual prompt, we incorporate the domain token for the domain discrepancies from different recording settings and the attack type token that describes the fine-grained types of attacks. First, the context-aware adversarial domain-generalized prompt learning is to train the domain prompt learner given the input of facial visual feature. A CLIP score style domain prompt alignment mechanism and an adversarial learning strategy are designed to learn a domain-generalized prompt. Second, the textual prompt of fine-grained attack types aims to boost discriminability by looking through the subtle details of different attack instruments." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.622, + 0.704, + 0.637 + ], + "angle": 0, + "content": "4.1 Context-aware Domain Generalized Prompt Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.647, + 0.788, + 0.767 + ], + "angle": 0, + "content": "Recording Domain Generalized Prompt Learning. Although text prompt can effectively assist in classifying concrete instances in upper-stream tasks, it encounters difficulties in describing the recording domain with various complex factors, including camera qualities, lighting conditions, and blurriness in FAS. Therefore we proposed the to learn domain-generalized prompt from the visual embedding. Inspired by [68], we employ vision domain features as conditional inputs to generate learnable prompts. The primary objective of the component is learning domain-invariant prompts to mitigate the impact of domain gaps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.768, + 0.788, + 0.813 + ], + "angle": 0, + "content": "Specifically, we propose a Prompt learner \\(PL\\) to obtain the domain prompt from the image \\(\\mathbf{x}\\). \\(PL\\) utilizes the features from \\(i\\)-th layer of \\(f_{v}\\) as the input. The domain feature \\(D\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.429, + 0.826, + 0.786, + 0.842 + ], + "angle": 0, + "content": "\\[\nD (\\boldsymbol {x}) = P L \\left(f _ {v, i} (\\boldsymbol {x})\\right), \\tag {2}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "S.Q. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.144, + 0.77, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.405, + 0.789, + 0.492 + ], + "angle": 0, + "content": "Fig. 2: The proposed Bottom-Up Domain Prompt Tuning method (BUDoPT) covers the different levels of domain variance for generalized FAS, including the domain of recording settings and domain of attack types. The context-aware adversarial domain-generalized prompt learning can learn domain-invariant prompt for different recording settings. The textual prompt of fine-grained attack types can boost the discriminability by looking through the subtle details of different attack instruments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.506, + 0.534 + ], + "angle": 0, + "content": "where \\(f_{v,i}\\) denotes \\(i\\)-layer features of \\(f_v\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.789, + 0.597 + ], + "angle": 0, + "content": "To accurately represent the domain feature while minimizing influences from the content feature of a single image, we utilize all images' domain features in a mini-batch to generate the domain prompt. Given the training mini-batch \\(\\{\\pmb{x}_n\\}_{n=1}^N\\) for domain \\(m\\), the domain prompt \\(\\pmb{p}_{D_m}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.606, + 0.787, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} _ {D _ {m}} = f _ {t} \\left(\\frac {1}{N} \\sum_ {n = 1} ^ {N} D \\left(\\boldsymbol {x} _ {n} ^ {m}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.657, + 0.731, + 0.672 + ], + "angle": 0, + "content": "where \\(N\\) is the mini-batch size and \\(x_{n}^{m}\\) denotes the \\(n\\)-th image from \\(m\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.672, + 0.789, + 0.749 + ], + "angle": 0, + "content": "Inspired by the CLIP style contrastive learning objective, we design a CLIP score-based domain generalization learning strategy that pulls together image embeddings with a domain-invariant prompt and pushes away the image embeddings with their domain-specific prompt. Given a batch of training data with \\(M\\) domains, the generalized domain prompt \\(\\pmb{p}_{GD}\\) can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.759, + 0.787, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} _ {G D} = f _ {t} \\left(\\frac {1}{N * M} \\sum_ {m = 1} ^ {M} \\sum_ {n = 1} ^ {N} D \\left(\\boldsymbol {x} _ {n} ^ {m}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "As the Figure 3 shows, only the labels belonging to the generalized domain can be marked as positive while other domain-specific prompts are always labeled" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.254, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.145, + 0.756, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.247, + 0.788, + 0.29 + ], + "angle": 0, + "content": "Fig. 3: Illustration of the CLIP score-based domain generalization learning strategy, where we pull together image embeddings with a domain-generalized prompt (▲) and push away the image embeddings with their domain-specific prompt" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.316, + 0.784, + 0.347 + ], + "angle": 0, + "content": "as negative. We denote \\(\\pmb{p}\\) as the set of prompts and then calculate the image cross-entropy loss \\(\\mathcal{L}_{ce}\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.301, + 0.354, + 0.786, + 0.397 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c e} = - \\sum_ {z = 1} ^ {2 \\times (M + 1)} y \\log \\left(\\frac {\\exp \\left(\\left\\langle f _ {v} (\\boldsymbol {x}) , f _ {t} \\left(\\boldsymbol {p} _ {z}\\right) \\right\\rangle\\right)}{\\sum_ {k = 1} ^ {2 \\times (M + 1)} \\exp \\left(\\left\\langle f _ {v} (\\boldsymbol {x}), f _ {t} \\left(\\boldsymbol {p} _ {k}\\right) \\right\\rangle\\right)}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.784, + 0.452 + ], + "angle": 0, + "content": "where \\(y\\) is the ground truth label, \\(k\\) and \\(z\\) denote the indices in \\(\\pmb{p}\\). Note that we have \\(2 \\times (M + 1)\\) classes in total, i.e., \\(M\\) domains plus \\(\\pmb{p}_{GD}\\) for each class since the prompt set \\(\\pmb{p} = \\{\\pmb{p}_{D_1}^{real}, \\dots, \\pmb{p}_{D_M}^{real}, \\pmb{p}_{GD}^{real}, \\pmb{p}_{D_1}^{fake}, \\dots, \\pmb{p}_{D_M}^{fake}, \\pmb{p}_{GD}^{fake}\\}\\)," + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.452, + 0.784, + 0.496 + ], + "angle": 0, + "content": "During inference, we assume that all samples in the batch are sourced from the same domain and possess the identical prompt. The model selects the class prompt with the highest similarity score as the final prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.788, + 0.655 + ], + "angle": 0, + "content": "Context-aware Domain Clue. To better learn the domain-generalized prompt, we try to extract more domain-specific information from the visual embedding as it is treated as the negative label to be pushed away. Considering the discrepancies of recording domain are low-level features that scatter in the entire image, we extract domain-specific information from both the facial content and background context regions. Specifically, we want to highlight the context information from the surrounding background regions with purer domain information without disturbance of facial content. Consequently, we classify the output patches of \\( f_{v} \\) into two types: context patches \\( c^{ctx} \\) and content patches \\( c^{ctn} \\). Then the Equation (2) can be rewritten as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.415, + 0.662, + 0.786, + 0.681 + ], + "angle": 0, + "content": "\\[\nD (x) = P L ([ c _ {i} ^ {c t x}, c _ {i} ^ {c t n} ]), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.788, + 0.792 + ], + "angle": 0, + "content": "where \\( c_{i}^{ctx} \\) denotes the context patch obtained from \\( f_{v,i} \\). To enhance the representation of domain features while reducing the influence of facial content, we extract context patches \\( c_{s}^{ctx} \\) from a shallower layer (\\( s < i \\)) of \\( f_{v} \\) and replace the original \\( c_{i}^{ctx} \\), as shown in Figure 2(green block). For the VIT visual encoder in CLIP, the context patches in shallower layers preserve low-level recording domain information as they have less interaction with content patches. Hence, the domain feature is formed as:" + }, + { + "type": "equation", + "bbox": [ + 0.415, + 0.8, + 0.786, + 0.818 + ], + "angle": 0, + "content": "\\[\nD (x) = P L \\left(\\left[ c _ {s} ^ {c t x}, c _ {i} ^ {c t n} \\right]\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.825, + 0.66, + 0.841 + ], + "angle": 0, + "content": "to emphasize the domain-specific information in the prompts." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "S.Q. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.236 + ], + "angle": 0, + "content": "Adversarial Domain Prompt Learning. We propose the Adversarial Domain Prompt Learning strategy to further ensure that the prompt learner can learn the domain-generalized information. A Discriminator \\( DC \\) is constructed to differentiate the domain features and determine their belonging domain. \\( PL \\), as the generator, is required to minimize the domain-specific trace in the output to fool \\( DC \\):" + }, + { + "type": "equation", + "bbox": [ + 0.337, + 0.247, + 0.786, + 0.288 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a d v} = - \\frac {1}{N * M} \\sum_ {m = 1} ^ {M} \\sum_ {n = 1} ^ {N} y _ {m} \\log \\left(D C \\left(D \\left(\\boldsymbol {x} _ {n} ^ {m}\\right)\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.297, + 0.789, + 0.405 + ], + "angle": 0, + "content": "\\(y_{m}\\) is the domain label. \\(\\mathcal{L}_{adv}\\) encourages discriminator \\(DC\\) to try its best to find the domain-specific vestige so that domain prompt learner can learn the domain-invariant information. Empirically, the adversarial learning strategy aggravates the unstable loss convergence. To ensure a more consistent learning direction, we employ a distribution constraint between domain features themselves. The Jensen-Shannon divergence is adopted to pull every two features from \\(M\\) domains together in a training batch:" + }, + { + "type": "equation", + "bbox": [ + 0.356, + 0.414, + 0.786, + 0.449 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d i s} = \\sum_ {j, t = 1, \\dots , N * M, j < t} \\operatorname {J S} \\left(D \\left(\\boldsymbol {x} _ {j}\\right), D \\left(\\boldsymbol {x} _ {t}\\right)\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.458, + 0.612, + 0.473 + ], + "angle": 0, + "content": "Here \\(N * M\\) is the total number of samples in a batch." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.495, + 0.667, + 0.511 + ], + "angle": 0, + "content": "4.2 Spoofing Domain Discriminative Prompt Tuning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.788, + 0.7 + ], + "angle": 0, + "content": "In contrast to the visual recording domain generalized prompt that aims to avoid bias, the attack types of spoofing domain can be utilized to boost the discriminability. Specific artifacts are highly correlated with the attack instruments, which can be regarded as spoofing clues, such as the saturation change in printing attacks, the moiré pattern in video replay attacks, and the edge defects of 3D masks. Different from the low-level recording domain prompt, the attack types can be clearly described with the (high-level) textual prompt. Regularize the visual embedding with finer grained attack types text helps the visual encoder learn more discriminative FAS features. This will also set clearer boundaries between genuine faces and spoofing attacks thereby enhancing model generalizability to unseen attack types. Therefore, we propose the Spoof Domain Prompt tuning strategy to encourage the model to differentiate finer-grained attack types." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.701, + 0.788, + 0.792 + ], + "angle": 0, + "content": "Specifically, we design a textual prompt template, a structured sentence that describes spoofing face photos with a placeholder of the attack types, such as \"a {fake} face photo {and it is the printing attack}\" Given \\( N * M \\) samples in a training batch, we generate the same amount of text prompts \\( \\boldsymbol{p}\\boldsymbol{a} \\). Spoofing Domain Discriminative Prompt Tuning updates the parameters of \\( f_{v} \\) with the text cross-entropy loss \\( \\mathcal{L}_t \\), which is calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.329, + 0.804, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t} = - \\sum_ {o = 1} ^ {N * M} o \\log \\left(\\frac {\\exp \\left(\\left\\langle f _ {v} (\\boldsymbol {x}) , f _ {t} \\left(\\boldsymbol {p} a _ {o}\\right) \\right\\rangle\\right)}{\\sum_ {w = 1} ^ {N * M} \\exp \\left(\\left\\langle f _ {v} (\\boldsymbol {x}) , f _ {t} \\left(\\boldsymbol {p} a _ {w}\\right) \\right\\rangle\\right)}\\right), \\tag {10}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.254, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.483, + 0.162 + ], + "angle": 0, + "content": "where \\(o\\) and \\(w\\) are the indices in \\(\\pmb{pa}\\)." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.164, + 0.606, + 0.178 + ], + "angle": 0, + "content": "The final loss function of our method is formed as:" + }, + { + "type": "equation", + "bbox": [ + 0.383, + 0.195, + 0.785, + 0.21 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\alpha \\mathcal {L} _ {c e} + \\beta \\mathcal {L} _ {t} - \\gamma \\mathcal {L} _ {a d v} + \\delta \\mathcal {L} _ {d i s}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.225, + 0.644, + 0.24 + ], + "angle": 0, + "content": "where \\(\\alpha, \\beta, \\gamma, \\delta\\) are hyperparameters that balance each loss." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.271, + 0.376, + 0.287 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.31, + 0.436, + 0.325 + ], + "angle": 0, + "content": "5.1 Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.462 + ], + "angle": 0, + "content": "Databases and Protocols. We evaluate our method with six widely used FAS databases: MSU-MFSD(M) [58], CASIA-FASD(C) [67], Idiap Replay attack (I) [8], OULU-NPU(O) [2], WMCA(W) [10] and CelebA-Sproof [66]. Five protocols that cover various domain generalization settings with different amounts of source domains are involved to comprehensively evaluate our method's performance. They are leave-one-out(P1) and one-to-one(P2) cross-domain generalization evaluation, different domain partition evaluation(P3), unseen attack type evaluation(P4) and various devices domain evaluation(P5)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.538 + ], + "angle": 0, + "content": "Metrics. Following [49], we evaluate the model performance using three metrics: Half Total Error Rate (HTER), Area Under the Receiver Operating Characteristic Curve (AUC) and True Positive Rate (TPR) at a False Positive Rate (FPR) \\(1\\%\\) (TPR@FPR=1%). HTER and AUC reflect the theoretical performance. TPR@FPR=1% evaluates the model performance in practice." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.788, + 0.644 + ], + "angle": 0, + "content": "Implementation Details. We utilize FAN [4] to obtain facial landmarks, aligning the images based on nose and eye landmarks before cropping them to a size of 224x224. We employ the pre-trained CLIP-ViT-B-16 [45] as the foundation model. An eleven-layer VIT following a two-layer MLP is adopted as the prompt learner \\(PL\\). We extract content patches from the 5th layer of \\(f_{v}\\) and context patches from the 1st layer. For adversarial learning an eight-layer 1D ResNet is adopted as the discriminator \\(DC\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Training strategy. We divide the training process into two stages for the Bottom-Up domain prompt tuning. In the first training stage, we conduct the Spoofing Domain Discriminative Prompt Tuning to boost the discriminability of \\( f_{v} \\). Adam optimizer is adopted with a learning rate of \\( 10^{-6} \\). In the second training stage, we freeze \\( f_{v} \\) and train the \\( PL \\) for the Recording Domain Generalization Prompt Learning. The optimizer for both \\( PL \\) and \\( DC \\) are Adam with learning rate \\( 10^{-3} \\) and \\( 3 * 10^{-4} \\). Before the adversarial learning, we first train at least one epoch for each domain source to warm up(exclude supplementary training data). Then, we take turns freezing the \\( PL \\) and \\( DC \\) for 50 iterations. For each stage we train 4000 iterations to ensure the largest dataset is traversed. The mini-batch size is \\( 4(N = 4) \\) for each domain. \\( M \\) varies according to the number of datasets used for training under different evaluation protocols. The hyperparameter \\( \\alpha \\), \\( \\beta \\), \\( \\gamma \\), and \\( \\delta \\) are set as 1, 1, 1, and 0.5, respectively." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "S.Q. Liu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.787, + 0.16 + ], + "angle": 0, + "content": "Table 1: Evaluation of cross-domain performance under leave-one-out protocol (P1)" + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.181, + 0.787, + 0.435 + ], + "angle": 0, + "content": "
MethodsOCI→MOMI→COCM→IICM→OAvg.
HTERAUCTPR@ FPR=1%HTERAUCTPR@ FPR=1%HTERAUCTPR@ FPR=1%HTERAUCTPR@ FPR=1%
MADDG [47]17.6988.06-24.5084.51-22.1984.99-27.9880.02-23.09
MDDR [53]17.0290.10-19.6887.43-20.8786.72-25.02581.47-20.64
NAS-FAS [64]16.8590.42-15.2192.64-11.6396.98-13.1694.18-14.21
RFMeta [48]13.8993.98-20.2788.16-17.3090.48-16.4591.16-16.97
D²AM [7]12.7095.6-20.9885.58-15.4391.22-15.2790.87-16.09
DRDG [28]12.4395.81-19.0588.79-15.5691.79-15.6391.75-15.66
Self-DA [55]15.4091.80-24.5084.40-15.6090.10-23.1084.30-19.65
ANRL [27]10.8396.75-17.8589.26-16.0391.04-15.6791.90-15.09
FGHV [26]9.1796.92-12.4793.47-16.2990.11-13.5893.55-12.87
SSDG-R [20]7.3897.17-10.4495.94-11.7196.59-15.6191.54-11.28
SSAB-R [57]6.6798.75-10.0096.67-8.8896.79-13.7293.65-9.80
PatchNet [51]7.1098.46-11.3394.58-13.4095.67-11.8295.07-10.90
GDA [71]9.2098.00-12.2093.00-10.0096.00-14.4092.60-11.45
LDCformer [16]6.4398.39-8.1196.67-8.5797.09-11.1795.58-8.57
TTDG-V [70]4.1698.48-7.5998.18-9.6298.18-1096.15-7.84
BUDoPT(ours)0.9599.7087.122.8598.0355.004.498.5486.462.2698.7855.252.62
ViT†[15]1.5899.6896.675.7098.9788.579.2597.1551.547.4798.4269.306.00
FLIP-V†[49]3.7999.3187.991.2799.7595.854.7198.8075.844.1598.7666.473.48
FLIP-IT†[49]5.2798.4279.330.4499.9899.862.9499.4284.623.6199.1584.763.06
FLIP-MCL†[49]4.9598.1174.670.5499.98100.004.2599.0784.622.3199.6392.283.01
BUDoPT-CoOp†1.9899.5480.631.9399.5888.063.1899.390.43.3398.9189.292.61
BUDoPT-CoCoOp†1.5199.4375.551.5599.5684.743.5399.292.533.5198.9887.982.53
BUDoPT(ours)†0.4099.9999.670.2699.9699.921.3899.6998.11.6099.5197.180.91
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.283, + 0.436, + 0.725, + 0.448 + ], + "angle": 0, + "content": "\\(\\dagger\\) indicates CelebA-Sproof [66] is employed as supplementary training dataset" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.5, + 0.352, + 0.515 + ], + "angle": 0, + "content": "5.2 Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.787, + 0.719 + ], + "angle": 0, + "content": "Cross-domain generalization evaluation. Following [49], the leave-one-out (P1) and one-to-one (P2) on M, C, I and O are conducted to evaluate cross-dataset performance. \\(\\mathrm{OCM} \\rightarrow \\mathrm{I}\\) refers to the model trains on O, C, M and tests on I. For a fair comparison with [49] and other models, we evaluate the performance of our method with and without the supplementary training data CelebA-Sproof [66]. For simplicity, we regard dataset as the domain partition in our proposed method. As shown in Table 1, without supplementary training data, the proposed BUDoPT outperforms SOTA in all settings. With supplementary training data, BUDoPT also surpasses the FLIP family model in HTER for all settings and achieves more than \\(69\\%\\) improvement on average. In terms of TPR@FPR \\(= 1\\%\\), our method performs better in three out of four settings. Although the performance slightly drops in TPR@FPR \\(= 1\\% (\\mathrm{C})\\) and AUC(C,O), the margin is relatively small as they already exceed \\(99\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In P2 of low-data regime [49] the model is trained separately on M, C, I and O on the remaining datasets, resulting in a total of 12 different scenario combinations. Since the model is trained on one dataset only, domain partition is based on capturing devices(C,M), environments(I), and sessions(O) for different datasets. We also test our method with and without supplementary CelebA-Sproof. CelebA-Sproof is regarded as a domain if it is involved. As illustrated in Table 2, without help of CelebA-Sproof, BUDoPT performs better than SOTA in nine out of twelve settings. We also outperform domain adap" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.254, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.224, + 0.145, + 0.777, + 0.16 + ], + "angle": 0, + "content": "Table 2: Evaluation of cross-domain performance under one-to-one protocol (P2)" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.182, + 0.787, + 0.332 + ], + "angle": 0, + "content": "
MethodsUnseenC → IC → MC → OI → CI → MOM → CM → IM → OO → CO → IO → MAvg.
ADDA [50]No41.836.6-49.835.1-39.035.2----39.6
DRCN [11]No44.427.4-48.942.0-28.936.8----38.1
DupGAN [14]No42.133.4-46.536.2-27.135.4----36.8
KSA [23]No39.315.1-12.333.3-9.134.9----24.0
DR-UDA [54]No15.69.028.734.229.03.516.83.030.219.525.427.423.1
ADA [52]No17.59.329.141.530.539.617.75.131.219.826.831.525.0
USDAN-Un [21]No16.09.2-30.225.8-13.33.4----16.3
GDA [71]No15.105.8-29.720.8-12.22.5----14.4
CDFTN-L [65]No1.78.129.911.99.629.98.81.325.619.15.86.313.2
BUDoPT(ours)Yes9.654.376.768.265.647.736.072.96.164.812.972.945.69
FLIP-V†[49]Yes15.0813.7312.344.309.687.870.563.964.792.095.016.007.12
FLIP-IT†[49]Yes12.3315.187.981.128.376.980.195.214.960.164.275.636.03
FLIP-MCL†[49]Yes10.577.153.910.687.224.220.195.883.950.195.698.404.84
BUDoPT(ours)†Yes4.332.624.980.481.834.1402.451.870.442.531.432.46
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.305, + 0.333, + 0.703, + 0.344 + ], + "angle": 0, + "content": "\\(\\dagger\\) indicates CelebA-Sproof is employed as supplementary training data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.785, + 0.43 + ], + "angle": 0, + "content": "tive models (Unseen=No) even without any information of target domain. With CelebA-Sproof, BUDoPT surpasses FLIP family models in ten settings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.788, + 0.597 + ], + "angle": 0, + "content": "Effect of Domain Partition. We set \\(\\mathrm{MI} \\rightarrow \\mathrm{O}\\) and \\(\\mathrm{MI} \\rightarrow \\mathrm{C}\\) to evaluate our model with different amounts of training data sources. Here we treat dataset as domain. Results in Table 3 show that with less training data, BUDoPT outperforms competitors by a larger margin. However, we also note that comparing with the \\(\\mathrm{M} \\rightarrow \\mathrm{O}\\) results in P2 the performance increase is limited. We believe this is due to the difference in domain partition strategies. For \\(\\mathrm{M} \\rightarrow \\mathrm{O}\\) the domain is defined based on the device while for \\(\\mathrm{MI} \\rightarrow \\mathrm{O}\\) it is based on dataset. The results indicate that a fine-grained partition strategy may help achieve better performance for our method even with a limited data scale. More analysis of partition strategy is in supplementary Sec 2.1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.72 + ], + "angle": 0, + "content": "Generalize to Unseen attack type. We employ \\(\\mathbf{W}\\) with high-quality 3D silicon mask attack samples to evaluate the model generalizability to unseen attack types (P4). Different from the baselines trained on IMCO, we only utilize \\(\\mathbf{C}\\) as the source dataset. Same domain partition strategy in P2 is adopted. As shown in Table 4, with less training data BUDoPT still outperforms the others. This result not only demonstrates our generalizability to the unseen attack types but also shows the learning efficiency of our method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Device Variance Only. In P5, we utilize the O's intra protocol 3 and protocol 4 [2] to evaluate the performance with device variance only. The domain gaps in P5 are located on the unseen device settings and parameters and other underlying factors are similar. Same domain partition strategy in P2 is adopted. As shown in Table 5, under constraint setting BUDoPT can handle the domain gap of devices well. This indicates the usability in real applications where some local samples can be available for fine-tuning." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "S.Q. Liu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.477, + 0.173 + ], + "angle": 0, + "content": "Table 3: Evaluation of different domain partition (P3)" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.185, + 0.477, + 0.297 + ], + "angle": 0, + "content": "
MethodsMI → CMI → O
HTERAUCHTERAUC
SSAN-R [57]25.5683.8924.4482.86
HFN+MP [6]30.8972.4820.9486.71
CIFAS [35]22.6783.8924.6381.48
DiVT-M [24]20.1186.7123.6185.73
SSDG-R [20]19.8686.4627.9278.72
DGUA-FAS [13]19.2286.8120.0588.75
BUDoPT(ours)5.3398.925.9498.37
" + }, + { + "type": "table_caption", + "bbox": [ + 0.527, + 0.145, + 0.786, + 0.173 + ], + "angle": 0, + "content": "Table 4: Evaluation of unseen attack type performance (P4)" + }, + { + "type": "table", + "bbox": [ + 0.537, + 0.185, + 0.779, + 0.275 + ], + "angle": 0, + "content": "
MethodsIMCO → WMCA
HTERAUC
DiVT-M [24]22.3686.82
DGUA-FAS [13]20.6288.07
C → WMCA
BUDoPT(ours)9.2095.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.344, + 0.476, + 0.372 + ], + "angle": 0, + "content": "Table 5: Evaluation of Device Variance Only (P5) on O. P refers protocol" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.384, + 0.476, + 0.504 + ], + "angle": 0, + "content": "
MethodsPACPERBCPERACER
DC-CDN [63]32.2±2.81.6±2.11.9±1.1
RAEDFL [17]31.38±1.780.28±0.680.83±0.86
TranFAS [56]30.6±0.71.1±2.50.9±1.1
LDCFormer [16]32.35±2.050.28±0.681.31±1.03
BUDoPT(ours)30±00±00±0
DC-CDN [63]45.4±3.32.5±4.24.0±3.1
RAEDFL [17]45.41±6.402.50±2.743.96±3.90
TranFAS [56]42.1±2.23.8±3.52.9±2.4
LDCFormer [16]41.08±1.281.17±1.941.13±1.02
BUDoPT(ours)40±00±00±0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.527, + 0.344, + 0.785, + 0.372 + ], + "angle": 0, + "content": "Table 6: Ablation study of major components of the proposed BUDoPT" + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.384, + 0.781, + 0.472 + ], + "angle": 0, + "content": "
SDSDDRDGADCDCICM → O
✓----3.51
-✓---5.78
✓-✓--3.01
✓-✓✓-2.50
✓-✓✓✓2.26
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.559, + 0.402, + 0.572 + ], + "angle": 0, + "content": "5.3 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.787, + 0.642 + ], + "angle": 0, + "content": "We analyze the effectiveness of major components(Table 6), test the intuitive textual recording domain prompt tuning(Table 6), and compared with generic visual prompt tuning CoOp [69], CoCoOp [68](Table 1)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Table 6 reports the results of the ablation studies of our major components. We set Spoofing Domain Discriminative Prompt Tuning(SD) as the baseline and evaluate performance with and without Recording Domain Generalization Prompt Learning (RDG), Adversarial Domain Learning(AD) and Context-aware Domain Clue(CDC). With Recording Domain Generalization Prompt Learning, the HTER is improved by \\(14\\%\\). With Adversarial Domain Learning, the HTER decreases by around \\(17\\%\\). The Context-aware Domain Clue contributes around \\(10\\%\\) improvement. The ablation studies indicate that each component within our method is crucial for improving the domain generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Intuitive textual recording domain prompt tuning. Considering the success of employing context information as a prompt in upper-stream tasks, we evaluate an intuitive domain prompt tuning strategy, i.e., using the textual domain description as the prompt. We mark the textual domain prompts with the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.254, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.147, + 0.785, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.361 + ], + "angle": 0, + "content": "Fig. 4: The t-SNE visualizations of original visual embedding (first row) and the learned domain-invariant prompt embedding (second row) under P1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.788, + 0.464 + ], + "angle": 0, + "content": "Spoofing Domain Discriminative Prompt Tuning as SDD2. As shown in Table 6, adding the intuitive textual recording domain prompt (SDD) harms the model's performance. This aligns with our analysis that the textual domain prompt of context encourages the model to extract more domain-specific information and contradicts the goal of domain-generalized FAS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.788, + 0.6 + ], + "angle": 0, + "content": "Generic visual prompt tuning. Since the proposed method follows the idea of visual prompt tuning, the classical visual prompt learning methods: CoOp and CoCoOp are compared. For a fair comparison, the CLIP visual encoder of the two models is finetuned with the Spoofing Domain Discriminative Prompt Learning module in our framework so we regard them as BUDoPT-CoOp and BUDoPT-CoCoOp. As the results shown in Table 1, by comparing with the FLIP-family models, BUDoPT-CoOp, and BUDoPT-CoCoOp, we comprehensively verify the effectiveness of our proposed method that adapts CLIP to FAS with better generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.787, + 0.632 + ], + "angle": 0, + "content": "We also conduct hyperparameter analysis of loss function. Results are summarized in Eq. (11) in the supplementary Section 2.2." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.654, + 0.379, + 0.669 + ], + "angle": 0, + "content": "5.4 Visualizations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.68, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Domain Prompt Visualization. Figure 4 shows the t-SNE [36] visualization of original visual embedding (first row) and the learned domain-invariant prompt embedding (second row) under P1. The original visual embedding naturally forms multiple clusters of domain-specific information. Each cluster shares some low-level underlying factors, such as the camera qualities (see supplementary material for details). Our BUDoPT learns a domain-generalized prompt embedding where samples from different domains are grouped together in a single cluster, thereby guiding the model to minimize the influences of domain gaps. Attention Visualization. We utilize the GradCam [46] to visualize attention" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.565, + 0.84 + ], + "angle": 0, + "content": "2 Please refer to supplementary material for details" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "S.Q. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.142, + 0.364, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.143, + 0.505, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.143, + 0.646, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.143, + 0.787, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.278, + 0.266, + 0.725, + 0.281 + ], + "angle": 0, + "content": "Fig. 5: Typical attention maps of real and fake samples under P1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.419 + ], + "angle": 0, + "content": "maps for genuine and spoofing samples in P1. Typical results are visualized in Figure 5. We can observe that BUDoPT takes different attention for live and spoof faces. Specifically, the model focuses on face regions for real samples while highlights spoof clues of fake faces. With the help of Spoofing Domain Discriminative Prompt Tuning, the model is able to detect the intrinsic spoofing clues for different types of attack, including moiré patterns(M, O), paper mask edges(C) and paper texture(I)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.448, + 0.36, + 0.464 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.789, + 0.728 + ], + "angle": 0, + "content": "In this work, we aim to improve the generalization ability of CLIP-based FAS from a prompt learning perspective. A Bottom-Up Domain Prompt Tuning method (BUDoPT) that covers the different levels of domain variance, including the domain of recording settings and domain of attack types is proposed. It contains a context-aware adversarial domain-generalized prompt learning strategy that can learn domain-invariant prompt for different recording settings and a fine-grained textual prompt that guides CLIP to look through the subtle details of different attack instruments. Extensive experiments are conducted on five FAS datasets with a large number of variations (camera types, resolutions, image qualities, lighting conditions, and recording environments). The effectiveness of our proposed method is evaluated with different amounts of source domains from multiple angles, where we boost the generalizability compared with the state of the arts with multiple training datasets or with only one dataset. We also observe the limitations of our methods with different domain partitioning strategies. Designing an adaptive domain partition approach is a potential way to further improve the generalizability of our method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.757, + 0.427, + 0.774 + ], + "angle": 0, + "content": "7 Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "This work is supported by GuangDong Basic and Applied Basic Research Foundation 2023A1515110706, RGC-NSFC joint project N_HKBU212/23, and NSFC Young Scientist Fund Application No. 6240075068." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.254, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. Bao, W., Li, H., Li, N., Jiang, W.: A liveness detection method for face recognition based on optical flow field. In: IASP (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.232 + ], + "angle": 0, + "content": "2. Boulkenafet, Z., Komulainen, J., Li, L., Feng, X., Hadid, A.: OULU-NPU: A mobile face presentation attack database with real-world variations. In: FG (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.233, + 0.785, + 0.272 + ], + "angle": 0, + "content": "3. Boulkenafet, Z., Komulainen, J., Hadid, A.: Face spoofing detection using colour texture analysis. IEEE Trans. on Information Forensics and Security 11(8), 1818-1830 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.785, + 0.314 + ], + "angle": 0, + "content": "4. Bulat, A., Tzimiropoulos, G.: How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In: International Conference on Computer Vision (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.315, + 0.785, + 0.357 + ], + "angle": 0, + "content": "5. Cai, R., Cui, Y., Li, Z., Yu, Z., Li, H., Hu, Y., Kot, A.: Rehearsal-free domain continual face anti-spoofing: Generalize more and forget less. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8037-8048 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.358, + 0.785, + 0.397 + ], + "angle": 0, + "content": "6. Cai, R., Li, Z., Wan, R., Li, H., Hu, Y., Kot, A.C.: Learning meta pattern for face anti-spoofing. IEEE Transactions on Information Forensics and Security 17, 1201-1213 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.398, + 0.785, + 0.44 + ], + "angle": 0, + "content": "7. Chen, Z., Yao, T., Sheng, K., Ding, S., Tai, Y., Li, J., Huang, F., Jin, X.: Generalizable representation learning for mixture domain face anti-spoofing. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 1132-1139 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.441, + 0.785, + 0.467 + ], + "angle": 0, + "content": "8. Chingovska, I., Anjos, A., Marcel, S.: On the effectiveness of local binary patterns in face anti-spoofing. In: BIOSIG. pp. 1-7 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.785, + 0.494 + ], + "angle": 0, + "content": "9. Ge, H., Tu, X., Ai, W., Luo, Y., Ma, Z., Xie, M.: Face anti-spoofing by the enhancement of temporal motion. In: CTISC. pp. 106-111. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.496, + 0.785, + 0.536 + ], + "angle": 0, + "content": "0. George, A., Mostaani, Z., Geissenbuhler, D., Nikisins, O., Anjos, A., Marcel, S.: Biometric face presentation attack detection with multi-channel convolutional neural network. IEEE Trans. on Information Forensics and Security (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.785, + 0.591 + ], + "angle": 0, + "content": "1. Ghifary, M., Kleijn, W.B., Zhang, M., Balduzzi, D., Li, W.: Deep reconstruction-classification networks for unsupervised domain adaptation. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14. pp. 597-613. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.593, + 0.785, + 0.619 + ], + "angle": 0, + "content": "2. Guo, J., Zhu, X., Xiao, J., Lei, Z., Wan, G., Li, S.Z.: Improving face anti-spoofing by 3d virtual synthesis. arXiv (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.785, + 0.66 + ], + "angle": 0, + "content": "3. Hong, Z.W., Lin, Y.C., Liu, H.T., Yeh, Y.R., Chen, C.S.: Domain-generalized face anti-spoofing with unknown attacks. In: 2023 IEEE International Conference on Image Processing (ICIP). pp. 820-824. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "4. Hu, L., Kan, M., Shan, S., Chen, X.: Duplex generative adversarial network for unsupervised domain adaptation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1498-1507 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.785, + 0.744 + ], + "angle": 0, + "content": "5. Huang, H.P., Sun, D., Liu, Y., Chu, W.S., Xiao, T., Yuan, J., Adam, H., Yang, M.H.: Adaptive transformers for robust few-shot cross-domain face anti-spoofing. In: European Conference on Computer Vision. pp. 37-54. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.799 + ], + "angle": 0, + "content": "6. Huang, P.K., Chiang, C.H., Chong, J.X., Chen, T.H., Ni, H.Y., Hsu, C.T.: Ldcformer: Incorporating learnable descriptive convolution to vision transformer for face anti-spoofing. In: 2023 IEEE International Conference on Image Processing (ICIP). pp. 121-125. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Huang, P.K., Chin, M.C., Hsu, C.T.: Face anti-spoofing via robust auxiliary estimation and discriminative feature learning. In: Asian Conference on Pattern Recognition. pp. 443-458. Springer (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "S.Q. Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "18. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: Proceedings of the IEEE international conference on computer vision. pp. 1501-1510 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "19. Jee, H.K., Jung, S.U., Yoo, J.H.: Liveness detection for embedded face recognition system. International Journal of Biological and Medical Sciences 1(4), 235-238 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.785, + 0.273 + ], + "angle": 0, + "content": "20. Jia, Y., Zhang, J., Shan, S., Chen, X.: Single-side domain generalization for face anti-spoofing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8484-8493 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "21. Jia, Y., Zhang, J., Shan, S., Chen, X.: Unified unsupervised and semi-supervised domain adaptation network for cross-scenario face anti-spoofing. Pattern Recognition 115, 107888 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "22. Jourabloo, A., Liu, Y., Liu, X.: Face de-spoofing: Anti-spoofing via noise modeling. In: ECCV. pp. 290-306 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "23. Li, H., Li, W., Cao, H., Wang, S., Huang, F., Kot, A.C.: Unsupervised domain adaptation for face anti-spoofing. IEEE Transactions on Information Forensics and Security 13(7), 1794-1809 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.438 + ], + "angle": 0, + "content": "24. Liao, C.H., Chen, W.C., Liu, H.T., Yeh, Y.R., Hu, M.C., Chen, C.S.: Domain invariant vision transformer learning for face anti-spoofing. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6098-6107 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.493 + ], + "angle": 0, + "content": "25. Liu, A., Xue, S., Gan, J., Wan, J., Liang, Y., Deng, J., Escalera, S., Lei, Z.: Cfplfas: Class free prompt learning for generalizable face anti-spoofing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 222-232 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.535 + ], + "angle": 0, + "content": "26. Liu, S., Lu, S., Xu, H., Yang, J., Ding, S., Ma, L.: Feature generation and hypothesis verification for reliable face anti-spoofing. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 1782-1791 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.59 + ], + "angle": 0, + "content": "27. Liu, S., Zhang, K.Y., Yao, T., Bi, M., Ding, S., Li, J., Huang, F., Ma, L.: Adaptive normalized representation learning for generalizable face anti-spoofing. In: Proceedings of the 29th ACM international conference on multimedia. pp. 1469-1477 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.633 + ], + "angle": 0, + "content": "28. Liu, S., Zhang, K.Y., Yao, T., Sheng, K., Ding, S., Tai, Y., Li, J., Xie, Y., Ma, L.: Dual reweighting domain generalization for face presentation attack detection. arXiv preprint arXiv:2106.16128 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.785, + 0.674 + ], + "angle": 0, + "content": "29. Liu, S.Q., Lan, X., Yuen, P.C.: Multi-channel remote photoplethysmography correspondence feature for 3d mask face presentation attack detection. IEEE Trans. on Information Forensics and Security 16, 2683-2696 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "30. Liu, S.Q., Lan, X., Yuen, P.C.: Learning temporal similarity of remote photoplethysmography for fast 3d mask face presentation attack detection. IEEE Transactions on Information Forensics and Security 17, 3195-3210 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.744 + ], + "angle": 0, + "content": "31. Liu, Y., Tai, Y., Li, J., Ding, S., Wang, C., Huang, F., Li, D., Qi, W., Ji, R.: Aurora guard: Real-time face anti-spoofing via light reflection. arXiv (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.785, + 0.771 + ], + "angle": 0, + "content": "32. Liu, Y., Jourabloo, A., Liu, X.: Learning deep models for face anti-spoofing: Binary or auxiliary supervision. In: CVPR. pp. 389-398 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "33. Liu, Y., Liu, X.: Spoof trace disentanglement for generic face anti-spoofing. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(3), 3813-3830 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "34. Liu, Y., Stehouwer, J., Jourabloo, A., Liu, X.: Deep tree learning for zero-shot face anti-spoofing. In: CVPR (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.254, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "35. Liu, Y., Chen, Y., Dai, W., Li, C., Zou, J., Xiong, H.: Causal intervention for generalizable face anti-spoofing. In: 2022 IEEE International Conference on Multimedia and Expo (ICME). pp. 01-06. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.217 + ], + "angle": 0, + "content": "36. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "37. Määttra, J., Hadid, A., Pietikainen, M.: Face spoofing detection from single images using micro-texture analysis. In: IJCB. pp. 1-7 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.286 + ], + "angle": 0, + "content": "38. Menotti, D., Chiachia, G., Pinto, A., Schwartz, W.R., Pedrini, H., Falcao, A.X., Rocha, A.: Deep representations for iris, face, and fingerprint spoofing detection. IEEE Trans. on Information Forensics and Security 10(4), 864-879 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "39. Muhammad, U., Holmberg, T., de Melo, W.C., Hadid, A.: Face anti-spoofing via sample learning based recurrent neural network (rnn). In: BMVC. p. 113 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "40. Pan, G., Sun, L., Wu, Z., Lao, S.: Eyeblink-based anti-spoofing in face recognition from a generic webcam camera. In: ICCV. pp. 1-8 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "41. Patel, K., Han, H., Jain, A.K.: Secure face unlock: Spoof detection on smartphones. IEEE transactions on information forensics and security 11(10), 2268-2283 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "42. Patel, K., Han, H., Jain, A.K., Ott, G.: Live face video vs. spoof face video: Use of moiré patterns to detect replay video attacks. In: ICB. pp. 98-105 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "43. Petroni, F., Rocktäschel, T., Lewis, P., Bakhtin, A., Wu, Y., Miller, A.H., Riedel, S.: Language models as knowledge bases? arXiv preprint arXiv:1909.01066 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "44. Qin, Y., Yu, Z., Yan, L., Wang, Z., Zhao, C., Lei, Z.: Meta-teacher for face antispoofing. IEEE transactions on pattern analysis and machine intelligence 44(10), 6311-6326 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.522 + ], + "angle": 0, + "content": "45. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.577 + ], + "angle": 0, + "content": "46. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.632 + ], + "angle": 0, + "content": "47. Shao, R., Lan, X., Li, J., Yuen, P.C.: Multi-adversarial discriminative deep domain generalization for face presentation attack detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10023-10031 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "48. Shao, R., Lan, X., Yuen, P.C.: Regularized fine-grained meta face anti-spoofing. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 11974-11981 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "49. Srivatsan, K., Naseer, M., Nandakumar, K.: Flip: Cross-domain face anti-spoofing with language guidance. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19685-19696 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "50. Tzeng, E., Hoffman, J., Saenko, K., Darrell, T.: Adversarial discriminative domain adaptation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7167-7176 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "51. Wang, C.Y., Lu, Y.D., Yang, S.T., Lai, S.H.: Patchnet: A simple face anti-spoofing framework via fine-grained patch recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20281-20290 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "52. Wang, G., Han, H., Shan, S., Chen, X.: Improving cross-database face presentation attack detection via adversarial domain adaptation. In: 2019 International Conference on Biometrics (ICB). pp. 1-8. IEEE (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.129 + ], + "angle": 0, + "content": "S.Q. Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.203 + ], + "angle": 0, + "content": "53. Wang, G., Han, H., Shan, S., Chen, X.: Cross-domain face presentation attack detection via multi-domain disentangled representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6678-6687 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.786, + 0.245 + ], + "angle": 0, + "content": "54. Wang, G., Han, H., Shan, S., Chen, X.: Unsupervised adversarial domain adaptation for cross-domain face presentation attack detection. IEEE Transactions on Information Forensics and Security 16, 56-69 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.786, + 0.286 + ], + "angle": 0, + "content": "55. Wang, J., Zhang, J., Bian, Y., Cai, Y., Wang, C., Pu, S.: Self-domain adaptation for face anti-spoofing. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 2746-2754 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.786, + 0.328 + ], + "angle": 0, + "content": "56. Wang, Z., Wang, Q., Deng, W., Guo, G.: Face anti-spoofing using transformers with relation-aware mechanism. IEEE Transactions on Biometrics, Behavior, and Identity Science 4(3), 439-450 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.786, + 0.383 + ], + "angle": 0, + "content": "57. Wang, Z., Wang, Z., Yu, Z., Deng, W., Li, J., Gao, T., Wang, Z.: Domain generalization via shuffled style assembly for face anti-spoofing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4123-4133 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.786, + 0.412 + ], + "angle": 0, + "content": "58. Wen, D., Han, H., Jain, A.K.: Face spoof detection with image distortion analysis. IEEE Trans. on Information Forensics and Security 10(4), 746-761 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.786, + 0.439 + ], + "angle": 0, + "content": "59. Yang, J., Lei, Z., Li, S.Z.: Learn convolutional neural network for face anti-spoofing. arXiv preprint arXiv:1408.5601 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.786, + 0.481 + ], + "angle": 0, + "content": "60. Yang, J., Lei, Z., Liao, S., Li, S.Z.: Face liveness detection with component dependent descriptor. In: 2013 International Conference on Biometrics (ICB). pp. 1-6. IEEE (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.786, + 0.509 + ], + "angle": 0, + "content": "61. Yu, Z., Li, X., Niu, X., Shi, J., Zhao, G.: Face anti-spoofing with human material perception. In: ECCV. pp. 557-575. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.786, + 0.55 + ], + "angle": 0, + "content": "62. Yu, Z., Qin, Y., Li, X., Zhao, C., Lei, Z., Zhao, G.: Deep learning for face antispoofing: A survey. IEEE transactions on pattern analysis and machine intelligence 45(5), 5609-5631 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.786, + 0.578 + ], + "angle": 0, + "content": "63. Yu, Z., Qin, Y., Zhao, H., Li, X., Zhao, G.: Dual-cross central difference network for face anti-spoofing. arXiv preprint arXiv:2105.01290 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.786, + 0.619 + ], + "angle": 0, + "content": "64. Yu, Z., Wan, J., Qin, Y., Li, X., Li, S.Z., Zhao, G.: Nas-fas: Static-dynamic central difference network search for face anti-spoofing. IEEE Trans. on Pattern Analysis and Machine Intelligence (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.786, + 0.661 + ], + "angle": 0, + "content": "65. Yue, H., Wang, K., Zhang, G., Feng, H., Han, J., Ding, E., Wang, J.: Cyclically disentangled feature translation for face anti-spoofing. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 3358-3366 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.786, + 0.702 + ], + "angle": 0, + "content": "66. Zhang, Y., Yin, Z., Li, Y., Yin, G., Yan, J., Shao, J., Liu, Z.: Celeba-spoof: Large-scale face anti-spoofing dataset with rich annotations. In: European Conference on Computer Vision (ECCV) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.786, + 0.73 + ], + "angle": 0, + "content": "67. Zhang, Z., Yan, J., Liu, S., Lei, Z., Yi, D., Li, S.Z.: A face antispoofing database with diverse attacks. In: ICB (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.786, + 0.772 + ], + "angle": 0, + "content": "68. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.786, + 0.8 + ], + "angle": 0, + "content": "69. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.786, + 0.842 + ], + "angle": 0, + "content": "70. Zhou, Q., Zhang, K.Y., Yao, T., Lu, X., Ding, S., Ma, L.: Test-time domain generalization for face anti-spoofing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 175-187 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.254, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.128 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.788, + 0.19 + ], + "angle": 0, + "content": "71. Zhou, Q., Zhang, K.Y., Yao, T., Yi, R., Sheng, K., Ding, S., Ma, L.: Generative domain adaptation for face anti-spoofing. In: European Conference on Computer Vision. pp. 335-356. Springer (2022)" + } + ] +] \ No newline at end of file diff --git a/2024/Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing/3e12da5b-7b20-4ba8-b65d-4a8790f34fde_origin.pdf b/2024/Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing/3e12da5b-7b20-4ba8-b65d-4a8790f34fde_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ccbb990fe7b9b91b03e725fb5cc0899456781cfb --- /dev/null +++ b/2024/Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing/3e12da5b-7b20-4ba8-b65d-4a8790f34fde_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b8a75d0b2f528b0252e5237f680cad3ed5bb5a77fe5544a2e928fb93c8d80db +size 6691947 diff --git a/2024/Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing/full.md b/2024/Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..21a89c5a31a709802766c461618238d66c61b447 --- /dev/null +++ b/2024/Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing/full.md @@ -0,0 +1,330 @@ +# Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing + +Si-Qi Liu $^{1,\ast,\boxtimes}$ , Qirui Wang $^{1,\ast}$ , and Pong C. Yuen $^{2}$ + +1 Shenzhen Research Institute of Big Data + +$^{2}$ Department of Computer Science, Hong Kong Baptist University siqiliu@sribd.cn, qwangcw@connect.ust.hk, pcyuen@comp.hkbu.edu.hk + +Abstract. Face anti-spoofing (FAS) which plays an important role in securing face recognition systems has been attracting increasing attention. Recently, vision-language model CLIP has been proven to be effective for FAS, where outstanding performance can be achieved by simply transferring the class label into textual prompt. In this work, we aim to improve the generalization ability of CLIP-based FAS from a prompt learning perspective. Specifically, a Bottom-Up Domain Prompt Tuning method (BUDoPT) that covers the different levels of domain variance, including the domain of recording settings and domain of attack types is proposed. To handle domain discrepancies of recording settings, we design a context-aware adversarial domain-generalized prompt learning strategy that can learn domain-invariant prompt. For spoofing domain with different attack types, we construct a fine-grained textual prompt that guides CLIP to look through the subtle details of different attack instruments. Extensive experiments are conducted on five FAS datasets with variations of camera types, resolutions, image qualities, lighting conditions, and recording environments. The effectiveness of our proposed method is evaluated with different amounts of source domains from multiple angles, where we boost the generalizability compared with the state of the arts with multiple or limited numbers of training datasets. + +Keywords: Face Anti-spoofing $\cdot$ Face presentation attack detection $\cdot$ Face liveness detection + +# 1 Introduction + +With the wide deployment of face recognition in real-world scenarios, the security concerns of FR systems attract increasing attention. Face images can be easily acquired from social networks, and the costs of conducting a face spoofing attack can be very low. Print or screens of mobile phones or tablets are the two common low-cost spoofing instruments. 3D mask attack could be more challenging since both the 3D shape and skin texture can be imitated. To counter these presentation attacks, a series of face anti-spoofing (FAS) methods based on different liveness cues are proposed, various from the hand-crafted features [1,19,40] to the deep-learning-based representations [34,62,64]. + +Early works are mainly developed for intra-domain testing where the recording environment, devices, and attack types are identical. Since the main difference between genuine and attack samples is located on small vestiges such as the texture and color tone, the subtle variations introduced by recording cameras, lighting conditions, and attack mediums form the domain gaps in FAS. + +To improve the robustness and generalizability, the generic domain adaptation and domain generalization approaches are applied or customized in the context of FAS [52,53,57]. Few-shot domain adaptation and unsupervised domain adaptation methods tend to have better results since they can learn from the testing data with or without the labels [21,23,54]. The zero-shot domain generalization setting is more practical as Since the testing data may not be accessible in real scenarios. Popular techniques such as adversarial learning, contrastive learning, or meta-learning are adopted to align representations from different domains and keep on boosting the performances [5,44]. Directly constructing generalized liveness cues can also be effective. The facial heartbeat signal (rPPG) based methods exhibit strong generalizability to different types of attacks but existing methods still require long observation time [29,30]. + +Recently, vision-language models like CLIP exhibited astonishing generalizability and zero-shot performance with the joint embedding of image-text pairs. The first trial of CLIP for FAS (FLIP) that transforms binary class labels into text and measures their similarity to image embeddings also achieves high performance [49]. However, since CLIP is trained for generic tasks with image pairs from the internet, the domain discrepancy of lighting conditions, camera qualities, or attack types in FAS can hardly be well addressed by simply finetuning it with the image-to-class-text contrastive loss. Intuitively, we may enrich the class text prompt with domain labels like the camera types, lighting conditions, or recording scene and construct hand-crafted prompts such as {a photo of [CLS] captured under [Recording Setting]} to learn a fine-grained representation. However, this approach may not work since 1) the detailed domain information in FAS can hardly be defined with text since it is related to multiple factors; 2) learning fine-grained representation that can tell domain discrepancies may contradict the final classification of genuine and spoofing samples. + +In this work, we aim to improve the generalization ability of CLIP-based FAS from a prompt learning perspective. Specifically, we proposed to learn a domain-invariant prompt from the visual embedding that can guide CLIP to avoid interference from the specific domain. Inspired by the CLIP style contrastive learning objective which pulls images to their textual descriptions while pushes away unmatched pairs, we achieve domain-generalized prompt learning by pulling together image embeddings with a domain-invariant prompt and pushing away the image embeddings with their domain-specific prompt. Considering the discrepancies of recording domain are low-level features that scatter in the entire image, we extract domain-specific information from both the facial content and background context regions as the domain-specific negative prompt (push away). The domain-generalized prompt embedding extract from multiple domains is regarded as the positive prompt (pull together). An adversarial + +![](images/4747e110a4a9b784448f68ce50222b72f696944a1937a2f8c6d7f0779f9d8c8e.jpg) +(a) Intuitive solution of CLIP for Face Anti-Spoofing + +![](images/6dfca787fa20bd353a89eaca2669cb1e90e7bd5210b8a12d28402ea5b90d5fd8.jpg) +(b) Bottom-Up Domain-Generalized Prompt Tuning for FAS (ours) +Fig. 1: The comparison between (a) the intuitive solution of CLIP for face anti-spoofing and (b) the proposed bottom-up domain-generalized prompt tuning. The Domain Token prompt guides the model to learn domain-generalized FAS features for different recording domains. The Attack Type prompt boosts the discriminability by encouraging the model to differentiate fine-grained attack instruments. The Domain Token is learned from the visual embedding through the Domain Prompt Learner. + +learning strategy is designed to further ensure the prompt learner can learn the domain-generalized information. The various attack instruments for spoofing samples can be regarded as another perspective of domains. Different from the treatment of recording domain prompt learning, we want the visual encoder to distinguish the attack types so that it can learn more discriminative FAS features [20, 24]. In this case, we form a textual prompt that guides the generic visual encoder to adapt to the FAS task and look through the fine-grained difference for various attack types. We refer to this newly proposed method that learns the low-level domain-generalized visual prompt and tunes high-level attack-type textual prompt as the Bottom-Up Domain Prompt Tuning (BUDoPT) for generalized face anti-spoofing. + +Extensive experiments are conducted on five FAS datasets with a large number of variations (camera types, resolutions, image qualities, lighting conditions, and recording environments). The effectiveness of our proposed method is evaluated with different amounts of source domains from multiple angles, where we boost the generalizability compared with the state of the arts with multiple training datasets or with only one dataset. + +In sum, the main contributions of this work are: + +- We improve the domain generalizability of FAS from a prompt tuning perspective within the context of vision-language model CLIP. A Bottom-Up Domain Prompt Tuning method (BUDoPT) that covers the different levels of domain variance, including the domain of recording settings and domain of attack types is proposed. +- We design 1) a context-aware adversarial domain-generalized prompt learning strategy that can learn domain-invariant prompt for different recording settings, 2) an effective textual prompt that guides CLIP to adapt to FAS task and look through the subtle details of different attack types. +- We improve the generalizability compared with the state of the arts with different scales of training domains, and give in-depth analysis with extensive experiments on five FAS datasets that cover typical variations in FAS. + +# 2 Related Work + +Face Anti-Spoofing. Face anti-spoofing has attracted increasing attention in the last decades, developed from the early handcrafted liveness feature construction [37, 60] to the latest deep learning-based design [64]. Early works focus on searching the discriminative liveness cues such as the appearance artifacts [31, 42, 58], and the motion difference between genuine faces and static attack instruments. Accordingly, hand-crafted features, such as SIFT [41], HoG [60], and LBP [3] are used to characterize spoof patterns. More discriminative features can be learned with Deep learning-based [38, 59]. Auxiliary tasks, such as the regression of facial depth map [32], 3D structure [12], and material [61], encourage the network to learn fine-grained appearance details. There are also classification-based methods [34, 51, 59] and Generative models [22, 33, 57]. Spatiotemporal information can also be learned through CNN and LSTM framework [9, 39]. + +Generalized FAS. Generalized FAS that can perform robustly on different domains are necessary for real-world application scenarios. Recent works proposed to learn the FAS model on multiple source domains and evaluated a target domain. Under the domain adaptation settings where the target domain is accessible, aligning the source and target feature distributions is the typical solution [21,52,53,65,71]. Adversarial learning strategies or the triplet loss encourage the model to learn shared feature for generalization [20,47,57]. Meta-learning and continue learning approaches are also proposed by [5,6,27,44,48] to simulate the domain shift during training. + +Prompt Learning. Prompt learning originated from the field of Natural Language Processing (NLP) motives to shape the output of pre-trained language models by providing contextual information or constraints, thereby enhancing their ability to adapt to downstream tasks [43]. Prompt engineering was extended from NLP to Visual-Language Models [45], such as CLIP which classifies images by matching them with class prompts. CoOp [69] is designed to turn the handcrafted textual prompt into a learnable visual prompt. CoCoOp [68] incorporates the image as the condition for prompts and improves the model's generalization capabilities. Within the visual-textual contrastive learning framework, generalizability can be improved with a smaller data scale. CFPL1 [25] and this work aims to improve the generalizability for FAS with prompt learning. Different from CFPL which regards style (in AdaIN [18]) as the domain-specific cue, in this work we learn the domain-invariant prompt from an intrinsic CLIP-style contrastive learning perspective. + +# 3 Preliminary on CLIP for FAS + +CLIP [45] comprises an image encoder denoted $f_{v}$ and a text encoder denoted $f_{t}$ . In the context of the FAS, the two class text prompt $p_{k}$ encompasses the + +descriptions of real and spoof faces, representing two classes. CLIP determines the similarity between the input image $\pmb{x}$ and $\pmb{p}$ instead of directly training the classification head. The model predicts the image's category $\hat{y}$ by selecting the class prompt with the highest response: + +$$ +\hat {y} = \underset {k \in \{\mathrm {r}, \mathrm {s} \}} {\operatorname {a r g m a x}} \left\langle f _ {v} (\boldsymbol {x}), f _ {t} \left(\boldsymbol {p} _ {k}\right) \right\rangle \tag {1} +$$ + +where $\langle ,\rangle$ denotes the cosine similarity. $p_r$ and $p_s$ represent the prompt with real and spoof face descriptions. + +With such learning framework, pretrained CLIP exhibits great performance in FAS with the standard finetuning process [49]. However, since CLIP is trained for generic tasks with image pairs from the internet, when handling the low-level feature in cross-domain settings in FAS, the domain discrepancy of lighting conditions, camera qualities, or attack types can hardly be well addressed. + +# 4 Proposed method + +Overview. The Bottom-Up Domain Prompt Tuning method (BUDoPT) is proposed to improve the generalization ability of FAS in the context of the vision language model CLIP. On top of the intuitive CLIP-based face anti-spoofing which turns the class label into textual prompt, we incorporate the domain token for the domain discrepancies from different recording settings and the attack type token that describes the fine-grained types of attacks. First, the context-aware adversarial domain-generalized prompt learning is to train the domain prompt learner given the input of facial visual feature. A CLIP score style domain prompt alignment mechanism and an adversarial learning strategy are designed to learn a domain-generalized prompt. Second, the textual prompt of fine-grained attack types aims to boost discriminability by looking through the subtle details of different attack instruments. + +# 4.1 Context-aware Domain Generalized Prompt Learning + +Recording Domain Generalized Prompt Learning. Although text prompt can effectively assist in classifying concrete instances in upper-stream tasks, it encounters difficulties in describing the recording domain with various complex factors, including camera qualities, lighting conditions, and blurriness in FAS. Therefore we proposed the to learn domain-generalized prompt from the visual embedding. Inspired by [68], we employ vision domain features as conditional inputs to generate learnable prompts. The primary objective of the component is learning domain-invariant prompts to mitigate the impact of domain gaps. + +Specifically, we propose a Prompt learner $PL$ to obtain the domain prompt from the image $\mathbf{x}$ . $PL$ utilizes the features from $i$ -th layer of $f_{v}$ as the input. The domain feature $D$ is defined as: + +$$ +D (\boldsymbol {x}) = P L \left(f _ {v, i} (\boldsymbol {x})\right), \tag {2} +$$ + +![](images/9b496ce7dbef0d00662e55612e172142ea398b9c5e9f237694769ceff3d568a9.jpg) +Fig. 2: The proposed Bottom-Up Domain Prompt Tuning method (BUDoPT) covers the different levels of domain variance for generalized FAS, including the domain of recording settings and domain of attack types. The context-aware adversarial domain-generalized prompt learning can learn domain-invariant prompt for different recording settings. The textual prompt of fine-grained attack types can boost the discriminability by looking through the subtle details of different attack instruments. + +where $f_{v,i}$ denotes $i$ -layer features of $f_v$ + +To accurately represent the domain feature while minimizing influences from the content feature of a single image, we utilize all images' domain features in a mini-batch to generate the domain prompt. Given the training mini-batch $\{\pmb{x}_n\}_{n=1}^N$ for domain $m$ , the domain prompt $\pmb{p}_{D_m}$ is defined as: + +$$ +\boldsymbol {p} _ {D _ {m}} = f _ {t} \left(\frac {1}{N} \sum_ {n = 1} ^ {N} D \left(\boldsymbol {x} _ {n} ^ {m}\right)\right), \tag {3} +$$ + +where $N$ is the mini-batch size and $x_{n}^{m}$ denotes the $n$ -th image from $m$ . + +Inspired by the CLIP style contrastive learning objective, we design a CLIP score-based domain generalization learning strategy that pulls together image embeddings with a domain-invariant prompt and pushes away the image embeddings with their domain-specific prompt. Given a batch of training data with $M$ domains, the generalized domain prompt $\pmb{p}_{GD}$ can be defined as: + +$$ +\boldsymbol {p} _ {G D} = f _ {t} \left(\frac {1}{N * M} \sum_ {m = 1} ^ {M} \sum_ {n = 1} ^ {N} D \left(\boldsymbol {x} _ {n} ^ {m}\right)\right), \tag {4} +$$ + +As the Figure 3 shows, only the labels belonging to the generalized domain can be marked as positive while other domain-specific prompts are always labeled + +![](images/90a415da0d07628e0f3f4cde1ba8e925ba73509daa935f6f5ad413bffdb4275a.jpg) +Fig. 3: Illustration of the CLIP score-based domain generalization learning strategy, where we pull together image embeddings with a domain-generalized prompt (▲) and push away the image embeddings with their domain-specific prompt + +as negative. We denote $\pmb{p}$ as the set of prompts and then calculate the image cross-entropy loss $\mathcal{L}_{ce}$ by: + +$$ +\mathcal {L} _ {c e} = - \sum_ {z = 1} ^ {2 \times (M + 1)} y \log \left(\frac {\exp \left(\left\langle f _ {v} (\boldsymbol {x}) , f _ {t} \left(\boldsymbol {p} _ {z}\right) \right\rangle\right)}{\sum_ {k = 1} ^ {2 \times (M + 1)} \exp \left(\left\langle f _ {v} (\boldsymbol {x}), f _ {t} \left(\boldsymbol {p} _ {k}\right) \right\rangle\right)}\right), \tag {5} +$$ + +where $y$ is the ground truth label, $k$ and $z$ denote the indices in $\pmb{p}$ . Note that we have $2 \times (M + 1)$ classes in total, i.e., $M$ domains plus $\pmb{p}_{GD}$ for each class since the prompt set $\pmb{p} = \{\pmb{p}_{D_1}^{real}, \dots, \pmb{p}_{D_M}^{real}, \pmb{p}_{GD}^{real}, \pmb{p}_{D_1}^{fake}, \dots, \pmb{p}_{D_M}^{fake}, \pmb{p}_{GD}^{fake}\}$ , + +During inference, we assume that all samples in the batch are sourced from the same domain and possess the identical prompt. The model selects the class prompt with the highest similarity score as the final prediction. + +Context-aware Domain Clue. To better learn the domain-generalized prompt, we try to extract more domain-specific information from the visual embedding as it is treated as the negative label to be pushed away. Considering the discrepancies of recording domain are low-level features that scatter in the entire image, we extract domain-specific information from both the facial content and background context regions. Specifically, we want to highlight the context information from the surrounding background regions with purer domain information without disturbance of facial content. Consequently, we classify the output patches of $f_{v}$ into two types: context patches $c^{ctx}$ and content patches $c^{ctn}$ . Then the Equation (2) can be rewritten as follows: + +$$ +D (x) = P L ([ c _ {i} ^ {c t x}, c _ {i} ^ {c t n} ]), \tag {6} +$$ + +where $c_{i}^{ctx}$ denotes the context patch obtained from $f_{v,i}$ . To enhance the representation of domain features while reducing the influence of facial content, we extract context patches $c_{s}^{ctx}$ from a shallower layer ( $s < i$ ) of $f_{v}$ and replace the original $c_{i}^{ctx}$ , as shown in Figure 2(green block). For the VIT visual encoder in CLIP, the context patches in shallower layers preserve low-level recording domain information as they have less interaction with content patches. Hence, the domain feature is formed as: + +$$ +D (x) = P L \left(\left[ c _ {s} ^ {c t x}, c _ {i} ^ {c t n} \right]\right), \tag {7} +$$ + +to emphasize the domain-specific information in the prompts. + +Adversarial Domain Prompt Learning. We propose the Adversarial Domain Prompt Learning strategy to further ensure that the prompt learner can learn the domain-generalized information. A Discriminator $DC$ is constructed to differentiate the domain features and determine their belonging domain. $PL$ , as the generator, is required to minimize the domain-specific trace in the output to fool $DC$ : + +$$ +\mathcal {L} _ {a d v} = - \frac {1}{N * M} \sum_ {m = 1} ^ {M} \sum_ {n = 1} ^ {N} y _ {m} \log \left(D C \left(D \left(\boldsymbol {x} _ {n} ^ {m}\right)\right)\right), \tag {8} +$$ + +$y_{m}$ is the domain label. $\mathcal{L}_{adv}$ encourages discriminator $DC$ to try its best to find the domain-specific vestige so that domain prompt learner can learn the domain-invariant information. Empirically, the adversarial learning strategy aggravates the unstable loss convergence. To ensure a more consistent learning direction, we employ a distribution constraint between domain features themselves. The Jensen-Shannon divergence is adopted to pull every two features from $M$ domains together in a training batch: + +$$ +\mathcal {L} _ {d i s} = \sum_ {j, t = 1, \dots , N * M, j < t} \operatorname {J S} \left(D \left(\boldsymbol {x} _ {j}\right), D \left(\boldsymbol {x} _ {t}\right)\right). \tag {9} +$$ + +Here $N * M$ is the total number of samples in a batch. + +# 4.2 Spoofing Domain Discriminative Prompt Tuning. + +In contrast to the visual recording domain generalized prompt that aims to avoid bias, the attack types of spoofing domain can be utilized to boost the discriminability. Specific artifacts are highly correlated with the attack instruments, which can be regarded as spoofing clues, such as the saturation change in printing attacks, the moiré pattern in video replay attacks, and the edge defects of 3D masks. Different from the low-level recording domain prompt, the attack types can be clearly described with the (high-level) textual prompt. Regularize the visual embedding with finer grained attack types text helps the visual encoder learn more discriminative FAS features. This will also set clearer boundaries between genuine faces and spoofing attacks thereby enhancing model generalizability to unseen attack types. Therefore, we propose the Spoof Domain Prompt tuning strategy to encourage the model to differentiate finer-grained attack types. + +Specifically, we design a textual prompt template, a structured sentence that describes spoofing face photos with a placeholder of the attack types, such as "a {fake} face photo {and it is the printing attack}" Given $N * M$ samples in a training batch, we generate the same amount of text prompts $\boldsymbol{p}\boldsymbol{a}$ . Spoofing Domain Discriminative Prompt Tuning updates the parameters of $f_{v}$ with the text cross-entropy loss $\mathcal{L}_t$ , which is calculated by: + +$$ +\mathcal {L} _ {t} = - \sum_ {o = 1} ^ {N * M} o \log \left(\frac {\exp \left(\left\langle f _ {v} (\boldsymbol {x}) , f _ {t} \left(\boldsymbol {p} a _ {o}\right) \right\rangle\right)}{\sum_ {w = 1} ^ {N * M} \exp \left(\left\langle f _ {v} (\boldsymbol {x}) , f _ {t} \left(\boldsymbol {p} a _ {w}\right) \right\rangle\right)}\right), \tag {10} +$$ + +where $o$ and $w$ are the indices in $\pmb{pa}$ . + +The final loss function of our method is formed as: + +$$ +\mathcal {L} = \alpha \mathcal {L} _ {c e} + \beta \mathcal {L} _ {t} - \gamma \mathcal {L} _ {a d v} + \delta \mathcal {L} _ {d i s}. \tag {11} +$$ + +where $\alpha, \beta, \gamma, \delta$ are hyperparameters that balance each loss. + +# 5 Experiments + +# 5.1 Experimental Setups + +Databases and Protocols. We evaluate our method with six widely used FAS databases: MSU-MFSD(M) [58], CASIA-FASD(C) [67], Idiap Replay attack (I) [8], OULU-NPU(O) [2], WMCA(W) [10] and CelebA-Sproof [66]. Five protocols that cover various domain generalization settings with different amounts of source domains are involved to comprehensively evaluate our method's performance. They are leave-one-out(P1) and one-to-one(P2) cross-domain generalization evaluation, different domain partition evaluation(P3), unseen attack type evaluation(P4) and various devices domain evaluation(P5). + +Metrics. Following [49], we evaluate the model performance using three metrics: Half Total Error Rate (HTER), Area Under the Receiver Operating Characteristic Curve (AUC) and True Positive Rate (TPR) at a False Positive Rate (FPR) $1\%$ (TPR@FPR=1%). HTER and AUC reflect the theoretical performance. TPR@FPR=1% evaluates the model performance in practice. + +Implementation Details. We utilize FAN [4] to obtain facial landmarks, aligning the images based on nose and eye landmarks before cropping them to a size of 224x224. We employ the pre-trained CLIP-ViT-B-16 [45] as the foundation model. An eleven-layer VIT following a two-layer MLP is adopted as the prompt learner $PL$ . We extract content patches from the 5th layer of $f_{v}$ and context patches from the 1st layer. For adversarial learning an eight-layer 1D ResNet is adopted as the discriminator $DC$ . + +Training strategy. We divide the training process into two stages for the Bottom-Up domain prompt tuning. In the first training stage, we conduct the Spoofing Domain Discriminative Prompt Tuning to boost the discriminability of $f_{v}$ . Adam optimizer is adopted with a learning rate of $10^{-6}$ . In the second training stage, we freeze $f_{v}$ and train the $PL$ for the Recording Domain Generalization Prompt Learning. The optimizer for both $PL$ and $DC$ are Adam with learning rate $10^{-3}$ and $3 * 10^{-4}$ . Before the adversarial learning, we first train at least one epoch for each domain source to warm up(exclude supplementary training data). Then, we take turns freezing the $PL$ and $DC$ for 50 iterations. For each stage we train 4000 iterations to ensure the largest dataset is traversed. The mini-batch size is $4(N = 4)$ for each domain. $M$ varies according to the number of datasets used for training under different evaluation protocols. The hyperparameter $\alpha$ , $\beta$ , $\gamma$ , and $\delta$ are set as 1, 1, 1, and 0.5, respectively. + +Table 1: Evaluation of cross-domain performance under leave-one-out protocol (P1) + +
MethodsOCI→MOMI→COCM→IICM→OAvg.
HTERAUCTPR@ FPR=1%HTERAUCTPR@ FPR=1%HTERAUCTPR@ FPR=1%HTERAUCTPR@ FPR=1%
MADDG [47]17.6988.06-24.5084.51-22.1984.99-27.9880.02-23.09
MDDR [53]17.0290.10-19.6887.43-20.8786.72-25.02581.47-20.64
NAS-FAS [64]16.8590.42-15.2192.64-11.6396.98-13.1694.18-14.21
RFMeta [48]13.8993.98-20.2788.16-17.3090.48-16.4591.16-16.97
D²AM [7]12.7095.6-20.9885.58-15.4391.22-15.2790.87-16.09
DRDG [28]12.4395.81-19.0588.79-15.5691.79-15.6391.75-15.66
Self-DA [55]15.4091.80-24.5084.40-15.6090.10-23.1084.30-19.65
ANRL [27]10.8396.75-17.8589.26-16.0391.04-15.6791.90-15.09
FGHV [26]9.1796.92-12.4793.47-16.2990.11-13.5893.55-12.87
SSDG-R [20]7.3897.17-10.4495.94-11.7196.59-15.6191.54-11.28
SSAB-R [57]6.6798.75-10.0096.67-8.8896.79-13.7293.65-9.80
PatchNet [51]7.1098.46-11.3394.58-13.4095.67-11.8295.07-10.90
GDA [71]9.2098.00-12.2093.00-10.0096.00-14.4092.60-11.45
LDCformer [16]6.4398.39-8.1196.67-8.5797.09-11.1795.58-8.57
TTDG-V [70]4.1698.48-7.5998.18-9.6298.18-1096.15-7.84
BUDoPT(ours)0.9599.7087.122.8598.0355.004.498.5486.462.2698.7855.252.62
ViT†[15]1.5899.6896.675.7098.9788.579.2597.1551.547.4798.4269.306.00
FLIP-V†[49]3.7999.3187.991.2799.7595.854.7198.8075.844.1598.7666.473.48
FLIP-IT†[49]5.2798.4279.330.4499.9899.862.9499.4284.623.6199.1584.763.06
FLIP-MCL†[49]4.9598.1174.670.5499.98100.004.2599.0784.622.3199.6392.283.01
BUDoPT-CoOp†1.9899.5480.631.9399.5888.063.1899.390.43.3398.9189.292.61
BUDoPT-CoCoOp†1.5199.4375.551.5599.5684.743.5399.292.533.5198.9887.982.53
BUDoPT(ours)†0.4099.9999.670.2699.9699.921.3899.6998.11.6099.5197.180.91
+ +$\dagger$ indicates CelebA-Sproof [66] is employed as supplementary training dataset + +# 5.2 Evaluation + +Cross-domain generalization evaluation. Following [49], the leave-one-out (P1) and one-to-one (P2) on M, C, I and O are conducted to evaluate cross-dataset performance. $\mathrm{OCM} \rightarrow \mathrm{I}$ refers to the model trains on O, C, M and tests on I. For a fair comparison with [49] and other models, we evaluate the performance of our method with and without the supplementary training data CelebA-Sproof [66]. For simplicity, we regard dataset as the domain partition in our proposed method. As shown in Table 1, without supplementary training data, the proposed BUDoPT outperforms SOTA in all settings. With supplementary training data, BUDoPT also surpasses the FLIP family model in HTER for all settings and achieves more than $69\%$ improvement on average. In terms of TPR@FPR $= 1\%$ , our method performs better in three out of four settings. Although the performance slightly drops in TPR@FPR $= 1\% (\mathrm{C})$ and AUC(C,O), the margin is relatively small as they already exceed $99\%$ . + +In P2 of low-data regime [49] the model is trained separately on M, C, I and O on the remaining datasets, resulting in a total of 12 different scenario combinations. Since the model is trained on one dataset only, domain partition is based on capturing devices(C,M), environments(I), and sessions(O) for different datasets. We also test our method with and without supplementary CelebA-Sproof. CelebA-Sproof is regarded as a domain if it is involved. As illustrated in Table 2, without help of CelebA-Sproof, BUDoPT performs better than SOTA in nine out of twelve settings. We also outperform domain adap + +Table 2: Evaluation of cross-domain performance under one-to-one protocol (P2) + +
MethodsUnseenC → IC → MC → OI → CI → MOM → CM → IM → OO → CO → IO → MAvg.
ADDA [50]No41.836.6-49.835.1-39.035.2----39.6
DRCN [11]No44.427.4-48.942.0-28.936.8----38.1
DupGAN [14]No42.133.4-46.536.2-27.135.4----36.8
KSA [23]No39.315.1-12.333.3-9.134.9----24.0
DR-UDA [54]No15.69.028.734.229.03.516.83.030.219.525.427.423.1
ADA [52]No17.59.329.141.530.539.617.75.131.219.826.831.525.0
USDAN-Un [21]No16.09.2-30.225.8-13.33.4----16.3
GDA [71]No15.105.8-29.720.8-12.22.5----14.4
CDFTN-L [65]No1.78.129.911.99.629.98.81.325.619.15.86.313.2
BUDoPT(ours)Yes9.654.376.768.265.647.736.072.96.164.812.972.945.69
FLIP-V†[49]Yes15.0813.7312.344.309.687.870.563.964.792.095.016.007.12
FLIP-IT†[49]Yes12.3315.187.981.128.376.980.195.214.960.164.275.636.03
FLIP-MCL†[49]Yes10.577.153.910.687.224.220.195.883.950.195.698.404.84
BUDoPT(ours)†Yes4.332.624.980.481.834.1402.451.870.442.531.432.46
+ +$\dagger$ indicates CelebA-Sproof is employed as supplementary training data. + +tive models (Unseen=No) even without any information of target domain. With CelebA-Sproof, BUDoPT surpasses FLIP family models in ten settings. + +Effect of Domain Partition. We set $\mathrm{MI} \rightarrow \mathrm{O}$ and $\mathrm{MI} \rightarrow \mathrm{C}$ to evaluate our model with different amounts of training data sources. Here we treat dataset as domain. Results in Table 3 show that with less training data, BUDoPT outperforms competitors by a larger margin. However, we also note that comparing with the $\mathrm{M} \rightarrow \mathrm{O}$ results in P2 the performance increase is limited. We believe this is due to the difference in domain partition strategies. For $\mathrm{M} \rightarrow \mathrm{O}$ the domain is defined based on the device while for $\mathrm{MI} \rightarrow \mathrm{O}$ it is based on dataset. The results indicate that a fine-grained partition strategy may help achieve better performance for our method even with a limited data scale. More analysis of partition strategy is in supplementary Sec 2.1. + +Generalize to Unseen attack type. We employ $\mathbf{W}$ with high-quality 3D silicon mask attack samples to evaluate the model generalizability to unseen attack types (P4). Different from the baselines trained on IMCO, we only utilize $\mathbf{C}$ as the source dataset. Same domain partition strategy in P2 is adopted. As shown in Table 4, with less training data BUDoPT still outperforms the others. This result not only demonstrates our generalizability to the unseen attack types but also shows the learning efficiency of our method. + +Device Variance Only. In P5, we utilize the O's intra protocol 3 and protocol 4 [2] to evaluate the performance with device variance only. The domain gaps in P5 are located on the unseen device settings and parameters and other underlying factors are similar. Same domain partition strategy in P2 is adopted. As shown in Table 5, under constraint setting BUDoPT can handle the domain gap of devices well. This indicates the usability in real applications where some local samples can be available for fine-tuning. + +Table 3: Evaluation of different domain partition (P3) + +
MethodsMI → CMI → O
HTERAUCHTERAUC
SSAN-R [57]25.5683.8924.4482.86
HFN+MP [6]30.8972.4820.9486.71
CIFAS [35]22.6783.8924.6381.48
DiVT-M [24]20.1186.7123.6185.73
SSDG-R [20]19.8686.4627.9278.72
DGUA-FAS [13]19.2286.8120.0588.75
BUDoPT(ours)5.3398.925.9498.37
+ +Table 4: Evaluation of unseen attack type performance (P4) + +
MethodsIMCO → WMCA
HTERAUC
DiVT-M [24]22.3686.82
DGUA-FAS [13]20.6288.07
C → WMCA
BUDoPT(ours)9.2095.93
+ +Table 5: Evaluation of Device Variance Only (P5) on O. P refers protocol + +
MethodsPACPERBCPERACER
DC-CDN [63]32.2±2.81.6±2.11.9±1.1
RAEDFL [17]31.38±1.780.28±0.680.83±0.86
TranFAS [56]30.6±0.71.1±2.50.9±1.1
LDCFormer [16]32.35±2.050.28±0.681.31±1.03
BUDoPT(ours)30±00±00±0
DC-CDN [63]45.4±3.32.5±4.24.0±3.1
RAEDFL [17]45.41±6.402.50±2.743.96±3.90
TranFAS [56]42.1±2.23.8±3.52.9±2.4
LDCFormer [16]41.08±1.281.17±1.941.13±1.02
BUDoPT(ours)40±00±00±0
+ +Table 6: Ablation study of major components of the proposed BUDoPT + +
SDSDDRDGADCDCICM → O
✓----3.51
-✓---5.78
✓-✓--3.01
✓-✓✓-2.50
✓-✓✓✓2.26
+ +# 5.3 Ablation Studies + +We analyze the effectiveness of major components(Table 6), test the intuitive textual recording domain prompt tuning(Table 6), and compared with generic visual prompt tuning CoOp [69], CoCoOp [68](Table 1). + +Table 6 reports the results of the ablation studies of our major components. We set Spoofing Domain Discriminative Prompt Tuning(SD) as the baseline and evaluate performance with and without Recording Domain Generalization Prompt Learning (RDG), Adversarial Domain Learning(AD) and Context-aware Domain Clue(CDC). With Recording Domain Generalization Prompt Learning, the HTER is improved by $14\%$ . With Adversarial Domain Learning, the HTER decreases by around $17\%$ . The Context-aware Domain Clue contributes around $10\%$ improvement. The ablation studies indicate that each component within our method is crucial for improving the domain generalization ability. + +Intuitive textual recording domain prompt tuning. Considering the success of employing context information as a prompt in upper-stream tasks, we evaluate an intuitive domain prompt tuning strategy, i.e., using the textual domain description as the prompt. We mark the textual domain prompts with the + +![](images/f81e81fb30622b2a61c02f1227aee00a55c968ebf485b4387670203becc4f325.jpg) +Fig. 4: The t-SNE visualizations of original visual embedding (first row) and the learned domain-invariant prompt embedding (second row) under P1. + +Spoofing Domain Discriminative Prompt Tuning as SDD2. As shown in Table 6, adding the intuitive textual recording domain prompt (SDD) harms the model's performance. This aligns with our analysis that the textual domain prompt of context encourages the model to extract more domain-specific information and contradicts the goal of domain-generalized FAS. + +Generic visual prompt tuning. Since the proposed method follows the idea of visual prompt tuning, the classical visual prompt learning methods: CoOp and CoCoOp are compared. For a fair comparison, the CLIP visual encoder of the two models is finetuned with the Spoofing Domain Discriminative Prompt Learning module in our framework so we regard them as BUDoPT-CoOp and BUDoPT-CoCoOp. As the results shown in Table 1, by comparing with the FLIP-family models, BUDoPT-CoOp, and BUDoPT-CoCoOp, we comprehensively verify the effectiveness of our proposed method that adapts CLIP to FAS with better generalization ability. + +We also conduct hyperparameter analysis of loss function. Results are summarized in Eq. (11) in the supplementary Section 2.2. + +# 5.4 Visualizations + +Domain Prompt Visualization. Figure 4 shows the t-SNE [36] visualization of original visual embedding (first row) and the learned domain-invariant prompt embedding (second row) under P1. The original visual embedding naturally forms multiple clusters of domain-specific information. Each cluster shares some low-level underlying factors, such as the camera qualities (see supplementary material for details). Our BUDoPT learns a domain-generalized prompt embedding where samples from different domains are grouped together in a single cluster, thereby guiding the model to minimize the influences of domain gaps. Attention Visualization. We utilize the GradCam [46] to visualize attention + +![](images/ea9e1286e03706a17b62f75e42ccd2ae27fa58fe1e7e780a2492e397a300f9ae.jpg) +Fig. 5: Typical attention maps of real and fake samples under P1. + +![](images/13948c45e91ce012c3b2fc4aae866a1b3f37457b3e5801b836cde11a51cc8343.jpg) + +![](images/46d5899b36ace9e81d326b04129d87f6a4f6c228bf197a47658c5e88ad66c5d2.jpg) + +![](images/7ec57ffabc1d59b0bb99eb0b60c9e903e1c609f9b2da4654cc4bbb9a5de02cda.jpg) + +maps for genuine and spoofing samples in P1. Typical results are visualized in Figure 5. We can observe that BUDoPT takes different attention for live and spoof faces. Specifically, the model focuses on face regions for real samples while highlights spoof clues of fake faces. With the help of Spoofing Domain Discriminative Prompt Tuning, the model is able to detect the intrinsic spoofing clues for different types of attack, including moiré patterns(M, O), paper mask edges(C) and paper texture(I). + +# 6 Conclusion + +In this work, we aim to improve the generalization ability of CLIP-based FAS from a prompt learning perspective. A Bottom-Up Domain Prompt Tuning method (BUDoPT) that covers the different levels of domain variance, including the domain of recording settings and domain of attack types is proposed. It contains a context-aware adversarial domain-generalized prompt learning strategy that can learn domain-invariant prompt for different recording settings and a fine-grained textual prompt that guides CLIP to look through the subtle details of different attack instruments. Extensive experiments are conducted on five FAS datasets with a large number of variations (camera types, resolutions, image qualities, lighting conditions, and recording environments). The effectiveness of our proposed method is evaluated with different amounts of source domains from multiple angles, where we boost the generalizability compared with the state of the arts with multiple training datasets or with only one dataset. We also observe the limitations of our methods with different domain partitioning strategies. Designing an adaptive domain partition approach is a potential way to further improve the generalizability of our method. + +# 7 Acknowledgement + +This work is supported by GuangDong Basic and Applied Basic Research Foundation 2023A1515110706, RGC-NSFC joint project N_HKBU212/23, and NSFC Young Scientist Fund Application No. 6240075068. + +# References + +1. Bao, W., Li, H., Li, N., Jiang, W.: A liveness detection method for face recognition based on optical flow field. In: IASP (2009) +2. Boulkenafet, Z., Komulainen, J., Li, L., Feng, X., Hadid, A.: OULU-NPU: A mobile face presentation attack database with real-world variations. In: FG (2017) +3. Boulkenafet, Z., Komulainen, J., Hadid, A.: Face spoofing detection using colour texture analysis. IEEE Trans. on Information Forensics and Security 11(8), 1818-1830 (2016) +4. Bulat, A., Tzimiropoulos, G.: How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In: International Conference on Computer Vision (2017) +5. Cai, R., Cui, Y., Li, Z., Yu, Z., Li, H., Hu, Y., Kot, A.: Rehearsal-free domain continual face anti-spoofing: Generalize more and forget less. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8037-8048 (2023) +6. Cai, R., Li, Z., Wan, R., Li, H., Hu, Y., Kot, A.C.: Learning meta pattern for face anti-spoofing. IEEE Transactions on Information Forensics and Security 17, 1201-1213 (2022) +7. Chen, Z., Yao, T., Sheng, K., Ding, S., Tai, Y., Li, J., Huang, F., Jin, X.: Generalizable representation learning for mixture domain face anti-spoofing. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 1132-1139 (2021) +8. Chingovska, I., Anjos, A., Marcel, S.: On the effectiveness of local binary patterns in face anti-spoofing. In: BIOSIG. pp. 1-7 (2012) +9. Ge, H., Tu, X., Ai, W., Luo, Y., Ma, Z., Xie, M.: Face anti-spoofing by the enhancement of temporal motion. In: CTISC. pp. 106-111. IEEE (2020) +0. George, A., Mostaani, Z., Geissenbuhler, D., Nikisins, O., Anjos, A., Marcel, S.: Biometric face presentation attack detection with multi-channel convolutional neural network. IEEE Trans. on Information Forensics and Security (2019) +1. Ghifary, M., Kleijn, W.B., Zhang, M., Balduzzi, D., Li, W.: Deep reconstruction-classification networks for unsupervised domain adaptation. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14. pp. 597-613. Springer (2016) +2. Guo, J., Zhu, X., Xiao, J., Lei, Z., Wan, G., Li, S.Z.: Improving face anti-spoofing by 3d virtual synthesis. arXiv (2019) +3. Hong, Z.W., Lin, Y.C., Liu, H.T., Yeh, Y.R., Chen, C.S.: Domain-generalized face anti-spoofing with unknown attacks. In: 2023 IEEE International Conference on Image Processing (ICIP). pp. 820-824. IEEE (2023) +4. Hu, L., Kan, M., Shan, S., Chen, X.: Duplex generative adversarial network for unsupervised domain adaptation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1498-1507 (2018) +5. Huang, H.P., Sun, D., Liu, Y., Chu, W.S., Xiao, T., Yuan, J., Adam, H., Yang, M.H.: Adaptive transformers for robust few-shot cross-domain face anti-spoofing. In: European Conference on Computer Vision. pp. 37-54. Springer (2022) +6. Huang, P.K., Chiang, C.H., Chong, J.X., Chen, T.H., Ni, H.Y., Hsu, C.T.: Ldcformer: Incorporating learnable descriptive convolution to vision transformer for face anti-spoofing. In: 2023 IEEE International Conference on Image Processing (ICIP). pp. 121-125. IEEE (2023) +7. Huang, P.K., Chin, M.C., Hsu, C.T.: Face anti-spoofing via robust auxiliary estimation and discriminative feature learning. In: Asian Conference on Pattern Recognition. pp. 443-458. Springer (2021) + +18. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: Proceedings of the IEEE international conference on computer vision. pp. 1501-1510 (2017) +19. Jee, H.K., Jung, S.U., Yoo, J.H.: Liveness detection for embedded face recognition system. International Journal of Biological and Medical Sciences 1(4), 235-238 (2006) +20. Jia, Y., Zhang, J., Shan, S., Chen, X.: Single-side domain generalization for face anti-spoofing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8484-8493 (2020) +21. Jia, Y., Zhang, J., Shan, S., Chen, X.: Unified unsupervised and semi-supervised domain adaptation network for cross-scenario face anti-spoofing. Pattern Recognition 115, 107888 (2021) +22. Jourabloo, A., Liu, Y., Liu, X.: Face de-spoofing: Anti-spoofing via noise modeling. In: ECCV. pp. 290-306 (2018) +23. Li, H., Li, W., Cao, H., Wang, S., Huang, F., Kot, A.C.: Unsupervised domain adaptation for face anti-spoofing. IEEE Transactions on Information Forensics and Security 13(7), 1794-1809 (2018) +24. Liao, C.H., Chen, W.C., Liu, H.T., Yeh, Y.R., Hu, M.C., Chen, C.S.: Domain invariant vision transformer learning for face anti-spoofing. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6098-6107 (2023) +25. Liu, A., Xue, S., Gan, J., Wan, J., Liang, Y., Deng, J., Escalera, S., Lei, Z.: Cfplfas: Class free prompt learning for generalizable face anti-spoofing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 222-232 (2024) +26. Liu, S., Lu, S., Xu, H., Yang, J., Ding, S., Ma, L.: Feature generation and hypothesis verification for reliable face anti-spoofing. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 1782-1791 (2022) +27. Liu, S., Zhang, K.Y., Yao, T., Bi, M., Ding, S., Li, J., Huang, F., Ma, L.: Adaptive normalized representation learning for generalizable face anti-spoofing. In: Proceedings of the 29th ACM international conference on multimedia. pp. 1469-1477 (2021) +28. Liu, S., Zhang, K.Y., Yao, T., Sheng, K., Ding, S., Tai, Y., Li, J., Xie, Y., Ma, L.: Dual reweighting domain generalization for face presentation attack detection. arXiv preprint arXiv:2106.16128 (2021) +29. Liu, S.Q., Lan, X., Yuen, P.C.: Multi-channel remote photoplethysmography correspondence feature for 3d mask face presentation attack detection. IEEE Trans. on Information Forensics and Security 16, 2683-2696 (2021) +30. Liu, S.Q., Lan, X., Yuen, P.C.: Learning temporal similarity of remote photoplethysmography for fast 3d mask face presentation attack detection. IEEE Transactions on Information Forensics and Security 17, 3195-3210 (2022) +31. Liu, Y., Tai, Y., Li, J., Ding, S., Wang, C., Huang, F., Li, D., Qi, W., Ji, R.: Aurora guard: Real-time face anti-spoofing via light reflection. arXiv (2019) +32. Liu, Y., Jourabloo, A., Liu, X.: Learning deep models for face anti-spoofing: Binary or auxiliary supervision. In: CVPR. pp. 389-398 (2018) +33. Liu, Y., Liu, X.: Spoof trace disentanglement for generic face anti-spoofing. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(3), 3813-3830 (2022) +34. Liu, Y., Stehouwer, J., Jourabloo, A., Liu, X.: Deep tree learning for zero-shot face anti-spoofing. In: CVPR (2019) + +35. Liu, Y., Chen, Y., Dai, W., Li, C., Zou, J., Xiong, H.: Causal intervention for generalizable face anti-spoofing. In: 2022 IEEE International Conference on Multimedia and Expo (ICME). pp. 01-06. IEEE (2022) +36. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008) +37. Määttra, J., Hadid, A., Pietikainen, M.: Face spoofing detection from single images using micro-texture analysis. In: IJCB. pp. 1-7 (2011) +38. Menotti, D., Chiachia, G., Pinto, A., Schwartz, W.R., Pedrini, H., Falcao, A.X., Rocha, A.: Deep representations for iris, face, and fingerprint spoofing detection. IEEE Trans. on Information Forensics and Security 10(4), 864-879 (2015) +39. Muhammad, U., Holmberg, T., de Melo, W.C., Hadid, A.: Face anti-spoofing via sample learning based recurrent neural network (rnn). In: BMVC. p. 113 (2019) +40. Pan, G., Sun, L., Wu, Z., Lao, S.: Eyeblink-based anti-spoofing in face recognition from a generic webcam camera. In: ICCV. pp. 1-8 (2007) +41. Patel, K., Han, H., Jain, A.K.: Secure face unlock: Spoof detection on smartphones. IEEE transactions on information forensics and security 11(10), 2268-2283 (2016) +42. Patel, K., Han, H., Jain, A.K., Ott, G.: Live face video vs. spoof face video: Use of moiré patterns to detect replay video attacks. In: ICB. pp. 98-105 (2015) +43. Petroni, F., Rocktäschel, T., Lewis, P., Bakhtin, A., Wu, Y., Miller, A.H., Riedel, S.: Language models as knowledge bases? arXiv preprint arXiv:1909.01066 (2019) +44. Qin, Y., Yu, Z., Yan, L., Wang, Z., Zhao, C., Lei, Z.: Meta-teacher for face antispoofing. IEEE transactions on pattern analysis and machine intelligence 44(10), 6311-6326 (2021) +45. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +46. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017) +47. Shao, R., Lan, X., Li, J., Yuen, P.C.: Multi-adversarial discriminative deep domain generalization for face presentation attack detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10023-10031 (2019) +48. Shao, R., Lan, X., Yuen, P.C.: Regularized fine-grained meta face anti-spoofing. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 11974-11981 (2020) +49. Srivatsan, K., Naseer, M., Nandakumar, K.: Flip: Cross-domain face anti-spoofing with language guidance. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19685-19696 (2023) +50. Tzeng, E., Hoffman, J., Saenko, K., Darrell, T.: Adversarial discriminative domain adaptation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7167-7176 (2017) +51. Wang, C.Y., Lu, Y.D., Yang, S.T., Lai, S.H.: Patchnet: A simple face anti-spoofing framework via fine-grained patch recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20281-20290 (2022) +52. Wang, G., Han, H., Shan, S., Chen, X.: Improving cross-database face presentation attack detection via adversarial domain adaptation. In: 2019 International Conference on Biometrics (ICB). pp. 1-8. IEEE (2019) + +53. Wang, G., Han, H., Shan, S., Chen, X.: Cross-domain face presentation attack detection via multi-domain disentangled representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6678-6687 (2020) +54. Wang, G., Han, H., Shan, S., Chen, X.: Unsupervised adversarial domain adaptation for cross-domain face presentation attack detection. IEEE Transactions on Information Forensics and Security 16, 56-69 (2020) +55. Wang, J., Zhang, J., Bian, Y., Cai, Y., Wang, C., Pu, S.: Self-domain adaptation for face anti-spoofing. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 2746-2754 (2021) +56. Wang, Z., Wang, Q., Deng, W., Guo, G.: Face anti-spoofing using transformers with relation-aware mechanism. IEEE Transactions on Biometrics, Behavior, and Identity Science 4(3), 439-450 (2022) +57. Wang, Z., Wang, Z., Yu, Z., Deng, W., Li, J., Gao, T., Wang, Z.: Domain generalization via shuffled style assembly for face anti-spoofing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4123-4133 (2022) +58. Wen, D., Han, H., Jain, A.K.: Face spoof detection with image distortion analysis. IEEE Trans. on Information Forensics and Security 10(4), 746-761 (2015) +59. Yang, J., Lei, Z., Li, S.Z.: Learn convolutional neural network for face anti-spoofing. arXiv preprint arXiv:1408.5601 (2014) +60. Yang, J., Lei, Z., Liao, S., Li, S.Z.: Face liveness detection with component dependent descriptor. In: 2013 International Conference on Biometrics (ICB). pp. 1-6. IEEE (2013) +61. Yu, Z., Li, X., Niu, X., Shi, J., Zhao, G.: Face anti-spoofing with human material perception. In: ECCV. pp. 557-575. Springer (2020) +62. Yu, Z., Qin, Y., Li, X., Zhao, C., Lei, Z., Zhao, G.: Deep learning for face antispoofing: A survey. IEEE transactions on pattern analysis and machine intelligence 45(5), 5609-5631 (2022) +63. Yu, Z., Qin, Y., Zhao, H., Li, X., Zhao, G.: Dual-cross central difference network for face anti-spoofing. arXiv preprint arXiv:2105.01290 (2021) +64. Yu, Z., Wan, J., Qin, Y., Li, X., Li, S.Z., Zhao, G.: Nas-fas: Static-dynamic central difference network search for face anti-spoofing. IEEE Trans. on Pattern Analysis and Machine Intelligence (2020) +65. Yue, H., Wang, K., Zhang, G., Feng, H., Han, J., Ding, E., Wang, J.: Cyclically disentangled feature translation for face anti-spoofing. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 3358-3366 (2023) +66. Zhang, Y., Yin, Z., Li, Y., Yin, G., Yan, J., Shao, J., Liu, Z.: Celeba-spoof: Large-scale face anti-spoofing dataset with rich annotations. In: European Conference on Computer Vision (ECCV) (2020) +67. Zhang, Z., Yan, J., Liu, S., Lei, Z., Yi, D., Li, S.Z.: A face antispoofing database with diverse attacks. In: ICB (2012) +68. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022) +69. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022) +70. Zhou, Q., Zhang, K.Y., Yao, T., Lu, X., Ding, S., Ma, L.: Test-time domain generalization for face anti-spoofing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 175-187 (2024) + +71. Zhou, Q., Zhang, K.Y., Yao, T., Yi, R., Sheng, K., Ding, S., Ma, L.: Generative domain adaptation for face anti-spoofing. In: European Conference on Computer Vision. pp. 335-356. Springer (2022) \ No newline at end of file diff --git a/2024/Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing/images.zip b/2024/Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..522bcbd4335474c6aedaa43ec1ed3ea603f7cb51 --- /dev/null +++ b/2024/Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:339212c28083619733a01e351544846d3b561055e170c79ed363b4589e2f6f90 +size 655964 diff --git a/2024/Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing/layout.json b/2024/Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e2d996e48ba9fe4f2eac2dad95af4566fd8e9a62 --- /dev/null +++ b/2024/Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing/layout.json @@ -0,0 +1,10145 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 168, + 111, + 445, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 111, + 445, + 148 + ], + "spans": [ + { + "bbox": [ + 168, + 111, + 445, + 148 + ], + "type": "text", + "content": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 196, + 167, + 416, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 167, + 416, + 180 + ], + "spans": [ + { + "bbox": [ + 196, + 167, + 416, + 180 + ], + "type": "text", + "content": "Si-Qi Liu" + }, + { + "bbox": [ + 196, + 167, + 416, + 180 + ], + "type": "inline_equation", + "content": "^{1,\\ast,\\boxtimes}" + }, + { + "bbox": [ + 196, + 167, + 416, + 180 + ], + "type": "text", + "content": ", Qirui Wang" + }, + { + "bbox": [ + 196, + 167, + 416, + 180 + ], + "type": "inline_equation", + "content": "^{1,\\ast}" + }, + { + "bbox": [ + 196, + 167, + 416, + 180 + ], + "type": "text", + "content": ", and Pong C. Yuen" + }, + { + "bbox": [ + 196, + 167, + 416, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 219, + 190, + 395, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 190, + 395, + 201 + ], + "spans": [ + { + "bbox": [ + 219, + 190, + 395, + 201 + ], + "type": "text", + "content": "1 Shenzhen Research Institute of Big Data" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 156, + 201, + 457, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 201, + 457, + 223 + ], + "spans": [ + { + "bbox": [ + 156, + 201, + 457, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 156, + 201, + 457, + 223 + ], + "type": "text", + "content": " Department of Computer Science, Hong Kong Baptist University siqiliu@sribd.cn, qwangcw@connect.ust.hk, pcyuen@comp.hkbu.edu.hk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 247, + 455, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 247, + 455, + 467 + ], + "spans": [ + { + "bbox": [ + 159, + 247, + 455, + 467 + ], + "type": "text", + "content": "Abstract. Face anti-spoofing (FAS) which plays an important role in securing face recognition systems has been attracting increasing attention. Recently, vision-language model CLIP has been proven to be effective for FAS, where outstanding performance can be achieved by simply transferring the class label into textual prompt. In this work, we aim to improve the generalization ability of CLIP-based FAS from a prompt learning perspective. Specifically, a Bottom-Up Domain Prompt Tuning method (BUDoPT) that covers the different levels of domain variance, including the domain of recording settings and domain of attack types is proposed. To handle domain discrepancies of recording settings, we design a context-aware adversarial domain-generalized prompt learning strategy that can learn domain-invariant prompt. For spoofing domain with different attack types, we construct a fine-grained textual prompt that guides CLIP to look through the subtle details of different attack instruments. Extensive experiments are conducted on five FAS datasets with variations of camera types, resolutions, image qualities, lighting conditions, and recording environments. The effectiveness of our proposed method is evaluated with different amounts of source domains from multiple angles, where we boost the generalizability compared with the state of the arts with multiple or limited numbers of training datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 476, + 453, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 476, + 453, + 498 + ], + "spans": [ + { + "bbox": [ + 159, + 476, + 453, + 498 + ], + "type": "text", + "content": "Keywords: Face Anti-spoofing " + }, + { + "bbox": [ + 159, + 476, + 453, + 498 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 476, + 453, + 498 + ], + "type": "text", + "content": " Face presentation attack detection " + }, + { + "bbox": [ + 159, + 476, + 453, + 498 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 476, + 453, + 498 + ], + "type": "text", + "content": " Face liveness detection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 517, + 230, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 517, + 230, + 529 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 230, + 529 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 540, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 482, + 647 + ], + "type": "text", + "content": "With the wide deployment of face recognition in real-world scenarios, the security concerns of FR systems attract increasing attention. Face images can be easily acquired from social networks, and the costs of conducting a face spoofing attack can be very low. Print or screens of mobile phones or tablets are the two common low-cost spoofing instruments. 3D mask attack could be more challenging since both the 3D shape and skin texture can be imitated. To counter these presentation attacks, a series of face anti-spoofing (FAS) methods based on different liveness cues are proposed, various from the hand-crafted features [1,19,40] to the deep-learning-based representations [34,62,64]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 332, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 332, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 332, + 665 + ], + "type": "text", + "content": "* Equal contribution, ⌒ Corresponding author" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "Early works are mainly developed for intra-domain testing where the recording environment, devices, and attack types are identical. Since the main difference between genuine and attack samples is located on small vestiges such as the texture and color tone, the subtle variations introduced by recording cameras, lighting conditions, and attack mediums form the domain gaps in FAS." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 179, + 480, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 179, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 179, + 480, + 323 + ], + "type": "text", + "content": "To improve the robustness and generalizability, the generic domain adaptation and domain generalization approaches are applied or customized in the context of FAS [52,53,57]. Few-shot domain adaptation and unsupervised domain adaptation methods tend to have better results since they can learn from the testing data with or without the labels [21,23,54]. The zero-shot domain generalization setting is more practical as Since the testing data may not be accessible in real scenarios. Popular techniques such as adversarial learning, contrastive learning, or meta-learning are adopted to align representations from different domains and keep on boosting the performances [5,44]. Directly constructing generalized liveness cues can also be effective. The facial heartbeat signal (rPPG) based methods exhibit strong generalizability to different types of attacks but existing methods still require long observation time [29,30]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 327, + 481, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 481, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 481, + 506 + ], + "type": "text", + "content": "Recently, vision-language models like CLIP exhibited astonishing generalizability and zero-shot performance with the joint embedding of image-text pairs. The first trial of CLIP for FAS (FLIP) that transforms binary class labels into text and measures their similarity to image embeddings also achieves high performance [49]. However, since CLIP is trained for generic tasks with image pairs from the internet, the domain discrepancy of lighting conditions, camera qualities, or attack types in FAS can hardly be well addressed by simply finetuning it with the image-to-class-text contrastive loss. Intuitively, we may enrich the class text prompt with domain labels like the camera types, lighting conditions, or recording scene and construct hand-crafted prompts such as {a photo of [CLS] captured under [Recording Setting]} to learn a fine-grained representation. However, this approach may not work since 1) the detailed domain information in FAS can hardly be defined with text since it is related to multiple factors; 2) learning fine-grained representation that can tell domain discrepancies may contradict the final classification of genuine and spoofing samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 666 + ], + "type": "text", + "content": "In this work, we aim to improve the generalization ability of CLIP-based FAS from a prompt learning perspective. Specifically, we proposed to learn a domain-invariant prompt from the visual embedding that can guide CLIP to avoid interference from the specific domain. Inspired by the CLIP style contrastive learning objective which pulls images to their textual descriptions while pushes away unmatched pairs, we achieve domain-generalized prompt learning by pulling together image embeddings with a domain-invariant prompt and pushing away the image embeddings with their domain-specific prompt. Considering the discrepancies of recording domain are low-level features that scatter in the entire image, we extract domain-specific information from both the facial content and background context regions as the domain-specific negative prompt (push away). The domain-generalized prompt embedding extract from multiple domains is regarded as the positive prompt (pull together). An adversarial" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "S.Q. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 272, + 175 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 272, + 175 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 272, + 175 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 272, + 175 + ], + "type": "image", + "image_path": "4747e110a4a9b784448f68ce50222b72f696944a1937a2f8c6d7f0779f9d8c8e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 179, + 279, + 187 + ], + "lines": [ + { + "bbox": [ + 143, + 179, + 279, + 187 + ], + "spans": [ + { + "bbox": [ + 143, + 179, + 279, + 187 + ], + "type": "text", + "content": "(a) Intuitive solution of CLIP for Face Anti-Spoofing" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 286, + 114, + 472, + 175 + ], + "blocks": [ + { + "bbox": [ + 286, + 114, + 472, + 175 + ], + "lines": [ + { + "bbox": [ + 286, + 114, + 472, + 175 + ], + "spans": [ + { + "bbox": [ + 286, + 114, + 472, + 175 + ], + "type": "image", + "image_path": "6dfca787fa20bd353a89eaca2669cb1e90e7bd5210b8a12d28402ea5b90d5fd8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 297, + 179, + 470, + 188 + ], + "lines": [ + { + "bbox": [ + 297, + 179, + 470, + 188 + ], + "spans": [ + { + "bbox": [ + 297, + 179, + 470, + 188 + ], + "type": "text", + "content": "(b) Bottom-Up Domain-Generalized Prompt Tuning for FAS (ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 198, + 482, + 264 + ], + "lines": [ + { + "bbox": [ + 130, + 198, + 482, + 264 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 482, + 264 + ], + "type": "text", + "content": "Fig. 1: The comparison between (a) the intuitive solution of CLIP for face anti-spoofing and (b) the proposed bottom-up domain-generalized prompt tuning. The Domain Token prompt guides the model to learn domain-generalized FAS features for different recording domains. The Attack Type prompt boosts the discriminability by encouraging the model to differentiate fine-grained attack instruments. The Domain Token is learned from the visual embedding through the Domain Prompt Learner." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 289, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 422 + ], + "type": "text", + "content": "learning strategy is designed to further ensure the prompt learner can learn the domain-generalized information. The various attack instruments for spoofing samples can be regarded as another perspective of domains. Different from the treatment of recording domain prompt learning, we want the visual encoder to distinguish the attack types so that it can learn more discriminative FAS features [20, 24]. In this case, we form a textual prompt that guides the generic visual encoder to adapt to the FAS task and look through the fine-grained difference for various attack types. We refer to this newly proposed method that learns the low-level domain-generalized visual prompt and tunes high-level attack-type textual prompt as the Bottom-Up Domain Prompt Tuning (BUDoPT) for generalized face anti-spoofing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 423, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 495 + ], + "type": "text", + "content": "Extensive experiments are conducted on five FAS datasets with a large number of variations (camera types, resolutions, image qualities, lighting conditions, and recording environments). The effectiveness of our proposed method is evaluated with different amounts of source domains from multiple angles, where we boost the generalizability compared with the state of the arts with multiple training datasets or with only one dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 146, + 496, + 361, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 496, + 361, + 507 + ], + "spans": [ + { + "bbox": [ + 146, + 496, + 361, + 507 + ], + "type": "text", + "content": "In sum, the main contributions of this work are:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 519, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 138, + 519, + 481, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 519, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 138, + 519, + 481, + 579 + ], + "type": "text", + "content": "- We improve the domain generalizability of FAS from a prompt tuning perspective within the context of vision-language model CLIP. A Bottom-Up Domain Prompt Tuning method (BUDoPT) that covers the different levels of domain variance, including the domain of recording settings and domain of attack types is proposed." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 580, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 580, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 138, + 580, + 481, + 628 + ], + "type": "text", + "content": "- We design 1) a context-aware adversarial domain-generalized prompt learning strategy that can learn domain-invariant prompt for different recording settings, 2) an effective textual prompt that guides CLIP to adapt to FAS task and look through the subtle details of different attack types." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 137, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 629, + 481, + 665 + ], + "type": "text", + "content": "- We improve the generalizability compared with the state of the arts with different scales of training domains, and give in-depth analysis with extensive experiments on five FAS datasets that cover typical variations in FAS." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 155, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 155, + 91, + 448, + 103 + ], + "type": "text", + "content": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 285 + ], + "type": "text", + "content": "Face Anti-Spoofing. Face anti-spoofing has attracted increasing attention in the last decades, developed from the early handcrafted liveness feature construction [37, 60] to the latest deep learning-based design [64]. Early works focus on searching the discriminative liveness cues such as the appearance artifacts [31, 42, 58], and the motion difference between genuine faces and static attack instruments. Accordingly, hand-crafted features, such as SIFT [41], HoG [60], and LBP [3] are used to characterize spoof patterns. More discriminative features can be learned with Deep learning-based [38, 59]. Auxiliary tasks, such as the regression of facial depth map [32], 3D structure [12], and material [61], encourage the network to learn fine-grained appearance details. There are also classification-based methods [34, 51, 59] and Generative models [22, 33, 57]. Spatiotemporal information can also be learned through CNN and LSTM framework [9, 39]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 293, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 402 + ], + "type": "text", + "content": "Generalized FAS. Generalized FAS that can perform robustly on different domains are necessary for real-world application scenarios. Recent works proposed to learn the FAS model on multiple source domains and evaluated a target domain. Under the domain adaptation settings where the target domain is accessible, aligning the source and target feature distributions is the typical solution [21,52,53,65,71]. Adversarial learning strategies or the triplet loss encourage the model to learn shared feature for generalization [20,47,57]. Meta-learning and continue learning approaches are also proposed by [5,6,27,44,48] to simulate the domain shift during training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 411, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 579 + ], + "type": "text", + "content": "Prompt Learning. Prompt learning originated from the field of Natural Language Processing (NLP) motives to shape the output of pre-trained language models by providing contextual information or constraints, thereby enhancing their ability to adapt to downstream tasks [43]. Prompt engineering was extended from NLP to Visual-Language Models [45], such as CLIP which classifies images by matching them with class prompts. CoOp [69] is designed to turn the handcrafted textual prompt into a learnable visual prompt. CoCoOp [68] incorporates the image as the condition for prompts and improves the model's generalization capabilities. Within the visual-textual contrastive learning framework, generalizability can be improved with a smaller data scale. CFPL1 [25] and this work aims to improve the generalizability for FAS with prompt learning. Different from CFPL which regards style (in AdaIN [18]) as the domain-specific cue, in this work we learn the domain-invariant prompt from an intrinsic CLIP-style contrastive learning perspective." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 597, + 331, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 331, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 331, + 611 + ], + "type": "text", + "content": "3 Preliminary on CLIP for FAS" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 622, + 481, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 622, + 481, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 622, + 481, + 647 + ], + "type": "text", + "content": "CLIP [45] comprises an image encoder denoted " + }, + { + "bbox": [ + 130, + 622, + 481, + 647 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 130, + 622, + 481, + 647 + ], + "type": "text", + "content": " and a text encoder denoted " + }, + { + "bbox": [ + 130, + 622, + 481, + 647 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 130, + 622, + 481, + 647 + ], + "type": "text", + "content": ". In the context of the FAS, the two class text prompt " + }, + { + "bbox": [ + 130, + 622, + 481, + 647 + ], + "type": "inline_equation", + "content": "p_{k}" + }, + { + "bbox": [ + 130, + 622, + 481, + 647 + ], + "type": "text", + "content": " encompasses the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "S.Q. Liu et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 421, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 421, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 421, + 665 + ], + "type": "text", + "content": "1 CFPL was published after this manuscript was submitted for review." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "descriptions of real and spoof faces, representing two classes. CLIP determines the similarity between the input image " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": " instead of directly training the classification head. The model predicts the image's category " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": " by selecting the class prompt with the highest response:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 249, + 175, + 481, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 175, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 249, + 175, + 481, + 194 + ], + "type": "interline_equation", + "content": "\\hat {y} = \\underset {k \\in \\{\\mathrm {r}, \\mathrm {s} \\}} {\\operatorname {a r g m a x}} \\left\\langle f _ {v} (\\boldsymbol {x}), f _ {t} \\left(\\boldsymbol {p} _ {k}\\right) \\right\\rangle \\tag {1}", + "image_path": "3c6b7eb30a5111a067c152e92ab127f72cb5539d0910c01bf9770b3f6ef08dff.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 203, + 480, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 203, + 480, + 227 + ], + "spans": [ + { + "bbox": [ + 130, + 203, + 480, + 227 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 203, + 480, + 227 + ], + "type": "inline_equation", + "content": "\\langle ,\\rangle" + }, + { + "bbox": [ + 130, + 203, + 480, + 227 + ], + "type": "text", + "content": " denotes the cosine similarity. " + }, + { + "bbox": [ + 130, + 203, + 480, + 227 + ], + "type": "inline_equation", + "content": "p_r" + }, + { + "bbox": [ + 130, + 203, + 480, + 227 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 203, + 480, + 227 + ], + "type": "inline_equation", + "content": "p_s" + }, + { + "bbox": [ + 130, + 203, + 480, + 227 + ], + "type": "text", + "content": " represent the prompt with real and spoof face descriptions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 228, + 481, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 481, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 481, + 288 + ], + "type": "text", + "content": "With such learning framework, pretrained CLIP exhibits great performance in FAS with the standard finetuning process [49]. However, since CLIP is trained for generic tasks with image pairs from the internet, when handling the low-level feature in cross-domain settings in FAS, the domain discrepancy of lighting conditions, camera qualities, or attack types can hardly be well addressed." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 306, + 259, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 306, + 259, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 306, + 259, + 319 + ], + "type": "text", + "content": "4 Proposed method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 331, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 475 + ], + "type": "text", + "content": "Overview. The Bottom-Up Domain Prompt Tuning method (BUDoPT) is proposed to improve the generalization ability of FAS in the context of the vision language model CLIP. On top of the intuitive CLIP-based face anti-spoofing which turns the class label into textual prompt, we incorporate the domain token for the domain discrepancies from different recording settings and the attack type token that describes the fine-grained types of attacks. First, the context-aware adversarial domain-generalized prompt learning is to train the domain prompt learner given the input of facial visual feature. A CLIP score style domain prompt alignment mechanism and an adversarial learning strategy are designed to learn a domain-generalized prompt. Second, the textual prompt of fine-grained attack types aims to boost discriminability by looking through the subtle details of different attack instruments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 492, + 430, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 492, + 430, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 492, + 430, + 504 + ], + "type": "text", + "content": "4.1 Context-aware Domain Generalized Prompt Learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 512, + 482, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 482, + 607 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 482, + 607 + ], + "type": "text", + "content": "Recording Domain Generalized Prompt Learning. Although text prompt can effectively assist in classifying concrete instances in upper-stream tasks, it encounters difficulties in describing the recording domain with various complex factors, including camera qualities, lighting conditions, and blurriness in FAS. Therefore we proposed the to learn domain-generalized prompt from the visual embedding. Inspired by [68], we employ vision domain features as conditional inputs to generate learnable prompts. The primary objective of the component is learning domain-invariant prompts to mitigate the impact of domain gaps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 608, + 482, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 608, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 608, + 482, + 643 + ], + "type": "text", + "content": "Specifically, we propose a Prompt learner " + }, + { + "bbox": [ + 130, + 608, + 482, + 643 + ], + "type": "inline_equation", + "content": "PL" + }, + { + "bbox": [ + 130, + 608, + 482, + 643 + ], + "type": "text", + "content": " to obtain the domain prompt from the image " + }, + { + "bbox": [ + 130, + 608, + 482, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 608, + 482, + 643 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 608, + 482, + 643 + ], + "type": "inline_equation", + "content": "PL" + }, + { + "bbox": [ + 130, + 608, + 482, + 643 + ], + "type": "text", + "content": " utilizes the features from " + }, + { + "bbox": [ + 130, + 608, + 482, + 643 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 608, + 482, + 643 + ], + "type": "text", + "content": "-th layer of " + }, + { + "bbox": [ + 130, + 608, + 482, + 643 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 130, + 608, + 482, + 643 + ], + "type": "text", + "content": " as the input. The domain feature " + }, + { + "bbox": [ + 130, + 608, + 482, + 643 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 608, + 482, + 643 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 262, + 654, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 654, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 262, + 654, + 481, + 666 + ], + "type": "interline_equation", + "content": "D (\\boldsymbol {x}) = P L \\left(f _ {v, i} (\\boldsymbol {x})\\right), \\tag {2}", + "image_path": "453580577ebd23d6ea7ebf6cc537e11e466174e58f02532a5a5835fccc1fc3a7.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 155, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 155, + 91, + 448, + 103 + ], + "type": "text", + "content": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 471, + 312 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 471, + 312 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 471, + 312 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 471, + 312 + ], + "type": "image", + "image_path": "9b496ce7dbef0d00662e55612e172142ea398b9c5e9f237694769ceff3d568a9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 320, + 482, + 389 + ], + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 389 + ], + "type": "text", + "content": "Fig. 2: The proposed Bottom-Up Domain Prompt Tuning method (BUDoPT) covers the different levels of domain variance for generalized FAS, including the domain of recording settings and domain of attack types. The context-aware adversarial domain-generalized prompt learning can learn domain-invariant prompt for different recording settings. The textual prompt of fine-grained attack types can boost the discriminability by looking through the subtle details of different attack instruments." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 411, + 309, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 309, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 309, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 411, + 309, + 422 + ], + "type": "inline_equation", + "content": "f_{v,i}" + }, + { + "bbox": [ + 130, + 411, + 309, + 422 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 130, + 411, + 309, + 422 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 411, + 309, + 422 + ], + "type": "text", + "content": "-layer features of " + }, + { + "bbox": [ + 130, + 411, + 309, + 422 + ], + "type": "inline_equation", + "content": "f_v" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 422, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 472 + ], + "type": "text", + "content": "To accurately represent the domain feature while minimizing influences from the content feature of a single image, we utilize all images' domain features in a mini-batch to generate the domain prompt. Given the training mini-batch " + }, + { + "bbox": [ + 130, + 422, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\{\\pmb{x}_n\\}_{n=1}^N" + }, + { + "bbox": [ + 130, + 422, + 482, + 472 + ], + "type": "text", + "content": " for domain " + }, + { + "bbox": [ + 130, + 422, + 482, + 472 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 422, + 482, + 472 + ], + "type": "text", + "content": ", the domain prompt " + }, + { + "bbox": [ + 130, + 422, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{D_m}" + }, + { + "bbox": [ + 130, + 422, + 482, + 472 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 247, + 479, + 481, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 479, + 481, + 513 + ], + "spans": [ + { + "bbox": [ + 247, + 479, + 481, + 513 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} _ {D _ {m}} = f _ {t} \\left(\\frac {1}{N} \\sum_ {n = 1} ^ {N} D \\left(\\boldsymbol {x} _ {n} ^ {m}\\right)\\right), \\tag {3}", + "image_path": "584612f4d6543ef986b3211aada0d1adb45e8b7274e01c9eba5ce509940d4e4d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 520, + 447, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 447, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 447, + 532 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 520, + 447, + 532 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 520, + 447, + 532 + ], + "type": "text", + "content": " is the mini-batch size and " + }, + { + "bbox": [ + 130, + 520, + 447, + 532 + ], + "type": "inline_equation", + "content": "x_{n}^{m}" + }, + { + "bbox": [ + 130, + 520, + 447, + 532 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 130, + 520, + 447, + 532 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 520, + 447, + 532 + ], + "type": "text", + "content": "-th image from " + }, + { + "bbox": [ + 130, + 520, + 447, + 532 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 520, + 447, + 532 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 532, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 532, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 532, + 482, + 593 + ], + "type": "text", + "content": "Inspired by the CLIP style contrastive learning objective, we design a CLIP score-based domain generalization learning strategy that pulls together image embeddings with a domain-invariant prompt and pushes away the image embeddings with their domain-specific prompt. Given a batch of training data with " + }, + { + "bbox": [ + 130, + 532, + 482, + 593 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 532, + 482, + 593 + ], + "type": "text", + "content": " domains, the generalized domain prompt " + }, + { + "bbox": [ + 130, + 532, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{GD}" + }, + { + "bbox": [ + 130, + 532, + 482, + 593 + ], + "type": "text", + "content": " can be defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 228, + 601, + 481, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 601, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 228, + 601, + 481, + 634 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} _ {G D} = f _ {t} \\left(\\frac {1}{N * M} \\sum_ {m = 1} ^ {M} \\sum_ {n = 1} ^ {N} D \\left(\\boldsymbol {x} _ {n} ^ {m}\\right)\\right), \\tag {4}", + "image_path": "7342a2a09ee06b8f9c801eba714ab4b6bc8659c9dceff188cb736bf271ecca22.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "As the Figure 3 shows, only the labels belonging to the generalized domain can be marked as positive while other domain-specific prompts are always labeled" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 225, + 102 + ], + "type": "text", + "content": "S.Q. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 462, + 187 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 462, + 187 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 462, + 187 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 462, + 187 + ], + "type": "image", + "image_path": "90a415da0d07628e0f3f4cde1ba8e925ba73509daa935f6f5ad413bffdb4275a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 195, + 482, + 229 + ], + "lines": [ + { + "bbox": [ + 131, + 195, + 482, + 229 + ], + "spans": [ + { + "bbox": [ + 131, + 195, + 482, + 229 + ], + "type": "text", + "content": "Fig. 3: Illustration of the CLIP score-based domain generalization learning strategy, where we pull together image embeddings with a domain-generalized prompt (▲) and push away the image embeddings with their domain-specific prompt" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 250, + 479, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 250, + 479, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 479, + 274 + ], + "type": "text", + "content": "as negative. We denote " + }, + { + "bbox": [ + 130, + 250, + 479, + 274 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 130, + 250, + 479, + 274 + ], + "type": "text", + "content": " as the set of prompts and then calculate the image cross-entropy loss " + }, + { + "bbox": [ + 130, + 250, + 479, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 130, + 250, + 479, + 274 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 184, + 280, + 481, + 314 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 280, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 184, + 280, + 481, + 314 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c e} = - \\sum_ {z = 1} ^ {2 \\times (M + 1)} y \\log \\left(\\frac {\\exp \\left(\\left\\langle f _ {v} (\\boldsymbol {x}) , f _ {t} \\left(\\boldsymbol {p} _ {z}\\right) \\right\\rangle\\right)}{\\sum_ {k = 1} ^ {2 \\times (M + 1)} \\exp \\left(\\left\\langle f _ {v} (\\boldsymbol {x}), f _ {t} \\left(\\boldsymbol {p} _ {k}\\right) \\right\\rangle\\right)}\\right), \\tag {5}", + "image_path": "99178d12f9a6c5aa29bb746b1081aff3ef35b1267a401cf1bf31088783cbb68c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 319, + 479, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 479, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 479, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 319, + 479, + 357 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 319, + 479, + 357 + ], + "type": "text", + "content": " is the ground truth label, " + }, + { + "bbox": [ + 130, + 319, + 479, + 357 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 319, + 479, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 319, + 479, + 357 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 319, + 479, + 357 + ], + "type": "text", + "content": " denote the indices in " + }, + { + "bbox": [ + 130, + 319, + 479, + 357 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 130, + 319, + 479, + 357 + ], + "type": "text", + "content": ". Note that we have " + }, + { + "bbox": [ + 130, + 319, + 479, + 357 + ], + "type": "inline_equation", + "content": "2 \\times (M + 1)" + }, + { + "bbox": [ + 130, + 319, + 479, + 357 + ], + "type": "text", + "content": " classes in total, i.e., " + }, + { + "bbox": [ + 130, + 319, + 479, + 357 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 319, + 479, + 357 + ], + "type": "text", + "content": " domains plus " + }, + { + "bbox": [ + 130, + 319, + 479, + 357 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{GD}" + }, + { + "bbox": [ + 130, + 319, + 479, + 357 + ], + "type": "text", + "content": " for each class since the prompt set " + }, + { + "bbox": [ + 130, + 319, + 479, + 357 + ], + "type": "inline_equation", + "content": "\\pmb{p} = \\{\\pmb{p}_{D_1}^{real}, \\dots, \\pmb{p}_{D_M}^{real}, \\pmb{p}_{GD}^{real}, \\pmb{p}_{D_1}^{fake}, \\dots, \\pmb{p}_{D_M}^{fake}, \\pmb{p}_{GD}^{fake}\\}" + }, + { + "bbox": [ + 130, + 319, + 479, + 357 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 357, + 479, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 479, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 479, + 392 + ], + "type": "text", + "content": "During inference, we assume that all samples in the batch are sourced from the same domain and possess the identical prompt. The model selects the class prompt with the highest similarity score as the final prediction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 399, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 518 + ], + "type": "text", + "content": "Context-aware Domain Clue. To better learn the domain-generalized prompt, we try to extract more domain-specific information from the visual embedding as it is treated as the negative label to be pushed away. Considering the discrepancies of recording domain are low-level features that scatter in the entire image, we extract domain-specific information from both the facial content and background context regions. Specifically, we want to highlight the context information from the surrounding background regions with purer domain information without disturbance of facial content. Consequently, we classify the output patches of " + }, + { + "bbox": [ + 130, + 399, + 482, + 518 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 130, + 399, + 482, + 518 + ], + "type": "text", + "content": " into two types: context patches " + }, + { + "bbox": [ + 130, + 399, + 482, + 518 + ], + "type": "inline_equation", + "content": "c^{ctx}" + }, + { + "bbox": [ + 130, + 399, + 482, + 518 + ], + "type": "text", + "content": " and content patches " + }, + { + "bbox": [ + 130, + 399, + 482, + 518 + ], + "type": "inline_equation", + "content": "c^{ctn}" + }, + { + "bbox": [ + 130, + 399, + 482, + 518 + ], + "type": "text", + "content": ". Then the Equation (2) can be rewritten as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 253, + 524, + 481, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 524, + 481, + 539 + ], + "spans": [ + { + "bbox": [ + 253, + 524, + 481, + 539 + ], + "type": "interline_equation", + "content": "D (x) = P L ([ c _ {i} ^ {c t x}, c _ {i} ^ {c t n} ]), \\tag {6}", + "image_path": "cb6712439c9ffe7812068f021aebdcd4a8df801ed62958d64714d1826dc89a5a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 544, + 482, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 627 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 627 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 544, + 482, + 627 + ], + "type": "inline_equation", + "content": "c_{i}^{ctx}" + }, + { + "bbox": [ + 130, + 544, + 482, + 627 + ], + "type": "text", + "content": " denotes the context patch obtained from " + }, + { + "bbox": [ + 130, + 544, + 482, + 627 + ], + "type": "inline_equation", + "content": "f_{v,i}" + }, + { + "bbox": [ + 130, + 544, + 482, + 627 + ], + "type": "text", + "content": ". To enhance the representation of domain features while reducing the influence of facial content, we extract context patches " + }, + { + "bbox": [ + 130, + 544, + 482, + 627 + ], + "type": "inline_equation", + "content": "c_{s}^{ctx}" + }, + { + "bbox": [ + 130, + 544, + 482, + 627 + ], + "type": "text", + "content": " from a shallower layer (" + }, + { + "bbox": [ + 130, + 544, + 482, + 627 + ], + "type": "inline_equation", + "content": "s < i" + }, + { + "bbox": [ + 130, + 544, + 482, + 627 + ], + "type": "text", + "content": ") of " + }, + { + "bbox": [ + 130, + 544, + 482, + 627 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 130, + 544, + 482, + 627 + ], + "type": "text", + "content": " and replace the original " + }, + { + "bbox": [ + 130, + 544, + 482, + 627 + ], + "type": "inline_equation", + "content": "c_{i}^{ctx}" + }, + { + "bbox": [ + 130, + 544, + 482, + 627 + ], + "type": "text", + "content": ", as shown in Figure 2(green block). For the VIT visual encoder in CLIP, the context patches in shallower layers preserve low-level recording domain information as they have less interaction with content patches. Hence, the domain feature is formed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 253, + 633, + 481, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 633, + 481, + 647 + ], + "spans": [ + { + "bbox": [ + 253, + 633, + 481, + 647 + ], + "type": "interline_equation", + "content": "D (x) = P L \\left(\\left[ c _ {s} ^ {c t x}, c _ {i} ^ {c t n} \\right]\\right), \\tag {7}", + "image_path": "41e3a501d5fc486f1e90a904b3dc489a336fd0128b2748d4f619a21430b5097d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 653, + 403, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 653, + 403, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 653, + 403, + 666 + ], + "type": "text", + "content": "to emphasize the domain-specific information in the prompts." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 155, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 155, + 91, + 448, + 103 + ], + "type": "text", + "content": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "text", + "content": "Adversarial Domain Prompt Learning. We propose the Adversarial Domain Prompt Learning strategy to further ensure that the prompt learner can learn the domain-generalized information. A Discriminator " + }, + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "inline_equation", + "content": "DC" + }, + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "text", + "content": " is constructed to differentiate the domain features and determine their belonging domain. " + }, + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "inline_equation", + "content": "PL" + }, + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "text", + "content": ", as the generator, is required to minimize the domain-specific trace in the output to fool " + }, + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "inline_equation", + "content": "DC" + }, + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 206, + 195, + 481, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 195, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 206, + 195, + 481, + 228 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a d v} = - \\frac {1}{N * M} \\sum_ {m = 1} ^ {M} \\sum_ {n = 1} ^ {N} y _ {m} \\log \\left(D C \\left(D \\left(\\boldsymbol {x} _ {n} ^ {m}\\right)\\right)\\right), \\tag {8}", + "image_path": "59ac306ee0b3d873d9f9f66b9748f56e2b9d4b9b0bf87246babc494a37949cfb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 235, + 482, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 235, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 131, + 235, + 482, + 320 + ], + "type": "inline_equation", + "content": "y_{m}" + }, + { + "bbox": [ + 131, + 235, + 482, + 320 + ], + "type": "text", + "content": " is the domain label. " + }, + { + "bbox": [ + 131, + 235, + 482, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adv}" + }, + { + "bbox": [ + 131, + 235, + 482, + 320 + ], + "type": "text", + "content": " encourages discriminator " + }, + { + "bbox": [ + 131, + 235, + 482, + 320 + ], + "type": "inline_equation", + "content": "DC" + }, + { + "bbox": [ + 131, + 235, + 482, + 320 + ], + "type": "text", + "content": " to try its best to find the domain-specific vestige so that domain prompt learner can learn the domain-invariant information. Empirically, the adversarial learning strategy aggravates the unstable loss convergence. To ensure a more consistent learning direction, we employ a distribution constraint between domain features themselves. The Jensen-Shannon divergence is adopted to pull every two features from " + }, + { + "bbox": [ + 131, + 235, + 482, + 320 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 235, + 482, + 320 + ], + "type": "text", + "content": " domains together in a training batch:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 217, + 327, + 481, + 355 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 327, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 217, + 327, + 481, + 355 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d i s} = \\sum_ {j, t = 1, \\dots , N * M, j < t} \\operatorname {J S} \\left(D \\left(\\boldsymbol {x} _ {j}\\right), D \\left(\\boldsymbol {x} _ {t}\\right)\\right). \\tag {9}", + "image_path": "b3775b1393e309e6f6dd3aac4ef7f2b05b43e51ca591fada5cd0d541f90d0bfa.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 362, + 374, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 362, + 374, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 362, + 374, + 374 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 132, + 362, + 374, + 374 + ], + "type": "inline_equation", + "content": "N * M" + }, + { + "bbox": [ + 132, + 362, + 374, + 374 + ], + "type": "text", + "content": " is the total number of samples in a batch." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 392, + 408, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 392, + 408, + 404 + ], + "spans": [ + { + "bbox": [ + 131, + 392, + 408, + 404 + ], + "type": "text", + "content": "4.2 Spoofing Domain Discriminative Prompt Tuning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 411, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 554 + ], + "type": "text", + "content": "In contrast to the visual recording domain generalized prompt that aims to avoid bias, the attack types of spoofing domain can be utilized to boost the discriminability. Specific artifacts are highly correlated with the attack instruments, which can be regarded as spoofing clues, such as the saturation change in printing attacks, the moiré pattern in video replay attacks, and the edge defects of 3D masks. Different from the low-level recording domain prompt, the attack types can be clearly described with the (high-level) textual prompt. Regularize the visual embedding with finer grained attack types text helps the visual encoder learn more discriminative FAS features. This will also set clearer boundaries between genuine faces and spoofing attacks thereby enhancing model generalizability to unseen attack types. Therefore, we propose the Spoof Domain Prompt tuning strategy to encourage the model to differentiate finer-grained attack types." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 555, + 482, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 555, + 482, + 627 + ], + "spans": [ + { + "bbox": [ + 131, + 555, + 482, + 627 + ], + "type": "text", + "content": "Specifically, we design a textual prompt template, a structured sentence that describes spoofing face photos with a placeholder of the attack types, such as \"a {fake} face photo {and it is the printing attack}\" Given " + }, + { + "bbox": [ + 131, + 555, + 482, + 627 + ], + "type": "inline_equation", + "content": "N * M" + }, + { + "bbox": [ + 131, + 555, + 482, + 627 + ], + "type": "text", + "content": " samples in a training batch, we generate the same amount of text prompts " + }, + { + "bbox": [ + 131, + 555, + 482, + 627 + ], + "type": "inline_equation", + "content": "\\boldsymbol{p}\\boldsymbol{a}" + }, + { + "bbox": [ + 131, + 555, + 482, + 627 + ], + "type": "text", + "content": ". Spoofing Domain Discriminative Prompt Tuning updates the parameters of " + }, + { + "bbox": [ + 131, + 555, + 482, + 627 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 131, + 555, + 482, + 627 + ], + "type": "text", + "content": " with the text cross-entropy loss " + }, + { + "bbox": [ + 131, + 555, + 482, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 131, + 555, + 482, + 627 + ], + "type": "text", + "content": ", which is calculated by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 201, + 636, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 636, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 201, + 636, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t} = - \\sum_ {o = 1} ^ {N * M} o \\log \\left(\\frac {\\exp \\left(\\left\\langle f _ {v} (\\boldsymbol {x}) , f _ {t} \\left(\\boldsymbol {p} a _ {o}\\right) \\right\\rangle\\right)}{\\sum_ {w = 1} ^ {N * M} \\exp \\left(\\left\\langle f _ {v} (\\boldsymbol {x}) , f _ {t} \\left(\\boldsymbol {p} a _ {w}\\right) \\right\\rangle\\right)}\\right), \\tag {10}", + "image_path": "782ab7a6210c97335a7c15eaa779487a1fe62adb5875c40742f203add7544786.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "S.Q. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 295, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 295, + 128 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 116, + 295, + 128 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 132, + 116, + 295, + 128 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 116, + 295, + 128 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 132, + 116, + 295, + 128 + ], + "type": "text", + "content": " are the indices in " + }, + { + "bbox": [ + 132, + 116, + 295, + 128 + ], + "type": "inline_equation", + "content": "\\pmb{pa}" + }, + { + "bbox": [ + 132, + 116, + 295, + 128 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 129, + 370, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 129, + 370, + 140 + ], + "spans": [ + { + "bbox": [ + 146, + 129, + 370, + 140 + ], + "type": "text", + "content": "The final loss function of our method is formed as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 234, + 154, + 480, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 154, + 480, + 166 + ], + "spans": [ + { + "bbox": [ + 234, + 154, + 480, + 166 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\alpha \\mathcal {L} _ {c e} + \\beta \\mathcal {L} _ {t} - \\gamma \\mathcal {L} _ {a d v} + \\delta \\mathcal {L} _ {d i s}. \\tag {11}", + "image_path": "16b4de3384304d11d3b15017e5af938c90dc36823603146b33ea447b52398390.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 178, + 394, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 178, + 394, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 394, + 190 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 178, + 394, + 190 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta, \\gamma, \\delta" + }, + { + "bbox": [ + 132, + 178, + 394, + 190 + ], + "type": "text", + "content": " are hyperparameters that balance each loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 214, + 230, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 230, + 227 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 230, + 227 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 245, + 266, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 245, + 266, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 245, + 266, + 257 + ], + "type": "text", + "content": "5.1 Experimental Setups" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 270, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 365 + ], + "type": "text", + "content": "Databases and Protocols. We evaluate our method with six widely used FAS databases: MSU-MFSD(M) [58], CASIA-FASD(C) [67], Idiap Replay attack (I) [8], OULU-NPU(O) [2], WMCA(W) [10] and CelebA-Sproof [66]. Five protocols that cover various domain generalization settings with different amounts of source domains are involved to comprehensively evaluate our method's performance. They are leave-one-out(P1) and one-to-one(P2) cross-domain generalization evaluation, different domain partition evaluation(P3), unseen attack type evaluation(P4) and various devices domain evaluation(P5)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 366, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 426 + ], + "type": "text", + "content": "Metrics. Following [49], we evaluate the model performance using three metrics: Half Total Error Rate (HTER), Area Under the Receiver Operating Characteristic Curve (AUC) and True Positive Rate (TPR) at a False Positive Rate (FPR) " + }, + { + "bbox": [ + 130, + 366, + 482, + 426 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 366, + 482, + 426 + ], + "type": "text", + "content": " (TPR@FPR=1%). HTER and AUC reflect the theoretical performance. TPR@FPR=1% evaluates the model performance in practice." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "text", + "content": "Implementation Details. We utilize FAN [4] to obtain facial landmarks, aligning the images based on nose and eye landmarks before cropping them to a size of 224x224. We employ the pre-trained CLIP-ViT-B-16 [45] as the foundation model. An eleven-layer VIT following a two-layer MLP is adopted as the prompt learner " + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "PL" + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "text", + "content": ". We extract content patches from the 5th layer of " + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "text", + "content": " and context patches from the 1st layer. For adversarial learning an eight-layer 1D ResNet is adopted as the discriminator " + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "DC" + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Training strategy. We divide the training process into two stages for the Bottom-Up domain prompt tuning. In the first training stage, we conduct the Spoofing Domain Discriminative Prompt Tuning to boost the discriminability of " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ". Adam optimizer is adopted with a learning rate of " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ". In the second training stage, we freeze " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " and train the " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "PL" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " for the Recording Domain Generalization Prompt Learning. The optimizer for both " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "PL" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "DC" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " are Adam with learning rate " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "3 * 10^{-4}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ". Before the adversarial learning, we first train at least one epoch for each domain source to warm up(exclude supplementary training data). Then, we take turns freezing the " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "PL" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "DC" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " for 50 iterations. For each stage we train 4000 iterations to ensure the largest dataset is traversed. The mini-batch size is " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "4(N = 4)" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " for each domain. " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " varies according to the number of datasets used for training under different evaluation protocols. The hyperparameter " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " are set as 1, 1, 1, and 0.5, respectively." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 155, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 155, + 91, + 448, + 103 + ], + "type": "text", + "content": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 143, + 481, + 344 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 126 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 126 + ], + "type": "text", + "content": "Table 1: Evaluation of cross-domain performance under leave-one-out protocol (P1)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 143, + 481, + 344 + ], + "lines": [ + { + "bbox": [ + 132, + 143, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 132, + 143, + 481, + 344 + ], + "type": "table", + "html": "
MethodsOCI→MOMI→COCM→IICM→OAvg.
HTERAUCTPR@ FPR=1%HTERAUCTPR@ FPR=1%HTERAUCTPR@ FPR=1%HTERAUCTPR@ FPR=1%
MADDG [47]17.6988.06-24.5084.51-22.1984.99-27.9880.02-23.09
MDDR [53]17.0290.10-19.6887.43-20.8786.72-25.02581.47-20.64
NAS-FAS [64]16.8590.42-15.2192.64-11.6396.98-13.1694.18-14.21
RFMeta [48]13.8993.98-20.2788.16-17.3090.48-16.4591.16-16.97
D²AM [7]12.7095.6-20.9885.58-15.4391.22-15.2790.87-16.09
DRDG [28]12.4395.81-19.0588.79-15.5691.79-15.6391.75-15.66
Self-DA [55]15.4091.80-24.5084.40-15.6090.10-23.1084.30-19.65
ANRL [27]10.8396.75-17.8589.26-16.0391.04-15.6791.90-15.09
FGHV [26]9.1796.92-12.4793.47-16.2990.11-13.5893.55-12.87
SSDG-R [20]7.3897.17-10.4495.94-11.7196.59-15.6191.54-11.28
SSAB-R [57]6.6798.75-10.0096.67-8.8896.79-13.7293.65-9.80
PatchNet [51]7.1098.46-11.3394.58-13.4095.67-11.8295.07-10.90
GDA [71]9.2098.00-12.2093.00-10.0096.00-14.4092.60-11.45
LDCformer [16]6.4398.39-8.1196.67-8.5797.09-11.1795.58-8.57
TTDG-V [70]4.1698.48-7.5998.18-9.6298.18-1096.15-7.84
BUDoPT(ours)0.9599.7087.122.8598.0355.004.498.5486.462.2698.7855.252.62
ViT†[15]1.5899.6896.675.7098.9788.579.2597.1551.547.4798.4269.306.00
FLIP-V†[49]3.7999.3187.991.2799.7595.854.7198.8075.844.1598.7666.473.48
FLIP-IT†[49]5.2798.4279.330.4499.9899.862.9499.4284.623.6199.1584.763.06
FLIP-MCL†[49]4.9598.1174.670.5499.98100.004.2599.0784.622.3199.6392.283.01
BUDoPT-CoOp†1.9899.5480.631.9399.5888.063.1899.390.43.3398.9189.292.61
BUDoPT-CoCoOp†1.5199.4375.551.5599.5684.743.5399.292.533.5198.9887.982.53
BUDoPT(ours)†0.4099.9999.670.2699.9699.921.3899.6998.11.6099.5197.180.91
", + "image_path": "37af7debb983da7c646e5a954cf72f3989ff97f547882201bd367ced3fbb18d1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 173, + 345, + 443, + 354 + ], + "lines": [ + { + "bbox": [ + 173, + 345, + 443, + 354 + ], + "spans": [ + { + "bbox": [ + 173, + 345, + 443, + 354 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 173, + 345, + 443, + 354 + ], + "type": "text", + "content": " indicates CelebA-Sproof [66] is employed as supplementary training dataset" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 396, + 215, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 215, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 215, + 407 + ], + "type": "text", + "content": "5.2 Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 414, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 481, + 569 + ], + "type": "text", + "content": "Cross-domain generalization evaluation. Following [49], the leave-one-out (P1) and one-to-one (P2) on M, C, I and O are conducted to evaluate cross-dataset performance. " + }, + { + "bbox": [ + 130, + 414, + 481, + 569 + ], + "type": "inline_equation", + "content": "\\mathrm{OCM} \\rightarrow \\mathrm{I}" + }, + { + "bbox": [ + 130, + 414, + 481, + 569 + ], + "type": "text", + "content": " refers to the model trains on O, C, M and tests on I. For a fair comparison with [49] and other models, we evaluate the performance of our method with and without the supplementary training data CelebA-Sproof [66]. For simplicity, we regard dataset as the domain partition in our proposed method. As shown in Table 1, without supplementary training data, the proposed BUDoPT outperforms SOTA in all settings. With supplementary training data, BUDoPT also surpasses the FLIP family model in HTER for all settings and achieves more than " + }, + { + "bbox": [ + 130, + 414, + 481, + 569 + ], + "type": "inline_equation", + "content": "69\\%" + }, + { + "bbox": [ + 130, + 414, + 481, + 569 + ], + "type": "text", + "content": " improvement on average. In terms of TPR@FPR " + }, + { + "bbox": [ + 130, + 414, + 481, + 569 + ], + "type": "inline_equation", + "content": "= 1\\%" + }, + { + "bbox": [ + 130, + 414, + 481, + 569 + ], + "type": "text", + "content": ", our method performs better in three out of four settings. Although the performance slightly drops in TPR@FPR " + }, + { + "bbox": [ + 130, + 414, + 481, + 569 + ], + "type": "inline_equation", + "content": "= 1\\% (\\mathrm{C})" + }, + { + "bbox": [ + 130, + 414, + 481, + 569 + ], + "type": "text", + "content": " and AUC(C,O), the margin is relatively small as they already exceed " + }, + { + "bbox": [ + 130, + 414, + 481, + 569 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 130, + 414, + 481, + 569 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": "In P2 of low-data regime [49] the model is trained separately on M, C, I and O on the remaining datasets, resulting in a total of 12 different scenario combinations. Since the model is trained on one dataset only, domain partition is based on capturing devices(C,M), environments(I), and sessions(O) for different datasets. We also test our method with and without supplementary CelebA-Sproof. CelebA-Sproof is regarded as a domain if it is involved. As illustrated in Table 2, without help of CelebA-Sproof, BUDoPT performs better than SOTA in nine out of twelve settings. We also outperform domain adap" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "S.Q. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 144, + 481, + 262 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 475, + 126 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 475, + 126 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 475, + 126 + ], + "type": "text", + "content": "Table 2: Evaluation of cross-domain performance under one-to-one protocol (P2)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 144, + 481, + 262 + ], + "lines": [ + { + "bbox": [ + 132, + 144, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 144, + 481, + 262 + ], + "type": "table", + "html": "
MethodsUnseenC → IC → MC → OI → CI → MOM → CM → IM → OO → CO → IO → MAvg.
ADDA [50]No41.836.6-49.835.1-39.035.2----39.6
DRCN [11]No44.427.4-48.942.0-28.936.8----38.1
DupGAN [14]No42.133.4-46.536.2-27.135.4----36.8
KSA [23]No39.315.1-12.333.3-9.134.9----24.0
DR-UDA [54]No15.69.028.734.229.03.516.83.030.219.525.427.423.1
ADA [52]No17.59.329.141.530.539.617.75.131.219.826.831.525.0
USDAN-Un [21]No16.09.2-30.225.8-13.33.4----16.3
GDA [71]No15.105.8-29.720.8-12.22.5----14.4
CDFTN-L [65]No1.78.129.911.99.629.98.81.325.619.15.86.313.2
BUDoPT(ours)Yes9.654.376.768.265.647.736.072.96.164.812.972.945.69
FLIP-V†[49]Yes15.0813.7312.344.309.687.870.563.964.792.095.016.007.12
FLIP-IT†[49]Yes12.3315.187.981.128.376.980.195.214.960.164.275.636.03
FLIP-MCL†[49]Yes10.577.153.910.687.224.220.195.883.950.195.698.404.84
BUDoPT(ours)†Yes4.332.624.980.481.834.1402.451.870.442.531.432.46
", + "image_path": "f68470aed5e821ad9dbed19fa7d16356d9567a83bb1ad998cd0941639254541e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 186, + 263, + 430, + 272 + ], + "lines": [ + { + "bbox": [ + 186, + 263, + 430, + 272 + ], + "spans": [ + { + "bbox": [ + 186, + 263, + 430, + 272 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 186, + 263, + 430, + 272 + ], + "type": "text", + "content": " indicates CelebA-Sproof is employed as supplementary training data." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 316, + 480, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 480, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 480, + 340 + ], + "type": "text", + "content": "tive models (Unseen=No) even without any information of target domain. With CelebA-Sproof, BUDoPT surpasses FLIP family models in ten settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 353, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 482, + 472 + ], + "type": "text", + "content": "Effect of Domain Partition. We set " + }, + { + "bbox": [ + 130, + 353, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\mathrm{MI} \\rightarrow \\mathrm{O}" + }, + { + "bbox": [ + 130, + 353, + 482, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 353, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\mathrm{MI} \\rightarrow \\mathrm{C}" + }, + { + "bbox": [ + 130, + 353, + 482, + 472 + ], + "type": "text", + "content": " to evaluate our model with different amounts of training data sources. Here we treat dataset as domain. Results in Table 3 show that with less training data, BUDoPT outperforms competitors by a larger margin. However, we also note that comparing with the " + }, + { + "bbox": [ + 130, + 353, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\mathrm{M} \\rightarrow \\mathrm{O}" + }, + { + "bbox": [ + 130, + 353, + 482, + 472 + ], + "type": "text", + "content": " results in P2 the performance increase is limited. We believe this is due to the difference in domain partition strategies. For " + }, + { + "bbox": [ + 130, + 353, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\mathrm{M} \\rightarrow \\mathrm{O}" + }, + { + "bbox": [ + 130, + 353, + 482, + 472 + ], + "type": "text", + "content": " the domain is defined based on the device while for " + }, + { + "bbox": [ + 130, + 353, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\mathrm{MI} \\rightarrow \\mathrm{O}" + }, + { + "bbox": [ + 130, + 353, + 482, + 472 + ], + "type": "text", + "content": " it is based on dataset. The results indicate that a fine-grained partition strategy may help achieve better performance for our method even with a limited data scale. More analysis of partition strategy is in supplementary Sec 2.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 485, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 570 + ], + "type": "text", + "content": "Generalize to Unseen attack type. We employ " + }, + { + "bbox": [ + 130, + 485, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 130, + 485, + 482, + 570 + ], + "type": "text", + "content": " with high-quality 3D silicon mask attack samples to evaluate the model generalizability to unseen attack types (P4). Different from the baselines trained on IMCO, we only utilize " + }, + { + "bbox": [ + 130, + 485, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 130, + 485, + 482, + 570 + ], + "type": "text", + "content": " as the source dataset. Same domain partition strategy in P2 is adopted. As shown in Table 4, with less training data BUDoPT still outperforms the others. This result not only demonstrates our generalizability to the unseen attack types but also shows the learning efficiency of our method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Device Variance Only. In P5, we utilize the O's intra protocol 3 and protocol 4 [2] to evaluate the performance with device variance only. The domain gaps in P5 are located on the unseen device settings and parameters and other underlying factors are similar. Same domain partition strategy in P2 is adopted. As shown in Table 5, under constraint setting BUDoPT can handle the domain gap of devices well. This indicates the usability in real applications where some local samples can be available for fine-tuning." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 155, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 155, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 146, + 291, + 235 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 291, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 291, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 291, + 137 + ], + "type": "text", + "content": "Table 3: Evaluation of different domain partition (P3)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 146, + 291, + 235 + ], + "lines": [ + { + "bbox": [ + 132, + 146, + 291, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 146, + 291, + 235 + ], + "type": "table", + "html": "
MethodsMI → CMI → O
HTERAUCHTERAUC
SSAN-R [57]25.5683.8924.4482.86
HFN+MP [6]30.8972.4820.9486.71
CIFAS [35]22.6783.8924.6381.48
DiVT-M [24]20.1186.7123.6185.73
SSDG-R [20]19.8686.4627.9278.72
DGUA-FAS [13]19.2286.8120.0588.75
BUDoPT(ours)5.3398.925.9498.37
", + "image_path": "acf5ac2bd119ed2b0e4456b6169b258d48e0bd8456656f1d0e1f4b2c69791e08.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 328, + 146, + 476, + 217 + ], + "blocks": [ + { + "bbox": [ + 322, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 322, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 322, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 4: Evaluation of unseen attack type performance (P4)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 146, + 476, + 217 + ], + "lines": [ + { + "bbox": [ + 328, + 146, + 476, + 217 + ], + "spans": [ + { + "bbox": [ + 328, + 146, + 476, + 217 + ], + "type": "table", + "html": "
MethodsIMCO → WMCA
HTERAUC
DiVT-M [24]22.3686.82
DGUA-FAS [13]20.6288.07
C → WMCA
BUDoPT(ours)9.2095.93
", + "image_path": "8167bbbea841be6070670685a00a75d7d6a391bf94f9739635433567664bc1b7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 132, + 304, + 291, + 399 + ], + "blocks": [ + { + "bbox": [ + 132, + 272, + 291, + 294 + ], + "lines": [ + { + "bbox": [ + 132, + 272, + 291, + 294 + ], + "spans": [ + { + "bbox": [ + 132, + 272, + 291, + 294 + ], + "type": "text", + "content": "Table 5: Evaluation of Device Variance Only (P5) on O. P refers protocol" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 304, + 291, + 399 + ], + "lines": [ + { + "bbox": [ + 132, + 304, + 291, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 291, + 399 + ], + "type": "table", + "html": "
MethodsPACPERBCPERACER
DC-CDN [63]32.2±2.81.6±2.11.9±1.1
RAEDFL [17]31.38±1.780.28±0.680.83±0.86
TranFAS [56]30.6±0.71.1±2.50.9±1.1
LDCFormer [16]32.35±2.050.28±0.681.31±1.03
BUDoPT(ours)30±00±00±0
DC-CDN [63]45.4±3.32.5±4.24.0±3.1
RAEDFL [17]45.41±6.402.50±2.743.96±3.90
TranFAS [56]42.1±2.23.8±3.52.9±2.4
LDCFormer [16]41.08±1.281.17±1.941.13±1.02
BUDoPT(ours)40±00±00±0
", + "image_path": "39a98a64a92a85b1326296e46b760e2517fee4432959f0f513d9cce7c6be3d81.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 326, + 304, + 477, + 373 + ], + "blocks": [ + { + "bbox": [ + 322, + 272, + 480, + 294 + ], + "lines": [ + { + "bbox": [ + 322, + 272, + 480, + 294 + ], + "spans": [ + { + "bbox": [ + 322, + 272, + 480, + 294 + ], + "type": "text", + "content": "Table 6: Ablation study of major components of the proposed BUDoPT" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 304, + 477, + 373 + ], + "lines": [ + { + "bbox": [ + 326, + 304, + 477, + 373 + ], + "spans": [ + { + "bbox": [ + 326, + 304, + 477, + 373 + ], + "type": "table", + "html": "
SDSDDRDGADCDCICM → O
✓----3.51
-✓---5.78
✓-✓--3.01
✓-✓✓-2.50
✓-✓✓✓2.26
", + "image_path": "720cb935fe98a72f8b5e1cc04745018e9f24ac591c6ad9b155450f81bbf60022.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 442, + 246, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 246, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 246, + 453 + ], + "type": "text", + "content": "5.3 Ablation Studies" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 472, + 481, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 481, + 508 + ], + "type": "text", + "content": "We analyze the effectiveness of major components(Table 6), test the intuitive textual recording domain prompt tuning(Table 6), and compared with generic visual prompt tuning CoOp [69], CoCoOp [68](Table 1)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "text", + "content": "Table 6 reports the results of the ablation studies of our major components. We set Spoofing Domain Discriminative Prompt Tuning(SD) as the baseline and evaluate performance with and without Recording Domain Generalization Prompt Learning (RDG), Adversarial Domain Learning(AD) and Context-aware Domain Clue(CDC). With Recording Domain Generalization Prompt Learning, the HTER is improved by " + }, + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "text", + "content": ". With Adversarial Domain Learning, the HTER decreases by around " + }, + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "inline_equation", + "content": "17\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "text", + "content": ". The Context-aware Domain Clue contributes around " + }, + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 617 + ], + "type": "text", + "content": " improvement. The ablation studies indicate that each component within our method is crucial for improving the domain generalization ability." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Intuitive textual recording domain prompt tuning. Considering the success of employing context information as a prompt in upper-stream tasks, we evaluate an intuitive domain prompt tuning strategy, i.e., using the textual domain description as the prompt. We mark the textual domain prompts with the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "S.Q. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 116, + 480, + 252 + ], + "blocks": [ + { + "bbox": [ + 135, + 116, + 480, + 252 + ], + "lines": [ + { + "bbox": [ + 135, + 116, + 480, + 252 + ], + "spans": [ + { + "bbox": [ + 135, + 116, + 480, + 252 + ], + "type": "image", + "image_path": "f81e81fb30622b2a61c02f1227aee00a55c968ebf485b4387670203becc4f325.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 261, + 482, + 285 + ], + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 285 + ], + "type": "text", + "content": "Fig. 4: The t-SNE visualizations of original visual embedding (first row) and the learned domain-invariant prompt embedding (second row) under P1." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 308, + 482, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 367 + ], + "type": "text", + "content": "Spoofing Domain Discriminative Prompt Tuning as SDD2. As shown in Table 6, adding the intuitive textual recording domain prompt (SDD) harms the model's performance. This aligns with our analysis that the textual domain prompt of context encourages the model to extract more domain-specific information and contradicts the goal of domain-generalized FAS." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 368, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 475 + ], + "type": "text", + "content": "Generic visual prompt tuning. Since the proposed method follows the idea of visual prompt tuning, the classical visual prompt learning methods: CoOp and CoCoOp are compared. For a fair comparison, the CLIP visual encoder of the two models is finetuned with the Spoofing Domain Discriminative Prompt Learning module in our framework so we regard them as BUDoPT-CoOp and BUDoPT-CoCoOp. As the results shown in Table 1, by comparing with the FLIP-family models, BUDoPT-CoOp, and BUDoPT-CoCoOp, we comprehensively verify the effectiveness of our proposed method that adapts CLIP to FAS with better generalization ability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 475, + 481, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 481, + 500 + ], + "type": "text", + "content": "We also conduct hyperparameter analysis of loss function. Results are summarized in Eq. (11) in the supplementary Section 2.2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 517, + 231, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 517, + 231, + 529 + ], + "spans": [ + { + "bbox": [ + 131, + 517, + 231, + 529 + ], + "type": "text", + "content": "5.4 Visualizations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 538, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 538, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 538, + 482, + 647 + ], + "type": "text", + "content": "Domain Prompt Visualization. Figure 4 shows the t-SNE [36] visualization of original visual embedding (first row) and the learned domain-invariant prompt embedding (second row) under P1. The original visual embedding naturally forms multiple clusters of domain-specific information. Each cluster shares some low-level underlying factors, such as the camera qualities (see supplementary material for details). Our BUDoPT learns a domain-generalized prompt embedding where samples from different domains are grouped together in a single cluster, thereby guiding the model to minimize the influences of domain gaps. Attention Visualization. We utilize the GradCam [46] to visualize attention" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 155, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 155, + 91, + 448, + 103 + ], + "type": "text", + "content": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 345, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 345, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 345, + 665 + ], + "type": "text", + "content": "2 Please refer to supplementary material for details" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 112, + 222, + 203 + ], + "blocks": [ + { + "bbox": [ + 134, + 112, + 222, + 203 + ], + "lines": [ + { + "bbox": [ + 134, + 112, + 222, + 203 + ], + "spans": [ + { + "bbox": [ + 134, + 112, + 222, + 203 + ], + "type": "image", + "image_path": "ea9e1286e03706a17b62f75e42ccd2ae27fa58fe1e7e780a2492e397a300f9ae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 210, + 443, + 222 + ], + "lines": [ + { + "bbox": [ + 170, + 210, + 443, + 222 + ], + "spans": [ + { + "bbox": [ + 170, + 210, + 443, + 222 + ], + "type": "text", + "content": "Fig. 5: Typical attention maps of real and fake samples under P1." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 228, + 113, + 309, + 204 + ], + "blocks": [ + { + "bbox": [ + 228, + 113, + 309, + 204 + ], + "lines": [ + { + "bbox": [ + 228, + 113, + 309, + 204 + ], + "spans": [ + { + "bbox": [ + 228, + 113, + 309, + 204 + ], + "type": "image", + "image_path": "13948c45e91ce012c3b2fc4aae866a1b3f37457b3e5801b836cde11a51cc8343.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 314, + 113, + 395, + 203 + ], + "blocks": [ + { + "bbox": [ + 314, + 113, + 395, + 203 + ], + "lines": [ + { + "bbox": [ + 314, + 113, + 395, + 203 + ], + "spans": [ + { + "bbox": [ + 314, + 113, + 395, + 203 + ], + "type": "image", + "image_path": "46d5899b36ace9e81d326b04129d87f6a4f6c228bf197a47658c5e88ad66c5d2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 400, + 113, + 481, + 204 + ], + "blocks": [ + { + "bbox": [ + 400, + 113, + 481, + 204 + ], + "lines": [ + { + "bbox": [ + 400, + 113, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 400, + 113, + 481, + 204 + ], + "type": "image", + "image_path": "7ec57ffabc1d59b0bb99eb0b60c9e903e1c609f9b2da4654cc4bbb9a5de02cda.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 331 + ], + "type": "text", + "content": "maps for genuine and spoofing samples in P1. Typical results are visualized in Figure 5. We can observe that BUDoPT takes different attention for live and spoof faces. Specifically, the model focuses on face regions for real samples while highlights spoof clues of fake faces. With the help of Spoofing Domain Discriminative Prompt Tuning, the model is able to detect the intrinsic spoofing clues for different types of attack, including moiré patterns(M, O), paper mask edges(C) and paper texture(I)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 354, + 220, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 220, + 367 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 220, + 367 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 384, + 482, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 576 + ], + "type": "text", + "content": "In this work, we aim to improve the generalization ability of CLIP-based FAS from a prompt learning perspective. A Bottom-Up Domain Prompt Tuning method (BUDoPT) that covers the different levels of domain variance, including the domain of recording settings and domain of attack types is proposed. It contains a context-aware adversarial domain-generalized prompt learning strategy that can learn domain-invariant prompt for different recording settings and a fine-grained textual prompt that guides CLIP to look through the subtle details of different attack instruments. Extensive experiments are conducted on five FAS datasets with a large number of variations (camera types, resolutions, image qualities, lighting conditions, and recording environments). The effectiveness of our proposed method is evaluated with different amounts of source domains from multiple angles, where we boost the generalizability compared with the state of the arts with multiple training datasets or with only one dataset. We also observe the limitations of our methods with different domain partitioning strategies. Designing an adaptive domain partition approach is a potential way to further improve the generalizability of our method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 599, + 261, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 261, + 613 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 261, + 613 + ], + "type": "text", + "content": "7 Acknowledgement" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "This work is supported by GuangDong Basic and Applied Basic Research Foundation 2023A1515110706, RGC-NSFC joint project N_HKBU212/23, and NSFC Young Scientist Fund Application No. 6240075068." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "S.Q. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "text", + "content": "1. Bao, W., Li, H., Li, N., Jiang, W.: A liveness detection method for face recognition based on optical flow field. In: IASP (2009)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "type": "text", + "content": "2. Boulkenafet, Z., Komulainen, J., Li, L., Feng, X., Hadid, A.: OULU-NPU: A mobile face presentation attack database with real-world variations. In: FG (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 184, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 480, + 215 + ], + "type": "text", + "content": "3. Boulkenafet, Z., Komulainen, J., Hadid, A.: Face spoofing detection using colour texture analysis. IEEE Trans. on Information Forensics and Security 11(8), 1818-1830 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 217, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 480, + 248 + ], + "type": "text", + "content": "4. Bulat, A., Tzimiropoulos, G.: How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In: International Conference on Computer Vision (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 249, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 249, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 249, + 480, + 282 + ], + "type": "text", + "content": "5. Cai, R., Cui, Y., Li, Z., Yu, Z., Li, H., Hu, Y., Kot, A.: Rehearsal-free domain continual face anti-spoofing: Generalize more and forget less. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8037-8048 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 283, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 283, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 138, + 283, + 480, + 314 + ], + "type": "text", + "content": "6. Cai, R., Li, Z., Wan, R., Li, H., Hu, Y., Kot, A.C.: Learning meta pattern for face anti-spoofing. IEEE Transactions on Information Forensics and Security 17, 1201-1213 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 315, + 480, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 315, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 315, + 480, + 348 + ], + "type": "text", + "content": "7. Chen, Z., Yao, T., Sheng, K., Ding, S., Tai, Y., Li, J., Huang, F., Jin, X.: Generalizable representation learning for mixture domain face anti-spoofing. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 1132-1139 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 349, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 349, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 349, + 480, + 369 + ], + "type": "text", + "content": "8. Chingovska, I., Anjos, A., Marcel, S.: On the effectiveness of local binary patterns in face anti-spoofing. In: BIOSIG. pp. 1-7 (2012)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 370, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 480, + 391 + ], + "type": "text", + "content": "9. Ge, H., Tu, X., Ai, W., Luo, Y., Ma, Z., Xie, M.: Face anti-spoofing by the enhancement of temporal motion. In: CTISC. pp. 106-111. IEEE (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 392, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 424 + ], + "type": "text", + "content": "0. George, A., Mostaani, Z., Geissenbuhler, D., Nikisins, O., Anjos, A., Marcel, S.: Biometric face presentation attack detection with multi-channel convolutional neural network. IEEE Trans. on Information Forensics and Security (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 425, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 480, + 468 + ], + "type": "text", + "content": "1. Ghifary, M., Kleijn, W.B., Zhang, M., Balduzzi, D., Li, W.: Deep reconstruction-classification networks for unsupervised domain adaptation. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14. pp. 597-613. Springer (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 469, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 469, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 469, + 480, + 490 + ], + "type": "text", + "content": "2. Guo, J., Zhu, X., Xiao, J., Lei, Z., Wan, G., Li, S.Z.: Improving face anti-spoofing by 3d virtual synthesis. arXiv (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 522 + ], + "type": "text", + "content": "3. Hong, Z.W., Lin, Y.C., Liu, H.T., Yeh, Y.R., Chen, C.S.: Domain-generalized face anti-spoofing with unknown attacks. In: 2023 IEEE International Conference on Image Processing (ICIP). pp. 820-824. IEEE (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "text", + "content": "4. Hu, L., Kan, M., Shan, S., Chen, X.: Duplex generative adversarial network for unsupervised domain adaptation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1498-1507 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 556, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 480, + 589 + ], + "type": "text", + "content": "5. Huang, H.P., Sun, D., Liu, Y., Chu, W.S., Xiao, T., Yuan, J., Adam, H., Yang, M.H.: Adaptive transformers for robust few-shot cross-domain face anti-spoofing. In: European Conference on Computer Vision. pp. 37-54. Springer (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "type": "text", + "content": "6. Huang, P.K., Chiang, C.H., Chong, J.X., Chen, T.H., Ni, H.Y., Hsu, C.T.: Ldcformer: Incorporating learnable descriptive convolution to vision transformer for face anti-spoofing. In: 2023 IEEE International Conference on Image Processing (ICIP). pp. 121-125. IEEE (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "7. Huang, P.K., Chin, M.C., Hsu, C.T.: Face anti-spoofing via robust auxiliary estimation and discriminative feature learning. In: Asian Conference on Pattern Recognition. pp. 443-458. Springer (2021)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 155, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 155, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "18. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: Proceedings of the IEEE international conference on computer vision. pp. 1501-1510 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "19. Jee, H.K., Jung, S.U., Yoo, J.H.: Liveness detection for embedded face recognition system. International Journal of Biological and Medical Sciences 1(4), 235-238 (2006)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 480, + 216 + ], + "type": "text", + "content": "20. Jia, Y., Zhang, J., Shan, S., Chen, X.: Single-side domain generalization for face anti-spoofing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8484-8493 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "text", + "content": "21. Jia, Y., Zhang, J., Shan, S., Chen, X.: Unified unsupervised and semi-supervised domain adaptation network for cross-scenario face anti-spoofing. Pattern Recognition 115, 107888 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "22. Jourabloo, A., Liu, Y., Liu, X.: Face de-spoofing: Anti-spoofing via noise modeling. In: ECCV. pp. 290-306 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "text", + "content": "23. Li, H., Li, W., Cao, H., Wang, S., Huang, F., Kot, A.C.: Unsupervised domain adaptation for face anti-spoofing. IEEE Transactions on Information Forensics and Security 13(7), 1794-1809 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 346 + ], + "type": "text", + "content": "24. Liao, C.H., Chen, W.C., Liu, H.T., Yeh, Y.R., Hu, M.C., Chen, C.S.: Domain invariant vision transformer learning for face anti-spoofing. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6098-6107 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 390 + ], + "type": "text", + "content": "25. Liu, A., Xue, S., Gan, J., Wan, J., Liang, Y., Deng, J., Escalera, S., Lei, Z.: Cfplfas: Class free prompt learning for generalizable face anti-spoofing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 222-232 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "text", + "content": "26. Liu, S., Lu, S., Xu, H., Yang, J., Ding, S., Ma, L.: Feature generation and hypothesis verification for reliable face anti-spoofing. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 1782-1791 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 480, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 467 + ], + "type": "text", + "content": "27. Liu, S., Zhang, K.Y., Yao, T., Bi, M., Ding, S., Li, J., Huang, F., Ma, L.: Adaptive normalized representation learning for generalizable face anti-spoofing. In: Proceedings of the 29th ACM international conference on multimedia. pp. 1469-1477 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "text", + "content": "28. Liu, S., Zhang, K.Y., Yao, T., Sheng, K., Ding, S., Tai, Y., Li, J., Xie, Y., Ma, L.: Dual reweighting domain generalization for face presentation attack detection. arXiv preprint arXiv:2106.16128 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "text", + "content": "29. Liu, S.Q., Lan, X., Yuen, P.C.: Multi-channel remote photoplethysmography correspondence feature for 3d mask face presentation attack detection. IEEE Trans. on Information Forensics and Security 16, 2683-2696 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "30. Liu, S.Q., Lan, X., Yuen, P.C.: Learning temporal similarity of remote photoplethysmography for fast 3d mask face presentation attack detection. IEEE Transactions on Information Forensics and Security 17, 3195-3210 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 589 + ], + "type": "text", + "content": "31. Liu, Y., Tai, Y., Li, J., Ding, S., Wang, C., Huang, F., Li, D., Qi, W., Ji, R.: Aurora guard: Real-time face anti-spoofing via light reflection. arXiv (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 590, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 610 + ], + "type": "text", + "content": "32. Liu, Y., Jourabloo, A., Liu, X.: Learning deep models for face anti-spoofing: Binary or auxiliary supervision. In: CVPR. pp. 389-398 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "33. Liu, Y., Liu, X.: Spoof trace disentanglement for generic face anti-spoofing. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(3), 3813-3830 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "34. Liu, Y., Stehouwer, J., Jourabloo, A., Liu, X.: Deep tree learning for zero-shot face anti-spoofing. In: CVPR (2019)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "S.Q. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "35. Liu, Y., Chen, Y., Dai, W., Li, C., Zou, J., Xiong, H.: Causal intervention for generalizable face anti-spoofing. In: 2022 IEEE International Conference on Multimedia and Expo (ICME). pp. 01-06. IEEE (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "text", + "content": "36. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "37. Määttra, J., Hadid, A., Pietikainen, M.: Face spoofing detection from single images using micro-texture analysis. In: IJCB. pp. 1-7 (2011)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "38. Menotti, D., Chiachia, G., Pinto, A., Schwartz, W.R., Pedrini, H., Falcao, A.X., Rocha, A.: Deep representations for iris, face, and fingerprint spoofing detection. IEEE Trans. on Information Forensics and Security 10(4), 864-879 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "39. Muhammad, U., Holmberg, T., de Melo, W.C., Hadid, A.: Face anti-spoofing via sample learning based recurrent neural network (rnn). In: BMVC. p. 113 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "40. Pan, G., Sun, L., Wu, Z., Lao, S.: Eyeblink-based anti-spoofing in face recognition from a generic webcam camera. In: ICCV. pp. 1-8 (2007)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "text", + "content": "41. Patel, K., Han, H., Jain, A.K.: Secure face unlock: Spoof detection on smartphones. IEEE transactions on information forensics and security 11(10), 2268-2283 (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "42. Patel, K., Han, H., Jain, A.K., Ott, G.: Live face video vs. spoof face video: Use of moiré patterns to detect replay video attacks. In: ICB. pp. 98-105 (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "43. Petroni, F., Rocktäschel, T., Lewis, P., Bakhtin, A., Wu, Y., Miller, A.H., Riedel, S.: Language models as knowledge bases? arXiv preprint arXiv:1909.01066 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "44. Qin, Y., Yu, Z., Yan, L., Wang, Z., Zhao, C., Lei, Z.: Meta-teacher for face antispoofing. IEEE transactions on pattern analysis and machine intelligence 44(10), 6311-6326 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 369, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 413 + ], + "type": "text", + "content": "45. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 414, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 456 + ], + "type": "text", + "content": "46. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618-626 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 457, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 500 + ], + "type": "text", + "content": "47. Shao, R., Lan, X., Li, J., Yuen, P.C.: Multi-adversarial discriminative deep domain generalization for face presentation attack detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10023-10031 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "text", + "content": "48. Shao, R., Lan, X., Yuen, P.C.: Regularized fine-grained meta face anti-spoofing. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 11974-11981 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "49. Srivatsan, K., Naseer, M., Nandakumar, K.: Flip: Cross-domain face anti-spoofing with language guidance. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19685-19696 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "50. Tzeng, E., Hoffman, J., Saenko, K., Darrell, T.: Adversarial discriminative domain adaptation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7167-7176 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "51. Wang, C.Y., Lu, Y.D., Yang, S.T., Lai, S.H.: Patchnet: A simple face anti-spoofing framework via fine-grained patch recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20281-20290 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "52. Wang, G., Han, H., Shan, S., Chen, X.: Improving cross-database face presentation attack detection via adversarial domain adaptation. In: 2019 International Conference on Biometrics (ICB). pp. 1-8. IEEE (2019)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 155, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 155, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 160 + ], + "type": "text", + "content": "53. Wang, G., Han, H., Shan, S., Chen, X.: Cross-domain face presentation attack detection via multi-domain disentangled representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6678-6687 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "text", + "content": "54. Wang, G., Han, H., Shan, S., Chen, X.: Unsupervised adversarial domain adaptation for cross-domain face presentation attack detection. IEEE Transactions on Information Forensics and Security 16, 56-69 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "55. Wang, J., Zhang, J., Bian, Y., Cai, Y., Wang, C., Pu, S.: Self-domain adaptation for face anti-spoofing. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 2746-2754 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "text", + "content": "56. Wang, Z., Wang, Q., Deng, W., Guo, G.: Face anti-spoofing using transformers with relation-aware mechanism. IEEE Transactions on Biometrics, Behavior, and Identity Science 4(3), 439-450 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "type": "text", + "content": "57. Wang, Z., Wang, Z., Yu, Z., Deng, W., Li, J., Gao, T., Wang, Z.: Domain generalization via shuffled style assembly for face anti-spoofing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4123-4133 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 303, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 481, + 326 + ], + "type": "text", + "content": "58. Wen, D., Han, H., Jain, A.K.: Face spoof detection with image distortion analysis. IEEE Trans. on Information Forensics and Security 10(4), 746-761 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "text", + "content": "59. Yang, J., Lei, Z., Li, S.Z.: Learn convolutional neural network for face anti-spoofing. arXiv preprint arXiv:1408.5601 (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "60. Yang, J., Lei, Z., Liao, S., Li, S.Z.: Face liveness detection with component dependent descriptor. In: 2013 International Conference on Biometrics (ICB). pp. 1-6. IEEE (2013)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 403 + ], + "type": "text", + "content": "61. Yu, Z., Li, X., Niu, X., Shi, J., Zhao, G.: Face anti-spoofing with human material perception. In: ECCV. pp. 557-575. Springer (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "text", + "content": "62. Yu, Z., Qin, Y., Li, X., Zhao, C., Lei, Z., Zhao, G.: Deep learning for face antispoofing: A survey. IEEE transactions on pattern analysis and machine intelligence 45(5), 5609-5631 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 457 + ], + "type": "text", + "content": "63. Yu, Z., Qin, Y., Zhao, H., Li, X., Zhao, G.: Dual-cross central difference network for face anti-spoofing. arXiv preprint arXiv:2105.01290 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "text", + "content": "64. Yu, Z., Wan, J., Qin, Y., Li, X., Li, S.Z., Zhao, G.: Nas-fas: Static-dynamic central difference network search for face anti-spoofing. IEEE Trans. on Pattern Analysis and Machine Intelligence (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 490, + 481, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 523 + ], + "type": "text", + "content": "65. Yue, H., Wang, K., Zhang, G., Feng, H., Han, J., Ding, E., Wang, J.: Cyclically disentangled feature translation for face anti-spoofing. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 3358-3366 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "text", + "content": "66. Zhang, Y., Yin, Z., Li, Y., Yin, G., Yan, J., Shao, J., Liu, Z.: Celeba-spoof: Large-scale face anti-spoofing dataset with rich annotations. In: European Conference on Computer Vision (ECCV) (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 578 + ], + "type": "text", + "content": "67. Zhang, Z., Yan, J., Liu, S., Lei, Z., Yi, D., Li, S.Z.: A face antispoofing database with diverse attacks. In: ICB (2012)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "text", + "content": "68. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 611, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 633 + ], + "type": "text", + "content": "69. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "text", + "content": "70. Zhou, Q., Zhang, K.Y., Yao, T., Lu, X., Ding, S., Ma, L.: Test-time domain generalization for face anti-spoofing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 175-187 (2024)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "S.Q. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 482, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 482, + 150 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 482, + 150 + ], + "type": "text", + "content": "71. Zhou, Q., Zhang, K.Y., Yao, T., Yi, R., Sheng, K., Ding, S., Ma, L.: Generative domain adaptation for face anti-spoofing. In: European Conference on Computer Vision. pp. 335-356. Springer (2022)" + } + ] + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 155, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 155, + 91, + 447, + 103 + ], + "type": "text", + "content": "Bottom-Up Domain Prompt Tuning for Generalized Face Anti-Spoofing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Brain Netflix_ Scaling Data to Reconstruct Videos from Brain Signals/9af6f00c-d99e-4633-82c3-bc4c9c80afb6_content_list.json b/2024/Brain Netflix_ Scaling Data to Reconstruct Videos from Brain Signals/9af6f00c-d99e-4633-82c3-bc4c9c80afb6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f9de38dc2f49349c16595e20ed5141fea7546e67 --- /dev/null +++ b/2024/Brain Netflix_ Scaling Data to Reconstruct Videos from Brain Signals/9af6f00c-d99e-4633-82c3-bc4c9c80afb6_content_list.json @@ -0,0 +1,1366 @@ +[ + { + "type": "text", + "text": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals", + "text_level": 1, + "bbox": [ + 256, + 140, + 746, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Camilo Fosco\\*1, Benjamin Lahner\\*1, Bowen Pan $^{1}$ , Alex Andonian $^{1}$ , Emilie Josephs $^{1}$ , Alex Lascelles $^{1}$ , and Aude Oliva $^{1}$", + "bbox": [ + 228, + 212, + 771, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "CSAIL, Massachusetts Institute of Technology, MA, USA [camilolu, blahner, bpan, andonian, ejosephs, alexlasc, oliva]@mit.edu", + "bbox": [ + 228, + 255, + 772, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The field of brain-to-stimuli reconstruction has seen significant progress in the last few years, but techniques continue to be subject-specific and are usually tested on a single dataset. In this work, we present a novel technique to reconstruct videos from functional Magnetic Resonance Imaging (fMRI) signals designed for performance across datasets and across human participants. Our pipeline accurately generates 2 and 3-second video clips from brain activity coming from distinct participants and different datasets by leveraging multi-dataset and multi-subject training. This helps us regress key latent and conditioning vectors for pretrained text-to-video and video-to-video models to reconstruct accurate videos that match the original stimuli observed by the participant. Key to our pipeline is the introduction of a 3-stage approach that first aligns fMRI signals to semantic embeddings, then regresses important vectors, and finally generates videos with those estimations. Our method demonstrates state-of-the-art reconstruction capabilities verified by qualitative and quantitative analyses, including crowd-sourced human evaluation. We showcase performance improvements across two datasets, as well as in multi-subject setups. Our ablation studies shed light on how different alignment strategies and data scaling decisions impact reconstruction performance, and we hint at a future for zero-shot reconstruction by analyzing how performance evolves as more subject data is leveraged.", + "bbox": [ + 261, + 330, + 743, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 666, + 375, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The human brain processes several million bits of information per second and compresses most of this information in ways that are still, in great part, unknown. What if we could partially recover this information by reading out brain signals? Reconstructing visual stimuli from brain activity is a promising way to investigate what information content is encoded in the human brain and how it may differ between participants. Reconstruction efforts typically require large amounts of neural-stimuli pairs to train high-parameter computational models, and inter-subject analyses additionally require repeated experiments across subjects. However, these neural-stimuli pairs are often not abundantly available due", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to their expensive and time-consuming collection process. Reconstructing dynamic videos, as opposed to static images, poses the additional challenges of temporal continuity.", + "bbox": [ + 212, + 146, + 785, + 191 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous fMRI (Functional Magnetic Resonance Imaging) work has shown that a greater extent of cortex responds to videos than still images [53] [38] [2], that some cortical regions are selective for motion features [2] [17] [27], and that inter-subject correlations to movie viewings are high [16] [13], all suggesting that fMRI reliably encodes dynamic content. Humans do not all perceive videos in the same way, however; in-the-wild longform videos, such as an unedited video of the happenings in a park, show markedly lower inter-subject correlations than edited cinematic films [12] [13]. These higher inter-subject correlations in cinematic films are largely due to production effects, such as camera angle, editing, and lighting, that structure everyone's attention and interpretation of the content in a similar way [42]. The everyday human visual experience includes events sampled at both ends of this continuum, from unstructured meanderings down a busy street to a structured viewing of a work presentation. Thus, while videos can be reconstructed from brain activity [4], cross-subject reconstructions and the structure of the video stimulus set have yet to be systematically studied in the video reconstruction domain.", + "bbox": [ + 212, + 191, + 787, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Here, we leverage two recent computational developments (latent diffusion models (LDMs) for video and masked-brain modeling) and several large-scale fMRI datasets to achieve high fidelity reconstructions of previously viewed 2- and 3-second videos. Latent diffusion models [34] are a class of generative models that generate high-fidelity images or videos, conditioned on inputs such as text, through a denoising diffusion process over a latent representation of the pixels. Masked brain modeling (MBM), an extension of masked training methods from natural language processing and computer vision to brain data, is a technique to learn robust self-supervised representations of brain signals by masking the signal and training a encoder-decoder network to reconstruct it. It has been shown to be a powerful way for a model to learn compressed representations of fMRI brain activity [4] [3].", + "bbox": [ + 212, + 433, + 787, + 616 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We leverage these two concepts to propose a novel 3-stage pipeline that a) learns robust fMRI representations through masked brain modeling and alignment, b) regresses conditioning vectors from these representations, and c) reconstructs target videos with high fidelity, leveraging the estimated conditioning vectors. Importantly, our alignment and training strategy combines information from multiple subjects, as opposed to previous works that train one model per subject. This allows our model to not only improve its general reconstruction quality, but also to be able to generalize from one subject to another with much higher performance. Our contributions are as follows:", + "bbox": [ + 212, + 614, + 787, + 752 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce a 3-stage pipeline for high-quality video reconstruction that leverages multiple fMRI datasets and aligns inter-subject data, outperforming existing approaches.", + "2. We showcase a detailed ablation study that illustrates how the model performs in different conditions, including modeling and data variants." + ], + "bbox": [ + 222, + 763, + 782, + 838 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "C. Fosco et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. We perform multi-subject and zero-shot reconstruction experiments to explore the possibility of a subject-agnostic visual reconstruction technology.", + "bbox": [ + 220, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Together, this work offers both modeling and neuroscience insights for video reconstruction from brain activity.", + "bbox": [ + 212, + 189, + 782, + 220 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/27d56e2ed6556e21b6d50e4823ec760757794086a7c22bcfc8a56a829193f247.jpg", + "image_caption": [ + "A", + "B" + ], + "image_footnote": [], + "bbox": [ + 238, + 262, + 766, + 369 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1b2ca13a3fb458028d0bd623528670bbaa519343b5670004bb535c22a6fe67f7.jpg", + "image_caption": [ + "Fig. 1: Reconstruction of short videos from fMRI activity A) Brain responses of subjects viewing video clips was measured with fMRI and then used to reconstruct the seen video. B) We show three examples of the ground truth video (top row) and our reconstruction using our approach (bottom row)." + ], + "image_footnote": [], + "bbox": [ + 243, + 386, + 779, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 580, + 385, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models: Latent diffusion models (LDMs) [34] are a class of generative model that achieves photorealistic generations through a process of denoising latent representations. This denoising process can be conditioned on other inputs, such as text, in order to guide the model toward generating outputs with specific semantic meaning of visual appearance. These models have been successful in reconstruction tasks, including image generation, super-resolution and recoloring [6, 35-37, 43], audio generation [23], among others. Recent text-to-video latent diffusion models have generated high fidelity reconstructions of short videos faithful to both content and temporal dynamics [1]. Here, we leverage these advances in video generation to form the backbone of our framework: we use Zeroscope V2, which creates 3 second videos based on text conditioning.", + "bbox": [ + 212, + 612, + 785, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Masked Brain modeling: Masked Brain Modeling [3] (MBM) is a recent method for pre-training fMRI reconstruction models. It is based on Masked Signal Modeling (MSM), a self-supervised learning task commonly used for pretraining in natural language processing and vision applications [5,14,50,52]. In", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals", + "bbox": [ + 267, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MSM, some amount of the signal is masked during training, and an autoencoder is trained to predict the masked content, which can be fine tuned with additional training. In MBM, a percentage of the brain activity in masked and the model is tasked with reconstructing the complete input, which pushes the MBM model to learn compressed representations of fMRI brain activity.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video reconstruction from brain activity: An early naturalistic video reconstruction work used a voxelwise motion-energy model to recover videos with similar spatiotemporal energy, showing that the BOLD signal captures dynamic visual information [29]. Other approaches used convolutional neural networks (CNNs) to reconstruct videos from fMRI data [51] [21], with one even leveraging a convolutional encoder-decoder framework to synthesize high framerate fMRI-video frame pairs [18]. A work using GANs employed dual spatial and temporal discriminators to achieve spatiotemporally faithful reconstructions [47], and variational autoencoders have been used to reconstruct videos frame-by-frame from a hidden latent space [11]. Most similar to our approach, [4] uses masked-brain modeling to train a vision transformer encoder that then feeds latent representations into a diffusion model to reconstruct the video.", + "bbox": [ + 212, + 222, + 787, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Video fMRI Datasets", + "text_level": 1, + "bbox": [ + 215, + 428, + 460, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We train and evaluate our reconstruction method on a compilation of four large-scale fMRI datasets [19] [56] [45] [51] to obtain a large quantity and diversity of brain responses. In order to effectively perform inter-subject and inter-dataset analyses we analyze all fMRI responses in Cifti fsLR32k space (see data preprocessing details in the supplement.) [10] [7]. This preprocessing pipeline accurately registers gray matter voxels to a shared cortical surface mesh, thereby allowing the model to learn informative spatial activation patterns across datasets. From the cortical surface, we select a subset of 41 regions of interest (ROIs) from the Glasser Atlas (see supplement for details) [9]. This subset of cortex is isolates brain regions that might respond to dynamic stimuli [46] [24] [20] [8] [33] [40] [30] [31] [49] as well as reduce the computational load of model training. All datasets combined, our task-based video fMRI data draws from 43 subjects, over 28,100 short video segments, and over 123,000 fMRI response trials, and our resting state fMRI data totals over one thousand hours across 1084 subjects.", + "bbox": [ + 212, + 460, + 787, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human Connectome Project Dataset (HCP). The 1200-subject release of the Human Connectome Project [45] includes nearly an hour of resting state scans on 1084 subjects. During acquisition of resting state data, subjects were instructed to remain awake and fixate on a cross-hair but completed no task nor viewed any other visual stimulus. Resting state fMRI brain activity captures temporally correlated functional networks that can be useful for understanding cortical organization [41]. We utilize this time series data to pre-train our Masked Brain Modeling encoder (Figure 2).", + "bbox": [ + 212, + 674, + 787, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "BOLD Moments Dataset (BMD). The BOLD Moments Dataset (BMD) [19] consists of fMRI responses to 1,102 3-second videos for ten human subjects collected in a rapid event-related design. Beta values were estimated at each", + "bbox": [ + 214, + 794, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "C. Fosco et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "cortical vertex for each video using a general linear model (GLM). The stimuli were divided into a non-overlapping 1,000 video training set and a 102 video testing set. Each subject viewed the training set videos 3 times and the testing set videos 10 times to facilitate both inter-subject and intra-subject analyses. The 1,102 videos were sampled from the 1 million video Moments in Time dataset [25] (overlapping with the Multi-Moments in Time [26] and Memento10k datasets [28]) and capture naturalistic, in-the-wild content like home videos. We use BMD's five sentence text description metadata provided for each video to generate ground truth text embeddings that our pipeline learns to estimate. To the best of our knowledge, this work is the first time BMD has been used for video reconstruction.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Human Actions Dataset (HAD). The Human Actions Dataset (HAD) [56] is composed of 30 human subjects each viewing 720 two second videos sampled from the Human Action Clips and Segments (HACS) dataset [55]. Similar to BMD, data were collected in a rapid event-related design and beta values were estimated across cortex for each video. Each video is annotated with one of 180 human-centric actions, and each subject viewed four (different) videos from each of the 180 action categories. Video presentations were not repeated within or across subjects. For each subject, we define a testing/training split by randomly selecting one of the four videos per action category as a testing video. We use this dataset primarily to train Stage 1 in our pipeline. To the best of our knowledge, this work is the first time HAD has been used in video reconstruction.", + "bbox": [ + 212, + 311, + 787, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CC2017 Dataset (CC2017). In the CC2017 dataset [51], three subjects viewed continuous training and testing movies composed of shorter 10-15 second movie clips under a longform viewing paradigm. A training movie was composed from 374 video clips and randomly split into 18 eight minute segments. A testing movie was composed of 598 video clips and randomly split into five eight minute segments. Thus, each eight minute segment was composed of a concatenation of shorter video clips varying in duration. The clips are sourced from professionally shot videos, such as cinematic movies or stock advertising footage. The training and testing segments were repeated two and ten times per subject, respectively. Here, we divide the 8 minute segments into 2 second clips for reconstruction. We compose fMRI-video pairs by sampling the fMRI time series at an offset of 4 seconds (to account for the hemodynamic lag in the BOLD response) to capture peak BOLD activation corresponding to short snippets of the movie (2 second snippets, same as CC2017's acquisition TR). As this dataset is one of the main benchmarks used in previous work [4], we utilize this data extensively to train our regressors and provide reconstructions and evaluation metrics.", + "bbox": [ + 212, + 494, + 787, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Reconstruction Pipeline", + "text_level": 1, + "bbox": [ + 215, + 761, + 485, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Approach Overview. Our reconstruction pipeline aims to regress key latent and conditioning vectors from brain signals to guide a pretrained video generation model towards reconstructing the original stimulus. This pipeline consists of", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals", + "bbox": [ + 267, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ed4062954afc56f8fb606c03ca8236756801a8f01a1e4734036b28da16d5f687.jpg", + "image_caption": [ + "Fig. 2: Our proposed reconstruction framework. Stage 1: We train an encoder-decoder structure to reconstruct fMRI signals from a compressed representation. We then fine-tune the MBM encoder to align with a CLIP embedding of the video. Stage 2: we regress the latent vector (z) and blip embedding (b) of the target video from stage 1's aligned embedding. Stage 3: we use the regressed conditioning vectors to generate the video. The estimated vector z is artificially noised and then denoised with a pre-trained denoising U-Net, following a video-to-video procedure. The estimated vector b is sent to a pretrained BLIP model to generate a caption, which is used as a conditioning input to the U-Net." + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 772, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3 stages. First, in an initial alignment stage, an encoder-decoder architecture is trained to reconstruct masked fMRI embeddings while simultaneously following a CLIP-like contrastive learning objective. This setup forces the encoder to map fMRI inputs into an embedding that captures both knowledge about the raw signal and knowledge about the semantic properties of the original stimuli. Second, the output of stage 1's encoder is used in a regression stage: we regress the necessary inputs required by our generative model with two MLPs that estimate an initial latent vector $(z)$ and BLIP embeddings $(b)$ [22]. The BLIP embeddings are used to infer a conditioning caption. Third, our reconstruction stage uses the predicted latent vectors and BLIP embeddings to generate a video: Given a new fMRI input, we pass them through the regressors to predict the latent and conditioning vectors $(z$ and $b)$ , re-noise $z$ and use $b$ to generate a caption. We feed the noised $z$ and caption to a pretrained denoising U-Net (Zeroscope v2 [48]) to estimate the final reconstruction. We describe each stage in detail below and show a pseudocode implementation in the supplement Algorithm 1.", + "bbox": [ + 212, + 551, + 787, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Stage 1: MBM and Alignment. Our model employs a straightforward encoder-decoder framework aimed at reconstructing the input fMRI signal by compressing it into a 1024-dimensional latent vector. During this phase, the input fMRI signal is divided into several patches, and certain patches are masked.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "C. Fosco et al.", + "bbox": [ + 271, + 114, + 369, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The encoder processes the signal with masked patches to produce a sequence of 1024-dimensional latent vectors. The decoder then attempts to transform these vectors back into the fMRI signal. Throughout the training process, the encoder-decoder structure is supervised through its ability to accurately reconstruct the masked patches. This approach, which we call Masked Brain Modeling (MBM), draws inspiration from the masked image modeling task [14] typical in the self-supervision literature. Its primary purpose is to acquaint the model with the spatial structure of the fMRI signal, rather than to imbue it with any semantic understanding. We first train our encoder-decoder pipeline on the masked modeling task alone with the large-scale HCP dataset [45] and HAD dataset [56], which ensures our embeddings learn basic fMRI structure. We utilize a simple MSE loss over the masked patches.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After this step, we further finetune the MBM encoder through a contrastive learning task. Contrastive learning is a powerful self-supervised technique to learn representations across modalities. We extract CLIP embeddings from captions describing each video, and we apply a contrastive loss over batches of encoder embeddings and CLIP embeddings. This contrastive loss follows [32]: for each batch, the loss enforces high cosine similarity between an fMRI embedding and its matching CLIP embedding (positive pair), while minimizing the similarity between that same embedding and all other CLIP embeddings in the batch (negative pairs):", + "bbox": [ + 212, + 329, + 787, + 465 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n t r a s t i v e}} = - \\sum_ {i = 1} ^ {N} \\log \\left(\\frac {\\exp \\left(\\frac {f _ {i} * c _ {i}}{\\tau}\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\frac {f _ {j} * c _ {j}}{\\tau}\\right)}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 479, + 656, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, $f_{i}$ is the fMRI embedding (output from the MBM encoder), $c_{i}$ is the CLIP-text embedding, and $\\tau$ is a temperature hyperparameter. We train this model by following the training details in [3] for MBM and [39] for the alignment procedure, where we use a encoder-decoder model with patch size of 16, hidden dimension of 1024 and 24 layers in the encoder. We use a temperature of 0.9 and a $75\\%$ mask ratio following previous work [3].", + "bbox": [ + 212, + 536, + 787, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Stage 2: Regression. To map from embeddings generated by the encoding model to latent and conditioning vectors suitable for the video generation pipeline, we utilize a multi-target regression model with regularization. Our regression models, instantiated as both a Ridge regression and a regularized MLP (see section 5.2 for comparisons), map outputs from Stage 1 into a latent vector $z$ of shape (4, 15, 33, 33) and a BLIP [22] embedding $b$ of shape (226, 768). The decision to map to BLIP vectors is rooted in how video generators handle noise in their conditioning inputs: similar to [44], we found that feeding a raw caption, even if imperfect, to our denoising U-Net performed better than feeding a regressed (and thus noisy) conditioning vector (comparisons in section 5.2). We hypothesize that the performance of video generators is susceptible to noise in the conditioning vector, and thus encapsulating the regression's imperfections at the level of BLIP's inputs allows the generator to receive clean (but imprecise) text tokens, which improves performance.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals", + "bbox": [ + 267, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our MLPs are composed of an initial linear layer, followed by 3 residual blocks and a final output layer. The initial layer and the residual blocks contain 2048 units, dropout regularization with $p = 0.3$ , a GELU [15] activation and Batch Normalization. They are trained with an MSE loss and predict flattened vectors, which are then reshaped and used in Stage 3. Perfect regression at this stage equates to obtaining exact latent and conditioning vectors, which would result in a near-perfect reconstruction when fed to Stage 3.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Stage 3: Reconstruction. During video reconstruction, we extract the latent vector $z$ and BLIP embedding $b$ using the pretrained MBM encoder followed by the $z$ and $b$ regressors from stages 1 and 2. $z$ undergoes a re-noising process, followed by a denoising procedure using a pretrained U-Net that follows [48]. We renoise the latents with a strength of 0.8 and apply 40 denoising steps, following insights from previous work [44]. Concurrently, $b$ is decoded using BLIP's decoder to produce a semantically relevant caption that conditions the U-Net during the denoising process. We observed during experiments that enforcing more descriptive and less repetitive captions tended to improve the quality of BLIP outputs, so we implement decoding with a high repetition penalty of 6, a minimum length of 4 and a maximum length of 20. The integration of $z$ and $b$ helps facilitate video generation that is both visually and contextually aligned with the video represented in the original fMRI signal. Our video generation model is frozen during this process.", + "bbox": [ + 212, + 252, + 789, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Experiments and Results", + "text_level": 1, + "bbox": [ + 215, + 489, + 496, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1 Comparisons to previous work", + "text_level": 1, + "bbox": [ + 215, + 523, + 511, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To showcase the quality of Brainflix, we obtain reconstructions from our full pipeline and compare to previous fMRI-to-video techniques. We report results over BMD and CC2017.", + "bbox": [ + 212, + 551, + 785, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation Details. We use the large-scale fMRI datasets introduced in section 3 for model training and finetuning. For all datasets, videos are downsampled to 15 FPS and resized to $224 \\times 224$ . We first train our alignment stage over the masked brain reconstruction task leveraging the HCP, HAD and BMD datasets for 200 epochs with a batch size of 300. We then finetune this model with a CLIP-like alignment objective over BMD and CC2017 simultaneously for 50 epochs with a batch size of 120. This aligns the output of the MBM encoder with embeddings of the captions describing each video. The captions are obtained through human labelers for BMD and synthetically for the rest of the datasets, leveraging the EILEV [54] video-to-caption model. Training for Stage 1 is done over 6 V100 GPUs, while training for Stage 2, a smaller scale endeavor given the smaller size of the MLPs, is performed over 2 Titan RTX GPUs.", + "bbox": [ + 212, + 597, + 787, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following previous work [4, 44], we utilize 3 main evaluation metrics that aim to understand different characteristics of performance. To measure pixel-level reconstruction quality, we utilize Structural Similarity (SSIM). For semantic evaluation, we use the N-way top-k classification approach", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "C. Fosco et al.", + "bbox": [ + 271, + 114, + 369, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/09eed28a4e56c47cecef472aea1371b54072c06f8284931bb376c07df7058a8e.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1: Quantitative comparison of our reconstruction performance against previous video reconstruction methods. Our methodology achieves state of the art on most metrics. Results from Kupershmidt and Mind-video on BMD are obtained through a reimplementation, as their code is not readily available." + ], + "table_body": "
MethodsCC2017BMD
SSIMMSE2-way50-waySSIMMSE2-way50-way
Kupershmidt0.128---0.0314.5610.5140.004
Mind-Video0.186-0.8530.2020.1760.7630.7110.101
Brainflux (Ours)0.1890.6810.8710.2030.1830.6910.7670.145
Brainfix multi-subject (Ours)0.1950.6550.8880.2210.1900.6710.8160.165
", + "bbox": [ + 215, + 143, + 787, + 241 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "from [4], which measures how often the classification of a simple ImageNet classifier over the reconstruction and ground truth video match, limiting the output of the classifier to N classes. We declare a successful trial if the ground truth class is within the top-k probabilities outputted over the reconstruction, and repeat the test 100 times to report average success rates. We also report MSE results over our target latent embeddings $z$ and BLIP embeddings $b$ , to observe how close our regressed embeddings are to ground truth.", + "bbox": [ + 212, + 340, + 785, + 445 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Datasets While we train on all datasets introduced in Section 3, we evaluate our reconstruction results on two large-scale video datasets, CC2017 [51] and BMD [19], to compare our method with previous work [4] [18] [47], as well as to introduce an additional dataset, BMD, with different subjects and video stimuli. These two datasets exhibit reasonable variety, complexity and semantic diversity, which makes them strong options for evaluation exercises.", + "bbox": [ + 212, + 446, + 785, + 537 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Results. We showcase our quantitative results over BMD and CC2017 in Table 1. Results are reported for Subject 1 in both datasets. Both our single-subject and our multi-subject approaches outperform previous work. Our multi-subject model, trained on all subjects from a given dataset before predicting on one, showcases larger boosts in performance which provides evidence towards the advantages of multi-subject training. We show qualitative results in Figure 3 and 4: we observe that our model is able to reconstruct examples from BMD and CC2017 with strong structural reliability. We hypothesize that our model's emphasis on regressing an accurate latent, a component that previous approaches lack, enforces accurate structural patterns that can then lead to reliable object positioning.", + "bbox": [ + 212, + 537, + 785, + 703 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Human Evaluation. To further measure how preferable our reconstructions are, we collect qualitative evaluations of semantic fidelity using human judgments. We designed a simple web experiment where participants indicated which of 6 reconstructed videos was the best semantic match to a real video. One of the 6 videos was a reconstruction of the reference video. We calculated the percent of trials where the selected video corresponded to the reconstruction of the reference video. Table 2 reports Brainflux results for two datasets. Our best performing method reached a semantic fidelity score of $86\\%$ on BMD and $84\\%$ on CC2017. This substantially exceeds chance $(16.66\\%)$ , but remains below the", + "bbox": [ + 212, + 704, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals", + "bbox": [ + 269, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/883d668a79294f0947ffd41c6abe46ab6491301664cfd33a1457c6b4f1a448be.jpg", + "image_caption": [ + "Fig. 3: Comparison to previous works on examples from CC2017. We show the ground truth videos in the first row, our results in the second row, and previous approaches in the following rows. Our model manages to capture structural similarity in a better way than previous approaches." + ], + "image_footnote": [], + "bbox": [ + 218, + 147, + 784, + 333 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/231d08431dc66d2382428bd3894b353224e0836142bca5831af098f93fe0c6ac.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBMDCC2017
Baseline (Oracle)94.192.8
MindVideo [4]80.383.9
Brainflix (Ours)86.184.7
", + "bbox": [ + 215, + 415, + 436, + 481 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/9c4292b8e142d00f6507e150148bb31cfe6c551295755141a26a25e1849df48d.jpg", + "table_caption": [ + "Table 2: Human evaluation results. Accuracy on a 6-way retrieval task, where subjects select the ground truth video that most closely matches a reconstructed reference." + ], + "table_footnote": [], + "table_body": "
Training dataBMDCC2017
SSIMMSESSIMMSE
Subject 1 only0.1230.7210.6990.127
Subject 1 + all subs0.1290.7020.7100.131
Subject 2 only0.1200.7350.6800.122
Subject 2 + all subs0.1280.7220.7080.126
", + "bbox": [ + 468, + 412, + 789, + 512 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 3: Multi-subject vs single-subject results. We observe that training on all subjects is value-adding for both datasets and both metrics. This effect is measurable across subjects.", + "bbox": [ + 468, + 512, + 787, + 568 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "best possible performance expected from oracle reconstructions given our backbone model (Zeroscope V2 [48]), which ranged from $87.9\\%$ to $98.0\\%$ depending on caption and seed (mean $= 94.1\\%$ ).", + "bbox": [ + 212, + 609, + 787, + 657 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Ablation studies.", + "text_level": 1, + "bbox": [ + 215, + 678, + 401, + 691 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To test the validity of our proposed pipeline, we perform a thorough set of ablation studies to showcase how each modeling decision impacts our final method. Our ablations cover two main dimensions: Modeling strategies and data management. Modeling strategies include details related to the inner working of our pipeline's stages, model components, and supervision paradigms. Data management relates to the inclusion of data from additional datasets and different subjects. We report results over BMD, but utilize CC2017 and HAD as pretraining datasets in some experiments as described below. We analyze the impact of the following features:", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "C. Fosco et al.", + "bbox": [ + 271, + 114, + 369, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/7aeb05e4c3bc3cb8c6c09109fdec7f1bc36f44c1d66759748eef930e7f2d0246.jpg", + "image_caption": [ + "Fig. 4: Success and failure cases of reconstructions. A) Additional success cases from BMD. B) Additional failure cases from BMD. C) Specific failure cases from BMD exemplifying semantic dominance (left), high frequency noise (middle), and superposition of indiscernible shapes onto correct reconstruction (right). D) Additional success cases of CC2017. E) Additional CC2017 failure cases. F) Failure cases on CC2017 exemplifying semantic dominance (left), poor facial reconstruction (middle), high frequency noise (right). G) CC2017 failure case: Sudden failure to reconstruct video despite showing previous success in the segment." + ], + "image_footnote": [], + "bbox": [ + 250, + 148, + 750, + 767 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals", + "bbox": [ + 267, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Impact of masked brain modeling: we measure how performance varies when we skip stage 1 and train a regressor from scratch to estimate the latent and conditioning vectors.", + "- Impact of semantic alignment: we observe how the contrastive learning objective affects the regression performance.", + "- Impact of regression backbone: we compare a linear regressor and a modern MLP.", + "- Impact of conditioning target: we analyze the difference between a direct regression to a 257x1024 text conditioning vector compared to regressing a BLIP vector used to build an estimated caption.", + "- Impact of data scaling: we observe how adding an additional pretraining dataset and multiple subjects affect the results." + ], + "bbox": [ + 225, + 146, + 787, + 327 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We report results in Table 4. Our simple pipeline without alignment is the worst performer: the simple regression model struggles to correctly estimate conditioning vectors, showcasing a large MSE. When switching to an MLP, we observe improved MSE, but the reconstructed vectors are still not good enough to correctly retrieve the video. The boost introduced by adding Stage 1, even if it consists only of MBM training, is quite significant and hints at the importance of a structural pretraining stage.", + "bbox": [ + 212, + 338, + 787, + 444 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Interestingly, regressing BLIP embeddings to then produce captions performs better than regressing raw conditioning vectors. The conditioning vectors accepted by zeroscope are hidden states from a CLIP-text model, of shape (257x1024). The BLIP embeddings, although of similar complexity, give a performance boost: we hypothesize that feeding the U-Net a clean caption, which BLIP produces even in the presence of a noisy input, allows the generation model to work in a familiar part of its latent space. This enables less deteriorated generations which translates to better metrics. Finally, scaling data showcases large improvements as well: both including HAD in Stage 1's pretraining and adding training data from other subjects gives boosts across metrics.", + "bbox": [ + 212, + 444, + 787, + 594 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Multi-subject vs. single subject training. Most previous work showcases reconstruction performance on a per-subject basis, as brain responses even to the same stimulus differs between observers [16] [13]. These discrepancies arise mostly from individual differences in brain composition, thought patterns, and the sensitivity of fMRI data collection. As indicated in Section 4, we explicitly train with multiple subjects in our pretraining and regression stages, attempting to build a model that can leverage training data not only from subject $s_0$ , but also $s_1 \\ldots s_n$ to improve test set reconstruction performance for $s_0$ . To shed light on the effect of training on multiple subjects, we report results over subject 1 and subject 2, comparing performance between a pipeline trained solely on subject N's data and a pipeline trained on all subjects. We show results in Table 3.", + "bbox": [ + 212, + 595, + 787, + 762 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Zero-shot reconstruction", + "text_level": 1, + "bbox": [ + 215, + 784, + 465, + 797 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we ask to what extent a model could reconstruct one's visual experience without training on their brain activity. This is a very challenging", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "C. Fosco et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/d7cd44c810529489ba0cc40335e4472b11d933447572d9634c0d34aeab8acc76.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4: Ablation results and additional baselines on BMD's test set. We show results using: an Oracle baseline, where we feed the original video's ground truth $z$ and $b$ vectors to stage 3; a \"random\" baseline, where we instantiate random fMRI vectors and attempt to map those to $z$ and $b$ vectors, an image-only baseline, where, instead of using a video generator, we generate an image with an image generator (Stable Diffusion XL) from the fMRI, and duplicate the frame to match our videos' frame count; a pipeline without stage 1, using a linear regression for stage 2; a pipeline without stage 1, but with our final MLP backbone; a pipeline with a stage 1 comprised of masked-brain-modeling only; a pipeline with CLIP alignment, but still regressing a raw conditioning vector; a pipeline where we replace that regression target (raw conditioning vector) with a BLIP embedding; a pipeline with HAD pretraining, and our full pipeline, trained on multiple subjects. We observe that all design decisions tend to improve performance, and we see a significant boost from the addition of CLIP alignment." + ], + "table_body": "
MethodBMD
SSIMMSE2-way50-way
Oracle0.891-1.00.980
Random0.0511.6770.5010.041
Image-only reconstruction0.0410.7210.6870.125
Linear regression0.1110.8280.6710.120
MLP regressor0.1230.7210.6990.127
+ MBM, no CLIP alignment0.1380.7020.7010.131
+ CLIP alignment0.1780.7610.7150.134
+ BLIP instead of raw cond vector0.1800.7180.7210.140
+ HAD pretraining0.1830.7000.7670.145
Full pipeline (Brainfix multi-subject)0.1900.6710.8160.165
", + "bbox": [ + 263, + 143, + 738, + 323 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "task, further complicated by the structure of the stimulus to reconstruct (see [13] [42] [12]). Current work, including this paper, reports results by training and testing on the same subject, but future applications of this technology would greatly benefit if they could make zero-shot predictions. We analyze the impact of scaling data in this regime: we first train a model on data from subject 2, and attempt to reconstruct test data from Subject 1. As expected, performance deteriorates. However, we find that adding data from more subjects and datasets improves regression performance on subject 1 test data, without ever seeing training examples from that subject. We showcase these results in Figure 5. To further understand this effect, we compute between-subject correlations on the fMRI signals, observing that there are significant clusters of highly correlated responses even over brains from different subjects. We expand on these results in Section 4 of the supplement.", + "bbox": [ + 212, + 550, + 787, + 750 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Limitations.", + "text_level": 1, + "bbox": [ + 214, + 768, + 362, + 784 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Figure 4 shows instances where our reconstruction method fails. Reconstructing complex, high spatial frequency videos (e.g., a crowd of people) remains challenging. In some cases, semantic reconstruction fails: we hypothesize that this is due", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals", + "bbox": [ + 267, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f71a145a1648bbc81410dfdd99a2693d1f7cd4107d47b443e056137ad4eb6f68.jpg", + "image_caption": [ + "Fig. 5: fMRI Correlations and Zero-shot regression performance over latent vector $\\mathbf{z}$ . A) Between subject pairwise correlations for subject 01 are plotted in an inflated brain and flattened brain, showing left hemisphere (LH) and right hemisphere (RH). Both BMD and CC2017 show areas of high correlation across subjects, mostly clustered around visual areas. We hypothesize this correlation might enable zero-shot strategies. B) Zero-shot performance as data from more subjects are added. The left-most bar corresponds to MSE when training on subject 01's train set and evaluating on subject 01's test set. Subsequent bars train on different combinations of subjects except 01, and test on subject 01's test set. We observe that as we add more subjects and datasets, MSE decreases in both examples, hinting at the possibility of developing accurate zero-shot reconstruction with the right scale." + ], + "image_footnote": [], + "bbox": [ + 263, + 146, + 748, + 388 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "to issues in the fMRI data, as our system is vulnerable to situations where the subject's mind wanders and loses focus on the stimulus while data is recorded. Temporally faithful reconstructions are also limited by the temporal capabilities of the available video generation models.", + "bbox": [ + 212, + 588, + 787, + 650 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 679, + 359, + 696 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We present a generative framework to reconstruct short clips from human brain responses. We leverage large pre-trained text-to-video models and two fMRI datasets to train a multi-stage pipeline to reconstruct a video observed by a given subject. Future generative neurotechnology in this vein could help determining the information recovery limits of a brain region, an uncharted terrain of neuroscience. Such technology could foster the development of therapeutic treatments to restore or compensate for sensory processing deprivation, as well as allow for more efficient human-machine communication.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "C. Fosco et al.", + "bbox": [ + 271, + 114, + 369, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research was funded by the Multidisciplinary University Research Initiative (MURI) award by the Army Research Office (grant No. W911NF-23-1-0277) to A.O; the MIT EECS MathWorks Fellowship to B.L.; the MIT EECS MathWorks Fellowship to C.F.", + "bbox": [ + 215, + 172, + 785, + 233 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 253, + 321, + 268 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22563-22575 (June 2023)", + "2. Buccino, G., Binkofski, F., Fink, G.R., Fadiga, L., Fogassi, L., Gallese, V., Seitz, R.J., Zilles, K., Rizzolatti, G., Freund, H.J.: Action observation activates premotor and parietal areas in a somatotopic manner: an fmri study. European journal of neuroscience 13(2), 400-404 (2001)", + "3. Chen, Z., Qing, J., Xiang, T., Yue, W.L., Zhou, J.H.: Seeing beyond the brain: Conditional diffusion model with sparse masked modeling for vision decoding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22710-22720 (2023)", + "4. Chen, Z., Qing, J., Zhou, J.H.: Cinematic mindscapes: High-quality video reconstruction from brain activity. arXiv preprint arXiv:2305.11675 (2023)", + "5. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)", + "6. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)", + "7. Dickie, E.W., Anticevic, A., Smith, D.E., Coalson, T.S., Manogaran, M., Calarco, N., Viviano, J.D., Glasser, M.F., Van Essen, D.C., Voineskos, A.N.: Ciftify: A framework for surface-based analysis of legacy mr acquisitions. Neuroimage 197, 818-826 (2019)", + "8. Gazzola, V., Keysers, C.: The observation and execution of actions share motor and somatosensory voxels in all tested subjects: single-subject analyses of unsmoothed fmri data. Cerebral cortex 19(6), 1239-1255 (2009)", + "9. Glasser, M.F., Coalson, T.S., Robinson, E.C., Hacker, C.D., Harwell, J., Yacoub, E., Ugurbil, K., Andersson, J., Beckmann, C.F., Jenkinson, M., et al.: A multimodal parcellation of human cerebral cortex. Nature 536(7615), 171-178 (2016)", + "10. Glasser, M.F., Sotiropoulos, S.N., Wilson, J.A., Coalson, T.S., Fischl, B., Andersson, J.L., Xu, J., Jbabdi, S., Webster, M., Polimeni, J.R., et al.: The minimal preprocessing pipelines for the human connectome project. Neuroimage 80, 105-124 (2013)", + "11. Han, K., Wen, H., Shi, J., Lu, K.H., Zhang, Y., Fu, D., Liu, Z.: Variational autoencoder: An unsupervised model for encoding and decoding fmri activity in visual cortex. NeuroImage 198, 125-136 (2019)", + "12. Hasson, U., Landesman, O., Knappmeyer, B., Vallines, I., Rubin, N., Heeger, D.J.: Neurocinematics: The neuroscience of film. Projections 2(1), 1-26 (2008)", + "13. Hasson, U., Nir, Y., Levy, I., Fuhrmann, G., Malach, R.: Intersubject synchronization of cortical activity during natural vision. science 303(5664), 1634-1640 (2004)" + ], + "bbox": [ + 218, + 281, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals", + "bbox": [ + 267, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)", + "15. Hendrycks, D., Gimpel, K.: Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415 (2016)", + "16. Kauppi, J.P., Jaëskeläinen, I.P., Sams, M., Tohka, J.: Inter-subject correlation of brain hemodynamic responses during watching a movie: localization in space and frequency. Frontiers in neuroinformatics 4, 669 (2010)", + "17. Konen, C.S., Kastner, S.: Representation of eye movements and stimulus motion in topographically organized areas of human posterior parietal cortex. Journal of Neuroscience 28(33), 8361-8375 (2008)", + "18. Kupershmidt, G., Belyi, R., Gaziv, G., Irani, M.: A penny for your (visual) thoughts: Self-supervised reconstruction of natural movies from brain activity. arXiv preprint arXiv:2206.03544 (2022)", + "19. Lahner, B., Dwivedi, K., Iamshchinina, P., Graumann, M., Lascelles, A., Roig, G., Gifford, A.T., Pan, B., Jin, S., Ratan Murty, N.A., et al.: Bold moments: modeling short visual events through a video fmri dataset and metadata. bioRxiv pp. 2023-03 (2023)", + "20. Le, A., Vesia, M., Yan, X., Crawford, J.D., Niemeier, M.: Parietal area ba7 integrates motor programs for reaching, grasping, and bimanual coordination. Journal of Neurophysiology (2017)", + "21. Le, L., Ambrogioni, L., Seeliger, K., Güçlütürk, Y., van Gerven, M., Güçlü, U.: Brain2pix: Fully convolutional naturalistic video frame reconstruction from brain activity. Frontiers in Neuroscience 16, 940972 (2022)", + "22. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)", + "23. Liu, Z., Guo, Y., Yu, K.: Diffvoice: Text-to-speech with latent diffusion. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)", + "24. Logothetis, N.K., Sheinberg, D.L.: Visual object recognition. Annual review of neuroscience 19(1), 577-621 (1996)", + "25. Monfort, M., Andonian, A., Zhou, B., Ramakrishnan, K., Bargal, S.A., Yan, T., Brown, L., Fan, Q., Gutfreund, D., Vondrick, C., et al.: Moments in time dataset: one million videos for event understanding. IEEE transactions on pattern analysis and machine intelligence 42(2), 502-508 (2019)", + "26. Monfort, M., Pan, B., Ramakrishnan, K., Andonian, A., McNamara, B.A., Lascelles, A., Fan, Q., Gutfreund, D., Feris, R.S., Oliva, A.: Multi-moments in time: Learning and interpreting models for multi-action video understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9434-9445 (2021)", + "27. Morrone, M.C., Tosetti, M., Montanaro, D., Fiorentini, A., Cioni, G., Burr, D.: A cortical area that responds specifically to optic flow, revealed by fmri. Nature neuroscience 3(12), 1322-1328 (2000)", + "28. Newman, A., Fosco, C., Casser, V., Lee, A., McNamara, B., Oliva, A.: Multimodal memorability: Modeling effects of semantics and decay on video memorability. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16. pp. 223-240. Springer (2020)", + "29. Nishimoto, S., Vu, A.T., Naselaris, T., Benjamini, Y., Yu, B., Gallant, J.L.: Reconstructing visual experiences from brain activity evoked by natural movies. Current biology 21(19), 1641-1646 (2011)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "C. Fosco et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Peeters, R., Simone, L., Nelissen, K., Fabbri-Destro, M., Vanduffel, W., Rizzolatti, G., Orban, G.A.: The representation of tool use in humans and monkeys: common and uniquely human features. Journal of Neuroscience 29(37), 11523-11539 (2009)", + "31. Peeters, R.R., Rizzolatti, G., Orban, G.A.: Functional properties of the left parietal tool use region. Neuroimage 78, 83-93 (2013)", + "32. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "33. Rizzolatti, G., Sinigaglia, C.: The functional role of the parieto-frontal mirror circuit: interpretations and misinterpretations. Nature reviews neuroscience 11(4), 264-274 (2010)", + "34. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "35. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022)", + "36. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)", + "37. Sahara, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(4), 4713-4726 (2022)", + "38. Schultz, J., Pilz, K.S.: Natural facial motion enhances cortical responses to faces. Experimental brain research 194, 465-475 (2009)", + "39. Scotti, P., Banerjee, A., Goode, J., Shabalin, S., Nguyen, A., Dempster, A., Verlinde, N., Yundler, E., Weisberg, D., Norman, K., et al.: Reconstructing the mind's eye: fmri-to-image with contrastive learning and diffusion priors. Advances in Neural Information Processing Systems 36 (2024)", + "40. Silver, M.A., Kastner, S.: Topographic maps in human frontal and parietal cortex. Trends in cognitive sciences 13(11), 488-495 (2009)", + "41. Smith, S.M., Vidaurre, D., Beckmann, C.F., Glasser, M.F., Jenkinson, M., Miller, K.L., Nichols, T.E., Robinson, E.C., Salimi-Khorshidi, G., Woolrich, M.W., et al.: Functional connectomics from resting-state fmri. Trends in cognitive sciences 17(12), 666-682 (2013)", + "42. Smith, T.J., Levin, D., Cutting, J.E.: A window on reality: Perceiving edited moving images. Current Directions in Psychological Science 21(2), 107-113 (2012)", + "43. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)", + "44. Takagi, Y., Nishimoto, S.: High-resolution image reconstruction with latent diffusion models from human brain activity. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14453-14463 (2023)", + "45. Van Essen, D.C., Smith, S.M., Barch, D.M., Behrens, T.E., Yacoub, E., Ugurbil, K., Consortium, W.M.H., et al.: The wu-minn human connectome project: an overview. Neuroimage 80, 62-79 (2013)", + "46. VanRullen, R., Thorpe, S.J.: The time course of visual processing: from early perception to decision-making. Journal of cognitive neuroscience 13(4), 454-461 (2001)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals", + "bbox": [ + 267, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "47. Wang, C., Yan, H., Huang, W., Li, J., Wang, Y., Fan, Y.S., Sheng, W., Liu, T., Li, R., Chen, H.: Reconstructing rapid natural vision with fmri-conditional video generative adversarial network. Cerebral Cortex 32(20), 4502-4511 (2022)", + "48. Wang, J., Yuan, H., Chen, D., Zhang, Y., Wang, X., Zhang, S.: Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571 (2023)", + "49. Wang, L., Mruczek, R.E., Arcaro, M.J., Kastner, S.: Probabilistic maps of visual topography in human cortex. Cerebral cortex 25(10), 3911-3931 (2015)", + "50. Wei, C., Fan, H., Xie, S., Wu, C.Y., Yuille, A., Feichtenhofer, C.: Masked feature prediction for self-supervised visual pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14668-14678 (2022)", + "51. Wen, H., Shi, J., Zhang, Y., Lu, K.H., Cao, J., Liu, Z.: Neural encoding and decoding with deep learning for dynamic natural vision. Cerebral cortex 28(12), 4136-4160 (2018)", + "52. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: A simple framework for masked image modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9653-9663 (2022)", + "53. Yildirim, I., Wu, J., Kanwisher, N., Tenenbaum, J.: An integrative computational architecture for object-driven cortex. Current opinion in neurobiology 55, 73-81 (2019)", + "54. Yu, K.P., Zhang, Z., Hu, F., Chai, J.: Efficient in-context learning in vision-language models for egocentric videos. arXiv preprint arXiv:2311.17041 (2023)", + "55. Zhao, H., Torralba, A., Torresani, L., Yan, Z.: Hacs: Human action clips and segments dataset for recognition and temporal localization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8668-8678 (2019)", + "56. Zhou, M., Gong, Z., Dai, Y., Wen, Y., Liu, Y., Zhen, Z.: A large-scale fmri dataset for human action recognition. Scientific Data 10(1), 415 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 507 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "C. Fosco et al.", + "bbox": [ + 271, + 114, + 369, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Brain Netflix_ Scaling Data to Reconstruct Videos from Brain Signals/9af6f00c-d99e-4633-82c3-bc4c9c80afb6_model.json b/2024/Brain Netflix_ Scaling Data to Reconstruct Videos from Brain Signals/9af6f00c-d99e-4633-82c3-bc4c9c80afb6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1d4d5e71a5e32f30847772e2d8bc40df259071ac --- /dev/null +++ b/2024/Brain Netflix_ Scaling Data to Reconstruct Videos from Brain Signals/9af6f00c-d99e-4633-82c3-bc4c9c80afb6_model.json @@ -0,0 +1,2073 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.258, + 0.141, + 0.747, + 0.187 + ], + "angle": 0, + "content": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.213, + 0.772, + 0.245 + ], + "angle": 0, + "content": "Camilo Fosco\\*1, Benjamin Lahner\\*1, Bowen Pan\\(^{1}\\), Alex Andonian\\(^{1}\\), Emilie Josephs\\(^{1}\\), Alex Lascelles\\(^{1}\\), and Aude Oliva\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.256, + 0.774, + 0.285 + ], + "angle": 0, + "content": "CSAIL, Massachusetts Institute of Technology, MA, USA [camilolu, blahner, bpan, andonian, ejosephs, alexlasc, oliva]@mit.edu" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.331, + 0.744, + 0.637 + ], + "angle": 0, + "content": "Abstract. The field of brain-to-stimuli reconstruction has seen significant progress in the last few years, but techniques continue to be subject-specific and are usually tested on a single dataset. In this work, we present a novel technique to reconstruct videos from functional Magnetic Resonance Imaging (fMRI) signals designed for performance across datasets and across human participants. Our pipeline accurately generates 2 and 3-second video clips from brain activity coming from distinct participants and different datasets by leveraging multi-dataset and multi-subject training. This helps us regress key latent and conditioning vectors for pretrained text-to-video and video-to-video models to reconstruct accurate videos that match the original stimuli observed by the participant. Key to our pipeline is the introduction of a 3-stage approach that first aligns fMRI signals to semantic embeddings, then regresses important vectors, and finally generates videos with those estimations. Our method demonstrates state-of-the-art reconstruction capabilities verified by qualitative and quantitative analyses, including crowd-sourced human evaluation. We showcase performance improvements across two datasets, as well as in multi-subject setups. Our ablation studies shed light on how different alignment strategies and data scaling decisions impact reconstruction performance, and we hint at a future for zero-shot reconstruction by analyzing how performance evolves as more subject data is leveraged." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.667, + 0.377, + 0.683 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The human brain processes several million bits of information per second and compresses most of this information in ways that are still, in great part, unknown. What if we could partially recover this information by reading out brain signals? Reconstructing visual stimuli from brain activity is a promising way to investigate what information content is encoded in the human brain and how it may differ between participants. Reconstruction efforts typically require large amounts of neural-stimuli pairs to train high-parameter computational models, and inter-subject analyses additionally require repeated experiments across subjects. However, these neural-stimuli pairs are often not abundantly available due" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "C. Fosco et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.787, + 0.192 + ], + "angle": 0, + "content": "to their expensive and time-consuming collection process. Reconstructing dynamic videos, as opposed to static images, poses the additional challenges of temporal continuity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.788, + 0.433 + ], + "angle": 0, + "content": "Previous fMRI (Functional Magnetic Resonance Imaging) work has shown that a greater extent of cortex responds to videos than still images [53] [38] [2], that some cortical regions are selective for motion features [2] [17] [27], and that inter-subject correlations to movie viewings are high [16] [13], all suggesting that fMRI reliably encodes dynamic content. Humans do not all perceive videos in the same way, however; in-the-wild longform videos, such as an unedited video of the happenings in a park, show markedly lower inter-subject correlations than edited cinematic films [12] [13]. These higher inter-subject correlations in cinematic films are largely due to production effects, such as camera angle, editing, and lighting, that structure everyone's attention and interpretation of the content in a similar way [42]. The everyday human visual experience includes events sampled at both ends of this continuum, from unstructured meanderings down a busy street to a structured viewing of a work presentation. Thus, while videos can be reconstructed from brain activity [4], cross-subject reconstructions and the structure of the video stimulus set have yet to be systematically studied in the video reconstruction domain." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.617 + ], + "angle": 0, + "content": "Here, we leverage two recent computational developments (latent diffusion models (LDMs) for video and masked-brain modeling) and several large-scale fMRI datasets to achieve high fidelity reconstructions of previously viewed 2- and 3-second videos. Latent diffusion models [34] are a class of generative models that generate high-fidelity images or videos, conditioned on inputs such as text, through a denoising diffusion process over a latent representation of the pixels. Masked brain modeling (MBM), an extension of masked training methods from natural language processing and computer vision to brain data, is a technique to learn robust self-supervised representations of brain signals by masking the signal and training a encoder-decoder network to reconstruct it. It has been shown to be a powerful way for a model to learn compressed representations of fMRI brain activity [4] [3]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.616, + 0.788, + 0.753 + ], + "angle": 0, + "content": "We leverage these two concepts to propose a novel 3-stage pipeline that a) learns robust fMRI representations through masked brain modeling and alignment, b) regresses conditioning vectors from these representations, and c) reconstructs target videos with high fidelity, leveraging the estimated conditioning vectors. Importantly, our alignment and training strategy combines information from multiple subjects, as opposed to previous works that train one model per subject. This allows our model to not only improve its general reconstruction quality, but also to be able to generalize from one subject to another with much higher performance. Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.765, + 0.784, + 0.809 + ], + "angle": 0, + "content": "1. We introduce a 3-stage pipeline for high-quality video reconstruction that leverages multiple fMRI datasets and aligns inter-subject data, outperforming existing approaches." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.81, + 0.784, + 0.839 + ], + "angle": 0, + "content": "2. We showcase a detailed ablation study that illustrates how the model performs in different conditions, including modeling and data variants." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.765, + 0.784, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.268, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "3. We perform multi-subject and zero-shot reconstruction experiments to explore the possibility of a subject-agnostic visual reconstruction technology." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.19, + 0.784, + 0.221 + ], + "angle": 0, + "content": "Together, this work offers both modeling and neuroscience insights for video reconstruction from brain activity." + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.252, + 0.239, + 0.263 + ], + "angle": 0, + "content": "A" + }, + { + "type": "image", + "bbox": [ + 0.239, + 0.263, + 0.767, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.369, + 0.238, + 0.38 + ], + "angle": 0, + "content": "B" + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.387, + 0.781, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.47, + 0.785, + 0.526 + ], + "angle": 0, + "content": "Fig. 1: Reconstruction of short videos from fMRI activity A) Brain responses of subjects viewing video clips was measured with fMRI and then used to reconstruct the seen video. B) We show three examples of the ground truth video (top row) and our reconstruction using our approach (bottom row)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.581, + 0.387, + 0.596 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.787, + 0.779 + ], + "angle": 0, + "content": "Diffusion models: Latent diffusion models (LDMs) [34] are a class of generative model that achieves photorealistic generations through a process of denoising latent representations. This denoising process can be conditioned on other inputs, such as text, in order to guide the model toward generating outputs with specific semantic meaning of visual appearance. These models have been successful in reconstruction tasks, including image generation, super-resolution and recoloring [6, 35-37, 43], audio generation [23], among others. Recent text-to-video latent diffusion models have generated high fidelity reconstructions of short videos faithful to both content and temporal dynamics [1]. Here, we leverage these advances in video generation to form the backbone of our framework: we use Zeroscope V2, which creates 3 second videos based on text conditioning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.786, + 0.841 + ], + "angle": 0, + "content": "Masked Brain modeling: Masked Brain Modeling [3] (MBM) is a recent method for pre-training fMRI reconstruction models. It is based on Masked Signal Modeling (MSM), a self-supervised learning task commonly used for pretraining in natural language processing and vision applications [5,14,50,52]. In" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "C. Fosco et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.221 + ], + "angle": 0, + "content": "MSM, some amount of the signal is masked during training, and an autoencoder is trained to predict the masked content, which can be fine tuned with additional training. In MBM, a percentage of the brain activity in masked and the model is tasked with reconstructing the complete input, which pushes the MBM model to learn compressed representations of fMRI brain activity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.788, + 0.404 + ], + "angle": 0, + "content": "Video reconstruction from brain activity: An early naturalistic video reconstruction work used a voxelwise motion-energy model to recover videos with similar spatiotemporal energy, showing that the BOLD signal captures dynamic visual information [29]. Other approaches used convolutional neural networks (CNNs) to reconstruct videos from fMRI data [51] [21], with one even leveraging a convolutional encoder-decoder framework to synthesize high framerate fMRI-video frame pairs [18]. A work using GANs employed dual spatial and temporal discriminators to achieve spatiotemporally faithful reconstructions [47], and variational autoencoders have been used to reconstruct videos frame-by-frame from a hidden latent space [11]. Most similar to our approach, [4] uses masked-brain modeling to train a vision transformer encoder that then feeds latent representations into a diffusion model to reconstruct the video." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.429, + 0.461, + 0.446 + ], + "angle": 0, + "content": "3 Video fMRI Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.674 + ], + "angle": 0, + "content": "We train and evaluate our reconstruction method on a compilation of four large-scale fMRI datasets [19] [56] [45] [51] to obtain a large quantity and diversity of brain responses. In order to effectively perform inter-subject and inter-dataset analyses we analyze all fMRI responses in Cifti fsLR32k space (see data preprocessing details in the supplement.) [10] [7]. This preprocessing pipeline accurately registers gray matter voxels to a shared cortical surface mesh, thereby allowing the model to learn informative spatial activation patterns across datasets. From the cortical surface, we select a subset of 41 regions of interest (ROIs) from the Glasser Atlas (see supplement for details) [9]. This subset of cortex is isolates brain regions that might respond to dynamic stimuli [46] [24] [20] [8] [33] [40] [30] [31] [49] as well as reduce the computational load of model training. All datasets combined, our task-based video fMRI data draws from 43 subjects, over 28,100 short video segments, and over 123,000 fMRI response trials, and our resting state fMRI data totals over one thousand hours across 1084 subjects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Human Connectome Project Dataset (HCP). The 1200-subject release of the Human Connectome Project [45] includes nearly an hour of resting state scans on 1084 subjects. During acquisition of resting state data, subjects were instructed to remain awake and fixate on a cross-hair but completed no task nor viewed any other visual stimulus. Resting state fMRI brain activity captures temporally correlated functional networks that can be useful for understanding cortical organization [41]. We utilize this time series data to pre-train our Masked Brain Modeling encoder (Figure 2)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "BOLD Moments Dataset (BMD). The BOLD Moments Dataset (BMD) [19] consists of fMRI responses to 1,102 3-second videos for ten human subjects collected in a rapid event-related design. Beta values were estimated at each" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.269, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.312 + ], + "angle": 0, + "content": "cortical vertex for each video using a general linear model (GLM). The stimuli were divided into a non-overlapping 1,000 video training set and a 102 video testing set. Each subject viewed the training set videos 3 times and the testing set videos 10 times to facilitate both inter-subject and intra-subject analyses. The 1,102 videos were sampled from the 1 million video Moments in Time dataset [25] (overlapping with the Multi-Moments in Time [26] and Memento10k datasets [28]) and capture naturalistic, in-the-wild content like home videos. We use BMD's five sentence text description metadata provided for each video to generate ground truth text embeddings that our pipeline learns to estimate. To the best of our knowledge, this work is the first time BMD has been used for video reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Human Actions Dataset (HAD). The Human Actions Dataset (HAD) [56] is composed of 30 human subjects each viewing 720 two second videos sampled from the Human Action Clips and Segments (HACS) dataset [55]. Similar to BMD, data were collected in a rapid event-related design and beta values were estimated across cortex for each video. Each video is annotated with one of 180 human-centric actions, and each subject viewed four (different) videos from each of the 180 action categories. Video presentations were not repeated within or across subjects. For each subject, we define a testing/training split by randomly selecting one of the four videos per action category as a testing video. We use this dataset primarily to train Stage 1 in our pipeline. To the best of our knowledge, this work is the first time HAD has been used in video reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.788, + 0.737 + ], + "angle": 0, + "content": "CC2017 Dataset (CC2017). In the CC2017 dataset [51], three subjects viewed continuous training and testing movies composed of shorter 10-15 second movie clips under a longform viewing paradigm. A training movie was composed from 374 video clips and randomly split into 18 eight minute segments. A testing movie was composed of 598 video clips and randomly split into five eight minute segments. Thus, each eight minute segment was composed of a concatenation of shorter video clips varying in duration. The clips are sourced from professionally shot videos, such as cinematic movies or stock advertising footage. The training and testing segments were repeated two and ten times per subject, respectively. Here, we divide the 8 minute segments into 2 second clips for reconstruction. We compose fMRI-video pairs by sampling the fMRI time series at an offset of 4 seconds (to account for the hemodynamic lag in the BOLD response) to capture peak BOLD activation corresponding to short snippets of the movie (2 second snippets, same as CC2017's acquisition TR). As this dataset is one of the main benchmarks used in previous work [4], we utilize this data extensively to train our regressors and provide reconstructions and evaluation metrics." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.762, + 0.486, + 0.779 + ], + "angle": 0, + "content": "4 Reconstruction Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Approach Overview. Our reconstruction pipeline aims to regress key latent and conditioning vectors from brain signals to guide a pretrained video generation model towards reconstructing the original stimulus. This pipeline consists of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.128 + ], + "angle": 0, + "content": "C. Fosco et al." + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.145, + 0.773, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.395, + 0.788, + 0.52 + ], + "angle": 0, + "content": "Fig. 2: Our proposed reconstruction framework. Stage 1: We train an encoder-decoder structure to reconstruct fMRI signals from a compressed representation. We then fine-tune the MBM encoder to align with a CLIP embedding of the video. Stage 2: we regress the latent vector (z) and blip embedding (b) of the target video from stage 1's aligned embedding. Stage 3: we use the regressed conditioning vectors to generate the video. The estimated vector z is artificially noised and then denoised with a pre-trained denoising U-Net, following a video-to-video procedure. The estimated vector b is sent to a pretrained BLIP model to generate a caption, which is used as a conditioning input to the U-Net." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.779 + ], + "angle": 0, + "content": "3 stages. First, in an initial alignment stage, an encoder-decoder architecture is trained to reconstruct masked fMRI embeddings while simultaneously following a CLIP-like contrastive learning objective. This setup forces the encoder to map fMRI inputs into an embedding that captures both knowledge about the raw signal and knowledge about the semantic properties of the original stimuli. Second, the output of stage 1's encoder is used in a regression stage: we regress the necessary inputs required by our generative model with two MLPs that estimate an initial latent vector \\((z)\\) and BLIP embeddings \\((b)\\) [22]. The BLIP embeddings are used to infer a conditioning caption. Third, our reconstruction stage uses the predicted latent vectors and BLIP embeddings to generate a video: Given a new fMRI input, we pass them through the regressors to predict the latent and conditioning vectors \\((z\\) and \\(b)\\), re-noise \\(z\\) and use \\(b\\) to generate a caption. We feed the noised \\(z\\) and caption to a pretrained denoising U-Net (Zeroscope v2 [48]) to estimate the final reconstruction. We describe each stage in detail below and show a pseudocode implementation in the supplement Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Stage 1: MBM and Alignment. Our model employs a straightforward encoder-decoder framework aimed at reconstructing the input fMRI signal by compressing it into a 1024-dimensional latent vector. During this phase, the input fMRI signal is divided into several patches, and certain patches are masked." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.268, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "The encoder processes the signal with masked patches to produce a sequence of 1024-dimensional latent vectors. The decoder then attempts to transform these vectors back into the fMRI signal. Throughout the training process, the encoder-decoder structure is supervised through its ability to accurately reconstruct the masked patches. This approach, which we call Masked Brain Modeling (MBM), draws inspiration from the masked image modeling task [14] typical in the self-supervision literature. Its primary purpose is to acquaint the model with the spatial structure of the fMRI signal, rather than to imbue it with any semantic understanding. We first train our encoder-decoder pipeline on the masked modeling task alone with the large-scale HCP dataset [45] and HAD dataset [56], which ensures our embeddings learn basic fMRI structure. We utilize a simple MSE loss over the masked patches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.466 + ], + "angle": 0, + "content": "After this step, we further finetune the MBM encoder through a contrastive learning task. Contrastive learning is a powerful self-supervised technique to learn representations across modalities. We extract CLIP embeddings from captions describing each video, and we apply a contrastive loss over batches of encoder embeddings and CLIP embeddings. This contrastive loss follows [32]: for each batch, the loss enforces high cosine similarity between an fMRI embedding and its matching CLIP embedding (positive pair), while minimizing the similarity between that same embedding and all other CLIP embeddings in the batch (negative pairs):" + }, + { + "type": "equation", + "bbox": [ + 0.348, + 0.48, + 0.657, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n t r a s t i v e}} = - \\sum_ {i = 1} ^ {N} \\log \\left(\\frac {\\exp \\left(\\frac {f _ {i} * c _ {i}}{\\tau}\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\frac {f _ {j} * c _ {j}}{\\tau}\\right)}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.628 + ], + "angle": 0, + "content": "Here, \\( f_{i} \\) is the fMRI embedding (output from the MBM encoder), \\( c_{i} \\) is the CLIP-text embedding, and \\( \\tau \\) is a temperature hyperparameter. We train this model by following the training details in [3] for MBM and [39] for the alignment procedure, where we use a encoder-decoder model with patch size of 16, hidden dimension of 1024 and 24 layers in the encoder. We use a temperature of 0.9 and a \\( 75\\% \\) mask ratio following previous work [3]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Stage 2: Regression. To map from embeddings generated by the encoding model to latent and conditioning vectors suitable for the video generation pipeline, we utilize a multi-target regression model with regularization. Our regression models, instantiated as both a Ridge regression and a regularized MLP (see section 5.2 for comparisons), map outputs from Stage 1 into a latent vector \\( z \\) of shape (4, 15, 33, 33) and a BLIP [22] embedding \\( b \\) of shape (226, 768). The decision to map to BLIP vectors is rooted in how video generators handle noise in their conditioning inputs: similar to [44], we found that feeding a raw caption, even if imperfect, to our denoising U-Net performed better than feeding a regressed (and thus noisy) conditioning vector (comparisons in section 5.2). We hypothesize that the performance of video generators is susceptible to noise in the conditioning vector, and thus encapsulating the regression's imperfections at the level of BLIP's inputs allows the generator to receive clean (but imprecise) text tokens, which improves performance." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.128 + ], + "angle": 0, + "content": "C. Fosco et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "Our MLPs are composed of an initial linear layer, followed by 3 residual blocks and a final output layer. The initial layer and the residual blocks contain 2048 units, dropout regularization with \\( p = 0.3 \\), a GELU [15] activation and Batch Normalization. They are trained with an MSE loss and predict flattened vectors, which are then reshaped and used in Stage 3. Perfect regression at this stage equates to obtaining exact latent and conditioning vectors, which would result in a near-perfect reconstruction when fed to Stage 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.79, + 0.466 + ], + "angle": 0, + "content": "Stage 3: Reconstruction. During video reconstruction, we extract the latent vector \\( z \\) and BLIP embedding \\( b \\) using the pretrained MBM encoder followed by the \\( z \\) and \\( b \\) regressors from stages 1 and 2. \\( z \\) undergoes a re-noising process, followed by a denoising procedure using a pretrained U-Net that follows [48]. We renoise the latents with a strength of 0.8 and apply 40 denoising steps, following insights from previous work [44]. Concurrently, \\( b \\) is decoded using BLIP's decoder to produce a semantically relevant caption that conditions the U-Net during the denoising process. We observed during experiments that enforcing more descriptive and less repetitive captions tended to improve the quality of BLIP outputs, so we implement decoding with a high repetition penalty of 6, a minimum length of 4 and a maximum length of 20. The integration of \\( z \\) and \\( b \\) helps facilitate video generation that is both visually and contextually aligned with the video represented in the original fMRI signal. Our video generation model is frozen during this process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.491, + 0.497, + 0.508 + ], + "angle": 0, + "content": "5 Experiments and Results" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.525, + 0.513, + 0.54 + ], + "angle": 0, + "content": "5.1 Comparisons to previous work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.787, + 0.597 + ], + "angle": 0, + "content": "To showcase the quality of Brainflix, we obtain reconstructions from our full pipeline and compare to previous fMRI-to-video techniques. We report results over BMD and CC2017." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Implementation Details. We use the large-scale fMRI datasets introduced in section 3 for model training and finetuning. For all datasets, videos are downsampled to 15 FPS and resized to \\(224 \\times 224\\). We first train our alignment stage over the masked brain reconstruction task leveraging the HCP, HAD and BMD datasets for 200 epochs with a batch size of 300. We then finetune this model with a CLIP-like alignment objective over BMD and CC2017 simultaneously for 50 epochs with a batch size of 120. This aligns the output of the MBM encoder with embeddings of the captions describing each video. The captions are obtained through human labelers for BMD and synthetically for the rest of the datasets, leveraging the EILEV [54] video-to-caption model. Training for Stage 1 is done over 6 V100 GPUs, while training for Stage 2, a smaller scale endeavor given the smaller size of the MLPs, is performed over 2 Titan RTX GPUs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Evaluation Metrics. Following previous work [4, 44], we utilize 3 main evaluation metrics that aim to understand different characteristics of performance. To measure pixel-level reconstruction quality, we utilize Structural Similarity (SSIM). For semantic evaluation, we use the N-way top-k classification approach" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.144, + 0.788, + 0.242 + ], + "angle": 0, + "content": "
MethodsCC2017BMD
SSIMMSE2-way50-waySSIMMSE2-way50-way
Kupershmidt0.128---0.0314.5610.5140.004
Mind-Video0.186-0.8530.2020.1760.7630.7110.101
Brainflux (Ours)0.1890.6810.8710.2030.1830.6910.7670.145
Brainfix multi-subject (Ours)0.1950.6550.8880.2210.1900.6710.8160.165
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.217, + 0.243, + 0.788, + 0.298 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison of our reconstruction performance against previous video reconstruction methods. Our methodology achieves state of the art on most metrics. Results from Kupershmidt and Mind-video on BMD are obtained through a reimplementation, as their code is not readily available." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.786, + 0.446 + ], + "angle": 0, + "content": "from [4], which measures how often the classification of a simple ImageNet classifier over the reconstruction and ground truth video match, limiting the output of the classifier to N classes. We declare a successful trial if the ground truth class is within the top-k probabilities outputted over the reconstruction, and repeat the test 100 times to report average success rates. We also report MSE results over our target latent embeddings \\( z \\) and BLIP embeddings \\( b \\), to observe how close our regressed embeddings are to ground truth." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.787, + 0.538 + ], + "angle": 0, + "content": "Evaluation Datasets While we train on all datasets introduced in Section 3, we evaluate our reconstruction results on two large-scale video datasets, CC2017 [51] and BMD [19], to compare our method with previous work [4] [18] [47], as well as to introduce an additional dataset, BMD, with different subjects and video stimuli. These two datasets exhibit reasonable variety, complexity and semantic diversity, which makes them strong options for evaluation exercises." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.787, + 0.704 + ], + "angle": 0, + "content": "Results. We showcase our quantitative results over BMD and CC2017 in Table 1. Results are reported for Subject 1 in both datasets. Both our single-subject and our multi-subject approaches outperform previous work. Our multi-subject model, trained on all subjects from a given dataset before predicting on one, showcases larger boosts in performance which provides evidence towards the advantages of multi-subject training. We show qualitative results in Figure 3 and 4: we observe that our model is able to reconstruct examples from BMD and CC2017 with strong structural reliability. We hypothesize that our model's emphasis on regressing an accurate latent, a component that previous approaches lack, enforces accurate structural patterns that can then lead to reliable object positioning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Human Evaluation. To further measure how preferable our reconstructions are, we collect qualitative evaluations of semantic fidelity using human judgments. We designed a simple web experiment where participants indicated which of 6 reconstructed videos was the best semantic match to a real video. One of the 6 videos was a reconstruction of the reference video. We calculated the percent of trials where the selected video corresponded to the reconstruction of the reference video. Table 2 reports Brainflux results for two datasets. Our best performing method reached a semantic fidelity score of \\(86\\%\\) on BMD and \\(84\\%\\) on CC2017. This substantially exceeds chance \\((16.66\\%)\\), but remains below the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.128 + ], + "angle": 0, + "content": "C. Fosco et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.148, + 0.785, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.401 + ], + "angle": 0, + "content": "Fig. 3: Comparison to previous works on examples from CC2017. We show the ground truth videos in the first row, our results in the second row, and previous approaches in the following rows. Our model manages to capture structural similarity in a better way than previous approaches." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.416, + 0.437, + 0.482 + ], + "angle": 0, + "content": "
MethodBMDCC2017
Baseline (Oracle)94.192.8
MindVideo [4]80.383.9
Brainflix (Ours)86.184.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.482, + 0.421, + 0.564 + ], + "angle": 0, + "content": "Table 2: Human evaluation results. Accuracy on a 6-way retrieval task, where subjects select the ground truth video that most closely matches a reconstructed reference." + }, + { + "type": "table", + "bbox": [ + 0.47, + 0.414, + 0.79, + 0.513 + ], + "angle": 0, + "content": "
Training dataBMDCC2017
SSIMMSESSIMMSE
Subject 1 only0.1230.7210.6990.127
Subject 1 + all subs0.1290.7020.7100.131
Subject 2 only0.1200.7350.6800.122
Subject 2 + all subs0.1280.7220.7080.126
" + }, + { + "type": "table_caption", + "bbox": [ + 0.469, + 0.513, + 0.788, + 0.569 + ], + "angle": 0, + "content": "Table 3: Multi-subject vs single-subject results. We observe that training on all subjects is value-adding for both datasets and both metrics. This effect is measurable across subjects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.788, + 0.658 + ], + "angle": 0, + "content": "best possible performance expected from oracle reconstructions given our backbone model (Zeroscope V2 [48]), which ranged from \\(87.9\\%\\) to \\(98.0\\%\\) depending on caption and seed (mean \\(= 94.1\\%\\))." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.679, + 0.403, + 0.693 + ], + "angle": 0, + "content": "5.2 Ablation studies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.841 + ], + "angle": 0, + "content": "To test the validity of our proposed pipeline, we perform a thorough set of ablation studies to showcase how each modeling decision impacts our final method. Our ablations cover two main dimensions: Modeling strategies and data management. Modeling strategies include details related to the inner working of our pipeline's stages, model components, and supervision paradigms. Data management relates to the inclusion of data from additional datasets and different subjects. We report results over BMD, but utilize CC2017 and HAD as pretraining datasets in some experiments as described below. We analyze the impact of the following features:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.269, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.149, + 0.75, + 0.768 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.782, + 0.788, + 0.894 + ], + "angle": 0, + "content": "Fig. 4: Success and failure cases of reconstructions. A) Additional success cases from BMD. B) Additional failure cases from BMD. C) Specific failure cases from BMD exemplifying semantic dominance (left), high frequency noise (middle), and superposition of indiscernible shapes onto correct reconstruction (right). D) Additional success cases of CC2017. E) Additional CC2017 failure cases. F) Failure cases on CC2017 exemplifying semantic dominance (left), poor facial reconstruction (middle), high frequency noise (right). G) CC2017 failure case: Sudden failure to reconstruct video despite showing previous success in the segment." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "C. Fosco et al." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.147, + 0.788, + 0.192 + ], + "angle": 0, + "content": "- Impact of masked brain modeling: we measure how performance varies when we skip stage 1 and train a regressor from scratch to estimate the latent and conditioning vectors." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.193, + 0.788, + 0.223 + ], + "angle": 0, + "content": "- Impact of semantic alignment: we observe how the contrastive learning objective affects the regression performance." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.223, + 0.788, + 0.252 + ], + "angle": 0, + "content": "- Impact of regression backbone: we compare a linear regressor and a modern MLP." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.253, + 0.788, + 0.298 + ], + "angle": 0, + "content": "- Impact of conditioning target: we analyze the difference between a direct regression to a 257x1024 text conditioning vector compared to regressing a BLIP vector used to build an estimated caption." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.299, + 0.788, + 0.328 + ], + "angle": 0, + "content": "- Impact of data scaling: we observe how adding an additional pretraining dataset and multiple subjects affect the results." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.788, + 0.445 + ], + "angle": 0, + "content": "We report results in Table 4. Our simple pipeline without alignment is the worst performer: the simple regression model struggles to correctly estimate conditioning vectors, showcasing a large MSE. When switching to an MLP, we observe improved MSE, but the reconstructed vectors are still not good enough to correctly retrieve the video. The boost introduced by adding Stage 1, even if it consists only of MBM training, is quite significant and hints at the importance of a structural pretraining stage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.595 + ], + "angle": 0, + "content": "Interestingly, regressing BLIP embeddings to then produce captions performs better than regressing raw conditioning vectors. The conditioning vectors accepted by zeroscope are hidden states from a CLIP-text model, of shape (257x1024). The BLIP embeddings, although of similar complexity, give a performance boost: we hypothesize that feeding the U-Net a clean caption, which BLIP produces even in the presence of a noisy input, allows the generation model to work in a familiar part of its latent space. This enables less deteriorated generations which translates to better metrics. Finally, scaling data showcases large improvements as well: both including HAD in Stage 1's pretraining and adding training data from other subjects gives boosts across metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Multi-subject vs. single subject training. Most previous work showcases reconstruction performance on a per-subject basis, as brain responses even to the same stimulus differs between observers [16] [13]. These discrepancies arise mostly from individual differences in brain composition, thought patterns, and the sensitivity of fMRI data collection. As indicated in Section 4, we explicitly train with multiple subjects in our pretraining and regression stages, attempting to build a model that can leverage training data not only from subject \\( s_0 \\), but also \\( s_1 \\ldots s_n \\) to improve test set reconstruction performance for \\( s_0 \\). To shed light on the effect of training on multiple subjects, we report results over subject 1 and subject 2, comparing performance between a pipeline trained solely on subject N's data and a pipeline trained on all subjects. We show results in Table 3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.785, + 0.466, + 0.799 + ], + "angle": 0, + "content": "5.3 Zero-shot reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "In this section, we ask to what extent a model could reconstruct one's visual experience without training on their brain activity. This is a very challenging" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.269, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.264, + 0.144, + 0.74, + 0.324 + ], + "angle": 0, + "content": "
MethodBMD
SSIMMSE2-way50-way
Oracle0.891-1.00.980
Random0.0511.6770.5010.041
Image-only reconstruction0.0410.7210.6870.125
Linear regression0.1110.8280.6710.120
MLP regressor0.1230.7210.6990.127
+ MBM, no CLIP alignment0.1380.7020.7010.131
+ CLIP alignment0.1780.7610.7150.134
+ BLIP instead of raw cond vector0.1800.7180.7210.140
+ HAD pretraining0.1830.7000.7670.145
Full pipeline (Brainfix multi-subject)0.1900.6710.8160.165
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.214, + 0.327, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Table 4: Ablation results and additional baselines on BMD's test set. We show results using: an Oracle baseline, where we feed the original video's ground truth \\( z \\) and \\( b \\) vectors to stage 3; a \"random\" baseline, where we instantiate random fMRI vectors and attempt to map those to \\( z \\) and \\( b \\) vectors, an image-only baseline, where, instead of using a video generator, we generate an image with an image generator (Stable Diffusion XL) from the fMRI, and duplicate the frame to match our videos' frame count; a pipeline without stage 1, using a linear regression for stage 2; a pipeline without stage 1, but with our final MLP backbone; a pipeline with a stage 1 comprised of masked-brain-modeling only; a pipeline with CLIP alignment, but still regressing a raw conditioning vector; a pipeline where we replace that regression target (raw conditioning vector) with a BLIP embedding; a pipeline with HAD pretraining, and our full pipeline, trained on multiple subjects. We observe that all design decisions tend to improve performance, and we see a significant boost from the addition of CLIP alignment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.75 + ], + "angle": 0, + "content": "task, further complicated by the structure of the stimulus to reconstruct (see [13] [42] [12]). Current work, including this paper, reports results by training and testing on the same subject, but future applications of this technology would greatly benefit if they could make zero-shot predictions. We analyze the impact of scaling data in this regime: we first train a model on data from subject 2, and attempt to reconstruct test data from Subject 1. As expected, performance deteriorates. However, we find that adding data from more subjects and datasets improves regression performance on subject 1 test data, without ever seeing training examples from that subject. We showcase these results in Figure 5. To further understand this effect, we compute between-subject correlations on the fMRI signals, observing that there are significant clusters of highly correlated responses even over brains from different subjects. We expand on these results in Section 4 of the supplement." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.77, + 0.363, + 0.785 + ], + "angle": 0, + "content": "5.4 Limitations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Figure 4 shows instances where our reconstruction method fails. Reconstructing complex, high spatial frequency videos (e.g., a crowd of people) remains challenging. In some cases, semantic reconstruction fails: we hypothesize that this is due" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.128 + ], + "angle": 0, + "content": "C. Fosco et al." + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.147, + 0.749, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.403, + 0.789, + 0.557 + ], + "angle": 0, + "content": "Fig. 5: fMRI Correlations and Zero-shot regression performance over latent vector \\(\\mathbf{z}\\). A) Between subject pairwise correlations for subject 01 are plotted in an inflated brain and flattened brain, showing left hemisphere (LH) and right hemisphere (RH). Both BMD and CC2017 show areas of high correlation across subjects, mostly clustered around visual areas. We hypothesize this correlation might enable zero-shot strategies. B) Zero-shot performance as data from more subjects are added. The left-most bar corresponds to MSE when training on subject 01's train set and evaluating on subject 01's test set. Subsequent bars train on different combinations of subjects except 01, and test on subject 01's test set. We observe that as we add more subjects and datasets, MSE decreases in both examples, hinting at the possibility of developing accurate zero-shot reconstruction with the right scale." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.589, + 0.789, + 0.651 + ], + "angle": 0, + "content": "to issues in the fMRI data, as our system is vulnerable to situations where the subject's mind wanders and loses focus on the stimulus while data is recorded. Temporally faithful reconstructions are also limited by the temporal capabilities of the available video generation models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.68, + 0.36, + 0.697 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.84 + ], + "angle": 0, + "content": "We present a generative framework to reconstruct short clips from human brain responses. We leverage large pre-trained text-to-video models and two fMRI datasets to train a multi-stage pipeline to reconstruct a video observed by a given subject. Future generative neurotechnology in this vein could help determining the information recovery limits of a brain region, an uncharted terrain of neuroscience. Such technology could foster the development of therapeutic treatments to restore or compensate for sensory processing deprivation, as well as allow for more efficient human-machine communication." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.269, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.174, + 0.787, + 0.234 + ], + "angle": 0, + "content": "This research was funded by the Multidisciplinary University Research Initiative (MURI) award by the Army Research Office (grant No. W911NF-23-1-0277) to A.O; the MIT EECS MathWorks Fellowship to B.L.; the MIT EECS MathWorks Fellowship to C.F." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.254, + 0.323, + 0.27 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.282, + 0.785, + 0.337 + ], + "angle": 0, + "content": "1. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22563-22575 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.785, + 0.392 + ], + "angle": 0, + "content": "2. Buccino, G., Binkofski, F., Fink, G.R., Fadiga, L., Fogassi, L., Gallese, V., Seitz, R.J., Zilles, K., Rizzolatti, G., Freund, H.J.: Action observation activates premotor and parietal areas in a somatotopic manner: an fmri study. European journal of neuroscience 13(2), 400-404 (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.393, + 0.785, + 0.447 + ], + "angle": 0, + "content": "3. Chen, Z., Qing, J., Xiang, T., Yue, W.L., Zhou, J.H.: Seeing beyond the brain: Conditional diffusion model with sparse masked modeling for vision decoding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22710-22720 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.448, + 0.785, + 0.473 + ], + "angle": 0, + "content": "4. Chen, Z., Qing, J., Zhou, J.H.: Cinematic mindscapes: High-quality video reconstruction from brain activity. arXiv preprint arXiv:2305.11675 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.474, + 0.785, + 0.514 + ], + "angle": 0, + "content": "5. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.515, + 0.785, + 0.54 + ], + "angle": 0, + "content": "6. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.542, + 0.785, + 0.595 + ], + "angle": 0, + "content": "7. Dickie, E.W., Anticevic, A., Smith, D.E., Coalson, T.S., Manogaran, M., Calarco, N., Viviano, J.D., Glasser, M.F., Van Essen, D.C., Voineskos, A.N.: Ciftify: A framework for surface-based analysis of legacy mr acquisitions. Neuroimage 197, 818-826 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.597, + 0.785, + 0.637 + ], + "angle": 0, + "content": "8. Gazzola, V., Keysers, C.: The observation and execution of actions share motor and somatosensory voxels in all tested subjects: single-subject analyses of unsmoothed fmri data. Cerebral cortex 19(6), 1239-1255 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.638, + 0.785, + 0.677 + ], + "angle": 0, + "content": "9. Glasser, M.F., Coalson, T.S., Robinson, E.C., Hacker, C.D., Harwell, J., Yacoub, E., Ugurbil, K., Andersson, J., Beckmann, C.F., Jenkinson, M., et al.: A multimodal parcellation of human cerebral cortex. Nature 536(7615), 171-178 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.679, + 0.785, + 0.732 + ], + "angle": 0, + "content": "10. Glasser, M.F., Sotiropoulos, S.N., Wilson, J.A., Coalson, T.S., Fischl, B., Andersson, J.L., Xu, J., Jbabdi, S., Webster, M., Polimeni, J.R., et al.: The minimal preprocessing pipelines for the human connectome project. Neuroimage 80, 105-124 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.733, + 0.785, + 0.772 + ], + "angle": 0, + "content": "11. Han, K., Wen, H., Shi, J., Lu, K.H., Zhang, Y., Fu, D., Liu, Z.: Variational autoencoder: An unsupervised model for encoding and decoding fmri activity in visual cortex. NeuroImage 198, 125-136 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.773, + 0.785, + 0.799 + ], + "angle": 0, + "content": "12. Hasson, U., Landesman, O., Knappmeyer, B., Vallines, I., Rubin, N., Heeger, D.J.: Neurocinematics: The neuroscience of film. Projections 2(1), 1-26 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. Hasson, U., Nir, Y., Levy, I., Fuhrmann, G., Malach, R.: Intersubject synchronization of cortical activity during natural vision. science 303(5664), 1634-1640 (2004)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.282, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "C. Fosco et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "14. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.217 + ], + "angle": 0, + "content": "15. Hendrycks, D., Gimpel, K.: Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.787, + 0.257 + ], + "angle": 0, + "content": "16. Kauppi, J.P., Jaëskeläinen, I.P., Sams, M., Tohka, J.: Inter-subject correlation of brain hemodynamic responses during watching a movie: localization in space and frequency. Frontiers in neuroinformatics 4, 669 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.257, + 0.787, + 0.298 + ], + "angle": 0, + "content": "17. Konen, C.S., Kastner, S.: Representation of eye movements and stimulus motion in topographically organized areas of human posterior parietal cortex. Journal of Neuroscience 28(33), 8361-8375 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.298, + 0.787, + 0.338 + ], + "angle": 0, + "content": "18. Kupershmidt, G., Belyi, R., Gaziv, G., Irani, M.: A penny for your (visual) thoughts: Self-supervised reconstruction of natural movies from brain activity. arXiv preprint arXiv:2206.03544 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.338, + 0.787, + 0.392 + ], + "angle": 0, + "content": "19. Lahner, B., Dwivedi, K., Iamshchinina, P., Graumann, M., Lascelles, A., Roig, G., Gifford, A.T., Pan, B., Jin, S., Ratan Murty, N.A., et al.: Bold moments: modeling short visual events through a video fmri dataset and metadata. bioRxiv pp. 2023-03 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.392, + 0.787, + 0.433 + ], + "angle": 0, + "content": "20. Le, A., Vesia, M., Yan, X., Crawford, J.D., Niemeier, M.: Parietal area ba7 integrates motor programs for reaching, grasping, and bimanual coordination. Journal of Neurophysiology (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.787, + 0.473 + ], + "angle": 0, + "content": "21. Le, L., Ambrogioni, L., Seeliger, K., Güçlütürk, Y., van Gerven, M., Güçlü, U.: Brain2pix: Fully convolutional naturalistic video frame reconstruction from brain activity. Frontiers in Neuroscience 16, 940972 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.514 + ], + "angle": 0, + "content": "22. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.787, + 0.555 + ], + "angle": 0, + "content": "23. Liu, Z., Guo, Y., Yu, K.: Diffvoice: Text-to-speech with latent diffusion. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.787, + 0.582 + ], + "angle": 0, + "content": "24. Logothetis, N.K., Sheinberg, D.L.: Visual object recognition. Annual review of neuroscience 19(1), 577-621 (1996)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.787, + 0.636 + ], + "angle": 0, + "content": "25. Monfort, M., Andonian, A., Zhou, B., Ramakrishnan, K., Bargal, S.A., Yan, T., Brown, L., Fan, Q., Gutfreund, D., Vondrick, C., et al.: Moments in time dataset: one million videos for event understanding. IEEE transactions on pattern analysis and machine intelligence 42(2), 502-508 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.636, + 0.787, + 0.704 + ], + "angle": 0, + "content": "26. Monfort, M., Pan, B., Ramakrishnan, K., Andonian, A., McNamara, B.A., Lascelles, A., Fan, Q., Gutfreund, D., Feris, R.S., Oliva, A.: Multi-moments in time: Learning and interpreting models for multi-action video understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9434-9445 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.787, + 0.744 + ], + "angle": 0, + "content": "27. Morrone, M.C., Tosetti, M., Montanaro, D., Fiorentini, A., Cioni, G., Burr, D.: A cortical area that responds specifically to optic flow, revealed by fmri. Nature neuroscience 3(12), 1322-1328 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.787, + 0.8 + ], + "angle": 0, + "content": "28. Newman, A., Fosco, C., Casser, V., Lee, A., McNamara, B., Oliva, A.: Multimodal memorability: Modeling effects of semantics and decay on video memorability. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16. pp. 223-240. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.841 + ], + "angle": 0, + "content": "29. Nishimoto, S., Vu, A.T., Naselaris, T., Benjamini, Y., Yu, B., Gallant, J.L.: Reconstructing visual experiences from brain activity evoked by natural movies. Current biology 21(19), 1641-1646 (2011)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.269, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "30. Peeters, R., Simone, L., Nelissen, K., Fabbri-Destro, M., Vanduffel, W., Rizzolatti, G., Orban, G.A.: The representation of tool use in humans and monkeys: common and uniquely human features. Journal of Neuroscience 29(37), 11523-11539 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.217 + ], + "angle": 0, + "content": "31. Peeters, R.R., Rizzolatti, G., Orban, G.A.: Functional properties of the left parietal tool use region. Neuroimage 78, 83-93 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.217, + 0.788, + 0.271 + ], + "angle": 0, + "content": "32. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.271, + 0.788, + 0.311 + ], + "angle": 0, + "content": "33. Rizzolatti, G., Sinigaglia, C.: The functional role of the parieto-frontal mirror circuit: interpretations and misinterpretations. Nature reviews neuroscience 11(4), 264-274 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.311, + 0.788, + 0.352 + ], + "angle": 0, + "content": "34. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.352, + 0.788, + 0.393 + ], + "angle": 0, + "content": "35. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.393, + 0.788, + 0.447 + ], + "angle": 0, + "content": "36. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.488 + ], + "angle": 0, + "content": "37. Sahara, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(4), 4713-4726 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.515 + ], + "angle": 0, + "content": "38. Schultz, J., Pilz, K.S.: Natural facial motion enhances cortical responses to faces. Experimental brain research 194, 465-475 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.515, + 0.788, + 0.569 + ], + "angle": 0, + "content": "39. Scotti, P., Banerjee, A., Goode, J., Shabalin, S., Nguyen, A., Dempster, A., Verlinde, N., Yundler, E., Weisberg, D., Norman, K., et al.: Reconstructing the mind's eye: fmri-to-image with contrastive learning and diffusion priors. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.597 + ], + "angle": 0, + "content": "40. Silver, M.A., Kastner, S.: Topographic maps in human frontal and parietal cortex. Trends in cognitive sciences 13(11), 488-495 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.597, + 0.788, + 0.651 + ], + "angle": 0, + "content": "41. Smith, S.M., Vidaurre, D., Beckmann, C.F., Glasser, M.F., Jenkinson, M., Miller, K.L., Nichols, T.E., Robinson, E.C., Salimi-Khorshidi, G., Woolrich, M.W., et al.: Functional connectomics from resting-state fmri. Trends in cognitive sciences 17(12), 666-682 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.651, + 0.788, + 0.678 + ], + "angle": 0, + "content": "42. Smith, T.J., Levin, D., Cutting, J.E.: A window on reality: Perceiving edited moving images. Current Directions in Psychological Science 21(2), 107-113 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.678, + 0.788, + 0.718 + ], + "angle": 0, + "content": "43. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.759 + ], + "angle": 0, + "content": "44. Takagi, Y., Nishimoto, S.: High-resolution image reconstruction with latent diffusion models from human brain activity. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14453-14463 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.759, + 0.788, + 0.8 + ], + "angle": 0, + "content": "45. Van Essen, D.C., Smith, S.M., Barch, D.M., Behrens, T.E., Yacoub, E., Ugurbil, K., Consortium, W.M.H., et al.: The wu-minn human connectome project: an overview. Neuroimage 80, 62-79 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.8, + 0.788, + 0.841 + ], + "angle": 0, + "content": "46. VanRullen, R., Thorpe, S.J.: The time course of visual processing: from early perception to decision-making. Journal of cognitive neuroscience 13(4), 454-461 (2001)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.128 + ], + "angle": 0, + "content": "C. Fosco et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "47. Wang, C., Yan, H., Huang, W., Li, J., Wang, Y., Fan, Y.S., Sheng, W., Liu, T., Li, R., Chen, H.: Reconstructing rapid natural vision with fmri-conditional video generative adversarial network. Cerebral Cortex 32(20), 4502-4511 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.786, + 0.217 + ], + "angle": 0, + "content": "48. Wang, J., Yuan, H., Chen, D., Zhang, Y., Wang, X., Zhang, S.: Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.244 + ], + "angle": 0, + "content": "49. Wang, L., Mruczek, R.E., Arcaro, M.J., Kastner, S.: Probabilistic maps of visual topography in human cortex. Cerebral cortex 25(10), 3911-3931 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.787, + 0.286 + ], + "angle": 0, + "content": "50. Wei, C., Fan, H., Xie, S., Wu, C.Y., Yuille, A., Feichtenhofer, C.: Masked feature prediction for self-supervised visual pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14668-14678 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.327 + ], + "angle": 0, + "content": "51. Wen, H., Shi, J., Zhang, Y., Lu, K.H., Cao, J., Liu, Z.: Neural encoding and decoding with deep learning for dynamic natural vision. Cerebral cortex 28(12), 4136-4160 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "52. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: A simple framework for masked image modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9653-9663 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.41 + ], + "angle": 0, + "content": "53. Yildirim, I., Wu, J., Kanwisher, N., Tenenbaum, J.: An integrative computational architecture for object-driven cortex. Current opinion in neurobiology 55, 73-81 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.787, + 0.439 + ], + "angle": 0, + "content": "54. Yu, K.P., Zhang, Z., Hu, F., Chai, J.: Efficient in-context learning in vision-language models for egocentric videos. arXiv preprint arXiv:2311.17041 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.787, + 0.481 + ], + "angle": 0, + "content": "55. Zhao, H., Torralba, A., Torresani, L., Yan, Z.: Hacs: Human action clips and segments dataset for recognition and temporal localization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8668-8678 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.508 + ], + "angle": 0, + "content": "56. Zhou, M., Gong, Z., Dai, Y., Wen, Y., Liu, Y., Zhen, Z.: A large-scale fmri dataset for human action recognition. Scientific Data 10(1), 415 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.508 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Brain Netflix_ Scaling Data to Reconstruct Videos from Brain Signals/9af6f00c-d99e-4633-82c3-bc4c9c80afb6_origin.pdf b/2024/Brain Netflix_ Scaling Data to Reconstruct Videos from Brain Signals/9af6f00c-d99e-4633-82c3-bc4c9c80afb6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d5040fd939dabffe423d0b5e673d243a7055f227 --- /dev/null +++ b/2024/Brain Netflix_ Scaling Data to Reconstruct Videos from Brain Signals/9af6f00c-d99e-4633-82c3-bc4c9c80afb6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d3351d48237745c2df4136b7ee5d41e0eab6198a204806d0b16f20bd8eeea41 +size 5802039 diff --git a/2024/Brain Netflix_ Scaling Data to Reconstruct Videos from Brain Signals/full.md b/2024/Brain Netflix_ Scaling Data to Reconstruct Videos from Brain Signals/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4acf46bbefad54e2e43e40a93d35e31d9537f867 --- /dev/null +++ b/2024/Brain Netflix_ Scaling Data to Reconstruct Videos from Brain Signals/full.md @@ -0,0 +1,227 @@ +# Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals + +Camilo Fosco\*1, Benjamin Lahner\*1, Bowen Pan $^{1}$ , Alex Andonian $^{1}$ , Emilie Josephs $^{1}$ , Alex Lascelles $^{1}$ , and Aude Oliva $^{1}$ + +CSAIL, Massachusetts Institute of Technology, MA, USA [camilolu, blahner, bpan, andonian, ejosephs, alexlasc, oliva]@mit.edu + +Abstract. The field of brain-to-stimuli reconstruction has seen significant progress in the last few years, but techniques continue to be subject-specific and are usually tested on a single dataset. In this work, we present a novel technique to reconstruct videos from functional Magnetic Resonance Imaging (fMRI) signals designed for performance across datasets and across human participants. Our pipeline accurately generates 2 and 3-second video clips from brain activity coming from distinct participants and different datasets by leveraging multi-dataset and multi-subject training. This helps us regress key latent and conditioning vectors for pretrained text-to-video and video-to-video models to reconstruct accurate videos that match the original stimuli observed by the participant. Key to our pipeline is the introduction of a 3-stage approach that first aligns fMRI signals to semantic embeddings, then regresses important vectors, and finally generates videos with those estimations. Our method demonstrates state-of-the-art reconstruction capabilities verified by qualitative and quantitative analyses, including crowd-sourced human evaluation. We showcase performance improvements across two datasets, as well as in multi-subject setups. Our ablation studies shed light on how different alignment strategies and data scaling decisions impact reconstruction performance, and we hint at a future for zero-shot reconstruction by analyzing how performance evolves as more subject data is leveraged. + +# 1 Introduction + +The human brain processes several million bits of information per second and compresses most of this information in ways that are still, in great part, unknown. What if we could partially recover this information by reading out brain signals? Reconstructing visual stimuli from brain activity is a promising way to investigate what information content is encoded in the human brain and how it may differ between participants. Reconstruction efforts typically require large amounts of neural-stimuli pairs to train high-parameter computational models, and inter-subject analyses additionally require repeated experiments across subjects. However, these neural-stimuli pairs are often not abundantly available due + +to their expensive and time-consuming collection process. Reconstructing dynamic videos, as opposed to static images, poses the additional challenges of temporal continuity. + +Previous fMRI (Functional Magnetic Resonance Imaging) work has shown that a greater extent of cortex responds to videos than still images [53] [38] [2], that some cortical regions are selective for motion features [2] [17] [27], and that inter-subject correlations to movie viewings are high [16] [13], all suggesting that fMRI reliably encodes dynamic content. Humans do not all perceive videos in the same way, however; in-the-wild longform videos, such as an unedited video of the happenings in a park, show markedly lower inter-subject correlations than edited cinematic films [12] [13]. These higher inter-subject correlations in cinematic films are largely due to production effects, such as camera angle, editing, and lighting, that structure everyone's attention and interpretation of the content in a similar way [42]. The everyday human visual experience includes events sampled at both ends of this continuum, from unstructured meanderings down a busy street to a structured viewing of a work presentation. Thus, while videos can be reconstructed from brain activity [4], cross-subject reconstructions and the structure of the video stimulus set have yet to be systematically studied in the video reconstruction domain. + +Here, we leverage two recent computational developments (latent diffusion models (LDMs) for video and masked-brain modeling) and several large-scale fMRI datasets to achieve high fidelity reconstructions of previously viewed 2- and 3-second videos. Latent diffusion models [34] are a class of generative models that generate high-fidelity images or videos, conditioned on inputs such as text, through a denoising diffusion process over a latent representation of the pixels. Masked brain modeling (MBM), an extension of masked training methods from natural language processing and computer vision to brain data, is a technique to learn robust self-supervised representations of brain signals by masking the signal and training a encoder-decoder network to reconstruct it. It has been shown to be a powerful way for a model to learn compressed representations of fMRI brain activity [4] [3]. + +We leverage these two concepts to propose a novel 3-stage pipeline that a) learns robust fMRI representations through masked brain modeling and alignment, b) regresses conditioning vectors from these representations, and c) reconstructs target videos with high fidelity, leveraging the estimated conditioning vectors. Importantly, our alignment and training strategy combines information from multiple subjects, as opposed to previous works that train one model per subject. This allows our model to not only improve its general reconstruction quality, but also to be able to generalize from one subject to another with much higher performance. Our contributions are as follows: + +1. We introduce a 3-stage pipeline for high-quality video reconstruction that leverages multiple fMRI datasets and aligns inter-subject data, outperforming existing approaches. +2. We showcase a detailed ablation study that illustrates how the model performs in different conditions, including modeling and data variants. + +3. We perform multi-subject and zero-shot reconstruction experiments to explore the possibility of a subject-agnostic visual reconstruction technology. + +Together, this work offers both modeling and neuroscience insights for video reconstruction from brain activity. + +![](images/27d56e2ed6556e21b6d50e4823ec760757794086a7c22bcfc8a56a829193f247.jpg) +A +B + +![](images/1b2ca13a3fb458028d0bd623528670bbaa519343b5670004bb535c22a6fe67f7.jpg) +Fig. 1: Reconstruction of short videos from fMRI activity A) Brain responses of subjects viewing video clips was measured with fMRI and then used to reconstruct the seen video. B) We show three examples of the ground truth video (top row) and our reconstruction using our approach (bottom row). + +# 2 Related Work + +Diffusion models: Latent diffusion models (LDMs) [34] are a class of generative model that achieves photorealistic generations through a process of denoising latent representations. This denoising process can be conditioned on other inputs, such as text, in order to guide the model toward generating outputs with specific semantic meaning of visual appearance. These models have been successful in reconstruction tasks, including image generation, super-resolution and recoloring [6, 35-37, 43], audio generation [23], among others. Recent text-to-video latent diffusion models have generated high fidelity reconstructions of short videos faithful to both content and temporal dynamics [1]. Here, we leverage these advances in video generation to form the backbone of our framework: we use Zeroscope V2, which creates 3 second videos based on text conditioning. + +Masked Brain modeling: Masked Brain Modeling [3] (MBM) is a recent method for pre-training fMRI reconstruction models. It is based on Masked Signal Modeling (MSM), a self-supervised learning task commonly used for pretraining in natural language processing and vision applications [5,14,50,52]. In + +MSM, some amount of the signal is masked during training, and an autoencoder is trained to predict the masked content, which can be fine tuned with additional training. In MBM, a percentage of the brain activity in masked and the model is tasked with reconstructing the complete input, which pushes the MBM model to learn compressed representations of fMRI brain activity. + +Video reconstruction from brain activity: An early naturalistic video reconstruction work used a voxelwise motion-energy model to recover videos with similar spatiotemporal energy, showing that the BOLD signal captures dynamic visual information [29]. Other approaches used convolutional neural networks (CNNs) to reconstruct videos from fMRI data [51] [21], with one even leveraging a convolutional encoder-decoder framework to synthesize high framerate fMRI-video frame pairs [18]. A work using GANs employed dual spatial and temporal discriminators to achieve spatiotemporally faithful reconstructions [47], and variational autoencoders have been used to reconstruct videos frame-by-frame from a hidden latent space [11]. Most similar to our approach, [4] uses masked-brain modeling to train a vision transformer encoder that then feeds latent representations into a diffusion model to reconstruct the video. + +# 3 Video fMRI Datasets + +We train and evaluate our reconstruction method on a compilation of four large-scale fMRI datasets [19] [56] [45] [51] to obtain a large quantity and diversity of brain responses. In order to effectively perform inter-subject and inter-dataset analyses we analyze all fMRI responses in Cifti fsLR32k space (see data preprocessing details in the supplement.) [10] [7]. This preprocessing pipeline accurately registers gray matter voxels to a shared cortical surface mesh, thereby allowing the model to learn informative spatial activation patterns across datasets. From the cortical surface, we select a subset of 41 regions of interest (ROIs) from the Glasser Atlas (see supplement for details) [9]. This subset of cortex is isolates brain regions that might respond to dynamic stimuli [46] [24] [20] [8] [33] [40] [30] [31] [49] as well as reduce the computational load of model training. All datasets combined, our task-based video fMRI data draws from 43 subjects, over 28,100 short video segments, and over 123,000 fMRI response trials, and our resting state fMRI data totals over one thousand hours across 1084 subjects. + +Human Connectome Project Dataset (HCP). The 1200-subject release of the Human Connectome Project [45] includes nearly an hour of resting state scans on 1084 subjects. During acquisition of resting state data, subjects were instructed to remain awake and fixate on a cross-hair but completed no task nor viewed any other visual stimulus. Resting state fMRI brain activity captures temporally correlated functional networks that can be useful for understanding cortical organization [41]. We utilize this time series data to pre-train our Masked Brain Modeling encoder (Figure 2). + +BOLD Moments Dataset (BMD). The BOLD Moments Dataset (BMD) [19] consists of fMRI responses to 1,102 3-second videos for ten human subjects collected in a rapid event-related design. Beta values were estimated at each + +cortical vertex for each video using a general linear model (GLM). The stimuli were divided into a non-overlapping 1,000 video training set and a 102 video testing set. Each subject viewed the training set videos 3 times and the testing set videos 10 times to facilitate both inter-subject and intra-subject analyses. The 1,102 videos were sampled from the 1 million video Moments in Time dataset [25] (overlapping with the Multi-Moments in Time [26] and Memento10k datasets [28]) and capture naturalistic, in-the-wild content like home videos. We use BMD's five sentence text description metadata provided for each video to generate ground truth text embeddings that our pipeline learns to estimate. To the best of our knowledge, this work is the first time BMD has been used for video reconstruction. + +Human Actions Dataset (HAD). The Human Actions Dataset (HAD) [56] is composed of 30 human subjects each viewing 720 two second videos sampled from the Human Action Clips and Segments (HACS) dataset [55]. Similar to BMD, data were collected in a rapid event-related design and beta values were estimated across cortex for each video. Each video is annotated with one of 180 human-centric actions, and each subject viewed four (different) videos from each of the 180 action categories. Video presentations were not repeated within or across subjects. For each subject, we define a testing/training split by randomly selecting one of the four videos per action category as a testing video. We use this dataset primarily to train Stage 1 in our pipeline. To the best of our knowledge, this work is the first time HAD has been used in video reconstruction. + +CC2017 Dataset (CC2017). In the CC2017 dataset [51], three subjects viewed continuous training and testing movies composed of shorter 10-15 second movie clips under a longform viewing paradigm. A training movie was composed from 374 video clips and randomly split into 18 eight minute segments. A testing movie was composed of 598 video clips and randomly split into five eight minute segments. Thus, each eight minute segment was composed of a concatenation of shorter video clips varying in duration. The clips are sourced from professionally shot videos, such as cinematic movies or stock advertising footage. The training and testing segments were repeated two and ten times per subject, respectively. Here, we divide the 8 minute segments into 2 second clips for reconstruction. We compose fMRI-video pairs by sampling the fMRI time series at an offset of 4 seconds (to account for the hemodynamic lag in the BOLD response) to capture peak BOLD activation corresponding to short snippets of the movie (2 second snippets, same as CC2017's acquisition TR). As this dataset is one of the main benchmarks used in previous work [4], we utilize this data extensively to train our regressors and provide reconstructions and evaluation metrics. + +# 4 Reconstruction Pipeline + +Approach Overview. Our reconstruction pipeline aims to regress key latent and conditioning vectors from brain signals to guide a pretrained video generation model towards reconstructing the original stimulus. This pipeline consists of + +![](images/ed4062954afc56f8fb606c03ca8236756801a8f01a1e4734036b28da16d5f687.jpg) +Fig. 2: Our proposed reconstruction framework. Stage 1: We train an encoder-decoder structure to reconstruct fMRI signals from a compressed representation. We then fine-tune the MBM encoder to align with a CLIP embedding of the video. Stage 2: we regress the latent vector (z) and blip embedding (b) of the target video from stage 1's aligned embedding. Stage 3: we use the regressed conditioning vectors to generate the video. The estimated vector z is artificially noised and then denoised with a pre-trained denoising U-Net, following a video-to-video procedure. The estimated vector b is sent to a pretrained BLIP model to generate a caption, which is used as a conditioning input to the U-Net. + +3 stages. First, in an initial alignment stage, an encoder-decoder architecture is trained to reconstruct masked fMRI embeddings while simultaneously following a CLIP-like contrastive learning objective. This setup forces the encoder to map fMRI inputs into an embedding that captures both knowledge about the raw signal and knowledge about the semantic properties of the original stimuli. Second, the output of stage 1's encoder is used in a regression stage: we regress the necessary inputs required by our generative model with two MLPs that estimate an initial latent vector $(z)$ and BLIP embeddings $(b)$ [22]. The BLIP embeddings are used to infer a conditioning caption. Third, our reconstruction stage uses the predicted latent vectors and BLIP embeddings to generate a video: Given a new fMRI input, we pass them through the regressors to predict the latent and conditioning vectors $(z$ and $b)$ , re-noise $z$ and use $b$ to generate a caption. We feed the noised $z$ and caption to a pretrained denoising U-Net (Zeroscope v2 [48]) to estimate the final reconstruction. We describe each stage in detail below and show a pseudocode implementation in the supplement Algorithm 1. + +Stage 1: MBM and Alignment. Our model employs a straightforward encoder-decoder framework aimed at reconstructing the input fMRI signal by compressing it into a 1024-dimensional latent vector. During this phase, the input fMRI signal is divided into several patches, and certain patches are masked. + +The encoder processes the signal with masked patches to produce a sequence of 1024-dimensional latent vectors. The decoder then attempts to transform these vectors back into the fMRI signal. Throughout the training process, the encoder-decoder structure is supervised through its ability to accurately reconstruct the masked patches. This approach, which we call Masked Brain Modeling (MBM), draws inspiration from the masked image modeling task [14] typical in the self-supervision literature. Its primary purpose is to acquaint the model with the spatial structure of the fMRI signal, rather than to imbue it with any semantic understanding. We first train our encoder-decoder pipeline on the masked modeling task alone with the large-scale HCP dataset [45] and HAD dataset [56], which ensures our embeddings learn basic fMRI structure. We utilize a simple MSE loss over the masked patches. + +After this step, we further finetune the MBM encoder through a contrastive learning task. Contrastive learning is a powerful self-supervised technique to learn representations across modalities. We extract CLIP embeddings from captions describing each video, and we apply a contrastive loss over batches of encoder embeddings and CLIP embeddings. This contrastive loss follows [32]: for each batch, the loss enforces high cosine similarity between an fMRI embedding and its matching CLIP embedding (positive pair), while minimizing the similarity between that same embedding and all other CLIP embeddings in the batch (negative pairs): + +$$ +\mathcal {L} _ {\text {c o n t r a s t i v e}} = - \sum_ {i = 1} ^ {N} \log \left(\frac {\exp \left(\frac {f _ {i} * c _ {i}}{\tau}\right)}{\sum_ {j = 1} ^ {N} \exp \left(\frac {f _ {j} * c _ {j}}{\tau}\right)}\right) +$$ + +Here, $f_{i}$ is the fMRI embedding (output from the MBM encoder), $c_{i}$ is the CLIP-text embedding, and $\tau$ is a temperature hyperparameter. We train this model by following the training details in [3] for MBM and [39] for the alignment procedure, where we use a encoder-decoder model with patch size of 16, hidden dimension of 1024 and 24 layers in the encoder. We use a temperature of 0.9 and a $75\%$ mask ratio following previous work [3]. + +Stage 2: Regression. To map from embeddings generated by the encoding model to latent and conditioning vectors suitable for the video generation pipeline, we utilize a multi-target regression model with regularization. Our regression models, instantiated as both a Ridge regression and a regularized MLP (see section 5.2 for comparisons), map outputs from Stage 1 into a latent vector $z$ of shape (4, 15, 33, 33) and a BLIP [22] embedding $b$ of shape (226, 768). The decision to map to BLIP vectors is rooted in how video generators handle noise in their conditioning inputs: similar to [44], we found that feeding a raw caption, even if imperfect, to our denoising U-Net performed better than feeding a regressed (and thus noisy) conditioning vector (comparisons in section 5.2). We hypothesize that the performance of video generators is susceptible to noise in the conditioning vector, and thus encapsulating the regression's imperfections at the level of BLIP's inputs allows the generator to receive clean (but imprecise) text tokens, which improves performance. + +Our MLPs are composed of an initial linear layer, followed by 3 residual blocks and a final output layer. The initial layer and the residual blocks contain 2048 units, dropout regularization with $p = 0.3$ , a GELU [15] activation and Batch Normalization. They are trained with an MSE loss and predict flattened vectors, which are then reshaped and used in Stage 3. Perfect regression at this stage equates to obtaining exact latent and conditioning vectors, which would result in a near-perfect reconstruction when fed to Stage 3. + +Stage 3: Reconstruction. During video reconstruction, we extract the latent vector $z$ and BLIP embedding $b$ using the pretrained MBM encoder followed by the $z$ and $b$ regressors from stages 1 and 2. $z$ undergoes a re-noising process, followed by a denoising procedure using a pretrained U-Net that follows [48]. We renoise the latents with a strength of 0.8 and apply 40 denoising steps, following insights from previous work [44]. Concurrently, $b$ is decoded using BLIP's decoder to produce a semantically relevant caption that conditions the U-Net during the denoising process. We observed during experiments that enforcing more descriptive and less repetitive captions tended to improve the quality of BLIP outputs, so we implement decoding with a high repetition penalty of 6, a minimum length of 4 and a maximum length of 20. The integration of $z$ and $b$ helps facilitate video generation that is both visually and contextually aligned with the video represented in the original fMRI signal. Our video generation model is frozen during this process. + +# 5 Experiments and Results + +# 5.1 Comparisons to previous work + +To showcase the quality of Brainflix, we obtain reconstructions from our full pipeline and compare to previous fMRI-to-video techniques. We report results over BMD and CC2017. + +Implementation Details. We use the large-scale fMRI datasets introduced in section 3 for model training and finetuning. For all datasets, videos are downsampled to 15 FPS and resized to $224 \times 224$ . We first train our alignment stage over the masked brain reconstruction task leveraging the HCP, HAD and BMD datasets for 200 epochs with a batch size of 300. We then finetune this model with a CLIP-like alignment objective over BMD and CC2017 simultaneously for 50 epochs with a batch size of 120. This aligns the output of the MBM encoder with embeddings of the captions describing each video. The captions are obtained through human labelers for BMD and synthetically for the rest of the datasets, leveraging the EILEV [54] video-to-caption model. Training for Stage 1 is done over 6 V100 GPUs, while training for Stage 2, a smaller scale endeavor given the smaller size of the MLPs, is performed over 2 Titan RTX GPUs. + +Evaluation Metrics. Following previous work [4, 44], we utilize 3 main evaluation metrics that aim to understand different characteristics of performance. To measure pixel-level reconstruction quality, we utilize Structural Similarity (SSIM). For semantic evaluation, we use the N-way top-k classification approach + +
MethodsCC2017BMD
SSIMMSE2-way50-waySSIMMSE2-way50-way
Kupershmidt0.128---0.0314.5610.5140.004
Mind-Video0.186-0.8530.2020.1760.7630.7110.101
Brainflux (Ours)0.1890.6810.8710.2030.1830.6910.7670.145
Brainfix multi-subject (Ours)0.1950.6550.8880.2210.1900.6710.8160.165
+ +Table 1: Quantitative comparison of our reconstruction performance against previous video reconstruction methods. Our methodology achieves state of the art on most metrics. Results from Kupershmidt and Mind-video on BMD are obtained through a reimplementation, as their code is not readily available. + +from [4], which measures how often the classification of a simple ImageNet classifier over the reconstruction and ground truth video match, limiting the output of the classifier to N classes. We declare a successful trial if the ground truth class is within the top-k probabilities outputted over the reconstruction, and repeat the test 100 times to report average success rates. We also report MSE results over our target latent embeddings $z$ and BLIP embeddings $b$ , to observe how close our regressed embeddings are to ground truth. + +Evaluation Datasets While we train on all datasets introduced in Section 3, we evaluate our reconstruction results on two large-scale video datasets, CC2017 [51] and BMD [19], to compare our method with previous work [4] [18] [47], as well as to introduce an additional dataset, BMD, with different subjects and video stimuli. These two datasets exhibit reasonable variety, complexity and semantic diversity, which makes them strong options for evaluation exercises. + +Results. We showcase our quantitative results over BMD and CC2017 in Table 1. Results are reported for Subject 1 in both datasets. Both our single-subject and our multi-subject approaches outperform previous work. Our multi-subject model, trained on all subjects from a given dataset before predicting on one, showcases larger boosts in performance which provides evidence towards the advantages of multi-subject training. We show qualitative results in Figure 3 and 4: we observe that our model is able to reconstruct examples from BMD and CC2017 with strong structural reliability. We hypothesize that our model's emphasis on regressing an accurate latent, a component that previous approaches lack, enforces accurate structural patterns that can then lead to reliable object positioning. + +Human Evaluation. To further measure how preferable our reconstructions are, we collect qualitative evaluations of semantic fidelity using human judgments. We designed a simple web experiment where participants indicated which of 6 reconstructed videos was the best semantic match to a real video. One of the 6 videos was a reconstruction of the reference video. We calculated the percent of trials where the selected video corresponded to the reconstruction of the reference video. Table 2 reports Brainflux results for two datasets. Our best performing method reached a semantic fidelity score of $86\%$ on BMD and $84\%$ on CC2017. This substantially exceeds chance $(16.66\%)$ , but remains below the + +![](images/883d668a79294f0947ffd41c6abe46ab6491301664cfd33a1457c6b4f1a448be.jpg) +Fig. 3: Comparison to previous works on examples from CC2017. We show the ground truth videos in the first row, our results in the second row, and previous approaches in the following rows. Our model manages to capture structural similarity in a better way than previous approaches. + +
MethodBMDCC2017
Baseline (Oracle)94.192.8
MindVideo [4]80.383.9
Brainflix (Ours)86.184.7
+ +Table 2: Human evaluation results. Accuracy on a 6-way retrieval task, where subjects select the ground truth video that most closely matches a reconstructed reference. + +
Training dataBMDCC2017
SSIMMSESSIMMSE
Subject 1 only0.1230.7210.6990.127
Subject 1 + all subs0.1290.7020.7100.131
Subject 2 only0.1200.7350.6800.122
Subject 2 + all subs0.1280.7220.7080.126
+ +Table 3: Multi-subject vs single-subject results. We observe that training on all subjects is value-adding for both datasets and both metrics. This effect is measurable across subjects. + +best possible performance expected from oracle reconstructions given our backbone model (Zeroscope V2 [48]), which ranged from $87.9\%$ to $98.0\%$ depending on caption and seed (mean $= 94.1\%$ ). + +# 5.2 Ablation studies. + +To test the validity of our proposed pipeline, we perform a thorough set of ablation studies to showcase how each modeling decision impacts our final method. Our ablations cover two main dimensions: Modeling strategies and data management. Modeling strategies include details related to the inner working of our pipeline's stages, model components, and supervision paradigms. Data management relates to the inclusion of data from additional datasets and different subjects. We report results over BMD, but utilize CC2017 and HAD as pretraining datasets in some experiments as described below. We analyze the impact of the following features: + +![](images/7aeb05e4c3bc3cb8c6c09109fdec7f1bc36f44c1d66759748eef930e7f2d0246.jpg) +Fig. 4: Success and failure cases of reconstructions. A) Additional success cases from BMD. B) Additional failure cases from BMD. C) Specific failure cases from BMD exemplifying semantic dominance (left), high frequency noise (middle), and superposition of indiscernible shapes onto correct reconstruction (right). D) Additional success cases of CC2017. E) Additional CC2017 failure cases. F) Failure cases on CC2017 exemplifying semantic dominance (left), poor facial reconstruction (middle), high frequency noise (right). G) CC2017 failure case: Sudden failure to reconstruct video despite showing previous success in the segment. + +- Impact of masked brain modeling: we measure how performance varies when we skip stage 1 and train a regressor from scratch to estimate the latent and conditioning vectors. +- Impact of semantic alignment: we observe how the contrastive learning objective affects the regression performance. +- Impact of regression backbone: we compare a linear regressor and a modern MLP. +- Impact of conditioning target: we analyze the difference between a direct regression to a 257x1024 text conditioning vector compared to regressing a BLIP vector used to build an estimated caption. +- Impact of data scaling: we observe how adding an additional pretraining dataset and multiple subjects affect the results. + +We report results in Table 4. Our simple pipeline without alignment is the worst performer: the simple regression model struggles to correctly estimate conditioning vectors, showcasing a large MSE. When switching to an MLP, we observe improved MSE, but the reconstructed vectors are still not good enough to correctly retrieve the video. The boost introduced by adding Stage 1, even if it consists only of MBM training, is quite significant and hints at the importance of a structural pretraining stage. + +Interestingly, regressing BLIP embeddings to then produce captions performs better than regressing raw conditioning vectors. The conditioning vectors accepted by zeroscope are hidden states from a CLIP-text model, of shape (257x1024). The BLIP embeddings, although of similar complexity, give a performance boost: we hypothesize that feeding the U-Net a clean caption, which BLIP produces even in the presence of a noisy input, allows the generation model to work in a familiar part of its latent space. This enables less deteriorated generations which translates to better metrics. Finally, scaling data showcases large improvements as well: both including HAD in Stage 1's pretraining and adding training data from other subjects gives boosts across metrics. + +Multi-subject vs. single subject training. Most previous work showcases reconstruction performance on a per-subject basis, as brain responses even to the same stimulus differs between observers [16] [13]. These discrepancies arise mostly from individual differences in brain composition, thought patterns, and the sensitivity of fMRI data collection. As indicated in Section 4, we explicitly train with multiple subjects in our pretraining and regression stages, attempting to build a model that can leverage training data not only from subject $s_0$ , but also $s_1 \ldots s_n$ to improve test set reconstruction performance for $s_0$ . To shed light on the effect of training on multiple subjects, we report results over subject 1 and subject 2, comparing performance between a pipeline trained solely on subject N's data and a pipeline trained on all subjects. We show results in Table 3. + +# 5.3 Zero-shot reconstruction + +In this section, we ask to what extent a model could reconstruct one's visual experience without training on their brain activity. This is a very challenging + +
MethodBMD
SSIMMSE2-way50-way
Oracle0.891-1.00.980
Random0.0511.6770.5010.041
Image-only reconstruction0.0410.7210.6870.125
Linear regression0.1110.8280.6710.120
MLP regressor0.1230.7210.6990.127
+ MBM, no CLIP alignment0.1380.7020.7010.131
+ CLIP alignment0.1780.7610.7150.134
+ BLIP instead of raw cond vector0.1800.7180.7210.140
+ HAD pretraining0.1830.7000.7670.145
Full pipeline (Brainfix multi-subject)0.1900.6710.8160.165
+ +Table 4: Ablation results and additional baselines on BMD's test set. We show results using: an Oracle baseline, where we feed the original video's ground truth $z$ and $b$ vectors to stage 3; a "random" baseline, where we instantiate random fMRI vectors and attempt to map those to $z$ and $b$ vectors, an image-only baseline, where, instead of using a video generator, we generate an image with an image generator (Stable Diffusion XL) from the fMRI, and duplicate the frame to match our videos' frame count; a pipeline without stage 1, using a linear regression for stage 2; a pipeline without stage 1, but with our final MLP backbone; a pipeline with a stage 1 comprised of masked-brain-modeling only; a pipeline with CLIP alignment, but still regressing a raw conditioning vector; a pipeline where we replace that regression target (raw conditioning vector) with a BLIP embedding; a pipeline with HAD pretraining, and our full pipeline, trained on multiple subjects. We observe that all design decisions tend to improve performance, and we see a significant boost from the addition of CLIP alignment. + +task, further complicated by the structure of the stimulus to reconstruct (see [13] [42] [12]). Current work, including this paper, reports results by training and testing on the same subject, but future applications of this technology would greatly benefit if they could make zero-shot predictions. We analyze the impact of scaling data in this regime: we first train a model on data from subject 2, and attempt to reconstruct test data from Subject 1. As expected, performance deteriorates. However, we find that adding data from more subjects and datasets improves regression performance on subject 1 test data, without ever seeing training examples from that subject. We showcase these results in Figure 5. To further understand this effect, we compute between-subject correlations on the fMRI signals, observing that there are significant clusters of highly correlated responses even over brains from different subjects. We expand on these results in Section 4 of the supplement. + +# 5.4 Limitations. + +Figure 4 shows instances where our reconstruction method fails. Reconstructing complex, high spatial frequency videos (e.g., a crowd of people) remains challenging. In some cases, semantic reconstruction fails: we hypothesize that this is due + +![](images/f71a145a1648bbc81410dfdd99a2693d1f7cd4107d47b443e056137ad4eb6f68.jpg) +Fig. 5: fMRI Correlations and Zero-shot regression performance over latent vector $\mathbf{z}$ . A) Between subject pairwise correlations for subject 01 are plotted in an inflated brain and flattened brain, showing left hemisphere (LH) and right hemisphere (RH). Both BMD and CC2017 show areas of high correlation across subjects, mostly clustered around visual areas. We hypothesize this correlation might enable zero-shot strategies. B) Zero-shot performance as data from more subjects are added. The left-most bar corresponds to MSE when training on subject 01's train set and evaluating on subject 01's test set. Subsequent bars train on different combinations of subjects except 01, and test on subject 01's test set. We observe that as we add more subjects and datasets, MSE decreases in both examples, hinting at the possibility of developing accurate zero-shot reconstruction with the right scale. + +to issues in the fMRI data, as our system is vulnerable to situations where the subject's mind wanders and loses focus on the stimulus while data is recorded. Temporally faithful reconstructions are also limited by the temporal capabilities of the available video generation models. + +# 6 Conclusion + +We present a generative framework to reconstruct short clips from human brain responses. We leverage large pre-trained text-to-video models and two fMRI datasets to train a multi-stage pipeline to reconstruct a video observed by a given subject. Future generative neurotechnology in this vein could help determining the information recovery limits of a brain region, an uncharted terrain of neuroscience. Such technology could foster the development of therapeutic treatments to restore or compensate for sensory processing deprivation, as well as allow for more efficient human-machine communication. + +# Acknowledgements + +This research was funded by the Multidisciplinary University Research Initiative (MURI) award by the Army Research Office (grant No. W911NF-23-1-0277) to A.O; the MIT EECS MathWorks Fellowship to B.L.; the MIT EECS MathWorks Fellowship to C.F. + +# References + +1. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22563-22575 (June 2023) +2. Buccino, G., Binkofski, F., Fink, G.R., Fadiga, L., Fogassi, L., Gallese, V., Seitz, R.J., Zilles, K., Rizzolatti, G., Freund, H.J.: Action observation activates premotor and parietal areas in a somatotopic manner: an fmri study. European journal of neuroscience 13(2), 400-404 (2001) +3. Chen, Z., Qing, J., Xiang, T., Yue, W.L., Zhou, J.H.: Seeing beyond the brain: Conditional diffusion model with sparse masked modeling for vision decoding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22710-22720 (2023) +4. Chen, Z., Qing, J., Zhou, J.H.: Cinematic mindscapes: High-quality video reconstruction from brain activity. arXiv preprint arXiv:2305.11675 (2023) +5. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) +6. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021) +7. Dickie, E.W., Anticevic, A., Smith, D.E., Coalson, T.S., Manogaran, M., Calarco, N., Viviano, J.D., Glasser, M.F., Van Essen, D.C., Voineskos, A.N.: Ciftify: A framework for surface-based analysis of legacy mr acquisitions. Neuroimage 197, 818-826 (2019) +8. Gazzola, V., Keysers, C.: The observation and execution of actions share motor and somatosensory voxels in all tested subjects: single-subject analyses of unsmoothed fmri data. Cerebral cortex 19(6), 1239-1255 (2009) +9. Glasser, M.F., Coalson, T.S., Robinson, E.C., Hacker, C.D., Harwell, J., Yacoub, E., Ugurbil, K., Andersson, J., Beckmann, C.F., Jenkinson, M., et al.: A multimodal parcellation of human cerebral cortex. Nature 536(7615), 171-178 (2016) +10. Glasser, M.F., Sotiropoulos, S.N., Wilson, J.A., Coalson, T.S., Fischl, B., Andersson, J.L., Xu, J., Jbabdi, S., Webster, M., Polimeni, J.R., et al.: The minimal preprocessing pipelines for the human connectome project. Neuroimage 80, 105-124 (2013) +11. Han, K., Wen, H., Shi, J., Lu, K.H., Zhang, Y., Fu, D., Liu, Z.: Variational autoencoder: An unsupervised model for encoding and decoding fmri activity in visual cortex. NeuroImage 198, 125-136 (2019) +12. Hasson, U., Landesman, O., Knappmeyer, B., Vallines, I., Rubin, N., Heeger, D.J.: Neurocinematics: The neuroscience of film. Projections 2(1), 1-26 (2008) +13. Hasson, U., Nir, Y., Levy, I., Fuhrmann, G., Malach, R.: Intersubject synchronization of cortical activity during natural vision. science 303(5664), 1634-1640 (2004) + +14. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022) +15. Hendrycks, D., Gimpel, K.: Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415 (2016) +16. Kauppi, J.P., Jaëskeläinen, I.P., Sams, M., Tohka, J.: Inter-subject correlation of brain hemodynamic responses during watching a movie: localization in space and frequency. Frontiers in neuroinformatics 4, 669 (2010) +17. Konen, C.S., Kastner, S.: Representation of eye movements and stimulus motion in topographically organized areas of human posterior parietal cortex. Journal of Neuroscience 28(33), 8361-8375 (2008) +18. Kupershmidt, G., Belyi, R., Gaziv, G., Irani, M.: A penny for your (visual) thoughts: Self-supervised reconstruction of natural movies from brain activity. arXiv preprint arXiv:2206.03544 (2022) +19. Lahner, B., Dwivedi, K., Iamshchinina, P., Graumann, M., Lascelles, A., Roig, G., Gifford, A.T., Pan, B., Jin, S., Ratan Murty, N.A., et al.: Bold moments: modeling short visual events through a video fmri dataset and metadata. bioRxiv pp. 2023-03 (2023) +20. Le, A., Vesia, M., Yan, X., Crawford, J.D., Niemeier, M.: Parietal area ba7 integrates motor programs for reaching, grasping, and bimanual coordination. Journal of Neurophysiology (2017) +21. Le, L., Ambrogioni, L., Seeliger, K., Güçlütürk, Y., van Gerven, M., Güçlü, U.: Brain2pix: Fully convolutional naturalistic video frame reconstruction from brain activity. Frontiers in Neuroscience 16, 940972 (2022) +22. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022) +23. Liu, Z., Guo, Y., Yu, K.: Diffvoice: Text-to-speech with latent diffusion. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023) +24. Logothetis, N.K., Sheinberg, D.L.: Visual object recognition. Annual review of neuroscience 19(1), 577-621 (1996) +25. Monfort, M., Andonian, A., Zhou, B., Ramakrishnan, K., Bargal, S.A., Yan, T., Brown, L., Fan, Q., Gutfreund, D., Vondrick, C., et al.: Moments in time dataset: one million videos for event understanding. IEEE transactions on pattern analysis and machine intelligence 42(2), 502-508 (2019) +26. Monfort, M., Pan, B., Ramakrishnan, K., Andonian, A., McNamara, B.A., Lascelles, A., Fan, Q., Gutfreund, D., Feris, R.S., Oliva, A.: Multi-moments in time: Learning and interpreting models for multi-action video understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9434-9445 (2021) +27. Morrone, M.C., Tosetti, M., Montanaro, D., Fiorentini, A., Cioni, G., Burr, D.: A cortical area that responds specifically to optic flow, revealed by fmri. Nature neuroscience 3(12), 1322-1328 (2000) +28. Newman, A., Fosco, C., Casser, V., Lee, A., McNamara, B., Oliva, A.: Multimodal memorability: Modeling effects of semantics and decay on video memorability. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16. pp. 223-240. Springer (2020) +29. Nishimoto, S., Vu, A.T., Naselaris, T., Benjamini, Y., Yu, B., Gallant, J.L.: Reconstructing visual experiences from brain activity evoked by natural movies. Current biology 21(19), 1641-1646 (2011) + +30. Peeters, R., Simone, L., Nelissen, K., Fabbri-Destro, M., Vanduffel, W., Rizzolatti, G., Orban, G.A.: The representation of tool use in humans and monkeys: common and uniquely human features. Journal of Neuroscience 29(37), 11523-11539 (2009) +31. Peeters, R.R., Rizzolatti, G., Orban, G.A.: Functional properties of the left parietal tool use region. Neuroimage 78, 83-93 (2013) +32. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +33. Rizzolatti, G., Sinigaglia, C.: The functional role of the parieto-frontal mirror circuit: interpretations and misinterpretations. Nature reviews neuroscience 11(4), 264-274 (2010) +34. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +35. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022) +36. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) +37. Sahara, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(4), 4713-4726 (2022) +38. Schultz, J., Pilz, K.S.: Natural facial motion enhances cortical responses to faces. Experimental brain research 194, 465-475 (2009) +39. Scotti, P., Banerjee, A., Goode, J., Shabalin, S., Nguyen, A., Dempster, A., Verlinde, N., Yundler, E., Weisberg, D., Norman, K., et al.: Reconstructing the mind's eye: fmri-to-image with contrastive learning and diffusion priors. Advances in Neural Information Processing Systems 36 (2024) +40. Silver, M.A., Kastner, S.: Topographic maps in human frontal and parietal cortex. Trends in cognitive sciences 13(11), 488-495 (2009) +41. Smith, S.M., Vidaurre, D., Beckmann, C.F., Glasser, M.F., Jenkinson, M., Miller, K.L., Nichols, T.E., Robinson, E.C., Salimi-Khorshidi, G., Woolrich, M.W., et al.: Functional connectomics from resting-state fmri. Trends in cognitive sciences 17(12), 666-682 (2013) +42. Smith, T.J., Levin, D., Cutting, J.E.: A window on reality: Perceiving edited moving images. Current Directions in Psychological Science 21(2), 107-113 (2012) +43. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020) +44. Takagi, Y., Nishimoto, S.: High-resolution image reconstruction with latent diffusion models from human brain activity. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14453-14463 (2023) +45. Van Essen, D.C., Smith, S.M., Barch, D.M., Behrens, T.E., Yacoub, E., Ugurbil, K., Consortium, W.M.H., et al.: The wu-minn human connectome project: an overview. Neuroimage 80, 62-79 (2013) +46. VanRullen, R., Thorpe, S.J.: The time course of visual processing: from early perception to decision-making. Journal of cognitive neuroscience 13(4), 454-461 (2001) + +47. Wang, C., Yan, H., Huang, W., Li, J., Wang, Y., Fan, Y.S., Sheng, W., Liu, T., Li, R., Chen, H.: Reconstructing rapid natural vision with fmri-conditional video generative adversarial network. Cerebral Cortex 32(20), 4502-4511 (2022) +48. Wang, J., Yuan, H., Chen, D., Zhang, Y., Wang, X., Zhang, S.: Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571 (2023) +49. Wang, L., Mruczek, R.E., Arcaro, M.J., Kastner, S.: Probabilistic maps of visual topography in human cortex. Cerebral cortex 25(10), 3911-3931 (2015) +50. Wei, C., Fan, H., Xie, S., Wu, C.Y., Yuille, A., Feichtenhofer, C.: Masked feature prediction for self-supervised visual pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14668-14678 (2022) +51. Wen, H., Shi, J., Zhang, Y., Lu, K.H., Cao, J., Liu, Z.: Neural encoding and decoding with deep learning for dynamic natural vision. Cerebral cortex 28(12), 4136-4160 (2018) +52. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: A simple framework for masked image modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9653-9663 (2022) +53. Yildirim, I., Wu, J., Kanwisher, N., Tenenbaum, J.: An integrative computational architecture for object-driven cortex. Current opinion in neurobiology 55, 73-81 (2019) +54. Yu, K.P., Zhang, Z., Hu, F., Chai, J.: Efficient in-context learning in vision-language models for egocentric videos. arXiv preprint arXiv:2311.17041 (2023) +55. Zhao, H., Torralba, A., Torresani, L., Yan, Z.: Hacs: Human action clips and segments dataset for recognition and temporal localization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8668-8678 (2019) +56. Zhou, M., Gong, Z., Dai, Y., Wen, Y., Liu, Y., Zhen, Z.: A large-scale fmri dataset for human action recognition. Scientific Data 10(1), 415 (2023) \ No newline at end of file diff --git a/2024/Brain Netflix_ Scaling Data to Reconstruct Videos from Brain Signals/images.zip b/2024/Brain Netflix_ Scaling Data to Reconstruct Videos from Brain Signals/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3d757b9ca9b7668835ecdceb3dc3ae3d9b63017c --- /dev/null +++ b/2024/Brain Netflix_ Scaling Data to Reconstruct Videos from Brain Signals/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f3e0b2db98c6574ae597cf0d6c8b352f2821359cb0cd724cc2cbb3655cfeacc3 +size 637151 diff --git a/2024/Brain Netflix_ Scaling Data to Reconstruct Videos from Brain Signals/layout.json b/2024/Brain Netflix_ Scaling Data to Reconstruct Videos from Brain Signals/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8732a4d25b896e80c9bc265a57dc36ec4df634a5 --- /dev/null +++ b/2024/Brain Netflix_ Scaling Data to Reconstruct Videos from Brain Signals/layout.json @@ -0,0 +1,7189 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 157, + 111, + 457, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 111, + 457, + 148 + ], + "spans": [ + { + "bbox": [ + 157, + 111, + 457, + 148 + ], + "type": "text", + "content": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 168, + 472, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 168, + 472, + 194 + ], + "spans": [ + { + "bbox": [ + 140, + 168, + 472, + 194 + ], + "type": "text", + "content": "Camilo Fosco\\*1, Benjamin Lahner\\*1, Bowen Pan" + }, + { + "bbox": [ + 140, + 168, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 168, + 472, + 194 + ], + "type": "text", + "content": ", Alex Andonian" + }, + { + "bbox": [ + 140, + 168, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 168, + 472, + 194 + ], + "type": "text", + "content": ", Emilie Josephs" + }, + { + "bbox": [ + 140, + 168, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 168, + 472, + 194 + ], + "type": "text", + "content": ", Alex Lascelles" + }, + { + "bbox": [ + 140, + 168, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 168, + 472, + 194 + ], + "type": "text", + "content": ", and Aude Oliva" + }, + { + "bbox": [ + 140, + 168, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 140, + 202, + 473, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 202, + 473, + 225 + ], + "spans": [ + { + "bbox": [ + 140, + 202, + 473, + 225 + ], + "type": "text", + "content": "CSAIL, Massachusetts Institute of Technology, MA, USA [camilolu, blahner, bpan, andonian, ejosephs, alexlasc, oliva]@mit.edu" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 262, + 455, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 262, + 455, + 504 + ], + "spans": [ + { + "bbox": [ + 160, + 262, + 455, + 504 + ], + "type": "text", + "content": "Abstract. The field of brain-to-stimuli reconstruction has seen significant progress in the last few years, but techniques continue to be subject-specific and are usually tested on a single dataset. In this work, we present a novel technique to reconstruct videos from functional Magnetic Resonance Imaging (fMRI) signals designed for performance across datasets and across human participants. Our pipeline accurately generates 2 and 3-second video clips from brain activity coming from distinct participants and different datasets by leveraging multi-dataset and multi-subject training. This helps us regress key latent and conditioning vectors for pretrained text-to-video and video-to-video models to reconstruct accurate videos that match the original stimuli observed by the participant. Key to our pipeline is the introduction of a 3-stage approach that first aligns fMRI signals to semantic embeddings, then regresses important vectors, and finally generates videos with those estimations. Our method demonstrates state-of-the-art reconstruction capabilities verified by qualitative and quantitative analyses, including crowd-sourced human evaluation. We showcase performance improvements across two datasets, as well as in multi-subject setups. Our ablation studies shed light on how different alignment strategies and data scaling decisions impact reconstruction performance, and we hint at a future for zero-shot reconstruction by analyzing how performance evolves as more subject data is leveraged." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 528, + 230, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 528, + 230, + 540 + ], + "spans": [ + { + "bbox": [ + 133, + 528, + 230, + 540 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "The human brain processes several million bits of information per second and compresses most of this information in ways that are still, in great part, unknown. What if we could partially recover this information by reading out brain signals? Reconstructing visual stimuli from brain activity is a promising way to investigate what information content is encoded in the human brain and how it may differ between participants. Reconstruction efforts typically require large amounts of neural-stimuli pairs to train high-parameter computational models, and inter-subject analyses additionally require repeated experiments across subjects. However, these neural-stimuli pairs are often not abundantly available due" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "text", + "content": "to their expensive and time-consuming collection process. Reconstructing dynamic videos, as opposed to static images, poses the additional challenges of temporal continuity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 342 + ], + "type": "text", + "content": "Previous fMRI (Functional Magnetic Resonance Imaging) work has shown that a greater extent of cortex responds to videos than still images [53] [38] [2], that some cortical regions are selective for motion features [2] [17] [27], and that inter-subject correlations to movie viewings are high [16] [13], all suggesting that fMRI reliably encodes dynamic content. Humans do not all perceive videos in the same way, however; in-the-wild longform videos, such as an unedited video of the happenings in a park, show markedly lower inter-subject correlations than edited cinematic films [12] [13]. These higher inter-subject correlations in cinematic films are largely due to production effects, such as camera angle, editing, and lighting, that structure everyone's attention and interpretation of the content in a similar way [42]. The everyday human visual experience includes events sampled at both ends of this continuum, from unstructured meanderings down a busy street to a structured viewing of a work presentation. Thus, while videos can be reconstructed from brain activity [4], cross-subject reconstructions and the structure of the video stimulus set have yet to be systematically studied in the video reconstruction domain." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 343, + 482, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 488 + ], + "type": "text", + "content": "Here, we leverage two recent computational developments (latent diffusion models (LDMs) for video and masked-brain modeling) and several large-scale fMRI datasets to achieve high fidelity reconstructions of previously viewed 2- and 3-second videos. Latent diffusion models [34] are a class of generative models that generate high-fidelity images or videos, conditioned on inputs such as text, through a denoising diffusion process over a latent representation of the pixels. Masked brain modeling (MBM), an extension of masked training methods from natural language processing and computer vision to brain data, is a technique to learn robust self-supervised representations of brain signals by masking the signal and training a encoder-decoder network to reconstruct it. It has been shown to be a powerful way for a model to learn compressed representations of fMRI brain activity [4] [3]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 487, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 596 + ], + "type": "text", + "content": "We leverage these two concepts to propose a novel 3-stage pipeline that a) learns robust fMRI representations through masked brain modeling and alignment, b) regresses conditioning vectors from these representations, and c) reconstructs target videos with high fidelity, leveraging the estimated conditioning vectors. Importantly, our alignment and training strategy combines information from multiple subjects, as opposed to previous works that train one model per subject. This allows our model to not only improve its general reconstruction quality, but also to be able to generalize from one subject to another with much higher performance. Our contributions are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 605, + 479, + 664 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 136, + 605, + 479, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 605, + 479, + 640 + ], + "spans": [ + { + "bbox": [ + 136, + 605, + 479, + 640 + ], + "type": "text", + "content": "1. We introduce a 3-stage pipeline for high-quality video reconstruction that leverages multiple fMRI datasets and aligns inter-subject data, outperforming existing approaches." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 641, + 479, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 641, + 479, + 664 + ], + "spans": [ + { + "bbox": [ + 136, + 641, + 479, + 664 + ], + "type": "text", + "content": "2. We showcase a detailed ablation study that illustrates how the model performs in different conditions, including modeling and data variants." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "C. Fosco et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 135, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 135, + 116, + 479, + 140 + ], + "type": "text", + "content": "3. We perform multi-subject and zero-shot reconstruction experiments to explore the possibility of a subject-agnostic visual reconstruction technology." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 479, + 175 + ], + "type": "text", + "content": "Together, this work offers both modeling and neuroscience insights for video reconstruction from brain activity." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 146, + 208, + 469, + 293 + ], + "blocks": [ + { + "bbox": [ + 137, + 199, + 146, + 208 + ], + "lines": [ + { + "bbox": [ + 137, + 199, + 146, + 208 + ], + "spans": [ + { + "bbox": [ + 137, + 199, + 146, + 208 + ], + "type": "text", + "content": "A" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 146, + 208, + 469, + 293 + ], + "lines": [ + { + "bbox": [ + 146, + 208, + 469, + 293 + ], + "spans": [ + { + "bbox": [ + 146, + 208, + 469, + 293 + ], + "type": "image", + "image_path": "27d56e2ed6556e21b6d50e4823ec760757794086a7c22bcfc8a56a829193f247.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 292, + 145, + 300 + ], + "lines": [ + { + "bbox": [ + 137, + 292, + 145, + 300 + ], + "spans": [ + { + "bbox": [ + 137, + 292, + 145, + 300 + ], + "type": "text", + "content": "B" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 149, + 306, + 477, + 360 + ], + "blocks": [ + { + "bbox": [ + 149, + 306, + 477, + 360 + ], + "lines": [ + { + "bbox": [ + 149, + 306, + 477, + 360 + ], + "spans": [ + { + "bbox": [ + 149, + 306, + 477, + 360 + ], + "type": "image", + "image_path": "1b2ca13a3fb458028d0bd623528670bbaa519343b5670004bb535c22a6fe67f7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 372, + 480, + 416 + ], + "lines": [ + { + "bbox": [ + 130, + 372, + 480, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 480, + 416 + ], + "type": "text", + "content": "Fig. 1: Reconstruction of short videos from fMRI activity A) Brain responses of subjects viewing video clips was measured with fMRI and then used to reconstruct the seen video. B) We show three examples of the ground truth video (top row) and our reconstruction using our approach (bottom row)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 460, + 236, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 236, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 236, + 472 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 485, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 481, + 616 + ], + "type": "text", + "content": "Diffusion models: Latent diffusion models (LDMs) [34] are a class of generative model that achieves photorealistic generations through a process of denoising latent representations. This denoising process can be conditioned on other inputs, such as text, in order to guide the model toward generating outputs with specific semantic meaning of visual appearance. These models have been successful in reconstruction tasks, including image generation, super-resolution and recoloring [6, 35-37, 43], audio generation [23], among others. Recent text-to-video latent diffusion models have generated high fidelity reconstructions of short videos faithful to both content and temporal dynamics [1]. Here, we leverage these advances in video generation to form the backbone of our framework: we use Zeroscope V2, which creates 3 second videos based on text conditioning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Masked Brain modeling: Masked Brain Modeling [3] (MBM) is a recent method for pre-training fMRI reconstruction models. It is based on Masked Signal Modeling (MSM), a self-supervised learning task commonly used for pretraining in natural language processing and vision applications [5,14,50,52]. In" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 164, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 164, + 91, + 447, + 102 + ], + "type": "text", + "content": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "MSM, some amount of the signal is masked during training, and an autoencoder is trained to predict the masked content, which can be fine tuned with additional training. In MBM, a percentage of the brain activity in masked and the model is tasked with reconstructing the complete input, which pushes the MBM model to learn compressed representations of fMRI brain activity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 482, + 319 + ], + "type": "text", + "content": "Video reconstruction from brain activity: An early naturalistic video reconstruction work used a voxelwise motion-energy model to recover videos with similar spatiotemporal energy, showing that the BOLD signal captures dynamic visual information [29]. Other approaches used convolutional neural networks (CNNs) to reconstruct videos from fMRI data [51] [21], with one even leveraging a convolutional encoder-decoder framework to synthesize high framerate fMRI-video frame pairs [18]. A work using GANs employed dual spatial and temporal discriminators to achieve spatiotemporally faithful reconstructions [47], and variational autoencoders have been used to reconstruct videos frame-by-frame from a hidden latent space [11]. Most similar to our approach, [4] uses masked-brain modeling to train a vision transformer encoder that then feeds latent representations into a diffusion model to reconstruct the video." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 339, + 282, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 339, + 282, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 339, + 282, + 353 + ], + "type": "text", + "content": "3 Video fMRI Datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 365, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 533 + ], + "type": "text", + "content": "We train and evaluate our reconstruction method on a compilation of four large-scale fMRI datasets [19] [56] [45] [51] to obtain a large quantity and diversity of brain responses. In order to effectively perform inter-subject and inter-dataset analyses we analyze all fMRI responses in Cifti fsLR32k space (see data preprocessing details in the supplement.) [10] [7]. This preprocessing pipeline accurately registers gray matter voxels to a shared cortical surface mesh, thereby allowing the model to learn informative spatial activation patterns across datasets. From the cortical surface, we select a subset of 41 regions of interest (ROIs) from the Glasser Atlas (see supplement for details) [9]. This subset of cortex is isolates brain regions that might respond to dynamic stimuli [46] [24] [20] [8] [33] [40] [30] [31] [49] as well as reduce the computational load of model training. All datasets combined, our task-based video fMRI data draws from 43 subjects, over 28,100 short video segments, and over 123,000 fMRI response trials, and our resting state fMRI data totals over one thousand hours across 1084 subjects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 534, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 629 + ], + "type": "text", + "content": "Human Connectome Project Dataset (HCP). The 1200-subject release of the Human Connectome Project [45] includes nearly an hour of resting state scans on 1084 subjects. During acquisition of resting state data, subjects were instructed to remain awake and fixate on a cross-hair but completed no task nor viewed any other visual stimulus. Resting state fMRI brain activity captures temporally correlated functional networks that can be useful for understanding cortical organization [41]. We utilize this time series data to pre-train our Masked Brain Modeling encoder (Figure 2)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 481, + 665 + ], + "type": "text", + "content": "BOLD Moments Dataset (BMD). The BOLD Moments Dataset (BMD) [19] consists of fMRI responses to 1,102 3-second videos for ten human subjects collected in a rapid event-related design. Beta values were estimated at each" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "C. Fosco et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "cortical vertex for each video using a general linear model (GLM). The stimuli were divided into a non-overlapping 1,000 video training set and a 102 video testing set. Each subject viewed the training set videos 3 times and the testing set videos 10 times to facilitate both inter-subject and intra-subject analyses. The 1,102 videos were sampled from the 1 million video Moments in Time dataset [25] (overlapping with the Multi-Moments in Time [26] and Memento10k datasets [28]) and capture naturalistic, in-the-wild content like home videos. We use BMD's five sentence text description metadata provided for each video to generate ground truth text embeddings that our pipeline learns to estimate. To the best of our knowledge, this work is the first time BMD has been used for video reconstruction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 247, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 390 + ], + "type": "text", + "content": "Human Actions Dataset (HAD). The Human Actions Dataset (HAD) [56] is composed of 30 human subjects each viewing 720 two second videos sampled from the Human Action Clips and Segments (HACS) dataset [55]. Similar to BMD, data were collected in a rapid event-related design and beta values were estimated across cortex for each video. Each video is annotated with one of 180 human-centric actions, and each subject viewed four (different) videos from each of the 180 action categories. Video presentations were not repeated within or across subjects. For each subject, we define a testing/training split by randomly selecting one of the four videos per action category as a testing video. We use this dataset primarily to train Stage 1 in our pipeline. To the best of our knowledge, this work is the first time HAD has been used in video reconstruction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 392, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 583 + ], + "type": "text", + "content": "CC2017 Dataset (CC2017). In the CC2017 dataset [51], three subjects viewed continuous training and testing movies composed of shorter 10-15 second movie clips under a longform viewing paradigm. A training movie was composed from 374 video clips and randomly split into 18 eight minute segments. A testing movie was composed of 598 video clips and randomly split into five eight minute segments. Thus, each eight minute segment was composed of a concatenation of shorter video clips varying in duration. The clips are sourced from professionally shot videos, such as cinematic movies or stock advertising footage. The training and testing segments were repeated two and ten times per subject, respectively. Here, we divide the 8 minute segments into 2 second clips for reconstruction. We compose fMRI-video pairs by sampling the fMRI time series at an offset of 4 seconds (to account for the hemodynamic lag in the BOLD response) to capture peak BOLD activation corresponding to short snippets of the movie (2 second snippets, same as CC2017's acquisition TR). As this dataset is one of the main benchmarks used in previous work [4], we utilize this data extensively to train our regressors and provide reconstructions and evaluation metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 603, + 297, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 603, + 297, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 603, + 297, + 616 + ], + "type": "text", + "content": "4 Reconstruction Pipeline" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Approach Overview. Our reconstruction pipeline aims to regress key latent and conditioning vectors from brain signals to guide a pretrained video generation model towards reconstructing the original stimulus. This pipeline consists of" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 164, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 164, + 91, + 447, + 102 + ], + "type": "text", + "content": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 473, + 304 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 473, + 304 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 473, + 304 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 473, + 304 + ], + "type": "image", + "image_path": "ed4062954afc56f8fb606c03ca8236756801a8f01a1e4734036b28da16d5f687.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 312, + 482, + 411 + ], + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 411 + ], + "type": "text", + "content": "Fig. 2: Our proposed reconstruction framework. Stage 1: We train an encoder-decoder structure to reconstruct fMRI signals from a compressed representation. We then fine-tune the MBM encoder to align with a CLIP embedding of the video. Stage 2: we regress the latent vector (z) and blip embedding (b) of the target video from stage 1's aligned embedding. Stage 3: we use the regressed conditioning vectors to generate the video. The estimated vector z is artificially noised and then denoised with a pre-trained denoising U-Net, following a video-to-video procedure. The estimated vector b is sent to a pretrained BLIP model to generate a caption, which is used as a conditioning input to the U-Net." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": "3 stages. First, in an initial alignment stage, an encoder-decoder architecture is trained to reconstruct masked fMRI embeddings while simultaneously following a CLIP-like contrastive learning objective. This setup forces the encoder to map fMRI inputs into an embedding that captures both knowledge about the raw signal and knowledge about the semantic properties of the original stimuli. Second, the output of stage 1's encoder is used in a regression stage: we regress the necessary inputs required by our generative model with two MLPs that estimate an initial latent vector " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "(z)" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": " and BLIP embeddings " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "(b)" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": " [22]. The BLIP embeddings are used to infer a conditioning caption. Third, our reconstruction stage uses the predicted latent vectors and BLIP embeddings to generate a video: Given a new fMRI input, we pass them through the regressors to predict the latent and conditioning vectors " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "(z" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "b)" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": ", re-noise " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": " and use " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": " to generate a caption. We feed the noised " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": " and caption to a pretrained denoising U-Net (Zeroscope v2 [48]) to estimate the final reconstruction. We describe each stage in detail below and show a pseudocode implementation in the supplement Algorithm 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Stage 1: MBM and Alignment. Our model employs a straightforward encoder-decoder framework aimed at reconstructing the input fMRI signal by compressing it into a 1024-dimensional latent vector. During this phase, the input fMRI signal is divided into several patches, and certain patches are masked." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "text", + "content": "C. Fosco et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "The encoder processes the signal with masked patches to produce a sequence of 1024-dimensional latent vectors. The decoder then attempts to transform these vectors back into the fMRI signal. Throughout the training process, the encoder-decoder structure is supervised through its ability to accurately reconstruct the masked patches. This approach, which we call Masked Brain Modeling (MBM), draws inspiration from the masked image modeling task [14] typical in the self-supervision literature. Its primary purpose is to acquaint the model with the spatial structure of the fMRI signal, rather than to imbue it with any semantic understanding. We first train our encoder-decoder pipeline on the masked modeling task alone with the large-scale HCP dataset [45] and HAD dataset [56], which ensures our embeddings learn basic fMRI structure. We utilize a simple MSE loss over the masked patches." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 261, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 369 + ], + "type": "text", + "content": "After this step, we further finetune the MBM encoder through a contrastive learning task. Contrastive learning is a powerful self-supervised technique to learn representations across modalities. We extract CLIP embeddings from captions describing each video, and we apply a contrastive loss over batches of encoder embeddings and CLIP embeddings. This contrastive loss follows [32]: for each batch, the loss enforces high cosine similarity between an fMRI embedding and its matching CLIP embedding (positive pair), while minimizing the similarity between that same embedding and all other CLIP embeddings in the batch (negative pairs):" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 212, + 380, + 402, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 380, + 402, + 416 + ], + "spans": [ + { + "bbox": [ + 212, + 380, + 402, + 416 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n t r a s t i v e}} = - \\sum_ {i = 1} ^ {N} \\log \\left(\\frac {\\exp \\left(\\frac {f _ {i} * c _ {i}}{\\tau}\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\frac {f _ {j} * c _ {j}}{\\tau}\\right)}\\right)", + "image_path": "a4569a00cf1816cc807a14decc13f0973876b9f5e9f2634f678167810c64ad0e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 425, + 482, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 497 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 425, + 482, + 497 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 130, + 425, + 482, + 497 + ], + "type": "text", + "content": " is the fMRI embedding (output from the MBM encoder), " + }, + { + "bbox": [ + 130, + 425, + 482, + 497 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 425, + 482, + 497 + ], + "type": "text", + "content": " is the CLIP-text embedding, and " + }, + { + "bbox": [ + 130, + 425, + 482, + 497 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 425, + 482, + 497 + ], + "type": "text", + "content": " is a temperature hyperparameter. We train this model by following the training details in [3] for MBM and [39] for the alignment procedure, where we use a encoder-decoder model with patch size of 16, hidden dimension of 1024 and 24 layers in the encoder. We use a temperature of 0.9 and a " + }, + { + "bbox": [ + 130, + 425, + 482, + 497 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 130, + 425, + 482, + 497 + ], + "type": "text", + "content": " mask ratio following previous work [3]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Stage 2: Regression. To map from embeddings generated by the encoding model to latent and conditioning vectors suitable for the video generation pipeline, we utilize a multi-target regression model with regularization. Our regression models, instantiated as both a Ridge regression and a regularized MLP (see section 5.2 for comparisons), map outputs from Stage 1 into a latent vector " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " of shape (4, 15, 33, 33) and a BLIP [22] embedding " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " of shape (226, 768). The decision to map to BLIP vectors is rooted in how video generators handle noise in their conditioning inputs: similar to [44], we found that feeding a raw caption, even if imperfect, to our denoising U-Net performed better than feeding a regressed (and thus noisy) conditioning vector (comparisons in section 5.2). We hypothesize that the performance of video generators is susceptible to noise in the conditioning vector, and thus encapsulating the regression's imperfections at the level of BLIP's inputs allows the generator to receive clean (but imprecise) text tokens, which improves performance." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 164, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 164, + 91, + 447, + 102 + ], + "type": "text", + "content": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "Our MLPs are composed of an initial linear layer, followed by 3 residual blocks and a final output layer. The initial layer and the residual blocks contain 2048 units, dropout regularization with " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "p = 0.3" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ", a GELU [15] activation and Batch Normalization. They are trained with an MSE loss and predict flattened vectors, which are then reshaped and used in Stage 3. Perfect regression at this stage equates to obtaining exact latent and conditioning vectors, which would result in a near-perfect reconstruction when fed to Stage 3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 483, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 483, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 483, + 369 + ], + "type": "text", + "content": "Stage 3: Reconstruction. During video reconstruction, we extract the latent vector " + }, + { + "bbox": [ + 130, + 200, + 483, + 369 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 200, + 483, + 369 + ], + "type": "text", + "content": " and BLIP embedding " + }, + { + "bbox": [ + 130, + 200, + 483, + 369 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 200, + 483, + 369 + ], + "type": "text", + "content": " using the pretrained MBM encoder followed by the " + }, + { + "bbox": [ + 130, + 200, + 483, + 369 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 200, + 483, + 369 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 200, + 483, + 369 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 200, + 483, + 369 + ], + "type": "text", + "content": " regressors from stages 1 and 2. " + }, + { + "bbox": [ + 130, + 200, + 483, + 369 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 200, + 483, + 369 + ], + "type": "text", + "content": " undergoes a re-noising process, followed by a denoising procedure using a pretrained U-Net that follows [48]. We renoise the latents with a strength of 0.8 and apply 40 denoising steps, following insights from previous work [44]. Concurrently, " + }, + { + "bbox": [ + 130, + 200, + 483, + 369 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 200, + 483, + 369 + ], + "type": "text", + "content": " is decoded using BLIP's decoder to produce a semantically relevant caption that conditions the U-Net during the denoising process. We observed during experiments that enforcing more descriptive and less repetitive captions tended to improve the quality of BLIP outputs, so we implement decoding with a high repetition penalty of 6, a minimum length of 4 and a maximum length of 20. The integration of " + }, + { + "bbox": [ + 130, + 200, + 483, + 369 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 200, + 483, + 369 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 200, + 483, + 369 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 200, + 483, + 369 + ], + "type": "text", + "content": " helps facilitate video generation that is both visually and contextually aligned with the video represented in the original fMRI signal. Our video generation model is frozen during this process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 388, + 304, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 388, + 304, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 388, + 304, + 402 + ], + "type": "text", + "content": "5 Experiments and Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 415, + 313, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 415, + 313, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 415, + 313, + 427 + ], + "type": "text", + "content": "5.1 Comparisons to previous work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 481, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 481, + 472 + ], + "type": "text", + "content": "To showcase the quality of Brainflix, we obtain reconstructions from our full pipeline and compare to previous fMRI-to-video techniques. We report results over BMD and CC2017." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 473, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 617 + ], + "type": "text", + "content": "Implementation Details. We use the large-scale fMRI datasets introduced in section 3 for model training and finetuning. For all datasets, videos are downsampled to 15 FPS and resized to " + }, + { + "bbox": [ + 130, + 473, + 482, + 617 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 473, + 482, + 617 + ], + "type": "text", + "content": ". We first train our alignment stage over the masked brain reconstruction task leveraging the HCP, HAD and BMD datasets for 200 epochs with a batch size of 300. We then finetune this model with a CLIP-like alignment objective over BMD and CC2017 simultaneously for 50 epochs with a batch size of 120. This aligns the output of the MBM encoder with embeddings of the captions describing each video. The captions are obtained through human labelers for BMD and synthetically for the rest of the datasets, leveraging the EILEV [54] video-to-caption model. Training for Stage 1 is done over 6 V100 GPUs, while training for Stage 2, a smaller scale endeavor given the smaller size of the MLPs, is performed over 2 Titan RTX GPUs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Evaluation Metrics. Following previous work [4, 44], we utilize 3 main evaluation metrics that aim to understand different characteristics of performance. To measure pixel-level reconstruction quality, we utilize Structural Similarity (SSIM). For semantic evaluation, we use the N-way top-k classification approach" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "text", + "content": "C. Fosco et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 114, + 482, + 191 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 191 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 191 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 191 + ], + "type": "table", + "html": "
MethodsCC2017BMD
SSIMMSE2-way50-waySSIMMSE2-way50-way
Kupershmidt0.128---0.0314.5610.5140.004
Mind-Video0.186-0.8530.2020.1760.7630.7110.101
Brainflux (Ours)0.1890.6810.8710.2030.1830.6910.7670.145
Brainfix multi-subject (Ours)0.1950.6550.8880.2210.1900.6710.8160.165
", + "image_path": "09eed28a4e56c47cecef472aea1371b54072c06f8284931bb376c07df7058a8e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 132, + 192, + 482, + 236 + ], + "lines": [ + { + "bbox": [ + 132, + 192, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 482, + 236 + ], + "type": "text", + "content": "Table 1: Quantitative comparison of our reconstruction performance against previous video reconstruction methods. Our methodology achieves state of the art on most metrics. Results from Kupershmidt and Mind-video on BMD are obtained through a reimplementation, as their code is not readily available." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 270, + 481, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 481, + 353 + ], + "type": "text", + "content": "from [4], which measures how often the classification of a simple ImageNet classifier over the reconstruction and ground truth video match, limiting the output of the classifier to N classes. We declare a successful trial if the ground truth class is within the top-k probabilities outputted over the reconstruction, and repeat the test 100 times to report average success rates. We also report MSE results over our target latent embeddings " + }, + { + "bbox": [ + 130, + 270, + 481, + 353 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 270, + 481, + 353 + ], + "type": "text", + "content": " and BLIP embeddings " + }, + { + "bbox": [ + 130, + 270, + 481, + 353 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 270, + 481, + 353 + ], + "type": "text", + "content": ", to observe how close our regressed embeddings are to ground truth." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 354, + 481, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 481, + 426 + ], + "type": "text", + "content": "Evaluation Datasets While we train on all datasets introduced in Section 3, we evaluate our reconstruction results on two large-scale video datasets, CC2017 [51] and BMD [19], to compare our method with previous work [4] [18] [47], as well as to introduce an additional dataset, BMD, with different subjects and video stimuli. These two datasets exhibit reasonable variety, complexity and semantic diversity, which makes them strong options for evaluation exercises." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 481, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 481, + 557 + ], + "type": "text", + "content": "Results. We showcase our quantitative results over BMD and CC2017 in Table 1. Results are reported for Subject 1 in both datasets. Both our single-subject and our multi-subject approaches outperform previous work. Our multi-subject model, trained on all subjects from a given dataset before predicting on one, showcases larger boosts in performance which provides evidence towards the advantages of multi-subject training. We show qualitative results in Figure 3 and 4: we observe that our model is able to reconstruct examples from BMD and CC2017 with strong structural reliability. We hypothesize that our model's emphasis on regressing an accurate latent, a component that previous approaches lack, enforces accurate structural patterns that can then lead to reliable object positioning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": "Human Evaluation. To further measure how preferable our reconstructions are, we collect qualitative evaluations of semantic fidelity using human judgments. We designed a simple web experiment where participants indicated which of 6 reconstructed videos was the best semantic match to a real video. One of the 6 videos was a reconstruction of the reference video. We calculated the percent of trials where the selected video corresponded to the reconstruction of the reference video. Table 2 reports Brainflux results for two datasets. Our best performing method reached a semantic fidelity score of " + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "inline_equation", + "content": "86\\%" + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": " on BMD and " + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "inline_equation", + "content": "84\\%" + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": " on CC2017. This substantially exceeds chance " + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "inline_equation", + "content": "(16.66\\%)" + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": ", but remains below the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "text", + "content": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 117, + 480, + 264 + ], + "blocks": [ + { + "bbox": [ + 134, + 117, + 480, + 264 + ], + "lines": [ + { + "bbox": [ + 134, + 117, + 480, + 264 + ], + "spans": [ + { + "bbox": [ + 134, + 117, + 480, + 264 + ], + "type": "image", + "image_path": "883d668a79294f0947ffd41c6abe46ab6491301664cfd33a1457c6b4f1a448be.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 272, + 482, + 317 + ], + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 317 + ], + "type": "text", + "content": "Fig. 3: Comparison to previous works on examples from CC2017. We show the ground truth videos in the first row, our results in the second row, and previous approaches in the following rows. Our model manages to capture structural similarity in a better way than previous approaches." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 132, + 329, + 267, + 381 + ], + "blocks": [ + { + "bbox": [ + 132, + 329, + 267, + 381 + ], + "lines": [ + { + "bbox": [ + 132, + 329, + 267, + 381 + ], + "spans": [ + { + "bbox": [ + 132, + 329, + 267, + 381 + ], + "type": "table", + "html": "
MethodBMDCC2017
Baseline (Oracle)94.192.8
MindVideo [4]80.383.9
Brainflix (Ours)86.184.7
", + "image_path": "231d08431dc66d2382428bd3894b353224e0836142bca5831af098f93fe0c6ac.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 287, + 327, + 483, + 406 + ], + "blocks": [ + { + "bbox": [ + 132, + 381, + 257, + 446 + ], + "lines": [ + { + "bbox": [ + 132, + 381, + 257, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 381, + 257, + 446 + ], + "type": "text", + "content": "Table 2: Human evaluation results. Accuracy on a 6-way retrieval task, where subjects select the ground truth video that most closely matches a reconstructed reference." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 287, + 327, + 483, + 406 + ], + "lines": [ + { + "bbox": [ + 287, + 327, + 483, + 406 + ], + "spans": [ + { + "bbox": [ + 287, + 327, + 483, + 406 + ], + "type": "table", + "html": "
Training dataBMDCC2017
SSIMMSESSIMMSE
Subject 1 only0.1230.7210.6990.127
Subject 1 + all subs0.1290.7020.7100.131
Subject 2 only0.1200.7350.6800.122
Subject 2 + all subs0.1280.7220.7080.126
", + "image_path": "9c4292b8e142d00f6507e150148bb31cfe6c551295755141a26a25e1849df48d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 287, + 406, + 482, + 450 + ], + "lines": [ + { + "bbox": [ + 287, + 406, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 287, + 406, + 482, + 450 + ], + "type": "text", + "content": "Table 3: Multi-subject vs single-subject results. We observe that training on all subjects is value-adding for both datasets and both metrics. This effect is measurable across subjects." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 483, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 521 + ], + "type": "text", + "content": "best possible performance expected from oracle reconstructions given our backbone model (Zeroscope V2 [48]), which ranged from " + }, + { + "bbox": [ + 130, + 483, + 482, + 521 + ], + "type": "inline_equation", + "content": "87.9\\%" + }, + { + "bbox": [ + 130, + 483, + 482, + 521 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 483, + 482, + 521 + ], + "type": "inline_equation", + "content": "98.0\\%" + }, + { + "bbox": [ + 130, + 483, + 482, + 521 + ], + "type": "text", + "content": " depending on caption and seed (mean " + }, + { + "bbox": [ + 130, + 483, + 482, + 521 + ], + "type": "inline_equation", + "content": "= 94.1\\%" + }, + { + "bbox": [ + 130, + 483, + 482, + 521 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 537, + 246, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 537, + 246, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 537, + 246, + 548 + ], + "type": "text", + "content": "5.2 Ablation studies." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "To test the validity of our proposed pipeline, we perform a thorough set of ablation studies to showcase how each modeling decision impacts our final method. Our ablations cover two main dimensions: Modeling strategies and data management. Modeling strategies include details related to the inner working of our pipeline's stages, model components, and supervision paradigms. Data management relates to the inclusion of data from additional datasets and different subjects. We report results over BMD, but utilize CC2017 and HAD as pretraining datasets in some experiments as described below. We analyze the impact of the following features:" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "text", + "content": "C. Fosco et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 118, + 459, + 608 + ], + "blocks": [ + { + "bbox": [ + 153, + 118, + 459, + 608 + ], + "lines": [ + { + "bbox": [ + 153, + 118, + 459, + 608 + ], + "spans": [ + { + "bbox": [ + 153, + 118, + 459, + 608 + ], + "type": "image", + "image_path": "7aeb05e4c3bc3cb8c6c09109fdec7f1bc36f44c1d66759748eef930e7f2d0246.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 619, + 482, + 708 + ], + "lines": [ + { + "bbox": [ + 130, + 619, + 482, + 708 + ], + "spans": [ + { + "bbox": [ + 130, + 619, + 482, + 708 + ], + "type": "text", + "content": "Fig. 4: Success and failure cases of reconstructions. A) Additional success cases from BMD. B) Additional failure cases from BMD. C) Specific failure cases from BMD exemplifying semantic dominance (left), high frequency noise (middle), and superposition of indiscernible shapes onto correct reconstruction (right). D) Additional success cases of CC2017. E) Additional CC2017 failure cases. F) Failure cases on CC2017 exemplifying semantic dominance (left), poor facial reconstruction (middle), high frequency noise (right). G) CC2017 failure case: Sudden failure to reconstruct video despite showing previous success in the segment." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 164, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 164, + 91, + 448, + 102 + ], + "type": "text", + "content": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 482, + 259 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 482, + 152 + ], + "type": "text", + "content": "- Impact of masked brain modeling: we measure how performance varies when we skip stage 1 and train a regressor from scratch to estimate the latent and conditioning vectors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 152, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 152, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 138, + 152, + 482, + 176 + ], + "type": "text", + "content": "- Impact of semantic alignment: we observe how the contrastive learning objective affects the regression performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 176, + 482, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 482, + 199 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 482, + 199 + ], + "type": "text", + "content": "- Impact of regression backbone: we compare a linear regressor and a modern MLP." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 200, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 200, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 138, + 200, + 482, + 236 + ], + "type": "text", + "content": "- Impact of conditioning target: we analyze the difference between a direct regression to a 257x1024 text conditioning vector compared to regressing a BLIP vector used to build an estimated caption." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 236, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 236, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 236, + 482, + 259 + ], + "type": "text", + "content": "- Impact of data scaling: we observe how adding an additional pretraining dataset and multiple subjects affect the results." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 268, + 482, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 482, + 352 + ], + "type": "text", + "content": "We report results in Table 4. Our simple pipeline without alignment is the worst performer: the simple regression model struggles to correctly estimate conditioning vectors, showcasing a large MSE. When switching to an MLP, we observe improved MSE, but the reconstructed vectors are still not good enough to correctly retrieve the video. The boost introduced by adding Stage 1, even if it consists only of MBM training, is quite significant and hints at the importance of a structural pretraining stage." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 352, + 482, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 471 + ], + "type": "text", + "content": "Interestingly, regressing BLIP embeddings to then produce captions performs better than regressing raw conditioning vectors. The conditioning vectors accepted by zeroscope are hidden states from a CLIP-text model, of shape (257x1024). The BLIP embeddings, although of similar complexity, give a performance boost: we hypothesize that feeding the U-Net a clean caption, which BLIP produces even in the presence of a noisy input, allows the generation model to work in a familiar part of its latent space. This enables less deteriorated generations which translates to better metrics. Finally, scaling data showcases large improvements as well: both including HAD in Stage 1's pretraining and adding training data from other subjects gives boosts across metrics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "text", + "content": "Multi-subject vs. single subject training. Most previous work showcases reconstruction performance on a per-subject basis, as brain responses even to the same stimulus differs between observers [16] [13]. These discrepancies arise mostly from individual differences in brain composition, thought patterns, and the sensitivity of fMRI data collection. As indicated in Section 4, we explicitly train with multiple subjects in our pretraining and regression stages, attempting to build a model that can leverage training data not only from subject " + }, + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "inline_equation", + "content": "s_0" + }, + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "text", + "content": ", but also " + }, + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "inline_equation", + "content": "s_1 \\ldots s_n" + }, + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "text", + "content": " to improve test set reconstruction performance for " + }, + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "inline_equation", + "content": "s_0" + }, + { + "bbox": [ + 130, + 472, + 482, + 604 + ], + "type": "text", + "content": ". To shed light on the effect of training on multiple subjects, we report results over subject 1 and subject 2, comparing performance between a pipeline trained solely on subject N's data and a pipeline trained on all subjects. We show results in Table 3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 621, + 285, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 285, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 285, + 632 + ], + "type": "text", + "content": "5.3 Zero-shot reconstruction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "In this section, we ask to what extent a model could reconstruct one's visual experience without training on their brain activity. This is a very challenging" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "C. Fosco et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 161, + 114, + 452, + 256 + ], + "blocks": [ + { + "bbox": [ + 161, + 114, + 452, + 256 + ], + "lines": [ + { + "bbox": [ + 161, + 114, + 452, + 256 + ], + "spans": [ + { + "bbox": [ + 161, + 114, + 452, + 256 + ], + "type": "table", + "html": "
MethodBMD
SSIMMSE2-way50-way
Oracle0.891-1.00.980
Random0.0511.6770.5010.041
Image-only reconstruction0.0410.7210.6870.125
Linear regression0.1110.8280.6710.120
MLP regressor0.1230.7210.6990.127
+ MBM, no CLIP alignment0.1380.7020.7010.131
+ CLIP alignment0.1780.7610.7150.134
+ BLIP instead of raw cond vector0.1800.7180.7210.140
+ HAD pretraining0.1830.7000.7670.145
Full pipeline (Brainfix multi-subject)0.1900.6710.8160.165
", + "image_path": "d7cd44c810529489ba0cc40335e4472b11d933447572d9634c0d34aeab8acc76.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "type": "text", + "content": "Table 4: Ablation results and additional baselines on BMD's test set. We show results using: an Oracle baseline, where we feed the original video's ground truth " + }, + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "type": "text", + "content": " vectors to stage 3; a \"random\" baseline, where we instantiate random fMRI vectors and attempt to map those to " + }, + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "type": "text", + "content": " vectors, an image-only baseline, where, instead of using a video generator, we generate an image with an image generator (Stable Diffusion XL) from the fMRI, and duplicate the frame to match our videos' frame count; a pipeline without stage 1, using a linear regression for stage 2; a pipeline without stage 1, but with our final MLP backbone; a pipeline with a stage 1 comprised of masked-brain-modeling only; a pipeline with CLIP alignment, but still regressing a raw conditioning vector; a pipeline where we replace that regression target (raw conditioning vector) with a BLIP embedding; a pipeline with HAD pretraining, and our full pipeline, trained on multiple subjects. We observe that all design decisions tend to improve performance, and we see a significant boost from the addition of CLIP alignment." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 436, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 594 + ], + "type": "text", + "content": "task, further complicated by the structure of the stimulus to reconstruct (see [13] [42] [12]). Current work, including this paper, reports results by training and testing on the same subject, but future applications of this technology would greatly benefit if they could make zero-shot predictions. We analyze the impact of scaling data in this regime: we first train a model on data from subject 2, and attempt to reconstruct test data from Subject 1. As expected, performance deteriorates. However, we find that adding data from more subjects and datasets improves regression performance on subject 1 test data, without ever seeing training examples from that subject. We showcase these results in Figure 5. To further understand this effect, we compute between-subject correlations on the fMRI signals, observing that there are significant clusters of highly correlated responses even over brains from different subjects. We expand on these results in Section 4 of the supplement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 609, + 222, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 222, + 621 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 222, + 621 + ], + "type": "text", + "content": "5.4 Limitations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Figure 4 shows instances where our reconstruction method fails. Reconstructing complex, high spatial frequency videos (e.g., a crowd of people) remains challenging. In some cases, semantic reconstruction fails: we hypothesize that this is due" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 164, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 164, + 91, + 447, + 102 + ], + "type": "text", + "content": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 116, + 458, + 308 + ], + "blocks": [ + { + "bbox": [ + 161, + 116, + 458, + 308 + ], + "lines": [ + { + "bbox": [ + 161, + 116, + 458, + 308 + ], + "spans": [ + { + "bbox": [ + 161, + 116, + 458, + 308 + ], + "type": "image", + "image_path": "f71a145a1648bbc81410dfdd99a2693d1f7cd4107d47b443e056137ad4eb6f68.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 319, + 482, + 441 + ], + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 441 + ], + "type": "text", + "content": "Fig. 5: fMRI Correlations and Zero-shot regression performance over latent vector " + }, + { + "bbox": [ + 130, + 319, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 319, + 482, + 441 + ], + "type": "text", + "content": ". A) Between subject pairwise correlations for subject 01 are plotted in an inflated brain and flattened brain, showing left hemisphere (LH) and right hemisphere (RH). Both BMD and CC2017 show areas of high correlation across subjects, mostly clustered around visual areas. We hypothesize this correlation might enable zero-shot strategies. B) Zero-shot performance as data from more subjects are added. The left-most bar corresponds to MSE when training on subject 01's train set and evaluating on subject 01's test set. Subsequent bars train on different combinations of subjects except 01, and test on subject 01's test set. We observe that as we add more subjects and datasets, MSE decreases in both examples, hinting at the possibility of developing accurate zero-shot reconstruction with the right scale." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 466, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 466, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 466, + 482, + 515 + ], + "type": "text", + "content": "to issues in the fMRI data, as our system is vulnerable to situations where the subject's mind wanders and loses focus on the stimulus while data is recorded. Temporally faithful reconstructions are also limited by the temporal capabilities of the available video generation models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 538, + 220, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 538, + 220, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 538, + 220, + 552 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "We present a generative framework to reconstruct short clips from human brain responses. We leverage large pre-trained text-to-video models and two fMRI datasets to train a multi-stage pipeline to reconstruct a video observed by a given subject. Future generative neurotechnology in this vein could help determining the information recovery limits of a brain region, an uncharted terrain of neuroscience. Such technology could foster the development of therapeutic treatments to restore or compensate for sensory processing deprivation, as well as allow for more efficient human-machine communication." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "text", + "content": "C. Fosco et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 481, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 481, + 185 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 481, + 185 + ], + "type": "text", + "content": "This research was funded by the Multidisciplinary University Research Initiative (MURI) award by the Army Research Office (grant No. W911NF-23-1-0277) to A.O; the MIT EECS MathWorks Fellowship to B.L.; the MIT EECS MathWorks Fellowship to C.F." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 201, + 197, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 201, + 197, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 201, + 197, + 213 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 223, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 223, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 223, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 223, + 480, + 266 + ], + "type": "text", + "content": "1. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22563-22575 (June 2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 267, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 480, + 310 + ], + "type": "text", + "content": "2. Buccino, G., Binkofski, F., Fink, G.R., Fadiga, L., Fogassi, L., Gallese, V., Seitz, R.J., Zilles, K., Rizzolatti, G., Freund, H.J.: Action observation activates premotor and parietal areas in a somatotopic manner: an fmri study. European journal of neuroscience 13(2), 400-404 (2001)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 311, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 311, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 311, + 480, + 354 + ], + "type": "text", + "content": "3. Chen, Z., Qing, J., Xiang, T., Yue, W.L., Zhou, J.H.: Seeing beyond the brain: Conditional diffusion model with sparse masked modeling for vision decoding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22710-22720 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 354, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 480, + 374 + ], + "type": "text", + "content": "4. Chen, Z., Qing, J., Zhou, J.H.: Cinematic mindscapes: High-quality video reconstruction from brain activity. arXiv preprint arXiv:2305.11675 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 375, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 375, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 375, + 480, + 407 + ], + "type": "text", + "content": "5. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 407, + 480, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 480, + 427 + ], + "type": "text", + "content": "6. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 429, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 480, + 471 + ], + "type": "text", + "content": "7. Dickie, E.W., Anticevic, A., Smith, D.E., Coalson, T.S., Manogaran, M., Calarco, N., Viviano, J.D., Glasser, M.F., Van Essen, D.C., Voineskos, A.N.: Ciftify: A framework for surface-based analysis of legacy mr acquisitions. Neuroimage 197, 818-826 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 472, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 480, + 504 + ], + "type": "text", + "content": "8. Gazzola, V., Keysers, C.: The observation and execution of actions share motor and somatosensory voxels in all tested subjects: single-subject analyses of unsmoothed fmri data. Cerebral cortex 19(6), 1239-1255 (2009)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 505, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 505, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 505, + 480, + 536 + ], + "type": "text", + "content": "9. Glasser, M.F., Coalson, T.S., Robinson, E.C., Hacker, C.D., Harwell, J., Yacoub, E., Ugurbil, K., Andersson, J., Beckmann, C.F., Jenkinson, M., et al.: A multimodal parcellation of human cerebral cortex. Nature 536(7615), 171-178 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 537, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 537, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 134, + 537, + 480, + 579 + ], + "type": "text", + "content": "10. Glasser, M.F., Sotiropoulos, S.N., Wilson, J.A., Coalson, T.S., Fischl, B., Andersson, J.L., Xu, J., Jbabdi, S., Webster, M., Polimeni, J.R., et al.: The minimal preprocessing pipelines for the human connectome project. Neuroimage 80, 105-124 (2013)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 580, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 580, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 134, + 580, + 480, + 611 + ], + "type": "text", + "content": "11. Han, K., Wen, H., Shi, J., Lu, K.H., Zhang, Y., Fu, D., Liu, Z.: Variational autoencoder: An unsupervised model for encoding and decoding fmri activity in visual cortex. NeuroImage 198, 125-136 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 612, + 480, + 632 + ], + "type": "text", + "content": "12. Hasson, U., Landesman, O., Knappmeyer, B., Vallines, I., Rubin, N., Heeger, D.J.: Neurocinematics: The neuroscience of film. Projections 2(1), 1-26 (2008)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "13. Hasson, U., Nir, Y., Levy, I., Fuhrmann, G., Malach, R.: Intersubject synchronization of cortical activity during natural vision. science 303(5664), 1634-1640 (2004)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 164, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 164, + 91, + 447, + 102 + ], + "type": "text", + "content": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "14. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 171 + ], + "type": "text", + "content": "15. Hendrycks, D., Gimpel, K.: Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415 (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 481, + 203 + ], + "type": "text", + "content": "16. Kauppi, J.P., Jaëskeläinen, I.P., Sams, M., Tohka, J.: Inter-subject correlation of brain hemodynamic responses during watching a movie: localization in space and frequency. Frontiers in neuroinformatics 4, 669 (2010)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 203, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 203, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 203, + 481, + 236 + ], + "type": "text", + "content": "17. Konen, C.S., Kastner, S.: Representation of eye movements and stimulus motion in topographically organized areas of human posterior parietal cortex. Journal of Neuroscience 28(33), 8361-8375 (2008)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 236, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 481, + 267 + ], + "type": "text", + "content": "18. Kupershmidt, G., Belyi, R., Gaziv, G., Irani, M.: A penny for your (visual) thoughts: Self-supervised reconstruction of natural movies from brain activity. arXiv preprint arXiv:2206.03544 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 267, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 267, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 133, + 267, + 481, + 310 + ], + "type": "text", + "content": "19. Lahner, B., Dwivedi, K., Iamshchinina, P., Graumann, M., Lascelles, A., Roig, G., Gifford, A.T., Pan, B., Jin, S., Ratan Murty, N.A., et al.: Bold moments: modeling short visual events through a video fmri dataset and metadata. bioRxiv pp. 2023-03 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 310, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 481, + 342 + ], + "type": "text", + "content": "20. Le, A., Vesia, M., Yan, X., Crawford, J.D., Niemeier, M.: Parietal area ba7 integrates motor programs for reaching, grasping, and bimanual coordination. Journal of Neurophysiology (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "type": "text", + "content": "21. Le, L., Ambrogioni, L., Seeliger, K., Güçlütürk, Y., van Gerven, M., Güçlü, U.: Brain2pix: Fully convolutional naturalistic video frame reconstruction from brain activity. Frontiers in Neuroscience 16, 940972 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "text", + "content": "22. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "type": "text", + "content": "23. Liu, Z., Guo, Y., Yu, K.: Diffvoice: Text-to-speech with latent diffusion. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 439, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 460 + ], + "type": "text", + "content": "24. Logothetis, N.K., Sheinberg, D.L.: Visual object recognition. Annual review of neuroscience 19(1), 577-621 (1996)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 460, + 481, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 481, + 503 + ], + "type": "text", + "content": "25. Monfort, M., Andonian, A., Zhou, B., Ramakrishnan, K., Bargal, S.A., Yan, T., Brown, L., Fan, Q., Gutfreund, D., Vondrick, C., et al.: Moments in time dataset: one million videos for event understanding. IEEE transactions on pattern analysis and machine intelligence 42(2), 502-508 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 503, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 481, + 557 + ], + "type": "text", + "content": "26. Monfort, M., Pan, B., Ramakrishnan, K., Andonian, A., McNamara, B.A., Lascelles, A., Fan, Q., Gutfreund, D., Feris, R.S., Oliva, A.: Multi-moments in time: Learning and interpreting models for multi-action video understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9434-9445 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 557, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 481, + 589 + ], + "type": "text", + "content": "27. Morrone, M.C., Tosetti, M., Montanaro, D., Fiorentini, A., Cioni, G., Burr, D.: A cortical area that responds specifically to optic flow, revealed by fmri. Nature neuroscience 3(12), 1322-1328 (2000)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 589, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 633 + ], + "type": "text", + "content": "28. Newman, A., Fosco, C., Casser, V., Lee, A., McNamara, B., Oliva, A.: Multimodal memorability: Modeling effects of semantics and decay on video memorability. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16. pp. 223-240. Springer (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "text", + "content": "29. Nishimoto, S., Vu, A.T., Naselaris, T., Benjamini, Y., Yu, B., Gallant, J.L.: Reconstructing visual experiences from brain activity evoked by natural movies. Current biology 21(19), 1641-1646 (2011)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "C. Fosco et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "30. Peeters, R., Simone, L., Nelissen, K., Fabbri-Destro, M., Vanduffel, W., Rizzolatti, G., Orban, G.A.: The representation of tool use in humans and monkeys: common and uniquely human features. Journal of Neuroscience 29(37), 11523-11539 (2009)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 171 + ], + "type": "text", + "content": "31. Peeters, R.R., Rizzolatti, G., Orban, G.A.: Functional properties of the left parietal tool use region. Neuroimage 78, 83-93 (2013)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 171, + 482, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 171, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 171, + 482, + 214 + ], + "type": "text", + "content": "32. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 214, + 482, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 246 + ], + "type": "text", + "content": "33. Rizzolatti, G., Sinigaglia, C.: The functional role of the parieto-frontal mirror circuit: interpretations and misinterpretations. Nature reviews neuroscience 11(4), 264-274 (2010)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 246, + 482, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 278 + ], + "type": "text", + "content": "34. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 278, + 482, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 311 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 311 + ], + "type": "text", + "content": "35. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 311, + 482, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 354 + ], + "type": "text", + "content": "36. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 354, + 482, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 386 + ], + "type": "text", + "content": "37. Sahara, C., Ho, J., Chan, W., Salimans, T., Fleet, D.J., Norouzi, M.: Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(4), 4713-4726 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 386, + 482, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 407 + ], + "type": "text", + "content": "38. Schultz, J., Pilz, K.S.: Natural facial motion enhances cortical responses to faces. Experimental brain research 194, 465-475 (2009)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 407, + 482, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 450 + ], + "type": "text", + "content": "39. Scotti, P., Banerjee, A., Goode, J., Shabalin, S., Nguyen, A., Dempster, A., Verlinde, N., Yundler, E., Weisberg, D., Norman, K., et al.: Reconstructing the mind's eye: fmri-to-image with contrastive learning and diffusion priors. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 450, + 482, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 472 + ], + "type": "text", + "content": "40. Silver, M.A., Kastner, S.: Topographic maps in human frontal and parietal cortex. Trends in cognitive sciences 13(11), 488-495 (2009)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 472, + 482, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 515 + ], + "type": "text", + "content": "41. Smith, S.M., Vidaurre, D., Beckmann, C.F., Glasser, M.F., Jenkinson, M., Miller, K.L., Nichols, T.E., Robinson, E.C., Salimi-Khorshidi, G., Woolrich, M.W., et al.: Functional connectomics from resting-state fmri. Trends in cognitive sciences 17(12), 666-682 (2013)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 515, + 482, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 482, + 536 + ], + "type": "text", + "content": "42. Smith, T.J., Levin, D., Cutting, J.E.: A window on reality: Perceiving edited moving images. Current Directions in Psychological Science 21(2), 107-113 (2012)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "type": "text", + "content": "43. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "type": "text", + "content": "44. Takagi, Y., Nishimoto, S.: High-resolution image reconstruction with latent diffusion models from human brain activity. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14453-14463 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "type": "text", + "content": "45. Van Essen, D.C., Smith, S.M., Barch, D.M., Behrens, T.E., Yacoub, E., Ugurbil, K., Consortium, W.M.H., et al.: The wu-minn human connectome project: an overview. Neuroimage 80, 62-79 (2013)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 633, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 633, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 633, + 482, + 666 + ], + "type": "text", + "content": "46. VanRullen, R., Thorpe, S.J.: The time course of visual processing: from early perception to decision-making. Journal of cognitive neuroscience 13(4), 454-461 (2001)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 164, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 164, + 91, + 447, + 102 + ], + "type": "text", + "content": "Brain Netflix: Scaling Data to Reconstruct Videos from Brain Signals" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 402 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "47. Wang, C., Yan, H., Huang, W., Li, J., Wang, Y., Fan, Y.S., Sheng, W., Liu, T., Li, R., Chen, H.: Reconstructing rapid natural vision with fmri-conditional video generative adversarial network. Cerebral Cortex 32(20), 4502-4511 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "type": "text", + "content": "48. Wang, J., Yuan, H., Chen, D., Zhang, Y., Wang, X., Zhang, S.: Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "text", + "content": "49. Wang, L., Mruczek, R.E., Arcaro, M.J., Kastner, S.: Probabilistic maps of visual topography in human cortex. Cerebral cortex 25(10), 3911-3931 (2015)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "50. Wei, C., Fan, H., Xie, S., Wu, C.Y., Yuille, A., Feichtenhofer, C.: Masked feature prediction for self-supervised visual pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14668-14678 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "51. Wen, H., Shi, J., Zhang, Y., Lu, K.H., Cao, J., Liu, Z.: Neural encoding and decoding with deep learning for dynamic natural vision. Cerebral cortex 28(12), 4136-4160 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "52. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: A simple framework for masked image modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9653-9663 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 324 + ], + "type": "text", + "content": "53. Yildirim, I., Wu, J., Kanwisher, N., Tenenbaum, J.: An integrative computational architecture for object-driven cortex. Current opinion in neurobiology 55, 73-81 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 347 + ], + "type": "text", + "content": "54. Yu, K.P., Zhang, Z., Hu, F., Chai, J.: Efficient in-context learning in vision-language models for egocentric videos. arXiv preprint arXiv:2311.17041 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "type": "text", + "content": "55. Zhao, H., Torralba, A., Torresani, L., Yan, Z.: Hacs: Human action clips and segments dataset for recognition and temporal localization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8668-8678 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "text", + "content": "56. Zhou, M., Gong, Z., Dai, Y., Wen, Y., Liu, Y., Zhen, Z.: A large-scale fmri dataset for human action recognition. Scientific Data 10(1), 415 (2023)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "text", + "content": "C. Fosco et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Brain-ID_ Learning Contrast-agnostic Anatomical Representations for Brain Imaging/b53a82d6-784c-4409-ad49-9bb017717031_content_list.json b/2024/Brain-ID_ Learning Contrast-agnostic Anatomical Representations for Brain Imaging/b53a82d6-784c-4409-ad49-9bb017717031_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..387f05800068ae356d59ff03f26e3c87ef2bbfb1 --- /dev/null +++ b/2024/Brain-ID_ Learning Contrast-agnostic Anatomical Representations for Brain Imaging/b53a82d6-784c-4409-ad49-9bb017717031_content_list.json @@ -0,0 +1,1675 @@ +[ + { + "type": "text", + "text": "Brain-ID: Learning Contrast-agnostic Anatomical Representations for Brain Imaging", + "text_level": 1, + "bbox": [ + 220, + 141, + 782, + 188 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Peirong Liu1 Oula Puonti1 Xiaoling Hu1", + "bbox": [ + 320, + 210, + 684, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Daniel C. Alexander $^{2}$ \nJuan E. Iglesias $^{1,2,3}$", + "bbox": [ + 330, + 227, + 671, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Harvard Medical School and Massachusetts General Hospital $^{2}$ University College London $^{3}$ Massachusetts Institute of Technology", + "bbox": [ + 267, + 253, + 733, + 284 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/24561250013a44bfa91f12541944ec764831343688bad5919067e368166efbd7.jpg", + "image_caption": [ + "Fig. 1: Brain-ID features serve as the distinctive identity of each subject (ID) regardless of appearance (contrast, deformation, corruptions, etc) and quickly adapt to downstream tasks – either contrast-independent (anatomy Reconstruction, Segmentation), or contrast-dependent (Super-Resolution, Bias Field estimation), all through one layer." + ], + "image_footnote": [], + "bbox": [ + 220, + 308, + 784, + 478 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recent learning-based approaches have made astonishing advances in calibrated medical imaging like computerized tomography (CT). Yet, they struggle to generalize in uncalibrated modalities – notably magnetic resonance (MR) imaging, where performance is highly sensitive to the differences in MR contrast, resolution, and orientation. This prevents broad applicability to diverse real-world clinical protocols. We introduce Brain-ID, an anatomical representation learning model for brain imaging. With the proposed \"mild-to-severe\" intra-subject generation, Brain-ID is robust to the subject-specific brain anatomy regardless of the appearance of acquired images. Trained entirely on synthetic inputs, Brain-ID readily adapts to various downstream tasks through one layer. We present new metrics to validate the intra/inter-subject robustness of Brain-ID features, and evaluate their performance on four downstream applications, covering contrast-independent (anatomy reconstruction, brain segmentation), and contrast-dependent (super-resolution, bias field estimation) tasks (Fig. 1). Extensive experiments on six public datasets demonstrate that Brain-ID achieves state-of-the-art performance in all tasks on different MR contrasts and CT, and more importantly, preserves its performance on low-resolution and small datasets. Code is available at https://github.com/peirong26/Brain-ID.", + "bbox": [ + 259, + 561, + 743, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 143, + 374, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Magnetic resonance imaging (MRI) enables in vivo noninvasive imaging of the human brain with exquisite and tunable soft-tissue contrast [8]. Recent machine learning methods have achieved great improvements in faster and more accurate MRI analysis [27], such as segmentation [18,32,42,47], registration [3,49,55,64], super-resolution [52, 53], and connectivity studies [45]. However, most existing MRI analysis methods are specific to certain MR contrast(s) and require near-isotropic acquisitions. Therefore, models face sharp performance drops when voxel size and anisotropy increase, or are being used for a different contrast than training [57]. This reduces model generalizability and results in duplicate training efforts. Resorting to synthetic data, recent contrast-agnostic models [4, 25, 27, 28, 36, 39, 40] achieve impressive results and largely extend model applicability. However, these models are only applicable to the tasks they were trained for.", + "bbox": [ + 212, + 172, + 787, + 352 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meanwhile, task-agnostic foundation models in computer vision and natural language processing have witnessed remarkable success, along with the fast developments of large-scale datasets [2,7,10,14,34]. However, due to different acquisition/processing protocols and privacy requirements across institutions, large-scale medical imaging datasets require significantly more effort than natural imaging/language. Thus, medical foundation models are not as well developed. MONAI [11] includes a pre-trained model zoo, yet all are highly task-oriented and sensitive to contrasts. Zhou et al. [66] constructed a medical foundation model, designed for eye and systemic health condition detections from retinal scans, and only works on the modalities of color fundus photography and optical coherence tomography. Lately, generalist biomedical AI systems [44, 50, 54] have shown great potential in biomedical tasks under vision-language context, e.g., (visual) question answering, image classification, radiology report generation and summarization. However, they have not explored the challenging vision tasks such as reconstruction, segmentation, super-resolution, and registration.", + "bbox": [ + 212, + 354, + 787, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce Brain-ID, a contrast-agnostic feature representation approach trained on synthetic inputs, that can adapt to various tasks through one layer.", + "bbox": [ + 215, + 580, + 785, + 611 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) We introduce an on-the-fly, intra-subject data generator capable of synthesizing any contrast, with a mild-to-severe corruption strategy (Fig. 2, Sec. 3.2). Unlike real-world datasets constrained by the images acquired per subject, Brain-ID learns in a much more expansive and diverse space.", + "2) We design a feature representation learning framework guided by the unique anatomy of each subject. We validate that Brain-ID's high-resolution features are consistently robust to the superficial perturbations of an image's appearance (Sec. 4.2). Straightforward one-layer adaptions are further presented to adapt Brain-ID features to downstream tasks (Sec. 3.3).", + "3) We evaluate Brain-ID on both contrast-independent (anatomy reconstruction, brain segmentation) and contrast-dependent (super-resolution, bias field estimation) tasks, across six public datasets ( $\\approx$ 8,000 images) including MR (T1w, T2w, FLAIR) and CT. Brain-ID achieves state-of-the-art performance on all tasks (Tab. 2), and further maintains its high performance on low-resolution images (Fig. 6) and limited-size datasets (Fig. 7)." + ], + "bbox": [ + 220, + 614, + 784, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "P. Liu et al.", + "bbox": [ + 271, + 114, + 352, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 143, + 382, + 160 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature Representation in Medical Imaging As mentioned in Sec. 1, general feature representation learning in medical imaging can be more challenging than in the natural image domain, due to limited data availability. Xu et al. [62] introduced a multiple-instance learning model for feature representation, but it only applies to histopathology image classification. You et al. [65] presented CVRL, a semi-supervised representations approach designed for image segmentation, but it requires CT to extract anatomical information. SAM [63] encodes anatomical information in feature embeddings, which has shown to be effective for registration [38]. However, same as CVRL, SAM only works on CT. BrainPrint [56] is a compact and discriminative representation of brain morphology, which is specifically designed for cortical surface analyses. To our best knowledge, CIFL [15] is the only existing work on learning contrast-agnostic and task-independent brain feature representations. CIFL relies on contrastive learning alone, and is insufficient to outperform task/contrast-specific supervised models in downstream applications as shown by our experiments (Tab. 2).", + "bbox": [ + 212, + 172, + 787, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrast-invariant Learning for MR Images MRI scans acquired across sites vary substantially in contrasts, resolutions, orientations, etc. When given a new dataset, heterogeneity leads to duplicate training efforts for approaches that are sensitive to specific combinations of MR contrast, resolution, or orientation. Classical brain segmentation models used Bayesian inference for contrast-robustness [20,37], which requires a long processing time and struggles with low or anisotropic resolutions [27, 46]. Recently, there have been works using synthetic data to achieve contrast-invariance in tasks like image segmentation [4,36], registration [25], super-resolution [28], and skull-stripping [26]. However, all the above-mentioned methods are trained in a task-specific manner, their features therefore cannot be readily applied to other domains.", + "bbox": [ + 212, + 416, + 787, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Brain-ID: Learning Anatomy-specific Brain Features", + "text_level": 1, + "bbox": [ + 214, + 601, + 754, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As discussed in Sec. 1, the main challenges to obtain a general and robust feature representation for MR images lie in $(i)$ the practical restrictions of building large-scale datasets with diverse contrasts; and $(ii)$ the nature of most medical imaging models that are task-oriented and specific to data type (contrast, resolution, orientation, etc). We aim to learn a brain feature representation that is:", + "bbox": [ + 212, + 628, + 785, + 702 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(i) Robust: features should be robust to each subject's distinct anatomy, unaffected by variations in poses/deformations, contrasts, resolutions, or artifacts.", + "(ii) Expressive: features should also exhibit high expressiveness, containing rich information that facilitates easy and effective adaptation to diverse downstream tasks, eliminating the necessity for extensive training data." + ], + "bbox": [ + 214, + 704, + 785, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first introduce Brain-ID's data generator (Sec. 3.1) and training framework (Sec. 3.2) to achieve the above two aims. Then, we present our one-layer solutions for adapting Brain-ID features to downstream tasks that could be either dependent or independent of the contrast of input images (Sec. 3.3).", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Brain-ID", + "bbox": [ + 665, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/68490469bf1f48f04ec51c79a8e01d79643addf15b0050a8c34fe4585f03a669.jpg", + "image_caption": [ + "Fig.2: Brain-ID's data generator on the fly. Given the brain segmentation labels of a subject, we randomly generate a deformation field, and synthesize intra-subject samples featuring various contrast intensities and corruption levels (Sec. 3.1)." + ], + "image_footnote": [], + "bbox": [ + 228, + 143, + 781, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Enrich the Intra-subject Learning Space", + "text_level": 1, + "bbox": [ + 214, + 330, + 593, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A good representation relies on large-scale data, however, different acquisition protocols, processing pipelines, and privacy requirements across institutions make large-scale data less accessible and require significantly more effort [1, 17, 48]. Moreover, for each subject, there are usually only limited acquired scans available. This lack of data consistency is a significant barrier to obtaining a robust, contrast-invariant feature representation. Brain-ID avoids these barriers through the use of synthetic data.", + "bbox": [ + 212, + 352, + 787, + 457 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To generate images with complex brain structures, we start with brain segmentation labels of anatomical structures ( $L$ in Fig. 2). The generation consists of three steps, $(i)$ deformation generation, $(ii)$ contrast synthesis, and $(iii)$ data corruption (including lower resolution resampling). For simplicity, $\\Theta$ denotes the parameter group of the generation process described below.", + "bbox": [ + 212, + 458, + 787, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Deformation Generation We first generate a random deformation field $(\\phi|_{\\theta_{\\phi}})$ consisting of an affine transformation and a non-linear displacement field:", + "bbox": [ + 214, + 537, + 785, + 568 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi | _ {\\theta_ {\\phi}} = \\mathcal {T} | _ {\\theta_ {\\phi}} \\circ \\mathcal {A} | _ {\\theta_ {\\phi}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 428, + 577, + 785, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{A}|_{\\theta_{\\phi}}$ denotes an affine transformation matrix which includes linear (rotation, scaling, shearing) transformation and translation, $\\mathcal{T}|_{\\theta_{\\phi}}$ refers to a nonlinear displacement field computed as the integration of a stationary velocity field (SVF) that is smooth and invertible everywhere and thus preserves the topology of the brain anatomy [28]. $\\theta_{\\phi} \\in \\Theta$ controls the transformation ranges.", + "bbox": [ + 212, + 599, + 787, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MR Contrast Synthesis Then, we synthesize images $(I(x), x \\in \\Omega)$ by randomly \"painting\" intensities on the segmentation maps according to their brain structure labels $(l \\in L)$ . Specifically, the regional intensities are generated by separately sampling a Gaussian distribution on each labeled region:", + "bbox": [ + 212, + 679, + 787, + 741 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} I (x) \\sim \\mathcal {N} \\left(\\mu_ {l}, \\sigma_ {l}\\right), \\quad l \\in L, \\\\ \\mu_ {l} \\sim \\mathcal {N} \\left(0, 1 \\mid \\theta_ {\\mu}, \\theta_ {l}\\right), \\sigma_ {l} \\sim \\mathcal {N} \\left(0, 1 \\mid \\theta_ {\\sigma}, \\theta_ {l}\\right), \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 747, + 785, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu_{l}$ and $\\sigma_{l}$ refer to the mean and standard deviation of each segmentation label $l$ , and are independently sampled from Gaussian distributions at each voxel, with $\\theta_{l}, \\theta_{\\mu}, \\theta_{\\sigma} \\in \\Theta$ controlling the shifts and scales of their values.", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "P. Liu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2b3785bbaec27ab54c9bbb8b9bf80ed99204e9da07ebd41278ea716e6d1c6fe1.jpg", + "image_caption": [ + "Fig. 3: Brain-ID's contrast-agnostic anatomical representation learning framework." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 782, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Resolution Simulation and Data Corruption Given a deformed, contrast-synthesized image $(I)$ , we adopt the standard data corruption pipeline [27], which further augments images with different levels of resolution and scanning artifacts that are commonly found in real-world clinical protocols.", + "bbox": [ + 212, + 335, + 782, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, we are able to generate an infinite number of variations from a single subject with its unique brain anatomy. Generating images with randomized contrast/resolution/orientation on the fly for each subject enormously enriches the learning space for a robust representation, and helps focus the learning on intrinsic subject-specific features rather than superficial aspects of images that depend on acquisition parameters and conditions.", + "bbox": [ + 212, + 396, + 784, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Extract Robust and Expressive Subject-specific Features", + "text_level": 1, + "bbox": [ + 214, + 508, + 728, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As mentioned above, we would like Brain-ID features to be robust to intra-subject variations and expressive to potential downstream tasks. In this section, we introduce Brain-ID's learning framework to achieve the two goals.", + "bbox": [ + 212, + 534, + 782, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intra-subject Data Generation In order to learn a feature representation that is distinctive to each subject and robust to varying MR contrasts, we observe that enriching intra-subject samples leads to better performance (Tab. 3). Specifically, instead of including multiple subjects for a mini-batch during each training iteration as in usual practice, Brain-ID focuses on maximizing the intra-subject variance to improve the subject-robustness of resulting features. As described in Alg. 1 (line 4-10), for each training iteration, after randomly selecting a subject and generating its deformation, Brain-ID generates a mini-batch of intra-subject samples $\\{(S_1,\\ldots ,S_N)\\}$ with randomly synthesized contrasts, resolutions and corruptions (Sec. 3.1). As will be introduced below, Brain-ID collects losses from all intra-subject samples and conducts back-propagation at once, to encourage the subject-specific robustness of its learned features.", + "bbox": [ + 212, + 583, + 784, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We set the intra-subject samples within a mini-batch to have random contrasts and \"mild-to-severe\", increasing level of corruptions (Fig. 3 (left)), to maximize the intra-subject variance while ensuring the stability of the training process against extreme corruption levels. (In Sec. 4.5, we compare various data generation designs, and provide insights on preventing unstable training.)", + "bbox": [ + 212, + 765, + 784, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Brain-ID", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Anatomy-guided Feature Representation A richer learning space helps improve the representation robustness against the variance of sample appearances, but not enough for extracting expressive features. For example, a mapping that projects all inputs to zero is perfectly robust, but does not provide useful information for potential downstream tasks. Therefore, proper guidance is crucial for feature representation as well. In CIFL [15], the authors use a contrastive loss on feature channels to encourage a more discriminative feature representation. However, this method is insufficient, given the extremely complex nature of the human brain anatomy (See comparisons in Sec. 4). Instead, we propose to use the standard T1w high-resolution structural MR contrast for brain morphometry, i.e., MP-RAGE (magnetization-prepared rapid gradient-echo), as the unique anatomy target to guide Brain-ID's feature representation learning.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 3, a feature extraction backbone $(\\mathcal{F})$ first maps the input mini-batch of intra-subject generated samples, $\\{S_1,\\dots,S_N\\}$ , to their corresponding feature space, $\\{\\mathbf{F}_1,\\ldots ,\\mathbf{F}_N\\}$ . A linear activation layer $(\\mathcal{L})$ then projects the features to the current subject's standard contrast (MP-RAGE) space, $T$ . The training loss is obtained by summing over the anatomy reconstruction loss of all intra-subject samples in the current mini-batch:", + "bbox": [ + 212, + 327, + 787, + 417 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL = \\sum_ {i} ^ {N} | \\mathcal {L} (\\mathbf {F} _ {i}) - T | + \\lambda | \\nabla \\mathcal {L} (\\mathbf {F} _ {i}) - \\nabla T |, \\lambda \\in R ^ {+}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 426, + 784, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $T$ , as the high-resolution, unique anatomy target for all intra-subject samples, encourages the similarities of the features extracted from the same subject via reconstruction ( $1^{\\mathrm{st}}$ term) and gradient difference ( $2^{\\mathrm{nd}}$ term) L1 loss. Brain-ID uses MP-RAGE as the learning target, which formulates both super-resolution and contrast-synthesis problems, and encourages a richer feature representation. (Sec. 4.5 provides further insights on how different choices of anatomy guidance would affect the resulting features' properties.)", + "bbox": [ + 214, + 458, + 787, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Adapting Brain-ID to Downstream Tasks by One Layer", + "text_level": 1, + "bbox": [ + 214, + 584, + 714, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With the intra-subject data generation and anatomy-guided feature learning design introduced in Sec. 3.2, a well-trained Brain-ID model is able to extract robust, high-resolution anatomical features from images with varying deformations, contrasts, resolutions, and artifacts. To minimize the modifications, we propose straightforward one-layer solutions adapting Brain-ID features to various brain imaging applications. We later demonstrate that the simple adaptions are effective enough for Brain-ID to achieve state-of-the-art performance across all downstream tasks (Sec. 4.3), even for small datasets (Sec. 4.4).", + "bbox": [ + 212, + 606, + 787, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Contrast-independent Tasks For tasks where the output should be independent of the input MR contrast, e.g., brain segmentation, we simply add an additional layer following Brain-ID features $(\\mathbf{F})$ , and fine-tune the model:", + "bbox": [ + 214, + 729, + 787, + 777 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL = \\operatorname {t a s k} _ {-} \\operatorname {l o s s} _ {-} \\operatorname {f u n c} \\left(\\mathcal {L} (\\mathbf {F}), T\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 785, + 784, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}$ and $T$ refer to the task-specific activation layer and the contrast-independent ground truth, respectively.", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "P. Liu et al.", + "bbox": [ + 271, + 114, + 352, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Contrast-dependent Tasks Since Brain-ID features are contrast-agnostic and robust to artifacts, for tasks relevant to the input's contrast/quality, e.g., super-resolution, we concatenate the input image $(I)$ that contains the original contrast information with its high-resolution Brain-ID features along the channel dimension, before forwarding into the task-specific activation layer:", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL = \\operatorname {t a s k} _ {-} \\operatorname {l o s s} _ {-} \\operatorname {f u n c} \\left(\\mathcal {L} (\\mathbf {F} \\oplus I), T\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 234, + 782, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{L}$ and $T$ refer to the task-specific activation layer and the ground truth.", + "bbox": [ + 214, + 260, + 779, + 275 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 297, + 374, + 314 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We demonstrate the two properties of Brain-ID as claimed in Sec. 3. (i) Robustness (Sec. 4.2): we propose \"canonical\" and \"atlas-registered\" features to assess the intra/inter-subject feature robustness; (ii) Expressiveness (Sec. 4.3): we evaluate the performance of adapting Brain-ID features to various brain imaging applications. We further challenge Brain-ID with reduced-size datasets to explore its ability when limited real data is available for training (Sec. 4.4).", + "bbox": [ + 212, + 328, + 782, + 419 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Datasets, Metrics and Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 441, + 643, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets For Brain-ID features pre-training, we use 2045 3D segmentations (\"anatomies\", or IDs) from the public ADNI dataset [30]. For downstream evaluation, we use six public datasets covering T1w and T2w MRI, FLAIR MRI, and CT: ADNI [30], ADNI3 [60], HCP [19], ADHD200 [9], AIBL [23], OASIS3 [35].", + "bbox": [ + 212, + 469, + 782, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics We evaluate individual tasks from different aspects. For feature similarity measurements, we use L1 distance, and (MS-)SSIM (multi-scale structural similarity) [41,58,59]. For reconstruction and super-resolution, we use L1, PSNR (peak signal-to-noise ratio) and (MS-)SSIM. For segmentation, we use Dice scores. For bias field estimation, we use the normalized L2 distance (norm-L2) to avoid possible arbitrary scalings from nonuniformity correction [16].", + "bbox": [ + 212, + 532, + 782, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details As a general feature representation model, Brain-ID can use any backbone to extract brain features. For fairer comparison, we adopt the same five-level 3D UNet [47] (with 64 feature channels in the last layer) as utilized in state-of-the-art models we compare with in Sec. 4.3. During feature pre-training (Sec. 3.2), a linear regression layer is added for anatomy supervision. For downstream adaptions, A task-specific adaption layer is added (Sec. 3.3).", + "bbox": [ + 212, + 628, + 782, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Intra/Inter-subject Feature Robustness", + "text_level": 1, + "bbox": [ + 214, + 739, + 586, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we examine the robustness of Brain-ID features, i.e., the 64 features from the last layer. For comparable and reproducible results, we use our data generator (Sec. 3.1) to prepare (deform and corrupt) 1000 testing samples from 100 randomly selected subjects in ADNI [30] (T1w) testing set, with 10 intra-subject samples for each subject.", + "bbox": [ + 212, + 763, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Brain-ID", + "bbox": [ + 665, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/437282f39cfc164e1f6371617cc3902ae0ce5b40c0baadf767663eab7eabf17b.jpg", + "image_caption": [ + "Fig. 4: (a) Intra-subject and (b) inter-subject robustness of Brain-ID features." + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 781, + 531 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare our features with CIFL [15] which is, to our best knowledge, the only similar work to Brain-ID, that also aims to learn contrast-agnostic brain features. Note the original CIFL method has only experimented on 2D images, and does not have our intra-subject data generation design. For a fairer comparison, we trained CIFL with its contrastive-learning design, while using the same data generator and network architecture as Brain-ID.", + "bbox": [ + 212, + 595, + 787, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Intra-subject Feature Robustness Ideally, Brain-ID features computed from the same subject should be structurally identical, regardless of their appearances (contrast, resolution, noises, etc). To assess the intra-subject feature robustness, we first define their corresponding \"canonical\" features.", + "bbox": [ + 212, + 694, + 787, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "\"Canonical\" Features. We compute $\\phi^{-1}$ , the inverse of $S$ 's generated deformation as in Eq. (1), and map back each sample's $(S)$ Brain-ID features $(\\mathbf{F} = \\mathbf{F}(S|\\phi))$ to their original domain without any deformation:", + "bbox": [ + 212, + 758, + 787, + 805 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} ^ {- 1} = \\phi^ {- 1} \\circ \\mathbf {F} (S \\mid \\phi). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 821, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "P. Liu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9f06611b63ae01e3351e07b97af6e05ca298bb8b01e247087c83acbae2d157ee.jpg", + "table_caption": [ + "Table 1: Evaluations on intra/inter-subject feature robustness." + ], + "table_footnote": [], + "table_body": "
ModeMethodL1 (↓)SSIM (↑)MS-SSIM (↑)
IntraCIFL [15]0.122 (±0.031)0.511 (±0.298)0.531 (±0.367)
Brain-ID0.011 (±0.001)0.858 (±0.041)0.921 (±0.008)
InterCIFL [15]0.230 (±0.039)0.552 (±0.240)0.524 (±0.291)
Brain-ID0.014 (±0.013)0.843 (±0.044)0.913 (±0.011)
", + "bbox": [ + 318, + 169, + 683, + 242 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As shown in Fig. 4-(a), for the same subject (ID-i), although the inputs (row) are processed by varying deformations and corruptions, their Brain-ID's \"canonical\" features are of similar looking along each feature channel (column). In Tab. 1, we quantify the feature similarity: For each testing subject, we first generate a sample without any deformation or corruption and use its Brain-ID feature as the gold-standard reference for all other \"canonical\" features computed from intra-subject samples that may have different levels of deformation and corruptions. The final results are obtained by scores averaged over all 64 feature channels of all the intra-subject samples.", + "bbox": [ + 212, + 266, + 787, + 402 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Inter-subject Feature Robustness We further assess the feature robustness among anatomies, based on the assumption that features from different subjects should ideally be structurally similar, once being registered to a common domain. \"Atlas-registered\" Features. We register [43] features $(\\mathbf{F} = \\mathbf{F}(S|\\phi))$ from all subjects to a common brain atlas [21, 22]:", + "bbox": [ + 212, + 406, + 787, + 482 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathbf {F}}} = \\psi_ {S \\rightarrow A} \\circ \\mathbf {F} (S \\mid \\phi), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 491, + 785, + 508 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\psi_{S\\to A}$ denotes the mapping from $S$ to the atlas, $A$", + "bbox": [ + 212, + 518, + 632, + 532 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As shown in Fig. 4-(b), even though the input samples (row) from different subjects have varying anatomical structures, after being registered to a common domain, their \"atlas-registered\" Brain-ID features still appear consistent (column). In Tab. 1, we further provide quantitative results of the structural similarity measured between features from different subjects. Similar to Sec. 4.2, we randomly select a testing subject and generate a sample without any deformation or corruption. We use its \"atlas-registered\" Brain-ID feature as the gold standard for all \"atlas-registered\" features computed from other samples that may have different levels of deformation and corruption. The final results are obtained by scores averaged over all samples of the 100 testing subjects.", + "bbox": [ + 212, + 534, + 787, + 685 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In both intra/inter-subject aspects, Brain-ID outperforms CIFL by a large margin, which validates its feature robustness against deformations, contrasts, resolutions, artifacts (intra-subject), and anatomies (inter-subject).", + "bbox": [ + 212, + 685, + 787, + 731 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Downstream Evaluation", + "text_level": 1, + "bbox": [ + 214, + 751, + 459, + 766 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Following the standard evaluation protocol for representation learning [12], we fine-tune Brain-ID features for downstream tasks through one-layer (Sec. 3.3). Models We compare the downstream performance of Brain-ID features with: $(i - ii)$ SAMSEG [13,46] and FastSurfer [24] (only works on T1w), state-of-the-art", + "bbox": [ + 212, + 775, + 787, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Brain-ID", + "bbox": [ + 665, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/7c5afb057c19828fd2a976c65d9e9ba4b880faf378ece7de0b0fdaa45c68eecf.jpg", + "image_caption": [ + "Fig. 5: Qualitative comparisons on downstream tasks of reconstruction (Recon), superresolution (SR), and segmentation (Seg), between Brain-ID, the baseline SCRATCH, and the state-of-the-art methods CIFL [15], SynthSR [27] (Recon and SR), SAMSEG [13] (Seg). The visualized testing examples are taken from: AIBL-FLAIR [23] for Recon, AIBLT1w [23] for SR, and OASIS-CT [35] for Seg. The mint circles highlight some details." + ], + "image_footnote": [], + "bbox": [ + 225, + 146, + 776, + 349 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "classical and machine-learning-based brain segmentation models, respectively; (iii) SynthSR [27], state-of-the-art, contrast-agnostic T1w synthesis model. We also provide fine-tuned SynthSR (SynthSR-FT) results for further comparison; (iv) SCRAW: a baseline with the same data generation/architecture as Brain-ID, yet trained from scratch. SCRAW is used to validate Brain-ID's effectiveness. (v) CIFL [15]: a baseline with the same data generation/architecture as Brain-ID and SCRAW, yet initialized with CIFL's pre-trained features. We use CIFL to demonstrate the superiority of Brain-ID's anatomy-guided representation.", + "bbox": [ + 212, + 467, + 787, + 588 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Contrast-independent Applications", + "text_level": 1, + "bbox": [ + 214, + 593, + 508, + 608 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "- Anatomy Reconstruction/Contrast Synthesis As one of the most common types of MRI scans, T1w MR images highlight the differences between gray and white matter, and are mostly often utilized to image the anatomy of the brain, spinal cord, bones and joints [51]. For each dataset, we train models to reconstruct their T1w MRI counterparts with the L1 loss.", + "bbox": [ + 212, + 609, + 787, + 685 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "- Brain Segmentation For each dataset, we use SynthSeg [4] to obtain the segmentation labels with 30 brain anatomical regions [20], as the gold standard segmentation target. We train our compared models to predict the brain segmentation labels, with the soft dice loss and cross-entropy loss [4].", + "bbox": [ + 212, + 686, + 787, + 748 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As shown in Fig. 5, Brain-ID clearly outperforms CIFL and SCRATCH in the reconstruction task. As highlighted by the mint circles, it reveals the anatomy details and produces more fine-grained results than the strong state-of-the-art method SynthSR [27]. Brain-ID also achieves better segmentation results, especially in the smaller and more challenging regions. Quantitative comparisons are in Tab. 2, where Brain-ID obtains the best scores across most metrics. No-", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "P. Liu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/3be05f652aedbb23a49b7a62b3f73f8b9196c33df03b14fbe221da40b3b28502.jpg", + "table_caption": [ + "Table 2: Comparisons of Brain-ID with state-of-the-art approaches on downstream applications. (\"train/test\" refers to the number of subjects in the training/testing set.)" + ], + "table_footnote": [ + "(1) \"means not applicable: FastSurfer [24] and SAMSEG [13] are for segmentation-only; SynthSR [27] only synthesize T1w MRI; CT does not have bias fields. (2) SynthSR ET; SynthSR fine tuned on each task/dataset, for a fairer comparison." + ], + "table_body": "
Modality (Contrast)Dataset (Train/Test)MethodReconstructionSegmentationSuper-resolution (from 3 / 5 / 7 mm)Bias Field norm-12 (↓)
L1 (↓)PSNR (↑)SSIM (↑)MS-SSIM (↑)Dice (↑)PSNR (↑)SSIM (↑)MS-SSIM (↑)MS-SSIM (↑)
MR (T1w)ADNI [30] (1841/204)SAMSEG [13]---0.795------
FastSurfer [24]---0.803------
SynthSR [27]0.01426.780.9800.980-27.01 / 25.81 / 21.790.977 / 0.943 / 0.8570.981 / 0.968 / 0.869--
SynthSR-FT0.01029.540.9840.9200.82530.46 / 28.03 / 23.910.979 / 0.960 / 0.8980.985 / 0.973 / 0.900-0.051
CIFL [15]0.01326.970.9730.9610.82030.12 / 27.21 / 23.800.978 / 0.912 / 0.8870.979 / 0.950 / 0.842-0.053
SCRATCH0.01127.540.9790.9570.81630.29 / 27.94 / 24.570.975 / 0.959 / 0.9200.982 / 0.964 / 0.866-0.056
Brain-ID0.01233.820.9930.9890.83731.30 / 28.62 / 24.680.983 / 0.961 / 0.9280.987 / 0.980 / 0.947-0.048
MR (T1w)HCP [19] (808/87)SAMSEG [13]---0.810------
FastSurfer [24]---0.819------
SynthSR [27]0.03322.130.8540.901-22.21 / 20.83 / 19.930.848 / 0.802 / 0.7470.899 / 0.864 / 0.789--
SynthSR-FT0.02527.110.9350.9170.81025.90 / 25.45 / 25.100.889 / 0.872 / 0.8310.919 / 0.911 / 0.889-0.052
CIFL [15]0.02926.420.9320.9130.80425.98 / 25.11 / 23.990.930 / 0.883 / 0.8710.936 / 0.927 / 0.880-0.059
SCRATCH0.02327.320.9230.9090.79226.69 / 26.41 / 25.050.932 / 0.896 / 0.8480.942 / 0.930 / 0.892-0.069
Brain-ID0.02027.470.9570.9290.84329.70 / 27.90 / 26.120.957 / 0.901 / 0.8830.973 / 0.938 / 0.890-0.047
SAMSEG [13]---0.769------
FastSurfer [24]---0.796------
SynthSR [27]0.02323.600.9280.909-23.13 / 22.40 / 22.270.921 / 0.907 / 0.8760.903 / 0.892 / 0.871--
SynthSR-FT0.02127.110.9500.9690.82027.95 / 27.34 / 26.180.920 / 0.893 / 0.9220.890 / 0.890 / 0.879-0.134
CIFL [15]0.02828.520.9610.9700.81927.81 / 27.10 / 26.320.900 / 0.877 / 0.8640.812 / 0.793 / 0.769-0.102
SCRATCH0.03327.280.9570.9630.81627.99 / 27.37 / 26.630.893 / 0.889 / 0.8760.794 / 0.785 / 0.753-0.128
Brain-ID0.02129.890.9660.9760.84330.17 / 28.23 / 26.890.962 / 0.923 / 0.8920.973 / 0.921 / 0.883-0.093
SAMSEG [13]---0.784------
FastSurfer [24]---0.801------
SynthSR [27]0.03521.670.8820.853-21.42 / 21.13 / 21.410.876 / 0.846 / 0.8090.851 / 0.831 / 0.805--
SynthSR-FT0.01626.900.9170.9050.81228.55 / 27.43 / 25.940.880 / 0.859 / 0.8230.860 / 0.845 / 0.811-0.110
CIFL [15]0.01328.690.9320.9210.81028.11 / 26.12 / 25.110.865 / 0.848 / 0.8310.828 / 0.800 / 0.789-0.109
SCRATCH0.01131.550.9860.9850.79629.41 / 27.89 / 26.580.854 / 0.847 / 0.8140.853 / 0.845 / 0.827-0.113
Brain-ID0.01132.480.9960.9960.84729.50 / 28.56 / 26.870.898 / 0.862 / 0.8110.887 / 0.850 / 0.823-0.107
SAMSEG [13]---0.799------
FastSurfer [24]---0.802------
SynthSR [27]0.02622.950.9160.912-22.57 / 21.82 / 21.780.917 / 0.892 / 0.8560.906 / 0.893 / 0.869--
SynthSR-FT0.01429.890.9410.9220.81029.15 / 27.89 / 26.770.932 / 0.906 / 0.8780.927 / 0.900 / 0.871-0.118
CIFL [15]0.01130.120.9380.9250.81629.46 / 27.97 / 26.730.907 / 0.889 / 0.8700.891 / 0.872 / 0.857-0.094
SCRATCH0.01130.530.9320.9130.81429.43 / 28.10 / 26.820.900 / 0.886 / 0.8630.884 / 0.870 / 0.853-0.128
Brain-ID0.00933.730.9720.9630.85129.75 / 28.58 / 27.090.957 / 0.922 / 0.8900.975 / 0.956 / 0.937-0.088
SAMSEG [13]---0.782------
FastSurfer [24]---0.805------
SynthSR [27]0.02826.100.8940.8980.75528.43 / 25.36 / 22.100.942 / 0.851 / 0.8010.977 / 0.892 / 0.798-0.140
SynthSR-FT0.03625.120.8910.8790.78728.26 / 25.78 / 21.730.959 / 0.893 / 0.7970.982 / 0.906 / 0.742-0.138
SCRATCH0.03824.990.8730.8660.75628.12 / 24.52 / 21.580.945 / 0.863 / 0.7830.980 / 0.893 / 0.738-0.136
Brain-ID0.01628.100.9340.9350.80530.26 / 26.11 / 24.100.959 / 0.902 / 0.8320.987 / 0.953 / 0.912-0.136
SAMSEG [13]---0.763------
FastSurfer [24]---0.801------
SynthSR [27]0.03320.080.8050.831------
SynthSR-FT0.02422.930.8150.8400.71930.15 / 28.09 / 26.920.940 / 0.912 / 0.8930.966 / 0.948 / 0.922-0.200
CIFL [15]0.02223.710.8200.8390.72129.98 / 27.01 / 25.660.931 / 0.899 / 0.8610.967 / 0.941 / 0.905-0.193
SCRATCH0.02022.270.8490.8510.71431.91 / 29.20 / 27.350.954 / 0.934 / 0.8960.982 / 0.969 / 0.949 / 0.936-0.197
Brain-ID0.02223.990.8610.8500.78232.26 / 29.90 / 27.090.937 / 0.930 / 0.9080.988 / 0.971 / 0.948-0.148
SAMSEG [13]---0.718------
SynthSR [27]0.02622.770.9190.895-30.11 / 28.70 / 25.320.930 / 0.899 / 0.8640.948 / 0.910 / 0.878-0.251
SynthSR-FT0.02123.340.9210.9000.75332.72 / 29.00 / 26.990.949 / 0.906 / 0.8730.953 / 0.919 / 0.889-0.237
CIFL [15]0.02021.290.9110.8970.76132.72 / 28.71 / 25.860.945 / 0.915 / 0.8770.941 / 0.936 / 0.917-0.136
SCRATCH0.02520.800.9010.8620.75932.36 / 28.71 / 25.840.945 / 0.915 / 0.8770.981 / 0.953 / 0.917-0.236
Brain-ID0.01726.440.9270.8920.78632.18 / 30.00 / 28.280.959 / 0.921 / 0.8830.982 / 0.965 / 0.921-0.227
SAMSEG [13]---0.710------
FastSurfer [24]---0.795------
SynthSR [27]0.02921.770.9020.892------
SynthSR-FT0.02425.800.9140.9000.73526.20 / 24.91 / 23.170.860 / 0.832 / 0.7950.933 / 0.910 / 0.894-0.126
CIFL [15]0.02627.110.8910.8700.72127.72 / 26.85 / 25.950.821 / 0.858 / 0.8140.969 / 0.939 / 0.905 / 0.901-0.109
SCRATCH0.02326.840.8980.8670.72028.02 / 26.52 / 25.430.878 / 0.838 / 0.7990.943 / 0.901 / 0.891-0.115
Brain-ID0.01927.250.9360.9120.76726.69 / 24.01 / 23.69 / 23.09 / 0.918 / 0.806 / 0.966 / 0.871 / 0.925 / 0.916---
", + "bbox": [ + 238, + 183, + 763, + 618 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "tably, Brain-ID achieves greater gains over baseline models, SCRATCH and CIFL, particularly on smaller datasets (e.g., ADNI3-T1/FLAIR, AIBL-FLAIR, where Brain-ID's robust and rich features quickly adapt to specific tasks/datasets.", + "bbox": [ + 215, + 671, + 782, + 717 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Contrast-dependent Applications We also evaluate Brain-ID on two brain imaging tasks that are dependent on the input modality/quality.", + "bbox": [ + 215, + 727, + 782, + 758 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "- Image Super-resolution We use the standard 1mm isotropic images from all datasets as the ground truth high-resolution target, and follow the strategy in SynthSR [27] where the input samples are randomly resampled and corrupted during training. For inference, we downsample the original images into 3mm, 5mm, and 7mm isotropic images as inputs, and evaluate the output quality.", + "bbox": [ + 215, + 763, + 782, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Brain-ID", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/fbd526bb28e75ae9e993bbda0e2a7d4230d8705df90091974255f28ff632932f.jpg", + "image_caption": [ + "Fig. 6: SCRATCH, which is well trained on HF T1w scans, produces highly descriptive features for HF T1w images ( $1^{\\mathrm{st}}$ row), but does not preserve the same high quality useful for downstream tasks when handling LF ( $2^{\\mathrm{nd}}$ row) or other contrasts ( $3^{\\mathrm{rd}}$ row)." + ], + "image_footnote": [], + "bbox": [ + 238, + 145, + 769, + 361 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "- Bias Field Estimation The bias field is a smooth, low-frequency multiplicative signal that corrupts MRI images, which affects image analysis tasks such as segmentation or texture analysis [31]. Bias field estimation is often needed as a pre-processing step to correct corrupted MRI images [16]. We apply randomly simulated bias fields to the input samples (Sec. 3.1), and train all models with the L2 loss. During inference, we pre-generate and apply the bias fields on the testing data for reproducibility, and evaluate the bias field estimation performance.", + "bbox": [ + 212, + 443, + 787, + 549 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "With a simple one-layer adaption (Sec. 3.3), Brain-ID's contrast-agnostic anatomical features achieve state-of-the-art performance on contrast-dependent tasks (Tab. 2, Fig. 5). As shown in Fig. 6, Brain-ID obtains contrast/resolution-robustness that cannot be achieved by models trained from real images (due to the limited variability in their appearance), regardless of the backbone choices. Such robust features greatly improve super-resolution, resulting in higher (MS-)SSIM scores. Brain-ID's gains are less obvious for bias field estimation, probably because the bias field is approximately independent of brain anatomy.", + "bbox": [ + 212, + 549, + 787, + 671 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Practical Value and Broader Impact", + "text_level": 1, + "bbox": [ + 214, + 691, + 560, + 707 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Low-resolution data Trained on synthetic images, Brain-ID features are not only robust to various contrasts/modalities, but also extremely robust to low-resolution data (Fig. 6) and provide huge potential in clinical MRI (including big retrospective data [5]) and portable low-field MRI [29].", + "bbox": [ + 212, + 715, + 785, + 775 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Small-size Datasets To investigate the effectiveness of Brain-ID features for limited-size datasets, we assess its performance across all four downstream tasks on subsets of ADNI3 [60] training set (298 cases originally). As shown in Fig. 7, we reduce the training set size percentage from $100\\%$ to an extreme $10\\%$ . With", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "P. Liu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/86e67cf7953cd9c059f570d091c039fe9b914201ffba548255a97a1d818ec3ae.jpg", + "image_caption": [ + "(c) Super-resolution", + "(d) Bias Field Estimation", + "Fig. 7: Adapting Brain-ID to small datasets. The horizontal (vertical) axes indicate training epochs (evaluation scores of corresponding tasks). Results are obtained by evaluating models collected throughout the epochs, on ADNI3 [60] full testing set." + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 769, + 463 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "only $20\\%$ of the data (mint line), Brain-ID still achieves better performance compared to its full-sized baseline, “ $100\\%$ ” SCRATCH (yellow line), over all tasks. 30 training samples ( $10\\%$ , blue line) may be at the edge of obtaining acceptable results for Brain-ID, as the model becomes less stable and the performance drops occur more frequently – especially for brain segmentation (Fig. 7-(b)).", + "bbox": [ + 212, + 561, + 787, + 638 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Additional Experiments", + "text_level": 1, + "bbox": [ + 214, + 660, + 459, + 676 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Choice of Anatomy Guidance We explore other anatomical learning targets than MP-RAGE (Sec. 3.2): $(i)$ segmentation labels; $(ii)$ both segmentation labels and MP-RAGE. As shown in Tab. 3, adding segmentation guidance improves robustness, yet inhibits the feature expressiveness and affects the downstream performance; we observe features from segmentation guidance produce less high-frequency texture than Brain-ID, especially within the same-label regions.", + "bbox": [ + 212, + 685, + 787, + 776 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Data Generation Design As shown in Tab. 3, training with all mildly-corrupted samples results in reduced robustness and harms the downstream performance, yet using all extremely corrupted samples leads to unstable training and model collapse. Brain-ID uses samples of gradually increased corruptions (Fig. 3 (left)).", + "bbox": [ + 212, + 779, + 792, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Brain-ID", + "bbox": [ + 666, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/409e8ef00ccd4cc441473b773a196e4945d01ec1e8f15aefcfbd8d6320a64534.jpg", + "table_caption": [ + "Table 3: Comparison between Brain-ID and its variants." + ], + "table_footnote": [], + "table_body": "
Model Setup [Comparison Target]Feature Robustness (Intra)Downstream (Reconstruction)
SSIM (↑)MS-SSIM (↑)L1 (↓)PSNR (↑)SSIM (↑)
[Representation guidance: supervision target for feature learning]
Segmentation0.8910.9630.02928.130.958
Segmentation + MP-RAGE0.8630.9400.02329.050.964
MP-RAGE (Brain-ID)0.8580.9210.02129.890.966
[Data generation design: corruption levels in intra-subject mini-batches]
All Mild (\\( \\sigma_{noise} = 1, \\cdots \\))0.7920.8130.02828.300.960
All Medium (\\( \\sigma_{noise} = 5, \\cdots \\))0.8310.8990.02229.670.964
All Severe (\\( \\sigma_{noise} = 10, \\cdots \\))N/AN/AN/AN/AN/A
Mild to Severe (Brain-ID)0.8580.9210.02129.890.966
[Mini-batch size: number of intra-subject samples]
20.8260.8970.02528.830.959
30.8410.9020.02229.300.962
4 (Brain-ID)0.8580.9210.02129.890.966
50.8600.9290.02129.930.965
", + "bbox": [ + 289, + 169, + 712, + 330 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Batch Size Brain-ID computes training loss of all intra-subject samples in a mini-batch at once (Eq. (3)). We observe that larger batches help improve feature robustness (Tab. 3), yet do not further enhance the downstream performance.", + "bbox": [ + 212, + 357, + 787, + 402 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.6 Further Discussions", + "text_level": 1, + "bbox": [ + 215, + 428, + 426, + 441 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations Although Brain-ID features excel in multiple tasks, they may not be optimal for tasks not involving voxel-wise prediction, e.g., functional brain mapping. Further, as Brain-ID generates data exclusively from anatomies, it tends to struggle with images with extensive pathology. Our future research will focus on robust representations that are encoded with pathology information.", + "bbox": [ + 212, + 455, + 785, + 532 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Data Privacy and Ethics Since Brain-ID features can potentially be used to reconstruct a high-resolution MRI scan, publicly sharing such features should be regarded as sharing the original scans from a subject [33]. Therefore, one should follow the same anonymization protocols as when sharing MRI data for research [61] – not only the removal of identifiers, but also defacing [6] or skull-stripping [26] to prevent the 3D reconstruction of faces.", + "bbox": [ + 212, + 536, + 785, + 628 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 652, + 359, + 669 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduced Brain-ID, a contrast-agnostic feature representation learning approach for brain imaging, which is distinctive to brain anatomy and resilient to variations in image appearances. Brain-ID is trained on synthetic inputs, and quickly adapts to downstream tasks through one layer. We validated Brain-ID features' robustness, and effectiveness on four downstream applications (reconstruction, segmentation, super-resolution, bias field estimation), and six public datasets, covering MR and CT. Brain-ID achieves state-of-the-art performances across all tasks and modalities, it further preserves high performance on low-resolution and small-size datasets. We believe Brain-ID unlocks the great potential of robust foundation models, especially for non-calibrated modalities.", + "bbox": [ + 212, + 689, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "P. Liu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Primarily supported by NIH under grant 1RF1AG080371. Additional support from NIH 1RF1AG080371, 1R01EB031114, 1R01AG070988, 1RF1MH123195. OP is supported by Lundbeckfonden R360-2021-395. Wellcome Trust award 221915/Z/20/Z and the NIHR UCLH Biomedical Research Centre support DCA on this topic.", + "bbox": [ + 212, + 176, + 787, + 253 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 277, + 321, + 292 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Adnan, M., Kalra, S., Cresswell, J.C., Taylor, G.W., Tizhoosh, H.R.: Federated learning and differential privacy for medical image analysis. Scientific reports (2022)", + "2. Awais, M., Naseer, M., Khan, S.S., Anwer, R.M., Cholakkal, H., Shah, M., et al.: Foundational models defining a new era in vision: A survey and outlook. arXiv abs/2307.13721 (2023)", + "3. Balakrishnan, G., Zhao, A., Sabuncu, M.R., Guttag, J.V., Dalca, A.V.: VoxelMorph: A learning framework for deformable medical image registration. IEEE Transactions on Medical Imaging (2018)", + "4. Billot, B., Greve, D.N., Puonti, O., Thielscher, A., Leemput, K.V., Fischl, B., et al.: SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining. Medical Image Analysis (2021)", + "5. Billot, B., Magdamo, C., Cheng, Y., Arnold, S.E., Das, S., Iglesias, J.E.: Robust machine learning segmentation for large-scale analysis of heterogeneous clinical brain MRI datasets. Proceedings of the National Academy of Sciences (2023)", + "6. Bischoff-Grethe, A., Ozyurt, I.B., Busa, E., Quinn, B.T., Fennema-Notestine, C., Clark, C.P., et al.: A technique for the deidentification of structural brain MR images. Human Brain Mapping (2007)", + "7. Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., et al.: On the opportunities and risks of foundation models. arXiv abs/2108.07258 (2021)", + "8. Brant-Zawadzki, M.N., Gillan, G.D., Nitz, W.R.: MP RAGE: A three-dimensional, T1-weighted, gradient-echo sequence-initial experience in the brain. Radiology (1992)", + "9. Brown, M.R.G., Sidhu, G.P.S., Greiner, R., Asgarian, N., Bastani, M., Silverstone, P.H., et al.: ADHD-200 global competition: diagnosing ADHD using personal characteristic data can outperform resting state fMRI measurements. Frontiers in Systems Neuroscience (2012)", + "10. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., et al.: Language models are few-shot learners. In: NeurIPS (2020)", + "11. Cardoso, M.J., Li, W., Brown, R., Ma, N., Kerfoot, E., Wang, Y., et al.: MONAI: An open-source framework for deep learning in healthcare. arXiv abs/2211.02701 (2022)", + "12. Caron, M., Touvron, H., Misra, I., J'egou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021)", + "13. Cerri, S., Puonti, O., Meier, D.S., Wuerfel, J., Mühlau, M., Siebner, H.R., Leemput, K.V.: A contrast-adaptive method for simultaneous whole-brain and lesion segmentation in multiple sclerosis. NeuroImage (2020)" + ], + "bbox": [ + 218, + 310, + 784, + 838 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Brain-ID", + "bbox": [ + 665, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., et al.: PaLM: Scaling language modeling with pathways. JMLR (2022)", + "15. Chua, Y.Z.R., Dalca, A.V.: Contrast invariant feature representations for segmentation and registration of medical images. In: MIDL (2023)", + "16. Chua, Z.Y., Zheng, W., Chee, M.W.L., Zagorodnov, V.: Evaluation of performance metrics for bias field correction in MR brain images. Journal of Magnetic Resonance Imaging (2009)", + "17. Darzidehkalani, E., Ghasemi-Rad, M., van Ooijen, P.: Federated learning in medical imaging: Part I: Toward multicentral health care ecosystems. Journal of the American College of Radiology (2022)", + "18. Ding, Z., Han, X., Liu, P., Niethammer, M.: Local temperature scaling for probability calibration. In: ICCV (2021)", + "19. Essen, D.C.V., Ugurbil, K., Auerbach, E.J., Barch, D.M., Behrens, T.E.J., Bucholz, R.D., et al.: The human connectome project: A data acquisition perspective. NeuroImage (2012)", + "20. Fischl, B., Salat, D.H., Busa, E., Albert, M.S., Dieterich, M., Haselgrove, C., et al.: Whole brain segmentation automated labeling of neuroanatomical structures in the human brain. Neuron (2002)", + "21. Fonov, V.S., Evans, A.C., Botteron, K., Almli, C.R., McKinstry, R.C., Collins, D.L., et al.: Unbiased average age-appropriate atlases for pediatric studies. Neuroimage (2011)", + "22. Fonov, V.S., Evans, A.C., McKinstry, R.C., Almli, C.R., Collins, D.: Unbiased nonlinear average age-appropriate brain templates from birth to adulthood. NeuroImage (2009)", + "23. Fowler, C., Rainey-Smith, S.R., Bird, S., Bomke, J., Bourgeat, P., Brown, B.M., et al.: Fifteen years of the australian imaging, biomarkers and lifestyle (AIBL) study: Progress and observations from 2,359 older adults spanning the spectrum from cognitive normality to Alzheimer's disease. Journal of Alzheimer's disease reports (2021)", + "24. Henschel, L., Kugler, D., Reuter, M.: FastSurferVINN: Building resolution-independence into deep learning segmentation methods — A solution for highres brain MRI. NeuroImage (2022)", + "25. Hoffmann, M., Billot, B., Greve, D.N., Iglesias, J.E., Fischl, B., Dalca, A.V.: SynthMorph: Learning contrast-invariant registration without acquired images. IEEE Transactions on Medical Imaging (2020)", + "26. Hoopes, A., Mora, J.S., Dalca, A.V., Fischl, B., Hoffmann, M.: SynthStrip: skullstripping for any brain image. NeuroImage (2022)", + "27. Iglesias, J.E., Billot, B., Balbastre, Y., Magdamo, C.G., Arnold, S., Das, S., et al.: SynthSR: A public AI tool to turn heterogeneous clinical brain scans into high-resolution T1-weighted images for 3D morphometry. Science Advances (2023)", + "28. Iglesias, J.E., Billot, B., Balbastre, Y., Tabari, A., Conklin, J., Alexander, D.C., et al.: Joint super-resolution and synthesis of $1\\mathrm{mm}$ isotropic MP-RAGE volumes from clinical MRI exams with scans of different orientation, resolution and contrast. NeuroImage (2020)", + "29. Iglesias, J.E., Schleicher, R., Laguna, S., Billot, B., Schaefer, P., McKaig, B., et al.: Quantitative brain morphometry of portable low-field-strength MRI using superresolution machine learning. Radiology (2022)", + "30. Jack, C.R., Bernstein, M.A., Fox, N.C., Thompson, P.M., Alexander, G.E., Harvey, D.J., et al.: The Alzheimer's disease neuroimaging initiative (ADNI): MRI methods. Journal of Magnetic Resonance Imaging (2008)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "P. Liu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Juntu, J., Sijbers, J., Dyck, D.V., Gielen, J.: Bias field correction for MRI images. In: CORES (2005)", + "32. Kamnitsas, K., Ledig, C., Newcombe, V.F.J., Simpson, J.P., Kane, A.D., Menon, D.K., et al.: Efficient multi-scale 3D CNN with fully connected CRF for accurate brain lesion segmentation. Medical Image Analysis (2016)", + "33. Kapkiç, M., Sağiroğlu, Ş.: Privacy issues in magnetic resonance images. International Journal of Information Security Science (2023)", + "34. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., et al.: Segment anything. arXiv abs/2304.02643 (2023)", + "35. LaMontagne, P.J., Keefe, S.J., Lauren, W., Xiong, C., Grant, E.A., Moulder, K.L., et al.: OASIS-3: Longitudinal neuroimaging, clinical, and cognitive dataset for normal aging and Alzheimer's disease. Alzheimer's & Dementia (2018)", + "36. Laso, P., Cerri, S., Sorby-Adams, A., Guo, J., Mateen, F., Goebl, P., et al.: Quantifying white matter hyperintensity and brain volumes in heterogeneous clinical and low-field portable MRI. arXiv abs/2312.05119 (2023)", + "37. Leemput, K.V., Maes, F., Vandermeulen, D., Suetens, P.: A unifying framework for partial volume segmentation of brain MR images. IEEE Transactions on Medical Imaging (2003)", + "38. Li, Z., Tian, L., Mok, T.C.W., Bai, X., Wang, P., Ge, J., et al.: SAMConvex: Fast discrete optimization for CT registration using self-supervised anatomical embedding and correlation pyramid. In: MICCAI (2023)", + "39. Liu, P., Lee, Y., Aylward, S., Niethammer, M.: Deep decomposition for stochastic normal-abnormal transport. In: CVPR (2022)", + "40. Liu, P., Tian, L., Zhang, Y., Aylward, S., Lee, Y., Niethammer, M.: Discovering hidden physics behind transport dynamics. In: CVPR (2021)", + "41. Liu, P., Wang, R., Cao, X., Zhou, Y., Shah, A., Lim, S.N.: Differential motion evolution for fine-grained motion deformation in unsupervised image animation. arXiv abs/2110.04658 (2021)", + "42. Miletari, F., Navab, N., Ahmadi, S.A.: V-Net: Fully convolutional neural networks for volumetric medical image segmentation. 3DV (2016)", + "43. Modat, M., Ridgway, G.R., Taylor, Z.A., Lehmann, M., Barnes, J., Hawkes, D.J., et al.: Fast free-form deformation using graphics processing units. Computer Methods and Programs in Biomedicine (2010)", + "44. Moor, M., Banerjee, O., Abad, Z.F.H., Krumholz, H.M., Leskovec, J., Topol, E.J., Rajpurkar, P.: Foundation models for generalist medical artificial intelligence. Nature (2023)", + "45. Müller, R.A., Shih, P., Keehn, B., Deyoe, J.R., Leyden, K.M., Shukla, D.: Underconnected, but how? A survey of functional connectivity MRI studies in autism spectrum disorders. Cerebral Cortex (2011)", + "46. Puonti, O., Iglesias, J.E., Leemput, K.V.: Fast and sequence-adaptive whole-brain segmentation using parametric bayesian modeling. NeuroImage (2016)", + "47. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015)", + "48. Sheller, M.J., Edwards, B., Reina, G.A., Martin, J., Pati, S., Kotrotsou, A., et al.: Federated learning in medicine: Facilitating multi-institutional collaborations without sharing patient data. Scientific reports (2020)", + "49. Shen, Z., Feydy, J., Liu, P., Curiale, A.H., San Jose Estepar, R., San Jose Estepar, R., Niethammer, M.: Accurate point cloud registration with robust optimal transport. In: NeurIPS (2021)", + "50. Singhal, K., Azizi, S., Tu, T., Mahdavi, S., Wei, J., Chung, H.W., et al.: Large language models encode clinical knowledge. Nature (2022)" + ], + "bbox": [ + 212, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Brain-ID", + "bbox": [ + 665, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "51. Symms, M.R., Jäger, H.R., Schmierer, K., Yousry, T.A.: A review of structural magnetic resonance neuroimaging. Journal of Neurology, Neurosurgery & Psychiatry (2004)", + "52. Tanno, R., Worrall, D.E., Kaden, E., Alexander, D.C.: Uncertainty modelling in deep learning for safer neuroimage enhancement: Demonstration in diffusion MRI. NeuroImage (2020)", + "53. Tian, Q., Bilgiç, B., Fan, Q., Ngamsombat, C., Zaretskaya, N., Fultz, N.E., et al.: Improving in vivo human cerebral cortical surface reconstruction using data-driven super-resolution. Cerebral Cortex (2020)", + "54. Tu, T., Azizi, S., Driess, D., Schaekermann, M., Amin, M., Chang, P.C., et al.: Towards generalist biomedical AI. arXiv abs/2307.14334 (2023)", + "55. de Vos, B.D., Berendsen, F.F., Viergever, M.A., Sokooti, H., Staring, M., Isgum, I.: A deep learning framework for unsupervised affine and deformable image registration. Medical Image Analysis (2019)", + "56. Wachinger, C., Golland, P., Reuter, M.: BrainPrint: Identifying subjects by their brain. In: MICCAI (2014)", + "57. Wang, M., Deng, W.: Deep visual domain adaptation: A survey. Neurocomputing (2018)", + "58. Wang, T.C., Mallya, A., Liu, M.Y.: One-shot free-view neural talking-head synthesis for video conferencing. In: CVPR (2021)", + "59. Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for image quality assessment. In: ACSSC (2003)", + "60. Weiner, M.W., Veitch, D.P., Aisen, P.S., Beckett, L.A., Cairns, N.J., Green, R.C., et al.: The Alzheimer's disease neuroimaging initiative 3: Continued innovation for clinical trial improvement. Alzheimer's & Dementia (2017)", + "61. White, T., Blok, E., Calhoun, V.D.: Data sharing and privacy issues in neuroimaging research: Opportunities, obstacles, challenges, and monsters under the bed. Human Brain Mapping (2022)", + "62. Xu, Y., Mo, T., Feng, Q., Zhong, P., Lai, M., Chang, E.I.C.: Deep learning of feature representation with multiple instance learning for medical image analysis. In: ICASSP (2014)", + "63. Yan, K., Cai, J., Jin, D., Miao, S., Harrison, A.P., Guo, D., et al.: SAM: Self-supervised learning of pixel-wise anatomical embeddings in radiological images. IEEE Transactions on Medical Imaging (2020)", + "64. Yang, X., Kwitt, R., Niethammer, M.: Quicksilver: Fast predictive image registration - a deep learning approach. NeuroImage (2017)", + "65. You, C., Zhao, R., Staib, L.H., Duncan, J.S.: Momentum contrastive voxel-wise representation learning for semi-supervised volumetric medical image segmentation. In: MICCAI (2021)", + "66. Zhou, Y., Chia, M.A., Wagner, S.K., Ayhan, M.S., Williamson, D.J., Struyven, R.R., et al.: A foundation model for generalizable disease detection from retinal images. Nature (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 729 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "P. Liu et al.", + "bbox": [ + 271, + 114, + 352, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Brain-ID_ Learning Contrast-agnostic Anatomical Representations for Brain Imaging/b53a82d6-784c-4409-ad49-9bb017717031_model.json b/2024/Brain-ID_ Learning Contrast-agnostic Anatomical Representations for Brain Imaging/b53a82d6-784c-4409-ad49-9bb017717031_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a82a934781e068db512ee2ce1dcc756e50c43d4a --- /dev/null +++ b/2024/Brain-ID_ Learning Contrast-agnostic Anatomical Representations for Brain Imaging/b53a82d6-784c-4409-ad49-9bb017717031_model.json @@ -0,0 +1,2480 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.221, + 0.142, + 0.784, + 0.189 + ], + "angle": 0, + "content": "Brain-ID: Learning Contrast-agnostic Anatomical Representations for Brain Imaging" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.212, + 0.685, + 0.228 + ], + "angle": 0, + "content": "Peirong Liu1 Oula Puonti1 Xiaoling Hu1" + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.228, + 0.673, + 0.244 + ], + "angle": 0, + "content": "Daniel C. Alexander\\(^{2}\\) \nJuan E. Iglesias\\(^{1,2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.255, + 0.734, + 0.285 + ], + "angle": 0, + "content": "\\(^{1}\\)Harvard Medical School and Massachusetts General Hospital \\(^{2}\\)University College London \\(^{3}\\)Massachusetts Institute of Technology" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.309, + 0.785, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.489, + 0.788, + 0.547 + ], + "angle": 0, + "content": "Fig. 1: Brain-ID features serve as the distinctive identity of each subject (ID) regardless of appearance (contrast, deformation, corruptions, etc) and quickly adapt to downstream tasks – either contrast-independent (anatomy Reconstruction, Segmentation), or contrast-dependent (Super-Resolution, Bias Field estimation), all through one layer." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.563, + 0.744, + 0.841 + ], + "angle": 0, + "content": "Abstract. Recent learning-based approaches have made astonishing advances in calibrated medical imaging like computerized tomography (CT). Yet, they struggle to generalize in uncalibrated modalities – notably magnetic resonance (MR) imaging, where performance is highly sensitive to the differences in MR contrast, resolution, and orientation. This prevents broad applicability to diverse real-world clinical protocols. We introduce Brain-ID, an anatomical representation learning model for brain imaging. With the proposed \"mild-to-severe\" intra-subject generation, Brain-ID is robust to the subject-specific brain anatomy regardless of the appearance of acquired images. Trained entirely on synthetic inputs, Brain-ID readily adapts to various downstream tasks through one layer. We present new metrics to validate the intra/inter-subject robustness of Brain-ID features, and evaluate their performance on four downstream applications, covering contrast-independent (anatomy reconstruction, brain segmentation), and contrast-dependent (super-resolution, bias field estimation) tasks (Fig. 1). Extensive experiments on six public datasets demonstrate that Brain-ID achieves state-of-the-art performance in all tasks on different MR contrasts and CT, and more importantly, preserves its performance on low-resolution and small datasets. Code is available at https://github.com/peirong26/Brain-ID." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.128 + ], + "angle": 0, + "content": "P. Liu et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.375, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.353 + ], + "angle": 0, + "content": "Magnetic resonance imaging (MRI) enables in vivo noninvasive imaging of the human brain with exquisite and tunable soft-tissue contrast [8]. Recent machine learning methods have achieved great improvements in faster and more accurate MRI analysis [27], such as segmentation [18,32,42,47], registration [3,49,55,64], super-resolution [52, 53], and connectivity studies [45]. However, most existing MRI analysis methods are specific to certain MR contrast(s) and require near-isotropic acquisitions. Therefore, models face sharp performance drops when voxel size and anisotropy increase, or are being used for a different contrast than training [57]. This reduces model generalizability and results in duplicate training efforts. Resorting to synthetic data, recent contrast-agnostic models [4, 25, 27, 28, 36, 39, 40] achieve impressive results and largely extend model applicability. However, these models are only applicable to the tasks they were trained for." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.788, + 0.581 + ], + "angle": 0, + "content": "Meanwhile, task-agnostic foundation models in computer vision and natural language processing have witnessed remarkable success, along with the fast developments of large-scale datasets [2,7,10,14,34]. However, due to different acquisition/processing protocols and privacy requirements across institutions, large-scale medical imaging datasets require significantly more effort than natural imaging/language. Thus, medical foundation models are not as well developed. MONAI [11] includes a pre-trained model zoo, yet all are highly task-oriented and sensitive to contrasts. Zhou et al. [66] constructed a medical foundation model, designed for eye and systemic health condition detections from retinal scans, and only works on the modalities of color fundus photography and optical coherence tomography. Lately, generalist biomedical AI systems [44, 50, 54] have shown great potential in biomedical tasks under vision-language context, e.g., (visual) question answering, image classification, radiology report generation and summarization. However, they have not explored the challenging vision tasks such as reconstruction, segmentation, super-resolution, and registration." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.582, + 0.787, + 0.612 + ], + "angle": 0, + "content": "We introduce Brain-ID, a contrast-agnostic feature representation approach trained on synthetic inputs, that can adapt to various tasks through one layer." + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.616, + 0.785, + 0.675 + ], + "angle": 0, + "content": "1) We introduce an on-the-fly, intra-subject data generator capable of synthesizing any contrast, with a mild-to-severe corruption strategy (Fig. 2, Sec. 3.2). Unlike real-world datasets constrained by the images acquired per subject, Brain-ID learns in a much more expansive and diverse space." + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.676, + 0.785, + 0.749 + ], + "angle": 0, + "content": "2) We design a feature representation learning framework guided by the unique anatomy of each subject. We validate that Brain-ID's high-resolution features are consistently robust to the superficial perturbations of an image's appearance (Sec. 4.2). Straightforward one-layer adaptions are further presented to adapt Brain-ID features to downstream tasks (Sec. 3.3)." + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.75, + 0.785, + 0.84 + ], + "angle": 0, + "content": "3) We evaluate Brain-ID on both contrast-independent (anatomy reconstruction, brain segmentation) and contrast-dependent (super-resolution, bias field estimation) tasks, across six public datasets (\\(\\approx\\) 8,000 images) including MR (T1w, T2w, FLAIR) and CT. Brain-ID achieves state-of-the-art performance on all tasks (Tab. 2), and further maintains its high performance on low-resolution images (Fig. 6) and limited-size datasets (Fig. 7)." + }, + { + "type": "list", + "bbox": [ + 0.221, + 0.616, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.666, + 0.116, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Brain-ID" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.383, + 0.161 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.4 + ], + "angle": 0, + "content": "Feature Representation in Medical Imaging As mentioned in Sec. 1, general feature representation learning in medical imaging can be more challenging than in the natural image domain, due to limited data availability. Xu et al. [62] introduced a multiple-instance learning model for feature representation, but it only applies to histopathology image classification. You et al. [65] presented CVRL, a semi-supervised representations approach designed for image segmentation, but it requires CT to extract anatomical information. SAM [63] encodes anatomical information in feature embeddings, which has shown to be effective for registration [38]. However, same as CVRL, SAM only works on CT. BrainPrint [56] is a compact and discriminative representation of brain morphology, which is specifically designed for cortical surface analyses. To our best knowledge, CIFL [15] is the only existing work on learning contrast-agnostic and task-independent brain feature representations. CIFL relies on contrastive learning alone, and is insufficient to outperform task/contrast-specific supervised models in downstream applications as shown by our experiments (Tab. 2)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.788, + 0.584 + ], + "angle": 0, + "content": "Contrast-invariant Learning for MR Images MRI scans acquired across sites vary substantially in contrasts, resolutions, orientations, etc. When given a new dataset, heterogeneity leads to duplicate training efforts for approaches that are sensitive to specific combinations of MR contrast, resolution, or orientation. Classical brain segmentation models used Bayesian inference for contrast-robustness [20,37], which requires a long processing time and struggles with low or anisotropic resolutions [27, 46]. Recently, there have been works using synthetic data to achieve contrast-invariance in tasks like image segmentation [4,36], registration [25], super-resolution [28], and skull-stripping [26]. However, all the above-mentioned methods are trained in a task-specific manner, their features therefore cannot be readily applied to other domains." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.602, + 0.755, + 0.62 + ], + "angle": 0, + "content": "3 Brain-ID: Learning Anatomy-specific Brain Features" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.787, + 0.703 + ], + "angle": 0, + "content": "As discussed in Sec. 1, the main challenges to obtain a general and robust feature representation for MR images lie in \\((i)\\) the practical restrictions of building large-scale datasets with diverse contrasts; and \\((ii)\\) the nature of most medical imaging models that are task-oriented and specific to data type (contrast, resolution, orientation, etc). We aim to learn a brain feature representation that is:" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.705, + 0.786, + 0.733 + ], + "angle": 0, + "content": "(i) Robust: features should be robust to each subject's distinct anatomy, unaffected by variations in poses/deformations, contrasts, resolutions, or artifacts." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.735, + 0.786, + 0.779 + ], + "angle": 0, + "content": "(ii) Expressive: features should also exhibit high expressiveness, containing rich information that facilitates easy and effective adaptation to diverse downstream tasks, eliminating the necessity for extensive training data." + }, + { + "type": "list", + "bbox": [ + 0.215, + 0.705, + 0.786, + 0.779 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "We first introduce Brain-ID's data generator (Sec. 3.1) and training framework (Sec. 3.2) to achieve the above two aims. Then, we present our one-layer solutions for adapting Brain-ID features to downstream tasks that could be either dependent or independent of the contrast of input images (Sec. 3.3)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "P. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.144, + 0.782, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.262, + 0.788, + 0.304 + ], + "angle": 0, + "content": "Fig.2: Brain-ID's data generator on the fly. Given the brain segmentation labels of a subject, we randomly generate a deformation field, and synthesize intra-subject samples featuring various contrast intensities and corruption levels (Sec. 3.1)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.331, + 0.594, + 0.347 + ], + "angle": 0, + "content": "3.1 Enrich the Intra-subject Learning Space" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.788, + 0.458 + ], + "angle": 0, + "content": "A good representation relies on large-scale data, however, different acquisition protocols, processing pipelines, and privacy requirements across institutions make large-scale data less accessible and require significantly more effort [1, 17, 48]. Moreover, for each subject, there are usually only limited acquired scans available. This lack of data consistency is a significant barrier to obtaining a robust, contrast-invariant feature representation. Brain-ID avoids these barriers through the use of synthetic data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.535 + ], + "angle": 0, + "content": "To generate images with complex brain structures, we start with brain segmentation labels of anatomical structures (\\(L\\) in Fig. 2). The generation consists of three steps, \\((i)\\) deformation generation, \\((ii)\\) contrast synthesis, and \\((iii)\\) data corruption (including lower resolution resampling). For simplicity, \\(\\Theta\\) denotes the parameter group of the generation process described below." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.538, + 0.787, + 0.569 + ], + "angle": 0, + "content": "Deformation Generation We first generate a random deformation field \\((\\phi|_{\\theta_{\\phi}})\\) consisting of an affine transformation and a non-linear displacement field:" + }, + { + "type": "equation", + "bbox": [ + 0.429, + 0.578, + 0.786, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\phi | _ {\\theta_ {\\phi}} = \\mathcal {T} | _ {\\theta_ {\\phi}} \\circ \\mathcal {A} | _ {\\theta_ {\\phi}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.788, + 0.678 + ], + "angle": 0, + "content": "where \\(\\mathcal{A}|_{\\theta_{\\phi}}\\) denotes an affine transformation matrix which includes linear (rotation, scaling, shearing) transformation and translation, \\(\\mathcal{T}|_{\\theta_{\\phi}}\\) refers to a nonlinear displacement field computed as the integration of a stationary velocity field (SVF) that is smooth and invertible everywhere and thus preserves the topology of the brain anatomy [28]. \\(\\theta_{\\phi} \\in \\Theta\\) controls the transformation ranges." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.68, + 0.788, + 0.742 + ], + "angle": 0, + "content": "MR Contrast Synthesis Then, we synthesize images \\((I(x), x \\in \\Omega)\\) by randomly \"painting\" intensities on the segmentation maps according to their brain structure labels \\((l \\in L)\\). Specifically, the regional intensities are generated by separately sampling a Gaussian distribution on each labeled region:" + }, + { + "type": "equation", + "bbox": [ + 0.332, + 0.748, + 0.786, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} I (x) \\sim \\mathcal {N} \\left(\\mu_ {l}, \\sigma_ {l}\\right), \\quad l \\in L, \\\\ \\mu_ {l} \\sim \\mathcal {N} \\left(0, 1 \\mid \\theta_ {\\mu}, \\theta_ {l}\\right), \\sigma_ {l} \\sim \\mathcal {N} \\left(0, 1 \\mid \\theta_ {\\sigma}, \\theta_ {l}\\right), \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\(\\mu_{l}\\) and \\(\\sigma_{l}\\) refer to the mean and standard deviation of each segmentation label \\(l\\), and are independently sampled from Gaussian distributions at each voxel, with \\(\\theta_{l}, \\theta_{\\mu}, \\theta_{\\sigma} \\in \\Theta\\) controlling the shifts and scales of their values." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.667, + 0.116, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Brain-ID" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.144, + 0.784, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.222, + 0.292, + 0.779, + 0.307 + ], + "angle": 0, + "content": "Fig. 3: Brain-ID's contrast-agnostic anatomical representation learning framework." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.336, + 0.784, + 0.396 + ], + "angle": 0, + "content": "Resolution Simulation and Data Corruption Given a deformed, contrast-synthesized image \\((I)\\), we adopt the standard data corruption pipeline [27], which further augments images with different levels of resolution and scanning artifacts that are commonly found in real-world clinical protocols." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.785, + 0.487 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, we are able to generate an infinite number of variations from a single subject with its unique brain anatomy. Generating images with randomized contrast/resolution/orientation on the fly for each subject enormously enriches the learning space for a robust representation, and helps focus the learning on intrinsic subject-specific features rather than superficial aspects of images that depend on acquisition parameters and conditions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.51, + 0.729, + 0.525 + ], + "angle": 0, + "content": "3.2 Extract Robust and Expressive Subject-specific Features" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.784, + 0.58 + ], + "angle": 0, + "content": "As mentioned above, we would like Brain-ID features to be robust to intra-subject variations and expressive to potential downstream tasks. In this section, we introduce Brain-ID's learning framework to achieve the two goals." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.785, + 0.764 + ], + "angle": 0, + "content": "Intra-subject Data Generation In order to learn a feature representation that is distinctive to each subject and robust to varying MR contrasts, we observe that enriching intra-subject samples leads to better performance (Tab. 3). Specifically, instead of including multiple subjects for a mini-batch during each training iteration as in usual practice, Brain-ID focuses on maximizing the intra-subject variance to improve the subject-robustness of resulting features. As described in Alg. 1 (line 4-10), for each training iteration, after randomly selecting a subject and generating its deformation, Brain-ID generates a mini-batch of intra-subject samples \\(\\{(S_1,\\ldots ,S_N)\\}\\) with randomly synthesized contrasts, resolutions and corruptions (Sec. 3.1). As will be introduced below, Brain-ID collects losses from all intra-subject samples and conducts back-propagation at once, to encourage the subject-specific robustness of its learned features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.785, + 0.841 + ], + "angle": 0, + "content": "We set the intra-subject samples within a mini-batch to have random contrasts and \"mild-to-severe\", increasing level of corruptions (Fig. 3 (left)), to maximize the intra-subject variance while ensuring the stability of the training process against extreme corruption levels. (In Sec. 4.5, we compare various data generation designs, and provide insights on preventing unstable training.)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.128 + ], + "angle": 0, + "content": "P. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "Anatomy-guided Feature Representation A richer learning space helps improve the representation robustness against the variance of sample appearances, but not enough for extracting expressive features. For example, a mapping that projects all inputs to zero is perfectly robust, but does not provide useful information for potential downstream tasks. Therefore, proper guidance is crucial for feature representation as well. In CIFL [15], the authors use a contrastive loss on feature channels to encourage a more discriminative feature representation. However, this method is insufficient, given the extremely complex nature of the human brain anatomy (See comparisons in Sec. 4). Instead, we propose to use the standard T1w high-resolution structural MR contrast for brain morphometry, i.e., MP-RAGE (magnetization-prepared rapid gradient-echo), as the unique anatomy target to guide Brain-ID's feature representation learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.419 + ], + "angle": 0, + "content": "As shown in Fig. 3, a feature extraction backbone \\((\\mathcal{F})\\) first maps the input mini-batch of intra-subject generated samples, \\(\\{S_1,\\dots,S_N\\}\\), to their corresponding feature space, \\(\\{\\mathbf{F}_1,\\ldots ,\\mathbf{F}_N\\}\\). A linear activation layer \\((\\mathcal{L})\\) then projects the features to the current subject's standard contrast (MP-RAGE) space, \\(T\\). The training loss is obtained by summing over the anatomy reconstruction loss of all intra-subject samples in the current mini-batch:" + }, + { + "type": "equation", + "bbox": [ + 0.315, + 0.428, + 0.785, + 0.453 + ], + "angle": 0, + "content": "\\[\nL = \\sum_ {i} ^ {N} | \\mathcal {L} (\\mathbf {F} _ {i}) - T | + \\lambda | \\nabla \\mathcal {L} (\\mathbf {F} _ {i}) - \\nabla T |, \\lambda \\in R ^ {+}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.459, + 0.788, + 0.567 + ], + "angle": 0, + "content": "where \\(T\\), as the high-resolution, unique anatomy target for all intra-subject samples, encourages the similarities of the features extracted from the same subject via reconstruction (\\(1^{\\mathrm{st}}\\) term) and gradient difference (\\(2^{\\mathrm{nd}}\\) term) L1 loss. Brain-ID uses MP-RAGE as the learning target, which formulates both super-resolution and contrast-synthesis problems, and encourages a richer feature representation. (Sec. 4.5 provides further insights on how different choices of anatomy guidance would affect the resulting features' properties.)" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.585, + 0.715, + 0.601 + ], + "angle": 0, + "content": "3.3 Adapting Brain-ID to Downstream Tasks by One Layer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.788, + 0.729 + ], + "angle": 0, + "content": "With the intra-subject data generation and anatomy-guided feature learning design introduced in Sec. 3.2, a well-trained Brain-ID model is able to extract robust, high-resolution anatomical features from images with varying deformations, contrasts, resolutions, and artifacts. To minimize the modifications, we propose straightforward one-layer solutions adapting Brain-ID features to various brain imaging applications. We later demonstrate that the simple adaptions are effective enough for Brain-ID to achieve state-of-the-art performance across all downstream tasks (Sec. 4.3), even for small datasets (Sec. 4.4)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.731, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Contrast-independent Tasks For tasks where the output should be independent of the input MR contrast, e.g., brain segmentation, we simply add an additional layer following Brain-ID features \\((\\mathbf{F})\\), and fine-tune the model:" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.786, + 0.785, + 0.803 + ], + "angle": 0, + "content": "\\[\nL = \\operatorname {t a s k} _ {-} \\operatorname {l o s s} _ {-} \\operatorname {f u n c} \\left(\\mathcal {L} (\\mathbf {F}), T\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}\\) and \\(T\\) refer to the task-specific activation layer and the contrast-independent ground truth, respectively." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.666, + 0.116, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Brain-ID" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.222 + ], + "angle": 0, + "content": "Contrast-dependent Tasks Since Brain-ID features are contrast-agnostic and robust to artifacts, for tasks relevant to the input's contrast/quality, e.g., super-resolution, we concatenate the input image \\((I)\\) that contains the original contrast information with its high-resolution Brain-ID features along the channel dimension, before forwarding into the task-specific activation layer:" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.235, + 0.784, + 0.25 + ], + "angle": 0, + "content": "\\[\nL = \\operatorname {t a s k} _ {-} \\operatorname {l o s s} _ {-} \\operatorname {f u n c} \\left(\\mathcal {L} (\\mathbf {F} \\oplus I), T\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.261, + 0.78, + 0.276 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}\\) and \\(T\\) refer to the task-specific activation layer and the ground truth." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.298, + 0.375, + 0.315 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.784, + 0.42 + ], + "angle": 0, + "content": "We demonstrate the two properties of Brain-ID as claimed in Sec. 3. (i) Robustness (Sec. 4.2): we propose \"canonical\" and \"atlas-registered\" features to assess the intra/inter-subject feature robustness; (ii) Expressiveness (Sec. 4.3): we evaluate the performance of adapting Brain-ID features to various brain imaging applications. We further challenge Brain-ID with reduced-size datasets to explore its ability when limited real data is available for training (Sec. 4.4)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.442, + 0.644, + 0.457 + ], + "angle": 0, + "content": "4.1 Datasets, Metrics and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.47, + 0.784, + 0.53 + ], + "angle": 0, + "content": "Datasets For Brain-ID features pre-training, we use 2045 3D segmentations (\"anatomies\", or IDs) from the public ADNI dataset [30]. For downstream evaluation, we use six public datasets covering T1w and T2w MRI, FLAIR MRI, and CT: ADNI [30], ADNI3 [60], HCP [19], ADHD200 [9], AIBL [23], OASIS3 [35]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.784, + 0.623 + ], + "angle": 0, + "content": "Metrics We evaluate individual tasks from different aspects. For feature similarity measurements, we use L1 distance, and (MS-)SSIM (multi-scale structural similarity) [41,58,59]. For reconstruction and super-resolution, we use L1, PSNR (peak signal-to-noise ratio) and (MS-)SSIM. For segmentation, we use Dice scores. For bias field estimation, we use the normalized L2 distance (norm-L2) to avoid possible arbitrary scalings from nonuniformity correction [16]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.784, + 0.718 + ], + "angle": 0, + "content": "Implementation Details As a general feature representation model, Brain-ID can use any backbone to extract brain features. For fairer comparison, we adopt the same five-level 3D UNet [47] (with 64 feature channels in the last layer) as utilized in state-of-the-art models we compare with in Sec. 4.3. During feature pre-training (Sec. 3.2), a linear regression layer is added for anatomy supervision. For downstream adaptions, A task-specific adaption layer is added (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.741, + 0.587, + 0.755 + ], + "angle": 0, + "content": "4.2 Intra/Inter-subject Feature Robustness" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.84 + ], + "angle": 0, + "content": "In this section, we examine the robustness of Brain-ID features, i.e., the 64 features from the last layer. For comparable and reproducible results, we use our data generator (Sec. 3.1) to prepare (deform and corrupt) 1000 testing samples from 100 randomly selected subjects in ADNI [30] (T1w) testing set, with 10 intra-subject samples for each subject." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "P. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.145, + 0.782, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.543, + 0.765, + 0.558 + ], + "angle": 0, + "content": "Fig. 4: (a) Intra-subject and (b) inter-subject robustness of Brain-ID features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.687 + ], + "angle": 0, + "content": "We compare our features with CIFL [15] which is, to our best knowledge, the only similar work to Brain-ID, that also aims to learn contrast-agnostic brain features. Note the original CIFL method has only experimented on 2D images, and does not have our intra-subject data generation design. For a fairer comparison, we trained CIFL with its contrastive-learning design, while using the same data generator and network architecture as Brain-ID." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.695, + 0.788, + 0.755 + ], + "angle": 0, + "content": "Intra-subject Feature Robustness Ideally, Brain-ID features computed from the same subject should be structurally identical, regardless of their appearances (contrast, resolution, noises, etc). To assess the intra-subject feature robustness, we first define their corresponding \"canonical\" features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.759, + 0.788, + 0.806 + ], + "angle": 0, + "content": "\"Canonical\" Features. We compute \\(\\phi^{-1}\\), the inverse of \\(S\\)'s generated deformation as in Eq. (1), and map back each sample's \\((S)\\) Brain-ID features \\((\\mathbf{F} = \\mathbf{F}(S|\\phi))\\) to their original domain without any deformation:" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.823, + 0.786, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} ^ {- 1} = \\phi^ {- 1} \\circ \\mathbf {F} (S \\mid \\phi). \\tag {6}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.666, + 0.116, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Brain-ID" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.289, + 0.145, + 0.712, + 0.159 + ], + "angle": 0, + "content": "Table 1: Evaluations on intra/inter-subject feature robustness." + }, + { + "type": "table", + "bbox": [ + 0.32, + 0.17, + 0.684, + 0.243 + ], + "angle": 0, + "content": "
ModeMethodL1 (↓)SSIM (↑)MS-SSIM (↑)
IntraCIFL [15]0.122 (±0.031)0.511 (±0.298)0.531 (±0.367)
Brain-ID0.011 (±0.001)0.858 (±0.041)0.921 (±0.008)
InterCIFL [15]0.230 (±0.039)0.552 (±0.240)0.524 (±0.291)
Brain-ID0.014 (±0.013)0.843 (±0.044)0.913 (±0.011)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.267, + 0.788, + 0.404 + ], + "angle": 0, + "content": "As shown in Fig. 4-(a), for the same subject (ID-i), although the inputs (row) are processed by varying deformations and corruptions, their Brain-ID's \"canonical\" features are of similar looking along each feature channel (column). In Tab. 1, we quantify the feature similarity: For each testing subject, we first generate a sample without any deformation or corruption and use its Brain-ID feature as the gold-standard reference for all other \"canonical\" features computed from intra-subject samples that may have different levels of deformation and corruptions. The final results are obtained by scores averaged over all 64 feature channels of all the intra-subject samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.788, + 0.483 + ], + "angle": 0, + "content": "Inter-subject Feature Robustness We further assess the feature robustness among anatomies, based on the assumption that features from different subjects should ideally be structurally similar, once being registered to a common domain. \"Atlas-registered\" Features. We register [43] features \\((\\mathbf{F} = \\mathbf{F}(S|\\phi))\\) from all subjects to a common brain atlas [21, 22]:" + }, + { + "type": "equation", + "bbox": [ + 0.421, + 0.492, + 0.786, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathbf {F}}} = \\psi_ {S \\rightarrow A} \\circ \\mathbf {F} (S \\mid \\phi), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.633, + 0.534 + ], + "angle": 0, + "content": "where \\(\\psi_{S\\to A}\\) denotes the mapping from \\(S\\) to the atlas, \\(A\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.686 + ], + "angle": 0, + "content": "As shown in Fig. 4-(b), even though the input samples (row) from different subjects have varying anatomical structures, after being registered to a common domain, their \"atlas-registered\" Brain-ID features still appear consistent (column). In Tab. 1, we further provide quantitative results of the structural similarity measured between features from different subjects. Similar to Sec. 4.2, we randomly select a testing subject and generate a sample without any deformation or corruption. We use its \"atlas-registered\" Brain-ID feature as the gold standard for all \"atlas-registered\" features computed from other samples that may have different levels of deformation and corruption. The final results are obtained by scores averaged over all samples of the 100 testing subjects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.686, + 0.788, + 0.732 + ], + "angle": 0, + "content": "In both intra/inter-subject aspects, Brain-ID outperforms CIFL by a large margin, which validates its feature robustness against deformations, contrasts, resolutions, artifacts (intra-subject), and anatomies (inter-subject)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.752, + 0.46, + 0.767 + ], + "angle": 0, + "content": "4.3 Downstream Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.776, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Following the standard evaluation protocol for representation learning [12], we fine-tune Brain-ID features for downstream tasks through one-layer (Sec. 3.3). Models We compare the downstream performance of Brain-ID features with: \\((i - ii)\\) SAMSEG [13,46] and FastSurfer [24] (only works on T1w), state-of-the-art" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "P. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.147, + 0.777, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.364, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparisons on downstream tasks of reconstruction (Recon), superresolution (SR), and segmentation (Seg), between Brain-ID, the baseline SCRATCH, and the state-of-the-art methods CIFL [15], SynthSR [27] (Recon and SR), SAMSEG [13] (Seg). The visualized testing examples are taken from: AIBL-FLAIR [23] for Recon, AIBLT1w [23] for SR, and OASIS-CT [35] for Seg. The mint circles highlight some details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.468, + 0.788, + 0.589 + ], + "angle": 0, + "content": "classical and machine-learning-based brain segmentation models, respectively; (iii) SynthSR [27], state-of-the-art, contrast-agnostic T1w synthesis model. We also provide fine-tuned SynthSR (SynthSR-FT) results for further comparison; (iv) SCRAW: a baseline with the same data generation/architecture as Brain-ID, yet trained from scratch. SCRAW is used to validate Brain-ID's effectiveness. (v) CIFL [15]: a baseline with the same data generation/architecture as Brain-ID and SCRAW, yet initialized with CIFL's pre-trained features. We use CIFL to demonstrate the superiority of Brain-ID's anatomy-guided representation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.594, + 0.509, + 0.609 + ], + "angle": 0, + "content": "Contrast-independent Applications" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.788, + 0.686 + ], + "angle": 0, + "content": "- Anatomy Reconstruction/Contrast Synthesis As one of the most common types of MRI scans, T1w MR images highlight the differences between gray and white matter, and are mostly often utilized to image the anatomy of the brain, spinal cord, bones and joints [51]. For each dataset, we train models to reconstruct their T1w MRI counterparts with the L1 loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.788, + 0.749 + ], + "angle": 0, + "content": "- Brain Segmentation For each dataset, we use SynthSeg [4] to obtain the segmentation labels with 30 brain anatomical regions [20], as the gold standard segmentation target. We train our compared models to predict the brain segmentation labels, with the soft dice loss and cross-entropy loss [4]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "As shown in Fig. 5, Brain-ID clearly outperforms CIFL and SCRATCH in the reconstruction task. As highlighted by the mint circles, it reveals the anatomy details and produces more fine-grained results than the strong state-of-the-art method SynthSR [27]. Brain-ID also achieves better segmentation results, especially in the smaller and more challenging regions. Quantitative comparisons are in Tab. 2, where Brain-ID obtains the best scores across most metrics. No-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.667, + 0.116, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Brain-ID" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 2: Comparisons of Brain-ID with state-of-the-art approaches on downstream applications. (\"train/test\" refers to the number of subjects in the training/testing set.)" + }, + { + "type": "table", + "bbox": [ + 0.239, + 0.184, + 0.764, + 0.619 + ], + "angle": 0, + "content": "
Modality (Contrast)Dataset (Train/Test)MethodReconstructionSegmentationSuper-resolution (from 3 / 5 / 7 mm)Bias Field norm-12 (↓)
L1 (↓)PSNR (↑)SSIM (↑)MS-SSIM (↑)Dice (↑)PSNR (↑)SSIM (↑)MS-SSIM (↑)MS-SSIM (↑)
MR (T1w)ADNI [30] (1841/204)SAMSEG [13]---0.795------
FastSurfer [24]---0.803------
SynthSR [27]0.01426.780.9800.980-27.01 / 25.81 / 21.790.977 / 0.943 / 0.8570.981 / 0.968 / 0.869--
SynthSR-FT0.01029.540.9840.9200.82530.46 / 28.03 / 23.910.979 / 0.960 / 0.8980.985 / 0.973 / 0.900-0.051
CIFL [15]0.01326.970.9730.9610.82030.12 / 27.21 / 23.800.978 / 0.912 / 0.8870.979 / 0.950 / 0.842-0.053
SCRATCH0.01127.540.9790.9570.81630.29 / 27.94 / 24.570.975 / 0.959 / 0.9200.982 / 0.964 / 0.866-0.056
Brain-ID0.01233.820.9930.9890.83731.30 / 28.62 / 24.680.983 / 0.961 / 0.9280.987 / 0.980 / 0.947-0.048
MR (T1w)HCP [19] (808/87)SAMSEG [13]---0.810------
FastSurfer [24]---0.819------
SynthSR [27]0.03322.130.8540.901-22.21 / 20.83 / 19.930.848 / 0.802 / 0.7470.899 / 0.864 / 0.789--
SynthSR-FT0.02527.110.9350.9170.81025.90 / 25.45 / 25.100.889 / 0.872 / 0.8310.919 / 0.911 / 0.889-0.052
CIFL [15]0.02926.420.9320.9130.80425.98 / 25.11 / 23.990.930 / 0.883 / 0.8710.936 / 0.927 / 0.880-0.059
SCRATCH0.02327.320.9230.9090.79226.69 / 26.41 / 25.050.932 / 0.896 / 0.8480.942 / 0.930 / 0.892-0.069
Brain-ID0.02027.470.9570.9290.84329.70 / 27.90 / 26.120.957 / 0.901 / 0.8830.973 / 0.938 / 0.890-0.047
SAMSEG [13]---0.769------
FastSurfer [24]---0.796------
SynthSR [27]0.02323.600.9280.909-23.13 / 22.40 / 22.270.921 / 0.907 / 0.8760.903 / 0.892 / 0.871--
SynthSR-FT0.02127.110.9500.9690.82027.95 / 27.34 / 26.180.920 / 0.893 / 0.9220.890 / 0.890 / 0.879-0.134
CIFL [15]0.02828.520.9610.9700.81927.81 / 27.10 / 26.320.900 / 0.877 / 0.8640.812 / 0.793 / 0.769-0.102
SCRATCH0.03327.280.9570.9630.81627.99 / 27.37 / 26.630.893 / 0.889 / 0.8760.794 / 0.785 / 0.753-0.128
Brain-ID0.02129.890.9660.9760.84330.17 / 28.23 / 26.890.962 / 0.923 / 0.8920.973 / 0.921 / 0.883-0.093
SAMSEG [13]---0.784------
FastSurfer [24]---0.801------
SynthSR [27]0.03521.670.8820.853-21.42 / 21.13 / 21.410.876 / 0.846 / 0.8090.851 / 0.831 / 0.805--
SynthSR-FT0.01626.900.9170.9050.81228.55 / 27.43 / 25.940.880 / 0.859 / 0.8230.860 / 0.845 / 0.811-0.110
CIFL [15]0.01328.690.9320.9210.81028.11 / 26.12 / 25.110.865 / 0.848 / 0.8310.828 / 0.800 / 0.789-0.109
SCRATCH0.01131.550.9860.9850.79629.41 / 27.89 / 26.580.854 / 0.847 / 0.8140.853 / 0.845 / 0.827-0.113
Brain-ID0.01132.480.9960.9960.84729.50 / 28.56 / 26.870.898 / 0.862 / 0.8110.887 / 0.850 / 0.823-0.107
SAMSEG [13]---0.799------
FastSurfer [24]---0.802------
SynthSR [27]0.02622.950.9160.912-22.57 / 21.82 / 21.780.917 / 0.892 / 0.8560.906 / 0.893 / 0.869--
SynthSR-FT0.01429.890.9410.9220.81029.15 / 27.89 / 26.770.932 / 0.906 / 0.8780.927 / 0.900 / 0.871-0.118
CIFL [15]0.01130.120.9380.9250.81629.46 / 27.97 / 26.730.907 / 0.889 / 0.8700.891 / 0.872 / 0.857-0.094
SCRATCH0.01130.530.9320.9130.81429.43 / 28.10 / 26.820.900 / 0.886 / 0.8630.884 / 0.870 / 0.853-0.128
Brain-ID0.00933.730.9720.9630.85129.75 / 28.58 / 27.090.957 / 0.922 / 0.8900.975 / 0.956 / 0.937-0.088
SAMSEG [13]---0.782------
FastSurfer [24]---0.805------
SynthSR [27]0.02826.100.8940.8980.75528.43 / 25.36 / 22.100.942 / 0.851 / 0.8010.977 / 0.892 / 0.798-0.140
SynthSR-FT0.03625.120.8910.8790.78728.26 / 25.78 / 21.730.959 / 0.893 / 0.7970.982 / 0.906 / 0.742-0.138
SCRATCH0.03824.990.8730.8660.75628.12 / 24.52 / 21.580.945 / 0.863 / 0.7830.980 / 0.893 / 0.738-0.136
Brain-ID0.01628.100.9340.9350.80530.26 / 26.11 / 24.100.959 / 0.902 / 0.8320.987 / 0.953 / 0.912-0.136
SAMSEG [13]---0.763------
FastSurfer [24]---0.801------
SynthSR [27]0.03320.080.8050.831------
SynthSR-FT0.02422.930.8150.8400.71930.15 / 28.09 / 26.920.940 / 0.912 / 0.8930.966 / 0.948 / 0.922-0.200
CIFL [15]0.02223.710.8200.8390.72129.98 / 27.01 / 25.660.931 / 0.899 / 0.8610.967 / 0.941 / 0.905-0.193
SCRATCH0.02022.270.8490.8510.71431.91 / 29.20 / 27.350.954 / 0.934 / 0.8960.982 / 0.969 / 0.949 / 0.936-0.197
Brain-ID0.02223.990.8610.8500.78232.26 / 29.90 / 27.090.937 / 0.930 / 0.9080.988 / 0.971 / 0.948-0.148
SAMSEG [13]---0.718------
SynthSR [27]0.02622.770.9190.895-30.11 / 28.70 / 25.320.930 / 0.899 / 0.8640.948 / 0.910 / 0.878-0.251
SynthSR-FT0.02123.340.9210.9000.75332.72 / 29.00 / 26.990.949 / 0.906 / 0.8730.953 / 0.919 / 0.889-0.237
CIFL [15]0.02021.290.9110.8970.76132.72 / 28.71 / 25.860.945 / 0.915 / 0.8770.941 / 0.936 / 0.917-0.136
SCRATCH0.02520.800.9010.8620.75932.36 / 28.71 / 25.840.945 / 0.915 / 0.8770.981 / 0.953 / 0.917-0.236
Brain-ID0.01726.440.9270.8920.78632.18 / 30.00 / 28.280.959 / 0.921 / 0.8830.982 / 0.965 / 0.921-0.227
SAMSEG [13]---0.710------
FastSurfer [24]---0.795------
SynthSR [27]0.02921.770.9020.892------
SynthSR-FT0.02425.800.9140.9000.73526.20 / 24.91 / 23.170.860 / 0.832 / 0.7950.933 / 0.910 / 0.894-0.126
CIFL [15]0.02627.110.8910.8700.72127.72 / 26.85 / 25.950.821 / 0.858 / 0.8140.969 / 0.939 / 0.905 / 0.901-0.109
SCRATCH0.02326.840.8980.8670.72028.02 / 26.52 / 25.430.878 / 0.838 / 0.7990.943 / 0.901 / 0.891-0.115
Brain-ID0.01927.250.9360.9120.76726.69 / 24.01 / 23.69 / 23.09 / 0.918 / 0.806 / 0.966 / 0.871 / 0.925 / 0.916---
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.241, + 0.62, + 0.696, + 0.627 + ], + "angle": 0, + "content": "(1) \"means not applicable: FastSurfer [24] and SAMSEG [13] are for segmentation-only; SynthSR [27] only synthesize T1w MRI; CT does not have bias fields. (2) SynthSR ET; SynthSR fine tuned on each task/dataset, for a fairer comparison." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.672, + 0.784, + 0.718 + ], + "angle": 0, + "content": "tably, Brain-ID achieves greater gains over baseline models, SCRATCH and CIFL, particularly on smaller datasets (e.g., ADNI3-T1/FLAIR, AIBL-FLAIR, where Brain-ID's robust and rich features quickly adapt to specific tasks/datasets." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.728, + 0.784, + 0.759 + ], + "angle": 0, + "content": "Contrast-dependent Applications We also evaluate Brain-ID on two brain imaging tasks that are dependent on the input modality/quality." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.765, + 0.784, + 0.841 + ], + "angle": 0, + "content": "- Image Super-resolution We use the standard 1mm isotropic images from all datasets as the ground truth high-resolution target, and follow the strategy in SynthSR [27] where the input samples are randomly resampled and corrupted during training. For inference, we downsample the original images into 3mm, 5mm, and 7mm isotropic images as inputs, and evaluate the output quality." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "P. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.239, + 0.146, + 0.77, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.373, + 0.788, + 0.416 + ], + "angle": 0, + "content": "Fig. 6: SCRATCH, which is well trained on HF T1w scans, produces highly descriptive features for HF T1w images (\\(1^{\\mathrm{st}}\\) row), but does not preserve the same high quality useful for downstream tasks when handling LF (\\(2^{\\mathrm{nd}}\\) row) or other contrasts (\\(3^{\\mathrm{rd}}\\) row)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.788, + 0.55 + ], + "angle": 0, + "content": "- Bias Field Estimation The bias field is a smooth, low-frequency multiplicative signal that corrupts MRI images, which affects image analysis tasks such as segmentation or texture analysis [31]. Bias field estimation is often needed as a pre-processing step to correct corrupted MRI images [16]. We apply randomly simulated bias fields to the input samples (Sec. 3.1), and train all models with the L2 loss. During inference, we pre-generate and apply the bias fields on the testing data for reproducibility, and evaluate the bias field estimation performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.672 + ], + "angle": 0, + "content": "With a simple one-layer adaption (Sec. 3.3), Brain-ID's contrast-agnostic anatomical features achieve state-of-the-art performance on contrast-dependent tasks (Tab. 2, Fig. 5). As shown in Fig. 6, Brain-ID obtains contrast/resolution-robustness that cannot be achieved by models trained from real images (due to the limited variability in their appearance), regardless of the backbone choices. Such robust features greatly improve super-resolution, resulting in higher (MS-)SSIM scores. Brain-ID's gains are less obvious for bias field estimation, probably because the bias field is approximately independent of brain anatomy." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.692, + 0.562, + 0.708 + ], + "angle": 0, + "content": "4.4 Practical Value and Broader Impact" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.716, + 0.787, + 0.776 + ], + "angle": 0, + "content": "Low-resolution data Trained on synthetic images, Brain-ID features are not only robust to various contrasts/modalities, but also extremely robust to low-resolution data (Fig. 6) and provide huge potential in clinical MRI (including big retrospective data [5]) and portable low-field MRI [29]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Small-size Datasets To investigate the effectiveness of Brain-ID features for limited-size datasets, we assess its performance across all four downstream tasks on subsets of ADNI3 [60] training set (298 cases originally). As shown in Fig. 7, we reduce the training set size percentage from \\(100\\%\\) to an extreme \\(10\\%\\). With" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.667, + 0.116, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Brain-ID" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.145, + 0.77, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.467, + 0.495, + 0.478 + ], + "angle": 0, + "content": "(c) Super-resolution" + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.467, + 0.728, + 0.477 + ], + "angle": 0, + "content": "(d) Bias Field Estimation" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.491, + 0.788, + 0.534 + ], + "angle": 0, + "content": "Fig. 7: Adapting Brain-ID to small datasets. The horizontal (vertical) axes indicate training epochs (evaluation scores of corresponding tasks). Results are obtained by evaluating models collected throughout the epochs, on ADNI3 [60] full testing set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.563, + 0.788, + 0.64 + ], + "angle": 0, + "content": "only \\(20\\%\\) of the data (mint line), Brain-ID still achieves better performance compared to its full-sized baseline, “\\(100\\%\\)” SCRATCH (yellow line), over all tasks. 30 training samples (\\(10\\%\\), blue line) may be at the edge of obtaining acceptable results for Brain-ID, as the model becomes less stable and the performance drops occur more frequently – especially for brain segmentation (Fig. 7-(b))." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.661, + 0.46, + 0.677 + ], + "angle": 0, + "content": "4.5 Additional Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.686, + 0.788, + 0.777 + ], + "angle": 0, + "content": "Choice of Anatomy Guidance We explore other anatomical learning targets than MP-RAGE (Sec. 3.2): \\((i)\\) segmentation labels; \\((ii)\\) both segmentation labels and MP-RAGE. As shown in Tab. 3, adding segmentation guidance improves robustness, yet inhibits the feature expressiveness and affects the downstream performance; we observe features from segmentation guidance produce less high-frequency texture than Brain-ID, especially within the same-label regions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.794, + 0.842 + ], + "angle": 0, + "content": "Data Generation Design As shown in Tab. 3, training with all mildly-corrupted samples results in reduced robustness and harms the downstream performance, yet using all extremely corrupted samples leads to unstable training and model collapse. Brain-ID uses samples of gradually increased corruptions (Fig. 3 (left))." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "P. Liu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.308, + 0.145, + 0.694, + 0.159 + ], + "angle": 0, + "content": "Table 3: Comparison between Brain-ID and its variants." + }, + { + "type": "table", + "bbox": [ + 0.29, + 0.17, + 0.714, + 0.332 + ], + "angle": 0, + "content": "
Model Setup [Comparison Target]Feature Robustness (Intra)Downstream (Reconstruction)
SSIM (↑)MS-SSIM (↑)L1 (↓)PSNR (↑)SSIM (↑)
[Representation guidance: supervision target for feature learning]
Segmentation0.8910.9630.02928.130.958
Segmentation + MP-RAGE0.8630.9400.02329.050.964
MP-RAGE (Brain-ID)0.8580.9210.02129.890.966
[Data generation design: corruption levels in intra-subject mini-batches]
All Mild (\\( \\sigma_{noise} = 1, \\cdots \\))0.7920.8130.02828.300.960
All Medium (\\( \\sigma_{noise} = 5, \\cdots \\))0.8310.8990.02229.670.964
All Severe (\\( \\sigma_{noise} = 10, \\cdots \\))N/AN/AN/AN/AN/A
Mild to Severe (Brain-ID)0.8580.9210.02129.890.966
[Mini-batch size: number of intra-subject samples]
20.8260.8970.02528.830.959
30.8410.9020.02229.300.962
4 (Brain-ID)0.8580.9210.02129.890.966
50.8600.9290.02129.930.965
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.404 + ], + "angle": 0, + "content": "Batch Size Brain-ID computes training loss of all intra-subject samples in a mini-batch at once (Eq. (3)). We observe that larger batches help improve feature robustness (Tab. 3), yet do not further enhance the downstream performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.429, + 0.427, + 0.443 + ], + "angle": 0, + "content": "4.6 Further Discussions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.787, + 0.534 + ], + "angle": 0, + "content": "Limitations Although Brain-ID features excel in multiple tasks, they may not be optimal for tasks not involving voxel-wise prediction, e.g., functional brain mapping. Further, as Brain-ID generates data exclusively from anatomies, it tends to struggle with images with extensive pathology. Our future research will focus on robust representations that are encoded with pathology information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.787, + 0.629 + ], + "angle": 0, + "content": "Data Privacy and Ethics Since Brain-ID features can potentially be used to reconstruct a high-resolution MRI scan, publicly sharing such features should be regarded as sharing the original scans from a subject [33]. Therefore, one should follow the same anonymization protocols as when sharing MRI data for research [61] – not only the removal of identifiers, but also defacing [6] or skull-stripping [26] to prevent the 3D reconstruction of faces." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.654, + 0.36, + 0.67 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.787, + 0.841 + ], + "angle": 0, + "content": "We introduced Brain-ID, a contrast-agnostic feature representation learning approach for brain imaging, which is distinctive to brain anatomy and resilient to variations in image appearances. Brain-ID is trained on synthetic inputs, and quickly adapts to downstream tasks through one layer. We validated Brain-ID features' robustness, and effectiveness on four downstream applications (reconstruction, segmentation, super-resolution, bias field estimation), and six public datasets, covering MR and CT. Brain-ID achieves state-of-the-art performances across all tasks and modalities, it further preserves high performance on low-resolution and small-size datasets. We believe Brain-ID unlocks the great potential of robust foundation models, especially for non-calibrated modalities." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.666, + 0.116, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Brain-ID" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.255 + ], + "angle": 0, + "content": "Primarily supported by NIH under grant 1RF1AG080371. Additional support from NIH 1RF1AG080371, 1R01EB031114, 1R01AG070988, 1RF1MH123195. OP is supported by Lundbeckfonden R360-2021-395. Wellcome Trust award 221915/Z/20/Z and the NIHR UCLH Biomedical Research Centre support DCA on this topic." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.279, + 0.323, + 0.294 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.785, + 0.351 + ], + "angle": 0, + "content": "1. Adnan, M., Kalra, S., Cresswell, J.C., Taylor, G.W., Tizhoosh, H.R.: Federated learning and differential privacy for medical image analysis. Scientific reports (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.354, + 0.785, + 0.393 + ], + "angle": 0, + "content": "2. Awais, M., Naseer, M., Khan, S.S., Anwer, R.M., Cholakkal, H., Shah, M., et al.: Foundational models defining a new era in vision: A survey and outlook. arXiv abs/2307.13721 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.396, + 0.785, + 0.435 + ], + "angle": 0, + "content": "3. Balakrishnan, G., Zhao, A., Sabuncu, M.R., Guttag, J.V., Dalca, A.V.: VoxelMorph: A learning framework for deformable medical image registration. IEEE Transactions on Medical Imaging (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.437, + 0.785, + 0.476 + ], + "angle": 0, + "content": "4. Billot, B., Greve, D.N., Puonti, O., Thielscher, A., Leemput, K.V., Fischl, B., et al.: SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining. Medical Image Analysis (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.479, + 0.785, + 0.518 + ], + "angle": 0, + "content": "5. Billot, B., Magdamo, C., Cheng, Y., Arnold, S.E., Das, S., Iglesias, J.E.: Robust machine learning segmentation for large-scale analysis of heterogeneous clinical brain MRI datasets. Proceedings of the National Academy of Sciences (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.521, + 0.785, + 0.56 + ], + "angle": 0, + "content": "6. Bischoff-Grethe, A., Ozyurt, I.B., Busa, E., Quinn, B.T., Fennema-Notestine, C., Clark, C.P., et al.: A technique for the deidentification of structural brain MR images. Human Brain Mapping (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.563, + 0.785, + 0.602 + ], + "angle": 0, + "content": "7. Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., et al.: On the opportunities and risks of foundation models. arXiv abs/2108.07258 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.604, + 0.785, + 0.643 + ], + "angle": 0, + "content": "8. Brant-Zawadzki, M.N., Gillan, G.D., Nitz, W.R.: MP RAGE: A three-dimensional, T1-weighted, gradient-echo sequence-initial experience in the brain. Radiology (1992)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.646, + 0.785, + 0.699 + ], + "angle": 0, + "content": "9. Brown, M.R.G., Sidhu, G.P.S., Greiner, R., Asgarian, N., Bastani, M., Silverstone, P.H., et al.: ADHD-200 global competition: diagnosing ADHD using personal characteristic data can outperform resting state fMRI measurements. Frontiers in Systems Neuroscience (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.702, + 0.785, + 0.728 + ], + "angle": 0, + "content": "10. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., et al.: Language models are few-shot learners. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.73, + 0.785, + 0.769 + ], + "angle": 0, + "content": "11. Cardoso, M.J., Li, W., Brown, R., Ma, N., Kerfoot, E., Wang, Y., et al.: MONAI: An open-source framework for deep learning in healthcare. arXiv abs/2211.02701 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.771, + 0.785, + 0.797 + ], + "angle": 0, + "content": "12. Caron, M., Touvron, H., Misra, I., J'egou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.785, + 0.839 + ], + "angle": 0, + "content": "13. Cerri, S., Puonti, O., Meier, D.S., Wuerfel, J., Mühlau, M., Siebner, H.R., Leemput, K.V.: A contrast-adaptive method for simultaneous whole-brain and lesion segmentation in multiple sclerosis. NeuroImage (2020)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.311, + 0.785, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "P. Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "14. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., et al.: PaLM: Scaling language modeling with pathways. JMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "15. Chua, Y.Z.R., Dalca, A.V.: Contrast invariant feature representations for segmentation and registration of medical images. In: MIDL (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.245 + ], + "angle": 0, + "content": "16. Chua, Z.Y., Zheng, W., Chee, M.W.L., Zagorodnov, V.: Evaluation of performance metrics for bias field correction in MR brain images. Journal of Magnetic Resonance Imaging (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "17. Darzidehkalani, E., Ghasemi-Rad, M., van Ooijen, P.: Federated learning in medical imaging: Part I: Toward multicentral health care ecosystems. Journal of the American College of Radiology (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "18. Ding, Z., Han, X., Liu, P., Niethammer, M.: Local temperature scaling for probability calibration. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "19. Essen, D.C.V., Ugurbil, K., Auerbach, E.J., Barch, D.M., Behrens, T.E.J., Bucholz, R.D., et al.: The human connectome project: A data acquisition perspective. NeuroImage (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.396 + ], + "angle": 0, + "content": "20. Fischl, B., Salat, D.H., Busa, E., Albert, M.S., Dieterich, M., Haselgrove, C., et al.: Whole brain segmentation automated labeling of neuroanatomical structures in the human brain. Neuron (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "21. Fonov, V.S., Evans, A.C., Botteron, K., Almli, C.R., McKinstry, R.C., Collins, D.L., et al.: Unbiased average age-appropriate atlases for pediatric studies. Neuroimage (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "22. Fonov, V.S., Evans, A.C., McKinstry, R.C., Almli, C.R., Collins, D.: Unbiased nonlinear average age-appropriate brain templates from birth to adulthood. NeuroImage (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.549 + ], + "angle": 0, + "content": "23. Fowler, C., Rainey-Smith, S.R., Bird, S., Bomke, J., Bourgeat, P., Brown, B.M., et al.: Fifteen years of the australian imaging, biomarkers and lifestyle (AIBL) study: Progress and observations from 2,359 older adults spanning the spectrum from cognitive normality to Alzheimer's disease. Journal of Alzheimer's disease reports (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "24. Henschel, L., Kugler, D., Reuter, M.: FastSurferVINN: Building resolution-independence into deep learning segmentation methods — A solution for highres brain MRI. NeuroImage (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.632 + ], + "angle": 0, + "content": "25. Hoffmann, M., Billot, B., Greve, D.N., Iglesias, J.E., Fischl, B., Dalca, A.V.: SynthMorph: Learning contrast-invariant registration without acquired images. IEEE Transactions on Medical Imaging (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.66 + ], + "angle": 0, + "content": "26. Hoopes, A., Mora, J.S., Dalca, A.V., Fischl, B., Hoffmann, M.: SynthStrip: skullstripping for any brain image. NeuroImage (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "27. Iglesias, J.E., Billot, B., Balbastre, Y., Magdamo, C.G., Arnold, S., Das, S., et al.: SynthSR: A public AI tool to turn heterogeneous clinical brain scans into high-resolution T1-weighted images for 3D morphometry. Science Advances (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.757 + ], + "angle": 0, + "content": "28. Iglesias, J.E., Billot, B., Balbastre, Y., Tabari, A., Conklin, J., Alexander, D.C., et al.: Joint super-resolution and synthesis of \\(1\\mathrm{mm}\\) isotropic MP-RAGE volumes from clinical MRI exams with scans of different orientation, resolution and contrast. NeuroImage (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "29. Iglesias, J.E., Schleicher, R., Laguna, S., Billot, B., Schaefer, P., McKaig, B., et al.: Quantitative brain morphometry of portable low-field-strength MRI using superresolution machine learning. Radiology (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "30. Jack, C.R., Bernstein, M.A., Fox, N.C., Thompson, P.M., Alexander, G.E., Harvey, D.J., et al.: The Alzheimer's disease neuroimaging initiative (ADNI): MRI methods. Journal of Magnetic Resonance Imaging (2008)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.666, + 0.116, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Brain-ID" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "31. Juntu, J., Sijbers, J., Dyck, D.V., Gielen, J.: Bias field correction for MRI images. In: CORES (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.787, + 0.217 + ], + "angle": 0, + "content": "32. Kamnitsas, K., Ledig, C., Newcombe, V.F.J., Simpson, J.P., Kane, A.D., Menon, D.K., et al.: Efficient multi-scale 3D CNN with fully connected CRF for accurate brain lesion segmentation. Medical Image Analysis (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.786, + 0.242 + ], + "angle": 0, + "content": "33. Kapkiç, M., Sağiroğlu, Ş.: Privacy issues in magnetic resonance images. International Journal of Information Security Science (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.786, + 0.27 + ], + "angle": 0, + "content": "34. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., et al.: Segment anything. arXiv abs/2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.27, + 0.786, + 0.311 + ], + "angle": 0, + "content": "35. LaMontagne, P.J., Keefe, S.J., Lauren, W., Xiong, C., Grant, E.A., Moulder, K.L., et al.: OASIS-3: Longitudinal neuroimaging, clinical, and cognitive dataset for normal aging and Alzheimer's disease. Alzheimer's & Dementia (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.786, + 0.352 + ], + "angle": 0, + "content": "36. Laso, P., Cerri, S., Sorby-Adams, A., Guo, J., Mateen, F., Goebl, P., et al.: Quantifying white matter hyperintensity and brain volumes in heterogeneous clinical and low-field portable MRI. arXiv abs/2312.05119 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.786, + 0.393 + ], + "angle": 0, + "content": "37. Leemput, K.V., Maes, F., Vandermeulen, D., Suetens, P.: A unifying framework for partial volume segmentation of brain MR images. IEEE Transactions on Medical Imaging (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.786, + 0.434 + ], + "angle": 0, + "content": "38. Li, Z., Tian, L., Mok, T.C.W., Bai, X., Wang, P., Ge, J., et al.: SAMConvex: Fast discrete optimization for CT registration using self-supervised anatomical embedding and correlation pyramid. In: MICCAI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.786, + 0.46 + ], + "angle": 0, + "content": "39. Liu, P., Lee, Y., Aylward, S., Niethammer, M.: Deep decomposition for stochastic normal-abnormal transport. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.786, + 0.487 + ], + "angle": 0, + "content": "40. Liu, P., Tian, L., Zhang, Y., Aylward, S., Lee, Y., Niethammer, M.: Discovering hidden physics behind transport dynamics. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.786, + 0.528 + ], + "angle": 0, + "content": "41. Liu, P., Wang, R., Cao, X., Zhou, Y., Shah, A., Lim, S.N.: Differential motion evolution for fine-grained motion deformation in unsupervised image animation. arXiv abs/2110.04658 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.786, + 0.555 + ], + "angle": 0, + "content": "42. Miletari, F., Navab, N., Ahmadi, S.A.: V-Net: Fully convolutional neural networks for volumetric medical image segmentation. 3DV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.786, + 0.596 + ], + "angle": 0, + "content": "43. Modat, M., Ridgway, G.R., Taylor, Z.A., Lehmann, M., Barnes, J., Hawkes, D.J., et al.: Fast free-form deformation using graphics processing units. Computer Methods and Programs in Biomedicine (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.786, + 0.636 + ], + "angle": 0, + "content": "44. Moor, M., Banerjee, O., Abad, Z.F.H., Krumholz, H.M., Leskovec, J., Topol, E.J., Rajpurkar, P.: Foundation models for generalist medical artificial intelligence. Nature (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.636, + 0.786, + 0.678 + ], + "angle": 0, + "content": "45. Müller, R.A., Shih, P., Keehn, B., Deyoe, J.R., Leyden, K.M., Shukla, D.: Underconnected, but how? A survey of functional connectivity MRI studies in autism spectrum disorders. Cerebral Cortex (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.786, + 0.704 + ], + "angle": 0, + "content": "46. Puonti, O., Iglesias, J.E., Leemput, K.V.: Fast and sequence-adaptive whole-brain segmentation using parametric bayesian modeling. NeuroImage (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.786, + 0.731 + ], + "angle": 0, + "content": "47. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.786, + 0.772 + ], + "angle": 0, + "content": "48. Sheller, M.J., Edwards, B., Reina, G.A., Martin, J., Pati, S., Kotrotsou, A., et al.: Federated learning in medicine: Facilitating multi-institutional collaborations without sharing patient data. Scientific reports (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.786, + 0.813 + ], + "angle": 0, + "content": "49. Shen, Z., Feydy, J., Liu, P., Curiale, A.H., San Jose Estepar, R., San Jose Estepar, R., Niethammer, M.: Accurate point cloud registration with robust optimal transport. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "50. Singhal, K., Azizi, S., Tu, T., Mahdavi, S., Wei, J., Chung, H.W., et al.: Large language models encode clinical knowledge. Nature (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.128 + ], + "angle": 0, + "content": "P. Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "51. Symms, M.R., Jäger, H.R., Schmierer, K., Yousry, T.A.: A review of structural magnetic resonance neuroimaging. Journal of Neurology, Neurosurgery & Psychiatry (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "52. Tanno, R., Worrall, D.E., Kaden, E., Alexander, D.C.: Uncertainty modelling in deep learning for safer neuroimage enhancement: Demonstration in diffusion MRI. NeuroImage (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "53. Tian, Q., Bilgiç, B., Fan, Q., Ngamsombat, C., Zaretskaya, N., Fultz, N.E., et al.: Improving in vivo human cerebral cortical surface reconstruction using data-driven super-resolution. Cerebral Cortex (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "54. Tu, T., Azizi, S., Driess, D., Schaekermann, M., Amin, M., Chang, P.C., et al.: Towards generalist biomedical AI. arXiv abs/2307.14334 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "55. de Vos, B.D., Berendsen, F.F., Viergever, M.A., Sokooti, H., Staring, M., Isgum, I.: A deep learning framework for unsupervised affine and deformable image registration. Medical Image Analysis (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.37 + ], + "angle": 0, + "content": "56. Wachinger, C., Golland, P., Reuter, M.: BrainPrint: Identifying subjects by their brain. In: MICCAI (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "57. Wang, M., Deng, W.: Deep visual domain adaptation: A survey. Neurocomputing (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "58. Wang, T.C., Mallya, A., Liu, M.Y.: One-shot free-view neural talking-head synthesis for video conferencing. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "59. Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for image quality assessment. In: ACSSC (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.493 + ], + "angle": 0, + "content": "60. Weiner, M.W., Veitch, D.P., Aisen, P.S., Beckett, L.A., Cairns, N.J., Green, R.C., et al.: The Alzheimer's disease neuroimaging initiative 3: Continued innovation for clinical trial improvement. Alzheimer's & Dementia (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.536 + ], + "angle": 0, + "content": "61. White, T., Blok, E., Calhoun, V.D.: Data sharing and privacy issues in neuroimaging research: Opportunities, obstacles, challenges, and monsters under the bed. Human Brain Mapping (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "62. Xu, Y., Mo, T., Feng, Q., Zhong, P., Lai, M., Chang, E.I.C.: Deep learning of feature representation with multiple instance learning for medical image analysis. In: ICASSP (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.619 + ], + "angle": 0, + "content": "63. Yan, K., Cai, J., Jin, D., Miao, S., Harrison, A.P., Guo, D., et al.: SAM: Self-supervised learning of pixel-wise anatomical embeddings in radiological images. IEEE Transactions on Medical Imaging (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.646 + ], + "angle": 0, + "content": "64. Yang, X., Kwitt, R., Niethammer, M.: Quicksilver: Fast predictive image registration - a deep learning approach. NeuroImage (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "65. You, C., Zhao, R., Staib, L.H., Duncan, J.S.: Momentum contrastive voxel-wise representation learning for semi-supervised volumetric medical image segmentation. In: MICCAI (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.73 + ], + "angle": 0, + "content": "66. Zhou, Y., Chia, M.A., Wagner, S.K., Ayhan, M.S., Williamson, D.J., Struyven, R.R., et al.: A foundation model for generalizable disease detection from retinal images. Nature (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.73 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Brain-ID_ Learning Contrast-agnostic Anatomical Representations for Brain Imaging/b53a82d6-784c-4409-ad49-9bb017717031_origin.pdf b/2024/Brain-ID_ Learning Contrast-agnostic Anatomical Representations for Brain Imaging/b53a82d6-784c-4409-ad49-9bb017717031_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..796536fc8ddde4c1cb75d0fbc03c73a3475fbef3 --- /dev/null +++ b/2024/Brain-ID_ Learning Contrast-agnostic Anatomical Representations for Brain Imaging/b53a82d6-784c-4409-ad49-9bb017717031_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c8d47f794e8897c873a8dd963d0fa305d4cac1a17200274ca01b1c6dce8e439 +size 3479178 diff --git a/2024/Brain-ID_ Learning Contrast-agnostic Anatomical Representations for Brain Imaging/full.md b/2024/Brain-ID_ Learning Contrast-agnostic Anatomical Representations for Brain Imaging/full.md new file mode 100644 index 0000000000000000000000000000000000000000..725127a1a8c591f737594115bd4cbfb7cb41387a --- /dev/null +++ b/2024/Brain-ID_ Learning Contrast-agnostic Anatomical Representations for Brain Imaging/full.md @@ -0,0 +1,304 @@ +# Brain-ID: Learning Contrast-agnostic Anatomical Representations for Brain Imaging + +Peirong Liu1 Oula Puonti1 Xiaoling Hu1 + +Daniel C. Alexander $^{2}$ +Juan E. Iglesias $^{1,2,3}$ + +$^{1}$ Harvard Medical School and Massachusetts General Hospital $^{2}$ University College London $^{3}$ Massachusetts Institute of Technology + +![](images/24561250013a44bfa91f12541944ec764831343688bad5919067e368166efbd7.jpg) +Fig. 1: Brain-ID features serve as the distinctive identity of each subject (ID) regardless of appearance (contrast, deformation, corruptions, etc) and quickly adapt to downstream tasks – either contrast-independent (anatomy Reconstruction, Segmentation), or contrast-dependent (Super-Resolution, Bias Field estimation), all through one layer. + +Abstract. Recent learning-based approaches have made astonishing advances in calibrated medical imaging like computerized tomography (CT). Yet, they struggle to generalize in uncalibrated modalities – notably magnetic resonance (MR) imaging, where performance is highly sensitive to the differences in MR contrast, resolution, and orientation. This prevents broad applicability to diverse real-world clinical protocols. We introduce Brain-ID, an anatomical representation learning model for brain imaging. With the proposed "mild-to-severe" intra-subject generation, Brain-ID is robust to the subject-specific brain anatomy regardless of the appearance of acquired images. Trained entirely on synthetic inputs, Brain-ID readily adapts to various downstream tasks through one layer. We present new metrics to validate the intra/inter-subject robustness of Brain-ID features, and evaluate their performance on four downstream applications, covering contrast-independent (anatomy reconstruction, brain segmentation), and contrast-dependent (super-resolution, bias field estimation) tasks (Fig. 1). Extensive experiments on six public datasets demonstrate that Brain-ID achieves state-of-the-art performance in all tasks on different MR contrasts and CT, and more importantly, preserves its performance on low-resolution and small datasets. Code is available at https://github.com/peirong26/Brain-ID. + +# 1 Introduction + +Magnetic resonance imaging (MRI) enables in vivo noninvasive imaging of the human brain with exquisite and tunable soft-tissue contrast [8]. Recent machine learning methods have achieved great improvements in faster and more accurate MRI analysis [27], such as segmentation [18,32,42,47], registration [3,49,55,64], super-resolution [52, 53], and connectivity studies [45]. However, most existing MRI analysis methods are specific to certain MR contrast(s) and require near-isotropic acquisitions. Therefore, models face sharp performance drops when voxel size and anisotropy increase, or are being used for a different contrast than training [57]. This reduces model generalizability and results in duplicate training efforts. Resorting to synthetic data, recent contrast-agnostic models [4, 25, 27, 28, 36, 39, 40] achieve impressive results and largely extend model applicability. However, these models are only applicable to the tasks they were trained for. + +Meanwhile, task-agnostic foundation models in computer vision and natural language processing have witnessed remarkable success, along with the fast developments of large-scale datasets [2,7,10,14,34]. However, due to different acquisition/processing protocols and privacy requirements across institutions, large-scale medical imaging datasets require significantly more effort than natural imaging/language. Thus, medical foundation models are not as well developed. MONAI [11] includes a pre-trained model zoo, yet all are highly task-oriented and sensitive to contrasts. Zhou et al. [66] constructed a medical foundation model, designed for eye and systemic health condition detections from retinal scans, and only works on the modalities of color fundus photography and optical coherence tomography. Lately, generalist biomedical AI systems [44, 50, 54] have shown great potential in biomedical tasks under vision-language context, e.g., (visual) question answering, image classification, radiology report generation and summarization. However, they have not explored the challenging vision tasks such as reconstruction, segmentation, super-resolution, and registration. + +We introduce Brain-ID, a contrast-agnostic feature representation approach trained on synthetic inputs, that can adapt to various tasks through one layer. + +1) We introduce an on-the-fly, intra-subject data generator capable of synthesizing any contrast, with a mild-to-severe corruption strategy (Fig. 2, Sec. 3.2). Unlike real-world datasets constrained by the images acquired per subject, Brain-ID learns in a much more expansive and diverse space. +2) We design a feature representation learning framework guided by the unique anatomy of each subject. We validate that Brain-ID's high-resolution features are consistently robust to the superficial perturbations of an image's appearance (Sec. 4.2). Straightforward one-layer adaptions are further presented to adapt Brain-ID features to downstream tasks (Sec. 3.3). +3) We evaluate Brain-ID on both contrast-independent (anatomy reconstruction, brain segmentation) and contrast-dependent (super-resolution, bias field estimation) tasks, across six public datasets ( $\approx$ 8,000 images) including MR (T1w, T2w, FLAIR) and CT. Brain-ID achieves state-of-the-art performance on all tasks (Tab. 2), and further maintains its high performance on low-resolution images (Fig. 6) and limited-size datasets (Fig. 7). + +# 2 Related work + +Feature Representation in Medical Imaging As mentioned in Sec. 1, general feature representation learning in medical imaging can be more challenging than in the natural image domain, due to limited data availability. Xu et al. [62] introduced a multiple-instance learning model for feature representation, but it only applies to histopathology image classification. You et al. [65] presented CVRL, a semi-supervised representations approach designed for image segmentation, but it requires CT to extract anatomical information. SAM [63] encodes anatomical information in feature embeddings, which has shown to be effective for registration [38]. However, same as CVRL, SAM only works on CT. BrainPrint [56] is a compact and discriminative representation of brain morphology, which is specifically designed for cortical surface analyses. To our best knowledge, CIFL [15] is the only existing work on learning contrast-agnostic and task-independent brain feature representations. CIFL relies on contrastive learning alone, and is insufficient to outperform task/contrast-specific supervised models in downstream applications as shown by our experiments (Tab. 2). + +Contrast-invariant Learning for MR Images MRI scans acquired across sites vary substantially in contrasts, resolutions, orientations, etc. When given a new dataset, heterogeneity leads to duplicate training efforts for approaches that are sensitive to specific combinations of MR contrast, resolution, or orientation. Classical brain segmentation models used Bayesian inference for contrast-robustness [20,37], which requires a long processing time and struggles with low or anisotropic resolutions [27, 46]. Recently, there have been works using synthetic data to achieve contrast-invariance in tasks like image segmentation [4,36], registration [25], super-resolution [28], and skull-stripping [26]. However, all the above-mentioned methods are trained in a task-specific manner, their features therefore cannot be readily applied to other domains. + +# 3 Brain-ID: Learning Anatomy-specific Brain Features + +As discussed in Sec. 1, the main challenges to obtain a general and robust feature representation for MR images lie in $(i)$ the practical restrictions of building large-scale datasets with diverse contrasts; and $(ii)$ the nature of most medical imaging models that are task-oriented and specific to data type (contrast, resolution, orientation, etc). We aim to learn a brain feature representation that is: + +(i) Robust: features should be robust to each subject's distinct anatomy, unaffected by variations in poses/deformations, contrasts, resolutions, or artifacts. +(ii) Expressive: features should also exhibit high expressiveness, containing rich information that facilitates easy and effective adaptation to diverse downstream tasks, eliminating the necessity for extensive training data. + +We first introduce Brain-ID's data generator (Sec. 3.1) and training framework (Sec. 3.2) to achieve the above two aims. Then, we present our one-layer solutions for adapting Brain-ID features to downstream tasks that could be either dependent or independent of the contrast of input images (Sec. 3.3). + +![](images/68490469bf1f48f04ec51c79a8e01d79643addf15b0050a8c34fe4585f03a669.jpg) +Fig.2: Brain-ID's data generator on the fly. Given the brain segmentation labels of a subject, we randomly generate a deformation field, and synthesize intra-subject samples featuring various contrast intensities and corruption levels (Sec. 3.1). + +# 3.1 Enrich the Intra-subject Learning Space + +A good representation relies on large-scale data, however, different acquisition protocols, processing pipelines, and privacy requirements across institutions make large-scale data less accessible and require significantly more effort [1, 17, 48]. Moreover, for each subject, there are usually only limited acquired scans available. This lack of data consistency is a significant barrier to obtaining a robust, contrast-invariant feature representation. Brain-ID avoids these barriers through the use of synthetic data. + +To generate images with complex brain structures, we start with brain segmentation labels of anatomical structures ( $L$ in Fig. 2). The generation consists of three steps, $(i)$ deformation generation, $(ii)$ contrast synthesis, and $(iii)$ data corruption (including lower resolution resampling). For simplicity, $\Theta$ denotes the parameter group of the generation process described below. + +Deformation Generation We first generate a random deformation field $(\phi|_{\theta_{\phi}})$ consisting of an affine transformation and a non-linear displacement field: + +$$ +\phi | _ {\theta_ {\phi}} = \mathcal {T} | _ {\theta_ {\phi}} \circ \mathcal {A} | _ {\theta_ {\phi}}, \tag {1} +$$ + +where $\mathcal{A}|_{\theta_{\phi}}$ denotes an affine transformation matrix which includes linear (rotation, scaling, shearing) transformation and translation, $\mathcal{T}|_{\theta_{\phi}}$ refers to a nonlinear displacement field computed as the integration of a stationary velocity field (SVF) that is smooth and invertible everywhere and thus preserves the topology of the brain anatomy [28]. $\theta_{\phi} \in \Theta$ controls the transformation ranges. + +MR Contrast Synthesis Then, we synthesize images $(I(x), x \in \Omega)$ by randomly "painting" intensities on the segmentation maps according to their brain structure labels $(l \in L)$ . Specifically, the regional intensities are generated by separately sampling a Gaussian distribution on each labeled region: + +$$ +\left\{ \begin{array}{l} I (x) \sim \mathcal {N} \left(\mu_ {l}, \sigma_ {l}\right), \quad l \in L, \\ \mu_ {l} \sim \mathcal {N} \left(0, 1 \mid \theta_ {\mu}, \theta_ {l}\right), \sigma_ {l} \sim \mathcal {N} \left(0, 1 \mid \theta_ {\sigma}, \theta_ {l}\right), \end{array} \right. \tag {2} +$$ + +where $\mu_{l}$ and $\sigma_{l}$ refer to the mean and standard deviation of each segmentation label $l$ , and are independently sampled from Gaussian distributions at each voxel, with $\theta_{l}, \theta_{\mu}, \theta_{\sigma} \in \Theta$ controlling the shifts and scales of their values. + +![](images/2b3785bbaec27ab54c9bbb8b9bf80ed99204e9da07ebd41278ea716e6d1c6fe1.jpg) +Fig. 3: Brain-ID's contrast-agnostic anatomical representation learning framework. + +Resolution Simulation and Data Corruption Given a deformed, contrast-synthesized image $(I)$ , we adopt the standard data corruption pipeline [27], which further augments images with different levels of resolution and scanning artifacts that are commonly found in real-world clinical protocols. + +As illustrated in Fig. 2, we are able to generate an infinite number of variations from a single subject with its unique brain anatomy. Generating images with randomized contrast/resolution/orientation on the fly for each subject enormously enriches the learning space for a robust representation, and helps focus the learning on intrinsic subject-specific features rather than superficial aspects of images that depend on acquisition parameters and conditions. + +# 3.2 Extract Robust and Expressive Subject-specific Features + +As mentioned above, we would like Brain-ID features to be robust to intra-subject variations and expressive to potential downstream tasks. In this section, we introduce Brain-ID's learning framework to achieve the two goals. + +Intra-subject Data Generation In order to learn a feature representation that is distinctive to each subject and robust to varying MR contrasts, we observe that enriching intra-subject samples leads to better performance (Tab. 3). Specifically, instead of including multiple subjects for a mini-batch during each training iteration as in usual practice, Brain-ID focuses on maximizing the intra-subject variance to improve the subject-robustness of resulting features. As described in Alg. 1 (line 4-10), for each training iteration, after randomly selecting a subject and generating its deformation, Brain-ID generates a mini-batch of intra-subject samples $\{(S_1,\ldots ,S_N)\}$ with randomly synthesized contrasts, resolutions and corruptions (Sec. 3.1). As will be introduced below, Brain-ID collects losses from all intra-subject samples and conducts back-propagation at once, to encourage the subject-specific robustness of its learned features. + +We set the intra-subject samples within a mini-batch to have random contrasts and "mild-to-severe", increasing level of corruptions (Fig. 3 (left)), to maximize the intra-subject variance while ensuring the stability of the training process against extreme corruption levels. (In Sec. 4.5, we compare various data generation designs, and provide insights on preventing unstable training.) + +Anatomy-guided Feature Representation A richer learning space helps improve the representation robustness against the variance of sample appearances, but not enough for extracting expressive features. For example, a mapping that projects all inputs to zero is perfectly robust, but does not provide useful information for potential downstream tasks. Therefore, proper guidance is crucial for feature representation as well. In CIFL [15], the authors use a contrastive loss on feature channels to encourage a more discriminative feature representation. However, this method is insufficient, given the extremely complex nature of the human brain anatomy (See comparisons in Sec. 4). Instead, we propose to use the standard T1w high-resolution structural MR contrast for brain morphometry, i.e., MP-RAGE (magnetization-prepared rapid gradient-echo), as the unique anatomy target to guide Brain-ID's feature representation learning. + +As shown in Fig. 3, a feature extraction backbone $(\mathcal{F})$ first maps the input mini-batch of intra-subject generated samples, $\{S_1,\dots,S_N\}$ , to their corresponding feature space, $\{\mathbf{F}_1,\ldots ,\mathbf{F}_N\}$ . A linear activation layer $(\mathcal{L})$ then projects the features to the current subject's standard contrast (MP-RAGE) space, $T$ . The training loss is obtained by summing over the anatomy reconstruction loss of all intra-subject samples in the current mini-batch: + +$$ +L = \sum_ {i} ^ {N} | \mathcal {L} (\mathbf {F} _ {i}) - T | + \lambda | \nabla \mathcal {L} (\mathbf {F} _ {i}) - \nabla T |, \lambda \in R ^ {+}, \tag {3} +$$ + +where $T$ , as the high-resolution, unique anatomy target for all intra-subject samples, encourages the similarities of the features extracted from the same subject via reconstruction ( $1^{\mathrm{st}}$ term) and gradient difference ( $2^{\mathrm{nd}}$ term) L1 loss. Brain-ID uses MP-RAGE as the learning target, which formulates both super-resolution and contrast-synthesis problems, and encourages a richer feature representation. (Sec. 4.5 provides further insights on how different choices of anatomy guidance would affect the resulting features' properties.) + +# 3.3 Adapting Brain-ID to Downstream Tasks by One Layer + +With the intra-subject data generation and anatomy-guided feature learning design introduced in Sec. 3.2, a well-trained Brain-ID model is able to extract robust, high-resolution anatomical features from images with varying deformations, contrasts, resolutions, and artifacts. To minimize the modifications, we propose straightforward one-layer solutions adapting Brain-ID features to various brain imaging applications. We later demonstrate that the simple adaptions are effective enough for Brain-ID to achieve state-of-the-art performance across all downstream tasks (Sec. 4.3), even for small datasets (Sec. 4.4). + +Contrast-independent Tasks For tasks where the output should be independent of the input MR contrast, e.g., brain segmentation, we simply add an additional layer following Brain-ID features $(\mathbf{F})$ , and fine-tune the model: + +$$ +L = \operatorname {t a s k} _ {-} \operatorname {l o s s} _ {-} \operatorname {f u n c} \left(\mathcal {L} (\mathbf {F}), T\right), \tag {4} +$$ + +where $\mathcal{L}$ and $T$ refer to the task-specific activation layer and the contrast-independent ground truth, respectively. + +Contrast-dependent Tasks Since Brain-ID features are contrast-agnostic and robust to artifacts, for tasks relevant to the input's contrast/quality, e.g., super-resolution, we concatenate the input image $(I)$ that contains the original contrast information with its high-resolution Brain-ID features along the channel dimension, before forwarding into the task-specific activation layer: + +$$ +L = \operatorname {t a s k} _ {-} \operatorname {l o s s} _ {-} \operatorname {f u n c} \left(\mathcal {L} (\mathbf {F} \oplus I), T\right), \tag {5} +$$ + +where $\mathcal{L}$ and $T$ refer to the task-specific activation layer and the ground truth. + +# 4 Experiments + +We demonstrate the two properties of Brain-ID as claimed in Sec. 3. (i) Robustness (Sec. 4.2): we propose "canonical" and "atlas-registered" features to assess the intra/inter-subject feature robustness; (ii) Expressiveness (Sec. 4.3): we evaluate the performance of adapting Brain-ID features to various brain imaging applications. We further challenge Brain-ID with reduced-size datasets to explore its ability when limited real data is available for training (Sec. 4.4). + +# 4.1 Datasets, Metrics and Implementation Details + +Datasets For Brain-ID features pre-training, we use 2045 3D segmentations ("anatomies", or IDs) from the public ADNI dataset [30]. For downstream evaluation, we use six public datasets covering T1w and T2w MRI, FLAIR MRI, and CT: ADNI [30], ADNI3 [60], HCP [19], ADHD200 [9], AIBL [23], OASIS3 [35]. + +Metrics We evaluate individual tasks from different aspects. For feature similarity measurements, we use L1 distance, and (MS-)SSIM (multi-scale structural similarity) [41,58,59]. For reconstruction and super-resolution, we use L1, PSNR (peak signal-to-noise ratio) and (MS-)SSIM. For segmentation, we use Dice scores. For bias field estimation, we use the normalized L2 distance (norm-L2) to avoid possible arbitrary scalings from nonuniformity correction [16]. + +Implementation Details As a general feature representation model, Brain-ID can use any backbone to extract brain features. For fairer comparison, we adopt the same five-level 3D UNet [47] (with 64 feature channels in the last layer) as utilized in state-of-the-art models we compare with in Sec. 4.3. During feature pre-training (Sec. 3.2), a linear regression layer is added for anatomy supervision. For downstream adaptions, A task-specific adaption layer is added (Sec. 3.3). + +# 4.2 Intra/Inter-subject Feature Robustness + +In this section, we examine the robustness of Brain-ID features, i.e., the 64 features from the last layer. For comparable and reproducible results, we use our data generator (Sec. 3.1) to prepare (deform and corrupt) 1000 testing samples from 100 randomly selected subjects in ADNI [30] (T1w) testing set, with 10 intra-subject samples for each subject. + +![](images/437282f39cfc164e1f6371617cc3902ae0ce5b40c0baadf767663eab7eabf17b.jpg) +Fig. 4: (a) Intra-subject and (b) inter-subject robustness of Brain-ID features. + +We compare our features with CIFL [15] which is, to our best knowledge, the only similar work to Brain-ID, that also aims to learn contrast-agnostic brain features. Note the original CIFL method has only experimented on 2D images, and does not have our intra-subject data generation design. For a fairer comparison, we trained CIFL with its contrastive-learning design, while using the same data generator and network architecture as Brain-ID. + +Intra-subject Feature Robustness Ideally, Brain-ID features computed from the same subject should be structurally identical, regardless of their appearances (contrast, resolution, noises, etc). To assess the intra-subject feature robustness, we first define their corresponding "canonical" features. + +"Canonical" Features. We compute $\phi^{-1}$ , the inverse of $S$ 's generated deformation as in Eq. (1), and map back each sample's $(S)$ Brain-ID features $(\mathbf{F} = \mathbf{F}(S|\phi))$ to their original domain without any deformation: + +$$ +\mathbf {F} ^ {- 1} = \phi^ {- 1} \circ \mathbf {F} (S \mid \phi). \tag {6} +$$ + +Table 1: Evaluations on intra/inter-subject feature robustness. + +
ModeMethodL1 (↓)SSIM (↑)MS-SSIM (↑)
IntraCIFL [15]0.122 (±0.031)0.511 (±0.298)0.531 (±0.367)
Brain-ID0.011 (±0.001)0.858 (±0.041)0.921 (±0.008)
InterCIFL [15]0.230 (±0.039)0.552 (±0.240)0.524 (±0.291)
Brain-ID0.014 (±0.013)0.843 (±0.044)0.913 (±0.011)
+ +As shown in Fig. 4-(a), for the same subject (ID-i), although the inputs (row) are processed by varying deformations and corruptions, their Brain-ID's "canonical" features are of similar looking along each feature channel (column). In Tab. 1, we quantify the feature similarity: For each testing subject, we first generate a sample without any deformation or corruption and use its Brain-ID feature as the gold-standard reference for all other "canonical" features computed from intra-subject samples that may have different levels of deformation and corruptions. The final results are obtained by scores averaged over all 64 feature channels of all the intra-subject samples. + +Inter-subject Feature Robustness We further assess the feature robustness among anatomies, based on the assumption that features from different subjects should ideally be structurally similar, once being registered to a common domain. "Atlas-registered" Features. We register [43] features $(\mathbf{F} = \mathbf{F}(S|\phi))$ from all subjects to a common brain atlas [21, 22]: + +$$ +\overline {{\mathbf {F}}} = \psi_ {S \rightarrow A} \circ \mathbf {F} (S \mid \phi), \tag {7} +$$ + +where $\psi_{S\to A}$ denotes the mapping from $S$ to the atlas, $A$ + +As shown in Fig. 4-(b), even though the input samples (row) from different subjects have varying anatomical structures, after being registered to a common domain, their "atlas-registered" Brain-ID features still appear consistent (column). In Tab. 1, we further provide quantitative results of the structural similarity measured between features from different subjects. Similar to Sec. 4.2, we randomly select a testing subject and generate a sample without any deformation or corruption. We use its "atlas-registered" Brain-ID feature as the gold standard for all "atlas-registered" features computed from other samples that may have different levels of deformation and corruption. The final results are obtained by scores averaged over all samples of the 100 testing subjects. + +In both intra/inter-subject aspects, Brain-ID outperforms CIFL by a large margin, which validates its feature robustness against deformations, contrasts, resolutions, artifacts (intra-subject), and anatomies (inter-subject). + +# 4.3 Downstream Evaluation + +Following the standard evaluation protocol for representation learning [12], we fine-tune Brain-ID features for downstream tasks through one-layer (Sec. 3.3). Models We compare the downstream performance of Brain-ID features with: $(i - ii)$ SAMSEG [13,46] and FastSurfer [24] (only works on T1w), state-of-the-art + +![](images/7c5afb057c19828fd2a976c65d9e9ba4b880faf378ece7de0b0fdaa45c68eecf.jpg) +Fig. 5: Qualitative comparisons on downstream tasks of reconstruction (Recon), superresolution (SR), and segmentation (Seg), between Brain-ID, the baseline SCRATCH, and the state-of-the-art methods CIFL [15], SynthSR [27] (Recon and SR), SAMSEG [13] (Seg). The visualized testing examples are taken from: AIBL-FLAIR [23] for Recon, AIBLT1w [23] for SR, and OASIS-CT [35] for Seg. The mint circles highlight some details. + +classical and machine-learning-based brain segmentation models, respectively; (iii) SynthSR [27], state-of-the-art, contrast-agnostic T1w synthesis model. We also provide fine-tuned SynthSR (SynthSR-FT) results for further comparison; (iv) SCRAW: a baseline with the same data generation/architecture as Brain-ID, yet trained from scratch. SCRAW is used to validate Brain-ID's effectiveness. (v) CIFL [15]: a baseline with the same data generation/architecture as Brain-ID and SCRAW, yet initialized with CIFL's pre-trained features. We use CIFL to demonstrate the superiority of Brain-ID's anatomy-guided representation. + +# Contrast-independent Applications + +- Anatomy Reconstruction/Contrast Synthesis As one of the most common types of MRI scans, T1w MR images highlight the differences between gray and white matter, and are mostly often utilized to image the anatomy of the brain, spinal cord, bones and joints [51]. For each dataset, we train models to reconstruct their T1w MRI counterparts with the L1 loss. + +- Brain Segmentation For each dataset, we use SynthSeg [4] to obtain the segmentation labels with 30 brain anatomical regions [20], as the gold standard segmentation target. We train our compared models to predict the brain segmentation labels, with the soft dice loss and cross-entropy loss [4]. + +As shown in Fig. 5, Brain-ID clearly outperforms CIFL and SCRATCH in the reconstruction task. As highlighted by the mint circles, it reveals the anatomy details and produces more fine-grained results than the strong state-of-the-art method SynthSR [27]. Brain-ID also achieves better segmentation results, especially in the smaller and more challenging regions. Quantitative comparisons are in Tab. 2, where Brain-ID obtains the best scores across most metrics. No- + +Table 2: Comparisons of Brain-ID with state-of-the-art approaches on downstream applications. ("train/test" refers to the number of subjects in the training/testing set.) + +
Modality (Contrast)Dataset (Train/Test)MethodReconstructionSegmentationSuper-resolution (from 3 / 5 / 7 mm)Bias Field norm-12 (↓)
L1 (↓)PSNR (↑)SSIM (↑)MS-SSIM (↑)Dice (↑)PSNR (↑)SSIM (↑)MS-SSIM (↑)MS-SSIM (↑)
MR (T1w)ADNI [30] (1841/204)SAMSEG [13]---0.795------
FastSurfer [24]---0.803------
SynthSR [27]0.01426.780.9800.980-27.01 / 25.81 / 21.790.977 / 0.943 / 0.8570.981 / 0.968 / 0.869--
SynthSR-FT0.01029.540.9840.9200.82530.46 / 28.03 / 23.910.979 / 0.960 / 0.8980.985 / 0.973 / 0.900-0.051
CIFL [15]0.01326.970.9730.9610.82030.12 / 27.21 / 23.800.978 / 0.912 / 0.8870.979 / 0.950 / 0.842-0.053
SCRATCH0.01127.540.9790.9570.81630.29 / 27.94 / 24.570.975 / 0.959 / 0.9200.982 / 0.964 / 0.866-0.056
Brain-ID0.01233.820.9930.9890.83731.30 / 28.62 / 24.680.983 / 0.961 / 0.9280.987 / 0.980 / 0.947-0.048
MR (T1w)HCP [19] (808/87)SAMSEG [13]---0.810------
FastSurfer [24]---0.819------
SynthSR [27]0.03322.130.8540.901-22.21 / 20.83 / 19.930.848 / 0.802 / 0.7470.899 / 0.864 / 0.789--
SynthSR-FT0.02527.110.9350.9170.81025.90 / 25.45 / 25.100.889 / 0.872 / 0.8310.919 / 0.911 / 0.889-0.052
CIFL [15]0.02926.420.9320.9130.80425.98 / 25.11 / 23.990.930 / 0.883 / 0.8710.936 / 0.927 / 0.880-0.059
SCRATCH0.02327.320.9230.9090.79226.69 / 26.41 / 25.050.932 / 0.896 / 0.8480.942 / 0.930 / 0.892-0.069
Brain-ID0.02027.470.9570.9290.84329.70 / 27.90 / 26.120.957 / 0.901 / 0.8830.973 / 0.938 / 0.890-0.047
SAMSEG [13]---0.769------
FastSurfer [24]---0.796------
SynthSR [27]0.02323.600.9280.909-23.13 / 22.40 / 22.270.921 / 0.907 / 0.8760.903 / 0.892 / 0.871--
SynthSR-FT0.02127.110.9500.9690.82027.95 / 27.34 / 26.180.920 / 0.893 / 0.9220.890 / 0.890 / 0.879-0.134
CIFL [15]0.02828.520.9610.9700.81927.81 / 27.10 / 26.320.900 / 0.877 / 0.8640.812 / 0.793 / 0.769-0.102
SCRATCH0.03327.280.9570.9630.81627.99 / 27.37 / 26.630.893 / 0.889 / 0.8760.794 / 0.785 / 0.753-0.128
Brain-ID0.02129.890.9660.9760.84330.17 / 28.23 / 26.890.962 / 0.923 / 0.8920.973 / 0.921 / 0.883-0.093
SAMSEG [13]---0.784------
FastSurfer [24]---0.801------
SynthSR [27]0.03521.670.8820.853-21.42 / 21.13 / 21.410.876 / 0.846 / 0.8090.851 / 0.831 / 0.805--
SynthSR-FT0.01626.900.9170.9050.81228.55 / 27.43 / 25.940.880 / 0.859 / 0.8230.860 / 0.845 / 0.811-0.110
CIFL [15]0.01328.690.9320.9210.81028.11 / 26.12 / 25.110.865 / 0.848 / 0.8310.828 / 0.800 / 0.789-0.109
SCRATCH0.01131.550.9860.9850.79629.41 / 27.89 / 26.580.854 / 0.847 / 0.8140.853 / 0.845 / 0.827-0.113
Brain-ID0.01132.480.9960.9960.84729.50 / 28.56 / 26.870.898 / 0.862 / 0.8110.887 / 0.850 / 0.823-0.107
SAMSEG [13]---0.799------
FastSurfer [24]---0.802------
SynthSR [27]0.02622.950.9160.912-22.57 / 21.82 / 21.780.917 / 0.892 / 0.8560.906 / 0.893 / 0.869--
SynthSR-FT0.01429.890.9410.9220.81029.15 / 27.89 / 26.770.932 / 0.906 / 0.8780.927 / 0.900 / 0.871-0.118
CIFL [15]0.01130.120.9380.9250.81629.46 / 27.97 / 26.730.907 / 0.889 / 0.8700.891 / 0.872 / 0.857-0.094
SCRATCH0.01130.530.9320.9130.81429.43 / 28.10 / 26.820.900 / 0.886 / 0.8630.884 / 0.870 / 0.853-0.128
Brain-ID0.00933.730.9720.9630.85129.75 / 28.58 / 27.090.957 / 0.922 / 0.8900.975 / 0.956 / 0.937-0.088
SAMSEG [13]---0.782------
FastSurfer [24]---0.805------
SynthSR [27]0.02826.100.8940.8980.75528.43 / 25.36 / 22.100.942 / 0.851 / 0.8010.977 / 0.892 / 0.798-0.140
SynthSR-FT0.03625.120.8910.8790.78728.26 / 25.78 / 21.730.959 / 0.893 / 0.7970.982 / 0.906 / 0.742-0.138
SCRATCH0.03824.990.8730.8660.75628.12 / 24.52 / 21.580.945 / 0.863 / 0.7830.980 / 0.893 / 0.738-0.136
Brain-ID0.01628.100.9340.9350.80530.26 / 26.11 / 24.100.959 / 0.902 / 0.8320.987 / 0.953 / 0.912-0.136
SAMSEG [13]---0.763------
FastSurfer [24]---0.801------
SynthSR [27]0.03320.080.8050.831------
SynthSR-FT0.02422.930.8150.8400.71930.15 / 28.09 / 26.920.940 / 0.912 / 0.8930.966 / 0.948 / 0.922-0.200
CIFL [15]0.02223.710.8200.8390.72129.98 / 27.01 / 25.660.931 / 0.899 / 0.8610.967 / 0.941 / 0.905-0.193
SCRATCH0.02022.270.8490.8510.71431.91 / 29.20 / 27.350.954 / 0.934 / 0.8960.982 / 0.969 / 0.949 / 0.936-0.197
Brain-ID0.02223.990.8610.8500.78232.26 / 29.90 / 27.090.937 / 0.930 / 0.9080.988 / 0.971 / 0.948-0.148
SAMSEG [13]---0.718------
SynthSR [27]0.02622.770.9190.895-30.11 / 28.70 / 25.320.930 / 0.899 / 0.8640.948 / 0.910 / 0.878-0.251
SynthSR-FT0.02123.340.9210.9000.75332.72 / 29.00 / 26.990.949 / 0.906 / 0.8730.953 / 0.919 / 0.889-0.237
CIFL [15]0.02021.290.9110.8970.76132.72 / 28.71 / 25.860.945 / 0.915 / 0.8770.941 / 0.936 / 0.917-0.136
SCRATCH0.02520.800.9010.8620.75932.36 / 28.71 / 25.840.945 / 0.915 / 0.8770.981 / 0.953 / 0.917-0.236
Brain-ID0.01726.440.9270.8920.78632.18 / 30.00 / 28.280.959 / 0.921 / 0.8830.982 / 0.965 / 0.921-0.227
SAMSEG [13]---0.710------
FastSurfer [24]---0.795------
SynthSR [27]0.02921.770.9020.892------
SynthSR-FT0.02425.800.9140.9000.73526.20 / 24.91 / 23.170.860 / 0.832 / 0.7950.933 / 0.910 / 0.894-0.126
CIFL [15]0.02627.110.8910.8700.72127.72 / 26.85 / 25.950.821 / 0.858 / 0.8140.969 / 0.939 / 0.905 / 0.901-0.109
SCRATCH0.02326.840.8980.8670.72028.02 / 26.52 / 25.430.878 / 0.838 / 0.7990.943 / 0.901 / 0.891-0.115
Brain-ID0.01927.250.9360.9120.76726.69 / 24.01 / 23.69 / 23.09 / 0.918 / 0.806 / 0.966 / 0.871 / 0.925 / 0.916---
+ +(1) "means not applicable: FastSurfer [24] and SAMSEG [13] are for segmentation-only; SynthSR [27] only synthesize T1w MRI; CT does not have bias fields. (2) SynthSR ET; SynthSR fine tuned on each task/dataset, for a fairer comparison. + +tably, Brain-ID achieves greater gains over baseline models, SCRATCH and CIFL, particularly on smaller datasets (e.g., ADNI3-T1/FLAIR, AIBL-FLAIR, where Brain-ID's robust and rich features quickly adapt to specific tasks/datasets. + +Contrast-dependent Applications We also evaluate Brain-ID on two brain imaging tasks that are dependent on the input modality/quality. + +- Image Super-resolution We use the standard 1mm isotropic images from all datasets as the ground truth high-resolution target, and follow the strategy in SynthSR [27] where the input samples are randomly resampled and corrupted during training. For inference, we downsample the original images into 3mm, 5mm, and 7mm isotropic images as inputs, and evaluate the output quality. + +![](images/fbd526bb28e75ae9e993bbda0e2a7d4230d8705df90091974255f28ff632932f.jpg) +Fig. 6: SCRATCH, which is well trained on HF T1w scans, produces highly descriptive features for HF T1w images ( $1^{\mathrm{st}}$ row), but does not preserve the same high quality useful for downstream tasks when handling LF ( $2^{\mathrm{nd}}$ row) or other contrasts ( $3^{\mathrm{rd}}$ row). + +- Bias Field Estimation The bias field is a smooth, low-frequency multiplicative signal that corrupts MRI images, which affects image analysis tasks such as segmentation or texture analysis [31]. Bias field estimation is often needed as a pre-processing step to correct corrupted MRI images [16]. We apply randomly simulated bias fields to the input samples (Sec. 3.1), and train all models with the L2 loss. During inference, we pre-generate and apply the bias fields on the testing data for reproducibility, and evaluate the bias field estimation performance. + +With a simple one-layer adaption (Sec. 3.3), Brain-ID's contrast-agnostic anatomical features achieve state-of-the-art performance on contrast-dependent tasks (Tab. 2, Fig. 5). As shown in Fig. 6, Brain-ID obtains contrast/resolution-robustness that cannot be achieved by models trained from real images (due to the limited variability in their appearance), regardless of the backbone choices. Such robust features greatly improve super-resolution, resulting in higher (MS-)SSIM scores. Brain-ID's gains are less obvious for bias field estimation, probably because the bias field is approximately independent of brain anatomy. + +# 4.4 Practical Value and Broader Impact + +Low-resolution data Trained on synthetic images, Brain-ID features are not only robust to various contrasts/modalities, but also extremely robust to low-resolution data (Fig. 6) and provide huge potential in clinical MRI (including big retrospective data [5]) and portable low-field MRI [29]. + +Small-size Datasets To investigate the effectiveness of Brain-ID features for limited-size datasets, we assess its performance across all four downstream tasks on subsets of ADNI3 [60] training set (298 cases originally). As shown in Fig. 7, we reduce the training set size percentage from $100\%$ to an extreme $10\%$ . With + +![](images/86e67cf7953cd9c059f570d091c039fe9b914201ffba548255a97a1d818ec3ae.jpg) +(c) Super-resolution +(d) Bias Field Estimation +Fig. 7: Adapting Brain-ID to small datasets. The horizontal (vertical) axes indicate training epochs (evaluation scores of corresponding tasks). Results are obtained by evaluating models collected throughout the epochs, on ADNI3 [60] full testing set. + +only $20\%$ of the data (mint line), Brain-ID still achieves better performance compared to its full-sized baseline, “ $100\%$ ” SCRATCH (yellow line), over all tasks. 30 training samples ( $10\%$ , blue line) may be at the edge of obtaining acceptable results for Brain-ID, as the model becomes less stable and the performance drops occur more frequently – especially for brain segmentation (Fig. 7-(b)). + +# 4.5 Additional Experiments + +Choice of Anatomy Guidance We explore other anatomical learning targets than MP-RAGE (Sec. 3.2): $(i)$ segmentation labels; $(ii)$ both segmentation labels and MP-RAGE. As shown in Tab. 3, adding segmentation guidance improves robustness, yet inhibits the feature expressiveness and affects the downstream performance; we observe features from segmentation guidance produce less high-frequency texture than Brain-ID, especially within the same-label regions. + +Data Generation Design As shown in Tab. 3, training with all mildly-corrupted samples results in reduced robustness and harms the downstream performance, yet using all extremely corrupted samples leads to unstable training and model collapse. Brain-ID uses samples of gradually increased corruptions (Fig. 3 (left)). + +Table 3: Comparison between Brain-ID and its variants. + +
Model Setup [Comparison Target]Feature Robustness (Intra)Downstream (Reconstruction)
SSIM (↑)MS-SSIM (↑)L1 (↓)PSNR (↑)SSIM (↑)
[Representation guidance: supervision target for feature learning]
Segmentation0.8910.9630.02928.130.958
Segmentation + MP-RAGE0.8630.9400.02329.050.964
MP-RAGE (Brain-ID)0.8580.9210.02129.890.966
[Data generation design: corruption levels in intra-subject mini-batches]
All Mild (\( \sigma_{noise} = 1, \cdots \))0.7920.8130.02828.300.960
All Medium (\( \sigma_{noise} = 5, \cdots \))0.8310.8990.02229.670.964
All Severe (\( \sigma_{noise} = 10, \cdots \))N/AN/AN/AN/AN/A
Mild to Severe (Brain-ID)0.8580.9210.02129.890.966
[Mini-batch size: number of intra-subject samples]
20.8260.8970.02528.830.959
30.8410.9020.02229.300.962
4 (Brain-ID)0.8580.9210.02129.890.966
50.8600.9290.02129.930.965
+ +Batch Size Brain-ID computes training loss of all intra-subject samples in a mini-batch at once (Eq. (3)). We observe that larger batches help improve feature robustness (Tab. 3), yet do not further enhance the downstream performance. + +# 4.6 Further Discussions + +Limitations Although Brain-ID features excel in multiple tasks, they may not be optimal for tasks not involving voxel-wise prediction, e.g., functional brain mapping. Further, as Brain-ID generates data exclusively from anatomies, it tends to struggle with images with extensive pathology. Our future research will focus on robust representations that are encoded with pathology information. + +Data Privacy and Ethics Since Brain-ID features can potentially be used to reconstruct a high-resolution MRI scan, publicly sharing such features should be regarded as sharing the original scans from a subject [33]. Therefore, one should follow the same anonymization protocols as when sharing MRI data for research [61] – not only the removal of identifiers, but also defacing [6] or skull-stripping [26] to prevent the 3D reconstruction of faces. + +# 5 Conclusion + +We introduced Brain-ID, a contrast-agnostic feature representation learning approach for brain imaging, which is distinctive to brain anatomy and resilient to variations in image appearances. Brain-ID is trained on synthetic inputs, and quickly adapts to downstream tasks through one layer. We validated Brain-ID features' robustness, and effectiveness on four downstream applications (reconstruction, segmentation, super-resolution, bias field estimation), and six public datasets, covering MR and CT. Brain-ID achieves state-of-the-art performances across all tasks and modalities, it further preserves high performance on low-resolution and small-size datasets. We believe Brain-ID unlocks the great potential of robust foundation models, especially for non-calibrated modalities. + +# Acknowledgements + +Primarily supported by NIH under grant 1RF1AG080371. Additional support from NIH 1RF1AG080371, 1R01EB031114, 1R01AG070988, 1RF1MH123195. OP is supported by Lundbeckfonden R360-2021-395. Wellcome Trust award 221915/Z/20/Z and the NIHR UCLH Biomedical Research Centre support DCA on this topic. + +# References + +1. Adnan, M., Kalra, S., Cresswell, J.C., Taylor, G.W., Tizhoosh, H.R.: Federated learning and differential privacy for medical image analysis. Scientific reports (2022) +2. Awais, M., Naseer, M., Khan, S.S., Anwer, R.M., Cholakkal, H., Shah, M., et al.: Foundational models defining a new era in vision: A survey and outlook. arXiv abs/2307.13721 (2023) +3. Balakrishnan, G., Zhao, A., Sabuncu, M.R., Guttag, J.V., Dalca, A.V.: VoxelMorph: A learning framework for deformable medical image registration. IEEE Transactions on Medical Imaging (2018) +4. Billot, B., Greve, D.N., Puonti, O., Thielscher, A., Leemput, K.V., Fischl, B., et al.: SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining. Medical Image Analysis (2021) +5. Billot, B., Magdamo, C., Cheng, Y., Arnold, S.E., Das, S., Iglesias, J.E.: Robust machine learning segmentation for large-scale analysis of heterogeneous clinical brain MRI datasets. Proceedings of the National Academy of Sciences (2023) +6. Bischoff-Grethe, A., Ozyurt, I.B., Busa, E., Quinn, B.T., Fennema-Notestine, C., Clark, C.P., et al.: A technique for the deidentification of structural brain MR images. Human Brain Mapping (2007) +7. Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., et al.: On the opportunities and risks of foundation models. arXiv abs/2108.07258 (2021) +8. Brant-Zawadzki, M.N., Gillan, G.D., Nitz, W.R.: MP RAGE: A three-dimensional, T1-weighted, gradient-echo sequence-initial experience in the brain. Radiology (1992) +9. Brown, M.R.G., Sidhu, G.P.S., Greiner, R., Asgarian, N., Bastani, M., Silverstone, P.H., et al.: ADHD-200 global competition: diagnosing ADHD using personal characteristic data can outperform resting state fMRI measurements. Frontiers in Systems Neuroscience (2012) +10. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., et al.: Language models are few-shot learners. In: NeurIPS (2020) +11. Cardoso, M.J., Li, W., Brown, R., Ma, N., Kerfoot, E., Wang, Y., et al.: MONAI: An open-source framework for deep learning in healthcare. arXiv abs/2211.02701 (2022) +12. Caron, M., Touvron, H., Misra, I., J'egou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021) +13. Cerri, S., Puonti, O., Meier, D.S., Wuerfel, J., Mühlau, M., Siebner, H.R., Leemput, K.V.: A contrast-adaptive method for simultaneous whole-brain and lesion segmentation in multiple sclerosis. NeuroImage (2020) + +14. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., et al.: PaLM: Scaling language modeling with pathways. JMLR (2022) +15. Chua, Y.Z.R., Dalca, A.V.: Contrast invariant feature representations for segmentation and registration of medical images. In: MIDL (2023) +16. Chua, Z.Y., Zheng, W., Chee, M.W.L., Zagorodnov, V.: Evaluation of performance metrics for bias field correction in MR brain images. Journal of Magnetic Resonance Imaging (2009) +17. Darzidehkalani, E., Ghasemi-Rad, M., van Ooijen, P.: Federated learning in medical imaging: Part I: Toward multicentral health care ecosystems. Journal of the American College of Radiology (2022) +18. Ding, Z., Han, X., Liu, P., Niethammer, M.: Local temperature scaling for probability calibration. In: ICCV (2021) +19. Essen, D.C.V., Ugurbil, K., Auerbach, E.J., Barch, D.M., Behrens, T.E.J., Bucholz, R.D., et al.: The human connectome project: A data acquisition perspective. NeuroImage (2012) +20. Fischl, B., Salat, D.H., Busa, E., Albert, M.S., Dieterich, M., Haselgrove, C., et al.: Whole brain segmentation automated labeling of neuroanatomical structures in the human brain. Neuron (2002) +21. Fonov, V.S., Evans, A.C., Botteron, K., Almli, C.R., McKinstry, R.C., Collins, D.L., et al.: Unbiased average age-appropriate atlases for pediatric studies. Neuroimage (2011) +22. Fonov, V.S., Evans, A.C., McKinstry, R.C., Almli, C.R., Collins, D.: Unbiased nonlinear average age-appropriate brain templates from birth to adulthood. NeuroImage (2009) +23. Fowler, C., Rainey-Smith, S.R., Bird, S., Bomke, J., Bourgeat, P., Brown, B.M., et al.: Fifteen years of the australian imaging, biomarkers and lifestyle (AIBL) study: Progress and observations from 2,359 older adults spanning the spectrum from cognitive normality to Alzheimer's disease. Journal of Alzheimer's disease reports (2021) +24. Henschel, L., Kugler, D., Reuter, M.: FastSurferVINN: Building resolution-independence into deep learning segmentation methods — A solution for highres brain MRI. NeuroImage (2022) +25. Hoffmann, M., Billot, B., Greve, D.N., Iglesias, J.E., Fischl, B., Dalca, A.V.: SynthMorph: Learning contrast-invariant registration without acquired images. IEEE Transactions on Medical Imaging (2020) +26. Hoopes, A., Mora, J.S., Dalca, A.V., Fischl, B., Hoffmann, M.: SynthStrip: skullstripping for any brain image. NeuroImage (2022) +27. Iglesias, J.E., Billot, B., Balbastre, Y., Magdamo, C.G., Arnold, S., Das, S., et al.: SynthSR: A public AI tool to turn heterogeneous clinical brain scans into high-resolution T1-weighted images for 3D morphometry. Science Advances (2023) +28. Iglesias, J.E., Billot, B., Balbastre, Y., Tabari, A., Conklin, J., Alexander, D.C., et al.: Joint super-resolution and synthesis of $1\mathrm{mm}$ isotropic MP-RAGE volumes from clinical MRI exams with scans of different orientation, resolution and contrast. NeuroImage (2020) +29. Iglesias, J.E., Schleicher, R., Laguna, S., Billot, B., Schaefer, P., McKaig, B., et al.: Quantitative brain morphometry of portable low-field-strength MRI using superresolution machine learning. Radiology (2022) +30. Jack, C.R., Bernstein, M.A., Fox, N.C., Thompson, P.M., Alexander, G.E., Harvey, D.J., et al.: The Alzheimer's disease neuroimaging initiative (ADNI): MRI methods. Journal of Magnetic Resonance Imaging (2008) + +31. Juntu, J., Sijbers, J., Dyck, D.V., Gielen, J.: Bias field correction for MRI images. In: CORES (2005) +32. Kamnitsas, K., Ledig, C., Newcombe, V.F.J., Simpson, J.P., Kane, A.D., Menon, D.K., et al.: Efficient multi-scale 3D CNN with fully connected CRF for accurate brain lesion segmentation. Medical Image Analysis (2016) +33. Kapkiç, M., Sağiroğlu, Ş.: Privacy issues in magnetic resonance images. International Journal of Information Security Science (2023) +34. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., et al.: Segment anything. arXiv abs/2304.02643 (2023) +35. LaMontagne, P.J., Keefe, S.J., Lauren, W., Xiong, C., Grant, E.A., Moulder, K.L., et al.: OASIS-3: Longitudinal neuroimaging, clinical, and cognitive dataset for normal aging and Alzheimer's disease. Alzheimer's & Dementia (2018) +36. Laso, P., Cerri, S., Sorby-Adams, A., Guo, J., Mateen, F., Goebl, P., et al.: Quantifying white matter hyperintensity and brain volumes in heterogeneous clinical and low-field portable MRI. arXiv abs/2312.05119 (2023) +37. Leemput, K.V., Maes, F., Vandermeulen, D., Suetens, P.: A unifying framework for partial volume segmentation of brain MR images. IEEE Transactions on Medical Imaging (2003) +38. Li, Z., Tian, L., Mok, T.C.W., Bai, X., Wang, P., Ge, J., et al.: SAMConvex: Fast discrete optimization for CT registration using self-supervised anatomical embedding and correlation pyramid. In: MICCAI (2023) +39. Liu, P., Lee, Y., Aylward, S., Niethammer, M.: Deep decomposition for stochastic normal-abnormal transport. In: CVPR (2022) +40. Liu, P., Tian, L., Zhang, Y., Aylward, S., Lee, Y., Niethammer, M.: Discovering hidden physics behind transport dynamics. In: CVPR (2021) +41. Liu, P., Wang, R., Cao, X., Zhou, Y., Shah, A., Lim, S.N.: Differential motion evolution for fine-grained motion deformation in unsupervised image animation. arXiv abs/2110.04658 (2021) +42. Miletari, F., Navab, N., Ahmadi, S.A.: V-Net: Fully convolutional neural networks for volumetric medical image segmentation. 3DV (2016) +43. Modat, M., Ridgway, G.R., Taylor, Z.A., Lehmann, M., Barnes, J., Hawkes, D.J., et al.: Fast free-form deformation using graphics processing units. Computer Methods and Programs in Biomedicine (2010) +44. Moor, M., Banerjee, O., Abad, Z.F.H., Krumholz, H.M., Leskovec, J., Topol, E.J., Rajpurkar, P.: Foundation models for generalist medical artificial intelligence. Nature (2023) +45. Müller, R.A., Shih, P., Keehn, B., Deyoe, J.R., Leyden, K.M., Shukla, D.: Underconnected, but how? A survey of functional connectivity MRI studies in autism spectrum disorders. Cerebral Cortex (2011) +46. Puonti, O., Iglesias, J.E., Leemput, K.V.: Fast and sequence-adaptive whole-brain segmentation using parametric bayesian modeling. NeuroImage (2016) +47. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015) +48. Sheller, M.J., Edwards, B., Reina, G.A., Martin, J., Pati, S., Kotrotsou, A., et al.: Federated learning in medicine: Facilitating multi-institutional collaborations without sharing patient data. Scientific reports (2020) +49. Shen, Z., Feydy, J., Liu, P., Curiale, A.H., San Jose Estepar, R., San Jose Estepar, R., Niethammer, M.: Accurate point cloud registration with robust optimal transport. In: NeurIPS (2021) +50. Singhal, K., Azizi, S., Tu, T., Mahdavi, S., Wei, J., Chung, H.W., et al.: Large language models encode clinical knowledge. Nature (2022) + +51. Symms, M.R., Jäger, H.R., Schmierer, K., Yousry, T.A.: A review of structural magnetic resonance neuroimaging. Journal of Neurology, Neurosurgery & Psychiatry (2004) +52. Tanno, R., Worrall, D.E., Kaden, E., Alexander, D.C.: Uncertainty modelling in deep learning for safer neuroimage enhancement: Demonstration in diffusion MRI. NeuroImage (2020) +53. Tian, Q., Bilgiç, B., Fan, Q., Ngamsombat, C., Zaretskaya, N., Fultz, N.E., et al.: Improving in vivo human cerebral cortical surface reconstruction using data-driven super-resolution. Cerebral Cortex (2020) +54. Tu, T., Azizi, S., Driess, D., Schaekermann, M., Amin, M., Chang, P.C., et al.: Towards generalist biomedical AI. arXiv abs/2307.14334 (2023) +55. de Vos, B.D., Berendsen, F.F., Viergever, M.A., Sokooti, H., Staring, M., Isgum, I.: A deep learning framework for unsupervised affine and deformable image registration. Medical Image Analysis (2019) +56. Wachinger, C., Golland, P., Reuter, M.: BrainPrint: Identifying subjects by their brain. In: MICCAI (2014) +57. Wang, M., Deng, W.: Deep visual domain adaptation: A survey. Neurocomputing (2018) +58. Wang, T.C., Mallya, A., Liu, M.Y.: One-shot free-view neural talking-head synthesis for video conferencing. In: CVPR (2021) +59. Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for image quality assessment. In: ACSSC (2003) +60. Weiner, M.W., Veitch, D.P., Aisen, P.S., Beckett, L.A., Cairns, N.J., Green, R.C., et al.: The Alzheimer's disease neuroimaging initiative 3: Continued innovation for clinical trial improvement. Alzheimer's & Dementia (2017) +61. White, T., Blok, E., Calhoun, V.D.: Data sharing and privacy issues in neuroimaging research: Opportunities, obstacles, challenges, and monsters under the bed. Human Brain Mapping (2022) +62. Xu, Y., Mo, T., Feng, Q., Zhong, P., Lai, M., Chang, E.I.C.: Deep learning of feature representation with multiple instance learning for medical image analysis. In: ICASSP (2014) +63. Yan, K., Cai, J., Jin, D., Miao, S., Harrison, A.P., Guo, D., et al.: SAM: Self-supervised learning of pixel-wise anatomical embeddings in radiological images. IEEE Transactions on Medical Imaging (2020) +64. Yang, X., Kwitt, R., Niethammer, M.: Quicksilver: Fast predictive image registration - a deep learning approach. NeuroImage (2017) +65. You, C., Zhao, R., Staib, L.H., Duncan, J.S.: Momentum contrastive voxel-wise representation learning for semi-supervised volumetric medical image segmentation. In: MICCAI (2021) +66. Zhou, Y., Chia, M.A., Wagner, S.K., Ayhan, M.S., Williamson, D.J., Struyven, R.R., et al.: A foundation model for generalizable disease detection from retinal images. Nature (2023) \ No newline at end of file diff --git a/2024/Brain-ID_ Learning Contrast-agnostic Anatomical Representations for Brain Imaging/images.zip b/2024/Brain-ID_ Learning Contrast-agnostic Anatomical Representations for Brain Imaging/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8f6e6b20dc799bd706c958e2d1835937b5b3b092 --- /dev/null +++ b/2024/Brain-ID_ Learning Contrast-agnostic Anatomical Representations for Brain Imaging/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:86bef457b3375c8ff1741f7692ee9dbba443da98c956aaa5768d03a127e5d259 +size 696766 diff --git a/2024/Brain-ID_ Learning Contrast-agnostic Anatomical Representations for Brain Imaging/layout.json b/2024/Brain-ID_ Learning Contrast-agnostic Anatomical Representations for Brain Imaging/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2da348b69301a17311dfef03988f85aa747c2ca8 --- /dev/null +++ b/2024/Brain-ID_ Learning Contrast-agnostic Anatomical Representations for Brain Imaging/layout.json @@ -0,0 +1,8695 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 112, + 479, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 112, + 479, + 149 + ], + "spans": [ + { + "bbox": [ + 135, + 112, + 479, + 149 + ], + "type": "text", + "content": "Brain-ID: Learning Contrast-agnostic Anatomical Representations for Brain Imaging" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 196, + 167, + 419, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 167, + 419, + 180 + ], + "spans": [ + { + "bbox": [ + 196, + 167, + 419, + 180 + ], + "type": "text", + "content": "Peirong Liu1 Oula Puonti1 Xiaoling Hu1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 202, + 180, + 411, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 180, + 411, + 193 + ], + "spans": [ + { + "bbox": [ + 202, + 180, + 411, + 193 + ], + "type": "text", + "content": "Daniel C. Alexander" + }, + { + "bbox": [ + 202, + 180, + 411, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 202, + 180, + 411, + 193 + ], + "type": "text", + "content": " \nJuan E. Iglesias" + }, + { + "bbox": [ + 202, + 180, + 411, + 193 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 164, + 201, + 449, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 201, + 449, + 225 + ], + "spans": [ + { + "bbox": [ + 164, + 201, + 449, + 225 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 164, + 201, + 449, + 225 + ], + "type": "text", + "content": "Harvard Medical School and Massachusetts General Hospital " + }, + { + "bbox": [ + 164, + 201, + 449, + 225 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 164, + 201, + 449, + 225 + ], + "type": "text", + "content": "University College London " + }, + { + "bbox": [ + 164, + 201, + 449, + 225 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 164, + 201, + 449, + 225 + ], + "type": "text", + "content": "Massachusetts Institute of Technology" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 135, + 244, + 480, + 379 + ], + "blocks": [ + { + "bbox": [ + 135, + 244, + 480, + 379 + ], + "lines": [ + { + "bbox": [ + 135, + 244, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 135, + 244, + 480, + 379 + ], + "type": "image", + "image_path": "24561250013a44bfa91f12541944ec764831343688bad5919067e368166efbd7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 387, + 482, + 433 + ], + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 433 + ], + "type": "text", + "content": "Fig. 1: Brain-ID features serve as the distinctive identity of each subject (ID) regardless of appearance (contrast, deformation, corruptions, etc) and quickly adapt to downstream tasks – either contrast-independent (anatomy Reconstruction, Segmentation), or contrast-dependent (Super-Resolution, Bias Field estimation), all through one layer." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 445, + 455, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 445, + 455, + 666 + ], + "spans": [ + { + "bbox": [ + 159, + 445, + 455, + 666 + ], + "type": "text", + "content": "Abstract. Recent learning-based approaches have made astonishing advances in calibrated medical imaging like computerized tomography (CT). Yet, they struggle to generalize in uncalibrated modalities – notably magnetic resonance (MR) imaging, where performance is highly sensitive to the differences in MR contrast, resolution, and orientation. This prevents broad applicability to diverse real-world clinical protocols. We introduce Brain-ID, an anatomical representation learning model for brain imaging. With the proposed \"mild-to-severe\" intra-subject generation, Brain-ID is robust to the subject-specific brain anatomy regardless of the appearance of acquired images. Trained entirely on synthetic inputs, Brain-ID readily adapts to various downstream tasks through one layer. We present new metrics to validate the intra/inter-subject robustness of Brain-ID features, and evaluate their performance on four downstream applications, covering contrast-independent (anatomy reconstruction, brain segmentation), and contrast-dependent (super-resolution, bias field estimation) tasks (Fig. 1). Extensive experiments on six public datasets demonstrate that Brain-ID achieves state-of-the-art performance in all tasks on different MR contrasts and CT, and more importantly, preserves its performance on low-resolution and small datasets. Code is available at https://github.com/peirong26/Brain-ID." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 229, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 229, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 229, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 279 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 279 + ], + "type": "text", + "content": "Magnetic resonance imaging (MRI) enables in vivo noninvasive imaging of the human brain with exquisite and tunable soft-tissue contrast [8]. Recent machine learning methods have achieved great improvements in faster and more accurate MRI analysis [27], such as segmentation [18,32,42,47], registration [3,49,55,64], super-resolution [52, 53], and connectivity studies [45]. However, most existing MRI analysis methods are specific to certain MR contrast(s) and require near-isotropic acquisitions. Therefore, models face sharp performance drops when voxel size and anisotropy increase, or are being used for a different contrast than training [57]. This reduces model generalizability and results in duplicate training efforts. Resorting to synthetic data, recent contrast-agnostic models [4, 25, 27, 28, 36, 39, 40] achieve impressive results and largely extend model applicability. However, these models are only applicable to the tasks they were trained for." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 281, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 460 + ], + "type": "text", + "content": "Meanwhile, task-agnostic foundation models in computer vision and natural language processing have witnessed remarkable success, along with the fast developments of large-scale datasets [2,7,10,14,34]. However, due to different acquisition/processing protocols and privacy requirements across institutions, large-scale medical imaging datasets require significantly more effort than natural imaging/language. Thus, medical foundation models are not as well developed. MONAI [11] includes a pre-trained model zoo, yet all are highly task-oriented and sensitive to contrasts. Zhou et al. [66] constructed a medical foundation model, designed for eye and systemic health condition detections from retinal scans, and only works on the modalities of color fundus photography and optical coherence tomography. Lately, generalist biomedical AI systems [44, 50, 54] have shown great potential in biomedical tasks under vision-language context, e.g., (visual) question answering, image classification, radiology report generation and summarization. However, they have not explored the challenging vision tasks such as reconstruction, segmentation, super-resolution, and registration." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 460, + 481, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 481, + 484 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 481, + 484 + ], + "type": "text", + "content": "We introduce Brain-ID, a contrast-agnostic feature representation approach trained on synthetic inputs, that can adapt to various tasks through one layer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 487, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 135, + 487, + 480, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 487, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 135, + 487, + 480, + 534 + ], + "type": "text", + "content": "1) We introduce an on-the-fly, intra-subject data generator capable of synthesizing any contrast, with a mild-to-severe corruption strategy (Fig. 2, Sec. 3.2). Unlike real-world datasets constrained by the images acquired per subject, Brain-ID learns in a much more expansive and diverse space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 535, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 535, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 135, + 535, + 480, + 593 + ], + "type": "text", + "content": "2) We design a feature representation learning framework guided by the unique anatomy of each subject. We validate that Brain-ID's high-resolution features are consistently robust to the superficial perturbations of an image's appearance (Sec. 4.2). Straightforward one-layer adaptions are further presented to adapt Brain-ID features to downstream tasks (Sec. 3.3)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 135, + 594, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 594, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 594, + 480, + 665 + ], + "type": "text", + "content": "3) We evaluate Brain-ID on both contrast-independent (anatomy reconstruction, brain segmentation) and contrast-dependent (super-resolution, bias field estimation) tasks, across six public datasets (" + }, + { + "bbox": [ + 135, + 594, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\approx" + }, + { + "bbox": [ + 135, + 594, + 480, + 665 + ], + "type": "text", + "content": " 8,000 images) including MR (T1w, T2w, FLAIR) and CT. Brain-ID achieves state-of-the-art performance on all tasks (Tab. 2), and further maintains its high performance on low-resolution images (Fig. 6) and limited-size datasets (Fig. 7)." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "text", + "content": "P. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 316 + ], + "type": "text", + "content": "Feature Representation in Medical Imaging As mentioned in Sec. 1, general feature representation learning in medical imaging can be more challenging than in the natural image domain, due to limited data availability. Xu et al. [62] introduced a multiple-instance learning model for feature representation, but it only applies to histopathology image classification. You et al. [65] presented CVRL, a semi-supervised representations approach designed for image segmentation, but it requires CT to extract anatomical information. SAM [63] encodes anatomical information in feature embeddings, which has shown to be effective for registration [38]. However, same as CVRL, SAM only works on CT. BrainPrint [56] is a compact and discriminative representation of brain morphology, which is specifically designed for cortical surface analyses. To our best knowledge, CIFL [15] is the only existing work on learning contrast-agnostic and task-independent brain feature representations. CIFL relies on contrastive learning alone, and is insufficient to outperform task/contrast-specific supervised models in downstream applications as shown by our experiments (Tab. 2)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 330, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 462 + ], + "type": "text", + "content": "Contrast-invariant Learning for MR Images MRI scans acquired across sites vary substantially in contrasts, resolutions, orientations, etc. When given a new dataset, heterogeneity leads to duplicate training efforts for approaches that are sensitive to specific combinations of MR contrast, resolution, or orientation. Classical brain segmentation models used Bayesian inference for contrast-robustness [20,37], which requires a long processing time and struggles with low or anisotropic resolutions [27, 46]. Recently, there have been works using synthetic data to achieve contrast-invariance in tasks like image segmentation [4,36], registration [25], super-resolution [28], and skull-stripping [26]. However, all the above-mentioned methods are trained in a task-specific manner, their features therefore cannot be readily applied to other domains." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 476, + 462, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 476, + 462, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 476, + 462, + 491 + ], + "type": "text", + "content": "3 Brain-ID: Learning Anatomy-specific Brain Features" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 481, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 556 + ], + "type": "text", + "content": "As discussed in Sec. 1, the main challenges to obtain a general and robust feature representation for MR images lie in " + }, + { + "bbox": [ + 130, + 498, + 481, + 556 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 130, + 498, + 481, + 556 + ], + "type": "text", + "content": " the practical restrictions of building large-scale datasets with diverse contrasts; and " + }, + { + "bbox": [ + 130, + 498, + 481, + 556 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 130, + 498, + 481, + 556 + ], + "type": "text", + "content": " the nature of most medical imaging models that are task-oriented and specific to data type (contrast, resolution, orientation, etc). We aim to learn a brain feature representation that is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 558, + 481, + 616 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 131, + 558, + 481, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 558, + 481, + 580 + ], + "spans": [ + { + "bbox": [ + 131, + 558, + 481, + 580 + ], + "type": "text", + "content": "(i) Robust: features should be robust to each subject's distinct anatomy, unaffected by variations in poses/deformations, contrasts, resolutions, or artifacts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 582, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 582, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 131, + 582, + 481, + 616 + ], + "type": "text", + "content": "(ii) Expressive: features should also exhibit high expressiveness, containing rich information that facilitates easy and effective adaptation to diverse downstream tasks, eliminating the necessity for extensive training data." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "We first introduce Brain-ID's data generator (Sec. 3.1) and training framework (Sec. 3.2) to achieve the above two aims. Then, we present our one-layer solutions for adapting Brain-ID features to downstream tasks that could be either dependent or independent of the contrast of input images (Sec. 3.3)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "type": "text", + "content": "Brain-ID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 114, + 478, + 196 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 478, + 196 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 478, + 196 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 478, + 196 + ], + "type": "image", + "image_path": "68490469bf1f48f04ec51c79a8e01d79643addf15b0050a8c34fe4585f03a669.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 207, + 482, + 240 + ], + "lines": [ + { + "bbox": [ + 130, + 207, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 130, + 207, + 482, + 240 + ], + "type": "text", + "content": "Fig.2: Brain-ID's data generator on the fly. Given the brain segmentation labels of a subject, we randomly generate a deformation field, and synthesize intra-subject samples featuring various contrast intensities and corruption levels (Sec. 3.1)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 262, + 363, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 262, + 363, + 274 + ], + "spans": [ + { + "bbox": [ + 131, + 262, + 363, + 274 + ], + "type": "text", + "content": "3.1 Enrich the Intra-subject Learning Space" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 279, + 482, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 482, + 362 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 482, + 362 + ], + "type": "text", + "content": "A good representation relies on large-scale data, however, different acquisition protocols, processing pipelines, and privacy requirements across institutions make large-scale data less accessible and require significantly more effort [1, 17, 48]. Moreover, for each subject, there are usually only limited acquired scans available. This lack of data consistency is a significant barrier to obtaining a robust, contrast-invariant feature representation. Brain-ID avoids these barriers through the use of synthetic data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 363, + 482, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 423 + ], + "type": "text", + "content": "To generate images with complex brain structures, we start with brain segmentation labels of anatomical structures (" + }, + { + "bbox": [ + 130, + 363, + 482, + 423 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 363, + 482, + 423 + ], + "type": "text", + "content": " in Fig. 2). The generation consists of three steps, " + }, + { + "bbox": [ + 130, + 363, + 482, + 423 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 130, + 363, + 482, + 423 + ], + "type": "text", + "content": " deformation generation, " + }, + { + "bbox": [ + 130, + 363, + 482, + 423 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 130, + 363, + 482, + 423 + ], + "type": "text", + "content": " contrast synthesis, and " + }, + { + "bbox": [ + 130, + 363, + 482, + 423 + ], + "type": "inline_equation", + "content": "(iii)" + }, + { + "bbox": [ + 130, + 363, + 482, + 423 + ], + "type": "text", + "content": " data corruption (including lower resolution resampling). For simplicity, " + }, + { + "bbox": [ + 130, + 363, + 482, + 423 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 363, + 482, + 423 + ], + "type": "text", + "content": " denotes the parameter group of the generation process described below." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 426, + 481, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 426, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 131, + 426, + 481, + 450 + ], + "type": "text", + "content": "Deformation Generation We first generate a random deformation field " + }, + { + "bbox": [ + 131, + 426, + 481, + 450 + ], + "type": "inline_equation", + "content": "(\\phi|_{\\theta_{\\phi}})" + }, + { + "bbox": [ + 131, + 426, + 481, + 450 + ], + "type": "text", + "content": " consisting of an affine transformation and a non-linear displacement field:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 262, + 457, + 481, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 457, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 262, + 457, + 481, + 471 + ], + "type": "interline_equation", + "content": "\\phi | _ {\\theta_ {\\phi}} = \\mathcal {T} | _ {\\theta_ {\\phi}} \\circ \\mathcal {A} | _ {\\theta_ {\\phi}}, \\tag {1}", + "image_path": "9b0ab02e485bc845d95c2a7c1888d657c8cc7958195288e2b87f7ceeac71edc6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{A}|_{\\theta_{\\phi}}" + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "content": " denotes an affine transformation matrix which includes linear (rotation, scaling, shearing) transformation and translation, " + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{T}|_{\\theta_{\\phi}}" + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "content": " refers to a nonlinear displacement field computed as the integration of a stationary velocity field (SVF) that is smooth and invertible everywhere and thus preserves the topology of the brain anatomy [28]. " + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\theta_{\\phi} \\in \\Theta" + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "content": " controls the transformation ranges." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 538, + 482, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 538, + 482, + 587 + ], + "spans": [ + { + "bbox": [ + 130, + 538, + 482, + 587 + ], + "type": "text", + "content": "MR Contrast Synthesis Then, we synthesize images " + }, + { + "bbox": [ + 130, + 538, + 482, + 587 + ], + "type": "inline_equation", + "content": "(I(x), x \\in \\Omega)" + }, + { + "bbox": [ + 130, + 538, + 482, + 587 + ], + "type": "text", + "content": " by randomly \"painting\" intensities on the segmentation maps according to their brain structure labels " + }, + { + "bbox": [ + 130, + 538, + 482, + 587 + ], + "type": "inline_equation", + "content": "(l \\in L)" + }, + { + "bbox": [ + 130, + 538, + 482, + 587 + ], + "type": "text", + "content": ". Specifically, the regional intensities are generated by separately sampling a Gaussian distribution on each labeled region:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 203, + 592, + 481, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 592, + 481, + 624 + ], + "spans": [ + { + "bbox": [ + 203, + 592, + 481, + 624 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} I (x) \\sim \\mathcal {N} \\left(\\mu_ {l}, \\sigma_ {l}\\right), \\quad l \\in L, \\\\ \\mu_ {l} \\sim \\mathcal {N} \\left(0, 1 \\mid \\theta_ {\\mu}, \\theta_ {l}\\right), \\sigma_ {l} \\sim \\mathcal {N} \\left(0, 1 \\mid \\theta_ {\\sigma}, \\theta_ {l}\\right), \\end{array} \\right. \\tag {2}", + "image_path": "c6bc7ecdffd081d8040d484bf4db8ea79b204e224902aa390656ba39267f15dc.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mu_{l}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\sigma_{l}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " refer to the mean and standard deviation of each segmentation label " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": ", and are independently sampled from Gaussian distributions at each voxel, with " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\theta_{l}, \\theta_{\\mu}, \\theta_{\\sigma} \\in \\Theta" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " controlling the shifts and scales of their values." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "P. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 479, + 222 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 479, + 222 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 479, + 222 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 479, + 222 + ], + "type": "image", + "image_path": "2b3785bbaec27ab54c9bbb8b9bf80ed99204e9da07ebd41278ea716e6d1c6fe1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 231, + 476, + 243 + ], + "lines": [ + { + "bbox": [ + 135, + 231, + 476, + 243 + ], + "spans": [ + { + "bbox": [ + 135, + 231, + 476, + 243 + ], + "type": "text", + "content": "Fig. 3: Brain-ID's contrast-agnostic anatomical representation learning framework." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 266, + 479, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 479, + 313 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 479, + 313 + ], + "type": "text", + "content": "Resolution Simulation and Data Corruption Given a deformed, contrast-synthesized image " + }, + { + "bbox": [ + 130, + 266, + 479, + 313 + ], + "type": "inline_equation", + "content": "(I)" + }, + { + "bbox": [ + 130, + 266, + 479, + 313 + ], + "type": "text", + "content": ", we adopt the standard data corruption pipeline [27], which further augments images with different levels of resolution and scanning artifacts that are commonly found in real-world clinical protocols." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 314, + 480, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 480, + 385 + ], + "type": "text", + "content": "As illustrated in Fig. 2, we are able to generate an infinite number of variations from a single subject with its unique brain anatomy. Generating images with randomized contrast/resolution/orientation on the fly for each subject enormously enriches the learning space for a robust representation, and helps focus the learning on intrinsic subject-specific features rather than superficial aspects of images that depend on acquisition parameters and conditions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 403, + 446, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 403, + 446, + 415 + ], + "spans": [ + { + "bbox": [ + 131, + 403, + 446, + 415 + ], + "type": "text", + "content": "3.2 Extract Robust and Expressive Subject-specific Features" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 423, + 479, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 479, + 459 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 479, + 459 + ], + "type": "text", + "content": "As mentioned above, we would like Brain-ID features to be robust to intra-subject variations and expressive to potential downstream tasks. In this section, we introduce Brain-ID's learning framework to achieve the two goals." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 462, + 480, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 480, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 480, + 605 + ], + "type": "text", + "content": "Intra-subject Data Generation In order to learn a feature representation that is distinctive to each subject and robust to varying MR contrasts, we observe that enriching intra-subject samples leads to better performance (Tab. 3). Specifically, instead of including multiple subjects for a mini-batch during each training iteration as in usual practice, Brain-ID focuses on maximizing the intra-subject variance to improve the subject-robustness of resulting features. As described in Alg. 1 (line 4-10), for each training iteration, after randomly selecting a subject and generating its deformation, Brain-ID generates a mini-batch of intra-subject samples " + }, + { + "bbox": [ + 130, + 462, + 480, + 605 + ], + "type": "inline_equation", + "content": "\\{(S_1,\\ldots ,S_N)\\}" + }, + { + "bbox": [ + 130, + 462, + 480, + 605 + ], + "type": "text", + "content": " with randomly synthesized contrasts, resolutions and corruptions (Sec. 3.1). As will be introduced below, Brain-ID collects losses from all intra-subject samples and conducts back-propagation at once, to encourage the subject-specific robustness of its learned features." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 606, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 480, + 666 + ], + "type": "text", + "content": "We set the intra-subject samples within a mini-batch to have random contrasts and \"mild-to-severe\", increasing level of corruptions (Fig. 3 (left)), to maximize the intra-subject variance while ensuring the stability of the training process against extreme corruption levels. (In Sec. 4.5, we compare various data generation designs, and provide insights on preventing unstable training.)" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "Brain-ID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "Anatomy-guided Feature Representation A richer learning space helps improve the representation robustness against the variance of sample appearances, but not enough for extracting expressive features. For example, a mapping that projects all inputs to zero is perfectly robust, but does not provide useful information for potential downstream tasks. Therefore, proper guidance is crucial for feature representation as well. In CIFL [15], the authors use a contrastive loss on feature channels to encourage a more discriminative feature representation. However, this method is insufficient, given the extremely complex nature of the human brain anatomy (See comparisons in Sec. 4). Instead, we propose to use the standard T1w high-resolution structural MR contrast for brain morphometry, i.e., MP-RAGE (magnetization-prepared rapid gradient-echo), as the unique anatomy target to guide Brain-ID's feature representation learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 259, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 331 + ], + "type": "text", + "content": "As shown in Fig. 3, a feature extraction backbone " + }, + { + "bbox": [ + 130, + 259, + 482, + 331 + ], + "type": "inline_equation", + "content": "(\\mathcal{F})" + }, + { + "bbox": [ + 130, + 259, + 482, + 331 + ], + "type": "text", + "content": " first maps the input mini-batch of intra-subject generated samples, " + }, + { + "bbox": [ + 130, + 259, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\{S_1,\\dots,S_N\\}" + }, + { + "bbox": [ + 130, + 259, + 482, + 331 + ], + "type": "text", + "content": ", to their corresponding feature space, " + }, + { + "bbox": [ + 130, + 259, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}_1,\\ldots ,\\mathbf{F}_N\\}" + }, + { + "bbox": [ + 130, + 259, + 482, + 331 + ], + "type": "text", + "content": ". A linear activation layer " + }, + { + "bbox": [ + 130, + 259, + 482, + 331 + ], + "type": "inline_equation", + "content": "(\\mathcal{L})" + }, + { + "bbox": [ + 130, + 259, + 482, + 331 + ], + "type": "text", + "content": " then projects the features to the current subject's standard contrast (MP-RAGE) space, " + }, + { + "bbox": [ + 130, + 259, + 482, + 331 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 259, + 482, + 331 + ], + "type": "text", + "content": ". The training loss is obtained by summing over the anatomy reconstruction loss of all intra-subject samples in the current mini-batch:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 192, + 338, + 480, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 338, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 192, + 338, + 480, + 358 + ], + "type": "interline_equation", + "content": "L = \\sum_ {i} ^ {N} | \\mathcal {L} (\\mathbf {F} _ {i}) - T | + \\lambda | \\nabla \\mathcal {L} (\\mathbf {F} _ {i}) - \\nabla T |, \\lambda \\in R ^ {+}, \\tag {3}", + "image_path": "90bbb657c29b573effc4872f4cad5173366a416922841ca03e09107bb749b449.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 363, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 363, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 131, + 363, + 482, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 363, + 482, + 449 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 363, + 482, + 449 + ], + "type": "text", + "content": ", as the high-resolution, unique anatomy target for all intra-subject samples, encourages the similarities of the features extracted from the same subject via reconstruction (" + }, + { + "bbox": [ + 131, + 363, + 482, + 449 + ], + "type": "inline_equation", + "content": "1^{\\mathrm{st}}" + }, + { + "bbox": [ + 131, + 363, + 482, + 449 + ], + "type": "text", + "content": " term) and gradient difference (" + }, + { + "bbox": [ + 131, + 363, + 482, + 449 + ], + "type": "inline_equation", + "content": "2^{\\mathrm{nd}}" + }, + { + "bbox": [ + 131, + 363, + 482, + 449 + ], + "type": "text", + "content": " term) L1 loss. Brain-ID uses MP-RAGE as the learning target, which formulates both super-resolution and contrast-synthesis problems, and encourages a richer feature representation. (Sec. 4.5 provides further insights on how different choices of anatomy guidance would affect the resulting features' properties.)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 463, + 437, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 463, + 437, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 463, + 437, + 475 + ], + "type": "text", + "content": "3.3 Adapting Brain-ID to Downstream Tasks by One Layer" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 480, + 482, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 482, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 482, + 577 + ], + "type": "text", + "content": "With the intra-subject data generation and anatomy-guided feature learning design introduced in Sec. 3.2, a well-trained Brain-ID model is able to extract robust, high-resolution anatomical features from images with varying deformations, contrasts, resolutions, and artifacts. To minimize the modifications, we propose straightforward one-layer solutions adapting Brain-ID features to various brain imaging applications. We later demonstrate that the simple adaptions are effective enough for Brain-ID to achieve state-of-the-art performance across all downstream tasks (Sec. 4.3), even for small datasets (Sec. 4.4)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 578, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 578, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 131, + 578, + 482, + 616 + ], + "type": "text", + "content": "Contrast-independent Tasks For tasks where the output should be independent of the input MR contrast, e.g., brain segmentation, we simply add an additional layer following Brain-ID features " + }, + { + "bbox": [ + 131, + 578, + 482, + 616 + ], + "type": "inline_equation", + "content": "(\\mathbf{F})" + }, + { + "bbox": [ + 131, + 578, + 482, + 616 + ], + "type": "text", + "content": ", and fine-tune the model:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 235, + 622, + 480, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 622, + 480, + 635 + ], + "spans": [ + { + "bbox": [ + 235, + 622, + 480, + 635 + ], + "type": "interline_equation", + "content": "L = \\operatorname {t a s k} _ {-} \\operatorname {l o s s} _ {-} \\operatorname {f u n c} \\left(\\mathcal {L} (\\mathbf {F}), T\\right), \\tag {4}", + "image_path": "f1fbdb5b0927faca4fa46b13eb24bc7fbc24f3016d087f18b3fb86be06c9d6dd.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " refer to the task-specific activation layer and the contrast-independent ground truth, respectively." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "text", + "content": "P. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "Contrast-dependent Tasks Since Brain-ID features are contrast-agnostic and robust to artifacts, for tasks relevant to the input's contrast/quality, e.g., super-resolution, we concatenate the input image " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "(I)" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " that contains the original contrast information with its high-resolution Brain-ID features along the channel dimension, before forwarding into the task-specific activation layer:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 227, + 186, + 479, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 186, + 479, + 198 + ], + "spans": [ + { + "bbox": [ + 227, + 186, + 479, + 198 + ], + "type": "interline_equation", + "content": "L = \\operatorname {t a s k} _ {-} \\operatorname {l o s s} _ {-} \\operatorname {f u n c} \\left(\\mathcal {L} (\\mathbf {F} \\oplus I), T\\right), \\tag {5}", + "image_path": "431f5025d200cb69aa05bc0647bfcbc1ba8e8cc781c5138a88cd080696fa9c21.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 206, + 477, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 206, + 477, + 218 + ], + "spans": [ + { + "bbox": [ + 131, + 206, + 477, + 218 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 206, + 477, + 218 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 131, + 206, + 477, + 218 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 206, + 477, + 218 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 206, + 477, + 218 + ], + "type": "text", + "content": " refer to the task-specific activation layer and the ground truth." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 236, + 229, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 236, + 229, + 249 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 229, + 249 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 260, + 479, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 479, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 479, + 332 + ], + "type": "text", + "content": "We demonstrate the two properties of Brain-ID as claimed in Sec. 3. (i) Robustness (Sec. 4.2): we propose \"canonical\" and \"atlas-registered\" features to assess the intra/inter-subject feature robustness; (ii) Expressiveness (Sec. 4.3): we evaluate the performance of adapting Brain-ID features to various brain imaging applications. We further challenge Brain-ID with reduced-size datasets to explore its ability when limited real data is available for training (Sec. 4.4)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 350, + 394, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 350, + 394, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 350, + 394, + 361 + ], + "type": "text", + "content": "4.1 Datasets, Metrics and Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 372, + 479, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 372, + 479, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 479, + 419 + ], + "type": "text", + "content": "Datasets For Brain-ID features pre-training, we use 2045 3D segmentations (\"anatomies\", or IDs) from the public ADNI dataset [30]. For downstream evaluation, we use six public datasets covering T1w and T2w MRI, FLAIR MRI, and CT: ADNI [30], ADNI3 [60], HCP [19], ADHD200 [9], AIBL [23], OASIS3 [35]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 422, + 479, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 479, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 479, + 493 + ], + "type": "text", + "content": "Metrics We evaluate individual tasks from different aspects. For feature similarity measurements, we use L1 distance, and (MS-)SSIM (multi-scale structural similarity) [41,58,59]. For reconstruction and super-resolution, we use L1, PSNR (peak signal-to-noise ratio) and (MS-)SSIM. For segmentation, we use Dice scores. For bias field estimation, we use the normalized L2 distance (norm-L2) to avoid possible arbitrary scalings from nonuniformity correction [16]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 498, + 479, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 479, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 479, + 568 + ], + "type": "text", + "content": "Implementation Details As a general feature representation model, Brain-ID can use any backbone to extract brain features. For fairer comparison, we adopt the same five-level 3D UNet [47] (with 64 feature channels in the last layer) as utilized in state-of-the-art models we compare with in Sec. 4.3. During feature pre-training (Sec. 3.2), a linear regression layer is added for anatomy supervision. For downstream adaptions, A task-specific adaption layer is added (Sec. 3.3)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 586, + 359, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 359, + 597 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 359, + 597 + ], + "type": "text", + "content": "4.2 Intra/Inter-subject Feature Robustness" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": "In this section, we examine the robustness of Brain-ID features, i.e., the 64 features from the last layer. For comparable and reproducible results, we use our data generator (Sec. 3.1) to prepare (deform and corrupt) 1000 testing samples from 100 randomly selected subjects in ADNI [30] (T1w) testing set, with 10 intra-subject samples for each subject." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "type": "text", + "content": "Brain-ID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 478, + 421 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 478, + 421 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 478, + 421 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 478, + 421 + ], + "type": "image", + "image_path": "437282f39cfc164e1f6371617cc3902ae0ce5b40c0baadf767663eab7eabf17b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 430, + 468, + 441 + ], + "lines": [ + { + "bbox": [ + 144, + 430, + 468, + 441 + ], + "spans": [ + { + "bbox": [ + 144, + 430, + 468, + 441 + ], + "type": "text", + "content": "Fig. 4: (a) Intra-subject and (b) inter-subject robustness of Brain-ID features." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "text", + "content": "We compare our features with CIFL [15] which is, to our best knowledge, the only similar work to Brain-ID, that also aims to learn contrast-agnostic brain features. Note the original CIFL method has only experimented on 2D images, and does not have our intra-subject data generation design. For a fairer comparison, we trained CIFL with its contrastive-learning design, while using the same data generator and network architecture as Brain-ID." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 550, + 482, + 597 + ], + "type": "text", + "content": "Intra-subject Feature Robustness Ideally, Brain-ID features computed from the same subject should be structurally identical, regardless of their appearances (contrast, resolution, noises, etc). To assess the intra-subject feature robustness, we first define their corresponding \"canonical\" features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 601, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 482, + 638 + ], + "type": "text", + "content": "\"Canonical\" Features. We compute " + }, + { + "bbox": [ + 130, + 601, + 482, + 638 + ], + "type": "inline_equation", + "content": "\\phi^{-1}" + }, + { + "bbox": [ + 130, + 601, + 482, + 638 + ], + "type": "text", + "content": ", the inverse of " + }, + { + "bbox": [ + 130, + 601, + 482, + 638 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 601, + 482, + 638 + ], + "type": "text", + "content": "'s generated deformation as in Eq. (1), and map back each sample's " + }, + { + "bbox": [ + 130, + 601, + 482, + 638 + ], + "type": "inline_equation", + "content": "(S)" + }, + { + "bbox": [ + 130, + 601, + 482, + 638 + ], + "type": "text", + "content": " Brain-ID features " + }, + { + "bbox": [ + 130, + 601, + 482, + 638 + ], + "type": "inline_equation", + "content": "(\\mathbf{F} = \\mathbf{F}(S|\\phi))" + }, + { + "bbox": [ + 130, + 601, + 482, + 638 + ], + "type": "text", + "content": " to their original domain without any deformation:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 256, + 651, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 651, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 256, + 651, + 481, + 666 + ], + "type": "interline_equation", + "content": "\\mathbf {F} ^ {- 1} = \\phi^ {- 1} \\circ \\mathbf {F} (S \\mid \\phi). \\tag {6}", + "image_path": "c73a523bc763e65a8f9b947ca7b43b7e1ef5bedeb24c044e09f97cd8e68f6e62.jpg" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "P. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 195, + 134, + 418, + 192 + ], + "blocks": [ + { + "bbox": [ + 176, + 114, + 435, + 125 + ], + "lines": [ + { + "bbox": [ + 176, + 114, + 435, + 125 + ], + "spans": [ + { + "bbox": [ + 176, + 114, + 435, + 125 + ], + "type": "text", + "content": "Table 1: Evaluations on intra/inter-subject feature robustness." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 195, + 134, + 418, + 192 + ], + "lines": [ + { + "bbox": [ + 195, + 134, + 418, + 192 + ], + "spans": [ + { + "bbox": [ + 195, + 134, + 418, + 192 + ], + "type": "table", + "html": "
ModeMethodL1 (↓)SSIM (↑)MS-SSIM (↑)
IntraCIFL [15]0.122 (±0.031)0.511 (±0.298)0.531 (±0.367)
Brain-ID0.011 (±0.001)0.858 (±0.041)0.921 (±0.008)
InterCIFL [15]0.230 (±0.039)0.552 (±0.240)0.524 (±0.291)
Brain-ID0.014 (±0.013)0.843 (±0.044)0.913 (±0.011)
", + "image_path": "9f06611b63ae01e3351e07b97af6e05ca298bb8b01e247087c83acbae2d157ee.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 211, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 211, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 482, + 319 + ], + "type": "text", + "content": "As shown in Fig. 4-(a), for the same subject (ID-i), although the inputs (row) are processed by varying deformations and corruptions, their Brain-ID's \"canonical\" features are of similar looking along each feature channel (column). In Tab. 1, we quantify the feature similarity: For each testing subject, we first generate a sample without any deformation or corruption and use its Brain-ID feature as the gold-standard reference for all other \"canonical\" features computed from intra-subject samples that may have different levels of deformation and corruptions. The final results are obtained by scores averaged over all 64 feature channels of all the intra-subject samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 322, + 482, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 482, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 482, + 382 + ], + "type": "text", + "content": "Inter-subject Feature Robustness We further assess the feature robustness among anatomies, based on the assumption that features from different subjects should ideally be structurally similar, once being registered to a common domain. \"Atlas-registered\" Features. We register [43] features " + }, + { + "bbox": [ + 130, + 322, + 482, + 382 + ], + "type": "inline_equation", + "content": "(\\mathbf{F} = \\mathbf{F}(S|\\phi))" + }, + { + "bbox": [ + 130, + 322, + 482, + 382 + ], + "type": "text", + "content": " from all subjects to a common brain atlas [21, 22]:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 257, + 389, + 481, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 389, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 257, + 389, + 481, + 403 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathbf {F}}} = \\psi_ {S \\rightarrow A} \\circ \\mathbf {F} (S \\mid \\phi), \\tag {7}", + "image_path": "5ac0bee6ce9847b83eaea8b2cbed3294fe454f2af6160384cbeb345644308f17.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 411, + 387, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 387, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 387, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 411, + 387, + 422 + ], + "type": "inline_equation", + "content": "\\psi_{S\\to A}" + }, + { + "bbox": [ + 130, + 411, + 387, + 422 + ], + "type": "text", + "content": " denotes the mapping from " + }, + { + "bbox": [ + 130, + 411, + 387, + 422 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 411, + 387, + 422 + ], + "type": "text", + "content": " to the atlas, " + }, + { + "bbox": [ + 130, + 411, + 387, + 422 + ], + "type": "inline_equation", + "content": "A" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 423, + 482, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 543 + ], + "type": "text", + "content": "As shown in Fig. 4-(b), even though the input samples (row) from different subjects have varying anatomical structures, after being registered to a common domain, their \"atlas-registered\" Brain-ID features still appear consistent (column). In Tab. 1, we further provide quantitative results of the structural similarity measured between features from different subjects. Similar to Sec. 4.2, we randomly select a testing subject and generate a sample without any deformation or corruption. We use its \"atlas-registered\" Brain-ID feature as the gold standard for all \"atlas-registered\" features computed from other samples that may have different levels of deformation and corruption. The final results are obtained by scores averaged over all samples of the 100 testing subjects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 543, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 543, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 543, + 482, + 579 + ], + "type": "text", + "content": "In both intra/inter-subject aspects, Brain-ID outperforms CIFL by a large margin, which validates its feature robustness against deformations, contrasts, resolutions, artifacts (intra-subject), and anatomies (inter-subject)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 595, + 281, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 595, + 281, + 607 + ], + "spans": [ + { + "bbox": [ + 131, + 595, + 281, + 607 + ], + "type": "text", + "content": "4.3 Downstream Evaluation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 614, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 614, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 614, + 482, + 667 + ], + "type": "text", + "content": "Following the standard evaluation protocol for representation learning [12], we fine-tune Brain-ID features for downstream tasks through one-layer (Sec. 3.3). Models We compare the downstream performance of Brain-ID features with: " + }, + { + "bbox": [ + 130, + 614, + 482, + 667 + ], + "type": "inline_equation", + "content": "(i - ii)" + }, + { + "bbox": [ + 130, + 614, + 482, + 667 + ], + "type": "text", + "content": " SAMSEG [13,46] and FastSurfer [24] (only works on T1w), state-of-the-art" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "type": "text", + "content": "Brain-ID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 116, + 475, + 277 + ], + "blocks": [ + { + "bbox": [ + 138, + 116, + 475, + 277 + ], + "lines": [ + { + "bbox": [ + 138, + 116, + 475, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 475, + 277 + ], + "type": "image", + "image_path": "7c5afb057c19828fd2a976c65d9e9ba4b880faf378ece7de0b0fdaa45c68eecf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 288, + 482, + 344 + ], + "lines": [ + { + "bbox": [ + 130, + 288, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 482, + 344 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparisons on downstream tasks of reconstruction (Recon), superresolution (SR), and segmentation (Seg), between Brain-ID, the baseline SCRATCH, and the state-of-the-art methods CIFL [15], SynthSR [27] (Recon and SR), SAMSEG [13] (Seg). The visualized testing examples are taken from: AIBL-FLAIR [23] for Recon, AIBLT1w [23] for SR, and OASIS-CT [35] for Seg. The mint circles highlight some details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 370, + 482, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 370, + 482, + 466 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 482, + 466 + ], + "type": "text", + "content": "classical and machine-learning-based brain segmentation models, respectively; (iii) SynthSR [27], state-of-the-art, contrast-agnostic T1w synthesis model. We also provide fine-tuned SynthSR (SynthSR-FT) results for further comparison; (iv) SCRAW: a baseline with the same data generation/architecture as Brain-ID, yet trained from scratch. SCRAW is used to validate Brain-ID's effectiveness. (v) CIFL [15]: a baseline with the same data generation/architecture as Brain-ID and SCRAW, yet initialized with CIFL's pre-trained features. We use CIFL to demonstrate the superiority of Brain-ID's anatomy-guided representation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 470, + 311, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 470, + 311, + 482 + ], + "spans": [ + { + "bbox": [ + 131, + 470, + 311, + 482 + ], + "type": "text", + "content": "Contrast-independent Applications" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 483, + 482, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 543 + ], + "type": "text", + "content": "- Anatomy Reconstruction/Contrast Synthesis As one of the most common types of MRI scans, T1w MR images highlight the differences between gray and white matter, and are mostly often utilized to image the anatomy of the brain, spinal cord, bones and joints [51]. For each dataset, we train models to reconstruct their T1w MRI counterparts with the L1 loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 544, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 593 + ], + "type": "text", + "content": "- Brain Segmentation For each dataset, we use SynthSeg [4] to obtain the segmentation labels with 30 brain anatomical regions [20], as the gold standard segmentation target. We train our compared models to predict the brain segmentation labels, with the soft dice loss and cross-entropy loss [4]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "As shown in Fig. 5, Brain-ID clearly outperforms CIFL and SCRATCH in the reconstruction task. As highlighted by the mint circles, it reveals the anatomy details and produces more fine-grained results than the strong state-of-the-art method SynthSR [27]. Brain-ID also achieves better segmentation results, especially in the smaller and more challenging regions. Quantitative comparisons are in Tab. 2, where Brain-ID obtains the best scores across most metrics. No-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "P. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 146, + 145, + 467, + 490 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 2: Comparisons of Brain-ID with state-of-the-art approaches on downstream applications. (\"train/test\" refers to the number of subjects in the training/testing set.)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 146, + 145, + 467, + 490 + ], + "lines": [ + { + "bbox": [ + 146, + 145, + 467, + 490 + ], + "spans": [ + { + "bbox": [ + 146, + 145, + 467, + 490 + ], + "type": "table", + "html": "
Modality (Contrast)Dataset (Train/Test)MethodReconstructionSegmentationSuper-resolution (from 3 / 5 / 7 mm)Bias Field norm-12 (↓)
L1 (↓)PSNR (↑)SSIM (↑)MS-SSIM (↑)Dice (↑)PSNR (↑)SSIM (↑)MS-SSIM (↑)MS-SSIM (↑)
MR (T1w)ADNI [30] (1841/204)SAMSEG [13]---0.795------
FastSurfer [24]---0.803------
SynthSR [27]0.01426.780.9800.980-27.01 / 25.81 / 21.790.977 / 0.943 / 0.8570.981 / 0.968 / 0.869--
SynthSR-FT0.01029.540.9840.9200.82530.46 / 28.03 / 23.910.979 / 0.960 / 0.8980.985 / 0.973 / 0.900-0.051
CIFL [15]0.01326.970.9730.9610.82030.12 / 27.21 / 23.800.978 / 0.912 / 0.8870.979 / 0.950 / 0.842-0.053
SCRATCH0.01127.540.9790.9570.81630.29 / 27.94 / 24.570.975 / 0.959 / 0.9200.982 / 0.964 / 0.866-0.056
Brain-ID0.01233.820.9930.9890.83731.30 / 28.62 / 24.680.983 / 0.961 / 0.9280.987 / 0.980 / 0.947-0.048
MR (T1w)HCP [19] (808/87)SAMSEG [13]---0.810------
FastSurfer [24]---0.819------
SynthSR [27]0.03322.130.8540.901-22.21 / 20.83 / 19.930.848 / 0.802 / 0.7470.899 / 0.864 / 0.789--
SynthSR-FT0.02527.110.9350.9170.81025.90 / 25.45 / 25.100.889 / 0.872 / 0.8310.919 / 0.911 / 0.889-0.052
CIFL [15]0.02926.420.9320.9130.80425.98 / 25.11 / 23.990.930 / 0.883 / 0.8710.936 / 0.927 / 0.880-0.059
SCRATCH0.02327.320.9230.9090.79226.69 / 26.41 / 25.050.932 / 0.896 / 0.8480.942 / 0.930 / 0.892-0.069
Brain-ID0.02027.470.9570.9290.84329.70 / 27.90 / 26.120.957 / 0.901 / 0.8830.973 / 0.938 / 0.890-0.047
SAMSEG [13]---0.769------
FastSurfer [24]---0.796------
SynthSR [27]0.02323.600.9280.909-23.13 / 22.40 / 22.270.921 / 0.907 / 0.8760.903 / 0.892 / 0.871--
SynthSR-FT0.02127.110.9500.9690.82027.95 / 27.34 / 26.180.920 / 0.893 / 0.9220.890 / 0.890 / 0.879-0.134
CIFL [15]0.02828.520.9610.9700.81927.81 / 27.10 / 26.320.900 / 0.877 / 0.8640.812 / 0.793 / 0.769-0.102
SCRATCH0.03327.280.9570.9630.81627.99 / 27.37 / 26.630.893 / 0.889 / 0.8760.794 / 0.785 / 0.753-0.128
Brain-ID0.02129.890.9660.9760.84330.17 / 28.23 / 26.890.962 / 0.923 / 0.8920.973 / 0.921 / 0.883-0.093
SAMSEG [13]---0.784------
FastSurfer [24]---0.801------
SynthSR [27]0.03521.670.8820.853-21.42 / 21.13 / 21.410.876 / 0.846 / 0.8090.851 / 0.831 / 0.805--
SynthSR-FT0.01626.900.9170.9050.81228.55 / 27.43 / 25.940.880 / 0.859 / 0.8230.860 / 0.845 / 0.811-0.110
CIFL [15]0.01328.690.9320.9210.81028.11 / 26.12 / 25.110.865 / 0.848 / 0.8310.828 / 0.800 / 0.789-0.109
SCRATCH0.01131.550.9860.9850.79629.41 / 27.89 / 26.580.854 / 0.847 / 0.8140.853 / 0.845 / 0.827-0.113
Brain-ID0.01132.480.9960.9960.84729.50 / 28.56 / 26.870.898 / 0.862 / 0.8110.887 / 0.850 / 0.823-0.107
SAMSEG [13]---0.799------
FastSurfer [24]---0.802------
SynthSR [27]0.02622.950.9160.912-22.57 / 21.82 / 21.780.917 / 0.892 / 0.8560.906 / 0.893 / 0.869--
SynthSR-FT0.01429.890.9410.9220.81029.15 / 27.89 / 26.770.932 / 0.906 / 0.8780.927 / 0.900 / 0.871-0.118
CIFL [15]0.01130.120.9380.9250.81629.46 / 27.97 / 26.730.907 / 0.889 / 0.8700.891 / 0.872 / 0.857-0.094
SCRATCH0.01130.530.9320.9130.81429.43 / 28.10 / 26.820.900 / 0.886 / 0.8630.884 / 0.870 / 0.853-0.128
Brain-ID0.00933.730.9720.9630.85129.75 / 28.58 / 27.090.957 / 0.922 / 0.8900.975 / 0.956 / 0.937-0.088
SAMSEG [13]---0.782------
FastSurfer [24]---0.805------
SynthSR [27]0.02826.100.8940.8980.75528.43 / 25.36 / 22.100.942 / 0.851 / 0.8010.977 / 0.892 / 0.798-0.140
SynthSR-FT0.03625.120.8910.8790.78728.26 / 25.78 / 21.730.959 / 0.893 / 0.7970.982 / 0.906 / 0.742-0.138
SCRATCH0.03824.990.8730.8660.75628.12 / 24.52 / 21.580.945 / 0.863 / 0.7830.980 / 0.893 / 0.738-0.136
Brain-ID0.01628.100.9340.9350.80530.26 / 26.11 / 24.100.959 / 0.902 / 0.8320.987 / 0.953 / 0.912-0.136
SAMSEG [13]---0.763------
FastSurfer [24]---0.801------
SynthSR [27]0.03320.080.8050.831------
SynthSR-FT0.02422.930.8150.8400.71930.15 / 28.09 / 26.920.940 / 0.912 / 0.8930.966 / 0.948 / 0.922-0.200
CIFL [15]0.02223.710.8200.8390.72129.98 / 27.01 / 25.660.931 / 0.899 / 0.8610.967 / 0.941 / 0.905-0.193
SCRATCH0.02022.270.8490.8510.71431.91 / 29.20 / 27.350.954 / 0.934 / 0.8960.982 / 0.969 / 0.949 / 0.936-0.197
Brain-ID0.02223.990.8610.8500.78232.26 / 29.90 / 27.090.937 / 0.930 / 0.9080.988 / 0.971 / 0.948-0.148
SAMSEG [13]---0.718------
SynthSR [27]0.02622.770.9190.895-30.11 / 28.70 / 25.320.930 / 0.899 / 0.8640.948 / 0.910 / 0.878-0.251
SynthSR-FT0.02123.340.9210.9000.75332.72 / 29.00 / 26.990.949 / 0.906 / 0.8730.953 / 0.919 / 0.889-0.237
CIFL [15]0.02021.290.9110.8970.76132.72 / 28.71 / 25.860.945 / 0.915 / 0.8770.941 / 0.936 / 0.917-0.136
SCRATCH0.02520.800.9010.8620.75932.36 / 28.71 / 25.840.945 / 0.915 / 0.8770.981 / 0.953 / 0.917-0.236
Brain-ID0.01726.440.9270.8920.78632.18 / 30.00 / 28.280.959 / 0.921 / 0.8830.982 / 0.965 / 0.921-0.227
SAMSEG [13]---0.710------
FastSurfer [24]---0.795------
SynthSR [27]0.02921.770.9020.892------
SynthSR-FT0.02425.800.9140.9000.73526.20 / 24.91 / 23.170.860 / 0.832 / 0.7950.933 / 0.910 / 0.894-0.126
CIFL [15]0.02627.110.8910.8700.72127.72 / 26.85 / 25.950.821 / 0.858 / 0.8140.969 / 0.939 / 0.905 / 0.901-0.109
SCRATCH0.02326.840.8980.8670.72028.02 / 26.52 / 25.430.878 / 0.838 / 0.7990.943 / 0.901 / 0.891-0.115
Brain-ID0.01927.250.9360.9120.76726.69 / 24.01 / 23.69 / 23.09 / 0.918 / 0.806 / 0.966 / 0.871 / 0.925 / 0.916---
", + "image_path": "3be05f652aedbb23a49b7a62b3f73f8b9196c33df03b14fbe221da40b3b28502.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 147, + 491, + 425, + 496 + ], + "lines": [ + { + "bbox": [ + 147, + 491, + 425, + 496 + ], + "spans": [ + { + "bbox": [ + 147, + 491, + 425, + 496 + ], + "type": "text", + "content": "(1) \"means not applicable: FastSurfer [24] and SAMSEG [13] are for segmentation-only; SynthSR [27] only synthesize T1w MRI; CT does not have bias fields. (2) SynthSR ET; SynthSR fine tuned on each task/dataset, for a fairer comparison." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 532, + 479, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 479, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 479, + 568 + ], + "type": "text", + "content": "tably, Brain-ID achieves greater gains over baseline models, SCRATCH and CIFL, particularly on smaller datasets (e.g., ADNI3-T1/FLAIR, AIBL-FLAIR, where Brain-ID's robust and rich features quickly adapt to specific tasks/datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 576, + 479, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 479, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 479, + 601 + ], + "type": "text", + "content": "Contrast-dependent Applications We also evaluate Brain-ID on two brain imaging tasks that are dependent on the input modality/quality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 605, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 479, + 666 + ], + "type": "text", + "content": "- Image Super-resolution We use the standard 1mm isotropic images from all datasets as the ground truth high-resolution target, and follow the strategy in SynthSR [27] where the input samples are randomly resampled and corrupted during training. For inference, we downsample the original images into 3mm, 5mm, and 7mm isotropic images as inputs, and evaluate the output quality." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "Brain-ID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 146, + 115, + 471, + 286 + ], + "blocks": [ + { + "bbox": [ + 146, + 115, + 471, + 286 + ], + "lines": [ + { + "bbox": [ + 146, + 115, + 471, + 286 + ], + "spans": [ + { + "bbox": [ + 146, + 115, + 471, + 286 + ], + "type": "image", + "image_path": "fbd526bb28e75ae9e993bbda0e2a7d4230d8705df90091974255f28ff632932f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 295, + 482, + 329 + ], + "lines": [ + { + "bbox": [ + 131, + 295, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 131, + 295, + 482, + 329 + ], + "type": "text", + "content": "Fig. 6: SCRATCH, which is well trained on HF T1w scans, produces highly descriptive features for HF T1w images (" + }, + { + "bbox": [ + 131, + 295, + 482, + 329 + ], + "type": "inline_equation", + "content": "1^{\\mathrm{st}}" + }, + { + "bbox": [ + 131, + 295, + 482, + 329 + ], + "type": "text", + "content": " row), but does not preserve the same high quality useful for downstream tasks when handling LF (" + }, + { + "bbox": [ + 131, + 295, + 482, + 329 + ], + "type": "inline_equation", + "content": "2^{\\mathrm{nd}}" + }, + { + "bbox": [ + 131, + 295, + 482, + 329 + ], + "type": "text", + "content": " row) or other contrasts (" + }, + { + "bbox": [ + 131, + 295, + 482, + 329 + ], + "type": "inline_equation", + "content": "3^{\\mathrm{rd}}" + }, + { + "bbox": [ + 131, + 295, + 482, + 329 + ], + "type": "text", + "content": " row)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 351, + 482, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 435 + ], + "type": "text", + "content": "- Bias Field Estimation The bias field is a smooth, low-frequency multiplicative signal that corrupts MRI images, which affects image analysis tasks such as segmentation or texture analysis [31]. Bias field estimation is often needed as a pre-processing step to correct corrupted MRI images [16]. We apply randomly simulated bias fields to the input samples (Sec. 3.1), and train all models with the L2 loss. During inference, we pre-generate and apply the bias fields on the testing data for reproducibility, and evaluate the bias field estimation performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 435, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 532 + ], + "type": "text", + "content": "With a simple one-layer adaption (Sec. 3.3), Brain-ID's contrast-agnostic anatomical features achieve state-of-the-art performance on contrast-dependent tasks (Tab. 2, Fig. 5). As shown in Fig. 6, Brain-ID obtains contrast/resolution-robustness that cannot be achieved by models trained from real images (due to the limited variability in their appearance), regardless of the backbone choices. Such robust features greatly improve super-resolution, resulting in higher (MS-)SSIM scores. Brain-ID's gains are less obvious for bias field estimation, probably because the bias field is approximately independent of brain anatomy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 548, + 343, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 343, + 560 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 343, + 560 + ], + "type": "text", + "content": "4.4 Practical Value and Broader Impact" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 567, + 481, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 481, + 614 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 481, + 614 + ], + "type": "text", + "content": "Low-resolution data Trained on synthetic images, Brain-ID features are not only robust to various contrasts/modalities, but also extremely robust to low-resolution data (Fig. 6) and provide huge potential in clinical MRI (including big retrospective data [5]) and portable low-field MRI [29]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "Small-size Datasets To investigate the effectiveness of Brain-ID features for limited-size datasets, we assess its performance across all four downstream tasks on subsets of ADNI3 [60] training set (298 cases originally). As shown in Fig. 7, we reduce the training set size percentage from " + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": " to an extreme " + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": ". With" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "P. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 471, + 367 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 471, + 367 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 471, + 367 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 471, + 367 + ], + "type": "image", + "image_path": "86e67cf7953cd9c059f570d091c039fe9b914201ffba548255a97a1d818ec3ae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 369, + 302, + 378 + ], + "lines": [ + { + "bbox": [ + 246, + 369, + 302, + 378 + ], + "spans": [ + { + "bbox": [ + 246, + 369, + 302, + 378 + ], + "type": "text", + "content": "(c) Super-resolution" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 369, + 369, + 445, + 377 + ], + "lines": [ + { + "bbox": [ + 369, + 369, + 445, + 377 + ], + "spans": [ + { + "bbox": [ + 369, + 369, + 445, + 377 + ], + "type": "text", + "content": "(d) Bias Field Estimation" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 388, + 482, + 422 + ], + "lines": [ + { + "bbox": [ + 131, + 388, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 131, + 388, + 482, + 422 + ], + "type": "text", + "content": "Fig. 7: Adapting Brain-ID to small datasets. The horizontal (vertical) axes indicate training epochs (evaluation scores of corresponding tasks). Results are obtained by evaluating models collected throughout the epochs, on ADNI3 [60] full testing set." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "text", + "content": "only " + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "text", + "content": " of the data (mint line), Brain-ID still achieves better performance compared to its full-sized baseline, “" + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "text", + "content": "” SCRATCH (yellow line), over all tasks. 30 training samples (" + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 445, + 482, + 506 + ], + "type": "text", + "content": ", blue line) may be at the edge of obtaining acceptable results for Brain-ID, as the model becomes less stable and the performance drops occur more frequently – especially for brain segmentation (Fig. 7-(b))." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 523, + 281, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 523, + 281, + 536 + ], + "spans": [ + { + "bbox": [ + 131, + 523, + 281, + 536 + ], + "type": "text", + "content": "4.5 Additional Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 543, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 543, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 543, + 482, + 615 + ], + "type": "text", + "content": "Choice of Anatomy Guidance We explore other anatomical learning targets than MP-RAGE (Sec. 3.2): " + }, + { + "bbox": [ + 130, + 543, + 482, + 615 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 130, + 543, + 482, + 615 + ], + "type": "text", + "content": " segmentation labels; " + }, + { + "bbox": [ + 130, + 543, + 482, + 615 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 130, + 543, + 482, + 615 + ], + "type": "text", + "content": " both segmentation labels and MP-RAGE. As shown in Tab. 3, adding segmentation guidance improves robustness, yet inhibits the feature expressiveness and affects the downstream performance; we observe features from segmentation guidance produce less high-frequency texture than Brain-ID, especially within the same-label regions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 485, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 485, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 485, + 666 + ], + "type": "text", + "content": "Data Generation Design As shown in Tab. 3, training with all mildly-corrupted samples results in reduced robustness and harms the downstream performance, yet using all extremely corrupted samples leads to unstable training and model collapse. Brain-ID uses samples of gradually increased corruptions (Fig. 3 (left))." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 100 + ], + "type": "text", + "content": "Brain-ID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 177, + 134, + 436, + 262 + ], + "blocks": [ + { + "bbox": [ + 188, + 114, + 424, + 125 + ], + "lines": [ + { + "bbox": [ + 188, + 114, + 424, + 125 + ], + "spans": [ + { + "bbox": [ + 188, + 114, + 424, + 125 + ], + "type": "text", + "content": "Table 3: Comparison between Brain-ID and its variants." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 177, + 134, + 436, + 262 + ], + "lines": [ + { + "bbox": [ + 177, + 134, + 436, + 262 + ], + "spans": [ + { + "bbox": [ + 177, + 134, + 436, + 262 + ], + "type": "table", + "html": "
Model Setup [Comparison Target]Feature Robustness (Intra)Downstream (Reconstruction)
SSIM (↑)MS-SSIM (↑)L1 (↓)PSNR (↑)SSIM (↑)
[Representation guidance: supervision target for feature learning]
Segmentation0.8910.9630.02928.130.958
Segmentation + MP-RAGE0.8630.9400.02329.050.964
MP-RAGE (Brain-ID)0.8580.9210.02129.890.966
[Data generation design: corruption levels in intra-subject mini-batches]
All Mild (\\( \\sigma_{noise} = 1, \\cdots \\))0.7920.8130.02828.300.960
All Medium (\\( \\sigma_{noise} = 5, \\cdots \\))0.8310.8990.02229.670.964
All Severe (\\( \\sigma_{noise} = 10, \\cdots \\))N/AN/AN/AN/AN/A
Mild to Severe (Brain-ID)0.8580.9210.02129.890.966
[Mini-batch size: number of intra-subject samples]
20.8260.8970.02528.830.959
30.8410.9020.02229.300.962
4 (Brain-ID)0.8580.9210.02129.890.966
50.8600.9290.02129.930.965
", + "image_path": "409e8ef00ccd4cc441473b773a196e4945d01ec1e8f15aefcfbd8d6320a64534.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 283, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 319 + ], + "type": "text", + "content": "Batch Size Brain-ID computes training loss of all intra-subject samples in a mini-batch at once (Eq. (3)). We observe that larger batches help improve feature robustness (Tab. 3), yet do not further enhance the downstream performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 339, + 261, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 339, + 261, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 339, + 261, + 350 + ], + "type": "text", + "content": "4.6 Further Discussions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 361, + 481, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 481, + 422 + ], + "type": "text", + "content": "Limitations Although Brain-ID features excel in multiple tasks, they may not be optimal for tasks not involving voxel-wise prediction, e.g., functional brain mapping. Further, as Brain-ID generates data exclusively from anatomies, it tends to struggle with images with extensive pathology. Our future research will focus on robust representations that are encoded with pathology information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 425, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 481, + 498 + ], + "type": "text", + "content": "Data Privacy and Ethics Since Brain-ID features can potentially be used to reconstruct a high-resolution MRI scan, publicly sharing such features should be regarded as sharing the original scans from a subject [33]. Therefore, one should follow the same anonymization protocols as when sharing MRI data for research [61] – not only the removal of identifiers, but also defacing [6] or skull-stripping [26] to prevent the 3D reconstruction of faces." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 517, + 220, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 517, + 220, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 220, + 530 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 481, + 666 + ], + "type": "text", + "content": "We introduced Brain-ID, a contrast-agnostic feature representation learning approach for brain imaging, which is distinctive to brain anatomy and resilient to variations in image appearances. Brain-ID is trained on synthetic inputs, and quickly adapts to downstream tasks through one layer. We validated Brain-ID features' robustness, and effectiveness on four downstream applications (reconstruction, segmentation, super-resolution, bias field estimation), and six public datasets, covering MR and CT. Brain-ID achieves state-of-the-art performances across all tasks and modalities, it further preserves high performance on low-resolution and small-size datasets. We believe Brain-ID unlocks the great potential of robust foundation models, especially for non-calibrated modalities." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "P. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "type": "text", + "content": "Primarily supported by NIH under grant 1RF1AG080371. Additional support from NIH 1RF1AG080371, 1R01EB031114, 1R01AG070988, 1RF1MH123195. OP is supported by Lundbeckfonden R360-2021-395. Wellcome Trust award 221915/Z/20/Z and the NIHR UCLH Biomedical Research Centre support DCA on this topic." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 220, + 197, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 220, + 197, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 220, + 197, + 232 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 246, + 480, + 664 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 246, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 480, + 277 + ], + "type": "text", + "content": "1. Adnan, M., Kalra, S., Cresswell, J.C., Taylor, G.W., Tizhoosh, H.R.: Federated learning and differential privacy for medical image analysis. Scientific reports (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 280, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 280, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 280, + 480, + 311 + ], + "type": "text", + "content": "2. Awais, M., Naseer, M., Khan, S.S., Anwer, R.M., Cholakkal, H., Shah, M., et al.: Foundational models defining a new era in vision: A survey and outlook. arXiv abs/2307.13721 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 313, + 480, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 313, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 313, + 480, + 344 + ], + "type": "text", + "content": "3. Balakrishnan, G., Zhao, A., Sabuncu, M.R., Guttag, J.V., Dalca, A.V.: VoxelMorph: A learning framework for deformable medical image registration. IEEE Transactions on Medical Imaging (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 346, + 480, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 480, + 376 + ], + "type": "text", + "content": "4. Billot, B., Greve, D.N., Puonti, O., Thielscher, A., Leemput, K.V., Fischl, B., et al.: SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining. Medical Image Analysis (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 379, + 480, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 379, + 480, + 410 + ], + "spans": [ + { + "bbox": [ + 138, + 379, + 480, + 410 + ], + "type": "text", + "content": "5. Billot, B., Magdamo, C., Cheng, Y., Arnold, S.E., Das, S., Iglesias, J.E.: Robust machine learning segmentation for large-scale analysis of heterogeneous clinical brain MRI datasets. Proceedings of the National Academy of Sciences (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 412, + 480, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 412, + 480, + 443 + ], + "spans": [ + { + "bbox": [ + 138, + 412, + 480, + 443 + ], + "type": "text", + "content": "6. Bischoff-Grethe, A., Ozyurt, I.B., Busa, E., Quinn, B.T., Fennema-Notestine, C., Clark, C.P., et al.: A technique for the deidentification of structural brain MR images. Human Brain Mapping (2007)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 445, + 480, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 480, + 476 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 480, + 476 + ], + "type": "text", + "content": "7. Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., et al.: On the opportunities and risks of foundation models. arXiv abs/2108.07258 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 478, + 480, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 478, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 138, + 478, + 480, + 509 + ], + "type": "text", + "content": "8. Brant-Zawadzki, M.N., Gillan, G.D., Nitz, W.R.: MP RAGE: A three-dimensional, T1-weighted, gradient-echo sequence-initial experience in the brain. Radiology (1992)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 511, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 511, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 511, + 480, + 553 + ], + "type": "text", + "content": "9. Brown, M.R.G., Sidhu, G.P.S., Greiner, R., Asgarian, N., Bastani, M., Silverstone, P.H., et al.: ADHD-200 global competition: diagnosing ADHD using personal characteristic data can outperform resting state fMRI measurements. Frontiers in Systems Neuroscience (2012)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 555, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 555, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 134, + 555, + 480, + 576 + ], + "type": "text", + "content": "10. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., et al.: Language models are few-shot learners. In: NeurIPS (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 578, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 578, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 134, + 578, + 480, + 609 + ], + "type": "text", + "content": "11. Cardoso, M.J., Li, W., Brown, R., Ma, N., Kerfoot, E., Wang, Y., et al.: MONAI: An open-source framework for deep learning in healthcare. arXiv abs/2211.02701 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 610, + 480, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 610, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 134, + 610, + 480, + 631 + ], + "type": "text", + "content": "12. Caron, M., Touvron, H., Misra, I., J'egou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 632, + 480, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 480, + 664 + ], + "type": "text", + "content": "13. Cerri, S., Puonti, O., Meier, D.S., Wuerfel, J., Mühlau, M., Siebner, H.R., Leemput, K.V.: A contrast-adaptive method for simultaneous whole-brain and lesion segmentation in multiple sclerosis. NeuroImage (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "type": "text", + "content": "Brain-ID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "14. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., et al.: PaLM: Scaling language modeling with pathways. JMLR (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "text", + "content": "15. Chua, Y.Z.R., Dalca, A.V.: Contrast invariant feature representations for segmentation and registration of medical images. In: MIDL (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "type": "text", + "content": "16. Chua, Z.Y., Zheng, W., Chee, M.W.L., Zagorodnov, V.: Evaluation of performance metrics for bias field correction in MR brain images. Journal of Magnetic Resonance Imaging (2009)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "text", + "content": "17. Darzidehkalani, E., Ghasemi-Rad, M., van Ooijen, P.: Federated learning in medical imaging: Part I: Toward multicentral health care ecosystems. Journal of the American College of Radiology (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 227, + 480, + 248 + ], + "type": "text", + "content": "18. Ding, Z., Han, X., Liu, P., Niethammer, M.: Local temperature scaling for probability calibration. In: ICCV (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 133, + 249, + 480, + 281 + ], + "type": "text", + "content": "19. Essen, D.C.V., Ugurbil, K., Auerbach, E.J., Barch, D.M., Behrens, T.E.J., Bucholz, R.D., et al.: The human connectome project: A data acquisition perspective. NeuroImage (2012)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 282, + 480, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 313 + ], + "type": "text", + "content": "20. Fischl, B., Salat, D.H., Busa, E., Albert, M.S., Dieterich, M., Haselgrove, C., et al.: Whole brain segmentation automated labeling of neuroanatomical structures in the human brain. Neuron (2002)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "21. Fonov, V.S., Evans, A.C., Botteron, K., Almli, C.R., McKinstry, R.C., Collins, D.L., et al.: Unbiased average age-appropriate atlases for pediatric studies. Neuroimage (2011)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "22. Fonov, V.S., Evans, A.C., McKinstry, R.C., Almli, C.R., Collins, D.: Unbiased nonlinear average age-appropriate brain templates from birth to adulthood. NeuroImage (2009)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 434 + ], + "type": "text", + "content": "23. Fowler, C., Rainey-Smith, S.R., Bird, S., Bomke, J., Bourgeat, P., Brown, B.M., et al.: Fifteen years of the australian imaging, biomarkers and lifestyle (AIBL) study: Progress and observations from 2,359 older adults spanning the spectrum from cognitive normality to Alzheimer's disease. Journal of Alzheimer's disease reports (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "24. Henschel, L., Kugler, D., Reuter, M.: FastSurferVINN: Building resolution-independence into deep learning segmentation methods — A solution for highres brain MRI. NeuroImage (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "text", + "content": "25. Hoffmann, M., Billot, B., Greve, D.N., Iglesias, J.E., Fischl, B., Dalca, A.V.: SynthMorph: Learning contrast-invariant registration without acquired images. IEEE Transactions on Medical Imaging (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "text", + "content": "26. Hoopes, A., Mora, J.S., Dalca, A.V., Fischl, B., Hoffmann, M.: SynthStrip: skullstripping for any brain image. NeuroImage (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "text", + "content": "27. Iglesias, J.E., Billot, B., Balbastre, Y., Magdamo, C.G., Arnold, S., Das, S., et al.: SynthSR: A public AI tool to turn heterogeneous clinical brain scans into high-resolution T1-weighted images for 3D morphometry. Science Advances (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 599 + ], + "type": "text", + "content": "28. Iglesias, J.E., Billot, B., Balbastre, Y., Tabari, A., Conklin, J., Alexander, D.C., et al.: Joint super-resolution and synthesis of " + }, + { + "bbox": [ + 132, + 555, + 480, + 599 + ], + "type": "inline_equation", + "content": "1\\mathrm{mm}" + }, + { + "bbox": [ + 132, + 555, + 480, + 599 + ], + "type": "text", + "content": " isotropic MP-RAGE volumes from clinical MRI exams with scans of different orientation, resolution and contrast. NeuroImage (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "29. Iglesias, J.E., Schleicher, R., Laguna, S., Billot, B., Schaefer, P., McKaig, B., et al.: Quantitative brain morphometry of portable low-field-strength MRI using superresolution machine learning. Radiology (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "30. Jack, C.R., Bernstein, M.A., Fox, N.C., Thompson, P.M., Alexander, G.E., Harvey, D.J., et al.: The Alzheimer's disease neuroimaging initiative (ADNI): MRI methods. Journal of Magnetic Resonance Imaging (2008)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "P. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "31. Juntu, J., Sijbers, J., Dyck, D.V., Gielen, J.: Bias field correction for MRI images. In: CORES (2005)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 481, + 171 + ], + "type": "text", + "content": "32. Kamnitsas, K., Ledig, C., Newcombe, V.F.J., Simpson, J.P., Kane, A.D., Menon, D.K., et al.: Efficient multi-scale 3D CNN with fully connected CRF for accurate brain lesion segmentation. Medical Image Analysis (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 481, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 481, + 191 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 481, + 191 + ], + "type": "text", + "content": "33. Kapkiç, M., Sağiroğlu, Ş.: Privacy issues in magnetic resonance images. International Journal of Information Security Science (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 481, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 481, + 213 + ], + "type": "text", + "content": "34. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., et al.: Segment anything. arXiv abs/2304.02643 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 213, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 481, + 246 + ], + "type": "text", + "content": "35. LaMontagne, P.J., Keefe, S.J., Lauren, W., Xiong, C., Grant, E.A., Moulder, K.L., et al.: OASIS-3: Longitudinal neuroimaging, clinical, and cognitive dataset for normal aging and Alzheimer's disease. Alzheimer's & Dementia (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "text", + "content": "36. Laso, P., Cerri, S., Sorby-Adams, A., Guo, J., Mateen, F., Goebl, P., et al.: Quantifying white matter hyperintensity and brain volumes in heterogeneous clinical and low-field portable MRI. arXiv abs/2312.05119 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 278, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 481, + 311 + ], + "type": "text", + "content": "37. Leemput, K.V., Maes, F., Vandermeulen, D., Suetens, P.: A unifying framework for partial volume segmentation of brain MR images. IEEE Transactions on Medical Imaging (2003)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 311, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 481, + 343 + ], + "type": "text", + "content": "38. Li, Z., Tian, L., Mok, T.C.W., Bai, X., Wang, P., Ge, J., et al.: SAMConvex: Fast discrete optimization for CT registration using self-supervised anatomical embedding and correlation pyramid. In: MICCAI (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 343, + 481, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 481, + 364 + ], + "type": "text", + "content": "39. Liu, P., Lee, Y., Aylward, S., Niethammer, M.: Deep decomposition for stochastic normal-abnormal transport. In: CVPR (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 364, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 481, + 385 + ], + "type": "text", + "content": "40. Liu, P., Tian, L., Zhang, Y., Aylward, S., Lee, Y., Niethammer, M.: Discovering hidden physics behind transport dynamics. In: CVPR (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "text", + "content": "41. Liu, P., Wang, R., Cao, X., Zhou, Y., Shah, A., Lim, S.N.: Differential motion evolution for fine-grained motion deformation in unsupervised image animation. arXiv abs/2110.04658 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 418, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 439 + ], + "type": "text", + "content": "42. Miletari, F., Navab, N., Ahmadi, S.A.: V-Net: Fully convolutional neural networks for volumetric medical image segmentation. 3DV (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "type": "text", + "content": "43. Modat, M., Ridgway, G.R., Taylor, Z.A., Lehmann, M., Barnes, J., Hawkes, D.J., et al.: Fast free-form deformation using graphics processing units. Computer Methods and Programs in Biomedicine (2010)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 472, + 481, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 503 + ], + "type": "text", + "content": "44. Moor, M., Banerjee, O., Abad, Z.F.H., Krumholz, H.M., Leskovec, J., Topol, E.J., Rajpurkar, P.: Foundation models for generalist medical artificial intelligence. Nature (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 503, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 481, + 536 + ], + "type": "text", + "content": "45. Müller, R.A., Shih, P., Keehn, B., Deyoe, J.R., Leyden, K.M., Shukla, D.: Underconnected, but how? A survey of functional connectivity MRI studies in autism spectrum disorders. Cerebral Cortex (2011)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "type": "text", + "content": "46. Puonti, O., Iglesias, J.E., Leemput, K.V.: Fast and sequence-adaptive whole-brain segmentation using parametric bayesian modeling. NeuroImage (2016)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 558, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 481, + 578 + ], + "type": "text", + "content": "47. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: MICCAI (2015)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "text", + "content": "48. Sheller, M.J., Edwards, B., Reina, G.A., Martin, J., Pati, S., Kotrotsou, A., et al.: Federated learning in medicine: Facilitating multi-institutional collaborations without sharing patient data. Scientific reports (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "49. Shen, Z., Feydy, J., Liu, P., Curiale, A.H., San Jose Estepar, R., San Jose Estepar, R., Niethammer, M.: Accurate point cloud registration with robust optimal transport. In: NeurIPS (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "50. Singhal, K., Azizi, S., Tu, T., Mahdavi, S., Wei, J., Chung, H.W., et al.: Large language models encode clinical knowledge. Nature (2022)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "type": "text", + "content": "Brain-ID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 578 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "51. Symms, M.R., Jäger, H.R., Schmierer, K., Yousry, T.A.: A review of structural magnetic resonance neuroimaging. Journal of Neurology, Neurosurgery & Psychiatry (2004)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "52. Tanno, R., Worrall, D.E., Kaden, E., Alexander, D.C.: Uncertainty modelling in deep learning for safer neuroimage enhancement: Demonstration in diffusion MRI. NeuroImage (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "53. Tian, Q., Bilgiç, B., Fan, Q., Ngamsombat, C., Zaretskaya, N., Fultz, N.E., et al.: Improving in vivo human cerebral cortical surface reconstruction using data-driven super-resolution. Cerebral Cortex (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "54. Tu, T., Azizi, S., Driess, D., Schaekermann, M., Amin, M., Chang, P.C., et al.: Towards generalist biomedical AI. arXiv abs/2307.14334 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "55. de Vos, B.D., Berendsen, F.F., Viergever, M.A., Sokooti, H., Staring, M., Isgum, I.: A deep learning framework for unsupervised affine and deformable image registration. Medical Image Analysis (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "text", + "content": "56. Wachinger, C., Golland, P., Reuter, M.: BrainPrint: Identifying subjects by their brain. In: MICCAI (2014)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "57. Wang, M., Deng, W.: Deep visual domain adaptation: A survey. Neurocomputing (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "58. Wang, T.C., Mallya, A., Liu, M.Y.: One-shot free-view neural talking-head synthesis for video conferencing. In: CVPR (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "59. Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for image quality assessment. In: ACSSC (2003)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "text", + "content": "60. Weiner, M.W., Veitch, D.P., Aisen, P.S., Beckett, L.A., Cairns, N.J., Green, R.C., et al.: The Alzheimer's disease neuroimaging initiative 3: Continued innovation for clinical trial improvement. Alzheimer's & Dementia (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 391, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 424 + ], + "type": "text", + "content": "61. White, T., Blok, E., Calhoun, V.D.: Data sharing and privacy issues in neuroimaging research: Opportunities, obstacles, challenges, and monsters under the bed. Human Brain Mapping (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "62. Xu, Y., Mo, T., Feng, Q., Zhong, P., Lai, M., Chang, E.I.C.: Deep learning of feature representation with multiple instance learning for medical image analysis. In: ICASSP (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 457, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 490 + ], + "type": "text", + "content": "63. Yan, K., Cai, J., Jin, D., Miao, S., Harrison, A.P., Guo, D., et al.: SAM: Self-supervised learning of pixel-wise anatomical embeddings in radiological images. IEEE Transactions on Medical Imaging (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "text", + "content": "64. Yang, X., Kwitt, R., Niethammer, M.: Quicksilver: Fast predictive image registration - a deep learning approach. NeuroImage (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "65. You, C., Zhao, R., Staib, L.H., Duncan, J.S.: Momentum contrastive voxel-wise representation learning for semi-supervised volumetric medical image segmentation. In: MICCAI (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "type": "text", + "content": "66. Zhou, Y., Chia, M.A., Wagner, S.K., Ayhan, M.S., Williamson, D.J., Struyven, R.R., et al.: A foundation model for generalizable disease detection from retinal images. Nature (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "text", + "content": "P. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Bridge Past and Future_ Overcoming Information Asymmetry in Incremental Object Detection/1c449442-a7f3-4df8-9bc6-973b26235638_content_list.json b/2024/Bridge Past and Future_ Overcoming Information Asymmetry in Incremental Object Detection/1c449442-a7f3-4df8-9bc6-973b26235638_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c9d8cc0ce6775b88b0275133c91cde095a2b43ba --- /dev/null +++ b/2024/Bridge Past and Future_ Overcoming Information Asymmetry in Incremental Object Detection/1c449442-a7f3-4df8-9bc6-973b26235638_content_list.json @@ -0,0 +1,1791 @@ +[ + { + "type": "text", + "text": "Bridge Past and Future: Overcoming Information Asymmetry in Incremental Object Detection", + "text_level": 1, + "bbox": [ + 218, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qijie Mo $^{1,3}$ , Yipeng Gao $^{1,3}$ , Shenghao Fu $^{1,3}$ , Junkai Yan $^{1,3}$ , Ancong Wu $^{1,3,\\star}$ , and Wei-Shi Zheng $^{1,2,3,\\star}$", + "bbox": [ + 258, + 210, + 743, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Engineering, Sun Yat-sen University, China \n $^{2}$ Peng Cheng Laboratory, Shenzhen, China", + "bbox": [ + 238, + 253, + 761, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China", + "bbox": [ + 269, + 282, + 730, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{moqj3,gaoyp23,fushh7,yanjk3}@mail2.sysu.edu.cn,wuanc@mail.sysu.edu.cn, wszheng@ieee.org", + "bbox": [ + 316, + 311, + 684, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In incremental object detection, knowledge distillation has been proven to be an effective way to alleviate catastrophic forgetting. However, previous works focused on preserving the knowledge of old models, ignoring that images could simultaneously contain categories from past, present, and future stages. The co-occurrence of objects makes the optimization objectives inconsistent across different stages since the definition for foreground objects differs across various stages, which limits the model's performance greatly. To overcome this problem, we propose a method called \"Bridge Past and Future\" (BPF), which aligns models across stages, ensuring consistent optimization directions. In addition, we propose a novel Distillation with Future (DwF) loss, fully leveraging the background probability to mitigate the forgetting of old classes while ensuring a high level of adaptability in learning new classes. Extensive experiments are conducted on both Pascal VOC and MS COCO benchmarks. Without memory, BPF outperforms current state-of-the-art methods under various settings. The code is available at https://github.com/iSEE-Laboratory/BPF.", + "bbox": [ + 261, + 376, + 743, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Object Detection $\\cdot$ Incremental Learning $\\cdot$ Knowledge Distillation", + "bbox": [ + 259, + 625, + 740, + 651 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 678, + 375, + 695 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object detection [3,14,36,53,61,68] is a fundamental computer vision task and has significantly given rise to many sub-directions [2, 15, 16, 19, 27, 44, 47, 63]. General detection models typically assume full access to all interest classes during training and require exhaustive labeled data from the start. However, in dynamic real-world applications, where new object categories may appear continually, object detection models are expected to adapt to these ongoing changing classes, especially when there is limited storage to save all the data from the beginning", + "bbox": [ + 212, + 710, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding Author", + "bbox": [ + 217, + 824, + 385, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aa666c9761dbf89963229ec6d7447721c272172efbfccb6dbd2a6b4c86a16591.jpg", + "image_caption": [ + "(a) Distribution of instances in the second stage of the 10-5-5 setting on VOC 2007 dataset" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 439, + 246 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f3a2f32a8be1939a244179e29f7af2bda6521c50c3ed594b7bde39ff7f6eaadc.jpg", + "image_caption": [ + "Fig. 1: Challenges faced by the IOD task. Classes from previous, current, and potential future stages may appear arbitrarily in the current training stage, while only the annotations of current classes are available to train the detector. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 444, + 146, + 612, + 244 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/168029cc099ccf2ea6d2fa6f540f61497bab82dfb5381267fd314dbf1e49bc09.jpg", + "image_caption": [ + "(b) Visualization of some typical training samples in IOD" + ], + "image_footnote": [], + "bbox": [ + 614, + 147, + 784, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to the present or when data privacy concerns are encountered. To this end, incremental object detection (IOD) [5,26,41,65,67,69] has caught progressive attention, which aims to continually detect new objects without forgetting the previously learned ones.", + "bbox": [ + 212, + 354, + 782, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike incremental learning in classification, the incremental object detection task faces a dual challenge caused by the concurrence of different classes: (1) Foreground objects identified in previous stages are unlabeled in the current stage, and thus, the model currently trained may regard them as background. (2) The current stage's background may also contain objects that will be recognized as the foreground in future stages. Taking the 10-5-5 setting on the VOC dataset [12] as an example, as shown in Figure 1, a large proportion of instances belonging to classes of the previous and future stages occur concurrently in the training images in the current stage. This asymmetry in information between the past and current stages causes the current model to erroneously classify objects of old categories as background, aggravating the catastrophic forgetting problem. Similarly, the asymmetry between the current and future stages leads the current model to classify objects of potential future categories as background erroneously, which requires models trained in future stages to rectify their misperception, thus hindering the learning of new classes. As a result, the optimization objectives across different learning stages are inconsistent, significantly limiting the generalization of current incremental detectors. Previous works [5, 65, 67] concentrate on introducing strong regularization to prevent catastrophic forgetting but ignore the impact of concurrence of different classes, resulting in limited performance on both old and new classes.", + "bbox": [ + 212, + 415, + 787, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This work aims to tackle the critical information asymmetry challenge in IOD by utilizing the abundant concurrence information within an image. We introduce a novel approach named Bridge Past and Future (BPF), which is designed to connect the past and future stages with the current stage, ensuring the model adheres to consistent optimization objectives throughout the entire incremental learning process. For the past stages, we utilize the previous model as a reliable labeler to construct some pseudo labels for old classes and combine them with current annotations for training the current model symmetrically.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Q. Mo et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Simultaneously, we re-identify some salient regions that may contain potential future objects out of the background and exclude them from negative samples in the current stage to avoid classifying them into the background at this stage and disturbing future learning, which is expected to achieve symmetry with future stages. These two novel designs alleviate the impact of inconsistent training objectives across different stages from the aspects of the past and the future, thus easing the training difficulty of incremental detectors.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition, we propose a distillation technique to augment the detector in the current stage, where we take a step back by viewing the current stage as a clear future stage for previous stages and propose a Distillation with Future (DwF) loss. The distillation is carried out by two teachers, i.e., the detector trained on the previous stage and an expert detector trained with only current data. The student detector absorbs the knowledge of old and current classes from the old and expert models, respectively, in a class-by-class manner, preventing the detector from catastrophic forgetting and facilitating learning of current classes.", + "bbox": [ + 212, + 252, + 787, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We conduct extensive experiments on the PASCAL VOC [12] and COCO [37] datasets to evaluate the effectiveness of our BPF, which outperforms other state-of-the-art (SOTA) methods under multiple settings in a memory-free way. Moreover, we also conduct comprehensive ablation studies and visualizations to help better understand how each component of the BPF works.", + "bbox": [ + 212, + 373, + 787, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 470, + 397, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object Detection. Traditional object detectors can be primarily classified into two streams: one-stage [36,52,58,61] and two-stage detectors [18,22,35,53]. Two-stage detectors first predict several coarse candidate proposals via region proposal extractors [18,53] and then adopt a region of interest (RoI) head to refine these proposals and output final predictions. Unlike two-stage detectors, one-stage counterparts directly generate final outputs without predicting candidate proposals. Despite performing well under the standard training setup, both of them fail to generalize to the incremental training setup due to lacking previous training data. Without losing generality, this paper concentrates on enhancing two-stage detectors, i.e., Faster R-CNN [53], enabling it to learn new classes incrementally while retaining previously acquired knowledge.", + "bbox": [ + 212, + 501, + 787, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Incremental Learning. Incremental learning methods can be mainly divided into three categories: memory-based [4, 24, 32, 33, 48, 51, 56, 59, 62], regularization-based [6, 8, 30, 34, 66], and structure-based [42, 45, 46, 54, 60]. Memory-based methods store a handful of samples for replaying [24, 40, 41] or suggest generating samples [29, 62] for data compensation in new stages. Some regularization-based methods try to eliminate the effect of new tasks on previously learned knowledge by identifying the most crucial parameters [30, 43, 62], while some other regularization-based methods propose to distill knowledge [23] from old to current models by transiting knowledge on logits, final features or intermediate features [8, 10, 11, 24, 34]. Structure-based methods dedicate specific parameters to each task, freezing them to mitigate forgetting. In this study, we focus on the", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Bridge Past and Future", + "bbox": [ + 571, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "regularization-based knowledge distillation approach for object detection and introduce a novel distillation with future information.", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Incremental Object Detection. Most previous works [9,28,31,40,50,55] focus on distilling old classes' knowledge of intermediate features [5,21,39,41,49,65,67], region proposal network [5, 49, 67], or RoI head [13] to prevent forgetting. A stream of methods also pays attention to classification loss. PPAS [67] introduces a pseudo-positive-aware sampling algorithm to identify regions corresponding to old classes and prevents them from being sampled as background. MMA [5] proposes unbiased classification loss, consolidating the background and all old classes into one entity, which aims to minimize the optimization objective conflict between the current stage's background and the past stages' foreground. However, it diminishes the ability to distinguish old classes from the background. Other methods [1,20,25,26,41,64] focus on rehearsal to maintain the previous stage knowledge, either performing replay of the intermediate features [1], the images [25,26,40], or the instances [41]. Incdet [38] presents a parameter isolation strategy that builds upon EWC [30]. In this study, in addition to preserving knowledge for old classes, we also take future classes into account, aligning optimization objectives consistently across all stages.", + "bbox": [ + 215, + 181, + 785, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 444, + 328, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 472, + 372, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Problem Formulation. In incremental object detection, the training is performed over multiple learning stages, each one introducing a new set of classes to be detected. Let $\\mathcal{C}$ denote the set of classes that are incrementally introduced to the object detector $\\mathcal{M}$ . In the $t$ -th training stage, a grouping of classes $\\mathcal{C}_t$ are introduced to the detector: $\\mathcal{C}_t \\subset \\mathcal{C}$ , such that $\\mathcal{C}_i \\cap \\mathcal{C}_j = \\emptyset$ , for any $i \\neq j$ and $i, j \\leq t$ . Let $\\mathcal{D}_t$ denote the images containing annotated objects of classes in $\\mathcal{C}_t$ . Each image can contain multiple objects of different classes, but annotations $\\mathcal{V}_t$ are available only for those object instances that belong to classes in $\\mathcal{C}_t$ . The challenge in class incremental object detection is to continually update $\\mathcal{M}_t$ to $\\mathcal{M}_{t+1}$ by learning $\\mathcal{C}_{t+1}$ , without access to $\\{\\mathcal{D}_0, \\dots, \\mathcal{D}_t\\}$ while maintaining original performance on $\\{\\mathcal{C}_0, \\dots, \\mathcal{C}_t\\}$ .", + "bbox": [ + 215, + 494, + 785, + 661 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "General Detector Training Pipeline. This work starts with the representative Faster RCNN-like detectors [53] $f = \\{f_b, f_{rpn}, f_{roi}\\}$ , which generally includes a backbone network $f_b$ , Region Proposal Network (RPN) $f_{rpn}$ and Region of Interest Head (RoI Head) $f_{roi}$ . During training, all training proposals are categorized into positive, negative, and ignored samples for supervised training based on their IoUs with the ground truth. Positive samples are assigned to predict the class of their matched ground truth, whereas negative samples are designated as background. Thus, the objects without annotations in the current stage will be classified as background, hindering the learning of incremental models.", + "bbox": [ + 215, + 667, + 785, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Common Distillation Methods for Incremental Object Detection. To prevent forgetting in IOD, a widely adopted solution involves knowledge distil-", + "bbox": [ + 215, + 809, + 784, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Q. Mo et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lation in RPN and RoI head [5,49,57,67]:", + "bbox": [ + 214, + 146, + 517, + 162 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i s t}} = \\mathcal {L} _ {\\text {d i s t}} ^ {r p n} + \\mathcal {L} _ {\\text {d i s t}} ^ {r o i}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 169, + 785, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The distillation in the RoI head includes the L2 loss between the box coordinates and the Kullback-Leibler divergence for class probabilities. MMA [5] notices the missing old annotations and proposes Unbiased Knowledge Distillation (UKD):", + "bbox": [ + 214, + 191, + 782, + 238 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i s t}, \\text {c l s}} ^ {\\text {r o i}} (i) = \\frac {1}{| \\mathcal {C} _ {t - 1} | + 1} \\left(p _ {i} ^ {b, t - 1} \\log \\left(p _ {i} ^ {b, t} + \\sum_ {c \\in \\mathcal {C} _ {t}} p _ {i} ^ {c, t}\\right) + \\sum_ {c \\in \\mathcal {C} _ {1: t - 1}} p _ {i} ^ {c, t - 1} \\log \\left(p _ {i} ^ {c, t}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 227, + 244, + 785, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p_i^{c,t-1}$ and $p_i^{c,t}$ indicate the classification output for the proposal $i$ and class $c$ of the teacher (old) model and student (current) model, respectively, and $b$ is the background class. As a common practice, 64 proposals are used for distillation, which are randomly selected out of the top 128 proposals with the highest objectness scores from the RPN network of the old model. However, given the class-agnostic characteristic of the RPN, the proposals for distillation cover objects of both old and current classes. In such cases, the old model, without seeing current classes, cannot provide useful knowledge for the current classes and even hinders their learning. Further, treating the background probability and current class probabilities as a unified entity during distillation inevitably diminishes the ability to differentiate between current classes and background.", + "bbox": [ + 214, + 289, + 787, + 457 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Overall Framework", + "text_level": 1, + "bbox": [ + 214, + 476, + 419, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike classification tasks with a single label per input, incremental object detection presents a scenario where an image $\\mathcal{I}_t$ can encompass objects from the current class set $\\mathcal{C}_t$ , previous class sets $\\mathcal{C}_{1:t-1}$ and future class sets $\\mathcal{C}_{t+1:\\infty}$ . Nonetheless, the annotations $\\mathcal{Y}_t$ are limited to bounding boxes and class labels for objects in $\\mathcal{C}_t$ , while objects of other classes are regarded as background.", + "bbox": [ + 212, + 497, + 782, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address the challenge of inconsistent optimization objectives, we introduce a novel strategy termed Bridge Past and Future (BPF). As shown in Figure 2, the method is divided into two parts. From the perspective of supervised learning, we use high-confidence predictions of old classes given by the old model as pseudo labels to bridge the information from the past stage (Section 3.3) and discard some potential objects from the background to bridge the future stage (Section 3.4). Regarding the distillation learning, we propose a novel Distillation with Future (DwF) loss by considering current classes in the distillation. The distillation probabilities are combined from two teacher models: the old model $\\mathcal{M}_{t-1}$ for old classes and an expert model for current classes. By utilizing abundant information in background probability, the distillation takes both previous stages and the current stage into consideration (Section 3.5).", + "bbox": [ + 214, + 573, + 787, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Bridging the Past", + "text_level": 1, + "bbox": [ + 214, + 772, + 408, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address the inconsistency in optimization objectives between the previous and current models, we bridge the past information to the current stage by incorporating pseudo supervision signals from past stages into the current model's", + "bbox": [ + 212, + 794, + 784, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Bridge Past and Future", + "bbox": [ + 571, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7649c9c1029ff8c148ec97a2b94c23bdbd993fd84fdbb5931ef097af45ffde81.jpg", + "image_caption": [ + "Fig. 2: The overall framework of our method. The top side illustrates the Bridge Past and Future (BPF) procedure, which identifies objects of past classes and excludes several potential objects of future classes to ensure consistent optimization during the entire training process. The bottom side shows the Distillation with Future (DwF) process, which employs both the old model $\\mathcal{M}_{t-1}$ adept at detecting old categories and the interim model $\\mathcal{M}_t^{im}$ trained on $\\mathcal{D}_t$ and specialized in new categories, to conduct a comprehensive distillation across all categories for the current model $\\mathcal{M}_t$ ." + ], + "image_footnote": [], + "bbox": [ + 256, + 147, + 751, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "optimization process. The old model was trained under the supervision of human annotations, which is a high-quality pseudo labeler for old classes. We take the high-confidence inference results of the old model $\\mathcal{M}_{t-1}$ on current training images as pseudo labels and combine them with current annotations to train the current detector, thereby aligning the supervision of old classes with previous stages. The left part of Figure 3 shows the pipeline of bridging the past strategy.", + "bbox": [ + 212, + 511, + 787, + 601 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Formally, given predictions $\\hat{y}^{old} \\in \\mathcal{P}$ of the old model $\\mathcal{M}_{t-1}$ on current images, we first select some high-confidence regions of previous classes $\\mathcal{C}_{1:t-1}$ with a confidence threshold $\\eta$ , followed by the Non-Maximum Suppression (NMS) operation to reduce duplication:", + "bbox": [ + 212, + 601, + 787, + 662 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {U} = \\{j \\in \\mathcal {P}: \\max _ {c \\in \\mathcal {C} _ {1: t - 1}} \\hat {p} _ {j} ^ {\\text {o l d}} (c) > \\eta \\}, \\quad \\mathcal {E} = \\mathrm {N M S} (\\mathcal {U}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 671, + 785, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, we further narrow down the predictions to a subset $\\mathcal{W} \\subset \\mathcal{E}$ that do not overlap with the ground-truth labels of the new categories via an IoU threshold $\\lambda_{1}$ , ensuring a clear distinction between past and present classes:", + "bbox": [ + 212, + 705, + 787, + 753 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {W} = \\{j \\in \\mathcal {E}: \\forall i \\in \\mathcal {Y} _ {t}, \\operatorname {I o U} \\left(\\hat {\\boldsymbol {b}} _ {j} ^ {\\text {o l d}}, \\boldsymbol {b} _ {i}\\right) \\leq \\lambda_ {1} \\}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 762, + 785, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By modeling the previous supervision signals from the old model, we bridge the past stages to the current one, ensuring the current model's optimization direction encompasses the objectives of earlier stages, significantly mitigating", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Q. Mo et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/70c935fc92d0f32026cb7ebe198ff5f50ad55a53e2c2a73dc6d00fec1b1e5b56.jpg", + "image_caption": [ + "Fig. 3: Overview of Bridge Past and Future. We adopt the previous model $\\mathcal{M}_{t-1}$ to predict some pseudo labels for past classes to complement their missing supervision in the current stage. Additionally, we exclude several proposals that are likely to be an object but are not included in the current ground truth and pseudo labels from the background to avoid classifying them into background mistakenly." + ], + "image_footnote": [], + "bbox": [ + 272, + 145, + 728, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the forgetting problem. MMA [5] treats the background probability and old class probabilities as a unified entity in the classification loss, making old classes hard to separate from the background. While our method explicitly models the old classes in the current stage out of the background.", + "bbox": [ + 212, + 405, + 782, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Bridging the Future", + "text_level": 1, + "bbox": [ + 215, + 489, + 426, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Bridging the future aims to rectify the misalignment between the optimization goals of the current and upcoming stages, as objects from future class sets $\\mathcal{C}_{t + 1:\\infty}$ in the current dataset $\\mathcal{D}_t$ are classified as background. The core design is to find some salient objects in the background and exclude them from the negative samples during the training. By separating pure background from regions likely to contain future category objects in the current stage, we align the optimization objectives for background treatment across both current and future models.", + "bbox": [ + 212, + 517, + 784, + 621 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, we find that the activation in feature maps is a good indicator for salient objects. As illustrated in Figure 3, in the absence of annotations for old and future classes, foreground and background features demonstrate notable differences in spatial attention. We produce the attention map $A_{i} \\in \\mathbb{R}^{H \\times W}$ from the backbone feature map $F_{i} = f_{b}(I_{i}) \\in \\mathbb{R}^{H \\times W \\times C}$ :", + "bbox": [ + 212, + 623, + 784, + 699 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nA _ {i} = \\operatorname {S o f t m a x} \\left(\\sum_ {c = 1} ^ {C} \\left| F _ {i} \\right| ^ {p}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 710, + 785, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $H$ , $W$ , and $C$ denote the feature's height, width, and channel. We compute the attention score for each region, indicating the likelihood of containing foreground objects. The attention score $a_{i,j}^{roi}$ for region $r_j$ is calculated as follows:", + "bbox": [ + 214, + 762, + 782, + 810 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\na _ {i, j} ^ {\\text {r o i}} = \\operatorname {A v g} \\left(\\operatorname {R o I P o o l} \\left(A _ {i}, r _ {j}\\right)\\right) \\in \\mathbb {R}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 823, + 784, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Bridge Past and Future", + "bbox": [ + 571, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d5787d9ef4e0988e0987e9d0f93fb93ceefe6246d848e1ab7214142d5058d50.jpg", + "image_caption": [ + "Fig. 4: The illustration of Distillation with Future strategy. An intermediate teacher model trained on the current dataset is used to compensate for the lack of current class information in the old model. For proposals overlapping ground truth of the current stage, since the intermediate teacher is specialized in detecting current classes, we directly inherit its probabilities on current classes and use the old model to enrich its background probability with old class knowledge. On the contrary, for proposals that do not overlap with GT, the old model is preferred, and the intermediate model is used as compensation. Combining two teachers makes the distillation class by class." + ], + "image_footnote": [], + "bbox": [ + 222, + 145, + 782, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Regions with high attention scores $a_{i,j}^{roi}$ from feature maps and objectness scores $o_j$ from class-agnostic RPN imply a greater chance of being future category objects. We discard these RoIs when sampling negative samples when training the RoI head, thus maintaining the model's consistency with future stage background definitions. Regions with lower scores or having a considerable IoU with ground truth are considered reliable backgrounds.", + "bbox": [ + 212, + 521, + 787, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Combining Current Classes into Distillation", + "text_level": 1, + "bbox": [ + 214, + 633, + 624, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Distillation is an effective way to prevent forgetting in incremental object detection. However, distilling from the teacher model, which is biased toward old classes, inevitably hinders the learning for current classes. To address this challenge, we introduce the Distillation with Future (DwF) loss to distill from the teacher model in a more fine-grained and adaptive way. Different from MMA [5], which aligns the current classes and the background as a whole with the old model's background, we distill each class one by one, preserving the distinction between current classes and the background. However, the model $\\mathcal{M}_{t-1}$ from the previous stage has not been trained on current classes, thus the distillation on current classes can not be performed. We introduce another intermediate teacher model $\\mathcal{M}_t^{im}$ , which is trained using the dataset from the current stage $\\mathcal{D}_t$ in a fully supervised way, as a supplement. Taking the intermediate model $\\mathcal{M}_t^{im}$ as a", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Q. Mo et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "complementary teacher, we explicitly consider the current stage $t$ as the future stage of stage $t - 1$ , making the distillation future-aware.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As the old model $\\mathcal{M}_{t-1}$ performs well on $\\mathcal{C}_{1:t-1}$ while the intermediate model $\\mathcal{M}_t^{im}$ is the expert in $\\mathcal{C}_t$ , we distill different regions with different combination of teacher models, as shown in Figure 4. Specifically, we divide the regions for distillation $\\mathcal{R}$ into two subsets $\\mathcal{R}_1, \\mathcal{R}_2 \\subset \\mathcal{R}$ that based on their intersection over union with the ground truth labels for the new categories $\\mathcal{C}_t$ :", + "bbox": [ + 212, + 176, + 784, + 252 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {1} = \\{j \\in \\mathcal {R}: \\forall i \\in \\mathcal {Y} _ {t}, \\operatorname {I o U} (b _ {j}, b _ {i}) \\leq \\lambda_ {2.} \\},\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 262, + 648, + 277 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {R} _ {1} = \\left\\{j \\in \\mathcal {K}: \\forall i \\in \\mathcal {Y} _ {t}, \\operatorname {I o U} \\left(b _ {j}, b _ {i}\\right) \\geq \\lambda_ {2}. \\right\\}, \\\\ \\mathcal {R} _ {2} = \\left\\{j \\in \\mathcal {R}: \\forall i \\in \\mathcal {Y} _ {t}, \\operatorname {I o U} \\left(b _ {j}, b _ {i}\\right) > \\lambda_ {2}. \\right\\}. \\end{array} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 268, + 784, + 296 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For regions $r_i \\in \\mathcal{R}_1$ , which are likely to be the regions for old classes, we take $\\mathcal{M}_{t-1}$ as the primary model for distillation and reconstruct its background representation with the model $\\mathcal{M}_t^{im}$ :", + "bbox": [ + 212, + 304, + 782, + 349 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {p} _ {i} ^ {c, i m} = p _ {i} ^ {c, i m} \\times p _ {i} ^ {b, t - 1}, \\quad r _ {i} \\in \\mathcal {R} _ {1}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 359, + 784, + 378 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $p_i^{b,t-1}$ and $p_i^{c,im}$ are the classification probabilities of background in the model $\\mathcal{M}_{t-1}$ and the classification probabilities for current classes and background $c \\in \\mathcal{C}_t \\cup \\mathcal{B}$ in the model $\\mathcal{M}_t^{im}$ for the region $r_i$ respectively. After weighting, $\\sum_{c=1}^{\\mathcal{C}_t \\cup \\mathcal{B}} \\hat{p}_i^{c,im} = p_i^{b,t-1}$ . The final distillation probabilities for regions in $\\mathcal{R}_1$ are $[p_i^{C_{1:t-1},t-1}, \\hat{p}_i^{C_t \\cup \\mathcal{B},im}] \\in \\mathbb{R}^{|C_{1:t}|+1}$ .", + "bbox": [ + 212, + 386, + 784, + 482 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "On the contrary, for regions $r_i \\in \\mathcal{R}_2$ , they are regions for current classes; thus, the intermediate model $\\mathcal{M}_t^{im}$ is taken as the primary model for distillation, and its background representation is reconstructed by the model $\\mathcal{M}_{t-1}$ :", + "bbox": [ + 212, + 483, + 784, + 527 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {p} _ {i} ^ {c, t - 1} = p _ {i} ^ {c, t - 1} \\times p _ {i} ^ {b, i m}, \\quad r _ {i} \\in \\mathcal {R} _ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 536, + 784, + 554 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\sum_{c=1}^{\\mathcal{C}_{1:t-1} \\cup \\mathcal{B}} \\hat{p}_i^{c,t-1} = p_i^{b,im}$ . The distillation probabilities for regions in $\\mathcal{R}_2$ are $[\\hat{p}_i^{\\mathcal{C}_{1:t-1},t-1}, p_i^{\\mathcal{C}_t,im}, \\hat{p}_i^{\\mathcal{B},t-1}] \\in \\mathbb{R}^{|C_{1:t}|+1}$ .", + "bbox": [ + 212, + 563, + 782, + 614 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "With the expanded probabilities from the complimentary teacher models, the distillation on the classification head can be performed using a conventional Kullback-Leibler divergence. Regarding the box distillation, we use the output boxes from the old model $\\mathcal{M}_{t-1}$ for regions in $\\mathcal{R}_1$ and the intermediate model $\\mathcal{M}_t^{im}$ for $\\mathcal{R}_2$ . As a result, the complementary knowledge from two teachers, the expansion of background probability, and the combination of adaptive probability for different regions not only prevent the student from catastrophic forgetting but also facilitate the learning for current classes.", + "bbox": [ + 212, + 614, + 784, + 734 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 757, + 375, + 773 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experiment Settings", + "text_level": 1, + "bbox": [ + 214, + 786, + 431, + 801 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets and Evaluation Metrics. Following previous works [5,20,25,26,41, 49,57,65,67], we evaluate our method on PASCAL VOC 2007 [12] and MS COCO", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Bridge Past and Future", + "bbox": [ + 571, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/a203671c09b13247dcfd6fb18836b63ee49cc805ebe1c145acb82611e5291cb8.jpg", + "table_caption": [ + "Table 1: mAP@0.5 results on single incremental step on PASCAL VOC 2007. The best performance in each is presented with **bold**, and the second best is presented with **underline**. Methods with * use exemplars." + ], + "table_footnote": [], + "table_body": "
Method19-115-510-105-15
1-19201-20Avg1-1516-201-20Avg1-1011-201-20Avg1-55-151-20Avg
Joint Training76.076.776.176.478.070.476.174.275.976.376.176.172.477.376.174.9
Fine-tuning12.062.814.537.414.259.225.436.79.562.536.036.06.963.149.135.0
ORE* [25]69.460.168.964.771.858.768.565.260.468.864.664.6----
OW-DETR* [20]70.262.069.866.172.259.869.166.063.567.965.765.7----
ILOD-Meta* [26]70.957.670.264.271.755.967.863.868.464.366.366.3----
ABR* [41]71.069.770.970.473.065.171.069.171.272.872.072.064.771.069.467.9
Faster ILOD [53]68.961.168.565.071.656.967.964.369.854.562.162.162.037.143.349.6
PPAS [67]70.553.069.261.8----63.560.061.861.8----
MVC [65]70.260.669.765.469.457.966.563.766.266.066.166.1----
PROB [69]73.948.572.661.573.560.870.167.066.067.266.566.5----
PseudoRM [64]72.967.372.670.173.460.970.366.969.168.668.968.9----
MMA [5]71.163.470.767.273.060.569.966.769.363.966.666.666.857.259.662.0
BPF (Ours)74.565.374.169.975.963.072.769.571.774.072.972.966.475.373.070.9
", + "bbox": [ + 218, + 198, + 787, + 388 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "2017 [37] datasets. PASCAL VOC 2007 dataset comprises 9,963 images across 20 categories. The COCO 2017 dataset encompasses objects from 80 categories, with around $118\\mathrm{k}$ images for training and 5,000 images for validation. The mean average precision at the 0.5 IoU threshold $(\\mathrm{mAP}@0.5)$ is used as the primary evaluation metric for the VOC dataset, and the mean average precision ranging from 0.5 to 0.95 is the main evaluation metric for the COCO dataset.", + "bbox": [ + 215, + 420, + 785, + 510 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For each incremental setting (A-B), the first number A denotes the number of classes in the first stage and the second number is the number of classes newly introduced in each new stage. Note that the columns with gray background in the table of experimental results represent the average AP among all classes.", + "bbox": [ + 215, + 512, + 785, + 573 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details. Similar to [5,25,26,41,49], we build our incremental object detector based on Faster R-CNN [53] with R50. Our method can easily be adapted to transformer-based detectors [3,17,68]. We conduct the experiments under a strict rehearsal-free setting, where no memory is used. We set $\\eta = 0.75$ , $\\lambda_{1} = 0.7$ , $\\lambda_{2} = 0.5$ . For $\\mathcal{W}$ in Equation (4), we use an IOU threshold of 0.3 to divide it into two sets. For the set with IoU $< 0.3$ , the supervision signal weights are set to 1.0, while the other set is assigned a weight of 0.3.", + "bbox": [ + 215, + 580, + 785, + 686 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 217, + 715, + 457, + 732 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Following previous work [5,7,41,49,57,65,67], our method is evaluated on settings with a range of initial classes and incorporating one or more incremental tasks. We benchmark our method against two baselines: Fine-Tuning, where the model is incrementally trained on new data without any regularization strategy or data replay, and Joint Training, which involves training the model on the complete dataset using all annotations.", + "bbox": [ + 215, + 748, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Q. Mo et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/efafb6105bcc185fc98598180939623aad7486bfee200166883cc4cb65f6e3d8.jpg", + "table_caption": [ + "Table 2: mAP@0.5 results on multiple incremental steps on PASCAL VOC 2007. The best performance in each is presented with **bold**, and the second best is presented with **underline**. Methods with * use exemplars." + ], + "table_footnote": [], + "table_body": "
Method10-5 (3 tasks)5-5 (4 tasks)10-2 (6 tasks)15-1 (6 tasks)10-1 (10 tasks)
1-1011-201-201-56-201-201-1011-201-201-1516-201-201-1011-201-20
Joint Training75.976.376.172.477.376.175.976.376.178.070.476.175.976.376.1
Fine-tuning [41]5.330.618.00.518.313.83.813.68.70.010.55.30.05.12.6
ABR* [41]68.767.167.964.756.458.467.058.162.668.756.765.762.055.758.9
Faster ILOD [53]68.357.963.155.716.025.964.248.656.466.944.561.352.941.547.2
MMA [5]66.761.864.262.331.238.965.053.159.168.354.364.159.248.353.8
BPF (Ours)69.168.268.760.663.162.568.756.362.571.553.166.962.248.355.2
", + "bbox": [ + 218, + 199, + 785, + 296 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/22833711d601c2e587fa126eba80438a3045eb0658988d2ff73871eda0b9cd0a.jpg", + "table_caption": [ + "Table 3: mAP results on COCO2017. Methods with * use exemplars." + ], + "table_footnote": [], + "table_body": "
Method40-4070-10
APAP50AP75APAP50AP75
Joint Training36.757.839.836.757.839.8
Fine-tuning [41]19.031.220.45.68.66.2
ILOD-Meta* [26]23.840.524.4---
ABR* [41]34.557.835.231.152.932.7
Faster ILOD [49]20.640.1-21.339.9-
PseudoRM [64]25.344.4----
MMA [5]33.056.634.630.252.131.5
BPF (Ours)34.454.337.336.256.838.9
", + "bbox": [ + 243, + 332, + 527, + 468 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/d54406906e15d6154fcce269cb255555688f09f03d9d86d3f0d2f7602fe46438.jpg", + "table_caption": [ + "Table 4: Ablation study of various combinations of teacher models." + ], + "table_footnote": [], + "table_body": "
DistillationVOC(10-10)
Lroi dist,clsLroi dist,bbox1-1011-201-20
λ2= 1.0 part boxes71.573.372.4
λ2= 0.5 part boxes71.774.072.9
λ2= 0.5 all boxes71.374.472.9
", + "bbox": [ + 555, + 335, + 745, + 401 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/d99e2ef21274c104781bba063b2293a0c49e7f1cb04ff0c318faac95165b59ef.jpg", + "table_caption": [ + "Table 5: Effect of BF." + ], + "table_footnote": [], + "table_body": "
BFVOC(5-15)VOC(10-10)VOC(15-5)
1-56-201-201-1011-201-201-1516-20
×66.374.472.471.273.372.375.662.8
√66.475.373.071.774.072.975.963.0
", + "bbox": [ + 549, + 426, + 751, + 465 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "PASCAL VOC 2007. For PASCAL VOC 2007, we order the classes alphabetically and evaluate our method with one or multiple training steps. We perform our experiments by adding 1 (19-1), 5 (15-5), 10 (10-10), or 15 (5-15) classes in a single incremental step. For multi-step incremental settings, we evaluate 10-5, 5-5, 10-2, 15-1, and 10-1 settings, where we add 5, 5, 2, 1, and 1 classes respectively at every step until all 20 classes are seen.", + "bbox": [ + 212, + 474, + 784, + 566 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "- Single-step Incremental Settings: Table 1 shows our BPF methods against the existing methods using rehearsal or not. Rehearsal-based methods are not compared fairly with our BPF since we do not store old samples and use replay memory. As shown in Table 1, BPF consistently outperforms all previous methods, including those designed to combat forgetting using exemplars, validating the superiority of our approach. In particular, in the 19-1, 15-5, 10-10, and 5-15 settings, BPF significantly improved over MMA [5] by $3.4\\%$ , $2.8\\%$ , $6.3\\%$ , and $13.4\\%$ on mAP@0.5 across all classes. Similarly, BPF outperforms the best rehearsal-based method ABR [41] by $3.2\\%$ , $1.7\\%$ , $0.9\\%$ , and $3.6\\%$ . The Avg metric equally averages old and new classes mAP, which straightly reports the incremental ability without the influence of the number of classes. BPF also outperforms most methods on the Avg metric, demonstrating BPF's adaptiveness in learning new classes and preserving the knowledge of old classes.", + "bbox": [ + 212, + 566, + 785, + 763 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "- Multi-step Incremental Settings: The issues of inconsistent optimization objectives across multiple stages and catastrophic forgetting are more crucial under the longer incremental settings. As shown in Table 2, BPF consistently outperforms MMA [5] across all the settings. Specifically, BPF improves over MMA by $1.4\\%$ , $2.8\\%$ , $3.4\\%$ , $4.5\\%$ , and even $23.6\\%$ at 1-20 mAP@0.5 under", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Bridge Past and Future", + "bbox": [ + 571, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/14a1be442ad249c9ae2b29655314ffda415ee6a42ee535ac7a4503044eaa487b.jpg", + "table_caption": [ + "Table 6: Ablation study on each component." + ], + "table_footnote": [], + "table_body": "
ModelBridge the PastBridge the futureDistillation with FutureVOC(10-10)VOC(10-5)
1-1011-201-201-1011-1516-20
(a)58.172.465.354.469.959.3
(b)✓71.272.171.769.173.560.0
(c)✓61.073.367.154.870.758.7
(d)✓✓71.972.772.370.473.758.7
(e)✓✓✓71.774.072.969.175.261.2
", + "bbox": [ + 228, + 171, + 772, + 266 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "the 10-1, 15-1, 10-2, 10-5, and 5-5 settings and enjoy improvement across all learning stages. Moreover, we find that even without storing memory, BPF still outperforms ABR [41] by $0.8\\%$ and $4.1\\%$ on overall mAP@0.5, $1.1\\%$ and $6.7\\%$ mAP@0.5 on new classes under 10-5 and 5-5 settings. In the settings of 10-2 and 10-1, limited by the small incremental data, BPF is inferior to ABR, but considering we do not require memory, these losses are acceptable.", + "bbox": [ + 212, + 291, + 782, + 383 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "MS COCO 2017. On the COCO2017 dataset, we perform experiments on 40-40 and 70-10 settings, adding 40 and 10 classes, respectively, following [41]. As illustrated in Table 3, our method improves over MMA on average $AP$ by $1.4\\%$ on 40-40 settings and by $6.0\\%$ on 70-10 settings. These results once again confirm the effectiveness of our method.", + "bbox": [ + 212, + 388, + 784, + 464 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Analysis and Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 484, + 500, + 500 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We examine the contributions of the \"Bridge the Past\", \"Bridge the Future\", and \"Distillation with Future\" in Table 6 within the VOC 10-10 and 10-5 settings. We take the unbiased knowledge distillation proposed by [5] as the baseline model. By bridging the past, our model (b) aligns its optimization objectives with earlier ones, effectively reducing catastrophic forgetting of old classes, greatly enhancing the performance on detecting old classes. Compared to the baseline (a), it significantly improved by $13.1\\%$ in the old classes and by $6.4\\%$ for all classes on the 10-10 setting. Owing to bridging the future, our model (c) maintains consistent optimization objectives with future models regarding the background, making it easier to incrementally learn new classes. Compared to the baseline (a) on the 10-10 setting, it improves by $2.9\\%$ and $0.9\\%$ on the old classes and the new classes, respectively. Combining Past and Future (d), the model outperforms consistently on each stage. In the Distillation with Future strategy, we leverage the intermediate model to aid the old model in modeling new classes, conducting distillation across all categories for the current model. Table 6 (e) shows that a comprehensive joint teacher model facilitates improved learning of all categories through knowledge distillation, with a significant improvement $(+1.3\\%)$ AP) in new classes. Similar results can be found in the 10-5 setting.", + "bbox": [ + 212, + 507, + 785, + 779 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We further verify the \"Bridge the future\" (BF) in Table 5. There is a clear trend that as the number of considered future classes increases, the BF shows increasing improvement in future classes ( $+0.9\\%$ in 5-15) without degrading the old classes' performance.", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Q. Mo et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6c89ff87d065a6f9ab29ccdd0fa7c6cc10189fddcc952b34b2014ea161568258.jpg", + "image_caption": [ + "BPF" + ], + "image_footnote": [], + "bbox": [ + 238, + 143, + 370, + 220 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1900f9ab1924fb3c62d64e4c1e50eb0952d142c0a041630e48aad19db45b424c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 143, + 501, + 220 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5fba540b8bad63031c31db4f4ed03720484fa76007fe7ece8774f19623a4d289.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 143, + 635, + 220 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2ae0cd7cf71505a1319cf2b46972aca1a345b6ac0dccad71432251cc02a6b3ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 143, + 784, + 220 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a446ce0cddd61694c5dca1fbd3446b6fee22ed55502d5a6dafa54842652e7db0.jpg", + "image_caption": [ + "MMA" + ], + "image_footnote": [], + "bbox": [ + 238, + 222, + 370, + 297 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/44113d95ec2825f062d1059c5fe1e4c46b0c7443dc8da63717d267f6f0e27573.jpg", + "image_caption": [ + "Fig. 5: Qualitative results for the model trained under the $10 + 10$ setting on the VOC 2007 test set. 'boat', 'cat', 'chair', and 'car' are old classes from the first stage, and 'person' and 'dog' are the classes from the second stage. Compared with MMA (bottom row), our BPF (top row) can produce reliable predictions on both old and new classes." + ], + "image_footnote": [], + "bbox": [ + 372, + 222, + 501, + 297 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/21b6e2735b4c0f063c09312d9073f5d84cd6cb3e372007ddf6e71039537e4d49.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 222, + 635, + 297 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c577aad62b4b6b7a6ea9da1c0f502309023b1b355eb0af3dc4ad9d8453645b37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 222, + 784, + 297 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We also conducted a quantitative analysis of Bridge Past and Future. Under the second stage of the VOC 10-5-5 (3 tasks), the original background boxes have a $92.6\\%$ and $83.9\\%$ Recall50 rate for old and future classes, while the Recall50 of pseudo labels for old classes and discarded boxes for future classes are $66.1\\%$ and $17.6\\%$ respectively, demonstrating the necessity and effectiveness of Bridge the Past and Bridge the Future.", + "bbox": [ + 212, + 397, + 784, + 487 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 4 presents experiments on the effect of different combinations of teacher models. $\\lambda_{2}$ in Equation (7) determines using which teacher model as the primary. Using the intermediate expert model as the primary on regions overlapping with gt $(\\lambda_{2} = 0.5)$ outperforms using the old model $(\\lambda_{2} = 1.0)$ , showing that distilling new class objects with the old model hinders their learning. For box distillation, we find that distilling boxes on primary classes (part boxes, i.e., only boxes for old classes in $\\mathcal{R}_1$ and boxes for new classes in $\\mathcal{R}_2$ participate in distillation while others are ignored) performs similarly with all classes (all boxes).", + "bbox": [ + 212, + 489, + 785, + 609 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Visualization", + "text_level": 1, + "bbox": [ + 215, + 637, + 370, + 651 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Visualization of Detection Results. We visualize the detection results in Figure 5 to illustrate the significant improvement compared to the previous methods qualitatively. Our method accurately detects both new and old class objects simultaneously. While MMA fails to detect old classes accurately, suffering from catastrophic forgetting.", + "bbox": [ + 212, + 666, + 784, + 742 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Visualization of Bridge Past and Future. We visualized the module of Bridge the Past and Bridge the Future separately in Figure 6. We effectively model the missing annotations for old class objects to bridge the past. As demonstrated in Figure 6(b), the attention maps clearly differentiate the foreground from the background, regardless of the presence of annotations. The discarded boxes (the third row) validate the effectiveness of our method.", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Bridge Past and Future", + "bbox": [ + 571, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d2dc35795df87af9ec32adb2e58f457e3b6338d686352c39360985d616d1abd6.jpg", + "image_caption": [ + "Fig. 6: Visualization of Bridge Past and Future. Boxes in red represent the ground truth in the current stage. (a) In Bridge the Past, we effectively constructed pseudo labels of past classes. (b) In Bridge the Future, salient objects (marked in green boxes) can be easily detected from the attention maps and are excluded from the background regions. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 496, + 358 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/b041acdd977e3de8e6adaf39f524c1d89cbc9da25a537bfbc8212a357c10c8c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 143, + 784, + 357 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 463, + 367, + 479 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. In our Bridge the Past procedure, we assume several objects of old classes may appear in the current training data. However, when the number of incremental classes is limited, e.g., increasing a single class in each stage, objects of old classes may rarely occur due to limited training images. This is expected to be alleviated by generating samples of old classes via the copy-paste strategy as in ABR [41], while it may introduce little stored samples. Detailed discussion can be found in the Supplement.", + "bbox": [ + 212, + 496, + 787, + 603 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Conclusions. In this work, we find that the concurrence of classes from different learning stages causes a severe information asymmetry, not only causing catastrophic forgetting for old classes but also hindering the learning of new classes. To tackle the problem, we propose the Bridge Past and Future method, which uses pseudo labels from the old model to fill in the missing annotations and exclude some potential future objects from the background, keeping the learning consistent across all stages. Further, Distillation with Future loss is proposed to solve the problem of the old teacher model's lack of knowledge of new classes. By combining the knowledge from the past to the future, our method consistently outperforms others across different learning stages in most incremental settings. To the best of our knowledge, we are the first to consider future classes during the learning, shedding light on a new aspect of Incremental Object Detection.", + "bbox": [ + 212, + 607, + 787, + 790 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported partially by the National Key Research and Development Program of China (2023YFA1008503), NSFC(U21A20471), Guangdong NSF Project (No. 2023B1515040025, 2020B1515120085).", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Q. Mo et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Acharya, M., Hayes, T.L., Kanan, C.: Rodeo: Replay for online object detection. arXiv preprint arXiv:2008.06439 (2020)", + "2. Cao, S., Joshi, D., Gui, L.Y., Wang, Y.X.: Contrastive mean teacher for domain adaptive object detectors. In: CVPR (2023)", + "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020)", + "4. Castro, F.M., Marin-Jiménez, M.J., Guil, N., Schmid, C., Alahari, K.: End-to-end incremental learning. In: ECCV (2018)", + "5. Cermelli, F., Geraci, A., Fontanel, D., Caputo, B.: Modeling missing annotations for incremental learning in object detection. In: CVPR (2022)", + "6. Chaudhry, A., Dokania, P.K., Ajanthan, T., Torr, P.H.: Riemannian walk for incremental learning: Understanding forgetting and intransigence. In: ECCV (2018)", + "7. Chen, L., Yu, C., Chen, L.: A new knowledge distillation for incremental object detection. In: IJCNN (2019)", + "8. Dhar, P., Singh, R.V., Peng, K.C., Wu, Z., Chellappa, R.: Learning without memorizing. In: CVPR (2019)", + "9. Dong, N., Zhang, Y., Ding, M., Bai, Y.: Class-incremental object detection. PR (2023)", + "0. Douillard, A., Cord, M., Ollion, C., Robert, T., Valle, E.: Podnet: Pooled outputs distillation for small-tasks incremental learning. In: ECCV (2020)", + "1. Douillard, A., Rame, A., Couairon, G., Cord, M.: Dytox: Transformers for continual learning with dynamic token expansion. In: CVPR (2022)", + "2. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. Springer IJCV (2010)", + "3. Feng, T., Wang, M., Yuan, H.: Overcoming catastrophic forgetting in incremental object detection via elastic response distillation. In: CVPR (2022)", + "4. Fu, S., Yan, J., Gao, Y., Xie, X., Zheng, W.S.: Asag: Building strong one-decoder-layer sparse detectors via adaptive sparse anchor generation. In: ICCV (2023)", + "5. Gao, Y., Lin, K.Y., Yan, J., Wang, Y., Zheng, W.S.: Asyfod: An asymmetric adaptation paradigm for few-shot domain adaptive object detection. In: CVPR (2023)", + "6. Gao, Y., Yang, L., Huang, Y., Xie, S., Li, S., Zheng, W.S.: Acrofod: An adaptive method for cross-domain few-shot object detection. In: ECCV (2022)", + "7. Gao, Z., Wang, L., Han, B., Guo, S.: Adamixer: A fast-converging query-based object detector. In: CVPR (2022)", + "8. Girshick, R.: Fast r-cnn. In: ICCV (2015)", + "9. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921 (2021)", + "20. Gupta, A., Narayan, S., Joseph, K., Khan, S., Khan, F.S., Shah, M.: Ow-detr: Open-world detection transformer. In: CVPR (2022)", + "21. Hao, Y., Fu, Y., Jiang, Y.G., Tian, Q.: An end-to-end architecture for class incremental object detection with knowledge distillation. In: ICME (2019)", + "22. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask r-cnn. In: ICCV (2017).", + "23. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)", + "24. Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incrementally via rebalancing. In: CVPR (2019)", + "25. Joseph, K., Khan, S., Khan, F.S., Balasubramanian, V.N.: Towards open world object detection. In: CVPR (2021)" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Bridge Past and Future", + "bbox": [ + 571, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "26. Joseph, K., Rajasegaran, J., Khan, S., Khan, F.S., Balasubramanian, V.N.: Incremental object detection via meta-learning. IEEE TPAMI (2021)", + "27. Kang, B., Liu, Z., Wang, X., Yu, F., Feng, J., Darrell, T.: Few-shot object detection via feature reweighting. In: ICCV (2019)", + "28. Kang, M., Zhang, J., Zhang, J., Wang, X., Chen, Y., Ma, Z., Huang, X.: Alleviating catastrophic forgetting of incremental object detection via within-class and between-class knowledge distillation. In: CVPR (2023)", + "29. Kemker, R., Kanan, C.: Fearnet: Brain-inspired model for incremental learning. arXiv preprint arXiv:1711.10563 (2017)", + "30. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences (2017)", + "31. Li, D., Tasci, S., Ghosh, S., Zhu, J., Zhang, J., Heck, L.: Rilod: Near real-time incremental learning for object detection at the edge. In: Proceedings of the 4th ACM/IEEE Symposium on Edge Computing (2019)", + "32. Li, M., Cong, Y., Liu, Y., Sun, G.: Class-incremental gesture recognition learning with out-of-distribution detection. In: IROS (2022)", + "33. Li, Y.M., Zeng, L.A., Meng, J.K., Zheng, W.S.: Continual action assessment via task-consistent score-discriminative feature distribution modeling. TCVT (2024)", + "34. Li, Z., Hoiem, D.: Learning without forgetting. IEEE TPAMI (2017)", + "35. Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: CVPR (2017)", + "36. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: ICCV (2017)", + "37. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014)", + "38. Liu, L., Kuang, Z., Chen, Y., Xue, J.H., Yang, W., Zhang, W.: Incdet: In defense of elastic weight consolidation for incremental object detection. IEEE TNNLS (2020)", + "39. Liu, X., Yang, H., Ravichandran, A., Bhotika, R., Soatto, S.: Multi-task incremental learning for object detection. arXiv preprint arXiv:2002.05347 (2020)", + "40. Liu, Y., Schiele, B., Vedaldi, A., Rupprecht, C.: Continual detection transformer for incremental object detection. In: CVPR (2023)", + "41. Liu, Y., Cong, Y., Goswami, D., Liu, X., van de Weijer, J.: Augmented box replay: Overcoming foreground shift for incremental object detection. In: ICCV (2023)", + "42. Liu, Y., Cong, Y., Sun, G., Zhang, T., Dong, J., Liu, H.: L3doc: Lifelong 3d object classification. IEEE TIP (2021)", + "43. Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. In: NeurIPS (2017)", + "44. Ma, J., Niu, Y., Xu, J., Huang, S., Han, G., Chang, S.F.: Digeo: Discriminative geometry-aware learning for generalized few-shot object detection. In: CVPR (2023)", + "45. Mallya, A., Davis, D., Lazebnik, S.: Piggyback: Adapting a single network to multiple tasks by learning to mask weights. In: ECCV (2018)", + "46. Mallya, A., Lazebnik, S.: Packnet: Adding multiple tasks to a single network by iterative pruning. In: CVPR (2018)", + "47. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., et al.: Simple open-vocabulary object detection. In: ECCV (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Q. Mo et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Ostapenko, O., Puscas, M., Klein, T., Jahnichen, P., Nabi, M.: Learning to remember: A synaptic plasticity driven framework for continual learning. In: CVPR (2019)", + "49. Peng, C., Zhao, K., Lovell, B.C.: Faster ilod: Incremental learning for object detectors based on faster rcnn. PR (2020)", + "50. Peng, C., Zhao, K., Maksoud, S., Wang, T., Lovell, B.C.: Diode: dilatable incremental object detection. PR (2023)", + "51. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: CVPR (2017)", + "52. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018)", + "53. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. In: NeurIPS (2015)", + "54. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016)", + "55. Shieh, J.L., Haq, Q.M.u., Haq, M.A., Karam, S., Chondro, P., Gao, D.Q., Ruan, S.J.: Continual learning strategy in one-stage object detection framework based on experience replay for autonomous driving vehicle. Sensors (2020)", + "56. Shin, H., Lee, J.K., Kim, J., Kim, J.: Continual learning with deep generative replay. In: NeurIPS (2017)", + "57. Shmelkov, K., Schmid, C., Alahari, K.: Incremental learning of object detectors without catastrophic forgetting. In: ICCV (2017)", + "58. Tan, M., Pang, R., Le, Q.V.: Efficientdet: Scalable and efficient object detection. In: CVPR (2020)", + "59. Tang, Y.M., Peng, Y.X., Zheng, W.S.: Learning to imagine: Diversify memory for incremental learning using unlabeled data. In: CVPR (2022)", + "60. Tang, Y.M., Peng, Y.X., Zheng, W.S.: When prompt-based incremental learning does not meet strong pretraining. In: ICCV (2023)", + "61. Tian, Z., Shen, C., Chen, H., He, T.: Fcos: Fully convolutional one-stage object detection. In: ICCV (2019)", + "62. Wu, C., Herranz, L., Liu, X., Van De Weijer, J., Raducanu, B., et al.: Memory replay gans: Learning to generate new categories without forgetting. In: NeurIPS (2018)", + "63. Yan, J., Yang, L., Gao, Y., Zheng, W.S.: Self-supervised cross-stage regional contrastive learning for object detection. In: ICME (2023)", + "64. Yang, D., Zhou, Y., Hong, X., Zhang, A., Wei, X., Zeng, L., Qiao, Z., Wang, W.: Pseudo object replay and mining for incremental object detection. In: ACM MM (2023)", + "65. Yang, D., Zhou, Y., Zhang, A., Sun, X., Wu, D., Wang, W., Ye, Q.: Multi-view correlation distillation for incremental object detection. PR (2022)", + "66. Zenke, F., Poole, B., Ganguli, S.: Continual learning through synaptic intelligence. In: ICML (2017)", + "67. Zhou, W., Chang, S., Sosa, N., Hamann, H., Cox, D.: Lifelong object detection. arXiv preprint arXiv:2009.01129 (2020)", + "68. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: ICLR (2021)", + "69. Zohar, O., Wang, K.C., Yeung, S.: Prob: Probabilistic objectness for open world object detection. In: CVPR (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 825 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Bridge Past and Future", + "bbox": [ + 571, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Bridge Past and Future_ Overcoming Information Asymmetry in Incremental Object Detection/1c449442-a7f3-4df8-9bc6-973b26235638_model.json b/2024/Bridge Past and Future_ Overcoming Information Asymmetry in Incremental Object Detection/1c449442-a7f3-4df8-9bc6-973b26235638_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8d160810b4b1dc96593fd950d9d7713f9fbd3746 --- /dev/null +++ b/2024/Bridge Past and Future_ Overcoming Information Asymmetry in Incremental Object Detection/1c449442-a7f3-4df8-9bc6-973b26235638_model.json @@ -0,0 +1,2577 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.22, + 0.141, + 0.784, + 0.187 + ], + "angle": 0, + "content": "Bridge Past and Future: Overcoming Information Asymmetry in Incremental Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.212, + 0.744, + 0.245 + ], + "angle": 0, + "content": "Qijie Mo\\(^{1,3}\\), Yipeng Gao\\(^{1,3}\\), Shenghao Fu\\(^{1,3}\\), Junkai Yan\\(^{1,3}\\), Ancong Wu\\(^{1,3,\\star}\\), and Wei-Shi Zheng\\(^{1,2,3,\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.255, + 0.763, + 0.282 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Computer Science and Engineering, Sun Yat-sen University, China \n\\(^{2}\\) Peng Cheng Laboratory, Shenzhen, China" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.283, + 0.732, + 0.31 + ], + "angle": 0, + "content": "3 Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.312, + 0.686, + 0.34 + ], + "angle": 0, + "content": "{moqj3,gaoyp23,fushh7,yanjk3}@mail2.sysu.edu.cn,wuanc@mail.sysu.edu.cn, wszheng@ieee.org" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.377, + 0.744, + 0.613 + ], + "angle": 0, + "content": "Abstract. In incremental object detection, knowledge distillation has been proven to be an effective way to alleviate catastrophic forgetting. However, previous works focused on preserving the knowledge of old models, ignoring that images could simultaneously contain categories from past, present, and future stages. The co-occurrence of objects makes the optimization objectives inconsistent across different stages since the definition for foreground objects differs across various stages, which limits the model's performance greatly. To overcome this problem, we propose a method called \"Bridge Past and Future\" (BPF), which aligns models across stages, ensuring consistent optimization directions. In addition, we propose a novel Distillation with Future (DwF) loss, fully leveraging the background probability to mitigate the forgetting of old classes while ensuring a high level of adaptability in learning new classes. Extensive experiments are conducted on both Pascal VOC and MS COCO benchmarks. Without memory, BPF outperforms current state-of-the-art methods under various settings. The code is available at https://github.com/iSEE-Laboratory/BPF." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.625, + 0.741, + 0.652 + ], + "angle": 0, + "content": "Keywords: Object Detection \\(\\cdot\\) Incremental Learning \\(\\cdot\\) Knowledge Distillation" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.679, + 0.377, + 0.696 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.711, + 0.788, + 0.818 + ], + "angle": 0, + "content": "Object detection [3,14,36,53,61,68] is a fundamental computer vision task and has significantly given rise to many sub-directions [2, 15, 16, 19, 27, 44, 47, 63]. General detection models typically assume full access to all interest classes during training and require exhaustive labeled data from the start. However, in dynamic real-world applications, where new object categories may appear continually, object detection models are expected to adapt to these ongoing changing classes, especially when there is limited storage to save all the data from the beginning" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.387, + 0.84 + ], + "angle": 0, + "content": "* Corresponding Author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Q. Mo et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.44, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.249, + 0.439, + 0.27 + ], + "angle": 0, + "content": "(a) Distribution of instances in the second stage of the 10-5-5 setting on VOC 2007 dataset" + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.147, + 0.613, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.484, + 0.249, + 0.74, + 0.262 + ], + "angle": 0, + "content": "(b) Visualization of some typical training samples in IOD" + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.148, + 0.785, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.281, + 0.788, + 0.325 + ], + "angle": 0, + "content": "Fig. 1: Challenges faced by the IOD task. Classes from previous, current, and potential future stages may appear arbitrarily in the current training stage, while only the annotations of current classes are available to train the detector. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.784, + 0.414 + ], + "angle": 0, + "content": "to the present or when data privacy concerns are encountered. To this end, incremental object detection (IOD) [5,26,41,65,67,69] has caught progressive attention, which aims to continually detect new objects without forgetting the previously learned ones." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.789, + 0.719 + ], + "angle": 0, + "content": "Unlike incremental learning in classification, the incremental object detection task faces a dual challenge caused by the concurrence of different classes: (1) Foreground objects identified in previous stages are unlabeled in the current stage, and thus, the model currently trained may regard them as background. (2) The current stage's background may also contain objects that will be recognized as the foreground in future stages. Taking the 10-5-5 setting on the VOC dataset [12] as an example, as shown in Figure 1, a large proportion of instances belonging to classes of the previous and future stages occur concurrently in the training images in the current stage. This asymmetry in information between the past and current stages causes the current model to erroneously classify objects of old categories as background, aggravating the catastrophic forgetting problem. Similarly, the asymmetry between the current and future stages leads the current model to classify objects of potential future categories as background erroneously, which requires models trained in future stages to rectify their misperception, thus hindering the learning of new classes. As a result, the optimization objectives across different learning stages are inconsistent, significantly limiting the generalization of current incremental detectors. Previous works [5, 65, 67] concentrate on introducing strong regularization to prevent catastrophic forgetting but ignore the impact of concurrence of different classes, resulting in limited performance on both old and new classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This work aims to tackle the critical information asymmetry challenge in IOD by utilizing the abundant concurrence information within an image. We introduce a novel approach named Bridge Past and Future (BPF), which is designed to connect the past and future stages with the current stage, ensuring the model adheres to consistent optimization objectives throughout the entire incremental learning process. For the past stages, we utilize the previous model as a reliable labeler to construct some pseudo labels for old classes and combine them with current annotations for training the current model symmetrically." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.572, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bridge Past and Future" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "Simultaneously, we re-identify some salient regions that may contain potential future objects out of the background and exclude them from negative samples in the current stage to avoid classifying them into the background at this stage and disturbing future learning, which is expected to achieve symmetry with future stages. These two novel designs alleviate the impact of inconsistent training objectives across different stages from the aspects of the past and the future, thus easing the training difficulty of incremental detectors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.373 + ], + "angle": 0, + "content": "In addition, we propose a distillation technique to augment the detector in the current stage, where we take a step back by viewing the current stage as a clear future stage for previous stages and propose a Distillation with Future (DwF) loss. The distillation is carried out by two teachers, i.e., the detector trained on the previous stage and an expert detector trained with only current data. The student detector absorbs the knowledge of old and current classes from the old and expert models, respectively, in a class-by-class manner, preventing the detector from catastrophic forgetting and facilitating learning of current classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.449 + ], + "angle": 0, + "content": "We conduct extensive experiments on the PASCAL VOC [12] and COCO [37] datasets to evaluate the effectiveness of our BPF, which outperforms other state-of-the-art (SOTA) methods under multiple settings in a memory-free way. Moreover, we also conduct comprehensive ablation studies and visualizations to help better understand how each component of the BPF works." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.471, + 0.398, + 0.487 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.788, + 0.669 + ], + "angle": 0, + "content": "Object Detection. Traditional object detectors can be primarily classified into two streams: one-stage [36,52,58,61] and two-stage detectors [18,22,35,53]. Two-stage detectors first predict several coarse candidate proposals via region proposal extractors [18,53] and then adopt a region of interest (RoI) head to refine these proposals and output final predictions. Unlike two-stage detectors, one-stage counterparts directly generate final outputs without predicting candidate proposals. Despite performing well under the standard training setup, both of them fail to generalize to the incremental training setup due to lacking previous training data. Without losing generality, this paper concentrates on enhancing two-stage detectors, i.e., Faster R-CNN [53], enabling it to learn new classes incrementally while retaining previously acquired knowledge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Incremental Learning. Incremental learning methods can be mainly divided into three categories: memory-based [4, 24, 32, 33, 48, 51, 56, 59, 62], regularization-based [6, 8, 30, 34, 66], and structure-based [42, 45, 46, 54, 60]. Memory-based methods store a handful of samples for replaying [24, 40, 41] or suggest generating samples [29, 62] for data compensation in new stages. Some regularization-based methods try to eliminate the effect of new tasks on previously learned knowledge by identifying the most crucial parameters [30, 43, 62], while some other regularization-based methods propose to distill knowledge [23] from old to current models by transiting knowledge on logits, final features or intermediate features [8, 10, 11, 24, 34]. Structure-based methods dedicate specific parameters to each task, freezing them to mitigate forgetting. In this study, we focus on the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Q. Mo et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.177 + ], + "angle": 0, + "content": "regularization-based knowledge distillation approach for object detection and introduce a novel distillation with future information." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.183, + 0.787, + 0.426 + ], + "angle": 0, + "content": "Incremental Object Detection. Most previous works [9,28,31,40,50,55] focus on distilling old classes' knowledge of intermediate features [5,21,39,41,49,65,67], region proposal network [5, 49, 67], or RoI head [13] to prevent forgetting. A stream of methods also pays attention to classification loss. PPAS [67] introduces a pseudo-positive-aware sampling algorithm to identify regions corresponding to old classes and prevents them from being sampled as background. MMA [5] proposes unbiased classification loss, consolidating the background and all old classes into one entity, which aims to minimize the optimization objective conflict between the current stage's background and the past stages' foreground. However, it diminishes the ability to distinguish old classes from the background. Other methods [1,20,25,26,41,64] focus on rehearsal to maintain the previous stage knowledge, either performing replay of the intermediate features [1], the images [25,26,40], or the instances [41]. Incdet [38] presents a parameter isolation strategy that builds upon EWC [30]. In this study, in addition to preserving knowledge for old classes, we also take future classes into account, aligning optimization objectives consistently across all stages." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.445, + 0.33, + 0.46 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.473, + 0.373, + 0.488 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.495, + 0.787, + 0.662 + ], + "angle": 0, + "content": "Problem Formulation. In incremental object detection, the training is performed over multiple learning stages, each one introducing a new set of classes to be detected. Let \\(\\mathcal{C}\\) denote the set of classes that are incrementally introduced to the object detector \\(\\mathcal{M}\\). In the \\(t\\)-th training stage, a grouping of classes \\(\\mathcal{C}_t\\) are introduced to the detector: \\(\\mathcal{C}_t \\subset \\mathcal{C}\\), such that \\(\\mathcal{C}_i \\cap \\mathcal{C}_j = \\emptyset\\), for any \\(i \\neq j\\) and \\(i, j \\leq t\\). Let \\(\\mathcal{D}_t\\) denote the images containing annotated objects of classes in \\(\\mathcal{C}_t\\). Each image can contain multiple objects of different classes, but annotations \\(\\mathcal{V}_t\\) are available only for those object instances that belong to classes in \\(\\mathcal{C}_t\\). The challenge in class incremental object detection is to continually update \\(\\mathcal{M}_t\\) to \\(\\mathcal{M}_{t+1}\\) by learning \\(\\mathcal{C}_{t+1}\\), without access to \\(\\{\\mathcal{D}_0, \\dots, \\mathcal{D}_t\\}\\) while maintaining original performance on \\(\\{\\mathcal{C}_0, \\dots, \\mathcal{C}_t\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.668, + 0.787, + 0.804 + ], + "angle": 0, + "content": "General Detector Training Pipeline. This work starts with the representative Faster RCNN-like detectors [53] \\( f = \\{f_b, f_{rpn}, f_{roi}\\} \\), which generally includes a backbone network \\( f_b \\), Region Proposal Network (RPN) \\( f_{rpn} \\) and Region of Interest Head (RoI Head) \\( f_{roi} \\). During training, all training proposals are categorized into positive, negative, and ignored samples for supervised training based on their IoUs with the ground truth. Positive samples are assigned to predict the class of their matched ground truth, whereas negative samples are designated as background. Thus, the objects without annotations in the current stage will be classified as background, hindering the learning of incremental models." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Common Distillation Methods for Incremental Object Detection. To prevent forgetting in IOD, a widely adopted solution involves knowledge distil-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.572, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bridge Past and Future" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.519, + 0.163 + ], + "angle": 0, + "content": "lation in RPN and RoI head [5,49,57,67]:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.17, + 0.786, + 0.188 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i s t}} = \\mathcal {L} _ {\\text {d i s t}} ^ {r p n} + \\mathcal {L} _ {\\text {d i s t}} ^ {r o i}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.193, + 0.784, + 0.239 + ], + "angle": 0, + "content": "The distillation in the RoI head includes the L2 loss between the box coordinates and the Kullback-Leibler divergence for class probabilities. MMA [5] notices the missing old annotations and proposes Unbiased Knowledge Distillation (UKD):" + }, + { + "type": "equation", + "bbox": [ + 0.228, + 0.246, + 0.786, + 0.282 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i s t}, \\text {c l s}} ^ {\\text {r o i}} (i) = \\frac {1}{| \\mathcal {C} _ {t - 1} | + 1} \\left(p _ {i} ^ {b, t - 1} \\log \\left(p _ {i} ^ {b, t} + \\sum_ {c \\in \\mathcal {C} _ {t}} p _ {i} ^ {c, t}\\right) + \\sum_ {c \\in \\mathcal {C} _ {1: t - 1}} p _ {i} ^ {c, t - 1} \\log \\left(p _ {i} ^ {c, t}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.29, + 0.788, + 0.458 + ], + "angle": 0, + "content": "where \\( p_i^{c,t-1} \\) and \\( p_i^{c,t} \\) indicate the classification output for the proposal \\( i \\) and class \\( c \\) of the teacher (old) model and student (current) model, respectively, and \\( b \\) is the background class. As a common practice, 64 proposals are used for distillation, which are randomly selected out of the top 128 proposals with the highest objectness scores from the RPN network of the old model. However, given the class-agnostic characteristic of the RPN, the proposals for distillation cover objects of both old and current classes. In such cases, the old model, without seeing current classes, cannot provide useful knowledge for the current classes and even hinders their learning. Further, treating the background probability and current class probabilities as a unified entity during distillation inevitably diminishes the ability to differentiate between current classes and background." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.477, + 0.421, + 0.491 + ], + "angle": 0, + "content": "3.2 Overall Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.784, + 0.573 + ], + "angle": 0, + "content": "Unlike classification tasks with a single label per input, incremental object detection presents a scenario where an image \\(\\mathcal{I}_t\\) can encompass objects from the current class set \\(\\mathcal{C}_t\\), previous class sets \\(\\mathcal{C}_{1:t-1}\\) and future class sets \\(\\mathcal{C}_{t+1:\\infty}\\). Nonetheless, the annotations \\(\\mathcal{Y}_t\\) are limited to bounding boxes and class labels for objects in \\(\\mathcal{C}_t\\), while objects of other classes are regarded as background." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.574, + 0.788, + 0.755 + ], + "angle": 0, + "content": "To address the challenge of inconsistent optimization objectives, we introduce a novel strategy termed Bridge Past and Future (BPF). As shown in Figure 2, the method is divided into two parts. From the perspective of supervised learning, we use high-confidence predictions of old classes given by the old model as pseudo labels to bridge the information from the past stage (Section 3.3) and discard some potential objects from the background to bridge the future stage (Section 3.4). Regarding the distillation learning, we propose a novel Distillation with Future (DwF) loss by considering current classes in the distillation. The distillation probabilities are combined from two teacher models: the old model \\(\\mathcal{M}_{t-1}\\) for old classes and an expert model for current classes. By utilizing abundant information in background probability, the distillation takes both previous stages and the current stage into consideration (Section 3.5)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.773, + 0.41, + 0.789 + ], + "angle": 0, + "content": "3.3 Bridging the Past" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.841 + ], + "angle": 0, + "content": "To address the inconsistency in optimization objectives between the previous and current models, we bridge the past information to the current stage by incorporating pseudo supervision signals from past stages into the current model's" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Q. Mo et al." + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.148, + 0.752, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.384, + 0.789, + 0.482 + ], + "angle": 0, + "content": "Fig. 2: The overall framework of our method. The top side illustrates the Bridge Past and Future (BPF) procedure, which identifies objects of past classes and excludes several potential objects of future classes to ensure consistent optimization during the entire training process. The bottom side shows the Distillation with Future (DwF) process, which employs both the old model \\(\\mathcal{M}_{t-1}\\) adept at detecting old categories and the interim model \\(\\mathcal{M}_t^{im}\\) trained on \\(\\mathcal{D}_t\\) and specialized in new categories, to conduct a comprehensive distillation across all categories for the current model \\(\\mathcal{M}_t\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.789, + 0.602 + ], + "angle": 0, + "content": "optimization process. The old model was trained under the supervision of human annotations, which is a high-quality pseudo labeler for old classes. We take the high-confidence inference results of the old model \\(\\mathcal{M}_{t-1}\\) on current training images as pseudo labels and combine them with current annotations to train the current detector, thereby aligning the supervision of old classes with previous stages. The left part of Figure 3 shows the pipeline of bridging the past strategy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.602, + 0.789, + 0.663 + ], + "angle": 0, + "content": "Formally, given predictions \\(\\hat{y}^{old} \\in \\mathcal{P}\\) of the old model \\(\\mathcal{M}_{t-1}\\) on current images, we first select some high-confidence regions of previous classes \\(\\mathcal{C}_{1:t-1}\\) with a confidence threshold \\(\\eta\\), followed by the Non-Maximum Suppression (NMS) operation to reduce duplication:" + }, + { + "type": "equation", + "bbox": [ + 0.319, + 0.672, + 0.787, + 0.697 + ], + "angle": 0, + "content": "\\[\n\\mathcal {U} = \\{j \\in \\mathcal {P}: \\max _ {c \\in \\mathcal {C} _ {1: t - 1}} \\hat {p} _ {j} ^ {\\text {o l d}} (c) > \\eta \\}, \\quad \\mathcal {E} = \\mathrm {N M S} (\\mathcal {U}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.707, + 0.789, + 0.754 + ], + "angle": 0, + "content": "Then, we further narrow down the predictions to a subset \\(\\mathcal{W} \\subset \\mathcal{E}\\) that do not overlap with the ground-truth labels of the new categories via an IoU threshold \\(\\lambda_{1}\\), ensuring a clear distinction between past and present classes:" + }, + { + "type": "equation", + "bbox": [ + 0.35, + 0.763, + 0.787, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathcal {W} = \\{j \\in \\mathcal {E}: \\forall i \\in \\mathcal {Y} _ {t}, \\operatorname {I o U} \\left(\\hat {\\boldsymbol {b}} _ {j} ^ {\\text {o l d}}, \\boldsymbol {b} _ {i}\\right) \\leq \\lambda_ {1} \\}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.843 + ], + "angle": 0, + "content": "By modeling the previous supervision signals from the old model, we bridge the past stages to the current one, ensuring the current model's optimization direction encompasses the objectives of earlier stages, significantly mitigating" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.572, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bridge Past and Future" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.146, + 0.729, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.306, + 0.788, + 0.377 + ], + "angle": 0, + "content": "Fig. 3: Overview of Bridge Past and Future. We adopt the previous model \\(\\mathcal{M}_{t-1}\\) to predict some pseudo labels for past classes to complement their missing supervision in the current stage. Additionally, we exclude several proposals that are likely to be an object but are not included in the current ground truth and pseudo labels from the background to avoid classifying them into background mistakenly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.784, + 0.468 + ], + "angle": 0, + "content": "the forgetting problem. MMA [5] treats the background probability and old class probabilities as a unified entity in the classification loss, making old classes hard to separate from the background. While our method explicitly models the old classes in the current stage out of the background." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.491, + 0.428, + 0.507 + ], + "angle": 0, + "content": "3.4 Bridging the Future" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.785, + 0.622 + ], + "angle": 0, + "content": "Bridging the future aims to rectify the misalignment between the optimization goals of the current and upcoming stages, as objects from future class sets \\(\\mathcal{C}_{t + 1:\\infty}\\) in the current dataset \\(\\mathcal{D}_t\\) are classified as background. The core design is to find some salient objects in the background and exclude them from the negative samples during the training. By separating pure background from regions likely to contain future category objects in the current stage, we align the optimization objectives for background treatment across both current and future models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.785, + 0.7 + ], + "angle": 0, + "content": "Specifically, we find that the activation in feature maps is a good indicator for salient objects. As illustrated in Figure 3, in the absence of annotations for old and future classes, foreground and background features demonstrate notable differences in spatial attention. We produce the attention map \\( A_{i} \\in \\mathbb{R}^{H \\times W} \\) from the backbone feature map \\( F_{i} = f_{b}(I_{i}) \\in \\mathbb{R}^{H \\times W \\times C} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.712, + 0.786, + 0.754 + ], + "angle": 0, + "content": "\\[\nA _ {i} = \\operatorname {S o f t m a x} \\left(\\sum_ {c = 1} ^ {C} \\left| F _ {i} \\right| ^ {p}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.763, + 0.784, + 0.811 + ], + "angle": 0, + "content": "where \\( H \\), \\( W \\), and \\( C \\) denote the feature's height, width, and channel. We compute the attention score for each region, indicating the likelihood of containing foreground objects. The attention score \\( a_{i,j}^{roi} \\) for region \\( r_j \\) is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.824, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\na _ {i, j} ^ {\\text {r o i}} = \\operatorname {A v g} \\left(\\operatorname {R o I P o o l} \\left(A _ {i}, r _ {j}\\right)\\right) \\in \\mathbb {R}. \\tag {6}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Q. Mo et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.146, + 0.784, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.381, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Fig. 4: The illustration of Distillation with Future strategy. An intermediate teacher model trained on the current dataset is used to compensate for the lack of current class information in the old model. For proposals overlapping ground truth of the current stage, since the intermediate teacher is specialized in detecting current classes, we directly inherit its probabilities on current classes and use the old model to enrich its background probability with old class knowledge. On the contrary, for proposals that do not overlap with GT, the old model is preferred, and the intermediate model is used as compensation. Combining two teachers makes the distillation class by class." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.614 + ], + "angle": 0, + "content": "Regions with high attention scores \\( a_{i,j}^{roi} \\) from feature maps and objectness scores \\( o_j \\) from class-agnostic RPN imply a greater chance of being future category objects. We discard these RoIs when sampling negative samples when training the RoI head, thus maintaining the model's consistency with future stage background definitions. Regions with lower scores or having a considerable IoU with ground truth are considered reliable backgrounds." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.634, + 0.625, + 0.65 + ], + "angle": 0, + "content": "3.5 Combining Current Classes into Distillation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Distillation is an effective way to prevent forgetting in incremental object detection. However, distilling from the teacher model, which is biased toward old classes, inevitably hinders the learning for current classes. To address this challenge, we introduce the Distillation with Future (DwF) loss to distill from the teacher model in a more fine-grained and adaptive way. Different from MMA [5], which aligns the current classes and the background as a whole with the old model's background, we distill each class one by one, preserving the distinction between current classes and the background. However, the model \\(\\mathcal{M}_{t-1}\\) from the previous stage has not been trained on current classes, thus the distillation on current classes can not be performed. We introduce another intermediate teacher model \\(\\mathcal{M}_t^{im}\\), which is trained using the dataset from the current stage \\(\\mathcal{D}_t\\) in a fully supervised way, as a supplement. Taking the intermediate model \\(\\mathcal{M}_t^{im}\\) as a" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.572, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bridge Past and Future" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "complementary teacher, we explicitly consider the current stage \\( t \\) as the future stage of stage \\( t - 1 \\), making the distillation future-aware." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.253 + ], + "angle": 0, + "content": "As the old model \\(\\mathcal{M}_{t-1}\\) performs well on \\(\\mathcal{C}_{1:t-1}\\) while the intermediate model \\(\\mathcal{M}_t^{im}\\) is the expert in \\(\\mathcal{C}_t\\), we distill different regions with different combination of teacher models, as shown in Figure 4. Specifically, we divide the regions for distillation \\(\\mathcal{R}\\) into two subsets \\(\\mathcal{R}_1, \\mathcal{R}_2 \\subset \\mathcal{R}\\) that based on their intersection over union with the ground truth labels for the new categories \\(\\mathcal{C}_t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.263, + 0.65, + 0.279 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {1} = \\{j \\in \\mathcal {R}: \\forall i \\in \\mathcal {Y} _ {t}, \\operatorname {I o U} (b _ {j}, b _ {i}) \\leq \\lambda_ {2.} \\},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.27, + 0.785, + 0.297 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {R} _ {1} = \\left\\{j \\in \\mathcal {K}: \\forall i \\in \\mathcal {Y} _ {t}, \\operatorname {I o U} \\left(b _ {j}, b _ {i}\\right) \\geq \\lambda_ {2}. \\right\\}, \\\\ \\mathcal {R} _ {2} = \\left\\{j \\in \\mathcal {R}: \\forall i \\in \\mathcal {Y} _ {t}, \\operatorname {I o U} \\left(b _ {j}, b _ {i}\\right) > \\lambda_ {2}. \\right\\}. \\end{array} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.305, + 0.784, + 0.351 + ], + "angle": 0, + "content": "For regions \\( r_i \\in \\mathcal{R}_1 \\), which are likely to be the regions for old classes, we take \\( \\mathcal{M}_{t-1} \\) as the primary model for distillation and reconstruct its background representation with the model \\( \\mathcal{M}_t^{im} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.36, + 0.785, + 0.379 + ], + "angle": 0, + "content": "\\[\n\\hat {p} _ {i} ^ {c, i m} = p _ {i} ^ {c, i m} \\times p _ {i} ^ {b, t - 1}, \\quad r _ {i} \\in \\mathcal {R} _ {1}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.785, + 0.483 + ], + "angle": 0, + "content": "where \\( p_i^{b,t-1} \\) and \\( p_i^{c,im} \\) are the classification probabilities of background in the model \\( \\mathcal{M}_{t-1} \\) and the classification probabilities for current classes and background \\( c \\in \\mathcal{C}_t \\cup \\mathcal{B} \\) in the model \\( \\mathcal{M}_t^{im} \\) for the region \\( r_i \\) respectively. After weighting, \\( \\sum_{c=1}^{\\mathcal{C}_t \\cup \\mathcal{B}} \\hat{p}_i^{c,im} = p_i^{b,t-1} \\). The final distillation probabilities for regions in \\( \\mathcal{R}_1 \\) are \\( [p_i^{C_{1:t-1},t-1}, \\hat{p}_i^{C_t \\cup \\mathcal{B},im}] \\in \\mathbb{R}^{|C_{1:t}|+1} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.484, + 0.785, + 0.529 + ], + "angle": 0, + "content": "On the contrary, for regions \\( r_i \\in \\mathcal{R}_2 \\), they are regions for current classes; thus, the intermediate model \\( \\mathcal{M}_t^{im} \\) is taken as the primary model for distillation, and its background representation is reconstructed by the model \\( \\mathcal{M}_{t-1} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.537, + 0.785, + 0.555 + ], + "angle": 0, + "content": "\\[\n\\hat {p} _ {i} ^ {c, t - 1} = p _ {i} ^ {c, t - 1} \\times p _ {i} ^ {b, i m}, \\quad r _ {i} \\in \\mathcal {R} _ {2}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.784, + 0.615 + ], + "angle": 0, + "content": "where \\(\\sum_{c=1}^{\\mathcal{C}_{1:t-1} \\cup \\mathcal{B}} \\hat{p}_i^{c,t-1} = p_i^{b,im}\\). The distillation probabilities for regions in \\(\\mathcal{R}_2\\) are \\([\\hat{p}_i^{\\mathcal{C}_{1:t-1},t-1}, p_i^{\\mathcal{C}_t,im}, \\hat{p}_i^{\\mathcal{B},t-1}] \\in \\mathbb{R}^{|C_{1:t}|+1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.616, + 0.785, + 0.735 + ], + "angle": 0, + "content": "With the expanded probabilities from the complimentary teacher models, the distillation on the classification head can be performed using a conventional Kullback-Leibler divergence. Regarding the box distillation, we use the output boxes from the old model \\(\\mathcal{M}_{t-1}\\) for regions in \\(\\mathcal{R}_1\\) and the intermediate model \\(\\mathcal{M}_t^{im}\\) for \\(\\mathcal{R}_2\\). As a result, the complementary knowledge from two teachers, the expansion of background probability, and the combination of adaptive probability for different regions not only prevent the student from catastrophic forgetting but also facilitate the learning for current classes." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.758, + 0.376, + 0.775 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.787, + 0.432, + 0.803 + ], + "angle": 0, + "content": "4.1 Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Datasets and Evaluation Metrics. Following previous works [5,20,25,26,41, 49,57,65,67], we evaluate our method on PASCAL VOC 2007 [12] and MS COCO" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Q. Mo et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.787, + 0.187 + ], + "angle": 0, + "content": "Table 1: mAP@0.5 results on single incremental step on PASCAL VOC 2007. The best performance in each is presented with **bold**, and the second best is presented with **underline**. Methods with * use exemplars." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.199, + 0.789, + 0.389 + ], + "angle": 0, + "content": "
Method19-115-510-105-15
1-19201-20Avg1-1516-201-20Avg1-1011-201-20Avg1-55-151-20Avg
Joint Training76.076.776.176.478.070.476.174.275.976.376.176.172.477.376.174.9
Fine-tuning12.062.814.537.414.259.225.436.79.562.536.036.06.963.149.135.0
ORE* [25]69.460.168.964.771.858.768.565.260.468.864.664.6----
OW-DETR* [20]70.262.069.866.172.259.869.166.063.567.965.765.7----
ILOD-Meta* [26]70.957.670.264.271.755.967.863.868.464.366.366.3----
ABR* [41]71.069.770.970.473.065.171.069.171.272.872.072.064.771.069.467.9
Faster ILOD [53]68.961.168.565.071.656.967.964.369.854.562.162.162.037.143.349.6
PPAS [67]70.553.069.261.8----63.560.061.861.8----
MVC [65]70.260.669.765.469.457.966.563.766.266.066.166.1----
PROB [69]73.948.572.661.573.560.870.167.066.067.266.566.5----
PseudoRM [64]72.967.372.670.173.460.970.366.969.168.668.968.9----
MMA [5]71.163.470.767.273.060.569.966.769.363.966.666.666.857.259.662.0
BPF (Ours)74.565.374.169.975.963.072.769.571.774.072.972.966.475.373.070.9
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.421, + 0.787, + 0.511 + ], + "angle": 0, + "content": "2017 [37] datasets. PASCAL VOC 2007 dataset comprises 9,963 images across 20 categories. The COCO 2017 dataset encompasses objects from 80 categories, with around \\(118\\mathrm{k}\\) images for training and 5,000 images for validation. The mean average precision at the 0.5 IoU threshold \\((\\mathrm{mAP}@0.5)\\) is used as the primary evaluation metric for the VOC dataset, and the mean average precision ranging from 0.5 to 0.95 is the main evaluation metric for the COCO dataset." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.513, + 0.787, + 0.574 + ], + "angle": 0, + "content": "For each incremental setting (A-B), the first number A denotes the number of classes in the first stage and the second number is the number of classes newly introduced in each new stage. Note that the columns with gray background in the table of experimental results represent the average AP among all classes." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.581, + 0.787, + 0.687 + ], + "angle": 0, + "content": "Implementation Details. Similar to [5,25,26,41,49], we build our incremental object detector based on Faster R-CNN [53] with R50. Our method can easily be adapted to transformer-based detectors [3,17,68]. We conduct the experiments under a strict rehearsal-free setting, where no memory is used. We set \\(\\eta = 0.75\\), \\(\\lambda_{1} = 0.7\\), \\(\\lambda_{2} = 0.5\\). For \\(\\mathcal{W}\\) in Equation (4), we use an IOU threshold of 0.3 to divide it into two sets. For the set with IoU \\(< 0.3\\), the supervision signal weights are set to 1.0, while the other set is assigned a weight of 0.3." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.717, + 0.459, + 0.733 + ], + "angle": 0, + "content": "4.2 Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.749, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Following previous work [5,7,41,49,57,65,67], our method is evaluated on settings with a range of initial classes and incorporating one or more incremental tasks. We benchmark our method against two baselines: Fine-Tuning, where the model is incrementally trained on new data without any regularization strategy or data replay, and Joint Training, which involves training the model on the complete dataset using all annotations." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.572, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bridge Past and Future" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: mAP@0.5 results on multiple incremental steps on PASCAL VOC 2007. The best performance in each is presented with **bold**, and the second best is presented with **underline**. Methods with * use exemplars." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.2, + 0.787, + 0.297 + ], + "angle": 0, + "content": "
Method10-5 (3 tasks)5-5 (4 tasks)10-2 (6 tasks)15-1 (6 tasks)10-1 (10 tasks)
1-1011-201-201-56-201-201-1011-201-201-1516-201-201-1011-201-20
Joint Training75.976.376.172.477.376.175.976.376.178.070.476.175.976.376.1
Fine-tuning [41]5.330.618.00.518.313.83.813.68.70.010.55.30.05.12.6
ABR* [41]68.767.167.964.756.458.467.058.162.668.756.765.762.055.758.9
Faster ILOD [53]68.357.963.155.716.025.964.248.656.466.944.561.352.941.547.2
MMA [5]66.761.864.262.331.238.965.053.159.168.354.364.159.248.353.8
BPF (Ours)69.168.268.760.663.162.568.756.362.571.553.166.962.248.355.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.241, + 0.304, + 0.528, + 0.331 + ], + "angle": 0, + "content": "Table 3: mAP results on COCO2017. Methods with * use exemplars." + }, + { + "type": "table", + "bbox": [ + 0.245, + 0.333, + 0.528, + 0.469 + ], + "angle": 0, + "content": "
Method40-4070-10
APAP50AP75APAP50AP75
Joint Training36.757.839.836.757.839.8
Fine-tuning [41]19.031.220.45.68.66.2
ILOD-Meta* [26]23.840.524.4---
ABR* [41]34.557.835.231.152.932.7
Faster ILOD [49]20.640.1-21.339.9-
PseudoRM [64]25.344.4----
MMA [5]33.056.634.630.252.131.5
BPF (Ours)34.454.337.336.256.838.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.537, + 0.307, + 0.768, + 0.334 + ], + "angle": 0, + "content": "Table 4: Ablation study of various combinations of teacher models." + }, + { + "type": "table", + "bbox": [ + 0.556, + 0.337, + 0.746, + 0.402 + ], + "angle": 0, + "content": "
DistillationVOC(10-10)
Lroi dist,clsLroi dist,bbox1-1011-201-20
λ2= 1.0 part boxes71.573.372.4
λ2= 0.5 part boxes71.774.072.9
λ2= 0.5 all boxes71.374.472.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.576, + 0.414, + 0.73, + 0.426 + ], + "angle": 0, + "content": "Table 5: Effect of BF." + }, + { + "type": "table", + "bbox": [ + 0.55, + 0.428, + 0.753, + 0.467 + ], + "angle": 0, + "content": "
BFVOC(5-15)VOC(10-10)VOC(15-5)
1-56-201-201-1011-201-201-1516-20
×66.374.472.471.273.372.375.662.8
√66.475.373.071.774.072.975.963.0
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.785, + 0.567 + ], + "angle": 0, + "content": "PASCAL VOC 2007. For PASCAL VOC 2007, we order the classes alphabetically and evaluate our method with one or multiple training steps. We perform our experiments by adding 1 (19-1), 5 (15-5), 10 (10-10), or 15 (5-15) classes in a single incremental step. For multi-step incremental settings, we evaluate 10-5, 5-5, 10-2, 15-1, and 10-1 settings, where we add 5, 5, 2, 1, and 1 classes respectively at every step until all 20 classes are seen." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.787, + 0.764 + ], + "angle": 0, + "content": "- Single-step Incremental Settings: Table 1 shows our BPF methods against the existing methods using rehearsal or not. Rehearsal-based methods are not compared fairly with our BPF since we do not store old samples and use replay memory. As shown in Table 1, BPF consistently outperforms all previous methods, including those designed to combat forgetting using exemplars, validating the superiority of our approach. In particular, in the 19-1, 15-5, 10-10, and 5-15 settings, BPF significantly improved over MMA [5] by \\(3.4\\%\\), \\(2.8\\%\\), \\(6.3\\%\\), and \\(13.4\\%\\) on mAP@0.5 across all classes. Similarly, BPF outperforms the best rehearsal-based method ABR [41] by \\(3.2\\%\\), \\(1.7\\%\\), \\(0.9\\%\\), and \\(3.6\\%\\). The Avg metric equally averages old and new classes mAP, which straightly reports the incremental ability without the influence of the number of classes. BPF also outperforms most methods on the Avg metric, demonstrating BPF's adaptiveness in learning new classes and preserving the knowledge of old classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- Multi-step Incremental Settings: The issues of inconsistent optimization objectives across multiple stages and catastrophic forgetting are more crucial under the longer incremental settings. As shown in Table 2, BPF consistently outperforms MMA [5] across all the settings. Specifically, BPF improves over MMA by \\(1.4\\%\\), \\(2.8\\%\\), \\(3.4\\%\\), \\(4.5\\%\\), and even \\(23.6\\%\\) at 1-20 mAP@0.5 under" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Q. Mo et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.348, + 0.145, + 0.654, + 0.16 + ], + "angle": 0, + "content": "Table 6: Ablation study on each component." + }, + { + "type": "table", + "bbox": [ + 0.23, + 0.172, + 0.773, + 0.267 + ], + "angle": 0, + "content": "
ModelBridge the PastBridge the futureDistillation with FutureVOC(10-10)VOC(10-5)
1-1011-201-201-1011-1516-20
(a)58.172.465.354.469.959.3
(b)✓71.272.171.769.173.560.0
(c)✓61.073.367.154.870.758.7
(d)✓✓71.972.772.370.473.758.7
(e)✓✓✓71.774.072.969.175.261.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.292, + 0.784, + 0.385 + ], + "angle": 0, + "content": "the 10-1, 15-1, 10-2, 10-5, and 5-5 settings and enjoy improvement across all learning stages. Moreover, we find that even without storing memory, BPF still outperforms ABR [41] by \\(0.8\\%\\) and \\(4.1\\%\\) on overall mAP@0.5, \\(1.1\\%\\) and \\(6.7\\%\\) mAP@0.5 on new classes under 10-5 and 5-5 settings. In the settings of 10-2 and 10-1, limited by the small incremental data, BPF is inferior to ABR, but considering we do not require memory, these losses are acceptable." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.785, + 0.465 + ], + "angle": 0, + "content": "MS COCO 2017. On the COCO2017 dataset, we perform experiments on 40-40 and 70-10 settings, adding 40 and 10 classes, respectively, following [41]. As illustrated in Table 3, our method improves over MMA on average \\(AP\\) by \\(1.4\\%\\) on 40-40 settings and by \\(6.0\\%\\) on 70-10 settings. These results once again confirm the effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.485, + 0.5, + 0.5 + ], + "angle": 0, + "content": "4.3 Analysis and Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.787, + 0.78 + ], + "angle": 0, + "content": "We examine the contributions of the \"Bridge the Past\", \"Bridge the Future\", and \"Distillation with Future\" in Table 6 within the VOC 10-10 and 10-5 settings. We take the unbiased knowledge distillation proposed by [5] as the baseline model. By bridging the past, our model (b) aligns its optimization objectives with earlier ones, effectively reducing catastrophic forgetting of old classes, greatly enhancing the performance on detecting old classes. Compared to the baseline (a), it significantly improved by \\(13.1\\%\\) in the old classes and by \\(6.4\\%\\) for all classes on the 10-10 setting. Owing to bridging the future, our model (c) maintains consistent optimization objectives with future models regarding the background, making it easier to incrementally learn new classes. Compared to the baseline (a) on the 10-10 setting, it improves by \\(2.9\\%\\) and \\(0.9\\%\\) on the old classes and the new classes, respectively. Combining Past and Future (d), the model outperforms consistently on each stage. In the Distillation with Future strategy, we leverage the intermediate model to aid the old model in modeling new classes, conducting distillation across all categories for the current model. Table 6 (e) shows that a comprehensive joint teacher model facilitates improved learning of all categories through knowledge distillation, with a significant improvement \\((+1.3\\%)\\) AP) in new classes. Similar results can be found in the 10-5 setting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.786, + 0.84 + ], + "angle": 0, + "content": "We further verify the \"Bridge the future\" (BF) in Table 5. There is a clear trend that as the number of considered future classes increases, the BF shows increasing improvement in future classes (\\(+0.9\\%\\) in 5-15) without degrading the old classes' performance." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.572, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bridge Past and Future" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.174, + 0.237, + 0.192 + ], + "angle": 0, + "content": "BPF" + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.144, + 0.371, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.144, + 0.503, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.144, + 0.636, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.144, + 0.785, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.249, + 0.237, + 0.273 + ], + "angle": 0, + "content": "MMA" + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.223, + 0.372, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.223, + 0.503, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.223, + 0.636, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.223, + 0.785, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.31, + 0.788, + 0.368 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results for the model trained under the \\(10 + 10\\) setting on the VOC 2007 test set. 'boat', 'cat', 'chair', and 'car' are old classes from the first stage, and 'person' and 'dog' are the classes from the second stage. Compared with MMA (bottom row), our BPF (top row) can produce reliable predictions on both old and new classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.785, + 0.488 + ], + "angle": 0, + "content": "We also conducted a quantitative analysis of Bridge Past and Future. Under the second stage of the VOC 10-5-5 (3 tasks), the original background boxes have a \\(92.6\\%\\) and \\(83.9\\%\\) Recall50 rate for old and future classes, while the Recall50 of pseudo labels for old classes and discarded boxes for future classes are \\(66.1\\%\\) and \\(17.6\\%\\) respectively, demonstrating the necessity and effectiveness of Bridge the Past and Bridge the Future." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.787, + 0.611 + ], + "angle": 0, + "content": "Table 4 presents experiments on the effect of different combinations of teacher models. \\(\\lambda_{2}\\) in Equation (7) determines using which teacher model as the primary. Using the intermediate expert model as the primary on regions overlapping with gt \\((\\lambda_{2} = 0.5)\\) outperforms using the old model \\((\\lambda_{2} = 1.0)\\), showing that distilling new class objects with the old model hinders their learning. For box distillation, we find that distilling boxes on primary classes (part boxes, i.e., only boxes for old classes in \\(\\mathcal{R}_1\\) and boxes for new classes in \\(\\mathcal{R}_2\\) participate in distillation while others are ignored) performs similarly with all classes (all boxes)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.638, + 0.371, + 0.652 + ], + "angle": 0, + "content": "4.4 Visualization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.785, + 0.743 + ], + "angle": 0, + "content": "Visualization of Detection Results. We visualize the detection results in Figure 5 to illustrate the significant improvement compared to the previous methods qualitatively. Our method accurately detects both new and old class objects simultaneously. While MMA fails to detect old classes accurately, suffering from catastrophic forgetting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Visualization of Bridge Past and Future. We visualized the module of Bridge the Past and Bridge the Future separately in Figure 6. We effectively model the missing annotations for old class objects to bridge the past. As demonstrated in Figure 6(b), the attention maps clearly differentiate the foreground from the background, regardless of the presence of annotations. The discarded boxes (the third row) validate the effectiveness of our method." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Q. Mo et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.497, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.144, + 0.785, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.366, + 0.788, + 0.437 + ], + "angle": 0, + "content": "Fig. 6: Visualization of Bridge Past and Future. Boxes in red represent the ground truth in the current stage. (a) In Bridge the Past, we effectively constructed pseudo labels of past classes. (b) In Bridge the Future, salient objects (marked in green boxes) can be easily detected from the attention maps and are excluded from the background regions. Best viewed in color." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.464, + 0.369, + 0.481 + ], + "angle": 0, + "content": "5 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.497, + 0.788, + 0.604 + ], + "angle": 0, + "content": "Limitations. In our Bridge the Past procedure, we assume several objects of old classes may appear in the current training data. However, when the number of incremental classes is limited, e.g., increasing a single class in each stage, objects of old classes may rarely occur due to limited training images. This is expected to be alleviated by generating samples of old classes via the copy-paste strategy as in ABR [41], while it may introduce little stored samples. Detailed discussion can be found in the Supplement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.788, + 0.791 + ], + "angle": 0, + "content": "Conclusions. In this work, we find that the concurrence of classes from different learning stages causes a severe information asymmetry, not only causing catastrophic forgetting for old classes but also hindering the learning of new classes. To tackle the problem, we propose the Bridge Past and Future method, which uses pseudo labels from the old model to fill in the missing annotations and exclude some potential future objects from the background, keeping the learning consistent across all stages. Further, Distillation with Future loss is proposed to solve the problem of the old teacher model's lack of knowledge of new classes. By combining the knowledge from the past to the future, our method consistently outperforms others across different learning stages in most incremental settings. To the best of our knowledge, we are the first to consider future classes during the learning, shedding light on a new aspect of Incremental Object Detection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported partially by the National Key Research and Development Program of China (2023YFA1008503), NSFC(U21A20471), Guangdong NSF Project (No. 2023B1515040025, 2020B1515120085)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.572, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bridge Past and Future" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.784, + 0.204 + ], + "angle": 0, + "content": "1. Acharya, M., Hayes, T.L., Kanan, C.: Rodeo: Replay for online object detection. arXiv preprint arXiv:2008.06439 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.232 + ], + "angle": 0, + "content": "2. Cao, S., Joshi, D., Gui, L.Y., Wang, Y.X.: Contrastive mean teacher for domain adaptive object detectors. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.233, + 0.785, + 0.259 + ], + "angle": 0, + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.785, + 0.287 + ], + "angle": 0, + "content": "4. Castro, F.M., Marin-Jiménez, M.J., Guil, N., Schmid, C., Alahari, K.: End-to-end incremental learning. In: ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.315 + ], + "angle": 0, + "content": "5. Cermelli, F., Geraci, A., Fontanel, D., Caputo, B.: Modeling missing annotations for incremental learning in object detection. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.316, + 0.785, + 0.343 + ], + "angle": 0, + "content": "6. Chaudhry, A., Dokania, P.K., Ajanthan, T., Torr, P.H.: Riemannian walk for incremental learning: Understanding forgetting and intransigence. In: ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.344, + 0.785, + 0.37 + ], + "angle": 0, + "content": "7. Chen, L., Yu, C., Chen, L.: A new knowledge distillation for incremental object detection. In: IJCNN (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.398 + ], + "angle": 0, + "content": "8. Dhar, P., Singh, R.V., Peng, K.C., Wu, Z., Chellappa, R.: Learning without memorizing. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.785, + 0.425 + ], + "angle": 0, + "content": "9. Dong, N., Zhang, Y., Ding, M., Bai, Y.: Class-incremental object detection. PR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.785, + 0.453 + ], + "angle": 0, + "content": "0. Douillard, A., Cord, M., Ollion, C., Robert, T., Valle, E.: Podnet: Pooled outputs distillation for small-tasks incremental learning. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.455, + 0.785, + 0.48 + ], + "angle": 0, + "content": "1. Douillard, A., Rame, A., Couairon, G., Cord, M.: Dytox: Transformers for continual learning with dynamic token expansion. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.481, + 0.785, + 0.509 + ], + "angle": 0, + "content": "2. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. Springer IJCV (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.51, + 0.785, + 0.536 + ], + "angle": 0, + "content": "3. Feng, T., Wang, M., Yuan, H.: Overcoming catastrophic forgetting in incremental object detection via elastic response distillation. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.785, + 0.564 + ], + "angle": 0, + "content": "4. Fu, S., Yan, J., Gao, Y., Xie, X., Zheng, W.S.: Asag: Building strong one-decoder-layer sparse detectors via adaptive sparse anchor generation. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.592 + ], + "angle": 0, + "content": "5. Gao, Y., Lin, K.Y., Yan, J., Wang, Y., Zheng, W.S.: Asyfod: An asymmetric adaptation paradigm for few-shot domain adaptive object detection. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.593, + 0.785, + 0.619 + ], + "angle": 0, + "content": "6. Gao, Y., Yang, L., Huang, Y., Xie, S., Li, S., Zheng, W.S.: Acrofod: An adaptive method for cross-domain few-shot object detection. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.785, + 0.646 + ], + "angle": 0, + "content": "7. Gao, Z., Wang, L., Han, B., Guo, S.: Adamixer: A fast-converging query-based object detector. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.52, + 0.66 + ], + "angle": 0, + "content": "8. Girshick, R.: Fast r-cnn. In: ICCV (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.688 + ], + "angle": 0, + "content": "9. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "20. Gupta, A., Narayan, S., Joseph, K., Khan, S., Khan, F.S., Shah, M.: Ow-detr: Open-world detection transformer. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.744 + ], + "angle": 0, + "content": "21. Hao, Y., Fu, Y., Jiang, Y.G., Tian, Q.: An end-to-end architecture for class incremental object detection with knowledge distillation. In: ICME (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.743, + 0.757 + ], + "angle": 0, + "content": "22. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask r-cnn. In: ICCV (2017)." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "23. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "24. Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incrementally via rebalancing. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "25. Joseph, K., Khan, S., Khan, F.S., Balasubramanian, V.N.: Towards open world object detection. In: CVPR (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Q. Mo et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "26. Joseph, K., Rajasegaran, J., Khan, S., Khan, F.S., Balasubramanian, V.N.: Incremental object detection via meta-learning. IEEE TPAMI (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "27. Kang, B., Liu, Z., Wang, X., Yu, F., Feng, J., Darrell, T.: Few-shot object detection via feature reweighting. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.785, + 0.245 + ], + "angle": 0, + "content": "28. Kang, M., Zhang, J., Zhang, J., Wang, X., Chen, Y., Ma, Z., Huang, X.: Alleviating catastrophic forgetting of incremental object detection via within-class and between-class knowledge distillation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.247, + 0.785, + 0.274 + ], + "angle": 0, + "content": "29. Kemker, R., Kanan, C.: Fearnet: Brain-inspired model for incremental learning. arXiv preprint arXiv:1711.10563 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.787, + 0.33 + ], + "angle": 0, + "content": "30. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.785, + 0.373 + ], + "angle": 0, + "content": "31. Li, D., Tasci, S., Ghosh, S., Zhu, J., Zhang, J., Heck, L.: Rilod: Near real-time incremental learning for object detection at the edge. In: Proceedings of the 4th ACM/IEEE Symposium on Edge Computing (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.374, + 0.785, + 0.401 + ], + "angle": 0, + "content": "32. Li, M., Cong, Y., Liu, Y., Sun, G.: Class-incremental gesture recognition learning with out-of-distribution detection. In: IROS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.402, + 0.785, + 0.429 + ], + "angle": 0, + "content": "33. Li, Y.M., Zeng, L.A., Meng, J.K., Zheng, W.S.: Continual action assessment via task-consistent score-discriminative feature distribution modeling. TCVT (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.43, + 0.699, + 0.444 + ], + "angle": 0, + "content": "34. Li, Z., Hoiem, D.: Learning without forgetting. IEEE TPAMI (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.785, + 0.471 + ], + "angle": 0, + "content": "35. Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.5 + ], + "angle": 0, + "content": "36. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.529 + ], + "angle": 0, + "content": "37. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.785, + 0.557 + ], + "angle": 0, + "content": "38. Liu, L., Kuang, Z., Chen, Y., Xue, J.H., Yang, W., Zhang, W.: Incdet: In defense of elastic weight consolidation for incremental object detection. IEEE TNNLS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.558, + 0.785, + 0.586 + ], + "angle": 0, + "content": "39. Liu, X., Yang, H., Ravichandran, A., Bhotika, R., Soatto, S.: Multi-task incremental learning for object detection. arXiv preprint arXiv:2002.05347 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.785, + 0.613 + ], + "angle": 0, + "content": "40. Liu, Y., Schiele, B., Vedaldi, A., Rupprecht, C.: Continual detection transformer for incremental object detection. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.785, + 0.642 + ], + "angle": 0, + "content": "41. Liu, Y., Cong, Y., Goswami, D., Liu, X., van de Weijer, J.: Augmented box replay: Overcoming foreground shift for incremental object detection. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.785, + 0.671 + ], + "angle": 0, + "content": "42. Liu, Y., Cong, Y., Sun, G., Zhang, T., Dong, J., Liu, H.: L3doc: Lifelong 3d object classification. IEEE TIP (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.785, + 0.699 + ], + "angle": 0, + "content": "43. Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. In: NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.785, + 0.741 + ], + "angle": 0, + "content": "44. Ma, J., Niu, Y., Xu, J., Huang, S., Han, G., Chang, S.F.: Digeo: Discriminative geometry-aware learning for generalized few-shot object detection. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.785, + 0.769 + ], + "angle": 0, + "content": "45. Mallya, A., Davis, D., Lazebnik, S.: Piggyback: Adapting a single network to multiple tasks by learning to mask weights. In: ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.77, + 0.785, + 0.798 + ], + "angle": 0, + "content": "46. Mallya, A., Lazebnik, S.: Packnet: Adding multiple tasks to a single network by iterative pruning. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "47. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., et al.: Simple open-vocabulary object detection. In: ECCV (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.572, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bridge Past and Future" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "48. Ostapenko, O., Puscas, M., Klein, T., Jahnichen, P., Nabi, M.: Learning to remember: A synaptic plasticity driven framework for continual learning. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "49. Peng, C., Zhao, K., Lovell, B.C.: Faster ilod: Incremental learning for object detectors based on faster rcnn. PR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.218, + 0.787, + 0.245 + ], + "angle": 0, + "content": "50. Peng, C., Zhao, K., Maksoud, S., Wang, T., Lovell, B.C.: Diode: dilatable incremental object detection. PR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.245, + 0.787, + 0.273 + ], + "angle": 0, + "content": "51. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "52. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.787, + 0.328 + ], + "angle": 0, + "content": "53. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. In: NeurIPS (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "54. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.787, + 0.412 + ], + "angle": 0, + "content": "55. Shieh, J.L., Haq, Q.M.u., Haq, M.A., Karam, S., Chondro, P., Gao, D.Q., Ruan, S.J.: Continual learning strategy in one-stage object detection framework based on experience replay for autonomous driving vehicle. Sensors (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.412, + 0.787, + 0.439 + ], + "angle": 0, + "content": "56. Shin, H., Lee, J.K., Kim, J., Kim, J.: Continual learning with deep generative replay. In: NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.787, + 0.467 + ], + "angle": 0, + "content": "57. Shmelkov, K., Schmid, C., Alahari, K.: Incremental learning of object detectors without catastrophic forgetting. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.467, + 0.787, + 0.494 + ], + "angle": 0, + "content": "58. Tan, M., Pang, R., Le, Q.V.: Efficientdet: Scalable and efficient object detection. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.494, + 0.787, + 0.522 + ], + "angle": 0, + "content": "59. Tang, Y.M., Peng, Y.X., Zheng, W.S.: Learning to imagine: Diversify memory for incremental learning using unlabeled data. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.787, + 0.55 + ], + "angle": 0, + "content": "60. Tang, Y.M., Peng, Y.X., Zheng, W.S.: When prompt-based incremental learning does not meet strong pretraining. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.787, + 0.577 + ], + "angle": 0, + "content": "61. Tian, Z., Shen, C., Chen, H., He, T.: Fcos: Fully convolutional one-stage object detection. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.578, + 0.787, + 0.618 + ], + "angle": 0, + "content": "62. Wu, C., Herranz, L., Liu, X., Van De Weijer, J., Raducanu, B., et al.: Memory replay gans: Learning to generate new categories without forgetting. In: NeurIPS (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.619, + 0.787, + 0.646 + ], + "angle": 0, + "content": "63. Yan, J., Yang, L., Gao, Y., Zheng, W.S.: Self-supervised cross-stage regional contrastive learning for object detection. In: ICME (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.646, + 0.787, + 0.688 + ], + "angle": 0, + "content": "64. Yang, D., Zhou, Y., Hong, X., Zhang, A., Wei, X., Zeng, L., Qiao, Z., Wang, W.: Pseudo object replay and mining for incremental object detection. In: ACM MM (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.688, + 0.787, + 0.716 + ], + "angle": 0, + "content": "65. Yang, D., Zhou, Y., Zhang, A., Sun, X., Wu, D., Wang, W., Ye, Q.: Multi-view correlation distillation for incremental object detection. PR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.716, + 0.787, + 0.743 + ], + "angle": 0, + "content": "66. Zenke, F., Poole, B., Ganguli, S.: Continual learning through synaptic intelligence. In: ICML (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.743, + 0.787, + 0.772 + ], + "angle": 0, + "content": "67. Zhou, W., Chang, S., Sosa, N., Hamann, H., Cox, D.: Lifelong object detection. arXiv preprint arXiv:2009.01129 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "68. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.787, + 0.827 + ], + "angle": 0, + "content": "69. Zohar, O., Wang, K.C., Yeung, S.: Prob: Probabilistic objectness for open world object detection. In: CVPR (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.827 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Bridge Past and Future_ Overcoming Information Asymmetry in Incremental Object Detection/1c449442-a7f3-4df8-9bc6-973b26235638_origin.pdf b/2024/Bridge Past and Future_ Overcoming Information Asymmetry in Incremental Object Detection/1c449442-a7f3-4df8-9bc6-973b26235638_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6b1bd99f3ea8245ad2f5fbdb352ae4f15864d2f4 --- /dev/null +++ b/2024/Bridge Past and Future_ Overcoming Information Asymmetry in Incremental Object Detection/1c449442-a7f3-4df8-9bc6-973b26235638_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:767dc83fd9e360cea405460e7f75eba9cda73c48c0a6c3b268e22378de6ec5a0 +size 4258843 diff --git a/2024/Bridge Past and Future_ Overcoming Information Asymmetry in Incremental Object Detection/full.md b/2024/Bridge Past and Future_ Overcoming Information Asymmetry in Incremental Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..75ceff4087d336b72f5f42eb1b55df76a5c079e8 --- /dev/null +++ b/2024/Bridge Past and Future_ Overcoming Information Asymmetry in Incremental Object Detection/full.md @@ -0,0 +1,334 @@ +# Bridge Past and Future: Overcoming Information Asymmetry in Incremental Object Detection + +Qijie Mo $^{1,3}$ , Yipeng Gao $^{1,3}$ , Shenghao Fu $^{1,3}$ , Junkai Yan $^{1,3}$ , Ancong Wu $^{1,3,\star}$ , and Wei-Shi Zheng $^{1,2,3,\star}$ + +$^{1}$ School of Computer Science and Engineering, Sun Yat-sen University, China + $^{2}$ Peng Cheng Laboratory, Shenzhen, China + +3 Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China + +{moqj3,gaoyp23,fushh7,yanjk3}@mail2.sysu.edu.cn,wuanc@mail.sysu.edu.cn, wszheng@ieee.org + +Abstract. In incremental object detection, knowledge distillation has been proven to be an effective way to alleviate catastrophic forgetting. However, previous works focused on preserving the knowledge of old models, ignoring that images could simultaneously contain categories from past, present, and future stages. The co-occurrence of objects makes the optimization objectives inconsistent across different stages since the definition for foreground objects differs across various stages, which limits the model's performance greatly. To overcome this problem, we propose a method called "Bridge Past and Future" (BPF), which aligns models across stages, ensuring consistent optimization directions. In addition, we propose a novel Distillation with Future (DwF) loss, fully leveraging the background probability to mitigate the forgetting of old classes while ensuring a high level of adaptability in learning new classes. Extensive experiments are conducted on both Pascal VOC and MS COCO benchmarks. Without memory, BPF outperforms current state-of-the-art methods under various settings. The code is available at https://github.com/iSEE-Laboratory/BPF. + +Keywords: Object Detection $\cdot$ Incremental Learning $\cdot$ Knowledge Distillation + +# 1 Introduction + +Object detection [3,14,36,53,61,68] is a fundamental computer vision task and has significantly given rise to many sub-directions [2, 15, 16, 19, 27, 44, 47, 63]. General detection models typically assume full access to all interest classes during training and require exhaustive labeled data from the start. However, in dynamic real-world applications, where new object categories may appear continually, object detection models are expected to adapt to these ongoing changing classes, especially when there is limited storage to save all the data from the beginning + +![](images/aa666c9761dbf89963229ec6d7447721c272172efbfccb6dbd2a6b4c86a16591.jpg) +(a) Distribution of instances in the second stage of the 10-5-5 setting on VOC 2007 dataset + +![](images/f3a2f32a8be1939a244179e29f7af2bda6521c50c3ed594b7bde39ff7f6eaadc.jpg) +Fig. 1: Challenges faced by the IOD task. Classes from previous, current, and potential future stages may appear arbitrarily in the current training stage, while only the annotations of current classes are available to train the detector. Best viewed in color. + +![](images/168029cc099ccf2ea6d2fa6f540f61497bab82dfb5381267fd314dbf1e49bc09.jpg) +(b) Visualization of some typical training samples in IOD + +to the present or when data privacy concerns are encountered. To this end, incremental object detection (IOD) [5,26,41,65,67,69] has caught progressive attention, which aims to continually detect new objects without forgetting the previously learned ones. + +Unlike incremental learning in classification, the incremental object detection task faces a dual challenge caused by the concurrence of different classes: (1) Foreground objects identified in previous stages are unlabeled in the current stage, and thus, the model currently trained may regard them as background. (2) The current stage's background may also contain objects that will be recognized as the foreground in future stages. Taking the 10-5-5 setting on the VOC dataset [12] as an example, as shown in Figure 1, a large proportion of instances belonging to classes of the previous and future stages occur concurrently in the training images in the current stage. This asymmetry in information between the past and current stages causes the current model to erroneously classify objects of old categories as background, aggravating the catastrophic forgetting problem. Similarly, the asymmetry between the current and future stages leads the current model to classify objects of potential future categories as background erroneously, which requires models trained in future stages to rectify their misperception, thus hindering the learning of new classes. As a result, the optimization objectives across different learning stages are inconsistent, significantly limiting the generalization of current incremental detectors. Previous works [5, 65, 67] concentrate on introducing strong regularization to prevent catastrophic forgetting but ignore the impact of concurrence of different classes, resulting in limited performance on both old and new classes. + +This work aims to tackle the critical information asymmetry challenge in IOD by utilizing the abundant concurrence information within an image. We introduce a novel approach named Bridge Past and Future (BPF), which is designed to connect the past and future stages with the current stage, ensuring the model adheres to consistent optimization objectives throughout the entire incremental learning process. For the past stages, we utilize the previous model as a reliable labeler to construct some pseudo labels for old classes and combine them with current annotations for training the current model symmetrically. + +Simultaneously, we re-identify some salient regions that may contain potential future objects out of the background and exclude them from negative samples in the current stage to avoid classifying them into the background at this stage and disturbing future learning, which is expected to achieve symmetry with future stages. These two novel designs alleviate the impact of inconsistent training objectives across different stages from the aspects of the past and the future, thus easing the training difficulty of incremental detectors. + +In addition, we propose a distillation technique to augment the detector in the current stage, where we take a step back by viewing the current stage as a clear future stage for previous stages and propose a Distillation with Future (DwF) loss. The distillation is carried out by two teachers, i.e., the detector trained on the previous stage and an expert detector trained with only current data. The student detector absorbs the knowledge of old and current classes from the old and expert models, respectively, in a class-by-class manner, preventing the detector from catastrophic forgetting and facilitating learning of current classes. + +We conduct extensive experiments on the PASCAL VOC [12] and COCO [37] datasets to evaluate the effectiveness of our BPF, which outperforms other state-of-the-art (SOTA) methods under multiple settings in a memory-free way. Moreover, we also conduct comprehensive ablation studies and visualizations to help better understand how each component of the BPF works. + +# 2 Related Works + +Object Detection. Traditional object detectors can be primarily classified into two streams: one-stage [36,52,58,61] and two-stage detectors [18,22,35,53]. Two-stage detectors first predict several coarse candidate proposals via region proposal extractors [18,53] and then adopt a region of interest (RoI) head to refine these proposals and output final predictions. Unlike two-stage detectors, one-stage counterparts directly generate final outputs without predicting candidate proposals. Despite performing well under the standard training setup, both of them fail to generalize to the incremental training setup due to lacking previous training data. Without losing generality, this paper concentrates on enhancing two-stage detectors, i.e., Faster R-CNN [53], enabling it to learn new classes incrementally while retaining previously acquired knowledge. + +Incremental Learning. Incremental learning methods can be mainly divided into three categories: memory-based [4, 24, 32, 33, 48, 51, 56, 59, 62], regularization-based [6, 8, 30, 34, 66], and structure-based [42, 45, 46, 54, 60]. Memory-based methods store a handful of samples for replaying [24, 40, 41] or suggest generating samples [29, 62] for data compensation in new stages. Some regularization-based methods try to eliminate the effect of new tasks on previously learned knowledge by identifying the most crucial parameters [30, 43, 62], while some other regularization-based methods propose to distill knowledge [23] from old to current models by transiting knowledge on logits, final features or intermediate features [8, 10, 11, 24, 34]. Structure-based methods dedicate specific parameters to each task, freezing them to mitigate forgetting. In this study, we focus on the + +regularization-based knowledge distillation approach for object detection and introduce a novel distillation with future information. + +Incremental Object Detection. Most previous works [9,28,31,40,50,55] focus on distilling old classes' knowledge of intermediate features [5,21,39,41,49,65,67], region proposal network [5, 49, 67], or RoI head [13] to prevent forgetting. A stream of methods also pays attention to classification loss. PPAS [67] introduces a pseudo-positive-aware sampling algorithm to identify regions corresponding to old classes and prevents them from being sampled as background. MMA [5] proposes unbiased classification loss, consolidating the background and all old classes into one entity, which aims to minimize the optimization objective conflict between the current stage's background and the past stages' foreground. However, it diminishes the ability to distinguish old classes from the background. Other methods [1,20,25,26,41,64] focus on rehearsal to maintain the previous stage knowledge, either performing replay of the intermediate features [1], the images [25,26,40], or the instances [41]. Incdet [38] presents a parameter isolation strategy that builds upon EWC [30]. In this study, in addition to preserving knowledge for old classes, we also take future classes into account, aligning optimization objectives consistently across all stages. + +# 3 Method + +# 3.1 Preliminaries + +Problem Formulation. In incremental object detection, the training is performed over multiple learning stages, each one introducing a new set of classes to be detected. Let $\mathcal{C}$ denote the set of classes that are incrementally introduced to the object detector $\mathcal{M}$ . In the $t$ -th training stage, a grouping of classes $\mathcal{C}_t$ are introduced to the detector: $\mathcal{C}_t \subset \mathcal{C}$ , such that $\mathcal{C}_i \cap \mathcal{C}_j = \emptyset$ , for any $i \neq j$ and $i, j \leq t$ . Let $\mathcal{D}_t$ denote the images containing annotated objects of classes in $\mathcal{C}_t$ . Each image can contain multiple objects of different classes, but annotations $\mathcal{V}_t$ are available only for those object instances that belong to classes in $\mathcal{C}_t$ . The challenge in class incremental object detection is to continually update $\mathcal{M}_t$ to $\mathcal{M}_{t+1}$ by learning $\mathcal{C}_{t+1}$ , without access to $\{\mathcal{D}_0, \dots, \mathcal{D}_t\}$ while maintaining original performance on $\{\mathcal{C}_0, \dots, \mathcal{C}_t\}$ . + +General Detector Training Pipeline. This work starts with the representative Faster RCNN-like detectors [53] $f = \{f_b, f_{rpn}, f_{roi}\}$ , which generally includes a backbone network $f_b$ , Region Proposal Network (RPN) $f_{rpn}$ and Region of Interest Head (RoI Head) $f_{roi}$ . During training, all training proposals are categorized into positive, negative, and ignored samples for supervised training based on their IoUs with the ground truth. Positive samples are assigned to predict the class of their matched ground truth, whereas negative samples are designated as background. Thus, the objects without annotations in the current stage will be classified as background, hindering the learning of incremental models. + +Common Distillation Methods for Incremental Object Detection. To prevent forgetting in IOD, a widely adopted solution involves knowledge distil- + +lation in RPN and RoI head [5,49,57,67]: + +$$ +\mathcal {L} _ {\text {d i s t}} = \mathcal {L} _ {\text {d i s t}} ^ {r p n} + \mathcal {L} _ {\text {d i s t}} ^ {r o i}. \tag {1} +$$ + +The distillation in the RoI head includes the L2 loss between the box coordinates and the Kullback-Leibler divergence for class probabilities. MMA [5] notices the missing old annotations and proposes Unbiased Knowledge Distillation (UKD): + +$$ +\mathcal {L} _ {\text {d i s t}, \text {c l s}} ^ {\text {r o i}} (i) = \frac {1}{| \mathcal {C} _ {t - 1} | + 1} \left(p _ {i} ^ {b, t - 1} \log \left(p _ {i} ^ {b, t} + \sum_ {c \in \mathcal {C} _ {t}} p _ {i} ^ {c, t}\right) + \sum_ {c \in \mathcal {C} _ {1: t - 1}} p _ {i} ^ {c, t - 1} \log \left(p _ {i} ^ {c, t}\right)\right), \tag {2} +$$ + +where $p_i^{c,t-1}$ and $p_i^{c,t}$ indicate the classification output for the proposal $i$ and class $c$ of the teacher (old) model and student (current) model, respectively, and $b$ is the background class. As a common practice, 64 proposals are used for distillation, which are randomly selected out of the top 128 proposals with the highest objectness scores from the RPN network of the old model. However, given the class-agnostic characteristic of the RPN, the proposals for distillation cover objects of both old and current classes. In such cases, the old model, without seeing current classes, cannot provide useful knowledge for the current classes and even hinders their learning. Further, treating the background probability and current class probabilities as a unified entity during distillation inevitably diminishes the ability to differentiate between current classes and background. + +# 3.2 Overall Framework + +Unlike classification tasks with a single label per input, incremental object detection presents a scenario where an image $\mathcal{I}_t$ can encompass objects from the current class set $\mathcal{C}_t$ , previous class sets $\mathcal{C}_{1:t-1}$ and future class sets $\mathcal{C}_{t+1:\infty}$ . Nonetheless, the annotations $\mathcal{Y}_t$ are limited to bounding boxes and class labels for objects in $\mathcal{C}_t$ , while objects of other classes are regarded as background. + +To address the challenge of inconsistent optimization objectives, we introduce a novel strategy termed Bridge Past and Future (BPF). As shown in Figure 2, the method is divided into two parts. From the perspective of supervised learning, we use high-confidence predictions of old classes given by the old model as pseudo labels to bridge the information from the past stage (Section 3.3) and discard some potential objects from the background to bridge the future stage (Section 3.4). Regarding the distillation learning, we propose a novel Distillation with Future (DwF) loss by considering current classes in the distillation. The distillation probabilities are combined from two teacher models: the old model $\mathcal{M}_{t-1}$ for old classes and an expert model for current classes. By utilizing abundant information in background probability, the distillation takes both previous stages and the current stage into consideration (Section 3.5). + +# 3.3 Bridging the Past + +To address the inconsistency in optimization objectives between the previous and current models, we bridge the past information to the current stage by incorporating pseudo supervision signals from past stages into the current model's + +![](images/7649c9c1029ff8c148ec97a2b94c23bdbd993fd84fdbb5931ef097af45ffde81.jpg) +Fig. 2: The overall framework of our method. The top side illustrates the Bridge Past and Future (BPF) procedure, which identifies objects of past classes and excludes several potential objects of future classes to ensure consistent optimization during the entire training process. The bottom side shows the Distillation with Future (DwF) process, which employs both the old model $\mathcal{M}_{t-1}$ adept at detecting old categories and the interim model $\mathcal{M}_t^{im}$ trained on $\mathcal{D}_t$ and specialized in new categories, to conduct a comprehensive distillation across all categories for the current model $\mathcal{M}_t$ . + +optimization process. The old model was trained under the supervision of human annotations, which is a high-quality pseudo labeler for old classes. We take the high-confidence inference results of the old model $\mathcal{M}_{t-1}$ on current training images as pseudo labels and combine them with current annotations to train the current detector, thereby aligning the supervision of old classes with previous stages. The left part of Figure 3 shows the pipeline of bridging the past strategy. + +Formally, given predictions $\hat{y}^{old} \in \mathcal{P}$ of the old model $\mathcal{M}_{t-1}$ on current images, we first select some high-confidence regions of previous classes $\mathcal{C}_{1:t-1}$ with a confidence threshold $\eta$ , followed by the Non-Maximum Suppression (NMS) operation to reduce duplication: + +$$ +\mathcal {U} = \{j \in \mathcal {P}: \max _ {c \in \mathcal {C} _ {1: t - 1}} \hat {p} _ {j} ^ {\text {o l d}} (c) > \eta \}, \quad \mathcal {E} = \mathrm {N M S} (\mathcal {U}). \tag {3} +$$ + +Then, we further narrow down the predictions to a subset $\mathcal{W} \subset \mathcal{E}$ that do not overlap with the ground-truth labels of the new categories via an IoU threshold $\lambda_{1}$ , ensuring a clear distinction between past and present classes: + +$$ +\mathcal {W} = \{j \in \mathcal {E}: \forall i \in \mathcal {Y} _ {t}, \operatorname {I o U} \left(\hat {\boldsymbol {b}} _ {j} ^ {\text {o l d}}, \boldsymbol {b} _ {i}\right) \leq \lambda_ {1} \}. \tag {4} +$$ + +By modeling the previous supervision signals from the old model, we bridge the past stages to the current one, ensuring the current model's optimization direction encompasses the objectives of earlier stages, significantly mitigating + +![](images/70c935fc92d0f32026cb7ebe198ff5f50ad55a53e2c2a73dc6d00fec1b1e5b56.jpg) +Fig. 3: Overview of Bridge Past and Future. We adopt the previous model $\mathcal{M}_{t-1}$ to predict some pseudo labels for past classes to complement their missing supervision in the current stage. Additionally, we exclude several proposals that are likely to be an object but are not included in the current ground truth and pseudo labels from the background to avoid classifying them into background mistakenly. + +the forgetting problem. MMA [5] treats the background probability and old class probabilities as a unified entity in the classification loss, making old classes hard to separate from the background. While our method explicitly models the old classes in the current stage out of the background. + +# 3.4 Bridging the Future + +Bridging the future aims to rectify the misalignment between the optimization goals of the current and upcoming stages, as objects from future class sets $\mathcal{C}_{t + 1:\infty}$ in the current dataset $\mathcal{D}_t$ are classified as background. The core design is to find some salient objects in the background and exclude them from the negative samples during the training. By separating pure background from regions likely to contain future category objects in the current stage, we align the optimization objectives for background treatment across both current and future models. + +Specifically, we find that the activation in feature maps is a good indicator for salient objects. As illustrated in Figure 3, in the absence of annotations for old and future classes, foreground and background features demonstrate notable differences in spatial attention. We produce the attention map $A_{i} \in \mathbb{R}^{H \times W}$ from the backbone feature map $F_{i} = f_{b}(I_{i}) \in \mathbb{R}^{H \times W \times C}$ : + +$$ +A _ {i} = \operatorname {S o f t m a x} \left(\sum_ {c = 1} ^ {C} \left| F _ {i} \right| ^ {p}\right), \tag {5} +$$ + +where $H$ , $W$ , and $C$ denote the feature's height, width, and channel. We compute the attention score for each region, indicating the likelihood of containing foreground objects. The attention score $a_{i,j}^{roi}$ for region $r_j$ is calculated as follows: + +$$ +a _ {i, j} ^ {\text {r o i}} = \operatorname {A v g} \left(\operatorname {R o I P o o l} \left(A _ {i}, r _ {j}\right)\right) \in \mathbb {R}. \tag {6} +$$ + +![](images/1d5787d9ef4e0988e0987e9d0f93fb93ceefe6246d848e1ab7214142d5058d50.jpg) +Fig. 4: The illustration of Distillation with Future strategy. An intermediate teacher model trained on the current dataset is used to compensate for the lack of current class information in the old model. For proposals overlapping ground truth of the current stage, since the intermediate teacher is specialized in detecting current classes, we directly inherit its probabilities on current classes and use the old model to enrich its background probability with old class knowledge. On the contrary, for proposals that do not overlap with GT, the old model is preferred, and the intermediate model is used as compensation. Combining two teachers makes the distillation class by class. + +Regions with high attention scores $a_{i,j}^{roi}$ from feature maps and objectness scores $o_j$ from class-agnostic RPN imply a greater chance of being future category objects. We discard these RoIs when sampling negative samples when training the RoI head, thus maintaining the model's consistency with future stage background definitions. Regions with lower scores or having a considerable IoU with ground truth are considered reliable backgrounds. + +# 3.5 Combining Current Classes into Distillation + +Distillation is an effective way to prevent forgetting in incremental object detection. However, distilling from the teacher model, which is biased toward old classes, inevitably hinders the learning for current classes. To address this challenge, we introduce the Distillation with Future (DwF) loss to distill from the teacher model in a more fine-grained and adaptive way. Different from MMA [5], which aligns the current classes and the background as a whole with the old model's background, we distill each class one by one, preserving the distinction between current classes and the background. However, the model $\mathcal{M}_{t-1}$ from the previous stage has not been trained on current classes, thus the distillation on current classes can not be performed. We introduce another intermediate teacher model $\mathcal{M}_t^{im}$ , which is trained using the dataset from the current stage $\mathcal{D}_t$ in a fully supervised way, as a supplement. Taking the intermediate model $\mathcal{M}_t^{im}$ as a + +complementary teacher, we explicitly consider the current stage $t$ as the future stage of stage $t - 1$ , making the distillation future-aware. + +As the old model $\mathcal{M}_{t-1}$ performs well on $\mathcal{C}_{1:t-1}$ while the intermediate model $\mathcal{M}_t^{im}$ is the expert in $\mathcal{C}_t$ , we distill different regions with different combination of teacher models, as shown in Figure 4. Specifically, we divide the regions for distillation $\mathcal{R}$ into two subsets $\mathcal{R}_1, \mathcal{R}_2 \subset \mathcal{R}$ that based on their intersection over union with the ground truth labels for the new categories $\mathcal{C}_t$ : + +$$ +\mathcal {R} _ {1} = \{j \in \mathcal {R}: \forall i \in \mathcal {Y} _ {t}, \operatorname {I o U} (b _ {j}, b _ {i}) \leq \lambda_ {2.} \}, +$$ + +$$ +\begin{array}{l} \mathcal {R} _ {1} = \left\{j \in \mathcal {K}: \forall i \in \mathcal {Y} _ {t}, \operatorname {I o U} \left(b _ {j}, b _ {i}\right) \geq \lambda_ {2}. \right\}, \\ \mathcal {R} _ {2} = \left\{j \in \mathcal {R}: \forall i \in \mathcal {Y} _ {t}, \operatorname {I o U} \left(b _ {j}, b _ {i}\right) > \lambda_ {2}. \right\}. \end{array} \tag {7} +$$ + +For regions $r_i \in \mathcal{R}_1$ , which are likely to be the regions for old classes, we take $\mathcal{M}_{t-1}$ as the primary model for distillation and reconstruct its background representation with the model $\mathcal{M}_t^{im}$ : + +$$ +\hat {p} _ {i} ^ {c, i m} = p _ {i} ^ {c, i m} \times p _ {i} ^ {b, t - 1}, \quad r _ {i} \in \mathcal {R} _ {1}, \tag {8} +$$ + +where $p_i^{b,t-1}$ and $p_i^{c,im}$ are the classification probabilities of background in the model $\mathcal{M}_{t-1}$ and the classification probabilities for current classes and background $c \in \mathcal{C}_t \cup \mathcal{B}$ in the model $\mathcal{M}_t^{im}$ for the region $r_i$ respectively. After weighting, $\sum_{c=1}^{\mathcal{C}_t \cup \mathcal{B}} \hat{p}_i^{c,im} = p_i^{b,t-1}$ . The final distillation probabilities for regions in $\mathcal{R}_1$ are $[p_i^{C_{1:t-1},t-1}, \hat{p}_i^{C_t \cup \mathcal{B},im}] \in \mathbb{R}^{|C_{1:t}|+1}$ . + +On the contrary, for regions $r_i \in \mathcal{R}_2$ , they are regions for current classes; thus, the intermediate model $\mathcal{M}_t^{im}$ is taken as the primary model for distillation, and its background representation is reconstructed by the model $\mathcal{M}_{t-1}$ : + +$$ +\hat {p} _ {i} ^ {c, t - 1} = p _ {i} ^ {c, t - 1} \times p _ {i} ^ {b, i m}, \quad r _ {i} \in \mathcal {R} _ {2}, \tag {9} +$$ + +where $\sum_{c=1}^{\mathcal{C}_{1:t-1} \cup \mathcal{B}} \hat{p}_i^{c,t-1} = p_i^{b,im}$ . The distillation probabilities for regions in $\mathcal{R}_2$ are $[\hat{p}_i^{\mathcal{C}_{1:t-1},t-1}, p_i^{\mathcal{C}_t,im}, \hat{p}_i^{\mathcal{B},t-1}] \in \mathbb{R}^{|C_{1:t}|+1}$ . + +With the expanded probabilities from the complimentary teacher models, the distillation on the classification head can be performed using a conventional Kullback-Leibler divergence. Regarding the box distillation, we use the output boxes from the old model $\mathcal{M}_{t-1}$ for regions in $\mathcal{R}_1$ and the intermediate model $\mathcal{M}_t^{im}$ for $\mathcal{R}_2$ . As a result, the complementary knowledge from two teachers, the expansion of background probability, and the combination of adaptive probability for different regions not only prevent the student from catastrophic forgetting but also facilitate the learning for current classes. + +# 4 Experiments + +# 4.1 Experiment Settings + +Datasets and Evaluation Metrics. Following previous works [5,20,25,26,41, 49,57,65,67], we evaluate our method on PASCAL VOC 2007 [12] and MS COCO + +Table 1: mAP@0.5 results on single incremental step on PASCAL VOC 2007. The best performance in each is presented with **bold**, and the second best is presented with **underline**. Methods with * use exemplars. + +
Method19-115-510-105-15
1-19201-20Avg1-1516-201-20Avg1-1011-201-20Avg1-55-151-20Avg
Joint Training76.076.776.176.478.070.476.174.275.976.376.176.172.477.376.174.9
Fine-tuning12.062.814.537.414.259.225.436.79.562.536.036.06.963.149.135.0
ORE* [25]69.460.168.964.771.858.768.565.260.468.864.664.6----
OW-DETR* [20]70.262.069.866.172.259.869.166.063.567.965.765.7----
ILOD-Meta* [26]70.957.670.264.271.755.967.863.868.464.366.366.3----
ABR* [41]71.069.770.970.473.065.171.069.171.272.872.072.064.771.069.467.9
Faster ILOD [53]68.961.168.565.071.656.967.964.369.854.562.162.162.037.143.349.6
PPAS [67]70.553.069.261.8----63.560.061.861.8----
MVC [65]70.260.669.765.469.457.966.563.766.266.066.166.1----
PROB [69]73.948.572.661.573.560.870.167.066.067.266.566.5----
PseudoRM [64]72.967.372.670.173.460.970.366.969.168.668.968.9----
MMA [5]71.163.470.767.273.060.569.966.769.363.966.666.666.857.259.662.0
BPF (Ours)74.565.374.169.975.963.072.769.571.774.072.972.966.475.373.070.9
+ +2017 [37] datasets. PASCAL VOC 2007 dataset comprises 9,963 images across 20 categories. The COCO 2017 dataset encompasses objects from 80 categories, with around $118\mathrm{k}$ images for training and 5,000 images for validation. The mean average precision at the 0.5 IoU threshold $(\mathrm{mAP}@0.5)$ is used as the primary evaluation metric for the VOC dataset, and the mean average precision ranging from 0.5 to 0.95 is the main evaluation metric for the COCO dataset. + +For each incremental setting (A-B), the first number A denotes the number of classes in the first stage and the second number is the number of classes newly introduced in each new stage. Note that the columns with gray background in the table of experimental results represent the average AP among all classes. + +Implementation Details. Similar to [5,25,26,41,49], we build our incremental object detector based on Faster R-CNN [53] with R50. Our method can easily be adapted to transformer-based detectors [3,17,68]. We conduct the experiments under a strict rehearsal-free setting, where no memory is used. We set $\eta = 0.75$ , $\lambda_{1} = 0.7$ , $\lambda_{2} = 0.5$ . For $\mathcal{W}$ in Equation (4), we use an IOU threshold of 0.3 to divide it into two sets. For the set with IoU $< 0.3$ , the supervision signal weights are set to 1.0, while the other set is assigned a weight of 0.3. + +# 4.2 Quantitative Evaluation + +Following previous work [5,7,41,49,57,65,67], our method is evaluated on settings with a range of initial classes and incorporating one or more incremental tasks. We benchmark our method against two baselines: Fine-Tuning, where the model is incrementally trained on new data without any regularization strategy or data replay, and Joint Training, which involves training the model on the complete dataset using all annotations. + +Table 2: mAP@0.5 results on multiple incremental steps on PASCAL VOC 2007. The best performance in each is presented with **bold**, and the second best is presented with **underline**. Methods with * use exemplars. + +
Method10-5 (3 tasks)5-5 (4 tasks)10-2 (6 tasks)15-1 (6 tasks)10-1 (10 tasks)
1-1011-201-201-56-201-201-1011-201-201-1516-201-201-1011-201-20
Joint Training75.976.376.172.477.376.175.976.376.178.070.476.175.976.376.1
Fine-tuning [41]5.330.618.00.518.313.83.813.68.70.010.55.30.05.12.6
ABR* [41]68.767.167.964.756.458.467.058.162.668.756.765.762.055.758.9
Faster ILOD [53]68.357.963.155.716.025.964.248.656.466.944.561.352.941.547.2
MMA [5]66.761.864.262.331.238.965.053.159.168.354.364.159.248.353.8
BPF (Ours)69.168.268.760.663.162.568.756.362.571.553.166.962.248.355.2
+ +Table 3: mAP results on COCO2017. Methods with * use exemplars. + +
Method40-4070-10
APAP50AP75APAP50AP75
Joint Training36.757.839.836.757.839.8
Fine-tuning [41]19.031.220.45.68.66.2
ILOD-Meta* [26]23.840.524.4---
ABR* [41]34.557.835.231.152.932.7
Faster ILOD [49]20.640.1-21.339.9-
PseudoRM [64]25.344.4----
MMA [5]33.056.634.630.252.131.5
BPF (Ours)34.454.337.336.256.838.9
+ +Table 4: Ablation study of various combinations of teacher models. + +
DistillationVOC(10-10)
Lroi dist,clsLroi dist,bbox1-1011-201-20
λ2= 1.0 part boxes71.573.372.4
λ2= 0.5 part boxes71.774.072.9
λ2= 0.5 all boxes71.374.472.9
+ +Table 5: Effect of BF. + +
BFVOC(5-15)VOC(10-10)VOC(15-5)
1-56-201-201-1011-201-201-1516-20
×66.374.472.471.273.372.375.662.8
√66.475.373.071.774.072.975.963.0
+ +PASCAL VOC 2007. For PASCAL VOC 2007, we order the classes alphabetically and evaluate our method with one or multiple training steps. We perform our experiments by adding 1 (19-1), 5 (15-5), 10 (10-10), or 15 (5-15) classes in a single incremental step. For multi-step incremental settings, we evaluate 10-5, 5-5, 10-2, 15-1, and 10-1 settings, where we add 5, 5, 2, 1, and 1 classes respectively at every step until all 20 classes are seen. + +- Single-step Incremental Settings: Table 1 shows our BPF methods against the existing methods using rehearsal or not. Rehearsal-based methods are not compared fairly with our BPF since we do not store old samples and use replay memory. As shown in Table 1, BPF consistently outperforms all previous methods, including those designed to combat forgetting using exemplars, validating the superiority of our approach. In particular, in the 19-1, 15-5, 10-10, and 5-15 settings, BPF significantly improved over MMA [5] by $3.4\%$ , $2.8\%$ , $6.3\%$ , and $13.4\%$ on mAP@0.5 across all classes. Similarly, BPF outperforms the best rehearsal-based method ABR [41] by $3.2\%$ , $1.7\%$ , $0.9\%$ , and $3.6\%$ . The Avg metric equally averages old and new classes mAP, which straightly reports the incremental ability without the influence of the number of classes. BPF also outperforms most methods on the Avg metric, demonstrating BPF's adaptiveness in learning new classes and preserving the knowledge of old classes. + +- Multi-step Incremental Settings: The issues of inconsistent optimization objectives across multiple stages and catastrophic forgetting are more crucial under the longer incremental settings. As shown in Table 2, BPF consistently outperforms MMA [5] across all the settings. Specifically, BPF improves over MMA by $1.4\%$ , $2.8\%$ , $3.4\%$ , $4.5\%$ , and even $23.6\%$ at 1-20 mAP@0.5 under + +Table 6: Ablation study on each component. + +
ModelBridge the PastBridge the futureDistillation with FutureVOC(10-10)VOC(10-5)
1-1011-201-201-1011-1516-20
(a)58.172.465.354.469.959.3
(b)✓71.272.171.769.173.560.0
(c)✓61.073.367.154.870.758.7
(d)✓✓71.972.772.370.473.758.7
(e)✓✓✓71.774.072.969.175.261.2
+ +the 10-1, 15-1, 10-2, 10-5, and 5-5 settings and enjoy improvement across all learning stages. Moreover, we find that even without storing memory, BPF still outperforms ABR [41] by $0.8\%$ and $4.1\%$ on overall mAP@0.5, $1.1\%$ and $6.7\%$ mAP@0.5 on new classes under 10-5 and 5-5 settings. In the settings of 10-2 and 10-1, limited by the small incremental data, BPF is inferior to ABR, but considering we do not require memory, these losses are acceptable. + +MS COCO 2017. On the COCO2017 dataset, we perform experiments on 40-40 and 70-10 settings, adding 40 and 10 classes, respectively, following [41]. As illustrated in Table 3, our method improves over MMA on average $AP$ by $1.4\%$ on 40-40 settings and by $6.0\%$ on 70-10 settings. These results once again confirm the effectiveness of our method. + +# 4.3 Analysis and Ablation Study + +We examine the contributions of the "Bridge the Past", "Bridge the Future", and "Distillation with Future" in Table 6 within the VOC 10-10 and 10-5 settings. We take the unbiased knowledge distillation proposed by [5] as the baseline model. By bridging the past, our model (b) aligns its optimization objectives with earlier ones, effectively reducing catastrophic forgetting of old classes, greatly enhancing the performance on detecting old classes. Compared to the baseline (a), it significantly improved by $13.1\%$ in the old classes and by $6.4\%$ for all classes on the 10-10 setting. Owing to bridging the future, our model (c) maintains consistent optimization objectives with future models regarding the background, making it easier to incrementally learn new classes. Compared to the baseline (a) on the 10-10 setting, it improves by $2.9\%$ and $0.9\%$ on the old classes and the new classes, respectively. Combining Past and Future (d), the model outperforms consistently on each stage. In the Distillation with Future strategy, we leverage the intermediate model to aid the old model in modeling new classes, conducting distillation across all categories for the current model. Table 6 (e) shows that a comprehensive joint teacher model facilitates improved learning of all categories through knowledge distillation, with a significant improvement $(+1.3\%)$ AP) in new classes. Similar results can be found in the 10-5 setting. + +We further verify the "Bridge the future" (BF) in Table 5. There is a clear trend that as the number of considered future classes increases, the BF shows increasing improvement in future classes ( $+0.9\%$ in 5-15) without degrading the old classes' performance. + +![](images/6c89ff87d065a6f9ab29ccdd0fa7c6cc10189fddcc952b34b2014ea161568258.jpg) +BPF + +![](images/1900f9ab1924fb3c62d64e4c1e50eb0952d142c0a041630e48aad19db45b424c.jpg) + +![](images/5fba540b8bad63031c31db4f4ed03720484fa76007fe7ece8774f19623a4d289.jpg) + +![](images/2ae0cd7cf71505a1319cf2b46972aca1a345b6ac0dccad71432251cc02a6b3ee.jpg) + +![](images/a446ce0cddd61694c5dca1fbd3446b6fee22ed55502d5a6dafa54842652e7db0.jpg) +MMA + +![](images/44113d95ec2825f062d1059c5fe1e4c46b0c7443dc8da63717d267f6f0e27573.jpg) +Fig. 5: Qualitative results for the model trained under the $10 + 10$ setting on the VOC 2007 test set. 'boat', 'cat', 'chair', and 'car' are old classes from the first stage, and 'person' and 'dog' are the classes from the second stage. Compared with MMA (bottom row), our BPF (top row) can produce reliable predictions on both old and new classes. + +![](images/21b6e2735b4c0f063c09312d9073f5d84cd6cb3e372007ddf6e71039537e4d49.jpg) + +![](images/c577aad62b4b6b7a6ea9da1c0f502309023b1b355eb0af3dc4ad9d8453645b37.jpg) + +We also conducted a quantitative analysis of Bridge Past and Future. Under the second stage of the VOC 10-5-5 (3 tasks), the original background boxes have a $92.6\%$ and $83.9\%$ Recall50 rate for old and future classes, while the Recall50 of pseudo labels for old classes and discarded boxes for future classes are $66.1\%$ and $17.6\%$ respectively, demonstrating the necessity and effectiveness of Bridge the Past and Bridge the Future. + +Table 4 presents experiments on the effect of different combinations of teacher models. $\lambda_{2}$ in Equation (7) determines using which teacher model as the primary. Using the intermediate expert model as the primary on regions overlapping with gt $(\lambda_{2} = 0.5)$ outperforms using the old model $(\lambda_{2} = 1.0)$ , showing that distilling new class objects with the old model hinders their learning. For box distillation, we find that distilling boxes on primary classes (part boxes, i.e., only boxes for old classes in $\mathcal{R}_1$ and boxes for new classes in $\mathcal{R}_2$ participate in distillation while others are ignored) performs similarly with all classes (all boxes). + +# 4.4 Visualization + +Visualization of Detection Results. We visualize the detection results in Figure 5 to illustrate the significant improvement compared to the previous methods qualitatively. Our method accurately detects both new and old class objects simultaneously. While MMA fails to detect old classes accurately, suffering from catastrophic forgetting. + +Visualization of Bridge Past and Future. We visualized the module of Bridge the Past and Bridge the Future separately in Figure 6. We effectively model the missing annotations for old class objects to bridge the past. As demonstrated in Figure 6(b), the attention maps clearly differentiate the foreground from the background, regardless of the presence of annotations. The discarded boxes (the third row) validate the effectiveness of our method. + +![](images/d2dc35795df87af9ec32adb2e58f457e3b6338d686352c39360985d616d1abd6.jpg) +Fig. 6: Visualization of Bridge Past and Future. Boxes in red represent the ground truth in the current stage. (a) In Bridge the Past, we effectively constructed pseudo labels of past classes. (b) In Bridge the Future, salient objects (marked in green boxes) can be easily detected from the attention maps and are excluded from the background regions. Best viewed in color. + +![](images/b041acdd977e3de8e6adaf39f524c1d89cbc9da25a537bfbc8212a357c10c8c8.jpg) + +# 5 Conclusions + +Limitations. In our Bridge the Past procedure, we assume several objects of old classes may appear in the current training data. However, when the number of incremental classes is limited, e.g., increasing a single class in each stage, objects of old classes may rarely occur due to limited training images. This is expected to be alleviated by generating samples of old classes via the copy-paste strategy as in ABR [41], while it may introduce little stored samples. Detailed discussion can be found in the Supplement. + +Conclusions. In this work, we find that the concurrence of classes from different learning stages causes a severe information asymmetry, not only causing catastrophic forgetting for old classes but also hindering the learning of new classes. To tackle the problem, we propose the Bridge Past and Future method, which uses pseudo labels from the old model to fill in the missing annotations and exclude some potential future objects from the background, keeping the learning consistent across all stages. Further, Distillation with Future loss is proposed to solve the problem of the old teacher model's lack of knowledge of new classes. By combining the knowledge from the past to the future, our method consistently outperforms others across different learning stages in most incremental settings. To the best of our knowledge, we are the first to consider future classes during the learning, shedding light on a new aspect of Incremental Object Detection. + +Acknowledgments. This work was supported partially by the National Key Research and Development Program of China (2023YFA1008503), NSFC(U21A20471), Guangdong NSF Project (No. 2023B1515040025, 2020B1515120085). + +# References + +1. Acharya, M., Hayes, T.L., Kanan, C.: Rodeo: Replay for online object detection. arXiv preprint arXiv:2008.06439 (2020) +2. Cao, S., Joshi, D., Gui, L.Y., Wang, Y.X.: Contrastive mean teacher for domain adaptive object detectors. In: CVPR (2023) +3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020) +4. Castro, F.M., Marin-Jiménez, M.J., Guil, N., Schmid, C., Alahari, K.: End-to-end incremental learning. In: ECCV (2018) +5. Cermelli, F., Geraci, A., Fontanel, D., Caputo, B.: Modeling missing annotations for incremental learning in object detection. In: CVPR (2022) +6. Chaudhry, A., Dokania, P.K., Ajanthan, T., Torr, P.H.: Riemannian walk for incremental learning: Understanding forgetting and intransigence. In: ECCV (2018) +7. Chen, L., Yu, C., Chen, L.: A new knowledge distillation for incremental object detection. In: IJCNN (2019) +8. Dhar, P., Singh, R.V., Peng, K.C., Wu, Z., Chellappa, R.: Learning without memorizing. In: CVPR (2019) +9. Dong, N., Zhang, Y., Ding, M., Bai, Y.: Class-incremental object detection. PR (2023) +0. Douillard, A., Cord, M., Ollion, C., Robert, T., Valle, E.: Podnet: Pooled outputs distillation for small-tasks incremental learning. In: ECCV (2020) +1. Douillard, A., Rame, A., Couairon, G., Cord, M.: Dytox: Transformers for continual learning with dynamic token expansion. In: CVPR (2022) +2. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. Springer IJCV (2010) +3. Feng, T., Wang, M., Yuan, H.: Overcoming catastrophic forgetting in incremental object detection via elastic response distillation. In: CVPR (2022) +4. Fu, S., Yan, J., Gao, Y., Xie, X., Zheng, W.S.: Asag: Building strong one-decoder-layer sparse detectors via adaptive sparse anchor generation. In: ICCV (2023) +5. Gao, Y., Lin, K.Y., Yan, J., Wang, Y., Zheng, W.S.: Asyfod: An asymmetric adaptation paradigm for few-shot domain adaptive object detection. In: CVPR (2023) +6. Gao, Y., Yang, L., Huang, Y., Xie, S., Li, S., Zheng, W.S.: Acrofod: An adaptive method for cross-domain few-shot object detection. In: ECCV (2022) +7. Gao, Z., Wang, L., Han, B., Guo, S.: Adamixer: A fast-converging query-based object detector. In: CVPR (2022) +8. Girshick, R.: Fast r-cnn. In: ICCV (2015) +9. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921 (2021) +20. Gupta, A., Narayan, S., Joseph, K., Khan, S., Khan, F.S., Shah, M.: Ow-detr: Open-world detection transformer. In: CVPR (2022) +21. Hao, Y., Fu, Y., Jiang, Y.G., Tian, Q.: An end-to-end architecture for class incremental object detection with knowledge distillation. In: ICME (2019) +22. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask r-cnn. In: ICCV (2017). +23. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) +24. Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incrementally via rebalancing. In: CVPR (2019) +25. Joseph, K., Khan, S., Khan, F.S., Balasubramanian, V.N.: Towards open world object detection. In: CVPR (2021) + +26. Joseph, K., Rajasegaran, J., Khan, S., Khan, F.S., Balasubramanian, V.N.: Incremental object detection via meta-learning. IEEE TPAMI (2021) +27. Kang, B., Liu, Z., Wang, X., Yu, F., Feng, J., Darrell, T.: Few-shot object detection via feature reweighting. In: ICCV (2019) +28. Kang, M., Zhang, J., Zhang, J., Wang, X., Chen, Y., Ma, Z., Huang, X.: Alleviating catastrophic forgetting of incremental object detection via within-class and between-class knowledge distillation. In: CVPR (2023) +29. Kemker, R., Kanan, C.: Fearnet: Brain-inspired model for incremental learning. arXiv preprint arXiv:1711.10563 (2017) +30. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences (2017) +31. Li, D., Tasci, S., Ghosh, S., Zhu, J., Zhang, J., Heck, L.: Rilod: Near real-time incremental learning for object detection at the edge. In: Proceedings of the 4th ACM/IEEE Symposium on Edge Computing (2019) +32. Li, M., Cong, Y., Liu, Y., Sun, G.: Class-incremental gesture recognition learning with out-of-distribution detection. In: IROS (2022) +33. Li, Y.M., Zeng, L.A., Meng, J.K., Zheng, W.S.: Continual action assessment via task-consistent score-discriminative feature distribution modeling. TCVT (2024) +34. Li, Z., Hoiem, D.: Learning without forgetting. IEEE TPAMI (2017) +35. Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: CVPR (2017) +36. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: ICCV (2017) +37. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014) +38. Liu, L., Kuang, Z., Chen, Y., Xue, J.H., Yang, W., Zhang, W.: Incdet: In defense of elastic weight consolidation for incremental object detection. IEEE TNNLS (2020) +39. Liu, X., Yang, H., Ravichandran, A., Bhotika, R., Soatto, S.: Multi-task incremental learning for object detection. arXiv preprint arXiv:2002.05347 (2020) +40. Liu, Y., Schiele, B., Vedaldi, A., Rupprecht, C.: Continual detection transformer for incremental object detection. In: CVPR (2023) +41. Liu, Y., Cong, Y., Goswami, D., Liu, X., van de Weijer, J.: Augmented box replay: Overcoming foreground shift for incremental object detection. In: ICCV (2023) +42. Liu, Y., Cong, Y., Sun, G., Zhang, T., Dong, J., Liu, H.: L3doc: Lifelong 3d object classification. IEEE TIP (2021) +43. Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. In: NeurIPS (2017) +44. Ma, J., Niu, Y., Xu, J., Huang, S., Han, G., Chang, S.F.: Digeo: Discriminative geometry-aware learning for generalized few-shot object detection. In: CVPR (2023) +45. Mallya, A., Davis, D., Lazebnik, S.: Piggyback: Adapting a single network to multiple tasks by learning to mask weights. In: ECCV (2018) +46. Mallya, A., Lazebnik, S.: Packnet: Adding multiple tasks to a single network by iterative pruning. In: CVPR (2018) +47. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., et al.: Simple open-vocabulary object detection. In: ECCV (2022) + +48. Ostapenko, O., Puscas, M., Klein, T., Jahnichen, P., Nabi, M.: Learning to remember: A synaptic plasticity driven framework for continual learning. In: CVPR (2019) +49. Peng, C., Zhao, K., Lovell, B.C.: Faster ilod: Incremental learning for object detectors based on faster rcnn. PR (2020) +50. Peng, C., Zhao, K., Maksoud, S., Wang, T., Lovell, B.C.: Diode: dilatable incremental object detection. PR (2023) +51. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: CVPR (2017) +52. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018) +53. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. In: NeurIPS (2015) +54. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016) +55. Shieh, J.L., Haq, Q.M.u., Haq, M.A., Karam, S., Chondro, P., Gao, D.Q., Ruan, S.J.: Continual learning strategy in one-stage object detection framework based on experience replay for autonomous driving vehicle. Sensors (2020) +56. Shin, H., Lee, J.K., Kim, J., Kim, J.: Continual learning with deep generative replay. In: NeurIPS (2017) +57. Shmelkov, K., Schmid, C., Alahari, K.: Incremental learning of object detectors without catastrophic forgetting. In: ICCV (2017) +58. Tan, M., Pang, R., Le, Q.V.: Efficientdet: Scalable and efficient object detection. In: CVPR (2020) +59. Tang, Y.M., Peng, Y.X., Zheng, W.S.: Learning to imagine: Diversify memory for incremental learning using unlabeled data. In: CVPR (2022) +60. Tang, Y.M., Peng, Y.X., Zheng, W.S.: When prompt-based incremental learning does not meet strong pretraining. In: ICCV (2023) +61. Tian, Z., Shen, C., Chen, H., He, T.: Fcos: Fully convolutional one-stage object detection. In: ICCV (2019) +62. Wu, C., Herranz, L., Liu, X., Van De Weijer, J., Raducanu, B., et al.: Memory replay gans: Learning to generate new categories without forgetting. In: NeurIPS (2018) +63. Yan, J., Yang, L., Gao, Y., Zheng, W.S.: Self-supervised cross-stage regional contrastive learning for object detection. In: ICME (2023) +64. Yang, D., Zhou, Y., Hong, X., Zhang, A., Wei, X., Zeng, L., Qiao, Z., Wang, W.: Pseudo object replay and mining for incremental object detection. In: ACM MM (2023) +65. Yang, D., Zhou, Y., Zhang, A., Sun, X., Wu, D., Wang, W., Ye, Q.: Multi-view correlation distillation for incremental object detection. PR (2022) +66. Zenke, F., Poole, B., Ganguli, S.: Continual learning through synaptic intelligence. In: ICML (2017) +67. Zhou, W., Chang, S., Sosa, N., Hamann, H., Cox, D.: Lifelong object detection. arXiv preprint arXiv:2009.01129 (2020) +68. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: ICLR (2021) +69. Zohar, O., Wang, K.C., Yeung, S.: Prob: Probabilistic objectness for open world object detection. In: CVPR (2023) \ No newline at end of file diff --git a/2024/Bridge Past and Future_ Overcoming Information Asymmetry in Incremental Object Detection/images.zip b/2024/Bridge Past and Future_ Overcoming Information Asymmetry in Incremental Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..eb35a245f3c4122d533fbbc9f0469c96b628ebf5 --- /dev/null +++ b/2024/Bridge Past and Future_ Overcoming Information Asymmetry in Incremental Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a82fc3c0c618c03f875260dcb2771fb4c7b17434fdcd81d116e78cee4413d07b +size 610557 diff --git a/2024/Bridge Past and Future_ Overcoming Information Asymmetry in Incremental Object Detection/layout.json b/2024/Bridge Past and Future_ Overcoming Information Asymmetry in Incremental Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b3ea96cf6f33af16808cb4dbd72bdf8fbfe86497 --- /dev/null +++ b/2024/Bridge Past and Future_ Overcoming Information Asymmetry in Incremental Object Detection/layout.json @@ -0,0 +1,10899 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "text", + "content": "Bridge Past and Future: Overcoming Information Asymmetry in Incremental Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 158, + 167, + 455, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 167, + 455, + 194 + ], + "spans": [ + { + "bbox": [ + 158, + 167, + 455, + 194 + ], + "type": "text", + "content": "Qijie Mo" + }, + { + "bbox": [ + 158, + 167, + 455, + 194 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 158, + 167, + 455, + 194 + ], + "type": "text", + "content": ", Yipeng Gao" + }, + { + "bbox": [ + 158, + 167, + 455, + 194 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 158, + 167, + 455, + 194 + ], + "type": "text", + "content": ", Shenghao Fu" + }, + { + "bbox": [ + 158, + 167, + 455, + 194 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 158, + 167, + 455, + 194 + ], + "type": "text", + "content": ", Junkai Yan" + }, + { + "bbox": [ + 158, + 167, + 455, + 194 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 158, + 167, + 455, + 194 + ], + "type": "text", + "content": ", Ancong Wu" + }, + { + "bbox": [ + 158, + 167, + 455, + 194 + ], + "type": "inline_equation", + "content": "^{1,3,\\star}" + }, + { + "bbox": [ + 158, + 167, + 455, + 194 + ], + "type": "text", + "content": ", and Wei-Shi Zheng" + }, + { + "bbox": [ + 158, + 167, + 455, + 194 + ], + "type": "inline_equation", + "content": "^{1,2,3,\\star}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 146, + 201, + 466, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 201, + 466, + 223 + ], + "spans": [ + { + "bbox": [ + 146, + 201, + 466, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 201, + 466, + 223 + ], + "type": "text", + "content": " School of Computer Science and Engineering, Sun Yat-sen University, China \n" + }, + { + "bbox": [ + 146, + 201, + 466, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 201, + 466, + 223 + ], + "type": "text", + "content": " Peng Cheng Laboratory, Shenzhen, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 165, + 224, + 447, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 224, + 447, + 245 + ], + "spans": [ + { + "bbox": [ + 165, + 224, + 447, + 245 + ], + "type": "text", + "content": "3 Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 194, + 247, + 419, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 247, + 419, + 269 + ], + "spans": [ + { + "bbox": [ + 194, + 247, + 419, + 269 + ], + "type": "text", + "content": "{moqj3,gaoyp23,fushh7,yanjk3}@mail2.sysu.edu.cn,wuanc@mail.sysu.edu.cn, wszheng@ieee.org" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 298, + 455, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 298, + 455, + 485 + ], + "spans": [ + { + "bbox": [ + 160, + 298, + 455, + 485 + ], + "type": "text", + "content": "Abstract. In incremental object detection, knowledge distillation has been proven to be an effective way to alleviate catastrophic forgetting. However, previous works focused on preserving the knowledge of old models, ignoring that images could simultaneously contain categories from past, present, and future stages. The co-occurrence of objects makes the optimization objectives inconsistent across different stages since the definition for foreground objects differs across various stages, which limits the model's performance greatly. To overcome this problem, we propose a method called \"Bridge Past and Future\" (BPF), which aligns models across stages, ensuring consistent optimization directions. In addition, we propose a novel Distillation with Future (DwF) loss, fully leveraging the background probability to mitigate the forgetting of old classes while ensuring a high level of adaptability in learning new classes. Extensive experiments are conducted on both Pascal VOC and MS COCO benchmarks. Without memory, BPF outperforms current state-of-the-art methods under various settings. The code is available at https://github.com/iSEE-Laboratory/BPF." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 495, + 453, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 495, + 453, + 516 + ], + "spans": [ + { + "bbox": [ + 159, + 495, + 453, + 516 + ], + "type": "text", + "content": "Keywords: Object Detection " + }, + { + "bbox": [ + 159, + 495, + 453, + 516 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 495, + 453, + 516 + ], + "type": "text", + "content": " Incremental Learning " + }, + { + "bbox": [ + 159, + 495, + 453, + 516 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 495, + 453, + 516 + ], + "type": "text", + "content": " Knowledge Distillation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 537, + 230, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 537, + 230, + 551 + ], + "spans": [ + { + "bbox": [ + 132, + 537, + 230, + 551 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "content": "Object detection [3,14,36,53,61,68] is a fundamental computer vision task and has significantly given rise to many sub-directions [2, 15, 16, 19, 27, 44, 47, 63]. General detection models typically assume full access to all interest classes during training and require exhaustive labeled data from the start. However, in dynamic real-world applications, where new object categories may appear continually, object detection models are expected to adapt to these ongoing changing classes, especially when there is limited storage to save all the data from the beginning" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "text", + "content": "* Corresponding Author" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 269, + 195 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 269, + 195 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 269, + 195 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 269, + 195 + ], + "type": "image", + "image_path": "aa666c9761dbf89963229ec6d7447721c272172efbfccb6dbd2a6b4c86a16591.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 197, + 268, + 213 + ], + "lines": [ + { + "bbox": [ + 138, + 197, + 268, + 213 + ], + "spans": [ + { + "bbox": [ + 138, + 197, + 268, + 213 + ], + "type": "text", + "content": "(a) Distribution of instances in the second stage of the 10-5-5 setting on VOC 2007 dataset" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 272, + 116, + 375, + 194 + ], + "blocks": [ + { + "bbox": [ + 272, + 116, + 375, + 194 + ], + "lines": [ + { + "bbox": [ + 272, + 116, + 375, + 194 + ], + "spans": [ + { + "bbox": [ + 272, + 116, + 375, + 194 + ], + "type": "image", + "image_path": "f3a2f32a8be1939a244179e29f7af2bda6521c50c3ed594b7bde39ff7f6eaadc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 222, + 482, + 257 + ], + "lines": [ + { + "bbox": [ + 130, + 222, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 222, + 482, + 257 + ], + "type": "text", + "content": "Fig. 1: Challenges faced by the IOD task. Classes from previous, current, and potential future stages may appear arbitrarily in the current training stage, while only the annotations of current classes are available to train the detector. Best viewed in color." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 376, + 117, + 480, + 194 + ], + "blocks": [ + { + "bbox": [ + 296, + 197, + 452, + 207 + ], + "lines": [ + { + "bbox": [ + 296, + 197, + 452, + 207 + ], + "spans": [ + { + "bbox": [ + 296, + 197, + 452, + 207 + ], + "type": "text", + "content": "(b) Visualization of some typical training samples in IOD" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 376, + 117, + 480, + 194 + ], + "lines": [ + { + "bbox": [ + 376, + 117, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 376, + 117, + 480, + 194 + ], + "type": "image", + "image_path": "168029cc099ccf2ea6d2fa6f540f61497bab82dfb5381267fd314dbf1e49bc09.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 281, + 479, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 479, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 479, + 327 + ], + "type": "text", + "content": "to the present or when data privacy concerns are encountered. To this end, incremental object detection (IOD) [5,26,41,65,67,69] has caught progressive attention, which aims to continually detect new objects without forgetting the previously learned ones." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 329, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 569 + ], + "type": "text", + "content": "Unlike incremental learning in classification, the incremental object detection task faces a dual challenge caused by the concurrence of different classes: (1) Foreground objects identified in previous stages are unlabeled in the current stage, and thus, the model currently trained may regard them as background. (2) The current stage's background may also contain objects that will be recognized as the foreground in future stages. Taking the 10-5-5 setting on the VOC dataset [12] as an example, as shown in Figure 1, a large proportion of instances belonging to classes of the previous and future stages occur concurrently in the training images in the current stage. This asymmetry in information between the past and current stages causes the current model to erroneously classify objects of old categories as background, aggravating the catastrophic forgetting problem. Similarly, the asymmetry between the current and future stages leads the current model to classify objects of potential future categories as background erroneously, which requires models trained in future stages to rectify their misperception, thus hindering the learning of new classes. As a result, the optimization objectives across different learning stages are inconsistent, significantly limiting the generalization of current incremental detectors. Previous works [5, 65, 67] concentrate on introducing strong regularization to prevent catastrophic forgetting but ignore the impact of concurrence of different classes, resulting in limited performance on both old and new classes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "This work aims to tackle the critical information asymmetry challenge in IOD by utilizing the abundant concurrence information within an image. We introduce a novel approach named Bridge Past and Future (BPF), which is designed to connect the past and future stages with the current stage, ensuring the model adheres to consistent optimization objectives throughout the entire incremental learning process. For the past stages, we utilize the previous model as a reliable labeler to construct some pseudo labels for old classes and combine them with current annotations for training the current model symmetrically." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Mo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "Simultaneously, we re-identify some salient regions that may contain potential future objects out of the background and exclude them from negative samples in the current stage to avoid classifying them into the background at this stage and disturbing future learning, which is expected to achieve symmetry with future stages. These two novel designs alleviate the impact of inconsistent training objectives across different stages from the aspects of the past and the future, thus easing the training difficulty of incremental detectors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 295 + ], + "type": "text", + "content": "In addition, we propose a distillation technique to augment the detector in the current stage, where we take a step back by viewing the current stage as a clear future stage for previous stages and propose a Distillation with Future (DwF) loss. The distillation is carried out by two teachers, i.e., the detector trained on the previous stage and an expert detector trained with only current data. The student detector absorbs the knowledge of old and current classes from the old and expert models, respectively, in a class-by-class manner, preventing the detector from catastrophic forgetting and facilitating learning of current classes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 296, + 482, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 355 + ], + "type": "text", + "content": "We conduct extensive experiments on the PASCAL VOC [12] and COCO [37] datasets to evaluate the effectiveness of our BPF, which outperforms other state-of-the-art (SOTA) methods under multiple settings in a memory-free way. Moreover, we also conduct comprehensive ablation studies and visualizations to help better understand how each component of the BPF works." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 373, + 243, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 373, + 243, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 373, + 243, + 385 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 397, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 482, + 529 + ], + "type": "text", + "content": "Object Detection. Traditional object detectors can be primarily classified into two streams: one-stage [36,52,58,61] and two-stage detectors [18,22,35,53]. Two-stage detectors first predict several coarse candidate proposals via region proposal extractors [18,53] and then adopt a region of interest (RoI) head to refine these proposals and output final predictions. Unlike two-stage detectors, one-stage counterparts directly generate final outputs without predicting candidate proposals. Despite performing well under the standard training setup, both of them fail to generalize to the incremental training setup due to lacking previous training data. Without losing generality, this paper concentrates on enhancing two-stage detectors, i.e., Faster R-CNN [53], enabling it to learn new classes incrementally while retaining previously acquired knowledge." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Incremental Learning. Incremental learning methods can be mainly divided into three categories: memory-based [4, 24, 32, 33, 48, 51, 56, 59, 62], regularization-based [6, 8, 30, 34, 66], and structure-based [42, 45, 46, 54, 60]. Memory-based methods store a handful of samples for replaying [24, 40, 41] or suggest generating samples [29, 62] for data compensation in new stages. Some regularization-based methods try to eliminate the effect of new tasks on previously learned knowledge by identifying the most crucial parameters [30, 43, 62], while some other regularization-based methods propose to distill knowledge [23] from old to current models by transiting knowledge on logits, final features or intermediate features [8, 10, 11, 24, 34]. Structure-based methods dedicate specific parameters to each task, freezing them to mitigate forgetting. In this study, we focus on the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bridge Past and Future" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "regularization-based knowledge distillation approach for object detection and introduce a novel distillation with future information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 144, + 481, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 144, + 481, + 337 + ], + "spans": [ + { + "bbox": [ + 132, + 144, + 481, + 337 + ], + "type": "text", + "content": "Incremental Object Detection. Most previous works [9,28,31,40,50,55] focus on distilling old classes' knowledge of intermediate features [5,21,39,41,49,65,67], region proposal network [5, 49, 67], or RoI head [13] to prevent forgetting. A stream of methods also pays attention to classification loss. PPAS [67] introduces a pseudo-positive-aware sampling algorithm to identify regions corresponding to old classes and prevents them from being sampled as background. MMA [5] proposes unbiased classification loss, consolidating the background and all old classes into one entity, which aims to minimize the optimization objective conflict between the current stage's background and the past stages' foreground. However, it diminishes the ability to distinguish old classes from the background. Other methods [1,20,25,26,41,64] focus on rehearsal to maintain the previous stage knowledge, either performing replay of the intermediate features [1], the images [25,26,40], or the instances [41]. Incdet [38] presents a parameter isolation strategy that builds upon EWC [30]. In this study, in addition to preserving knowledge for old classes, we also take future classes into account, aligning optimization objectives consistently across all stages." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 352, + 201, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 201, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 201, + 364 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 374, + 228, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 228, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 228, + 386 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "text", + "content": "Problem Formulation. In incremental object detection, the training is performed over multiple learning stages, each one introducing a new set of classes to be detected. Let " + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "text", + "content": " denote the set of classes that are incrementally introduced to the object detector " + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "text", + "content": ". In the " + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "text", + "content": "-th training stage, a grouping of classes " + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_t" + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "text", + "content": " are introduced to the detector: " + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_t \\subset \\mathcal{C}" + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i \\cap \\mathcal{C}_j = \\emptyset" + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "text", + "content": ", for any " + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "inline_equation", + "content": "i, j \\leq t" + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "text", + "content": " denote the images containing annotated objects of classes in " + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_t" + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "text", + "content": ". Each image can contain multiple objects of different classes, but annotations " + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_t" + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "text", + "content": " are available only for those object instances that belong to classes in " + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_t" + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "text", + "content": ". The challenge in class incremental object detection is to continually update " + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t" + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{t+1}" + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "text", + "content": " by learning " + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{t+1}" + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "text", + "content": ", without access to " + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{D}_0, \\dots, \\mathcal{D}_t\\}" + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "text", + "content": " while maintaining original performance on " + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{C}_0, \\dots, \\mathcal{C}_t\\}" + }, + { + "bbox": [ + 132, + 392, + 481, + 524 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 529, + 481, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 529, + 481, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 529, + 481, + 636 + ], + "type": "text", + "content": "General Detector Training Pipeline. This work starts with the representative Faster RCNN-like detectors [53] " + }, + { + "bbox": [ + 132, + 529, + 481, + 636 + ], + "type": "inline_equation", + "content": "f = \\{f_b, f_{rpn}, f_{roi}\\}" + }, + { + "bbox": [ + 132, + 529, + 481, + 636 + ], + "type": "text", + "content": ", which generally includes a backbone network " + }, + { + "bbox": [ + 132, + 529, + 481, + 636 + ], + "type": "inline_equation", + "content": "f_b" + }, + { + "bbox": [ + 132, + 529, + 481, + 636 + ], + "type": "text", + "content": ", Region Proposal Network (RPN) " + }, + { + "bbox": [ + 132, + 529, + 481, + 636 + ], + "type": "inline_equation", + "content": "f_{rpn}" + }, + { + "bbox": [ + 132, + 529, + 481, + 636 + ], + "type": "text", + "content": " and Region of Interest Head (RoI Head) " + }, + { + "bbox": [ + 132, + 529, + 481, + 636 + ], + "type": "inline_equation", + "content": "f_{roi}" + }, + { + "bbox": [ + 132, + 529, + 481, + 636 + ], + "type": "text", + "content": ". During training, all training proposals are categorized into positive, negative, and ignored samples for supervised training based on their IoUs with the ground truth. Positive samples are assigned to predict the class of their matched ground truth, whereas negative samples are designated as background. Thus, the objects without annotations in the current stage will be classified as background, hindering the learning of incremental models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 480, + 665 + ], + "type": "text", + "content": "Common Distillation Methods for Incremental Object Detection. To prevent forgetting in IOD, a widely adopted solution involves knowledge distil-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Mo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 317, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 317, + 129 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 317, + 129 + ], + "type": "text", + "content": "lation in RPN and RoI head [5,49,57,67]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 261, + 134, + 481, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 134, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 261, + 134, + 481, + 148 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i s t}} = \\mathcal {L} _ {\\text {d i s t}} ^ {r p n} + \\mathcal {L} _ {\\text {d i s t}} ^ {r o i}. \\tag {1}", + "image_path": "ea8362e34ffe17c8428e0ea81235958278bb03b2cfbbc6d955e8cc99b56d7349.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 152, + 479, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 152, + 479, + 189 + ], + "spans": [ + { + "bbox": [ + 131, + 152, + 479, + 189 + ], + "type": "text", + "content": "The distillation in the RoI head includes the L2 loss between the box coordinates and the Kullback-Leibler divergence for class probabilities. MMA [5] notices the missing old annotations and proposes Unbiased Knowledge Distillation (UKD):" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 194, + 481, + 223 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 194, + 481, + 223 + ], + "spans": [ + { + "bbox": [ + 139, + 194, + 481, + 223 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i s t}, \\text {c l s}} ^ {\\text {r o i}} (i) = \\frac {1}{| \\mathcal {C} _ {t - 1} | + 1} \\left(p _ {i} ^ {b, t - 1} \\log \\left(p _ {i} ^ {b, t} + \\sum_ {c \\in \\mathcal {C} _ {t}} p _ {i} ^ {c, t}\\right) + \\sum_ {c \\in \\mathcal {C} _ {1: t - 1}} p _ {i} ^ {c, t - 1} \\log \\left(p _ {i} ^ {c, t}\\right)\\right), \\tag {2}", + "image_path": "e94ed7124105f0a37900a3054ffe23ad82dd6c2fad7f0a60a6888f3870ca3a14.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 229, + 482, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 229, + 482, + 362 + ], + "spans": [ + { + "bbox": [ + 131, + 229, + 482, + 362 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 229, + 482, + 362 + ], + "type": "inline_equation", + "content": "p_i^{c,t-1}" + }, + { + "bbox": [ + 131, + 229, + 482, + 362 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 229, + 482, + 362 + ], + "type": "inline_equation", + "content": "p_i^{c,t}" + }, + { + "bbox": [ + 131, + 229, + 482, + 362 + ], + "type": "text", + "content": " indicate the classification output for the proposal " + }, + { + "bbox": [ + 131, + 229, + 482, + 362 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 229, + 482, + 362 + ], + "type": "text", + "content": " and class " + }, + { + "bbox": [ + 131, + 229, + 482, + 362 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 131, + 229, + 482, + 362 + ], + "type": "text", + "content": " of the teacher (old) model and student (current) model, respectively, and " + }, + { + "bbox": [ + 131, + 229, + 482, + 362 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 131, + 229, + 482, + 362 + ], + "type": "text", + "content": " is the background class. As a common practice, 64 proposals are used for distillation, which are randomly selected out of the top 128 proposals with the highest objectness scores from the RPN network of the old model. However, given the class-agnostic characteristic of the RPN, the proposals for distillation cover objects of both old and current classes. In such cases, the old model, without seeing current classes, cannot provide useful knowledge for the current classes and even hinders their learning. Further, treating the background probability and current class probabilities as a unified entity during distillation inevitably diminishes the ability to differentiate between current classes and background." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 377, + 257, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 377, + 257, + 388 + ], + "spans": [ + { + "bbox": [ + 131, + 377, + 257, + 388 + ], + "type": "text", + "content": "3.2 Overall Framework" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 394, + 479, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 479, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 479, + 453 + ], + "type": "text", + "content": "Unlike classification tasks with a single label per input, incremental object detection presents a scenario where an image " + }, + { + "bbox": [ + 130, + 394, + 479, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_t" + }, + { + "bbox": [ + 130, + 394, + 479, + 453 + ], + "type": "text", + "content": " can encompass objects from the current class set " + }, + { + "bbox": [ + 130, + 394, + 479, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_t" + }, + { + "bbox": [ + 130, + 394, + 479, + 453 + ], + "type": "text", + "content": ", previous class sets " + }, + { + "bbox": [ + 130, + 394, + 479, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{1:t-1}" + }, + { + "bbox": [ + 130, + 394, + 479, + 453 + ], + "type": "text", + "content": " and future class sets " + }, + { + "bbox": [ + 130, + 394, + 479, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{t+1:\\infty}" + }, + { + "bbox": [ + 130, + 394, + 479, + 453 + ], + "type": "text", + "content": ". Nonetheless, the annotations " + }, + { + "bbox": [ + 130, + 394, + 479, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_t" + }, + { + "bbox": [ + 130, + 394, + 479, + 453 + ], + "type": "text", + "content": " are limited to bounding boxes and class labels for objects in " + }, + { + "bbox": [ + 130, + 394, + 479, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_t" + }, + { + "bbox": [ + 130, + 394, + 479, + 453 + ], + "type": "text", + "content": ", while objects of other classes are regarded as background." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 454, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 454, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 131, + 454, + 482, + 597 + ], + "type": "text", + "content": "To address the challenge of inconsistent optimization objectives, we introduce a novel strategy termed Bridge Past and Future (BPF). As shown in Figure 2, the method is divided into two parts. From the perspective of supervised learning, we use high-confidence predictions of old classes given by the old model as pseudo labels to bridge the information from the past stage (Section 3.3) and discard some potential objects from the background to bridge the future stage (Section 3.4). Regarding the distillation learning, we propose a novel Distillation with Future (DwF) loss by considering current classes in the distillation. The distillation probabilities are combined from two teacher models: the old model " + }, + { + "bbox": [ + 131, + 454, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{t-1}" + }, + { + "bbox": [ + 131, + 454, + 482, + 597 + ], + "type": "text", + "content": " for old classes and an expert model for current classes. By utilizing abundant information in background probability, the distillation takes both previous stages and the current stage into consideration (Section 3.5)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 612, + 250, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 612, + 250, + 624 + ], + "spans": [ + { + "bbox": [ + 131, + 612, + 250, + 624 + ], + "type": "text", + "content": "3.3 Bridging the Past" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 666 + ], + "type": "text", + "content": "To address the inconsistency in optimization objectives between the previous and current models, we bridge the past information to the current stage by incorporating pseudo supervision signals from past stages into the current model's" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bridge Past and Future" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 117, + 460, + 293 + ], + "blocks": [ + { + "bbox": [ + 157, + 117, + 460, + 293 + ], + "lines": [ + { + "bbox": [ + 157, + 117, + 460, + 293 + ], + "spans": [ + { + "bbox": [ + 157, + 117, + 460, + 293 + ], + "type": "image", + "image_path": "7649c9c1029ff8c148ec97a2b94c23bdbd993fd84fdbb5931ef097af45ffde81.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 304, + 482, + 381 + ], + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 381 + ], + "type": "text", + "content": "Fig. 2: The overall framework of our method. The top side illustrates the Bridge Past and Future (BPF) procedure, which identifies objects of past classes and excludes several potential objects of future classes to ensure consistent optimization during the entire training process. The bottom side shows the Distillation with Future (DwF) process, which employs both the old model " + }, + { + "bbox": [ + 130, + 304, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{t-1}" + }, + { + "bbox": [ + 130, + 304, + 482, + 381 + ], + "type": "text", + "content": " adept at detecting old categories and the interim model " + }, + { + "bbox": [ + 130, + 304, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t^{im}" + }, + { + "bbox": [ + 130, + 304, + 482, + 381 + ], + "type": "text", + "content": " trained on " + }, + { + "bbox": [ + 130, + 304, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 130, + 304, + 482, + 381 + ], + "type": "text", + "content": " and specialized in new categories, to conduct a comprehensive distillation across all categories for the current model " + }, + { + "bbox": [ + 130, + 304, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t" + }, + { + "bbox": [ + 130, + 304, + 482, + 381 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 405, + 482, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 482, + 476 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 482, + 476 + ], + "type": "text", + "content": "optimization process. The old model was trained under the supervision of human annotations, which is a high-quality pseudo labeler for old classes. We take the high-confidence inference results of the old model " + }, + { + "bbox": [ + 130, + 405, + 482, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{t-1}" + }, + { + "bbox": [ + 130, + 405, + 482, + 476 + ], + "type": "text", + "content": " on current training images as pseudo labels and combine them with current annotations to train the current detector, thereby aligning the supervision of old classes with previous stages. The left part of Figure 3 shows the pipeline of bridging the past strategy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 476, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 476, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 476, + 482, + 525 + ], + "type": "text", + "content": "Formally, given predictions " + }, + { + "bbox": [ + 130, + 476, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\hat{y}^{old} \\in \\mathcal{P}" + }, + { + "bbox": [ + 130, + 476, + 482, + 525 + ], + "type": "text", + "content": " of the old model " + }, + { + "bbox": [ + 130, + 476, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{t-1}" + }, + { + "bbox": [ + 130, + 476, + 482, + 525 + ], + "type": "text", + "content": " on current images, we first select some high-confidence regions of previous classes " + }, + { + "bbox": [ + 130, + 476, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{1:t-1}" + }, + { + "bbox": [ + 130, + 476, + 482, + 525 + ], + "type": "text", + "content": " with a confidence threshold " + }, + { + "bbox": [ + 130, + 476, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 476, + 482, + 525 + ], + "type": "text", + "content": ", followed by the Non-Maximum Suppression (NMS) operation to reduce duplication:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 195, + 532, + 481, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 532, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 195, + 532, + 481, + 552 + ], + "type": "interline_equation", + "content": "\\mathcal {U} = \\{j \\in \\mathcal {P}: \\max _ {c \\in \\mathcal {C} _ {1: t - 1}} \\hat {p} _ {j} ^ {\\text {o l d}} (c) > \\eta \\}, \\quad \\mathcal {E} = \\mathrm {N M S} (\\mathcal {U}). \\tag {3}", + "image_path": "f293e798bcbafc6468aadea3f4fb5fad4999ebe21b69171db14a201dcefae344.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 559, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 559, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 559, + 482, + 597 + ], + "type": "text", + "content": "Then, we further narrow down the predictions to a subset " + }, + { + "bbox": [ + 130, + 559, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\mathcal{W} \\subset \\mathcal{E}" + }, + { + "bbox": [ + 130, + 559, + 482, + 597 + ], + "type": "text", + "content": " that do not overlap with the ground-truth labels of the new categories via an IoU threshold " + }, + { + "bbox": [ + 130, + 559, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 559, + 482, + 597 + ], + "type": "text", + "content": ", ensuring a clear distinction between past and present classes:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 214, + 604, + 481, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 604, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 214, + 604, + 481, + 622 + ], + "type": "interline_equation", + "content": "\\mathcal {W} = \\{j \\in \\mathcal {E}: \\forall i \\in \\mathcal {Y} _ {t}, \\operatorname {I o U} \\left(\\hat {\\boldsymbol {b}} _ {j} ^ {\\text {o l d}}, \\boldsymbol {b} _ {i}\\right) \\leq \\lambda_ {1} \\}. \\tag {4}", + "image_path": "e65b5f15a1d1a561cd47cd81fecc77f3f1c66da3df2fa596adb4cea7a48a0e5f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "By modeling the previous supervision signals from the old model, we bridge the past stages to the current one, ensuring the current model's optimization direction encompasses the objectives of earlier stages, significantly mitigating" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Mo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 115, + 446, + 232 + ], + "blocks": [ + { + "bbox": [ + 167, + 115, + 446, + 232 + ], + "lines": [ + { + "bbox": [ + 167, + 115, + 446, + 232 + ], + "spans": [ + { + "bbox": [ + 167, + 115, + 446, + 232 + ], + "type": "image", + "image_path": "70c935fc92d0f32026cb7ebe198ff5f50ad55a53e2c2a73dc6d00fec1b1e5b56.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 242, + 482, + 298 + ], + "lines": [ + { + "bbox": [ + 131, + 242, + 482, + 298 + ], + "spans": [ + { + "bbox": [ + 131, + 242, + 482, + 298 + ], + "type": "text", + "content": "Fig. 3: Overview of Bridge Past and Future. We adopt the previous model " + }, + { + "bbox": [ + 131, + 242, + 482, + 298 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{t-1}" + }, + { + "bbox": [ + 131, + 242, + 482, + 298 + ], + "type": "text", + "content": " to predict some pseudo labels for past classes to complement their missing supervision in the current stage. Additionally, we exclude several proposals that are likely to be an object but are not included in the current ground truth and pseudo labels from the background to avoid classifying them into background mistakenly." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 321, + 479, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 479, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 479, + 370 + ], + "type": "text", + "content": "the forgetting problem. MMA [5] treats the background probability and old class probabilities as a unified entity in the classification loss, making old classes hard to separate from the background. While our method explicitly models the old classes in the current stage out of the background." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 388, + 261, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 388, + 261, + 401 + ], + "spans": [ + { + "bbox": [ + 132, + 388, + 261, + 401 + ], + "type": "text", + "content": "3.4 Bridging the Future" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 410, + 480, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 480, + 492 + ], + "type": "text", + "content": "Bridging the future aims to rectify the misalignment between the optimization goals of the current and upcoming stages, as objects from future class sets " + }, + { + "bbox": [ + 130, + 410, + 480, + 492 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{t + 1:\\infty}" + }, + { + "bbox": [ + 130, + 410, + 480, + 492 + ], + "type": "text", + "content": " in the current dataset " + }, + { + "bbox": [ + 130, + 410, + 480, + 492 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 130, + 410, + 480, + 492 + ], + "type": "text", + "content": " are classified as background. The core design is to find some salient objects in the background and exclude them from the negative samples during the training. By separating pure background from regions likely to contain future category objects in the current stage, we align the optimization objectives for background treatment across both current and future models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 494, + 480, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 480, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 480, + 554 + ], + "type": "text", + "content": "Specifically, we find that the activation in feature maps is a good indicator for salient objects. As illustrated in Figure 3, in the absence of annotations for old and future classes, foreground and background features demonstrate notable differences in spatial attention. We produce the attention map " + }, + { + "bbox": [ + 130, + 494, + 480, + 554 + ], + "type": "inline_equation", + "content": "A_{i} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 130, + 494, + 480, + 554 + ], + "type": "text", + "content": " from the backbone feature map " + }, + { + "bbox": [ + 130, + 494, + 480, + 554 + ], + "type": "inline_equation", + "content": "F_{i} = f_{b}(I_{i}) \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 130, + 494, + 480, + 554 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 246, + 563, + 481, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 563, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 246, + 563, + 481, + 597 + ], + "type": "interline_equation", + "content": "A _ {i} = \\operatorname {S o f t m a x} \\left(\\sum_ {c = 1} ^ {C} \\left| F _ {i} \\right| ^ {p}\\right), \\tag {5}", + "image_path": "9efaac5f7732cb0d0f225053af560260140a1c27253a2d0f71d87ec5cdbb30b6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 604, + 479, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 604, + 479, + 642 + ], + "spans": [ + { + "bbox": [ + 131, + 604, + 479, + 642 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 604, + 479, + 642 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 131, + 604, + 479, + 642 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 604, + 479, + 642 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 131, + 604, + 479, + 642 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 604, + 479, + 642 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 131, + 604, + 479, + 642 + ], + "type": "text", + "content": " denote the feature's height, width, and channel. We compute the attention score for each region, indicating the likelihood of containing foreground objects. The attention score " + }, + { + "bbox": [ + 131, + 604, + 479, + 642 + ], + "type": "inline_equation", + "content": "a_{i,j}^{roi}" + }, + { + "bbox": [ + 131, + 604, + 479, + 642 + ], + "type": "text", + "content": " for region " + }, + { + "bbox": [ + 131, + 604, + 479, + 642 + ], + "type": "inline_equation", + "content": "r_j" + }, + { + "bbox": [ + 131, + 604, + 479, + 642 + ], + "type": "text", + "content": " is calculated as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 231, + 652, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 652, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 231, + 652, + 480, + 667 + ], + "type": "interline_equation", + "content": "a _ {i, j} ^ {\\text {r o i}} = \\operatorname {A v g} \\left(\\operatorname {R o I P o o l} \\left(A _ {i}, r _ {j}\\right)\\right) \\in \\mathbb {R}. \\tag {6}", + "image_path": "c477e42d7a9c7218263e9f532b1d2415100d0621f6845bc693c3f66aebd28cdf.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bridge Past and Future" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 115, + 479, + 293 + ], + "blocks": [ + { + "bbox": [ + 136, + 115, + 479, + 293 + ], + "lines": [ + { + "bbox": [ + 136, + 115, + 479, + 293 + ], + "spans": [ + { + "bbox": [ + 136, + 115, + 479, + 293 + ], + "type": "image", + "image_path": "1d5787d9ef4e0988e0987e9d0f93fb93ceefe6246d848e1ab7214142d5058d50.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 301, + 482, + 390 + ], + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 390 + ], + "type": "text", + "content": "Fig. 4: The illustration of Distillation with Future strategy. An intermediate teacher model trained on the current dataset is used to compensate for the lack of current class information in the old model. For proposals overlapping ground truth of the current stage, since the intermediate teacher is specialized in detecting current classes, we directly inherit its probabilities on current classes and use the old model to enrich its background probability with old class knowledge. On the contrary, for proposals that do not overlap with GT, the old model is preferred, and the intermediate model is used as compensation. Combining two teachers makes the distillation class by class." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 413, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 486 + ], + "type": "text", + "content": "Regions with high attention scores " + }, + { + "bbox": [ + 130, + 413, + 482, + 486 + ], + "type": "inline_equation", + "content": "a_{i,j}^{roi}" + }, + { + "bbox": [ + 130, + 413, + 482, + 486 + ], + "type": "text", + "content": " from feature maps and objectness scores " + }, + { + "bbox": [ + 130, + 413, + 482, + 486 + ], + "type": "inline_equation", + "content": "o_j" + }, + { + "bbox": [ + 130, + 413, + 482, + 486 + ], + "type": "text", + "content": " from class-agnostic RPN imply a greater chance of being future category objects. We discard these RoIs when sampling negative samples when training the RoI head, thus maintaining the model's consistency with future stage background definitions. Regions with lower scores or having a considerable IoU with ground truth are considered reliable backgrounds." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 502, + 382, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 382, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 382, + 514 + ], + "type": "text", + "content": "3.5 Combining Current Classes into Distillation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Distillation is an effective way to prevent forgetting in incremental object detection. However, distilling from the teacher model, which is biased toward old classes, inevitably hinders the learning for current classes. To address this challenge, we introduce the Distillation with Future (DwF) loss to distill from the teacher model in a more fine-grained and adaptive way. Different from MMA [5], which aligns the current classes and the background as a whole with the old model's background, we distill each class one by one, preserving the distinction between current classes and the background. However, the model " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{t-1}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " from the previous stage has not been trained on current classes, thus the distillation on current classes can not be performed. We introduce another intermediate teacher model " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t^{im}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", which is trained using the dataset from the current stage " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " in a fully supervised way, as a supplement. Taking the intermediate model " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t^{im}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " as a" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Mo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "complementary teacher, we explicitly consider the current stage " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " as the future stage of stage " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": ", making the distillation future-aware." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": "As the old model " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{t-1}" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": " performs well on " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{1:t-1}" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": " while the intermediate model " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t^{im}" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": " is the expert in " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_t" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": ", we distill different regions with different combination of teacher models, as shown in Figure 4. Specifically, we divide the regions for distillation " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": " into two subsets " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_1, \\mathcal{R}_2 \\subset \\mathcal{R}" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": " that based on their intersection over union with the ground truth labels for the new categories " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_t" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 217, + 208, + 397, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 208, + 397, + 220 + ], + "spans": [ + { + "bbox": [ + 217, + 208, + 397, + 220 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {1} = \\{j \\in \\mathcal {R}: \\forall i \\in \\mathcal {Y} _ {t}, \\operatorname {I o U} (b _ {j}, b _ {i}) \\leq \\lambda_ {2.} \\},", + "image_path": "83342365d51e38e95efd93a2d9936bf58b959829aea105f03b315d0f1f4750fa.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 217, + 213, + 480, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 213, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 217, + 213, + 480, + 235 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {R} _ {1} = \\left\\{j \\in \\mathcal {K}: \\forall i \\in \\mathcal {Y} _ {t}, \\operatorname {I o U} \\left(b _ {j}, b _ {i}\\right) \\geq \\lambda_ {2}. \\right\\}, \\\\ \\mathcal {R} _ {2} = \\left\\{j \\in \\mathcal {R}: \\forall i \\in \\mathcal {Y} _ {t}, \\operatorname {I o U} \\left(b _ {j}, b _ {i}\\right) > \\lambda_ {2}. \\right\\}. \\end{array} \\tag {7}", + "image_path": "40213edaa9ac3febb7a3eb2b8221ee67262faeac4900a8198f228ce1393d74c6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 241, + 479, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 241, + 479, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 479, + 277 + ], + "type": "text", + "content": "For regions " + }, + { + "bbox": [ + 130, + 241, + 479, + 277 + ], + "type": "inline_equation", + "content": "r_i \\in \\mathcal{R}_1" + }, + { + "bbox": [ + 130, + 241, + 479, + 277 + ], + "type": "text", + "content": ", which are likely to be the regions for old classes, we take " + }, + { + "bbox": [ + 130, + 241, + 479, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{t-1}" + }, + { + "bbox": [ + 130, + 241, + 479, + 277 + ], + "type": "text", + "content": " as the primary model for distillation and reconstruct its background representation with the model " + }, + { + "bbox": [ + 130, + 241, + 479, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t^{im}" + }, + { + "bbox": [ + 130, + 241, + 479, + 277 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 235, + 285, + 480, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 285, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 235, + 285, + 480, + 300 + ], + "type": "interline_equation", + "content": "\\hat {p} _ {i} ^ {c, i m} = p _ {i} ^ {c, i m} \\times p _ {i} ^ {b, t - 1}, \\quad r _ {i} \\in \\mathcal {R} _ {1}, \\tag {8}", + "image_path": "3660768497477e9d52b3398742b5415c2ca80e72280733cfc0f5da28ea763466.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 306, + 480, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 480, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 480, + 382 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 306, + 480, + 382 + ], + "type": "inline_equation", + "content": "p_i^{b,t-1}" + }, + { + "bbox": [ + 130, + 306, + 480, + 382 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 306, + 480, + 382 + ], + "type": "inline_equation", + "content": "p_i^{c,im}" + }, + { + "bbox": [ + 130, + 306, + 480, + 382 + ], + "type": "text", + "content": " are the classification probabilities of background in the model " + }, + { + "bbox": [ + 130, + 306, + 480, + 382 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{t-1}" + }, + { + "bbox": [ + 130, + 306, + 480, + 382 + ], + "type": "text", + "content": " and the classification probabilities for current classes and background " + }, + { + "bbox": [ + 130, + 306, + 480, + 382 + ], + "type": "inline_equation", + "content": "c \\in \\mathcal{C}_t \\cup \\mathcal{B}" + }, + { + "bbox": [ + 130, + 306, + 480, + 382 + ], + "type": "text", + "content": " in the model " + }, + { + "bbox": [ + 130, + 306, + 480, + 382 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t^{im}" + }, + { + "bbox": [ + 130, + 306, + 480, + 382 + ], + "type": "text", + "content": " for the region " + }, + { + "bbox": [ + 130, + 306, + 480, + 382 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 130, + 306, + 480, + 382 + ], + "type": "text", + "content": " respectively. After weighting, " + }, + { + "bbox": [ + 130, + 306, + 480, + 382 + ], + "type": "inline_equation", + "content": "\\sum_{c=1}^{\\mathcal{C}_t \\cup \\mathcal{B}} \\hat{p}_i^{c,im} = p_i^{b,t-1}" + }, + { + "bbox": [ + 130, + 306, + 480, + 382 + ], + "type": "text", + "content": ". The final distillation probabilities for regions in " + }, + { + "bbox": [ + 130, + 306, + 480, + 382 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_1" + }, + { + "bbox": [ + 130, + 306, + 480, + 382 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 130, + 306, + 480, + 382 + ], + "type": "inline_equation", + "content": "[p_i^{C_{1:t-1},t-1}, \\hat{p}_i^{C_t \\cup \\mathcal{B},im}] \\in \\mathbb{R}^{|C_{1:t}|+1}" + }, + { + "bbox": [ + 130, + 306, + 480, + 382 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 383, + 480, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 383, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 383, + 480, + 418 + ], + "type": "text", + "content": "On the contrary, for regions " + }, + { + "bbox": [ + 130, + 383, + 480, + 418 + ], + "type": "inline_equation", + "content": "r_i \\in \\mathcal{R}_2" + }, + { + "bbox": [ + 130, + 383, + 480, + 418 + ], + "type": "text", + "content": ", they are regions for current classes; thus, the intermediate model " + }, + { + "bbox": [ + 130, + 383, + 480, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t^{im}" + }, + { + "bbox": [ + 130, + 383, + 480, + 418 + ], + "type": "text", + "content": " is taken as the primary model for distillation, and its background representation is reconstructed by the model " + }, + { + "bbox": [ + 130, + 383, + 480, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{t-1}" + }, + { + "bbox": [ + 130, + 383, + 480, + 418 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 233, + 425, + 480, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 425, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 233, + 425, + 480, + 439 + ], + "type": "interline_equation", + "content": "\\hat {p} _ {i} ^ {c, t - 1} = p _ {i} ^ {c, t - 1} \\times p _ {i} ^ {b, i m}, \\quad r _ {i} \\in \\mathcal {R} _ {2}, \\tag {9}", + "image_path": "2bb34c1c80e44a51578e7f33d53b1d07bd636e85271a7d29c9cb2fd909db5483.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 446, + 479, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 479, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 479, + 487 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 446, + 479, + 487 + ], + "type": "inline_equation", + "content": "\\sum_{c=1}^{\\mathcal{C}_{1:t-1} \\cup \\mathcal{B}} \\hat{p}_i^{c,t-1} = p_i^{b,im}" + }, + { + "bbox": [ + 130, + 446, + 479, + 487 + ], + "type": "text", + "content": ". The distillation probabilities for regions in " + }, + { + "bbox": [ + 130, + 446, + 479, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_2" + }, + { + "bbox": [ + 130, + 446, + 479, + 487 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 130, + 446, + 479, + 487 + ], + "type": "inline_equation", + "content": "[\\hat{p}_i^{\\mathcal{C}_{1:t-1},t-1}, p_i^{\\mathcal{C}_t,im}, \\hat{p}_i^{\\mathcal{B},t-1}] \\in \\mathbb{R}^{|C_{1:t}|+1}" + }, + { + "bbox": [ + 130, + 446, + 479, + 487 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 487, + 480, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 480, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 480, + 582 + ], + "type": "text", + "content": "With the expanded probabilities from the complimentary teacher models, the distillation on the classification head can be performed using a conventional Kullback-Leibler divergence. Regarding the box distillation, we use the output boxes from the old model " + }, + { + "bbox": [ + 130, + 487, + 480, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{t-1}" + }, + { + "bbox": [ + 130, + 487, + 480, + 582 + ], + "type": "text", + "content": " for regions in " + }, + { + "bbox": [ + 130, + 487, + 480, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_1" + }, + { + "bbox": [ + 130, + 487, + 480, + 582 + ], + "type": "text", + "content": " and the intermediate model " + }, + { + "bbox": [ + 130, + 487, + 480, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t^{im}" + }, + { + "bbox": [ + 130, + 487, + 480, + 582 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 487, + 480, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_2" + }, + { + "bbox": [ + 130, + 487, + 480, + 582 + ], + "type": "text", + "content": ". As a result, the complementary knowledge from two teachers, the expansion of background probability, and the combination of adaptive probability for different regions not only prevent the student from catastrophic forgetting but also facilitate the learning for current classes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 600, + 230, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 600, + 230, + 613 + ], + "spans": [ + { + "bbox": [ + 131, + 600, + 230, + 613 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 623, + 264, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 623, + 264, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 623, + 264, + 635 + ], + "type": "text", + "content": "4.1 Experiment Settings" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Datasets and Evaluation Metrics. Following previous works [5,20,25,26,41, 49,57,65,67], we evaluate our method on PASCAL VOC 2007 [12] and MS COCO" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bridge Past and Future" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 157, + 482, + 308 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 148 + ], + "type": "text", + "content": "Table 1: mAP@0.5 results on single incremental step on PASCAL VOC 2007. The best performance in each is presented with **bold**, and the second best is presented with **underline**. Methods with * use exemplars." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 157, + 482, + 308 + ], + "lines": [ + { + "bbox": [ + 134, + 157, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 134, + 157, + 482, + 308 + ], + "type": "table", + "html": "
Method19-115-510-105-15
1-19201-20Avg1-1516-201-20Avg1-1011-201-20Avg1-55-151-20Avg
Joint Training76.076.776.176.478.070.476.174.275.976.376.176.172.477.376.174.9
Fine-tuning12.062.814.537.414.259.225.436.79.562.536.036.06.963.149.135.0
ORE* [25]69.460.168.964.771.858.768.565.260.468.864.664.6----
OW-DETR* [20]70.262.069.866.172.259.869.166.063.567.965.765.7----
ILOD-Meta* [26]70.957.670.264.271.755.967.863.868.464.366.366.3----
ABR* [41]71.069.770.970.473.065.171.069.171.272.872.072.064.771.069.467.9
Faster ILOD [53]68.961.168.565.071.656.967.964.369.854.562.162.162.037.143.349.6
PPAS [67]70.553.069.261.8----63.560.061.861.8----
MVC [65]70.260.669.765.469.457.966.563.766.266.066.166.1----
PROB [69]73.948.572.661.573.560.870.167.066.067.266.566.5----
PseudoRM [64]72.967.372.670.173.460.970.366.969.168.668.968.9----
MMA [5]71.163.470.767.273.060.569.966.769.363.966.666.666.857.259.662.0
BPF (Ours)74.565.374.169.975.963.072.769.571.774.072.972.966.475.373.070.9
", + "image_path": "a203671c09b13247dcfd6fb18836b63ee49cc805ebe1c145acb82611e5291cb8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 333, + 481, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 481, + 404 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 481, + 404 + ], + "type": "text", + "content": "2017 [37] datasets. PASCAL VOC 2007 dataset comprises 9,963 images across 20 categories. The COCO 2017 dataset encompasses objects from 80 categories, with around " + }, + { + "bbox": [ + 132, + 333, + 481, + 404 + ], + "type": "inline_equation", + "content": "118\\mathrm{k}" + }, + { + "bbox": [ + 132, + 333, + 481, + 404 + ], + "type": "text", + "content": " images for training and 5,000 images for validation. The mean average precision at the 0.5 IoU threshold " + }, + { + "bbox": [ + 132, + 333, + 481, + 404 + ], + "type": "inline_equation", + "content": "(\\mathrm{mAP}@0.5)" + }, + { + "bbox": [ + 132, + 333, + 481, + 404 + ], + "type": "text", + "content": " is used as the primary evaluation metric for the VOC dataset, and the mean average precision ranging from 0.5 to 0.95 is the main evaluation metric for the COCO dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 406, + 481, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 406, + 481, + 454 + ], + "spans": [ + { + "bbox": [ + 132, + 406, + 481, + 454 + ], + "type": "text", + "content": "For each incremental setting (A-B), the first number A denotes the number of classes in the first stage and the second number is the number of classes newly introduced in each new stage. Note that the columns with gray background in the table of experimental results represent the average AP among all classes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 460, + 481, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 481, + 544 + ], + "type": "text", + "content": "Implementation Details. Similar to [5,25,26,41,49], we build our incremental object detector based on Faster R-CNN [53] with R50. Our method can easily be adapted to transformer-based detectors [3,17,68]. We conduct the experiments under a strict rehearsal-free setting, where no memory is used. We set " + }, + { + "bbox": [ + 132, + 460, + 481, + 544 + ], + "type": "inline_equation", + "content": "\\eta = 0.75" + }, + { + "bbox": [ + 132, + 460, + 481, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 460, + 481, + 544 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 0.7" + }, + { + "bbox": [ + 132, + 460, + 481, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 460, + 481, + 544 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 0.5" + }, + { + "bbox": [ + 132, + 460, + 481, + 544 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 132, + 460, + 481, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 132, + 460, + 481, + 544 + ], + "type": "text", + "content": " in Equation (4), we use an IOU threshold of 0.3 to divide it into two sets. For the set with IoU " + }, + { + "bbox": [ + 132, + 460, + 481, + 544 + ], + "type": "inline_equation", + "content": "< 0.3" + }, + { + "bbox": [ + 132, + 460, + 481, + 544 + ], + "type": "text", + "content": ", the supervision signal weights are set to 1.0, while the other set is assigned a weight of 0.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 567, + 280, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 567, + 280, + 580 + ], + "spans": [ + { + "bbox": [ + 133, + 567, + 280, + 580 + ], + "type": "text", + "content": "4.2 Quantitative Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 593, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 593, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 593, + 481, + 665 + ], + "type": "text", + "content": "Following previous work [5,7,41,49,57,65,67], our method is evaluated on settings with a range of initial classes and incorporating one or more incremental tasks. We benchmark our method against two baselines: Fine-Tuning, where the model is incrementally trained on new data without any regularization strategy or data replay, and Joint Training, which involves training the model on the complete dataset using all annotations." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Mo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 158, + 481, + 235 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: mAP@0.5 results on multiple incremental steps on PASCAL VOC 2007. The best performance in each is presented with **bold**, and the second best is presented with **underline**. Methods with * use exemplars." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 158, + 481, + 235 + ], + "lines": [ + { + "bbox": [ + 134, + 158, + 481, + 235 + ], + "spans": [ + { + "bbox": [ + 134, + 158, + 481, + 235 + ], + "type": "table", + "html": "
Method10-5 (3 tasks)5-5 (4 tasks)10-2 (6 tasks)15-1 (6 tasks)10-1 (10 tasks)
1-1011-201-201-56-201-201-1011-201-201-1516-201-201-1011-201-20
Joint Training75.976.376.172.477.376.175.976.376.178.070.476.175.976.376.1
Fine-tuning [41]5.330.618.00.518.313.83.813.68.70.010.55.30.05.12.6
ABR* [41]68.767.167.964.756.458.467.058.162.668.756.765.762.055.758.9
Faster ILOD [53]68.357.963.155.716.025.964.248.656.466.944.561.352.941.547.2
MMA [5]66.761.864.262.331.238.965.053.159.168.354.364.159.248.353.8
BPF (Ours)69.168.268.760.663.162.568.756.362.571.553.166.962.248.355.2
", + "image_path": "efafb6105bcc185fc98598180939623aad7486bfee200166883cc4cb65f6e3d8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 149, + 263, + 323, + 371 + ], + "blocks": [ + { + "bbox": [ + 147, + 240, + 323, + 262 + ], + "lines": [ + { + "bbox": [ + 147, + 240, + 323, + 262 + ], + "spans": [ + { + "bbox": [ + 147, + 240, + 323, + 262 + ], + "type": "text", + "content": "Table 3: mAP results on COCO2017. Methods with * use exemplars." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 263, + 323, + 371 + ], + "lines": [ + { + "bbox": [ + 149, + 263, + 323, + 371 + ], + "spans": [ + { + "bbox": [ + 149, + 263, + 323, + 371 + ], + "type": "table", + "html": "
Method40-4070-10
APAP50AP75APAP50AP75
Joint Training36.757.839.836.757.839.8
Fine-tuning [41]19.031.220.45.68.66.2
ILOD-Meta* [26]23.840.524.4---
ABR* [41]34.557.835.231.152.932.7
Faster ILOD [49]20.640.1-21.339.9-
PseudoRM [64]25.344.4----
MMA [5]33.056.634.630.252.131.5
BPF (Ours)34.454.337.336.256.838.9
", + "image_path": "22833711d601c2e587fa126eba80438a3045eb0658988d2ff73871eda0b9cd0a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 340, + 266, + 456, + 318 + ], + "blocks": [ + { + "bbox": [ + 328, + 243, + 470, + 264 + ], + "lines": [ + { + "bbox": [ + 328, + 243, + 470, + 264 + ], + "spans": [ + { + "bbox": [ + 328, + 243, + 470, + 264 + ], + "type": "text", + "content": "Table 4: Ablation study of various combinations of teacher models." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 340, + 266, + 456, + 318 + ], + "lines": [ + { + "bbox": [ + 340, + 266, + 456, + 318 + ], + "spans": [ + { + "bbox": [ + 340, + 266, + 456, + 318 + ], + "type": "table", + "html": "
DistillationVOC(10-10)
Lroi dist,clsLroi dist,bbox1-1011-201-20
λ2= 1.0 part boxes71.573.372.4
λ2= 0.5 part boxes71.774.072.9
λ2= 0.5 all boxes71.374.472.9
", + "image_path": "d54406906e15d6154fcce269cb255555688f09f03d9d86d3f0d2f7602fe46438.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 336, + 338, + 460, + 369 + ], + "blocks": [ + { + "bbox": [ + 352, + 327, + 446, + 337 + ], + "lines": [ + { + "bbox": [ + 352, + 327, + 446, + 337 + ], + "spans": [ + { + "bbox": [ + 352, + 327, + 446, + 337 + ], + "type": "text", + "content": "Table 5: Effect of BF." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 336, + 338, + 460, + 369 + ], + "lines": [ + { + "bbox": [ + 336, + 338, + 460, + 369 + ], + "spans": [ + { + "bbox": [ + 336, + 338, + 460, + 369 + ], + "type": "table", + "html": "
BFVOC(5-15)VOC(10-10)VOC(15-5)
1-56-201-201-1011-201-201-1516-20
×66.374.472.471.273.372.375.662.8
√66.475.373.071.774.072.975.963.0
", + "image_path": "d99e2ef21274c104781bba063b2293a0c49e7f1cb04ff0c318faac95165b59ef.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 376, + 480, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 480, + 449 + ], + "type": "text", + "content": "PASCAL VOC 2007. For PASCAL VOC 2007, we order the classes alphabetically and evaluate our method with one or multiple training steps. We perform our experiments by adding 1 (19-1), 5 (15-5), 10 (10-10), or 15 (5-15) classes in a single incremental step. For multi-step incremental settings, we evaluate 10-5, 5-5, 10-2, 15-1, and 10-1 settings, where we add 5, 5, 2, 1, and 1 classes respectively at every step until all 20 classes are seen." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 449, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 481, + 605 + ], + "type": "text", + "content": "- Single-step Incremental Settings: Table 1 shows our BPF methods against the existing methods using rehearsal or not. Rehearsal-based methods are not compared fairly with our BPF since we do not store old samples and use replay memory. As shown in Table 1, BPF consistently outperforms all previous methods, including those designed to combat forgetting using exemplars, validating the superiority of our approach. In particular, in the 19-1, 15-5, 10-10, and 5-15 settings, BPF significantly improved over MMA [5] by " + }, + { + "bbox": [ + 130, + 449, + 481, + 605 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 130, + 449, + 481, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 449, + 481, + 605 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 130, + 449, + 481, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 449, + 481, + 605 + ], + "type": "inline_equation", + "content": "6.3\\%" + }, + { + "bbox": [ + 130, + 449, + 481, + 605 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 449, + 481, + 605 + ], + "type": "inline_equation", + "content": "13.4\\%" + }, + { + "bbox": [ + 130, + 449, + 481, + 605 + ], + "type": "text", + "content": " on mAP@0.5 across all classes. Similarly, BPF outperforms the best rehearsal-based method ABR [41] by " + }, + { + "bbox": [ + 130, + 449, + 481, + 605 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 130, + 449, + 481, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 449, + 481, + 605 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 130, + 449, + 481, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 449, + 481, + 605 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 130, + 449, + 481, + 605 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 449, + 481, + 605 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 130, + 449, + 481, + 605 + ], + "type": "text", + "content": ". The Avg metric equally averages old and new classes mAP, which straightly reports the incremental ability without the influence of the number of classes. BPF also outperforms most methods on the Avg metric, demonstrating BPF's adaptiveness in learning new classes and preserving the knowledge of old classes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "- Multi-step Incremental Settings: The issues of inconsistent optimization objectives across multiple stages and catastrophic forgetting are more crucial under the longer incremental settings. As shown in Table 2, BPF consistently outperforms MMA [5] across all the settings. Specifically, BPF improves over MMA by " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": ", and even " + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "inline_equation", + "content": "23.6\\%" + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": " at 1-20 mAP@0.5 under" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bridge Past and Future" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 136, + 473, + 211 + ], + "blocks": [ + { + "bbox": [ + 212, + 114, + 400, + 126 + ], + "lines": [ + { + "bbox": [ + 212, + 114, + 400, + 126 + ], + "spans": [ + { + "bbox": [ + 212, + 114, + 400, + 126 + ], + "type": "text", + "content": "Table 6: Ablation study on each component." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 136, + 473, + 211 + ], + "lines": [ + { + "bbox": [ + 140, + 136, + 473, + 211 + ], + "spans": [ + { + "bbox": [ + 140, + 136, + 473, + 211 + ], + "type": "table", + "html": "
ModelBridge the PastBridge the futureDistillation with FutureVOC(10-10)VOC(10-5)
1-1011-201-201-1011-1516-20
(a)58.172.465.354.469.959.3
(b)✓71.272.171.769.173.560.0
(c)✓61.073.367.154.870.758.7
(d)✓✓71.972.772.370.473.758.7
(e)✓✓✓71.774.072.969.175.261.2
", + "image_path": "14a1be442ad249c9ae2b29655314ffda415ee6a42ee535ac7a4503044eaa487b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 231, + 479, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 231, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 479, + 304 + ], + "type": "text", + "content": "the 10-1, 15-1, 10-2, 10-5, and 5-5 settings and enjoy improvement across all learning stages. Moreover, we find that even without storing memory, BPF still outperforms ABR [41] by " + }, + { + "bbox": [ + 130, + 231, + 479, + 304 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 130, + 231, + 479, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 231, + 479, + 304 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 130, + 231, + 479, + 304 + ], + "type": "text", + "content": " on overall mAP@0.5, " + }, + { + "bbox": [ + 130, + 231, + 479, + 304 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 130, + 231, + 479, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 231, + 479, + 304 + ], + "type": "inline_equation", + "content": "6.7\\%" + }, + { + "bbox": [ + 130, + 231, + 479, + 304 + ], + "type": "text", + "content": " mAP@0.5 on new classes under 10-5 and 5-5 settings. In the settings of 10-2 and 10-1, limited by the small incremental data, BPF is inferior to ABR, but considering we do not require memory, these losses are acceptable." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 308, + 480, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 480, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 480, + 368 + ], + "type": "text", + "content": "MS COCO 2017. On the COCO2017 dataset, we perform experiments on 40-40 and 70-10 settings, adding 40 and 10 classes, respectively, following [41]. As illustrated in Table 3, our method improves over MMA on average " + }, + { + "bbox": [ + 130, + 308, + 480, + 368 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 130, + 308, + 480, + 368 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 130, + 308, + 480, + 368 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 130, + 308, + 480, + 368 + ], + "type": "text", + "content": " on 40-40 settings and by " + }, + { + "bbox": [ + 130, + 308, + 480, + 368 + ], + "type": "inline_equation", + "content": "6.0\\%" + }, + { + "bbox": [ + 130, + 308, + 480, + 368 + ], + "type": "text", + "content": " on 70-10 settings. These results once again confirm the effectiveness of our method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 384, + 306, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 384, + 306, + 396 + ], + "spans": [ + { + "bbox": [ + 131, + 384, + 306, + 396 + ], + "type": "text", + "content": "4.3 Analysis and Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 402, + 481, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 481, + 617 + ], + "type": "text", + "content": "We examine the contributions of the \"Bridge the Past\", \"Bridge the Future\", and \"Distillation with Future\" in Table 6 within the VOC 10-10 and 10-5 settings. We take the unbiased knowledge distillation proposed by [5] as the baseline model. By bridging the past, our model (b) aligns its optimization objectives with earlier ones, effectively reducing catastrophic forgetting of old classes, greatly enhancing the performance on detecting old classes. Compared to the baseline (a), it significantly improved by " + }, + { + "bbox": [ + 130, + 402, + 481, + 617 + ], + "type": "inline_equation", + "content": "13.1\\%" + }, + { + "bbox": [ + 130, + 402, + 481, + 617 + ], + "type": "text", + "content": " in the old classes and by " + }, + { + "bbox": [ + 130, + 402, + 481, + 617 + ], + "type": "inline_equation", + "content": "6.4\\%" + }, + { + "bbox": [ + 130, + 402, + 481, + 617 + ], + "type": "text", + "content": " for all classes on the 10-10 setting. Owing to bridging the future, our model (c) maintains consistent optimization objectives with future models regarding the background, making it easier to incrementally learn new classes. Compared to the baseline (a) on the 10-10 setting, it improves by " + }, + { + "bbox": [ + 130, + 402, + 481, + 617 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 130, + 402, + 481, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 402, + 481, + 617 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 130, + 402, + 481, + 617 + ], + "type": "text", + "content": " on the old classes and the new classes, respectively. Combining Past and Future (d), the model outperforms consistently on each stage. In the Distillation with Future strategy, we leverage the intermediate model to aid the old model in modeling new classes, conducting distillation across all categories for the current model. Table 6 (e) shows that a comprehensive joint teacher model facilitates improved learning of all categories through knowledge distillation, with a significant improvement " + }, + { + "bbox": [ + 130, + 402, + 481, + 617 + ], + "type": "inline_equation", + "content": "(+1.3\\%)" + }, + { + "bbox": [ + 130, + 402, + 481, + 617 + ], + "type": "text", + "content": " AP) in new classes. Similar results can be found in the 10-5 setting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "We further verify the \"Bridge the future\" (BF) in Table 5. There is a clear trend that as the number of considered future classes increases, the BF shows increasing improvement in future classes (" + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "+0.9\\%" + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": " in 5-15) without degrading the old classes' performance." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Mo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 146, + 114, + 227, + 175 + ], + "blocks": [ + { + "bbox": [ + 136, + 137, + 145, + 152 + ], + "lines": [ + { + "bbox": [ + 136, + 137, + 145, + 152 + ], + "spans": [ + { + "bbox": [ + 136, + 137, + 145, + 152 + ], + "type": "text", + "content": "BPF" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 146, + 114, + 227, + 175 + ], + "lines": [ + { + "bbox": [ + 146, + 114, + 227, + 175 + ], + "spans": [ + { + "bbox": [ + 146, + 114, + 227, + 175 + ], + "type": "image", + "image_path": "6c89ff87d065a6f9ab29ccdd0fa7c6cc10189fddcc952b34b2014ea161568258.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 228, + 114, + 307, + 175 + ], + "blocks": [ + { + "bbox": [ + 228, + 114, + 307, + 175 + ], + "lines": [ + { + "bbox": [ + 228, + 114, + 307, + 175 + ], + "spans": [ + { + "bbox": [ + 228, + 114, + 307, + 175 + ], + "type": "image", + "image_path": "1900f9ab1924fb3c62d64e4c1e50eb0952d142c0a041630e48aad19db45b424c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 389, + 175 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 389, + 175 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 389, + 175 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 389, + 175 + ], + "type": "image", + "image_path": "5fba540b8bad63031c31db4f4ed03720484fa76007fe7ece8774f19623a4d289.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 391, + 114, + 480, + 175 + ], + "blocks": [ + { + "bbox": [ + 391, + 114, + 480, + 175 + ], + "lines": [ + { + "bbox": [ + 391, + 114, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 391, + 114, + 480, + 175 + ], + "type": "image", + "image_path": "2ae0cd7cf71505a1319cf2b46972aca1a345b6ac0dccad71432251cc02a6b3ee.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 146, + 176, + 227, + 236 + ], + "blocks": [ + { + "bbox": [ + 137, + 197, + 145, + 216 + ], + "lines": [ + { + "bbox": [ + 137, + 197, + 145, + 216 + ], + "spans": [ + { + "bbox": [ + 137, + 197, + 145, + 216 + ], + "type": "text", + "content": "MMA" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 146, + 176, + 227, + 236 + ], + "lines": [ + { + "bbox": [ + 146, + 176, + 227, + 236 + ], + "spans": [ + { + "bbox": [ + 146, + 176, + 227, + 236 + ], + "type": "image", + "image_path": "a446ce0cddd61694c5dca1fbd3446b6fee22ed55502d5a6dafa54842652e7db0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 228, + 176, + 307, + 236 + ], + "blocks": [ + { + "bbox": [ + 228, + 176, + 307, + 236 + ], + "lines": [ + { + "bbox": [ + 228, + 176, + 307, + 236 + ], + "spans": [ + { + "bbox": [ + 228, + 176, + 307, + 236 + ], + "type": "image", + "image_path": "44113d95ec2825f062d1059c5fe1e4c46b0c7443dc8da63717d267f6f0e27573.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 245, + 482, + 291 + ], + "lines": [ + { + "bbox": [ + 131, + 245, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 131, + 245, + 482, + 291 + ], + "type": "text", + "content": "Fig. 5: Qualitative results for the model trained under the " + }, + { + "bbox": [ + 131, + 245, + 482, + 291 + ], + "type": "inline_equation", + "content": "10 + 10" + }, + { + "bbox": [ + 131, + 245, + 482, + 291 + ], + "type": "text", + "content": " setting on the VOC 2007 test set. 'boat', 'cat', 'chair', and 'car' are old classes from the first stage, and 'person' and 'dog' are the classes from the second stage. Compared with MMA (bottom row), our BPF (top row) can produce reliable predictions on both old and new classes." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 176, + 389, + 236 + ], + "blocks": [ + { + "bbox": [ + 309, + 176, + 389, + 236 + ], + "lines": [ + { + "bbox": [ + 309, + 176, + 389, + 236 + ], + "spans": [ + { + "bbox": [ + 309, + 176, + 389, + 236 + ], + "type": "image", + "image_path": "21b6e2735b4c0f063c09312d9073f5d84cd6cb3e372007ddf6e71039537e4d49.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 391, + 176, + 480, + 236 + ], + "blocks": [ + { + "bbox": [ + 391, + 176, + 480, + 236 + ], + "lines": [ + { + "bbox": [ + 391, + 176, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 391, + 176, + 480, + 236 + ], + "type": "image", + "image_path": "c577aad62b4b6b7a6ea9da1c0f502309023b1b355eb0af3dc4ad9d8453645b37.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 315, + 480, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 480, + 386 + ], + "type": "text", + "content": "We also conducted a quantitative analysis of Bridge Past and Future. Under the second stage of the VOC 10-5-5 (3 tasks), the original background boxes have a " + }, + { + "bbox": [ + 130, + 315, + 480, + 386 + ], + "type": "inline_equation", + "content": "92.6\\%" + }, + { + "bbox": [ + 130, + 315, + 480, + 386 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 315, + 480, + 386 + ], + "type": "inline_equation", + "content": "83.9\\%" + }, + { + "bbox": [ + 130, + 315, + 480, + 386 + ], + "type": "text", + "content": " Recall50 rate for old and future classes, while the Recall50 of pseudo labels for old classes and discarded boxes for future classes are " + }, + { + "bbox": [ + 130, + 315, + 480, + 386 + ], + "type": "inline_equation", + "content": "66.1\\%" + }, + { + "bbox": [ + 130, + 315, + 480, + 386 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 315, + 480, + 386 + ], + "type": "inline_equation", + "content": "17.6\\%" + }, + { + "bbox": [ + 130, + 315, + 480, + 386 + ], + "type": "text", + "content": " respectively, demonstrating the necessity and effectiveness of Bridge the Past and Bridge the Future." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 388, + 481, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 481, + 483 + ], + "type": "text", + "content": "Table 4 presents experiments on the effect of different combinations of teacher models. " + }, + { + "bbox": [ + 130, + 388, + 481, + 483 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 388, + 481, + 483 + ], + "type": "text", + "content": " in Equation (7) determines using which teacher model as the primary. Using the intermediate expert model as the primary on regions overlapping with gt " + }, + { + "bbox": [ + 130, + 388, + 481, + 483 + ], + "type": "inline_equation", + "content": "(\\lambda_{2} = 0.5)" + }, + { + "bbox": [ + 130, + 388, + 481, + 483 + ], + "type": "text", + "content": " outperforms using the old model " + }, + { + "bbox": [ + 130, + 388, + 481, + 483 + ], + "type": "inline_equation", + "content": "(\\lambda_{2} = 1.0)" + }, + { + "bbox": [ + 130, + 388, + 481, + 483 + ], + "type": "text", + "content": ", showing that distilling new class objects with the old model hinders their learning. For box distillation, we find that distilling boxes on primary classes (part boxes, i.e., only boxes for old classes in " + }, + { + "bbox": [ + 130, + 388, + 481, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_1" + }, + { + "bbox": [ + 130, + 388, + 481, + 483 + ], + "type": "text", + "content": " and boxes for new classes in " + }, + { + "bbox": [ + 130, + 388, + 481, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_2" + }, + { + "bbox": [ + 130, + 388, + 481, + 483 + ], + "type": "text", + "content": " participate in distillation while others are ignored) performs similarly with all classes (all boxes)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 505, + 227, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 505, + 227, + 516 + ], + "spans": [ + { + "bbox": [ + 132, + 505, + 227, + 516 + ], + "type": "text", + "content": "4.4 Visualization" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 528, + 480, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 480, + 588 + ], + "type": "text", + "content": "Visualization of Detection Results. We visualize the detection results in Figure 5 to illustrate the significant improvement compared to the previous methods qualitatively. Our method accurately detects both new and old class objects simultaneously. While MMA fails to detect old classes accurately, suffering from catastrophic forgetting." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "Visualization of Bridge Past and Future. We visualized the module of Bridge the Past and Bridge the Future separately in Figure 6. We effectively model the missing annotations for old class objects to bridge the past. As demonstrated in Figure 6(b), the attention maps clearly differentiate the foreground from the background, regardless of the presence of annotations. The discarded boxes (the third row) validate the effectiveness of our method." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bridge Past and Future" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 304, + 284 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 304, + 284 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 304, + 284 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 304, + 284 + ], + "type": "image", + "image_path": "d2dc35795df87af9ec32adb2e58f457e3b6338d686352c39360985d616d1abd6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 289, + 482, + 346 + ], + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 346 + ], + "type": "text", + "content": "Fig. 6: Visualization of Bridge Past and Future. Boxes in red represent the ground truth in the current stage. (a) In Bridge the Past, we effectively constructed pseudo labels of past classes. (b) In Bridge the Future, salient objects (marked in green boxes) can be easily detected from the attention maps and are excluded from the background regions. Best viewed in color." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 306, + 114, + 480, + 283 + ], + "blocks": [ + { + "bbox": [ + 306, + 114, + 480, + 283 + ], + "lines": [ + { + "bbox": [ + 306, + 114, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 306, + 114, + 480, + 283 + ], + "type": "image", + "image_path": "b041acdd977e3de8e6adaf39f524c1d89cbc9da25a537bfbc8212a357c10c8c8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 367, + 225, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 367, + 225, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 367, + 225, + 380 + ], + "type": "text", + "content": "5 Conclusions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 393, + 482, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 393, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 393, + 482, + 478 + ], + "type": "text", + "content": "Limitations. In our Bridge the Past procedure, we assume several objects of old classes may appear in the current training data. However, when the number of incremental classes is limited, e.g., increasing a single class in each stage, objects of old classes may rarely occur due to limited training images. This is expected to be alleviated by generating samples of old classes via the copy-paste strategy as in ABR [41], while it may introduce little stored samples. Detailed discussion can be found in the Supplement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 481, + 482, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 482, + 626 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 482, + 626 + ], + "type": "text", + "content": "Conclusions. In this work, we find that the concurrence of classes from different learning stages causes a severe information asymmetry, not only causing catastrophic forgetting for old classes but also hindering the learning of new classes. To tackle the problem, we propose the Bridge Past and Future method, which uses pseudo labels from the old model to fill in the missing annotations and exclude some potential future objects from the background, keeping the learning consistent across all stages. Further, Distillation with Future loss is proposed to solve the problem of the old teacher model's lack of knowledge of new classes. By combining the knowledge from the past to the future, our method consistently outperforms others across different learning stages in most incremental settings. To the best of our knowledge, we are the first to consider future classes during the learning, shedding light on a new aspect of Incremental Object Detection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Acknowledgments. This work was supported partially by the National Key Research and Development Program of China (2023YFA1008503), NSFC(U21A20471), Guangdong NSF Project (No. 2023B1515040025, 2020B1515120085)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Mo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 138, + 140, + 479, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 479, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 479, + 161 + ], + "type": "text", + "content": "1. Acharya, M., Hayes, T.L., Kanan, C.: Rodeo: Replay for online object detection. arXiv preprint arXiv:2008.06439 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "type": "text", + "content": "2. Cao, S., Joshi, D., Gui, L.Y., Wang, Y.X.: Contrastive mean teacher for domain adaptive object detectors. In: CVPR (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 184, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 480, + 205 + ], + "type": "text", + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 205, + 480, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 480, + 227 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 480, + 227 + ], + "type": "text", + "content": "4. Castro, F.M., Marin-Jiménez, M.J., Guil, N., Schmid, C., Alahari, K.: End-to-end incremental learning. In: ECCV (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 228, + 480, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 249 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 249 + ], + "type": "text", + "content": "5. Cermelli, F., Geraci, A., Fontanel, D., Caputo, B.: Modeling missing annotations for incremental learning in object detection. In: CVPR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 250, + 480, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 250, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 138, + 250, + 480, + 271 + ], + "type": "text", + "content": "6. Chaudhry, A., Dokania, P.K., Ajanthan, T., Torr, P.H.: Riemannian walk for incremental learning: Understanding forgetting and intransigence. In: ECCV (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 272, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 272, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 272, + 480, + 293 + ], + "type": "text", + "content": "7. Chen, L., Yu, C., Chen, L.: A new knowledge distillation for incremental object detection. In: IJCNN (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 293, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 315 + ], + "type": "text", + "content": "8. Dhar, P., Singh, R.V., Peng, K.C., Wu, Z., Chellappa, R.: Learning without memorizing. In: CVPR (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 316, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 480, + 336 + ], + "type": "text", + "content": "9. Dong, N., Zhang, Y., Ding, M., Bai, Y.: Class-incremental object detection. PR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 337, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 480, + 358 + ], + "type": "text", + "content": "0. Douillard, A., Cord, M., Ollion, C., Robert, T., Valle, E.: Podnet: Pooled outputs distillation for small-tasks incremental learning. In: ECCV (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 360, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 360, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 360, + 480, + 380 + ], + "type": "text", + "content": "1. Douillard, A., Rame, A., Couairon, G., Cord, M.: Dytox: Transformers for continual learning with dynamic token expansion. In: CVPR (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 380, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 480, + 403 + ], + "type": "text", + "content": "2. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. Springer IJCV (2010)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "type": "text", + "content": "3. Feng, T., Wang, M., Yuan, H.: Overcoming catastrophic forgetting in incremental object detection via elastic response distillation. In: CVPR (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 425, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 480, + 446 + ], + "type": "text", + "content": "4. Fu, S., Yan, J., Gao, Y., Xie, X., Zheng, W.S.: Asag: Building strong one-decoder-layer sparse detectors via adaptive sparse anchor generation. In: ICCV (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "text", + "content": "5. Gao, Y., Lin, K.Y., Yan, J., Wang, Y., Zheng, W.S.: Asyfod: An asymmetric adaptation paradigm for few-shot domain adaptive object detection. In: CVPR (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 469, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 469, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 469, + 480, + 490 + ], + "type": "text", + "content": "6. Gao, Y., Yang, L., Huang, Y., Xie, S., Li, S., Zheng, W.S.: Acrofod: An adaptive method for cross-domain few-shot object detection. In: ECCV (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "type": "text", + "content": "7. Gao, Z., Wang, L., Han, B., Guo, S.: Adamixer: A fast-converging query-based object detector. In: CVPR (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 512, + 318, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 318, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 318, + 522 + ], + "type": "text", + "content": "8. Girshick, R.: Fast r-cnn. In: ICCV (2015)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "text", + "content": "9. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921 (2021)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "text", + "content": "20. Gupta, A., Narayan, S., Joseph, K., Khan, S., Khan, F.S., Shah, M.: Ow-detr: Open-world detection transformer. In: CVPR (2022)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "text", + "content": "21. Hao, Y., Fu, Y., Jiang, Y.G., Tian, Q.: An end-to-end architecture for class incremental object detection with knowledge distillation. In: ICME (2019)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 138, + 590, + 454, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 454, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 454, + 599 + ], + "type": "text", + "content": "22. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask r-cnn. In: ICCV (2017)." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "type": "text", + "content": "23. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "text", + "content": "24. Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incrementally via rebalancing. In: CVPR (2019)" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "text", + "content": "25. Joseph, K., Khan, S., Khan, F.S., Balasubramanian, V.N.: Towards open world object detection. In: CVPR (2021)" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bridge Past and Future" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "26. Joseph, K., Rajasegaran, J., Khan, S., Khan, F.S., Balasubramanian, V.N.: Incremental object detection via meta-learning. IEEE TPAMI (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "27. Kang, B., Liu, Z., Wang, X., Yu, F., Feng, J., Darrell, T.: Few-shot object detection via feature reweighting. In: ICCV (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 162, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 480, + 194 + ], + "type": "text", + "content": "28. Kang, M., Zhang, J., Zhang, J., Wang, X., Chen, Y., Ma, Z., Huang, X.: Alleviating catastrophic forgetting of incremental object detection via within-class and between-class knowledge distillation. In: CVPR (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 195, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 480, + 217 + ], + "type": "text", + "content": "29. Kemker, R., Kanan, C.: Fearnet: Brain-inspired model for incremental learning. arXiv preprint arXiv:1711.10563 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 218, + 481, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 261 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 261 + ], + "type": "text", + "content": "30. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 262, + 480, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 480, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 480, + 295 + ], + "type": "text", + "content": "31. Li, D., Tasci, S., Ghosh, S., Zhu, J., Zhang, J., Heck, L.: Rilod: Near real-time incremental learning for object detection at the edge. In: Proceedings of the 4th ACM/IEEE Symposium on Edge Computing (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 296, + 480, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 296, + 480, + 317 + ], + "spans": [ + { + "bbox": [ + 132, + 296, + 480, + 317 + ], + "type": "text", + "content": "32. Li, M., Cong, Y., Liu, Y., Sun, G.: Class-incremental gesture recognition learning with out-of-distribution detection. In: IROS (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 318, + 480, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 318, + 480, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 480, + 339 + ], + "type": "text", + "content": "33. Li, Y.M., Zeng, L.A., Meng, J.K., Zheng, W.S.: Continual action assessment via task-consistent score-discriminative feature distribution modeling. TCVT (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 340, + 427, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 340, + 427, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 340, + 427, + 351 + ], + "type": "text", + "content": "34. Li, Z., Hoiem, D.: Learning without forgetting. IEEE TPAMI (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 352, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 480, + 373 + ], + "type": "text", + "content": "35. Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: CVPR (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "type": "text", + "content": "36. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: ICCV (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "type": "text", + "content": "37. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 419, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 480, + 441 + ], + "type": "text", + "content": "38. Liu, L., Kuang, Z., Chen, Y., Xue, J.H., Yang, W., Zhang, W.: Incdet: In defense of elastic weight consolidation for incremental object detection. IEEE TNNLS (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 441, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 480, + 464 + ], + "type": "text", + "content": "39. Liu, X., Yang, H., Ravichandran, A., Bhotika, R., Soatto, S.: Multi-task incremental learning for object detection. arXiv preprint arXiv:2002.05347 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 464, + 480, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 480, + 485 + ], + "type": "text", + "content": "40. Liu, Y., Schiele, B., Vedaldi, A., Rupprecht, C.: Continual detection transformer for incremental object detection. In: CVPR (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 487, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 480, + 508 + ], + "type": "text", + "content": "41. Liu, Y., Cong, Y., Goswami, D., Liu, X., van de Weijer, J.: Augmented box replay: Overcoming foreground shift for incremental object detection. In: ICCV (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 509, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 480, + 531 + ], + "type": "text", + "content": "42. Liu, Y., Cong, Y., Sun, G., Zhang, T., Dong, J., Liu, H.: L3doc: Lifelong 3d object classification. IEEE TIP (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 532, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 480, + 553 + ], + "type": "text", + "content": "43. Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. In: NeurIPS (2017)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 554, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 480, + 586 + ], + "type": "text", + "content": "44. Ma, J., Niu, Y., Xu, J., Huang, S., Han, G., Chang, S.F.: Digeo: Discriminative geometry-aware learning for generalized few-shot object detection. In: CVPR (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 587, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 480, + 609 + ], + "type": "text", + "content": "45. Mallya, A., Davis, D., Lazebnik, S.: Piggyback: Adapting a single network to multiple tasks by learning to mask weights. In: ECCV (2018)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 609, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 480, + 632 + ], + "type": "text", + "content": "46. Mallya, A., Lazebnik, S.: Packnet: Adding multiple tasks to a single network by iterative pruning. In: CVPR (2018)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "47. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., et al.: Simple open-vocabulary object detection. In: ECCV (2022)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Q. Mo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 654 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "48. Ostapenko, O., Puscas, M., Klein, T., Jahnichen, P., Nabi, M.: Learning to remember: A synaptic plasticity driven framework for continual learning. In: CVPR (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "text", + "content": "49. Peng, C., Zhao, K., Lovell, B.C.: Faster ilod: Incremental learning for object detectors based on faster rcnn. PR (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 172, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 172, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 172, + 481, + 194 + ], + "type": "text", + "content": "50. Peng, C., Zhao, K., Maksoud, S., Wang, T., Lovell, B.C.: Diode: dilatable incremental object detection. PR (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 481, + 216 + ], + "type": "text", + "content": "51. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: CVPR (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 481, + 237 + ], + "type": "text", + "content": "52. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 237, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 481, + 259 + ], + "type": "text", + "content": "53. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. In: NeurIPS (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 481, + 293 + ], + "type": "text", + "content": "54. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 293, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 481, + 326 + ], + "type": "text", + "content": "55. Shieh, J.L., Haq, Q.M.u., Haq, M.A., Karam, S., Chondro, P., Gao, D.Q., Ruan, S.J.: Continual learning strategy in one-stage object detection framework based on experience replay for autonomous driving vehicle. Sensors (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 326, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 481, + 347 + ], + "type": "text", + "content": "56. Shin, H., Lee, J.K., Kim, J., Kim, J.: Continual learning with deep generative replay. In: NeurIPS (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 347, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 481, + 369 + ], + "type": "text", + "content": "57. Shmelkov, K., Schmid, C., Alahari, K.: Incremental learning of object detectors without catastrophic forgetting. In: ICCV (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 369, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 481, + 391 + ], + "type": "text", + "content": "58. Tan, M., Pang, R., Le, Q.V.: Efficientdet: Scalable and efficient object detection. In: CVPR (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 391, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 481, + 413 + ], + "type": "text", + "content": "59. Tang, Y.M., Peng, Y.X., Zheng, W.S.: Learning to imagine: Diversify memory for incremental learning using unlabeled data. In: CVPR (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 413, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 481, + 435 + ], + "type": "text", + "content": "60. Tang, Y.M., Peng, Y.X., Zheng, W.S.: When prompt-based incremental learning does not meet strong pretraining. In: ICCV (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 435, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 481, + 456 + ], + "type": "text", + "content": "61. Tian, Z., Shen, C., Chen, H., He, T.: Fcos: Fully convolutional one-stage object detection. In: ICCV (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 457, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 481, + 489 + ], + "type": "text", + "content": "62. Wu, C., Herranz, L., Liu, X., Van De Weijer, J., Raducanu, B., et al.: Memory replay gans: Learning to generate new categories without forgetting. In: NeurIPS (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 490, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 481, + 511 + ], + "type": "text", + "content": "63. Yan, J., Yang, L., Gao, Y., Zheng, W.S.: Self-supervised cross-stage regional contrastive learning for object detection. In: ICME (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 511, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 481, + 544 + ], + "type": "text", + "content": "64. Yang, D., Zhou, Y., Hong, X., Zhang, A., Wei, X., Zeng, L., Qiao, Z., Wang, W.: Pseudo object replay and mining for incremental object detection. In: ACM MM (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 544, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 481, + 567 + ], + "type": "text", + "content": "65. Yang, D., Zhou, Y., Zhang, A., Sun, X., Wu, D., Wang, W., Ye, Q.: Multi-view correlation distillation for incremental object detection. PR (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 567, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 481, + 588 + ], + "type": "text", + "content": "66. Zenke, F., Poole, B., Ganguli, S.: Continual learning through synaptic intelligence. In: ICML (2017)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 588, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 588, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 588, + 481, + 611 + ], + "type": "text", + "content": "67. Zhou, W., Chang, S., Sosa, N., Hamann, H., Cox, D.: Lifelong object detection. arXiv preprint arXiv:2009.01129 (2020)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 481, + 632 + ], + "type": "text", + "content": "68. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. In: ICLR (2021)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 632, + 481, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 481, + 654 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 481, + 654 + ], + "type": "text", + "content": "69. Zohar, O., Wang, K.C., Yeung, S.: Prob: Probabilistic objectness for open world object detection. In: CVPR (2023)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bridge Past and Future" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Bridging Different Language Models and Generative Vision Models for Text-to-Image Generation/1764d397-410b-451c-a099-8966f6b5772c_content_list.json b/2024/Bridging Different Language Models and Generative Vision Models for Text-to-Image Generation/1764d397-410b-451c-a099-8966f6b5772c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a011ec5c6c0ec2b823011e6deed61a524c182e60 --- /dev/null +++ b/2024/Bridging Different Language Models and Generative Vision Models for Text-to-Image Generation/1764d397-410b-451c-a099-8966f6b5772c_content_list.json @@ -0,0 +1,1447 @@ +[ + { + "type": "text", + "text": "Bridging Different Language Models and Generative Vision Models for Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 269, + 140, + 736, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shihao Zhao $^{1}$ , Shaozhe Hao $^{1}$ , Bojia Zi $^{2}$ , Huaizhe Xu $^{3}$ , and Kwan-Yee K. Wong $^{1\\star}$", + "bbox": [ + 277, + 233, + 725, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The University of Hong Kong", + "bbox": [ + 393, + 276, + 607, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{shzhao, szhao, kykwong}@cs.hku.hk", + "bbox": [ + 377, + 292, + 625, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 The Chinese University of Hong Kong", + "bbox": [ + 367, + 305, + 635, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "bjzi@se.cuhk.edu.hk", + "bbox": [ + 426, + 320, + 575, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 The Hong Kong University of Science and Technology", + "bbox": [ + 313, + 333, + 687, + 347 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hxubr@connect.ust.hk", + "bbox": [ + 421, + 349, + 580, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Text-to-image generation has made significant advancements with the introduction of text-to-image diffusion models. These models typically consist of a language model that interprets user prompts and a vision model that generates corresponding images. As language and vision models continue to progress in their respective domains, there is a great potential in exploring the replacement of components in text-to-image diffusion models with more advanced counterparts. A broader research objective would therefore be to investigate the integration of any two unrelated language and generative vision models for text-to-image generation. In this paper, we explore this objective and propose LaVi-Bridge, a pipeline that enables the integration of diverse pre-trained language models and generative vision models for text-to-image generation. By leveraging LoRA and adapters, LaVi-Bridge offers a flexible and plug-and-play approach without requiring modifications to the original weights of the language and vision models. Our pipeline is compatible with various language models and generative vision models, accommodating different structures. Within this framework, we demonstrate that incorporating superior modules, such as more advanced language models or generative vision models, results in notable improvements in capabilities like text alignment or image quality. Extensive evaluations have been conducted to verify the effectiveness of LaVi-Bridge. Code is available at https://github.com/ShihaoZhaoZSH/LaVi-Bridge.", + "bbox": [ + 261, + 396, + 743, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Diffusion model $\\cdot$ Text-to-image generation", + "bbox": [ + 261, + 714, + 638, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 753, + 375, + 771 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, there have been remarkable advancements in the field of text-to-image generation, specifically through the use of diffusion models [9, 18, 43, 45].", + "bbox": [ + 212, + 785, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding authors", + "bbox": [ + 217, + 824, + 388, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5ec3a7ca92f91d7be5173d94c1dd33629d5b59caefa36e7e9386410ff58731c5.jpg", + "image_caption": [ + "Fig. 1: Overview of LaVi-Bridge. LaVi-Bridge is capable of integrating various language models and generative vision models. On the left side, we keep the vision model fixed and experiment with different language models in our pipeline. On the right side, we keep the language model fixed and try out different vision models. We display the visualization results alongside each combination." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 309 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These models have made significant contributions and have gained considerable attention for their exceptional performance. By leveraging large-scale training datasets alongside large deep models, text-to-image diffusion models are capable of producing high-quality images that faithfully align with the textual descriptions provided by users. This has rendered them highly applicable in real-world scenarios such as content creation and architectural design.", + "bbox": [ + 212, + 426, + 787, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-image diffusion models [2, 6, 29, 35, 38, 41, 50] typically consist of two key components, namely a language model and a generative vision model. The language model is responsible for comprehending the input prompts, whereas the vision model is tasked with generating images that align with the extracted context. Existing text-to-image diffusion models employ various language models and generative vision models and have gained widespread usage. For instance, Stable Diffusion (SD) [38] is a highly popular text-to-image diffusion model that employs the CLIP text encoder [31] as its language model and a U-Net [39] as its generative vision model. Another example is PixArt [6], a recently proposed text-to-image diffusion model that adopts the T5 [34] as its language model and a Vision Transformer (ViT) [11] as its generative vision model. These models are trained on a vast amount of text-image pairs, enabling seamless collaboration between their language modules and vision modules.", + "bbox": [ + 212, + 518, + 787, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The advancements in deep language models and deep vision models have witnessed rapid progress in recent years, with both fields experiencing continuous developments and the introduction of more powerful models. However, this rapid development poses a challenge for the research in text-to-image generation when it comes to integrating more advanced language or vision models into existing text-to-image diffusion models. The problem of how to integrate any two unrelated language and vision models is unexplored, and the impact of newly developed models on text-to-image generation capabilities also remains", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. Zhao et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "uncertain. The current situation highlights the presence of a gap between the language or vision modules within text-to-image diffusion models and the state-of-the-art models in their respective domains. Therefore, it has become crucial to address this gap and explore ways to incorporate more advanced language or vision models into existing text-to-image diffusion models. Furthermore, the broader challenge of integrating any pre-trained language model with any generative vision model deserves a thorough investigation.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, our objective is to delve into the aforementioned problem. We propose LaVi-Bridge, a flexible framework that facilitates the integration of diverse well-trained language models and generative vision models to achieve text-to-image generation. Our framework enables the integration of two unrelated language and vision models that have not been previously trained together, as shown in Fig. 1. Importantly, LaVi-Bridge does not require modifying the original weights of the language and vision models. Instead, it injects LoRA [20] into the language and vision models separately and utilizes an adapter to bridge these two modules. Moreover, LaVi-Bridge only necessitates a relatively small dataset to integrate different language models and generative vision models for text-to-image generation.", + "bbox": [ + 212, + 252, + 787, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We summarize the advantages and features of LaVi-Bridge as follows:", + "bbox": [ + 238, + 419, + 741, + 434 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. LaVi-Bridge is designed for text-to-image diffusion models and serves as a bridge, capable of connecting various pre-trained language models and generative vision models. Our framework can accommodate different model structures, including encoder-only, encoder-decoder, and decoder-only language models, as well as U-Net-based and Transformer-based generative vision models.", + "2. LaVi-Bridge utilizes LoRA and adapters, eliminating the need to modify the original weights of the models. It is more flexible and requires relatively small computing resources compared to training the entire diffusion model.", + "3. We evaluated various text-image alignment and image quality metrics on short prompts, long prompts, and compositional prompts. We also conducted extensive visualization. We then drew several conclusions. For instance, integrating superior models leads to improved performance in the corresponding modality, such as enhanced semantic understanding with advanced language models or improved image quality with more powerful generative vision models. Additionally, the diffusion model utilizing Llama-2 demonstrates exceptional semantic understanding, while the diffusion model utilizing the transformer in PixArt yields images with enhanced aesthetics." + ], + "bbox": [ + 222, + 446, + 785, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 746, + 387, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Language Models and Generative Vision Models", + "text_level": 1, + "bbox": [ + 212, + 781, + 663, + 796 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The mainstream Large Language Models (LLMs) [8, 32-34] are built based on the transformer structure [48], with three main types of architectures, namely", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "LaVi-Bridge", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "encoder-only, encoder-decoder, and decoder-only. All these three belong to Sequence to Sequence (Seq2Seq) [46]. The encoder-only architecture is exemplified by BERT [8]. CLIP text encoder [31] is based on BERT and further trained to align with the image domain. Models of this type excel at understanding the content of the input and generating outputs tailored to specific tasks. On the other hand, the encoder-decoder framework is adept at handling tasks that involve complex mappings between input and output sequences. Examples include T5 [34] and BART [24]. Recently, due to the tremendous success of ChatGPT, attention has been drawn to models that consist solely of a decoder, like GPT-3 [4] and Llama-2 [47]. The decoder-only architecture demonstrates exceptional performance in semantic understanding. For text-to-image generation, all three types of LLMs can provide effective semantic information to serve as conditions for image generation in diffusion models. In this paper, we explore and compare all these three types of language models.", + "bbox": [ + 212, + 146, + 787, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A generative vision model refers to a vision model with the ability to generate images or visual contents. There are two common types of structures, namely U-Net-based [39] and Transformer-based [11]. Generative Adversarial Networks (GANs) [14, 37, 51] employ a framework consisting of a discriminator and a generator, with the generator's structure based on U-Net. On the other hand, motivated by the success of GPT models, recent works have attempted to use the Transformer architecture for image generation in an autoregressive manner, with notable examples being DALLE [36] and CogView [10]. Another popular class of generative models is diffusion models [7, 18, 43, 45], which are based on the diffusion process and gradually denoise to produce natural images. Early diffusion models often employed U-Net as their generative vision model, such as Stable Diffusion, which scaled up the Latent Diffusion Model (LDM) [38] with larger data scales. Some recent works have started to replace the U-Net in diffusion models with Vision Transformer and have made significant progress, such as DiT [30], U-ViT [1] and PixArt [6]. In this paper, we focus on diffusion models and explore both U-Net-based and Transformer-based vision models.", + "bbox": [ + 212, + 357, + 787, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Text-to-Image Diffusion Models", + "text_level": 1, + "bbox": [ + 215, + 619, + 524, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Text-to-image diffusion models [2,9,29,38,41,44] are capable of generating images based on user prompts. These models consist of two main components, namely a language model and a vision model. The language module is responsible for understanding the text input provided by the user, extracting contextual information, and injecting it into the vision module to generate the desired image. Text-to-image diffusion models have paved the way for various exciting research areas, including image editing [3,22,27], controllable image generation [28,52,53], personalized object generation [12, 16, 40], as well as other interesting applications [5,13,15]. In the extensive exploration of diffusion models, researchers have utilized different language models and vision models. For instance, Stable Diffusion [38] employs CLIP text encoder [31] as its language model and U-Net as its vision model.Imagen [41] utilizes T5 [34] as its language model, which claims to enhance both sample fidelity and image-text alignment. ParaDiffusion [49]", + "bbox": [ + 212, + 643, + 787, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. Zhao et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "focuses on paragraph-to-image generation and leverages the powerful semantic understanding capability of Llama-2 [47] to comprehend lengthy sentences. PixArt [6], on the other hand, utilizes a ViT [11] as its vision model and achieves high image fidelity while being trained at a lower cost.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After training on a large dataset of text-image pairs [42], the language and vision models in the text-to-image diffusion model become closely intertwined. This tight coupling ensures a strong alignment between the provided text description and the generated image, but at the same time also limits the flexibility of the diffusion model. For instance, if a more advanced language or vision model becomes available, it may have the potential to enhance the text-to-image task. However, decoupling the language and vision modules in existing text-to-image diffusion models and replacing a module with a new one is nontrivial. Therefore, this paper explores the dilemma faced by text-to-image generation and proposes a framework that enables efficient integration of various language models and generative vision models.", + "bbox": [ + 212, + 207, + 787, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 393, + 330, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminary", + "text_level": 1, + "bbox": [ + 215, + 420, + 361, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A diffusion model is based on the diffusion process for image generation. This process consists of two stages, namely the forward process and the reverse process. During the forward process, Gaussian noise is progressively added to a natural image until the image becomes completely noisy. After that, during the reverse process, the noise is gradually eliminated over a series of time steps, resulting in a natural image. In the reverse process, a trainable vision model is used to predict and remove the noise. By employing this denoising model, we are able to obtain a natural image from Gaussian noise by denoising. Within a text-to-image diffusion model, there are two components at each time step, namely a language model $f$ and a vision model $g$ . The language model converts user input text $y$ into embeddings, which capture the semantic meaning of the text. On the other hand, the vision model, which is the denoising model aforementioned, encodes image features $z$ , extracting relevant visual information from the input images. The interaction between text embeddings and image features is achieved through cross-attention layers, which can be formulated as", + "bbox": [ + 212, + 441, + 787, + 667 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nc = f (y), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 439, + 676, + 784, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ = W _ {q} (z), K = W _ {k} (c), V = W _ {v} (c), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 703, + 784, + 719 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC r o s s A t t e n t i o n (Q, K, V) = \\text {s o f t m a x} (Q \\cdot K ^ {T}) \\cdot V, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 729, + 784, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W_{q}, W_{k}$ and $W_{v}$ are projection matrices.", + "bbox": [ + 214, + 753, + 557, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Language and Vision Alignment", + "text_level": 1, + "bbox": [ + 214, + 787, + 527, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LaVi-Bridge enables the integration of any two pre-trained language and generative vision models, even though these models are not related and have been", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "LaVi-Bridge", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7f8a055761e674f935c0550440915fa22f8ac8f5c5c47e96a7db6dc70b8c09aa.jpg", + "image_caption": [ + "Fig. 2: Pipeline of LaVi-Bridge. We select one model each from the language and vision model pools. We then freeze the pre-trained language and vision models and incorporate LoRA into both models. The connection between the language and vision models is established through an adapter. The only weights we need to train are the ones introduced by LoRA and the adapter." + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 784, + 300 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "trained separately. Here, we denote the language model as $f$ and the vision model as $g$ , as mentioned previously. If we directly interact the textual information and image information using Eq. (1), considering that $f$ and $g$ are trained independently, the parameters in the cross-attention layers of $g$ cannot comprehend the text embedding output by $f$ , resulting in meaningless model outputs.", + "bbox": [ + 212, + 422, + 782, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To establish a connection between them, LaVi-Bridge keeps the pre-trained language and vision models fixed and utilizes LoRA to introduce trainable parameters $\\Delta \\theta$ into both the language model and the vision model. In this context, we denote the language and vision models with LoRA as $f^{\\theta_1 + \\Delta \\theta_1}$ and $g^{\\theta_2 + \\Delta \\theta_2}$ , where $\\theta_{1}$ and $\\theta_{2}$ are the original parameters of $f$ and $g$ , respectively. Furthermore, we introduce an adapter as a bridge between the language model and vision model to facilitate better alignment. The adapter consists of stacked feedforward layers, denoted as $h$ . Consequently, the cross-attention layer can be expressed as", + "bbox": [ + 212, + 503, + 787, + 625 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nc = f ^ {\\theta_ {1} + \\Delta \\theta_ {1}} (y), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 444, + 660, + 785, + 679 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nQ = W _ {q} ^ {\\theta_ {2} + \\Delta \\theta_ {2}} (z), K = W _ {k} ^ {\\theta_ {2} + \\Delta \\theta_ {2}} (h (c)), V = W _ {v} ^ {\\theta_ {2} + \\Delta \\theta_ {2}} (h (c)), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 282, + 689, + 785, + 710 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nC r o s s A t t e n t i o n (Q, K, V) = s o f t m a x (Q \\cdot K ^ {T}) \\cdot V. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 718, + 785, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Now, we only need to train $\\Delta \\theta_{1}$ , $\\Delta \\theta_{2}$ , and $h$ on a relatively small amount of text-image pairs. After training, the language and generative vision models can effectively collaborate to generate meaningful images. We present the framework of LaVi-Bridge in Fig. 2. LaVi-Bridge is very straightforward, with both LoRA and the adapter being its crucial and indispensable components.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. Zhao et al.", + "bbox": [ + 269, + 114, + 367, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Design Details", + "text_level": 1, + "bbox": [ + 215, + 146, + 382, + 161 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LaVi-Bridge is designed to accommodate a wide range of language model structures, including encoder-only, encoder-decoder, decoder-only, as well as generative vision model structures such as U-Net and ViT. In the language model, we inject LoRA into all linear layers of the attention layers. Likewise, in a transformer-based vision model, LoRA is injected into all linear layers of the attention layers. In a U-Net-based vision model, LoRA is injected into all linear layers and convolutional layers of the ResBlocks, attention, and cross-attention layers. To address the dimension disparity between the output embedding of the language model and the dimensions handled by the cross-attention of the vision model, we employ two feedforward layers for the adapter. The input dimension of the adapter matches the output text embedding dimension of the language model, while the output dimension aligns with the dimensions received by the cross-attention of the vision model.", + "bbox": [ + 212, + 172, + 787, + 367 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For training, we first select the language and generative vision models that we choose to integrate. We keep their original weights fixed and train LoRA and the adapter on text-image pairs following the design mentioned above. The trained LoRA and adapter have fewer parameters compared to the original model weights, which makes LaVi-Bridge highly flexible. For evaluation, we used various metrics to assess text alignment and image quality across short prompts, long prompts, and compositional prompts.", + "bbox": [ + 212, + 369, + 787, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 500, + 375, + 515 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 215, + 531, + 446, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we explored the performance of different language models and generative vision models under LaVi-Bridge. We also tested the impact of LoRA and adapters. We trained on a dataset consisting of a total of 1 million text-image pairs, including around 600k text-image pairs from the COCO2017 [25] train set and 400k text-image pairs from an internal dataset with high-quality images and captions. For each setting, we set the LoRA rank to 32, image resolution to $512 \\times 512$ and the batch size to 256. We used the AdamW optimizer [26] with a learning rate of $1 \\times 10^{-4}$ and trained for a total of 50k steps. During inference, we employed the DDIM sampler [43] for sampling with the number of time steps set to 50 and the classifier free guidance scale [19] set to 7.5.", + "bbox": [ + 212, + 556, + 787, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As mentioned above, we conducted our quantitative evaluation on short prompts, long prompts, and compositional prompts. Specifically,", + "bbox": [ + 215, + 709, + 785, + 739 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. For short prompts, we evaluated using the COCO2014 [25] validation set. We randomly sampled 30k images and tested image quality and text alignment within this subset. We used FID [17] and aesthetic score [23] as evaluation metrics for image quality and CLIP score for text alignment.", + "2. For long prompts, we employed the same 30k-subset of COCO2014 and utilized Llama-2 to generate expanded captions ranging from 20 to 70 words to" + ], + "bbox": [ + 220, + 748, + 782, + 837 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "LaVi-Bridge", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/033d20f43b41fcf5baa497aa5d811d38aef26228fac6bffd8c173e86cf1fa8f4.jpg", + "image_caption": [ + "Fig. 3: Visualization results of LaVi-Bridge with different language models. The first row to the fifth row present the results with CLIP text encoder, T5-Small, T5-Base, T5-Large, and Llama-2, respectively. The prompts are displayed at the top or bottom of each column." + ], + "image_footnote": [], + "bbox": [ + 218, + 146, + 784, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "construct a dataset of 30k long prompts. Since the caption expansion process does not refer to the content of the reference image, we solely used aesthetic score to evaluate image quality and CLIP score for text alignment.", + "bbox": [ + 240, + 556, + 785, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3. For compositional prompts, we utilized the benchmark proposed by Comp-bench [21]. Compositional prompts were mainly used to test the model's understanding of textual attributes, such as generating correct object properties like color and shape, as well as accurate relationships between objects, such as spatial positioning.", + "bbox": [ + 220, + 604, + 785, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conducted a user study on different combinations of language and vision models. For each combination, we evaluated two metrics, namely image quality and text alignment. Users were asked to rank the generated images based on these evaluation criteria. The image ranked last received a score of 1, the second-to-last received a score of 2, and so on. We then calculated the percentage of scores for each model. We selected 20 prompts and included 30 users participated in the testing. In addition to quantitative evaluation and user study, we provided ample visualization results in each section to offer a more intuitive understanding of the performance of each model.", + "bbox": [ + 212, + 703, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. Zhao et al.", + "bbox": [ + 269, + 114, + 367, + 126 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/12ec0ec11c9d5b3857cf30132dc0a41ee5887a0035e18dd3a993e806f203ce15.jpg", + "table_caption": [ + "Table 1: Quantitative evaluation of LaVi-Bridge with different language models. \"Short\", \"Long\" and \"Comp\" denote short prompts, long prompts, and compositional prompts respectively. The best results are in bold." + ], + "table_footnote": [], + "table_body": "
CLIPT5-SmallT5-BaseT5-LargeLlama-2
Short - FID23.5722.9822.6223.1121.80
Short - Aesthetics5.6095.8135.8885.8815.883
Short - CLIP Score0.31020.31220.31490.31560.3172
Long - Aesthetics6.0036.2066.2846.3056.355
Long - CLIP Score0.31200.31110.31790.31930.3231
Comp - Color0.35780.33680.38560.38890.4859
Comp - Shape0.37520.29620.32660.35520.4285
Comp - Texture0.45060.37280.41320.45240.5055
Comp - Spatial0.12960.14560.15690.15820.1914
Comp - Non-Spatial0.30090.29840.30540.30680.3106
Comp - Complex0.29850.27280.30550.30720.3094
", + "bbox": [ + 225, + 198, + 782, + 395 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Evaluation on Different Language Models", + "text_level": 1, + "bbox": [ + 215, + 426, + 604, + 441 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This section evaluates the performance of LaVi-Bridge with different language models. We fixed the vision model to the U-Net of Stable Diffusion V1.4 and integrated it with different language models under LaVi-Bridge. We considered CLIP text encoder, based on the encoder-only framework, T5 series (T5-Small, T5-Base, T5-Large), based on the encoder-decoder framework, and Llama-2-7B, based on the decoder-only framework. We present the visualization results in Fig. 3, quantitative evaluation in Tab. 1, and user study in Figure 1 in the supplementary material.", + "bbox": [ + 212, + 460, + 787, + 582 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Visualization From Fig. 3, we can observe that with LaVi-Bridge, all these language models can effectively integrate with U-Net of Stable Diffusion V1.4 and generate meaningful results, such as cases of the cat and living room in Fig. 3. This demonstrates the great generalization ability of LaVi-Bridge for various language models. Additionally, we notice that the performance of different model structures varies when the provided prompts contain more complex semantics. We find that the text-to-image diffusion model corresponding to Llama-2 can perfectly describe semantic information. For example, in the third column, Llama-2's generated result effectively integrates a woman into the sea of fragmented porcelain. In the fourth column, it correctly understands and generates both the girl and the cat in a paper craft art. In the seventh column, it even portrays an entire beach scene using yarn. These examples surpass the capabilities of those models with CLIP and T5. Furthermore, we observe that T5-Large and Llama-2 accurately generate food and wine in the case of Iron Man, and in the last column, they successfully generate \"an ancient stone with eyes in dark yellow and emerald\". Models with CLIP text encoder, T5-Small, and T5-Base are not able to capture these cases accurately.", + "bbox": [ + 212, + 583, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "LaVi-Bridge", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/d46fa202ef10d74063e853d6c281bafbbafc1f1b726440f99ca8312eacdee910.jpg", + "image_caption": [ + "Fig. 4: Visualization results of LaVi-Bridge under different generative vision models. The first row to the third row present the results with U-Net in Latent Diffusion Model, U-Net in Stable Diffusion V1.4 and transformer in PixArt, respectively. The prompts are displayed at the top or bottom of each column." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 784, + 344 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Quantitative Evaluation From Tab. 1, we can observe that Llama-2 achieves the best results for all the metrics used to evaluate text alignment ability, under the setting of all the short prompts, long prompts, and compositional prompts. Besides, Llama-2 also performs the best on most of the metrics evaluating image quality. On the other hand, as the model capacity increases, in general circumstances T5-Large usually outperforms T5-Base, and T5-Base outperforms T5-Small in the area of Natural Language Processing. This conclusion also holds true for LaVi-Bridge. For all the metrics used to evaluate text alignment ability in Tab. 1, T5-Large is superior to T5-Base, and T5-Base is superior to T5-Small. This tells us that incorporating a better language model into the text-to-image diffusion model under LaVi-Bridge can lead to improved text alignment. This makes one of the motivations of LaVi-Bridge meaningful, which is that replacing the model in the existing text-to-image diffusion model with a better model can lead to performance improvements.", + "bbox": [ + 212, + 441, + 787, + 652 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "User Study We follow the settings described in Sec. 4.1, and the results are presented in Section A in the supplementary material.", + "bbox": [ + 212, + 652, + 787, + 683 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Evaluation on Different Vision Models", + "text_level": 1, + "bbox": [ + 214, + 707, + 578, + 722 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "This section evaluates the performance of LaVi-Bridge with different vision models. We fixed the language model to T5-Large and integrated it with different generative vision models under LaVi-Bridge. We considered the well-trained U-Nets in the Latent Diffusion Model and Stable Diffusion V1.4, as well as the Vision Transformer in PixArt, totally three models. We present the visualization results in Fig. 4, quantitative evaluation in Tab. 2, and user study in Figure 1 in the supplementary material.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. Zhao et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/8ed4514aea38668104dc10dda60d59bc29bc9b597bf18bafc96be8b03736c125.jpg", + "table_caption": [ + "Table 2: Quantitative evaluation of LaVi-Bridge under different generative vision models. \"Short\", \"Long\" and \"Comp\" denote short prompts, long prompts, and compositional prompts respectively. The best results are in bold." + ], + "table_footnote": [], + "table_body": "
U-Net(LDM)U-Net(SD)Transformer(PixArt)
Short - FID25.9423.1123.02
Short - Aesthetics5.7035.8816.145
Short - CLIP Score0.31260.31560.3172
Long - Aesthetics6.1226.3056.406
Long - CLIP Score0.31890.31930.3210
Comp - Color0.40990.38890.3689
Comp - Shape0.37240.35520.3316
Comp - Texture0.50460.45240.4553
Comp - Spatial0.15500.15820.1725
Comp - Non-Spatial0.30040.30680.3098
Comp - Complex0.30600.30720.3014
", + "bbox": [ + 228, + 198, + 779, + 393 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Visualization From Fig. 4, we can see that all these three vision models integrate well with T5-Large and generate relatively accurate images based on the given text prompts. From these cases, we can observe that the images generated by the transformer model based on PixArt exhibit richer details compared to the images generated by the other two models based on U-Net. For example, the forest in the first column, the hull of the pirate ship in the third column, and the bushes at the foot of the mountain in the sixth column are very intricate and realistic. Additionally, we can observe from these cases that the images generated by the model with U-Net of Stable Diffusion V1.4 have more detailed features compared to the images generated by the model with U-Net of Latent Diffusion Model. Furthermore, we can also find that, for the PixArt-based model, text alignment is better in some cases. For instance, in the image of the fifth column, only the model that is based on the transformer of PixArt generates the \"aged car\" mentioned in the prompt. Similarly, in the seventh column, the garden warrior holding a sword and shield is highly consistent with the prompt description.", + "bbox": [ + 217, + 439, + 785, + 680 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative Evaluation From Tab. 2, it can be observed that for all the metrics measuring image quality, LaVi-Bridge with the PixArt vision model achieves the best results. Additionally, PixArt also achieves the best text alignment for both short and long prompts. This reflects the use of PixArt's transformer as a vision model can also improve the model's understanding of semantics to some extent. Additionally, it is noteworthy that the U-Net in Stable Diffusion, an enhanced version of the U-Net utilized in the Latent Diffusion Model, still outperforms Latent Diffusion Model's U-Net under LaVi-Bridge on all the metrics measuring image quality. This aligns with our previous discussion in Sec. 4.2 and further validates the underlying motivation behind our proposed LaVi-Bridge.", + "bbox": [ + 217, + 688, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "LaVi-Bridge", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "User Study We follow the settings described in Sec. 4.1, and the results are presented in Section A in the supplementary material.", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 196, + 387, + 213 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we investigate two sets of ablation experiments. The first set aims to explore the impact of training LaVi-Bridge on the original pre-trained text-to-image diffusion model. The second set of experiments is to study the effects of LoRA and adapters in LaVi-Bridge. For both sets of experiments, we present visualization results in Fig. 5 and provide quantitative evaluations in Tab. 3.", + "bbox": [ + 215, + 220, + 785, + 295 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Training with LaVi-Bridge We investigate the impact of our LaVi-Bridge training framework on the original pre-trained text-to-image diffusion model. Specifically, we consider Stable Diffusion V1.4 which adopts CLIP text encoder as its language model and U-Net as its vision model. We incorporate LoRA and an adapter and apply LaVi-Bridge to the same language and vision models with identical structures and weights to those in Stable Diffusion V1.4. We then compare the performance of the model under LaVi-Bridge with the original Stable Diffusion V1.4.", + "bbox": [ + 215, + 296, + 785, + 416 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The visualization results are shown in the first two rows of Fig. 5. For these two models, there is no significant difference in image quality and text alignment, varying on a case-by-case basis. In some cases, Stable Diffusion performs better, such as in the third column, where Stable Diffusion successfully generates the case of a \"Fox bracelet made of buckskin with fox features\", while the model trained under LaVi-Bridge only generates the fox and fails to understand the bracelet made of buckskin. Similarly, in the case of Marvel's Hulk playing basketball, Stable Diffusion generates a slam dunk action following the prompt, whereas the model trained under LaVi-Bridge does not. However, in the second column, the model trained under LaVi-Bridge correctly understands the quantity and successfully generates two elephants, while Stable Diffusion only generates one. Moreover, in the last column, the model trained under LaVi-Bridge accurately describes a frog in a spacesuit, while Stable Diffusion fails.", + "bbox": [ + 215, + 417, + 785, + 612 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The left two columns of Tab. 3 present the quantitative evaluation results. It can be observed that Stable Diffusion achieves the best image quality and text alignment for both short prompts and long prompts. However, for compositional prompts, the model trained under LaVi-Bridge outperforms Stable Diffusion in four out of six settings.", + "bbox": [ + 215, + 613, + 785, + 688 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Based on the visualization results and quantitative evaluations, we can conclude that overall there is no significant improvement or decline in text alignment. Regarding image quality, it should be noted that training with LaVi-Bridge may result in a decrease compared to the original text-to-image diffusion model, if the same models and weights are used. However, it is important to understand that the main purpose of LaVi-Bridge is to establish connections between different language and vision models, enabling the utilization of more advanced models for performance enhancement. It is not intended to be directly applied to the original text-to-image diffusion models using the same models and weights.", + "bbox": [ + 215, + 689, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. Zhao et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/040ac1ab6d6724c09ae0a959bc2f24eec8d97e401580e199eb4c4f4fd2af4e98.jpg", + "image_caption": [ + "Fig. 5: Visualization results of the ablation study. The top two rows show the impact of LaVi-Bridge on the original pre-trained text-to-image diffusion models. The bottom three rows illustrate the influence of the adapter and LoRA. The prompts are displayed at the top or bottom of each column." + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 784, + 452 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "LoRA and Adapter Here, we investigate the role of LoRA and adapters in LaVi-Bridge. We use T5-Large as the language model and Stable Diffusion V1.4's U-Net as the vision model. For the LoRA experiments, we kept the language and vision models fixed without introducing LoRA, and only trained the adapter. For the adapter experiments, considering the mismatch in the dimensions of text embeddings from the language model and the input embeddings acceptable by the vision model, we aligned the dimensions between the language and vision models using a single linear layer instead of stacked feedforward layers which include non-linear activation layers. Under this setting, we trained both LoRA and this linear layer.", + "bbox": [ + 212, + 551, + 787, + 702 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The visualization results are shown in the bottom three rows of Fig. 5. We can observe that both image quality and text alignment are significantly affected when LoRA and adapters are not used. For example, in the case of \"Bull Fit Athlete\" in the seventh column, without LoRA or the adapter, the model cannot understand and integrate these two less related elements, and the image quality is much lower compared to results generated by the original setting. We also found that the results without LoRA are worse than those without the adapter. For instance in the fourth column, there is not even a ferret present in the image in the absence of LoRA.", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "LaVi-Bridge", + "bbox": [ + 647, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/0768d007a995b8e6c459bde1d8510ae578df1c6843f611a465c345de5076ac2e.jpg", + "table_caption": [ + "Table 3: Quantitative evaluation of the ablation study. The left two columns present the impact of LaVi-Bridge on the original pre-trained text-to-image diffusion models. The right three columns demonstrate the influence of the adapter and LoRA. \"Short\", \"Long\" and \"Comp\" denote short prompts, long prompts, and compositional prompts respectively. The best results are in bold." + ], + "table_footnote": [], + "table_body": "
SDCLIP+U-Netw/o Adapterw/o LoRAT5+U-Net
Short - FID20.3223.5723.8122.3523.11
Short - Aesthetics5.8995.6095.8075.8295.881
Short - CLIP Score0.31320.31020.31470.31070.3156
Long - Aesthetics6.1206.0036.1316.2736.305
Long - CLIP Score0.31710.31200.31060.30970.3193
Comp - Color0.35700.35780.35500.24850.3889
Comp - Shape0.35630.37520.30440.29440.3552
Comp - Texture0.40280.45060.40010.31900.4524
Comp - Spatial0.12250.12960.16510.09560.1582
Comp - Non-Spatial0.31040.30090.30650.29980.3068
Comp - Complex0.30420.29850.28780.26870.3072
", + "bbox": [ + 230, + 226, + 777, + 421 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The right three columns of Tab. 3 present the quantitative evaluation results. We find that our default setting, which utilizes both LoRA and the adapter, achieves the best performance in most cases. Additionally, overall, the absence of LoRA has a significant impact on text alignment, with many text alignment evaluation metrics being much lower compared to the absence of the adapter.", + "bbox": [ + 212, + 453, + 787, + 530 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 559, + 359, + 575 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose LaVi-Bridge, which works on text-to-image diffusion models. LaVi-Bridge is capable of connecting various language models and generative vision models for text-to-image generation. It is highly versatile and can adapt to different structures. LaVi-Bridge is also flexible, as it achieves integration without modifying the original weights of language and vision models. Instead, it utilizes LoRA and an adapter for fine-tuning. Additionally, under LaVi-Bridge, using superior language or vision models can enhance the text comprehension capability or image quality. These advantages enable LaVi-Bridge to help text-to-image diffusion models leverage the latest advancements in the areas of Natural Language Processing and Computer Vision, to enhance text-to-image generation. We believe that this task holds significant research value and requires further exploration. LaVi-Bridge allows designers, artists, and others to flexibly utilize existing language and vision models to achieve their creative goals. It is of utmost importance to avoid misuse and mitigate potential negative social impacts. In practical deployment, it is crucial to standardize its usage, improve model transparency.", + "bbox": [ + 212, + 597, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. Zhao et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bao, F., Li, C., Cao, Y., Zhu, J.: All are worth words: a vit backbone for score-based diffusion models. CVPR (2023)", + "2. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., Manassra, W., Dhariwal, P., Chu, C., Jiao, Y., Ramesh, A.: Improving image generation with better captions. https://cdn.openai.com/papers/dall-e-3.pdf (2023)", + "3. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. CVPR (2023)", + "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. NeurIPS 2020", + "5. Chen, J., Huang, Y., Lv, T., Cui, L., Chen, Q., Wei, F.: Textdiffuser: Diffusion models as text painters. NeurIPS (2024)", + "6. Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wu, Y., Wang, Z., Kwok, J., Luo, P., Lu, H., et al.: Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. ICLR (2024)", + "7. Croitoru, F.A., Hondru, V., Ionescu, R.T., Shah, M.: Diffusion models in vision: A survey. TPAMI (2023)", + "8. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. NAACL-HLT (2018)", + "9. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. NeurIPS (2021)", + "0. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.: Cogview: Mastering text-to-image generation via transformers. NeurIPS (2021)", + "1. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021)", + "2. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618 (2022)", + "3. Ge, S., Park, T., Zhu, J.Y., Huang, J.B.: Expressive text-to-image generation with rich text. ICCV (2023)", + "4. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. NeurIPS (2014)", + "5. Guo, Y., Yang, C., Rao, A., Liang, Z., Wang, Y., Qiao, Y., Agrawala, M., Lin, D., Dai, B.: Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. ICLR (2024)", + "6. Hao, S., Han, K., Zhao, S., Wong, K.Y.K.: Vico: Detail-preserving visual condition for personalized text-to-image generation. arXiv preprint arXiv:2306.00971 (2023)", + "7. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS (2017)", + "8. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS (2020)", + "9. Ho, J., Salimans, T.: Classifier-free diffusion guidance. NeurIPS Workshop on Deep Generative Models and Downstream Applications (2021)" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "LaVi-Bridge", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "20. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. ICLR (2022)", + "21. Huang, K., Sun, K., Xie, E., Li, Z., Liu, X.: T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. arXiv preprint arXiv: 2307.06350 (2023)", + "22. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. CVPR (2023)", + "23. LAION-AI: aesthetic-predictor. https://github.com/LAION-AI/aesthetic-predictor (2022)", + "24. Lewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mohamed, A., Levy, O., Stoyanov, V., Zettlemoyer, L.: Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461 (2019)", + "25. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. ECCV (2014)", + "26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. ICLR (2019)", + "27. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. ICLR (2022)", + "28. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023)", + "29. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. ICML (2021)", + "30. Peebles, W., Xie, S.: Scalable diffusion models with transformers. ICCV (2023)", + "31. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. ICML (2021)", + "32. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training. OpenAI blog (2018)", + "33. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog (2019)", + "34. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR (2020)", + "35. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)", + "36. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. ICML (2021)", + "37. Reed, S., Akata, Z., Yan, X., Logeswaran, L., Schiele, B., Lee, H.: Generative adversarial text to image synthesis. ICML (2016)", + "38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. CVPR (2022)", + "39. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. MICCAI (2015)", + "40. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. CVPR (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. Zhao et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "41. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022)", + "42. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)", + "43. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. ICLR (2021)", + "44. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. ICML (2023)", + "45. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. ICLR (2021)", + "46. Sutskever, I., Vinyals, O., Le, Q.V.: Sequence to sequence learning with neural networks. NeurIPS (2014)", + "47. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)", + "48. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS (2017)", + "49. Wu, W., Li, Z., He, Y., Shou, M.Z., Shen, C., Cheng, L., Li, Y., Gao, T., Zhang, D., Wang, Z.: Paragraph-to-image generation with information-enriched diffusion model. arXiv preprint arXiv:2311.14284 (2023)", + "50. Xue, Z., Song, G., Guo, Q., Liu, B., Zong, Z., Liu, Y., Luo, P.: Raphael: Text-to-image generation via large mixture of diffusion paths. NeurIPS (2024)", + "51. Zhang, H., Xu, T., Li, H., Zhang, S., Wang, X., Huang, X., Metaxas, D.N.: Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. ICCV (2017)", + "52. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. ICCV (2023)", + "53. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Unicontrolnet: All-in-one control to text-to-image diffusion models. NeurIPS (2024)" + ], + "bbox": [ + 212, + 146, + 787, + 549 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "LaVi-Bridge", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Bridging Different Language Models and Generative Vision Models for Text-to-Image Generation/1764d397-410b-451c-a099-8966f6b5772c_model.json b/2024/Bridging Different Language Models and Generative Vision Models for Text-to-Image Generation/1764d397-410b-451c-a099-8966f6b5772c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1acbf5cd7a1b880aad9a1d0f9f86ab3a9792767a --- /dev/null +++ b/2024/Bridging Different Language Models and Generative Vision Models for Text-to-Image Generation/1764d397-410b-451c-a099-8966f6b5772c_model.json @@ -0,0 +1,2082 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.27, + 0.141, + 0.737, + 0.21 + ], + "angle": 0, + "content": "Bridging Different Language Models and Generative Vision Models for Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.279, + 0.234, + 0.726, + 0.266 + ], + "angle": 0, + "content": "Shihao Zhao\\(^{1}\\), Shaozhe Hao\\(^{1}\\), Bojia Zi\\(^{2}\\), Huaizhe Xu\\(^{3}\\), and Kwan-Yee K. Wong\\(^{1\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.394, + 0.277, + 0.608, + 0.292 + ], + "angle": 0, + "content": "1 The University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.378, + 0.293, + 0.627, + 0.306 + ], + "angle": 0, + "content": "{shzhao, szhao, kykwong}@cs.hku.hk" + }, + { + "type": "text", + "bbox": [ + 0.369, + 0.306, + 0.636, + 0.32 + ], + "angle": 0, + "content": "2 The Chinese University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.428, + 0.321, + 0.576, + 0.334 + ], + "angle": 0, + "content": "bjzi@se.cuhk.edu.hk" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.334, + 0.688, + 0.348 + ], + "angle": 0, + "content": "3 The Hong Kong University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.423, + 0.35, + 0.581, + 0.36 + ], + "angle": 0, + "content": "hxubr@connect.ust.hk" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.397, + 0.744, + 0.705 + ], + "angle": 0, + "content": "Abstract. Text-to-image generation has made significant advancements with the introduction of text-to-image diffusion models. These models typically consist of a language model that interprets user prompts and a vision model that generates corresponding images. As language and vision models continue to progress in their respective domains, there is a great potential in exploring the replacement of components in text-to-image diffusion models with more advanced counterparts. A broader research objective would therefore be to investigate the integration of any two unrelated language and generative vision models for text-to-image generation. In this paper, we explore this objective and propose LaVi-Bridge, a pipeline that enables the integration of diverse pre-trained language models and generative vision models for text-to-image generation. By leveraging LoRA and adapters, LaVi-Bridge offers a flexible and plug-and-play approach without requiring modifications to the original weights of the language and vision models. Our pipeline is compatible with various language models and generative vision models, accommodating different structures. Within this framework, we demonstrate that incorporating superior modules, such as more advanced language models or generative vision models, results in notable improvements in capabilities like text alignment or image quality. Extensive evaluations have been conducted to verify the effectiveness of LaVi-Bridge. Code is available at https://github.com/ShihaoZhaoZSH/LaVi-Bridge." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.715, + 0.64, + 0.73 + ], + "angle": 0, + "content": "Keywords: Diffusion model \\(\\cdot\\) Text-to-image generation" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.755, + 0.377, + 0.772 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.786, + 0.788, + 0.817 + ], + "angle": 0, + "content": "In recent years, there have been remarkable advancements in the field of text-to-image generation, specifically through the use of diffusion models [9, 18, 43, 45]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.39, + 0.84 + ], + "angle": 0, + "content": "* Corresponding authors" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "S. Zhao et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.321, + 0.788, + 0.391 + ], + "angle": 0, + "content": "Fig. 1: Overview of LaVi-Bridge. LaVi-Bridge is capable of integrating various language models and generative vision models. On the left side, we keep the vision model fixed and experiment with different language models in our pipeline. On the right side, we keep the language model fixed and try out different vision models. We display the visualization results alongside each combination." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.788, + 0.517 + ], + "angle": 0, + "content": "These models have made significant contributions and have gained considerable attention for their exceptional performance. By leveraging large-scale training datasets alongside large deep models, text-to-image diffusion models are capable of producing high-quality images that faithfully align with the textual descriptions provided by users. This has rendered them highly applicable in real-world scenarios such as content creation and architectural design." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.788, + 0.717 + ], + "angle": 0, + "content": "Text-to-image diffusion models [2, 6, 29, 35, 38, 41, 50] typically consist of two key components, namely a language model and a generative vision model. The language model is responsible for comprehending the input prompts, whereas the vision model is tasked with generating images that align with the extracted context. Existing text-to-image diffusion models employ various language models and generative vision models and have gained widespread usage. For instance, Stable Diffusion (SD) [38] is a highly popular text-to-image diffusion model that employs the CLIP text encoder [31] as its language model and a U-Net [39] as its generative vision model. Another example is PixArt [6], a recently proposed text-to-image diffusion model that adopts the T5 [34] as its language model and a Vision Transformer (ViT) [11] as its generative vision model. These models are trained on a vast amount of text-image pairs, enabling seamless collaboration between their language modules and vision modules." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "The advancements in deep language models and deep vision models have witnessed rapid progress in recent years, with both fields experiencing continuous developments and the introduction of more powerful models. However, this rapid development poses a challenge for the research in text-to-image generation when it comes to integrating more advanced language or vision models into existing text-to-image diffusion models. The problem of how to integrate any two unrelated language and vision models is unexplored, and the impact of newly developed models on text-to-image generation capabilities also remains" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LaVi-Bridge" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "uncertain. The current situation highlights the presence of a gap between the language or vision modules within text-to-image diffusion models and the state-of-the-art models in their respective domains. Therefore, it has become crucial to address this gap and explore ways to incorporate more advanced language or vision models into existing text-to-image diffusion models. Furthermore, the broader challenge of integrating any pre-trained language model with any generative vision model deserves a thorough investigation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.419 + ], + "angle": 0, + "content": "In this paper, our objective is to delve into the aforementioned problem. We propose LaVi-Bridge, a flexible framework that facilitates the integration of diverse well-trained language models and generative vision models to achieve text-to-image generation. Our framework enables the integration of two unrelated language and vision models that have not been previously trained together, as shown in Fig. 1. Importantly, LaVi-Bridge does not require modifying the original weights of the language and vision models. Instead, it injects LoRA [20] into the language and vision models separately and utilizes an adapter to bridge these two modules. Moreover, LaVi-Bridge only necessitates a relatively small dataset to integrate different language models and generative vision models for text-to-image generation." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.42, + 0.742, + 0.435 + ], + "angle": 0, + "content": "We summarize the advantages and features of LaVi-Bridge as follows:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.448, + 0.787, + 0.537 + ], + "angle": 0, + "content": "1. LaVi-Bridge is designed for text-to-image diffusion models and serves as a bridge, capable of connecting various pre-trained language models and generative vision models. Our framework can accommodate different model structures, including encoder-only, encoder-decoder, and decoder-only language models, as well as U-Net-based and Transformer-based generative vision models." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.539, + 0.787, + 0.584 + ], + "angle": 0, + "content": "2. LaVi-Bridge utilizes LoRA and adapters, eliminating the need to modify the original weights of the models. It is more flexible and requires relatively small computing resources compared to training the entire diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.585, + 0.787, + 0.721 + ], + "angle": 0, + "content": "3. We evaluated various text-image alignment and image quality metrics on short prompts, long prompts, and compositional prompts. We also conducted extensive visualization. We then drew several conclusions. For instance, integrating superior models leads to improved performance in the corresponding modality, such as enhanced semantic understanding with advanced language models or improved image quality with more powerful generative vision models. Additionally, the diffusion model utilizing Llama-2 demonstrates exceptional semantic understanding, while the diffusion model utilizing the transformer in PixArt yields images with enhanced aesthetics." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.448, + 0.787, + 0.721 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.747, + 0.388, + 0.763 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.782, + 0.664, + 0.797 + ], + "angle": 0, + "content": "2.1 Language Models and Generative Vision Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "The mainstream Large Language Models (LLMs) [8, 32-34] are built based on the transformer structure [48], with three main types of architectures, namely" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.127 + ], + "angle": 0, + "content": "S. Zhao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.357 + ], + "angle": 0, + "content": "encoder-only, encoder-decoder, and decoder-only. All these three belong to Sequence to Sequence (Seq2Seq) [46]. The encoder-only architecture is exemplified by BERT [8]. CLIP text encoder [31] is based on BERT and further trained to align with the image domain. Models of this type excel at understanding the content of the input and generating outputs tailored to specific tasks. On the other hand, the encoder-decoder framework is adept at handling tasks that involve complex mappings between input and output sequences. Examples include T5 [34] and BART [24]. Recently, due to the tremendous success of ChatGPT, attention has been drawn to models that consist solely of a decoder, like GPT-3 [4] and Llama-2 [47]. The decoder-only architecture demonstrates exceptional performance in semantic understanding. For text-to-image generation, all three types of LLMs can provide effective semantic information to serve as conditions for image generation in diffusion models. In this paper, we explore and compare all these three types of language models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.601 + ], + "angle": 0, + "content": "A generative vision model refers to a vision model with the ability to generate images or visual contents. There are two common types of structures, namely U-Net-based [39] and Transformer-based [11]. Generative Adversarial Networks (GANs) [14, 37, 51] employ a framework consisting of a discriminator and a generator, with the generator's structure based on U-Net. On the other hand, motivated by the success of GPT models, recent works have attempted to use the Transformer architecture for image generation in an autoregressive manner, with notable examples being DALLE [36] and CogView [10]. Another popular class of generative models is diffusion models [7, 18, 43, 45], which are based on the diffusion process and gradually denoise to produce natural images. Early diffusion models often employed U-Net as their generative vision model, such as Stable Diffusion, which scaled up the Latent Diffusion Model (LDM) [38] with larger data scales. Some recent works have started to replace the U-Net in diffusion models with Vision Transformer and have made significant progress, such as DiT [30], U-ViT [1] and PixArt [6]. In this paper, we focus on diffusion models and explore both U-Net-based and Transformer-based vision models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.621, + 0.525, + 0.637 + ], + "angle": 0, + "content": "2.2 Text-to-Image Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Text-to-image diffusion models [2,9,29,38,41,44] are capable of generating images based on user prompts. These models consist of two main components, namely a language model and a vision model. The language module is responsible for understanding the text input provided by the user, extracting contextual information, and injecting it into the vision module to generate the desired image. Text-to-image diffusion models have paved the way for various exciting research areas, including image editing [3,22,27], controllable image generation [28,52,53], personalized object generation [12, 16, 40], as well as other interesting applications [5,13,15]. In the extensive exploration of diffusion models, researchers have utilized different language models and vision models. For instance, Stable Diffusion [38] employs CLIP text encoder [31] as its language model and U-Net as its vision model.Imagen [41] utilizes T5 [34] as its language model, which claims to enhance both sample fidelity and image-text alignment. ParaDiffusion [49]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LaVi-Bridge" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "focuses on paragraph-to-image generation and leverages the powerful semantic understanding capability of Llama-2 [47] to comprehend lengthy sentences. PixArt [6], on the other hand, utilizes a ViT [11] as its vision model and achieves high image fidelity while being trained at a lower cost." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.788, + 0.374 + ], + "angle": 0, + "content": "After training on a large dataset of text-image pairs [42], the language and vision models in the text-to-image diffusion model become closely intertwined. This tight coupling ensures a strong alignment between the provided text description and the generated image, but at the same time also limits the flexibility of the diffusion model. For instance, if a more advanced language or vision model becomes available, it may have the potential to enhance the text-to-image task. However, decoupling the language and vision modules in existing text-to-image diffusion models and replacing a module with a new one is nontrivial. Therefore, this paper explores the dilemma faced by text-to-image generation and proposes a framework that enables efficient integration of various language models and generative vision models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.394, + 0.331, + 0.41 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.421, + 0.362, + 0.436 + ], + "angle": 0, + "content": "3.1 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.669 + ], + "angle": 0, + "content": "A diffusion model is based on the diffusion process for image generation. This process consists of two stages, namely the forward process and the reverse process. During the forward process, Gaussian noise is progressively added to a natural image until the image becomes completely noisy. After that, during the reverse process, the noise is gradually eliminated over a series of time steps, resulting in a natural image. In the reverse process, a trainable vision model is used to predict and remove the noise. By employing this denoising model, we are able to obtain a natural image from Gaussian noise by denoising. Within a text-to-image diffusion model, there are two components at each time step, namely a language model \\( f \\) and a vision model \\( g \\). The language model converts user input text \\( y \\) into embeddings, which capture the semantic meaning of the text. On the other hand, the vision model, which is the denoising model aforementioned, encodes image features \\( z \\), extracting relevant visual information from the input images. The interaction between text embeddings and image features is achieved through cross-attention layers, which can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.441, + 0.678, + 0.785, + 0.693 + ], + "angle": 0, + "content": "\\[\nc = f (y), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.704, + 0.785, + 0.72 + ], + "angle": 0, + "content": "\\[\nQ = W _ {q} (z), K = W _ {k} (c), V = W _ {v} (c), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.323, + 0.731, + 0.785, + 0.748 + ], + "angle": 0, + "content": "\\[\nC r o s s A t t e n t i o n (Q, K, V) = \\text {s o f t m a x} (Q \\cdot K ^ {T}) \\cdot V, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.755, + 0.558, + 0.772 + ], + "angle": 0, + "content": "where \\(W_{q}, W_{k}\\) and \\(W_{v}\\) are projection matrices." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.789, + 0.528, + 0.804 + ], + "angle": 0, + "content": "3.2 Language and Vision Alignment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "LaVi-Bridge enables the integration of any two pre-trained language and generative vision models, even though these models are not related and have been" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "S. Zhao et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.146, + 0.785, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.312, + 0.788, + 0.382 + ], + "angle": 0, + "content": "Fig. 2: Pipeline of LaVi-Bridge. We select one model each from the language and vision model pools. We then freeze the pre-trained language and vision models and incorporate LoRA into both models. The connection between the language and vision models is established through an adapter. The only weights we need to train are the ones introduced by LoRA and the adapter." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.423, + 0.784, + 0.5 + ], + "angle": 0, + "content": "trained separately. Here, we denote the language model as \\( f \\) and the vision model as \\( g \\), as mentioned previously. If we directly interact the textual information and image information using Eq. (1), considering that \\( f \\) and \\( g \\) are trained independently, the parameters in the cross-attention layers of \\( g \\) cannot comprehend the text embedding output by \\( f \\), resulting in meaningless model outputs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.788, + 0.626 + ], + "angle": 0, + "content": "To establish a connection between them, LaVi-Bridge keeps the pre-trained language and vision models fixed and utilizes LoRA to introduce trainable parameters \\(\\Delta \\theta\\) into both the language model and the vision model. In this context, we denote the language and vision models with LoRA as \\(f^{\\theta_1 + \\Delta \\theta_1}\\) and \\(g^{\\theta_2 + \\Delta \\theta_2}\\), where \\(\\theta_{1}\\) and \\(\\theta_{2}\\) are the original parameters of \\(f\\) and \\(g\\), respectively. Furthermore, we introduce an adapter as a bridge between the language model and vision model to facilitate better alignment. The adapter consists of stacked feedforward layers, denoted as \\(h\\). Consequently, the cross-attention layer can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.446, + 0.661, + 0.786, + 0.68 + ], + "angle": 0, + "content": "\\[\nc = f ^ {\\theta_ {1} + \\Delta \\theta_ {1}} (y), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.284, + 0.69, + 0.787, + 0.711 + ], + "angle": 0, + "content": "\\[\nQ = W _ {q} ^ {\\theta_ {2} + \\Delta \\theta_ {2}} (z), K = W _ {k} ^ {\\theta_ {2} + \\Delta \\theta_ {2}} (h (c)), V = W _ {v} ^ {\\theta_ {2} + \\Delta \\theta_ {2}} (h (c)), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.312, + 0.719, + 0.787, + 0.739 + ], + "angle": 0, + "content": "\\[\nC r o s s A t t e n t i o n (Q, K, V) = s o f t m a x (Q \\cdot K ^ {T}) \\cdot V. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Now, we only need to train \\(\\Delta \\theta_{1}\\), \\(\\Delta \\theta_{2}\\), and \\(h\\) on a relatively small amount of text-image pairs. After training, the language and generative vision models can effectively collaborate to generate meaningful images. We present the framework of LaVi-Bridge in Fig. 2. LaVi-Bridge is very straightforward, with both LoRA and the adapter being its crucial and indispensable components." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LaVi-Bridge" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.383, + 0.162 + ], + "angle": 0, + "content": "3.3 Design Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.368 + ], + "angle": 0, + "content": "LaVi-Bridge is designed to accommodate a wide range of language model structures, including encoder-only, encoder-decoder, decoder-only, as well as generative vision model structures such as U-Net and ViT. In the language model, we inject LoRA into all linear layers of the attention layers. Likewise, in a transformer-based vision model, LoRA is injected into all linear layers of the attention layers. In a U-Net-based vision model, LoRA is injected into all linear layers and convolutional layers of the ResBlocks, attention, and cross-attention layers. To address the dimension disparity between the output embedding of the language model and the dimensions handled by the cross-attention of the vision model, we employ two feedforward layers for the adapter. The input dimension of the adapter matches the output text embedding dimension of the language model, while the output dimension aligns with the dimensions received by the cross-attention of the vision model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.476 + ], + "angle": 0, + "content": "For training, we first select the language and generative vision models that we choose to integrate. We keep their original weights fixed and train LoRA and the adapter on text-image pairs following the design mentioned above. The trained LoRA and adapter have fewer parameters compared to the original model weights, which makes LaVi-Bridge highly flexible. For evaluation, we used various metrics to assess text alignment and image quality across short prompts, long prompts, and compositional prompts." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.5, + 0.377, + 0.516 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.532, + 0.447, + 0.548 + ], + "angle": 0, + "content": "4.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.558, + 0.788, + 0.709 + ], + "angle": 0, + "content": "In this section, we explored the performance of different language models and generative vision models under LaVi-Bridge. We also tested the impact of LoRA and adapters. We trained on a dataset consisting of a total of 1 million text-image pairs, including around 600k text-image pairs from the COCO2017 [25] train set and 400k text-image pairs from an internal dataset with high-quality images and captions. For each setting, we set the LoRA rank to 32, image resolution to \\(512 \\times 512\\) and the batch size to 256. We used the AdamW optimizer [26] with a learning rate of \\(1 \\times 10^{-4}\\) and trained for a total of 50k steps. During inference, we employed the DDIM sampler [43] for sampling with the number of time steps set to 50 and the classifier free guidance scale [19] set to 7.5." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.71, + 0.787, + 0.74 + ], + "angle": 0, + "content": "As mentioned above, we conducted our quantitative evaluation on short prompts, long prompts, and compositional prompts. Specifically," + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.749, + 0.784, + 0.807 + ], + "angle": 0, + "content": "1. For short prompts, we evaluated using the COCO2014 [25] validation set. We randomly sampled 30k images and tested image quality and text alignment within this subset. We used FID [17] and aesthetic score [23] as evaluation metrics for image quality and CLIP score for text alignment." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.81, + 0.784, + 0.838 + ], + "angle": 0, + "content": "2. For long prompts, we employed the same 30k-subset of COCO2014 and utilized Llama-2 to generate expanded captions ranging from 20 to 70 words to" + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.749, + 0.784, + 0.838 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "S. Zhao et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.147, + 0.785, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.517 + ], + "angle": 0, + "content": "Fig. 3: Visualization results of LaVi-Bridge with different language models. The first row to the fifth row present the results with CLIP text encoder, T5-Small, T5-Base, T5-Large, and Llama-2, respectively. The prompts are displayed at the top or bottom of each column." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.557, + 0.787, + 0.603 + ], + "angle": 0, + "content": "construct a dataset of 30k long prompts. Since the caption expansion process does not refer to the content of the reference image, we solely used aesthetic score to evaluate image quality and CLIP score for text alignment." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.606, + 0.787, + 0.682 + ], + "angle": 0, + "content": "3. For compositional prompts, we utilized the benchmark proposed by Comp-bench [21]. Compositional prompts were mainly used to test the model's understanding of textual attributes, such as generating correct object properties like color and shape, as well as accurate relationships between objects, such as spatial positioning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.84 + ], + "angle": 0, + "content": "We conducted a user study on different combinations of language and vision models. For each combination, we evaluated two metrics, namely image quality and text alignment. Users were asked to rank the generated images based on these evaluation criteria. The image ranked last received a score of 1, the second-to-last received a score of 2, and so on. We then calculated the percentage of scores for each model. We selected 20 prompts and included 30 users participated in the testing. In addition to quantitative evaluation and user study, we provided ample visualization results in each section to offer a more intuitive understanding of the performance of each model." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LaVi-Bridge" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Quantitative evaluation of LaVi-Bridge with different language models. \"Short\", \"Long\" and \"Comp\" denote short prompts, long prompts, and compositional prompts respectively. The best results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.199, + 0.784, + 0.396 + ], + "angle": 0, + "content": "
CLIPT5-SmallT5-BaseT5-LargeLlama-2
Short - FID23.5722.9822.6223.1121.80
Short - Aesthetics5.6095.8135.8885.8815.883
Short - CLIP Score0.31020.31220.31490.31560.3172
Long - Aesthetics6.0036.2066.2846.3056.355
Long - CLIP Score0.31200.31110.31790.31930.3231
Comp - Color0.35780.33680.38560.38890.4859
Comp - Shape0.37520.29620.32660.35520.4285
Comp - Texture0.45060.37280.41320.45240.5055
Comp - Spatial0.12960.14560.15690.15820.1914
Comp - Non-Spatial0.30090.29840.30540.30680.3106
Comp - Complex0.29850.27280.30550.30720.3094
" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.427, + 0.606, + 0.442 + ], + "angle": 0, + "content": "4.2 Evaluation on Different Language Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.788, + 0.583 + ], + "angle": 0, + "content": "This section evaluates the performance of LaVi-Bridge with different language models. We fixed the vision model to the U-Net of Stable Diffusion V1.4 and integrated it with different language models under LaVi-Bridge. We considered CLIP text encoder, based on the encoder-only framework, T5 series (T5-Small, T5-Base, T5-Large), based on the encoder-decoder framework, and Llama-2-7B, based on the decoder-only framework. We present the visualization results in Fig. 3, quantitative evaluation in Tab. 1, and user study in Figure 1 in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Visualization From Fig. 3, we can observe that with LaVi-Bridge, all these language models can effectively integrate with U-Net of Stable Diffusion V1.4 and generate meaningful results, such as cases of the cat and living room in Fig. 3. This demonstrates the great generalization ability of LaVi-Bridge for various language models. Additionally, we notice that the performance of different model structures varies when the provided prompts contain more complex semantics. We find that the text-to-image diffusion model corresponding to Llama-2 can perfectly describe semantic information. For example, in the third column, Llama-2's generated result effectively integrates a woman into the sea of fragmented porcelain. In the fourth column, it correctly understands and generates both the girl and the cat in a paper craft art. In the seventh column, it even portrays an entire beach scene using yarn. These examples surpass the capabilities of those models with CLIP and T5. Furthermore, we observe that T5-Large and Llama-2 accurately generate food and wine in the case of Iron Man, and in the last column, they successfully generate \"an ancient stone with eyes in dark yellow and emerald\". Models with CLIP text encoder, T5-Small, and T5-Base are not able to capture these cases accurately." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "S. Zhao et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.785, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.355, + 0.788, + 0.412 + ], + "angle": 0, + "content": "Fig. 4: Visualization results of LaVi-Bridge under different generative vision models. The first row to the third row present the results with U-Net in Latent Diffusion Model, U-Net in Stable Diffusion V1.4 and transformer in PixArt, respectively. The prompts are displayed at the top or bottom of each column." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.788, + 0.653 + ], + "angle": 0, + "content": "Quantitative Evaluation From Tab. 1, we can observe that Llama-2 achieves the best results for all the metrics used to evaluate text alignment ability, under the setting of all the short prompts, long prompts, and compositional prompts. Besides, Llama-2 also performs the best on most of the metrics evaluating image quality. On the other hand, as the model capacity increases, in general circumstances T5-Large usually outperforms T5-Base, and T5-Base outperforms T5-Small in the area of Natural Language Processing. This conclusion also holds true for LaVi-Bridge. For all the metrics used to evaluate text alignment ability in Tab. 1, T5-Large is superior to T5-Base, and T5-Base is superior to T5-Small. This tells us that incorporating a better language model into the text-to-image diffusion model under LaVi-Bridge can lead to improved text alignment. This makes one of the motivations of LaVi-Bridge meaningful, which is that replacing the model in the existing text-to-image diffusion model with a better model can lead to performance improvements." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.654, + 0.788, + 0.684 + ], + "angle": 0, + "content": "User Study We follow the settings described in Sec. 4.1, and the results are presented in Section A in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.708, + 0.58, + 0.723 + ], + "angle": 0, + "content": "4.3 Evaluation on Different Vision Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This section evaluates the performance of LaVi-Bridge with different vision models. We fixed the language model to T5-Large and integrated it with different generative vision models under LaVi-Bridge. We considered the well-trained U-Nets in the Latent Diffusion Model and Stable Diffusion V1.4, as well as the Vision Transformer in PixArt, totally three models. We present the visualization results in Fig. 4, quantitative evaluation in Tab. 2, and user study in Figure 1 in the supplementary material." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LaVi-Bridge" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.187 + ], + "angle": 0, + "content": "Table 2: Quantitative evaluation of LaVi-Bridge under different generative vision models. \"Short\", \"Long\" and \"Comp\" denote short prompts, long prompts, and compositional prompts respectively. The best results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.199, + 0.78, + 0.395 + ], + "angle": 0, + "content": "
U-Net(LDM)U-Net(SD)Transformer(PixArt)
Short - FID25.9423.1123.02
Short - Aesthetics5.7035.8816.145
Short - CLIP Score0.31260.31560.3172
Long - Aesthetics6.1226.3056.406
Long - CLIP Score0.31890.31930.3210
Comp - Color0.40990.38890.3689
Comp - Shape0.37240.35520.3316
Comp - Texture0.50460.45240.4553
Comp - Spatial0.15500.15820.1725
Comp - Non-Spatial0.30040.30680.3098
Comp - Complex0.30600.30720.3014
" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.44, + 0.786, + 0.681 + ], + "angle": 0, + "content": "Visualization From Fig. 4, we can see that all these three vision models integrate well with T5-Large and generate relatively accurate images based on the given text prompts. From these cases, we can observe that the images generated by the transformer model based on PixArt exhibit richer details compared to the images generated by the other two models based on U-Net. For example, the forest in the first column, the hull of the pirate ship in the third column, and the bushes at the foot of the mountain in the sixth column are very intricate and realistic. Additionally, we can observe from these cases that the images generated by the model with U-Net of Stable Diffusion V1.4 have more detailed features compared to the images generated by the model with U-Net of Latent Diffusion Model. Furthermore, we can also find that, for the PixArt-based model, text alignment is better in some cases. For instance, in the image of the fifth column, only the model that is based on the transformer of PixArt generates the \"aged car\" mentioned in the prompt. Similarly, in the seventh column, the garden warrior holding a sword and shield is highly consistent with the prompt description." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.689, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Quantitative Evaluation From Tab. 2, it can be observed that for all the metrics measuring image quality, LaVi-Bridge with the PixArt vision model achieves the best results. Additionally, PixArt also achieves the best text alignment for both short and long prompts. This reflects the use of PixArt's transformer as a vision model can also improve the model's understanding of semantics to some extent. Additionally, it is noteworthy that the U-Net in Stable Diffusion, an enhanced version of the U-Net utilized in the Latent Diffusion Model, still outperforms Latent Diffusion Model's U-Net under LaVi-Bridge on all the metrics measuring image quality. This aligns with our previous discussion in Sec. 4.2 and further validates the underlying motivation behind our proposed LaVi-Bridge." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.127 + ], + "angle": 0, + "content": "S. Zhao et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.177 + ], + "angle": 0, + "content": "User Study We follow the settings described in Sec. 4.1, and the results are presented in Section A in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.198, + 0.388, + 0.214 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.222, + 0.786, + 0.296 + ], + "angle": 0, + "content": "In this section, we investigate two sets of ablation experiments. The first set aims to explore the impact of training LaVi-Bridge on the original pre-trained text-to-image diffusion model. The second set of experiments is to study the effects of LoRA and adapters in LaVi-Bridge. For both sets of experiments, we present visualization results in Fig. 5 and provide quantitative evaluations in Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.297, + 0.786, + 0.417 + ], + "angle": 0, + "content": "Training with LaVi-Bridge We investigate the impact of our LaVi-Bridge training framework on the original pre-trained text-to-image diffusion model. Specifically, we consider Stable Diffusion V1.4 which adopts CLIP text encoder as its language model and U-Net as its vision model. We incorporate LoRA and an adapter and apply LaVi-Bridge to the same language and vision models with identical structures and weights to those in Stable Diffusion V1.4. We then compare the performance of the model under LaVi-Bridge with the original Stable Diffusion V1.4." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.418, + 0.786, + 0.613 + ], + "angle": 0, + "content": "The visualization results are shown in the first two rows of Fig. 5. For these two models, there is no significant difference in image quality and text alignment, varying on a case-by-case basis. In some cases, Stable Diffusion performs better, such as in the third column, where Stable Diffusion successfully generates the case of a \"Fox bracelet made of buckskin with fox features\", while the model trained under LaVi-Bridge only generates the fox and fails to understand the bracelet made of buckskin. Similarly, in the case of Marvel's Hulk playing basketball, Stable Diffusion generates a slam dunk action following the prompt, whereas the model trained under LaVi-Bridge does not. However, in the second column, the model trained under LaVi-Bridge correctly understands the quantity and successfully generates two elephants, while Stable Diffusion only generates one. Moreover, in the last column, the model trained under LaVi-Bridge accurately describes a frog in a spacesuit, while Stable Diffusion fails." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.614, + 0.786, + 0.689 + ], + "angle": 0, + "content": "The left two columns of Tab. 3 present the quantitative evaluation results. It can be observed that Stable Diffusion achieves the best image quality and text alignment for both short prompts and long prompts. However, for compositional prompts, the model trained under LaVi-Bridge outperforms Stable Diffusion in four out of six settings." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.69, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Based on the visualization results and quantitative evaluations, we can conclude that overall there is no significant improvement or decline in text alignment. Regarding image quality, it should be noted that training with LaVi-Bridge may result in a decrease compared to the original text-to-image diffusion model, if the same models and weights are used. However, it is important to understand that the main purpose of LaVi-Bridge is to establish connections between different language and vision models, enabling the utilization of more advanced models for performance enhancement. It is not intended to be directly applied to the original text-to-image diffusion models using the same models and weights." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.648, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LaVi-Bridge" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.146, + 0.785, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.464, + 0.788, + 0.52 + ], + "angle": 0, + "content": "Fig. 5: Visualization results of the ablation study. The top two rows show the impact of LaVi-Bridge on the original pre-trained text-to-image diffusion models. The bottom three rows illustrate the influence of the adapter and LoRA. The prompts are displayed at the top or bottom of each column." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.703 + ], + "angle": 0, + "content": "LoRA and Adapter Here, we investigate the role of LoRA and adapters in LaVi-Bridge. We use T5-Large as the language model and Stable Diffusion V1.4's U-Net as the vision model. For the LoRA experiments, we kept the language and vision models fixed without introducing LoRA, and only trained the adapter. For the adapter experiments, considering the mismatch in the dimensions of text embeddings from the language model and the input embeddings acceptable by the vision model, we aligned the dimensions between the language and vision models using a single linear layer instead of stacked feedforward layers which include non-linear activation layers. Under this setting, we trained both LoRA and this linear layer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.84 + ], + "angle": 0, + "content": "The visualization results are shown in the bottom three rows of Fig. 5. We can observe that both image quality and text alignment are significantly affected when LoRA and adapters are not used. For example, in the case of \"Bull Fit Athlete\" in the seventh column, without LoRA or the adapter, the model cannot understand and integrate these two less related elements, and the image quality is much lower compared to results generated by the original setting. We also found that the results without LoRA are worse than those without the adapter. For instance in the fourth column, there is not even a ferret present in the image in the absence of LoRA." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.127 + ], + "angle": 0, + "content": "S. Zhao et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 3: Quantitative evaluation of the ablation study. The left two columns present the impact of LaVi-Bridge on the original pre-trained text-to-image diffusion models. The right three columns demonstrate the influence of the adapter and LoRA. \"Short\", \"Long\" and \"Comp\" denote short prompts, long prompts, and compositional prompts respectively. The best results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.231, + 0.227, + 0.779, + 0.422 + ], + "angle": 0, + "content": "
SDCLIP+U-Netw/o Adapterw/o LoRAT5+U-Net
Short - FID20.3223.5723.8122.3523.11
Short - Aesthetics5.8995.6095.8075.8295.881
Short - CLIP Score0.31320.31020.31470.31070.3156
Long - Aesthetics6.1206.0036.1316.2736.305
Long - CLIP Score0.31710.31200.31060.30970.3193
Comp - Color0.35700.35780.35500.24850.3889
Comp - Shape0.35630.37520.30440.29440.3552
Comp - Texture0.40280.45060.40010.31900.4524
Comp - Spatial0.12250.12960.16510.09560.1582
Comp - Non-Spatial0.31040.30090.30650.29980.3068
Comp - Complex0.30420.29850.28780.26870.3072
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.454, + 0.788, + 0.531 + ], + "angle": 0, + "content": "The right three columns of Tab. 3 present the quantitative evaluation results. We find that our default setting, which utilizes both LoRA and the adapter, achieves the best performance in most cases. Additionally, overall, the absence of LoRA has a significant impact on text alignment, with many text alignment evaluation metrics being much lower compared to the absence of the adapter." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.56, + 0.36, + 0.577 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this paper, we propose LaVi-Bridge, which works on text-to-image diffusion models. LaVi-Bridge is capable of connecting various language models and generative vision models for text-to-image generation. It is highly versatile and can adapt to different structures. LaVi-Bridge is also flexible, as it achieves integration without modifying the original weights of language and vision models. Instead, it utilizes LoRA and an adapter for fine-tuning. Additionally, under LaVi-Bridge, using superior language or vision models can enhance the text comprehension capability or image quality. These advantages enable LaVi-Bridge to help text-to-image diffusion models leverage the latest advancements in the areas of Natural Language Processing and Computer Vision, to enhance text-to-image generation. We believe that this task holds significant research value and requires further exploration. LaVi-Bridge allows designers, artists, and others to flexibly utilize existing language and vision models to achieve their creative goals. It is of utmost importance to avoid misuse and mitigate potential negative social impacts. In practical deployment, it is crucial to standardize its usage, improve model transparency." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LaVi-Bridge" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. Bao, F., Li, C., Cao, Y., Zhu, J.: All are worth words: a vit backbone for score-based diffusion models. CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.26 + ], + "angle": 0, + "content": "2. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., Manassra, W., Dhariwal, P., Chu, C., Jiao, Y., Ramesh, A.: Improving image generation with better captions. https://cdn.openai.com/papers/dall-e-3.pdf (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.286 + ], + "angle": 0, + "content": "3. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.327 + ], + "angle": 0, + "content": "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. NeurIPS 2020" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.329, + 0.785, + 0.356 + ], + "angle": 0, + "content": "5. Chen, J., Huang, Y., Lv, T., Cui, L., Chen, Q., Wei, F.: Textdiffuser: Diffusion models as text painters. NeurIPS (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.357, + 0.785, + 0.398 + ], + "angle": 0, + "content": "6. Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wu, Y., Wang, Z., Kwok, J., Luo, P., Lu, H., et al.: Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.785, + 0.425 + ], + "angle": 0, + "content": "7. Croitoru, F.A., Hondru, V., Ionescu, R.T., Shah, M.: Diffusion models in vision: A survey. TPAMI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.785, + 0.453 + ], + "angle": 0, + "content": "8. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. NAACL-HLT (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.454, + 0.785, + 0.479 + ], + "angle": 0, + "content": "9. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "0. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.: Cogview: Mastering text-to-image generation via transformers. NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.785, + 0.577 + ], + "angle": 0, + "content": "1. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.578, + 0.785, + 0.619 + ], + "angle": 0, + "content": "2. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.785, + 0.646 + ], + "angle": 0, + "content": "3. Ge, S., Park, T., Zhu, J.Y., Huang, J.B.: Expressive text-to-image generation with rich text. ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "4. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. NeurIPS (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.716 + ], + "angle": 0, + "content": "5. Guo, Y., Yang, C., Rao, A., Liang, Z., Wang, Y., Qiao, Y., Agrawala, M., Lin, D., Dai, B.: Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.744 + ], + "angle": 0, + "content": "6. Hao, S., Han, K., Zhao, S., Wong, K.Y.K.: Vico: Detail-preserving visual condition for personalized text-to-image generation. arXiv preprint arXiv:2306.00971 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.785 + ], + "angle": 0, + "content": "7. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.812 + ], + "angle": 0, + "content": "8. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "9. Ho, J., Salimans, T.: Classifier-free diffusion guidance. NeurIPS Workshop on Deep Generative Models and Downstream Applications (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.127 + ], + "angle": 0, + "content": "S. Zhao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "20. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "21. Huang, K., Sun, K., Xie, E., Li, Z., Liu, X.: T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. arXiv preprint arXiv: 2307.06350 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.219, + 0.785, + 0.247 + ], + "angle": 0, + "content": "22. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.248, + 0.785, + 0.275 + ], + "angle": 0, + "content": "23. LAION-AI: aesthetic-predictor. https://github.com/LAION-AI/aesthetic-predictor (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.785, + 0.331 + ], + "angle": 0, + "content": "24. Lewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mohamed, A., Levy, O., Stoyanov, V., Zettlemoyer, L.: Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.785, + 0.36 + ], + "angle": 0, + "content": "25. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. ECCV (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.767, + 0.374 + ], + "angle": 0, + "content": "26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. ICLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.785, + 0.403 + ], + "angle": 0, + "content": "27. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.785, + 0.445 + ], + "angle": 0, + "content": "28. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.785, + 0.487 + ], + "angle": 0, + "content": "29. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.767, + 0.501 + ], + "angle": 0, + "content": "30. Peebles, W., Xie, S.: Scalable diffusion models with transformers. ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.543 + ], + "angle": 0, + "content": "31. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.785, + 0.572 + ], + "angle": 0, + "content": "32. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training. OpenAI blog (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.785, + 0.601 + ], + "angle": 0, + "content": "33. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.785, + 0.642 + ], + "angle": 0, + "content": "34. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.785, + 0.685 + ], + "angle": 0, + "content": "35. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.785, + 0.713 + ], + "angle": 0, + "content": "36. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.785, + 0.742 + ], + "angle": 0, + "content": "37. Reed, S., Akata, Z., Yan, X., Logeswaran, L., Schiele, B., Lee, H.: Generative adversarial text to image synthesis. ICML (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.77 + ], + "angle": 0, + "content": "38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.785, + 0.799 + ], + "angle": 0, + "content": "39. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. MICCAI (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "40. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. CVPR (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "LaVi-Bridge" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "41. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.231 + ], + "angle": 0, + "content": "42. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.786, + 0.245 + ], + "angle": 0, + "content": "43. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.245, + 0.786, + 0.259 + ], + "angle": 0, + "content": "44. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.786, + 0.286 + ], + "angle": 0, + "content": "45. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.786, + 0.313 + ], + "angle": 0, + "content": "46. Sutskever, I., Vinyals, O., Le, Q.V.: Sequence to sequence learning with neural networks. NeurIPS (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.786, + 0.355 + ], + "angle": 0, + "content": "47. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.786, + 0.383 + ], + "angle": 0, + "content": "48. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.383, + 0.786, + 0.424 + ], + "angle": 0, + "content": "49. Wu, W., Li, Z., He, Y., Shou, M.Z., Shen, C., Cheng, L., Li, Y., Gao, T., Zhang, D., Wang, Z.: Paragraph-to-image generation with information-enriched diffusion model. arXiv preprint arXiv:2311.14284 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.424, + 0.786, + 0.452 + ], + "angle": 0, + "content": "50. Xue, Z., Song, G., Guo, Q., Liu, B., Zong, Z., Liu, Y., Luo, P.: Raphael: Text-to-image generation via large mixture of diffusion paths. NeurIPS (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.452, + 0.786, + 0.493 + ], + "angle": 0, + "content": "51. Zhang, H., Xu, T., Li, H., Zhang, S., Wang, X., Huang, X., Metaxas, D.N.: Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.493, + 0.786, + 0.522 + ], + "angle": 0, + "content": "52. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.786, + 0.55 + ], + "angle": 0, + "content": "53. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Unicontrolnet: All-in-one control to text-to-image diffusion models. NeurIPS (2024)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.55 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Bridging Different Language Models and Generative Vision Models for Text-to-Image Generation/1764d397-410b-451c-a099-8966f6b5772c_origin.pdf b/2024/Bridging Different Language Models and Generative Vision Models for Text-to-Image Generation/1764d397-410b-451c-a099-8966f6b5772c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..23fdef16479823a53c1b6b7c6b4f949a69b7f14a --- /dev/null +++ b/2024/Bridging Different Language Models and Generative Vision Models for Text-to-Image Generation/1764d397-410b-451c-a099-8966f6b5772c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd821d2aba42f83fb31d207a3f7cc68bc1c9244c7ce9ae0db626574622ea93de +size 7232339 diff --git a/2024/Bridging Different Language Models and Generative Vision Models for Text-to-Image Generation/full.md b/2024/Bridging Different Language Models and Generative Vision Models for Text-to-Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0cb0d91e9d18dd8d4eb31a90dcbcda59e10f6dbd --- /dev/null +++ b/2024/Bridging Different Language Models and Generative Vision Models for Text-to-Image Generation/full.md @@ -0,0 +1,249 @@ +# Bridging Different Language Models and Generative Vision Models for Text-to-Image Generation + +Shihao Zhao $^{1}$ , Shaozhe Hao $^{1}$ , Bojia Zi $^{2}$ , Huaizhe Xu $^{3}$ , and Kwan-Yee K. Wong $^{1\star}$ + +1 The University of Hong Kong + +{shzhao, szhao, kykwong}@cs.hku.hk + +2 The Chinese University of Hong Kong + +bjzi@se.cuhk.edu.hk + +3 The Hong Kong University of Science and Technology + +hxubr@connect.ust.hk + +Abstract. Text-to-image generation has made significant advancements with the introduction of text-to-image diffusion models. These models typically consist of a language model that interprets user prompts and a vision model that generates corresponding images. As language and vision models continue to progress in their respective domains, there is a great potential in exploring the replacement of components in text-to-image diffusion models with more advanced counterparts. A broader research objective would therefore be to investigate the integration of any two unrelated language and generative vision models for text-to-image generation. In this paper, we explore this objective and propose LaVi-Bridge, a pipeline that enables the integration of diverse pre-trained language models and generative vision models for text-to-image generation. By leveraging LoRA and adapters, LaVi-Bridge offers a flexible and plug-and-play approach without requiring modifications to the original weights of the language and vision models. Our pipeline is compatible with various language models and generative vision models, accommodating different structures. Within this framework, we demonstrate that incorporating superior modules, such as more advanced language models or generative vision models, results in notable improvements in capabilities like text alignment or image quality. Extensive evaluations have been conducted to verify the effectiveness of LaVi-Bridge. Code is available at https://github.com/ShihaoZhaoZSH/LaVi-Bridge. + +Keywords: Diffusion model $\cdot$ Text-to-image generation + +# 1 Introduction + +In recent years, there have been remarkable advancements in the field of text-to-image generation, specifically through the use of diffusion models [9, 18, 43, 45]. + +![](images/5ec3a7ca92f91d7be5173d94c1dd33629d5b59caefa36e7e9386410ff58731c5.jpg) +Fig. 1: Overview of LaVi-Bridge. LaVi-Bridge is capable of integrating various language models and generative vision models. On the left side, we keep the vision model fixed and experiment with different language models in our pipeline. On the right side, we keep the language model fixed and try out different vision models. We display the visualization results alongside each combination. + +These models have made significant contributions and have gained considerable attention for their exceptional performance. By leveraging large-scale training datasets alongside large deep models, text-to-image diffusion models are capable of producing high-quality images that faithfully align with the textual descriptions provided by users. This has rendered them highly applicable in real-world scenarios such as content creation and architectural design. + +Text-to-image diffusion models [2, 6, 29, 35, 38, 41, 50] typically consist of two key components, namely a language model and a generative vision model. The language model is responsible for comprehending the input prompts, whereas the vision model is tasked with generating images that align with the extracted context. Existing text-to-image diffusion models employ various language models and generative vision models and have gained widespread usage. For instance, Stable Diffusion (SD) [38] is a highly popular text-to-image diffusion model that employs the CLIP text encoder [31] as its language model and a U-Net [39] as its generative vision model. Another example is PixArt [6], a recently proposed text-to-image diffusion model that adopts the T5 [34] as its language model and a Vision Transformer (ViT) [11] as its generative vision model. These models are trained on a vast amount of text-image pairs, enabling seamless collaboration between their language modules and vision modules. + +The advancements in deep language models and deep vision models have witnessed rapid progress in recent years, with both fields experiencing continuous developments and the introduction of more powerful models. However, this rapid development poses a challenge for the research in text-to-image generation when it comes to integrating more advanced language or vision models into existing text-to-image diffusion models. The problem of how to integrate any two unrelated language and vision models is unexplored, and the impact of newly developed models on text-to-image generation capabilities also remains + +uncertain. The current situation highlights the presence of a gap between the language or vision modules within text-to-image diffusion models and the state-of-the-art models in their respective domains. Therefore, it has become crucial to address this gap and explore ways to incorporate more advanced language or vision models into existing text-to-image diffusion models. Furthermore, the broader challenge of integrating any pre-trained language model with any generative vision model deserves a thorough investigation. + +In this paper, our objective is to delve into the aforementioned problem. We propose LaVi-Bridge, a flexible framework that facilitates the integration of diverse well-trained language models and generative vision models to achieve text-to-image generation. Our framework enables the integration of two unrelated language and vision models that have not been previously trained together, as shown in Fig. 1. Importantly, LaVi-Bridge does not require modifying the original weights of the language and vision models. Instead, it injects LoRA [20] into the language and vision models separately and utilizes an adapter to bridge these two modules. Moreover, LaVi-Bridge only necessitates a relatively small dataset to integrate different language models and generative vision models for text-to-image generation. + +We summarize the advantages and features of LaVi-Bridge as follows: + +1. LaVi-Bridge is designed for text-to-image diffusion models and serves as a bridge, capable of connecting various pre-trained language models and generative vision models. Our framework can accommodate different model structures, including encoder-only, encoder-decoder, and decoder-only language models, as well as U-Net-based and Transformer-based generative vision models. +2. LaVi-Bridge utilizes LoRA and adapters, eliminating the need to modify the original weights of the models. It is more flexible and requires relatively small computing resources compared to training the entire diffusion model. +3. We evaluated various text-image alignment and image quality metrics on short prompts, long prompts, and compositional prompts. We also conducted extensive visualization. We then drew several conclusions. For instance, integrating superior models leads to improved performance in the corresponding modality, such as enhanced semantic understanding with advanced language models or improved image quality with more powerful generative vision models. Additionally, the diffusion model utilizing Llama-2 demonstrates exceptional semantic understanding, while the diffusion model utilizing the transformer in PixArt yields images with enhanced aesthetics. + +# 2 Related Work + +# 2.1 Language Models and Generative Vision Models + +The mainstream Large Language Models (LLMs) [8, 32-34] are built based on the transformer structure [48], with three main types of architectures, namely + +encoder-only, encoder-decoder, and decoder-only. All these three belong to Sequence to Sequence (Seq2Seq) [46]. The encoder-only architecture is exemplified by BERT [8]. CLIP text encoder [31] is based on BERT and further trained to align with the image domain. Models of this type excel at understanding the content of the input and generating outputs tailored to specific tasks. On the other hand, the encoder-decoder framework is adept at handling tasks that involve complex mappings between input and output sequences. Examples include T5 [34] and BART [24]. Recently, due to the tremendous success of ChatGPT, attention has been drawn to models that consist solely of a decoder, like GPT-3 [4] and Llama-2 [47]. The decoder-only architecture demonstrates exceptional performance in semantic understanding. For text-to-image generation, all three types of LLMs can provide effective semantic information to serve as conditions for image generation in diffusion models. In this paper, we explore and compare all these three types of language models. + +A generative vision model refers to a vision model with the ability to generate images or visual contents. There are two common types of structures, namely U-Net-based [39] and Transformer-based [11]. Generative Adversarial Networks (GANs) [14, 37, 51] employ a framework consisting of a discriminator and a generator, with the generator's structure based on U-Net. On the other hand, motivated by the success of GPT models, recent works have attempted to use the Transformer architecture for image generation in an autoregressive manner, with notable examples being DALLE [36] and CogView [10]. Another popular class of generative models is diffusion models [7, 18, 43, 45], which are based on the diffusion process and gradually denoise to produce natural images. Early diffusion models often employed U-Net as their generative vision model, such as Stable Diffusion, which scaled up the Latent Diffusion Model (LDM) [38] with larger data scales. Some recent works have started to replace the U-Net in diffusion models with Vision Transformer and have made significant progress, such as DiT [30], U-ViT [1] and PixArt [6]. In this paper, we focus on diffusion models and explore both U-Net-based and Transformer-based vision models. + +# 2.2 Text-to-Image Diffusion Models + +Text-to-image diffusion models [2,9,29,38,41,44] are capable of generating images based on user prompts. These models consist of two main components, namely a language model and a vision model. The language module is responsible for understanding the text input provided by the user, extracting contextual information, and injecting it into the vision module to generate the desired image. Text-to-image diffusion models have paved the way for various exciting research areas, including image editing [3,22,27], controllable image generation [28,52,53], personalized object generation [12, 16, 40], as well as other interesting applications [5,13,15]. In the extensive exploration of diffusion models, researchers have utilized different language models and vision models. For instance, Stable Diffusion [38] employs CLIP text encoder [31] as its language model and U-Net as its vision model.Imagen [41] utilizes T5 [34] as its language model, which claims to enhance both sample fidelity and image-text alignment. ParaDiffusion [49] + +focuses on paragraph-to-image generation and leverages the powerful semantic understanding capability of Llama-2 [47] to comprehend lengthy sentences. PixArt [6], on the other hand, utilizes a ViT [11] as its vision model and achieves high image fidelity while being trained at a lower cost. + +After training on a large dataset of text-image pairs [42], the language and vision models in the text-to-image diffusion model become closely intertwined. This tight coupling ensures a strong alignment between the provided text description and the generated image, but at the same time also limits the flexibility of the diffusion model. For instance, if a more advanced language or vision model becomes available, it may have the potential to enhance the text-to-image task. However, decoupling the language and vision modules in existing text-to-image diffusion models and replacing a module with a new one is nontrivial. Therefore, this paper explores the dilemma faced by text-to-image generation and proposes a framework that enables efficient integration of various language models and generative vision models. + +# 3 Method + +# 3.1 Preliminary + +A diffusion model is based on the diffusion process for image generation. This process consists of two stages, namely the forward process and the reverse process. During the forward process, Gaussian noise is progressively added to a natural image until the image becomes completely noisy. After that, during the reverse process, the noise is gradually eliminated over a series of time steps, resulting in a natural image. In the reverse process, a trainable vision model is used to predict and remove the noise. By employing this denoising model, we are able to obtain a natural image from Gaussian noise by denoising. Within a text-to-image diffusion model, there are two components at each time step, namely a language model $f$ and a vision model $g$ . The language model converts user input text $y$ into embeddings, which capture the semantic meaning of the text. On the other hand, the vision model, which is the denoising model aforementioned, encodes image features $z$ , extracting relevant visual information from the input images. The interaction between text embeddings and image features is achieved through cross-attention layers, which can be formulated as + +$$ +c = f (y), \tag {1} +$$ + +$$ +Q = W _ {q} (z), K = W _ {k} (c), V = W _ {v} (c), \tag {2} +$$ + +$$ +C r o s s A t t e n t i o n (Q, K, V) = \text {s o f t m a x} (Q \cdot K ^ {T}) \cdot V, \tag {3} +$$ + +where $W_{q}, W_{k}$ and $W_{v}$ are projection matrices. + +# 3.2 Language and Vision Alignment + +LaVi-Bridge enables the integration of any two pre-trained language and generative vision models, even though these models are not related and have been + +![](images/7f8a055761e674f935c0550440915fa22f8ac8f5c5c47e96a7db6dc70b8c09aa.jpg) +Fig. 2: Pipeline of LaVi-Bridge. We select one model each from the language and vision model pools. We then freeze the pre-trained language and vision models and incorporate LoRA into both models. The connection between the language and vision models is established through an adapter. The only weights we need to train are the ones introduced by LoRA and the adapter. + +trained separately. Here, we denote the language model as $f$ and the vision model as $g$ , as mentioned previously. If we directly interact the textual information and image information using Eq. (1), considering that $f$ and $g$ are trained independently, the parameters in the cross-attention layers of $g$ cannot comprehend the text embedding output by $f$ , resulting in meaningless model outputs. + +To establish a connection between them, LaVi-Bridge keeps the pre-trained language and vision models fixed and utilizes LoRA to introduce trainable parameters $\Delta \theta$ into both the language model and the vision model. In this context, we denote the language and vision models with LoRA as $f^{\theta_1 + \Delta \theta_1}$ and $g^{\theta_2 + \Delta \theta_2}$ , where $\theta_{1}$ and $\theta_{2}$ are the original parameters of $f$ and $g$ , respectively. Furthermore, we introduce an adapter as a bridge between the language model and vision model to facilitate better alignment. The adapter consists of stacked feedforward layers, denoted as $h$ . Consequently, the cross-attention layer can be expressed as + +$$ +c = f ^ {\theta_ {1} + \Delta \theta_ {1}} (y), \tag {4} +$$ + +$$ +Q = W _ {q} ^ {\theta_ {2} + \Delta \theta_ {2}} (z), K = W _ {k} ^ {\theta_ {2} + \Delta \theta_ {2}} (h (c)), V = W _ {v} ^ {\theta_ {2} + \Delta \theta_ {2}} (h (c)), \tag {5} +$$ + +$$ +C r o s s A t t e n t i o n (Q, K, V) = s o f t m a x (Q \cdot K ^ {T}) \cdot V. \tag {6} +$$ + +Now, we only need to train $\Delta \theta_{1}$ , $\Delta \theta_{2}$ , and $h$ on a relatively small amount of text-image pairs. After training, the language and generative vision models can effectively collaborate to generate meaningful images. We present the framework of LaVi-Bridge in Fig. 2. LaVi-Bridge is very straightforward, with both LoRA and the adapter being its crucial and indispensable components. + +# 3.3 Design Details + +LaVi-Bridge is designed to accommodate a wide range of language model structures, including encoder-only, encoder-decoder, decoder-only, as well as generative vision model structures such as U-Net and ViT. In the language model, we inject LoRA into all linear layers of the attention layers. Likewise, in a transformer-based vision model, LoRA is injected into all linear layers of the attention layers. In a U-Net-based vision model, LoRA is injected into all linear layers and convolutional layers of the ResBlocks, attention, and cross-attention layers. To address the dimension disparity between the output embedding of the language model and the dimensions handled by the cross-attention of the vision model, we employ two feedforward layers for the adapter. The input dimension of the adapter matches the output text embedding dimension of the language model, while the output dimension aligns with the dimensions received by the cross-attention of the vision model. + +For training, we first select the language and generative vision models that we choose to integrate. We keep their original weights fixed and train LoRA and the adapter on text-image pairs following the design mentioned above. The trained LoRA and adapter have fewer parameters compared to the original model weights, which makes LaVi-Bridge highly flexible. For evaluation, we used various metrics to assess text alignment and image quality across short prompts, long prompts, and compositional prompts. + +# 4 Experiments + +# 4.1 Experimental Settings + +In this section, we explored the performance of different language models and generative vision models under LaVi-Bridge. We also tested the impact of LoRA and adapters. We trained on a dataset consisting of a total of 1 million text-image pairs, including around 600k text-image pairs from the COCO2017 [25] train set and 400k text-image pairs from an internal dataset with high-quality images and captions. For each setting, we set the LoRA rank to 32, image resolution to $512 \times 512$ and the batch size to 256. We used the AdamW optimizer [26] with a learning rate of $1 \times 10^{-4}$ and trained for a total of 50k steps. During inference, we employed the DDIM sampler [43] for sampling with the number of time steps set to 50 and the classifier free guidance scale [19] set to 7.5. + +As mentioned above, we conducted our quantitative evaluation on short prompts, long prompts, and compositional prompts. Specifically, + +1. For short prompts, we evaluated using the COCO2014 [25] validation set. We randomly sampled 30k images and tested image quality and text alignment within this subset. We used FID [17] and aesthetic score [23] as evaluation metrics for image quality and CLIP score for text alignment. +2. For long prompts, we employed the same 30k-subset of COCO2014 and utilized Llama-2 to generate expanded captions ranging from 20 to 70 words to + +![](images/033d20f43b41fcf5baa497aa5d811d38aef26228fac6bffd8c173e86cf1fa8f4.jpg) +Fig. 3: Visualization results of LaVi-Bridge with different language models. The first row to the fifth row present the results with CLIP text encoder, T5-Small, T5-Base, T5-Large, and Llama-2, respectively. The prompts are displayed at the top or bottom of each column. + +construct a dataset of 30k long prompts. Since the caption expansion process does not refer to the content of the reference image, we solely used aesthetic score to evaluate image quality and CLIP score for text alignment. + +3. For compositional prompts, we utilized the benchmark proposed by Comp-bench [21]. Compositional prompts were mainly used to test the model's understanding of textual attributes, such as generating correct object properties like color and shape, as well as accurate relationships between objects, such as spatial positioning. + +We conducted a user study on different combinations of language and vision models. For each combination, we evaluated two metrics, namely image quality and text alignment. Users were asked to rank the generated images based on these evaluation criteria. The image ranked last received a score of 1, the second-to-last received a score of 2, and so on. We then calculated the percentage of scores for each model. We selected 20 prompts and included 30 users participated in the testing. In addition to quantitative evaluation and user study, we provided ample visualization results in each section to offer a more intuitive understanding of the performance of each model. + +Table 1: Quantitative evaluation of LaVi-Bridge with different language models. "Short", "Long" and "Comp" denote short prompts, long prompts, and compositional prompts respectively. The best results are in bold. + +
CLIPT5-SmallT5-BaseT5-LargeLlama-2
Short - FID23.5722.9822.6223.1121.80
Short - Aesthetics5.6095.8135.8885.8815.883
Short - CLIP Score0.31020.31220.31490.31560.3172
Long - Aesthetics6.0036.2066.2846.3056.355
Long - CLIP Score0.31200.31110.31790.31930.3231
Comp - Color0.35780.33680.38560.38890.4859
Comp - Shape0.37520.29620.32660.35520.4285
Comp - Texture0.45060.37280.41320.45240.5055
Comp - Spatial0.12960.14560.15690.15820.1914
Comp - Non-Spatial0.30090.29840.30540.30680.3106
Comp - Complex0.29850.27280.30550.30720.3094
+ +# 4.2 Evaluation on Different Language Models + +This section evaluates the performance of LaVi-Bridge with different language models. We fixed the vision model to the U-Net of Stable Diffusion V1.4 and integrated it with different language models under LaVi-Bridge. We considered CLIP text encoder, based on the encoder-only framework, T5 series (T5-Small, T5-Base, T5-Large), based on the encoder-decoder framework, and Llama-2-7B, based on the decoder-only framework. We present the visualization results in Fig. 3, quantitative evaluation in Tab. 1, and user study in Figure 1 in the supplementary material. + +Visualization From Fig. 3, we can observe that with LaVi-Bridge, all these language models can effectively integrate with U-Net of Stable Diffusion V1.4 and generate meaningful results, such as cases of the cat and living room in Fig. 3. This demonstrates the great generalization ability of LaVi-Bridge for various language models. Additionally, we notice that the performance of different model structures varies when the provided prompts contain more complex semantics. We find that the text-to-image diffusion model corresponding to Llama-2 can perfectly describe semantic information. For example, in the third column, Llama-2's generated result effectively integrates a woman into the sea of fragmented porcelain. In the fourth column, it correctly understands and generates both the girl and the cat in a paper craft art. In the seventh column, it even portrays an entire beach scene using yarn. These examples surpass the capabilities of those models with CLIP and T5. Furthermore, we observe that T5-Large and Llama-2 accurately generate food and wine in the case of Iron Man, and in the last column, they successfully generate "an ancient stone with eyes in dark yellow and emerald". Models with CLIP text encoder, T5-Small, and T5-Base are not able to capture these cases accurately. + +![](images/d46fa202ef10d74063e853d6c281bafbbafc1f1b726440f99ca8312eacdee910.jpg) +Fig. 4: Visualization results of LaVi-Bridge under different generative vision models. The first row to the third row present the results with U-Net in Latent Diffusion Model, U-Net in Stable Diffusion V1.4 and transformer in PixArt, respectively. The prompts are displayed at the top or bottom of each column. + +Quantitative Evaluation From Tab. 1, we can observe that Llama-2 achieves the best results for all the metrics used to evaluate text alignment ability, under the setting of all the short prompts, long prompts, and compositional prompts. Besides, Llama-2 also performs the best on most of the metrics evaluating image quality. On the other hand, as the model capacity increases, in general circumstances T5-Large usually outperforms T5-Base, and T5-Base outperforms T5-Small in the area of Natural Language Processing. This conclusion also holds true for LaVi-Bridge. For all the metrics used to evaluate text alignment ability in Tab. 1, T5-Large is superior to T5-Base, and T5-Base is superior to T5-Small. This tells us that incorporating a better language model into the text-to-image diffusion model under LaVi-Bridge can lead to improved text alignment. This makes one of the motivations of LaVi-Bridge meaningful, which is that replacing the model in the existing text-to-image diffusion model with a better model can lead to performance improvements. + +User Study We follow the settings described in Sec. 4.1, and the results are presented in Section A in the supplementary material. + +# 4.3 Evaluation on Different Vision Models + +This section evaluates the performance of LaVi-Bridge with different vision models. We fixed the language model to T5-Large and integrated it with different generative vision models under LaVi-Bridge. We considered the well-trained U-Nets in the Latent Diffusion Model and Stable Diffusion V1.4, as well as the Vision Transformer in PixArt, totally three models. We present the visualization results in Fig. 4, quantitative evaluation in Tab. 2, and user study in Figure 1 in the supplementary material. + +Table 2: Quantitative evaluation of LaVi-Bridge under different generative vision models. "Short", "Long" and "Comp" denote short prompts, long prompts, and compositional prompts respectively. The best results are in bold. + +
U-Net(LDM)U-Net(SD)Transformer(PixArt)
Short - FID25.9423.1123.02
Short - Aesthetics5.7035.8816.145
Short - CLIP Score0.31260.31560.3172
Long - Aesthetics6.1226.3056.406
Long - CLIP Score0.31890.31930.3210
Comp - Color0.40990.38890.3689
Comp - Shape0.37240.35520.3316
Comp - Texture0.50460.45240.4553
Comp - Spatial0.15500.15820.1725
Comp - Non-Spatial0.30040.30680.3098
Comp - Complex0.30600.30720.3014
+ +Visualization From Fig. 4, we can see that all these three vision models integrate well with T5-Large and generate relatively accurate images based on the given text prompts. From these cases, we can observe that the images generated by the transformer model based on PixArt exhibit richer details compared to the images generated by the other two models based on U-Net. For example, the forest in the first column, the hull of the pirate ship in the third column, and the bushes at the foot of the mountain in the sixth column are very intricate and realistic. Additionally, we can observe from these cases that the images generated by the model with U-Net of Stable Diffusion V1.4 have more detailed features compared to the images generated by the model with U-Net of Latent Diffusion Model. Furthermore, we can also find that, for the PixArt-based model, text alignment is better in some cases. For instance, in the image of the fifth column, only the model that is based on the transformer of PixArt generates the "aged car" mentioned in the prompt. Similarly, in the seventh column, the garden warrior holding a sword and shield is highly consistent with the prompt description. + +Quantitative Evaluation From Tab. 2, it can be observed that for all the metrics measuring image quality, LaVi-Bridge with the PixArt vision model achieves the best results. Additionally, PixArt also achieves the best text alignment for both short and long prompts. This reflects the use of PixArt's transformer as a vision model can also improve the model's understanding of semantics to some extent. Additionally, it is noteworthy that the U-Net in Stable Diffusion, an enhanced version of the U-Net utilized in the Latent Diffusion Model, still outperforms Latent Diffusion Model's U-Net under LaVi-Bridge on all the metrics measuring image quality. This aligns with our previous discussion in Sec. 4.2 and further validates the underlying motivation behind our proposed LaVi-Bridge. + +User Study We follow the settings described in Sec. 4.1, and the results are presented in Section A in the supplementary material. + +# 4.4 Ablation Study + +In this section, we investigate two sets of ablation experiments. The first set aims to explore the impact of training LaVi-Bridge on the original pre-trained text-to-image diffusion model. The second set of experiments is to study the effects of LoRA and adapters in LaVi-Bridge. For both sets of experiments, we present visualization results in Fig. 5 and provide quantitative evaluations in Tab. 3. + +Training with LaVi-Bridge We investigate the impact of our LaVi-Bridge training framework on the original pre-trained text-to-image diffusion model. Specifically, we consider Stable Diffusion V1.4 which adopts CLIP text encoder as its language model and U-Net as its vision model. We incorporate LoRA and an adapter and apply LaVi-Bridge to the same language and vision models with identical structures and weights to those in Stable Diffusion V1.4. We then compare the performance of the model under LaVi-Bridge with the original Stable Diffusion V1.4. + +The visualization results are shown in the first two rows of Fig. 5. For these two models, there is no significant difference in image quality and text alignment, varying on a case-by-case basis. In some cases, Stable Diffusion performs better, such as in the third column, where Stable Diffusion successfully generates the case of a "Fox bracelet made of buckskin with fox features", while the model trained under LaVi-Bridge only generates the fox and fails to understand the bracelet made of buckskin. Similarly, in the case of Marvel's Hulk playing basketball, Stable Diffusion generates a slam dunk action following the prompt, whereas the model trained under LaVi-Bridge does not. However, in the second column, the model trained under LaVi-Bridge correctly understands the quantity and successfully generates two elephants, while Stable Diffusion only generates one. Moreover, in the last column, the model trained under LaVi-Bridge accurately describes a frog in a spacesuit, while Stable Diffusion fails. + +The left two columns of Tab. 3 present the quantitative evaluation results. It can be observed that Stable Diffusion achieves the best image quality and text alignment for both short prompts and long prompts. However, for compositional prompts, the model trained under LaVi-Bridge outperforms Stable Diffusion in four out of six settings. + +Based on the visualization results and quantitative evaluations, we can conclude that overall there is no significant improvement or decline in text alignment. Regarding image quality, it should be noted that training with LaVi-Bridge may result in a decrease compared to the original text-to-image diffusion model, if the same models and weights are used. However, it is important to understand that the main purpose of LaVi-Bridge is to establish connections between different language and vision models, enabling the utilization of more advanced models for performance enhancement. It is not intended to be directly applied to the original text-to-image diffusion models using the same models and weights. + +![](images/040ac1ab6d6724c09ae0a959bc2f24eec8d97e401580e199eb4c4f4fd2af4e98.jpg) +Fig. 5: Visualization results of the ablation study. The top two rows show the impact of LaVi-Bridge on the original pre-trained text-to-image diffusion models. The bottom three rows illustrate the influence of the adapter and LoRA. The prompts are displayed at the top or bottom of each column. + +LoRA and Adapter Here, we investigate the role of LoRA and adapters in LaVi-Bridge. We use T5-Large as the language model and Stable Diffusion V1.4's U-Net as the vision model. For the LoRA experiments, we kept the language and vision models fixed without introducing LoRA, and only trained the adapter. For the adapter experiments, considering the mismatch in the dimensions of text embeddings from the language model and the input embeddings acceptable by the vision model, we aligned the dimensions between the language and vision models using a single linear layer instead of stacked feedforward layers which include non-linear activation layers. Under this setting, we trained both LoRA and this linear layer. + +The visualization results are shown in the bottom three rows of Fig. 5. We can observe that both image quality and text alignment are significantly affected when LoRA and adapters are not used. For example, in the case of "Bull Fit Athlete" in the seventh column, without LoRA or the adapter, the model cannot understand and integrate these two less related elements, and the image quality is much lower compared to results generated by the original setting. We also found that the results without LoRA are worse than those without the adapter. For instance in the fourth column, there is not even a ferret present in the image in the absence of LoRA. + +Table 3: Quantitative evaluation of the ablation study. The left two columns present the impact of LaVi-Bridge on the original pre-trained text-to-image diffusion models. The right three columns demonstrate the influence of the adapter and LoRA. "Short", "Long" and "Comp" denote short prompts, long prompts, and compositional prompts respectively. The best results are in bold. + +
SDCLIP+U-Netw/o Adapterw/o LoRAT5+U-Net
Short - FID20.3223.5723.8122.3523.11
Short - Aesthetics5.8995.6095.8075.8295.881
Short - CLIP Score0.31320.31020.31470.31070.3156
Long - Aesthetics6.1206.0036.1316.2736.305
Long - CLIP Score0.31710.31200.31060.30970.3193
Comp - Color0.35700.35780.35500.24850.3889
Comp - Shape0.35630.37520.30440.29440.3552
Comp - Texture0.40280.45060.40010.31900.4524
Comp - Spatial0.12250.12960.16510.09560.1582
Comp - Non-Spatial0.31040.30090.30650.29980.3068
Comp - Complex0.30420.29850.28780.26870.3072
+ +The right three columns of Tab. 3 present the quantitative evaluation results. We find that our default setting, which utilizes both LoRA and the adapter, achieves the best performance in most cases. Additionally, overall, the absence of LoRA has a significant impact on text alignment, with many text alignment evaluation metrics being much lower compared to the absence of the adapter. + +# 5 Conclusion + +In this paper, we propose LaVi-Bridge, which works on text-to-image diffusion models. LaVi-Bridge is capable of connecting various language models and generative vision models for text-to-image generation. It is highly versatile and can adapt to different structures. LaVi-Bridge is also flexible, as it achieves integration without modifying the original weights of language and vision models. Instead, it utilizes LoRA and an adapter for fine-tuning. Additionally, under LaVi-Bridge, using superior language or vision models can enhance the text comprehension capability or image quality. These advantages enable LaVi-Bridge to help text-to-image diffusion models leverage the latest advancements in the areas of Natural Language Processing and Computer Vision, to enhance text-to-image generation. We believe that this task holds significant research value and requires further exploration. LaVi-Bridge allows designers, artists, and others to flexibly utilize existing language and vision models to achieve their creative goals. It is of utmost importance to avoid misuse and mitigate potential negative social impacts. In practical deployment, it is crucial to standardize its usage, improve model transparency. + +# References + +1. Bao, F., Li, C., Cao, Y., Zhu, J.: All are worth words: a vit backbone for score-based diffusion models. CVPR (2023) +2. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., Manassra, W., Dhariwal, P., Chu, C., Jiao, Y., Ramesh, A.: Improving image generation with better captions. https://cdn.openai.com/papers/dall-e-3.pdf (2023) +3. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. CVPR (2023) +4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. NeurIPS 2020 +5. Chen, J., Huang, Y., Lv, T., Cui, L., Chen, Q., Wei, F.: Textdiffuser: Diffusion models as text painters. NeurIPS (2024) +6. Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wu, Y., Wang, Z., Kwok, J., Luo, P., Lu, H., et al.: Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. ICLR (2024) +7. Croitoru, F.A., Hondru, V., Ionescu, R.T., Shah, M.: Diffusion models in vision: A survey. TPAMI (2023) +8. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. NAACL-HLT (2018) +9. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. NeurIPS (2021) +0. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.: Cogview: Mastering text-to-image generation via transformers. NeurIPS (2021) +1. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021) +2. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618 (2022) +3. Ge, S., Park, T., Zhu, J.Y., Huang, J.B.: Expressive text-to-image generation with rich text. ICCV (2023) +4. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. NeurIPS (2014) +5. Guo, Y., Yang, C., Rao, A., Liang, Z., Wang, Y., Qiao, Y., Agrawala, M., Lin, D., Dai, B.: Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. ICLR (2024) +6. Hao, S., Han, K., Zhao, S., Wong, K.Y.K.: Vico: Detail-preserving visual condition for personalized text-to-image generation. arXiv preprint arXiv:2306.00971 (2023) +7. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS (2017) +8. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS (2020) +9. Ho, J., Salimans, T.: Classifier-free diffusion guidance. NeurIPS Workshop on Deep Generative Models and Downstream Applications (2021) + +20. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. ICLR (2022) +21. Huang, K., Sun, K., Xie, E., Li, Z., Liu, X.: T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. arXiv preprint arXiv: 2307.06350 (2023) +22. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. CVPR (2023) +23. LAION-AI: aesthetic-predictor. https://github.com/LAION-AI/aesthetic-predictor (2022) +24. Lewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mohamed, A., Levy, O., Stoyanov, V., Zettlemoyer, L.: Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461 (2019) +25. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. ECCV (2014) +26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. ICLR (2019) +27. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. ICLR (2022) +28. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023) +29. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. ICML (2021) +30. Peebles, W., Xie, S.: Scalable diffusion models with transformers. ICCV (2023) +31. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. ICML (2021) +32. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training. OpenAI blog (2018) +33. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog (2019) +34. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR (2020) +35. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022) +36. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. ICML (2021) +37. Reed, S., Akata, Z., Yan, X., Logeswaran, L., Schiele, B., Lee, H.: Generative adversarial text to image synthesis. ICML (2016) +38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. CVPR (2022) +39. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. MICCAI (2015) +40. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. CVPR (2023) + +41. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022) +42. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021) +43. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. ICLR (2021) +44. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. ICML (2023) +45. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. ICLR (2021) +46. Sutskever, I., Vinyals, O., Le, Q.V.: Sequence to sequence learning with neural networks. NeurIPS (2014) +47. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) +48. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS (2017) +49. Wu, W., Li, Z., He, Y., Shou, M.Z., Shen, C., Cheng, L., Li, Y., Gao, T., Zhang, D., Wang, Z.: Paragraph-to-image generation with information-enriched diffusion model. arXiv preprint arXiv:2311.14284 (2023) +50. Xue, Z., Song, G., Guo, Q., Liu, B., Zong, Z., Liu, Y., Luo, P.: Raphael: Text-to-image generation via large mixture of diffusion paths. NeurIPS (2024) +51. Zhang, H., Xu, T., Li, H., Zhang, S., Wang, X., Huang, X., Metaxas, D.N.: Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. ICCV (2017) +52. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. ICCV (2023) +53. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Unicontrolnet: All-in-one control to text-to-image diffusion models. NeurIPS (2024) \ No newline at end of file diff --git a/2024/Bridging Different Language Models and Generative Vision Models for Text-to-Image Generation/images.zip b/2024/Bridging Different Language Models and Generative Vision Models for Text-to-Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a0cef36fab8f65aca62a4655760a3b73222cdc03 --- /dev/null +++ b/2024/Bridging Different Language Models and Generative Vision Models for Text-to-Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49e131bd9b83e36c747025a94ad6bec09f623775c24c018efc25c83bd38553d2 +size 761396 diff --git a/2024/Bridging Different Language Models and Generative Vision Models for Text-to-Image Generation/layout.json b/2024/Bridging Different Language Models and Generative Vision Models for Text-to-Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fac2651d8f53ac758fa8b17d8be8cf1646effc32 --- /dev/null +++ b/2024/Bridging Different Language Models and Generative Vision Models for Text-to-Image Generation/layout.json @@ -0,0 +1,6951 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 165, + 111, + 451, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 111, + 451, + 166 + ], + "spans": [ + { + "bbox": [ + 165, + 111, + 451, + 166 + ], + "type": "text", + "content": "Bridging Different Language Models and Generative Vision Models for Text-to-Image Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 170, + 185, + 444, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 185, + 444, + 210 + ], + "spans": [ + { + "bbox": [ + 170, + 185, + 444, + 210 + ], + "type": "text", + "content": "Shihao Zhao" + }, + { + "bbox": [ + 170, + 185, + 444, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 170, + 185, + 444, + 210 + ], + "type": "text", + "content": ", Shaozhe Hao" + }, + { + "bbox": [ + 170, + 185, + 444, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 170, + 185, + 444, + 210 + ], + "type": "text", + "content": ", Bojia Zi" + }, + { + "bbox": [ + 170, + 185, + 444, + 210 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 170, + 185, + 444, + 210 + ], + "type": "text", + "content": ", Huaizhe Xu" + }, + { + "bbox": [ + 170, + 185, + 444, + 210 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 170, + 185, + 444, + 210 + ], + "type": "text", + "content": ", and Kwan-Yee K. Wong" + }, + { + "bbox": [ + 170, + 185, + 444, + 210 + ], + "type": "inline_equation", + "content": "^{1\\star}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 241, + 219, + 372, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 219, + 372, + 231 + ], + "spans": [ + { + "bbox": [ + 241, + 219, + 372, + 231 + ], + "type": "text", + "content": "1 The University of Hong Kong" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 231, + 232, + 383, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 232, + 383, + 242 + ], + "spans": [ + { + "bbox": [ + 231, + 232, + 383, + 242 + ], + "type": "text", + "content": "{shzhao, szhao, kykwong}@cs.hku.hk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 225, + 242, + 389, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 242, + 389, + 253 + ], + "spans": [ + { + "bbox": [ + 225, + 242, + 389, + 253 + ], + "type": "text", + "content": "2 The Chinese University of Hong Kong" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 261, + 254, + 352, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 254, + 352, + 264 + ], + "spans": [ + { + "bbox": [ + 261, + 254, + 352, + 264 + ], + "type": "text", + "content": "bjzi@se.cuhk.edu.hk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 192, + 264, + 421, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 264, + 421, + 275 + ], + "spans": [ + { + "bbox": [ + 192, + 264, + 421, + 275 + ], + "type": "text", + "content": "3 The Hong Kong University of Science and Technology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 258, + 277, + 355, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 277, + 355, + 285 + ], + "spans": [ + { + "bbox": [ + 258, + 277, + 355, + 285 + ], + "type": "text", + "content": "hxubr@connect.ust.hk" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 160, + 314, + 455, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 314, + 455, + 558 + ], + "spans": [ + { + "bbox": [ + 160, + 314, + 455, + 558 + ], + "type": "text", + "content": "Abstract. Text-to-image generation has made significant advancements with the introduction of text-to-image diffusion models. These models typically consist of a language model that interprets user prompts and a vision model that generates corresponding images. As language and vision models continue to progress in their respective domains, there is a great potential in exploring the replacement of components in text-to-image diffusion models with more advanced counterparts. A broader research objective would therefore be to investigate the integration of any two unrelated language and generative vision models for text-to-image generation. In this paper, we explore this objective and propose LaVi-Bridge, a pipeline that enables the integration of diverse pre-trained language models and generative vision models for text-to-image generation. By leveraging LoRA and adapters, LaVi-Bridge offers a flexible and plug-and-play approach without requiring modifications to the original weights of the language and vision models. Our pipeline is compatible with various language models and generative vision models, accommodating different structures. Within this framework, we demonstrate that incorporating superior modules, such as more advanced language models or generative vision models, results in notable improvements in capabilities like text alignment or image quality. Extensive evaluations have been conducted to verify the effectiveness of LaVi-Bridge. Code is available at https://github.com/ShihaoZhaoZSH/LaVi-Bridge." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 160, + 566, + 391, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 566, + 391, + 578 + ], + "spans": [ + { + "bbox": [ + 160, + 566, + 391, + 578 + ], + "type": "text", + "content": "Keywords: Diffusion model " + }, + { + "bbox": [ + 160, + 566, + 391, + 578 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 566, + 391, + 578 + ], + "type": "text", + "content": " Text-to-image generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 597, + 230, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 230, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 230, + 611 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 622, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 622, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 622, + 482, + 647 + ], + "type": "text", + "content": "In recent years, there have been remarkable advancements in the field of text-to-image generation, specifically through the use of diffusion models [9, 18, 43, 45]." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 238, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 238, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 238, + 665 + ], + "type": "text", + "content": "* Corresponding authors" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 245 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 245 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 245 + ], + "type": "image", + "image_path": "5ec3a7ca92f91d7be5173d94c1dd33629d5b59caefa36e7e9386410ff58731c5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 254, + 482, + 309 + ], + "lines": [ + { + "bbox": [ + 130, + 254, + 482, + 309 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 482, + 309 + ], + "type": "text", + "content": "Fig. 1: Overview of LaVi-Bridge. LaVi-Bridge is capable of integrating various language models and generative vision models. On the left side, we keep the vision model fixed and experiment with different language models in our pipeline. On the right side, we keep the language model fixed and try out different vision models. We display the visualization results alongside each combination." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 338, + 482, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 409 + ], + "type": "text", + "content": "These models have made significant contributions and have gained considerable attention for their exceptional performance. By leveraging large-scale training datasets alongside large deep models, text-to-image diffusion models are capable of producing high-quality images that faithfully align with the textual descriptions provided by users. This has rendered them highly applicable in real-world scenarios such as content creation and architectural design." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 411, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 567 + ], + "type": "text", + "content": "Text-to-image diffusion models [2, 6, 29, 35, 38, 41, 50] typically consist of two key components, namely a language model and a generative vision model. The language model is responsible for comprehending the input prompts, whereas the vision model is tasked with generating images that align with the extracted context. Existing text-to-image diffusion models employ various language models and generative vision models and have gained widespread usage. For instance, Stable Diffusion (SD) [38] is a highly popular text-to-image diffusion model that employs the CLIP text encoder [31] as its language model and a U-Net [39] as its generative vision model. Another example is PixArt [6], a recently proposed text-to-image diffusion model that adopts the T5 [34] as its language model and a Vision Transformer (ViT) [11] as its generative vision model. These models are trained on a vast amount of text-image pairs, enabling seamless collaboration between their language modules and vision modules." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "The advancements in deep language models and deep vision models have witnessed rapid progress in recent years, with both fields experiencing continuous developments and the introduction of more powerful models. However, this rapid development poses a challenge for the research in text-to-image generation when it comes to integrating more advanced language or vision models into existing text-to-image diffusion models. The problem of how to integrate any two unrelated language and vision models is unexplored, and the impact of newly developed models on text-to-image generation capabilities also remains" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "S. Zhao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "uncertain. The current situation highlights the presence of a gap between the language or vision modules within text-to-image diffusion models and the state-of-the-art models in their respective domains. Therefore, it has become crucial to address this gap and explore ways to incorporate more advanced language or vision models into existing text-to-image diffusion models. Furthermore, the broader challenge of integrating any pre-trained language model with any generative vision model deserves a thorough investigation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 331 + ], + "type": "text", + "content": "In this paper, our objective is to delve into the aforementioned problem. We propose LaVi-Bridge, a flexible framework that facilitates the integration of diverse well-trained language models and generative vision models to achieve text-to-image generation. Our framework enables the integration of two unrelated language and vision models that have not been previously trained together, as shown in Fig. 1. Importantly, LaVi-Bridge does not require modifying the original weights of the language and vision models. Instead, it injects LoRA [20] into the language and vision models separately and utilizes an adapter to bridge these two modules. Moreover, LaVi-Bridge only necessitates a relatively small dataset to integrate different language models and generative vision models for text-to-image generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 332, + 454, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 332, + 454, + 344 + ], + "spans": [ + { + "bbox": [ + 146, + 332, + 454, + 344 + ], + "type": "text", + "content": "We summarize the advantages and features of LaVi-Bridge as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 354, + 481, + 571 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 136, + 354, + 481, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 354, + 481, + 425 + ], + "spans": [ + { + "bbox": [ + 136, + 354, + 481, + 425 + ], + "type": "text", + "content": "1. LaVi-Bridge is designed for text-to-image diffusion models and serves as a bridge, capable of connecting various pre-trained language models and generative vision models. Our framework can accommodate different model structures, including encoder-only, encoder-decoder, and decoder-only language models, as well as U-Net-based and Transformer-based generative vision models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 426, + 481, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 426, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 136, + 426, + 481, + 462 + ], + "type": "text", + "content": "2. LaVi-Bridge utilizes LoRA and adapters, eliminating the need to modify the original weights of the models. It is more flexible and requires relatively small computing resources compared to training the entire diffusion model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 463, + 481, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 463, + 481, + 571 + ], + "spans": [ + { + "bbox": [ + 136, + 463, + 481, + 571 + ], + "type": "text", + "content": "3. We evaluated various text-image alignment and image quality metrics on short prompts, long prompts, and compositional prompts. We also conducted extensive visualization. We then drew several conclusions. For instance, integrating superior models leads to improved performance in the corresponding modality, such as enhanced semantic understanding with advanced language models or improved image quality with more powerful generative vision models. Additionally, the diffusion model utilizing Llama-2 demonstrates exceptional semantic understanding, while the diffusion model utilizing the transformer in PixArt yields images with enhanced aesthetics." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 591, + 237, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 591, + 237, + 604 + ], + "spans": [ + { + "bbox": [ + 132, + 591, + 237, + 604 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 619, + 406, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 619, + 406, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 619, + 406, + 631 + ], + "type": "text", + "content": "2.1 Language Models and Generative Vision Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "The mainstream Large Language Models (LLMs) [8, 32-34] are built based on the transformer structure [48], with three main types of architectures, namely" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "LaVi-Bridge" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": "encoder-only, encoder-decoder, and decoder-only. All these three belong to Sequence to Sequence (Seq2Seq) [46]. The encoder-only architecture is exemplified by BERT [8]. CLIP text encoder [31] is based on BERT and further trained to align with the image domain. Models of this type excel at understanding the content of the input and generating outputs tailored to specific tasks. On the other hand, the encoder-decoder framework is adept at handling tasks that involve complex mappings between input and output sequences. Examples include T5 [34] and BART [24]. Recently, due to the tremendous success of ChatGPT, attention has been drawn to models that consist solely of a decoder, like GPT-3 [4] and Llama-2 [47]. The decoder-only architecture demonstrates exceptional performance in semantic understanding. For text-to-image generation, all three types of LLMs can provide effective semantic information to serve as conditions for image generation in diffusion models. In this paper, we explore and compare all these three types of language models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 283, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 475 + ], + "type": "text", + "content": "A generative vision model refers to a vision model with the ability to generate images or visual contents. There are two common types of structures, namely U-Net-based [39] and Transformer-based [11]. Generative Adversarial Networks (GANs) [14, 37, 51] employ a framework consisting of a discriminator and a generator, with the generator's structure based on U-Net. On the other hand, motivated by the success of GPT models, recent works have attempted to use the Transformer architecture for image generation in an autoregressive manner, with notable examples being DALLE [36] and CogView [10]. Another popular class of generative models is diffusion models [7, 18, 43, 45], which are based on the diffusion process and gradually denoise to produce natural images. Early diffusion models often employed U-Net as their generative vision model, such as Stable Diffusion, which scaled up the Latent Diffusion Model (LDM) [38] with larger data scales. Some recent works have started to replace the U-Net in diffusion models with Vision Transformer and have made significant progress, such as DiT [30], U-ViT [1] and PixArt [6]. In this paper, we focus on diffusion models and explore both U-Net-based and Transformer-based vision models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 491, + 321, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 321, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 321, + 504 + ], + "type": "text", + "content": "2.2 Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "text", + "content": "Text-to-image diffusion models [2,9,29,38,41,44] are capable of generating images based on user prompts. These models consist of two main components, namely a language model and a vision model. The language module is responsible for understanding the text input provided by the user, extracting contextual information, and injecting it into the vision module to generate the desired image. Text-to-image diffusion models have paved the way for various exciting research areas, including image editing [3,22,27], controllable image generation [28,52,53], personalized object generation [12, 16, 40], as well as other interesting applications [5,13,15]. In the extensive exploration of diffusion models, researchers have utilized different language models and vision models. For instance, Stable Diffusion [38] employs CLIP text encoder [31] as its language model and U-Net as its vision model.Imagen [41] utilizes T5 [34] as its language model, which claims to enhance both sample fidelity and image-text alignment. ParaDiffusion [49]" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "S. Zhao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "focuses on paragraph-to-image generation and leverages the powerful semantic understanding capability of Llama-2 [47] to comprehend lengthy sentences. PixArt [6], on the other hand, utilizes a ViT [11] as its vision model and achieves high image fidelity while being trained at a lower cost." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 482, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 482, + 296 + ], + "type": "text", + "content": "After training on a large dataset of text-image pairs [42], the language and vision models in the text-to-image diffusion model become closely intertwined. This tight coupling ensures a strong alignment between the provided text description and the generated image, but at the same time also limits the flexibility of the diffusion model. For instance, if a more advanced language or vision model becomes available, it may have the potential to enhance the text-to-image task. However, decoupling the language and vision modules in existing text-to-image diffusion models and replacing a module with a new one is nontrivial. Therefore, this paper explores the dilemma faced by text-to-image generation and proposes a framework that enables efficient integration of various language models and generative vision models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 312, + 202, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 202, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 202, + 324 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 333, + 221, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 221, + 345 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 221, + 345 + ], + "type": "text", + "content": "3.1 Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 350, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 529 + ], + "type": "text", + "content": "A diffusion model is based on the diffusion process for image generation. This process consists of two stages, namely the forward process and the reverse process. During the forward process, Gaussian noise is progressively added to a natural image until the image becomes completely noisy. After that, during the reverse process, the noise is gradually eliminated over a series of time steps, resulting in a natural image. In the reverse process, a trainable vision model is used to predict and remove the noise. By employing this denoising model, we are able to obtain a natural image from Gaussian noise by denoising. Within a text-to-image diffusion model, there are two components at each time step, namely a language model " + }, + { + "bbox": [ + 130, + 350, + 482, + 529 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 350, + 482, + 529 + ], + "type": "text", + "content": " and a vision model " + }, + { + "bbox": [ + 130, + 350, + 482, + 529 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 350, + 482, + 529 + ], + "type": "text", + "content": ". The language model converts user input text " + }, + { + "bbox": [ + 130, + 350, + 482, + 529 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 350, + 482, + 529 + ], + "type": "text", + "content": " into embeddings, which capture the semantic meaning of the text. On the other hand, the vision model, which is the denoising model aforementioned, encodes image features " + }, + { + "bbox": [ + 130, + 350, + 482, + 529 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 350, + 482, + 529 + ], + "type": "text", + "content": ", extracting relevant visual information from the input images. The interaction between text embeddings and image features is achieved through cross-attention layers, which can be formulated as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 269, + 536, + 480, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 536, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 269, + 536, + 480, + 548 + ], + "type": "interline_equation", + "content": "c = f (y), \\tag {1}", + "image_path": "ade042ab05a77db91348869c6471d26694c5f11697b2cad70afc3879b2ebfbb0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 219, + 557, + 480, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 557, + 480, + 570 + ], + "spans": [ + { + "bbox": [ + 219, + 557, + 480, + 570 + ], + "type": "interline_equation", + "content": "Q = W _ {q} (z), K = W _ {k} (c), V = W _ {v} (c), \\tag {2}", + "image_path": "b0117d787f4a979b8bdca5d2ba336660e11f8f46db2fff1c79bb2117e48ef356.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 197, + 578, + 480, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 578, + 480, + 592 + ], + "spans": [ + { + "bbox": [ + 197, + 578, + 480, + 592 + ], + "type": "interline_equation", + "content": "C r o s s A t t e n t i o n (Q, K, V) = \\text {s o f t m a x} (Q \\cdot K ^ {T}) \\cdot V, \\tag {3}", + "image_path": "33cfbb90aca9e5d702d8234d82483bdaf9aa88a6954754d4047e4fdacc75053d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 597, + 341, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 597, + 341, + 611 + ], + "spans": [ + { + "bbox": [ + 131, + 597, + 341, + 611 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 597, + 341, + 611 + ], + "type": "inline_equation", + "content": "W_{q}, W_{k}" + }, + { + "bbox": [ + 131, + 597, + 341, + 611 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 597, + 341, + 611 + ], + "type": "inline_equation", + "content": "W_{v}" + }, + { + "bbox": [ + 131, + 597, + 341, + 611 + ], + "type": "text", + "content": " are projection matrices." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 624, + 323, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 323, + 636 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 323, + 636 + ], + "type": "text", + "content": "3.2 Language and Vision Alignment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "LaVi-Bridge enables the integration of any two pre-trained language and generative vision models, even though these models are not related and have been" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "LaVi-Bridge" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 480, + 238 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 480, + 238 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 480, + 238 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 480, + 238 + ], + "type": "image", + "image_path": "7f8a055761e674f935c0550440915fa22f8ac8f5c5c47e96a7db6dc70b8c09aa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 247, + 482, + 302 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 302 + ], + "type": "text", + "content": "Fig. 2: Pipeline of LaVi-Bridge. We select one model each from the language and vision model pools. We then freeze the pre-trained language and vision models and incorporate LoRA into both models. The connection between the language and vision models is established through an adapter. The only weights we need to train are the ones introduced by LoRA and the adapter." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 335, + 479, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 479, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 479, + 396 + ], + "type": "text", + "content": "trained separately. Here, we denote the language model as " + }, + { + "bbox": [ + 130, + 335, + 479, + 396 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 335, + 479, + 396 + ], + "type": "text", + "content": " and the vision model as " + }, + { + "bbox": [ + 130, + 335, + 479, + 396 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 335, + 479, + 396 + ], + "type": "text", + "content": ", as mentioned previously. If we directly interact the textual information and image information using Eq. (1), considering that " + }, + { + "bbox": [ + 130, + 335, + 479, + 396 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 335, + 479, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 335, + 479, + 396 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 335, + 479, + 396 + ], + "type": "text", + "content": " are trained independently, the parameters in the cross-attention layers of " + }, + { + "bbox": [ + 130, + 335, + 479, + 396 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 335, + 479, + 396 + ], + "type": "text", + "content": " cannot comprehend the text embedding output by " + }, + { + "bbox": [ + 130, + 335, + 479, + 396 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 335, + 479, + 396 + ], + "type": "text", + "content": ", resulting in meaningless model outputs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "text", + "content": "To establish a connection between them, LaVi-Bridge keeps the pre-trained language and vision models fixed and utilizes LoRA to introduce trainable parameters " + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\Delta \\theta" + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "text", + "content": " into both the language model and the vision model. In this context, we denote the language and vision models with LoRA as " + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "inline_equation", + "content": "f^{\\theta_1 + \\Delta \\theta_1}" + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "inline_equation", + "content": "g^{\\theta_2 + \\Delta \\theta_2}" + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\theta_{1}" + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\theta_{2}" + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "text", + "content": " are the original parameters of " + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "text", + "content": ", respectively. Furthermore, we introduce an adapter as a bridge between the language model and vision model to facilitate better alignment. The adapter consists of stacked feedforward layers, denoted as " + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 399, + 482, + 495 + ], + "type": "text", + "content": ". Consequently, the cross-attention layer can be expressed as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 272, + 523, + 481, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 523, + 481, + 538 + ], + "spans": [ + { + "bbox": [ + 272, + 523, + 481, + 538 + ], + "type": "interline_equation", + "content": "c = f ^ {\\theta_ {1} + \\Delta \\theta_ {1}} (y), \\tag {4}", + "image_path": "06382c7d2f4731e6d94acae6186d3b5b62c130deda5da6da74405de0abed023d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 173, + 546, + 481, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 546, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 173, + 546, + 481, + 563 + ], + "type": "interline_equation", + "content": "Q = W _ {q} ^ {\\theta_ {2} + \\Delta \\theta_ {2}} (z), K = W _ {k} ^ {\\theta_ {2} + \\Delta \\theta_ {2}} (h (c)), V = W _ {v} ^ {\\theta_ {2} + \\Delta \\theta_ {2}} (h (c)), \\tag {5}", + "image_path": "c54a7f6d78db4dc18b609eb1ee9ffa9671e46d09140a67186241ac5686b3dc1f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 190, + 569, + 481, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 569, + 481, + 585 + ], + "spans": [ + { + "bbox": [ + 190, + 569, + 481, + 585 + ], + "type": "interline_equation", + "content": "C r o s s A t t e n t i o n (Q, K, V) = s o f t m a x (Q \\cdot K ^ {T}) \\cdot V. \\tag {6}", + "image_path": "070c00c25f3062854b3f32e946da815e27809e5dabf37445b224987b0b8906d1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Now, we only need to train " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_{1}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_{2}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " on a relatively small amount of text-image pairs. After training, the language and generative vision models can effectively collaborate to generate meaningful images. We present the framework of LaVi-Bridge in Fig. 2. LaVi-Bridge is very straightforward, with both LoRA and the adapter being its crucial and indispensable components." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 225, + 101 + ], + "type": "text", + "content": "S. Zhao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 234, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 234, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 234, + 128 + ], + "type": "text", + "content": "3.3 Design Details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 291 + ], + "type": "text", + "content": "LaVi-Bridge is designed to accommodate a wide range of language model structures, including encoder-only, encoder-decoder, decoder-only, as well as generative vision model structures such as U-Net and ViT. In the language model, we inject LoRA into all linear layers of the attention layers. Likewise, in a transformer-based vision model, LoRA is injected into all linear layers of the attention layers. In a U-Net-based vision model, LoRA is injected into all linear layers and convolutional layers of the ResBlocks, attention, and cross-attention layers. To address the dimension disparity between the output embedding of the language model and the dimensions handled by the cross-attention of the vision model, we employ two feedforward layers for the adapter. The input dimension of the adapter matches the output text embedding dimension of the language model, while the output dimension aligns with the dimensions received by the cross-attention of the vision model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 376 + ], + "type": "text", + "content": "For training, we first select the language and generative vision models that we choose to integrate. We keep their original weights fixed and train LoRA and the adapter on text-image pairs following the design mentioned above. The trained LoRA and adapter have fewer parameters compared to the original model weights, which makes LaVi-Bridge highly flexible. For evaluation, we used various metrics to assess text alignment and image quality across short prompts, long prompts, and compositional prompts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 396, + 230, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 230, + 408 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 230, + 408 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 421, + 273, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 421, + 273, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 421, + 273, + 434 + ], + "type": "text", + "content": "4.1 Experimental Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 441, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 561 + ], + "type": "text", + "content": "In this section, we explored the performance of different language models and generative vision models under LaVi-Bridge. We also tested the impact of LoRA and adapters. We trained on a dataset consisting of a total of 1 million text-image pairs, including around 600k text-image pairs from the COCO2017 [25] train set and 400k text-image pairs from an internal dataset with high-quality images and captions. For each setting, we set the LoRA rank to 32, image resolution to " + }, + { + "bbox": [ + 130, + 441, + 482, + 561 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 441, + 482, + 561 + ], + "type": "text", + "content": " and the batch size to 256. We used the AdamW optimizer [26] with a learning rate of " + }, + { + "bbox": [ + 130, + 441, + 482, + 561 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 441, + 482, + 561 + ], + "type": "text", + "content": " and trained for a total of 50k steps. During inference, we employed the DDIM sampler [43] for sampling with the number of time steps set to 50 and the classifier free guidance scale [19] set to 7.5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 562, + 481, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 562, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 562, + 481, + 586 + ], + "type": "text", + "content": "As mentioned above, we conducted our quantitative evaluation on short prompts, long prompts, and compositional prompts. Specifically," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 135, + 593, + 479, + 663 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 135, + 593, + 479, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 593, + 479, + 639 + ], + "spans": [ + { + "bbox": [ + 135, + 593, + 479, + 639 + ], + "type": "text", + "content": "1. For short prompts, we evaluated using the COCO2014 [25] validation set. We randomly sampled 30k images and tested image quality and text alignment within this subset. We used FID [17] and aesthetic score [23] as evaluation metrics for image quality and CLIP score for text alignment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 135, + 641, + 479, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 641, + 479, + 663 + ], + "spans": [ + { + "bbox": [ + 135, + 641, + 479, + 663 + ], + "type": "text", + "content": "2. For long prompts, we employed the same 30k-subset of COCO2014 and utilized Llama-2 to generate expanded captions ranging from 20 to 70 words to" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "LaVi-Bridge" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 116, + 480, + 357 + ], + "blocks": [ + { + "bbox": [ + 134, + 116, + 480, + 357 + ], + "lines": [ + { + "bbox": [ + 134, + 116, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 480, + 357 + ], + "type": "image", + "image_path": "033d20f43b41fcf5baa497aa5d811d38aef26228fac6bffd8c173e86cf1fa8f4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 365, + 482, + 409 + ], + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 409 + ], + "type": "text", + "content": "Fig. 3: Visualization results of LaVi-Bridge with different language models. The first row to the fifth row present the results with CLIP text encoder, T5-Small, T5-Base, T5-Large, and Llama-2, respectively. The prompts are displayed at the top or bottom of each column." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 441, + 481, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 441, + 481, + 477 + ], + "spans": [ + { + "bbox": [ + 147, + 441, + 481, + 477 + ], + "type": "text", + "content": "construct a dataset of 30k long prompts. Since the caption expansion process does not refer to the content of the reference image, we solely used aesthetic score to evaluate image quality and CLIP score for text alignment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 479, + 481, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 479, + 481, + 540 + ], + "spans": [ + { + "bbox": [ + 135, + 479, + 481, + 540 + ], + "type": "text", + "content": "3. For compositional prompts, we utilized the benchmark proposed by Comp-bench [21]. Compositional prompts were mainly used to test the model's understanding of textual attributes, such as generating correct object properties like color and shape, as well as accurate relationships between objects, such as spatial positioning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": "We conducted a user study on different combinations of language and vision models. For each combination, we evaluated two metrics, namely image quality and text alignment. Users were asked to rank the generated images based on these evaluation criteria. The image ranked last received a score of 1, the second-to-last received a score of 2, and so on. We then calculated the percentage of scores for each model. We selected 20 prompts and included 30 users participated in the testing. In addition to quantitative evaluation and user study, we provided ample visualization results in each section to offer a more intuitive understanding of the performance of each model." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 225, + 100 + ], + "type": "text", + "content": "S. Zhao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 157, + 479, + 313 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Quantitative evaluation of LaVi-Bridge with different language models. \"Short\", \"Long\" and \"Comp\" denote short prompts, long prompts, and compositional prompts respectively. The best results are in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 157, + 479, + 313 + ], + "lines": [ + { + "bbox": [ + 138, + 157, + 479, + 313 + ], + "spans": [ + { + "bbox": [ + 138, + 157, + 479, + 313 + ], + "type": "table", + "html": "
CLIPT5-SmallT5-BaseT5-LargeLlama-2
Short - FID23.5722.9822.6223.1121.80
Short - Aesthetics5.6095.8135.8885.8815.883
Short - CLIP Score0.31020.31220.31490.31560.3172
Long - Aesthetics6.0036.2066.2846.3056.355
Long - CLIP Score0.31200.31110.31790.31930.3231
Comp - Color0.35780.33680.38560.38890.4859
Comp - Shape0.37520.29620.32660.35520.4285
Comp - Texture0.45060.37280.41320.45240.5055
Comp - Spatial0.12960.14560.15690.15820.1914
Comp - Non-Spatial0.30090.29840.30540.30680.3106
Comp - Complex0.29850.27280.30550.30720.3094
", + "image_path": "12ec0ec11c9d5b3857cf30132dc0a41ee5887a0035e18dd3a993e806f203ce15.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 338, + 370, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 338, + 370, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 338, + 370, + 350 + ], + "type": "text", + "content": "4.2 Evaluation on Different Language Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "text", + "content": "This section evaluates the performance of LaVi-Bridge with different language models. We fixed the vision model to the U-Net of Stable Diffusion V1.4 and integrated it with different language models under LaVi-Bridge. We considered CLIP text encoder, based on the encoder-only framework, T5 series (T5-Small, T5-Base, T5-Large), based on the encoder-decoder framework, and Llama-2-7B, based on the decoder-only framework. We present the visualization results in Fig. 3, quantitative evaluation in Tab. 1, and user study in Figure 1 in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": "Visualization From Fig. 3, we can observe that with LaVi-Bridge, all these language models can effectively integrate with U-Net of Stable Diffusion V1.4 and generate meaningful results, such as cases of the cat and living room in Fig. 3. This demonstrates the great generalization ability of LaVi-Bridge for various language models. Additionally, we notice that the performance of different model structures varies when the provided prompts contain more complex semantics. We find that the text-to-image diffusion model corresponding to Llama-2 can perfectly describe semantic information. For example, in the third column, Llama-2's generated result effectively integrates a woman into the sea of fragmented porcelain. In the fourth column, it correctly understands and generates both the girl and the cat in a paper craft art. In the seventh column, it even portrays an entire beach scene using yarn. These examples surpass the capabilities of those models with CLIP and T5. Furthermore, we observe that T5-Large and Llama-2 accurately generate food and wine in the case of Iron Man, and in the last column, they successfully generate \"an ancient stone with eyes in dark yellow and emerald\". Models with CLIP text encoder, T5-Small, and T5-Base are not able to capture these cases accurately." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "LaVi-Bridge" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 480, + 273 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 480, + 273 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 480, + 273 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 480, + 273 + ], + "type": "image", + "image_path": "d46fa202ef10d74063e853d6c281bafbbafc1f1b726440f99ca8312eacdee910.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 281, + 482, + 326 + ], + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 326 + ], + "type": "text", + "content": "Fig. 4: Visualization results of LaVi-Bridge under different generative vision models. The first row to the third row present the results with U-Net in Latent Diffusion Model, U-Net in Stable Diffusion V1.4 and transformer in PixArt, respectively. The prompts are displayed at the top or bottom of each column." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 350, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 517 + ], + "type": "text", + "content": "Quantitative Evaluation From Tab. 1, we can observe that Llama-2 achieves the best results for all the metrics used to evaluate text alignment ability, under the setting of all the short prompts, long prompts, and compositional prompts. Besides, Llama-2 also performs the best on most of the metrics evaluating image quality. On the other hand, as the model capacity increases, in general circumstances T5-Large usually outperforms T5-Base, and T5-Base outperforms T5-Small in the area of Natural Language Processing. This conclusion also holds true for LaVi-Bridge. For all the metrics used to evaluate text alignment ability in Tab. 1, T5-Large is superior to T5-Base, and T5-Base is superior to T5-Small. This tells us that incorporating a better language model into the text-to-image diffusion model under LaVi-Bridge can lead to improved text alignment. This makes one of the motivations of LaVi-Bridge meaningful, which is that replacing the model in the existing text-to-image diffusion model with a better model can lead to performance improvements." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 517, + 482, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 482, + 541 + ], + "type": "text", + "content": "User Study We follow the settings described in Sec. 4.1, and the results are presented in Section A in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 560, + 354, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 560, + 354, + 572 + ], + "spans": [ + { + "bbox": [ + 131, + 560, + 354, + 572 + ], + "type": "text", + "content": "4.3 Evaluation on Different Vision Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "This section evaluates the performance of LaVi-Bridge with different vision models. We fixed the language model to T5-Large and integrated it with different generative vision models under LaVi-Bridge. We considered the well-trained U-Nets in the Latent Diffusion Model and Stable Diffusion V1.4, as well as the Vision Transformer in PixArt, totally three models. We present the visualization results in Fig. 4, quantitative evaluation in Tab. 2, and user study in Figure 1 in the supplementary material." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "S. Zhao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 157, + 477, + 312 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "type": "text", + "content": "Table 2: Quantitative evaluation of LaVi-Bridge under different generative vision models. \"Short\", \"Long\" and \"Comp\" denote short prompts, long prompts, and compositional prompts respectively. The best results are in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 157, + 477, + 312 + ], + "lines": [ + { + "bbox": [ + 140, + 157, + 477, + 312 + ], + "spans": [ + { + "bbox": [ + 140, + 157, + 477, + 312 + ], + "type": "table", + "html": "
U-Net(LDM)U-Net(SD)Transformer(PixArt)
Short - FID25.9423.1123.02
Short - Aesthetics5.7035.8816.145
Short - CLIP Score0.31260.31560.3172
Long - Aesthetics6.1226.3056.406
Long - CLIP Score0.31890.31930.3210
Comp - Color0.40990.38890.3689
Comp - Shape0.37240.35520.3316
Comp - Texture0.50460.45240.4553
Comp - Spatial0.15500.15820.1725
Comp - Non-Spatial0.30040.30680.3098
Comp - Complex0.30600.30720.3014
", + "image_path": "8ed4514aea38668104dc10dda60d59bc29bc9b597bf18bafc96be8b03736c125.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 348, + 481, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 348, + 481, + 539 + ], + "spans": [ + { + "bbox": [ + 133, + 348, + 481, + 539 + ], + "type": "text", + "content": "Visualization From Fig. 4, we can see that all these three vision models integrate well with T5-Large and generate relatively accurate images based on the given text prompts. From these cases, we can observe that the images generated by the transformer model based on PixArt exhibit richer details compared to the images generated by the other two models based on U-Net. For example, the forest in the first column, the hull of the pirate ship in the third column, and the bushes at the foot of the mountain in the sixth column are very intricate and realistic. Additionally, we can observe from these cases that the images generated by the model with U-Net of Stable Diffusion V1.4 have more detailed features compared to the images generated by the model with U-Net of Latent Diffusion Model. Furthermore, we can also find that, for the PixArt-based model, text alignment is better in some cases. For instance, in the image of the fifth column, only the model that is based on the transformer of PixArt generates the \"aged car\" mentioned in the prompt. Similarly, in the seventh column, the garden warrior holding a sword and shield is highly consistent with the prompt description." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 545, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 545, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 545, + 481, + 665 + ], + "type": "text", + "content": "Quantitative Evaluation From Tab. 2, it can be observed that for all the metrics measuring image quality, LaVi-Bridge with the PixArt vision model achieves the best results. Additionally, PixArt also achieves the best text alignment for both short and long prompts. This reflects the use of PixArt's transformer as a vision model can also improve the model's understanding of semantics to some extent. Additionally, it is noteworthy that the U-Net in Stable Diffusion, an enhanced version of the U-Net utilized in the Latent Diffusion Model, still outperforms Latent Diffusion Model's U-Net under LaVi-Bridge on all the metrics measuring image quality. This aligns with our previous discussion in Sec. 4.2 and further validates the underlying motivation behind our proposed LaVi-Bridge." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "LaVi-Bridge" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "User Study We follow the settings described in Sec. 4.1, and the results are presented in Section A in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 156, + 237, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 156, + 237, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 156, + 237, + 169 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 175, + 481, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 175, + 481, + 234 + ], + "spans": [ + { + "bbox": [ + 132, + 175, + 481, + 234 + ], + "type": "text", + "content": "In this section, we investigate two sets of ablation experiments. The first set aims to explore the impact of training LaVi-Bridge on the original pre-trained text-to-image diffusion model. The second set of experiments is to study the effects of LoRA and adapters in LaVi-Bridge. For both sets of experiments, we present visualization results in Fig. 5 and provide quantitative evaluations in Tab. 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 235, + 481, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 235, + 481, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 235, + 481, + 330 + ], + "type": "text", + "content": "Training with LaVi-Bridge We investigate the impact of our LaVi-Bridge training framework on the original pre-trained text-to-image diffusion model. Specifically, we consider Stable Diffusion V1.4 which adopts CLIP text encoder as its language model and U-Net as its vision model. We incorporate LoRA and an adapter and apply LaVi-Bridge to the same language and vision models with identical structures and weights to those in Stable Diffusion V1.4. We then compare the performance of the model under LaVi-Bridge with the original Stable Diffusion V1.4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 331, + 481, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 481, + 485 + ], + "type": "text", + "content": "The visualization results are shown in the first two rows of Fig. 5. For these two models, there is no significant difference in image quality and text alignment, varying on a case-by-case basis. In some cases, Stable Diffusion performs better, such as in the third column, where Stable Diffusion successfully generates the case of a \"Fox bracelet made of buckskin with fox features\", while the model trained under LaVi-Bridge only generates the fox and fails to understand the bracelet made of buckskin. Similarly, in the case of Marvel's Hulk playing basketball, Stable Diffusion generates a slam dunk action following the prompt, whereas the model trained under LaVi-Bridge does not. However, in the second column, the model trained under LaVi-Bridge correctly understands the quantity and successfully generates two elephants, while Stable Diffusion only generates one. Moreover, in the last column, the model trained under LaVi-Bridge accurately describes a frog in a spacesuit, while Stable Diffusion fails." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 486, + 481, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 481, + 545 + ], + "type": "text", + "content": "The left two columns of Tab. 3 present the quantitative evaluation results. It can be observed that Stable Diffusion achieves the best image quality and text alignment for both short prompts and long prompts. However, for compositional prompts, the model trained under LaVi-Bridge outperforms Stable Diffusion in four out of six settings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "type": "text", + "content": "Based on the visualization results and quantitative evaluations, we can conclude that overall there is no significant improvement or decline in text alignment. Regarding image quality, it should be noted that training with LaVi-Bridge may result in a decrease compared to the original text-to-image diffusion model, if the same models and weights are used. However, it is important to understand that the main purpose of LaVi-Bridge is to establish connections between different language and vision models, enabling the utilization of more advanced models for performance enhancement. It is not intended to be directly applied to the original text-to-image diffusion models using the same models and weights." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "S. Zhao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 480, + 358 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 480, + 358 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 480, + 358 + ], + "type": "image", + "image_path": "040ac1ab6d6724c09ae0a959bc2f24eec8d97e401580e199eb4c4f4fd2af4e98.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 367, + 482, + 411 + ], + "lines": [ + { + "bbox": [ + 130, + 367, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 482, + 411 + ], + "type": "text", + "content": "Fig. 5: Visualization results of the ablation study. The top two rows show the impact of LaVi-Bridge on the original pre-trained text-to-image diffusion models. The bottom three rows illustrate the influence of the adapter and LoRA. The prompts are displayed at the top or bottom of each column." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 437, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 556 + ], + "type": "text", + "content": "LoRA and Adapter Here, we investigate the role of LoRA and adapters in LaVi-Bridge. We use T5-Large as the language model and Stable Diffusion V1.4's U-Net as the vision model. For the LoRA experiments, we kept the language and vision models fixed without introducing LoRA, and only trained the adapter. For the adapter experiments, considering the mismatch in the dimensions of text embeddings from the language model and the input embeddings acceptable by the vision model, we aligned the dimensions between the language and vision models using a single linear layer instead of stacked feedforward layers which include non-linear activation layers. Under this setting, we trained both LoRA and this linear layer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "The visualization results are shown in the bottom three rows of Fig. 5. We can observe that both image quality and text alignment are significantly affected when LoRA and adapters are not used. For example, in the case of \"Bull Fit Athlete\" in the seventh column, without LoRA or the adapter, the model cannot understand and integrate these two less related elements, and the image quality is much lower compared to results generated by the original setting. We also found that the results without LoRA are worse than those without the adapter. For instance in the fourth column, there is not even a ferret present in the image in the absence of LoRA." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 396, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 396, + 91, + 447, + 102 + ], + "type": "text", + "content": "LaVi-Bridge" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 141, + 179, + 476, + 334 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 3: Quantitative evaluation of the ablation study. The left two columns present the impact of LaVi-Bridge on the original pre-trained text-to-image diffusion models. The right three columns demonstrate the influence of the adapter and LoRA. \"Short\", \"Long\" and \"Comp\" denote short prompts, long prompts, and compositional prompts respectively. The best results are in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 179, + 476, + 334 + ], + "lines": [ + { + "bbox": [ + 141, + 179, + 476, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 179, + 476, + 334 + ], + "type": "table", + "html": "
SDCLIP+U-Netw/o Adapterw/o LoRAT5+U-Net
Short - FID20.3223.5723.8122.3523.11
Short - Aesthetics5.8995.6095.8075.8295.881
Short - CLIP Score0.31320.31020.31470.31070.3156
Long - Aesthetics6.1206.0036.1316.2736.305
Long - CLIP Score0.31710.31200.31060.30970.3193
Comp - Color0.35700.35780.35500.24850.3889
Comp - Shape0.35630.37520.30440.29440.3552
Comp - Texture0.40280.45060.40010.31900.4524
Comp - Spatial0.12250.12960.16510.09560.1582
Comp - Non-Spatial0.31040.30090.30650.29980.3068
Comp - Complex0.30420.29850.28780.26870.3072
", + "image_path": "0768d007a995b8e6c459bde1d8510ae578df1c6843f611a465c345de5076ac2e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 359, + 482, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 359, + 482, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 482, + 420 + ], + "type": "text", + "content": "The right three columns of Tab. 3 present the quantitative evaluation results. We find that our default setting, which utilizes both LoRA and the adapter, achieves the best performance in most cases. Additionally, overall, the absence of LoRA has a significant impact on text alignment, with many text alignment evaluation metrics being much lower compared to the absence of the adapter." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 443, + 220, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 443, + 220, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 443, + 220, + 456 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we propose LaVi-Bridge, which works on text-to-image diffusion models. LaVi-Bridge is capable of connecting various language models and generative vision models for text-to-image generation. It is highly versatile and can adapt to different structures. LaVi-Bridge is also flexible, as it achieves integration without modifying the original weights of language and vision models. Instead, it utilizes LoRA and an adapter for fine-tuning. Additionally, under LaVi-Bridge, using superior language or vision models can enhance the text comprehension capability or image quality. These advantages enable LaVi-Bridge to help text-to-image diffusion models leverage the latest advancements in the areas of Natural Language Processing and Computer Vision, to enhance text-to-image generation. We believe that this task holds significant research value and requires further exploration. LaVi-Bridge allows designers, artists, and others to flexibly utilize existing language and vision models to achieve their creative goals. It is of utmost importance to avoid misuse and mitigate potential negative social impacts. In practical deployment, it is crucial to standardize its usage, improve model transparency." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "S. Zhao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "text", + "content": "1. Bao, F., Li, C., Cao, Y., Zhu, J.: All are worth words: a vit backbone for score-based diffusion models. CVPR (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 205 + ], + "type": "text", + "content": "2. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., Manassra, W., Dhariwal, P., Chu, C., Jiao, Y., Ramesh, A.: Improving image generation with better captions. https://cdn.openai.com/papers/dall-e-3.pdf (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 206, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 480, + 226 + ], + "type": "text", + "content": "3. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. CVPR (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 258 + ], + "type": "text", + "content": "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. NeurIPS 2020" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 260, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 260, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 260, + 480, + 281 + ], + "type": "text", + "content": "5. Chen, J., Huang, Y., Lv, T., Cui, L., Chen, Q., Wei, F.: Textdiffuser: Diffusion models as text painters. NeurIPS (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 282, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 282, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 282, + 480, + 315 + ], + "type": "text", + "content": "6. Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wu, Y., Wang, Z., Kwok, J., Luo, P., Lu, H., et al.: Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. ICLR (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 316, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 480, + 336 + ], + "type": "text", + "content": "7. Croitoru, F.A., Hondru, V., Ionescu, R.T., Shah, M.: Diffusion models in vision: A survey. TPAMI (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 337, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 480, + 358 + ], + "type": "text", + "content": "8. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. NAACL-HLT (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 359, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 359, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 138, + 359, + 480, + 379 + ], + "type": "text", + "content": "9. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. NeurIPS (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 480, + 413 + ], + "type": "text", + "content": "0. Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et al.: Cogview: Mastering text-to-image generation via transformers. NeurIPS (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 414, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 480, + 456 + ], + "type": "text", + "content": "1. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 457, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 457, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 480, + 490 + ], + "type": "text", + "content": "2. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "type": "text", + "content": "3. Ge, S., Park, T., Zhu, J.Y., Huang, J.B.: Expressive text-to-image generation with rich text. ICCV (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "type": "text", + "content": "4. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. NeurIPS (2014)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 534, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 567 + ], + "type": "text", + "content": "5. Guo, Y., Yang, C., Rao, A., Liang, Z., Wang, Y., Qiao, Y., Agrawala, M., Lin, D., Dai, B.: Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. ICLR (2024)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "text", + "content": "6. Hao, S., Han, K., Zhao, S., Wong, K.Y.K.: Vico: Detail-preserving visual condition for personalized text-to-image generation. arXiv preprint arXiv:2306.00971 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 590, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 621 + ], + "type": "text", + "content": "7. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS (2017)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "text", + "content": "8. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS (2020)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "9. Ho, J., Salimans, T.: Classifier-free diffusion guidance. NeurIPS Workshop on Deep Generative Models and Downstream Applications (2021)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "LaVi-Bridge" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "20. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. ICLR (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 172 + ], + "type": "text", + "content": "21. Huang, K., Sun, K., Xie, E., Li, Z., Liu, X.: T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. arXiv preprint arXiv: 2307.06350 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 173, + 480, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 173, + 480, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 480, + 195 + ], + "type": "text", + "content": "22. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. CVPR (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 196, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 480, + 217 + ], + "type": "text", + "content": "23. LAION-AI: aesthetic-predictor. https://github.com/LAION-AI/aesthetic-predictor (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 218, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 480, + 262 + ], + "type": "text", + "content": "24. Lewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mohamed, A., Levy, O., Stoyanov, V., Zettlemoyer, L.: Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 262, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 480, + 285 + ], + "type": "text", + "content": "25. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. ECCV (2014)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 285, + 469, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 469, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 469, + 296 + ], + "type": "text", + "content": "26. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. ICLR (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 297, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 480, + 319 + ], + "type": "text", + "content": "27. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. ICLR (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 319, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 480, + 352 + ], + "type": "text", + "content": "28. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 353, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 480, + 385 + ], + "type": "text", + "content": "29. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. ICML (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 386, + 469, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 469, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 469, + 396 + ], + "type": "text", + "content": "30. Peebles, W., Xie, S.: Scalable diffusion models with transformers. ICCV (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 397, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 430 + ], + "type": "text", + "content": "31. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. ICML (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 430, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 480, + 453 + ], + "type": "text", + "content": "32. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training. OpenAI blog (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 453, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 480, + 475 + ], + "type": "text", + "content": "33. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 475, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 480, + 508 + ], + "type": "text", + "content": "34. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 509, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 480, + 542 + ], + "type": "text", + "content": "35. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 543, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 480, + 564 + ], + "type": "text", + "content": "36. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. ICML (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 565, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 480, + 587 + ], + "type": "text", + "content": "37. Reed, S., Akata, Z., Yan, X., Logeswaran, L., Schiele, B., Lee, H.: Generative adversarial text to image synthesis. ICML (2016)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 588, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 609 + ], + "type": "text", + "content": "38. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. CVPR (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "type": "text", + "content": "39. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. MICCAI (2015)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "40. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. CVPR (2023)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "S. Zhao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 435 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "41. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 182 + ], + "type": "text", + "content": "42. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 481, + 194 + ], + "type": "text", + "content": "43. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. ICLR (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 194, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 481, + 205 + ], + "type": "text", + "content": "44. Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. ICML (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 205, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 481, + 226 + ], + "type": "text", + "content": "45. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. ICLR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 226, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 481, + 247 + ], + "type": "text", + "content": "46. Sutskever, I., Vinyals, O., Le, Q.V.: Sequence to sequence learning with neural networks. NeurIPS (2014)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 248, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 481, + 281 + ], + "type": "text", + "content": "47. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 281, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 481, + 303 + ], + "type": "text", + "content": "48. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 303, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 481, + 335 + ], + "type": "text", + "content": "49. Wu, W., Li, Z., He, Y., Shou, M.Z., Shen, C., Cheng, L., Li, Y., Gao, T., Zhang, D., Wang, Z.: Paragraph-to-image generation with information-enriched diffusion model. arXiv preprint arXiv:2311.14284 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 335, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 481, + 357 + ], + "type": "text", + "content": "50. Xue, Z., Song, G., Guo, Q., Liu, B., Zong, Z., Liu, Y., Luo, P.: Raphael: Text-to-image generation via large mixture of diffusion paths. NeurIPS (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 357, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 481, + 390 + ], + "type": "text", + "content": "51. Zhang, H., Xu, T., Li, H., Zhang, S., Wang, X., Huang, X., Metaxas, D.N.: Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. ICCV (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 390, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 481, + 413 + ], + "type": "text", + "content": "52. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. ICCV (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 413, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 481, + 435 + ], + "type": "text", + "content": "53. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Unicontrolnet: All-in-one control to text-to-image diffusion models. NeurIPS (2024)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "LaVi-Bridge" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Bridging Synthetic and Real Worlds for Pre-training Scene Text Detectors/80dda151-f601-4dfe-a0ed-8536e8b997b9_content_list.json b/2024/Bridging Synthetic and Real Worlds for Pre-training Scene Text Detectors/80dda151-f601-4dfe-a0ed-8536e8b997b9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c37f9819bfab5eeb102d7581028cba424688368e --- /dev/null +++ b/2024/Bridging Synthetic and Real Worlds for Pre-training Scene Text Detectors/80dda151-f601-4dfe-a0ed-8536e8b997b9_content_list.json @@ -0,0 +1,1633 @@ +[ + { + "type": "text", + "text": "Bridging Synthetic and Real Worlds for Pre-training Scene Text Detectors", + "text_level": 1, + "bbox": [ + 274, + 140, + 728, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tongkun Guan $^{1}$ , Wei Shen $^{1(\\boxtimes)}$ , Xue Yang $^{2}$ , Xuehui Wang $^{1}$ , and Xiaokang Yang $^{1}$", + "bbox": [ + 341, + 210, + 658, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University {gtk0615,wei.shen}@sjtu.edu.cn", + "bbox": [ + 215, + 253, + 785, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Shanghai AI Laboratory yangxue@pjlab.org.cn", + "bbox": [ + 411, + 282, + 589, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Existing scene text detection methods typically rely on extensive real data for training. Due to the lack of annotated real images, recent works have attempted to exploit large-scale labeled synthetic data (LSD) for pre-training text detectors. However, a synth-to-real domain gap emerges, further limiting the performance of text detectors. Differently, in this work, we propose FreeReal, a real-domain-aligned pretraining paradigm that enables the complementary strengths of both LSD and unlabeled real data (URD). Specifically, to bridge real and synthetic worlds for pre-training, a glyph-based mixing mechanism (GlyphMix) is tailored for text images. GlyphMix delineates the character structures of synthetic images and embeds them as graffiti-like units onto real images. Without introducing real domain drift, GlyphMix freely yields real-world images with partial annotations derived from synthetic labels. Furthermore, when given free fine-grained synthetic labels, GlyphMix can effectively bridge the linguistic domain gap stemming from English-dominated LSD to URD in various languages. Without bells and whistles, FreeReal achieves average gains of $1.97\\%$ , $3.90\\%$ , $3.85\\%$ , and $4.56\\%$ in improving the performance of FCENet, PSENet, PANet, and DBNet methods, respectively, consistently outperforming previous pre-training methods by a substantial margin across four public datasets. Code will be available at https://github.com/SJTU-DeepVisionLab/FreeReal.", + "bbox": [ + 261, + 348, + 738, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Scene text detection $\\cdot$ Pre-training paradigm $\\cdot$ Domain-bridging method", + "bbox": [ + 259, + 652, + 740, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 707, + 375, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scene text detection aims to predict the bounding box or polygon for each arbitrarily shaped text instance within an image. This technology has diverse applications, including online education, instant translation, and product search, owing to the prevalence of text-laden images. However, despite their considerable progress, supervised methods [8,10,22,46,50,68] still struggle with an increasing", + "bbox": [ + 212, + 739, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author.", + "bbox": [ + 228, + 823, + 397, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9d8f019c33ab1a50b0a35f02d6e7fe250242854abac78561c91effbcdec1003a.jpg", + "image_caption": [ + "Fig. 1: We present different pre-training paradigms for enhancing text detectors." + ], + "image_footnote": [], + "bbox": [ + 276, + 145, + 727, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "demand for real data with labor-intensive annotations, especially when detecting complicated text images.", + "bbox": [ + 212, + 332, + 784, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, recent research efforts [39, 43, 54] have focused on pre-training scene text detectors using large-scale labeled synthetic data (LSD), which leverage both visual and textual knowledge to extract text feature representations, as illustrated in Fig. 1(a). For example, VLPT [39] formulates three cross-modality pretext tasks to enhance the performance of the backbone, and oCLIP [54] focuses on aligning visual and textual information to assist text detectors. Nevertheless, these well-curated designs are limited to using synthetic images due to their reliance on labeled word transcriptions. This limitation will lead to a noticeable synth-to-real domain gap when models are fine-tuned on real data, negatively impacting text detection performance. Therefore, exploring the potential of unlabeled real data (URD) is of great importance as they are readily available. However, the pre-training on URD has seldom been studied. One of the few works [45] attempts to pre-train text detectors on URD via masked image modeling [13] (step 1) and on LSD via supervised learning (step 2), which will re-transfers pre-trained weights from the real domain to the synthetic domain.", + "bbox": [ + 212, + 366, + 787, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Differently, our objective is to jointly leverage LSD and URD by bridging their gap for enhancing the text detector pre-training, as shown in Fig. 1(b). The primary challenge is how to effectively modulate these URD and LSD without encountering real domain drift. In light of this, we initially investigate the adaptation capabilities of existing domain-bridging methods [32, 62, 65]. This involves training a classifier to distinguish between URD and LSD and subsequently assessing the confidence of these domain-bridging methods in aligning with the real domain. As shown in Fig. 2(b), their highest score is only $78.9\\%$ . The observation implies that these domain-bridging methods may introduce intermediate domains and disrupt the real domain.", + "bbox": [ + 212, + 595, + 787, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To conquer this domain drift issue, we propose FreeReal, a real-domain-aligned pre-training paradigm, leveraging synthetic labels to guide text detector learning on real-domain-aligned images. Specifically, a glyph-based mixing mechanism, named GlyphMix, is tailored for text images to create real-world images with annotations derived from synthetic labels. This involves delineating the glyphs of synthetic images based on their internal structure and embedding", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 227, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "T. Guan et al.", + "bbox": [ + 271, + 114, + 369, + 127 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1e4e3829745eba0bcbf9d190b1ee69de0e3c64de5186e462dc25dcd57a0715e8.jpg", + "image_caption": [ + "Fig. 2: We present the domain classification task and the performances of different domain bridging ways between LSD and URD. In the text detection area, experiments demonstrate that our domain bridging method significantly performs better in aligning with the real domain. This is easy to explain: (I) Mixup may lead to unexpected pixel-wise ambiguities; (II) CutMix may cause the absence of many text regions and incomplete semantics of texts; (III) ClassMix will introduce salient boundary priors; (IV) GlyphMix (ours) effectively preserves the semantic information of text without causing real domain drift and bringing extra boundary priors." + ], + "image_footnote": [], + "bbox": [ + 277, + 145, + 723, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "them as graffiti-like units onto any unlabeled real images. GlyphMix offers several advantages: 1) extracting glyph structures in a cost-effective way without introducing extra boundary priors; 2) leveraging synthetic labels for free while greatly minimizing real domain drift (achieving a confidence score of $95.7\\%$ ).", + "bbox": [ + 212, + 411, + 784, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As a byproduct, FreeReal can bridge the language-to-language gap, when adapting features learned from English-dominated LSD to URD in various languages. Specifically, given free character-level synthetic labels, FreeReal enables the pre-trained model to easily learn character features, since characters are the underlying unit of any language. Subsequently, during fine-tuning, these learned character features are adaptively combined according to the rules (e.g., word-and sentence-level) provided by available benchmarks.", + "bbox": [ + 212, + 474, + 787, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building upon these insights, we break the trend of recent pre-training methods for scene text detectors (PSTD) that combine increasingly complex pretext task designs or more training modules to improve performance. Without bells and whistles, we focus on exploring the potential of integrating the LSD and URD. Remarkably, despite using $90\\%$ less LSD than recent SOTAs, our method still attains a $1.2\\%$ gain, effectively underscoring the superiority of our pretraining paradigm. As shown in Table 1, FreeReal consistently surpasses recent SOTAs by a large margin. The main contributions are summarized as follows:", + "bbox": [ + 212, + 582, + 787, + 703 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We are the first to simultaneously integrate labeled synthetic data and unlabeled real data for enhancing the pre-training of text detectors. FreeReal shows its superiority and significant performance gains over other complex alternatives, serving as a simple yet strong baseline for future PSTD studies.", + "- We effectively address the challenges of synthetic-to-real and language-to-language domain adaptation. Without bringing boundary priors and causing real domain drift, we freely harness the strengths of real-world text images with annotations derived from synthetic labels." + ], + "bbox": [ + 225, + 719, + 782, + 838 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "FreeReal for Pre-training STDs", + "bbox": [ + 521, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 387, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scene Text Detection. Existing STD methods delineate text regions by regression-based or segmentation-based representations. The regression-based methods [8, 14,24,34,49,67,68] directly regress the bounding points to enclose text regions, and gradually increase the number of points as required to detect complex scenes, e.g., arbitrary texts. The segmentation-based methods regard texts of arbitrary shapes by grouping pixel-level [22, 23, 40, 48], character-level [1, 9], segment-level [26,37] or contour-level [35,44,50] segmentation results by post-processing.", + "bbox": [ + 212, + 175, + 810, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pre-training for Scene Text Detector. In early research, [12, 21, 27] proposed to pre-train scene text detectors by generating synthetic data. These data engine tools follow the same paradigm: they first stack multiple off-the-shelf models to perform semantic segmentation and depth estimation analysis on background images, and then blend pre-processed texts into these images. However, the multi-stage, pre-generated methods incur large computational costs (GPU and CPU resources) and substantial latency, making them unsuitable for online training tasks where immediate data processing is essential. Inspired by recent advancements in pre-training techniques [3, 4, 11, 36], the integration of text and visual cross-modal information into scene text detectors has garnered increasing attention. One representative work involves enhancing the backbone with text transcriptions. For example, STKM [43] proposes a pioneering work that pretrains the backbone through an image-level text recognition task. VLPT [39] formulates three cross-modality pretext tasks to boost better text representations derived from the backbone. Another representative work is to endow the entire detector with pre-trained vision-text knowledge. Inherited from CLIP [36], TCM incorporates the Image and Text encoders into the text detector to refine text regions using cross-modal visual-language priors. A persistent challenge still remains—the synth-to-real domain gap limits the detection performance.", + "bbox": [ + 212, + 281, + 785, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semi-Supervised Learning. Due to the scarcity of expensive annotations, semi-supervised learning (SSL) methods are becoming increasingly important. These methods extract knowledge from a reduced amount of annotated images in a supervised manner, and from a larger amount of unlabeled images in an unsupervised way. Commonly, a self-training framework for SSL is employed, which encourages a model (student) to learn unlabeled images from pseudo-labels generated by the up-to-date optimized model (teacher). To avoid overfitting incorrect pseudo-labels, recent advancements have focused on pseudo-label refinement by improving the reliability [38,64] and pseudo-label regularization by cross-view consistency [19, 20, 42, 52]. Specifically, some domain adaption works [5, 51, 63] for STD focus on minimizing domain discrepancies in the feature space, including self-training and adversarial alignment methods. The self-training methods adopt the source data to pre-train a model for generating pseudo labels on target data, then use them to retrain the model. Adversarial alignment approaches aim to align the feature distribution and learn the domain-invariant features between the source and target domains. Recently, they have been replaced by the student-teacher framework, which significantly reduces domain gaps through dual-branch knowledge transfer [56].", + "bbox": [ + 212, + 568, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "T. Guan et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/17e5610e84a7fe2b5752b87e9a748a27a728c9dad340779b7aa990968d215e99.jpg", + "image_caption": [ + "(a) FreeReal" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 483, + 248 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/15ce7f2392943d4f04eb7bee305daaea034e94cfca3b855001b19fd2cd547267.jpg", + "image_caption": [ + "(b) GlyphMix", + "Fig. 3: Our network pipeline in a student-teacher framework." + ], + "image_footnote": [], + "bbox": [ + 490, + 147, + 784, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 319, + 379, + 337 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overview. Given a labeled synthetic set $\\mathcal{D}_l = \\{(\\mathbf{X}_l^{(i)},\\mathcal{B}_l^{(i)})\\}_{i = 1}^{N_l}$ with $N_{l}$ samples, $\\mathbf{X}_l^{(i)}\\in \\mathbb{R}^{H\\times W\\times 3}$ represents an image and $\\mathcal{B}_l^{(i)} = \\{(p_{lt}^i,p_{lb}^i,p_{rt}^i,p_{rb}^i)\\}_{i = 1}^n$ is its corresponding set of bounding boxes. $p$ refers to the values of the vertice along the abscissa and ordinate, and $n$ denotes the number of bounding boxes. $H$ and $W$ are the height and width of the image. Similarly, the unlabeled real set is denoted as $\\mathcal{D}_u = \\{\\mathbf{X}_u^{(i)}\\}_{i = 1}^{N_u}$ with $N_{u}$ samples. Our objective is to jointly leverage labeled synthetic data $\\mathcal{D}_l$ and unlabeled real data $\\mathcal{D}_u$ by bridging their gap for enhancing the pre-training of scene text detectors.", + "bbox": [ + 212, + 354, + 787, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The student-teacher framework exhibits substantial efficacy due to its dual capacity for knowledge transfer between its two branches, which is typically employed to mitigate the domain gap (e.g., data and multimodal). Accordingly, we customize it for aligning real and synthetic text data, as illustrated in Fig. 3. For notation simplification in the following introduction, we omit the superscript $i$ as mentioned above.", + "bbox": [ + 212, + 483, + 787, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, the segmentation-based text detector DBNet [22] is employed as our baseline model, primarily due to its impressive real-time performance, versatility, and practical applicability. In the teacher branch, the teacher model parameterized by $\\mathbf{f}_t$ , is responsible for generating truncated-gradient logits $\\hat{\\mathbf{Y}}_u \\in \\mathbb{R}^{H \\times W}$ from a weakly augmented unlabeled real image $w(\\mathbf{X}_u)$ . The pseudo label $\\mathbf{Y}_u \\in \\{0,1\\}^{H \\times W}$ is then obtained by binarizing the logit map $\\hat{\\mathbf{Y}}_u$ . In the student branch, the process begins by generating a pixel-wise binary map $\\mathbf{Y}_l \\in \\{0,1\\}^{H \\times W}$ , where 1 is assigned to locations within all bounding boxes from annotation $\\mathcal{B}_l$ , and 0 otherwise. The labeled synthetic image $\\mathbf{X}_l$ and the strongly augmented unlabeled real image $s(\\mathbf{X}_u)$ are combined to create $\\mathbf{X}_g = \\mathsf{A}(s(\\mathbf{X}_u), \\mathbf{X}_l)$ , with $\\mathsf{A}(\\cdot)$ denoting our proposed GlyphMix mechanism, which will be introduced later. Subsequently, the mixed image $\\mathbf{X}_g$ serves as input to train the student model, under the supervision of mixed pseudo-labels $\\mathbf{Y}_g = \\mathsf{A}(\\mathbf{Y}_u, \\mathbf{Y}_l)$ , by minimizing loss in the following:", + "bbox": [ + 212, + 574, + 787, + 786 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\ell_ {D B} \\left(\\hat {\\mathbf {Y}} _ {g}, \\mathbf {Y} _ {g}, \\mathbf {E}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 787, + 784, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\ell_{DB}$ denotes the original loss function provided by DBNet [22] and $\\hat{\\mathbf{Y}}_g$ represents the prediction logits generated by the student model. The matrix", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "FreeReal for Pre-training STDs", + "bbox": [ + 521, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 782, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathbf{E} \\in \\mathbb{R}^{H \\times W}$ , which indicates positive regions for loss calculation, is set as an all-one matrix by default.", + "bbox": [ + 212, + 145, + 782, + 175 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To filter out the unreliable predictions in mixed pseudo-labels $\\mathbf{Y}_g$ , we adaptively adjust the matrix $\\mathbf{E}$ during pre-training. Specifically, for the labeled synthetic image $\\mathbf{X}_l$ , the positive regions $\\mathbf{E}_l$ are an all-one matrix. For the unlabeled real image $\\mathbf{X}_u$ , we compute the pixel-wise entropy map on $\\hat{\\mathbf{Y}}_u$ and then apply a dynamic entropy threshold $\\zeta$ to obtain its positive regions $\\mathbf{E}_u$ as follows:", + "bbox": [ + 212, + 176, + 784, + 252 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} _ {u} = \\mathbb {I} \\left(- \\hat {\\mathbf {Y}} _ {u} \\log \\hat {\\mathbf {Y}} _ {u} \\leq \\zeta\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 255, + 784, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbb{I}(\\cdot)$ denotes the indicator function. The matrix is subsequently formulated as $\\mathbf{E} = \\mathsf{A}(\\mathbf{E}_u,\\mathbf{E}_l)$ . In our experiments, the threshold $\\zeta$ is determined by the value corresponding to the top $\\gamma$ percent of the entropy map. As the reliability of pseudo-label gradually improves, we smoothly update $\\gamma$ using np.linspace(80, 20, t), where $t$ denotes the total number of training iterations.", + "bbox": [ + 212, + 273, + 784, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, the teacher model $\\mathbf{f}_t$ gradually updates itself via the exponential moving averaging (EMA) of the student weights $\\mathbf{f}_s$ as follows:", + "bbox": [ + 212, + 349, + 782, + 378 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {t} = \\alpha \\mathbf {f} _ {t} + (1 - \\alpha) \\mathbf {f} _ {s}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 382, + 784, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha$ is a momentum parameter, set as 0.996 by default.", + "bbox": [ + 214, + 400, + 643, + 415 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 GlyphMix", + "text_level": 1, + "bbox": [ + 214, + 440, + 346, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Although some efforts [32, 62, 65] attempt to aggregate the advantages of synthetic and unlabeled images through versatile transformations, they inevitably introduce intermediate domains, leading to real domain drift. In this section, we present a domain-bridging way tailored for text images, allowing for the free use of synthetic labels while ensuring consistent alignment with the real domain.", + "bbox": [ + 212, + 468, + 784, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1) Graffiti-like Inter-domain Mixing. Given an input synthetic image $\\mathbf{X}_l$ , its corresponding bounding box annotations $\\mathcal{B}_l$ and converted pixel-wise binary map $\\mathbf{Y}_l$ , our goal is to perform character segmentation. This process identifies all character structure regions and produces a mask $\\mathbf{M}_G$ .", + "bbox": [ + 212, + 544, + 784, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To achieve this, we begin with an unsupervised clustering task based on the grayscale values of pixels, avoiding the introduction of extra pixel-level annotation costs. This involves aligning and cropping the image according to the coordinates of each bounding box, yielding a set of image patches $\\mathcal{T} = \\{\\mathbf{T}^{(i)}|\\mathbf{T}^{(i)}\\in \\mathbb{R}^{w_i\\times h_i\\times 3}\\}_{i = 1}^n$ . Subsequently, within each image patch $\\mathbf{T}^{(i)}$ , we cluster the pixels into two regions: the glyph region representing character structures and the background region. For simplicity, we focus our study on K-means to implement the clustering task, but other clustering approaches with predefined categories can be used. Following this, the glyph regions from all image patches are pasted to an all-zero image to create the character structure mask $\\mathbf{M}_G$ .", + "bbox": [ + 212, + 606, + 785, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the pseudo label $\\mathbf{Y}_u$ produced by the teacher model on an unlabeled real image $\\mathbf{X}_u$ , we use $\\mathbf{M}_G$ to generate the inter-domain mixing image and its corresponding label. This process can be formulated as follows:", + "bbox": [ + 212, + 756, + 782, + 801 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\mathbb {A} ^ {G} \\left(\\mathbf {X} _ {u}, \\mathbf {X} _ {l}\\right) \\leftarrow \\mathbf {M} _ {G} \\odot \\mathbf {X} _ {l} + \\left(\\mathbf {1} - \\mathbf {M} _ {G}\\right) \\odot \\mathbf {X} _ {u}, \\\\ \\mathbb {A} ^ {G} \\left(\\mathbf {Y} _ {u}, \\mathbf {Y} _ {l}\\right) \\leftarrow \\mathbf {M} _ {G} \\odot \\mathbf {Y} _ {l} + \\left(\\mathbf {1} - \\mathbf {M} _ {G}\\right) \\odot \\mathbf {Y} _ {u}, \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 804, + 784, + 843 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 215, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "T. Guan et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{1}$ is an all-one matrix and $\\odot$ denotes the element-wise multiplication operation. $\\mathbf{M}_G$ indicates pixels occupied by the glyph need to be copied from the synth domain and pasted to the real domain. Thus the mixed image closely aligns with the real domain, taking advantage of the freely available synthetic labels. Put simply, this process produces real-world images with partial annotations derived from synthetic labels.", + "bbox": [ + 212, + 146, + 782, + 234 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "2) Text-balanced Intra-domain Mixing. Real-world text images inherently exhibit complexity, and the collected unlabeled data often struggles to comprehensively represent the full spectrum of real-world scenarios. Within a student-teacher framework, intra-domain mixing serves as an effective perturbation technique, enriching the diversity of real data and enhancing feature consistency and robustness. However, when randomly generating and applying binary masks to construct intra-domain mixing images and their corresponding labels, there is a risk of missing many text regions, especially considering that text regions typically occupy only a small fraction of an image. To mitigate this challenge, our objective is to retain more text regions while still maintaining randomness.", + "bbox": [ + 212, + 237, + 784, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In light of this, we initiate the process by randomly selecting two unlabeled real images, denoted as $(\\mathbf{X}_u^{(1)},\\mathbf{X}_u^{(2)})$ , and obtaining the corresponding segmentation pseudo-labels, $(\\mathbf{Y}_u^{(1)},\\mathbf{Y}_u^{(2)})$ , produced by the teacher model. Subsequently, $k$ rectangular boxes $\\{(p_{lt}^i,p_{lb}^i,p_{rt}^i,p_{rb}^i)\\}_{i = 1}^k$ are randomly generated to crop the pseudo-label $\\mathbf{Y}_u^{(2)}$ based on their coordinates. This operation yields a set of local pseudo-label regions $S = \\{\\mathbf{S}^{(i)}|\\mathbf{S}^{(i)}\\in \\mathbb{R}^{w_i\\times h_i}\\}_{i = 1}^k$ . Following this, we identify the rectangular box that contains the maximum text regions by calculating $\\arg \\max_{i}\\sum_{j = 1}^{w_{i}\\times h_{i}}\\mathbf{S}_{j}^{(i)}$ , and convert it into a binary mask, denoted as $\\mathbf{M}_T$ . Finally, we employ $\\mathbf{M}_T$ to generate the intra-domain mixed image and its corresponding label, following the formulated procedure:", + "bbox": [ + 212, + 388, + 784, + 547 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\mathbf {A} ^ {T} (\\mathbf {X} _ {u}) \\leftarrow \\mathbf {M} _ {T} \\odot \\mathbf {X} _ {u} ^ {(2)} + \\left(\\mathbf {1} - \\mathbf {M} _ {T}\\right) \\odot \\mathbf {X} _ {u} ^ {(1)}, \\\\ \\mathbf {A} ^ {T} (\\mathbf {Y} _ {u}) \\leftarrow \\mathbf {M} _ {T} \\odot \\mathbf {Y} _ {u} ^ {(2)} + \\left(\\mathbf {1} - \\mathbf {M} _ {T}\\right) \\odot \\mathbf {Y} _ {u} ^ {(1)}, \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 551, + 784, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{1}$ is an all-one matrix and $\\odot$ represents the element-wise multiplication.", + "bbox": [ + 212, + 590, + 779, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3) Intra- and Inter-domain Mixing. When these two strategies, tailored for scene text images, are integrated within the student-teacher framework, they collectively form a comprehensive method that effectively harnesses the strengths of both synthetic and real data. The whole process can be formulated as:", + "bbox": [ + 212, + 607, + 782, + 666 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\mathbf {A} (\\mathbf {X}) \\leftarrow \\mathbf {A} ^ {G} \\left(\\mathbf {A} ^ {T} \\left(\\mathbf {X} _ {u}\\right), \\mathbf {X} _ {l}\\right), \\\\ \\mathbf {A} (\\mathbf {Y}) \\leftarrow \\mathbf {A} ^ {G} \\left(\\mathbf {A} ^ {T} \\left(\\mathbf {Y} _ {u}\\right), \\mathbf {Y} _ {l}\\right) \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 669, + 784, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Consequently, the synergy maximizes the utilization of synthetic labels as a valuable resource, effectively tackling the intricacies of real-world scenarios.", + "bbox": [ + 212, + 709, + 782, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Discussion on Character Region Awareness", + "text_level": 1, + "bbox": [ + 212, + 765, + 617, + 780 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Existing LSD only contains English texts with tightly spaced characters, while URD may encompass text properties from various languages (such as Japanese, Korean, Chinese, etc). Consequently, even though the model predictions effec", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "FreeReal for Pre-training STDs", + "bbox": [ + 521, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tively cover text regions, we encounter challenges in parsing these predictions into word-level bounding boxes or polygons.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Indeed, when using a single language (e.g., English) as the source domain and adapting it to various language domains, it appears that no matter how sophisticated the models and loss functions are, they struggle to accurately and effectively parse the complex text layout and semantic context. This challenge primarily arises from variations in the spacing and arrangement of characters, words, and sentences across languages. This observation leads us to revisit the learning objective of scene text detectors during pre-training.", + "bbox": [ + 212, + 176, + 785, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We emphasize an alternative perspective — utilizing characters as the fundamental learning unit for text detection. The strategy presents several advantages: 1) Like word-level bounding boxes (WBB), character-level bounding box (CBB) annotations can be readily acquired from the existing synthetic dataset (SynthText) at no additional cost; 2) Instead of parsing complex layouts and spacing disparities across languages during pre-training, decoding them into individual characters is straightforward and attainable; 3) The pre-trained model with character-level region awareness easily integrates its learned character features to adapt text attribute rules provided by different language text datasets during fine-tuning. To this end, we select CBB annotations as $\\mathcal{B}_l$ .", + "bbox": [ + 212, + 282, + 787, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 215, + 458, + 364, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Dataset", + "text_level": 1, + "bbox": [ + 215, + 491, + 326, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Labeled Synthetic Data (LSD). The SynthText [12] serves as our source of LSD, consisting of 800K images. Each image is provided with character-level and word-level bounding boxes, along with corresponding text transcriptions.", + "bbox": [ + 212, + 517, + 785, + 563 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unlabeled Real Data (URD). URD is collected from a wide array of natural scenes, encompassing a diverse range of fonts, languages, and text shapes. Specifically, URD includes 430K images from LSVT [41], 9K from MLT17 [30], 19657 from MLT19 [29], 229 from ICDAR2013 [18], and 500 from USTB-SV1K [59].", + "bbox": [ + 212, + 563, + 785, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "**Benchmarks.** Four publicly available benchmarks are used to evaluate the performance, including 1) Total-Text (TT) [6], 2) CTW 1500 (CTW) [25], 3) IC-DAR2015 (IC15) [17], and 4) MSRA-TD500 (TD) [57].", + "bbox": [ + 212, + 625, + 785, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 691, + 455, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Referring to recent works [39, 43, 54], we combine representative text detection methods [22, 46, 48, 68] to conduct experiments. URD is used in Tabs. 1-3 and the same data is used in Tabs. 4-7. All experiments are implemented on a server with 2 NVIDIA A800 GPUs, using PyTorch. 1) Pre-Training. The pre-training experiments are conducted on LSD and URD, utilizing the open-source codes provided by these text detection methods. For synthetic images, we adopt consistent augmentation settings as prescribed by these methods for a fair comparison. For unlabeled real images, we employ a standard weak-to-strong augmentation", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 227, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "T. Guan et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "technique [66] to enrich diversity. Our method is trained using the AdamW optimizer [28], along with a cosine learning rate scheduler [7]. This includes a warm-up phrase for 1 epoch in a total of 10 epochs. The base learning rate is set to $0.003 \\times$ batchsize/256. At each iteration, we equally sample LSD and URD with a total batch size of 24. 2) Fine-Tuning. For a fair comparison, we employ the same fine-tuning configurations as these detection methods and then evaluate them on these publicly available benchmarks: TT, CTW, IC15 and TD, following the standard metric: Precision (P), Recall (R), and F-measure (F).", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Experiment Results", + "text_level": 1, + "bbox": [ + 214, + 287, + 426, + 303 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To ensure a comprehensive comparison, we collect results from public evaluation benchmarks reported by existing pre-training methods, and organize them in Table 1. The sources of these results are clearly recorded in the table's caption using different symbols for easy reference. Specifically, these methods can be classified into three pre-training paradigms. In Paradigm A, models (i.e., $+ST$ , $+TCM$ ) are initially pre-trained on LSD and then fine-tuned on real datasets directly. In Paradigm B, models (i.e., $+STKM$ , $+VLPT$ , $+oCLIP$ ) are first pretrained on LSD by various pretext tasks, and the well-learned backbone weights are transferred into scene text detectors when fine-tuning on real datasets. In Paradigm C, models (i.e., $+MTM$ ) are pre-trained on URD by pretext tasks and on LSD by supervised learning, and then fine-tuned on real datasets.", + "bbox": [ + 212, + 311, + 787, + 476 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) Results on DBNet. As indicated in the second group of rows in Table 1, all pre-training methods both report their cooperative performance with DBNet. When compared to methods using paradigm A, differing from $DBNet + ST$ , we exclude the use of a deformable convolutional network (DCN) in our model. Despite these differences, our method still achieves impressive results, yielding improvements of $4.2\\%$ , $4.5\\%$ , $4.6\\%$ , and $5.2\\%$ on TT, CTW, IC15, and TD datasets, respectively. Compared to $DBNet + TCM$ methods, our method achieves SOTA detection performance on four datasets, particularly on curved text datasets (TT and CTW), with gains of $3.0\\%$ and $2.8\\%$ , respectively. In comparison with methods following Paradigm B, our method consistently outperforms previous SOTA works by substantial margins, achieving performance improvements of $2.6\\%$ , $5.9\\%$ , $3.5\\%$ and $1.6\\%$ on TT, CTW, IC15, and TD datasets, respectively. In comparison with $DBNet + MTM$ following Paradigm C, our method gets $3.2\\%$ and $4.5\\%$ gains on TT and TD datasets, respectively.", + "2) Results on PANet. As shown in the third group of rows in Table 1, our method outperforms $PANet + ST$ on four benchmarks, achieving a gain of $3.85\\%$ in terms of average F-measure, while using the same parameters and labeled data for pre-training and fine-tuning. Furthermore, our method exhibits its prominent superiority, particularly on the CTW and TD datasets, surpassing the $PANet + TCM$ method by a substantial margin of $4.1\\%$ and $4.6\\%$ , respectively.", + "3) Results on PSENet. FreeReal brings $0.7\\%$ , $2.8\\%$ , and $0.9\\%$ gains on TT, CTW, and IC15 datasets, respectively.", + "4) Results on FCENet. FreeReal further refreshes the best text detection results and improves $2.1\\%$ , $1.3\\%$ , $0.6\\%$ and $2.1\\%$ on TT, CTW, IC15, and TD" + ], + "bbox": [ + 212, + 477, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "FreeReal for Pre-training STDs", + "bbox": [ + 521, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/ae00a87c75e810eaf91477012409dc48f5020a3d0d84ad6ecdfb4ebbfbb9786f.jpg", + "table_caption": [ + "Table 1: Comparison with SOTA pre-training methods for text detectors on TT, CTW, IC15, and TD. F-measure (\\%) is reported. * denotes the use of extra labeled real data. For easy reference, we mark these symbols $\\dagger$ , $*$ , $\\ddagger$ , and $*$ separately to indicate the results from [61], [39], [54], and [45], within the scope of this paper. The best results are shown in bold font. The second-best results are underlined." + ], + "table_footnote": [], + "table_body": "
IndexMethodVenueTTCTWIC15TD
1PSENet-1s [46]CVPR'1980.982.285.7-
PAN++ [47]PAMI'2185.384.084.587.9
DBNet++ [23]PAMI'2286.085.387.387.2
DeepSOLO* [58]CVPR'2387.6-89.8-
ESTextSpotter* [15]ICCV'2390.0-91.089.5
SIR [33]ACMMM'2388.285.587.889.6
2DBNet+ST† [22]AAAI'2084.783.485.484.9
DBNet+STKM† [43]CVPR'2185.5-86.185.9
DBNet+VLPT† [39]CVPR'2286.3-86.588.5
DBNet+oCLIP† [54]ECCV'2284.182.085.4-
DBNet+TCM† [60]CVPR'2385.985.189.488.8
DBNet+MTM* [45]ACMMM'2385.7--85.6
DBNet+FreeReal (ours)-88.9(+2.6)87.9(+2.8)90.0(+0.6)90.1(+1.3)
3PANet+ST† [48]ICCV'1985.083.782.984.1
PANet+TCM† [60]CVPR'23-84.384.685.3
PANet+FreeReal (ours)-87.4(+2.4)88.4(+4.1)86.0(+1.4)89.9(+4.6)
4PSENet+ST* [46]CVPR'1983.981.681.3-
PSENet+STKM* [43]CVPR'2184.382.983.7-
PSENet+VLPT* [39]CVPR'2286.183.384.3-
PSENet+oCLIP† [54]ECCV'2285.582.8--
PSENet+FreeReal (ours)-86.8(+0.7)86.1(+2.8)85.2(+0.9)86.0
5FCENet+ST† [68]CVPR'2185.885.586.285.4
FCENet+oCLIP† [54]ECCV'22-85.686.7-
FCENet+TCM† [60]CVPR'23-85.987.186.9
FCENet+FreeReal (ours)-87.9(+2.1)87.2(+1.3)87.7(+0.6)89.0(+2.1)
", + "bbox": [ + 236, + 226, + 759, + 525 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "datasets, respectively. These experimental results demonstrate the robustness of FreeReal, even when cooperated with regression-based methods.", + "bbox": [ + 212, + 573, + 787, + 603 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5) Paradigm Comparison. In this section, we aim to investigate which kinds of pre-training paradigms for text detectors are important and meaningful for future studies. To this end, we conduct a comparative analysis with three mainstream pre-training paradigms mentioned above, as summarized in Table 2. When compared to these well-curated designs at the cost of introducing additional auxiliary pretext tasks or more training modules, our proposed paradigm achieves a significant improvement of $1.2\\%$ (87.5 vs. 86.3), despite utilizing $10\\%$ of the labeled synthetic data and $10\\%$ of the unlabeled real data. This outcome underscores our ability to harness the potential of unlabeled real images. Moreover, when we gradually increase the proportion of the total training data, the text detection performance continues to improve. Specifically, our method achieves gains of $1.8\\%$ , $2.0\\%$ , and $2.6\\%$ at data fractions of $20\\%$ , $50\\%$ , and $100\\%$ , respectively, compared to the recent SOTA method. This outcome fully demonstrates the superiority of our pre-training paradigm, which jointly leverage LSD and URD by bridging their gap for enhancing the text detector pre-training.", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "T. Guan et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/8931c343d196f89063d5238c473a3cf8bc7869482f3fb1b3d41df91f0ab7077f.jpg", + "table_caption": [ + "Table 2: Comparison with existing pre-training paradigms when fine-tuning on the TotalText dataset. These symbols $\\dagger$ , $\\ddagger$ , and $*$ indicate the results from [61], [54] and [45], respectively. $\\Delta F$ is the improvement relative to the baseline." + ], + "table_footnote": [], + "table_body": "
ParadigmMethodVenueDataF-measureΔ F(%)
LSDURD
-DBNet‡ [22]AA AI'20--78.2-
ADBNet+ST† [22]AA AI'20800K-84.7+6.5%
DBNet+TCM† [60]CVPR'23800K-85.9+7.7%
BDBNet+STKM† [43]CVPR'21800K-85.5+7.3%
DBNet+VLPT† [39]CVPR'22800K-86.3+8.1%
DBNet+oCLIP‡ [54]ECCV'22800K-84.1+5.9%
CDBNet+MTM* [45]ACM MM'23800K41K85.7+7.5%
DOurs (9%)-71K41K87.4+9.2%
Ours (10%)-80K46K87.5+9.3%
Ours (20%)-160K92K88.1+9.9%
Ours (50%)-400K230K88.3+10.1%
Ours (100%)-800K460K88.9+10.7%
", + "bbox": [ + 264, + 199, + 733, + 364 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "7) Results on Unseen Language Text. Our pre-trained model is capable of learning the basic character feature representations of languages and then combining them during fine-tuning. To show its robustness, we choose to evaluate on unseen language text dataset VinText [31]. FreeReal achieves an improvement of $0.7\\%$ , as shown in Table 3.", + "bbox": [ + 215, + 395, + 784, + 469 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5 Ablations and analysis", + "text_level": 1, + "bbox": [ + 217, + 498, + 467, + 515 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Unless otherwise stated, all ablation experiments are implemented in collaboration with DBNet method [22] using $20\\%$ of LSD and $20\\%$ of URD and evaluated on the curved TotalText (TT) dataset for efficiency.", + "bbox": [ + 215, + 535, + 784, + 580 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Effectiveness of Network Components. To comprehensively evaluate the effectiveness of our proposed method, we divide it into three components and evaluate each one individually, as shown in Table 4. \"Stu-Tea\", \"CRA\", \"GlyphMix\" denote the standard student-to-teacher framework, Character Region Awareness introduced in Sec.3.2 to bridge the linguistic domain gap, and the glyph-based mixing mechanism, respectively. First, we enhance the baseline DBNet model by incorporating a standard student-teacher framework, yielding a performance gain of $1.31\\%$ . Second, within the student-teacher framework, we evaluate the individual contributions of the CRA and GlyphMix components, achieving $1.73\\%$ and $3.12\\%$ performance improvements, respectively. Finally, upon combining all plug-ins, the overall performance reaches $88.1\\%$ , showing a substantial gain of $3.40\\%$ compared to the baseline. These experiments consistently demonstrate that FreeReal effectively bridges both the synth-to-real and language-to-language domain gaps when leveraging the intrinsic qualities of unlabeled real data.", + "bbox": [ + 215, + 582, + 785, + 792 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Discussion on Using WBB or CBB. The SynthText dataset offers word-level and character-level annotations, with the latter being exploited in our method to bridge the linguistic domain gap, bringing a $0.4\\%$ gain as shown in the third row", + "bbox": [ + 215, + 794, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "FreeReal for Pre-training STDs", + "bbox": [ + 521, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/80aadfe79a831300963621c93bd0c1c5054d6151a76861fc738166e05656d5a3.jpg", + "table_caption": [ + "Table 3: Comparison with supervised SOTA work on unseen language text. \"R50\" and \"DeT\" denote a ResNet50 and Deformable Transformer structure, respectively. * denotes the use of extra labeled real data." + ], + "table_footnote": [], + "table_body": "
MethodVenueBackbonePara.FPSF
ESTextSpotter*ICCV'23R50+DeT56M3.489.6
DBNet+FreeReal (ours)-R5025M1190.3
", + "bbox": [ + 313, + 199, + 684, + 241 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/0941c58a0156c2be48dcaba74891d895b3751781e37ccd214dc99b621c506693.jpg", + "table_caption": [ + "Table 4: Evaluate the effectiveness of the proposed modules. $\\Delta F$ is the improvement of F-measure relative to baseline." + ], + "table_footnote": [], + "table_body": "
Stu-TeaCRAGlyphMixΔ F(%)PRF
×××-87.182.584.7
✓××+1.31%88.8083.3986.01
✓✓×+1.73%87.7685.1586.43
✓×✓+3.12%89.1086.5887.82
✓✓✓+3.40%88.8187.4088.10
", + "bbox": [ + 315, + 297, + 683, + 380 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "of Table 4. Notably, even without the use of CRA, our method still achieves an F-measure of $87.82\\%$ (in the fourth row), significantly surpassing recent SOTAs.", + "bbox": [ + 212, + 411, + 785, + 441 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison with pre-training methods using URD. Bridging the synth-to-real gap to simultaneously leverage LSD and URD for pre-training is our major contribution. This capability, not present in other PSTD methods, enables seamless incorporation of the huge number of URD to enhance pre-training. Thus to compare to other PSTD methods with complex pretext task designs, we use different ratios of URD to explore the effectiveness of our pre-training paradigm. Even when using $10\\%$ LSD and $10\\%$ URD in Table 2, our method still achieves SOTA with a $1.2\\%$ gain, which exhibits our method can serve as a novel baseline for PSTD. Besides, we add the comparison using additional URD to train other PSTD methods. In this case, these mentioned methods can be viewed the same as MTM [45]. As a result, FreeReal outperforms MTM by $1.7\\%$ in Table 2.", + "bbox": [ + 212, + 443, + 787, + 609 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effectiveness of GlyphMix Components. In Table 5, GlyphMix comprises two modules, inter-domain GIM and intra-domain TIM, contributing gains of $1.28\\%$ and $1.22\\%$ to the final performance, respectively. Besides, when compared to the data engine tool [12], GIM achieves intra-domain mixing in a cost-effective way, which yields real-world images online via K-means, a more flexible, faster and direct process than the multi-stage, pre-generated method. Specifically, GIM attains a $10.7\\mathrm{x}$ speedup (2.6h vs 27.7h) in generating 92K images, which is more suitable for real-time online training tasks. Additionally, GIM can avoid improper text blends on URD by leveraging the learned knowledge (boxes/maps) from our pre-training framework online. We also conduct the ablations and get $0.5\\%$ $(85.3\\%)$ vs $84.8\\%$ ) performance gains, when using $5\\%$ training data.", + "bbox": [ + 212, + 611, + 787, + 777 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison with Bridging-domain Methods. The goal of generating mixed images through domain-bridging methods is to bring synthetic images closer to real images. Consequently, the adaptation capability can be viewed as the likelihood of mixed images belonging to the real domain. A classification task is em", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "T. Guan et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/cfd035502271b86bd9fe924d4c9ba125c44283c33f635812246b7448ba8ec475.jpg", + "table_caption": [ + "Table 5: Evaluate the effectiveness of GlyphMix components. \"GIM\" and \"TIM\" denote the Graffiti-like Inter-domain Mixing and Text-balanced Intra-domain Mixing." + ], + "table_footnote": [], + "table_body": "
GIMTIMΔ F(%)PRF
××-87.7685.1586.43
×✓+1.22%88.6486.6887.65
✓×+1.28%88.2987.1387.71
", + "bbox": [ + 352, + 184, + 645, + 244 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/c435c040338b80ee89d256b36df7176ceb78c35dc726ef04bd298995da558379.jpg", + "table_caption": [ + "Table 6: Comparison of the effectiveness with existing domain bridging methods. \"DCA\" refers to Domain Classification Accuracy in aligning with the real domain." + ], + "table_footnote": [], + "table_body": "
methodDCAPRFΔ F(%)
Stu-Tea-88.8083.3986.01-
Mix up [65]69.589.1784.1286.57+0.56%
CutMix [62]72.887.9985.6486.80+0.79%
ClassMix [32]78.990.6583.6186.99+0.98%
FreeReal (ours)95.788.8187.4088.10+2.09%
", + "bbox": [ + 328, + 301, + 669, + 381 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/2f8418d03f6e0cf66911d6449bd6394da9ddf811846c19bd5488147895ae8152.jpg", + "table_caption": [ + "Table 7: Comparison of the effectiveness with SSL methods." + ], + "table_footnote": [], + "table_body": "
MethodVenuePRFΔ F(%)
Stu-Tea-88.8083.3986.01-
GTA [16]NeurIPS'2287.1881.0884.02-1.99%
BCP [2]CVPR'2386.7285.8286.27+0.26%
AugSeg [66]CVPR'2388.4885.6587.04+1.03%
Unimatch [55]CVPR'2388.9485.5687.22+1.21%
FreeReal (ours)-88.8187.4088.10+2.09%
", + "bbox": [ + 316, + 422, + 679, + 515 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "ployed to assess the capability, whose the evaluation protocol is defined: Given a well-learned domain classifier $\\phi$ (trained by $n$ synthetic images $x_{l} \\in X_{L}$ and $n$ real images $x_{u} \\in X_{U}$ ) and domain bridging method $\\xi$ , the classifier predicts the domain label (0 for synthetic and 1 for real) of mixed images and evaluates the accuracy as $\\frac{\\sum_{x_{l} \\in X_{L}, x_{u} \\in X_{U}} \\phi(\\xi(x_{l}, x_{u}))}{n^{2}}$ . To accomplish this, we employ a network architecture, consisting of resnet18 as a feature extractor, followed by a binary classification head, for simplicity. Subsequently, the performance of different domain bridging methods is summarized in Table 6. Remarkably, when compared to these methods, GlyphMix achieves an outstanding performance score of $95.7\\%$ , surpassing the second-best method by a substantial margin of $16.8\\%$ . This result demonstrates that our method is able to generate arbitrarily realistic images without domain drift.", + "bbox": [ + 215, + 546, + 785, + 731 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To be more convincing, we also conduct text detection experiments to assess the effectiveness. As depicted in Table 6, our method consistently outperforms others, achieving the best F-measure value. Furthermore, the observed positive correlation between DCA and F-measure reinforces the idea that bridging the domain gap is the crucial factor in improving text detection performance.", + "bbox": [ + 215, + 733, + 785, + 808 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Comparison with Semi-supervised Methods. In this section, we compared the performance with recent SSL methods. These methods tend to adopt in-", + "bbox": [ + 215, + 809, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "FreeReal for Pre-training STDs", + "bbox": [ + 521, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "creasingly complex designs, involving advanced contrastive learning techniques, additional network components, multiple ensemble models, and complex training procedures. In contrast, we aim to eliminate domain gaps between LSD and URD. Despite adopting the standard student-teacher framework, our method still exceeds the previous SOTA by $0.88\\%$ in Table 7.", + "bbox": [ + 212, + 146, + 782, + 219 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effectiveness of $\\mathbf{M}_G$ . We evaluate the accuracy of character segmentation results $\\mathbf{M}_G$ using TotalText [6] and TextSeg [53] datasets, both of which provide pixel-level annotations. Specifically, compared to conducting the cluster task on the whole image, we crop it into several text instances and then perform clustering to form $\\mathbf{M}_G$ . While the accuracy of the former is $32.35\\%$ for TotalText and $45.54\\%$ for TextSeg, $\\mathbf{M}_G$ achieves $80.43\\%$ and $84.14\\%$ , respectively. Notably, the underlying morphological representations of glyphs exhibit relative invariance to slight structural changes, such as thicker or thinner. This reduces the dependence on pixel-level high-precision segmentation with expensive costs. Furthermore, although the clustering process introduces some noise, leading to inaccuracies in $\\mathbf{M}_G$ , it can still be viewed as a way of adding data noise in the image-mixing process, enhancing the robustness of text detection.", + "bbox": [ + 212, + 220, + 782, + 402 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ablation Study on Hyper-parameters. We study the impact of the parameter $\\gamma$ on the pseudos generated by the teacher model during pre-training. This parameter controls the proportion of unreliable and reliable pixels. With $\\gamma$ set at 80, 50, and 20, the F-measure are $87.94\\%$ , $87.68\\%$ , and $87.06\\%$ , respectively. Finally, we employ a linear warm-up strategy for $\\gamma$ that gradually reduces from 80 to 20, yielding the best result with an F-measure of $88.10\\%$ .", + "bbox": [ + 212, + 402, + 782, + 492 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 517, + 357, + 534 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We propose an effective pre-training paradigm for scene text detection, situated within a standard student-teacher framework. The paradigm leverages the complementary strengths of both LSD and URD, standing out from existing intricate pre-training paradigms in text detection due to its simplicity and significant advancements. Moreover, we identify and address prominent challenges, i.e., synth-to-real and language-to-language domain gaps. The proposed solutions, namely GlyphMix and character region awareness, have shown their effectiveness in mitigating these challenges and improving text detection performance, as substantiated by extensive experimental results. More importantly, our method exhibits the ability to harness the intrinsic attributes of URD, achieving substantial performance improvements even with a small fraction of LSD. This highlights the significance of exploring the potential of unlabeled real data in scene text detection, ultimately advancing the field and its practical applications.", + "bbox": [ + 212, + 547, + 782, + 743 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgment. This work was supported by the NSFC under Grant 62322604 and 62176159, and in part by the Shanghai Municipal Science, Technology Major Project under Grant 2021SHZDZX0102, and National Key R&D Program of China (NO. 2022ZD0160100). The computations in this paper were run on the AI for Science Platform supported by the Artificial Intelligence Institute at Shanghai Jiao Tong University.", + "bbox": [ + 212, + 744, + 792, + 835 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "T. Guan et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Baek, Y., Lee, B., Han, D., Yun, S., Lee, H.: Character region awareness for text detection. In: CVPR. pp. 9365-9374 (2019)", + "2. Bai, Y., Chen, D., Li, Q., Shen, W., Wang, Y.: Bidirectional copy-paste for semi-supervised medical image segmentation. In: CVPR. pp. 11514-11524 (2023)", + "3. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV. pp. 9650-9660 (2021)", + "4. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML. pp. 1597-1607 (2020)", + "5. Chen, Y., Wang, W., Zhou, Y., Yang, F., Yang, D., Wang, W.: Self-training for domain adaptive scene text detection. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 850-857. IEEE (2021)", + "6. Ch'ng, C.K., Chan, C.S., Liu, C.L.: Total-text: toward orientation robustness in scene text detection. IJDAR 23(1), 31-52 (2020)", + "7. Grill, J., Strub, F., Altché, F., et al.: Bootstrap your own latent - A new approach to self-supervised learning. In: NeurIPS. pp. 21271-21284 (2020)", + "8. Guan, T., Gu, C., Lu, C., Tu, J., Feng, Q., Wu, K., Guan, X.: Industrial scene text detection with refined feature-attentive network. IEEE TCSVT 32(9), 6073-6085 (2022)", + "9. Guan, T., Gu, C., Tu, J., Yang, X., Feng, Q., Zhao, Y., Shen, W.: Self-supervised implicit glyph attention for text recognition. In: CVPR. pp. 15285-15294 (2023)", + "10. Guan, T., Lin, C., Shen, W., Yang, X.: Posformer: Recognizing complex handwritten mathematical expression with position forest transformer. arXiv preprint arXiv:2407.07764 (2024)", + "11. Guan, T., Shen, W., Yang, X., Feng, Q., Jiang, Z., Yang, X.: Self-supervised character-to-character distillation for text recognition. In: ICCV. pp. 19473-19484 (2023)", + "12. Gupta, A., Vedaldi, A., Zisserman, A.: Synthetic data for text localisation in natural images. In: CVPR. pp. 2315-2324 (2016)", + "13. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR. pp. 16000-16009 (2022)", + "14. He, M., Liao, M., Yang, Z., Zhong, H., Tang, J., Cheng, W., Yao, C., Wang, Y., Bai, X.: Most: A multi-oriented scene text detector with localization refinement. In: CVPR. pp. 8813-8822 (2021)", + "15. Huang, M., Zhang, J., Peng, D., Lu, H., Huang, C., Liu, Y., Bai, X., Jin, L.: Estextspotter: Towards better scene text spotting with explicit synergy in transformer. In: ICCV. pp. 19495-19505 (2023)", + "16. Jin, Y., Wang, J., Lin, D.: Semi-supervised semantic segmentation via gentle teaching assistant. In: NeurIPS. vol. 35, pp. 2803-2816 (2022)", + "17. Karatzas, D., Gomez-Bigorda, L., Nicolaou, A., et al.: Icdar 2015 competition on robust reading. In: ICDAR. pp. 1156-1160 (2015)", + "18. Karatzas, D., Shafait, F., Uchida, S., et al.: Icdar 2013 robust reading competition. In: ICDAR. pp. 1484-1493 (2013)", + "19. Laine, S., Aila, T.: Temporal ensembling for semi-supervised learning. In: arXiv:1610.02242 (2016)", + "20. Li, W., Gao, H.a., Gao, M., Tian, B., Zhi, R., Zhao, H.: Training-free model merging for multi-target domain adaptation. In: ECCV. Springer (2024)" + ], + "bbox": [ + 217, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "FreeReal for Pre-training STDs", + "bbox": [ + 521, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "21. Liao, M., Song, B., Long, S., He, M., Yao, C., Bai, X.: Synthtext3d: synthesizing scene text images from 3d virtual worlds. Science China Information Sciences 63, 1-14 (2020)", + "22. Liao, M., Wan, Z., Yao, C., Chen, K., Bai, X.: Real-time scene text detection with differentiable binarization. In: AAAI. pp. 11474-11481 (2020)", + "23. Liao, M., Zou, Z., Wan, Z., Yao, C., Bai, X.: Real-time scene text detection with differentiable binarization and adaptive scale fusion. IEEE TPAMI 45(1), 919-931 (2022)", + "24. Liu, Y., Chen, H., Shen, C., He, T., Jin, L., Wang, L.: Abcnet: Real-time scene text spotting with adaptiveBezier-curve network. In: CVPR. pp. 9809-9818 (2020)", + "25. Liu, Y., Jin, L., Zhang, S., Luo, C., Zhang, S.: Curved scene text detection via transverse and longitudinal sequence connection. PR 90, 337-345 (2019)", + "26. Long, S., Ruan, J., Zhang, W., He, X., Wu, W., Yao, C.: Textsnake: A flexible representation for detecting text of arbitrary shapes. In: ECCV. pp. 20-36 (2018)", + "27. Long, S., Yao, C.: Unrealtext: Synthesizing realistic scene text images from the unreal world. arXiv:2003.10608 (2020)", + "28. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv:1711.05101 (2017)", + "29. Nayef, N., Patel, Y., Busta, M., et al.: Icdar2019 robust reading challenge on multilingual scene text detection and recognition—rrc-plt-2019. In: ICDAR. pp. 1582-1587 (2019)", + "30. Nayef, N., Yin, F., Bizid, I., et al.: Icdar2017 robust reading challenge on multilingual scene text detection and script identification-rrc-mlt. In: ICDAR. vol. 1, pp. 1454-1459 (2017)", + "31. Nguyen, N., Nguyen, T., Tran, V., et al.: Dictionary-guided scene text recognition. In: CVPR. pp. 7383-7392 (2021)", + "32. Olsson, V., Tranheden, W., Pinto, J., Svensson, L.: Classmix: Segmentation-based data augmentation for semi-supervised learning. In: WACV. pp. 1369-1378 (2021)", + "33. Qin, X., Lyu, P., Zhang, C., Zhou, Y., Yao, K., Zhang, P., Lin, H., Wang, W.: Towards robust real-time scene text detection: From semantic to instance representation learning. In: ACM MM. pp. 2025-2034 (2023)", + "34. Qin, X., Zhou, Y., Guo, Y., Wu, D., Tian, Z., Jiang, N., Wang, H., Wang, W.: Mask is all you need: Rethinking mask r-cnn for dense and arbitrary-shaped scene text detection. In: ACM MM. pp. 414-423 (2021)", + "35. Qin, X., Zhou, Y., Guo, Y., Wu, D., Wang, W.: Fc 2 rn: a fully convolutional corner refinement network for accurate multi-oriented scene text detection. In: ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 4350-4354. IEEE (2021)", + "36. Radford, A., Kim, J.W., Hallacy, C., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021)", + "37. Shi, B., Bai, X., Belongie, S.: Detecting oriented text in natural images by linking segments. In: CVPR. pp. 2550-2558 (2017)", + "38. Sohn, K., Berthelot, D., Carlini, N., Zhang, Z., Zhang, H., Raffel, C.A., Cubuk, E.D., Kurakin, A., Li, C.L.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In: NeurIPS. vol. 33, pp. 596-608 (2020)", + "39. Song, S., Wan, J., Yang, Z., Tang, J., Cheng, W., Bai, X., Yao, C.: Vision-language pre-training for boosting scene text detectors. In: CVPR. pp. 15681-15691 (2022)", + "40. Sun, X., Lyu, J., Zhang, Y., Zeng, G., Fang, B., Zhou, Y., Xie, E., Ma, C.: Feature enhancement with text-specific region contrast for scene text detection. In: Chinese Conference on Pattern Recognition and Computer Vision (PRCV). pp. 3-14 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "T. Guan et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "41. Sun, Y., Ni, Z., Chng, C.K., et al.: Icdar 2019 competition on large-scale street view text with partial labeling-rrc-lsvt. In: ICDAR. pp. 1557-1562 (2019)", + "42. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In: NeurIPS. vol. 30 (2017)", + "43. Wan, Q., Ji, H., Shen, L.: Self-attention based text knowledge mining for text detection. In: CVPR. pp. 5983-5992 (2021)", + "44. Wang, F., Chen, Y., Wu, F., Li, X.: Texray: Contour-based geometric modeling for arbitrary-shaped scene text detection. In: ACM MM. pp. 111-119 (2020)", + "45. Wang, K., Xie, H., Wang, Y., Zhang, D., Qu, Y., Gao, Z., Zhang, Y.: Masked text modeling: A self-supervised pre-training method for scene text detection. In: ACM MM. pp. 2006-2015 (2023)", + "46. Wang, W., Xie, E., Li, X., Hou, W., Lu, T., Yu, G., Shao, S.: Shape robust text detection with progressive scale expansion network. In: CVPR. pp. 9336-9345 (2019)", + "47. Wang, W., Xie, E., Li, X., Liu, X., Liang, D., Yang, Z., Lu, T., Shen, C.: Pan++: Towards efficient and accurate end-to-end spotting of arbitrarily-shaped text. IEEE TPAMI 44(9), 5349-5367 (2022)", + "48. Wang, W., Xie, E., Song, X., Zang, Y., Wang, W., Lu, T., Yu, G., Shen, C.: Efficient and accurate arbitrary-shaped text detection with pixel aggregation network. In: ICCV. pp. 8440-8449 (2019)", + "49. Wang, X., Jiang, Y., Luo, Z., Liu, C.L., Choi, H., Kim, S.: Arbitrary shape scene text detection with adaptive text region representation. In: CVPR. pp. 6449-6458 (2019)", + "50. Wang, Y., Xie, H., Zha, Z.J., Xing, M., Fu, Z., Zhang, Y.: Contournet: Taking a further step toward accurate arbitrary-shaped scene text detection. In: CVPR. pp. 11753-11762 (2020)", + "51. Wu, W., Lu, N., Xie, E., Wang, Y., Yu, W., Yang, C., Zhou, H.: Synthetic-to-real unsupervised domain adaptation for scene text detection in the wild. In: Proceedings of the Asian Conference on Computer Vision (2020)", + "52. Xu, H., Liu, L., Bian, Q., Yang, Z.: Semi-supervised semantic segmentation with prototype-based consistency regularization. In: NeurIPS. vol. 35, pp. 26007-26020 (2022)", + "53. Xu, X., Zhang, Z., Wang, Z., Price, B., Wang, Z., Shi, H.: Rethinking text segmentation: A novel dataset and a text-specific refinement approach. In: CVPR. pp. 12045-12055 (2021)", + "54. Xue, C., Zhang, W., Hao, Y., Lu, S., Torr, P.H., Bai, S.: Language matters: A weakly supervised vision-language pre-training approach for scene text detection and spotting. In: ECCV. pp. 284-302. Springer (2022)", + "55. Yang, L., Qi, L., Feng, L., Zhang, W., Shi, Y.: Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In: CVPR. pp. 7236-7246 (2023)", + "56. Yang, X., Song, Z., King, I., Xu, Z.: A survey on deep semi-supervised learning. IEEE Transactions on Knowledge and Data Engineering 35(9), 8934-8954 (2022)", + "57. Yao, C., Bai, X., Liu, W., Ma, Y., Tu, Z.: Detecting texts of arbitrary orientations in natural images. In: CVPR. pp. 1083-1090 (2012)", + "58. Ye, M., Zhang, J., Zhao, S., Liu, J., Liu, T., Du, B., Tao, D.: Deepsolo: Let transformer decoder with explicit points solo for text spotting. In: CVPR. pp. 19348-19357 (2023)", + "59. Yin, X.C., Pei, W.Y., Zhang, J., Hao, H.W.: Multi-orientation scene text detection with adaptive clustering. IEEE TPAMI 37(9), 1930-1937 (2015)", + "60. Yu, W., Liu, Y., Hua, W., Jiang, D., Ren, B., Bai, X.: Turning a clip model into a scene text detector. In: CVPR. pp. 6978-6988 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "FreeReal for Pre-training STDs", + "bbox": [ + 521, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "61. Yu, W., Liu, Y., Zhu, X., Cao, H., Sun, X., Bai, X.: Turning a clip model into a scene text spotter. IEEE TPAMI (2024)", + "62. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: ICCV. pp. 6023-6032 (2019)", + "63. Zhan, F., Xue, C., Lu, S.: Ga-dan: Geometry-aware domain adaptation network for scene text detection and recognition. In: ICCV. pp. 9105-9115 (2019)", + "64. Zhang, B., Wang, Y., Hou, W., Wu, H., Wang, J., Okumura, M., Shinozaki, T.: Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In: NeurIPS. vol. 34, pp. 18408-18419 (2021)", + "65. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv:1710.09412 (2017)", + "66. Zhao, Z., Yang, L., Long, S., Pi, J., Zhou, L., Wang, J.: Augmentation matters: A simple-yet-effective approach to semi-supervised semantic segmentation. In: CVPR. pp. 11350-11359 (2023)", + "67. Zhou, X., Yao, C., Wen, H., Wang, Y., Zhou, S., He, W., Liang, J.: East: an efficient and accurate scene text detector. In: CVPR. pp. 5551-5560 (2017)", + "68. Zhu, Y., Chen, J., Liang, L., Kuang, Z., Jin, L., Zhang, W.: Fourier contour embedding for arbitrary-shaped text detection. In: CVPR. pp. 3123-3131 (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 410 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "T. Guan et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Bridging Synthetic and Real Worlds for Pre-training Scene Text Detectors/80dda151-f601-4dfe-a0ed-8536e8b997b9_model.json b/2024/Bridging Synthetic and Real Worlds for Pre-training Scene Text Detectors/80dda151-f601-4dfe-a0ed-8536e8b997b9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7b6f1008d809083352405932d67894f0e18ad0ab --- /dev/null +++ b/2024/Bridging Synthetic and Real Worlds for Pre-training Scene Text Detectors/80dda151-f601-4dfe-a0ed-8536e8b997b9_model.json @@ -0,0 +1,2458 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.276, + 0.141, + 0.73, + 0.187 + ], + "angle": 0, + "content": "Bridging Synthetic and Real Worlds for Pre-training Scene Text Detectors" + }, + { + "type": "text", + "bbox": [ + 0.343, + 0.212, + 0.66, + 0.244 + ], + "angle": 0, + "content": "Tongkun Guan\\(^{1}\\), Wei Shen\\(^{1(\\boxtimes)}\\), Xue Yang\\(^{2}\\), Xuehui Wang\\(^{1}\\), and Xiaokang Yang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.254, + 0.786, + 0.282 + ], + "angle": 0, + "content": "1 MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University {gtk0615,wei.shen}@sjtu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.412, + 0.283, + 0.59, + 0.312 + ], + "angle": 0, + "content": "\\(^{2}\\) Shanghai AI Laboratory yangxue@pjlab.org.cn" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.349, + 0.74, + 0.64 + ], + "angle": 0, + "content": "Abstract. Existing scene text detection methods typically rely on extensive real data for training. Due to the lack of annotated real images, recent works have attempted to exploit large-scale labeled synthetic data (LSD) for pre-training text detectors. However, a synth-to-real domain gap emerges, further limiting the performance of text detectors. Differently, in this work, we propose FreeReal, a real-domain-aligned pretraining paradigm that enables the complementary strengths of both LSD and unlabeled real data (URD). Specifically, to bridge real and synthetic worlds for pre-training, a glyph-based mixing mechanism (GlyphMix) is tailored for text images. GlyphMix delineates the character structures of synthetic images and embeds them as graffiti-like units onto real images. Without introducing real domain drift, GlyphMix freely yields real-world images with partial annotations derived from synthetic labels. Furthermore, when given free fine-grained synthetic labels, GlyphMix can effectively bridge the linguistic domain gap stemming from English-dominated LSD to URD in various languages. Without bells and whistles, FreeReal achieves average gains of \\(1.97\\%\\), \\(3.90\\%\\), \\(3.85\\%\\), and \\(4.56\\%\\) in improving the performance of FCENet, PSENet, PANet, and DBNet methods, respectively, consistently outperforming previous pre-training methods by a substantial margin across four public datasets. Code will be available at https://github.com/SJTU-DeepVisionLab/FreeReal." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.653, + 0.741, + 0.682 + ], + "angle": 0, + "content": "Keywords: Scene text detection \\(\\cdot\\) Pre-training paradigm \\(\\cdot\\) Domain-bridging method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.708, + 0.377, + 0.723 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.74, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Scene text detection aims to predict the bounding box or polygon for each arbitrarily shaped text instance within an image. This technology has diverse applications, including online education, instant translation, and product search, owing to the prevalence of text-laden images. However, despite their considerable progress, supervised methods [8,10,22,46,50,68] still struggle with an increasing" + }, + { + "type": "page_footnote", + "bbox": [ + 0.23, + 0.824, + 0.398, + 0.841 + ], + "angle": 0, + "content": "Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.228, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.128 + ], + "angle": 0, + "content": "T. Guan et al." + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.146, + 0.728, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.231, + 0.283, + 0.772, + 0.298 + ], + "angle": 0, + "content": "Fig. 1: We present different pre-training paradigms for enhancing text detectors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.333, + 0.785, + 0.364 + ], + "angle": 0, + "content": "demand for real data with labor-intensive annotations, especially when detecting complicated text images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.593 + ], + "angle": 0, + "content": "In contrast, recent research efforts [39, 43, 54] have focused on pre-training scene text detectors using large-scale labeled synthetic data (LSD), which leverage both visual and textual knowledge to extract text feature representations, as illustrated in Fig. 1(a). For example, VLPT [39] formulates three cross-modality pretext tasks to enhance the performance of the backbone, and oCLIP [54] focuses on aligning visual and textual information to assist text detectors. Nevertheless, these well-curated designs are limited to using synthetic images due to their reliance on labeled word transcriptions. This limitation will lead to a noticeable synth-to-real domain gap when models are fine-tuned on real data, negatively impacting text detection performance. Therefore, exploring the potential of unlabeled real data (URD) is of great importance as they are readily available. However, the pre-training on URD has seldom been studied. One of the few works [45] attempts to pre-train text detectors on URD via masked image modeling [13] (step 1) and on LSD via supervised learning (step 2), which will re-transfers pre-trained weights from the real domain to the synthetic domain." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.746 + ], + "angle": 0, + "content": "Differently, our objective is to jointly leverage LSD and URD by bridging their gap for enhancing the text detector pre-training, as shown in Fig. 1(b). The primary challenge is how to effectively modulate these URD and LSD without encountering real domain drift. In light of this, we initially investigate the adaptation capabilities of existing domain-bridging methods [32, 62, 65]. This involves training a classifier to distinguish between URD and LSD and subsequently assessing the confidence of these domain-bridging methods in aligning with the real domain. As shown in Fig. 2(b), their highest score is only \\(78.9\\%\\). The observation implies that these domain-bridging methods may introduce intermediate domains and disrupt the real domain." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To conquer this domain drift issue, we propose FreeReal, a real-domain-aligned pre-training paradigm, leveraging synthetic labels to guide text detector learning on real-domain-aligned images. Specifically, a glyph-based mixing mechanism, named GlyphMix, is tailored for text images to create real-world images with annotations derived from synthetic labels. This involves delineating the glyphs of synthetic images based on their internal structure and embedding" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.522, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "FreeReal for Pre-training STDs" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.146, + 0.725, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.266, + 0.788, + 0.378 + ], + "angle": 0, + "content": "Fig. 2: We present the domain classification task and the performances of different domain bridging ways between LSD and URD. In the text detection area, experiments demonstrate that our domain bridging method significantly performs better in aligning with the real domain. This is easy to explain: (I) Mixup may lead to unexpected pixel-wise ambiguities; (II) CutMix may cause the absence of many text regions and incomplete semantics of texts; (III) ClassMix will introduce salient boundary priors; (IV) GlyphMix (ours) effectively preserves the semantic information of text without causing real domain drift and bringing extra boundary priors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.785, + 0.472 + ], + "angle": 0, + "content": "them as graffiti-like units onto any unlabeled real images. GlyphMix offers several advantages: 1) extracting glyph structures in a cost-effective way without introducing extra boundary priors; 2) leveraging synthetic labels for free while greatly minimizing real domain drift (achieving a confidence score of \\(95.7\\%\\))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.788, + 0.581 + ], + "angle": 0, + "content": "As a byproduct, FreeReal can bridge the language-to-language gap, when adapting features learned from English-dominated LSD to URD in various languages. Specifically, given free character-level synthetic labels, FreeReal enables the pre-trained model to easily learn character features, since characters are the underlying unit of any language. Subsequently, during fine-tuning, these learned character features are adaptively combined according to the rules (e.g., word-and sentence-level) provided by available benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.704 + ], + "angle": 0, + "content": "Building upon these insights, we break the trend of recent pre-training methods for scene text detectors (PSTD) that combine increasingly complex pretext task designs or more training modules to improve performance. Without bells and whistles, we focus on exploring the potential of integrating the LSD and URD. Remarkably, despite using \\(90\\%\\) less LSD than recent SOTAs, our method still attains a \\(1.2\\%\\) gain, effectively underscoring the superiority of our pretraining paradigm. As shown in Table 1, FreeReal consistently surpasses recent SOTAs by a large margin. The main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.72, + 0.784, + 0.779 + ], + "angle": 0, + "content": "- We are the first to simultaneously integrate labeled synthetic data and unlabeled real data for enhancing the pre-training of text detectors. FreeReal shows its superiority and significant performance gains over other complex alternatives, serving as a simple yet strong baseline for future PSTD studies." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.78, + 0.784, + 0.839 + ], + "angle": 0, + "content": "- We effectively address the challenges of synthetic-to-real and language-to-language domain adaptation. Without bringing boundary priors and causing real domain drift, we freely harness the strengths of real-world text images with annotations derived from synthetic labels." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.72, + 0.784, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "T. Guan et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.388, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.812, + 0.281 + ], + "angle": 0, + "content": "Scene Text Detection. Existing STD methods delineate text regions by regression-based or segmentation-based representations. The regression-based methods [8, 14,24,34,49,67,68] directly regress the bounding points to enclose text regions, and gradually increase the number of points as required to detect complex scenes, e.g., arbitrary texts. The segmentation-based methods regard texts of arbitrary shapes by grouping pixel-level [22, 23, 40, 48], character-level [1, 9], segment-level [26,37] or contour-level [35,44,50] segmentation results by post-processing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.282, + 0.787, + 0.569 + ], + "angle": 0, + "content": "Pre-training for Scene Text Detector. In early research, [12, 21, 27] proposed to pre-train scene text detectors by generating synthetic data. These data engine tools follow the same paradigm: they first stack multiple off-the-shelf models to perform semantic segmentation and depth estimation analysis on background images, and then blend pre-processed texts into these images. However, the multi-stage, pre-generated methods incur large computational costs (GPU and CPU resources) and substantial latency, making them unsuitable for online training tasks where immediate data processing is essential. Inspired by recent advancements in pre-training techniques [3, 4, 11, 36], the integration of text and visual cross-modal information into scene text detectors has garnered increasing attention. One representative work involves enhancing the backbone with text transcriptions. For example, STKM [43] proposes a pioneering work that pretrains the backbone through an image-level text recognition task. VLPT [39] formulates three cross-modality pretext tasks to boost better text representations derived from the backbone. Another representative work is to endow the entire detector with pre-trained vision-text knowledge. Inherited from CLIP [36], TCM incorporates the Image and Text encoders into the text detector to refine text regions using cross-modal visual-language priors. A persistent challenge still remains—the synth-to-real domain gap limits the detection performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Semi-Supervised Learning. Due to the scarcity of expensive annotations, semi-supervised learning (SSL) methods are becoming increasingly important. These methods extract knowledge from a reduced amount of annotated images in a supervised manner, and from a larger amount of unlabeled images in an unsupervised way. Commonly, a self-training framework for SSL is employed, which encourages a model (student) to learn unlabeled images from pseudo-labels generated by the up-to-date optimized model (teacher). To avoid overfitting incorrect pseudo-labels, recent advancements have focused on pseudo-label refinement by improving the reliability [38,64] and pseudo-label regularization by cross-view consistency [19, 20, 42, 52]. Specifically, some domain adaption works [5, 51, 63] for STD focus on minimizing domain discrepancies in the feature space, including self-training and adversarial alignment methods. The self-training methods adopt the source data to pre-train a model for generating pseudo labels on target data, then use them to retrain the model. Adversarial alignment approaches aim to align the feature distribution and learn the domain-invariant features between the source and target domains. Recently, they have been replaced by the student-teacher framework, which significantly reduces domain gaps through dual-branch knowledge transfer [56]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.522, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "FreeReal for Pre-training STDs" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.484, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.253, + 0.379, + 0.262 + ], + "angle": 0, + "content": "(a) FreeReal" + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.148, + 0.785, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.253, + 0.669, + 0.264 + ], + "angle": 0, + "content": "(b) GlyphMix" + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.276, + 0.705, + 0.29 + ], + "angle": 0, + "content": "Fig. 3: Our network pipeline in a student-teacher framework." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.32, + 0.38, + 0.338 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.788, + 0.483 + ], + "angle": 0, + "content": "Overview. Given a labeled synthetic set \\(\\mathcal{D}_l = \\{(\\mathbf{X}_l^{(i)},\\mathcal{B}_l^{(i)})\\}_{i = 1}^{N_l}\\) with \\(N_{l}\\) samples, \\(\\mathbf{X}_l^{(i)}\\in \\mathbb{R}^{H\\times W\\times 3}\\) represents an image and \\(\\mathcal{B}_l^{(i)} = \\{(p_{lt}^i,p_{lb}^i,p_{rt}^i,p_{rb}^i)\\}_{i = 1}^n\\) is its corresponding set of bounding boxes. \\(p\\) refers to the values of the vertice along the abscissa and ordinate, and \\(n\\) denotes the number of bounding boxes. \\(H\\) and \\(W\\) are the height and width of the image. Similarly, the unlabeled real set is denoted as \\(\\mathcal{D}_u = \\{\\mathbf{X}_u^{(i)}\\}_{i = 1}^{N_u}\\) with \\(N_{u}\\) samples. Our objective is to jointly leverage labeled synthetic data \\(\\mathcal{D}_l\\) and unlabeled real data \\(\\mathcal{D}_u\\) by bridging their gap for enhancing the pre-training of scene text detectors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.484, + 0.788, + 0.573 + ], + "angle": 0, + "content": "The student-teacher framework exhibits substantial efficacy due to its dual capacity for knowledge transfer between its two branches, which is typically employed to mitigate the domain gap (e.g., data and multimodal). Accordingly, we customize it for aligning real and synthetic text data, as illustrated in Fig. 3. For notation simplification in the following introduction, we omit the superscript \\(i\\) as mentioned above." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.788, + 0.787 + ], + "angle": 0, + "content": "Specifically, the segmentation-based text detector DBNet [22] is employed as our baseline model, primarily due to its impressive real-time performance, versatility, and practical applicability. In the teacher branch, the teacher model parameterized by \\(\\mathbf{f}_t\\), is responsible for generating truncated-gradient logits \\(\\hat{\\mathbf{Y}}_u \\in \\mathbb{R}^{H \\times W}\\) from a weakly augmented unlabeled real image \\(w(\\mathbf{X}_u)\\). The pseudo label \\(\\mathbf{Y}_u \\in \\{0,1\\}^{H \\times W}\\) is then obtained by binarizing the logit map \\(\\hat{\\mathbf{Y}}_u\\). In the student branch, the process begins by generating a pixel-wise binary map \\(\\mathbf{Y}_l \\in \\{0,1\\}^{H \\times W}\\), where 1 is assigned to locations within all bounding boxes from annotation \\(\\mathcal{B}_l\\), and 0 otherwise. The labeled synthetic image \\(\\mathbf{X}_l\\) and the strongly augmented unlabeled real image \\(s(\\mathbf{X}_u)\\) are combined to create \\(\\mathbf{X}_g = \\mathsf{A}(s(\\mathbf{X}_u), \\mathbf{X}_l)\\), with \\(\\mathsf{A}(\\cdot)\\) denoting our proposed GlyphMix mechanism, which will be introduced later. Subsequently, the mixed image \\(\\mathbf{X}_g\\) serves as input to train the student model, under the supervision of mixed pseudo-labels \\(\\mathbf{Y}_g = \\mathsf{A}(\\mathbf{Y}_u, \\mathbf{Y}_l)\\), by minimizing loss in the following:" + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.788, + 0.785, + 0.806 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\ell_ {D B} \\left(\\hat {\\mathbf {Y}} _ {g}, \\mathbf {Y} _ {g}, \\mathbf {E}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "where \\(\\ell_{DB}\\) denotes the original loss function provided by DBNet [22] and \\(\\hat{\\mathbf{Y}}_g\\) represents the prediction logits generated by the student model. The matrix" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.217, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "T. Guan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.176 + ], + "angle": 0, + "content": "\\(\\mathbf{E} \\in \\mathbb{R}^{H \\times W}\\), which indicates positive regions for loss calculation, is set as an all-one matrix by default." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.253 + ], + "angle": 0, + "content": "To filter out the unreliable predictions in mixed pseudo-labels \\(\\mathbf{Y}_g\\), we adaptively adjust the matrix \\(\\mathbf{E}\\) during pre-training. Specifically, for the labeled synthetic image \\(\\mathbf{X}_l\\), the positive regions \\(\\mathbf{E}_l\\) are an all-one matrix. For the unlabeled real image \\(\\mathbf{X}_u\\), we compute the pixel-wise entropy map on \\(\\hat{\\mathbf{Y}}_u\\) and then apply a dynamic entropy threshold \\(\\zeta\\) to obtain its positive regions \\(\\mathbf{E}_u\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.256, + 0.785, + 0.272 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} _ {u} = \\mathbb {I} \\left(- \\hat {\\mathbf {Y}} _ {u} \\log \\hat {\\mathbf {Y}} _ {u} \\leq \\zeta\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.274, + 0.785, + 0.349 + ], + "angle": 0, + "content": "where \\(\\mathbb{I}(\\cdot)\\) denotes the indicator function. The matrix is subsequently formulated as \\(\\mathbf{E} = \\mathsf{A}(\\mathbf{E}_u,\\mathbf{E}_l)\\). In our experiments, the threshold \\(\\zeta\\) is determined by the value corresponding to the top \\(\\gamma\\) percent of the entropy map. As the reliability of pseudo-label gradually improves, we smoothly update \\(\\gamma\\) using np.linspace(80, 20, t), where \\(t\\) denotes the total number of training iterations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.784, + 0.38 + ], + "angle": 0, + "content": "Finally, the teacher model \\(\\mathbf{f}_t\\) gradually updates itself via the exponential moving averaging (EMA) of the student weights \\(\\mathbf{f}_s\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.383, + 0.785, + 0.398 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {t} = \\alpha \\mathbf {f} _ {t} + (1 - \\alpha) \\mathbf {f} _ {s}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.401, + 0.644, + 0.416 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a momentum parameter, set as 0.996 by default." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.441, + 0.347, + 0.456 + ], + "angle": 0, + "content": "3.1 GlyphMix" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.785, + 0.544 + ], + "angle": 0, + "content": "Although some efforts [32, 62, 65] attempt to aggregate the advantages of synthetic and unlabeled images through versatile transformations, they inevitably introduce intermediate domains, leading to real domain drift. In this section, we present a domain-bridging way tailored for text images, allowing for the free use of synthetic labels while ensuring consistent alignment with the real domain." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.545, + 0.785, + 0.605 + ], + "angle": 0, + "content": "1) Graffiti-like Inter-domain Mixing. Given an input synthetic image \\(\\mathbf{X}_l\\), its corresponding bounding box annotations \\(\\mathcal{B}_l\\) and converted pixel-wise binary map \\(\\mathbf{Y}_l\\), our goal is to perform character segmentation. This process identifies all character structure regions and produces a mask \\(\\mathbf{M}_G\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.786, + 0.756 + ], + "angle": 0, + "content": "To achieve this, we begin with an unsupervised clustering task based on the grayscale values of pixels, avoiding the introduction of extra pixel-level annotation costs. This involves aligning and cropping the image according to the coordinates of each bounding box, yielding a set of image patches \\(\\mathcal{T} = \\{\\mathbf{T}^{(i)}|\\mathbf{T}^{(i)}\\in \\mathbb{R}^{w_i\\times h_i\\times 3}\\}_{i = 1}^n\\). Subsequently, within each image patch \\(\\mathbf{T}^{(i)}\\), we cluster the pixels into two regions: the glyph region representing character structures and the background region. For simplicity, we focus our study on K-means to implement the clustering task, but other clustering approaches with predefined categories can be used. Following this, the glyph regions from all image patches are pasted to an all-zero image to create the character structure mask \\(\\mathbf{M}_G\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.757, + 0.784, + 0.802 + ], + "angle": 0, + "content": "Given the pseudo label \\(\\mathbf{Y}_u\\) produced by the teacher model on an unlabeled real image \\(\\mathbf{X}_u\\), we use \\(\\mathbf{M}_G\\) to generate the inter-domain mixing image and its corresponding label. This process can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.805, + 0.785, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\mathbb {A} ^ {G} \\left(\\mathbf {X} _ {u}, \\mathbf {X} _ {l}\\right) \\leftarrow \\mathbf {M} _ {G} \\odot \\mathbf {X} _ {l} + \\left(\\mathbf {1} - \\mathbf {M} _ {G}\\right) \\odot \\mathbf {X} _ {u}, \\\\ \\mathbb {A} ^ {G} \\left(\\mathbf {Y} _ {u}, \\mathbf {Y} _ {l}\\right) \\leftarrow \\mathbf {M} _ {G} \\odot \\mathbf {Y} _ {l} + \\left(\\mathbf {1} - \\mathbf {M} _ {G}\\right) \\odot \\mathbf {Y} _ {u}, \\end{array} \\right. \\tag {4}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.522, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "FreeReal for Pre-training STDs" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.236 + ], + "angle": 0, + "content": "where \\(\\mathbf{1}\\) is an all-one matrix and \\(\\odot\\) denotes the element-wise multiplication operation. \\(\\mathbf{M}_G\\) indicates pixels occupied by the glyph need to be copied from the synth domain and pasted to the real domain. Thus the mixed image closely aligns with the real domain, taking advantage of the freely available synthetic labels. Put simply, this process produces real-world images with partial annotations derived from synthetic labels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.785, + 0.389 + ], + "angle": 0, + "content": "2) Text-balanced Intra-domain Mixing. Real-world text images inherently exhibit complexity, and the collected unlabeled data often struggles to comprehensively represent the full spectrum of real-world scenarios. Within a student-teacher framework, intra-domain mixing serves as an effective perturbation technique, enriching the diversity of real data and enhancing feature consistency and robustness. However, when randomly generating and applying binary masks to construct intra-domain mixing images and their corresponding labels, there is a risk of missing many text regions, especially considering that text regions typically occupy only a small fraction of an image. To mitigate this challenge, our objective is to retain more text regions while still maintaining randomness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.785, + 0.549 + ], + "angle": 0, + "content": "In light of this, we initiate the process by randomly selecting two unlabeled real images, denoted as \\((\\mathbf{X}_u^{(1)},\\mathbf{X}_u^{(2)})\\), and obtaining the corresponding segmentation pseudo-labels, \\((\\mathbf{Y}_u^{(1)},\\mathbf{Y}_u^{(2)})\\), produced by the teacher model. Subsequently, \\(k\\) rectangular boxes \\(\\{(p_{lt}^i,p_{lb}^i,p_{rt}^i,p_{rb}^i)\\}_{i = 1}^k\\) are randomly generated to crop the pseudo-label \\(\\mathbf{Y}_u^{(2)}\\) based on their coordinates. This operation yields a set of local pseudo-label regions \\(S = \\{\\mathbf{S}^{(i)}|\\mathbf{S}^{(i)}\\in \\mathbb{R}^{w_i\\times h_i}\\}_{i = 1}^k\\). Following this, we identify the rectangular box that contains the maximum text regions by calculating \\(\\arg \\max_{i}\\sum_{j = 1}^{w_{i}\\times h_{i}}\\mathbf{S}_{j}^{(i)}\\), and convert it into a binary mask, denoted as \\(\\mathbf{M}_T\\). Finally, we employ \\(\\mathbf{M}_T\\) to generate the intra-domain mixed image and its corresponding label, following the formulated procedure:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.552, + 0.785, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\mathbf {A} ^ {T} (\\mathbf {X} _ {u}) \\leftarrow \\mathbf {M} _ {T} \\odot \\mathbf {X} _ {u} ^ {(2)} + \\left(\\mathbf {1} - \\mathbf {M} _ {T}\\right) \\odot \\mathbf {X} _ {u} ^ {(1)}, \\\\ \\mathbf {A} ^ {T} (\\mathbf {Y} _ {u}) \\leftarrow \\mathbf {M} _ {T} \\odot \\mathbf {Y} _ {u} ^ {(2)} + \\left(\\mathbf {1} - \\mathbf {M} _ {T}\\right) \\odot \\mathbf {Y} _ {u} ^ {(1)}, \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.78, + 0.607 + ], + "angle": 0, + "content": "where \\(\\mathbf{1}\\) is an all-one matrix and \\(\\odot\\) represents the element-wise multiplication." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.784, + 0.667 + ], + "angle": 0, + "content": "3) Intra- and Inter-domain Mixing. When these two strategies, tailored for scene text images, are integrated within the student-teacher framework, they collectively form a comprehensive method that effectively harnesses the strengths of both synthetic and real data. The whole process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.398, + 0.67, + 0.785, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\mathbf {A} (\\mathbf {X}) \\leftarrow \\mathbf {A} ^ {G} \\left(\\mathbf {A} ^ {T} \\left(\\mathbf {X} _ {u}\\right), \\mathbf {X} _ {l}\\right), \\\\ \\mathbf {A} (\\mathbf {Y}) \\leftarrow \\mathbf {A} ^ {G} \\left(\\mathbf {A} ^ {T} \\left(\\mathbf {Y} _ {u}\\right), \\mathbf {Y} _ {l}\\right) \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.71, + 0.784, + 0.74 + ], + "angle": 0, + "content": "Consequently, the synergy maximizes the utilization of synthetic labels as a valuable resource, effectively tackling the intricacies of real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.766, + 0.618, + 0.781 + ], + "angle": 0, + "content": "3.2 Discussion on Character Region Awareness" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Existing LSD only contains English texts with tightly spaced characters, while URD may encompass text properties from various languages (such as Japanese, Korean, Chinese, etc). Consequently, even though the model predictions effec" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.228, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "T. Guan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "tively cover text regions, we encounter challenges in parsing these predictions into word-level bounding boxes or polygons." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.284 + ], + "angle": 0, + "content": "Indeed, when using a single language (e.g., English) as the source domain and adapting it to various language domains, it appears that no matter how sophisticated the models and loss functions are, they struggle to accurately and effectively parse the complex text layout and semantic context. This challenge primarily arises from variations in the spacing and arrangement of characters, words, and sentences across languages. This observation leads us to revisit the learning objective of scene text detectors during pre-training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.284, + 0.788, + 0.435 + ], + "angle": 0, + "content": "We emphasize an alternative perspective — utilizing characters as the fundamental learning unit for text detection. The strategy presents several advantages: 1) Like word-level bounding boxes (WBB), character-level bounding box (CBB) annotations can be readily acquired from the existing synthetic dataset (SynthText) at no additional cost; 2) Instead of parsing complex layouts and spacing disparities across languages during pre-training, decoding them into individual characters is straightforward and attainable; 3) The pre-trained model with character-level region awareness easily integrates its learned character features to adapt text attribute rules provided by different language text datasets during fine-tuning. To this end, we select CBB annotations as \\(\\mathcal{B}_l\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.459, + 0.366, + 0.475 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.492, + 0.327, + 0.506 + ], + "angle": 0, + "content": "4.1 Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.787, + 0.564 + ], + "angle": 0, + "content": "Labeled Synthetic Data (LSD). The SynthText [12] serves as our source of LSD, consisting of 800K images. Each image is provided with character-level and word-level bounding boxes, along with corresponding text transcriptions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.787, + 0.625 + ], + "angle": 0, + "content": "Unlabeled Real Data (URD). URD is collected from a wide array of natural scenes, encompassing a diverse range of fonts, languages, and text shapes. Specifically, URD includes 430K images from LSVT [41], 9K from MLT17 [30], 19657 from MLT19 [29], 229 from ICDAR2013 [18], and 500 from USTB-SV1K [59]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.787, + 0.672 + ], + "angle": 0, + "content": "**Benchmarks.** Four publicly available benchmarks are used to evaluate the performance, including 1) Total-Text (TT) [6], 2) CTW 1500 (CTW) [25], 3) IC-DAR2015 (IC15) [17], and 4) MSRA-TD500 (TD) [57]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.693, + 0.457, + 0.708 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Referring to recent works [39, 43, 54], we combine representative text detection methods [22, 46, 48, 68] to conduct experiments. URD is used in Tabs. 1-3 and the same data is used in Tabs. 4-7. All experiments are implemented on a server with 2 NVIDIA A800 GPUs, using PyTorch. 1) Pre-Training. The pre-training experiments are conducted on LSD and URD, utilizing the open-source codes provided by these text detection methods. For synthetic images, we adopt consistent augmentation settings as prescribed by these methods for a fair comparison. For unlabeled real images, we employ a standard weak-to-strong augmentation" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.522, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "FreeReal for Pre-training STDs" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "technique [66] to enrich diversity. Our method is trained using the AdamW optimizer [28], along with a cosine learning rate scheduler [7]. This includes a warm-up phrase for 1 epoch in a total of 10 epochs. The base learning rate is set to \\(0.003 \\times\\) batchsize/256. At each iteration, we equally sample LSD and URD with a total batch size of 24. 2) Fine-Tuning. For a fair comparison, we employ the same fine-tuning configurations as these detection methods and then evaluate them on these publicly available benchmarks: TT, CTW, IC15 and TD, following the standard metric: Precision (P), Recall (R), and F-measure (F)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.289, + 0.427, + 0.304 + ], + "angle": 0, + "content": "4.3 Experiment Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.789, + 0.477 + ], + "angle": 0, + "content": "To ensure a comprehensive comparison, we collect results from public evaluation benchmarks reported by existing pre-training methods, and organize them in Table 1. The sources of these results are clearly recorded in the table's caption using different symbols for easy reference. Specifically, these methods can be classified into three pre-training paradigms. In Paradigm A, models (i.e., \\(+ST\\), \\(+TCM\\)) are initially pre-trained on LSD and then fine-tuned on real datasets directly. In Paradigm B, models (i.e., \\(+STKM\\), \\(+VLPT\\), \\(+oCLIP\\)) are first pretrained on LSD by various pretext tasks, and the well-learned backbone weights are transferred into scene text detectors when fine-tuning on real datasets. In Paradigm C, models (i.e., \\(+MTM\\)) are pre-trained on URD by pretext tasks and on LSD by supervised learning, and then fine-tuned on real datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.789, + 0.689 + ], + "angle": 0, + "content": "1) Results on DBNet. As indicated in the second group of rows in Table 1, all pre-training methods both report their cooperative performance with DBNet. When compared to methods using paradigm A, differing from \\(DBNet + ST\\), we exclude the use of a deformable convolutional network (DCN) in our model. Despite these differences, our method still achieves impressive results, yielding improvements of \\(4.2\\%\\), \\(4.5\\%\\), \\(4.6\\%\\), and \\(5.2\\%\\) on TT, CTW, IC15, and TD datasets, respectively. Compared to \\(DBNet + TCM\\) methods, our method achieves SOTA detection performance on four datasets, particularly on curved text datasets (TT and CTW), with gains of \\(3.0\\%\\) and \\(2.8\\%\\), respectively. In comparison with methods following Paradigm B, our method consistently outperforms previous SOTA works by substantial margins, achieving performance improvements of \\(2.6\\%\\), \\(5.9\\%\\), \\(3.5\\%\\) and \\(1.6\\%\\) on TT, CTW, IC15, and TD datasets, respectively. In comparison with \\(DBNet + MTM\\) following Paradigm C, our method gets \\(3.2\\%\\) and \\(4.5\\%\\) gains on TT and TD datasets, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.781 + ], + "angle": 0, + "content": "2) Results on PANet. As shown in the third group of rows in Table 1, our method outperforms \\(PANet + ST\\) on four benchmarks, achieving a gain of \\(3.85\\%\\) in terms of average F-measure, while using the same parameters and labeled data for pre-training and fine-tuning. Furthermore, our method exhibits its prominent superiority, particularly on the CTW and TD datasets, surpassing the \\(PANet + TCM\\) method by a substantial margin of \\(4.1\\%\\) and \\(4.6\\%\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.81 + ], + "angle": 0, + "content": "3) Results on PSENet. FreeReal brings \\(0.7\\%\\), \\(2.8\\%\\), and \\(0.9\\%\\) gains on TT, CTW, and IC15 datasets, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "4) Results on FCENet. FreeReal further refreshes the best text detection results and improves \\(2.1\\%\\), \\(1.3\\%\\), \\(0.6\\%\\) and \\(2.1\\%\\) on TT, CTW, IC15, and TD" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.478, + 0.789, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "T. Guan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.214 + ], + "angle": 0, + "content": "Table 1: Comparison with SOTA pre-training methods for text detectors on TT, CTW, IC15, and TD. F-measure (\\%) is reported. * denotes the use of extra labeled real data. For easy reference, we mark these symbols \\(\\dagger\\), \\(*\\), \\(\\ddagger\\), and \\(*\\) separately to indicate the results from [61], [39], [54], and [45], within the scope of this paper. The best results are shown in bold font. The second-best results are underlined." + }, + { + "type": "table", + "bbox": [ + 0.238, + 0.227, + 0.761, + 0.526 + ], + "angle": 0, + "content": "
IndexMethodVenueTTCTWIC15TD
1PSENet-1s [46]CVPR'1980.982.285.7-
PAN++ [47]PAMI'2185.384.084.587.9
DBNet++ [23]PAMI'2286.085.387.387.2
DeepSOLO* [58]CVPR'2387.6-89.8-
ESTextSpotter* [15]ICCV'2390.0-91.089.5
SIR [33]ACMMM'2388.285.587.889.6
2DBNet+ST† [22]AAAI'2084.783.485.484.9
DBNet+STKM† [43]CVPR'2185.5-86.185.9
DBNet+VLPT† [39]CVPR'2286.3-86.588.5
DBNet+oCLIP† [54]ECCV'2284.182.085.4-
DBNet+TCM† [60]CVPR'2385.985.189.488.8
DBNet+MTM* [45]ACMMM'2385.7--85.6
DBNet+FreeReal (ours)-88.9(+2.6)87.9(+2.8)90.0(+0.6)90.1(+1.3)
3PANet+ST† [48]ICCV'1985.083.782.984.1
PANet+TCM† [60]CVPR'23-84.384.685.3
PANet+FreeReal (ours)-87.4(+2.4)88.4(+4.1)86.0(+1.4)89.9(+4.6)
4PSENet+ST* [46]CVPR'1983.981.681.3-
PSENet+STKM* [43]CVPR'2184.382.983.7-
PSENet+VLPT* [39]CVPR'2286.183.384.3-
PSENet+oCLIP† [54]ECCV'2285.582.8--
PSENet+FreeReal (ours)-86.8(+0.7)86.1(+2.8)85.2(+0.9)86.0
5FCENet+ST† [68]CVPR'2185.885.586.285.4
FCENet+oCLIP† [54]ECCV'22-85.686.7-
FCENet+TCM† [60]CVPR'23-85.987.186.9
FCENet+FreeReal (ours)-87.9(+2.1)87.2(+1.3)87.7(+0.6)89.0(+2.1)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.574, + 0.788, + 0.604 + ], + "angle": 0, + "content": "datasets, respectively. These experimental results demonstrate the robustness of FreeReal, even when cooperated with regression-based methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.842 + ], + "angle": 0, + "content": "5) Paradigm Comparison. In this section, we aim to investigate which kinds of pre-training paradigms for text detectors are important and meaningful for future studies. To this end, we conduct a comparative analysis with three mainstream pre-training paradigms mentioned above, as summarized in Table 2. When compared to these well-curated designs at the cost of introducing additional auxiliary pretext tasks or more training modules, our proposed paradigm achieves a significant improvement of \\(1.2\\%\\) (87.5 vs. 86.3), despite utilizing \\(10\\%\\) of the labeled synthetic data and \\(10\\%\\) of the unlabeled real data. This outcome underscores our ability to harness the potential of unlabeled real images. Moreover, when we gradually increase the proportion of the total training data, the text detection performance continues to improve. Specifically, our method achieves gains of \\(1.8\\%\\), \\(2.0\\%\\), and \\(2.6\\%\\) at data fractions of \\(20\\%\\), \\(50\\%\\), and \\(100\\%\\), respectively, compared to the recent SOTA method. This outcome fully demonstrates the superiority of our pre-training paradigm, which jointly leverage LSD and URD by bridging their gap for enhancing the text detector pre-training." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.522, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "FreeReal for Pre-training STDs" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.186 + ], + "angle": 0, + "content": "Table 2: Comparison with existing pre-training paradigms when fine-tuning on the TotalText dataset. These symbols \\(\\dagger\\), \\(\\ddagger\\), and \\(*\\) indicate the results from [61], [54] and [45], respectively. \\(\\Delta F\\) is the improvement relative to the baseline." + }, + { + "type": "table", + "bbox": [ + 0.266, + 0.2, + 0.734, + 0.365 + ], + "angle": 0, + "content": "
ParadigmMethodVenueDataF-measureΔ F(%)
LSDURD
-DBNet‡ [22]AA AI'20--78.2-
ADBNet+ST† [22]AA AI'20800K-84.7+6.5%
DBNet+TCM† [60]CVPR'23800K-85.9+7.7%
BDBNet+STKM† [43]CVPR'21800K-85.5+7.3%
DBNet+VLPT† [39]CVPR'22800K-86.3+8.1%
DBNet+oCLIP‡ [54]ECCV'22800K-84.1+5.9%
CDBNet+MTM* [45]ACM MM'23800K41K85.7+7.5%
DOurs (9%)-71K41K87.4+9.2%
Ours (10%)-80K46K87.5+9.3%
Ours (20%)-160K92K88.1+9.9%
Ours (50%)-400K230K88.3+10.1%
Ours (100%)-800K460K88.9+10.7%
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.396, + 0.785, + 0.47 + ], + "angle": 0, + "content": "7) Results on Unseen Language Text. Our pre-trained model is capable of learning the basic character feature representations of languages and then combining them during fine-tuning. To show its robustness, we choose to evaluate on unseen language text dataset VinText [31]. FreeReal achieves an improvement of \\(0.7\\%\\), as shown in Table 3." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.499, + 0.468, + 0.516 + ], + "angle": 0, + "content": "5 Ablations and analysis" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.581 + ], + "angle": 0, + "content": "Unless otherwise stated, all ablation experiments are implemented in collaboration with DBNet method [22] using \\(20\\%\\) of LSD and \\(20\\%\\) of URD and evaluated on the curved TotalText (TT) dataset for efficiency." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.583, + 0.786, + 0.793 + ], + "angle": 0, + "content": "Effectiveness of Network Components. To comprehensively evaluate the effectiveness of our proposed method, we divide it into three components and evaluate each one individually, as shown in Table 4. \"Stu-Tea\", \"CRA\", \"GlyphMix\" denote the standard student-to-teacher framework, Character Region Awareness introduced in Sec.3.2 to bridge the linguistic domain gap, and the glyph-based mixing mechanism, respectively. First, we enhance the baseline DBNet model by incorporating a standard student-teacher framework, yielding a performance gain of \\(1.31\\%\\). Second, within the student-teacher framework, we evaluate the individual contributions of the CRA and GlyphMix components, achieving \\(1.73\\%\\) and \\(3.12\\%\\) performance improvements, respectively. Finally, upon combining all plug-ins, the overall performance reaches \\(88.1\\%\\), showing a substantial gain of \\(3.40\\%\\) compared to the baseline. These experiments consistently demonstrate that FreeReal effectively bridges both the synth-to-real and language-to-language domain gaps when leveraging the intrinsic qualities of unlabeled real data." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Discussion on Using WBB or CBB. The SynthText dataset offers word-level and character-level annotations, with the latter being exploited in our method to bridge the linguistic domain gap, bringing a \\(0.4\\%\\) gain as shown in the third row" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.128 + ], + "angle": 0, + "content": "T. Guan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 3: Comparison with supervised SOTA work on unseen language text. \"R50\" and \"DeT\" denote a ResNet50 and Deformable Transformer structure, respectively. * denotes the use of extra labeled real data." + }, + { + "type": "table", + "bbox": [ + 0.314, + 0.2, + 0.686, + 0.242 + ], + "angle": 0, + "content": "
MethodVenueBackbonePara.FPSF
ESTextSpotter*ICCV'23R50+DeT56M3.489.6
DBNet+FreeReal (ours)-R5025M1190.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.259, + 0.784, + 0.286 + ], + "angle": 0, + "content": "Table 4: Evaluate the effectiveness of the proposed modules. \\(\\Delta F\\) is the improvement of F-measure relative to baseline." + }, + { + "type": "table", + "bbox": [ + 0.316, + 0.299, + 0.684, + 0.381 + ], + "angle": 0, + "content": "
Stu-TeaCRAGlyphMixΔ F(%)PRF
×××-87.182.584.7
✓××+1.31%88.8083.3986.01
✓✓×+1.73%87.7685.1586.43
✓×✓+3.12%89.1086.5887.82
✓✓✓+3.40%88.8187.4088.10
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.787, + 0.443 + ], + "angle": 0, + "content": "of Table 4. Notably, even without the use of CRA, our method still achieves an F-measure of \\(87.82\\%\\) (in the fourth row), significantly surpassing recent SOTAs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.788, + 0.61 + ], + "angle": 0, + "content": "Comparison with pre-training methods using URD. Bridging the synth-to-real gap to simultaneously leverage LSD and URD for pre-training is our major contribution. This capability, not present in other PSTD methods, enables seamless incorporation of the huge number of URD to enhance pre-training. Thus to compare to other PSTD methods with complex pretext task designs, we use different ratios of URD to explore the effectiveness of our pre-training paradigm. Even when using \\(10\\%\\) LSD and \\(10\\%\\) URD in Table 2, our method still achieves SOTA with a \\(1.2\\%\\) gain, which exhibits our method can serve as a novel baseline for PSTD. Besides, we add the comparison using additional URD to train other PSTD methods. In this case, these mentioned methods can be viewed the same as MTM [45]. As a result, FreeReal outperforms MTM by \\(1.7\\%\\) in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Effectiveness of GlyphMix Components. In Table 5, GlyphMix comprises two modules, inter-domain GIM and intra-domain TIM, contributing gains of \\(1.28\\%\\) and \\(1.22\\%\\) to the final performance, respectively. Besides, when compared to the data engine tool [12], GIM achieves intra-domain mixing in a cost-effective way, which yields real-world images online via K-means, a more flexible, faster and direct process than the multi-stage, pre-generated method. Specifically, GIM attains a \\(10.7\\mathrm{x}\\) speedup (2.6h vs 27.7h) in generating 92K images, which is more suitable for real-time online training tasks. Additionally, GIM can avoid improper text blends on URD by leveraging the learned knowledge (boxes/maps) from our pre-training framework online. We also conduct the ablations and get \\(0.5\\%\\) \\((85.3\\%)\\) vs \\(84.8\\%\\) ) performance gains, when using \\(5\\%\\) training data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Comparison with Bridging-domain Methods. The goal of generating mixed images through domain-bridging methods is to bring synthetic images closer to real images. Consequently, the adaptation capability can be viewed as the likelihood of mixed images belonging to the real domain. A classification task is em" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.522, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "FreeReal for Pre-training STDs" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 5: Evaluate the effectiveness of GlyphMix components. \"GIM\" and \"TIM\" denote the Graffiti-like Inter-domain Mixing and Text-balanced Intra-domain Mixing." + }, + { + "type": "table", + "bbox": [ + 0.353, + 0.185, + 0.646, + 0.246 + ], + "angle": 0, + "content": "
GIMTIMΔ F(%)PRF
××-87.7685.1586.43
×✓+1.22%88.6486.6887.65
✓×+1.28%88.2987.1387.71
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.26, + 0.784, + 0.289 + ], + "angle": 0, + "content": "Table 6: Comparison of the effectiveness with existing domain bridging methods. \"DCA\" refers to Domain Classification Accuracy in aligning with the real domain." + }, + { + "type": "table", + "bbox": [ + 0.329, + 0.302, + 0.67, + 0.382 + ], + "angle": 0, + "content": "
methodDCAPRFΔ F(%)
Stu-Tea-88.8083.3986.01-
Mix up [65]69.589.1784.1286.57+0.56%
CutMix [62]72.887.9985.6486.80+0.79%
ClassMix [32]78.990.6583.6186.99+0.98%
FreeReal (ours)95.788.8187.4088.10+2.09%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.298, + 0.397, + 0.704, + 0.411 + ], + "angle": 0, + "content": "Table 7: Comparison of the effectiveness with SSL methods." + }, + { + "type": "table", + "bbox": [ + 0.317, + 0.424, + 0.681, + 0.516 + ], + "angle": 0, + "content": "
MethodVenuePRFΔ F(%)
Stu-Tea-88.8083.3986.01-
GTA [16]NeurIPS'2287.1881.0884.02-1.99%
BCP [2]CVPR'2386.7285.8286.27+0.26%
AugSeg [66]CVPR'2388.4885.6587.04+1.03%
Unimatch [55]CVPR'2388.9485.5687.22+1.21%
FreeReal (ours)-88.8187.4088.10+2.09%
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.547, + 0.787, + 0.732 + ], + "angle": 0, + "content": "ployed to assess the capability, whose the evaluation protocol is defined: Given a well-learned domain classifier \\(\\phi\\) (trained by \\(n\\) synthetic images \\(x_{l} \\in X_{L}\\) and \\(n\\) real images \\(x_{u} \\in X_{U}\\)) and domain bridging method \\(\\xi\\), the classifier predicts the domain label (0 for synthetic and 1 for real) of mixed images and evaluates the accuracy as \\(\\frac{\\sum_{x_{l} \\in X_{L}, x_{u} \\in X_{U}} \\phi(\\xi(x_{l}, x_{u}))}{n^{2}}\\). To accomplish this, we employ a network architecture, consisting of resnet18 as a feature extractor, followed by a binary classification head, for simplicity. Subsequently, the performance of different domain bridging methods is summarized in Table 6. Remarkably, when compared to these methods, GlyphMix achieves an outstanding performance score of \\(95.7\\%\\), surpassing the second-best method by a substantial margin of \\(16.8\\%\\). This result demonstrates that our method is able to generate arbitrarily realistic images without domain drift." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.734, + 0.786, + 0.809 + ], + "angle": 0, + "content": "To be more convincing, we also conduct text detection experiments to assess the effectiveness. As depicted in Table 6, our method consistently outperforms others, achieving the best F-measure value. Furthermore, the observed positive correlation between DCA and F-measure reinforces the idea that bridging the domain gap is the crucial factor in improving text detection performance." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.81, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Comparison with Semi-supervised Methods. In this section, we compared the performance with recent SSL methods. These methods tend to adopt in-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "T. Guan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.22 + ], + "angle": 0, + "content": "creasingly complex designs, involving advanced contrastive learning techniques, additional network components, multiple ensemble models, and complex training procedures. In contrast, we aim to eliminate domain gaps between LSD and URD. Despite adopting the standard student-teacher framework, our method still exceeds the previous SOTA by \\(0.88\\%\\) in Table 7." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.222, + 0.784, + 0.403 + ], + "angle": 0, + "content": "Effectiveness of \\(\\mathbf{M}_G\\). We evaluate the accuracy of character segmentation results \\(\\mathbf{M}_G\\) using TotalText [6] and TextSeg [53] datasets, both of which provide pixel-level annotations. Specifically, compared to conducting the cluster task on the whole image, we crop it into several text instances and then perform clustering to form \\(\\mathbf{M}_G\\). While the accuracy of the former is \\(32.35\\%\\) for TotalText and \\(45.54\\%\\) for TextSeg, \\(\\mathbf{M}_G\\) achieves \\(80.43\\%\\) and \\(84.14\\%\\), respectively. Notably, the underlying morphological representations of glyphs exhibit relative invariance to slight structural changes, such as thicker or thinner. This reduces the dependence on pixel-level high-precision segmentation with expensive costs. Furthermore, although the clustering process introduces some noise, leading to inaccuracies in \\(\\mathbf{M}_G\\), it can still be viewed as a way of adding data noise in the image-mixing process, enhancing the robustness of text detection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.784, + 0.493 + ], + "angle": 0, + "content": "Ablation Study on Hyper-parameters. We study the impact of the parameter \\(\\gamma\\) on the pseudos generated by the teacher model during pre-training. This parameter controls the proportion of unreliable and reliable pixels. With \\(\\gamma\\) set at 80, 50, and 20, the F-measure are \\(87.94\\%\\), \\(87.68\\%\\), and \\(87.06\\%\\), respectively. Finally, we employ a linear warm-up strategy for \\(\\gamma\\) that gradually reduces from 80 to 20, yielding the best result with an F-measure of \\(88.10\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.518, + 0.358, + 0.535 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.784, + 0.744 + ], + "angle": 0, + "content": "We propose an effective pre-training paradigm for scene text detection, situated within a standard student-teacher framework. The paradigm leverages the complementary strengths of both LSD and URD, standing out from existing intricate pre-training paradigms in text detection due to its simplicity and significant advancements. Moreover, we identify and address prominent challenges, i.e., synth-to-real and language-to-language domain gaps. The proposed solutions, namely GlyphMix and character region awareness, have shown their effectiveness in mitigating these challenges and improving text detection performance, as substantiated by extensive experimental results. More importantly, our method exhibits the ability to harness the intrinsic attributes of URD, achieving substantial performance improvements even with a small fraction of LSD. This highlights the significance of exploring the potential of unlabeled real data in scene text detection, ultimately advancing the field and its practical applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.746, + 0.793, + 0.836 + ], + "angle": 0, + "content": "Acknowledgment. This work was supported by the NSFC under Grant 62322604 and 62176159, and in part by the Shanghai Municipal Science, Technology Major Project under Grant 2021SHZDZX0102, and National Key R&D Program of China (NO. 2022ZD0160100). The computations in this paper were run on the AI for Science Platform supported by the Artificial Intelligence Institute at Shanghai Jiao Tong University." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.522, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "FreeReal for Pre-training STDs" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.208 + ], + "angle": 0, + "content": "1. Baek, Y., Lee, B., Han, D., Yun, S., Lee, H.: Character region awareness for text detection. In: CVPR. pp. 9365-9374 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.209, + 0.785, + 0.236 + ], + "angle": 0, + "content": "2. Bai, Y., Chen, D., Li, Q., Shen, W., Wang, Y.: Bidirectional copy-paste for semi-supervised medical image segmentation. In: CVPR. pp. 11514-11524 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.237, + 0.785, + 0.277 + ], + "angle": 0, + "content": "3. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV. pp. 9650-9660 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.279, + 0.785, + 0.306 + ], + "angle": 0, + "content": "4. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML. pp. 1597-1607 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.307, + 0.785, + 0.348 + ], + "angle": 0, + "content": "5. Chen, Y., Wang, W., Zhou, Y., Yang, F., Yang, D., Wang, W.: Self-training for domain adaptive scene text detection. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 850-857. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.349, + 0.785, + 0.376 + ], + "angle": 0, + "content": "6. Ch'ng, C.K., Chan, C.S., Liu, C.L.: Total-text: toward orientation robustness in scene text detection. IJDAR 23(1), 31-52 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.377, + 0.785, + 0.404 + ], + "angle": 0, + "content": "7. Grill, J., Strub, F., Altché, F., et al.: Bootstrap your own latent - A new approach to self-supervised learning. In: NeurIPS. pp. 21271-21284 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.785, + 0.446 + ], + "angle": 0, + "content": "8. Guan, T., Gu, C., Lu, C., Tu, J., Feng, Q., Wu, K., Guan, X.: Industrial scene text detection with refined feature-attentive network. IEEE TCSVT 32(9), 6073-6085 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.447, + 0.785, + 0.474 + ], + "angle": 0, + "content": "9. Guan, T., Gu, C., Tu, J., Yang, X., Feng, Q., Zhao, Y., Shen, W.: Self-supervised implicit glyph attention for text recognition. In: CVPR. pp. 15285-15294 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.476, + 0.785, + 0.516 + ], + "angle": 0, + "content": "10. Guan, T., Lin, C., Shen, W., Yang, X.: Posformer: Recognizing complex handwritten mathematical expression with position forest transformer. arXiv preprint arXiv:2407.07764 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.517, + 0.785, + 0.558 + ], + "angle": 0, + "content": "11. Guan, T., Shen, W., Yang, X., Feng, Q., Jiang, Z., Yang, X.: Self-supervised character-to-character distillation for text recognition. In: ICCV. pp. 19473-19484 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.56, + 0.785, + 0.587 + ], + "angle": 0, + "content": "12. Gupta, A., Vedaldi, A., Zisserman, A.: Synthetic data for text localisation in natural images. In: CVPR. pp. 2315-2324 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.588, + 0.785, + 0.615 + ], + "angle": 0, + "content": "13. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR. pp. 16000-16009 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.616, + 0.785, + 0.657 + ], + "angle": 0, + "content": "14. He, M., Liao, M., Yang, Z., Zhong, H., Tang, J., Cheng, W., Yao, C., Wang, Y., Bai, X.: Most: A multi-oriented scene text detector with localization refinement. In: CVPR. pp. 8813-8822 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.658, + 0.785, + 0.699 + ], + "angle": 0, + "content": "15. Huang, M., Zhang, J., Peng, D., Lu, H., Huang, C., Liu, Y., Bai, X., Jin, L.: Estextspotter: Towards better scene text spotting with explicit synergy in transformer. In: ICCV. pp. 19495-19505 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.7, + 0.785, + 0.728 + ], + "angle": 0, + "content": "16. Jin, Y., Wang, J., Lin, D.: Semi-supervised semantic segmentation via gentle teaching assistant. In: NeurIPS. vol. 35, pp. 2803-2816 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.729, + 0.785, + 0.756 + ], + "angle": 0, + "content": "17. Karatzas, D., Gomez-Bigorda, L., Nicolaou, A., et al.: Icdar 2015 competition on robust reading. In: ICDAR. pp. 1156-1160 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.757, + 0.785, + 0.784 + ], + "angle": 0, + "content": "18. Karatzas, D., Shafait, F., Uchida, S., et al.: Icdar 2013 robust reading competition. In: ICDAR. pp. 1484-1493 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.785, + 0.785, + 0.812 + ], + "angle": 0, + "content": "19. Laine, S., Aila, T.: Temporal ensembling for semi-supervised learning. In: arXiv:1610.02242 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "20. Li, W., Gao, H.a., Gao, M., Tian, B., Zhi, R., Zhao, H.: Training-free model merging for multi-target domain adaptation. In: ECCV. Springer (2024)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "T. Guan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.188 + ], + "angle": 0, + "content": "21. Liao, M., Song, B., Long, S., He, M., Yao, C., Bai, X.: Synthtext3d: synthesizing scene text images from 3d virtual worlds. Science China Information Sciences 63, 1-14 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.786, + 0.217 + ], + "angle": 0, + "content": "22. Liao, M., Wan, Z., Yao, C., Chen, K., Bai, X.: Real-time scene text detection with differentiable binarization. In: AAAI. pp. 11474-11481 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.786, + 0.258 + ], + "angle": 0, + "content": "23. Liao, M., Zou, Z., Wan, Z., Yao, C., Bai, X.: Real-time scene text detection with differentiable binarization and adaptive scale fusion. IEEE TPAMI 45(1), 919-931 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.786, + 0.286 + ], + "angle": 0, + "content": "24. Liu, Y., Chen, H., Shen, C., He, T., Jin, L., Wang, L.: Abcnet: Real-time scene text spotting with adaptiveBezier-curve network. In: CVPR. pp. 9809-9818 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.786, + 0.314 + ], + "angle": 0, + "content": "25. Liu, Y., Jin, L., Zhang, S., Luo, C., Zhang, S.: Curved scene text detection via transverse and longitudinal sequence connection. PR 90, 337-345 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.786, + 0.342 + ], + "angle": 0, + "content": "26. Long, S., Ruan, J., Zhang, W., He, X., Wu, W., Yao, C.: Textsnake: A flexible representation for detecting text of arbitrary shapes. In: ECCV. pp. 20-36 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.786, + 0.369 + ], + "angle": 0, + "content": "27. Long, S., Yao, C.: Unrealtext: Synthesizing realistic scene text images from the unreal world. arXiv:2003.10608 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.786, + 0.396 + ], + "angle": 0, + "content": "28. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv:1711.05101 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.786, + 0.438 + ], + "angle": 0, + "content": "29. Nayef, N., Patel, Y., Busta, M., et al.: Icdar2019 robust reading challenge on multilingual scene text detection and recognition—rrc-plt-2019. In: ICDAR. pp. 1582-1587 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.786, + 0.481 + ], + "angle": 0, + "content": "30. Nayef, N., Yin, F., Bizid, I., et al.: Icdar2017 robust reading challenge on multilingual scene text detection and script identification-rrc-mlt. In: ICDAR. vol. 1, pp. 1454-1459 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.786, + 0.508 + ], + "angle": 0, + "content": "31. Nguyen, N., Nguyen, T., Tran, V., et al.: Dictionary-guided scene text recognition. In: CVPR. pp. 7383-7392 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.786, + 0.536 + ], + "angle": 0, + "content": "32. Olsson, V., Tranheden, W., Pinto, J., Svensson, L.: Classmix: Segmentation-based data augmentation for semi-supervised learning. In: WACV. pp. 1369-1378 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.786, + 0.577 + ], + "angle": 0, + "content": "33. Qin, X., Lyu, P., Zhang, C., Zhou, Y., Yao, K., Zhang, P., Lin, H., Wang, W.: Towards robust real-time scene text detection: From semantic to instance representation learning. In: ACM MM. pp. 2025-2034 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.786, + 0.619 + ], + "angle": 0, + "content": "34. Qin, X., Zhou, Y., Guo, Y., Wu, D., Tian, Z., Jiang, N., Wang, H., Wang, W.: Mask is all you need: Rethinking mask r-cnn for dense and arbitrary-shaped scene text detection. In: ACM MM. pp. 414-423 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.786, + 0.674 + ], + "angle": 0, + "content": "35. Qin, X., Zhou, Y., Guo, Y., Wu, D., Wang, W.: Fc 2 rn: a fully convolutional corner refinement network for accurate multi-oriented scene text detection. In: ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 4350-4354. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.786, + 0.702 + ], + "angle": 0, + "content": "36. Radford, A., Kim, J.W., Hallacy, C., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.786, + 0.729 + ], + "angle": 0, + "content": "37. Shi, B., Bai, X., Belongie, S.: Detecting oriented text in natural images by linking segments. In: CVPR. pp. 2550-2558 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.786, + 0.771 + ], + "angle": 0, + "content": "38. Sohn, K., Berthelot, D., Carlini, N., Zhang, Z., Zhang, H., Raffel, C.A., Cubuk, E.D., Kurakin, A., Li, C.L.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In: NeurIPS. vol. 33, pp. 596-608 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.786, + 0.799 + ], + "angle": 0, + "content": "39. Song, S., Wan, J., Yang, Z., Tang, J., Cheng, W., Bai, X., Yao, C.: Vision-language pre-training for boosting scene text detectors. In: CVPR. pp. 15681-15691 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.786, + 0.841 + ], + "angle": 0, + "content": "40. Sun, X., Lyu, J., Zhang, Y., Zeng, G., Fang, B., Zhou, Y., Xie, E., Ma, C.: Feature enhancement with text-specific region contrast for scene text detection. In: Chinese Conference on Pattern Recognition and Computer Vision (PRCV). pp. 3-14 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.522, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "FreeReal for Pre-training STDs" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "41. Sun, Y., Ni, Z., Chng, C.K., et al.: Icdar 2019 competition on large-scale street view text with partial labeling-rrc-lsvt. In: ICDAR. pp. 1557-1562 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.213, + 0.175, + 0.788, + 0.215 + ], + "angle": 0, + "content": "42. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In: NeurIPS. vol. 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.213, + 0.215, + 0.788, + 0.242 + ], + "angle": 0, + "content": "43. Wan, Q., Ji, H., Shen, L.: Self-attention based text knowledge mining for text detection. In: CVPR. pp. 5983-5992 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.213, + 0.242, + 0.788, + 0.27 + ], + "angle": 0, + "content": "44. Wang, F., Chen, Y., Wu, F., Li, X.: Texray: Contour-based geometric modeling for arbitrary-shaped scene text detection. In: ACM MM. pp. 111-119 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.213, + 0.27, + 0.788, + 0.311 + ], + "angle": 0, + "content": "45. Wang, K., Xie, H., Wang, Y., Zhang, D., Qu, Y., Gao, Z., Zhang, Y.: Masked text modeling: A self-supervised pre-training method for scene text detection. In: ACM MM. pp. 2006-2015 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.213, + 0.311, + 0.788, + 0.338 + ], + "angle": 0, + "content": "46. Wang, W., Xie, E., Li, X., Hou, W., Lu, T., Yu, G., Shao, S.: Shape robust text detection with progressive scale expansion network. In: CVPR. pp. 9336-9345 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.213, + 0.338, + 0.788, + 0.378 + ], + "angle": 0, + "content": "47. Wang, W., Xie, E., Li, X., Liu, X., Liang, D., Yang, Z., Lu, T., Shen, C.: Pan++: Towards efficient and accurate end-to-end spotting of arbitrarily-shaped text. IEEE TPAMI 44(9), 5349-5367 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.213, + 0.379, + 0.788, + 0.419 + ], + "angle": 0, + "content": "48. Wang, W., Xie, E., Song, X., Zang, Y., Wang, W., Lu, T., Yu, G., Shen, C.: Efficient and accurate arbitrary-shaped text detection with pixel aggregation network. In: ICCV. pp. 8440-8449 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.213, + 0.42, + 0.788, + 0.459 + ], + "angle": 0, + "content": "49. Wang, X., Jiang, Y., Luo, Z., Liu, C.L., Choi, H., Kim, S.: Arbitrary shape scene text detection with adaptive text region representation. In: CVPR. pp. 6449-6458 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.213, + 0.459, + 0.788, + 0.5 + ], + "angle": 0, + "content": "50. Wang, Y., Xie, H., Zha, Z.J., Xing, M., Fu, Z., Zhang, Y.: Contournet: Taking a further step toward accurate arbitrary-shaped scene text detection. In: CVPR. pp. 11753-11762 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.213, + 0.501, + 0.788, + 0.542 + ], + "angle": 0, + "content": "51. Wu, W., Lu, N., Xie, E., Wang, Y., Yu, W., Yang, C., Zhou, H.: Synthetic-to-real unsupervised domain adaptation for scene text detection in the wild. In: Proceedings of the Asian Conference on Computer Vision (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.213, + 0.542, + 0.788, + 0.582 + ], + "angle": 0, + "content": "52. Xu, H., Liu, L., Bian, Q., Yang, Z.: Semi-supervised semantic segmentation with prototype-based consistency regularization. In: NeurIPS. vol. 35, pp. 26007-26020 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.213, + 0.583, + 0.788, + 0.622 + ], + "angle": 0, + "content": "53. Xu, X., Zhang, Z., Wang, Z., Price, B., Wang, Z., Shi, H.: Rethinking text segmentation: A novel dataset and a text-specific refinement approach. In: CVPR. pp. 12045-12055 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.213, + 0.622, + 0.788, + 0.664 + ], + "angle": 0, + "content": "54. Xue, C., Zhang, W., Hao, Y., Lu, S., Torr, P.H., Bai, S.: Language matters: A weakly supervised vision-language pre-training approach for scene text detection and spotting. In: ECCV. pp. 284-302. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.213, + 0.664, + 0.788, + 0.692 + ], + "angle": 0, + "content": "55. Yang, L., Qi, L., Feng, L., Zhang, W., Shi, Y.: Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In: CVPR. pp. 7236-7246 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.213, + 0.692, + 0.788, + 0.718 + ], + "angle": 0, + "content": "56. Yang, X., Song, Z., King, I., Xu, Z.: A survey on deep semi-supervised learning. IEEE Transactions on Knowledge and Data Engineering 35(9), 8934-8954 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.213, + 0.718, + 0.788, + 0.746 + ], + "angle": 0, + "content": "57. Yao, C., Bai, X., Liu, W., Ma, Y., Tu, Z.: Detecting texts of arbitrary orientations in natural images. In: CVPR. pp. 1083-1090 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.213, + 0.746, + 0.788, + 0.786 + ], + "angle": 0, + "content": "58. Ye, M., Zhang, J., Zhao, S., Liu, J., Liu, T., Du, B., Tao, D.: Deepsolo: Let transformer decoder with explicit points solo for text spotting. In: CVPR. pp. 19348-19357 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.213, + 0.786, + 0.788, + 0.813 + ], + "angle": 0, + "content": "59. Yin, X.C., Pei, W.Y., Zhang, J., Hao, H.W.: Multi-orientation scene text detection with adaptive clustering. IEEE TPAMI 37(9), 1930-1937 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.213, + 0.813, + 0.788, + 0.84 + ], + "angle": 0, + "content": "60. Yu, W., Liu, Y., Hua, W., Jiang, D., Ren, B., Bai, X.: Turning a clip model into a scene text detector. In: CVPR. pp. 6978-6988 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.213, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "T. Guan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "61. Yu, W., Liu, Y., Zhu, X., Cao, H., Sun, X., Bai, X.: Turning a clip model into a scene text spotter. IEEE TPAMI (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.217 + ], + "angle": 0, + "content": "62. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: ICCV. pp. 6023-6032 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "63. Zhan, F., Xue, C., Lu, S.: Ga-dan: Geometry-aware domain adaptation network for scene text detection and recognition. In: ICCV. pp. 9105-9115 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.286 + ], + "angle": 0, + "content": "64. Zhang, B., Wang, Y., Hou, W., Wu, H., Wang, J., Okumura, M., Shinozaki, T.: Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In: NeurIPS. vol. 34, pp. 18408-18419 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "65. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv:1710.09412 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "66. Zhao, Z., Yang, L., Long, S., Pi, J., Zhou, L., Wang, J.: Augmentation matters: A simple-yet-effective approach to semi-supervised semantic segmentation. In: CVPR. pp. 11350-11359 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.384 + ], + "angle": 0, + "content": "67. Zhou, X., Yao, C., Wen, H., Wang, Y., Zhou, S., He, W., Liang, J.: East: an efficient and accurate scene text detector. In: CVPR. pp. 5551-5560 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "68. Zhu, Y., Chen, J., Liang, L., Kuang, Z., Jin, L., Zhang, W.: Fourier contour embedding for arbitrary-shaped text detection. In: CVPR. pp. 3123-3131 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.411 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Bridging Synthetic and Real Worlds for Pre-training Scene Text Detectors/80dda151-f601-4dfe-a0ed-8536e8b997b9_origin.pdf b/2024/Bridging Synthetic and Real Worlds for Pre-training Scene Text Detectors/80dda151-f601-4dfe-a0ed-8536e8b997b9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a275dfbe933466e94bd1abf9f20b45d397396a2f --- /dev/null +++ b/2024/Bridging Synthetic and Real Worlds for Pre-training Scene Text Detectors/80dda151-f601-4dfe-a0ed-8536e8b997b9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a5769955d1d127a1ec176cb561baf22745361c73e78e0e63d6073820b7129fd0 +size 777354 diff --git a/2024/Bridging Synthetic and Real Worlds for Pre-training Scene Text Detectors/full.md b/2024/Bridging Synthetic and Real Worlds for Pre-training Scene Text Detectors/full.md new file mode 100644 index 0000000000000000000000000000000000000000..056663c48c1537f181ade1a862f6d9d08c00d4f9 --- /dev/null +++ b/2024/Bridging Synthetic and Real Worlds for Pre-training Scene Text Detectors/full.md @@ -0,0 +1,296 @@ +# Bridging Synthetic and Real Worlds for Pre-training Scene Text Detectors + +Tongkun Guan $^{1}$ , Wei Shen $^{1(\boxtimes)}$ , Xue Yang $^{2}$ , Xuehui Wang $^{1}$ , and Xiaokang Yang $^{1}$ + +1 MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University {gtk0615,wei.shen}@sjtu.edu.cn + +$^{2}$ Shanghai AI Laboratory yangxue@pjlab.org.cn + +Abstract. Existing scene text detection methods typically rely on extensive real data for training. Due to the lack of annotated real images, recent works have attempted to exploit large-scale labeled synthetic data (LSD) for pre-training text detectors. However, a synth-to-real domain gap emerges, further limiting the performance of text detectors. Differently, in this work, we propose FreeReal, a real-domain-aligned pretraining paradigm that enables the complementary strengths of both LSD and unlabeled real data (URD). Specifically, to bridge real and synthetic worlds for pre-training, a glyph-based mixing mechanism (GlyphMix) is tailored for text images. GlyphMix delineates the character structures of synthetic images and embeds them as graffiti-like units onto real images. Without introducing real domain drift, GlyphMix freely yields real-world images with partial annotations derived from synthetic labels. Furthermore, when given free fine-grained synthetic labels, GlyphMix can effectively bridge the linguistic domain gap stemming from English-dominated LSD to URD in various languages. Without bells and whistles, FreeReal achieves average gains of $1.97\%$ , $3.90\%$ , $3.85\%$ , and $4.56\%$ in improving the performance of FCENet, PSENet, PANet, and DBNet methods, respectively, consistently outperforming previous pre-training methods by a substantial margin across four public datasets. Code will be available at https://github.com/SJTU-DeepVisionLab/FreeReal. + +Keywords: Scene text detection $\cdot$ Pre-training paradigm $\cdot$ Domain-bridging method + +# 1 Introduction + +Scene text detection aims to predict the bounding box or polygon for each arbitrarily shaped text instance within an image. This technology has diverse applications, including online education, instant translation, and product search, owing to the prevalence of text-laden images. However, despite their considerable progress, supervised methods [8,10,22,46,50,68] still struggle with an increasing + +![](images/9d8f019c33ab1a50b0a35f02d6e7fe250242854abac78561c91effbcdec1003a.jpg) +Fig. 1: We present different pre-training paradigms for enhancing text detectors. + +demand for real data with labor-intensive annotations, especially when detecting complicated text images. + +In contrast, recent research efforts [39, 43, 54] have focused on pre-training scene text detectors using large-scale labeled synthetic data (LSD), which leverage both visual and textual knowledge to extract text feature representations, as illustrated in Fig. 1(a). For example, VLPT [39] formulates three cross-modality pretext tasks to enhance the performance of the backbone, and oCLIP [54] focuses on aligning visual and textual information to assist text detectors. Nevertheless, these well-curated designs are limited to using synthetic images due to their reliance on labeled word transcriptions. This limitation will lead to a noticeable synth-to-real domain gap when models are fine-tuned on real data, negatively impacting text detection performance. Therefore, exploring the potential of unlabeled real data (URD) is of great importance as they are readily available. However, the pre-training on URD has seldom been studied. One of the few works [45] attempts to pre-train text detectors on URD via masked image modeling [13] (step 1) and on LSD via supervised learning (step 2), which will re-transfers pre-trained weights from the real domain to the synthetic domain. + +Differently, our objective is to jointly leverage LSD and URD by bridging their gap for enhancing the text detector pre-training, as shown in Fig. 1(b). The primary challenge is how to effectively modulate these URD and LSD without encountering real domain drift. In light of this, we initially investigate the adaptation capabilities of existing domain-bridging methods [32, 62, 65]. This involves training a classifier to distinguish between URD and LSD and subsequently assessing the confidence of these domain-bridging methods in aligning with the real domain. As shown in Fig. 2(b), their highest score is only $78.9\%$ . The observation implies that these domain-bridging methods may introduce intermediate domains and disrupt the real domain. + +To conquer this domain drift issue, we propose FreeReal, a real-domain-aligned pre-training paradigm, leveraging synthetic labels to guide text detector learning on real-domain-aligned images. Specifically, a glyph-based mixing mechanism, named GlyphMix, is tailored for text images to create real-world images with annotations derived from synthetic labels. This involves delineating the glyphs of synthetic images based on their internal structure and embedding + +![](images/1e4e3829745eba0bcbf9d190b1ee69de0e3c64de5186e462dc25dcd57a0715e8.jpg) +Fig. 2: We present the domain classification task and the performances of different domain bridging ways between LSD and URD. In the text detection area, experiments demonstrate that our domain bridging method significantly performs better in aligning with the real domain. This is easy to explain: (I) Mixup may lead to unexpected pixel-wise ambiguities; (II) CutMix may cause the absence of many text regions and incomplete semantics of texts; (III) ClassMix will introduce salient boundary priors; (IV) GlyphMix (ours) effectively preserves the semantic information of text without causing real domain drift and bringing extra boundary priors. + +them as graffiti-like units onto any unlabeled real images. GlyphMix offers several advantages: 1) extracting glyph structures in a cost-effective way without introducing extra boundary priors; 2) leveraging synthetic labels for free while greatly minimizing real domain drift (achieving a confidence score of $95.7\%$ ). + +As a byproduct, FreeReal can bridge the language-to-language gap, when adapting features learned from English-dominated LSD to URD in various languages. Specifically, given free character-level synthetic labels, FreeReal enables the pre-trained model to easily learn character features, since characters are the underlying unit of any language. Subsequently, during fine-tuning, these learned character features are adaptively combined according to the rules (e.g., word-and sentence-level) provided by available benchmarks. + +Building upon these insights, we break the trend of recent pre-training methods for scene text detectors (PSTD) that combine increasingly complex pretext task designs or more training modules to improve performance. Without bells and whistles, we focus on exploring the potential of integrating the LSD and URD. Remarkably, despite using $90\%$ less LSD than recent SOTAs, our method still attains a $1.2\%$ gain, effectively underscoring the superiority of our pretraining paradigm. As shown in Table 1, FreeReal consistently surpasses recent SOTAs by a large margin. The main contributions are summarized as follows: + +- We are the first to simultaneously integrate labeled synthetic data and unlabeled real data for enhancing the pre-training of text detectors. FreeReal shows its superiority and significant performance gains over other complex alternatives, serving as a simple yet strong baseline for future PSTD studies. +- We effectively address the challenges of synthetic-to-real and language-to-language domain adaptation. Without bringing boundary priors and causing real domain drift, we freely harness the strengths of real-world text images with annotations derived from synthetic labels. + +# 2 Related Work + +Scene Text Detection. Existing STD methods delineate text regions by regression-based or segmentation-based representations. The regression-based methods [8, 14,24,34,49,67,68] directly regress the bounding points to enclose text regions, and gradually increase the number of points as required to detect complex scenes, e.g., arbitrary texts. The segmentation-based methods regard texts of arbitrary shapes by grouping pixel-level [22, 23, 40, 48], character-level [1, 9], segment-level [26,37] or contour-level [35,44,50] segmentation results by post-processing. + +Pre-training for Scene Text Detector. In early research, [12, 21, 27] proposed to pre-train scene text detectors by generating synthetic data. These data engine tools follow the same paradigm: they first stack multiple off-the-shelf models to perform semantic segmentation and depth estimation analysis on background images, and then blend pre-processed texts into these images. However, the multi-stage, pre-generated methods incur large computational costs (GPU and CPU resources) and substantial latency, making them unsuitable for online training tasks where immediate data processing is essential. Inspired by recent advancements in pre-training techniques [3, 4, 11, 36], the integration of text and visual cross-modal information into scene text detectors has garnered increasing attention. One representative work involves enhancing the backbone with text transcriptions. For example, STKM [43] proposes a pioneering work that pretrains the backbone through an image-level text recognition task. VLPT [39] formulates three cross-modality pretext tasks to boost better text representations derived from the backbone. Another representative work is to endow the entire detector with pre-trained vision-text knowledge. Inherited from CLIP [36], TCM incorporates the Image and Text encoders into the text detector to refine text regions using cross-modal visual-language priors. A persistent challenge still remains—the synth-to-real domain gap limits the detection performance. + +Semi-Supervised Learning. Due to the scarcity of expensive annotations, semi-supervised learning (SSL) methods are becoming increasingly important. These methods extract knowledge from a reduced amount of annotated images in a supervised manner, and from a larger amount of unlabeled images in an unsupervised way. Commonly, a self-training framework for SSL is employed, which encourages a model (student) to learn unlabeled images from pseudo-labels generated by the up-to-date optimized model (teacher). To avoid overfitting incorrect pseudo-labels, recent advancements have focused on pseudo-label refinement by improving the reliability [38,64] and pseudo-label regularization by cross-view consistency [19, 20, 42, 52]. Specifically, some domain adaption works [5, 51, 63] for STD focus on minimizing domain discrepancies in the feature space, including self-training and adversarial alignment methods. The self-training methods adopt the source data to pre-train a model for generating pseudo labels on target data, then use them to retrain the model. Adversarial alignment approaches aim to align the feature distribution and learn the domain-invariant features between the source and target domains. Recently, they have been replaced by the student-teacher framework, which significantly reduces domain gaps through dual-branch knowledge transfer [56]. + +![](images/17e5610e84a7fe2b5752b87e9a748a27a728c9dad340779b7aa990968d215e99.jpg) +(a) FreeReal + +![](images/15ce7f2392943d4f04eb7bee305daaea034e94cfca3b855001b19fd2cd547267.jpg) +(b) GlyphMix +Fig. 3: Our network pipeline in a student-teacher framework. + +# 3 Methodology + +Overview. Given a labeled synthetic set $\mathcal{D}_l = \{(\mathbf{X}_l^{(i)},\mathcal{B}_l^{(i)})\}_{i = 1}^{N_l}$ with $N_{l}$ samples, $\mathbf{X}_l^{(i)}\in \mathbb{R}^{H\times W\times 3}$ represents an image and $\mathcal{B}_l^{(i)} = \{(p_{lt}^i,p_{lb}^i,p_{rt}^i,p_{rb}^i)\}_{i = 1}^n$ is its corresponding set of bounding boxes. $p$ refers to the values of the vertice along the abscissa and ordinate, and $n$ denotes the number of bounding boxes. $H$ and $W$ are the height and width of the image. Similarly, the unlabeled real set is denoted as $\mathcal{D}_u = \{\mathbf{X}_u^{(i)}\}_{i = 1}^{N_u}$ with $N_{u}$ samples. Our objective is to jointly leverage labeled synthetic data $\mathcal{D}_l$ and unlabeled real data $\mathcal{D}_u$ by bridging their gap for enhancing the pre-training of scene text detectors. + +The student-teacher framework exhibits substantial efficacy due to its dual capacity for knowledge transfer between its two branches, which is typically employed to mitigate the domain gap (e.g., data and multimodal). Accordingly, we customize it for aligning real and synthetic text data, as illustrated in Fig. 3. For notation simplification in the following introduction, we omit the superscript $i$ as mentioned above. + +Specifically, the segmentation-based text detector DBNet [22] is employed as our baseline model, primarily due to its impressive real-time performance, versatility, and practical applicability. In the teacher branch, the teacher model parameterized by $\mathbf{f}_t$ , is responsible for generating truncated-gradient logits $\hat{\mathbf{Y}}_u \in \mathbb{R}^{H \times W}$ from a weakly augmented unlabeled real image $w(\mathbf{X}_u)$ . The pseudo label $\mathbf{Y}_u \in \{0,1\}^{H \times W}$ is then obtained by binarizing the logit map $\hat{\mathbf{Y}}_u$ . In the student branch, the process begins by generating a pixel-wise binary map $\mathbf{Y}_l \in \{0,1\}^{H \times W}$ , where 1 is assigned to locations within all bounding boxes from annotation $\mathcal{B}_l$ , and 0 otherwise. The labeled synthetic image $\mathbf{X}_l$ and the strongly augmented unlabeled real image $s(\mathbf{X}_u)$ are combined to create $\mathbf{X}_g = \mathsf{A}(s(\mathbf{X}_u), \mathbf{X}_l)$ , with $\mathsf{A}(\cdot)$ denoting our proposed GlyphMix mechanism, which will be introduced later. Subsequently, the mixed image $\mathbf{X}_g$ serves as input to train the student model, under the supervision of mixed pseudo-labels $\mathbf{Y}_g = \mathsf{A}(\mathbf{Y}_u, \mathbf{Y}_l)$ , by minimizing loss in the following: + +$$ +\mathcal {L} = \ell_ {D B} \left(\hat {\mathbf {Y}} _ {g}, \mathbf {Y} _ {g}, \mathbf {E}\right), \tag {1} +$$ + +where $\ell_{DB}$ denotes the original loss function provided by DBNet [22] and $\hat{\mathbf{Y}}_g$ represents the prediction logits generated by the student model. The matrix + +$\mathbf{E} \in \mathbb{R}^{H \times W}$ , which indicates positive regions for loss calculation, is set as an all-one matrix by default. + +To filter out the unreliable predictions in mixed pseudo-labels $\mathbf{Y}_g$ , we adaptively adjust the matrix $\mathbf{E}$ during pre-training. Specifically, for the labeled synthetic image $\mathbf{X}_l$ , the positive regions $\mathbf{E}_l$ are an all-one matrix. For the unlabeled real image $\mathbf{X}_u$ , we compute the pixel-wise entropy map on $\hat{\mathbf{Y}}_u$ and then apply a dynamic entropy threshold $\zeta$ to obtain its positive regions $\mathbf{E}_u$ as follows: + +$$ +\mathbf {E} _ {u} = \mathbb {I} \left(- \hat {\mathbf {Y}} _ {u} \log \hat {\mathbf {Y}} _ {u} \leq \zeta\right), \tag {2} +$$ + +where $\mathbb{I}(\cdot)$ denotes the indicator function. The matrix is subsequently formulated as $\mathbf{E} = \mathsf{A}(\mathbf{E}_u,\mathbf{E}_l)$ . In our experiments, the threshold $\zeta$ is determined by the value corresponding to the top $\gamma$ percent of the entropy map. As the reliability of pseudo-label gradually improves, we smoothly update $\gamma$ using np.linspace(80, 20, t), where $t$ denotes the total number of training iterations. + +Finally, the teacher model $\mathbf{f}_t$ gradually updates itself via the exponential moving averaging (EMA) of the student weights $\mathbf{f}_s$ as follows: + +$$ +\mathbf {f} _ {t} = \alpha \mathbf {f} _ {t} + (1 - \alpha) \mathbf {f} _ {s}, \tag {3} +$$ + +where $\alpha$ is a momentum parameter, set as 0.996 by default. + +# 3.1 GlyphMix + +Although some efforts [32, 62, 65] attempt to aggregate the advantages of synthetic and unlabeled images through versatile transformations, they inevitably introduce intermediate domains, leading to real domain drift. In this section, we present a domain-bridging way tailored for text images, allowing for the free use of synthetic labels while ensuring consistent alignment with the real domain. + +1) Graffiti-like Inter-domain Mixing. Given an input synthetic image $\mathbf{X}_l$ , its corresponding bounding box annotations $\mathcal{B}_l$ and converted pixel-wise binary map $\mathbf{Y}_l$ , our goal is to perform character segmentation. This process identifies all character structure regions and produces a mask $\mathbf{M}_G$ . + +To achieve this, we begin with an unsupervised clustering task based on the grayscale values of pixels, avoiding the introduction of extra pixel-level annotation costs. This involves aligning and cropping the image according to the coordinates of each bounding box, yielding a set of image patches $\mathcal{T} = \{\mathbf{T}^{(i)}|\mathbf{T}^{(i)}\in \mathbb{R}^{w_i\times h_i\times 3}\}_{i = 1}^n$ . Subsequently, within each image patch $\mathbf{T}^{(i)}$ , we cluster the pixels into two regions: the glyph region representing character structures and the background region. For simplicity, we focus our study on K-means to implement the clustering task, but other clustering approaches with predefined categories can be used. Following this, the glyph regions from all image patches are pasted to an all-zero image to create the character structure mask $\mathbf{M}_G$ . + +Given the pseudo label $\mathbf{Y}_u$ produced by the teacher model on an unlabeled real image $\mathbf{X}_u$ , we use $\mathbf{M}_G$ to generate the inter-domain mixing image and its corresponding label. This process can be formulated as follows: + +$$ +\left\{ \begin{array}{l} \mathbb {A} ^ {G} \left(\mathbf {X} _ {u}, \mathbf {X} _ {l}\right) \leftarrow \mathbf {M} _ {G} \odot \mathbf {X} _ {l} + \left(\mathbf {1} - \mathbf {M} _ {G}\right) \odot \mathbf {X} _ {u}, \\ \mathbb {A} ^ {G} \left(\mathbf {Y} _ {u}, \mathbf {Y} _ {l}\right) \leftarrow \mathbf {M} _ {G} \odot \mathbf {Y} _ {l} + \left(\mathbf {1} - \mathbf {M} _ {G}\right) \odot \mathbf {Y} _ {u}, \end{array} \right. \tag {4} +$$ + +where $\mathbf{1}$ is an all-one matrix and $\odot$ denotes the element-wise multiplication operation. $\mathbf{M}_G$ indicates pixels occupied by the glyph need to be copied from the synth domain and pasted to the real domain. Thus the mixed image closely aligns with the real domain, taking advantage of the freely available synthetic labels. Put simply, this process produces real-world images with partial annotations derived from synthetic labels. + +2) Text-balanced Intra-domain Mixing. Real-world text images inherently exhibit complexity, and the collected unlabeled data often struggles to comprehensively represent the full spectrum of real-world scenarios. Within a student-teacher framework, intra-domain mixing serves as an effective perturbation technique, enriching the diversity of real data and enhancing feature consistency and robustness. However, when randomly generating and applying binary masks to construct intra-domain mixing images and their corresponding labels, there is a risk of missing many text regions, especially considering that text regions typically occupy only a small fraction of an image. To mitigate this challenge, our objective is to retain more text regions while still maintaining randomness. + +In light of this, we initiate the process by randomly selecting two unlabeled real images, denoted as $(\mathbf{X}_u^{(1)},\mathbf{X}_u^{(2)})$ , and obtaining the corresponding segmentation pseudo-labels, $(\mathbf{Y}_u^{(1)},\mathbf{Y}_u^{(2)})$ , produced by the teacher model. Subsequently, $k$ rectangular boxes $\{(p_{lt}^i,p_{lb}^i,p_{rt}^i,p_{rb}^i)\}_{i = 1}^k$ are randomly generated to crop the pseudo-label $\mathbf{Y}_u^{(2)}$ based on their coordinates. This operation yields a set of local pseudo-label regions $S = \{\mathbf{S}^{(i)}|\mathbf{S}^{(i)}\in \mathbb{R}^{w_i\times h_i}\}_{i = 1}^k$ . Following this, we identify the rectangular box that contains the maximum text regions by calculating $\arg \max_{i}\sum_{j = 1}^{w_{i}\times h_{i}}\mathbf{S}_{j}^{(i)}$ , and convert it into a binary mask, denoted as $\mathbf{M}_T$ . Finally, we employ $\mathbf{M}_T$ to generate the intra-domain mixed image and its corresponding label, following the formulated procedure: + +$$ +\left\{ \begin{array}{l} \mathbf {A} ^ {T} (\mathbf {X} _ {u}) \leftarrow \mathbf {M} _ {T} \odot \mathbf {X} _ {u} ^ {(2)} + \left(\mathbf {1} - \mathbf {M} _ {T}\right) \odot \mathbf {X} _ {u} ^ {(1)}, \\ \mathbf {A} ^ {T} (\mathbf {Y} _ {u}) \leftarrow \mathbf {M} _ {T} \odot \mathbf {Y} _ {u} ^ {(2)} + \left(\mathbf {1} - \mathbf {M} _ {T}\right) \odot \mathbf {Y} _ {u} ^ {(1)}, \end{array} \right. \tag {5} +$$ + +where $\mathbf{1}$ is an all-one matrix and $\odot$ represents the element-wise multiplication. + +3) Intra- and Inter-domain Mixing. When these two strategies, tailored for scene text images, are integrated within the student-teacher framework, they collectively form a comprehensive method that effectively harnesses the strengths of both synthetic and real data. The whole process can be formulated as: + +$$ +\left\{ \begin{array}{l} \mathbf {A} (\mathbf {X}) \leftarrow \mathbf {A} ^ {G} \left(\mathbf {A} ^ {T} \left(\mathbf {X} _ {u}\right), \mathbf {X} _ {l}\right), \\ \mathbf {A} (\mathbf {Y}) \leftarrow \mathbf {A} ^ {G} \left(\mathbf {A} ^ {T} \left(\mathbf {Y} _ {u}\right), \mathbf {Y} _ {l}\right) \end{array} \right. \tag {6} +$$ + +Consequently, the synergy maximizes the utilization of synthetic labels as a valuable resource, effectively tackling the intricacies of real-world scenarios. + +# 3.2 Discussion on Character Region Awareness + +Existing LSD only contains English texts with tightly spaced characters, while URD may encompass text properties from various languages (such as Japanese, Korean, Chinese, etc). Consequently, even though the model predictions effec + +tively cover text regions, we encounter challenges in parsing these predictions into word-level bounding boxes or polygons. + +Indeed, when using a single language (e.g., English) as the source domain and adapting it to various language domains, it appears that no matter how sophisticated the models and loss functions are, they struggle to accurately and effectively parse the complex text layout and semantic context. This challenge primarily arises from variations in the spacing and arrangement of characters, words, and sentences across languages. This observation leads us to revisit the learning objective of scene text detectors during pre-training. + +We emphasize an alternative perspective — utilizing characters as the fundamental learning unit for text detection. The strategy presents several advantages: 1) Like word-level bounding boxes (WBB), character-level bounding box (CBB) annotations can be readily acquired from the existing synthetic dataset (SynthText) at no additional cost; 2) Instead of parsing complex layouts and spacing disparities across languages during pre-training, decoding them into individual characters is straightforward and attainable; 3) The pre-trained model with character-level region awareness easily integrates its learned character features to adapt text attribute rules provided by different language text datasets during fine-tuning. To this end, we select CBB annotations as $\mathcal{B}_l$ . + +# 4 Experiment + +# 4.1 Dataset + +Labeled Synthetic Data (LSD). The SynthText [12] serves as our source of LSD, consisting of 800K images. Each image is provided with character-level and word-level bounding boxes, along with corresponding text transcriptions. + +Unlabeled Real Data (URD). URD is collected from a wide array of natural scenes, encompassing a diverse range of fonts, languages, and text shapes. Specifically, URD includes 430K images from LSVT [41], 9K from MLT17 [30], 19657 from MLT19 [29], 229 from ICDAR2013 [18], and 500 from USTB-SV1K [59]. + +**Benchmarks.** Four publicly available benchmarks are used to evaluate the performance, including 1) Total-Text (TT) [6], 2) CTW 1500 (CTW) [25], 3) IC-DAR2015 (IC15) [17], and 4) MSRA-TD500 (TD) [57]. + +# 4.2 Implementation Details + +Referring to recent works [39, 43, 54], we combine representative text detection methods [22, 46, 48, 68] to conduct experiments. URD is used in Tabs. 1-3 and the same data is used in Tabs. 4-7. All experiments are implemented on a server with 2 NVIDIA A800 GPUs, using PyTorch. 1) Pre-Training. The pre-training experiments are conducted on LSD and URD, utilizing the open-source codes provided by these text detection methods. For synthetic images, we adopt consistent augmentation settings as prescribed by these methods for a fair comparison. For unlabeled real images, we employ a standard weak-to-strong augmentation + +technique [66] to enrich diversity. Our method is trained using the AdamW optimizer [28], along with a cosine learning rate scheduler [7]. This includes a warm-up phrase for 1 epoch in a total of 10 epochs. The base learning rate is set to $0.003 \times$ batchsize/256. At each iteration, we equally sample LSD and URD with a total batch size of 24. 2) Fine-Tuning. For a fair comparison, we employ the same fine-tuning configurations as these detection methods and then evaluate them on these publicly available benchmarks: TT, CTW, IC15 and TD, following the standard metric: Precision (P), Recall (R), and F-measure (F). + +# 4.3 Experiment Results + +To ensure a comprehensive comparison, we collect results from public evaluation benchmarks reported by existing pre-training methods, and organize them in Table 1. The sources of these results are clearly recorded in the table's caption using different symbols for easy reference. Specifically, these methods can be classified into three pre-training paradigms. In Paradigm A, models (i.e., $+ST$ , $+TCM$ ) are initially pre-trained on LSD and then fine-tuned on real datasets directly. In Paradigm B, models (i.e., $+STKM$ , $+VLPT$ , $+oCLIP$ ) are first pretrained on LSD by various pretext tasks, and the well-learned backbone weights are transferred into scene text detectors when fine-tuning on real datasets. In Paradigm C, models (i.e., $+MTM$ ) are pre-trained on URD by pretext tasks and on LSD by supervised learning, and then fine-tuned on real datasets. + +1) Results on DBNet. As indicated in the second group of rows in Table 1, all pre-training methods both report their cooperative performance with DBNet. When compared to methods using paradigm A, differing from $DBNet + ST$ , we exclude the use of a deformable convolutional network (DCN) in our model. Despite these differences, our method still achieves impressive results, yielding improvements of $4.2\%$ , $4.5\%$ , $4.6\%$ , and $5.2\%$ on TT, CTW, IC15, and TD datasets, respectively. Compared to $DBNet + TCM$ methods, our method achieves SOTA detection performance on four datasets, particularly on curved text datasets (TT and CTW), with gains of $3.0\%$ and $2.8\%$ , respectively. In comparison with methods following Paradigm B, our method consistently outperforms previous SOTA works by substantial margins, achieving performance improvements of $2.6\%$ , $5.9\%$ , $3.5\%$ and $1.6\%$ on TT, CTW, IC15, and TD datasets, respectively. In comparison with $DBNet + MTM$ following Paradigm C, our method gets $3.2\%$ and $4.5\%$ gains on TT and TD datasets, respectively. +2) Results on PANet. As shown in the third group of rows in Table 1, our method outperforms $PANet + ST$ on four benchmarks, achieving a gain of $3.85\%$ in terms of average F-measure, while using the same parameters and labeled data for pre-training and fine-tuning. Furthermore, our method exhibits its prominent superiority, particularly on the CTW and TD datasets, surpassing the $PANet + TCM$ method by a substantial margin of $4.1\%$ and $4.6\%$ , respectively. +3) Results on PSENet. FreeReal brings $0.7\%$ , $2.8\%$ , and $0.9\%$ gains on TT, CTW, and IC15 datasets, respectively. +4) Results on FCENet. FreeReal further refreshes the best text detection results and improves $2.1\%$ , $1.3\%$ , $0.6\%$ and $2.1\%$ on TT, CTW, IC15, and TD + +Table 1: Comparison with SOTA pre-training methods for text detectors on TT, CTW, IC15, and TD. F-measure (\%) is reported. * denotes the use of extra labeled real data. For easy reference, we mark these symbols $\dagger$ , $*$ , $\ddagger$ , and $*$ separately to indicate the results from [61], [39], [54], and [45], within the scope of this paper. The best results are shown in bold font. The second-best results are underlined. + +
IndexMethodVenueTTCTWIC15TD
1PSENet-1s [46]CVPR'1980.982.285.7-
PAN++ [47]PAMI'2185.384.084.587.9
DBNet++ [23]PAMI'2286.085.387.387.2
DeepSOLO* [58]CVPR'2387.6-89.8-
ESTextSpotter* [15]ICCV'2390.0-91.089.5
SIR [33]ACMMM'2388.285.587.889.6
2DBNet+ST† [22]AAAI'2084.783.485.484.9
DBNet+STKM† [43]CVPR'2185.5-86.185.9
DBNet+VLPT† [39]CVPR'2286.3-86.588.5
DBNet+oCLIP† [54]ECCV'2284.182.085.4-
DBNet+TCM† [60]CVPR'2385.985.189.488.8
DBNet+MTM* [45]ACMMM'2385.7--85.6
DBNet+FreeReal (ours)-88.9(+2.6)87.9(+2.8)90.0(+0.6)90.1(+1.3)
3PANet+ST† [48]ICCV'1985.083.782.984.1
PANet+TCM† [60]CVPR'23-84.384.685.3
PANet+FreeReal (ours)-87.4(+2.4)88.4(+4.1)86.0(+1.4)89.9(+4.6)
4PSENet+ST* [46]CVPR'1983.981.681.3-
PSENet+STKM* [43]CVPR'2184.382.983.7-
PSENet+VLPT* [39]CVPR'2286.183.384.3-
PSENet+oCLIP† [54]ECCV'2285.582.8--
PSENet+FreeReal (ours)-86.8(+0.7)86.1(+2.8)85.2(+0.9)86.0
5FCENet+ST† [68]CVPR'2185.885.586.285.4
FCENet+oCLIP† [54]ECCV'22-85.686.7-
FCENet+TCM† [60]CVPR'23-85.987.186.9
FCENet+FreeReal (ours)-87.9(+2.1)87.2(+1.3)87.7(+0.6)89.0(+2.1)
+ +datasets, respectively. These experimental results demonstrate the robustness of FreeReal, even when cooperated with regression-based methods. + +5) Paradigm Comparison. In this section, we aim to investigate which kinds of pre-training paradigms for text detectors are important and meaningful for future studies. To this end, we conduct a comparative analysis with three mainstream pre-training paradigms mentioned above, as summarized in Table 2. When compared to these well-curated designs at the cost of introducing additional auxiliary pretext tasks or more training modules, our proposed paradigm achieves a significant improvement of $1.2\%$ (87.5 vs. 86.3), despite utilizing $10\%$ of the labeled synthetic data and $10\%$ of the unlabeled real data. This outcome underscores our ability to harness the potential of unlabeled real images. Moreover, when we gradually increase the proportion of the total training data, the text detection performance continues to improve. Specifically, our method achieves gains of $1.8\%$ , $2.0\%$ , and $2.6\%$ at data fractions of $20\%$ , $50\%$ , and $100\%$ , respectively, compared to the recent SOTA method. This outcome fully demonstrates the superiority of our pre-training paradigm, which jointly leverage LSD and URD by bridging their gap for enhancing the text detector pre-training. + +Table 2: Comparison with existing pre-training paradigms when fine-tuning on the TotalText dataset. These symbols $\dagger$ , $\ddagger$ , and $*$ indicate the results from [61], [54] and [45], respectively. $\Delta F$ is the improvement relative to the baseline. + +
ParadigmMethodVenueDataF-measureΔ F(%)
LSDURD
-DBNet‡ [22]AA AI'20--78.2-
ADBNet+ST† [22]AA AI'20800K-84.7+6.5%
DBNet+TCM† [60]CVPR'23800K-85.9+7.7%
BDBNet+STKM† [43]CVPR'21800K-85.5+7.3%
DBNet+VLPT† [39]CVPR'22800K-86.3+8.1%
DBNet+oCLIP‡ [54]ECCV'22800K-84.1+5.9%
CDBNet+MTM* [45]ACM MM'23800K41K85.7+7.5%
DOurs (9%)-71K41K87.4+9.2%
Ours (10%)-80K46K87.5+9.3%
Ours (20%)-160K92K88.1+9.9%
Ours (50%)-400K230K88.3+10.1%
Ours (100%)-800K460K88.9+10.7%
+ +7) Results on Unseen Language Text. Our pre-trained model is capable of learning the basic character feature representations of languages and then combining them during fine-tuning. To show its robustness, we choose to evaluate on unseen language text dataset VinText [31]. FreeReal achieves an improvement of $0.7\%$ , as shown in Table 3. + +# 5 Ablations and analysis + +Unless otherwise stated, all ablation experiments are implemented in collaboration with DBNet method [22] using $20\%$ of LSD and $20\%$ of URD and evaluated on the curved TotalText (TT) dataset for efficiency. + +Effectiveness of Network Components. To comprehensively evaluate the effectiveness of our proposed method, we divide it into three components and evaluate each one individually, as shown in Table 4. "Stu-Tea", "CRA", "GlyphMix" denote the standard student-to-teacher framework, Character Region Awareness introduced in Sec.3.2 to bridge the linguistic domain gap, and the glyph-based mixing mechanism, respectively. First, we enhance the baseline DBNet model by incorporating a standard student-teacher framework, yielding a performance gain of $1.31\%$ . Second, within the student-teacher framework, we evaluate the individual contributions of the CRA and GlyphMix components, achieving $1.73\%$ and $3.12\%$ performance improvements, respectively. Finally, upon combining all plug-ins, the overall performance reaches $88.1\%$ , showing a substantial gain of $3.40\%$ compared to the baseline. These experiments consistently demonstrate that FreeReal effectively bridges both the synth-to-real and language-to-language domain gaps when leveraging the intrinsic qualities of unlabeled real data. + +Discussion on Using WBB or CBB. The SynthText dataset offers word-level and character-level annotations, with the latter being exploited in our method to bridge the linguistic domain gap, bringing a $0.4\%$ gain as shown in the third row + +Table 3: Comparison with supervised SOTA work on unseen language text. "R50" and "DeT" denote a ResNet50 and Deformable Transformer structure, respectively. * denotes the use of extra labeled real data. + +
MethodVenueBackbonePara.FPSF
ESTextSpotter*ICCV'23R50+DeT56M3.489.6
DBNet+FreeReal (ours)-R5025M1190.3
+ +Table 4: Evaluate the effectiveness of the proposed modules. $\Delta F$ is the improvement of F-measure relative to baseline. + +
Stu-TeaCRAGlyphMixΔ F(%)PRF
×××-87.182.584.7
✓××+1.31%88.8083.3986.01
✓✓×+1.73%87.7685.1586.43
✓×✓+3.12%89.1086.5887.82
✓✓✓+3.40%88.8187.4088.10
+ +of Table 4. Notably, even without the use of CRA, our method still achieves an F-measure of $87.82\%$ (in the fourth row), significantly surpassing recent SOTAs. + +Comparison with pre-training methods using URD. Bridging the synth-to-real gap to simultaneously leverage LSD and URD for pre-training is our major contribution. This capability, not present in other PSTD methods, enables seamless incorporation of the huge number of URD to enhance pre-training. Thus to compare to other PSTD methods with complex pretext task designs, we use different ratios of URD to explore the effectiveness of our pre-training paradigm. Even when using $10\%$ LSD and $10\%$ URD in Table 2, our method still achieves SOTA with a $1.2\%$ gain, which exhibits our method can serve as a novel baseline for PSTD. Besides, we add the comparison using additional URD to train other PSTD methods. In this case, these mentioned methods can be viewed the same as MTM [45]. As a result, FreeReal outperforms MTM by $1.7\%$ in Table 2. + +Effectiveness of GlyphMix Components. In Table 5, GlyphMix comprises two modules, inter-domain GIM and intra-domain TIM, contributing gains of $1.28\%$ and $1.22\%$ to the final performance, respectively. Besides, when compared to the data engine tool [12], GIM achieves intra-domain mixing in a cost-effective way, which yields real-world images online via K-means, a more flexible, faster and direct process than the multi-stage, pre-generated method. Specifically, GIM attains a $10.7\mathrm{x}$ speedup (2.6h vs 27.7h) in generating 92K images, which is more suitable for real-time online training tasks. Additionally, GIM can avoid improper text blends on URD by leveraging the learned knowledge (boxes/maps) from our pre-training framework online. We also conduct the ablations and get $0.5\%$ $(85.3\%)$ vs $84.8\%$ ) performance gains, when using $5\%$ training data. + +Comparison with Bridging-domain Methods. The goal of generating mixed images through domain-bridging methods is to bring synthetic images closer to real images. Consequently, the adaptation capability can be viewed as the likelihood of mixed images belonging to the real domain. A classification task is em + +Table 5: Evaluate the effectiveness of GlyphMix components. "GIM" and "TIM" denote the Graffiti-like Inter-domain Mixing and Text-balanced Intra-domain Mixing. + +
GIMTIMΔ F(%)PRF
××-87.7685.1586.43
×✓+1.22%88.6486.6887.65
✓×+1.28%88.2987.1387.71
+ +Table 6: Comparison of the effectiveness with existing domain bridging methods. "DCA" refers to Domain Classification Accuracy in aligning with the real domain. + +
methodDCAPRFΔ F(%)
Stu-Tea-88.8083.3986.01-
Mix up [65]69.589.1784.1286.57+0.56%
CutMix [62]72.887.9985.6486.80+0.79%
ClassMix [32]78.990.6583.6186.99+0.98%
FreeReal (ours)95.788.8187.4088.10+2.09%
+ +Table 7: Comparison of the effectiveness with SSL methods. + +
MethodVenuePRFΔ F(%)
Stu-Tea-88.8083.3986.01-
GTA [16]NeurIPS'2287.1881.0884.02-1.99%
BCP [2]CVPR'2386.7285.8286.27+0.26%
AugSeg [66]CVPR'2388.4885.6587.04+1.03%
Unimatch [55]CVPR'2388.9485.5687.22+1.21%
FreeReal (ours)-88.8187.4088.10+2.09%
+ +ployed to assess the capability, whose the evaluation protocol is defined: Given a well-learned domain classifier $\phi$ (trained by $n$ synthetic images $x_{l} \in X_{L}$ and $n$ real images $x_{u} \in X_{U}$ ) and domain bridging method $\xi$ , the classifier predicts the domain label (0 for synthetic and 1 for real) of mixed images and evaluates the accuracy as $\frac{\sum_{x_{l} \in X_{L}, x_{u} \in X_{U}} \phi(\xi(x_{l}, x_{u}))}{n^{2}}$ . To accomplish this, we employ a network architecture, consisting of resnet18 as a feature extractor, followed by a binary classification head, for simplicity. Subsequently, the performance of different domain bridging methods is summarized in Table 6. Remarkably, when compared to these methods, GlyphMix achieves an outstanding performance score of $95.7\%$ , surpassing the second-best method by a substantial margin of $16.8\%$ . This result demonstrates that our method is able to generate arbitrarily realistic images without domain drift. + +To be more convincing, we also conduct text detection experiments to assess the effectiveness. As depicted in Table 6, our method consistently outperforms others, achieving the best F-measure value. Furthermore, the observed positive correlation between DCA and F-measure reinforces the idea that bridging the domain gap is the crucial factor in improving text detection performance. + +Comparison with Semi-supervised Methods. In this section, we compared the performance with recent SSL methods. These methods tend to adopt in- + +creasingly complex designs, involving advanced contrastive learning techniques, additional network components, multiple ensemble models, and complex training procedures. In contrast, we aim to eliminate domain gaps between LSD and URD. Despite adopting the standard student-teacher framework, our method still exceeds the previous SOTA by $0.88\%$ in Table 7. + +Effectiveness of $\mathbf{M}_G$ . We evaluate the accuracy of character segmentation results $\mathbf{M}_G$ using TotalText [6] and TextSeg [53] datasets, both of which provide pixel-level annotations. Specifically, compared to conducting the cluster task on the whole image, we crop it into several text instances and then perform clustering to form $\mathbf{M}_G$ . While the accuracy of the former is $32.35\%$ for TotalText and $45.54\%$ for TextSeg, $\mathbf{M}_G$ achieves $80.43\%$ and $84.14\%$ , respectively. Notably, the underlying morphological representations of glyphs exhibit relative invariance to slight structural changes, such as thicker or thinner. This reduces the dependence on pixel-level high-precision segmentation with expensive costs. Furthermore, although the clustering process introduces some noise, leading to inaccuracies in $\mathbf{M}_G$ , it can still be viewed as a way of adding data noise in the image-mixing process, enhancing the robustness of text detection. + +Ablation Study on Hyper-parameters. We study the impact of the parameter $\gamma$ on the pseudos generated by the teacher model during pre-training. This parameter controls the proportion of unreliable and reliable pixels. With $\gamma$ set at 80, 50, and 20, the F-measure are $87.94\%$ , $87.68\%$ , and $87.06\%$ , respectively. Finally, we employ a linear warm-up strategy for $\gamma$ that gradually reduces from 80 to 20, yielding the best result with an F-measure of $88.10\%$ . + +# 6 Conclusion + +We propose an effective pre-training paradigm for scene text detection, situated within a standard student-teacher framework. The paradigm leverages the complementary strengths of both LSD and URD, standing out from existing intricate pre-training paradigms in text detection due to its simplicity and significant advancements. Moreover, we identify and address prominent challenges, i.e., synth-to-real and language-to-language domain gaps. The proposed solutions, namely GlyphMix and character region awareness, have shown their effectiveness in mitigating these challenges and improving text detection performance, as substantiated by extensive experimental results. More importantly, our method exhibits the ability to harness the intrinsic attributes of URD, achieving substantial performance improvements even with a small fraction of LSD. This highlights the significance of exploring the potential of unlabeled real data in scene text detection, ultimately advancing the field and its practical applications. + +Acknowledgment. This work was supported by the NSFC under Grant 62322604 and 62176159, and in part by the Shanghai Municipal Science, Technology Major Project under Grant 2021SHZDZX0102, and National Key R&D Program of China (NO. 2022ZD0160100). The computations in this paper were run on the AI for Science Platform supported by the Artificial Intelligence Institute at Shanghai Jiao Tong University. + +# References + +1. Baek, Y., Lee, B., Han, D., Yun, S., Lee, H.: Character region awareness for text detection. In: CVPR. pp. 9365-9374 (2019) +2. Bai, Y., Chen, D., Li, Q., Shen, W., Wang, Y.: Bidirectional copy-paste for semi-supervised medical image segmentation. In: CVPR. pp. 11514-11524 (2023) +3. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV. pp. 9650-9660 (2021) +4. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML. pp. 1597-1607 (2020) +5. Chen, Y., Wang, W., Zhou, Y., Yang, F., Yang, D., Wang, W.: Self-training for domain adaptive scene text detection. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 850-857. IEEE (2021) +6. Ch'ng, C.K., Chan, C.S., Liu, C.L.: Total-text: toward orientation robustness in scene text detection. IJDAR 23(1), 31-52 (2020) +7. Grill, J., Strub, F., Altché, F., et al.: Bootstrap your own latent - A new approach to self-supervised learning. In: NeurIPS. pp. 21271-21284 (2020) +8. Guan, T., Gu, C., Lu, C., Tu, J., Feng, Q., Wu, K., Guan, X.: Industrial scene text detection with refined feature-attentive network. IEEE TCSVT 32(9), 6073-6085 (2022) +9. Guan, T., Gu, C., Tu, J., Yang, X., Feng, Q., Zhao, Y., Shen, W.: Self-supervised implicit glyph attention for text recognition. In: CVPR. pp. 15285-15294 (2023) +10. Guan, T., Lin, C., Shen, W., Yang, X.: Posformer: Recognizing complex handwritten mathematical expression with position forest transformer. arXiv preprint arXiv:2407.07764 (2024) +11. Guan, T., Shen, W., Yang, X., Feng, Q., Jiang, Z., Yang, X.: Self-supervised character-to-character distillation for text recognition. In: ICCV. pp. 19473-19484 (2023) +12. Gupta, A., Vedaldi, A., Zisserman, A.: Synthetic data for text localisation in natural images. In: CVPR. pp. 2315-2324 (2016) +13. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR. pp. 16000-16009 (2022) +14. He, M., Liao, M., Yang, Z., Zhong, H., Tang, J., Cheng, W., Yao, C., Wang, Y., Bai, X.: Most: A multi-oriented scene text detector with localization refinement. In: CVPR. pp. 8813-8822 (2021) +15. Huang, M., Zhang, J., Peng, D., Lu, H., Huang, C., Liu, Y., Bai, X., Jin, L.: Estextspotter: Towards better scene text spotting with explicit synergy in transformer. In: ICCV. pp. 19495-19505 (2023) +16. Jin, Y., Wang, J., Lin, D.: Semi-supervised semantic segmentation via gentle teaching assistant. In: NeurIPS. vol. 35, pp. 2803-2816 (2022) +17. Karatzas, D., Gomez-Bigorda, L., Nicolaou, A., et al.: Icdar 2015 competition on robust reading. In: ICDAR. pp. 1156-1160 (2015) +18. Karatzas, D., Shafait, F., Uchida, S., et al.: Icdar 2013 robust reading competition. In: ICDAR. pp. 1484-1493 (2013) +19. Laine, S., Aila, T.: Temporal ensembling for semi-supervised learning. In: arXiv:1610.02242 (2016) +20. Li, W., Gao, H.a., Gao, M., Tian, B., Zhi, R., Zhao, H.: Training-free model merging for multi-target domain adaptation. In: ECCV. Springer (2024) + +21. Liao, M., Song, B., Long, S., He, M., Yao, C., Bai, X.: Synthtext3d: synthesizing scene text images from 3d virtual worlds. Science China Information Sciences 63, 1-14 (2020) +22. Liao, M., Wan, Z., Yao, C., Chen, K., Bai, X.: Real-time scene text detection with differentiable binarization. In: AAAI. pp. 11474-11481 (2020) +23. Liao, M., Zou, Z., Wan, Z., Yao, C., Bai, X.: Real-time scene text detection with differentiable binarization and adaptive scale fusion. IEEE TPAMI 45(1), 919-931 (2022) +24. Liu, Y., Chen, H., Shen, C., He, T., Jin, L., Wang, L.: Abcnet: Real-time scene text spotting with adaptiveBezier-curve network. In: CVPR. pp. 9809-9818 (2020) +25. Liu, Y., Jin, L., Zhang, S., Luo, C., Zhang, S.: Curved scene text detection via transverse and longitudinal sequence connection. PR 90, 337-345 (2019) +26. Long, S., Ruan, J., Zhang, W., He, X., Wu, W., Yao, C.: Textsnake: A flexible representation for detecting text of arbitrary shapes. In: ECCV. pp. 20-36 (2018) +27. Long, S., Yao, C.: Unrealtext: Synthesizing realistic scene text images from the unreal world. arXiv:2003.10608 (2020) +28. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv:1711.05101 (2017) +29. Nayef, N., Patel, Y., Busta, M., et al.: Icdar2019 robust reading challenge on multilingual scene text detection and recognition—rrc-plt-2019. In: ICDAR. pp. 1582-1587 (2019) +30. Nayef, N., Yin, F., Bizid, I., et al.: Icdar2017 robust reading challenge on multilingual scene text detection and script identification-rrc-mlt. In: ICDAR. vol. 1, pp. 1454-1459 (2017) +31. Nguyen, N., Nguyen, T., Tran, V., et al.: Dictionary-guided scene text recognition. In: CVPR. pp. 7383-7392 (2021) +32. Olsson, V., Tranheden, W., Pinto, J., Svensson, L.: Classmix: Segmentation-based data augmentation for semi-supervised learning. In: WACV. pp. 1369-1378 (2021) +33. Qin, X., Lyu, P., Zhang, C., Zhou, Y., Yao, K., Zhang, P., Lin, H., Wang, W.: Towards robust real-time scene text detection: From semantic to instance representation learning. In: ACM MM. pp. 2025-2034 (2023) +34. Qin, X., Zhou, Y., Guo, Y., Wu, D., Tian, Z., Jiang, N., Wang, H., Wang, W.: Mask is all you need: Rethinking mask r-cnn for dense and arbitrary-shaped scene text detection. In: ACM MM. pp. 414-423 (2021) +35. Qin, X., Zhou, Y., Guo, Y., Wu, D., Wang, W.: Fc 2 rn: a fully convolutional corner refinement network for accurate multi-oriented scene text detection. In: ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 4350-4354. IEEE (2021) +36. Radford, A., Kim, J.W., Hallacy, C., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021) +37. Shi, B., Bai, X., Belongie, S.: Detecting oriented text in natural images by linking segments. In: CVPR. pp. 2550-2558 (2017) +38. Sohn, K., Berthelot, D., Carlini, N., Zhang, Z., Zhang, H., Raffel, C.A., Cubuk, E.D., Kurakin, A., Li, C.L.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In: NeurIPS. vol. 33, pp. 596-608 (2020) +39. Song, S., Wan, J., Yang, Z., Tang, J., Cheng, W., Bai, X., Yao, C.: Vision-language pre-training for boosting scene text detectors. In: CVPR. pp. 15681-15691 (2022) +40. Sun, X., Lyu, J., Zhang, Y., Zeng, G., Fang, B., Zhou, Y., Xie, E., Ma, C.: Feature enhancement with text-specific region contrast for scene text detection. In: Chinese Conference on Pattern Recognition and Computer Vision (PRCV). pp. 3-14 (2023) + +41. Sun, Y., Ni, Z., Chng, C.K., et al.: Icdar 2019 competition on large-scale street view text with partial labeling-rrc-lsvt. In: ICDAR. pp. 1557-1562 (2019) +42. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In: NeurIPS. vol. 30 (2017) +43. Wan, Q., Ji, H., Shen, L.: Self-attention based text knowledge mining for text detection. In: CVPR. pp. 5983-5992 (2021) +44. Wang, F., Chen, Y., Wu, F., Li, X.: Texray: Contour-based geometric modeling for arbitrary-shaped scene text detection. In: ACM MM. pp. 111-119 (2020) +45. Wang, K., Xie, H., Wang, Y., Zhang, D., Qu, Y., Gao, Z., Zhang, Y.: Masked text modeling: A self-supervised pre-training method for scene text detection. In: ACM MM. pp. 2006-2015 (2023) +46. Wang, W., Xie, E., Li, X., Hou, W., Lu, T., Yu, G., Shao, S.: Shape robust text detection with progressive scale expansion network. In: CVPR. pp. 9336-9345 (2019) +47. Wang, W., Xie, E., Li, X., Liu, X., Liang, D., Yang, Z., Lu, T., Shen, C.: Pan++: Towards efficient and accurate end-to-end spotting of arbitrarily-shaped text. IEEE TPAMI 44(9), 5349-5367 (2022) +48. Wang, W., Xie, E., Song, X., Zang, Y., Wang, W., Lu, T., Yu, G., Shen, C.: Efficient and accurate arbitrary-shaped text detection with pixel aggregation network. In: ICCV. pp. 8440-8449 (2019) +49. Wang, X., Jiang, Y., Luo, Z., Liu, C.L., Choi, H., Kim, S.: Arbitrary shape scene text detection with adaptive text region representation. In: CVPR. pp. 6449-6458 (2019) +50. Wang, Y., Xie, H., Zha, Z.J., Xing, M., Fu, Z., Zhang, Y.: Contournet: Taking a further step toward accurate arbitrary-shaped scene text detection. In: CVPR. pp. 11753-11762 (2020) +51. Wu, W., Lu, N., Xie, E., Wang, Y., Yu, W., Yang, C., Zhou, H.: Synthetic-to-real unsupervised domain adaptation for scene text detection in the wild. In: Proceedings of the Asian Conference on Computer Vision (2020) +52. Xu, H., Liu, L., Bian, Q., Yang, Z.: Semi-supervised semantic segmentation with prototype-based consistency regularization. In: NeurIPS. vol. 35, pp. 26007-26020 (2022) +53. Xu, X., Zhang, Z., Wang, Z., Price, B., Wang, Z., Shi, H.: Rethinking text segmentation: A novel dataset and a text-specific refinement approach. In: CVPR. pp. 12045-12055 (2021) +54. Xue, C., Zhang, W., Hao, Y., Lu, S., Torr, P.H., Bai, S.: Language matters: A weakly supervised vision-language pre-training approach for scene text detection and spotting. In: ECCV. pp. 284-302. Springer (2022) +55. Yang, L., Qi, L., Feng, L., Zhang, W., Shi, Y.: Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In: CVPR. pp. 7236-7246 (2023) +56. Yang, X., Song, Z., King, I., Xu, Z.: A survey on deep semi-supervised learning. IEEE Transactions on Knowledge and Data Engineering 35(9), 8934-8954 (2022) +57. Yao, C., Bai, X., Liu, W., Ma, Y., Tu, Z.: Detecting texts of arbitrary orientations in natural images. In: CVPR. pp. 1083-1090 (2012) +58. Ye, M., Zhang, J., Zhao, S., Liu, J., Liu, T., Du, B., Tao, D.: Deepsolo: Let transformer decoder with explicit points solo for text spotting. In: CVPR. pp. 19348-19357 (2023) +59. Yin, X.C., Pei, W.Y., Zhang, J., Hao, H.W.: Multi-orientation scene text detection with adaptive clustering. IEEE TPAMI 37(9), 1930-1937 (2015) +60. Yu, W., Liu, Y., Hua, W., Jiang, D., Ren, B., Bai, X.: Turning a clip model into a scene text detector. In: CVPR. pp. 6978-6988 (2023) + +61. Yu, W., Liu, Y., Zhu, X., Cao, H., Sun, X., Bai, X.: Turning a clip model into a scene text spotter. IEEE TPAMI (2024) +62. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: ICCV. pp. 6023-6032 (2019) +63. Zhan, F., Xue, C., Lu, S.: Ga-dan: Geometry-aware domain adaptation network for scene text detection and recognition. In: ICCV. pp. 9105-9115 (2019) +64. Zhang, B., Wang, Y., Hou, W., Wu, H., Wang, J., Okumura, M., Shinozaki, T.: Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In: NeurIPS. vol. 34, pp. 18408-18419 (2021) +65. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv:1710.09412 (2017) +66. Zhao, Z., Yang, L., Long, S., Pi, J., Zhou, L., Wang, J.: Augmentation matters: A simple-yet-effective approach to semi-supervised semantic segmentation. In: CVPR. pp. 11350-11359 (2023) +67. Zhou, X., Yao, C., Wen, H., Wang, Y., Zhou, S., He, W., Liang, J.: East: an efficient and accurate scene text detector. In: CVPR. pp. 5551-5560 (2017) +68. Zhu, Y., Chen, J., Liang, L., Kuang, Z., Jin, L., Zhang, W.: Fourier contour embedding for arbitrary-shaped text detection. In: CVPR. pp. 3123-3131 (2021) \ No newline at end of file diff --git a/2024/Bridging Synthetic and Real Worlds for Pre-training Scene Text Detectors/images.zip b/2024/Bridging Synthetic and Real Worlds for Pre-training Scene Text Detectors/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bd124903a2f9a8c9ec8de5479df638ec48ae2078 --- /dev/null +++ b/2024/Bridging Synthetic and Real Worlds for Pre-training Scene Text Detectors/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89bde06becd707637f5f6b40f72681d2efc3065ab0c59dde43fc253dea4045da +size 422259 diff --git a/2024/Bridging Synthetic and Real Worlds for Pre-training Scene Text Detectors/layout.json b/2024/Bridging Synthetic and Real Worlds for Pre-training Scene Text Detectors/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2e8d7f2e2f12c3204075aeb0edbdb0073d49436b --- /dev/null +++ b/2024/Bridging Synthetic and Real Worlds for Pre-training Scene Text Detectors/layout.json @@ -0,0 +1,11175 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 168, + 111, + 446, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 111, + 446, + 148 + ], + "spans": [ + { + "bbox": [ + 168, + 111, + 446, + 148 + ], + "type": "text", + "content": "Bridging Synthetic and Real Worlds for Pre-training Scene Text Detectors" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 209, + 167, + 403, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 167, + 403, + 193 + ], + "spans": [ + { + "bbox": [ + 209, + 167, + 403, + 193 + ], + "type": "text", + "content": "Tongkun Guan" + }, + { + "bbox": [ + 209, + 167, + 403, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 209, + 167, + 403, + 193 + ], + "type": "text", + "content": ", Wei Shen" + }, + { + "bbox": [ + 209, + 167, + 403, + 193 + ], + "type": "inline_equation", + "content": "^{1(\\boxtimes)}" + }, + { + "bbox": [ + 209, + 167, + 403, + 193 + ], + "type": "text", + "content": ", Xue Yang" + }, + { + "bbox": [ + 209, + 167, + 403, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 209, + 167, + 403, + 193 + ], + "type": "text", + "content": ", Xuehui Wang" + }, + { + "bbox": [ + 209, + 167, + 403, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 209, + 167, + 403, + 193 + ], + "type": "text", + "content": ", and Xiaokang Yang" + }, + { + "bbox": [ + 209, + 167, + 403, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 201, + 481, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 201, + 481, + 223 + ], + "spans": [ + { + "bbox": [ + 132, + 201, + 481, + 223 + ], + "type": "text", + "content": "1 MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University {gtk0615,wei.shen}@sjtu.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 252, + 224, + 361, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 224, + 361, + 247 + ], + "spans": [ + { + "bbox": [ + 252, + 224, + 361, + 247 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 252, + 224, + 361, + 247 + ], + "type": "text", + "content": " Shanghai AI Laboratory yangxue@pjlab.org.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 276, + 452, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 276, + 452, + 506 + ], + "spans": [ + { + "bbox": [ + 160, + 276, + 452, + 506 + ], + "type": "text", + "content": "Abstract. Existing scene text detection methods typically rely on extensive real data for training. Due to the lack of annotated real images, recent works have attempted to exploit large-scale labeled synthetic data (LSD) for pre-training text detectors. However, a synth-to-real domain gap emerges, further limiting the performance of text detectors. Differently, in this work, we propose FreeReal, a real-domain-aligned pretraining paradigm that enables the complementary strengths of both LSD and unlabeled real data (URD). Specifically, to bridge real and synthetic worlds for pre-training, a glyph-based mixing mechanism (GlyphMix) is tailored for text images. GlyphMix delineates the character structures of synthetic images and embeds them as graffiti-like units onto real images. Without introducing real domain drift, GlyphMix freely yields real-world images with partial annotations derived from synthetic labels. Furthermore, when given free fine-grained synthetic labels, GlyphMix can effectively bridge the linguistic domain gap stemming from English-dominated LSD to URD in various languages. Without bells and whistles, FreeReal achieves average gains of " + }, + { + "bbox": [ + 160, + 276, + 452, + 506 + ], + "type": "inline_equation", + "content": "1.97\\%" + }, + { + "bbox": [ + 160, + 276, + 452, + 506 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 160, + 276, + 452, + 506 + ], + "type": "inline_equation", + "content": "3.90\\%" + }, + { + "bbox": [ + 160, + 276, + 452, + 506 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 160, + 276, + 452, + 506 + ], + "type": "inline_equation", + "content": "3.85\\%" + }, + { + "bbox": [ + 160, + 276, + 452, + 506 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 160, + 276, + 452, + 506 + ], + "type": "inline_equation", + "content": "4.56\\%" + }, + { + "bbox": [ + 160, + 276, + 452, + 506 + ], + "type": "text", + "content": " in improving the performance of FCENet, PSENet, PANet, and DBNet methods, respectively, consistently outperforming previous pre-training methods by a substantial margin across four public datasets. Code will be available at https://github.com/SJTU-DeepVisionLab/FreeReal." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 517, + 453, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 517, + 453, + 540 + ], + "spans": [ + { + "bbox": [ + 159, + 517, + 453, + 540 + ], + "type": "text", + "content": "Keywords: Scene text detection " + }, + { + "bbox": [ + 159, + 517, + 453, + 540 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 517, + 453, + 540 + ], + "type": "text", + "content": " Pre-training paradigm " + }, + { + "bbox": [ + 159, + 517, + 453, + 540 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 517, + 453, + 540 + ], + "type": "text", + "content": " Domain-bridging method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 560, + 230, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 560, + 230, + 572 + ], + "spans": [ + { + "bbox": [ + 132, + 560, + 230, + 572 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": "Scene text detection aims to predict the bounding box or polygon for each arbitrarily shaped text instance within an image. This technology has diverse applications, including online education, instant translation, and product search, owing to the prevalence of text-laden images. However, despite their considerable progress, supervised methods [8,10,22,46,50,68] still struggle with an increasing" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 140, + 652, + 243, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 652, + 243, + 666 + ], + "spans": [ + { + "bbox": [ + 140, + 652, + 243, + 666 + ], + "type": "text", + "content": "Corresponding author." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 169, + 115, + 445, + 216 + ], + "blocks": [ + { + "bbox": [ + 169, + 115, + 445, + 216 + ], + "lines": [ + { + "bbox": [ + 169, + 115, + 445, + 216 + ], + "spans": [ + { + "bbox": [ + 169, + 115, + 445, + 216 + ], + "type": "image", + "image_path": "9d8f019c33ab1a50b0a35f02d6e7fe250242854abac78561c91effbcdec1003a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 224, + 472, + 236 + ], + "lines": [ + { + "bbox": [ + 141, + 224, + 472, + 236 + ], + "spans": [ + { + "bbox": [ + 141, + 224, + 472, + 236 + ], + "type": "text", + "content": "Fig. 1: We present different pre-training paradigms for enhancing text detectors." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 263, + 480, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 263, + 480, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 480, + 288 + ], + "type": "text", + "content": "demand for real data with labor-intensive annotations, especially when detecting complicated text images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 290, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 469 + ], + "type": "text", + "content": "In contrast, recent research efforts [39, 43, 54] have focused on pre-training scene text detectors using large-scale labeled synthetic data (LSD), which leverage both visual and textual knowledge to extract text feature representations, as illustrated in Fig. 1(a). For example, VLPT [39] formulates three cross-modality pretext tasks to enhance the performance of the backbone, and oCLIP [54] focuses on aligning visual and textual information to assist text detectors. Nevertheless, these well-curated designs are limited to using synthetic images due to their reliance on labeled word transcriptions. This limitation will lead to a noticeable synth-to-real domain gap when models are fine-tuned on real data, negatively impacting text detection performance. Therefore, exploring the potential of unlabeled real data (URD) is of great importance as they are readily available. However, the pre-training on URD has seldom been studied. One of the few works [45] attempts to pre-train text detectors on URD via masked image modeling [13] (step 1) and on LSD via supervised learning (step 2), which will re-transfers pre-trained weights from the real domain to the synthetic domain." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 472, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 590 + ], + "type": "text", + "content": "Differently, our objective is to jointly leverage LSD and URD by bridging their gap for enhancing the text detector pre-training, as shown in Fig. 1(b). The primary challenge is how to effectively modulate these URD and LSD without encountering real domain drift. In light of this, we initially investigate the adaptation capabilities of existing domain-bridging methods [32, 62, 65]. This involves training a classifier to distinguish between URD and LSD and subsequently assessing the confidence of these domain-bridging methods in aligning with the real domain. As shown in Fig. 2(b), their highest score is only " + }, + { + "bbox": [ + 130, + 472, + 482, + 590 + ], + "type": "inline_equation", + "content": "78.9\\%" + }, + { + "bbox": [ + 130, + 472, + 482, + 590 + ], + "type": "text", + "content": ". The observation implies that these domain-bridging methods may introduce intermediate domains and disrupt the real domain." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "To conquer this domain drift issue, we propose FreeReal, a real-domain-aligned pre-training paradigm, leveraging synthetic labels to guide text detector learning on real-domain-aligned images. Specifically, a glyph-based mixing mechanism, named GlyphMix, is tailored for text images to create real-world images with annotations derived from synthetic labels. This involves delineating the glyphs of synthetic images based on their internal structure and embedding" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "text", + "content": "T. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 170, + 115, + 443, + 200 + ], + "blocks": [ + { + "bbox": [ + 170, + 115, + 443, + 200 + ], + "lines": [ + { + "bbox": [ + 170, + 115, + 443, + 200 + ], + "spans": [ + { + "bbox": [ + 170, + 115, + 443, + 200 + ], + "type": "image", + "image_path": "1e4e3829745eba0bcbf9d190b1ee69de0e3c64de5186e462dc25dcd57a0715e8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 210, + 482, + 299 + ], + "lines": [ + { + "bbox": [ + 130, + 210, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 210, + 482, + 299 + ], + "type": "text", + "content": "Fig. 2: We present the domain classification task and the performances of different domain bridging ways between LSD and URD. In the text detection area, experiments demonstrate that our domain bridging method significantly performs better in aligning with the real domain. This is easy to explain: (I) Mixup may lead to unexpected pixel-wise ambiguities; (II) CutMix may cause the absence of many text regions and incomplete semantics of texts; (III) ClassMix will introduce salient boundary priors; (IV) GlyphMix (ours) effectively preserves the semantic information of text without causing real domain drift and bringing extra boundary priors." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 326, + 480, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 480, + 373 + ], + "type": "text", + "content": "them as graffiti-like units onto any unlabeled real images. GlyphMix offers several advantages: 1) extracting glyph structures in a cost-effective way without introducing extra boundary priors; 2) leveraging synthetic labels for free while greatly minimizing real domain drift (achieving a confidence score of " + }, + { + "bbox": [ + 130, + 326, + 480, + 373 + ], + "type": "inline_equation", + "content": "95.7\\%" + }, + { + "bbox": [ + 130, + 326, + 480, + 373 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 376, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 460 + ], + "type": "text", + "content": "As a byproduct, FreeReal can bridge the language-to-language gap, when adapting features learned from English-dominated LSD to URD in various languages. Specifically, given free character-level synthetic labels, FreeReal enables the pre-trained model to easily learn character features, since characters are the underlying unit of any language. Subsequently, during fine-tuning, these learned character features are adaptively combined according to the rules (e.g., word-and sentence-level) provided by available benchmarks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 461, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 557 + ], + "type": "text", + "content": "Building upon these insights, we break the trend of recent pre-training methods for scene text detectors (PSTD) that combine increasingly complex pretext task designs or more training modules to improve performance. Without bells and whistles, we focus on exploring the potential of integrating the LSD and URD. Remarkably, despite using " + }, + { + "bbox": [ + 130, + 461, + 482, + 557 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 130, + 461, + 482, + 557 + ], + "type": "text", + "content": " less LSD than recent SOTAs, our method still attains a " + }, + { + "bbox": [ + 130, + 461, + 482, + 557 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 130, + 461, + 482, + 557 + ], + "type": "text", + "content": " gain, effectively underscoring the superiority of our pretraining paradigm. As shown in Table 1, FreeReal consistently surpasses recent SOTAs by a large margin. The main contributions are summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 570, + 479, + 664 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 570, + 479, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 570, + 479, + 616 + ], + "spans": [ + { + "bbox": [ + 138, + 570, + 479, + 616 + ], + "type": "text", + "content": "- We are the first to simultaneously integrate labeled synthetic data and unlabeled real data for enhancing the pre-training of text detectors. FreeReal shows its superiority and significant performance gains over other complex alternatives, serving as a simple yet strong baseline for future PSTD studies." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 617, + 479, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 617, + 479, + 664 + ], + "spans": [ + { + "bbox": [ + 138, + 617, + 479, + 664 + ], + "type": "text", + "content": "- We effectively address the challenges of synthetic-to-real and language-to-language domain adaptation. Without bringing boundary priors and causing real domain drift, we freely harness the strengths of real-world text images with annotations derived from synthetic labels." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 319, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 319, + 91, + 447, + 102 + ], + "type": "text", + "content": "FreeReal for Pre-training STDs" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 496, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 496, + 222 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 496, + 222 + ], + "type": "text", + "content": "Scene Text Detection. Existing STD methods delineate text regions by regression-based or segmentation-based representations. The regression-based methods [8, 14,24,34,49,67,68] directly regress the bounding points to enclose text regions, and gradually increase the number of points as required to detect complex scenes, e.g., arbitrary texts. The segmentation-based methods regard texts of arbitrary shapes by grouping pixel-level [22, 23, 40, 48], character-level [1, 9], segment-level [26,37] or contour-level [35,44,50] segmentation results by post-processing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 223, + 481, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 223, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 481, + 450 + ], + "type": "text", + "content": "Pre-training for Scene Text Detector. In early research, [12, 21, 27] proposed to pre-train scene text detectors by generating synthetic data. These data engine tools follow the same paradigm: they first stack multiple off-the-shelf models to perform semantic segmentation and depth estimation analysis on background images, and then blend pre-processed texts into these images. However, the multi-stage, pre-generated methods incur large computational costs (GPU and CPU resources) and substantial latency, making them unsuitable for online training tasks where immediate data processing is essential. Inspired by recent advancements in pre-training techniques [3, 4, 11, 36], the integration of text and visual cross-modal information into scene text detectors has garnered increasing attention. One representative work involves enhancing the backbone with text transcriptions. For example, STKM [43] proposes a pioneering work that pretrains the backbone through an image-level text recognition task. VLPT [39] formulates three cross-modality pretext tasks to boost better text representations derived from the backbone. Another representative work is to endow the entire detector with pre-trained vision-text knowledge. Inherited from CLIP [36], TCM incorporates the Image and Text encoders into the text detector to refine text regions using cross-modal visual-language priors. A persistent challenge still remains—the synth-to-real domain gap limits the detection performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 450, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 481, + 666 + ], + "type": "text", + "content": "Semi-Supervised Learning. Due to the scarcity of expensive annotations, semi-supervised learning (SSL) methods are becoming increasingly important. These methods extract knowledge from a reduced amount of annotated images in a supervised manner, and from a larger amount of unlabeled images in an unsupervised way. Commonly, a self-training framework for SSL is employed, which encourages a model (student) to learn unlabeled images from pseudo-labels generated by the up-to-date optimized model (teacher). To avoid overfitting incorrect pseudo-labels, recent advancements have focused on pseudo-label refinement by improving the reliability [38,64] and pseudo-label regularization by cross-view consistency [19, 20, 42, 52]. Specifically, some domain adaption works [5, 51, 63] for STD focus on minimizing domain discrepancies in the feature space, including self-training and adversarial alignment methods. The self-training methods adopt the source data to pre-train a model for generating pseudo labels on target data, then use them to retrain the model. Adversarial alignment approaches aim to align the feature distribution and learn the domain-invariant features between the source and target domains. Recently, they have been replaced by the student-teacher framework, which significantly reduces domain gaps through dual-branch knowledge transfer [56]." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "T. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 296, + 197 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 296, + 197 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 296, + 197 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 296, + 197 + ], + "type": "image", + "image_path": "17e5610e84a7fe2b5752b87e9a748a27a728c9dad340779b7aa990968d215e99.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 200, + 231, + 207 + ], + "lines": [ + { + "bbox": [ + 198, + 200, + 231, + 207 + ], + "spans": [ + { + "bbox": [ + 198, + 200, + 231, + 207 + ], + "type": "text", + "content": "(a) FreeReal" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 300, + 117, + 480, + 196 + ], + "blocks": [ + { + "bbox": [ + 300, + 117, + 480, + 196 + ], + "lines": [ + { + "bbox": [ + 300, + 117, + 480, + 196 + ], + "spans": [ + { + "bbox": [ + 300, + 117, + 480, + 196 + ], + "type": "image", + "image_path": "15ce7f2392943d4f04eb7bee305daaea034e94cfca3b855001b19fd2cd547267.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 200, + 409, + 209 + ], + "lines": [ + { + "bbox": [ + 374, + 200, + 409, + 209 + ], + "spans": [ + { + "bbox": [ + 374, + 200, + 409, + 209 + ], + "type": "text", + "content": "(b) GlyphMix" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 180, + 218, + 431, + 229 + ], + "lines": [ + { + "bbox": [ + 180, + 218, + 431, + 229 + ], + "spans": [ + { + "bbox": [ + 180, + 218, + 431, + 229 + ], + "type": "text", + "content": "Fig. 3: Our network pipeline in a student-teacher framework." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 253, + 232, + 267 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 253, + 232, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 253, + 232, + 267 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "text", + "content": "Overview. Given a labeled synthetic set " + }, + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_l = \\{(\\mathbf{X}_l^{(i)},\\mathcal{B}_l^{(i)})\\}_{i = 1}^{N_l}" + }, + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "inline_equation", + "content": "N_{l}" + }, + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "text", + "content": " samples, " + }, + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_l^{(i)}\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "text", + "content": " represents an image and " + }, + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_l^{(i)} = \\{(p_{lt}^i,p_{lb}^i,p_{rt}^i,p_{rb}^i)\\}_{i = 1}^n" + }, + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "text", + "content": " is its corresponding set of bounding boxes. " + }, + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "text", + "content": " refers to the values of the vertice along the abscissa and ordinate, and " + }, + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "text", + "content": " denotes the number of bounding boxes. " + }, + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "text", + "content": " are the height and width of the image. Similarly, the unlabeled real set is denoted as " + }, + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_u = \\{\\mathbf{X}_u^{(i)}\\}_{i = 1}^{N_u}" + }, + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "inline_equation", + "content": "N_{u}" + }, + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "text", + "content": " samples. Our objective is to jointly leverage labeled synthetic data " + }, + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_l" + }, + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "text", + "content": " and unlabeled real data " + }, + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_u" + }, + { + "bbox": [ + 130, + 281, + 482, + 382 + ], + "type": "text", + "content": " by bridging their gap for enhancing the pre-training of scene text detectors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 383, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 383, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 383, + 482, + 453 + ], + "type": "text", + "content": "The student-teacher framework exhibits substantial efficacy due to its dual capacity for knowledge transfer between its two branches, which is typically employed to mitigate the domain gap (e.g., data and multimodal). Accordingly, we customize it for aligning real and synthetic text data, as illustrated in Fig. 3. For notation simplification in the following introduction, we omit the superscript " + }, + { + "bbox": [ + 130, + 383, + 482, + 453 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 383, + 482, + 453 + ], + "type": "text", + "content": " as mentioned above." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "text", + "content": "Specifically, the segmentation-based text detector DBNet [22] is employed as our baseline model, primarily due to its impressive real-time performance, versatility, and practical applicability. In the teacher branch, the teacher model parameterized by " + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t" + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "text", + "content": ", is responsible for generating truncated-gradient logits " + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}_u \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "text", + "content": " from a weakly augmented unlabeled real image " + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "inline_equation", + "content": "w(\\mathbf{X}_u)" + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "text", + "content": ". The pseudo label " + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_u \\in \\{0,1\\}^{H \\times W}" + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "text", + "content": " is then obtained by binarizing the logit map " + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}_u" + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "text", + "content": ". In the student branch, the process begins by generating a pixel-wise binary map " + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_l \\in \\{0,1\\}^{H \\times W}" + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "text", + "content": ", where 1 is assigned to locations within all bounding boxes from annotation " + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_l" + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "text", + "content": ", and 0 otherwise. The labeled synthetic image " + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_l" + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "text", + "content": " and the strongly augmented unlabeled real image " + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "inline_equation", + "content": "s(\\mathbf{X}_u)" + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "text", + "content": " are combined to create " + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_g = \\mathsf{A}(s(\\mathbf{X}_u), \\mathbf{X}_l)" + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "inline_equation", + "content": "\\mathsf{A}(\\cdot)" + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "text", + "content": " denoting our proposed GlyphMix mechanism, which will be introduced later. Subsequently, the mixed image " + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_g" + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "text", + "content": " serves as input to train the student model, under the supervision of mixed pseudo-labels " + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_g = \\mathsf{A}(\\mathbf{Y}_u, \\mathbf{Y}_l)" + }, + { + "bbox": [ + 130, + 455, + 482, + 623 + ], + "type": "text", + "content": ", by minimizing loss in the following:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 254, + 624, + 480, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 624, + 480, + 638 + ], + "spans": [ + { + "bbox": [ + 254, + 624, + 480, + 638 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\ell_ {D B} \\left(\\hat {\\mathbf {Y}} _ {g}, \\mathbf {Y} _ {g}, \\mathbf {E}\\right), \\tag {1}", + "image_path": "81facaea8689a171bd330f629e6640388a02c08669100bf7beb0c9451303867b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\ell_{DB}" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " denotes the original loss function provided by DBNet [22] and " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}_g" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " represents the prediction logits generated by the student model. The matrix" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 319, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 319, + 91, + 447, + 102 + ], + "type": "text", + "content": "FreeReal for Pre-training STDs" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 479, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 139 + ], + "type": "inline_equation", + "content": "\\mathbf{E} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 130, + 115, + 479, + 139 + ], + "type": "text", + "content": ", which indicates positive regions for loss calculation, is set as an all-one matrix by default." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": "To filter out the unreliable predictions in mixed pseudo-labels " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_g" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": ", we adaptively adjust the matrix " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": " during pre-training. Specifically, for the labeled synthetic image " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_l" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": ", the positive regions " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_l" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": " are an all-one matrix. For the unlabeled real image " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_u" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": ", we compute the pixel-wise entropy map on " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}_u" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": " and then apply a dynamic entropy threshold " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": " to obtain its positive regions " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_u" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 251, + 202, + 480, + 215 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 202, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 251, + 202, + 480, + 215 + ], + "type": "interline_equation", + "content": "\\mathbf {E} _ {u} = \\mathbb {I} \\left(- \\hat {\\mathbf {Y}} _ {u} \\log \\hat {\\mathbf {Y}} _ {u} \\leq \\zeta\\right), \\tag {2}", + "image_path": "d31cee6b3d29e447be6d8e688131a9fa49e909b500d87898f44a9e3e59427fa7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 217, + 480, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 480, + 276 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 480, + 276 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 217, + 480, + 276 + ], + "type": "inline_equation", + "content": "\\mathbb{I}(\\cdot)" + }, + { + "bbox": [ + 130, + 217, + 480, + 276 + ], + "type": "text", + "content": " denotes the indicator function. The matrix is subsequently formulated as " + }, + { + "bbox": [ + 130, + 217, + 480, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{E} = \\mathsf{A}(\\mathbf{E}_u,\\mathbf{E}_l)" + }, + { + "bbox": [ + 130, + 217, + 480, + 276 + ], + "type": "text", + "content": ". In our experiments, the threshold " + }, + { + "bbox": [ + 130, + 217, + 480, + 276 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 130, + 217, + 480, + 276 + ], + "type": "text", + "content": " is determined by the value corresponding to the top " + }, + { + "bbox": [ + 130, + 217, + 480, + 276 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 217, + 480, + 276 + ], + "type": "text", + "content": " percent of the entropy map. As the reliability of pseudo-label gradually improves, we smoothly update " + }, + { + "bbox": [ + 130, + 217, + 480, + 276 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 217, + 480, + 276 + ], + "type": "text", + "content": " using np.linspace(80, 20, t), where " + }, + { + "bbox": [ + 130, + 217, + 480, + 276 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 217, + 480, + 276 + ], + "type": "text", + "content": " denotes the total number of training iterations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 277, + 479, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 479, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 479, + 300 + ], + "type": "text", + "content": "Finally, the teacher model " + }, + { + "bbox": [ + 130, + 277, + 479, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t" + }, + { + "bbox": [ + 130, + 277, + 479, + 300 + ], + "type": "text", + "content": " gradually updates itself via the exponential moving averaging (EMA) of the student weights " + }, + { + "bbox": [ + 130, + 277, + 479, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_s" + }, + { + "bbox": [ + 130, + 277, + 479, + 300 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 261, + 303, + 480, + 315 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 303, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 261, + 303, + 480, + 315 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {t} = \\alpha \\mathbf {f} _ {t} + (1 - \\alpha) \\mathbf {f} _ {s}, \\tag {3}", + "image_path": "c078f5bb902acd7c2fd3e9b9ebc9c91499b2e735ee6a5c5a09344768061085d0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 317, + 394, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 317, + 394, + 329 + ], + "spans": [ + { + "bbox": [ + 131, + 317, + 394, + 329 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 317, + 394, + 329 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 317, + 394, + 329 + ], + "type": "text", + "content": " is a momentum parameter, set as 0.996 by default." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 349, + 212, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 349, + 212, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 349, + 212, + 361 + ], + "type": "text", + "content": "3.1 GlyphMix" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 371, + 480, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 480, + 430 + ], + "type": "text", + "content": "Although some efforts [32, 62, 65] attempt to aggregate the advantages of synthetic and unlabeled images through versatile transformations, they inevitably introduce intermediate domains, leading to real domain drift. In this section, we present a domain-bridging way tailored for text images, allowing for the free use of synthetic labels while ensuring consistent alignment with the real domain." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 431, + 480, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 431, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 431, + 480, + 479 + ], + "type": "text", + "content": "1) Graffiti-like Inter-domain Mixing. Given an input synthetic image " + }, + { + "bbox": [ + 130, + 431, + 480, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_l" + }, + { + "bbox": [ + 130, + 431, + 480, + 479 + ], + "type": "text", + "content": ", its corresponding bounding box annotations " + }, + { + "bbox": [ + 130, + 431, + 480, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_l" + }, + { + "bbox": [ + 130, + 431, + 480, + 479 + ], + "type": "text", + "content": " and converted pixel-wise binary map " + }, + { + "bbox": [ + 130, + 431, + 480, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_l" + }, + { + "bbox": [ + 130, + 431, + 480, + 479 + ], + "type": "text", + "content": ", our goal is to perform character segmentation. This process identifies all character structure regions and produces a mask " + }, + { + "bbox": [ + 130, + 431, + 480, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_G" + }, + { + "bbox": [ + 130, + 431, + 480, + 479 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 480, + 481, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 481, + 598 + ], + "type": "text", + "content": "To achieve this, we begin with an unsupervised clustering task based on the grayscale values of pixels, avoiding the introduction of extra pixel-level annotation costs. This involves aligning and cropping the image according to the coordinates of each bounding box, yielding a set of image patches " + }, + { + "bbox": [ + 130, + 480, + 481, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{\\mathbf{T}^{(i)}|\\mathbf{T}^{(i)}\\in \\mathbb{R}^{w_i\\times h_i\\times 3}\\}_{i = 1}^n" + }, + { + "bbox": [ + 130, + 480, + 481, + 598 + ], + "type": "text", + "content": ". Subsequently, within each image patch " + }, + { + "bbox": [ + 130, + 480, + 481, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{T}^{(i)}" + }, + { + "bbox": [ + 130, + 480, + 481, + 598 + ], + "type": "text", + "content": ", we cluster the pixels into two regions: the glyph region representing character structures and the background region. For simplicity, we focus our study on K-means to implement the clustering task, but other clustering approaches with predefined categories can be used. Following this, the glyph regions from all image patches are pasted to an all-zero image to create the character structure mask " + }, + { + "bbox": [ + 130, + 480, + 481, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_G" + }, + { + "bbox": [ + 130, + 480, + 481, + 598 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 599, + 479, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 599, + 479, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 599, + 479, + 635 + ], + "type": "text", + "content": "Given the pseudo label " + }, + { + "bbox": [ + 130, + 599, + 479, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_u" + }, + { + "bbox": [ + 130, + 599, + 479, + 635 + ], + "type": "text", + "content": " produced by the teacher model on an unlabeled real image " + }, + { + "bbox": [ + 130, + 599, + 479, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_u" + }, + { + "bbox": [ + 130, + 599, + 479, + 635 + ], + "type": "text", + "content": ", we use " + }, + { + "bbox": [ + 130, + 599, + 479, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_G" + }, + { + "bbox": [ + 130, + 599, + 479, + 635 + ], + "type": "text", + "content": " to generate the inter-domain mixing image and its corresponding label. This process can be formulated as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 203, + 637, + 480, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 637, + 480, + 668 + ], + "spans": [ + { + "bbox": [ + 203, + 637, + 480, + 668 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\mathbb {A} ^ {G} \\left(\\mathbf {X} _ {u}, \\mathbf {X} _ {l}\\right) \\leftarrow \\mathbf {M} _ {G} \\odot \\mathbf {X} _ {l} + \\left(\\mathbf {1} - \\mathbf {M} _ {G}\\right) \\odot \\mathbf {X} _ {u}, \\\\ \\mathbb {A} ^ {G} \\left(\\mathbf {Y} _ {u}, \\mathbf {Y} _ {l}\\right) \\leftarrow \\mathbf {M} _ {G} \\odot \\mathbf {Y} _ {l} + \\left(\\mathbf {1} - \\mathbf {M} _ {G}\\right) \\odot \\mathbf {Y} _ {u}, \\end{array} \\right. \\tag {4}", + "image_path": "0e2a5233aef22e36d53dd2a14fe3f63314115b58fb3c67281d29921587317dc5.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 132, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "T. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "inline_equation", + "content": "\\mathbf{1}" + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": " is an all-one matrix and " + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": " denotes the element-wise multiplication operation. " + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_G" + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": " indicates pixels occupied by the glyph need to be copied from the synth domain and pasted to the real domain. Thus the mixed image closely aligns with the real domain, taking advantage of the freely available synthetic labels. Put simply, this process produces real-world images with partial annotations derived from synthetic labels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 480, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 480, + 308 + ], + "type": "text", + "content": "2) Text-balanced Intra-domain Mixing. Real-world text images inherently exhibit complexity, and the collected unlabeled data often struggles to comprehensively represent the full spectrum of real-world scenarios. Within a student-teacher framework, intra-domain mixing serves as an effective perturbation technique, enriching the diversity of real data and enhancing feature consistency and robustness. However, when randomly generating and applying binary masks to construct intra-domain mixing images and their corresponding labels, there is a risk of missing many text regions, especially considering that text regions typically occupy only a small fraction of an image. To mitigate this challenge, our objective is to retain more text regions while still maintaining randomness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 480, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 480, + 434 + ], + "type": "text", + "content": "In light of this, we initiate the process by randomly selecting two unlabeled real images, denoted as " + }, + { + "bbox": [ + 130, + 308, + 480, + 434 + ], + "type": "inline_equation", + "content": "(\\mathbf{X}_u^{(1)},\\mathbf{X}_u^{(2)})" + }, + { + "bbox": [ + 130, + 308, + 480, + 434 + ], + "type": "text", + "content": ", and obtaining the corresponding segmentation pseudo-labels, " + }, + { + "bbox": [ + 130, + 308, + 480, + 434 + ], + "type": "inline_equation", + "content": "(\\mathbf{Y}_u^{(1)},\\mathbf{Y}_u^{(2)})" + }, + { + "bbox": [ + 130, + 308, + 480, + 434 + ], + "type": "text", + "content": ", produced by the teacher model. Subsequently, " + }, + { + "bbox": [ + 130, + 308, + 480, + 434 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 308, + 480, + 434 + ], + "type": "text", + "content": " rectangular boxes " + }, + { + "bbox": [ + 130, + 308, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\{(p_{lt}^i,p_{lb}^i,p_{rt}^i,p_{rb}^i)\\}_{i = 1}^k" + }, + { + "bbox": [ + 130, + 308, + 480, + 434 + ], + "type": "text", + "content": " are randomly generated to crop the pseudo-label " + }, + { + "bbox": [ + 130, + 308, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_u^{(2)}" + }, + { + "bbox": [ + 130, + 308, + 480, + 434 + ], + "type": "text", + "content": " based on their coordinates. This operation yields a set of local pseudo-label regions " + }, + { + "bbox": [ + 130, + 308, + 480, + 434 + ], + "type": "inline_equation", + "content": "S = \\{\\mathbf{S}^{(i)}|\\mathbf{S}^{(i)}\\in \\mathbb{R}^{w_i\\times h_i}\\}_{i = 1}^k" + }, + { + "bbox": [ + 130, + 308, + 480, + 434 + ], + "type": "text", + "content": ". Following this, we identify the rectangular box that contains the maximum text regions by calculating " + }, + { + "bbox": [ + 130, + 308, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\arg \\max_{i}\\sum_{j = 1}^{w_{i}\\times h_{i}}\\mathbf{S}_{j}^{(i)}" + }, + { + "bbox": [ + 130, + 308, + 480, + 434 + ], + "type": "text", + "content": ", and convert it into a binary mask, denoted as " + }, + { + "bbox": [ + 130, + 308, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_T" + }, + { + "bbox": [ + 130, + 308, + 480, + 434 + ], + "type": "text", + "content": ". Finally, we employ " + }, + { + "bbox": [ + 130, + 308, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_T" + }, + { + "bbox": [ + 130, + 308, + 480, + 434 + ], + "type": "text", + "content": " to generate the intra-domain mixed image and its corresponding label, following the formulated procedure:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 205, + 437, + 480, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 437, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 205, + 437, + 480, + 467 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\mathbf {A} ^ {T} (\\mathbf {X} _ {u}) \\leftarrow \\mathbf {M} _ {T} \\odot \\mathbf {X} _ {u} ^ {(2)} + \\left(\\mathbf {1} - \\mathbf {M} _ {T}\\right) \\odot \\mathbf {X} _ {u} ^ {(1)}, \\\\ \\mathbf {A} ^ {T} (\\mathbf {Y} _ {u}) \\leftarrow \\mathbf {M} _ {T} \\odot \\mathbf {Y} _ {u} ^ {(2)} + \\left(\\mathbf {1} - \\mathbf {M} _ {T}\\right) \\odot \\mathbf {Y} _ {u} ^ {(1)}, \\end{array} \\right. \\tag {5}", + "image_path": "81bd2e48635a4068a42acf2d1b89e9a34f7b25cf815fa95e4ccb52564ecce7d6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 468, + 477, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 477, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 477, + 480 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 468, + 477, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{1}" + }, + { + "bbox": [ + 130, + 468, + 477, + 480 + ], + "type": "text", + "content": " is an all-one matrix and " + }, + { + "bbox": [ + 130, + 468, + 477, + 480 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 468, + 477, + 480 + ], + "type": "text", + "content": " represents the element-wise multiplication." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 481, + 479, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 479, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 479, + 528 + ], + "type": "text", + "content": "3) Intra- and Inter-domain Mixing. When these two strategies, tailored for scene text images, are integrated within the student-teacher framework, they collectively form a comprehensive method that effectively harnesses the strengths of both synthetic and real data. The whole process can be formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 243, + 530, + 480, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 530, + 480, + 561 + ], + "spans": [ + { + "bbox": [ + 243, + 530, + 480, + 561 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\mathbf {A} (\\mathbf {X}) \\leftarrow \\mathbf {A} ^ {G} \\left(\\mathbf {A} ^ {T} \\left(\\mathbf {X} _ {u}\\right), \\mathbf {X} _ {l}\\right), \\\\ \\mathbf {A} (\\mathbf {Y}) \\leftarrow \\mathbf {A} ^ {G} \\left(\\mathbf {A} ^ {T} \\left(\\mathbf {Y} _ {u}\\right), \\mathbf {Y} _ {l}\\right) \\end{array} \\right. \\tag {6}", + "image_path": "c84a109de9eb702d127444e6ad0d2bf3feeba762b13e9076de946d867171fd0d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 562, + 479, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 562, + 479, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 562, + 479, + 586 + ], + "type": "text", + "content": "Consequently, the synergy maximizes the utilization of synthetic labels as a valuable resource, effectively tackling the intricacies of real-world scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 606, + 378, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 378, + 618 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 378, + 618 + ], + "type": "text", + "content": "3.2 Discussion on Character Region Awareness" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "Existing LSD only contains English texts with tightly spaced characters, while URD may encompass text properties from various languages (such as Japanese, Korean, Chinese, etc). Consequently, even though the model predictions effec" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 319, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 319, + 91, + 447, + 102 + ], + "type": "text", + "content": "FreeReal for Pre-training STDs" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "tively cover text regions, we encounter challenges in parsing these predictions into word-level bounding boxes or polygons." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 224 + ], + "type": "text", + "content": "Indeed, when using a single language (e.g., English) as the source domain and adapting it to various language domains, it appears that no matter how sophisticated the models and loss functions are, they struggle to accurately and effectively parse the complex text layout and semantic context. This challenge primarily arises from variations in the spacing and arrangement of characters, words, and sentences across languages. This observation leads us to revisit the learning objective of scene text detectors during pre-training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 344 + ], + "type": "text", + "content": "We emphasize an alternative perspective — utilizing characters as the fundamental learning unit for text detection. The strategy presents several advantages: 1) Like word-level bounding boxes (WBB), character-level bounding box (CBB) annotations can be readily acquired from the existing synthetic dataset (SynthText) at no additional cost; 2) Instead of parsing complex layouts and spacing disparities across languages during pre-training, decoding them into individual characters is straightforward and attainable; 3) The pre-trained model with character-level region awareness easily integrates its learned character features to adapt text attribute rules provided by different language text datasets during fine-tuning. To this end, we select CBB annotations as " + }, + { + "bbox": [ + 130, + 224, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_l" + }, + { + "bbox": [ + 130, + 224, + 482, + 344 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 363, + 223, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 223, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 223, + 376 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 389, + 200, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 389, + 200, + 400 + ], + "spans": [ + { + "bbox": [ + 132, + 389, + 200, + 400 + ], + "type": "text", + "content": "4.1 Dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 410, + 481, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 481, + 446 + ], + "type": "text", + "content": "Labeled Synthetic Data (LSD). The SynthText [12] serves as our source of LSD, consisting of 800K images. Each image is provided with character-level and word-level bounding boxes, along with corresponding text transcriptions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 446, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 481, + 495 + ], + "type": "text", + "content": "Unlabeled Real Data (URD). URD is collected from a wide array of natural scenes, encompassing a diverse range of fonts, languages, and text shapes. Specifically, URD includes 430K images from LSVT [41], 9K from MLT17 [30], 19657 from MLT19 [29], 229 from ICDAR2013 [18], and 500 from USTB-SV1K [59]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 495, + 481, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 481, + 532 + ], + "type": "text", + "content": "**Benchmarks.** Four publicly available benchmarks are used to evaluate the performance, including 1) Total-Text (TT) [6], 2) CTW 1500 (CTW) [25], 3) IC-DAR2015 (IC15) [17], and 4) MSRA-TD500 (TD) [57]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 548, + 279, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 279, + 560 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 279, + 560 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "Referring to recent works [39, 43, 54], we combine representative text detection methods [22, 46, 48, 68] to conduct experiments. URD is used in Tabs. 1-3 and the same data is used in Tabs. 4-7. All experiments are implemented on a server with 2 NVIDIA A800 GPUs, using PyTorch. 1) Pre-Training. The pre-training experiments are conducted on LSD and URD, utilizing the open-source codes provided by these text detection methods. For synthetic images, we adopt consistent augmentation settings as prescribed by these methods for a fair comparison. For unlabeled real images, we employ a standard weak-to-strong augmentation" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "T. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "technique [66] to enrich diversity. Our method is trained using the AdamW optimizer [28], along with a cosine learning rate scheduler [7]. This includes a warm-up phrase for 1 epoch in a total of 10 epochs. The base learning rate is set to " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "0.003 \\times" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " batchsize/256. At each iteration, we equally sample LSD and URD with a total batch size of 24. 2) Fine-Tuning. For a fair comparison, we employ the same fine-tuning configurations as these detection methods and then evaluate them on these publicly available benchmarks: TT, CTW, IC15 and TD, following the standard metric: Precision (P), Recall (R), and F-measure (F)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 228, + 261, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 228, + 261, + 240 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 261, + 240 + ], + "type": "text", + "content": "4.3 Experiment Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 482, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 377 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 377 + ], + "type": "text", + "content": "To ensure a comprehensive comparison, we collect results from public evaluation benchmarks reported by existing pre-training methods, and organize them in Table 1. The sources of these results are clearly recorded in the table's caption using different symbols for easy reference. Specifically, these methods can be classified into three pre-training paradigms. In Paradigm A, models (i.e., " + }, + { + "bbox": [ + 130, + 247, + 482, + 377 + ], + "type": "inline_equation", + "content": "+ST" + }, + { + "bbox": [ + 130, + 247, + 482, + 377 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 247, + 482, + 377 + ], + "type": "inline_equation", + "content": "+TCM" + }, + { + "bbox": [ + 130, + 247, + 482, + 377 + ], + "type": "text", + "content": ") are initially pre-trained on LSD and then fine-tuned on real datasets directly. In Paradigm B, models (i.e., " + }, + { + "bbox": [ + 130, + 247, + 482, + 377 + ], + "type": "inline_equation", + "content": "+STKM" + }, + { + "bbox": [ + 130, + 247, + 482, + 377 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 247, + 482, + 377 + ], + "type": "inline_equation", + "content": "+VLPT" + }, + { + "bbox": [ + 130, + 247, + 482, + 377 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 247, + 482, + 377 + ], + "type": "inline_equation", + "content": "+oCLIP" + }, + { + "bbox": [ + 130, + 247, + 482, + 377 + ], + "type": "text", + "content": ") are first pretrained on LSD by various pretext tasks, and the well-learned backbone weights are transferred into scene text detectors when fine-tuning on real datasets. In Paradigm C, models (i.e., " + }, + { + "bbox": [ + 130, + 247, + 482, + 377 + ], + "type": "inline_equation", + "content": "+MTM" + }, + { + "bbox": [ + 130, + 247, + 482, + 377 + ], + "type": "text", + "content": ") are pre-trained on URD by pretext tasks and on LSD by supervised learning, and then fine-tuned on real datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 378, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "text", + "content": "1) Results on DBNet. As indicated in the second group of rows in Table 1, all pre-training methods both report their cooperative performance with DBNet. When compared to methods using paradigm A, differing from " + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "inline_equation", + "content": "DBNet + ST" + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "text", + "content": ", we exclude the use of a deformable convolutional network (DCN) in our model. Despite these differences, our method still achieves impressive results, yielding improvements of " + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "text", + "content": " on TT, CTW, IC15, and TD datasets, respectively. Compared to " + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "inline_equation", + "content": "DBNet + TCM" + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "text", + "content": " methods, our method achieves SOTA detection performance on four datasets, particularly on curved text datasets (TT and CTW), with gains of " + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "inline_equation", + "content": "3.0\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "text", + "content": ", respectively. In comparison with methods following Paradigm B, our method consistently outperforms previous SOTA works by substantial margins, achieving performance improvements of " + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "inline_equation", + "content": "5.9\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "text", + "content": " on TT, CTW, IC15, and TD datasets, respectively. In comparison with " + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "inline_equation", + "content": "DBNet + MTM" + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "text", + "content": " following Paradigm C, our method gets " + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "text", + "content": " gains on TT and TD datasets, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 618 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 618 + ], + "type": "text", + "content": "2) Results on PANet. As shown in the third group of rows in Table 1, our method outperforms " + }, + { + "bbox": [ + 130, + 545, + 482, + 618 + ], + "type": "inline_equation", + "content": "PANet + ST" + }, + { + "bbox": [ + 130, + 545, + 482, + 618 + ], + "type": "text", + "content": " on four benchmarks, achieving a gain of " + }, + { + "bbox": [ + 130, + 545, + 482, + 618 + ], + "type": "inline_equation", + "content": "3.85\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 618 + ], + "type": "text", + "content": " in terms of average F-measure, while using the same parameters and labeled data for pre-training and fine-tuning. Furthermore, our method exhibits its prominent superiority, particularly on the CTW and TD datasets, surpassing the " + }, + { + "bbox": [ + 130, + 545, + 482, + 618 + ], + "type": "inline_equation", + "content": "PANet + TCM" + }, + { + "bbox": [ + 130, + 545, + 482, + 618 + ], + "type": "text", + "content": " method by a substantial margin of " + }, + { + "bbox": [ + 130, + 545, + 482, + 618 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 618 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 482, + 618 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 618 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 641 + ], + "type": "text", + "content": "3) Results on PSENet. FreeReal brings " + }, + { + "bbox": [ + 130, + 617, + 481, + 641 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 130, + 617, + 481, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 617, + 481, + 641 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 130, + 617, + 481, + 641 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 617, + 481, + 641 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 130, + 617, + 481, + 641 + ], + "type": "text", + "content": " gains on TT, CTW, and IC15 datasets, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "4) Results on FCENet. FreeReal further refreshes the best text detection results and improves " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " on TT, CTW, IC15, and TD" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 319, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 319, + 91, + 447, + 102 + ], + "type": "text", + "content": "FreeReal for Pre-training STDs" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 145, + 179, + 465, + 416 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": "Table 1: Comparison with SOTA pre-training methods for text detectors on TT, CTW, IC15, and TD. F-measure (\\%) is reported. * denotes the use of extra labeled real data. For easy reference, we mark these symbols " + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": " separately to indicate the results from [61], [39], [54], and [45], within the scope of this paper. The best results are shown in bold font. The second-best results are underlined." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 145, + 179, + 465, + 416 + ], + "lines": [ + { + "bbox": [ + 145, + 179, + 465, + 416 + ], + "spans": [ + { + "bbox": [ + 145, + 179, + 465, + 416 + ], + "type": "table", + "html": "
IndexMethodVenueTTCTWIC15TD
1PSENet-1s [46]CVPR'1980.982.285.7-
PAN++ [47]PAMI'2185.384.084.587.9
DBNet++ [23]PAMI'2286.085.387.387.2
DeepSOLO* [58]CVPR'2387.6-89.8-
ESTextSpotter* [15]ICCV'2390.0-91.089.5
SIR [33]ACMMM'2388.285.587.889.6
2DBNet+ST† [22]AAAI'2084.783.485.484.9
DBNet+STKM† [43]CVPR'2185.5-86.185.9
DBNet+VLPT† [39]CVPR'2286.3-86.588.5
DBNet+oCLIP† [54]ECCV'2284.182.085.4-
DBNet+TCM† [60]CVPR'2385.985.189.488.8
DBNet+MTM* [45]ACMMM'2385.7--85.6
DBNet+FreeReal (ours)-88.9(+2.6)87.9(+2.8)90.0(+0.6)90.1(+1.3)
3PANet+ST† [48]ICCV'1985.083.782.984.1
PANet+TCM† [60]CVPR'23-84.384.685.3
PANet+FreeReal (ours)-87.4(+2.4)88.4(+4.1)86.0(+1.4)89.9(+4.6)
4PSENet+ST* [46]CVPR'1983.981.681.3-
PSENet+STKM* [43]CVPR'2184.382.983.7-
PSENet+VLPT* [39]CVPR'2286.183.384.3-
PSENet+oCLIP† [54]ECCV'2285.582.8--
PSENet+FreeReal (ours)-86.8(+0.7)86.1(+2.8)85.2(+0.9)86.0
5FCENet+ST† [68]CVPR'2185.885.586.285.4
FCENet+oCLIP† [54]ECCV'22-85.686.7-
FCENet+TCM† [60]CVPR'23-85.987.186.9
FCENet+FreeReal (ours)-87.9(+2.1)87.2(+1.3)87.7(+0.6)89.0(+2.1)
", + "image_path": "ae00a87c75e810eaf91477012409dc48f5020a3d0d84ad6ecdfb4ebbfbb9786f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 454, + 482, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 454, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 454, + 482, + 478 + ], + "type": "text", + "content": "datasets, respectively. These experimental results demonstrate the robustness of FreeReal, even when cooperated with regression-based methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "5) Paradigm Comparison. In this section, we aim to investigate which kinds of pre-training paradigms for text detectors are important and meaningful for future studies. To this end, we conduct a comparative analysis with three mainstream pre-training paradigms mentioned above, as summarized in Table 2. When compared to these well-curated designs at the cost of introducing additional auxiliary pretext tasks or more training modules, our proposed paradigm achieves a significant improvement of " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " (87.5 vs. 86.3), despite utilizing " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " of the labeled synthetic data and " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " of the unlabeled real data. This outcome underscores our ability to harness the potential of unlabeled real images. Moreover, when we gradually increase the proportion of the total training data, the text detection performance continues to improve. Specifically, our method achieves gains of " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " at data fractions of " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ", respectively, compared to the recent SOTA method. This outcome fully demonstrates the superiority of our pre-training paradigm, which jointly leverage LSD and URD by bridging their gap for enhancing the text detector pre-training." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "T. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 162, + 158, + 449, + 289 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 147 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 147 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 147 + ], + "type": "text", + "content": "Table 2: Comparison with existing pre-training paradigms when fine-tuning on the TotalText dataset. These symbols " + }, + { + "bbox": [ + 132, + 114, + 480, + 147 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 132, + 114, + 480, + 147 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 114, + 480, + 147 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 132, + 114, + 480, + 147 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 132, + 114, + 480, + 147 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 132, + 114, + 480, + 147 + ], + "type": "text", + "content": " indicate the results from [61], [54] and [45], respectively. " + }, + { + "bbox": [ + 132, + 114, + 480, + 147 + ], + "type": "inline_equation", + "content": "\\Delta F" + }, + { + "bbox": [ + 132, + 114, + 480, + 147 + ], + "type": "text", + "content": " is the improvement relative to the baseline." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 162, + 158, + 449, + 289 + ], + "lines": [ + { + "bbox": [ + 162, + 158, + 449, + 289 + ], + "spans": [ + { + "bbox": [ + 162, + 158, + 449, + 289 + ], + "type": "table", + "html": "
ParadigmMethodVenueDataF-measureΔ F(%)
LSDURD
-DBNet‡ [22]AA AI'20--78.2-
ADBNet+ST† [22]AA AI'20800K-84.7+6.5%
DBNet+TCM† [60]CVPR'23800K-85.9+7.7%
BDBNet+STKM† [43]CVPR'21800K-85.5+7.3%
DBNet+VLPT† [39]CVPR'22800K-86.3+8.1%
DBNet+oCLIP‡ [54]ECCV'22800K-84.1+5.9%
CDBNet+MTM* [45]ACM MM'23800K41K85.7+7.5%
DOurs (9%)-71K41K87.4+9.2%
Ours (10%)-80K46K87.5+9.3%
Ours (20%)-160K92K88.1+9.9%
Ours (50%)-400K230K88.3+10.1%
Ours (100%)-800K460K88.9+10.7%
", + "image_path": "8931c343d196f89063d5238c473a3cf8bc7869482f3fb1b3d41df91f0ab7077f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 313, + 480, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 313, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 132, + 313, + 480, + 372 + ], + "type": "text", + "content": "7) Results on Unseen Language Text. Our pre-trained model is capable of learning the basic character feature representations of languages and then combining them during fine-tuning. To show its robustness, we choose to evaluate on unseen language text dataset VinText [31]. FreeReal achieves an improvement of " + }, + { + "bbox": [ + 132, + 313, + 480, + 372 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 132, + 313, + 480, + 372 + ], + "type": "text", + "content": ", as shown in Table 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 395, + 286, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 395, + 286, + 408 + ], + "spans": [ + { + "bbox": [ + 133, + 395, + 286, + 408 + ], + "type": "text", + "content": "5 Ablations and analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 424, + 480, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 460 + ], + "type": "text", + "content": "Unless otherwise stated, all ablation experiments are implemented in collaboration with DBNet method [22] using " + }, + { + "bbox": [ + 132, + 424, + 480, + 460 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 132, + 424, + 480, + 460 + ], + "type": "text", + "content": " of LSD and " + }, + { + "bbox": [ + 132, + 424, + 480, + 460 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 132, + 424, + 480, + 460 + ], + "type": "text", + "content": " of URD and evaluated on the curved TotalText (TT) dataset for efficiency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 461, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 481, + 628 + ], + "type": "text", + "content": "Effectiveness of Network Components. To comprehensively evaluate the effectiveness of our proposed method, we divide it into three components and evaluate each one individually, as shown in Table 4. \"Stu-Tea\", \"CRA\", \"GlyphMix\" denote the standard student-to-teacher framework, Character Region Awareness introduced in Sec.3.2 to bridge the linguistic domain gap, and the glyph-based mixing mechanism, respectively. First, we enhance the baseline DBNet model by incorporating a standard student-teacher framework, yielding a performance gain of " + }, + { + "bbox": [ + 132, + 461, + 481, + 628 + ], + "type": "inline_equation", + "content": "1.31\\%" + }, + { + "bbox": [ + 132, + 461, + 481, + 628 + ], + "type": "text", + "content": ". Second, within the student-teacher framework, we evaluate the individual contributions of the CRA and GlyphMix components, achieving " + }, + { + "bbox": [ + 132, + 461, + 481, + 628 + ], + "type": "inline_equation", + "content": "1.73\\%" + }, + { + "bbox": [ + 132, + 461, + 481, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 461, + 481, + 628 + ], + "type": "inline_equation", + "content": "3.12\\%" + }, + { + "bbox": [ + 132, + 461, + 481, + 628 + ], + "type": "text", + "content": " performance improvements, respectively. Finally, upon combining all plug-ins, the overall performance reaches " + }, + { + "bbox": [ + 132, + 461, + 481, + 628 + ], + "type": "inline_equation", + "content": "88.1\\%" + }, + { + "bbox": [ + 132, + 461, + 481, + 628 + ], + "type": "text", + "content": ", showing a substantial gain of " + }, + { + "bbox": [ + 132, + 461, + 481, + 628 + ], + "type": "inline_equation", + "content": "3.40\\%" + }, + { + "bbox": [ + 132, + 461, + 481, + 628 + ], + "type": "text", + "content": " compared to the baseline. These experiments consistently demonstrate that FreeReal effectively bridges both the synth-to-real and language-to-language domain gaps when leveraging the intrinsic qualities of unlabeled real data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 480, + 665 + ], + "type": "text", + "content": "Discussion on Using WBB or CBB. The SynthText dataset offers word-level and character-level annotations, with the latter being exploited in our method to bridge the linguistic domain gap, bringing a " + }, + { + "bbox": [ + 132, + 629, + 480, + 665 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 132, + 629, + 480, + 665 + ], + "type": "text", + "content": " gain as shown in the third row" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 319, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 319, + 91, + 447, + 102 + ], + "type": "text", + "content": "FreeReal for Pre-training STDs" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 192, + 158, + 419, + 191 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 3: Comparison with supervised SOTA work on unseen language text. \"R50\" and \"DeT\" denote a ResNet50 and Deformable Transformer structure, respectively. * denotes the use of extra labeled real data." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 192, + 158, + 419, + 191 + ], + "lines": [ + { + "bbox": [ + 192, + 158, + 419, + 191 + ], + "spans": [ + { + "bbox": [ + 192, + 158, + 419, + 191 + ], + "type": "table", + "html": "
MethodVenueBackbonePara.FPSF
ESTextSpotter*ICCV'23R50+DeT56M3.489.6
DBNet+FreeReal (ours)-R5025M1190.3
", + "image_path": "80aadfe79a831300963621c93bd0c1c5054d6151a76861fc738166e05656d5a3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 193, + 236, + 418, + 301 + ], + "blocks": [ + { + "bbox": [ + 130, + 205, + 479, + 226 + ], + "lines": [ + { + "bbox": [ + 130, + 205, + 479, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 479, + 226 + ], + "type": "text", + "content": "Table 4: Evaluate the effectiveness of the proposed modules. " + }, + { + "bbox": [ + 130, + 205, + 479, + 226 + ], + "type": "inline_equation", + "content": "\\Delta F" + }, + { + "bbox": [ + 130, + 205, + 479, + 226 + ], + "type": "text", + "content": " is the improvement of F-measure relative to baseline." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 193, + 236, + 418, + 301 + ], + "lines": [ + { + "bbox": [ + 193, + 236, + 418, + 301 + ], + "spans": [ + { + "bbox": [ + 193, + 236, + 418, + 301 + ], + "type": "table", + "html": "
Stu-TeaCRAGlyphMixΔ F(%)PRF
×××-87.182.584.7
✓××+1.31%88.8083.3986.01
✓✓×+1.73%87.7685.1586.43
✓×✓+3.12%89.1086.5887.82
✓✓✓+3.40%88.8187.4088.10
", + "image_path": "0941c58a0156c2be48dcaba74891d895b3751781e37ccd214dc99b621c506693.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 326, + 481, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 481, + 350 + ], + "type": "text", + "content": "of Table 4. Notably, even without the use of CRA, our method still achieves an F-measure of " + }, + { + "bbox": [ + 130, + 326, + 481, + 350 + ], + "type": "inline_equation", + "content": "87.82\\%" + }, + { + "bbox": [ + 130, + 326, + 481, + 350 + ], + "type": "text", + "content": " (in the fourth row), significantly surpassing recent SOTAs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 351, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 483 + ], + "type": "text", + "content": "Comparison with pre-training methods using URD. Bridging the synth-to-real gap to simultaneously leverage LSD and URD for pre-training is our major contribution. This capability, not present in other PSTD methods, enables seamless incorporation of the huge number of URD to enhance pre-training. Thus to compare to other PSTD methods with complex pretext task designs, we use different ratios of URD to explore the effectiveness of our pre-training paradigm. Even when using " + }, + { + "bbox": [ + 130, + 351, + 482, + 483 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 351, + 482, + 483 + ], + "type": "text", + "content": " LSD and " + }, + { + "bbox": [ + 130, + 351, + 482, + 483 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 351, + 482, + 483 + ], + "type": "text", + "content": " URD in Table 2, our method still achieves SOTA with a " + }, + { + "bbox": [ + 130, + 351, + 482, + 483 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 130, + 351, + 482, + 483 + ], + "type": "text", + "content": " gain, which exhibits our method can serve as a novel baseline for PSTD. Besides, we add the comparison using additional URD to train other PSTD methods. In this case, these mentioned methods can be viewed the same as MTM [45]. As a result, FreeReal outperforms MTM by " + }, + { + "bbox": [ + 130, + 351, + 482, + 483 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 130, + 351, + 482, + 483 + ], + "type": "text", + "content": " in Table 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 484, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 616 + ], + "type": "text", + "content": "Effectiveness of GlyphMix Components. In Table 5, GlyphMix comprises two modules, inter-domain GIM and intra-domain TIM, contributing gains of " + }, + { + "bbox": [ + 130, + 484, + 482, + 616 + ], + "type": "inline_equation", + "content": "1.28\\%" + }, + { + "bbox": [ + 130, + 484, + 482, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 484, + 482, + 616 + ], + "type": "inline_equation", + "content": "1.22\\%" + }, + { + "bbox": [ + 130, + 484, + 482, + 616 + ], + "type": "text", + "content": " to the final performance, respectively. Besides, when compared to the data engine tool [12], GIM achieves intra-domain mixing in a cost-effective way, which yields real-world images online via K-means, a more flexible, faster and direct process than the multi-stage, pre-generated method. Specifically, GIM attains a " + }, + { + "bbox": [ + 130, + 484, + 482, + 616 + ], + "type": "inline_equation", + "content": "10.7\\mathrm{x}" + }, + { + "bbox": [ + 130, + 484, + 482, + 616 + ], + "type": "text", + "content": " speedup (2.6h vs 27.7h) in generating 92K images, which is more suitable for real-time online training tasks. Additionally, GIM can avoid improper text blends on URD by leveraging the learned knowledge (boxes/maps) from our pre-training framework online. We also conduct the ablations and get " + }, + { + "bbox": [ + 130, + 484, + 482, + 616 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 130, + 484, + 482, + 616 + ], + "type": "inline_equation", + "content": "(85.3\\%)" + }, + { + "bbox": [ + 130, + 484, + 482, + 616 + ], + "type": "text", + "content": " vs " + }, + { + "bbox": [ + 130, + 484, + 482, + 616 + ], + "type": "inline_equation", + "content": "84.8\\%" + }, + { + "bbox": [ + 130, + 484, + 482, + 616 + ], + "type": "text", + "content": " ) performance gains, when using " + }, + { + "bbox": [ + 130, + 484, + 482, + 616 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 484, + 482, + 616 + ], + "type": "text", + "content": " training data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Comparison with Bridging-domain Methods. The goal of generating mixed images through domain-bridging methods is to bring synthetic images closer to real images. Consequently, the adaptation capability can be viewed as the likelihood of mixed images belonging to the real domain. A classification task is em" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "T. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 216, + 146, + 395, + 194 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 5: Evaluate the effectiveness of GlyphMix components. \"GIM\" and \"TIM\" denote the Graffiti-like Inter-domain Mixing and Text-balanced Intra-domain Mixing." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 216, + 146, + 395, + 194 + ], + "lines": [ + { + "bbox": [ + 216, + 146, + 395, + 194 + ], + "spans": [ + { + "bbox": [ + 216, + 146, + 395, + 194 + ], + "type": "table", + "html": "
GIMTIMΔ F(%)PRF
××-87.7685.1586.43
×✓+1.22%88.6486.6887.65
✓×+1.28%88.2987.1387.71
", + "image_path": "cfd035502271b86bd9fe924d4c9ba125c44283c33f635812246b7448ba8ec475.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 201, + 239, + 410, + 302 + ], + "blocks": [ + { + "bbox": [ + 132, + 205, + 479, + 228 + ], + "lines": [ + { + "bbox": [ + 132, + 205, + 479, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 479, + 228 + ], + "type": "text", + "content": "Table 6: Comparison of the effectiveness with existing domain bridging methods. \"DCA\" refers to Domain Classification Accuracy in aligning with the real domain." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 201, + 239, + 410, + 302 + ], + "lines": [ + { + "bbox": [ + 201, + 239, + 410, + 302 + ], + "spans": [ + { + "bbox": [ + 201, + 239, + 410, + 302 + ], + "type": "table", + "html": "
methodDCAPRFΔ F(%)
Stu-Tea-88.8083.3986.01-
Mix up [65]69.589.1784.1286.57+0.56%
CutMix [62]72.887.9985.6486.80+0.79%
ClassMix [32]78.990.6583.6186.99+0.98%
FreeReal (ours)95.788.8187.4088.10+2.09%
", + "image_path": "c435c040338b80ee89d256b36df7176ceb78c35dc726ef04bd298995da558379.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 194, + 335, + 416, + 408 + ], + "blocks": [ + { + "bbox": [ + 182, + 314, + 430, + 325 + ], + "lines": [ + { + "bbox": [ + 182, + 314, + 430, + 325 + ], + "spans": [ + { + "bbox": [ + 182, + 314, + 430, + 325 + ], + "type": "text", + "content": "Table 7: Comparison of the effectiveness with SSL methods." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 194, + 335, + 416, + 408 + ], + "lines": [ + { + "bbox": [ + 194, + 335, + 416, + 408 + ], + "spans": [ + { + "bbox": [ + 194, + 335, + 416, + 408 + ], + "type": "table", + "html": "
MethodVenuePRFΔ F(%)
Stu-Tea-88.8083.3986.01-
GTA [16]NeurIPS'2287.1881.0884.02-1.99%
BCP [2]CVPR'2386.7285.8286.27+0.26%
AugSeg [66]CVPR'2388.4885.6587.04+1.03%
Unimatch [55]CVPR'2388.9485.5687.22+1.21%
FreeReal (ours)-88.8187.4088.10+2.09%
", + "image_path": "2f8418d03f6e0cf66911d6449bd6394da9ddf811846c19bd5488147895ae8152.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 433, + 481, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 433, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 433, + 481, + 579 + ], + "type": "text", + "content": "ployed to assess the capability, whose the evaluation protocol is defined: Given a well-learned domain classifier " + }, + { + "bbox": [ + 132, + 433, + 481, + 579 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 132, + 433, + 481, + 579 + ], + "type": "text", + "content": " (trained by " + }, + { + "bbox": [ + 132, + 433, + 481, + 579 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 132, + 433, + 481, + 579 + ], + "type": "text", + "content": " synthetic images " + }, + { + "bbox": [ + 132, + 433, + 481, + 579 + ], + "type": "inline_equation", + "content": "x_{l} \\in X_{L}" + }, + { + "bbox": [ + 132, + 433, + 481, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 433, + 481, + 579 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 132, + 433, + 481, + 579 + ], + "type": "text", + "content": " real images " + }, + { + "bbox": [ + 132, + 433, + 481, + 579 + ], + "type": "inline_equation", + "content": "x_{u} \\in X_{U}" + }, + { + "bbox": [ + 132, + 433, + 481, + 579 + ], + "type": "text", + "content": ") and domain bridging method " + }, + { + "bbox": [ + 132, + 433, + 481, + 579 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 132, + 433, + 481, + 579 + ], + "type": "text", + "content": ", the classifier predicts the domain label (0 for synthetic and 1 for real) of mixed images and evaluates the accuracy as " + }, + { + "bbox": [ + 132, + 433, + 481, + 579 + ], + "type": "inline_equation", + "content": "\\frac{\\sum_{x_{l} \\in X_{L}, x_{u} \\in X_{U}} \\phi(\\xi(x_{l}, x_{u}))}{n^{2}}" + }, + { + "bbox": [ + 132, + 433, + 481, + 579 + ], + "type": "text", + "content": ". To accomplish this, we employ a network architecture, consisting of resnet18 as a feature extractor, followed by a binary classification head, for simplicity. Subsequently, the performance of different domain bridging methods is summarized in Table 6. Remarkably, when compared to these methods, GlyphMix achieves an outstanding performance score of " + }, + { + "bbox": [ + 132, + 433, + 481, + 579 + ], + "type": "inline_equation", + "content": "95.7\\%" + }, + { + "bbox": [ + 132, + 433, + 481, + 579 + ], + "type": "text", + "content": ", surpassing the second-best method by a substantial margin of " + }, + { + "bbox": [ + 132, + 433, + 481, + 579 + ], + "type": "inline_equation", + "content": "16.8\\%" + }, + { + "bbox": [ + 132, + 433, + 481, + 579 + ], + "type": "text", + "content": ". This result demonstrates that our method is able to generate arbitrarily realistic images without domain drift." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 581, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 581, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 481, + 640 + ], + "type": "text", + "content": "To be more convincing, we also conduct text detection experiments to assess the effectiveness. As depicted in Table 6, our method consistently outperforms others, achieving the best F-measure value. Furthermore, the observed positive correlation between DCA and F-measure reinforces the idea that bridging the domain gap is the crucial factor in improving text detection performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "content": "Comparison with Semi-supervised Methods. In this section, we compared the performance with recent SSL methods. These methods tend to adopt in-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 319, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 319, + 91, + 447, + 102 + ], + "type": "text", + "content": "FreeReal for Pre-training STDs" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 174 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 174 + ], + "type": "text", + "content": "creasingly complex designs, involving advanced contrastive learning techniques, additional network components, multiple ensemble models, and complex training procedures. In contrast, we aim to eliminate domain gaps between LSD and URD. Despite adopting the standard student-teacher framework, our method still exceeds the previous SOTA by " + }, + { + "bbox": [ + 130, + 116, + 479, + 174 + ], + "type": "inline_equation", + "content": "0.88\\%" + }, + { + "bbox": [ + 130, + 116, + 479, + 174 + ], + "type": "text", + "content": " in Table 7." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 175, + 479, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 479, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 479, + 319 + ], + "type": "text", + "content": "Effectiveness of " + }, + { + "bbox": [ + 130, + 175, + 479, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_G" + }, + { + "bbox": [ + 130, + 175, + 479, + 319 + ], + "type": "text", + "content": ". We evaluate the accuracy of character segmentation results " + }, + { + "bbox": [ + 130, + 175, + 479, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_G" + }, + { + "bbox": [ + 130, + 175, + 479, + 319 + ], + "type": "text", + "content": " using TotalText [6] and TextSeg [53] datasets, both of which provide pixel-level annotations. Specifically, compared to conducting the cluster task on the whole image, we crop it into several text instances and then perform clustering to form " + }, + { + "bbox": [ + 130, + 175, + 479, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_G" + }, + { + "bbox": [ + 130, + 175, + 479, + 319 + ], + "type": "text", + "content": ". While the accuracy of the former is " + }, + { + "bbox": [ + 130, + 175, + 479, + 319 + ], + "type": "inline_equation", + "content": "32.35\\%" + }, + { + "bbox": [ + 130, + 175, + 479, + 319 + ], + "type": "text", + "content": " for TotalText and " + }, + { + "bbox": [ + 130, + 175, + 479, + 319 + ], + "type": "inline_equation", + "content": "45.54\\%" + }, + { + "bbox": [ + 130, + 175, + 479, + 319 + ], + "type": "text", + "content": " for TextSeg, " + }, + { + "bbox": [ + 130, + 175, + 479, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_G" + }, + { + "bbox": [ + 130, + 175, + 479, + 319 + ], + "type": "text", + "content": " achieves " + }, + { + "bbox": [ + 130, + 175, + 479, + 319 + ], + "type": "inline_equation", + "content": "80.43\\%" + }, + { + "bbox": [ + 130, + 175, + 479, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 175, + 479, + 319 + ], + "type": "inline_equation", + "content": "84.14\\%" + }, + { + "bbox": [ + 130, + 175, + 479, + 319 + ], + "type": "text", + "content": ", respectively. Notably, the underlying morphological representations of glyphs exhibit relative invariance to slight structural changes, such as thicker or thinner. This reduces the dependence on pixel-level high-precision segmentation with expensive costs. Furthermore, although the clustering process introduces some noise, leading to inaccuracies in " + }, + { + "bbox": [ + 130, + 175, + 479, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_G" + }, + { + "bbox": [ + 130, + 175, + 479, + 319 + ], + "type": "text", + "content": ", it can still be viewed as a way of adding data noise in the image-mixing process, enhancing the robustness of text detection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 319, + 479, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 479, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 479, + 390 + ], + "type": "text", + "content": "Ablation Study on Hyper-parameters. We study the impact of the parameter " + }, + { + "bbox": [ + 130, + 319, + 479, + 390 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 319, + 479, + 390 + ], + "type": "text", + "content": " on the pseudos generated by the teacher model during pre-training. This parameter controls the proportion of unreliable and reliable pixels. With " + }, + { + "bbox": [ + 130, + 319, + 479, + 390 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 319, + 479, + 390 + ], + "type": "text", + "content": " set at 80, 50, and 20, the F-measure are " + }, + { + "bbox": [ + 130, + 319, + 479, + 390 + ], + "type": "inline_equation", + "content": "87.94\\%" + }, + { + "bbox": [ + 130, + 319, + 479, + 390 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 319, + 479, + 390 + ], + "type": "inline_equation", + "content": "87.68\\%" + }, + { + "bbox": [ + 130, + 319, + 479, + 390 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 319, + 479, + 390 + ], + "type": "inline_equation", + "content": "87.06\\%" + }, + { + "bbox": [ + 130, + 319, + 479, + 390 + ], + "type": "text", + "content": ", respectively. Finally, we employ a linear warm-up strategy for " + }, + { + "bbox": [ + 130, + 319, + 479, + 390 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 319, + 479, + 390 + ], + "type": "text", + "content": " that gradually reduces from 80 to 20, yielding the best result with an F-measure of " + }, + { + "bbox": [ + 130, + 319, + 479, + 390 + ], + "type": "inline_equation", + "content": "88.10\\%" + }, + { + "bbox": [ + 130, + 319, + 479, + 390 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 410, + 219, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 410, + 219, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 410, + 219, + 423 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 434, + 479, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 479, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 479, + 589 + ], + "type": "text", + "content": "We propose an effective pre-training paradigm for scene text detection, situated within a standard student-teacher framework. The paradigm leverages the complementary strengths of both LSD and URD, standing out from existing intricate pre-training paradigms in text detection due to its simplicity and significant advancements. Moreover, we identify and address prominent challenges, i.e., synth-to-real and language-to-language domain gaps. The proposed solutions, namely GlyphMix and character region awareness, have shown their effectiveness in mitigating these challenges and improving text detection performance, as substantiated by extensive experimental results. More importantly, our method exhibits the ability to harness the intrinsic attributes of URD, achieving substantial performance improvements even with a small fraction of LSD. This highlights the significance of exploring the potential of unlabeled real data in scene text detection, ultimately advancing the field and its practical applications." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 590, + 485, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 485, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 485, + 662 + ], + "type": "text", + "content": "Acknowledgment. This work was supported by the NSFC under Grant 62322604 and 62176159, and in part by the Shanghai Municipal Science, Technology Major Project under Grant 2021SHZDZX0102, and National Key R&D Program of China (NO. 2022ZD0160100). The computations in this paper were run on the AI for Science Platform supported by the Artificial Intelligence Institute at Shanghai Jiao Tong University." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "T. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "type": "text", + "content": "1. Baek, Y., Lee, B., Han, D., Yun, S., Lee, H.: Character region awareness for text detection. In: CVPR. pp. 9365-9374 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 165, + 480, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 480, + 186 + ], + "type": "text", + "content": "2. Bai, Y., Chen, D., Li, Q., Shen, W., Wang, Y.: Bidirectional copy-paste for semi-supervised medical image segmentation. In: CVPR. pp. 11514-11524 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 187, + 480, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 187, + 480, + 219 + ], + "spans": [ + { + "bbox": [ + 138, + 187, + 480, + 219 + ], + "type": "text", + "content": "3. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV. pp. 9650-9660 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 220, + 480, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 220, + 480, + 242 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 480, + 242 + ], + "type": "text", + "content": "4. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML. pp. 1597-1607 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 243, + 480, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 480, + 275 + ], + "type": "text", + "content": "5. Chen, Y., Wang, W., Zhou, Y., Yang, F., Yang, D., Wang, W.: Self-training for domain adaptive scene text detection. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 850-857. IEEE (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 276, + 480, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 276, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 276, + 480, + 297 + ], + "type": "text", + "content": "6. Ch'ng, C.K., Chan, C.S., Liu, C.L.: Total-text: toward orientation robustness in scene text detection. IJDAR 23(1), 31-52 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 298, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 298, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 298, + 480, + 319 + ], + "type": "text", + "content": "7. Grill, J., Strub, F., Altché, F., et al.: Bootstrap your own latent - A new approach to self-supervised learning. In: NeurIPS. pp. 21271-21284 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 320, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 480, + 353 + ], + "type": "text", + "content": "8. Guan, T., Gu, C., Lu, C., Tu, J., Feng, Q., Wu, K., Guan, X.: Industrial scene text detection with refined feature-attentive network. IEEE TCSVT 32(9), 6073-6085 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 354, + 480, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 480, + 375 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 480, + 375 + ], + "type": "text", + "content": "9. Guan, T., Gu, C., Tu, J., Yang, X., Feng, Q., Zhao, Y., Shen, W.: Self-supervised implicit glyph attention for text recognition. In: CVPR. pp. 15285-15294 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 376, + 480, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 376, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 134, + 376, + 480, + 408 + ], + "type": "text", + "content": "10. Guan, T., Lin, C., Shen, W., Yang, X.: Posformer: Recognizing complex handwritten mathematical expression with position forest transformer. arXiv preprint arXiv:2407.07764 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 409, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 409, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 133, + 409, + 480, + 441 + ], + "type": "text", + "content": "11. Guan, T., Shen, W., Yang, X., Feng, Q., Jiang, Z., Yang, X.: Self-supervised character-to-character distillation for text recognition. In: ICCV. pp. 19473-19484 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 443, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 443, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 133, + 443, + 480, + 464 + ], + "type": "text", + "content": "12. Gupta, A., Vedaldi, A., Zisserman, A.: Synthetic data for text localisation in natural images. In: CVPR. pp. 2315-2324 (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 465, + 480, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 465, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 133, + 465, + 480, + 487 + ], + "type": "text", + "content": "13. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: CVPR. pp. 16000-16009 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 487, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 487, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 133, + 487, + 480, + 520 + ], + "type": "text", + "content": "14. He, M., Liao, M., Yang, Z., Zhong, H., Tang, J., Cheng, W., Yao, C., Wang, Y., Bai, X.: Most: A multi-oriented scene text detector with localization refinement. In: CVPR. pp. 8813-8822 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 521, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 521, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 133, + 521, + 480, + 553 + ], + "type": "text", + "content": "15. Huang, M., Zhang, J., Peng, D., Lu, H., Huang, C., Liu, Y., Bai, X., Jin, L.: Estextspotter: Towards better scene text spotting with explicit synergy in transformer. In: ICCV. pp. 19495-19505 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 554, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 554, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 133, + 554, + 480, + 576 + ], + "type": "text", + "content": "16. Jin, Y., Wang, J., Lin, D.: Semi-supervised semantic segmentation via gentle teaching assistant. In: NeurIPS. vol. 35, pp. 2803-2816 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 577, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 577, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 133, + 577, + 480, + 598 + ], + "type": "text", + "content": "17. Karatzas, D., Gomez-Bigorda, L., Nicolaou, A., et al.: Icdar 2015 competition on robust reading. In: ICDAR. pp. 1156-1160 (2015)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 599, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 599, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 133, + 599, + 480, + 620 + ], + "type": "text", + "content": "18. Karatzas, D., Shafait, F., Uchida, S., et al.: Icdar 2013 robust reading competition. In: ICDAR. pp. 1484-1493 (2013)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 133, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 621, + 480, + 643 + ], + "type": "text", + "content": "19. Laine, S., Aila, T.: Temporal ensembling for semi-supervised learning. In: arXiv:1610.02242 (2016)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 480, + 665 + ], + "type": "text", + "content": "20. Li, W., Gao, H.a., Gao, M., Tian, B., Zhi, R., Zhao, H.: Training-free model merging for multi-target domain adaptation. In: ECCV. Springer (2024)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 319, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 319, + 91, + 447, + 102 + ], + "type": "text", + "content": "FreeReal for Pre-training STDs" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 148 + ], + "type": "text", + "content": "21. Liao, M., Song, B., Long, S., He, M., Yao, C., Bai, X.: Synthtext3d: synthesizing scene text images from 3d virtual worlds. Science China Information Sciences 63, 1-14 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 481, + 171 + ], + "type": "text", + "content": "22. Liao, M., Wan, Z., Yao, C., Chen, K., Bai, X.: Real-time scene text detection with differentiable binarization. In: AAAI. pp. 11474-11481 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 204 + ], + "type": "text", + "content": "23. Liao, M., Zou, Z., Wan, Z., Yao, C., Bai, X.: Real-time scene text detection with differentiable binarization and adaptive scale fusion. IEEE TPAMI 45(1), 919-931 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 226 + ], + "type": "text", + "content": "24. Liu, Y., Chen, H., Shen, C., He, T., Jin, L., Wang, L.: Abcnet: Real-time scene text spotting with adaptiveBezier-curve network. In: CVPR. pp. 9809-9818 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "text", + "content": "25. Liu, Y., Jin, L., Zhang, S., Luo, C., Zhang, S.: Curved scene text detection via transverse and longitudinal sequence connection. PR 90, 337-345 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "text", + "content": "26. Long, S., Ruan, J., Zhang, W., He, X., Wu, W., Yao, C.: Textsnake: A flexible representation for detecting text of arbitrary shapes. In: ECCV. pp. 20-36 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 481, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 292 + ], + "type": "text", + "content": "27. Long, S., Yao, C.: Unrealtext: Synthesizing realistic scene text images from the unreal world. arXiv:2003.10608 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 481, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 313 + ], + "type": "text", + "content": "28. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv:1711.05101 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "text", + "content": "29. Nayef, N., Patel, Y., Busta, M., et al.: Icdar2019 robust reading challenge on multilingual scene text detection and recognition—rrc-plt-2019. In: ICDAR. pp. 1582-1587 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "30. Nayef, N., Yin, F., Bizid, I., et al.: Icdar2017 robust reading challenge on multilingual scene text detection and script identification-rrc-mlt. In: ICDAR. vol. 1, pp. 1454-1459 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "text", + "content": "31. Nguyen, N., Nguyen, T., Tran, V., et al.: Dictionary-guided scene text recognition. In: CVPR. pp. 7383-7392 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "text", + "content": "32. Olsson, V., Tranheden, W., Pinto, J., Svensson, L.: Classmix: Segmentation-based data augmentation for semi-supervised learning. In: WACV. pp. 1369-1378 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 425, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 481, + 456 + ], + "type": "text", + "content": "33. Qin, X., Lyu, P., Zhang, C., Zhou, Y., Yao, K., Zhang, P., Lin, H., Wang, W.: Towards robust real-time scene text detection: From semantic to instance representation learning. In: ACM MM. pp. 2025-2034 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "text", + "content": "34. Qin, X., Zhou, Y., Guo, Y., Wu, D., Tian, Z., Jiang, N., Wang, H., Wang, W.: Mask is all you need: Rethinking mask r-cnn for dense and arbitrary-shaped scene text detection. In: ACM MM. pp. 414-423 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 491, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 533 + ], + "type": "text", + "content": "35. Qin, X., Zhou, Y., Guo, Y., Wu, D., Wang, W.: Fc 2 rn: a fully convolutional corner refinement network for accurate multi-oriented scene text detection. In: ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 4350-4354. IEEE (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "text", + "content": "36. Radford, A., Kim, J.W., Hallacy, C., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 555, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 577 + ], + "type": "text", + "content": "37. Shi, B., Bai, X., Belongie, S.: Detecting oriented text in natural images by linking segments. In: CVPR. pp. 2550-2558 (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 577, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 481, + 610 + ], + "type": "text", + "content": "38. Sohn, K., Berthelot, D., Carlini, N., Zhang, Z., Zhang, H., Raffel, C.A., Cubuk, E.D., Kurakin, A., Li, C.L.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In: NeurIPS. vol. 33, pp. 596-608 (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "text", + "content": "39. Song, S., Wan, J., Yang, Z., Tang, J., Cheng, W., Bai, X., Yao, C.: Vision-language pre-training for boosting scene text detectors. In: CVPR. pp. 15681-15691 (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "40. Sun, X., Lyu, J., Zhang, Y., Zeng, G., Fang, B., Zhou, Y., Xie, E., Ma, C.: Feature enhancement with text-specific region contrast for scene text detection. In: Chinese Conference on Pattern Recognition and Computer Vision (PRCV). pp. 3-14 (2023)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "T. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "41. Sun, Y., Ni, Z., Chng, C.K., et al.: Icdar 2019 competition on large-scale street view text with partial labeling-rrc-lsvt. In: ICDAR. pp. 1557-1562 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 170 + ], + "type": "text", + "content": "42. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In: NeurIPS. vol. 30 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 170, + 482, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 170, + 482, + 191 + ], + "spans": [ + { + "bbox": [ + 130, + 170, + 482, + 191 + ], + "type": "text", + "content": "43. Wan, Q., Ji, H., Shen, L.: Self-attention based text knowledge mining for text detection. In: CVPR. pp. 5983-5992 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 191, + 482, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 191, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 482, + 213 + ], + "type": "text", + "content": "44. Wang, F., Chen, Y., Wu, F., Li, X.: Texray: Contour-based geometric modeling for arbitrary-shaped scene text detection. In: ACM MM. pp. 111-119 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 213, + 482, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 246 + ], + "type": "text", + "content": "45. Wang, K., Xie, H., Wang, Y., Zhang, D., Qu, Y., Gao, Z., Zhang, Y.: Masked text modeling: A self-supervised pre-training method for scene text detection. In: ACM MM. pp. 2006-2015 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 246, + 482, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 267 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 267 + ], + "type": "text", + "content": "46. Wang, W., Xie, E., Li, X., Hou, W., Lu, T., Yu, G., Shao, S.: Shape robust text detection with progressive scale expansion network. In: CVPR. pp. 9336-9345 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 267, + 482, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 482, + 299 + ], + "type": "text", + "content": "47. Wang, W., Xie, E., Li, X., Liu, X., Liang, D., Yang, Z., Lu, T., Shen, C.: Pan++: Towards efficient and accurate end-to-end spotting of arbitrarily-shaped text. IEEE TPAMI 44(9), 5349-5367 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 300, + 482, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 331 + ], + "type": "text", + "content": "48. Wang, W., Xie, E., Song, X., Zang, Y., Wang, W., Lu, T., Yu, G., Shen, C.: Efficient and accurate arbitrary-shaped text detection with pixel aggregation network. In: ICCV. pp. 8440-8449 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 332, + 482, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 363 + ], + "type": "text", + "content": "49. Wang, X., Jiang, Y., Luo, Z., Liu, C.L., Choi, H., Kim, S.: Arbitrary shape scene text detection with adaptive text region representation. In: CVPR. pp. 6449-6458 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 363, + 482, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 396 + ], + "type": "text", + "content": "50. Wang, Y., Xie, H., Zha, Z.J., Xing, M., Fu, Z., Zhang, Y.: Contournet: Taking a further step toward accurate arbitrary-shaped scene text detection. In: CVPR. pp. 11753-11762 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 396, + 482, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 429 + ], + "type": "text", + "content": "51. Wu, W., Lu, N., Xie, E., Wang, Y., Yu, W., Yang, C., Zhou, H.: Synthetic-to-real unsupervised domain adaptation for scene text detection in the wild. In: Proceedings of the Asian Conference on Computer Vision (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 429, + 482, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 460 + ], + "type": "text", + "content": "52. Xu, H., Liu, L., Bian, Q., Yang, Z.: Semi-supervised semantic segmentation with prototype-based consistency regularization. In: NeurIPS. vol. 35, pp. 26007-26020 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 461, + 482, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 492 + ], + "type": "text", + "content": "53. Xu, X., Zhang, Z., Wang, Z., Price, B., Wang, Z., Shi, H.: Rethinking text segmentation: A novel dataset and a text-specific refinement approach. In: CVPR. pp. 12045-12055 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 492, + 482, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 482, + 525 + ], + "type": "text", + "content": "54. Xue, C., Zhang, W., Hao, Y., Lu, S., Torr, P.H., Bai, S.: Language matters: A weakly supervised vision-language pre-training approach for scene text detection and spotting. In: ECCV. pp. 284-302. Springer (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 525, + 482, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 548 + ], + "type": "text", + "content": "55. Yang, L., Qi, L., Feng, L., Zhang, W., Shi, Y.: Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In: CVPR. pp. 7236-7246 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 548, + 482, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 568 + ], + "type": "text", + "content": "56. Yang, X., Song, Z., King, I., Xu, Z.: A survey on deep semi-supervised learning. IEEE Transactions on Knowledge and Data Engineering 35(9), 8934-8954 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 568, + 482, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 590 + ], + "type": "text", + "content": "57. Yao, C., Bai, X., Liu, W., Ma, Y., Tu, Z.: Detecting texts of arbitrary orientations in natural images. In: CVPR. pp. 1083-1090 (2012)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 590, + 482, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 482, + 622 + ], + "type": "text", + "content": "58. Ye, M., Zhang, J., Zhao, S., Liu, J., Liu, T., Du, B., Tao, D.: Deepsolo: Let transformer decoder with explicit points solo for text spotting. In: CVPR. pp. 19348-19357 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 622, + 482, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 622, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 622, + 482, + 643 + ], + "type": "text", + "content": "59. Yin, X.C., Pei, W.Y., Zhang, J., Hao, H.W.: Multi-orientation scene text detection with adaptive clustering. IEEE TPAMI 37(9), 1930-1937 (2015)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 643, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 482, + 665 + ], + "type": "text", + "content": "60. Yu, W., Liu, Y., Hua, W., Jiang, D., Ren, B., Bai, X.: Turning a clip model into a scene text detector. In: CVPR. pp. 6978-6988 (2023)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 319, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 319, + 91, + 447, + 102 + ], + "type": "text", + "content": "FreeReal for Pre-training STDs" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 325 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "61. Yu, W., Liu, Y., Zhu, X., Cao, H., Sun, X., Bai, X.: Turning a clip model into a scene text spotter. IEEE TPAMI (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "type": "text", + "content": "62. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: ICCV. pp. 6023-6032 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "63. Zhan, F., Xue, C., Lu, S.: Ga-dan: Geometry-aware domain adaptation network for scene text detection and recognition. In: ICCV. pp. 9105-9115 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "64. Zhang, B., Wang, Y., Hou, W., Wu, H., Wang, J., Okumura, M., Shinozaki, T.: Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. In: NeurIPS. vol. 34, pp. 18408-18419 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "65. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk minimization. arXiv:1710.09412 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "66. Zhao, Z., Yang, L., Long, S., Pi, J., Zhou, L., Wang, J.: Augmentation matters: A simple-yet-effective approach to semi-supervised semantic segmentation. In: CVPR. pp. 11350-11359 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 304 + ], + "type": "text", + "content": "67. Zhou, X., Yao, C., Wen, H., Wang, Y., Zhou, S., He, W., Liang, J.: East: an efficient and accurate scene text detector. In: CVPR. pp. 5551-5560 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "68. Zhu, Y., Chen, J., Liang, L., Kuang, Z., Jin, L., Zhang, W.: Fourier contour embedding for arbitrary-shaped text detection. In: CVPR. pp. 3123-3131 (2021)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "T. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Bridging the Gap Between Human Motion and Action Semantics via Kinematics Phrases/ceda5829-f6a0-4ea2-94dd-2ddc6da8e6ed_content_list.json b/2024/Bridging the Gap Between Human Motion and Action Semantics via Kinematics Phrases/ceda5829-f6a0-4ea2-94dd-2ddc6da8e6ed_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..199c2905d19d44f7444f7141f598bb8c0b9a13cf --- /dev/null +++ b/2024/Bridging the Gap Between Human Motion and Action Semantics via Kinematics Phrases/ceda5829-f6a0-4ea2-94dd-2ddc6da8e6ed_content_list.json @@ -0,0 +1,1759 @@ +[ + { + "type": "text", + "text": "Bridging the Gap between Human Motion and Action Semantics via Kinematic Phrases", + "text_level": 1, + "bbox": [ + 233, + 140, + 772, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinpeng Liu $^{1}$ , Yong-Lu Li $^{1\\star}$ , Ailing Zeng $^{2}$ , Zizheng Zhou $^{1}$ , Yang You $^{3}$ , and Cewu Lu $^{1\\star}$", + "bbox": [ + 222, + 210, + 781, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai Jiao Tong University", + "bbox": [ + 388, + 253, + 612, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Tencent", + "bbox": [ + 465, + 268, + 535, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Stanford University", + "bbox": [ + 426, + 282, + 575, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xinpengliu0907,ailingzengzzz} $@$ gmail.com,", + "bbox": [ + 341, + 297, + 658, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yonglu_li, zhou_zz, lucewu}@sjtu.edu.cn, yangyou@stanford.edu", + "bbox": [ + 269, + 311, + 733, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Motion understanding aims to establish a reliable mapping between motion and action semantics, while it is a challenging many-to-many problem. An abstract action semantic (i.e., walk forwards) could be conveyed by perceptually diverse motions (walking with arms up or swinging). In contrast, a motion could carry different semantics w.r.t. its context and intention. This makes an elegant mapping between them difficult. Previous attempts adopted direct-mapping paradigms with limited reliability. Also, current automatic metrics fail to provide reliable assessments of the consistency between motions and action semantics. We identify the source of these problems as the significant gap between the two modalities. To alleviate this gap, we propose Kinematic Phrases (KP) that take the objective kinematic facts of human motion with proper abstraction, interpretability, and generality. Based on KP, we can unify a motion knowledge base and build a motion understanding system. Meanwhile, KP can be automatically converted from motions to text descriptions with no subjective bias, inspiring Kinematic Prompt Generation (KPG) as a novel white-box motion generation benchmark. In extensive experiments, our approach shows superiority over other methods. Our project is available at https://foruck.github.io/KP/.", + "bbox": [ + 259, + 362, + 743, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Motion Representation $\\cdot$ Motion Understanding $\\cdot$ Text-to-Motion Benchmark", + "bbox": [ + 259, + 638, + 743, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 693, + 377, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human motion understanding has a wide range of applications, including activity understanding [31], autonomous driving [36], robotics [27], and automatic animation [49], making it increasingly attractive. The core is to establish a mapping between the motion space and the action semantics space. The motion space indicates a space of sequential 3D human representations, e.g., 3D pose or SMPL [33]/SMPL-X [37] parameter sequence, while the action semantic space", + "bbox": [ + 212, + 724, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding authors.", + "bbox": [ + 217, + 824, + 392, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2f9a250c0b660f4e9de119af4ffb199a2a442a893c9c8c1ed7f35d6e24a7f298.jpg", + "image_caption": [ + "Fig. 1: The huge gap between motion and action semantics results in the many-to-many problem. We propose Kinematic Phrases (KP) as an intermediate to bridge the gap. KPs objectively capture human kinematic cues. It properly abstracts diverse motions with interpretability. As shown, the Phrases in the yellow box could capture key patterns of walk for diverse motions." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "can be represented as action categories or sentences described by natural language.", + "bbox": [ + 212, + 464, + 785, + 494 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, a growing focus has been on generative mapping from semantics to motion, including action category-based generation [38] and text-based generation [7, 14, 35, 39, 48, 54, 55]. Most of them typically build a mapping that links motion and semantics either directly or via motion latent, with understated concerns for intermediate motion-semantic structures. However, these models suffer from inferior reliability. They cannot guarantee they generated correct samples without human filtering. Additionally, the existing evaluation of motion generation is problematic. Widely adopted FID and R-Precision rely on the latent space from a black-box pre-trained model, which might fail to out-of-distribution (OOD) and over-fitting cases. There is a long-standing need for an evaluation method that can cheaply and reliably assess whether a generated motion is consistent with particular action semantics. We identify the essence of these as the significant gap between raw human motion and action semantics, which makes direct mapping hard to learn.", + "bbox": [ + 212, + 501, + 787, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As in Fig. 1, an action semantics can correspond to diverse motions. For instance, a person could walk in countless ways with diverse motions, either with arms up or swinging, while action semantics tend to abstract these away from a walking motion. Additionally, they are robust against small perturbations, while motion is more specific and complex, with representations changing vastly when perturbed or mis-captured. Moreover, a motion sequence could have diverse semantics w.r.t. contexts. Modeling this many-to-many mapping between motion and semantics is challenging.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To bridge this gap between motion and action semantics, we propose Kinematic Phrases (KP), an interpretable intermediate representation. KP focuses on the objective kinematic facts, which are usually omitted by general action semantics, like left-hand moving forwards then backward. KP is designed as qualitative categorical representations of these facts. For objectivity and actuality, KP captures sign changes with minimal pre-defined standards. Inspired by previous studies on kinematic human motion representation [5,28], KP is proposed", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "as six types shown in Fig. 1, covering joint positions, joint pair positions and distances, limb angles and directions, and global velocity. Note that, although KP can be described by natural language, a major difference is that KP is strictly dedicated to objective kinematic facts instead of coarse actions such as surrender or fine-grained actions like raise both hands.", + "bbox": [ + 212, + 253, + 486, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We highlight three advantages of KP. First, KP offers proper abstraction, which disentangles motion perturbations and semantics changes, easing the learning process. Even though the motion differs significantly, KP manages to capture walk patterns easily. Second, KP is inter", + "bbox": [ + 212, + 422, + 488, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pretable, as it can be viewed as instructions on executing the action, making it easily understandable to humans. Finally, KP is general, as it can be automatically extracted from different modalities of human motion, including skeleton and SMPL parameters. The conversion from KP to text is also effortless.", + "bbox": [ + 212, + 542, + 785, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With KP as an intermediate representation, a unified large-scale motion knowledge base is constructed. Then, to fully exploit KP and the knowledge base, we build a motion understanding system with KP mediation. In detail, we learn a motion-KP joint latent space for multiple motion understanding applications, including motion interpolation, modification, and generation. Moreover, leveraging the interpretability of KP, we propose a benchmark called Kinematic Prompts Generation (KPG), which generates motion from text prompts converted from KPs. Thanks to the consistency and convenience of the KP-to-text conversion, KPG enables efficient white-box motion generation evaluation. Given the bottle-necked evaluation of the current motion generation benchmarks [14,41] as shown in Fig. 2, KPG could be a promising complementary benchmark. Results in Tab. 2 show that KPG brings novel insights into the current text-to-motion progress.", + "bbox": [ + 212, + 608, + 787, + 805 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions are: (1) We propose KP as an intermediate representation to bridge the gap between motion and action semantics. (2) We build a novel", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b9aefb9afaed2df8cb29e3234b1a68e80a618a25056dc470d94a3be214c7a60d.jpg", + "image_caption": [ + "Fig. 2: HumanML3D [14] metric values are approaching even surpassing GT-level while being increasingly indistinguishable. The higher-than-GT R-Precision could be insufficient as a semantic consistency indicator." + ], + "image_footnote": [], + "bbox": [ + 506, + 282, + 779, + 425 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Kinematic Phrases", + "bbox": [ + 604, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "motion understanding system using KP and the aggregated large-scale knowledge base. (3) We propose KPG as a benchmark for efficient white-box motion generation evaluation. Promising results are achieved on motion interpolation and generation tasks. Moreover, extensive user studies are conducted, verifying the efficacy of our methods, and the consistency between KPG evaluation and human perception.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 258, + 397, + 275 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motion Representation An intuitive motion representation is a sequence of static pose representations, like joint locations and limb rotations. Efforts are paid to address the discontinuity of rotation for deep-learning methods [6,59]. Recent works on parametric body models [33,37] enable a more realistic body representation. Meanwhile, [42] proposed Posebits, representing pose with boolean geometric part relationships. [9, 10] translates Posebits into text descriptions. These abstract representations are flexible and insensitive to little perturbations, but their static nature ignores motion dynamics. [46] acquire similar fine-grained descriptions from human annotation, while [3, 51] adopted large-scale language models. However, few recognize their potential in bridging the low-level motion and the high-level action semantics. Phase functions [20], Labanotations [28], and learned Motion Words [1] were also explored, though limited to specific actions like locomotion and dancing.", + "bbox": [ + 212, + 289, + 789, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motion Generation can be conditioned by its prefix/suffix [2, 17, 19], action categories [16, 38, 52], or audio [29, 30]. Text-based motion generation has developed rapidly with the proposal of text-motion datasets [14, 43]. Early works [14, 39, 44] used VAEs, while some [21, 32, 47] extended the CLIP [45] space to motion. Diffusion models [8, 48, 50, 55] were also adopted. MAA [4] adopted a U-Net structure. Some [40, 56] explored retrieval-based methods. [25] aimed at controllable generation, while [53] introduced physical constraints. However, most approaches still suffer from the gap between motion and action semantics. Multiple methods [7, 15, 26, 35, 54, 58, 61] adopted (VQ-)VAE-compressed motion representation as mediation, while in the current data-limited situation, we identify that this single-modality compression might be sub-optimal. Instead, KP could alleviate this by introducing explicit semantic-geometric correlation.", + "bbox": [ + 212, + 506, + 789, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Kinematic Phrase Base", + "text_level": 1, + "bbox": [ + 215, + 709, + 478, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Kinematic Phrases", + "text_level": 1, + "bbox": [ + 215, + 739, + 419, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Kinematic Phrases abstract motion into objective kinematic facts like left-hand moves up qualitatively. We take inspiration from previous kinematic motion representations [28] and qualitative static pose representations [9,11,42], proposing six types of KP to comprehensively represent motion from different kinematic hierarchies: For joint movements, there are 34 Position Phrases (PPs). For joint", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ee41f23925b0bf02a9ea76e0fb59029b4cadfab8609702d49dde943777902438.jpg", + "image_caption": [ + "Fig.3: Six types of KP from four kinematic hierarchies are extracted from a motion sequence. A scalar indicator $s_i$ is calculated per Phrase per frame. Its sign categorizes the corresponding Phrase." + ], + "image_footnote": [], + "bbox": [ + 228, + 143, + 772, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "pair movements, there are 242 Pairwise Relative Position Phrases (PRPPs) and 81 Pairwise Distance Phrases (PDPs). For limb movements, there are 8 Limb Angle Phrases (LAPs) and 24 Limb Orientation Phrases (LOPs). For whole-body movements, there are 3 Global Velocity Phrases (GVPs). KP extraction is based on a skeleton sequence $X = \\{x_{i} | x_{i} \\in \\mathcal{R}^{n_{k} \\times 3}\\}_{i=1}^{t}$ , where $n_{k}$ is the number of joints ( $n_{k} = 17$ here), $x_{i}$ is the joint coordinates at $i$ -th frame, and $t$ is the sequence length. Note that $x_{i}^{0}$ indicates the pelvis/root joint. For each Phrase, a scalar indicator sequence is calculated from the skeleton sequence. Phrases are extracted as per-frame categorical representations w.r.t. indicator signs. Unlike previous efforts [9, 42], we limit the KP criteria as the indicator signs to minimize human-defined biases (e.g., numerical criteria on the closeness of two joints) for objectivity and actuality. Fig. 3 illustrated the procedure.", + "bbox": [ + 212, + 494, + 787, + 678 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reference Vectors are first constructed, indicating right, upward, and forward directions from a human cognitive view. We aim at the egocentric reference frames that human tends to use when performing actions. The negative direction of gravity is adopted as upward vector $r^u$ , the vector from left hip to right hip is adopted as right vector $r^r$ , and the forward vector is calculated as $r^f = r^u \\times r^r$ . These vectors of each frame are denoted as $R^{\\cdot} = \\{r_{i}^{\\cdot}\\}_{i=1}^{t}$ .", + "bbox": [ + 212, + 680, + 787, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Position Phrase (PP) focuses on the movement direction of joint $x^{j}$ w.r.t. reference vector $R$ . The indicator for PP at $i$ -th frame is calculated as", + "bbox": [ + 212, + 773, + 785, + 805 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i} ^ {(j, \\cdot)} = \\langle (x _ {i} ^ {j} - x _ {i} ^ {0}), r _ {i} ^ {\\cdot} \\rangle - \\langle (x _ {i - 1} ^ {j} - x _ {i - 1} ^ {0}), r _ {i - 1} ^ {\\cdot} \\rangle . \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 821, + 785, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Kinematic Phrases", + "bbox": [ + 604, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The sign of $s_i^{(j,\\cdot)}$ categorizes PP into moving along/against $R'$ , or relatively static along $R'$ for indicators with small amplitudes. 12 joints (except the pelvis, eyes, and hips) and 3 reference vectors resulted in originally 36 PPs. Specifically, shoulders are excluded for the leftward vector, and hips are excluded for the upward vector, resulting in 36-2=34 PPs.", + "bbox": [ + 212, + 145, + 782, + 222 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pairwise Relative Position Phrase (PRPP) describes the relative position of a pair of joints $(x^{j}, x^{k})$ w.r.t. reference vector $R$ . PRPP indicator at $i$ -th frame is", + "bbox": [ + 212, + 222, + 782, + 265 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i} ^ {(j, k, \\cdot)} = \\langle \\left(x _ {i} ^ {j} - x _ {i} ^ {k}\\right), r _ {i} ^ {\\cdot} \\rangle . \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 265, + 784, + 284 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For (L-Hand, R-Hand) and forward vector $R^f$ , PRPP could be L-Hand behind /in front of R-Hand according to the sign of $s_i^{(j,k,\\cdot)}$ . 136 joint pairs and 3 reference vectors composed 408 Phrases originally. 24 joint pairs linked by limbs are filtered out. 30 Eye-related pairs are filtered out except (left eye, right eye) due to the others are covered by head-related pairs. 4 triplets that the relationship barely changes along the reference vector (e.g., right knee, left hip, and leftward vector) are filtered out. These result in $408 - 24^{*}3 - 30^{*}3 - 4 = 242$ PRPPs.", + "bbox": [ + 212, + 287, + 784, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pairwise Distance Phrase (PDP) describes how the L2 distance between a pair of joints $(x^{j}, x^{k})$ changes. The indicator for PDP is calculated as", + "bbox": [ + 214, + 396, + 782, + 426 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i} ^ {(j, k)} = \\left\\| x _ {i} ^ {j} - x _ {i} ^ {k} \\right\\| _ {2} - \\left\\| x _ {i - 1} ^ {j} - x _ {i - 1} ^ {k} \\right\\| _ {2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 436, + 784, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The sign of $s_i^{(j,k)}$ categorizes PDP into moving closer/away, or relatively static. Among the original 105 joint pairs, 24 pairs that are in the skeleton topology are excluded, such as the hand and elbow, resulting in 81 PDPs.", + "bbox": [ + 214, + 465, + 782, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Limb Angle Phrase (LAP) targets at the change of bend angle between two connected limbs $(x^{j}, x^{k})$ and $(x^{j}, x^{l})$ . The LAP indicator is calculated as", + "bbox": [ + 214, + 513, + 782, + 544 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i} ^ {(j, k, l)} = \\operatorname {a r c c o s} \\left(\\left\\langle x _ {i} ^ {k} - x _ {i} ^ {j}, x _ {i} ^ {l} - x _ {i} ^ {j} \\right\\rangle\\right) - \\operatorname {a r c c o s} \\left(\\left\\langle x _ {i - 1} ^ {k} - x _ {i - 1} ^ {j}, x _ {i - 1} ^ {l} - x _ {i - 1} ^ {j} \\right\\rangle\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 232, + 553, + 784, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "8 LAPs describe the limb chain $(x^{j},x^{k}) - (x^{j},x^{l})$ as bending or unbending.", + "bbox": [ + 214, + 582, + 750, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Limb Orientation Phrase (LOP) describes the orientation of the limb $(x^{j}, x^{k})$ w.r.t. $R^{\\prime}$ , with $x^{k}$ being the distal limb. The LOP indicator is", + "bbox": [ + 214, + 598, + 782, + 628 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i} ^ {(j, k, \\cdot)} = \\left\\langle x _ {i} ^ {k} - x _ {i} ^ {j}, r _ {i} ^ {\\prime} \\right\\rangle . \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 423, + 638, + 784, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The sign of $s_i^{(j,k,\\cdot)}$ categorizes the LOP $(x^j, x^k)$ pointing along/against $R$ , or a placeholder category for those with little magnitude. 4 arms, 4 leg limbs, head, collarbones, hips, torsos, and upper body, paired with the 3 reference vectors, result in 45 LOPs in total. 21 pairs that barely change along the reference vector (e.g., left-right hip, leftward vector) are filtered out, resulting in 45-21=24 LOPs.", + "bbox": [ + 214, + 666, + 784, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Global Velocity Phrase (GVP) describes the direction of global velocity with respect to $R$ . The indicator is calculated as", + "bbox": [ + 214, + 744, + 782, + 773 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i} = \\left\\langle x _ {i + 1} ^ {0} - x _ {i} ^ {0}, r _ {i} ^ {\\prime} \\right\\rangle . \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 784, + 784, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The three categories are moving along/against $R$ , or static along $R$ according to the sign of $s_i$ .", + "bbox": [ + 214, + 809, + 782, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "These result in 392 Phrases in total, covering motion diversity and distribution from various levels. While we clarify that these Phrases do not rule out the possibility of other possible useful potentials.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Constructing Kinematic Phrase Base", + "text_level": 1, + "bbox": [ + 214, + 214, + 568, + 231 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "KP enables us to unify motion data with different formats to construct a large knowledge base containing motion, text, and KP thanks to its generality. Motion sequences of different representations are collected, including 3D skeleton sequences and SMPL [33] /SMPL-X [37] parameter sequences. The sequences are first re-sampled to $30\\mathrm{Hz}$ and rotated so that the negative direction of the z-axis is the gravity direction. Then, the sequences are converted into 3D skeleton sequences for KP extraction as in Sec. 3.1. Text annotations attached to the sequences are directly saved. For sequences with action category annotation, the category name is saved. For those with neither text nor action category, the text information is set from its attached additional information, like objects for SAMP [18]. Finally, we collect 140K motion sequences from 11 datasets. More details are included in the appendix.", + "bbox": [ + 212, + 239, + 787, + 421 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Applications of KP", + "text_level": 1, + "bbox": [ + 214, + 445, + 439, + 463 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Motion Understanding with KP", + "text_level": 1, + "bbox": [ + 214, + 477, + 527, + 493 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "An intuitive application of KP is to leverage it for better motion understanding. To achieve this, we first learn a motion-KP joint space, which exploits the clarity and interpretability of KP. Then, we adopt the learned latent space for multiple motion understanding tasks. An overview is illustrated in Fig. 4.", + "bbox": [ + 212, + 503, + 784, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Preliminaries. Motion is represented as a pose sequence with $n$ frames as $M = \\{m_i\\}_{i=1}^n$ . In detail, SMPL [33] pose parameters are transformed to the 6D representation [59], then concatenated with the velocity of the root, resulting in a 147-dim representation per frame. KP is represented by indicator signs as $C = \\{c_i\\}_{i=1}^n, c_i \\in \\{1,0,-1\\}^{392}$ .", + "bbox": [ + 214, + 585, + 787, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Joint Space Learning. The joint space is learned with Motion VAE $\\{\\mathcal{E}_m,\\mathcal{D}_m\\}$ and KP VAE $\\{\\mathcal{E}_p,\\mathcal{D}_p\\}$ . All VAEs follow the transformer-based architecture from [38]. Motion encoder $\\mathcal{E}_m$ takes motion $M$ and two distribution tokens $m_{\\mu},m_{\\sigma}$ as input, and the outputs corresponding to the distribution tokens are taken as the $\\mu_{m}$ and $\\sigma_{m}$ of the Gaussian distribution. Then, motion decoder $\\mathcal{D}_m$ takes $z_{m}\\sim \\mathcal{G}(\\mu_{m},\\sigma_{m})$ as $K,V$ , and a sinusoidal positional encoding of the expected duration as $Q$ . The output is fed into a linear layer to reconstruct motion $\\hat{M}$ . KP VAE $\\{\\mathcal{E}_p,\\mathcal{D}_p\\}$ operates similarly, with the sign of $\\mathcal{D}_p$ output as the predicted KP $\\hat{C}$ . Notice that the $\\mathcal{D}_m,\\mathcal{D}_p$ could take arbitrary combinations of $z_{m},z_{p}$ as input, outputting $\\hat{M},\\hat{C}$ . To learn the joint space, besides the conventional loss", + "bbox": [ + 214, + 685, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Kinematic Phrases", + "bbox": [ + 604, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/505be9df18e697cb5c0fde68195f4fdb4b059e0510cce4f160c06ab451bdbabb.jpg", + "image_caption": [ + "Fig. 4: We train a motion-KP joint latent space to exploit the clarity and interpretability of KP. The space is then applied to multiple tasks, including motion interpolation, modification, and generation." + ], + "image_footnote": [], + "bbox": [ + 253, + 143, + 759, + 348 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of reconstruction and KL divergence, we also align the two VAEs similarly to [39]. In this way, we guide the learned motion latent space with the structure of KP. Meanwhile, we randomly corrupt samples during training by setting a small portion of KP as 0. The idea is that the overall representation should be robust with a small portion of KP unknown. Also, the missing Phrases should be recovered from existing Phrases. This mines the correlation among Phrases and increases the robustness of the space. More details are in the supplementary.", + "bbox": [ + 212, + 431, + 787, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Motion Understanding Tasks with Joint Space. With the joint space, we can perform both low-level and high-level motion understanding tasks. We introduce three applications to show the capability of KP, as shown in Fig. 4. For motion interpolation given a corrupted motion sequence $\\tilde{M}$ , we extract its corresponding KP sequence $\\tilde{C}$ , then feed them to encoders $\\mathcal{E}_m, \\mathcal{E}_p$ and decoder $\\mathcal{D}_p$ , resulting in the estimated KP sequence $\\hat{C}$ . $\\hat{C}$ and $\\tilde{M}$ are fed into $\\mathcal{E}_m, \\mathcal{E}_p$ and $\\mathcal{D}_m$ , resulting in interpolated $\\hat{M}$ . For motion modification on motion $M$ , modifications could be made on the extracted KP sequence $C$ resulting in $\\tilde{C}$ . Motion frames to modify are then masked, getting $\\tilde{M}$ . $\\tilde{M}, \\tilde{C}$ are fed into $\\mathcal{E}_m, \\mathcal{E}_p$ and $\\mathcal{D}_m$ , getting the interpolated $\\hat{M}$ . For motion generation, we simply replace the motion latent space in MLD [7] with the latent space of our Motion VAE. Experiments show that KP manages to boost MLD with a considerable margin.", + "bbox": [ + 212, + 566, + 787, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Motion Benchmarking with Kinematic Prompt Generation", + "text_level": 1, + "bbox": [ + 212, + 779, + 748, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A more interesting and important application is benchmarking motion generation with the interpretability and objectivity of KP.", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb1cabaf2b56169e8c55ab971f4da7cf711db6c5e367379870bf8c9032d58b0d.jpg", + "image_caption": [ + "Fig. 5: Kinematic Prompt Generation. 7,796 prompts are converted from KP with templates, corresponding to certain KP patterns. We calculate the generation accuracy by checking the appearance of the patterns in KP extracted from the generated motion." + ], + "image_footnote": [], + "bbox": [ + 246, + 143, + 754, + 340 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Before that, we first analyze current benchmarks. A crucial aspect of motion generation evaluation is motion-semantic consistency. The gold standard is user study. However, it is expensive and inefficient to scale. Early metrics like MPJPE (Mean Per Joint Position Error) and MAE (Mean Angle Error) mechanically calculate the error between the generated and GT samples. These metrics fail to reveal the real ability of generative models: What if the models memorize GT samples? Or what if the samples are diverse from GT but also true? FID (Frechet Inception Distance) is adopted to mitigate this issue. However, it provides a macro view of the quality of all generated samples without guarantees for individual samples. [14] proposed R-Precision, using a pre-trained text-motion matching model to examine whether the generated samples carry true semantics. They both rely on the latent space from a black-box pre-trained model, which is not credible. Besides, models might learn short paths to over-fit the pre-trained model. Meanwhile, since automatic mapping from motion to semantics is still an unsettled problem, adopting it to evaluate motion generation is not a decent choice. Moreover, most current motion generation evaluations are performed on datasets [14,22,41] with considerable complex everyday actions, further increasing the difficulty. As shown in Fig. 2 and Tab. 1, the current evaluation metrics are becoming increasingly indistinguishable. Also, the higher-than-GT R-Precision could be insufficient as an indicator of semantic consistency.", + "bbox": [ + 212, + 426, + 787, + 729 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To this end, we propose a novel benchmark: Kinematic Prompts Generation (KPG). Instead of previous benchmarks focusing on everyday activities or sports, we step back in the complexity of action semantics. Based on KP, KPG focuses on evaluating whether the models could generate motion sequences consistent with specific kinematic facts given text prompts. An overview is given in Fig. 5.", + "bbox": [ + 212, + 731, + 787, + 808 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In detail, we convert KP into 7,796 text prompts with templates as in Fig. 5. The prompts could be categorized into 4 groups. First, each atomic prompt", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Kinematic Phrases", + "bbox": [ + 604, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "involves only 1 KP, such as \"left hand moves upwards\", which evaluates the basic semantic consistency of text-to-motion models. Notice that PRPP and LOP are excluded since they represent static states, resulting in 252 prompts. Second, each repetitive prompt involves a KP being repeated two or three times., like \"left arm bends twice\", evaluating the basic counting ability. PRPP, LOP, and GVP are excluded, bringing 492 prompts. Third, each sequential prompt involves two independent KP to be sequentially executed, like \"left arm bends, then right arm bends\". 3,912 sequential prompts are randomly sampled. Finally, each simultaneous prompt involves two independent KPs to be simultaneously executed, like \"left arm bends, and simultaneously, right arm bends.\" 3,120 prompts are randomly sampled. The sequential and simultaneous prompts examine whether the model can understand the composition of simple actions. Given prompt $T_{i} \\in T$ from Phrase $c_{i}$ (or Phrases $(c_{i}^{1}, c_{i}^{2})$ ), the model generates motion $\\hat{M}_{i}$ , along with extracted KP $\\hat{C}_{i}$ . Then, whether $T_{i}$ is hit could be inferred via designed protocols by identifying certain KP patterns in $\\hat{C}_{i}$ . In detail, the atomic prompt is identified as hit if $c_{i}$ appears for more than 5 consecutive frames in $\\hat{C}_{i}$ . The repetitive prompt is identified as hit if and only if $c_{i}$ appears for at least 5 consecutive frames in $\\hat{C}_{i}$ for corresponding times. The sequential prompt is identified as hit if $c_{i}^{2}$ appears for at least 5 consecutive frames after (maybe not immediately) $c_{i}^{1}$ does the same. The simultaneous prompt is identified as hit if $c_{i}^{1}$ and $c_{i}^{2}$ appear simultaneously for at least 5 consecutive frames. The calculation involves no black-box model thanks to KP, presenting a fully white-box evaluation. Also, with the effortless motion-to-KP conversion, the computation could be conducted automatically. More details are in the appendix.", + "bbox": [ + 212, + 146, + 787, + 511 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiment", + "text_level": 1, + "bbox": [ + 215, + 532, + 366, + 550 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details. HumanML3D [14] test split is held out for evaluation, with the rest of KPB for training. The batch size is set as 288, and an AdamW optimizer with a learning rate of 1e-4 is adopted. We randomly corrupt less than $20\\%$ of the Phrases for a sample. The Motion-KP joint space is trained for 6,000 epochs. While the text-to-motion latent diffusion model is trained for 3,000 epochs, with the joint space frozen. All experiments are conducted in 4 NVIDIA RTX 3090 GPUs. More details are in the appendix.", + "bbox": [ + 212, + 564, + 787, + 672 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Motion Interpolation", + "text_level": 1, + "bbox": [ + 215, + 691, + 437, + 708 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Following [23], $50\\%$ frames are randomly masked for interpolation. In Tab. 1, our method provides better FID. With KPB, Diversity is increased.", + "bbox": [ + 212, + 717, + 785, + 748 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Motion Generation", + "text_level": 1, + "bbox": [ + 215, + 768, + 419, + 784 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results on conventional text to motion are shown in Tab. 1. We adopt the HumanML3D test set [14] for conventional text-to-motion evaluation. The evaluation model from [14] is adopted to calculate R-Precision, FID, Diversity,", + "bbox": [ + 212, + 794, + 785, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/3aaf9d47c3d9aac7a9b1d750804dabb41d5e7e16cf0e6744ceb86f21a8cf4905.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsInterpolationGeneration
FID↓Diversity→R-P@1↑FID↓Diversity→Multimodality
GT0.0029.5030.5110.0029.503-
TEMOS [39]--0.4243.7348.9730.368
T2M [14]--0.4551.0679.1882.090
MotionDiffuse [55]--0.4910.6309.4100.730
MDM [48]2.6988.4200.3200.5449.5592.799
TM2T [15]--0.4241.5018.5892.424
MLD [7]--0.4810.4739.7242.413
T2M-GPT [54]--0.4920.1419.7221.831
Fg-T2M [50]--0.4920.2439.2781.614
M2DM [26]--0.4970.3529.9263.587
AttT2M [58]--0.4990.1129.7002.452
ReMoDiffuse [56]--0.5100.1039.0181.795
MotionGPT [23]0.2149.5600.4920.2329.5282.008
FineMoGen [57]--0.5040.1519.2632.696
GraphMotion [24]--0.5040.1169.6922.766
AvatarGPT [60]--0.5100.1689.624-
GUESS [12]--0.5030.1099.8262.430
MDD [34]--0.5150.0729.6831.869
MoMask [13]--0.5210.045-1.241
Ours0.1929.6230.4960.2759.9752.218
Ours w/ KPB0.21410.1420.4710.64110.2622.891
", + "bbox": [ + 218, + 143, + 560, + 343 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/876177e7dcfd3b55c44669db4f1509c105ef03d6a26523ffd1ec139a8e88abb8.jpg", + "image_caption": [ + "Table 1: Results on HumanML3D. R-P@1 is short Fig.6: R-Precision could contrafor R-Precision@1. Most methods show comparadict to user reviews, indicatingble performance. Also, the higher-than-GT phethat R-Precision could produce nomenon indicates that R-Precision might be insuf-unfaithful evaluations of semantic ficient as a semantic consistency indicator. consistency." + ], + "image_footnote": [], + "bbox": [ + 560, + 145, + 790, + 338 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "and Multimodality. We run the evaluation 20 times and report the average value. Our method is competitive without KPB. However, there exist some unsatisfying observations. First, KPB brings a counter-intuitive performance drop. Second, the metric values of most techniques are rather close (e.g., difference $< 0.03$ for R-P@1), and no substantial advantage is perceived for one method over others. Also, as discovered by MAA [4], despite its relatively poor R-Precision and FID, MAA manages to achieve better user reviews, indicating the potential loss of efficacy of these conventional metrics. This is further revealed in Fig. 6, where a contradiction is observed between human judgment and R-precision. Moreover, some methods show higher-than-GT R-P@1, which indicates that R-Precision might be insufficient as a semantic consistency indicator. Given these, we further switch to KPG and pursue further insights.", + "bbox": [ + 212, + 454, + 787, + 636 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results on KPG are shown in Tab. 2. Besides accuracy, diversity and R-precision are reported. We run the evaluation 20 times and report the average metric value. The white-box designation of KPG evaluation makes it possible to evaluate the different aspects of the models. Meanwhile, KPG is considered an easier task than conventional text-based motion generation since it is targeted at much less complex action semantics. Results show that previous methods are not performing as well as expected. Also, R-Precision fails to model the performance on KPG with trivially low values. Most methods perform well for atomic prompts, but when it comes to compositions, the accuracy drops. Early efforts like HMDM [48] and MotionDiffuse [55] are surprisingly competitive, even surpassing several more recent works. T2M-GPT [54] outperforms other methods by a considerable margin, but it fails for repetitive prompts. An unexpected", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Kinematic Phrases", + "bbox": [ + 604, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/4dfd63441d54cd627ed44ef79e9a1bf94e6bf9dc1c423a75061a7229cf06e34f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsAccuracyDiversityR-P@1
AtomicRepetitiveSequentialSimultaneousOverall
HMDM [48]93.2523.9856.1130.9045.165.9230.034
MotionDiffuse [55]93.2122.1551.5427.7641.495.2570.051
T2M-GPT [54]97.221.0270.2442.4755.596.1530.042
MLD [7]92.0622.7648.2628.0139.945.5260.061
ReMoDiffuse [56]53.173.0532.1631.0630.565.8090.048
MotionGPT [23]92.0616.8755.9831.7644.965.7760.039
MoMask [13]96.4328.8659.1032.6647.795.9650.060
Ours98.8026.0271.3243.0857.866.0480.053
", + "bbox": [ + 276, + 143, + 728, + 251 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 2: Results on KPG demonstrate the various strengths of different methods.", + "bbox": [ + 225, + 252, + 774, + 266 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "phenomenon is the failure of ReMoDiffuse [56], which might be constrained by its retrieval-based methodology since most prompts in KPG are not available in the retrieval database. Our method, which replaces the latent space of MLD [7], delivers substantial improvements, but the accuracy remains below $60\\%$ , which is far from satisfying. A considerable gap exists between existing efforts and ideal motion generation models. More analyses are in the appendix.", + "bbox": [ + 212, + 311, + 785, + 402 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "User Study. To reach a convincing performance comparison, we conduct user studies on both HumanML3D and KPG. Our user study is different from previous efforts in two aspects. First, instead of testing a small set of text prompts (less than 50 in previous works [7,48]), we randomly select 600 sentences from the HumanML3D test set. By scaling up, the result is convincing in reflecting the ability to generate motion for diverse text inputs. Second, without asking the volunteers to give a general rating for each sample or to choose between different samples, we ask them two questions: 1) Do the motion and the text match? and 2) Is the motion natural? For Q1, three choices are given as \"No, Partially, Yes\". For Q2, two choices are given as \"Yes, No\". In this way, we explicitly evaluate two aspects of text-to-motion tasks: semantic consistency and naturalness, corresponding to R-Precision and FID. For each prompt, we generate one sample considering the annotation cost. We claim that the models should generate natural text-matching motion most of the time so that the one-sample setting would not hurt the fidelity of our user study. 36 volunteers are invited, each reviewing 200 sequences. Thus each sequence receives 3 user reviews. Also, we compute R-precision@1 of the generated sequences.", + "bbox": [ + 212, + 430, + 787, + 686 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Besides our model, we select MDM [48] as a naive baseline, T2M-GPT [54] as the SOTA model given the results in Tab. 1 and Tab. 2. MLD [7] is also selected as a direct baseline of our proposed method. Results on HumanML3D [14] are shown in Fig. 7. Though our method is not superior in R-Precision, we receive competitive user reviews, showcasing the efficacy of our motion-KP joint space. Also please notice that our model only contains 45.1M parameters, while T2M-GPT contains 228M parameters. T2M-GPT and MLD present similar R-Precision, but only T2M-GPT manages to keep a good performance with user reviews. Moreover, the discrepancy between R-Precision and user reviews is revealed in both absolute value and trends, also in Fig. 6.", + "bbox": [ + 212, + 688, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e3fac0ccfdc387fd86d7b6828aa4ad633b3d6f5bbc3d86281ea1108948729ee4.jpg", + "image_caption": [ + "Fig. 7: User study on HumanML3D, with Fig. 8: User study on KPG, with \"Y\" for \"Y\" for Yes and \"P\" for partially. Yes and \"P\" for partially." + ], + "image_footnote": [], + "bbox": [ + 227, + 142, + 488, + 265 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2337b3ab3bcc685556299264bc486c0cef254659834cd7dd6fcf0362ba9726b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 146, + 777, + 263 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/550a8c39304d5e1005aa4a2da9d188b0cf6f100557600607c7695ad46a76d9ff.jpg", + "image_caption": [ + "Fig.9: Our model supports fine-grained modification on motion via KP." + ], + "image_footnote": [], + "bbox": [ + 225, + 318, + 398, + 444 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/99d5e32a546234b54428d7d13d7545aff8c409e2bfff36f076bae8af62106c73.jpg", + "image_caption": [ + "Fig. 10: Visualization of generated samples. Compared to T2M-GPT, our method provides a better response to prompts with explicit constraints on specific body parts." + ], + "image_footnote": [], + "bbox": [ + 426, + 321, + 769, + 441 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A similar user study on KPG is conducted in Fig. 7 with 100 randomly selected prompts from KPG involving T2M-GPT and our model. Fig. 8 demonstrates that KP-inferred Accuracy and user reviews share similar trends. Meanwhile, KP and user study give the same reviews for $84\\%$ of the samples. We believe KPG could thus be a first step towards white-box automatic motion generation evaluation for semantic consistency. More analyses are in the appendix.", + "bbox": [ + 212, + 526, + 787, + 617 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Visualization", + "text_level": 1, + "bbox": [ + 214, + 640, + 370, + 655 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A modification sample is shown in Fig. 9. By modifying KP, we could edit arbitrary motion at a fine-grained level. Fig. 10 compares generated samples of T2M-GPT and our methods. Our method properly responds to text prompts with constraints on specific body parts thanks to KP, which explicitly models the action semantics with body part kinematics cues. Note that T2M-GPT generates redundant motion for simple prompts, while our method provides more concise and precise results. More visualizations are in the appendix.", + "bbox": [ + 212, + 666, + 787, + 773 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 214, + 796, + 401, + 810 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "KPG ablation study results are in Tab. 3.", + "bbox": [ + 214, + 821, + 517, + 835 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Kinematic Phrases", + "bbox": [ + 604, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "KP latent. By using a latent space without KP, a substantial performance drop is perceived, revealing the significance of KP in conveying action semantics.", + "bbox": [ + 212, + 146, + 542, + 205 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Different KP sets. We examine the contribution of different KP sets: joint KP (PP), joint pair KP (PRPP, PDP), limb KP (LAP, LOP), and body KP (GVP). A leave-one-out evaluation shows the elimination of joint KP and joint pair KP results in notable perfor", + "bbox": [ + 212, + 207, + 542, + 297 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "mance degradation, while the influence of the rest is relatively subtle.", + "bbox": [ + 212, + 297, + 714, + 311 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/adf314bf9dce14b369b197017adac4a819c781e0a7e9f59bfd62329f9f00b0f3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsAcc. %↑Diversity
Ours57.866.048
w/o KP39.945.526
w/o Joint KP50.035.685
w/o Joint Pair KP47.245.772
w/o Limb KP55.925.934
w/o Body KP56.845.871
", + "bbox": [ + 558, + 148, + 787, + 241 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 3: KPG ablation results.", + "bbox": [ + 562, + 250, + 777, + 263 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Discussion", + "text_level": 1, + "bbox": [ + 214, + 335, + 354, + 351 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. First, the current KP criteria guarantee objectivity but overlook kinematic information like amplitude and speed. Also, due to the granularity of the adopted skeleton, fine-grained information on fingers is not well-preserved. These would be promising for further exploration. Second, KPB could be extended to datasets with other modalities, like 2D pose and egocentric action datasets. Although these modalities provide incomplete 3D information, we could extract KP that is credibly accessible across modalities. Third, with the convenient KP-to-text conversion, auxiliary motion descriptions could be automatically generated via KP. Fourth, KPG could be extended by paraphrasing existing prompts and combining different Phrases. Finally, the proposed motion generation pipeline is devised for quick evaluation, while the motion-KP latent space could be incorporated with more advanced algorithms for better exploitation.", + "bbox": [ + 212, + 368, + 787, + 563 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Broader Impacts. The proposed KP and KPG provide a white-box instrument to diagnose text-to-motion algorithms, which could help develop reliable motion tools. Better motion synthesis could advance fields including animation production, sports training, human-robot interaction, and rehabilitation. However, it could also be used for illegal media content or false information.", + "bbox": [ + 212, + 587, + 787, + 662 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 686, + 359, + 702 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we proposed an intermediate representation to bridge human motion and action semantics as the Kinematic Phrase. KP achieved proper abstraction, interpretability, and generality by focusing on objective kinematic facts of human motion. A motion understanding system based on KP was proposed and proven effective in motion interpolation, modification, and generation. Moreover, Kinematic Prompt Generation is proposed as a white-box text-to-motion benchmark, bringing new insights to the literature. We believe that KP has great potential to advance motion understanding.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported in part by the National Natural Science Foundation of China under Grants No.62306175, No.62302296, National Key Research and Development Project of China (No.2022ZD0160102, No.2021ZD0110704), Shanghai Artificial Intelligence Laboratory, XPLORER PRIZE grants.", + "bbox": [ + 212, + 178, + 787, + 239 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 263, + 321, + 277 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aristidou, A., Cohen-Or, D., Hodgins, J.K., Chrysanthou, Y., Shamir, A.: Deep motifs and motion signatures. ACM Trans. Graph. 37(6), 187:1-187:13 (Nov 2018). https://doi.org/10.1145/3272127.3275038, http://doi.acm.org/10.1145/3272127.3275038", + "2. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: Teach: Temporal action composition for 3d humans. In: 2022 International Conference on 3D Vision (3DV). pp. 414-423. IEEE (2022)", + "3. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: SINC: Spatial composition of 3D human motions for simultaneous action generation. In: ICCV (2023)", + "4. Azadi, S., Shah, A., Hayes, T., Parikh, D., Gupta, S.: Make-an-animation: Large-scale text-conditional 3d human motion generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15039-15048 (October 2023)", + "5. Bartlett, R.: Introduction to Sports Biomechanics. Introduction to Sports Biomechanics, E & FN Spon (1997), https://books.google.com.tw/books?id=-6Db8mgxsqQC", + "6. Brégier, R.: Deep regression on manifolds: a 3d rotation case study. In: 2021 International Conference on 3D Vision (3DV). pp. 166-174. IEEE (2021)", + "7. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18000-18010 (2023)", + "8. Dabral, R., Mughal, M.H., Golyanik, V., Theobalt, C.: Mofusion: A framework for denoising-diffusion-based motion synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9760-9770 (June 2023)", + "9. Delmas, G., Weinzaepfel, P., Lucas, T., Moreno-Noguer, F., Rogez, G.: Posescript: 3d human poses from natural language. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI. pp. 346-362. Springer (2022)", + "0. Delmas, G., Weinzaepfel, P., Moreno-Noguer, F., Rogez, G.: Posefix: Correcting 3d human poses with natural language. arXiv preprint arXiv:2309.08480 (2023)", + "1. Fang, H.S., Li, J., Tang, H., Xu, C., Zhu, H., Xiu, Y., Li, Y.L., Lu, C.: Alpha-posed: Whole-body regional multi-person pose estimation and tracking in real-time. TPAMI (2023)", + "12. Gao, X., Yang, Y., Xie, Z., Du, S., Sun, Z., Wu, Y.: Guess: Gradually enriching synthesis for text-driven human motion generation. IEEE Transactions on Visualization and Computer Graphics (2024)", + "13. Guo, C., Mu, Y., Javed, M.G., Wang, S., Cheng, L.: Momask: Generative masked modeling of 3d human motions. arXiv preprint arXiv:2312.00063 (2023)" + ], + "bbox": [ + 218, + 295, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Kinematic Phrases", + "bbox": [ + 604, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5152-5161 (June 2022)", + "15. Guo, C., Zuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In: European Conference on Computer Vision. pp. 580-597. Springer (2022)", + "16. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2021-2029 (2020)", + "17. Guo, W., Du, Y., Shen, X., Lepetit, V., Alameda-Pineda, X., Moreno-Noguer, F.: Back to mlp: A simple baseline for human motion prediction. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4809-4819 (2023)", + "18. Hassan, M., Ceylan, D., Villegas, R., Saito, J., Yang, J., Zhou, Y., Black, M.: Stochastic scene-aware motion prediction. In: Proceedings of the International Conference on Computer Vision 2021 (Oct 2021)", + "19. Hernandez, A., Gall, J., Moreno-Noguer, F.: Human motion prediction via spatiotemporal inpainting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7134-7143 (2019)", + "20. Holden, D., Kanoun, O., Perepichka, M., Popa, T.: Learned motion matching. ACM Transactions on Graphics (TOG) 39(4), 53-1 (2020)", + "21. Hong, F., Zhang, M., Pan, L., Cai, Z., Yang, L., Liu, Z.: Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535 (2022)", + "22. Ji, Y., Xu, F., Yang, Y., Shen, F., Shen, H.T., Zheng, W.S.: A large-scale rgb-d database for arbitrary-view human action recognition. In: Proceedings of the 26th ACM international Conference on Multimedia. pp. 1510-1518 (2018)", + "23. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795 (2023)", + "24. Jin, P., Wu, Y., Fan, Y., Sun, Z., Yang, W., Yuan, L.: Act as you wish: Fine-grained control of motion diffusion model with hierarchical semantic graphs. Advances in Neural Information Processing Systems 36 (2024)", + "25. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2151-2162 (October 2023)", + "26. Kong, H., Gong, K., Lian, D., Mi, M.B., Wang, X.: Priority-centric human motion generation in discrete latent space. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14806-14816 (October 2023)", + "27. Koppula, H.S., Saxena, A.: Anticipating human activities for reactive robotic response. In: IROS. p. 2071. Tokyo (2013)", + "28. von Laban, R., Lange, R.: Laban's Principles of Dance and Movement Notation. Macdonald & Evans (1975), https://books.google.com.tw/books?id=VrOAAAAAMAAJ", + "29. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13401-13412 (2021)", + "30. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Learn to dance with aist++: Music conditioned 3d dance generation (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Li, Y.L., Wu, X., Liu, X., Wang, Z., Dou, Y., Ji, Y., Zhang, J., Li, Y., Lu, X., Tan, J., et al.: From isolated islands to pangea: Unifying semantic space for human action understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16582-16592 (2024)", + "32. Lin, J., Chang, J., Liu, L., Li, G., Lin, L., Tian, Q., Chen, C.W.: Being comes from not-being: Open-vocabulary text-to-motion generation with wordless training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23222-23231 (June 2023)", + "33. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM transactions on graphics (TOG) 34(6), 1-16 (2015)", + "34. Lou, Y., Zhu, L., Wang, Y., Wang, X., Yang, Y.: Diversemotion: Towards diverse human motion generation via discrete diffusion. arXiv preprint arXiv:2309.01372 (2023)", + "35. Lucas, T., Baradel, F., Weinzaepfel, P., Rogez, G.: Posegpt: quantization-based 3d human motion generation and forecasting. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI. pp. 417-435. Springer (2022)", + "36. Paden, B., Cap, M., Yong, S.Z., Yershov, D., Frazzoli, E.: A survey of motion planning and control techniques for self-driving urban vehicles. IEEE Transactions on intelligent vehicles 1(1), 33-55 (2016)", + "37. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3D hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 10975-10985 (2019)", + "38. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10985-10995 (2021)", + "39. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII. pp. 480-497. Springer (2022)", + "40. Petrovich, M., Black, M.J., Varol, G.: Tmr: Text-to-motion retrieval using contrastive 3d human motion synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9488-9497 (October 2023)", + "41. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data 4(4), 236-252 (2016)", + "42. Pons-Moll, G., Fleet, D.J., Rosenhahn, B.: Posebits for monocular human pose estimation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2337-2344 (2014)", + "43. Punnakkal, A.R., Chandrasekaran, A., Athanasiou, N., Quiros-Ramirez, A., Black, M.J.: Babel: bodies, action and behavior with english labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 722-731 (2021)", + "44. Qian, Y., Urbanek, J., Hauptmann, A.G., Won, J.: Breaking the limits of text-conditioned 3d motion synthesis with elaborative descriptions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2306-2316 (October 2023)", + "45. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Kinematic Phrases", + "bbox": [ + 604, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "46. Tang, Y., Liu, J., Liu, A., Yang, B., Dai, W., Rao, Y., Lu, J., Zhou, J., Li, X.: Flag3d: A 3d fitness activity dataset with language instruction. arXiv preprint arXiv:2212.04638 (2022)", + "47. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII. pp. 358-374. Springer (2022)", + "48. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022)", + "49. Van Welbergen, H., Van Basten, B.J., Egges, A., Ruttkay, Z.M., Overmars, M.H.: Real time animation of virtual humans: a trade-off between naturalness and control. In: Computer Graphics Forum. vol. 29, pp. 2530-2554. Wiley Online Library (2010)", + "50. Wang, Y., Leng, Z., Li, F.W.B., Wu, S.C., Liang, X.: Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 22035-22044 (October 2023)", + "51. Xiang, W., Li, C., Zhou, Y., Wang, B., Zhang, L.: Language supervised training for skeleton-based action recognition. arXiv preprint arXiv:2208.05318 (2022)", + "52. Xu, L., Song, Z., Wang, D., Su, J., Fang, Z., Ding, C., Gan, W., Yan, Y., Jin, X., Yang, X., Zeng, W., Wu, W.: Actformer: A gan-based transformer towards general action-conditioned 3d human motion generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2228-2238 (October 2023)", + "53. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 16010-16021 (October 2023)", + "54. Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052 (2023)", + "55. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022)", + "56. Zhang, M., Guo, X., Pan, L., Cai, Z., Hong, F., Li, H., Yang, L., Liu, Z.: Remodiffuse: Retrieval-augmented motion diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 364-373 (October 2023)", + "57. Zhang, M., Li, H., Cai, Z., Ren, J., Yang, L., Liu, Z.: Finemogen: Fine-grained spatio-temporal motion generation and editing. Advances in Neural Information Processing Systems 36 (2024)", + "58. Zhong, C., Hu, L., Zhang, Z., Xia, S.: Attt2m: Text-driven human motion generation with multi-perspective attention mechanism. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 509-519 (October 2023)", + "59. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5745-5753 (2019)", + "60. Zhou, Z., Wan, Y., Wang, B.: Avatargpt: All-in-one framework for motion understanding, planning, generation and beyond. arXiv preprint arXiv:2311.16468 (2023)" + ], + "bbox": [ + 215, + 147, + 785, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 17 + }, + { + "type": "ref_text", + "text": "61. Zhou, Z., Wang, B.: Ude: A unified driving engine for human motion generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5632-5641 (June 2023)", + "bbox": [ + 217, + 147, + 784, + 189 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Kinematic Phrases", + "bbox": [ + 604, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Bridging the Gap Between Human Motion and Action Semantics via Kinematics Phrases/ceda5829-f6a0-4ea2-94dd-2ddc6da8e6ed_model.json b/2024/Bridging the Gap Between Human Motion and Action Semantics via Kinematics Phrases/ceda5829-f6a0-4ea2-94dd-2ddc6da8e6ed_model.json new file mode 100644 index 0000000000000000000000000000000000000000..eee9a04adc63ebc2c8c1363c391ffb0d67761d06 --- /dev/null +++ b/2024/Bridging the Gap Between Human Motion and Action Semantics via Kinematics Phrases/ceda5829-f6a0-4ea2-94dd-2ddc6da8e6ed_model.json @@ -0,0 +1,2427 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.235, + 0.141, + 0.773, + 0.187 + ], + "angle": 0, + "content": "Bridging the Gap between Human Motion and Action Semantics via Kinematic Phrases" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.212, + 0.782, + 0.245 + ], + "angle": 0, + "content": "Xinpeng Liu\\(^{1}\\), Yong-Lu Li\\(^{1\\star}\\), Ailing Zeng\\(^{2}\\), Zizheng Zhou\\(^{1}\\), Yang You\\(^{3}\\), and Cewu Lu\\(^{1\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.39, + 0.255, + 0.613, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\) Shanghai Jiao Tong University" + }, + { + "type": "text", + "bbox": [ + 0.467, + 0.269, + 0.536, + 0.282 + ], + "angle": 0, + "content": "2 Tencent" + }, + { + "type": "text", + "bbox": [ + 0.428, + 0.283, + 0.576, + 0.297 + ], + "angle": 0, + "content": "3 Stanford University" + }, + { + "type": "text", + "bbox": [ + 0.343, + 0.298, + 0.659, + 0.311 + ], + "angle": 0, + "content": "{xinpengliu0907,ailingzengzzz} \\(@\\) gmail.com," + }, + { + "type": "text", + "bbox": [ + 0.271, + 0.312, + 0.734, + 0.325 + ], + "angle": 0, + "content": "{yonglu_li, zhou_zz, lucewu}@sjtu.edu.cn, yangyou@stanford.edu" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.363, + 0.744, + 0.628 + ], + "angle": 0, + "content": "Abstract. Motion understanding aims to establish a reliable mapping between motion and action semantics, while it is a challenging many-to-many problem. An abstract action semantic (i.e., walk forwards) could be conveyed by perceptually diverse motions (walking with arms up or swinging). In contrast, a motion could carry different semantics w.r.t. its context and intention. This makes an elegant mapping between them difficult. Previous attempts adopted direct-mapping paradigms with limited reliability. Also, current automatic metrics fail to provide reliable assessments of the consistency between motions and action semantics. We identify the source of these problems as the significant gap between the two modalities. To alleviate this gap, we propose Kinematic Phrases (KP) that take the objective kinematic facts of human motion with proper abstraction, interpretability, and generality. Based on KP, we can unify a motion knowledge base and build a motion understanding system. Meanwhile, KP can be automatically converted from motions to text descriptions with no subjective bias, inspiring Kinematic Prompt Generation (KPG) as a novel white-box motion generation benchmark. In extensive experiments, our approach shows superiority over other methods. Our project is available at https://foruck.github.io/KP/." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.64, + 0.744, + 0.667 + ], + "angle": 0, + "content": "Keywords: Motion Representation \\(\\cdot\\) Motion Understanding \\(\\cdot\\) Text-to-Motion Benchmark" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.694, + 0.378, + 0.71 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.726, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Human motion understanding has a wide range of applications, including activity understanding [31], autonomous driving [36], robotics [27], and automatic animation [49], making it increasingly attractive. The core is to establish a mapping between the motion space and the action semantics space. The motion space indicates a space of sequential 3D human representations, e.g., 3D pose or SMPL [33]/SMPL-X [37] parameter sequence, while the action semantic space" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.393, + 0.84 + ], + "angle": 0, + "content": "* Corresponding authors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.352, + 0.788, + 0.422 + ], + "angle": 0, + "content": "Fig. 1: The huge gap between motion and action semantics results in the many-to-many problem. We propose Kinematic Phrases (KP) as an intermediate to bridge the gap. KPs objectively capture human kinematic cues. It properly abstracts diverse motions with interpretability. As shown, the Phrases in the yellow box could capture key patterns of walk for diverse motions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.786, + 0.496 + ], + "angle": 0, + "content": "can be represented as action categories or sentences described by natural language." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.788, + 0.713 + ], + "angle": 0, + "content": "Recently, a growing focus has been on generative mapping from semantics to motion, including action category-based generation [38] and text-based generation [7, 14, 35, 39, 48, 54, 55]. Most of them typically build a mapping that links motion and semantics either directly or via motion latent, with understated concerns for intermediate motion-semantic structures. However, these models suffer from inferior reliability. They cannot guarantee they generated correct samples without human filtering. Additionally, the existing evaluation of motion generation is problematic. Widely adopted FID and R-Precision rely on the latent space from a black-box pre-trained model, which might fail to out-of-distribution (OOD) and over-fitting cases. There is a long-standing need for an evaluation method that can cheaply and reliably assess whether a generated motion is consistent with particular action semantics. We identify the essence of these as the significant gap between raw human motion and action semantics, which makes direct mapping hard to learn." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As in Fig. 1, an action semantics can correspond to diverse motions. For instance, a person could walk in countless ways with diverse motions, either with arms up or swinging, while action semantics tend to abstract these away from a walking motion. Additionally, they are robust against small perturbations, while motion is more specific and complex, with representations changing vastly when perturbed or mis-captured. Moreover, a motion sequence could have diverse semantics w.r.t. contexts. Modeling this many-to-many mapping between motion and semantics is challenging." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.605, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Kinematic Phrases" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "To bridge this gap between motion and action semantics, we propose Kinematic Phrases (KP), an interpretable intermediate representation. KP focuses on the objective kinematic facts, which are usually omitted by general action semantics, like left-hand moving forwards then backward. KP is designed as qualitative categorical representations of these facts. For objectivity and actuality, KP captures sign changes with minimal pre-defined standards. Inspired by previous studies on kinematic human motion representation [5,28], KP is proposed" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.254, + 0.488, + 0.418 + ], + "angle": 0, + "content": "as six types shown in Fig. 1, covering joint positions, joint pair positions and distances, limb angles and directions, and global velocity. Note that, although KP can be described by natural language, a major difference is that KP is strictly dedicated to objective kinematic facts instead of coarse actions such as surrender or fine-grained actions like raise both hands." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.423, + 0.49, + 0.544 + ], + "angle": 0, + "content": "We highlight three advantages of KP. First, KP offers proper abstraction, which disentangles motion perturbations and semantics changes, easing the learning process. Even though the motion differs significantly, KP manages to capture walk patterns easily. Second, KP is inter" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.787, + 0.605 + ], + "angle": 0, + "content": "pretable, as it can be viewed as instructions on executing the action, making it easily understandable to humans. Finally, KP is general, as it can be automatically extracted from different modalities of human motion, including skeleton and SMPL parameters. The conversion from KP to text is also effortless." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.609, + 0.788, + 0.806 + ], + "angle": 0, + "content": "With KP as an intermediate representation, a unified large-scale motion knowledge base is constructed. Then, to fully exploit KP and the knowledge base, we build a motion understanding system with KP mediation. In detail, we learn a motion-KP joint latent space for multiple motion understanding applications, including motion interpolation, modification, and generation. Moreover, leveraging the interpretability of KP, we propose a benchmark called Kinematic Prompts Generation (KPG), which generates motion from text prompts converted from KPs. Thanks to the consistency and convenience of the KP-to-text conversion, KPG enables efficient white-box motion generation evaluation. Given the bottle-necked evaluation of the current motion generation benchmarks [14,41] as shown in Fig. 2, KPG could be a promising complementary benchmark. Results in Tab. 2 show that KPG brings novel insights into the current text-to-motion progress." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Our contributions are: (1) We propose KP as an intermediate representation to bridge the gap between motion and action semantics. (2) We build a novel" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.284, + 0.78, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.437, + 0.788, + 0.507 + ], + "angle": 0, + "content": "Fig. 2: HumanML3D [14] metric values are approaching even surpassing GT-level while being increasingly indistinguishable. The higher-than-GT R-Precision could be insufficient as a semantic consistency indicator." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.128 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "motion understanding system using KP and the aggregated large-scale knowledge base. (3) We propose KPG as a benchmark for efficient white-box motion generation evaluation. Promising results are achieved on motion interpolation and generation tasks. Moreover, extensive user studies are conducted, verifying the efficacy of our methods, and the consistency between KPG evaluation and human perception." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.26, + 0.398, + 0.276 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.29, + 0.79, + 0.488 + ], + "angle": 0, + "content": "Motion Representation An intuitive motion representation is a sequence of static pose representations, like joint locations and limb rotations. Efforts are paid to address the discontinuity of rotation for deep-learning methods [6,59]. Recent works on parametric body models [33,37] enable a more realistic body representation. Meanwhile, [42] proposed Posebits, representing pose with boolean geometric part relationships. [9, 10] translates Posebits into text descriptions. These abstract representations are flexible and insensitive to little perturbations, but their static nature ignores motion dynamics. [46] acquire similar fine-grained descriptions from human annotation, while [3, 51] adopted large-scale language models. However, few recognize their potential in bridging the low-level motion and the high-level action semantics. Phase functions [20], Labanotations [28], and learned Motion Words [1] were also explored, though limited to specific actions like locomotion and dancing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.79, + 0.691 + ], + "angle": 0, + "content": "Motion Generation can be conditioned by its prefix/suffix [2, 17, 19], action categories [16, 38, 52], or audio [29, 30]. Text-based motion generation has developed rapidly with the proposal of text-motion datasets [14, 43]. Early works [14, 39, 44] used VAEs, while some [21, 32, 47] extended the CLIP [45] space to motion. Diffusion models [8, 48, 50, 55] were also adopted. MAA [4] adopted a U-Net structure. Some [40, 56] explored retrieval-based methods. [25] aimed at controllable generation, while [53] introduced physical constraints. However, most approaches still suffer from the gap between motion and action semantics. Multiple methods [7, 15, 26, 35, 54, 58, 61] adopted (VQ-)VAE-compressed motion representation as mediation, while in the current data-limited situation, we identify that this single-modality compression might be sub-optimal. Instead, KP could alleviate this by introducing explicit semantic-geometric correlation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.71, + 0.48, + 0.727 + ], + "angle": 0, + "content": "3 Kinematic Phrase Base" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.741, + 0.421, + 0.756 + ], + "angle": 0, + "content": "3.1 Kinematic Phrases" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Kinematic Phrases abstract motion into objective kinematic facts like left-hand moves up qualitatively. We take inspiration from previous kinematic motion representations [28] and qualitative static pose representations [9,11,42], proposing six types of KP to comprehensively represent motion from different kinematic hierarchies: For joint movements, there are 34 Position Phrases (PPs). For joint" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.605, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Kinematic Phrases" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.145, + 0.774, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.417, + 0.788, + 0.46 + ], + "angle": 0, + "content": "Fig.3: Six types of KP from four kinematic hierarchies are extracted from a motion sequence. A scalar indicator \\( s_i \\) is calculated per Phrase per frame. Its sign categorizes the corresponding Phrase." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.496, + 0.788, + 0.679 + ], + "angle": 0, + "content": "pair movements, there are 242 Pairwise Relative Position Phrases (PRPPs) and 81 Pairwise Distance Phrases (PDPs). For limb movements, there are 8 Limb Angle Phrases (LAPs) and 24 Limb Orientation Phrases (LOPs). For whole-body movements, there are 3 Global Velocity Phrases (GVPs). KP extraction is based on a skeleton sequence \\( X = \\{x_{i} | x_{i} \\in \\mathcal{R}^{n_{k} \\times 3}\\}_{i=1}^{t} \\), where \\( n_{k} \\) is the number of joints (\\( n_{k} = 17 \\) here), \\( x_{i} \\) is the joint coordinates at \\( i \\)-th frame, and \\( t \\) is the sequence length. Note that \\( x_{i}^{0} \\) indicates the pelvis/root joint. For each Phrase, a scalar indicator sequence is calculated from the skeleton sequence. Phrases are extracted as per-frame categorical representations w.r.t. indicator signs. Unlike previous efforts [9, 42], we limit the KP criteria as the indicator signs to minimize human-defined biases (e.g., numerical criteria on the closeness of two joints) for objectivity and actuality. Fig. 3 illustrated the procedure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.681, + 0.788, + 0.773 + ], + "angle": 0, + "content": "Reference Vectors are first constructed, indicating right, upward, and forward directions from a human cognitive view. We aim at the egocentric reference frames that human tends to use when performing actions. The negative direction of gravity is adopted as upward vector \\( r^u \\), the vector from left hip to right hip is adopted as right vector \\( r^r \\), and the forward vector is calculated as \\( r^f = r^u \\times r^r \\). These vectors of each frame are denoted as \\( R^{\\cdot} = \\{r_{i}^{\\cdot}\\}_{i=1}^{t} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.775, + 0.787, + 0.806 + ], + "angle": 0, + "content": "Position Phrase (PP) focuses on the movement direction of joint \\( x^{j} \\) w.r.t. reference vector \\( R \\). The indicator for PP at \\( i \\)-th frame is calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.336, + 0.822, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\ns _ {i} ^ {(j, \\cdot)} = \\langle (x _ {i} ^ {j} - x _ {i} ^ {0}), r _ {i} ^ {\\cdot} \\rangle - \\langle (x _ {i - 1} ^ {j} - x _ {i - 1} ^ {0}), r _ {i - 1} ^ {\\cdot} \\rangle . \\tag {1}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.128 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.223 + ], + "angle": 0, + "content": "The sign of \\( s_i^{(j,\\cdot)} \\) categorizes PP into moving along/against \\( R' \\), or relatively static along \\( R' \\) for indicators with small amplitudes. 12 joints (except the pelvis, eyes, and hips) and 3 reference vectors resulted in originally 36 PPs. Specifically, shoulders are excluded for the leftward vector, and hips are excluded for the upward vector, resulting in 36-2=34 PPs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.784, + 0.266 + ], + "angle": 0, + "content": "Pairwise Relative Position Phrase (PRPP) describes the relative position of a pair of joints \\((x^{j}, x^{k})\\) w.r.t. reference vector \\(R\\). PRPP indicator at \\(i\\)-th frame is" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.266, + 0.785, + 0.285 + ], + "angle": 0, + "content": "\\[\ns _ {i} ^ {(j, k, \\cdot)} = \\langle \\left(x _ {i} ^ {j} - x _ {i} ^ {k}\\right), r _ {i} ^ {\\cdot} \\rangle . \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.289, + 0.785, + 0.395 + ], + "angle": 0, + "content": "For (L-Hand, R-Hand) and forward vector \\(R^f\\), PRPP could be L-Hand behind /in front of R-Hand according to the sign of \\(s_i^{(j,k,\\cdot)}\\). 136 joint pairs and 3 reference vectors composed 408 Phrases originally. 24 joint pairs linked by limbs are filtered out. 30 Eye-related pairs are filtered out except (left eye, right eye) due to the others are covered by head-related pairs. 4 triplets that the relationship barely changes along the reference vector (e.g., right knee, left hip, and leftward vector) are filtered out. These result in \\(408 - 24^{*}3 - 30^{*}3 - 4 = 242\\) PRPPs." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.397, + 0.784, + 0.427 + ], + "angle": 0, + "content": "Pairwise Distance Phrase (PDP) describes how the L2 distance between a pair of joints \\((x^{j}, x^{k})\\) changes. The indicator for PDP is calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.437, + 0.785, + 0.457 + ], + "angle": 0, + "content": "\\[\ns _ {i} ^ {(j, k)} = \\left\\| x _ {i} ^ {j} - x _ {i} ^ {k} \\right\\| _ {2} - \\left\\| x _ {i - 1} ^ {j} - x _ {i - 1} ^ {k} \\right\\| _ {2}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.467, + 0.784, + 0.514 + ], + "angle": 0, + "content": "The sign of \\( s_i^{(j,k)} \\) categorizes PDP into moving closer/away, or relatively static. Among the original 105 joint pairs, 24 pairs that are in the skeleton topology are excluded, such as the hand and elbow, resulting in 81 PDPs." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.514, + 0.784, + 0.545 + ], + "angle": 0, + "content": "Limb Angle Phrase (LAP) targets at the change of bend angle between two connected limbs \\((x^{j}, x^{k})\\) and \\((x^{j}, x^{l})\\). The LAP indicator is calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.233, + 0.554, + 0.785, + 0.574 + ], + "angle": 0, + "content": "\\[\ns _ {i} ^ {(j, k, l)} = \\operatorname {a r c c o s} \\left(\\left\\langle x _ {i} ^ {k} - x _ {i} ^ {j}, x _ {i} ^ {l} - x _ {i} ^ {j} \\right\\rangle\\right) - \\operatorname {a r c c o s} \\left(\\left\\langle x _ {i - 1} ^ {k} - x _ {i - 1} ^ {j}, x _ {i - 1} ^ {l} - x _ {i - 1} ^ {j} \\right\\rangle\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.583, + 0.75, + 0.598 + ], + "angle": 0, + "content": "8 LAPs describe the limb chain \\((x^{j},x^{k}) - (x^{j},x^{l})\\) as bending or unbending." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.599, + 0.784, + 0.629 + ], + "angle": 0, + "content": "Limb Orientation Phrase (LOP) describes the orientation of the limb \\((x^{j}, x^{k})\\) w.r.t. \\(R^{\\prime}\\), with \\(x^{k}\\) being the distal limb. The LOP indicator is" + }, + { + "type": "equation", + "bbox": [ + 0.424, + 0.639, + 0.785, + 0.658 + ], + "angle": 0, + "content": "\\[\ns _ {i} ^ {(j, k, \\cdot)} = \\left\\langle x _ {i} ^ {k} - x _ {i} ^ {j}, r _ {i} ^ {\\prime} \\right\\rangle . \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.667, + 0.785, + 0.745 + ], + "angle": 0, + "content": "The sign of \\( s_i^{(j,k,\\cdot)} \\) categorizes the LOP \\( (x^j, x^k) \\) pointing along/against \\( R \\), or a placeholder category for those with little magnitude. 4 arms, 4 leg limbs, head, collarbones, hips, torsos, and upper body, paired with the 3 reference vectors, result in 45 LOPs in total. 21 pairs that barely change along the reference vector (e.g., left-right hip, leftward vector) are filtered out, resulting in 45-21=24 LOPs." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.745, + 0.784, + 0.774 + ], + "angle": 0, + "content": "Global Velocity Phrase (GVP) describes the direction of global velocity with respect to \\( R \\). The indicator is calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.432, + 0.785, + 0.785, + 0.801 + ], + "angle": 0, + "content": "\\[\ns _ {i} = \\left\\langle x _ {i + 1} ^ {0} - x _ {i} ^ {0}, r _ {i} ^ {\\prime} \\right\\rangle . \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.841 + ], + "angle": 0, + "content": "The three categories are moving along/against \\( R \\), or static along \\( R \\) according to the sign of \\( s_i \\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.605, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Kinematic Phrases" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "These result in 392 Phrases in total, covering motion diversity and distribution from various levels. While we clarify that these Phrases do not rule out the possibility of other possible useful potentials." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.215, + 0.569, + 0.232 + ], + "angle": 0, + "content": "3.2 Constructing Kinematic Phrase Base" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.241, + 0.788, + 0.422 + ], + "angle": 0, + "content": "KP enables us to unify motion data with different formats to construct a large knowledge base containing motion, text, and KP thanks to its generality. Motion sequences of different representations are collected, including 3D skeleton sequences and SMPL [33] /SMPL-X [37] parameter sequences. The sequences are first re-sampled to \\(30\\mathrm{Hz}\\) and rotated so that the negative direction of the z-axis is the gravity direction. Then, the sequences are converted into 3D skeleton sequences for KP extraction as in Sec. 3.1. Text annotations attached to the sequences are directly saved. For sequences with action category annotation, the category name is saved. For those with neither text nor action category, the text information is set from its attached additional information, like objects for SAMP [18]. Finally, we collect 140K motion sequences from 11 datasets. More details are included in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.446, + 0.441, + 0.464 + ], + "angle": 0, + "content": "4 Applications of KP" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.478, + 0.528, + 0.494 + ], + "angle": 0, + "content": "4.1 Motion Understanding with KP" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.785, + 0.566 + ], + "angle": 0, + "content": "An intuitive application of KP is to leverage it for better motion understanding. To achieve this, we first learn a motion-KP joint space, which exploits the clarity and interpretability of KP. Then, we adopt the learned latent space for multiple motion understanding tasks. An overview is illustrated in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.587, + 0.788, + 0.664 + ], + "angle": 0, + "content": "Preliminaries. Motion is represented as a pose sequence with \\( n \\) frames as \\( M = \\{m_i\\}_{i=1}^n \\). In detail, SMPL [33] pose parameters are transformed to the 6D representation [59], then concatenated with the velocity of the root, resulting in a 147-dim representation per frame. KP is represented by indicator signs as \\( C = \\{c_i\\}_{i=1}^n, c_i \\in \\{1,0,-1\\}^{392} \\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.686, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Joint Space Learning. The joint space is learned with Motion VAE \\(\\{\\mathcal{E}_m,\\mathcal{D}_m\\}\\) and KP VAE \\(\\{\\mathcal{E}_p,\\mathcal{D}_p\\}\\). All VAEs follow the transformer-based architecture from [38]. Motion encoder \\(\\mathcal{E}_m\\) takes motion \\(M\\) and two distribution tokens \\(m_{\\mu},m_{\\sigma}\\) as input, and the outputs corresponding to the distribution tokens are taken as the \\(\\mu_{m}\\) and \\(\\sigma_{m}\\) of the Gaussian distribution. Then, motion decoder \\(\\mathcal{D}_m\\) takes \\(z_{m}\\sim \\mathcal{G}(\\mu_{m},\\sigma_{m})\\) as \\(K,V\\), and a sinusoidal positional encoding of the expected duration as \\(Q\\). The output is fed into a linear layer to reconstruct motion \\(\\hat{M}\\). KP VAE \\(\\{\\mathcal{E}_p,\\mathcal{D}_p\\}\\) operates similarly, with the sign of \\(\\mathcal{D}_p\\) output as the predicted KP \\(\\hat{C}\\). Notice that the \\(\\mathcal{D}_m,\\mathcal{D}_p\\) could take arbitrary combinations of \\(z_{m},z_{p}\\) as input, outputting \\(\\hat{M},\\hat{C}\\). To learn the joint space, besides the conventional loss" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.145, + 0.761, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.36, + 0.785, + 0.403 + ], + "angle": 0, + "content": "Fig. 4: We train a motion-KP joint latent space to exploit the clarity and interpretability of KP. The space is then applied to multiple tasks, including motion interpolation, modification, and generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.788, + 0.542 + ], + "angle": 0, + "content": "of reconstruction and KL divergence, we also align the two VAEs similarly to [39]. In this way, we guide the learned motion latent space with the structure of KP. Meanwhile, we randomly corrupt samples during training by setting a small portion of KP as 0. The idea is that the overall representation should be robust with a small portion of KP unknown. Also, the missing Phrases should be recovered from existing Phrases. This mines the correlation among Phrases and increases the robustness of the space. More details are in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.753 + ], + "angle": 0, + "content": "Motion Understanding Tasks with Joint Space. With the joint space, we can perform both low-level and high-level motion understanding tasks. We introduce three applications to show the capability of KP, as shown in Fig. 4. For motion interpolation given a corrupted motion sequence \\(\\tilde{M}\\), we extract its corresponding KP sequence \\(\\tilde{C}\\), then feed them to encoders \\(\\mathcal{E}_m, \\mathcal{E}_p\\) and decoder \\(\\mathcal{D}_p\\), resulting in the estimated KP sequence \\(\\hat{C}\\). \\(\\hat{C}\\) and \\(\\tilde{M}\\) are fed into \\(\\mathcal{E}_m, \\mathcal{E}_p\\) and \\(\\mathcal{D}_m\\), resulting in interpolated \\(\\hat{M}\\). For motion modification on motion \\(M\\), modifications could be made on the extracted KP sequence \\(C\\) resulting in \\(\\tilde{C}\\). Motion frames to modify are then masked, getting \\(\\tilde{M}\\). \\(\\tilde{M}, \\tilde{C}\\) are fed into \\(\\mathcal{E}_m, \\mathcal{E}_p\\) and \\(\\mathcal{D}_m\\), getting the interpolated \\(\\hat{M}\\). For motion generation, we simply replace the motion latent space in MLD [7] with the latent space of our Motion VAE. Experiments show that KP manages to boost MLD with a considerable margin." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.78, + 0.749, + 0.796 + ], + "angle": 0, + "content": "4.2 Motion Benchmarking with Kinematic Prompt Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "A more interesting and important application is benchmarking motion generation with the interpretability and objectivity of KP." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.606, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Kinematic Phrases" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.144, + 0.756, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.35, + 0.788, + 0.393 + ], + "angle": 0, + "content": "Fig. 5: Kinematic Prompt Generation. 7,796 prompts are converted from KP with templates, corresponding to certain KP patterns. We calculate the generation accuracy by checking the appearance of the patterns in KP extracted from the generated motion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.789, + 0.73 + ], + "angle": 0, + "content": "Before that, we first analyze current benchmarks. A crucial aspect of motion generation evaluation is motion-semantic consistency. The gold standard is user study. However, it is expensive and inefficient to scale. Early metrics like MPJPE (Mean Per Joint Position Error) and MAE (Mean Angle Error) mechanically calculate the error between the generated and GT samples. These metrics fail to reveal the real ability of generative models: What if the models memorize GT samples? Or what if the samples are diverse from GT but also true? FID (Frechet Inception Distance) is adopted to mitigate this issue. However, it provides a macro view of the quality of all generated samples without guarantees for individual samples. [14] proposed R-Precision, using a pre-trained text-motion matching model to examine whether the generated samples carry true semantics. They both rely on the latent space from a black-box pre-trained model, which is not credible. Besides, models might learn short paths to over-fit the pre-trained model. Meanwhile, since automatic mapping from motion to semantics is still an unsettled problem, adopting it to evaluate motion generation is not a decent choice. Moreover, most current motion generation evaluations are performed on datasets [14,22,41] with considerable complex everyday actions, further increasing the difficulty. As shown in Fig. 2 and Tab. 1, the current evaluation metrics are becoming increasingly indistinguishable. Also, the higher-than-GT R-Precision could be insufficient as an indicator of semantic consistency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.732, + 0.789, + 0.809 + ], + "angle": 0, + "content": "To this end, we propose a novel benchmark: Kinematic Prompts Generation (KPG). Instead of previous benchmarks focusing on everyday activities or sports, we step back in the complexity of action semantics. Based on KP, KPG focuses on evaluating whether the models could generate motion sequences consistent with specific kinematic facts given text prompts. An overview is given in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In detail, we convert KP into 7,796 text prompts with templates as in Fig. 5. The prompts could be categorized into 4 groups. First, each atomic prompt" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.512 + ], + "angle": 0, + "content": "involves only 1 KP, such as \"left hand moves upwards\", which evaluates the basic semantic consistency of text-to-motion models. Notice that PRPP and LOP are excluded since they represent static states, resulting in 252 prompts. Second, each repetitive prompt involves a KP being repeated two or three times., like \"left arm bends twice\", evaluating the basic counting ability. PRPP, LOP, and GVP are excluded, bringing 492 prompts. Third, each sequential prompt involves two independent KP to be sequentially executed, like \"left arm bends, then right arm bends\". 3,912 sequential prompts are randomly sampled. Finally, each simultaneous prompt involves two independent KPs to be simultaneously executed, like \"left arm bends, and simultaneously, right arm bends.\" 3,120 prompts are randomly sampled. The sequential and simultaneous prompts examine whether the model can understand the composition of simple actions. Given prompt \\( T_{i} \\in T \\) from Phrase \\( c_{i} \\) (or Phrases \\( (c_{i}^{1}, c_{i}^{2}) \\)), the model generates motion \\( \\hat{M}_{i} \\), along with extracted KP \\( \\hat{C}_{i} \\). Then, whether \\( T_{i} \\) is hit could be inferred via designed protocols by identifying certain KP patterns in \\( \\hat{C}_{i} \\). In detail, the atomic prompt is identified as hit if \\( c_{i} \\) appears for more than 5 consecutive frames in \\( \\hat{C}_{i} \\). The repetitive prompt is identified as hit if and only if \\( c_{i} \\) appears for at least 5 consecutive frames in \\( \\hat{C}_{i} \\) for corresponding times. The sequential prompt is identified as hit if \\( c_{i}^{2} \\) appears for at least 5 consecutive frames after (maybe not immediately) \\( c_{i}^{1} \\) does the same. The simultaneous prompt is identified as hit if \\( c_{i}^{1} \\) and \\( c_{i}^{2} \\) appear simultaneously for at least 5 consecutive frames. The calculation involves no black-box model thanks to KP, presenting a fully white-box evaluation. Also, with the effortless motion-to-KP conversion, the computation could be conducted automatically. More details are in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.534, + 0.367, + 0.551 + ], + "angle": 0, + "content": "5 Experiment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.788, + 0.673 + ], + "angle": 0, + "content": "Implementation Details. HumanML3D [14] test split is held out for evaluation, with the rest of KPB for training. The batch size is set as 288, and an AdamW optimizer with a learning rate of 1e-4 is adopted. We randomly corrupt less than \\(20\\%\\) of the Phrases for a sample. The Motion-KP joint space is trained for 6,000 epochs. While the text-to-motion latent diffusion model is trained for 3,000 epochs, with the joint space frozen. All experiments are conducted in 4 NVIDIA RTX 3090 GPUs. More details are in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.693, + 0.438, + 0.709 + ], + "angle": 0, + "content": "5.1 Motion Interpolation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.718, + 0.787, + 0.749 + ], + "angle": 0, + "content": "Following [23], \\(50\\%\\) frames are randomly masked for interpolation. In Tab. 1, our method provides better FID. With KPB, Diversity is increased." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.77, + 0.421, + 0.785 + ], + "angle": 0, + "content": "5.2 Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.843 + ], + "angle": 0, + "content": "Results on conventional text to motion are shown in Tab. 1. We adopt the HumanML3D test set [14] for conventional text-to-motion evaluation. The evaluation model from [14] is adopted to calculate R-Precision, FID, Diversity," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.605, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Kinematic Phrases" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.219, + 0.144, + 0.561, + 0.344 + ], + "angle": 0, + "content": "
MethodsInterpolationGeneration
FID↓Diversity→R-P@1↑FID↓Diversity→Multimodality
GT0.0029.5030.5110.0029.503-
TEMOS [39]--0.4243.7348.9730.368
T2M [14]--0.4551.0679.1882.090
MotionDiffuse [55]--0.4910.6309.4100.730
MDM [48]2.6988.4200.3200.5449.5592.799
TM2T [15]--0.4241.5018.5892.424
MLD [7]--0.4810.4739.7242.413
T2M-GPT [54]--0.4920.1419.7221.831
Fg-T2M [50]--0.4920.2439.2781.614
M2DM [26]--0.4970.3529.9263.587
AttT2M [58]--0.4990.1129.7002.452
ReMoDiffuse [56]--0.5100.1039.0181.795
MotionGPT [23]0.2149.5600.4920.2329.5282.008
FineMoGen [57]--0.5040.1519.2632.696
GraphMotion [24]--0.5040.1169.6922.766
AvatarGPT [60]--0.5100.1689.624-
GUESS [12]--0.5030.1099.8262.430
MDD [34]--0.5150.0729.6831.869
MoMask [13]--0.5210.045-1.241
Ours0.1929.6230.4960.2759.9752.218
Ours w/ KPB0.21410.1420.4710.64110.2622.891
" + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.146, + 0.791, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.342, + 0.793, + 0.412 + ], + "angle": 0, + "content": "Table 1: Results on HumanML3D. R-P@1 is short Fig.6: R-Precision could contrafor R-Precision@1. Most methods show comparadict to user reviews, indicatingble performance. Also, the higher-than-GT phethat R-Precision could produce nomenon indicates that R-Precision might be insuf-unfaithful evaluations of semantic ficient as a semantic consistency indicator. consistency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.788, + 0.637 + ], + "angle": 0, + "content": "and Multimodality. We run the evaluation 20 times and report the average value. Our method is competitive without KPB. However, there exist some unsatisfying observations. First, KPB brings a counter-intuitive performance drop. Second, the metric values of most techniques are rather close (e.g., difference \\(< 0.03\\) for R-P@1), and no substantial advantage is perceived for one method over others. Also, as discovered by MAA [4], despite its relatively poor R-Precision and FID, MAA manages to achieve better user reviews, indicating the potential loss of efficacy of these conventional metrics. This is further revealed in Fig. 6, where a contradiction is observed between human judgment and R-precision. Moreover, some methods show higher-than-GT R-P@1, which indicates that R-Precision might be insufficient as a semantic consistency indicator. Given these, we further switch to KPG and pursue further insights." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Results on KPG are shown in Tab. 2. Besides accuracy, diversity and R-precision are reported. We run the evaluation 20 times and report the average metric value. The white-box designation of KPG evaluation makes it possible to evaluate the different aspects of the models. Meanwhile, KPG is considered an easier task than conventional text-based motion generation since it is targeted at much less complex action semantics. Results show that previous methods are not performing as well as expected. Also, R-Precision fails to model the performance on KPG with trivially low values. Most methods perform well for atomic prompts, but when it comes to compositions, the accuracy drops. Early efforts like HMDM [48] and MotionDiffuse [55] are surprisingly competitive, even surpassing several more recent works. T2M-GPT [54] outperforms other methods by a considerable margin, but it fails for repetitive prompts. An unexpected" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "table", + "bbox": [ + 0.277, + 0.144, + 0.73, + 0.252 + ], + "angle": 0, + "content": "
MethodsAccuracyDiversityR-P@1
AtomicRepetitiveSequentialSimultaneousOverall
HMDM [48]93.2523.9856.1130.9045.165.9230.034
MotionDiffuse [55]93.2122.1551.5427.7641.495.2570.051
T2M-GPT [54]97.221.0270.2442.4755.596.1530.042
MLD [7]92.0622.7648.2628.0139.945.5260.061
ReMoDiffuse [56]53.173.0532.1631.0630.565.8090.048
MotionGPT [23]92.0616.8755.9831.7644.965.7760.039
MoMask [13]96.4328.8659.1032.6647.795.9650.060
Ours98.8026.0271.3243.0857.866.0480.053
" + }, + { + "type": "table_caption", + "bbox": [ + 0.227, + 0.253, + 0.776, + 0.267 + ], + "angle": 0, + "content": "Table 2: Results on KPG demonstrate the various strengths of different methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.787, + 0.404 + ], + "angle": 0, + "content": "phenomenon is the failure of ReMoDiffuse [56], which might be constrained by its retrieval-based methodology since most prompts in KPG are not available in the retrieval database. Our method, which replaces the latent space of MLD [7], delivers substantial improvements, but the accuracy remains below \\(60\\%\\), which is far from satisfying. A considerable gap exists between existing efforts and ideal motion generation models. More analyses are in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.688 + ], + "angle": 0, + "content": "User Study. To reach a convincing performance comparison, we conduct user studies on both HumanML3D and KPG. Our user study is different from previous efforts in two aspects. First, instead of testing a small set of text prompts (less than 50 in previous works [7,48]), we randomly select 600 sentences from the HumanML3D test set. By scaling up, the result is convincing in reflecting the ability to generate motion for diverse text inputs. Second, without asking the volunteers to give a general rating for each sample or to choose between different samples, we ask them two questions: 1) Do the motion and the text match? and 2) Is the motion natural? For Q1, three choices are given as \"No, Partially, Yes\". For Q2, two choices are given as \"Yes, No\". In this way, we explicitly evaluate two aspects of text-to-motion tasks: semantic consistency and naturalness, corresponding to R-Precision and FID. For each prompt, we generate one sample considering the annotation cost. We claim that the models should generate natural text-matching motion most of the time so that the one-sample setting would not hurt the fidelity of our user study. 36 volunteers are invited, each reviewing 200 sequences. Thus each sequence receives 3 user reviews. Also, we compute R-precision@1 of the generated sequences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Besides our model, we select MDM [48] as a naive baseline, T2M-GPT [54] as the SOTA model given the results in Tab. 1 and Tab. 2. MLD [7] is also selected as a direct baseline of our proposed method. Results on HumanML3D [14] are shown in Fig. 7. Though our method is not superior in R-Precision, we receive competitive user reviews, showcasing the efficacy of our motion-KP joint space. Also please notice that our model only contains 45.1M parameters, while T2M-GPT contains 228M parameters. T2M-GPT and MLD present similar R-Precision, but only T2M-GPT manages to keep a good performance with user reviews. Moreover, the discrepancy between R-Precision and user reviews is revealed in both absolute value and trends, also in Fig. 6." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.605, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Kinematic Phrases" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.143, + 0.49, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.147, + 0.778, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.275, + 0.79, + 0.307 + ], + "angle": 0, + "content": "Fig. 7: User study on HumanML3D, with Fig. 8: User study on KPG, with \"Y\" for \"Y\" for Yes and \"P\" for partially. Yes and \"P\" for partially." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.319, + 0.4, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.323, + 0.771, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.453, + 0.407, + 0.496 + ], + "angle": 0, + "content": "Fig.9: Our model supports fine-grained modification on motion via KP." + }, + { + "type": "image_caption", + "bbox": [ + 0.407, + 0.45, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Fig. 10: Visualization of generated samples. Compared to T2M-GPT, our method provides a better response to prompts with explicit constraints on specific body parts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.527, + 0.788, + 0.618 + ], + "angle": 0, + "content": "A similar user study on KPG is conducted in Fig. 7 with 100 randomly selected prompts from KPG involving T2M-GPT and our model. Fig. 8 demonstrates that KP-inferred Accuracy and user reviews share similar trends. Meanwhile, KP and user study give the same reviews for \\(84\\%\\) of the samples. We believe KPG could thus be a first step towards white-box automatic motion generation evaluation for semantic consistency. More analyses are in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.641, + 0.371, + 0.656 + ], + "angle": 0, + "content": "5.3 Visualization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.788, + 0.775 + ], + "angle": 0, + "content": "A modification sample is shown in Fig. 9. By modifying KP, we could edit arbitrary motion at a fine-grained level. Fig. 10 compares generated samples of T2M-GPT and our methods. Our method properly responds to text prompts with constraints on specific body parts thanks to KP, which explicitly models the action semantics with body part kinematics cues. Note that T2M-GPT generates redundant motion for simple prompts, while our method provides more concise and precise results. More visualizations are in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.797, + 0.402, + 0.811 + ], + "angle": 0, + "content": "5.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.823, + 0.519, + 0.837 + ], + "angle": 0, + "content": "KPG ablation study results are in Tab. 3." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.544, + 0.207 + ], + "angle": 0, + "content": "KP latent. By using a latent space without KP, a substantial performance drop is perceived, revealing the significance of KP in conveying action semantics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.544, + 0.298 + ], + "angle": 0, + "content": "Different KP sets. We examine the contribution of different KP sets: joint KP (PP), joint pair KP (PRPP, PDP), limb KP (LAP, LOP), and body KP (GVP). A leave-one-out evaluation shows the elimination of joint KP and joint pair KP results in notable perfor" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.715, + 0.313 + ], + "angle": 0, + "content": "mance degradation, while the influence of the rest is relatively subtle." + }, + { + "type": "table", + "bbox": [ + 0.56, + 0.149, + 0.788, + 0.242 + ], + "angle": 0, + "content": "
MethodsAcc. %↑Diversity
Ours57.866.048
w/o KP39.945.526
w/o Joint KP50.035.685
w/o Joint Pair KP47.245.772
w/o Limb KP55.925.934
w/o Body KP56.845.871
" + }, + { + "type": "table_caption", + "bbox": [ + 0.563, + 0.25, + 0.778, + 0.264 + ], + "angle": 0, + "content": "Table 3: KPG ablation results." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.337, + 0.355, + 0.352 + ], + "angle": 0, + "content": "6 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.788, + 0.564 + ], + "angle": 0, + "content": "Limitations. First, the current KP criteria guarantee objectivity but overlook kinematic information like amplitude and speed. Also, due to the granularity of the adopted skeleton, fine-grained information on fingers is not well-preserved. These would be promising for further exploration. Second, KPB could be extended to datasets with other modalities, like 2D pose and egocentric action datasets. Although these modalities provide incomplete 3D information, we could extract KP that is credibly accessible across modalities. Third, with the convenient KP-to-text conversion, auxiliary motion descriptions could be automatically generated via KP. Fourth, KPG could be extended by paraphrasing existing prompts and combining different Phrases. Finally, the proposed motion generation pipeline is devised for quick evaluation, while the motion-KP latent space could be incorporated with more advanced algorithms for better exploitation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.588, + 0.788, + 0.664 + ], + "angle": 0, + "content": "Broader Impacts. The proposed KP and KPG provide a white-box instrument to diagnose text-to-motion algorithms, which could help develop reliable motion tools. Better motion synthesis could advance fields including animation production, sports training, human-robot interaction, and rehabilitation. However, it could also be used for illegal media content or false information." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.687, + 0.36, + 0.703 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this paper, we proposed an intermediate representation to bridge human motion and action semantics as the Kinematic Phrase. KP achieved proper abstraction, interpretability, and generality by focusing on objective kinematic facts of human motion. A motion understanding system based on KP was proposed and proven effective in motion interpolation, modification, and generation. Moreover, Kinematic Prompt Generation is proposed as a white-box text-to-motion benchmark, bringing new insights to the literature. We believe that KP has great potential to advance motion understanding." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.605, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Kinematic Phrases" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.24 + ], + "angle": 0, + "content": "This work is supported in part by the National Natural Science Foundation of China under Grants No.62306175, No.62302296, National Key Research and Development Project of China (No.2022ZD0160102, No.2021ZD0110704), Shanghai Artificial Intelligence Laboratory, XPLORER PRIZE grants." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.264, + 0.323, + 0.279 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.786, + 0.351 + ], + "angle": 0, + "content": "1. Aristidou, A., Cohen-Or, D., Hodgins, J.K., Chrysanthou, Y., Shamir, A.: Deep motifs and motion signatures. ACM Trans. Graph. 37(6), 187:1-187:13 (Nov 2018). https://doi.org/10.1145/3272127.3275038, http://doi.acm.org/10.1145/3272127.3275038" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.354, + 0.786, + 0.394 + ], + "angle": 0, + "content": "2. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: Teach: Temporal action composition for 3d humans. In: 2022 International Conference on 3D Vision (3DV). pp. 414-423. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.395, + 0.786, + 0.422 + ], + "angle": 0, + "content": "3. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: SINC: Spatial composition of 3D human motions for simultaneous action generation. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.423, + 0.786, + 0.478 + ], + "angle": 0, + "content": "4. Azadi, S., Shah, A., Hayes, T., Parikh, D., Gupta, S.: Make-an-animation: Large-scale text-conditional 3d human motion generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15039-15048 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.479, + 0.786, + 0.519 + ], + "angle": 0, + "content": "5. Bartlett, R.: Introduction to Sports Biomechanics. Introduction to Sports Biomechanics, E & FN Spon (1997), https://books.google.com.tw/books?id=-6Db8mgxsqQC" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.52, + 0.786, + 0.548 + ], + "angle": 0, + "content": "6. Brégier, R.: Deep regression on manifolds: a 3d rotation case study. In: 2021 International Conference on 3D Vision (3DV). pp. 166-174. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.549, + 0.786, + 0.589 + ], + "angle": 0, + "content": "7. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18000-18010 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.591, + 0.786, + 0.645 + ], + "angle": 0, + "content": "8. Dabral, R., Mughal, M.H., Golyanik, V., Theobalt, C.: Mofusion: A framework for denoising-diffusion-based motion synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9760-9770 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.646, + 0.786, + 0.701 + ], + "angle": 0, + "content": "9. Delmas, G., Weinzaepfel, P., Lucas, T., Moreno-Noguer, F., Rogez, G.: Posescript: 3d human poses from natural language. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI. pp. 346-362. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.786, + 0.729 + ], + "angle": 0, + "content": "0. Delmas, G., Weinzaepfel, P., Moreno-Noguer, F., Rogez, G.: Posefix: Correcting 3d human poses with natural language. arXiv preprint arXiv:2309.08480 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.786, + 0.77 + ], + "angle": 0, + "content": "1. Fang, H.S., Li, J., Tang, H., Xu, C., Zhu, H., Xiu, Y., Li, Y.L., Lu, C.: Alpha-posed: Whole-body regional multi-person pose estimation and tracking in real-time. TPAMI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.771, + 0.786, + 0.812 + ], + "angle": 0, + "content": "12. Gao, X., Yang, Y., Xie, Z., Du, S., Sun, Z., Wu, Y.: Guess: Gradually enriching synthesis for text-driven human motion generation. IEEE Transactions on Visualization and Computer Graphics (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "13. Guo, C., Mu, Y., Javed, M.G., Wang, S., Cheng, L.: Momask: Generative masked modeling of 3d human motions. arXiv preprint arXiv:2312.00063 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.296, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "14. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5152-5161 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "15. Guo, C., Zuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In: European Conference on Computer Vision. pp. 580-597. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.787, + 0.286 + ], + "angle": 0, + "content": "16. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2021-2029 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.342 + ], + "angle": 0, + "content": "17. Guo, W., Du, Y., Shen, X., Lepetit, V., Alameda-Pineda, X., Moreno-Noguer, F.: Back to mlp: A simple baseline for human motion prediction. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4809-4819 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.383 + ], + "angle": 0, + "content": "18. Hassan, M., Ceylan, D., Villegas, R., Saito, J., Yang, J., Zhou, Y., Black, M.: Stochastic scene-aware motion prediction. In: Proceedings of the International Conference on Computer Vision 2021 (Oct 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.425 + ], + "angle": 0, + "content": "19. Hernandez, A., Gall, J., Moreno-Noguer, F.: Human motion prediction via spatiotemporal inpainting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7134-7143 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.426, + 0.787, + 0.452 + ], + "angle": 0, + "content": "20. Holden, D., Kanoun, O., Perepichka, M., Popa, T.: Learned motion matching. ACM Transactions on Graphics (TOG) 39(4), 53-1 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.493 + ], + "angle": 0, + "content": "21. Hong, F., Zhang, M., Pan, L., Cai, Z., Yang, L., Liu, Z.: Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.787, + 0.536 + ], + "angle": 0, + "content": "22. Ji, Y., Xu, F., Yang, Y., Shen, F., Shen, H.T., Zheng, W.S.: A large-scale rgb-d database for arbitrary-view human action recognition. In: Proceedings of the 26th ACM international Conference on Multimedia. pp. 1510-1518 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.787, + 0.563 + ], + "angle": 0, + "content": "23. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.787, + 0.605 + ], + "angle": 0, + "content": "24. Jin, P., Wu, Y., Fan, Y., Sun, Z., Yang, W., Yuan, L.: Act as you wish: Fine-grained control of motion diffusion model with hierarchical semantic graphs. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.66 + ], + "angle": 0, + "content": "25. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2151-2162 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.787, + 0.702 + ], + "angle": 0, + "content": "26. Kong, H., Gong, K., Lian, D., Mi, M.B., Wang, X.: Priority-centric human motion generation in discrete latent space. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14806-14816 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.787, + 0.729 + ], + "angle": 0, + "content": "27. Koppula, H.S., Saxena, A.: Anticipating human activities for reactive robotic response. In: IROS. p. 2071. Tokyo (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.787, + 0.769 + ], + "angle": 0, + "content": "28. von Laban, R., Lange, R.: Laban's Principles of Dance and Movement Notation. Macdonald & Evans (1975), https://books.google.com.tw/books?id=VrOAAAAAMAAJ" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.787, + 0.813 + ], + "angle": 0, + "content": "29. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13401-13412 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.787, + 0.84 + ], + "angle": 0, + "content": "30. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Learn to dance with aist++: Music conditioned 3d dance generation (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.605, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Kinematic Phrases" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "31. Li, Y.L., Wu, X., Liu, X., Wang, Z., Dou, Y., Ji, Y., Zhang, J., Li, Y., Lu, X., Tan, J., et al.: From isolated islands to pangea: Unifying semantic space for human action understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16582-16592 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.788, + 0.259 + ], + "angle": 0, + "content": "32. Lin, J., Chang, J., Liu, L., Li, G., Lin, L., Tian, Q., Chen, C.W.: Being comes from not-being: Open-vocabulary text-to-motion generation with wordless training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23222-23231 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.26, + 0.788, + 0.3 + ], + "angle": 0, + "content": "33. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM transactions on graphics (TOG) 34(6), 1-16 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.342 + ], + "angle": 0, + "content": "34. Lou, Y., Zhu, L., Wang, Y., Wang, X., Yang, Y.: Diversemotion: Towards diverse human motion generation via discrete diffusion. arXiv preprint arXiv:2309.01372 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.397 + ], + "angle": 0, + "content": "35. Lucas, T., Baradel, F., Weinzaepfel, P., Rogez, G.: Posegpt: quantization-based 3d human motion generation and forecasting. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI. pp. 417-435. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.439 + ], + "angle": 0, + "content": "36. Paden, B., Cap, M., Yong, S.Z., Yershov, D., Frazzoli, E.: A survey of motion planning and control techniques for self-driving urban vehicles. IEEE Transactions on intelligent vehicles 1(1), 33-55 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.495 + ], + "angle": 0, + "content": "37. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3D hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 10975-10985 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.495, + 0.788, + 0.536 + ], + "angle": 0, + "content": "38. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10985-10995 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.592 + ], + "angle": 0, + "content": "39. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII. pp. 480-497. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.592, + 0.788, + 0.633 + ], + "angle": 0, + "content": "40. Petrovich, M., Black, M.J., Varol, G.: Tmr: Text-to-motion retrieval using contrastive 3d human motion synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9488-9497 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.66 + ], + "angle": 0, + "content": "41. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data 4(4), 236-252 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.702 + ], + "angle": 0, + "content": "42. Pons-Moll, G., Fleet, D.J., Rosenhahn, B.: Posebits for monocular human pose estimation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2337-2344 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.702, + 0.788, + 0.757 + ], + "angle": 0, + "content": "43. Punnakkal, A.R., Chandrasekaran, A., Athanasiou, N., Quiros-Ramirez, A., Black, M.J.: Babel: bodies, action and behavior with english labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 722-731 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.758, + 0.788, + 0.813 + ], + "angle": 0, + "content": "44. Qian, Y., Urbanek, J., Hauptmann, A.G., Won, J.: Breaking the limits of text-conditioned 3d motion synthesis with elaborative descriptions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2306-2316 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.813, + 0.788, + 0.84 + ], + "angle": 0, + "content": "45. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.786, + 0.215 + ], + "angle": 0, + "content": "46. Tang, Y., Liu, J., Liu, A., Yang, B., Dai, W., Rao, Y., Lu, J., Zhou, J., Li, X.: Flag3d: A 3d fitness activity dataset with language instruction. arXiv preprint arXiv:2212.04638 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.216, + 0.786, + 0.271 + ], + "angle": 0, + "content": "47. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII. pp. 358-374. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.786, + 0.297 + ], + "angle": 0, + "content": "48. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.297, + 0.786, + 0.338 + ], + "angle": 0, + "content": "49. Van Welbergen, H., Van Basten, B.J., Egges, A., Ruttkay, Z.M., Overmars, M.H.: Real time animation of virtual humans: a trade-off between naturalness and control. In: Computer Graphics Forum. vol. 29, pp. 2530-2554. Wiley Online Library (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.786, + 0.392 + ], + "angle": 0, + "content": "50. Wang, Y., Leng, Z., Li, F.W.B., Wu, S.C., Liang, X.: Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 22035-22044 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.392, + 0.786, + 0.419 + ], + "angle": 0, + "content": "51. Xiang, W., Li, C., Zhou, Y., Wang, B., Zhang, L.: Language supervised training for skeleton-based action recognition. arXiv preprint arXiv:2208.05318 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.419, + 0.786, + 0.487 + ], + "angle": 0, + "content": "52. Xu, L., Song, Z., Wang, D., Su, J., Fang, Z., Ding, C., Gan, W., Yan, Y., Jin, X., Yang, X., Zeng, W., Wu, W.: Actformer: A gan-based transformer towards general action-conditioned 3d human motion generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2228-2238 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.487, + 0.786, + 0.528 + ], + "angle": 0, + "content": "53. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 16010-16021 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.786, + 0.569 + ], + "angle": 0, + "content": "54. Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.786, + 0.609 + ], + "angle": 0, + "content": "55. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.786, + 0.663 + ], + "angle": 0, + "content": "56. Zhang, M., Guo, X., Pan, L., Cai, Z., Hong, F., Li, H., Yang, L., Liu, Z.: Remodiffuse: Retrieval-augmented motion diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 364-373 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.663, + 0.786, + 0.704 + ], + "angle": 0, + "content": "57. Zhang, M., Li, H., Cai, Z., Ren, J., Yang, L., Liu, Z.: Finemogen: Fine-grained spatio-temporal motion generation and editing. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.786, + 0.759 + ], + "angle": 0, + "content": "58. Zhong, C., Hu, L., Zhang, Z., Xia, S.: Attt2m: Text-driven human motion generation with multi-perspective attention mechanism. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 509-519 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.786, + 0.8 + ], + "angle": 0, + "content": "59. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5745-5753 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.786, + 0.841 + ], + "angle": 0, + "content": "60. Zhou, Z., Wan, Y., Wang, B.: Avatargpt: All-in-one framework for motion understanding, planning, generation and beyond. arXiv preprint arXiv:2311.16468 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.786, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.605, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Kinematic Phrases" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.19 + ], + "angle": 0, + "content": "61. Zhou, Z., Wang, B.: Ude: A unified driving engine for human motion generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5632-5641 (June 2023)" + } + ] +] \ No newline at end of file diff --git a/2024/Bridging the Gap Between Human Motion and Action Semantics via Kinematics Phrases/ceda5829-f6a0-4ea2-94dd-2ddc6da8e6ed_origin.pdf b/2024/Bridging the Gap Between Human Motion and Action Semantics via Kinematics Phrases/ceda5829-f6a0-4ea2-94dd-2ddc6da8e6ed_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..20683dd1bb8bb22651d801c69dc2791d94fcdfc6 --- /dev/null +++ b/2024/Bridging the Gap Between Human Motion and Action Semantics via Kinematics Phrases/ceda5829-f6a0-4ea2-94dd-2ddc6da8e6ed_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1127cabfbb1aa376eff210f321ef7260668efbfa7eb68a7924ff3ef15494d079 +size 2880776 diff --git a/2024/Bridging the Gap Between Human Motion and Action Semantics via Kinematics Phrases/full.md b/2024/Bridging the Gap Between Human Motion and Action Semantics via Kinematics Phrases/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1c6f4c443857b456378f2673b53da9b3dd526448 --- /dev/null +++ b/2024/Bridging the Gap Between Human Motion and Action Semantics via Kinematics Phrases/full.md @@ -0,0 +1,297 @@ +# Bridging the Gap between Human Motion and Action Semantics via Kinematic Phrases + +Xinpeng Liu $^{1}$ , Yong-Lu Li $^{1\star}$ , Ailing Zeng $^{2}$ , Zizheng Zhou $^{1}$ , Yang You $^{3}$ , and Cewu Lu $^{1\star}$ + +$^{1}$ Shanghai Jiao Tong University + +2 Tencent + +3 Stanford University + +{xinpengliu0907,ailingzengzzz} $@$ gmail.com, + +{yonglu_li, zhou_zz, lucewu}@sjtu.edu.cn, yangyou@stanford.edu + +Abstract. Motion understanding aims to establish a reliable mapping between motion and action semantics, while it is a challenging many-to-many problem. An abstract action semantic (i.e., walk forwards) could be conveyed by perceptually diverse motions (walking with arms up or swinging). In contrast, a motion could carry different semantics w.r.t. its context and intention. This makes an elegant mapping between them difficult. Previous attempts adopted direct-mapping paradigms with limited reliability. Also, current automatic metrics fail to provide reliable assessments of the consistency between motions and action semantics. We identify the source of these problems as the significant gap between the two modalities. To alleviate this gap, we propose Kinematic Phrases (KP) that take the objective kinematic facts of human motion with proper abstraction, interpretability, and generality. Based on KP, we can unify a motion knowledge base and build a motion understanding system. Meanwhile, KP can be automatically converted from motions to text descriptions with no subjective bias, inspiring Kinematic Prompt Generation (KPG) as a novel white-box motion generation benchmark. In extensive experiments, our approach shows superiority over other methods. Our project is available at https://foruck.github.io/KP/. + +Keywords: Motion Representation $\cdot$ Motion Understanding $\cdot$ Text-to-Motion Benchmark + +# 1 Introduction + +Human motion understanding has a wide range of applications, including activity understanding [31], autonomous driving [36], robotics [27], and automatic animation [49], making it increasingly attractive. The core is to establish a mapping between the motion space and the action semantics space. The motion space indicates a space of sequential 3D human representations, e.g., 3D pose or SMPL [33]/SMPL-X [37] parameter sequence, while the action semantic space + +![](images/2f9a250c0b660f4e9de119af4ffb199a2a442a893c9c8c1ed7f35d6e24a7f298.jpg) +Fig. 1: The huge gap between motion and action semantics results in the many-to-many problem. We propose Kinematic Phrases (KP) as an intermediate to bridge the gap. KPs objectively capture human kinematic cues. It properly abstracts diverse motions with interpretability. As shown, the Phrases in the yellow box could capture key patterns of walk for diverse motions. + +can be represented as action categories or sentences described by natural language. + +Recently, a growing focus has been on generative mapping from semantics to motion, including action category-based generation [38] and text-based generation [7, 14, 35, 39, 48, 54, 55]. Most of them typically build a mapping that links motion and semantics either directly or via motion latent, with understated concerns for intermediate motion-semantic structures. However, these models suffer from inferior reliability. They cannot guarantee they generated correct samples without human filtering. Additionally, the existing evaluation of motion generation is problematic. Widely adopted FID and R-Precision rely on the latent space from a black-box pre-trained model, which might fail to out-of-distribution (OOD) and over-fitting cases. There is a long-standing need for an evaluation method that can cheaply and reliably assess whether a generated motion is consistent with particular action semantics. We identify the essence of these as the significant gap between raw human motion and action semantics, which makes direct mapping hard to learn. + +As in Fig. 1, an action semantics can correspond to diverse motions. For instance, a person could walk in countless ways with diverse motions, either with arms up or swinging, while action semantics tend to abstract these away from a walking motion. Additionally, they are robust against small perturbations, while motion is more specific and complex, with representations changing vastly when perturbed or mis-captured. Moreover, a motion sequence could have diverse semantics w.r.t. contexts. Modeling this many-to-many mapping between motion and semantics is challenging. + +To bridge this gap between motion and action semantics, we propose Kinematic Phrases (KP), an interpretable intermediate representation. KP focuses on the objective kinematic facts, which are usually omitted by general action semantics, like left-hand moving forwards then backward. KP is designed as qualitative categorical representations of these facts. For objectivity and actuality, KP captures sign changes with minimal pre-defined standards. Inspired by previous studies on kinematic human motion representation [5,28], KP is proposed + +as six types shown in Fig. 1, covering joint positions, joint pair positions and distances, limb angles and directions, and global velocity. Note that, although KP can be described by natural language, a major difference is that KP is strictly dedicated to objective kinematic facts instead of coarse actions such as surrender or fine-grained actions like raise both hands. + +We highlight three advantages of KP. First, KP offers proper abstraction, which disentangles motion perturbations and semantics changes, easing the learning process. Even though the motion differs significantly, KP manages to capture walk patterns easily. Second, KP is inter + +pretable, as it can be viewed as instructions on executing the action, making it easily understandable to humans. Finally, KP is general, as it can be automatically extracted from different modalities of human motion, including skeleton and SMPL parameters. The conversion from KP to text is also effortless. + +With KP as an intermediate representation, a unified large-scale motion knowledge base is constructed. Then, to fully exploit KP and the knowledge base, we build a motion understanding system with KP mediation. In detail, we learn a motion-KP joint latent space for multiple motion understanding applications, including motion interpolation, modification, and generation. Moreover, leveraging the interpretability of KP, we propose a benchmark called Kinematic Prompts Generation (KPG), which generates motion from text prompts converted from KPs. Thanks to the consistency and convenience of the KP-to-text conversion, KPG enables efficient white-box motion generation evaluation. Given the bottle-necked evaluation of the current motion generation benchmarks [14,41] as shown in Fig. 2, KPG could be a promising complementary benchmark. Results in Tab. 2 show that KPG brings novel insights into the current text-to-motion progress. + +Our contributions are: (1) We propose KP as an intermediate representation to bridge the gap between motion and action semantics. (2) We build a novel + +![](images/b9aefb9afaed2df8cb29e3234b1a68e80a618a25056dc470d94a3be214c7a60d.jpg) +Fig. 2: HumanML3D [14] metric values are approaching even surpassing GT-level while being increasingly indistinguishable. The higher-than-GT R-Precision could be insufficient as a semantic consistency indicator. + +motion understanding system using KP and the aggregated large-scale knowledge base. (3) We propose KPG as a benchmark for efficient white-box motion generation evaluation. Promising results are achieved on motion interpolation and generation tasks. Moreover, extensive user studies are conducted, verifying the efficacy of our methods, and the consistency between KPG evaluation and human perception. + +# 2 Related Works + +Motion Representation An intuitive motion representation is a sequence of static pose representations, like joint locations and limb rotations. Efforts are paid to address the discontinuity of rotation for deep-learning methods [6,59]. Recent works on parametric body models [33,37] enable a more realistic body representation. Meanwhile, [42] proposed Posebits, representing pose with boolean geometric part relationships. [9, 10] translates Posebits into text descriptions. These abstract representations are flexible and insensitive to little perturbations, but their static nature ignores motion dynamics. [46] acquire similar fine-grained descriptions from human annotation, while [3, 51] adopted large-scale language models. However, few recognize their potential in bridging the low-level motion and the high-level action semantics. Phase functions [20], Labanotations [28], and learned Motion Words [1] were also explored, though limited to specific actions like locomotion and dancing. + +Motion Generation can be conditioned by its prefix/suffix [2, 17, 19], action categories [16, 38, 52], or audio [29, 30]. Text-based motion generation has developed rapidly with the proposal of text-motion datasets [14, 43]. Early works [14, 39, 44] used VAEs, while some [21, 32, 47] extended the CLIP [45] space to motion. Diffusion models [8, 48, 50, 55] were also adopted. MAA [4] adopted a U-Net structure. Some [40, 56] explored retrieval-based methods. [25] aimed at controllable generation, while [53] introduced physical constraints. However, most approaches still suffer from the gap between motion and action semantics. Multiple methods [7, 15, 26, 35, 54, 58, 61] adopted (VQ-)VAE-compressed motion representation as mediation, while in the current data-limited situation, we identify that this single-modality compression might be sub-optimal. Instead, KP could alleviate this by introducing explicit semantic-geometric correlation. + +# 3 Kinematic Phrase Base + +# 3.1 Kinematic Phrases + +Kinematic Phrases abstract motion into objective kinematic facts like left-hand moves up qualitatively. We take inspiration from previous kinematic motion representations [28] and qualitative static pose representations [9,11,42], proposing six types of KP to comprehensively represent motion from different kinematic hierarchies: For joint movements, there are 34 Position Phrases (PPs). For joint + +![](images/ee41f23925b0bf02a9ea76e0fb59029b4cadfab8609702d49dde943777902438.jpg) +Fig.3: Six types of KP from four kinematic hierarchies are extracted from a motion sequence. A scalar indicator $s_i$ is calculated per Phrase per frame. Its sign categorizes the corresponding Phrase. + +pair movements, there are 242 Pairwise Relative Position Phrases (PRPPs) and 81 Pairwise Distance Phrases (PDPs). For limb movements, there are 8 Limb Angle Phrases (LAPs) and 24 Limb Orientation Phrases (LOPs). For whole-body movements, there are 3 Global Velocity Phrases (GVPs). KP extraction is based on a skeleton sequence $X = \{x_{i} | x_{i} \in \mathcal{R}^{n_{k} \times 3}\}_{i=1}^{t}$ , where $n_{k}$ is the number of joints ( $n_{k} = 17$ here), $x_{i}$ is the joint coordinates at $i$ -th frame, and $t$ is the sequence length. Note that $x_{i}^{0}$ indicates the pelvis/root joint. For each Phrase, a scalar indicator sequence is calculated from the skeleton sequence. Phrases are extracted as per-frame categorical representations w.r.t. indicator signs. Unlike previous efforts [9, 42], we limit the KP criteria as the indicator signs to minimize human-defined biases (e.g., numerical criteria on the closeness of two joints) for objectivity and actuality. Fig. 3 illustrated the procedure. + +Reference Vectors are first constructed, indicating right, upward, and forward directions from a human cognitive view. We aim at the egocentric reference frames that human tends to use when performing actions. The negative direction of gravity is adopted as upward vector $r^u$ , the vector from left hip to right hip is adopted as right vector $r^r$ , and the forward vector is calculated as $r^f = r^u \times r^r$ . These vectors of each frame are denoted as $R^{\cdot} = \{r_{i}^{\cdot}\}_{i=1}^{t}$ . + +Position Phrase (PP) focuses on the movement direction of joint $x^{j}$ w.r.t. reference vector $R$ . The indicator for PP at $i$ -th frame is calculated as + +$$ +s _ {i} ^ {(j, \cdot)} = \langle (x _ {i} ^ {j} - x _ {i} ^ {0}), r _ {i} ^ {\cdot} \rangle - \langle (x _ {i - 1} ^ {j} - x _ {i - 1} ^ {0}), r _ {i - 1} ^ {\cdot} \rangle . \tag {1} +$$ + +The sign of $s_i^{(j,\cdot)}$ categorizes PP into moving along/against $R'$ , or relatively static along $R'$ for indicators with small amplitudes. 12 joints (except the pelvis, eyes, and hips) and 3 reference vectors resulted in originally 36 PPs. Specifically, shoulders are excluded for the leftward vector, and hips are excluded for the upward vector, resulting in 36-2=34 PPs. + +Pairwise Relative Position Phrase (PRPP) describes the relative position of a pair of joints $(x^{j}, x^{k})$ w.r.t. reference vector $R$ . PRPP indicator at $i$ -th frame is + +$$ +s _ {i} ^ {(j, k, \cdot)} = \langle \left(x _ {i} ^ {j} - x _ {i} ^ {k}\right), r _ {i} ^ {\cdot} \rangle . \tag {2} +$$ + +For (L-Hand, R-Hand) and forward vector $R^f$ , PRPP could be L-Hand behind /in front of R-Hand according to the sign of $s_i^{(j,k,\cdot)}$ . 136 joint pairs and 3 reference vectors composed 408 Phrases originally. 24 joint pairs linked by limbs are filtered out. 30 Eye-related pairs are filtered out except (left eye, right eye) due to the others are covered by head-related pairs. 4 triplets that the relationship barely changes along the reference vector (e.g., right knee, left hip, and leftward vector) are filtered out. These result in $408 - 24^{*}3 - 30^{*}3 - 4 = 242$ PRPPs. + +Pairwise Distance Phrase (PDP) describes how the L2 distance between a pair of joints $(x^{j}, x^{k})$ changes. The indicator for PDP is calculated as + +$$ +s _ {i} ^ {(j, k)} = \left\| x _ {i} ^ {j} - x _ {i} ^ {k} \right\| _ {2} - \left\| x _ {i - 1} ^ {j} - x _ {i - 1} ^ {k} \right\| _ {2}. \tag {3} +$$ + +The sign of $s_i^{(j,k)}$ categorizes PDP into moving closer/away, or relatively static. Among the original 105 joint pairs, 24 pairs that are in the skeleton topology are excluded, such as the hand and elbow, resulting in 81 PDPs. + +Limb Angle Phrase (LAP) targets at the change of bend angle between two connected limbs $(x^{j}, x^{k})$ and $(x^{j}, x^{l})$ . The LAP indicator is calculated as + +$$ +s _ {i} ^ {(j, k, l)} = \operatorname {a r c c o s} \left(\left\langle x _ {i} ^ {k} - x _ {i} ^ {j}, x _ {i} ^ {l} - x _ {i} ^ {j} \right\rangle\right) - \operatorname {a r c c o s} \left(\left\langle x _ {i - 1} ^ {k} - x _ {i - 1} ^ {j}, x _ {i - 1} ^ {l} - x _ {i - 1} ^ {j} \right\rangle\right). \tag {4} +$$ + +8 LAPs describe the limb chain $(x^{j},x^{k}) - (x^{j},x^{l})$ as bending or unbending. + +Limb Orientation Phrase (LOP) describes the orientation of the limb $(x^{j}, x^{k})$ w.r.t. $R^{\prime}$ , with $x^{k}$ being the distal limb. The LOP indicator is + +$$ +s _ {i} ^ {(j, k, \cdot)} = \left\langle x _ {i} ^ {k} - x _ {i} ^ {j}, r _ {i} ^ {\prime} \right\rangle . \tag {5} +$$ + +The sign of $s_i^{(j,k,\cdot)}$ categorizes the LOP $(x^j, x^k)$ pointing along/against $R$ , or a placeholder category for those with little magnitude. 4 arms, 4 leg limbs, head, collarbones, hips, torsos, and upper body, paired with the 3 reference vectors, result in 45 LOPs in total. 21 pairs that barely change along the reference vector (e.g., left-right hip, leftward vector) are filtered out, resulting in 45-21=24 LOPs. + +Global Velocity Phrase (GVP) describes the direction of global velocity with respect to $R$ . The indicator is calculated as + +$$ +s _ {i} = \left\langle x _ {i + 1} ^ {0} - x _ {i} ^ {0}, r _ {i} ^ {\prime} \right\rangle . \tag {6} +$$ + +The three categories are moving along/against $R$ , or static along $R$ according to the sign of $s_i$ . + +These result in 392 Phrases in total, covering motion diversity and distribution from various levels. While we clarify that these Phrases do not rule out the possibility of other possible useful potentials. + +# 3.2 Constructing Kinematic Phrase Base + +KP enables us to unify motion data with different formats to construct a large knowledge base containing motion, text, and KP thanks to its generality. Motion sequences of different representations are collected, including 3D skeleton sequences and SMPL [33] /SMPL-X [37] parameter sequences. The sequences are first re-sampled to $30\mathrm{Hz}$ and rotated so that the negative direction of the z-axis is the gravity direction. Then, the sequences are converted into 3D skeleton sequences for KP extraction as in Sec. 3.1. Text annotations attached to the sequences are directly saved. For sequences with action category annotation, the category name is saved. For those with neither text nor action category, the text information is set from its attached additional information, like objects for SAMP [18]. Finally, we collect 140K motion sequences from 11 datasets. More details are included in the appendix. + +# 4 Applications of KP + +# 4.1 Motion Understanding with KP + +An intuitive application of KP is to leverage it for better motion understanding. To achieve this, we first learn a motion-KP joint space, which exploits the clarity and interpretability of KP. Then, we adopt the learned latent space for multiple motion understanding tasks. An overview is illustrated in Fig. 4. + +Preliminaries. Motion is represented as a pose sequence with $n$ frames as $M = \{m_i\}_{i=1}^n$ . In detail, SMPL [33] pose parameters are transformed to the 6D representation [59], then concatenated with the velocity of the root, resulting in a 147-dim representation per frame. KP is represented by indicator signs as $C = \{c_i\}_{i=1}^n, c_i \in \{1,0,-1\}^{392}$ . + +Joint Space Learning. The joint space is learned with Motion VAE $\{\mathcal{E}_m,\mathcal{D}_m\}$ and KP VAE $\{\mathcal{E}_p,\mathcal{D}_p\}$ . All VAEs follow the transformer-based architecture from [38]. Motion encoder $\mathcal{E}_m$ takes motion $M$ and two distribution tokens $m_{\mu},m_{\sigma}$ as input, and the outputs corresponding to the distribution tokens are taken as the $\mu_{m}$ and $\sigma_{m}$ of the Gaussian distribution. Then, motion decoder $\mathcal{D}_m$ takes $z_{m}\sim \mathcal{G}(\mu_{m},\sigma_{m})$ as $K,V$ , and a sinusoidal positional encoding of the expected duration as $Q$ . The output is fed into a linear layer to reconstruct motion $\hat{M}$ . KP VAE $\{\mathcal{E}_p,\mathcal{D}_p\}$ operates similarly, with the sign of $\mathcal{D}_p$ output as the predicted KP $\hat{C}$ . Notice that the $\mathcal{D}_m,\mathcal{D}_p$ could take arbitrary combinations of $z_{m},z_{p}$ as input, outputting $\hat{M},\hat{C}$ . To learn the joint space, besides the conventional loss + +![](images/505be9df18e697cb5c0fde68195f4fdb4b059e0510cce4f160c06ab451bdbabb.jpg) +Fig. 4: We train a motion-KP joint latent space to exploit the clarity and interpretability of KP. The space is then applied to multiple tasks, including motion interpolation, modification, and generation. + +of reconstruction and KL divergence, we also align the two VAEs similarly to [39]. In this way, we guide the learned motion latent space with the structure of KP. Meanwhile, we randomly corrupt samples during training by setting a small portion of KP as 0. The idea is that the overall representation should be robust with a small portion of KP unknown. Also, the missing Phrases should be recovered from existing Phrases. This mines the correlation among Phrases and increases the robustness of the space. More details are in the supplementary. + +Motion Understanding Tasks with Joint Space. With the joint space, we can perform both low-level and high-level motion understanding tasks. We introduce three applications to show the capability of KP, as shown in Fig. 4. For motion interpolation given a corrupted motion sequence $\tilde{M}$ , we extract its corresponding KP sequence $\tilde{C}$ , then feed them to encoders $\mathcal{E}_m, \mathcal{E}_p$ and decoder $\mathcal{D}_p$ , resulting in the estimated KP sequence $\hat{C}$ . $\hat{C}$ and $\tilde{M}$ are fed into $\mathcal{E}_m, \mathcal{E}_p$ and $\mathcal{D}_m$ , resulting in interpolated $\hat{M}$ . For motion modification on motion $M$ , modifications could be made on the extracted KP sequence $C$ resulting in $\tilde{C}$ . Motion frames to modify are then masked, getting $\tilde{M}$ . $\tilde{M}, \tilde{C}$ are fed into $\mathcal{E}_m, \mathcal{E}_p$ and $\mathcal{D}_m$ , getting the interpolated $\hat{M}$ . For motion generation, we simply replace the motion latent space in MLD [7] with the latent space of our Motion VAE. Experiments show that KP manages to boost MLD with a considerable margin. + +# 4.2 Motion Benchmarking with Kinematic Prompt Generation + +A more interesting and important application is benchmarking motion generation with the interpretability and objectivity of KP. + +![](images/eb1cabaf2b56169e8c55ab971f4da7cf711db6c5e367379870bf8c9032d58b0d.jpg) +Fig. 5: Kinematic Prompt Generation. 7,796 prompts are converted from KP with templates, corresponding to certain KP patterns. We calculate the generation accuracy by checking the appearance of the patterns in KP extracted from the generated motion. + +Before that, we first analyze current benchmarks. A crucial aspect of motion generation evaluation is motion-semantic consistency. The gold standard is user study. However, it is expensive and inefficient to scale. Early metrics like MPJPE (Mean Per Joint Position Error) and MAE (Mean Angle Error) mechanically calculate the error between the generated and GT samples. These metrics fail to reveal the real ability of generative models: What if the models memorize GT samples? Or what if the samples are diverse from GT but also true? FID (Frechet Inception Distance) is adopted to mitigate this issue. However, it provides a macro view of the quality of all generated samples without guarantees for individual samples. [14] proposed R-Precision, using a pre-trained text-motion matching model to examine whether the generated samples carry true semantics. They both rely on the latent space from a black-box pre-trained model, which is not credible. Besides, models might learn short paths to over-fit the pre-trained model. Meanwhile, since automatic mapping from motion to semantics is still an unsettled problem, adopting it to evaluate motion generation is not a decent choice. Moreover, most current motion generation evaluations are performed on datasets [14,22,41] with considerable complex everyday actions, further increasing the difficulty. As shown in Fig. 2 and Tab. 1, the current evaluation metrics are becoming increasingly indistinguishable. Also, the higher-than-GT R-Precision could be insufficient as an indicator of semantic consistency. + +To this end, we propose a novel benchmark: Kinematic Prompts Generation (KPG). Instead of previous benchmarks focusing on everyday activities or sports, we step back in the complexity of action semantics. Based on KP, KPG focuses on evaluating whether the models could generate motion sequences consistent with specific kinematic facts given text prompts. An overview is given in Fig. 5. + +In detail, we convert KP into 7,796 text prompts with templates as in Fig. 5. The prompts could be categorized into 4 groups. First, each atomic prompt + +involves only 1 KP, such as "left hand moves upwards", which evaluates the basic semantic consistency of text-to-motion models. Notice that PRPP and LOP are excluded since they represent static states, resulting in 252 prompts. Second, each repetitive prompt involves a KP being repeated two or three times., like "left arm bends twice", evaluating the basic counting ability. PRPP, LOP, and GVP are excluded, bringing 492 prompts. Third, each sequential prompt involves two independent KP to be sequentially executed, like "left arm bends, then right arm bends". 3,912 sequential prompts are randomly sampled. Finally, each simultaneous prompt involves two independent KPs to be simultaneously executed, like "left arm bends, and simultaneously, right arm bends." 3,120 prompts are randomly sampled. The sequential and simultaneous prompts examine whether the model can understand the composition of simple actions. Given prompt $T_{i} \in T$ from Phrase $c_{i}$ (or Phrases $(c_{i}^{1}, c_{i}^{2})$ ), the model generates motion $\hat{M}_{i}$ , along with extracted KP $\hat{C}_{i}$ . Then, whether $T_{i}$ is hit could be inferred via designed protocols by identifying certain KP patterns in $\hat{C}_{i}$ . In detail, the atomic prompt is identified as hit if $c_{i}$ appears for more than 5 consecutive frames in $\hat{C}_{i}$ . The repetitive prompt is identified as hit if and only if $c_{i}$ appears for at least 5 consecutive frames in $\hat{C}_{i}$ for corresponding times. The sequential prompt is identified as hit if $c_{i}^{2}$ appears for at least 5 consecutive frames after (maybe not immediately) $c_{i}^{1}$ does the same. The simultaneous prompt is identified as hit if $c_{i}^{1}$ and $c_{i}^{2}$ appear simultaneously for at least 5 consecutive frames. The calculation involves no black-box model thanks to KP, presenting a fully white-box evaluation. Also, with the effortless motion-to-KP conversion, the computation could be conducted automatically. More details are in the appendix. + +# 5 Experiment + +Implementation Details. HumanML3D [14] test split is held out for evaluation, with the rest of KPB for training. The batch size is set as 288, and an AdamW optimizer with a learning rate of 1e-4 is adopted. We randomly corrupt less than $20\%$ of the Phrases for a sample. The Motion-KP joint space is trained for 6,000 epochs. While the text-to-motion latent diffusion model is trained for 3,000 epochs, with the joint space frozen. All experiments are conducted in 4 NVIDIA RTX 3090 GPUs. More details are in the appendix. + +# 5.1 Motion Interpolation + +Following [23], $50\%$ frames are randomly masked for interpolation. In Tab. 1, our method provides better FID. With KPB, Diversity is increased. + +# 5.2 Motion Generation + +Results on conventional text to motion are shown in Tab. 1. We adopt the HumanML3D test set [14] for conventional text-to-motion evaluation. The evaluation model from [14] is adopted to calculate R-Precision, FID, Diversity, + +
MethodsInterpolationGeneration
FID↓Diversity→R-P@1↑FID↓Diversity→Multimodality
GT0.0029.5030.5110.0029.503-
TEMOS [39]--0.4243.7348.9730.368
T2M [14]--0.4551.0679.1882.090
MotionDiffuse [55]--0.4910.6309.4100.730
MDM [48]2.6988.4200.3200.5449.5592.799
TM2T [15]--0.4241.5018.5892.424
MLD [7]--0.4810.4739.7242.413
T2M-GPT [54]--0.4920.1419.7221.831
Fg-T2M [50]--0.4920.2439.2781.614
M2DM [26]--0.4970.3529.9263.587
AttT2M [58]--0.4990.1129.7002.452
ReMoDiffuse [56]--0.5100.1039.0181.795
MotionGPT [23]0.2149.5600.4920.2329.5282.008
FineMoGen [57]--0.5040.1519.2632.696
GraphMotion [24]--0.5040.1169.6922.766
AvatarGPT [60]--0.5100.1689.624-
GUESS [12]--0.5030.1099.8262.430
MDD [34]--0.5150.0729.6831.869
MoMask [13]--0.5210.045-1.241
Ours0.1929.6230.4960.2759.9752.218
Ours w/ KPB0.21410.1420.4710.64110.2622.891
+ +![](images/876177e7dcfd3b55c44669db4f1509c105ef03d6a26523ffd1ec139a8e88abb8.jpg) +Table 1: Results on HumanML3D. R-P@1 is short Fig.6: R-Precision could contrafor R-Precision@1. Most methods show comparadict to user reviews, indicatingble performance. Also, the higher-than-GT phethat R-Precision could produce nomenon indicates that R-Precision might be insuf-unfaithful evaluations of semantic ficient as a semantic consistency indicator. consistency. + +and Multimodality. We run the evaluation 20 times and report the average value. Our method is competitive without KPB. However, there exist some unsatisfying observations. First, KPB brings a counter-intuitive performance drop. Second, the metric values of most techniques are rather close (e.g., difference $< 0.03$ for R-P@1), and no substantial advantage is perceived for one method over others. Also, as discovered by MAA [4], despite its relatively poor R-Precision and FID, MAA manages to achieve better user reviews, indicating the potential loss of efficacy of these conventional metrics. This is further revealed in Fig. 6, where a contradiction is observed between human judgment and R-precision. Moreover, some methods show higher-than-GT R-P@1, which indicates that R-Precision might be insufficient as a semantic consistency indicator. Given these, we further switch to KPG and pursue further insights. + +Results on KPG are shown in Tab. 2. Besides accuracy, diversity and R-precision are reported. We run the evaluation 20 times and report the average metric value. The white-box designation of KPG evaluation makes it possible to evaluate the different aspects of the models. Meanwhile, KPG is considered an easier task than conventional text-based motion generation since it is targeted at much less complex action semantics. Results show that previous methods are not performing as well as expected. Also, R-Precision fails to model the performance on KPG with trivially low values. Most methods perform well for atomic prompts, but when it comes to compositions, the accuracy drops. Early efforts like HMDM [48] and MotionDiffuse [55] are surprisingly competitive, even surpassing several more recent works. T2M-GPT [54] outperforms other methods by a considerable margin, but it fails for repetitive prompts. An unexpected + +
MethodsAccuracyDiversityR-P@1
AtomicRepetitiveSequentialSimultaneousOverall
HMDM [48]93.2523.9856.1130.9045.165.9230.034
MotionDiffuse [55]93.2122.1551.5427.7641.495.2570.051
T2M-GPT [54]97.221.0270.2442.4755.596.1530.042
MLD [7]92.0622.7648.2628.0139.945.5260.061
ReMoDiffuse [56]53.173.0532.1631.0630.565.8090.048
MotionGPT [23]92.0616.8755.9831.7644.965.7760.039
MoMask [13]96.4328.8659.1032.6647.795.9650.060
Ours98.8026.0271.3243.0857.866.0480.053
+ +Table 2: Results on KPG demonstrate the various strengths of different methods. + +phenomenon is the failure of ReMoDiffuse [56], which might be constrained by its retrieval-based methodology since most prompts in KPG are not available in the retrieval database. Our method, which replaces the latent space of MLD [7], delivers substantial improvements, but the accuracy remains below $60\%$ , which is far from satisfying. A considerable gap exists between existing efforts and ideal motion generation models. More analyses are in the appendix. + +User Study. To reach a convincing performance comparison, we conduct user studies on both HumanML3D and KPG. Our user study is different from previous efforts in two aspects. First, instead of testing a small set of text prompts (less than 50 in previous works [7,48]), we randomly select 600 sentences from the HumanML3D test set. By scaling up, the result is convincing in reflecting the ability to generate motion for diverse text inputs. Second, without asking the volunteers to give a general rating for each sample or to choose between different samples, we ask them two questions: 1) Do the motion and the text match? and 2) Is the motion natural? For Q1, three choices are given as "No, Partially, Yes". For Q2, two choices are given as "Yes, No". In this way, we explicitly evaluate two aspects of text-to-motion tasks: semantic consistency and naturalness, corresponding to R-Precision and FID. For each prompt, we generate one sample considering the annotation cost. We claim that the models should generate natural text-matching motion most of the time so that the one-sample setting would not hurt the fidelity of our user study. 36 volunteers are invited, each reviewing 200 sequences. Thus each sequence receives 3 user reviews. Also, we compute R-precision@1 of the generated sequences. + +Besides our model, we select MDM [48] as a naive baseline, T2M-GPT [54] as the SOTA model given the results in Tab. 1 and Tab. 2. MLD [7] is also selected as a direct baseline of our proposed method. Results on HumanML3D [14] are shown in Fig. 7. Though our method is not superior in R-Precision, we receive competitive user reviews, showcasing the efficacy of our motion-KP joint space. Also please notice that our model only contains 45.1M parameters, while T2M-GPT contains 228M parameters. T2M-GPT and MLD present similar R-Precision, but only T2M-GPT manages to keep a good performance with user reviews. Moreover, the discrepancy between R-Precision and user reviews is revealed in both absolute value and trends, also in Fig. 6. + +![](images/e3fac0ccfdc387fd86d7b6828aa4ad633b3d6f5bbc3d86281ea1108948729ee4.jpg) +Fig. 7: User study on HumanML3D, with Fig. 8: User study on KPG, with "Y" for "Y" for Yes and "P" for partially. Yes and "P" for partially. + +![](images/2337b3ab3bcc685556299264bc486c0cef254659834cd7dd6fcf0362ba9726b8.jpg) + +![](images/550a8c39304d5e1005aa4a2da9d188b0cf6f100557600607c7695ad46a76d9ff.jpg) +Fig.9: Our model supports fine-grained modification on motion via KP. + +![](images/99d5e32a546234b54428d7d13d7545aff8c409e2bfff36f076bae8af62106c73.jpg) +Fig. 10: Visualization of generated samples. Compared to T2M-GPT, our method provides a better response to prompts with explicit constraints on specific body parts. + +A similar user study on KPG is conducted in Fig. 7 with 100 randomly selected prompts from KPG involving T2M-GPT and our model. Fig. 8 demonstrates that KP-inferred Accuracy and user reviews share similar trends. Meanwhile, KP and user study give the same reviews for $84\%$ of the samples. We believe KPG could thus be a first step towards white-box automatic motion generation evaluation for semantic consistency. More analyses are in the appendix. + +# 5.3 Visualization + +A modification sample is shown in Fig. 9. By modifying KP, we could edit arbitrary motion at a fine-grained level. Fig. 10 compares generated samples of T2M-GPT and our methods. Our method properly responds to text prompts with constraints on specific body parts thanks to KP, which explicitly models the action semantics with body part kinematics cues. Note that T2M-GPT generates redundant motion for simple prompts, while our method provides more concise and precise results. More visualizations are in the appendix. + +# 5.4 Ablation Studies + +KPG ablation study results are in Tab. 3. + +KP latent. By using a latent space without KP, a substantial performance drop is perceived, revealing the significance of KP in conveying action semantics. + +Different KP sets. We examine the contribution of different KP sets: joint KP (PP), joint pair KP (PRPP, PDP), limb KP (LAP, LOP), and body KP (GVP). A leave-one-out evaluation shows the elimination of joint KP and joint pair KP results in notable perfor + +mance degradation, while the influence of the rest is relatively subtle. + +
MethodsAcc. %↑Diversity
Ours57.866.048
w/o KP39.945.526
w/o Joint KP50.035.685
w/o Joint Pair KP47.245.772
w/o Limb KP55.925.934
w/o Body KP56.845.871
+ +Table 3: KPG ablation results. + +# 6 Discussion + +Limitations. First, the current KP criteria guarantee objectivity but overlook kinematic information like amplitude and speed. Also, due to the granularity of the adopted skeleton, fine-grained information on fingers is not well-preserved. These would be promising for further exploration. Second, KPB could be extended to datasets with other modalities, like 2D pose and egocentric action datasets. Although these modalities provide incomplete 3D information, we could extract KP that is credibly accessible across modalities. Third, with the convenient KP-to-text conversion, auxiliary motion descriptions could be automatically generated via KP. Fourth, KPG could be extended by paraphrasing existing prompts and combining different Phrases. Finally, the proposed motion generation pipeline is devised for quick evaluation, while the motion-KP latent space could be incorporated with more advanced algorithms for better exploitation. + +Broader Impacts. The proposed KP and KPG provide a white-box instrument to diagnose text-to-motion algorithms, which could help develop reliable motion tools. Better motion synthesis could advance fields including animation production, sports training, human-robot interaction, and rehabilitation. However, it could also be used for illegal media content or false information. + +# 7 Conclusion + +In this paper, we proposed an intermediate representation to bridge human motion and action semantics as the Kinematic Phrase. KP achieved proper abstraction, interpretability, and generality by focusing on objective kinematic facts of human motion. A motion understanding system based on KP was proposed and proven effective in motion interpolation, modification, and generation. Moreover, Kinematic Prompt Generation is proposed as a white-box text-to-motion benchmark, bringing new insights to the literature. We believe that KP has great potential to advance motion understanding. + +# Acknowledgments + +This work is supported in part by the National Natural Science Foundation of China under Grants No.62306175, No.62302296, National Key Research and Development Project of China (No.2022ZD0160102, No.2021ZD0110704), Shanghai Artificial Intelligence Laboratory, XPLORER PRIZE grants. + +# References + +1. Aristidou, A., Cohen-Or, D., Hodgins, J.K., Chrysanthou, Y., Shamir, A.: Deep motifs and motion signatures. ACM Trans. Graph. 37(6), 187:1-187:13 (Nov 2018). https://doi.org/10.1145/3272127.3275038, http://doi.acm.org/10.1145/3272127.3275038 +2. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: Teach: Temporal action composition for 3d humans. In: 2022 International Conference on 3D Vision (3DV). pp. 414-423. IEEE (2022) +3. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: SINC: Spatial composition of 3D human motions for simultaneous action generation. In: ICCV (2023) +4. Azadi, S., Shah, A., Hayes, T., Parikh, D., Gupta, S.: Make-an-animation: Large-scale text-conditional 3d human motion generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15039-15048 (October 2023) +5. Bartlett, R.: Introduction to Sports Biomechanics. Introduction to Sports Biomechanics, E & FN Spon (1997), https://books.google.com.tw/books?id=-6Db8mgxsqQC +6. Brégier, R.: Deep regression on manifolds: a 3d rotation case study. In: 2021 International Conference on 3D Vision (3DV). pp. 166-174. IEEE (2021) +7. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18000-18010 (2023) +8. Dabral, R., Mughal, M.H., Golyanik, V., Theobalt, C.: Mofusion: A framework for denoising-diffusion-based motion synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9760-9770 (June 2023) +9. Delmas, G., Weinzaepfel, P., Lucas, T., Moreno-Noguer, F., Rogez, G.: Posescript: 3d human poses from natural language. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI. pp. 346-362. Springer (2022) +0. Delmas, G., Weinzaepfel, P., Moreno-Noguer, F., Rogez, G.: Posefix: Correcting 3d human poses with natural language. arXiv preprint arXiv:2309.08480 (2023) +1. Fang, H.S., Li, J., Tang, H., Xu, C., Zhu, H., Xiu, Y., Li, Y.L., Lu, C.: Alpha-posed: Whole-body regional multi-person pose estimation and tracking in real-time. TPAMI (2023) +12. Gao, X., Yang, Y., Xie, Z., Du, S., Sun, Z., Wu, Y.: Guess: Gradually enriching synthesis for text-driven human motion generation. IEEE Transactions on Visualization and Computer Graphics (2024) +13. Guo, C., Mu, Y., Javed, M.G., Wang, S., Cheng, L.: Momask: Generative masked modeling of 3d human motions. arXiv preprint arXiv:2312.00063 (2023) + +14. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5152-5161 (June 2022) +15. Guo, C., Zuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In: European Conference on Computer Vision. pp. 580-597. Springer (2022) +16. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2021-2029 (2020) +17. Guo, W., Du, Y., Shen, X., Lepetit, V., Alameda-Pineda, X., Moreno-Noguer, F.: Back to mlp: A simple baseline for human motion prediction. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4809-4819 (2023) +18. Hassan, M., Ceylan, D., Villegas, R., Saito, J., Yang, J., Zhou, Y., Black, M.: Stochastic scene-aware motion prediction. In: Proceedings of the International Conference on Computer Vision 2021 (Oct 2021) +19. Hernandez, A., Gall, J., Moreno-Noguer, F.: Human motion prediction via spatiotemporal inpainting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7134-7143 (2019) +20. Holden, D., Kanoun, O., Perepichka, M., Popa, T.: Learned motion matching. ACM Transactions on Graphics (TOG) 39(4), 53-1 (2020) +21. Hong, F., Zhang, M., Pan, L., Cai, Z., Yang, L., Liu, Z.: Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535 (2022) +22. Ji, Y., Xu, F., Yang, Y., Shen, F., Shen, H.T., Zheng, W.S.: A large-scale rgb-d database for arbitrary-view human action recognition. In: Proceedings of the 26th ACM international Conference on Multimedia. pp. 1510-1518 (2018) +23. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795 (2023) +24. Jin, P., Wu, Y., Fan, Y., Sun, Z., Yang, W., Yuan, L.: Act as you wish: Fine-grained control of motion diffusion model with hierarchical semantic graphs. Advances in Neural Information Processing Systems 36 (2024) +25. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2151-2162 (October 2023) +26. Kong, H., Gong, K., Lian, D., Mi, M.B., Wang, X.: Priority-centric human motion generation in discrete latent space. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14806-14816 (October 2023) +27. Koppula, H.S., Saxena, A.: Anticipating human activities for reactive robotic response. In: IROS. p. 2071. Tokyo (2013) +28. von Laban, R., Lange, R.: Laban's Principles of Dance and Movement Notation. Macdonald & Evans (1975), https://books.google.com.tw/books?id=VrOAAAAAMAAJ +29. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13401-13412 (2021) +30. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Learn to dance with aist++: Music conditioned 3d dance generation (2021) + +31. Li, Y.L., Wu, X., Liu, X., Wang, Z., Dou, Y., Ji, Y., Zhang, J., Li, Y., Lu, X., Tan, J., et al.: From isolated islands to pangea: Unifying semantic space for human action understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16582-16592 (2024) +32. Lin, J., Chang, J., Liu, L., Li, G., Lin, L., Tian, Q., Chen, C.W.: Being comes from not-being: Open-vocabulary text-to-motion generation with wordless training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23222-23231 (June 2023) +33. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM transactions on graphics (TOG) 34(6), 1-16 (2015) +34. Lou, Y., Zhu, L., Wang, Y., Wang, X., Yang, Y.: Diversemotion: Towards diverse human motion generation via discrete diffusion. arXiv preprint arXiv:2309.01372 (2023) +35. Lucas, T., Baradel, F., Weinzaepfel, P., Rogez, G.: Posegpt: quantization-based 3d human motion generation and forecasting. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI. pp. 417-435. Springer (2022) +36. Paden, B., Cap, M., Yong, S.Z., Yershov, D., Frazzoli, E.: A survey of motion planning and control techniques for self-driving urban vehicles. IEEE Transactions on intelligent vehicles 1(1), 33-55 (2016) +37. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3D hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 10975-10985 (2019) +38. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10985-10995 (2021) +39. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII. pp. 480-497. Springer (2022) +40. Petrovich, M., Black, M.J., Varol, G.: Tmr: Text-to-motion retrieval using contrastive 3d human motion synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9488-9497 (October 2023) +41. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data 4(4), 236-252 (2016) +42. Pons-Moll, G., Fleet, D.J., Rosenhahn, B.: Posebits for monocular human pose estimation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2337-2344 (2014) +43. Punnakkal, A.R., Chandrasekaran, A., Athanasiou, N., Quiros-Ramirez, A., Black, M.J.: Babel: bodies, action and behavior with english labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 722-731 (2021) +44. Qian, Y., Urbanek, J., Hauptmann, A.G., Won, J.: Breaking the limits of text-conditioned 3d motion synthesis with elaborative descriptions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2306-2316 (October 2023) +45. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from + +natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +46. Tang, Y., Liu, J., Liu, A., Yang, B., Dai, W., Rao, Y., Lu, J., Zhou, J., Li, X.: Flag3d: A 3d fitness activity dataset with language instruction. arXiv preprint arXiv:2212.04638 (2022) +47. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII. pp. 358-374. Springer (2022) +48. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022) +49. Van Welbergen, H., Van Basten, B.J., Egges, A., Ruttkay, Z.M., Overmars, M.H.: Real time animation of virtual humans: a trade-off between naturalness and control. In: Computer Graphics Forum. vol. 29, pp. 2530-2554. Wiley Online Library (2010) +50. Wang, Y., Leng, Z., Li, F.W.B., Wu, S.C., Liang, X.: Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 22035-22044 (October 2023) +51. Xiang, W., Li, C., Zhou, Y., Wang, B., Zhang, L.: Language supervised training for skeleton-based action recognition. arXiv preprint arXiv:2208.05318 (2022) +52. Xu, L., Song, Z., Wang, D., Su, J., Fang, Z., Ding, C., Gan, W., Yan, Y., Jin, X., Yang, X., Zeng, W., Wu, W.: Actformer: A gan-based transformer towards general action-conditioned 3d human motion generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2228-2238 (October 2023) +53. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 16010-16021 (October 2023) +54. Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052 (2023) +55. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022) +56. Zhang, M., Guo, X., Pan, L., Cai, Z., Hong, F., Li, H., Yang, L., Liu, Z.: Remodiffuse: Retrieval-augmented motion diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 364-373 (October 2023) +57. Zhang, M., Li, H., Cai, Z., Ren, J., Yang, L., Liu, Z.: Finemogen: Fine-grained spatio-temporal motion generation and editing. Advances in Neural Information Processing Systems 36 (2024) +58. Zhong, C., Hu, L., Zhang, Z., Xia, S.: Attt2m: Text-driven human motion generation with multi-perspective attention mechanism. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 509-519 (October 2023) +59. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5745-5753 (2019) +60. Zhou, Z., Wan, Y., Wang, B.: Avatargpt: All-in-one framework for motion understanding, planning, generation and beyond. arXiv preprint arXiv:2311.16468 (2023) + +61. Zhou, Z., Wang, B.: Ude: A unified driving engine for human motion generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5632-5641 (June 2023) \ No newline at end of file diff --git a/2024/Bridging the Gap Between Human Motion and Action Semantics via Kinematics Phrases/images.zip b/2024/Bridging the Gap Between Human Motion and Action Semantics via Kinematics Phrases/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3bb18b51c96b4bc5e547229ec31ba5baf9266029 --- /dev/null +++ b/2024/Bridging the Gap Between Human Motion and Action Semantics via Kinematics Phrases/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80bd496e5d1de965f9fb77f7f8fa2ad5941eb982a9ce94efa6a25ac451308bc5 +size 497508 diff --git a/2024/Bridging the Gap Between Human Motion and Action Semantics via Kinematics Phrases/layout.json b/2024/Bridging the Gap Between Human Motion and Action Semantics via Kinematics Phrases/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c8c571fba7c96883f3bdeb779b42462e106c3c58 --- /dev/null +++ b/2024/Bridging the Gap Between Human Motion and Action Semantics via Kinematics Phrases/layout.json @@ -0,0 +1,9625 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 111, + 473, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 111, + 473, + 148 + ], + "spans": [ + { + "bbox": [ + 143, + 111, + 473, + 148 + ], + "type": "text", + "content": "Bridging the Gap between Human Motion and Action Semantics via Kinematic Phrases" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "spans": [ + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "text", + "content": "Xinpeng Liu" + }, + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "text", + "content": ", Yong-Lu Li" + }, + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "text", + "content": ", Ailing Zeng" + }, + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "text", + "content": ", Zizheng Zhou" + }, + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "text", + "content": ", Yang You" + }, + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "text", + "content": ", and Cewu Lu" + }, + { + "bbox": [ + 136, + 167, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{1\\star}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 238, + 201, + 375, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 201, + 375, + 213 + ], + "spans": [ + { + "bbox": [ + 238, + 201, + 375, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 238, + 201, + 375, + 213 + ], + "type": "text", + "content": " Shanghai Jiao Tong University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 285, + 213, + 328, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 285, + 213, + 328, + 223 + ], + "spans": [ + { + "bbox": [ + 285, + 213, + 328, + 223 + ], + "type": "text", + "content": "2 Tencent" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 261, + 224, + 352, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 224, + 352, + 235 + ], + "spans": [ + { + "bbox": [ + 261, + 224, + 352, + 235 + ], + "type": "text", + "content": "3 Stanford University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 209, + 236, + 403, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 236, + 403, + 246 + ], + "spans": [ + { + "bbox": [ + 209, + 236, + 403, + 246 + ], + "type": "text", + "content": "{xinpengliu0907,ailingzengzzz} " + }, + { + "bbox": [ + 209, + 236, + 403, + 246 + ], + "type": "inline_equation", + "content": "@" + }, + { + "bbox": [ + 209, + 236, + 403, + 246 + ], + "type": "text", + "content": " gmail.com," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 165, + 247, + 449, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 247, + 449, + 257 + ], + "spans": [ + { + "bbox": [ + 165, + 247, + 449, + 257 + ], + "type": "text", + "content": "{yonglu_li, zhou_zz, lucewu}@sjtu.edu.cn, yangyou@stanford.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 287, + 455, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 287, + 455, + 497 + ], + "spans": [ + { + "bbox": [ + 159, + 287, + 455, + 497 + ], + "type": "text", + "content": "Abstract. Motion understanding aims to establish a reliable mapping between motion and action semantics, while it is a challenging many-to-many problem. An abstract action semantic (i.e., walk forwards) could be conveyed by perceptually diverse motions (walking with arms up or swinging). In contrast, a motion could carry different semantics w.r.t. its context and intention. This makes an elegant mapping between them difficult. Previous attempts adopted direct-mapping paradigms with limited reliability. Also, current automatic metrics fail to provide reliable assessments of the consistency between motions and action semantics. We identify the source of these problems as the significant gap between the two modalities. To alleviate this gap, we propose Kinematic Phrases (KP) that take the objective kinematic facts of human motion with proper abstraction, interpretability, and generality. Based on KP, we can unify a motion knowledge base and build a motion understanding system. Meanwhile, KP can be automatically converted from motions to text descriptions with no subjective bias, inspiring Kinematic Prompt Generation (KPG) as a novel white-box motion generation benchmark. In extensive experiments, our approach shows superiority over other methods. Our project is available at https://foruck.github.io/KP/." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 506, + 455, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 506, + 455, + 528 + ], + "spans": [ + { + "bbox": [ + 159, + 506, + 455, + 528 + ], + "type": "text", + "content": "Keywords: Motion Representation " + }, + { + "bbox": [ + 159, + 506, + 455, + 528 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 506, + 455, + 528 + ], + "type": "text", + "content": " Motion Understanding " + }, + { + "bbox": [ + 159, + 506, + 455, + 528 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 506, + 455, + 528 + ], + "type": "text", + "content": " Text-to-Motion Benchmark" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 549, + 231, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 549, + 231, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 549, + 231, + 562 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "content": "Human motion understanding has a wide range of applications, including activity understanding [31], autonomous driving [36], robotics [27], and automatic animation [49], making it increasingly attractive. The core is to establish a mapping between the motion space and the action semantics space. The motion space indicates a space of sequential 3D human representations, e.g., 3D pose or SMPL [33]/SMPL-X [37] parameter sequence, while the action semantic space" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 240, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 240, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 240, + 665 + ], + "type": "text", + "content": "* Corresponding authors." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 270 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 270 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 270 + ], + "type": "image", + "image_path": "2f9a250c0b660f4e9de119af4ffb199a2a442a893c9c8c1ed7f35d6e24a7f298.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 278, + 482, + 334 + ], + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 334 + ], + "type": "text", + "content": "Fig. 1: The huge gap between motion and action semantics results in the many-to-many problem. We propose Kinematic Phrases (KP) as an intermediate to bridge the gap. KPs objectively capture human kinematic cues. It properly abstracts diverse motions with interpretability. As shown, the Phrases in the yellow box could capture key patterns of walk for diverse motions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 368, + 481, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 481, + 392 + ], + "type": "text", + "content": "can be represented as action categories or sentences described by natural language." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 397, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 482, + 564 + ], + "type": "text", + "content": "Recently, a growing focus has been on generative mapping from semantics to motion, including action category-based generation [38] and text-based generation [7, 14, 35, 39, 48, 54, 55]. Most of them typically build a mapping that links motion and semantics either directly or via motion latent, with understated concerns for intermediate motion-semantic structures. However, these models suffer from inferior reliability. They cannot guarantee they generated correct samples without human filtering. Additionally, the existing evaluation of motion generation is problematic. Widely adopted FID and R-Precision rely on the latent space from a black-box pre-trained model, which might fail to out-of-distribution (OOD) and over-fitting cases. There is a long-standing need for an evaluation method that can cheaply and reliably assess whether a generated motion is consistent with particular action semantics. We identify the essence of these as the significant gap between raw human motion and action semantics, which makes direct mapping hard to learn." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "As in Fig. 1, an action semantics can correspond to diverse motions. For instance, a person could walk in countless ways with diverse motions, either with arms up or swinging, while action semantics tend to abstract these away from a walking motion. Additionally, they are robust against small perturbations, while motion is more specific and complex, with representations changing vastly when perturbed or mis-captured. Moreover, a motion sequence could have diverse semantics w.r.t. contexts. Modeling this many-to-many mapping between motion and semantics is challenging." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "To bridge this gap between motion and action semantics, we propose Kinematic Phrases (KP), an interpretable intermediate representation. KP focuses on the objective kinematic facts, which are usually omitted by general action semantics, like left-hand moving forwards then backward. KP is designed as qualitative categorical representations of these facts. For objectivity and actuality, KP captures sign changes with minimal pre-defined standards. Inspired by previous studies on kinematic human motion representation [5,28], KP is proposed" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 201, + 298, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 298, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 298, + 331 + ], + "type": "text", + "content": "as six types shown in Fig. 1, covering joint positions, joint pair positions and distances, limb angles and directions, and global velocity. Note that, although KP can be described by natural language, a major difference is that KP is strictly dedicated to objective kinematic facts instead of coarse actions such as surrender or fine-grained actions like raise both hands." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 335, + 299, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 299, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 299, + 430 + ], + "type": "text", + "content": "We highlight three advantages of KP. First, KP offers proper abstraction, which disentangles motion perturbations and semantics changes, easing the learning process. Even though the motion differs significantly, KP manages to capture walk patterns easily. Second, KP is inter" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 430, + 481, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 481, + 479 + ], + "type": "text", + "content": "pretable, as it can be viewed as instructions on executing the action, making it easily understandable to humans. Finally, KP is general, as it can be automatically extracted from different modalities of human motion, including skeleton and SMPL parameters. The conversion from KP to text is also effortless." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 482, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 482, + 638 + ], + "type": "text", + "content": "With KP as an intermediate representation, a unified large-scale motion knowledge base is constructed. Then, to fully exploit KP and the knowledge base, we build a motion understanding system with KP mediation. In detail, we learn a motion-KP joint latent space for multiple motion understanding applications, including motion interpolation, modification, and generation. Moreover, leveraging the interpretability of KP, we propose a benchmark called Kinematic Prompts Generation (KPG), which generates motion from text prompts converted from KPs. Thanks to the consistency and convenience of the KP-to-text conversion, KPG enables efficient white-box motion generation evaluation. Given the bottle-necked evaluation of the current motion generation benchmarks [14,41] as shown in Fig. 2, KPG could be a promising complementary benchmark. Results in Tab. 2 show that KPG brings novel insights into the current text-to-motion progress." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Our contributions are: (1) We propose KP as an intermediate representation to bridge the gap between motion and action semantics. (2) We build a novel" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 224, + 477, + 337 + ], + "blocks": [ + { + "bbox": [ + 310, + 224, + 477, + 337 + ], + "lines": [ + { + "bbox": [ + 310, + 224, + 477, + 337 + ], + "spans": [ + { + "bbox": [ + 310, + 224, + 477, + 337 + ], + "type": "image", + "image_path": "b9aefb9afaed2df8cb29e3234b1a68e80a618a25056dc470d94a3be214c7a60d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 346, + 482, + 401 + ], + "lines": [ + { + "bbox": [ + 304, + 346, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 482, + 401 + ], + "type": "text", + "content": "Fig. 2: HumanML3D [14] metric values are approaching even surpassing GT-level while being increasingly indistinguishable. The higher-than-GT R-Precision could be insufficient as a semantic consistency indicator." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "type": "text", + "content": "Kinematic Phrases" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "motion understanding system using KP and the aggregated large-scale knowledge base. (3) We propose KPG as a benchmark for efficient white-box motion generation evaluation. Promising results are achieved on motion interpolation and generation tasks. Moreover, extensive user studies are conducted, verifying the efficacy of our methods, and the consistency between KPG evaluation and human perception." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 205, + 243, + 218 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 243, + 218 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 243, + 218 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 229, + 483, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 229, + 483, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 229, + 483, + 386 + ], + "type": "text", + "content": "Motion Representation An intuitive motion representation is a sequence of static pose representations, like joint locations and limb rotations. Efforts are paid to address the discontinuity of rotation for deep-learning methods [6,59]. Recent works on parametric body models [33,37] enable a more realistic body representation. Meanwhile, [42] proposed Posebits, representing pose with boolean geometric part relationships. [9, 10] translates Posebits into text descriptions. These abstract representations are flexible and insensitive to little perturbations, but their static nature ignores motion dynamics. [46] acquire similar fine-grained descriptions from human annotation, while [3, 51] adopted large-scale language models. However, few recognize their potential in bridging the low-level motion and the high-level action semantics. Phase functions [20], Labanotations [28], and learned Motion Words [1] were also explored, though limited to specific actions like locomotion and dancing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 401, + 483, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 483, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 483, + 547 + ], + "type": "text", + "content": "Motion Generation can be conditioned by its prefix/suffix [2, 17, 19], action categories [16, 38, 52], or audio [29, 30]. Text-based motion generation has developed rapidly with the proposal of text-motion datasets [14, 43]. Early works [14, 39, 44] used VAEs, while some [21, 32, 47] extended the CLIP [45] space to motion. Diffusion models [8, 48, 50, 55] were also adopted. MAA [4] adopted a U-Net structure. Some [40, 56] explored retrieval-based methods. [25] aimed at controllable generation, while [53] introduced physical constraints. However, most approaches still suffer from the gap between motion and action semantics. Multiple methods [7, 15, 26, 35, 54, 58, 61] adopted (VQ-)VAE-compressed motion representation as mediation, while in the current data-limited situation, we identify that this single-modality compression might be sub-optimal. Instead, KP could alleviate this by introducing explicit semantic-geometric correlation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 562, + 293, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 562, + 293, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 562, + 293, + 575 + ], + "type": "text", + "content": "3 Kinematic Phrase Base" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 586, + 257, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 257, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 257, + 598 + ], + "type": "text", + "content": "3.1 Kinematic Phrases" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Kinematic Phrases abstract motion into objective kinematic facts like left-hand moves up qualitatively. We take inspiration from previous kinematic motion representations [28] and qualitative static pose representations [9,11,42], proposing six types of KP to comprehensively represent motion from different kinematic hierarchies: For joint movements, there are 34 Position Phrases (PPs). For joint" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 101 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 114, + 473, + 322 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 473, + 322 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 473, + 322 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 473, + 322 + ], + "type": "image", + "image_path": "ee41f23925b0bf02a9ea76e0fb59029b4cadfab8609702d49dde943777902438.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 330, + 482, + 364 + ], + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 364 + ], + "type": "text", + "content": "Fig.3: Six types of KP from four kinematic hierarchies are extracted from a motion sequence. A scalar indicator " + }, + { + "bbox": [ + 130, + 330, + 482, + 364 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 130, + 330, + 482, + 364 + ], + "type": "text", + "content": " is calculated per Phrase per frame. Its sign categorizes the corresponding Phrase." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 392, + 482, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 537 + ], + "type": "text", + "content": "pair movements, there are 242 Pairwise Relative Position Phrases (PRPPs) and 81 Pairwise Distance Phrases (PDPs). For limb movements, there are 8 Limb Angle Phrases (LAPs) and 24 Limb Orientation Phrases (LOPs). For whole-body movements, there are 3 Global Velocity Phrases (GVPs). KP extraction is based on a skeleton sequence " + }, + { + "bbox": [ + 130, + 392, + 482, + 537 + ], + "type": "inline_equation", + "content": "X = \\{x_{i} | x_{i} \\in \\mathcal{R}^{n_{k} \\times 3}\\}_{i=1}^{t}" + }, + { + "bbox": [ + 130, + 392, + 482, + 537 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 392, + 482, + 537 + ], + "type": "inline_equation", + "content": "n_{k}" + }, + { + "bbox": [ + 130, + 392, + 482, + 537 + ], + "type": "text", + "content": " is the number of joints (" + }, + { + "bbox": [ + 130, + 392, + 482, + 537 + ], + "type": "inline_equation", + "content": "n_{k} = 17" + }, + { + "bbox": [ + 130, + 392, + 482, + 537 + ], + "type": "text", + "content": " here), " + }, + { + "bbox": [ + 130, + 392, + 482, + 537 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 392, + 482, + 537 + ], + "type": "text", + "content": " is the joint coordinates at " + }, + { + "bbox": [ + 130, + 392, + 482, + 537 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 392, + 482, + 537 + ], + "type": "text", + "content": "-th frame, and " + }, + { + "bbox": [ + 130, + 392, + 482, + 537 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 392, + 482, + 537 + ], + "type": "text", + "content": " is the sequence length. Note that " + }, + { + "bbox": [ + 130, + 392, + 482, + 537 + ], + "type": "inline_equation", + "content": "x_{i}^{0}" + }, + { + "bbox": [ + 130, + 392, + 482, + 537 + ], + "type": "text", + "content": " indicates the pelvis/root joint. For each Phrase, a scalar indicator sequence is calculated from the skeleton sequence. Phrases are extracted as per-frame categorical representations w.r.t. indicator signs. Unlike previous efforts [9, 42], we limit the KP criteria as the indicator signs to minimize human-defined biases (e.g., numerical criteria on the closeness of two joints) for objectivity and actuality. Fig. 3 illustrated the procedure." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 539, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 539, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 539, + 482, + 612 + ], + "type": "text", + "content": "Reference Vectors are first constructed, indicating right, upward, and forward directions from a human cognitive view. We aim at the egocentric reference frames that human tends to use when performing actions. The negative direction of gravity is adopted as upward vector " + }, + { + "bbox": [ + 130, + 539, + 482, + 612 + ], + "type": "inline_equation", + "content": "r^u" + }, + { + "bbox": [ + 130, + 539, + 482, + 612 + ], + "type": "text", + "content": ", the vector from left hip to right hip is adopted as right vector " + }, + { + "bbox": [ + 130, + 539, + 482, + 612 + ], + "type": "inline_equation", + "content": "r^r" + }, + { + "bbox": [ + 130, + 539, + 482, + 612 + ], + "type": "text", + "content": ", and the forward vector is calculated as " + }, + { + "bbox": [ + 130, + 539, + 482, + 612 + ], + "type": "inline_equation", + "content": "r^f = r^u \\times r^r" + }, + { + "bbox": [ + 130, + 539, + 482, + 612 + ], + "type": "text", + "content": ". These vectors of each frame are denoted as " + }, + { + "bbox": [ + 130, + 539, + 482, + 612 + ], + "type": "inline_equation", + "content": "R^{\\cdot} = \\{r_{i}^{\\cdot}\\}_{i=1}^{t}" + }, + { + "bbox": [ + 130, + 539, + 482, + 612 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 613, + 481, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 613, + 481, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 613, + 481, + 638 + ], + "type": "text", + "content": "Position Phrase (PP) focuses on the movement direction of joint " + }, + { + "bbox": [ + 130, + 613, + 481, + 638 + ], + "type": "inline_equation", + "content": "x^{j}" + }, + { + "bbox": [ + 130, + 613, + 481, + 638 + ], + "type": "text", + "content": " w.r.t. reference vector " + }, + { + "bbox": [ + 130, + 613, + 481, + 638 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 613, + 481, + 638 + ], + "type": "text", + "content": ". The indicator for PP at " + }, + { + "bbox": [ + 130, + 613, + 481, + 638 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 613, + 481, + 638 + ], + "type": "text", + "content": "-th frame is calculated as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 205, + 651, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 651, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 205, + 651, + 481, + 667 + ], + "type": "interline_equation", + "content": "s _ {i} ^ {(j, \\cdot)} = \\langle (x _ {i} ^ {j} - x _ {i} ^ {0}), r _ {i} ^ {\\cdot} \\rangle - \\langle (x _ {i - 1} ^ {j} - x _ {i - 1} ^ {0}), r _ {i - 1} ^ {\\cdot} \\rangle . \\tag {1}", + "image_path": "0e2df4a8381957fe200f8055144728fb3e1125a488807d20f1c77367158006cf.jpg" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 100 + ], + "type": "text", + "content": "Kinematic Phrases" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 176 + ], + "type": "text", + "content": "The sign of " + }, + { + "bbox": [ + 130, + 115, + 479, + 176 + ], + "type": "inline_equation", + "content": "s_i^{(j,\\cdot)}" + }, + { + "bbox": [ + 130, + 115, + 479, + 176 + ], + "type": "text", + "content": " categorizes PP into moving along/against " + }, + { + "bbox": [ + 130, + 115, + 479, + 176 + ], + "type": "inline_equation", + "content": "R'" + }, + { + "bbox": [ + 130, + 115, + 479, + 176 + ], + "type": "text", + "content": ", or relatively static along " + }, + { + "bbox": [ + 130, + 115, + 479, + 176 + ], + "type": "inline_equation", + "content": "R'" + }, + { + "bbox": [ + 130, + 115, + 479, + 176 + ], + "type": "text", + "content": " for indicators with small amplitudes. 12 joints (except the pelvis, eyes, and hips) and 3 reference vectors resulted in originally 36 PPs. Specifically, shoulders are excluded for the leftward vector, and hips are excluded for the upward vector, resulting in 36-2=34 PPs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 479, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 479, + 210 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 479, + 210 + ], + "type": "text", + "content": "Pairwise Relative Position Phrase (PRPP) describes the relative position of a pair of joints " + }, + { + "bbox": [ + 130, + 176, + 479, + 210 + ], + "type": "inline_equation", + "content": "(x^{j}, x^{k})" + }, + { + "bbox": [ + 130, + 176, + 479, + 210 + ], + "type": "text", + "content": " w.r.t. reference vector " + }, + { + "bbox": [ + 130, + 176, + 479, + 210 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 176, + 479, + 210 + ], + "type": "text", + "content": ". PRPP indicator at " + }, + { + "bbox": [ + 130, + 176, + 479, + 210 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 176, + 479, + 210 + ], + "type": "text", + "content": "-th frame is" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 255, + 210, + 480, + 225 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 210, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 255, + 210, + 480, + 225 + ], + "type": "interline_equation", + "content": "s _ {i} ^ {(j, k, \\cdot)} = \\langle \\left(x _ {i} ^ {j} - x _ {i} ^ {k}\\right), r _ {i} ^ {\\cdot} \\rangle . \\tag {2}", + "image_path": "c1ccb77a5e1b514026b56d33c3e8d786b5e4bbb9ed2e89b0c72fe8157e82ee05.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 228, + 480, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 480, + 312 + ], + "type": "text", + "content": "For (L-Hand, R-Hand) and forward vector " + }, + { + "bbox": [ + 130, + 228, + 480, + 312 + ], + "type": "inline_equation", + "content": "R^f" + }, + { + "bbox": [ + 130, + 228, + 480, + 312 + ], + "type": "text", + "content": ", PRPP could be L-Hand behind /in front of R-Hand according to the sign of " + }, + { + "bbox": [ + 130, + 228, + 480, + 312 + ], + "type": "inline_equation", + "content": "s_i^{(j,k,\\cdot)}" + }, + { + "bbox": [ + 130, + 228, + 480, + 312 + ], + "type": "text", + "content": ". 136 joint pairs and 3 reference vectors composed 408 Phrases originally. 24 joint pairs linked by limbs are filtered out. 30 Eye-related pairs are filtered out except (left eye, right eye) due to the others are covered by head-related pairs. 4 triplets that the relationship barely changes along the reference vector (e.g., right knee, left hip, and leftward vector) are filtered out. These result in " + }, + { + "bbox": [ + 130, + 228, + 480, + 312 + ], + "type": "inline_equation", + "content": "408 - 24^{*}3 - 30^{*}3 - 4 = 242" + }, + { + "bbox": [ + 130, + 228, + 480, + 312 + ], + "type": "text", + "content": " PRPPs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 314, + 479, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 314, + 479, + 338 + ], + "spans": [ + { + "bbox": [ + 131, + 314, + 479, + 338 + ], + "type": "text", + "content": "Pairwise Distance Phrase (PDP) describes how the L2 distance between a pair of joints " + }, + { + "bbox": [ + 131, + 314, + 479, + 338 + ], + "type": "inline_equation", + "content": "(x^{j}, x^{k})" + }, + { + "bbox": [ + 131, + 314, + 479, + 338 + ], + "type": "text", + "content": " changes. The indicator for PDP is calculated as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 225, + 346, + 480, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 346, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 225, + 346, + 480, + 361 + ], + "type": "interline_equation", + "content": "s _ {i} ^ {(j, k)} = \\left\\| x _ {i} ^ {j} - x _ {i} ^ {k} \\right\\| _ {2} - \\left\\| x _ {i - 1} ^ {j} - x _ {i - 1} ^ {k} \\right\\| _ {2}. \\tag {3}", + "image_path": "e1e5f0a712760e7872f0fc8507b827059b02b91d35e3f243808760cdbf5fd91f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 369, + 479, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 369, + 479, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 369, + 479, + 407 + ], + "type": "text", + "content": "The sign of " + }, + { + "bbox": [ + 131, + 369, + 479, + 407 + ], + "type": "inline_equation", + "content": "s_i^{(j,k)}" + }, + { + "bbox": [ + 131, + 369, + 479, + 407 + ], + "type": "text", + "content": " categorizes PDP into moving closer/away, or relatively static. Among the original 105 joint pairs, 24 pairs that are in the skeleton topology are excluded, such as the hand and elbow, resulting in 81 PDPs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 407, + 479, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 407, + 479, + 431 + ], + "spans": [ + { + "bbox": [ + 131, + 407, + 479, + 431 + ], + "type": "text", + "content": "Limb Angle Phrase (LAP) targets at the change of bend angle between two connected limbs " + }, + { + "bbox": [ + 131, + 407, + 479, + 431 + ], + "type": "inline_equation", + "content": "(x^{j}, x^{k})" + }, + { + "bbox": [ + 131, + 407, + 479, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 407, + 479, + 431 + ], + "type": "inline_equation", + "content": "(x^{j}, x^{l})" + }, + { + "bbox": [ + 131, + 407, + 479, + 431 + ], + "type": "text", + "content": ". The LAP indicator is calculated as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 438, + 480, + 454 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 438, + 480, + 454 + ], + "spans": [ + { + "bbox": [ + 142, + 438, + 480, + 454 + ], + "type": "interline_equation", + "content": "s _ {i} ^ {(j, k, l)} = \\operatorname {a r c c o s} \\left(\\left\\langle x _ {i} ^ {k} - x _ {i} ^ {j}, x _ {i} ^ {l} - x _ {i} ^ {j} \\right\\rangle\\right) - \\operatorname {a r c c o s} \\left(\\left\\langle x _ {i - 1} ^ {k} - x _ {i - 1} ^ {j}, x _ {i - 1} ^ {l} - x _ {i - 1} ^ {j} \\right\\rangle\\right). \\tag {4}", + "image_path": "f925b1f7a945b20dec52055318ad51e3c5b54be82c99b2c4c06178739c473517.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 461, + 459, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 461, + 459, + 473 + ], + "spans": [ + { + "bbox": [ + 131, + 461, + 459, + 473 + ], + "type": "text", + "content": "8 LAPs describe the limb chain " + }, + { + "bbox": [ + 131, + 461, + 459, + 473 + ], + "type": "inline_equation", + "content": "(x^{j},x^{k}) - (x^{j},x^{l})" + }, + { + "bbox": [ + 131, + 461, + 459, + 473 + ], + "type": "text", + "content": " as bending or unbending." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 474, + 479, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 474, + 479, + 498 + ], + "spans": [ + { + "bbox": [ + 131, + 474, + 479, + 498 + ], + "type": "text", + "content": "Limb Orientation Phrase (LOP) describes the orientation of the limb " + }, + { + "bbox": [ + 131, + 474, + 479, + 498 + ], + "type": "inline_equation", + "content": "(x^{j}, x^{k})" + }, + { + "bbox": [ + 131, + 474, + 479, + 498 + ], + "type": "text", + "content": " w.r.t. " + }, + { + "bbox": [ + 131, + 474, + 479, + 498 + ], + "type": "inline_equation", + "content": "R^{\\prime}" + }, + { + "bbox": [ + 131, + 474, + 479, + 498 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 131, + 474, + 479, + 498 + ], + "type": "inline_equation", + "content": "x^{k}" + }, + { + "bbox": [ + 131, + 474, + 479, + 498 + ], + "type": "text", + "content": " being the distal limb. The LOP indicator is" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 259, + 506, + 480, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 506, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 259, + 506, + 480, + 521 + ], + "type": "interline_equation", + "content": "s _ {i} ^ {(j, k, \\cdot)} = \\left\\langle x _ {i} ^ {k} - x _ {i} ^ {j}, r _ {i} ^ {\\prime} \\right\\rangle . \\tag {5}", + "image_path": "c72e51dc0c54bf33570f3175705710b6721cd05f84ec088ba56ea3ee9be4e1a1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 528, + 480, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 528, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 131, + 528, + 480, + 590 + ], + "type": "text", + "content": "The sign of " + }, + { + "bbox": [ + 131, + 528, + 480, + 590 + ], + "type": "inline_equation", + "content": "s_i^{(j,k,\\cdot)}" + }, + { + "bbox": [ + 131, + 528, + 480, + 590 + ], + "type": "text", + "content": " categorizes the LOP " + }, + { + "bbox": [ + 131, + 528, + 480, + 590 + ], + "type": "inline_equation", + "content": "(x^j, x^k)" + }, + { + "bbox": [ + 131, + 528, + 480, + 590 + ], + "type": "text", + "content": " pointing along/against " + }, + { + "bbox": [ + 131, + 528, + 480, + 590 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 131, + 528, + 480, + 590 + ], + "type": "text", + "content": ", or a placeholder category for those with little magnitude. 4 arms, 4 leg limbs, head, collarbones, hips, torsos, and upper body, paired with the 3 reference vectors, result in 45 LOPs in total. 21 pairs that barely change along the reference vector (e.g., left-right hip, leftward vector) are filtered out, resulting in 45-21=24 LOPs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 590, + 479, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 590, + 479, + 613 + ], + "spans": [ + { + "bbox": [ + 131, + 590, + 479, + 613 + ], + "type": "text", + "content": "Global Velocity Phrase (GVP) describes the direction of global velocity with respect to " + }, + { + "bbox": [ + 131, + 590, + 479, + 613 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 131, + 590, + 479, + 613 + ], + "type": "text", + "content": ". The indicator is calculated as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 264, + 621, + 480, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 621, + 480, + 634 + ], + "spans": [ + { + "bbox": [ + 264, + 621, + 480, + 634 + ], + "type": "interline_equation", + "content": "s _ {i} = \\left\\langle x _ {i + 1} ^ {0} - x _ {i} ^ {0}, r _ {i} ^ {\\prime} \\right\\rangle . \\tag {6}", + "image_path": "8aa5192c814bb75394d6cb0416274725ffd31a8a48266f21807d0d1bdad659e0.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "text", + "content": "The three categories are moving along/against " + }, + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "text", + "content": ", or static along " + }, + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "text", + "content": " according to the sign of " + }, + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 131, + 641, + 479, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 101 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "These result in 392 Phrases in total, covering motion diversity and distribution from various levels. While we clarify that these Phrases do not rule out the possibility of other possible useful potentials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 170, + 348, + 183 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 170, + 348, + 183 + ], + "spans": [ + { + "bbox": [ + 131, + 170, + 348, + 183 + ], + "type": "text", + "content": "3.2 Constructing Kinematic Phrase Base" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 190, + 482, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 482, + 334 + ], + "type": "text", + "content": "KP enables us to unify motion data with different formats to construct a large knowledge base containing motion, text, and KP thanks to its generality. Motion sequences of different representations are collected, including 3D skeleton sequences and SMPL [33] /SMPL-X [37] parameter sequences. The sequences are first re-sampled to " + }, + { + "bbox": [ + 130, + 190, + 482, + 334 + ], + "type": "inline_equation", + "content": "30\\mathrm{Hz}" + }, + { + "bbox": [ + 130, + 190, + 482, + 334 + ], + "type": "text", + "content": " and rotated so that the negative direction of the z-axis is the gravity direction. Then, the sequences are converted into 3D skeleton sequences for KP extraction as in Sec. 3.1. Text annotations attached to the sequences are directly saved. For sequences with action category annotation, the category name is saved. For those with neither text nor action category, the text information is set from its attached additional information, like objects for SAMP [18]. Finally, we collect 140K motion sequences from 11 datasets. More details are included in the appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 353, + 269, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 353, + 269, + 367 + ], + "spans": [ + { + "bbox": [ + 131, + 353, + 269, + 367 + ], + "type": "text", + "content": "4 Applications of KP" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 378, + 323, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 378, + 323, + 391 + ], + "spans": [ + { + "bbox": [ + 131, + 378, + 323, + 391 + ], + "type": "text", + "content": "4.1 Motion Understanding with KP" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 399, + 480, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 480, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 480, + 448 + ], + "type": "text", + "content": "An intuitive application of KP is to leverage it for better motion understanding. To achieve this, we first learn a motion-KP joint space, which exploits the clarity and interpretability of KP. Then, we adopt the learned latent space for multiple motion understanding tasks. An overview is illustrated in Fig. 4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 464, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 464, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 131, + 464, + 482, + 525 + ], + "type": "text", + "content": "Preliminaries. Motion is represented as a pose sequence with " + }, + { + "bbox": [ + 131, + 464, + 482, + 525 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 131, + 464, + 482, + 525 + ], + "type": "text", + "content": " frames as " + }, + { + "bbox": [ + 131, + 464, + 482, + 525 + ], + "type": "inline_equation", + "content": "M = \\{m_i\\}_{i=1}^n" + }, + { + "bbox": [ + 131, + 464, + 482, + 525 + ], + "type": "text", + "content": ". In detail, SMPL [33] pose parameters are transformed to the 6D representation [59], then concatenated with the velocity of the root, resulting in a 147-dim representation per frame. KP is represented by indicator signs as " + }, + { + "bbox": [ + 131, + 464, + 482, + 525 + ], + "type": "inline_equation", + "content": "C = \\{c_i\\}_{i=1}^n, c_i \\in \\{1,0,-1\\}^{392}" + }, + { + "bbox": [ + 131, + 464, + 482, + 525 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "text", + "content": "Joint Space Learning. The joint space is learned with Motion VAE " + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{E}_m,\\mathcal{D}_m\\}" + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "text", + "content": " and KP VAE " + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{E}_p,\\mathcal{D}_p\\}" + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "text", + "content": ". All VAEs follow the transformer-based architecture from [38]. Motion encoder " + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_m" + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "text", + "content": " takes motion " + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "text", + "content": " and two distribution tokens " + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "m_{\\mu},m_{\\sigma}" + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "text", + "content": " as input, and the outputs corresponding to the distribution tokens are taken as the " + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mu_{m}" + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sigma_{m}" + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "text", + "content": " of the Gaussian distribution. Then, motion decoder " + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_m" + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "text", + "content": " takes " + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "z_{m}\\sim \\mathcal{G}(\\mu_{m},\\sigma_{m})" + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "K,V" + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "text", + "content": ", and a sinusoidal positional encoding of the expected duration as " + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "text", + "content": ". The output is fed into a linear layer to reconstruct motion " + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "text", + "content": ". KP VAE " + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{E}_p,\\mathcal{D}_p\\}" + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "text", + "content": " operates similarly, with the sign of " + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_p" + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "text", + "content": " output as the predicted KP " + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{C}" + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "text", + "content": ". Notice that the " + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_m,\\mathcal{D}_p" + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "text", + "content": " could take arbitrary combinations of " + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "z_{m},z_{p}" + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "text", + "content": " as input, outputting " + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{M},\\hat{C}" + }, + { + "bbox": [ + 131, + 543, + 482, + 666 + ], + "type": "text", + "content": ". To learn the joint space, besides the conventional loss" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 101 + ], + "type": "text", + "content": "Kinematic Phrases" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 155, + 114, + 465, + 276 + ], + "blocks": [ + { + "bbox": [ + 155, + 114, + 465, + 276 + ], + "lines": [ + { + "bbox": [ + 155, + 114, + 465, + 276 + ], + "spans": [ + { + "bbox": [ + 155, + 114, + 465, + 276 + ], + "type": "image", + "image_path": "505be9df18e697cb5c0fde68195f4fdb4b059e0510cce4f160c06ab451bdbabb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 285, + 480, + 319 + ], + "lines": [ + { + "bbox": [ + 130, + 285, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 480, + 319 + ], + "type": "text", + "content": "Fig. 4: We train a motion-KP joint latent space to exploit the clarity and interpretability of KP. The space is then applied to multiple tasks, including motion interpolation, modification, and generation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 342, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 429 + ], + "type": "text", + "content": "of reconstruction and KL divergence, we also align the two VAEs similarly to [39]. In this way, we guide the learned motion latent space with the structure of KP. Meanwhile, we randomly corrupt samples during training by setting a small portion of KP as 0. The idea is that the overall representation should be robust with a small portion of KP unknown. Also, the missing Phrases should be recovered from existing Phrases. This mines the correlation among Phrases and increases the robustness of the space. More details are in the supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "text", + "content": "Motion Understanding Tasks with Joint Space. With the joint space, we can perform both low-level and high-level motion understanding tasks. We introduce three applications to show the capability of KP, as shown in Fig. 4. For motion interpolation given a corrupted motion sequence " + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\tilde{M}" + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "text", + "content": ", we extract its corresponding KP sequence " + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\tilde{C}" + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "text", + "content": ", then feed them to encoders " + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_m, \\mathcal{E}_p" + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_p" + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "text", + "content": ", resulting in the estimated KP sequence " + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\hat{C}" + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\hat{C}" + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\tilde{M}" + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "text", + "content": " are fed into " + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_m, \\mathcal{E}_p" + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_m" + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "text", + "content": ", resulting in interpolated " + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "text", + "content": ". For motion modification on motion " + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "text", + "content": ", modifications could be made on the extracted KP sequence " + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "text", + "content": " resulting in " + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\tilde{C}" + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "text", + "content": ". Motion frames to modify are then masked, getting " + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\tilde{M}" + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\tilde{M}, \\tilde{C}" + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "text", + "content": " are fed into " + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_m, \\mathcal{E}_p" + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_m" + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "text", + "content": ", getting the interpolated " + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 130, + 449, + 482, + 596 + ], + "type": "text", + "content": ". For motion generation, we simply replace the motion latent space in MLD [7] with the latent space of our Motion VAE. Experiments show that KP manages to boost MLD with a considerable margin." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 458, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 458, + 630 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 458, + 630 + ], + "type": "text", + "content": "4.2 Motion Benchmarking with Kinematic Prompt Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "A more interesting and important application is benchmarking motion generation with the interpretability and objectivity of KP." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 114, + 462, + 270 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 462, + 270 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 462, + 270 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 462, + 270 + ], + "type": "image", + "image_path": "eb1cabaf2b56169e8c55ab971f4da7cf711db6c5e367379870bf8c9032d58b0d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 277, + 482, + 311 + ], + "lines": [ + { + "bbox": [ + 131, + 277, + 482, + 311 + ], + "spans": [ + { + "bbox": [ + 131, + 277, + 482, + 311 + ], + "type": "text", + "content": "Fig. 5: Kinematic Prompt Generation. 7,796 prompts are converted from KP with templates, corresponding to certain KP patterns. We calculate the generation accuracy by checking the appearance of the patterns in KP extracted from the generated motion." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 338, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 578 + ], + "type": "text", + "content": "Before that, we first analyze current benchmarks. A crucial aspect of motion generation evaluation is motion-semantic consistency. The gold standard is user study. However, it is expensive and inefficient to scale. Early metrics like MPJPE (Mean Per Joint Position Error) and MAE (Mean Angle Error) mechanically calculate the error between the generated and GT samples. These metrics fail to reveal the real ability of generative models: What if the models memorize GT samples? Or what if the samples are diverse from GT but also true? FID (Frechet Inception Distance) is adopted to mitigate this issue. However, it provides a macro view of the quality of all generated samples without guarantees for individual samples. [14] proposed R-Precision, using a pre-trained text-motion matching model to examine whether the generated samples carry true semantics. They both rely on the latent space from a black-box pre-trained model, which is not credible. Besides, models might learn short paths to over-fit the pre-trained model. Meanwhile, since automatic mapping from motion to semantics is still an unsettled problem, adopting it to evaluate motion generation is not a decent choice. Moreover, most current motion generation evaluations are performed on datasets [14,22,41] with considerable complex everyday actions, further increasing the difficulty. As shown in Fig. 2 and Tab. 1, the current evaluation metrics are becoming increasingly indistinguishable. Also, the higher-than-GT R-Precision could be insufficient as an indicator of semantic consistency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 579, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 579, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 579, + 482, + 640 + ], + "type": "text", + "content": "To this end, we propose a novel benchmark: Kinematic Prompts Generation (KPG). Instead of previous benchmarks focusing on everyday activities or sports, we step back in the complexity of action semantics. Based on KP, KPG focuses on evaluating whether the models could generate motion sequences consistent with specific kinematic facts given text prompts. An overview is given in Fig. 5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "In detail, we convert KP into 7,796 text prompts with templates as in Fig. 5. The prompts could be categorized into 4 groups. First, each atomic prompt" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 100 + ], + "type": "text", + "content": "Kinematic Phrases" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "text", + "content": "involves only 1 KP, such as \"left hand moves upwards\", which evaluates the basic semantic consistency of text-to-motion models. Notice that PRPP and LOP are excluded since they represent static states, resulting in 252 prompts. Second, each repetitive prompt involves a KP being repeated two or three times., like \"left arm bends twice\", evaluating the basic counting ability. PRPP, LOP, and GVP are excluded, bringing 492 prompts. Third, each sequential prompt involves two independent KP to be sequentially executed, like \"left arm bends, then right arm bends\". 3,912 sequential prompts are randomly sampled. Finally, each simultaneous prompt involves two independent KPs to be simultaneously executed, like \"left arm bends, and simultaneously, right arm bends.\" 3,120 prompts are randomly sampled. The sequential and simultaneous prompts examine whether the model can understand the composition of simple actions. Given prompt " + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "inline_equation", + "content": "T_{i} \\in T" + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "text", + "content": " from Phrase " + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "text", + "content": " (or Phrases " + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "inline_equation", + "content": "(c_{i}^{1}, c_{i}^{2})" + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "text", + "content": "), the model generates motion " + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "inline_equation", + "content": "\\hat{M}_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "text", + "content": ", along with extracted KP " + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "inline_equation", + "content": "\\hat{C}_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "text", + "content": ". Then, whether " + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "text", + "content": " is hit could be inferred via designed protocols by identifying certain KP patterns in " + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "inline_equation", + "content": "\\hat{C}_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "text", + "content": ". In detail, the atomic prompt is identified as hit if " + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "text", + "content": " appears for more than 5 consecutive frames in " + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "inline_equation", + "content": "\\hat{C}_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "text", + "content": ". The repetitive prompt is identified as hit if and only if " + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "text", + "content": " appears for at least 5 consecutive frames in " + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "inline_equation", + "content": "\\hat{C}_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "text", + "content": " for corresponding times. The sequential prompt is identified as hit if " + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "inline_equation", + "content": "c_{i}^{2}" + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "text", + "content": " appears for at least 5 consecutive frames after (maybe not immediately) " + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "inline_equation", + "content": "c_{i}^{1}" + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "text", + "content": " does the same. The simultaneous prompt is identified as hit if " + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "inline_equation", + "content": "c_{i}^{1}" + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "inline_equation", + "content": "c_{i}^{2}" + }, + { + "bbox": [ + 130, + 116, + 482, + 405 + ], + "type": "text", + "content": " appear simultaneously for at least 5 consecutive frames. The calculation involves no black-box model thanks to KP, presenting a fully white-box evaluation. Also, with the effortless motion-to-KP conversion, the computation could be conducted automatically. More details are in the appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 422, + 224, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 422, + 224, + 436 + ], + "spans": [ + { + "bbox": [ + 132, + 422, + 224, + 436 + ], + "type": "text", + "content": "5 Experiment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 447, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 533 + ], + "type": "text", + "content": "Implementation Details. HumanML3D [14] test split is held out for evaluation, with the rest of KPB for training. The batch size is set as 288, and an AdamW optimizer with a learning rate of 1e-4 is adopted. We randomly corrupt less than " + }, + { + "bbox": [ + 130, + 447, + 482, + 533 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 447, + 482, + 533 + ], + "type": "text", + "content": " of the Phrases for a sample. The Motion-KP joint space is trained for 6,000 epochs. While the text-to-motion latent diffusion model is trained for 3,000 epochs, with the joint space frozen. All experiments are conducted in 4 NVIDIA RTX 3090 GPUs. More details are in the appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 548, + 268, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 268, + 561 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 268, + 561 + ], + "type": "text", + "content": "5.1 Motion Interpolation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 568, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 481, + 593 + ], + "type": "text", + "content": "Following [23], " + }, + { + "bbox": [ + 130, + 568, + 481, + 593 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 568, + 481, + 593 + ], + "type": "text", + "content": " frames are randomly masked for interpolation. In Tab. 1, our method provides better FID. With KPB, Diversity is increased." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 609, + 257, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 257, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 257, + 621 + ], + "type": "text", + "content": "5.2 Motion Generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": "Results on conventional text to motion are shown in Tab. 1. We adopt the HumanML3D test set [14] for conventional text-to-motion evaluation. The evaluation model from [14] is adopted to calculate R-Precision, FID, Diversity," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 114, + 343, + 272 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 343, + 272 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 343, + 272 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 343, + 272 + ], + "type": "table", + "html": "
MethodsInterpolationGeneration
FID↓Diversity→R-P@1↑FID↓Diversity→Multimodality
GT0.0029.5030.5110.0029.503-
TEMOS [39]--0.4243.7348.9730.368
T2M [14]--0.4551.0679.1882.090
MotionDiffuse [55]--0.4910.6309.4100.730
MDM [48]2.6988.4200.3200.5449.5592.799
TM2T [15]--0.4241.5018.5892.424
MLD [7]--0.4810.4739.7242.413
T2M-GPT [54]--0.4920.1419.7221.831
Fg-T2M [50]--0.4920.2439.2781.614
M2DM [26]--0.4970.3529.9263.587
AttT2M [58]--0.4990.1129.7002.452
ReMoDiffuse [56]--0.5100.1039.0181.795
MotionGPT [23]0.2149.5600.4920.2329.5282.008
FineMoGen [57]--0.5040.1519.2632.696
GraphMotion [24]--0.5040.1169.6922.766
AvatarGPT [60]--0.5100.1689.624-
GUESS [12]--0.5030.1099.8262.430
MDD [34]--0.5150.0729.6831.869
MoMask [13]--0.5210.045-1.241
Ours0.1929.6230.4960.2759.9752.218
Ours w/ KPB0.21410.1420.4710.64110.2622.891
", + "image_path": "3aaf9d47c3d9aac7a9b1d750804dabb41d5e7e16cf0e6744ceb86f21a8cf4905.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 343, + 115, + 484, + 268 + ], + "blocks": [ + { + "bbox": [ + 343, + 115, + 484, + 268 + ], + "lines": [ + { + "bbox": [ + 343, + 115, + 484, + 268 + ], + "spans": [ + { + "bbox": [ + 343, + 115, + 484, + 268 + ], + "type": "image", + "image_path": "876177e7dcfd3b55c44669db4f1509c105ef03d6a26523ffd1ec139a8e88abb8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 270, + 485, + 326 + ], + "lines": [ + { + "bbox": [ + 130, + 270, + 485, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 485, + 326 + ], + "type": "text", + "content": "Table 1: Results on HumanML3D. R-P@1 is short Fig.6: R-Precision could contrafor R-Precision@1. Most methods show comparadict to user reviews, indicatingble performance. Also, the higher-than-GT phethat R-Precision could produce nomenon indicates that R-Precision might be insuf-unfaithful evaluations of semantic ficient as a semantic consistency indicator. consistency." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "text", + "content": "and Multimodality. We run the evaluation 20 times and report the average value. Our method is competitive without KPB. However, there exist some unsatisfying observations. First, KPB brings a counter-intuitive performance drop. Second, the metric values of most techniques are rather close (e.g., difference " + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "inline_equation", + "content": "< 0.03" + }, + { + "bbox": [ + 130, + 360, + 482, + 504 + ], + "type": "text", + "content": " for R-P@1), and no substantial advantage is perceived for one method over others. Also, as discovered by MAA [4], despite its relatively poor R-Precision and FID, MAA manages to achieve better user reviews, indicating the potential loss of efficacy of these conventional metrics. This is further revealed in Fig. 6, where a contradiction is observed between human judgment and R-precision. Moreover, some methods show higher-than-GT R-P@1, which indicates that R-Precision might be insufficient as a semantic consistency indicator. Given these, we further switch to KPG and pursue further insights." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Results on KPG are shown in Tab. 2. Besides accuracy, diversity and R-precision are reported. We run the evaluation 20 times and report the average metric value. The white-box designation of KPG evaluation makes it possible to evaluate the different aspects of the models. Meanwhile, KPG is considered an easier task than conventional text-based motion generation since it is targeted at much less complex action semantics. Results show that previous methods are not performing as well as expected. Also, R-Precision fails to model the performance on KPG with trivially low values. Most methods perform well for atomic prompts, but when it comes to compositions, the accuracy drops. Early efforts like HMDM [48] and MotionDiffuse [55] are surprisingly competitive, even surpassing several more recent works. T2M-GPT [54] outperforms other methods by a considerable margin, but it fails for repetitive prompts. An unexpected" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 100 + ], + "type": "text", + "content": "Kinematic Phrases" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 169, + 114, + 446, + 199 + ], + "blocks": [ + { + "bbox": [ + 169, + 114, + 446, + 199 + ], + "lines": [ + { + "bbox": [ + 169, + 114, + 446, + 199 + ], + "spans": [ + { + "bbox": [ + 169, + 114, + 446, + 199 + ], + "type": "table", + "html": "
MethodsAccuracyDiversityR-P@1
AtomicRepetitiveSequentialSimultaneousOverall
HMDM [48]93.2523.9856.1130.9045.165.9230.034
MotionDiffuse [55]93.2122.1551.5427.7641.495.2570.051
T2M-GPT [54]97.221.0270.2442.4755.596.1530.042
MLD [7]92.0622.7648.2628.0139.945.5260.061
ReMoDiffuse [56]53.173.0532.1631.0630.565.8090.048
MotionGPT [23]92.0616.8755.9831.7644.965.7760.039
MoMask [13]96.4328.8659.1032.6647.795.9650.060
Ours98.8026.0271.3243.0857.866.0480.053
", + "image_path": "4dfd63441d54cd627ed44ef79e9a1bf94e6bf9dc1c423a75061a7229cf06e34f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 200, + 474, + 211 + ], + "lines": [ + { + "bbox": [ + 138, + 200, + 474, + 211 + ], + "spans": [ + { + "bbox": [ + 138, + 200, + 474, + 211 + ], + "type": "text", + "content": "Table 2: Results on KPG demonstrate the various strengths of different methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 247, + 481, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 481, + 319 + ], + "type": "text", + "content": "phenomenon is the failure of ReMoDiffuse [56], which might be constrained by its retrieval-based methodology since most prompts in KPG are not available in the retrieval database. Our method, which replaces the latent space of MLD [7], delivers substantial improvements, but the accuracy remains below " + }, + { + "bbox": [ + 130, + 247, + 481, + 319 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 247, + 481, + 319 + ], + "type": "text", + "content": ", which is far from satisfying. A considerable gap exists between existing efforts and ideal motion generation models. More analyses are in the appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 341, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 544 + ], + "type": "text", + "content": "User Study. To reach a convincing performance comparison, we conduct user studies on both HumanML3D and KPG. Our user study is different from previous efforts in two aspects. First, instead of testing a small set of text prompts (less than 50 in previous works [7,48]), we randomly select 600 sentences from the HumanML3D test set. By scaling up, the result is convincing in reflecting the ability to generate motion for diverse text inputs. Second, without asking the volunteers to give a general rating for each sample or to choose between different samples, we ask them two questions: 1) Do the motion and the text match? and 2) Is the motion natural? For Q1, three choices are given as \"No, Partially, Yes\". For Q2, two choices are given as \"Yes, No\". In this way, we explicitly evaluate two aspects of text-to-motion tasks: semantic consistency and naturalness, corresponding to R-Precision and FID. For each prompt, we generate one sample considering the annotation cost. We claim that the models should generate natural text-matching motion most of the time so that the one-sample setting would not hurt the fidelity of our user study. 36 volunteers are invited, each reviewing 200 sequences. Thus each sequence receives 3 user reviews. Also, we compute R-precision@1 of the generated sequences." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": "Besides our model, we select MDM [48] as a naive baseline, T2M-GPT [54] as the SOTA model given the results in Tab. 1 and Tab. 2. MLD [7] is also selected as a direct baseline of our proposed method. Results on HumanML3D [14] are shown in Fig. 7. Though our method is not superior in R-Precision, we receive competitive user reviews, showcasing the efficacy of our motion-KP joint space. Also please notice that our model only contains 45.1M parameters, while T2M-GPT contains 228M parameters. T2M-GPT and MLD present similar R-Precision, but only T2M-GPT manages to keep a good performance with user reviews. Moreover, the discrepancy between R-Precision and user reviews is revealed in both absolute value and trends, also in Fig. 6." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 139, + 113, + 299, + 210 + ], + "blocks": [ + { + "bbox": [ + 139, + 113, + 299, + 210 + ], + "lines": [ + { + "bbox": [ + 139, + 113, + 299, + 210 + ], + "spans": [ + { + "bbox": [ + 139, + 113, + 299, + 210 + ], + "type": "image", + "image_path": "e3fac0ccfdc387fd86d7b6828aa4ad633b3d6f5bbc3d86281ea1108948729ee4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 217, + 483, + 243 + ], + "lines": [ + { + "bbox": [ + 130, + 217, + 483, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 483, + 243 + ], + "type": "text", + "content": "Fig. 7: User study on HumanML3D, with Fig. 8: User study on KPG, with \"Y\" for \"Y\" for Yes and \"P\" for partially. Yes and \"P\" for partially." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 316, + 116, + 476, + 209 + ], + "blocks": [ + { + "bbox": [ + 316, + 116, + 476, + 209 + ], + "lines": [ + { + "bbox": [ + 316, + 116, + 476, + 209 + ], + "spans": [ + { + "bbox": [ + 316, + 116, + 476, + 209 + ], + "type": "image", + "image_path": "2337b3ab3bcc685556299264bc486c0cef254659834cd7dd6fcf0362ba9726b8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 138, + 252, + 244, + 352 + ], + "blocks": [ + { + "bbox": [ + 138, + 252, + 244, + 352 + ], + "lines": [ + { + "bbox": [ + 138, + 252, + 244, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 252, + 244, + 352 + ], + "type": "image", + "image_path": "550a8c39304d5e1005aa4a2da9d188b0cf6f100557600607c7695ad46a76d9ff.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 358, + 249, + 392 + ], + "lines": [ + { + "bbox": [ + 130, + 358, + 249, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 249, + 392 + ], + "type": "text", + "content": "Fig.9: Our model supports fine-grained modification on motion via KP." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 261, + 255, + 471, + 350 + ], + "blocks": [ + { + "bbox": [ + 261, + 255, + 471, + 350 + ], + "lines": [ + { + "bbox": [ + 261, + 255, + 471, + 350 + ], + "spans": [ + { + "bbox": [ + 261, + 255, + 471, + 350 + ], + "type": "image", + "image_path": "99d5e32a546234b54428d7d13d7545aff8c409e2bfff36f076bae8af62106c73.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 356, + 482, + 390 + ], + "lines": [ + { + "bbox": [ + 249, + 356, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 249, + 356, + 482, + 390 + ], + "type": "text", + "content": "Fig. 10: Visualization of generated samples. Compared to T2M-GPT, our method provides a better response to prompts with explicit constraints on specific body parts." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 417, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 482, + 489 + ], + "type": "text", + "content": "A similar user study on KPG is conducted in Fig. 7 with 100 randomly selected prompts from KPG involving T2M-GPT and our model. Fig. 8 demonstrates that KP-inferred Accuracy and user reviews share similar trends. Meanwhile, KP and user study give the same reviews for " + }, + { + "bbox": [ + 130, + 417, + 482, + 489 + ], + "type": "inline_equation", + "content": "84\\%" + }, + { + "bbox": [ + 130, + 417, + 482, + 489 + ], + "type": "text", + "content": " of the samples. We believe KPG could thus be a first step towards white-box automatic motion generation evaluation for semantic consistency. More analyses are in the appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 507, + 227, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 507, + 227, + 519 + ], + "spans": [ + { + "bbox": [ + 131, + 507, + 227, + 519 + ], + "type": "text", + "content": "5.3 Visualization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 528, + 482, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 482, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 482, + 613 + ], + "type": "text", + "content": "A modification sample is shown in Fig. 9. By modifying KP, we could edit arbitrary motion at a fine-grained level. Fig. 10 compares generated samples of T2M-GPT and our methods. Our method properly responds to text prompts with constraints on specific body parts thanks to KP, which explicitly models the action semantics with body part kinematics cues. Note that T2M-GPT generates redundant motion for simple prompts, while our method provides more concise and precise results. More visualizations are in the appendix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 631, + 246, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 631, + 246, + 642 + ], + "spans": [ + { + "bbox": [ + 131, + 631, + 246, + 642 + ], + "type": "text", + "content": "5.4 Ablation Studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 651, + 317, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 651, + 317, + 662 + ], + "spans": [ + { + "bbox": [ + 131, + 651, + 317, + 662 + ], + "type": "text", + "content": "KPG ablation study results are in Tab. 3." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 100 + ], + "type": "text", + "content": "Kinematic Phrases" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 332, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 332, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 332, + 163 + ], + "type": "text", + "content": "KP latent. By using a latent space without KP, a substantial performance drop is perceived, revealing the significance of KP in conveying action semantics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 332, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 332, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 332, + 236 + ], + "type": "text", + "content": "Different KP sets. We examine the contribution of different KP sets: joint KP (PP), joint pair KP (PRPP, PDP), limb KP (LAP, LOP), and body KP (GVP). A leave-one-out evaluation shows the elimination of joint KP and joint pair KP results in notable perfor" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 437, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 437, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 437, + 247 + ], + "type": "text", + "content": "mance degradation, while the influence of the rest is relatively subtle." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 342, + 118, + 482, + 191 + ], + "blocks": [ + { + "bbox": [ + 342, + 118, + 482, + 191 + ], + "lines": [ + { + "bbox": [ + 342, + 118, + 482, + 191 + ], + "spans": [ + { + "bbox": [ + 342, + 118, + 482, + 191 + ], + "type": "table", + "html": "
MethodsAcc. %↑Diversity
Ours57.866.048
w/o KP39.945.526
w/o Joint KP50.035.685
w/o Joint Pair KP47.245.772
w/o Limb KP55.925.934
w/o Body KP56.845.871
", + "image_path": "adf314bf9dce14b369b197017adac4a819c781e0a7e9f59bfd62329f9f00b0f3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 344, + 198, + 476, + 209 + ], + "lines": [ + { + "bbox": [ + 344, + 198, + 476, + 209 + ], + "spans": [ + { + "bbox": [ + 344, + 198, + 476, + 209 + ], + "type": "text", + "content": "Table 3: KPG ablation results." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 131, + 266, + 217, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 266, + 217, + 278 + ], + "spans": [ + { + "bbox": [ + 131, + 266, + 217, + 278 + ], + "type": "text", + "content": "6 Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 292, + 482, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 482, + 446 + ], + "type": "text", + "content": "Limitations. First, the current KP criteria guarantee objectivity but overlook kinematic information like amplitude and speed. Also, due to the granularity of the adopted skeleton, fine-grained information on fingers is not well-preserved. These would be promising for further exploration. Second, KPB could be extended to datasets with other modalities, like 2D pose and egocentric action datasets. Although these modalities provide incomplete 3D information, we could extract KP that is credibly accessible across modalities. Third, with the convenient KP-to-text conversion, auxiliary motion descriptions could be automatically generated via KP. Fourth, KPG could be extended by paraphrasing existing prompts and combining different Phrases. Finally, the proposed motion generation pipeline is devised for quick evaluation, while the motion-KP latent space could be incorporated with more advanced algorithms for better exploitation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 465, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 465, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 465, + 482, + 525 + ], + "type": "text", + "content": "Broader Impacts. The proposed KP and KPG provide a white-box instrument to diagnose text-to-motion algorithms, which could help develop reliable motion tools. Better motion synthesis could advance fields including animation production, sports training, human-robot interaction, and rehabilitation. However, it could also be used for illegal media content or false information." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 544, + 220, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 544, + 220, + 556 + ], + "spans": [ + { + "bbox": [ + 131, + 544, + 220, + 556 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "In this paper, we proposed an intermediate representation to bridge human motion and action semantics as the Kinematic Phrase. KP achieved proper abstraction, interpretability, and generality by focusing on objective kinematic facts of human motion. A motion understanding system based on KP was proposed and proven effective in motion interpolation, modification, and generation. Moreover, Kinematic Prompt Generation is proposed as a white-box text-to-motion benchmark, bringing new insights to the literature. We believe that KP has great potential to advance motion understanding." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 190 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 190 + ], + "type": "text", + "content": "This work is supported in part by the National Natural Science Foundation of China under Grants No.62306175, No.62302296, National Key Research and Development Project of China (No.2022ZD0160102, No.2021ZD0110704), Shanghai Artificial Intelligence Laboratory, XPLORER PRIZE grants." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 209, + 197, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 209, + 197, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 209, + 197, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 234, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 234, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 481, + 277 + ], + "type": "text", + "content": "1. Aristidou, A., Cohen-Or, D., Hodgins, J.K., Chrysanthou, Y., Shamir, A.: Deep motifs and motion signatures. ACM Trans. Graph. 37(6), 187:1-187:13 (Nov 2018). https://doi.org/10.1145/3272127.3275038, http://doi.acm.org/10.1145/3272127.3275038" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 280, + 481, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 280, + 481, + 312 + ], + "spans": [ + { + "bbox": [ + 138, + 280, + 481, + 312 + ], + "type": "text", + "content": "2. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: Teach: Temporal action composition for 3d humans. In: 2022 International Conference on 3D Vision (3DV). pp. 414-423. IEEE (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 312, + 481, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 481, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 481, + 334 + ], + "type": "text", + "content": "3. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: SINC: Spatial composition of 3D human motions for simultaneous action generation. In: ICCV (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 335, + 481, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 481, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 481, + 378 + ], + "type": "text", + "content": "4. Azadi, S., Shah, A., Hayes, T., Parikh, D., Gupta, S.: Make-an-animation: Large-scale text-conditional 3d human motion generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 15039-15048 (October 2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 379, + 481, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 379, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 379, + 481, + 411 + ], + "type": "text", + "content": "5. Bartlett, R.: Introduction to Sports Biomechanics. Introduction to Sports Biomechanics, E & FN Spon (1997), https://books.google.com.tw/books?id=-6Db8mgxsqQC" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 411, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 481, + 434 + ], + "type": "text", + "content": "6. Brégier, R.: Deep regression on manifolds: a 3d rotation case study. In: 2021 International Conference on 3D Vision (3DV). pp. 166-174. IEEE (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 434, + 481, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 481, + 466 + ], + "type": "text", + "content": "7. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18000-18010 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 468, + 481, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 481, + 510 + ], + "type": "text", + "content": "8. Dabral, R., Mughal, M.H., Golyanik, V., Theobalt, C.: Mofusion: A framework for denoising-diffusion-based motion synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9760-9770 (June 2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 511, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 511, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 511, + 481, + 555 + ], + "type": "text", + "content": "9. Delmas, G., Weinzaepfel, P., Lucas, T., Moreno-Noguer, F., Rogez, G.: Posescript: 3d human poses from natural language. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI. pp. 346-362. Springer (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 555, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 577 + ], + "type": "text", + "content": "0. Delmas, G., Weinzaepfel, P., Moreno-Noguer, F., Rogez, G.: Posefix: Correcting 3d human poses with natural language. arXiv preprint arXiv:2309.08480 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 578, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 481, + 609 + ], + "type": "text", + "content": "1. Fang, H.S., Li, J., Tang, H., Xu, C., Zhu, H., Xiu, Y., Li, Y.L., Lu, C.: Alpha-posed: Whole-body regional multi-person pose estimation and tracking in real-time. TPAMI (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 610, + 481, + 643 + ], + "type": "text", + "content": "12. Gao, X., Yang, Y., Xie, Z., Du, S., Sun, Z., Wu, Y.: Guess: Gradually enriching synthesis for text-driven human motion generation. IEEE Transactions on Visualization and Computer Graphics (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "text", + "content": "13. Guo, C., Mu, Y., Javed, M.G., Wang, S., Cheng, L.: Momask: Generative masked modeling of 3d human motions. arXiv preprint arXiv:2312.00063 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 100 + ], + "type": "text", + "content": "Kinematic Phrases" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "14. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5152-5161 (June 2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "type": "text", + "content": "15. Guo, C., Zuo, X., Wang, S., Cheng, L.: Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In: European Conference on Computer Vision. pp. 580-597. Springer (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 481, + 226 + ], + "type": "text", + "content": "16. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2021-2029 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "type": "text", + "content": "17. Guo, W., Du, Y., Shen, X., Lepetit, V., Alameda-Pineda, X., Moreno-Noguer, F.: Back to mlp: A simple baseline for human motion prediction. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4809-4819 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "text", + "content": "18. Hassan, M., Ceylan, D., Villegas, R., Saito, J., Yang, J., Zhou, Y., Black, M.: Stochastic scene-aware motion prediction. In: Proceedings of the International Conference on Computer Vision 2021 (Oct 2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "text", + "content": "19. Hernandez, A., Gall, J., Moreno-Noguer, F.: Human motion prediction via spatiotemporal inpainting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7134-7143 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 337, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 481, + 357 + ], + "type": "text", + "content": "20. Holden, D., Kanoun, O., Perepichka, M., Popa, T.: Learned motion matching. ACM Transactions on Graphics (TOG) 39(4), 53-1 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "text", + "content": "21. Hong, F., Zhang, M., Pan, L., Cai, Z., Yang, L., Liu, Z.: Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 391, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 481, + 424 + ], + "type": "text", + "content": "22. Ji, Y., Xu, F., Yang, Y., Shen, F., Shen, H.T., Zheng, W.S.: A large-scale rgb-d database for arbitrary-view human action recognition. In: Proceedings of the 26th ACM international Conference on Multimedia. pp. 1510-1518 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 425, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 481, + 445 + ], + "type": "text", + "content": "23. Jiang, B., Chen, X., Liu, W., Yu, J., Yu, G., Chen, T.: Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "text", + "content": "24. Jin, P., Wu, Y., Fan, Y., Sun, Z., Yang, W., Yuan, L.: Act as you wish: Fine-grained control of motion diffusion model with hierarchical semantic graphs. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 522 + ], + "type": "text", + "content": "25. Karunratanakul, K., Preechakul, K., Suwajanakorn, S., Tang, S.: Guided motion diffusion for controllable human motion synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2151-2162 (October 2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "text", + "content": "26. Kong, H., Gong, K., Lian, D., Mi, M.B., Wang, X.: Priority-centric human motion generation in discrete latent space. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14806-14816 (October 2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 556, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 481, + 577 + ], + "type": "text", + "content": "27. Koppula, H.S., Saxena, A.: Anticipating human activities for reactive robotic response. In: IROS. p. 2071. Tokyo (2013)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 578, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 609 + ], + "type": "text", + "content": "28. von Laban, R., Lange, R.: Laban's Principles of Dance and Movement Notation. Macdonald & Evans (1975), https://books.google.com.tw/books?id=VrOAAAAAMAAJ" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "text", + "content": "29. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13401-13412 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "text", + "content": "30. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Learn to dance with aist++: Music conditioned 3d dance generation (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "31. Li, Y.L., Wu, X., Liu, X., Wang, Z., Dou, Y., Ji, Y., Zhang, J., Li, Y., Lu, X., Tan, J., et al.: From isolated islands to pangea: Unifying semantic space for human action understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16582-16592 (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 482, + 205 + ], + "type": "text", + "content": "32. Lin, J., Chang, J., Liu, L., Li, G., Lin, L., Tian, Q., Chen, C.W.: Being comes from not-being: Open-vocabulary text-to-motion generation with wordless training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 23222-23231 (June 2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "text", + "content": "33. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM transactions on graphics (TOG) 34(6), 1-16 (2015)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "text", + "content": "34. Lou, Y., Zhu, L., Wang, Y., Wang, X., Yang, Y.: Diversemotion: Towards diverse human motion generation via discrete diffusion. arXiv preprint arXiv:2309.01372 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 270, + 482, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 314 + ], + "type": "text", + "content": "35. Lucas, T., Baradel, F., Weinzaepfel, P., Rogez, G.: Posegpt: quantization-based 3d human motion generation and forecasting. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI. pp. 417-435. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "type": "text", + "content": "36. Paden, B., Cap, M., Yong, S.Z., Yershov, D., Frazzoli, E.: A survey of motion planning and control techniques for self-driving urban vehicles. IEEE Transactions on intelligent vehicles 1(1), 33-55 (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 347, + 482, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 392 + ], + "type": "text", + "content": "37. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3D hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 10975-10985 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 392, + 482, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 424 + ], + "type": "text", + "content": "38. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10985-10995 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 424, + 482, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 468 + ], + "type": "text", + "content": "39. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII. pp. 480-497. Springer (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "type": "text", + "content": "40. Petrovich, M., Black, M.J., Varol, G.: Tmr: Text-to-motion retrieval using contrastive 3d human motion synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9488-9497 (October 2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 501, + 482, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 522 + ], + "type": "text", + "content": "41. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data 4(4), 236-252 (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "type": "text", + "content": "42. Pons-Moll, G., Fleet, D.J., Rosenhahn, B.: Posebits for monocular human pose estimation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2337-2344 (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 555, + 482, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 599 + ], + "type": "text", + "content": "43. Punnakkal, A.R., Chandrasekaran, A., Athanasiou, N., Quiros-Ramirez, A., Black, M.J.: Babel: bodies, action and behavior with english labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 722-731 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 600, + 482, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 600, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 600, + 482, + 643 + ], + "type": "text", + "content": "44. Qian, Y., Urbanek, J., Hauptmann, A.G., Won, J.: Breaking the limits of text-conditioned 3d motion synthesis with elaborative descriptions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2306-2316 (October 2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 643, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 482, + 665 + ], + "type": "text", + "content": "45. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 100 + ], + "type": "text", + "content": "Kinematic Phrases" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 147, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 117, + 480, + 138 + ], + "type": "text", + "content": "natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 481, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 481, + 170 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 481, + 170 + ], + "type": "text", + "content": "46. Tang, Y., Liu, J., Liu, A., Yang, B., Dai, W., Rao, Y., Lu, J., Zhou, J., Li, X.: Flag3d: A 3d fitness activity dataset with language instruction. arXiv preprint arXiv:2212.04638 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 481, + 214 + ], + "type": "text", + "content": "47. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII. pp. 358-374. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 481, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 481, + 235 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 481, + 235 + ], + "type": "text", + "content": "48. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 235, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 235, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 133, + 235, + 481, + 267 + ], + "type": "text", + "content": "49. Van Welbergen, H., Van Basten, B.J., Egges, A., Ruttkay, Z.M., Overmars, M.H.: Real time animation of virtual humans: a trade-off between naturalness and control. In: Computer Graphics Forum. vol. 29, pp. 2530-2554. Wiley Online Library (2010)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 267, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 481, + 310 + ], + "type": "text", + "content": "50. Wang, Y., Leng, Z., Li, F.W.B., Wu, S.C., Liang, X.: Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 22035-22044 (October 2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 310, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 481, + 331 + ], + "type": "text", + "content": "51. Xiang, W., Li, C., Zhou, Y., Wang, B., Zhang, L.: Language supervised training for skeleton-based action recognition. arXiv preprint arXiv:2208.05318 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 331, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 481, + 385 + ], + "type": "text", + "content": "52. Xu, L., Song, Z., Wang, D., Su, J., Fang, Z., Ding, C., Gan, W., Yan, Y., Jin, X., Yang, X., Zeng, W., Wu, W.: Actformer: A gan-based transformer towards general action-conditioned 3d human motion generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2228-2238 (October 2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 385, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 481, + 418 + ], + "type": "text", + "content": "53. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 16010-16021 (October 2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "type": "text", + "content": "54. Zhang, J., Zhang, Y., Cun, X., Huang, S., Zhang, Y., Zhao, H., Lu, H., Shen, X.: T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "type": "text", + "content": "55. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 482, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 481, + 525 + ], + "type": "text", + "content": "56. Zhang, M., Guo, X., Pan, L., Cai, Z., Hong, F., Li, H., Yang, L., Liu, Z.: Remodiffuse: Retrieval-augmented motion diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 364-373 (October 2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "text", + "content": "57. Zhang, M., Li, H., Cai, Z., Ren, J., Yang, L., Liu, Z.: Finemogen: Fine-grained spatio-temporal motion generation and editing. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 557, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 481, + 601 + ], + "type": "text", + "content": "58. Zhong, C., Hu, L., Zhang, Z., Xia, S.: Attt2m: Text-driven human motion generation with multi-perspective attention mechanism. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 509-519 (October 2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "text", + "content": "59. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5745-5753 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "text", + "content": "60. Zhou, Z., Wan, Y., Wang, B.: Avatargpt: All-in-one framework for motion understanding, planning, generation and beyond. arXiv preprint arXiv:2311.16468 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 150 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 150 + ], + "type": "text", + "content": "61. Zhou, Z., Wang, B.: Ude: A unified driving engine for human motion generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5632-5641 (June 2023)" + } + ] + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 370, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 370, + 91, + 447, + 100 + ], + "type": "text", + "content": "Kinematic Phrases" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Bridging the Gap_ Studio-like Avatar Creation from a Monocular Phone Capture/3e2c48bd-a83b-478a-9094-3e6d6e7a5043_content_list.json b/2024/Bridging the Gap_ Studio-like Avatar Creation from a Monocular Phone Capture/3e2c48bd-a83b-478a-9094-3e6d6e7a5043_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bc79aa54fd32a76b0c234541e9544337d70dda84 --- /dev/null +++ b/2024/Bridging the Gap_ Studio-like Avatar Creation from a Monocular Phone Capture/3e2c48bd-a83b-478a-9094-3e6d6e7a5043_content_list.json @@ -0,0 +1,1868 @@ +[ + { + "type": "text", + "text": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture", + "text_level": 1, + "bbox": [ + 228, + 140, + 774, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ShahRukh Athar $^{1,3*}$ Shunsuke Saito $^{2}$ Zhengyu Yang $^{2}$ Stanislav Pidhorskyi $^{2}$ Chen Cao $^{2}$", + "bbox": [ + 225, + 210, + 779, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Captions Research, New York", + "bbox": [ + 276, + 256, + 483, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Meta Reality Labs, Pittsburgh.", + "bbox": [ + 517, + 257, + 728, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stony Brook University, New York", + "bbox": [ + 397, + 287, + 638, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "shahrukh@nocapinc.com shunsuke.saito16@gmail.com", + "bbox": [ + 282, + 316, + 720, + 329 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "stpidhorskyi@meta.com zhengyu-yang@outlook.com", + "bbox": [ + 290, + 330, + 710, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zju.caochen@gmail.com", + "bbox": [ + 408, + 345, + 594, + 357 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/60cb2a326479dbe9b1e096bc14e396f5926296187b5943feb1e0de334faca02f.jpg", + "image_caption": [ + "Monocular phone capture" + ], + "image_footnote": [], + "bbox": [ + 225, + 422, + 289, + 477 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d82b72dd12f0a0880c5ab26afa11f26e8527e6235cb2e0f3ec5e81c46c63d462.jpg", + "image_caption": [ + "Fig. 1: From a low-resolution texture map obtained through monocular phone capture, our method produces a high-resolution texture map with precise facial details, studio-like illumination, and inpainted missing regions. This generated texture map can subsequently be utilized to create a high-quality, photorealistic avatar using the pretrained Universal Prior Model (UPM) from AVA [6]." + ], + "image_footnote": [], + "bbox": [ + 225, + 479, + 289, + 535 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/879fe4f17527076281cf808d47c10465417d11b6df9f5440ada1799aa717b6e6.jpg", + "image_caption": [ + "Extracted texture map" + ], + "image_footnote": [], + "bbox": [ + 294, + 422, + 439, + 534 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/47983b98dab4ec58dafc3d4015ba511b8b71652884c6895a8633b300568e62d4.jpg", + "image_caption": [ + "Generated highresolution texture n" + ], + "image_footnote": [], + "bbox": [ + 462, + 422, + 606, + 534 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ef2231891fc8d89d00f972c47097edf4d01c5112307364c2df8d0f24b419adc9.jpg", + "image_caption": [ + "High-quality reanimation and novel view synthesis" + ], + "image_footnote": [], + "bbox": [ + 674, + 422, + 764, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Creating photorealistic avatars for individuals traditionally involves extensive capture sessions with complex and expensive studio devices like the LightStage system. While recent strides in neural representations have enabled the generation of photorealistic and animatable 3D avatars from quick phone scans, they have the capture-time lighting baked-in, lack facial details and have missing regions in areas such as the back of the ears. Thus, they lag in quality compared to studio-captured avatars. In this paper, we propose a method that bridges this gap by generating studio-like illuminated texture maps from short, monocular phone captures. We do this by parameterizing the phone texture maps using the $W^{+}$ space of a StyleGAN2, enabling near-perfect reconstruction. Then, we finetune a StyleGAN2 by sampling in the $W^{+}$ parameterized space using a very small set of studio-captured textures as an adversarial training signal. To further enhance the realism and accuracy of facial details, we super-resolve", + "bbox": [ + 259, + 636, + 743, + 816 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "*Work done while interning at Meta Reality Labs", + "bbox": [ + 215, + 825, + 511, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the output of the StyleGAN2 using carefully designed diffusion model that is guided by image gradients of the phone-captured texture map. Once trained, our method excels at producing studio-like facial texture maps from casual monocular smartphone videos. Demonstrating its capabilities, we showcase the generation of photorealistic, uniformly lit, complete avatars from monocular phone captures.", + "bbox": [ + 261, + 147, + 743, + 218 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 261, + 356, + 277 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Photorealistic and animatable avatars are paramount for lifelike human-to-human interactions in AR/VR applications. Creating high-fidelity avatars often requires sophisticated devices like the LightStage capture system that capture human heads with a range of facial expressions [27, 28]. Typically, these captures occur in a studio-like environment, with uniform illumination and densely sampled views to reconstruct complete avatars with consistent lighting. While these methods excel in generating hyper-realistic avatars, they cannot be scaled to millions of people as capturing so many people with a LightStage-like capture system is impractical and non-trivial.", + "bbox": [ + 212, + 295, + 787, + 415 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, there has been extensive work in generating photorealistic avatars through a monocular capture [3,6,14]. These methods involve scanning various facial expressions and head poses of the user to reconstruct a 3D avatar that closely aligns with the captured data. However, a notable limitation of these methods is that the captured lighting is embedded in the avatars, causing their appearance to be heavily dependent on the capture devices and surrounding environments. Furthermore, due to the constraints of the capture setup, certain areas, such as the back of the head or ears, are never visible, leaving visible holes and artifacts when viewed during animation.", + "bbox": [ + 212, + 416, + 787, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "An innovative strategy to tackle this challenge is to use image-to-image translation to transform the phone data into studio-captured data. This transformation can be learnt through supervised training using paired data [18], or through unsupervised training using unpaired datasets [42]. Due to reasons mentioned earlier, creating a large-scale paired dataset of studio and phone captures is impractical, which rules out the possibility of using supervised image-to-image translation methods. On the flip side, current unsupervised image-to-image translation methods fail to preserve fine detail in the transformation process [38, 42] which is paramount to creating a high-fidelity avatar.", + "bbox": [ + 212, + 537, + 789, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce a method capable of generating studio-like, high-quality avatars from monocular phone captures. We do this by parameterizing a large-scale dataset of phone-captured face texture maps using the $W^{+}$ space of a StyleGAN2. Then, we finetune this StyleGAN2 by sampling from this parametrized $W^{+}$ space using a small set of unpaired studio-captured texture maps, to create a Studio-StyleGAN2 model. Our key insight is that sampling from the $W^{+}$ space instead of Z space leads to a more generalizable model as samples from the $W^{+}$ are, by construction, as diverse as the training set while samples from the Z space often suffer from mode-collapse. During inference, the given phone-captured texture map, is first inverted to the $W^{+}$ space of the StyleGAN2. This inverted $W^{+}$ vector is then given as input to the Studio-StyleGAN2 to generate a low-resolution studio-like texture map. Finally, a novel facial detail conditioned diffusion model is used to enhance facial details of the low-resolution", + "bbox": [ + 212, + 659, + 789, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Athar et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "studio-like texture map obtained in the previous step. Once trained, our method excels at producing high-quality studio-like face texture maps from monocular phone captures, which are then used as inputs to the universal face prior model (UPM) from Authentic Volumetric Avatars (AVA) [6] to generate photorealistic avatars. In summary, our key contributions are as follows:", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Introducing a groundbreaking method for creating studio-like, photorealistic avatars from monocular phone captures.", + "- Innovative finetuning of a pre-trained generative model using a minimal dataset from another domain by sampling in the inverted $W^{+}$ space, which facilitates the development of a generative model for a new domain while preserving the integrity of the $W^{+}$ latent space.", + "- A novel diffusion model conditioned on the phone texture gradient, that is designed to upsample studio textures, effectively enhancing facial details and contributing to the overall realism of the generated avatars." + ], + "bbox": [ + 225, + 229, + 784, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 387, + 366, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our paper aims to bridge the gap between a studio capture and a phone capture by creating a studio-like, photorealistic avatar through a monocular phone capture. In delineating the related work, we offer a comprehensive overview of key research domains, including studio-captured avatars, phone-captured avatars, and image-to-image translation.", + "bbox": [ + 212, + 417, + 785, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Studio-captured avatars. The reconstruction of high-fidelity static and dynamic models of the human head based on photometric measurements has a long-standing history in computer graphics and vision. Achieving a photorealistic human avatar often requires specialized hardware in high-end production, such as the LightStage system. To model the complex skin appearance, various physically-based models have been explored. Notably, subsurface scattering [5], linear polarization patterns [12] and fine-scale skin details [1, 2] have been investigated. For dynamic expression details, Jimenez et al. [20] compute dynamic skin appearances by blending hemoglobin distributions captured with different expressions. In their subsequent work [19], expression-dependent normal maps are interpolated to add realistic wrinkles to an animated face. Nagano et al. [32] synthesize skin microstructures based on local geometric features derived from high-precision microgeometry, acquired with an LED sphere and a skin deformer. While these methods have been instrumental for offline movie production, their substantial compute requirements make them less suitable for real-time applications. Despite recent efforts to enable real-time rendering of physically-based avatars [36], heavy compute remains challenge. In response to the challenges posed by complex physical computations, researchers have proposed a deep appearance model [27]. This model utilizes a coarse 3D triangle mesh in conjunction with view-dependent texture mapping. The texture is regressed by a neural network conditioned on viewpoint and expression latent codes. This conditioning accounts for view- and expression-dependent variations while compensating for the imperfect proxy geometry. Subsequent work extends the mesh-texture representation to a volumetric representation using a Mixture of Volumetric Primitives (MVP) [28], further enhancing the model's quality. Pixel Codec Avatars (PiCA), as demonstrated by [30], showcase the efficiency of rendering such models, even on mobile hardware platforms,", + "bbox": [ + 212, + 478, + 787, + 843 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture", + "bbox": [ + 256, + 114, + 730, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "by leveraging efficient per-pixel processing. Moreover, Bi et al. [4], based on relighting captured data, can also relight avatars with any novel point light or environment maps. While these methods can achieve hyper-realistic avatars, their studio requirements make them challenging to generalize to ordinary users.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Phone-captured avatars. There are several methods aimed at creating avatars in lightweight ways, even from a phone or a single image. Some of these approaches focus on generating stylized avatars based on different input requirements, ranging from capturing multi-view images using a phone [17] to utilizing a single monocular image [26, 29, 35, 37]. While these methods excel in producing animatable avatars, it is important to note that their appearance tends to be cartoonish and lacks realism.", + "bbox": [ + 212, + 222, + 787, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another category of methods is dedicated to creating realistic human avatars based on graphics pipelines [7, 8, 11, 15, 24, 25, 41]. While relying on traditional graphics methods, these approaches often result in avatars that appear uncanny. Building upon deep learning-based representations, researchers have proposed methods to generate avatars with increased realism, including notable works by Gafni et al. [10] and Grassal et al. [14]. Among these, the work of Cao et al. [6], Authentic Volumetric Avatars (AVA), stands out for its focus on creating photorealistic avatars from phone scans. The process involves training a Universal Prior Model (UPM) using studio-captured data, followed by personalizing this UPM using data from a phone scan of an unseen subject. While the method successfully produces avatars with realistic appearance and animation, it is worth noting that the lighting is baked into the personalized avatar, and certain details, such as the back of the ears, may be missing.", + "bbox": [ + 212, + 327, + 789, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image-to-image translation We use image-to-image translation to map images from the source domain (phone data) to the target domain (studio data). Isola et al. introduced Pix2Pix [18], a method that utilizes adversarial training strategies [13] to achieve this mapping. Additionally, Wang et al. [40] focused on increasing the resolution of generated results from semantic label maps. While these methods successfully map images between domains, it is essential to note that they require paired training data for effective implementation. In many real-world scenarios, obtaining paired training data can be challenging and expensive.", + "bbox": [ + 212, + 523, + 789, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this issue, unsupervised image-to-image translation has been introduced. Zhu et al. [42] proposed a novel cycle-consistency loss to ensure that translating an image from one domain to another and back again should result in the original image. This helps the model maintain consistency and produce more realistic translations. Subsequent methods have further improved unsupervised image-to-image translation from different perspectives, including a multimodal model [16], few-shot input for video-to-video translation [39], translation of images with human control [34], and translation of real images into different styles [38]. However, none of these methods are designed with preservation of facial identity and the generation of realistic facial details in mind, especially with such little training data. Later in the paper we show how this prior work compares to ours for transforming low-resolution phone-captured texture maps to high-resolution studio-captured texture maps.", + "bbox": [ + 212, + 657, + 789, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Athar et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1e0379f9076c396cc9e0ae4f9f76d160176b7b478e0ba1284fb2e6ad8f5bfb87.jpg", + "image_caption": [ + "Fig. 2: Method Overview. Our method employs a two-step process to train $\\mathcal{G}_{\\text{Studio}}$ . Initially, we train a StyleGAN2 on $12k$ neutral texture maps captured by phones, yielding $\\mathcal{G}_{\\text{phone}}$ . Subsequently, we initialize $\\mathcal{G}_{\\text{Studio}}$ with the weights of $\\mathcal{G}_{\\text{phone}}$ and fine-tune it by sampling from $S_{W^{+}} = \\{W_{\\mathcal{I}_0^{\\text{phone}}}, \\ldots, W_{\\mathcal{I}_{N-1}^{\\text{phone}}}^{+}\\}$ ; where $N = 12k$ , and $W_{\\mathcal{I}_i^{\\text{phone}}}^{+}$ represents the vector obtained by inverting the $i$ 'th phone-captured texture map, $\\mathcal{I}_0^{\\text{phone}}$ , in the $W^{+}$ space of $\\mathcal{G}_{\\text{phone}}$ . During inference, the given phone-captured texture map is inverted to $W^{+}$ , then passed to $\\mathcal{G}_{\\text{Studio}}$ to generate a low-resolution studio-lit texture map, $\\mathcal{I}^{*}$ . Accurate facial details are subsequently added using diffusion model $f_{\\phi}$ , conditioned on the gradient of the phone texture. This process results in the final high-resolution, studio-lit, and completed texture map, $\\mathcal{I}^{*}$ ." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 580, + 286 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/31a1533d559db413d9f515be8a5f77c709c0ef57be5ddbed4b846be033f77875.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 143, + 790, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 460, + 316, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present our method for generating studio-like avatars from the phone captures. Our approach consists of two key components: a StyleGAN2 for texture translation and a diffusion model for facial detail generation. In Sect 3.1, we describe the generation of a low-resolution texture map with studio-like lighting and missing regions inpainted. First, we pretrain a StyleGAN2 on 12k phone textures and then finetune it using a small set of studio-captured texture maps. In order to improve the generalization of the finetuned StyleGAN2, we optimize it by sampling in the $W^{+}$ space, instead of the W-space or Z-space, using a set of 12k $W^{+}$ vectors obtained by inverting the phone captured texture maps. In Sect 3.2, we introduce a diffusion model that generates facial details. The diffusion model takes the output from the aforementioned StyleGAN2 and generates a high-resolution neutral texture with realistic facial details. After obtaining a high-resolution studio-like neutral texture from our method, we use it to learn a color transform to transfer phone-captured expression textures to studio-lit expression textures. These expressions are subsequently utilized for animating a high-quality avatar (Sect 3.3). Notably, we intentionally avoid applying inpainting or super-resolution techniques to the expression textures. Fig. 2 provides an overview of our method.", + "bbox": [ + 212, + 491, + 789, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Illumination manipulation and inpainting", + "text_level": 1, + "bbox": [ + 215, + 755, + 547, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our initial step, we track the geometry from a monocular phone capture of the user's neutral face and extract the neutral texture $\\mathcal{L}^{phone}$ from the captured image, employing the method outlined in [6]. Subsequently, we translate this phone-captured texture with in-the-wild lighting and missing regions into a texture map with studio lighting and", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture", + "bbox": [ + 256, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "missing regions in-painted. This translation is accomplished by parametrizing $\\mathcal{I}^{phone}$ using the $W^{+}$ space of a StyleGAN2, as follows:", + "bbox": [ + 215, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nW _ {\\mathcal {I} ^ {p h o n e}} ^ {+} = \\underset {W ^ {+}} {\\operatorname {a r g m i n}} \\left\\| \\mathcal {G} _ {p h o n e} \\left(W ^ {+}\\right) - \\mathcal {I} ^ {p h o n e} \\right\\| _ {2} ^ {2} + L P I P S \\left(\\mathcal {G} _ {p h o n e} \\left(W ^ {+}\\right), \\mathcal {I} ^ {p h o n e}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 267, + 196, + 787, + 219 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{G}_{phone}$ is a StyleGAN2 trained on phone-captured textures, and $W_{\\mathcal{T}^{phone}}^{+}$ is the optimized vector in the $W^{+}$ space of $\\mathcal{G}_{phone}$ . Now, we generate the low-resolution studio-like texture as follows:", + "bbox": [ + 214, + 229, + 782, + 273 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} ^ {*} = \\mathcal {G} _ {\\text {S t u d i o}} \\left(W _ {\\mathcal {I} _ {\\text {p h o n e}}} ^ {+}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 429, + 284, + 787, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{G}_{\\text{Studio}}$ is a StyleGAN2 responsible for generating low-resolution studio-like textures, while $\\mathcal{I}^*$ refers to the studio-lit version of $\\mathcal{I}^{\\text{phone}}$ with the inpainted missing regions. Ideally, $\\mathcal{I}^*$ should retain the identity and semantics of $\\mathcal{I}^{\\text{phone}}$ , with the only distinctions being the illumination and the inpainting of missing regions. Below we describe the training procedure for $\\mathcal{G}_{\\text{Studio}}$ .", + "bbox": [ + 214, + 308, + 782, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We initialize the synthesis network of $\\mathcal{G}_{\\text{Studio}}$ using weights from $\\mathcal{G}_{\\text{phone}}$ , which we then finetune to generate texture maps with studio lighting and inpaint the missing regions. We choose to finetune $\\mathcal{G}_{\\text{phone}}$ for the following reasons: 1) Due to the immense costs of a studio capture, studio quality training data is typically very limited. In our case, we only have 383 studio-captured texture maps, making training from scratch hard [22]; 2) Since we want $\\mathcal{I}^*$ and $\\mathcal{I}^{phone}$ to have the same facial identity and semantics meaning, it is beneficial to learn a latent space that is shared between $\\mathcal{G}_{\\text{Studio}}$ and $\\mathcal{G}_{\\text{phone}}$ . For finetuning, we first invert the entire dataset of $N$ phone-captured texture maps using Eq. (1), giving us a set of $N$ vectors in the $W^{+}$ space: $S_{W^{+}} = \\{W_{\\mathcal{I}_0^{phone}}^{+}, \\ldots, W_{\\mathcal{I}_{N-1}^{phone}}^{+}\\}$ . Next, we randomly sample vectors from the inverted set, $S_{W^{+}}$ , and finetune the synthesis network of $\\mathcal{G}_{\\text{phone}}$ to give $\\mathcal{G}_{\\text{Studio}}$ . We choose to finetune the generator with samples from inverted $W^{+}$ instead of the traditional $W$ space, since they are, by construction, from the real data distribution and are consequently more diverse, leading to better generalization of $\\mathcal{G}_{\\text{Studio}}$ . The losses we use during finetuning are described as below.", + "bbox": [ + 214, + 383, + 787, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Studio-Discriminator loss. This loss uses a discriminator to ensure the texture maps generated by $\\mathcal{G}_{\\text{Studio}}$ are from the distribution of studio-captured texture maps:", + "bbox": [ + 215, + 601, + 782, + 630 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {A d v} = \\mathbb {E} _ {\\mathcal {T S t u d i o} \\sim P (\\mathcal {T S t u d i o})} [ \\min (0, - 1 + \\mathcal {D} _ {\\text {S t u d i o}}) ] \\\\ + \\mathbb {E} _ {W _ {\\mathcal {I} _ {i} ^ {p h o n e}} ^ {+} \\sim S _ {W ^ {+}}} \\left[ \\min (0, - 1 - \\mathcal {D} _ {\\text {S t u d i o}} \\left(\\mathcal {G} _ {\\text {S t u d i o}} \\left(W _ {\\mathcal {T} _ {i} ^ {p h o n e}} ^ {+}\\right)\\right) \\right], \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 305, + 631, + 785, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{D}_{\\text{Studio}}$ is a discriminator that initialized from the pretrained phone-captured texture StyleGAN2 Discriminator and finetuned using ground-truth studio-captured textures and textures generated by $\\mathcal{G}_{\\text{Studio}}$ .", + "bbox": [ + 214, + 681, + 782, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Face-Identity Loss This loss compares the face identity embeddings of the renders of texture maps generated from $\\mathcal{G}_{\\text{Studio}}$ and $\\mathcal{G}_{\\text{phone}}$ in order to ensure the identity is preserved:", + "bbox": [ + 214, + 727, + 782, + 770 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {F a c e I D}} = \\left\\| \\left| \\mathcal {F} \\left(\\mathcal {G} _ {\\text {S t u d i o}} \\left(W _ {\\mathcal {I} _ {i} ^ {\\text {p h o n e}}} ^ {+}\\right)\\right) - \\mathcal {F} \\left(\\mathcal {G} _ {\\text {p h o n e}} \\left(W _ {\\mathcal {I} _ {i} ^ {\\text {p h o n e}}} ^ {+}\\right)\\right) \\right| \\right\\| _ {2} ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 312, + 770, + 787, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{F}$ is a pretrained face recognition network1.", + "bbox": [ + 214, + 801, + 549, + 818 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Athar et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1We use the network from here", + "bbox": [ + 215, + 823, + 401, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Perceptual Loss This loss encourages the preservation of semantic features between the texture maps generated $\\mathcal{G}_{\\text{Studio}}$ and $\\mathcal{G}_{\\text{phone}}$ by minimizing their distance in the VGG feature space", + "bbox": [ + 214, + 146, + 787, + 190 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {P e r c p}} = \\operatorname {L P I P S} \\left(\\mathcal {G} _ {\\text {S t u d i o}} \\left(W _ {\\mathcal {I} _ {i} ^ {\\text {p h o n e}}} ^ {+}\\right), \\mathcal {G} _ {\\text {p h o n e}} \\left(W _ {\\mathcal {I} _ {i} ^ {\\text {p h o n e}}} ^ {+}\\right)\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 183, + 787, + 212 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Perceptual Reconstruction Loss This loss ensures that skintones are preserved using a small amount of paired data (i.e. subjects for whom we have both phone and studio textures maps i.e both $\\mathcal{I}^{phone}$ and $\\mathcal{I}^{Studio}$ )", + "bbox": [ + 214, + 218, + 787, + 262 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {P e r c p - R e c o n s}} = \\frac {1}{k} \\sum_ {i = 1} ^ {K} \\operatorname {L P I P S} \\left(\\mathcal {G} _ {\\text {S t u d i o}} \\left(W _ {\\mathcal {T} _ {i} ^ {\\text {p h o n e}}} ^ {+}\\right), \\mathcal {I} _ {i} ^ {\\text {S t u d i o}}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 267, + 787, + 296 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{I}_i^{Studio}$ and $\\mathcal{I}_i^{phone}$ are the ground-truth studio and ground-truth phone texture maps of the same person. We only have a very small amount of this data for training (K=83). R1-Regularization: We regularize $\\mathcal{D}_{Studio}$ using the R1 regularization [31] as follows:", + "bbox": [ + 214, + 306, + 787, + 366 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {R 1} = - \\frac {\\gamma}{2} \\mathbb {E} _ {\\mathcal {I} ^ {S t u d i o} \\sim P (\\mathcal {I} ^ {S t u d i o})} \\left[ \\left\\| \\nabla \\mathcal {D} _ {S t u d i o} (\\mathcal {I} ^ {S t u d i o}) \\right\\| _ {2} ^ {2} \\right], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 363, + 787, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\gamma = 10$", + "bbox": [ + 214, + 386, + 316, + 400 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In order to further encourage identity preservation, we leverage the multi-resolution architecture of StyleGAN2 [23] and only optimize the generator parameters after the $8 \\times 8$ resolution during the finetuning process. The intuition being that identity-specific information is mostly stored in the low-resolution maps, we ablate this choice in the Supplementary section of the paper. The final optimization for the joint finetuning of $\\mathcal{G}_{\\text{Studio}}$ and $\\mathcal{D}_{\\text{Studio}}$ is the following:", + "bbox": [ + 214, + 402, + 787, + 492 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathcal {G} _ {S t u d i o} ^ {\\theta (8 +)}} \\max _ {\\mathcal {D} _ {S t u d i o}} \\mathcal {L} _ {A d v} + \\mathcal {L} _ {R 1} + \\mathcal {L} _ {P e r c p - R e c o n s} + \\lambda_ {1} \\mathcal {L} _ {P e r c p} + \\lambda_ {2} \\mathcal {L} _ {F a c e I D}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 299, + 500, + 787, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\lambda_{1} = 0.5$ and $\\lambda_{2} = 10$ . We ablate $\\mathcal{L}_{Percp-Recons}, \\mathcal{L}_{Percp}, \\mathcal{L}_{FaceID}$ in the supplementary.", + "bbox": [ + 214, + 534, + 787, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Synthesis of accurate facial details", + "text_level": 1, + "bbox": [ + 214, + 588, + 496, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generating high-quality facial details. $\\mathcal{G}_{\\text{Studio}}$ can produce an inpainted texture map, $\\mathcal{I}^*$ , with illumination consistent with that of a studio capture. However, due to the limited amount of finetuning data, $\\mathcal{G}_{\\text{Studio}}$ often struggles to reproduce facial details that are visibly present in $\\mathcal{I}^{\\text{phone}}$ itself. This limitation results in oversmoothed and inaccurate avatars. Achieving realistic facial details necessitates an accurate modeling of the facial texture distribution, a task that the StyleGAN2 fails to do.", + "bbox": [ + 214, + 616, + 787, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Motivated by the recent success of latent diffusion models in accurately modeling data distributions, we propose the task of generating facial details as the reverse process of a Markov chain that transforms a low-resolution face texture map to a high-resolution one. We adopt the formulation from [43], where the residual between the low-res and high-res studio-captured texture maps, denoted as $e_0 = \\mathcal{I}_{LR}^{Studio} - \\mathcal{I}^{Studio}$ , is used to define the Markov process. More specifically, the forward process is defined as follows:", + "bbox": [ + 214, + 708, + 787, + 797 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathcal {I} _ {t} ^ {\\text {S t u d i o}} \\mid \\mathcal {I} _ {t - 1} ^ {\\text {S t u d i o}}, \\mathcal {I} _ {L R} ^ {\\text {S t u d i o}}\\right) = \\mathcal {N} \\left(\\mathcal {I} _ {t} ^ {\\text {S t u d i o}}; \\mathcal {I} _ {t - 1} ^ {\\text {S t u d i o}} + \\alpha_ {t} e _ {0}, \\kappa^ {2} \\alpha_ {t} I\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 797, + 787, + 828 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nt = 1, 2, \\dots , T,\n$$\n", + "text_format": "latex", + "bbox": [ + 493, + 829, + 583, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture", + "bbox": [ + 254, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\alpha_{t}$ controls the schedule with which the residual is added, $\\kappa$ is a hyperparameter controlling the noise schedule and $I$ is the identity matrix. The reverse process is defined as", + "bbox": [ + 212, + 146, + 787, + 189 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathcal {I} _ {t - 1} ^ {S t u d i o} \\mid \\mathcal {I} _ {\\{t, L R \\}} ^ {S t u d i o}, \\mathcal {I} ^ {S t u d i o}\\right) = \\mathcal {N} \\left(\\mathcal {I} _ {t - 1} ^ {S t u d i o} \\left| \\frac {\\eta_ {t - 1}}{\\eta_ {t}} \\mathcal {I} _ {t} ^ {S t u d i o} + \\frac {\\alpha_ {t}}{\\eta_ {t}} \\mathcal {I} ^ {S t u d i o}, \\kappa^ {2} \\frac {\\eta_ {t - 1}}{\\eta_ {t}} \\alpha_ {t} I\\right), \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 193, + 787, + 228 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\eta_{t} = \\alpha_{t} + \\eta_{t - 1}$ . We refer the reader to [43] for details regarding the formulation of both the forward and reverse process. Finally, the diffusion model, $f_{\\phi}$ , is trained to minimize the following objective", + "bbox": [ + 212, + 238, + 787, + 282 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\phi} \\sum_ {t} \\| f _ {\\phi} \\left(\\mathcal {I} _ {t} ^ {S t u d i o}, \\mathcal {I} _ {L R} ^ {S t u d i o}, t\\right) - \\mathcal {I} ^ {S t u d i o} \\| _ {2} ^ {2} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 287, + 787, + 310 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During inference, we use $f_{\\phi}$ to add realistic facial details to the low-resolution output of $\\mathcal{G}_{\\text{Studio}}$ as follows:", + "bbox": [ + 214, + 320, + 787, + 349 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {I}} ^ {*} = \\operatorname {R e v e r s e P r o c e s s} \\left(f _ {\\phi}, \\mathcal {I} ^ {*}\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 343, + 787, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Recovering accurate facial details. While $f_{\\phi}$ enhances $\\mathcal{I}^*$ by adding realistic facial details, it struggles to recover details already present in the phone-captured texture map $\\mathcal{I}^{phone}$ due to the inherent lack of such details in $\\mathcal{I}^*$ . The loss of details occurs during the illumination manipulation and inpainting process, as described in Eq. (1) and Eq. (2). To address this issue, we incorporate the image gradient from the phone-captured texture map into the low-resolution texture map during the training of the diffusion model, as follows:", + "bbox": [ + 212, + 366, + 787, + 470 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} _ {L R} ^ {\\text {S t u d i o *}} = \\mathcal {I} _ {L R} ^ {\\text {S t u d i o}} + \\mathbb {G} (\\mathcal {I} ^ {\\text {p h o n e}}) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 467, + 787, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{I}^{phone}$ represents the phone-captured texture map, and $\\mathcal{I}^{Studio}$ is the studio-captured texture map of the same person. $\\mathbb{G}$ denotes the operator used to calculate the image gradient. During training, $\\mathcal{I}_{LR}^{Studio*}$ replaces $\\mathcal{I}_{LR}^{Studio}$ in Eq. (9), Eq. (10), and Eq. (11). In the inference stage, we augment $\\mathcal{I}^*$ by adding the gradient of the phone-captured texture map as follows:", + "bbox": [ + 212, + 489, + 787, + 566 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {I}} ^ {*} = \\text {R e v e r s e P r o c e s s} \\left(f _ {\\phi}, \\mathcal {I} ^ {*} + \\mathbb {G} \\left(\\mathcal {I} ^ {\\text {p h o n e}}\\right)\\right) \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 571, + 787, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation details We utilize 83 paired texture maps, representing subjects for whom we have both $\\mathcal{I}^{Studio}$ and $\\mathcal{I}^{phone}$ , to train the diffusion model. Following the approach in [43], we employ a latent diffusion model. To prevent overfitting, training is conducted on random $512 \\times 512$ crops of the $1024 \\times 1024$ resolution texture map. During the inference stage, we employ the full-resolution texture map.", + "bbox": [ + 212, + 599, + 787, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Driving a high-quality avatar", + "text_level": 1, + "bbox": [ + 214, + 696, + 462, + 713 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With the studio-lit version $\\mathcal{I}^*$ now available for a given neutral phone-captured texture map $\\mathcal{I}^{phone}$ , we proceed to estimate a color transform mapping from the phone-captured texture map to the studio-lit texture map as follows:", + "bbox": [ + 212, + 723, + 787, + 768 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{G, B \\right\\} \\leftarrow \\underset {\\{G, B \\}} {\\operatorname {a r g m i n}} \\left| \\left| \\mathcal {I} ^ {*} - \\left(\\mathrm {R s z} (G) \\times \\mathcal {I} ^ {\\text {p h o n e}} + \\mathrm {R s z} (B)\\right) \\right| \\right|. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 773, + 787, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, $G$ and $B$ represent gain and bias maps of resolution $k \\times k$ and $\\mathbb{R} \\times \\mathbb{Z}$ is the resizing operator with bilinear interpolation. Utilizing $G$ and $B$ , we perform a transformation", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Athar et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "on phone-captured expression texture maps to achieve studio-like lighting, as outlined below:", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} _ {e x p} ^ {*} = \\operatorname {R s z} (G) \\times \\mathcal {I} _ {e x p} ^ {\\text {p h o n e}} + \\operatorname {R s z} (B). \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 176, + 787, + 193 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In our experiment, we select the value of $k = 32$ to efficiently transform the lighting while preserving the details.", + "bbox": [ + 212, + 199, + 782, + 228 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Now, the studio-like high-resolution neutral texture $\\hat{I}^*$ serves as conditioning data for the universal avatar prior from [6]. By combining the expression code generated by inputting $\\mathcal{I}_{exp}^{*}$ into the expression encoder from [6], we can render a high-quality avatar from any desired view $v$ as follows:", + "bbox": [ + 212, + 229, + 782, + 289 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} = \\operatorname {A V A} \\left(\\hat {\\mathcal {I}} ^ {*}, \\mathcal {I} _ {e x p} ^ {*} - \\mathcal {I} ^ {*}, v, \\mathbf {F}\\right), \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 301, + 785, + 316 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbf{I}$ represents the render of the avatar, and $\\mathbf{F}$ is the geometry generated during 3D face tracking for a monocular phone capture. AVA corresponds to the inference process of the universal prior model. For more details about the universal prior model, please refer to [6].", + "bbox": [ + 212, + 328, + 782, + 388 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Results", + "text_level": 1, + "bbox": [ + 214, + 412, + 310, + 428 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we introduce the dataset used in this paper, along with the baselines of our method. Subsequently, we present both quantitative and qualitative results of our method, comparing it to prior work. All phone-captured texture maps are generated using the mesh fitting procedure outlined in Authentic Volumetric Avatars (AVA) [6]. Furthermore, we utilize AVA to render avatars based on the texture maps generated by all the methods, facilitating the calculation of image space metrics, including face identity similarity.", + "bbox": [ + 212, + 443, + 540, + 609 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/589263fe66c9ae45cf37636cdfbc92187676efd5106280303853419c00ba1217.jpg", + "table_caption": [ + "Table 1: Quantitative results on paired phone-cum-studio captured texture maps. Our method outperforms prior work across all metrics. Best and Second Best scores are highlighted." + ], + "table_footnote": [], + "table_body": "
ModelsPSNR ↑SSIM ↑LPIPS ↓DISTS ↓
Ours22.760.7260.3640.163
AgileGAN [38]18.050.6570.4060.180
CUT [33]21.4400.6420.4020.169
CycleGAN [42]21.0870.6430.4000.175
", + "bbox": [ + 552, + 515, + 781, + 580 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/4d256fe2059e3de284af8d1b01f53d4732320b634406d5a98d347929ba606b67.jpg", + "table_caption": [ + "Table 2: FaceID results on unpaired phone captured texture maps. Our method better preserves identity than prior work without sacrificing the quality of the generated texture maps. Best and Second Best scores are highlighted." + ], + "table_footnote": [], + "table_body": "
ModelsOursAgileGANCUTCycleGAN
FaceID4.31e-41.36e-36.89e-45.19e-4
", + "bbox": [ + 217, + 737, + 447, + 772 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training and Evaluation Data. We utilize a dataset comprising 12,543 neutral phone-captured texture maps to train $\\mathcal{G}_{phone}$ , and 383 studio-captured texture maps for fine-tuning $\\mathcal{G}_{phone}$ to obtain $\\mathcal{G}_{Studio}$ . Among the 383 maps, 83 are paired neutral texture maps $\\{\\mathcal{I}^{phone},\\mathcal{I}^{Studio}\\}$ used for calculating the perceptual reconstruction loss, as described in Eq. (6). This paired dataset also serves for training the detail-preserving diffusion model. For quantitative and qualitative evaluation, we employ 10 paired phone-cum-studio captured texture maps and 31 phone-captured texture maps.", + "bbox": [ + 460, + 611, + 787, + 806 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baselines. We compare our method to the following prior works on unpaired image-to-image translation. 1) AgileGAN [38], which utilizes an aligned StyleGAN latent space", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture", + "bbox": [ + 254, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "to stylize images, even with very few examples; 2) Contrastive Unpaired Translation (CUT) [33], which employs a patch-based contrastive loss in a learned feature space, enabling domain adaptation, even for single images; 3) CycleGAN [42], which trains two generators based on a cycle-consistency loss to translate images between two domains. Since AgileGAN [38] has no code available, we implement it ourselves. We use publicly available code for CUT [33] and [42] for all experiments.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/01c936b38159dfa3bcf75574d26e9bf90d555086149ddf43078d61cb2bff9b37.jpg", + "image_caption": [ + "Fig. 3: Comparisons with baselines on paired phone-cum-studio texture data reveal that our method produces results closest to the ground truth. It achieves uniform studio-like lighting, well-reconstructed facial details visible in the phone capture, and effective inpainting of missing regions. In contrast, AgileGAN [38] fails to preserve identity, while CUT [33] and CycleGAN [42] introduce significant artifacts." + ], + "image_footnote": [], + "bbox": [ + 272, + 270, + 730, + 512 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Quantitative Results. In Table 1, we present the results of a quantitative evaluation comparing our method to the baselines using 10 paired texture maps captured with both phones and in a studio setting. The evaluation metrics include mean PSNR, SSIM, LPIPS [21], and DISTS [9]. DISTS and LPIPS are perceptual metrics aligned with human judgment, while PSNR and SSIM are pixel-based metrics. As shown in Table 1, our method consistently outperforms the baseline across all metrics. The improvement is not limited to perceptual metrics but also extends to PSNR and SSIM, highlighting the efficacy of our proposed approach. It is worth noting that the relatively low PSNR and SSIM scores can be attributed to their sensitivity to small pixel shifts between the studio and phone-captured textures.", + "bbox": [ + 212, + 625, + 787, + 776 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In Table 2, we display the face embedding distances, measured using Eq. (4), between avatars rendered from the phone-captured texture and the textures generated by various methods on the 31 unpaired phone captures. Our method exhibits the best preservation of facial identity.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Athar et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/5d5a5b29b3ae5edc8c829d17bda2c83d0a82badbeba3db073e3f249d7789f125.jpg", + "image_caption": [ + "Fig. 4: Comparisons on Unpaired Phone-Captured Data: In comparison to prior work, our method excels in generating texture maps with superior preservation of identity, enhanced photorealism in facial details, more uniform illumination, and improved inpainting of missing regions." + ], + "image_footnote": [], + "bbox": [ + 287, + 143, + 715, + 474 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Qualitative Results. In Fig 3, we provide some qualitative results on the paired test data. It is evident that our method generates a more plausible and photorealistic texture map compared to prior work, showcasing improvements in illumination transfer, facial details, and inpainting of missing regions.", + "bbox": [ + 212, + 570, + 787, + 632 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In Fig 4, we present qualitative results comparing our method with prior work. While AgileGAN [38] successfully changes the illumination to be studio-like and inpaints missing regions, it introduces a significant identity shift (quantitatively measured in Table 2) and lacks facial details. We attribute this to the use of the $Z^{+}$ space [38], which may not be flexible enough for high-fidelity inversion, and the absence of identity-preserving constraints, both in architecture and optimization, during training. Due to its contrastive training paradigm, CUT [33] preserves identity better than AgileGAN but introduces significant artifacts. It is also unable to inpaint the missing regions around the ears and corners of the head. Similarly, like CUT [33], CycleGAN [42] also preserves identity better than AgileGAN [38] but struggles with inpainting missing regions. The textures also contain numerous uncanny artifacts that are uncharacteristic of human skin. In contrast, our method generates a high-quality studio-illuminated texture map with accurate facial details and inpainted missing regions.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture", + "bbox": [ + 256, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5a8c586c7d97e1c4cbef53a27ac8700df4a6a5cd48875a36b4568969354b7e69.jpg", + "image_caption": [ + "Fig. 5: Avatar Reanimation. The top row showcases a multiview render of an avatar generated using studio-like neutral and expression texture maps, created through our method as described in Sect 3.3. In contrast, the bottom row utilizes phone-captured texture maps. Notably, the lip region, highlighted with pink rectangles, appears more realistic and less blurry when using studio-like texture maps generated by our method." + ], + "image_footnote": [], + "bbox": [ + 330, + 143, + 671, + 262 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/961325b567e324f045bca7a2d849bafe8e750a0a420f5ab6b43435330c5bde4b.jpg", + "image_caption": [ + "Fig. 6: Comparison of Avatars Generated Using [6]: In this comparison, we evaluate the quality of avatars generated by the Universal Prior Model (UPM) from AVA [6], using texture maps from various methods as input. Evidently, the avatars generated using texture maps from our method far exceed those generated by prior work." + ], + "image_footnote": [], + "bbox": [ + 388, + 344, + 617, + 556 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Reanimation Results. As explained in Sect 3.3, we can transform the phone expression texture maps $\\mathcal{T}_{exp}^{phone}$ to studio-like illumination using Eq. (15) and Eq. (16). We observe that utilizing studio-like illuminated expression texture maps $\\mathcal{T}^{*}_{exp}$ results in a modest improvement in the quality of reanimated avatars using the universal prior model from AVA [6]. In Fig 5, we present an example where the top row shows the render of an avatar using studio-like neutral and expression textures, while the bottom row displays a render using phone-captured neutral and expression textures. It is evident that the lip region appears more realistic and less blurry when reanimated with the studio-like neutral and expression textures compared to the phone-captured textures. We recommend that readers refer to our supplementary video for a more comprehensive comparison.", + "bbox": [ + 214, + 641, + 787, + 792 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.1 Ablations", + "text_level": 1, + "bbox": [ + 215, + 815, + 326, + 829 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Athar et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we conduct ablation studies to evaluate the various components of our method.", + "bbox": [ + 215, + 146, + 537, + 176 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Details Conditioned Diffusion.", + "text_level": 1, + "bbox": [ + 215, + 189, + 429, + 204 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We ablate the necessity of using facial details extracted from the phone capture texture map, in the form of an image gradient, to synthesize accurate facial details using a diffusion model. We explore three scenarios:", + "bbox": [ + 214, + 205, + 537, + 280 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "1) A Diffusion model that does not use any conditioning on phone-captured texture gradient (Vanilla Diffusion); 2) A Diffusion model that uses the phone-captured texture gradient only dur", + "bbox": [ + 214, + 282, + 539, + 343 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "ing inference but not during training; 3) Our model, which incorporates the phone-captured texture gradient during both training and inference.", + "bbox": [ + 214, + 343, + 785, + 372 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Fig 7, we present the results of each of the three scenarios. While the Vanilla Diffusion generates a realistic-looking facial texture, it fails to preserve facial details present in the phone-captured texture map. Prominent moles, marked by the blue boxes in the phone-captured texture map, are not generated by the Vanilla Diffusion model. When conditioning the reverse process using the phone-texture gradient only during inference, we observe that the model misses some details and transfers shading from the phone-captured texture map to the studio-lit texture map (marked by red boxes in Fig 7), which is undesirable. Ideally, we want facial details to be preserved while eliminating illumination-dependent effects, such as strong", + "bbox": [ + 214, + 375, + 539, + 599 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 3: Detail conditioning ablation. We calculate quantitative metrics on the 10 paired phone-cum-studio captured texture maps. It is evident that utilizing detail conditioning during both training and inference yields the best performance. Best and Second Best scores are highlighted.", + "bbox": [ + 547, + 146, + 787, + 258 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/92d09d3f4ba33a42b5d5d1a2adf5e7ae272f8748f3c2329e7fa790369d4b43ae.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsLPIPS ↓DISTS ↓
Vanilla Diffusion0.3830.179
Inference-only Details conditioning0.3760.183
Training + Inference Details conditioning (Ours)0.3640.163
", + "bbox": [ + 552, + 268, + 795, + 316 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 4: $W^{+}$ vs. $Z$ Space ablation. We utilize the 10 paired phone-cum-studio captured texture maps to compare the outcomes of training in the $W^{+}$ and $Z$ spaces. The results clearly demonstrate that training in the $W^{+}$ space yields significantly better outcomes.", + "bbox": [ + 547, + 429, + 787, + 526 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1e271e6deb7696597dee1211209fad9d085d0c47480fec2abf1160fb4f6845cd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsPSNR ↑SSIM ↑LPIPS ↓DISTS ↓
Z Space21.3680.7180.4010.172
W+ Space (Ours)22.760.7260.3640.163
", + "bbox": [ + 560, + 527, + 777, + 569 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "shading and shadows, from the studio-lit texture map. Our method conditions the diffusion model on the phone-texture gradient during both inference and training. As shown in Fig 7, this allows the model to learn to retain facial details while ignoring illumination-induced effects, such as strong shading, shadows, and specularities when synthesizing facial details. In Table 3, we quantitatively validate each of the aforementioned diffusion model designs on paired data using LPIPS [21] and DISTS [9]. As can be seen, conditioning the diffusion model on phone-texture gradients both during training and inference gives the best results.", + "bbox": [ + 214, + 599, + 787, + 720 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "$W^{+}$ space vs. $Z$ space for finetuning. We now evaluate the effectiveness of the $W^{+}$ space over the standard $Z$ space for fine-tuning $\\mathcal{G}_{\\text{Studio}}$ . For a quantitative comparison, we calculate the FaceID loss over the 31 unpaired evaluation texture maps and compute PSNR, SSIM, LPIPS [21], and DISTS [9] over the 10 paired studio-cum-phone texture maps. Qualitative results, along with the average FaceID distance over unpaired data, are shown in Fig 8, and the metrics on paired data are presented in Table 4. As evident, sampling in the $W^{+}$ space generalizes much better to unseen identities, exhibiting", + "bbox": [ + 214, + 733, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture", + "bbox": [ + 256, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/190ce0af12d939452e07756539f943fafe4677c1219e171617e66fc70c4e83eb.jpg", + "image_caption": [ + "Fig. 7: Ablation on Detail-conditioned Diffusion. When no detail conditioning is applied, facial details visible in the phone capture are not reproduced. When detail conditioning is employed only during inference, undesirable shading effects are transferred to the generated texture map. Optimal results, with the most accurate and plausible reproduction of facial details, are achieved when detail conditioning is applied during both training and inference. Please refer to the text for further details." + ], + "image_footnote": [], + "bbox": [ + 287, + 143, + 715, + 275 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/443bdafb5b111eace363ef8fe789a930e970f627223ae84a622e8346efa6040d.jpg", + "image_caption": [ + "Fig. 8: Qualitative $W^{+}$ space ablation. We observe that when $\\mathcal{G}_{\\text{Studio}}$ is finetuned using the $Z$ space, it fails to generalize to novel subjects and exhibits significant artifacts. In contrast, training $\\mathcal{G}_{\\text{Studio}}$ in the $W^{+}$ lends it far better generalization." + ], + "image_footnote": [], + "bbox": [ + 302, + 375, + 702, + 450 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "significantly fewer artifacts and yielding superior results in terms of FaceID, PSNR, SSIM, LPIPS, and DISTS. We posit that this is because samples in the $W^{+}$ space (i.e., samples from $S_{W^{+}}$ space) are near-perfect inversions of the training data, making them more diverse than those generated from the Gaussian-distributed $Z$ space. Consequently, as seen in Fig 8, this leads to better generalization to unseen subjects. The model trained using the $Z$ space exhibits uncanny artifacts in its results.", + "bbox": [ + 212, + 532, + 787, + 625 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion and Limitations", + "text_level": 1, + "bbox": [ + 214, + 628, + 483, + 643 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we present a method for generating studio-like, high-quality avatars from monocular phone captures, using a StyleGAN2-based image-to-image translation and diffusion-based image upsampling. Experiments show the effectiveness of our approach in manipulating lighting, inpainting missing parts and generating facial details thus enabling the creation of complete, studio-lit textures for rendering high-quality avatars. However, our method does have limitations. It struggles with input textures exhibiting extreme non-uniform lighting due to the constrained lighting conditions in our training data (we include examples in the supplementary). We also do not fine-tune the universal prior model to personalize the avatar based on phone capture data, as suggested in [6], thus the avatars lack personalized details for different facial expressions. Finally, our avatars are incomplete, featuring only the head. Future work involves extending the model to include the neck, shoulders, hands, and the entire body.", + "bbox": [ + 212, + 657, + 789, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Athar et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 310, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alexander, O., Rogers, M., Lambeth, W., Chiang, J., Ma, W., Wang, C., Debevec, P.: The digital emily project: Achieving a photoreal digital actor. IEEE Computer Graphics and Applications 30 (2009) 3", + "2. Alexander, O., Fyffe, G., Busch, J., Yu, X., Ichikari, R., Jones, A., Debevec, P., Jimenez, J., Danvoye, E., Antionazzi, B., et al.: Digital ira: Creating a real-time photoreal digital actor. In: ACM SIGGRAPH 2013 Posters, pp. 1-1 (2013) 3", + "3. Athar, S., Xu, Z., Sunkavalli, K., Shechtman, E., Shu, Z.: Rignerf: Fully controllable neural 3d portraits. In: CVPR (June 2022) 2", + "4. Bi, S., Lombardi, S., Saito, S., Simon, T., Wei, S.E., Mcphail, K., Ramamoorthi, R., Sheikh, Y., Saragih, J.: Deep reightable appearance models for animatable faces. ACM Transactions on Graphics (TOG) 40(4), 1-15 (2021) 4", + "5. Borshukov, G., Lewis, J.P.: Realistic human face rendering for\" the matrix reloaded\". In: ACM Siggraph 2005 Courses, pp. 13-es (2005) 3", + "6. Cao, C., Simon, T., Kim, J.K., Schwartz, G., Zollhoefer, M., Saito, S.S., Lombardi, S., Wei, S.E., Belko, D., Yu, S.I., Sheikh, Y., Saragih, J.: Authentic volumetric avatars from a phone scan. ACM Trans. Graph. (2022) 1, 2, 3, 4, 5, 9, 12, 14", + "7. Cao, C., Wu, H., Weng, Y., Shao, T., Zhou, K.: Real-time facial animation with image-based dynamic avatars. ACM Transactions on Graphics 35(4) (2016) 4", + "8. Casas, D., Alexander, O., Feng, A.W., Fyffe, G., Ichikari, R., Debevec, P., Wang, R., Suma, E., Shapiro, A.: Rapid photorealistic blendshapes from commodity rgb-d sensors. In: Proceedings of the 19th Symposium on Interactive 3D Graphics and Games. pp. 134-134 (2015) 4", + "9. Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image quality assessment: Unifying structure and texture similarity 44(5), 2567-2581 (2022). https://doi.org/10.1109/TPAMI.2020.3045810 10, 13", + "0. Gafni, G., Thies, J., Zollhofer, M., Nießner, M.: Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8649-8658 (2021) 4", + "1. Garrido, P., Zollhöfer, M., Casas, D., Valgaerts, L., Varanasi, K., Pérez, P., Theobalt, C.: Reconstruction of personalized 3d face rigs from monocular video. ACM Transactions on Graphics (TOG) 35(3), 1-15 (2016) 4", + "2. Ghosh, A., Fyffe, G., Tunwattanapong, B., Busch, J., Yu, X., Debevec, P.: Multiview face capture using polarized spherical gradient illumination. ACM Trans. Graph. (2011) 3", + "3. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020) 4", + "4. Grassal, P.W., Prinzler, M., Leistner, T., Rother, C., Nießner, M., Thies, J.: Neural head avatars from monocular rgb videos. In: CVPR. pp. 18653-18664 (June 2022) 2, 4", + "5. Hu, L., Saito, S., Wei, L., Nagano, K., Seo, J., Fursund, J., Sadeghi, I., Sun, C., Chen, Y.C., Li, H.: Avatar digitization from a single image for real-time rendering. ACM Transactions on Graphics (ToG) 36(6), 1-14 (2017) 4", + "6. Huang, X., Liu, M.Y., Belongie, S., Kautz, J.: Multimodal unsupervised image-to-image translation. In: Proceedings of the European conference on computer vision (ECCV). pp. 172-189 (2018) 4", + "7. Ichim, A.E., Bouaziz, S., Pauly, M.: Dynamic 3d avatar creation from hand-held video input. ACM Transactions on Graphics (ToG) 34(4), 1-14 (2015) 4", + "8. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: CVPR (2017) 2, 4" + ], + "bbox": [ + 225, + 178, + 785, + 838 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 769, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Jimenez, J., Echevarria, J.I., Oat, C., Gutierrez, D.: Practical and realistic facial wrinkles animation. In: GPU Pro 360 Guide to Geometry Manipulation, pp. 95-107. AK Peters/CRC Press (2018) 3", + "20. Jimenez, J., Scully, T., Barbosa, N., Donner, C., Alvarez, X., Vieira, T., Matts, P., Orvalho, V., Gutierrez, D., Weyrich, T.: A practical appearance model for dynamic facial color. In: ACM SIGGRAPH Asia 2010 papers, pp. 1-10 (2010) 3", + "21. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and superresolution. In: ECCV (2016) 10, 13", + "22. Karras, T., Aittala, M., Hellsten, J., Laine, S., Lehtinen, J., Aila, T.: Training generative adversarial networks with limited data. In: Proc. NeurIPS (2020) 6", + "23. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019) 7", + "24. Lattas, A., Moschoglou, S., Gecer, B., Ploumpis, S., Triantafyllou, V., Ghosh, A., Zafeiriou, S.: Avatarme: Realistically renderable 3d facial reconstruction\" in-the-wild\". In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 760-769 (2020) 4", + "25. Lattas, A., Moschoglou, S., Ploumpis, S., Gecer, B., Ghosh, A., Zafeiriou, S.: Avatarme++: Facial shape and brdf inference with photorealistic rendering-aware gans. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9269-9284 (2021) 4", + "26. Lin, J., Yuan, Y., Zou, Z.: Meingame: Create a game character face from a single portrait. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 311-319 (2021) 4", + "27. Lombardi, S., Saragih, J., Simon, T., Sheikh, Y.: Deep appearance models for face rendering. ACM Transactions on Graphics (ToG) 37(4), 1-13 (2018) 2, 3", + "28. Lombardi, S., Simon, T., Schwartz, G., Zollhoefer, M., Sheikh, Y., Saragih, J.: Mixture of volumetric primitives for efficient neural rendering. ACM Transactions on Graphics (ToG) 40(4), 1-13 (2021) 2, 3", + "29. Luo, H., Nagano, K., Kung, H., Xu, Q., Wang, Z., Wei, L., Hu, L., Li, H.: Normalized avatar synthesis using stylegan and perceptual refinement. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 4", + "30. Ma, S., Simon, T., Saragih, J., Wang, D., Li, Y., De La Torre, F., Sheikh, Y.: Pixel codec avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 64-73 (2021) 3", + "31. Mescheder, L., Geiger, A., Nowozin, S.: Which training methods for gans do actually converge? (2018) 7", + "32. Nagano, K., Fyffe, G., Alexander, O., Barbic, J., Li, H., Ghosh, A., Debevec, P.E.: Skin microstructure deformation with displacement map convolution. ACM Trans. Graph. 34(4), 109-1 (2015) 3", + "33. Park, T., Efros, A.A., Zhang, R., Zhu, J.Y.: Contrastive learning for unpaired image-to-image translation. In: European Conference on Computer Vision (2020) 9, 10, 11", + "34. Pinkney, J.N., Adler, D.: Resolution dependent gan interpolation for controllable image synthesis between domains. arXiv preprint arXiv:2010.05334 (2020) 4", + "35. Sang, S., Zhi, T., Song, G., Liu, M., Lai, C., Liu, J., Wen, X., Davis, J., Luo, L.: Agileavatar: Stylized 3d avatar creation via cascaded domain bridging. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-8 (2022) 4", + "36. Seymour, M., Evans, C., Libreri, K.: Meet mike: epic avatars. In: ACM SIGGRAPH 2017 VR Village, pp. 1-2 (2017) 3", + "37. Shi, T., Yuan, Y., Fan, C., Zou, Z., Shi, Z., Liu, Y.: Face-to-parameter translation for game character auto-creation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 161–170 (2019) 4" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Athar et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "38. Song, G., Luo, L., Liu, J., Ma, W.C., Lai, C., Zheng, C., Cham, T.J.: Agilegan: stylizing portraits by inversion-consistent transfer learning. ACM Transactions on Graphics (TOG) 40(4), 1-13 (2021) 2, 4, 9, 10, 11", + "39. Wang, T.C., Liu, M.Y., Tao, A., Liu, G., Kautz, J., Catanzaro, B.: Few-shot video-to-video synthesis (2019) 4", + "40. Wang, T.C., Liu, M.Y., Zhu, J.Y., Tao, A., Kautz, J., Catanzaro, B.: High-resolution image synthesis and semantic manipulation with conditional gans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8798-8807 (2018) 4", + "41. Yamaguchi, S., Saito, S., Nagano, K., Zhao, Y., Chen, W., Olszewski, K., Morishima, S., Li, H.: High-fidelity facial reflectance and geometry inference from an unconstrained image. ACM Transactions on Graphics (TOG) 37(4), 1-14 (2018) 4", + "42. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networkss. In: Computer Vision (ICCV), 2017 IEEE International Conference on (2017) 2, 4, 9, 10, 11", + "43. Zongsheng Yue, J.W., Loy, C.C.: Resshift: Efficient diffusion model for image super-resolution by residual shifting. In: Advances in Neural Information Processing Systems (NeurIPS) (2023) 7, 8" + ], + "bbox": [ + 215, + 146, + 787, + 382 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Bridging the Gap_ Studio-like Avatar Creation from a Monocular Phone Capture/3e2c48bd-a83b-478a-9094-3e6d6e7a5043_model.json b/2024/Bridging the Gap_ Studio-like Avatar Creation from a Monocular Phone Capture/3e2c48bd-a83b-478a-9094-3e6d6e7a5043_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a086b5929b1b6602e2a56abf88bf6bd00e1dff67 --- /dev/null +++ b/2024/Bridging the Gap_ Studio-like Avatar Creation from a Monocular Phone Capture/3e2c48bd-a83b-478a-9094-3e6d6e7a5043_model.json @@ -0,0 +1,2412 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.23, + 0.141, + 0.776, + 0.188 + ], + "angle": 0, + "content": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.212, + 0.78, + 0.245 + ], + "angle": 0, + "content": "ShahRukh Athar\\(^{1,3*}\\) Shunsuke Saito\\(^{2}\\) Zhengyu Yang\\(^{2}\\) Stanislav Pidhorskyi\\(^{2}\\) Chen Cao\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.257, + 0.484, + 0.274 + ], + "angle": 0, + "content": "\\(^{1}\\)Captions Research, New York" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.258, + 0.729, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\)Meta Reality Labs, Pittsburgh." + }, + { + "type": "text", + "bbox": [ + 0.398, + 0.288, + 0.639, + 0.305 + ], + "angle": 0, + "content": "Stony Brook University, New York" + }, + { + "type": "text", + "bbox": [ + 0.283, + 0.317, + 0.721, + 0.33 + ], + "angle": 0, + "content": "shahrukh@nocapinc.com shunsuke.saito16@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.292, + 0.331, + 0.712, + 0.344 + ], + "angle": 0, + "content": "stpidhorskyi@meta.com zhengyu-yang@outlook.com" + }, + { + "type": "text", + "bbox": [ + 0.409, + 0.346, + 0.596, + 0.358 + ], + "angle": 0, + "content": "zju.caochen@gmail.com" + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.396, + 0.292, + 0.42 + ], + "angle": 0, + "content": "Monocular phone capture" + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.407, + 0.422, + 0.419 + ], + "angle": 0, + "content": "Extracted texture map" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.423, + 0.29, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.48, + 0.29, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.295, + 0.424, + 0.44, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.402, + 0.576, + 0.423 + ], + "angle": 0, + "content": "Generated highresolution texture n" + }, + { + "type": "image", + "bbox": [ + 0.463, + 0.424, + 0.607, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.402, + 0.781, + 0.423 + ], + "angle": 0, + "content": "High-quality reanimation and novel view synthesis" + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.424, + 0.766, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.549, + 0.788, + 0.62 + ], + "angle": 0, + "content": "Fig. 1: From a low-resolution texture map obtained through monocular phone capture, our method produces a high-resolution texture map with precise facial details, studio-like illumination, and inpainted missing regions. This generated texture map can subsequently be utilized to create a high-quality, photorealistic avatar using the pretrained Universal Prior Model (UPM) from AVA [6]." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.637, + 0.744, + 0.818 + ], + "angle": 0, + "content": "Abstract. Creating photorealistic avatars for individuals traditionally involves extensive capture sessions with complex and expensive studio devices like the LightStage system. While recent strides in neural representations have enabled the generation of photorealistic and animatable 3D avatars from quick phone scans, they have the capture-time lighting baked-in, lack facial details and have missing regions in areas such as the back of the ears. Thus, they lag in quality compared to studio-captured avatars. In this paper, we propose a method that bridges this gap by generating studio-like illuminated texture maps from short, monocular phone captures. We do this by parameterizing the phone texture maps using the \\( W^{+} \\) space of a StyleGAN2, enabling near-perfect reconstruction. Then, we finetune a StyleGAN2 by sampling in the \\( W^{+} \\) parameterized space using a very small set of studio-captured textures as an adversarial training signal. To further enhance the realism and accuracy of facial details, we super-resolve" + }, + { + "type": "footer", + "bbox": [ + 0.217, + 0.826, + 0.513, + 0.841 + ], + "angle": 0, + "content": "*Work done while interning at Meta Reality Labs" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.341, + 0.128 + ], + "angle": 0, + "content": "Athar et al." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.148, + 0.744, + 0.219 + ], + "angle": 0, + "content": "the output of the StyleGAN2 using carefully designed diffusion model that is guided by image gradients of the phone-captured texture map. Once trained, our method excels at producing studio-like facial texture maps from casual monocular smartphone videos. Demonstrating its capabilities, we showcase the generation of photorealistic, uniformly lit, complete avatars from monocular phone captures." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.262, + 0.357, + 0.279 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.296, + 0.789, + 0.416 + ], + "angle": 0, + "content": "Photorealistic and animatable avatars are paramount for lifelike human-to-human interactions in AR/VR applications. Creating high-fidelity avatars often requires sophisticated devices like the LightStage capture system that capture human heads with a range of facial expressions [27, 28]. Typically, these captures occur in a studio-like environment, with uniform illumination and densely sampled views to reconstruct complete avatars with consistent lighting. While these methods excel in generating hyper-realistic avatars, they cannot be scaled to millions of people as capturing so many people with a LightStage-like capture system is impractical and non-trivial." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.789, + 0.538 + ], + "angle": 0, + "content": "Recently, there has been extensive work in generating photorealistic avatars through a monocular capture [3,6,14]. These methods involve scanning various facial expressions and head poses of the user to reconstruct a 3D avatar that closely aligns with the captured data. However, a notable limitation of these methods is that the captured lighting is embedded in the avatars, causing their appearance to be heavily dependent on the capture devices and surrounding environments. Furthermore, due to the constraints of the capture setup, certain areas, such as the back of the head or ears, are never visible, leaving visible holes and artifacts when viewed during animation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.79, + 0.659 + ], + "angle": 0, + "content": "An innovative strategy to tackle this challenge is to use image-to-image translation to transform the phone data into studio-captured data. This transformation can be learnt through supervised training using paired data [18], or through unsupervised training using unpaired datasets [42]. Due to reasons mentioned earlier, creating a large-scale paired dataset of studio and phone captures is impractical, which rules out the possibility of using supervised image-to-image translation methods. On the flip side, current unsupervised image-to-image translation methods fail to preserve fine detail in the transformation process [38, 42] which is paramount to creating a high-fidelity avatar." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.79, + 0.84 + ], + "angle": 0, + "content": "In this paper, we introduce a method capable of generating studio-like, high-quality avatars from monocular phone captures. We do this by parameterizing a large-scale dataset of phone-captured face texture maps using the \\( W^{+} \\) space of a StyleGAN2. Then, we finetune this StyleGAN2 by sampling from this parametrized \\( W^{+} \\) space using a small set of unpaired studio-captured texture maps, to create a Studio-StyleGAN2 model. Our key insight is that sampling from the \\( W^{+} \\) space instead of Z space leads to a more generalizable model as samples from the \\( W^{+} \\) are, by construction, as diverse as the training set while samples from the Z space often suffer from mode-collapse. During inference, the given phone-captured texture map, is first inverted to the \\( W^{+} \\) space of the StyleGAN2. This inverted \\( W^{+} \\) vector is then given as input to the Studio-StyleGAN2 to generate a low-resolution studio-like texture map. Finally, a novel facial detail conditioned diffusion model is used to enhance facial details of the low-resolution" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "studio-like texture map obtained in the previous step. Once trained, our method excels at producing high-quality studio-like face texture maps from monocular phone captures, which are then used as inputs to the universal face prior model (UPM) from Authentic Volumetric Avatars (AVA) [6] to generate photorealistic avatars. In summary, our key contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.231, + 0.784, + 0.259 + ], + "angle": 0, + "content": "- Introducing a groundbreaking method for creating studio-like, photorealistic avatars from monocular phone captures." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.318 + ], + "angle": 0, + "content": "- Innovative finetuning of a pre-trained generative model using a minimal dataset from another domain by sampling in the inverted \\( W^{+} \\) space, which facilitates the development of a generative model for a new domain while preserving the integrity of the \\( W^{+} \\) latent space." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.321, + 0.784, + 0.362 + ], + "angle": 0, + "content": "- A novel diffusion model conditioned on the phone texture gradient, that is designed to upsample studio textures, effectively enhancing facial details and contributing to the overall realism of the generated avatars." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.231, + 0.785, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.388, + 0.367, + 0.403 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.787, + 0.478 + ], + "angle": 0, + "content": "Our paper aims to bridge the gap between a studio capture and a phone capture by creating a studio-like, photorealistic avatar through a monocular phone capture. In delineating the related work, we offer a comprehensive overview of key research domains, including studio-captured avatars, phone-captured avatars, and image-to-image translation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Studio-captured avatars. The reconstruction of high-fidelity static and dynamic models of the human head based on photometric measurements has a long-standing history in computer graphics and vision. Achieving a photorealistic human avatar often requires specialized hardware in high-end production, such as the LightStage system. To model the complex skin appearance, various physically-based models have been explored. Notably, subsurface scattering [5], linear polarization patterns [12] and fine-scale skin details [1, 2] have been investigated. For dynamic expression details, Jimenez et al. [20] compute dynamic skin appearances by blending hemoglobin distributions captured with different expressions. In their subsequent work [19], expression-dependent normal maps are interpolated to add realistic wrinkles to an animated face. Nagano et al. [32] synthesize skin microstructures based on local geometric features derived from high-precision microgeometry, acquired with an LED sphere and a skin deformer. While these methods have been instrumental for offline movie production, their substantial compute requirements make them less suitable for real-time applications. Despite recent efforts to enable real-time rendering of physically-based avatars [36], heavy compute remains challenge. In response to the challenges posed by complex physical computations, researchers have proposed a deep appearance model [27]. This model utilizes a coarse 3D triangle mesh in conjunction with view-dependent texture mapping. The texture is regressed by a neural network conditioned on viewpoint and expression latent codes. This conditioning accounts for view- and expression-dependent variations while compensating for the imperfect proxy geometry. Subsequent work extends the mesh-texture representation to a volumetric representation using a Mixture of Volumetric Primitives (MVP) [28], further enhancing the model's quality. Pixel Codec Avatars (PiCA), as demonstrated by [30], showcase the efficiency of rendering such models, even on mobile hardware platforms," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.341, + 0.128 + ], + "angle": 0, + "content": "Athar et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "by leveraging efficient per-pixel processing. Moreover, Bi et al. [4], based on relighting captured data, can also relight avatars with any novel point light or environment maps. While these methods can achieve hyper-realistic avatars, their studio requirements make them challenging to generalize to ordinary users." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Phone-captured avatars. There are several methods aimed at creating avatars in lightweight ways, even from a phone or a single image. Some of these approaches focus on generating stylized avatars based on different input requirements, ranging from capturing multi-view images using a phone [17] to utilizing a single monocular image [26, 29, 35, 37]. While these methods excel in producing animatable avatars, it is important to note that their appearance tends to be cartoonish and lacks realism." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.79, + 0.51 + ], + "angle": 0, + "content": "Another category of methods is dedicated to creating realistic human avatars based on graphics pipelines [7, 8, 11, 15, 24, 25, 41]. While relying on traditional graphics methods, these approaches often result in avatars that appear uncanny. Building upon deep learning-based representations, researchers have proposed methods to generate avatars with increased realism, including notable works by Gafni et al. [10] and Grassal et al. [14]. Among these, the work of Cao et al. [6], Authentic Volumetric Avatars (AVA), stands out for its focus on creating photorealistic avatars from phone scans. The process involves training a Universal Prior Model (UPM) using studio-captured data, followed by personalizing this UPM using data from a phone scan of an unseen subject. While the method successfully produces avatars with realistic appearance and animation, it is worth noting that the lighting is baked into the personalized avatar, and certain details, such as the back of the ears, may be missing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.79, + 0.646 + ], + "angle": 0, + "content": "Image-to-image translation We use image-to-image translation to map images from the source domain (phone data) to the target domain (studio data). Isola et al. introduced Pix2Pix [18], a method that utilizes adversarial training strategies [13] to achieve this mapping. Additionally, Wang et al. [40] focused on increasing the resolution of generated results from semantic label maps. While these methods successfully map images between domains, it is essential to note that they require paired training data for effective implementation. In many real-world scenarios, obtaining paired training data can be challenging and expensive." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.79, + 0.842 + ], + "angle": 0, + "content": "To address this issue, unsupervised image-to-image translation has been introduced. Zhu et al. [42] proposed a novel cycle-consistency loss to ensure that translating an image from one domain to another and back again should result in the original image. This helps the model maintain consistency and produce more realistic translations. Subsequent methods have further improved unsupervised image-to-image translation from different perspectives, including a multimodal model [16], few-shot input for video-to-video translation [39], translation of images with human control [34], and translation of real images into different styles [38]. However, none of these methods are designed with preservation of facial identity and the generation of realistic facial details in mind, especially with such little training data. Later in the paper we show how this prior work compares to ours for transforming low-resolution phone-captured texture maps to high-resolution studio-captured texture maps." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.581, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.144, + 0.792, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.299, + 0.789, + 0.433 + ], + "angle": 0, + "content": "Fig. 2: Method Overview. Our method employs a two-step process to train \\(\\mathcal{G}_{\\text{Studio}}\\). Initially, we train a StyleGAN2 on \\(12k\\) neutral texture maps captured by phones, yielding \\(\\mathcal{G}_{\\text{phone}}\\). Subsequently, we initialize \\(\\mathcal{G}_{\\text{Studio}}\\) with the weights of \\(\\mathcal{G}_{\\text{phone}}\\) and fine-tune it by sampling from \\(S_{W^{+}} = \\{W_{\\mathcal{I}_0^{\\text{phone}}}, \\ldots, W_{\\mathcal{I}_{N-1}^{\\text{phone}}}^{+}\\}\\); where \\(N = 12k\\), and \\(W_{\\mathcal{I}_i^{\\text{phone}}}^{+}\\) represents the vector obtained by inverting the \\(i\\)'th phone-captured texture map, \\(\\mathcal{I}_0^{\\text{phone}}\\), in the \\(W^{+}\\) space of \\(\\mathcal{G}_{\\text{phone}}\\). During inference, the given phone-captured texture map is inverted to \\(W^{+}\\), then passed to \\(\\mathcal{G}_{\\text{Studio}}\\) to generate a low-resolution studio-lit texture map, \\(\\mathcal{I}^{*}\\). Accurate facial details are subsequently added using diffusion model \\(f_{\\phi}\\), conditioned on the gradient of the phone texture. This process results in the final high-resolution, studio-lit, and completed texture map, \\(\\mathcal{I}^{*}\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.461, + 0.317, + 0.476 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.79, + 0.734 + ], + "angle": 0, + "content": "In this section, we present our method for generating studio-like avatars from the phone captures. Our approach consists of two key components: a StyleGAN2 for texture translation and a diffusion model for facial detail generation. In Sect 3.1, we describe the generation of a low-resolution texture map with studio-like lighting and missing regions inpainted. First, we pretrain a StyleGAN2 on 12k phone textures and then finetune it using a small set of studio-captured texture maps. In order to improve the generalization of the finetuned StyleGAN2, we optimize it by sampling in the \\( W^{+} \\) space, instead of the W-space or Z-space, using a set of 12k \\( W^{+} \\) vectors obtained by inverting the phone captured texture maps. In Sect 3.2, we introduce a diffusion model that generates facial details. The diffusion model takes the output from the aforementioned StyleGAN2 and generates a high-resolution neutral texture with realistic facial details. After obtaining a high-resolution studio-like neutral texture from our method, we use it to learn a color transform to transfer phone-captured expression textures to studio-lit expression textures. These expressions are subsequently utilized for animating a high-quality avatar (Sect 3.3). Notably, we intentionally avoid applying inpainting or super-resolution techniques to the expression textures. Fig. 2 provides an overview of our method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.756, + 0.549, + 0.771 + ], + "angle": 0, + "content": "3.1 Illumination manipulation and inpainting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In our initial step, we track the geometry from a monocular phone capture of the user's neutral face and extract the neutral texture \\(\\mathcal{L}^{phone}\\) from the captured image, employing the method outlined in [6]. Subsequently, we translate this phone-captured texture with in-the-wild lighting and missing regions into a texture map with studio lighting and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.341, + 0.128 + ], + "angle": 0, + "content": "Athar et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "missing regions in-painted. This translation is accomplished by parametrizing \\(\\mathcal{I}^{phone}\\) using the \\(W^{+}\\) space of a StyleGAN2, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.268, + 0.198, + 0.788, + 0.22 + ], + "angle": 0, + "content": "\\[\nW _ {\\mathcal {I} ^ {p h o n e}} ^ {+} = \\underset {W ^ {+}} {\\operatorname {a r g m i n}} \\left\\| \\mathcal {G} _ {p h o n e} \\left(W ^ {+}\\right) - \\mathcal {I} ^ {p h o n e} \\right\\| _ {2} ^ {2} + L P I P S \\left(\\mathcal {G} _ {p h o n e} \\left(W ^ {+}\\right), \\mathcal {I} ^ {p h o n e}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.23, + 0.784, + 0.275 + ], + "angle": 0, + "content": "where \\(\\mathcal{G}_{phone}\\) is a StyleGAN2 trained on phone-captured textures, and \\(W_{\\mathcal{T}^{phone}}^{+}\\) is the optimized vector in the \\(W^{+}\\) space of \\(\\mathcal{G}_{phone}\\). Now, we generate the low-resolution studio-like texture as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.43, + 0.285, + 0.788, + 0.302 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} ^ {*} = \\mathcal {G} _ {\\text {S t u d i o}} \\left(W _ {\\mathcal {I} _ {\\text {p h o n e}}} ^ {+}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.309, + 0.784, + 0.384 + ], + "angle": 0, + "content": "where \\(\\mathcal{G}_{\\text{Studio}}\\) is a StyleGAN2 responsible for generating low-resolution studio-like textures, while \\(\\mathcal{I}^*\\) refers to the studio-lit version of \\(\\mathcal{I}^{\\text{phone}}\\) with the inpainted missing regions. Ideally, \\(\\mathcal{I}^*\\) should retain the identity and semantics of \\(\\mathcal{I}^{\\text{phone}}\\), with the only distinctions being the illumination and the inpainting of missing regions. Below we describe the training procedure for \\(\\mathcal{G}_{\\text{Studio}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.385, + 0.789, + 0.601 + ], + "angle": 0, + "content": "We initialize the synthesis network of \\(\\mathcal{G}_{\\text{Studio}}\\) using weights from \\(\\mathcal{G}_{\\text{phone}}\\), which we then finetune to generate texture maps with studio lighting and inpaint the missing regions. We choose to finetune \\(\\mathcal{G}_{\\text{phone}}\\) for the following reasons: 1) Due to the immense costs of a studio capture, studio quality training data is typically very limited. In our case, we only have 383 studio-captured texture maps, making training from scratch hard [22]; 2) Since we want \\(\\mathcal{I}^*\\) and \\(\\mathcal{I}^{phone}\\) to have the same facial identity and semantics meaning, it is beneficial to learn a latent space that is shared between \\(\\mathcal{G}_{\\text{Studio}}\\) and \\(\\mathcal{G}_{\\text{phone}}\\). For finetuning, we first invert the entire dataset of \\(N\\) phone-captured texture maps using Eq. (1), giving us a set of \\(N\\) vectors in the \\(W^{+}\\) space: \\(S_{W^{+}} = \\{W_{\\mathcal{I}_0^{phone}}^{+}, \\ldots, W_{\\mathcal{I}_{N-1}^{phone}}^{+}\\}\\). Next, we randomly sample vectors from the inverted set, \\(S_{W^{+}}\\), and finetune the synthesis network of \\(\\mathcal{G}_{\\text{phone}}\\) to give \\(\\mathcal{G}_{\\text{Studio}}\\). We choose to finetune the generator with samples from inverted \\(W^{+}\\) instead of the traditional \\(W\\) space, since they are, by construction, from the real data distribution and are consequently more diverse, leading to better generalization of \\(\\mathcal{G}_{\\text{Studio}}\\). The losses we use during finetuning are described as below." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.602, + 0.784, + 0.631 + ], + "angle": 0, + "content": "Studio-Discriminator loss. This loss uses a discriminator to ensure the texture maps generated by \\(\\mathcal{G}_{\\text{Studio}}\\) are from the distribution of studio-captured texture maps:" + }, + { + "type": "equation", + "bbox": [ + 0.306, + 0.632, + 0.787, + 0.674 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {A d v} = \\mathbb {E} _ {\\mathcal {T S t u d i o} \\sim P (\\mathcal {T S t u d i o})} [ \\min (0, - 1 + \\mathcal {D} _ {\\text {S t u d i o}}) ] \\\\ + \\mathbb {E} _ {W _ {\\mathcal {I} _ {i} ^ {p h o n e}} ^ {+} \\sim S _ {W ^ {+}}} \\left[ \\min (0, - 1 - \\mathcal {D} _ {\\text {S t u d i o}} \\left(\\mathcal {G} _ {\\text {S t u d i o}} \\left(W _ {\\mathcal {T} _ {i} ^ {p h o n e}} ^ {+}\\right)\\right) \\right], \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.682, + 0.784, + 0.726 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}_{\\text{Studio}}\\) is a discriminator that initialized from the pretrained phone-captured texture StyleGAN2 Discriminator and finetuned using ground-truth studio-captured textures and textures generated by \\(\\mathcal{G}_{\\text{Studio}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.728, + 0.784, + 0.771 + ], + "angle": 0, + "content": "Face-Identity Loss This loss compares the face identity embeddings of the renders of texture maps generated from \\(\\mathcal{G}_{\\text{Studio}}\\) and \\(\\mathcal{G}_{\\text{phone}}\\) in order to ensure the identity is preserved:" + }, + { + "type": "equation", + "bbox": [ + 0.313, + 0.771, + 0.788, + 0.795 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {F a c e I D}} = \\left\\| \\left| \\mathcal {F} \\left(\\mathcal {G} _ {\\text {S t u d i o}} \\left(W _ {\\mathcal {I} _ {i} ^ {\\text {p h o n e}}} ^ {+}\\right)\\right) - \\mathcal {F} \\left(\\mathcal {G} _ {\\text {p h o n e}} \\left(W _ {\\mathcal {I} _ {i} ^ {\\text {p h o n e}}} ^ {+}\\right)\\right) \\right| \\right\\| _ {2} ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.802, + 0.55, + 0.819 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}\\) is a pretrained face recognition network1." + }, + { + "type": "page_footnote", + "bbox": [ + 0.216, + 0.824, + 0.403, + 0.839 + ], + "angle": 0, + "content": "1We use the network from here" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.256, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.191 + ], + "angle": 0, + "content": "Perceptual Loss This loss encourages the preservation of semantic features between the texture maps generated \\(\\mathcal{G}_{\\text{Studio}}\\) and \\(\\mathcal{G}_{\\text{phone}}\\) by minimizing their distance in the VGG feature space" + }, + { + "type": "equation", + "bbox": [ + 0.34, + 0.184, + 0.788, + 0.213 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {P e r c p}} = \\operatorname {L P I P S} \\left(\\mathcal {G} _ {\\text {S t u d i o}} \\left(W _ {\\mathcal {I} _ {i} ^ {\\text {p h o n e}}} ^ {+}\\right), \\mathcal {G} _ {\\text {p h o n e}} \\left(W _ {\\mathcal {I} _ {i} ^ {\\text {p h o n e}}} ^ {+}\\right)\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.219, + 0.789, + 0.263 + ], + "angle": 0, + "content": "Perceptual Reconstruction Loss This loss ensures that skintones are preserved using a small amount of paired data (i.e. subjects for whom we have both phone and studio textures maps i.e both \\(\\mathcal{I}^{phone}\\) and \\(\\mathcal{I}^{Studio}\\))" + }, + { + "type": "equation", + "bbox": [ + 0.322, + 0.268, + 0.788, + 0.297 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {P e r c p - R e c o n s}} = \\frac {1}{k} \\sum_ {i = 1} ^ {K} \\operatorname {L P I P S} \\left(\\mathcal {G} _ {\\text {S t u d i o}} \\left(W _ {\\mathcal {T} _ {i} ^ {\\text {p h o n e}}} ^ {+}\\right), \\mathcal {I} _ {i} ^ {\\text {S t u d i o}}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.308, + 0.789, + 0.367 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}_i^{Studio}\\) and \\(\\mathcal{I}_i^{phone}\\) are the ground-truth studio and ground-truth phone texture maps of the same person. We only have a very small amount of this data for training (K=83). R1-Regularization: We regularize \\(\\mathcal{D}_{Studio}\\) using the R1 regularization [31] as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.34, + 0.364, + 0.788, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {R 1} = - \\frac {\\gamma}{2} \\mathbb {E} _ {\\mathcal {I} ^ {S t u d i o} \\sim P (\\mathcal {I} ^ {S t u d i o})} \\left[ \\left\\| \\nabla \\mathcal {D} _ {S t u d i o} (\\mathcal {I} ^ {S t u d i o}) \\right\\| _ {2} ^ {2} \\right], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.387, + 0.317, + 0.401 + ], + "angle": 0, + "content": "where \\(\\gamma = 10\\)" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.403, + 0.789, + 0.493 + ], + "angle": 0, + "content": "In order to further encourage identity preservation, we leverage the multi-resolution architecture of StyleGAN2 [23] and only optimize the generator parameters after the \\(8 \\times 8\\) resolution during the finetuning process. The intuition being that identity-specific information is mostly stored in the low-resolution maps, we ablate this choice in the Supplementary section of the paper. The final optimization for the joint finetuning of \\(\\mathcal{G}_{\\text{Studio}}\\) and \\(\\mathcal{D}_{\\text{Studio}}\\) is the following:" + }, + { + "type": "equation", + "bbox": [ + 0.3, + 0.5, + 0.788, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathcal {G} _ {S t u d i o} ^ {\\theta (8 +)}} \\max _ {\\mathcal {D} _ {S t u d i o}} \\mathcal {L} _ {A d v} + \\mathcal {L} _ {R 1} + \\mathcal {L} _ {P e r c p - R e c o n s} + \\lambda_ {1} \\mathcal {L} _ {P e r c p} + \\lambda_ {2} \\mathcal {L} _ {F a c e I D}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.535, + 0.788, + 0.566 + ], + "angle": 0, + "content": "where \\(\\lambda_{1} = 0.5\\) and \\(\\lambda_{2} = 10\\). We ablate \\(\\mathcal{L}_{Percp-Recons}, \\mathcal{L}_{Percp}, \\mathcal{L}_{FaceID}\\) in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.589, + 0.498, + 0.604 + ], + "angle": 0, + "content": "3.2 Synthesis of accurate facial details" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.617, + 0.789, + 0.707 + ], + "angle": 0, + "content": "Generating high-quality facial details. \\( \\mathcal{G}_{\\text{Studio}} \\) can produce an inpainted texture map, \\( \\mathcal{I}^* \\), with illumination consistent with that of a studio capture. However, due to the limited amount of finetuning data, \\( \\mathcal{G}_{\\text{Studio}} \\) often struggles to reproduce facial details that are visibly present in \\( \\mathcal{I}^{\\text{phone}} \\) itself. This limitation results in oversmoothed and inaccurate avatars. Achieving realistic facial details necessitates an accurate modeling of the facial texture distribution, a task that the StyleGAN2 fails to do." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.709, + 0.789, + 0.798 + ], + "angle": 0, + "content": "Motivated by the recent success of latent diffusion models in accurately modeling data distributions, we propose the task of generating facial details as the reverse process of a Markov chain that transforms a low-resolution face texture map to a high-resolution one. We adopt the formulation from [43], where the residual between the low-res and high-res studio-captured texture maps, denoted as \\( e_0 = \\mathcal{I}_{LR}^{Studio} - \\mathcal{I}^{Studio} \\), is used to define the Markov process. More specifically, the forward process is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.31, + 0.799, + 0.788, + 0.829 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathcal {I} _ {t} ^ {\\text {S t u d i o}} \\mid \\mathcal {I} _ {t - 1} ^ {\\text {S t u d i o}}, \\mathcal {I} _ {L R} ^ {\\text {S t u d i o}}\\right) = \\mathcal {N} \\left(\\mathcal {I} _ {t} ^ {\\text {S t u d i o}}; \\mathcal {I} _ {t - 1} ^ {\\text {S t u d i o}} + \\alpha_ {t} e _ {0}, \\kappa^ {2} \\alpha_ {t} I\\right), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.495, + 0.83, + 0.584, + 0.842 + ], + "angle": 0, + "content": "\\[\nt = 1, 2, \\dots , T,\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.341, + 0.128 + ], + "angle": 0, + "content": "Athar et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.19 + ], + "angle": 0, + "content": "where \\(\\alpha_{t}\\) controls the schedule with which the residual is added, \\(\\kappa\\) is a hyperparameter controlling the noise schedule and \\(I\\) is the identity matrix. The reverse process is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.245, + 0.194, + 0.788, + 0.229 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathcal {I} _ {t - 1} ^ {S t u d i o} \\mid \\mathcal {I} _ {\\{t, L R \\}} ^ {S t u d i o}, \\mathcal {I} ^ {S t u d i o}\\right) = \\mathcal {N} \\left(\\mathcal {I} _ {t - 1} ^ {S t u d i o} \\left| \\frac {\\eta_ {t - 1}}{\\eta_ {t}} \\mathcal {I} _ {t} ^ {S t u d i o} + \\frac {\\alpha_ {t}}{\\eta_ {t}} \\mathcal {I} ^ {S t u d i o}, \\kappa^ {2} \\frac {\\eta_ {t - 1}}{\\eta_ {t}} \\alpha_ {t} I\\right), \\right. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.788, + 0.284 + ], + "angle": 0, + "content": "where \\(\\eta_{t} = \\alpha_{t} + \\eta_{t - 1}\\). We refer the reader to [43] for details regarding the formulation of both the forward and reverse process. Finally, the diffusion model, \\(f_{\\phi}\\), is trained to minimize the following objective" + }, + { + "type": "equation", + "bbox": [ + 0.372, + 0.289, + 0.788, + 0.311 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\phi} \\sum_ {t} \\| f _ {\\phi} \\left(\\mathcal {I} _ {t} ^ {S t u d i o}, \\mathcal {I} _ {L R} ^ {S t u d i o}, t\\right) - \\mathcal {I} ^ {S t u d i o} \\| _ {2} ^ {2} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.321, + 0.789, + 0.35 + ], + "angle": 0, + "content": "During inference, we use \\( f_{\\phi} \\) to add realistic facial details to the low-resolution output of \\( \\mathcal{G}_{\\text{Studio}} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.344, + 0.788, + 0.361 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {I}} ^ {*} = \\operatorname {R e v e r s e P r o c e s s} \\left(f _ {\\phi}, \\mathcal {I} ^ {*}\\right). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.789, + 0.472 + ], + "angle": 0, + "content": "Recovering accurate facial details. While \\( f_{\\phi} \\) enhances \\( \\mathcal{I}^* \\) by adding realistic facial details, it struggles to recover details already present in the phone-captured texture map \\( \\mathcal{I}^{phone} \\) due to the inherent lack of such details in \\( \\mathcal{I}^* \\). The loss of details occurs during the illumination manipulation and inpainting process, as described in Eq. (1) and Eq. (2). To address this issue, we incorporate the image gradient from the phone-captured texture map into the low-resolution texture map during the training of the diffusion model, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.468, + 0.788, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} _ {L R} ^ {\\text {S t u d i o *}} = \\mathcal {I} _ {L R} ^ {\\text {S t u d i o}} + \\mathbb {G} (\\mathcal {I} ^ {\\text {p h o n e}}) \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.789, + 0.567 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}^{phone}\\) represents the phone-captured texture map, and \\(\\mathcal{I}^{Studio}\\) is the studio-captured texture map of the same person. \\(\\mathbb{G}\\) denotes the operator used to calculate the image gradient. During training, \\(\\mathcal{I}_{LR}^{Studio*}\\) replaces \\(\\mathcal{I}_{LR}^{Studio}\\) in Eq. (9), Eq. (10), and Eq. (11). In the inference stage, we augment \\(\\mathcal{I}^*\\) by adding the gradient of the phone-captured texture map as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.572, + 0.788, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {I}} ^ {*} = \\text {R e v e r s e P r o c e s s} \\left(f _ {\\phi}, \\mathcal {I} ^ {*} + \\mathbb {G} \\left(\\mathcal {I} ^ {\\text {p h o n e}}\\right)\\right) \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.6, + 0.789, + 0.677 + ], + "angle": 0, + "content": "Implementation details We utilize 83 paired texture maps, representing subjects for whom we have both \\(\\mathcal{I}^{Studio}\\) and \\(\\mathcal{I}^{phone}\\), to train the diffusion model. Following the approach in [43], we employ a latent diffusion model. To prevent overfitting, training is conducted on random \\(512 \\times 512\\) crops of the \\(1024 \\times 1024\\) resolution texture map. During the inference stage, we employ the full-resolution texture map." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.698, + 0.464, + 0.714 + ], + "angle": 0, + "content": "3.3 Driving a high-quality avatar" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.724, + 0.788, + 0.77 + ], + "angle": 0, + "content": "With the studio-lit version \\(\\mathcal{I}^*\\) now available for a given neutral phone-captured texture map \\(\\mathcal{I}^{phone}\\), we proceed to estimate a color transform mapping from the phone-captured texture map to the studio-lit texture map as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.775, + 0.788, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\left\\{G, B \\right\\} \\leftarrow \\underset {\\{G, B \\}} {\\operatorname {a r g m i n}} \\left| \\left| \\mathcal {I} ^ {*} - \\left(\\mathrm {R s z} (G) \\times \\mathcal {I} ^ {\\text {p h o n e}} + \\mathrm {R s z} (B)\\right) \\right| \\right|. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Here, \\( G \\) and \\( B \\) represent gain and bias maps of resolution \\( k \\times k \\) and \\( \\mathbb{R} \\times \\mathbb{Z} \\) is the resizing operator with bilinear interpolation. Utilizing \\( G \\) and \\( B \\), we perform a transformation" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.256, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "on phone-captured expression texture maps to achieve studio-like lighting, as outlined below:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.178, + 0.788, + 0.194 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} _ {e x p} ^ {*} = \\operatorname {R s z} (G) \\times \\mathcal {I} _ {e x p} ^ {\\text {p h o n e}} + \\operatorname {R s z} (B). \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.2, + 0.784, + 0.229 + ], + "angle": 0, + "content": "In our experiment, we select the value of \\( k = 32 \\) to efficiently transform the lighting while preserving the details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.23, + 0.784, + 0.29 + ], + "angle": 0, + "content": "Now, the studio-like high-resolution neutral texture \\(\\hat{I}^*\\) serves as conditioning data for the universal avatar prior from [6]. By combining the expression code generated by inputting \\(\\mathcal{I}_{exp}^{*}\\) into the expression encoder from [6], we can render a high-quality avatar from any desired view \\(v\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.409, + 0.303, + 0.787, + 0.318 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} = \\operatorname {A V A} \\left(\\hat {\\mathcal {I}} ^ {*}, \\mathcal {I} _ {e x p} ^ {*} - \\mathcal {I} ^ {*}, v, \\mathbf {F}\\right), \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.784, + 0.39 + ], + "angle": 0, + "content": "where \\(\\mathbf{I}\\) represents the render of the avatar, and \\(\\mathbf{F}\\) is the geometry generated during 3D face tracking for a monocular phone capture. AVA corresponds to the inference process of the universal prior model. For more details about the universal prior model, please refer to [6]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.413, + 0.312, + 0.429 + ], + "angle": 0, + "content": "4 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.541, + 0.61 + ], + "angle": 0, + "content": "In this section, we introduce the dataset used in this paper, along with the baselines of our method. Subsequently, we present both quantitative and qualitative results of our method, comparing it to prior work. All phone-captured texture maps are generated using the mesh fitting procedure outlined in Authentic Volumetric Avatars (AVA) [6]. Furthermore, we utilize AVA to render avatars based on the texture maps generated by all the methods, facilitating the calculation of image space metrics, including face identity similarity." + }, + { + "type": "table_caption", + "bbox": [ + 0.549, + 0.441, + 0.788, + 0.514 + ], + "angle": 0, + "content": "Table 1: Quantitative results on paired phone-cum-studio captured texture maps. Our method outperforms prior work across all metrics. Best and Second Best scores are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.553, + 0.516, + 0.782, + 0.581 + ], + "angle": 0, + "content": "
ModelsPSNR ↑SSIM ↑LPIPS ↓DISTS ↓
Ours22.760.7260.3640.163
AgileGAN [38]18.050.6570.4060.180
CUT [33]21.4400.6420.4020.169
CycleGAN [42]21.0870.6430.4000.175
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.638, + 0.454, + 0.736 + ], + "angle": 0, + "content": "Table 2: FaceID results on unpaired phone captured texture maps. Our method better preserves identity than prior work without sacrificing the quality of the generated texture maps. Best and Second Best scores are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.738, + 0.449, + 0.773 + ], + "angle": 0, + "content": "
ModelsOursAgileGANCUTCycleGAN
FaceID4.31e-41.36e-36.89e-45.19e-4
" + }, + { + "type": "text", + "bbox": [ + 0.462, + 0.612, + 0.789, + 0.808 + ], + "angle": 0, + "content": "Training and Evaluation Data. We utilize a dataset comprising 12,543 neutral phone-captured texture maps to train \\(\\mathcal{G}_{phone}\\), and 383 studio-captured texture maps for fine-tuning \\(\\mathcal{G}_{phone}\\) to obtain \\(\\mathcal{G}_{Studio}\\). Among the 383 maps, 83 are paired neutral texture maps \\(\\{\\mathcal{I}^{phone},\\mathcal{I}^{Studio}\\}\\) used for calculating the perceptual reconstruction loss, as described in Eq. (6). This paired dataset also serves for training the detail-preserving diffusion model. For quantitative and qualitative evaluation, we employ 10 paired phone-cum-studio captured texture maps and 31 phone-captured texture maps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Baselines. We compare our method to the following prior works on unpaired image-to-image translation. 1) AgileGAN [38], which utilizes an aligned StyleGAN latent space" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.341, + 0.127 + ], + "angle": 0, + "content": "Athar et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "to stylize images, even with very few examples; 2) Contrastive Unpaired Translation (CUT) [33], which employs a patch-based contrastive loss in a learned feature space, enabling domain adaptation, even for single images; 3) CycleGAN [42], which trains two generators based on a cycle-consistency loss to translate images between two domains. Since AgileGAN [38] has no code available, we implement it ourselves. We use publicly available code for CUT [33] and [42] for all experiments." + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.271, + 0.731, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.523, + 0.789, + 0.592 + ], + "angle": 0, + "content": "Fig. 3: Comparisons with baselines on paired phone-cum-studio texture data reveal that our method produces results closest to the ground truth. It achieves uniform studio-like lighting, well-reconstructed facial details visible in the phone capture, and effective inpainting of missing regions. In contrast, AgileGAN [38] fails to preserve identity, while CUT [33] and CycleGAN [42] introduce significant artifacts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.789, + 0.777 + ], + "angle": 0, + "content": "Quantitative Results. In Table 1, we present the results of a quantitative evaluation comparing our method to the baselines using 10 paired texture maps captured with both phones and in a studio setting. The evaluation metrics include mean PSNR, SSIM, LPIPS [21], and DISTS [9]. DISTS and LPIPS are perceptual metrics aligned with human judgment, while PSNR and SSIM are pixel-based metrics. As shown in Table 1, our method consistently outperforms the baseline across all metrics. The improvement is not limited to perceptual metrics but also extends to PSNR and SSIM, highlighting the efficacy of our proposed approach. It is worth noting that the relatively low PSNR and SSIM scores can be attributed to their sensitivity to small pixel shifts between the studio and phone-captured textures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In Table 2, we display the face embedding distances, measured using Eq. (4), between avatars rendered from the phone-captured texture and the textures generated by various methods on the 31 unpaired phone captures. Our method exhibits the best preservation of facial identity." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.145, + 0.716, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.486, + 0.788, + 0.528 + ], + "angle": 0, + "content": "Fig. 4: Comparisons on Unpaired Phone-Captured Data: In comparison to prior work, our method excels in generating texture maps with superior preservation of identity, enhanced photorealism in facial details, more uniform illumination, and improved inpainting of missing regions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.571, + 0.788, + 0.633 + ], + "angle": 0, + "content": "Qualitative Results. In Fig 3, we provide some qualitative results on the paired test data. It is evident that our method generates a more plausible and photorealistic texture map compared to prior work, showcasing improvements in illumination transfer, facial details, and inpainting of missing regions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In Fig 4, we present qualitative results comparing our method with prior work. While AgileGAN [38] successfully changes the illumination to be studio-like and inpaints missing regions, it introduces a significant identity shift (quantitatively measured in Table 2) and lacks facial details. We attribute this to the use of the \\( Z^{+} \\) space [38], which may not be flexible enough for high-fidelity inversion, and the absence of identity-preserving constraints, both in architecture and optimization, during training. Due to its contrastive training paradigm, CUT [33] preserves identity better than AgileGAN but introduces significant artifacts. It is also unable to inpaint the missing regions around the ears and corners of the head. Similarly, like CUT [33], CycleGAN [42] also preserves identity better than AgileGAN [38] but struggles with inpainting missing regions. The textures also contain numerous uncanny artifacts that are uncharacteristic of human skin. In contrast, our method generates a high-quality studio-illuminated texture map with accurate facial details and inpainted missing regions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.341, + 0.127 + ], + "angle": 0, + "content": "Athar et al." + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.144, + 0.673, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.272, + 0.789, + 0.342 + ], + "angle": 0, + "content": "Fig. 5: Avatar Reanimation. The top row showcases a multiview render of an avatar generated using studio-like neutral and expression texture maps, created through our method as described in Sect 3.3. In contrast, the bottom row utilizes phone-captured texture maps. Notably, the lip region, highlighted with pink rectangles, appears more realistic and less blurry when using studio-like texture maps generated by our method." + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.345, + 0.619, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.567, + 0.789, + 0.625 + ], + "angle": 0, + "content": "Fig. 6: Comparison of Avatars Generated Using [6]: In this comparison, we evaluate the quality of avatars generated by the Universal Prior Model (UPM) from AVA [6], using texture maps from various methods as input. Evidently, the avatars generated using texture maps from our method far exceed those generated by prior work." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.642, + 0.789, + 0.794 + ], + "angle": 0, + "content": "Reanimation Results. As explained in Sect 3.3, we can transform the phone expression texture maps \\(\\mathcal{T}_{exp}^{phone}\\) to studio-like illumination using Eq. (15) and Eq. (16). We observe that utilizing studio-like illuminated expression texture maps \\(\\mathcal{T}^{*}_{exp}\\) results in a modest improvement in the quality of reanimated avatars using the universal prior model from AVA [6]. In Fig 5, we present an example where the top row shows the render of an avatar using studio-like neutral and expression textures, while the bottom row displays a render using phone-captured neutral and expression textures. It is evident that the lip region appears more realistic and less blurry when reanimated with the studio-like neutral and expression textures compared to the phone-captured textures. We recommend that readers refer to our supplementary video for a more comprehensive comparison." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.816, + 0.327, + 0.83 + ], + "angle": 0, + "content": "4.1 Ablations" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.538, + 0.178 + ], + "angle": 0, + "content": "In this section, we conduct ablation studies to evaluate the various components of our method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.19, + 0.431, + 0.205 + ], + "angle": 0, + "content": "Details Conditioned Diffusion." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.206, + 0.538, + 0.281 + ], + "angle": 0, + "content": "We ablate the necessity of using facial details extracted from the phone capture texture map, in the form of an image gradient, to synthesize accurate facial details using a diffusion model. We explore three scenarios:" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.283, + 0.54, + 0.344 + ], + "angle": 0, + "content": "1) A Diffusion model that does not use any conditioning on phone-captured texture gradient (Vanilla Diffusion); 2) A Diffusion model that uses the phone-captured texture gradient only dur" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.344, + 0.787, + 0.373 + ], + "angle": 0, + "content": "ing inference but not during training; 3) Our model, which incorporates the phone-captured texture gradient during both training and inference." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.375, + 0.54, + 0.6 + ], + "angle": 0, + "content": "In Fig 7, we present the results of each of the three scenarios. While the Vanilla Diffusion generates a realistic-looking facial texture, it fails to preserve facial details present in the phone-captured texture map. Prominent moles, marked by the blue boxes in the phone-captured texture map, are not generated by the Vanilla Diffusion model. When conditioning the reverse process using the phone-texture gradient only during inference, we observe that the model misses some details and transfers shading from the phone-captured texture map to the studio-lit texture map (marked by red boxes in Fig 7), which is undesirable. Ideally, we want facial details to be preserved while eliminating illumination-dependent effects, such as strong" + }, + { + "type": "text", + "bbox": [ + 0.548, + 0.147, + 0.789, + 0.259 + ], + "angle": 0, + "content": "Table 3: Detail conditioning ablation. We calculate quantitative metrics on the 10 paired phone-cum-studio captured texture maps. It is evident that utilizing detail conditioning during both training and inference yields the best performance. Best and Second Best scores are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.553, + 0.269, + 0.797, + 0.317 + ], + "angle": 0, + "content": "
ModelsLPIPS ↓DISTS ↓
Vanilla Diffusion0.3830.179
Inference-only Details conditioning0.3760.183
Training + Inference Details conditioning (Ours)0.3640.163
" + }, + { + "type": "text", + "bbox": [ + 0.548, + 0.43, + 0.788, + 0.527 + ], + "angle": 0, + "content": "Table 4: \\( W^{+} \\) vs. \\( Z \\) Space ablation. We utilize the 10 paired phone-cum-studio captured texture maps to compare the outcomes of training in the \\( W^{+} \\) and \\( Z \\) spaces. The results clearly demonstrate that training in the \\( W^{+} \\) space yields significantly better outcomes." + }, + { + "type": "table", + "bbox": [ + 0.562, + 0.528, + 0.778, + 0.57 + ], + "angle": 0, + "content": "
ModelsPSNR ↑SSIM ↑LPIPS ↓DISTS ↓
Z Space21.3680.7180.4010.172
W+ Space (Ours)22.760.7260.3640.163
" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.601, + 0.789, + 0.722 + ], + "angle": 0, + "content": "shading and shadows, from the studio-lit texture map. Our method conditions the diffusion model on the phone-texture gradient during both inference and training. As shown in Fig 7, this allows the model to learn to retain facial details while ignoring illumination-induced effects, such as strong shading, shadows, and specularities when synthesizing facial details. In Table 3, we quantitatively validate each of the aforementioned diffusion model designs on paired data using LPIPS [21] and DISTS [9]. As can be seen, conditioning the diffusion model on phone-texture gradients both during training and inference gives the best results." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.734, + 0.789, + 0.842 + ], + "angle": 0, + "content": "\\(W^{+}\\) space vs. \\(Z\\) space for finetuning. We now evaluate the effectiveness of the \\(W^{+}\\) space over the standard \\(Z\\) space for fine-tuning \\(\\mathcal{G}_{\\text{Studio}}\\). For a quantitative comparison, we calculate the FaceID loss over the 31 unpaired evaluation texture maps and compute PSNR, SSIM, LPIPS [21], and DISTS [9] over the 10 paired studio-cum-phone texture maps. Qualitative results, along with the average FaceID distance over unpaired data, are shown in Fig 8, and the metrics on paired data are presented in Table 4. As evident, sampling in the \\(W^{+}\\) space generalizes much better to unseen identities, exhibiting" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.341, + 0.127 + ], + "angle": 0, + "content": "Athar et al." + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.145, + 0.716, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.285, + 0.789, + 0.367 + ], + "angle": 0, + "content": "Fig. 7: Ablation on Detail-conditioned Diffusion. When no detail conditioning is applied, facial details visible in the phone capture are not reproduced. When detail conditioning is employed only during inference, undesirable shading effects are transferred to the generated texture map. Optimal results, with the most accurate and plausible reproduction of facial details, are achieved when detail conditioning is applied during both training and inference. Please refer to the text for further details." + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.375, + 0.703, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.462, + 0.789, + 0.505 + ], + "angle": 0, + "content": "Fig. 8: Qualitative \\( W^{+} \\) space ablation. We observe that when \\( \\mathcal{G}_{\\text{Studio}} \\) is finetuned using the \\( Z \\) space, it fails to generalize to novel subjects and exhibits significant artifacts. In contrast, training \\( \\mathcal{G}_{\\text{Studio}} \\) in the \\( W^{+} \\) lends it far better generalization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.789, + 0.625 + ], + "angle": 0, + "content": "significantly fewer artifacts and yielding superior results in terms of FaceID, PSNR, SSIM, LPIPS, and DISTS. We posit that this is because samples in the \\( W^{+} \\) space (i.e., samples from \\( S_{W^{+}} \\) space) are near-perfect inversions of the training data, making them more diverse than those generated from the Gaussian-distributed \\( Z \\) space. Consequently, as seen in Fig 8, this leads to better generalization to unseen subjects. The model trained using the \\( Z \\) space exhibits uncanny artifacts in its results." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.629, + 0.484, + 0.645 + ], + "angle": 0, + "content": "5 Conclusion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.79, + 0.841 + ], + "angle": 0, + "content": "In this paper, we present a method for generating studio-like, high-quality avatars from monocular phone captures, using a StyleGAN2-based image-to-image translation and diffusion-based image upsampling. Experiments show the effectiveness of our approach in manipulating lighting, inpainting missing parts and generating facial details thus enabling the creation of complete, studio-lit textures for rendering high-quality avatars. However, our method does have limitations. It struggles with input textures exhibiting extreme non-uniform lighting due to the constrained lighting conditions in our training data (we include examples in the supplementary). We also do not fine-tune the universal prior model to personalize the avatar based on phone capture data, as suggested in [6], thus the avatars lack personalized details for different facial expressions. Finally, our avatars are incomplete, featuring only the head. Future work involves extending the model to include the neck, shoulders, hands, and the entire body." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.312, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.786, + 0.221 + ], + "angle": 0, + "content": "1. Alexander, O., Rogers, M., Lambeth, W., Chiang, J., Ma, W., Wang, C., Debevec, P.: The digital emily project: Achieving a photoreal digital actor. IEEE Computer Graphics and Applications 30 (2009) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.223, + 0.787, + 0.263 + ], + "angle": 0, + "content": "2. Alexander, O., Fyffe, G., Busch, J., Yu, X., Ichikari, R., Jones, A., Debevec, P., Jimenez, J., Danvoye, E., Antionazzi, B., et al.: Digital ira: Creating a real-time photoreal digital actor. In: ACM SIGGRAPH 2013 Posters, pp. 1-1 (2013) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.265, + 0.786, + 0.29 + ], + "angle": 0, + "content": "3. Athar, S., Xu, Z., Sunkavalli, K., Shechtman, E., Shu, Z.: Rignerf: Fully controllable neural 3d portraits. In: CVPR (June 2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.293, + 0.787, + 0.333 + ], + "angle": 0, + "content": "4. Bi, S., Lombardi, S., Saito, S., Simon, T., Wei, S.E., Mcphail, K., Ramamoorthi, R., Sheikh, Y., Saragih, J.: Deep reightable appearance models for animatable faces. ACM Transactions on Graphics (TOG) 40(4), 1-15 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.335, + 0.787, + 0.362 + ], + "angle": 0, + "content": "5. Borshukov, G., Lewis, J.P.: Realistic human face rendering for\" the matrix reloaded\". In: ACM Siggraph 2005 Courses, pp. 13-es (2005) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.364, + 0.787, + 0.403 + ], + "angle": 0, + "content": "6. Cao, C., Simon, T., Kim, J.K., Schwartz, G., Zollhoefer, M., Saito, S.S., Lombardi, S., Wei, S.E., Belko, D., Yu, S.I., Sheikh, Y., Saragih, J.: Authentic volumetric avatars from a phone scan. ACM Trans. Graph. (2022) 1, 2, 3, 4, 5, 9, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.787, + 0.432 + ], + "angle": 0, + "content": "7. Cao, C., Wu, H., Weng, Y., Shao, T., Zhou, K.: Real-time facial animation with image-based dynamic avatars. ACM Transactions on Graphics 35(4) (2016) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.787, + 0.473 + ], + "angle": 0, + "content": "8. Casas, D., Alexander, O., Feng, A.W., Fyffe, G., Ichikari, R., Debevec, P., Wang, R., Suma, E., Shapiro, A.: Rapid photorealistic blendshapes from commodity rgb-d sensors. In: Proceedings of the 19th Symposium on Interactive 3D Graphics and Games. pp. 134-134 (2015) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.476, + 0.787, + 0.515 + ], + "angle": 0, + "content": "9. Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image quality assessment: Unifying structure and texture similarity 44(5), 2567-2581 (2022). https://doi.org/10.1109/TPAMI.2020.3045810 10, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.518, + 0.787, + 0.558 + ], + "angle": 0, + "content": "0. Gafni, G., Thies, J., Zollhofer, M., Nießner, M.: Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8649-8658 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.56, + 0.787, + 0.6 + ], + "angle": 0, + "content": "1. Garrido, P., Zollhöfer, M., Casas, D., Valgaerts, L., Varanasi, K., Pérez, P., Theobalt, C.: Reconstruction of personalized 3d face rigs from monocular video. ACM Transactions on Graphics (TOG) 35(3), 1-15 (2016) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.602, + 0.787, + 0.629 + ], + "angle": 0, + "content": "2. Ghosh, A., Fyffe, G., Tunwattanapong, B., Busch, J., Yu, X., Debevec, P.: Multiview face capture using polarized spherical gradient illumination. ACM Trans. Graph. (2011) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.63, + 0.787, + 0.67 + ], + "angle": 0, + "content": "3. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.672, + 0.787, + 0.699 + ], + "angle": 0, + "content": "4. Grassal, P.W., Prinzler, M., Leistner, T., Rother, C., Nießner, M., Thies, J.: Neural head avatars from monocular rgb videos. In: CVPR. pp. 18653-18664 (June 2022) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.701, + 0.787, + 0.741 + ], + "angle": 0, + "content": "5. Hu, L., Saito, S., Wei, L., Nagano, K., Seo, J., Fursund, J., Sadeghi, I., Sun, C., Chen, Y.C., Li, H.: Avatar digitization from a single image for real-time rendering. ACM Transactions on Graphics (ToG) 36(6), 1-14 (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.743, + 0.787, + 0.782 + ], + "angle": 0, + "content": "6. Huang, X., Liu, M.Y., Belongie, S., Kautz, J.: Multimodal unsupervised image-to-image translation. In: Proceedings of the European conference on computer vision (ECCV). pp. 172-189 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.787, + 0.811 + ], + "angle": 0, + "content": "7. Ichim, A.E., Bouaziz, S., Pauly, M.: Dynamic 3d avatar creation from hand-held video input. ACM Transactions on Graphics (ToG) 34(4), 1-14 (2015) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.787, + 0.839 + ], + "angle": 0, + "content": "8. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: CVPR (2017) 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.787, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.34, + 0.127 + ], + "angle": 0, + "content": "Athar et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.188 + ], + "angle": 0, + "content": "19. Jimenez, J., Echevarria, J.I., Oat, C., Gutierrez, D.: Practical and realistic facial wrinkles animation. In: GPU Pro 360 Guide to Geometry Manipulation, pp. 95-107. AK Peters/CRC Press (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "20. Jimenez, J., Scully, T., Barbosa, N., Donner, C., Alvarez, X., Vieira, T., Matts, P., Orvalho, V., Gutierrez, D., Weyrich, T.: A practical appearance model for dynamic facial color. In: ACM SIGGRAPH Asia 2010 papers, pp. 1-10 (2010) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.787, + 0.26 + ], + "angle": 0, + "content": "21. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and superresolution. In: ECCV (2016) 10, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.787, + 0.288 + ], + "angle": 0, + "content": "22. Karras, T., Aittala, M., Hellsten, J., Laine, S., Lehtinen, J., Aila, T.: Training generative adversarial networks with limited data. In: Proc. NeurIPS (2020) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.787, + 0.316 + ], + "angle": 0, + "content": "23. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.787, + 0.372 + ], + "angle": 0, + "content": "24. Lattas, A., Moschoglou, S., Gecer, B., Ploumpis, S., Triantafyllou, V., Ghosh, A., Zafeiriou, S.: Avatarme: Realistically renderable 3d facial reconstruction\" in-the-wild\". In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 760-769 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.787, + 0.415 + ], + "angle": 0, + "content": "25. Lattas, A., Moschoglou, S., Ploumpis, S., Gecer, B., Ghosh, A., Zafeiriou, S.: Avatarme++: Facial shape and brdf inference with photorealistic rendering-aware gans. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9269-9284 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.787, + 0.444 + ], + "angle": 0, + "content": "26. Lin, J., Yuan, Y., Zou, Z.: Meingame: Create a game character face from a single portrait. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 311-319 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.787, + 0.472 + ], + "angle": 0, + "content": "27. Lombardi, S., Saragih, J., Simon, T., Sheikh, Y.: Deep appearance models for face rendering. ACM Transactions on Graphics (ToG) 37(4), 1-13 (2018) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.787, + 0.514 + ], + "angle": 0, + "content": "28. Lombardi, S., Simon, T., Schwartz, G., Zollhoefer, M., Sheikh, Y., Saragih, J.: Mixture of volumetric primitives for efficient neural rendering. ACM Transactions on Graphics (ToG) 40(4), 1-13 (2021) 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.787, + 0.557 + ], + "angle": 0, + "content": "29. Luo, H., Nagano, K., Kung, H., Xu, Q., Wang, Z., Wei, L., Hu, L., Li, H.: Normalized avatar synthesis using stylegan and perceptual refinement. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.558, + 0.787, + 0.599 + ], + "angle": 0, + "content": "30. Ma, S., Simon, T., Saragih, J., Wang, D., Li, Y., De La Torre, F., Sheikh, Y.: Pixel codec avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 64-73 (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.787, + 0.627 + ], + "angle": 0, + "content": "31. Mescheder, L., Geiger, A., Nowozin, S.: Which training methods for gans do actually converge? (2018) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.787, + 0.67 + ], + "angle": 0, + "content": "32. Nagano, K., Fyffe, G., Alexander, O., Barbic, J., Li, H., Ghosh, A., Debevec, P.E.: Skin microstructure deformation with displacement map convolution. ACM Trans. Graph. 34(4), 109-1 (2015) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.787, + 0.698 + ], + "angle": 0, + "content": "33. Park, T., Efros, A.A., Zhang, R., Zhu, J.Y.: Contrastive learning for unpaired image-to-image translation. In: European Conference on Computer Vision (2020) 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.787, + 0.727 + ], + "angle": 0, + "content": "34. Pinkney, J.N., Adler, D.: Resolution dependent gan interpolation for controllable image synthesis between domains. arXiv preprint arXiv:2010.05334 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.787, + 0.769 + ], + "angle": 0, + "content": "35. Sang, S., Zhi, T., Song, G., Liu, M., Lai, C., Liu, J., Wen, X., Davis, J., Luo, L.: Agileavatar: Stylized 3d avatar creation via cascaded domain bridging. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-8 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.787, + 0.797 + ], + "angle": 0, + "content": "36. Seymour, M., Evans, C., Libreri, K.: Meet mike: epic avatars. In: ACM SIGGRAPH 2017 VR Village, pp. 1-2 (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "37. Shi, T., Yuan, Y., Fan, C., Zou, Z., Shi, Z., Liu, Y.: Face-to-parameter translation for game character auto-creation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 161–170 (2019) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "38. Song, G., Luo, L., Liu, J., Ma, W.C., Lai, C., Zheng, C., Cham, T.J.: Agilegan: stylizing portraits by inversion-consistent transfer learning. ACM Transactions on Graphics (TOG) 40(4), 1-13 (2021) 2, 4, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.218 + ], + "angle": 0, + "content": "39. Wang, T.C., Liu, M.Y., Tao, A., Liu, G., Kautz, J., Catanzaro, B.: Few-shot video-to-video synthesis (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.788, + 0.259 + ], + "angle": 0, + "content": "40. Wang, T.C., Liu, M.Y., Zhu, J.Y., Tao, A., Kautz, J., Catanzaro, B.: High-resolution image synthesis and semantic manipulation with conditional gans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8798-8807 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.788, + 0.3 + ], + "angle": 0, + "content": "41. Yamaguchi, S., Saito, S., Nagano, K., Zhao, Y., Chen, W., Olszewski, K., Morishima, S., Li, H.: High-fidelity facial reflectance and geometry inference from an unconstrained image. ACM Transactions on Graphics (TOG) 37(4), 1-14 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.788, + 0.342 + ], + "angle": 0, + "content": "42. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networkss. In: Computer Vision (ICCV), 2017 IEEE International Conference on (2017) 2, 4, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.788, + 0.383 + ], + "angle": 0, + "content": "43. Zongsheng Yue, J.W., Loy, C.C.: Resshift: Efficient diffusion model for image super-resolution by residual shifting. In: Advances in Neural Information Processing Systems (NeurIPS) (2023) 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.383 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Bridging the Gap_ Studio-like Avatar Creation from a Monocular Phone Capture/3e2c48bd-a83b-478a-9094-3e6d6e7a5043_origin.pdf b/2024/Bridging the Gap_ Studio-like Avatar Creation from a Monocular Phone Capture/3e2c48bd-a83b-478a-9094-3e6d6e7a5043_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..491330bf4c22b0c813596b482ddb4f2841d25017 --- /dev/null +++ b/2024/Bridging the Gap_ Studio-like Avatar Creation from a Monocular Phone Capture/3e2c48bd-a83b-478a-9094-3e6d6e7a5043_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c458555f147981eff42632288f9721cc855480795993d8cda62279848fdd00d4 +size 87690704 diff --git a/2024/Bridging the Gap_ Studio-like Avatar Creation from a Monocular Phone Capture/full.md b/2024/Bridging the Gap_ Studio-like Avatar Creation from a Monocular Phone Capture/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5222f508e35c0ea221760328eda3b22f67ced77f --- /dev/null +++ b/2024/Bridging the Gap_ Studio-like Avatar Creation from a Monocular Phone Capture/full.md @@ -0,0 +1,337 @@ +# Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture + +ShahRukh Athar $^{1,3*}$ Shunsuke Saito $^{2}$ Zhengyu Yang $^{2}$ Stanislav Pidhorskyi $^{2}$ Chen Cao $^{2}$ + +$^{1}$ Captions Research, New York + +$^{2}$ Meta Reality Labs, Pittsburgh. + +Stony Brook University, New York + +shahrukh@nocapinc.com shunsuke.saito16@gmail.com + +stpidhorskyi@meta.com zhengyu-yang@outlook.com + +zju.caochen@gmail.com + +![](images/60cb2a326479dbe9b1e096bc14e396f5926296187b5943feb1e0de334faca02f.jpg) +Monocular phone capture + +![](images/d82b72dd12f0a0880c5ab26afa11f26e8527e6235cb2e0f3ec5e81c46c63d462.jpg) +Fig. 1: From a low-resolution texture map obtained through monocular phone capture, our method produces a high-resolution texture map with precise facial details, studio-like illumination, and inpainted missing regions. This generated texture map can subsequently be utilized to create a high-quality, photorealistic avatar using the pretrained Universal Prior Model (UPM) from AVA [6]. + +![](images/879fe4f17527076281cf808d47c10465417d11b6df9f5440ada1799aa717b6e6.jpg) +Extracted texture map + +![](images/47983b98dab4ec58dafc3d4015ba511b8b71652884c6895a8633b300568e62d4.jpg) +Generated highresolution texture n + +![](images/ef2231891fc8d89d00f972c47097edf4d01c5112307364c2df8d0f24b419adc9.jpg) +High-quality reanimation and novel view synthesis + +Abstract. Creating photorealistic avatars for individuals traditionally involves extensive capture sessions with complex and expensive studio devices like the LightStage system. While recent strides in neural representations have enabled the generation of photorealistic and animatable 3D avatars from quick phone scans, they have the capture-time lighting baked-in, lack facial details and have missing regions in areas such as the back of the ears. Thus, they lag in quality compared to studio-captured avatars. In this paper, we propose a method that bridges this gap by generating studio-like illuminated texture maps from short, monocular phone captures. We do this by parameterizing the phone texture maps using the $W^{+}$ space of a StyleGAN2, enabling near-perfect reconstruction. Then, we finetune a StyleGAN2 by sampling in the $W^{+}$ parameterized space using a very small set of studio-captured textures as an adversarial training signal. To further enhance the realism and accuracy of facial details, we super-resolve + +the output of the StyleGAN2 using carefully designed diffusion model that is guided by image gradients of the phone-captured texture map. Once trained, our method excels at producing studio-like facial texture maps from casual monocular smartphone videos. Demonstrating its capabilities, we showcase the generation of photorealistic, uniformly lit, complete avatars from monocular phone captures. + +# 1 Introduction + +Photorealistic and animatable avatars are paramount for lifelike human-to-human interactions in AR/VR applications. Creating high-fidelity avatars often requires sophisticated devices like the LightStage capture system that capture human heads with a range of facial expressions [27, 28]. Typically, these captures occur in a studio-like environment, with uniform illumination and densely sampled views to reconstruct complete avatars with consistent lighting. While these methods excel in generating hyper-realistic avatars, they cannot be scaled to millions of people as capturing so many people with a LightStage-like capture system is impractical and non-trivial. + +Recently, there has been extensive work in generating photorealistic avatars through a monocular capture [3,6,14]. These methods involve scanning various facial expressions and head poses of the user to reconstruct a 3D avatar that closely aligns with the captured data. However, a notable limitation of these methods is that the captured lighting is embedded in the avatars, causing their appearance to be heavily dependent on the capture devices and surrounding environments. Furthermore, due to the constraints of the capture setup, certain areas, such as the back of the head or ears, are never visible, leaving visible holes and artifacts when viewed during animation. + +An innovative strategy to tackle this challenge is to use image-to-image translation to transform the phone data into studio-captured data. This transformation can be learnt through supervised training using paired data [18], or through unsupervised training using unpaired datasets [42]. Due to reasons mentioned earlier, creating a large-scale paired dataset of studio and phone captures is impractical, which rules out the possibility of using supervised image-to-image translation methods. On the flip side, current unsupervised image-to-image translation methods fail to preserve fine detail in the transformation process [38, 42] which is paramount to creating a high-fidelity avatar. + +In this paper, we introduce a method capable of generating studio-like, high-quality avatars from monocular phone captures. We do this by parameterizing a large-scale dataset of phone-captured face texture maps using the $W^{+}$ space of a StyleGAN2. Then, we finetune this StyleGAN2 by sampling from this parametrized $W^{+}$ space using a small set of unpaired studio-captured texture maps, to create a Studio-StyleGAN2 model. Our key insight is that sampling from the $W^{+}$ space instead of Z space leads to a more generalizable model as samples from the $W^{+}$ are, by construction, as diverse as the training set while samples from the Z space often suffer from mode-collapse. During inference, the given phone-captured texture map, is first inverted to the $W^{+}$ space of the StyleGAN2. This inverted $W^{+}$ vector is then given as input to the Studio-StyleGAN2 to generate a low-resolution studio-like texture map. Finally, a novel facial detail conditioned diffusion model is used to enhance facial details of the low-resolution + +studio-like texture map obtained in the previous step. Once trained, our method excels at producing high-quality studio-like face texture maps from monocular phone captures, which are then used as inputs to the universal face prior model (UPM) from Authentic Volumetric Avatars (AVA) [6] to generate photorealistic avatars. In summary, our key contributions are as follows: + +- Introducing a groundbreaking method for creating studio-like, photorealistic avatars from monocular phone captures. +- Innovative finetuning of a pre-trained generative model using a minimal dataset from another domain by sampling in the inverted $W^{+}$ space, which facilitates the development of a generative model for a new domain while preserving the integrity of the $W^{+}$ latent space. +- A novel diffusion model conditioned on the phone texture gradient, that is designed to upsample studio textures, effectively enhancing facial details and contributing to the overall realism of the generated avatars. + +# 2 Related Work + +Our paper aims to bridge the gap between a studio capture and a phone capture by creating a studio-like, photorealistic avatar through a monocular phone capture. In delineating the related work, we offer a comprehensive overview of key research domains, including studio-captured avatars, phone-captured avatars, and image-to-image translation. + +Studio-captured avatars. The reconstruction of high-fidelity static and dynamic models of the human head based on photometric measurements has a long-standing history in computer graphics and vision. Achieving a photorealistic human avatar often requires specialized hardware in high-end production, such as the LightStage system. To model the complex skin appearance, various physically-based models have been explored. Notably, subsurface scattering [5], linear polarization patterns [12] and fine-scale skin details [1, 2] have been investigated. For dynamic expression details, Jimenez et al. [20] compute dynamic skin appearances by blending hemoglobin distributions captured with different expressions. In their subsequent work [19], expression-dependent normal maps are interpolated to add realistic wrinkles to an animated face. Nagano et al. [32] synthesize skin microstructures based on local geometric features derived from high-precision microgeometry, acquired with an LED sphere and a skin deformer. While these methods have been instrumental for offline movie production, their substantial compute requirements make them less suitable for real-time applications. Despite recent efforts to enable real-time rendering of physically-based avatars [36], heavy compute remains challenge. In response to the challenges posed by complex physical computations, researchers have proposed a deep appearance model [27]. This model utilizes a coarse 3D triangle mesh in conjunction with view-dependent texture mapping. The texture is regressed by a neural network conditioned on viewpoint and expression latent codes. This conditioning accounts for view- and expression-dependent variations while compensating for the imperfect proxy geometry. Subsequent work extends the mesh-texture representation to a volumetric representation using a Mixture of Volumetric Primitives (MVP) [28], further enhancing the model's quality. Pixel Codec Avatars (PiCA), as demonstrated by [30], showcase the efficiency of rendering such models, even on mobile hardware platforms, + +by leveraging efficient per-pixel processing. Moreover, Bi et al. [4], based on relighting captured data, can also relight avatars with any novel point light or environment maps. While these methods can achieve hyper-realistic avatars, their studio requirements make them challenging to generalize to ordinary users. + +Phone-captured avatars. There are several methods aimed at creating avatars in lightweight ways, even from a phone or a single image. Some of these approaches focus on generating stylized avatars based on different input requirements, ranging from capturing multi-view images using a phone [17] to utilizing a single monocular image [26, 29, 35, 37]. While these methods excel in producing animatable avatars, it is important to note that their appearance tends to be cartoonish and lacks realism. + +Another category of methods is dedicated to creating realistic human avatars based on graphics pipelines [7, 8, 11, 15, 24, 25, 41]. While relying on traditional graphics methods, these approaches often result in avatars that appear uncanny. Building upon deep learning-based representations, researchers have proposed methods to generate avatars with increased realism, including notable works by Gafni et al. [10] and Grassal et al. [14]. Among these, the work of Cao et al. [6], Authentic Volumetric Avatars (AVA), stands out for its focus on creating photorealistic avatars from phone scans. The process involves training a Universal Prior Model (UPM) using studio-captured data, followed by personalizing this UPM using data from a phone scan of an unseen subject. While the method successfully produces avatars with realistic appearance and animation, it is worth noting that the lighting is baked into the personalized avatar, and certain details, such as the back of the ears, may be missing. + +Image-to-image translation We use image-to-image translation to map images from the source domain (phone data) to the target domain (studio data). Isola et al. introduced Pix2Pix [18], a method that utilizes adversarial training strategies [13] to achieve this mapping. Additionally, Wang et al. [40] focused on increasing the resolution of generated results from semantic label maps. While these methods successfully map images between domains, it is essential to note that they require paired training data for effective implementation. In many real-world scenarios, obtaining paired training data can be challenging and expensive. + +To address this issue, unsupervised image-to-image translation has been introduced. Zhu et al. [42] proposed a novel cycle-consistency loss to ensure that translating an image from one domain to another and back again should result in the original image. This helps the model maintain consistency and produce more realistic translations. Subsequent methods have further improved unsupervised image-to-image translation from different perspectives, including a multimodal model [16], few-shot input for video-to-video translation [39], translation of images with human control [34], and translation of real images into different styles [38]. However, none of these methods are designed with preservation of facial identity and the generation of realistic facial details in mind, especially with such little training data. Later in the paper we show how this prior work compares to ours for transforming low-resolution phone-captured texture maps to high-resolution studio-captured texture maps. + +![](images/1e0379f9076c396cc9e0ae4f9f76d160176b7b478e0ba1284fb2e6ad8f5bfb87.jpg) +Fig. 2: Method Overview. Our method employs a two-step process to train $\mathcal{G}_{\text{Studio}}$ . Initially, we train a StyleGAN2 on $12k$ neutral texture maps captured by phones, yielding $\mathcal{G}_{\text{phone}}$ . Subsequently, we initialize $\mathcal{G}_{\text{Studio}}$ with the weights of $\mathcal{G}_{\text{phone}}$ and fine-tune it by sampling from $S_{W^{+}} = \{W_{\mathcal{I}_0^{\text{phone}}}, \ldots, W_{\mathcal{I}_{N-1}^{\text{phone}}}^{+}\}$ ; where $N = 12k$ , and $W_{\mathcal{I}_i^{\text{phone}}}^{+}$ represents the vector obtained by inverting the $i$ 'th phone-captured texture map, $\mathcal{I}_0^{\text{phone}}$ , in the $W^{+}$ space of $\mathcal{G}_{\text{phone}}$ . During inference, the given phone-captured texture map is inverted to $W^{+}$ , then passed to $\mathcal{G}_{\text{Studio}}$ to generate a low-resolution studio-lit texture map, $\mathcal{I}^{*}$ . Accurate facial details are subsequently added using diffusion model $f_{\phi}$ , conditioned on the gradient of the phone texture. This process results in the final high-resolution, studio-lit, and completed texture map, $\mathcal{I}^{*}$ . + +![](images/31a1533d559db413d9f515be8a5f77c709c0ef57be5ddbed4b846be033f77875.jpg) + +# 3 Method + +In this section, we present our method for generating studio-like avatars from the phone captures. Our approach consists of two key components: a StyleGAN2 for texture translation and a diffusion model for facial detail generation. In Sect 3.1, we describe the generation of a low-resolution texture map with studio-like lighting and missing regions inpainted. First, we pretrain a StyleGAN2 on 12k phone textures and then finetune it using a small set of studio-captured texture maps. In order to improve the generalization of the finetuned StyleGAN2, we optimize it by sampling in the $W^{+}$ space, instead of the W-space or Z-space, using a set of 12k $W^{+}$ vectors obtained by inverting the phone captured texture maps. In Sect 3.2, we introduce a diffusion model that generates facial details. The diffusion model takes the output from the aforementioned StyleGAN2 and generates a high-resolution neutral texture with realistic facial details. After obtaining a high-resolution studio-like neutral texture from our method, we use it to learn a color transform to transfer phone-captured expression textures to studio-lit expression textures. These expressions are subsequently utilized for animating a high-quality avatar (Sect 3.3). Notably, we intentionally avoid applying inpainting or super-resolution techniques to the expression textures. Fig. 2 provides an overview of our method. + +# 3.1 Illumination manipulation and inpainting + +In our initial step, we track the geometry from a monocular phone capture of the user's neutral face and extract the neutral texture $\mathcal{L}^{phone}$ from the captured image, employing the method outlined in [6]. Subsequently, we translate this phone-captured texture with in-the-wild lighting and missing regions into a texture map with studio lighting and + +missing regions in-painted. This translation is accomplished by parametrizing $\mathcal{I}^{phone}$ using the $W^{+}$ space of a StyleGAN2, as follows: + +$$ +W _ {\mathcal {I} ^ {p h o n e}} ^ {+} = \underset {W ^ {+}} {\operatorname {a r g m i n}} \left\| \mathcal {G} _ {p h o n e} \left(W ^ {+}\right) - \mathcal {I} ^ {p h o n e} \right\| _ {2} ^ {2} + L P I P S \left(\mathcal {G} _ {p h o n e} \left(W ^ {+}\right), \mathcal {I} ^ {p h o n e}\right), \tag {1} +$$ + +where $\mathcal{G}_{phone}$ is a StyleGAN2 trained on phone-captured textures, and $W_{\mathcal{T}^{phone}}^{+}$ is the optimized vector in the $W^{+}$ space of $\mathcal{G}_{phone}$ . Now, we generate the low-resolution studio-like texture as follows: + +$$ +\mathcal {I} ^ {*} = \mathcal {G} _ {\text {S t u d i o}} \left(W _ {\mathcal {I} _ {\text {p h o n e}}} ^ {+}\right), \tag {2} +$$ + +where $\mathcal{G}_{\text{Studio}}$ is a StyleGAN2 responsible for generating low-resolution studio-like textures, while $\mathcal{I}^*$ refers to the studio-lit version of $\mathcal{I}^{\text{phone}}$ with the inpainted missing regions. Ideally, $\mathcal{I}^*$ should retain the identity and semantics of $\mathcal{I}^{\text{phone}}$ , with the only distinctions being the illumination and the inpainting of missing regions. Below we describe the training procedure for $\mathcal{G}_{\text{Studio}}$ . + +We initialize the synthesis network of $\mathcal{G}_{\text{Studio}}$ using weights from $\mathcal{G}_{\text{phone}}$ , which we then finetune to generate texture maps with studio lighting and inpaint the missing regions. We choose to finetune $\mathcal{G}_{\text{phone}}$ for the following reasons: 1) Due to the immense costs of a studio capture, studio quality training data is typically very limited. In our case, we only have 383 studio-captured texture maps, making training from scratch hard [22]; 2) Since we want $\mathcal{I}^*$ and $\mathcal{I}^{phone}$ to have the same facial identity and semantics meaning, it is beneficial to learn a latent space that is shared between $\mathcal{G}_{\text{Studio}}$ and $\mathcal{G}_{\text{phone}}$ . For finetuning, we first invert the entire dataset of $N$ phone-captured texture maps using Eq. (1), giving us a set of $N$ vectors in the $W^{+}$ space: $S_{W^{+}} = \{W_{\mathcal{I}_0^{phone}}^{+}, \ldots, W_{\mathcal{I}_{N-1}^{phone}}^{+}\}$ . Next, we randomly sample vectors from the inverted set, $S_{W^{+}}$ , and finetune the synthesis network of $\mathcal{G}_{\text{phone}}$ to give $\mathcal{G}_{\text{Studio}}$ . We choose to finetune the generator with samples from inverted $W^{+}$ instead of the traditional $W$ space, since they are, by construction, from the real data distribution and are consequently more diverse, leading to better generalization of $\mathcal{G}_{\text{Studio}}$ . The losses we use during finetuning are described as below. + +Studio-Discriminator loss. This loss uses a discriminator to ensure the texture maps generated by $\mathcal{G}_{\text{Studio}}$ are from the distribution of studio-captured texture maps: + +$$ +\begin{array}{l} \mathcal {L} _ {A d v} = \mathbb {E} _ {\mathcal {T S t u d i o} \sim P (\mathcal {T S t u d i o})} [ \min (0, - 1 + \mathcal {D} _ {\text {S t u d i o}}) ] \\ + \mathbb {E} _ {W _ {\mathcal {I} _ {i} ^ {p h o n e}} ^ {+} \sim S _ {W ^ {+}}} \left[ \min (0, - 1 - \mathcal {D} _ {\text {S t u d i o}} \left(\mathcal {G} _ {\text {S t u d i o}} \left(W _ {\mathcal {T} _ {i} ^ {p h o n e}} ^ {+}\right)\right) \right], \tag {3} \\ \end{array} +$$ + +where $\mathcal{D}_{\text{Studio}}$ is a discriminator that initialized from the pretrained phone-captured texture StyleGAN2 Discriminator and finetuned using ground-truth studio-captured textures and textures generated by $\mathcal{G}_{\text{Studio}}$ . + +Face-Identity Loss This loss compares the face identity embeddings of the renders of texture maps generated from $\mathcal{G}_{\text{Studio}}$ and $\mathcal{G}_{\text{phone}}$ in order to ensure the identity is preserved: + +$$ +\mathcal {L} _ {\text {F a c e I D}} = \left\| \left| \mathcal {F} \left(\mathcal {G} _ {\text {S t u d i o}} \left(W _ {\mathcal {I} _ {i} ^ {\text {p h o n e}}} ^ {+}\right)\right) - \mathcal {F} \left(\mathcal {G} _ {\text {p h o n e}} \left(W _ {\mathcal {I} _ {i} ^ {\text {p h o n e}}} ^ {+}\right)\right) \right| \right\| _ {2} ^ {2}, \tag {4} +$$ + +where $\mathcal{F}$ is a pretrained face recognition network1. + +Perceptual Loss This loss encourages the preservation of semantic features between the texture maps generated $\mathcal{G}_{\text{Studio}}$ and $\mathcal{G}_{\text{phone}}$ by minimizing their distance in the VGG feature space + +$$ +\mathcal {L} _ {\text {P e r c p}} = \operatorname {L P I P S} \left(\mathcal {G} _ {\text {S t u d i o}} \left(W _ {\mathcal {I} _ {i} ^ {\text {p h o n e}}} ^ {+}\right), \mathcal {G} _ {\text {p h o n e}} \left(W _ {\mathcal {I} _ {i} ^ {\text {p h o n e}}} ^ {+}\right)\right). \tag {5} +$$ + +Perceptual Reconstruction Loss This loss ensures that skintones are preserved using a small amount of paired data (i.e. subjects for whom we have both phone and studio textures maps i.e both $\mathcal{I}^{phone}$ and $\mathcal{I}^{Studio}$ ) + +$$ +\mathcal {L} _ {\text {P e r c p - R e c o n s}} = \frac {1}{k} \sum_ {i = 1} ^ {K} \operatorname {L P I P S} \left(\mathcal {G} _ {\text {S t u d i o}} \left(W _ {\mathcal {T} _ {i} ^ {\text {p h o n e}}} ^ {+}\right), \mathcal {I} _ {i} ^ {\text {S t u d i o}}\right), \tag {6} +$$ + +where $\mathcal{I}_i^{Studio}$ and $\mathcal{I}_i^{phone}$ are the ground-truth studio and ground-truth phone texture maps of the same person. We only have a very small amount of this data for training (K=83). R1-Regularization: We regularize $\mathcal{D}_{Studio}$ using the R1 regularization [31] as follows: + +$$ +\mathcal {L} _ {R 1} = - \frac {\gamma}{2} \mathbb {E} _ {\mathcal {I} ^ {S t u d i o} \sim P (\mathcal {I} ^ {S t u d i o})} \left[ \left\| \nabla \mathcal {D} _ {S t u d i o} (\mathcal {I} ^ {S t u d i o}) \right\| _ {2} ^ {2} \right], \tag {7} +$$ + +where $\gamma = 10$ + +In order to further encourage identity preservation, we leverage the multi-resolution architecture of StyleGAN2 [23] and only optimize the generator parameters after the $8 \times 8$ resolution during the finetuning process. The intuition being that identity-specific information is mostly stored in the low-resolution maps, we ablate this choice in the Supplementary section of the paper. The final optimization for the joint finetuning of $\mathcal{G}_{\text{Studio}}$ and $\mathcal{D}_{\text{Studio}}$ is the following: + +$$ +\min _ {\mathcal {G} _ {S t u d i o} ^ {\theta (8 +)}} \max _ {\mathcal {D} _ {S t u d i o}} \mathcal {L} _ {A d v} + \mathcal {L} _ {R 1} + \mathcal {L} _ {P e r c p - R e c o n s} + \lambda_ {1} \mathcal {L} _ {P e r c p} + \lambda_ {2} \mathcal {L} _ {F a c e I D}, \tag {8} +$$ + +where $\lambda_{1} = 0.5$ and $\lambda_{2} = 10$ . We ablate $\mathcal{L}_{Percp-Recons}, \mathcal{L}_{Percp}, \mathcal{L}_{FaceID}$ in the supplementary. + +# 3.2 Synthesis of accurate facial details + +Generating high-quality facial details. $\mathcal{G}_{\text{Studio}}$ can produce an inpainted texture map, $\mathcal{I}^*$ , with illumination consistent with that of a studio capture. However, due to the limited amount of finetuning data, $\mathcal{G}_{\text{Studio}}$ often struggles to reproduce facial details that are visibly present in $\mathcal{I}^{\text{phone}}$ itself. This limitation results in oversmoothed and inaccurate avatars. Achieving realistic facial details necessitates an accurate modeling of the facial texture distribution, a task that the StyleGAN2 fails to do. + +Motivated by the recent success of latent diffusion models in accurately modeling data distributions, we propose the task of generating facial details as the reverse process of a Markov chain that transforms a low-resolution face texture map to a high-resolution one. We adopt the formulation from [43], where the residual between the low-res and high-res studio-captured texture maps, denoted as $e_0 = \mathcal{I}_{LR}^{Studio} - \mathcal{I}^{Studio}$ , is used to define the Markov process. More specifically, the forward process is defined as follows: + +$$ +q \left(\mathcal {I} _ {t} ^ {\text {S t u d i o}} \mid \mathcal {I} _ {t - 1} ^ {\text {S t u d i o}}, \mathcal {I} _ {L R} ^ {\text {S t u d i o}}\right) = \mathcal {N} \left(\mathcal {I} _ {t} ^ {\text {S t u d i o}}; \mathcal {I} _ {t - 1} ^ {\text {S t u d i o}} + \alpha_ {t} e _ {0}, \kappa^ {2} \alpha_ {t} I\right), \tag {9} +$$ + +$$ +t = 1, 2, \dots , T, +$$ + +where $\alpha_{t}$ controls the schedule with which the residual is added, $\kappa$ is a hyperparameter controlling the noise schedule and $I$ is the identity matrix. The reverse process is defined as + +$$ +q \left(\mathcal {I} _ {t - 1} ^ {S t u d i o} \mid \mathcal {I} _ {\{t, L R \}} ^ {S t u d i o}, \mathcal {I} ^ {S t u d i o}\right) = \mathcal {N} \left(\mathcal {I} _ {t - 1} ^ {S t u d i o} \left| \frac {\eta_ {t - 1}}{\eta_ {t}} \mathcal {I} _ {t} ^ {S t u d i o} + \frac {\alpha_ {t}}{\eta_ {t}} \mathcal {I} ^ {S t u d i o}, \kappa^ {2} \frac {\eta_ {t - 1}}{\eta_ {t}} \alpha_ {t} I\right), \right. \tag {10} +$$ + +where $\eta_{t} = \alpha_{t} + \eta_{t - 1}$ . We refer the reader to [43] for details regarding the formulation of both the forward and reverse process. Finally, the diffusion model, $f_{\phi}$ , is trained to minimize the following objective + +$$ +\min _ {\phi} \sum_ {t} \| f _ {\phi} \left(\mathcal {I} _ {t} ^ {S t u d i o}, \mathcal {I} _ {L R} ^ {S t u d i o}, t\right) - \mathcal {I} ^ {S t u d i o} \| _ {2} ^ {2} \tag {11} +$$ + +During inference, we use $f_{\phi}$ to add realistic facial details to the low-resolution output of $\mathcal{G}_{\text{Studio}}$ as follows: + +$$ +\hat {\mathcal {I}} ^ {*} = \operatorname {R e v e r s e P r o c e s s} \left(f _ {\phi}, \mathcal {I} ^ {*}\right). \tag {12} +$$ + +Recovering accurate facial details. While $f_{\phi}$ enhances $\mathcal{I}^*$ by adding realistic facial details, it struggles to recover details already present in the phone-captured texture map $\mathcal{I}^{phone}$ due to the inherent lack of such details in $\mathcal{I}^*$ . The loss of details occurs during the illumination manipulation and inpainting process, as described in Eq. (1) and Eq. (2). To address this issue, we incorporate the image gradient from the phone-captured texture map into the low-resolution texture map during the training of the diffusion model, as follows: + +$$ +\mathcal {I} _ {L R} ^ {\text {S t u d i o *}} = \mathcal {I} _ {L R} ^ {\text {S t u d i o}} + \mathbb {G} (\mathcal {I} ^ {\text {p h o n e}}) \tag {13} +$$ + +where $\mathcal{I}^{phone}$ represents the phone-captured texture map, and $\mathcal{I}^{Studio}$ is the studio-captured texture map of the same person. $\mathbb{G}$ denotes the operator used to calculate the image gradient. During training, $\mathcal{I}_{LR}^{Studio*}$ replaces $\mathcal{I}_{LR}^{Studio}$ in Eq. (9), Eq. (10), and Eq. (11). In the inference stage, we augment $\mathcal{I}^*$ by adding the gradient of the phone-captured texture map as follows: + +$$ +\hat {\mathcal {I}} ^ {*} = \text {R e v e r s e P r o c e s s} \left(f _ {\phi}, \mathcal {I} ^ {*} + \mathbb {G} \left(\mathcal {I} ^ {\text {p h o n e}}\right)\right) \tag {14} +$$ + +Implementation details We utilize 83 paired texture maps, representing subjects for whom we have both $\mathcal{I}^{Studio}$ and $\mathcal{I}^{phone}$ , to train the diffusion model. Following the approach in [43], we employ a latent diffusion model. To prevent overfitting, training is conducted on random $512 \times 512$ crops of the $1024 \times 1024$ resolution texture map. During the inference stage, we employ the full-resolution texture map. + +# 3.3 Driving a high-quality avatar + +With the studio-lit version $\mathcal{I}^*$ now available for a given neutral phone-captured texture map $\mathcal{I}^{phone}$ , we proceed to estimate a color transform mapping from the phone-captured texture map to the studio-lit texture map as follows: + +$$ +\left\{G, B \right\} \leftarrow \underset {\{G, B \}} {\operatorname {a r g m i n}} \left| \left| \mathcal {I} ^ {*} - \left(\mathrm {R s z} (G) \times \mathcal {I} ^ {\text {p h o n e}} + \mathrm {R s z} (B)\right) \right| \right|. \tag {15} +$$ + +Here, $G$ and $B$ represent gain and bias maps of resolution $k \times k$ and $\mathbb{R} \times \mathbb{Z}$ is the resizing operator with bilinear interpolation. Utilizing $G$ and $B$ , we perform a transformation + +on phone-captured expression texture maps to achieve studio-like lighting, as outlined below: + +$$ +\mathcal {I} _ {e x p} ^ {*} = \operatorname {R s z} (G) \times \mathcal {I} _ {e x p} ^ {\text {p h o n e}} + \operatorname {R s z} (B). \tag {16} +$$ + +In our experiment, we select the value of $k = 32$ to efficiently transform the lighting while preserving the details. + +Now, the studio-like high-resolution neutral texture $\hat{I}^*$ serves as conditioning data for the universal avatar prior from [6]. By combining the expression code generated by inputting $\mathcal{I}_{exp}^{*}$ into the expression encoder from [6], we can render a high-quality avatar from any desired view $v$ as follows: + +$$ +\mathbf {I} = \operatorname {A V A} \left(\hat {\mathcal {I}} ^ {*}, \mathcal {I} _ {e x p} ^ {*} - \mathcal {I} ^ {*}, v, \mathbf {F}\right), \tag {17} +$$ + +where $\mathbf{I}$ represents the render of the avatar, and $\mathbf{F}$ is the geometry generated during 3D face tracking for a monocular phone capture. AVA corresponds to the inference process of the universal prior model. For more details about the universal prior model, please refer to [6]. + +# 4 Results + +In this section, we introduce the dataset used in this paper, along with the baselines of our method. Subsequently, we present both quantitative and qualitative results of our method, comparing it to prior work. All phone-captured texture maps are generated using the mesh fitting procedure outlined in Authentic Volumetric Avatars (AVA) [6]. Furthermore, we utilize AVA to render avatars based on the texture maps generated by all the methods, facilitating the calculation of image space metrics, including face identity similarity. + +Table 1: Quantitative results on paired phone-cum-studio captured texture maps. Our method outperforms prior work across all metrics. Best and Second Best scores are highlighted. + +
ModelsPSNR ↑SSIM ↑LPIPS ↓DISTS ↓
Ours22.760.7260.3640.163
AgileGAN [38]18.050.6570.4060.180
CUT [33]21.4400.6420.4020.169
CycleGAN [42]21.0870.6430.4000.175
+ +Table 2: FaceID results on unpaired phone captured texture maps. Our method better preserves identity than prior work without sacrificing the quality of the generated texture maps. Best and Second Best scores are highlighted. + +
ModelsOursAgileGANCUTCycleGAN
FaceID4.31e-41.36e-36.89e-45.19e-4
+ +Training and Evaluation Data. We utilize a dataset comprising 12,543 neutral phone-captured texture maps to train $\mathcal{G}_{phone}$ , and 383 studio-captured texture maps for fine-tuning $\mathcal{G}_{phone}$ to obtain $\mathcal{G}_{Studio}$ . Among the 383 maps, 83 are paired neutral texture maps $\{\mathcal{I}^{phone},\mathcal{I}^{Studio}\}$ used for calculating the perceptual reconstruction loss, as described in Eq. (6). This paired dataset also serves for training the detail-preserving diffusion model. For quantitative and qualitative evaluation, we employ 10 paired phone-cum-studio captured texture maps and 31 phone-captured texture maps. + +Baselines. We compare our method to the following prior works on unpaired image-to-image translation. 1) AgileGAN [38], which utilizes an aligned StyleGAN latent space + +to stylize images, even with very few examples; 2) Contrastive Unpaired Translation (CUT) [33], which employs a patch-based contrastive loss in a learned feature space, enabling domain adaptation, even for single images; 3) CycleGAN [42], which trains two generators based on a cycle-consistency loss to translate images between two domains. Since AgileGAN [38] has no code available, we implement it ourselves. We use publicly available code for CUT [33] and [42] for all experiments. + +![](images/01c936b38159dfa3bcf75574d26e9bf90d555086149ddf43078d61cb2bff9b37.jpg) +Fig. 3: Comparisons with baselines on paired phone-cum-studio texture data reveal that our method produces results closest to the ground truth. It achieves uniform studio-like lighting, well-reconstructed facial details visible in the phone capture, and effective inpainting of missing regions. In contrast, AgileGAN [38] fails to preserve identity, while CUT [33] and CycleGAN [42] introduce significant artifacts. + +Quantitative Results. In Table 1, we present the results of a quantitative evaluation comparing our method to the baselines using 10 paired texture maps captured with both phones and in a studio setting. The evaluation metrics include mean PSNR, SSIM, LPIPS [21], and DISTS [9]. DISTS and LPIPS are perceptual metrics aligned with human judgment, while PSNR and SSIM are pixel-based metrics. As shown in Table 1, our method consistently outperforms the baseline across all metrics. The improvement is not limited to perceptual metrics but also extends to PSNR and SSIM, highlighting the efficacy of our proposed approach. It is worth noting that the relatively low PSNR and SSIM scores can be attributed to their sensitivity to small pixel shifts between the studio and phone-captured textures. + +In Table 2, we display the face embedding distances, measured using Eq. (4), between avatars rendered from the phone-captured texture and the textures generated by various methods on the 31 unpaired phone captures. Our method exhibits the best preservation of facial identity. + +![](images/5d5a5b29b3ae5edc8c829d17bda2c83d0a82badbeba3db073e3f249d7789f125.jpg) +Fig. 4: Comparisons on Unpaired Phone-Captured Data: In comparison to prior work, our method excels in generating texture maps with superior preservation of identity, enhanced photorealism in facial details, more uniform illumination, and improved inpainting of missing regions. + +Qualitative Results. In Fig 3, we provide some qualitative results on the paired test data. It is evident that our method generates a more plausible and photorealistic texture map compared to prior work, showcasing improvements in illumination transfer, facial details, and inpainting of missing regions. + +In Fig 4, we present qualitative results comparing our method with prior work. While AgileGAN [38] successfully changes the illumination to be studio-like and inpaints missing regions, it introduces a significant identity shift (quantitatively measured in Table 2) and lacks facial details. We attribute this to the use of the $Z^{+}$ space [38], which may not be flexible enough for high-fidelity inversion, and the absence of identity-preserving constraints, both in architecture and optimization, during training. Due to its contrastive training paradigm, CUT [33] preserves identity better than AgileGAN but introduces significant artifacts. It is also unable to inpaint the missing regions around the ears and corners of the head. Similarly, like CUT [33], CycleGAN [42] also preserves identity better than AgileGAN [38] but struggles with inpainting missing regions. The textures also contain numerous uncanny artifacts that are uncharacteristic of human skin. In contrast, our method generates a high-quality studio-illuminated texture map with accurate facial details and inpainted missing regions. + +![](images/5a8c586c7d97e1c4cbef53a27ac8700df4a6a5cd48875a36b4568969354b7e69.jpg) +Fig. 5: Avatar Reanimation. The top row showcases a multiview render of an avatar generated using studio-like neutral and expression texture maps, created through our method as described in Sect 3.3. In contrast, the bottom row utilizes phone-captured texture maps. Notably, the lip region, highlighted with pink rectangles, appears more realistic and less blurry when using studio-like texture maps generated by our method. + +![](images/961325b567e324f045bca7a2d849bafe8e750a0a420f5ab6b43435330c5bde4b.jpg) +Fig. 6: Comparison of Avatars Generated Using [6]: In this comparison, we evaluate the quality of avatars generated by the Universal Prior Model (UPM) from AVA [6], using texture maps from various methods as input. Evidently, the avatars generated using texture maps from our method far exceed those generated by prior work. + +Reanimation Results. As explained in Sect 3.3, we can transform the phone expression texture maps $\mathcal{T}_{exp}^{phone}$ to studio-like illumination using Eq. (15) and Eq. (16). We observe that utilizing studio-like illuminated expression texture maps $\mathcal{T}^{*}_{exp}$ results in a modest improvement in the quality of reanimated avatars using the universal prior model from AVA [6]. In Fig 5, we present an example where the top row shows the render of an avatar using studio-like neutral and expression textures, while the bottom row displays a render using phone-captured neutral and expression textures. It is evident that the lip region appears more realistic and less blurry when reanimated with the studio-like neutral and expression textures compared to the phone-captured textures. We recommend that readers refer to our supplementary video for a more comprehensive comparison. + +# 4.1 Ablations + +In this section, we conduct ablation studies to evaluate the various components of our method. + +# Details Conditioned Diffusion. + +We ablate the necessity of using facial details extracted from the phone capture texture map, in the form of an image gradient, to synthesize accurate facial details using a diffusion model. We explore three scenarios: + +1) A Diffusion model that does not use any conditioning on phone-captured texture gradient (Vanilla Diffusion); 2) A Diffusion model that uses the phone-captured texture gradient only dur + +ing inference but not during training; 3) Our model, which incorporates the phone-captured texture gradient during both training and inference. + +In Fig 7, we present the results of each of the three scenarios. While the Vanilla Diffusion generates a realistic-looking facial texture, it fails to preserve facial details present in the phone-captured texture map. Prominent moles, marked by the blue boxes in the phone-captured texture map, are not generated by the Vanilla Diffusion model. When conditioning the reverse process using the phone-texture gradient only during inference, we observe that the model misses some details and transfers shading from the phone-captured texture map to the studio-lit texture map (marked by red boxes in Fig 7), which is undesirable. Ideally, we want facial details to be preserved while eliminating illumination-dependent effects, such as strong + +Table 3: Detail conditioning ablation. We calculate quantitative metrics on the 10 paired phone-cum-studio captured texture maps. It is evident that utilizing detail conditioning during both training and inference yields the best performance. Best and Second Best scores are highlighted. + +
ModelsLPIPS ↓DISTS ↓
Vanilla Diffusion0.3830.179
Inference-only Details conditioning0.3760.183
Training + Inference Details conditioning (Ours)0.3640.163
+ +Table 4: $W^{+}$ vs. $Z$ Space ablation. We utilize the 10 paired phone-cum-studio captured texture maps to compare the outcomes of training in the $W^{+}$ and $Z$ spaces. The results clearly demonstrate that training in the $W^{+}$ space yields significantly better outcomes. + +
ModelsPSNR ↑SSIM ↑LPIPS ↓DISTS ↓
Z Space21.3680.7180.4010.172
W+ Space (Ours)22.760.7260.3640.163
+ +shading and shadows, from the studio-lit texture map. Our method conditions the diffusion model on the phone-texture gradient during both inference and training. As shown in Fig 7, this allows the model to learn to retain facial details while ignoring illumination-induced effects, such as strong shading, shadows, and specularities when synthesizing facial details. In Table 3, we quantitatively validate each of the aforementioned diffusion model designs on paired data using LPIPS [21] and DISTS [9]. As can be seen, conditioning the diffusion model on phone-texture gradients both during training and inference gives the best results. + +$W^{+}$ space vs. $Z$ space for finetuning. We now evaluate the effectiveness of the $W^{+}$ space over the standard $Z$ space for fine-tuning $\mathcal{G}_{\text{Studio}}$ . For a quantitative comparison, we calculate the FaceID loss over the 31 unpaired evaluation texture maps and compute PSNR, SSIM, LPIPS [21], and DISTS [9] over the 10 paired studio-cum-phone texture maps. Qualitative results, along with the average FaceID distance over unpaired data, are shown in Fig 8, and the metrics on paired data are presented in Table 4. As evident, sampling in the $W^{+}$ space generalizes much better to unseen identities, exhibiting + +![](images/190ce0af12d939452e07756539f943fafe4677c1219e171617e66fc70c4e83eb.jpg) +Fig. 7: Ablation on Detail-conditioned Diffusion. When no detail conditioning is applied, facial details visible in the phone capture are not reproduced. When detail conditioning is employed only during inference, undesirable shading effects are transferred to the generated texture map. Optimal results, with the most accurate and plausible reproduction of facial details, are achieved when detail conditioning is applied during both training and inference. Please refer to the text for further details. + +![](images/443bdafb5b111eace363ef8fe789a930e970f627223ae84a622e8346efa6040d.jpg) +Fig. 8: Qualitative $W^{+}$ space ablation. We observe that when $\mathcal{G}_{\text{Studio}}$ is finetuned using the $Z$ space, it fails to generalize to novel subjects and exhibits significant artifacts. In contrast, training $\mathcal{G}_{\text{Studio}}$ in the $W^{+}$ lends it far better generalization. + +significantly fewer artifacts and yielding superior results in terms of FaceID, PSNR, SSIM, LPIPS, and DISTS. We posit that this is because samples in the $W^{+}$ space (i.e., samples from $S_{W^{+}}$ space) are near-perfect inversions of the training data, making them more diverse than those generated from the Gaussian-distributed $Z$ space. Consequently, as seen in Fig 8, this leads to better generalization to unseen subjects. The model trained using the $Z$ space exhibits uncanny artifacts in its results. + +# 5 Conclusion and Limitations + +In this paper, we present a method for generating studio-like, high-quality avatars from monocular phone captures, using a StyleGAN2-based image-to-image translation and diffusion-based image upsampling. Experiments show the effectiveness of our approach in manipulating lighting, inpainting missing parts and generating facial details thus enabling the creation of complete, studio-lit textures for rendering high-quality avatars. However, our method does have limitations. It struggles with input textures exhibiting extreme non-uniform lighting due to the constrained lighting conditions in our training data (we include examples in the supplementary). We also do not fine-tune the universal prior model to personalize the avatar based on phone capture data, as suggested in [6], thus the avatars lack personalized details for different facial expressions. Finally, our avatars are incomplete, featuring only the head. Future work involves extending the model to include the neck, shoulders, hands, and the entire body. + +# References + +1. Alexander, O., Rogers, M., Lambeth, W., Chiang, J., Ma, W., Wang, C., Debevec, P.: The digital emily project: Achieving a photoreal digital actor. IEEE Computer Graphics and Applications 30 (2009) 3 +2. Alexander, O., Fyffe, G., Busch, J., Yu, X., Ichikari, R., Jones, A., Debevec, P., Jimenez, J., Danvoye, E., Antionazzi, B., et al.: Digital ira: Creating a real-time photoreal digital actor. In: ACM SIGGRAPH 2013 Posters, pp. 1-1 (2013) 3 +3. Athar, S., Xu, Z., Sunkavalli, K., Shechtman, E., Shu, Z.: Rignerf: Fully controllable neural 3d portraits. In: CVPR (June 2022) 2 +4. Bi, S., Lombardi, S., Saito, S., Simon, T., Wei, S.E., Mcphail, K., Ramamoorthi, R., Sheikh, Y., Saragih, J.: Deep reightable appearance models for animatable faces. ACM Transactions on Graphics (TOG) 40(4), 1-15 (2021) 4 +5. Borshukov, G., Lewis, J.P.: Realistic human face rendering for" the matrix reloaded". In: ACM Siggraph 2005 Courses, pp. 13-es (2005) 3 +6. Cao, C., Simon, T., Kim, J.K., Schwartz, G., Zollhoefer, M., Saito, S.S., Lombardi, S., Wei, S.E., Belko, D., Yu, S.I., Sheikh, Y., Saragih, J.: Authentic volumetric avatars from a phone scan. ACM Trans. Graph. (2022) 1, 2, 3, 4, 5, 9, 12, 14 +7. Cao, C., Wu, H., Weng, Y., Shao, T., Zhou, K.: Real-time facial animation with image-based dynamic avatars. ACM Transactions on Graphics 35(4) (2016) 4 +8. Casas, D., Alexander, O., Feng, A.W., Fyffe, G., Ichikari, R., Debevec, P., Wang, R., Suma, E., Shapiro, A.: Rapid photorealistic blendshapes from commodity rgb-d sensors. In: Proceedings of the 19th Symposium on Interactive 3D Graphics and Games. pp. 134-134 (2015) 4 +9. Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image quality assessment: Unifying structure and texture similarity 44(5), 2567-2581 (2022). https://doi.org/10.1109/TPAMI.2020.3045810 10, 13 +0. Gafni, G., Thies, J., Zollhofer, M., Nießner, M.: Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8649-8658 (2021) 4 +1. Garrido, P., Zollhöfer, M., Casas, D., Valgaerts, L., Varanasi, K., Pérez, P., Theobalt, C.: Reconstruction of personalized 3d face rigs from monocular video. ACM Transactions on Graphics (TOG) 35(3), 1-15 (2016) 4 +2. Ghosh, A., Fyffe, G., Tunwattanapong, B., Busch, J., Yu, X., Debevec, P.: Multiview face capture using polarized spherical gradient illumination. ACM Trans. Graph. (2011) 3 +3. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020) 4 +4. Grassal, P.W., Prinzler, M., Leistner, T., Rother, C., Nießner, M., Thies, J.: Neural head avatars from monocular rgb videos. In: CVPR. pp. 18653-18664 (June 2022) 2, 4 +5. Hu, L., Saito, S., Wei, L., Nagano, K., Seo, J., Fursund, J., Sadeghi, I., Sun, C., Chen, Y.C., Li, H.: Avatar digitization from a single image for real-time rendering. ACM Transactions on Graphics (ToG) 36(6), 1-14 (2017) 4 +6. Huang, X., Liu, M.Y., Belongie, S., Kautz, J.: Multimodal unsupervised image-to-image translation. In: Proceedings of the European conference on computer vision (ECCV). pp. 172-189 (2018) 4 +7. Ichim, A.E., Bouaziz, S., Pauly, M.: Dynamic 3d avatar creation from hand-held video input. ACM Transactions on Graphics (ToG) 34(4), 1-14 (2015) 4 +8. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: CVPR (2017) 2, 4 + +19. Jimenez, J., Echevarria, J.I., Oat, C., Gutierrez, D.: Practical and realistic facial wrinkles animation. In: GPU Pro 360 Guide to Geometry Manipulation, pp. 95-107. AK Peters/CRC Press (2018) 3 +20. Jimenez, J., Scully, T., Barbosa, N., Donner, C., Alvarez, X., Vieira, T., Matts, P., Orvalho, V., Gutierrez, D., Weyrich, T.: A practical appearance model for dynamic facial color. In: ACM SIGGRAPH Asia 2010 papers, pp. 1-10 (2010) 3 +21. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and superresolution. In: ECCV (2016) 10, 13 +22. Karras, T., Aittala, M., Hellsten, J., Laine, S., Lehtinen, J., Aila, T.: Training generative adversarial networks with limited data. In: Proc. NeurIPS (2020) 6 +23. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019) 7 +24. Lattas, A., Moschoglou, S., Gecer, B., Ploumpis, S., Triantafyllou, V., Ghosh, A., Zafeiriou, S.: Avatarme: Realistically renderable 3d facial reconstruction" in-the-wild". In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 760-769 (2020) 4 +25. Lattas, A., Moschoglou, S., Ploumpis, S., Gecer, B., Ghosh, A., Zafeiriou, S.: Avatarme++: Facial shape and brdf inference with photorealistic rendering-aware gans. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9269-9284 (2021) 4 +26. Lin, J., Yuan, Y., Zou, Z.: Meingame: Create a game character face from a single portrait. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 311-319 (2021) 4 +27. Lombardi, S., Saragih, J., Simon, T., Sheikh, Y.: Deep appearance models for face rendering. ACM Transactions on Graphics (ToG) 37(4), 1-13 (2018) 2, 3 +28. Lombardi, S., Simon, T., Schwartz, G., Zollhoefer, M., Sheikh, Y., Saragih, J.: Mixture of volumetric primitives for efficient neural rendering. ACM Transactions on Graphics (ToG) 40(4), 1-13 (2021) 2, 3 +29. Luo, H., Nagano, K., Kung, H., Xu, Q., Wang, Z., Wei, L., Hu, L., Li, H.: Normalized avatar synthesis using stylegan and perceptual refinement. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 4 +30. Ma, S., Simon, T., Saragih, J., Wang, D., Li, Y., De La Torre, F., Sheikh, Y.: Pixel codec avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 64-73 (2021) 3 +31. Mescheder, L., Geiger, A., Nowozin, S.: Which training methods for gans do actually converge? (2018) 7 +32. Nagano, K., Fyffe, G., Alexander, O., Barbic, J., Li, H., Ghosh, A., Debevec, P.E.: Skin microstructure deformation with displacement map convolution. ACM Trans. Graph. 34(4), 109-1 (2015) 3 +33. Park, T., Efros, A.A., Zhang, R., Zhu, J.Y.: Contrastive learning for unpaired image-to-image translation. In: European Conference on Computer Vision (2020) 9, 10, 11 +34. Pinkney, J.N., Adler, D.: Resolution dependent gan interpolation for controllable image synthesis between domains. arXiv preprint arXiv:2010.05334 (2020) 4 +35. Sang, S., Zhi, T., Song, G., Liu, M., Lai, C., Liu, J., Wen, X., Davis, J., Luo, L.: Agileavatar: Stylized 3d avatar creation via cascaded domain bridging. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-8 (2022) 4 +36. Seymour, M., Evans, C., Libreri, K.: Meet mike: epic avatars. In: ACM SIGGRAPH 2017 VR Village, pp. 1-2 (2017) 3 +37. Shi, T., Yuan, Y., Fan, C., Zou, Z., Shi, Z., Liu, Y.: Face-to-parameter translation for game character auto-creation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 161–170 (2019) 4 + +38. Song, G., Luo, L., Liu, J., Ma, W.C., Lai, C., Zheng, C., Cham, T.J.: Agilegan: stylizing portraits by inversion-consistent transfer learning. ACM Transactions on Graphics (TOG) 40(4), 1-13 (2021) 2, 4, 9, 10, 11 +39. Wang, T.C., Liu, M.Y., Tao, A., Liu, G., Kautz, J., Catanzaro, B.: Few-shot video-to-video synthesis (2019) 4 +40. Wang, T.C., Liu, M.Y., Zhu, J.Y., Tao, A., Kautz, J., Catanzaro, B.: High-resolution image synthesis and semantic manipulation with conditional gans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8798-8807 (2018) 4 +41. Yamaguchi, S., Saito, S., Nagano, K., Zhao, Y., Chen, W., Olszewski, K., Morishima, S., Li, H.: High-fidelity facial reflectance and geometry inference from an unconstrained image. ACM Transactions on Graphics (TOG) 37(4), 1-14 (2018) 4 +42. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networkss. In: Computer Vision (ICCV), 2017 IEEE International Conference on (2017) 2, 4, 9, 10, 11 +43. Zongsheng Yue, J.W., Loy, C.C.: Resshift: Efficient diffusion model for image super-resolution by residual shifting. In: Advances in Neural Information Processing Systems (NeurIPS) (2023) 7, 8 \ No newline at end of file diff --git a/2024/Bridging the Gap_ Studio-like Avatar Creation from a Monocular Phone Capture/images.zip b/2024/Bridging the Gap_ Studio-like Avatar Creation from a Monocular Phone Capture/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..65e33518b86b47fb0a14f466fc4b40d17d3f4c9f --- /dev/null +++ b/2024/Bridging the Gap_ Studio-like Avatar Creation from a Monocular Phone Capture/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89f8780e5e95254ed7c12d567d3a36676ebefaf31d074466c56f468b2883be52 +size 485399 diff --git a/2024/Bridging the Gap_ Studio-like Avatar Creation from a Monocular Phone Capture/layout.json b/2024/Bridging the Gap_ Studio-like Avatar Creation from a Monocular Phone Capture/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..67f65f64f28e41100ba3de3aca9f748b1074269e --- /dev/null +++ b/2024/Bridging the Gap_ Studio-like Avatar Creation from a Monocular Phone Capture/layout.json @@ -0,0 +1,10481 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "text", + "content": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "text", + "content": "ShahRukh Athar" + }, + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "text", + "content": " Shunsuke Saito" + }, + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "text", + "content": " Zhengyu Yang" + }, + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "text", + "content": " Stanislav Pidhorskyi" + }, + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "text", + "content": " Chen Cao" + }, + { + "bbox": [ + 138, + 167, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 169, + 203, + 296, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 203, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 169, + 203, + 296, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 169, + 203, + 296, + 217 + ], + "type": "text", + "content": "Captions Research, New York" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 317, + 204, + 446, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 204, + 446, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 204, + 446, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 317, + 204, + 446, + 217 + ], + "type": "text", + "content": "Meta Reality Labs, Pittsburgh." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 243, + 228, + 391, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 228, + 391, + 241 + ], + "spans": [ + { + "bbox": [ + 243, + 228, + 391, + 241 + ], + "type": "text", + "content": "Stony Brook University, New York" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 173, + 251, + 441, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 251, + 441, + 261 + ], + "spans": [ + { + "bbox": [ + 173, + 251, + 441, + 261 + ], + "type": "text", + "content": "shahrukh@nocapinc.com shunsuke.saito16@gmail.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 178, + 262, + 435, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 262, + 435, + 272 + ], + "spans": [ + { + "bbox": [ + 178, + 262, + 435, + 272 + ], + "type": "text", + "content": "stpidhorskyi@meta.com zhengyu-yang@outlook.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 250, + 274, + 364, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 274, + 364, + 283 + ], + "spans": [ + { + "bbox": [ + 250, + 274, + 364, + 283 + ], + "type": "text", + "content": "zju.caochen@gmail.com" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 138, + 335, + 177, + 378 + ], + "blocks": [ + { + "bbox": [ + 138, + 313, + 178, + 332 + ], + "lines": [ + { + "bbox": [ + 138, + 313, + 178, + 332 + ], + "spans": [ + { + "bbox": [ + 138, + 313, + 178, + 332 + ], + "type": "text", + "content": "Monocular phone capture" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 335, + 177, + 378 + ], + "lines": [ + { + "bbox": [ + 138, + 335, + 177, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 177, + 378 + ], + "type": "image", + "image_path": "60cb2a326479dbe9b1e096bc14e396f5926296187b5943feb1e0de334faca02f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 138, + 380, + 177, + 424 + ], + "blocks": [ + { + "bbox": [ + 138, + 380, + 177, + 424 + ], + "lines": [ + { + "bbox": [ + 138, + 380, + 177, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 177, + 424 + ], + "type": "image", + "image_path": "d82b72dd12f0a0880c5ab26afa11f26e8527e6235cb2e0f3ec5e81c46c63d462.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 434, + 482, + 491 + ], + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 491 + ], + "type": "text", + "content": "Fig. 1: From a low-resolution texture map obtained through monocular phone capture, our method produces a high-resolution texture map with precise facial details, studio-like illumination, and inpainted missing regions. This generated texture map can subsequently be utilized to create a high-quality, photorealistic avatar using the pretrained Universal Prior Model (UPM) from AVA [6]." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 180, + 335, + 269, + 423 + ], + "blocks": [ + { + "bbox": [ + 179, + 322, + 258, + 331 + ], + "lines": [ + { + "bbox": [ + 179, + 322, + 258, + 331 + ], + "spans": [ + { + "bbox": [ + 179, + 322, + 258, + 331 + ], + "type": "text", + "content": "Extracted texture map" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 180, + 335, + 269, + 423 + ], + "lines": [ + { + "bbox": [ + 180, + 335, + 269, + 423 + ], + "spans": [ + { + "bbox": [ + 180, + 335, + 269, + 423 + ], + "type": "image", + "image_path": "879fe4f17527076281cf808d47c10465417d11b6df9f5440ada1799aa717b6e6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 283, + 335, + 371, + 423 + ], + "blocks": [ + { + "bbox": [ + 304, + 318, + 352, + 335 + ], + "lines": [ + { + "bbox": [ + 304, + 318, + 352, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 352, + 335 + ], + "type": "text", + "content": "Generated highresolution texture n" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 283, + 335, + 371, + 423 + ], + "lines": [ + { + "bbox": [ + 283, + 335, + 371, + 423 + ], + "spans": [ + { + "bbox": [ + 283, + 335, + 371, + 423 + ], + "type": "image", + "image_path": "47983b98dab4ec58dafc3d4015ba511b8b71652884c6895a8633b300568e62d4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 413, + 335, + 468, + 423 + ], + "blocks": [ + { + "bbox": [ + 398, + 318, + 477, + 335 + ], + "lines": [ + { + "bbox": [ + 398, + 318, + 477, + 335 + ], + "spans": [ + { + "bbox": [ + 398, + 318, + 477, + 335 + ], + "type": "text", + "content": "High-quality reanimation and novel view synthesis" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 413, + 335, + 468, + 423 + ], + "lines": [ + { + "bbox": [ + 413, + 335, + 468, + 423 + ], + "spans": [ + { + "bbox": [ + 413, + 335, + 468, + 423 + ], + "type": "image", + "image_path": "ef2231891fc8d89d00f972c47097edf4d01c5112307364c2df8d0f24b419adc9.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 159, + 504, + 455, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 504, + 455, + 647 + ], + "spans": [ + { + "bbox": [ + 159, + 504, + 455, + 647 + ], + "type": "text", + "content": "Abstract. Creating photorealistic avatars for individuals traditionally involves extensive capture sessions with complex and expensive studio devices like the LightStage system. While recent strides in neural representations have enabled the generation of photorealistic and animatable 3D avatars from quick phone scans, they have the capture-time lighting baked-in, lack facial details and have missing regions in areas such as the back of the ears. Thus, they lag in quality compared to studio-captured avatars. In this paper, we propose a method that bridges this gap by generating studio-like illuminated texture maps from short, monocular phone captures. We do this by parameterizing the phone texture maps using the " + }, + { + "bbox": [ + 159, + 504, + 455, + 647 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 159, + 504, + 455, + 647 + ], + "type": "text", + "content": " space of a StyleGAN2, enabling near-perfect reconstruction. Then, we finetune a StyleGAN2 by sampling in the " + }, + { + "bbox": [ + 159, + 504, + 455, + 647 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 159, + 504, + 455, + 647 + ], + "type": "text", + "content": " parameterized space using a very small set of studio-captured textures as an adversarial training signal. To further enhance the realism and accuracy of facial details, we super-resolve" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 654, + 313, + 666 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 654, + 313, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 654, + 313, + 666 + ], + "type": "text", + "content": "*Work done while interning at Meta Reality Labs" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 160, + 117, + 455, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 117, + 455, + 173 + ], + "spans": [ + { + "bbox": [ + 160, + 117, + 455, + 173 + ], + "type": "text", + "content": "the output of the StyleGAN2 using carefully designed diffusion model that is guided by image gradients of the phone-captured texture map. Once trained, our method excels at producing studio-like facial texture maps from casual monocular smartphone videos. Demonstrating its capabilities, we showcase the generation of photorealistic, uniformly lit, complete avatars from monocular phone captures." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 207, + 218, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 207, + 218, + 220 + ], + "spans": [ + { + "bbox": [ + 133, + 207, + 218, + 220 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 234, + 482, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 234, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 482, + 329 + ], + "type": "text", + "content": "Photorealistic and animatable avatars are paramount for lifelike human-to-human interactions in AR/VR applications. Creating high-fidelity avatars often requires sophisticated devices like the LightStage capture system that capture human heads with a range of facial expressions [27, 28]. Typically, these captures occur in a studio-like environment, with uniform illumination and densely sampled views to reconstruct complete avatars with consistent lighting. While these methods excel in generating hyper-realistic avatars, they cannot be scaled to millions of people as capturing so many people with a LightStage-like capture system is impractical and non-trivial." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 330, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 426 + ], + "type": "text", + "content": "Recently, there has been extensive work in generating photorealistic avatars through a monocular capture [3,6,14]. These methods involve scanning various facial expressions and head poses of the user to reconstruct a 3D avatar that closely aligns with the captured data. However, a notable limitation of these methods is that the captured lighting is embedded in the avatars, causing their appearance to be heavily dependent on the capture devices and surrounding environments. Furthermore, due to the constraints of the capture setup, certain areas, such as the back of the head or ears, are never visible, leaving visible holes and artifacts when viewed during animation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 483, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 483, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 483, + 521 + ], + "type": "text", + "content": "An innovative strategy to tackle this challenge is to use image-to-image translation to transform the phone data into studio-captured data. This transformation can be learnt through supervised training using paired data [18], or through unsupervised training using unpaired datasets [42]. Due to reasons mentioned earlier, creating a large-scale paired dataset of studio and phone captures is impractical, which rules out the possibility of using supervised image-to-image translation methods. On the flip side, current unsupervised image-to-image translation methods fail to preserve fine detail in the transformation process [38, 42] which is paramount to creating a high-fidelity avatar." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 522, + 483, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 483, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 483, + 665 + ], + "type": "text", + "content": "In this paper, we introduce a method capable of generating studio-like, high-quality avatars from monocular phone captures. We do this by parameterizing a large-scale dataset of phone-captured face texture maps using the " + }, + { + "bbox": [ + 130, + 522, + 483, + 665 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 130, + 522, + 483, + 665 + ], + "type": "text", + "content": " space of a StyleGAN2. Then, we finetune this StyleGAN2 by sampling from this parametrized " + }, + { + "bbox": [ + 130, + 522, + 483, + 665 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 130, + 522, + 483, + 665 + ], + "type": "text", + "content": " space using a small set of unpaired studio-captured texture maps, to create a Studio-StyleGAN2 model. Our key insight is that sampling from the " + }, + { + "bbox": [ + 130, + 522, + 483, + 665 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 130, + 522, + 483, + 665 + ], + "type": "text", + "content": " space instead of Z space leads to a more generalizable model as samples from the " + }, + { + "bbox": [ + 130, + 522, + 483, + 665 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 130, + 522, + 483, + 665 + ], + "type": "text", + "content": " are, by construction, as diverse as the training set while samples from the Z space often suffer from mode-collapse. During inference, the given phone-captured texture map, is first inverted to the " + }, + { + "bbox": [ + 130, + 522, + 483, + 665 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 130, + 522, + 483, + 665 + ], + "type": "text", + "content": " space of the StyleGAN2. This inverted " + }, + { + "bbox": [ + 130, + 522, + 483, + 665 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 130, + 522, + 483, + 665 + ], + "type": "text", + "content": " vector is then given as input to the Studio-StyleGAN2 to generate a low-resolution studio-like texture map. Finally, a novel facial detail conditioned diffusion model is used to enhance facial details of the low-resolution" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "Athar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "studio-like texture map obtained in the previous step. Once trained, our method excels at producing high-quality studio-like face texture maps from monocular phone captures, which are then used as inputs to the universal face prior model (UPM) from Authentic Volumetric Avatars (AVA) [6] to generate photorealistic avatars. In summary, our key contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 182, + 480, + 286 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 182, + 479, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 182, + 479, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 182, + 479, + 205 + ], + "type": "text", + "content": "- Introducing a groundbreaking method for creating studio-like, photorealistic avatars from monocular phone captures." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 206, + 480, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 480, + 251 + ], + "type": "text", + "content": "- Innovative finetuning of a pre-trained generative model using a minimal dataset from another domain by sampling in the inverted " + }, + { + "bbox": [ + 138, + 206, + 480, + 251 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 138, + 206, + 480, + 251 + ], + "type": "text", + "content": " space, which facilitates the development of a generative model for a new domain while preserving the integrity of the " + }, + { + "bbox": [ + 138, + 206, + 480, + 251 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 138, + 206, + 480, + 251 + ], + "type": "text", + "content": " latent space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 254, + 479, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 254, + 479, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 254, + 479, + 286 + ], + "type": "text", + "content": "- A novel diffusion model conditioned on the phone texture gradient, that is designed to upsample studio textures, effectively enhancing facial details and contributing to the overall realism of the generated avatars." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 307, + 224, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 307, + 224, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 224, + 319 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 331, + 481, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 481, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 481, + 378 + ], + "type": "text", + "content": "Our paper aims to bridge the gap between a studio capture and a phone capture by creating a studio-like, photorealistic avatar through a monocular phone capture. In delineating the related work, we offer a comprehensive overview of key research domains, including studio-captured avatars, phone-captured avatars, and image-to-image translation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 379, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 668 + ], + "type": "text", + "content": "Studio-captured avatars. The reconstruction of high-fidelity static and dynamic models of the human head based on photometric measurements has a long-standing history in computer graphics and vision. Achieving a photorealistic human avatar often requires specialized hardware in high-end production, such as the LightStage system. To model the complex skin appearance, various physically-based models have been explored. Notably, subsurface scattering [5], linear polarization patterns [12] and fine-scale skin details [1, 2] have been investigated. For dynamic expression details, Jimenez et al. [20] compute dynamic skin appearances by blending hemoglobin distributions captured with different expressions. In their subsequent work [19], expression-dependent normal maps are interpolated to add realistic wrinkles to an animated face. Nagano et al. [32] synthesize skin microstructures based on local geometric features derived from high-precision microgeometry, acquired with an LED sphere and a skin deformer. While these methods have been instrumental for offline movie production, their substantial compute requirements make them less suitable for real-time applications. Despite recent efforts to enable real-time rendering of physically-based avatars [36], heavy compute remains challenge. In response to the challenges posed by complex physical computations, researchers have proposed a deep appearance model [27]. This model utilizes a coarse 3D triangle mesh in conjunction with view-dependent texture mapping. The texture is regressed by a neural network conditioned on viewpoint and expression latent codes. This conditioning accounts for view- and expression-dependent variations while compensating for the imperfect proxy geometry. Subsequent work extends the mesh-texture representation to a volumetric representation using a Mixture of Volumetric Primitives (MVP) [28], further enhancing the model's quality. Pixel Codec Avatars (PiCA), as demonstrated by [30], showcase the efficiency of rendering such models, even on mobile hardware platforms," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 103 + ], + "type": "text", + "content": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "by leveraging efficient per-pixel processing. Moreover, Bi et al. [4], based on relighting captured data, can also relight avatars with any novel point light or environment maps. While these methods can achieve hyper-realistic avatars, their studio requirements make them challenging to generalize to ordinary users." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 482, + 248 + ], + "type": "text", + "content": "Phone-captured avatars. There are several methods aimed at creating avatars in lightweight ways, even from a phone or a single image. Some of these approaches focus on generating stylized avatars based on different input requirements, ranging from capturing multi-view images using a phone [17] to utilizing a single monocular image [26, 29, 35, 37]. While these methods excel in producing animatable avatars, it is important to note that their appearance tends to be cartoonish and lacks realism." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 259, + 483, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 483, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 483, + 403 + ], + "type": "text", + "content": "Another category of methods is dedicated to creating realistic human avatars based on graphics pipelines [7, 8, 11, 15, 24, 25, 41]. While relying on traditional graphics methods, these approaches often result in avatars that appear uncanny. Building upon deep learning-based representations, researchers have proposed methods to generate avatars with increased realism, including notable works by Gafni et al. [10] and Grassal et al. [14]. Among these, the work of Cao et al. [6], Authentic Volumetric Avatars (AVA), stands out for its focus on creating photorealistic avatars from phone scans. The process involves training a Universal Prior Model (UPM) using studio-captured data, followed by personalizing this UPM using data from a phone scan of an unseen subject. While the method successfully produces avatars with realistic appearance and animation, it is worth noting that the lighting is baked into the personalized avatar, and certain details, such as the back of the ears, may be missing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 415, + 483, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 483, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 483, + 511 + ], + "type": "text", + "content": "Image-to-image translation We use image-to-image translation to map images from the source domain (phone data) to the target domain (studio data). Isola et al. introduced Pix2Pix [18], a method that utilizes adversarial training strategies [13] to achieve this mapping. Additionally, Wang et al. [40] focused on increasing the resolution of generated results from semantic label maps. While these methods successfully map images between domains, it is essential to note that they require paired training data for effective implementation. In many real-world scenarios, obtaining paired training data can be challenging and expensive." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "text", + "content": "To address this issue, unsupervised image-to-image translation has been introduced. Zhu et al. [42] proposed a novel cycle-consistency loss to ensure that translating an image from one domain to another and back again should result in the original image. This helps the model maintain consistency and produce more realistic translations. Subsequent methods have further improved unsupervised image-to-image translation from different perspectives, including a multimodal model [16], few-shot input for video-to-video translation [39], translation of images with human control [34], and translation of real images into different styles [38]. However, none of these methods are designed with preservation of facial identity and the generation of realistic facial details in mind, especially with such little training data. Later in the paper we show how this prior work compares to ours for transforming low-resolution phone-captured texture maps to high-resolution studio-captured texture maps." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "Athar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 355, + 227 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 355, + 227 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 355, + 227 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 355, + 227 + ], + "type": "image", + "image_path": "1e0379f9076c396cc9e0ae4f9f76d160176b7b478e0ba1284fb2e6ad8f5bfb87.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "lines": [ + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "text", + "content": "Fig. 2: Method Overview. Our method employs a two-step process to train " + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{Studio}}" + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "text", + "content": ". Initially, we train a StyleGAN2 on " + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "inline_equation", + "content": "12k" + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "text", + "content": " neutral texture maps captured by phones, yielding " + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{phone}}" + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "text", + "content": ". Subsequently, we initialize " + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{Studio}}" + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "text", + "content": " with the weights of " + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{phone}}" + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "text", + "content": " and fine-tune it by sampling from " + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "inline_equation", + "content": "S_{W^{+}} = \\{W_{\\mathcal{I}_0^{\\text{phone}}}, \\ldots, W_{\\mathcal{I}_{N-1}^{\\text{phone}}}^{+}\\}" + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "text", + "content": "; where " + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "inline_equation", + "content": "N = 12k" + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "inline_equation", + "content": "W_{\\mathcal{I}_i^{\\text{phone}}}^{+}" + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "text", + "content": " represents the vector obtained by inverting the " + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "text", + "content": "'th phone-captured texture map, " + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_0^{\\text{phone}}" + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "text", + "content": ", in the " + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "text", + "content": " space of " + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{phone}}" + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "text", + "content": ". During inference, the given phone-captured texture map is inverted to " + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "text", + "content": ", then passed to " + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{Studio}}" + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "text", + "content": " to generate a low-resolution studio-lit texture map, " + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{*}" + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "text", + "content": ". Accurate facial details are subsequently added using diffusion model " + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "inline_equation", + "content": "f_{\\phi}" + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "text", + "content": ", conditioned on the gradient of the phone texture. This process results in the final high-resolution, studio-lit, and completed texture map, " + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{*}" + }, + { + "bbox": [ + 131, + 236, + 482, + 342 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 361, + 114, + 484, + 227 + ], + "blocks": [ + { + "bbox": [ + 361, + 114, + 484, + 227 + ], + "lines": [ + { + "bbox": [ + 361, + 114, + 484, + 227 + ], + "spans": [ + { + "bbox": [ + 361, + 114, + 484, + 227 + ], + "type": "image", + "image_path": "31a1533d559db413d9f515be8a5f77c709c0ef57be5ddbed4b846be033f77875.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 365, + 194, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 194, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 194, + 376 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 389, + 483, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 483, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 483, + 581 + ], + "type": "text", + "content": "In this section, we present our method for generating studio-like avatars from the phone captures. Our approach consists of two key components: a StyleGAN2 for texture translation and a diffusion model for facial detail generation. In Sect 3.1, we describe the generation of a low-resolution texture map with studio-like lighting and missing regions inpainted. First, we pretrain a StyleGAN2 on 12k phone textures and then finetune it using a small set of studio-captured texture maps. In order to improve the generalization of the finetuned StyleGAN2, we optimize it by sampling in the " + }, + { + "bbox": [ + 130, + 389, + 483, + 581 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 130, + 389, + 483, + 581 + ], + "type": "text", + "content": " space, instead of the W-space or Z-space, using a set of 12k " + }, + { + "bbox": [ + 130, + 389, + 483, + 581 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 130, + 389, + 483, + 581 + ], + "type": "text", + "content": " vectors obtained by inverting the phone captured texture maps. In Sect 3.2, we introduce a diffusion model that generates facial details. The diffusion model takes the output from the aforementioned StyleGAN2 and generates a high-resolution neutral texture with realistic facial details. After obtaining a high-resolution studio-like neutral texture from our method, we use it to learn a color transform to transfer phone-captured expression textures to studio-lit expression textures. These expressions are subsequently utilized for animating a high-quality avatar (Sect 3.3). Notably, we intentionally avoid applying inpainting or super-resolution techniques to the expression textures. Fig. 2 provides an overview of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 598, + 335, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 335, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 335, + 610 + ], + "type": "text", + "content": "3.1 Illumination manipulation and inpainting" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "In our initial step, we track the geometry from a monocular phone capture of the user's neutral face and extract the neutral texture " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{phone}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " from the captured image, employing the method outlined in [6]. Subsequently, we translate this phone-captured texture with in-the-wild lighting and missing regions into a texture map with studio lighting and" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 448, + 102 + ], + "type": "text", + "content": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "content": "missing regions in-painted. This translation is accomplished by parametrizing " + }, + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{phone}" + }, + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "content": " using the " + }, + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "content": " space of a StyleGAN2, as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 164, + 156, + 482, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 156, + 482, + 174 + ], + "spans": [ + { + "bbox": [ + 164, + 156, + 482, + 174 + ], + "type": "interline_equation", + "content": "W _ {\\mathcal {I} ^ {p h o n e}} ^ {+} = \\underset {W ^ {+}} {\\operatorname {a r g m i n}} \\left\\| \\mathcal {G} _ {p h o n e} \\left(W ^ {+}\\right) - \\mathcal {I} ^ {p h o n e} \\right\\| _ {2} ^ {2} + L P I P S \\left(\\mathcal {G} _ {p h o n e} \\left(W ^ {+}\\right), \\mathcal {I} ^ {p h o n e}\\right), \\tag {1}", + "image_path": "440e6415487922f178f0d6ddf85e21ab593c9db18e9b8fc2eaf0b601500a4976.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 182, + 479, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 182, + 479, + 217 + ], + "spans": [ + { + "bbox": [ + 131, + 182, + 479, + 217 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 182, + 479, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{phone}" + }, + { + "bbox": [ + 131, + 182, + 479, + 217 + ], + "type": "text", + "content": " is a StyleGAN2 trained on phone-captured textures, and " + }, + { + "bbox": [ + 131, + 182, + 479, + 217 + ], + "type": "inline_equation", + "content": "W_{\\mathcal{T}^{phone}}^{+}" + }, + { + "bbox": [ + 131, + 182, + 479, + 217 + ], + "type": "text", + "content": " is the optimized vector in the " + }, + { + "bbox": [ + 131, + 182, + 479, + 217 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 131, + 182, + 479, + 217 + ], + "type": "text", + "content": " space of " + }, + { + "bbox": [ + 131, + 182, + 479, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{phone}" + }, + { + "bbox": [ + 131, + 182, + 479, + 217 + ], + "type": "text", + "content": ". Now, we generate the low-resolution studio-like texture as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 263, + 225, + 482, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 225, + 482, + 239 + ], + "spans": [ + { + "bbox": [ + 263, + 225, + 482, + 239 + ], + "type": "interline_equation", + "content": "\\mathcal {I} ^ {*} = \\mathcal {G} _ {\\text {S t u d i o}} \\left(W _ {\\mathcal {I} _ {\\text {p h o n e}}} ^ {+}\\right), \\tag {2}", + "image_path": "e069a64098e331dbb32e0ce336c0e13fea797c8c9e38181f0be37570f334ec93.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{Studio}}" + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "text", + "content": " is a StyleGAN2 responsible for generating low-resolution studio-like textures, while " + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^*" + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "text", + "content": " refers to the studio-lit version of " + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{\\text{phone}}" + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "text", + "content": " with the inpainted missing regions. Ideally, " + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^*" + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "text", + "content": " should retain the identity and semantics of " + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{\\text{phone}}" + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "text", + "content": ", with the only distinctions being the illumination and the inpainting of missing regions. Below we describe the training procedure for " + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{Studio}}" + }, + { + "bbox": [ + 131, + 244, + 479, + 304 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "text", + "content": "We initialize the synthesis network of " + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{Studio}}" + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "text", + "content": " using weights from " + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{phone}}" + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "text", + "content": ", which we then finetune to generate texture maps with studio lighting and inpaint the missing regions. We choose to finetune " + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{phone}}" + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "text", + "content": " for the following reasons: 1) Due to the immense costs of a studio capture, studio quality training data is typically very limited. In our case, we only have 383 studio-captured texture maps, making training from scratch hard [22]; 2) Since we want " + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^*" + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{phone}" + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "text", + "content": " to have the same facial identity and semantics meaning, it is beneficial to learn a latent space that is shared between " + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{Studio}}" + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{phone}}" + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "text", + "content": ". For finetuning, we first invert the entire dataset of " + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "text", + "content": " phone-captured texture maps using Eq. (1), giving us a set of " + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "text", + "content": " vectors in the " + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "text", + "content": " space: " + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "inline_equation", + "content": "S_{W^{+}} = \\{W_{\\mathcal{I}_0^{phone}}^{+}, \\ldots, W_{\\mathcal{I}_{N-1}^{phone}}^{+}\\}" + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "text", + "content": ". Next, we randomly sample vectors from the inverted set, " + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "inline_equation", + "content": "S_{W^{+}}" + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "text", + "content": ", and finetune the synthesis network of " + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{phone}}" + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "text", + "content": " to give " + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{Studio}}" + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "text", + "content": ". We choose to finetune the generator with samples from inverted " + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "text", + "content": " instead of the traditional " + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "text", + "content": " space, since they are, by construction, from the real data distribution and are consequently more diverse, leading to better generalization of " + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{Studio}}" + }, + { + "bbox": [ + 131, + 304, + 482, + 475 + ], + "type": "text", + "content": ". The losses we use during finetuning are described as below." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 476, + 479, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 479, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 479, + 499 + ], + "type": "text", + "content": "Studio-Discriminator loss. This loss uses a discriminator to ensure the texture maps generated by " + }, + { + "bbox": [ + 132, + 476, + 479, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{Studio}}" + }, + { + "bbox": [ + 132, + 476, + 479, + 499 + ], + "type": "text", + "content": " are from the distribution of studio-captured texture maps:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 187, + 500, + 481, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 500, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 187, + 500, + 481, + 533 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {A d v} = \\mathbb {E} _ {\\mathcal {T S t u d i o} \\sim P (\\mathcal {T S t u d i o})} [ \\min (0, - 1 + \\mathcal {D} _ {\\text {S t u d i o}}) ] \\\\ + \\mathbb {E} _ {W _ {\\mathcal {I} _ {i} ^ {p h o n e}} ^ {+} \\sim S _ {W ^ {+}}} \\left[ \\min (0, - 1 - \\mathcal {D} _ {\\text {S t u d i o}} \\left(\\mathcal {G} _ {\\text {S t u d i o}} \\left(W _ {\\mathcal {T} _ {i} ^ {p h o n e}} ^ {+}\\right)\\right) \\right], \\tag {3} \\\\ \\end{array}", + "image_path": "b28f870152c5154ab6966a2f4f122bf3ae1cf775c0978001f0fa71adc9270398.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 540, + 479, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 540, + 479, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 540, + 479, + 574 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 540, + 479, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\text{Studio}}" + }, + { + "bbox": [ + 131, + 540, + 479, + 574 + ], + "type": "text", + "content": " is a discriminator that initialized from the pretrained phone-captured texture StyleGAN2 Discriminator and finetuned using ground-truth studio-captured textures and textures generated by " + }, + { + "bbox": [ + 131, + 540, + 479, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{Studio}}" + }, + { + "bbox": [ + 131, + 540, + 479, + 574 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 576, + 479, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 576, + 479, + 610 + ], + "spans": [ + { + "bbox": [ + 131, + 576, + 479, + 610 + ], + "type": "text", + "content": "Face-Identity Loss This loss compares the face identity embeddings of the renders of texture maps generated from " + }, + { + "bbox": [ + 131, + 576, + 479, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{Studio}}" + }, + { + "bbox": [ + 131, + 576, + 479, + 610 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 576, + 479, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{phone}}" + }, + { + "bbox": [ + 131, + 576, + 479, + 610 + ], + "type": "text", + "content": " in order to ensure the identity is preserved:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 191, + 610, + 482, + 629 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 191, + 610, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 191, + 610, + 482, + 629 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {F a c e I D}} = \\left\\| \\left| \\mathcal {F} \\left(\\mathcal {G} _ {\\text {S t u d i o}} \\left(W _ {\\mathcal {I} _ {i} ^ {\\text {p h o n e}}} ^ {+}\\right)\\right) - \\mathcal {F} \\left(\\mathcal {G} _ {\\text {p h o n e}} \\left(W _ {\\mathcal {I} _ {i} ^ {\\text {p h o n e}}} ^ {+}\\right)\\right) \\right| \\right\\| _ {2} ^ {2}, \\tag {4}", + "image_path": "27a5e70a46a08361c0b14ad75231e32fe482bde56addb47cad631d98240d2be2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 635, + 336, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 635, + 336, + 648 + ], + "spans": [ + { + "bbox": [ + 131, + 635, + 336, + 648 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 635, + 336, + 648 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 131, + 635, + 336, + 648 + ], + "type": "text", + "content": " is a pretrained face recognition network1." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "Athar et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 652, + 246, + 664 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 652, + 246, + 664 + ], + "spans": [ + { + "bbox": [ + 132, + 652, + 246, + 664 + ], + "type": "text", + "content": "1We use the network from here" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 151 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 151 + ], + "type": "text", + "content": "Perceptual Loss This loss encourages the preservation of semantic features between the texture maps generated " + }, + { + "bbox": [ + 131, + 116, + 482, + 151 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{Studio}}" + }, + { + "bbox": [ + 131, + 116, + 482, + 151 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 116, + 482, + 151 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{phone}}" + }, + { + "bbox": [ + 131, + 116, + 482, + 151 + ], + "type": "text", + "content": " by minimizing their distance in the VGG feature space" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 208, + 145, + 482, + 168 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 145, + 482, + 168 + ], + "spans": [ + { + "bbox": [ + 208, + 145, + 482, + 168 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {P e r c p}} = \\operatorname {L P I P S} \\left(\\mathcal {G} _ {\\text {S t u d i o}} \\left(W _ {\\mathcal {I} _ {i} ^ {\\text {p h o n e}}} ^ {+}\\right), \\mathcal {G} _ {\\text {p h o n e}} \\left(W _ {\\mathcal {I} _ {i} ^ {\\text {p h o n e}}} ^ {+}\\right)\\right). \\tag {5}", + "image_path": "fbb3dd43f3cf0d836740a5f96d7657d006af0d9963b97a214f3099b90ba3c9cb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 173, + 482, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 173, + 482, + 208 + ], + "spans": [ + { + "bbox": [ + 131, + 173, + 482, + 208 + ], + "type": "text", + "content": "Perceptual Reconstruction Loss This loss ensures that skintones are preserved using a small amount of paired data (i.e. subjects for whom we have both phone and studio textures maps i.e both " + }, + { + "bbox": [ + 131, + 173, + 482, + 208 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{phone}" + }, + { + "bbox": [ + 131, + 173, + 482, + 208 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 173, + 482, + 208 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{Studio}" + }, + { + "bbox": [ + 131, + 173, + 482, + 208 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 197, + 212, + 482, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 212, + 482, + 235 + ], + "spans": [ + { + "bbox": [ + 197, + 212, + 482, + 235 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {P e r c p - R e c o n s}} = \\frac {1}{k} \\sum_ {i = 1} ^ {K} \\operatorname {L P I P S} \\left(\\mathcal {G} _ {\\text {S t u d i o}} \\left(W _ {\\mathcal {T} _ {i} ^ {\\text {p h o n e}}} ^ {+}\\right), \\mathcal {I} _ {i} ^ {\\text {S t u d i o}}\\right), \\tag {6}", + "image_path": "6f603b3b9e66429b437524d99e6e9a15ef31ce9a659e870846532565875fc718.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 243, + 482, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 243, + 482, + 290 + ], + "spans": [ + { + "bbox": [ + 131, + 243, + 482, + 290 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 243, + 482, + 290 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_i^{Studio}" + }, + { + "bbox": [ + 131, + 243, + 482, + 290 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 243, + 482, + 290 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_i^{phone}" + }, + { + "bbox": [ + 131, + 243, + 482, + 290 + ], + "type": "text", + "content": " are the ground-truth studio and ground-truth phone texture maps of the same person. We only have a very small amount of this data for training (K=83). R1-Regularization: We regularize " + }, + { + "bbox": [ + 131, + 243, + 482, + 290 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{Studio}" + }, + { + "bbox": [ + 131, + 243, + 482, + 290 + ], + "type": "text", + "content": " using the R1 regularization [31] as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 208, + 288, + 482, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 288, + 482, + 301 + ], + "spans": [ + { + "bbox": [ + 208, + 288, + 482, + 301 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {R 1} = - \\frac {\\gamma}{2} \\mathbb {E} _ {\\mathcal {I} ^ {S t u d i o} \\sim P (\\mathcal {I} ^ {S t u d i o})} \\left[ \\left\\| \\nabla \\mathcal {D} _ {S t u d i o} (\\mathcal {I} ^ {S t u d i o}) \\right\\| _ {2} ^ {2} \\right], \\tag {7}", + "image_path": "88fe115ecd51c13d741f8d18974adef137a4974b2420aac3350b459bfb6c6764.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 306, + 194, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 306, + 194, + 317 + ], + "spans": [ + { + "bbox": [ + 131, + 306, + 194, + 317 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 306, + 194, + 317 + ], + "type": "inline_equation", + "content": "\\gamma = 10" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 319, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 319, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 131, + 319, + 482, + 390 + ], + "type": "text", + "content": "In order to further encourage identity preservation, we leverage the multi-resolution architecture of StyleGAN2 [23] and only optimize the generator parameters after the " + }, + { + "bbox": [ + 131, + 319, + 482, + 390 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 131, + 319, + 482, + 390 + ], + "type": "text", + "content": " resolution during the finetuning process. The intuition being that identity-specific information is mostly stored in the low-resolution maps, we ablate this choice in the Supplementary section of the paper. The final optimization for the joint finetuning of " + }, + { + "bbox": [ + 131, + 319, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{Studio}}" + }, + { + "bbox": [ + 131, + 319, + 482, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 319, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\text{Studio}}" + }, + { + "bbox": [ + 131, + 319, + 482, + 390 + ], + "type": "text", + "content": " is the following:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 183, + 396, + 482, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 396, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 183, + 396, + 482, + 416 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathcal {G} _ {S t u d i o} ^ {\\theta (8 +)}} \\max _ {\\mathcal {D} _ {S t u d i o}} \\mathcal {L} _ {A d v} + \\mathcal {L} _ {R 1} + \\mathcal {L} _ {P e r c p - R e c o n s} + \\lambda_ {1} \\mathcal {L} _ {P e r c p} + \\lambda_ {2} \\mathcal {L} _ {F a c e I D}, \\tag {8}", + "image_path": "6aad5fbf820602b46eafc60d3e9028f94b9527db6dad74a613b53d90d5f4a7c1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 423, + 482, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 423, + 482, + 448 + ], + "spans": [ + { + "bbox": [ + 131, + 423, + 482, + 448 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 423, + 482, + 448 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 0.5" + }, + { + "bbox": [ + 131, + 423, + 482, + 448 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 423, + 482, + 448 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 10" + }, + { + "bbox": [ + 131, + 423, + 482, + 448 + ], + "type": "text", + "content": ". We ablate " + }, + { + "bbox": [ + 131, + 423, + 482, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Percp-Recons}, \\mathcal{L}_{Percp}, \\mathcal{L}_{FaceID}" + }, + { + "bbox": [ + 131, + 423, + 482, + 448 + ], + "type": "text", + "content": " in the supplementary." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 466, + 304, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 466, + 304, + 478 + ], + "spans": [ + { + "bbox": [ + 131, + 466, + 304, + 478 + ], + "type": "text", + "content": "3.2 Synthesis of accurate facial details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 488, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 488, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 131, + 488, + 482, + 559 + ], + "type": "text", + "content": "Generating high-quality facial details. " + }, + { + "bbox": [ + 131, + 488, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{Studio}}" + }, + { + "bbox": [ + 131, + 488, + 482, + 559 + ], + "type": "text", + "content": " can produce an inpainted texture map, " + }, + { + "bbox": [ + 131, + 488, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^*" + }, + { + "bbox": [ + 131, + 488, + 482, + 559 + ], + "type": "text", + "content": ", with illumination consistent with that of a studio capture. However, due to the limited amount of finetuning data, " + }, + { + "bbox": [ + 131, + 488, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{Studio}}" + }, + { + "bbox": [ + 131, + 488, + 482, + 559 + ], + "type": "text", + "content": " often struggles to reproduce facial details that are visibly present in " + }, + { + "bbox": [ + 131, + 488, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{\\text{phone}}" + }, + { + "bbox": [ + 131, + 488, + 482, + 559 + ], + "type": "text", + "content": " itself. This limitation results in oversmoothed and inaccurate avatars. Achieving realistic facial details necessitates an accurate modeling of the facial texture distribution, a task that the StyleGAN2 fails to do." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 561, + 482, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 561, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 131, + 561, + 482, + 632 + ], + "type": "text", + "content": "Motivated by the recent success of latent diffusion models in accurately modeling data distributions, we propose the task of generating facial details as the reverse process of a Markov chain that transforms a low-resolution face texture map to a high-resolution one. We adopt the formulation from [43], where the residual between the low-res and high-res studio-captured texture maps, denoted as " + }, + { + "bbox": [ + 131, + 561, + 482, + 632 + ], + "type": "inline_equation", + "content": "e_0 = \\mathcal{I}_{LR}^{Studio} - \\mathcal{I}^{Studio}" + }, + { + "bbox": [ + 131, + 561, + 482, + 632 + ], + "type": "text", + "content": ", is used to define the Markov process. More specifically, the forward process is defined as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 189, + 632, + 482, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 632, + 482, + 656 + ], + "spans": [ + { + "bbox": [ + 189, + 632, + 482, + 656 + ], + "type": "interline_equation", + "content": "q \\left(\\mathcal {I} _ {t} ^ {\\text {S t u d i o}} \\mid \\mathcal {I} _ {t - 1} ^ {\\text {S t u d i o}}, \\mathcal {I} _ {L R} ^ {\\text {S t u d i o}}\\right) = \\mathcal {N} \\left(\\mathcal {I} _ {t} ^ {\\text {S t u d i o}}; \\mathcal {I} _ {t - 1} ^ {\\text {S t u d i o}} + \\alpha_ {t} e _ {0}, \\kappa^ {2} \\alpha_ {t} I\\right), \\tag {9}", + "image_path": "8fde6610b0726a09a31111cd2becb89d793f5614ed1d4317b15e19486b75d2ea.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 657, + 357, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 657, + 357, + 666 + ], + "spans": [ + { + "bbox": [ + 302, + 657, + 357, + 666 + ], + "type": "interline_equation", + "content": "t = 1, 2, \\dots , T,", + "image_path": "38066217a1501f425bfd8f4c6a2bb6d79826823bc01e26b739e4b77ff06ba696.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 156, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 156, + 91, + 448, + 103 + ], + "type": "text", + "content": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "text", + "content": " controls the schedule with which the residual is added, " + }, + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "text", + "content": " is a hyperparameter controlling the noise schedule and " + }, + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "text", + "content": " is the identity matrix. The reverse process is defined as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 153, + 482, + 181 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 153, + 482, + 181 + ], + "spans": [ + { + "bbox": [ + 149, + 153, + 482, + 181 + ], + "type": "interline_equation", + "content": "q \\left(\\mathcal {I} _ {t - 1} ^ {S t u d i o} \\mid \\mathcal {I} _ {\\{t, L R \\}} ^ {S t u d i o}, \\mathcal {I} ^ {S t u d i o}\\right) = \\mathcal {N} \\left(\\mathcal {I} _ {t - 1} ^ {S t u d i o} \\left| \\frac {\\eta_ {t - 1}}{\\eta_ {t}} \\mathcal {I} _ {t} ^ {S t u d i o} + \\frac {\\alpha_ {t}}{\\eta_ {t}} \\mathcal {I} ^ {S t u d i o}, \\kappa^ {2} \\frac {\\eta_ {t - 1}}{\\eta_ {t}} \\alpha_ {t} I\\right), \\right. \\tag {10}", + "image_path": "9b527fdb753048c8879272cefab18a1d8127e2161fd50706d0fcf47973f8520a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 189, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 482, + 224 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 189, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\eta_{t} = \\alpha_{t} + \\eta_{t - 1}" + }, + { + "bbox": [ + 130, + 189, + 482, + 224 + ], + "type": "text", + "content": ". We refer the reader to [43] for details regarding the formulation of both the forward and reverse process. Finally, the diffusion model, " + }, + { + "bbox": [ + 130, + 189, + 482, + 224 + ], + "type": "inline_equation", + "content": "f_{\\phi}" + }, + { + "bbox": [ + 130, + 189, + 482, + 224 + ], + "type": "text", + "content": ", is trained to minimize the following objective" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 227, + 228, + 482, + 246 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 228, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 227, + 228, + 482, + 246 + ], + "type": "interline_equation", + "content": "\\min _ {\\phi} \\sum_ {t} \\| f _ {\\phi} \\left(\\mathcal {I} _ {t} ^ {S t u d i o}, \\mathcal {I} _ {L R} ^ {S t u d i o}, t\\right) - \\mathcal {I} ^ {S t u d i o} \\| _ {2} ^ {2} \\tag {11}", + "image_path": "3b29992fb3c43f75ef945863a1270150dff6f21a0d0fba34350f20710691b3e0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 254, + 482, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 254, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 131, + 254, + 482, + 277 + ], + "type": "text", + "content": "During inference, we use " + }, + { + "bbox": [ + 131, + 254, + 482, + 277 + ], + "type": "inline_equation", + "content": "f_{\\phi}" + }, + { + "bbox": [ + 131, + 254, + 482, + 277 + ], + "type": "text", + "content": " to add realistic facial details to the low-resolution output of " + }, + { + "bbox": [ + 131, + 254, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{Studio}}" + }, + { + "bbox": [ + 131, + 254, + 482, + 277 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 252, + 272, + 482, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 272, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 252, + 272, + 482, + 285 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {I}} ^ {*} = \\operatorname {R e v e r s e P r o c e s s} \\left(f _ {\\phi}, \\mathcal {I} ^ {*}\\right). \\tag {12}", + "image_path": "dd7d66d0d3b6e5c4a8238700f1557e0d9e6a07d431ada1c01a4befb4e162ca24.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "text", + "content": "Recovering accurate facial details. While " + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "inline_equation", + "content": "f_{\\phi}" + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "text", + "content": " enhances " + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^*" + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "text", + "content": " by adding realistic facial details, it struggles to recover details already present in the phone-captured texture map " + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{phone}" + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "text", + "content": " due to the inherent lack of such details in " + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^*" + }, + { + "bbox": [ + 130, + 290, + 482, + 373 + ], + "type": "text", + "content": ". The loss of details occurs during the illumination manipulation and inpainting process, as described in Eq. (1) and Eq. (2). To address this issue, we incorporate the image gradient from the phone-captured texture map into the low-resolution texture map during the training of the diffusion model, as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 249, + 370, + 482, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 370, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 249, + 370, + 482, + 384 + ], + "type": "interline_equation", + "content": "\\mathcal {I} _ {L R} ^ {\\text {S t u d i o *}} = \\mathcal {I} _ {L R} ^ {\\text {S t u d i o}} + \\mathbb {G} (\\mathcal {I} ^ {\\text {p h o n e}}) \\tag {13}", + "image_path": "294c62b50304b8bab66dc4e7cd9d2c02c1a5654a73f140366b35730373aadbf9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 388, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 388, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{phone}" + }, + { + "bbox": [ + 130, + 388, + 482, + 449 + ], + "type": "text", + "content": " represents the phone-captured texture map, and " + }, + { + "bbox": [ + 130, + 388, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{Studio}" + }, + { + "bbox": [ + 130, + 388, + 482, + 449 + ], + "type": "text", + "content": " is the studio-captured texture map of the same person. " + }, + { + "bbox": [ + 130, + 388, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\mathbb{G}" + }, + { + "bbox": [ + 130, + 388, + 482, + 449 + ], + "type": "text", + "content": " denotes the operator used to calculate the image gradient. During training, " + }, + { + "bbox": [ + 130, + 388, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{LR}^{Studio*}" + }, + { + "bbox": [ + 130, + 388, + 482, + 449 + ], + "type": "text", + "content": " replaces " + }, + { + "bbox": [ + 130, + 388, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{LR}^{Studio}" + }, + { + "bbox": [ + 130, + 388, + 482, + 449 + ], + "type": "text", + "content": " in Eq. (9), Eq. (10), and Eq. (11). In the inference stage, we augment " + }, + { + "bbox": [ + 130, + 388, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^*" + }, + { + "bbox": [ + 130, + 388, + 482, + 449 + ], + "type": "text", + "content": " by adding the gradient of the phone-captured texture map as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 230, + 453, + 482, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 453, + 482, + 466 + ], + "spans": [ + { + "bbox": [ + 230, + 453, + 482, + 466 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {I}} ^ {*} = \\text {R e v e r s e P r o c e s s} \\left(f _ {\\phi}, \\mathcal {I} ^ {*} + \\mathbb {G} \\left(\\mathcal {I} ^ {\\text {p h o n e}}\\right)\\right) \\tag {14}", + "image_path": "e97a80dae7cd35097b3bf532e8eb0a13169b3b973c61973d5558e6e55e37272e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "content": "Implementation details We utilize 83 paired texture maps, representing subjects for whom we have both " + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{Studio}" + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{phone}" + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "content": ", to train the diffusion model. Following the approach in [43], we employ a latent diffusion model. To prevent overfitting, training is conducted on random " + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "content": " crops of the " + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "content": " resolution texture map. During the inference stage, we employ the full-resolution texture map." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 552, + 283, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 552, + 283, + 565 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 283, + 565 + ], + "type": "text", + "content": "3.3 Driving a high-quality avatar" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 573, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 573, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 573, + 482, + 609 + ], + "type": "text", + "content": "With the studio-lit version " + }, + { + "bbox": [ + 130, + 573, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^*" + }, + { + "bbox": [ + 130, + 573, + 482, + 609 + ], + "type": "text", + "content": " now available for a given neutral phone-captured texture map " + }, + { + "bbox": [ + 130, + 573, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{phone}" + }, + { + "bbox": [ + 130, + 573, + 482, + 609 + ], + "type": "text", + "content": ", we proceed to estimate a color transform mapping from the phone-captured texture map to the studio-lit texture map as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 203, + 613, + 482, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 613, + 482, + 633 + ], + "spans": [ + { + "bbox": [ + 203, + 613, + 482, + 633 + ], + "type": "interline_equation", + "content": "\\left\\{G, B \\right\\} \\leftarrow \\underset {\\{G, B \\}} {\\operatorname {a r g m i n}} \\left| \\left| \\mathcal {I} ^ {*} - \\left(\\mathrm {R s z} (G) \\times \\mathcal {I} ^ {\\text {p h o n e}} + \\mathrm {R s z} (B)\\right) \\right| \\right|. \\tag {15}", + "image_path": "01aa90bd93c49500b4e857333ebfbf2358bb1a063c4e1e519ff4e8266afa7bf2.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " represent gain and bias maps of resolution " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "k \\times k" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{R} \\times \\mathbb{Z}" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " is the resizing operator with bilinear interpolation. Utilizing " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": ", we perform a transformation" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "Athar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "on phone-captured expression texture maps to achieve studio-like lighting, as outlined below:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 238, + 140, + 482, + 153 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 140, + 482, + 153 + ], + "spans": [ + { + "bbox": [ + 238, + 140, + 482, + 153 + ], + "type": "interline_equation", + "content": "\\mathcal {I} _ {e x p} ^ {*} = \\operatorname {R s z} (G) \\times \\mathcal {I} _ {e x p} ^ {\\text {p h o n e}} + \\operatorname {R s z} (B). \\tag {16}", + "image_path": "b3a985034c9235988f1352bd5f16df107e23ec7c20bc7590e1f18f4379dfb1e4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 158, + 479, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 158, + 479, + 181 + ], + "spans": [ + { + "bbox": [ + 130, + 158, + 479, + 181 + ], + "type": "text", + "content": "In our experiment, we select the value of " + }, + { + "bbox": [ + 130, + 158, + 479, + 181 + ], + "type": "inline_equation", + "content": "k = 32" + }, + { + "bbox": [ + 130, + 158, + 479, + 181 + ], + "type": "text", + "content": " to efficiently transform the lighting while preserving the details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 182, + 479, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 479, + 229 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 479, + 229 + ], + "type": "text", + "content": "Now, the studio-like high-resolution neutral texture " + }, + { + "bbox": [ + 130, + 182, + 479, + 229 + ], + "type": "inline_equation", + "content": "\\hat{I}^*" + }, + { + "bbox": [ + 130, + 182, + 479, + 229 + ], + "type": "text", + "content": " serves as conditioning data for the universal avatar prior from [6]. By combining the expression code generated by inputting " + }, + { + "bbox": [ + 130, + 182, + 479, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{exp}^{*}" + }, + { + "bbox": [ + 130, + 182, + 479, + 229 + ], + "type": "text", + "content": " into the expression encoder from [6], we can render a high-quality avatar from any desired view " + }, + { + "bbox": [ + 130, + 182, + 479, + 229 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 182, + 479, + 229 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 250, + 239, + 481, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 239, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 250, + 239, + 481, + 251 + ], + "type": "interline_equation", + "content": "\\mathbf {I} = \\operatorname {A V A} \\left(\\hat {\\mathcal {I}} ^ {*}, \\mathcal {I} _ {e x p} ^ {*} - \\mathcal {I} ^ {*}, v, \\mathbf {F}\\right), \\tag {17}", + "image_path": "b6c3c20458425100c46b8a8f3e39c104c6dd43a7a4999623dc1f8227b047170b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 260, + 479, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 479, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 479, + 308 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 260, + 479, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 260, + 479, + 308 + ], + "type": "text", + "content": " represents the render of the avatar, and " + }, + { + "bbox": [ + 130, + 260, + 479, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 130, + 260, + 479, + 308 + ], + "type": "text", + "content": " is the geometry generated during 3D face tracking for a monocular phone capture. AVA corresponds to the inference process of the universal prior model. For more details about the universal prior model, please refer to [6]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 327, + 190, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 327, + 190, + 339 + ], + "spans": [ + { + "bbox": [ + 131, + 327, + 190, + 339 + ], + "type": "text", + "content": "4 Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 351, + 331, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 331, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 331, + 483 + ], + "type": "text", + "content": "In this section, we introduce the dataset used in this paper, along with the baselines of our method. Subsequently, we present both quantitative and qualitative results of our method, comparing it to prior work. All phone-captured texture maps are generated using the mesh fitting procedure outlined in Authentic Volumetric Avatars (AVA) [6]. Furthermore, we utilize AVA to render avatars based on the texture maps generated by all the methods, facilitating the calculation of image space metrics, including face identity similarity." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 338, + 408, + 478, + 460 + ], + "blocks": [ + { + "bbox": [ + 335, + 349, + 482, + 407 + ], + "lines": [ + { + "bbox": [ + 335, + 349, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 335, + 349, + 482, + 407 + ], + "type": "text", + "content": "Table 1: Quantitative results on paired phone-cum-studio captured texture maps. Our method outperforms prior work across all metrics. Best and Second Best scores are highlighted." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 338, + 408, + 478, + 460 + ], + "lines": [ + { + "bbox": [ + 338, + 408, + 478, + 460 + ], + "spans": [ + { + "bbox": [ + 338, + 408, + 478, + 460 + ], + "type": "table", + "html": "
ModelsPSNR ↑SSIM ↑LPIPS ↓DISTS ↓
Ours22.760.7260.3640.163
AgileGAN [38]18.050.6570.4060.180
CUT [33]21.4400.6420.4020.169
CycleGAN [42]21.0870.6430.4000.175
", + "image_path": "589263fe66c9ae45cf37636cdfbc92187676efd5106280303853419c00ba1217.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 133, + 584, + 274, + 612 + ], + "blocks": [ + { + "bbox": [ + 130, + 505, + 277, + 582 + ], + "lines": [ + { + "bbox": [ + 130, + 505, + 277, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 505, + 277, + 582 + ], + "type": "text", + "content": "Table 2: FaceID results on unpaired phone captured texture maps. Our method better preserves identity than prior work without sacrificing the quality of the generated texture maps. Best and Second Best scores are highlighted." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 584, + 274, + 612 + ], + "lines": [ + { + "bbox": [ + 133, + 584, + 274, + 612 + ], + "spans": [ + { + "bbox": [ + 133, + 584, + 274, + 612 + ], + "type": "table", + "html": "
ModelsOursAgileGANCUTCycleGAN
FaceID4.31e-41.36e-36.89e-45.19e-4
", + "image_path": "4d256fe2059e3de284af8d1b01f53d4732320b634406d5a98d347929ba606b67.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 282, + 484, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 484, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 282, + 484, + 482, + 639 + ], + "type": "text", + "content": "Training and Evaluation Data. We utilize a dataset comprising 12,543 neutral phone-captured texture maps to train " + }, + { + "bbox": [ + 282, + 484, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{phone}" + }, + { + "bbox": [ + 282, + 484, + 482, + 639 + ], + "type": "text", + "content": ", and 383 studio-captured texture maps for fine-tuning " + }, + { + "bbox": [ + 282, + 484, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{phone}" + }, + { + "bbox": [ + 282, + 484, + 482, + 639 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 282, + 484, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{Studio}" + }, + { + "bbox": [ + 282, + 484, + 482, + 639 + ], + "type": "text", + "content": ". Among the 383 maps, 83 are paired neutral texture maps " + }, + { + "bbox": [ + 282, + 484, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{I}^{phone},\\mathcal{I}^{Studio}\\}" + }, + { + "bbox": [ + 282, + 484, + 482, + 639 + ], + "type": "text", + "content": " used for calculating the perceptual reconstruction loss, as described in Eq. (6). This paired dataset also serves for training the detail-preserving diffusion model. For quantitative and qualitative evaluation, we employ 10 paired phone-cum-studio captured texture maps and 31 phone-captured texture maps." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Baselines. We compare our method to the following prior works on unpaired image-to-image translation. 1) AgileGAN [38], which utilizes an aligned StyleGAN latent space" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 156, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 156, + 91, + 448, + 103 + ], + "type": "text", + "content": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "to stylize images, even with very few examples; 2) Contrastive Unpaired Translation (CUT) [33], which employs a patch-based contrastive loss in a learned feature space, enabling domain adaptation, even for single images; 3) CycleGAN [42], which trains two generators based on a cycle-consistency loss to translate images between two domains. Since AgileGAN [38] has no code available, we implement it ourselves. We use publicly available code for CUT [33] and [42] for all experiments." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 167, + 214, + 447, + 406 + ], + "blocks": [ + { + "bbox": [ + 167, + 214, + 447, + 406 + ], + "lines": [ + { + "bbox": [ + 167, + 214, + 447, + 406 + ], + "spans": [ + { + "bbox": [ + 167, + 214, + 447, + 406 + ], + "type": "image", + "image_path": "01c936b38159dfa3bcf75574d26e9bf90d555086149ddf43078d61cb2bff9b37.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 414, + 482, + 468 + ], + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 468 + ], + "type": "text", + "content": "Fig. 3: Comparisons with baselines on paired phone-cum-studio texture data reveal that our method produces results closest to the ground truth. It achieves uniform studio-like lighting, well-reconstructed facial details visible in the phone capture, and effective inpainting of missing regions. In contrast, AgileGAN [38] fails to preserve identity, while CUT [33] and CycleGAN [42] introduce significant artifacts." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 495, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 615 + ], + "type": "text", + "content": "Quantitative Results. In Table 1, we present the results of a quantitative evaluation comparing our method to the baselines using 10 paired texture maps captured with both phones and in a studio setting. The evaluation metrics include mean PSNR, SSIM, LPIPS [21], and DISTS [9]. DISTS and LPIPS are perceptual metrics aligned with human judgment, while PSNR and SSIM are pixel-based metrics. As shown in Table 1, our method consistently outperforms the baseline across all metrics. The improvement is not limited to perceptual metrics but also extends to PSNR and SSIM, highlighting the efficacy of our proposed approach. It is worth noting that the relatively low PSNR and SSIM scores can be attributed to their sensitivity to small pixel shifts between the studio and phone-captured textures." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "In Table 2, we display the face embedding distances, measured using Eq. (4), between avatars rendered from the phone-captured texture and the textures generated by various methods on the 31 unpaired phone captures. Our method exhibits the best preservation of facial identity." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "Athar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 176, + 114, + 438, + 376 + ], + "blocks": [ + { + "bbox": [ + 176, + 114, + 438, + 376 + ], + "lines": [ + { + "bbox": [ + 176, + 114, + 438, + 376 + ], + "spans": [ + { + "bbox": [ + 176, + 114, + 438, + 376 + ], + "type": "image", + "image_path": "5d5a5b29b3ae5edc8c829d17bda2c83d0a82badbeba3db073e3f249d7789f125.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 384, + 482, + 418 + ], + "lines": [ + { + "bbox": [ + 131, + 384, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 131, + 384, + 482, + 418 + ], + "type": "text", + "content": "Fig. 4: Comparisons on Unpaired Phone-Captured Data: In comparison to prior work, our method excels in generating texture maps with superior preservation of identity, enhanced photorealism in facial details, more uniform illumination, and improved inpainting of missing regions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 452, + 482, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 452, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 452, + 482, + 501 + ], + "type": "text", + "content": "Qualitative Results. In Fig 3, we provide some qualitative results on the paired test data. It is evident that our method generates a more plausible and photorealistic texture map compared to prior work, showcasing improvements in illumination transfer, facial details, and inpainting of missing regions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "In Fig 4, we present qualitative results comparing our method with prior work. While AgileGAN [38] successfully changes the illumination to be studio-like and inpaints missing regions, it introduces a significant identity shift (quantitatively measured in Table 2) and lacks facial details. We attribute this to the use of the " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "Z^{+}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " space [38], which may not be flexible enough for high-fidelity inversion, and the absence of identity-preserving constraints, both in architecture and optimization, during training. Due to its contrastive training paradigm, CUT [33] preserves identity better than AgileGAN but introduces significant artifacts. It is also unable to inpaint the missing regions around the ears and corners of the head. Similarly, like CUT [33], CycleGAN [42] also preserves identity better than AgileGAN [38] but struggles with inpainting missing regions. The textures also contain numerous uncanny artifacts that are uncharacteristic of human skin. In contrast, our method generates a high-quality studio-illuminated texture map with accurate facial details and inpainted missing regions." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 448, + 102 + ], + "type": "text", + "content": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 202, + 114, + 411, + 208 + ], + "blocks": [ + { + "bbox": [ + 202, + 114, + 411, + 208 + ], + "lines": [ + { + "bbox": [ + 202, + 114, + 411, + 208 + ], + "spans": [ + { + "bbox": [ + 202, + 114, + 411, + 208 + ], + "type": "image", + "image_path": "5a8c586c7d97e1c4cbef53a27ac8700df4a6a5cd48875a36b4568969354b7e69.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 215, + 482, + 270 + ], + "lines": [ + { + "bbox": [ + 131, + 215, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 131, + 215, + 482, + 270 + ], + "type": "text", + "content": "Fig. 5: Avatar Reanimation. The top row showcases a multiview render of an avatar generated using studio-like neutral and expression texture maps, created through our method as described in Sect 3.3. In contrast, the bottom row utilizes phone-captured texture maps. Notably, the lip region, highlighted with pink rectangles, appears more realistic and less blurry when using studio-like texture maps generated by our method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 238, + 273, + 378, + 441 + ], + "blocks": [ + { + "bbox": [ + 238, + 273, + 378, + 441 + ], + "lines": [ + { + "bbox": [ + 238, + 273, + 378, + 441 + ], + "spans": [ + { + "bbox": [ + 238, + 273, + 378, + 441 + ], + "type": "image", + "image_path": "961325b567e324f045bca7a2d849bafe8e750a0a420f5ab6b43435330c5bde4b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 449, + 482, + 495 + ], + "lines": [ + { + "bbox": [ + 131, + 449, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 131, + 449, + 482, + 495 + ], + "type": "text", + "content": "Fig. 6: Comparison of Avatars Generated Using [6]: In this comparison, we evaluate the quality of avatars generated by the Universal Prior Model (UPM) from AVA [6], using texture maps from various methods as input. Evidently, the avatars generated using texture maps from our method far exceed those generated by prior work." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 508, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 508, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 131, + 508, + 482, + 628 + ], + "type": "text", + "content": "Reanimation Results. As explained in Sect 3.3, we can transform the phone expression texture maps " + }, + { + "bbox": [ + 131, + 508, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{exp}^{phone}" + }, + { + "bbox": [ + 131, + 508, + 482, + 628 + ], + "type": "text", + "content": " to studio-like illumination using Eq. (15) and Eq. (16). We observe that utilizing studio-like illuminated expression texture maps " + }, + { + "bbox": [ + 131, + 508, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{*}_{exp}" + }, + { + "bbox": [ + 131, + 508, + 482, + 628 + ], + "type": "text", + "content": " results in a modest improvement in the quality of reanimated avatars using the universal prior model from AVA [6]. In Fig 5, we present an example where the top row shows the render of an avatar using studio-like neutral and expression textures, while the bottom row displays a render using phone-captured neutral and expression textures. It is evident that the lip region appears more realistic and less blurry when reanimated with the studio-like neutral and expression textures compared to the phone-captured textures. We recommend that readers refer to our supplementary video for a more comprehensive comparison." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 646, + 200, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 646, + 200, + 657 + ], + "spans": [ + { + "bbox": [ + 132, + 646, + 200, + 657 + ], + "type": "text", + "content": "4.1 Ablations" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "Athar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 329, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 329, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 329, + 140 + ], + "type": "text", + "content": "In this section, we conduct ablation studies to evaluate the various components of our method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 263, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 263, + 162 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 263, + 162 + ], + "type": "text", + "content": "Details Conditioned Diffusion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 163, + 329, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 163, + 329, + 222 + ], + "spans": [ + { + "bbox": [ + 131, + 163, + 329, + 222 + ], + "type": "text", + "content": "We ablate the necessity of using facial details extracted from the phone capture texture map, in the form of an image gradient, to synthesize accurate facial details using a diffusion model. We explore three scenarios:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 224, + 330, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 224, + 330, + 272 + ], + "spans": [ + { + "bbox": [ + 131, + 224, + 330, + 272 + ], + "type": "text", + "content": "1) A Diffusion model that does not use any conditioning on phone-captured texture gradient (Vanilla Diffusion); 2) A Diffusion model that uses the phone-captured texture gradient only dur" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 272, + 481, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 272, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 131, + 272, + 481, + 295 + ], + "type": "text", + "content": "ing inference but not during training; 3) Our model, which incorporates the phone-captured texture gradient during both training and inference." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 297, + 330, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 297, + 330, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 297, + 330, + 475 + ], + "type": "text", + "content": "In Fig 7, we present the results of each of the three scenarios. While the Vanilla Diffusion generates a realistic-looking facial texture, it fails to preserve facial details present in the phone-captured texture map. Prominent moles, marked by the blue boxes in the phone-captured texture map, are not generated by the Vanilla Diffusion model. When conditioning the reverse process using the phone-texture gradient only during inference, we observe that the model misses some details and transfers shading from the phone-captured texture map to the studio-lit texture map (marked by red boxes in Fig 7), which is undesirable. Ideally, we want facial details to be preserved while eliminating illumination-dependent effects, such as strong" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 335, + 116, + 482, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 116, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 335, + 116, + 482, + 205 + ], + "type": "text", + "content": "Table 3: Detail conditioning ablation. We calculate quantitative metrics on the 10 paired phone-cum-studio captured texture maps. It is evident that utilizing detail conditioning during both training and inference yields the best performance. Best and Second Best scores are highlighted." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 338, + 213, + 487, + 251 + ], + "blocks": [ + { + "bbox": [ + 338, + 213, + 487, + 251 + ], + "lines": [ + { + "bbox": [ + 338, + 213, + 487, + 251 + ], + "spans": [ + { + "bbox": [ + 338, + 213, + 487, + 251 + ], + "type": "table", + "html": "
ModelsLPIPS ↓DISTS ↓
Vanilla Diffusion0.3830.179
Inference-only Details conditioning0.3760.183
Training + Inference Details conditioning (Ours)0.3640.163
", + "image_path": "92d09d3f4ba33a42b5d5d1a2adf5e7ae272f8748f3c2329e7fa790369d4b43ae.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 335, + 340, + 482, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 340, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 335, + 340, + 482, + 417 + ], + "type": "text", + "content": "Table 4: " + }, + { + "bbox": [ + 335, + 340, + 482, + 417 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 335, + 340, + 482, + 417 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 335, + 340, + 482, + 417 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 335, + 340, + 482, + 417 + ], + "type": "text", + "content": " Space ablation. We utilize the 10 paired phone-cum-studio captured texture maps to compare the outcomes of training in the " + }, + { + "bbox": [ + 335, + 340, + 482, + 417 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 335, + 340, + 482, + 417 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 335, + 340, + 482, + 417 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 335, + 340, + 482, + 417 + ], + "type": "text", + "content": " spaces. The results clearly demonstrate that training in the " + }, + { + "bbox": [ + 335, + 340, + 482, + 417 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 335, + 340, + 482, + 417 + ], + "type": "text", + "content": " space yields significantly better outcomes." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 343, + 418, + 476, + 451 + ], + "blocks": [ + { + "bbox": [ + 343, + 418, + 476, + 451 + ], + "lines": [ + { + "bbox": [ + 343, + 418, + 476, + 451 + ], + "spans": [ + { + "bbox": [ + 343, + 418, + 476, + 451 + ], + "type": "table", + "html": "
ModelsPSNR ↑SSIM ↑LPIPS ↓DISTS ↓
Z Space21.3680.7180.4010.172
W+ Space (Ours)22.760.7260.3640.163
", + "image_path": "1e271e6deb7696597dee1211209fad9d085d0c47480fec2abf1160fb4f6845cd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 475, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 475, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 131, + 475, + 482, + 571 + ], + "type": "text", + "content": "shading and shadows, from the studio-lit texture map. Our method conditions the diffusion model on the phone-texture gradient during both inference and training. As shown in Fig 7, this allows the model to learn to retain facial details while ignoring illumination-induced effects, such as strong shading, shadows, and specularities when synthesizing facial details. In Table 3, we quantitatively validate each of the aforementioned diffusion model designs on paired data using LPIPS [21] and DISTS [9]. As can be seen, conditioning the diffusion model on phone-texture gradients both during training and inference gives the best results." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "content": " space vs. " + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "content": " space for finetuning. We now evaluate the effectiveness of the " + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "content": " space over the standard " + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "content": " space for fine-tuning " + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{Studio}}" + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "content": ". For a quantitative comparison, we calculate the FaceID loss over the 31 unpaired evaluation texture maps and compute PSNR, SSIM, LPIPS [21], and DISTS [9] over the 10 paired studio-cum-phone texture maps. Qualitative results, along with the average FaceID distance over unpaired data, are shown in Fig 8, and the metrics on paired data are presented in Table 4. As evident, sampling in the " + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "content": " space generalizes much better to unseen identities, exhibiting" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 448, + 102 + ], + "type": "text", + "content": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 176, + 114, + 438, + 218 + ], + "blocks": [ + { + "bbox": [ + 176, + 114, + 438, + 218 + ], + "lines": [ + { + "bbox": [ + 176, + 114, + 438, + 218 + ], + "spans": [ + { + "bbox": [ + 176, + 114, + 438, + 218 + ], + "type": "image", + "image_path": "190ce0af12d939452e07756539f943fafe4677c1219e171617e66fc70c4e83eb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 225, + 482, + 290 + ], + "lines": [ + { + "bbox": [ + 130, + 225, + 482, + 290 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 482, + 290 + ], + "type": "text", + "content": "Fig. 7: Ablation on Detail-conditioned Diffusion. When no detail conditioning is applied, facial details visible in the phone capture are not reproduced. When detail conditioning is employed only during inference, undesirable shading effects are transferred to the generated texture map. Optimal results, with the most accurate and plausible reproduction of facial details, are achieved when detail conditioning is applied during both training and inference. Please refer to the text for further details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 185, + 297, + 430, + 357 + ], + "blocks": [ + { + "bbox": [ + 185, + 297, + 430, + 357 + ], + "lines": [ + { + "bbox": [ + 185, + 297, + 430, + 357 + ], + "spans": [ + { + "bbox": [ + 185, + 297, + 430, + 357 + ], + "type": "image", + "image_path": "443bdafb5b111eace363ef8fe789a930e970f627223ae84a622e8346efa6040d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 365, + 482, + 399 + ], + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 399 + ], + "type": "text", + "content": "Fig. 8: Qualitative " + }, + { + "bbox": [ + 130, + 365, + 482, + 399 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 130, + 365, + 482, + 399 + ], + "type": "text", + "content": " space ablation. We observe that when " + }, + { + "bbox": [ + 130, + 365, + 482, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{Studio}}" + }, + { + "bbox": [ + 130, + 365, + 482, + 399 + ], + "type": "text", + "content": " is finetuned using the " + }, + { + "bbox": [ + 130, + 365, + 482, + 399 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 365, + 482, + 399 + ], + "type": "text", + "content": " space, it fails to generalize to novel subjects and exhibits significant artifacts. In contrast, training " + }, + { + "bbox": [ + 130, + 365, + 482, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{Studio}}" + }, + { + "bbox": [ + 130, + 365, + 482, + 399 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 130, + 365, + 482, + 399 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 130, + 365, + 482, + 399 + ], + "type": "text", + "content": " lends it far better generalization." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "type": "text", + "content": "significantly fewer artifacts and yielding superior results in terms of FaceID, PSNR, SSIM, LPIPS, and DISTS. We posit that this is because samples in the " + }, + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "type": "text", + "content": " space (i.e., samples from " + }, + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "type": "inline_equation", + "content": "S_{W^{+}}" + }, + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "type": "text", + "content": " space) are near-perfect inversions of the training data, making them more diverse than those generated from the Gaussian-distributed " + }, + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "type": "text", + "content": " space. Consequently, as seen in Fig 8, this leads to better generalization to unseen subjects. The model trained using the " + }, + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "type": "text", + "content": " space exhibits uncanny artifacts in its results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 498, + 296, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 296, + 510 + ], + "type": "text", + "content": "5 Conclusion and Limitations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 483, + 666 + ], + "type": "text", + "content": "In this paper, we present a method for generating studio-like, high-quality avatars from monocular phone captures, using a StyleGAN2-based image-to-image translation and diffusion-based image upsampling. Experiments show the effectiveness of our approach in manipulating lighting, inpainting missing parts and generating facial details thus enabling the creation of complete, studio-lit textures for rendering high-quality avatars. However, our method does have limitations. It struggles with input textures exhibiting extreme non-uniform lighting due to the constrained lighting conditions in our training data (we include examples in the supplementary). We also do not fine-tune the universal prior model to personalize the avatar based on phone capture data, as suggested in [6], thus the avatars lack personalized details for different facial expressions. Finally, our avatars are incomplete, featuring only the head. Future work involves extending the model to include the neck, shoulders, hands, and the entire body." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "Athar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 481, + 664 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 141, + 481, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 481, + 175 + ], + "type": "text", + "content": "1. Alexander, O., Rogers, M., Lambeth, W., Chiang, J., Ma, W., Wang, C., Debevec, P.: The digital emily project: Achieving a photoreal digital actor. IEEE Computer Graphics and Applications 30 (2009) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 176, + 481, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 481, + 208 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 481, + 208 + ], + "type": "text", + "content": "2. Alexander, O., Fyffe, G., Busch, J., Yu, X., Ichikari, R., Jones, A., Debevec, P., Jimenez, J., Danvoye, E., Antionazzi, B., et al.: Digital ira: Creating a real-time photoreal digital actor. In: ACM SIGGRAPH 2013 Posters, pp. 1-1 (2013) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 209, + 481, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 481, + 229 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 481, + 229 + ], + "type": "text", + "content": "3. Athar, S., Xu, Z., Sunkavalli, K., Shechtman, E., Shu, Z.: Rignerf: Fully controllable neural 3d portraits. In: CVPR (June 2022) 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 232, + 481, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 481, + 263 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 481, + 263 + ], + "type": "text", + "content": "4. Bi, S., Lombardi, S., Saito, S., Simon, T., Wei, S.E., Mcphail, K., Ramamoorthi, R., Sheikh, Y., Saragih, J.: Deep reightable appearance models for animatable faces. ACM Transactions on Graphics (TOG) 40(4), 1-15 (2021) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 265, + 481, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 265, + 481, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 265, + 481, + 286 + ], + "type": "text", + "content": "5. Borshukov, G., Lewis, J.P.: Realistic human face rendering for\" the matrix reloaded\". In: ACM Siggraph 2005 Courses, pp. 13-es (2005) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 288, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 288, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 288, + 481, + 319 + ], + "type": "text", + "content": "6. Cao, C., Simon, T., Kim, J.K., Schwartz, G., Zollhoefer, M., Saito, S.S., Lombardi, S., Wei, S.E., Belko, D., Yu, S.I., Sheikh, Y., Saragih, J.: Authentic volumetric avatars from a phone scan. ACM Trans. Graph. (2022) 1, 2, 3, 4, 5, 9, 12, 14" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 321, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 481, + 342 + ], + "type": "text", + "content": "7. Cao, C., Wu, H., Weng, Y., Shao, T., Zhou, K.: Real-time facial animation with image-based dynamic avatars. ACM Transactions on Graphics 35(4) (2016) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 343, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 481, + 374 + ], + "type": "text", + "content": "8. Casas, D., Alexander, O., Feng, A.W., Fyffe, G., Ichikari, R., Debevec, P., Wang, R., Suma, E., Shapiro, A.: Rapid photorealistic blendshapes from commodity rgb-d sensors. In: Proceedings of the 19th Symposium on Interactive 3D Graphics and Games. pp. 134-134 (2015) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 376, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 481, + 407 + ], + "type": "text", + "content": "9. Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image quality assessment: Unifying structure and texture similarity 44(5), 2567-2581 (2022). https://doi.org/10.1109/TPAMI.2020.3045810 10, 13" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 410, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 410, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 138, + 410, + 481, + 441 + ], + "type": "text", + "content": "0. Gafni, G., Thies, J., Zollhofer, M., Nießner, M.: Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8649-8658 (2021) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 443, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 443, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 443, + 481, + 475 + ], + "type": "text", + "content": "1. Garrido, P., Zollhöfer, M., Casas, D., Valgaerts, L., Varanasi, K., Pérez, P., Theobalt, C.: Reconstruction of personalized 3d face rigs from monocular video. ACM Transactions on Graphics (TOG) 35(3), 1-15 (2016) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 476, + 481, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 476, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 476, + 481, + 498 + ], + "type": "text", + "content": "2. Ghosh, A., Fyffe, G., Tunwattanapong, B., Busch, J., Yu, X., Debevec, P.: Multiview face capture using polarized spherical gradient illumination. ACM Trans. Graph. (2011) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 498, + 481, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 481, + 530 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 481, + 530 + ], + "type": "text", + "content": "3. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139-144 (2020) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 532, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 532, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 532, + 481, + 553 + ], + "type": "text", + "content": "4. Grassal, P.W., Prinzler, M., Leistner, T., Rother, C., Nießner, M., Thies, J.: Neural head avatars from monocular rgb videos. In: CVPR. pp. 18653-18664 (June 2022) 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 555, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 586 + ], + "type": "text", + "content": "5. Hu, L., Saito, S., Wei, L., Nagano, K., Seo, J., Fursund, J., Sadeghi, I., Sun, C., Chen, Y.C., Li, H.: Avatar digitization from a single image for real-time rendering. ACM Transactions on Graphics (ToG) 36(6), 1-14 (2017) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 588, + 481, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 481, + 619 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 481, + 619 + ], + "type": "text", + "content": "6. Huang, X., Liu, M.Y., Belongie, S., Kautz, J.: Multimodal unsupervised image-to-image translation. In: Proceedings of the European conference on computer vision (ECCV). pp. 172-189 (2018) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 621, + 481, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 481, + 642 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 481, + 642 + ], + "type": "text", + "content": "7. Ichim, A.E., Bouaziz, S., Pauly, M.: Dynamic 3d avatar creation from hand-held video input. ACM Transactions on Graphics (ToG) 34(4), 1-14 (2015) 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 643, + 481, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 664 + ], + "type": "text", + "content": "8. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: CVPR (2017) 2, 4" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 148 + ], + "type": "text", + "content": "19. Jimenez, J., Echevarria, J.I., Oat, C., Gutierrez, D.: Practical and realistic facial wrinkles animation. In: GPU Pro 360 Guide to Geometry Manipulation, pp. 95-107. AK Peters/CRC Press (2018) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "20. Jimenez, J., Scully, T., Barbosa, N., Donner, C., Alvarez, X., Vieira, T., Matts, P., Orvalho, V., Gutierrez, D., Weyrich, T.: A practical appearance model for dynamic facial color. In: ACM SIGGRAPH Asia 2010 papers, pp. 1-10 (2010) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 481, + 205 + ], + "type": "text", + "content": "21. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and superresolution. In: ECCV (2016) 10, 13" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 206, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 481, + 228 + ], + "type": "text", + "content": "22. Karras, T., Aittala, M., Hellsten, J., Laine, S., Lehtinen, J., Aila, T.: Training generative adversarial networks with limited data. In: Proc. NeurIPS (2020) 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 229, + 481, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 481, + 250 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 481, + 250 + ], + "type": "text", + "content": "23. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: CVPR (2019) 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 251, + 481, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 294 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 294 + ], + "type": "text", + "content": "24. Lattas, A., Moschoglou, S., Gecer, B., Ploumpis, S., Triantafyllou, V., Ghosh, A., Zafeiriou, S.: Avatarme: Realistically renderable 3d facial reconstruction\" in-the-wild\". In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 760-769 (2020) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 297, + 481, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 328 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 328 + ], + "type": "text", + "content": "25. Lattas, A., Moschoglou, S., Ploumpis, S., Gecer, B., Ghosh, A., Zafeiriou, S.: Avatarme++: Facial shape and brdf inference with photorealistic rendering-aware gans. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9269-9284 (2021) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 330, + 481, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 481, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 481, + 351 + ], + "type": "text", + "content": "26. Lin, J., Yuan, Y., Zou, Z.: Meingame: Create a game character face from a single portrait. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 311-319 (2021) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 353, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 481, + 373 + ], + "type": "text", + "content": "27. Lombardi, S., Saragih, J., Simon, T., Sheikh, Y.: Deep appearance models for face rendering. ACM Transactions on Graphics (ToG) 37(4), 1-13 (2018) 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "type": "text", + "content": "28. Lombardi, S., Simon, T., Schwartz, G., Zollhoefer, M., Sheikh, Y., Saragih, J.: Mixture of volumetric primitives for efficient neural rendering. ACM Transactions on Graphics (ToG) 40(4), 1-13 (2021) 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "type": "text", + "content": "29. Luo, H., Nagano, K., Kung, H., Xu, Q., Wang, Z., Wei, L., Hu, L., Li, H.: Normalized avatar synthesis using stylegan and perceptual refinement. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 441, + 481, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 481, + 474 + ], + "type": "text", + "content": "30. Ma, S., Simon, T., Saragih, J., Wang, D., Li, Y., De La Torre, F., Sheikh, Y.: Pixel codec avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 64-73 (2021) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 475, + 481, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 496 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 496 + ], + "type": "text", + "content": "31. Mescheder, L., Geiger, A., Nowozin, S.: Which training methods for gans do actually converge? (2018) 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "type": "text", + "content": "32. Nagano, K., Fyffe, G., Alexander, O., Barbic, J., Li, H., Ghosh, A., Debevec, P.E.: Skin microstructure deformation with displacement map convolution. ACM Trans. Graph. 34(4), 109-1 (2015) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 532, + 481, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 481, + 552 + ], + "type": "text", + "content": "33. Park, T., Efros, A.A., Zhang, R., Zhu, J.Y.: Contrastive learning for unpaired image-to-image translation. In: European Conference on Computer Vision (2020) 9, 10, 11" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 554, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 481, + 575 + ], + "type": "text", + "content": "34. Pinkney, J.N., Adler, D.: Resolution dependent gan interpolation for controllable image synthesis between domains. arXiv preprint arXiv:2010.05334 (2020) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 576, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 481, + 609 + ], + "type": "text", + "content": "35. Sang, S., Zhi, T., Song, G., Liu, M., Lai, C., Liu, J., Wen, X., Davis, J., Luo, L.: Agileavatar: Stylized 3d avatar creation via cascaded domain bridging. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-8 (2022) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 610, + 481, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 631 + ], + "type": "text", + "content": "36. Seymour, M., Evans, C., Libreri, K.: Meet mike: epic avatars. In: ACM SIGGRAPH 2017 VR Village, pp. 1-2 (2017) 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "37. Shi, T., Yuan, Y., Fan, C., Zou, Z., Shi, Z., Liu, Y.: Face-to-parameter translation for game character auto-creation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 161–170 (2019) 4" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "Athar et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 303 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "38. Song, G., Luo, L., Liu, J., Ma, W.C., Lai, C., Zheng, C., Cham, T.J.: Agilegan: stylizing portraits by inversion-consistent transfer learning. ACM Transactions on Graphics (TOG) 40(4), 1-13 (2021) 2, 4, 9, 10, 11" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 172 + ], + "type": "text", + "content": "39. Wang, T.C., Liu, M.Y., Tao, A., Liu, G., Kautz, J., Catanzaro, B.: Few-shot video-to-video synthesis (2019) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 482, + 205 + ], + "type": "text", + "content": "40. Wang, T.C., Liu, M.Y., Zhu, J.Y., Tao, A., Kautz, J., Catanzaro, B.: High-resolution image synthesis and semantic manipulation with conditional gans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8798-8807 (2018) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 482, + 237 + ], + "type": "text", + "content": "41. Yamaguchi, S., Saito, S., Nagano, K., Zhao, Y., Chen, W., Olszewski, K., Morishima, S., Li, H.: High-fidelity facial reflectance and geometry inference from an unconstrained image. ACM Transactions on Graphics (TOG) 37(4), 1-14 (2018) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 482, + 270 + ], + "type": "text", + "content": "42. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networkss. In: Computer Vision (ICCV), 2017 IEEE International Conference on (2017) 2, 4, 9, 10, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 482, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 482, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 482, + 303 + ], + "type": "text", + "content": "43. Zongsheng Yue, J.W., Loy, C.C.: Resshift: Efficient diffusion model for image super-resolution by residual shifting. In: Advances in Neural Information Processing Systems (NeurIPS) (2023) 7, 8" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bridging the Gap: Studio-like Avatar Creation from a Monocular Phone Capture" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Bridging the Pathology Domain Gap_ Efficiently Adapting CLIP for Pathology Image Analysis with Limited Labeled Data/3b6fb56d-1649-4e06-bdb2-706d8b03ddb5_content_list.json b/2024/Bridging the Pathology Domain Gap_ Efficiently Adapting CLIP for Pathology Image Analysis with Limited Labeled Data/3b6fb56d-1649-4e06-bdb2-706d8b03ddb5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..553bd95ab4ab024feae874c6083556980568ca30 --- /dev/null +++ b/2024/Bridging the Pathology Domain Gap_ Efficiently Adapting CLIP for Pathology Image Analysis with Limited Labeled Data/3b6fb56d-1649-4e06-bdb2-706d8b03ddb5_content_list.json @@ -0,0 +1,1552 @@ +[ + { + "type": "text", + "text": "Bridging the Pathology Domain Gap: Efficiently Adapting CLIP for Pathology Image Analysis with Limited Labeled Data", + "text_level": 1, + "bbox": [ + 217, + 140, + 785, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhengfeng Lai1, Joohi Chauhan2, Brittany N. Dugger1, and Chen-Nee Chuah1", + "bbox": [ + 328, + 241, + 671, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of California, Davis \n2 Motilal Nehru National Institute of Technology Allahabad {lzhengfeng, bndugger, chuah}@ucdavis.edu joohi@mnnit.ac.in", + "bbox": [ + 300, + 282, + 700, + 340 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Contrastive Language-Image Pre-training (CLIP) has shown its proficiency in acquiring distinctive visual representations and exhibiting strong generalization across diverse vision tasks. However, its effectiveness in pathology image analysis, particularly with limited labeled data, remains an ongoing area of investigation due to challenges associated with significant domain shifts and catastrophic forgetting. Thus, it is imperative to devise efficient adaptation strategies in this domain to enable scalable analysis. In this study, we introduce Path-CLIP, a framework tailored for a swift adaptation of CLIP to various pathology tasks. Firstly, we propose Residual Feature Refinement (RFR) with a dynamically adjustable ratio to effectively integrate and balance source and task-specific knowledge. Secondly, we introduce Hidden Representation Perturbation (HRP) and Dual-view Vision Contrastive (DVC) techniques to mitigate overfitting issues. Finally, we present the Doublet Multimodal Contrastive Loss (DMCL) for fine-tuning CLIP for pathology tasks. We demonstrate that Path-CLIP adeptly adapts pre-trained CLIP to downstream pathology tasks, yielding competitive results. Specifically, Path-CLIP achieves over $+19\\%$ improvement in accuracy when utilizing mere $0.1\\%$ of labeled data in PCam with only 10 minutes of fine-tuning while running on a single GPU.", + "bbox": [ + 261, + 359, + 743, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 659, + 375, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in deep learning, bolstered by superior network architectures and the availability of extensive, meticulously curated datasets, have led to substantial enhancements in pathology image analysis tasks [17]. Nevertheless, the process of collecting high-quality datasets with reliable annotations for each specific vision task remains a laborious and time-intensive endeavor [18,53]. This may prevent the broad adoption of advanced deep learning techniques. To relieve the reliance on such datasets, pre-training and fine-tuning methods (TFT) have been studied in vision tasks: pre-train the model on a large-scale dataset and then fine-tune the model on different downstream tasks [9]. There are several challenges of such methods: 1) they may still require a large amount of labeled", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/035939840df9d49f945cdec888d3339b46dc0c353203c6d467ff41819d128dd2.jpg", + "image_caption": [ + "(a) CLIP" + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 504, + 281 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3ac9c8ba0cb6d4a4780890b6a0397c57575cec8aac5b5ff4b35b60f5a6766d8e.jpg", + "image_caption": [ + "(b) Path-CLIP", + "Fig. 1: Illustration of (a) zero-shot CLIP, and (b) our proposed Path-CLIP. We aim to freeze the encoders but only train the Residual Feature Refinement (RFR) module to balance the original knowledge from CLIP and new task-specific knowledge." + ], + "image_footnote": [], + "bbox": [ + 513, + 143, + 784, + 281 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "set to avoid the overfitting issue when fine-tuning the model for the downstream task [21, 40]; 2) the fine-tuning may not yield satisfactory performance in the target domain due to the existence of a large domain gap between the pre-trained data and pathology images [32]. Consequently, a pivotal question arises: how can the large-scale pre-trained model be efficiently tailored to a new task, minimizing overfitting, particularly when the new dataset is limited in size?", + "bbox": [ + 212, + 383, + 787, + 476 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To fill the performance gap due to domain shift, Contrastive Language-Image Pre-training (CLIP) [26] has shown its power in learning generic and distinctive visual representations via language supervision. As shown in Fig. 1(a), it aligns images and texts in the same feature space and uses a contrastive loss to formulate the learning objective. CLIP uses two separate encoders for images and texts, then maximizes the similarity score of positive pairs of images and texts while minimizing it of negative pairs [26,54]. It achieves promising results on various image classification tasks without needing any annotated data, i.e., zero-shot transfer settings. As a language-vision model, CLIP uses prompts as the supervision, where the visual labels are entered into the hand-crafted template. By pre-training the model at a large scale, models can learn the visual contents and easily be transferred to downstream tasks through the prompt-based zero-shot transfer.", + "bbox": [ + 212, + 483, + 789, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the manual design of prompts can be a non-trivial and time-consuming task. [54] found that even a slight change in the prompt (e.g., one word) can make a big difference. They introduced Context Optimization (CoOp) to automate prompt engineering to generate continuous soft prompts instead of using hand-chosen hard prompts [54]. However, CoOp requires substantial computing resources, and the results of CoOp are not interpretable. Furthermore, such prompt tuning faces performance degradation when there is a significant domain shift [39], e.g., from natural images to pathology images, rendering it less adaptable to medical imaging tasks. On the other hand, medical image analysis may involve multiple learning tasks simultaneously. Conversely, in medical im", + "bbox": [ + 212, + 689, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Lai et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "age analysis, multiple learning tasks may be concurrently required. It becomes impractical if fine-tuning necessitates substantial computational resources. For instance, Lai et al. [17] quantify and visualize distinct pathologies in grey/white matter (GM/WM) regions within gigapixel images. To accomplish this, the task is divided into two learning objectives: pathology detection and GM/WM segmentation. They collected and annotated two datasets and then trained two models from scratch to achieve it. Therefore, it will be beneficial if we can propose an effective way to adapt a big model into multiple downstream tasks quickly so the final analysis objective can be reached in a reasonable time manner.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we aim to fine-tune CLIP efficiently with light computing resources for pathology image analysis tasks. There are several challenges in fine-tuning CLIP. First, overfitting is a severe issue if we adapt CLIP to the downstream tasks directly since CLIP is pre-trained on a 400M dataset while the new domain dataset can be small [10]. Second, CLIP has shown its promising zero-shot generalization ability, directly fine-tuning CLIP may lose this capability and face the \"catastrophic forgetting\" issue. In other words, it is unclear how to effectively learn the new knowledge while retaining the original pre-trained knowledge so that the generalization of CLIP can be maintained. Third, there are limited studies on benchmarking CLIP's transferability in the pathology domain; hence, its applicability remains unclear. To address the above issues, we first study the applicability of CLIP on two public pathology datasets and benchmark the zero-shot ability of CLIP on them. Then, we did empirical experiments to verify the importance of the CLIP's pre-trained knowledge. To address the \"catastrophic forgetting\" issue, we propose Residual Feature Refinement (RFR) as a lightweight approach for adapting CLIP to pathology images. As a shallow network, RFR demonstrates proficiency in extracting crucial spatial features and patterns from pathology images. Within RFR, we introduce a self-adaptive module designed to seamlessly fuse and balance the original knowledge from CLIP with the newly acquired insights from the specific pathological task, requiring only a minimal number of additional trainable weights, as opposed to retraining the entire encoders in CLIP. To further mitigate overfitting, we also incorporate Hidden Representation Perturbation (HRP) and Dual-view Vision Contrastive (DVC). Finally, we establish an end-to-end adaptation of CLIP, named Path-CLIP (as shown in Fig. 1(b)), tailored for pathology image analysis tasks. Path-CLIP incorporates a newly proposed loss function, Doublet Multimodal Contrastive Loss (DMCL).", + "bbox": [ + 212, + 297, + 787, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We summarize our contributions as follows:", + "bbox": [ + 214, + 719, + 532, + 734 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We explore the applicability of CLIP on pathology images and highlight the catastrophic forgetting issues when adapting CLIP to downstream pathology tasks.", + "- We propose Path-CLIP to alleviate the overfitting and forgetting issues and effectively adapt CLIP for pathology tasks." + ], + "bbox": [ + 225, + 750, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Path-CLIP: Efficiently Adapting CLIP for Pathology Image Analysis", + "bbox": [ + 272, + 114, + 732, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Extensive experiments demonstrate the efficiency and robustness of PathCLIP. For example, it achieves $+19\\%$ on PCam when only using $0.1\\%$ of labeled data.", + "bbox": [ + 223, + 146, + 789, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 215, + 387, + 231 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Language-vision model (LVM). LVMs have shown promising performance in learning generic visual representations [15, 22, 26, 48]. Recent models advance via text representation learning with large-scale Transformers [30] and even web-scale training datasets [54]. Transformer-based multimodal learning has achieved great success in such large-scale datasets [23, 29, 36]. For example, CLIP [26] was trained on 400 million image-caption pairs and achieved state-of-the-art results in many fields [26, 44, 46, 47]. In CLIP's architecture, it has two encoders: the vision encoder can be ResNet [12] or ViT [7] while the text encoder can be Transformer (e.g., BERT [6]). A recent work [27] fine-tuned CLIP for video data and performed competitively to more complicated methods designed for video processing. PointCLIP [50] applied CLIP to 3D recognition. CLIP is also applied for image generation tasks [8] and shows its power to reduce required data collection. However, the best way to adapt CLIP for downstream tasks is still being studied, especially when the new domain is far from the pre-trained domains, e.g., the medical domain.", + "bbox": [ + 212, + 247, + 787, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CLIP Fine-tuning and adaptation. Although recent multimodal models (e.g., CLIP [26]) perform astonishingly well in standard vision tasks without fine-tuning, the gap between CLIP and supervised learning is still noticeable. Therefore, many works focus on fine-tuning CLIP with a few samples from the new task provided. For example, CoOp [54] kept the entire pre-trained parameters fixed but optimized the continuous prompt through a learnable vector. This method closes the gap between CLIP and supervised learning but the performance gain is limited as it only focuses on the perspective of prompt design instead of the vision part for the image recognition tasks. On the other hand, CLIP-Adapter [9] and Tip-Adapter [51] freeze the pre-trained weights and train additive module, which significantly improves the CLIP's performance with limited training data in the new task. These methods require a lower computational cost. However, both CLIP-Adapter [9] and Tip-Adapter [51] involve tuning a ratio (hyper-parameter) to balance the original and fine-tuned knowledge manually, which may not be practical when we have multiple target learning tasks in the medical domain.", + "bbox": [ + 212, + 481, + 789, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 731, + 380, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce the proposed Path-CLIP. We summarize the end-to-end pipeline in Fig. 1(b): a self-adaptive residual ratio $\\alpha$ will be incorporated into Residual Feature Refinement (Section 3.2) to fuse and balance the original knowledge with the task-specific knowledge from the new downstream task. We introduce Hidden Representation Perturbation (HRP) and Dual-view Vision", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Lai et al.", + "bbox": [ + 271, + 114, + 352, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fa6cef420246b5d56be80fa190e37fe1d539dca58ddea36076b5ca2aaeb916b5.jpg", + "image_caption": [ + "Fig. 2: Overview of the Residual Feature Refinement (RFR): a downsampling-upsampling shallow architecture of fully connected layers. RFR blends the fine-tuned knowledge with the original knowledge from CLIP's vision encoder $(F(\\cdot))$ via a self-adaptive residual ratio $\\alpha$ . HRP refers to Hidden Representation Perturbation and TFT refers to pre-training fine-tuning. We use blue color to denote the original visual features and orange color to denote the fine-tuned features." + ], + "image_footnote": [], + "bbox": [ + 243, + 142, + 759, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Contrastive (Section 3.3) to alleviate the overfitting issue. Finally, we propose Doublet Multimodal Contrastive Loss (DMCL) in Section 3.4 as the loss function for fine-tuning.", + "bbox": [ + 212, + 470, + 787, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Language-Vision Pre-training", + "text_level": 1, + "bbox": [ + 214, + 556, + 504, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first revisit language-vision pre-training and use CLIP [26] as one example. Note that our framework is applicable to broader language-vision models. CLIP [26] has a vision encoder $F(\\cdot)$ and a text encoder $G(\\cdot)$ . The vision encoder maps a high-dimensional image into low-dimensional image embeddings. The vision encoder can be a CNN like ResNet [12] or a Vision Transformer (ViT) [7]. The text encoder is built on Transformer [30] and generates text embeddings from the prompt. Specifically, \"prompt\" here refers to a sequence of words (tokens), such as \"a neuropathological image of grey matter\".", + "bbox": [ + 212, + 597, + 787, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. CLIP jointly trains $F(\\cdot)$ and $G(\\cdot)$ to optimize the similarity score (e.g., symmetric cross-entropy loss [33]) between the visual and textual embeddings for each batch. The input consists of an image and its corresponding prompt (e.g., \"this is a photo of grey matter\"). They formulate the learning objective as a contrastive loss to align the visual and textual embedding spaces. Given a batch of image-prompt pairs, CLIP maximizes the similarity score for positive pairs while minimizing it for negative pairs.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Path-CLIP: Efficiently Adapting CLIP for Pathology Image Analysis", + "bbox": [ + 272, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Contrastive loss function. In CLIP training, the agreement between a positive pair of vision and text embeddings is maximized by InfoNCE loss [24]:", + "bbox": [ + 215, + 146, + 787, + 176 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nl _ {i} ^ {v \\rightarrow t} = - \\log \\frac {\\exp \\left(\\mathbf {s i m} \\left(v _ {i} , t _ {i}\\right) / \\gamma\\right)}{\\sum_ {k = 1} ^ {B} \\exp \\left(\\mathbf {s i m} \\left(v _ {i} , t _ {k}\\right) / \\gamma\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 186, + 787, + 222 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $v, t$ are normalized vectors from vision and text encoders. $(v_{i}, t_{i})$ is a positive pair. $\\mathbf{sim}$ is a function to measure the similarity of two vectors. $B$ is the mini-batch size of the image-text pairs and $\\gamma$ is the learnable temperature. Therefore, there is only 1 positive text for the input $i^{th}$ image but $N - 1$ negative texts. $l_{i}^{v \\rightarrow t}$ means the InfoNCE loss from an image $i$ to texts, so as the text loss $l_{i}^{t \\rightarrow v}$ . Then the loss function for CLIP is as follows:", + "bbox": [ + 214, + 229, + 787, + 321 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C L I P}} = \\frac {1}{2 B} \\sum_ {i = 1} ^ {B} \\left(l _ {i} ^ {v \\rightarrow t} + l _ {i} ^ {t \\rightarrow v}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 400, + 334, + 787, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference for classification tasks. For the inference process, an image $x$ is transformed into a feature manifold $f \\in \\mathbb{R}^D$ , where $D$ is the feature dimension. Then, $f$ is multiplied with a classifier weight matrix $\\mathbf{W} \\in \\mathbb{R}^{D \\times K}$ , where $K$ is the number of classes in the learning task. We get a $K$ -dimensional logit after matrix multiplication. Then we apply softmax to convert this logit into a probability vector $p \\in \\mathbb{R}^K$ over the $K$ classes. The whole process can be summarized as the following equation:", + "bbox": [ + 214, + 383, + 787, + 489 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np _ {i} = \\frac {\\exp \\left(\\mathbf {W} _ {\\mathbf {i}} ^ {T} \\times f\\right) / \\gamma}{\\sum_ {i = 1} ^ {K} \\exp \\left(\\mathbf {W} _ {\\mathbf {i}} ^ {T} \\times f\\right) / \\gamma}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 405, + 489, + 787, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\gamma$ is a temperature parameter learned by CLIP during training and $\\mathbf{W_i}$ is the prototype weight vector for class- $i$ .", + "bbox": [ + 214, + 530, + 787, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Transferability. Compared to traditional classifier learning methods where the supervision is a fixed one-hot vector, language-vision pre-training utilizes an open-set visual concept that can be explored through the text encoder to learn a broader semantic space, which in turn makes the learned representations more diverse and adaptable to various downstream tasks (e.g., zero-shot learning). Therefore, we argue that such language-vision pre-training is helpful for medical problems where the domain is very distinctive from the natural images.", + "bbox": [ + 214, + 566, + 787, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Residual Feature Refinement", + "text_level": 1, + "bbox": [ + 215, + 695, + 501, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this subsection, we introduce Residual Feature Refinement (RFR) to learn the task-relevant context when we adapt CLIP to the downstream learning tasks. Unlike CoOp's prompt tuning, which may not address the domain shift issue between the natural and pathology images, we focus on fine-tuning the visual features $f$ . However, simple fine-tuning of the entire network may fail in a new pathological task due to 1) the overfitting issue caused by a large number of parameters in CLIP and a shortage of training samples [13]; 2) catastrophic forgetting issue.", + "bbox": [ + 214, + 719, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Lai et al.", + "bbox": [ + 271, + 114, + 352, + 127 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1c9f581b851326ba35fabd2f734b3d811117394d617297dec3bf36d9dba63fda.jpg", + "table_caption": [ + "Table 1: Baseline. We fine-tune CLIP (ResNet-50) on PCam [31] with $1\\%$ training data. \"No fine-tuning\" is the best baseline due to the catastrophic forgetting issue." + ], + "table_footnote": [], + "table_body": "
# ConfigurationAccuracy
1No fine-tuning56.5%
2Fine-tune 10 epochs40.3%
3Fine-tune 100 epochs35.9%
4With a lower learning rate on #345.6%
", + "bbox": [ + 341, + 184, + 655, + 271 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Importance of the CLIP's pre-trained knowledge. CLIP has shown its zero-shot ability and transferability to a wide range of downstream tasks. We argue that preserving pre-trained knowledge of CLIP is crucial to maintain its generalization and alleviate the overfitting issue when the new dataset is small. There are two reasons behind this. First, one recent NLP work [11] shows large pre-trained language models (PLMs) are well-calibrated on the masked language modeling tasks. However, such property cannot be retained through direct fine-tuning due to catastrophic forgetting. Specifically, the pre-trained features are only used for the initialization and will be distorted by the fine-tuning [11]. Inspired by this, we investigate whether this issue also appears in language-vision models. As shown in Table 1, \"no fine-tuning\" but zero-shot gets the best results compared to fine-tuning CLIP on $1\\%$ of training data in PCam [31]. We also attempt to lower the learning rate to $1 \\times 10^{-6}$ , but the results are still far from satisfactory. Therefore, we conclude direct fine-tuning may distort the pre-trained knowledge and result in a severe catastrophic forgetting issue. Besides the forgetting issue, overfitting is another common challenge when adapting a big model to a new task. To alleviate the above two issues, we propose a residual refinement architecture to preserve the pre-trained features and acquire rich semantic information for pathology image analysis.", + "bbox": [ + 212, + 305, + 789, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Architecture overview. As shown in Fig. 2, given an image $x$ , we get the visual feature $f$ from the image encoder and compute the classifier weight $\\mathbf{W}$ from the text encoder. Then we design trainable adaptive layers $L$ to convert $f$ into $L(f)$ . $L$ can be multiple layers of linear transformations in a \"down-and-up\" architecture. As shown in Fig. 2, we have four layers to transform the feature dimension as \"1024-256-64-256-1024\". This is one example: the layers are flexible with other dimensions. After this architecture, $L(f)$ can be of the same size of $f$ and blended with $f$ as follows:", + "bbox": [ + 212, + 606, + 789, + 728 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} ^ {*} = \\left(1 - \\alpha_ {i}\\right) L \\left(f _ {i}\\right) + \\alpha_ {i} f _ {i}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 746, + 785, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $i = \\{1,2,3,\\dots n\\}$ and $\\alpha$ is the self-adaptive residual ratio to balance the fine-tuned knowledge and the original CLIP's knowledge. Specifically, the first term adaptively learns knowledge from the new dataset, while the second term preserves the original knowledge from CLIP.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Path-CLIP: Efficiently Adapting CLIP for Pathology Image Analysis", + "bbox": [ + 272, + 113, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Then we adopt Eq. (3) with the new $f^*$ to get the class probability vector and predict the category. During fine-tuning, the weights of the trainable layers are optimized through the symmetric contrastive loss used in CLIP [26]. The prompt can be \"this is a photo of [], where \"+\" is filled with the class name. We can incorporate pathological descriptions into the prompt to provide more domain information, e.g., \"this is a pathological breast photo of []\".", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hidden Representation Perturbation (HRP). Fine-tuning a pre-trained language-vision model on a small dataset can lead to potential issues such as overfitting or representation collapse [45]. To mitigate these issues, we propose Hidden Representation Perturbation (HRP) that is strategically integrated with the trainable layers. To balance the model's complexity and efficiency, we adopt a \"four-layer\" architecture with HRP, steering clear of deeper architectures to ensure a controlled number of trainable parameters. This precaution is crucial as an excess of trainable parameters may exacerbate overfitting, especially in scenarios characterized by limited resources, where only a sparse dataset is available for the new task [52].", + "bbox": [ + 212, + 244, + 787, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To alleviate the overfitting issue in low-resource settings, we focus on adding noise during the training process, a widely-used NLP strategy for large language models to smoothen the optimization landscape. For example, NoisyTune [35] adds the perturbations to the pre-trained weights to improve the model's generalization. Inspired by these NLP strategies [19, 35, 45], we propose to add random noise to the feature $f$ before trainable adapting layers fine-tuning it. The random noise can be Gaussian noise $\\varepsilon$ generated from $\\mathcal{N}(0, \\sigma^2)$ . After this, the perturbed feature $f$ is converted into $\\hat{f} = f + \\varepsilon$ . Therefore, Eq. (4) is converted into:", + "bbox": [ + 212, + 404, + 789, + 527 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} ^ {*} = \\left(1 - \\alpha_ {i}\\right) L \\left(\\hat {f} _ {i}\\right) + \\alpha_ {i} f _ {i}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 541, + 787, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Therefore, the residual feature will be perturbed with the Gaussian noise for the refinement, designed to alleviate the overfitting issue.", + "bbox": [ + 212, + 569, + 787, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Self-adaptive residual ratio", + "text_level": 1, + "bbox": [ + 214, + 625, + 485, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Section 3.2, we introduce residual feature refinement and the residual ratio $\\alpha$ . $\\alpha$ helps modify the degree of preserving the original knowledge for better performance. Empirically, $\\alpha$ should be small if the domain gap between the source data and the downstream task is enormous, so more knowledge from the downstream task $(L(f))$ can be considered. However, if $\\alpha$ is too small, the model can be overfitting as it will heavily rely on the fine-tuned features and disregard the pre-trained knowledge from CLIP. On the other hand, the model can also be underfitting if $\\alpha$ is too large, as it may prevent learning from the fine-tuned features. Therefore, it is crucial to quantify the residual ratio $\\alpha$ so the model can be well-trained instead of underfitting or overfitting.", + "bbox": [ + 212, + 650, + 787, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Recent works such as Tip-Adapter [51] and CLIP-Adapter [9] involve feature refinement while they set such ratio as a fixed hyper-parameter that needs to be", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Lai et al.", + "bbox": [ + 271, + 114, + 352, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tuned during the training process. We argue the residual ratio should be dynamic and self-adaptive during training instead of fixed, as the learning status changes along the training process. Hence, we propose a self-adaptive residual ratio in Path-CLIP.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "DVC: Dual-view Vision Contrastive. Inspired by self-supervised learning (e.g., SimCLR [4]), the composition of data augmentations is relevant to define effective predictive tasks. Therefore, we propose Dual-view Vision Contrastive (DVC) to measure the model's learning status. Given an image $x$ , we leverage two types of augmentations: \"weak\" and \"strong\" [28]. \"weak\" refers to the standard flip-and-shift augmentation strategy while \"strong\" refers to perturb the appearance (RandAugment [5] and CTAugment [2]). We propose to measure the distance between representations from both augmentations, e.g., cosine similarity has shown good performance in both visual contrastive [4] and textual contrastive [3] tasks.", + "bbox": [ + 212, + 214, + 784, + 364 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "With the input $x$ , we will get $w(x)$ from the weak augmentation and $s(x)$ from the strong augmentation. We obtain $F(w(x))$ and $F(s(x))$ after the vision encoder $F$ . We define the distance as $\\tau$ :", + "bbox": [ + 212, + 373, + 785, + 417 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\tau = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\cos \\left(F \\left(w \\left(x _ {i}\\right)\\right), F \\left(s \\left(x _ {i}\\right)\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 422, + 785, + 457 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $N$ is the number of samples in the training set, and $\\cos$ is a cosine similarity. The model's learning status is relevant to its consistency across different perturbations. We traverse the entire training set and calculate its mean value to represent the model's performance.", + "bbox": [ + 212, + 462, + 782, + 523 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Bound constraints. In addition to DVC which measures the predictive consistency of the model, we also select \"accuracy\" on the training set as a second metric to measure the learning status, denoted as $T$ . Empirically, \"accuracy\" alone cannot fully tell if the model is well-fitted: 100% of accuracy may indicate overfitting, which is the biggest challenge when we adapt a large model to a downstream task where the training set is very small. However, \"accuracy\" is a good metric to show if the model is underfitting. Therefore, we propose to fuse DVC and \"accuracy\" to represent the model's status during training.", + "bbox": [ + 212, + 531, + 787, + 652 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "If we look at $\\alpha$ and set it fixed, we can find the model is prone to overfit if $\\alpha$ is too small as it will heavily rely on the fine-tuned knowledge; the model tends to underfit if $\\alpha$ is too large as it prevents the model from learning new knowledge (directly use the original pre-trained knowledge). Therefore, we propose a dynamic way with lower and upper bounds to obtain $\\alpha$ during training:", + "bbox": [ + 212, + 659, + 782, + 736 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha = \\min (\\max (\\tau , T), 1 - \\tau). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 742, + 785, + 758 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The bounds are associated with predictive consistency (DVC) and predictive accuracy on the training set. For example, if the model is overfitting, $T$ will be large, and $\\tau$ will be relatively small. Then $\\alpha$ will be smaller in the next epoch to use less fine-tuned features. Therefore, our proposed bound constraints can effectively adjust $\\alpha$ to alleviate the overfitting issue.", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Path-CLIP: Efficiently Adapting CLIP for Pathology Image Analysis", + "bbox": [ + 272, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c2e7f96d3618a13c1d9605bfa1ac5f219d04fc819c37596898e29877a1210be4.jpg", + "image_caption": [ + "Fig. 3: Doublet Multimodal Contrastive Loss (DMCL) that utilizes the traditional image-text pair and the proposed image-image pair with weak/strong augmentation criteria." + ], + "image_footnote": [], + "bbox": [ + 272, + 143, + 733, + 262 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3.4 Path-CLIP: End-to-end adaptation of CLIP to pathology tasks", + "text_level": 1, + "bbox": [ + 214, + 347, + 779, + 364 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To efficiently adapt CLIP to pathology image tasks, we propose RFR with a self-adaptive residual ratio to balance the original and fine-tuned knowledge based on the learning status. To alleviate the overfitting issue, we incorporate HRP and design bound constraints for the residual ratio. Besides these, to distill more visual representations from the new image domain, we add a contrastive loss to the loss function so that the trainable portion has a better classifying ability. We follow SimCLR [4] to define the Contrastive Loss (CL) for one image in the batch as:", + "bbox": [ + 212, + 378, + 787, + 500 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nl _ {i} ^ {c} = \\left\\{ \\begin{array}{l} - \\log \\frac {\\exp (\\mathbf {s i m} (v _ {i} , v _ {i + B}) / \\gamma)}{\\sum_ {k \\neq i} ^ {2 B} \\exp (\\mathbf {s i m} (v _ {i} , v _ {k}) / \\gamma)}, \\text {i f} i < B, \\\\ - \\log \\frac {\\exp (\\mathbf {s i m} (v _ {i} , v _ {B - i}) / \\gamma)}{\\sum_ {k \\neq i} ^ {2 B} \\exp (\\mathbf {s i m} (v _ {i} , v _ {k}) / \\gamma)}, \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 343, + 511, + 787, + 575 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Differing from SimCLR [4] and RepRec [38], as shown in Fig. 3, here we set image $i$ as the weakly augmented image and image $i + B$ as the strongly augmented image instead of two randomly augmented images as the positive pair. We named this loss as $\\mathcal{L}_{\\mathrm{SWAPP}}$ . SWAPP here refers to Strong-Weak Augmented Positive Pair. Finally, the total loss for fine-tuning CLIP on our pathology datasets is as follows:", + "bbox": [ + 212, + 595, + 787, + 686 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {t o t a l}} = \\mathcal {L} _ {\\mathrm {C L I P}} + \\mathcal {L} _ {\\mathrm {S W A P P}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 689, + 578, + 705 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C L I P}} \\rightarrow \\frac {1}{2 B} \\sum_ {i = 1} ^ {B} \\left(l _ {i} ^ {v \\rightarrow t} + l _ {i} ^ {t \\rightarrow v}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 708, + 785, + 744 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {S W A P P}} \\rightarrow \\frac {1}{2 B} \\sum_ {i = 1} ^ {2 B} l _ {i} ^ {c}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 747, + 542, + 784 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We define this loss as Doublet Multimodal Contrastive Loss (DMCL) because the items involve both intra-contrastive (image and text) and inter-contrastive (image and image). The algorithm can be found in the Appendix.", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Lai et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/29ddc90df9772bd44a77b694abda18227524cb8922c69ea8e76bd34b83748077.jpg", + "table_caption": [ + "Table 2: Quantitative comparison on PCam. Precision, Recall, and F1-score refer to the macro-averaged values from all classes" + ], + "table_footnote": [], + "table_body": "
Learning MannerLabeled RatioAlgorithmAccuracyPrecisionRecallF1-scoreAUC
Supervised100%-92.8 ± 0.3092.9 ± 0.2092.6 ± 0.2192.8 ± 0.200.95 ± 0.01
Supervised0.5%-45.1 ± 2.2742.0 ± 2.1433.6 ± 3.3037.3 ± 2.860.51 ± 0.01
Self-supervised0.5%SimCLR [4]60.4 ± 1.4563.3 ± 1.9058.5 ± 2.0160.8 ± 1.200.61 ± 0.04
Semi-supervised0.5%Pseudo-Label [20]55.7 ± 2.0559.2 ± 2.2354.5 ± 1.9756.8 ± 1.550.58 ± 0.03
FixMatch [28]73.2 ± 0.7677.5 ± 0.5073.7 ± 0.2575.6 ± 0.370.84 ± 0.04
Dash [37]71.0 ± 0.9875.3 ± 0.7870.2 ± 0.4572.7 ± 0.500.82 ± 0.02
FlexMatch [49]74.0 ± 0.8078.1 ± 1.1573.0 ± 0.9075.5 ± 0.840.85 ± 0.03
Multimodal CLIP0.5%Zero-shot [26]56.5 ± 0.0057.4 ± 0.0050.3 ± 0.0053.7 ± 0.000.60 ± 0.00
CoOp [54]63.6 ± 0.2563.9 ± 0.4262.5 ± 0.3563.0 ± 0.290.67 ± 0.02
CLIP-Adapter [9]72.3 ± 1.0277.2 ± 0.9563.2 ± 0.5669.4 ± 0.840.81 ± 0.02
TIP-Adapter [51]73.9 ± 0.5070.5 ± 0.8578.0 ± 0.4974.0 ± 0.950.84 ± 0.02
Path-CLIP81.9 ± 0.7879.4 ± 0.3585.0 ± 0.8082.1 ± 0.950.89 ± 0.02
Supervised1.5%-62.8 ± 1.3169.5 ± 0.6754.1 ± 2.2760.8 ± 1.680.63 ± 0.02
Self-supervised1.5%SimCLR [4]64.9 ± 1.0265.2 ± 1.4560.1 ± 1.7062.5 ± 1.050.64 ± 0.03
Semi-supervised1.5%Pseudo-Label [20]67.9 ± 1.8570.5 ± 1.9062.2 ± 1.7666.1 ± 1.420.67 ± 0.04
FixMatch [28]83.9 ± 1.2985.1 ± 0.8584.1 ± 0.9584.6 ± 0.900.87 ± 0.01
Dash [37]82.8 ± 1.0983.0 ± 0.5581.2 ± 0.6082.1 ± 0.780.85 ± 0.03
FlexMatch [49]84.1 ± 1.0284.9 ± 0.7584.4 ± 0.8084.6 ± 0.750.86 ± 0.04
Multimodal CLIP1.5%CoOp [54]61.5 ± 0.7560.5 ± 1.0365.6 ± 0.9063.1 ± 0.850.65 ± 0.02
CLIP-Adapter [9]74.1 ± 0.8879.1 ± 0.3465.4 ± 0.8971.6 ± 1.020.82 ± 0.01
TIP-Adapter [51]75.2 ± 0.9472.3 ± 0.6979.4 ± 0.8675.7 ± 0.730.83 ± 0.02
Path-CLIP83.1 ± 0.6782.8 ± 0.9582.0 ± 0.7582.4 ± 0.460.91 ± 0.02
", + "bbox": [ + 230, + 184, + 769, + 395 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 420, + 375, + 436 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Datasets. We select two classification datasets (PCam [31] and MHIST [34]) and a gigapixel grey/white matter segmentation dataset to evaluate Path-CLIP. The details of these datasets can be found in Appendix.", + "bbox": [ + 212, + 449, + 787, + 494 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Experimental setup. For a fair comparison, we use ResNet-50 [12] as the backbone in the vision encoder $f(\\cdot)$ and BERT [6] as the textual encoder. We follow CLIP [26] to use gradient scaling to facilitate mixed-precision training. We set the learning rate as 0.0001. The batch size is 32. We use Adam [16] as the optimizer. To compare the computational complexity, we conduct all experiments on a single GPU (Nvidia RTX 2080Ti).", + "bbox": [ + 212, + 502, + 787, + 594 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Baselines. We first implement traditional supervised learning (SL) methods with the entire dataset to explore the upper-bound performance for each task. Then we reduce the data usage and investigate the performance change of these SL methods. Then we compare Path-CLIP with original CLIP [26] and three recent CLIP-based fine-tuning methods: CoOp [54], Tip-Adapter [51] and CLIP-Adapter [9]. We compare them under the limited data scenarios to test their ability to do downstream tasks when the new dataset is small. Besides that, we also incorporate recent self-supervised learning (SimCLR [4]) and semi-supervised learning (Pseudo-Label [20], FixMatch [28], Dash [37], and FlexMatch [49]). Semi-supervised training requires a large amount of unlabeled data, which involves data collection efforts [41-43].", + "bbox": [ + 212, + 599, + 787, + 768 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Main results", + "text_level": 1, + "bbox": [ + 215, + 787, + 366, + 801 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Classification task: PCam [31] and MHIST [34]. The main results for two classification benchmarks are summarized in Table 2 and Table 3. First, CLIP", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Path-CLIP: Efficiently Adapting CLIP for Pathology Image Analysis", + "bbox": [ + 272, + 114, + 730, + 130 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/c8206ae796a872e1f93e3452dd12f1aa8ebe79d70210508515d1c9395f541b02.jpg", + "table_caption": [ + "Table 3: Quantitative comparison on MHIST. Precision, Recall, and F1-score refer to the macro-averaged values from all classes" + ], + "table_footnote": [], + "table_body": "
Learning MannerLabeled RatioAlgorithmAccuracyPrecisionRecallF1-scoreAUC
Supervised100%-86.0 ± 0.3684.8 ± 0.6183.5 ± 0.5584.1 ± 0.580.89 ± 0.01
Self-supervised10%SimCLR [4]63.2 ± 0.9563.8 ± 1.2062.2 ± 1.7063.0 ± 0.850.66 ± 0.04
Semi-supervised10%Pseudo-Label [20]65.4 ± 1.0466.5 ± 0.9865.9 ± 0.8566.2 ± 1.420.75 ± 0.03
FixMatch [28]70.5 ± 1.5568.8 ± 2.7669.5 ± 2.9769.1 ± 2.860.79 ± 0.02
Dash [37]70.2 ± 1.2170.0 ± 0.8571.3 ± 0.6970.6 ± 1.480.81 ± 0.04
FlexMatch [49]73.1 ± 0.8474.6 ± 1.0575.1 ± 0.9074.8 ± 0.890.82 ± 0.03
Multimodal CLIP10%Zero-shot [26]36.9 ± 0.0036.9 ± 0.00100 ± 0.0053.9 ± 0.000.50 ± 0.00
CoOp [54]68.9 ± 0.4056.0 ± 0.6572.8 ± 0.5363.3 ± 0.860.76 ± 0.01
CLIP-Adapter [9]65.9 ± 1.4552.2 ± 1.2387.2 ± 0.8565.3 ± 0.950.77 ± 0.02
TIP-Adapter [51]63.1 ± 0.7849.9 ± 0.9094.7 ± 1.0265.4 ± 0.560.78 ± 0.03
Path-CLIP74.8 ± 1.9563.3 ± 1.4075.6 ± 0.9868.9 ± 1.500.84 ± 0.02
", + "bbox": [ + 230, + 184, + 767, + 300 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "shows its promising zero-shot ability: when we directly test CLIP on PCam [31], it can get $56.5\\%$ accuracy. Then, we compare Path-CLIP with other CLIP-based fine-tuning methods. In PCam [31], we present $0.5\\%$ and $1.5\\%$ data usage in the training set to fine-tune CLIP, respectively. We find all of these fine-tuning algorithms [9, 51, 54] can improve the performance of the original CLIP with the new data from the downstream task presented. Path-CLIP can outperform them in both settings. Specifically, Path-CLIP outperforms TIP-Adapter [51] with $8\\%$ and $7.9\\%$ accuracy when only $0.5\\%$ and $1.5\\%$ data are used, respectively. Path-CLIP also shows similar improvement in MHIST [34] (Table 3). We also implemented several semi-supervised learning where the rest of the data in the training set are used as an unlabeled set. Although Path-CLIP does not use unlabeled data, it can achieve competitive results with semi-supervised learning.", + "bbox": [ + 212, + 335, + 787, + 517 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Besides these ratios of data usage, we also visualize the results by varying the ratio to show the performance of each fine-tuning algorithm. As shown in Fig. 4, CoOp [54] may not improve PCam when more data are used to fine-tune the prompt. CoOp freezes the encoders and only fine-tunes the language prompt part, while other algorithms also fine-tune the vision part. We conclude that visual fine-tuning is more critical compared to prompt tuning.", + "bbox": [ + 212, + 529, + 787, + 621 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/859cb8c799e940407e8f279e525652182d3daf3db9a0cbd9f05860f9ee502eaf.jpg", + "image_caption": [ + "(a) PCam" + ], + "image_footnote": [], + "bbox": [ + 241, + 654, + 493, + 789 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/47c058cc12f1284798227d7cbeb7ace6e0a56b57a4298d346f97b41d9c3cbc07.jpg", + "image_caption": [ + "(b) MHIST", + "Fig. 4: Comparison of different fine-tuning methods by varying data usage." + ], + "image_footnote": [], + "bbox": [ + 501, + 654, + 751, + 789 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Lai et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/10ee9e23232ed5b0e484192175653909a2e8c14e165e3d07c7808797f2a58dac.jpg", + "table_caption": [ + "Table 4: Performance comparison on GM/WM segmentation." + ], + "table_footnote": [], + "table_body": "
Data UsageAlgorithmGM_IoUGM_DICEWM_IoUWM_DICE
100% (SL)FCN [1]71.7283.2149.0966.08
U-Net [25]91.5295.1782.0288.15
0.1% (SL)FCN [1]48.2525.0124.5430.02
U-Net [25]46.1350.105.7512.23
0.1% (CLIP)CoOp [54]62.2365.7039.9248.65
CLIP-Adapter [9]67.8074.3243.2554.61
TIP-Adapter [51]69.9275.5046.2356.60
Path-CLIP78.0583.7054.5063.23
1% (SL)FCN [1]51.3430.2031.0235.12
U-Net [25]50.2453.2315.4325.58
1% (CLIP)CoOp [54]63.4564.1237.8545.34
CLIP-Adapter [9]69.9976.5447.9558.34
TIP-Adapter [51]72.1277.3055.6563.60
Path-CLIP80.2188.7060.5070.34
", + "bbox": [ + 272, + 171, + 727, + 380 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "GM/WM segmentation [17]. Besides two classification tasks, we follow [17] to generalize our method to a weakly-supervised segmentation problem. We select IoU and DICE scores as the performance metrics. The results are summarized in Table 4. We select FCN [1] and U-Net [25] as two SL baselines: we first train them on all data (100%) in a supervised manner, then we reduce the dataset to 0.1% and 1% setting and train the model again. We can see both SL algorithms face performance degradation when the training data is scarce. They require an extensive and well-curated labeled dataset to get satisfactory results. We then focus on fine-tuning CLIP algorithms. Table 4 shows Path-CLIP can achieve superior results than other fine-tuning techniques.", + "bbox": [ + 212, + 410, + 787, + 564 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.2 Ablation studies", + "text_level": 1, + "bbox": [ + 215, + 589, + 398, + 604 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Importance of alleviating overfitting. Since the new task may have a very limited training set, it is crucial to alleviate the potential overfitting issue. Besides the RFR with the self-adaptive ratio, we also propose HRP, DVC, and DMCL to cope with the overfitting issue. Table 5 summarizes the contribution of each proposed module: HRP can improve with $+2.2\\%$ accuracy and DVC can improve with $+4.3\\%$ . DMCL inspired by self-supervised learning can also aid in great improvements. We conclude that contrastive learning, perturbations on data and hidden representations are helpful to the overfitting issue.", + "bbox": [ + 212, + 618, + 787, + 741 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Computational efficiency. When we range the data usage from $0.1\\%$ to $1\\%$ , Path-CLIP can obtain $+5.5\\%$ gain in accuracy while only introducing 1 minute of additional training time. However, CoOp [54] has the overfitting issue and obtains a lower score while using 53 minutes, almost five times compared to the time used by RFR. The situation remains similar under the $10\\%$ of data usage. We conclude our proposed RFR can get over $19\\%$ improvement on CLIP by only", + "bbox": [ + 212, + 748, + 789, + 842 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Path-CLIP: Efficiently Adapting CLIP for Pathology Image Analysis", + "bbox": [ + 272, + 113, + 730, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 127 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/6b7f0952896b94d05b3621a3b719f31c6efbff3424de5b3cec7c1099640f1367.jpg", + "table_caption": [ + "Table 5: Ablation study on PCam using $0.5\\%$ data for fine-tuning." + ], + "table_footnote": [], + "table_body": "
#RFCHRPDVCDMCLAccuracyF1-scoreAUC
1XXXX56.553.70.60
2✓XXX72.070.50.69
3✓✓XX74.275.50.78
4✓✓✓X78.579.30.84
5✓✓✓✓81.982.10.89
", + "bbox": [ + 357, + 171, + 643, + 250 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "using 10 minutes for the fine-tuning on a single GPU, which is promising for digital pathology research. The details can be found in Table 8 in the Appendix.", + "bbox": [ + 212, + 276, + 785, + 308 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Generalization to pathology foundation model: PLIP [14]. The results of using PLIP's weights are in Appendix 3.6: PLIP is pre-trained on a large-scale pathology dataset and achieves strong zero-shot performance. With our method, we can further boost the performance and outperform other tuning methods.", + "bbox": [ + 212, + 313, + 785, + 375 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Discussion", + "text_level": 1, + "bbox": [ + 214, + 395, + 354, + 411 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we explore the generalization of CLIP in pathology image analysis. To alleviate the overfitting and forgetting issues when adapting CLIP to small downstream datasets, we propose self-adaptive RFR with HRP. We propose Path-CLIP with DMCL as an end-to-end pipeline for CLIP adaptation in pathology tasks. The results reveal the superiority and effectiveness of Path-CLIP compared to the existing fine-tuning methods. In addition, the less computational time improves the efficiency and accessibility of Path-CLIP. Hence, we conclude Path-CLIP is more robust and has the potential to bridge the domain shift between the pre-trained natural images and pathology images in an efficient and cost-effective manner.", + "bbox": [ + 212, + 424, + 787, + 575 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 215, + 595, + 393, + 614 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was supported by the Noyce Initiative UC Partnerships in Computational Transformation Grant and Child Family Endowed Professorship. Resources for this study were funded in part by grants from the National Institute on Aging of the National Institutes of Health under Award Numbers R01AG062517, P30AG072972, and R01AG056519.", + "bbox": [ + 212, + 625, + 785, + 702 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 722, + 323, + 738 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bandy, P., van de Loo, R., Intezar, M., Geijs, D., Ciompi, F., van Ginneken, B., van der Laak, J., Litjens, G.: Comparison of different methods for tissue segmentation in histopathological whole-slide images. In: Proc. 2017 IEEE Int. Symp. Biomed. Imaging. pp. 591-595 (2017)", + "2. Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., Raffel, C.A.: MixMatch: A holistic approach to semi-supervised learning. In: NeurIPS. vol. 32 (2019)" + ], + "bbox": [ + 222, + 742, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Lai et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "3. Chen, F., Zhang, H., Li, Z., Dou, J., Mo, S., Chen, H., Zhang, Y., Ahmed, U., Zhu, C., Savvides, M.: Unitail: Detecting, reading, and matching in retail scene. In: ECCV. pp. 705-722. Springer (2022)", + "4. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML. pp. 1597-1607 (2020)", + "5. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Workshop. pp. 702-703 (2020)", + "6. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)", + "7. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)", + "8. Gal, R., Patashnik, O., Maron, H., Bermano, A.H., Chechik, G., Cohen-Or, D.: Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG) 41(4), 1-13 (2022)", + "9. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544 (2021)", + "0. Ge, C., Huang, R., Xie, M., Lai, Z., Song, S., Li, S., Huang, G.: Domain adaptation via prompt learning. arXiv preprint arXiv:2202.06687 (2022)", + "1. He, G., Chen, J., Zhu, J.: Preserving pre-trained features helps calibrate fine-tuned language models. In: ICLR (2023)", + "2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)", + "3. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: ICML. pp. 2790-2799. PMLR (2019)", + "4. Huang, Z., other: A visual-language foundation model for pathology image analysis using medical twitter. Nature medicine (2023)", + "5. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML. pp. 4904-4916 (2021)", + "6. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (Poster) (2015)", + "7. Lai, Z., Oliveira, L.C., Guo, R., Xu, W., Hu, Z., Mifflin, K., Decarli, C., Cheung, S.C., Chuah, C.N., Dugger, B.N.: Brainsec: Automated brain tissue segmentation pipeline for scalable neuropathological analysis. IEEE Access 10, 49064-49079 (2022)", + "8. Lai, Z., Wang, C., Oliveira, L.C., Dugger, B.N., Cheung, S.C., Chuah, C.N.: Joint semi-supervised and active learning for segmentation of gigapixel pathology images with cost-effective labeling. In: ICCV Workshop. pp. 591-600 (2021)", + "9. Lam, S.Y., Zeng, Q., Zhang, K., You, C., Voigt, R.: Large language models are partially primed in pronoun interpretation. arXiv preprint arXiv:2305.16917 (2023)", + "20. Lee, D.H.: Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In: Workshop on challenges in representation learning (ICML) (2013)", + "21. Liu, M., Hu, L., Tang, Y., Wang, C., He, Y., Zeng, C., Lin, K., He, Z., Huo, W.: A deep learning method for breast cancer classification in the pathology images. IEEE J. Biomed. Health. Inf. 26(10), 5025-5032 (2022)" + ], + "bbox": [ + 222, + 146, + 787, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Path-CLIP: Efficiently Adapting CLIP for Pathology Image Analysis", + "bbox": [ + 272, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "22. Lu, Y., Liu, J., Zhang, Y., Liu, Y., Tian, X.: Prompt distribution learning. In: CVPR. pp. 5206-5215 (2022)", + "23. Lyu, W., Dong, X., Wong, R., Zheng, S., Abell-Hart, K., Wang, F., Chen, C.: A multimodal transformer: Fusing clinical notes with structured ehr data for interpretable in-hospital mortality prediction. In: AMIA Annual Symposium Proceedings. vol. 2022, p. 719. American Medical Informatics Association (2022)", + "24. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)", + "25. Oskal, K.R., Risdal, M., Janssen, E.A., Undersrud, E.S., Gulsrud, T.O.: A U-net based approach to epidermal tissue segmentation in whole slide histopathological images. SN Appl. Sci. 1(7), 1-12 (2019)", + "26. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021)", + "27. Rasheed, H., khattak, M.U., Maaz, M., Khan, S., Khan, F.S.: Finetuned clip models are efficient video learners. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)", + "28. Sohn, K., Berthelot, D., Carlini, N., Zhang, Z., Zhang, H., Raffel, C.A., Cubuk, E.D., Kurakin, A., Li, C.L.: FixMatch: Simplifying semi-supervised learning with consistency and confidence. In: NeurIPS. vol. 33, pp. 596-608 (2020)", + "29. Tian, Y., Newsam, S., Boakye, K.: Fashion image retrieval with text feedback by additive attention compositional learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1011-1021 (2023)", + "30. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017)", + "31. Veeling, B.S., Linmans, J., Winkens, J., Cohen, T., Welling, M.: Rotation equivariant cnns for digital pathology. In: MICCAI. pp. 210-218. Springer (2018)", + "32. Wang, X., Yang, S., Zhang, J., Wang, M., Zhang, J., Huang, J., Yang, W., Han, X.: Transpath: Transformer-based self-supervised learning for histopathological image classification. In: MICCAI. pp. 186-195. Springer (2021)", + "33. Wang, Y., Ma, X., Chen, Z., Luo, Y., Yi, J., Bailey, J.: Symmetric cross entropy for robust learning with noisy labels. In: ICCV (2019)", + "34. Wei, J., Suriawinata, A., Ren, B., Liu, X., Lisovsky, M., Vaickus, L., Brown, C., Baker, M., Tomita, N., Torresani, L., et al.: A petri dish for histopathology image analysis. In: Int. Conf. Arti. Intell. in Medi. pp. 11-24. Springer (2021)", + "35. Wu, C., Wu, F., Qi, T., Huang, Y.: Noisytune: A little noise can help you finetune pretrained language models better. In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). pp. 680-685 (2022)", + "36. Xu, P., Zhu, X., Clifton, D.A.: Multimodal learning with transformers: A survey. arXiv preprint arXiv:2206.06488 (2022)", + "37. Xu, Y., Shang, L., Ye, J., Qian, Q., Li, Y.F., Sun, B., Li, H., Jin, R.: Dash: Semi-supervised learning with dynamic thresholding. In: International Conference on Machine Learning. pp. 11525-11536. PMLR (2021)", + "38. Yan, X., Naushad, J., Sun, S., Han, K., Tang, H., Kong, D., Ma, H., You, C., Xie, X.: Representation recovering for self-supervised pre-training on medical images. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2685-2695 (2023)", + "39. Yang, B., Liu, F., Li, Z., Yin, Q., You, C., Yin, B., Zou, Y.: Multimodal prompt learning for product title generation with extremely limited labels. arXiv preprint arXiv:2307.01969 (2023)" + ], + "bbox": [ + 215, + 146, + 787, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Lai et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "40. Yang, J., Chen, H., Liang, Y., Huang, J., He, L., Yao, J.: Concl: Concept contrastive learning for dense prediction pre-training in pathology images. In: ECCV. pp. 523-539. Springer (2022)", + "41. You, C., Dai, W., Min, Y., Liu, F., Zhang, X., Feng, C., Clifton, D.A., Zhou, S.K., Staib, L.H., Duncan, J.S.: Rethinking semi-supervised medical image segmentation: A variance-reduction perspective. In: NeurIPS (2023)", + "42. You, C., Dai, W., Min, Y., Staib, L., Duncan, J.S.: Bootstrapping semi-supervised medical image segmentation with anatomical-aware contrastive distillation. In: International Conference on Information Processing in Medical Imaging. pp. 641-653. Springer (2023)", + "43. You, C., Dai, W., Min, Y., Staib, L., Sekhon, J., Duncan, J.S.: Action++: Improving semi-supervised medical image segmentation with adaptive anatomical contrast. arXiv preprint arXiv:2304.02689 (2023)", + "44. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917 (2022)", + "45. Yuan, H., Yuan, Z., Tan, C., Huang, F., Huang, S.: Hype: Better pre-trained language model fine-tuning with hidden representation perturbation. arXiv preprint arXiv:2212.08853 (2022)", + "46. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021)", + "47. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12104-12113 (June 2022)", + "48. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: Lit: Zero-shot transfer with locked-image text tuning. In: CVPR. pp. 18123-18133 (2022)", + "49. Zhang, B., Wang, Y., Hou, W., Wu, H., Wang, J., Okumura, M., Shinozaki, T.: Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. NeurIPS 34 (2021)", + "50. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8552-8562 (2022)", + "51. Zhang, R., Zhang, W., Fang, R., Gao, P., Li, K., Dai, J., Qiao, Y., Li, H.: Tip-adapter: Training-free adaption of clip for few-shot classification. In: ECCV. pp. 493-510. Springer (2022)", + "52. Zhang, T., Wu, F., Katiyar, A., Weinberger, K.Q., Artzi, Y.: Revisiting few-sample bert fine-tuning. In: ICLR (2021)", + "53. Zhang, W., Zhu, L., Hallinan, J., Zhang, S., Makmur, A., Cai, Q., Ooi, B.C.: Boostmis: Boosting medical image semi-supervised learning with adaptive pseudo labeling and informative active annotation. In: CVPR. pp. 20666-20676 (2022)", + "54. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. Int. J. Comput. Vision 130(9), 2337-2348 (2022)" + ], + "bbox": [ + 215, + 146, + 787, + 756 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Path-CLIP: Efficiently Adapting CLIP for Pathology Image Analysis", + "bbox": [ + 272, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Bridging the Pathology Domain Gap_ Efficiently Adapting CLIP for Pathology Image Analysis with Limited Labeled Data/3b6fb56d-1649-4e06-bdb2-706d8b03ddb5_model.json b/2024/Bridging the Pathology Domain Gap_ Efficiently Adapting CLIP for Pathology Image Analysis with Limited Labeled Data/3b6fb56d-1649-4e06-bdb2-706d8b03ddb5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..65413ffaa242c65be2855923bc80f8d46574ece4 --- /dev/null +++ b/2024/Bridging the Pathology Domain Gap_ Efficiently Adapting CLIP for Pathology Image Analysis with Limited Labeled Data/3b6fb56d-1649-4e06-bdb2-706d8b03ddb5_model.json @@ -0,0 +1,2203 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.141, + 0.787, + 0.207 + ], + "angle": 0, + "content": "Bridging the Pathology Domain Gap: Efficiently Adapting CLIP for Pathology Image Analysis with Limited Labeled Data" + }, + { + "type": "text", + "bbox": [ + 0.33, + 0.242, + 0.673, + 0.274 + ], + "angle": 0, + "content": "Zhengfeng Lai1, Joohi Chauhan2, Brittany N. Dugger1, and Chen-Nee Chuah1" + }, + { + "type": "text", + "bbox": [ + 0.301, + 0.284, + 0.702, + 0.341 + ], + "angle": 0, + "content": "1 University of California, Davis \n2 Motilal Nehru National Institute of Technology Allahabad {lzhengfeng, bndugger, chuah}@ucdavis.edu joohi@mnnit.ac.in" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.361, + 0.744, + 0.639 + ], + "angle": 0, + "content": "Abstract. Contrastive Language-Image Pre-training (CLIP) has shown its proficiency in acquiring distinctive visual representations and exhibiting strong generalization across diverse vision tasks. However, its effectiveness in pathology image analysis, particularly with limited labeled data, remains an ongoing area of investigation due to challenges associated with significant domain shifts and catastrophic forgetting. Thus, it is imperative to devise efficient adaptation strategies in this domain to enable scalable analysis. In this study, we introduce Path-CLIP, a framework tailored for a swift adaptation of CLIP to various pathology tasks. Firstly, we propose Residual Feature Refinement (RFR) with a dynamically adjustable ratio to effectively integrate and balance source and task-specific knowledge. Secondly, we introduce Hidden Representation Perturbation (HRP) and Dual-view Vision Contrastive (DVC) techniques to mitigate overfitting issues. Finally, we present the Doublet Multimodal Contrastive Loss (DMCL) for fine-tuning CLIP for pathology tasks. We demonstrate that Path-CLIP adeptly adapts pre-trained CLIP to downstream pathology tasks, yielding competitive results. Specifically, Path-CLIP achieves over \\(+19\\%\\) improvement in accuracy when utilizing mere \\(0.1\\%\\) of labeled data in PCam with only 10 minutes of fine-tuning while running on a single GPU." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.66, + 0.377, + 0.675 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Recent advancements in deep learning, bolstered by superior network architectures and the availability of extensive, meticulously curated datasets, have led to substantial enhancements in pathology image analysis tasks [17]. Nevertheless, the process of collecting high-quality datasets with reliable annotations for each specific vision task remains a laborious and time-intensive endeavor [18,53]. This may prevent the broad adoption of advanced deep learning techniques. To relieve the reliance on such datasets, pre-training and fine-tuning methods (TFT) have been studied in vision tasks: pre-train the model on a large-scale dataset and then fine-tune the model on different downstream tasks [9]. There are several challenges of such methods: 1) they may still require a large amount of labeled" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Z. Lai et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.506, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.284, + 0.376, + 0.295 + ], + "angle": 0, + "content": "(a) CLIP" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.145, + 0.785, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.285, + 0.667, + 0.296 + ], + "angle": 0, + "content": "(b) Path-CLIP" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.311, + 0.789, + 0.355 + ], + "angle": 0, + "content": "Fig. 1: Illustration of (a) zero-shot CLIP, and (b) our proposed Path-CLIP. We aim to freeze the encoders but only train the Residual Feature Refinement (RFR) module to balance the original knowledge from CLIP and new task-specific knowledge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.789, + 0.477 + ], + "angle": 0, + "content": "set to avoid the overfitting issue when fine-tuning the model for the downstream task [21, 40]; 2) the fine-tuning may not yield satisfactory performance in the target domain due to the existence of a large domain gap between the pre-trained data and pathology images [32]. Consequently, a pivotal question arises: how can the large-scale pre-trained model be efficiently tailored to a new task, minimizing overfitting, particularly when the new dataset is limited in size?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.484, + 0.79, + 0.68 + ], + "angle": 0, + "content": "To fill the performance gap due to domain shift, Contrastive Language-Image Pre-training (CLIP) [26] has shown its power in learning generic and distinctive visual representations via language supervision. As shown in Fig. 1(a), it aligns images and texts in the same feature space and uses a contrastive loss to formulate the learning objective. CLIP uses two separate encoders for images and texts, then maximizes the similarity score of positive pairs of images and texts while minimizing it of negative pairs [26,54]. It achieves promising results on various image classification tasks without needing any annotated data, i.e., zero-shot transfer settings. As a language-vision model, CLIP uses prompts as the supervision, where the visual labels are entered into the hand-crafted template. By pre-training the model at a large scale, models can learn the visual contents and easily be transferred to downstream tasks through the prompt-based zero-shot transfer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.79, + 0.842 + ], + "angle": 0, + "content": "However, the manual design of prompts can be a non-trivial and time-consuming task. [54] found that even a slight change in the prompt (e.g., one word) can make a big difference. They introduced Context Optimization (CoOp) to automate prompt engineering to generate continuous soft prompts instead of using hand-chosen hard prompts [54]. However, CoOp requires substantial computing resources, and the results of CoOp are not interpretable. Furthermore, such prompt tuning faces performance degradation when there is a significant domain shift [39], e.g., from natural images to pathology images, rendering it less adaptable to medical imaging tasks. On the other hand, medical image analysis may involve multiple learning tasks simultaneously. Conversely, in medical im" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.273, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Path-CLIP: Efficiently Adapting CLIP for Pathology Image Analysis" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "age analysis, multiple learning tasks may be concurrently required. It becomes impractical if fine-tuning necessitates substantial computational resources. For instance, Lai et al. [17] quantify and visualize distinct pathologies in grey/white matter (GM/WM) regions within gigapixel images. To accomplish this, the task is divided into two learning objectives: pathology detection and GM/WM segmentation. They collected and annotated two datasets and then trained two models from scratch to achieve it. Therefore, it will be beneficial if we can propose an effective way to adapt a big model into multiple downstream tasks quickly so the final analysis objective can be reached in a reasonable time manner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.789, + 0.706 + ], + "angle": 0, + "content": "In this work, we aim to fine-tune CLIP efficiently with light computing resources for pathology image analysis tasks. There are several challenges in fine-tuning CLIP. First, overfitting is a severe issue if we adapt CLIP to the downstream tasks directly since CLIP is pre-trained on a 400M dataset while the new domain dataset can be small [10]. Second, CLIP has shown its promising zero-shot generalization ability, directly fine-tuning CLIP may lose this capability and face the \"catastrophic forgetting\" issue. In other words, it is unclear how to effectively learn the new knowledge while retaining the original pre-trained knowledge so that the generalization of CLIP can be maintained. Third, there are limited studies on benchmarking CLIP's transferability in the pathology domain; hence, its applicability remains unclear. To address the above issues, we first study the applicability of CLIP on two public pathology datasets and benchmark the zero-shot ability of CLIP on them. Then, we did empirical experiments to verify the importance of the CLIP's pre-trained knowledge. To address the \"catastrophic forgetting\" issue, we propose Residual Feature Refinement (RFR) as a lightweight approach for adapting CLIP to pathology images. As a shallow network, RFR demonstrates proficiency in extracting crucial spatial features and patterns from pathology images. Within RFR, we introduce a self-adaptive module designed to seamlessly fuse and balance the original knowledge from CLIP with the newly acquired insights from the specific pathological task, requiring only a minimal number of additional trainable weights, as opposed to retraining the entire encoders in CLIP. To further mitigate overfitting, we also incorporate Hidden Representation Perturbation (HRP) and Dual-view Vision Contrastive (DVC). Finally, we establish an end-to-end adaptation of CLIP, named Path-CLIP (as shown in Fig. 1(b)), tailored for pathology image analysis tasks. Path-CLIP incorporates a newly proposed loss function, Doublet Multimodal Contrastive Loss (DMCL)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.72, + 0.533, + 0.735 + ], + "angle": 0, + "content": "We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.75, + 0.787, + 0.794 + ], + "angle": 0, + "content": "- We explore the applicability of CLIP on pathology images and highlight the catastrophic forgetting issues when adapting CLIP to downstream pathology tasks." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- We propose Path-CLIP to alleviate the overfitting and forgetting issues and effectively adapt CLIP for pathology tasks." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.128 + ], + "angle": 0, + "content": "Z. Lai et al." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.147, + 0.79, + 0.192 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate the efficiency and robustness of PathCLIP. For example, it achieves \\(+19\\%\\) on PCam when only using \\(0.1\\%\\) of labeled data." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.216, + 0.388, + 0.232 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.248, + 0.789, + 0.475 + ], + "angle": 0, + "content": "Language-vision model (LVM). LVMs have shown promising performance in learning generic visual representations [15, 22, 26, 48]. Recent models advance via text representation learning with large-scale Transformers [30] and even web-scale training datasets [54]. Transformer-based multimodal learning has achieved great success in such large-scale datasets [23, 29, 36]. For example, CLIP [26] was trained on 400 million image-caption pairs and achieved state-of-the-art results in many fields [26, 44, 46, 47]. In CLIP's architecture, it has two encoders: the vision encoder can be ResNet [12] or ViT [7] while the text encoder can be Transformer (e.g., BERT [6]). A recent work [27] fine-tuned CLIP for video data and performed competitively to more complicated methods designed for video processing. PointCLIP [50] applied CLIP to 3D recognition. CLIP is also applied for image generation tasks [8] and shows its power to reduce required data collection. However, the best way to adapt CLIP for downstream tasks is still being studied, especially when the new domain is far from the pre-trained domains, e.g., the medical domain." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.79, + 0.71 + ], + "angle": 0, + "content": "CLIP Fine-tuning and adaptation. Although recent multimodal models (e.g., CLIP [26]) perform astonishingly well in standard vision tasks without fine-tuning, the gap between CLIP and supervised learning is still noticeable. Therefore, many works focus on fine-tuning CLIP with a few samples from the new task provided. For example, CoOp [54] kept the entire pre-trained parameters fixed but optimized the continuous prompt through a learnable vector. This method closes the gap between CLIP and supervised learning but the performance gain is limited as it only focuses on the perspective of prompt design instead of the vision part for the image recognition tasks. On the other hand, CLIP-Adapter [9] and Tip-Adapter [51] freeze the pre-trained weights and train additive module, which significantly improves the CLIP's performance with limited training data in the new task. These methods require a lower computational cost. However, both CLIP-Adapter [9] and Tip-Adapter [51] involve tuning a ratio (hyper-parameter) to balance the original and fine-tuned knowledge manually, which may not be practical when we have multiple target learning tasks in the medical domain." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.732, + 0.381, + 0.751 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In this section, we introduce the proposed Path-CLIP. We summarize the end-to-end pipeline in Fig. 1(b): a self-adaptive residual ratio \\(\\alpha\\) will be incorporated into Residual Feature Refinement (Section 3.2) to fuse and balance the original knowledge with the task-specific knowledge from the new downstream task. We introduce Hidden Representation Perturbation (HRP) and Dual-view Vision" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.273, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Path-CLIP: Efficiently Adapting CLIP for Pathology Image Analysis" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.143, + 0.761, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.349, + 0.79, + 0.434 + ], + "angle": 0, + "content": "Fig. 2: Overview of the Residual Feature Refinement (RFR): a downsampling-upsampling shallow architecture of fully connected layers. RFR blends the fine-tuned knowledge with the original knowledge from CLIP's vision encoder \\((F(\\cdot))\\) via a self-adaptive residual ratio \\(\\alpha\\). HRP refers to Hidden Representation Perturbation and TFT refers to pre-training fine-tuning. We use blue color to denote the original visual features and orange color to denote the fine-tuned features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.471, + 0.788, + 0.517 + ], + "angle": 0, + "content": "Contrastive (Section 3.3) to alleviate the overfitting issue. Finally, we propose Doublet Multimodal Contrastive Loss (DMCL) in Section 3.4 as the loss function for fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.557, + 0.506, + 0.573 + ], + "angle": 0, + "content": "3.1 Language-Vision Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.789, + 0.72 + ], + "angle": 0, + "content": "We first revisit language-vision pre-training and use CLIP [26] as one example. Note that our framework is applicable to broader language-vision models. CLIP [26] has a vision encoder \\( F(\\cdot) \\) and a text encoder \\( G(\\cdot) \\). The vision encoder maps a high-dimensional image into low-dimensional image embeddings. The vision encoder can be a CNN like ResNet [12] or a Vision Transformer (ViT) [7]. The text encoder is built on Transformer [30] and generates text embeddings from the prompt. Specifically, \"prompt\" here refers to a sequence of words (tokens), such as \"a neuropathological image of grey matter\"." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Training. CLIP jointly trains \\( F(\\cdot) \\) and \\( G(\\cdot) \\) to optimize the similarity score (e.g., symmetric cross-entropy loss [33]) between the visual and textual embeddings for each batch. The input consists of an image and its corresponding prompt (e.g., \"this is a photo of grey matter\"). They formulate the learning objective as a contrastive loss to align the visual and textual embedding spaces. Given a batch of image-prompt pairs, CLIP maximizes the similarity score for positive pairs while minimizing it for negative pairs." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.128 + ], + "angle": 0, + "content": "Z. Lai et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.178 + ], + "angle": 0, + "content": "Contrastive loss function. In CLIP training, the agreement between a positive pair of vision and text embeddings is maximized by InfoNCE loss [24]:" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.188, + 0.788, + 0.223 + ], + "angle": 0, + "content": "\\[\nl _ {i} ^ {v \\rightarrow t} = - \\log \\frac {\\exp \\left(\\mathbf {s i m} \\left(v _ {i} , t _ {i}\\right) / \\gamma\\right)}{\\sum_ {k = 1} ^ {B} \\exp \\left(\\mathbf {s i m} \\left(v _ {i} , t _ {k}\\right) / \\gamma\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.231, + 0.789, + 0.323 + ], + "angle": 0, + "content": "where \\( v, t \\) are normalized vectors from vision and text encoders. \\( (v_{i}, t_{i}) \\) is a positive pair. \\( \\mathbf{sim} \\) is a function to measure the similarity of two vectors. \\( B \\) is the mini-batch size of the image-text pairs and \\( \\gamma \\) is the learnable temperature. Therefore, there is only 1 positive text for the input \\( i^{th} \\) image but \\( N - 1 \\) negative texts. \\( l_{i}^{v \\rightarrow t} \\) means the InfoNCE loss from an image \\( i \\) to texts, so as the text loss \\( l_{i}^{t \\rightarrow v} \\). Then the loss function for CLIP is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.401, + 0.335, + 0.788, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C L I P}} = \\frac {1}{2 B} \\sum_ {i = 1} ^ {B} \\left(l _ {i} ^ {v \\rightarrow t} + l _ {i} ^ {t \\rightarrow v}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.385, + 0.789, + 0.49 + ], + "angle": 0, + "content": "Inference for classification tasks. For the inference process, an image \\( x \\) is transformed into a feature manifold \\( f \\in \\mathbb{R}^D \\), where \\( D \\) is the feature dimension. Then, \\( f \\) is multiplied with a classifier weight matrix \\( \\mathbf{W} \\in \\mathbb{R}^{D \\times K} \\), where \\( K \\) is the number of classes in the learning task. We get a \\( K \\)-dimensional logit after matrix multiplication. Then we apply softmax to convert this logit into a probability vector \\( p \\in \\mathbb{R}^K \\) over the \\( K \\) classes. The whole process can be summarized as the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.406, + 0.491, + 0.788, + 0.527 + ], + "angle": 0, + "content": "\\[\np _ {i} = \\frac {\\exp \\left(\\mathbf {W} _ {\\mathbf {i}} ^ {T} \\times f\\right) / \\gamma}{\\sum_ {i = 1} ^ {K} \\exp \\left(\\mathbf {W} _ {\\mathbf {i}} ^ {T} \\times f\\right) / \\gamma}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.531, + 0.788, + 0.561 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is a temperature parameter learned by CLIP during training and \\(\\mathbf{W_i}\\) is the prototype weight vector for class- \\(i\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.568, + 0.789, + 0.675 + ], + "angle": 0, + "content": "Transferability. Compared to traditional classifier learning methods where the supervision is a fixed one-hot vector, language-vision pre-training utilizes an open-set visual concept that can be explored through the text encoder to learn a broader semantic space, which in turn makes the learned representations more diverse and adaptable to various downstream tasks (e.g., zero-shot learning). Therefore, we argue that such language-vision pre-training is helpful for medical problems where the domain is very distinctive from the natural images." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.696, + 0.502, + 0.71 + ], + "angle": 0, + "content": "3.2 Residual Feature Refinement" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this subsection, we introduce Residual Feature Refinement (RFR) to learn the task-relevant context when we adapt CLIP to the downstream learning tasks. Unlike CoOp's prompt tuning, which may not address the domain shift issue between the natural and pathology images, we focus on fine-tuning the visual features \\( f \\). However, simple fine-tuning of the entire network may fail in a new pathological task due to 1) the overfitting issue caused by a large number of parameters in CLIP and a shortage of training samples [13]; 2) catastrophic forgetting issue." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.273, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Path-CLIP: Efficiently Adapting CLIP for Pathology Image Analysis" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.144, + 0.788, + 0.175 + ], + "angle": 0, + "content": "Table 1: Baseline. We fine-tune CLIP (ResNet-50) on PCam [31] with \\(1\\%\\) training data. \"No fine-tuning\" is the best baseline due to the catastrophic forgetting issue." + }, + { + "type": "table", + "bbox": [ + 0.343, + 0.185, + 0.656, + 0.272 + ], + "angle": 0, + "content": "
# ConfigurationAccuracy
1No fine-tuning56.5%
2Fine-tune 10 epochs40.3%
3Fine-tune 100 epochs35.9%
4With a lower learning rate on #345.6%
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.306, + 0.79, + 0.594 + ], + "angle": 0, + "content": "Importance of the CLIP's pre-trained knowledge. CLIP has shown its zero-shot ability and transferability to a wide range of downstream tasks. We argue that preserving pre-trained knowledge of CLIP is crucial to maintain its generalization and alleviate the overfitting issue when the new dataset is small. There are two reasons behind this. First, one recent NLP work [11] shows large pre-trained language models (PLMs) are well-calibrated on the masked language modeling tasks. However, such property cannot be retained through direct fine-tuning due to catastrophic forgetting. Specifically, the pre-trained features are only used for the initialization and will be distorted by the fine-tuning [11]. Inspired by this, we investigate whether this issue also appears in language-vision models. As shown in Table 1, \"no fine-tuning\" but zero-shot gets the best results compared to fine-tuning CLIP on \\(1\\%\\) of training data in PCam [31]. We also attempt to lower the learning rate to \\(1 \\times 10^{-6}\\), but the results are still far from satisfactory. Therefore, we conclude direct fine-tuning may distort the pre-trained knowledge and result in a severe catastrophic forgetting issue. Besides the forgetting issue, overfitting is another common challenge when adapting a big model to a new task. To alleviate the above two issues, we propose a residual refinement architecture to preserve the pre-trained features and acquire rich semantic information for pathology image analysis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.79, + 0.729 + ], + "angle": 0, + "content": "Architecture overview. As shown in Fig. 2, given an image \\( x \\), we get the visual feature \\( f \\) from the image encoder and compute the classifier weight \\( \\mathbf{W} \\) from the text encoder. Then we design trainable adaptive layers \\( L \\) to convert \\( f \\) into \\( L(f) \\). \\( L \\) can be multiple layers of linear transformations in a \"down-and-up\" architecture. As shown in Fig. 2, we have four layers to transform the feature dimension as \"1024-256-64-256-1024\". This is one example: the layers are flexible with other dimensions. After this architecture, \\( L(f) \\) can be of the same size of \\( f \\) and blended with \\( f \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.747, + 0.787, + 0.763 + ], + "angle": 0, + "content": "\\[\nf _ {i} ^ {*} = \\left(1 - \\alpha_ {i}\\right) L \\left(f _ {i}\\right) + \\alpha_ {i} f _ {i}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\( i = \\{1,2,3,\\dots n\\} \\) and \\( \\alpha \\) is the self-adaptive residual ratio to balance the fine-tuned knowledge and the original CLIP's knowledge. Specifically, the first term adaptively learns knowledge from the new dataset, while the second term preserves the original knowledge from CLIP." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.128 + ], + "angle": 0, + "content": "Z. Lai et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Then we adopt Eq. (3) with the new \\( f^* \\) to get the class probability vector and predict the category. During fine-tuning, the weights of the trainable layers are optimized through the symmetric contrastive loss used in CLIP [26]. The prompt can be \"this is a photo of [], where \"+\" is filled with the class name. We can incorporate pathological descriptions into the prompt to provide more domain information, e.g., \"this is a pathological breast photo of []\"." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.246, + 0.788, + 0.399 + ], + "angle": 0, + "content": "Hidden Representation Perturbation (HRP). Fine-tuning a pre-trained language-vision model on a small dataset can lead to potential issues such as overfitting or representation collapse [45]. To mitigate these issues, we propose Hidden Representation Perturbation (HRP) that is strategically integrated with the trainable layers. To balance the model's complexity and efficiency, we adopt a \"four-layer\" architecture with HRP, steering clear of deeper architectures to ensure a controlled number of trainable parameters. This precaution is crucial as an excess of trainable parameters may exacerbate overfitting, especially in scenarios characterized by limited resources, where only a sparse dataset is available for the new task [52]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.79, + 0.528 + ], + "angle": 0, + "content": "To alleviate the overfitting issue in low-resource settings, we focus on adding noise during the training process, a widely-used NLP strategy for large language models to smoothen the optimization landscape. For example, NoisyTune [35] adds the perturbations to the pre-trained weights to improve the model's generalization. Inspired by these NLP strategies [19, 35, 45], we propose to add random noise to the feature \\( f \\) before trainable adapting layers fine-tuning it. The random noise can be Gaussian noise \\( \\varepsilon \\) generated from \\( \\mathcal{N}(0, \\sigma^2) \\). After this, the perturbed feature \\( f \\) is converted into \\( \\hat{f} = f + \\varepsilon \\). Therefore, Eq. (4) is converted into:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.542, + 0.788, + 0.561 + ], + "angle": 0, + "content": "\\[\nf _ {i} ^ {*} = \\left(1 - \\alpha_ {i}\\right) L \\left(\\hat {f} _ {i}\\right) + \\alpha_ {i} f _ {i}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.788, + 0.601 + ], + "angle": 0, + "content": "Therefore, the residual feature will be perturbed with the Gaussian noise for the refinement, designed to alleviate the overfitting issue." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.625, + 0.486, + 0.64 + ], + "angle": 0, + "content": "3.3 Self-adaptive residual ratio" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.651, + 0.788, + 0.803 + ], + "angle": 0, + "content": "In Section 3.2, we introduce residual feature refinement and the residual ratio \\(\\alpha\\). \\(\\alpha\\) helps modify the degree of preserving the original knowledge for better performance. Empirically, \\(\\alpha\\) should be small if the domain gap between the source data and the downstream task is enormous, so more knowledge from the downstream task \\((L(f))\\) can be considered. However, if \\(\\alpha\\) is too small, the model can be overfitting as it will heavily rely on the fine-tuned features and disregard the pre-trained knowledge from CLIP. On the other hand, the model can also be underfitting if \\(\\alpha\\) is too large, as it may prevent learning from the fine-tuned features. Therefore, it is crucial to quantify the residual ratio \\(\\alpha\\) so the model can be well-trained instead of underfitting or overfitting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Recent works such as Tip-Adapter [51] and CLIP-Adapter [9] involve feature refinement while they set such ratio as a fixed hyper-parameter that needs to be" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.273, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Path-CLIP: Efficiently Adapting CLIP for Pathology Image Analysis" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "tuned during the training process. We argue the residual ratio should be dynamic and self-adaptive during training instead of fixed, as the learning status changes along the training process. Hence, we propose a self-adaptive residual ratio in Path-CLIP." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.215, + 0.785, + 0.366 + ], + "angle": 0, + "content": "DVC: Dual-view Vision Contrastive. Inspired by self-supervised learning (e.g., SimCLR [4]), the composition of data augmentations is relevant to define effective predictive tasks. Therefore, we propose Dual-view Vision Contrastive (DVC) to measure the model's learning status. Given an image \\( x \\), we leverage two types of augmentations: \"weak\" and \"strong\" [28]. \"weak\" refers to the standard flip-and-shift augmentation strategy while \"strong\" refers to perturb the appearance (RandAugment [5] and CTAugment [2]). We propose to measure the distance between representations from both augmentations, e.g., cosine similarity has shown good performance in both visual contrastive [4] and textual contrastive [3] tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.787, + 0.418 + ], + "angle": 0, + "content": "With the input \\( x \\), we will get \\( w(x) \\) from the weak augmentation and \\( s(x) \\) from the strong augmentation. We obtain \\( F(w(x)) \\) and \\( F(s(x)) \\) after the vision encoder \\( F \\). We define the distance as \\( \\tau \\):" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.424, + 0.787, + 0.458 + ], + "angle": 0, + "content": "\\[\n\\tau = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\cos \\left(F \\left(w \\left(x _ {i}\\right)\\right), F \\left(s \\left(x _ {i}\\right)\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.784, + 0.524 + ], + "angle": 0, + "content": "where \\( N \\) is the number of samples in the training set, and \\( \\cos \\) is a cosine similarity. The model's learning status is relevant to its consistency across different perturbations. We traverse the entire training set and calculate its mean value to represent the model's performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.532, + 0.788, + 0.653 + ], + "angle": 0, + "content": "Bound constraints. In addition to DVC which measures the predictive consistency of the model, we also select \"accuracy\" on the training set as a second metric to measure the learning status, denoted as \\( T \\). Empirically, \"accuracy\" alone cannot fully tell if the model is well-fitted: 100% of accuracy may indicate overfitting, which is the biggest challenge when we adapt a large model to a downstream task where the training set is very small. However, \"accuracy\" is a good metric to show if the model is underfitting. Therefore, we propose to fuse DVC and \"accuracy\" to represent the model's status during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.784, + 0.737 + ], + "angle": 0, + "content": "If we look at \\(\\alpha\\) and set it fixed, we can find the model is prone to overfit if \\(\\alpha\\) is too small as it will heavily rely on the fine-tuned knowledge; the model tends to underfit if \\(\\alpha\\) is too large as it prevents the model from learning new knowledge (directly use the original pre-trained knowledge). Therefore, we propose a dynamic way with lower and upper bounds to obtain \\(\\alpha\\) during training:" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.743, + 0.786, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\alpha = \\min (\\max (\\tau , T), 1 - \\tau). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.786, + 0.841 + ], + "angle": 0, + "content": "The bounds are associated with predictive consistency (DVC) and predictive accuracy on the training set. For example, if the model is overfitting, \\( T \\) will be large, and \\( \\tau \\) will be relatively small. Then \\( \\alpha \\) will be smaller in the next epoch to use less fine-tuned features. Therefore, our proposed bound constraints can effectively adjust \\( \\alpha \\) to alleviate the overfitting issue." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Z. Lai et al." + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.144, + 0.734, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.315 + ], + "angle": 0, + "content": "Fig. 3: Doublet Multimodal Contrastive Loss (DMCL) that utilizes the traditional image-text pair and the proposed image-image pair with weak/strong augmentation criteria." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.348, + 0.78, + 0.365 + ], + "angle": 0, + "content": "3.4 Path-CLIP: End-to-end adaptation of CLIP to pathology tasks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.789, + 0.501 + ], + "angle": 0, + "content": "To efficiently adapt CLIP to pathology image tasks, we propose RFR with a self-adaptive residual ratio to balance the original and fine-tuned knowledge based on the learning status. To alleviate the overfitting issue, we incorporate HRP and design bound constraints for the residual ratio. Besides these, to distill more visual representations from the new image domain, we add a contrastive loss to the loss function so that the trainable portion has a better classifying ability. We follow SimCLR [4] to define the Contrastive Loss (CL) for one image in the batch as:" + }, + { + "type": "equation", + "bbox": [ + 0.344, + 0.512, + 0.788, + 0.577 + ], + "angle": 0, + "content": "\\[\nl _ {i} ^ {c} = \\left\\{ \\begin{array}{l} - \\log \\frac {\\exp (\\mathbf {s i m} (v _ {i} , v _ {i + B}) / \\gamma)}{\\sum_ {k \\neq i} ^ {2 B} \\exp (\\mathbf {s i m} (v _ {i} , v _ {k}) / \\gamma)}, \\text {i f} i < B, \\\\ - \\log \\frac {\\exp (\\mathbf {s i m} (v _ {i} , v _ {B - i}) / \\gamma)}{\\sum_ {k \\neq i} ^ {2 B} \\exp (\\mathbf {s i m} (v _ {i} , v _ {k}) / \\gamma)}, \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.788, + 0.687 + ], + "angle": 0, + "content": "Differing from SimCLR [4] and RepRec [38], as shown in Fig. 3, here we set image \\(i\\) as the weakly augmented image and image \\(i + B\\) as the strongly augmented image instead of two randomly augmented images as the positive pair. We named this loss as \\(\\mathcal{L}_{\\mathrm{SWAPP}}\\). SWAPP here refers to Strong-Weak Augmented Positive Pair. Finally, the total loss for fine-tuning CLIP on our pathology datasets is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.394, + 0.69, + 0.579, + 0.707 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {t o t a l}} = \\mathcal {L} _ {\\mathrm {C L I P}} + \\mathcal {L} _ {\\mathrm {S W A P P}}.\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.394, + 0.709, + 0.787, + 0.745 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C L I P}} \\rightarrow \\frac {1}{2 B} \\sum_ {i = 1} ^ {B} \\left(l _ {i} ^ {v \\rightarrow t} + l _ {i} ^ {t \\rightarrow v}\\right), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.748, + 0.543, + 0.785 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {S W A P P}} \\rightarrow \\frac {1}{2 B} \\sum_ {i = 1} ^ {2 B} l _ {i} ^ {c}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.843 + ], + "angle": 0, + "content": "We define this loss as Doublet Multimodal Contrastive Loss (DMCL) because the items involve both intra-contrastive (image and text) and inter-contrastive (image and image). The algorithm can be found in the Appendix." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.273, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Path-CLIP: Efficiently Adapting CLIP for Pathology Image Analysis" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.787, + 0.173 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison on PCam. Precision, Recall, and F1-score refer to the macro-averaged values from all classes" + }, + { + "type": "table", + "bbox": [ + 0.231, + 0.185, + 0.77, + 0.396 + ], + "angle": 0, + "content": "
Learning MannerLabeled RatioAlgorithmAccuracyPrecisionRecallF1-scoreAUC
Supervised100%-92.8 ± 0.3092.9 ± 0.2092.6 ± 0.2192.8 ± 0.200.95 ± 0.01
Supervised0.5%-45.1 ± 2.2742.0 ± 2.1433.6 ± 3.3037.3 ± 2.860.51 ± 0.01
Self-supervised0.5%SimCLR [4]60.4 ± 1.4563.3 ± 1.9058.5 ± 2.0160.8 ± 1.200.61 ± 0.04
Semi-supervised0.5%Pseudo-Label [20]55.7 ± 2.0559.2 ± 2.2354.5 ± 1.9756.8 ± 1.550.58 ± 0.03
FixMatch [28]73.2 ± 0.7677.5 ± 0.5073.7 ± 0.2575.6 ± 0.370.84 ± 0.04
Dash [37]71.0 ± 0.9875.3 ± 0.7870.2 ± 0.4572.7 ± 0.500.82 ± 0.02
FlexMatch [49]74.0 ± 0.8078.1 ± 1.1573.0 ± 0.9075.5 ± 0.840.85 ± 0.03
Multimodal CLIP0.5%Zero-shot [26]56.5 ± 0.0057.4 ± 0.0050.3 ± 0.0053.7 ± 0.000.60 ± 0.00
CoOp [54]63.6 ± 0.2563.9 ± 0.4262.5 ± 0.3563.0 ± 0.290.67 ± 0.02
CLIP-Adapter [9]72.3 ± 1.0277.2 ± 0.9563.2 ± 0.5669.4 ± 0.840.81 ± 0.02
TIP-Adapter [51]73.9 ± 0.5070.5 ± 0.8578.0 ± 0.4974.0 ± 0.950.84 ± 0.02
Path-CLIP81.9 ± 0.7879.4 ± 0.3585.0 ± 0.8082.1 ± 0.950.89 ± 0.02
Supervised1.5%-62.8 ± 1.3169.5 ± 0.6754.1 ± 2.2760.8 ± 1.680.63 ± 0.02
Self-supervised1.5%SimCLR [4]64.9 ± 1.0265.2 ± 1.4560.1 ± 1.7062.5 ± 1.050.64 ± 0.03
Semi-supervised1.5%Pseudo-Label [20]67.9 ± 1.8570.5 ± 1.9062.2 ± 1.7666.1 ± 1.420.67 ± 0.04
FixMatch [28]83.9 ± 1.2985.1 ± 0.8584.1 ± 0.9584.6 ± 0.900.87 ± 0.01
Dash [37]82.8 ± 1.0983.0 ± 0.5581.2 ± 0.6082.1 ± 0.780.85 ± 0.03
FlexMatch [49]84.1 ± 1.0284.9 ± 0.7584.4 ± 0.8084.6 ± 0.750.86 ± 0.04
Multimodal CLIP1.5%CoOp [54]61.5 ± 0.7560.5 ± 1.0365.6 ± 0.9063.1 ± 0.850.65 ± 0.02
CLIP-Adapter [9]74.1 ± 0.8879.1 ± 0.3465.4 ± 0.8971.6 ± 1.020.82 ± 0.01
TIP-Adapter [51]75.2 ± 0.9472.3 ± 0.6979.4 ± 0.8675.7 ± 0.730.83 ± 0.02
Path-CLIP83.1 ± 0.6782.8 ± 0.9582.0 ± 0.7582.4 ± 0.460.91 ± 0.02
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.421, + 0.377, + 0.438 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.496 + ], + "angle": 0, + "content": "Datasets. We select two classification datasets (PCam [31] and MHIST [34]) and a gigapixel grey/white matter segmentation dataset to evaluate Path-CLIP. The details of these datasets can be found in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.788, + 0.595 + ], + "angle": 0, + "content": "Experimental setup. For a fair comparison, we use ResNet-50 [12] as the backbone in the vision encoder \\( f(\\cdot) \\) and BERT [6] as the textual encoder. We follow CLIP [26] to use gradient scaling to facilitate mixed-precision training. We set the learning rate as 0.0001. The batch size is 32. We use Adam [16] as the optimizer. To compare the computational complexity, we conduct all experiments on a single GPU (Nvidia RTX 2080Ti)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.789, + 0.769 + ], + "angle": 0, + "content": "Baselines. We first implement traditional supervised learning (SL) methods with the entire dataset to explore the upper-bound performance for each task. Then we reduce the data usage and investigate the performance change of these SL methods. Then we compare Path-CLIP with original CLIP [26] and three recent CLIP-based fine-tuning methods: CoOp [54], Tip-Adapter [51] and CLIP-Adapter [9]. We compare them under the limited data scenarios to test their ability to do downstream tasks when the new dataset is small. Besides that, we also incorporate recent self-supervised learning (SimCLR [4]) and semi-supervised learning (Pseudo-Label [20], FixMatch [28], Dash [37], and FlexMatch [49]). Semi-supervised training requires a large amount of unlabeled data, which involves data collection efforts [41-43]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.367, + 0.803 + ], + "angle": 0, + "content": "4.1 Main results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Classification task: PCam [31] and MHIST [34]. The main results for two classification benchmarks are summarized in Table 2 and Table 3. First, CLIP" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "Z. Lai et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 3: Quantitative comparison on MHIST. Precision, Recall, and F1-score refer to the macro-averaged values from all classes" + }, + { + "type": "table", + "bbox": [ + 0.232, + 0.185, + 0.769, + 0.301 + ], + "angle": 0, + "content": "
Learning MannerLabeled RatioAlgorithmAccuracyPrecisionRecallF1-scoreAUC
Supervised100%-86.0 ± 0.3684.8 ± 0.6183.5 ± 0.5584.1 ± 0.580.89 ± 0.01
Self-supervised10%SimCLR [4]63.2 ± 0.9563.8 ± 1.2062.2 ± 1.7063.0 ± 0.850.66 ± 0.04
Semi-supervised10%Pseudo-Label [20]65.4 ± 1.0466.5 ± 0.9865.9 ± 0.8566.2 ± 1.420.75 ± 0.03
FixMatch [28]70.5 ± 1.5568.8 ± 2.7669.5 ± 2.9769.1 ± 2.860.79 ± 0.02
Dash [37]70.2 ± 1.2170.0 ± 0.8571.3 ± 0.6970.6 ± 1.480.81 ± 0.04
FlexMatch [49]73.1 ± 0.8474.6 ± 1.0575.1 ± 0.9074.8 ± 0.890.82 ± 0.03
Multimodal CLIP10%Zero-shot [26]36.9 ± 0.0036.9 ± 0.00100 ± 0.0053.9 ± 0.000.50 ± 0.00
CoOp [54]68.9 ± 0.4056.0 ± 0.6572.8 ± 0.5363.3 ± 0.860.76 ± 0.01
CLIP-Adapter [9]65.9 ± 1.4552.2 ± 1.2387.2 ± 0.8565.3 ± 0.950.77 ± 0.02
TIP-Adapter [51]63.1 ± 0.7849.9 ± 0.9094.7 ± 1.0265.4 ± 0.560.78 ± 0.03
Path-CLIP74.8 ± 1.9563.3 ± 1.4075.6 ± 0.9868.9 ± 1.500.84 ± 0.02
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.336, + 0.788, + 0.518 + ], + "angle": 0, + "content": "shows its promising zero-shot ability: when we directly test CLIP on PCam [31], it can get \\(56.5\\%\\) accuracy. Then, we compare Path-CLIP with other CLIP-based fine-tuning methods. In PCam [31], we present \\(0.5\\%\\) and \\(1.5\\%\\) data usage in the training set to fine-tune CLIP, respectively. We find all of these fine-tuning algorithms [9, 51, 54] can improve the performance of the original CLIP with the new data from the downstream task presented. Path-CLIP can outperform them in both settings. Specifically, Path-CLIP outperforms TIP-Adapter [51] with \\(8\\%\\) and \\(7.9\\%\\) accuracy when only \\(0.5\\%\\) and \\(1.5\\%\\) data are used, respectively. Path-CLIP also shows similar improvement in MHIST [34] (Table 3). We also implemented several semi-supervised learning where the rest of the data in the training set are used as an unlabeled set. Although Path-CLIP does not use unlabeled data, it can achieve competitive results with semi-supervised learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.53, + 0.788, + 0.622 + ], + "angle": 0, + "content": "Besides these ratios of data usage, we also visualize the results by varying the ratio to show the performance of each fine-tuning algorithm. As shown in Fig. 4, CoOp [54] may not improve PCam when more data are used to fine-tune the prompt. CoOp freezes the encoders and only fine-tunes the language prompt part, while other algorithms also fine-tune the vision part. We conclude that visual fine-tuning is more critical compared to prompt tuning." + }, + { + "type": "image", + "bbox": [ + 0.243, + 0.655, + 0.495, + 0.79 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.34, + 0.796, + 0.406, + 0.81 + ], + "angle": 0, + "content": "(a) PCam" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.655, + 0.753, + 0.79 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.795, + 0.665, + 0.81 + ], + "angle": 0, + "content": "(b) MHIST" + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.822, + 0.751, + 0.837 + ], + "angle": 0, + "content": "Fig. 4: Comparison of different fine-tuning methods by varying data usage." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.273, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Path-CLIP: Efficiently Adapting CLIP for Pathology Image Analysis" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.128 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.292, + 0.145, + 0.709, + 0.16 + ], + "angle": 0, + "content": "Table 4: Performance comparison on GM/WM segmentation." + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.172, + 0.728, + 0.381 + ], + "angle": 0, + "content": "
Data UsageAlgorithmGM_IoUGM_DICEWM_IoUWM_DICE
100% (SL)FCN [1]71.7283.2149.0966.08
U-Net [25]91.5295.1782.0288.15
0.1% (SL)FCN [1]48.2525.0124.5430.02
U-Net [25]46.1350.105.7512.23
0.1% (CLIP)CoOp [54]62.2365.7039.9248.65
CLIP-Adapter [9]67.8074.3243.2554.61
TIP-Adapter [51]69.9275.5046.2356.60
Path-CLIP78.0583.7054.5063.23
1% (SL)FCN [1]51.3430.2031.0235.12
U-Net [25]50.2453.2315.4325.58
1% (CLIP)CoOp [54]63.4564.1237.8545.34
CLIP-Adapter [9]69.9976.5447.9558.34
TIP-Adapter [51]72.1277.3055.6563.60
Path-CLIP80.2188.7060.5070.34
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.789, + 0.565 + ], + "angle": 0, + "content": "GM/WM segmentation [17]. Besides two classification tasks, we follow [17] to generalize our method to a weakly-supervised segmentation problem. We select IoU and DICE scores as the performance metrics. The results are summarized in Table 4. We select FCN [1] and U-Net [25] as two SL baselines: we first train them on all data (100%) in a supervised manner, then we reduce the dataset to 0.1% and 1% setting and train the model again. We can see both SL algorithms face performance degradation when the training data is scarce. They require an extensive and well-curated labeled dataset to get satisfactory results. We then focus on fine-tuning CLIP algorithms. Table 4 shows Path-CLIP can achieve superior results than other fine-tuning techniques." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.59, + 0.4, + 0.605 + ], + "angle": 0, + "content": "4.2 Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.619, + 0.789, + 0.742 + ], + "angle": 0, + "content": "Importance of alleviating overfitting. Since the new task may have a very limited training set, it is crucial to alleviate the potential overfitting issue. Besides the RFR with the self-adaptive ratio, we also propose HRP, DVC, and DMCL to cope with the overfitting issue. Table 5 summarizes the contribution of each proposed module: HRP can improve with \\(+2.2\\%\\) accuracy and DVC can improve with \\(+4.3\\%\\). DMCL inspired by self-supervised learning can also aid in great improvements. We conclude that contrastive learning, perturbations on data and hidden representations are helpful to the overfitting issue." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.79, + 0.843 + ], + "angle": 0, + "content": "Computational efficiency. When we range the data usage from \\(0.1\\%\\) to \\(1\\%\\), Path-CLIP can obtain \\(+5.5\\%\\) gain in accuracy while only introducing 1 minute of additional training time. However, CoOp [54] has the overfitting issue and obtains a lower score while using 53 minutes, almost five times compared to the time used by RFR. The situation remains similar under the \\(10\\%\\) of data usage. We conclude our proposed RFR can get over \\(19\\%\\) improvement on CLIP by only" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Z. Lai et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.275, + 0.145, + 0.726, + 0.16 + ], + "angle": 0, + "content": "Table 5: Ablation study on PCam using \\(0.5\\%\\) data for fine-tuning." + }, + { + "type": "table", + "bbox": [ + 0.359, + 0.172, + 0.645, + 0.251 + ], + "angle": 0, + "content": "
#RFCHRPDVCDMCLAccuracyF1-scoreAUC
1XXXX56.553.70.60
2✓XXX72.070.50.69
3✓✓XX74.275.50.78
4✓✓✓X78.579.30.84
5✓✓✓✓81.982.10.89
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.277, + 0.787, + 0.309 + ], + "angle": 0, + "content": "using 10 minutes for the fine-tuning on a single GPU, which is promising for digital pathology research. The details can be found in Table 8 in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.787, + 0.375 + ], + "angle": 0, + "content": "Generalization to pathology foundation model: PLIP [14]. The results of using PLIP's weights are in Appendix 3.6: PLIP is pre-trained on a large-scale pathology dataset and achieves strong zero-shot performance. With our method, we can further boost the performance and outperform other tuning methods." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.396, + 0.356, + 0.412 + ], + "angle": 0, + "content": "5 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.425, + 0.789, + 0.577 + ], + "angle": 0, + "content": "In this work, we explore the generalization of CLIP in pathology image analysis. To alleviate the overfitting and forgetting issues when adapting CLIP to small downstream datasets, we propose self-adaptive RFR with HRP. We propose Path-CLIP with DMCL as an end-to-end pipeline for CLIP adaptation in pathology tasks. The results reveal the superiority and effectiveness of Path-CLIP compared to the existing fine-tuning methods. In addition, the less computational time improves the efficiency and accessibility of Path-CLIP. Hence, we conclude Path-CLIP is more robust and has the potential to bridge the domain shift between the pre-trained natural images and pathology images in an efficient and cost-effective manner." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.597, + 0.395, + 0.615 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.787, + 0.703 + ], + "angle": 0, + "content": "This work was supported by the Noyce Initiative UC Partnerships in Computational Transformation Grant and Child Family Endowed Professorship. Resources for this study were funded in part by grants from the National Institute on Aging of the National Institutes of Health under Award Numbers R01AG062517, P30AG072972, and R01AG056519." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.723, + 0.324, + 0.739 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.743, + 0.788, + 0.799 + ], + "angle": 0, + "content": "1. Bandy, P., van de Loo, R., Intezar, M., Geijs, D., Ciompi, F., van Ginneken, B., van der Laak, J., Litjens, G.: Comparison of different methods for tissue segmentation in histopathological whole-slide images. In: Proc. 2017 IEEE Int. Symp. Biomed. Imaging. pp. 591-595 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.799, + 0.788, + 0.84 + ], + "angle": 0, + "content": "2. Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., Raffel, C.A.: MixMatch: A holistic approach to semi-supervised learning. In: NeurIPS. vol. 32 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.743, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.273, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Path-CLIP: Efficiently Adapting CLIP for Pathology Image Analysis" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "3. Chen, F., Zhang, H., Li, Z., Dou, J., Mo, S., Chen, H., Zhang, Y., Ahmed, U., Zhu, C., Savvides, M.: Unitail: Detecting, reading, and matching in retail scene. In: ECCV. pp. 705-722. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.19, + 0.787, + 0.217 + ], + "angle": 0, + "content": "4. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML. pp. 1597-1607 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.217, + 0.787, + 0.257 + ], + "angle": 0, + "content": "5. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Workshop. pp. 702-703 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.257, + 0.787, + 0.298 + ], + "angle": 0, + "content": "6. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.298, + 0.787, + 0.338 + ], + "angle": 0, + "content": "7. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.787, + 0.379 + ], + "angle": 0, + "content": "8. Gal, R., Patashnik, O., Maron, H., Bermano, A.H., Chechik, G., Cohen-Or, D.: Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG) 41(4), 1-13 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.379, + 0.787, + 0.42 + ], + "angle": 0, + "content": "9. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.787, + 0.447 + ], + "angle": 0, + "content": "0. Ge, C., Huang, R., Xie, M., Lai, Z., Song, S., Li, S., Huang, G.: Domain adaptation via prompt learning. arXiv preprint arXiv:2202.06687 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.447, + 0.787, + 0.473 + ], + "angle": 0, + "content": "1. He, G., Chen, J., Zhu, J.: Preserving pre-trained features helps calibrate fine-tuned language models. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.787, + 0.5 + ], + "angle": 0, + "content": "2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.501, + 0.787, + 0.542 + ], + "angle": 0, + "content": "3. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: ICML. pp. 2790-2799. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.542, + 0.787, + 0.568 + ], + "angle": 0, + "content": "4. Huang, Z., other: A visual-language foundation model for pathology image analysis using medical twitter. Nature medicine (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.568, + 0.787, + 0.609 + ], + "angle": 0, + "content": "5. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML. pp. 4904-4916 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.609, + 0.787, + 0.637 + ], + "angle": 0, + "content": "6. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (Poster) (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.637, + 0.787, + 0.69 + ], + "angle": 0, + "content": "7. Lai, Z., Oliveira, L.C., Guo, R., Xu, W., Hu, Z., Mifflin, K., Decarli, C., Cheung, S.C., Chuah, C.N., Dugger, B.N.: Brainsec: Automated brain tissue segmentation pipeline for scalable neuropathological analysis. IEEE Access 10, 49064-49079 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.787, + 0.732 + ], + "angle": 0, + "content": "8. Lai, Z., Wang, C., Oliveira, L.C., Dugger, B.N., Cheung, S.C., Chuah, C.N.: Joint semi-supervised and active learning for segmentation of gigapixel pathology images with cost-effective labeling. In: ICCV Workshop. pp. 591-600 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.787, + 0.759 + ], + "angle": 0, + "content": "9. Lam, S.Y., Zeng, Q., Zhang, K., You, C., Voigt, R.: Large language models are partially primed in pronoun interpretation. arXiv preprint arXiv:2305.16917 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "20. Lee, D.H.: Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In: Workshop on challenges in representation learning (ICML) (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "21. Liu, M., Hu, L., Tang, Y., Wang, C., He, Y., Zeng, C., Lin, K., He, Z., Huo, W.: A deep learning method for breast cancer classification in the pathology images. IEEE J. Biomed. Health. Inf. 26(10), 5025-5032 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "Z. Lai et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "22. Lu, Y., Liu, J., Zhang, Y., Liu, Y., Tian, X.: Prompt distribution learning. In: CVPR. pp. 5206-5215 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.788, + 0.23 + ], + "angle": 0, + "content": "23. Lyu, W., Dong, X., Wong, R., Zheng, S., Abell-Hart, K., Wang, F., Chen, C.: A multimodal transformer: Fusing clinical notes with structured ehr data for interpretable in-hospital mortality prediction. In: AMIA Annual Symposium Proceedings. vol. 2022, p. 719. American Medical Informatics Association (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.787, + 0.257 + ], + "angle": 0, + "content": "24. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.787, + 0.298 + ], + "angle": 0, + "content": "25. Oskal, K.R., Risdal, M., Janssen, E.A., Undersrud, E.S., Gulsrud, T.O.: A U-net based approach to epidermal tissue segmentation in whole slide histopathological images. SN Appl. Sci. 1(7), 1-12 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.787, + 0.338 + ], + "angle": 0, + "content": "26. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.787, + 0.379 + ], + "angle": 0, + "content": "27. Rasheed, H., khattak, M.U., Maaz, M., Khan, S., Khan, F.S.: Finetuned clip models are efficient video learners. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.787, + 0.42 + ], + "angle": 0, + "content": "28. Sohn, K., Berthelot, D., Carlini, N., Zhang, Z., Zhang, H., Raffel, C.A., Cubuk, E.D., Kurakin, A., Li, C.L.: FixMatch: Simplifying semi-supervised learning with consistency and confidence. In: NeurIPS. vol. 33, pp. 596-608 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.787, + 0.461 + ], + "angle": 0, + "content": "29. Tian, Y., Newsam, S., Boakye, K.: Fashion image retrieval with text feedback by additive attention compositional learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1011-1021 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.787, + 0.488 + ], + "angle": 0, + "content": "30. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.514 + ], + "angle": 0, + "content": "31. Veeling, B.S., Linmans, J., Winkens, J., Cohen, T., Welling, M.: Rotation equivariant cnns for digital pathology. In: MICCAI. pp. 210-218. Springer (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.787, + 0.555 + ], + "angle": 0, + "content": "32. Wang, X., Yang, S., Zhang, J., Wang, M., Zhang, J., Huang, J., Yang, W., Han, X.: Transpath: Transformer-based self-supervised learning for histopathological image classification. In: MICCAI. pp. 186-195. Springer (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.787, + 0.582 + ], + "angle": 0, + "content": "33. Wang, Y., Ma, X., Chen, Z., Luo, Y., Yi, J., Bailey, J.: Symmetric cross entropy for robust learning with noisy labels. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.787, + 0.623 + ], + "angle": 0, + "content": "34. Wei, J., Suriawinata, A., Ren, B., Liu, X., Lisovsky, M., Vaickus, L., Brown, C., Baker, M., Tomita, N., Torresani, L., et al.: A petri dish for histopathology image analysis. In: Int. Conf. Arti. Intell. in Medi. pp. 11-24. Springer (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.787, + 0.677 + ], + "angle": 0, + "content": "35. Wu, C., Wu, F., Qi, T., Huang, Y.: Noisytune: A little noise can help you finetune pretrained language models better. In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). pp. 680-685 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.787, + 0.704 + ], + "angle": 0, + "content": "36. Xu, P., Zhu, X., Clifton, D.A.: Multimodal learning with transformers: A survey. arXiv preprint arXiv:2206.06488 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.787, + 0.744 + ], + "angle": 0, + "content": "37. Xu, Y., Shang, L., Ye, J., Qian, Q., Li, Y.F., Sun, B., Li, H., Jin, R.: Dash: Semi-supervised learning with dynamic thresholding. In: International Conference on Machine Learning. pp. 11525-11536. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.787, + 0.8 + ], + "angle": 0, + "content": "38. Yan, X., Naushad, J., Sun, S., Han, K., Tang, H., Kong, D., Ma, H., You, C., Xie, X.: Representation recovering for self-supervised pre-training on medical images. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2685-2695 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "39. Yang, B., Liu, F., Li, Z., Yin, Q., You, C., Yin, B., Zou, Y.: Multimodal prompt learning for product title generation with extremely limited labels. arXiv preprint arXiv:2307.01969 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.273, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Path-CLIP: Efficiently Adapting CLIP for Pathology Image Analysis" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "40. Yang, J., Chen, H., Liang, Y., Huang, J., He, L., Yao, J.: Concl: Concept contrastive learning for dense prediction pre-training in pathology images. In: ECCV. pp. 523-539. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "41. You, C., Dai, W., Min, Y., Liu, F., Zhang, X., Feng, C., Clifton, D.A., Zhou, S.K., Staib, L.H., Duncan, J.S.: Rethinking semi-supervised medical image segmentation: A variance-reduction perspective. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.788, + 0.286 + ], + "angle": 0, + "content": "42. You, C., Dai, W., Min, Y., Staib, L., Duncan, J.S.: Bootstrapping semi-supervised medical image segmentation with anatomical-aware contrastive distillation. In: International Conference on Information Processing in Medical Imaging. pp. 641-653. Springer (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.788, + 0.328 + ], + "angle": 0, + "content": "43. You, C., Dai, W., Min, Y., Staib, L., Sekhon, J., Duncan, J.S.: Action++: Improving semi-supervised medical image segmentation with adaptive anatomical contrast. arXiv preprint arXiv:2304.02689 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.788, + 0.37 + ], + "angle": 0, + "content": "44. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.788, + 0.411 + ], + "angle": 0, + "content": "45. Yuan, H., Yuan, Z., Tan, C., Huang, F., Huang, S.: Hype: Better pre-trained language model fine-tuning with hidden representation perturbation. arXiv preprint arXiv:2212.08853 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.788, + 0.453 + ], + "angle": 0, + "content": "46. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.788, + 0.495 + ], + "angle": 0, + "content": "47. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12104-12113 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.788, + 0.535 + ], + "angle": 0, + "content": "48. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: Lit: Zero-shot transfer with locked-image text tuning. In: CVPR. pp. 18123-18133 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.535, + 0.788, + 0.577 + ], + "angle": 0, + "content": "49. Zhang, B., Wang, Y., Hou, W., Wu, H., Wang, J., Okumura, M., Shinozaki, T.: Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. NeurIPS 34 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.577, + 0.788, + 0.619 + ], + "angle": 0, + "content": "50. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8552-8562 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.788, + 0.66 + ], + "angle": 0, + "content": "51. Zhang, R., Zhang, W., Fang, R., Gao, P., Li, K., Dai, J., Qiao, Y., Li, H.: Tip-adapter: Training-free adaption of clip for few-shot classification. In: ECCV. pp. 493-510. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.788, + 0.688 + ], + "angle": 0, + "content": "52. Zhang, T., Wu, F., Katiyar, A., Weinberger, K.Q., Artzi, Y.: Revisiting few-sample bert fine-tuning. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.788, + 0.73 + ], + "angle": 0, + "content": "53. Zhang, W., Zhu, L., Hallinan, J., Zhang, S., Makmur, A., Cai, Q., Ooi, B.C.: Boostmis: Boosting medical image semi-supervised learning with adaptive pseudo labeling and informative active annotation. In: CVPR. pp. 20666-20676 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.788, + 0.757 + ], + "angle": 0, + "content": "54. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. Int. J. Comput. Vision 130(9), 2337-2348 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.757 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Bridging the Pathology Domain Gap_ Efficiently Adapting CLIP for Pathology Image Analysis with Limited Labeled Data/3b6fb56d-1649-4e06-bdb2-706d8b03ddb5_origin.pdf b/2024/Bridging the Pathology Domain Gap_ Efficiently Adapting CLIP for Pathology Image Analysis with Limited Labeled Data/3b6fb56d-1649-4e06-bdb2-706d8b03ddb5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1737e902597ab8520f7831082fb6cb7fe386dc47 --- /dev/null +++ b/2024/Bridging the Pathology Domain Gap_ Efficiently Adapting CLIP for Pathology Image Analysis with Limited Labeled Data/3b6fb56d-1649-4e06-bdb2-706d8b03ddb5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:115bdaa60037e8e3559842144ca78ec82001f93ac50bc08544ac015ad5953e11 +size 2529629 diff --git a/2024/Bridging the Pathology Domain Gap_ Efficiently Adapting CLIP for Pathology Image Analysis with Limited Labeled Data/full.md b/2024/Bridging the Pathology Domain Gap_ Efficiently Adapting CLIP for Pathology Image Analysis with Limited Labeled Data/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b8d2a7a499fbc550ddf88dc69f8301c99e7b0381 --- /dev/null +++ b/2024/Bridging the Pathology Domain Gap_ Efficiently Adapting CLIP for Pathology Image Analysis with Limited Labeled Data/full.md @@ -0,0 +1,283 @@ +# Bridging the Pathology Domain Gap: Efficiently Adapting CLIP for Pathology Image Analysis with Limited Labeled Data + +Zhengfeng Lai1, Joohi Chauhan2, Brittany N. Dugger1, and Chen-Nee Chuah1 + +1 University of California, Davis +2 Motilal Nehru National Institute of Technology Allahabad {lzhengfeng, bndugger, chuah}@ucdavis.edu joohi@mnnit.ac.in + +Abstract. Contrastive Language-Image Pre-training (CLIP) has shown its proficiency in acquiring distinctive visual representations and exhibiting strong generalization across diverse vision tasks. However, its effectiveness in pathology image analysis, particularly with limited labeled data, remains an ongoing area of investigation due to challenges associated with significant domain shifts and catastrophic forgetting. Thus, it is imperative to devise efficient adaptation strategies in this domain to enable scalable analysis. In this study, we introduce Path-CLIP, a framework tailored for a swift adaptation of CLIP to various pathology tasks. Firstly, we propose Residual Feature Refinement (RFR) with a dynamically adjustable ratio to effectively integrate and balance source and task-specific knowledge. Secondly, we introduce Hidden Representation Perturbation (HRP) and Dual-view Vision Contrastive (DVC) techniques to mitigate overfitting issues. Finally, we present the Doublet Multimodal Contrastive Loss (DMCL) for fine-tuning CLIP for pathology tasks. We demonstrate that Path-CLIP adeptly adapts pre-trained CLIP to downstream pathology tasks, yielding competitive results. Specifically, Path-CLIP achieves over $+19\%$ improvement in accuracy when utilizing mere $0.1\%$ of labeled data in PCam with only 10 minutes of fine-tuning while running on a single GPU. + +# 1 Introduction + +Recent advancements in deep learning, bolstered by superior network architectures and the availability of extensive, meticulously curated datasets, have led to substantial enhancements in pathology image analysis tasks [17]. Nevertheless, the process of collecting high-quality datasets with reliable annotations for each specific vision task remains a laborious and time-intensive endeavor [18,53]. This may prevent the broad adoption of advanced deep learning techniques. To relieve the reliance on such datasets, pre-training and fine-tuning methods (TFT) have been studied in vision tasks: pre-train the model on a large-scale dataset and then fine-tune the model on different downstream tasks [9]. There are several challenges of such methods: 1) they may still require a large amount of labeled + +![](images/035939840df9d49f945cdec888d3339b46dc0c353203c6d467ff41819d128dd2.jpg) +(a) CLIP + +![](images/3ac9c8ba0cb6d4a4780890b6a0397c57575cec8aac5b5ff4b35b60f5a6766d8e.jpg) +(b) Path-CLIP +Fig. 1: Illustration of (a) zero-shot CLIP, and (b) our proposed Path-CLIP. We aim to freeze the encoders but only train the Residual Feature Refinement (RFR) module to balance the original knowledge from CLIP and new task-specific knowledge. + +set to avoid the overfitting issue when fine-tuning the model for the downstream task [21, 40]; 2) the fine-tuning may not yield satisfactory performance in the target domain due to the existence of a large domain gap between the pre-trained data and pathology images [32]. Consequently, a pivotal question arises: how can the large-scale pre-trained model be efficiently tailored to a new task, minimizing overfitting, particularly when the new dataset is limited in size? + +To fill the performance gap due to domain shift, Contrastive Language-Image Pre-training (CLIP) [26] has shown its power in learning generic and distinctive visual representations via language supervision. As shown in Fig. 1(a), it aligns images and texts in the same feature space and uses a contrastive loss to formulate the learning objective. CLIP uses two separate encoders for images and texts, then maximizes the similarity score of positive pairs of images and texts while minimizing it of negative pairs [26,54]. It achieves promising results on various image classification tasks without needing any annotated data, i.e., zero-shot transfer settings. As a language-vision model, CLIP uses prompts as the supervision, where the visual labels are entered into the hand-crafted template. By pre-training the model at a large scale, models can learn the visual contents and easily be transferred to downstream tasks through the prompt-based zero-shot transfer. + +However, the manual design of prompts can be a non-trivial and time-consuming task. [54] found that even a slight change in the prompt (e.g., one word) can make a big difference. They introduced Context Optimization (CoOp) to automate prompt engineering to generate continuous soft prompts instead of using hand-chosen hard prompts [54]. However, CoOp requires substantial computing resources, and the results of CoOp are not interpretable. Furthermore, such prompt tuning faces performance degradation when there is a significant domain shift [39], e.g., from natural images to pathology images, rendering it less adaptable to medical imaging tasks. On the other hand, medical image analysis may involve multiple learning tasks simultaneously. Conversely, in medical im + +age analysis, multiple learning tasks may be concurrently required. It becomes impractical if fine-tuning necessitates substantial computational resources. For instance, Lai et al. [17] quantify and visualize distinct pathologies in grey/white matter (GM/WM) regions within gigapixel images. To accomplish this, the task is divided into two learning objectives: pathology detection and GM/WM segmentation. They collected and annotated two datasets and then trained two models from scratch to achieve it. Therefore, it will be beneficial if we can propose an effective way to adapt a big model into multiple downstream tasks quickly so the final analysis objective can be reached in a reasonable time manner. + +In this work, we aim to fine-tune CLIP efficiently with light computing resources for pathology image analysis tasks. There are several challenges in fine-tuning CLIP. First, overfitting is a severe issue if we adapt CLIP to the downstream tasks directly since CLIP is pre-trained on a 400M dataset while the new domain dataset can be small [10]. Second, CLIP has shown its promising zero-shot generalization ability, directly fine-tuning CLIP may lose this capability and face the "catastrophic forgetting" issue. In other words, it is unclear how to effectively learn the new knowledge while retaining the original pre-trained knowledge so that the generalization of CLIP can be maintained. Third, there are limited studies on benchmarking CLIP's transferability in the pathology domain; hence, its applicability remains unclear. To address the above issues, we first study the applicability of CLIP on two public pathology datasets and benchmark the zero-shot ability of CLIP on them. Then, we did empirical experiments to verify the importance of the CLIP's pre-trained knowledge. To address the "catastrophic forgetting" issue, we propose Residual Feature Refinement (RFR) as a lightweight approach for adapting CLIP to pathology images. As a shallow network, RFR demonstrates proficiency in extracting crucial spatial features and patterns from pathology images. Within RFR, we introduce a self-adaptive module designed to seamlessly fuse and balance the original knowledge from CLIP with the newly acquired insights from the specific pathological task, requiring only a minimal number of additional trainable weights, as opposed to retraining the entire encoders in CLIP. To further mitigate overfitting, we also incorporate Hidden Representation Perturbation (HRP) and Dual-view Vision Contrastive (DVC). Finally, we establish an end-to-end adaptation of CLIP, named Path-CLIP (as shown in Fig. 1(b)), tailored for pathology image analysis tasks. Path-CLIP incorporates a newly proposed loss function, Doublet Multimodal Contrastive Loss (DMCL). + +We summarize our contributions as follows: + +- We explore the applicability of CLIP on pathology images and highlight the catastrophic forgetting issues when adapting CLIP to downstream pathology tasks. +- We propose Path-CLIP to alleviate the overfitting and forgetting issues and effectively adapt CLIP for pathology tasks. + +- Extensive experiments demonstrate the efficiency and robustness of PathCLIP. For example, it achieves $+19\%$ on PCam when only using $0.1\%$ of labeled data. + +# 2 Related Work + +Language-vision model (LVM). LVMs have shown promising performance in learning generic visual representations [15, 22, 26, 48]. Recent models advance via text representation learning with large-scale Transformers [30] and even web-scale training datasets [54]. Transformer-based multimodal learning has achieved great success in such large-scale datasets [23, 29, 36]. For example, CLIP [26] was trained on 400 million image-caption pairs and achieved state-of-the-art results in many fields [26, 44, 46, 47]. In CLIP's architecture, it has two encoders: the vision encoder can be ResNet [12] or ViT [7] while the text encoder can be Transformer (e.g., BERT [6]). A recent work [27] fine-tuned CLIP for video data and performed competitively to more complicated methods designed for video processing. PointCLIP [50] applied CLIP to 3D recognition. CLIP is also applied for image generation tasks [8] and shows its power to reduce required data collection. However, the best way to adapt CLIP for downstream tasks is still being studied, especially when the new domain is far from the pre-trained domains, e.g., the medical domain. + +CLIP Fine-tuning and adaptation. Although recent multimodal models (e.g., CLIP [26]) perform astonishingly well in standard vision tasks without fine-tuning, the gap between CLIP and supervised learning is still noticeable. Therefore, many works focus on fine-tuning CLIP with a few samples from the new task provided. For example, CoOp [54] kept the entire pre-trained parameters fixed but optimized the continuous prompt through a learnable vector. This method closes the gap between CLIP and supervised learning but the performance gain is limited as it only focuses on the perspective of prompt design instead of the vision part for the image recognition tasks. On the other hand, CLIP-Adapter [9] and Tip-Adapter [51] freeze the pre-trained weights and train additive module, which significantly improves the CLIP's performance with limited training data in the new task. These methods require a lower computational cost. However, both CLIP-Adapter [9] and Tip-Adapter [51] involve tuning a ratio (hyper-parameter) to balance the original and fine-tuned knowledge manually, which may not be practical when we have multiple target learning tasks in the medical domain. + +# 3 Methodology + +In this section, we introduce the proposed Path-CLIP. We summarize the end-to-end pipeline in Fig. 1(b): a self-adaptive residual ratio $\alpha$ will be incorporated into Residual Feature Refinement (Section 3.2) to fuse and balance the original knowledge with the task-specific knowledge from the new downstream task. We introduce Hidden Representation Perturbation (HRP) and Dual-view Vision + +![](images/fa6cef420246b5d56be80fa190e37fe1d539dca58ddea36076b5ca2aaeb916b5.jpg) +Fig. 2: Overview of the Residual Feature Refinement (RFR): a downsampling-upsampling shallow architecture of fully connected layers. RFR blends the fine-tuned knowledge with the original knowledge from CLIP's vision encoder $(F(\cdot))$ via a self-adaptive residual ratio $\alpha$ . HRP refers to Hidden Representation Perturbation and TFT refers to pre-training fine-tuning. We use blue color to denote the original visual features and orange color to denote the fine-tuned features. + +Contrastive (Section 3.3) to alleviate the overfitting issue. Finally, we propose Doublet Multimodal Contrastive Loss (DMCL) in Section 3.4 as the loss function for fine-tuning. + +# 3.1 Language-Vision Pre-training + +We first revisit language-vision pre-training and use CLIP [26] as one example. Note that our framework is applicable to broader language-vision models. CLIP [26] has a vision encoder $F(\cdot)$ and a text encoder $G(\cdot)$ . The vision encoder maps a high-dimensional image into low-dimensional image embeddings. The vision encoder can be a CNN like ResNet [12] or a Vision Transformer (ViT) [7]. The text encoder is built on Transformer [30] and generates text embeddings from the prompt. Specifically, "prompt" here refers to a sequence of words (tokens), such as "a neuropathological image of grey matter". + +Training. CLIP jointly trains $F(\cdot)$ and $G(\cdot)$ to optimize the similarity score (e.g., symmetric cross-entropy loss [33]) between the visual and textual embeddings for each batch. The input consists of an image and its corresponding prompt (e.g., "this is a photo of grey matter"). They formulate the learning objective as a contrastive loss to align the visual and textual embedding spaces. Given a batch of image-prompt pairs, CLIP maximizes the similarity score for positive pairs while minimizing it for negative pairs. + +Contrastive loss function. In CLIP training, the agreement between a positive pair of vision and text embeddings is maximized by InfoNCE loss [24]: + +$$ +l _ {i} ^ {v \rightarrow t} = - \log \frac {\exp \left(\mathbf {s i m} \left(v _ {i} , t _ {i}\right) / \gamma\right)}{\sum_ {k = 1} ^ {B} \exp \left(\mathbf {s i m} \left(v _ {i} , t _ {k}\right) / \gamma\right)}, \tag {1} +$$ + +where $v, t$ are normalized vectors from vision and text encoders. $(v_{i}, t_{i})$ is a positive pair. $\mathbf{sim}$ is a function to measure the similarity of two vectors. $B$ is the mini-batch size of the image-text pairs and $\gamma$ is the learnable temperature. Therefore, there is only 1 positive text for the input $i^{th}$ image but $N - 1$ negative texts. $l_{i}^{v \rightarrow t}$ means the InfoNCE loss from an image $i$ to texts, so as the text loss $l_{i}^{t \rightarrow v}$ . Then the loss function for CLIP is as follows: + +$$ +\mathcal {L} _ {\mathrm {C L I P}} = \frac {1}{2 B} \sum_ {i = 1} ^ {B} \left(l _ {i} ^ {v \rightarrow t} + l _ {i} ^ {t \rightarrow v}\right). \tag {2} +$$ + +Inference for classification tasks. For the inference process, an image $x$ is transformed into a feature manifold $f \in \mathbb{R}^D$ , where $D$ is the feature dimension. Then, $f$ is multiplied with a classifier weight matrix $\mathbf{W} \in \mathbb{R}^{D \times K}$ , where $K$ is the number of classes in the learning task. We get a $K$ -dimensional logit after matrix multiplication. Then we apply softmax to convert this logit into a probability vector $p \in \mathbb{R}^K$ over the $K$ classes. The whole process can be summarized as the following equation: + +$$ +p _ {i} = \frac {\exp \left(\mathbf {W} _ {\mathbf {i}} ^ {T} \times f\right) / \gamma}{\sum_ {i = 1} ^ {K} \exp \left(\mathbf {W} _ {\mathbf {i}} ^ {T} \times f\right) / \gamma}, \tag {3} +$$ + +where $\gamma$ is a temperature parameter learned by CLIP during training and $\mathbf{W_i}$ is the prototype weight vector for class- $i$ . + +Transferability. Compared to traditional classifier learning methods where the supervision is a fixed one-hot vector, language-vision pre-training utilizes an open-set visual concept that can be explored through the text encoder to learn a broader semantic space, which in turn makes the learned representations more diverse and adaptable to various downstream tasks (e.g., zero-shot learning). Therefore, we argue that such language-vision pre-training is helpful for medical problems where the domain is very distinctive from the natural images. + +# 3.2 Residual Feature Refinement + +In this subsection, we introduce Residual Feature Refinement (RFR) to learn the task-relevant context when we adapt CLIP to the downstream learning tasks. Unlike CoOp's prompt tuning, which may not address the domain shift issue between the natural and pathology images, we focus on fine-tuning the visual features $f$ . However, simple fine-tuning of the entire network may fail in a new pathological task due to 1) the overfitting issue caused by a large number of parameters in CLIP and a shortage of training samples [13]; 2) catastrophic forgetting issue. + +Table 1: Baseline. We fine-tune CLIP (ResNet-50) on PCam [31] with $1\%$ training data. "No fine-tuning" is the best baseline due to the catastrophic forgetting issue. + +
# ConfigurationAccuracy
1No fine-tuning56.5%
2Fine-tune 10 epochs40.3%
3Fine-tune 100 epochs35.9%
4With a lower learning rate on #345.6%
+ +Importance of the CLIP's pre-trained knowledge. CLIP has shown its zero-shot ability and transferability to a wide range of downstream tasks. We argue that preserving pre-trained knowledge of CLIP is crucial to maintain its generalization and alleviate the overfitting issue when the new dataset is small. There are two reasons behind this. First, one recent NLP work [11] shows large pre-trained language models (PLMs) are well-calibrated on the masked language modeling tasks. However, such property cannot be retained through direct fine-tuning due to catastrophic forgetting. Specifically, the pre-trained features are only used for the initialization and will be distorted by the fine-tuning [11]. Inspired by this, we investigate whether this issue also appears in language-vision models. As shown in Table 1, "no fine-tuning" but zero-shot gets the best results compared to fine-tuning CLIP on $1\%$ of training data in PCam [31]. We also attempt to lower the learning rate to $1 \times 10^{-6}$ , but the results are still far from satisfactory. Therefore, we conclude direct fine-tuning may distort the pre-trained knowledge and result in a severe catastrophic forgetting issue. Besides the forgetting issue, overfitting is another common challenge when adapting a big model to a new task. To alleviate the above two issues, we propose a residual refinement architecture to preserve the pre-trained features and acquire rich semantic information for pathology image analysis. + +Architecture overview. As shown in Fig. 2, given an image $x$ , we get the visual feature $f$ from the image encoder and compute the classifier weight $\mathbf{W}$ from the text encoder. Then we design trainable adaptive layers $L$ to convert $f$ into $L(f)$ . $L$ can be multiple layers of linear transformations in a "down-and-up" architecture. As shown in Fig. 2, we have four layers to transform the feature dimension as "1024-256-64-256-1024". This is one example: the layers are flexible with other dimensions. After this architecture, $L(f)$ can be of the same size of $f$ and blended with $f$ as follows: + +$$ +f _ {i} ^ {*} = \left(1 - \alpha_ {i}\right) L \left(f _ {i}\right) + \alpha_ {i} f _ {i}, \tag {4} +$$ + +where $i = \{1,2,3,\dots n\}$ and $\alpha$ is the self-adaptive residual ratio to balance the fine-tuned knowledge and the original CLIP's knowledge. Specifically, the first term adaptively learns knowledge from the new dataset, while the second term preserves the original knowledge from CLIP. + +Then we adopt Eq. (3) with the new $f^*$ to get the class probability vector and predict the category. During fine-tuning, the weights of the trainable layers are optimized through the symmetric contrastive loss used in CLIP [26]. The prompt can be "this is a photo of [], where "+" is filled with the class name. We can incorporate pathological descriptions into the prompt to provide more domain information, e.g., "this is a pathological breast photo of []". + +Hidden Representation Perturbation (HRP). Fine-tuning a pre-trained language-vision model on a small dataset can lead to potential issues such as overfitting or representation collapse [45]. To mitigate these issues, we propose Hidden Representation Perturbation (HRP) that is strategically integrated with the trainable layers. To balance the model's complexity and efficiency, we adopt a "four-layer" architecture with HRP, steering clear of deeper architectures to ensure a controlled number of trainable parameters. This precaution is crucial as an excess of trainable parameters may exacerbate overfitting, especially in scenarios characterized by limited resources, where only a sparse dataset is available for the new task [52]. + +To alleviate the overfitting issue in low-resource settings, we focus on adding noise during the training process, a widely-used NLP strategy for large language models to smoothen the optimization landscape. For example, NoisyTune [35] adds the perturbations to the pre-trained weights to improve the model's generalization. Inspired by these NLP strategies [19, 35, 45], we propose to add random noise to the feature $f$ before trainable adapting layers fine-tuning it. The random noise can be Gaussian noise $\varepsilon$ generated from $\mathcal{N}(0, \sigma^2)$ . After this, the perturbed feature $f$ is converted into $\hat{f} = f + \varepsilon$ . Therefore, Eq. (4) is converted into: + +$$ +f _ {i} ^ {*} = \left(1 - \alpha_ {i}\right) L \left(\hat {f} _ {i}\right) + \alpha_ {i} f _ {i}. \tag {5} +$$ + +Therefore, the residual feature will be perturbed with the Gaussian noise for the refinement, designed to alleviate the overfitting issue. + +# 3.3 Self-adaptive residual ratio + +In Section 3.2, we introduce residual feature refinement and the residual ratio $\alpha$ . $\alpha$ helps modify the degree of preserving the original knowledge for better performance. Empirically, $\alpha$ should be small if the domain gap between the source data and the downstream task is enormous, so more knowledge from the downstream task $(L(f))$ can be considered. However, if $\alpha$ is too small, the model can be overfitting as it will heavily rely on the fine-tuned features and disregard the pre-trained knowledge from CLIP. On the other hand, the model can also be underfitting if $\alpha$ is too large, as it may prevent learning from the fine-tuned features. Therefore, it is crucial to quantify the residual ratio $\alpha$ so the model can be well-trained instead of underfitting or overfitting. + +Recent works such as Tip-Adapter [51] and CLIP-Adapter [9] involve feature refinement while they set such ratio as a fixed hyper-parameter that needs to be + +tuned during the training process. We argue the residual ratio should be dynamic and self-adaptive during training instead of fixed, as the learning status changes along the training process. Hence, we propose a self-adaptive residual ratio in Path-CLIP. + +DVC: Dual-view Vision Contrastive. Inspired by self-supervised learning (e.g., SimCLR [4]), the composition of data augmentations is relevant to define effective predictive tasks. Therefore, we propose Dual-view Vision Contrastive (DVC) to measure the model's learning status. Given an image $x$ , we leverage two types of augmentations: "weak" and "strong" [28]. "weak" refers to the standard flip-and-shift augmentation strategy while "strong" refers to perturb the appearance (RandAugment [5] and CTAugment [2]). We propose to measure the distance between representations from both augmentations, e.g., cosine similarity has shown good performance in both visual contrastive [4] and textual contrastive [3] tasks. + +With the input $x$ , we will get $w(x)$ from the weak augmentation and $s(x)$ from the strong augmentation. We obtain $F(w(x))$ and $F(s(x))$ after the vision encoder $F$ . We define the distance as $\tau$ : + +$$ +\tau = \frac {1}{N} \sum_ {i = 1} ^ {N} \cos \left(F \left(w \left(x _ {i}\right)\right), F \left(s \left(x _ {i}\right)\right)\right), \tag {6} +$$ + +where $N$ is the number of samples in the training set, and $\cos$ is a cosine similarity. The model's learning status is relevant to its consistency across different perturbations. We traverse the entire training set and calculate its mean value to represent the model's performance. + +Bound constraints. In addition to DVC which measures the predictive consistency of the model, we also select "accuracy" on the training set as a second metric to measure the learning status, denoted as $T$ . Empirically, "accuracy" alone cannot fully tell if the model is well-fitted: 100% of accuracy may indicate overfitting, which is the biggest challenge when we adapt a large model to a downstream task where the training set is very small. However, "accuracy" is a good metric to show if the model is underfitting. Therefore, we propose to fuse DVC and "accuracy" to represent the model's status during training. + +If we look at $\alpha$ and set it fixed, we can find the model is prone to overfit if $\alpha$ is too small as it will heavily rely on the fine-tuned knowledge; the model tends to underfit if $\alpha$ is too large as it prevents the model from learning new knowledge (directly use the original pre-trained knowledge). Therefore, we propose a dynamic way with lower and upper bounds to obtain $\alpha$ during training: + +$$ +\alpha = \min (\max (\tau , T), 1 - \tau). \tag {7} +$$ + +The bounds are associated with predictive consistency (DVC) and predictive accuracy on the training set. For example, if the model is overfitting, $T$ will be large, and $\tau$ will be relatively small. Then $\alpha$ will be smaller in the next epoch to use less fine-tuned features. Therefore, our proposed bound constraints can effectively adjust $\alpha$ to alleviate the overfitting issue. + +![](images/c2e7f96d3618a13c1d9605bfa1ac5f219d04fc819c37596898e29877a1210be4.jpg) +Fig. 3: Doublet Multimodal Contrastive Loss (DMCL) that utilizes the traditional image-text pair and the proposed image-image pair with weak/strong augmentation criteria. + +# 3.4 Path-CLIP: End-to-end adaptation of CLIP to pathology tasks + +To efficiently adapt CLIP to pathology image tasks, we propose RFR with a self-adaptive residual ratio to balance the original and fine-tuned knowledge based on the learning status. To alleviate the overfitting issue, we incorporate HRP and design bound constraints for the residual ratio. Besides these, to distill more visual representations from the new image domain, we add a contrastive loss to the loss function so that the trainable portion has a better classifying ability. We follow SimCLR [4] to define the Contrastive Loss (CL) for one image in the batch as: + +$$ +l _ {i} ^ {c} = \left\{ \begin{array}{l} - \log \frac {\exp (\mathbf {s i m} (v _ {i} , v _ {i + B}) / \gamma)}{\sum_ {k \neq i} ^ {2 B} \exp (\mathbf {s i m} (v _ {i} , v _ {k}) / \gamma)}, \text {i f} i < B, \\ - \log \frac {\exp (\mathbf {s i m} (v _ {i} , v _ {B - i}) / \gamma)}{\sum_ {k \neq i} ^ {2 B} \exp (\mathbf {s i m} (v _ {i} , v _ {k}) / \gamma)}, \text {o t h e r w i s e .} \end{array} \right. \tag {8} +$$ + +Differing from SimCLR [4] and RepRec [38], as shown in Fig. 3, here we set image $i$ as the weakly augmented image and image $i + B$ as the strongly augmented image instead of two randomly augmented images as the positive pair. We named this loss as $\mathcal{L}_{\mathrm{SWAPP}}$ . SWAPP here refers to Strong-Weak Augmented Positive Pair. Finally, the total loss for fine-tuning CLIP on our pathology datasets is as follows: + +$$ +\mathcal {L} _ {\mathrm {t o t a l}} = \mathcal {L} _ {\mathrm {C L I P}} + \mathcal {L} _ {\mathrm {S W A P P}}. +$$ + +$$ +\mathcal {L} _ {\mathrm {C L I P}} \rightarrow \frac {1}{2 B} \sum_ {i = 1} ^ {B} \left(l _ {i} ^ {v \rightarrow t} + l _ {i} ^ {t \rightarrow v}\right), \tag {9} +$$ + +$$ +\mathcal {L} _ {\mathrm {S W A P P}} \rightarrow \frac {1}{2 B} \sum_ {i = 1} ^ {2 B} l _ {i} ^ {c} +$$ + +We define this loss as Doublet Multimodal Contrastive Loss (DMCL) because the items involve both intra-contrastive (image and text) and inter-contrastive (image and image). The algorithm can be found in the Appendix. + +Table 2: Quantitative comparison on PCam. Precision, Recall, and F1-score refer to the macro-averaged values from all classes + +
Learning MannerLabeled RatioAlgorithmAccuracyPrecisionRecallF1-scoreAUC
Supervised100%-92.8 ± 0.3092.9 ± 0.2092.6 ± 0.2192.8 ± 0.200.95 ± 0.01
Supervised0.5%-45.1 ± 2.2742.0 ± 2.1433.6 ± 3.3037.3 ± 2.860.51 ± 0.01
Self-supervised0.5%SimCLR [4]60.4 ± 1.4563.3 ± 1.9058.5 ± 2.0160.8 ± 1.200.61 ± 0.04
Semi-supervised0.5%Pseudo-Label [20]55.7 ± 2.0559.2 ± 2.2354.5 ± 1.9756.8 ± 1.550.58 ± 0.03
FixMatch [28]73.2 ± 0.7677.5 ± 0.5073.7 ± 0.2575.6 ± 0.370.84 ± 0.04
Dash [37]71.0 ± 0.9875.3 ± 0.7870.2 ± 0.4572.7 ± 0.500.82 ± 0.02
FlexMatch [49]74.0 ± 0.8078.1 ± 1.1573.0 ± 0.9075.5 ± 0.840.85 ± 0.03
Multimodal CLIP0.5%Zero-shot [26]56.5 ± 0.0057.4 ± 0.0050.3 ± 0.0053.7 ± 0.000.60 ± 0.00
CoOp [54]63.6 ± 0.2563.9 ± 0.4262.5 ± 0.3563.0 ± 0.290.67 ± 0.02
CLIP-Adapter [9]72.3 ± 1.0277.2 ± 0.9563.2 ± 0.5669.4 ± 0.840.81 ± 0.02
TIP-Adapter [51]73.9 ± 0.5070.5 ± 0.8578.0 ± 0.4974.0 ± 0.950.84 ± 0.02
Path-CLIP81.9 ± 0.7879.4 ± 0.3585.0 ± 0.8082.1 ± 0.950.89 ± 0.02
Supervised1.5%-62.8 ± 1.3169.5 ± 0.6754.1 ± 2.2760.8 ± 1.680.63 ± 0.02
Self-supervised1.5%SimCLR [4]64.9 ± 1.0265.2 ± 1.4560.1 ± 1.7062.5 ± 1.050.64 ± 0.03
Semi-supervised1.5%Pseudo-Label [20]67.9 ± 1.8570.5 ± 1.9062.2 ± 1.7666.1 ± 1.420.67 ± 0.04
FixMatch [28]83.9 ± 1.2985.1 ± 0.8584.1 ± 0.9584.6 ± 0.900.87 ± 0.01
Dash [37]82.8 ± 1.0983.0 ± 0.5581.2 ± 0.6082.1 ± 0.780.85 ± 0.03
FlexMatch [49]84.1 ± 1.0284.9 ± 0.7584.4 ± 0.8084.6 ± 0.750.86 ± 0.04
Multimodal CLIP1.5%CoOp [54]61.5 ± 0.7560.5 ± 1.0365.6 ± 0.9063.1 ± 0.850.65 ± 0.02
CLIP-Adapter [9]74.1 ± 0.8879.1 ± 0.3465.4 ± 0.8971.6 ± 1.020.82 ± 0.01
TIP-Adapter [51]75.2 ± 0.9472.3 ± 0.6979.4 ± 0.8675.7 ± 0.730.83 ± 0.02
Path-CLIP83.1 ± 0.6782.8 ± 0.9582.0 ± 0.7582.4 ± 0.460.91 ± 0.02
+ +# 4 Experiments + +Datasets. We select two classification datasets (PCam [31] and MHIST [34]) and a gigapixel grey/white matter segmentation dataset to evaluate Path-CLIP. The details of these datasets can be found in Appendix. + +Experimental setup. For a fair comparison, we use ResNet-50 [12] as the backbone in the vision encoder $f(\cdot)$ and BERT [6] as the textual encoder. We follow CLIP [26] to use gradient scaling to facilitate mixed-precision training. We set the learning rate as 0.0001. The batch size is 32. We use Adam [16] as the optimizer. To compare the computational complexity, we conduct all experiments on a single GPU (Nvidia RTX 2080Ti). + +Baselines. We first implement traditional supervised learning (SL) methods with the entire dataset to explore the upper-bound performance for each task. Then we reduce the data usage and investigate the performance change of these SL methods. Then we compare Path-CLIP with original CLIP [26] and three recent CLIP-based fine-tuning methods: CoOp [54], Tip-Adapter [51] and CLIP-Adapter [9]. We compare them under the limited data scenarios to test their ability to do downstream tasks when the new dataset is small. Besides that, we also incorporate recent self-supervised learning (SimCLR [4]) and semi-supervised learning (Pseudo-Label [20], FixMatch [28], Dash [37], and FlexMatch [49]). Semi-supervised training requires a large amount of unlabeled data, which involves data collection efforts [41-43]. + +# 4.1 Main results + +Classification task: PCam [31] and MHIST [34]. The main results for two classification benchmarks are summarized in Table 2 and Table 3. First, CLIP + +Table 3: Quantitative comparison on MHIST. Precision, Recall, and F1-score refer to the macro-averaged values from all classes + +
Learning MannerLabeled RatioAlgorithmAccuracyPrecisionRecallF1-scoreAUC
Supervised100%-86.0 ± 0.3684.8 ± 0.6183.5 ± 0.5584.1 ± 0.580.89 ± 0.01
Self-supervised10%SimCLR [4]63.2 ± 0.9563.8 ± 1.2062.2 ± 1.7063.0 ± 0.850.66 ± 0.04
Semi-supervised10%Pseudo-Label [20]65.4 ± 1.0466.5 ± 0.9865.9 ± 0.8566.2 ± 1.420.75 ± 0.03
FixMatch [28]70.5 ± 1.5568.8 ± 2.7669.5 ± 2.9769.1 ± 2.860.79 ± 0.02
Dash [37]70.2 ± 1.2170.0 ± 0.8571.3 ± 0.6970.6 ± 1.480.81 ± 0.04
FlexMatch [49]73.1 ± 0.8474.6 ± 1.0575.1 ± 0.9074.8 ± 0.890.82 ± 0.03
Multimodal CLIP10%Zero-shot [26]36.9 ± 0.0036.9 ± 0.00100 ± 0.0053.9 ± 0.000.50 ± 0.00
CoOp [54]68.9 ± 0.4056.0 ± 0.6572.8 ± 0.5363.3 ± 0.860.76 ± 0.01
CLIP-Adapter [9]65.9 ± 1.4552.2 ± 1.2387.2 ± 0.8565.3 ± 0.950.77 ± 0.02
TIP-Adapter [51]63.1 ± 0.7849.9 ± 0.9094.7 ± 1.0265.4 ± 0.560.78 ± 0.03
Path-CLIP74.8 ± 1.9563.3 ± 1.4075.6 ± 0.9868.9 ± 1.500.84 ± 0.02
+ +shows its promising zero-shot ability: when we directly test CLIP on PCam [31], it can get $56.5\%$ accuracy. Then, we compare Path-CLIP with other CLIP-based fine-tuning methods. In PCam [31], we present $0.5\%$ and $1.5\%$ data usage in the training set to fine-tune CLIP, respectively. We find all of these fine-tuning algorithms [9, 51, 54] can improve the performance of the original CLIP with the new data from the downstream task presented. Path-CLIP can outperform them in both settings. Specifically, Path-CLIP outperforms TIP-Adapter [51] with $8\%$ and $7.9\%$ accuracy when only $0.5\%$ and $1.5\%$ data are used, respectively. Path-CLIP also shows similar improvement in MHIST [34] (Table 3). We also implemented several semi-supervised learning where the rest of the data in the training set are used as an unlabeled set. Although Path-CLIP does not use unlabeled data, it can achieve competitive results with semi-supervised learning. + +Besides these ratios of data usage, we also visualize the results by varying the ratio to show the performance of each fine-tuning algorithm. As shown in Fig. 4, CoOp [54] may not improve PCam when more data are used to fine-tune the prompt. CoOp freezes the encoders and only fine-tunes the language prompt part, while other algorithms also fine-tune the vision part. We conclude that visual fine-tuning is more critical compared to prompt tuning. + +![](images/859cb8c799e940407e8f279e525652182d3daf3db9a0cbd9f05860f9ee502eaf.jpg) +(a) PCam + +![](images/47c058cc12f1284798227d7cbeb7ace6e0a56b57a4298d346f97b41d9c3cbc07.jpg) +(b) MHIST +Fig. 4: Comparison of different fine-tuning methods by varying data usage. + +Table 4: Performance comparison on GM/WM segmentation. + +
Data UsageAlgorithmGM_IoUGM_DICEWM_IoUWM_DICE
100% (SL)FCN [1]71.7283.2149.0966.08
U-Net [25]91.5295.1782.0288.15
0.1% (SL)FCN [1]48.2525.0124.5430.02
U-Net [25]46.1350.105.7512.23
0.1% (CLIP)CoOp [54]62.2365.7039.9248.65
CLIP-Adapter [9]67.8074.3243.2554.61
TIP-Adapter [51]69.9275.5046.2356.60
Path-CLIP78.0583.7054.5063.23
1% (SL)FCN [1]51.3430.2031.0235.12
U-Net [25]50.2453.2315.4325.58
1% (CLIP)CoOp [54]63.4564.1237.8545.34
CLIP-Adapter [9]69.9976.5447.9558.34
TIP-Adapter [51]72.1277.3055.6563.60
Path-CLIP80.2188.7060.5070.34
+ +GM/WM segmentation [17]. Besides two classification tasks, we follow [17] to generalize our method to a weakly-supervised segmentation problem. We select IoU and DICE scores as the performance metrics. The results are summarized in Table 4. We select FCN [1] and U-Net [25] as two SL baselines: we first train them on all data (100%) in a supervised manner, then we reduce the dataset to 0.1% and 1% setting and train the model again. We can see both SL algorithms face performance degradation when the training data is scarce. They require an extensive and well-curated labeled dataset to get satisfactory results. We then focus on fine-tuning CLIP algorithms. Table 4 shows Path-CLIP can achieve superior results than other fine-tuning techniques. + +# 4.2 Ablation studies + +Importance of alleviating overfitting. Since the new task may have a very limited training set, it is crucial to alleviate the potential overfitting issue. Besides the RFR with the self-adaptive ratio, we also propose HRP, DVC, and DMCL to cope with the overfitting issue. Table 5 summarizes the contribution of each proposed module: HRP can improve with $+2.2\%$ accuracy and DVC can improve with $+4.3\%$ . DMCL inspired by self-supervised learning can also aid in great improvements. We conclude that contrastive learning, perturbations on data and hidden representations are helpful to the overfitting issue. + +Computational efficiency. When we range the data usage from $0.1\%$ to $1\%$ , Path-CLIP can obtain $+5.5\%$ gain in accuracy while only introducing 1 minute of additional training time. However, CoOp [54] has the overfitting issue and obtains a lower score while using 53 minutes, almost five times compared to the time used by RFR. The situation remains similar under the $10\%$ of data usage. We conclude our proposed RFR can get over $19\%$ improvement on CLIP by only + +Table 5: Ablation study on PCam using $0.5\%$ data for fine-tuning. + +
#RFCHRPDVCDMCLAccuracyF1-scoreAUC
1XXXX56.553.70.60
2✓XXX72.070.50.69
3✓✓XX74.275.50.78
4✓✓✓X78.579.30.84
5✓✓✓✓81.982.10.89
+ +using 10 minutes for the fine-tuning on a single GPU, which is promising for digital pathology research. The details can be found in Table 8 in the Appendix. + +Generalization to pathology foundation model: PLIP [14]. The results of using PLIP's weights are in Appendix 3.6: PLIP is pre-trained on a large-scale pathology dataset and achieves strong zero-shot performance. With our method, we can further boost the performance and outperform other tuning methods. + +# 5 Discussion + +In this work, we explore the generalization of CLIP in pathology image analysis. To alleviate the overfitting and forgetting issues when adapting CLIP to small downstream datasets, we propose self-adaptive RFR with HRP. We propose Path-CLIP with DMCL as an end-to-end pipeline for CLIP adaptation in pathology tasks. The results reveal the superiority and effectiveness of Path-CLIP compared to the existing fine-tuning methods. In addition, the less computational time improves the efficiency and accessibility of Path-CLIP. Hence, we conclude Path-CLIP is more robust and has the potential to bridge the domain shift between the pre-trained natural images and pathology images in an efficient and cost-effective manner. + +# Acknowledgments + +This work was supported by the Noyce Initiative UC Partnerships in Computational Transformation Grant and Child Family Endowed Professorship. Resources for this study were funded in part by grants from the National Institute on Aging of the National Institutes of Health under Award Numbers R01AG062517, P30AG072972, and R01AG056519. + +# References + +1. Bandy, P., van de Loo, R., Intezar, M., Geijs, D., Ciompi, F., van Ginneken, B., van der Laak, J., Litjens, G.: Comparison of different methods for tissue segmentation in histopathological whole-slide images. In: Proc. 2017 IEEE Int. Symp. Biomed. Imaging. pp. 591-595 (2017) +2. Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., Raffel, C.A.: MixMatch: A holistic approach to semi-supervised learning. In: NeurIPS. vol. 32 (2019) + +3. Chen, F., Zhang, H., Li, Z., Dou, J., Mo, S., Chen, H., Zhang, Y., Ahmed, U., Zhu, C., Savvides, M.: Unitail: Detecting, reading, and matching in retail scene. In: ECCV. pp. 705-722. Springer (2022) +4. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML. pp. 1597-1607 (2020) +5. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Workshop. pp. 702-703 (2020) +6. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) +7. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) +8. Gal, R., Patashnik, O., Maron, H., Bermano, A.H., Chechik, G., Cohen-Or, D.: Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG) 41(4), 1-13 (2022) +9. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544 (2021) +0. Ge, C., Huang, R., Xie, M., Lai, Z., Song, S., Li, S., Huang, G.: Domain adaptation via prompt learning. arXiv preprint arXiv:2202.06687 (2022) +1. He, G., Chen, J., Zhu, J.: Preserving pre-trained features helps calibrate fine-tuned language models. In: ICLR (2023) +2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016) +3. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: ICML. pp. 2790-2799. PMLR (2019) +4. Huang, Z., other: A visual-language foundation model for pathology image analysis using medical twitter. Nature medicine (2023) +5. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML. pp. 4904-4916 (2021) +6. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (Poster) (2015) +7. Lai, Z., Oliveira, L.C., Guo, R., Xu, W., Hu, Z., Mifflin, K., Decarli, C., Cheung, S.C., Chuah, C.N., Dugger, B.N.: Brainsec: Automated brain tissue segmentation pipeline for scalable neuropathological analysis. IEEE Access 10, 49064-49079 (2022) +8. Lai, Z., Wang, C., Oliveira, L.C., Dugger, B.N., Cheung, S.C., Chuah, C.N.: Joint semi-supervised and active learning for segmentation of gigapixel pathology images with cost-effective labeling. In: ICCV Workshop. pp. 591-600 (2021) +9. Lam, S.Y., Zeng, Q., Zhang, K., You, C., Voigt, R.: Large language models are partially primed in pronoun interpretation. arXiv preprint arXiv:2305.16917 (2023) +20. Lee, D.H.: Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In: Workshop on challenges in representation learning (ICML) (2013) +21. Liu, M., Hu, L., Tang, Y., Wang, C., He, Y., Zeng, C., Lin, K., He, Z., Huo, W.: A deep learning method for breast cancer classification in the pathology images. IEEE J. Biomed. Health. Inf. 26(10), 5025-5032 (2022) + +22. Lu, Y., Liu, J., Zhang, Y., Liu, Y., Tian, X.: Prompt distribution learning. In: CVPR. pp. 5206-5215 (2022) +23. Lyu, W., Dong, X., Wong, R., Zheng, S., Abell-Hart, K., Wang, F., Chen, C.: A multimodal transformer: Fusing clinical notes with structured ehr data for interpretable in-hospital mortality prediction. In: AMIA Annual Symposium Proceedings. vol. 2022, p. 719. American Medical Informatics Association (2022) +24. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) +25. Oskal, K.R., Risdal, M., Janssen, E.A., Undersrud, E.S., Gulsrud, T.O.: A U-net based approach to epidermal tissue segmentation in whole slide histopathological images. SN Appl. Sci. 1(7), 1-12 (2019) +26. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021) +27. Rasheed, H., khattak, M.U., Maaz, M., Khan, S., Khan, F.S.: Finetuned clip models are efficient video learners. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) +28. Sohn, K., Berthelot, D., Carlini, N., Zhang, Z., Zhang, H., Raffel, C.A., Cubuk, E.D., Kurakin, A., Li, C.L.: FixMatch: Simplifying semi-supervised learning with consistency and confidence. In: NeurIPS. vol. 33, pp. 596-608 (2020) +29. Tian, Y., Newsam, S., Boakye, K.: Fashion image retrieval with text feedback by additive attention compositional learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1011-1021 (2023) +30. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017) +31. Veeling, B.S., Linmans, J., Winkens, J., Cohen, T., Welling, M.: Rotation equivariant cnns for digital pathology. In: MICCAI. pp. 210-218. Springer (2018) +32. Wang, X., Yang, S., Zhang, J., Wang, M., Zhang, J., Huang, J., Yang, W., Han, X.: Transpath: Transformer-based self-supervised learning for histopathological image classification. In: MICCAI. pp. 186-195. Springer (2021) +33. Wang, Y., Ma, X., Chen, Z., Luo, Y., Yi, J., Bailey, J.: Symmetric cross entropy for robust learning with noisy labels. In: ICCV (2019) +34. Wei, J., Suriawinata, A., Ren, B., Liu, X., Lisovsky, M., Vaickus, L., Brown, C., Baker, M., Tomita, N., Torresani, L., et al.: A petri dish for histopathology image analysis. In: Int. Conf. Arti. Intell. in Medi. pp. 11-24. Springer (2021) +35. Wu, C., Wu, F., Qi, T., Huang, Y.: Noisytune: A little noise can help you finetune pretrained language models better. In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). pp. 680-685 (2022) +36. Xu, P., Zhu, X., Clifton, D.A.: Multimodal learning with transformers: A survey. arXiv preprint arXiv:2206.06488 (2022) +37. Xu, Y., Shang, L., Ye, J., Qian, Q., Li, Y.F., Sun, B., Li, H., Jin, R.: Dash: Semi-supervised learning with dynamic thresholding. In: International Conference on Machine Learning. pp. 11525-11536. PMLR (2021) +38. Yan, X., Naushad, J., Sun, S., Han, K., Tang, H., Kong, D., Ma, H., You, C., Xie, X.: Representation recovering for self-supervised pre-training on medical images. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2685-2695 (2023) +39. Yang, B., Liu, F., Li, Z., Yin, Q., You, C., Yin, B., Zou, Y.: Multimodal prompt learning for product title generation with extremely limited labels. arXiv preprint arXiv:2307.01969 (2023) + +40. Yang, J., Chen, H., Liang, Y., Huang, J., He, L., Yao, J.: Concl: Concept contrastive learning for dense prediction pre-training in pathology images. In: ECCV. pp. 523-539. Springer (2022) +41. You, C., Dai, W., Min, Y., Liu, F., Zhang, X., Feng, C., Clifton, D.A., Zhou, S.K., Staib, L.H., Duncan, J.S.: Rethinking semi-supervised medical image segmentation: A variance-reduction perspective. In: NeurIPS (2023) +42. You, C., Dai, W., Min, Y., Staib, L., Duncan, J.S.: Bootstrapping semi-supervised medical image segmentation with anatomical-aware contrastive distillation. In: International Conference on Information Processing in Medical Imaging. pp. 641-653. Springer (2023) +43. You, C., Dai, W., Min, Y., Staib, L., Sekhon, J., Duncan, J.S.: Action++: Improving semi-supervised medical image segmentation with adaptive anatomical contrast. arXiv preprint arXiv:2304.02689 (2023) +44. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917 (2022) +45. Yuan, H., Yuan, Z., Tan, C., Huang, F., Huang, S.: Hype: Better pre-trained language model fine-tuning with hidden representation perturbation. arXiv preprint arXiv:2212.08853 (2022) +46. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021) +47. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12104-12113 (June 2022) +48. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: Lit: Zero-shot transfer with locked-image text tuning. In: CVPR. pp. 18123-18133 (2022) +49. Zhang, B., Wang, Y., Hou, W., Wu, H., Wang, J., Okumura, M., Shinozaki, T.: Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. NeurIPS 34 (2021) +50. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8552-8562 (2022) +51. Zhang, R., Zhang, W., Fang, R., Gao, P., Li, K., Dai, J., Qiao, Y., Li, H.: Tip-adapter: Training-free adaption of clip for few-shot classification. In: ECCV. pp. 493-510. Springer (2022) +52. Zhang, T., Wu, F., Katiyar, A., Weinberger, K.Q., Artzi, Y.: Revisiting few-sample bert fine-tuning. In: ICLR (2021) +53. Zhang, W., Zhu, L., Hallinan, J., Zhang, S., Makmur, A., Cai, Q., Ooi, B.C.: Boostmis: Boosting medical image semi-supervised learning with adaptive pseudo labeling and informative active annotation. In: CVPR. pp. 20666-20676 (2022) +54. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. Int. J. Comput. Vision 130(9), 2337-2348 (2022) \ No newline at end of file diff --git a/2024/Bridging the Pathology Domain Gap_ Efficiently Adapting CLIP for Pathology Image Analysis with Limited Labeled Data/images.zip b/2024/Bridging the Pathology Domain Gap_ Efficiently Adapting CLIP for Pathology Image Analysis with Limited Labeled Data/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3ee245b6108b9c2b15e50ae7b156ce7e638a3b8c --- /dev/null +++ b/2024/Bridging the Pathology Domain Gap_ Efficiently Adapting CLIP for Pathology Image Analysis with Limited Labeled Data/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37998f7b90889e4de37579d67a1a63ecb229c99130a83938cabdb9c4b37615d2 +size 468739 diff --git a/2024/Bridging the Pathology Domain Gap_ Efficiently Adapting CLIP for Pathology Image Analysis with Limited Labeled Data/layout.json b/2024/Bridging the Pathology Domain Gap_ Efficiently Adapting CLIP for Pathology Image Analysis with Limited Labeled Data/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..eac8e2f8093e2989d33eceb91781b9fc252a84e8 --- /dev/null +++ b/2024/Bridging the Pathology Domain Gap_ Efficiently Adapting CLIP for Pathology Image Analysis with Limited Labeled Data/layout.json @@ -0,0 +1,8746 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 111, + 481, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 111, + 481, + 163 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 481, + 163 + ], + "type": "text", + "content": "Bridging the Pathology Domain Gap: Efficiently Adapting CLIP for Pathology Image Analysis with Limited Labeled Data" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 201, + 191, + 411, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 191, + 411, + 217 + ], + "spans": [ + { + "bbox": [ + 201, + 191, + 411, + 217 + ], + "type": "text", + "content": "Zhengfeng Lai1, Joohi Chauhan2, Brittany N. Dugger1, and Chen-Nee Chuah1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 184, + 224, + 429, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 224, + 429, + 270 + ], + "spans": [ + { + "bbox": [ + 184, + 224, + 429, + 270 + ], + "type": "text", + "content": "1 University of California, Davis \n2 Motilal Nehru National Institute of Technology Allahabad {lzhengfeng, bndugger, chuah}@ucdavis.edu joohi@mnnit.ac.in" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 285, + 455, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 285, + 455, + 506 + ], + "spans": [ + { + "bbox": [ + 160, + 285, + 455, + 506 + ], + "type": "text", + "content": "Abstract. Contrastive Language-Image Pre-training (CLIP) has shown its proficiency in acquiring distinctive visual representations and exhibiting strong generalization across diverse vision tasks. However, its effectiveness in pathology image analysis, particularly with limited labeled data, remains an ongoing area of investigation due to challenges associated with significant domain shifts and catastrophic forgetting. Thus, it is imperative to devise efficient adaptation strategies in this domain to enable scalable analysis. In this study, we introduce Path-CLIP, a framework tailored for a swift adaptation of CLIP to various pathology tasks. Firstly, we propose Residual Feature Refinement (RFR) with a dynamically adjustable ratio to effectively integrate and balance source and task-specific knowledge. Secondly, we introduce Hidden Representation Perturbation (HRP) and Dual-view Vision Contrastive (DVC) techniques to mitigate overfitting issues. Finally, we present the Doublet Multimodal Contrastive Loss (DMCL) for fine-tuning CLIP for pathology tasks. We demonstrate that Path-CLIP adeptly adapts pre-trained CLIP to downstream pathology tasks, yielding competitive results. Specifically, Path-CLIP achieves over " + }, + { + "bbox": [ + 160, + 285, + 455, + 506 + ], + "type": "inline_equation", + "content": "+19\\%" + }, + { + "bbox": [ + 160, + 285, + 455, + 506 + ], + "type": "text", + "content": " improvement in accuracy when utilizing mere " + }, + { + "bbox": [ + 160, + 285, + 455, + 506 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 160, + 285, + 455, + 506 + ], + "type": "text", + "content": " of labeled data in PCam with only 10 minutes of fine-tuning while running on a single GPU." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 522, + 230, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 230, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 230, + 534 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Recent advancements in deep learning, bolstered by superior network architectures and the availability of extensive, meticulously curated datasets, have led to substantial enhancements in pathology image analysis tasks [17]. Nevertheless, the process of collecting high-quality datasets with reliable annotations for each specific vision task remains a laborious and time-intensive endeavor [18,53]. This may prevent the broad adoption of advanced deep learning techniques. To relieve the reliance on such datasets, pre-training and fine-tuning methods (TFT) have been studied in vision tasks: pre-train the model on a large-scale dataset and then fine-tune the model on different downstream tasks [9]. There are several challenges of such methods: 1) they may still require a large amount of labeled" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 309, + 223 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 309, + 223 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 309, + 223 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 309, + 223 + ], + "type": "image", + "image_path": "035939840df9d49f945cdec888d3339b46dc0c353203c6d467ff41819d128dd2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 224, + 230, + 233 + ], + "lines": [ + { + "bbox": [ + 204, + 224, + 230, + 233 + ], + "spans": [ + { + "bbox": [ + 204, + 224, + 230, + 233 + ], + "type": "text", + "content": "(a) CLIP" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 314, + 114, + 480, + 223 + ], + "blocks": [ + { + "bbox": [ + 314, + 114, + 480, + 223 + ], + "lines": [ + { + "bbox": [ + 314, + 114, + 480, + 223 + ], + "spans": [ + { + "bbox": [ + 314, + 114, + 480, + 223 + ], + "type": "image", + "image_path": "3ac9c8ba0cb6d4a4780890b6a0397c57575cec8aac5b5ff4b35b60f5a6766d8e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 225, + 408, + 234 + ], + "lines": [ + { + "bbox": [ + 367, + 225, + 408, + 234 + ], + "spans": [ + { + "bbox": [ + 367, + 225, + 408, + 234 + ], + "type": "text", + "content": "(b) Path-CLIP" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 246, + 482, + 281 + ], + "lines": [ + { + "bbox": [ + 131, + 246, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 131, + 246, + 482, + 281 + ], + "type": "text", + "content": "Fig. 1: Illustration of (a) zero-shot CLIP, and (b) our proposed Path-CLIP. We aim to freeze the encoders but only train the Residual Feature Refinement (RFR) module to balance the original knowledge from CLIP and new task-specific knowledge." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 304, + 482, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 377 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 377 + ], + "type": "text", + "content": "set to avoid the overfitting issue when fine-tuning the model for the downstream task [21, 40]; 2) the fine-tuning may not yield satisfactory performance in the target domain due to the existence of a large domain gap between the pre-trained data and pathology images [32]. Consequently, a pivotal question arises: how can the large-scale pre-trained model be efficiently tailored to a new task, minimizing overfitting, particularly when the new dataset is limited in size?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 383, + 483, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 383, + 483, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 383, + 483, + 538 + ], + "type": "text", + "content": "To fill the performance gap due to domain shift, Contrastive Language-Image Pre-training (CLIP) [26] has shown its power in learning generic and distinctive visual representations via language supervision. As shown in Fig. 1(a), it aligns images and texts in the same feature space and uses a contrastive loss to formulate the learning objective. CLIP uses two separate encoders for images and texts, then maximizes the similarity score of positive pairs of images and texts while minimizing it of negative pairs [26,54]. It achieves promising results on various image classification tasks without needing any annotated data, i.e., zero-shot transfer settings. As a language-vision model, CLIP uses prompts as the supervision, where the visual labels are entered into the hand-crafted template. By pre-training the model at a large scale, models can learn the visual contents and easily be transferred to downstream tasks through the prompt-based zero-shot transfer." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 546, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 483, + 666 + ], + "type": "text", + "content": "However, the manual design of prompts can be a non-trivial and time-consuming task. [54] found that even a slight change in the prompt (e.g., one word) can make a big difference. They introduced Context Optimization (CoOp) to automate prompt engineering to generate continuous soft prompts instead of using hand-chosen hard prompts [54]. However, CoOp requires substantial computing resources, and the results of CoOp are not interpretable. Furthermore, such prompt tuning faces performance degradation when there is a significant domain shift [39], e.g., from natural images to pathology images, rendering it less adaptable to medical imaging tasks. On the other hand, medical image analysis may involve multiple learning tasks simultaneously. Conversely, in medical im" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Z. Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "age analysis, multiple learning tasks may be concurrently required. It becomes impractical if fine-tuning necessitates substantial computational resources. For instance, Lai et al. [17] quantify and visualize distinct pathologies in grey/white matter (GM/WM) regions within gigapixel images. To accomplish this, the task is divided into two learning objectives: pathology detection and GM/WM segmentation. They collected and annotated two datasets and then trained two models from scratch to achieve it. Therefore, it will be beneficial if we can propose an effective way to adapt a big model into multiple downstream tasks quickly so the final analysis objective can be reached in a reasonable time manner." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 559 + ], + "type": "text", + "content": "In this work, we aim to fine-tune CLIP efficiently with light computing resources for pathology image analysis tasks. There are several challenges in fine-tuning CLIP. First, overfitting is a severe issue if we adapt CLIP to the downstream tasks directly since CLIP is pre-trained on a 400M dataset while the new domain dataset can be small [10]. Second, CLIP has shown its promising zero-shot generalization ability, directly fine-tuning CLIP may lose this capability and face the \"catastrophic forgetting\" issue. In other words, it is unclear how to effectively learn the new knowledge while retaining the original pre-trained knowledge so that the generalization of CLIP can be maintained. Third, there are limited studies on benchmarking CLIP's transferability in the pathology domain; hence, its applicability remains unclear. To address the above issues, we first study the applicability of CLIP on two public pathology datasets and benchmark the zero-shot ability of CLIP on them. Then, we did empirical experiments to verify the importance of the CLIP's pre-trained knowledge. To address the \"catastrophic forgetting\" issue, we propose Residual Feature Refinement (RFR) as a lightweight approach for adapting CLIP to pathology images. As a shallow network, RFR demonstrates proficiency in extracting crucial spatial features and patterns from pathology images. Within RFR, we introduce a self-adaptive module designed to seamlessly fuse and balance the original knowledge from CLIP with the newly acquired insights from the specific pathological task, requiring only a minimal number of additional trainable weights, as opposed to retraining the entire encoders in CLIP. To further mitigate overfitting, we also incorporate Hidden Representation Perturbation (HRP) and Dual-view Vision Contrastive (DVC). Finally, we establish an end-to-end adaptation of CLIP, named Path-CLIP (as shown in Fig. 1(b)), tailored for pathology image analysis tasks. Path-CLIP incorporates a newly proposed loss function, Doublet Multimodal Contrastive Loss (DMCL)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 570, + 326, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 570, + 326, + 582 + ], + "spans": [ + { + "bbox": [ + 131, + 570, + 326, + 582 + ], + "type": "text", + "content": "We summarize our contributions as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 594, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 594, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 594, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 138, + 594, + 481, + 628 + ], + "type": "text", + "content": "- We explore the applicability of CLIP on pathology images and highlight the catastrophic forgetting issues when adapting CLIP to downstream pathology tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 641, + 481, + 665 + ], + "type": "text", + "content": "- We propose Path-CLIP to alleviate the overfitting and forgetting issues and effectively adapt CLIP for pathology tasks." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 448, + 103 + ], + "type": "text", + "content": "Path-CLIP: Efficiently Adapting CLIP for Pathology Image Analysis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 483, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 483, + 152 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 483, + 152 + ], + "type": "text", + "content": "- Extensive experiments demonstrate the efficiency and robustness of PathCLIP. For example, it achieves " + }, + { + "bbox": [ + 137, + 116, + 483, + 152 + ], + "type": "inline_equation", + "content": "+19\\%" + }, + { + "bbox": [ + 137, + 116, + 483, + 152 + ], + "type": "text", + "content": " on PCam when only using " + }, + { + "bbox": [ + 137, + 116, + 483, + 152 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 137, + 116, + 483, + 152 + ], + "type": "text", + "content": " of labeled data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 171, + 237, + 183 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 237, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 237, + 183 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 196, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 196, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 196, + 482, + 376 + ], + "type": "text", + "content": "Language-vision model (LVM). LVMs have shown promising performance in learning generic visual representations [15, 22, 26, 48]. Recent models advance via text representation learning with large-scale Transformers [30] and even web-scale training datasets [54]. Transformer-based multimodal learning has achieved great success in such large-scale datasets [23, 29, 36]. For example, CLIP [26] was trained on 400 million image-caption pairs and achieved state-of-the-art results in many fields [26, 44, 46, 47]. In CLIP's architecture, it has two encoders: the vision encoder can be ResNet [12] or ViT [7] while the text encoder can be Transformer (e.g., BERT [6]). A recent work [27] fine-tuned CLIP for video data and performed competitively to more complicated methods designed for video processing. PointCLIP [50] applied CLIP to 3D recognition. CLIP is also applied for image generation tasks [8] and shows its power to reduce required data collection. However, the best way to adapt CLIP for downstream tasks is still being studied, especially when the new domain is far from the pre-trained domains, e.g., the medical domain." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 381, + 483, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 483, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 483, + 562 + ], + "type": "text", + "content": "CLIP Fine-tuning and adaptation. Although recent multimodal models (e.g., CLIP [26]) perform astonishingly well in standard vision tasks without fine-tuning, the gap between CLIP and supervised learning is still noticeable. Therefore, many works focus on fine-tuning CLIP with a few samples from the new task provided. For example, CoOp [54] kept the entire pre-trained parameters fixed but optimized the continuous prompt through a learnable vector. This method closes the gap between CLIP and supervised learning but the performance gain is limited as it only focuses on the perspective of prompt design instead of the vision part for the image recognition tasks. On the other hand, CLIP-Adapter [9] and Tip-Adapter [51] freeze the pre-trained weights and train additive module, which significantly improves the CLIP's performance with limited training data in the new task. These methods require a lower computational cost. However, both CLIP-Adapter [9] and Tip-Adapter [51] involve tuning a ratio (hyper-parameter) to balance the original and fine-tuned knowledge manually, which may not be practical when we have multiple target learning tasks in the medical domain." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 579, + 233, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 233, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 233, + 594 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "In this section, we introduce the proposed Path-CLIP. We summarize the end-to-end pipeline in Fig. 1(b): a self-adaptive residual ratio " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " will be incorporated into Residual Feature Refinement (Section 3.2) to fuse and balance the original knowledge with the task-specific knowledge from the new downstream task. We introduce Hidden Representation Perturbation (HRP) and Dual-view Vision" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "text", + "content": "Z. Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 113, + 465, + 270 + ], + "blocks": [ + { + "bbox": [ + 149, + 113, + 465, + 270 + ], + "lines": [ + { + "bbox": [ + 149, + 113, + 465, + 270 + ], + "spans": [ + { + "bbox": [ + 149, + 113, + 465, + 270 + ], + "type": "image", + "image_path": "fa6cef420246b5d56be80fa190e37fe1d539dca58ddea36076b5ca2aaeb916b5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 276, + 483, + 343 + ], + "lines": [ + { + "bbox": [ + 130, + 276, + 483, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 483, + 343 + ], + "type": "text", + "content": "Fig. 2: Overview of the Residual Feature Refinement (RFR): a downsampling-upsampling shallow architecture of fully connected layers. RFR blends the fine-tuned knowledge with the original knowledge from CLIP's vision encoder " + }, + { + "bbox": [ + 130, + 276, + 483, + 343 + ], + "type": "inline_equation", + "content": "(F(\\cdot))" + }, + { + "bbox": [ + 130, + 276, + 483, + 343 + ], + "type": "text", + "content": " via a self-adaptive residual ratio " + }, + { + "bbox": [ + 130, + 276, + 483, + 343 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 276, + 483, + 343 + ], + "type": "text", + "content": ". HRP refers to Hidden Representation Perturbation and TFT refers to pre-training fine-tuning. We use blue color to denote the original visual features and orange color to denote the fine-tuned features." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 373, + 482, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 409 + ], + "type": "text", + "content": "Contrastive (Section 3.3) to alleviate the overfitting issue. Finally, we propose Doublet Multimodal Contrastive Loss (DMCL) in Section 3.4 as the loss function for fine-tuning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 441, + 309, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 441, + 309, + 453 + ], + "spans": [ + { + "bbox": [ + 131, + 441, + 309, + 453 + ], + "type": "text", + "content": "3.1 Language-Vision Pre-training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 473, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 570 + ], + "type": "text", + "content": "We first revisit language-vision pre-training and use CLIP [26] as one example. Note that our framework is applicable to broader language-vision models. CLIP [26] has a vision encoder " + }, + { + "bbox": [ + 130, + 473, + 482, + 570 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 130, + 473, + 482, + 570 + ], + "type": "text", + "content": " and a text encoder " + }, + { + "bbox": [ + 130, + 473, + 482, + 570 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 130, + 473, + 482, + 570 + ], + "type": "text", + "content": ". The vision encoder maps a high-dimensional image into low-dimensional image embeddings. The vision encoder can be a CNN like ResNet [12] or a Vision Transformer (ViT) [7]. The text encoder is built on Transformer [30] and generates text embeddings from the prompt. Specifically, \"prompt\" here refers to a sequence of words (tokens), such as \"a neuropathological image of grey matter\"." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Training. CLIP jointly trains " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " to optimize the similarity score (e.g., symmetric cross-entropy loss [33]) between the visual and textual embeddings for each batch. The input consists of an image and its corresponding prompt (e.g., \"this is a photo of grey matter\"). They formulate the learning objective as a contrastive loss to align the visual and textual embedding spaces. Given a batch of image-prompt pairs, CLIP maximizes the similarity score for positive pairs while minimizing it for negative pairs." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 448, + 103 + ], + "type": "text", + "content": "Path-CLIP: Efficiently Adapting CLIP for Pathology Image Analysis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "content": "Contrastive loss function. In CLIP training, the agreement between a positive pair of vision and text embeddings is maximized by InfoNCE loss [24]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 227, + 148, + 482, + 176 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 148, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 227, + 148, + 482, + 176 + ], + "type": "interline_equation", + "content": "l _ {i} ^ {v \\rightarrow t} = - \\log \\frac {\\exp \\left(\\mathbf {s i m} \\left(v _ {i} , t _ {i}\\right) / \\gamma\\right)}{\\sum_ {k = 1} ^ {B} \\exp \\left(\\mathbf {s i m} \\left(v _ {i} , t _ {k}\\right) / \\gamma\\right)}, \\tag {1}", + "image_path": "f937d63e99f01c9bc08a0fb412afd83659ce0187264f9690c67e302d41899132.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 182, + 482, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 182, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 131, + 182, + 482, + 255 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 182, + 482, + 255 + ], + "type": "inline_equation", + "content": "v, t" + }, + { + "bbox": [ + 131, + 182, + 482, + 255 + ], + "type": "text", + "content": " are normalized vectors from vision and text encoders. " + }, + { + "bbox": [ + 131, + 182, + 482, + 255 + ], + "type": "inline_equation", + "content": "(v_{i}, t_{i})" + }, + { + "bbox": [ + 131, + 182, + 482, + 255 + ], + "type": "text", + "content": " is a positive pair. " + }, + { + "bbox": [ + 131, + 182, + 482, + 255 + ], + "type": "inline_equation", + "content": "\\mathbf{sim}" + }, + { + "bbox": [ + 131, + 182, + 482, + 255 + ], + "type": "text", + "content": " is a function to measure the similarity of two vectors. " + }, + { + "bbox": [ + 131, + 182, + 482, + 255 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 131, + 182, + 482, + 255 + ], + "type": "text", + "content": " is the mini-batch size of the image-text pairs and " + }, + { + "bbox": [ + 131, + 182, + 482, + 255 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 182, + 482, + 255 + ], + "type": "text", + "content": " is the learnable temperature. Therefore, there is only 1 positive text for the input " + }, + { + "bbox": [ + 131, + 182, + 482, + 255 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 131, + 182, + 482, + 255 + ], + "type": "text", + "content": " image but " + }, + { + "bbox": [ + 131, + 182, + 482, + 255 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 131, + 182, + 482, + 255 + ], + "type": "text", + "content": " negative texts. " + }, + { + "bbox": [ + 131, + 182, + 482, + 255 + ], + "type": "inline_equation", + "content": "l_{i}^{v \\rightarrow t}" + }, + { + "bbox": [ + 131, + 182, + 482, + 255 + ], + "type": "text", + "content": " means the InfoNCE loss from an image " + }, + { + "bbox": [ + 131, + 182, + 482, + 255 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 182, + 482, + 255 + ], + "type": "text", + "content": " to texts, so as the text loss " + }, + { + "bbox": [ + 131, + 182, + 482, + 255 + ], + "type": "inline_equation", + "content": "l_{i}^{t \\rightarrow v}" + }, + { + "bbox": [ + 131, + 182, + 482, + 255 + ], + "type": "text", + "content": ". Then the loss function for CLIP is as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 245, + 265, + 482, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 265, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 245, + 265, + 482, + 291 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C L I P}} = \\frac {1}{2 B} \\sum_ {i = 1} ^ {B} \\left(l _ {i} ^ {v \\rightarrow t} + l _ {i} ^ {t \\rightarrow v}\\right). \\tag {2}", + "image_path": "4baef87dc3cc9e8d9221bdb2ecc8e7e62a8a8d370f1a901f81e38f19a60c9e02.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 304, + 482, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 304, + 482, + 388 + ], + "spans": [ + { + "bbox": [ + 131, + 304, + 482, + 388 + ], + "type": "text", + "content": "Inference for classification tasks. For the inference process, an image " + }, + { + "bbox": [ + 131, + 304, + 482, + 388 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 131, + 304, + 482, + 388 + ], + "type": "text", + "content": " is transformed into a feature manifold " + }, + { + "bbox": [ + 131, + 304, + 482, + 388 + ], + "type": "inline_equation", + "content": "f \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 131, + 304, + 482, + 388 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 304, + 482, + 388 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 131, + 304, + 482, + 388 + ], + "type": "text", + "content": " is the feature dimension. Then, " + }, + { + "bbox": [ + 131, + 304, + 482, + 388 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 131, + 304, + 482, + 388 + ], + "type": "text", + "content": " is multiplied with a classifier weight matrix " + }, + { + "bbox": [ + 131, + 304, + 482, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{D \\times K}" + }, + { + "bbox": [ + 131, + 304, + 482, + 388 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 304, + 482, + 388 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 304, + 482, + 388 + ], + "type": "text", + "content": " is the number of classes in the learning task. We get a " + }, + { + "bbox": [ + 131, + 304, + 482, + 388 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 304, + 482, + 388 + ], + "type": "text", + "content": "-dimensional logit after matrix multiplication. Then we apply softmax to convert this logit into a probability vector " + }, + { + "bbox": [ + 131, + 304, + 482, + 388 + ], + "type": "inline_equation", + "content": "p \\in \\mathbb{R}^K" + }, + { + "bbox": [ + 131, + 304, + 482, + 388 + ], + "type": "text", + "content": " over the " + }, + { + "bbox": [ + 131, + 304, + 482, + 388 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 304, + 482, + 388 + ], + "type": "text", + "content": " classes. The whole process can be summarized as the following equation:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 248, + 388, + 482, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 388, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 248, + 388, + 482, + 417 + ], + "type": "interline_equation", + "content": "p _ {i} = \\frac {\\exp \\left(\\mathbf {W} _ {\\mathbf {i}} ^ {T} \\times f\\right) / \\gamma}{\\sum_ {i = 1} ^ {K} \\exp \\left(\\mathbf {W} _ {\\mathbf {i}} ^ {T} \\times f\\right) / \\gamma}, \\tag {3}", + "image_path": "38c8692e17cee3cafe8a195d25c1ade7f31fafdc678c69be2597260355cc7f20.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 420, + 482, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 420, + 482, + 444 + ], + "spans": [ + { + "bbox": [ + 131, + 420, + 482, + 444 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 420, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 420, + 482, + 444 + ], + "type": "text", + "content": " is a temperature parameter learned by CLIP during training and " + }, + { + "bbox": [ + 131, + 420, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\mathbf{W_i}" + }, + { + "bbox": [ + 131, + 420, + 482, + 444 + ], + "type": "text", + "content": " is the prototype weight vector for class- " + }, + { + "bbox": [ + 131, + 420, + 482, + 444 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 420, + 482, + 444 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 449, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 449, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 131, + 449, + 482, + 534 + ], + "type": "text", + "content": "Transferability. Compared to traditional classifier learning methods where the supervision is a fixed one-hot vector, language-vision pre-training utilizes an open-set visual concept that can be explored through the text encoder to learn a broader semantic space, which in turn makes the learned representations more diverse and adaptable to various downstream tasks (e.g., zero-shot learning). Therefore, we argue that such language-vision pre-training is helpful for medical problems where the domain is very distinctive from the natural images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 551, + 307, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 551, + 307, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 551, + 307, + 562 + ], + "type": "text", + "content": "3.2 Residual Feature Refinement" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 570, + 482, + 666 + ], + "type": "text", + "content": "In this subsection, we introduce Residual Feature Refinement (RFR) to learn the task-relevant context when we adapt CLIP to the downstream learning tasks. Unlike CoOp's prompt tuning, which may not address the domain shift issue between the natural and pathology images, we focus on fine-tuning the visual features " + }, + { + "bbox": [ + 131, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 131, + 570, + 482, + 666 + ], + "type": "text", + "content": ". However, simple fine-tuning of the entire network may fail in a new pathological task due to 1) the overfitting issue caused by a large number of parameters in CLIP and a shortage of training samples [13]; 2) catastrophic forgetting issue." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "text", + "content": "Z. Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 209, + 146, + 401, + 215 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 138 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 138 + ], + "type": "text", + "content": "Table 1: Baseline. We fine-tune CLIP (ResNet-50) on PCam [31] with " + }, + { + "bbox": [ + 130, + 114, + 482, + 138 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 138 + ], + "type": "text", + "content": " training data. \"No fine-tuning\" is the best baseline due to the catastrophic forgetting issue." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 209, + 146, + 401, + 215 + ], + "lines": [ + { + "bbox": [ + 209, + 146, + 401, + 215 + ], + "spans": [ + { + "bbox": [ + 209, + 146, + 401, + 215 + ], + "type": "table", + "html": "
# ConfigurationAccuracy
1No fine-tuning56.5%
2Fine-tune 10 epochs40.3%
3Fine-tune 100 epochs35.9%
4With a lower learning rate on #345.6%
", + "image_path": "1c9f581b851326ba35fabd2f734b3d811117394d617297dec3bf36d9dba63fda.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 242, + 483, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 242, + 483, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 483, + 470 + ], + "type": "text", + "content": "Importance of the CLIP's pre-trained knowledge. CLIP has shown its zero-shot ability and transferability to a wide range of downstream tasks. We argue that preserving pre-trained knowledge of CLIP is crucial to maintain its generalization and alleviate the overfitting issue when the new dataset is small. There are two reasons behind this. First, one recent NLP work [11] shows large pre-trained language models (PLMs) are well-calibrated on the masked language modeling tasks. However, such property cannot be retained through direct fine-tuning due to catastrophic forgetting. Specifically, the pre-trained features are only used for the initialization and will be distorted by the fine-tuning [11]. Inspired by this, we investigate whether this issue also appears in language-vision models. As shown in Table 1, \"no fine-tuning\" but zero-shot gets the best results compared to fine-tuning CLIP on " + }, + { + "bbox": [ + 130, + 242, + 483, + 470 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 242, + 483, + 470 + ], + "type": "text", + "content": " of training data in PCam [31]. We also attempt to lower the learning rate to " + }, + { + "bbox": [ + 130, + 242, + 483, + 470 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-6}" + }, + { + "bbox": [ + 130, + 242, + 483, + 470 + ], + "type": "text", + "content": ", but the results are still far from satisfactory. Therefore, we conclude direct fine-tuning may distort the pre-trained knowledge and result in a severe catastrophic forgetting issue. Besides the forgetting issue, overfitting is another common challenge when adapting a big model to a new task. To alleviate the above two issues, we propose a residual refinement architecture to preserve the pre-trained features and acquire rich semantic information for pathology image analysis." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 480, + 483, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 483, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 483, + 577 + ], + "type": "text", + "content": "Architecture overview. As shown in Fig. 2, given an image " + }, + { + "bbox": [ + 130, + 480, + 483, + 577 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 480, + 483, + 577 + ], + "type": "text", + "content": ", we get the visual feature " + }, + { + "bbox": [ + 130, + 480, + 483, + 577 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 480, + 483, + 577 + ], + "type": "text", + "content": " from the image encoder and compute the classifier weight " + }, + { + "bbox": [ + 130, + 480, + 483, + 577 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 130, + 480, + 483, + 577 + ], + "type": "text", + "content": " from the text encoder. Then we design trainable adaptive layers " + }, + { + "bbox": [ + 130, + 480, + 483, + 577 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 480, + 483, + 577 + ], + "type": "text", + "content": " to convert " + }, + { + "bbox": [ + 130, + 480, + 483, + 577 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 480, + 483, + 577 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 480, + 483, + 577 + ], + "type": "inline_equation", + "content": "L(f)" + }, + { + "bbox": [ + 130, + 480, + 483, + 577 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 480, + 483, + 577 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 480, + 483, + 577 + ], + "type": "text", + "content": " can be multiple layers of linear transformations in a \"down-and-up\" architecture. As shown in Fig. 2, we have four layers to transform the feature dimension as \"1024-256-64-256-1024\". This is one example: the layers are flexible with other dimensions. After this architecture, " + }, + { + "bbox": [ + 130, + 480, + 483, + 577 + ], + "type": "inline_equation", + "content": "L(f)" + }, + { + "bbox": [ + 130, + 480, + 483, + 577 + ], + "type": "text", + "content": " can be of the same size of " + }, + { + "bbox": [ + 130, + 480, + 483, + 577 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 480, + 483, + 577 + ], + "type": "text", + "content": " and blended with " + }, + { + "bbox": [ + 130, + 480, + 483, + 577 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 480, + 483, + 577 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 251, + 591, + 481, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 591, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 251, + 591, + 481, + 604 + ], + "type": "interline_equation", + "content": "f _ {i} ^ {*} = \\left(1 - \\alpha_ {i}\\right) L \\left(f _ {i}\\right) + \\alpha_ {i} f _ {i}, \\tag {4}", + "image_path": "e881ac6859a442ebd91b3dcee8a42e04d38da1b3438f375c2182a6139365bc52.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "i = \\{1,2,3,\\dots n\\}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " is the self-adaptive residual ratio to balance the fine-tuned knowledge and the original CLIP's knowledge. Specifically, the first term adaptively learns knowledge from the new dataset, while the second term preserves the original knowledge from CLIP." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 167, + 90, + 448, + 103 + ], + "type": "text", + "content": "Path-CLIP: Efficiently Adapting CLIP for Pathology Image Analysis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "Then we adopt Eq. (3) with the new " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "f^*" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " to get the class probability vector and predict the category. During fine-tuning, the weights of the trainable layers are optimized through the symmetric contrastive loss used in CLIP [26]. The prompt can be \"this is a photo of [], where \"+\" is filled with the class name. We can incorporate pathological descriptions into the prompt to provide more domain information, e.g., \"this is a pathological breast photo of []\"." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 194, + 482, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 316 + ], + "type": "text", + "content": "Hidden Representation Perturbation (HRP). Fine-tuning a pre-trained language-vision model on a small dataset can lead to potential issues such as overfitting or representation collapse [45]. To mitigate these issues, we propose Hidden Representation Perturbation (HRP) that is strategically integrated with the trainable layers. To balance the model's complexity and efficiency, we adopt a \"four-layer\" architecture with HRP, steering clear of deeper architectures to ensure a controlled number of trainable parameters. This precaution is crucial as an excess of trainable parameters may exacerbate overfitting, especially in scenarios characterized by limited resources, where only a sparse dataset is available for the new task [52]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 320, + 483, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 483, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 483, + 418 + ], + "type": "text", + "content": "To alleviate the overfitting issue in low-resource settings, we focus on adding noise during the training process, a widely-used NLP strategy for large language models to smoothen the optimization landscape. For example, NoisyTune [35] adds the perturbations to the pre-trained weights to improve the model's generalization. Inspired by these NLP strategies [19, 35, 45], we propose to add random noise to the feature " + }, + { + "bbox": [ + 130, + 320, + 483, + 418 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 320, + 483, + 418 + ], + "type": "text", + "content": " before trainable adapting layers fine-tuning it. The random noise can be Gaussian noise " + }, + { + "bbox": [ + 130, + 320, + 483, + 418 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 130, + 320, + 483, + 418 + ], + "type": "text", + "content": " generated from " + }, + { + "bbox": [ + 130, + 320, + 483, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0, \\sigma^2)" + }, + { + "bbox": [ + 130, + 320, + 483, + 418 + ], + "type": "text", + "content": ". After this, the perturbed feature " + }, + { + "bbox": [ + 130, + 320, + 483, + 418 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 320, + 483, + 418 + ], + "type": "text", + "content": " is converted into " + }, + { + "bbox": [ + 130, + 320, + 483, + 418 + ], + "type": "inline_equation", + "content": "\\hat{f} = f + \\varepsilon" + }, + { + "bbox": [ + 130, + 320, + 483, + 418 + ], + "type": "text", + "content": ". Therefore, Eq. (4) is converted into:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 252, + 429, + 482, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 429, + 482, + 444 + ], + "spans": [ + { + "bbox": [ + 252, + 429, + 482, + 444 + ], + "type": "interline_equation", + "content": "f _ {i} ^ {*} = \\left(1 - \\alpha_ {i}\\right) L \\left(\\hat {f} _ {i}\\right) + \\alpha_ {i} f _ {i}. \\tag {5}", + "image_path": "d5bf2ee3af6bcf35d4954381f4c42f853a61871e9bbc1a66982a51231767a927.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 451, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 475 + ], + "type": "text", + "content": "Therefore, the residual feature will be perturbed with the Gaussian noise for the refinement, designed to alleviate the overfitting issue." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 495, + 297, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 495, + 297, + 506 + ], + "spans": [ + { + "bbox": [ + 131, + 495, + 297, + 506 + ], + "type": "text", + "content": "3.3 Self-adaptive residual ratio" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 515, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 482, + 635 + ], + "type": "text", + "content": "In Section 3.2, we introduce residual feature refinement and the residual ratio " + }, + { + "bbox": [ + 130, + 515, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 515, + 482, + 635 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 515, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 515, + 482, + 635 + ], + "type": "text", + "content": " helps modify the degree of preserving the original knowledge for better performance. Empirically, " + }, + { + "bbox": [ + 130, + 515, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 515, + 482, + 635 + ], + "type": "text", + "content": " should be small if the domain gap between the source data and the downstream task is enormous, so more knowledge from the downstream task " + }, + { + "bbox": [ + 130, + 515, + 482, + 635 + ], + "type": "inline_equation", + "content": "(L(f))" + }, + { + "bbox": [ + 130, + 515, + 482, + 635 + ], + "type": "text", + "content": " can be considered. However, if " + }, + { + "bbox": [ + 130, + 515, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 515, + 482, + 635 + ], + "type": "text", + "content": " is too small, the model can be overfitting as it will heavily rely on the fine-tuned features and disregard the pre-trained knowledge from CLIP. On the other hand, the model can also be underfitting if " + }, + { + "bbox": [ + 130, + 515, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 515, + 482, + 635 + ], + "type": "text", + "content": " is too large, as it may prevent learning from the fine-tuned features. Therefore, it is crucial to quantify the residual ratio " + }, + { + "bbox": [ + 130, + 515, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 515, + 482, + 635 + ], + "type": "text", + "content": " so the model can be well-trained instead of underfitting or overfitting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Recent works such as Tip-Adapter [51] and CLIP-Adapter [9] involve feature refinement while they set such ratio as a fixed hyper-parameter that needs to be" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "text", + "content": "Z. Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "tuned during the training process. We argue the residual ratio should be dynamic and self-adaptive during training instead of fixed, as the learning status changes along the training process. Hence, we propose a self-adaptive residual ratio in Path-CLIP." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 170, + 480, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 170, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 170, + 480, + 289 + ], + "type": "text", + "content": "DVC: Dual-view Vision Contrastive. Inspired by self-supervised learning (e.g., SimCLR [4]), the composition of data augmentations is relevant to define effective predictive tasks. Therefore, we propose Dual-view Vision Contrastive (DVC) to measure the model's learning status. Given an image " + }, + { + "bbox": [ + 130, + 170, + 480, + 289 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 170, + 480, + 289 + ], + "type": "text", + "content": ", we leverage two types of augmentations: \"weak\" and \"strong\" [28]. \"weak\" refers to the standard flip-and-shift augmentation strategy while \"strong\" refers to perturb the appearance (RandAugment [5] and CTAugment [2]). We propose to measure the distance between representations from both augmentations, e.g., cosine similarity has shown good performance in both visual contrastive [4] and textual contrastive [3] tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 296, + 481, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 481, + 331 + ], + "type": "text", + "content": "With the input " + }, + { + "bbox": [ + 130, + 296, + 481, + 331 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 296, + 481, + 331 + ], + "type": "text", + "content": ", we will get " + }, + { + "bbox": [ + 130, + 296, + 481, + 331 + ], + "type": "inline_equation", + "content": "w(x)" + }, + { + "bbox": [ + 130, + 296, + 481, + 331 + ], + "type": "text", + "content": " from the weak augmentation and " + }, + { + "bbox": [ + 130, + 296, + 481, + 331 + ], + "type": "inline_equation", + "content": "s(x)" + }, + { + "bbox": [ + 130, + 296, + 481, + 331 + ], + "type": "text", + "content": " from the strong augmentation. We obtain " + }, + { + "bbox": [ + 130, + 296, + 481, + 331 + ], + "type": "inline_equation", + "content": "F(w(x))" + }, + { + "bbox": [ + 130, + 296, + 481, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 296, + 481, + 331 + ], + "type": "inline_equation", + "content": "F(s(x))" + }, + { + "bbox": [ + 130, + 296, + 481, + 331 + ], + "type": "text", + "content": " after the vision encoder " + }, + { + "bbox": [ + 130, + 296, + 481, + 331 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 296, + 481, + 331 + ], + "type": "text", + "content": ". We define the distance as " + }, + { + "bbox": [ + 130, + 296, + 481, + 331 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 296, + 481, + 331 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 231, + 335, + 481, + 362 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 335, + 481, + 362 + ], + "spans": [ + { + "bbox": [ + 231, + 335, + 481, + 362 + ], + "type": "interline_equation", + "content": "\\tau = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\cos \\left(F \\left(w \\left(x _ {i}\\right)\\right), F \\left(s \\left(x _ {i}\\right)\\right)\\right), \\tag {6}", + "image_path": "697d0d05e7de3c42a6f01e4abd7aca3dd08853f57158b187abcba173ab31c269.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 366, + 479, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 479, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 479, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 366, + 479, + 415 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 366, + 479, + 415 + ], + "type": "text", + "content": " is the number of samples in the training set, and " + }, + { + "bbox": [ + 130, + 366, + 479, + 415 + ], + "type": "inline_equation", + "content": "\\cos" + }, + { + "bbox": [ + 130, + 366, + 479, + 415 + ], + "type": "text", + "content": " is a cosine similarity. The model's learning status is relevant to its consistency across different perturbations. We traverse the entire training set and calculate its mean value to represent the model's performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 421, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 421, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 421, + 482, + 517 + ], + "type": "text", + "content": "Bound constraints. In addition to DVC which measures the predictive consistency of the model, we also select \"accuracy\" on the training set as a second metric to measure the learning status, denoted as " + }, + { + "bbox": [ + 130, + 421, + 482, + 517 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 421, + 482, + 517 + ], + "type": "text", + "content": ". Empirically, \"accuracy\" alone cannot fully tell if the model is well-fitted: 100% of accuracy may indicate overfitting, which is the biggest challenge when we adapt a large model to a downstream task where the training set is very small. However, \"accuracy\" is a good metric to show if the model is underfitting. Therefore, we propose to fuse DVC and \"accuracy\" to represent the model's status during training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 522, + 479, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 479, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 479, + 583 + ], + "type": "text", + "content": "If we look at " + }, + { + "bbox": [ + 130, + 522, + 479, + 583 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 522, + 479, + 583 + ], + "type": "text", + "content": " and set it fixed, we can find the model is prone to overfit if " + }, + { + "bbox": [ + 130, + 522, + 479, + 583 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 522, + 479, + 583 + ], + "type": "text", + "content": " is too small as it will heavily rely on the fine-tuned knowledge; the model tends to underfit if " + }, + { + "bbox": [ + 130, + 522, + 479, + 583 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 522, + 479, + 583 + ], + "type": "text", + "content": " is too large as it prevents the model from learning new knowledge (directly use the original pre-trained knowledge). Therefore, we propose a dynamic way with lower and upper bounds to obtain " + }, + { + "bbox": [ + 130, + 522, + 479, + 583 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 522, + 479, + 583 + ], + "type": "text", + "content": " during training:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 250, + 588, + 481, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 588, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 250, + 588, + 481, + 601 + ], + "type": "interline_equation", + "content": "\\alpha = \\min (\\max (\\tau , T), 1 - \\tau). \\tag {7}", + "image_path": "271c0aa0efc3681504787e99ad28670f31c63f59061c679f80496ad8ab32da47.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "The bounds are associated with predictive consistency (DVC) and predictive accuracy on the training set. For example, if the model is overfitting, " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": " will be large, and " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": " will be relatively small. Then " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": " will be smaller in the next epoch to use less fine-tuned features. Therefore, our proposed bound constraints can effectively adjust " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": " to alleviate the overfitting issue." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 448, + 103 + ], + "type": "text", + "content": "Path-CLIP: Efficiently Adapting CLIP for Pathology Image Analysis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 114, + 449, + 208 + ], + "blocks": [ + { + "bbox": [ + 167, + 114, + 449, + 208 + ], + "lines": [ + { + "bbox": [ + 167, + 114, + 449, + 208 + ], + "spans": [ + { + "bbox": [ + 167, + 114, + 449, + 208 + ], + "type": "image", + "image_path": "c2e7f96d3618a13c1d9605bfa1ac5f219d04fc819c37596898e29877a1210be4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 216, + 482, + 249 + ], + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 249 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 249 + ], + "type": "text", + "content": "Fig. 3: Doublet Multimodal Contrastive Loss (DMCL) that utilizes the traditional image-text pair and the proposed image-image pair with weak/strong augmentation criteria." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 275, + 477, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 275, + 477, + 289 + ], + "spans": [ + { + "bbox": [ + 131, + 275, + 477, + 289 + ], + "type": "text", + "content": "3.4 Path-CLIP: End-to-end adaptation of CLIP to pathology tasks" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 300, + 482, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 396 + ], + "type": "text", + "content": "To efficiently adapt CLIP to pathology image tasks, we propose RFR with a self-adaptive residual ratio to balance the original and fine-tuned knowledge based on the learning status. To alleviate the overfitting issue, we incorporate HRP and design bound constraints for the residual ratio. Besides these, to distill more visual representations from the new image domain, we add a contrastive loss to the loss function so that the trainable portion has a better classifying ability. We follow SimCLR [4] to define the Contrastive Loss (CL) for one image in the batch as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 210, + 405, + 482, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 405, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 210, + 405, + 482, + 456 + ], + "type": "interline_equation", + "content": "l _ {i} ^ {c} = \\left\\{ \\begin{array}{l} - \\log \\frac {\\exp (\\mathbf {s i m} (v _ {i} , v _ {i + B}) / \\gamma)}{\\sum_ {k \\neq i} ^ {2 B} \\exp (\\mathbf {s i m} (v _ {i} , v _ {k}) / \\gamma)}, \\text {i f} i < B, \\\\ - \\log \\frac {\\exp (\\mathbf {s i m} (v _ {i} , v _ {B - i}) / \\gamma)}{\\sum_ {k \\neq i} ^ {2 B} \\exp (\\mathbf {s i m} (v _ {i} , v _ {k}) / \\gamma)}, \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {8}", + "image_path": "ea5988c5072f23e1aee35e8fbc861ff6d263a07977130b305af0e960363af447.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "text", + "content": "Differing from SimCLR [4] and RepRec [38], as shown in Fig. 3, here we set image " + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "text", + "content": " as the weakly augmented image and image " + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "inline_equation", + "content": "i + B" + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "text", + "content": " as the strongly augmented image instead of two randomly augmented images as the positive pair. We named this loss as " + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SWAPP}}" + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "text", + "content": ". SWAPP here refers to Strong-Weak Augmented Positive Pair. Finally, the total loss for fine-tuning CLIP on our pathology datasets is as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 241, + 546, + 354, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 546, + 354, + 559 + ], + "spans": [ + { + "bbox": [ + 241, + 546, + 354, + 559 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {t o t a l}} = \\mathcal {L} _ {\\mathrm {C L I P}} + \\mathcal {L} _ {\\mathrm {S W A P P}}.", + "image_path": "82747cd60211a9253265c41cdc8ee1e823f96d5d977718825e31049a44557118.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 241, + 561, + 481, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 561, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 241, + 561, + 481, + 590 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C L I P}} \\rightarrow \\frac {1}{2 B} \\sum_ {i = 1} ^ {B} \\left(l _ {i} ^ {v \\rightarrow t} + l _ {i} ^ {t \\rightarrow v}\\right), \\tag {9}", + "image_path": "a127925da0016ed1d850f5c97ec6062891aa4f8da7b200ca7feb28d92a84ccf6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 241, + 592, + 332, + 621 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 592, + 332, + 621 + ], + "spans": [ + { + "bbox": [ + 241, + 592, + 332, + 621 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {S W A P P}} \\rightarrow \\frac {1}{2 B} \\sum_ {i = 1} ^ {2 B} l _ {i} ^ {c}", + "image_path": "25990cd1241978100063f1882de3c7629041d48e95e4c2f8658b8f163a1b3952.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "We define this loss as Doublet Multimodal Contrastive Loss (DMCL) because the items involve both intra-contrastive (image and text) and inter-contrastive (image and image). The algorithm can be found in the Appendix." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Z. Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 141, + 146, + 471, + 313 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 2: Quantitative comparison on PCam. Precision, Recall, and F1-score refer to the macro-averaged values from all classes" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 146, + 471, + 313 + ], + "lines": [ + { + "bbox": [ + 141, + 146, + 471, + 313 + ], + "spans": [ + { + "bbox": [ + 141, + 146, + 471, + 313 + ], + "type": "table", + "html": "
Learning MannerLabeled RatioAlgorithmAccuracyPrecisionRecallF1-scoreAUC
Supervised100%-92.8 ± 0.3092.9 ± 0.2092.6 ± 0.2192.8 ± 0.200.95 ± 0.01
Supervised0.5%-45.1 ± 2.2742.0 ± 2.1433.6 ± 3.3037.3 ± 2.860.51 ± 0.01
Self-supervised0.5%SimCLR [4]60.4 ± 1.4563.3 ± 1.9058.5 ± 2.0160.8 ± 1.200.61 ± 0.04
Semi-supervised0.5%Pseudo-Label [20]55.7 ± 2.0559.2 ± 2.2354.5 ± 1.9756.8 ± 1.550.58 ± 0.03
FixMatch [28]73.2 ± 0.7677.5 ± 0.5073.7 ± 0.2575.6 ± 0.370.84 ± 0.04
Dash [37]71.0 ± 0.9875.3 ± 0.7870.2 ± 0.4572.7 ± 0.500.82 ± 0.02
FlexMatch [49]74.0 ± 0.8078.1 ± 1.1573.0 ± 0.9075.5 ± 0.840.85 ± 0.03
Multimodal CLIP0.5%Zero-shot [26]56.5 ± 0.0057.4 ± 0.0050.3 ± 0.0053.7 ± 0.000.60 ± 0.00
CoOp [54]63.6 ± 0.2563.9 ± 0.4262.5 ± 0.3563.0 ± 0.290.67 ± 0.02
CLIP-Adapter [9]72.3 ± 1.0277.2 ± 0.9563.2 ± 0.5669.4 ± 0.840.81 ± 0.02
TIP-Adapter [51]73.9 ± 0.5070.5 ± 0.8578.0 ± 0.4974.0 ± 0.950.84 ± 0.02
Path-CLIP81.9 ± 0.7879.4 ± 0.3585.0 ± 0.8082.1 ± 0.950.89 ± 0.02
Supervised1.5%-62.8 ± 1.3169.5 ± 0.6754.1 ± 2.2760.8 ± 1.680.63 ± 0.02
Self-supervised1.5%SimCLR [4]64.9 ± 1.0265.2 ± 1.4560.1 ± 1.7062.5 ± 1.050.64 ± 0.03
Semi-supervised1.5%Pseudo-Label [20]67.9 ± 1.8570.5 ± 1.9062.2 ± 1.7666.1 ± 1.420.67 ± 0.04
FixMatch [28]83.9 ± 1.2985.1 ± 0.8584.1 ± 0.9584.6 ± 0.900.87 ± 0.01
Dash [37]82.8 ± 1.0983.0 ± 0.5581.2 ± 0.6082.1 ± 0.780.85 ± 0.03
FlexMatch [49]84.1 ± 1.0284.9 ± 0.7584.4 ± 0.8084.6 ± 0.750.86 ± 0.04
Multimodal CLIP1.5%CoOp [54]61.5 ± 0.7560.5 ± 1.0365.6 ± 0.9063.1 ± 0.850.65 ± 0.02
CLIP-Adapter [9]74.1 ± 0.8879.1 ± 0.3465.4 ± 0.8971.6 ± 1.020.82 ± 0.01
TIP-Adapter [51]75.2 ± 0.9472.3 ± 0.6979.4 ± 0.8675.7 ± 0.730.83 ± 0.02
Path-CLIP83.1 ± 0.6782.8 ± 0.9582.0 ± 0.7582.4 ± 0.460.91 ± 0.02
", + "image_path": "29ddc90df9772bd44a77b694abda18227524cb8922c69ea8e76bd34b83748077.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 333, + 230, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 230, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 230, + 346 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 356, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 392 + ], + "type": "text", + "content": "Datasets. We select two classification datasets (PCam [31] and MHIST [34]) and a gigapixel grey/white matter segmentation dataset to evaluate Path-CLIP. The details of these datasets can be found in Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 398, + 482, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 482, + 471 + ], + "type": "text", + "content": "Experimental setup. For a fair comparison, we use ResNet-50 [12] as the backbone in the vision encoder " + }, + { + "bbox": [ + 130, + 398, + 482, + 471 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 130, + 398, + 482, + 471 + ], + "type": "text", + "content": " and BERT [6] as the textual encoder. We follow CLIP [26] to use gradient scaling to facilitate mixed-precision training. We set the learning rate as 0.0001. The batch size is 32. We use Adam [16] as the optimizer. To compare the computational complexity, we conduct all experiments on a single GPU (Nvidia RTX 2080Ti)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 475, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 609 + ], + "type": "text", + "content": "Baselines. We first implement traditional supervised learning (SL) methods with the entire dataset to explore the upper-bound performance for each task. Then we reduce the data usage and investigate the performance change of these SL methods. Then we compare Path-CLIP with original CLIP [26] and three recent CLIP-based fine-tuning methods: CoOp [54], Tip-Adapter [51] and CLIP-Adapter [9]. We compare them under the limited data scenarios to test their ability to do downstream tasks when the new dataset is small. Besides that, we also incorporate recent self-supervised learning (SimCLR [4]) and semi-supervised learning (Pseudo-Label [20], FixMatch [28], Dash [37], and FlexMatch [49]). Semi-supervised training requires a large amount of unlabeled data, which involves data collection efforts [41-43]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 624, + 224, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 224, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 224, + 635 + ], + "type": "text", + "content": "4.1 Main results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Classification task: PCam [31] and MHIST [34]. The main results for two classification benchmarks are summarized in Table 2 and Table 3. First, CLIP" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 447, + 103 + ], + "type": "text", + "content": "Path-CLIP: Efficiently Adapting CLIP for Pathology Image Analysis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 141, + 146, + 470, + 238 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 3: Quantitative comparison on MHIST. Precision, Recall, and F1-score refer to the macro-averaged values from all classes" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 146, + 470, + 238 + ], + "lines": [ + { + "bbox": [ + 141, + 146, + 470, + 238 + ], + "spans": [ + { + "bbox": [ + 141, + 146, + 470, + 238 + ], + "type": "table", + "html": "
Learning MannerLabeled RatioAlgorithmAccuracyPrecisionRecallF1-scoreAUC
Supervised100%-86.0 ± 0.3684.8 ± 0.6183.5 ± 0.5584.1 ± 0.580.89 ± 0.01
Self-supervised10%SimCLR [4]63.2 ± 0.9563.8 ± 1.2062.2 ± 1.7063.0 ± 0.850.66 ± 0.04
Semi-supervised10%Pseudo-Label [20]65.4 ± 1.0466.5 ± 0.9865.9 ± 0.8566.2 ± 1.420.75 ± 0.03
FixMatch [28]70.5 ± 1.5568.8 ± 2.7669.5 ± 2.9769.1 ± 2.860.79 ± 0.02
Dash [37]70.2 ± 1.2170.0 ± 0.8571.3 ± 0.6970.6 ± 1.480.81 ± 0.04
FlexMatch [49]73.1 ± 0.8474.6 ± 1.0575.1 ± 0.9074.8 ± 0.890.82 ± 0.03
Multimodal CLIP10%Zero-shot [26]36.9 ± 0.0036.9 ± 0.00100 ± 0.0053.9 ± 0.000.50 ± 0.00
CoOp [54]68.9 ± 0.4056.0 ± 0.6572.8 ± 0.5363.3 ± 0.860.76 ± 0.01
CLIP-Adapter [9]65.9 ± 1.4552.2 ± 1.2387.2 ± 0.8565.3 ± 0.950.77 ± 0.02
TIP-Adapter [51]63.1 ± 0.7849.9 ± 0.9094.7 ± 1.0265.4 ± 0.560.78 ± 0.03
Path-CLIP74.8 ± 1.9563.3 ± 1.4075.6 ± 0.9868.9 ± 1.500.84 ± 0.02
", + "image_path": "c8206ae796a872e1f93e3452dd12f1aa8ebe79d70210508515d1c9395f541b02.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "content": "shows its promising zero-shot ability: when we directly test CLIP on PCam [31], it can get " + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "inline_equation", + "content": "56.5\\%" + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "content": " accuracy. Then, we compare Path-CLIP with other CLIP-based fine-tuning methods. In PCam [31], we present " + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "content": " data usage in the training set to fine-tune CLIP, respectively. We find all of these fine-tuning algorithms [9, 51, 54] can improve the performance of the original CLIP with the new data from the downstream task presented. Path-CLIP can outperform them in both settings. Specifically, Path-CLIP outperforms TIP-Adapter [51] with " + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "inline_equation", + "content": "7.9\\%" + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "content": " accuracy when only " + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "content": " data are used, respectively. Path-CLIP also shows similar improvement in MHIST [34] (Table 3). We also implemented several semi-supervised learning where the rest of the data in the training set are used as an unlabeled set. Although Path-CLIP does not use unlabeled data, it can achieve competitive results with semi-supervised learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 419, + 482, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 482, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 482, + 492 + ], + "type": "text", + "content": "Besides these ratios of data usage, we also visualize the results by varying the ratio to show the performance of each fine-tuning algorithm. As shown in Fig. 4, CoOp [54] may not improve PCam when more data are used to fine-tune the prompt. CoOp freezes the encoders and only fine-tunes the language prompt part, while other algorithms also fine-tune the vision part. We conclude that visual fine-tuning is more critical compared to prompt tuning." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 148, + 518, + 302, + 625 + ], + "blocks": [ + { + "bbox": [ + 148, + 518, + 302, + 625 + ], + "lines": [ + { + "bbox": [ + 148, + 518, + 302, + 625 + ], + "spans": [ + { + "bbox": [ + 148, + 518, + 302, + 625 + ], + "type": "image", + "image_path": "859cb8c799e940407e8f279e525652182d3daf3db9a0cbd9f05860f9ee502eaf.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 630, + 248, + 641 + ], + "lines": [ + { + "bbox": [ + 208, + 630, + 248, + 641 + ], + "spans": [ + { + "bbox": [ + 208, + 630, + 248, + 641 + ], + "type": "text", + "content": "(a) PCam" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 518, + 460, + 625 + ], + "blocks": [ + { + "bbox": [ + 307, + 518, + 460, + 625 + ], + "lines": [ + { + "bbox": [ + 307, + 518, + 460, + 625 + ], + "spans": [ + { + "bbox": [ + 307, + 518, + 460, + 625 + ], + "type": "image", + "image_path": "47c058cc12f1284798227d7cbeb7ace6e0a56b57a4298d346f97b41d9c3cbc07.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 629, + 406, + 641 + ], + "lines": [ + { + "bbox": [ + 360, + 629, + 406, + 641 + ], + "spans": [ + { + "bbox": [ + 360, + 629, + 406, + 641 + ], + "type": "text", + "content": "(b) MHIST" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 152, + 651, + 459, + 662 + ], + "lines": [ + { + "bbox": [ + 152, + 651, + 459, + 662 + ], + "spans": [ + { + "bbox": [ + 152, + 651, + 459, + 662 + ], + "type": "text", + "content": "Fig. 4: Comparison of different fine-tuning methods by varying data usage." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "Z. Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 136, + 445, + 301 + ], + "blocks": [ + { + "bbox": [ + 178, + 114, + 433, + 126 + ], + "lines": [ + { + "bbox": [ + 178, + 114, + 433, + 126 + ], + "spans": [ + { + "bbox": [ + 178, + 114, + 433, + 126 + ], + "type": "text", + "content": "Table 4: Performance comparison on GM/WM segmentation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 136, + 445, + 301 + ], + "lines": [ + { + "bbox": [ + 167, + 136, + 445, + 301 + ], + "spans": [ + { + "bbox": [ + 167, + 136, + 445, + 301 + ], + "type": "table", + "html": "
Data UsageAlgorithmGM_IoUGM_DICEWM_IoUWM_DICE
100% (SL)FCN [1]71.7283.2149.0966.08
U-Net [25]91.5295.1782.0288.15
0.1% (SL)FCN [1]48.2525.0124.5430.02
U-Net [25]46.1350.105.7512.23
0.1% (CLIP)CoOp [54]62.2365.7039.9248.65
CLIP-Adapter [9]67.8074.3243.2554.61
TIP-Adapter [51]69.9275.5046.2356.60
Path-CLIP78.0583.7054.5063.23
1% (SL)FCN [1]51.3430.2031.0235.12
U-Net [25]50.2453.2315.4325.58
1% (CLIP)CoOp [54]63.4564.1237.8545.34
CLIP-Adapter [9]69.9976.5447.9558.34
TIP-Adapter [51]72.1277.3055.6563.60
Path-CLIP80.2188.7060.5070.34
", + "image_path": "10ee9e23232ed5b0e484192175653909a2e8c14e165e3d07c7808797f2a58dac.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 325, + 482, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 447 + ], + "type": "text", + "content": "GM/WM segmentation [17]. Besides two classification tasks, we follow [17] to generalize our method to a weakly-supervised segmentation problem. We select IoU and DICE scores as the performance metrics. The results are summarized in Table 4. We select FCN [1] and U-Net [25] as two SL baselines: we first train them on all data (100%) in a supervised manner, then we reduce the dataset to 0.1% and 1% setting and train the model again. We can see both SL algorithms face performance degradation when the training data is scarce. They require an extensive and well-curated labeled dataset to get satisfactory results. We then focus on fine-tuning CLIP algorithms. Table 4 shows Path-CLIP can achieve superior results than other fine-tuning techniques." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 467, + 244, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 467, + 244, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 467, + 244, + 479 + ], + "type": "text", + "content": "4.2 Ablation studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 490, + 482, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 587 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 587 + ], + "type": "text", + "content": "Importance of alleviating overfitting. Since the new task may have a very limited training set, it is crucial to alleviate the potential overfitting issue. Besides the RFR with the self-adaptive ratio, we also propose HRP, DVC, and DMCL to cope with the overfitting issue. Table 5 summarizes the contribution of each proposed module: HRP can improve with " + }, + { + "bbox": [ + 130, + 490, + 482, + 587 + ], + "type": "inline_equation", + "content": "+2.2\\%" + }, + { + "bbox": [ + 130, + 490, + 482, + 587 + ], + "type": "text", + "content": " accuracy and DVC can improve with " + }, + { + "bbox": [ + 130, + 490, + 482, + 587 + ], + "type": "inline_equation", + "content": "+4.3\\%" + }, + { + "bbox": [ + 130, + 490, + 482, + 587 + ], + "type": "text", + "content": ". DMCL inspired by self-supervised learning can also aid in great improvements. We conclude that contrastive learning, perturbations on data and hidden representations are helpful to the overfitting issue." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 593, + 483, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 483, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 483, + 667 + ], + "type": "text", + "content": "Computational efficiency. When we range the data usage from " + }, + { + "bbox": [ + 130, + 593, + 483, + 667 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 130, + 593, + 483, + 667 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 593, + 483, + 667 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 593, + 483, + 667 + ], + "type": "text", + "content": ", Path-CLIP can obtain " + }, + { + "bbox": [ + 130, + 593, + 483, + 667 + ], + "type": "inline_equation", + "content": "+5.5\\%" + }, + { + "bbox": [ + 130, + 593, + 483, + 667 + ], + "type": "text", + "content": " gain in accuracy while only introducing 1 minute of additional training time. However, CoOp [54] has the overfitting issue and obtains a lower score while using 53 minutes, almost five times compared to the time used by RFR. The situation remains similar under the " + }, + { + "bbox": [ + 130, + 593, + 483, + 667 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 593, + 483, + 667 + ], + "type": "text", + "content": " of data usage. We conclude our proposed RFR can get over " + }, + { + "bbox": [ + 130, + 593, + 483, + 667 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 130, + 593, + 483, + 667 + ], + "type": "text", + "content": " improvement on CLIP by only" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 167, + 90, + 447, + 103 + ], + "type": "text", + "content": "Path-CLIP: Efficiently Adapting CLIP for Pathology Image Analysis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 219, + 136, + 394, + 198 + ], + "blocks": [ + { + "bbox": [ + 168, + 114, + 444, + 126 + ], + "lines": [ + { + "bbox": [ + 168, + 114, + 444, + 126 + ], + "spans": [ + { + "bbox": [ + 168, + 114, + 444, + 126 + ], + "type": "text", + "content": "Table 5: Ablation study on PCam using " + }, + { + "bbox": [ + 168, + 114, + 444, + 126 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 168, + 114, + 444, + 126 + ], + "type": "text", + "content": " data for fine-tuning." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 219, + 136, + 394, + 198 + ], + "lines": [ + { + "bbox": [ + 219, + 136, + 394, + 198 + ], + "spans": [ + { + "bbox": [ + 219, + 136, + 394, + 198 + ], + "type": "table", + "html": "
#RFCHRPDVCDMCLAccuracyF1-scoreAUC
1XXXX56.553.70.60
2✓XXX72.070.50.69
3✓✓XX74.275.50.78
4✓✓✓X78.579.30.84
5✓✓✓✓81.982.10.89
", + "image_path": "6b7f0952896b94d05b3621a3b719f31c6efbff3424de5b3cec7c1099640f1367.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 219, + 481, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 219, + 481, + 244 + ], + "spans": [ + { + "bbox": [ + 130, + 219, + 481, + 244 + ], + "type": "text", + "content": "using 10 minutes for the fine-tuning on a single GPU, which is promising for digital pathology research. The details can be found in Table 8 in the Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 248, + 481, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 481, + 297 + ], + "type": "text", + "content": "Generalization to pathology foundation model: PLIP [14]. The results of using PLIP's weights are in Appendix 3.6: PLIP is pre-trained on a large-scale pathology dataset and achieves strong zero-shot performance. With our method, we can further boost the performance and outperform other tuning methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 313, + 217, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 313, + 217, + 326 + ], + "spans": [ + { + "bbox": [ + 131, + 313, + 217, + 326 + ], + "type": "text", + "content": "5 Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 336, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 456 + ], + "type": "text", + "content": "In this work, we explore the generalization of CLIP in pathology image analysis. To alleviate the overfitting and forgetting issues when adapting CLIP to small downstream datasets, we propose self-adaptive RFR with HRP. We propose Path-CLIP with DMCL as an end-to-end pipeline for CLIP adaptation in pathology tasks. The results reveal the superiority and effectiveness of Path-CLIP compared to the existing fine-tuning methods. In addition, the less computational time improves the efficiency and accessibility of Path-CLIP. Hence, we conclude Path-CLIP is more robust and has the potential to bridge the domain shift between the pre-trained natural images and pathology images in an efficient and cost-effective manner." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 472, + 241, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 241, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 241, + 487 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 495, + 481, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 481, + 556 + ], + "type": "text", + "content": "This work was supported by the Noyce Initiative UC Partnerships in Computational Transformation Grant and Child Family Endowed Professorship. Resources for this study were funded in part by grants from the National Institute on Aging of the National Institutes of Health under Award Numbers R01AG062517, P30AG072972, and R01AG056519." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 572, + 198, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 572, + 198, + 585 + ], + "spans": [ + { + "bbox": [ + 132, + 572, + 198, + 585 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 136, + 588, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 136, + 588, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 588, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 136, + 588, + 482, + 632 + ], + "type": "text", + "content": "1. Bandy, P., van de Loo, R., Intezar, M., Geijs, D., Ciompi, F., van Ginneken, B., van der Laak, J., Litjens, G.: Comparison of different methods for tissue segmentation in histopathological whole-slide images. In: Proc. 2017 IEEE Int. Symp. Biomed. Imaging. pp. 591-595 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 136, + 632, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 632, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 136, + 632, + 482, + 665 + ], + "type": "text", + "content": "2. Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., Raffel, C.A.: MixMatch: A holistic approach to semi-supervised learning. In: NeurIPS. vol. 32 (2019)" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Z. Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 136, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 136, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 482, + 149 + ], + "type": "text", + "content": "3. Chen, F., Zhang, H., Li, Z., Dou, J., Mo, S., Chen, H., Zhang, Y., Ahmed, U., Zhu, C., Savvides, M.: Unitail: Detecting, reading, and matching in retail scene. In: ECCV. pp. 705-722. Springer (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 138, + 150, + 481, + 171 + ], + "type": "text", + "content": "4. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML. pp. 1597-1607 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 171, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 171, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 138, + 171, + 481, + 203 + ], + "type": "text", + "content": "5. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Workshop. pp. 702-703 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 203, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 203, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 138, + 203, + 481, + 236 + ], + "type": "text", + "content": "6. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 236, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 236, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 236, + 481, + 267 + ], + "type": "text", + "content": "7. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 267, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 481, + 300 + ], + "type": "text", + "content": "8. Gal, R., Patashnik, O., Maron, H., Bermano, A.H., Chechik, G., Cohen-Or, D.: Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG) 41(4), 1-13 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 300, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 481, + 332 + ], + "type": "text", + "content": "9. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 332, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 481, + 354 + ], + "type": "text", + "content": "0. Ge, C., Huang, R., Xie, M., Lai, Z., Song, S., Li, S., Huang, G.: Domain adaptation via prompt learning. arXiv preprint arXiv:2202.06687 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 354, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 481, + 374 + ], + "type": "text", + "content": "1. He, G., Chen, J., Zhu, J.: Preserving pre-trained features helps calibrate fine-tuned language models. In: ICLR (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 374, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 481, + 396 + ], + "type": "text", + "content": "2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 396, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 481, + 429 + ], + "type": "text", + "content": "3. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: ICML. pp. 2790-2799. PMLR (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 429, + 481, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 481, + 449 + ], + "type": "text", + "content": "4. Huang, Z., other: A visual-language foundation model for pathology image analysis using medical twitter. Nature medicine (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 449, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 481, + 482 + ], + "type": "text", + "content": "5. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML. pp. 4904-4916 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 482, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 482, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 138, + 482, + 481, + 504 + ], + "type": "text", + "content": "6. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (Poster) (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 504, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 504, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 504, + 481, + 546 + ], + "type": "text", + "content": "7. Lai, Z., Oliveira, L.C., Guo, R., Xu, W., Hu, Z., Mifflin, K., Decarli, C., Cheung, S.C., Chuah, C.N., Dugger, B.N.: Brainsec: Automated brain tissue segmentation pipeline for scalable neuropathological analysis. IEEE Access 10, 49064-49079 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 547, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 481, + 579 + ], + "type": "text", + "content": "8. Lai, Z., Wang, C., Oliveira, L.C., Dugger, B.N., Cheung, S.C., Chuah, C.N.: Joint semi-supervised and active learning for segmentation of gigapixel pathology images with cost-effective labeling. In: ICCV Workshop. pp. 591-600 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 579, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 481, + 601 + ], + "type": "text", + "content": "9. Lam, S.Y., Zeng, Q., Zhang, K., You, C., Voigt, R.: Large language models are partially primed in pronoun interpretation. arXiv preprint arXiv:2305.16917 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "type": "text", + "content": "20. Lee, D.H.: Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In: Workshop on challenges in representation learning (ICML) (2013)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "text", + "content": "21. Liu, M., Hu, L., Tang, Y., Wang, C., He, Y., Zeng, C., Lin, K., He, Z., Huo, W.: A deep learning method for breast cancer classification in the pathology images. IEEE J. Biomed. Health. Inf. 26(10), 5025-5032 (2022)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "text", + "content": "Path-CLIP: Efficiently Adapting CLIP for Pathology Image Analysis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 138 + ], + "type": "text", + "content": "22. Lu, Y., Liu, J., Zhang, Y., Liu, Y., Tian, X.: Prompt distribution learning. In: CVPR. pp. 5206-5215 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 482, + 182 + ], + "type": "text", + "content": "23. Lyu, W., Dong, X., Wong, R., Zheng, S., Abell-Hart, K., Wang, F., Chen, C.: A multimodal transformer: Fusing clinical notes with structured ehr data for interpretable in-hospital mortality prediction. In: AMIA Annual Symposium Proceedings. vol. 2022, p. 719. American Medical Informatics Association (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 203 + ], + "type": "text", + "content": "24. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "type": "text", + "content": "25. Oskal, K.R., Risdal, M., Janssen, E.A., Undersrud, E.S., Gulsrud, T.O.: A U-net based approach to epidermal tissue segmentation in whole slide histopathological images. SN Appl. Sci. 1(7), 1-12 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "type": "text", + "content": "26. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "type": "text", + "content": "27. Rasheed, H., khattak, M.U., Maaz, M., Khan, S., Khan, F.S.: Finetuned clip models are efficient video learners. In: The IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "text", + "content": "28. Sohn, K., Berthelot, D., Carlini, N., Zhang, Z., Zhang, H., Raffel, C.A., Cubuk, E.D., Kurakin, A., Li, C.L.: FixMatch: Simplifying semi-supervised learning with consistency and confidence. In: NeurIPS. vol. 33, pp. 596-608 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "type": "text", + "content": "29. Tian, Y., Newsam, S., Boakye, K.: Fashion image retrieval with text feedback by additive attention compositional learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1011-1021 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 365, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 386 + ], + "type": "text", + "content": "30. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "text", + "content": "31. Veeling, B.S., Linmans, J., Winkens, J., Cohen, T., Welling, M.: Rotation equivariant cnns for digital pathology. In: MICCAI. pp. 210-218. Springer (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "type": "text", + "content": "32. Wang, X., Yang, S., Zhang, J., Wang, M., Zhang, J., Huang, J., Yang, W., Han, X.: Transpath: Transformer-based self-supervised learning for histopathological image classification. In: MICCAI. pp. 186-195. Springer (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 439, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 460 + ], + "type": "text", + "content": "33. Wang, Y., Ma, X., Chen, Z., Luo, Y., Yi, J., Bailey, J.: Symmetric cross entropy for robust learning with noisy labels. In: ICCV (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 460, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 481, + 493 + ], + "type": "text", + "content": "34. Wei, J., Suriawinata, A., Ren, B., Liu, X., Lisovsky, M., Vaickus, L., Brown, C., Baker, M., Tomita, N., Torresani, L., et al.: A petri dish for histopathology image analysis. In: Int. Conf. Arti. Intell. in Medi. pp. 11-24. Springer (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 493, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 481, + 536 + ], + "type": "text", + "content": "35. Wu, C., Wu, F., Qi, T., Huang, Y.: Noisytune: A little noise can help you finetune pretrained language models better. In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). pp. 680-685 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "type": "text", + "content": "36. Xu, P., Zhu, X., Clifton, D.A.: Multimodal learning with transformers: A survey. arXiv preprint arXiv:2206.06488 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 557, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 481, + 589 + ], + "type": "text", + "content": "37. Xu, Y., Shang, L., Ye, J., Qian, Q., Li, Y.F., Sun, B., Li, H., Jin, R.: Dash: Semi-supervised learning with dynamic thresholding. In: International Conference on Machine Learning. pp. 11525-11536. PMLR (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 589, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 633 + ], + "type": "text", + "content": "38. Yan, X., Naushad, J., Sun, S., Han, K., Tang, H., Kong, D., Ma, H., You, C., Xie, X.: Representation recovering for self-supervised pre-training on medical images. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2685-2695 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "39. Yang, B., Liu, F., Li, Z., Yin, Q., You, C., Yin, B., Zou, Y.: Multimodal prompt learning for product title generation with extremely limited labels. arXiv preprint arXiv:2307.01969 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "Z. Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 599 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "40. Yang, J., Chen, H., Liang, Y., Huang, J., He, L., Yao, J.: Concl: Concept contrastive learning for dense prediction pre-training in pathology images. In: ECCV. pp. 523-539. Springer (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "text", + "content": "41. You, C., Dai, W., Min, Y., Liu, F., Zhang, X., Feng, C., Clifton, D.A., Zhou, S.K., Staib, L.H., Duncan, J.S.: Rethinking semi-supervised medical image segmentation: A variance-reduction perspective. In: NeurIPS (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 482, + 226 + ], + "type": "text", + "content": "42. You, C., Dai, W., Min, Y., Staib, L., Duncan, J.S.: Bootstrapping semi-supervised medical image segmentation with anatomical-aware contrastive distillation. In: International Conference on Information Processing in Medical Imaging. pp. 641-653. Springer (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 482, + 259 + ], + "type": "text", + "content": "43. You, C., Dai, W., Min, Y., Staib, L., Sekhon, J., Duncan, J.S.: Action++: Improving semi-supervised medical image segmentation with adaptive anatomical contrast. arXiv preprint arXiv:2304.02689 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 482, + 293 + ], + "type": "text", + "content": "44. Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., Wu, Y.: Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 482, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 482, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 482, + 325 + ], + "type": "text", + "content": "45. Yuan, H., Yuan, Z., Tan, C., Huang, F., Huang, S.: Hype: Better pre-trained language model fine-tuning with hidden representation perturbation. arXiv preprint arXiv:2212.08853 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 325, + 482, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 482, + 358 + ], + "type": "text", + "content": "46. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 482, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 482, + 392 + ], + "type": "text", + "content": "47. Zhai, X., Kolesnikov, A., Houlsby, N., Beyer, L.: Scaling vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12104-12113 (June 2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 482, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 482, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 482, + 423 + ], + "type": "text", + "content": "48. Zhai, X., Wang, X., Mustafa, B., Steiner, A., Keysers, D., Kolesnikov, A., Beyer, L.: Lit: Zero-shot transfer with locked-image text tuning. In: CVPR. pp. 18123-18133 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 423, + 482, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 423, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 423, + 482, + 456 + ], + "type": "text", + "content": "49. Zhang, B., Wang, Y., Hou, W., Wu, H., Wang, J., Okumura, M., Shinozaki, T.: Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. NeurIPS 34 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 456, + 482, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 482, + 490 + ], + "type": "text", + "content": "50. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8552-8562 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 490, + 482, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 482, + 522 + ], + "type": "text", + "content": "51. Zhang, R., Zhang, W., Fang, R., Gao, P., Li, K., Dai, J., Qiao, Y., Li, H.: Tip-adapter: Training-free adaption of clip for few-shot classification. In: ECCV. pp. 493-510. Springer (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 522, + 482, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 482, + 544 + ], + "type": "text", + "content": "52. Zhang, T., Wu, F., Katiyar, A., Weinberger, K.Q., Artzi, Y.: Revisiting few-sample bert fine-tuning. In: ICLR (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 544, + 482, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 482, + 578 + ], + "type": "text", + "content": "53. Zhang, W., Zhu, L., Hallinan, J., Zhang, S., Makmur, A., Cai, Q., Ooi, B.C.: Boostmis: Boosting medical image semi-supervised learning with adaptive pseudo labeling and informative active annotation. In: CVPR. pp. 20666-20676 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 578, + 482, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 482, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 482, + 599 + ], + "type": "text", + "content": "54. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. Int. J. Comput. Vision 130(9), 2337-2348 (2022)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "text", + "content": "Path-CLIP: Efficiently Adapting CLIP for Pathology Image Analysis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/BrushNet_ A Plug-and-Play Image Inpainting Model with Decomposed Dual-Branch Diffusion/8cdf68e6-b7ca-45f7-9a98-9bdc1df71997_content_list.json b/2024/BrushNet_ A Plug-and-Play Image Inpainting Model with Decomposed Dual-Branch Diffusion/8cdf68e6-b7ca-45f7-9a98-9bdc1df71997_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8bca8ca17e2a4fd55d5f30c90fcde2087efecfa7 --- /dev/null +++ b/2024/BrushNet_ A Plug-and-Play Image Inpainting Model with Decomposed Dual-Branch Diffusion/8cdf68e6-b7ca-45f7-9a98-9bdc1df71997_content_list.json @@ -0,0 +1,1885 @@ +[ + { + "type": "text", + "text": "BrushNet: A Plug-and-Play Image Inpainting Model with Decomposed Dual-Branch Diffusion", + "text_level": 1, + "bbox": [ + 222, + 141, + 772, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xuan $\\mathrm{Ju}^{1,2}$ , Xian Liu $^{1,2}$ , Xintao Wang $^{1\\star}$ , Yuxuan Bian $^{2}$ , Ying Shan $^{1}$ , and Qiang $\\mathrm{Xu}^{2\\star}$", + "bbox": [ + 215, + 198, + 784, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ ARC Lab, Tencent PCG $^{2}$ The Chinese University of Hong Kong", + "bbox": [ + 279, + 217, + 723, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/TencentARC/BrushNet", + "bbox": [ + 352, + 234, + 648, + 247 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9a73c4b7f103f0cd858bc52bb37b16167d1524cc4cd0eaf471b5413c80b009c8.jpg", + "image_caption": [ + "Fig. 1: Performance comparisons of BrushNet and previous image inpainting methods across various inpainting tasks: (I) Random Mask (< 50% masked), (II) Random Mask (> 50% masked), (III) Segmentation Mask Inside-Inpainting, (IV) Segmentation Mask Outside-Inpainting. Each group of results contains an artificial image (left) and a natural image (right) with 6 inpainting methods: (b) Blended Latent Diffusion (BLD) [1], (c) Stable Diffusion Inpainting (SDI) [33], (d) HD-Painter (HDP) [25], (e) PowerPaint (PP) [56], (f) ControlNet-Inpainting (CNI) [51], and (g) Ours." + ], + "image_footnote": [], + "bbox": [ + 228, + 253, + 764, + 438 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Image inpainting, the process of restoring corrupted images, has seen significant advancements with the advent of diffusion models (DMs). Despite these advancements, current DM adaptations for inpainting, which involve modifications to the sampling strategy or the development of inpainting-specific DMs, frequently suffer from semantic inconsistencies and reduced image quality. Addressing these challenges, our work introduces a novel paradigm: the division of masked image features and noisy latent into separate branches. This division dramatically diminishes the model's learning load, facilitating a nuanced incorporation of essential masked image information in a hierarchical fashion. Herein, we present BrushNet, a novel plug-and-play dual-branch model engineered to embed pixel-level masked image features into any pre-trained DM, guaranteeing coherent and enhanced image inpainting outcomes. Additionally, we introduce BrushData and BrushBench to facilitate segmentation-based inpainting training and performance assessment. Our extensive experimental analysis demonstrates BrushNet's superior performance over existing models across seven key metrics, including image quality, mask region preservation, and textual coherence.", + "bbox": [ + 259, + 547, + 743, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Image Inpainting $\\cdot$ Diffusion Models $\\cdot$ Image Generation", + "bbox": [ + 261, + 805, + 723, + 820 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 218, + 825, + 385, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image inpainting [45] aims at restoring the missing regions of an image while maintaining the overall coherence. As a long-standing computer vision problem, it facilitates numerous applications such as virtual try-on [18] and image editing [15]. Recently, diffusion models [12,36] have demonstrated impressive performance in image generation, enabling flexible user control with semantic and structural conditions [33,51]. To this end, researchers resort to diffusion-based pipelines for high-quality image inpainting that aligns with given text prompts.", + "bbox": [ + 212, + 170, + 787, + 276 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Commonly used diffusion-based text-guided inpainting methods can be roughly divided into two categories: (1) Sampling strategy modification [1,2,6,20,23,47, 50], which modifies the standard denoising process by sampling the masked regions from a pre-trained diffusion model, and the unmasked areas are simply copy-pasted from the given image in each denoising step. Although they can be used in arbitrary diffusion backbones, the limited perceptual knowledge of mask boundaries and the unmasked image region context leads to incoherent inpainting results. (2) Dedicated inpainting models [4, 33, 37, 42, 43, 46, 49, 56], which fine-tune a specially designed image inpainting model by expanding the input channel dimension of base diffusion models to incorporate provided corrupted image and mask. While they enable the diffusion model to generate more satisfying results with specialized content-aware and shape-aware models, we argue, is this architecture the best fit for diffusion-based inpainting?", + "bbox": [ + 212, + 277, + 792, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Fig. 2, dedicated inpainting models fuse noisy latent, masked image latent, mask, and text at an early stage. This architectural design makes the masked image feature easily influenced by the text embedding, preventing subsequent layers in the UNet from obtaining pure masked image features due to the influence of text. Additionally, handling the condition and generation in a single branch imposes extra burdens on the UNet framework. These approaches also necessitate fine-tuning in different variations of diffusion backbones, which can be time-consuming with limited transferability.", + "bbox": [ + 212, + 474, + 787, + 595 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c3c37ed7f16079a11ef5a942c52021579f3d07987efc14995ca68236ffa87cb2.jpg", + "image_caption": [ + "Fig. 2: Comparison of previous inpainting architectures and BrushNet." + ], + "image_footnote": [], + "bbox": [ + 230, + 603, + 772, + 770 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adding an additional branch dedicated to masked image feature extraction could be a promising approach to tackle the issue above. However, the existing solutions such as ControlNet [51] lead to inadequate information extraction", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Xuan Ju, et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and insertion when directly applied to inpainting, which originates from differences between controllable image generation and inpainting: inpainting requires pixel-to-pixel constraints with strong restrictive information instead of sparse structural control relying on text for content completion. As a result, ControlNet yields unsatisfactory results compared to dedicated inpainting models.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this issue, we propose BrushNet, which introduces an additional branch to the diffusion framework, creating a more suitable architecture for image inpainting. Specifically, our designs are three-fold: (1) To improve the extraction of image features for adaptation to the UNet distribution, we use a VAE encoder instead of randomly initialized convolution layers to process the masked image. (2) To enable dense per-pixel control, we adopt a hierarchical approach by gradually incorporating the full UNet feature layer-by-layer into the pre-trained UNet. (3) To ensure pure image information is considered in the additional branch, we remove text cross-attention from UNet. This design further offers plug-and-play capabilities and flexible unmasked region controllability to the inpainting process. For better consistency and a larger range of unmasked region controllability, we additionally propose a blurred blending strategy.", + "bbox": [ + 212, + 222, + 784, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To ensure a comprehensive evaluation for real-life applications, we categorize inpainting tasks into two distinct types based on mask shape: random brush masks and segmentation-based masks. We utilize EditBench [37] as the comparing benchmark for random brush mask inpainting. Additionally, we introduce a new training dataset BrushData and a new benchmark BrushBench for training and evaluating segmentation-based mask inpainting. Results show BrushNet achieve state-of-the-art performance across 7 metrics encompassing image quality, masked region preservation, and text alignment.", + "bbox": [ + 212, + 404, + 785, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 535, + 387, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image inpainting is a classic problem in computer vision, aiming to restore masked regions of an image with plausible and natural content [29, 45]. Previous methods based on traditional techniques [3, 7], Variational Auto-Encoders (VAEs) [27, 54] and Generative Adversarial Networks (GANs) [21, 53, 55] often require auxiliary hand-engineered features but yield poor results. Recently, diffusion-based methods [1, 2, 20, 23, 31, 42] gain popularity due to their impressive high-quality generation, fine-grained control, and output diversity [12, 13, 33].", + "bbox": [ + 212, + 553, + 784, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Initial attempts at utilizing diffusion models for text-guided inpainting [1, 2,6,20,23,47,50], such as Blended Latent Diffusion, modifies the standard denoising strategy by sampling the masked regions from a pre-trained diffusion model and the unmasked areas from the given image, which is commonly used as the default inpainting choice in widely-used image generation libraries like Diffusers [28]. Although these methods demonstrate satisfactory results in simple image inpainting tasks and can be plug-and-play to any diffusion model, they struggle with complex mask shapes, image contents, and text prompts, leading to results that lack coherence. This is primarily attributed to their limited perceptual knowledge of mask boundaries and the unmasked image region context.", + "bbox": [ + 212, + 659, + 785, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous works [4,33,37,42,43,46,49,56] address this issue by fine-tuning the base models into content-aware and shape-aware models specifically designed", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "BrushNet", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for image inpainting. Specifically, SmartBrush [42] augments the diffusion U-Net with object-mask prediction to guide the sampling process with mask boundaries information. Stable Diffusion Inpainting [33] fine-tunes a diffusion model specifically designed for inpainting tasks, taking the mask, masked image, and noisy latent as inputs to UNet architecture. HD-Painter [25] and PowerPaint [56] build upon the foundation of Stable Diffusion Inpainting, separately enhancing the generation quality and enabling the model to perform multiple tasks.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, these approaches make it difficult to effectively transfer their inpainting ability to arbitrary pre-trained models, restricting their applicability. To enable any diffusion model with inpainting capabilities, the community fine-tunes ControlNet [51] on inpainting image pairs. However, the model design of ControlNet exhibits limitations in its perceptual understanding of masks and masked images, which consequently leads to unsatisfactory outcomes. Compared with previous methods (shown in Tab. 1), BrushNet is plug-and-play, content-aware, and shape-aware, with a flexible preserving degree for unmasked regions.", + "bbox": [ + 212, + 252, + 787, + 373 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/f64c023054c88432abf10266339c6ce7e8e626b01edbd5ef9ce70979fa106262.jpg", + "table_caption": [ + "Table 1: Comparison of BrushNet with Previous Image Inpainting Methods. BrushNet offers the advantage of being plug-and-play with any pretrained diffusion model. Moreover, it allows for flexible control over the scale of inpainting and is designed to be aware of both the mask shape and the unmasked content. Note that we only list commonly used text-guided diffusion methods in this table." + ], + "table_footnote": [], + "table_body": "
ModelPlug-and-PlayFlexible-ScaleContent-AwareShape-Aware
Blended Diffusion [1,2]✓
SmartBrush [42]✓
SD Inpainting [33]✓✓
PowerPaint [56]✓✓
HD-Painter [25]✓✓
ReplaceAnything [4]✓✓
Imagen [37]✓✓
ControlNet-Inpainting [51]✓✓✓
BrushNet✓✓✓✓
", + "bbox": [ + 232, + 452, + 781, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Preliminaries and Motivation", + "text_level": 1, + "bbox": [ + 214, + 593, + 535, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we will first introduce diffusion models in Sec. 3.1. Then, Sec. 3.2 would review previous inpainting techniques based on sampling strategy modification and special training. Finally, the motivation is outlined in Section 3.3.", + "bbox": [ + 212, + 612, + 785, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Diffusion Models", + "text_level": 1, + "bbox": [ + 214, + 662, + 401, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion models include a forward process that adds Gaussian noise $\\epsilon$ to convert clean sample $z_0$ to noise sample $z_T$ , and a backward process that iteratively performs denoising from $z_T$ to $z_0$ , where $\\epsilon \\sim \\mathcal{N}(0,1)$ , and $T$ represents the total number of timesteps. The forward process can be formulated as:", + "bbox": [ + 212, + 683, + 787, + 743 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} = \\sqrt {\\alpha_ {t}} z _ {0} + \\sqrt {1 - \\alpha_ {t}} \\epsilon \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 757, + 785, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$z_{t}$ is the noised feature at step $t$ with $t\\sim [1,T]$ , and $\\alpha$ is a hyper-parameter.", + "bbox": [ + 238, + 779, + 785, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the backward process, given input noise $z_{T}$ sampled from a random Gaussian distribution, learnable network $\\epsilon_{\\theta}$ estimates noise at each step $t$ conditioned on $C$ . After $T$ progressively refining iterations, $z_{0}$ is derived as the output sample:", + "bbox": [ + 212, + 795, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Xuan Ju, et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t - 1} = \\frac {\\sqrt {\\alpha_ {t - 1}}}{\\sqrt {\\alpha_ {t}}} z _ {t} + \\sqrt {\\alpha_ {t - 1}} \\left(\\sqrt {\\frac {1}{\\alpha_ {t - 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {t}} - 1}\\right) \\epsilon_ {\\theta} \\left(z _ {t}, t, C\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 272, + 170, + 785, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The training of diffusion models revolves around optimizing the denoiser network $\\epsilon_{\\theta}$ to conduct denoising with condition $C$ , guided by the objective:", + "bbox": [ + 215, + 212, + 782, + 242 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\theta} {\\min } E _ {z _ {0}, \\epsilon \\sim \\mathcal {N} (0, I), t \\sim U (1, T)} \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, C) \\| \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 251, + 785, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Previous Inpainting Models", + "text_level": 1, + "bbox": [ + 215, + 280, + 491, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sampling Strategy Modification. This line of research achieves inpainting by gradually blending masked images with the generated results. The most used method among them is Blended Latent Diffusion (BLD) [1], serving as the default choice for inpainting in widely-used diffusion-based image generation libraries (e.g., Diffusers [28]). Given a binary mask $m$ and a masked image $x_0^{masked}$ , BLD first extracts the latent representation $z_0^{masked}$ of the masked image using VAE. Subsequently, the mask $m$ is resized to $m^{resized}$ to match the size of the latent representation. To formulate the inpainting process, BLD adds Gaussian noise to $z_0^{masked}$ for $T$ steps and gets $z_t^{masked}$ , where $t \\sim [1,T]$ . Then, denoising steps start from $z_T^{masked}$ , where each sampling step in eq. 2 is followed by:", + "bbox": [ + 212, + 301, + 784, + 452 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t - 1} \\leftarrow z _ {t - 1} \\cdot \\left(1 - m ^ {\\text {r e s i z e d}}\\right) + z _ {t - 1} ^ {\\text {m a s k e d}} \\cdot m ^ {\\text {r e s i z e d}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 462, + 784, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Despite its simplicity in implementation, BLD exhibits suboptimal performance in terms of both unmasked region preservation and generation content alignment. This is due to (1) the resize of the mask preventing it from correctly blending the noisy latent, (2) the diffusion model lacking perceptual knowledge of mask boundaries and the unmasked image region context.", + "bbox": [ + 212, + 484, + 782, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dedicated Inpainting Models. To enhance the performance of inpainting, previous works fine-tune the base models by expanding the input UNet channel to include the mask and masked image inputs, turning it into an architecture specifically designed for image inpainting. Though having better generation results compared with BLD, they still have several drawbacks: (1) These models merge the noisy latent, masked image latent, and mask at the initial convolution layer of the UNet architecture, where they are collectively influenced by the text embedding. Consequently, subsequent layers in the UNet model struggle to obtain pure masked image features due to the text's influence. (2) Incorporating both the condition processing and generation within a single branch places additional burdens on the UNet framework. (3) These approaches require extensive fine-tuning across various variations of diffusion backbones, which is computationally intensive and lacks transferability to custom diffusion models.", + "bbox": [ + 212, + 568, + 787, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Motivation", + "text_level": 1, + "bbox": [ + 215, + 775, + 354, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the analysis presented in Section 3.2, a more effective architecture design of inpainting would be introducing an additional branch specifically dedicated to masked image processing. ControlNet [51] is one of the widely adopted", + "bbox": [ + 212, + 794, + 784, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "BrushNet", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "strategies that exemplifies this idea. However, it should be noted that directly fine-tuning ControlNet, which is originally designed for controllable image generation, on the inpainting task yields unsatisfactory results. ControlNet designs a lightweight encoder to incorporate out-of-domain structural conditions (e.g., skeleton) and relies on text guidance for content generation, which is unsuitable for pixel-level inpainting image feature injection. Furthermore, ControlNet typically relies on sparse control, meaning that merely adding control to residuals in the UNet framework would be sufficient, while inpainting requires pixel-to-pixel constraints with strong restrictive information. Thus a new architecture specifically designed for inpainting is urgently needed.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 215, + 308, + 330, + 324 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "An overview of BrushNet is shown in Fig. 3. We employ a dual-branch strategy for masked image guidance insertion (Sec. 4.1). Blending operations with a blurred mask is used to ensure better-unmasked region preservation (Sec. 4.2). Notably, BrushNet can achieve flexible control by adjusting the added scale.", + "bbox": [ + 212, + 333, + 785, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/adc686765481ed163c4ad9e77e71ec36827bf43a5ff44056a215663840e6bdec.jpg", + "image_caption": [ + "Fig. 3: Model overview. Our model outputs an inpainted image given the mask and masked image input. Firstly, we downsample the mask to accommodate the size of the latent, and input the masked image to the VAE encoder to align the distribution of latent space. Then, noisy latent, masked image latent, and downsampled mask are concatenated as the input of BrushNet. The feature extracted from BrushNet is added to pretrained UNet layer by layer after a zero convolution block [51]. After denoising, the generated image and masked image are blended with a blurred mask." + ], + "image_footnote": [], + "bbox": [ + 218, + 410, + 784, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Masked Image Guidance", + "text_level": 1, + "bbox": [ + 215, + 744, + 465, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The insertion of the masked image feature into the pre-trained diffusion network is accomplished through an additional branch, which explicitly separates the feature extraction of masked images from the image-generating processes. The input to the additional branch includes noisy latent, masked image latent, and downsampled mask, which are concatenated together to form the input.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Xuan Ju, et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, the noisy latent provides generation information during the present generation process, helping BrushNet enhance the semantic coherence of the masked image feature. The masked image latent is extracted from the masked image using VAE, which is aligned with the data distribution of the pre-trained UNet. To ensure the alignment of the mask size with the noisy latent and masked image latent, we employ cubic interpolation to downsample the mask.", + "bbox": [ + 212, + 146, + 782, + 234 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To process the masked image features, BrushNet utilizes a clone of the pre-trained diffusion model while excluding its cross-attention layers. The pretrained weights of the diffusion model serve as a strong prior for extracting the masked image features, while the removal of the cross-attention layers ensures that only pure image information is considered within this additional branch. BrushNet feature is inserted into the frozen diffusion model layer-by-layer, enabling dense per-pixel control hierarchically. Similar to ControlNet [51], we employ zero convolution layers to establish a connection between the locked model and the trainable BrushNet. This ensures that harmful noise does not influence the hidden states in the trainable copy during the initial stages of training.", + "bbox": [ + 212, + 238, + 784, + 387 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The feature insertion operation is shown in Eq. 5. Specifically, $\\epsilon_{\\theta}(z_t,t,C)_i$ indicates the feature of the $i$ -th layer in network $\\epsilon_{\\theta}$ with $i\\sim [1,n]$ , where $n$ is the number of layers. The same notation applies to $\\epsilon_{\\theta}^{BrushNet}$ . $\\epsilon_{\\theta}^{BrushNet}$ takes the concatenated noisy latent $z_{t}$ , masked image latent $z_0^{masked}$ , and downsampled mask $m^{resized}$ as input, with the concatenation operation denoted as $[\\cdot]$ . $\\mathcal{Z}$ is the zero convolution operation. $w$ is the preservation scale used to adjust the influence of BrushNet on pretrained diffusion model.", + "bbox": [ + 212, + 388, + 784, + 494 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {\\theta} \\left(z _ {t}, t, C\\right) _ {i} = \\epsilon_ {\\theta} \\left(z _ {t}, t, C\\right) _ {i} + w \\cdot \\mathcal {Z} \\left(\\epsilon_ {\\theta} ^ {\\text {B r u s h N e t}} \\left(\\left[ z _ {t}, z _ {0} ^ {\\text {m a s k e d}}, m ^ {\\text {r e s i z e d}} \\right], t\\right) _ {i}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 507, + 784, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Blending Operation", + "text_level": 1, + "bbox": [ + 214, + 542, + 426, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As mentioned in Section 4.2, the blending operation conducted in latent space can result in inaccuracies due to the resizing of the mask. Similarly, in our approach, a similar issue arises as we resize the mask to match the size of the latent space, which can introduce potential inaccuracies. Additionally, it is important to acknowledge that VAE encoding and decoding operations have inherent limitations and may not ensure complete image reconstruction.", + "bbox": [ + 212, + 566, + 782, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To ensure a fully consistent image reconstruction of the unmasked region, previous works have explored different techniques. Some approaches [4, 56], utilize past-and-copy methods, where the unmasked region is directly copied from the original image. However, this can result in a lack of semantic coherence in the final generation results. On the other hand, adopting latent blending operations inspired by BLD [1, 33] has been observed to face challenges in effectively preserving the desired information in the unmasked regions.", + "bbox": [ + 212, + 657, + 784, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this work, we present a simple pixel space solution to address this issue by first blurring the mask and then performing copy-and-paste using the blurred mask. Although this approach may result in a slight loss of accuracy in preserving the details of the mask boundary, the error is nearly imperceptible to the naked eye and results in significantly improved coherence in the mask boundary.", + "bbox": [ + 212, + 765, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "BrushNet", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Flexible Control", + "text_level": 1, + "bbox": [ + 215, + 146, + 397, + 160 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The architecture design of BrushNet inherently makes it suitable for seamless plug-and-play integration to various pretrained diffusion models and enables flexible preservation scale. Specifically, the flexible control of our proposed BrushNet includes: (1) Since BrushNet does not modify the weights of the pretrained diffusion model, it can be readily integrated as a plug-and-play component with any community fine-tuned diffusion models. This allows for easy adoption and experimentation with different pretrained models. (2) Preservation Scale Adjustment: The preservation scale of the unmasked region can be controlled by incorporating BrushNet features into the frozen diffusion model with the weight $w$ . This weight determines the influence of BrushNet on the preservation scale, offering the ability to adjust the desired level of preservation. (3) Blurring Scale and Blending Operation: By adjusting the scale of blurring and deciding whether to apply the blending operation, the preservation scale of the unmasked region can be further customized. These features allow for flexible fine-grained control over the inpainting process. More explanation can be found in Sec. 5.5.", + "bbox": [ + 212, + 167, + 787, + 395 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 405, + 375, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1 Evaluation Benchmark and Metrics", + "text_level": 1, + "bbox": [ + 215, + 426, + 553, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Benchmark. Previous commonly used datasets in the image inpainting field include CelebA [22], CelebA-HQ [14], ImageNet [8], MSCOCO [19], Open Images [17], and LSUN-Bedroom [48]. However, these datasets either primarily focus on a small area, such as human faces, or predominantly consist of low-quality, cluttered real-life scene data. As a result, these datasets are not well-suited for training and evaluating diffusion-based inpainting models, which can generate high-quality diverse images that align with text prompts.", + "bbox": [ + 212, + 446, + 787, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Recently proposed EditBench [37] serves as a benchmark specifically designed for text-guided image inpainting for diffusion models. This benchmark consists of a collection of 240 images comprising an equal ratio of natural images and generated images, with mask and caption annotation for each image. However, the annotated masks in EditBench are mostly random shapes without specific object information, neglecting the practical application of inpainting in real scenarios such as replacing an object with an external mask, as commonly seen in E-commerce product displays and image editing.", + "bbox": [ + 212, + 553, + 787, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To fill the gap, we propose BrushBench for segmentation-based inpainting, as shown in Fig. 4. BrushBench comprises a total of 600 images, with each image accompanied by the human-annotated mask and caption annotation. The images in BrushBench are evenly distributed between natural images and artificial images, such as paintings. Furthermore, the dataset ensures an equal distribution among different categories, including humans, animals, indoor scenarios, and outdoor scenarios. This balanced distribution enables a fair evaluation across various categories, promoting better evaluation equity.", + "bbox": [ + 212, + 674, + 787, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further enhance the analysis of the inpainting task, we categorize it into two distinct types based on the masks used: random brush masks and segmentation-based masks. We use EditBench as the comparison benchmark", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Xuan Ju, et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9e81c8cb4347892949d92e8b70253f27259f3ee930f7f7ddf10488f1d0df5858.jpg", + "image_caption": [ + "Fig. 4: Benchmark overview. I and II separately show natural and artificial images, masks, and caption of BrushBench. (a) to (d) show images of humans, animals, indoor scenarios, and outdoor scenarios. Each group of images shows the original image, inside-inpainting mask, and outside-inpainting mask, with an image caption on the top. III show image, mask, and caption from EditBench [37], with (e) for generated images and (f) for natural images. The images are randomly selected from both benchmarks." + ], + "image_footnote": [], + "bbox": [ + 222, + 145, + 779, + 344 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "for random brush masks and use BrushBench for segmentation-based masks. Regarding inpainting with segmentation-based masks, we refine the task by considering two specific scenarios: segmentation mask inside-inpainting and segmentation mask outside-inpainting. By separating these two subtasks, we can better understand the inpainting performance in different image regions.", + "bbox": [ + 212, + 434, + 787, + 511 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dataset. To train segmentation-based mask inpainting, we annotate segmentation mask on Laion-Aesthetic [34] dataset, called BrushData. We employ the Grounded-SAM [32] to annotate open-world masks and then filter the masks based on their confidence score, ensuring that only masks with relatively higher confidence scores are retained. Additionally, we consider factors such as reasonable mask size and good mask continuity during the filtering process.", + "bbox": [ + 212, + 513, + 787, + 603 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Metrics. We consider 7 metrics from three aspects: image generation quality, masked region preservation, and text alignment.", + "bbox": [ + 212, + 606, + 785, + 636 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Image Generation Quality. Metrics most used by previous inpainting methods (e.g., FID [11] and KID [5]) show a poor representation of the rich and varied content generated by modern text-to-image models [16]. Thus, we use Image Reward (IR) [44], HPS v2 (HPS) [41], and Aesthetic Score (AS) [34] that align with human perception. Specifically, ImageReward and HPS v2 are text-to-image human preference evaluation models trained on large-scale datasets of human preference choices on generated images. Aesthetic Score is a linear model trained on image quality rating pairs of real images.", + "- Masked Region Preservation. We follow previous works using standard Peak Signal-to-Noise Ratio (PSNR) [39], Learned Perceptual Image Patch Similarity (LPIPS) [52], and Mean Squared Error (MSE) [38] in the unmasked region among the generated image and the original image." + ], + "bbox": [ + 225, + 638, + 787, + 818 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "BrushNet", + "bbox": [ + 665, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "1 The proposed BrushData and BrushBench will be released along with the codes.", + "bbox": [ + 217, + 824, + 769, + 840 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "- Text Alignment. We use CLIP Similarity (CLIP Sim) [40] to evaluate text-image consistency between the generated images and corresponding text prompts. CLIP Similarity projects text and images to the same shared space with the CLIP model [30] and evaluates the similarity of their embeddings.", + "bbox": [ + 225, + 145, + 787, + 208 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 224, + 457, + 239 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We perform the inference of different inpainting methods in the same setting unless specifically clarified, i.e., on NVIDIA Tesla V100 following their open-source code with a base model of Stabe Diffusion v1.5 in 50 steps, with a guidance scale of 7.5. We keep the recommended hyper-parameter for each inpainting method in all images for fair comparison. BrushNet and all ablation models are trained for 430 thousands steps on 8 NVIDIA Tesla V100 GPUs, which takes around 3 days. For comparison on BrushBench, we use BrushNet trained on BrushData. For comparison on EditBench, we use the model trained on LAION-5B with random masks. Details can be found in the provided codes.", + "bbox": [ + 212, + 248, + 787, + 383 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.3 Quantitative Comparison", + "text_level": 1, + "bbox": [ + 215, + 396, + 470, + 411 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/4bb420381fc7d24dc3243c9038638bfb3021604638161e41f09c9521f4cfa0fa.jpg", + "table_caption": [ + "Table 2: Quantitative comparisons among BrushNet and other diffusion-based inpainting models in BrushBench: Blended Latent Diffusion (BLD) [1], Stable Diffusion Inpainting (SDI) [33], HD-Painter (HDP) [25], PowerPaint (PP) [56], and ControlNet-Inpainting (CNI) [51]. Metrics encompassing image quality, masked region preservation, and text alignment (Text Align) for inside-inpainting and outside-inpainting are shown in the table. All models use Stable Diffusion V1.5 as base model. Red stands for the best result, Blue stands for the second best result." + ], + "table_footnote": [ + "* with blending operation", + "Tab. 2 and Tab. 3 show the quantitative comparison on BrushBench and EditBench [37]. We compare the inpainting results of sampling strategy modification method Blended Latent Diffusion [1], dedicated inpainting models Stable Diffusion Inpainting [33], HD-Painter [25], and PowerPaint [56], as well as plug-and-play method ControlNet [51] trained on inpainting data." + ], + "table_body": "
MetricsImage QualityMasked Region PreservationText Align
ModelsIR×10↑HPS×102↑AS↑PSNR↑MSE×103↓LPIPS×103↓CLIP Sim↑
InsideBLD [1]9.7825.876.1721.339.7649.2626.15
SDI [33]11.7227.066.5021.5213.8748.3926.17
HDP [25]11.6826.906.4222.619.9543.5026.37
PP [56]11.4627.356.2421.4332.7348.4326.48
CNI [51]9.926.026.5312.3978.78243.6226.47
CNI* [51]11.2126.926.3922.7324.5843.4926.22
Ours12.3627.406.5321.659.3148.2826.48
Ours*12.6427.786.5131.940.8018.6726.39
OutsideBLD [1]7.8126.776.2315.8535.8621.4026.73
SDI [33]10.2727.996.5518.0419.8715.1327.21
HDP [25]9.6627.796.4618.0322.9915.2226.96
PP [56]7.4528.016.2618.0431.7815.1326.72
CNI [51]9.2627.686.4211.9183.0358.1627.29
CNI* [51]9.5727.766.2817.5037.7219.9526.92
Ours10.8228.026.6418.0622.8615.0827.33
Ours*10.8828.096.6427.822.254.6327.22
", + "bbox": [ + 225, + 523, + 777, + 741 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Xuan Ju, et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results demonstrate BrushNet's effectiveness across image quality, masked region preservation, and image-text alignment. Blended Latent Diffusion [1] shows the poorest results in image quality and text alignment, derived from the incoherence between the generated masked and the unmasked given image. At the same time, its performance in masked region preservation is also not satisfactory because of the loss incurred by resized mask blending operation in the latent space. Modified from Stable Diffusion Inpainting [33], HD-Painter [25] and PowerPaint [56] demonstrate comparable performance to Stable Diffusion Inpainting in the task of inside-inpainting. However, when it comes to outside-inpainting, their results in terms of image quality and text alignment are significantly poorer compared to Stable Diffusion Inpainting, which can be attributed to their exclusive emphasis on the inside-inpainting task.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "ControlNet [51] trained on inpainting has the most similar experimental configuration to ours. Due to the mismatch between its model design and the inpainting task, ControlNet shows poor results in masked region preservation and image quality, necessitating its combination with Blended Latent Diffusion [1] to generate satisfying inpainted images. However, even with this combination, it still falls short compared to dedicated inpainting models and BrushNet.", + "bbox": [ + 212, + 329, + 787, + 420 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/a010d12ca81ff675be16f7515ddfebacbd68e649f8b47632cce3f9a16d7025ab.jpg", + "table_caption": [ + "Table 3: Quantitative comparisons among BrushNet and other diffusion-based inpainting models in EditBench. A detailed explanation of compared methods and metrics can be found in the caption of Tab. 2. Red stands for the best result, Blue stands for the second best result." + ], + "table_footnote": [ + "* with blending operation" + ], + "table_body": "
MetricsImage QualityMasked Region PreservationText Align
ModelsIR×10↑HPS×102↑AS↑PSNR↑MSE×103↓LPIPS×103↓CLIP Sim↑
BLD [1]0.9023.815.4420.8910.9331.9028.62
SDI [33]1.8624.245.6923.256.9424.3028.00
HDP [25]1.7424.205.6423.076.7024.3228.34
PP [56]1.2424.505.4423.3420.1224.1227.80
CNI [51]1.4924.465.8212.7169.42159.7128.16
CNI* [51]0.9023.795.4622.6135.9326.1427.74
Ours4.4025.105.8423.356.8124.1128.67
Ours*4.4625.245.8233.660.6310.1228.87
", + "bbox": [ + 233, + 488, + 766, + 638 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The performance on the EditBench is roughly consistent with the overall performance on BrushBench, which similarly shows BrushNet's superior performance. This indicates that our method exhibits strong performance across a range of inpainting tasks with various mask types, including random masks, inside-inpainting masks, and outside-inpainting masks.", + "bbox": [ + 212, + 664, + 787, + 742 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.4 Qualitative Comparison", + "text_level": 1, + "bbox": [ + 214, + 753, + 459, + 771 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The qualitative comparison with previous image inpainting methods is shown in Fig. 1. We provide results on artificial images and natural images across various inpainting tasks, including random mask inpainting, segmentation mask inside-inpainting, and segmentation mask outside-inpainting. BrushNet consistently", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "BrushNet", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "show exceptional results in the coherent of generated region and unmasked region, considering content (I, II right, III right, IV), color (II left), and text (III left). Interestingly, Fig. 1 III left requires the model to generate a cat and a goldfish. All previous methods fail to recognize that a goldfish is already present in the masked image, resulting in the generation of an additional fish within the masked region. BrushNet successfully realized the awareness of background information due to the design of the dual-branch decoupling.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/51b2813a57d023f97be77d4b34e3c2fb7e895b8b4b20ce7618e651880ca3b5d3.jpg", + "image_caption": [ + "Fig. 5: Comparison of previous inpainting methods and BrushNet on various image domain. A detailed explanation of compared methods is in Fig. 1." + ], + "image_footnote": [], + "bbox": [ + 217, + 257, + 777, + 535 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The untouched pre-trained diffusion branch also provides the advantage of better coverage across different data domains, such as painting and anime. As shown in Fig. 5, BrushNet demonstrates superior performance across various image categories, including natural image (I, II), pencil painting (III), anime (IV), illustration (V), digital art (VI left), and watercolor (VI right). Due to the page limit, more qualitative comparison results are in supplementary files.", + "bbox": [ + 212, + 575, + 787, + 667 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.5 Flexible Control Ability", + "text_level": 1, + "bbox": [ + 214, + 675, + 460, + 690 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Fig. 6 and Fig. 7 illustrate the flexible control provided by BrushNet in two aspects: base diffusion model selection and controlling scale. In Fig. 6, we showcase the ability to combine BrushNet with different diffusion models fine-tuned by the community. This allows users to select a specific model that best suits their inpainting requirements, enabling users to achieve the desired inpainting results based on their specific needs. Fig. 7 demonstrates the adjustment of the control scale of BrushNet. This control scale parameter allows users to effectively control the extent of unmasked region protection during the inpainting process. By manipulating the scale parameter, users can achieve fine-grained control over the inpainting process, enabling precise and customizable inpainting.", + "bbox": [ + 212, + 700, + 787, + 853 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Xuan Ju, et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/76228267e1ed329a83533c4805e99107400c5df7e9b93689366e220d820d07b8.jpg", + "image_caption": [ + "A single red rose is laying on top of snow" + ], + "image_footnote": [], + "bbox": [ + 236, + 157, + 372, + 195 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ac187c77211becdb3dcd38a5f8135018bf4e8efdadf8d999fb857d70df92928c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 157, + 506, + 195 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/bf93efd098fa90b639fdba74566759e61b1f29078ad7f469d2bd2291bc6f9c45.jpg", + "image_caption": [ + "A cartoon boy driving a car" + ], + "image_footnote": [], + "bbox": [ + 509, + 157, + 692, + 195 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b48bd04a24f3d4e7b0df40bdbd04ed1e197e6bf9206905c72ef58092cdbd6cd7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 157, + 733, + 195 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0621ab133f65bdddaf98800b2cb4eccd7292e587b6d8f978574208d634121a40.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 159, + 785, + 195 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/bd949013803c54c98582c8e6f25572344811acf7a0e54fefcfbc4d1ddb753ac2.jpg", + "image_caption": [ + "A little girl sitting on the grass with her dog", + "(a) Mask", + "(b) DS", + "(c) ER", + "(d) HR (e)", + "Fig. 6: Integrating BrushNet to community fine-tuned diffusion models. We use five popular community diffusion models fine-tuned from stable diffusion v1.5: DreamShaper (DS) [24], epiCRealism (ER) [9], Henmix_Real (HR) [10], MeinaMix (MM) [26], and Realistic Vision (RV) [35]. MM is specifically designed for anime images." + ], + "image_footnote": [], + "bbox": [ + 236, + 209, + 372, + 246 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/31ad5eab82caf1d088f4f5ce436feac6fc4d5e4362776b5167b3bdeebadca15c.jpg", + "image_caption": [ + ")MM", + "(f) RV" + ], + "image_footnote": [], + "bbox": [ + 372, + 209, + 506, + 246 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3e63eabc35e89d86671afca7305a914e87ef044fa33d92f63aef3afd5717e5ec.jpg", + "image_caption": [ + "A woman with her hands on her face", + "(a) Mask", + "(b) DS", + "(c) ER", + "(d) HR", + "(e) MM" + ], + "image_footnote": [], + "bbox": [ + 509, + 210, + 692, + 246 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/87142fa1b19038c9e7a5ff017ae480f0b28790eedb6a717d48f1c59df5fb07bb.jpg", + "image_caption": [ + "(f) RV" + ], + "image_footnote": [], + "bbox": [ + 692, + 210, + 782, + 246 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b911b418c06f7d9da12f655394c7a885f3843045c2947495d0207f2ecb5cf4be.jpg", + "image_caption": [ + "Book, coffee and purple flower", + "A cupcake with a lit candle on top" + ], + "image_footnote": [], + "bbox": [ + 218, + 352, + 500, + 380 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/abe2cfc0f166cf0321d04f9e8cd038fac82001cdecf00afe15a7b20a16e910a4.jpg", + "image_caption": [ + "(a) Mask", + "(b) 1.0", + "(c) 0.8", + "(d) 0.6", + "(e) 0", + "(1)0.4", + "(g) 0.3", + "Fig. 7: Flexible control scale of BrushNet. (a) shows the given masked image, (b)-(h) show adding BrushNet with control scale $w$ from 1.0 to 0.2. Results show a gradually diminishing controllable ability from precise to rough control." + ], + "image_footnote": [], + "bbox": [ + 218, + 391, + 500, + 419 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/66c7432ed6d7764ba3dd97737f94396a9c3f103455cc16eeb8b2fd5530993580.jpg", + "image_caption": [ + "A small island with a tree and clouds" + ], + "image_footnote": [], + "bbox": [ + 503, + 352, + 782, + 380 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/599b7b1af0165525307e84eac87130cf4c592dc8bf4149a778b39fa2275e9dff.jpg", + "image_caption": [ + "A white rabbit sitting on the ground in the grass", + "(a) Mask (b) 1.0", + "(c) 0.8", + "(d) 0.6", + "(e) 0.5", + "(f) 0.4", + "(g) 0.3" + ], + "image_footnote": [], + "bbox": [ + 503, + 391, + 782, + 419 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.6 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 497, + 388, + 512 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 4: Ablation on dual-branch design. Stable Diffusion Inpainting (SDI) use single-branch design, where the entire UNet is fine-tuned. We conducted an ablation analysis by training a dual-branch model with two variations: one with the base UNet fine-tuned, and another with the base UNet forzened. Results demonstrate the superior performance achieved by adopting the dual-branch design. Red is the best result.", + "bbox": [ + 214, + 518, + 785, + 589 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/45fbcbccd03fc4eb714088d1e9ece8cd84bcc629eace8eefd0c0a334eb917449.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MetricsImage QualityMasked Region PreservationText Align
ModelIR×10↑HPS×102↑AS↑PSNR↑MSE×102↓LPIPS×103↓CLIP Sim↑
SDI11.0027.536.5319.7816.8731.7626.69
w/o fine-tune11.5927.716.5919.8616.0931.6826.91
w/ fine-tune11.6327.736.6020.1315.8431.5726.93
", + "bbox": [ + 225, + 597, + 772, + 676 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We conducted ablation studies to investigate the impact of different model designs. Tab. 4 compares the dual-branch and single-branch designs. Tab. 5 shows the ablation study focusing on the additional branch architecture. The ablation studies are conducted on BrushBench, averaging the performance of inside-inpainting and outside-inpainting. The results presented in Tab. 4 demonstrate that the dual-branch design significantly outperforms the single-branch design. Additionally, fine-tuning the base diffusion model in dual-branch design yields better results compared to freezing it. However, fine-tuning the base diffusion model may restrict flexibility and control over the model. Considering this trade-off between performance and flexibility, we decide to adopt the frozen", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "BrushNet", + "bbox": [ + 665, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "aside_text", + "text": "I", + "bbox": [ + 218, + 172, + 227, + 181 + ], + "page_idx": 12 + }, + { + "type": "aside_text", + "text": "I", + "bbox": [ + 218, + 223, + 227, + 233 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "dual-branch design as our model design. Tab. 5 presents the rationale behind the design choices for (1) using a VAE encoder instead of randomly initialized convolution layers to process the masked image, (2) incorporating the full UNet feature layer-by-layer into the pre-trained UNet, (3) removing text cross-attention in BrushNet, which avoids masked image features influenced by text.", + "bbox": [ + 212, + 146, + 782, + 222 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 5: Ablation on model architecture. We ablate on the following components: the image encoder (Enc), selected from a random initialized convolution (Conv) and a VAE; the inclusion of mask in input (Mask), chosen from adding (w/) and not adding (w/o); the presence of cross-attention layers (Attn), chosen from adding (w/) and not adding (w/o); the type of UNet feature addition (UNet), selected from adding the full UNet feature (full), adding half of the UNet feature (half), and adding the feature like ControlNet (CN); and finally, the blending operation (Blend), chosen from not adding (w/o), direct pasting (paste), and blurred blending (blur). Red is the best result.", + "bbox": [ + 212, + 224, + 787, + 335 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/dbfb03d1d183d84c9b643cea81c48652c2797dcdc8cf7adc3b624156db4bf959.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MetricsImage QualityMasked Region PreservationText Align
EncMaskAttnUNetBlendIR×10↑HPS×102↑AS↑PSNR↑MSE×102↓LPIPS×103↓CLIP Sim↑
Convw/w/ofullw/o11.0526.236.5514.8937.2364.5426.76
VAEw/ow/ofullw/o11.5527.706.5717.9626.3849.3326.87
VAEw/w/fullw/o11.2527.626.5618.6919.4434.2826.63
Convw/w/CNw/o9.5826.856.4712.1580.91150.8926.88
VAEw/w/CNw/o10.5327.426.5918.2824.3641.6326.89
VAEw/w/oCNw/o11.4227.696.5818.4924.0936.3326.86
VAEw/w/ohalfw/o11.4727.706.5719.0123.7733.5726.87
VAEw/w/ofullw/o11.5927.716.5919.8616.0931.6826.91
VAEw/w/ofullpaste11.7227.936.58---26.80
VAEw/w/ofullblur11.7627.946.5829.881.5311.6526.81
", + "bbox": [ + 228, + 342, + 766, + 476 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Discussion", + "text_level": 1, + "bbox": [ + 215, + 486, + 354, + 502 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Conclusion. This paper proposes a plug-and-play image inpainting method BrushNet with a pixel-level masked image feature insertion architectural design. Quantitative and qualitative results on our proposed benchmark, BrushBench, and EditBench show the superior performance of BrushNet considering image generation quality, masked region preservation, and image-text alignment.", + "bbox": [ + 212, + 505, + 782, + 580 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations and Future Work. However, BrushNet still has some limitations: (1) The quality and content generated by our model are heavily dependent on the chosen base model. (2) Even with BrushNet, we still observe poor generation results in cases where the given mask has an unusually shaped or irregular form, or when the given text does not align well with the masked image. In our future work, we will continue to address these challenges.", + "bbox": [ + 212, + 584, + 782, + 674 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Negative Social Impact. Image inpainting models present exciting opportunities for content creation, but they also carry potential risks to individuals and society. Their reliance on internet-collected training data can amplify social biases, and there is a specific risk of generating persuasive misinformation by manipulating human images with offensive elements. To address these concerns, it is crucial to emphasize responsible use and establish ethical guidelines when utilizing these models. This is also a key focus for our future model releases.", + "bbox": [ + 212, + 678, + 782, + 782 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "*Acknowledgment.", + "text_level": 1, + "bbox": [ + 215, + 794, + 372, + 809 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work is supported in part by Research Matching Grant(CSE-7-2022)-RMG01.", + "bbox": [ + 215, + 809, + 792, + 824 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Xuan Ju, et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. ACM transactions on graphics (TOG) 42(4), 1-11 (2023)", + "2. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18208-18218 (2022)", + "3. Bertalmio, M., Sapiro, G., Caseles, V., Ballester, C.: Image inpainting. In: International Conference and Exhibition on Computer Graphics and Interactive Techniques (SIGGRAPH). pp. 417-424 (2000)", + "4. Binghui, C., Chao, L., Chongyang, Z., Wangmeng, X., Yifeng, G., Xuansong, X.: Replaceanything as you want: Ultra-high quality content replacement (2023), https://aigcdesigngroup.github.io/replace-anything/", + "5. Binkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying MMD GANs. arXiv preprint arXiv:1801.01401 (2018)", + "6. Corneanu, C., Gadde, R., Martinez, A.M.: Latentpaint: Image inpainting in latent space with diffusion models. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 4334-4343 (2024)", + "7. Criminisi, A., Pérez, P., Toyama, K.: Region filling and object removal by exemplar-based image inpainting. IEEE Transactions on Image Processing 13(9), 1200-1212 (2004)", + "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 248-255. IEEE (2009)", + "9. epinikion: epicrealism (2023), https://civitai.com/models/25694?modelVersionId=143906", + "0. hemi29833: Henmixreal (2024), https://civitai.com/models/20282?modelVersionId=305687", + "1. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs trained by a two time-scale update rule converge to a local Nash equilibrium. Advances in Neural Information Processing Systems (NIPS) 30 (2017)", + "2. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NIPS) 33, 6840-6851 (2020)", + "3. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)", + "4. Huang, H., He, R., Sun, Z., Tan, T., et al.: Introvae: Introspective variational autoencoders for photographic image synthesis. Advances in Neural Information Processing Systems (NIPS) 31 (2018)", + "5. Huang, Y., Huang, J., Liu, Y., Yan, M., Lv, J., Liu, J., Xiong, W., Zhang, H., Chen, S., Cao, L.: Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525 (2024)", + "6. Jayasumana, S., Ramalingam, S., Veit, A., Glasner, D., Chakrabarti, A., Kumar, S.: Rethinking fid: Towards a better evaluation metric for image generation. arXiv preprint arXiv:2401.09603 (2023)", + "7. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision (IJCV) 128(7), 1956-1981 (2020)", + "8. Li, Z., Wei, P., Yin, X., Ma, Z., Kot, A.C.: Virtual try-on with pose-garment keypoints guided inpainting. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 22788-22797 (2023)" + ], + "bbox": [ + 225, + 172, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "BrushNet", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision (ECCV). pp. 740-755. Springer (2014)", + "20. Liu, A., Niepert, M., Broeck, G.V.d.: Image inpainting via tractable steering of diffusion models. arXiv preprint arXiv:2401.03349 (2023)", + "21. Liu, H., Wan, Z., Huang, W., Song, Y., Han, X., Liao, J.: Pd-GAN: Probabilistic diverse GAN for image inpainting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9371-9381 (2021)", + "22. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: IEEE/CVF International Conference on Computer Vision (ICCV) (December 2015)", + "23. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11461-11471 (2022)", + "24. Lykon: Dreamshaper (2022), https://civitai.com/models/4384?modelVersionId=128713", + "25. Manukyan, H., Sargsyan, A., Atanyan, B., Wang, Z., Navasardyan, S., Shi, H.: Hd-painter: High-resolution and prompt-faithful text-guided image inpainting with diffusion models. arXiv preprint arXiv:2312.14091 (2023)", + "26. Meina: Meinamix (2023), https://civitai.com/models/7240?modelVersionId=119057", + "27. Peng, J., Liu, D., Xu, S., Li, H.: Generating diverse structure for image inpainting with hierarchical vq-vae. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10775-10784 (2021)", + "28. von Platen, P., Patil, S., Lozhkov, A., Cuenca, P., Lambert, N., Rasul, K., Davaadorj, M., Wolf, T.: Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers (2022)", + "29. Quan, W., Chen, J., Liu, Y., Yan, D.M., Wonka, P.: Deep learning-based image and video inpainting: A survey. International Journal of Computer Vision (IJCV) pp. 1-34 (2024)", + "30. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (ICML). pp. 8748-8763. PMLR (2021)", + "31. Razzhigaev, A., Shakhmatov, A., Maltseva, A., Arkhipkin, V., Pavlov, I., Ryabov, I., Kuts, A., Panchenko, A., Kuznetsov, A., Dimitrov, D.: Kandinsky: an improved text-to-image synthesis with image prior and latent diffusion. arXiv preprint arXiv:2310.03502 (2023)", + "32. Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., et al.: Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159 (2024)", + "33. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022)", + "34. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems (NIPS) 35, 25278–25294 (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Xuan Ju, et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. SG161222: Realisticvision (2023), https://civitai.com/models/4201?modelVersionId=130072", + "36. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)", + "37. Wang, S., Sahara, C., Montgomery, C., Pont-Tuset, J., Noy, S., Pellegrini, S., Onoe, Y., Laszlo, S., Fleet, D.J., Soricut, R., et al.: Imagen editor and editbench: Advancing and evaluating text-guided image inpainting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18359-18369 (2023)", + "38. Wikipedia contributors: Mean squared error — Wikipedia, the free encyclopedia (2024), https://en.wikipedia.org/w/index.php?title=Mean_squared_error&oldid=1207422018, [Online; accessed 4-March-2024]", + "39. Wikipedia contributors: Peak signal-to-noise ratio — Wikipedia, the free encyclopedia (2024), https://en.wikipedia.org/w/index.php?title=Peak_signal-to-noise_ratio&oldid=1210897995, [Online; accessed 4-March-2024]", + "40. Wu, C., Huang, L., Zhang, Q., Li, B., Ji, L., Yang, F., Sapiro, G., Duan, N.: GO-DIVA: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806 (2021)", + "41. Wu, X., Hao, Y., Sun, K., Chen, Y., Zhu, F., Zhao, R., Li, H.: Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341 (2023)", + "42. Xie, S., Zhang, Z., Lin, Z., Hinz, T., Zhang, K.: Smartbrush: Text and shape guided object inpainting with diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22428-22437 (2023)", + "43. Xie, S., Zhao, Y., Xiao, Z., Chan, K.C., Li, Y., Xu, Y., Zhang, K., Hou, T.: Dreaminpainter: Text-guided subject-driven image inpainting with diffusion models. arXiv preprint arXiv:2312.03771 (2023)", + "44. Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagereward: Learning and evaluating human preferences for text-to-image generation (2023)", + "45. Xu, Z., Zhang, X., Chen, W., Yao, M., Liu, J., Xu, T., Wang, Z.: A review of image inpainting methods based on deep learning. Applied Sciences 13(20), 11189 (2023)", + "46. Yang, S., Chen, X., Liao, J.: Uni-paint: A unified framework for multimodal image inpainting with pretrained diffusion model. In: ACM International Conference on Multimedia (MM). pp. 3190-3199 (2023)", + "47. Yang, S., Zhang, L., Ma, L., Liu, Y., Fu, J., He, Y.: Magicremover: Tuning-free text-guided image inpainting with diffusion models. arXiv preprint arXiv:2310.02848 (2023)", + "48. Yu, F., Seff, A., Zhang, Y., Song, S., Funkhouser, T., Xiao, J.: Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365 (2015)", + "49. Yu, T., Feng, R., Feng, R., Liu, J., Jin, X., Zeng, W., Chen, Z.: Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790 (2023)", + "50. Zhang, G., Ji, J., Zhang, Y., Yu, M., Jaakkola, T., Chang, S.: Towards coherent image inpainting using denoising diffusion implicit models (2023)", + "51. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "BrushNet", + "bbox": [ + 663, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "52. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 586-595 (2018)", + "53. Zhao, S., Cui, J., Sheng, Y., Dong, Y., Liang, X., Chang, E.I., Xu, Y.: Large scale image completion via co-modulated generative adversarial networks. arXiv preprint arXiv:2103.10428 (2021)", + "54. Zheng, C., Cham, T.J., Cai, J.: Pluralistic image completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1438-1447 (2019)", + "55. Zheng, H., Lin, Z., Lu, J., Cohen, S., Shechtman, E., Barnes, C., Zhang, J., Xu, N., Amirghodsi, S., Luo, J.: Image inpainting with cascaded modulation GAN and object-aware training. In: European Conference on Computer Vision (ECCV). pp. 277-296. Springer (2022)", + "56. Zhuang, J., Zeng, Y., Liu, W., Yuan, C., Chen, K.: A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. arXiv preprint arXiv:2312.03594 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 383 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Xuan Ju, et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/BrushNet_ A Plug-and-Play Image Inpainting Model with Decomposed Dual-Branch Diffusion/8cdf68e6-b7ca-45f7-9a98-9bdc1df71997_model.json b/2024/BrushNet_ A Plug-and-Play Image Inpainting Model with Decomposed Dual-Branch Diffusion/8cdf68e6-b7ca-45f7-9a98-9bdc1df71997_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cd084e8e41d37258d1d02aaf6fbdd09796ebfa3d --- /dev/null +++ b/2024/BrushNet_ A Plug-and-Play Image Inpainting Model with Decomposed Dual-Branch Diffusion/8cdf68e6-b7ca-45f7-9a98-9bdc1df71997_model.json @@ -0,0 +1,2865 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.223, + 0.142, + 0.773, + 0.187 + ], + "angle": 0, + "content": "BrushNet: A Plug-and-Play Image Inpainting Model with Decomposed Dual-Branch Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.199, + 0.785, + 0.215 + ], + "angle": 0, + "content": "Xuan \\(\\mathrm{Ju}^{1,2}\\), Xian Liu\\(^{1,2}\\), Xintao Wang\\(^{1\\star}\\), Yuxuan Bian\\(^{2}\\), Ying Shan\\(^{1}\\), and Qiang \\(\\mathrm{Xu}^{2\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.218, + 0.724, + 0.233 + ], + "angle": 0, + "content": "\\(^{1}\\)ARC Lab, Tencent PCG \\(^{2}\\)The Chinese University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.353, + 0.235, + 0.65, + 0.248 + ], + "angle": 0, + "content": "https://github.com/TencentARC/BrushNet" + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.254, + 0.766, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.441, + 0.788, + 0.541 + ], + "angle": 0, + "content": "Fig. 1: Performance comparisons of BrushNet and previous image inpainting methods across various inpainting tasks: (I) Random Mask (< 50% masked), (II) Random Mask (> 50% masked), (III) Segmentation Mask Inside-Inpainting, (IV) Segmentation Mask Outside-Inpainting. Each group of results contains an artificial image (left) and a natural image (right) with 6 inpainting methods: (b) Blended Latent Diffusion (BLD) [1], (c) Stable Diffusion Inpainting (SDI) [33], (d) HD-Painter (HDP) [25], (e) PowerPaint (PP) [56], (f) ControlNet-Inpainting (CNI) [51], and (g) Ours." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.548, + 0.744, + 0.798 + ], + "angle": 0, + "content": "Abstract. Image inpainting, the process of restoring corrupted images, has seen significant advancements with the advent of diffusion models (DMs). Despite these advancements, current DM adaptations for inpainting, which involve modifications to the sampling strategy or the development of inpainting-specific DMs, frequently suffer from semantic inconsistencies and reduced image quality. Addressing these challenges, our work introduces a novel paradigm: the division of masked image features and noisy latent into separate branches. This division dramatically diminishes the model's learning load, facilitating a nuanced incorporation of essential masked image information in a hierarchical fashion. Herein, we present BrushNet, a novel plug-and-play dual-branch model engineered to embed pixel-level masked image features into any pre-trained DM, guaranteeing coherent and enhanced image inpainting outcomes. Additionally, we introduce BrushData and BrushBench to facilitate segmentation-based inpainting training and performance assessment. Our extensive experimental analysis demonstrates BrushNet's superior performance over existing models across seven key metrics, including image quality, mask region preservation, and textual coherence." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.806, + 0.725, + 0.821 + ], + "angle": 0, + "content": "Keywords: Image Inpainting \\(\\cdot\\) Diffusion Models \\(\\cdot\\) Image Generation" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.386, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.128 + ], + "angle": 0, + "content": "Xuan Ju, et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.376, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.171, + 0.788, + 0.277 + ], + "angle": 0, + "content": "Image inpainting [45] aims at restoring the missing regions of an image while maintaining the overall coherence. As a long-standing computer vision problem, it facilitates numerous applications such as virtual try-on [18] and image editing [15]. Recently, diffusion models [12,36] have demonstrated impressive performance in image generation, enabling flexible user control with semantic and structural conditions [33,51]. To this end, researchers resort to diffusion-based pipelines for high-quality image inpainting that aligns with given text prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.278, + 0.794, + 0.475 + ], + "angle": 0, + "content": "Commonly used diffusion-based text-guided inpainting methods can be roughly divided into two categories: (1) Sampling strategy modification [1,2,6,20,23,47, 50], which modifies the standard denoising process by sampling the masked regions from a pre-trained diffusion model, and the unmasked areas are simply copy-pasted from the given image in each denoising step. Although they can be used in arbitrary diffusion backbones, the limited perceptual knowledge of mask boundaries and the unmasked image region context leads to incoherent inpainting results. (2) Dedicated inpainting models [4, 33, 37, 42, 43, 46, 49, 56], which fine-tune a specially designed image inpainting model by expanding the input channel dimension of base diffusion models to incorporate provided corrupted image and mask. While they enable the diffusion model to generate more satisfying results with specialized content-aware and shape-aware models, we argue, is this architecture the best fit for diffusion-based inpainting?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.789, + 0.596 + ], + "angle": 0, + "content": "As shown in Fig. 2, dedicated inpainting models fuse noisy latent, masked image latent, mask, and text at an early stage. This architectural design makes the masked image feature easily influenced by the text embedding, preventing subsequent layers in the UNet from obtaining pure masked image features due to the influence of text. Additionally, handling the condition and generation in a single branch imposes extra burdens on the UNet framework. These approaches also necessitate fine-tuning in different variations of diffusion backbones, which can be time-consuming with limited transferability." + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.604, + 0.773, + 0.771 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.231, + 0.774, + 0.773, + 0.789 + ], + "angle": 0, + "content": "Fig. 2: Comparison of previous inpainting architectures and BrushNet." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Adding an additional branch dedicated to masked image feature extraction could be a promising approach to tackle the issue above. However, the existing solutions such as ControlNet [51] lead to inadequate information extraction" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.665, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BrushNet" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.221 + ], + "angle": 0, + "content": "and insertion when directly applied to inpainting, which originates from differences between controllable image generation and inpainting: inpainting requires pixel-to-pixel constraints with strong restrictive information instead of sparse structural control relying on text for content completion. As a result, ControlNet yields unsatisfactory results compared to dedicated inpainting models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.403 + ], + "angle": 0, + "content": "To address this issue, we propose BrushNet, which introduces an additional branch to the diffusion framework, creating a more suitable architecture for image inpainting. Specifically, our designs are three-fold: (1) To improve the extraction of image features for adaptation to the UNet distribution, we use a VAE encoder instead of randomly initialized convolution layers to process the masked image. (2) To enable dense per-pixel control, we adopt a hierarchical approach by gradually incorporating the full UNet feature layer-by-layer into the pre-trained UNet. (3) To ensure pure image information is considered in the additional branch, we remove text cross-attention from UNet. This design further offers plug-and-play capabilities and flexible unmasked region controllability to the inpainting process. For better consistency and a larger range of unmasked region controllability, we additionally propose a blurred blending strategy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.787, + 0.525 + ], + "angle": 0, + "content": "To ensure a comprehensive evaluation for real-life applications, we categorize inpainting tasks into two distinct types based on mask shape: random brush masks and segmentation-based masks. We utilize EditBench [37] as the comparing benchmark for random brush mask inpainting. Additionally, we introduce a new training dataset BrushData and a new benchmark BrushBench for training and evaluating segmentation-based mask inpainting. Results show BrushNet achieve state-of-the-art performance across 7 metrics encompassing image quality, masked region preservation, and text alignment." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.536, + 0.388, + 0.551 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.785, + 0.659 + ], + "angle": 0, + "content": "Image inpainting is a classic problem in computer vision, aiming to restore masked regions of an image with plausible and natural content [29, 45]. Previous methods based on traditional techniques [3, 7], Variational Auto-Encoders (VAEs) [27, 54] and Generative Adversarial Networks (GANs) [21, 53, 55] often require auxiliary hand-engineered features but yield poor results. Recently, diffusion-based methods [1, 2, 20, 23, 31, 42] gain popularity due to their impressive high-quality generation, fine-grained control, and output diversity [12, 13, 33]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.786, + 0.809 + ], + "angle": 0, + "content": "Initial attempts at utilizing diffusion models for text-guided inpainting [1, 2,6,20,23,47,50], such as Blended Latent Diffusion, modifies the standard denoising strategy by sampling the masked regions from a pre-trained diffusion model and the unmasked areas from the given image, which is commonly used as the default inpainting choice in widely-used image generation libraries like Diffusers [28]. Although these methods demonstrate satisfactory results in simple image inpainting tasks and can be plug-and-play to any diffusion model, they struggle with complex mask shapes, image contents, and text prompts, leading to results that lack coherence. This is primarily attributed to their limited perceptual knowledge of mask boundaries and the unmasked image region context." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Previous works [4,33,37,42,43,46,49,56] address this issue by fine-tuning the base models into content-aware and shape-aware models specifically designed" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.128 + ], + "angle": 0, + "content": "Xuan Ju, et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "for image inpainting. Specifically, SmartBrush [42] augments the diffusion U-Net with object-mask prediction to guide the sampling process with mask boundaries information. Stable Diffusion Inpainting [33] fine-tunes a diffusion model specifically designed for inpainting tasks, taking the mask, masked image, and noisy latent as inputs to UNet architecture. HD-Painter [25] and PowerPaint [56] build upon the foundation of Stable Diffusion Inpainting, separately enhancing the generation quality and enabling the model to perform multiple tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.374 + ], + "angle": 0, + "content": "However, these approaches make it difficult to effectively transfer their inpainting ability to arbitrary pre-trained models, restricting their applicability. To enable any diffusion model with inpainting capabilities, the community fine-tunes ControlNet [51] on inpainting image pairs. However, the model design of ControlNet exhibits limitations in its perceptual understanding of masks and masked images, which consequently leads to unsatisfactory outcomes. Compared with previous methods (shown in Tab. 1), BrushNet is plug-and-play, content-aware, and shape-aware, with a flexible preserving degree for unmasked regions." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.381, + 0.788, + 0.451 + ], + "angle": 0, + "content": "Table 1: Comparison of BrushNet with Previous Image Inpainting Methods. BrushNet offers the advantage of being plug-and-play with any pretrained diffusion model. Moreover, it allows for flexible control over the scale of inpainting and is designed to be aware of both the mask shape and the unmasked content. Note that we only list commonly used text-guided diffusion methods in this table." + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.453, + 0.782, + 0.585 + ], + "angle": 0, + "content": "
ModelPlug-and-PlayFlexible-ScaleContent-AwareShape-Aware
Blended Diffusion [1,2]✓
SmartBrush [42]✓
SD Inpainting [33]✓✓
PowerPaint [56]✓✓
HD-Painter [25]✓✓
ReplaceAnything [4]✓✓
Imagen [37]✓✓
ControlNet-Inpainting [51]✓✓✓
BrushNet✓✓✓✓
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.594, + 0.536, + 0.609 + ], + "angle": 0, + "content": "3 Preliminaries and Motivation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.787, + 0.658 + ], + "angle": 0, + "content": "In this section, we will first introduce diffusion models in Sec. 3.1. Then, Sec. 3.2 would review previous inpainting techniques based on sampling strategy modification and special training. Finally, the motivation is outlined in Section 3.3." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.664, + 0.403, + 0.678 + ], + "angle": 0, + "content": "3.1 Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.684, + 0.788, + 0.744 + ], + "angle": 0, + "content": "Diffusion models include a forward process that adds Gaussian noise \\(\\epsilon\\) to convert clean sample \\(z_0\\) to noise sample \\(z_T\\), and a backward process that iteratively performs denoising from \\(z_T\\) to \\(z_0\\), where \\(\\epsilon \\sim \\mathcal{N}(0,1)\\), and \\(T\\) represents the total number of timesteps. The forward process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.758, + 0.786, + 0.775 + ], + "angle": 0, + "content": "\\[\nz _ {t} = \\sqrt {\\alpha_ {t}} z _ {0} + \\sqrt {1 - \\alpha_ {t}} \\epsilon \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.78, + 0.786, + 0.795 + ], + "angle": 0, + "content": "\\(z_{t}\\) is the noised feature at step \\(t\\) with \\(t\\sim [1,T]\\), and \\(\\alpha\\) is a hyper-parameter." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In the backward process, given input noise \\( z_{T} \\) sampled from a random Gaussian distribution, learnable network \\( \\epsilon_{\\theta} \\) estimates noise at each step \\( t \\) conditioned on \\( C \\). After \\( T \\) progressively refining iterations, \\( z_{0} \\) is derived as the output sample:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.665, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BrushNet" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "equation", + "bbox": [ + 0.274, + 0.171, + 0.786, + 0.211 + ], + "angle": 0, + "content": "\\[\nz _ {t - 1} = \\frac {\\sqrt {\\alpha_ {t - 1}}}{\\sqrt {\\alpha_ {t}}} z _ {t} + \\sqrt {\\alpha_ {t - 1}} \\left(\\sqrt {\\frac {1}{\\alpha_ {t - 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {t}} - 1}\\right) \\epsilon_ {\\theta} \\left(z _ {t}, t, C\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.213, + 0.784, + 0.243 + ], + "angle": 0, + "content": "The training of diffusion models revolves around optimizing the denoiser network \\(\\epsilon_{\\theta}\\) to conduct denoising with condition \\(C\\), guided by the objective:" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.252, + 0.786, + 0.272 + ], + "angle": 0, + "content": "\\[\n\\underset {\\theta} {\\min } E _ {z _ {0}, \\epsilon \\sim \\mathcal {N} (0, I), t \\sim U (1, T)} \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, C) \\| \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.281, + 0.492, + 0.296 + ], + "angle": 0, + "content": "3.2 Previous Inpainting Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.302, + 0.785, + 0.453 + ], + "angle": 0, + "content": "Sampling Strategy Modification. This line of research achieves inpainting by gradually blending masked images with the generated results. The most used method among them is Blended Latent Diffusion (BLD) [1], serving as the default choice for inpainting in widely-used diffusion-based image generation libraries (e.g., Diffusers [28]). Given a binary mask \\( m \\) and a masked image \\( x_0^{masked} \\), BLD first extracts the latent representation \\( z_0^{masked} \\) of the masked image using VAE. Subsequently, the mask \\( m \\) is resized to \\( m^{resized} \\) to match the size of the latent representation. To formulate the inpainting process, BLD adds Gaussian noise to \\( z_0^{masked} \\) for \\( T \\) steps and gets \\( z_t^{masked} \\), where \\( t \\sim [1,T] \\). Then, denoising steps start from \\( z_T^{masked} \\), where each sampling step in eq. 2 is followed by:" + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.463, + 0.785, + 0.481 + ], + "angle": 0, + "content": "\\[\nz _ {t - 1} \\leftarrow z _ {t - 1} \\cdot \\left(1 - m ^ {\\text {r e s i z e d}}\\right) + z _ {t - 1} ^ {\\text {m a s k e d}} \\cdot m ^ {\\text {r e s i z e d}} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.784, + 0.562 + ], + "angle": 0, + "content": "Despite its simplicity in implementation, BLD exhibits suboptimal performance in terms of both unmasked region preservation and generation content alignment. This is due to (1) the resize of the mask preventing it from correctly blending the noisy latent, (2) the diffusion model lacking perceptual knowledge of mask boundaries and the unmasked image region context." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.766 + ], + "angle": 0, + "content": "Dedicated Inpainting Models. To enhance the performance of inpainting, previous works fine-tune the base models by expanding the input UNet channel to include the mask and masked image inputs, turning it into an architecture specifically designed for image inpainting. Though having better generation results compared with BLD, they still have several drawbacks: (1) These models merge the noisy latent, masked image latent, and mask at the initial convolution layer of the UNet architecture, where they are collectively influenced by the text embedding. Consequently, subsequent layers in the UNet model struggle to obtain pure masked image features due to the text's influence. (2) Incorporating both the condition processing and generation within a single branch places additional burdens on the UNet framework. (3) These approaches require extensive fine-tuning across various variations of diffusion backbones, which is computationally intensive and lacks transferability to custom diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.776, + 0.355, + 0.79 + ], + "angle": 0, + "content": "3.3 Motivation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Based on the analysis presented in Section 3.2, a more effective architecture design of inpainting would be introducing an additional branch specifically dedicated to masked image processing. ControlNet [51] is one of the widely adopted" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.128 + ], + "angle": 0, + "content": "Xuan Ju, et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "strategies that exemplifies this idea. However, it should be noted that directly fine-tuning ControlNet, which is originally designed for controllable image generation, on the inpainting task yields unsatisfactory results. ControlNet designs a lightweight encoder to incorporate out-of-domain structural conditions (e.g., skeleton) and relies on text guidance for content generation, which is unsuitable for pixel-level inpainting image feature injection. Furthermore, ControlNet typically relies on sparse control, meaning that merely adding control to residuals in the UNet framework would be sufficient, while inpainting requires pixel-to-pixel constraints with strong restrictive information. Thus a new architecture specifically designed for inpainting is urgently needed." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.309, + 0.331, + 0.325 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.334, + 0.786, + 0.395 + ], + "angle": 0, + "content": "An overview of BrushNet is shown in Fig. 3. We employ a dual-branch strategy for masked image guidance insertion (Sec. 4.1). Blending operations with a blurred mask is used to ensure better-unmasked region preservation (Sec. 4.2). Notably, BrushNet can achieve flexible control by adjusting the added scale." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.411, + 0.785, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.634, + 0.788, + 0.733 + ], + "angle": 0, + "content": "Fig. 3: Model overview. Our model outputs an inpainted image given the mask and masked image input. Firstly, we downsample the mask to accommodate the size of the latent, and input the masked image to the VAE encoder to align the distribution of latent space. Then, noisy latent, masked image latent, and downsampled mask are concatenated as the input of BrushNet. The feature extracted from BrushNet is added to pretrained UNet layer by layer after a zero convolution block [51]. After denoising, the generated image and masked image are blended with a blurred mask." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.745, + 0.467, + 0.761 + ], + "angle": 0, + "content": "4.1 Masked Image Guidance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The insertion of the masked image feature into the pre-trained diffusion network is accomplished through an additional branch, which explicitly separates the feature extraction of masked images from the image-generating processes. The input to the additional branch includes noisy latent, masked image latent, and downsampled mask, which are concatenated together to form the input." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.665, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BrushNet" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.236 + ], + "angle": 0, + "content": "Specifically, the noisy latent provides generation information during the present generation process, helping BrushNet enhance the semantic coherence of the masked image feature. The masked image latent is extracted from the masked image using VAE, which is aligned with the data distribution of the pre-trained UNet. To ensure the alignment of the mask size with the noisy latent and masked image latent, we employ cubic interpolation to downsample the mask." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.785, + 0.388 + ], + "angle": 0, + "content": "To process the masked image features, BrushNet utilizes a clone of the pre-trained diffusion model while excluding its cross-attention layers. The pretrained weights of the diffusion model serve as a strong prior for extracting the masked image features, while the removal of the cross-attention layers ensures that only pure image information is considered within this additional branch. BrushNet feature is inserted into the frozen diffusion model layer-by-layer, enabling dense per-pixel control hierarchically. Similar to ControlNet [51], we employ zero convolution layers to establish a connection between the locked model and the trainable BrushNet. This ensures that harmful noise does not influence the hidden states in the trainable copy during the initial stages of training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.785, + 0.495 + ], + "angle": 0, + "content": "The feature insertion operation is shown in Eq. 5. Specifically, \\(\\epsilon_{\\theta}(z_t,t,C)_i\\) indicates the feature of the \\(i\\)-th layer in network \\(\\epsilon_{\\theta}\\) with \\(i\\sim [1,n]\\), where \\(n\\) is the number of layers. The same notation applies to \\(\\epsilon_{\\theta}^{BrushNet}\\). \\(\\epsilon_{\\theta}^{BrushNet}\\) takes the concatenated noisy latent \\(z_{t}\\), masked image latent \\(z_0^{masked}\\), and downsampled mask \\(m^{resized}\\) as input, with the concatenation operation denoted as \\([\\cdot]\\). \\(\\mathcal{Z}\\) is the zero convolution operation. \\(w\\) is the preservation scale used to adjust the influence of BrushNet on pretrained diffusion model." + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.508, + 0.785, + 0.527 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {\\theta} \\left(z _ {t}, t, C\\right) _ {i} = \\epsilon_ {\\theta} \\left(z _ {t}, t, C\\right) _ {i} + w \\cdot \\mathcal {Z} \\left(\\epsilon_ {\\theta} ^ {\\text {B r u s h N e t}} \\left(\\left[ z _ {t}, z _ {0} ^ {\\text {m a s k e d}}, m ^ {\\text {r e s i z e d}} \\right], t\\right) _ {i}\\right) \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.544, + 0.427, + 0.56 + ], + "angle": 0, + "content": "4.2 Blending Operation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.784, + 0.658 + ], + "angle": 0, + "content": "As mentioned in Section 4.2, the blending operation conducted in latent space can result in inaccuracies due to the resizing of the mask. Similarly, in our approach, a similar issue arises as we resize the mask to match the size of the latent space, which can introduce potential inaccuracies. Additionally, it is important to acknowledge that VAE encoding and decoding operations have inherent limitations and may not ensure complete image reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.785, + 0.764 + ], + "angle": 0, + "content": "To ensure a fully consistent image reconstruction of the unmasked region, previous works have explored different techniques. Some approaches [4, 56], utilize past-and-copy methods, where the unmasked region is directly copied from the original image. However, this can result in a lack of semantic coherence in the final generation results. On the other hand, adopting latent blending operations inspired by BLD [1, 33] has been observed to face challenges in effectively preserving the desired information in the unmasked regions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.784, + 0.84 + ], + "angle": 0, + "content": "In this work, we present a simple pixel space solution to address this issue by first blurring the mask and then performing copy-and-paste using the blurred mask. Although this approach may result in a slight loss of accuracy in preserving the details of the mask boundary, the error is nearly imperceptible to the naked eye and results in significantly improved coherence in the mask boundary." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.128 + ], + "angle": 0, + "content": "Xuan Ju, et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.398, + 0.161 + ], + "angle": 0, + "content": "4.3 Flexible Control" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.168, + 0.788, + 0.396 + ], + "angle": 0, + "content": "The architecture design of BrushNet inherently makes it suitable for seamless plug-and-play integration to various pretrained diffusion models and enables flexible preservation scale. Specifically, the flexible control of our proposed BrushNet includes: (1) Since BrushNet does not modify the weights of the pretrained diffusion model, it can be readily integrated as a plug-and-play component with any community fine-tuned diffusion models. This allows for easy adoption and experimentation with different pretrained models. (2) Preservation Scale Adjustment: The preservation scale of the unmasked region can be controlled by incorporating BrushNet features into the frozen diffusion model with the weight \\( w \\). This weight determines the influence of BrushNet on the preservation scale, offering the ability to adjust the desired level of preservation. (3) Blurring Scale and Blending Operation: By adjusting the scale of blurring and deciding whether to apply the blending operation, the preservation scale of the unmasked region can be further customized. These features allow for flexible fine-grained control over the inpainting process. More explanation can be found in Sec. 5.5." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.406, + 0.377, + 0.423 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.428, + 0.555, + 0.442 + ], + "angle": 0, + "content": "5.1 Evaluation Benchmark and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.553 + ], + "angle": 0, + "content": "Benchmark. Previous commonly used datasets in the image inpainting field include CelebA [22], CelebA-HQ [14], ImageNet [8], MSCOCO [19], Open Images [17], and LSUN-Bedroom [48]. However, these datasets either primarily focus on a small area, such as human faces, or predominantly consist of low-quality, cluttered real-life scene data. As a result, these datasets are not well-suited for training and evaluating diffusion-based inpainting models, which can generate high-quality diverse images that align with text prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Recently proposed EditBench [37] serves as a benchmark specifically designed for text-guided image inpainting for diffusion models. This benchmark consists of a collection of 240 images comprising an equal ratio of natural images and generated images, with mask and caption annotation for each image. However, the annotated masks in EditBench are mostly random shapes without specific object information, neglecting the practical application of inpainting in real scenarios such as replacing an object with an external mask, as commonly seen in E-commerce product displays and image editing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.795 + ], + "angle": 0, + "content": "To fill the gap, we propose BrushBench for segmentation-based inpainting, as shown in Fig. 4. BrushBench comprises a total of 600 images, with each image accompanied by the human-annotated mask and caption annotation. The images in BrushBench are evenly distributed between natural images and artificial images, such as paintings. Furthermore, the dataset ensures an equal distribution among different categories, including humans, animals, indoor scenarios, and outdoor scenarios. This balanced distribution enables a fair evaluation across various categories, promoting better evaluation equity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To further enhance the analysis of the inpainting task, we categorize it into two distinct types based on the masks used: random brush masks and segmentation-based masks. We use EditBench as the comparison benchmark" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.666, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BrushNet" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.146, + 0.78, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.349, + 0.788, + 0.432 + ], + "angle": 0, + "content": "Fig. 4: Benchmark overview. I and II separately show natural and artificial images, masks, and caption of BrushBench. (a) to (d) show images of humans, animals, indoor scenarios, and outdoor scenarios. Each group of images shows the original image, inside-inpainting mask, and outside-inpainting mask, with an image caption on the top. III show image, mask, and caption from EditBench [37], with (e) for generated images and (f) for natural images. The images are randomly selected from both benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.788, + 0.512 + ], + "angle": 0, + "content": "for random brush masks and use BrushBench for segmentation-based masks. Regarding inpainting with segmentation-based masks, we refine the task by considering two specific scenarios: segmentation mask inside-inpainting and segmentation mask outside-inpainting. By separating these two subtasks, we can better understand the inpainting performance in different image regions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.514, + 0.788, + 0.604 + ], + "angle": 0, + "content": "Dataset. To train segmentation-based mask inpainting, we annotate segmentation mask on Laion-Aesthetic [34] dataset, called BrushData. We employ the Grounded-SAM [32] to annotate open-world masks and then filter the masks based on their confidence score, ensuring that only masks with relatively higher confidence scores are retained. Additionally, we consider factors such as reasonable mask size and good mask continuity during the filtering process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.787, + 0.637 + ], + "angle": 0, + "content": "Metrics. We consider 7 metrics from three aspects: image generation quality, masked region preservation, and text alignment." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.639, + 0.788, + 0.759 + ], + "angle": 0, + "content": "- Image Generation Quality. Metrics most used by previous inpainting methods (e.g., FID [11] and KID [5]) show a poor representation of the rich and varied content generated by modern text-to-image models [16]. Thus, we use Image Reward (IR) [44], HPS v2 (HPS) [41], and Aesthetic Score (AS) [34] that align with human perception. Specifically, ImageReward and HPS v2 are text-to-image human preference evaluation models trained on large-scale datasets of human preference choices on generated images. Aesthetic Score is a linear model trained on image quality rating pairs of real images." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.759, + 0.788, + 0.819 + ], + "angle": 0, + "content": "- Masked Region Preservation. We follow previous works using standard Peak Signal-to-Noise Ratio (PSNR) [39], Learned Perceptual Image Patch Similarity (LPIPS) [52], and Mean Squared Error (MSE) [38] in the unmasked region among the generated image and the original image." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.639, + 0.788, + 0.819 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.771, + 0.841 + ], + "angle": 0, + "content": "1 The proposed BrushData and BrushBench will be released along with the codes." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.128 + ], + "angle": 0, + "content": "Xuan Ju, et al." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.146, + 0.788, + 0.209 + ], + "angle": 0, + "content": "- Text Alignment. We use CLIP Similarity (CLIP Sim) [40] to evaluate text-image consistency between the generated images and corresponding text prompts. CLIP Similarity projects text and images to the same shared space with the CLIP model [30] and evaluates the similarity of their embeddings." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.225, + 0.458, + 0.24 + ], + "angle": 0, + "content": "5.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.249, + 0.788, + 0.385 + ], + "angle": 0, + "content": "We perform the inference of different inpainting methods in the same setting unless specifically clarified, i.e., on NVIDIA Tesla V100 following their open-source code with a base model of Stabe Diffusion v1.5 in 50 steps, with a guidance scale of 7.5. We keep the recommended hyper-parameter for each inpainting method in all images for fair comparison. BrushNet and all ablation models are trained for 430 thousands steps on 8 NVIDIA Tesla V100 GPUs, which takes around 3 days. For comparison on BrushBench, we use BrushNet trained on BrushData. For comparison on EditBench, we use the model trained on LAION-5B with random masks. Details can be found in the provided codes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.397, + 0.472, + 0.412 + ], + "angle": 0, + "content": "5.3 Quantitative Comparison" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.421, + 0.788, + 0.517 + ], + "angle": 0, + "content": "Table 2: Quantitative comparisons among BrushNet and other diffusion-based inpainting models in BrushBench: Blended Latent Diffusion (BLD) [1], Stable Diffusion Inpainting (SDI) [33], HD-Painter (HDP) [25], PowerPaint (PP) [56], and ControlNet-Inpainting (CNI) [51]. Metrics encompassing image quality, masked region preservation, and text alignment (Text Align) for inside-inpainting and outside-inpainting are shown in the table. All models use Stable Diffusion V1.5 as base model. Red stands for the best result, Blue stands for the second best result." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.525, + 0.779, + 0.742 + ], + "angle": 0, + "content": "
MetricsImage QualityMasked Region PreservationText Align
ModelsIR×10↑HPS×102↑AS↑PSNR↑MSE×103↓LPIPS×103↓CLIP Sim↑
InsideBLD [1]9.7825.876.1721.339.7649.2626.15
SDI [33]11.7227.066.5021.5213.8748.3926.17
HDP [25]11.6826.906.4222.619.9543.5026.37
PP [56]11.4627.356.2421.4332.7348.4326.48
CNI [51]9.926.026.5312.3978.78243.6226.47
CNI* [51]11.2126.926.3922.7324.5843.4926.22
Ours12.3627.406.5321.659.3148.2826.48
Ours*12.6427.786.5131.940.8018.6726.39
OutsideBLD [1]7.8126.776.2315.8535.8621.4026.73
SDI [33]10.2727.996.5518.0419.8715.1327.21
HDP [25]9.6627.796.4618.0322.9915.2226.96
PP [56]7.4528.016.2618.0431.7815.1326.72
CNI [51]9.2627.686.4211.9183.0358.1627.29
CNI* [51]9.5727.766.2817.5037.7219.9526.92
Ours10.8228.026.6418.0622.8615.0827.33
Ours*10.8828.096.6427.822.254.6327.22
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.236, + 0.744, + 0.409, + 0.76 + ], + "angle": 0, + "content": "* with blending operation" + }, + { + "type": "table_footnote", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Tab. 2 and Tab. 3 show the quantitative comparison on BrushBench and EditBench [37]. We compare the inpainting results of sampling strategy modification method Blended Latent Diffusion [1], dedicated inpainting models Stable Diffusion Inpainting [33], HD-Painter [25], and PowerPaint [56], as well as plug-and-play method ControlNet [51] trained on inpainting data." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.744, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.665, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BrushNet" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "Results demonstrate BrushNet's effectiveness across image quality, masked region preservation, and image-text alignment. Blended Latent Diffusion [1] shows the poorest results in image quality and text alignment, derived from the incoherence between the generated masked and the unmasked given image. At the same time, its performance in masked region preservation is also not satisfactory because of the loss incurred by resized mask blending operation in the latent space. Modified from Stable Diffusion Inpainting [33], HD-Painter [25] and PowerPaint [56] demonstrate comparable performance to Stable Diffusion Inpainting in the task of inside-inpainting. However, when it comes to outside-inpainting, their results in terms of image quality and text alignment are significantly poorer compared to Stable Diffusion Inpainting, which can be attributed to their exclusive emphasis on the inside-inpainting task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.421 + ], + "angle": 0, + "content": "ControlNet [51] trained on inpainting has the most similar experimental configuration to ours. Due to the mismatch between its model design and the inpainting task, ControlNet shows poor results in masked region preservation and image quality, necessitating its combination with Blended Latent Diffusion [1] to generate satisfying inpainted images. However, even with this combination, it still falls short compared to dedicated inpainting models and BrushNet." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.43, + 0.79, + 0.486 + ], + "angle": 0, + "content": "Table 3: Quantitative comparisons among BrushNet and other diffusion-based inpainting models in EditBench. A detailed explanation of compared methods and metrics can be found in the caption of Tab. 2. Red stands for the best result, Blue stands for the second best result." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.489, + 0.767, + 0.639 + ], + "angle": 0, + "content": "
MetricsImage QualityMasked Region PreservationText Align
ModelsIR×10↑HPS×102↑AS↑PSNR↑MSE×103↓LPIPS×103↓CLIP Sim↑
BLD [1]0.9023.815.4420.8910.9331.9028.62
SDI [33]1.8624.245.6923.256.9424.3028.00
HDP [25]1.7424.205.6423.076.7024.3228.34
PP [56]1.2424.505.4423.3420.1224.1227.80
CNI [51]1.4924.465.8212.7169.42159.7128.16
CNI* [51]0.9023.795.4622.6135.9326.1427.74
Ours4.4025.105.8423.356.8124.1128.67
Ours*4.4625.245.8233.660.6310.1228.87
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.245, + 0.641, + 0.418, + 0.656 + ], + "angle": 0, + "content": "* with blending operation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.665, + 0.788, + 0.743 + ], + "angle": 0, + "content": "The performance on the EditBench is roughly consistent with the overall performance on BrushBench, which similarly shows BrushNet's superior performance. This indicates that our method exhibits strong performance across a range of inpainting tasks with various mask types, including random masks, inside-inpainting masks, and outside-inpainting masks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.755, + 0.46, + 0.772 + ], + "angle": 0, + "content": "5.4 Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The qualitative comparison with previous image inpainting methods is shown in Fig. 1. We provide results on artificial images and natural images across various inpainting tasks, including random mask inpainting, segmentation mask inside-inpainting, and segmentation mask outside-inpainting. BrushNet consistently" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.128 + ], + "angle": 0, + "content": "Xuan Ju, et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "show exceptional results in the coherent of generated region and unmasked region, considering content (I, II right, III right, IV), color (II left), and text (III left). Interestingly, Fig. 1 III left requires the model to generate a cat and a goldfish. All previous methods fail to recognize that a goldfish is already present in the masked image, resulting in the generation of an additional fish within the masked region. BrushNet successfully realized the awareness of background information due to the design of the dual-branch decoupling." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.258, + 0.778, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.543, + 0.788, + 0.573 + ], + "angle": 0, + "content": "Fig. 5: Comparison of previous inpainting methods and BrushNet on various image domain. A detailed explanation of compared methods is in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.577, + 0.788, + 0.669 + ], + "angle": 0, + "content": "The untouched pre-trained diffusion branch also provides the advantage of better coverage across different data domains, such as painting and anime. As shown in Fig. 5, BrushNet demonstrates superior performance across various image categories, including natural image (I, II), pencil painting (III), anime (IV), illustration (V), digital art (VI left), and watercolor (VI right). Due to the page limit, more qualitative comparison results are in supplementary files." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.676, + 0.461, + 0.691 + ], + "angle": 0, + "content": "5.5 Flexible Control Ability" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.701, + 0.789, + 0.854 + ], + "angle": 0, + "content": "Fig. 6 and Fig. 7 illustrate the flexible control provided by BrushNet in two aspects: base diffusion model selection and controlling scale. In Fig. 6, we showcase the ability to combine BrushNet with different diffusion models fine-tuned by the community. This allows users to select a specific model that best suits their inpainting requirements, enabling users to achieve the desired inpainting results based on their specific needs. Fig. 7 demonstrates the adjustment of the control scale of BrushNet. This control scale parameter allows users to effectively control the extent of unmasked region protection during the inpainting process. By manipulating the scale parameter, users can achieve fine-grained control over the inpainting process, enabling precise and customizable inpainting." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.666, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BrushNet" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "aside_text", + "bbox": [ + 0.22, + 0.173, + 0.228, + 0.183 + ], + "angle": 0, + "content": "I" + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.147, + 0.486, + 0.159 + ], + "angle": 0, + "content": "A single red rose is laying on top of snow" + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.159, + 0.373, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.159, + 0.508, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.147, + 0.679, + 0.159 + ], + "angle": 0, + "content": "A cartoon boy driving a car" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.159, + 0.693, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.159, + 0.735, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.16, + 0.787, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "aside_text", + "bbox": [ + 0.22, + 0.224, + 0.228, + 0.234 + ], + "angle": 0, + "content": "I" + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.198, + 0.503, + 0.21 + ], + "angle": 0, + "content": "A little girl sitting on the grass with her dog" + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.21, + 0.373, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.21, + 0.508, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.2, + 0.73, + 0.21 + ], + "angle": 0, + "content": "A woman with her hands on her face" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.211, + 0.694, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.211, + 0.784, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.249, + 0.282, + 0.258 + ], + "angle": 0, + "content": "(a) Mask" + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.249, + 0.316, + 0.258 + ], + "angle": 0, + "content": "(b) DS" + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.249, + 0.365, + 0.258 + ], + "angle": 0, + "content": "(c) ER" + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.249, + 0.426, + 0.258 + ], + "angle": 0, + "content": "(d) HR (e)" + }, + { + "type": "image_caption", + "bbox": [ + 0.427, + 0.249, + 0.454, + 0.258 + ], + "angle": 0, + "content": ")MM" + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.249, + 0.497, + 0.258 + ], + "angle": 0, + "content": "(f) RV" + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.249, + 0.556, + 0.259 + ], + "angle": 0, + "content": "(a) Mask" + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.249, + 0.595, + 0.258 + ], + "angle": 0, + "content": "(b) DS" + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.249, + 0.639, + 0.259 + ], + "angle": 0, + "content": "(c) ER" + }, + { + "type": "image_caption", + "bbox": [ + 0.65, + 0.249, + 0.684, + 0.258 + ], + "angle": 0, + "content": "(d) HR" + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.249, + 0.729, + 0.258 + ], + "angle": 0, + "content": "(e) MM" + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.249, + 0.772, + 0.258 + ], + "angle": 0, + "content": "(f) RV" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.271, + 0.787, + 0.328 + ], + "angle": 0, + "content": "Fig. 6: Integrating BrushNet to community fine-tuned diffusion models. We use five popular community diffusion models fine-tuned from stable diffusion v1.5: DreamShaper (DS) [24], epiCRealism (ER) [9], Henmix_Real (HR) [10], MeinaMix (MM) [26], and Realistic Vision (RV) [35]. MM is specifically designed for anime images." + }, + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.342, + 0.364, + 0.353 + ], + "angle": 0, + "content": "Book, coffee and purple flower" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.353, + 0.5, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.382, + 0.382, + 0.391 + ], + "angle": 0, + "content": "A cupcake with a lit candle on top" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.392, + 0.5, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.222, + 0.42, + 0.256, + 0.428 + ], + "angle": 0, + "content": "(a) Mask" + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.42, + 0.285, + 0.428 + ], + "angle": 0, + "content": "(b) 1.0" + }, + { + "type": "image_caption", + "bbox": [ + 0.287, + 0.42, + 0.32, + 0.428 + ], + "angle": 0, + "content": "(c) 0.8" + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.42, + 0.355, + 0.428 + ], + "angle": 0, + "content": "(d) 0.6" + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.42, + 0.382, + 0.428 + ], + "angle": 0, + "content": "(e) 0" + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.42, + 0.423, + 0.428 + ], + "angle": 0, + "content": "(1)0.4" + }, + { + "type": "image_caption", + "bbox": [ + 0.434, + 0.42, + 0.461, + 0.428 + ], + "angle": 0, + "content": "(g) 0.3" + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.342, + 0.686, + 0.352 + ], + "angle": 0, + "content": "A small island with a tree and clouds" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.353, + 0.784, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.382, + 0.739, + 0.392 + ], + "angle": 0, + "content": "A white rabbit sitting on the ground in the grass" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.392, + 0.784, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.505, + 0.42, + 0.568, + 0.428 + ], + "angle": 0, + "content": "(a) Mask (b) 1.0" + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.42, + 0.604, + 0.428 + ], + "angle": 0, + "content": "(c) 0.8" + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.42, + 0.64, + 0.428 + ], + "angle": 0, + "content": "(d) 0.6" + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.42, + 0.676, + 0.428 + ], + "angle": 0, + "content": "(e) 0.5" + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.42, + 0.716, + 0.428 + ], + "angle": 0, + "content": "(f) 0.4" + }, + { + "type": "image_caption", + "bbox": [ + 0.72, + 0.42, + 0.748, + 0.428 + ], + "angle": 0, + "content": "(g) 0.3" + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.441, + 0.787, + 0.484 + ], + "angle": 0, + "content": "Fig. 7: Flexible control scale of BrushNet. (a) shows the given masked image, (b)-(h) show adding BrushNet with control scale \\( w \\) from 1.0 to 0.2. Results show a gradually diminishing controllable ability from precise to rough control." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.498, + 0.389, + 0.513 + ], + "angle": 0, + "content": "5.6 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.52, + 0.787, + 0.59 + ], + "angle": 0, + "content": "Table 4: Ablation on dual-branch design. Stable Diffusion Inpainting (SDI) use single-branch design, where the entire UNet is fine-tuned. We conducted an ablation analysis by training a dual-branch model with two variations: one with the base UNet fine-tuned, and another with the base UNet forzened. Results demonstrate the superior performance achieved by adopting the dual-branch design. Red is the best result." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.598, + 0.774, + 0.678 + ], + "angle": 0, + "content": "
MetricsImage QualityMasked Region PreservationText Align
ModelIR×10↑HPS×102↑AS↑PSNR↑MSE×102↓LPIPS×103↓CLIP Sim↑
SDI11.0027.536.5319.7816.8731.7626.69
w/o fine-tune11.5927.716.5919.8616.0931.6826.91
w/ fine-tune11.6327.736.6020.1315.8431.5726.93
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We conducted ablation studies to investigate the impact of different model designs. Tab. 4 compares the dual-branch and single-branch designs. Tab. 5 shows the ablation study focusing on the additional branch architecture. The ablation studies are conducted on BrushBench, averaging the performance of inside-inpainting and outside-inpainting. The results presented in Tab. 4 demonstrate that the dual-branch design significantly outperforms the single-branch design. Additionally, fine-tuning the base diffusion model in dual-branch design yields better results compared to freezing it. However, fine-tuning the base diffusion model may restrict flexibility and control over the model. Considering this trade-off between performance and flexibility, we decide to adopt the frozen" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.128 + ], + "angle": 0, + "content": "Xuan Ju, et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.223 + ], + "angle": 0, + "content": "dual-branch design as our model design. Tab. 5 presents the rationale behind the design choices for (1) using a VAE encoder instead of randomly initialized convolution layers to process the masked image, (2) incorporating the full UNet feature layer-by-layer into the pre-trained UNet, (3) removing text cross-attention in BrushNet, which avoids masked image features influenced by text." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.225, + 0.788, + 0.337 + ], + "angle": 0, + "content": "Table 5: Ablation on model architecture. We ablate on the following components: the image encoder (Enc), selected from a random initialized convolution (Conv) and a VAE; the inclusion of mask in input (Mask), chosen from adding (w/) and not adding (w/o); the presence of cross-attention layers (Attn), chosen from adding (w/) and not adding (w/o); the type of UNet feature addition (UNet), selected from adding the full UNet feature (full), adding half of the UNet feature (half), and adding the feature like ControlNet (CN); and finally, the blending operation (Blend), chosen from not adding (w/o), direct pasting (paste), and blurred blending (blur). Red is the best result." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.343, + 0.767, + 0.477 + ], + "angle": 0, + "content": "
MetricsImage QualityMasked Region PreservationText Align
EncMaskAttnUNetBlendIR×10↑HPS×102↑AS↑PSNR↑MSE×102↓LPIPS×103↓CLIP Sim↑
Convw/w/ofullw/o11.0526.236.5514.8937.2364.5426.76
VAEw/ow/ofullw/o11.5527.706.5717.9626.3849.3326.87
VAEw/w/fullw/o11.2527.626.5618.6919.4434.2826.63
Convw/w/CNw/o9.5826.856.4712.1580.91150.8926.88
VAEw/w/CNw/o10.5327.426.5918.2824.3641.6326.89
VAEw/w/oCNw/o11.4227.696.5818.4924.0936.3326.86
VAEw/w/ohalfw/o11.4727.706.5719.0123.7733.5726.87
VAEw/w/ofullw/o11.5927.716.5919.8616.0931.6826.91
VAEw/w/ofullpaste11.7227.936.58---26.80
VAEw/w/ofullblur11.7627.946.5829.881.5311.6526.81
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.487, + 0.355, + 0.503 + ], + "angle": 0, + "content": "6 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.784, + 0.581 + ], + "angle": 0, + "content": "Conclusion. This paper proposes a plug-and-play image inpainting method BrushNet with a pixel-level masked image feature insertion architectural design. Quantitative and qualitative results on our proposed benchmark, BrushBench, and EditBench show the superior performance of BrushNet considering image generation quality, masked region preservation, and image-text alignment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.784, + 0.675 + ], + "angle": 0, + "content": "Limitations and Future Work. However, BrushNet still has some limitations: (1) The quality and content generated by our model are heavily dependent on the chosen base model. (2) Even with BrushNet, we still observe poor generation results in cases where the given mask has an unusually shaped or irregular form, or when the given text does not align well with the masked image. In our future work, we will continue to address these challenges." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.679, + 0.784, + 0.784 + ], + "angle": 0, + "content": "Negative Social Impact. Image inpainting models present exciting opportunities for content creation, but they also carry potential risks to individuals and society. Their reliance on internet-collected training data can amplify social biases, and there is a specific risk of generating persuasive misinformation by manipulating human images with offensive elements. To address these concerns, it is crucial to emphasize responsible use and establish ethical guidelines when utilizing these models. This is also a key focus for our future model releases." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.795, + 0.373, + 0.81 + ], + "angle": 0, + "content": "*Acknowledgment." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.794, + 0.825 + ], + "angle": 0, + "content": "This work is supported in part by Research Matching Grant(CSE-7-2022)-RMG01." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.665, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BrushNet" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.2 + ], + "angle": 0, + "content": "1. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. ACM transactions on graphics (TOG) 42(4), 1-11 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.201, + 0.785, + 0.241 + ], + "angle": 0, + "content": "2. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18208-18218 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.242, + 0.785, + 0.283 + ], + "angle": 0, + "content": "3. Bertalmio, M., Sapiro, G., Caseles, V., Ballester, C.: Image inpainting. In: International Conference and Exhibition on Computer Graphics and Interactive Techniques (SIGGRAPH). pp. 417-424 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.283, + 0.785, + 0.324 + ], + "angle": 0, + "content": "4. Binghui, C., Chao, L., Chongyang, Z., Wangmeng, X., Yifeng, G., Xuansong, X.: Replaceanything as you want: Ultra-high quality content replacement (2023), https://aigcdesigngroup.github.io/replace-anything/" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.325, + 0.785, + 0.35 + ], + "angle": 0, + "content": "5. Binkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying MMD GANs. arXiv preprint arXiv:1801.01401 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.351, + 0.787, + 0.392 + ], + "angle": 0, + "content": "6. Corneanu, C., Gadde, R., Martinez, A.M.: Latentpaint: Image inpainting in latent space with diffusion models. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 4334-4343 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.392, + 0.785, + 0.432 + ], + "angle": 0, + "content": "7. Criminisi, A., Pérez, P., Toyama, K.: Region filling and object removal by exemplar-based image inpainting. IEEE Transactions on Image Processing 13(9), 1200-1212 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.785, + 0.472 + ], + "angle": 0, + "content": "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.785, + 0.499 + ], + "angle": 0, + "content": "9. epinikion: epicrealism (2023), https://civitai.com/models/25694?modelVersionId=143906" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.5, + 0.785, + 0.525 + ], + "angle": 0, + "content": "0. hemi29833: Henmixreal (2024), https://civitai.com/models/20282?modelVersionId=305687" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.527, + 0.785, + 0.568 + ], + "angle": 0, + "content": "1. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs trained by a two time-scale update rule converge to a local Nash equilibrium. Advances in Neural Information Processing Systems (NIPS) 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.568, + 0.785, + 0.595 + ], + "angle": 0, + "content": "2. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NIPS) 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.596, + 0.785, + 0.621 + ], + "angle": 0, + "content": "3. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.622, + 0.785, + 0.662 + ], + "angle": 0, + "content": "4. Huang, H., He, R., Sun, Z., Tan, T., et al.: Introvae: Introspective variational autoencoders for photographic image synthesis. Advances in Neural Information Processing Systems (NIPS) 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.663, + 0.785, + 0.703 + ], + "angle": 0, + "content": "5. Huang, Y., Huang, J., Liu, Y., Yan, M., Lv, J., Liu, J., Xiong, W., Zhang, H., Chen, S., Cao, L.: Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.704, + 0.785, + 0.744 + ], + "angle": 0, + "content": "6. Jayasumana, S., Ramalingam, S., Veit, A., Glasner, D., Chakrabarti, A., Kumar, S.: Rethinking fid: Towards a better evaluation metric for image generation. arXiv preprint arXiv:2401.09603 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.799 + ], + "angle": 0, + "content": "7. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision (IJCV) 128(7), 1956-1981 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "8. Li, Z., Wei, P., Yin, X., Ma, Z., Kot, A.C.: Virtual try-on with pose-garment keypoints guided inpainting. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 22788-22797 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.128 + ], + "angle": 0, + "content": "Xuan Ju, et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "19. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision (ECCV). pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.217 + ], + "angle": 0, + "content": "20. Liu, A., Niepert, M., Broeck, G.V.d.: Image inpainting via tractable steering of diffusion models. arXiv preprint arXiv:2401.03349 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "21. Liu, H., Wan, Z., Huang, W., Song, Y., Han, X., Liao, J.: Pd-GAN: Probabilistic diverse GAN for image inpainting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9371-9381 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.299 + ], + "angle": 0, + "content": "22. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: IEEE/CVF International Conference on Computer Vision (ICCV) (December 2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.356 + ], + "angle": 0, + "content": "23. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11461-11471 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.383 + ], + "angle": 0, + "content": "24. Lykon: Dreamshaper (2022), https://civitai.com/models/4384?modelVersionId=128713" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.424 + ], + "angle": 0, + "content": "25. Manukyan, H., Sargsyan, A., Atanyan, B., Wang, Z., Navasardyan, S., Shi, H.: Hd-painter: High-resolution and prompt-faithful text-guided image inpainting with diffusion models. arXiv preprint arXiv:2312.14091 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.451 + ], + "angle": 0, + "content": "26. Meina: Meinamix (2023), https://civitai.com/models/7240?modelVersionId=119057" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.493 + ], + "angle": 0, + "content": "27. Peng, J., Liu, D., Xu, S., Li, H.: Generating diverse structure for image inpainting with hierarchical vq-vae. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10775-10784 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.535 + ], + "angle": 0, + "content": "28. von Platen, P., Patil, S., Lozhkov, A., Cuenca, P., Lambert, N., Rasul, K., Davaadorj, M., Wolf, T.: Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.787, + 0.577 + ], + "angle": 0, + "content": "29. Quan, W., Chen, J., Liu, Y., Yan, D.M., Wonka, P.: Deep learning-based image and video inpainting: A survey. International Journal of Computer Vision (IJCV) pp. 1-34 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.632 + ], + "angle": 0, + "content": "30. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (ICML). pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.687 + ], + "angle": 0, + "content": "31. Razzhigaev, A., Shakhmatov, A., Maltseva, A., Arkhipkin, V., Pavlov, I., Ryabov, I., Kuts, A., Panchenko, A., Kuznetsov, A., Dimitrov, D.: Kandinsky: an improved text-to-image synthesis with image prior and latent diffusion. arXiv preprint arXiv:2310.03502 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.787, + 0.729 + ], + "angle": 0, + "content": "32. Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., et al.: Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.787, + 0.784 + ], + "angle": 0, + "content": "33. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "34. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems (NIPS) 35, 25278–25294 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.665, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BrushNet" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "35. SG161222: Realisticvision (2023), https://civitai.com/models/4201?modelVersionId=130072" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "36. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.206, + 0.785, + 0.274 + ], + "angle": 0, + "content": "37. Wang, S., Sahara, C., Montgomery, C., Pont-Tuset, J., Noy, S., Pellegrini, S., Onoe, Y., Laszlo, S., Fleet, D.J., Soricut, R., et al.: Imagen editor and editbench: Advancing and evaluating text-guided image inpainting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18359-18369 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.785, + 0.317 + ], + "angle": 0, + "content": "38. Wikipedia contributors: Mean squared error — Wikipedia, the free encyclopedia (2024), https://en.wikipedia.org/w/index.php?title=Mean_squared_error&oldid=1207422018, [Online; accessed 4-March-2024]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.785, + 0.358 + ], + "angle": 0, + "content": "39. Wikipedia contributors: Peak signal-to-noise ratio — Wikipedia, the free encyclopedia (2024), https://en.wikipedia.org/w/index.php?title=Peak_signal-to-noise_ratio&oldid=1210897995, [Online; accessed 4-March-2024]" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.785, + 0.401 + ], + "angle": 0, + "content": "40. Wu, C., Huang, L., Zhang, Q., Li, B., Ji, L., Yang, F., Sapiro, G., Duan, N.: GO-DIVA: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.785, + 0.444 + ], + "angle": 0, + "content": "41. Wu, X., Hao, Y., Sun, K., Chen, Y., Zhu, F., Zhao, R., Li, H.: Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.785, + 0.5 + ], + "angle": 0, + "content": "42. Xie, S., Zhang, Z., Lin, Z., Hinz, T., Zhang, K.: Smartbrush: Text and shape guided object inpainting with diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22428-22437 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.542 + ], + "angle": 0, + "content": "43. Xie, S., Zhao, Y., Xiao, Z., Chan, K.C., Li, Y., Xu, Y., Zhang, K., Hou, T.: Dreaminpainter: Text-guided subject-driven image inpainting with diffusion models. arXiv preprint arXiv:2312.03771 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.785, + 0.584 + ], + "angle": 0, + "content": "44. Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagereward: Learning and evaluating human preferences for text-to-image generation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.586, + 0.785, + 0.613 + ], + "angle": 0, + "content": "45. Xu, Z., Zhang, X., Chen, W., Yao, M., Liu, J., Xu, T., Wang, Z.: A review of image inpainting methods based on deep learning. Applied Sciences 13(20), 11189 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.785, + 0.655 + ], + "angle": 0, + "content": "46. Yang, S., Chen, X., Liao, J.: Uni-paint: A unified framework for multimodal image inpainting with pretrained diffusion model. In: ACM International Conference on Multimedia (MM). pp. 3190-3199 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.785, + 0.697 + ], + "angle": 0, + "content": "47. Yang, S., Zhang, L., Ma, L., Liu, Y., Fu, J., He, Y.: Magicremover: Tuning-free text-guided image inpainting with diffusion models. arXiv preprint arXiv:2310.02848 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.785, + 0.74 + ], + "angle": 0, + "content": "48. Yu, F., Seff, A., Zhang, Y., Song, S., Funkhouser, T., Xiao, J.: Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.785, + 0.769 + ], + "angle": 0, + "content": "49. Yu, T., Feng, R., Feng, R., Liu, J., Jin, X., Zeng, W., Chen, Z.: Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.785, + 0.797 + ], + "angle": 0, + "content": "50. Zhang, G., Ji, J., Zhang, Y., Yu, M., Jaakkola, T., Chang, S.: Towards coherent image inpainting using denoising diffusion implicit models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "51. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.128 + ], + "angle": 0, + "content": "Xuan Ju, et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "52. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 586-595 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "53. Zhao, S., Cui, J., Sheng, Y., Dong, Y., Liang, X., Chang, E.I., Xu, Y.: Large scale image completion via co-modulated generative adversarial networks. arXiv preprint arXiv:2103.10428 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.787, + 0.286 + ], + "angle": 0, + "content": "54. Zheng, C., Cham, T.J., Cai, J.: Pluralistic image completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1438-1447 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.286, + 0.787, + 0.342 + ], + "angle": 0, + "content": "55. Zheng, H., Lin, Z., Lu, J., Cohen, S., Shechtman, E., Barnes, C., Zhang, J., Xu, N., Amirghodsi, S., Luo, J.: Image inpainting with cascaded modulation GAN and object-aware training. In: European Conference on Computer Vision (ECCV). pp. 277-296. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.787, + 0.384 + ], + "angle": 0, + "content": "56. Zhuang, J., Zeng, Y., Liu, W., Yuan, C., Chen, K.: A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. arXiv preprint arXiv:2312.03594 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.384 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/BrushNet_ A Plug-and-Play Image Inpainting Model with Decomposed Dual-Branch Diffusion/8cdf68e6-b7ca-45f7-9a98-9bdc1df71997_origin.pdf b/2024/BrushNet_ A Plug-and-Play Image Inpainting Model with Decomposed Dual-Branch Diffusion/8cdf68e6-b7ca-45f7-9a98-9bdc1df71997_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b361a5c83c976bd509f953ac3ca8fb310a0dc003 --- /dev/null +++ b/2024/BrushNet_ A Plug-and-Play Image Inpainting Model with Decomposed Dual-Branch Diffusion/8cdf68e6-b7ca-45f7-9a98-9bdc1df71997_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6b05e9b2be891eb44a41c574c1830328d03aa1adf7f36ce1761d06fde09eb214 +size 20959295 diff --git a/2024/BrushNet_ A Plug-and-Play Image Inpainting Model with Decomposed Dual-Branch Diffusion/full.md b/2024/BrushNet_ A Plug-and-Play Image Inpainting Model with Decomposed Dual-Branch Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0c13ae03a787d25a52d7daed78e12e9b73f8dbc4 --- /dev/null +++ b/2024/BrushNet_ A Plug-and-Play Image Inpainting Model with Decomposed Dual-Branch Diffusion/full.md @@ -0,0 +1,341 @@ +# BrushNet: A Plug-and-Play Image Inpainting Model with Decomposed Dual-Branch Diffusion + +Xuan $\mathrm{Ju}^{1,2}$ , Xian Liu $^{1,2}$ , Xintao Wang $^{1\star}$ , Yuxuan Bian $^{2}$ , Ying Shan $^{1}$ , and Qiang $\mathrm{Xu}^{2\star}$ + +$^{1}$ ARC Lab, Tencent PCG $^{2}$ The Chinese University of Hong Kong + +https://github.com/TencentARC/BrushNet + +![](images/9a73c4b7f103f0cd858bc52bb37b16167d1524cc4cd0eaf471b5413c80b009c8.jpg) +Fig. 1: Performance comparisons of BrushNet and previous image inpainting methods across various inpainting tasks: (I) Random Mask (< 50% masked), (II) Random Mask (> 50% masked), (III) Segmentation Mask Inside-Inpainting, (IV) Segmentation Mask Outside-Inpainting. Each group of results contains an artificial image (left) and a natural image (right) with 6 inpainting methods: (b) Blended Latent Diffusion (BLD) [1], (c) Stable Diffusion Inpainting (SDI) [33], (d) HD-Painter (HDP) [25], (e) PowerPaint (PP) [56], (f) ControlNet-Inpainting (CNI) [51], and (g) Ours. + +Abstract. Image inpainting, the process of restoring corrupted images, has seen significant advancements with the advent of diffusion models (DMs). Despite these advancements, current DM adaptations for inpainting, which involve modifications to the sampling strategy or the development of inpainting-specific DMs, frequently suffer from semantic inconsistencies and reduced image quality. Addressing these challenges, our work introduces a novel paradigm: the division of masked image features and noisy latent into separate branches. This division dramatically diminishes the model's learning load, facilitating a nuanced incorporation of essential masked image information in a hierarchical fashion. Herein, we present BrushNet, a novel plug-and-play dual-branch model engineered to embed pixel-level masked image features into any pre-trained DM, guaranteeing coherent and enhanced image inpainting outcomes. Additionally, we introduce BrushData and BrushBench to facilitate segmentation-based inpainting training and performance assessment. Our extensive experimental analysis demonstrates BrushNet's superior performance over existing models across seven key metrics, including image quality, mask region preservation, and textual coherence. + +Keywords: Image Inpainting $\cdot$ Diffusion Models $\cdot$ Image Generation + +# 1 Introduction + +Image inpainting [45] aims at restoring the missing regions of an image while maintaining the overall coherence. As a long-standing computer vision problem, it facilitates numerous applications such as virtual try-on [18] and image editing [15]. Recently, diffusion models [12,36] have demonstrated impressive performance in image generation, enabling flexible user control with semantic and structural conditions [33,51]. To this end, researchers resort to diffusion-based pipelines for high-quality image inpainting that aligns with given text prompts. + +Commonly used diffusion-based text-guided inpainting methods can be roughly divided into two categories: (1) Sampling strategy modification [1,2,6,20,23,47, 50], which modifies the standard denoising process by sampling the masked regions from a pre-trained diffusion model, and the unmasked areas are simply copy-pasted from the given image in each denoising step. Although they can be used in arbitrary diffusion backbones, the limited perceptual knowledge of mask boundaries and the unmasked image region context leads to incoherent inpainting results. (2) Dedicated inpainting models [4, 33, 37, 42, 43, 46, 49, 56], which fine-tune a specially designed image inpainting model by expanding the input channel dimension of base diffusion models to incorporate provided corrupted image and mask. While they enable the diffusion model to generate more satisfying results with specialized content-aware and shape-aware models, we argue, is this architecture the best fit for diffusion-based inpainting? + +As shown in Fig. 2, dedicated inpainting models fuse noisy latent, masked image latent, mask, and text at an early stage. This architectural design makes the masked image feature easily influenced by the text embedding, preventing subsequent layers in the UNet from obtaining pure masked image features due to the influence of text. Additionally, handling the condition and generation in a single branch imposes extra burdens on the UNet framework. These approaches also necessitate fine-tuning in different variations of diffusion backbones, which can be time-consuming with limited transferability. + +![](images/c3c37ed7f16079a11ef5a942c52021579f3d07987efc14995ca68236ffa87cb2.jpg) +Fig. 2: Comparison of previous inpainting architectures and BrushNet. + +Adding an additional branch dedicated to masked image feature extraction could be a promising approach to tackle the issue above. However, the existing solutions such as ControlNet [51] lead to inadequate information extraction + +and insertion when directly applied to inpainting, which originates from differences between controllable image generation and inpainting: inpainting requires pixel-to-pixel constraints with strong restrictive information instead of sparse structural control relying on text for content completion. As a result, ControlNet yields unsatisfactory results compared to dedicated inpainting models. + +To address this issue, we propose BrushNet, which introduces an additional branch to the diffusion framework, creating a more suitable architecture for image inpainting. Specifically, our designs are three-fold: (1) To improve the extraction of image features for adaptation to the UNet distribution, we use a VAE encoder instead of randomly initialized convolution layers to process the masked image. (2) To enable dense per-pixel control, we adopt a hierarchical approach by gradually incorporating the full UNet feature layer-by-layer into the pre-trained UNet. (3) To ensure pure image information is considered in the additional branch, we remove text cross-attention from UNet. This design further offers plug-and-play capabilities and flexible unmasked region controllability to the inpainting process. For better consistency and a larger range of unmasked region controllability, we additionally propose a blurred blending strategy. + +To ensure a comprehensive evaluation for real-life applications, we categorize inpainting tasks into two distinct types based on mask shape: random brush masks and segmentation-based masks. We utilize EditBench [37] as the comparing benchmark for random brush mask inpainting. Additionally, we introduce a new training dataset BrushData and a new benchmark BrushBench for training and evaluating segmentation-based mask inpainting. Results show BrushNet achieve state-of-the-art performance across 7 metrics encompassing image quality, masked region preservation, and text alignment. + +# 2 Related Work + +Image inpainting is a classic problem in computer vision, aiming to restore masked regions of an image with plausible and natural content [29, 45]. Previous methods based on traditional techniques [3, 7], Variational Auto-Encoders (VAEs) [27, 54] and Generative Adversarial Networks (GANs) [21, 53, 55] often require auxiliary hand-engineered features but yield poor results. Recently, diffusion-based methods [1, 2, 20, 23, 31, 42] gain popularity due to their impressive high-quality generation, fine-grained control, and output diversity [12, 13, 33]. + +Initial attempts at utilizing diffusion models for text-guided inpainting [1, 2,6,20,23,47,50], such as Blended Latent Diffusion, modifies the standard denoising strategy by sampling the masked regions from a pre-trained diffusion model and the unmasked areas from the given image, which is commonly used as the default inpainting choice in widely-used image generation libraries like Diffusers [28]. Although these methods demonstrate satisfactory results in simple image inpainting tasks and can be plug-and-play to any diffusion model, they struggle with complex mask shapes, image contents, and text prompts, leading to results that lack coherence. This is primarily attributed to their limited perceptual knowledge of mask boundaries and the unmasked image region context. + +Previous works [4,33,37,42,43,46,49,56] address this issue by fine-tuning the base models into content-aware and shape-aware models specifically designed + +for image inpainting. Specifically, SmartBrush [42] augments the diffusion U-Net with object-mask prediction to guide the sampling process with mask boundaries information. Stable Diffusion Inpainting [33] fine-tunes a diffusion model specifically designed for inpainting tasks, taking the mask, masked image, and noisy latent as inputs to UNet architecture. HD-Painter [25] and PowerPaint [56] build upon the foundation of Stable Diffusion Inpainting, separately enhancing the generation quality and enabling the model to perform multiple tasks. + +However, these approaches make it difficult to effectively transfer their inpainting ability to arbitrary pre-trained models, restricting their applicability. To enable any diffusion model with inpainting capabilities, the community fine-tunes ControlNet [51] on inpainting image pairs. However, the model design of ControlNet exhibits limitations in its perceptual understanding of masks and masked images, which consequently leads to unsatisfactory outcomes. Compared with previous methods (shown in Tab. 1), BrushNet is plug-and-play, content-aware, and shape-aware, with a flexible preserving degree for unmasked regions. + +Table 1: Comparison of BrushNet with Previous Image Inpainting Methods. BrushNet offers the advantage of being plug-and-play with any pretrained diffusion model. Moreover, it allows for flexible control over the scale of inpainting and is designed to be aware of both the mask shape and the unmasked content. Note that we only list commonly used text-guided diffusion methods in this table. + +
ModelPlug-and-PlayFlexible-ScaleContent-AwareShape-Aware
Blended Diffusion [1,2]✓
SmartBrush [42]✓
SD Inpainting [33]✓✓
PowerPaint [56]✓✓
HD-Painter [25]✓✓
ReplaceAnything [4]✓✓
Imagen [37]✓✓
ControlNet-Inpainting [51]✓✓✓
BrushNet✓✓✓✓
+ +# 3 Preliminaries and Motivation + +In this section, we will first introduce diffusion models in Sec. 3.1. Then, Sec. 3.2 would review previous inpainting techniques based on sampling strategy modification and special training. Finally, the motivation is outlined in Section 3.3. + +# 3.1 Diffusion Models + +Diffusion models include a forward process that adds Gaussian noise $\epsilon$ to convert clean sample $z_0$ to noise sample $z_T$ , and a backward process that iteratively performs denoising from $z_T$ to $z_0$ , where $\epsilon \sim \mathcal{N}(0,1)$ , and $T$ represents the total number of timesteps. The forward process can be formulated as: + +$$ +z _ {t} = \sqrt {\alpha_ {t}} z _ {0} + \sqrt {1 - \alpha_ {t}} \epsilon \tag {1} +$$ + +$z_{t}$ is the noised feature at step $t$ with $t\sim [1,T]$ , and $\alpha$ is a hyper-parameter. + +In the backward process, given input noise $z_{T}$ sampled from a random Gaussian distribution, learnable network $\epsilon_{\theta}$ estimates noise at each step $t$ conditioned on $C$ . After $T$ progressively refining iterations, $z_{0}$ is derived as the output sample: + +$$ +z _ {t - 1} = \frac {\sqrt {\alpha_ {t - 1}}}{\sqrt {\alpha_ {t}}} z _ {t} + \sqrt {\alpha_ {t - 1}} \left(\sqrt {\frac {1}{\alpha_ {t - 1}} - 1} - \sqrt {\frac {1}{\alpha_ {t}} - 1}\right) \epsilon_ {\theta} \left(z _ {t}, t, C\right) \tag {2} +$$ + +The training of diffusion models revolves around optimizing the denoiser network $\epsilon_{\theta}$ to conduct denoising with condition $C$ , guided by the objective: + +$$ +\underset {\theta} {\min } E _ {z _ {0}, \epsilon \sim \mathcal {N} (0, I), t \sim U (1, T)} \| \epsilon - \epsilon_ {\theta} (z _ {t}, t, C) \| \tag {3} +$$ + +# 3.2 Previous Inpainting Models + +Sampling Strategy Modification. This line of research achieves inpainting by gradually blending masked images with the generated results. The most used method among them is Blended Latent Diffusion (BLD) [1], serving as the default choice for inpainting in widely-used diffusion-based image generation libraries (e.g., Diffusers [28]). Given a binary mask $m$ and a masked image $x_0^{masked}$ , BLD first extracts the latent representation $z_0^{masked}$ of the masked image using VAE. Subsequently, the mask $m$ is resized to $m^{resized}$ to match the size of the latent representation. To formulate the inpainting process, BLD adds Gaussian noise to $z_0^{masked}$ for $T$ steps and gets $z_t^{masked}$ , where $t \sim [1,T]$ . Then, denoising steps start from $z_T^{masked}$ , where each sampling step in eq. 2 is followed by: + +$$ +z _ {t - 1} \leftarrow z _ {t - 1} \cdot \left(1 - m ^ {\text {r e s i z e d}}\right) + z _ {t - 1} ^ {\text {m a s k e d}} \cdot m ^ {\text {r e s i z e d}} \tag {4} +$$ + +Despite its simplicity in implementation, BLD exhibits suboptimal performance in terms of both unmasked region preservation and generation content alignment. This is due to (1) the resize of the mask preventing it from correctly blending the noisy latent, (2) the diffusion model lacking perceptual knowledge of mask boundaries and the unmasked image region context. + +Dedicated Inpainting Models. To enhance the performance of inpainting, previous works fine-tune the base models by expanding the input UNet channel to include the mask and masked image inputs, turning it into an architecture specifically designed for image inpainting. Though having better generation results compared with BLD, they still have several drawbacks: (1) These models merge the noisy latent, masked image latent, and mask at the initial convolution layer of the UNet architecture, where they are collectively influenced by the text embedding. Consequently, subsequent layers in the UNet model struggle to obtain pure masked image features due to the text's influence. (2) Incorporating both the condition processing and generation within a single branch places additional burdens on the UNet framework. (3) These approaches require extensive fine-tuning across various variations of diffusion backbones, which is computationally intensive and lacks transferability to custom diffusion models. + +# 3.3 Motivation + +Based on the analysis presented in Section 3.2, a more effective architecture design of inpainting would be introducing an additional branch specifically dedicated to masked image processing. ControlNet [51] is one of the widely adopted + +strategies that exemplifies this idea. However, it should be noted that directly fine-tuning ControlNet, which is originally designed for controllable image generation, on the inpainting task yields unsatisfactory results. ControlNet designs a lightweight encoder to incorporate out-of-domain structural conditions (e.g., skeleton) and relies on text guidance for content generation, which is unsuitable for pixel-level inpainting image feature injection. Furthermore, ControlNet typically relies on sparse control, meaning that merely adding control to residuals in the UNet framework would be sufficient, while inpainting requires pixel-to-pixel constraints with strong restrictive information. Thus a new architecture specifically designed for inpainting is urgently needed. + +# 4 Method + +An overview of BrushNet is shown in Fig. 3. We employ a dual-branch strategy for masked image guidance insertion (Sec. 4.1). Blending operations with a blurred mask is used to ensure better-unmasked region preservation (Sec. 4.2). Notably, BrushNet can achieve flexible control by adjusting the added scale. + +![](images/adc686765481ed163c4ad9e77e71ec36827bf43a5ff44056a215663840e6bdec.jpg) +Fig. 3: Model overview. Our model outputs an inpainted image given the mask and masked image input. Firstly, we downsample the mask to accommodate the size of the latent, and input the masked image to the VAE encoder to align the distribution of latent space. Then, noisy latent, masked image latent, and downsampled mask are concatenated as the input of BrushNet. The feature extracted from BrushNet is added to pretrained UNet layer by layer after a zero convolution block [51]. After denoising, the generated image and masked image are blended with a blurred mask. + +# 4.1 Masked Image Guidance + +The insertion of the masked image feature into the pre-trained diffusion network is accomplished through an additional branch, which explicitly separates the feature extraction of masked images from the image-generating processes. The input to the additional branch includes noisy latent, masked image latent, and downsampled mask, which are concatenated together to form the input. + +Specifically, the noisy latent provides generation information during the present generation process, helping BrushNet enhance the semantic coherence of the masked image feature. The masked image latent is extracted from the masked image using VAE, which is aligned with the data distribution of the pre-trained UNet. To ensure the alignment of the mask size with the noisy latent and masked image latent, we employ cubic interpolation to downsample the mask. + +To process the masked image features, BrushNet utilizes a clone of the pre-trained diffusion model while excluding its cross-attention layers. The pretrained weights of the diffusion model serve as a strong prior for extracting the masked image features, while the removal of the cross-attention layers ensures that only pure image information is considered within this additional branch. BrushNet feature is inserted into the frozen diffusion model layer-by-layer, enabling dense per-pixel control hierarchically. Similar to ControlNet [51], we employ zero convolution layers to establish a connection between the locked model and the trainable BrushNet. This ensures that harmful noise does not influence the hidden states in the trainable copy during the initial stages of training. + +The feature insertion operation is shown in Eq. 5. Specifically, $\epsilon_{\theta}(z_t,t,C)_i$ indicates the feature of the $i$ -th layer in network $\epsilon_{\theta}$ with $i\sim [1,n]$ , where $n$ is the number of layers. The same notation applies to $\epsilon_{\theta}^{BrushNet}$ . $\epsilon_{\theta}^{BrushNet}$ takes the concatenated noisy latent $z_{t}$ , masked image latent $z_0^{masked}$ , and downsampled mask $m^{resized}$ as input, with the concatenation operation denoted as $[\cdot]$ . $\mathcal{Z}$ is the zero convolution operation. $w$ is the preservation scale used to adjust the influence of BrushNet on pretrained diffusion model. + +$$ +\epsilon_ {\theta} \left(z _ {t}, t, C\right) _ {i} = \epsilon_ {\theta} \left(z _ {t}, t, C\right) _ {i} + w \cdot \mathcal {Z} \left(\epsilon_ {\theta} ^ {\text {B r u s h N e t}} \left(\left[ z _ {t}, z _ {0} ^ {\text {m a s k e d}}, m ^ {\text {r e s i z e d}} \right], t\right) _ {i}\right) \tag {5} +$$ + +# 4.2 Blending Operation + +As mentioned in Section 4.2, the blending operation conducted in latent space can result in inaccuracies due to the resizing of the mask. Similarly, in our approach, a similar issue arises as we resize the mask to match the size of the latent space, which can introduce potential inaccuracies. Additionally, it is important to acknowledge that VAE encoding and decoding operations have inherent limitations and may not ensure complete image reconstruction. + +To ensure a fully consistent image reconstruction of the unmasked region, previous works have explored different techniques. Some approaches [4, 56], utilize past-and-copy methods, where the unmasked region is directly copied from the original image. However, this can result in a lack of semantic coherence in the final generation results. On the other hand, adopting latent blending operations inspired by BLD [1, 33] has been observed to face challenges in effectively preserving the desired information in the unmasked regions. + +In this work, we present a simple pixel space solution to address this issue by first blurring the mask and then performing copy-and-paste using the blurred mask. Although this approach may result in a slight loss of accuracy in preserving the details of the mask boundary, the error is nearly imperceptible to the naked eye and results in significantly improved coherence in the mask boundary. + +# 4.3 Flexible Control + +The architecture design of BrushNet inherently makes it suitable for seamless plug-and-play integration to various pretrained diffusion models and enables flexible preservation scale. Specifically, the flexible control of our proposed BrushNet includes: (1) Since BrushNet does not modify the weights of the pretrained diffusion model, it can be readily integrated as a plug-and-play component with any community fine-tuned diffusion models. This allows for easy adoption and experimentation with different pretrained models. (2) Preservation Scale Adjustment: The preservation scale of the unmasked region can be controlled by incorporating BrushNet features into the frozen diffusion model with the weight $w$ . This weight determines the influence of BrushNet on the preservation scale, offering the ability to adjust the desired level of preservation. (3) Blurring Scale and Blending Operation: By adjusting the scale of blurring and deciding whether to apply the blending operation, the preservation scale of the unmasked region can be further customized. These features allow for flexible fine-grained control over the inpainting process. More explanation can be found in Sec. 5.5. + +# 5 Experiments + +# 5.1 Evaluation Benchmark and Metrics + +Benchmark. Previous commonly used datasets in the image inpainting field include CelebA [22], CelebA-HQ [14], ImageNet [8], MSCOCO [19], Open Images [17], and LSUN-Bedroom [48]. However, these datasets either primarily focus on a small area, such as human faces, or predominantly consist of low-quality, cluttered real-life scene data. As a result, these datasets are not well-suited for training and evaluating diffusion-based inpainting models, which can generate high-quality diverse images that align with text prompts. + +Recently proposed EditBench [37] serves as a benchmark specifically designed for text-guided image inpainting for diffusion models. This benchmark consists of a collection of 240 images comprising an equal ratio of natural images and generated images, with mask and caption annotation for each image. However, the annotated masks in EditBench are mostly random shapes without specific object information, neglecting the practical application of inpainting in real scenarios such as replacing an object with an external mask, as commonly seen in E-commerce product displays and image editing. + +To fill the gap, we propose BrushBench for segmentation-based inpainting, as shown in Fig. 4. BrushBench comprises a total of 600 images, with each image accompanied by the human-annotated mask and caption annotation. The images in BrushBench are evenly distributed between natural images and artificial images, such as paintings. Furthermore, the dataset ensures an equal distribution among different categories, including humans, animals, indoor scenarios, and outdoor scenarios. This balanced distribution enables a fair evaluation across various categories, promoting better evaluation equity. + +To further enhance the analysis of the inpainting task, we categorize it into two distinct types based on the masks used: random brush masks and segmentation-based masks. We use EditBench as the comparison benchmark + +![](images/9e81c8cb4347892949d92e8b70253f27259f3ee930f7f7ddf10488f1d0df5858.jpg) +Fig. 4: Benchmark overview. I and II separately show natural and artificial images, masks, and caption of BrushBench. (a) to (d) show images of humans, animals, indoor scenarios, and outdoor scenarios. Each group of images shows the original image, inside-inpainting mask, and outside-inpainting mask, with an image caption on the top. III show image, mask, and caption from EditBench [37], with (e) for generated images and (f) for natural images. The images are randomly selected from both benchmarks. + +for random brush masks and use BrushBench for segmentation-based masks. Regarding inpainting with segmentation-based masks, we refine the task by considering two specific scenarios: segmentation mask inside-inpainting and segmentation mask outside-inpainting. By separating these two subtasks, we can better understand the inpainting performance in different image regions. + +Dataset. To train segmentation-based mask inpainting, we annotate segmentation mask on Laion-Aesthetic [34] dataset, called BrushData. We employ the Grounded-SAM [32] to annotate open-world masks and then filter the masks based on their confidence score, ensuring that only masks with relatively higher confidence scores are retained. Additionally, we consider factors such as reasonable mask size and good mask continuity during the filtering process. + +Metrics. We consider 7 metrics from three aspects: image generation quality, masked region preservation, and text alignment. + +- Image Generation Quality. Metrics most used by previous inpainting methods (e.g., FID [11] and KID [5]) show a poor representation of the rich and varied content generated by modern text-to-image models [16]. Thus, we use Image Reward (IR) [44], HPS v2 (HPS) [41], and Aesthetic Score (AS) [34] that align with human perception. Specifically, ImageReward and HPS v2 are text-to-image human preference evaluation models trained on large-scale datasets of human preference choices on generated images. Aesthetic Score is a linear model trained on image quality rating pairs of real images. +- Masked Region Preservation. We follow previous works using standard Peak Signal-to-Noise Ratio (PSNR) [39], Learned Perceptual Image Patch Similarity (LPIPS) [52], and Mean Squared Error (MSE) [38] in the unmasked region among the generated image and the original image. + +- Text Alignment. We use CLIP Similarity (CLIP Sim) [40] to evaluate text-image consistency between the generated images and corresponding text prompts. CLIP Similarity projects text and images to the same shared space with the CLIP model [30] and evaluates the similarity of their embeddings. + +# 5.2 Implementation Details + +We perform the inference of different inpainting methods in the same setting unless specifically clarified, i.e., on NVIDIA Tesla V100 following their open-source code with a base model of Stabe Diffusion v1.5 in 50 steps, with a guidance scale of 7.5. We keep the recommended hyper-parameter for each inpainting method in all images for fair comparison. BrushNet and all ablation models are trained for 430 thousands steps on 8 NVIDIA Tesla V100 GPUs, which takes around 3 days. For comparison on BrushBench, we use BrushNet trained on BrushData. For comparison on EditBench, we use the model trained on LAION-5B with random masks. Details can be found in the provided codes. + +# 5.3 Quantitative Comparison + +Table 2: Quantitative comparisons among BrushNet and other diffusion-based inpainting models in BrushBench: Blended Latent Diffusion (BLD) [1], Stable Diffusion Inpainting (SDI) [33], HD-Painter (HDP) [25], PowerPaint (PP) [56], and ControlNet-Inpainting (CNI) [51]. Metrics encompassing image quality, masked region preservation, and text alignment (Text Align) for inside-inpainting and outside-inpainting are shown in the table. All models use Stable Diffusion V1.5 as base model. Red stands for the best result, Blue stands for the second best result. + +
MetricsImage QualityMasked Region PreservationText Align
ModelsIR×10↑HPS×102↑AS↑PSNR↑MSE×103↓LPIPS×103↓CLIP Sim↑
InsideBLD [1]9.7825.876.1721.339.7649.2626.15
SDI [33]11.7227.066.5021.5213.8748.3926.17
HDP [25]11.6826.906.4222.619.9543.5026.37
PP [56]11.4627.356.2421.4332.7348.4326.48
CNI [51]9.926.026.5312.3978.78243.6226.47
CNI* [51]11.2126.926.3922.7324.5843.4926.22
Ours12.3627.406.5321.659.3148.2826.48
Ours*12.6427.786.5131.940.8018.6726.39
OutsideBLD [1]7.8126.776.2315.8535.8621.4026.73
SDI [33]10.2727.996.5518.0419.8715.1327.21
HDP [25]9.6627.796.4618.0322.9915.2226.96
PP [56]7.4528.016.2618.0431.7815.1326.72
CNI [51]9.2627.686.4211.9183.0358.1627.29
CNI* [51]9.5727.766.2817.5037.7219.9526.92
Ours10.8228.026.6418.0622.8615.0827.33
Ours*10.8828.096.6427.822.254.6327.22
+ +* with blending operation +Tab. 2 and Tab. 3 show the quantitative comparison on BrushBench and EditBench [37]. We compare the inpainting results of sampling strategy modification method Blended Latent Diffusion [1], dedicated inpainting models Stable Diffusion Inpainting [33], HD-Painter [25], and PowerPaint [56], as well as plug-and-play method ControlNet [51] trained on inpainting data. + +Results demonstrate BrushNet's effectiveness across image quality, masked region preservation, and image-text alignment. Blended Latent Diffusion [1] shows the poorest results in image quality and text alignment, derived from the incoherence between the generated masked and the unmasked given image. At the same time, its performance in masked region preservation is also not satisfactory because of the loss incurred by resized mask blending operation in the latent space. Modified from Stable Diffusion Inpainting [33], HD-Painter [25] and PowerPaint [56] demonstrate comparable performance to Stable Diffusion Inpainting in the task of inside-inpainting. However, when it comes to outside-inpainting, their results in terms of image quality and text alignment are significantly poorer compared to Stable Diffusion Inpainting, which can be attributed to their exclusive emphasis on the inside-inpainting task. + +ControlNet [51] trained on inpainting has the most similar experimental configuration to ours. Due to the mismatch between its model design and the inpainting task, ControlNet shows poor results in masked region preservation and image quality, necessitating its combination with Blended Latent Diffusion [1] to generate satisfying inpainted images. However, even with this combination, it still falls short compared to dedicated inpainting models and BrushNet. + +Table 3: Quantitative comparisons among BrushNet and other diffusion-based inpainting models in EditBench. A detailed explanation of compared methods and metrics can be found in the caption of Tab. 2. Red stands for the best result, Blue stands for the second best result. + +
MetricsImage QualityMasked Region PreservationText Align
ModelsIR×10↑HPS×102↑AS↑PSNR↑MSE×103↓LPIPS×103↓CLIP Sim↑
BLD [1]0.9023.815.4420.8910.9331.9028.62
SDI [33]1.8624.245.6923.256.9424.3028.00
HDP [25]1.7424.205.6423.076.7024.3228.34
PP [56]1.2424.505.4423.3420.1224.1227.80
CNI [51]1.4924.465.8212.7169.42159.7128.16
CNI* [51]0.9023.795.4622.6135.9326.1427.74
Ours4.4025.105.8423.356.8124.1128.67
Ours*4.4625.245.8233.660.6310.1228.87
+ +* with blending operation + +The performance on the EditBench is roughly consistent with the overall performance on BrushBench, which similarly shows BrushNet's superior performance. This indicates that our method exhibits strong performance across a range of inpainting tasks with various mask types, including random masks, inside-inpainting masks, and outside-inpainting masks. + +# 5.4 Qualitative Comparison + +The qualitative comparison with previous image inpainting methods is shown in Fig. 1. We provide results on artificial images and natural images across various inpainting tasks, including random mask inpainting, segmentation mask inside-inpainting, and segmentation mask outside-inpainting. BrushNet consistently + +show exceptional results in the coherent of generated region and unmasked region, considering content (I, II right, III right, IV), color (II left), and text (III left). Interestingly, Fig. 1 III left requires the model to generate a cat and a goldfish. All previous methods fail to recognize that a goldfish is already present in the masked image, resulting in the generation of an additional fish within the masked region. BrushNet successfully realized the awareness of background information due to the design of the dual-branch decoupling. + +![](images/51b2813a57d023f97be77d4b34e3c2fb7e895b8b4b20ce7618e651880ca3b5d3.jpg) +Fig. 5: Comparison of previous inpainting methods and BrushNet on various image domain. A detailed explanation of compared methods is in Fig. 1. + +The untouched pre-trained diffusion branch also provides the advantage of better coverage across different data domains, such as painting and anime. As shown in Fig. 5, BrushNet demonstrates superior performance across various image categories, including natural image (I, II), pencil painting (III), anime (IV), illustration (V), digital art (VI left), and watercolor (VI right). Due to the page limit, more qualitative comparison results are in supplementary files. + +# 5.5 Flexible Control Ability + +Fig. 6 and Fig. 7 illustrate the flexible control provided by BrushNet in two aspects: base diffusion model selection and controlling scale. In Fig. 6, we showcase the ability to combine BrushNet with different diffusion models fine-tuned by the community. This allows users to select a specific model that best suits their inpainting requirements, enabling users to achieve the desired inpainting results based on their specific needs. Fig. 7 demonstrates the adjustment of the control scale of BrushNet. This control scale parameter allows users to effectively control the extent of unmasked region protection during the inpainting process. By manipulating the scale parameter, users can achieve fine-grained control over the inpainting process, enabling precise and customizable inpainting. + +![](images/76228267e1ed329a83533c4805e99107400c5df7e9b93689366e220d820d07b8.jpg) +A single red rose is laying on top of snow + +![](images/ac187c77211becdb3dcd38a5f8135018bf4e8efdadf8d999fb857d70df92928c.jpg) + +![](images/bf93efd098fa90b639fdba74566759e61b1f29078ad7f469d2bd2291bc6f9c45.jpg) +A cartoon boy driving a car + +![](images/b48bd04a24f3d4e7b0df40bdbd04ed1e197e6bf9206905c72ef58092cdbd6cd7.jpg) + +![](images/0621ab133f65bdddaf98800b2cb4eccd7292e587b6d8f978574208d634121a40.jpg) + +![](images/bd949013803c54c98582c8e6f25572344811acf7a0e54fefcfbc4d1ddb753ac2.jpg) +A little girl sitting on the grass with her dog +(a) Mask +(b) DS +(c) ER +(d) HR (e) +Fig. 6: Integrating BrushNet to community fine-tuned diffusion models. We use five popular community diffusion models fine-tuned from stable diffusion v1.5: DreamShaper (DS) [24], epiCRealism (ER) [9], Henmix_Real (HR) [10], MeinaMix (MM) [26], and Realistic Vision (RV) [35]. MM is specifically designed for anime images. + +![](images/31ad5eab82caf1d088f4f5ce436feac6fc4d5e4362776b5167b3bdeebadca15c.jpg) +)MM +(f) RV + +![](images/3e63eabc35e89d86671afca7305a914e87ef044fa33d92f63aef3afd5717e5ec.jpg) +A woman with her hands on her face +(a) Mask +(b) DS +(c) ER +(d) HR +(e) MM + +![](images/87142fa1b19038c9e7a5ff017ae480f0b28790eedb6a717d48f1c59df5fb07bb.jpg) +(f) RV + +![](images/b911b418c06f7d9da12f655394c7a885f3843045c2947495d0207f2ecb5cf4be.jpg) +Book, coffee and purple flower +A cupcake with a lit candle on top + +![](images/abe2cfc0f166cf0321d04f9e8cd038fac82001cdecf00afe15a7b20a16e910a4.jpg) +(a) Mask +(b) 1.0 +(c) 0.8 +(d) 0.6 +(e) 0 +(1)0.4 +(g) 0.3 +Fig. 7: Flexible control scale of BrushNet. (a) shows the given masked image, (b)-(h) show adding BrushNet with control scale $w$ from 1.0 to 0.2. Results show a gradually diminishing controllable ability from precise to rough control. + +![](images/66c7432ed6d7764ba3dd97737f94396a9c3f103455cc16eeb8b2fd5530993580.jpg) +A small island with a tree and clouds + +![](images/599b7b1af0165525307e84eac87130cf4c592dc8bf4149a778b39fa2275e9dff.jpg) +A white rabbit sitting on the ground in the grass +(a) Mask (b) 1.0 +(c) 0.8 +(d) 0.6 +(e) 0.5 +(f) 0.4 +(g) 0.3 + +# 5.6 Ablation Study + +Table 4: Ablation on dual-branch design. Stable Diffusion Inpainting (SDI) use single-branch design, where the entire UNet is fine-tuned. We conducted an ablation analysis by training a dual-branch model with two variations: one with the base UNet fine-tuned, and another with the base UNet forzened. Results demonstrate the superior performance achieved by adopting the dual-branch design. Red is the best result. + +
MetricsImage QualityMasked Region PreservationText Align
ModelIR×10↑HPS×102↑AS↑PSNR↑MSE×102↓LPIPS×103↓CLIP Sim↑
SDI11.0027.536.5319.7816.8731.7626.69
w/o fine-tune11.5927.716.5919.8616.0931.6826.91
w/ fine-tune11.6327.736.6020.1315.8431.5726.93
+ +We conducted ablation studies to investigate the impact of different model designs. Tab. 4 compares the dual-branch and single-branch designs. Tab. 5 shows the ablation study focusing on the additional branch architecture. The ablation studies are conducted on BrushBench, averaging the performance of inside-inpainting and outside-inpainting. The results presented in Tab. 4 demonstrate that the dual-branch design significantly outperforms the single-branch design. Additionally, fine-tuning the base diffusion model in dual-branch design yields better results compared to freezing it. However, fine-tuning the base diffusion model may restrict flexibility and control over the model. Considering this trade-off between performance and flexibility, we decide to adopt the frozen + +dual-branch design as our model design. Tab. 5 presents the rationale behind the design choices for (1) using a VAE encoder instead of randomly initialized convolution layers to process the masked image, (2) incorporating the full UNet feature layer-by-layer into the pre-trained UNet, (3) removing text cross-attention in BrushNet, which avoids masked image features influenced by text. + +Table 5: Ablation on model architecture. We ablate on the following components: the image encoder (Enc), selected from a random initialized convolution (Conv) and a VAE; the inclusion of mask in input (Mask), chosen from adding (w/) and not adding (w/o); the presence of cross-attention layers (Attn), chosen from adding (w/) and not adding (w/o); the type of UNet feature addition (UNet), selected from adding the full UNet feature (full), adding half of the UNet feature (half), and adding the feature like ControlNet (CN); and finally, the blending operation (Blend), chosen from not adding (w/o), direct pasting (paste), and blurred blending (blur). Red is the best result. + +
MetricsImage QualityMasked Region PreservationText Align
EncMaskAttnUNetBlendIR×10↑HPS×102↑AS↑PSNR↑MSE×102↓LPIPS×103↓CLIP Sim↑
Convw/w/ofullw/o11.0526.236.5514.8937.2364.5426.76
VAEw/ow/ofullw/o11.5527.706.5717.9626.3849.3326.87
VAEw/w/fullw/o11.2527.626.5618.6919.4434.2826.63
Convw/w/CNw/o9.5826.856.4712.1580.91150.8926.88
VAEw/w/CNw/o10.5327.426.5918.2824.3641.6326.89
VAEw/w/oCNw/o11.4227.696.5818.4924.0936.3326.86
VAEw/w/ohalfw/o11.4727.706.5719.0123.7733.5726.87
VAEw/w/ofullw/o11.5927.716.5919.8616.0931.6826.91
VAEw/w/ofullpaste11.7227.936.58---26.80
VAEw/w/ofullblur11.7627.946.5829.881.5311.6526.81
+ +# 6 Discussion + +Conclusion. This paper proposes a plug-and-play image inpainting method BrushNet with a pixel-level masked image feature insertion architectural design. Quantitative and qualitative results on our proposed benchmark, BrushBench, and EditBench show the superior performance of BrushNet considering image generation quality, masked region preservation, and image-text alignment. + +Limitations and Future Work. However, BrushNet still has some limitations: (1) The quality and content generated by our model are heavily dependent on the chosen base model. (2) Even with BrushNet, we still observe poor generation results in cases where the given mask has an unusually shaped or irregular form, or when the given text does not align well with the masked image. In our future work, we will continue to address these challenges. + +Negative Social Impact. Image inpainting models present exciting opportunities for content creation, but they also carry potential risks to individuals and society. Their reliance on internet-collected training data can amplify social biases, and there is a specific risk of generating persuasive misinformation by manipulating human images with offensive elements. To address these concerns, it is crucial to emphasize responsible use and establish ethical guidelines when utilizing these models. This is also a key focus for our future model releases. + +# *Acknowledgment. + +This work is supported in part by Research Matching Grant(CSE-7-2022)-RMG01. + +# References + +1. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. ACM transactions on graphics (TOG) 42(4), 1-11 (2023) +2. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18208-18218 (2022) +3. Bertalmio, M., Sapiro, G., Caseles, V., Ballester, C.: Image inpainting. In: International Conference and Exhibition on Computer Graphics and Interactive Techniques (SIGGRAPH). pp. 417-424 (2000) +4. Binghui, C., Chao, L., Chongyang, Z., Wangmeng, X., Yifeng, G., Xuansong, X.: Replaceanything as you want: Ultra-high quality content replacement (2023), https://aigcdesigngroup.github.io/replace-anything/ +5. Binkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying MMD GANs. arXiv preprint arXiv:1801.01401 (2018) +6. Corneanu, C., Gadde, R., Martinez, A.M.: Latentpaint: Image inpainting in latent space with diffusion models. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 4334-4343 (2024) +7. Criminisi, A., Pérez, P., Toyama, K.: Region filling and object removal by exemplar-based image inpainting. IEEE Transactions on Image Processing 13(9), 1200-1212 (2004) +8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 248-255. IEEE (2009) +9. epinikion: epicrealism (2023), https://civitai.com/models/25694?modelVersionId=143906 +0. hemi29833: Henmixreal (2024), https://civitai.com/models/20282?modelVersionId=305687 +1. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs trained by a two time-scale update rule converge to a local Nash equilibrium. Advances in Neural Information Processing Systems (NIPS) 30 (2017) +2. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NIPS) 33, 6840-6851 (2020) +3. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) +4. Huang, H., He, R., Sun, Z., Tan, T., et al.: Introvae: Introspective variational autoencoders for photographic image synthesis. Advances in Neural Information Processing Systems (NIPS) 31 (2018) +5. Huang, Y., Huang, J., Liu, Y., Yan, M., Lv, J., Liu, J., Xiong, W., Zhang, H., Chen, S., Cao, L.: Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525 (2024) +6. Jayasumana, S., Ramalingam, S., Veit, A., Glasner, D., Chakrabarti, A., Kumar, S.: Rethinking fid: Towards a better evaluation metric for image generation. arXiv preprint arXiv:2401.09603 (2023) +7. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision (IJCV) 128(7), 1956-1981 (2020) +8. Li, Z., Wei, P., Yin, X., Ma, Z., Kot, A.C.: Virtual try-on with pose-garment keypoints guided inpainting. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 22788-22797 (2023) + +19. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision (ECCV). pp. 740-755. Springer (2014) +20. Liu, A., Niepert, M., Broeck, G.V.d.: Image inpainting via tractable steering of diffusion models. arXiv preprint arXiv:2401.03349 (2023) +21. Liu, H., Wan, Z., Huang, W., Song, Y., Han, X., Liao, J.: Pd-GAN: Probabilistic diverse GAN for image inpainting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9371-9381 (2021) +22. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: IEEE/CVF International Conference on Computer Vision (ICCV) (December 2015) +23. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11461-11471 (2022) +24. Lykon: Dreamshaper (2022), https://civitai.com/models/4384?modelVersionId=128713 +25. Manukyan, H., Sargsyan, A., Atanyan, B., Wang, Z., Navasardyan, S., Shi, H.: Hd-painter: High-resolution and prompt-faithful text-guided image inpainting with diffusion models. arXiv preprint arXiv:2312.14091 (2023) +26. Meina: Meinamix (2023), https://civitai.com/models/7240?modelVersionId=119057 +27. Peng, J., Liu, D., Xu, S., Li, H.: Generating diverse structure for image inpainting with hierarchical vq-vae. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10775-10784 (2021) +28. von Platen, P., Patil, S., Lozhkov, A., Cuenca, P., Lambert, N., Rasul, K., Davaadorj, M., Wolf, T.: Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers (2022) +29. Quan, W., Chen, J., Liu, Y., Yan, D.M., Wonka, P.: Deep learning-based image and video inpainting: A survey. International Journal of Computer Vision (IJCV) pp. 1-34 (2024) +30. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (ICML). pp. 8748-8763. PMLR (2021) +31. Razzhigaev, A., Shakhmatov, A., Maltseva, A., Arkhipkin, V., Pavlov, I., Ryabov, I., Kuts, A., Panchenko, A., Kuznetsov, A., Dimitrov, D.: Kandinsky: an improved text-to-image synthesis with image prior and latent diffusion. arXiv preprint arXiv:2310.03502 (2023) +32. Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., et al.: Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159 (2024) +33. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022) +34. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems (NIPS) 35, 25278–25294 (2022) + +35. SG161222: Realisticvision (2023), https://civitai.com/models/4201?modelVersionId=130072 +36. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) +37. Wang, S., Sahara, C., Montgomery, C., Pont-Tuset, J., Noy, S., Pellegrini, S., Onoe, Y., Laszlo, S., Fleet, D.J., Soricut, R., et al.: Imagen editor and editbench: Advancing and evaluating text-guided image inpainting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18359-18369 (2023) +38. Wikipedia contributors: Mean squared error — Wikipedia, the free encyclopedia (2024), https://en.wikipedia.org/w/index.php?title=Mean_squared_error&oldid=1207422018, [Online; accessed 4-March-2024] +39. Wikipedia contributors: Peak signal-to-noise ratio — Wikipedia, the free encyclopedia (2024), https://en.wikipedia.org/w/index.php?title=Peak_signal-to-noise_ratio&oldid=1210897995, [Online; accessed 4-March-2024] +40. Wu, C., Huang, L., Zhang, Q., Li, B., Ji, L., Yang, F., Sapiro, G., Duan, N.: GO-DIVA: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806 (2021) +41. Wu, X., Hao, Y., Sun, K., Chen, Y., Zhu, F., Zhao, R., Li, H.: Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341 (2023) +42. Xie, S., Zhang, Z., Lin, Z., Hinz, T., Zhang, K.: Smartbrush: Text and shape guided object inpainting with diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22428-22437 (2023) +43. Xie, S., Zhao, Y., Xiao, Z., Chan, K.C., Li, Y., Xu, Y., Zhang, K., Hou, T.: Dreaminpainter: Text-guided subject-driven image inpainting with diffusion models. arXiv preprint arXiv:2312.03771 (2023) +44. Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagereward: Learning and evaluating human preferences for text-to-image generation (2023) +45. Xu, Z., Zhang, X., Chen, W., Yao, M., Liu, J., Xu, T., Wang, Z.: A review of image inpainting methods based on deep learning. Applied Sciences 13(20), 11189 (2023) +46. Yang, S., Chen, X., Liao, J.: Uni-paint: A unified framework for multimodal image inpainting with pretrained diffusion model. In: ACM International Conference on Multimedia (MM). pp. 3190-3199 (2023) +47. Yang, S., Zhang, L., Ma, L., Liu, Y., Fu, J., He, Y.: Magicremover: Tuning-free text-guided image inpainting with diffusion models. arXiv preprint arXiv:2310.02848 (2023) +48. Yu, F., Seff, A., Zhang, Y., Song, S., Funkhouser, T., Xiao, J.: Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365 (2015) +49. Yu, T., Feng, R., Feng, R., Liu, J., Jin, X., Zeng, W., Chen, Z.: Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790 (2023) +50. Zhang, G., Ji, J., Zhang, Y., Yu, M., Jaakkola, T., Chang, S.: Towards coherent image inpainting using denoising diffusion implicit models (2023) +51. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2023) + +52. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 586-595 (2018) +53. Zhao, S., Cui, J., Sheng, Y., Dong, Y., Liang, X., Chang, E.I., Xu, Y.: Large scale image completion via co-modulated generative adversarial networks. arXiv preprint arXiv:2103.10428 (2021) +54. Zheng, C., Cham, T.J., Cai, J.: Pluralistic image completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1438-1447 (2019) +55. Zheng, H., Lin, Z., Lu, J., Cohen, S., Shechtman, E., Barnes, C., Zhang, J., Xu, N., Amirghodsi, S., Luo, J.: Image inpainting with cascaded modulation GAN and object-aware training. In: European Conference on Computer Vision (ECCV). pp. 277-296. Springer (2022) +56. Zhuang, J., Zeng, Y., Liu, W., Yuan, C., Chen, K.: A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. arXiv preprint arXiv:2312.03594 (2023) \ No newline at end of file diff --git a/2024/BrushNet_ A Plug-and-Play Image Inpainting Model with Decomposed Dual-Branch Diffusion/images.zip b/2024/BrushNet_ A Plug-and-Play Image Inpainting Model with Decomposed Dual-Branch Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fc9df8a6bbb208a56049c65b28a2e4cad2fc0599 --- /dev/null +++ b/2024/BrushNet_ A Plug-and-Play Image Inpainting Model with Decomposed Dual-Branch Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:049b5527d47f082c5335873ab7488b2232523122b3effed120adf6d5fd9ec998 +size 824263 diff --git a/2024/BrushNet_ A Plug-and-Play Image Inpainting Model with Decomposed Dual-Branch Diffusion/layout.json b/2024/BrushNet_ A Plug-and-Play Image Inpainting Model with Decomposed Dual-Branch Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1b26893e984218117defbc47d4f82b2d73ce4154 --- /dev/null +++ b/2024/BrushNet_ A Plug-and-Play Image Inpainting Model with Decomposed Dual-Branch Diffusion/layout.json @@ -0,0 +1,9893 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 112, + 473, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 112, + 473, + 148 + ], + "spans": [ + { + "bbox": [ + 136, + 112, + 473, + 148 + ], + "type": "text", + "content": "BrushNet: A Plug-and-Play Image Inpainting Model with Decomposed Dual-Branch Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 157, + 480, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 157, + 480, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 157, + 480, + 170 + ], + "type": "text", + "content": "Xuan " + }, + { + "bbox": [ + 132, + 157, + 480, + 170 + ], + "type": "inline_equation", + "content": "\\mathrm{Ju}^{1,2}" + }, + { + "bbox": [ + 132, + 157, + 480, + 170 + ], + "type": "text", + "content": ", Xian Liu" + }, + { + "bbox": [ + 132, + 157, + 480, + 170 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 132, + 157, + 480, + 170 + ], + "type": "text", + "content": ", Xintao Wang" + }, + { + "bbox": [ + 132, + 157, + 480, + 170 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 132, + 157, + 480, + 170 + ], + "type": "text", + "content": ", Yuxuan Bian" + }, + { + "bbox": [ + 132, + 157, + 480, + 170 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 157, + 480, + 170 + ], + "type": "text", + "content": ", Ying Shan" + }, + { + "bbox": [ + 132, + 157, + 480, + 170 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 157, + 480, + 170 + ], + "type": "text", + "content": ", and Qiang " + }, + { + "bbox": [ + 132, + 157, + 480, + 170 + ], + "type": "inline_equation", + "content": "\\mathrm{Xu}^{2\\star}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 171, + 172, + 443, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 172, + 443, + 184 + ], + "spans": [ + { + "bbox": [ + 171, + 172, + 443, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 171, + 172, + 443, + 184 + ], + "type": "text", + "content": "ARC Lab, Tencent PCG " + }, + { + "bbox": [ + 171, + 172, + 443, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 171, + 172, + 443, + 184 + ], + "type": "text", + "content": "The Chinese University of Hong Kong" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 216, + 186, + 397, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 186, + 397, + 196 + ], + "spans": [ + { + "bbox": [ + 216, + 186, + 397, + 196 + ], + "type": "text", + "content": "https://github.com/TencentARC/BrushNet" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 140, + 201, + 468, + 347 + ], + "blocks": [ + { + "bbox": [ + 140, + 201, + 468, + 347 + ], + "lines": [ + { + "bbox": [ + 140, + 201, + 468, + 347 + ], + "spans": [ + { + "bbox": [ + 140, + 201, + 468, + 347 + ], + "type": "image", + "image_path": "9a73c4b7f103f0cd858bc52bb37b16167d1524cc4cd0eaf471b5413c80b009c8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 349, + 482, + 428 + ], + "lines": [ + { + "bbox": [ + 130, + 349, + 482, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 482, + 428 + ], + "type": "text", + "content": "Fig. 1: Performance comparisons of BrushNet and previous image inpainting methods across various inpainting tasks: (I) Random Mask (< 50% masked), (II) Random Mask (> 50% masked), (III) Segmentation Mask Inside-Inpainting, (IV) Segmentation Mask Outside-Inpainting. Each group of results contains an artificial image (left) and a natural image (right) with 6 inpainting methods: (b) Blended Latent Diffusion (BLD) [1], (c) Stable Diffusion Inpainting (SDI) [33], (d) HD-Painter (HDP) [25], (e) PowerPaint (PP) [56], (f) ControlNet-Inpainting (CNI) [51], and (g) Ours." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 434, + 455, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 434, + 455, + 632 + ], + "spans": [ + { + "bbox": [ + 159, + 434, + 455, + 632 + ], + "type": "text", + "content": "Abstract. Image inpainting, the process of restoring corrupted images, has seen significant advancements with the advent of diffusion models (DMs). Despite these advancements, current DM adaptations for inpainting, which involve modifications to the sampling strategy or the development of inpainting-specific DMs, frequently suffer from semantic inconsistencies and reduced image quality. Addressing these challenges, our work introduces a novel paradigm: the division of masked image features and noisy latent into separate branches. This division dramatically diminishes the model's learning load, facilitating a nuanced incorporation of essential masked image information in a hierarchical fashion. Herein, we present BrushNet, a novel plug-and-play dual-branch model engineered to embed pixel-level masked image features into any pre-trained DM, guaranteeing coherent and enhanced image inpainting outcomes. Additionally, we introduce BrushData and BrushBench to facilitate segmentation-based inpainting training and performance assessment. Our extensive experimental analysis demonstrates BrushNet's superior performance over existing models across seven key metrics, including image quality, mask region preservation, and textual coherence." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 638, + 443, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 638, + 443, + 650 + ], + "spans": [ + { + "bbox": [ + 160, + 638, + 443, + 650 + ], + "type": "text", + "content": "Keywords: Image Inpainting " + }, + { + "bbox": [ + 160, + 638, + 443, + 650 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 638, + 443, + 650 + ], + "type": "text", + "content": " Diffusion Models " + }, + { + "bbox": [ + 160, + 638, + 443, + 650 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 638, + 443, + 650 + ], + "type": "text", + "content": " Image Generation" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 654, + 236, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 236, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 236, + 665 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 135, + 482, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 135, + 482, + 219 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 482, + 219 + ], + "type": "text", + "content": "Image inpainting [45] aims at restoring the missing regions of an image while maintaining the overall coherence. As a long-standing computer vision problem, it facilitates numerous applications such as virtual try-on [18] and image editing [15]. Recently, diffusion models [12,36] have demonstrated impressive performance in image generation, enabling flexible user control with semantic and structural conditions [33,51]. To this end, researchers resort to diffusion-based pipelines for high-quality image inpainting that aligns with given text prompts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 220, + 485, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 485, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 485, + 376 + ], + "type": "text", + "content": "Commonly used diffusion-based text-guided inpainting methods can be roughly divided into two categories: (1) Sampling strategy modification [1,2,6,20,23,47, 50], which modifies the standard denoising process by sampling the masked regions from a pre-trained diffusion model, and the unmasked areas are simply copy-pasted from the given image in each denoising step. Although they can be used in arbitrary diffusion backbones, the limited perceptual knowledge of mask boundaries and the unmasked image region context leads to incoherent inpainting results. (2) Dedicated inpainting models [4, 33, 37, 42, 43, 46, 49, 56], which fine-tune a specially designed image inpainting model by expanding the input channel dimension of base diffusion models to incorporate provided corrupted image and mask. While they enable the diffusion model to generate more satisfying results with specialized content-aware and shape-aware models, we argue, is this architecture the best fit for diffusion-based inpainting?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 472 + ], + "type": "text", + "content": "As shown in Fig. 2, dedicated inpainting models fuse noisy latent, masked image latent, mask, and text at an early stage. This architectural design makes the masked image feature easily influenced by the text embedding, preventing subsequent layers in the UNet from obtaining pure masked image features due to the influence of text. Additionally, handling the condition and generation in a single branch imposes extra burdens on the UNet framework. These approaches also necessitate fine-tuning in different variations of diffusion backbones, which can be time-consuming with limited transferability." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 141, + 478, + 473, + 610 + ], + "blocks": [ + { + "bbox": [ + 141, + 478, + 473, + 610 + ], + "lines": [ + { + "bbox": [ + 141, + 478, + 473, + 610 + ], + "spans": [ + { + "bbox": [ + 141, + 478, + 473, + 610 + ], + "type": "image", + "image_path": "c3c37ed7f16079a11ef5a942c52021579f3d07987efc14995ca68236ffa87cb2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 613, + 473, + 624 + ], + "lines": [ + { + "bbox": [ + 141, + 613, + 473, + 624 + ], + "spans": [ + { + "bbox": [ + 141, + 613, + 473, + 624 + ], + "type": "text", + "content": "Fig. 2: Comparison of previous inpainting architectures and BrushNet." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Adding an additional branch dedicated to masked image feature extraction could be a promising approach to tackle the issue above. However, the existing solutions such as ControlNet [51] lead to inadequate information extraction" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "Xuan Ju, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "and insertion when directly applied to inpainting, which originates from differences between controllable image generation and inpainting: inpainting requires pixel-to-pixel constraints with strong restrictive information instead of sparse structural control relying on text for content completion. As a result, ControlNet yields unsatisfactory results compared to dedicated inpainting models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 319 + ], + "type": "text", + "content": "To address this issue, we propose BrushNet, which introduces an additional branch to the diffusion framework, creating a more suitable architecture for image inpainting. Specifically, our designs are three-fold: (1) To improve the extraction of image features for adaptation to the UNet distribution, we use a VAE encoder instead of randomly initialized convolution layers to process the masked image. (2) To enable dense per-pixel control, we adopt a hierarchical approach by gradually incorporating the full UNet feature layer-by-layer into the pre-trained UNet. (3) To ensure pure image information is considered in the additional branch, we remove text cross-attention from UNet. This design further offers plug-and-play capabilities and flexible unmasked region controllability to the inpainting process. For better consistency and a larger range of unmasked region controllability, we additionally propose a blurred blending strategy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 320, + 481, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 481, + 415 + ], + "type": "text", + "content": "To ensure a comprehensive evaluation for real-life applications, we categorize inpainting tasks into two distinct types based on mask shape: random brush masks and segmentation-based masks. We utilize EditBench [37] as the comparing benchmark for random brush mask inpainting. Additionally, we introduce a new training dataset BrushData and a new benchmark BrushBench for training and evaluating segmentation-based mask inpainting. Results show BrushNet achieve state-of-the-art performance across 7 metrics encompassing image quality, masked region preservation, and text alignment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 424, + 237, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 424, + 237, + 436 + ], + "spans": [ + { + "bbox": [ + 131, + 424, + 237, + 436 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 438, + 480, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 480, + 521 + ], + "type": "text", + "content": "Image inpainting is a classic problem in computer vision, aiming to restore masked regions of an image with plausible and natural content [29, 45]. Previous methods based on traditional techniques [3, 7], Variational Auto-Encoders (VAEs) [27, 54] and Generative Adversarial Networks (GANs) [21, 53, 55] often require auxiliary hand-engineered features but yield poor results. Recently, diffusion-based methods [1, 2, 20, 23, 31, 42] gain popularity due to their impressive high-quality generation, fine-grained control, and output diversity [12, 13, 33]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 522, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 481, + 640 + ], + "type": "text", + "content": "Initial attempts at utilizing diffusion models for text-guided inpainting [1, 2,6,20,23,47,50], such as Blended Latent Diffusion, modifies the standard denoising strategy by sampling the masked regions from a pre-trained diffusion model and the unmasked areas from the given image, which is commonly used as the default inpainting choice in widely-used image generation libraries like Diffusers [28]. Although these methods demonstrate satisfactory results in simple image inpainting tasks and can be plug-and-play to any diffusion model, they struggle with complex mask shapes, image contents, and text prompts, leading to results that lack coherence. This is primarily attributed to their limited perceptual knowledge of mask boundaries and the unmasked image region context." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Previous works [4,33,37,42,43,46,49,56] address this issue by fine-tuning the base models into content-aware and shape-aware models specifically designed" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "BrushNet" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "for image inpainting. Specifically, SmartBrush [42] augments the diffusion U-Net with object-mask prediction to guide the sampling process with mask boundaries information. Stable Diffusion Inpainting [33] fine-tunes a diffusion model specifically designed for inpainting tasks, taking the mask, masked image, and noisy latent as inputs to UNet architecture. HD-Painter [25] and PowerPaint [56] build upon the foundation of Stable Diffusion Inpainting, separately enhancing the generation quality and enabling the model to perform multiple tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 296 + ], + "type": "text", + "content": "However, these approaches make it difficult to effectively transfer their inpainting ability to arbitrary pre-trained models, restricting their applicability. To enable any diffusion model with inpainting capabilities, the community fine-tunes ControlNet [51] on inpainting image pairs. However, the model design of ControlNet exhibits limitations in its perceptual understanding of masks and masked images, which consequently leads to unsatisfactory outcomes. Compared with previous methods (shown in Tab. 1), BrushNet is plug-and-play, content-aware, and shape-aware, with a flexible preserving degree for unmasked regions." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 142, + 358, + 478, + 463 + ], + "blocks": [ + { + "bbox": [ + 131, + 301, + 482, + 357 + ], + "lines": [ + { + "bbox": [ + 131, + 301, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 131, + 301, + 482, + 357 + ], + "type": "text", + "content": "Table 1: Comparison of BrushNet with Previous Image Inpainting Methods. BrushNet offers the advantage of being plug-and-play with any pretrained diffusion model. Moreover, it allows for flexible control over the scale of inpainting and is designed to be aware of both the mask shape and the unmasked content. Note that we only list commonly used text-guided diffusion methods in this table." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 142, + 358, + 478, + 463 + ], + "lines": [ + { + "bbox": [ + 142, + 358, + 478, + 463 + ], + "spans": [ + { + "bbox": [ + 142, + 358, + 478, + 463 + ], + "type": "table", + "html": "
ModelPlug-and-PlayFlexible-ScaleContent-AwareShape-Aware
Blended Diffusion [1,2]✓
SmartBrush [42]✓
SD Inpainting [33]✓✓
PowerPaint [56]✓✓
HD-Painter [25]✓✓
ReplaceAnything [4]✓✓
Imagen [37]✓✓
ControlNet-Inpainting [51]✓✓✓
BrushNet✓✓✓✓
", + "image_path": "f64c023054c88432abf10266339c6ce7e8e626b01edbd5ef9ce70979fa106262.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 470, + 328, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 470, + 328, + 482 + ], + "spans": [ + { + "bbox": [ + 131, + 470, + 328, + 482 + ], + "type": "text", + "content": "3 Preliminaries and Motivation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 481, + 521 + ], + "type": "text", + "content": "In this section, we will first introduce diffusion models in Sec. 3.1. Then, Sec. 3.2 would review previous inpainting techniques based on sampling strategy modification and special training. Finally, the motivation is outlined in Section 3.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 525, + 246, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 525, + 246, + 536 + ], + "spans": [ + { + "bbox": [ + 131, + 525, + 246, + 536 + ], + "type": "text", + "content": "3.1 Diffusion Models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "text", + "content": "Diffusion models include a forward process that adds Gaussian noise " + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "text", + "content": " to convert clean sample " + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "text", + "content": " to noise sample " + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "inline_equation", + "content": "z_T" + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "text", + "content": ", and a backward process that iteratively performs denoising from " + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "inline_equation", + "content": "z_T" + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 541, + 482, + 589 + ], + "type": "text", + "content": " represents the total number of timesteps. The forward process can be formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 255, + 600, + 481, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 600, + 481, + 613 + ], + "spans": [ + { + "bbox": [ + 255, + 600, + 481, + 613 + ], + "type": "interline_equation", + "content": "z _ {t} = \\sqrt {\\alpha_ {t}} z _ {0} + \\sqrt {1 - \\alpha_ {t}} \\epsilon \\tag {1}", + "image_path": "619983396016ec01e5b7ee3b8b00ec6d89e44d6a61294074b0ce154e4d0135c2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 146, + 617, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 617, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 146, + 617, + 481, + 629 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 146, + 617, + 481, + 629 + ], + "type": "text", + "content": " is the noised feature at step " + }, + { + "bbox": [ + 146, + 617, + 481, + 629 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 146, + 617, + 481, + 629 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 146, + 617, + 481, + 629 + ], + "type": "inline_equation", + "content": "t\\sim [1,T]" + }, + { + "bbox": [ + 146, + 617, + 481, + 629 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 146, + 617, + 481, + 629 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 146, + 617, + 481, + 629 + ], + "type": "text", + "content": " is a hyper-parameter." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "content": "In the backward process, given input noise " + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "content": " sampled from a random Gaussian distribution, learnable network " + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "content": " estimates noise at each step " + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "content": " conditioned on " + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "content": ". After " + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "content": " progressively refining iterations, " + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "content": " is derived as the output sample:" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "Xuan Ju, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 167, + 135, + 481, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 135, + 481, + 167 + ], + "spans": [ + { + "bbox": [ + 167, + 135, + 481, + 167 + ], + "type": "interline_equation", + "content": "z _ {t - 1} = \\frac {\\sqrt {\\alpha_ {t - 1}}}{\\sqrt {\\alpha_ {t}}} z _ {t} + \\sqrt {\\alpha_ {t - 1}} \\left(\\sqrt {\\frac {1}{\\alpha_ {t - 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {t}} - 1}\\right) \\epsilon_ {\\theta} \\left(z _ {t}, t, C\\right) \\tag {2}", + "image_path": "09489b8514b7e9b49d2540cc6ea0feda3e65bfd4a915ed53b5dfb010cd694dff.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 168, + 479, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 168, + 479, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 168, + 479, + 192 + ], + "type": "text", + "content": "The training of diffusion models revolves around optimizing the denoiser network " + }, + { + "bbox": [ + 132, + 168, + 479, + 192 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 132, + 168, + 479, + 192 + ], + "type": "text", + "content": " to conduct denoising with condition " + }, + { + "bbox": [ + 132, + 168, + 479, + 192 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 132, + 168, + 479, + 192 + ], + "type": "text", + "content": ", guided by the objective:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 217, + 199, + 481, + 215 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 199, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 217, + 199, + 481, + 215 + ], + "type": "interline_equation", + "content": "\\underset {\\theta} {\\min } E _ {z _ {0}, \\epsilon \\sim \\mathcal {N} (0, I), t \\sim U (1, T)} \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, C) \\| \\tag {3}", + "image_path": "be767126215582cc6bc81d141243f702bbb65c4518d1653d76ac2a0bf0741965.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 222, + 301, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 222, + 301, + 234 + ], + "spans": [ + { + "bbox": [ + 132, + 222, + 301, + 234 + ], + "type": "text", + "content": "3.2 Previous Inpainting Models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 239, + 480, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 480, + 358 + ], + "type": "text", + "content": "Sampling Strategy Modification. This line of research achieves inpainting by gradually blending masked images with the generated results. The most used method among them is Blended Latent Diffusion (BLD) [1], serving as the default choice for inpainting in widely-used diffusion-based image generation libraries (e.g., Diffusers [28]). Given a binary mask " + }, + { + "bbox": [ + 130, + 239, + 480, + 358 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 239, + 480, + 358 + ], + "type": "text", + "content": " and a masked image " + }, + { + "bbox": [ + 130, + 239, + 480, + 358 + ], + "type": "inline_equation", + "content": "x_0^{masked}" + }, + { + "bbox": [ + 130, + 239, + 480, + 358 + ], + "type": "text", + "content": ", BLD first extracts the latent representation " + }, + { + "bbox": [ + 130, + 239, + 480, + 358 + ], + "type": "inline_equation", + "content": "z_0^{masked}" + }, + { + "bbox": [ + 130, + 239, + 480, + 358 + ], + "type": "text", + "content": " of the masked image using VAE. Subsequently, the mask " + }, + { + "bbox": [ + 130, + 239, + 480, + 358 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 239, + 480, + 358 + ], + "type": "text", + "content": " is resized to " + }, + { + "bbox": [ + 130, + 239, + 480, + 358 + ], + "type": "inline_equation", + "content": "m^{resized}" + }, + { + "bbox": [ + 130, + 239, + 480, + 358 + ], + "type": "text", + "content": " to match the size of the latent representation. To formulate the inpainting process, BLD adds Gaussian noise to " + }, + { + "bbox": [ + 130, + 239, + 480, + 358 + ], + "type": "inline_equation", + "content": "z_0^{masked}" + }, + { + "bbox": [ + 130, + 239, + 480, + 358 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 239, + 480, + 358 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 239, + 480, + 358 + ], + "type": "text", + "content": " steps and gets " + }, + { + "bbox": [ + 130, + 239, + 480, + 358 + ], + "type": "inline_equation", + "content": "z_t^{masked}" + }, + { + "bbox": [ + 130, + 239, + 480, + 358 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 239, + 480, + 358 + ], + "type": "inline_equation", + "content": "t \\sim [1,T]" + }, + { + "bbox": [ + 130, + 239, + 480, + 358 + ], + "type": "text", + "content": ". Then, denoising steps start from " + }, + { + "bbox": [ + 130, + 239, + 480, + 358 + ], + "type": "inline_equation", + "content": "z_T^{masked}" + }, + { + "bbox": [ + 130, + 239, + 480, + 358 + ], + "type": "text", + "content": ", where each sampling step in eq. 2 is followed by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 202, + 366, + 480, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 366, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 202, + 366, + 480, + 380 + ], + "type": "interline_equation", + "content": "z _ {t - 1} \\leftarrow z _ {t - 1} \\cdot \\left(1 - m ^ {\\text {r e s i z e d}}\\right) + z _ {t - 1} ^ {\\text {m a s k e d}} \\cdot m ^ {\\text {r e s i z e d}} \\tag {4}", + "image_path": "6459b3b8675c6d57bfe2b0adf647a014e0e0f11362d4d290fbb58c20a4edfbd7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 384, + 479, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 479, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 479, + 445 + ], + "type": "text", + "content": "Despite its simplicity in implementation, BLD exhibits suboptimal performance in terms of both unmasked region preservation and generation content alignment. This is due to (1) the resize of the mask preventing it from correctly blending the noisy latent, (2) the diffusion model lacking perceptual knowledge of mask boundaries and the unmasked image region context." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 450, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 606 + ], + "type": "text", + "content": "Dedicated Inpainting Models. To enhance the performance of inpainting, previous works fine-tune the base models by expanding the input UNet channel to include the mask and masked image inputs, turning it into an architecture specifically designed for image inpainting. Though having better generation results compared with BLD, they still have several drawbacks: (1) These models merge the noisy latent, masked image latent, and mask at the initial convolution layer of the UNet architecture, where they are collectively influenced by the text embedding. Consequently, subsequent layers in the UNet model struggle to obtain pure masked image features due to the text's influence. (2) Incorporating both the condition processing and generation within a single branch places additional burdens on the UNet framework. (3) These approaches require extensive fine-tuning across various variations of diffusion backbones, which is computationally intensive and lacks transferability to custom diffusion models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 614, + 217, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 614, + 217, + 625 + ], + "spans": [ + { + "bbox": [ + 132, + 614, + 217, + 625 + ], + "type": "text", + "content": "3.3 Motivation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": "Based on the analysis presented in Section 3.2, a more effective architecture design of inpainting would be introducing an additional branch specifically dedicated to masked image processing. ControlNet [51] is one of the widely adopted" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "BrushNet" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "strategies that exemplifies this idea. However, it should be noted that directly fine-tuning ControlNet, which is originally designed for controllable image generation, on the inpainting task yields unsatisfactory results. ControlNet designs a lightweight encoder to incorporate out-of-domain structural conditions (e.g., skeleton) and relies on text guidance for content generation, which is unsuitable for pixel-level inpainting image feature injection. Furthermore, ControlNet typically relies on sparse control, meaning that merely adding control to residuals in the UNet framework would be sufficient, while inpainting requires pixel-to-pixel constraints with strong restrictive information. Thus a new architecture specifically designed for inpainting is urgently needed." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 244, + 202, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 244, + 202, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 244, + 202, + 257 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 264, + 481, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 264, + 481, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 481, + 312 + ], + "type": "text", + "content": "An overview of BrushNet is shown in Fig. 3. We employ a dual-branch strategy for masked image guidance insertion (Sec. 4.1). Blending operations with a blurred mask is used to ensure better-unmasked region preservation (Sec. 4.2). Notably, BrushNet can achieve flexible control by adjusting the added scale." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 134, + 325, + 480, + 483 + ], + "blocks": [ + { + "bbox": [ + 134, + 325, + 480, + 483 + ], + "lines": [ + { + "bbox": [ + 134, + 325, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 134, + 325, + 480, + 483 + ], + "type": "image", + "image_path": "adc686765481ed163c4ad9e77e71ec36827bf43a5ff44056a215663840e6bdec.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 502, + 482, + 580 + ], + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 580 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 580 + ], + "type": "text", + "content": "Fig. 3: Model overview. Our model outputs an inpainted image given the mask and masked image input. Firstly, we downsample the mask to accommodate the size of the latent, and input the masked image to the VAE encoder to align the distribution of latent space. Then, noisy latent, masked image latent, and downsampled mask are concatenated as the input of BrushNet. The feature extracted from BrushNet is added to pretrained UNet layer by layer after a zero convolution block [51]. After denoising, the generated image and masked image are blended with a blurred mask." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 590, + 285, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 285, + 602 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 285, + 602 + ], + "type": "text", + "content": "4.1 Masked Image Guidance" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "The insertion of the masked image feature into the pre-trained diffusion network is accomplished through an additional branch, which explicitly separates the feature extraction of masked images from the image-generating processes. The input to the additional branch includes noisy latent, masked image latent, and downsampled mask, which are concatenated together to form the input." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "Xuan Ju, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": "Specifically, the noisy latent provides generation information during the present generation process, helping BrushNet enhance the semantic coherence of the masked image feature. The masked image latent is extracted from the masked image using VAE, which is aligned with the data distribution of the pre-trained UNet. To ensure the alignment of the mask size with the noisy latent and masked image latent, we employ cubic interpolation to downsample the mask." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 189, + 480, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 480, + 307 + ], + "type": "text", + "content": "To process the masked image features, BrushNet utilizes a clone of the pre-trained diffusion model while excluding its cross-attention layers. The pretrained weights of the diffusion model serve as a strong prior for extracting the masked image features, while the removal of the cross-attention layers ensures that only pure image information is considered within this additional branch. BrushNet feature is inserted into the frozen diffusion model layer-by-layer, enabling dense per-pixel control hierarchically. Similar to ControlNet [51], we employ zero convolution layers to establish a connection between the locked model and the trainable BrushNet. This ensures that harmful noise does not influence the hidden states in the trainable copy during the initial stages of training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "text", + "content": "The feature insertion operation is shown in Eq. 5. Specifically, " + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t,t,C)_i" + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "text", + "content": " indicates the feature of the " + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "text", + "content": "-th layer in network " + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "inline_equation", + "content": "i\\sim [1,n]" + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "text", + "content": " is the number of layers. The same notation applies to " + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{BrushNet}" + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{BrushNet}" + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "text", + "content": " takes the concatenated noisy latent " + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "text", + "content": ", masked image latent " + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "inline_equation", + "content": "z_0^{masked}" + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "text", + "content": ", and downsampled mask " + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "inline_equation", + "content": "m^{resized}" + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "text", + "content": " as input, with the concatenation operation denoted as " + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "inline_equation", + "content": "[\\cdot]" + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "text", + "content": " is the zero convolution operation. " + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 308, + 480, + 392 + ], + "type": "text", + "content": " is the preservation scale used to adjust the influence of BrushNet on pretrained diffusion model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 402, + 480, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 402, + 480, + 417 + ], + "spans": [ + { + "bbox": [ + 141, + 402, + 480, + 417 + ], + "type": "interline_equation", + "content": "\\epsilon_ {\\theta} \\left(z _ {t}, t, C\\right) _ {i} = \\epsilon_ {\\theta} \\left(z _ {t}, t, C\\right) _ {i} + w \\cdot \\mathcal {Z} \\left(\\epsilon_ {\\theta} ^ {\\text {B r u s h N e t}} \\left(\\left[ z _ {t}, z _ {0} ^ {\\text {m a s k e d}}, m ^ {\\text {r e s i z e d}} \\right], t\\right) _ {i}\\right) \\tag {5}", + "image_path": "5c425c67eebe5e0c166bcad3b07071fb0a4d540d547f9ef51e10ef329ec08cff.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 430, + 261, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 430, + 261, + 443 + ], + "spans": [ + { + "bbox": [ + 131, + 430, + 261, + 443 + ], + "type": "text", + "content": "4.2 Blending Operation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 449, + 479, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 479, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 479, + 521 + ], + "type": "text", + "content": "As mentioned in Section 4.2, the blending operation conducted in latent space can result in inaccuracies due to the resizing of the mask. Similarly, in our approach, a similar issue arises as we resize the mask to match the size of the latent space, which can introduce potential inaccuracies. Additionally, it is important to acknowledge that VAE encoding and decoding operations have inherent limitations and may not ensure complete image reconstruction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 480, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 480, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 480, + 605 + ], + "type": "text", + "content": "To ensure a fully consistent image reconstruction of the unmasked region, previous works have explored different techniques. Some approaches [4, 56], utilize past-and-copy methods, where the unmasked region is directly copied from the original image. However, this can result in a lack of semantic coherence in the final generation results. On the other hand, adopting latent blending operations inspired by BLD [1, 33] has been observed to face challenges in effectively preserving the desired information in the unmasked regions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "text", + "content": "In this work, we present a simple pixel space solution to address this issue by first blurring the mask and then performing copy-and-paste using the blurred mask. Although this approach may result in a slight loss of accuracy in preserving the details of the mask boundary, the error is nearly imperceptible to the naked eye and results in significantly improved coherence in the mask boundary." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "BrushNet" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 243, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 243, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 243, + 127 + ], + "type": "text", + "content": "4.3 Flexible Control" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 482, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 482, + 313 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 482, + 313 + ], + "type": "text", + "content": "The architecture design of BrushNet inherently makes it suitable for seamless plug-and-play integration to various pretrained diffusion models and enables flexible preservation scale. Specifically, the flexible control of our proposed BrushNet includes: (1) Since BrushNet does not modify the weights of the pretrained diffusion model, it can be readily integrated as a plug-and-play component with any community fine-tuned diffusion models. This allows for easy adoption and experimentation with different pretrained models. (2) Preservation Scale Adjustment: The preservation scale of the unmasked region can be controlled by incorporating BrushNet features into the frozen diffusion model with the weight " + }, + { + "bbox": [ + 130, + 133, + 482, + 313 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 133, + 482, + 313 + ], + "type": "text", + "content": ". This weight determines the influence of BrushNet on the preservation scale, offering the ability to adjust the desired level of preservation. (3) Blurring Scale and Blending Operation: By adjusting the scale of blurring and deciding whether to apply the blending operation, the preservation scale of the unmasked region can be further customized. These features allow for flexible fine-grained control over the inpainting process. More explanation can be found in Sec. 5.5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 321, + 230, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 230, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 230, + 335 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 338, + 339, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 338, + 339, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 338, + 339, + 350 + ], + "type": "text", + "content": "5.1 Evaluation Benchmark and Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "text", + "content": "Benchmark. Previous commonly used datasets in the image inpainting field include CelebA [22], CelebA-HQ [14], ImageNet [8], MSCOCO [19], Open Images [17], and LSUN-Bedroom [48]. However, these datasets either primarily focus on a small area, such as human faces, or predominantly consist of low-quality, cluttered real-life scene data. As a result, these datasets are not well-suited for training and evaluating diffusion-based inpainting models, which can generate high-quality diverse images that align with text prompts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 438, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 533 + ], + "type": "text", + "content": "Recently proposed EditBench [37] serves as a benchmark specifically designed for text-guided image inpainting for diffusion models. This benchmark consists of a collection of 240 images comprising an equal ratio of natural images and generated images, with mask and caption annotation for each image. However, the annotated masks in EditBench are mostly random shapes without specific object information, neglecting the practical application of inpainting in real scenarios such as replacing an object with an external mask, as commonly seen in E-commerce product displays and image editing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 534, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 629 + ], + "type": "text", + "content": "To fill the gap, we propose BrushBench for segmentation-based inpainting, as shown in Fig. 4. BrushBench comprises a total of 600 images, with each image accompanied by the human-annotated mask and caption annotation. The images in BrushBench are evenly distributed between natural images and artificial images, such as paintings. Furthermore, the dataset ensures an equal distribution among different categories, including humans, animals, indoor scenarios, and outdoor scenarios. This balanced distribution enables a fair evaluation across various categories, promoting better evaluation equity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "To further enhance the analysis of the inpainting task, we categorize it into two distinct types based on the masks used: random brush masks and segmentation-based masks. We use EditBench as the comparison benchmark" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "Xuan Ju, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 115, + 477, + 273 + ], + "blocks": [ + { + "bbox": [ + 136, + 115, + 477, + 273 + ], + "lines": [ + { + "bbox": [ + 136, + 115, + 477, + 273 + ], + "spans": [ + { + "bbox": [ + 136, + 115, + 477, + 273 + ], + "type": "image", + "image_path": "9e81c8cb4347892949d92e8b70253f27259f3ee930f7f7ddf10488f1d0df5858.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 276, + 482, + 342 + ], + "lines": [ + { + "bbox": [ + 130, + 276, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 482, + 342 + ], + "type": "text", + "content": "Fig. 4: Benchmark overview. I and II separately show natural and artificial images, masks, and caption of BrushBench. (a) to (d) show images of humans, animals, indoor scenarios, and outdoor scenarios. Each group of images shows the original image, inside-inpainting mask, and outside-inpainting mask, with an image caption on the top. III show image, mask, and caption from EditBench [37], with (e) for generated images and (f) for natural images. The images are randomly selected from both benchmarks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 344, + 482, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 482, + 405 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 482, + 405 + ], + "type": "text", + "content": "for random brush masks and use BrushBench for segmentation-based masks. Regarding inpainting with segmentation-based masks, we refine the task by considering two specific scenarios: segmentation mask inside-inpainting and segmentation mask outside-inpainting. By separating these two subtasks, we can better understand the inpainting performance in different image regions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 407, + 482, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 478 + ], + "type": "text", + "content": "Dataset. To train segmentation-based mask inpainting, we annotate segmentation mask on Laion-Aesthetic [34] dataset, called BrushData. We employ the Grounded-SAM [32] to annotate open-world masks and then filter the masks based on their confidence score, ensuring that only masks with relatively higher confidence scores are retained. Additionally, we consider factors such as reasonable mask size and good mask continuity during the filtering process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 480, + 481, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 481, + 504 + ], + "type": "text", + "content": "Metrics. We consider 7 metrics from three aspects: image generation quality, masked region preservation, and text alignment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 506, + 482, + 648 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 506, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 506, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 506, + 482, + 601 + ], + "type": "text", + "content": "- Image Generation Quality. Metrics most used by previous inpainting methods (e.g., FID [11] and KID [5]) show a poor representation of the rich and varied content generated by modern text-to-image models [16]. Thus, we use Image Reward (IR) [44], HPS v2 (HPS) [41], and Aesthetic Score (AS) [34] that align with human perception. Specifically, ImageReward and HPS v2 are text-to-image human preference evaluation models trained on large-scale datasets of human preference choices on generated images. Aesthetic Score is a linear model trained on image quality rating pairs of real images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 139, + 601, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 601, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 139, + 601, + 482, + 648 + ], + "type": "text", + "content": "- Masked Region Preservation. We follow previous works using standard Peak Signal-to-Noise Ratio (PSNR) [39], Learned Perceptual Image Patch Similarity (LPIPS) [52], and Mean Squared Error (MSE) [38] in the unmasked region among the generated image and the original image." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "type": "text", + "content": "BrushNet" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 471, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 471, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 471, + 666 + ], + "type": "text", + "content": "1 The proposed BrushData and BrushBench will be released along with the codes." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 115, + 482, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 115, + 482, + 165 + ], + "spans": [ + { + "bbox": [ + 138, + 115, + 482, + 165 + ], + "type": "text", + "content": "- Text Alignment. We use CLIP Similarity (CLIP Sim) [40] to evaluate text-image consistency between the generated images and corresponding text prompts. CLIP Similarity projects text and images to the same shared space with the CLIP model [30] and evaluates the similarity of their embeddings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 178, + 280, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 178, + 280, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 280, + 190 + ], + "type": "text", + "content": "5.2 Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 197, + 482, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 197, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 197, + 482, + 304 + ], + "type": "text", + "content": "We perform the inference of different inpainting methods in the same setting unless specifically clarified, i.e., on NVIDIA Tesla V100 following their open-source code with a base model of Stabe Diffusion v1.5 in 50 steps, with a guidance scale of 7.5. We keep the recommended hyper-parameter for each inpainting method in all images for fair comparison. BrushNet and all ablation models are trained for 430 thousands steps on 8 NVIDIA Tesla V100 GPUs, which takes around 3 days. For comparison on BrushBench, we use BrushNet trained on BrushData. For comparison on EditBench, we use the model trained on LAION-5B with random masks. Details can be found in the provided codes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 314, + 288, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 314, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 288, + 326 + ], + "type": "text", + "content": "5.3 Quantitative Comparison" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 138, + 415, + 476, + 587 + ], + "blocks": [ + { + "bbox": [ + 130, + 333, + 482, + 409 + ], + "lines": [ + { + "bbox": [ + 130, + 333, + 482, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 482, + 409 + ], + "type": "text", + "content": "Table 2: Quantitative comparisons among BrushNet and other diffusion-based inpainting models in BrushBench: Blended Latent Diffusion (BLD) [1], Stable Diffusion Inpainting (SDI) [33], HD-Painter (HDP) [25], PowerPaint (PP) [56], and ControlNet-Inpainting (CNI) [51]. Metrics encompassing image quality, masked region preservation, and text alignment (Text Align) for inside-inpainting and outside-inpainting are shown in the table. All models use Stable Diffusion V1.5 as base model. Red stands for the best result, Blue stands for the second best result." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 415, + 476, + 587 + ], + "lines": [ + { + "bbox": [ + 138, + 415, + 476, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 476, + 587 + ], + "type": "table", + "html": "
MetricsImage QualityMasked Region PreservationText Align
ModelsIR×10↑HPS×102↑AS↑PSNR↑MSE×103↓LPIPS×103↓CLIP Sim↑
InsideBLD [1]9.7825.876.1721.339.7649.2626.15
SDI [33]11.7227.066.5021.5213.8748.3926.17
HDP [25]11.6826.906.4222.619.9543.5026.37
PP [56]11.4627.356.2421.4332.7348.4326.48
CNI [51]9.926.026.5312.3978.78243.6226.47
CNI* [51]11.2126.926.3922.7324.5843.4926.22
Ours12.3627.406.5321.659.3148.2826.48
Ours*12.6427.786.5131.940.8018.6726.39
OutsideBLD [1]7.8126.776.2315.8535.8621.4026.73
SDI [33]10.2727.996.5518.0419.8715.1327.21
HDP [25]9.6627.796.4618.0322.9915.2226.96
PP [56]7.4528.016.2618.0431.7815.1326.72
CNI [51]9.2627.686.4211.9183.0358.1627.29
CNI* [51]9.5727.766.2817.5037.7219.9526.92
Ours10.8228.026.6418.0622.8615.0827.33
Ours*10.8828.096.6427.822.254.6327.22
", + "image_path": "4bb420381fc7d24dc3243c9038638bfb3021604638161e41f09c9521f4cfa0fa.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 144, + 589, + 250, + 601 + ], + "lines": [ + { + "bbox": [ + 144, + 589, + 250, + 601 + ], + "spans": [ + { + "bbox": [ + 144, + 589, + 250, + 601 + ], + "type": "text", + "content": "* with blending operation" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Tab. 2 and Tab. 3 show the quantitative comparison on BrushBench and EditBench [37]. We compare the inpainting results of sampling strategy modification method Blended Latent Diffusion [1], dedicated inpainting models Stable Diffusion Inpainting [33], HD-Painter [25], and PowerPaint [56], as well as plug-and-play method ControlNet [51] trained on inpainting data." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "Xuan Ju, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "Results demonstrate BrushNet's effectiveness across image quality, masked region preservation, and image-text alignment. Blended Latent Diffusion [1] shows the poorest results in image quality and text alignment, derived from the incoherence between the generated masked and the unmasked given image. At the same time, its performance in masked region preservation is also not satisfactory because of the loss incurred by resized mask blending operation in the latent space. Modified from Stable Diffusion Inpainting [33], HD-Painter [25] and PowerPaint [56] demonstrate comparable performance to Stable Diffusion Inpainting in the task of inside-inpainting. However, when it comes to outside-inpainting, their results in terms of image quality and text alignment are significantly poorer compared to Stable Diffusion Inpainting, which can be attributed to their exclusive emphasis on the inside-inpainting task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 261, + 482, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 333 + ], + "type": "text", + "content": "ControlNet [51] trained on inpainting has the most similar experimental configuration to ours. Due to the mismatch between its model design and the inpainting task, ControlNet shows poor results in masked region preservation and image quality, necessitating its combination with Blended Latent Diffusion [1] to generate satisfying inpainted images. However, even with this combination, it still falls short compared to dedicated inpainting models and BrushNet." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 143, + 387, + 469, + 506 + ], + "blocks": [ + { + "bbox": [ + 130, + 340, + 483, + 384 + ], + "lines": [ + { + "bbox": [ + 130, + 340, + 483, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 483, + 384 + ], + "type": "text", + "content": "Table 3: Quantitative comparisons among BrushNet and other diffusion-based inpainting models in EditBench. A detailed explanation of compared methods and metrics can be found in the caption of Tab. 2. Red stands for the best result, Blue stands for the second best result." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 387, + 469, + 506 + ], + "lines": [ + { + "bbox": [ + 143, + 387, + 469, + 506 + ], + "spans": [ + { + "bbox": [ + 143, + 387, + 469, + 506 + ], + "type": "table", + "html": "
MetricsImage QualityMasked Region PreservationText Align
ModelsIR×10↑HPS×102↑AS↑PSNR↑MSE×103↓LPIPS×103↓CLIP Sim↑
BLD [1]0.9023.815.4420.8910.9331.9028.62
SDI [33]1.8624.245.6923.256.9424.3028.00
HDP [25]1.7424.205.6423.076.7024.3228.34
PP [56]1.2424.505.4423.3420.1224.1227.80
CNI [51]1.4924.465.8212.7169.42159.7128.16
CNI* [51]0.9023.795.4622.6135.9326.1427.74
Ours4.4025.105.8423.356.8124.1128.67
Ours*4.4625.245.8233.660.6310.1228.87
", + "image_path": "a010d12ca81ff675be16f7515ddfebacbd68e649f8b47632cce3f9a16d7025ab.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 149, + 507, + 255, + 519 + ], + "lines": [ + { + "bbox": [ + 149, + 507, + 255, + 519 + ], + "spans": [ + { + "bbox": [ + 149, + 507, + 255, + 519 + ], + "type": "text", + "content": "* with blending operation" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 526, + 482, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 526, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 526, + 482, + 588 + ], + "type": "text", + "content": "The performance on the EditBench is roughly consistent with the overall performance on BrushBench, which similarly shows BrushNet's superior performance. This indicates that our method exhibits strong performance across a range of inpainting tasks with various mask types, including random masks, inside-inpainting masks, and outside-inpainting masks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 597, + 281, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 597, + 281, + 611 + ], + "spans": [ + { + "bbox": [ + 131, + 597, + 281, + 611 + ], + "type": "text", + "content": "5.4 Qualitative Comparison" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "The qualitative comparison with previous image inpainting methods is shown in Fig. 1. We provide results on artificial images and natural images across various inpainting tasks, including random mask inpainting, segmentation mask inside-inpainting, and segmentation mask outside-inpainting. BrushNet consistently" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "BrushNet" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "show exceptional results in the coherent of generated region and unmasked region, considering content (I, II right, III right, IV), color (II left), and text (III left). Interestingly, Fig. 1 III left requires the model to generate a cat and a goldfish. All previous methods fail to recognize that a goldfish is already present in the masked image, resulting in the generation of an additional fish within the masked region. BrushNet successfully realized the awareness of background information due to the design of the dual-branch decoupling." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 204, + 476, + 424 + ], + "blocks": [ + { + "bbox": [ + 133, + 204, + 476, + 424 + ], + "lines": [ + { + "bbox": [ + 133, + 204, + 476, + 424 + ], + "spans": [ + { + "bbox": [ + 133, + 204, + 476, + 424 + ], + "type": "image", + "image_path": "51b2813a57d023f97be77d4b34e3c2fb7e895b8b4b20ce7618e651880ca3b5d3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 430, + 482, + 453 + ], + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 453 + ], + "type": "text", + "content": "Fig. 5: Comparison of previous inpainting methods and BrushNet on various image domain. A detailed explanation of compared methods is in Fig. 1." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 456, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 529 + ], + "type": "text", + "content": "The untouched pre-trained diffusion branch also provides the advantage of better coverage across different data domains, such as painting and anime. As shown in Fig. 5, BrushNet demonstrates superior performance across various image categories, including natural image (I, II), pencil painting (III), anime (IV), illustration (V), digital art (VI left), and watercolor (VI right). Due to the page limit, more qualitative comparison results are in supplementary files." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 535, + 282, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 535, + 282, + 547 + ], + "spans": [ + { + "bbox": [ + 131, + 535, + 282, + 547 + ], + "type": "text", + "content": "5.5 Flexible Control Ability" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 555, + 482, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 676 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 676 + ], + "type": "text", + "content": "Fig. 6 and Fig. 7 illustrate the flexible control provided by BrushNet in two aspects: base diffusion model selection and controlling scale. In Fig. 6, we showcase the ability to combine BrushNet with different diffusion models fine-tuned by the community. This allows users to select a specific model that best suits their inpainting requirements, enabling users to achieve the desired inpainting results based on their specific needs. Fig. 7 demonstrates the adjustment of the control scale of BrushNet. This control scale parameter allows users to effectively control the extent of unmasked region protection during the inpainting process. By manipulating the scale parameter, users can achieve fine-grained control over the inpainting process, enabling precise and customizable inpainting." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "Xuan Ju, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 125, + 228, + 155 + ], + "blocks": [ + { + "bbox": [ + 145, + 116, + 297, + 125 + ], + "lines": [ + { + "bbox": [ + 145, + 116, + 297, + 125 + ], + "spans": [ + { + "bbox": [ + 145, + 116, + 297, + 125 + ], + "type": "text", + "content": "A single red rose is laying on top of snow" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 145, + 125, + 228, + 155 + ], + "lines": [ + { + "bbox": [ + 145, + 125, + 228, + 155 + ], + "spans": [ + { + "bbox": [ + 145, + 125, + 228, + 155 + ], + "type": "image", + "image_path": "76228267e1ed329a83533c4805e99107400c5df7e9b93689366e220d820d07b8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 228, + 125, + 310, + 155 + ], + "blocks": [ + { + "bbox": [ + 228, + 125, + 310, + 155 + ], + "lines": [ + { + "bbox": [ + 228, + 125, + 310, + 155 + ], + "spans": [ + { + "bbox": [ + 228, + 125, + 310, + 155 + ], + "type": "image", + "image_path": "ac187c77211becdb3dcd38a5f8135018bf4e8efdadf8d999fb857d70df92928c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 312, + 125, + 424, + 155 + ], + "blocks": [ + { + "bbox": [ + 312, + 116, + 415, + 125 + ], + "lines": [ + { + "bbox": [ + 312, + 116, + 415, + 125 + ], + "spans": [ + { + "bbox": [ + 312, + 116, + 415, + 125 + ], + "type": "text", + "content": "A cartoon boy driving a car" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 125, + 424, + 155 + ], + "lines": [ + { + "bbox": [ + 312, + 125, + 424, + 155 + ], + "spans": [ + { + "bbox": [ + 312, + 125, + 424, + 155 + ], + "type": "image", + "image_path": "bf93efd098fa90b639fdba74566759e61b1f29078ad7f469d2bd2291bc6f9c45.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 424, + 125, + 449, + 155 + ], + "blocks": [ + { + "bbox": [ + 424, + 125, + 449, + 155 + ], + "lines": [ + { + "bbox": [ + 424, + 125, + 449, + 155 + ], + "spans": [ + { + "bbox": [ + 424, + 125, + 449, + 155 + ], + "type": "image", + "image_path": "b48bd04a24f3d4e7b0df40bdbd04ed1e197e6bf9206905c72ef58092cdbd6cd7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 449, + 126, + 481, + 155 + ], + "blocks": [ + { + "bbox": [ + 449, + 126, + 481, + 155 + ], + "lines": [ + { + "bbox": [ + 449, + 126, + 481, + 155 + ], + "spans": [ + { + "bbox": [ + 449, + 126, + 481, + 155 + ], + "type": "image", + "image_path": "0621ab133f65bdddaf98800b2cb4eccd7292e587b6d8f978574208d634121a40.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 145, + 166, + 228, + 195 + ], + "blocks": [ + { + "bbox": [ + 145, + 156, + 307, + 166 + ], + "lines": [ + { + "bbox": [ + 145, + 156, + 307, + 166 + ], + "spans": [ + { + "bbox": [ + 145, + 156, + 307, + 166 + ], + "type": "text", + "content": "A little girl sitting on the grass with her dog" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 145, + 166, + 228, + 195 + ], + "lines": [ + { + "bbox": [ + 145, + 166, + 228, + 195 + ], + "spans": [ + { + "bbox": [ + 145, + 166, + 228, + 195 + ], + "type": "image", + "image_path": "bd949013803c54c98582c8e6f25572344811acf7a0e54fefcfbc4d1ddb753ac2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 197, + 172, + 204 + ], + "lines": [ + { + "bbox": [ + 145, + 197, + 172, + 204 + ], + "spans": [ + { + "bbox": [ + 145, + 197, + 172, + 204 + ], + "type": "text", + "content": "(a) Mask" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 173, + 197, + 193, + 204 + ], + "lines": [ + { + "bbox": [ + 173, + 197, + 193, + 204 + ], + "spans": [ + { + "bbox": [ + 173, + 197, + 193, + 204 + ], + "type": "text", + "content": "(b) DS" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 203, + 197, + 223, + 204 + ], + "lines": [ + { + "bbox": [ + 203, + 197, + 223, + 204 + ], + "spans": [ + { + "bbox": [ + 203, + 197, + 223, + 204 + ], + "type": "text", + "content": "(c) ER" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 228, + 197, + 260, + 204 + ], + "lines": [ + { + "bbox": [ + 228, + 197, + 260, + 204 + ], + "spans": [ + { + "bbox": [ + 228, + 197, + 260, + 204 + ], + "type": "text", + "content": "(d) HR (e)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 214, + 481, + 259 + ], + "lines": [ + { + "bbox": [ + 132, + 214, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 481, + 259 + ], + "type": "text", + "content": "Fig. 6: Integrating BrushNet to community fine-tuned diffusion models. We use five popular community diffusion models fine-tuned from stable diffusion v1.5: DreamShaper (DS) [24], epiCRealism (ER) [9], Henmix_Real (HR) [10], MeinaMix (MM) [26], and Realistic Vision (RV) [35]. MM is specifically designed for anime images." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 228, + 166, + 310, + 195 + ], + "blocks": [ + { + "bbox": [ + 228, + 166, + 310, + 195 + ], + "lines": [ + { + "bbox": [ + 228, + 166, + 310, + 195 + ], + "spans": [ + { + "bbox": [ + 228, + 166, + 310, + 195 + ], + "type": "image", + "image_path": "31ad5eab82caf1d088f4f5ce436feac6fc4d5e4362776b5167b3bdeebadca15c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 197, + 277, + 204 + ], + "lines": [ + { + "bbox": [ + 261, + 197, + 277, + 204 + ], + "spans": [ + { + "bbox": [ + 261, + 197, + 277, + 204 + ], + "type": "text", + "content": ")MM" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 286, + 197, + 304, + 204 + ], + "lines": [ + { + "bbox": [ + 286, + 197, + 304, + 204 + ], + "spans": [ + { + "bbox": [ + 286, + 197, + 304, + 204 + ], + "type": "text", + "content": "(f) RV" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 312, + 167, + 424, + 195 + ], + "blocks": [ + { + "bbox": [ + 312, + 158, + 446, + 166 + ], + "lines": [ + { + "bbox": [ + 312, + 158, + 446, + 166 + ], + "spans": [ + { + "bbox": [ + 312, + 158, + 446, + 166 + ], + "type": "text", + "content": "A woman with her hands on her face" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 167, + 424, + 195 + ], + "lines": [ + { + "bbox": [ + 312, + 167, + 424, + 195 + ], + "spans": [ + { + "bbox": [ + 312, + 167, + 424, + 195 + ], + "type": "image", + "image_path": "3e63eabc35e89d86671afca7305a914e87ef044fa33d92f63aef3afd5717e5ec.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 197, + 340, + 205 + ], + "lines": [ + { + "bbox": [ + 314, + 197, + 340, + 205 + ], + "spans": [ + { + "bbox": [ + 314, + 197, + 340, + 205 + ], + "type": "text", + "content": "(a) Mask" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 345, + 197, + 364, + 204 + ], + "lines": [ + { + "bbox": [ + 345, + 197, + 364, + 204 + ], + "spans": [ + { + "bbox": [ + 345, + 197, + 364, + 204 + ], + "type": "text", + "content": "(b) DS" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 371, + 197, + 391, + 205 + ], + "lines": [ + { + "bbox": [ + 371, + 197, + 391, + 205 + ], + "spans": [ + { + "bbox": [ + 371, + 197, + 391, + 205 + ], + "type": "text", + "content": "(c) ER" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 397, + 197, + 418, + 204 + ], + "lines": [ + { + "bbox": [ + 397, + 197, + 418, + 204 + ], + "spans": [ + { + "bbox": [ + 397, + 197, + 418, + 204 + ], + "type": "text", + "content": "(d) HR" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 419, + 197, + 446, + 204 + ], + "lines": [ + { + "bbox": [ + 419, + 197, + 446, + 204 + ], + "spans": [ + { + "bbox": [ + 419, + 197, + 446, + 204 + ], + "type": "text", + "content": "(e) MM" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 424, + 167, + 479, + 195 + ], + "blocks": [ + { + "bbox": [ + 424, + 167, + 479, + 195 + ], + "lines": [ + { + "bbox": [ + 424, + 167, + 479, + 195 + ], + "spans": [ + { + "bbox": [ + 424, + 167, + 479, + 195 + ], + "type": "image", + "image_path": "87142fa1b19038c9e7a5ff017ae480f0b28790eedb6a717d48f1c59df5fb07bb.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 197, + 472, + 204 + ], + "lines": [ + { + "bbox": [ + 454, + 197, + 472, + 204 + ], + "spans": [ + { + "bbox": [ + 454, + 197, + 472, + 204 + ], + "type": "text", + "content": "(f) RV" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 134, + 279, + 306, + 301 + ], + "blocks": [ + { + "bbox": [ + 134, + 270, + 222, + 279 + ], + "lines": [ + { + "bbox": [ + 134, + 270, + 222, + 279 + ], + "spans": [ + { + "bbox": [ + 134, + 270, + 222, + 279 + ], + "type": "text", + "content": "Book, coffee and purple flower" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 279, + 306, + 301 + ], + "lines": [ + { + "bbox": [ + 134, + 279, + 306, + 301 + ], + "spans": [ + { + "bbox": [ + 134, + 279, + 306, + 301 + ], + "type": "image", + "image_path": "b911b418c06f7d9da12f655394c7a885f3843045c2947495d0207f2ecb5cf4be.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 302, + 233, + 309 + ], + "lines": [ + { + "bbox": [ + 135, + 302, + 233, + 309 + ], + "spans": [ + { + "bbox": [ + 135, + 302, + 233, + 309 + ], + "type": "text", + "content": "A cupcake with a lit candle on top" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 134, + 310, + 306, + 332 + ], + "blocks": [ + { + "bbox": [ + 134, + 310, + 306, + 332 + ], + "lines": [ + { + "bbox": [ + 134, + 310, + 306, + 332 + ], + "spans": [ + { + "bbox": [ + 134, + 310, + 306, + 332 + ], + "type": "image", + "image_path": "abe2cfc0f166cf0321d04f9e8cd038fac82001cdecf00afe15a7b20a16e910a4.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 332, + 156, + 338 + ], + "lines": [ + { + "bbox": [ + 135, + 332, + 156, + 338 + ], + "spans": [ + { + "bbox": [ + 135, + 332, + 156, + 338 + ], + "type": "text", + "content": "(a) Mask" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 157, + 332, + 174, + 338 + ], + "lines": [ + { + "bbox": [ + 157, + 332, + 174, + 338 + ], + "spans": [ + { + "bbox": [ + 157, + 332, + 174, + 338 + ], + "type": "text", + "content": "(b) 1.0" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 175, + 332, + 195, + 338 + ], + "lines": [ + { + "bbox": [ + 175, + 332, + 195, + 338 + ], + "spans": [ + { + "bbox": [ + 175, + 332, + 195, + 338 + ], + "type": "text", + "content": "(c) 0.8" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 203, + 332, + 217, + 338 + ], + "lines": [ + { + "bbox": [ + 203, + 332, + 217, + 338 + ], + "spans": [ + { + "bbox": [ + 203, + 332, + 217, + 338 + ], + "type": "text", + "content": "(d) 0.6" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 223, + 332, + 233, + 338 + ], + "lines": [ + { + "bbox": [ + 223, + 332, + 233, + 338 + ], + "spans": [ + { + "bbox": [ + 223, + 332, + 233, + 338 + ], + "type": "text", + "content": "(e) 0" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 238, + 332, + 258, + 338 + ], + "lines": [ + { + "bbox": [ + 238, + 332, + 258, + 338 + ], + "spans": [ + { + "bbox": [ + 238, + 332, + 258, + 338 + ], + "type": "text", + "content": "(1)0.4" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 265, + 332, + 282, + 338 + ], + "lines": [ + { + "bbox": [ + 265, + 332, + 282, + 338 + ], + "spans": [ + { + "bbox": [ + 265, + 332, + 282, + 338 + ], + "type": "text", + "content": "(g) 0.3" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 349, + 481, + 383 + ], + "lines": [ + { + "bbox": [ + 133, + 349, + 481, + 383 + ], + "spans": [ + { + "bbox": [ + 133, + 349, + 481, + 383 + ], + "type": "text", + "content": "Fig. 7: Flexible control scale of BrushNet. (a) shows the given masked image, (b)-(h) show adding BrushNet with control scale " + }, + { + "bbox": [ + 133, + 349, + 481, + 383 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 133, + 349, + 481, + 383 + ], + "type": "text", + "content": " from 1.0 to 0.2. Results show a gradually diminishing controllable ability from precise to rough control." + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 308, + 279, + 479, + 301 + ], + "blocks": [ + { + "bbox": [ + 308, + 270, + 419, + 278 + ], + "lines": [ + { + "bbox": [ + 308, + 270, + 419, + 278 + ], + "spans": [ + { + "bbox": [ + 308, + 270, + 419, + 278 + ], + "type": "text", + "content": "A small island with a tree and clouds" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 279, + 479, + 301 + ], + "lines": [ + { + "bbox": [ + 308, + 279, + 479, + 301 + ], + "spans": [ + { + "bbox": [ + 308, + 279, + 479, + 301 + ], + "type": "image", + "image_path": "66c7432ed6d7764ba3dd97737f94396a9c3f103455cc16eeb8b2fd5530993580.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 308, + 310, + 479, + 332 + ], + "blocks": [ + { + "bbox": [ + 308, + 302, + 452, + 310 + ], + "lines": [ + { + "bbox": [ + 308, + 302, + 452, + 310 + ], + "spans": [ + { + "bbox": [ + 308, + 302, + 452, + 310 + ], + "type": "text", + "content": "A white rabbit sitting on the ground in the grass" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 310, + 479, + 332 + ], + "lines": [ + { + "bbox": [ + 308, + 310, + 479, + 332 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 479, + 332 + ], + "type": "image", + "image_path": "599b7b1af0165525307e84eac87130cf4c592dc8bf4149a778b39fa2275e9dff.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 332, + 347, + 338 + ], + "lines": [ + { + "bbox": [ + 309, + 332, + 347, + 338 + ], + "spans": [ + { + "bbox": [ + 309, + 332, + 347, + 338 + ], + "type": "text", + "content": "(a) Mask (b) 1.0" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 349, + 332, + 369, + 338 + ], + "lines": [ + { + "bbox": [ + 349, + 332, + 369, + 338 + ], + "spans": [ + { + "bbox": [ + 349, + 332, + 369, + 338 + ], + "type": "text", + "content": "(c) 0.8" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 371, + 332, + 391, + 338 + ], + "lines": [ + { + "bbox": [ + 371, + 332, + 391, + 338 + ], + "spans": [ + { + "bbox": [ + 371, + 332, + 391, + 338 + ], + "type": "text", + "content": "(d) 0.6" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 393, + 332, + 413, + 338 + ], + "lines": [ + { + "bbox": [ + 393, + 332, + 413, + 338 + ], + "spans": [ + { + "bbox": [ + 393, + 332, + 413, + 338 + ], + "type": "text", + "content": "(e) 0.5" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 332, + 438, + 338 + ], + "lines": [ + { + "bbox": [ + 417, + 332, + 438, + 338 + ], + "spans": [ + { + "bbox": [ + 417, + 332, + 438, + 338 + ], + "type": "text", + "content": "(f) 0.4" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 440, + 332, + 457, + 338 + ], + "lines": [ + { + "bbox": [ + 440, + 332, + 457, + 338 + ], + "spans": [ + { + "bbox": [ + 440, + 332, + 457, + 338 + ], + "type": "text", + "content": "(g) 0.3" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "bbox": [ + 132, + 394, + 238, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 394, + 238, + 406 + ], + "spans": [ + { + "bbox": [ + 132, + 394, + 238, + 406 + ], + "type": "text", + "content": "5.6 Ablation Study" + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 131, + 411, + 481, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 411, + 481, + 467 + ], + "spans": [ + { + "bbox": [ + 131, + 411, + 481, + 467 + ], + "type": "text", + "content": "Table 4: Ablation on dual-branch design. Stable Diffusion Inpainting (SDI) use single-branch design, where the entire UNet is fine-tuned. We conducted an ablation analysis by training a dual-branch model with two variations: one with the base UNet fine-tuned, and another with the base UNet forzened. Results demonstrate the superior performance achieved by adopting the dual-branch design. Red is the best result." + } + ] + } + ], + "index": 53 + }, + { + "type": "table", + "bbox": [ + 138, + 473, + 473, + 536 + ], + "blocks": [ + { + "bbox": [ + 138, + 473, + 473, + 536 + ], + "lines": [ + { + "bbox": [ + 138, + 473, + 473, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 473, + 473, + 536 + ], + "type": "table", + "html": "
MetricsImage QualityMasked Region PreservationText Align
ModelIR×10↑HPS×102↑AS↑PSNR↑MSE×102↓LPIPS×103↓CLIP Sim↑
SDI11.0027.536.5319.7816.8731.7626.69
w/o fine-tune11.5927.716.5919.8616.0931.6826.91
w/ fine-tune11.6327.736.6020.1315.8431.5726.93
", + "image_path": "45fbcbccd03fc4eb714088d1e9ece8cd84bcc629eace8eefd0c0a334eb917449.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "table_body" + } + ], + "index": 54 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "We conducted ablation studies to investigate the impact of different model designs. Tab. 4 compares the dual-branch and single-branch designs. Tab. 5 shows the ablation study focusing on the additional branch architecture. The ablation studies are conducted on BrushBench, averaging the performance of inside-inpainting and outside-inpainting. The results presented in Tab. 4 demonstrate that the dual-branch design significantly outperforms the single-branch design. Additionally, fine-tuning the base diffusion model in dual-branch design yields better results compared to freezing it. However, fine-tuning the base diffusion model may restrict flexibility and control over the model. Considering this trade-off between performance and flexibility, we decide to adopt the frozen" + } + ] + } + ], + "index": 55 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 407, + 91, + 447, + 100 + ], + "type": "text", + "content": "BrushNet" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 134, + 137, + 139, + 144 + ], + "type": "aside_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 137, + 139, + 144 + ], + "spans": [ + { + "bbox": [ + 134, + 137, + 139, + 144 + ], + "type": "text", + "content": "I" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 177, + 139, + 185 + ], + "type": "aside_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 177, + 139, + 185 + ], + "spans": [ + { + "bbox": [ + 134, + 177, + 139, + 185 + ], + "type": "text", + "content": "I" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": "dual-branch design as our model design. Tab. 5 presents the rationale behind the design choices for (1) using a VAE encoder instead of randomly initialized convolution layers to process the masked image, (2) incorporating the full UNet feature layer-by-layer into the pre-trained UNet, (3) removing text cross-attention in BrushNet, which avoids masked image features influenced by text." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 178, + 482, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 178, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 482, + 266 + ], + "type": "text", + "content": "Table 5: Ablation on model architecture. We ablate on the following components: the image encoder (Enc), selected from a random initialized convolution (Conv) and a VAE; the inclusion of mask in input (Mask), chosen from adding (w/) and not adding (w/o); the presence of cross-attention layers (Attn), chosen from adding (w/) and not adding (w/o); the type of UNet feature addition (UNet), selected from adding the full UNet feature (full), adding half of the UNet feature (half), and adding the feature like ControlNet (CN); and finally, the blending operation (Blend), chosen from not adding (w/o), direct pasting (paste), and blurred blending (blur). Red is the best result." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 140, + 271, + 469, + 377 + ], + "blocks": [ + { + "bbox": [ + 140, + 271, + 469, + 377 + ], + "lines": [ + { + "bbox": [ + 140, + 271, + 469, + 377 + ], + "spans": [ + { + "bbox": [ + 140, + 271, + 469, + 377 + ], + "type": "table", + "html": "
MetricsImage QualityMasked Region PreservationText Align
EncMaskAttnUNetBlendIR×10↑HPS×102↑AS↑PSNR↑MSE×102↓LPIPS×103↓CLIP Sim↑
Convw/w/ofullw/o11.0526.236.5514.8937.2364.5426.76
VAEw/ow/ofullw/o11.5527.706.5717.9626.3849.3326.87
VAEw/w/fullw/o11.2527.626.5618.6919.4434.2826.63
Convw/w/CNw/o9.5826.856.4712.1580.91150.8926.88
VAEw/w/CNw/o10.5327.426.5918.2824.3641.6326.89
VAEw/w/oCNw/o11.4227.696.5818.4924.0936.3326.86
VAEw/w/ohalfw/o11.4727.706.5719.0123.7733.5726.87
VAEw/w/ofullw/o11.5927.716.5919.8616.0931.6826.91
VAEw/w/ofullpaste11.7227.936.58---26.80
VAEw/w/ofullblur11.7627.946.5829.881.5311.6526.81
", + "image_path": "dbfb03d1d183d84c9b643cea81c48652c2797dcdc8cf7adc3b624156db4bf959.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 385, + 217, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 217, + 398 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 217, + 398 + ], + "type": "text", + "content": "6 Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 400, + 479, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 479, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 479, + 460 + ], + "type": "text", + "content": "Conclusion. This paper proposes a plug-and-play image inpainting method BrushNet with a pixel-level masked image feature insertion architectural design. Quantitative and qualitative results on our proposed benchmark, BrushBench, and EditBench show the superior performance of BrushNet considering image generation quality, masked region preservation, and image-text alignment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 463, + 479, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 479, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 479, + 534 + ], + "type": "text", + "content": "Limitations and Future Work. However, BrushNet still has some limitations: (1) The quality and content generated by our model are heavily dependent on the chosen base model. (2) Even with BrushNet, we still observe poor generation results in cases where the given mask has an unusually shaped or irregular form, or when the given text does not align well with the masked image. In our future work, we will continue to address these challenges." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 537, + 479, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 537, + 479, + 620 + ], + "spans": [ + { + "bbox": [ + 130, + 537, + 479, + 620 + ], + "type": "text", + "content": "Negative Social Impact. Image inpainting models present exciting opportunities for content creation, but they also carry potential risks to individuals and society. Their reliance on internet-collected training data can amplify social biases, and there is a specific risk of generating persuasive misinformation by manipulating human images with offensive elements. To address these concerns, it is crucial to emphasize responsible use and establish ethical guidelines when utilizing these models. This is also a key focus for our future model releases." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 629, + 228, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 228, + 641 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 228, + 641 + ], + "type": "text", + "content": "*Acknowledgment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 641, + 485, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 485, + 653 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 485, + 653 + ], + "type": "text", + "content": "This work is supported in part by Research Matching Grant(CSE-7-2022)-RMG01." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "Xuan Ju, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "text", + "content": "1. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. ACM transactions on graphics (TOG) 42(4), 1-11 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 159, + 480, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 159, + 480, + 190 + ], + "spans": [ + { + "bbox": [ + 138, + 159, + 480, + 190 + ], + "type": "text", + "content": "2. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18208-18218 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 191, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 191, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 191, + 480, + 224 + ], + "type": "text", + "content": "3. Bertalmio, M., Sapiro, G., Caseles, V., Ballester, C.: Image inpainting. In: International Conference and Exhibition on Computer Graphics and Interactive Techniques (SIGGRAPH). pp. 417-424 (2000)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 224, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 480, + 256 + ], + "type": "text", + "content": "4. Binghui, C., Chao, L., Chongyang, Z., Wangmeng, X., Yifeng, G., Xuansong, X.: Replaceanything as you want: Ultra-high quality content replacement (2023), https://aigcdesigngroup.github.io/replace-anything/" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 257, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 257, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 257, + 480, + 277 + ], + "type": "text", + "content": "5. Binkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying MMD GANs. arXiv preprint arXiv:1801.01401 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 277, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 481, + 310 + ], + "type": "text", + "content": "6. Corneanu, C., Gadde, R., Martinez, A.M.: Latentpaint: Image inpainting in latent space with diffusion models. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 4334-4343 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 310, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 310, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 310, + 480, + 342 + ], + "type": "text", + "content": "7. Criminisi, A., Pérez, P., Toyama, K.: Region filling and object removal by exemplar-based image inpainting. IEEE Transactions on Image Processing 13(9), 1200-1212 (2004)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "type": "text", + "content": "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 374, + 480, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 480, + 395 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 480, + 395 + ], + "type": "text", + "content": "9. epinikion: epicrealism (2023), https://civitai.com/models/25694?modelVersionId=143906" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 396, + 480, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 480, + 415 + ], + "type": "text", + "content": "0. hemi29833: Henmixreal (2024), https://civitai.com/models/20282?modelVersionId=305687" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 417, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 417, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 417, + 480, + 449 + ], + "type": "text", + "content": "1. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs trained by a two time-scale update rule converge to a local Nash equilibrium. Advances in Neural Information Processing Systems (NIPS) 30 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 449, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 480, + 471 + ], + "type": "text", + "content": "2. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NIPS) 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 472, + 480, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 480, + 491 + ], + "type": "text", + "content": "3. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 492, + 480, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 492, + 480, + 524 + ], + "spans": [ + { + "bbox": [ + 138, + 492, + 480, + 524 + ], + "type": "text", + "content": "4. Huang, H., He, R., Sun, Z., Tan, T., et al.: Introvae: Introspective variational autoencoders for photographic image synthesis. Advances in Neural Information Processing Systems (NIPS) 31 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 525, + 480, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 556 + ], + "type": "text", + "content": "5. Huang, Y., Huang, J., Liu, Y., Yan, M., Lv, J., Liu, J., Xiong, W., Zhang, H., Chen, S., Cao, L.: Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525 (2024)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 557, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 557, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 557, + 480, + 589 + ], + "type": "text", + "content": "6. Jayasumana, S., Ramalingam, S., Veit, A., Glasner, D., Chakrabarti, A., Kumar, S.: Rethinking fid: Towards a better evaluation metric for image generation. arXiv preprint arXiv:2401.09603 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "type": "text", + "content": "7. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision (IJCV) 128(7), 1956-1981 (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "8. Li, Z., Wei, P., Yin, X., Ma, Z., Kot, A.C.: Virtual try-on with pose-garment keypoints guided inpainting. In: IEEE/CVF International Conference on Computer Vision (ICCV). pp. 22788-22797 (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "BrushNet" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "19. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision (ECCV). pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "type": "text", + "content": "20. Liu, A., Niepert, M., Broeck, G.V.d.: Image inpainting via tractable steering of diffusion models. arXiv preprint arXiv:2401.03349 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "21. Liu, H., Wan, Z., Huang, W., Song, Y., Han, X., Liao, J.: Pd-GAN: Probabilistic diverse GAN for image inpainting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9371-9381 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 236 + ], + "type": "text", + "content": "22. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: IEEE/CVF International Conference on Computer Vision (ICCV) (December 2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 281 + ], + "type": "text", + "content": "23. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11461-11471 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "type": "text", + "content": "24. Lykon: Dreamshaper (2022), https://civitai.com/models/4384?modelVersionId=128713" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "text", + "content": "25. Manukyan, H., Sargsyan, A., Atanyan, B., Wang, Z., Navasardyan, S., Shi, H.: Hd-painter: High-resolution and prompt-faithful text-guided image inpainting with diffusion models. arXiv preprint arXiv:2312.14091 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "text", + "content": "26. Meina: Meinamix (2023), https://civitai.com/models/7240?modelVersionId=119057" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "text", + "content": "27. Peng, J., Liu, D., Xu, S., Li, H.: Generating diverse structure for image inpainting with hierarchical vq-vae. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10775-10784 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "text", + "content": "28. von Platen, P., Patil, S., Lozhkov, A., Cuenca, P., Lambert, N., Rasul, K., Davaadorj, M., Wolf, T.: Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "text", + "content": "29. Quan, W., Chen, J., Liu, Y., Yan, D.M., Wonka, P.: Deep learning-based image and video inpainting: A survey. International Journal of Computer Vision (IJCV) pp. 1-34 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 500 + ], + "type": "text", + "content": "30. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (ICML). pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 501, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 544 + ], + "type": "text", + "content": "31. Razzhigaev, A., Shakhmatov, A., Maltseva, A., Arkhipkin, V., Pavlov, I., Ryabov, I., Kuts, A., Panchenko, A., Kuznetsov, A., Dimitrov, D.: Kandinsky: an improved text-to-image synthesis with image prior and latent diffusion. arXiv preprint arXiv:2310.03502 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 544, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 577 + ], + "type": "text", + "content": "32. Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., et al.: Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159 (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 577, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 481, + 620 + ], + "type": "text", + "content": "33. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "34. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems (NIPS) 35, 25278–25294 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "Xuan Ju, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "35. SG161222: Realisticvision (2023), https://civitai.com/models/4201?modelVersionId=130072" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "type": "text", + "content": "36. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 163, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 163, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 163, + 480, + 217 + ], + "type": "text", + "content": "37. Wang, S., Sahara, C., Montgomery, C., Pont-Tuset, J., Noy, S., Pellegrini, S., Onoe, Y., Laszlo, S., Fleet, D.J., Soricut, R., et al.: Imagen editor and editbench: Advancing and evaluating text-guided image inpainting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18359-18369 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 480, + 251 + ], + "type": "text", + "content": "38. Wikipedia contributors: Mean squared error — Wikipedia, the free encyclopedia (2024), https://en.wikipedia.org/w/index.php?title=Mean_squared_error&oldid=1207422018, [Online; accessed 4-March-2024]" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 480, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 480, + 283 + ], + "type": "text", + "content": "39. Wikipedia contributors: Peak signal-to-noise ratio — Wikipedia, the free encyclopedia (2024), https://en.wikipedia.org/w/index.php?title=Peak_signal-to-noise_ratio&oldid=1210897995, [Online; accessed 4-March-2024]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 285, + 480, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 317 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 317 + ], + "type": "text", + "content": "40. Wu, C., Huang, L., Zhang, Q., Li, B., Ji, L., Yang, F., Sapiro, G., Duan, N.: GO-DIVA: Generating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 319, + 480, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 480, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 480, + 351 + ], + "type": "text", + "content": "41. Wu, X., Hao, Y., Sun, K., Chen, Y., Zhu, F., Zhao, R., Li, H.: Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 353, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 480, + 396 + ], + "type": "text", + "content": "42. Xie, S., Zhang, Z., Lin, Z., Hinz, T., Zhang, K.: Smartbrush: Text and shape guided object inpainting with diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22428-22437 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 397, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 429 + ], + "type": "text", + "content": "43. Xie, S., Zhao, Y., Xiao, Z., Chan, K.C., Li, Y., Xu, Y., Zhang, K., Hou, T.: Dreaminpainter: Text-guided subject-driven image inpainting with diffusion models. arXiv preprint arXiv:2312.03771 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 430, + 480, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 480, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 480, + 462 + ], + "type": "text", + "content": "44. Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagereward: Learning and evaluating human preferences for text-to-image generation (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 464, + 480, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 480, + 485 + ], + "type": "text", + "content": "45. Xu, Z., Zhang, X., Chen, W., Yao, M., Liu, J., Xu, T., Wang, Z.: A review of image inpainting methods based on deep learning. Applied Sciences 13(20), 11189 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 487, + 480, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 480, + 518 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 480, + 518 + ], + "type": "text", + "content": "46. Yang, S., Chen, X., Liao, J.: Uni-paint: A unified framework for multimodal image inpainting with pretrained diffusion model. In: ACM International Conference on Multimedia (MM). pp. 3190-3199 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 520, + 480, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 480, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 480, + 552 + ], + "type": "text", + "content": "47. Yang, S., Zhang, L., Ma, L., Liu, Y., Fu, J., He, Y.: Magicremover: Tuning-free text-guided image inpainting with diffusion models. arXiv preprint arXiv:2310.02848 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 554, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 480, + 586 + ], + "type": "text", + "content": "48. Yu, F., Seff, A., Zhang, Y., Song, S., Funkhouser, T., Xiao, J.: Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 587, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 480, + 609 + ], + "type": "text", + "content": "49. Yu, T., Feng, R., Feng, R., Liu, J., Jin, X., Zeng, W., Chen, Z.: Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 610, + 480, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 480, + 631 + ], + "type": "text", + "content": "50. Zhang, G., Ji, J., Zhang, Y., Yu, M., Jaakkola, T., Chang, S.: Towards coherent image inpainting using denoising diffusion implicit models (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "51. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 447, + 100 + ], + "type": "text", + "content": "BrushNet" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 304 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "type": "text", + "content": "52. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 586-595 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "text", + "content": "53. Zhao, S., Cui, J., Sheng, Y., Dong, Y., Liang, X., Chang, E.I., Xu, Y.: Large scale image completion via co-modulated generative adversarial networks. arXiv preprint arXiv:2103.10428 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "54. Zheng, C., Cham, T.J., Cai, J.: Pluralistic image completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1438-1447 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 226, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 226, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 226, + 481, + 270 + ], + "type": "text", + "content": "55. Zheng, H., Lin, Z., Lu, J., Cohen, S., Shechtman, E., Barnes, C., Zhang, J., Xu, N., Amirghodsi, S., Luo, J.: Image inpainting with cascaded modulation GAN and object-aware training. In: European Conference on Computer Vision (ECCV). pp. 277-296. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 270, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 304 + ], + "type": "text", + "content": "56. Zhuang, J., Zeng, Y., Liu, W., Yuan, C., Chen, K.: A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. arXiv preprint arXiv:2312.03594 (2023)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "Xuan Ju, et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Bucketed Ranking-based Losses for Efficient Training of Object Detectors/7957777c-4d0a-4feb-89c6-b57e75289486_content_list.json b/2024/Bucketed Ranking-based Losses for Efficient Training of Object Detectors/7957777c-4d0a-4feb-89c6-b57e75289486_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..318d19031f5f53589fb7d261513de3bffad4f9d2 --- /dev/null +++ b/2024/Bucketed Ranking-based Losses for Efficient Training of Object Detectors/7957777c-4d0a-4feb-89c6-b57e75289486_content_list.json @@ -0,0 +1,1874 @@ +[ + { + "type": "text", + "text": "Bucketed Ranking-based Losses for Efficient Training of Object Detectors", + "text_level": 1, + "bbox": [ + 285, + 140, + 715, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Feyza Yavuz1, Baris Can Cam1, Adnan Harun Dogan1, Kemal Oksuz2, Emre Akbas1,3,*, and Sinan Kalkan1,3,*", + "bbox": [ + 223, + 210, + 779, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Computer Engineering, METU, Ankara, Türkiye $^{2}$ Five AI Ltd., United Kingdom", + "bbox": [ + 279, + 253, + 720, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 METU ROMER Robotics Center, Ankara, Türkiye * Equal senior authorship", + "bbox": [ + 323, + 282, + 678, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Ranking-based loss functions, such as Average Precision Loss and Rank&Sort Loss, outperform widely used score-based losses in object detection. These loss functions better align with the evaluation criteria, have fewer hyperparameters, and offer robustness against the imbalance between positive and negative classes. However, they require pairwise comparisons among $P$ positive and $N$ negative predictions, introducing a time complexity of $\\mathcal{O}(PN)$ , which is prohibitive since $N$ is often large (e.g., $10^8$ in ATSS). Despite their advantages, the widespread adoption of ranking-based losses has been hindered by their high time and space complexities. In this paper, we focus on improving the efficiency of ranking-based loss functions. To this end, we propose Bucketed Ranking-based Losses which group negative predictions into $B$ buckets ( $B \\ll N$ ) in order to reduce the number of pairwise comparisons so that time complexity can be reduced. Our method enhances the time complexity, reducing it to $\\mathcal{O}(\\max(N\\log(N),P^2))$ . To validate our method and show its generality, we conducted experiments on 2 different tasks, 3 different datasets, 7 different detectors. We show that Bucketed Ranking-based (BR) Losses yield the same accuracy with the unbucketed versions and provide $2 \\times$ faster training on average. We also train, for the first time, transformer-based object detectors using ranking-based losses, thanks to the efficiency of our BR. When we train CoDETR, a state-of-the-art transformer-based object detector, using our BR Loss, we consistently outperform its original results over several different backbones. Code will be released.", + "bbox": [ + 261, + 349, + 743, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Object Detection $\\cdot$ Ranking Losses $\\cdot$ Efficient Ranking Losses", + "bbox": [ + 215, + 707, + 751, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 747, + 375, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Training an object detector incurs many challenges unique to object detection, including (i) assigning dynamically a large number of object hypotheses with the objects [2, 12, 13, 25, 29, 33, 34], (ii) sampling among these hypotheses to ensure that the background class does not dominate training [19, 27, 29, 32] and", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c805ee1cbcd735452d8b8b83680e27b7722fce63da5db9af6fe80d48307ebeb3.jpg", + "image_caption": [ + "(a) Comparisons in AP Loss" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 519, + 186 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7a617a6d94d316ca388a4502024453c37d0122189af09f0f7a40290b7354215b.jpg", + "image_caption": [ + "(b) Comparisons in Bucketed AP Loss (ours)" + ], + "image_footnote": [], + "bbox": [ + 217, + 222, + 519, + 271 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6c56eb8e3f6cbc738baa91872c80d850eb019b86c314cc556fcb9af123a3a762.jpg", + "image_caption": [ + "(c) Average Precision (AP) vs. average training iteration time for baselines, RS Loss and Bucketed RS Loss (ours)", + "Fig. 1: (a) Existing ranking-based losses (i.e., AP Loss [5,6]) incur significant overhead owing to pairwise comparisons between positives and negatives. (b) We propose bucketing negatives to decrease the number of comparisons, and hence the complexity. Under certain assumptions, our bucketing approach provides equal gradients with conventional ranking-based losses such as AP Loss in (a). (c) Accuracy and efficiency comparison. Our BRS formulation facilitates faster (between $1.9 \\times$ and $6.0 \\times$ ) training of visual detectors with similar AP." + ], + "image_footnote": [], + "bbox": [ + 526, + 143, + 785, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(iii) minimizing a multi-task objective function [17,23,24,29]. While the choices on assignment and sampling generally vary depending on the trained object detector, it is very common to combine Cross Entropy or Focal Loss [17] with a regression loss [30,36] as the multi-task loss function. Recently proposed ranking-based loss functions [5,6,23,24] offer an alternative approach for addressing these challenges by formulating the training objective based on the rank of positive examples over negative examples.", + "bbox": [ + 212, + 448, + 787, + 554 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Benefits of ranking-based losses. First, they are inherently robust to imbalance [23] and hence, do not require any sampling mechanism under very challenging scenarios [24], e.g., even when the background foreground ratio is 10K for LVIS [9]. Second, they are shown to generalize over different detectors with diverse architectures - with the exception of transformer-based ones [14,37,38], since ranking-based losses further slow down the training of transformer-based detectors, which we address in this paper. Such losses also offer significant performance gain over their score-based counterparts, and having less hyperparameters, they are much easier to tune [23,24].", + "bbox": [ + 212, + 561, + 787, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The drawback of ranking-based losses. Compared to score-based losses, ranking-based losses are more inefficient as the ranking operation inherently requires each pair of object hypotheses to be compared against each other (Fig. 1(a)), inducing a quadratic complexity on the large number of object hypotheses (e.g., $10^{8}$ for ATSS [34] on COCO [18]). As a result, vectorized implementations for parallel GPU processing are infeasible as such large matrices (e.g., with $\\sim$ $10^{16}$ entries for ATSS) do not fit into GPU memories. This has driven researchers towards alternative ways of computing these matrices, which, in the end, saves from the storage complexity but results in more inefficient algorithms [5,6].", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Yavuz et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach. The main cause for the aforementioned complexity is the background class, i.e., negatives, forming up to $99.9\\%$ of the hypotheses [26]. Considering this, we introduce a novel bucketing approach (Fig. 1(b)) on these negatives to address the main shortcoming of the ranking-based loss functions. Specifically, we first sort all examples based on their scores and bucket the negatives located between successive positives. The negatives within the same bucket are treated as a single example and represented by their average score. We then incorporate our bucketing approach into AP Loss [5,6] and RS Loss [24], and introduce novel mechanisms for calculating the gradients correctly.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The significance of our approach. Theoretically, we prove that our bucketing approach provides the same gradients as the original ranking-based loss. Practically, we show we can reduce loss computation time by up to $\\sim 40\\times$ and training time of a detector by up to $6\\times$ (Fig. 1(c)), addressing the main drawback of ranking-based loss functions. Hence, we show that the gap between the score-based and ranking-based loss functions in terms of training time vanishes, enabling us to train strong transformer-based object detectors such as Co-DETR using our loss functions. As ranking-based loss functions now surpass score-based ones or perform on par in every aspect such as accuracy, tuning simplicity, robustness to imbalance and training time, we believe that our work will make their prevalence more dominant in the upcoming object detectors.", + "bbox": [ + 212, + 282, + 787, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contributions. Our contributions can be summarized as follows:", + "bbox": [ + 215, + 449, + 691, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We propose a novel bucketing approach to improve the efficiency of computationally-expensive ranking-based losses to train object detectors. Theoretically, our approach yields the same gradients with the original ranking-based losses while decreasing their time complexity. Practically, we enable up to $6 \\times$ faster training using our bucketing approach with no accuracy loss.", + "bbox": [ + 225, + 470, + 785, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- For the first time, we incorporate ranking-based loss functions to transformer-based detectors. Specifically, we construct BRS-DETR by replacing the training objective of the state-of-the-art transformer-based object detector CoDETR [38] by our Bucketed RS (BRS) Loss.", + "bbox": [ + 225, + 547, + 785, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Our comprehensive experiments on detection and instance segmentation on 3 different challenging datasets, 5 backbones and 7 detectors show the effectiveness and generalizability of our approach. Our BRS-DETR reaches 50.1 AP on COCO val set with only 12 epochs and ResNet-50 backbone, outperforming all existing detectors using the same backbone and 300 queries. BRS-DETR also outperforms CoDETR with Swin-T and Swin-L backbones, reaching 57.2 AP on the COCO dataset.", + "bbox": [ + 225, + 607, + 785, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 734, + 387, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although score-based loss functions, such as cross-entropy and focal loss, are widely used to train both CNN-based [2, 17, 35] and transformer-based object detectors [14, 37, 38], ranking-based losses for visual detection offer advantages. The primary advantage is their robustness to class imbalance [23]. They yield on par or better performance compared to score-based losses, without the need", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Bucketed Ranking-based Losses", + "bbox": [ + 519, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for extensive tuning of class weights or multistage sampling techniques. Another advantage they offer is the ease of balancing losses among different tasks. In AP Loss [5, 6], aLRP Loss [23] and RS Loss [24], the authors use the ratio of classification and regression losses to balance these tasks. As a result, ranking-based losses have much fewer hyperparameters during training, compared to score-based ones.", + "bbox": [ + 212, + 146, + 787, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Other Ranking-based Losses. The literature has explored other ranking-based loss functions with different goals and assumptions. For example, DR Loss [28] constrains a margin to be satisfied between positives and negatives, not taking the recall into account. Smooth AP [1] introduces a smooth, differentiable approximation of AP, with the assumption that $|\\mathcal{N}|$ is not large. Several studies focused on improving the efficiency of ranking-based losses for use in Support Vector Machines (SVMs) [21, 22]. However, these solutions have been limited to training linear SVMs and hence, problems that can be solved using linear functions. In this paper, we focus on improving ranking-based loss functions [5, 6, 24] that can be used for training complex deep networks such as object detectors.", + "bbox": [ + 212, + 238, + 787, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Comparative Summary. Ranking-based training objective for visual detectors proved their superiority over score-based loss functions in visual detection tasks [5, 6, 23, 24]. However, due to the requirement of pairwise ranking their time complexity is higher than the score-based alternatives. Such a disadvantage of ranking-based loss functions prevents their applicability to larger visual detectors (i.e., transformer-based object detectors).", + "bbox": [ + 214, + 406, + 789, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Background on Ranking-based Losses", + "text_level": 1, + "bbox": [ + 215, + 527, + 614, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ranking-based losses for object detection rely on pairwise comparisons between the scores of different detections to determine the rank of a detection among positives and negatives. Denoting the score of the $i^{th}$ detection by $s_i$ , we can compare the scores of two detections $i$ and $j$ with a simple difference transform: $x_{ij} = s_j - s_i$ . By counting the number of instances with $x_{ij} > 0$ , we can determine $i^{th}$ detection's rank among positives $(\\mathcal{P})$ and all detections (positives and negatives: $\\mathcal{P} \\cup \\mathcal{N}$ ) as follows:", + "bbox": [ + 214, + 569, + 787, + 676 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {r a n k} ^ {+} (i) = \\sum_ {j \\in \\mathcal {P}} \\bar {H} \\left(x _ {i j}\\right), \\text {a n d} \\operatorname {r a n k} (i) = \\sum_ {j \\in \\mathcal {P} \\cup \\mathcal {N}} \\bar {H} \\left(x _ {i j}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 691, + 787, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{H}(x)$ is one if $x > 0$ and zero otherwise. Since $\\mathrm{d}\\bar{H}(x) / \\mathrm{d}x$ is either infinite (at $x = 0$ ) or zero (for $x \\neq 0$ ), a smoothed version is used ( $\\delta$ : a hyper-parameter):", + "bbox": [ + 215, + 737, + 787, + 771 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nH (x) = \\left\\{ \\begin{array}{l l} 0, & x < - \\delta \\\\ \\frac {x}{2 \\delta} + 0. 5, & - \\delta \\leq x \\leq \\delta \\\\ 1, & \\delta < x \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 785, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Yavuz et al.", + "bbox": [ + 271, + 114, + 352, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Revisiting Average Precision (AP) Loss", + "text_level": 1, + "bbox": [ + 215, + 146, + 586, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the definitions in Eq. 1, AP Loss [5,6] can be defined as:", + "bbox": [ + 215, + 172, + 671, + 186 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {A P} = 1 - \\mathrm {A P} = 1 - \\frac {1}{| \\mathcal {P} |} \\sum_ {i \\in \\mathcal {P}} \\operatorname {p r e c i s i o n} (i) = 1 - \\frac {1}{| \\mathcal {P} |} \\sum_ {i \\in \\mathcal {P}} \\frac {\\operatorname {r a n k} ^ {+} (i)}{\\operatorname {r a n k} (i)}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 285, + 199, + 785, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Chen et al. [5,6] simplified Eq. 3 by rewriting it in terms of pairwise comparisons between positive and negative detections:", + "bbox": [ + 215, + 241, + 782, + 272 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {A P} = \\frac {1}{| \\mathcal {P} |} \\sum_ {i \\in \\mathcal {P}} \\sum_ {j \\in \\mathcal {N}} \\frac {H (x _ {i j})}{\\operatorname {r a n k} (i)} = \\frac {1}{| \\mathcal {P} |} \\sum_ {i \\in \\mathcal {P}} \\sum_ {j \\in \\mathcal {N}} L _ {i j} ^ {A P}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 282, + 785, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L_{ij}^{AP} = H(x_{ij}) / \\mathrm{rank}(i)$ is called the primary term. Note that $L_{ij}^{AP}$ is non-differentiable since the step function $(H(\\cdot))$ applied on $x_{ij}$ is non-differentiable, which will be discussed next.", + "bbox": [ + 215, + 325, + 782, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Revisiting Identity Update", + "text_level": 1, + "bbox": [ + 215, + 395, + 483, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Oksuz et al. [24] showed that various ranking-based losses (including AP Loss in Eqs. 3 & 4) can be written in a general form as:", + "bbox": [ + 215, + 421, + 782, + 450 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\frac {1}{Z} \\sum_ {i \\in \\mathcal {P} \\cup \\mathcal {N}} (\\ell (i) - \\ell^ {*} (i)), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 464, + 785, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\ell(i)$ is the ranking-based error (e.g., precision error) computed on the $i$ th detection, $\\ell^{*}(i)$ is the target ranking-based error (the lowest error possible) and $Z$ is the normalization constant. The loss in Eq. 5 can be computed and optimized as follows [24]:", + "bbox": [ + 215, + 510, + 782, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1. Computation of the Loss. First, each pair of logits $(s_i$ and $s_j)$ are compared by calculating their difference transforms as $x_{ij} = s_j - s_i$ . With the step function $H(\\cdot)$ (Eq. 2), the number of detections higher than $s_i$ (and therefore its precision error, $\\ell(i)$ ) can be easily calculated (see Sect. 3.1). Given $x_{ij}$ , the loss can be rewritten and calculated in terms of primary terms $L_{ij}$ as $\\mathcal{L} = \\frac{1}{Z}\\sum_{i\\in \\mathcal{P}\\cup \\mathcal{N}}\\sum_{j\\in \\mathcal{P}\\cup \\mathcal{N}}L_{ij}$ , by taking:", + "bbox": [ + 215, + 571, + 784, + 670 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {i j} = \\left(\\ell (i) - \\ell^ {*} (i)\\right) p (j | i), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 685, + 784, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p(j|i)$ is a probability mass function (pmf) that distributes the error computed on the $i$ th example over the $j$ th example in order to determine the pairwise primary term $L_{ij}$ . $p(j|i)$ is commonly taken as a uniform distribution [23, 24].", + "bbox": [ + 215, + 710, + 782, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2. Optimization of the Loss. The gradient of the primary term wrt. to the difference transform $(\\partial L_{ij} / \\partial x_{ij})$ is non-differentiable. Denoting this term by $\\Delta x_{ij}$ , we have [5]:", + "bbox": [ + 215, + 757, + 782, + 803 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\mathcal {L}}{\\partial s _ {i}} = \\sum_ {j, k} \\frac {\\partial \\mathcal {L}}{\\partial L _ {j k}} \\Delta x _ {j k} \\frac {\\partial x _ {j k}}{\\partial s _ {i}} = \\frac {1}{Z} \\left(\\sum_ {j} \\Delta x _ {j i} - \\sum_ {j} \\Delta x _ {i j}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 815, + 784, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Bucketed Ranking-based Losses", + "bbox": [ + 517, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, optimizing a ranking-based loss function reduces to determining $\\Delta x_{ij}$ . Inspired by Chen et al. [5], Oksuz et al. [24] employ Perceptron Learning [31] and show that $\\Delta x_{ij}$ in Eq. 7 is simply the primary term itself: $\\Delta x_{ij} = -(L_{ij}^{*} - L_{ij}) = -(0 - L_{ij}) = L_{ij}$ , hence the name Identity Update. Plugging this into Eq. 7 yields (see Oksuz et al. [24] for the steps of the derivation):", + "bbox": [ + 215, + 146, + 787, + 223 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\mathcal {L}}{\\partial s _ {i}} = \\frac {1}{Z} \\left(\\sum_ {j} L _ {j i} - \\sum_ {j} L _ {i j}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 234, + 785, + 262 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Therefore, both computing and optimizing the loss reduces determining the primary term $L_{ij}$ .", + "bbox": [ + 215, + 272, + 785, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Revisiting AP Loss and RS Loss with Identity Update", + "text_level": 1, + "bbox": [ + 215, + 325, + 710, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "AP Loss [5, 6]. For defining AP Loss with identity update, the errors in Eq. 5 can be derived as $\\ell_{\\mathrm{R}}(i) = \\frac{N_{\\mathrm{FP}}(i)}{\\mathrm{rank}(i)}, \\ell_{\\mathrm{R}}^{*}(i) = 0$ and $Z = |\\mathcal{P}|$ . Furthermore, defining $p_R(j|i)$ as a uniform pmf, that is $p_R(j|i) = \\frac{H(x_{ij})}{N_{\\mathrm{FP}}(i)}$ , we can obtain the primary terms of AP Loss using Identity Update, completing the derivation for computation and optimization (Eq. 8):", + "bbox": [ + 214, + 349, + 787, + 436 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {i j} ^ {A P} = \\left\\{ \\begin{array}{l l} \\left(\\ell_ {\\mathrm {R}} (i) - \\ell_ {\\mathrm {R}} ^ {*} (i)\\right) p _ {R} (j | i), & \\text {f o r} i \\in \\mathcal {P}, j \\in \\mathcal {N} \\\\ 0, & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 351, + 448, + 785, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "RS Loss [24]. RS Loss includes an additional sorting objective $(\\ell_S(i))$ to promote better-localized positives to be ranked higher than other positives:", + "bbox": [ + 215, + 484, + 785, + 516 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {R S}} = \\frac {1}{| \\mathcal {P} |} \\sum_ {i \\in \\mathcal {P}} \\left[ \\underbrace {\\left(\\ell_ {\\mathrm {R}} (i) - \\ell_ {\\mathrm {R}} ^ {*} (i)\\right)} _ {\\text {r a n k i n g}} + \\underbrace {\\left(\\ell_ {\\mathrm {S}} (i) - \\ell_ {\\mathrm {S}} ^ {*} (i)\\right)} _ {\\text {s o r t i n g}} \\right]. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 351, + 526, + 785, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The primary terms of RS Loss can be obtained in a similar manner to AP Loss (see the Supp.Mat. for derivations), yielding:", + "bbox": [ + 215, + 577, + 785, + 608 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {i j} ^ {R S} = \\left\\{ \\begin{array}{l l} \\left(\\ell_ {\\mathrm {R}} (i) - \\ell_ {\\mathrm {R}} ^ {*} (i)\\right) p _ {R} (j | i), & \\text {f o r} i \\in \\mathcal {P}, j \\in \\mathcal {N} \\\\ \\left(\\ell_ {\\mathrm {S}} (i) - \\ell_ {\\mathrm {S}} ^ {*} (i)\\right) p _ {S} (j | i), & \\text {f o r} i \\in \\mathcal {P}, j \\in \\mathcal {P}, \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 619, + 785, + 659 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $p_S(j|i)$ is the sorting pmf. Note that RS Loss has the same primary terms with AP Loss for $i \\in \\mathcal{P}, j \\in \\mathcal{N}$ . The primary terms for $i \\in \\mathcal{P}, j \\in \\mathcal{P}$ directly targets promoting the better-localized positives.", + "bbox": [ + 215, + 670, + 785, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 Complexity of Ranking-based Loss Functions", + "text_level": 1, + "bbox": [ + 215, + 739, + 633, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "These loss functions originally have the space complexity of $\\mathcal{O}((|\\mathcal{P}| + |\\mathcal{N}|)^2)$ [5,6] as they need to compare each pair. This quadratic space complexity makes it infeasible to compute these loss functions using modern GPUs as the number of logits in object detection is very large. To alleviate that, Chen et al. [5] introduced two tricks at the expense of making the computation inefficient: (1) Loop", + "bbox": [ + 214, + 763, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Yavuz et al.", + "bbox": [ + 271, + 114, + 352, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "on Positives: By implementing a loop over the positive examples, Chen et al. managed to reduce the time complexity to $\\mathcal{O}(|\\mathcal{P}|(|\\mathcal{P}| + |\\mathcal{N}|)) \\approx \\mathcal{O}(|\\mathcal{P}||\\mathcal{N}|)$ since $\\mathcal{N} \\gg \\mathcal{P}$ and the space complexity to $\\mathcal{O}(|\\mathcal{P}| + |\\mathcal{N}|)$ . (2) Discard Trivial Negatives: AP Loss considers negative examples with lower rank than all positives as trivial and disregards them.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Bucketed Ranking-based Loss Functions for Efficient Optimisation of Object Detectors", + "text_level": 1, + "bbox": [ + 214, + 244, + 764, + 280 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training object detectors with existing ranking losses has several advantages as we presented before, however, they suffer from significantly large training time. We now address this by our Bucketing Approach, and propose our Bucketed AP and RS Losses.", + "bbox": [ + 212, + 294, + 787, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Bucketing Negatives for Efficient Optimisation", + "text_level": 1, + "bbox": [ + 214, + 377, + 645, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our main intuition is to bucket the sequential negative examples considering that they will have very similar or equal gradients. Fig. 1(a) demonstrates this phenomenon in which $n_1, n_2$ both are assigned equal gradients as well as each negative among $n_3, n_4, n_5, n_6$ in the standard AP Loss. Formally, we sort all given logits $s_i \\in \\mathcal{P} \\cup \\mathcal{N}$ using a conventional sorting methods. Let us denote this sorted permutation of the logits by $\\hat{s}_1, \\hat{s}_2, \\dots, \\hat{s}_{|\\mathcal{P} \\cup \\mathcal{N}|}$ , i.e., $\\hat{s}_1 > \\hat{s}_2 > \\dots > \\hat{s}_{|S|}$ . Given these sorted logits and denoting the $i$ th positive logit in the ordering by $\\hat{s}_i^+$ , the buckets of negatives $B_1, \\dots, B_{|\\mathcal{P} + 1|}$ can be obtained by:", + "bbox": [ + 212, + 402, + 787, + 525 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nB _ {1} > \\hat {s} _ {1} ^ {+} > B _ {2} > \\hat {s} _ {2} ^ {+} > \\dots > \\hat {s} _ {| \\mathcal {P} |} ^ {+} > B _ {| \\mathcal {P} + 1 |}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 536, + 784, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We will denote the size of the $i$ th negative bucket by $b_{i}$ . Having created the buckets, we now determine a single logit value for each bucket, which we call as the prototype logit as it is necessary while assigning the ranking error. We denote the prototype logit for the $i^{\\mathrm{th}}$ bucket by $s_i^b$ . Note that if $\\delta = 0$ in Eq. 2, then any logit satisfying $\\hat{s}_{i - 1}^{+} > s_{i}^{b} > \\hat{s}_{i}^{+}$ can be a prototype logit as the ranking stays the same. However, in the case of $\\delta > 0$ , the boundaries between the logits are smoothed. That's why, practically, we find it effective to use the mean logit of a bucket as its prototype logit.", + "bbox": [ + 212, + 565, + 787, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that this bucketing approach reduces the number of logits (positive and prototype negative) to a maximum of $2|\\mathcal{P}| + 1$ . As a result, the pairwise errors can now fit into the memory. Consequently, the loop in the red box of Alg. S1 (in Suppl. Mat.) is no longer necessary, which gives rise to efficient ranking-based loss functions which we discuss in the following.", + "bbox": [ + 212, + 686, + 787, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Bucketed Ranking-based Loss Functions", + "text_level": 1, + "bbox": [ + 214, + 784, + 594, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, we introduce Bucketed versions of AP and RS Losses. Please refer to Alg. S2 (in Suppl. Mat.) for the algorithm.", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Bucketed Ranking-based Losses", + "bbox": [ + 517, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Bucketed AP Loss Definition. To define Bucketed version of AP Loss, we need to define current $(\\ell_{\\mathrm{R}}^{b}(i))$ and target ranking errors $(\\ell_{\\mathrm{R}}^{b,*}(i))$ for the $i$ th positive. Similar to previous works [5,6,24], we use a target ranking error of 0, i.e., $\\ell_{\\mathrm{R}}^{b,*}(i) = 0$ . Defining the current ranking error, similar to Eq. 3, requires $\\operatorname{rank}(i)$ and $N_{\\mathrm{FP}}(i)$ . These two quantities can easily be defined using the bucket size $b_i$ and the prototype logit. Different from the conventional AP Loss, we obtain the pairwise relation $H(x_{ij}^{b})$ using the prototype logit $s_i^b$ . Particularly, with the ordering in Eq. 12, $\\operatorname{rank}(i) = \\sum_{j=1}^{i} H(x_{ij}) + H(x_{ij}^{b}) b_j$ and $N_{\\mathrm{FP}}(i) = \\sum_{j=1}^{i} H(x_{ij}^{b}) b_j$ where $x_{ij}^{b} = s_j^b - s_i$ . Then, the resulting ranking error computed on $i$ th positive is:", + "bbox": [ + 212, + 146, + 787, + 306 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\mathrm {R}} ^ {b} (i) = \\frac {N _ {\\mathrm {F P}} (i)}{\\operatorname {r a n k} (i)} = \\frac {\\sum_ {j = 1} ^ {i} H \\left(x _ {i j} ^ {b}\\right) b _ {j}}{\\sum_ {j = 1} ^ {i} H \\left(x _ {i j}\\right) + H \\left(x _ {i j} ^ {b}\\right) b _ {j}}, \\quad i \\in \\mathcal {P}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 311, + 785, + 348 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Optimisation. Here, we need to define the primary terms, as the product between the error and the pmf following Identity Update [24]. Unlike the previous work, the weights of each prototype negative are not equal as a bucket includes varying number of negatives. Therefore, we use a weighted pmf while distributing the ranking error over negatives. Formally, if $j$ is the $j$ th prototype negative, then we define the pmf as $p(j^b|i) = b_j / N_{\\mathrm{FP}}(i)$ . The resulting primary term between the $i$ th positive and $j^b$ th negative is then:", + "bbox": [ + 214, + 352, + 787, + 460 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {i j} ^ {b} = \\frac {\\sum_ {j = 1} ^ {i} H \\left(x _ {i j} ^ {b}\\right) b _ {j}}{\\sum_ {j = 1} ^ {i} H \\left(x _ {i j}\\right) + H \\left(x _ {i j} ^ {b}\\right) b _ {j}} \\times \\frac {b _ {j}}{N _ {\\mathrm {F P}} (i)}, \\quad i \\in \\mathcal {P}, j \\in \\tilde {\\mathcal {N}} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 465, + 785, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathcal{N}}$ is the set of prototype negatives. However, we need the primary terms for the $i$ th negative, not for the prototype ones. Given $L_{ij}^{b}$ , the gradients of the prototype negatives and actual negatives can easily be obtained following Identity Update. However, we still need to find the gradients of the actual negatives. To do so, we simply normalise the prototype gradient by its bucket size $b_{j}$ and distribute the gradients to actual negatives, completing our method. In Theorem 1, we establish that our Bucketed AP Loss provides exactly the same gradients with AP Loss when $\\delta = 0$ .", + "bbox": [ + 212, + 507, + 787, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Bucketed RS Loss Given that we obtained Bucketed AP Loss, converting RS Loss into a bucketed form is more straightforward. This is because (i) the primary term and its gradients of RS Loss are equal to those of AP Loss if $i \\in \\mathcal{P}$ , $j \\in \\mathcal{N}$ and we simply use $L_{ij}^{b}$ in such cases; and (ii) if $i \\in \\mathcal{P}$ , $j \\in \\mathcal{P}$ , then an additional term is included in RS Loss to estimate the pairwise relations between positives (Eq. 11), which is not affected by bucketing.", + "bbox": [ + 212, + 647, + 787, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 Theoretical Discussion", + "text_level": 1, + "bbox": [ + 214, + 757, + 447, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The first theorem ensures that the gradients provided by our loss functions are identical with their conventional counterparts under certain circumstances. The second states that our algorithm is theoretically faster than the conventional AP and RS Losses.", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Yavuz et al.", + "bbox": [ + 271, + 114, + 352, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Theorem 1. Bucketed AP Loss and Bucketed RS Loss provide exactly the same gradients with AP and RS Losses respectively when $\\delta = 0$ .", + "bbox": [ + 215, + 146, + 787, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Theorem 2. Bucketed RS and Bucketed AP Losses have $\\mathcal{O}(\\max((|\\mathcal{P}| + |\\mathcal{N}|) \\log(|\\mathcal{P}| + |\\mathcal{N}|), |\\mathcal{P}|^2))$ time complexity.", + "bbox": [ + 215, + 186, + 816, + 219 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As a result, we observe the same performance empirically in less amount of time. Proofs are provided in Supp. Mat.", + "bbox": [ + 215, + 228, + 784, + 260 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Bucketed Rank-Sort (BRS) DETR", + "text_level": 1, + "bbox": [ + 215, + 282, + 588, + 301 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "While transformer-based detectors [8,14,16,37,38] have been providing the best performance in several object detection benchmarks, optimizing such detectors using ranking-based losses based on performance measures has not been investigated. Here, we address this gap by incorporating our BRS Loss into CoDETR [38] as the current SOTA detector on the common COCO benchmark [18].", + "bbox": [ + 212, + 313, + 782, + 388 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Co-DETR increases the number and variation of the positive examples in the transformer head by leveraging one-to-many assignment strategies in anchor-based detectors (such as ATSS [34] and Faster R-CNN [29]) as auxiliary heads. As the number of proposals is large in such detectors, they can easily provide more and diverse examples, resulting in better performance of the transformer head. Co-DETR is originally supervised with the following loss function:", + "bbox": [ + 212, + 390, + 784, + 479 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {l = 1} ^ {L} \\left(\\tilde {\\mathcal {L}} _ {l} ^ {D e c} + \\lambda_ {1} \\sum_ {k = 1} ^ {K} \\tilde {\\mathcal {L}} _ {k, l} ^ {A u x} + \\lambda_ {2} \\sum_ {k = 1} ^ {K} \\mathcal {L} _ {k, l} ^ {A u x}\\right), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 489, + 784, + 527 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ weigh each loss component. The first two components, $\\tilde{\\mathcal{L}}_l^{Dec}$ and $\\tilde{\\mathcal{L}}_{k,l}^{Aux}$ , are the losses of the $l^{th}$ decoder layer in the following form:", + "bbox": [ + 215, + 537, + 784, + 571 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {c l s} \\mathcal {L} _ {c l s} + \\lambda_ {b b o x} \\mathcal {L} _ {b b o x} + \\lambda_ {I o U} \\mathcal {L} _ {I o U}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 585, + 784, + 599 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "and $\\mathcal{L}_{cls}$ , $\\mathcal{L}_{bbox}$ and $\\mathcal{L}_{IoU}$ are Focal Loss, L1 Loss and GIoU Loss respectively [38]. $\\tilde{\\mathcal{L}}_l^{Dec}$ and $\\tilde{\\mathcal{L}}_{k,l}^{Aux}$ differ in their input queries. That is, while $\\tilde{\\mathcal{L}}_l^{Dec}$ follows standard DETR-based models [14, 37] with one-to-one matching of the queries, $\\tilde{\\mathcal{L}}_{k,l}^{Aux}$ is computed based on positive anchors of $k^{th}$ auxiliary head with one-to-many assignment. And $\\mathcal{L}_{k,l}^{Aux}$ is the conventional loss of the $k^{th}$ auxiliary head, e.g., the weighted sum of classification, localisation and centerness losses for ATSS [34].", + "bbox": [ + 215, + 611, + 784, + 705 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In order to align the training and evaluation objectives better, we replace $\\tilde{\\mathcal{L}}_l^{Dec}$ and $\\tilde{\\mathcal{L}}_{k,l}^{Aux}$ in Eq. 16 by:", + "bbox": [ + 215, + 707, + 784, + 738 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {B R S} + \\lambda_ {b b o x} \\mathcal {L} _ {b b o x} + \\lambda_ {I o U} \\mathcal {L} _ {I o U}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 753, + 784, + 767 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "and set $\\lambda_{bbox}$ and $\\lambda_{IoU}$ dynamically during training to $\\mathcal{L}_{BRS} / \\mathcal{L}_{bbox}$ and $\\mathcal{L}_{BRS} / \\mathcal{L}_{IoU}$ to simplify the hyperparameter tuning following Oksuz et al. [24]. Similarly, we replace the loss functions of the auxiliary ATSS [34] and Faster R-CNN [29] heads ( $\\mathcal{L}_{k,l}^{Aux}$ in Eq. 15) by our BRS Loss. This modification also aligns with the", + "bbox": [ + 212, + 779, + 799, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Bucketed Ranking-based Losses", + "bbox": [ + 517, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "objective of the auxiliary heads by the performance measure, hence results in more accurate auxiliary heads [24]. Furthermore, as using BRS Loss with Faster R-CNN does not require sampling thanks to its robustness to imbalance, no limitation is imposed on the number of positives in Faster R-CNN. Hence, the main aim of Co-DETR, that is to introduce more positive examples to the transformer head, is corroborated. As a result, our BRS DETR enables significantly efficient training compared to RS Loss also by improving the performance of Co-DETR.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6 Experiments", + "text_level": 1, + "bbox": [ + 215, + 276, + 375, + 292 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We analyze our contributions in three primary sections. First, we evaluate the effectiveness of our bucketing approach by contrasting it with RS Loss and AP Loss across different CNN-based visual detectors, on object detection and instance segmentation tasks. Next, we thoroughly examine how our bucketed RS Loss performs on transformer-based object detectors with different backbones. This is the first time a ranking-based loss is applied to transformer-based object detectors, thanks to the time-efficiency of our bucketing method. Finally, we present comparisons with score-based loss functions and other ranking-based loss functions in the literature. Our experiments clearly show that in terms of training time, accuracy and ease of tuning, our BRS Loss is more preferable over any existing loss function to train object detectors.", + "bbox": [ + 212, + 308, + 787, + 473 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dataset and Performance Measures. Unless otherwise noted, we train all models on COCO trainval35k (115k images) and test them on minival (5k images). We use COCO-style Average Precision (AP) and also report $\\mathrm{AP}_{50}$ , $\\mathrm{AP}_{75}$ as the APs at IoUs 0.50 and 0.75; and $\\mathrm{AP_S}$ , $\\mathrm{AP_M}$ and $\\mathrm{AP_L}$ to present the accuracy on small, medium and large objects.", + "bbox": [ + 212, + 474, + 787, + 551 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6.1 Efficiency and Effectiveness of Bucketed Ranking Losses", + "text_level": 1, + "bbox": [ + 214, + 571, + 723, + 588 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To comprehensively analyze the effectiveness of our bucketing method, we design experiments on real-world data as well as synthetic data. For the former, we aim to present that our method significantly decreases the training time of the detection and segmentation methods (up to $\\sim 5\\times$ ). As training the entire detector does not isolate the runtime of the loss function, on which our main contribution is, we also design an experiment with synthetic data. This set of experiments show that our bucketing approach, in fact, decreases the loss function runtime by up to $40\\times$ , thereby resulting in shorter training time of the detectors.", + "bbox": [ + 212, + 598, + 787, + 718 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Experimental Setup. We follow the experimental setup used in RSLoss [24] only by replacing the loss function by our bucketed loss function to ensure comparability and consistency of results. We report our findings on efficiency and accuracy using the models trained on 4 Tesla A100 GPUs with a batch size of 16, i.e., 4 images/GPU. Unless stated otherwise, we train the models for 12 epochs and test them on images with a size of $1333 \\times 800$ following the common convention [12, 24, 34]. In order to comprehensively demonstrate the efficiency of our approach, we use five different detectors: Faster R-CNN [29], Cascade R-CNN [2],", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Yavuz et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/f1cd41a8792a9a2992f472c73adff8e8f8b1151369c791eb8977b8cc89f00467.jpg", + "table_caption": [ + "Table 1: Bucketed Losses (BRS, BAP) vs. RS & AP Losses on COCO object detection for object detectors. R-X: ResNet-X. Models with R-101 are trained 36 epochs." + ], + "table_footnote": [], + "table_body": "
MethodDetectorBBLTimeAccuracy
\\(T_{iter}(s) \\downarrow\\)AP↑\\(\\mathrm{AP}_{50} \\uparrow\\)\\(\\mathrm{AP}_{75} \\uparrow\\)
Multi StageFaster R-CNNR-50RS0.5039.459.543.0
BRS\\(0.17 3.0x \\downarrow\\)39.559.542.8
R-101RS0.7547.367.451.2
BRS\\(0.38 2.0x \\downarrow\\)47.767.851.5
Cascade R-CNNR-50RS1.2841.158.744.1
BRS\\(0.24 5.3x \\downarrow\\)41.158.844.2
One StageATSSR-50AP0.3238.158.241.0
BAP\\(0.15 2.1x \\downarrow\\)38.558.641.3
RS0.3639.858.942.6
BRS\\(0.15 2.4x \\downarrow\\)39.858.842.9
PAAAP0.4537.354.341.2
BAP\\(0.30 1.5x \\downarrow\\)37.256.240.2
RS0.5740.858.844.6
BRS\\(0.30 1.9x \\downarrow\\)40.959.044.4
", + "bbox": [ + 232, + 185, + 772, + 378 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "ATSS [34] and PAA [12] for object detection as well as Mask R-CNN [10] for instance segmentation.", + "bbox": [ + 215, + 419, + 784, + 450 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Bucketed Ranking Based Losses Significantly Decrease the Training Time with Similar Accuracy. Multi-stage detectors: Among multi-stage detectors, we train the commonly-used Faster R-CNN and Cascade R-CNN with RS Loss and BRS Loss, and present average iteration time as well as AP of the trained models. In order to evaluate our contribution comprehensively, we train Faster R-CNN also with a stronger setting, in which, we use ResNet-101, train it for 36 epochs using multi-scale training similar to [24]. The results are presented in Table 1, in which we can clearly see that BRS Loss consistently reduces the training time of all three detectors as well as preserves (or slightly improve) their performance. Especially for Cascade R-CNN, which is still a very popular and strong object detector along with its variants such as HTC [4], the training time decreases by $5.3 \\times$ . This is because the loss function is applied to the logits for three times based on the cascaded nature of this detector, and therefore, our contribution can be easily noticed. Moreover, it might seem that the efficiency gain decreases once the size of the backbone increases from R-50 to R-101 in Faster R-CNN. However, this is an expected result as the overall feature extraction time increases due to the larger number of parameters in the backbone, and hence, as we will discuss in the synthetic experiments, this is independent from our bucketing approach, operating on the loss function after the feature extraction.", + "bbox": [ + 215, + 455, + 785, + 758 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "One-stage detectors: To show that our gains generalize to one-stage detectors and to AP Loss, we train the common ATSS and PAA detectors with our BAP and BRS Losses, and compare them with AP and RS Losses respectively. Table 1 validates our previous claims: Our bucketed losses obtain similar performance in around half of the training time required for AP and RS Losses.", + "bbox": [ + 215, + 765, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Bucketed Ranking-based Losses", + "bbox": [ + 519, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/6ec77a103cd2cee55d123e834d1899025f9200a94c1be60e6beea36b91f6f8c4.jpg", + "table_caption": [ + "Table 2: Comparison with RS Loss on instance segmentation task on COCO val using Mask R-CNN. BB: Backbone." + ], + "table_footnote": [], + "table_body": "
BBLEfficiencyAccuracy
\\(T_{iter}(s)\\)↓AP↑\\(AP_{50}\\)↑\\(AP_{75}\\)↑
R-50RS0.5636.357.238.8
BRS0.24 2.3x ↓36.257.238.8
R-101RS0.5940.261.843.5
BRS0.27 2.2x ↓40.362.043.8
", + "bbox": [ + 220, + 199, + 483, + 268 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/1fcd948665081d87c25ee3e558b8d1d69d7726ac966936af413286c418dcdb3d.jpg", + "table_caption": [ + "Table 3: Comparison on different inst. segm. datasets using Mask R-CNN. $\\mathrm{AP}_{75}$ is N/A as it is not used for Cityscapes." + ], + "table_footnote": [], + "table_body": "
DatasetLEfficiencyAccuracy
\\(T_{iter}(s) \\downarrow\\)AP↑\\(AP_{50} \\uparrow\\)\\(AP_{75} \\uparrow\\)
CityscapesRS0.4343.568.1N/A
BRS\\(0.19 2.3x \\downarrow\\)43.367.7N/A
LVISRS0.8725.639.227.3
BRS\\(0.35 2.5x \\downarrow\\)25.839.627.4
", + "bbox": [ + 514, + 199, + 785, + 268 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/0f041e36a3aabba76a9d302524394365ffca9675b43300e2790045164f1e9900.jpg", + "table_caption": [ + "Table 4: BRS Loss further improves performance with similar training budget. We report mask AP for instance segmentation" + ], + "table_footnote": [], + "table_body": "
TaskDetectorLAP↑AP50↑AP75↑
Object Det.C.R-CNNRS41.158.744.1
BRS42.3+1.260.245.2
Inst.Seg.M.R-CNNRS36.357.238.8
BRS37.3+1.058.440.2
", + "bbox": [ + 331, + 324, + 669, + 386 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Instance Segmentation: Given that our loss functions is efficient in object detectors, one simple extension is to see their generalization to instance segmentation methods. To show that, we train Mask R-CNN, a common baseline, with our BRS Loss on three different datasets from various domains: (i) COCO (Table 2), (ii) Cityscapes [7] as an autonomous driving dataset (Table 3) and LVIS [9] a long-tailed dataset with more than 1K classes (Table 3). The results confirm our previous finding that our BRS Loss decreases the training time significantly by up to $\\sim 2.5\\times$ compared to RS Loss by preserving its accuracy.", + "bbox": [ + 212, + 415, + 784, + 536 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Given Similar Training Budget, Bucketed Ranking Based Loss Functions Significantly Improve the Accuracy. There might be several benefits of reducing the training time of the detector as more GPU time is saved. Here, we show a use-case in which we ask the following question: what would happen if we allocate the similar amount of training time to both bucketed and non-bucketed losses? To answer that, we train Cascade R-CNN, a detector, and Mask R-CNN, an instance segmentation method, using our BRS Loss. Considering our speed-ups on these models $(6.0\\times$ and $2.3\\times)$ , we simply increased their training epochs from 12 to 36 and 27 respectively. Note that, for Cascade R-CNN, we still spend significantly less amount of training time. Table 4 shows that models trained with the BRS Loss outperform (with $+1.2$ and $1.0\\mathrm{AP}$ ) their counterparts trained with the RS Loss in both cases.", + "bbox": [ + 212, + 537, + 785, + 717 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Experiments on Synthetic Data. Up to now, to measure any speed-up, we considered the entire pipeline and did not decouple the loss computation step. However, we only improve the loss computation step. Hence, an analysis focusing only on this part can reveal our contribution more clearly. Therefore, we design a simple experiment using synthetic data. Particularly, we randomly generate $L$ logits such that $m\\%$ of these logits are positive (see Supp.Mat. for data generation details). To cover the wide range of existing detectors, we set $L \\in \\{10K, 100K, 1M\\}$ and $m \\in \\{0.1, 1.0, 2.0, 5.0\\}$ . Different choices for $L$ and $m$", + "bbox": [ + 212, + 719, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Yavuz et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/9fe7d0684473471f44bc12e4c8ca762d1442820a603b2888d8191196fb376881.jpg", + "table_caption": [ + "Table 5: Comparison of our BRS-DETR with Co-DETR (trained with its original loss function) on different backbones on COCO val set." + ], + "table_footnote": [], + "table_body": "
BackboneDetectorAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_s \\)\\( AP_m \\)\\( AP_1 \\)
ResNet50Co-DETR49.367.254.032.152.663.8
BRS-DETR50.167.454.631.953.965.0
Swin-TCo-DETR51.769.656.434.454.966.8
BRS-DETR52.369.557.134.655.768.0
Swin-LCo-DETR56.975.562.640.161.273.3
BRS-DETR57.275.062.539.461.774.1
", + "bbox": [ + 220, + 213, + 496, + 292 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/14fd7e3f1c0cad8f7980c47c9f2e20063758a197f00fab60002dc10940c7a60a.jpg", + "table_caption": [ + "Table 6: Comparison of our BRS-DETR with DETR variants (trained with their original loss functions) w ResNet-50 on COCO val set. Qu: Queries. Ep: Epochs." + ], + "table_footnote": [], + "table_body": "
DetectorQuEpAP\\( {\\mathrm{{AP}}}_{50} \\)\\( {\\mathrm{{AP}}}_{75} \\)\\( {\\mathrm{{AP}}}_{\\mathrm{s}} \\)\\( {\\mathrm{{AP}}}_{\\mathrm{m}} \\)\\( {\\mathrm{{AP}}}_{1} \\)
DETR [14]10030042.062.444.220.545.861.1
DN-DETR [15]3005048.667.452.731.052.063.7
DINO [3]9001249.466.953.832.352.563.9
H-DETR [11]3001248.766.452.931.251.563.5
Co-DETR [38]3001249.367.254.032.152.663.8
BRS-DETR3001250.167.454.631.953.965.0
", + "bbox": [ + 513, + 214, + 784, + 290 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "correspond to the logit distributions of common object detectors. We compute BAP and AP Loss on this synthetic data and compare their runtime to focus on loss computation time. We observed that our BAP Loss has a significant, up to $\\sim 40\\times$ , speed-up compared to AP Loss. The speed-up becomes more pronounced with increasing number of logits. Supp. Mat. presents detailed results.", + "bbox": [ + 212, + 325, + 784, + 402 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.2 The Effectiveness of Our BRS-DETR", + "text_level": 1, + "bbox": [ + 215, + 440, + 570, + 455 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "While we incorporate our BRS Loss into Co-DETR, we use all original settings unless otherwise noted. Please refer to Supp.Mat. for the details.", + "bbox": [ + 212, + 482, + 784, + 513 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "BRS-DETR Outperforms Existing Transformer-based Detectors Consistently. Here, we first employ ResNet-50 backbone as it has been used by many detectors, enabling us to compare our method with many different DETR-based manner in a consistent manner. In this fair comparison setting, our BRS-DETR outperforms all existing DETR variants reaching 50.1 AP as shown in Table 6. For example, our BRS DETR outperforms DN-DETR by 1.5AP with less training epochs and improves Co-DETR as we discuss next.", + "bbox": [ + 212, + 515, + 785, + 621 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Training Co-DETR with RS Loss takes 6x less time compared to using RS Loss. Finally, we also compare the training efficiencies of BRS Loss and RS Loss on Co-DETR. We observe that our BRS Loss decreases the training time of Co-DETR by $6.0 \\times$ (from 4.14s per iteration to 0.69s) in comparison to RS Loss. We note that this is the largest training time gain of our BRS Loss. This is because Co-DETR consists of multiple transformer-based as well as auxiliary heads, requiring multiple loss estimations.", + "bbox": [ + 212, + 625, + 785, + 731 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "BRS-DETR Improves Co-DETR over Different Backbones Consistently. In order to show the effectiveness of our BRS Loss, we compare our results with Co-DETR in Table 5 using different backbones. We note that we improve the baseline Co-DETR consistently in all settings. For example, our improvement on ResNet-50 is 0.8AP, which is a notable improvement. However, as the backbone gets larger (e.g., Swin-L [20]), our gains decrease, which is expected and commonly observed in the literature, e.g., [12].", + "bbox": [ + 212, + 733, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Bucketed Ranking-based Losses", + "bbox": [ + 517, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/27eb2f1e8afb1a3fe26b3b13199a149fc746375105f54f66bc6c370661fcd876.jpg", + "table_caption": [ + "Table 7: Comparison with the score-based losses. CE: Cross-entropy. #H: Number of hyperparameters. Ours has very similar $T_{iter}(s)$ also by being more accurate and simple-to-tune." + ], + "table_footnote": [], + "table_body": "
DetectorLTiter(s) ↓AP ↑#H
Faster R-CNNCE+L10.1437.69
RS0.5039.43
BRS (Ours)0.1739.53
ATSSFocal L. +G IoU0.1439.35
RS0.3639.81
BRS (Ours)0.1539.81
", + "bbox": [ + 220, + 226, + 501, + 310 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/e253b80241448a19f850ad5c391bced8d0a1f398ded517acf27f0cfb36074a5b.jpg", + "table_caption": [ + "Table 8: Comparison with other ranking-based losses. Our approach performs better on both accuracy and efficiency." + ], + "table_footnote": [], + "table_body": "
DetectorL\\(T_{iter}(s) \\downarrow\\)AP ↑#H
Faster R-CNNaLRP [23]0.2837.43
BRS (Ours)0.1739.53
ATSSAP [5]0.3238.15
DR [28]0.2038.15
aLRP [23]0.3237.71
BRS (Ours)0.1539.81
RetinaNetDR [28]0.2937.45
BRS (Ours)0.2338.31
", + "bbox": [ + 514, + 200, + 782, + 308 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6.3 Comparison with Score-based and Other Ranking-based Losses", + "text_level": 1, + "bbox": [ + 215, + 342, + 781, + 357 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In previous sections, we showed the efficiency of our loss functions compared to AP and RS Losses as their counterparts. Here, we compare our bucketed losses with the score based losses, i.e., Cross-entropy and Focal Loss, and other ranking based losses including DR Loss [28] and aLRP Loss [23]. To do so, in Tables 7 and 8, we report average training time, AP for accuracy and number of hyperparameters to capture the tuning simplicity of the loss functions. Compared to score-based loss functions in Table 7, our BRS Loss improves performance of both approaches, is significantly simpler-to-tune with a low number of hyperparameters and has similar training iteration time. Hence, our BRS Loss is either superior or on par with existing score-based losses. As for other ranking-based losses in Table 8, our BRS Loss is the most efficient and accurate ranking-based loss function for all three detectors. As an example, compared to ATSS trained with DR Loss, our BRS Loss yields 1.8AP better accuracy with $25\\%$ less training time and significantly less number of hyperparameters. Therefore, our bucketed loss functions are now promising alternatives to train object detectors.", + "bbox": [ + 212, + 375, + 787, + 602 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 633, + 359, + 648 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduced a novel method to improve the efficiency of ranking-based loss functions by binning negatives into buckets and implementing positive-negative pairwise comparisons between positives and buckets of negatives. We showed that this method improves the computational complexity to a level where pairwise comparisons can be stored as a matrix in memory and running time becomes very close to that of score-based loss functions. Also, for the first time, we integrated a ranking-based loss to Co-DETR, a DETR-based detector, which was possible thanks to our method's lower complexity, and reported improvements on different backbones. Our comprehensive experiments comprising 2 different tasks, 3 different dataset and 7 different detectors - all with positive outcomes - show the general applicability of our approach.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Yavuz et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We gratefully acknowledge the computational resources provided by TÜBITAK ULAKBIM High Performance and Grid Computing Center (TRUBA), METU Center for Robotics and Artificial Intelligence (METU-ROMER) and METU Image Processing Laboratory. Dr. Akbas is supported by the \"Young Scientist Awards Program (BAGEP)\" of Science Academy, Türkiye.", + "bbox": [ + 212, + 179, + 787, + 256 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 282, + 321, + 299 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Brown, A., Xie, W., Kalogeiton, V., Zisserman, A.: Smooth-ap: Smoothing the path towards large-scale image retrieval. In: European Conference on Computer Vision. pp. 677-694. Springer (2020)", + "2. Cai, Z., Vasconcelos, N.: Cascade R-CNN: Delving into high quality object detection. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018)", + "3. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers (2021)", + "4. Chen, K., Pang, J., Wang, J., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Shi, J., Ouyang, W., Loy, C.C., Lin, D.: Hybrid task cascade for instance segmentation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)", + "5. Chen, K., Li, J., Lin, W., See, J., Wang, J., Duan, L., Chen, Z., He, C., Zou, J.: Towards accurate one-stage object detection with ap-loss. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)", + "6. Chen, K., Lin, W., li, J., See, J., Wang, J., Zou, J.: Ap-loss for accurate one-stage object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) pp. 1-1 (2020)", + "7. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016)", + "8. Dai, Z., Cai, B., Lin, Y., Chen, J.: Unsupervised pre-training for detection transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence p. 1-11 (2022). https://doi.org/10.1109/tpami.2022.3216514, http://dx.doi.org/10.1109/TPAMI.2022.3216514", + "9. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)", + "0. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask R-CNN. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2017)", + "1. Jia, D., Yuan, Y., He, H., Wu, X., Yu, H., Lin, W., Sun, L., Zhang, C., Hu, H.: Detrs with hybrid matching (2023)", + "2. Kim, K., Lee, H.S.: Probabilistic anchor assignment with iou prediction for object detection. In: The European Conference on Computer Vision (ECCV) (2020)", + "3. Law, H., Deng, J.: Cornernet: Detecting objects as paired keypoints. In: The European Conference on Computer Vision (ECCV) (2018)" + ], + "bbox": [ + 225, + 316, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Bucketed Ranking-based Losses", + "bbox": [ + 519, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13619-13627 (June 2022)", + "15. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising (2022)", + "16. Li*, L.H., Zhang*, P., Zhang*, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., Chang, K.W., Gao, J.: Grounded language-image pretraining. In: CVPR (2022)", + "17. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 42(2), 318-327 (2020)", + "18. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common Objects in Context. In: The European Conference on Computer Vision (ECCV) (2014)", + "19. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S.E., Fu, C., Berg, A.C.: SSD: single shot multibox detector. In: The European Conference on Computer Vision (ECCV) (2016)", + "20. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows (2021)", + "21. Mohapatra, P., Jawahar, C., Kumar, M.P.: Efficient optimization for average precision svm. Advances in Neural Information Processing Systems 27 (2014)", + "22. Mohapatra, P., Rolínek, M., Jawahar, C., Kolmogorov, V., Pawan Kumar, M.: Efficient optimization for rank-based loss functions. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018)", + "23. Oksuz, K., Cam, B.C., Akbas, E., Kalkan, S.: A ranking-based, balanced loss function unifying classification and localisation in object detection. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)", + "24. Oksuz, K., Cam, B.C., Akbas, E., Kalkan, S.: Rank & sort loss for object detection and instance segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3009-3018 (2021)", + "25. Oksuz, K., Cam, B.C., Kahraman, F., Baltaci, Z.S., Kalkan, S., Akbas, E.: Mask-awareiou for anchor assignment in real-time instance segmentation. In: The British Machine Vision Conference (BMVC) (2021)", + "26. Oksuz, K., Cam, B.C., Kalkan, S., Akbas, E.: Imbalance problems in object detection: A review. IEEE transactions on pattern analysis and machine intelligence 43(10), 3388-3415 (2020)", + "27. Pang, J., Chen, K., Shi, J., Feng, H., Ouyang, W., Lin, D.: Libra R-CNN: Towards balanced learning for object detection. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)", + "28. Qian, Q., Chen, L., Li, H., Jin, R.: Dr loss: Improving object detection by distributional ranking. In: CVPR (2020)", + "29. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 39(6), 1137-1149 (2017)", + "30. Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I., Savarese, S.: Generalized intersection over union: A metric and a loss for bounding box regression. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)", + "31. Rosenblatt, F.: The perceptron: A probabilistic model for information storage and organization in the brain. Psychological Review pp. 65-386 (1958)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Yavuz et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Shrivastava, A., Gupta, A., Girshick, R.: Training region-based object detectors with online hard example mining. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2016)", + "33. Sun, P., Zhang, R., Jiang, Y., Kong, T., Xu, C., Zhan, W., Tomizuka, M., Li, L., Yuan, Z., Wang, C., Luo, P.: SparseR-CNN: End-to-end object detection with learnable proposals. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018)", + "34. Zhang, S., Chi, C., Yao, Y., Lei, Z., Li, S.Z.: Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)", + "35. Zhang, X., Wan, F., Liu, C., Ji, R., Ye, Q.: Freeanchor: Learning to match anchors for visual object detection. In: Advances in Neural Information Processing Systems (NeurIPS) (2019)", + "36. Zheng, Z., Wang, P., Liu, W., Li, J., Rongguang, Y., Dongwei, R.: Distance-iou loss: Faster and better learning for bounding box regression. In: AAAI Conference on Artificial Intelligence (2020)", + "37. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. ArXiv abs/2010.04159 (2020), https://api-semanticscholar.org/CorpusID:222208633", + "38. Zong, Z., Song, G., Liu, Y.: Detrs with collaborative hybrid assignments training (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 438 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Bucketed Ranking-based Losses", + "bbox": [ + 519, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Bucketed Ranking-based Losses for Efficient Training of Object Detectors/7957777c-4d0a-4feb-89c6-b57e75289486_model.json b/2024/Bucketed Ranking-based Losses for Efficient Training of Object Detectors/7957777c-4d0a-4feb-89c6-b57e75289486_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e5006edf2840866278e5700a33d86b180f566758 --- /dev/null +++ b/2024/Bucketed Ranking-based Losses for Efficient Training of Object Detectors/7957777c-4d0a-4feb-89c6-b57e75289486_model.json @@ -0,0 +1,2324 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.287, + 0.141, + 0.717, + 0.187 + ], + "angle": 0, + "content": "Bucketed Ranking-based Losses for Efficient Training of Object Detectors" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.212, + 0.781, + 0.244 + ], + "angle": 0, + "content": "Feyza Yavuz1, Baris Can Cam1, Adnan Harun Dogan1, Kemal Oksuz2, Emre Akbas1,3,*, and Sinan Kalkan1,3,*" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.254, + 0.721, + 0.282 + ], + "angle": 0, + "content": "\\(^{1}\\) Department of Computer Engineering, METU, Ankara, Türkiye \\(^{2}\\) Five AI Ltd., United Kingdom" + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.283, + 0.679, + 0.311 + ], + "angle": 0, + "content": "3 METU ROMER Robotics Center, Ankara, Türkiye * Equal senior authorship" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.35, + 0.744, + 0.683 + ], + "angle": 0, + "content": "Abstract. Ranking-based loss functions, such as Average Precision Loss and Rank&Sort Loss, outperform widely used score-based losses in object detection. These loss functions better align with the evaluation criteria, have fewer hyperparameters, and offer robustness against the imbalance between positive and negative classes. However, they require pairwise comparisons among \\( P \\) positive and \\( N \\) negative predictions, introducing a time complexity of \\( \\mathcal{O}(PN) \\), which is prohibitive since \\( N \\) is often large (e.g., \\( 10^8 \\) in ATSS). Despite their advantages, the widespread adoption of ranking-based losses has been hindered by their high time and space complexities. In this paper, we focus on improving the efficiency of ranking-based loss functions. To this end, we propose Bucketed Ranking-based Losses which group negative predictions into \\( B \\) buckets (\\( B \\ll N \\)) in order to reduce the number of pairwise comparisons so that time complexity can be reduced. Our method enhances the time complexity, reducing it to \\( \\mathcal{O}(\\max(N\\log(N),P^2)) \\). To validate our method and show its generality, we conducted experiments on 2 different tasks, 3 different datasets, 7 different detectors. We show that Bucketed Ranking-based (BR) Losses yield the same accuracy with the unbucketed versions and provide \\( 2 \\times \\) faster training on average. We also train, for the first time, transformer-based object detectors using ranking-based losses, thanks to the efficiency of our BR. When we train CoDETR, a state-of-the-art transformer-based object detector, using our BR Loss, we consistently outperform its original results over several different backbones. Code will be released." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.708, + 0.753, + 0.725 + ], + "angle": 0, + "content": "Keywords: Object Detection \\(\\cdot\\) Ranking Losses \\(\\cdot\\) Efficient Ranking Losses" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.748, + 0.377, + 0.765 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Training an object detector incurs many challenges unique to object detection, including (i) assigning dynamically a large number of object hypotheses with the objects [2, 12, 13, 25, 29, 33, 34], (ii) sampling among these hypotheses to ensure that the background class does not dominate training [19, 27, 29, 32] and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Yavuz et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.52, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.195, + 0.436, + 0.206 + ], + "angle": 0, + "content": "(a) Comparisons in AP Loss" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.223, + 0.52, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.282, + 0.467, + 0.293 + ], + "angle": 0, + "content": "(b) Comparisons in Bucketed AP Loss (ours)" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.144, + 0.787, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.278, + 0.782, + 0.298 + ], + "angle": 0, + "content": "(c) Average Precision (AP) vs. average training iteration time for baselines, RS Loss and Bucketed RS Loss (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.308, + 0.788, + 0.406 + ], + "angle": 0, + "content": "Fig. 1: (a) Existing ranking-based losses (i.e., AP Loss [5,6]) incur significant overhead owing to pairwise comparisons between positives and negatives. (b) We propose bucketing negatives to decrease the number of comparisons, and hence the complexity. Under certain assumptions, our bucketing approach provides equal gradients with conventional ranking-based losses such as AP Loss in (a). (c) Accuracy and efficiency comparison. Our BRS formulation facilitates faster (between \\(1.9 \\times\\) and \\(6.0 \\times\\)) training of visual detectors with similar AP." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.449, + 0.788, + 0.555 + ], + "angle": 0, + "content": "(iii) minimizing a multi-task objective function [17,23,24,29]. While the choices on assignment and sampling generally vary depending on the trained object detector, it is very common to combine Cross Entropy or Focal Loss [17] with a regression loss [30,36] as the multi-task loss function. Recently proposed ranking-based loss functions [5,6,23,24] offer an alternative approach for addressing these challenges by formulating the training objective based on the rank of positive examples over negative examples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.562, + 0.788, + 0.699 + ], + "angle": 0, + "content": "Benefits of ranking-based losses. First, they are inherently robust to imbalance [23] and hence, do not require any sampling mechanism under very challenging scenarios [24], e.g., even when the background foreground ratio is 10K for LVIS [9]. Second, they are shown to generalize over different detectors with diverse architectures - with the exception of transformer-based ones [14,37,38], since ranking-based losses further slow down the training of transformer-based detectors, which we address in this paper. Such losses also offer significant performance gain over their score-based counterparts, and having less hyperparameters, they are much easier to tune [23,24]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The drawback of ranking-based losses. Compared to score-based losses, ranking-based losses are more inefficient as the ranking operation inherently requires each pair of object hypotheses to be compared against each other (Fig. 1(a)), inducing a quadratic complexity on the large number of object hypotheses (e.g., \\(10^{8}\\) for ATSS [34] on COCO [18]). As a result, vectorized implementations for parallel GPU processing are infeasible as such large matrices (e.g., with \\(\\sim\\) \\(10^{16}\\) entries for ATSS) do not fit into GPU memories. This has driven researchers towards alternative ways of computing these matrices, which, in the end, saves from the storage complexity but results in more inefficient algorithms [5,6]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.52, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bucketed Ranking-based Losses" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "Our approach. The main cause for the aforementioned complexity is the background class, i.e., negatives, forming up to \\(99.9\\%\\) of the hypotheses [26]. Considering this, we introduce a novel bucketing approach (Fig. 1(b)) on these negatives to address the main shortcoming of the ranking-based loss functions. Specifically, we first sort all examples based on their scores and bucket the negatives located between successive positives. The negatives within the same bucket are treated as a single example and represented by their average score. We then incorporate our bucketing approach into AP Loss [5,6] and RS Loss [24], and introduce novel mechanisms for calculating the gradients correctly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.449 + ], + "angle": 0, + "content": "The significance of our approach. Theoretically, we prove that our bucketing approach provides the same gradients as the original ranking-based loss. Practically, we show we can reduce loss computation time by up to \\(\\sim 40\\times\\) and training time of a detector by up to \\(6\\times\\) (Fig. 1(c)), addressing the main drawback of ranking-based loss functions. Hence, we show that the gap between the score-based and ranking-based loss functions in terms of training time vanishes, enabling us to train strong transformer-based object detectors such as Co-DETR using our loss functions. As ranking-based loss functions now surpass score-based ones or perform on par in every aspect such as accuracy, tuning simplicity, robustness to imbalance and training time, we believe that our work will make their prevalence more dominant in the upcoming object detectors." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.45, + 0.692, + 0.463 + ], + "angle": 0, + "content": "Contributions. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.472, + 0.787, + 0.547 + ], + "angle": 0, + "content": "- We propose a novel bucketing approach to improve the efficiency of computationally-expensive ranking-based losses to train object detectors. Theoretically, our approach yields the same gradients with the original ranking-based losses while decreasing their time complexity. Practically, we enable up to \\(6 \\times\\) faster training using our bucketing approach with no accuracy loss." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.548, + 0.787, + 0.607 + ], + "angle": 0, + "content": "- For the first time, we incorporate ranking-based loss functions to transformer-based detectors. Specifically, we construct BRS-DETR by replacing the training objective of the state-of-the-art transformer-based object detector CoDETR [38] by our Bucketed RS (BRS) Loss." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.608, + 0.787, + 0.713 + ], + "angle": 0, + "content": "- Our comprehensive experiments on detection and instance segmentation on 3 different challenging datasets, 5 backbones and 7 detectors show the effectiveness and generalizability of our approach. Our BRS-DETR reaches 50.1 AP on COCO val set with only 12 epochs and ResNet-50 backbone, outperforming all existing detectors using the same backbone and 300 queries. BRS-DETR also outperforms CoDETR with Swin-T and Swin-L backbones, reaching 57.2 AP on the COCO dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.735, + 0.388, + 0.751 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Although score-based loss functions, such as cross-entropy and focal loss, are widely used to train both CNN-based [2, 17, 35] and transformer-based object detectors [14, 37, 38], ranking-based losses for visual detection offer advantages. The primary advantage is their robustness to class imbalance [23]. They yield on par or better performance compared to score-based losses, without the need" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.128 + ], + "angle": 0, + "content": "Yavuz et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.236 + ], + "angle": 0, + "content": "for extensive tuning of class weights or multistage sampling techniques. Another advantage they offer is the ease of balancing losses among different tasks. In AP Loss [5, 6], aLRP Loss [23] and RS Loss [24], the authors use the ratio of classification and regression losses to balance these tasks. As a result, ranking-based losses have much fewer hyperparameters during training, compared to score-based ones." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.789, + 0.405 + ], + "angle": 0, + "content": "Other Ranking-based Losses. The literature has explored other ranking-based loss functions with different goals and assumptions. For example, DR Loss [28] constrains a margin to be satisfied between positives and negatives, not taking the recall into account. Smooth AP [1] introduces a smooth, differentiable approximation of AP, with the assumption that \\( |\\mathcal{N}| \\) is not large. Several studies focused on improving the efficiency of ranking-based losses for use in Support Vector Machines (SVMs) [21, 22]. However, these solutions have been limited to training linear SVMs and hence, problems that can be solved using linear functions. In this paper, we focus on improving ranking-based loss functions [5, 6, 24] that can be used for training complex deep networks such as object detectors." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.407, + 0.79, + 0.499 + ], + "angle": 0, + "content": "Comparative Summary. Ranking-based training objective for visual detectors proved their superiority over score-based loss functions in visual detection tasks [5, 6, 23, 24]. However, due to the requirement of pairwise ranking their time complexity is higher than the score-based alternatives. Such a disadvantage of ranking-based loss functions prevents their applicability to larger visual detectors (i.e., transformer-based object detectors)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.529, + 0.616, + 0.548 + ], + "angle": 0, + "content": "3 Background on Ranking-based Losses" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.57, + 0.789, + 0.678 + ], + "angle": 0, + "content": "Ranking-based losses for object detection rely on pairwise comparisons between the scores of different detections to determine the rank of a detection among positives and negatives. Denoting the score of the \\(i^{th}\\) detection by \\(s_i\\), we can compare the scores of two detections \\(i\\) and \\(j\\) with a simple difference transform: \\(x_{ij} = s_j - s_i\\). By counting the number of instances with \\(x_{ij} > 0\\), we can determine \\(i^{th}\\) detection's rank among positives \\((\\mathcal{P})\\) and all detections (positives and negatives: \\(\\mathcal{P} \\cup \\mathcal{N}\\)) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.692, + 0.788, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\operatorname {r a n k} ^ {+} (i) = \\sum_ {j \\in \\mathcal {P}} \\bar {H} \\left(x _ {i j}\\right), \\text {a n d} \\operatorname {r a n k} (i) = \\sum_ {j \\in \\mathcal {P} \\cup \\mathcal {N}} \\bar {H} \\left(x _ {i j}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.738, + 0.788, + 0.772 + ], + "angle": 0, + "content": "where \\(\\bar{H}(x)\\) is one if \\(x > 0\\) and zero otherwise. Since \\(\\mathrm{d}\\bar{H}(x) / \\mathrm{d}x\\) is either infinite (at \\(x = 0\\)) or zero (for \\(x \\neq 0\\)), a smoothed version is used (\\(\\delta\\): a hyper-parameter):" + }, + { + "type": "equation", + "bbox": [ + 0.391, + 0.786, + 0.788, + 0.843 + ], + "angle": 0, + "content": "\\[\nH (x) = \\left\\{ \\begin{array}{l l} 0, & x < - \\delta \\\\ \\frac {x}{2 \\delta} + 0. 5, & - \\delta \\leq x \\leq \\delta \\\\ 1, & \\delta < x \\end{array} \\right. \\tag {2}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.519, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bucketed Ranking-based Losses" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.588, + 0.163 + ], + "angle": 0, + "content": "3.1 Revisiting Average Precision (AP) Loss" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.173, + 0.672, + 0.188 + ], + "angle": 0, + "content": "Given the definitions in Eq. 1, AP Loss [5,6] can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.287, + 0.2, + 0.786, + 0.233 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {A P} = 1 - \\mathrm {A P} = 1 - \\frac {1}{| \\mathcal {P} |} \\sum_ {i \\in \\mathcal {P}} \\operatorname {p r e c i s i o n} (i) = 1 - \\frac {1}{| \\mathcal {P} |} \\sum_ {i \\in \\mathcal {P}} \\frac {\\operatorname {r a n k} ^ {+} (i)}{\\operatorname {r a n k} (i)}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.242, + 0.784, + 0.273 + ], + "angle": 0, + "content": "Chen et al. [5,6] simplified Eq. 3 by rewriting it in terms of pairwise comparisons between positive and negative detections:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.283, + 0.786, + 0.317 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {A P} = \\frac {1}{| \\mathcal {P} |} \\sum_ {i \\in \\mathcal {P}} \\sum_ {j \\in \\mathcal {N}} \\frac {H (x _ {i j})}{\\operatorname {r a n k} (i)} = \\frac {1}{| \\mathcal {P} |} \\sum_ {i \\in \\mathcal {P}} \\sum_ {j \\in \\mathcal {N}} L _ {i j} ^ {A P}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.326, + 0.784, + 0.371 + ], + "angle": 0, + "content": "where \\( L_{ij}^{AP} = H(x_{ij}) / \\mathrm{rank}(i) \\) is called the primary term. Note that \\( L_{ij}^{AP} \\) is non-differentiable since the step function \\( (H(\\cdot)) \\) applied on \\( x_{ij} \\) is non-differentiable, which will be discussed next." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.396, + 0.484, + 0.412 + ], + "angle": 0, + "content": "3.2 Revisiting Identity Update" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.422, + 0.784, + 0.452 + ], + "angle": 0, + "content": "Oksuz et al. [24] showed that various ranking-based losses (including AP Loss in Eqs. 3 & 4) can be written in a general form as:" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.465, + 0.786, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\frac {1}{Z} \\sum_ {i \\in \\mathcal {P} \\cup \\mathcal {N}} (\\ell (i) - \\ell^ {*} (i)), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.511, + 0.784, + 0.572 + ], + "angle": 0, + "content": "where \\(\\ell(i)\\) is the ranking-based error (e.g., precision error) computed on the \\(i\\)th detection, \\(\\ell^{*}(i)\\) is the target ranking-based error (the lowest error possible) and \\(Z\\) is the normalization constant. The loss in Eq. 5 can be computed and optimized as follows [24]:" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.572, + 0.785, + 0.671 + ], + "angle": 0, + "content": "1. Computation of the Loss. First, each pair of logits \\((s_i\\) and \\(s_j)\\) are compared by calculating their difference transforms as \\(x_{ij} = s_j - s_i\\). With the step function \\(H(\\cdot)\\) (Eq. 2), the number of detections higher than \\(s_i\\) (and therefore its precision error, \\(\\ell(i)\\)) can be easily calculated (see Sect. 3.1). Given \\(x_{ij}\\), the loss can be rewritten and calculated in terms of primary terms \\(L_{ij}\\) as \\(\\mathcal{L} = \\frac{1}{Z}\\sum_{i\\in \\mathcal{P}\\cup \\mathcal{N}}\\sum_{j\\in \\mathcal{P}\\cup \\mathcal{N}}L_{ij}\\), by taking:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.686, + 0.785, + 0.701 + ], + "angle": 0, + "content": "\\[\nL _ {i j} = \\left(\\ell (i) - \\ell^ {*} (i)\\right) p (j | i), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.712, + 0.784, + 0.758 + ], + "angle": 0, + "content": "where \\( p(j|i) \\) is a probability mass function (pmf) that distributes the error computed on the \\( i \\)th example over the \\( j \\)th example in order to determine the pairwise primary term \\( L_{ij} \\). \\( p(j|i) \\) is commonly taken as a uniform distribution [23, 24]." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.758, + 0.784, + 0.804 + ], + "angle": 0, + "content": "2. Optimization of the Loss. The gradient of the primary term wrt. to the difference transform \\((\\partial L_{ij} / \\partial x_{ij})\\) is non-differentiable. Denoting this term by \\(\\Delta x_{ij}\\), we have [5]:" + }, + { + "type": "equation", + "bbox": [ + 0.341, + 0.816, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\mathcal {L}}{\\partial s _ {i}} = \\sum_ {j, k} \\frac {\\partial \\mathcal {L}}{\\partial L _ {j k}} \\Delta x _ {j k} \\frac {\\partial x _ {j k}}{\\partial s _ {i}} = \\frac {1}{Z} \\left(\\sum_ {j} \\Delta x _ {j i} - \\sum_ {j} \\Delta x _ {i j}\\right). \\tag {7}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.128 + ], + "angle": 0, + "content": "Yavuz et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "Therefore, optimizing a ranking-based loss function reduces to determining \\(\\Delta x_{ij}\\). Inspired by Chen et al. [5], Oksuz et al. [24] employ Perceptron Learning [31] and show that \\(\\Delta x_{ij}\\) in Eq. 7 is simply the primary term itself: \\(\\Delta x_{ij} = -(L_{ij}^{*} - L_{ij}) = -(0 - L_{ij}) = L_{ij}\\), hence the name Identity Update. Plugging this into Eq. 7 yields (see Oksuz et al. [24] for the steps of the derivation):" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.235, + 0.786, + 0.263 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\mathcal {L}}{\\partial s _ {i}} = \\frac {1}{Z} \\left(\\sum_ {j} L _ {j i} - \\sum_ {j} L _ {i j}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.273, + 0.786, + 0.304 + ], + "angle": 0, + "content": "Therefore, both computing and optimizing the loss reduces determining the primary term \\( L_{ij} \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.326, + 0.712, + 0.342 + ], + "angle": 0, + "content": "3.3 Revisiting AP Loss and RS Loss with Identity Update" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.351, + 0.788, + 0.437 + ], + "angle": 0, + "content": "AP Loss [5, 6]. For defining AP Loss with identity update, the errors in Eq. 5 can be derived as \\(\\ell_{\\mathrm{R}}(i) = \\frac{N_{\\mathrm{FP}}(i)}{\\mathrm{rank}(i)}, \\ell_{\\mathrm{R}}^{*}(i) = 0\\) and \\(Z = |\\mathcal{P}|\\). Furthermore, defining \\(p_R(j|i)\\) as a uniform pmf, that is \\(p_R(j|i) = \\frac{H(x_{ij})}{N_{\\mathrm{FP}}(i)}\\), we can obtain the primary terms of AP Loss using Identity Update, completing the derivation for computation and optimization (Eq. 8):" + }, + { + "type": "equation", + "bbox": [ + 0.352, + 0.449, + 0.786, + 0.475 + ], + "angle": 0, + "content": "\\[\nL _ {i j} ^ {A P} = \\left\\{ \\begin{array}{l l} \\left(\\ell_ {\\mathrm {R}} (i) - \\ell_ {\\mathrm {R}} ^ {*} (i)\\right) p _ {R} (j | i), & \\text {f o r} i \\in \\mathcal {P}, j \\in \\mathcal {N} \\\\ 0, & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.485, + 0.787, + 0.517 + ], + "angle": 0, + "content": "RS Loss [24]. RS Loss includes an additional sorting objective \\((\\ell_S(i))\\) to promote better-localized positives to be ranked higher than other positives:" + }, + { + "type": "equation", + "bbox": [ + 0.352, + 0.527, + 0.786, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {R S}} = \\frac {1}{| \\mathcal {P} |} \\sum_ {i \\in \\mathcal {P}} \\left[ \\underbrace {\\left(\\ell_ {\\mathrm {R}} (i) - \\ell_ {\\mathrm {R}} ^ {*} (i)\\right)} _ {\\text {r a n k i n g}} + \\underbrace {\\left(\\ell_ {\\mathrm {S}} (i) - \\ell_ {\\mathrm {S}} ^ {*} (i)\\right)} _ {\\text {s o r t i n g}} \\right]. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.578, + 0.787, + 0.609 + ], + "angle": 0, + "content": "The primary terms of RS Loss can be obtained in a similar manner to AP Loss (see the Supp.Mat. for derivations), yielding:" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.621, + 0.786, + 0.66 + ], + "angle": 0, + "content": "\\[\nL _ {i j} ^ {R S} = \\left\\{ \\begin{array}{l l} \\left(\\ell_ {\\mathrm {R}} (i) - \\ell_ {\\mathrm {R}} ^ {*} (i)\\right) p _ {R} (j | i), & \\text {f o r} i \\in \\mathcal {P}, j \\in \\mathcal {N} \\\\ \\left(\\ell_ {\\mathrm {S}} (i) - \\ell_ {\\mathrm {S}} ^ {*} (i)\\right) p _ {S} (j | i), & \\text {f o r} i \\in \\mathcal {P}, j \\in \\mathcal {P}, \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.671, + 0.787, + 0.717 + ], + "angle": 0, + "content": "where \\( p_S(j|i) \\) is the sorting pmf. Note that RS Loss has the same primary terms with AP Loss for \\( i \\in \\mathcal{P}, j \\in \\mathcal{N} \\). The primary terms for \\( i \\in \\mathcal{P}, j \\in \\mathcal{P} \\) directly targets promoting the better-localized positives." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.74, + 0.634, + 0.755 + ], + "angle": 0, + "content": "3.4 Complexity of Ranking-based Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "These loss functions originally have the space complexity of \\(\\mathcal{O}((|\\mathcal{P}| + |\\mathcal{N}|)^2)\\) [5,6] as they need to compare each pair. This quadratic space complexity makes it infeasible to compute these loss functions using modern GPUs as the number of logits in object detection is very large. To alleviate that, Chen et al. [5] introduced two tricks at the expense of making the computation inefficient: (1) Loop" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.519, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bucketed Ranking-based Losses" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "on Positives: By implementing a loop over the positive examples, Chen et al. managed to reduce the time complexity to \\(\\mathcal{O}(|\\mathcal{P}|(|\\mathcal{P}| + |\\mathcal{N}|)) \\approx \\mathcal{O}(|\\mathcal{P}||\\mathcal{N}|)\\) since \\(\\mathcal{N} \\gg \\mathcal{P}\\) and the space complexity to \\(\\mathcal{O}(|\\mathcal{P}| + |\\mathcal{N}|)\\). (2) Discard Trivial Negatives: AP Loss considers negative examples with lower rank than all positives as trivial and disregards them." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.245, + 0.765, + 0.281 + ], + "angle": 0, + "content": "4 Bucketed Ranking-based Loss Functions for Efficient Optimisation of Object Detectors" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.295, + 0.789, + 0.355 + ], + "angle": 0, + "content": "Training object detectors with existing ranking losses has several advantages as we presented before, however, they suffer from significantly large training time. We now address this by our Bucketing Approach, and propose our Bucketed AP and RS Losses." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.378, + 0.646, + 0.394 + ], + "angle": 0, + "content": "4.1 Bucketing Negatives for Efficient Optimisation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.789, + 0.526 + ], + "angle": 0, + "content": "Our main intuition is to bucket the sequential negative examples considering that they will have very similar or equal gradients. Fig. 1(a) demonstrates this phenomenon in which \\( n_1, n_2 \\) both are assigned equal gradients as well as each negative among \\( n_3, n_4, n_5, n_6 \\) in the standard AP Loss. Formally, we sort all given logits \\( s_i \\in \\mathcal{P} \\cup \\mathcal{N} \\) using a conventional sorting methods. Let us denote this sorted permutation of the logits by \\( \\hat{s}_1, \\hat{s}_2, \\dots, \\hat{s}_{|\\mathcal{P} \\cup \\mathcal{N}|} \\), i.e., \\( \\hat{s}_1 > \\hat{s}_2 > \\dots > \\hat{s}_{|S|} \\). Given these sorted logits and denoting the \\( i \\)th positive logit in the ordering by \\( \\hat{s}_i^+ \\), the buckets of negatives \\( B_1, \\dots, B_{|\\mathcal{P} + 1|} \\) can be obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.537, + 0.785, + 0.557 + ], + "angle": 0, + "content": "\\[\nB _ {1} > \\hat {s} _ {1} ^ {+} > B _ {2} > \\hat {s} _ {2} ^ {+} > \\dots > \\hat {s} _ {| \\mathcal {P} |} ^ {+} > B _ {| \\mathcal {P} + 1 |}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.789, + 0.687 + ], + "angle": 0, + "content": "We will denote the size of the \\(i\\)th negative bucket by \\(b_{i}\\). Having created the buckets, we now determine a single logit value for each bucket, which we call as the prototype logit as it is necessary while assigning the ranking error. We denote the prototype logit for the \\(i^{\\mathrm{th}}\\) bucket by \\(s_i^b\\). Note that if \\(\\delta = 0\\) in Eq. 2, then any logit satisfying \\(\\hat{s}_{i - 1}^{+} > s_{i}^{b} > \\hat{s}_{i}^{+}\\) can be a prototype logit as the ranking stays the same. However, in the case of \\(\\delta > 0\\), the boundaries between the logits are smoothed. That's why, practically, we find it effective to use the mean logit of a bucket as its prototype logit." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.789, + 0.764 + ], + "angle": 0, + "content": "Note that this bucketing approach reduces the number of logits (positive and prototype negative) to a maximum of \\(2|\\mathcal{P}| + 1\\). As a result, the pairwise errors can now fit into the memory. Consequently, the loop in the red box of Alg. S1 (in Suppl. Mat.) is no longer necessary, which gives rise to efficient ranking-based loss functions which we discuss in the following." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.785, + 0.595, + 0.8 + ], + "angle": 0, + "content": "4.2 Bucketed Ranking-based Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Here, we introduce Bucketed versions of AP and RS Losses. Please refer to Alg. S2 (in Suppl. Mat.) for the algorithm." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.128 + ], + "angle": 0, + "content": "Yavuz et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.307 + ], + "angle": 0, + "content": "Bucketed AP Loss Definition. To define Bucketed version of AP Loss, we need to define current \\((\\ell_{\\mathrm{R}}^{b}(i))\\) and target ranking errors \\((\\ell_{\\mathrm{R}}^{b,*}(i))\\) for the \\(i\\)th positive. Similar to previous works [5,6,24], we use a target ranking error of 0, i.e., \\(\\ell_{\\mathrm{R}}^{b,*}(i) = 0\\). Defining the current ranking error, similar to Eq. 3, requires \\(\\operatorname{rank}(i)\\) and \\(N_{\\mathrm{FP}}(i)\\). These two quantities can easily be defined using the bucket size \\(b_i\\) and the prototype logit. Different from the conventional AP Loss, we obtain the pairwise relation \\(H(x_{ij}^{b})\\) using the prototype logit \\(s_i^b\\). Particularly, with the ordering in Eq. 12, \\(\\operatorname{rank}(i) = \\sum_{j=1}^{i} H(x_{ij}) + H(x_{ij}^{b}) b_j\\) and \\(N_{\\mathrm{FP}}(i) = \\sum_{j=1}^{i} H(x_{ij}^{b}) b_j\\) where \\(x_{ij}^{b} = s_j^b - s_i\\). Then, the resulting ranking error computed on \\(i\\)th positive is:" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.313, + 0.786, + 0.349 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\mathrm {R}} ^ {b} (i) = \\frac {N _ {\\mathrm {F P}} (i)}{\\operatorname {r a n k} (i)} = \\frac {\\sum_ {j = 1} ^ {i} H \\left(x _ {i j} ^ {b}\\right) b _ {j}}{\\sum_ {j = 1} ^ {i} H \\left(x _ {i j}\\right) + H \\left(x _ {i j} ^ {b}\\right) b _ {j}}, \\quad i \\in \\mathcal {P}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.353, + 0.788, + 0.461 + ], + "angle": 0, + "content": "Optimisation. Here, we need to define the primary terms, as the product between the error and the pmf following Identity Update [24]. Unlike the previous work, the weights of each prototype negative are not equal as a bucket includes varying number of negatives. Therefore, we use a weighted pmf while distributing the ranking error over negatives. Formally, if \\( j \\) is the \\( j \\)th prototype negative, then we define the pmf as \\( p(j^b|i) = b_j / N_{\\mathrm{FP}}(i) \\). The resulting primary term between the \\( i \\)th positive and \\( j^b \\)th negative is then:" + }, + { + "type": "equation", + "bbox": [ + 0.314, + 0.466, + 0.786, + 0.502 + ], + "angle": 0, + "content": "\\[\nL _ {i j} ^ {b} = \\frac {\\sum_ {j = 1} ^ {i} H \\left(x _ {i j} ^ {b}\\right) b _ {j}}{\\sum_ {j = 1} ^ {i} H \\left(x _ {i j}\\right) + H \\left(x _ {i j} ^ {b}\\right) b _ {j}} \\times \\frac {b _ {j}}{N _ {\\mathrm {F P}} (i)}, \\quad i \\in \\mathcal {P}, j \\in \\tilde {\\mathcal {N}} \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.63 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathcal{N}}\\) is the set of prototype negatives. However, we need the primary terms for the \\(i\\)th negative, not for the prototype ones. Given \\(L_{ij}^{b}\\), the gradients of the prototype negatives and actual negatives can easily be obtained following Identity Update. However, we still need to find the gradients of the actual negatives. To do so, we simply normalise the prototype gradient by its bucket size \\(b_{j}\\) and distribute the gradients to actual negatives, completing our method. In Theorem 1, we establish that our Bucketed AP Loss provides exactly the same gradients with AP Loss when \\(\\delta = 0\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.648, + 0.788, + 0.741 + ], + "angle": 0, + "content": "Bucketed RS Loss Given that we obtained Bucketed AP Loss, converting RS Loss into a bucketed form is more straightforward. This is because (i) the primary term and its gradients of RS Loss are equal to those of AP Loss if \\( i \\in \\mathcal{P} \\), \\( j \\in \\mathcal{N} \\) and we simply use \\( L_{ij}^{b} \\) in such cases; and (ii) if \\( i \\in \\mathcal{P} \\), \\( j \\in \\mathcal{P} \\), then an additional term is included in RS Loss to estimate the pairwise relations between positives (Eq. 11), which is not affected by bucketing." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.758, + 0.449, + 0.772 + ], + "angle": 0, + "content": "4.3 Theoretical Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "The first theorem ensures that the gradients provided by our loss functions are identical with their conventional counterparts under certain circumstances. The second states that our algorithm is theoretically faster than the conventional AP and RS Losses." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.519, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bucketed Ranking-based Losses" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.178 + ], + "angle": 0, + "content": "Theorem 1. Bucketed AP Loss and Bucketed RS Loss provide exactly the same gradients with AP and RS Losses respectively when \\(\\delta = 0\\)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.187, + 0.817, + 0.22 + ], + "angle": 0, + "content": "Theorem 2. Bucketed RS and Bucketed AP Losses have \\(\\mathcal{O}(\\max((|\\mathcal{P}| + |\\mathcal{N}|) \\log(|\\mathcal{P}| + |\\mathcal{N}|), |\\mathcal{P}|^2))\\) time complexity." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.229, + 0.785, + 0.261 + ], + "angle": 0, + "content": "As a result, we observe the same performance empirically in less amount of time. Proofs are provided in Supp. Mat." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.283, + 0.589, + 0.302 + ], + "angle": 0, + "content": "5 Bucketed Rank-Sort (BRS) DETR" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.784, + 0.39 + ], + "angle": 0, + "content": "While transformer-based detectors [8,14,16,37,38] have been providing the best performance in several object detection benchmarks, optimizing such detectors using ranking-based losses based on performance measures has not been investigated. Here, we address this gap by incorporating our BRS Loss into CoDETR [38] as the current SOTA detector on the common COCO benchmark [18]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.785, + 0.481 + ], + "angle": 0, + "content": "Co-DETR increases the number and variation of the positive examples in the transformer head by leveraging one-to-many assignment strategies in anchor-based detectors (such as ATSS [34] and Faster R-CNN [29]) as auxiliary heads. As the number of proposals is large in such detectors, they can easily provide more and diverse examples, resulting in better performance of the transformer head. Co-DETR is originally supervised with the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.491, + 0.785, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\sum_ {l = 1} ^ {L} \\left(\\tilde {\\mathcal {L}} _ {l} ^ {D e c} + \\lambda_ {1} \\sum_ {k = 1} ^ {K} \\tilde {\\mathcal {L}} _ {k, l} ^ {A u x} + \\lambda_ {2} \\sum_ {k = 1} ^ {K} \\mathcal {L} _ {k, l} ^ {A u x}\\right), \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.538, + 0.785, + 0.573 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) weigh each loss component. The first two components, \\(\\tilde{\\mathcal{L}}_l^{Dec}\\) and \\(\\tilde{\\mathcal{L}}_{k,l}^{Aux}\\), are the losses of the \\(l^{th}\\) decoder layer in the following form:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.586, + 0.785, + 0.601 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {c l s} \\mathcal {L} _ {c l s} + \\lambda_ {b b o x} \\mathcal {L} _ {b b o x} + \\lambda_ {I o U} \\mathcal {L} _ {I o U}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.612, + 0.785, + 0.707 + ], + "angle": 0, + "content": "and \\(\\mathcal{L}_{cls}\\), \\(\\mathcal{L}_{bbox}\\) and \\(\\mathcal{L}_{IoU}\\) are Focal Loss, L1 Loss and GIoU Loss respectively [38]. \\(\\tilde{\\mathcal{L}}_l^{Dec}\\) and \\(\\tilde{\\mathcal{L}}_{k,l}^{Aux}\\) differ in their input queries. That is, while \\(\\tilde{\\mathcal{L}}_l^{Dec}\\) follows standard DETR-based models [14, 37] with one-to-one matching of the queries, \\(\\tilde{\\mathcal{L}}_{k,l}^{Aux}\\) is computed based on positive anchors of \\(k^{th}\\) auxiliary head with one-to-many assignment. And \\(\\mathcal{L}_{k,l}^{Aux}\\) is the conventional loss of the \\(k^{th}\\) auxiliary head, e.g., the weighted sum of classification, localisation and centerness losses for ATSS [34]." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.708, + 0.785, + 0.739 + ], + "angle": 0, + "content": "In order to align the training and evaluation objectives better, we replace \\(\\tilde{\\mathcal{L}}_l^{Dec}\\) and \\(\\tilde{\\mathcal{L}}_{k,l}^{Aux}\\) in Eq. 16 by:" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.754, + 0.785, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {B R S} + \\lambda_ {b b o x} \\mathcal {L} _ {b b o x} + \\lambda_ {I o U} \\mathcal {L} _ {I o U}, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.8, + 0.843 + ], + "angle": 0, + "content": "and set \\(\\lambda_{bbox}\\) and \\(\\lambda_{IoU}\\) dynamically during training to \\(\\mathcal{L}_{BRS} / \\mathcal{L}_{bbox}\\) and \\(\\mathcal{L}_{BRS} / \\mathcal{L}_{IoU}\\) to simplify the hyperparameter tuning following Oksuz et al. [24]. Similarly, we replace the loss functions of the auxiliary ATSS [34] and Faster R-CNN [29] heads (\\(\\mathcal{L}_{k,l}^{Aux}\\) in Eq. 15) by our BRS Loss. This modification also aligns with the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Yavuz et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "objective of the auxiliary heads by the performance measure, hence results in more accurate auxiliary heads [24]. Furthermore, as using BRS Loss with Faster R-CNN does not require sampling thanks to its robustness to imbalance, no limitation is imposed on the number of positives in Faster R-CNN. Hence, the main aim of Co-DETR, that is to introduce more positive examples to the transformer head, is corroborated. As a result, our BRS DETR enables significantly efficient training compared to RS Loss also by improving the performance of Co-DETR." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.277, + 0.376, + 0.294 + ], + "angle": 0, + "content": "6 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.309, + 0.789, + 0.474 + ], + "angle": 0, + "content": "We analyze our contributions in three primary sections. First, we evaluate the effectiveness of our bucketing approach by contrasting it with RS Loss and AP Loss across different CNN-based visual detectors, on object detection and instance segmentation tasks. Next, we thoroughly examine how our bucketed RS Loss performs on transformer-based object detectors with different backbones. This is the first time a ranking-based loss is applied to transformer-based object detectors, thanks to the time-efficiency of our bucketing method. Finally, we present comparisons with score-based loss functions and other ranking-based loss functions in the literature. Our experiments clearly show that in terms of training time, accuracy and ease of tuning, our BRS Loss is more preferable over any existing loss function to train object detectors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.789, + 0.552 + ], + "angle": 0, + "content": "Dataset and Performance Measures. Unless otherwise noted, we train all models on COCO trainval35k (115k images) and test them on minival (5k images). We use COCO-style Average Precision (AP) and also report \\(\\mathrm{AP}_{50}\\), \\(\\mathrm{AP}_{75}\\) as the APs at IoUs 0.50 and 0.75; and \\(\\mathrm{AP_S}\\), \\(\\mathrm{AP_M}\\) and \\(\\mathrm{AP_L}\\) to present the accuracy on small, medium and large objects." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.573, + 0.725, + 0.589 + ], + "angle": 0, + "content": "6.1 Efficiency and Effectiveness of Bucketed Ranking Losses" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.719 + ], + "angle": 0, + "content": "To comprehensively analyze the effectiveness of our bucketing method, we design experiments on real-world data as well as synthetic data. For the former, we aim to present that our method significantly decreases the training time of the detection and segmentation methods (up to \\(\\sim 5\\times\\)). As training the entire detector does not isolate the runtime of the loss function, on which our main contribution is, we also design an experiment with synthetic data. This set of experiments show that our bucketing approach, in fact, decreases the loss function runtime by up to \\(40\\times\\), thereby resulting in shorter training time of the detectors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Experimental Setup. We follow the experimental setup used in RSLoss [24] only by replacing the loss function by our bucketed loss function to ensure comparability and consistency of results. We report our findings on efficiency and accuracy using the models trained on 4 Tesla A100 GPUs with a batch size of 16, i.e., 4 images/GPU. Unless stated otherwise, we train the models for 12 epochs and test them on images with a size of \\(1333 \\times 800\\) following the common convention [12, 24, 34]. In order to comprehensively demonstrate the efficiency of our approach, we use five different detectors: Faster R-CNN [29], Cascade R-CNN [2]," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.52, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bucketed Ranking-based Losses" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 1: Bucketed Losses (BRS, BAP) vs. RS & AP Losses on COCO object detection for object detectors. R-X: ResNet-X. Models with R-101 are trained 36 epochs." + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.186, + 0.773, + 0.379 + ], + "angle": 0, + "content": "
MethodDetectorBBLTimeAccuracy
\\(T_{iter}(s) \\downarrow\\)AP↑\\(\\mathrm{AP}_{50} \\uparrow\\)\\(\\mathrm{AP}_{75} \\uparrow\\)
Multi StageFaster R-CNNR-50RS0.5039.459.543.0
BRS\\(0.17 3.0x \\downarrow\\)39.559.542.8
R-101RS0.7547.367.451.2
BRS\\(0.38 2.0x \\downarrow\\)47.767.851.5
Cascade R-CNNR-50RS1.2841.158.744.1
BRS\\(0.24 5.3x \\downarrow\\)41.158.844.2
One StageATSSR-50AP0.3238.158.241.0
BAP\\(0.15 2.1x \\downarrow\\)38.558.641.3
RS0.3639.858.942.6
BRS\\(0.15 2.4x \\downarrow\\)39.858.842.9
PAAAP0.4537.354.341.2
BAP\\(0.30 1.5x \\downarrow\\)37.256.240.2
RS0.5740.858.844.6
BRS\\(0.30 1.9x \\downarrow\\)40.959.044.4
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.451 + ], + "angle": 0, + "content": "ATSS [34] and PAA [12] for object detection as well as Mask R-CNN [10] for instance segmentation." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.457, + 0.787, + 0.759 + ], + "angle": 0, + "content": "Bucketed Ranking Based Losses Significantly Decrease the Training Time with Similar Accuracy. Multi-stage detectors: Among multi-stage detectors, we train the commonly-used Faster R-CNN and Cascade R-CNN with RS Loss and BRS Loss, and present average iteration time as well as AP of the trained models. In order to evaluate our contribution comprehensively, we train Faster R-CNN also with a stronger setting, in which, we use ResNet-101, train it for 36 epochs using multi-scale training similar to [24]. The results are presented in Table 1, in which we can clearly see that BRS Loss consistently reduces the training time of all three detectors as well as preserves (or slightly improve) their performance. Especially for Cascade R-CNN, which is still a very popular and strong object detector along with its variants such as HTC [4], the training time decreases by \\(5.3 \\times\\). This is because the loss function is applied to the logits for three times based on the cascaded nature of this detector, and therefore, our contribution can be easily noticed. Moreover, it might seem that the efficiency gain decreases once the size of the backbone increases from R-50 to R-101 in Faster R-CNN. However, this is an expected result as the overall feature extraction time increases due to the larger number of parameters in the backbone, and hence, as we will discuss in the synthetic experiments, this is independent from our bucketing approach, operating on the loss function after the feature extraction." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.766, + 0.787, + 0.84 + ], + "angle": 0, + "content": "One-stage detectors: To show that our gains generalize to one-stage detectors and to AP Loss, we train the common ATSS and PAA detectors with our BAP and BRS Losses, and compare them with AP and RS Losses respectively. Table 1 validates our previous claims: Our bucketed losses obtain similar performance in around half of the training time required for AP and RS Losses." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Yavuz et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.492, + 0.187 + ], + "angle": 0, + "content": "Table 2: Comparison with RS Loss on instance segmentation task on COCO val using Mask R-CNN. BB: Backbone." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.2, + 0.484, + 0.27 + ], + "angle": 0, + "content": "
BBLEfficiencyAccuracy
\\(T_{iter}(s)\\)↓AP↑\\(AP_{50}\\)↑\\(AP_{75}\\)↑
R-50RS0.5636.357.238.8
BRS0.24 2.3x ↓36.257.238.8
R-101RS0.5940.261.843.5
BRS0.27 2.2x ↓40.362.043.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.145, + 0.785, + 0.187 + ], + "angle": 0, + "content": "Table 3: Comparison on different inst. segm. datasets using Mask R-CNN. \\(\\mathrm{AP}_{75}\\) is N/A as it is not used for Cityscapes." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.2, + 0.787, + 0.269 + ], + "angle": 0, + "content": "
DatasetLEfficiencyAccuracy
\\(T_{iter}(s) \\downarrow\\)AP↑\\(AP_{50} \\uparrow\\)\\(AP_{75} \\uparrow\\)
CityscapesRS0.4343.568.1N/A
BRS\\(0.19 2.3x \\downarrow\\)43.367.7N/A
LVISRS0.8725.639.227.3
BRS\\(0.35 2.5x \\downarrow\\)25.839.627.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.284, + 0.785, + 0.312 + ], + "angle": 0, + "content": "Table 4: BRS Loss further improves performance with similar training budget. We report mask AP for instance segmentation" + }, + { + "type": "table", + "bbox": [ + 0.333, + 0.325, + 0.67, + 0.387 + ], + "angle": 0, + "content": "
TaskDetectorLAP↑AP50↑AP75↑
Object Det.C.R-CNNRS41.158.744.1
BRS42.3+1.260.245.2
Inst.Seg.M.R-CNNRS36.357.238.8
BRS37.3+1.058.440.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.785, + 0.537 + ], + "angle": 0, + "content": "Instance Segmentation: Given that our loss functions is efficient in object detectors, one simple extension is to see their generalization to instance segmentation methods. To show that, we train Mask R-CNN, a common baseline, with our BRS Loss on three different datasets from various domains: (i) COCO (Table 2), (ii) Cityscapes [7] as an autonomous driving dataset (Table 3) and LVIS [9] a long-tailed dataset with more than 1K classes (Table 3). The results confirm our previous finding that our BRS Loss decreases the training time significantly by up to \\(\\sim 2.5\\times\\) compared to RS Loss by preserving its accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.786, + 0.718 + ], + "angle": 0, + "content": "Given Similar Training Budget, Bucketed Ranking Based Loss Functions Significantly Improve the Accuracy. There might be several benefits of reducing the training time of the detector as more GPU time is saved. Here, we show a use-case in which we ask the following question: what would happen if we allocate the similar amount of training time to both bucketed and non-bucketed losses? To answer that, we train Cascade R-CNN, a detector, and Mask R-CNN, an instance segmentation method, using our BRS Loss. Considering our speed-ups on these models \\((6.0\\times\\) and \\(2.3\\times)\\), we simply increased their training epochs from 12 to 36 and 27 respectively. Note that, for Cascade R-CNN, we still spend significantly less amount of training time. Table 4 shows that models trained with the BRS Loss outperform (with \\(+1.2\\) and \\(1.0\\mathrm{AP}\\)) their counterparts trained with the RS Loss in both cases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Experiments on Synthetic Data. Up to now, to measure any speed-up, we considered the entire pipeline and did not decouple the loss computation step. However, we only improve the loss computation step. Hence, an analysis focusing only on this part can reveal our contribution more clearly. Therefore, we design a simple experiment using synthetic data. Particularly, we randomly generate \\( L \\) logits such that \\( m\\% \\) of these logits are positive (see Supp.Mat. for data generation details). To cover the wide range of existing detectors, we set \\( L \\in \\{10K, 100K, 1M\\} \\) and \\( m \\in \\{0.1, 1.0, 2.0, 5.0\\} \\). Different choices for \\( L \\) and \\( m \\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.519, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bucketed Ranking-based Losses" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.493, + 0.2 + ], + "angle": 0, + "content": "Table 5: Comparison of our BRS-DETR with Co-DETR (trained with its original loss function) on different backbones on COCO val set." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.214, + 0.497, + 0.293 + ], + "angle": 0, + "content": "
BackboneDetectorAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_s \\)\\( AP_m \\)\\( AP_1 \\)
ResNet50Co-DETR49.367.254.032.152.663.8
BRS-DETR50.167.454.631.953.965.0
Swin-TCo-DETR51.769.656.434.454.966.8
BRS-DETR52.369.557.134.655.768.0
Swin-LCo-DETR56.975.562.640.161.273.3
BRS-DETR57.275.062.539.461.774.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "Table 6: Comparison of our BRS-DETR with DETR variants (trained with their original loss functions) w ResNet-50 on COCO val set. Qu: Queries. Ep: Epochs." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.215, + 0.785, + 0.291 + ], + "angle": 0, + "content": "
DetectorQuEpAP\\( {\\mathrm{{AP}}}_{50} \\)\\( {\\mathrm{{AP}}}_{75} \\)\\( {\\mathrm{{AP}}}_{\\mathrm{s}} \\)\\( {\\mathrm{{AP}}}_{\\mathrm{m}} \\)\\( {\\mathrm{{AP}}}_{1} \\)
DETR [14]10030042.062.444.220.545.861.1
DN-DETR [15]3005048.667.452.731.052.063.7
DINO [3]9001249.466.953.832.352.563.9
H-DETR [11]3001248.766.452.931.251.563.5
Co-DETR [38]3001249.367.254.032.152.663.8
BRS-DETR3001250.167.454.631.953.965.0
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.785, + 0.404 + ], + "angle": 0, + "content": "correspond to the logit distributions of common object detectors. We compute BAP and AP Loss on this synthetic data and compare their runtime to focus on loss computation time. We observed that our BAP Loss has a significant, up to \\(\\sim 40\\times\\), speed-up compared to AP Loss. The speed-up becomes more pronounced with increasing number of logits. Supp. Mat. presents detailed results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.441, + 0.571, + 0.456 + ], + "angle": 0, + "content": "6.2 The Effectiveness of Our BRS-DETR" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.785, + 0.514 + ], + "angle": 0, + "content": "While we incorporate our BRS Loss into Co-DETR, we use all original settings unless otherwise noted. Please refer to Supp.Mat. for the details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.787, + 0.622 + ], + "angle": 0, + "content": "BRS-DETR Outperforms Existing Transformer-based Detectors Consistently. Here, we first employ ResNet-50 backbone as it has been used by many detectors, enabling us to compare our method with many different DETR-based manner in a consistent manner. In this fair comparison setting, our BRS-DETR outperforms all existing DETR variants reaching 50.1 AP as shown in Table 6. For example, our BRS DETR outperforms DN-DETR by 1.5AP with less training epochs and improves Co-DETR as we discuss next." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.787, + 0.732 + ], + "angle": 0, + "content": "Training Co-DETR with RS Loss takes 6x less time compared to using RS Loss. Finally, we also compare the training efficiencies of BRS Loss and RS Loss on Co-DETR. We observe that our BRS Loss decreases the training time of Co-DETR by \\(6.0 \\times\\) (from 4.14s per iteration to 0.69s) in comparison to RS Loss. We note that this is the largest training time gain of our BRS Loss. This is because Co-DETR consists of multiple transformer-based as well as auxiliary heads, requiring multiple loss estimations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.787, + 0.842 + ], + "angle": 0, + "content": "BRS-DETR Improves Co-DETR over Different Backbones Consistently. In order to show the effectiveness of our BRS Loss, we compare our results with Co-DETR in Table 5 using different backbones. We note that we improve the baseline Co-DETR consistently in all settings. For example, our improvement on ResNet-50 is 0.8AP, which is a notable improvement. However, as the backbone gets larger (e.g., Swin-L [20]), our gains decrease, which is expected and commonly observed in the literature, e.g., [12]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Yavuz et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.493, + 0.215 + ], + "angle": 0, + "content": "Table 7: Comparison with the score-based losses. CE: Cross-entropy. #H: Number of hyperparameters. Ours has very similar \\( T_{iter}(s) \\) also by being more accurate and simple-to-tune." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.227, + 0.502, + 0.311 + ], + "angle": 0, + "content": "
DetectorLTiter(s) ↓AP ↑#H
Faster R-CNNCE+L10.1437.69
RS0.5039.43
BRS (Ours)0.1739.53
ATSSFocal L. +G IoU0.1439.35
RS0.3639.81
BRS (Ours)0.1539.81
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "Table 8: Comparison with other ranking-based losses. Our approach performs better on both accuracy and efficiency." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.202, + 0.784, + 0.309 + ], + "angle": 0, + "content": "
DetectorL\\(T_{iter}(s) \\downarrow\\)AP ↑#H
Faster R-CNNaLRP [23]0.2837.43
BRS (Ours)0.1739.53
ATSSAP [5]0.3238.15
DR [28]0.2038.15
aLRP [23]0.3237.71
BRS (Ours)0.1539.81
RetinaNetDR [28]0.2937.45
BRS (Ours)0.2338.31
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.343, + 0.782, + 0.358 + ], + "angle": 0, + "content": "6.3 Comparison with Score-based and Other Ranking-based Losses" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.788, + 0.603 + ], + "angle": 0, + "content": "In previous sections, we showed the efficiency of our loss functions compared to AP and RS Losses as their counterparts. Here, we compare our bucketed losses with the score based losses, i.e., Cross-entropy and Focal Loss, and other ranking based losses including DR Loss [28] and aLRP Loss [23]. To do so, in Tables 7 and 8, we report average training time, AP for accuracy and number of hyperparameters to capture the tuning simplicity of the loss functions. Compared to score-based loss functions in Table 7, our BRS Loss improves performance of both approaches, is significantly simpler-to-tune with a low number of hyperparameters and has similar training iteration time. Hence, our BRS Loss is either superior or on par with existing score-based losses. As for other ranking-based losses in Table 8, our BRS Loss is the most efficient and accurate ranking-based loss function for all three detectors. As an example, compared to ATSS trained with DR Loss, our BRS Loss yields 1.8AP better accuracy with \\(25\\%\\) less training time and significantly less number of hyperparameters. Therefore, our bucketed loss functions are now promising alternatives to train object detectors." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.634, + 0.36, + 0.65 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we introduced a novel method to improve the efficiency of ranking-based loss functions by binning negatives into buckets and implementing positive-negative pairwise comparisons between positives and buckets of negatives. We showed that this method improves the computational complexity to a level where pairwise comparisons can be stored as a matrix in memory and running time becomes very close to that of score-based loss functions. Also, for the first time, we integrated a ranking-based loss to Co-DETR, a DETR-based detector, which was possible thanks to our method's lower complexity, and reported improvements on different backbones. Our comprehensive experiments comprising 2 different tasks, 3 different dataset and 7 different detectors - all with positive outcomes - show the general applicability of our approach." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.52, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bucketed Ranking-based Losses" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.18, + 0.788, + 0.257 + ], + "angle": 0, + "content": "We gratefully acknowledge the computational resources provided by TÜBITAK ULAKBIM High Performance and Grid Computing Center (TRUBA), METU Center for Robotics and Artificial Intelligence (METU-ROMER) and METU Image Processing Laboratory. Dr. Akbas is supported by the \"Young Scientist Awards Program (BAGEP)\" of Science Academy, Türkiye." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.284, + 0.323, + 0.3 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.36 + ], + "angle": 0, + "content": "1. Brown, A., Xie, W., Kalogeiton, V., Zisserman, A.: Smooth-ap: Smoothing the path towards large-scale image retrieval. In: European Conference on Computer Vision. pp. 677-694. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.362, + 0.785, + 0.403 + ], + "angle": 0, + "content": "2. Cai, Z., Vasconcelos, N.: Cascade R-CNN: Delving into high quality object detection. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.785, + 0.431 + ], + "angle": 0, + "content": "3. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.785, + 0.487 + ], + "angle": 0, + "content": "4. Chen, K., Pang, J., Wang, J., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Shi, J., Ouyang, W., Loy, C.C., Lin, D.: Hybrid task cascade for instance segmentation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.489, + 0.785, + 0.53 + ], + "angle": 0, + "content": "5. Chen, K., Li, J., Lin, W., See, J., Wang, J., Duan, L., Chen, Z., He, C., Zou, J.: Towards accurate one-stage object detection with ap-loss. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.532, + 0.785, + 0.572 + ], + "angle": 0, + "content": "6. Chen, K., Lin, W., li, J., See, J., Wang, J., Zou, J.: Ap-loss for accurate one-stage object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) pp. 1-1 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.574, + 0.785, + 0.627 + ], + "angle": 0, + "content": "7. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.63, + 0.785, + 0.683 + ], + "angle": 0, + "content": "8. Dai, Z., Cai, B., Lin, Y., Chen, J.: Unsupervised pre-training for detection transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence p. 1-11 (2022). https://doi.org/10.1109/tpami.2022.3216514, http://dx.doi.org/10.1109/TPAMI.2022.3216514" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.686, + 0.785, + 0.726 + ], + "angle": 0, + "content": "9. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.728, + 0.785, + 0.754 + ], + "angle": 0, + "content": "0. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask R-CNN. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.783 + ], + "angle": 0, + "content": "1. Jia, D., Yuan, Y., He, H., Wu, X., Yu, H., Lin, W., Sun, L., Zhang, C., Hu, H.: Detrs with hybrid matching (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.812 + ], + "angle": 0, + "content": "2. Kim, K., Lee, H.S.: Probabilistic anchor assignment with iou prediction for object detection. In: The European Conference on Computer Vision (ECCV) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "3. Law, H., Deng, J.: Cornernet: Detecting objects as paired keypoints. In: The European Conference on Computer Vision (ECCV) (2018)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "Yavuz et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "14. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13619-13627 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.787, + 0.23 + ], + "angle": 0, + "content": "15. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.787, + 0.271 + ], + "angle": 0, + "content": "16. Li*, L.H., Zhang*, P., Zhang*, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., Chang, K.W., Gao, J.: Grounded language-image pretraining. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.787, + 0.311 + ], + "angle": 0, + "content": "17. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 42(2), 318-327 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.311, + 0.787, + 0.352 + ], + "angle": 0, + "content": "18. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common Objects in Context. In: The European Conference on Computer Vision (ECCV) (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.352, + 0.787, + 0.393 + ], + "angle": 0, + "content": "19. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S.E., Fu, C., Berg, A.C.: SSD: single shot multibox detector. In: The European Conference on Computer Vision (ECCV) (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.787, + 0.42 + ], + "angle": 0, + "content": "20. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.787, + 0.447 + ], + "angle": 0, + "content": "21. Mohapatra, P., Jawahar, C., Kumar, M.P.: Efficient optimization for average precision svm. Advances in Neural Information Processing Systems 27 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.787, + 0.488 + ], + "angle": 0, + "content": "22. Mohapatra, P., Rolínek, M., Jawahar, C., Kolmogorov, V., Pawan Kumar, M.: Efficient optimization for rank-based loss functions. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.528 + ], + "angle": 0, + "content": "23. Oksuz, K., Cam, B.C., Akbas, E., Kalkan, S.: A ranking-based, balanced loss function unifying classification and localisation in object detection. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.787, + 0.569 + ], + "angle": 0, + "content": "24. Oksuz, K., Cam, B.C., Akbas, E., Kalkan, S.: Rank & sort loss for object detection and instance segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3009-3018 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.787, + 0.609 + ], + "angle": 0, + "content": "25. Oksuz, K., Cam, B.C., Kahraman, F., Baltaci, Z.S., Kalkan, S., Akbas, E.: Mask-awareiou for anchor assignment in real-time instance segmentation. In: The British Machine Vision Conference (BMVC) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.787, + 0.651 + ], + "angle": 0, + "content": "26. Oksuz, K., Cam, B.C., Kalkan, S., Akbas, E.: Imbalance problems in object detection: A review. IEEE transactions on pattern analysis and machine intelligence 43(10), 3388-3415 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.787, + 0.691 + ], + "angle": 0, + "content": "27. Pang, J., Chen, K., Shi, J., Feng, H., Ouyang, W., Lin, D.: Libra R-CNN: Towards balanced learning for object detection. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.787, + 0.718 + ], + "angle": 0, + "content": "28. Qian, Q., Chen, L., Li, H., Jin, R.: Dr loss: Improving object detection by distributional ranking. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.787, + 0.759 + ], + "angle": 0, + "content": "29. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 39(6), 1137-1149 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.813 + ], + "angle": 0, + "content": "30. Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I., Savarese, S.: Generalized intersection over union: A metric and a loss for bounding box regression. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.841 + ], + "angle": 0, + "content": "31. Rosenblatt, F.: The perceptron: A probabilistic model for information storage and organization in the brain. Psychological Review pp. 65-386 (1958)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.52, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Bucketed Ranking-based Losses" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "32. Shrivastava, A., Gupta, A., Girshick, R.: Training region-based object detectors with online hard example mining. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.245 + ], + "angle": 0, + "content": "33. Sun, P., Zhang, R., Jiang, Y., Kong, T., Xu, C., Zhan, W., Tomizuka, M., Li, L., Yuan, Z., Wang, C., Luo, P.: SparseR-CNN: End-to-end object detection with learnable proposals. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.245, + 0.788, + 0.286 + ], + "angle": 0, + "content": "34. Zhang, S., Chi, C., Yao, Y., Lei, Z., Li, S.Z.: Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.328 + ], + "angle": 0, + "content": "35. Zhang, X., Wan, F., Liu, C., Ji, R., Ye, Q.: Freeanchor: Learning to match anchors for visual object detection. In: Advances in Neural Information Processing Systems (NeurIPS) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.37 + ], + "angle": 0, + "content": "36. Zheng, Z., Wang, P., Liu, W., Li, J., Rongguang, Y., Dongwei, R.: Distance-iou loss: Faster and better learning for bounding box regression. In: AAAI Conference on Artificial Intelligence (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.412 + ], + "angle": 0, + "content": "37. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. ArXiv abs/2010.04159 (2020), https://api-semanticscholar.org/CorpusID:222208633" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.412, + 0.788, + 0.439 + ], + "angle": 0, + "content": "38. Zong, Z., Song, G., Liu, Y.: Detrs with collaborative hybrid assignments training (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.439 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Bucketed Ranking-based Losses for Efficient Training of Object Detectors/7957777c-4d0a-4feb-89c6-b57e75289486_origin.pdf b/2024/Bucketed Ranking-based Losses for Efficient Training of Object Detectors/7957777c-4d0a-4feb-89c6-b57e75289486_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..15d4a0403a65b6512a33cd92cda5974dfc974581 --- /dev/null +++ b/2024/Bucketed Ranking-based Losses for Efficient Training of Object Detectors/7957777c-4d0a-4feb-89c6-b57e75289486_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0790b6a889d7433c0c708c88c80b9c5ede8486bb160ebee3e7f54bef26686cea +size 731682 diff --git a/2024/Bucketed Ranking-based Losses for Efficient Training of Object Detectors/full.md b/2024/Bucketed Ranking-based Losses for Efficient Training of Object Detectors/full.md new file mode 100644 index 0000000000000000000000000000000000000000..70150d4c96b16a35f2cc84abd018b53585a25956 --- /dev/null +++ b/2024/Bucketed Ranking-based Losses for Efficient Training of Object Detectors/full.md @@ -0,0 +1,340 @@ +# Bucketed Ranking-based Losses for Efficient Training of Object Detectors + +Feyza Yavuz1, Baris Can Cam1, Adnan Harun Dogan1, Kemal Oksuz2, Emre Akbas1,3,*, and Sinan Kalkan1,3,* + +$^{1}$ Department of Computer Engineering, METU, Ankara, Türkiye $^{2}$ Five AI Ltd., United Kingdom + +3 METU ROMER Robotics Center, Ankara, Türkiye * Equal senior authorship + +Abstract. Ranking-based loss functions, such as Average Precision Loss and Rank&Sort Loss, outperform widely used score-based losses in object detection. These loss functions better align with the evaluation criteria, have fewer hyperparameters, and offer robustness against the imbalance between positive and negative classes. However, they require pairwise comparisons among $P$ positive and $N$ negative predictions, introducing a time complexity of $\mathcal{O}(PN)$ , which is prohibitive since $N$ is often large (e.g., $10^8$ in ATSS). Despite their advantages, the widespread adoption of ranking-based losses has been hindered by their high time and space complexities. In this paper, we focus on improving the efficiency of ranking-based loss functions. To this end, we propose Bucketed Ranking-based Losses which group negative predictions into $B$ buckets ( $B \ll N$ ) in order to reduce the number of pairwise comparisons so that time complexity can be reduced. Our method enhances the time complexity, reducing it to $\mathcal{O}(\max(N\log(N),P^2))$ . To validate our method and show its generality, we conducted experiments on 2 different tasks, 3 different datasets, 7 different detectors. We show that Bucketed Ranking-based (BR) Losses yield the same accuracy with the unbucketed versions and provide $2 \times$ faster training on average. We also train, for the first time, transformer-based object detectors using ranking-based losses, thanks to the efficiency of our BR. When we train CoDETR, a state-of-the-art transformer-based object detector, using our BR Loss, we consistently outperform its original results over several different backbones. Code will be released. + +Keywords: Object Detection $\cdot$ Ranking Losses $\cdot$ Efficient Ranking Losses + +# 1 Introduction + +Training an object detector incurs many challenges unique to object detection, including (i) assigning dynamically a large number of object hypotheses with the objects [2, 12, 13, 25, 29, 33, 34], (ii) sampling among these hypotheses to ensure that the background class does not dominate training [19, 27, 29, 32] and + +![](images/c805ee1cbcd735452d8b8b83680e27b7722fce63da5db9af6fe80d48307ebeb3.jpg) +(a) Comparisons in AP Loss + +![](images/7a617a6d94d316ca388a4502024453c37d0122189af09f0f7a40290b7354215b.jpg) +(b) Comparisons in Bucketed AP Loss (ours) + +![](images/6c56eb8e3f6cbc738baa91872c80d850eb019b86c314cc556fcb9af123a3a762.jpg) +(c) Average Precision (AP) vs. average training iteration time for baselines, RS Loss and Bucketed RS Loss (ours) +Fig. 1: (a) Existing ranking-based losses (i.e., AP Loss [5,6]) incur significant overhead owing to pairwise comparisons between positives and negatives. (b) We propose bucketing negatives to decrease the number of comparisons, and hence the complexity. Under certain assumptions, our bucketing approach provides equal gradients with conventional ranking-based losses such as AP Loss in (a). (c) Accuracy and efficiency comparison. Our BRS formulation facilitates faster (between $1.9 \times$ and $6.0 \times$ ) training of visual detectors with similar AP. + +(iii) minimizing a multi-task objective function [17,23,24,29]. While the choices on assignment and sampling generally vary depending on the trained object detector, it is very common to combine Cross Entropy or Focal Loss [17] with a regression loss [30,36] as the multi-task loss function. Recently proposed ranking-based loss functions [5,6,23,24] offer an alternative approach for addressing these challenges by formulating the training objective based on the rank of positive examples over negative examples. + +Benefits of ranking-based losses. First, they are inherently robust to imbalance [23] and hence, do not require any sampling mechanism under very challenging scenarios [24], e.g., even when the background foreground ratio is 10K for LVIS [9]. Second, they are shown to generalize over different detectors with diverse architectures - with the exception of transformer-based ones [14,37,38], since ranking-based losses further slow down the training of transformer-based detectors, which we address in this paper. Such losses also offer significant performance gain over their score-based counterparts, and having less hyperparameters, they are much easier to tune [23,24]. + +The drawback of ranking-based losses. Compared to score-based losses, ranking-based losses are more inefficient as the ranking operation inherently requires each pair of object hypotheses to be compared against each other (Fig. 1(a)), inducing a quadratic complexity on the large number of object hypotheses (e.g., $10^{8}$ for ATSS [34] on COCO [18]). As a result, vectorized implementations for parallel GPU processing are infeasible as such large matrices (e.g., with $\sim$ $10^{16}$ entries for ATSS) do not fit into GPU memories. This has driven researchers towards alternative ways of computing these matrices, which, in the end, saves from the storage complexity but results in more inefficient algorithms [5,6]. + +Our approach. The main cause for the aforementioned complexity is the background class, i.e., negatives, forming up to $99.9\%$ of the hypotheses [26]. Considering this, we introduce a novel bucketing approach (Fig. 1(b)) on these negatives to address the main shortcoming of the ranking-based loss functions. Specifically, we first sort all examples based on their scores and bucket the negatives located between successive positives. The negatives within the same bucket are treated as a single example and represented by their average score. We then incorporate our bucketing approach into AP Loss [5,6] and RS Loss [24], and introduce novel mechanisms for calculating the gradients correctly. + +The significance of our approach. Theoretically, we prove that our bucketing approach provides the same gradients as the original ranking-based loss. Practically, we show we can reduce loss computation time by up to $\sim 40\times$ and training time of a detector by up to $6\times$ (Fig. 1(c)), addressing the main drawback of ranking-based loss functions. Hence, we show that the gap between the score-based and ranking-based loss functions in terms of training time vanishes, enabling us to train strong transformer-based object detectors such as Co-DETR using our loss functions. As ranking-based loss functions now surpass score-based ones or perform on par in every aspect such as accuracy, tuning simplicity, robustness to imbalance and training time, we believe that our work will make their prevalence more dominant in the upcoming object detectors. + +Contributions. Our contributions can be summarized as follows: + +- We propose a novel bucketing approach to improve the efficiency of computationally-expensive ranking-based losses to train object detectors. Theoretically, our approach yields the same gradients with the original ranking-based losses while decreasing their time complexity. Practically, we enable up to $6 \times$ faster training using our bucketing approach with no accuracy loss. + +- For the first time, we incorporate ranking-based loss functions to transformer-based detectors. Specifically, we construct BRS-DETR by replacing the training objective of the state-of-the-art transformer-based object detector CoDETR [38] by our Bucketed RS (BRS) Loss. + +- Our comprehensive experiments on detection and instance segmentation on 3 different challenging datasets, 5 backbones and 7 detectors show the effectiveness and generalizability of our approach. Our BRS-DETR reaches 50.1 AP on COCO val set with only 12 epochs and ResNet-50 backbone, outperforming all existing detectors using the same backbone and 300 queries. BRS-DETR also outperforms CoDETR with Swin-T and Swin-L backbones, reaching 57.2 AP on the COCO dataset. + +# 2 Related Work + +Although score-based loss functions, such as cross-entropy and focal loss, are widely used to train both CNN-based [2, 17, 35] and transformer-based object detectors [14, 37, 38], ranking-based losses for visual detection offer advantages. The primary advantage is their robustness to class imbalance [23]. They yield on par or better performance compared to score-based losses, without the need + +for extensive tuning of class weights or multistage sampling techniques. Another advantage they offer is the ease of balancing losses among different tasks. In AP Loss [5, 6], aLRP Loss [23] and RS Loss [24], the authors use the ratio of classification and regression losses to balance these tasks. As a result, ranking-based losses have much fewer hyperparameters during training, compared to score-based ones. + +Other Ranking-based Losses. The literature has explored other ranking-based loss functions with different goals and assumptions. For example, DR Loss [28] constrains a margin to be satisfied between positives and negatives, not taking the recall into account. Smooth AP [1] introduces a smooth, differentiable approximation of AP, with the assumption that $|\mathcal{N}|$ is not large. Several studies focused on improving the efficiency of ranking-based losses for use in Support Vector Machines (SVMs) [21, 22]. However, these solutions have been limited to training linear SVMs and hence, problems that can be solved using linear functions. In this paper, we focus on improving ranking-based loss functions [5, 6, 24] that can be used for training complex deep networks such as object detectors. + +Comparative Summary. Ranking-based training objective for visual detectors proved their superiority over score-based loss functions in visual detection tasks [5, 6, 23, 24]. However, due to the requirement of pairwise ranking their time complexity is higher than the score-based alternatives. Such a disadvantage of ranking-based loss functions prevents their applicability to larger visual detectors (i.e., transformer-based object detectors). + +# 3 Background on Ranking-based Losses + +Ranking-based losses for object detection rely on pairwise comparisons between the scores of different detections to determine the rank of a detection among positives and negatives. Denoting the score of the $i^{th}$ detection by $s_i$ , we can compare the scores of two detections $i$ and $j$ with a simple difference transform: $x_{ij} = s_j - s_i$ . By counting the number of instances with $x_{ij} > 0$ , we can determine $i^{th}$ detection's rank among positives $(\mathcal{P})$ and all detections (positives and negatives: $\mathcal{P} \cup \mathcal{N}$ ) as follows: + +$$ +\operatorname {r a n k} ^ {+} (i) = \sum_ {j \in \mathcal {P}} \bar {H} \left(x _ {i j}\right), \text {a n d} \operatorname {r a n k} (i) = \sum_ {j \in \mathcal {P} \cup \mathcal {N}} \bar {H} \left(x _ {i j}\right), \tag {1} +$$ + +where $\bar{H}(x)$ is one if $x > 0$ and zero otherwise. Since $\mathrm{d}\bar{H}(x) / \mathrm{d}x$ is either infinite (at $x = 0$ ) or zero (for $x \neq 0$ ), a smoothed version is used ( $\delta$ : a hyper-parameter): + +$$ +H (x) = \left\{ \begin{array}{l l} 0, & x < - \delta \\ \frac {x}{2 \delta} + 0. 5, & - \delta \leq x \leq \delta \\ 1, & \delta < x \end{array} \right. \tag {2} +$$ + +# 3.1 Revisiting Average Precision (AP) Loss + +Given the definitions in Eq. 1, AP Loss [5,6] can be defined as: + +$$ +\mathcal {L} _ {A P} = 1 - \mathrm {A P} = 1 - \frac {1}{| \mathcal {P} |} \sum_ {i \in \mathcal {P}} \operatorname {p r e c i s i o n} (i) = 1 - \frac {1}{| \mathcal {P} |} \sum_ {i \in \mathcal {P}} \frac {\operatorname {r a n k} ^ {+} (i)}{\operatorname {r a n k} (i)}. \tag {3} +$$ + +Chen et al. [5,6] simplified Eq. 3 by rewriting it in terms of pairwise comparisons between positive and negative detections: + +$$ +\mathcal {L} _ {A P} = \frac {1}{| \mathcal {P} |} \sum_ {i \in \mathcal {P}} \sum_ {j \in \mathcal {N}} \frac {H (x _ {i j})}{\operatorname {r a n k} (i)} = \frac {1}{| \mathcal {P} |} \sum_ {i \in \mathcal {P}} \sum_ {j \in \mathcal {N}} L _ {i j} ^ {A P}, \tag {4} +$$ + +where $L_{ij}^{AP} = H(x_{ij}) / \mathrm{rank}(i)$ is called the primary term. Note that $L_{ij}^{AP}$ is non-differentiable since the step function $(H(\cdot))$ applied on $x_{ij}$ is non-differentiable, which will be discussed next. + +# 3.2 Revisiting Identity Update + +Oksuz et al. [24] showed that various ranking-based losses (including AP Loss in Eqs. 3 & 4) can be written in a general form as: + +$$ +\mathcal {L} = \frac {1}{Z} \sum_ {i \in \mathcal {P} \cup \mathcal {N}} (\ell (i) - \ell^ {*} (i)), \tag {5} +$$ + +where $\ell(i)$ is the ranking-based error (e.g., precision error) computed on the $i$ th detection, $\ell^{*}(i)$ is the target ranking-based error (the lowest error possible) and $Z$ is the normalization constant. The loss in Eq. 5 can be computed and optimized as follows [24]: + +1. Computation of the Loss. First, each pair of logits $(s_i$ and $s_j)$ are compared by calculating their difference transforms as $x_{ij} = s_j - s_i$ . With the step function $H(\cdot)$ (Eq. 2), the number of detections higher than $s_i$ (and therefore its precision error, $\ell(i)$ ) can be easily calculated (see Sect. 3.1). Given $x_{ij}$ , the loss can be rewritten and calculated in terms of primary terms $L_{ij}$ as $\mathcal{L} = \frac{1}{Z}\sum_{i\in \mathcal{P}\cup \mathcal{N}}\sum_{j\in \mathcal{P}\cup \mathcal{N}}L_{ij}$ , by taking: + +$$ +L _ {i j} = \left(\ell (i) - \ell^ {*} (i)\right) p (j | i), \tag {6} +$$ + +where $p(j|i)$ is a probability mass function (pmf) that distributes the error computed on the $i$ th example over the $j$ th example in order to determine the pairwise primary term $L_{ij}$ . $p(j|i)$ is commonly taken as a uniform distribution [23, 24]. + +2. Optimization of the Loss. The gradient of the primary term wrt. to the difference transform $(\partial L_{ij} / \partial x_{ij})$ is non-differentiable. Denoting this term by $\Delta x_{ij}$ , we have [5]: + +$$ +\frac {\partial \mathcal {L}}{\partial s _ {i}} = \sum_ {j, k} \frac {\partial \mathcal {L}}{\partial L _ {j k}} \Delta x _ {j k} \frac {\partial x _ {j k}}{\partial s _ {i}} = \frac {1}{Z} \left(\sum_ {j} \Delta x _ {j i} - \sum_ {j} \Delta x _ {i j}\right). \tag {7} +$$ + +Therefore, optimizing a ranking-based loss function reduces to determining $\Delta x_{ij}$ . Inspired by Chen et al. [5], Oksuz et al. [24] employ Perceptron Learning [31] and show that $\Delta x_{ij}$ in Eq. 7 is simply the primary term itself: $\Delta x_{ij} = -(L_{ij}^{*} - L_{ij}) = -(0 - L_{ij}) = L_{ij}$ , hence the name Identity Update. Plugging this into Eq. 7 yields (see Oksuz et al. [24] for the steps of the derivation): + +$$ +\frac {\partial \mathcal {L}}{\partial s _ {i}} = \frac {1}{Z} \left(\sum_ {j} L _ {j i} - \sum_ {j} L _ {i j}\right). \tag {8} +$$ + +Therefore, both computing and optimizing the loss reduces determining the primary term $L_{ij}$ . + +# 3.3 Revisiting AP Loss and RS Loss with Identity Update + +AP Loss [5, 6]. For defining AP Loss with identity update, the errors in Eq. 5 can be derived as $\ell_{\mathrm{R}}(i) = \frac{N_{\mathrm{FP}}(i)}{\mathrm{rank}(i)}, \ell_{\mathrm{R}}^{*}(i) = 0$ and $Z = |\mathcal{P}|$ . Furthermore, defining $p_R(j|i)$ as a uniform pmf, that is $p_R(j|i) = \frac{H(x_{ij})}{N_{\mathrm{FP}}(i)}$ , we can obtain the primary terms of AP Loss using Identity Update, completing the derivation for computation and optimization (Eq. 8): + +$$ +L _ {i j} ^ {A P} = \left\{ \begin{array}{l l} \left(\ell_ {\mathrm {R}} (i) - \ell_ {\mathrm {R}} ^ {*} (i)\right) p _ {R} (j | i), & \text {f o r} i \in \mathcal {P}, j \in \mathcal {N} \\ 0, & \text {o t h e r w i s e ,} \end{array} \right. \tag {9} +$$ + +RS Loss [24]. RS Loss includes an additional sorting objective $(\ell_S(i))$ to promote better-localized positives to be ranked higher than other positives: + +$$ +\mathcal {L} _ {\mathrm {R S}} = \frac {1}{| \mathcal {P} |} \sum_ {i \in \mathcal {P}} \left[ \underbrace {\left(\ell_ {\mathrm {R}} (i) - \ell_ {\mathrm {R}} ^ {*} (i)\right)} _ {\text {r a n k i n g}} + \underbrace {\left(\ell_ {\mathrm {S}} (i) - \ell_ {\mathrm {S}} ^ {*} (i)\right)} _ {\text {s o r t i n g}} \right]. \tag {10} +$$ + +The primary terms of RS Loss can be obtained in a similar manner to AP Loss (see the Supp.Mat. for derivations), yielding: + +$$ +L _ {i j} ^ {R S} = \left\{ \begin{array}{l l} \left(\ell_ {\mathrm {R}} (i) - \ell_ {\mathrm {R}} ^ {*} (i)\right) p _ {R} (j | i), & \text {f o r} i \in \mathcal {P}, j \in \mathcal {N} \\ \left(\ell_ {\mathrm {S}} (i) - \ell_ {\mathrm {S}} ^ {*} (i)\right) p _ {S} (j | i), & \text {f o r} i \in \mathcal {P}, j \in \mathcal {P}, \\ 0, & \text {o t h e r w i s e}, \end{array} \right. \tag {11} +$$ + +where $p_S(j|i)$ is the sorting pmf. Note that RS Loss has the same primary terms with AP Loss for $i \in \mathcal{P}, j \in \mathcal{N}$ . The primary terms for $i \in \mathcal{P}, j \in \mathcal{P}$ directly targets promoting the better-localized positives. + +# 3.4 Complexity of Ranking-based Loss Functions + +These loss functions originally have the space complexity of $\mathcal{O}((|\mathcal{P}| + |\mathcal{N}|)^2)$ [5,6] as they need to compare each pair. This quadratic space complexity makes it infeasible to compute these loss functions using modern GPUs as the number of logits in object detection is very large. To alleviate that, Chen et al. [5] introduced two tricks at the expense of making the computation inefficient: (1) Loop + +on Positives: By implementing a loop over the positive examples, Chen et al. managed to reduce the time complexity to $\mathcal{O}(|\mathcal{P}|(|\mathcal{P}| + |\mathcal{N}|)) \approx \mathcal{O}(|\mathcal{P}||\mathcal{N}|)$ since $\mathcal{N} \gg \mathcal{P}$ and the space complexity to $\mathcal{O}(|\mathcal{P}| + |\mathcal{N}|)$ . (2) Discard Trivial Negatives: AP Loss considers negative examples with lower rank than all positives as trivial and disregards them. + +# 4 Bucketed Ranking-based Loss Functions for Efficient Optimisation of Object Detectors + +Training object detectors with existing ranking losses has several advantages as we presented before, however, they suffer from significantly large training time. We now address this by our Bucketing Approach, and propose our Bucketed AP and RS Losses. + +# 4.1 Bucketing Negatives for Efficient Optimisation + +Our main intuition is to bucket the sequential negative examples considering that they will have very similar or equal gradients. Fig. 1(a) demonstrates this phenomenon in which $n_1, n_2$ both are assigned equal gradients as well as each negative among $n_3, n_4, n_5, n_6$ in the standard AP Loss. Formally, we sort all given logits $s_i \in \mathcal{P} \cup \mathcal{N}$ using a conventional sorting methods. Let us denote this sorted permutation of the logits by $\hat{s}_1, \hat{s}_2, \dots, \hat{s}_{|\mathcal{P} \cup \mathcal{N}|}$ , i.e., $\hat{s}_1 > \hat{s}_2 > \dots > \hat{s}_{|S|}$ . Given these sorted logits and denoting the $i$ th positive logit in the ordering by $\hat{s}_i^+$ , the buckets of negatives $B_1, \dots, B_{|\mathcal{P} + 1|}$ can be obtained by: + +$$ +B _ {1} > \hat {s} _ {1} ^ {+} > B _ {2} > \hat {s} _ {2} ^ {+} > \dots > \hat {s} _ {| \mathcal {P} |} ^ {+} > B _ {| \mathcal {P} + 1 |}. \tag {12} +$$ + +We will denote the size of the $i$ th negative bucket by $b_{i}$ . Having created the buckets, we now determine a single logit value for each bucket, which we call as the prototype logit as it is necessary while assigning the ranking error. We denote the prototype logit for the $i^{\mathrm{th}}$ bucket by $s_i^b$ . Note that if $\delta = 0$ in Eq. 2, then any logit satisfying $\hat{s}_{i - 1}^{+} > s_{i}^{b} > \hat{s}_{i}^{+}$ can be a prototype logit as the ranking stays the same. However, in the case of $\delta > 0$ , the boundaries between the logits are smoothed. That's why, practically, we find it effective to use the mean logit of a bucket as its prototype logit. + +Note that this bucketing approach reduces the number of logits (positive and prototype negative) to a maximum of $2|\mathcal{P}| + 1$ . As a result, the pairwise errors can now fit into the memory. Consequently, the loop in the red box of Alg. S1 (in Suppl. Mat.) is no longer necessary, which gives rise to efficient ranking-based loss functions which we discuss in the following. + +# 4.2 Bucketed Ranking-based Loss Functions + +Here, we introduce Bucketed versions of AP and RS Losses. Please refer to Alg. S2 (in Suppl. Mat.) for the algorithm. + +Bucketed AP Loss Definition. To define Bucketed version of AP Loss, we need to define current $(\ell_{\mathrm{R}}^{b}(i))$ and target ranking errors $(\ell_{\mathrm{R}}^{b,*}(i))$ for the $i$ th positive. Similar to previous works [5,6,24], we use a target ranking error of 0, i.e., $\ell_{\mathrm{R}}^{b,*}(i) = 0$ . Defining the current ranking error, similar to Eq. 3, requires $\operatorname{rank}(i)$ and $N_{\mathrm{FP}}(i)$ . These two quantities can easily be defined using the bucket size $b_i$ and the prototype logit. Different from the conventional AP Loss, we obtain the pairwise relation $H(x_{ij}^{b})$ using the prototype logit $s_i^b$ . Particularly, with the ordering in Eq. 12, $\operatorname{rank}(i) = \sum_{j=1}^{i} H(x_{ij}) + H(x_{ij}^{b}) b_j$ and $N_{\mathrm{FP}}(i) = \sum_{j=1}^{i} H(x_{ij}^{b}) b_j$ where $x_{ij}^{b} = s_j^b - s_i$ . Then, the resulting ranking error computed on $i$ th positive is: + +$$ +\ell_ {\mathrm {R}} ^ {b} (i) = \frac {N _ {\mathrm {F P}} (i)}{\operatorname {r a n k} (i)} = \frac {\sum_ {j = 1} ^ {i} H \left(x _ {i j} ^ {b}\right) b _ {j}}{\sum_ {j = 1} ^ {i} H \left(x _ {i j}\right) + H \left(x _ {i j} ^ {b}\right) b _ {j}}, \quad i \in \mathcal {P}. \tag {13} +$$ + +Optimisation. Here, we need to define the primary terms, as the product between the error and the pmf following Identity Update [24]. Unlike the previous work, the weights of each prototype negative are not equal as a bucket includes varying number of negatives. Therefore, we use a weighted pmf while distributing the ranking error over negatives. Formally, if $j$ is the $j$ th prototype negative, then we define the pmf as $p(j^b|i) = b_j / N_{\mathrm{FP}}(i)$ . The resulting primary term between the $i$ th positive and $j^b$ th negative is then: + +$$ +L _ {i j} ^ {b} = \frac {\sum_ {j = 1} ^ {i} H \left(x _ {i j} ^ {b}\right) b _ {j}}{\sum_ {j = 1} ^ {i} H \left(x _ {i j}\right) + H \left(x _ {i j} ^ {b}\right) b _ {j}} \times \frac {b _ {j}}{N _ {\mathrm {F P}} (i)}, \quad i \in \mathcal {P}, j \in \tilde {\mathcal {N}} \tag {14} +$$ + +where $\tilde{\mathcal{N}}$ is the set of prototype negatives. However, we need the primary terms for the $i$ th negative, not for the prototype ones. Given $L_{ij}^{b}$ , the gradients of the prototype negatives and actual negatives can easily be obtained following Identity Update. However, we still need to find the gradients of the actual negatives. To do so, we simply normalise the prototype gradient by its bucket size $b_{j}$ and distribute the gradients to actual negatives, completing our method. In Theorem 1, we establish that our Bucketed AP Loss provides exactly the same gradients with AP Loss when $\delta = 0$ . + +Bucketed RS Loss Given that we obtained Bucketed AP Loss, converting RS Loss into a bucketed form is more straightforward. This is because (i) the primary term and its gradients of RS Loss are equal to those of AP Loss if $i \in \mathcal{P}$ , $j \in \mathcal{N}$ and we simply use $L_{ij}^{b}$ in such cases; and (ii) if $i \in \mathcal{P}$ , $j \in \mathcal{P}$ , then an additional term is included in RS Loss to estimate the pairwise relations between positives (Eq. 11), which is not affected by bucketing. + +# 4.3 Theoretical Discussion + +The first theorem ensures that the gradients provided by our loss functions are identical with their conventional counterparts under certain circumstances. The second states that our algorithm is theoretically faster than the conventional AP and RS Losses. + +Theorem 1. Bucketed AP Loss and Bucketed RS Loss provide exactly the same gradients with AP and RS Losses respectively when $\delta = 0$ . + +Theorem 2. Bucketed RS and Bucketed AP Losses have $\mathcal{O}(\max((|\mathcal{P}| + |\mathcal{N}|) \log(|\mathcal{P}| + |\mathcal{N}|), |\mathcal{P}|^2))$ time complexity. + +As a result, we observe the same performance empirically in less amount of time. Proofs are provided in Supp. Mat. + +# 5 Bucketed Rank-Sort (BRS) DETR + +While transformer-based detectors [8,14,16,37,38] have been providing the best performance in several object detection benchmarks, optimizing such detectors using ranking-based losses based on performance measures has not been investigated. Here, we address this gap by incorporating our BRS Loss into CoDETR [38] as the current SOTA detector on the common COCO benchmark [18]. + +Co-DETR increases the number and variation of the positive examples in the transformer head by leveraging one-to-many assignment strategies in anchor-based detectors (such as ATSS [34] and Faster R-CNN [29]) as auxiliary heads. As the number of proposals is large in such detectors, they can easily provide more and diverse examples, resulting in better performance of the transformer head. Co-DETR is originally supervised with the following loss function: + +$$ +\sum_ {l = 1} ^ {L} \left(\tilde {\mathcal {L}} _ {l} ^ {D e c} + \lambda_ {1} \sum_ {k = 1} ^ {K} \tilde {\mathcal {L}} _ {k, l} ^ {A u x} + \lambda_ {2} \sum_ {k = 1} ^ {K} \mathcal {L} _ {k, l} ^ {A u x}\right), \tag {15} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ weigh each loss component. The first two components, $\tilde{\mathcal{L}}_l^{Dec}$ and $\tilde{\mathcal{L}}_{k,l}^{Aux}$ , are the losses of the $l^{th}$ decoder layer in the following form: + +$$ +\lambda_ {c l s} \mathcal {L} _ {c l s} + \lambda_ {b b o x} \mathcal {L} _ {b b o x} + \lambda_ {I o U} \mathcal {L} _ {I o U}, \tag {16} +$$ + +and $\mathcal{L}_{cls}$ , $\mathcal{L}_{bbox}$ and $\mathcal{L}_{IoU}$ are Focal Loss, L1 Loss and GIoU Loss respectively [38]. $\tilde{\mathcal{L}}_l^{Dec}$ and $\tilde{\mathcal{L}}_{k,l}^{Aux}$ differ in their input queries. That is, while $\tilde{\mathcal{L}}_l^{Dec}$ follows standard DETR-based models [14, 37] with one-to-one matching of the queries, $\tilde{\mathcal{L}}_{k,l}^{Aux}$ is computed based on positive anchors of $k^{th}$ auxiliary head with one-to-many assignment. And $\mathcal{L}_{k,l}^{Aux}$ is the conventional loss of the $k^{th}$ auxiliary head, e.g., the weighted sum of classification, localisation and centerness losses for ATSS [34]. + +In order to align the training and evaluation objectives better, we replace $\tilde{\mathcal{L}}_l^{Dec}$ and $\tilde{\mathcal{L}}_{k,l}^{Aux}$ in Eq. 16 by: + +$$ +\mathcal {L} _ {B R S} + \lambda_ {b b o x} \mathcal {L} _ {b b o x} + \lambda_ {I o U} \mathcal {L} _ {I o U}, \tag {17} +$$ + +and set $\lambda_{bbox}$ and $\lambda_{IoU}$ dynamically during training to $\mathcal{L}_{BRS} / \mathcal{L}_{bbox}$ and $\mathcal{L}_{BRS} / \mathcal{L}_{IoU}$ to simplify the hyperparameter tuning following Oksuz et al. [24]. Similarly, we replace the loss functions of the auxiliary ATSS [34] and Faster R-CNN [29] heads ( $\mathcal{L}_{k,l}^{Aux}$ in Eq. 15) by our BRS Loss. This modification also aligns with the + +objective of the auxiliary heads by the performance measure, hence results in more accurate auxiliary heads [24]. Furthermore, as using BRS Loss with Faster R-CNN does not require sampling thanks to its robustness to imbalance, no limitation is imposed on the number of positives in Faster R-CNN. Hence, the main aim of Co-DETR, that is to introduce more positive examples to the transformer head, is corroborated. As a result, our BRS DETR enables significantly efficient training compared to RS Loss also by improving the performance of Co-DETR. + +# 6 Experiments + +We analyze our contributions in three primary sections. First, we evaluate the effectiveness of our bucketing approach by contrasting it with RS Loss and AP Loss across different CNN-based visual detectors, on object detection and instance segmentation tasks. Next, we thoroughly examine how our bucketed RS Loss performs on transformer-based object detectors with different backbones. This is the first time a ranking-based loss is applied to transformer-based object detectors, thanks to the time-efficiency of our bucketing method. Finally, we present comparisons with score-based loss functions and other ranking-based loss functions in the literature. Our experiments clearly show that in terms of training time, accuracy and ease of tuning, our BRS Loss is more preferable over any existing loss function to train object detectors. + +Dataset and Performance Measures. Unless otherwise noted, we train all models on COCO trainval35k (115k images) and test them on minival (5k images). We use COCO-style Average Precision (AP) and also report $\mathrm{AP}_{50}$ , $\mathrm{AP}_{75}$ as the APs at IoUs 0.50 and 0.75; and $\mathrm{AP_S}$ , $\mathrm{AP_M}$ and $\mathrm{AP_L}$ to present the accuracy on small, medium and large objects. + +# 6.1 Efficiency and Effectiveness of Bucketed Ranking Losses + +To comprehensively analyze the effectiveness of our bucketing method, we design experiments on real-world data as well as synthetic data. For the former, we aim to present that our method significantly decreases the training time of the detection and segmentation methods (up to $\sim 5\times$ ). As training the entire detector does not isolate the runtime of the loss function, on which our main contribution is, we also design an experiment with synthetic data. This set of experiments show that our bucketing approach, in fact, decreases the loss function runtime by up to $40\times$ , thereby resulting in shorter training time of the detectors. + +Experimental Setup. We follow the experimental setup used in RSLoss [24] only by replacing the loss function by our bucketed loss function to ensure comparability and consistency of results. We report our findings on efficiency and accuracy using the models trained on 4 Tesla A100 GPUs with a batch size of 16, i.e., 4 images/GPU. Unless stated otherwise, we train the models for 12 epochs and test them on images with a size of $1333 \times 800$ following the common convention [12, 24, 34]. In order to comprehensively demonstrate the efficiency of our approach, we use five different detectors: Faster R-CNN [29], Cascade R-CNN [2], + +Table 1: Bucketed Losses (BRS, BAP) vs. RS & AP Losses on COCO object detection for object detectors. R-X: ResNet-X. Models with R-101 are trained 36 epochs. + +
MethodDetectorBBLTimeAccuracy
\(T_{iter}(s) \downarrow\)AP↑\(\mathrm{AP}_{50} \uparrow\)\(\mathrm{AP}_{75} \uparrow\)
Multi StageFaster R-CNNR-50RS0.5039.459.543.0
BRS\(0.17 3.0x \downarrow\)39.559.542.8
R-101RS0.7547.367.451.2
BRS\(0.38 2.0x \downarrow\)47.767.851.5
Cascade R-CNNR-50RS1.2841.158.744.1
BRS\(0.24 5.3x \downarrow\)41.158.844.2
One StageATSSR-50AP0.3238.158.241.0
BAP\(0.15 2.1x \downarrow\)38.558.641.3
RS0.3639.858.942.6
BRS\(0.15 2.4x \downarrow\)39.858.842.9
PAAAP0.4537.354.341.2
BAP\(0.30 1.5x \downarrow\)37.256.240.2
RS0.5740.858.844.6
BRS\(0.30 1.9x \downarrow\)40.959.044.4
+ +ATSS [34] and PAA [12] for object detection as well as Mask R-CNN [10] for instance segmentation. + +Bucketed Ranking Based Losses Significantly Decrease the Training Time with Similar Accuracy. Multi-stage detectors: Among multi-stage detectors, we train the commonly-used Faster R-CNN and Cascade R-CNN with RS Loss and BRS Loss, and present average iteration time as well as AP of the trained models. In order to evaluate our contribution comprehensively, we train Faster R-CNN also with a stronger setting, in which, we use ResNet-101, train it for 36 epochs using multi-scale training similar to [24]. The results are presented in Table 1, in which we can clearly see that BRS Loss consistently reduces the training time of all three detectors as well as preserves (or slightly improve) their performance. Especially for Cascade R-CNN, which is still a very popular and strong object detector along with its variants such as HTC [4], the training time decreases by $5.3 \times$ . This is because the loss function is applied to the logits for three times based on the cascaded nature of this detector, and therefore, our contribution can be easily noticed. Moreover, it might seem that the efficiency gain decreases once the size of the backbone increases from R-50 to R-101 in Faster R-CNN. However, this is an expected result as the overall feature extraction time increases due to the larger number of parameters in the backbone, and hence, as we will discuss in the synthetic experiments, this is independent from our bucketing approach, operating on the loss function after the feature extraction. + +One-stage detectors: To show that our gains generalize to one-stage detectors and to AP Loss, we train the common ATSS and PAA detectors with our BAP and BRS Losses, and compare them with AP and RS Losses respectively. Table 1 validates our previous claims: Our bucketed losses obtain similar performance in around half of the training time required for AP and RS Losses. + +Table 2: Comparison with RS Loss on instance segmentation task on COCO val using Mask R-CNN. BB: Backbone. + +
BBLEfficiencyAccuracy
\(T_{iter}(s)\)↓AP↑\(AP_{50}\)↑\(AP_{75}\)↑
R-50RS0.5636.357.238.8
BRS0.24 2.3x ↓36.257.238.8
R-101RS0.5940.261.843.5
BRS0.27 2.2x ↓40.362.043.8
+ +Table 3: Comparison on different inst. segm. datasets using Mask R-CNN. $\mathrm{AP}_{75}$ is N/A as it is not used for Cityscapes. + +
DatasetLEfficiencyAccuracy
\(T_{iter}(s) \downarrow\)AP↑\(AP_{50} \uparrow\)\(AP_{75} \uparrow\)
CityscapesRS0.4343.568.1N/A
BRS\(0.19 2.3x \downarrow\)43.367.7N/A
LVISRS0.8725.639.227.3
BRS\(0.35 2.5x \downarrow\)25.839.627.4
+ +Table 4: BRS Loss further improves performance with similar training budget. We report mask AP for instance segmentation + +
TaskDetectorLAP↑AP50↑AP75↑
Object Det.C.R-CNNRS41.158.744.1
BRS42.3+1.260.245.2
Inst.Seg.M.R-CNNRS36.357.238.8
BRS37.3+1.058.440.2
+ +Instance Segmentation: Given that our loss functions is efficient in object detectors, one simple extension is to see their generalization to instance segmentation methods. To show that, we train Mask R-CNN, a common baseline, with our BRS Loss on three different datasets from various domains: (i) COCO (Table 2), (ii) Cityscapes [7] as an autonomous driving dataset (Table 3) and LVIS [9] a long-tailed dataset with more than 1K classes (Table 3). The results confirm our previous finding that our BRS Loss decreases the training time significantly by up to $\sim 2.5\times$ compared to RS Loss by preserving its accuracy. + +Given Similar Training Budget, Bucketed Ranking Based Loss Functions Significantly Improve the Accuracy. There might be several benefits of reducing the training time of the detector as more GPU time is saved. Here, we show a use-case in which we ask the following question: what would happen if we allocate the similar amount of training time to both bucketed and non-bucketed losses? To answer that, we train Cascade R-CNN, a detector, and Mask R-CNN, an instance segmentation method, using our BRS Loss. Considering our speed-ups on these models $(6.0\times$ and $2.3\times)$ , we simply increased their training epochs from 12 to 36 and 27 respectively. Note that, for Cascade R-CNN, we still spend significantly less amount of training time. Table 4 shows that models trained with the BRS Loss outperform (with $+1.2$ and $1.0\mathrm{AP}$ ) their counterparts trained with the RS Loss in both cases. + +Experiments on Synthetic Data. Up to now, to measure any speed-up, we considered the entire pipeline and did not decouple the loss computation step. However, we only improve the loss computation step. Hence, an analysis focusing only on this part can reveal our contribution more clearly. Therefore, we design a simple experiment using synthetic data. Particularly, we randomly generate $L$ logits such that $m\%$ of these logits are positive (see Supp.Mat. for data generation details). To cover the wide range of existing detectors, we set $L \in \{10K, 100K, 1M\}$ and $m \in \{0.1, 1.0, 2.0, 5.0\}$ . Different choices for $L$ and $m$ + +Table 5: Comparison of our BRS-DETR with Co-DETR (trained with its original loss function) on different backbones on COCO val set. + +
BackboneDetectorAP\( AP_{50} \)\( AP_{75} \)\( AP_s \)\( AP_m \)\( AP_1 \)
ResNet50Co-DETR49.367.254.032.152.663.8
BRS-DETR50.167.454.631.953.965.0
Swin-TCo-DETR51.769.656.434.454.966.8
BRS-DETR52.369.557.134.655.768.0
Swin-LCo-DETR56.975.562.640.161.273.3
BRS-DETR57.275.062.539.461.774.1
+ +Table 6: Comparison of our BRS-DETR with DETR variants (trained with their original loss functions) w ResNet-50 on COCO val set. Qu: Queries. Ep: Epochs. + +
DetectorQuEpAP\( {\mathrm{{AP}}}_{50} \)\( {\mathrm{{AP}}}_{75} \)\( {\mathrm{{AP}}}_{\mathrm{s}} \)\( {\mathrm{{AP}}}_{\mathrm{m}} \)\( {\mathrm{{AP}}}_{1} \)
DETR [14]10030042.062.444.220.545.861.1
DN-DETR [15]3005048.667.452.731.052.063.7
DINO [3]9001249.466.953.832.352.563.9
H-DETR [11]3001248.766.452.931.251.563.5
Co-DETR [38]3001249.367.254.032.152.663.8
BRS-DETR3001250.167.454.631.953.965.0
+ +correspond to the logit distributions of common object detectors. We compute BAP and AP Loss on this synthetic data and compare their runtime to focus on loss computation time. We observed that our BAP Loss has a significant, up to $\sim 40\times$ , speed-up compared to AP Loss. The speed-up becomes more pronounced with increasing number of logits. Supp. Mat. presents detailed results. + +# 6.2 The Effectiveness of Our BRS-DETR + +While we incorporate our BRS Loss into Co-DETR, we use all original settings unless otherwise noted. Please refer to Supp.Mat. for the details. + +BRS-DETR Outperforms Existing Transformer-based Detectors Consistently. Here, we first employ ResNet-50 backbone as it has been used by many detectors, enabling us to compare our method with many different DETR-based manner in a consistent manner. In this fair comparison setting, our BRS-DETR outperforms all existing DETR variants reaching 50.1 AP as shown in Table 6. For example, our BRS DETR outperforms DN-DETR by 1.5AP with less training epochs and improves Co-DETR as we discuss next. + +Training Co-DETR with RS Loss takes 6x less time compared to using RS Loss. Finally, we also compare the training efficiencies of BRS Loss and RS Loss on Co-DETR. We observe that our BRS Loss decreases the training time of Co-DETR by $6.0 \times$ (from 4.14s per iteration to 0.69s) in comparison to RS Loss. We note that this is the largest training time gain of our BRS Loss. This is because Co-DETR consists of multiple transformer-based as well as auxiliary heads, requiring multiple loss estimations. + +BRS-DETR Improves Co-DETR over Different Backbones Consistently. In order to show the effectiveness of our BRS Loss, we compare our results with Co-DETR in Table 5 using different backbones. We note that we improve the baseline Co-DETR consistently in all settings. For example, our improvement on ResNet-50 is 0.8AP, which is a notable improvement. However, as the backbone gets larger (e.g., Swin-L [20]), our gains decrease, which is expected and commonly observed in the literature, e.g., [12]. + +Table 7: Comparison with the score-based losses. CE: Cross-entropy. #H: Number of hyperparameters. Ours has very similar $T_{iter}(s)$ also by being more accurate and simple-to-tune. + +
DetectorLTiter(s) ↓AP ↑#H
Faster R-CNNCE+L10.1437.69
RS0.5039.43
BRS (Ours)0.1739.53
ATSSFocal L. +G IoU0.1439.35
RS0.3639.81
BRS (Ours)0.1539.81
+ +Table 8: Comparison with other ranking-based losses. Our approach performs better on both accuracy and efficiency. + +
DetectorL\(T_{iter}(s) \downarrow\)AP ↑#H
Faster R-CNNaLRP [23]0.2837.43
BRS (Ours)0.1739.53
ATSSAP [5]0.3238.15
DR [28]0.2038.15
aLRP [23]0.3237.71
BRS (Ours)0.1539.81
RetinaNetDR [28]0.2937.45
BRS (Ours)0.2338.31
+ +# 6.3 Comparison with Score-based and Other Ranking-based Losses + +In previous sections, we showed the efficiency of our loss functions compared to AP and RS Losses as their counterparts. Here, we compare our bucketed losses with the score based losses, i.e., Cross-entropy and Focal Loss, and other ranking based losses including DR Loss [28] and aLRP Loss [23]. To do so, in Tables 7 and 8, we report average training time, AP for accuracy and number of hyperparameters to capture the tuning simplicity of the loss functions. Compared to score-based loss functions in Table 7, our BRS Loss improves performance of both approaches, is significantly simpler-to-tune with a low number of hyperparameters and has similar training iteration time. Hence, our BRS Loss is either superior or on par with existing score-based losses. As for other ranking-based losses in Table 8, our BRS Loss is the most efficient and accurate ranking-based loss function for all three detectors. As an example, compared to ATSS trained with DR Loss, our BRS Loss yields 1.8AP better accuracy with $25\%$ less training time and significantly less number of hyperparameters. Therefore, our bucketed loss functions are now promising alternatives to train object detectors. + +# 7 Conclusion + +In this paper, we introduced a novel method to improve the efficiency of ranking-based loss functions by binning negatives into buckets and implementing positive-negative pairwise comparisons between positives and buckets of negatives. We showed that this method improves the computational complexity to a level where pairwise comparisons can be stored as a matrix in memory and running time becomes very close to that of score-based loss functions. Also, for the first time, we integrated a ranking-based loss to Co-DETR, a DETR-based detector, which was possible thanks to our method's lower complexity, and reported improvements on different backbones. Our comprehensive experiments comprising 2 different tasks, 3 different dataset and 7 different detectors - all with positive outcomes - show the general applicability of our approach. + +# Acknowledgements + +We gratefully acknowledge the computational resources provided by TÜBITAK ULAKBIM High Performance and Grid Computing Center (TRUBA), METU Center for Robotics and Artificial Intelligence (METU-ROMER) and METU Image Processing Laboratory. Dr. Akbas is supported by the "Young Scientist Awards Program (BAGEP)" of Science Academy, Türkiye. + +# References + +1. Brown, A., Xie, W., Kalogeiton, V., Zisserman, A.: Smooth-ap: Smoothing the path towards large-scale image retrieval. In: European Conference on Computer Vision. pp. 677-694. Springer (2020) +2. Cai, Z., Vasconcelos, N.: Cascade R-CNN: Delving into high quality object detection. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018) +3. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers (2021) +4. Chen, K., Pang, J., Wang, J., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Shi, J., Ouyang, W., Loy, C.C., Lin, D.: Hybrid task cascade for instance segmentation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019) +5. Chen, K., Li, J., Lin, W., See, J., Wang, J., Duan, L., Chen, Z., He, C., Zou, J.: Towards accurate one-stage object detection with ap-loss. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019) +6. Chen, K., Lin, W., li, J., See, J., Wang, J., Zou, J.: Ap-loss for accurate one-stage object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) pp. 1-1 (2020) +7. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016) +8. Dai, Z., Cai, B., Lin, Y., Chen, J.: Unsupervised pre-training for detection transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence p. 1-11 (2022). https://doi.org/10.1109/tpami.2022.3216514, http://dx.doi.org/10.1109/TPAMI.2022.3216514 +9. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019) +0. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask R-CNN. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2017) +1. Jia, D., Yuan, Y., He, H., Wu, X., Yu, H., Lin, W., Sun, L., Zhang, C., Hu, H.: Detrs with hybrid matching (2023) +2. Kim, K., Lee, H.S.: Probabilistic anchor assignment with iou prediction for object detection. In: The European Conference on Computer Vision (ECCV) (2020) +3. Law, H., Deng, J.: Cornernet: Detecting objects as paired keypoints. In: The European Conference on Computer Vision (ECCV) (2018) + +14. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13619-13627 (June 2022) +15. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising (2022) +16. Li*, L.H., Zhang*, P., Zhang*, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., Chang, K.W., Gao, J.: Grounded language-image pretraining. In: CVPR (2022) +17. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 42(2), 318-327 (2020) +18. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common Objects in Context. In: The European Conference on Computer Vision (ECCV) (2014) +19. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S.E., Fu, C., Berg, A.C.: SSD: single shot multibox detector. In: The European Conference on Computer Vision (ECCV) (2016) +20. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows (2021) +21. Mohapatra, P., Jawahar, C., Kumar, M.P.: Efficient optimization for average precision svm. Advances in Neural Information Processing Systems 27 (2014) +22. Mohapatra, P., Rolínek, M., Jawahar, C., Kolmogorov, V., Pawan Kumar, M.: Efficient optimization for rank-based loss functions. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018) +23. Oksuz, K., Cam, B.C., Akbas, E., Kalkan, S.: A ranking-based, balanced loss function unifying classification and localisation in object detection. In: Advances in Neural Information Processing Systems (NeurIPS) (2020) +24. Oksuz, K., Cam, B.C., Akbas, E., Kalkan, S.: Rank & sort loss for object detection and instance segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3009-3018 (2021) +25. Oksuz, K., Cam, B.C., Kahraman, F., Baltaci, Z.S., Kalkan, S., Akbas, E.: Mask-awareiou for anchor assignment in real-time instance segmentation. In: The British Machine Vision Conference (BMVC) (2021) +26. Oksuz, K., Cam, B.C., Kalkan, S., Akbas, E.: Imbalance problems in object detection: A review. IEEE transactions on pattern analysis and machine intelligence 43(10), 3388-3415 (2020) +27. Pang, J., Chen, K., Shi, J., Feng, H., Ouyang, W., Lin, D.: Libra R-CNN: Towards balanced learning for object detection. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019) +28. Qian, Q., Chen, L., Li, H., Jin, R.: Dr loss: Improving object detection by distributional ranking. In: CVPR (2020) +29. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 39(6), 1137-1149 (2017) +30. Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I., Savarese, S.: Generalized intersection over union: A metric and a loss for bounding box regression. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019) +31. Rosenblatt, F.: The perceptron: A probabilistic model for information storage and organization in the brain. Psychological Review pp. 65-386 (1958) + +32. Shrivastava, A., Gupta, A., Girshick, R.: Training region-based object detectors with online hard example mining. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2016) +33. Sun, P., Zhang, R., Jiang, Y., Kong, T., Xu, C., Zhan, W., Tomizuka, M., Li, L., Yuan, Z., Wang, C., Luo, P.: SparseR-CNN: End-to-end object detection with learnable proposals. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018) +34. Zhang, S., Chi, C., Yao, Y., Lei, Z., Li, S.Z.: Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020) +35. Zhang, X., Wan, F., Liu, C., Ji, R., Ye, Q.: Freeanchor: Learning to match anchors for visual object detection. In: Advances in Neural Information Processing Systems (NeurIPS) (2019) +36. Zheng, Z., Wang, P., Liu, W., Li, J., Rongguang, Y., Dongwei, R.: Distance-iou loss: Faster and better learning for bounding box regression. In: AAAI Conference on Artificial Intelligence (2020) +37. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. ArXiv abs/2010.04159 (2020), https://api-semanticscholar.org/CorpusID:222208633 +38. Zong, Z., Song, G., Liu, Y.: Detrs with collaborative hybrid assignments training (2022) \ No newline at end of file diff --git a/2024/Bucketed Ranking-based Losses for Efficient Training of Object Detectors/images.zip b/2024/Bucketed Ranking-based Losses for Efficient Training of Object Detectors/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8dc118053309e635fbb87189bcaa647121a3f320 --- /dev/null +++ b/2024/Bucketed Ranking-based Losses for Efficient Training of Object Detectors/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:959c29819d6d6f6ec6e1b87b3791c4f985e41846361b71e3f85bfcca1f4de1d3 +size 380659 diff --git a/2024/Bucketed Ranking-based Losses for Efficient Training of Object Detectors/layout.json b/2024/Bucketed Ranking-based Losses for Efficient Training of Object Detectors/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..aa6167007b100121f846b6b9f478ae587c4c9f19 --- /dev/null +++ b/2024/Bucketed Ranking-based Losses for Efficient Training of Object Detectors/layout.json @@ -0,0 +1,10546 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 175, + 111, + 438, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 111, + 438, + 148 + ], + "spans": [ + { + "bbox": [ + 175, + 111, + 438, + 148 + ], + "type": "text", + "content": "Bucketed Ranking-based Losses for Efficient Training of Object Detectors" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "spans": [ + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "content": "Feyza Yavuz1, Baris Can Cam1, Adnan Harun Dogan1, Kemal Oksuz2, Emre Akbas1,3,*, and Sinan Kalkan1,3,*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 171, + 201, + 441, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 201, + 441, + 223 + ], + "spans": [ + { + "bbox": [ + 171, + 201, + 441, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 171, + 201, + 441, + 223 + ], + "type": "text", + "content": " Department of Computer Engineering, METU, Ankara, Türkiye " + }, + { + "bbox": [ + 171, + 201, + 441, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 171, + 201, + 441, + 223 + ], + "type": "text", + "content": " Five AI Ltd., United Kingdom" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 198, + 224, + 415, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 224, + 415, + 246 + ], + "spans": [ + { + "bbox": [ + 198, + 224, + 415, + 246 + ], + "type": "text", + "content": "3 METU ROMER Robotics Center, Ankara, Türkiye * Equal senior authorship" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 277, + 455, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 277, + 455, + 540 + ], + "spans": [ + { + "bbox": [ + 160, + 277, + 455, + 540 + ], + "type": "text", + "content": "Abstract. Ranking-based loss functions, such as Average Precision Loss and Rank&Sort Loss, outperform widely used score-based losses in object detection. These loss functions better align with the evaluation criteria, have fewer hyperparameters, and offer robustness against the imbalance between positive and negative classes. However, they require pairwise comparisons among " + }, + { + "bbox": [ + 160, + 277, + 455, + 540 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 160, + 277, + 455, + 540 + ], + "type": "text", + "content": " positive and " + }, + { + "bbox": [ + 160, + 277, + 455, + 540 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 160, + 277, + 455, + 540 + ], + "type": "text", + "content": " negative predictions, introducing a time complexity of " + }, + { + "bbox": [ + 160, + 277, + 455, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(PN)" + }, + { + "bbox": [ + 160, + 277, + 455, + 540 + ], + "type": "text", + "content": ", which is prohibitive since " + }, + { + "bbox": [ + 160, + 277, + 455, + 540 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 160, + 277, + 455, + 540 + ], + "type": "text", + "content": " is often large (e.g., " + }, + { + "bbox": [ + 160, + 277, + 455, + 540 + ], + "type": "inline_equation", + "content": "10^8" + }, + { + "bbox": [ + 160, + 277, + 455, + 540 + ], + "type": "text", + "content": " in ATSS). Despite their advantages, the widespread adoption of ranking-based losses has been hindered by their high time and space complexities. In this paper, we focus on improving the efficiency of ranking-based loss functions. To this end, we propose Bucketed Ranking-based Losses which group negative predictions into " + }, + { + "bbox": [ + 160, + 277, + 455, + 540 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 160, + 277, + 455, + 540 + ], + "type": "text", + "content": " buckets (" + }, + { + "bbox": [ + 160, + 277, + 455, + 540 + ], + "type": "inline_equation", + "content": "B \\ll N" + }, + { + "bbox": [ + 160, + 277, + 455, + 540 + ], + "type": "text", + "content": ") in order to reduce the number of pairwise comparisons so that time complexity can be reduced. Our method enhances the time complexity, reducing it to " + }, + { + "bbox": [ + 160, + 277, + 455, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\max(N\\log(N),P^2))" + }, + { + "bbox": [ + 160, + 277, + 455, + 540 + ], + "type": "text", + "content": ". To validate our method and show its generality, we conducted experiments on 2 different tasks, 3 different datasets, 7 different detectors. We show that Bucketed Ranking-based (BR) Losses yield the same accuracy with the unbucketed versions and provide " + }, + { + "bbox": [ + 160, + 277, + 455, + 540 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 160, + 277, + 455, + 540 + ], + "type": "text", + "content": " faster training on average. We also train, for the first time, transformer-based object detectors using ranking-based losses, thanks to the efficiency of our BR. When we train CoDETR, a state-of-the-art transformer-based object detector, using our BR Loss, we consistently outperform its original results over several different backbones. Code will be released." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 560, + 460, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 560, + 460, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 560, + 460, + 574 + ], + "type": "text", + "content": "Keywords: Object Detection " + }, + { + "bbox": [ + 132, + 560, + 460, + 574 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 132, + 560, + 460, + 574 + ], + "type": "text", + "content": " Ranking Losses " + }, + { + "bbox": [ + 132, + 560, + 460, + 574 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 132, + 560, + 460, + 574 + ], + "type": "text", + "content": " Efficient Ranking Losses" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 592, + 230, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 592, + 230, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 592, + 230, + 605 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Training an object detector incurs many challenges unique to object detection, including (i) assigning dynamically a large number of object hypotheses with the objects [2, 12, 13, 25, 29, 33, 34], (ii) sampling among these hypotheses to ensure that the background class does not dominate training [19, 27, 29, 32] and" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 318, + 148 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 318, + 148 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 318, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 318, + 148 + ], + "type": "image", + "image_path": "c805ee1cbcd735452d8b8b83680e27b7722fce63da5db9af6fe80d48307ebeb3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 154, + 266, + 163 + ], + "lines": [ + { + "bbox": [ + 194, + 154, + 266, + 163 + ], + "spans": [ + { + "bbox": [ + 194, + 154, + 266, + 163 + ], + "type": "text", + "content": "(a) Comparisons in AP Loss" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 176, + 318, + 215 + ], + "blocks": [ + { + "bbox": [ + 133, + 176, + 318, + 215 + ], + "lines": [ + { + "bbox": [ + 133, + 176, + 318, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 176, + 318, + 215 + ], + "type": "image", + "image_path": "7a617a6d94d316ca388a4502024453c37d0122189af09f0f7a40290b7354215b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 223, + 285, + 232 + ], + "lines": [ + { + "bbox": [ + 170, + 223, + 285, + 232 + ], + "spans": [ + { + "bbox": [ + 170, + 223, + 285, + 232 + ], + "type": "text", + "content": "(b) Comparisons in Bucketed AP Loss (ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 322, + 114, + 481, + 213 + ], + "blocks": [ + { + "bbox": [ + 322, + 114, + 481, + 213 + ], + "lines": [ + { + "bbox": [ + 322, + 114, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 322, + 114, + 481, + 213 + ], + "type": "image", + "image_path": "6c56eb8e3f6cbc738baa91872c80d850eb019b86c314cc556fcb9af123a3a762.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 220, + 478, + 236 + ], + "lines": [ + { + "bbox": [ + 333, + 220, + 478, + 236 + ], + "spans": [ + { + "bbox": [ + 333, + 220, + 478, + 236 + ], + "type": "text", + "content": "(c) Average Precision (AP) vs. average training iteration time for baselines, RS Loss and Bucketed RS Loss (ours)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 243, + 482, + 321 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 321 + ], + "type": "text", + "content": "Fig. 1: (a) Existing ranking-based losses (i.e., AP Loss [5,6]) incur significant overhead owing to pairwise comparisons between positives and negatives. (b) We propose bucketing negatives to decrease the number of comparisons, and hence the complexity. Under certain assumptions, our bucketing approach provides equal gradients with conventional ranking-based losses such as AP Loss in (a). (c) Accuracy and efficiency comparison. Our BRS formulation facilitates faster (between " + }, + { + "bbox": [ + 130, + 243, + 482, + 321 + ], + "type": "inline_equation", + "content": "1.9 \\times" + }, + { + "bbox": [ + 130, + 243, + 482, + 321 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 243, + 482, + 321 + ], + "type": "inline_equation", + "content": "6.0 \\times" + }, + { + "bbox": [ + 130, + 243, + 482, + 321 + ], + "type": "text", + "content": ") training of visual detectors with similar AP." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 355, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 355, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 355, + 482, + 439 + ], + "type": "text", + "content": "(iii) minimizing a multi-task objective function [17,23,24,29]. While the choices on assignment and sampling generally vary depending on the trained object detector, it is very common to combine Cross Entropy or Focal Loss [17] with a regression loss [30,36] as the multi-task loss function. Recently proposed ranking-based loss functions [5,6,23,24] offer an alternative approach for addressing these challenges by formulating the training objective based on the rank of positive examples over negative examples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 445, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 553 + ], + "type": "text", + "content": "Benefits of ranking-based losses. First, they are inherently robust to imbalance [23] and hence, do not require any sampling mechanism under very challenging scenarios [24], e.g., even when the background foreground ratio is 10K for LVIS [9]. Second, they are shown to generalize over different detectors with diverse architectures - with the exception of transformer-based ones [14,37,38], since ranking-based losses further slow down the training of transformer-based detectors, which we address in this paper. Such losses also offer significant performance gain over their score-based counterparts, and having less hyperparameters, they are much easier to tune [23,24]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "The drawback of ranking-based losses. Compared to score-based losses, ranking-based losses are more inefficient as the ranking operation inherently requires each pair of object hypotheses to be compared against each other (Fig. 1(a)), inducing a quadratic complexity on the large number of object hypotheses (e.g., " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "10^{8}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " for ATSS [34] on COCO [18]). As a result, vectorized implementations for parallel GPU processing are infeasible as such large matrices (e.g., with " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "10^{16}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " entries for ATSS) do not fit into GPU memories. This has driven researchers towards alternative ways of computing these matrices, which, in the end, saves from the storage complexity but results in more inefficient algorithms [5,6]." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Yavuz et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "Our approach. The main cause for the aforementioned complexity is the background class, i.e., negatives, forming up to " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "99.9\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " of the hypotheses [26]. Considering this, we introduce a novel bucketing approach (Fig. 1(b)) on these negatives to address the main shortcoming of the ranking-based loss functions. Specifically, we first sort all examples based on their scores and bucket the negatives located between successive positives. The negatives within the same bucket are treated as a single example and represented by their average score. We then incorporate our bucketing approach into AP Loss [5,6] and RS Loss [24], and introduce novel mechanisms for calculating the gradients correctly." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 355 + ], + "type": "text", + "content": "The significance of our approach. Theoretically, we prove that our bucketing approach provides the same gradients as the original ranking-based loss. Practically, we show we can reduce loss computation time by up to " + }, + { + "bbox": [ + 130, + 224, + 482, + 355 + ], + "type": "inline_equation", + "content": "\\sim 40\\times" + }, + { + "bbox": [ + 130, + 224, + 482, + 355 + ], + "type": "text", + "content": " and training time of a detector by up to " + }, + { + "bbox": [ + 130, + 224, + 482, + 355 + ], + "type": "inline_equation", + "content": "6\\times" + }, + { + "bbox": [ + 130, + 224, + 482, + 355 + ], + "type": "text", + "content": " (Fig. 1(c)), addressing the main drawback of ranking-based loss functions. Hence, we show that the gap between the score-based and ranking-based loss functions in terms of training time vanishes, enabling us to train strong transformer-based object detectors such as Co-DETR using our loss functions. As ranking-based loss functions now surpass score-based ones or perform on par in every aspect such as accuracy, tuning simplicity, robustness to imbalance and training time, we believe that our work will make their prevalence more dominant in the upcoming object detectors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 356, + 423, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 356, + 423, + 366 + ], + "spans": [ + { + "bbox": [ + 132, + 356, + 423, + 366 + ], + "type": "text", + "content": "Contributions. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 373, + 481, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 373, + 481, + 433 + ], + "spans": [ + { + "bbox": [ + 138, + 373, + 481, + 433 + ], + "type": "text", + "content": "- We propose a novel bucketing approach to improve the efficiency of computationally-expensive ranking-based losses to train object detectors. Theoretically, our approach yields the same gradients with the original ranking-based losses while decreasing their time complexity. Practically, we enable up to " + }, + { + "bbox": [ + 138, + 373, + 481, + 433 + ], + "type": "inline_equation", + "content": "6 \\times" + }, + { + "bbox": [ + 138, + 373, + 481, + 433 + ], + "type": "text", + "content": " faster training using our bucketing approach with no accuracy loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 434, + 481, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 481, + 480 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 481, + 480 + ], + "type": "text", + "content": "- For the first time, we incorporate ranking-based loss functions to transformer-based detectors. Specifically, we construct BRS-DETR by replacing the training objective of the state-of-the-art transformer-based object detector CoDETR [38] by our Bucketed RS (BRS) Loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 481, + 481, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 481, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 481, + 481, + 564 + ], + "type": "text", + "content": "- Our comprehensive experiments on detection and instance segmentation on 3 different challenging datasets, 5 backbones and 7 detectors show the effectiveness and generalizability of our approach. Our BRS-DETR reaches 50.1 AP on COCO val set with only 12 epochs and ResNet-50 backbone, outperforming all existing detectors using the same backbone and 300 queries. BRS-DETR also outperforms CoDETR with Swin-T and Swin-L backbones, reaching 57.2 AP on the COCO dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 582, + 237, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 582, + 237, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 582, + 237, + 594 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Although score-based loss functions, such as cross-entropy and focal loss, are widely used to train both CNN-based [2, 17, 35] and transformer-based object detectors [14, 37, 38], ranking-based losses for visual detection offer advantages. The primary advantage is their robustness to class imbalance [23]. They yield on par or better performance compared to score-based losses, without the need" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 318, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 318, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bucketed Ranking-based Losses" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "text", + "content": "for extensive tuning of class weights or multistage sampling techniques. Another advantage they offer is the ease of balancing losses among different tasks. In AP Loss [5, 6], aLRP Loss [23] and RS Loss [24], the authors use the ratio of classification and regression losses to balance these tasks. As a result, ranking-based losses have much fewer hyperparameters during training, compared to score-based ones." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 189, + 482, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 482, + 320 + ], + "type": "text", + "content": "Other Ranking-based Losses. The literature has explored other ranking-based loss functions with different goals and assumptions. For example, DR Loss [28] constrains a margin to be satisfied between positives and negatives, not taking the recall into account. Smooth AP [1] introduces a smooth, differentiable approximation of AP, with the assumption that " + }, + { + "bbox": [ + 130, + 189, + 482, + 320 + ], + "type": "inline_equation", + "content": "|\\mathcal{N}|" + }, + { + "bbox": [ + 130, + 189, + 482, + 320 + ], + "type": "text", + "content": " is not large. Several studies focused on improving the efficiency of ranking-based losses for use in Support Vector Machines (SVMs) [21, 22]. However, these solutions have been limited to training linear SVMs and hence, problems that can be solved using linear functions. In this paper, we focus on improving ranking-based loss functions [5, 6, 24] that can be used for training complex deep networks such as object detectors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 322, + 483, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 322, + 483, + 395 + ], + "spans": [ + { + "bbox": [ + 131, + 322, + 483, + 395 + ], + "type": "text", + "content": "Comparative Summary. Ranking-based training objective for visual detectors proved their superiority over score-based loss functions in visual detection tasks [5, 6, 23, 24]. However, due to the requirement of pairwise ranking their time complexity is higher than the score-based alternatives. Such a disadvantage of ranking-based loss functions prevents their applicability to larger visual detectors (i.e., transformer-based object detectors)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 418, + 376, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 376, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 376, + 434 + ], + "type": "text", + "content": "3 Background on Ranking-based Losses" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 451, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 451, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 131, + 451, + 482, + 536 + ], + "type": "text", + "content": "Ranking-based losses for object detection rely on pairwise comparisons between the scores of different detections to determine the rank of a detection among positives and negatives. Denoting the score of the " + }, + { + "bbox": [ + 131, + 451, + 482, + 536 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 131, + 451, + 482, + 536 + ], + "type": "text", + "content": " detection by " + }, + { + "bbox": [ + 131, + 451, + 482, + 536 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 131, + 451, + 482, + 536 + ], + "type": "text", + "content": ", we can compare the scores of two detections " + }, + { + "bbox": [ + 131, + 451, + 482, + 536 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 451, + 482, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 451, + 482, + 536 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 131, + 451, + 482, + 536 + ], + "type": "text", + "content": " with a simple difference transform: " + }, + { + "bbox": [ + 131, + 451, + 482, + 536 + ], + "type": "inline_equation", + "content": "x_{ij} = s_j - s_i" + }, + { + "bbox": [ + 131, + 451, + 482, + 536 + ], + "type": "text", + "content": ". By counting the number of instances with " + }, + { + "bbox": [ + 131, + 451, + 482, + 536 + ], + "type": "inline_equation", + "content": "x_{ij} > 0" + }, + { + "bbox": [ + 131, + 451, + 482, + 536 + ], + "type": "text", + "content": ", we can determine " + }, + { + "bbox": [ + 131, + 451, + 482, + 536 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 131, + 451, + 482, + 536 + ], + "type": "text", + "content": " detection's rank among positives " + }, + { + "bbox": [ + 131, + 451, + 482, + 536 + ], + "type": "inline_equation", + "content": "(\\mathcal{P})" + }, + { + "bbox": [ + 131, + 451, + 482, + 536 + ], + "type": "text", + "content": " and all detections (positives and negatives: " + }, + { + "bbox": [ + 131, + 451, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\cup \\mathcal{N}" + }, + { + "bbox": [ + 131, + 451, + 482, + 536 + ], + "type": "text", + "content": ") as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 194, + 548, + 482, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 548, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 194, + 548, + 482, + 574 + ], + "type": "interline_equation", + "content": "\\operatorname {r a n k} ^ {+} (i) = \\sum_ {j \\in \\mathcal {P}} \\bar {H} \\left(x _ {i j}\\right), \\text {a n d} \\operatorname {r a n k} (i) = \\sum_ {j \\in \\mathcal {P} \\cup \\mathcal {N}} \\bar {H} \\left(x _ {i j}\\right), \\tag {1}", + "image_path": "a75899dba54653d739a0d5b5748cbc2088f1f16f7ea7f9873d919f75c6786865.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 584, + 482, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 584, + 482, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 584, + 482, + 611 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 584, + 482, + 611 + ], + "type": "inline_equation", + "content": "\\bar{H}(x)" + }, + { + "bbox": [ + 132, + 584, + 482, + 611 + ], + "type": "text", + "content": " is one if " + }, + { + "bbox": [ + 132, + 584, + 482, + 611 + ], + "type": "inline_equation", + "content": "x > 0" + }, + { + "bbox": [ + 132, + 584, + 482, + 611 + ], + "type": "text", + "content": " and zero otherwise. Since " + }, + { + "bbox": [ + 132, + 584, + 482, + 611 + ], + "type": "inline_equation", + "content": "\\mathrm{d}\\bar{H}(x) / \\mathrm{d}x" + }, + { + "bbox": [ + 132, + 584, + 482, + 611 + ], + "type": "text", + "content": " is either infinite (at " + }, + { + "bbox": [ + 132, + 584, + 482, + 611 + ], + "type": "inline_equation", + "content": "x = 0" + }, + { + "bbox": [ + 132, + 584, + 482, + 611 + ], + "type": "text", + "content": ") or zero (for " + }, + { + "bbox": [ + 132, + 584, + 482, + 611 + ], + "type": "inline_equation", + "content": "x \\neq 0" + }, + { + "bbox": [ + 132, + 584, + 482, + 611 + ], + "type": "text", + "content": "), a smoothed version is used (" + }, + { + "bbox": [ + 132, + 584, + 482, + 611 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 132, + 584, + 482, + 611 + ], + "type": "text", + "content": ": a hyper-parameter):" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 239, + 622, + 482, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 622, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 239, + 622, + 482, + 667 + ], + "type": "interline_equation", + "content": "H (x) = \\left\\{ \\begin{array}{l l} 0, & x < - \\delta \\\\ \\frac {x}{2 \\delta} + 0. 5, & - \\delta \\leq x \\leq \\delta \\\\ 1, & \\delta < x \\end{array} \\right. \\tag {2}", + "image_path": "5ed0a4fc25689f0a13088772b4c2a7ca409f86a646d414408081b489cc7eefee.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "text", + "content": "Yavuz et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 359, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 359, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 359, + 129 + ], + "type": "text", + "content": "3.1 Revisiting Average Precision (AP) Loss" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 411, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 411, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 411, + 148 + ], + "type": "text", + "content": "Given the definitions in Eq. 1, AP Loss [5,6] can be defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 175, + 158, + 481, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 158, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 175, + 158, + 481, + 184 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {A P} = 1 - \\mathrm {A P} = 1 - \\frac {1}{| \\mathcal {P} |} \\sum_ {i \\in \\mathcal {P}} \\operatorname {p r e c i s i o n} (i) = 1 - \\frac {1}{| \\mathcal {P} |} \\sum_ {i \\in \\mathcal {P}} \\frac {\\operatorname {r a n k} ^ {+} (i)}{\\operatorname {r a n k} (i)}. \\tag {3}", + "image_path": "7b53cd7be7d747b398c90ec6880ed467bb87b11e151d9a0a5fedb14421208a0f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 191, + 479, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 191, + 479, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 191, + 479, + 216 + ], + "type": "text", + "content": "Chen et al. [5,6] simplified Eq. 3 by rewriting it in terms of pairwise comparisons between positive and negative detections:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 212, + 224, + 481, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 224, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 212, + 224, + 481, + 251 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {A P} = \\frac {1}{| \\mathcal {P} |} \\sum_ {i \\in \\mathcal {P}} \\sum_ {j \\in \\mathcal {N}} \\frac {H (x _ {i j})}{\\operatorname {r a n k} (i)} = \\frac {1}{| \\mathcal {P} |} \\sum_ {i \\in \\mathcal {P}} \\sum_ {j \\in \\mathcal {N}} L _ {i j} ^ {A P}, \\tag {4}", + "image_path": "812a8963519c0ce215d328a7186a7a5ddbe8172c377c790b785ba8b662fc578d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 258, + 479, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 258, + 479, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 258, + 479, + 293 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 258, + 479, + 293 + ], + "type": "inline_equation", + "content": "L_{ij}^{AP} = H(x_{ij}) / \\mathrm{rank}(i)" + }, + { + "bbox": [ + 132, + 258, + 479, + 293 + ], + "type": "text", + "content": " is called the primary term. Note that " + }, + { + "bbox": [ + 132, + 258, + 479, + 293 + ], + "type": "inline_equation", + "content": "L_{ij}^{AP}" + }, + { + "bbox": [ + 132, + 258, + 479, + 293 + ], + "type": "text", + "content": " is non-differentiable since the step function " + }, + { + "bbox": [ + 132, + 258, + 479, + 293 + ], + "type": "inline_equation", + "content": "(H(\\cdot))" + }, + { + "bbox": [ + 132, + 258, + 479, + 293 + ], + "type": "text", + "content": " applied on " + }, + { + "bbox": [ + 132, + 258, + 479, + 293 + ], + "type": "inline_equation", + "content": "x_{ij}" + }, + { + "bbox": [ + 132, + 258, + 479, + 293 + ], + "type": "text", + "content": " is non-differentiable, which will be discussed next." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 313, + 296, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 313, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 313, + 296, + 326 + ], + "type": "text", + "content": "3.2 Revisiting Identity Update" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 334, + 479, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 334, + 479, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 334, + 479, + 357 + ], + "type": "text", + "content": "Oksuz et al. [24] showed that various ranking-based losses (including AP Loss in Eqs. 3 & 4) can be written in a general form as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 249, + 368, + 481, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 368, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 249, + 368, + 481, + 396 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\frac {1}{Z} \\sum_ {i \\in \\mathcal {P} \\cup \\mathcal {N}} (\\ell (i) - \\ell^ {*} (i)), \\tag {5}", + "image_path": "8d4f612c609d7806c423257fbff093ca5f4ef95197dcf0ef3729de5b3728c094.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 404, + 479, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 404, + 479, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 404, + 479, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 404, + 479, + 453 + ], + "type": "inline_equation", + "content": "\\ell(i)" + }, + { + "bbox": [ + 132, + 404, + 479, + 453 + ], + "type": "text", + "content": " is the ranking-based error (e.g., precision error) computed on the " + }, + { + "bbox": [ + 132, + 404, + 479, + 453 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 132, + 404, + 479, + 453 + ], + "type": "text", + "content": "th detection, " + }, + { + "bbox": [ + 132, + 404, + 479, + 453 + ], + "type": "inline_equation", + "content": "\\ell^{*}(i)" + }, + { + "bbox": [ + 132, + 404, + 479, + 453 + ], + "type": "text", + "content": " is the target ranking-based error (the lowest error possible) and " + }, + { + "bbox": [ + 132, + 404, + 479, + 453 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 132, + 404, + 479, + 453 + ], + "type": "text", + "content": " is the normalization constant. The loss in Eq. 5 can be computed and optimized as follows [24]:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 453, + 480, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 480, + 531 + ], + "type": "text", + "content": "1. Computation of the Loss. First, each pair of logits " + }, + { + "bbox": [ + 132, + 453, + 480, + 531 + ], + "type": "inline_equation", + "content": "(s_i" + }, + { + "bbox": [ + 132, + 453, + 480, + 531 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 453, + 480, + 531 + ], + "type": "inline_equation", + "content": "s_j)" + }, + { + "bbox": [ + 132, + 453, + 480, + 531 + ], + "type": "text", + "content": " are compared by calculating their difference transforms as " + }, + { + "bbox": [ + 132, + 453, + 480, + 531 + ], + "type": "inline_equation", + "content": "x_{ij} = s_j - s_i" + }, + { + "bbox": [ + 132, + 453, + 480, + 531 + ], + "type": "text", + "content": ". With the step function " + }, + { + "bbox": [ + 132, + 453, + 480, + 531 + ], + "type": "inline_equation", + "content": "H(\\cdot)" + }, + { + "bbox": [ + 132, + 453, + 480, + 531 + ], + "type": "text", + "content": " (Eq. 2), the number of detections higher than " + }, + { + "bbox": [ + 132, + 453, + 480, + 531 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 132, + 453, + 480, + 531 + ], + "type": "text", + "content": " (and therefore its precision error, " + }, + { + "bbox": [ + 132, + 453, + 480, + 531 + ], + "type": "inline_equation", + "content": "\\ell(i)" + }, + { + "bbox": [ + 132, + 453, + 480, + 531 + ], + "type": "text", + "content": ") can be easily calculated (see Sect. 3.1). Given " + }, + { + "bbox": [ + 132, + 453, + 480, + 531 + ], + "type": "inline_equation", + "content": "x_{ij}" + }, + { + "bbox": [ + 132, + 453, + 480, + 531 + ], + "type": "text", + "content": ", the loss can be rewritten and calculated in terms of primary terms " + }, + { + "bbox": [ + 132, + 453, + 480, + 531 + ], + "type": "inline_equation", + "content": "L_{ij}" + }, + { + "bbox": [ + 132, + 453, + 480, + 531 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 132, + 453, + 480, + 531 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\frac{1}{Z}\\sum_{i\\in \\mathcal{P}\\cup \\mathcal{N}}\\sum_{j\\in \\mathcal{P}\\cup \\mathcal{N}}L_{ij}" + }, + { + "bbox": [ + 132, + 453, + 480, + 531 + ], + "type": "text", + "content": ", by taking:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 252, + 543, + 480, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 543, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 252, + 543, + 480, + 555 + ], + "type": "interline_equation", + "content": "L _ {i j} = \\left(\\ell (i) - \\ell^ {*} (i)\\right) p (j | i), \\tag {6}", + "image_path": "25f8b6db8facdbdd4d66fef63ab154bea2b7a035f95299666bb7d48dc27cca76.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 563, + 479, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 563, + 479, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 563, + 479, + 600 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 563, + 479, + 600 + ], + "type": "inline_equation", + "content": "p(j|i)" + }, + { + "bbox": [ + 132, + 563, + 479, + 600 + ], + "type": "text", + "content": " is a probability mass function (pmf) that distributes the error computed on the " + }, + { + "bbox": [ + 132, + 563, + 479, + 600 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 132, + 563, + 479, + 600 + ], + "type": "text", + "content": "th example over the " + }, + { + "bbox": [ + 132, + 563, + 479, + 600 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 132, + 563, + 479, + 600 + ], + "type": "text", + "content": "th example in order to determine the pairwise primary term " + }, + { + "bbox": [ + 132, + 563, + 479, + 600 + ], + "type": "inline_equation", + "content": "L_{ij}" + }, + { + "bbox": [ + 132, + 563, + 479, + 600 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 132, + 563, + 479, + 600 + ], + "type": "inline_equation", + "content": "p(j|i)" + }, + { + "bbox": [ + 132, + 563, + 479, + 600 + ], + "type": "text", + "content": " is commonly taken as a uniform distribution [23, 24]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 600, + 479, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 479, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 479, + 636 + ], + "type": "text", + "content": "2. Optimization of the Loss. The gradient of the primary term wrt. to the difference transform " + }, + { + "bbox": [ + 132, + 600, + 479, + 636 + ], + "type": "inline_equation", + "content": "(\\partial L_{ij} / \\partial x_{ij})" + }, + { + "bbox": [ + 132, + 600, + 479, + 636 + ], + "type": "text", + "content": " is non-differentiable. Denoting this term by " + }, + { + "bbox": [ + 132, + 600, + 479, + 636 + ], + "type": "inline_equation", + "content": "\\Delta x_{ij}" + }, + { + "bbox": [ + 132, + 600, + 479, + 636 + ], + "type": "text", + "content": ", we have [5]:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 208, + 646, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 646, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 208, + 646, + 480, + 667 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\mathcal {L}}{\\partial s _ {i}} = \\sum_ {j, k} \\frac {\\partial \\mathcal {L}}{\\partial L _ {j k}} \\Delta x _ {j k} \\frac {\\partial x _ {j k}}{\\partial s _ {i}} = \\frac {1}{Z} \\left(\\sum_ {j} \\Delta x _ {j i} - \\sum_ {j} \\Delta x _ {i j}\\right). \\tag {7}", + "image_path": "8c2ebe00bc2a672e86276b1cd3888781b19f75df423f257ffa8b061b51026f12.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bucketed Ranking-based Losses" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 177 + ], + "type": "text", + "content": "Therefore, optimizing a ranking-based loss function reduces to determining " + }, + { + "bbox": [ + 132, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\Delta x_{ij}" + }, + { + "bbox": [ + 132, + 116, + 482, + 177 + ], + "type": "text", + "content": ". Inspired by Chen et al. [5], Oksuz et al. [24] employ Perceptron Learning [31] and show that " + }, + { + "bbox": [ + 132, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\Delta x_{ij}" + }, + { + "bbox": [ + 132, + 116, + 482, + 177 + ], + "type": "text", + "content": " in Eq. 7 is simply the primary term itself: " + }, + { + "bbox": [ + 132, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\Delta x_{ij} = -(L_{ij}^{*} - L_{ij}) = -(0 - L_{ij}) = L_{ij}" + }, + { + "bbox": [ + 132, + 116, + 482, + 177 + ], + "type": "text", + "content": ", hence the name Identity Update. Plugging this into Eq. 7 yields (see Oksuz et al. [24] for the steps of the derivation):" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 255, + 186, + 481, + 208 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 186, + 481, + 208 + ], + "spans": [ + { + "bbox": [ + 255, + 186, + 481, + 208 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\mathcal {L}}{\\partial s _ {i}} = \\frac {1}{Z} \\left(\\sum_ {j} L _ {j i} - \\sum_ {j} L _ {i j}\\right). \\tag {8}", + "image_path": "04658e268f0c5b6c1eb02ead34a6f74af7390833b2b6a7640136cfe54d2e98e7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 216, + 481, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 240 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 240 + ], + "type": "text", + "content": "Therefore, both computing and optimizing the loss reduces determining the primary term " + }, + { + "bbox": [ + 132, + 216, + 481, + 240 + ], + "type": "inline_equation", + "content": "L_{ij}" + }, + { + "bbox": [ + 132, + 216, + 481, + 240 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 258, + 435, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 258, + 435, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 258, + 435, + 270 + ], + "type": "text", + "content": "3.3 Revisiting AP Loss and RS Loss with Identity Update" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 277, + 482, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 277, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 131, + 277, + 482, + 346 + ], + "type": "text", + "content": "AP Loss [5, 6]. For defining AP Loss with identity update, the errors in Eq. 5 can be derived as " + }, + { + "bbox": [ + 131, + 277, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{R}}(i) = \\frac{N_{\\mathrm{FP}}(i)}{\\mathrm{rank}(i)}, \\ell_{\\mathrm{R}}^{*}(i) = 0" + }, + { + "bbox": [ + 131, + 277, + 482, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 277, + 482, + 346 + ], + "type": "inline_equation", + "content": "Z = |\\mathcal{P}|" + }, + { + "bbox": [ + 131, + 277, + 482, + 346 + ], + "type": "text", + "content": ". Furthermore, defining " + }, + { + "bbox": [ + 131, + 277, + 482, + 346 + ], + "type": "inline_equation", + "content": "p_R(j|i)" + }, + { + "bbox": [ + 131, + 277, + 482, + 346 + ], + "type": "text", + "content": " as a uniform pmf, that is " + }, + { + "bbox": [ + 131, + 277, + 482, + 346 + ], + "type": "inline_equation", + "content": "p_R(j|i) = \\frac{H(x_{ij})}{N_{\\mathrm{FP}}(i)}" + }, + { + "bbox": [ + 131, + 277, + 482, + 346 + ], + "type": "text", + "content": ", we can obtain the primary terms of AP Loss using Identity Update, completing the derivation for computation and optimization (Eq. 8):" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 215, + 355, + 481, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 355, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 215, + 355, + 481, + 376 + ], + "type": "interline_equation", + "content": "L _ {i j} ^ {A P} = \\left\\{ \\begin{array}{l l} \\left(\\ell_ {\\mathrm {R}} (i) - \\ell_ {\\mathrm {R}} ^ {*} (i)\\right) p _ {R} (j | i), & \\text {f o r} i \\in \\mathcal {P}, j \\in \\mathcal {N} \\\\ 0, & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {9}", + "image_path": "ebd4cb31a1384f23d55b1a796d56431ea82916b02ee2563b304c14b8907f64d7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 384, + 481, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 384, + 481, + 409 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 481, + 409 + ], + "type": "text", + "content": "RS Loss [24]. RS Loss includes an additional sorting objective " + }, + { + "bbox": [ + 132, + 384, + 481, + 409 + ], + "type": "inline_equation", + "content": "(\\ell_S(i))" + }, + { + "bbox": [ + 132, + 384, + 481, + 409 + ], + "type": "text", + "content": " to promote better-localized positives to be ranked higher than other positives:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 215, + 417, + 481, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 417, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 215, + 417, + 481, + 449 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {R S}} = \\frac {1}{| \\mathcal {P} |} \\sum_ {i \\in \\mathcal {P}} \\left[ \\underbrace {\\left(\\ell_ {\\mathrm {R}} (i) - \\ell_ {\\mathrm {R}} ^ {*} (i)\\right)} _ {\\text {r a n k i n g}} + \\underbrace {\\left(\\ell_ {\\mathrm {S}} (i) - \\ell_ {\\mathrm {S}} ^ {*} (i)\\right)} _ {\\text {s o r t i n g}} \\right]. \\tag {10}", + "image_path": "36731f0b36d73e95edfd5ddd971e20782259d74034cb15cc5b9bf30b89960bcb.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 457, + 481, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 482 + ], + "type": "text", + "content": "The primary terms of RS Loss can be obtained in a similar manner to AP Loss (see the Supp.Mat. for derivations), yielding:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 214, + 491, + 481, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 491, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 214, + 491, + 481, + 522 + ], + "type": "interline_equation", + "content": "L _ {i j} ^ {R S} = \\left\\{ \\begin{array}{l l} \\left(\\ell_ {\\mathrm {R}} (i) - \\ell_ {\\mathrm {R}} ^ {*} (i)\\right) p _ {R} (j | i), & \\text {f o r} i \\in \\mathcal {P}, j \\in \\mathcal {N} \\\\ \\left(\\ell_ {\\mathrm {S}} (i) - \\ell_ {\\mathrm {S}} ^ {*} (i)\\right) p _ {S} (j | i), & \\text {f o r} i \\in \\mathcal {P}, j \\in \\mathcal {P}, \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {11}", + "image_path": "c12488a125a4b2e071bd3350aa764a7d376cdc0af8570bfe3e95022c5fcf8c1f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 531, + 481, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 481, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 531, + 481, + 567 + ], + "type": "inline_equation", + "content": "p_S(j|i)" + }, + { + "bbox": [ + 132, + 531, + 481, + 567 + ], + "type": "text", + "content": " is the sorting pmf. Note that RS Loss has the same primary terms with AP Loss for " + }, + { + "bbox": [ + 132, + 531, + 481, + 567 + ], + "type": "inline_equation", + "content": "i \\in \\mathcal{P}, j \\in \\mathcal{N}" + }, + { + "bbox": [ + 132, + 531, + 481, + 567 + ], + "type": "text", + "content": ". The primary terms for " + }, + { + "bbox": [ + 132, + 531, + 481, + 567 + ], + "type": "inline_equation", + "content": "i \\in \\mathcal{P}, j \\in \\mathcal{P}" + }, + { + "bbox": [ + 132, + 531, + 481, + 567 + ], + "type": "text", + "content": " directly targets promoting the better-localized positives." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 586, + 388, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 388, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 388, + 597 + ], + "type": "text", + "content": "3.4 Complexity of Ranking-based Loss Functions" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "text", + "content": "These loss functions originally have the space complexity of " + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{O}((|\\mathcal{P}| + |\\mathcal{N}|)^2)" + }, + { + "bbox": [ + 131, + 605, + 482, + 667 + ], + "type": "text", + "content": " [5,6] as they need to compare each pair. This quadratic space complexity makes it infeasible to compute these loss functions using modern GPUs as the number of logits in object detection is very large. To alleviate that, Chen et al. [5] introduced two tricks at the expense of making the computation inefficient: (1) Loop" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "text", + "content": "Yavuz et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "on Positives: By implementing a loop over the positive examples, Chen et al. managed to reduce the time complexity to " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(|\\mathcal{P}|(|\\mathcal{P}| + |\\mathcal{N}|)) \\approx \\mathcal{O}(|\\mathcal{P}||\\mathcal{N}|)" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " since " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "\\mathcal{N} \\gg \\mathcal{P}" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " and the space complexity to " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(|\\mathcal{P}| + |\\mathcal{N}|)" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": ". (2) Discard Trivial Negatives: AP Loss considers negative examples with lower rank than all positives as trivial and disregards them." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 194, + 468, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 194, + 468, + 222 + ], + "spans": [ + { + "bbox": [ + 131, + 194, + 468, + 222 + ], + "type": "text", + "content": "4 Bucketed Ranking-based Loss Functions for Efficient Optimisation of Object Detectors" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 233, + 482, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 233, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 482, + 281 + ], + "type": "text", + "content": "Training object detectors with existing ranking losses has several advantages as we presented before, however, they suffer from significantly large training time. We now address this by our Bucketing Approach, and propose our Bucketed AP and RS Losses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 299, + 395, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 299, + 395, + 312 + ], + "spans": [ + { + "bbox": [ + 131, + 299, + 395, + 312 + ], + "type": "text", + "content": "4.1 Bucketing Negatives for Efficient Optimisation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 319, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 416 + ], + "type": "text", + "content": "Our main intuition is to bucket the sequential negative examples considering that they will have very similar or equal gradients. Fig. 1(a) demonstrates this phenomenon in which " + }, + { + "bbox": [ + 130, + 319, + 482, + 416 + ], + "type": "inline_equation", + "content": "n_1, n_2" + }, + { + "bbox": [ + 130, + 319, + 482, + 416 + ], + "type": "text", + "content": " both are assigned equal gradients as well as each negative among " + }, + { + "bbox": [ + 130, + 319, + 482, + 416 + ], + "type": "inline_equation", + "content": "n_3, n_4, n_5, n_6" + }, + { + "bbox": [ + 130, + 319, + 482, + 416 + ], + "type": "text", + "content": " in the standard AP Loss. Formally, we sort all given logits " + }, + { + "bbox": [ + 130, + 319, + 482, + 416 + ], + "type": "inline_equation", + "content": "s_i \\in \\mathcal{P} \\cup \\mathcal{N}" + }, + { + "bbox": [ + 130, + 319, + 482, + 416 + ], + "type": "text", + "content": " using a conventional sorting methods. Let us denote this sorted permutation of the logits by " + }, + { + "bbox": [ + 130, + 319, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\hat{s}_1, \\hat{s}_2, \\dots, \\hat{s}_{|\\mathcal{P} \\cup \\mathcal{N}|}" + }, + { + "bbox": [ + 130, + 319, + 482, + 416 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 319, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\hat{s}_1 > \\hat{s}_2 > \\dots > \\hat{s}_{|S|}" + }, + { + "bbox": [ + 130, + 319, + 482, + 416 + ], + "type": "text", + "content": ". Given these sorted logits and denoting the " + }, + { + "bbox": [ + 130, + 319, + 482, + 416 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 319, + 482, + 416 + ], + "type": "text", + "content": "th positive logit in the ordering by " + }, + { + "bbox": [ + 130, + 319, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\hat{s}_i^+" + }, + { + "bbox": [ + 130, + 319, + 482, + 416 + ], + "type": "text", + "content": ", the buckets of negatives " + }, + { + "bbox": [ + 130, + 319, + 482, + 416 + ], + "type": "inline_equation", + "content": "B_1, \\dots, B_{|\\mathcal{P} + 1|}" + }, + { + "bbox": [ + 130, + 319, + 482, + 416 + ], + "type": "text", + "content": " can be obtained by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 216, + 425, + 480, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 425, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 216, + 425, + 480, + 441 + ], + "type": "interline_equation", + "content": "B _ {1} > \\hat {s} _ {1} ^ {+} > B _ {2} > \\hat {s} _ {2} ^ {+} > \\dots > \\hat {s} _ {| \\mathcal {P} |} ^ {+} > B _ {| \\mathcal {P} + 1 |}. \\tag {12}", + "image_path": "e47843d89413c2d5ff902dd772ee7e25cb393f24929af638afb6e7292402bbf9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "text", + "content": "We will denote the size of the " + }, + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "text", + "content": "th negative bucket by " + }, + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "text", + "content": ". Having created the buckets, we now determine a single logit value for each bucket, which we call as the prototype logit as it is necessary while assigning the ranking error. We denote the prototype logit for the " + }, + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "text", + "content": " bucket by " + }, + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "inline_equation", + "content": "s_i^b" + }, + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "text", + "content": ". Note that if " + }, + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\delta = 0" + }, + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "text", + "content": " in Eq. 2, then any logit satisfying " + }, + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\hat{s}_{i - 1}^{+} > s_{i}^{b} > \\hat{s}_{i}^{+}" + }, + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "text", + "content": " can be a prototype logit as the ranking stays the same. However, in the case of " + }, + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\delta > 0" + }, + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "text", + "content": ", the boundaries between the logits are smoothed. That's why, practically, we find it effective to use the mean logit of a bucket as its prototype logit." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 544, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 605 + ], + "type": "text", + "content": "Note that this bucketing approach reduces the number of logits (positive and prototype negative) to a maximum of " + }, + { + "bbox": [ + 130, + 544, + 482, + 605 + ], + "type": "inline_equation", + "content": "2|\\mathcal{P}| + 1" + }, + { + "bbox": [ + 130, + 544, + 482, + 605 + ], + "type": "text", + "content": ". As a result, the pairwise errors can now fit into the memory. Consequently, the loop in the red box of Alg. S1 (in Suppl. Mat.) is no longer necessary, which gives rise to efficient ranking-based loss functions which we discuss in the following." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 621, + 364, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 621, + 364, + 633 + ], + "spans": [ + { + "bbox": [ + 131, + 621, + 364, + 633 + ], + "type": "text", + "content": "4.2 Bucketed Ranking-based Loss Functions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Here, we introduce Bucketed versions of AP and RS Losses. Please refer to Alg. S2 (in Suppl. Mat.) for the algorithm." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bucketed Ranking-based Losses" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "text", + "content": "Bucketed AP Loss Definition. To define Bucketed version of AP Loss, we need to define current " + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "inline_equation", + "content": "(\\ell_{\\mathrm{R}}^{b}(i))" + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "text", + "content": " and target ranking errors " + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "inline_equation", + "content": "(\\ell_{\\mathrm{R}}^{b,*}(i))" + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "text", + "content": "th positive. Similar to previous works [5,6,24], we use a target ranking error of 0, i.e., " + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{R}}^{b,*}(i) = 0" + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "text", + "content": ". Defining the current ranking error, similar to Eq. 3, requires " + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "inline_equation", + "content": "\\operatorname{rank}(i)" + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{FP}}(i)" + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "text", + "content": ". These two quantities can easily be defined using the bucket size " + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "inline_equation", + "content": "b_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "text", + "content": " and the prototype logit. Different from the conventional AP Loss, we obtain the pairwise relation " + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "inline_equation", + "content": "H(x_{ij}^{b})" + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "text", + "content": " using the prototype logit " + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "inline_equation", + "content": "s_i^b" + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "text", + "content": ". Particularly, with the ordering in Eq. 12, " + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "inline_equation", + "content": "\\operatorname{rank}(i) = \\sum_{j=1}^{i} H(x_{ij}) + H(x_{ij}^{b}) b_j" + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{FP}}(i) = \\sum_{j=1}^{i} H(x_{ij}^{b}) b_j" + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "inline_equation", + "content": "x_{ij}^{b} = s_j^b - s_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "text", + "content": ". Then, the resulting ranking error computed on " + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 482, + 243 + ], + "type": "text", + "content": "th positive is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 200, + 247, + 481, + 276 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 247, + 481, + 276 + ], + "spans": [ + { + "bbox": [ + 200, + 247, + 481, + 276 + ], + "type": "interline_equation", + "content": "\\ell_ {\\mathrm {R}} ^ {b} (i) = \\frac {N _ {\\mathrm {F P}} (i)}{\\operatorname {r a n k} (i)} = \\frac {\\sum_ {j = 1} ^ {i} H \\left(x _ {i j} ^ {b}\\right) b _ {j}}{\\sum_ {j = 1} ^ {i} H \\left(x _ {i j}\\right) + H \\left(x _ {i j} ^ {b}\\right) b _ {j}}, \\quad i \\in \\mathcal {P}. \\tag {13}", + "image_path": "255b0d543f2dc9b399dcfc95a6ae52bb2a1bf4df26b0dd4c62dcfae98c2e4b0c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 279, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 279, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 131, + 279, + 482, + 365 + ], + "type": "text", + "content": "Optimisation. Here, we need to define the primary terms, as the product between the error and the pmf following Identity Update [24]. Unlike the previous work, the weights of each prototype negative are not equal as a bucket includes varying number of negatives. Therefore, we use a weighted pmf while distributing the ranking error over negatives. Formally, if " + }, + { + "bbox": [ + 131, + 279, + 482, + 365 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 131, + 279, + 482, + 365 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 131, + 279, + 482, + 365 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 131, + 279, + 482, + 365 + ], + "type": "text", + "content": "th prototype negative, then we define the pmf as " + }, + { + "bbox": [ + 131, + 279, + 482, + 365 + ], + "type": "inline_equation", + "content": "p(j^b|i) = b_j / N_{\\mathrm{FP}}(i)" + }, + { + "bbox": [ + 131, + 279, + 482, + 365 + ], + "type": "text", + "content": ". The resulting primary term between the " + }, + { + "bbox": [ + 131, + 279, + 482, + 365 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 279, + 482, + 365 + ], + "type": "text", + "content": "th positive and " + }, + { + "bbox": [ + 131, + 279, + 482, + 365 + ], + "type": "inline_equation", + "content": "j^b" + }, + { + "bbox": [ + 131, + 279, + 482, + 365 + ], + "type": "text", + "content": "th negative is then:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 192, + 369, + 481, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 369, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 192, + 369, + 481, + 397 + ], + "type": "interline_equation", + "content": "L _ {i j} ^ {b} = \\frac {\\sum_ {j = 1} ^ {i} H \\left(x _ {i j} ^ {b}\\right) b _ {j}}{\\sum_ {j = 1} ^ {i} H \\left(x _ {i j}\\right) + H \\left(x _ {i j} ^ {b}\\right) b _ {j}} \\times \\frac {b _ {j}}{N _ {\\mathrm {F P}} (i)}, \\quad i \\in \\mathcal {P}, j \\in \\tilde {\\mathcal {N}} \\tag {14}", + "image_path": "23cad5efd7ee697a20cd02e13ee529361d7f261c4adf14f0c5c452e57e061037.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{N}}" + }, + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "type": "text", + "content": " is the set of prototype negatives. However, we need the primary terms for the " + }, + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "type": "text", + "content": "th negative, not for the prototype ones. Given " + }, + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "type": "inline_equation", + "content": "L_{ij}^{b}" + }, + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "type": "text", + "content": ", the gradients of the prototype negatives and actual negatives can easily be obtained following Identity Update. However, we still need to find the gradients of the actual negatives. To do so, we simply normalise the prototype gradient by its bucket size " + }, + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "type": "inline_equation", + "content": "b_{j}" + }, + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "type": "text", + "content": " and distribute the gradients to actual negatives, completing our method. In Theorem 1, we establish that our Bucketed AP Loss provides exactly the same gradients with AP Loss when " + }, + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\delta = 0" + }, + { + "bbox": [ + 130, + 402, + 482, + 498 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 513, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 513, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 513, + 482, + 586 + ], + "type": "text", + "content": "Bucketed RS Loss Given that we obtained Bucketed AP Loss, converting RS Loss into a bucketed form is more straightforward. This is because (i) the primary term and its gradients of RS Loss are equal to those of AP Loss if " + }, + { + "bbox": [ + 130, + 513, + 482, + 586 + ], + "type": "inline_equation", + "content": "i \\in \\mathcal{P}" + }, + { + "bbox": [ + 130, + 513, + 482, + 586 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 513, + 482, + 586 + ], + "type": "inline_equation", + "content": "j \\in \\mathcal{N}" + }, + { + "bbox": [ + 130, + 513, + 482, + 586 + ], + "type": "text", + "content": " and we simply use " + }, + { + "bbox": [ + 130, + 513, + 482, + 586 + ], + "type": "inline_equation", + "content": "L_{ij}^{b}" + }, + { + "bbox": [ + 130, + 513, + 482, + 586 + ], + "type": "text", + "content": " in such cases; and (ii) if " + }, + { + "bbox": [ + 130, + 513, + 482, + 586 + ], + "type": "inline_equation", + "content": "i \\in \\mathcal{P}" + }, + { + "bbox": [ + 130, + 513, + 482, + 586 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 513, + 482, + 586 + ], + "type": "inline_equation", + "content": "j \\in \\mathcal{P}" + }, + { + "bbox": [ + 130, + 513, + 482, + 586 + ], + "type": "text", + "content": ", then an additional term is included in RS Loss to estimate the pairwise relations between positives (Eq. 11), which is not affected by bucketing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 600, + 274, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 600, + 274, + 611 + ], + "spans": [ + { + "bbox": [ + 131, + 600, + 274, + 611 + ], + "type": "text", + "content": "4.3 Theoretical Discussion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "The first theorem ensures that the gradients provided by our loss functions are identical with their conventional counterparts under certain circumstances. The second states that our algorithm is theoretically faster than the conventional AP and RS Losses." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 101 + ], + "type": "text", + "content": "Yavuz et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "content": "Theorem 1. Bucketed AP Loss and Bucketed RS Loss provide exactly the same gradients with AP and RS Losses respectively when " + }, + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "inline_equation", + "content": "\\delta = 0" + }, + { + "bbox": [ + 132, + 116, + 482, + 140 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 148, + 500, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 148, + 500, + 174 + ], + "spans": [ + { + "bbox": [ + 132, + 148, + 500, + 174 + ], + "type": "text", + "content": "Theorem 2. Bucketed RS and Bucketed AP Losses have " + }, + { + "bbox": [ + 132, + 148, + 500, + 174 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\max((|\\mathcal{P}| + |\\mathcal{N}|) \\log(|\\mathcal{P}| + |\\mathcal{N}|), |\\mathcal{P}|^2))" + }, + { + "bbox": [ + 132, + 148, + 500, + 174 + ], + "type": "text", + "content": " time complexity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 181, + 480, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 181, + 480, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 181, + 480, + 206 + ], + "type": "text", + "content": "As a result, we observe the same performance empirically in less amount of time. Proofs are provided in Supp. Mat." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 224, + 360, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 360, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 360, + 239 + ], + "type": "text", + "content": "5 Bucketed Rank-Sort (BRS) DETR" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 248, + 479, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 479, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 479, + 308 + ], + "type": "text", + "content": "While transformer-based detectors [8,14,16,37,38] have been providing the best performance in several object detection benchmarks, optimizing such detectors using ranking-based losses based on performance measures has not been investigated. Here, we address this gap by incorporating our BRS Loss into CoDETR [38] as the current SOTA detector on the common COCO benchmark [18]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 309, + 480, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 480, + 380 + ], + "type": "text", + "content": "Co-DETR increases the number and variation of the positive examples in the transformer head by leveraging one-to-many assignment strategies in anchor-based detectors (such as ATSS [34] and Faster R-CNN [29]) as auxiliary heads. As the number of proposals is large in such detectors, they can easily provide more and diverse examples, resulting in better performance of the transformer head. Co-DETR is originally supervised with the following loss function:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 222, + 388, + 480, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 388, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 222, + 388, + 480, + 418 + ], + "type": "interline_equation", + "content": "\\sum_ {l = 1} ^ {L} \\left(\\tilde {\\mathcal {L}} _ {l} ^ {D e c} + \\lambda_ {1} \\sum_ {k = 1} ^ {K} \\tilde {\\mathcal {L}} _ {k, l} ^ {A u x} + \\lambda_ {2} \\sum_ {k = 1} ^ {K} \\mathcal {L} _ {k, l} ^ {A u x}\\right), \\tag {15}", + "image_path": "78fce66781103f3162e3bae0767c4fe9f935a2ecd8202f7c6ddd9073f40a6151.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 426, + 480, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 426, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 480, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 426, + 480, + 453 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 132, + 426, + 480, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 426, + 480, + 453 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 132, + 426, + 480, + 453 + ], + "type": "text", + "content": " weigh each loss component. The first two components, " + }, + { + "bbox": [ + 132, + 426, + 480, + 453 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{L}}_l^{Dec}" + }, + { + "bbox": [ + 132, + 426, + 480, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 426, + 480, + 453 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{L}}_{k,l}^{Aux}" + }, + { + "bbox": [ + 132, + 426, + 480, + 453 + ], + "type": "text", + "content": ", are the losses of the " + }, + { + "bbox": [ + 132, + 426, + 480, + 453 + ], + "type": "inline_equation", + "content": "l^{th}" + }, + { + "bbox": [ + 132, + 426, + 480, + 453 + ], + "type": "text", + "content": " decoder layer in the following form:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 233, + 464, + 480, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 464, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 233, + 464, + 480, + 475 + ], + "type": "interline_equation", + "content": "\\lambda_ {c l s} \\mathcal {L} _ {c l s} + \\lambda_ {b b o x} \\mathcal {L} _ {b b o x} + \\lambda_ {I o U} \\mathcal {L} _ {I o U}, \\tag {16}", + "image_path": "cbe9fcd35c4399c7c69b55a63e0a7708cd24ac87305e5593f8d593f4d1cc0664.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 484, + 480, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 484, + 480, + 559 + ], + "spans": [ + { + "bbox": [ + 132, + 484, + 480, + 559 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 132, + 484, + 480, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 132, + 484, + 480, + 559 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 484, + 480, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bbox}" + }, + { + "bbox": [ + 132, + 484, + 480, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 484, + 480, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{IoU}" + }, + { + "bbox": [ + 132, + 484, + 480, + 559 + ], + "type": "text", + "content": " are Focal Loss, L1 Loss and GIoU Loss respectively [38]. " + }, + { + "bbox": [ + 132, + 484, + 480, + 559 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{L}}_l^{Dec}" + }, + { + "bbox": [ + 132, + 484, + 480, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 484, + 480, + 559 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{L}}_{k,l}^{Aux}" + }, + { + "bbox": [ + 132, + 484, + 480, + 559 + ], + "type": "text", + "content": " differ in their input queries. That is, while " + }, + { + "bbox": [ + 132, + 484, + 480, + 559 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{L}}_l^{Dec}" + }, + { + "bbox": [ + 132, + 484, + 480, + 559 + ], + "type": "text", + "content": " follows standard DETR-based models [14, 37] with one-to-one matching of the queries, " + }, + { + "bbox": [ + 132, + 484, + 480, + 559 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{L}}_{k,l}^{Aux}" + }, + { + "bbox": [ + 132, + 484, + 480, + 559 + ], + "type": "text", + "content": " is computed based on positive anchors of " + }, + { + "bbox": [ + 132, + 484, + 480, + 559 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 132, + 484, + 480, + 559 + ], + "type": "text", + "content": " auxiliary head with one-to-many assignment. And " + }, + { + "bbox": [ + 132, + 484, + 480, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{k,l}^{Aux}" + }, + { + "bbox": [ + 132, + 484, + 480, + 559 + ], + "type": "text", + "content": " is the conventional loss of the " + }, + { + "bbox": [ + 132, + 484, + 480, + 559 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 132, + 484, + 480, + 559 + ], + "type": "text", + "content": " auxiliary head, e.g., the weighted sum of classification, localisation and centerness losses for ATSS [34]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 560, + 480, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 560, + 480, + 585 + ], + "spans": [ + { + "bbox": [ + 132, + 560, + 480, + 585 + ], + "type": "text", + "content": "In order to align the training and evaluation objectives better, we replace " + }, + { + "bbox": [ + 132, + 560, + 480, + 585 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{L}}_l^{Dec}" + }, + { + "bbox": [ + 132, + 560, + 480, + 585 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 560, + 480, + 585 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{L}}_{k,l}^{Aux}" + }, + { + "bbox": [ + 132, + 560, + 480, + 585 + ], + "type": "text", + "content": " in Eq. 16 by:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 237, + 597, + 480, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 597, + 480, + 608 + ], + "spans": [ + { + "bbox": [ + 237, + 597, + 480, + 608 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {B R S} + \\lambda_ {b b o x} \\mathcal {L} _ {b b o x} + \\lambda_ {I o U} \\mathcal {L} _ {I o U}, \\tag {17}", + "image_path": "069c29735a644d500a7ad79ed7ecd43baf6b0902b554a73eaafffd3a74982453.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 617, + 489, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 489, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 489, + 667 + ], + "type": "text", + "content": "and set " + }, + { + "bbox": [ + 130, + 617, + 489, + 667 + ], + "type": "inline_equation", + "content": "\\lambda_{bbox}" + }, + { + "bbox": [ + 130, + 617, + 489, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 617, + 489, + 667 + ], + "type": "inline_equation", + "content": "\\lambda_{IoU}" + }, + { + "bbox": [ + 130, + 617, + 489, + 667 + ], + "type": "text", + "content": " dynamically during training to " + }, + { + "bbox": [ + 130, + 617, + 489, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{BRS} / \\mathcal{L}_{bbox}" + }, + { + "bbox": [ + 130, + 617, + 489, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 617, + 489, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{BRS} / \\mathcal{L}_{IoU}" + }, + { + "bbox": [ + 130, + 617, + 489, + 667 + ], + "type": "text", + "content": " to simplify the hyperparameter tuning following Oksuz et al. [24]. Similarly, we replace the loss functions of the auxiliary ATSS [34] and Faster R-CNN [29] heads (" + }, + { + "bbox": [ + 130, + 617, + 489, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{k,l}^{Aux}" + }, + { + "bbox": [ + 130, + 617, + 489, + 667 + ], + "type": "text", + "content": " in Eq. 15) by our BRS Loss. This modification also aligns with the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bucketed Ranking-based Losses" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "objective of the auxiliary heads by the performance measure, hence results in more accurate auxiliary heads [24]. Furthermore, as using BRS Loss with Faster R-CNN does not require sampling thanks to its robustness to imbalance, no limitation is imposed on the number of positives in Faster R-CNN. Hence, the main aim of Co-DETR, that is to introduce more positive examples to the transformer head, is corroborated. As a result, our BRS DETR enables significantly efficient training compared to RS Loss also by improving the performance of Co-DETR." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 219, + 230, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 219, + 230, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 219, + 230, + 232 + ], + "type": "text", + "content": "6 Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 244, + 482, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 244, + 482, + 375 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 482, + 375 + ], + "type": "text", + "content": "We analyze our contributions in three primary sections. First, we evaluate the effectiveness of our bucketing approach by contrasting it with RS Loss and AP Loss across different CNN-based visual detectors, on object detection and instance segmentation tasks. Next, we thoroughly examine how our bucketed RS Loss performs on transformer-based object detectors with different backbones. This is the first time a ranking-based loss is applied to transformer-based object detectors, thanks to the time-efficiency of our bucketing method. Finally, we present comparisons with score-based loss functions and other ranking-based loss functions in the literature. Our experiments clearly show that in terms of training time, accuracy and ease of tuning, our BRS Loss is more preferable over any existing loss function to train object detectors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 376, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 437 + ], + "type": "text", + "content": "Dataset and Performance Measures. Unless otherwise noted, we train all models on COCO trainval35k (115k images) and test them on minival (5k images). We use COCO-style Average Precision (AP) and also report " + }, + { + "bbox": [ + 130, + 376, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 130, + 376, + 482, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 376, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{75}" + }, + { + "bbox": [ + 130, + 376, + 482, + 437 + ], + "type": "text", + "content": " as the APs at IoUs 0.50 and 0.75; and " + }, + { + "bbox": [ + 130, + 376, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathrm{AP_S}" + }, + { + "bbox": [ + 130, + 376, + 482, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 376, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathrm{AP_M}" + }, + { + "bbox": [ + 130, + 376, + 482, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 376, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathrm{AP_L}" + }, + { + "bbox": [ + 130, + 376, + 482, + 437 + ], + "type": "text", + "content": " to present the accuracy on small, medium and large objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 453, + 443, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 453, + 443, + 466 + ], + "spans": [ + { + "bbox": [ + 131, + 453, + 443, + 466 + ], + "type": "text", + "content": "6.1 Efficiency and Effectiveness of Bucketed Ranking Losses" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 569 + ], + "type": "text", + "content": "To comprehensively analyze the effectiveness of our bucketing method, we design experiments on real-world data as well as synthetic data. For the former, we aim to present that our method significantly decreases the training time of the detection and segmentation methods (up to " + }, + { + "bbox": [ + 130, + 474, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\sim 5\\times" + }, + { + "bbox": [ + 130, + 474, + 482, + 569 + ], + "type": "text", + "content": "). As training the entire detector does not isolate the runtime of the loss function, on which our main contribution is, we also design an experiment with synthetic data. This set of experiments show that our bucketing approach, in fact, decreases the loss function runtime by up to " + }, + { + "bbox": [ + 130, + 474, + 482, + 569 + ], + "type": "inline_equation", + "content": "40\\times" + }, + { + "bbox": [ + 130, + 474, + 482, + 569 + ], + "type": "text", + "content": ", thereby resulting in shorter training time of the detectors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Experimental Setup. We follow the experimental setup used in RSLoss [24] only by replacing the loss function by our bucketed loss function to ensure comparability and consistency of results. We report our findings on efficiency and accuracy using the models trained on 4 Tesla A100 GPUs with a batch size of 16, i.e., 4 images/GPU. Unless stated otherwise, we train the models for 12 epochs and test them on images with a size of " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "1333 \\times 800" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " following the common convention [12, 24, 34]. In order to comprehensively demonstrate the efficiency of our approach, we use five different detectors: Faster R-CNN [29], Cascade R-CNN [2]," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Yavuz et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 142, + 147, + 473, + 300 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 1: Bucketed Losses (BRS, BAP) vs. RS & AP Losses on COCO object detection for object detectors. R-X: ResNet-X. Models with R-101 are trained 36 epochs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 142, + 147, + 473, + 300 + ], + "lines": [ + { + "bbox": [ + 142, + 147, + 473, + 300 + ], + "spans": [ + { + "bbox": [ + 142, + 147, + 473, + 300 + ], + "type": "table", + "html": "
MethodDetectorBBLTimeAccuracy
\\(T_{iter}(s) \\downarrow\\)AP↑\\(\\mathrm{AP}_{50} \\uparrow\\)\\(\\mathrm{AP}_{75} \\uparrow\\)
Multi StageFaster R-CNNR-50RS0.5039.459.543.0
BRS\\(0.17 3.0x \\downarrow\\)39.559.542.8
R-101RS0.7547.367.451.2
BRS\\(0.38 2.0x \\downarrow\\)47.767.851.5
Cascade R-CNNR-50RS1.2841.158.744.1
BRS\\(0.24 5.3x \\downarrow\\)41.158.844.2
One StageATSSR-50AP0.3238.158.241.0
BAP\\(0.15 2.1x \\downarrow\\)38.558.641.3
RS0.3639.858.942.6
BRS\\(0.15 2.4x \\downarrow\\)39.858.842.9
PAAAP0.4537.354.341.2
BAP\\(0.30 1.5x \\downarrow\\)37.256.240.2
RS0.5740.858.844.6
BRS\\(0.30 1.9x \\downarrow\\)40.959.044.4
", + "image_path": "f1cd41a8792a9a2992f472c73adff8e8f8b1151369c791eb8977b8cc89f00467.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 332, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 357 + ], + "type": "text", + "content": "ATSS [34] and PAA [12] for object detection as well as Mask R-CNN [10] for instance segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 361, + 481, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 481, + 601 + ], + "type": "text", + "content": "Bucketed Ranking Based Losses Significantly Decrease the Training Time with Similar Accuracy. Multi-stage detectors: Among multi-stage detectors, we train the commonly-used Faster R-CNN and Cascade R-CNN with RS Loss and BRS Loss, and present average iteration time as well as AP of the trained models. In order to evaluate our contribution comprehensively, we train Faster R-CNN also with a stronger setting, in which, we use ResNet-101, train it for 36 epochs using multi-scale training similar to [24]. The results are presented in Table 1, in which we can clearly see that BRS Loss consistently reduces the training time of all three detectors as well as preserves (or slightly improve) their performance. Especially for Cascade R-CNN, which is still a very popular and strong object detector along with its variants such as HTC [4], the training time decreases by " + }, + { + "bbox": [ + 132, + 361, + 481, + 601 + ], + "type": "inline_equation", + "content": "5.3 \\times" + }, + { + "bbox": [ + 132, + 361, + 481, + 601 + ], + "type": "text", + "content": ". This is because the loss function is applied to the logits for three times based on the cascaded nature of this detector, and therefore, our contribution can be easily noticed. Moreover, it might seem that the efficiency gain decreases once the size of the backbone increases from R-50 to R-101 in Faster R-CNN. However, this is an expected result as the overall feature extraction time increases due to the larger number of parameters in the backbone, and hence, as we will discuss in the synthetic experiments, this is independent from our bucketing approach, operating on the loss function after the feature extraction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 606, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 606, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 606, + 481, + 665 + ], + "type": "text", + "content": "One-stage detectors: To show that our gains generalize to one-stage detectors and to AP Loss, we train the common ATSS and PAA detectors with our BAP and BRS Losses, and compare them with AP and RS Losses respectively. Table 1 validates our previous claims: Our bucketed losses obtain similar performance in around half of the training time required for AP and RS Losses." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 318, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 318, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bucketed Ranking-based Losses" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 158, + 296, + 213 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 301, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 301, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 301, + 148 + ], + "type": "text", + "content": "Table 2: Comparison with RS Loss on instance segmentation task on COCO val using Mask R-CNN. BB: Backbone." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 158, + 296, + 213 + ], + "lines": [ + { + "bbox": [ + 135, + 158, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 135, + 158, + 296, + 213 + ], + "type": "table", + "html": "
BBLEfficiencyAccuracy
\\(T_{iter}(s)\\)↓AP↑\\(AP_{50}\\)↑\\(AP_{75}\\)↑
R-50RS0.5636.357.238.8
BRS0.24 2.3x ↓36.257.238.8
R-101RS0.5940.261.843.5
BRS0.27 2.2x ↓40.362.043.8
", + "image_path": "6ec77a103cd2cee55d123e834d1899025f9200a94c1be60e6beea36b91f6f8c4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 315, + 158, + 481, + 213 + ], + "blocks": [ + { + "bbox": [ + 312, + 114, + 480, + 148 + ], + "lines": [ + { + "bbox": [ + 312, + 114, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 312, + 114, + 480, + 148 + ], + "type": "text", + "content": "Table 3: Comparison on different inst. segm. datasets using Mask R-CNN. " + }, + { + "bbox": [ + 312, + 114, + 480, + 148 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{75}" + }, + { + "bbox": [ + 312, + 114, + 480, + 148 + ], + "type": "text", + "content": " is N/A as it is not used for Cityscapes." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 158, + 481, + 213 + ], + "lines": [ + { + "bbox": [ + 315, + 158, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 315, + 158, + 481, + 213 + ], + "type": "table", + "html": "
DatasetLEfficiencyAccuracy
\\(T_{iter}(s) \\downarrow\\)AP↑\\(AP_{50} \\uparrow\\)\\(AP_{75} \\uparrow\\)
CityscapesRS0.4343.568.1N/A
BRS\\(0.19 2.3x \\downarrow\\)43.367.7N/A
LVISRS0.8725.639.227.3
BRS\\(0.35 2.5x \\downarrow\\)25.839.627.4
", + "image_path": "1fcd948665081d87c25ee3e558b8d1d69d7726ac966936af413286c418dcdb3d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 203, + 257, + 410, + 306 + ], + "blocks": [ + { + "bbox": [ + 132, + 224, + 480, + 247 + ], + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 247 + ], + "type": "text", + "content": "Table 4: BRS Loss further improves performance with similar training budget. We report mask AP for instance segmentation" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 203, + 257, + 410, + 306 + ], + "lines": [ + { + "bbox": [ + 203, + 257, + 410, + 306 + ], + "spans": [ + { + "bbox": [ + 203, + 257, + 410, + 306 + ], + "type": "table", + "html": "
TaskDetectorLAP↑AP50↑AP75↑
Object Det.C.R-CNNRS41.158.744.1
BRS42.3+1.260.245.2
Inst.Seg.M.R-CNNRS36.357.238.8
BRS37.3+1.058.440.2
", + "image_path": "0f041e36a3aabba76a9d302524394365ffca9675b43300e2790045164f1e9900.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 329, + 480, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 480, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 480, + 425 + ], + "type": "text", + "content": "Instance Segmentation: Given that our loss functions is efficient in object detectors, one simple extension is to see their generalization to instance segmentation methods. To show that, we train Mask R-CNN, a common baseline, with our BRS Loss on three different datasets from various domains: (i) COCO (Table 2), (ii) Cityscapes [7] as an autonomous driving dataset (Table 3) and LVIS [9] a long-tailed dataset with more than 1K classes (Table 3). The results confirm our previous finding that our BRS Loss decreases the training time significantly by up to " + }, + { + "bbox": [ + 130, + 329, + 480, + 425 + ], + "type": "inline_equation", + "content": "\\sim 2.5\\times" + }, + { + "bbox": [ + 130, + 329, + 480, + 425 + ], + "type": "text", + "content": " compared to RS Loss by preserving its accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 426, + 481, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 481, + 568 + ], + "type": "text", + "content": "Given Similar Training Budget, Bucketed Ranking Based Loss Functions Significantly Improve the Accuracy. There might be several benefits of reducing the training time of the detector as more GPU time is saved. Here, we show a use-case in which we ask the following question: what would happen if we allocate the similar amount of training time to both bucketed and non-bucketed losses? To answer that, we train Cascade R-CNN, a detector, and Mask R-CNN, an instance segmentation method, using our BRS Loss. Considering our speed-ups on these models " + }, + { + "bbox": [ + 130, + 426, + 481, + 568 + ], + "type": "inline_equation", + "content": "(6.0\\times" + }, + { + "bbox": [ + 130, + 426, + 481, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 481, + 568 + ], + "type": "inline_equation", + "content": "2.3\\times)" + }, + { + "bbox": [ + 130, + 426, + 481, + 568 + ], + "type": "text", + "content": ", we simply increased their training epochs from 12 to 36 and 27 respectively. Note that, for Cascade R-CNN, we still spend significantly less amount of training time. Table 4 shows that models trained with the BRS Loss outperform (with " + }, + { + "bbox": [ + 130, + 426, + 481, + 568 + ], + "type": "inline_equation", + "content": "+1.2" + }, + { + "bbox": [ + 130, + 426, + 481, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 481, + 568 + ], + "type": "inline_equation", + "content": "1.0\\mathrm{AP}" + }, + { + "bbox": [ + 130, + 426, + 481, + 568 + ], + "type": "text", + "content": ") their counterparts trained with the RS Loss in both cases." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": "Experiments on Synthetic Data. Up to now, to measure any speed-up, we considered the entire pipeline and did not decouple the loss computation step. However, we only improve the loss computation step. Hence, an analysis focusing only on this part can reveal our contribution more clearly. Therefore, we design a simple experiment using synthetic data. Particularly, we randomly generate " + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": " logits such that " + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "inline_equation", + "content": "m\\%" + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": " of these logits are positive (see Supp.Mat. for data generation details). To cover the wide range of existing detectors, we set " + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "inline_equation", + "content": "L \\in \\{10K, 100K, 1M\\}" + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "inline_equation", + "content": "m \\in \\{0.1, 1.0, 2.0, 5.0\\}" + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": ". Different choices for " + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "inline_equation", + "content": "m" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Yavuz et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 169, + 304, + 232 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 301, + 158 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 301, + 158 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 301, + 158 + ], + "type": "text", + "content": "Table 5: Comparison of our BRS-DETR with Co-DETR (trained with its original loss function) on different backbones on COCO val set." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 169, + 304, + 232 + ], + "lines": [ + { + "bbox": [ + 135, + 169, + 304, + 232 + ], + "spans": [ + { + "bbox": [ + 135, + 169, + 304, + 232 + ], + "type": "table", + "html": "
BackboneDetectorAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_s \\)\\( AP_m \\)\\( AP_1 \\)
ResNet50Co-DETR49.367.254.032.152.663.8
BRS-DETR50.167.454.631.953.965.0
Swin-TCo-DETR51.769.656.434.454.966.8
BRS-DETR52.369.557.134.655.768.0
Swin-LCo-DETR56.975.562.640.161.273.3
BRS-DETR57.275.062.539.461.774.1
", + "image_path": "9fe7d0684473471f44bc12e4c8ca762d1442820a603b2888d8191196fb376881.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 314, + 170, + 480, + 230 + ], + "blocks": [ + { + "bbox": [ + 312, + 116, + 481, + 160 + ], + "lines": [ + { + "bbox": [ + 312, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 312, + 116, + 481, + 160 + ], + "type": "text", + "content": "Table 6: Comparison of our BRS-DETR with DETR variants (trained with their original loss functions) w ResNet-50 on COCO val set. Qu: Queries. Ep: Epochs." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 170, + 480, + 230 + ], + "lines": [ + { + "bbox": [ + 314, + 170, + 480, + 230 + ], + "spans": [ + { + "bbox": [ + 314, + 170, + 480, + 230 + ], + "type": "table", + "html": "
DetectorQuEpAP\\( {\\mathrm{{AP}}}_{50} \\)\\( {\\mathrm{{AP}}}_{75} \\)\\( {\\mathrm{{AP}}}_{\\mathrm{s}} \\)\\( {\\mathrm{{AP}}}_{\\mathrm{m}} \\)\\( {\\mathrm{{AP}}}_{1} \\)
DETR [14]10030042.062.444.220.545.861.1
DN-DETR [15]3005048.667.452.731.052.063.7
DINO [3]9001249.466.953.832.352.563.9
H-DETR [11]3001248.766.452.931.251.563.5
Co-DETR [38]3001249.367.254.032.152.663.8
BRS-DETR3001250.167.454.631.953.965.0
", + "image_path": "14fd7e3f1c0cad8f7980c47c9f2e20063758a197f00fab60002dc10940c7a60a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 258, + 480, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 480, + 319 + ], + "type": "text", + "content": "correspond to the logit distributions of common object detectors. We compute BAP and AP Loss on this synthetic data and compare their runtime to focus on loss computation time. We observed that our BAP Loss has a significant, up to " + }, + { + "bbox": [ + 130, + 258, + 480, + 319 + ], + "type": "inline_equation", + "content": "\\sim 40\\times" + }, + { + "bbox": [ + 130, + 258, + 480, + 319 + ], + "type": "text", + "content": ", speed-up compared to AP Loss. The speed-up becomes more pronounced with increasing number of logits. Supp. Mat. presents detailed results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 349, + 349, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 349, + 349, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 349, + 349, + 361 + ], + "type": "text", + "content": "6.2 The Effectiveness of Our BRS-DETR" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 382, + 480, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 480, + 407 + ], + "type": "text", + "content": "While we incorporate our BRS Loss into Co-DETR, we use all original settings unless otherwise noted. Please refer to Supp.Mat. for the details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 408, + 481, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 481, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 481, + 492 + ], + "type": "text", + "content": "BRS-DETR Outperforms Existing Transformer-based Detectors Consistently. Here, we first employ ResNet-50 backbone as it has been used by many detectors, enabling us to compare our method with many different DETR-based manner in a consistent manner. In this fair comparison setting, our BRS-DETR outperforms all existing DETR variants reaching 50.1 AP as shown in Table 6. For example, our BRS DETR outperforms DN-DETR by 1.5AP with less training epochs and improves Co-DETR as we discuss next." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 495, + 481, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 481, + 579 + ], + "type": "text", + "content": "Training Co-DETR with RS Loss takes 6x less time compared to using RS Loss. Finally, we also compare the training efficiencies of BRS Loss and RS Loss on Co-DETR. We observe that our BRS Loss decreases the training time of Co-DETR by " + }, + { + "bbox": [ + 130, + 495, + 481, + 579 + ], + "type": "inline_equation", + "content": "6.0 \\times" + }, + { + "bbox": [ + 130, + 495, + 481, + 579 + ], + "type": "text", + "content": " (from 4.14s per iteration to 0.69s) in comparison to RS Loss. We note that this is the largest training time gain of our BRS Loss. This is because Co-DETR consists of multiple transformer-based as well as auxiliary heads, requiring multiple loss estimations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "type": "text", + "content": "BRS-DETR Improves Co-DETR over Different Backbones Consistently. In order to show the effectiveness of our BRS Loss, we compare our results with Co-DETR in Table 5 using different backbones. We note that we improve the baseline Co-DETR consistently in all settings. For example, our improvement on ResNet-50 is 0.8AP, which is a notable improvement. However, as the backbone gets larger (e.g., Swin-L [20]), our gains decrease, which is expected and commonly observed in the literature, e.g., [12]." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bucketed Ranking-based Losses" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 179, + 307, + 246 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 301, + 170 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 301, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 301, + 170 + ], + "type": "text", + "content": "Table 7: Comparison with the score-based losses. CE: Cross-entropy. #H: Number of hyperparameters. Ours has very similar " + }, + { + "bbox": [ + 132, + 114, + 301, + 170 + ], + "type": "inline_equation", + "content": "T_{iter}(s)" + }, + { + "bbox": [ + 132, + 114, + 301, + 170 + ], + "type": "text", + "content": " also by being more accurate and simple-to-tune." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 179, + 307, + 246 + ], + "lines": [ + { + "bbox": [ + 135, + 179, + 307, + 246 + ], + "spans": [ + { + "bbox": [ + 135, + 179, + 307, + 246 + ], + "type": "table", + "html": "
DetectorLTiter(s) ↓AP ↑#H
Faster R-CNNCE+L10.1437.69
RS0.5039.43
BRS (Ours)0.1739.53
ATSSFocal L. +G IoU0.1439.35
RS0.3639.81
BRS (Ours)0.1539.81
", + "image_path": "27eb2f1e8afb1a3fe26b3b13199a149fc746375105f54f66bc6c370661fcd876.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 315, + 159, + 479, + 244 + ], + "blocks": [ + { + "bbox": [ + 312, + 116, + 481, + 149 + ], + "lines": [ + { + "bbox": [ + 312, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 312, + 116, + 481, + 149 + ], + "type": "text", + "content": "Table 8: Comparison with other ranking-based losses. Our approach performs better on both accuracy and efficiency." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 159, + 479, + 244 + ], + "lines": [ + { + "bbox": [ + 315, + 159, + 479, + 244 + ], + "spans": [ + { + "bbox": [ + 315, + 159, + 479, + 244 + ], + "type": "table", + "html": "
DetectorL\\(T_{iter}(s) \\downarrow\\)AP ↑#H
Faster R-CNNaLRP [23]0.2837.43
BRS (Ours)0.1739.53
ATSSAP [5]0.3238.15
DR [28]0.2038.15
aLRP [23]0.3237.71
BRS (Ours)0.1539.81
RetinaNetDR [28]0.2937.45
BRS (Ours)0.2338.31
", + "image_path": "e253b80241448a19f850ad5c391bced8d0a1f398ded517acf27f0cfb36074a5b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 478, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 478, + 283 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 478, + 283 + ], + "type": "text", + "content": "6.3 Comparison with Score-based and Other Ranking-based Losses" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 297, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 477 + ], + "type": "text", + "content": "In previous sections, we showed the efficiency of our loss functions compared to AP and RS Losses as their counterparts. Here, we compare our bucketed losses with the score based losses, i.e., Cross-entropy and Focal Loss, and other ranking based losses including DR Loss [28] and aLRP Loss [23]. To do so, in Tables 7 and 8, we report average training time, AP for accuracy and number of hyperparameters to capture the tuning simplicity of the loss functions. Compared to score-based loss functions in Table 7, our BRS Loss improves performance of both approaches, is significantly simpler-to-tune with a low number of hyperparameters and has similar training iteration time. Hence, our BRS Loss is either superior or on par with existing score-based losses. As for other ranking-based losses in Table 8, our BRS Loss is the most efficient and accurate ranking-based loss function for all three detectors. As an example, compared to ATSS trained with DR Loss, our BRS Loss yields 1.8AP better accuracy with " + }, + { + "bbox": [ + 130, + 297, + 482, + 477 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 130, + 297, + 482, + 477 + ], + "type": "text", + "content": " less training time and significantly less number of hyperparameters. Therefore, our bucketed loss functions are now promising alternatives to train object detectors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 502, + 220, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 220, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 220, + 514 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we introduced a novel method to improve the efficiency of ranking-based loss functions by binning negatives into buckets and implementing positive-negative pairwise comparisons between positives and buckets of negatives. We showed that this method improves the computational complexity to a level where pairwise comparisons can be stored as a matrix in memory and running time becomes very close to that of score-based loss functions. Also, for the first time, we integrated a ranking-based loss to Co-DETR, a DETR-based detector, which was possible thanks to our method's lower complexity, and reported improvements on different backbones. Our comprehensive experiments comprising 2 different tasks, 3 different dataset and 7 different detectors - all with positive outcomes - show the general applicability of our approach." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Yavuz et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 482, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 482, + 203 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 482, + 203 + ], + "type": "text", + "content": "We gratefully acknowledge the computational resources provided by TÜBITAK ULAKBIM High Performance and Grid Computing Center (TRUBA), METU Center for Robotics and Artificial Intelligence (METU-ROMER) and METU Image Processing Laboratory. Dr. Akbas is supported by the \"Young Scientist Awards Program (BAGEP)\" of Science Academy, Türkiye." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 224, + 197, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 197, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 197, + 237 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 251, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 251, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 285 + ], + "type": "text", + "content": "1. Brown, A., Xie, W., Kalogeiton, V., Zisserman, A.: Smooth-ap: Smoothing the path towards large-scale image retrieval. In: European Conference on Computer Vision. pp. 677-694. Springer (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 286, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 286, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 286, + 480, + 319 + ], + "type": "text", + "content": "2. Cai, Z., Vasconcelos, N.: Cascade R-CNN: Delving into high quality object detection. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 320, + 480, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 480, + 341 + ], + "type": "text", + "content": "3. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 342, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 480, + 385 + ], + "type": "text", + "content": "4. Chen, K., Pang, J., Wang, J., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Shi, J., Ouyang, W., Loy, C.C., Lin, D.: Hybrid task cascade for instance segmentation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 387, + 480, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 387, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 138, + 387, + 480, + 419 + ], + "type": "text", + "content": "5. Chen, K., Li, J., Lin, W., See, J., Wang, J., Duan, L., Chen, Z., He, C., Zou, J.: Towards accurate one-stage object detection with ap-loss. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 421, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 421, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 421, + 480, + 453 + ], + "type": "text", + "content": "6. Chen, K., Lin, W., li, J., See, J., Wang, J., Zou, J.: Ap-loss for accurate one-stage object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) pp. 1-1 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 454, + 480, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 454, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 138, + 454, + 480, + 496 + ], + "type": "text", + "content": "7. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 498, + 480, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 480, + 540 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 480, + 540 + ], + "type": "text", + "content": "8. Dai, Z., Cai, B., Lin, Y., Chen, J.: Unsupervised pre-training for detection transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence p. 1-11 (2022). https://doi.org/10.1109/tpami.2022.3216514, http://dx.doi.org/10.1109/TPAMI.2022.3216514" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 543, + 480, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 543, + 480, + 574 + ], + "spans": [ + { + "bbox": [ + 138, + 543, + 480, + 574 + ], + "type": "text", + "content": "9. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 576, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 576, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 576, + 480, + 597 + ], + "type": "text", + "content": "0. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask R-CNN. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "type": "text", + "content": "1. Jia, D., Yuan, Y., He, H., Wu, X., Yu, H., Lin, W., Sun, L., Zhang, C., Hu, H.: Detrs with hybrid matching (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "text", + "content": "2. Kim, K., Lee, H.S.: Probabilistic anchor assignment with iou prediction for object detection. In: The European Conference on Computer Vision (ECCV) (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "text", + "content": "3. Law, H., Deng, J.: Cornernet: Detecting objects as paired keypoints. In: The European Conference on Computer Vision (ECCV) (2018)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 318, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 318, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bucketed Ranking-based Losses" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "14. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13619-13627 (June 2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 481, + 182 + ], + "type": "text", + "content": "15. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "type": "text", + "content": "16. Li*, L.H., Zhang*, P., Zhang*, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., Chang, K.W., Gao, J.: Grounded language-image pretraining. In: CVPR (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 481, + 246 + ], + "type": "text", + "content": "17. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 42(2), 318-327 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 246, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 246, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 133, + 246, + 481, + 278 + ], + "type": "text", + "content": "18. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common Objects in Context. In: The European Conference on Computer Vision (ECCV) (2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 278, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 278, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 133, + 278, + 481, + 311 + ], + "type": "text", + "content": "19. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S.E., Fu, C., Berg, A.C.: SSD: single shot multibox detector. In: The European Conference on Computer Vision (ECCV) (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 311, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 481, + 332 + ], + "type": "text", + "content": "20. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 332, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 481, + 354 + ], + "type": "text", + "content": "21. Mohapatra, P., Jawahar, C., Kumar, M.P.: Efficient optimization for average precision svm. Advances in Neural Information Processing Systems 27 (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "text", + "content": "22. Mohapatra, P., Rolínek, M., Jawahar, C., Kolmogorov, V., Pawan Kumar, M.: Efficient optimization for rank-based loss functions. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "text", + "content": "23. Oksuz, K., Cam, B.C., Akbas, E., Kalkan, S.: A ranking-based, balanced loss function unifying classification and localisation in object detection. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "type": "text", + "content": "24. Oksuz, K., Cam, B.C., Akbas, E., Kalkan, S.: Rank & sort loss for object detection and instance segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3009-3018 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "type": "text", + "content": "25. Oksuz, K., Cam, B.C., Kahraman, F., Baltaci, Z.S., Kalkan, S., Akbas, E.: Mask-awareiou for anchor assignment in real-time instance segmentation. In: The British Machine Vision Conference (BMVC) (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 482, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 481, + 515 + ], + "type": "text", + "content": "26. Oksuz, K., Cam, B.C., Kalkan, S., Akbas, E.: Imbalance problems in object detection: A review. IEEE transactions on pattern analysis and machine intelligence 43(10), 3388-3415 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 515, + 481, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 481, + 547 + ], + "type": "text", + "content": "27. Pang, J., Chen, K., Shi, J., Feng, H., Ouyang, W., Lin, D.: Libra R-CNN: Towards balanced learning for object detection. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 547, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 481, + 568 + ], + "type": "text", + "content": "28. Qian, Q., Chen, L., Li, H., Jin, R.: Dr loss: Improving object detection by distributional ranking. In: CVPR (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "text", + "content": "29. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 39(6), 1137-1149 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 601, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 643 + ], + "type": "text", + "content": "30. Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I., Savarese, S.: Generalized intersection over union: A metric and a loss for bounding box regression. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "31. Rosenblatt, F.: The perceptron: A probabilistic model for information storage and organization in the brain. Psychological Review pp. 65-386 (1958)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "Yavuz et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 347 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "32. Shrivastava, A., Gupta, A., Girshick, R.: Training region-based object detectors with online hard example mining. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 194 + ], + "type": "text", + "content": "33. Sun, P., Zhang, R., Jiang, Y., Kong, T., Xu, C., Zhan, W., Tomizuka, M., Li, L., Yuan, Z., Wang, C., Luo, P.: SparseR-CNN: End-to-end object detection with learnable proposals. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "text", + "content": "34. Zhang, S., Chi, C., Yao, Y., Lei, Z., Li, S.Z.: Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "type": "text", + "content": "35. Zhang, X., Wan, F., Liu, C., Ji, R., Ye, Q.: Freeanchor: Learning to match anchors for visual object detection. In: Advances in Neural Information Processing Systems (NeurIPS) (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "text", + "content": "36. Zheng, Z., Wang, P., Liu, W., Li, J., Rongguang, Y., Dongwei, R.: Distance-iou loss: Faster and better learning for bounding box regression. In: AAAI Conference on Artificial Intelligence (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 482, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 326 + ], + "type": "text", + "content": "37. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. ArXiv abs/2010.04159 (2020), https://api-semanticscholar.org/CorpusID:222208633" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 326, + 482, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 482, + 347 + ], + "type": "text", + "content": "38. Zong, Z., Song, G., Liu, Y.: Detrs with collaborative hybrid assignments training (2022)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 318, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 318, + 91, + 447, + 102 + ], + "type": "text", + "content": "Bucketed Ranking-based Losses" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/BugNIST - a Large Volumetric Dataset for Detection under Domain Shift/fceefb33-6fd3-4a63-9cb1-f0d439dc4d49_content_list.json b/2024/BugNIST - a Large Volumetric Dataset for Detection under Domain Shift/fceefb33-6fd3-4a63-9cb1-f0d439dc4d49_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3440f9751a60bb60f45c4bc745b6eceaaf238142 --- /dev/null +++ b/2024/BugNIST - a Large Volumetric Dataset for Detection under Domain Shift/fceefb33-6fd3-4a63-9cb1-f0d439dc4d49_content_list.json @@ -0,0 +1,1605 @@ +[ + { + "type": "text", + "text": "BugNIST - a Large Volumetric Dataset for Object Detection under Domain Shift", + "text_level": 1, + "bbox": [ + 254, + 140, + 750, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Patrick Møller Jensen, Vedrana Andersen Dahl, Rebecca Engberg, Carsten Gundlach, Hans Marin Kjer, and Anders Bjorholm Dahl", + "bbox": [ + 245, + 212, + 754, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Technical University of Denmark, Kgs. Lyngby, Denmark {patmjen, vand, reen}@dtu.dk, cagu@fysik.dtu.dk, {hmjk, abda}@dtu.dk", + "bbox": [ + 256, + 255, + 746, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Domain shift significantly influences the performance of deep learning algorithms, particularly for object detection within volumetric 3D images. Annotated training data is essential for deep learning-based object detection. However, annotating densely packed objects is time-consuming and costly. Instead, we suggest training models on individually scanned objects, causing a domain shift between training and detection data. To address this challenge, we introduce the BugNIST dataset, comprising 9154 micro-CT volumes of 12 bug types and 388 volumes of tightly packed bug mixtures. This dataset is characterized by having objects with the same appearance in the source and target domains, which is uncommon for other benchmark datasets for domain shift. During training, individual bug volumes labeled by class are utilized, while testing employs mixtures with center point annotations and bug type labels. Together with the dataset, we provide a baseline detection analysis, with the aim of advancing the field of 3D object detection methods.", + "bbox": [ + 261, + 323, + 740, + 531 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Volumetric Dataset, Benchmark, Volumetric Object Detection, Domain Shift.", + "bbox": [ + 261, + 545, + 740, + 574 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 625, + 375, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our work on domain shift in volumetric 3D images is motivated by the need for labeled data to train supervised deep learning models for volumetric imaging. The problem that we aim to solve is object detection and classification. We propose to label images of objects scanned as isolated entities as the source domain and use these as a basis to train models for object detection and classification in a complex context of mixed objects and other materials as the target domain. The effort needed for obtaining labels in the two domains is significantly different. If objects are isolated, they can be automatically labeled whereas mixed objects require expensive manual labeling. Automating the labeling based on isolated objects, however, leads to a domain shift between the data in the source domain for training and the data in the target domain for detection and classification. This domain shift is special because the appearance of the objects is the same", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ba200f363445f099ca69ad44d26396a72c9d667e3c50bc5e33df3352a7dc9d47.jpg", + "image_caption": [ + "(a) 9154 volumes of individual bugs with class labels." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 475, + 276 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/265b46fe7ce8a15fda6f80c106c3407d9da56e3764376b9caa07cd0a91feea43.jpg", + "image_caption": [ + "(b) 388 volumes of bug mixtures with center point and class labels.", + "Fig. 1: Overview of the two parts of the dataset. In (a), representative individual bug volumes are displayed with a class abbreviation, name, and number of volumes. In (b), a few volumes of bug mixtures are shown. Each bug is annotated with bug class and center point." + ], + "image_footnote": [], + "bbox": [ + 480, + 143, + 790, + 275 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in the two domains, but the surrounding context is different. We use the term context shift for this type of domain shift.", + "bbox": [ + 212, + 419, + 784, + 448 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To promote the development of new deep learning-based methods that can handle context shift, we contribute the volumetric dataset called BugNIST1. The dataset contains 9154 micro-CT volumes with individual bugs and 388 micro-CT volumes containing a mixture of several bugs as well as other materials such as leaves, wood shavings, etc.; see Fig. 1. The challenge is to use the individually scanned bugs to detect and classify the bugs in the mixtures. The dataset is intended to be a relatively simple test case for starting to develop methods for detecting objects in volumetric images under context shift, and bugs are chosen due to their complex shape and not for solving a problem in entomology.", + "bbox": [ + 212, + 450, + 787, + 585 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As a basis for training an accurate deep learning model, a training dataset for object detection and classification must capture the appearance distribution of the object classes that are in the dataset. Since the appearance of a CT-scanned object changes minimally if it is scanned individually or as part of a mixture, the appearance distribution can be covered by only considering the individual scans. This is different from 2D photos. Here additional variations in scale, viewpoint, and illumination influence the appearance of objects, and it would be more difficult to acquire representative images of individual objects for a dataset based on 2D photos.", + "bbox": [ + 212, + 587, + 787, + 720 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since BugNIST is a test case for context shift, we wanted to ensure that each class is covered well. By scanning systematically arranged bugs, we have generated 9154 volumes of individual bugs, each labeled with its corresponding class. Furthermore, we scanned 388 mixed volumes, each containing between 4 and 45 bugs alongside other materials. For evaluation purposes, each bug in the mixtures is manually annotated with its center position and class label. In total,", + "bbox": [ + 212, + 723, + 787, + 813 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "P.M. Jensen et al.", + "bbox": [ + 271, + 114, + 392, + 126 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1 https://abdahl.github.io/bugnist/", + "bbox": [ + 217, + 823, + 491, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "9542 volumes make up the BugNIST dataset. The numbers in each class are shown in Fig. 1. Since the bugs in each class are the same age, we expect to cover the biological variation sufficiently to be representative of the class.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Solutions to the type of context shift, that we present here, namely a domain shift where only the context changes, and not the appearance of the objects, have wide applications in volumetric imaging. The field of volumetric imaging is dominated by medical applications [39], but there is a growing use beyond the medical area based on e.g. micro-CT [58, 63]. In all these application areas, there is a need to employ supervised deep learning to solve problems that require quantification of depicted structures. When the use case is novel and there is no directly applicable training data available — a typical scenario in volumetric imaging of e.g. materials microstructure, but also in specialized biomedical applications — it is necessary to create training data. If labor-intensive manual labeling can be limited or avoided, there is potentially a huge gain in productivity. Examples of applications are detecting and classifying objects in scans of marine sediments such as microplastics [55, 68], museum specimens [20], or detecting threat items in baggage security screening [56].", + "bbox": [ + 212, + 194, + 784, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In volumetric scans, manual labeling is significantly more demanding than in 2D. For instance, annotating a bounding box of a single object in 3D requires traversing slices of the volume and making a mental model of the object. If the volume contains densely packed complex objects, determining their bounding boxes becomes difficult. Further, bounding boxes may overlap to such an extent that they stop being a useful way to locate objects. This is why we employ center points as the annotation format in the mixtures.", + "bbox": [ + 212, + 409, + 785, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The name BugNIST is a homage to the influential MNIST dataset [29] which has played a pivotal role in the development and evaluation of deep learning methods. Similar to the way that MNIST has been used, BugNIST is intended to be a benchmark dataset that represents a well-defined problem, where the development of solutions will make a basis for solutions that can advance the field. Specifically for BugNIST, solutions will enable object detection and classification with significantly less effort put into creating the data foundation for training. This calls for methods that can learn to predict objects in a complex context from volumes of individual objects, namely solving the context shift problem. To our knowledge, no existing datasets have this property.", + "bbox": [ + 212, + 517, + 785, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The term context shift refers to a shift in the data distribution between training and inference. Typical shifts in data distributions arise from changes in data acquisition systems, e.g. the use of different scanners, or from variations in the objects of interest between datasets. Such problems have been addressed using domain adaptation [18,59]. In our setting, there is almost no change in the appearance of the bugs between volumes of individual bugs and mixtures, but the change is in what surrounds the bugs, i.e. the change is in the context.", + "bbox": [ + 212, + 671, + 785, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We hypothesize that methods capable of learning under context shift might also prove capable of learning from very few examples, similar to how humans learn [67]. Instead of relying on seeing objects in every imaginable context, such methods could instead isolate objects from their contexts. If this hypothesis holds", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "BugNIST - Object Detection under Domain Shift", + "bbox": [ + 398, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "true, learning under context shift may reduce the need for extensive training data and the associated training costs. This is in contrast to the current trend, where larger models and datasets as seen in LLMs [69] and computer vision [28, 35], lead to better performance, but also limit who can effectively train such models.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Contributions We contribute BugNIST, a new volumetric image dataset for benchmarking domain shift. In this paper, we use BugNIST for benchmarking detection methods under context shift, trained only on individual bugs and tested on mixtures. We provide several baselines and investigate different training strategies for existing volumetric detection methods.", + "bbox": [ + 212, + 218, + 784, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our results show that volumetric object detection under context shift is difficult. We obtain reasonable results in localizing the bugs, but when a class label is assigned, there is a significant drop in performance. This is even though the classification of individual bugs is easy. Our experiments demonstrate great potential for developing new approaches for 3D object detection of complex objects.", + "bbox": [ + 212, + 292, + 785, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 388, + 382, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In computer vision, the term 3D data refers to different data types, including point clouds, surfaces, volumes, and depth images. Here, we are only concerned with 3D volumes, i.e. image data represented on a voxel grid.", + "bbox": [ + 212, + 417, + 782, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "BugNIST is created to advance deep learning methods for object detection and classification under domain shift. Especially in medical imaging, analysis of 3D volumes is in great demand. A review of medical challenges illustrates this [36]. Their study shows that detection makes up $7\\%$ of the tasks, classification makes up $10\\%$ of the tasks, whereas segmentation accounts for $70\\%$ of the tasks. Detection is important for example for diagnosis [7, 19, 27, 49, 71]. Outside the medical domain, object detection problems include baggage security screening [15, 61, 62], although the problem and the available data are still primarily in the 2D domain [56], food inspection [12] such as fruit quality [24], grain classification [6], foreign object detection [13], biological applications such as analysis of weeds [64] and other uses driven by developments in X-ray micro-CT imaging [42], etc.", + "bbox": [ + 212, + 462, + 785, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The domain shift problem presented in BugNIST, where only the context changes, and not the appearance of the objects of interest, is not widely studied. Most research on domain shifts focuses primarily on problems where the appearance of objects changes. Solutions focus on domain adaptation, where labeled data in the source domain are available together with data without labels from the target domain [40], or domain generalization, where the inference domain is not known at training time [37, 60]. We focus on domain adaptation, which is also the typical use case for volumetric imaging.", + "bbox": [ + 212, + 628, + 785, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Several widely used datasets are available for studying and developing methods for domain shift. For 2D image classification problems the datasets PACS [30], Digits-DG [70], Office-Home [57], and DomainNet [41] developed to classify images from domains such as photos, art paintings, sketches, etc. Domain shift between artificially generated and real images e.g. for segmentation is another area that receives much attention [48] with standard datasets such as", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "P.M. Jensen et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SYNTHEIA [47], CITYSCAPES [9], and GTA-5 [45], or completely synthetic domain shift like the SHIFT dataset [52] for autonomous driving.", + "bbox": [ + 212, + 145, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In 3D, domain shift is often studied in medical applications where the same type of object is imaged with different scanning modalities [32]. It can be MRI (magnetic resonance imaging), CT, OCT (optical coherence tomography), etc. Standard datasets include the CrossMoDA 2021 dataset [11] and the BRATS dataset [3], both of different MRI modalities. There are many other datasets used for domain adaptation for volumetric medical imaging [32], but all for adapting between modalities or between the same modality but different scanners [18].", + "bbox": [ + 212, + 179, + 785, + 285 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The structure of the imaged objects in BugNIST is different from most existing volumetric deep learning benchmark datasets. The morphological complexity of bugs is, however, representative of volumetric imaging data of e.g. 3D microstructure, where there is much less use of deep learning than in e.g. the medical domain [63].", + "bbox": [ + 212, + 287, + 785, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Platforms like Zenodo [14], TomoBank [10], or the Human Organ Atlas [58] also contain a large amount of volumetric image datasets. Different from e.g. medical challenges, this data generally comes with no clear problem formulation, well-defined tasks or even labels, that allow for a comparison of methods. Further, their large volume sizes are difficult to access. It can also be difficult to judge the complexity of the data and if deep learning is needed. Some images can easily be analyzed using a threshold whereas others are complex and can greatly benefit from deep learning. Therefore, most of this data is not attractive to use as a benchmark for developing new deep learning methods. Some datasets are emerging, e.g. the vessel segmentation challenge [23], but they remain exceptions.", + "bbox": [ + 212, + 364, + 785, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "There is extensive research in deep learning-based object detection, but primarily in 2D. Most methods either use a two-stage detector such as R-CNN [16, 17, 44], with a region proposal and a classifier to choose among the proposed regions, or a one-stage detector such asYOLO [43], where regions and classes are proposed in a single network.", + "bbox": [ + 212, + 518, + 785, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Object detection in volumetric images is less common. One approach has been to apply a 3D U-Net [8] segmentation network and use each connected component of the foreground label class as a detection [22]. An example of using this approach is lung nodule detection in [65]. Employing segmentation for object detection comes with the problem of separating objects that may be connected or combining separated segments into one object. To avoid this problem, Jaeger et al. [22] propose the Retina U-Net: a one-shot 3D detection algorithm that combines the Retina Net [33] and U-Net [46] giving detections with 3D bounding boxes. Despite the good performance of Retina U-Net, changing to a new 3D detection problem requires a time-consuming method configuration. This has been addressed in nnDetection [4]. Inspired by the nnU-Net (no new U-Net) [21], they propose a no new 3D detection method. On the BugNIST data, we test detection with 3D U-Net [8], the nnDetection method [4], and a 3D adaption of the Faster R-CNN model [44].", + "bbox": [ + 212, + 595, + 785, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmark datasets for volumetric object detection are primarily medical. Examples include Baumgartner et al. [4], which uses 13 datasets to develop the", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "BugNIST - Object Detection under Domain Shift", + "bbox": [ + 398, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "nnDetection method. Two of these datasets were originally targeted 3D detection and include the LUNA16 for nodule detection [50], which is a subset of the LIDC dataset [2] and contains 888 CT scans of lungs, and the ADAM dataset [53] for detecting aneurysms in MRI scans. ADAM is made up of 254 MRI scans. The rest of the detection datasets are medical segmentation datasets such as data from the Medical Segmentation Decathlon [51], the rib fracture detection [26], the PROSTATEx dataset [1], and others. The segmentation datasets have been transformed into detection data by identifying center positions and bounding boxes for the segments. Within baggage security screening, some 3D datasets are obtainable, for example, the ALERT datasets used in [62]. Due to the sensitive nature of the problem, these datasets have copyright restrictions and limited availability under a non-disclosure agreement.", + "bbox": [ + 212, + 146, + 787, + 328 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3 Data acquisition and processing", + "text_level": 1, + "bbox": [ + 215, + 353, + 560, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 1 gives an overview of the BugNIST dataset. Each individual bug volume contains exactly one bug located near the center of the volume. The total of 9154 individual bug volumes have between 713 and 946 volumes of each class distributed across 12 bug classes (species and stages). Bug mixtures contain several bugs, leaves, cotton, shredded paper, and wood shaving. There are in total 388 mixtures. The mixtures contain 4 to 45 bugs, totaling 5011 bugs with an average of 13 bugs per mixture, see Fig. 2. In bug mixtures, we have manually marked the center positions of each bug and provided it with a class label.", + "bbox": [ + 212, + 385, + 787, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All scans were acquired using a laboratory micro-CT scanner (Nikon metrology scanner) that was set to record volumes of 2000 voxels cubed with an isotropic voxel size of $26.62\\mu \\mathrm{m}$ . Being able to acquire such large volumes allowed us to scan multiple bugs and bug mixtures at once, and then crop them into smaller volumes. The dataset was collected through nearly 250 scans, each taking 30 minutes, over 90 days. Care has been taken that the scanning parameters stay unchanged for all scans, and we have calibrated the scanner daily.", + "bbox": [ + 212, + 507, + 787, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All bugs were placed in tubes, either as individuals or in mixtures. We used cotton, which is almost transparent to X-rays, to separate the bugs, either as individual bugs or mixtures. For the mixtures, we varied the difficulty of the bug detection task by adding leaves, shredded paper, or wood shavings to some of the mixtures, and also varied the number of bugs, their class, and the packing density.", + "bbox": [ + 212, + 613, + 787, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The cylindrical shape of the tubes was easy to segment using a circular Hough-transform [66] and sparse layered graphs [25], leaving us with a cylinder of only bugs and cotton. Due to the low intensity of cotton compared to the bugs, we could easily divide each cylinder into volumes containing one bug each. The volumes have been manually inspected and cleaned to correct cases where something went wrong, e.g. parts of another bug showing in the periphery of the volume. The final volumes of scans of individual bugs are $450 \\times 450 \\times 900$ voxels, the size which was chosen to accommodate the biggest bugs. The mixtures were segmented similarly and the resulting volumes are $650 \\times 650 \\times 900$ voxels. All", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "P.M. Jensen et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/249e81cfd94ad453faa8285871dd5e92a0e132261ca3aa141934562277587715.jpg", + "image_caption": [ + "(a) Class distribution for bugs from mixture volumes." + ], + "image_footnote": [], + "bbox": [ + 230, + 146, + 472, + 204 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4b0d018cfc209628bed82289d0347f3ab40db26b9b666bb8624bf56be349bb4b.jpg", + "image_caption": [ + "Fig. 2: Summary of volumes from scans of bug mixtures." + ], + "image_footnote": [], + "bbox": [ + 496, + 146, + 633, + 214 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b5103bb97060de34a90c6d8d5b6a0968158601f4e6acb9dd4044e2e4f785c45f.jpg", + "image_caption": [ + "(b) Distribution of the number of bugs per volume and the number of classes per volume." + ], + "image_footnote": [], + "bbox": [ + 638, + 148, + 771, + 214 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "volumes are stored as 8-bit unsigned integers with their original spatial resolution of $26.62\\mu \\mathrm{m}$ (physical voxel size). Our dataset comprises 1.8 trillion voxels, approximately a quarter of the RGB values in ImageNet.", + "bbox": [ + 212, + 295, + 782, + 340 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The bugs scanned for the BugNIST dataset are bred as pet food and have been frozen before being packed for scanning. Despite careful packing, some scanned bugs may appear damaged, e.g. a few crickets have broken off antennas or miss a leg. We have chosen to accept these defects because they reflect some variation in a population and add to the challenge of the dataset.", + "bbox": [ + 212, + 340, + 784, + 416 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To create additional lightweight datasets, we created four downsized versions of all volumes. The individual bug volumes, originally 900 voxels long, have lengths of 512, 256, 128, and 64 voxels in the downscaled versions, and all the other lengths are scaled accordingly. The voxel size of the coarsest data is $374.3\\mu \\mathrm{m}$ , which is a scaling factor of approximately 14 compared to the original scan.", + "bbox": [ + 212, + 417, + 784, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 532, + 375, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To establish baselines for volumetric detection under context shift (domain shift with no change in object appearance), we have selected and adapted three existing detection models. Furthermore, we have applied three training strategies, yielding a total of nine baseline models. The diversity in model choices and training strategies contributes to a comprehensive set of baseline models that provide a strong foundation for object detection under context shift. We briefly describe the selected methods and how they were trained. Full details are provided in the supplementary material. To determine the difficulty of differentiating between the individual bugs, we also performed a classification experiment which is described in the supplementary material.", + "bbox": [ + 212, + 564, + 784, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All experiments are carried out using the x128 data, which is a compromise between memory needs and the preservation of structural details of the bugs.", + "bbox": [ + 212, + 715, + 782, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Training strategies for detection under context shift", + "text_level": 1, + "bbox": [ + 214, + 768, + 689, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To provide a simple baseline investigation for the BugNIST dataset, we investigate some strategies for solving the detection problem under context shift. Since the purpose of our paper is to initiate research in domain adaptation methods", + "bbox": [ + 212, + 794, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "BugNIST - Object Detection under Domain Shift", + "bbox": [ + 397, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/633471798fd54ad57ec676f5d61247ac367820706fe92530f776fc06a3cf80ae.jpg", + "image_caption": [ + "Fig. 3: The top line shows the automated annotation pipeline for individual bugs. By thresholding, bounding boxes and segmentation masks are easily obtained. The bottom shows the synthetic mixtures used for the more advanced training strategy." + ], + "image_footnote": [], + "bbox": [ + 250, + 152, + 748, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "for volumetric data under context shift, this baseline study is intended as an offset for future developments based on the BugNIST dataset. Our first attempt is to train models from the volumes of individual bugs. Our two other attempts are to use a data augmentation strategy by creating synthetic mixtures. Data augmentation has previously been shown to greatly benefit domain shift problems [60]. The automated annotation of the BugNIST data and the data that we use for training detection algorithms are illustrated in Fig. 3.", + "bbox": [ + 212, + 436, + 787, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Single bugs As a first stage, we train the models to detect the bugs in the individual scans. For models requiring ground truth segmentation labels, those are generated by thresholding the volumes and only keeping the largest connected component. The value of the class label then replaces the foreground voxels. This creates segmentation labels where each voxel has one of 13 values including background or one of the twelve bugs. For models requiring ground truth bounding boxes, those are extracted from each connected component. This process of preparing the training data using this simplest strategy is illustrated in the top line of Fig. 3.", + "bbox": [ + 212, + 555, + 787, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Synthetic mixtures As a second stage, we train the models on artificial mixes. These are generated by randomly sampling individual bug scans and adding them to a common volume. We start with an empty volume of $256^3$ and we set a random target number of bugs drawn from a normal distribution, $\\mathcal{N}(32,6)$ . A random bug from a random class (both uniform distributions) is drawn. To create variation, the selected sample is flipped around a random axis, rotated a random multiple (0-3) of 90 degrees around the first axis, and then transposed randomly. The bug volume is intensity normalized, smoothed with a Gaussian filter with $\\sigma = 0.5$ , and thresholded at 0.2. We draw a random voxel position", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "P.M. Jensen et al.", + "bbox": [ + 271, + 114, + 392, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "from the volume and only insert the bug if its bounding box does not exceed the volume boundaries nor overlaps with existing labeled voxels. We attempt a maximum of 1000 random positions to achieve a valid insertion before the bug is discarded. In total 300 synthetic mixtures are created.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Crowded synthetic mixtures Finally, since the bugs are packed very close together in the real scans, we create crowded synthetic mixtures. Here the bugs are packed closer by reducing the volume size to $200 \\times 200 \\times 200$ , adjusting the target number of bugs to $\\mathcal{N}(64,8)$ , and relaxing the insertion criteria such that no labeled voxels can be overwritten. This effectively allows for a larger overlap between bounding boxes and that bugs can touch each other. Again, a total of 300 synthetic mixtures are created.", + "bbox": [ + 212, + 220, + 789, + 327 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Baseline detection models", + "text_level": 1, + "bbox": [ + 215, + 349, + 478, + 363 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Despite important uses of detecting objects in 3D volumes, the number of methods addressing volumetric object detection is much smaller compared to 2D, and only a few methods are readily available for volumes. So, to establish baselines for detection under context shift, we have selected and adapted three detection approaches. They vary in how they utilize training data; relying on either ground truth segmentation labels, ground truth bounding boxes, or both.", + "bbox": [ + 212, + 373, + 787, + 465 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Detection U-Net Inspired by [4] we use U-Net [8] as one baseline detection model. U-Net is a dominating model for volumetric segmentation, making it a natural choice for adapting into a baseline detection model. Detection U-Net is trained on volumes accompanied by ground truth segmentation labels. It is first trained to predict the segmentation of the single bugs. Then, starting from the obtained weights, it is trained to predict the segmentation of the synthetic mixtures. Finally, again starting from the previous weights, it is trained to predict the segmentation in the crowded synthetic mixtures. For all training runs, we augment the images with random rotations, translations, geometric scaling, and intensity scaling. We train the models using the sum of Dice and cross entropy loss. During training we hold out a validation split which we evaluate on after every epoch. At the end of training, the model which scored the highest on the validation split is kept. Like [4], we extract bounding boxes by thresholding the softmax probabilities (at 0.75) and then extracting the connected components for each channel. Only components larger than $125(5^{3})$ voxels are kept. We use the U-Net implementation from MONAI [5].", + "bbox": [ + 212, + 478, + 787, + 720 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Faster R-CNN As an example of region proposal networks, we have chosen Faster R-CNN [44] with the modifications from [31]. We have adapted the implementation from torchvision [54] to allow for 3D detection. We train in the same sequence as the U-Net: single bugs, synthetic mixtures, and crowded synthetic mixtures. We use the same augmentations as for the U-Net. In addition, we also randomly perturb the bounding box corners. The model is only trained on bounding boxes and does not make use of the voxel labels. The model is", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "BugNIST - Object Detection under Domain Shift", + "bbox": [ + 398, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "trained with smooth L1 loss for bounding box regression and cross-entropy for objectness and class predictions. For full details, we refer to [31, 44]. As with U-Net, we hold out a validation split which is evaluated after each epoch. After training, the model scoring the highest on the validation split is kept. For inference, only bounding boxes with an objectness score over 0.05 are kept.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "nnDetection The third detection model, nnDetection, trains the RetinaUNet [22] detection model using a combination of segmentation masks and bounding boxes. The bounding box branch is trained using a generalized IoU loss for regression and cross-entropy for classification. The voxel segmentation branch is trained with a sum of Dice and cross-entropy loss. As nnDetection is a self-configuring framework, we refer to [4, 21] for further training and model details. We simply provide the data as instructed. We train a separate model for single bugs, synthetic mixtures, and crowded synthetic mixtures. nnDetection trains using 5-fold cross-validation, and the final predictions are made using an ensemble over the best model from each fold [4].", + "bbox": [ + 212, + 237, + 789, + 388 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Evaluation", + "text_level": 1, + "bbox": [ + 215, + 411, + 352, + 425 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We assess model performance by evaluating their detection accuracy on real scans of densely packed bugs, using manually annotated center points. These annotations are matched with predicted bounding boxe centers based on Euclidean distance, optimized using Munkres' algorithm [38]. Matches are only accepted if the annotated point falls within a $1.5 \\times$ inflated bounding box, as illustrated in Fig. 4. This inflation ensures a fair assessment for detections like in Fig. 4(b). This, combined with matching on the bounding box centers, helps ensure that large boxes are not favored since it places more weight on the location of the boxes. We compute precision to avoid favoring overdetections, recall to measure the proportion of correct matches, and F1-score to combine the two. Further, we first compute the metrics purely from the point matches without considering classes — thereby evaluating if the models can detect the presence of bugs vs. background material. Next, we recompute the matches where we only allow matches between points of equal class.", + "bbox": [ + 212, + 436, + 787, + 648 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Results", + "text_level": 1, + "bbox": [ + 215, + 672, + 326, + 686 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The results are summarized in Tab. 1, example results are shown in Fig. 5, and the distributions of the metrics without class information are shown in Fig. 6. U-Net performs relatively well, independent of what it was trained on, whereas Faster R-CNN performs worse on crowded mixtures and nnDetection on single bugs. If class labels are required, the performance is generally low.", + "bbox": [ + 212, + 704, + 787, + 779 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The distributions in Fig. 6 show that U-Net has a narrower spread around the mean for the F1-score, whereas the two others are spread more out. Wrt. the precision, many detections obtained by the nnDetection method are 1, but have a lower recall, meaning that those that it predicts are correct, but there", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "P.M. Jensen et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/54b7172581217d9d6bb33f28eb39c37bfe6bc886dea38ea23106e9fc679553c7.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 279, + 147, + 424, + 238 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/842f6f2ecd4543c290ad5358ab452b07011e497f1c29141f66909bfe570d35d3.jpg", + "image_caption": [ + "(b)", + "Fig. 4: Matching criterion for bounding box to center point. The circle shows the manually annotated center point, the predicted bounding box is shown with the full line, and the inflated bounding box is shown with the dashed line. In (a), the match is accepted by the predicted bounding box, in (b), the match is also accepted due to the inflated bounding box, whereas in (c), there is no match. Note that this illustration is 2D, but the detection is done in 3D." + ], + "image_footnote": [], + "bbox": [ + 426, + 147, + 573, + 238 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8d3a276c11a1f9dcfc671c18397288ad8af552f9e342ffb804e5cdb1ee37775a.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 576, + 147, + 722, + 238 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "are many potential detections that it misses. It is the opposite for the Faster R-CNN, where many datasets get a recall of 1, but a lower precision, meaning that it detects many of the bugs, but also has many false detections.", + "bbox": [ + 212, + 382, + 784, + 426 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To investigate the capabilities of the three methods without the domain shift between the synthetic and real mixtures, we included the results in Tab. 2. This table shows the results of the validation data for training the models on the volumes of individual bugs and the two synthetic mixtures with the class labels. Since we have the bounding boxes for these datasets, we compute the mean average precision (mAP), similar to, e.g., COCO detection [34]. We compute mAP with the IoU threshold running from 0.5 to 0.95 (shown as mAP) and fixed at 0.5 (mAP@50). We note that the values are not directly comparable with those in Tab. 1. We show example detections in the supplementary. The U-Net generally has high performance, Faster R-CNN has high performance for single bugs but has difficulty with the mixtures, whereas nnDetection performs better in the mixtures than for the individual bugs. Generally, the detection performance of bugs is far from perfect and can be improved, also when training and testing within the same domain.", + "bbox": [ + 212, + 429, + 787, + 640 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6 Discussion", + "text_level": 1, + "bbox": [ + 215, + 667, + 354, + 683 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The domain shift introduced by shifting the context surrounding the bugs significantly increases the difficulty of the analysis problem. While U-Net and nnDetection achieve high performance on their training tasks (at least for synthetic mixtures), they struggle when transferred to real densely packed scans. It is also interesting that the domain shift has such an effect on the ability of the models to infer the correct class even when accurately localizing the bug. This is surprising since detecting the bugs in the training domain can be done with high accuracy, as shown in Tab. 2, and pure classification (see supplementary) gave an accuracy of 0.97. Despite the models for classification being different than the detection", + "bbox": [ + 212, + 703, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "BugNIST - Object Detection under Domain Shift", + "bbox": [ + 397, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/53bd22c734e829ea8853fbd694369dc9e61fd9a5132ecb1ece9be7314efa7f38.jpg", + "table_caption": [ + "Table 1: Detection results on mixed scans. We show mean and std. for each metric computed over the evaluation images. Higher is better. Without class info refers to not considering the bug label and with class info expects the bug type to be classified as well as being detected." + ], + "table_footnote": [], + "table_body": "
Model- Trained onWithout class info.With class info.
F1-ScorePrecisionRecallF1-ScorePrecisionRecall
U-Net [8]- Single bugs0.59±0.180.68±0.170.60±0.250.11±0.160.13±0.190.11±0.16
- Synth. mixes0.60±0.120.52±0.170.82±0.190.11±0.080.09±0.080.14±0.10
- Crowded s.m.0.54±0.130.39±0.140.97±0.070.10±0.080.07±0.060.18±0.14
Faster R-CNN [44]- Single bugs0.56±0.201.00±0.040.42±0.210.11±0.100.21±0.190.08±0.08
- Synth. mixes0.68±0.160.63±0.210.83±0.220.16±0.140.15±0.140.19±0.18
- Crowded s.m.0.27±0.170.81±0.240.18±0.160.06±0.090.19±0.300.04±0.06
nn-Detection [4]- Single bugs0.03±0.020.02±0.010.91±0.150.00±0.010.00±0.000.09±0.16
- Synth. mixes0.50±0.140.41±0.170.80±0.220.10±0.060.08±0.060.16±0.11
- Crowded s.m.0.60±0.190.77±0.230.58±0.270.11±0.120.15±0.160.11±0.12
", + "bbox": [ + 215, + 212, + 787, + 364 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/0f82912a385e7a3a23be4a4fd19898251be9ae927ba9eaf332eb27301b12f894.jpg", + "table_caption": [ + "Table 2: Detection results on the training tasks obtained during training on the single bugs and the two synthetic mixtures. For U-Net and Faster R-CNN performance is evaluated on a validation split. Since nnDetection is an ensemble created from multiple train-validation splits, we evaluate on all images in the training task. We show mean average precision (mAP) between predicted and ground truth bounding boxes. Higher is better." + ], + "table_footnote": [], + "table_body": "
Model- Trained onmAPmAP@50
U-Net [8]- Single bugs0.720.72
- Synthetic mixes0.640.77
- Crowded synth. mixes0.750.86
Faster R-CNN [44]- Single bugs0.700.70
- Synthetic mixes0.020.02
- Crowded synth. mixes0.050.05
nnDetection [4]- Single bugs0.100.10
- Synthetic mixes0.590.70
- Crowded synth. mixes0.490.66
", + "bbox": [ + 215, + 477, + 789, + 609 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "models, these results show that even when the objects do not change appearance, the shift in the domain by changing the context surrounding the objects is sufficient for deep learning-based detection models to lose performance.", + "bbox": [ + 212, + 642, + 785, + 686 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In preparing the data, we have strived to get the scans of individual bugs and bugs in mixtures to have the same appearance. Since bugs are flexible, there will be differences in the pose of the bugs in mixtures and individual bugs. Larvae will sometimes bend differently and bugs with antennae or legs will bend differently. Also, the cotton is visible and can slightly influence the appearance. Despite these differences, the voxel intensity distributions in individual scanned bugs and mixtures are the same, and therefore the BugNIST dataset is a good basis for developing detection methods that can handle context shift. If the mixtures were (partly) labeled for segmentation, BugNIST could also be used to investigate segmentation algorithms under context shift.", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "P.M. Jensen et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/eb1ac95eecc14df2f0bc40673c32812a94a2999f3c09b7d423ad367754d0275d.jpg", + "image_caption": [ + "Fig. 5: Detection results on real mixed scans. The images are maximum-intensity projections of the volumes. Center points are marked with circles, detections are marked with bounding boxes, and color indicates the bug class. A line from a point to the upper left corner of a bounding box shows a match. If the center point and the bounding box have the same color, then the detection and classification are correct. The shown images are the volumes with the highest scores for the different models. For U-Net the results when trained on synthetic mixes, for Faster R-CNN when trained on synthetic mixes and nnDetection when trained on crowded synthetic mixes." + ], + "image_footnote": [], + "bbox": [ + 220, + 142, + 781, + 454 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/01c91c05303a061e584404fa4aebd0b5e77a0f9d8179303ed60136159284688b.jpg", + "image_caption": [ + "Fig. 6: Distributions of performance scores for the three detection methods." + ], + "image_footnote": [], + "bbox": [ + 228, + 595, + 776, + 659 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We chose to annotate only the center positions of the bugs. As the bugs are often elongated and at odd angles, axis-aligned bounding boxes would contain significant empty space and overlap between neighboring boxes—thus offering limited benefit over points. This could be solved with object aligned bounding boxes or full voxel annotations. However, this would drastically increase the annotation time. We chose to prioritize having a large number of annotations such that the test set can have a large variety of configurations. Furthermore, given that current methods show ample room for improvement (cf. Tab. 1), we", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "BugNIST - Object Detection under Domain Shift", + "bbox": [ + 398, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "find that point annotations combined with the chosen metric are sufficient to measure method performance and improvements of future approaches.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The methods and experiments are chosen as a basic baseline to illustrate the properties of the BugNIST dataset. From the experiments, we learned that data augmentation by composing synthetic mixtures is a feasible approach for improving detection performance. More investigations into packing algorithms and ways to combine the synthetic mixed bugs with other backgrounds can potentially improve performance further. But it is also clear that there is great potential for developing new methods for object detection in volumes. We chose Detection U-Net, nnDetection, and our adaption of Faster R-CNN to 3D volumes because these methods were available, widely used, and are based on complementary ideas. We did not find more recent alternatives for volumetric analysis.", + "bbox": [ + 212, + 176, + 785, + 327 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our results show that Faster R-CNN performs less than the two other methods. While it scores highest in Tab. 1, it is clear from Fig. 5 that the bounding boxes are of much lower quality. Furthermore, it does not score well on the training tasks, except when trained on individual bugs. To ensure the poor performance of Faster R-CNN was not just the result of too few training volumes, we redid the experiments with 1000 synthetic mixes instead of the original 300. Performance remained the same. It is also noteworthy that nnDetection scores overall lower than U-Net on both the training and evaluation tasks, even though it also has access to the segmentation masks for training.", + "bbox": [ + 212, + 328, + 787, + 464 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 484, + 359, + 502 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "With our BugNIST benchmark dataset for domain shift in volumetric data, we demonstrate how detecting objects, that have the same appearance in the source and target domain, fails when the objects in the two domains are placed in different contexts. By contexts, we refer to the appearance of the image surrounding the objects. To our knowledge, the BugNIST dataset is the only dataset designed for developing methods for context shift, i.e. domain shift where the objects stay the same but the surroundings change. Solutions will enable an automated annotation of data for volumetric object detection, which is an area that has not received much attention. There are, however, a range of applications where object detection is needed, including medical imaging and emerging areas such as micro-CT scanning of materials. The BugNIST dataset can also be used for investigating other problems such as segmentation under domain shift provided reference annotations are created.", + "bbox": [ + 212, + 515, + 787, + 710 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 215, + 734, + 393, + 752 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We acknowledge the 3D Imaging Center at DTU, the Infrastructure for Quantitative AI-based Tomography (QUAITOM), supported by the Novo Nordisk Foundation (grant number NNF21OC0069766) and STUDIOS: Segmenting Tomograms Using Different Interpretation of Simplicity funded by the Villum Foundation (grant number VIL50425).", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "P.M. Jensen et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Armato, S.G., Huisman, H., Drukker, K., Hadjiiski, L., Kirby, J.S., Petrick, N., Redmond, G., Giger, M.L., Cha, K., Mamonov, A., et al.: PROSTATEx challenges for computerized classification of prostate lesions from multiparametric magnetic resonance images. Journal of Medical Imaging 5(4), 044501-044501 (2018)", + "2. Armato, S.G., McLennan, G., Bidaut, L., McNitt-Gray, M.F., Meyer, C.R., Reeves, A.P., Zhao, B., Aberle, D.R., Henschke, C.I., Hoffman, E.A., et al.: The lung image database consortium (LIDC) and image database resource initiative (IDRI): a completed reference database of lung nodules on CT scans. Medical Physics 38(2), 915–931 (2011)", + "3. Baid, U., Ghodasara, S., Mohan, S., Bilello, M., Calabrese, E., Colak, E., Farahani, K., Kalpathy-Cramer, J., Kitamura, F.C., Pati, S., et al.: The RSNA-ASNR-MICCAI BraTS 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314 (2021)", + "4. Baumgartner, M., Jäger, P.F., Isensee, F., Maier-Hein, K.H.: nnDetection: a self-configuring method for medical object detection. In: Proceedings of the Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 530-539. Springer (2021)", + "5. Cardoso, M.J., Li, W., Brown, R., Ma, N., Kerfoot, E., Wang, Y., Murrey, B., Myronenko, A., Zhao, C., Yang, D., et al.: MONAI: An open-source framework for deep learning in healthcare. arXiv preprint arXiv:2211.02701 (2022)", + "6. Charytanowicz, M., Kulczycki, P., Kowalski, P.A., Lukasik, S., Czabak-Garbacz, R.: An evaluation of utilizing geometric features for wheat grain classification using X-ray images. Computers and Electronics in Agriculture 144, 260-268 (2018)", + "7. Chaudhary, S., Sadbhawna, S., Jakhetiya, V., Subudhi, B.N., Baid, U., Guntuku, S.C.: Detecting COVID-19 and community acquired pneumonia using chest CT scan images with deep learning. In: Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 8583-8587. IEEE (2021)", + "8. Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3D U-Net: learning dense volumetric segmentation from sparse annotation. In: Proceedings of the Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 424–432. Springer (2016)", + "9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3213-3223 (2016)", + "0. De Carlo, F., Gürsoy, D., Ching, D.J., Batenburg, K.J., Ludwig, W., Mancini, L., Marone, F., Mokso, R., Pelt, D.M., Sijbers, J., et al.: Tomobank: a tomographic data repository for computational x-ray science. Measurement Science and Technology 29(3), 034004 (2018)", + "1. Dorent, R., Kujawa, A., Ivory, M., Bakas, S., Rieke, N., Joutard, S., Glocker, B., Cardoso, J., Modat, M., Batmanghelich, K., et al.: Crossmoda 2021 challenge: Benchmark of cross-modality domain adaptation techniques for vestibular schwannoma and cochlea segmentation. Medical Image Analysis 83, 102628 (2023)", + "2. Du, Z., Hu, Y., Ali Buttar, N., Mahmood, A.: X-ray computed tomography for quality inspection of agricultural products: A review. Food Science & Nutrition 7(10), 3146-3160 (2019)" + ], + "bbox": [ + 225, + 179, + 785, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "BugNIST - Object Detection under Domain Shift", + "bbox": [ + 397, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Einarsdóttir, H., Emerson, M.J., Clemmensen, L.H., Scherer, K., Willer, K., Bech, M., Larsen, R., Ersbøll, B.K., Pfeiffer, F.: Novelty detection of foreign objects in food using multi-modal X-ray imaging. Food Control 67, 39-47 (2016)", + "14. European Organization For Nuclear Research, OpenAIRE: Zenodo (2013). https://doi.org/10.25495/7GXK-RD71, https://www.zenodo.org/", + "15. Flitton, G., Mouton, A., Breckon, T.P.: Object classification in 3D baggage security computed tomography imagery using visual codebooks. Pattern Recognition 48(8), 2489-2499 (2015)", + "16. Girshick, R.: Fast R-CNN. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 1440-1448 (2015)", + "17. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object detection and semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 580-587 (2014)", + "18. Guan, H., Liu, M.: Domain adaptation for medical image analysis: a survey. IEEE Transactions on Biomedical Engineering 69(3), 1173-1185 (2021)", + "19. Harmon, S.A., Sanford, T.H., Xu, S., Turkbey, E.B., Roth, H., Xu, Z., Yang, D., Myronenko, A., Anderson, V., Amalou, A., et al.: Artificial intelligence for the detection of COVID-19 pneumonia on chest CT using multinational datasets. Nature Communications 11(1), 4080 (2020)", + "20. Hipsley, C.A., Aguilar, R., Black, J.R., Hocknull, S.A.: High-throughput microCT scanning of small specimens: preparation, packing, parameters and post-processing. Scientific Reports 10(1), 13863 (2020)", + "21. Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods 18(2), 203-211 (2021)", + "22. Jaeger, P.F., Kohl, S.A., Bickelhaupt, S., Isensee, F., Kuder, T.A., Schlemmer, H.P., Maier-Hein, K.H.: Retina U-Net: Embarrassingly simple exploitation of segmentation supervision for medical object detection. In: Machine Learning for Health Workshop. pp. 171-183. PMLR (2020)", + "23. Jain, Y., Borner, K., Walsh, C., Ruschman, N., Lee, P.D., Weber, G.M., Holbrook, R., Howard, A.: SenNet + HOA - hacking the human vasculature in 3D (2023), https://kaggle.com/competitions/blood-vessel-segmentation", + "24. Jarolmasjed, S., Espinoza, C.Z., Sankaran, S., Khot, L.R.: Postharvest bitter pit detection and progression evaluation in 'Honeycrisp' apples using computed tomography images. Postharvest Biology and Technology 118, 35-42 (2016)", + "25. Jeppesen, N., Christensen, A.N., Dahl, V.A., Dahl, A.B.: Sparse layered graphs for multi-object segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12777-12785 (2020)", + "26. Jin, L., Yang, J., Kuang, K., Ni, B., Gao, Y., Sun, Y., Gao, P., Ma, W., Tan, M., Kang, H., et al.: Deep-learning-assisted detection and segmentation of rib fractures from CT scans: Development and validation of FracNet. EBioMedicine 62, 103106 (2020)", + "27. Ker, J., Singh, S.P., Bai, Y., Rao, J., Lim, T., Wang, L.: Image thresholding improves 3-dimensional convolutional neural network diagnosis of different acute brain hemorrhages on computed tomography scans. Sensors 19(9), 2167 (2019)", + "28. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (2023)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "P.M. Jensen et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998)", + "30. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.M.: Deeper, broader and artier domain generalization. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 5542-5550 (2017)", + "31. Li, Y., Xie, S., Chen, X., Dollar, P., He, K., Girshick, R.: Benchmarking detection transfer learning with vision transformers. arXiv preprint arXiv:2111.11429 (2021)", + "32. Li, Y., Fan, Y.: Medical image segmentation with domain adaptation: A survey. arXiv preprint arXiv:2311.01702 (2023)", + "33. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 2980-2988 (2017)", + "34. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 740-755. Springer (2014)", + "35. Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature Communications 15(1), 654 (2024)", + "36. Maier-Hein, L., Eisenmann, M., Reinke, A., Onogur, S., Stankovic, M., Scholz, P., Arbel, T., Bogunovic, H., Bradley, A.P., Carass, A., et al.: Why rankings of biomedical image analysis competitions should be interpreted with care. Nature Communications 9(1), 5217 (2018)", + "37. Muandet, K., Balduzzi, D., Scholkopf, B.: Domain generalization via invariant feature representation. In: International Conference on Machine Learning. pp. 10-18. PMLR (2013)", + "38. Munkres, J.: Algorithms for the assignment and transportation problems. Journal of the Society for Industrial and Applied Mathematics 5(1), 32-38 (1957)", + "39. OECD: Health at a Glance 2023 (2023). https://doi.org//10.1787/7a7afb35-en", + "40. Pan, S.J., Tsang, I.W., Kwok, J.T., Yang, Q.: Domain adaptation via transfer component analysis. IEEE Transactions on Neural Networks 22(2), 199-210 (2010)", + "41. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1406-1415 (2019)", + "42. Rawson, S.D., Maksimcuka, J., Withers, P.J., Cartmell, S.H.: X-ray computed tomography in life sciences. BMC biology 18(1), 1-15 (2020)", + "43. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 779-788 (2016)", + "44. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. Advances in neural information processing systems (NeurIPS) 28 (2015)", + "45. Richter, S.R., Vineet, V., Roth, S., Koltun, V.: Playing for data: Ground truth from computer games. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 102-118. Springer (2016)", + "46. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: Proceedings of the Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 234-241. Springer (2015)", + "47. Ros, G., Sellart, L., Materzynska, J., Vazquez, D., Lopez, A.M.: The synthia dataset: A large collection of synthetic images for semantic segmentation of urban" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "BugNIST - Object Detection under Domain Shift", + "bbox": [ + 398, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3234-3243 (2016)", + "48. Sankaranarayanan, S., Balaji, Y., Jain, A., Lim, S.N., Chellappa, R.: Learning from synthetic data: Addressing domain shift for semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3752-3761 (2018)", + "49. Serte, S., Demirel, H.: Deep learning for diagnosis of COVID-19 using 3D CT scans. Computers in biology and medicine 132, 104306 (2021)", + "50. Setio, A.A.A., Traverso, A., De Bel, T., Berens, M.S., Van Den Bogaard, C., Cerello, P., Chen, H., Dou, Q., Fantacci, M.E., Geurts, B., et al.: Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: the luna16 challenge. Medical Image Analysis 42, 1-13 (2017)", + "51. Simpson, A.L., Antonelli, M., Bakas, S., Bilello, M., Farahani, K., van Ginneken, B., Kopp-Schneider, A., Landman, B.A., Litjens, G., Menze, B.H., Ronneberger, O., Summers, R.M., Bilic, P., Christ, P.F., Do, R.K.G., Gollub, M., Golia-Pernicka, J., Heckers, S., Jarnagin, W.R., McHugo, M., Napel, S., Vorontsov, E., Maier-Hein, L., Cardoso, M.J.: A large annotated medical image dataset for the development and evaluation of segmentation algorithms. CoRR abs/1902.09063 (2019), http://arxiv.org/abs/1902.09063", + "52. Sun, T., Segu, M., Postels, J., Wang, Y., Van Gool, L., Schiele, B., Tombari, F., Yu, F.: Shift: a synthetic driving dataset for continuous multi-task domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21371-21382 (2022)", + "53. Timmins, K.M., van der Schaaf, I.C., Bennink, E., Ruigrok, Y.M., An, X., Baumgartner, M., Bourdon, P., De Feo, R., Di Noto, T., Dubost, F., et al.: Comparing methods of detecting and segmenting unruptured intracranial aneurysms on TOF-MRAS: The ADAM challenge. Neuroimage 238, 118216 (2021)", + "54. TorchVision maintainers and contributors: Torchvision: Pytorch's computer vision library. https://github.com/pytorch/vision (2016)", + "55. Trusler, M.M., Sturrock, C.J., Vane, C.H., Cook, S., Lomax, B.H.: X-ray computed tomography: A novel non-invasive approach for the detection of microplastics in sediments? Marine Pollution Bulletin 194, 115350 (2023)", + "56. Velayudhan, D., Hassan, T., Damiani, E., Werghi, N.: Recent advances in baggage threat detection: A comprehensive and systematic survey. ACM Computing Surveys 55(8) (dec 2022). https://doi.org/10.1145/3549932, https://doi.org/10.1145/3549932", + "57. Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5018-5027 (2017)", + "58. Walsh, C., Tafforeau, P., Wagner, W., Jafree, D., Bellier, A., Werlein, C., Kühnel, M., Boller, E., Walker-Samuel, S., Robertus, J., et al.: Imaging intact human organs with local resolution of cellular structures using hierarchical phase-contrast tomography. Nature Methods 18(12), 1532–1541 (2021)", + "59. Wang, M., Deng, W.: Deep visual domain adaptation: A survey. Neurocomputing 312, 135-153 (2018)", + "60. Wang, M., Liu, Y., Yuan, J., Wang, S., Wang, Z., Wang, W.: Inter-class and inter-domain semantic augmentation for domain generalization. IEEE Transactions on Image Processing (2024)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "P.M. Jensen et al.", + "bbox": [ + 271, + 114, + 392, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "61. Wang, Q., Bhowmik, N., Breckon, T.P.: Multi-class 3D object detection within volumetric 3D computed tomography baggage security screening imagery. In: Proceedings of the IEEE International Conference on Machine Learning and Applications (ICMLA). pp. 13-18. IEEE (2020)", + "62. Wang, Q., Breckon, T.P.: Contraband materials detection within volumetric 3D computed tomography baggage security screening imagery. In: Proceedings of the IEEE International Conference on Machine Learning and Applications (ICMLA). pp. 75-82. IEEE (2021)", + "63. Withers, P.J., Bouman, C., Carmignato, S., Cnudde, V., Grimaldi, D., Hagen, C.K., Maire, E., Manley, M., Du Plessis, A., Stock, S.R.: X-ray computed tomography. Nature Reviews Methods Primers 1(1), 18 (2021)", + "64. Wu, D., Wu, D., Feng, H., Duan, L., Dai, G., Liu, X., Wang, K., Yang, P., Chen, G., Gay, A.P., et al.: A deep learning-integrated micro-ct image analysis pipeline for quantifying rice lodging resistance-related traits. Plant Communications 2(2) (2021)", + "65. Xiao, Z., Liu, B., Geng, L., Zhang, F., Liu, Y.: Segmentation of lung nodules using improved 3d-unet neural network. Symmetry 12(11), 1787 (2020)", + "66. Xie, Y., Ji, Q.: A new efficient ellipse detection method. In: Proceedings of the International Conference on Pattern Recognition (ICPR). vol. 2, pp. 957-960. IEEE (2002)", + "67. Xu, F., Tenenbaum, J.B.: Word learning as bayesian inference. Psychological review 114(2), 245 (2007)", + "68. Yan, Y.T., Chua, S., DeCarlo, T.M., Kempf, P., Morgan, K.M., Switzer, A.D.: Core-ct: A matlab application for the quantitative analysis of sediment and coral cores from X-ray computed tomography (ct). Computers & Geosciences 156, 104871 (2021)", + "69. Zhao, W.X., Zhou, K., Li, J., Tang, T., Wang, X., Hou, Y., Min, Y., Zhang, B., Zhang, J., Dong, Z., et al.: A survey of large language models. arXiv preprint arXiv:2303.18223 (2023)", + "70. Zhou, K., Yang, Y., Hospedales, T., Xiang, T.: Deep domain-adversarial image generation for domain generalisation. Proceedings of the AAAI Conference on Artificial Intelligence 34(07), 13025-13032 (April 2020). https://doi.org/10.1609/aaii.v34i07.7003, https://ojs.aaii.org/index.php/AAAI/article/view/7003", + "71. Zhou, S.K., Greenspan, H., Davatzikos, C., Duncan, J.S., Van Ginneken, B., Madabhushi, A., Prince, J.L., Rueckert, D., Summers, R.M.: A review of deep learning in medical imaging: Imaging traits, technology trends, case studies with progress highlights, and future promises. Proceedings of the IEEE 109(5), 820-838 (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 674 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "BugNIST - Object Detection under Domain Shift", + "bbox": [ + 398, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/BugNIST - a Large Volumetric Dataset for Detection under Domain Shift/fceefb33-6fd3-4a63-9cb1-f0d439dc4d49_model.json b/2024/BugNIST - a Large Volumetric Dataset for Detection under Domain Shift/fceefb33-6fd3-4a63-9cb1-f0d439dc4d49_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7082d99d60c117ed96b5d1a01d0038cb07768979 --- /dev/null +++ b/2024/BugNIST - a Large Volumetric Dataset for Detection under Domain Shift/fceefb33-6fd3-4a63-9cb1-f0d439dc4d49_model.json @@ -0,0 +1,2449 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.255, + 0.141, + 0.75, + 0.187 + ], + "angle": 0, + "content": "BugNIST - a Large Volumetric Dataset for Object Detection under Domain Shift" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.213, + 0.756, + 0.245 + ], + "angle": 0, + "content": "Patrick Møller Jensen, Vedrana Andersen Dahl, Rebecca Engberg, Carsten Gundlach, Hans Marin Kjer, and Anders Bjorholm Dahl" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.256, + 0.747, + 0.285 + ], + "angle": 0, + "content": "Technical University of Denmark, Kgs. Lyngby, Denmark {patmjen, vand, reen}@dtu.dk, cagu@fysik.dtu.dk, {hmjk, abda}@dtu.dk" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.324, + 0.741, + 0.532 + ], + "angle": 0, + "content": "Abstract. Domain shift significantly influences the performance of deep learning algorithms, particularly for object detection within volumetric 3D images. Annotated training data is essential for deep learning-based object detection. However, annotating densely packed objects is time-consuming and costly. Instead, we suggest training models on individually scanned objects, causing a domain shift between training and detection data. To address this challenge, we introduce the BugNIST dataset, comprising 9154 micro-CT volumes of 12 bug types and 388 volumes of tightly packed bug mixtures. This dataset is characterized by having objects with the same appearance in the source and target domains, which is uncommon for other benchmark datasets for domain shift. During training, individual bug volumes labeled by class are utilized, while testing employs mixtures with center point annotations and bug type labels. Together with the dataset, we provide a baseline detection analysis, with the aim of advancing the field of 3D object detection methods." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.546, + 0.741, + 0.575 + ], + "angle": 0, + "content": "Keywords: Volumetric Dataset, Benchmark, Volumetric Object Detection, Domain Shift." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.626, + 0.377, + 0.642 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Our work on domain shift in volumetric 3D images is motivated by the need for labeled data to train supervised deep learning models for volumetric imaging. The problem that we aim to solve is object detection and classification. We propose to label images of objects scanned as isolated entities as the source domain and use these as a basis to train models for object detection and classification in a complex context of mixed objects and other materials as the target domain. The effort needed for obtaining labels in the two domains is significantly different. If objects are isolated, they can be automatically labeled whereas mixed objects require expensive manual labeling. Automating the labeling based on isolated objects, however, leads to a domain shift between the data in the source domain for training and the data in the target domain for detection and classification. This domain shift is special because the appearance of the objects is the same" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.127 + ], + "angle": 0, + "content": "P.M. Jensen et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.144, + 0.477, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.283, + 0.458, + 0.305 + ], + "angle": 0, + "content": "(a) 9154 volumes of individual bugs with class labels." + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.145, + 0.791, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.282, + 0.766, + 0.305 + ], + "angle": 0, + "content": "(b) 388 volumes of bug mixtures with center point and class labels." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.332, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Fig. 1: Overview of the two parts of the dataset. In (a), representative individual bug volumes are displayed with a class abbreviation, name, and number of volumes. In (b), a few volumes of bug mixtures are shown. Each bug is annotated with bug class and center point." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.785, + 0.449 + ], + "angle": 0, + "content": "in the two domains, but the surrounding context is different. We use the term context shift for this type of domain shift." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.451, + 0.788, + 0.587 + ], + "angle": 0, + "content": "To promote the development of new deep learning-based methods that can handle context shift, we contribute the volumetric dataset called BugNIST1. The dataset contains 9154 micro-CT volumes with individual bugs and 388 micro-CT volumes containing a mixture of several bugs as well as other materials such as leaves, wood shavings, etc.; see Fig. 1. The challenge is to use the individually scanned bugs to detect and classify the bugs in the mixtures. The dataset is intended to be a relatively simple test case for starting to develop methods for detecting objects in volumetric images under context shift, and bugs are chosen due to their complex shape and not for solving a problem in entomology." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.588, + 0.788, + 0.722 + ], + "angle": 0, + "content": "As a basis for training an accurate deep learning model, a training dataset for object detection and classification must capture the appearance distribution of the object classes that are in the dataset. Since the appearance of a CT-scanned object changes minimally if it is scanned individually or as part of a mixture, the appearance distribution can be covered by only considering the individual scans. This is different from 2D photos. Here additional variations in scale, viewpoint, and illumination influence the appearance of objects, and it would be more difficult to acquire representative images of individual objects for a dataset based on 2D photos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.724, + 0.788, + 0.814 + ], + "angle": 0, + "content": "Since BugNIST is a test case for context shift, we wanted to ensure that each class is covered well. By scanning systematically arranged bugs, we have generated 9154 volumes of individual bugs, each labeled with its corresponding class. Furthermore, we scanned 388 mixed volumes, each containing between 4 and 45 bugs alongside other materials. For evaluation purposes, each bug in the mixtures is manually annotated with its center position and class label. In total," + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.824, + 0.492, + 0.841 + ], + "angle": 0, + "content": "1 https://abdahl.github.io/bugnist/" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.399, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BugNIST - Object Detection under Domain Shift" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "9542 volumes make up the BugNIST dataset. The numbers in each class are shown in Fig. 1. Since the bugs in each class are the same age, we expect to cover the biological variation sufficiently to be representative of the class." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.195, + 0.785, + 0.407 + ], + "angle": 0, + "content": "Solutions to the type of context shift, that we present here, namely a domain shift where only the context changes, and not the appearance of the objects, have wide applications in volumetric imaging. The field of volumetric imaging is dominated by medical applications [39], but there is a growing use beyond the medical area based on e.g. micro-CT [58, 63]. In all these application areas, there is a need to employ supervised deep learning to solve problems that require quantification of depicted structures. When the use case is novel and there is no directly applicable training data available — a typical scenario in volumetric imaging of e.g. materials microstructure, but also in specialized biomedical applications — it is necessary to create training data. If labor-intensive manual labeling can be limited or avoided, there is potentially a huge gain in productivity. Examples of applications are detecting and classifying objects in scans of marine sediments such as microplastics [55, 68], museum specimens [20], or detecting threat items in baggage security screening [56]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.787, + 0.514 + ], + "angle": 0, + "content": "In volumetric scans, manual labeling is significantly more demanding than in 2D. For instance, annotating a bounding box of a single object in 3D requires traversing slices of the volume and making a mental model of the object. If the volume contains densely packed complex objects, determining their bounding boxes becomes difficult. Further, bounding boxes may overlap to such an extent that they stop being a useful way to locate objects. This is why we employ center points as the annotation format in the mixtures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.787, + 0.669 + ], + "angle": 0, + "content": "The name BugNIST is a homage to the influential MNIST dataset [29] which has played a pivotal role in the development and evaluation of deep learning methods. Similar to the way that MNIST has been used, BugNIST is intended to be a benchmark dataset that represents a well-defined problem, where the development of solutions will make a basis for solutions that can advance the field. Specifically for BugNIST, solutions will enable object detection and classification with significantly less effort put into creating the data foundation for training. This calls for methods that can learn to predict objects in a complex context from volumes of individual objects, namely solving the context shift problem. To our knowledge, no existing datasets have this property." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.672, + 0.787, + 0.777 + ], + "angle": 0, + "content": "The term context shift refers to a shift in the data distribution between training and inference. Typical shifts in data distributions arise from changes in data acquisition systems, e.g. the use of different scanners, or from variations in the objects of interest between datasets. Such problems have been addressed using domain adaptation [18,59]. In our setting, there is almost no change in the appearance of the bugs between volumes of individual bugs and mixtures, but the change is in what surrounds the bugs, i.e. the change is in the context." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We hypothesize that methods capable of learning under context shift might also prove capable of learning from very few examples, similar to how humans learn [67]. Instead of relying on seeing objects in every imaginable context, such methods could instead isolate objects from their contexts. If this hypothesis holds" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "P.M. Jensen et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "true, learning under context shift may reduce the need for extensive training data and the associated training costs. This is in contrast to the current trend, where larger models and datasets as seen in LLMs [69] and computer vision [28, 35], lead to better performance, but also limit who can effectively train such models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.219, + 0.785, + 0.293 + ], + "angle": 0, + "content": "Contributions We contribute BugNIST, a new volumetric image dataset for benchmarking domain shift. In this paper, we use BugNIST for benchmarking detection methods under context shift, trained only on individual bugs and tested on mixtures. We provide several baselines and investigate different training strategies for existing volumetric detection methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.294, + 0.787, + 0.37 + ], + "angle": 0, + "content": "Our results show that volumetric object detection under context shift is difficult. We obtain reasonable results in localizing the bugs, but when a class label is assigned, there is a significant drop in performance. This is even though the classification of individual bugs is easy. Our experiments demonstrate great potential for developing new approaches for 3D object detection of complex objects." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.39, + 0.383, + 0.405 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.784, + 0.461 + ], + "angle": 0, + "content": "In computer vision, the term 3D data refers to different data types, including point clouds, surfaces, volumes, and depth images. Here, we are only concerned with 3D volumes, i.e. image data represented on a voxel grid." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.787, + 0.629 + ], + "angle": 0, + "content": "BugNIST is created to advance deep learning methods for object detection and classification under domain shift. Especially in medical imaging, analysis of 3D volumes is in great demand. A review of medical challenges illustrates this [36]. Their study shows that detection makes up \\(7\\%\\) of the tasks, classification makes up \\(10\\%\\) of the tasks, whereas segmentation accounts for \\(70\\%\\) of the tasks. Detection is important for example for diagnosis [7, 19, 27, 49, 71]. Outside the medical domain, object detection problems include baggage security screening [15, 61, 62], although the problem and the available data are still primarily in the 2D domain [56], food inspection [12] such as fruit quality [24], grain classification [6], foreign object detection [13], biological applications such as analysis of weeds [64] and other uses driven by developments in X-ray micro-CT imaging [42], etc." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.787, + 0.749 + ], + "angle": 0, + "content": "The domain shift problem presented in BugNIST, where only the context changes, and not the appearance of the objects of interest, is not widely studied. Most research on domain shifts focuses primarily on problems where the appearance of objects changes. Solutions focus on domain adaptation, where labeled data in the source domain are available together with data without labels from the target domain [40], or domain generalization, where the inference domain is not known at training time [37, 60]. We focus on domain adaptation, which is also the typical use case for volumetric imaging." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Several widely used datasets are available for studying and developing methods for domain shift. For 2D image classification problems the datasets PACS [30], Digits-DG [70], Office-Home [57], and DomainNet [41] developed to classify images from domains such as photos, art paintings, sketches, etc. Domain shift between artificially generated and real images e.g. for segmentation is another area that receives much attention [48] with standard datasets such as" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.399, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BugNIST - Object Detection under Domain Shift" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.178 + ], + "angle": 0, + "content": "SYNTHEIA [47], CITYSCAPES [9], and GTA-5 [45], or completely synthetic domain shift like the SHIFT dataset [52] for autonomous driving." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.18, + 0.787, + 0.286 + ], + "angle": 0, + "content": "In 3D, domain shift is often studied in medical applications where the same type of object is imaged with different scanning modalities [32]. It can be MRI (magnetic resonance imaging), CT, OCT (optical coherence tomography), etc. Standard datasets include the CrossMoDA 2021 dataset [11] and the BRATS dataset [3], both of different MRI modalities. There are many other datasets used for domain adaptation for volumetric medical imaging [32], but all for adapting between modalities or between the same modality but different scanners [18]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.288, + 0.787, + 0.363 + ], + "angle": 0, + "content": "The structure of the imaged objects in BugNIST is different from most existing volumetric deep learning benchmark datasets. The morphological complexity of bugs is, however, representative of volumetric imaging data of e.g. 3D microstructure, where there is much less use of deep learning than in e.g. the medical domain [63]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.365, + 0.787, + 0.518 + ], + "angle": 0, + "content": "Platforms like Zenodo [14], TomoBank [10], or the Human Organ Atlas [58] also contain a large amount of volumetric image datasets. Different from e.g. medical challenges, this data generally comes with no clear problem formulation, well-defined tasks or even labels, that allow for a comparison of methods. Further, their large volume sizes are difficult to access. It can also be difficult to judge the complexity of the data and if deep learning is needed. Some images can easily be analyzed using a threshold whereas others are complex and can greatly benefit from deep learning. Therefore, most of this data is not attractive to use as a benchmark for developing new deep learning methods. Some datasets are emerging, e.g. the vessel segmentation challenge [23], but they remain exceptions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.787, + 0.594 + ], + "angle": 0, + "content": "There is extensive research in deep learning-based object detection, but primarily in 2D. Most methods either use a two-stage detector such as R-CNN [16, 17, 44], with a region proposal and a classifier to choose among the proposed regions, or a one-stage detector such asYOLO [43], where regions and classes are proposed in a single network." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.787, + 0.809 + ], + "angle": 0, + "content": "Object detection in volumetric images is less common. One approach has been to apply a 3D U-Net [8] segmentation network and use each connected component of the foreground label class as a detection [22]. An example of using this approach is lung nodule detection in [65]. Employing segmentation for object detection comes with the problem of separating objects that may be connected or combining separated segments into one object. To avoid this problem, Jaeger et al. [22] propose the Retina U-Net: a one-shot 3D detection algorithm that combines the Retina Net [33] and U-Net [46] giving detections with 3D bounding boxes. Despite the good performance of Retina U-Net, changing to a new 3D detection problem requires a time-consuming method configuration. This has been addressed in nnDetection [4]. Inspired by the nnU-Net (no new U-Net) [21], they propose a no new 3D detection method. On the BugNIST data, we test detection with 3D U-Net [8], the nnDetection method [4], and a 3D adaption of the Faster R-CNN model [44]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Benchmark datasets for volumetric object detection are primarily medical. Examples include Baumgartner et al. [4], which uses 13 datasets to develop the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "P.M. Jensen et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.329 + ], + "angle": 0, + "content": "nnDetection method. Two of these datasets were originally targeted 3D detection and include the LUNA16 for nodule detection [50], which is a subset of the LIDC dataset [2] and contains 888 CT scans of lungs, and the ADAM dataset [53] for detecting aneurysms in MRI scans. ADAM is made up of 254 MRI scans. The rest of the detection datasets are medical segmentation datasets such as data from the Medical Segmentation Decathlon [51], the rib fracture detection [26], the PROSTATEx dataset [1], and others. The segmentation datasets have been transformed into detection data by identifying center positions and bounding boxes for the segments. Within baggage security screening, some 3D datasets are obtainable, for example, the ALERT datasets used in [62]. Due to the sensitive nature of the problem, these datasets have copyright restrictions and limited availability under a non-disclosure agreement." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.354, + 0.562, + 0.373 + ], + "angle": 0, + "content": "3 Data acquisition and processing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.788, + 0.508 + ], + "angle": 0, + "content": "Fig. 1 gives an overview of the BugNIST dataset. Each individual bug volume contains exactly one bug located near the center of the volume. The total of 9154 individual bug volumes have between 713 and 946 volumes of each class distributed across 12 bug classes (species and stages). Bug mixtures contain several bugs, leaves, cotton, shredded paper, and wood shaving. There are in total 388 mixtures. The mixtures contain 4 to 45 bugs, totaling 5011 bugs with an average of 13 bugs per mixture, see Fig. 2. In bug mixtures, we have manually marked the center positions of each bug and provided it with a class label." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.613 + ], + "angle": 0, + "content": "All scans were acquired using a laboratory micro-CT scanner (Nikon metrology scanner) that was set to record volumes of 2000 voxels cubed with an isotropic voxel size of \\(26.62\\mu \\mathrm{m}\\). Being able to acquire such large volumes allowed us to scan multiple bugs and bug mixtures at once, and then crop them into smaller volumes. The dataset was collected through nearly 250 scans, each taking 30 minutes, over 90 days. Care has been taken that the scanning parameters stay unchanged for all scans, and we have calibrated the scanner daily." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.704 + ], + "angle": 0, + "content": "All bugs were placed in tubes, either as individuals or in mixtures. We used cotton, which is almost transparent to X-rays, to separate the bugs, either as individual bugs or mixtures. For the mixtures, we varied the difficulty of the bug detection task by adding leaves, shredded paper, or wood shavings to some of the mixtures, and also varied the number of bugs, their class, and the packing density." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.84 + ], + "angle": 0, + "content": "The cylindrical shape of the tubes was easy to segment using a circular Hough-transform [66] and sparse layered graphs [25], leaving us with a cylinder of only bugs and cotton. Due to the low intensity of cotton compared to the bugs, we could easily divide each cylinder into volumes containing one bug each. The volumes have been manually inspected and cleaned to correct cases where something went wrong, e.g. parts of another bug showing in the periphery of the volume. The final volumes of scans of individual bugs are \\(450 \\times 450 \\times 900\\) voxels, the size which was chosen to accommodate the biggest bugs. The mixtures were segmented similarly and the resulting volumes are \\(650 \\times 650 \\times 900\\) voxels. All" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.398, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "BugNIST - Object Detection under Domain Shift" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.147, + 0.473, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.217, + 0.466, + 0.237 + ], + "angle": 0, + "content": "(a) Class distribution for bugs from mixture volumes." + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.147, + 0.634, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.149, + 0.772, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.217, + 0.762, + 0.237 + ], + "angle": 0, + "content": "(b) Distribution of the number of bugs per volume and the number of classes per volume." + }, + { + "type": "image_caption", + "bbox": [ + 0.309, + 0.252, + 0.692, + 0.266 + ], + "angle": 0, + "content": "Fig. 2: Summary of volumes from scans of bug mixtures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.296, + 0.784, + 0.341 + ], + "angle": 0, + "content": "volumes are stored as 8-bit unsigned integers with their original spatial resolution of \\(26.62\\mu \\mathrm{m}\\) (physical voxel size). Our dataset comprises 1.8 trillion voxels, approximately a quarter of the RGB values in ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.785, + 0.417 + ], + "angle": 0, + "content": "The bugs scanned for the BugNIST dataset are bred as pet food and have been frozen before being packed for scanning. Despite careful packing, some scanned bugs may appear damaged, e.g. a few crickets have broken off antennas or miss a leg. We have chosen to accept these defects because they reflect some variation in a population and add to the challenge of the dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.785, + 0.507 + ], + "angle": 0, + "content": "To create additional lightweight datasets, we created four downsized versions of all volumes. The individual bug volumes, originally 900 voxels long, have lengths of 512, 256, 128, and 64 voxels in the downscaled versions, and all the other lengths are scaled accordingly. The voxel size of the coarsest data is \\(374.3\\mu \\mathrm{m}\\), which is a scaling factor of approximately 14 compared to the original scan." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.533, + 0.376, + 0.55 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.785, + 0.715 + ], + "angle": 0, + "content": "To establish baselines for volumetric detection under context shift (domain shift with no change in object appearance), we have selected and adapted three existing detection models. Furthermore, we have applied three training strategies, yielding a total of nine baseline models. The diversity in model choices and training strategies contributes to a comprehensive set of baseline models that provide a strong foundation for object detection under context shift. We briefly describe the selected methods and how they were trained. Full details are provided in the supplementary material. To determine the difficulty of differentiating between the individual bugs, we also performed a classification experiment which is described in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.716, + 0.784, + 0.746 + ], + "angle": 0, + "content": "All experiments are carried out using the x128 data, which is a compromise between memory needs and the preservation of structural details of the bugs." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.769, + 0.691, + 0.784 + ], + "angle": 0, + "content": "4.1 Training strategies for detection under context shift" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "To provide a simple baseline investigation for the BugNIST dataset, we investigate some strategies for solving the detection problem under context shift. Since the purpose of our paper is to initiate research in domain adaptation methods" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.127 + ], + "angle": 0, + "content": "P.M. Jensen et al." + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.154, + 0.749, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.366, + 0.788, + 0.41 + ], + "angle": 0, + "content": "Fig. 3: The top line shows the automated annotation pipeline for individual bugs. By thresholding, bounding boxes and segmentation masks are easily obtained. The bottom shows the synthetic mixtures used for the more advanced training strategy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.437, + 0.788, + 0.543 + ], + "angle": 0, + "content": "for volumetric data under context shift, this baseline study is intended as an offset for future developments based on the BugNIST dataset. Our first attempt is to train models from the volumes of individual bugs. Our two other attempts are to use a data augmentation strategy by creating synthetic mixtures. Data augmentation has previously been shown to greatly benefit domain shift problems [60]. The automated annotation of the BugNIST data and the data that we use for training detection algorithms are illustrated in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.788, + 0.692 + ], + "angle": 0, + "content": "Single bugs As a first stage, we train the models to detect the bugs in the individual scans. For models requiring ground truth segmentation labels, those are generated by thresholding the volumes and only keeping the largest connected component. The value of the class label then replaces the foreground voxels. This creates segmentation labels where each voxel has one of 13 values including background or one of the twelve bugs. For models requiring ground truth bounding boxes, those are extracted from each connected component. This process of preparing the training data using this simplest strategy is illustrated in the top line of Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Synthetic mixtures As a second stage, we train the models on artificial mixes. These are generated by randomly sampling individual bug scans and adding them to a common volume. We start with an empty volume of \\(256^3\\) and we set a random target number of bugs drawn from a normal distribution, \\(\\mathcal{N}(32,6)\\). A random bug from a random class (both uniform distributions) is drawn. To create variation, the selected sample is flipped around a random axis, rotated a random multiple (0-3) of 90 degrees around the first axis, and then transposed randomly. The bug volume is intensity normalized, smoothed with a Gaussian filter with \\(\\sigma = 0.5\\), and thresholded at 0.2. We draw a random voxel position" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.399, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "BugNIST - Object Detection under Domain Shift" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "from the volume and only insert the bug if its bounding box does not exceed the volume boundaries nor overlaps with existing labeled voxels. We attempt a maximum of 1000 random positions to achieve a valid insertion before the bug is discarded. In total 300 synthetic mixtures are created." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.222, + 0.79, + 0.328 + ], + "angle": 0, + "content": "Crowded synthetic mixtures Finally, since the bugs are packed very close together in the real scans, we create crowded synthetic mixtures. Here the bugs are packed closer by reducing the volume size to \\(200 \\times 200 \\times 200\\), adjusting the target number of bugs to \\(\\mathcal{N}(64,8)\\), and relaxing the insertion criteria such that no labeled voxels can be overwritten. This effectively allows for a larger overlap between bounding boxes and that bugs can touch each other. Again, a total of 300 synthetic mixtures are created." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.35, + 0.479, + 0.364 + ], + "angle": 0, + "content": "4.2 Baseline detection models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.466 + ], + "angle": 0, + "content": "Despite important uses of detecting objects in 3D volumes, the number of methods addressing volumetric object detection is much smaller compared to 2D, and only a few methods are readily available for volumes. So, to establish baselines for detection under context shift, we have selected and adapted three detection approaches. They vary in how they utilize training data; relying on either ground truth segmentation labels, ground truth bounding boxes, or both." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.789, + 0.721 + ], + "angle": 0, + "content": "Detection U-Net Inspired by [4] we use U-Net [8] as one baseline detection model. U-Net is a dominating model for volumetric segmentation, making it a natural choice for adapting into a baseline detection model. Detection U-Net is trained on volumes accompanied by ground truth segmentation labels. It is first trained to predict the segmentation of the single bugs. Then, starting from the obtained weights, it is trained to predict the segmentation of the synthetic mixtures. Finally, again starting from the previous weights, it is trained to predict the segmentation in the crowded synthetic mixtures. For all training runs, we augment the images with random rotations, translations, geometric scaling, and intensity scaling. We train the models using the sum of Dice and cross entropy loss. During training we hold out a validation split which we evaluate on after every epoch. At the end of training, the model which scored the highest on the validation split is kept. Like [4], we extract bounding boxes by thresholding the softmax probabilities (at 0.75) and then extracting the connected components for each channel. Only components larger than \\(125(5^{3})\\) voxels are kept. We use the U-Net implementation from MONAI [5]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Faster R-CNN As an example of region proposal networks, we have chosen Faster R-CNN [44] with the modifications from [31]. We have adapted the implementation from torchvision [54] to allow for 3D detection. We train in the same sequence as the U-Net: single bugs, synthetic mixtures, and crowded synthetic mixtures. We use the same augmentations as for the U-Net. In addition, we also randomly perturb the bounding box corners. The model is only trained on bounding boxes and does not make use of the voxel labels. The model is" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "P.M. Jensen et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "trained with smooth L1 loss for bounding box regression and cross-entropy for objectness and class predictions. For full details, we refer to [31, 44]. As with U-Net, we hold out a validation split which is evaluated after each epoch. After training, the model scoring the highest on the validation split is kept. For inference, only bounding boxes with an objectness score over 0.05 are kept." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.79, + 0.39 + ], + "angle": 0, + "content": "nnDetection The third detection model, nnDetection, trains the RetinaUNet [22] detection model using a combination of segmentation masks and bounding boxes. The bounding box branch is trained using a generalized IoU loss for regression and cross-entropy for classification. The voxel segmentation branch is trained with a sum of Dice and cross-entropy loss. As nnDetection is a self-configuring framework, we refer to [4, 21] for further training and model details. We simply provide the data as instructed. We train a separate model for single bugs, synthetic mixtures, and crowded synthetic mixtures. nnDetection trains using 5-fold cross-validation, and the final predictions are made using an ensemble over the best model from each fold [4]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.412, + 0.353, + 0.426 + ], + "angle": 0, + "content": "4.3 Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.789, + 0.649 + ], + "angle": 0, + "content": "We assess model performance by evaluating their detection accuracy on real scans of densely packed bugs, using manually annotated center points. These annotations are matched with predicted bounding boxe centers based on Euclidean distance, optimized using Munkres' algorithm [38]. Matches are only accepted if the annotated point falls within a \\(1.5 \\times\\) inflated bounding box, as illustrated in Fig. 4. This inflation ensures a fair assessment for detections like in Fig. 4(b). This, combined with matching on the bounding box centers, helps ensure that large boxes are not favored since it places more weight on the location of the boxes. We compute precision to avoid favoring overdetections, recall to measure the proportion of correct matches, and F1-score to combine the two. Further, we first compute the metrics purely from the point matches without considering classes — thereby evaluating if the models can detect the presence of bugs vs. background material. Next, we recompute the matches where we only allow matches between points of equal class." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.673, + 0.327, + 0.688 + ], + "angle": 0, + "content": "5 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.78 + ], + "angle": 0, + "content": "The results are summarized in Tab. 1, example results are shown in Fig. 5, and the distributions of the metrics without class information are shown in Fig. 6. U-Net performs relatively well, independent of what it was trained on, whereas Faster R-CNN performs worse on crowded mixtures and nnDetection on single bugs. If class labels are required, the performance is generally low." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The distributions in Fig. 6 show that U-Net has a narrower spread around the mean for the F1-score, whereas the two others are spread more out. Wrt. the precision, many detections obtained by the nnDetection method are 1, but have a lower recall, meaning that those that it predicts are correct, but there" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.398, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "BugNIST - Object Detection under Domain Shift" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.148, + 0.426, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.242, + 0.362, + 0.253 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.148, + 0.574, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.493, + 0.242, + 0.511, + 0.253 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.148, + 0.723, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.242, + 0.659, + 0.253 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.351 + ], + "angle": 0, + "content": "Fig. 4: Matching criterion for bounding box to center point. The circle shows the manually annotated center point, the predicted bounding box is shown with the full line, and the inflated bounding box is shown with the dashed line. In (a), the match is accepted by the predicted bounding box, in (b), the match is also accepted due to the inflated bounding box, whereas in (c), there is no match. Note that this illustration is 2D, but the detection is done in 3D." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.785, + 0.428 + ], + "angle": 0, + "content": "are many potential detections that it misses. It is the opposite for the Faster R-CNN, where many datasets get a recall of 1, but a lower precision, meaning that it detects many of the bugs, but also has many false detections." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.641 + ], + "angle": 0, + "content": "To investigate the capabilities of the three methods without the domain shift between the synthetic and real mixtures, we included the results in Tab. 2. This table shows the results of the validation data for training the models on the volumes of individual bugs and the two synthetic mixtures with the class labels. Since we have the bounding boxes for these datasets, we compute the mean average precision (mAP), similar to, e.g., COCO detection [34]. We compute mAP with the IoU threshold running from 0.5 to 0.95 (shown as mAP) and fixed at 0.5 (mAP@50). We note that the values are not directly comparable with those in Tab. 1. We show example detections in the supplementary. The U-Net generally has high performance, Faster R-CNN has high performance for single bugs but has difficulty with the mixtures, whereas nnDetection performs better in the mixtures than for the individual bugs. Generally, the detection performance of bugs is far from perfect and can be improved, also when training and testing within the same domain." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.668, + 0.355, + 0.684 + ], + "angle": 0, + "content": "6 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.84 + ], + "angle": 0, + "content": "The domain shift introduced by shifting the context surrounding the bugs significantly increases the difficulty of the analysis problem. While U-Net and nnDetection achieve high performance on their training tasks (at least for synthetic mixtures), they struggle when transferred to real densely packed scans. It is also interesting that the domain shift has such an effect on the ability of the models to infer the correct class even when accurately localizing the bug. This is surprising since detecting the bugs in the training domain can be done with high accuracy, as shown in Tab. 2, and pure classification (see supplementary) gave an accuracy of 0.97. Despite the models for classification being different than the detection" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "P.M. Jensen et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: Detection results on mixed scans. We show mean and std. for each metric computed over the evaluation images. Higher is better. Without class info refers to not considering the bug label and with class info expects the bug type to be classified as well as being detected." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.213, + 0.789, + 0.365 + ], + "angle": 0, + "content": "
Model- Trained onWithout class info.With class info.
F1-ScorePrecisionRecallF1-ScorePrecisionRecall
U-Net [8]- Single bugs0.59±0.180.68±0.170.60±0.250.11±0.160.13±0.190.11±0.16
- Synth. mixes0.60±0.120.52±0.170.82±0.190.11±0.080.09±0.080.14±0.10
- Crowded s.m.0.54±0.130.39±0.140.97±0.070.10±0.080.07±0.060.18±0.14
Faster R-CNN [44]- Single bugs0.56±0.201.00±0.040.42±0.210.11±0.100.21±0.190.08±0.08
- Synth. mixes0.68±0.160.63±0.210.83±0.220.16±0.140.15±0.140.19±0.18
- Crowded s.m.0.27±0.170.81±0.240.18±0.160.06±0.090.19±0.300.04±0.06
nn-Detection [4]- Single bugs0.03±0.020.02±0.010.91±0.150.00±0.010.00±0.000.09±0.16
- Synth. mixes0.50±0.140.41±0.170.80±0.220.10±0.060.08±0.060.16±0.11
- Crowded s.m.0.60±0.190.77±0.230.58±0.270.11±0.120.15±0.160.11±0.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.383, + 0.789, + 0.466 + ], + "angle": 0, + "content": "Table 2: Detection results on the training tasks obtained during training on the single bugs and the two synthetic mixtures. For U-Net and Faster R-CNN performance is evaluated on a validation split. Since nnDetection is an ensemble created from multiple train-validation splits, we evaluate on all images in the training task. We show mean average precision (mAP) between predicted and ground truth bounding boxes. Higher is better." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.478, + 0.79, + 0.61 + ], + "angle": 0, + "content": "
Model- Trained onmAPmAP@50
U-Net [8]- Single bugs0.720.72
- Synthetic mixes0.640.77
- Crowded synth. mixes0.750.86
Faster R-CNN [44]- Single bugs0.700.70
- Synthetic mixes0.020.02
- Crowded synth. mixes0.050.05
nnDetection [4]- Single bugs0.100.10
- Synthetic mixes0.590.70
- Crowded synth. mixes0.490.66
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.787, + 0.687 + ], + "angle": 0, + "content": "models, these results show that even when the objects do not change appearance, the shift in the domain by changing the context surrounding the objects is sufficient for deep learning-based detection models to lose performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In preparing the data, we have strived to get the scans of individual bugs and bugs in mixtures to have the same appearance. Since bugs are flexible, there will be differences in the pose of the bugs in mixtures and individual bugs. Larvae will sometimes bend differently and bugs with antennae or legs will bend differently. Also, the cotton is visible and can slightly influence the appearance. Despite these differences, the voxel intensity distributions in individual scanned bugs and mixtures are the same, and therefore the BugNIST dataset is a good basis for developing detection methods that can handle context shift. If the mixtures were (partly) labeled for segmentation, BugNIST could also be used to investigate segmentation algorithms under context shift." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.399, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "BugNIST - Object Detection under Domain Shift" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.143, + 0.782, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.466, + 0.788, + 0.577 + ], + "angle": 0, + "content": "Fig. 5: Detection results on real mixed scans. The images are maximum-intensity projections of the volumes. Center points are marked with circles, detections are marked with bounding boxes, and color indicates the bug class. A line from a point to the upper left corner of a bounding box shows a match. If the center point and the bounding box have the same color, then the detection and classification are correct. The shown images are the volumes with the highest scores for the different models. For U-Net the results when trained on synthetic mixes, for Faster R-CNN when trained on synthetic mixes and nnDetection when trained on crowded synthetic mixes." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.596, + 0.777, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.247, + 0.673, + 0.754, + 0.687 + ], + "angle": 0, + "content": "Fig. 6: Distributions of performance scores for the three detection methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We chose to annotate only the center positions of the bugs. As the bugs are often elongated and at odd angles, axis-aligned bounding boxes would contain significant empty space and overlap between neighboring boxes—thus offering limited benefit over points. This could be solved with object aligned bounding boxes or full voxel annotations. However, this would drastically increase the annotation time. We chose to prioritize having a large number of annotations such that the test set can have a large variety of configurations. Furthermore, given that current methods show ample room for improvement (cf. Tab. 1), we" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "P.M. Jensen et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "find that point annotations combined with the chosen metric are sufficient to measure method performance and improvements of future approaches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.328 + ], + "angle": 0, + "content": "The methods and experiments are chosen as a basic baseline to illustrate the properties of the BugNIST dataset. From the experiments, we learned that data augmentation by composing synthetic mixtures is a feasible approach for improving detection performance. More investigations into packing algorithms and ways to combine the synthetic mixed bugs with other backgrounds can potentially improve performance further. But it is also clear that there is great potential for developing new methods for object detection in volumes. We chose Detection U-Net, nnDetection, and our adaption of Faster R-CNN to 3D volumes because these methods were available, widely used, and are based on complementary ideas. We did not find more recent alternatives for volumetric analysis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.788, + 0.465 + ], + "angle": 0, + "content": "Our results show that Faster R-CNN performs less than the two other methods. While it scores highest in Tab. 1, it is clear from Fig. 5 that the bounding boxes are of much lower quality. Furthermore, it does not score well on the training tasks, except when trained on individual bugs. To ensure the poor performance of Faster R-CNN was not just the result of too few training volumes, we redid the experiments with 1000 synthetic mixes instead of the original 300. Performance remained the same. It is also noteworthy that nnDetection scores overall lower than U-Net on both the training and evaluation tasks, even though it also has access to the segmentation masks for training." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.486, + 0.36, + 0.503 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.788, + 0.712 + ], + "angle": 0, + "content": "With our BugNIST benchmark dataset for domain shift in volumetric data, we demonstrate how detecting objects, that have the same appearance in the source and target domain, fails when the objects in the two domains are placed in different contexts. By contexts, we refer to the appearance of the image surrounding the objects. To our knowledge, the BugNIST dataset is the only dataset designed for developing methods for context shift, i.e. domain shift where the objects stay the same but the surroundings change. Solutions will enable an automated annotation of data for volumetric object detection, which is an area that has not received much attention. There are, however, a range of applications where object detection is needed, including medical imaging and emerging areas such as micro-CT scanning of materials. The BugNIST dataset can also be used for investigating other problems such as segmentation under domain shift provided reference annotations are created." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.735, + 0.394, + 0.753 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.841 + ], + "angle": 0, + "content": "We acknowledge the 3D Imaging Center at DTU, the Infrastructure for Quantitative AI-based Tomography (QUAITOM), supported by the Novo Nordisk Foundation (grant number NNF21OC0069766) and STUDIOS: Segmenting Tomograms Using Different Interpretation of Simplicity funded by the Villum Foundation (grant number VIL50425)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.398, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BugNIST - Object Detection under Domain Shift" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.18, + 0.786, + 0.237 + ], + "angle": 0, + "content": "1. Armato, S.G., Huisman, H., Drukker, K., Hadjiiski, L., Kirby, J.S., Petrick, N., Redmond, G., Giger, M.L., Cha, K., Mamonov, A., et al.: PROSTATEx challenges for computerized classification of prostate lesions from multiparametric magnetic resonance images. Journal of Medical Imaging 5(4), 044501-044501 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.239, + 0.787, + 0.307 + ], + "angle": 0, + "content": "2. Armato, S.G., McLennan, G., Bidaut, L., McNitt-Gray, M.F., Meyer, C.R., Reeves, A.P., Zhao, B., Aberle, D.R., Henschke, C.I., Hoffman, E.A., et al.: The lung image database consortium (LIDC) and image database resource initiative (IDRI): a completed reference database of lung nodules on CT scans. Medical Physics 38(2), 915–931 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.308, + 0.787, + 0.363 + ], + "angle": 0, + "content": "3. Baid, U., Ghodasara, S., Mohan, S., Bilello, M., Calabrese, E., Colak, E., Farahani, K., Kalpathy-Cramer, J., Kitamura, F.C., Pati, S., et al.: The RSNA-ASNR-MICCAI BraTS 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.364, + 0.787, + 0.419 + ], + "angle": 0, + "content": "4. Baumgartner, M., Jäger, P.F., Isensee, F., Maier-Hein, K.H.: nnDetection: a self-configuring method for medical object detection. In: Proceedings of the Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 530-539. Springer (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.787, + 0.462 + ], + "angle": 0, + "content": "5. Cardoso, M.J., Li, W., Brown, R., Ma, N., Kerfoot, E., Wang, Y., Murrey, B., Myronenko, A., Zhao, C., Yang, D., et al.: MONAI: An open-source framework for deep learning in healthcare. arXiv preprint arXiv:2211.02701 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.463, + 0.787, + 0.504 + ], + "angle": 0, + "content": "6. Charytanowicz, M., Kulczycki, P., Kowalski, P.A., Lukasik, S., Czabak-Garbacz, R.: An evaluation of utilizing geometric features for wheat grain classification using X-ray images. Computers and Electronics in Agriculture 144, 260-268 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.505, + 0.787, + 0.574 + ], + "angle": 0, + "content": "7. Chaudhary, S., Sadbhawna, S., Jakhetiya, V., Subudhi, B.N., Baid, U., Guntuku, S.C.: Detecting COVID-19 and community acquired pneumonia using chest CT scan images with deep learning. In: Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 8583-8587. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.575, + 0.787, + 0.63 + ], + "angle": 0, + "content": "8. Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3D U-Net: learning dense volumetric segmentation from sparse annotation. In: Proceedings of the Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 424–432. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.631, + 0.787, + 0.686 + ], + "angle": 0, + "content": "9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3213-3223 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.687, + 0.787, + 0.742 + ], + "angle": 0, + "content": "0. De Carlo, F., Gürsoy, D., Ching, D.J., Batenburg, K.J., Ludwig, W., Mancini, L., Marone, F., Mokso, R., Pelt, D.M., Sijbers, J., et al.: Tomobank: a tomographic data repository for computational x-ray science. Measurement Science and Technology 29(3), 034004 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.743, + 0.787, + 0.798 + ], + "angle": 0, + "content": "1. Dorent, R., Kujawa, A., Ivory, M., Bakas, S., Rieke, N., Joutard, S., Glocker, B., Cardoso, J., Modat, M., Batmanghelich, K., et al.: Crossmoda 2021 challenge: Benchmark of cross-modality domain adaptation techniques for vestibular schwannoma and cochlea segmentation. Medical Image Analysis 83, 102628 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "2. Du, Z., Hu, Y., Ali Buttar, N., Mahmood, A.: X-ray computed tomography for quality inspection of agricultural products: A review. Food Science & Nutrition 7(10), 3146-3160 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.18, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "P.M. Jensen et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "13. Einarsdóttir, H., Emerson, M.J., Clemmensen, L.H., Scherer, K., Willer, K., Bech, M., Larsen, R., Ersbøll, B.K., Pfeiffer, F.: Novelty detection of foreign objects in food using multi-modal X-ray imaging. Food Control 67, 39-47 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "14. European Organization For Nuclear Research, OpenAIRE: Zenodo (2013). https://doi.org/10.25495/7GXK-RD71, https://www.zenodo.org/" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "15. Flitton, G., Mouton, A., Breckon, T.P.: Object classification in 3D baggage security computed tomography imagery using visual codebooks. Pattern Recognition 48(8), 2489-2499 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.286 + ], + "angle": 0, + "content": "16. Girshick, R.: Fast R-CNN. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 1440-1448 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.342 + ], + "angle": 0, + "content": "17. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object detection and semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 580-587 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.37 + ], + "angle": 0, + "content": "18. Guan, H., Liu, M.: Domain adaptation for medical image analysis: a survey. IEEE Transactions on Biomedical Engineering 69(3), 1173-1185 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.424 + ], + "angle": 0, + "content": "19. Harmon, S.A., Sanford, T.H., Xu, S., Turkbey, E.B., Roth, H., Xu, Z., Yang, D., Myronenko, A., Anderson, V., Amalou, A., et al.: Artificial intelligence for the detection of COVID-19 pneumonia on chest CT using multinational datasets. Nature Communications 11(1), 4080 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "20. Hipsley, C.A., Aguilar, R., Black, J.R., Hocknull, S.A.: High-throughput microCT scanning of small specimens: preparation, packing, parameters and post-processing. Scientific Reports 10(1), 13863 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "21. Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods 18(2), 203-211 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.563 + ], + "angle": 0, + "content": "22. Jaeger, P.F., Kohl, S.A., Bickelhaupt, S., Isensee, F., Kuder, T.A., Schlemmer, H.P., Maier-Hein, K.H.: Retina U-Net: Embarrassingly simple exploitation of segmentation supervision for medical object detection. In: Machine Learning for Health Workshop. pp. 171-183. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.605 + ], + "angle": 0, + "content": "23. Jain, Y., Borner, K., Walsh, C., Ruschman, N., Lee, P.D., Weber, G.M., Holbrook, R., Howard, A.: SenNet + HOA - hacking the human vasculature in 3D (2023), https://kaggle.com/competitions/blood-vessel-segmentation" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "24. Jarolmasjed, S., Espinoza, C.Z., Sankaran, S., Khot, L.R.: Postharvest bitter pit detection and progression evaluation in 'Honeycrisp' apples using computed tomography images. Postharvest Biology and Technology 118, 35-42 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "25. Jeppesen, N., Christensen, A.N., Dahl, V.A., Dahl, A.B.: Sparse layered graphs for multi-object segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12777-12785 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.743 + ], + "angle": 0, + "content": "26. Jin, L., Yang, J., Kuang, K., Ni, B., Gao, Y., Sun, Y., Gao, P., Ma, W., Tan, M., Kang, H., et al.: Deep-learning-assisted detection and segmentation of rib fractures from CT scans: Development and validation of FracNet. EBioMedicine 62, 103106 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.785 + ], + "angle": 0, + "content": "27. Ker, J., Singh, S.P., Bai, Y., Rao, J., Lim, T., Wang, L.: Image thresholding improves 3-dimensional convolutional neural network diagnosis of different acute brain hemorrhages on computed tomography scans. Sensors 19(9), 2167 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "28. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.399, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BugNIST - Object Detection under Domain Shift" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "29. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "30. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.M.: Deeper, broader and artier domain generalization. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 5542-5550 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.245 + ], + "angle": 0, + "content": "31. Li, Y., Xie, S., Chen, X., Dollar, P., He, K., Girshick, R.: Benchmarking detection transfer learning with vision transformers. arXiv preprint arXiv:2111.11429 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.273 + ], + "angle": 0, + "content": "32. Li, Y., Fan, Y.: Medical image segmentation with domain adaptation: A survey. arXiv preprint arXiv:2311.01702 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "33. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 2980-2988 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.357 + ], + "angle": 0, + "content": "34. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.384 + ], + "angle": 0, + "content": "35. Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature Communications 15(1), 654 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.438 + ], + "angle": 0, + "content": "36. Maier-Hein, L., Eisenmann, M., Reinke, A., Onogur, S., Stankovic, M., Scholz, P., Arbel, T., Bogunovic, H., Bradley, A.P., Carass, A., et al.: Why rankings of biomedical image analysis competitions should be interpreted with care. Nature Communications 9(1), 5217 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "37. Muandet, K., Balduzzi, D., Scholkopf, B.: Domain generalization via invariant feature representation. In: International Conference on Machine Learning. pp. 10-18. PMLR (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "38. Munkres, J.: Algorithms for the assignment and transportation problems. Journal of the Society for Industrial and Applied Mathematics 5(1), 32-38 (1957)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.522 + ], + "angle": 0, + "content": "39. OECD: Health at a Glance 2023 (2023). https://doi.org//10.1787/7a7afb35-en" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.55 + ], + "angle": 0, + "content": "40. Pan, S.J., Tsang, I.W., Kwok, J.T., Yang, Q.: Domain adaptation via transfer component analysis. IEEE Transactions on Neural Networks 22(2), 199-210 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "41. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1406-1415 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.619 + ], + "angle": 0, + "content": "42. Rawson, S.D., Maksimcuka, J., Withers, P.J., Cartmell, S.H.: X-ray computed tomography in life sciences. BMC biology 18(1), 1-15 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.66 + ], + "angle": 0, + "content": "43. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 779-788 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.702 + ], + "angle": 0, + "content": "44. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. Advances in neural information processing systems (NeurIPS) 28 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.758 + ], + "angle": 0, + "content": "45. Richter, S.R., Vineet, V., Roth, S., Koltun, V.: Playing for data: Ground truth from computer games. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 102-118. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.813 + ], + "angle": 0, + "content": "46. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: Proceedings of the Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 234-241. Springer (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "47. Ros, G., Sellart, L., Materzynska, J., Vazquez, D., Lopez, A.M.: The synthia dataset: A large collection of synthetic images for semantic segmentation of urban" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.127 + ], + "angle": 0, + "content": "P.M. Jensen et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.148, + 0.786, + 0.175 + ], + "angle": 0, + "content": "scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3234-3243 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.178, + 0.787, + 0.232 + ], + "angle": 0, + "content": "48. Sankaranarayanan, S., Balaji, Y., Jain, A., Lim, S.N., Chellappa, R.: Learning from synthetic data: Addressing domain shift for semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3752-3761 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.786, + 0.261 + ], + "angle": 0, + "content": "49. Serte, S., Demirel, H.: Deep learning for diagnosis of COVID-19 using 3D CT scans. Computers in biology and medicine 132, 104306 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.263, + 0.787, + 0.331 + ], + "angle": 0, + "content": "50. Setio, A.A.A., Traverso, A., De Bel, T., Berens, M.S., Van Den Bogaard, C., Cerello, P., Chen, H., Dou, Q., Fantacci, M.E., Geurts, B., et al.: Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: the luna16 challenge. Medical Image Analysis 42, 1-13 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.333, + 0.787, + 0.429 + ], + "angle": 0, + "content": "51. Simpson, A.L., Antonelli, M., Bakas, S., Bilello, M., Farahani, K., van Ginneken, B., Kopp-Schneider, A., Landman, B.A., Litjens, G., Menze, B.H., Ronneberger, O., Summers, R.M., Bilic, P., Christ, P.F., Do, R.K.G., Gollub, M., Golia-Pernicka, J., Heckers, S., Jarnagin, W.R., McHugo, M., Napel, S., Vorontsov, E., Maier-Hein, L., Cardoso, M.J.: A large annotated medical image dataset for the development and evaluation of segmentation algorithms. CoRR abs/1902.09063 (2019), http://arxiv.org/abs/1902.09063" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.787, + 0.486 + ], + "angle": 0, + "content": "52. Sun, T., Segu, M., Postels, J., Wang, Y., Van Gool, L., Schiele, B., Tombari, F., Yu, F.: Shift: a synthetic driving dataset for continuous multi-task domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21371-21382 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.487, + 0.787, + 0.543 + ], + "angle": 0, + "content": "53. Timmins, K.M., van der Schaaf, I.C., Bennink, E., Ruigrok, Y.M., An, X., Baumgartner, M., Bourdon, P., De Feo, R., Di Noto, T., Dubost, F., et al.: Comparing methods of detecting and segmenting unruptured intracranial aneurysms on TOF-MRAS: The ADAM challenge. Neuroimage 238, 118216 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.787, + 0.572 + ], + "angle": 0, + "content": "54. TorchVision maintainers and contributors: Torchvision: Pytorch's computer vision library. https://github.com/pytorch/vision (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.787, + 0.613 + ], + "angle": 0, + "content": "55. Trusler, M.M., Sturrock, C.J., Vane, C.H., Cook, S., Lomax, B.H.: X-ray computed tomography: A novel non-invasive approach for the detection of microplastics in sediments? Marine Pollution Bulletin 194, 115350 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.787, + 0.669 + ], + "angle": 0, + "content": "56. Velayudhan, D., Hassan, T., Damiani, E., Werghi, N.: Recent advances in baggage threat detection: A comprehensive and systematic survey. ACM Computing Surveys 55(8) (dec 2022). https://doi.org/10.1145/3549932, https://doi.org/10.1145/3549932" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.671, + 0.787, + 0.712 + ], + "angle": 0, + "content": "57. Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5018-5027 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.787, + 0.769 + ], + "angle": 0, + "content": "58. Walsh, C., Tafforeau, P., Wagner, W., Jafree, D., Bellier, A., Werlein, C., Kühnel, M., Boller, E., Walker-Samuel, S., Robertus, J., et al.: Imaging intact human organs with local resolution of cellular structures using hierarchical phase-contrast tomography. Nature Methods 18(12), 1532–1541 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.77, + 0.787, + 0.798 + ], + "angle": 0, + "content": "59. Wang, M., Deng, W.: Deep visual domain adaptation: A survey. Neurocomputing 312, 135-153 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "60. Wang, M., Liu, Y., Yuan, J., Wang, S., Wang, Z., Wang, W.: Inter-class and inter-domain semantic augmentation for domain generalization. IEEE Transactions on Image Processing (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.399, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "BugNIST - Object Detection under Domain Shift" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "61. Wang, Q., Bhowmik, N., Breckon, T.P.: Multi-class 3D object detection within volumetric 3D computed tomography baggage security screening imagery. In: Proceedings of the IEEE International Conference on Machine Learning and Applications (ICMLA). pp. 13-18. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.788, + 0.258 + ], + "angle": 0, + "content": "62. Wang, Q., Breckon, T.P.: Contraband materials detection within volumetric 3D computed tomography baggage security screening imagery. In: Proceedings of the IEEE International Conference on Machine Learning and Applications (ICMLA). pp. 75-82. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.3 + ], + "angle": 0, + "content": "63. Withers, P.J., Bouman, C., Carmignato, S., Cnudde, V., Grimaldi, D., Hagen, C.K., Maire, E., Manley, M., Du Plessis, A., Stock, S.R.: X-ray computed tomography. Nature Reviews Methods Primers 1(1), 18 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.355 + ], + "angle": 0, + "content": "64. Wu, D., Wu, D., Feng, H., Duan, L., Dai, G., Liu, X., Wang, K., Yang, P., Chen, G., Gay, A.P., et al.: A deep learning-integrated micro-ct image analysis pipeline for quantifying rice lodging resistance-related traits. Plant Communications 2(2) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.384 + ], + "angle": 0, + "content": "65. Xiao, Z., Liu, B., Geng, L., Zhang, F., Liu, Y.: Segmentation of lung nodules using improved 3d-unet neural network. Symmetry 12(11), 1787 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.424 + ], + "angle": 0, + "content": "66. Xie, Y., Ji, Q.: A new efficient ellipse detection method. In: Proceedings of the International Conference on Pattern Recognition (ICPR). vol. 2, pp. 957-960. IEEE (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.452 + ], + "angle": 0, + "content": "67. Xu, F., Tenenbaum, J.B.: Word learning as bayesian inference. Psychological review 114(2), 245 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.507 + ], + "angle": 0, + "content": "68. Yan, Y.T., Chua, S., DeCarlo, T.M., Kempf, P., Morgan, K.M., Switzer, A.D.: Core-ct: A matlab application for the quantitative analysis of sediment and coral cores from X-ray computed tomography (ct). Computers & Geosciences 156, 104871 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.55 + ], + "angle": 0, + "content": "69. Zhao, W.X., Zhou, K., Li, J., Tang, T., Wang, X., Hou, Y., Min, Y., Zhang, B., Zhang, J., Dong, Z., et al.: A survey of large language models. arXiv preprint arXiv:2303.18223 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.617 + ], + "angle": 0, + "content": "70. Zhou, K., Yang, Y., Hospedales, T., Xiang, T.: Deep domain-adversarial image generation for domain generalisation. Proceedings of the AAAI Conference on Artificial Intelligence 34(07), 13025-13032 (April 2020). https://doi.org/10.1609/aaii.v34i07.7003, https://ojs.aaii.org/index.php/AAAI/article/view/7003" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.618, + 0.788, + 0.675 + ], + "angle": 0, + "content": "71. Zhou, S.K., Greenspan, H., Davatzikos, C., Duncan, J.S., Van Ginneken, B., Madabhushi, A., Prince, J.L., Rueckert, D., Summers, R.M.: A review of deep learning in medical imaging: Imaging traits, technology trends, case studies with progress highlights, and future promises. Proceedings of the IEEE 109(5), 820-838 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.675 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/BugNIST - a Large Volumetric Dataset for Detection under Domain Shift/fceefb33-6fd3-4a63-9cb1-f0d439dc4d49_origin.pdf b/2024/BugNIST - a Large Volumetric Dataset for Detection under Domain Shift/fceefb33-6fd3-4a63-9cb1-f0d439dc4d49_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..32311296cdb6138d90c65dc3a8cab41da5273f9e --- /dev/null +++ b/2024/BugNIST - a Large Volumetric Dataset for Detection under Domain Shift/fceefb33-6fd3-4a63-9cb1-f0d439dc4d49_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf049efce4b687c7b62c86d9621cc93d9eca18297315d6c8e9338bc5832b8770 +size 2656188 diff --git a/2024/BugNIST - a Large Volumetric Dataset for Detection under Domain Shift/full.md b/2024/BugNIST - a Large Volumetric Dataset for Detection under Domain Shift/full.md new file mode 100644 index 0000000000000000000000000000000000000000..809e1d604f057236ea12da8437292fd9a439008b --- /dev/null +++ b/2024/BugNIST - a Large Volumetric Dataset for Detection under Domain Shift/full.md @@ -0,0 +1,273 @@ +# BugNIST - a Large Volumetric Dataset for Object Detection under Domain Shift + +Patrick Møller Jensen, Vedrana Andersen Dahl, Rebecca Engberg, Carsten Gundlach, Hans Marin Kjer, and Anders Bjorholm Dahl + +Technical University of Denmark, Kgs. Lyngby, Denmark {patmjen, vand, reen}@dtu.dk, cagu@fysik.dtu.dk, {hmjk, abda}@dtu.dk + +Abstract. Domain shift significantly influences the performance of deep learning algorithms, particularly for object detection within volumetric 3D images. Annotated training data is essential for deep learning-based object detection. However, annotating densely packed objects is time-consuming and costly. Instead, we suggest training models on individually scanned objects, causing a domain shift between training and detection data. To address this challenge, we introduce the BugNIST dataset, comprising 9154 micro-CT volumes of 12 bug types and 388 volumes of tightly packed bug mixtures. This dataset is characterized by having objects with the same appearance in the source and target domains, which is uncommon for other benchmark datasets for domain shift. During training, individual bug volumes labeled by class are utilized, while testing employs mixtures with center point annotations and bug type labels. Together with the dataset, we provide a baseline detection analysis, with the aim of advancing the field of 3D object detection methods. + +Keywords: Volumetric Dataset, Benchmark, Volumetric Object Detection, Domain Shift. + +# 1 Introduction + +Our work on domain shift in volumetric 3D images is motivated by the need for labeled data to train supervised deep learning models for volumetric imaging. The problem that we aim to solve is object detection and classification. We propose to label images of objects scanned as isolated entities as the source domain and use these as a basis to train models for object detection and classification in a complex context of mixed objects and other materials as the target domain. The effort needed for obtaining labels in the two domains is significantly different. If objects are isolated, they can be automatically labeled whereas mixed objects require expensive manual labeling. Automating the labeling based on isolated objects, however, leads to a domain shift between the data in the source domain for training and the data in the target domain for detection and classification. This domain shift is special because the appearance of the objects is the same + +![](images/ba200f363445f099ca69ad44d26396a72c9d667e3c50bc5e33df3352a7dc9d47.jpg) +(a) 9154 volumes of individual bugs with class labels. + +![](images/265b46fe7ce8a15fda6f80c106c3407d9da56e3764376b9caa07cd0a91feea43.jpg) +(b) 388 volumes of bug mixtures with center point and class labels. +Fig. 1: Overview of the two parts of the dataset. In (a), representative individual bug volumes are displayed with a class abbreviation, name, and number of volumes. In (b), a few volumes of bug mixtures are shown. Each bug is annotated with bug class and center point. + +in the two domains, but the surrounding context is different. We use the term context shift for this type of domain shift. + +To promote the development of new deep learning-based methods that can handle context shift, we contribute the volumetric dataset called BugNIST1. The dataset contains 9154 micro-CT volumes with individual bugs and 388 micro-CT volumes containing a mixture of several bugs as well as other materials such as leaves, wood shavings, etc.; see Fig. 1. The challenge is to use the individually scanned bugs to detect and classify the bugs in the mixtures. The dataset is intended to be a relatively simple test case for starting to develop methods for detecting objects in volumetric images under context shift, and bugs are chosen due to their complex shape and not for solving a problem in entomology. + +As a basis for training an accurate deep learning model, a training dataset for object detection and classification must capture the appearance distribution of the object classes that are in the dataset. Since the appearance of a CT-scanned object changes minimally if it is scanned individually or as part of a mixture, the appearance distribution can be covered by only considering the individual scans. This is different from 2D photos. Here additional variations in scale, viewpoint, and illumination influence the appearance of objects, and it would be more difficult to acquire representative images of individual objects for a dataset based on 2D photos. + +Since BugNIST is a test case for context shift, we wanted to ensure that each class is covered well. By scanning systematically arranged bugs, we have generated 9154 volumes of individual bugs, each labeled with its corresponding class. Furthermore, we scanned 388 mixed volumes, each containing between 4 and 45 bugs alongside other materials. For evaluation purposes, each bug in the mixtures is manually annotated with its center position and class label. In total, + +9542 volumes make up the BugNIST dataset. The numbers in each class are shown in Fig. 1. Since the bugs in each class are the same age, we expect to cover the biological variation sufficiently to be representative of the class. + +Solutions to the type of context shift, that we present here, namely a domain shift where only the context changes, and not the appearance of the objects, have wide applications in volumetric imaging. The field of volumetric imaging is dominated by medical applications [39], but there is a growing use beyond the medical area based on e.g. micro-CT [58, 63]. In all these application areas, there is a need to employ supervised deep learning to solve problems that require quantification of depicted structures. When the use case is novel and there is no directly applicable training data available — a typical scenario in volumetric imaging of e.g. materials microstructure, but also in specialized biomedical applications — it is necessary to create training data. If labor-intensive manual labeling can be limited or avoided, there is potentially a huge gain in productivity. Examples of applications are detecting and classifying objects in scans of marine sediments such as microplastics [55, 68], museum specimens [20], or detecting threat items in baggage security screening [56]. + +In volumetric scans, manual labeling is significantly more demanding than in 2D. For instance, annotating a bounding box of a single object in 3D requires traversing slices of the volume and making a mental model of the object. If the volume contains densely packed complex objects, determining their bounding boxes becomes difficult. Further, bounding boxes may overlap to such an extent that they stop being a useful way to locate objects. This is why we employ center points as the annotation format in the mixtures. + +The name BugNIST is a homage to the influential MNIST dataset [29] which has played a pivotal role in the development and evaluation of deep learning methods. Similar to the way that MNIST has been used, BugNIST is intended to be a benchmark dataset that represents a well-defined problem, where the development of solutions will make a basis for solutions that can advance the field. Specifically for BugNIST, solutions will enable object detection and classification with significantly less effort put into creating the data foundation for training. This calls for methods that can learn to predict objects in a complex context from volumes of individual objects, namely solving the context shift problem. To our knowledge, no existing datasets have this property. + +The term context shift refers to a shift in the data distribution between training and inference. Typical shifts in data distributions arise from changes in data acquisition systems, e.g. the use of different scanners, or from variations in the objects of interest between datasets. Such problems have been addressed using domain adaptation [18,59]. In our setting, there is almost no change in the appearance of the bugs between volumes of individual bugs and mixtures, but the change is in what surrounds the bugs, i.e. the change is in the context. + +We hypothesize that methods capable of learning under context shift might also prove capable of learning from very few examples, similar to how humans learn [67]. Instead of relying on seeing objects in every imaginable context, such methods could instead isolate objects from their contexts. If this hypothesis holds + +true, learning under context shift may reduce the need for extensive training data and the associated training costs. This is in contrast to the current trend, where larger models and datasets as seen in LLMs [69] and computer vision [28, 35], lead to better performance, but also limit who can effectively train such models. + +Contributions We contribute BugNIST, a new volumetric image dataset for benchmarking domain shift. In this paper, we use BugNIST for benchmarking detection methods under context shift, trained only on individual bugs and tested on mixtures. We provide several baselines and investigate different training strategies for existing volumetric detection methods. + +Our results show that volumetric object detection under context shift is difficult. We obtain reasonable results in localizing the bugs, but when a class label is assigned, there is a significant drop in performance. This is even though the classification of individual bugs is easy. Our experiments demonstrate great potential for developing new approaches for 3D object detection of complex objects. + +# 2 Related work + +In computer vision, the term 3D data refers to different data types, including point clouds, surfaces, volumes, and depth images. Here, we are only concerned with 3D volumes, i.e. image data represented on a voxel grid. + +BugNIST is created to advance deep learning methods for object detection and classification under domain shift. Especially in medical imaging, analysis of 3D volumes is in great demand. A review of medical challenges illustrates this [36]. Their study shows that detection makes up $7\%$ of the tasks, classification makes up $10\%$ of the tasks, whereas segmentation accounts for $70\%$ of the tasks. Detection is important for example for diagnosis [7, 19, 27, 49, 71]. Outside the medical domain, object detection problems include baggage security screening [15, 61, 62], although the problem and the available data are still primarily in the 2D domain [56], food inspection [12] such as fruit quality [24], grain classification [6], foreign object detection [13], biological applications such as analysis of weeds [64] and other uses driven by developments in X-ray micro-CT imaging [42], etc. + +The domain shift problem presented in BugNIST, where only the context changes, and not the appearance of the objects of interest, is not widely studied. Most research on domain shifts focuses primarily on problems where the appearance of objects changes. Solutions focus on domain adaptation, where labeled data in the source domain are available together with data without labels from the target domain [40], or domain generalization, where the inference domain is not known at training time [37, 60]. We focus on domain adaptation, which is also the typical use case for volumetric imaging. + +Several widely used datasets are available for studying and developing methods for domain shift. For 2D image classification problems the datasets PACS [30], Digits-DG [70], Office-Home [57], and DomainNet [41] developed to classify images from domains such as photos, art paintings, sketches, etc. Domain shift between artificially generated and real images e.g. for segmentation is another area that receives much attention [48] with standard datasets such as + +SYNTHEIA [47], CITYSCAPES [9], and GTA-5 [45], or completely synthetic domain shift like the SHIFT dataset [52] for autonomous driving. + +In 3D, domain shift is often studied in medical applications where the same type of object is imaged with different scanning modalities [32]. It can be MRI (magnetic resonance imaging), CT, OCT (optical coherence tomography), etc. Standard datasets include the CrossMoDA 2021 dataset [11] and the BRATS dataset [3], both of different MRI modalities. There are many other datasets used for domain adaptation for volumetric medical imaging [32], but all for adapting between modalities or between the same modality but different scanners [18]. + +The structure of the imaged objects in BugNIST is different from most existing volumetric deep learning benchmark datasets. The morphological complexity of bugs is, however, representative of volumetric imaging data of e.g. 3D microstructure, where there is much less use of deep learning than in e.g. the medical domain [63]. + +Platforms like Zenodo [14], TomoBank [10], or the Human Organ Atlas [58] also contain a large amount of volumetric image datasets. Different from e.g. medical challenges, this data generally comes with no clear problem formulation, well-defined tasks or even labels, that allow for a comparison of methods. Further, their large volume sizes are difficult to access. It can also be difficult to judge the complexity of the data and if deep learning is needed. Some images can easily be analyzed using a threshold whereas others are complex and can greatly benefit from deep learning. Therefore, most of this data is not attractive to use as a benchmark for developing new deep learning methods. Some datasets are emerging, e.g. the vessel segmentation challenge [23], but they remain exceptions. + +There is extensive research in deep learning-based object detection, but primarily in 2D. Most methods either use a two-stage detector such as R-CNN [16, 17, 44], with a region proposal and a classifier to choose among the proposed regions, or a one-stage detector such asYOLO [43], where regions and classes are proposed in a single network. + +Object detection in volumetric images is less common. One approach has been to apply a 3D U-Net [8] segmentation network and use each connected component of the foreground label class as a detection [22]. An example of using this approach is lung nodule detection in [65]. Employing segmentation for object detection comes with the problem of separating objects that may be connected or combining separated segments into one object. To avoid this problem, Jaeger et al. [22] propose the Retina U-Net: a one-shot 3D detection algorithm that combines the Retina Net [33] and U-Net [46] giving detections with 3D bounding boxes. Despite the good performance of Retina U-Net, changing to a new 3D detection problem requires a time-consuming method configuration. This has been addressed in nnDetection [4]. Inspired by the nnU-Net (no new U-Net) [21], they propose a no new 3D detection method. On the BugNIST data, we test detection with 3D U-Net [8], the nnDetection method [4], and a 3D adaption of the Faster R-CNN model [44]. + +Benchmark datasets for volumetric object detection are primarily medical. Examples include Baumgartner et al. [4], which uses 13 datasets to develop the + +nnDetection method. Two of these datasets were originally targeted 3D detection and include the LUNA16 for nodule detection [50], which is a subset of the LIDC dataset [2] and contains 888 CT scans of lungs, and the ADAM dataset [53] for detecting aneurysms in MRI scans. ADAM is made up of 254 MRI scans. The rest of the detection datasets are medical segmentation datasets such as data from the Medical Segmentation Decathlon [51], the rib fracture detection [26], the PROSTATEx dataset [1], and others. The segmentation datasets have been transformed into detection data by identifying center positions and bounding boxes for the segments. Within baggage security screening, some 3D datasets are obtainable, for example, the ALERT datasets used in [62]. Due to the sensitive nature of the problem, these datasets have copyright restrictions and limited availability under a non-disclosure agreement. + +# 3 Data acquisition and processing + +Fig. 1 gives an overview of the BugNIST dataset. Each individual bug volume contains exactly one bug located near the center of the volume. The total of 9154 individual bug volumes have between 713 and 946 volumes of each class distributed across 12 bug classes (species and stages). Bug mixtures contain several bugs, leaves, cotton, shredded paper, and wood shaving. There are in total 388 mixtures. The mixtures contain 4 to 45 bugs, totaling 5011 bugs with an average of 13 bugs per mixture, see Fig. 2. In bug mixtures, we have manually marked the center positions of each bug and provided it with a class label. + +All scans were acquired using a laboratory micro-CT scanner (Nikon metrology scanner) that was set to record volumes of 2000 voxels cubed with an isotropic voxel size of $26.62\mu \mathrm{m}$ . Being able to acquire such large volumes allowed us to scan multiple bugs and bug mixtures at once, and then crop them into smaller volumes. The dataset was collected through nearly 250 scans, each taking 30 minutes, over 90 days. Care has been taken that the scanning parameters stay unchanged for all scans, and we have calibrated the scanner daily. + +All bugs were placed in tubes, either as individuals or in mixtures. We used cotton, which is almost transparent to X-rays, to separate the bugs, either as individual bugs or mixtures. For the mixtures, we varied the difficulty of the bug detection task by adding leaves, shredded paper, or wood shavings to some of the mixtures, and also varied the number of bugs, their class, and the packing density. + +The cylindrical shape of the tubes was easy to segment using a circular Hough-transform [66] and sparse layered graphs [25], leaving us with a cylinder of only bugs and cotton. Due to the low intensity of cotton compared to the bugs, we could easily divide each cylinder into volumes containing one bug each. The volumes have been manually inspected and cleaned to correct cases where something went wrong, e.g. parts of another bug showing in the periphery of the volume. The final volumes of scans of individual bugs are $450 \times 450 \times 900$ voxels, the size which was chosen to accommodate the biggest bugs. The mixtures were segmented similarly and the resulting volumes are $650 \times 650 \times 900$ voxels. All + +![](images/249e81cfd94ad453faa8285871dd5e92a0e132261ca3aa141934562277587715.jpg) +(a) Class distribution for bugs from mixture volumes. + +![](images/4b0d018cfc209628bed82289d0347f3ab40db26b9b666bb8624bf56be349bb4b.jpg) +Fig. 2: Summary of volumes from scans of bug mixtures. + +![](images/b5103bb97060de34a90c6d8d5b6a0968158601f4e6acb9dd4044e2e4f785c45f.jpg) +(b) Distribution of the number of bugs per volume and the number of classes per volume. + +volumes are stored as 8-bit unsigned integers with their original spatial resolution of $26.62\mu \mathrm{m}$ (physical voxel size). Our dataset comprises 1.8 trillion voxels, approximately a quarter of the RGB values in ImageNet. + +The bugs scanned for the BugNIST dataset are bred as pet food and have been frozen before being packed for scanning. Despite careful packing, some scanned bugs may appear damaged, e.g. a few crickets have broken off antennas or miss a leg. We have chosen to accept these defects because they reflect some variation in a population and add to the challenge of the dataset. + +To create additional lightweight datasets, we created four downsized versions of all volumes. The individual bug volumes, originally 900 voxels long, have lengths of 512, 256, 128, and 64 voxels in the downscaled versions, and all the other lengths are scaled accordingly. The voxel size of the coarsest data is $374.3\mu \mathrm{m}$ , which is a scaling factor of approximately 14 compared to the original scan. + +# 4 Experiments + +To establish baselines for volumetric detection under context shift (domain shift with no change in object appearance), we have selected and adapted three existing detection models. Furthermore, we have applied three training strategies, yielding a total of nine baseline models. The diversity in model choices and training strategies contributes to a comprehensive set of baseline models that provide a strong foundation for object detection under context shift. We briefly describe the selected methods and how they were trained. Full details are provided in the supplementary material. To determine the difficulty of differentiating between the individual bugs, we also performed a classification experiment which is described in the supplementary material. + +All experiments are carried out using the x128 data, which is a compromise between memory needs and the preservation of structural details of the bugs. + +# 4.1 Training strategies for detection under context shift + +To provide a simple baseline investigation for the BugNIST dataset, we investigate some strategies for solving the detection problem under context shift. Since the purpose of our paper is to initiate research in domain adaptation methods + +![](images/633471798fd54ad57ec676f5d61247ac367820706fe92530f776fc06a3cf80ae.jpg) +Fig. 3: The top line shows the automated annotation pipeline for individual bugs. By thresholding, bounding boxes and segmentation masks are easily obtained. The bottom shows the synthetic mixtures used for the more advanced training strategy. + +for volumetric data under context shift, this baseline study is intended as an offset for future developments based on the BugNIST dataset. Our first attempt is to train models from the volumes of individual bugs. Our two other attempts are to use a data augmentation strategy by creating synthetic mixtures. Data augmentation has previously been shown to greatly benefit domain shift problems [60]. The automated annotation of the BugNIST data and the data that we use for training detection algorithms are illustrated in Fig. 3. + +Single bugs As a first stage, we train the models to detect the bugs in the individual scans. For models requiring ground truth segmentation labels, those are generated by thresholding the volumes and only keeping the largest connected component. The value of the class label then replaces the foreground voxels. This creates segmentation labels where each voxel has one of 13 values including background or one of the twelve bugs. For models requiring ground truth bounding boxes, those are extracted from each connected component. This process of preparing the training data using this simplest strategy is illustrated in the top line of Fig. 3. + +Synthetic mixtures As a second stage, we train the models on artificial mixes. These are generated by randomly sampling individual bug scans and adding them to a common volume. We start with an empty volume of $256^3$ and we set a random target number of bugs drawn from a normal distribution, $\mathcal{N}(32,6)$ . A random bug from a random class (both uniform distributions) is drawn. To create variation, the selected sample is flipped around a random axis, rotated a random multiple (0-3) of 90 degrees around the first axis, and then transposed randomly. The bug volume is intensity normalized, smoothed with a Gaussian filter with $\sigma = 0.5$ , and thresholded at 0.2. We draw a random voxel position + +from the volume and only insert the bug if its bounding box does not exceed the volume boundaries nor overlaps with existing labeled voxels. We attempt a maximum of 1000 random positions to achieve a valid insertion before the bug is discarded. In total 300 synthetic mixtures are created. + +Crowded synthetic mixtures Finally, since the bugs are packed very close together in the real scans, we create crowded synthetic mixtures. Here the bugs are packed closer by reducing the volume size to $200 \times 200 \times 200$ , adjusting the target number of bugs to $\mathcal{N}(64,8)$ , and relaxing the insertion criteria such that no labeled voxels can be overwritten. This effectively allows for a larger overlap between bounding boxes and that bugs can touch each other. Again, a total of 300 synthetic mixtures are created. + +# 4.2 Baseline detection models + +Despite important uses of detecting objects in 3D volumes, the number of methods addressing volumetric object detection is much smaller compared to 2D, and only a few methods are readily available for volumes. So, to establish baselines for detection under context shift, we have selected and adapted three detection approaches. They vary in how they utilize training data; relying on either ground truth segmentation labels, ground truth bounding boxes, or both. + +Detection U-Net Inspired by [4] we use U-Net [8] as one baseline detection model. U-Net is a dominating model for volumetric segmentation, making it a natural choice for adapting into a baseline detection model. Detection U-Net is trained on volumes accompanied by ground truth segmentation labels. It is first trained to predict the segmentation of the single bugs. Then, starting from the obtained weights, it is trained to predict the segmentation of the synthetic mixtures. Finally, again starting from the previous weights, it is trained to predict the segmentation in the crowded synthetic mixtures. For all training runs, we augment the images with random rotations, translations, geometric scaling, and intensity scaling. We train the models using the sum of Dice and cross entropy loss. During training we hold out a validation split which we evaluate on after every epoch. At the end of training, the model which scored the highest on the validation split is kept. Like [4], we extract bounding boxes by thresholding the softmax probabilities (at 0.75) and then extracting the connected components for each channel. Only components larger than $125(5^{3})$ voxels are kept. We use the U-Net implementation from MONAI [5]. + +Faster R-CNN As an example of region proposal networks, we have chosen Faster R-CNN [44] with the modifications from [31]. We have adapted the implementation from torchvision [54] to allow for 3D detection. We train in the same sequence as the U-Net: single bugs, synthetic mixtures, and crowded synthetic mixtures. We use the same augmentations as for the U-Net. In addition, we also randomly perturb the bounding box corners. The model is only trained on bounding boxes and does not make use of the voxel labels. The model is + +trained with smooth L1 loss for bounding box regression and cross-entropy for objectness and class predictions. For full details, we refer to [31, 44]. As with U-Net, we hold out a validation split which is evaluated after each epoch. After training, the model scoring the highest on the validation split is kept. For inference, only bounding boxes with an objectness score over 0.05 are kept. + +nnDetection The third detection model, nnDetection, trains the RetinaUNet [22] detection model using a combination of segmentation masks and bounding boxes. The bounding box branch is trained using a generalized IoU loss for regression and cross-entropy for classification. The voxel segmentation branch is trained with a sum of Dice and cross-entropy loss. As nnDetection is a self-configuring framework, we refer to [4, 21] for further training and model details. We simply provide the data as instructed. We train a separate model for single bugs, synthetic mixtures, and crowded synthetic mixtures. nnDetection trains using 5-fold cross-validation, and the final predictions are made using an ensemble over the best model from each fold [4]. + +# 4.3 Evaluation + +We assess model performance by evaluating their detection accuracy on real scans of densely packed bugs, using manually annotated center points. These annotations are matched with predicted bounding boxe centers based on Euclidean distance, optimized using Munkres' algorithm [38]. Matches are only accepted if the annotated point falls within a $1.5 \times$ inflated bounding box, as illustrated in Fig. 4. This inflation ensures a fair assessment for detections like in Fig. 4(b). This, combined with matching on the bounding box centers, helps ensure that large boxes are not favored since it places more weight on the location of the boxes. We compute precision to avoid favoring overdetections, recall to measure the proportion of correct matches, and F1-score to combine the two. Further, we first compute the metrics purely from the point matches without considering classes — thereby evaluating if the models can detect the presence of bugs vs. background material. Next, we recompute the matches where we only allow matches between points of equal class. + +# 5 Results + +The results are summarized in Tab. 1, example results are shown in Fig. 5, and the distributions of the metrics without class information are shown in Fig. 6. U-Net performs relatively well, independent of what it was trained on, whereas Faster R-CNN performs worse on crowded mixtures and nnDetection on single bugs. If class labels are required, the performance is generally low. + +The distributions in Fig. 6 show that U-Net has a narrower spread around the mean for the F1-score, whereas the two others are spread more out. Wrt. the precision, many detections obtained by the nnDetection method are 1, but have a lower recall, meaning that those that it predicts are correct, but there + +![](images/54b7172581217d9d6bb33f28eb39c37bfe6bc886dea38ea23106e9fc679553c7.jpg) +(a) + +![](images/842f6f2ecd4543c290ad5358ab452b07011e497f1c29141f66909bfe570d35d3.jpg) +(b) +Fig. 4: Matching criterion for bounding box to center point. The circle shows the manually annotated center point, the predicted bounding box is shown with the full line, and the inflated bounding box is shown with the dashed line. In (a), the match is accepted by the predicted bounding box, in (b), the match is also accepted due to the inflated bounding box, whereas in (c), there is no match. Note that this illustration is 2D, but the detection is done in 3D. + +![](images/8d3a276c11a1f9dcfc671c18397288ad8af552f9e342ffb804e5cdb1ee37775a.jpg) +(c) + +are many potential detections that it misses. It is the opposite for the Faster R-CNN, where many datasets get a recall of 1, but a lower precision, meaning that it detects many of the bugs, but also has many false detections. + +To investigate the capabilities of the three methods without the domain shift between the synthetic and real mixtures, we included the results in Tab. 2. This table shows the results of the validation data for training the models on the volumes of individual bugs and the two synthetic mixtures with the class labels. Since we have the bounding boxes for these datasets, we compute the mean average precision (mAP), similar to, e.g., COCO detection [34]. We compute mAP with the IoU threshold running from 0.5 to 0.95 (shown as mAP) and fixed at 0.5 (mAP@50). We note that the values are not directly comparable with those in Tab. 1. We show example detections in the supplementary. The U-Net generally has high performance, Faster R-CNN has high performance for single bugs but has difficulty with the mixtures, whereas nnDetection performs better in the mixtures than for the individual bugs. Generally, the detection performance of bugs is far from perfect and can be improved, also when training and testing within the same domain. + +# 6 Discussion + +The domain shift introduced by shifting the context surrounding the bugs significantly increases the difficulty of the analysis problem. While U-Net and nnDetection achieve high performance on their training tasks (at least for synthetic mixtures), they struggle when transferred to real densely packed scans. It is also interesting that the domain shift has such an effect on the ability of the models to infer the correct class even when accurately localizing the bug. This is surprising since detecting the bugs in the training domain can be done with high accuracy, as shown in Tab. 2, and pure classification (see supplementary) gave an accuracy of 0.97. Despite the models for classification being different than the detection + +Table 1: Detection results on mixed scans. We show mean and std. for each metric computed over the evaluation images. Higher is better. Without class info refers to not considering the bug label and with class info expects the bug type to be classified as well as being detected. + +
Model- Trained onWithout class info.With class info.
F1-ScorePrecisionRecallF1-ScorePrecisionRecall
U-Net [8]- Single bugs0.59±0.180.68±0.170.60±0.250.11±0.160.13±0.190.11±0.16
- Synth. mixes0.60±0.120.52±0.170.82±0.190.11±0.080.09±0.080.14±0.10
- Crowded s.m.0.54±0.130.39±0.140.97±0.070.10±0.080.07±0.060.18±0.14
Faster R-CNN [44]- Single bugs0.56±0.201.00±0.040.42±0.210.11±0.100.21±0.190.08±0.08
- Synth. mixes0.68±0.160.63±0.210.83±0.220.16±0.140.15±0.140.19±0.18
- Crowded s.m.0.27±0.170.81±0.240.18±0.160.06±0.090.19±0.300.04±0.06
nn-Detection [4]- Single bugs0.03±0.020.02±0.010.91±0.150.00±0.010.00±0.000.09±0.16
- Synth. mixes0.50±0.140.41±0.170.80±0.220.10±0.060.08±0.060.16±0.11
- Crowded s.m.0.60±0.190.77±0.230.58±0.270.11±0.120.15±0.160.11±0.12
+ +Table 2: Detection results on the training tasks obtained during training on the single bugs and the two synthetic mixtures. For U-Net and Faster R-CNN performance is evaluated on a validation split. Since nnDetection is an ensemble created from multiple train-validation splits, we evaluate on all images in the training task. We show mean average precision (mAP) between predicted and ground truth bounding boxes. Higher is better. + +
Model- Trained onmAPmAP@50
U-Net [8]- Single bugs0.720.72
- Synthetic mixes0.640.77
- Crowded synth. mixes0.750.86
Faster R-CNN [44]- Single bugs0.700.70
- Synthetic mixes0.020.02
- Crowded synth. mixes0.050.05
nnDetection [4]- Single bugs0.100.10
- Synthetic mixes0.590.70
- Crowded synth. mixes0.490.66
+ +models, these results show that even when the objects do not change appearance, the shift in the domain by changing the context surrounding the objects is sufficient for deep learning-based detection models to lose performance. + +In preparing the data, we have strived to get the scans of individual bugs and bugs in mixtures to have the same appearance. Since bugs are flexible, there will be differences in the pose of the bugs in mixtures and individual bugs. Larvae will sometimes bend differently and bugs with antennae or legs will bend differently. Also, the cotton is visible and can slightly influence the appearance. Despite these differences, the voxel intensity distributions in individual scanned bugs and mixtures are the same, and therefore the BugNIST dataset is a good basis for developing detection methods that can handle context shift. If the mixtures were (partly) labeled for segmentation, BugNIST could also be used to investigate segmentation algorithms under context shift. + +![](images/eb1ac95eecc14df2f0bc40673c32812a94a2999f3c09b7d423ad367754d0275d.jpg) +Fig. 5: Detection results on real mixed scans. The images are maximum-intensity projections of the volumes. Center points are marked with circles, detections are marked with bounding boxes, and color indicates the bug class. A line from a point to the upper left corner of a bounding box shows a match. If the center point and the bounding box have the same color, then the detection and classification are correct. The shown images are the volumes with the highest scores for the different models. For U-Net the results when trained on synthetic mixes, for Faster R-CNN when trained on synthetic mixes and nnDetection when trained on crowded synthetic mixes. + +![](images/01c91c05303a061e584404fa4aebd0b5e77a0f9d8179303ed60136159284688b.jpg) +Fig. 6: Distributions of performance scores for the three detection methods. + +We chose to annotate only the center positions of the bugs. As the bugs are often elongated and at odd angles, axis-aligned bounding boxes would contain significant empty space and overlap between neighboring boxes—thus offering limited benefit over points. This could be solved with object aligned bounding boxes or full voxel annotations. However, this would drastically increase the annotation time. We chose to prioritize having a large number of annotations such that the test set can have a large variety of configurations. Furthermore, given that current methods show ample room for improvement (cf. Tab. 1), we + +find that point annotations combined with the chosen metric are sufficient to measure method performance and improvements of future approaches. + +The methods and experiments are chosen as a basic baseline to illustrate the properties of the BugNIST dataset. From the experiments, we learned that data augmentation by composing synthetic mixtures is a feasible approach for improving detection performance. More investigations into packing algorithms and ways to combine the synthetic mixed bugs with other backgrounds can potentially improve performance further. But it is also clear that there is great potential for developing new methods for object detection in volumes. We chose Detection U-Net, nnDetection, and our adaption of Faster R-CNN to 3D volumes because these methods were available, widely used, and are based on complementary ideas. We did not find more recent alternatives for volumetric analysis. + +Our results show that Faster R-CNN performs less than the two other methods. While it scores highest in Tab. 1, it is clear from Fig. 5 that the bounding boxes are of much lower quality. Furthermore, it does not score well on the training tasks, except when trained on individual bugs. To ensure the poor performance of Faster R-CNN was not just the result of too few training volumes, we redid the experiments with 1000 synthetic mixes instead of the original 300. Performance remained the same. It is also noteworthy that nnDetection scores overall lower than U-Net on both the training and evaluation tasks, even though it also has access to the segmentation masks for training. + +# 7 Conclusion + +With our BugNIST benchmark dataset for domain shift in volumetric data, we demonstrate how detecting objects, that have the same appearance in the source and target domain, fails when the objects in the two domains are placed in different contexts. By contexts, we refer to the appearance of the image surrounding the objects. To our knowledge, the BugNIST dataset is the only dataset designed for developing methods for context shift, i.e. domain shift where the objects stay the same but the surroundings change. Solutions will enable an automated annotation of data for volumetric object detection, which is an area that has not received much attention. There are, however, a range of applications where object detection is needed, including medical imaging and emerging areas such as micro-CT scanning of materials. The BugNIST dataset can also be used for investigating other problems such as segmentation under domain shift provided reference annotations are created. + +# Acknowledgments + +We acknowledge the 3D Imaging Center at DTU, the Infrastructure for Quantitative AI-based Tomography (QUAITOM), supported by the Novo Nordisk Foundation (grant number NNF21OC0069766) and STUDIOS: Segmenting Tomograms Using Different Interpretation of Simplicity funded by the Villum Foundation (grant number VIL50425). + +# References + +1. Armato, S.G., Huisman, H., Drukker, K., Hadjiiski, L., Kirby, J.S., Petrick, N., Redmond, G., Giger, M.L., Cha, K., Mamonov, A., et al.: PROSTATEx challenges for computerized classification of prostate lesions from multiparametric magnetic resonance images. Journal of Medical Imaging 5(4), 044501-044501 (2018) +2. Armato, S.G., McLennan, G., Bidaut, L., McNitt-Gray, M.F., Meyer, C.R., Reeves, A.P., Zhao, B., Aberle, D.R., Henschke, C.I., Hoffman, E.A., et al.: The lung image database consortium (LIDC) and image database resource initiative (IDRI): a completed reference database of lung nodules on CT scans. Medical Physics 38(2), 915–931 (2011) +3. Baid, U., Ghodasara, S., Mohan, S., Bilello, M., Calabrese, E., Colak, E., Farahani, K., Kalpathy-Cramer, J., Kitamura, F.C., Pati, S., et al.: The RSNA-ASNR-MICCAI BraTS 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314 (2021) +4. Baumgartner, M., Jäger, P.F., Isensee, F., Maier-Hein, K.H.: nnDetection: a self-configuring method for medical object detection. In: Proceedings of the Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 530-539. Springer (2021) +5. Cardoso, M.J., Li, W., Brown, R., Ma, N., Kerfoot, E., Wang, Y., Murrey, B., Myronenko, A., Zhao, C., Yang, D., et al.: MONAI: An open-source framework for deep learning in healthcare. arXiv preprint arXiv:2211.02701 (2022) +6. Charytanowicz, M., Kulczycki, P., Kowalski, P.A., Lukasik, S., Czabak-Garbacz, R.: An evaluation of utilizing geometric features for wheat grain classification using X-ray images. Computers and Electronics in Agriculture 144, 260-268 (2018) +7. Chaudhary, S., Sadbhawna, S., Jakhetiya, V., Subudhi, B.N., Baid, U., Guntuku, S.C.: Detecting COVID-19 and community acquired pneumonia using chest CT scan images with deep learning. In: Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 8583-8587. IEEE (2021) +8. Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3D U-Net: learning dense volumetric segmentation from sparse annotation. In: Proceedings of the Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 424–432. Springer (2016) +9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3213-3223 (2016) +0. De Carlo, F., Gürsoy, D., Ching, D.J., Batenburg, K.J., Ludwig, W., Mancini, L., Marone, F., Mokso, R., Pelt, D.M., Sijbers, J., et al.: Tomobank: a tomographic data repository for computational x-ray science. Measurement Science and Technology 29(3), 034004 (2018) +1. Dorent, R., Kujawa, A., Ivory, M., Bakas, S., Rieke, N., Joutard, S., Glocker, B., Cardoso, J., Modat, M., Batmanghelich, K., et al.: Crossmoda 2021 challenge: Benchmark of cross-modality domain adaptation techniques for vestibular schwannoma and cochlea segmentation. Medical Image Analysis 83, 102628 (2023) +2. Du, Z., Hu, Y., Ali Buttar, N., Mahmood, A.: X-ray computed tomography for quality inspection of agricultural products: A review. Food Science & Nutrition 7(10), 3146-3160 (2019) + +13. Einarsdóttir, H., Emerson, M.J., Clemmensen, L.H., Scherer, K., Willer, K., Bech, M., Larsen, R., Ersbøll, B.K., Pfeiffer, F.: Novelty detection of foreign objects in food using multi-modal X-ray imaging. Food Control 67, 39-47 (2016) +14. European Organization For Nuclear Research, OpenAIRE: Zenodo (2013). https://doi.org/10.25495/7GXK-RD71, https://www.zenodo.org/ +15. Flitton, G., Mouton, A., Breckon, T.P.: Object classification in 3D baggage security computed tomography imagery using visual codebooks. Pattern Recognition 48(8), 2489-2499 (2015) +16. Girshick, R.: Fast R-CNN. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 1440-1448 (2015) +17. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object detection and semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 580-587 (2014) +18. Guan, H., Liu, M.: Domain adaptation for medical image analysis: a survey. IEEE Transactions on Biomedical Engineering 69(3), 1173-1185 (2021) +19. Harmon, S.A., Sanford, T.H., Xu, S., Turkbey, E.B., Roth, H., Xu, Z., Yang, D., Myronenko, A., Anderson, V., Amalou, A., et al.: Artificial intelligence for the detection of COVID-19 pneumonia on chest CT using multinational datasets. Nature Communications 11(1), 4080 (2020) +20. Hipsley, C.A., Aguilar, R., Black, J.R., Hocknull, S.A.: High-throughput microCT scanning of small specimens: preparation, packing, parameters and post-processing. Scientific Reports 10(1), 13863 (2020) +21. Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods 18(2), 203-211 (2021) +22. Jaeger, P.F., Kohl, S.A., Bickelhaupt, S., Isensee, F., Kuder, T.A., Schlemmer, H.P., Maier-Hein, K.H.: Retina U-Net: Embarrassingly simple exploitation of segmentation supervision for medical object detection. In: Machine Learning for Health Workshop. pp. 171-183. PMLR (2020) +23. Jain, Y., Borner, K., Walsh, C., Ruschman, N., Lee, P.D., Weber, G.M., Holbrook, R., Howard, A.: SenNet + HOA - hacking the human vasculature in 3D (2023), https://kaggle.com/competitions/blood-vessel-segmentation +24. Jarolmasjed, S., Espinoza, C.Z., Sankaran, S., Khot, L.R.: Postharvest bitter pit detection and progression evaluation in 'Honeycrisp' apples using computed tomography images. Postharvest Biology and Technology 118, 35-42 (2016) +25. Jeppesen, N., Christensen, A.N., Dahl, V.A., Dahl, A.B.: Sparse layered graphs for multi-object segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12777-12785 (2020) +26. Jin, L., Yang, J., Kuang, K., Ni, B., Gao, Y., Sun, Y., Gao, P., Ma, W., Tan, M., Kang, H., et al.: Deep-learning-assisted detection and segmentation of rib fractures from CT scans: Development and validation of FracNet. EBioMedicine 62, 103106 (2020) +27. Ker, J., Singh, S.P., Bai, Y., Rao, J., Lim, T., Wang, L.: Image thresholding improves 3-dimensional convolutional neural network diagnosis of different acute brain hemorrhages on computed tomography scans. Sensors 19(9), 2167 (2019) +28. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (2023) + +29. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998) +30. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.M.: Deeper, broader and artier domain generalization. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 5542-5550 (2017) +31. Li, Y., Xie, S., Chen, X., Dollar, P., He, K., Girshick, R.: Benchmarking detection transfer learning with vision transformers. arXiv preprint arXiv:2111.11429 (2021) +32. Li, Y., Fan, Y.: Medical image segmentation with domain adaptation: A survey. arXiv preprint arXiv:2311.01702 (2023) +33. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 2980-2988 (2017) +34. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 740-755. Springer (2014) +35. Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature Communications 15(1), 654 (2024) +36. Maier-Hein, L., Eisenmann, M., Reinke, A., Onogur, S., Stankovic, M., Scholz, P., Arbel, T., Bogunovic, H., Bradley, A.P., Carass, A., et al.: Why rankings of biomedical image analysis competitions should be interpreted with care. Nature Communications 9(1), 5217 (2018) +37. Muandet, K., Balduzzi, D., Scholkopf, B.: Domain generalization via invariant feature representation. In: International Conference on Machine Learning. pp. 10-18. PMLR (2013) +38. Munkres, J.: Algorithms for the assignment and transportation problems. Journal of the Society for Industrial and Applied Mathematics 5(1), 32-38 (1957) +39. OECD: Health at a Glance 2023 (2023). https://doi.org//10.1787/7a7afb35-en +40. Pan, S.J., Tsang, I.W., Kwok, J.T., Yang, Q.: Domain adaptation via transfer component analysis. IEEE Transactions on Neural Networks 22(2), 199-210 (2010) +41. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1406-1415 (2019) +42. Rawson, S.D., Maksimcuka, J., Withers, P.J., Cartmell, S.H.: X-ray computed tomography in life sciences. BMC biology 18(1), 1-15 (2020) +43. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 779-788 (2016) +44. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. Advances in neural information processing systems (NeurIPS) 28 (2015) +45. Richter, S.R., Vineet, V., Roth, S., Koltun, V.: Playing for data: Ground truth from computer games. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 102-118. Springer (2016) +46. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: Proceedings of the Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 234-241. Springer (2015) +47. Ros, G., Sellart, L., Materzynska, J., Vazquez, D., Lopez, A.M.: The synthia dataset: A large collection of synthetic images for semantic segmentation of urban + +scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3234-3243 (2016) +48. Sankaranarayanan, S., Balaji, Y., Jain, A., Lim, S.N., Chellappa, R.: Learning from synthetic data: Addressing domain shift for semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3752-3761 (2018) +49. Serte, S., Demirel, H.: Deep learning for diagnosis of COVID-19 using 3D CT scans. Computers in biology and medicine 132, 104306 (2021) +50. Setio, A.A.A., Traverso, A., De Bel, T., Berens, M.S., Van Den Bogaard, C., Cerello, P., Chen, H., Dou, Q., Fantacci, M.E., Geurts, B., et al.: Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: the luna16 challenge. Medical Image Analysis 42, 1-13 (2017) +51. Simpson, A.L., Antonelli, M., Bakas, S., Bilello, M., Farahani, K., van Ginneken, B., Kopp-Schneider, A., Landman, B.A., Litjens, G., Menze, B.H., Ronneberger, O., Summers, R.M., Bilic, P., Christ, P.F., Do, R.K.G., Gollub, M., Golia-Pernicka, J., Heckers, S., Jarnagin, W.R., McHugo, M., Napel, S., Vorontsov, E., Maier-Hein, L., Cardoso, M.J.: A large annotated medical image dataset for the development and evaluation of segmentation algorithms. CoRR abs/1902.09063 (2019), http://arxiv.org/abs/1902.09063 +52. Sun, T., Segu, M., Postels, J., Wang, Y., Van Gool, L., Schiele, B., Tombari, F., Yu, F.: Shift: a synthetic driving dataset for continuous multi-task domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21371-21382 (2022) +53. Timmins, K.M., van der Schaaf, I.C., Bennink, E., Ruigrok, Y.M., An, X., Baumgartner, M., Bourdon, P., De Feo, R., Di Noto, T., Dubost, F., et al.: Comparing methods of detecting and segmenting unruptured intracranial aneurysms on TOF-MRAS: The ADAM challenge. Neuroimage 238, 118216 (2021) +54. TorchVision maintainers and contributors: Torchvision: Pytorch's computer vision library. https://github.com/pytorch/vision (2016) +55. Trusler, M.M., Sturrock, C.J., Vane, C.H., Cook, S., Lomax, B.H.: X-ray computed tomography: A novel non-invasive approach for the detection of microplastics in sediments? Marine Pollution Bulletin 194, 115350 (2023) +56. Velayudhan, D., Hassan, T., Damiani, E., Werghi, N.: Recent advances in baggage threat detection: A comprehensive and systematic survey. ACM Computing Surveys 55(8) (dec 2022). https://doi.org/10.1145/3549932, https://doi.org/10.1145/3549932 +57. Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5018-5027 (2017) +58. Walsh, C., Tafforeau, P., Wagner, W., Jafree, D., Bellier, A., Werlein, C., Kühnel, M., Boller, E., Walker-Samuel, S., Robertus, J., et al.: Imaging intact human organs with local resolution of cellular structures using hierarchical phase-contrast tomography. Nature Methods 18(12), 1532–1541 (2021) +59. Wang, M., Deng, W.: Deep visual domain adaptation: A survey. Neurocomputing 312, 135-153 (2018) +60. Wang, M., Liu, Y., Yuan, J., Wang, S., Wang, Z., Wang, W.: Inter-class and inter-domain semantic augmentation for domain generalization. IEEE Transactions on Image Processing (2024) + +61. Wang, Q., Bhowmik, N., Breckon, T.P.: Multi-class 3D object detection within volumetric 3D computed tomography baggage security screening imagery. In: Proceedings of the IEEE International Conference on Machine Learning and Applications (ICMLA). pp. 13-18. IEEE (2020) +62. Wang, Q., Breckon, T.P.: Contraband materials detection within volumetric 3D computed tomography baggage security screening imagery. In: Proceedings of the IEEE International Conference on Machine Learning and Applications (ICMLA). pp. 75-82. IEEE (2021) +63. Withers, P.J., Bouman, C., Carmignato, S., Cnudde, V., Grimaldi, D., Hagen, C.K., Maire, E., Manley, M., Du Plessis, A., Stock, S.R.: X-ray computed tomography. Nature Reviews Methods Primers 1(1), 18 (2021) +64. Wu, D., Wu, D., Feng, H., Duan, L., Dai, G., Liu, X., Wang, K., Yang, P., Chen, G., Gay, A.P., et al.: A deep learning-integrated micro-ct image analysis pipeline for quantifying rice lodging resistance-related traits. Plant Communications 2(2) (2021) +65. Xiao, Z., Liu, B., Geng, L., Zhang, F., Liu, Y.: Segmentation of lung nodules using improved 3d-unet neural network. Symmetry 12(11), 1787 (2020) +66. Xie, Y., Ji, Q.: A new efficient ellipse detection method. In: Proceedings of the International Conference on Pattern Recognition (ICPR). vol. 2, pp. 957-960. IEEE (2002) +67. Xu, F., Tenenbaum, J.B.: Word learning as bayesian inference. Psychological review 114(2), 245 (2007) +68. Yan, Y.T., Chua, S., DeCarlo, T.M., Kempf, P., Morgan, K.M., Switzer, A.D.: Core-ct: A matlab application for the quantitative analysis of sediment and coral cores from X-ray computed tomography (ct). Computers & Geosciences 156, 104871 (2021) +69. Zhao, W.X., Zhou, K., Li, J., Tang, T., Wang, X., Hou, Y., Min, Y., Zhang, B., Zhang, J., Dong, Z., et al.: A survey of large language models. arXiv preprint arXiv:2303.18223 (2023) +70. Zhou, K., Yang, Y., Hospedales, T., Xiang, T.: Deep domain-adversarial image generation for domain generalisation. Proceedings of the AAAI Conference on Artificial Intelligence 34(07), 13025-13032 (April 2020). https://doi.org/10.1609/aaii.v34i07.7003, https://ojs.aaii.org/index.php/AAAI/article/view/7003 +71. Zhou, S.K., Greenspan, H., Davatzikos, C., Duncan, J.S., Van Ginneken, B., Madabhushi, A., Prince, J.L., Rueckert, D., Summers, R.M.: A review of deep learning in medical imaging: Imaging traits, technology trends, case studies with progress highlights, and future promises. Proceedings of the IEEE 109(5), 820-838 (2021) \ No newline at end of file diff --git a/2024/BugNIST - a Large Volumetric Dataset for Detection under Domain Shift/images.zip b/2024/BugNIST - a Large Volumetric Dataset for Detection under Domain Shift/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f8ca13824a4783abfa32f361f4917073d3e5ff5d --- /dev/null +++ b/2024/BugNIST - a Large Volumetric Dataset for Detection under Domain Shift/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:269b0a75ca2519165e6a7c26b540725187eb4ba8b31f0e37343674b963803774 +size 409029 diff --git a/2024/BugNIST - a Large Volumetric Dataset for Detection under Domain Shift/layout.json b/2024/BugNIST - a Large Volumetric Dataset for Detection under Domain Shift/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..97f6e00e1659c220883784edb4011985aa2c529f --- /dev/null +++ b/2024/BugNIST - a Large Volumetric Dataset for Detection under Domain Shift/layout.json @@ -0,0 +1,7804 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "spans": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "type": "text", + "content": "BugNIST - a Large Volumetric Dataset for Object Detection under Domain Shift" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 150, + 168, + 462, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 168, + 462, + 194 + ], + "spans": [ + { + "bbox": [ + 150, + 168, + 462, + 194 + ], + "type": "text", + "content": "Patrick Møller Jensen, Vedrana Andersen Dahl, Rebecca Engberg, Carsten Gundlach, Hans Marin Kjer, and Anders Bjorholm Dahl" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 157, + 202, + 457, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 202, + 457, + 225 + ], + "spans": [ + { + "bbox": [ + 157, + 202, + 457, + 225 + ], + "type": "text", + "content": "Technical University of Denmark, Kgs. Lyngby, Denmark {patmjen, vand, reen}@dtu.dk, cagu@fysik.dtu.dk, {hmjk, abda}@dtu.dk" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 256, + 453, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 256, + 453, + 421 + ], + "spans": [ + { + "bbox": [ + 160, + 256, + 453, + 421 + ], + "type": "text", + "content": "Abstract. Domain shift significantly influences the performance of deep learning algorithms, particularly for object detection within volumetric 3D images. Annotated training data is essential for deep learning-based object detection. However, annotating densely packed objects is time-consuming and costly. Instead, we suggest training models on individually scanned objects, causing a domain shift between training and detection data. To address this challenge, we introduce the BugNIST dataset, comprising 9154 micro-CT volumes of 12 bug types and 388 volumes of tightly packed bug mixtures. This dataset is characterized by having objects with the same appearance in the source and target domains, which is uncommon for other benchmark datasets for domain shift. During training, individual bug volumes labeled by class are utilized, while testing employs mixtures with center point annotations and bug type labels. Together with the dataset, we provide a baseline detection analysis, with the aim of advancing the field of 3D object detection methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 432, + 453, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 432, + 453, + 455 + ], + "spans": [ + { + "bbox": [ + 160, + 432, + 453, + 455 + ], + "type": "text", + "content": "Keywords: Volumetric Dataset, Benchmark, Volumetric Object Detection, Domain Shift." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 495, + 230, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 495, + 230, + 508 + ], + "spans": [ + { + "bbox": [ + 133, + 495, + 230, + 508 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Our work on domain shift in volumetric 3D images is motivated by the need for labeled data to train supervised deep learning models for volumetric imaging. The problem that we aim to solve is object detection and classification. We propose to label images of objects scanned as isolated entities as the source domain and use these as a basis to train models for object detection and classification in a complex context of mixed objects and other materials as the target domain. The effort needed for obtaining labels in the two domains is significantly different. If objects are isolated, they can be automatically labeled whereas mixed objects require expensive manual labeling. Automating the labeling based on isolated objects, however, leads to a domain shift between the data in the source domain for training and the data in the target domain for detection and classification. This domain shift is special because the appearance of the objects is the same" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 291, + 219 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 291, + 219 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 291, + 219 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 291, + 219 + ], + "type": "image", + "image_path": "ba200f363445f099ca69ad44d26396a72c9d667e3c50bc5e33df3352a7dc9d47.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 224, + 280, + 241 + ], + "lines": [ + { + "bbox": [ + 147, + 224, + 280, + 241 + ], + "spans": [ + { + "bbox": [ + 147, + 224, + 280, + 241 + ], + "type": "text", + "content": "(a) 9154 volumes of individual bugs with class labels." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 294, + 114, + 484, + 218 + ], + "blocks": [ + { + "bbox": [ + 294, + 114, + 484, + 218 + ], + "lines": [ + { + "bbox": [ + 294, + 114, + 484, + 218 + ], + "spans": [ + { + "bbox": [ + 294, + 114, + 484, + 218 + ], + "type": "image", + "image_path": "265b46fe7ce8a15fda6f80c106c3407d9da56e3764376b9caa07cd0a91feea43.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 223, + 468, + 241 + ], + "lines": [ + { + "bbox": [ + 310, + 223, + 468, + 241 + ], + "spans": [ + { + "bbox": [ + 310, + 223, + 468, + 241 + ], + "type": "text", + "content": "(b) 388 volumes of bug mixtures with center point and class labels." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 262, + 482, + 307 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 307 + ], + "type": "text", + "content": "Fig. 1: Overview of the two parts of the dataset. In (a), representative individual bug volumes are displayed with a class abbreviation, name, and number of volumes. In (b), a few volumes of bug mixtures are shown. Each bug is annotated with bug class and center point." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 332, + 480, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 480, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 480, + 355 + ], + "type": "text", + "content": "in the two domains, but the surrounding context is different. We use the term context shift for this type of domain shift." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 357, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 464 + ], + "type": "text", + "content": "To promote the development of new deep learning-based methods that can handle context shift, we contribute the volumetric dataset called BugNIST1. The dataset contains 9154 micro-CT volumes with individual bugs and 388 micro-CT volumes containing a mixture of several bugs as well as other materials such as leaves, wood shavings, etc.; see Fig. 1. The challenge is to use the individually scanned bugs to detect and classify the bugs in the mixtures. The dataset is intended to be a relatively simple test case for starting to develop methods for detecting objects in volumetric images under context shift, and bugs are chosen due to their complex shape and not for solving a problem in entomology." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 465, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 465, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 465, + 482, + 571 + ], + "type": "text", + "content": "As a basis for training an accurate deep learning model, a training dataset for object detection and classification must capture the appearance distribution of the object classes that are in the dataset. Since the appearance of a CT-scanned object changes minimally if it is scanned individually or as part of a mixture, the appearance distribution can be covered by only considering the individual scans. This is different from 2D photos. Here additional variations in scale, viewpoint, and illumination influence the appearance of objects, and it would be more difficult to acquire representative images of individual objects for a dataset based on 2D photos." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 573, + 482, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 573, + 482, + 644 + ], + "spans": [ + { + "bbox": [ + 130, + 573, + 482, + 644 + ], + "type": "text", + "content": "Since BugNIST is a test case for context shift, we wanted to ensure that each class is covered well. By scanning systematically arranged bugs, we have generated 9154 volumes of individual bugs, each labeled with its corresponding class. Furthermore, we scanned 388 mixed volumes, each containing between 4 and 45 bugs alongside other materials. For evaluation purposes, each bug in the mixtures is manually annotated with its center position and class label. In total," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "text", + "content": "P.M. Jensen et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 652, + 301, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 652, + 301, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 652, + 301, + 666 + ], + "type": "text", + "content": "1 https://abdahl.github.io/bugnist/" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "9542 volumes make up the BugNIST dataset. The numbers in each class are shown in Fig. 1. Since the bugs in each class are the same age, we expect to cover the biological variation sufficiently to be representative of the class." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 154, + 480, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 154, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 154, + 480, + 322 + ], + "type": "text", + "content": "Solutions to the type of context shift, that we present here, namely a domain shift where only the context changes, and not the appearance of the objects, have wide applications in volumetric imaging. The field of volumetric imaging is dominated by medical applications [39], but there is a growing use beyond the medical area based on e.g. micro-CT [58, 63]. In all these application areas, there is a need to employ supervised deep learning to solve problems that require quantification of depicted structures. When the use case is novel and there is no directly applicable training data available — a typical scenario in volumetric imaging of e.g. materials microstructure, but also in specialized biomedical applications — it is necessary to create training data. If labor-intensive manual labeling can be limited or avoided, there is potentially a huge gain in productivity. Examples of applications are detecting and classifying objects in scans of marine sediments such as microplastics [55, 68], museum specimens [20], or detecting threat items in baggage security screening [56]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 324, + 481, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 481, + 407 + ], + "type": "text", + "content": "In volumetric scans, manual labeling is significantly more demanding than in 2D. For instance, annotating a bounding box of a single object in 3D requires traversing slices of the volume and making a mental model of the object. If the volume contains densely packed complex objects, determining their bounding boxes becomes difficult. Further, bounding boxes may overlap to such an extent that they stop being a useful way to locate objects. This is why we employ center points as the annotation format in the mixtures." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 410, + 481, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 481, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 481, + 529 + ], + "type": "text", + "content": "The name BugNIST is a homage to the influential MNIST dataset [29] which has played a pivotal role in the development and evaluation of deep learning methods. Similar to the way that MNIST has been used, BugNIST is intended to be a benchmark dataset that represents a well-defined problem, where the development of solutions will make a basis for solutions that can advance the field. Specifically for BugNIST, solutions will enable object detection and classification with significantly less effort put into creating the data foundation for training. This calls for methods that can learn to predict objects in a complex context from volumes of individual objects, namely solving the context shift problem. To our knowledge, no existing datasets have this property." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 532, + 481, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 532, + 481, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 532, + 481, + 615 + ], + "type": "text", + "content": "The term context shift refers to a shift in the data distribution between training and inference. Typical shifts in data distributions arise from changes in data acquisition systems, e.g. the use of different scanners, or from variations in the objects of interest between datasets. Such problems have been addressed using domain adaptation [18,59]. In our setting, there is almost no change in the appearance of the bugs between volumes of individual bugs and mixtures, but the change is in what surrounds the bugs, i.e. the change is in the context." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "We hypothesize that methods capable of learning under context shift might also prove capable of learning from very few examples, similar to how humans learn [67]. Instead of relying on seeing objects in every imaginable context, such methods could instead isolate objects from their contexts. If this hypothesis holds" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 244, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 244, + 91, + 447, + 102 + ], + "type": "text", + "content": "BugNIST - Object Detection under Domain Shift" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "true, learning under context shift may reduce the need for extensive training data and the associated training costs. This is in contrast to the current trend, where larger models and datasets as seen in LLMs [69] and computer vision [28, 35], lead to better performance, but also limit who can effectively train such models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 173, + 480, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 173, + 480, + 232 + ], + "spans": [ + { + "bbox": [ + 130, + 173, + 480, + 232 + ], + "type": "text", + "content": "Contributions We contribute BugNIST, a new volumetric image dataset for benchmarking domain shift. In this paper, we use BugNIST for benchmarking detection methods under context shift, trained only on individual bugs and tested on mixtures. We provide several baselines and investigate different training strategies for existing volumetric detection methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 232, + 481, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 232, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 481, + 293 + ], + "type": "text", + "content": "Our results show that volumetric object detection under context shift is difficult. We obtain reasonable results in localizing the bugs, but when a class label is assigned, there is a significant drop in performance. This is even though the classification of individual bugs is easy. Our experiments demonstrate great potential for developing new approaches for 3D object detection of complex objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 308, + 234, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 234, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 234, + 320 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 331, + 479, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 479, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 479, + 365 + ], + "type": "text", + "content": "In computer vision, the term 3D data refers to different data types, including point clouds, surfaces, volumes, and depth images. Here, we are only concerned with 3D volumes, i.e. image data represented on a voxel grid." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 366, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 481, + 498 + ], + "type": "text", + "content": "BugNIST is created to advance deep learning methods for object detection and classification under domain shift. Especially in medical imaging, analysis of 3D volumes is in great demand. A review of medical challenges illustrates this [36]. Their study shows that detection makes up " + }, + { + "bbox": [ + 130, + 366, + 481, + 498 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 130, + 366, + 481, + 498 + ], + "type": "text", + "content": " of the tasks, classification makes up " + }, + { + "bbox": [ + 130, + 366, + 481, + 498 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 366, + 481, + 498 + ], + "type": "text", + "content": " of the tasks, whereas segmentation accounts for " + }, + { + "bbox": [ + 130, + 366, + 481, + 498 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 130, + 366, + 481, + 498 + ], + "type": "text", + "content": " of the tasks. Detection is important for example for diagnosis [7, 19, 27, 49, 71]. Outside the medical domain, object detection problems include baggage security screening [15, 61, 62], although the problem and the available data are still primarily in the 2D domain [56], food inspection [12] such as fruit quality [24], grain classification [6], foreign object detection [13], biological applications such as analysis of weeds [64] and other uses driven by developments in X-ray micro-CT imaging [42], etc." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 593 + ], + "type": "text", + "content": "The domain shift problem presented in BugNIST, where only the context changes, and not the appearance of the objects of interest, is not widely studied. Most research on domain shifts focuses primarily on problems where the appearance of objects changes. Solutions focus on domain adaptation, where labeled data in the source domain are available together with data without labels from the target domain [40], or domain generalization, where the inference domain is not known at training time [37, 60]. We focus on domain adaptation, which is also the typical use case for volumetric imaging." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "Several widely used datasets are available for studying and developing methods for domain shift. For 2D image classification problems the datasets PACS [30], Digits-DG [70], Office-Home [57], and DomainNet [41] developed to classify images from domains such as photos, art paintings, sketches, etc. Domain shift between artificially generated and real images e.g. for segmentation is another area that receives much attention [48] with standard datasets such as" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "P.M. Jensen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": "SYNTHEIA [47], CITYSCAPES [9], and GTA-5 [45], or completely synthetic domain shift like the SHIFT dataset [52] for autonomous driving." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 481, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 481, + 226 + ], + "type": "text", + "content": "In 3D, domain shift is often studied in medical applications where the same type of object is imaged with different scanning modalities [32]. It can be MRI (magnetic resonance imaging), CT, OCT (optical coherence tomography), etc. Standard datasets include the CrossMoDA 2021 dataset [11] and the BRATS dataset [3], both of different MRI modalities. There are many other datasets used for domain adaptation for volumetric medical imaging [32], but all for adapting between modalities or between the same modality but different scanners [18]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 228, + 481, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 481, + 287 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 481, + 287 + ], + "type": "text", + "content": "The structure of the imaged objects in BugNIST is different from most existing volumetric deep learning benchmark datasets. The morphological complexity of bugs is, however, representative of volumetric imaging data of e.g. 3D microstructure, where there is much less use of deep learning than in e.g. the medical domain [63]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 289, + 481, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 481, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 481, + 410 + ], + "type": "text", + "content": "Platforms like Zenodo [14], TomoBank [10], or the Human Organ Atlas [58] also contain a large amount of volumetric image datasets. Different from e.g. medical challenges, this data generally comes with no clear problem formulation, well-defined tasks or even labels, that allow for a comparison of methods. Further, their large volume sizes are difficult to access. It can also be difficult to judge the complexity of the data and if deep learning is needed. Some images can easily be analyzed using a threshold whereas others are complex and can greatly benefit from deep learning. Therefore, most of this data is not attractive to use as a benchmark for developing new deep learning methods. Some datasets are emerging, e.g. the vessel segmentation challenge [23], but they remain exceptions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 411, + 481, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 481, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 481, + 470 + ], + "type": "text", + "content": "There is extensive research in deep learning-based object detection, but primarily in 2D. Most methods either use a two-stage detector such as R-CNN [16, 17, 44], with a region proposal and a classifier to choose among the proposed regions, or a one-stage detector such asYOLO [43], where regions and classes are proposed in a single network." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 472, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 481, + 640 + ], + "type": "text", + "content": "Object detection in volumetric images is less common. One approach has been to apply a 3D U-Net [8] segmentation network and use each connected component of the foreground label class as a detection [22]. An example of using this approach is lung nodule detection in [65]. Employing segmentation for object detection comes with the problem of separating objects that may be connected or combining separated segments into one object. To avoid this problem, Jaeger et al. [22] propose the Retina U-Net: a one-shot 3D detection algorithm that combines the Retina Net [33] and U-Net [46] giving detections with 3D bounding boxes. Despite the good performance of Retina U-Net, changing to a new 3D detection problem requires a time-consuming method configuration. This has been addressed in nnDetection [4]. Inspired by the nnU-Net (no new U-Net) [21], they propose a no new 3D detection method. On the BugNIST data, we test detection with 3D U-Net [8], the nnDetection method [4], and a 3D adaption of the Faster R-CNN model [44]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "Benchmark datasets for volumetric object detection are primarily medical. Examples include Baumgartner et al. [4], which uses 13 datasets to develop the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 244, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 244, + 91, + 447, + 102 + ], + "type": "text", + "content": "BugNIST - Object Detection under Domain Shift" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": "nnDetection method. Two of these datasets were originally targeted 3D detection and include the LUNA16 for nodule detection [50], which is a subset of the LIDC dataset [2] and contains 888 CT scans of lungs, and the ADAM dataset [53] for detecting aneurysms in MRI scans. ADAM is made up of 254 MRI scans. The rest of the detection datasets are medical segmentation datasets such as data from the Medical Segmentation Decathlon [51], the rib fracture detection [26], the PROSTATEx dataset [1], and others. The segmentation datasets have been transformed into detection data by identifying center positions and bounding boxes for the segments. Within baggage security screening, some 3D datasets are obtainable, for example, the ALERT datasets used in [62]. Due to the sensitive nature of the problem, these datasets have copyright restrictions and limited availability under a non-disclosure agreement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 280, + 343, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 280, + 343, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 280, + 343, + 295 + ], + "type": "text", + "content": "3 Data acquisition and processing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 305, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 482, + 402 + ], + "type": "text", + "content": "Fig. 1 gives an overview of the BugNIST dataset. Each individual bug volume contains exactly one bug located near the center of the volume. The total of 9154 individual bug volumes have between 713 and 946 volumes of each class distributed across 12 bug classes (species and stages). Bug mixtures contain several bugs, leaves, cotton, shredded paper, and wood shaving. There are in total 388 mixtures. The mixtures contain 4 to 45 bugs, totaling 5011 bugs with an average of 13 bugs per mixture, see Fig. 2. In bug mixtures, we have manually marked the center positions of each bug and provided it with a class label." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 402, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 485 + ], + "type": "text", + "content": "All scans were acquired using a laboratory micro-CT scanner (Nikon metrology scanner) that was set to record volumes of 2000 voxels cubed with an isotropic voxel size of " + }, + { + "bbox": [ + 130, + 402, + 482, + 485 + ], + "type": "inline_equation", + "content": "26.62\\mu \\mathrm{m}" + }, + { + "bbox": [ + 130, + 402, + 482, + 485 + ], + "type": "text", + "content": ". Being able to acquire such large volumes allowed us to scan multiple bugs and bug mixtures at once, and then crop them into smaller volumes. The dataset was collected through nearly 250 scans, each taking 30 minutes, over 90 days. Care has been taken that the scanning parameters stay unchanged for all scans, and we have calibrated the scanner daily." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 557 + ], + "type": "text", + "content": "All bugs were placed in tubes, either as individuals or in mixtures. We used cotton, which is almost transparent to X-rays, to separate the bugs, either as individual bugs or mixtures. For the mixtures, we varied the difficulty of the bug detection task by adding leaves, shredded paper, or wood shavings to some of the mixtures, and also varied the number of bugs, their class, and the packing density." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "The cylindrical shape of the tubes was easy to segment using a circular Hough-transform [66] and sparse layered graphs [25], leaving us with a cylinder of only bugs and cotton. Due to the low intensity of cotton compared to the bugs, we could easily divide each cylinder into volumes containing one bug each. The volumes have been manually inspected and cleaned to correct cases where something went wrong, e.g. parts of another bug showing in the periphery of the volume. The final volumes of scans of individual bugs are " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "450 \\times 450 \\times 900" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": " voxels, the size which was chosen to accommodate the biggest bugs. The mixtures were segmented similarly and the resulting volumes are " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "650 \\times 650 \\times 900" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": " voxels. All" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "P.M. Jensen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 116, + 289, + 162 + ], + "blocks": [ + { + "bbox": [ + 141, + 116, + 289, + 162 + ], + "lines": [ + { + "bbox": [ + 141, + 116, + 289, + 162 + ], + "spans": [ + { + "bbox": [ + 141, + 116, + 289, + 162 + ], + "type": "image", + "image_path": "249e81cfd94ad453faa8285871dd5e92a0e132261ca3aa141934562277587715.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 171, + 285, + 187 + ], + "lines": [ + { + "bbox": [ + 144, + 171, + 285, + 187 + ], + "spans": [ + { + "bbox": [ + 144, + 171, + 285, + 187 + ], + "type": "text", + "content": "(a) Class distribution for bugs from mixture volumes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 304, + 116, + 388, + 170 + ], + "blocks": [ + { + "bbox": [ + 304, + 116, + 388, + 170 + ], + "lines": [ + { + "bbox": [ + 304, + 116, + 388, + 170 + ], + "spans": [ + { + "bbox": [ + 304, + 116, + 388, + 170 + ], + "type": "image", + "image_path": "4b0d018cfc209628bed82289d0347f3ab40db26b9b666bb8624bf56be349bb4b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 199, + 423, + 210 + ], + "lines": [ + { + "bbox": [ + 189, + 199, + 423, + 210 + ], + "spans": [ + { + "bbox": [ + 189, + 199, + 423, + 210 + ], + "type": "text", + "content": "Fig. 2: Summary of volumes from scans of bug mixtures." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 391, + 118, + 472, + 170 + ], + "blocks": [ + { + "bbox": [ + 391, + 118, + 472, + 170 + ], + "lines": [ + { + "bbox": [ + 391, + 118, + 472, + 170 + ], + "spans": [ + { + "bbox": [ + 391, + 118, + 472, + 170 + ], + "type": "image", + "image_path": "b5103bb97060de34a90c6d8d5b6a0968158601f4e6acb9dd4044e2e4f785c45f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 171, + 466, + 187 + ], + "lines": [ + { + "bbox": [ + 309, + 171, + 466, + 187 + ], + "spans": [ + { + "bbox": [ + 309, + 171, + 466, + 187 + ], + "type": "text", + "content": "(b) Distribution of the number of bugs per volume and the number of classes per volume." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 234, + 479, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 234, + 479, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 479, + 270 + ], + "type": "text", + "content": "volumes are stored as 8-bit unsigned integers with their original spatial resolution of " + }, + { + "bbox": [ + 130, + 234, + 479, + 270 + ], + "type": "inline_equation", + "content": "26.62\\mu \\mathrm{m}" + }, + { + "bbox": [ + 130, + 234, + 479, + 270 + ], + "type": "text", + "content": " (physical voxel size). Our dataset comprises 1.8 trillion voxels, approximately a quarter of the RGB values in ImageNet." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 270, + 480, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 480, + 330 + ], + "type": "text", + "content": "The bugs scanned for the BugNIST dataset are bred as pet food and have been frozen before being packed for scanning. Despite careful packing, some scanned bugs may appear damaged, e.g. a few crickets have broken off antennas or miss a leg. We have chosen to accept these defects because they reflect some variation in a population and add to the challenge of the dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 331, + 480, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 480, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 480, + 401 + ], + "type": "text", + "content": "To create additional lightweight datasets, we created four downsized versions of all volumes. The individual bug volumes, originally 900 voxels long, have lengths of 512, 256, 128, and 64 voxels in the downscaled versions, and all the other lengths are scaled accordingly. The voxel size of the coarsest data is " + }, + { + "bbox": [ + 130, + 331, + 480, + 401 + ], + "type": "inline_equation", + "content": "374.3\\mu \\mathrm{m}" + }, + { + "bbox": [ + 130, + 331, + 480, + 401 + ], + "type": "text", + "content": ", which is a scaling factor of approximately 14 compared to the original scan." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 422, + 230, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 422, + 230, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 422, + 230, + 435 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 447, + 480, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 480, + 566 + ], + "type": "text", + "content": "To establish baselines for volumetric detection under context shift (domain shift with no change in object appearance), we have selected and adapted three existing detection models. Furthermore, we have applied three training strategies, yielding a total of nine baseline models. The diversity in model choices and training strategies contributes to a comprehensive set of baseline models that provide a strong foundation for object detection under context shift. We briefly describe the selected methods and how they were trained. Full details are provided in the supplementary material. To determine the difficulty of differentiating between the individual bugs, we also performed a classification experiment which is described in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 567, + 479, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 479, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 479, + 590 + ], + "type": "text", + "content": "All experiments are carried out using the x128 data, which is a compromise between memory needs and the preservation of structural details of the bugs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 609, + 422, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 422, + 620 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 422, + 620 + ], + "type": "text", + "content": "4.1 Training strategies for detection under context shift" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": "To provide a simple baseline investigation for the BugNIST dataset, we investigate some strategies for solving the detection problem under context shift. Since the purpose of our paper is to initiate research in domain adaptation methods" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 243, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 243, + 91, + 448, + 102 + ], + "type": "text", + "content": "BugNIST - Object Detection under Domain Shift" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 121, + 458, + 268 + ], + "blocks": [ + { + "bbox": [ + 153, + 121, + 458, + 268 + ], + "lines": [ + { + "bbox": [ + 153, + 121, + 458, + 268 + ], + "spans": [ + { + "bbox": [ + 153, + 121, + 458, + 268 + ], + "type": "image", + "image_path": "633471798fd54ad57ec676f5d61247ac367820706fe92530f776fc06a3cf80ae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 289, + 482, + 324 + ], + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 324 + ], + "type": "text", + "content": "Fig. 3: The top line shows the automated annotation pipeline for individual bugs. By thresholding, bounding boxes and segmentation masks are easily obtained. The bottom shows the synthetic mixtures used for the more advanced training strategy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 346, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 430 + ], + "type": "text", + "content": "for volumetric data under context shift, this baseline study is intended as an offset for future developments based on the BugNIST dataset. Our first attempt is to train models from the volumes of individual bugs. Our two other attempts are to use a data augmentation strategy by creating synthetic mixtures. Data augmentation has previously been shown to greatly benefit domain shift problems [60]. The automated annotation of the BugNIST data and the data that we use for training detection algorithms are illustrated in Fig. 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 440, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 482, + 548 + ], + "type": "text", + "content": "Single bugs As a first stage, we train the models to detect the bugs in the individual scans. For models requiring ground truth segmentation labels, those are generated by thresholding the volumes and only keeping the largest connected component. The value of the class label then replaces the foreground voxels. This creates segmentation labels where each voxel has one of 13 values including background or one of the twelve bugs. For models requiring ground truth bounding boxes, those are extracted from each connected component. This process of preparing the training data using this simplest strategy is illustrated in the top line of Fig. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Synthetic mixtures As a second stage, we train the models on artificial mixes. These are generated by randomly sampling individual bug scans and adding them to a common volume. We start with an empty volume of " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "256^3" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " and we set a random target number of bugs drawn from a normal distribution, " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(32,6)" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ". A random bug from a random class (both uniform distributions) is drawn. To create variation, the selected sample is flipped around a random axis, rotated a random multiple (0-3) of 90 degrees around the first axis, and then transposed randomly. The bug volume is intensity normalized, smoothed with a Gaussian filter with " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sigma = 0.5" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ", and thresholded at 0.2. We draw a random voxel position" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "text", + "content": "P.M. Jensen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "from the volume and only insert the bug if its bounding box does not exceed the volume boundaries nor overlaps with existing labeled voxels. We attempt a maximum of 1000 random positions to achieve a valid insertion before the bug is discarded. In total 300 synthetic mixtures are created." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 175, + 483, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 483, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 483, + 259 + ], + "type": "text", + "content": "Crowded synthetic mixtures Finally, since the bugs are packed very close together in the real scans, we create crowded synthetic mixtures. Here the bugs are packed closer by reducing the volume size to " + }, + { + "bbox": [ + 130, + 175, + 483, + 259 + ], + "type": "inline_equation", + "content": "200 \\times 200 \\times 200" + }, + { + "bbox": [ + 130, + 175, + 483, + 259 + ], + "type": "text", + "content": ", adjusting the target number of bugs to " + }, + { + "bbox": [ + 130, + 175, + 483, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(64,8)" + }, + { + "bbox": [ + 130, + 175, + 483, + 259 + ], + "type": "text", + "content": ", and relaxing the insertion criteria such that no labeled voxels can be overwritten. This effectively allows for a larger overlap between bounding boxes and that bugs can touch each other. Again, a total of 300 synthetic mixtures are created." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 277, + 293, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 293, + 288 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 293, + 288 + ], + "type": "text", + "content": "4.2 Baseline detection models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 296, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 369 + ], + "type": "text", + "content": "Despite important uses of detecting objects in 3D volumes, the number of methods addressing volumetric object detection is much smaller compared to 2D, and only a few methods are readily available for volumes. So, to establish baselines for detection under context shift, we have selected and adapted three detection approaches. They vary in how they utilize training data; relying on either ground truth segmentation labels, ground truth bounding boxes, or both." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 379, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 571 + ], + "type": "text", + "content": "Detection U-Net Inspired by [4] we use U-Net [8] as one baseline detection model. U-Net is a dominating model for volumetric segmentation, making it a natural choice for adapting into a baseline detection model. Detection U-Net is trained on volumes accompanied by ground truth segmentation labels. It is first trained to predict the segmentation of the single bugs. Then, starting from the obtained weights, it is trained to predict the segmentation of the synthetic mixtures. Finally, again starting from the previous weights, it is trained to predict the segmentation in the crowded synthetic mixtures. For all training runs, we augment the images with random rotations, translations, geometric scaling, and intensity scaling. We train the models using the sum of Dice and cross entropy loss. During training we hold out a validation split which we evaluate on after every epoch. At the end of training, the model which scored the highest on the validation split is kept. Like [4], we extract bounding boxes by thresholding the softmax probabilities (at 0.75) and then extracting the connected components for each channel. Only components larger than " + }, + { + "bbox": [ + 130, + 379, + 482, + 571 + ], + "type": "inline_equation", + "content": "125(5^{3})" + }, + { + "bbox": [ + 130, + 379, + 482, + 571 + ], + "type": "text", + "content": " voxels are kept. We use the U-Net implementation from MONAI [5]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Faster R-CNN As an example of region proposal networks, we have chosen Faster R-CNN [44] with the modifications from [31]. We have adapted the implementation from torchvision [54] to allow for 3D detection. We train in the same sequence as the U-Net: single bugs, synthetic mixtures, and crowded synthetic mixtures. We use the same augmentations as for the U-Net. In addition, we also randomly perturb the bounding box corners. The model is only trained on bounding boxes and does not make use of the voxel labels. The model is" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 244, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 244, + 91, + 447, + 102 + ], + "type": "text", + "content": "BugNIST - Object Detection under Domain Shift" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "trained with smooth L1 loss for bounding box regression and cross-entropy for objectness and class predictions. For full details, we refer to [31, 44]. As with U-Net, we hold out a validation split which is evaluated after each epoch. After training, the model scoring the highest on the validation split is kept. For inference, only bounding boxes with an objectness score over 0.05 are kept." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 483, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 483, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 483, + 308 + ], + "type": "text", + "content": "nnDetection The third detection model, nnDetection, trains the RetinaUNet [22] detection model using a combination of segmentation masks and bounding boxes. The bounding box branch is trained using a generalized IoU loss for regression and cross-entropy for classification. The voxel segmentation branch is trained with a sum of Dice and cross-entropy loss. As nnDetection is a self-configuring framework, we refer to [4, 21] for further training and model details. We simply provide the data as instructed. We train a separate model for single bugs, synthetic mixtures, and crowded synthetic mixtures. nnDetection trains using 5-fold cross-validation, and the final predictions are made using an ensemble over the best model from each fold [4]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 326, + 216, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 216, + 337 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 216, + 337 + ], + "type": "text", + "content": "4.3 Evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 346, + 482, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 514 + ], + "type": "text", + "content": "We assess model performance by evaluating their detection accuracy on real scans of densely packed bugs, using manually annotated center points. These annotations are matched with predicted bounding boxe centers based on Euclidean distance, optimized using Munkres' algorithm [38]. Matches are only accepted if the annotated point falls within a " + }, + { + "bbox": [ + 130, + 346, + 482, + 514 + ], + "type": "inline_equation", + "content": "1.5 \\times" + }, + { + "bbox": [ + 130, + 346, + 482, + 514 + ], + "type": "text", + "content": " inflated bounding box, as illustrated in Fig. 4. This inflation ensures a fair assessment for detections like in Fig. 4(b). This, combined with matching on the bounding box centers, helps ensure that large boxes are not favored since it places more weight on the location of the boxes. We compute precision to avoid favoring overdetections, recall to measure the proportion of correct matches, and F1-score to combine the two. Further, we first compute the metrics purely from the point matches without considering classes — thereby evaluating if the models can detect the presence of bugs vs. background material. Next, we recompute the matches where we only allow matches between points of equal class." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 533, + 200, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 200, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 200, + 544 + ], + "type": "text", + "content": "5 Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 617 + ], + "type": "text", + "content": "The results are summarized in Tab. 1, example results are shown in Fig. 5, and the distributions of the metrics without class information are shown in Fig. 6. U-Net performs relatively well, independent of what it was trained on, whereas Faster R-CNN performs worse on crowded mixtures and nnDetection on single bugs. If class labels are required, the performance is generally low." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "The distributions in Fig. 6 show that U-Net has a narrower spread around the mean for the F1-score, whereas the two others are spread more out. Wrt. the precision, many detections obtained by the nnDetection method are 1, but have a lower recall, meaning that those that it predicts are correct, but there" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "P.M. Jensen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 171, + 117, + 260, + 189 + ], + "blocks": [ + { + "bbox": [ + 171, + 117, + 260, + 189 + ], + "lines": [ + { + "bbox": [ + 171, + 117, + 260, + 189 + ], + "spans": [ + { + "bbox": [ + 171, + 117, + 260, + 189 + ], + "type": "image", + "image_path": "54b7172581217d9d6bb33f28eb39c37bfe6bc886dea38ea23106e9fc679553c7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 191, + 221, + 200 + ], + "lines": [ + { + "bbox": [ + 211, + 191, + 221, + 200 + ], + "spans": [ + { + "bbox": [ + 211, + 191, + 221, + 200 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 261, + 117, + 351, + 189 + ], + "blocks": [ + { + "bbox": [ + 261, + 117, + 351, + 189 + ], + "lines": [ + { + "bbox": [ + 261, + 117, + 351, + 189 + ], + "spans": [ + { + "bbox": [ + 261, + 117, + 351, + 189 + ], + "type": "image", + "image_path": "842f6f2ecd4543c290ad5358ab452b07011e497f1c29141f66909bfe570d35d3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 191, + 312, + 200 + ], + "lines": [ + { + "bbox": [ + 301, + 191, + 312, + 200 + ], + "spans": [ + { + "bbox": [ + 301, + 191, + 312, + 200 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 212, + 482, + 277 + ], + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 277 + ], + "type": "text", + "content": "Fig. 4: Matching criterion for bounding box to center point. The circle shows the manually annotated center point, the predicted bounding box is shown with the full line, and the inflated bounding box is shown with the dashed line. In (a), the match is accepted by the predicted bounding box, in (b), the match is also accepted due to the inflated bounding box, whereas in (c), there is no match. Note that this illustration is 2D, but the detection is done in 3D." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 353, + 117, + 442, + 189 + ], + "blocks": [ + { + "bbox": [ + 353, + 117, + 442, + 189 + ], + "lines": [ + { + "bbox": [ + 353, + 117, + 442, + 189 + ], + "spans": [ + { + "bbox": [ + 353, + 117, + 442, + 189 + ], + "type": "image", + "image_path": "8d3a276c11a1f9dcfc671c18397288ad8af552f9e342ffb804e5cdb1ee37775a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 191, + 403, + 200 + ], + "lines": [ + { + "bbox": [ + 392, + 191, + 403, + 200 + ], + "spans": [ + { + "bbox": [ + 392, + 191, + 403, + 200 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 303, + 480, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 480, + 338 + ], + "type": "text", + "content": "are many potential detections that it misses. It is the opposite for the Faster R-CNN, where many datasets get a recall of 1, but a lower precision, meaning that it detects many of the bugs, but also has many false detections." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 340, + 482, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 507 + ], + "type": "text", + "content": "To investigate the capabilities of the three methods without the domain shift between the synthetic and real mixtures, we included the results in Tab. 2. This table shows the results of the validation data for training the models on the volumes of individual bugs and the two synthetic mixtures with the class labels. Since we have the bounding boxes for these datasets, we compute the mean average precision (mAP), similar to, e.g., COCO detection [34]. We compute mAP with the IoU threshold running from 0.5 to 0.95 (shown as mAP) and fixed at 0.5 (mAP@50). We note that the values are not directly comparable with those in Tab. 1. We show example detections in the supplementary. The U-Net generally has high performance, Faster R-CNN has high performance for single bugs but has difficulty with the mixtures, whereas nnDetection performs better in the mixtures than for the individual bugs. Generally, the detection performance of bugs is far from perfect and can be improved, also when training and testing within the same domain." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 529, + 217, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 529, + 217, + 541 + ], + "spans": [ + { + "bbox": [ + 132, + 529, + 217, + 541 + ], + "type": "text", + "content": "6 Discussion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": "The domain shift introduced by shifting the context surrounding the bugs significantly increases the difficulty of the analysis problem. While U-Net and nnDetection achieve high performance on their training tasks (at least for synthetic mixtures), they struggle when transferred to real densely packed scans. It is also interesting that the domain shift has such an effect on the ability of the models to infer the correct class even when accurately localizing the bug. This is surprising since detecting the bugs in the training domain can be done with high accuracy, as shown in Tab. 2, and pure classification (see supplementary) gave an accuracy of 0.97. Despite the models for classification being different than the detection" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "type": "text", + "content": "BugNIST - Object Detection under Domain Shift" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 168, + 482, + 289 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: Detection results on mixed scans. We show mean and std. for each metric computed over the evaluation images. Higher is better. Without class info refers to not considering the bug label and with class info expects the bug type to be classified as well as being detected." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 168, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 132, + 168, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 168, + 482, + 289 + ], + "type": "table", + "html": "
Model- Trained onWithout class info.With class info.
F1-ScorePrecisionRecallF1-ScorePrecisionRecall
U-Net [8]- Single bugs0.59±0.180.68±0.170.60±0.250.11±0.160.13±0.190.11±0.16
- Synth. mixes0.60±0.120.52±0.170.82±0.190.11±0.080.09±0.080.14±0.10
- Crowded s.m.0.54±0.130.39±0.140.97±0.070.10±0.080.07±0.060.18±0.14
Faster R-CNN [44]- Single bugs0.56±0.201.00±0.040.42±0.210.11±0.100.21±0.190.08±0.08
- Synth. mixes0.68±0.160.63±0.210.83±0.220.16±0.140.15±0.140.19±0.18
- Crowded s.m.0.27±0.170.81±0.240.18±0.160.06±0.090.19±0.300.04±0.06
nn-Detection [4]- Single bugs0.03±0.020.02±0.010.91±0.150.00±0.010.00±0.000.09±0.16
- Synth. mixes0.50±0.140.41±0.170.80±0.220.10±0.060.08±0.060.16±0.11
- Crowded s.m.0.60±0.190.77±0.230.58±0.270.11±0.120.15±0.160.11±0.12
", + "image_path": "53bd22c734e829ea8853fbd694369dc9e61fd9a5132ecb1ece9be7314efa7f38.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 132, + 378, + 483, + 483 + ], + "blocks": [ + { + "bbox": [ + 130, + 303, + 482, + 369 + ], + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 369 + ], + "type": "text", + "content": "Table 2: Detection results on the training tasks obtained during training on the single bugs and the two synthetic mixtures. For U-Net and Faster R-CNN performance is evaluated on a validation split. Since nnDetection is an ensemble created from multiple train-validation splits, we evaluate on all images in the training task. We show mean average precision (mAP) between predicted and ground truth bounding boxes. Higher is better." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 378, + 483, + 483 + ], + "lines": [ + { + "bbox": [ + 132, + 378, + 483, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 378, + 483, + 483 + ], + "type": "table", + "html": "
Model- Trained onmAPmAP@50
U-Net [8]- Single bugs0.720.72
- Synthetic mixes0.640.77
- Crowded synth. mixes0.750.86
Faster R-CNN [44]- Single bugs0.700.70
- Synthetic mixes0.020.02
- Crowded synth. mixes0.050.05
nnDetection [4]- Single bugs0.100.10
- Synthetic mixes0.590.70
- Crowded synth. mixes0.490.66
", + "image_path": "0f82912a385e7a3a23be4a4fd19898251be9ae927ba9eaf332eb27301b12f894.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 509, + 481, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 481, + 544 + ], + "type": "text", + "content": "models, these results show that even when the objects do not change appearance, the shift in the domain by changing the context surrounding the objects is sufficient for deep learning-based detection models to lose performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "In preparing the data, we have strived to get the scans of individual bugs and bugs in mixtures to have the same appearance. Since bugs are flexible, there will be differences in the pose of the bugs in mixtures and individual bugs. Larvae will sometimes bend differently and bugs with antennae or legs will bend differently. Also, the cotton is visible and can slightly influence the appearance. Despite these differences, the voxel intensity distributions in individual scanned bugs and mixtures are the same, and therefore the BugNIST dataset is a good basis for developing detection methods that can handle context shift. If the mixtures were (partly) labeled for segmentation, BugNIST could also be used to investigate segmentation algorithms under context shift." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "P.M. Jensen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 113, + 478, + 360 + ], + "blocks": [ + { + "bbox": [ + 135, + 113, + 478, + 360 + ], + "lines": [ + { + "bbox": [ + 135, + 113, + 478, + 360 + ], + "spans": [ + { + "bbox": [ + 135, + 113, + 478, + 360 + ], + "type": "image", + "image_path": "eb1ac95eecc14df2f0bc40673c32812a94a2999f3c09b7d423ad367754d0275d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 369, + 482, + 456 + ], + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 456 + ], + "type": "text", + "content": "Fig. 5: Detection results on real mixed scans. The images are maximum-intensity projections of the volumes. Center points are marked with circles, detections are marked with bounding boxes, and color indicates the bug class. A line from a point to the upper left corner of a bounding box shows a match. If the center point and the bounding box have the same color, then the detection and classification are correct. The shown images are the volumes with the highest scores for the different models. For U-Net the results when trained on synthetic mixes, for Faster R-CNN when trained on synthetic mixes and nnDetection when trained on crowded synthetic mixes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 140, + 472, + 475, + 522 + ], + "blocks": [ + { + "bbox": [ + 140, + 472, + 475, + 522 + ], + "lines": [ + { + "bbox": [ + 140, + 472, + 475, + 522 + ], + "spans": [ + { + "bbox": [ + 140, + 472, + 475, + 522 + ], + "type": "image", + "image_path": "01c91c05303a061e584404fa4aebd0b5e77a0f9d8179303ed60136159284688b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 533, + 461, + 544 + ], + "lines": [ + { + "bbox": [ + 151, + 533, + 461, + 544 + ], + "spans": [ + { + "bbox": [ + 151, + 533, + 461, + 544 + ], + "type": "text", + "content": "Fig. 6: Distributions of performance scores for the three detection methods." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "We chose to annotate only the center positions of the bugs. As the bugs are often elongated and at odd angles, axis-aligned bounding boxes would contain significant empty space and overlap between neighboring boxes—thus offering limited benefit over points. This could be solved with object aligned bounding boxes or full voxel annotations. However, this would drastically increase the annotation time. We chose to prioritize having a large number of annotations such that the test set can have a large variety of configurations. Furthermore, given that current methods show ample room for improvement (cf. Tab. 1), we" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 244, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 244, + 91, + 447, + 102 + ], + "type": "text", + "content": "BugNIST - Object Detection under Domain Shift" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "find that point annotations combined with the chosen metric are sufficient to measure method performance and improvements of future approaches." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 259 + ], + "type": "text", + "content": "The methods and experiments are chosen as a basic baseline to illustrate the properties of the BugNIST dataset. From the experiments, we learned that data augmentation by composing synthetic mixtures is a feasible approach for improving detection performance. More investigations into packing algorithms and ways to combine the synthetic mixed bugs with other backgrounds can potentially improve performance further. But it is also clear that there is great potential for developing new methods for object detection in volumes. We chose Detection U-Net, nnDetection, and our adaption of Faster R-CNN to 3D volumes because these methods were available, widely used, and are based on complementary ideas. We did not find more recent alternatives for volumetric analysis." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 260, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 368 + ], + "type": "text", + "content": "Our results show that Faster R-CNN performs less than the two other methods. While it scores highest in Tab. 1, it is clear from Fig. 5 that the bounding boxes are of much lower quality. Furthermore, it does not score well on the training tasks, except when trained on individual bugs. To ensure the poor performance of Faster R-CNN was not just the result of too few training volumes, we redid the experiments with 1000 synthetic mixes instead of the original 300. Performance remained the same. It is also noteworthy that nnDetection scores overall lower than U-Net on both the training and evaluation tasks, even though it also has access to the segmentation masks for training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 384, + 220, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 384, + 220, + 398 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 220, + 398 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 408, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 482, + 563 + ], + "type": "text", + "content": "With our BugNIST benchmark dataset for domain shift in volumetric data, we demonstrate how detecting objects, that have the same appearance in the source and target domain, fails when the objects in the two domains are placed in different contexts. By contexts, we refer to the appearance of the image surrounding the objects. To our knowledge, the BugNIST dataset is the only dataset designed for developing methods for context shift, i.e. domain shift where the objects stay the same but the surroundings change. Solutions will enable an automated annotation of data for volumetric object detection, which is an area that has not received much attention. There are, however, a range of applications where object detection is needed, including medical imaging and emerging areas such as micro-CT scanning of materials. The BugNIST dataset can also be used for investigating other problems such as segmentation under domain shift provided reference annotations are created." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 582, + 241, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 582, + 241, + 596 + ], + "spans": [ + { + "bbox": [ + 132, + 582, + 241, + 596 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "We acknowledge the 3D Imaging Center at DTU, the Infrastructure for Quantitative AI-based Tomography (QUAITOM), supported by the Novo Nordisk Foundation (grant number NNF21OC0069766) and STUDIOS: Segmenting Tomograms Using Different Interpretation of Simplicity funded by the Villum Foundation (grant number VIL50425)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "P.M. Jensen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 142, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 138, + 142, + 481, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 481, + 187 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 481, + 187 + ], + "type": "text", + "content": "1. Armato, S.G., Huisman, H., Drukker, K., Hadjiiski, L., Kirby, J.S., Petrick, N., Redmond, G., Giger, M.L., Cha, K., Mamonov, A., et al.: PROSTATEx challenges for computerized classification of prostate lesions from multiparametric magnetic resonance images. Journal of Medical Imaging 5(4), 044501-044501 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 189, + 481, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 189, + 481, + 243 + ], + "spans": [ + { + "bbox": [ + 138, + 189, + 481, + 243 + ], + "type": "text", + "content": "2. Armato, S.G., McLennan, G., Bidaut, L., McNitt-Gray, M.F., Meyer, C.R., Reeves, A.P., Zhao, B., Aberle, D.R., Henschke, C.I., Hoffman, E.A., et al.: The lung image database consortium (LIDC) and image database resource initiative (IDRI): a completed reference database of lung nodules on CT scans. Medical Physics 38(2), 915–931 (2011)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 243, + 481, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 481, + 287 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 481, + 287 + ], + "type": "text", + "content": "3. Baid, U., Ghodasara, S., Mohan, S., Bilello, M., Calabrese, E., Colak, E., Farahani, K., Kalpathy-Cramer, J., Kitamura, F.C., Pati, S., et al.: The RSNA-ASNR-MICCAI BraTS 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 288, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 288, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 288, + 481, + 331 + ], + "type": "text", + "content": "4. Baumgartner, M., Jäger, P.F., Isensee, F., Maier-Hein, K.H.: nnDetection: a self-configuring method for medical object detection. In: Proceedings of the Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 530-539. Springer (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 332, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 481, + 365 + ], + "type": "text", + "content": "5. Cardoso, M.J., Li, W., Brown, R., Ma, N., Kerfoot, E., Wang, Y., Murrey, B., Myronenko, A., Zhao, C., Yang, D., et al.: MONAI: An open-source framework for deep learning in healthcare. arXiv preprint arXiv:2211.02701 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 366, + 481, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 366, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 366, + 481, + 399 + ], + "type": "text", + "content": "6. Charytanowicz, M., Kulczycki, P., Kowalski, P.A., Lukasik, S., Czabak-Garbacz, R.: An evaluation of utilizing geometric features for wheat grain classification using X-ray images. Computers and Electronics in Agriculture 144, 260-268 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 399, + 481, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 481, + 454 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 481, + 454 + ], + "type": "text", + "content": "7. Chaudhary, S., Sadbhawna, S., Jakhetiya, V., Subudhi, B.N., Baid, U., Guntuku, S.C.: Detecting COVID-19 and community acquired pneumonia using chest CT scan images with deep learning. In: Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 8583-8587. IEEE (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 455, + 481, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 455, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 455, + 481, + 498 + ], + "type": "text", + "content": "8. Çiçek, Ö., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3D U-Net: learning dense volumetric segmentation from sparse annotation. In: Proceedings of the Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 424–432. Springer (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 499, + 481, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 499, + 481, + 543 + ], + "spans": [ + { + "bbox": [ + 138, + 499, + 481, + 543 + ], + "type": "text", + "content": "9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3213-3223 (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 544, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 481, + 587 + ], + "type": "text", + "content": "0. De Carlo, F., Gürsoy, D., Ching, D.J., Batenburg, K.J., Ludwig, W., Mancini, L., Marone, F., Mokso, R., Pelt, D.M., Sijbers, J., et al.: Tomobank: a tomographic data repository for computational x-ray science. Measurement Science and Technology 29(3), 034004 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 588, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 481, + 632 + ], + "type": "text", + "content": "1. Dorent, R., Kujawa, A., Ivory, M., Bakas, S., Rieke, N., Joutard, S., Glocker, B., Cardoso, J., Modat, M., Batmanghelich, K., et al.: Crossmoda 2021 challenge: Benchmark of cross-modality domain adaptation techniques for vestibular schwannoma and cochlea segmentation. Medical Image Analysis 83, 102628 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "type": "text", + "content": "2. Du, Z., Hu, Y., Ali Buttar, N., Mahmood, A.: X-ray computed tomography for quality inspection of agricultural products: A review. Food Science & Nutrition 7(10), 3146-3160 (2019)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 243, + 91, + 447, + 102 + ], + "type": "text", + "content": "BugNIST - Object Detection under Domain Shift" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "13. Einarsdóttir, H., Emerson, M.J., Clemmensen, L.H., Scherer, K., Willer, K., Bech, M., Larsen, R., Ersbøll, B.K., Pfeiffer, F.: Novelty detection of foreign objects in food using multi-modal X-ray imaging. Food Control 67, 39-47 (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "text", + "content": "14. European Organization For Nuclear Research, OpenAIRE: Zenodo (2013). https://doi.org/10.25495/7GXK-RD71, https://www.zenodo.org/" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "text", + "content": "15. Flitton, G., Mouton, A., Breckon, T.P.: Object classification in 3D baggage security computed tomography imagery using visual codebooks. Pattern Recognition 48(8), 2489-2499 (2015)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "16. Girshick, R.: Fast R-CNN. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 1440-1448 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "type": "text", + "content": "17. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object detection and semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 580-587 (2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "text", + "content": "18. Guan, H., Liu, M.: Domain adaptation for medical image analysis: a survey. IEEE Transactions on Biomedical Engineering 69(3), 1173-1185 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 335 + ], + "type": "text", + "content": "19. Harmon, S.A., Sanford, T.H., Xu, S., Turkbey, E.B., Roth, H., Xu, Z., Yang, D., Myronenko, A., Anderson, V., Amalou, A., et al.: Artificial intelligence for the detection of COVID-19 pneumonia on chest CT using multinational datasets. Nature Communications 11(1), 4080 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "20. Hipsley, C.A., Aguilar, R., Black, J.R., Hocknull, S.A.: High-throughput microCT scanning of small specimens: preparation, packing, parameters and post-processing. Scientific Reports 10(1), 13863 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "21. Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods 18(2), 203-211 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 445 + ], + "type": "text", + "content": "22. Jaeger, P.F., Kohl, S.A., Bickelhaupt, S., Isensee, F., Kuder, T.A., Schlemmer, H.P., Maier-Hein, K.H.: Retina U-Net: Embarrassingly simple exploitation of segmentation supervision for medical object detection. In: Machine Learning for Health Workshop. pp. 171-183. PMLR (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "text", + "content": "23. Jain, Y., Borner, K., Walsh, C., Ruschman, N., Lee, P.D., Weber, G.M., Holbrook, R., Howard, A.: SenNet + HOA - hacking the human vasculature in 3D (2023), https://kaggle.com/competitions/blood-vessel-segmentation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "text", + "content": "24. Jarolmasjed, S., Espinoza, C.Z., Sankaran, S., Khot, L.R.: Postharvest bitter pit detection and progression evaluation in 'Honeycrisp' apples using computed tomography images. Postharvest Biology and Technology 118, 35-42 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "25. Jeppesen, N., Christensen, A.N., Dahl, V.A., Dahl, A.B.: Sparse layered graphs for multi-object segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12777-12785 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 545, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 588 + ], + "type": "text", + "content": "26. Jin, L., Yang, J., Kuang, K., Ni, B., Gao, Y., Sun, Y., Gao, P., Ma, W., Tan, M., Kang, H., et al.: Deep-learning-assisted detection and segmentation of rib fractures from CT scans: Development and validation of FracNet. EBioMedicine 62, 103106 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 621 + ], + "type": "text", + "content": "27. Ker, J., Singh, S.P., Bai, Y., Rao, J., Lim, T., Wang, L.: Image thresholding improves 3-dimensional convolutional neural network diagnosis of different acute brain hemorrhages on computed tomography scans. Sensors 19(9), 2167 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "text", + "content": "28. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "P.M. Jensen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "29. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 171 + ], + "type": "text", + "content": "30. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.M.: Deeper, broader and artier domain generalization. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 5542-5550 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 194 + ], + "type": "text", + "content": "31. Li, Y., Xie, S., Chen, X., Dollar, P., He, K., Girshick, R.: Benchmarking detection transfer learning with vision transformers. arXiv preprint arXiv:2111.11429 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "text", + "content": "32. Li, Y., Fan, Y.: Medical image segmentation with domain adaptation: A survey. arXiv preprint arXiv:2311.01702 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "text", + "content": "33. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). pp. 2980-2988 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 282 + ], + "type": "text", + "content": "34. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "type": "text", + "content": "35. Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature Communications 15(1), 654 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 346 + ], + "type": "text", + "content": "36. Maier-Hein, L., Eisenmann, M., Reinke, A., Onogur, S., Stankovic, M., Scholz, P., Arbel, T., Bogunovic, H., Bradley, A.P., Carass, A., et al.: Why rankings of biomedical image analysis competitions should be interpreted with care. Nature Communications 9(1), 5217 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "37. Muandet, K., Balduzzi, D., Scholkopf, B.: Domain generalization via invariant feature representation. In: International Conference on Machine Learning. pp. 10-18. PMLR (2013)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "38. Munkres, J.: Algorithms for the assignment and transportation problems. Journal of the Society for Industrial and Applied Mathematics 5(1), 32-38 (1957)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 403, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 413 + ], + "type": "text", + "content": "39. OECD: Health at a Glance 2023 (2023). https://doi.org//10.1787/7a7afb35-en" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "type": "text", + "content": "40. Pan, S.J., Tsang, I.W., Kwok, J.T., Yang, Q.: Domain adaptation via transfer component analysis. IEEE Transactions on Neural Networks 22(2), 199-210 (2010)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "41. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1406-1415 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "text", + "content": "42. Rawson, S.D., Maksimcuka, J., Withers, P.J., Cartmell, S.H.: X-ray computed tomography in life sciences. BMC biology 18(1), 1-15 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "type": "text", + "content": "43. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 779-788 (2016)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "type": "text", + "content": "44. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. Advances in neural information processing systems (NeurIPS) 28 (2015)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 555, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 600 + ], + "type": "text", + "content": "45. Richter, S.R., Vineet, V., Roth, S., Koltun, V.: Playing for data: Ground truth from computer games. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 102-118. Springer (2016)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 600, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 643 + ], + "type": "text", + "content": "46. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: Proceedings of the Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). pp. 234-241. Springer (2015)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "47. Ros, G., Sellart, L., Materzynska, J., Vazquez, D., Lopez, A.M.: The synthia dataset: A large collection of synthetic images for semantic segmentation of urban" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 244, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 244, + 91, + 447, + 102 + ], + "type": "text", + "content": "BugNIST - Object Detection under Domain Shift" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 147, + 117, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 117, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 117, + 481, + 138 + ], + "type": "text", + "content": "scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3234-3243 (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 183 + ], + "type": "text", + "content": "48. Sankaranarayanan, S., Balaji, Y., Jain, A., Lim, S.N., Chellappa, R.: Learning from synthetic data: Addressing domain shift for semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3752-3761 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 481, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 481, + 206 + ], + "type": "text", + "content": "49. Serte, S., Demirel, H.: Deep learning for diagnosis of COVID-19 using 3D CT scans. Computers in biology and medicine 132, 104306 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 208, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 208, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 208, + 481, + 262 + ], + "type": "text", + "content": "50. Setio, A.A.A., Traverso, A., De Bel, T., Berens, M.S., Van Den Bogaard, C., Cerello, P., Chen, H., Dou, Q., Fantacci, M.E., Geurts, B., et al.: Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: the luna16 challenge. Medical Image Analysis 42, 1-13 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 263, + 481, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 481, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 481, + 339 + ], + "type": "text", + "content": "51. Simpson, A.L., Antonelli, M., Bakas, S., Bilello, M., Farahani, K., van Ginneken, B., Kopp-Schneider, A., Landman, B.A., Litjens, G., Menze, B.H., Ronneberger, O., Summers, R.M., Bilic, P., Christ, P.F., Do, R.K.G., Gollub, M., Golia-Pernicka, J., Heckers, S., Jarnagin, W.R., McHugo, M., Napel, S., Vorontsov, E., Maier-Hein, L., Cardoso, M.J.: A large annotated medical image dataset for the development and evaluation of segmentation algorithms. CoRR abs/1902.09063 (2019), http://arxiv.org/abs/1902.09063" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 341, + 481, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 481, + 384 + ], + "type": "text", + "content": "52. Sun, T., Segu, M., Postels, J., Wang, Y., Van Gool, L., Schiele, B., Tombari, F., Yu, F.: Shift: a synthetic driving dataset for continuous multi-task domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21371-21382 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 385, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 481, + 430 + ], + "type": "text", + "content": "53. Timmins, K.M., van der Schaaf, I.C., Bennink, E., Ruigrok, Y.M., An, X., Baumgartner, M., Bourdon, P., De Feo, R., Di Noto, T., Dubost, F., et al.: Comparing methods of detecting and segmenting unruptured intracranial aneurysms on TOF-MRAS: The ADAM challenge. Neuroimage 238, 118216 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 430, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 453 + ], + "type": "text", + "content": "54. TorchVision maintainers and contributors: Torchvision: Pytorch's computer vision library. https://github.com/pytorch/vision (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 453, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 485 + ], + "type": "text", + "content": "55. Trusler, M.M., Sturrock, C.J., Vane, C.H., Cook, S., Lomax, B.H.: X-ray computed tomography: A novel non-invasive approach for the detection of microplastics in sediments? Marine Pollution Bulletin 194, 115350 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 487, + 481, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 481, + 529 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 481, + 529 + ], + "type": "text", + "content": "56. Velayudhan, D., Hassan, T., Damiani, E., Werghi, N.: Recent advances in baggage threat detection: A comprehensive and systematic survey. ACM Computing Surveys 55(8) (dec 2022). https://doi.org/10.1145/3549932, https://doi.org/10.1145/3549932" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 531, + 481, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 481, + 563 + ], + "type": "text", + "content": "57. Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5018-5027 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 565, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 481, + 609 + ], + "type": "text", + "content": "58. Walsh, C., Tafforeau, P., Wagner, W., Jafree, D., Bellier, A., Werlein, C., Kühnel, M., Boller, E., Walker-Samuel, S., Robertus, J., et al.: Imaging intact human organs with local resolution of cellular structures using hierarchical phase-contrast tomography. Nature Methods 18(12), 1532–1541 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 609, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 481, + 632 + ], + "type": "text", + "content": "59. Wang, M., Deng, W.: Deep visual domain adaptation: A survey. Neurocomputing 312, 135-153 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "60. Wang, M., Liu, Y., Yuan, J., Wang, S., Wang, Z., Wang, W.: Inter-class and inter-domain semantic augmentation for domain generalization. IEEE Transactions on Image Processing (2024)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "text", + "content": "P.M. Jensen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 534 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "61. Wang, Q., Bhowmik, N., Breckon, T.P.: Multi-class 3D object detection within volumetric 3D computed tomography baggage security screening imagery. In: Proceedings of the IEEE International Conference on Machine Learning and Applications (ICMLA). pp. 13-18. IEEE (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 482, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 482, + 204 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 482, + 204 + ], + "type": "text", + "content": "62. Wang, Q., Breckon, T.P.: Contraband materials detection within volumetric 3D computed tomography baggage security screening imagery. In: Proceedings of the IEEE International Conference on Machine Learning and Applications (ICMLA). pp. 75-82. IEEE (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "text", + "content": "63. Withers, P.J., Bouman, C., Carmignato, S., Cnudde, V., Grimaldi, D., Hagen, C.K., Maire, E., Manley, M., Du Plessis, A., Stock, S.R.: X-ray computed tomography. Nature Reviews Methods Primers 1(1), 18 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 237, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 281 + ], + "type": "text", + "content": "64. Wu, D., Wu, D., Feng, H., Duan, L., Dai, G., Liu, X., Wang, K., Yang, P., Chen, G., Gay, A.P., et al.: A deep learning-integrated micro-ct image analysis pipeline for quantifying rice lodging resistance-related traits. Plant Communications 2(2) (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 281, + 482, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 304 + ], + "type": "text", + "content": "65. Xiao, Z., Liu, B., Geng, L., Zhang, F., Liu, Y.: Segmentation of lung nodules using improved 3d-unet neural network. Symmetry 12(11), 1787 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 304, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 335 + ], + "type": "text", + "content": "66. Xie, Y., Ji, Q.: A new efficient ellipse detection method. In: Proceedings of the International Conference on Pattern Recognition (ICPR). vol. 2, pp. 957-960. IEEE (2002)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 335, + 482, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 357 + ], + "type": "text", + "content": "67. Xu, F., Tenenbaum, J.B.: Word learning as bayesian inference. Psychological review 114(2), 245 (2007)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 358, + 482, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 401 + ], + "type": "text", + "content": "68. Yan, Y.T., Chua, S., DeCarlo, T.M., Kempf, P., Morgan, K.M., Switzer, A.D.: Core-ct: A matlab application for the quantitative analysis of sediment and coral cores from X-ray computed tomography (ct). Computers & Geosciences 156, 104871 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 402, + 482, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 435 + ], + "type": "text", + "content": "69. Zhao, W.X., Zhou, K., Li, J., Tang, T., Wang, X., Hou, Y., Min, Y., Zhang, B., Zhang, J., Dong, Z., et al.: A survey of large language models. arXiv preprint arXiv:2303.18223 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 435, + 482, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 488 + ], + "type": "text", + "content": "70. Zhou, K., Yang, Y., Hospedales, T., Xiang, T.: Deep domain-adversarial image generation for domain generalisation. Proceedings of the AAAI Conference on Artificial Intelligence 34(07), 13025-13032 (April 2020). https://doi.org/10.1609/aaii.v34i07.7003, https://ojs.aaii.org/index.php/AAAI/article/view/7003" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 489, + 482, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 489, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 489, + 482, + 534 + ], + "type": "text", + "content": "71. Zhou, S.K., Greenspan, H., Davatzikos, C., Duncan, J.S., Van Ginneken, B., Madabhushi, A., Prince, J.L., Rueckert, D., Summers, R.M.: A review of deep learning in medical imaging: Imaging traits, technology trends, case studies with progress highlights, and future promises. Proceedings of the IEEE 109(5), 820-838 (2021)" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 244, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 244, + 91, + 447, + 102 + ], + "type": "text", + "content": "BugNIST - Object Detection under Domain Shift" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/BurstM_ Deep Burst Multi-scale SR using Fourier Space with Optical Flow/608ca9aa-ec95-4beb-807e-424a93d0d485_content_list.json b/2024/BurstM_ Deep Burst Multi-scale SR using Fourier Space with Optical Flow/608ca9aa-ec95-4beb-807e-424a93d0d485_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9909b4b099d45eb1f5ca0fad1a4b6cfb8cb82b4a --- /dev/null +++ b/2024/BurstM_ Deep Burst Multi-scale SR using Fourier Space with Optical Flow/608ca9aa-ec95-4beb-807e-424a93d0d485_content_list.json @@ -0,0 +1,1773 @@ +[ + { + "type": "text", + "text": "BurstM: Deep Burst Multi-scale SR using Fourier Space with Optical Flow", + "text_level": 1, + "bbox": [ + 217, + 140, + 785, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "EungGu Kang¹ Byeonghun Lee² Sunghoon Im1* Kyong Hwan Jin2*", + "bbox": [ + 215, + 210, + 785, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Daegu Gyeongbuk Institute of Science and Technology (DGIST), Korea", + "bbox": [ + 254, + 238, + 746, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Korea University, Korea", + "bbox": [ + 413, + 253, + 589, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "eunggukang@gmail.com, sunghoonim@dgist.ac.kr, {byeonghun_lee,", + "bbox": [ + 263, + 268, + 733, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "kyong_jin}@korea.ac.kr", + "bbox": [ + 413, + 282, + 586, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Multi frame super-resolution (MFSR) achieves higher performance than single image super-resolution (SISR), because MFSR leverages abundant information from multiple frames. Recent MFSR approaches adapt the deformable convolution network (DCN) to align the frames. However, the existing MFSR suffers from misalignments between the reference and source frames due to the limitations of DCN, such as small receptive fields and the predefined number of kernels. From these problems, existing MFSR approaches struggle to represent high-frequency information. To this end, we propose Deep Burst Multi-scale SR using Fourier Space with Optical Flow (BurstM). The proposed method estimates the optical flow offset for accurate alignment and predicts the continuous Fourier coefficient of each frame for representing high-frequency textures. In addition, we have enhanced the network's flexibility by supporting various super-resolution (SR) scale factors with the unimodel. We demonstrate that our method has the highest performance and flexibility than the existing MFSR methods. Our source code is available at https://github.com/Egkang-Luis/burstm", + "bbox": [ + 261, + 297, + 740, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Super resolution $\\cdot$ Burst super resolution $\\cdot$ Implicit neural representation $\\cdot$ Fourier features", + "bbox": [ + 259, + 545, + 740, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 582, + 375, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single image super-resolution (SISR) is the major ill-posed problem in low-level computer vision. SISR aims to enhance the image quality of low-resolution by estimating sub-pixel information. After the success of deep learning, the most SISR [10,25,31,44,45] adapt the learning-based method to estimate high-resolution images. However, SISR has a limitation in high-frequency representation of the capacity due to a lack of information.", + "bbox": [ + 212, + 606, + 785, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Super-resolution researchers focus on multi-frame super-resolution (MFSR) due to the accurate representation of high-frequency textures. Multiple frames provide additional sub-pixel priors, and MFSR utilizes rich information with alignment between the reference and source frames. However, multiple frames take similar scenes with degradations such as noise and object movement. MFSR requires the accurate alignment performance to utilize low-quality frames, because the inaccurate alignment performance can lead to the blur artifacts in the prediction result. Thus, the alignment process is the critical challenge of MFSR.", + "bbox": [ + 212, + 696, + 785, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\* Co-corresponding authors.", + "bbox": [ + 218, + 825, + 410, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9b4a1f34c71f3a5e518e3e18bcdbf253510cd5e58ae6516a6cdb6a5aa6c1535d.jpg", + "image_caption": [ + "Fig. 1: Overview of proposed Burst SR network. The proposed method (BurstM) predicts high-resolution images from multiple low-resolution images. The neural warping performs high-precision warping with upscaling on Fourier space. Additionally, the neural warping contributes to the prediction of an accurate high-resolution images." + ], + "image_footnote": [], + "bbox": [ + 303, + 145, + 702, + 299 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Like SISR, the latest MFSR [3, 4, 11, 12, 29, 30] employs a deep learning method. They adapt similar architectures, which is complicated feature extraction methods using CNN or transformer [7, 24, 26] and alignment using DCN [9] to reconstruct the high-resolution images. Recent MFSR [11, 12, 29, 30] adapts multi-layer DCN structures to extend a small receptive field of DCN, but even this approach still has limited receptive fields. In addition, the predefined number of DCN kernels restricts representing performance during aligning features. The fixed SR scale is another limitation of existing MFSR due to their upsampling method. Most networks [11, 12, 29, 30] generally adapt pixel shuffle (PS) [37] as an upsampling module due to the effectiveness of PS. Nevertheless, PS restricts the network's flexibility, e.g., $\\times 4$ SR, because PS can conduct upsampling by a predetermined scale, which is the square of the channel number. If the network predicts a different SR scale, it must re-train with changing architecture or additional downsampling after a predefined SR scale.", + "bbox": [ + 212, + 378, + 787, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose Deep Burst Multi-scale SR using Fourier Space with Optical Flow (BurstM), which has a high representation capability and flexible scale SR $(\\times 2,\\times 3,\\times 4)$ with computational efficiency. BurstM adapts Optical Flow [16] as a solution for mitigating the limitations of DCN. Estimated optical flow provides correlated offsets between reference and source frames. In order to represent high-frequency textures using well-aligned information, our network estimates Fourier information. To address the fixed scale factor of MFSR, we employ INR [17] inspired by recent successful researches [21, 23].", + "bbox": [ + 212, + 590, + 789, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contribution is as follows:", + "bbox": [ + 238, + 710, + 566, + 726 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Our network conducts precise alignment using optical flow without DCN on the Fourier space.", + "- We solve the fixed-scale issue in MFSR and demonstrate our flexible networks of multi-scale SR.", + "- Extensive experiments demonstrate that our approach outperforms existing approaches in both image quality and computational efficiency." + ], + "bbox": [ + 225, + 729, + 782, + 819 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "EG Kang et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 388, + 160 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi Frame Super-Resolution (MFSR) The first MFSR from Tsai [43] restores high-resolution images in the frequency domain. The following approaches [1, 13, 14, 18, 33] have been introduced using the spatial domain and the maximum posterior. Deep learning-based MFSR approaches [3, 4, 11, 12, 29, 30] have arisen due to higher performance than the conventional methods. Furthermore, since the introduction of DCN [9] as the effective alignment method, many approaches [11, 12, 29, 30, 46] have adopted DCN because a biased position of DCN kernel contributes to outstanding alignment performance. However, the small kernel size and a fixed number of kernels restrict the alignment performance. A small kernel size $(3 \\times 3)$ of DCN has a small receptive field, which makes it impossible to cover global alignments. To address it, recent methods [11, 12, 29, 30] employ multi-layer DCN in varying spatial information, but it is not enough. Another issue is the fixed number of DCN kernels. Even if DCN cannot find the relevant information, DCN still needs to use the number of predefined kernels, leading to blurry and noisy images. Our BurstM employs optical flow to acquire a broader receptive field and utilize only one-to-one correlated information to tackle this issue.", + "bbox": [ + 212, + 205, + 787, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Optical Flow Optical flow is a popular method to describe the pixel-level correspondence of object displacements between frames. The conventional methods [15, 16, 42] address the data fidelity and the regularization term as an energy optimization problem. However, they often struggle to estimate accurate results in various scenarios, such as substantial object displacements or small-sized objects. Recent learning-based methods [35, 40, 41] outperform the conventional method's accuracy and computational cost using various structures such as hierarchical design and progressive estimation. Many tasks, such as video super-resolution [6, 22], action recognition [38], and autonomous driving [19], predict the optical flow to find pixel correspondences. However, MFSR task does not effectively utilize optical flow due to computational cost. On the other hand, the proposed network (BurstM) has a simple structure and lightweights, allowing us to incorporate an optical flow estimator into the network.", + "bbox": [ + 212, + 484, + 787, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Implicit Neural Representation (INR) One of the popular representation methods for the continuous-domain signal is INR [17]. Neural networks designed for INR learn and represent complex structures implicitly. The various tasks employ INR in image processes such as 2D and 3D shapes. Despite such success, INR parameterized by a standalone MLP with the ReLU has a spectral bias problem [32, 34, 39] that fails to represent high-frequency details known as a spectral bias. Recent approaches [8, 21, 23] tackle a spectral bias through positional encoding and Fourier feature mapping, demonstrating a flexible SR network. We have adopted INR method to utilize the flexible characteristic.", + "bbox": [ + 212, + 703, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "BurstM", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dc66da2460fb6f87ef3e7f26746873f6239f1226fab9ddf7dbb98475eeada02a.jpg", + "image_caption": [ + "Fig. 2: Proposed BurstM Network. The network estimates optical flow between the reference and source frames using FNet. The neural warping $\\mathbf{T}_{\\nu}$ emphasizes the high-frequency details and implements the warping on Fourier space. Reconstruction module $\\mathbf{R}_{\\varphi}$ includes the blender $\\mathbf{B}_{\\eta}$ and the decoder $\\mathbf{G}_{\\Theta}$ . The blender $\\mathbf{B}_{\\eta}$ matches color differences and merges multiple frames into a single frame. The Decoder $\\mathbf{G}_{\\Theta}$ predicts the final output. The gray color indicates a single operation such a convolution and LeakyReLU activation and the others include nonlinear activation" + ], + "image_footnote": [], + "bbox": [ + 220, + 142, + 787, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Problem Formulation", + "text_level": 1, + "bbox": [ + 215, + 415, + 459, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our purpose is to reconstruct the high-resolution image $\\mathbf{I}^{\\mathbf{SR}} \\in \\mathbb{R}^{3 \\times 2sH \\times 2sW}$ with multiple scale factors $s$ from the set of a low-resolution images $\\{\\mathbf{I}_i^{\\mathbf{L}}\\}_{i=1}^N$ , where $\\mathbf{I}_i^{\\mathbf{L}} \\in \\mathbb{R}^{4 \\times H \\times W}$ , $N$ is the number of frame and $H, W$ are the spatial sizes of low resolution image, and 4-channels are due to 'RGGB' domain. The dimension of a high-resolution image is increased by a factor of 2s, because the domain is changed from 4-channels 'RGGB' to 3-channels 'sRGB' in the network (in Figure 2).", + "bbox": [ + 212, + 440, + 787, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Optical Flow Estimator", + "text_level": 1, + "bbox": [ + 215, + 556, + 455, + 573 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We adopt FNet (optical flow estimator) of FRVSR [36]. FNet predicts optical flow on low-resolution grid as below:", + "bbox": [ + 212, + 582, + 785, + 612 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\pmb {\\delta} _ {i} ^ {\\mathbf {L}} = F N e t (\\mathbf {I} _ {i} ^ {\\mathbf {L}}), \\quad i = 1, \\dots , N, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 619, + 785, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\delta_{i}^{\\mathbf{L}}\\in \\mathbb{R}^{H\\times W\\times 2}$ is the set of optical flow offsets between a reference and source frames on low-resolution grid. Our model adds offsets $\\delta_{i}^{\\mathbf{L}}$ into regular coordinates $\\mathbf{x}_i^{\\mathbf{L}}\\in [-1,1]^{H\\times W\\times 2}$ and applies s scale upsampling.", + "bbox": [ + 212, + 650, + 787, + 699 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {i} ^ {\\mathbf {H}} = \\mathbf {B _ {s}} (\\pmb {\\delta} _ {i} ^ {\\mathbf {L}} + \\mathbf {x} _ {i} ^ {\\mathbf {L}}), \\quad i = 1, \\dots , N, \\qquad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 707, + 785, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{B}_{\\mathbf{s}}(\\cdot)$ is bilinear upsampling operator with a scale factor of s, and $\\hat{\\mathbf{x}}_i^{\\mathbf{H}} \\in \\mathbb{R}^{sH \\times sW \\times 2}$ is the upscaled coordinate with an offset defined by scale factor s.", + "bbox": [ + 212, + 734, + 785, + 767 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 Learnable Neural Warping", + "text_level": 1, + "bbox": [ + 215, + 776, + 482, + 791 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our encoder $(\\mathbf{E}_{\\psi}(\\cdot):\\mathbb{R}^4\\mapsto \\mathbb{R}^{2K})$ , which is a learnable super-resolution (SR) backbone [25] without upsampling, extracts latent vectors $\\mathbf{z}_i$ . NW $(\\mathbf{T}_{\\nu})$ takes the latent vectors $\\mathbf{z}_i$ for warping based on a Fourier information. NW consists", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "EG Kang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of three components: an amplitude estimator $(\\mathbf{g}_{\\mathbf{a}}(\\cdot):\\mathbb{R}^{2K}\\mapsto \\mathbb{R}^{2K})$ , a frequency estimator $(\\mathbf{g}_{\\mathbf{f}}(\\cdot):\\mathbb{R}^{2K}\\mapsto \\mathbb{R}^{2K})$ and, a phase estimator $(\\mathbf{g}_{\\mathbf{p}}(\\cdot):\\mathbb{R}^2\\mapsto \\mathbb{R}^K)$ . $2K$ denotes the number of channels in the latent vector. Amplitude and frequency estimators utilize the latent vector $\\mathbf{z}_i$ to estimate the Fourier coefficients. After that, the nearest interpolation $\\mathbf{N_s}$ is performed using estimated coordinates $\\hat{\\mathbf{x}}_i^{\\mathbf{L}}$ . The input of a phase estimator is a cell $\\mathbf{c}_i\\in \\mathbb{R}^{sH\\times sW\\times 2}$ , which is consist of the divisor of the input frame's spatial sizes $(1 / H,1 / W)$ . As in [8,20], we use relative coordinates $(\\delta_{i} = \\hat{\\mathbf{x}}_{i}^{\\mathrm{H}} - \\mathbf{x}_{i}^{\\mathrm{H}})$ known as a local grid to predict the continuous signal where $\\hat{\\mathbf{x}}_i^{\\mathrm{H}}\\in \\mathbb{R}^{sH\\times sW\\times 2}$ . The designed formulation is as follows:", + "bbox": [ + 212, + 145, + 787, + 282 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {i} = \\mathbf {E} _ {\\psi} \\left(\\mathbf {I} _ {i} ^ {\\mathbf {L}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 454, + 294, + 784, + 311 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {i} = \\mathbf {N} _ {\\mathbf {s}} \\left(\\mathbf {g} _ {\\mathbf {f}} \\left(\\mathbf {z} _ {i}\\right), \\hat {\\mathbf {x}} _ {i} ^ {\\mathbf {H}}\\right), \\quad \\mathbf {A} _ {i} = \\mathbf {N} _ {\\mathbf {s}} \\left(\\mathbf {g} _ {\\mathbf {a}} \\left(\\mathbf {z} _ {i}\\right), \\hat {\\mathbf {x}} _ {i} ^ {\\mathbf {H}}\\right), \\quad \\mathbf {p h} _ {i} = \\mathbf {g} _ {\\mathbf {p}} \\left(\\mathbf {c} _ {i}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 282, + 315, + 782, + 332 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {m} _ {i} = \\mathbf {A} _ {i} \\odot \\left[ \\begin{array}{l} \\cos (\\pi \\left(\\left\\langle \\mathbf {F} _ {i}, \\boldsymbol {\\delta} _ {i} \\right\\rangle + \\mathbf {p h} _ {i}\\right) \\\\ \\sin (\\pi \\left(\\left\\langle \\mathbf {F} _ {i}, \\boldsymbol {\\delta} _ {i} \\right\\rangle + \\mathbf {p h} _ {i}\\right) \\end{array} \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 335, + 782, + 368 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e ,} \\quad i = 1, \\dots , N, \\quad \\boldsymbol {\\nu} = [ \\mathbf {f}; \\mathbf {a}; \\mathbf {p} ].\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 372, + 630, + 387 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Reconstruction Module", + "text_level": 1, + "bbox": [ + 215, + 395, + 455, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To reconstruct the high-resolution image using estimated features in Fourier space, we employ the reconstruction module $\\mathbf{R}_{\\varphi}$ ; Blender $(\\mathbf{B}_{\\eta}(\\cdot):\\mathbb{R}^{N\\times 2K}\\mapsto \\mathbb{R}^{1\\times 2K})$ and MLP decoder $(\\mathbf{G}_{\\Theta}(\\cdot):\\mathbb{R}^{2K}\\mapsto \\mathbb{R}^{12})$ .", + "bbox": [ + 212, + 412, + 784, + 458 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} = \\mathbf {R} _ {\\varphi} \\left(\\left\\{\\mathbf {m} _ {i} \\right\\} _ {i = 1} ^ {N}\\right) = \\mathbf {G} _ {\\Theta} \\left(\\mathbf {B} _ {\\boldsymbol {\\eta}} \\left(\\left\\{\\mathbf {m} _ {i} \\right\\} _ {i = 1} ^ {N}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 467, + 784, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{h} \\in \\mathbb{R}^{1 \\times 12 \\times sH \\times sW}$ is the merged Fourier features of burst shots. The blender $\\mathbf{B}_{\\eta}$ , which is the same module with the encoder $\\mathbf{E}_{\\psi}$ , reduces color differences between frames and merges multiple frames. MLP decoder $(\\mathbf{G}_{\\Theta})$ estimates 'sRGB' information.", + "bbox": [ + 212, + 492, + 787, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4 Skip connection with upsampling", + "text_level": 1, + "bbox": [ + 214, + 561, + 535, + 577 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To prevent learning on the low-frequency only, we utilize the upscaled skip connection $(\\mathbf{Q}_{\\pmb{\\xi}}(\\cdot):\\mathbb{R}^{4\\times H\\times W}\\mapsto \\mathbb{R}^{12\\times sH\\times sW})$ about the reference frame $\\mathbf{I}_1^{\\mathbf{L}}$ . Finally, the spatial size is increased using pixel shuffle $(\\mathbf{P_2}(\\cdot):\\mathbb{R}^{12\\times sH\\times sW}\\mapsto \\mathbb{R}^{3\\times 2sH\\times 2sW})$ . The proposed BurstM can be formulated as follow:", + "bbox": [ + 212, + 579, + 787, + 638 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} ^ {\\mathbf {S R}} = \\mathbf {P} _ {2} (\\mathbf {h} + \\mathbf {Q} _ {\\xi} \\left(\\mathbf {I} _ {1} ^ {\\mathbf {L}}\\right)) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 648, + 784, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5 Loss function", + "text_level": 1, + "bbox": [ + 214, + 672, + 372, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use two loss functions for both SyntheticBurst [2] and BurstSR [3]. The MFSR datasets [2,3] does not provide the ground truth of the optical flow map between frames. Thus, we use L2 loss term to use the photometric loss.", + "bbox": [ + 212, + 689, + 785, + 736 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {O F}} = \\sum_ {i = 2} ^ {N} \\left\\| W a r p \\left(\\mathbf {I} _ {i} ^ {\\mathbf {L}}, \\boldsymbol {\\delta} _ {i} ^ {\\mathbf {L}}\\right) - \\mathbf {I} _ {1} ^ {\\mathbf {L}} \\right\\| _ {2} = \\sum_ {i = 2} ^ {N} \\left(W a r p \\left(\\mathbf {I} _ {i} ^ {\\mathbf {L}}, \\boldsymbol {\\delta} _ {i} ^ {\\mathbf {L}}\\right) - \\mathbf {I} _ {1} ^ {\\mathbf {L}}\\right) ^ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 279, + 744, + 784, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\text{Warp}(\\cdot, \\cdot)$ is the bilinear warping operator using optical flow from FNet, and $\\mathbf{I}_1^{\\mathbf{L}}$ is the reference frame.", + "bbox": [ + 212, + 794, + 784, + 825 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "BurstM", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SyntheticBurst dataset Input frames are generated from the ground truth. Thus, we update our model using L1 loss between predicted image and the ground truth.", + "bbox": [ + 212, + 146, + 782, + 190 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {S R}} = \\left\\| \\mathbf {I} ^ {\\mathbf {S R}} - \\mathbf {I} ^ {\\mathbf {G T}} \\right\\| _ {1}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 190, + 784, + 209 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Thus, the total loss function during training is as follows:", + "bbox": [ + 215, + 214, + 629, + 229 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {S y n t h e t i c}} = \\mathcal {L} _ {\\mathcal {S R}} + \\mathcal {L} _ {\\mathcal {O F}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 242, + 785, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BurstSR dataset We employs an aligned L1 loss for the BurstSR [3] dataset due to mismatching between low-resolution images and a ground truth image. An aligned L1 loss utilizes pre-trained weight of PwCNet [40] as follows:", + "bbox": [ + 214, + 268, + 782, + 315 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\delta} ^ {\\mathbf {A}} = P W C N e t \\left(\\mathbf {I} ^ {\\mathbf {S R}}, \\mathbf {I} ^ {\\mathbf {G T}}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 325, + 784, + 343 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {A l i g n e d S R}} = \\left\\| W a r p \\left(\\mathbf {I} ^ {\\mathbf {S R}}, \\boldsymbol {\\delta} ^ {\\mathbf {A}}\\right) - \\mathbf {I} ^ {\\mathbf {G T}} \\right\\| _ {1}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 347, + 784, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\delta^{\\mathbf{A}}$ is optical flow between the predicted image on high-resolution grid and ground truth using PwCNet [40]. Total loss function for BurstSR [3] dataset is as follows:", + "bbox": [ + 212, + 378, + 782, + 422 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {R e a l}} = \\mathcal {L} _ {\\text {A l i g n e d S R}} + \\mathcal {L} _ {\\mathcal {O F}}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 425, + 784, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 215, + 463, + 366, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Network details", + "text_level": 1, + "bbox": [ + 215, + 494, + 392, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Optical Flow Estimator (FNet) Our network adapts FNet from FRVSR [36]. To apply FNet into BurstSR task, we modify the stem of FNet from 'sRGB' to 'RGGB' images.", + "bbox": [ + 214, + 518, + 782, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Learnable Neural Warping $(\\mathbf{T}_{\\nu})$ NW estimates the amplitude, frequency, and phase from burst shots with a single convolution operation to represent the Fourier information. Amplitude and frequency estimators have 128 channels of the convolutions with $3\\times 3$ kernel, phase estimator has 64 channels of the convolutions with $1\\times 1$ kernel.", + "bbox": [ + 214, + 580, + 784, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Encoder $(\\mathbf{E}_{\\psi})$ and Reconstruction module $(\\mathbf{R}_{\\varphi})$ The reconstruction module consists of Blender $\\mathbf{B}_{\\eta}$ and MLP decoder $\\mathbf{G}_{\\Theta}$ . Both Encoder $\\mathbf{E}_{\\psi}$ and Blender $\\mathbf{B}_{\\eta}$ are EDSR [25] without upsampling module and 128 channels. MLP Decoder, $\\mathbf{G}_{\\Theta}$ , is organized by 4-layer of $1\\times 1$ convolution with 256 channels and ReLU activation.", + "bbox": [ + 214, + 671, + 784, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Implementation details", + "text_level": 1, + "bbox": [ + 215, + 768, + 452, + 784 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BurstM is an end-to-end trainable network and is possible to train without pre-trained weights. We apply an AdamW optimizer [28] with the Cosine annealing strategy [27] from $10^{-4}$ to $10^{-6}$ learning rate. For SyntheticBurst [2], we", + "bbox": [ + 212, + 794, + 782, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "EG Kang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "train BurstM with L1 loss for the output image and L2 loss for optical flows. For BurstSR [3], We fine-tune the pre-trained BurstM of SyntheticBurst with L2 loss for optical flow and aligned L1 loss [3] for the output image. Aligned L1 loss utilizes PWCNet [40] for the output image. The reason for the usage of the aligned L1 Loss is that there are no LR images aligned with the GT image because they were captured using different cameras. We train our model using four RTX3090 GPUs and training times are 10 days for SyntheticBurst and 8 hours for BurstSR. The summary of the training strategy is in Table 1.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3f793a559396feb8bc60202b5b3117387bfa06d8b30914c71861aa06cdf28624.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SyntheticBurst [2]BurstSR [3]
Training Epoch30025
OptimizerADAMW [28]
Learning rate10-4~10-6
LR schedulerCosine annealing [27]
Batch sizes4
Main LossL1Aligned L1
Photo metric LossL2
×2:96×96
Input sizes×3:64×64×4:80×80
×4:48×48
Training time10 days8 hours
", + "bbox": [ + 331, + 268, + 671, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1: Training Strategies. Training epoch and main loss, input sizes are different. Except for these factors, we use same training strategies for both SyntheticBurst [2] for BurstSR [3] datasets.", + "bbox": [ + 212, + 441, + 787, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SyntheticBurst Dataset [2] The dataset contains burst frames which are 46,839 set for training and 300 set for validation. Each burst set contains 14 low-resolution burst frames, which are generated from 'sRGB' as follows. We utilized the unprocessing [5]. Subsequently, we apply random rotations and translations. We employ random down-sampling $(\\times 2, \\times 3, \\times 4)$ to generate low-resolution images of multiple scales. Finally, we apply Bayer mosaicking and add random noise. We trained the proposed model with batch size 4 during 300 epochs, and the training patch size is changed per scale factors in Table 1.", + "bbox": [ + 212, + 498, + 787, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "BurstSR Dataset [3] Input and GT are captured by different cameras: Smartphone and DSLR, respectively. Both low and high resolution images have 200 raw burst sets. The low-resolution set contains 14 frames, and we extract 5,405 patches for training and 882 patches for validation. Unlike SyntehticBurst dataset [2], the scale factor between low-resolution and high-resolution is fixed. Thus, we train BurstM only for $\\times 4$ scale factor. Nevertheless, BurstM can accurately predict high-resolution images for multiple scale factors ( $\\times 2$ , $\\times 3$ , $\\times 4$ ), including clear representation of high-frequency textures without the boundary artifacts.", + "bbox": [ + 212, + 633, + 810, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Validation Results", + "text_level": 1, + "bbox": [ + 214, + 773, + 415, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show the qualitative comparison with the existing methods [3, 4, 11, 12, 29, 30] and the visual comparison with latest three methods [11, 12, 29] using the pre-trained weights from authors.", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "BurstM", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7d5af780815ec8c19bb7c61b3cbcd53de6ead3506849820d62afa153119b50fd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
# Params. MSyntheticBurstBurstSR x4 PSNR SSIM
x2 PSNRSSIMx3 PSNRSSIMx4 PSNR
Bicubic-38.300.94833.940.88633.020.862
DBSR [3]13.0140.510.96540.110.95940.760.959
MFIR [4]12.1341.250.97141.810.97241.560.964
BIPNet†[11]6.737.580.92840.830.95541.930.960
GMTNet†[30]-----42.360.961
Burststormer†[12]2.537.060.92540.260.95342.830.973
BSRT-Small†[29]4.9240.640.96642.300.97542.720.971
BSRT-Large†[29]20.7140.330.96542.870.97943.620.975
BurstM (Ours)14.046.010.98544.790.98242.870.973
", + "bbox": [ + 228, + 143, + 776, + 318 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2: Quantitative comparison on SyntheticBurst and BurstSR. Quantitative comparisons of BurstSR such as PSNR and SSIM can only be measured for $\\times 4$ , due to LR and HR being provided in fixed sizes. Red and Blue colors indicate the best and the second performance, respectively. $\\dagger$ indicates Transformer-based method.", + "bbox": [ + 215, + 319, + 784, + 373 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/574d07622ca736affaae30f009aa0ec1b7e009554ab9ca5474254a254d5ddbf6.jpg", + "image_caption": [ + "Fig. 3: Visual comparison for $\\times 4$ super-resolution for BurstSR. BurstM shows clearer textures and correct color than the others without the boundary artifacts." + ], + "image_footnote": [], + "bbox": [ + 222, + 380, + 779, + 804 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "EG Kang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SR on BurstSR [3] Table 2 is the quantitative comparison result of $\\times 4$ SR on BurstSR [3]. Ours shows a performance gain of $+0.17\\mathrm{dB}$ compared to the state-of-the-art baselines. As shown in Figure 3, BurstM predicts outperform high-frequency details without the boundary artifacts from low-resolution images of real-world dataset. Also, BurstM predicts multi-scale SR without an additional processing, different with existing models. However, the quantitative comparison of $\\times 2$ , $\\times 3$ is not possible, because there is no GT of both $\\times 2$ , $\\times 3$ SR of given dataset [3]. We demonstrate the robustness of our approach by representing high-frequency details in Figure 4 and computation time in Table 4. BurstM directly predicts multiple scales, while other methods implement $\\times 4$ SR followed by bicubic downsampling. In Table 4, BurstM shows shorter computation time than the existing methods.", + "bbox": [ + 212, + 146, + 787, + 325 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f42d53edad15dd700baa4f107b1b5759f6d8e8ab8fc757ca4eee07ce05b0254c.jpg", + "image_caption": [ + "Fig. 4: Visual comparison of multi-scale SR on BurstSR. BurstM predict several scaled-images with unimodel and the others are downsampling after $\\times 4$ SR." + ], + "image_footnote": [], + "bbox": [ + 228, + 327, + 776, + 518 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "SR on SyntheticBurst [2] Table 2 is the quantitative comparison results of multiple scale factors $(\\times 2, \\times 3, \\times 4)$ SR and Figure 5 is the visual comparison result of $\\times 4$ SR on SyntheticBurst [2]. Our model achieves second result with reasonable the number of parameters. In addition, we compare the performance on multiple scale factors in Table 2 and Figure 6. BurstM predicts high-resolution images on multiple scale factors with an unimodel and the others implement multiple SR with additional bicubic downsampling after $\\times 4$ SR. In case of BurstM, it is observed that PSNR gradually increases as the scale factor decrease. However, PSNR of existing methods decreases as the scale factor decreases because of the information loss during the additional bicubic downsampling process. Figure 6 shows the residual images, which means the loss of the high-frequency textures. Furthermore, as shown in Table 4, BurstM outperforms the other methods in terms of computation time.", + "bbox": [ + 212, + 551, + 787, + 750 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Fourier feature space Figure 7 demonstrates the estimation of high-frequency details on Fourier space. The result from each frame do not illustrate the DFT spectrum, but accumulated result of all frames shows a similar result with the DFT spectrum. Because BurstM leverages sub-pixel information from burst frames to represent the high-resolution images.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "BurstM", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a43c4c5dab9a0ffbb77b28c856b37399c91cb0fc4343b6c985c9fcccfc683345.jpg", + "image_caption": [ + "Fig. 5: Visual comparison for $\\times 4$ Super-Resolution for SynteticBurst." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 782, + 364 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/53409e3e489013298822216c8c9c7c94da78f30da1d6b313242a8ebbd441d37d.jpg", + "image_caption": [ + "Fig.6: Visual comparison for multi-scale SR on SyntheticBurst. The colorbar indicates a difference with GT. The existing methods lose high-frequency textures during additional down-sampling because the highlighted locations are a high-frequency area. But BurstM represent the similar high-frequency textures with GT by because of direct prediction on multiple scale factors." + ], + "image_footnote": [], + "bbox": [ + 212, + 391, + 787, + 661 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Aligning performances We conduct a comparative analysis of the alignment performance between DCN-based method and BurstM in Figure 8. As in [11, 12, 29, 30], we utilize multi-layer DCN with varied spatial sizes for DCN-based method. In contrast, BurstM implements a single image warping process using predicted optical flows. The 'Before align.' and 'After align.' residual maps in Figure 8 represent the differences with the reference frame at each position in", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "EG Kang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/1545ca80d0bf01163304d68d8226c7125c00d978fc38aef0d0ba0a967417df3c.jpg", + "image_caption": [ + "Fig.7: Fourier space analysis. Accumulated frequencies and coefficients from BurstM are analogous to DFT spectrum of GT (Red color in 'DFT' and 'ACCUM')." + ], + "image_footnote": [], + "bbox": [ + 240, + 142, + 759, + 368 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "flowchart. The red dots in 'DCN method flowchart' illustrate offsets to the center pixel. DCN uses an unrelated pixel information, because DCN requires to use the number of predefined kernels regardless of the correlation. Therefore, the aligned results are inaccurate and show high values in the residual map. Additionally, boundary artifacts are also observed due to the utilization of irrelevant information. The color map and white arrows in 'BurstM method flowchart' are the visualization results of optical flow. BurstM method shows one-to-one correspondences between pixels. Thus, the alignment results demonstrate clearer and sharper compared to the DCN-based method without the boundary artifacts.", + "bbox": [ + 212, + 402, + 787, + 539 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3754588c6c75114624fc2e0033f7944a0bef59938eb6d9bae41a23ba4be0a3b6.jpg", + "image_caption": [ + "DCN method flowchart [12] BurstM method (ours) flowchart" + ], + "image_footnote": [], + "bbox": [ + 256, + 542, + 493, + 613 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3587067fb3f61041baadd4da369409c8f6af3838944346d9d1d68a842f92bbf4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 542, + 735, + 613 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/17c08bbf5c1850793e2a9df5dcce235b3a53b72476b06336a44a4e1c8b8522dc.jpg", + "image_caption": [ + "Fig. 8: Feature alignment results. Flow charts indicate where we take features. The residual map is between the reference and source frames at each location in flowchart. DCN based-method has a big difference after an alignment because DCN take unrelated information during alignment. However, BurstM point out one-to-one correspondences between a reference and source frames. BurstM shows accurate align result without the boundary artifacts." + ], + "image_footnote": [], + "bbox": [ + 217, + 632, + 782, + 768 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "BurstM", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5 Ablation study", + "text_level": 1, + "bbox": [ + 215, + 143, + 397, + 162 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The comparison between with and without Fourier mapping INR [17] is the most representative method for representing continuous signals, but it has a spectral bias [32, 34, 39]. As shown in [21, 23], BurstM also addresses this issue through Fourier mapping. Table 3a compares the performance of BurstM with and without Fourier mapping. BurstM without Fourier mapping has significantly lower PSNR values. This is because the network without Fourier mapping learns more about low frequencies due to spectral bias during training.", + "bbox": [ + 212, + 165, + 787, + 272 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Relationship between flexibility and performance BurstM is supporting the multiple scale SR $(\\times 2, \\times 3, \\times 4)$ . The performance in Table 2 has trained with three multiple scale factors $(\\times 2, \\times 3, \\times 4)$ . We train with different number of scale factors such as one $(\\times 4)$ and two $(\\times 3, \\times 4)$ items. In such cases, we observed that PSNR increases on in-of-scale, but performance decreases on out-of-scale. When there are many instances of out-of-scale, the model's flexibility diminishes, so we proposed three scales satisfy both performance and flexibility.", + "bbox": [ + 212, + 286, + 787, + 393 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Different optical flow estimator While FNet [36] estimates an optical flow based on given image sizes, the other optical flow estimator [35] adjust image sizes according to its setting before input. This resizing of images introduces aliasing effects, resulting in a decline in SR performance. Additionally, BurstM with Spynet [35] shows longer inference time than the proposed method.", + "bbox": [ + 212, + 407, + 787, + 482 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/5bc604ef10517e9566840676a18f955282f4bd67a5a5060012c6c1843b21f5b7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BurstM w\\o Fourier mappingBurstM w\\Fourier mapping
×4 PSNR39.82 dB41.74 dB
", + "bbox": [ + 248, + 482, + 753, + 512 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/23191e0921df20aaca77a54952d99a3c61ec6ee4e7958065074b1d1b714406f1.jpg", + "table_caption": [ + "(a) The performance comparison between with and without Fourier mapping" + ], + "table_footnote": [], + "table_body": "
Training scale×2×3×4
×433.3635.9742.26
×3, ×430.8043.6141.82
×2, ×3, ×4 (Ours)44.7743.4541.74
", + "bbox": [ + 235, + 531, + 483, + 593 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/3dda7a082221a3b41ed67ff9e7396e756314bc8697654c72bd3d8e5948a1552c.jpg", + "table_caption": [ + "(b) Relationship between flexibility and performance (PSNR:dB)" + ], + "table_footnote": [], + "table_body": "
BurstM\nw\\SpynetBurstM\nw\\FNet (Ours)
×4 PSNR40.85 dB41.74 dB
Inference time14.3 ms11.6 ms
", + "bbox": [ + 491, + 531, + 782, + 592 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "(c) Performance comparison based on optical flow estimator", + "bbox": [ + 544, + 594, + 725, + 616 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 3: Ablation study summary", + "bbox": [ + 215, + 622, + 475, + 636 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(a) Fourier mapping contributes to the model performance.", + "(b) Red color indicates the result of in-scale. Less training scale has higher PSNR on in-scale but network flexibility is low due to increasing out-scale. Thus, we select our model with 3 scale factors $(\\times 2, \\times 3, \\times 4)$ considering both flexibility and performance.", + "(c) A different optical flow estimator leads to PSNR drop and an increase in inference time." + ], + "bbox": [ + 215, + 636, + 785, + 717 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6 Discussion", + "text_level": 1, + "bbox": [ + 215, + 727, + 354, + 742 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Boundary line artifacts As shown in Figure 9, the recent method [29] exhibits boundary artifacts despite utilizing optical flow, due to its continued reliance on DCN [9] for alignment. As illustrated in Figure 8 and 9, DCN [9] must utilize information corresponding to the predefined kernel size regardless of correlation, resulting in the incorporation of unnecessary information when most frames fail to provide essential data. In contrast, BurstM aligns frames without DCN [9],", + "bbox": [ + 212, + 748, + 787, + 842 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "EG Kang et al.", + "bbox": [ + 271, + 114, + 374, + 127 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "solely utilizing optical flow. Consequently, BurstM avoids the use of irrelevant information, leading to accurate predictions without boundary artifacts.", + "bbox": [ + 214, + 146, + 782, + 176 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ee27cf3636ffa531cec16670c108357ef31b6b4202b602606ccbb9d97ef6c4a7.jpg", + "image_caption": [ + "Fig.9: Boundary artifacts The most of frames didn't provide the information of edge. The existing method [29] suffers the boundary line artifacts, but BurstM represents clear images without the boundary line artifacts." + ], + "image_footnote": [], + "bbox": [ + 225, + 181, + 763, + 385 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Computation complexity Despite advancements in computational system, SR field still requires shorter computation times as the image resolutions increase. Table 4 compares the computation time usage of BurstM with the other methods [11,12,29]. We implements $\\times 4$ SR with $80 \\times 80$ frame sizes of BurstSR dataset [3] on NVIDIA RTX3090 24GB. While BurstM has a large number of parameters, BurstM demonstrates the shortest computation time.", + "bbox": [ + 212, + 436, + 787, + 527 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/bbea5824731e2b6578ed494a066e72cc7506d6a72027f753db81fd457d0b8e56.jpg", + "image_caption": [ + "Table 4: Computational complexity on one GPU (RTX3090). Despite a large number of parameters, BurstM shows the inference time (ms) and best PSNR (dB) performance. Red : best metrics. † indicates Transformer-based method." + ], + "image_footnote": [], + "bbox": [ + 217, + 527, + 539, + 662 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1d9a5ec120abbd17640dd932ecf6a8851a0eae27a1facf8961c8143513eca434.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SR scale: ×4Params.(M) / Mem.(GB)Inference time(ms)PSNR (dB)
BIPNet† [11]6.7 / 11.820.848.49
Burstormer† [12]2.5 / 12.020.848.82
BSRT-S† [29]4.92 / 1.0148.048.48
BSRT-L† [29]20.71 / 1.5266.148.57
BurstM (Ours)14.9 / 8.411.649.12
", + "bbox": [ + 540, + 527, + 792, + 659 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Super-Resolution with floating scale factors BurstM can estimate the multi-scale SR with floating scale factors such as $\\times 3.5$ without any architecture changes. However, the training of the floating scale factors ( $\\times 2\\sim \\times 4$ ) shows lower PSNR (39.44dB) than the training of the integer scale factors (41.74dB). The performance drop is caused by the demosaic process when generating low-resolution frames. To convert from 'sRGB' to 'RGGB', the dimensions (H, W) of the images are changed to (H/2, W/2). If the image size of 'sRGB' is not divisible by two when a floating scale factor is applied, the network apply an additional", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "BurstM", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "downsampling operation to make the dimension of $(\\mathrm{H} / 2, \\mathrm{W} / 2)$ . The double downsampling imposes a severe degradation of the generated low-resolution images, and it results in low PSNRs.", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "**Corner case** The major challenge on optical flow is occlusions which are caused by a new object that is not present in the reference frame. BurstM also has the occlusion issue (Figure 10) even though BurstM demonstrate accurate alignment performance than DCN based methods. Fortunately, BurstM utilize inaccurate optical flow of an occlusion as meaningful prior and reconstruct the high-quality images in a high-resolution grid.", + "bbox": [ + 212, + 196, + 787, + 287 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/04b3ada32e3a25a2b1902f784b7241a3a5860dc8209a631f673bb473c0f85768.jpg", + "image_caption": [ + "Fig. 10: Corner case. Inaccurate optical flow prediction due to occlusions which are caused by new objects (red arrow) not present in the reference image appear." + ], + "image_footnote": [], + "bbox": [ + 243, + 287, + 759, + 521 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 560, + 359, + 575 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we proposed Deep Burst Multi-scale SR using Fourier Space with Optical Flow (BurstM) to overcome the limitations of existing methods: SR of fixed scale factor $(\\times 4)$ and low-quality of real-world dataset [3]. Our network performs the precise alignment using optical flow and the advanced warping process based on estimated Fourier information. We demonstrated their effectiveness in representing the high-frequency textures with $+0.17\\mathrm{dB}$ PSNR gain in the real-world dataset [3] compared to the previous state-of-the-art model. Furthermore, BurstM successfully implements multi-scale SR $(\\times 2,\\times 3,\\times 4)$ of the unimodel by representing high-frequency details using the INR method. The function of multi-scale SR enhances our model's capacity compared to the other learning-based methods. In addition, we demonstrate our approach is faster inference time compared to the existing methods by the efficient computation.", + "bbox": [ + 212, + 583, + 787, + 763 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (RS-2021-II212068, Artificial Intelligence Innovation Hub) and the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (RS-2024-00335741).", + "bbox": [ + 212, + 765, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "EG Kang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bascle, B., Blake, A., Zisserman, A.: Motion deblurring and super-resolution from an image sequence. In: Computer Vision—ECCV'96: 4th European Conference on Computer Vision Cambridge, UK, April 15–18, 1996 Proceedings Volume II 4. pp. 571–582. Springer (1996)", + "2. Bhat, G., Danelljan, M., Timofte, R.: Ntire 2021 challenge on burst superresolution: Methods and results. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 613-626 (2021)", + "3. Bhat, G., Danelljan, M., Van Gool, L., Timofte, R.: Deep burst super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9209-9218 (2021)", + "4. Bhat, G., Danelljan, M., Yu, F., Van Gool, L., Timofte, R.: Deep reparametrization of multi-frame super-resolution and denoising. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2460-2470 (2021)", + "5. Brooks, T., Mildenhall, B., Xue, T., Chen, J., Sharlet, D., Barron, J.T.: Unprocessing images for learned raw denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11036-11045 (2019)", + "6. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basicvsr: The search for essential components in video super-resolution and beyond. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4947-4956 (2021)", + "7. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-Trained Image Processing Transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12299-12310 (2021)", + "8. Chen, Y., Liu, S., Wang, X.: Learning continuous image representation with local implicit image function. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8628-8638 (2021)", + "9. Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., Wei, Y.: Deformable convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 764-773 (2017)", + "0. Dai, T., Cai, J., Zhang, Y., Xia, S.T., Zhang, L.: Second-order attention network for single image super-resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11065-11074 (2019)", + "1. Dudhane, A., Zamir, S.W., Khan, S., Khan, F.S., Yang, M.H.: Burst image restoration and enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5759-5768 (2022)", + "2. Dudhane, A., Zamir, S.W., Khan, S., Khan, F.S., Yang, M.H.: Burstomer: Burst image restoration and enhancement transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5703-5712 (2023)", + "3. Elad, M., Feuer, A.: Restoration of a single superresolution image from several blurred, noisy, and undersampled measured images. IEEE transactions on image processing 6(12), 1646-1658 (1997)", + "4. Hardie, R.C., Barnard, K.J., Bognar, J.G., Armstrong, E.E., Watson, E.A.: High-resolution image reconstruction from a sequence of rotated and translated frames and its application to an infrared imaging system. Optical Engineering 37(1), 247-260 (1998)", + "5. Hirschmuller, H.: Stereo processing by semiglobal matching and mutual information. IEEE Transactions on pattern analysis and machine intelligence 30(2), 328-341 (2007)" + ], + "bbox": [ + 225, + 172, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "BurstM", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Horn, B.K., Schunck, B.G.: Determining optical flow. Artificial intelligence 17(1-3), 185-203 (1981)", + "17. Hornik, K., Stinchcombe, M., White, H.: Multilayer feedforward networks are universal approximators. Neural networks 2(5), 359-366 (1989)", + "18. Irani, M., Peleg, S.: Improving resolution by image registration. CVGIP: Graphical models and image processing 53(3), 231-239 (1991)", + "19. Janai, J., Güney, F., Behl, A., Geiger, A., et al.: Computer vision for autonomous vehicles: Problems, datasets and state of the art. Foundations and Trends in Computer Graphics and Vision 12(1-3), 1-308 (2020)", + "20. Jiang, C.M., Sud, A., Makadia, A., Huang, J., Niessner, M., Funkhouser, T.: Local Implicit Grid Representations for 3D Scenes. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)", + "21. Kim, M., Lee, J., Lee, B., Im, S., Jin, K.H.: Implicit neural image stitching with enhanced and blended feature reconstruction (2023)", + "22. Lai, W.S., Huang, J.B., Ahuja, N., Yang, M.H.: Deep laplacian pyramid networks for fast and accurate super-resolution. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 624-632 (2017)", + "23. Lee, J., Jin, K.H.: Local texture estimator for implicit representation function. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1929-1938 (2022)", + "24. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: SwinIR: Image Restoration Using Swin Transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 1833-1844 (2021)", + "25. Lim, B., Son, S., Kim, H., Nah, S., Mu Lee, K.: Enhanced deep residual networks for single image super-resolution. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 136-144 (2017)", + "26. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10012-10022 (2021)", + "27. Loshchilov, I., Hutter, F.: Stochastic gradient descent with warm restarts. In: Proceedings of the 5th Int. Conf. Learning Representations. pp. 1-16", + "28. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)", + "29. Luo, Z., Li, Y., Cheng, S., Yu, L., Wu, Q., Wen, Z., Fan, H., Sun, J., Liu, S.: BSRT: Improving burst super-resolution with swim transformer and flow-guided deformable alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 998-1008 (2022)", + "30. Mehta, N., Dudhane, A., Murala, S., Zamir, S.W., Khan, S., Khan, F.S.: Gated multi-resolution transfer network for burst restoration and enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22201-22210 (2023)", + "31. Mei, Y., Fan, Y., Zhou, Y.: Image super-resolution with non-local sparse attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3517-3526 (2021)", + "32. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)", + "33. Peleg, S., Keren, D., Schweitzer, L.: Improving image resolution using subpixel motion. Pattern recognition letters 5(3), 223-226 (1987)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "EG Kang et al.", + "bbox": [ + 271, + 114, + 374, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Rahaman, N., Baratin, A., Arpit, D., Draxler, F., Lin, M., Hamprecht, F., Bengio, Y., Courville, A.: On the Spectral Bias of Neural Networks. In: Chaudhuri, K., Salakhutdinov, R. (eds.) Proceedings of the 36th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 97, pp. 5301-5310. PMLR (2019)", + "35. Ranjan, A., Black, M.J.: Optical flow estimation using a spatial pyramid network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4161-4170 (2017)", + "36. Sajjadi, M.S., Vemulapalli, R., Brown, M.: Frame-recurrent video super-resolution. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6626-6634 (2018)", + "37. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1874-1883 (2016)", + "38. Simonyan, K., Zisserman, A.: Two-stream convolutional networks for action recognition in videos. Advances in neural information processing systems 27 (2014)", + "39. Sitzmann, V., Martel, J., Bergman, A., Lindell, D., Wetzstein, G.: Implicit Neural Representations with Periodic Activation Functions. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M.F., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 7462-7473. Curran Associates, Inc. (2020)", + "40. Sun, D., Yang, X., Liu, M.Y., Kautz, J.: Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8934-8943 (2018)", + "41. Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 402-419. Springer (2020)", + "42. Triggs, B., McLauchlan, P.F., Hartley, R.I., Fitzgibbon, A.W.: Bundle adjustment—a modern synthesis. In: Vision Algorithms: Theory and Practice: International Workshop on Vision Algorithms Corfu, Greece, September 21–22, 1999 Proceedings. pp. 298–372. Springer (2000)", + "43. Tsai, R.Y., Huang, T.S.: Multiframe image restoration and registration. Multiframe image restoration and registration 1, 317-339 (1984)", + "44. Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image super-resolution using very deep residual channel attention networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 286-301 (2018)", + "45. Zhang, Y., Tian, Y., Kong, Y., Zhong, B., Fu, Y.: Residual dense network for image super-resolution. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2472-2481 (2018)", + "46. Zhang, Z., Wang, R., Zhang, H., Chen, Y., Zuo, W.: Self-supervised learning for real-world super-resolution from dual zoomed observations. In: European Conference on Computer Vision. pp. 610-627. Springer (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 729 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "BurstM", + "bbox": [ + 676, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/BurstM_ Deep Burst Multi-scale SR using Fourier Space with Optical Flow/608ca9aa-ec95-4beb-807e-424a93d0d485_model.json b/2024/BurstM_ Deep Burst Multi-scale SR using Fourier Space with Optical Flow/608ca9aa-ec95-4beb-807e-424a93d0d485_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ce331f3e958affd66e3a9e40750008a5a4870d37 --- /dev/null +++ b/2024/BurstM_ Deep Burst Multi-scale SR using Fourier Space with Optical Flow/608ca9aa-ec95-4beb-807e-424a93d0d485_model.json @@ -0,0 +1,2368 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.141, + 0.787, + 0.187 + ], + "angle": 0, + "content": "BurstM: Deep Burst Multi-scale SR using Fourier Space with Optical Flow" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.212, + 0.786, + 0.229 + ], + "angle": 0, + "content": "EungGu Kang¹ Byeonghun Lee² Sunghoon Im1* Kyong Hwan Jin2*" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.239, + 0.747, + 0.253 + ], + "angle": 0, + "content": "1 Daegu Gyeongbuk Institute of Science and Technology (DGIST), Korea" + }, + { + "type": "text", + "bbox": [ + 0.415, + 0.254, + 0.59, + 0.268 + ], + "angle": 0, + "content": "2 Korea University, Korea" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.269, + 0.735, + 0.282 + ], + "angle": 0, + "content": "eunggukang@gmail.com, sunghoonim@dgist.ac.kr, {byeonghun_lee," + }, + { + "type": "text", + "bbox": [ + 0.415, + 0.284, + 0.588, + 0.296 + ], + "angle": 0, + "content": "kyong_jin}@korea.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.298, + 0.741, + 0.533 + ], + "angle": 0, + "content": "Abstract. Multi frame super-resolution (MFSR) achieves higher performance than single image super-resolution (SISR), because MFSR leverages abundant information from multiple frames. Recent MFSR approaches adapt the deformable convolution network (DCN) to align the frames. However, the existing MFSR suffers from misalignments between the reference and source frames due to the limitations of DCN, such as small receptive fields and the predefined number of kernels. From these problems, existing MFSR approaches struggle to represent high-frequency information. To this end, we propose Deep Burst Multi-scale SR using Fourier Space with Optical Flow (BurstM). The proposed method estimates the optical flow offset for accurate alignment and predicts the continuous Fourier coefficient of each frame for representing high-frequency textures. In addition, we have enhanced the network's flexibility by supporting various super-resolution (SR) scale factors with the unimodel. We demonstrate that our method has the highest performance and flexibility than the existing MFSR methods. Our source code is available at https://github.com/Egkang-Luis/burstm" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.546, + 0.741, + 0.573 + ], + "angle": 0, + "content": "Keywords: Super resolution \\(\\cdot\\) Burst super resolution \\(\\cdot\\) Implicit neural representation \\(\\cdot\\) Fourier features" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.583, + 0.377, + 0.599 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.787, + 0.698 + ], + "angle": 0, + "content": "Single image super-resolution (SISR) is the major ill-posed problem in low-level computer vision. SISR aims to enhance the image quality of low-resolution by estimating sub-pixel information. After the success of deep learning, the most SISR [10,25,31,44,45] adapt the learning-based method to estimate high-resolution images. However, SISR has a limitation in high-frequency representation of the capacity due to a lack of information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.698, + 0.787, + 0.819 + ], + "angle": 0, + "content": "Super-resolution researchers focus on multi-frame super-resolution (MFSR) due to the accurate representation of high-frequency textures. Multiple frames provide additional sub-pixel priors, and MFSR utilizes rich information with alignment between the reference and source frames. However, multiple frames take similar scenes with degradations such as noise and object movement. MFSR requires the accurate alignment performance to utilize low-quality frames, because the inaccurate alignment performance can lead to the blur artifacts in the prediction result. Thus, the alignment process is the critical challenge of MFSR." + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.411, + 0.84 + ], + "angle": 0, + "content": "\\* Co-corresponding authors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "EG Kang et al." + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.146, + 0.703, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.312, + 0.788, + 0.369 + ], + "angle": 0, + "content": "Fig. 1: Overview of proposed Burst SR network. The proposed method (BurstM) predicts high-resolution images from multiple low-resolution images. The neural warping performs high-precision warping with upscaling on Fourier space. Additionally, the neural warping contributes to the prediction of an accurate high-resolution images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.789, + 0.59 + ], + "angle": 0, + "content": "Like SISR, the latest MFSR [3, 4, 11, 12, 29, 30] employs a deep learning method. They adapt similar architectures, which is complicated feature extraction methods using CNN or transformer [7, 24, 26] and alignment using DCN [9] to reconstruct the high-resolution images. Recent MFSR [11, 12, 29, 30] adapts multi-layer DCN structures to extend a small receptive field of DCN, but even this approach still has limited receptive fields. In addition, the predefined number of DCN kernels restricts representing performance during aligning features. The fixed SR scale is another limitation of existing MFSR due to their upsampling method. Most networks [11, 12, 29, 30] generally adapt pixel shuffle (PS) [37] as an upsampling module due to the effectiveness of PS. Nevertheless, PS restricts the network's flexibility, e.g., \\(\\times 4\\) SR, because PS can conduct upsampling by a predetermined scale, which is the square of the channel number. If the network predicts a different SR scale, it must re-train with changing architecture or additional downsampling after a predefined SR scale." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.591, + 0.79, + 0.712 + ], + "angle": 0, + "content": "To this end, we propose Deep Burst Multi-scale SR using Fourier Space with Optical Flow (BurstM), which has a high representation capability and flexible scale SR \\((\\times 2,\\times 3,\\times 4)\\) with computational efficiency. BurstM adapts Optical Flow [16] as a solution for mitigating the limitations of DCN. Estimated optical flow provides correlated offsets between reference and source frames. In order to represent high-frequency textures using well-aligned information, our network estimates Fourier information. To address the fixed scale factor of MFSR, we employ INR [17] inspired by recent successful researches [21, 23]." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.712, + 0.568, + 0.727 + ], + "angle": 0, + "content": "To summarize, our contribution is as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.731, + 0.784, + 0.759 + ], + "angle": 0, + "content": "- Our network conducts precise alignment using optical flow without DCN on the Fourier space." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.761, + 0.784, + 0.788 + ], + "angle": 0, + "content": "- We solve the fixed-scale issue in MFSR and demonstrate our flexible networks of multi-scale SR." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.792, + 0.784, + 0.82 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that our approach outperforms existing approaches in both image quality and computational efficiency." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.731, + 0.784, + 0.82 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.678, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BurstM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.39, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.206, + 0.788, + 0.465 + ], + "angle": 0, + "content": "Multi Frame Super-Resolution (MFSR) The first MFSR from Tsai [43] restores high-resolution images in the frequency domain. The following approaches [1, 13, 14, 18, 33] have been introduced using the spatial domain and the maximum posterior. Deep learning-based MFSR approaches [3, 4, 11, 12, 29, 30] have arisen due to higher performance than the conventional methods. Furthermore, since the introduction of DCN [9] as the effective alignment method, many approaches [11, 12, 29, 30, 46] have adopted DCN because a biased position of DCN kernel contributes to outstanding alignment performance. However, the small kernel size and a fixed number of kernels restrict the alignment performance. A small kernel size \\((3 \\times 3)\\) of DCN has a small receptive field, which makes it impossible to cover global alignments. To address it, recent methods [11, 12, 29, 30] employ multi-layer DCN in varying spatial information, but it is not enough. Another issue is the fixed number of DCN kernels. Even if DCN cannot find the relevant information, DCN still needs to use the number of predefined kernels, leading to blurry and noisy images. Our BurstM employs optical flow to acquire a broader receptive field and utilize only one-to-one correlated information to tackle this issue." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.788, + 0.683 + ], + "angle": 0, + "content": "Optical Flow Optical flow is a popular method to describe the pixel-level correspondence of object displacements between frames. The conventional methods [15, 16, 42] address the data fidelity and the regularization term as an energy optimization problem. However, they often struggle to estimate accurate results in various scenarios, such as substantial object displacements or small-sized objects. Recent learning-based methods [35, 40, 41] outperform the conventional method's accuracy and computational cost using various structures such as hierarchical design and progressive estimation. Many tasks, such as video super-resolution [6, 22], action recognition [38], and autonomous driving [19], predict the optical flow to find pixel correspondences. However, MFSR task does not effectively utilize optical flow due to computational cost. On the other hand, the proposed network (BurstM) has a simple structure and lightweights, allowing us to incorporate an optical flow estimator into the network." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Implicit Neural Representation (INR) One of the popular representation methods for the continuous-domain signal is INR [17]. Neural networks designed for INR learn and represent complex structures implicitly. The various tasks employ INR in image processes such as 2D and 3D shapes. Despite such success, INR parameterized by a standalone MLP with the ReLU has a spectral bias problem [32, 34, 39] that fails to represent high-frequency details known as a spectral bias. Recent approaches [8, 21, 23] tackle a spectral bias through positional encoding and Fourier feature mapping, demonstrating a flexible SR network. We have adopted INR method to utilize the flexible characteristic." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "EG Kang et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.143, + 0.788, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.303, + 0.789, + 0.402 + ], + "angle": 0, + "content": "Fig. 2: Proposed BurstM Network. The network estimates optical flow between the reference and source frames using FNet. The neural warping \\(\\mathbf{T}_{\\nu}\\) emphasizes the high-frequency details and implements the warping on Fourier space. Reconstruction module \\(\\mathbf{R}_{\\varphi}\\) includes the blender \\(\\mathbf{B}_{\\eta}\\) and the decoder \\(\\mathbf{G}_{\\Theta}\\). The blender \\(\\mathbf{B}_{\\eta}\\) matches color differences and merges multiple frames into a single frame. The Decoder \\(\\mathbf{G}_{\\Theta}\\) predicts the final output. The gray color indicates a single operation such a convolution and LeakyReLU activation and the others include nonlinear activation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.416, + 0.46, + 0.433 + ], + "angle": 0, + "content": "3 Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.788, + 0.549 + ], + "angle": 0, + "content": "Our purpose is to reconstruct the high-resolution image \\(\\mathbf{I}^{\\mathbf{SR}} \\in \\mathbb{R}^{3 \\times 2sH \\times 2sW}\\) with multiple scale factors \\(s\\) from the set of a low-resolution images \\(\\{\\mathbf{I}_i^{\\mathbf{L}}\\}_{i=1}^N\\), where \\(\\mathbf{I}_i^{\\mathbf{L}} \\in \\mathbb{R}^{4 \\times H \\times W}\\), \\(N\\) is the number of frame and \\(H, W\\) are the spatial sizes of low resolution image, and 4-channels are due to 'RGGB' domain. The dimension of a high-resolution image is increased by a factor of 2s, because the domain is changed from 4-channels 'RGGB' to 3-channels 'sRGB' in the network (in Figure 2)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.558, + 0.457, + 0.574 + ], + "angle": 0, + "content": "3.1 Optical Flow Estimator" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.787, + 0.613 + ], + "angle": 0, + "content": "We adopt FNet (optical flow estimator) of FRVSR [36]. FNet predicts optical flow on low-resolution grid as below:" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.621, + 0.787, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\pmb {\\delta} _ {i} ^ {\\mathbf {L}} = F N e t (\\mathbf {I} _ {i} ^ {\\mathbf {L}}), \\quad i = 1, \\dots , N, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.651, + 0.788, + 0.7 + ], + "angle": 0, + "content": "where \\(\\delta_{i}^{\\mathbf{L}}\\in \\mathbb{R}^{H\\times W\\times 2}\\) is the set of optical flow offsets between a reference and source frames on low-resolution grid. Our model adds offsets \\(\\delta_{i}^{\\mathbf{L}}\\) into regular coordinates \\(\\mathbf{x}_i^{\\mathbf{L}}\\in [-1,1]^{H\\times W\\times 2}\\) and applies s scale upsampling." + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.708, + 0.787, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {i} ^ {\\mathbf {H}} = \\mathbf {B _ {s}} (\\pmb {\\delta} _ {i} ^ {\\mathbf {L}} + \\mathbf {x} _ {i} ^ {\\mathbf {L}}), \\quad i = 1, \\dots , N, \\qquad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.736, + 0.787, + 0.768 + ], + "angle": 0, + "content": "where \\(\\mathbf{B}_{\\mathbf{s}}(\\cdot)\\) is bilinear upsampling operator with a scale factor of s, and \\(\\hat{\\mathbf{x}}_i^{\\mathbf{H}} \\in \\mathbb{R}^{sH \\times sW \\times 2}\\) is the upscaled coordinate with an offset defined by scale factor s." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.777, + 0.483, + 0.792 + ], + "angle": 0, + "content": "3.2 Learnable Neural Warping" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Our encoder \\((\\mathbf{E}_{\\psi}(\\cdot):\\mathbb{R}^4\\mapsto \\mathbb{R}^{2K})\\), which is a learnable super-resolution (SR) backbone [25] without upsampling, extracts latent vectors \\(\\mathbf{z}_i\\). NW \\((\\mathbf{T}_{\\nu})\\) takes the latent vectors \\(\\mathbf{z}_i\\) for warping based on a Fourier information. NW consists" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.678, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BurstM" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.284 + ], + "angle": 0, + "content": "of three components: an amplitude estimator \\((\\mathbf{g}_{\\mathbf{a}}(\\cdot):\\mathbb{R}^{2K}\\mapsto \\mathbb{R}^{2K})\\), a frequency estimator \\((\\mathbf{g}_{\\mathbf{f}}(\\cdot):\\mathbb{R}^{2K}\\mapsto \\mathbb{R}^{2K})\\) and, a phase estimator \\((\\mathbf{g}_{\\mathbf{p}}(\\cdot):\\mathbb{R}^2\\mapsto \\mathbb{R}^K)\\). \\(2K\\) denotes the number of channels in the latent vector. Amplitude and frequency estimators utilize the latent vector \\(\\mathbf{z}_i\\) to estimate the Fourier coefficients. After that, the nearest interpolation \\(\\mathbf{N_s}\\) is performed using estimated coordinates \\(\\hat{\\mathbf{x}}_i^{\\mathbf{L}}\\). The input of a phase estimator is a cell \\(\\mathbf{c}_i\\in \\mathbb{R}^{sH\\times sW\\times 2}\\), which is consist of the divisor of the input frame's spatial sizes \\((1 / H,1 / W)\\). As in [8,20], we use relative coordinates \\((\\delta_{i} = \\hat{\\mathbf{x}}_{i}^{\\mathrm{H}} - \\mathbf{x}_{i}^{\\mathrm{H}})\\) known as a local grid to predict the continuous signal where \\(\\hat{\\mathbf{x}}_i^{\\mathrm{H}}\\in \\mathbb{R}^{sH\\times sW\\times 2}\\). The designed formulation is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.455, + 0.295, + 0.785, + 0.312 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {i} = \\mathbf {E} _ {\\psi} \\left(\\mathbf {I} _ {i} ^ {\\mathbf {L}}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.284, + 0.316, + 0.784, + 0.333 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {i} = \\mathbf {N} _ {\\mathbf {s}} \\left(\\mathbf {g} _ {\\mathbf {f}} \\left(\\mathbf {z} _ {i}\\right), \\hat {\\mathbf {x}} _ {i} ^ {\\mathbf {H}}\\right), \\quad \\mathbf {A} _ {i} = \\mathbf {N} _ {\\mathbf {s}} \\left(\\mathbf {g} _ {\\mathbf {a}} \\left(\\mathbf {z} _ {i}\\right), \\hat {\\mathbf {x}} _ {i} ^ {\\mathbf {H}}\\right), \\quad \\mathbf {p h} _ {i} = \\mathbf {g} _ {\\mathbf {p}} \\left(\\mathbf {c} _ {i}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.337, + 0.784, + 0.369 + ], + "angle": 0, + "content": "\\[\n\\mathbf {m} _ {i} = \\mathbf {A} _ {i} \\odot \\left[ \\begin{array}{l} \\cos (\\pi \\left(\\left\\langle \\mathbf {F} _ {i}, \\boldsymbol {\\delta} _ {i} \\right\\rangle + \\mathbf {p h} _ {i}\\right) \\\\ \\sin (\\pi \\left(\\left\\langle \\mathbf {F} _ {i}, \\boldsymbol {\\delta} _ {i} \\right\\rangle + \\mathbf {p h} _ {i}\\right) \\end{array} \\right], \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.373, + 0.632, + 0.388 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e ,} \\quad i = 1, \\dots , N, \\quad \\boldsymbol {\\nu} = [ \\mathbf {f}; \\mathbf {a}; \\mathbf {p} ].\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.396, + 0.457, + 0.409 + ], + "angle": 0, + "content": "3.3 Reconstruction Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.785, + 0.459 + ], + "angle": 0, + "content": "To reconstruct the high-resolution image using estimated features in Fourier space, we employ the reconstruction module \\(\\mathbf{R}_{\\varphi}\\); Blender \\((\\mathbf{B}_{\\eta}(\\cdot):\\mathbb{R}^{N\\times 2K}\\mapsto \\mathbb{R}^{1\\times 2K})\\) and MLP decoder \\((\\mathbf{G}_{\\Theta}(\\cdot):\\mathbb{R}^{2K}\\mapsto \\mathbb{R}^{12})\\)." + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.468, + 0.785, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} = \\mathbf {R} _ {\\varphi} \\left(\\left\\{\\mathbf {m} _ {i} \\right\\} _ {i = 1} ^ {N}\\right) = \\mathbf {G} _ {\\Theta} \\left(\\mathbf {B} _ {\\boldsymbol {\\eta}} \\left(\\left\\{\\mathbf {m} _ {i} \\right\\} _ {i = 1} ^ {N}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.553 + ], + "angle": 0, + "content": "where \\(\\mathbf{h} \\in \\mathbb{R}^{1 \\times 12 \\times sH \\times sW}\\) is the merged Fourier features of burst shots. The blender \\(\\mathbf{B}_{\\eta}\\), which is the same module with the encoder \\(\\mathbf{E}_{\\psi}\\), reduces color differences between frames and merges multiple frames. MLP decoder \\((\\mathbf{G}_{\\Theta})\\) estimates 'sRGB' information." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.562, + 0.537, + 0.578 + ], + "angle": 0, + "content": "3.4 Skip connection with upsampling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.788, + 0.64 + ], + "angle": 0, + "content": "To prevent learning on the low-frequency only, we utilize the upscaled skip connection \\((\\mathbf{Q}_{\\pmb{\\xi}}(\\cdot):\\mathbb{R}^{4\\times H\\times W}\\mapsto \\mathbb{R}^{12\\times sH\\times sW})\\) about the reference frame \\(\\mathbf{I}_1^{\\mathbf{L}}\\). Finally, the spatial size is increased using pixel shuffle \\((\\mathbf{P_2}(\\cdot):\\mathbb{R}^{12\\times sH\\times sW}\\mapsto \\mathbb{R}^{3\\times 2sH\\times 2sW})\\). The proposed BurstM can be formulated as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.649, + 0.785, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} ^ {\\mathbf {S R}} = \\mathbf {P} _ {2} (\\mathbf {h} + \\mathbf {Q} _ {\\xi} \\left(\\mathbf {I} _ {1} ^ {\\mathbf {L}}\\right)) \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.674, + 0.373, + 0.687 + ], + "angle": 0, + "content": "3.5 Loss function" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.787, + 0.737 + ], + "angle": 0, + "content": "We use two loss functions for both SyntheticBurst [2] and BurstSR [3]. The MFSR datasets [2,3] does not provide the ground truth of the optical flow map between frames. Thus, we use L2 loss term to use the photometric loss." + }, + { + "type": "equation", + "bbox": [ + 0.28, + 0.746, + 0.785, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {O F}} = \\sum_ {i = 2} ^ {N} \\left\\| W a r p \\left(\\mathbf {I} _ {i} ^ {\\mathbf {L}}, \\boldsymbol {\\delta} _ {i} ^ {\\mathbf {L}}\\right) - \\mathbf {I} _ {1} ^ {\\mathbf {L}} \\right\\| _ {2} = \\sum_ {i = 2} ^ {N} \\left(W a r p \\left(\\mathbf {I} _ {i} ^ {\\mathbf {L}}, \\boldsymbol {\\delta} _ {i} ^ {\\mathbf {L}}\\right) - \\mathbf {I} _ {1} ^ {\\mathbf {L}}\\right) ^ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.826 + ], + "angle": 0, + "content": "where \\( \\text{Warp}(\\cdot, \\cdot) \\) is the bilinear warping operator using optical flow from FNet, and \\( \\mathbf{I}_1^{\\mathbf{L}} \\) is the reference frame." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "EG Kang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "SyntheticBurst dataset Input frames are generated from the ground truth. Thus, we update our model using L1 loss between predicted image and the ground truth." + }, + { + "type": "equation", + "bbox": [ + 0.42, + 0.191, + 0.785, + 0.21 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {S R}} = \\left\\| \\mathbf {I} ^ {\\mathbf {S R}} - \\mathbf {I} ^ {\\mathbf {G T}} \\right\\| _ {1}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.215, + 0.63, + 0.23 + ], + "angle": 0, + "content": "Thus, the total loss function during training is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.243, + 0.786, + 0.26 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {S y n t h e t i c}} = \\mathcal {L} _ {\\mathcal {S R}} + \\mathcal {L} _ {\\mathcal {O F}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.27, + 0.784, + 0.316 + ], + "angle": 0, + "content": "BurstSR dataset We employs an aligned L1 loss for the BurstSR [3] dataset due to mismatching between low-resolution images and a ground truth image. An aligned L1 loss utilizes pre-trained weight of PwCNet [40] as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.404, + 0.327, + 0.785, + 0.344 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\delta} ^ {\\mathbf {A}} = P W C N e t \\left(\\mathbf {I} ^ {\\mathbf {S R}}, \\mathbf {I} ^ {\\mathbf {G T}}\\right), \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.356, + 0.348, + 0.785, + 0.366 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {A l i g n e d S R}} = \\left\\| W a r p \\left(\\mathbf {I} ^ {\\mathbf {S R}}, \\boldsymbol {\\delta} ^ {\\mathbf {A}}\\right) - \\mathbf {I} ^ {\\mathbf {G T}} \\right\\| _ {1}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.784, + 0.424 + ], + "angle": 0, + "content": "where \\(\\delta^{\\mathbf{A}}\\) is optical flow between the predicted image on high-resolution grid and ground truth using PwCNet [40]. Total loss function for BurstSR [3] dataset is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.404, + 0.426, + 0.785, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {R e a l}} = \\mathcal {L} _ {\\text {A l i g n e d S R}} + \\mathcal {L} _ {\\mathcal {O F}}. \\tag {13}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.464, + 0.367, + 0.481 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.495, + 0.393, + 0.51 + ], + "angle": 0, + "content": "4.1 Network details" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.52, + 0.784, + 0.566 + ], + "angle": 0, + "content": "Optical Flow Estimator (FNet) Our network adapts FNet from FRVSR [36]. To apply FNet into BurstSR task, we modify the stem of FNet from 'sRGB' to 'RGGB' images." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.581, + 0.785, + 0.656 + ], + "angle": 0, + "content": "Learnable Neural Warping \\((\\mathbf{T}_{\\nu})\\) NW estimates the amplitude, frequency, and phase from burst shots with a single convolution operation to represent the Fourier information. Amplitude and frequency estimators have 128 channels of the convolutions with \\(3\\times 3\\) kernel, phase estimator has 64 channels of the convolutions with \\(1\\times 1\\) kernel." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.672, + 0.785, + 0.746 + ], + "angle": 0, + "content": "Encoder \\((\\mathbf{E}_{\\psi})\\) and Reconstruction module \\((\\mathbf{R}_{\\varphi})\\) The reconstruction module consists of Blender \\(\\mathbf{B}_{\\eta}\\) and MLP decoder \\(\\mathbf{G}_{\\Theta}\\). Both Encoder \\(\\mathbf{E}_{\\psi}\\) and Blender \\(\\mathbf{B}_{\\eta}\\) are EDSR [25] without upsampling module and 128 channels. MLP Decoder, \\(\\mathbf{G}_{\\Theta}\\), is organized by 4-layer of \\(1\\times 1\\) convolution with 256 channels and ReLU activation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.77, + 0.454, + 0.785 + ], + "angle": 0, + "content": "4.2 Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.842 + ], + "angle": 0, + "content": "BurstM is an end-to-end trainable network and is possible to train without pre-trained weights. We apply an AdamW optimizer [28] with the Cosine annealing strategy [27] from \\(10^{-4}\\) to \\(10^{-6}\\) learning rate. For SyntheticBurst [2], we" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.678, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BurstM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "train BurstM with L1 loss for the output image and L2 loss for optical flows. For BurstSR [3], We fine-tune the pre-trained BurstM of SyntheticBurst with L2 loss for optical flow and aligned L1 loss [3] for the output image. Aligned L1 loss utilizes PWCNet [40] for the output image. The reason for the usage of the aligned L1 Loss is that there are no LR images aligned with the GT image because they were captured using different cameras. We train our model using four RTX3090 GPUs and training times are 10 days for SyntheticBurst and 8 hours for BurstSR. The summary of the training strategy is in Table 1." + }, + { + "type": "table", + "bbox": [ + 0.333, + 0.269, + 0.673, + 0.442 + ], + "angle": 0, + "content": "
SyntheticBurst [2]BurstSR [3]
Training Epoch30025
OptimizerADAMW [28]
Learning rate10-4~10-6
LR schedulerCosine annealing [27]
Batch sizes4
Main LossL1Aligned L1
Photo metric LossL2
×2:96×96
Input sizes×3:64×64×4:80×80
×4:48×48
Training time10 days8 hours
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.442, + 0.788, + 0.484 + ], + "angle": 0, + "content": "Table 1: Training Strategies. Training epoch and main loss, input sizes are different. Except for these factors, we use same training strategies for both SyntheticBurst [2] for BurstSR [3] datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.499, + 0.788, + 0.621 + ], + "angle": 0, + "content": "SyntheticBurst Dataset [2] The dataset contains burst frames which are 46,839 set for training and 300 set for validation. Each burst set contains 14 low-resolution burst frames, which are generated from 'sRGB' as follows. We utilized the unprocessing [5]. Subsequently, we apply random rotations and translations. We employ random down-sampling \\((\\times 2, \\times 3, \\times 4)\\) to generate low-resolution images of multiple scales. Finally, we apply Bayer mosaicking and add random noise. We trained the proposed model with batch size 4 during 300 epochs, and the training patch size is changed per scale factors in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.635, + 0.811, + 0.756 + ], + "angle": 0, + "content": "BurstSR Dataset [3] Input and GT are captured by different cameras: Smartphone and DSLR, respectively. Both low and high resolution images have 200 raw burst sets. The low-resolution set contains 14 frames, and we extract 5,405 patches for training and 882 patches for validation. Unlike SyntehticBurst dataset [2], the scale factor between low-resolution and high-resolution is fixed. Thus, we train BurstM only for \\(\\times 4\\) scale factor. Nevertheless, BurstM can accurately predict high-resolution images for multiple scale factors (\\(\\times 2\\), \\(\\times 3\\), \\(\\times 4\\)), including clear representation of high-frequency textures without the boundary artifacts." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.774, + 0.416, + 0.789 + ], + "angle": 0, + "content": "4.3 Validation Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We show the qualitative comparison with the existing methods [3, 4, 11, 12, 29, 30] and the visual comparison with latest three methods [11, 12, 29] using the pre-trained weights from authors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.229, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.375, + 0.129 + ], + "angle": 0, + "content": "EG Kang et al." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.145, + 0.777, + 0.319 + ], + "angle": 0, + "content": "
# Params. MSyntheticBurstBurstSR x4 PSNR SSIM
x2 PSNRSSIMx3 PSNRSSIMx4 PSNR
Bicubic-38.300.94833.940.88633.020.862
DBSR [3]13.0140.510.96540.110.95940.760.959
MFIR [4]12.1341.250.97141.810.97241.560.964
BIPNet†[11]6.737.580.92840.830.95541.930.960
GMTNet†[30]-----42.360.961
Burststormer†[12]2.537.060.92540.260.95342.830.973
BSRT-Small†[29]4.9240.640.96642.300.97542.720.971
BSRT-Large†[29]20.7140.330.96542.870.97943.620.975
BurstM (Ours)14.046.010.98544.790.98242.870.973
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.32, + 0.785, + 0.374 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison on SyntheticBurst and BurstSR. Quantitative comparisons of BurstSR such as PSNR and SSIM can only be measured for \\(\\times 4\\), due to LR and HR being provided in fixed sizes. Red and Blue colors indicate the best and the second performance, respectively. \\(\\dagger\\) indicates Transformer-based method." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.381, + 0.78, + 0.805 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.809, + 0.787, + 0.837 + ], + "angle": 0, + "content": "Fig. 3: Visual comparison for \\(\\times 4\\) super-resolution for BurstSR. BurstM shows clearer textures and correct color than the others without the boundary artifacts." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.678, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BurstM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.327 + ], + "angle": 0, + "content": "SR on BurstSR [3] Table 2 is the quantitative comparison result of \\(\\times 4\\) SR on BurstSR [3]. Ours shows a performance gain of \\(+0.17\\mathrm{dB}\\) compared to the state-of-the-art baselines. As shown in Figure 3, BurstM predicts outperform high-frequency details without the boundary artifacts from low-resolution images of real-world dataset. Also, BurstM predicts multi-scale SR without an additional processing, different with existing models. However, the quantitative comparison of \\(\\times 2\\), \\(\\times 3\\) is not possible, because there is no GT of both \\(\\times 2\\), \\(\\times 3\\) SR of given dataset [3]. We demonstrate the robustness of our approach by representing high-frequency details in Figure 4 and computation time in Table 4. BurstM directly predicts multiple scales, while other methods implement \\(\\times 4\\) SR followed by bicubic downsampling. In Table 4, BurstM shows shorter computation time than the existing methods." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.328, + 0.777, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.551 + ], + "angle": 0, + "content": "Fig. 4: Visual comparison of multi-scale SR on BurstSR. BurstM predict several scaled-images with unimodel and the others are downsampling after \\(\\times 4\\) SR." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.789, + 0.751 + ], + "angle": 0, + "content": "SR on SyntheticBurst [2] Table 2 is the quantitative comparison results of multiple scale factors \\((\\times 2, \\times 3, \\times 4)\\) SR and Figure 5 is the visual comparison result of \\(\\times 4\\) SR on SyntheticBurst [2]. Our model achieves second result with reasonable the number of parameters. In addition, we compare the performance on multiple scale factors in Table 2 and Figure 6. BurstM predicts high-resolution images on multiple scale factors with an unimodel and the others implement multiple SR with additional bicubic downsampling after \\(\\times 4\\) SR. In case of BurstM, it is observed that PSNR gradually increases as the scale factor decrease. However, PSNR of existing methods decreases as the scale factor decreases because of the information loss during the additional bicubic downsampling process. Figure 6 shows the residual images, which means the loss of the high-frequency textures. Furthermore, as shown in Table 4, BurstM outperforms the other methods in terms of computation time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Fourier feature space Figure 7 demonstrates the estimation of high-frequency details on Fourier space. The result from each frame do not illustrate the DFT spectrum, but accumulated result of all frames shows a similar result with the DFT spectrum. Because BurstM leverages sub-pixel information from burst frames to represent the high-resolution images." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "EG Kang et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.784, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.369, + 0.764, + 0.384 + ], + "angle": 0, + "content": "Fig. 5: Visual comparison for \\(\\times 4\\) Super-Resolution for SynteticBurst." + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.392, + 0.789, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.667, + 0.788, + 0.737 + ], + "angle": 0, + "content": "Fig.6: Visual comparison for multi-scale SR on SyntheticBurst. The colorbar indicates a difference with GT. The existing methods lose high-frequency textures during additional down-sampling because the highlighted locations are a high-frequency area. But BurstM represent the similar high-frequency textures with GT by because of direct prediction on multiple scale factors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Aligning performances We conduct a comparative analysis of the alignment performance between DCN-based method and BurstM in Figure 8. As in [11, 12, 29, 30], we utilize multi-layer DCN with varied spatial sizes for DCN-based method. In contrast, BurstM implements a single image warping process using predicted optical flows. The 'Before align.' and 'After align.' residual maps in Figure 8 represent the differences with the reference frame at each position in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.678, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BurstM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.143, + 0.761, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.399 + ], + "angle": 0, + "content": "Fig.7: Fourier space analysis. Accumulated frequencies and coefficients from BurstM are analogous to DFT spectrum of GT (Red color in 'DFT' and 'ACCUM')." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.789, + 0.54 + ], + "angle": 0, + "content": "flowchart. The red dots in 'DCN method flowchart' illustrate offsets to the center pixel. DCN uses an unrelated pixel information, because DCN requires to use the number of predefined kernels regardless of the correlation. Therefore, the aligned results are inaccurate and show high values in the residual map. Additionally, boundary artifacts are also observed due to the utilization of irrelevant information. The color map and white arrows in 'BurstM method flowchart' are the visualization results of optical flow. BurstM method shows one-to-one correspondences between pixels. Thus, the alignment results demonstrate clearer and sharper compared to the DCN-based method without the boundary artifacts." + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.544, + 0.495, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.544, + 0.736, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.617, + 0.749, + 0.631 + ], + "angle": 0, + "content": "DCN method flowchart [12] BurstM method (ours) flowchart" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.633, + 0.784, + 0.769 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.772, + 0.788, + 0.856 + ], + "angle": 0, + "content": "Fig. 8: Feature alignment results. Flow charts indicate where we take features. The residual map is between the reference and source frames at each location in flowchart. DCN based-method has a big difference after an alignment because DCN take unrelated information during alignment. However, BurstM point out one-to-one correspondences between a reference and source frames. BurstM shows accurate align result without the boundary artifacts." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.128 + ], + "angle": 0, + "content": "EG Kang et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.398, + 0.163 + ], + "angle": 0, + "content": "5 Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.166, + 0.788, + 0.273 + ], + "angle": 0, + "content": "The comparison between with and without Fourier mapping INR [17] is the most representative method for representing continuous signals, but it has a spectral bias [32, 34, 39]. As shown in [21, 23], BurstM also addresses this issue through Fourier mapping. Table 3a compares the performance of BurstM with and without Fourier mapping. BurstM without Fourier mapping has significantly lower PSNR values. This is because the network without Fourier mapping learns more about low frequencies due to spectral bias during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.287, + 0.789, + 0.395 + ], + "angle": 0, + "content": "Relationship between flexibility and performance BurstM is supporting the multiple scale SR \\((\\times 2, \\times 3, \\times 4)\\). The performance in Table 2 has trained with three multiple scale factors \\((\\times 2, \\times 3, \\times 4)\\). We train with different number of scale factors such as one \\((\\times 4)\\) and two \\((\\times 3, \\times 4)\\) items. In such cases, we observed that PSNR increases on in-of-scale, but performance decreases on out-of-scale. When there are many instances of out-of-scale, the model's flexibility diminishes, so we proposed three scales satisfy both performance and flexibility." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.789, + 0.483 + ], + "angle": 0, + "content": "Different optical flow estimator While FNet [36] estimates an optical flow based on given image sizes, the other optical flow estimator [35] adjust image sizes according to its setting before input. This resizing of images introduces aliasing effects, resulting in a decline in SR performance. Additionally, BurstM with Spynet [35] shows longer inference time than the proposed method." + }, + { + "type": "table", + "bbox": [ + 0.249, + 0.483, + 0.754, + 0.513 + ], + "angle": 0, + "content": "
BurstM w\\o Fourier mappingBurstM w\\Fourier mapping
×4 PSNR39.82 dB41.74 dB
" + }, + { + "type": "table_caption", + "bbox": [ + 0.28, + 0.514, + 0.725, + 0.527 + ], + "angle": 0, + "content": "(a) The performance comparison between with and without Fourier mapping" + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.532, + 0.484, + 0.594 + ], + "angle": 0, + "content": "
Training scale×2×3×4
×433.3635.9742.26
×3, ×430.8043.6141.82
×2, ×3, ×4 (Ours)44.7743.4541.74
" + }, + { + "type": "table_caption", + "bbox": [ + 0.245, + 0.595, + 0.477, + 0.618 + ], + "angle": 0, + "content": "(b) Relationship between flexibility and performance (PSNR:dB)" + }, + { + "type": "table", + "bbox": [ + 0.492, + 0.532, + 0.783, + 0.593 + ], + "angle": 0, + "content": "
BurstM\nw\\SpynetBurstM\nw\\FNet (Ours)
×4 PSNR40.85 dB41.74 dB
Inference time14.3 ms11.6 ms
" + }, + { + "type": "table_caption", + "bbox": [ + 0.545, + 0.595, + 0.726, + 0.617 + ], + "angle": 0, + "content": "(c) Performance comparison based on optical flow estimator" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.623, + 0.477, + 0.637 + ], + "angle": 0, + "content": "Table 3: Ablation study summary" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.637, + 0.611, + 0.65 + ], + "angle": 0, + "content": "(a) Fourier mapping contributes to the model performance." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.65, + 0.786, + 0.692 + ], + "angle": 0, + "content": "(b) Red color indicates the result of in-scale. Less training scale has higher PSNR on in-scale but network flexibility is low due to increasing out-scale. Thus, we select our model with 3 scale factors \\((\\times 2, \\times 3, \\times 4)\\) considering both flexibility and performance." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.692, + 0.786, + 0.718 + ], + "angle": 0, + "content": "(c) A different optical flow estimator leads to PSNR drop and an increase in inference time." + }, + { + "type": "list", + "bbox": [ + 0.216, + 0.637, + 0.786, + 0.718 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.728, + 0.355, + 0.743 + ], + "angle": 0, + "content": "6 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Boundary line artifacts As shown in Figure 9, the recent method [29] exhibits boundary artifacts despite utilizing optical flow, due to its continued reliance on DCN [9] for alignment. As illustrated in Figure 8 and 9, DCN [9] must utilize information corresponding to the predefined kernel size regardless of correlation, resulting in the incorporation of unnecessary information when most frames fail to provide essential data. In contrast, BurstM aligns frames without DCN [9]," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.678, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BurstM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "solely utilizing optical flow. Consequently, BurstM avoids the use of irrelevant information, leading to accurate predictions without boundary artifacts." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.183, + 0.764, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.39, + 0.788, + 0.434 + ], + "angle": 0, + "content": "Fig.9: Boundary artifacts The most of frames didn't provide the information of edge. The existing method [29] suffers the boundary line artifacts, but BurstM represents clear images without the boundary line artifacts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.788, + 0.528 + ], + "angle": 0, + "content": "Computation complexity Despite advancements in computational system, SR field still requires shorter computation times as the image resolutions increase. Table 4 compares the computation time usage of BurstM with the other methods [11,12,29]. We implements \\(\\times 4\\) SR with \\(80 \\times 80\\) frame sizes of BurstSR dataset [3] on NVIDIA RTX3090 24GB. While BurstM has a large number of parameters, BurstM demonstrates the shortest computation time." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.528, + 0.54, + 0.663 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.529, + 0.794, + 0.66 + ], + "angle": 0, + "content": "
SR scale: ×4Params.(M) / Mem.(GB)Inference time(ms)PSNR (dB)
BIPNet† [11]6.7 / 11.820.848.49
Burstormer† [12]2.5 / 12.020.848.82
BSRT-S† [29]4.92 / 1.0148.048.48
BSRT-L† [29]20.71 / 1.5266.148.57
BurstM (Ours)14.9 / 8.411.649.12
" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.663, + 0.788, + 0.706 + ], + "angle": 0, + "content": "Table 4: Computational complexity on one GPU (RTX3090). Despite a large number of parameters, BurstM shows the inference time (ms) and best PSNR (dB) performance. Red : best metrics. † indicates Transformer-based method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Super-Resolution with floating scale factors BurstM can estimate the multi-scale SR with floating scale factors such as \\(\\times 3.5\\) without any architecture changes. However, the training of the floating scale factors (\\(\\times 2\\sim \\times 4\\)) shows lower PSNR (39.44dB) than the training of the integer scale factors (41.74dB). The performance drop is caused by the demosaic process when generating low-resolution frames. To convert from 'sRGB' to 'RGGB', the dimensions (H, W) of the images are changed to (H/2, W/2). If the image size of 'sRGB' is not divisible by two when a floating scale factor is applied, the network apply an additional" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "EG Kang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "downsampling operation to make the dimension of \\((\\mathrm{H} / 2, \\mathrm{W} / 2)\\). The double downsampling imposes a severe degradation of the generated low-resolution images, and it results in low PSNRs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.198, + 0.788, + 0.288 + ], + "angle": 0, + "content": "**Corner case** The major challenge on optical flow is occlusions which are caused by a new object that is not present in the reference frame. BurstM also has the occlusion issue (Figure 10) even though BurstM demonstrate accurate alignment performance than DCN based methods. Fortunately, BurstM utilize inaccurate optical flow of an occlusion as meaningful prior and reconstruct the high-quality images in a high-resolution grid." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.288, + 0.761, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.553 + ], + "angle": 0, + "content": "Fig. 10: Corner case. Inaccurate optical flow prediction due to occlusions which are caused by new objects (red arrow) not present in the reference image appear." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.561, + 0.36, + 0.576 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.765 + ], + "angle": 0, + "content": "In this paper, we proposed Deep Burst Multi-scale SR using Fourier Space with Optical Flow (BurstM) to overcome the limitations of existing methods: SR of fixed scale factor \\((\\times 4)\\) and low-quality of real-world dataset [3]. Our network performs the precise alignment using optical flow and the advanced warping process based on estimated Fourier information. We demonstrated their effectiveness in representing the high-frequency textures with \\(+0.17\\mathrm{dB}\\) PSNR gain in the real-world dataset [3] compared to the previous state-of-the-art model. Furthermore, BurstM successfully implements multi-scale SR \\((\\times 2,\\times 3,\\times 4)\\) of the unimodel by representing high-frequency details using the INR method. The function of multi-scale SR enhances our model's capacity compared to the other learning-based methods. In addition, we demonstrate our approach is faster inference time compared to the existing methods by the efficient computation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Acknowledgement This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (RS-2021-II212068, Artificial Intelligence Innovation Hub) and the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (RS-2024-00335741)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.678, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BurstM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.786, + 0.228 + ], + "angle": 0, + "content": "1. Bascle, B., Blake, A., Zisserman, A.: Motion deblurring and super-resolution from an image sequence. In: Computer Vision—ECCV'96: 4th European Conference on Computer Vision Cambridge, UK, April 15–18, 1996 Proceedings Volume II 4. pp. 571–582. Springer (1996)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.229, + 0.786, + 0.269 + ], + "angle": 0, + "content": "2. Bhat, G., Danelljan, M., Timofte, R.: Ntire 2021 challenge on burst superresolution: Methods and results. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 613-626 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.27, + 0.786, + 0.31 + ], + "angle": 0, + "content": "3. Bhat, G., Danelljan, M., Van Gool, L., Timofte, R.: Deep burst super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9209-9218 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.786, + 0.351 + ], + "angle": 0, + "content": "4. Bhat, G., Danelljan, M., Yu, F., Van Gool, L., Timofte, R.: Deep reparametrization of multi-frame super-resolution and denoising. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2460-2470 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.351, + 0.786, + 0.392 + ], + "angle": 0, + "content": "5. Brooks, T., Mildenhall, B., Xue, T., Chen, J., Sharlet, D., Barron, J.T.: Unprocessing images for learned raw denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11036-11045 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.392, + 0.786, + 0.445 + ], + "angle": 0, + "content": "6. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basicvsr: The search for essential components in video super-resolution and beyond. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4947-4956 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.786, + 0.5 + ], + "angle": 0, + "content": "7. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-Trained Image Processing Transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12299-12310 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.501, + 0.786, + 0.542 + ], + "angle": 0, + "content": "8. Chen, Y., Liu, S., Wang, X.: Learning continuous image representation with local implicit image function. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8628-8638 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.542, + 0.786, + 0.582 + ], + "angle": 0, + "content": "9. Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., Wei, Y.: Deformable convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 764-773 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.583, + 0.786, + 0.623 + ], + "angle": 0, + "content": "0. Dai, T., Cai, J., Zhang, Y., Xia, S.T., Zhang, L.: Second-order attention network for single image super-resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11065-11074 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.623, + 0.786, + 0.663 + ], + "angle": 0, + "content": "1. Dudhane, A., Zamir, S.W., Khan, S., Khan, F.S., Yang, M.H.: Burst image restoration and enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5759-5768 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.663, + 0.786, + 0.704 + ], + "angle": 0, + "content": "2. Dudhane, A., Zamir, S.W., Khan, S., Khan, F.S., Yang, M.H.: Burstomer: Burst image restoration and enhancement transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5703-5712 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.705, + 0.786, + 0.745 + ], + "angle": 0, + "content": "3. Elad, M., Feuer, A.: Restoration of a single superresolution image from several blurred, noisy, and undersampled measured images. IEEE transactions on image processing 6(12), 1646-1658 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.746, + 0.786, + 0.799 + ], + "angle": 0, + "content": "4. Hardie, R.C., Barnard, K.J., Bognar, J.G., Armstrong, E.E., Watson, E.A.: High-resolution image reconstruction from a sequence of rotated and translated frames and its application to an infrared imaging system. Optical Engineering 37(1), 247-260 (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "5. Hirschmuller, H.: Stereo processing by semiglobal matching and mutual information. IEEE Transactions on pattern analysis and machine intelligence 30(2), 328-341 (2007)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.129 + ], + "angle": 0, + "content": "EG Kang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Horn, B.K., Schunck, B.G.: Determining optical flow. Artificial intelligence 17(1-3), 185-203 (1981)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.203 + ], + "angle": 0, + "content": "17. Hornik, K., Stinchcombe, M., White, H.: Multilayer feedforward networks are universal approximators. Neural networks 2(5), 359-366 (1989)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "18. Irani, M., Peleg, S.: Improving resolution by image registration. CVGIP: Graphical models and image processing 53(3), 231-239 (1991)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "19. Janai, J., Güney, F., Behl, A., Geiger, A., et al.: Computer vision for autonomous vehicles: Problems, datasets and state of the art. Foundations and Trends in Computer Graphics and Vision 12(1-3), 1-308 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "20. Jiang, C.M., Sud, A., Makadia, A., Huang, J., Niessner, M., Funkhouser, T.: Local Implicit Grid Representations for 3D Scenes. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "21. Kim, M., Lee, J., Lee, B., Im, S., Jin, K.H.: Implicit neural image stitching with enhanced and blended feature reconstruction (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "22. Lai, W.S., Huang, J.B., Ahuja, N., Yang, M.H.: Deep laplacian pyramid networks for fast and accurate super-resolution. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 624-632 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "23. Lee, J., Jin, K.H.: Local texture estimator for implicit representation function. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1929-1938 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "24. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: SwinIR: Image Restoration Using Swin Transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 1833-1844 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "25. Lim, B., Son, S., Kim, H., Nah, S., Mu Lee, K.: Enhanced deep residual networks for single image super-resolution. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 136-144 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.509, + 0.785, + 0.563 + ], + "angle": 0, + "content": "26. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10012-10022 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.591 + ], + "angle": 0, + "content": "27. Loshchilov, I., Hutter, F.: Stochastic gradient descent with warm restarts. In: Proceedings of the 5th Int. Conf. Learning Representations. pp. 1-16" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.618 + ], + "angle": 0, + "content": "28. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.619, + 0.785, + 0.674 + ], + "angle": 0, + "content": "29. Luo, Z., Li, Y., Cheng, S., Yu, L., Wu, Q., Wen, Z., Fan, H., Sun, J., Liu, S.: BSRT: Improving burst super-resolution with swim transformer and flow-guided deformable alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 998-1008 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.675, + 0.785, + 0.729 + ], + "angle": 0, + "content": "30. Mehta, N., Dudhane, A., Murala, S., Zamir, S.W., Khan, S., Khan, F.S.: Gated multi-resolution transfer network for burst restoration and enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22201-22210 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.73, + 0.785, + 0.771 + ], + "angle": 0, + "content": "31. Mei, Y., Fan, Y., Zhou, Y.: Image super-resolution with non-local sparse attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3517-3526 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "32. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "33. Peleg, S., Keren, D., Schweitzer, L.: Improving image resolution using subpixel motion. Pattern recognition letters 5(3), 223-226 (1987)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.678, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "BurstM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.218 + ], + "angle": 0, + "content": "34. Rahaman, N., Baratin, A., Arpit, D., Draxler, F., Lin, M., Hamprecht, F., Bengio, Y., Courville, A.: On the Spectral Bias of Neural Networks. In: Chaudhuri, K., Salakhutdinov, R. (eds.) Proceedings of the 36th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 97, pp. 5301-5310. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.786, + 0.259 + ], + "angle": 0, + "content": "35. Ranjan, A., Black, M.J.: Optical flow estimation using a spatial pyramid network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4161-4170 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.3 + ], + "angle": 0, + "content": "36. Sajjadi, M.S., Vemulapalli, R., Brown, M.: Frame-recurrent video super-resolution. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6626-6634 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.357 + ], + "angle": 0, + "content": "37. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1874-1883 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.384 + ], + "angle": 0, + "content": "38. Simonyan, K., Zisserman, A.: Two-stream convolutional networks for action recognition in videos. Advances in neural information processing systems 27 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.439 + ], + "angle": 0, + "content": "39. Sitzmann, V., Martel, J., Bergman, A., Lindell, D., Wetzstein, G.: Implicit Neural Representations with Periodic Activation Functions. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M.F., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 7462-7473. Curran Associates, Inc. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "40. Sun, D., Yang, X., Liu, M.Y., Kautz, J.: Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8934-8943 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "41. Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 402-419. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.578 + ], + "angle": 0, + "content": "42. Triggs, B., McLauchlan, P.F., Hartley, R.I., Fitzgibbon, A.W.: Bundle adjustment—a modern synthesis. In: Vision Algorithms: Theory and Practice: International Workshop on Vision Algorithms Corfu, Greece, September 21–22, 1999 Proceedings. pp. 298–372. Springer (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.605 + ], + "angle": 0, + "content": "43. Tsai, R.Y., Huang, T.S.: Multiframe image restoration and registration. Multiframe image restoration and registration 1, 317-339 (1984)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.787, + 0.647 + ], + "angle": 0, + "content": "44. Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image super-resolution using very deep residual channel attention networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 286-301 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "45. Zhang, Y., Tian, Y., Kong, Y., Zhong, B., Fu, Y.: Residual dense network for image super-resolution. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2472-2481 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.787, + 0.73 + ], + "angle": 0, + "content": "46. Zhang, Z., Wang, R., Zhang, H., Chen, Y., Zuo, W.: Self-supervised learning for real-world super-resolution from dual zoomed observations. In: European Conference on Computer Vision. pp. 610-627. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.73 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/BurstM_ Deep Burst Multi-scale SR using Fourier Space with Optical Flow/608ca9aa-ec95-4beb-807e-424a93d0d485_origin.pdf b/2024/BurstM_ Deep Burst Multi-scale SR using Fourier Space with Optical Flow/608ca9aa-ec95-4beb-807e-424a93d0d485_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b284a98a16460d895342df2f58ee8fa38aa92a9a --- /dev/null +++ b/2024/BurstM_ Deep Burst Multi-scale SR using Fourier Space with Optical Flow/608ca9aa-ec95-4beb-807e-424a93d0d485_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:940e6cd4cc6da2d147bbf800cce9653ce0732de618d1c4ac745fc338c9f720e5 +size 9095461 diff --git a/2024/BurstM_ Deep Burst Multi-scale SR using Fourier Space with Optical Flow/full.md b/2024/BurstM_ Deep Burst Multi-scale SR using Fourier Space with Optical Flow/full.md new file mode 100644 index 0000000000000000000000000000000000000000..60784dcbed762c04c45dccd1650905f4d9d2d6bc --- /dev/null +++ b/2024/BurstM_ Deep Burst Multi-scale SR using Fourier Space with Optical Flow/full.md @@ -0,0 +1,318 @@ +# BurstM: Deep Burst Multi-scale SR using Fourier Space with Optical Flow + +EungGu Kang¹ Byeonghun Lee² Sunghoon Im1* Kyong Hwan Jin2* + +1 Daegu Gyeongbuk Institute of Science and Technology (DGIST), Korea + +2 Korea University, Korea + +eunggukang@gmail.com, sunghoonim@dgist.ac.kr, {byeonghun_lee, + +kyong_jin}@korea.ac.kr + +Abstract. Multi frame super-resolution (MFSR) achieves higher performance than single image super-resolution (SISR), because MFSR leverages abundant information from multiple frames. Recent MFSR approaches adapt the deformable convolution network (DCN) to align the frames. However, the existing MFSR suffers from misalignments between the reference and source frames due to the limitations of DCN, such as small receptive fields and the predefined number of kernels. From these problems, existing MFSR approaches struggle to represent high-frequency information. To this end, we propose Deep Burst Multi-scale SR using Fourier Space with Optical Flow (BurstM). The proposed method estimates the optical flow offset for accurate alignment and predicts the continuous Fourier coefficient of each frame for representing high-frequency textures. In addition, we have enhanced the network's flexibility by supporting various super-resolution (SR) scale factors with the unimodel. We demonstrate that our method has the highest performance and flexibility than the existing MFSR methods. Our source code is available at https://github.com/Egkang-Luis/burstm + +Keywords: Super resolution $\cdot$ Burst super resolution $\cdot$ Implicit neural representation $\cdot$ Fourier features + +# 1 Introduction + +Single image super-resolution (SISR) is the major ill-posed problem in low-level computer vision. SISR aims to enhance the image quality of low-resolution by estimating sub-pixel information. After the success of deep learning, the most SISR [10,25,31,44,45] adapt the learning-based method to estimate high-resolution images. However, SISR has a limitation in high-frequency representation of the capacity due to a lack of information. + +Super-resolution researchers focus on multi-frame super-resolution (MFSR) due to the accurate representation of high-frequency textures. Multiple frames provide additional sub-pixel priors, and MFSR utilizes rich information with alignment between the reference and source frames. However, multiple frames take similar scenes with degradations such as noise and object movement. MFSR requires the accurate alignment performance to utilize low-quality frames, because the inaccurate alignment performance can lead to the blur artifacts in the prediction result. Thus, the alignment process is the critical challenge of MFSR. + +![](images/9b4a1f34c71f3a5e518e3e18bcdbf253510cd5e58ae6516a6cdb6a5aa6c1535d.jpg) +Fig. 1: Overview of proposed Burst SR network. The proposed method (BurstM) predicts high-resolution images from multiple low-resolution images. The neural warping performs high-precision warping with upscaling on Fourier space. Additionally, the neural warping contributes to the prediction of an accurate high-resolution images. + +Like SISR, the latest MFSR [3, 4, 11, 12, 29, 30] employs a deep learning method. They adapt similar architectures, which is complicated feature extraction methods using CNN or transformer [7, 24, 26] and alignment using DCN [9] to reconstruct the high-resolution images. Recent MFSR [11, 12, 29, 30] adapts multi-layer DCN structures to extend a small receptive field of DCN, but even this approach still has limited receptive fields. In addition, the predefined number of DCN kernels restricts representing performance during aligning features. The fixed SR scale is another limitation of existing MFSR due to their upsampling method. Most networks [11, 12, 29, 30] generally adapt pixel shuffle (PS) [37] as an upsampling module due to the effectiveness of PS. Nevertheless, PS restricts the network's flexibility, e.g., $\times 4$ SR, because PS can conduct upsampling by a predetermined scale, which is the square of the channel number. If the network predicts a different SR scale, it must re-train with changing architecture or additional downsampling after a predefined SR scale. + +To this end, we propose Deep Burst Multi-scale SR using Fourier Space with Optical Flow (BurstM), which has a high representation capability and flexible scale SR $(\times 2,\times 3,\times 4)$ with computational efficiency. BurstM adapts Optical Flow [16] as a solution for mitigating the limitations of DCN. Estimated optical flow provides correlated offsets between reference and source frames. In order to represent high-frequency textures using well-aligned information, our network estimates Fourier information. To address the fixed scale factor of MFSR, we employ INR [17] inspired by recent successful researches [21, 23]. + +To summarize, our contribution is as follows: + +- Our network conducts precise alignment using optical flow without DCN on the Fourier space. +- We solve the fixed-scale issue in MFSR and demonstrate our flexible networks of multi-scale SR. +- Extensive experiments demonstrate that our approach outperforms existing approaches in both image quality and computational efficiency. + +# 2 Related Work + +Multi Frame Super-Resolution (MFSR) The first MFSR from Tsai [43] restores high-resolution images in the frequency domain. The following approaches [1, 13, 14, 18, 33] have been introduced using the spatial domain and the maximum posterior. Deep learning-based MFSR approaches [3, 4, 11, 12, 29, 30] have arisen due to higher performance than the conventional methods. Furthermore, since the introduction of DCN [9] as the effective alignment method, many approaches [11, 12, 29, 30, 46] have adopted DCN because a biased position of DCN kernel contributes to outstanding alignment performance. However, the small kernel size and a fixed number of kernels restrict the alignment performance. A small kernel size $(3 \times 3)$ of DCN has a small receptive field, which makes it impossible to cover global alignments. To address it, recent methods [11, 12, 29, 30] employ multi-layer DCN in varying spatial information, but it is not enough. Another issue is the fixed number of DCN kernels. Even if DCN cannot find the relevant information, DCN still needs to use the number of predefined kernels, leading to blurry and noisy images. Our BurstM employs optical flow to acquire a broader receptive field and utilize only one-to-one correlated information to tackle this issue. + +Optical Flow Optical flow is a popular method to describe the pixel-level correspondence of object displacements between frames. The conventional methods [15, 16, 42] address the data fidelity and the regularization term as an energy optimization problem. However, they often struggle to estimate accurate results in various scenarios, such as substantial object displacements or small-sized objects. Recent learning-based methods [35, 40, 41] outperform the conventional method's accuracy and computational cost using various structures such as hierarchical design and progressive estimation. Many tasks, such as video super-resolution [6, 22], action recognition [38], and autonomous driving [19], predict the optical flow to find pixel correspondences. However, MFSR task does not effectively utilize optical flow due to computational cost. On the other hand, the proposed network (BurstM) has a simple structure and lightweights, allowing us to incorporate an optical flow estimator into the network. + +Implicit Neural Representation (INR) One of the popular representation methods for the continuous-domain signal is INR [17]. Neural networks designed for INR learn and represent complex structures implicitly. The various tasks employ INR in image processes such as 2D and 3D shapes. Despite such success, INR parameterized by a standalone MLP with the ReLU has a spectral bias problem [32, 34, 39] that fails to represent high-frequency details known as a spectral bias. Recent approaches [8, 21, 23] tackle a spectral bias through positional encoding and Fourier feature mapping, demonstrating a flexible SR network. We have adopted INR method to utilize the flexible characteristic. + +![](images/dc66da2460fb6f87ef3e7f26746873f6239f1226fab9ddf7dbb98475eeada02a.jpg) +Fig. 2: Proposed BurstM Network. The network estimates optical flow between the reference and source frames using FNet. The neural warping $\mathbf{T}_{\nu}$ emphasizes the high-frequency details and implements the warping on Fourier space. Reconstruction module $\mathbf{R}_{\varphi}$ includes the blender $\mathbf{B}_{\eta}$ and the decoder $\mathbf{G}_{\Theta}$ . The blender $\mathbf{B}_{\eta}$ matches color differences and merges multiple frames into a single frame. The Decoder $\mathbf{G}_{\Theta}$ predicts the final output. The gray color indicates a single operation such a convolution and LeakyReLU activation and the others include nonlinear activation + +# 3 Problem Formulation + +Our purpose is to reconstruct the high-resolution image $\mathbf{I}^{\mathbf{SR}} \in \mathbb{R}^{3 \times 2sH \times 2sW}$ with multiple scale factors $s$ from the set of a low-resolution images $\{\mathbf{I}_i^{\mathbf{L}}\}_{i=1}^N$ , where $\mathbf{I}_i^{\mathbf{L}} \in \mathbb{R}^{4 \times H \times W}$ , $N$ is the number of frame and $H, W$ are the spatial sizes of low resolution image, and 4-channels are due to 'RGGB' domain. The dimension of a high-resolution image is increased by a factor of 2s, because the domain is changed from 4-channels 'RGGB' to 3-channels 'sRGB' in the network (in Figure 2). + +# 3.1 Optical Flow Estimator + +We adopt FNet (optical flow estimator) of FRVSR [36]. FNet predicts optical flow on low-resolution grid as below: + +$$ +\pmb {\delta} _ {i} ^ {\mathbf {L}} = F N e t (\mathbf {I} _ {i} ^ {\mathbf {L}}), \quad i = 1, \dots , N, \tag {1} +$$ + +where $\delta_{i}^{\mathbf{L}}\in \mathbb{R}^{H\times W\times 2}$ is the set of optical flow offsets between a reference and source frames on low-resolution grid. Our model adds offsets $\delta_{i}^{\mathbf{L}}$ into regular coordinates $\mathbf{x}_i^{\mathbf{L}}\in [-1,1]^{H\times W\times 2}$ and applies s scale upsampling. + +$$ +\hat {\mathbf {x}} _ {i} ^ {\mathbf {H}} = \mathbf {B _ {s}} (\pmb {\delta} _ {i} ^ {\mathbf {L}} + \mathbf {x} _ {i} ^ {\mathbf {L}}), \quad i = 1, \dots , N, \qquad (2) +$$ + +where $\mathbf{B}_{\mathbf{s}}(\cdot)$ is bilinear upsampling operator with a scale factor of s, and $\hat{\mathbf{x}}_i^{\mathbf{H}} \in \mathbb{R}^{sH \times sW \times 2}$ is the upscaled coordinate with an offset defined by scale factor s. + +# 3.2 Learnable Neural Warping + +Our encoder $(\mathbf{E}_{\psi}(\cdot):\mathbb{R}^4\mapsto \mathbb{R}^{2K})$ , which is a learnable super-resolution (SR) backbone [25] without upsampling, extracts latent vectors $\mathbf{z}_i$ . NW $(\mathbf{T}_{\nu})$ takes the latent vectors $\mathbf{z}_i$ for warping based on a Fourier information. NW consists + +of three components: an amplitude estimator $(\mathbf{g}_{\mathbf{a}}(\cdot):\mathbb{R}^{2K}\mapsto \mathbb{R}^{2K})$ , a frequency estimator $(\mathbf{g}_{\mathbf{f}}(\cdot):\mathbb{R}^{2K}\mapsto \mathbb{R}^{2K})$ and, a phase estimator $(\mathbf{g}_{\mathbf{p}}(\cdot):\mathbb{R}^2\mapsto \mathbb{R}^K)$ . $2K$ denotes the number of channels in the latent vector. Amplitude and frequency estimators utilize the latent vector $\mathbf{z}_i$ to estimate the Fourier coefficients. After that, the nearest interpolation $\mathbf{N_s}$ is performed using estimated coordinates $\hat{\mathbf{x}}_i^{\mathbf{L}}$ . The input of a phase estimator is a cell $\mathbf{c}_i\in \mathbb{R}^{sH\times sW\times 2}$ , which is consist of the divisor of the input frame's spatial sizes $(1 / H,1 / W)$ . As in [8,20], we use relative coordinates $(\delta_{i} = \hat{\mathbf{x}}_{i}^{\mathrm{H}} - \mathbf{x}_{i}^{\mathrm{H}})$ known as a local grid to predict the continuous signal where $\hat{\mathbf{x}}_i^{\mathrm{H}}\in \mathbb{R}^{sH\times sW\times 2}$ . The designed formulation is as follows: + +$$ +\mathbf {z} _ {i} = \mathbf {E} _ {\psi} \left(\mathbf {I} _ {i} ^ {\mathbf {L}}\right), \tag {3} +$$ + +$$ +\mathbf {F} _ {i} = \mathbf {N} _ {\mathbf {s}} \left(\mathbf {g} _ {\mathbf {f}} \left(\mathbf {z} _ {i}\right), \hat {\mathbf {x}} _ {i} ^ {\mathbf {H}}\right), \quad \mathbf {A} _ {i} = \mathbf {N} _ {\mathbf {s}} \left(\mathbf {g} _ {\mathbf {a}} \left(\mathbf {z} _ {i}\right), \hat {\mathbf {x}} _ {i} ^ {\mathbf {H}}\right), \quad \mathbf {p h} _ {i} = \mathbf {g} _ {\mathbf {p}} \left(\mathbf {c} _ {i}\right), \tag {4} +$$ + +$$ +\mathbf {m} _ {i} = \mathbf {A} _ {i} \odot \left[ \begin{array}{l} \cos (\pi \left(\left\langle \mathbf {F} _ {i}, \boldsymbol {\delta} _ {i} \right\rangle + \mathbf {p h} _ {i}\right) \\ \sin (\pi \left(\left\langle \mathbf {F} _ {i}, \boldsymbol {\delta} _ {i} \right\rangle + \mathbf {p h} _ {i}\right) \end{array} \right], \tag {5} +$$ + +$$ +\text {w h e r e ,} \quad i = 1, \dots , N, \quad \boldsymbol {\nu} = [ \mathbf {f}; \mathbf {a}; \mathbf {p} ]. +$$ + +# 3.3 Reconstruction Module + +To reconstruct the high-resolution image using estimated features in Fourier space, we employ the reconstruction module $\mathbf{R}_{\varphi}$ ; Blender $(\mathbf{B}_{\eta}(\cdot):\mathbb{R}^{N\times 2K}\mapsto \mathbb{R}^{1\times 2K})$ and MLP decoder $(\mathbf{G}_{\Theta}(\cdot):\mathbb{R}^{2K}\mapsto \mathbb{R}^{12})$ . + +$$ +\mathbf {h} = \mathbf {R} _ {\varphi} \left(\left\{\mathbf {m} _ {i} \right\} _ {i = 1} ^ {N}\right) = \mathbf {G} _ {\Theta} \left(\mathbf {B} _ {\boldsymbol {\eta}} \left(\left\{\mathbf {m} _ {i} \right\} _ {i = 1} ^ {N}\right)\right), \tag {6} +$$ + +where $\mathbf{h} \in \mathbb{R}^{1 \times 12 \times sH \times sW}$ is the merged Fourier features of burst shots. The blender $\mathbf{B}_{\eta}$ , which is the same module with the encoder $\mathbf{E}_{\psi}$ , reduces color differences between frames and merges multiple frames. MLP decoder $(\mathbf{G}_{\Theta})$ estimates 'sRGB' information. + +# 3.4 Skip connection with upsampling + +To prevent learning on the low-frequency only, we utilize the upscaled skip connection $(\mathbf{Q}_{\pmb{\xi}}(\cdot):\mathbb{R}^{4\times H\times W}\mapsto \mathbb{R}^{12\times sH\times sW})$ about the reference frame $\mathbf{I}_1^{\mathbf{L}}$ . Finally, the spatial size is increased using pixel shuffle $(\mathbf{P_2}(\cdot):\mathbb{R}^{12\times sH\times sW}\mapsto \mathbb{R}^{3\times 2sH\times 2sW})$ . The proposed BurstM can be formulated as follow: + +$$ +\mathbf {I} ^ {\mathbf {S R}} = \mathbf {P} _ {2} (\mathbf {h} + \mathbf {Q} _ {\xi} \left(\mathbf {I} _ {1} ^ {\mathbf {L}}\right)) \tag {7} +$$ + +# 3.5 Loss function + +We use two loss functions for both SyntheticBurst [2] and BurstSR [3]. The MFSR datasets [2,3] does not provide the ground truth of the optical flow map between frames. Thus, we use L2 loss term to use the photometric loss. + +$$ +\mathcal {L} _ {\mathcal {O F}} = \sum_ {i = 2} ^ {N} \left\| W a r p \left(\mathbf {I} _ {i} ^ {\mathbf {L}}, \boldsymbol {\delta} _ {i} ^ {\mathbf {L}}\right) - \mathbf {I} _ {1} ^ {\mathbf {L}} \right\| _ {2} = \sum_ {i = 2} ^ {N} \left(W a r p \left(\mathbf {I} _ {i} ^ {\mathbf {L}}, \boldsymbol {\delta} _ {i} ^ {\mathbf {L}}\right) - \mathbf {I} _ {1} ^ {\mathbf {L}}\right) ^ {2}, \tag {8} +$$ + +where $\text{Warp}(\cdot, \cdot)$ is the bilinear warping operator using optical flow from FNet, and $\mathbf{I}_1^{\mathbf{L}}$ is the reference frame. + +SyntheticBurst dataset Input frames are generated from the ground truth. Thus, we update our model using L1 loss between predicted image and the ground truth. + +$$ +\mathcal {L} _ {\mathcal {S R}} = \left\| \mathbf {I} ^ {\mathbf {S R}} - \mathbf {I} ^ {\mathbf {G T}} \right\| _ {1}. \tag {9} +$$ + +Thus, the total loss function during training is as follows: + +$$ +\mathcal {L} _ {\text {S y n t h e t i c}} = \mathcal {L} _ {\mathcal {S R}} + \mathcal {L} _ {\mathcal {O F}}. \tag {10} +$$ + +BurstSR dataset We employs an aligned L1 loss for the BurstSR [3] dataset due to mismatching between low-resolution images and a ground truth image. An aligned L1 loss utilizes pre-trained weight of PwCNet [40] as follows: + +$$ +\boldsymbol {\delta} ^ {\mathbf {A}} = P W C N e t \left(\mathbf {I} ^ {\mathbf {S R}}, \mathbf {I} ^ {\mathbf {G T}}\right), \tag {11} +$$ + +$$ +\mathcal {L} _ {\text {A l i g n e d S R}} = \left\| W a r p \left(\mathbf {I} ^ {\mathbf {S R}}, \boldsymbol {\delta} ^ {\mathbf {A}}\right) - \mathbf {I} ^ {\mathbf {G T}} \right\| _ {1}, \tag {12} +$$ + +where $\delta^{\mathbf{A}}$ is optical flow between the predicted image on high-resolution grid and ground truth using PwCNet [40]. Total loss function for BurstSR [3] dataset is as follows: + +$$ +\mathcal {L} _ {\text {R e a l}} = \mathcal {L} _ {\text {A l i g n e d S R}} + \mathcal {L} _ {\mathcal {O F}}. \tag {13} +$$ + +# 4 Experiment + +# 4.1 Network details + +Optical Flow Estimator (FNet) Our network adapts FNet from FRVSR [36]. To apply FNet into BurstSR task, we modify the stem of FNet from 'sRGB' to 'RGGB' images. + +Learnable Neural Warping $(\mathbf{T}_{\nu})$ NW estimates the amplitude, frequency, and phase from burst shots with a single convolution operation to represent the Fourier information. Amplitude and frequency estimators have 128 channels of the convolutions with $3\times 3$ kernel, phase estimator has 64 channels of the convolutions with $1\times 1$ kernel. + +Encoder $(\mathbf{E}_{\psi})$ and Reconstruction module $(\mathbf{R}_{\varphi})$ The reconstruction module consists of Blender $\mathbf{B}_{\eta}$ and MLP decoder $\mathbf{G}_{\Theta}$ . Both Encoder $\mathbf{E}_{\psi}$ and Blender $\mathbf{B}_{\eta}$ are EDSR [25] without upsampling module and 128 channels. MLP Decoder, $\mathbf{G}_{\Theta}$ , is organized by 4-layer of $1\times 1$ convolution with 256 channels and ReLU activation. + +# 4.2 Implementation details + +BurstM is an end-to-end trainable network and is possible to train without pre-trained weights. We apply an AdamW optimizer [28] with the Cosine annealing strategy [27] from $10^{-4}$ to $10^{-6}$ learning rate. For SyntheticBurst [2], we + +train BurstM with L1 loss for the output image and L2 loss for optical flows. For BurstSR [3], We fine-tune the pre-trained BurstM of SyntheticBurst with L2 loss for optical flow and aligned L1 loss [3] for the output image. Aligned L1 loss utilizes PWCNet [40] for the output image. The reason for the usage of the aligned L1 Loss is that there are no LR images aligned with the GT image because they were captured using different cameras. We train our model using four RTX3090 GPUs and training times are 10 days for SyntheticBurst and 8 hours for BurstSR. The summary of the training strategy is in Table 1. + +
SyntheticBurst [2]BurstSR [3]
Training Epoch30025
OptimizerADAMW [28]
Learning rate10-4~10-6
LR schedulerCosine annealing [27]
Batch sizes4
Main LossL1Aligned L1
Photo metric LossL2
×2:96×96
Input sizes×3:64×64×4:80×80
×4:48×48
Training time10 days8 hours
+ +Table 1: Training Strategies. Training epoch and main loss, input sizes are different. Except for these factors, we use same training strategies for both SyntheticBurst [2] for BurstSR [3] datasets. + +SyntheticBurst Dataset [2] The dataset contains burst frames which are 46,839 set for training and 300 set for validation. Each burst set contains 14 low-resolution burst frames, which are generated from 'sRGB' as follows. We utilized the unprocessing [5]. Subsequently, we apply random rotations and translations. We employ random down-sampling $(\times 2, \times 3, \times 4)$ to generate low-resolution images of multiple scales. Finally, we apply Bayer mosaicking and add random noise. We trained the proposed model with batch size 4 during 300 epochs, and the training patch size is changed per scale factors in Table 1. + +BurstSR Dataset [3] Input and GT are captured by different cameras: Smartphone and DSLR, respectively. Both low and high resolution images have 200 raw burst sets. The low-resolution set contains 14 frames, and we extract 5,405 patches for training and 882 patches for validation. Unlike SyntehticBurst dataset [2], the scale factor between low-resolution and high-resolution is fixed. Thus, we train BurstM only for $\times 4$ scale factor. Nevertheless, BurstM can accurately predict high-resolution images for multiple scale factors ( $\times 2$ , $\times 3$ , $\times 4$ ), including clear representation of high-frequency textures without the boundary artifacts. + +# 4.3 Validation Results + +We show the qualitative comparison with the existing methods [3, 4, 11, 12, 29, 30] and the visual comparison with latest three methods [11, 12, 29] using the pre-trained weights from authors. + +
# Params. MSyntheticBurstBurstSR x4 PSNR SSIM
x2 PSNRSSIMx3 PSNRSSIMx4 PSNR
Bicubic-38.300.94833.940.88633.020.862
DBSR [3]13.0140.510.96540.110.95940.760.959
MFIR [4]12.1341.250.97141.810.97241.560.964
BIPNet†[11]6.737.580.92840.830.95541.930.960
GMTNet†[30]-----42.360.961
Burststormer†[12]2.537.060.92540.260.95342.830.973
BSRT-Small†[29]4.9240.640.96642.300.97542.720.971
BSRT-Large†[29]20.7140.330.96542.870.97943.620.975
BurstM (Ours)14.046.010.98544.790.98242.870.973
+ +Table 2: Quantitative comparison on SyntheticBurst and BurstSR. Quantitative comparisons of BurstSR such as PSNR and SSIM can only be measured for $\times 4$ , due to LR and HR being provided in fixed sizes. Red and Blue colors indicate the best and the second performance, respectively. $\dagger$ indicates Transformer-based method. + +![](images/574d07622ca736affaae30f009aa0ec1b7e009554ab9ca5474254a254d5ddbf6.jpg) +Fig. 3: Visual comparison for $\times 4$ super-resolution for BurstSR. BurstM shows clearer textures and correct color than the others without the boundary artifacts. + +SR on BurstSR [3] Table 2 is the quantitative comparison result of $\times 4$ SR on BurstSR [3]. Ours shows a performance gain of $+0.17\mathrm{dB}$ compared to the state-of-the-art baselines. As shown in Figure 3, BurstM predicts outperform high-frequency details without the boundary artifacts from low-resolution images of real-world dataset. Also, BurstM predicts multi-scale SR without an additional processing, different with existing models. However, the quantitative comparison of $\times 2$ , $\times 3$ is not possible, because there is no GT of both $\times 2$ , $\times 3$ SR of given dataset [3]. We demonstrate the robustness of our approach by representing high-frequency details in Figure 4 and computation time in Table 4. BurstM directly predicts multiple scales, while other methods implement $\times 4$ SR followed by bicubic downsampling. In Table 4, BurstM shows shorter computation time than the existing methods. + +![](images/f42d53edad15dd700baa4f107b1b5759f6d8e8ab8fc757ca4eee07ce05b0254c.jpg) +Fig. 4: Visual comparison of multi-scale SR on BurstSR. BurstM predict several scaled-images with unimodel and the others are downsampling after $\times 4$ SR. + +SR on SyntheticBurst [2] Table 2 is the quantitative comparison results of multiple scale factors $(\times 2, \times 3, \times 4)$ SR and Figure 5 is the visual comparison result of $\times 4$ SR on SyntheticBurst [2]. Our model achieves second result with reasonable the number of parameters. In addition, we compare the performance on multiple scale factors in Table 2 and Figure 6. BurstM predicts high-resolution images on multiple scale factors with an unimodel and the others implement multiple SR with additional bicubic downsampling after $\times 4$ SR. In case of BurstM, it is observed that PSNR gradually increases as the scale factor decrease. However, PSNR of existing methods decreases as the scale factor decreases because of the information loss during the additional bicubic downsampling process. Figure 6 shows the residual images, which means the loss of the high-frequency textures. Furthermore, as shown in Table 4, BurstM outperforms the other methods in terms of computation time. + +Fourier feature space Figure 7 demonstrates the estimation of high-frequency details on Fourier space. The result from each frame do not illustrate the DFT spectrum, but accumulated result of all frames shows a similar result with the DFT spectrum. Because BurstM leverages sub-pixel information from burst frames to represent the high-resolution images. + +![](images/a43c4c5dab9a0ffbb77b28c856b37399c91cb0fc4343b6c985c9fcccfc683345.jpg) +Fig. 5: Visual comparison for $\times 4$ Super-Resolution for SynteticBurst. + +![](images/53409e3e489013298822216c8c9c7c94da78f30da1d6b313242a8ebbd441d37d.jpg) +Fig.6: Visual comparison for multi-scale SR on SyntheticBurst. The colorbar indicates a difference with GT. The existing methods lose high-frequency textures during additional down-sampling because the highlighted locations are a high-frequency area. But BurstM represent the similar high-frequency textures with GT by because of direct prediction on multiple scale factors. + +Aligning performances We conduct a comparative analysis of the alignment performance between DCN-based method and BurstM in Figure 8. As in [11, 12, 29, 30], we utilize multi-layer DCN with varied spatial sizes for DCN-based method. In contrast, BurstM implements a single image warping process using predicted optical flows. The 'Before align.' and 'After align.' residual maps in Figure 8 represent the differences with the reference frame at each position in + +![](images/1545ca80d0bf01163304d68d8226c7125c00d978fc38aef0d0ba0a967417df3c.jpg) +Fig.7: Fourier space analysis. Accumulated frequencies and coefficients from BurstM are analogous to DFT spectrum of GT (Red color in 'DFT' and 'ACCUM'). + +flowchart. The red dots in 'DCN method flowchart' illustrate offsets to the center pixel. DCN uses an unrelated pixel information, because DCN requires to use the number of predefined kernels regardless of the correlation. Therefore, the aligned results are inaccurate and show high values in the residual map. Additionally, boundary artifacts are also observed due to the utilization of irrelevant information. The color map and white arrows in 'BurstM method flowchart' are the visualization results of optical flow. BurstM method shows one-to-one correspondences between pixels. Thus, the alignment results demonstrate clearer and sharper compared to the DCN-based method without the boundary artifacts. + +![](images/3754588c6c75114624fc2e0033f7944a0bef59938eb6d9bae41a23ba4be0a3b6.jpg) +DCN method flowchart [12] BurstM method (ours) flowchart + +![](images/3587067fb3f61041baadd4da369409c8f6af3838944346d9d1d68a842f92bbf4.jpg) + +![](images/17c08bbf5c1850793e2a9df5dcce235b3a53b72476b06336a44a4e1c8b8522dc.jpg) +Fig. 8: Feature alignment results. Flow charts indicate where we take features. The residual map is between the reference and source frames at each location in flowchart. DCN based-method has a big difference after an alignment because DCN take unrelated information during alignment. However, BurstM point out one-to-one correspondences between a reference and source frames. BurstM shows accurate align result without the boundary artifacts. + +# 5 Ablation study + +The comparison between with and without Fourier mapping INR [17] is the most representative method for representing continuous signals, but it has a spectral bias [32, 34, 39]. As shown in [21, 23], BurstM also addresses this issue through Fourier mapping. Table 3a compares the performance of BurstM with and without Fourier mapping. BurstM without Fourier mapping has significantly lower PSNR values. This is because the network without Fourier mapping learns more about low frequencies due to spectral bias during training. + +Relationship between flexibility and performance BurstM is supporting the multiple scale SR $(\times 2, \times 3, \times 4)$ . The performance in Table 2 has trained with three multiple scale factors $(\times 2, \times 3, \times 4)$ . We train with different number of scale factors such as one $(\times 4)$ and two $(\times 3, \times 4)$ items. In such cases, we observed that PSNR increases on in-of-scale, but performance decreases on out-of-scale. When there are many instances of out-of-scale, the model's flexibility diminishes, so we proposed three scales satisfy both performance and flexibility. + +Different optical flow estimator While FNet [36] estimates an optical flow based on given image sizes, the other optical flow estimator [35] adjust image sizes according to its setting before input. This resizing of images introduces aliasing effects, resulting in a decline in SR performance. Additionally, BurstM with Spynet [35] shows longer inference time than the proposed method. + +
BurstM w\o Fourier mappingBurstM w\Fourier mapping
×4 PSNR39.82 dB41.74 dB
+ +(a) The performance comparison between with and without Fourier mapping + +
Training scale×2×3×4
×433.3635.9742.26
×3, ×430.8043.6141.82
×2, ×3, ×4 (Ours)44.7743.4541.74
+ +(b) Relationship between flexibility and performance (PSNR:dB) + +
BurstM +w\SpynetBurstM +w\FNet (Ours)
×4 PSNR40.85 dB41.74 dB
Inference time14.3 ms11.6 ms
+ +(c) Performance comparison based on optical flow estimator + +Table 3: Ablation study summary + +(a) Fourier mapping contributes to the model performance. +(b) Red color indicates the result of in-scale. Less training scale has higher PSNR on in-scale but network flexibility is low due to increasing out-scale. Thus, we select our model with 3 scale factors $(\times 2, \times 3, \times 4)$ considering both flexibility and performance. +(c) A different optical flow estimator leads to PSNR drop and an increase in inference time. + +# 6 Discussion + +Boundary line artifacts As shown in Figure 9, the recent method [29] exhibits boundary artifacts despite utilizing optical flow, due to its continued reliance on DCN [9] for alignment. As illustrated in Figure 8 and 9, DCN [9] must utilize information corresponding to the predefined kernel size regardless of correlation, resulting in the incorporation of unnecessary information when most frames fail to provide essential data. In contrast, BurstM aligns frames without DCN [9], + +solely utilizing optical flow. Consequently, BurstM avoids the use of irrelevant information, leading to accurate predictions without boundary artifacts. + +![](images/ee27cf3636ffa531cec16670c108357ef31b6b4202b602606ccbb9d97ef6c4a7.jpg) +Fig.9: Boundary artifacts The most of frames didn't provide the information of edge. The existing method [29] suffers the boundary line artifacts, but BurstM represents clear images without the boundary line artifacts. + +Computation complexity Despite advancements in computational system, SR field still requires shorter computation times as the image resolutions increase. Table 4 compares the computation time usage of BurstM with the other methods [11,12,29]. We implements $\times 4$ SR with $80 \times 80$ frame sizes of BurstSR dataset [3] on NVIDIA RTX3090 24GB. While BurstM has a large number of parameters, BurstM demonstrates the shortest computation time. + +![](images/bbea5824731e2b6578ed494a066e72cc7506d6a72027f753db81fd457d0b8e56.jpg) +Table 4: Computational complexity on one GPU (RTX3090). Despite a large number of parameters, BurstM shows the inference time (ms) and best PSNR (dB) performance. Red : best metrics. † indicates Transformer-based method. + +
SR scale: ×4Params.(M) / Mem.(GB)Inference time(ms)PSNR (dB)
BIPNet† [11]6.7 / 11.820.848.49
Burstormer† [12]2.5 / 12.020.848.82
BSRT-S† [29]4.92 / 1.0148.048.48
BSRT-L† [29]20.71 / 1.5266.148.57
BurstM (Ours)14.9 / 8.411.649.12
+ +Super-Resolution with floating scale factors BurstM can estimate the multi-scale SR with floating scale factors such as $\times 3.5$ without any architecture changes. However, the training of the floating scale factors ( $\times 2\sim \times 4$ ) shows lower PSNR (39.44dB) than the training of the integer scale factors (41.74dB). The performance drop is caused by the demosaic process when generating low-resolution frames. To convert from 'sRGB' to 'RGGB', the dimensions (H, W) of the images are changed to (H/2, W/2). If the image size of 'sRGB' is not divisible by two when a floating scale factor is applied, the network apply an additional + +downsampling operation to make the dimension of $(\mathrm{H} / 2, \mathrm{W} / 2)$ . The double downsampling imposes a severe degradation of the generated low-resolution images, and it results in low PSNRs. + +**Corner case** The major challenge on optical flow is occlusions which are caused by a new object that is not present in the reference frame. BurstM also has the occlusion issue (Figure 10) even though BurstM demonstrate accurate alignment performance than DCN based methods. Fortunately, BurstM utilize inaccurate optical flow of an occlusion as meaningful prior and reconstruct the high-quality images in a high-resolution grid. + +![](images/04b3ada32e3a25a2b1902f784b7241a3a5860dc8209a631f673bb473c0f85768.jpg) +Fig. 10: Corner case. Inaccurate optical flow prediction due to occlusions which are caused by new objects (red arrow) not present in the reference image appear. + +# 7 Conclusion + +In this paper, we proposed Deep Burst Multi-scale SR using Fourier Space with Optical Flow (BurstM) to overcome the limitations of existing methods: SR of fixed scale factor $(\times 4)$ and low-quality of real-world dataset [3]. Our network performs the precise alignment using optical flow and the advanced warping process based on estimated Fourier information. We demonstrated their effectiveness in representing the high-frequency textures with $+0.17\mathrm{dB}$ PSNR gain in the real-world dataset [3] compared to the previous state-of-the-art model. Furthermore, BurstM successfully implements multi-scale SR $(\times 2,\times 3,\times 4)$ of the unimodel by representing high-frequency details using the INR method. The function of multi-scale SR enhances our model's capacity compared to the other learning-based methods. In addition, we demonstrate our approach is faster inference time compared to the existing methods by the efficient computation. + +Acknowledgement This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (RS-2021-II212068, Artificial Intelligence Innovation Hub) and the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (RS-2024-00335741). + +# References + +1. Bascle, B., Blake, A., Zisserman, A.: Motion deblurring and super-resolution from an image sequence. In: Computer Vision—ECCV'96: 4th European Conference on Computer Vision Cambridge, UK, April 15–18, 1996 Proceedings Volume II 4. pp. 571–582. Springer (1996) +2. Bhat, G., Danelljan, M., Timofte, R.: Ntire 2021 challenge on burst superresolution: Methods and results. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 613-626 (2021) +3. Bhat, G., Danelljan, M., Van Gool, L., Timofte, R.: Deep burst super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9209-9218 (2021) +4. Bhat, G., Danelljan, M., Yu, F., Van Gool, L., Timofte, R.: Deep reparametrization of multi-frame super-resolution and denoising. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2460-2470 (2021) +5. Brooks, T., Mildenhall, B., Xue, T., Chen, J., Sharlet, D., Barron, J.T.: Unprocessing images for learned raw denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11036-11045 (2019) +6. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basicvsr: The search for essential components in video super-resolution and beyond. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4947-4956 (2021) +7. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-Trained Image Processing Transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12299-12310 (2021) +8. Chen, Y., Liu, S., Wang, X.: Learning continuous image representation with local implicit image function. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8628-8638 (2021) +9. Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., Wei, Y.: Deformable convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 764-773 (2017) +0. Dai, T., Cai, J., Zhang, Y., Xia, S.T., Zhang, L.: Second-order attention network for single image super-resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11065-11074 (2019) +1. Dudhane, A., Zamir, S.W., Khan, S., Khan, F.S., Yang, M.H.: Burst image restoration and enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5759-5768 (2022) +2. Dudhane, A., Zamir, S.W., Khan, S., Khan, F.S., Yang, M.H.: Burstomer: Burst image restoration and enhancement transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5703-5712 (2023) +3. Elad, M., Feuer, A.: Restoration of a single superresolution image from several blurred, noisy, and undersampled measured images. IEEE transactions on image processing 6(12), 1646-1658 (1997) +4. Hardie, R.C., Barnard, K.J., Bognar, J.G., Armstrong, E.E., Watson, E.A.: High-resolution image reconstruction from a sequence of rotated and translated frames and its application to an infrared imaging system. Optical Engineering 37(1), 247-260 (1998) +5. Hirschmuller, H.: Stereo processing by semiglobal matching and mutual information. IEEE Transactions on pattern analysis and machine intelligence 30(2), 328-341 (2007) + +16. Horn, B.K., Schunck, B.G.: Determining optical flow. Artificial intelligence 17(1-3), 185-203 (1981) +17. Hornik, K., Stinchcombe, M., White, H.: Multilayer feedforward networks are universal approximators. Neural networks 2(5), 359-366 (1989) +18. Irani, M., Peleg, S.: Improving resolution by image registration. CVGIP: Graphical models and image processing 53(3), 231-239 (1991) +19. Janai, J., Güney, F., Behl, A., Geiger, A., et al.: Computer vision for autonomous vehicles: Problems, datasets and state of the art. Foundations and Trends in Computer Graphics and Vision 12(1-3), 1-308 (2020) +20. Jiang, C.M., Sud, A., Makadia, A., Huang, J., Niessner, M., Funkhouser, T.: Local Implicit Grid Representations for 3D Scenes. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020) +21. Kim, M., Lee, J., Lee, B., Im, S., Jin, K.H.: Implicit neural image stitching with enhanced and blended feature reconstruction (2023) +22. Lai, W.S., Huang, J.B., Ahuja, N., Yang, M.H.: Deep laplacian pyramid networks for fast and accurate super-resolution. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 624-632 (2017) +23. Lee, J., Jin, K.H.: Local texture estimator for implicit representation function. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1929-1938 (2022) +24. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: SwinIR: Image Restoration Using Swin Transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 1833-1844 (2021) +25. Lim, B., Son, S., Kim, H., Nah, S., Mu Lee, K.: Enhanced deep residual networks for single image super-resolution. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 136-144 (2017) +26. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10012-10022 (2021) +27. Loshchilov, I., Hutter, F.: Stochastic gradient descent with warm restarts. In: Proceedings of the 5th Int. Conf. Learning Representations. pp. 1-16 +28. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) +29. Luo, Z., Li, Y., Cheng, S., Yu, L., Wu, Q., Wen, Z., Fan, H., Sun, J., Liu, S.: BSRT: Improving burst super-resolution with swim transformer and flow-guided deformable alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 998-1008 (2022) +30. Mehta, N., Dudhane, A., Murala, S., Zamir, S.W., Khan, S., Khan, F.S.: Gated multi-resolution transfer network for burst restoration and enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22201-22210 (2023) +31. Mei, Y., Fan, Y., Zhou, Y.: Image super-resolution with non-local sparse attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3517-3526 (2021) +32. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) +33. Peleg, S., Keren, D., Schweitzer, L.: Improving image resolution using subpixel motion. Pattern recognition letters 5(3), 223-226 (1987) + +34. Rahaman, N., Baratin, A., Arpit, D., Draxler, F., Lin, M., Hamprecht, F., Bengio, Y., Courville, A.: On the Spectral Bias of Neural Networks. In: Chaudhuri, K., Salakhutdinov, R. (eds.) Proceedings of the 36th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 97, pp. 5301-5310. PMLR (2019) +35. Ranjan, A., Black, M.J.: Optical flow estimation using a spatial pyramid network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4161-4170 (2017) +36. Sajjadi, M.S., Vemulapalli, R., Brown, M.: Frame-recurrent video super-resolution. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6626-6634 (2018) +37. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1874-1883 (2016) +38. Simonyan, K., Zisserman, A.: Two-stream convolutional networks for action recognition in videos. Advances in neural information processing systems 27 (2014) +39. Sitzmann, V., Martel, J., Bergman, A., Lindell, D., Wetzstein, G.: Implicit Neural Representations with Periodic Activation Functions. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M.F., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 7462-7473. Curran Associates, Inc. (2020) +40. Sun, D., Yang, X., Liu, M.Y., Kautz, J.: Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8934-8943 (2018) +41. Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 402-419. Springer (2020) +42. Triggs, B., McLauchlan, P.F., Hartley, R.I., Fitzgibbon, A.W.: Bundle adjustment—a modern synthesis. In: Vision Algorithms: Theory and Practice: International Workshop on Vision Algorithms Corfu, Greece, September 21–22, 1999 Proceedings. pp. 298–372. Springer (2000) +43. Tsai, R.Y., Huang, T.S.: Multiframe image restoration and registration. Multiframe image restoration and registration 1, 317-339 (1984) +44. Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image super-resolution using very deep residual channel attention networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 286-301 (2018) +45. Zhang, Y., Tian, Y., Kong, Y., Zhong, B., Fu, Y.: Residual dense network for image super-resolution. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2472-2481 (2018) +46. Zhang, Z., Wang, R., Zhang, H., Chen, Y., Zuo, W.: Self-supervised learning for real-world super-resolution from dual zoomed observations. In: European Conference on Computer Vision. pp. 610-627. Springer (2022) \ No newline at end of file diff --git a/2024/BurstM_ Deep Burst Multi-scale SR using Fourier Space with Optical Flow/images.zip b/2024/BurstM_ Deep Burst Multi-scale SR using Fourier Space with Optical Flow/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..76e9a5f3f79bec77cbb4c15d77a0555395ef01bf --- /dev/null +++ b/2024/BurstM_ Deep Burst Multi-scale SR using Fourier Space with Optical Flow/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8dc3334e4eeafa6f378f2b20b6472dffe0c4575c75ffab2ed53761fb6d44a88 +size 980556 diff --git a/2024/BurstM_ Deep Burst Multi-scale SR using Fourier Space with Optical Flow/layout.json b/2024/BurstM_ Deep Burst Multi-scale SR using Fourier Space with Optical Flow/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bcffba65224582f9ee9dfd0b1a14dae74395f487 --- /dev/null +++ b/2024/BurstM_ Deep Burst Multi-scale SR using Fourier Space with Optical Flow/layout.json @@ -0,0 +1,9282 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "type": "text", + "content": "BurstM: Deep Burst Multi-scale SR using Fourier Space with Optical Flow" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 167, + 481, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 481, + 181 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 481, + 181 + ], + "type": "text", + "content": "EungGu Kang¹ Byeonghun Lee² Sunghoon Im1* Kyong Hwan Jin2*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 156, + 189, + 457, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 189, + 457, + 200 + ], + "spans": [ + { + "bbox": [ + 156, + 189, + 457, + 200 + ], + "type": "text", + "content": "1 Daegu Gyeongbuk Institute of Science and Technology (DGIST), Korea" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 253, + 201, + 361, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 201, + 361, + 212 + ], + "spans": [ + { + "bbox": [ + 253, + 201, + 361, + 212 + ], + "type": "text", + "content": "2 Korea University, Korea" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 161, + 213, + 449, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 213, + 449, + 223 + ], + "spans": [ + { + "bbox": [ + 161, + 213, + 449, + 223 + ], + "type": "text", + "content": "eunggukang@gmail.com, sunghoonim@dgist.ac.kr, {byeonghun_lee," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 253, + 224, + 359, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 224, + 359, + 234 + ], + "spans": [ + { + "bbox": [ + 253, + 224, + 359, + 234 + ], + "type": "text", + "content": "kyong_jin}@korea.ac.kr" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 236, + 453, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 236, + 453, + 422 + ], + "spans": [ + { + "bbox": [ + 160, + 236, + 453, + 422 + ], + "type": "text", + "content": "Abstract. Multi frame super-resolution (MFSR) achieves higher performance than single image super-resolution (SISR), because MFSR leverages abundant information from multiple frames. Recent MFSR approaches adapt the deformable convolution network (DCN) to align the frames. However, the existing MFSR suffers from misalignments between the reference and source frames due to the limitations of DCN, such as small receptive fields and the predefined number of kernels. From these problems, existing MFSR approaches struggle to represent high-frequency information. To this end, we propose Deep Burst Multi-scale SR using Fourier Space with Optical Flow (BurstM). The proposed method estimates the optical flow offset for accurate alignment and predicts the continuous Fourier coefficient of each frame for representing high-frequency textures. In addition, we have enhanced the network's flexibility by supporting various super-resolution (SR) scale factors with the unimodel. We demonstrate that our method has the highest performance and flexibility than the existing MFSR methods. Our source code is available at https://github.com/Egkang-Luis/burstm" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 432, + 453, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 432, + 453, + 453 + ], + "spans": [ + { + "bbox": [ + 159, + 432, + 453, + 453 + ], + "type": "text", + "content": "Keywords: Super resolution " + }, + { + "bbox": [ + 159, + 432, + 453, + 453 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 432, + 453, + 453 + ], + "type": "text", + "content": " Burst super resolution " + }, + { + "bbox": [ + 159, + 432, + 453, + 453 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 432, + 453, + 453 + ], + "type": "text", + "content": " Implicit neural representation " + }, + { + "bbox": [ + 159, + 432, + 453, + 453 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 432, + 453, + 453 + ], + "type": "text", + "content": " Fourier features" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 461, + 230, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 230, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 230, + 474 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 480, + 481, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 481, + 552 + ], + "type": "text", + "content": "Single image super-resolution (SISR) is the major ill-posed problem in low-level computer vision. SISR aims to enhance the image quality of low-resolution by estimating sub-pixel information. After the success of deep learning, the most SISR [10,25,31,44,45] adapt the learning-based method to estimate high-resolution images. However, SISR has a limitation in high-frequency representation of the capacity due to a lack of information." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 552, + 481, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 481, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 481, + 648 + ], + "type": "text", + "content": "Super-resolution researchers focus on multi-frame super-resolution (MFSR) due to the accurate representation of high-frequency textures. Multiple frames provide additional sub-pixel priors, and MFSR utilizes rich information with alignment between the reference and source frames. However, multiple frames take similar scenes with degradations such as noise and object movement. MFSR requires the accurate alignment performance to utilize low-quality frames, because the inaccurate alignment performance can lead to the blur artifacts in the prediction result. Thus, the alignment process is the critical challenge of MFSR." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 654, + 251, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 251, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 251, + 665 + ], + "type": "text", + "content": "\\* Co-corresponding authors." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 186, + 115, + 430, + 237 + ], + "blocks": [ + { + "bbox": [ + 186, + 115, + 430, + 237 + ], + "lines": [ + { + "bbox": [ + 186, + 115, + 430, + 237 + ], + "spans": [ + { + "bbox": [ + 186, + 115, + 430, + 237 + ], + "type": "image", + "image_path": "9b4a1f34c71f3a5e518e3e18bcdbf253510cd5e58ae6516a6cdb6a5aa6c1535d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 247, + 482, + 292 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 292 + ], + "type": "text", + "content": "Fig. 1: Overview of proposed Burst SR network. The proposed method (BurstM) predicts high-resolution images from multiple low-resolution images. The neural warping performs high-precision warping with upscaling on Fourier space. Additionally, the neural warping contributes to the prediction of an accurate high-resolution images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 300, + 482, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 467 + ], + "type": "text", + "content": "Like SISR, the latest MFSR [3, 4, 11, 12, 29, 30] employs a deep learning method. They adapt similar architectures, which is complicated feature extraction methods using CNN or transformer [7, 24, 26] and alignment using DCN [9] to reconstruct the high-resolution images. Recent MFSR [11, 12, 29, 30] adapts multi-layer DCN structures to extend a small receptive field of DCN, but even this approach still has limited receptive fields. In addition, the predefined number of DCN kernels restricts representing performance during aligning features. The fixed SR scale is another limitation of existing MFSR due to their upsampling method. Most networks [11, 12, 29, 30] generally adapt pixel shuffle (PS) [37] as an upsampling module due to the effectiveness of PS. Nevertheless, PS restricts the network's flexibility, e.g., " + }, + { + "bbox": [ + 130, + 300, + 482, + 467 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 130, + 300, + 482, + 467 + ], + "type": "text", + "content": " SR, because PS can conduct upsampling by a predetermined scale, which is the square of the channel number. If the network predicts a different SR scale, it must re-train with changing architecture or additional downsampling after a predefined SR scale." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 468, + 483, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 483, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 483, + 563 + ], + "type": "text", + "content": "To this end, we propose Deep Burst Multi-scale SR using Fourier Space with Optical Flow (BurstM), which has a high representation capability and flexible scale SR " + }, + { + "bbox": [ + 130, + 468, + 483, + 563 + ], + "type": "inline_equation", + "content": "(\\times 2,\\times 3,\\times 4)" + }, + { + "bbox": [ + 130, + 468, + 483, + 563 + ], + "type": "text", + "content": " with computational efficiency. BurstM adapts Optical Flow [16] as a solution for mitigating the limitations of DCN. Estimated optical flow provides correlated offsets between reference and source frames. In order to represent high-frequency textures using well-aligned information, our network estimates Fourier information. To address the fixed scale factor of MFSR, we employ INR [17] inspired by recent successful researches [21, 23]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 563, + 347, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 563, + 347, + 575 + ], + "spans": [ + { + "bbox": [ + 146, + 563, + 347, + 575 + ], + "type": "text", + "content": "To summarize, our contribution is as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 578, + 479, + 649 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 138, + 578, + 479, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 479, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 479, + 601 + ], + "type": "text", + "content": "- Our network conducts precise alignment using optical flow without DCN on the Fourier space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 602, + 479, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 602, + 479, + 624 + ], + "spans": [ + { + "bbox": [ + 138, + 602, + 479, + 624 + ], + "type": "text", + "content": "- We solve the fixed-scale issue in MFSR and demonstrate our flexible networks of multi-scale SR." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 627, + 479, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 627, + 479, + 649 + ], + "spans": [ + { + "bbox": [ + 138, + 627, + 479, + 649 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that our approach outperforms existing approaches in both image quality and computational efficiency." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "EG Kang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 163, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 163, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 163, + 482, + 368 + ], + "type": "text", + "content": "Multi Frame Super-Resolution (MFSR) The first MFSR from Tsai [43] restores high-resolution images in the frequency domain. The following approaches [1, 13, 14, 18, 33] have been introduced using the spatial domain and the maximum posterior. Deep learning-based MFSR approaches [3, 4, 11, 12, 29, 30] have arisen due to higher performance than the conventional methods. Furthermore, since the introduction of DCN [9] as the effective alignment method, many approaches [11, 12, 29, 30, 46] have adopted DCN because a biased position of DCN kernel contributes to outstanding alignment performance. However, the small kernel size and a fixed number of kernels restrict the alignment performance. A small kernel size " + }, + { + "bbox": [ + 130, + 163, + 482, + 368 + ], + "type": "inline_equation", + "content": "(3 \\times 3)" + }, + { + "bbox": [ + 130, + 163, + 482, + 368 + ], + "type": "text", + "content": " of DCN has a small receptive field, which makes it impossible to cover global alignments. To address it, recent methods [11, 12, 29, 30] employ multi-layer DCN in varying spatial information, but it is not enough. Another issue is the fixed number of DCN kernels. Even if DCN cannot find the relevant information, DCN still needs to use the number of predefined kernels, leading to blurry and noisy images. Our BurstM employs optical flow to acquire a broader receptive field and utilize only one-to-one correlated information to tackle this issue." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 384, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 540 + ], + "type": "text", + "content": "Optical Flow Optical flow is a popular method to describe the pixel-level correspondence of object displacements between frames. The conventional methods [15, 16, 42] address the data fidelity and the regularization term as an energy optimization problem. However, they often struggle to estimate accurate results in various scenarios, such as substantial object displacements or small-sized objects. Recent learning-based methods [35, 40, 41] outperform the conventional method's accuracy and computational cost using various structures such as hierarchical design and progressive estimation. Many tasks, such as video super-resolution [6, 22], action recognition [38], and autonomous driving [19], predict the optical flow to find pixel correspondences. However, MFSR task does not effectively utilize optical flow due to computational cost. On the other hand, the proposed network (BurstM) has a simple structure and lightweights, allowing us to incorporate an optical flow estimator into the network." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": "Implicit Neural Representation (INR) One of the popular representation methods for the continuous-domain signal is INR [17]. Neural networks designed for INR learn and represent complex structures implicitly. The various tasks employ INR in image processes such as 2D and 3D shapes. Despite such success, INR parameterized by a standalone MLP with the ReLU has a spectral bias problem [32, 34, 39] that fails to represent high-frequency details known as a spectral bias. Recent approaches [8, 21, 23] tackle a spectral bias through positional encoding and Fourier feature mapping, demonstrating a flexible SR network. We have adopted INR method to utilize the flexible characteristic." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "BurstM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 113, + 482, + 230 + ], + "blocks": [ + { + "bbox": [ + 135, + 113, + 482, + 230 + ], + "lines": [ + { + "bbox": [ + 135, + 113, + 482, + 230 + ], + "spans": [ + { + "bbox": [ + 135, + 113, + 482, + 230 + ], + "type": "image", + "image_path": "dc66da2460fb6f87ef3e7f26746873f6239f1226fab9ddf7dbb98475eeada02a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 239, + 482, + 318 + ], + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 318 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 318 + ], + "type": "text", + "content": "Fig. 2: Proposed BurstM Network. The network estimates optical flow between the reference and source frames using FNet. The neural warping " + }, + { + "bbox": [ + 130, + 239, + 482, + 318 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\nu}" + }, + { + "bbox": [ + 130, + 239, + 482, + 318 + ], + "type": "text", + "content": " emphasizes the high-frequency details and implements the warping on Fourier space. Reconstruction module " + }, + { + "bbox": [ + 130, + 239, + 482, + 318 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\varphi}" + }, + { + "bbox": [ + 130, + 239, + 482, + 318 + ], + "type": "text", + "content": " includes the blender " + }, + { + "bbox": [ + 130, + 239, + 482, + 318 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\eta}" + }, + { + "bbox": [ + 130, + 239, + 482, + 318 + ], + "type": "text", + "content": " and the decoder " + }, + { + "bbox": [ + 130, + 239, + 482, + 318 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\Theta}" + }, + { + "bbox": [ + 130, + 239, + 482, + 318 + ], + "type": "text", + "content": ". The blender " + }, + { + "bbox": [ + 130, + 239, + 482, + 318 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\eta}" + }, + { + "bbox": [ + 130, + 239, + 482, + 318 + ], + "type": "text", + "content": " matches color differences and merges multiple frames into a single frame. The Decoder " + }, + { + "bbox": [ + 130, + 239, + 482, + 318 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\Theta}" + }, + { + "bbox": [ + 130, + 239, + 482, + 318 + ], + "type": "text", + "content": " predicts the final output. The gray color indicates a single operation such a convolution and LeakyReLU activation and the others include nonlinear activation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 329, + 281, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 329, + 281, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 329, + 281, + 342 + ], + "type": "text", + "content": "3 Problem Formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 349, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 482, + 434 + ], + "type": "text", + "content": "Our purpose is to reconstruct the high-resolution image " + }, + { + "bbox": [ + 130, + 349, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathbf{SR}} \\in \\mathbb{R}^{3 \\times 2sH \\times 2sW}" + }, + { + "bbox": [ + 130, + 349, + 482, + 434 + ], + "type": "text", + "content": " with multiple scale factors " + }, + { + "bbox": [ + 130, + 349, + 482, + 434 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 349, + 482, + 434 + ], + "type": "text", + "content": " from the set of a low-resolution images " + }, + { + "bbox": [ + 130, + 349, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_i^{\\mathbf{L}}\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 349, + 482, + 434 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 349, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i^{\\mathbf{L}} \\in \\mathbb{R}^{4 \\times H \\times W}" + }, + { + "bbox": [ + 130, + 349, + 482, + 434 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 349, + 482, + 434 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 349, + 482, + 434 + ], + "type": "text", + "content": " is the number of frame and " + }, + { + "bbox": [ + 130, + 349, + 482, + 434 + ], + "type": "inline_equation", + "content": "H, W" + }, + { + "bbox": [ + 130, + 349, + 482, + 434 + ], + "type": "text", + "content": " are the spatial sizes of low resolution image, and 4-channels are due to 'RGGB' domain. The dimension of a high-resolution image is increased by a factor of 2s, because the domain is changed from 4-channels 'RGGB' to 3-channels 'sRGB' in the network (in Figure 2)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 441, + 279, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 279, + 454 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 279, + 454 + ], + "type": "text", + "content": "3.1 Optical Flow Estimator" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 461, + 481, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 481, + 485 + ], + "type": "text", + "content": "We adopt FNet (optical flow estimator) of FRVSR [36]. FNet predicts optical flow on low-resolution grid as below:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 235, + 491, + 481, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 491, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 235, + 491, + 481, + 509 + ], + "type": "interline_equation", + "content": "\\pmb {\\delta} _ {i} ^ {\\mathbf {L}} = F N e t (\\mathbf {I} _ {i} ^ {\\mathbf {L}}), \\quad i = 1, \\dots , N, \\tag {1}", + "image_path": "f4e4c0a292a476bef9f90bccf2b2126b24ef3d38ffb94b3f8432db8f8651f20a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 515, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 482, + 554 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 515, + 482, + 554 + ], + "type": "inline_equation", + "content": "\\delta_{i}^{\\mathbf{L}}\\in \\mathbb{R}^{H\\times W\\times 2}" + }, + { + "bbox": [ + 130, + 515, + 482, + 554 + ], + "type": "text", + "content": " is the set of optical flow offsets between a reference and source frames on low-resolution grid. Our model adds offsets " + }, + { + "bbox": [ + 130, + 515, + 482, + 554 + ], + "type": "inline_equation", + "content": "\\delta_{i}^{\\mathbf{L}}" + }, + { + "bbox": [ + 130, + 515, + 482, + 554 + ], + "type": "text", + "content": " into regular coordinates " + }, + { + "bbox": [ + 130, + 515, + 482, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^{\\mathbf{L}}\\in [-1,1]^{H\\times W\\times 2}" + }, + { + "bbox": [ + 130, + 515, + 482, + 554 + ], + "type": "text", + "content": " and applies s scale upsampling." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 228, + 560, + 481, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 560, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 228, + 560, + 481, + 576 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {i} ^ {\\mathbf {H}} = \\mathbf {B _ {s}} (\\pmb {\\delta} _ {i} ^ {\\mathbf {L}} + \\mathbf {x} _ {i} ^ {\\mathbf {L}}), \\quad i = 1, \\dots , N, \\qquad (2)", + "image_path": "52ec0d8bb0e9b33d768174fe5345abb8da28805a3b8ca7d49b020ce8e854927b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 582, + 481, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 608 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 582, + 481, + 608 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathbf{s}}(\\cdot)" + }, + { + "bbox": [ + 130, + 582, + 481, + 608 + ], + "type": "text", + "content": " is bilinear upsampling operator with a scale factor of s, and " + }, + { + "bbox": [ + 130, + 582, + 481, + 608 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_i^{\\mathbf{H}} \\in \\mathbb{R}^{sH \\times sW \\times 2}" + }, + { + "bbox": [ + 130, + 582, + 481, + 608 + ], + "type": "text", + "content": " is the upscaled coordinate with an offset defined by scale factor s." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 615, + 295, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 615, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 132, + 615, + 295, + 627 + ], + "type": "text", + "content": "3.2 Learnable Neural Warping" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Our encoder " + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "(\\mathbf{E}_{\\psi}(\\cdot):\\mathbb{R}^4\\mapsto \\mathbb{R}^{2K})" + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": ", which is a learnable super-resolution (SR) backbone [25] without upsampling, extracts latent vectors " + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": ". NW " + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "(\\mathbf{T}_{\\nu})" + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": " takes the latent vectors " + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": " for warping based on a Fourier information. NW consists" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "EG Kang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": "of three components: an amplitude estimator " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "(\\mathbf{g}_{\\mathbf{a}}(\\cdot):\\mathbb{R}^{2K}\\mapsto \\mathbb{R}^{2K})" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": ", a frequency estimator " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "(\\mathbf{g}_{\\mathbf{f}}(\\cdot):\\mathbb{R}^{2K}\\mapsto \\mathbb{R}^{2K})" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": " and, a phase estimator " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "(\\mathbf{g}_{\\mathbf{p}}(\\cdot):\\mathbb{R}^2\\mapsto \\mathbb{R}^K)" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "2K" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": " denotes the number of channels in the latent vector. Amplitude and frequency estimators utilize the latent vector " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": " to estimate the Fourier coefficients. After that, the nearest interpolation " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{N_s}" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": " is performed using estimated coordinates " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_i^{\\mathbf{L}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": ". The input of a phase estimator is a cell " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i\\in \\mathbb{R}^{sH\\times sW\\times 2}" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": ", which is consist of the divisor of the input frame's spatial sizes " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "(1 / H,1 / W)" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": ". As in [8,20], we use relative coordinates " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "(\\delta_{i} = \\hat{\\mathbf{x}}_{i}^{\\mathrm{H}} - \\mathbf{x}_{i}^{\\mathrm{H}})" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": " known as a local grid to predict the continuous signal where " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_i^{\\mathrm{H}}\\in \\mathbb{R}^{sH\\times sW\\times 2}" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": ". The designed formulation is as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 278, + 233, + 480, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 278, + 233, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 278, + 233, + 480, + 247 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {i} = \\mathbf {E} _ {\\psi} \\left(\\mathbf {I} _ {i} ^ {\\mathbf {L}}\\right), \\tag {3}", + "image_path": "8077d54c670506dbafe1bbcd837b9390e588b268a75279744553fe39634b7e0a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 173, + 250, + 479, + 263 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 250, + 479, + 263 + ], + "spans": [ + { + "bbox": [ + 173, + 250, + 479, + 263 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {i} = \\mathbf {N} _ {\\mathbf {s}} \\left(\\mathbf {g} _ {\\mathbf {f}} \\left(\\mathbf {z} _ {i}\\right), \\hat {\\mathbf {x}} _ {i} ^ {\\mathbf {H}}\\right), \\quad \\mathbf {A} _ {i} = \\mathbf {N} _ {\\mathbf {s}} \\left(\\mathbf {g} _ {\\mathbf {a}} \\left(\\mathbf {z} _ {i}\\right), \\hat {\\mathbf {x}} _ {i} ^ {\\mathbf {H}}\\right), \\quad \\mathbf {p h} _ {i} = \\mathbf {g} _ {\\mathbf {p}} \\left(\\mathbf {c} _ {i}\\right), \\tag {4}", + "image_path": "11924ef498d0632549f26926c9a5b4c7506617d743deab45e4c6f57637592d57.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 228, + 266, + 479, + 292 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 266, + 479, + 292 + ], + "spans": [ + { + "bbox": [ + 228, + 266, + 479, + 292 + ], + "type": "interline_equation", + "content": "\\mathbf {m} _ {i} = \\mathbf {A} _ {i} \\odot \\left[ \\begin{array}{l} \\cos (\\pi \\left(\\left\\langle \\mathbf {F} _ {i}, \\boldsymbol {\\delta} _ {i} \\right\\rangle + \\mathbf {p h} _ {i}\\right) \\\\ \\sin (\\pi \\left(\\left\\langle \\mathbf {F} _ {i}, \\boldsymbol {\\delta} _ {i} \\right\\rangle + \\mathbf {p h} _ {i}\\right) \\end{array} \\right], \\tag {5}", + "image_path": "a10db84098c5c611d445aa71fc67d86b9a6b57b91dbc835075af65bef5674503.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 227, + 295, + 386, + 307 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 295, + 386, + 307 + ], + "spans": [ + { + "bbox": [ + 227, + 295, + 386, + 307 + ], + "type": "interline_equation", + "content": "\\text {w h e r e ,} \\quad i = 1, \\dots , N, \\quad \\boldsymbol {\\nu} = [ \\mathbf {f}; \\mathbf {a}; \\mathbf {p} ].", + "image_path": "1357d9501af2f0ab73115a41fa92105e7a20dff8a8e71dc455962c294bcbc9c2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 313, + 279, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 313, + 279, + 323 + ], + "spans": [ + { + "bbox": [ + 132, + 313, + 279, + 323 + ], + "type": "text", + "content": "3.3 Reconstruction Module" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 327, + 480, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 480, + 363 + ], + "type": "text", + "content": "To reconstruct the high-resolution image using estimated features in Fourier space, we employ the reconstruction module " + }, + { + "bbox": [ + 130, + 327, + 480, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\varphi}" + }, + { + "bbox": [ + 130, + 327, + 480, + 363 + ], + "type": "text", + "content": "; Blender " + }, + { + "bbox": [ + 130, + 327, + 480, + 363 + ], + "type": "inline_equation", + "content": "(\\mathbf{B}_{\\eta}(\\cdot):\\mathbb{R}^{N\\times 2K}\\mapsto \\mathbb{R}^{1\\times 2K})" + }, + { + "bbox": [ + 130, + 327, + 480, + 363 + ], + "type": "text", + "content": " and MLP decoder " + }, + { + "bbox": [ + 130, + 327, + 480, + 363 + ], + "type": "inline_equation", + "content": "(\\mathbf{G}_{\\Theta}(\\cdot):\\mathbb{R}^{2K}\\mapsto \\mathbb{R}^{12})" + }, + { + "bbox": [ + 130, + 327, + 480, + 363 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 218, + 370, + 480, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 370, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 218, + 370, + 480, + 384 + ], + "type": "interline_equation", + "content": "\\mathbf {h} = \\mathbf {R} _ {\\varphi} \\left(\\left\\{\\mathbf {m} _ {i} \\right\\} _ {i = 1} ^ {N}\\right) = \\mathbf {G} _ {\\Theta} \\left(\\mathbf {B} _ {\\boldsymbol {\\eta}} \\left(\\left\\{\\mathbf {m} _ {i} \\right\\} _ {i = 1} ^ {N}\\right)\\right), \\tag {6}", + "image_path": "6e2e37cbf286dc5863e9bdf28ff4dbc065b5e78f07d8e5f6e519e8a2c57cb442.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 390, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 390, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{h} \\in \\mathbb{R}^{1 \\times 12 \\times sH \\times sW}" + }, + { + "bbox": [ + 130, + 390, + 482, + 437 + ], + "type": "text", + "content": " is the merged Fourier features of burst shots. The blender " + }, + { + "bbox": [ + 130, + 390, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\eta}" + }, + { + "bbox": [ + 130, + 390, + 482, + 437 + ], + "type": "text", + "content": ", which is the same module with the encoder " + }, + { + "bbox": [ + 130, + 390, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\psi}" + }, + { + "bbox": [ + 130, + 390, + 482, + 437 + ], + "type": "text", + "content": ", reduces color differences between frames and merges multiple frames. MLP decoder " + }, + { + "bbox": [ + 130, + 390, + 482, + 437 + ], + "type": "inline_equation", + "content": "(\\mathbf{G}_{\\Theta})" + }, + { + "bbox": [ + 130, + 390, + 482, + 437 + ], + "type": "text", + "content": " estimates 'sRGB' information." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 445, + 328, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 445, + 328, + 457 + ], + "spans": [ + { + "bbox": [ + 131, + 445, + 328, + 457 + ], + "type": "text", + "content": "3.4 Skip connection with upsampling" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 459, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 482, + 506 + ], + "type": "text", + "content": "To prevent learning on the low-frequency only, we utilize the upscaled skip connection " + }, + { + "bbox": [ + 130, + 459, + 482, + 506 + ], + "type": "inline_equation", + "content": "(\\mathbf{Q}_{\\pmb{\\xi}}(\\cdot):\\mathbb{R}^{4\\times H\\times W}\\mapsto \\mathbb{R}^{12\\times sH\\times sW})" + }, + { + "bbox": [ + 130, + 459, + 482, + 506 + ], + "type": "text", + "content": " about the reference frame " + }, + { + "bbox": [ + 130, + 459, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_1^{\\mathbf{L}}" + }, + { + "bbox": [ + 130, + 459, + 482, + 506 + ], + "type": "text", + "content": ". Finally, the spatial size is increased using pixel shuffle " + }, + { + "bbox": [ + 130, + 459, + 482, + 506 + ], + "type": "inline_equation", + "content": "(\\mathbf{P_2}(\\cdot):\\mathbb{R}^{12\\times sH\\times sW}\\mapsto \\mathbb{R}^{3\\times 2sH\\times 2sW})" + }, + { + "bbox": [ + 130, + 459, + 482, + 506 + ], + "type": "text", + "content": ". The proposed BurstM can be formulated as follow:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 256, + 514, + 480, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 514, + 480, + 528 + ], + "spans": [ + { + "bbox": [ + 256, + 514, + 480, + 528 + ], + "type": "interline_equation", + "content": "\\mathbf {I} ^ {\\mathbf {S R}} = \\mathbf {P} _ {2} (\\mathbf {h} + \\mathbf {Q} _ {\\xi} \\left(\\mathbf {I} _ {1} ^ {\\mathbf {L}}\\right)) \\tag {7}", + "image_path": "f5fd35697e0aefa7ca1352ec4e2d99fc4d0fc63ee3826b32fcc8626238b2853f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 533, + 228, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 533, + 228, + 544 + ], + "spans": [ + { + "bbox": [ + 131, + 533, + 228, + 544 + ], + "type": "text", + "content": "3.5 Loss function" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 546, + 481, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 481, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 481, + 583 + ], + "type": "text", + "content": "We use two loss functions for both SyntheticBurst [2] and BurstSR [3]. The MFSR datasets [2,3] does not provide the ground truth of the optical flow map between frames. Thus, we use L2 loss term to use the photometric loss." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 171, + 590, + 480, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 590, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 171, + 590, + 480, + 622 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {O F}} = \\sum_ {i = 2} ^ {N} \\left\\| W a r p \\left(\\mathbf {I} _ {i} ^ {\\mathbf {L}}, \\boldsymbol {\\delta} _ {i} ^ {\\mathbf {L}}\\right) - \\mathbf {I} _ {1} ^ {\\mathbf {L}} \\right\\| _ {2} = \\sum_ {i = 2} ^ {N} \\left(W a r p \\left(\\mathbf {I} _ {i} ^ {\\mathbf {L}}, \\boldsymbol {\\delta} _ {i} ^ {\\mathbf {L}}\\right) - \\mathbf {I} _ {1} ^ {\\mathbf {L}}\\right) ^ {2}, \\tag {8}", + "image_path": "d4815516c9f71bf3761f0e3ded430662b805073ebcd2be0278a532327e3f9e4a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 629, + 480, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 654 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 629, + 480, + 654 + ], + "type": "inline_equation", + "content": "\\text{Warp}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 629, + 480, + 654 + ], + "type": "text", + "content": " is the bilinear warping operator using optical flow from FNet, and " + }, + { + "bbox": [ + 130, + 629, + 480, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_1^{\\mathbf{L}}" + }, + { + "bbox": [ + 130, + 629, + 480, + 654 + ], + "type": "text", + "content": " is the reference frame." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "BurstM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": "SyntheticBurst dataset Input frames are generated from the ground truth. Thus, we update our model using L1 loss between predicted image and the ground truth." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 257, + 151, + 480, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 151, + 480, + 166 + ], + "spans": [ + { + "bbox": [ + 257, + 151, + 480, + 166 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {S R}} = \\left\\| \\mathbf {I} ^ {\\mathbf {S R}} - \\mathbf {I} ^ {\\mathbf {G T}} \\right\\| _ {1}. \\tag {9}", + "image_path": "51abf45c4f63ee6442d43f6daa2fcfa5d40b9e2245bd2e5e02393dc30ad916e6.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 170, + 385, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 170, + 385, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 170, + 385, + 182 + ], + "type": "text", + "content": "Thus, the total loss function during training is as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 251, + 192, + 481, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 192, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 251, + 192, + 481, + 205 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {S y n t h e t i c}} = \\mathcal {L} _ {\\mathcal {S R}} + \\mathcal {L} _ {\\mathcal {O F}}. \\tag {10}", + "image_path": "00449ad3f71a9b400c7c19e5545d582f97eac02f6e547f62fe8095a2758f0eda.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 213, + 479, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 213, + 479, + 250 + ], + "spans": [ + { + "bbox": [ + 131, + 213, + 479, + 250 + ], + "type": "text", + "content": "BurstSR dataset We employs an aligned L1 loss for the BurstSR [3] dataset due to mismatching between low-resolution images and a ground truth image. An aligned L1 loss utilizes pre-trained weight of PwCNet [40] as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 247, + 258, + 480, + 272 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 258, + 480, + 272 + ], + "spans": [ + { + "bbox": [ + 247, + 258, + 480, + 272 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\delta} ^ {\\mathbf {A}} = P W C N e t \\left(\\mathbf {I} ^ {\\mathbf {S R}}, \\mathbf {I} ^ {\\mathbf {G T}}\\right), \\tag {11}", + "image_path": "e2dcf8841d40289b97e151648a564103dcca37e6eaf829dc2f1dddd00ac050b2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 217, + 275, + 480, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 275, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 217, + 275, + 480, + 289 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {A l i g n e d S R}} = \\left\\| W a r p \\left(\\mathbf {I} ^ {\\mathbf {S R}}, \\boldsymbol {\\delta} ^ {\\mathbf {A}}\\right) - \\mathbf {I} ^ {\\mathbf {G T}} \\right\\| _ {1}, \\tag {12}", + "image_path": "64fdbc910550940c765ee6fc95f2200e34bde863cc9e1a77521fe08729957cb5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 300, + 479, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 479, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 479, + 335 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 300, + 479, + 335 + ], + "type": "inline_equation", + "content": "\\delta^{\\mathbf{A}}" + }, + { + "bbox": [ + 130, + 300, + 479, + 335 + ], + "type": "text", + "content": " is optical flow between the predicted image on high-resolution grid and ground truth using PwCNet [40]. Total loss function for BurstSR [3] dataset is as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 247, + 337, + 480, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 337, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 247, + 337, + 480, + 350 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {R e a l}} = \\mathcal {L} _ {\\text {A l i g n e d S R}} + \\mathcal {L} _ {\\mathcal {O F}}. \\tag {13}", + "image_path": "23882c12ad93f255724b73ec6d24d44e8ec11de45d686f5bf3ee3492b2170f4b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 367, + 224, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 367, + 224, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 367, + 224, + 380 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 392, + 240, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 240, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 240, + 403 + ], + "type": "text", + "content": "4.1 Network details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 411, + 479, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 411, + 479, + 448 + ], + "spans": [ + { + "bbox": [ + 131, + 411, + 479, + 448 + ], + "type": "text", + "content": "Optical Flow Estimator (FNet) Our network adapts FNet from FRVSR [36]. To apply FNet into BurstSR task, we modify the stem of FNet from 'sRGB' to 'RGGB' images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 460, + 480, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 460, + 480, + 519 + ], + "spans": [ + { + "bbox": [ + 131, + 460, + 480, + 519 + ], + "type": "text", + "content": "Learnable Neural Warping " + }, + { + "bbox": [ + 131, + 460, + 480, + 519 + ], + "type": "inline_equation", + "content": "(\\mathbf{T}_{\\nu})" + }, + { + "bbox": [ + 131, + 460, + 480, + 519 + ], + "type": "text", + "content": " NW estimates the amplitude, frequency, and phase from burst shots with a single convolution operation to represent the Fourier information. Amplitude and frequency estimators have 128 channels of the convolutions with " + }, + { + "bbox": [ + 131, + 460, + 480, + 519 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 131, + 460, + 480, + 519 + ], + "type": "text", + "content": " kernel, phase estimator has 64 channels of the convolutions with " + }, + { + "bbox": [ + 131, + 460, + 480, + 519 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 131, + 460, + 480, + 519 + ], + "type": "text", + "content": " kernel." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 532, + 480, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 532, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 131, + 532, + 480, + 590 + ], + "type": "text", + "content": "Encoder " + }, + { + "bbox": [ + 131, + 532, + 480, + 590 + ], + "type": "inline_equation", + "content": "(\\mathbf{E}_{\\psi})" + }, + { + "bbox": [ + 131, + 532, + 480, + 590 + ], + "type": "text", + "content": " and Reconstruction module " + }, + { + "bbox": [ + 131, + 532, + 480, + 590 + ], + "type": "inline_equation", + "content": "(\\mathbf{R}_{\\varphi})" + }, + { + "bbox": [ + 131, + 532, + 480, + 590 + ], + "type": "text", + "content": " The reconstruction module consists of Blender " + }, + { + "bbox": [ + 131, + 532, + 480, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\eta}" + }, + { + "bbox": [ + 131, + 532, + 480, + 590 + ], + "type": "text", + "content": " and MLP decoder " + }, + { + "bbox": [ + 131, + 532, + 480, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\Theta}" + }, + { + "bbox": [ + 131, + 532, + 480, + 590 + ], + "type": "text", + "content": ". Both Encoder " + }, + { + "bbox": [ + 131, + 532, + 480, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\psi}" + }, + { + "bbox": [ + 131, + 532, + 480, + 590 + ], + "type": "text", + "content": " and Blender " + }, + { + "bbox": [ + 131, + 532, + 480, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\eta}" + }, + { + "bbox": [ + 131, + 532, + 480, + 590 + ], + "type": "text", + "content": " are EDSR [25] without upsampling module and 128 channels. MLP Decoder, " + }, + { + "bbox": [ + 131, + 532, + 480, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\Theta}" + }, + { + "bbox": [ + 131, + 532, + 480, + 590 + ], + "type": "text", + "content": ", is organized by 4-layer of " + }, + { + "bbox": [ + 131, + 532, + 480, + 590 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 131, + 532, + 480, + 590 + ], + "type": "text", + "content": " convolution with 256 channels and ReLU activation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 609, + 277, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 277, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 277, + 621 + ], + "type": "text", + "content": "4.2 Implementation details" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "content": "BurstM is an end-to-end trainable network and is possible to train without pre-trained weights. We apply an AdamW optimizer [28] with the Cosine annealing strategy [27] from " + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "content": " learning rate. For SyntheticBurst [2], we" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "EG Kang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "train BurstM with L1 loss for the output image and L2 loss for optical flows. For BurstSR [3], We fine-tune the pre-trained BurstM of SyntheticBurst with L2 loss for optical flow and aligned L1 loss [3] for the output image. Aligned L1 loss utilizes PWCNet [40] for the output image. The reason for the usage of the aligned L1 Loss is that there are no LR images aligned with the GT image because they were captured using different cameras. We train our model using four RTX3090 GPUs and training times are 10 days for SyntheticBurst and 8 hours for BurstSR. The summary of the training strategy is in Table 1." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 203, + 213, + 411, + 350 + ], + "blocks": [ + { + "bbox": [ + 203, + 213, + 411, + 350 + ], + "lines": [ + { + "bbox": [ + 203, + 213, + 411, + 350 + ], + "spans": [ + { + "bbox": [ + 203, + 213, + 411, + 350 + ], + "type": "table", + "html": "
SyntheticBurst [2]BurstSR [3]
Training Epoch30025
OptimizerADAMW [28]
Learning rate10-4~10-6
LR schedulerCosine annealing [27]
Batch sizes4
Main LossL1Aligned L1
Photo metric LossL2
×2:96×96
Input sizes×3:64×64×4:80×80
×4:48×48
Training time10 days8 hours
", + "image_path": "3f793a559396feb8bc60202b5b3117387bfa06d8b30914c71861aa06cdf28624.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 350, + 482, + 383 + ], + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 383 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 383 + ], + "type": "text", + "content": "Table 1: Training Strategies. Training epoch and main loss, input sizes are different. Except for these factors, we use same training strategies for both SyntheticBurst [2] for BurstSR [3] datasets." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 395, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 395, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 395, + 482, + 491 + ], + "type": "text", + "content": "SyntheticBurst Dataset [2] The dataset contains burst frames which are 46,839 set for training and 300 set for validation. Each burst set contains 14 low-resolution burst frames, which are generated from 'sRGB' as follows. We utilized the unprocessing [5]. Subsequently, we apply random rotations and translations. We employ random down-sampling " + }, + { + "bbox": [ + 130, + 395, + 482, + 491 + ], + "type": "inline_equation", + "content": "(\\times 2, \\times 3, \\times 4)" + }, + { + "bbox": [ + 130, + 395, + 482, + 491 + ], + "type": "text", + "content": " to generate low-resolution images of multiple scales. Finally, we apply Bayer mosaicking and add random noise. We trained the proposed model with batch size 4 during 300 epochs, and the training patch size is changed per scale factors in Table 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 502, + 496, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 496, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 496, + 598 + ], + "type": "text", + "content": "BurstSR Dataset [3] Input and GT are captured by different cameras: Smartphone and DSLR, respectively. Both low and high resolution images have 200 raw burst sets. The low-resolution set contains 14 frames, and we extract 5,405 patches for training and 882 patches for validation. Unlike SyntehticBurst dataset [2], the scale factor between low-resolution and high-resolution is fixed. Thus, we train BurstM only for " + }, + { + "bbox": [ + 130, + 502, + 496, + 598 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 130, + 502, + 496, + 598 + ], + "type": "text", + "content": " scale factor. Nevertheless, BurstM can accurately predict high-resolution images for multiple scale factors (" + }, + { + "bbox": [ + 130, + 502, + 496, + 598 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 130, + 502, + 496, + 598 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 502, + 496, + 598 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 130, + 502, + 496, + 598 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 502, + 496, + 598 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 130, + 502, + 496, + 598 + ], + "type": "text", + "content": "), including clear representation of high-frequency textures without the boundary artifacts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 613, + 254, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 613, + 254, + 624 + ], + "spans": [ + { + "bbox": [ + 131, + 613, + 254, + 624 + ], + "type": "text", + "content": "4.3 Validation Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "We show the qualitative comparison with the existing methods [3, 4, 11, 12, 29, 30] and the visual comparison with latest three methods [11, 12, 29] using the pre-trained weights from authors." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "BurstM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 114, + 475, + 252 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 475, + 252 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 475, + 252 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 475, + 252 + ], + "type": "table", + "html": "
# Params. MSyntheticBurstBurstSR x4 PSNR SSIM
x2 PSNRSSIMx3 PSNRSSIMx4 PSNR
Bicubic-38.300.94833.940.88633.020.862
DBSR [3]13.0140.510.96540.110.95940.760.959
MFIR [4]12.1341.250.97141.810.97241.560.964
BIPNet†[11]6.737.580.92840.830.95541.930.960
GMTNet†[30]-----42.360.961
Burststormer†[12]2.537.060.92540.260.95342.830.973
BSRT-Small†[29]4.9240.640.96642.300.97542.720.971
BSRT-Large†[29]20.7140.330.96542.870.97943.620.975
BurstM (Ours)14.046.010.98544.790.98242.870.973
", + "image_path": "7d5af780815ec8c19bb7c61b3cbcd53de6ead3506849820d62afa153119b50fd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 253, + 480, + 296 + ], + "lines": [ + { + "bbox": [ + 132, + 253, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 253, + 480, + 296 + ], + "type": "text", + "content": "Table 2: Quantitative comparison on SyntheticBurst and BurstSR. Quantitative comparisons of BurstSR such as PSNR and SSIM can only be measured for " + }, + { + "bbox": [ + 132, + 253, + 480, + 296 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 132, + 253, + 480, + 296 + ], + "type": "text", + "content": ", due to LR and HR being provided in fixed sizes. Red and Blue colors indicate the best and the second performance, respectively. " + }, + { + "bbox": [ + 132, + 253, + 480, + 296 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 132, + 253, + 480, + 296 + ], + "type": "text", + "content": " indicates Transformer-based method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 136, + 301, + 477, + 637 + ], + "blocks": [ + { + "bbox": [ + 136, + 301, + 477, + 637 + ], + "lines": [ + { + "bbox": [ + 136, + 301, + 477, + 637 + ], + "spans": [ + { + "bbox": [ + 136, + 301, + 477, + 637 + ], + "type": "image", + "image_path": "574d07622ca736affaae30f009aa0ec1b7e009554ab9ca5474254a254d5ddbf6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 640, + 481, + 662 + ], + "lines": [ + { + "bbox": [ + 132, + 640, + 481, + 662 + ], + "spans": [ + { + "bbox": [ + 132, + 640, + 481, + 662 + ], + "type": "text", + "content": "Fig. 3: Visual comparison for " + }, + { + "bbox": [ + 132, + 640, + 481, + 662 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 132, + 640, + 481, + 662 + ], + "type": "text", + "content": " super-resolution for BurstSR. BurstM shows clearer textures and correct color than the others without the boundary artifacts." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "EG Kang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "text", + "content": "SR on BurstSR [3] Table 2 is the quantitative comparison result of " + }, + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "text", + "content": " SR on BurstSR [3]. Ours shows a performance gain of " + }, + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "inline_equation", + "content": "+0.17\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "text", + "content": " compared to the state-of-the-art baselines. As shown in Figure 3, BurstM predicts outperform high-frequency details without the boundary artifacts from low-resolution images of real-world dataset. Also, BurstM predicts multi-scale SR without an additional processing, different with existing models. However, the quantitative comparison of " + }, + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "text", + "content": " is not possible, because there is no GT of both " + }, + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "text", + "content": " SR of given dataset [3]. We demonstrate the robustness of our approach by representing high-frequency details in Figure 4 and computation time in Table 4. BurstM directly predicts multiple scales, while other methods implement " + }, + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 130, + 116, + 482, + 258 + ], + "type": "text", + "content": " SR followed by bicubic downsampling. In Table 4, BurstM shows shorter computation time than the existing methods." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 140, + 259, + 475, + 411 + ], + "blocks": [ + { + "bbox": [ + 140, + 259, + 475, + 411 + ], + "lines": [ + { + "bbox": [ + 140, + 259, + 475, + 411 + ], + "spans": [ + { + "bbox": [ + 140, + 259, + 475, + 411 + ], + "type": "image", + "image_path": "f42d53edad15dd700baa4f107b1b5759f6d8e8ab8fc757ca4eee07ce05b0254c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 413, + 482, + 436 + ], + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 436 + ], + "type": "text", + "content": "Fig. 4: Visual comparison of multi-scale SR on BurstSR. BurstM predict several scaled-images with unimodel and the others are downsampling after " + }, + { + "bbox": [ + 130, + 413, + 482, + 436 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 130, + 413, + 482, + 436 + ], + "type": "text", + "content": " SR." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 437, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 594 + ], + "type": "text", + "content": "SR on SyntheticBurst [2] Table 2 is the quantitative comparison results of multiple scale factors " + }, + { + "bbox": [ + 130, + 437, + 482, + 594 + ], + "type": "inline_equation", + "content": "(\\times 2, \\times 3, \\times 4)" + }, + { + "bbox": [ + 130, + 437, + 482, + 594 + ], + "type": "text", + "content": " SR and Figure 5 is the visual comparison result of " + }, + { + "bbox": [ + 130, + 437, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 130, + 437, + 482, + 594 + ], + "type": "text", + "content": " SR on SyntheticBurst [2]. Our model achieves second result with reasonable the number of parameters. In addition, we compare the performance on multiple scale factors in Table 2 and Figure 6. BurstM predicts high-resolution images on multiple scale factors with an unimodel and the others implement multiple SR with additional bicubic downsampling after " + }, + { + "bbox": [ + 130, + 437, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 130, + 437, + 482, + 594 + ], + "type": "text", + "content": " SR. In case of BurstM, it is observed that PSNR gradually increases as the scale factor decrease. However, PSNR of existing methods decreases as the scale factor decreases because of the information loss during the additional bicubic downsampling process. Figure 6 shows the residual images, which means the loss of the high-frequency textures. Furthermore, as shown in Table 4, BurstM outperforms the other methods in terms of computation time." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Fourier feature space Figure 7 demonstrates the estimation of high-frequency details on Fourier space. The result from each frame do not illustrate the DFT spectrum, but accumulated result of all frames shows a similar result with the DFT spectrum. Because BurstM leverages sub-pixel information from burst frames to represent the high-resolution images." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "BurstM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 479, + 289 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 479, + 289 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 479, + 289 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 479, + 289 + ], + "type": "image", + "image_path": "a43c4c5dab9a0ffbb77b28c856b37399c91cb0fc4343b6c985c9fcccfc683345.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 292, + 467, + 304 + ], + "lines": [ + { + "bbox": [ + 145, + 292, + 467, + 304 + ], + "spans": [ + { + "bbox": [ + 145, + 292, + 467, + 304 + ], + "type": "text", + "content": "Fig. 5: Visual comparison for " + }, + { + "bbox": [ + 145, + 292, + 467, + 304 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 145, + 292, + 467, + 304 + ], + "type": "text", + "content": " Super-Resolution for SynteticBurst." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 130, + 310, + 482, + 524 + ], + "blocks": [ + { + "bbox": [ + 130, + 310, + 482, + 524 + ], + "lines": [ + { + "bbox": [ + 130, + 310, + 482, + 524 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 482, + 524 + ], + "type": "image", + "image_path": "53409e3e489013298822216c8c9c7c94da78f30da1d6b313242a8ebbd441d37d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 528, + 482, + 583 + ], + "lines": [ + { + "bbox": [ + 130, + 528, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 482, + 583 + ], + "type": "text", + "content": "Fig.6: Visual comparison for multi-scale SR on SyntheticBurst. The colorbar indicates a difference with GT. The existing methods lose high-frequency textures during additional down-sampling because the highlighted locations are a high-frequency area. But BurstM represent the similar high-frequency textures with GT by because of direct prediction on multiple scale factors." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Aligning performances We conduct a comparative analysis of the alignment performance between DCN-based method and BurstM in Figure 8. As in [11, 12, 29, 30], we utilize multi-layer DCN with varied spatial sizes for DCN-based method. In contrast, BurstM implements a single image warping process using predicted optical flows. The 'Before align.' and 'After align.' residual maps in Figure 8 represent the differences with the reference frame at each position in" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "EG Kang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 147, + 113, + 465, + 292 + ], + "blocks": [ + { + "bbox": [ + 147, + 113, + 465, + 292 + ], + "lines": [ + { + "bbox": [ + 147, + 113, + 465, + 292 + ], + "spans": [ + { + "bbox": [ + 147, + 113, + 465, + 292 + ], + "type": "image", + "image_path": "1545ca80d0bf01163304d68d8226c7125c00d978fc38aef0d0ba0a967417df3c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 293, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 316 + ], + "type": "text", + "content": "Fig.7: Fourier space analysis. Accumulated frequencies and coefficients from BurstM are analogous to DFT spectrum of GT (Red color in 'DFT' and 'ACCUM')." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 319, + 482, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 427 + ], + "type": "text", + "content": "flowchart. The red dots in 'DCN method flowchart' illustrate offsets to the center pixel. DCN uses an unrelated pixel information, because DCN requires to use the number of predefined kernels regardless of the correlation. Therefore, the aligned results are inaccurate and show high values in the residual map. Additionally, boundary artifacts are also observed due to the utilization of irrelevant information. The color map and white arrows in 'BurstM method flowchart' are the visualization results of optical flow. BurstM method shows one-to-one correspondences between pixels. Thus, the alignment results demonstrate clearer and sharper compared to the DCN-based method without the boundary artifacts." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 157, + 430, + 302, + 486 + ], + "blocks": [ + { + "bbox": [ + 157, + 430, + 302, + 486 + ], + "lines": [ + { + "bbox": [ + 157, + 430, + 302, + 486 + ], + "spans": [ + { + "bbox": [ + 157, + 430, + 302, + 486 + ], + "type": "image", + "image_path": "3754588c6c75114624fc2e0033f7944a0bef59938eb6d9bae41a23ba4be0a3b6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 488, + 458, + 499 + ], + "lines": [ + { + "bbox": [ + 163, + 488, + 458, + 499 + ], + "spans": [ + { + "bbox": [ + 163, + 488, + 458, + 499 + ], + "type": "text", + "content": "DCN method flowchart [12] BurstM method (ours) flowchart" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 430, + 450, + 486 + ], + "blocks": [ + { + "bbox": [ + 309, + 430, + 450, + 486 + ], + "lines": [ + { + "bbox": [ + 309, + 430, + 450, + 486 + ], + "spans": [ + { + "bbox": [ + 309, + 430, + 450, + 486 + ], + "type": "image", + "image_path": "3587067fb3f61041baadd4da369409c8f6af3838944346d9d1d68a842f92bbf4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 133, + 501, + 479, + 609 + ], + "blocks": [ + { + "bbox": [ + 133, + 501, + 479, + 609 + ], + "lines": [ + { + "bbox": [ + 133, + 501, + 479, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 501, + 479, + 609 + ], + "type": "image", + "image_path": "17c08bbf5c1850793e2a9df5dcce235b3a53b72476b06336a44a4e1c8b8522dc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 611, + 482, + 677 + ], + "lines": [ + { + "bbox": [ + 130, + 611, + 482, + 677 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 482, + 677 + ], + "type": "text", + "content": "Fig. 8: Feature alignment results. Flow charts indicate where we take features. The residual map is between the reference and source frames at each location in flowchart. DCN based-method has a big difference after an alignment because DCN take unrelated information during alignment. However, BurstM point out one-to-one correspondences between a reference and source frames. BurstM shows accurate align result without the boundary artifacts." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "BurstM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 243, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 243, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 243, + 129 + ], + "type": "text", + "content": "5 Ablation study" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 131, + 482, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 131, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 131, + 482, + 216 + ], + "type": "text", + "content": "The comparison between with and without Fourier mapping INR [17] is the most representative method for representing continuous signals, but it has a spectral bias [32, 34, 39]. As shown in [21, 23], BurstM also addresses this issue through Fourier mapping. Table 3a compares the performance of BurstM with and without Fourier mapping. BurstM without Fourier mapping has significantly lower PSNR values. This is because the network without Fourier mapping learns more about low frequencies due to spectral bias during training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 227, + 482, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 227, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 482, + 312 + ], + "type": "text", + "content": "Relationship between flexibility and performance BurstM is supporting the multiple scale SR " + }, + { + "bbox": [ + 130, + 227, + 482, + 312 + ], + "type": "inline_equation", + "content": "(\\times 2, \\times 3, \\times 4)" + }, + { + "bbox": [ + 130, + 227, + 482, + 312 + ], + "type": "text", + "content": ". The performance in Table 2 has trained with three multiple scale factors " + }, + { + "bbox": [ + 130, + 227, + 482, + 312 + ], + "type": "inline_equation", + "content": "(\\times 2, \\times 3, \\times 4)" + }, + { + "bbox": [ + 130, + 227, + 482, + 312 + ], + "type": "text", + "content": ". We train with different number of scale factors such as one " + }, + { + "bbox": [ + 130, + 227, + 482, + 312 + ], + "type": "inline_equation", + "content": "(\\times 4)" + }, + { + "bbox": [ + 130, + 227, + 482, + 312 + ], + "type": "text", + "content": " and two " + }, + { + "bbox": [ + 130, + 227, + 482, + 312 + ], + "type": "inline_equation", + "content": "(\\times 3, \\times 4)" + }, + { + "bbox": [ + 130, + 227, + 482, + 312 + ], + "type": "text", + "content": " items. In such cases, we observed that PSNR increases on in-of-scale, but performance decreases on out-of-scale. When there are many instances of out-of-scale, the model's flexibility diminishes, so we proposed three scales satisfy both performance and flexibility." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 323, + 482, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 382 + ], + "type": "text", + "content": "Different optical flow estimator While FNet [36] estimates an optical flow based on given image sizes, the other optical flow estimator [35] adjust image sizes according to its setting before input. This resizing of images introduces aliasing effects, resulting in a decline in SR performance. Additionally, BurstM with Spynet [35] shows longer inference time than the proposed method." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 152, + 382, + 461, + 406 + ], + "blocks": [ + { + "bbox": [ + 152, + 382, + 461, + 406 + ], + "lines": [ + { + "bbox": [ + 152, + 382, + 461, + 406 + ], + "spans": [ + { + "bbox": [ + 152, + 382, + 461, + 406 + ], + "type": "table", + "html": "
BurstM w\\o Fourier mappingBurstM w\\Fourier mapping
×4 PSNR39.82 dB41.74 dB
", + "image_path": "5bc604ef10517e9566840676a18f955282f4bd67a5a5060012c6c1843b21f5b7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 144, + 421, + 296, + 470 + ], + "blocks": [ + { + "bbox": [ + 171, + 407, + 443, + 417 + ], + "lines": [ + { + "bbox": [ + 171, + 407, + 443, + 417 + ], + "spans": [ + { + "bbox": [ + 171, + 407, + 443, + 417 + ], + "type": "text", + "content": "(a) The performance comparison between with and without Fourier mapping" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 144, + 421, + 296, + 470 + ], + "lines": [ + { + "bbox": [ + 144, + 421, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 144, + 421, + 296, + 470 + ], + "type": "table", + "html": "
Training scale×2×3×4
×433.3635.9742.26
×3, ×430.8043.6141.82
×2, ×3, ×4 (Ours)44.7743.4541.74
", + "image_path": "23191e0921df20aaca77a54952d99a3c61ec6ee4e7958065074b1d1b714406f1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 301, + 421, + 479, + 469 + ], + "blocks": [ + { + "bbox": [ + 149, + 471, + 291, + 489 + ], + "lines": [ + { + "bbox": [ + 149, + 471, + 291, + 489 + ], + "spans": [ + { + "bbox": [ + 149, + 471, + 291, + 489 + ], + "type": "text", + "content": "(b) Relationship between flexibility and performance (PSNR:dB)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 301, + 421, + 479, + 469 + ], + "lines": [ + { + "bbox": [ + 301, + 421, + 479, + 469 + ], + "spans": [ + { + "bbox": [ + 301, + 421, + 479, + 469 + ], + "type": "table", + "html": "
BurstM\nw\\SpynetBurstM\nw\\FNet (Ours)
×4 PSNR40.85 dB41.74 dB
Inference time14.3 ms11.6 ms
", + "image_path": "3dda7a082221a3b41ed67ff9e7396e756314bc8697654c72bd3d8e5948a1552c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 333, + 471, + 444, + 488 + ], + "lines": [ + { + "bbox": [ + 333, + 471, + 444, + 488 + ], + "spans": [ + { + "bbox": [ + 333, + 471, + 444, + 488 + ], + "type": "text", + "content": "(c) Performance comparison based on optical flow estimator" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 493, + 291, + 504 + ], + "lines": [ + { + "bbox": [ + 132, + 493, + 291, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 291, + 504 + ], + "type": "text", + "content": "Table 3: Ablation study summary" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 504, + 481, + 568 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 132, + 504, + 373, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 373, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 373, + 514 + ], + "type": "text", + "content": "(a) Fourier mapping contributes to the model performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 514, + 481, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 481, + 548 + ], + "type": "text", + "content": "(b) Red color indicates the result of in-scale. Less training scale has higher PSNR on in-scale but network flexibility is low due to increasing out-scale. Thus, we select our model with 3 scale factors " + }, + { + "bbox": [ + 132, + 514, + 481, + 548 + ], + "type": "inline_equation", + "content": "(\\times 2, \\times 3, \\times 4)" + }, + { + "bbox": [ + 132, + 514, + 481, + 548 + ], + "type": "text", + "content": " considering both flexibility and performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 548, + 481, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 481, + 568 + ], + "type": "text", + "content": "(c) A different optical flow estimator leads to PSNR drop and an increase in inference time." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 576, + 217, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 217, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 217, + 588 + ], + "type": "text", + "content": "6 Discussion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": "Boundary line artifacts As shown in Figure 9, the recent method [29] exhibits boundary artifacts despite utilizing optical flow, due to its continued reliance on DCN [9] for alignment. As illustrated in Figure 8 and 9, DCN [9] must utilize information corresponding to the predefined kernel size regardless of correlation, resulting in the incorporation of unnecessary information when most frames fail to provide essential data. In contrast, BurstM aligns frames without DCN [9]," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 101 + ], + "type": "text", + "content": "EG Kang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": "solely utilizing optical flow. Consequently, BurstM avoids the use of irrelevant information, leading to accurate predictions without boundary artifacts." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 138, + 144, + 467, + 305 + ], + "blocks": [ + { + "bbox": [ + 138, + 144, + 467, + 305 + ], + "lines": [ + { + "bbox": [ + 138, + 144, + 467, + 305 + ], + "spans": [ + { + "bbox": [ + 138, + 144, + 467, + 305 + ], + "type": "image", + "image_path": "ee27cf3636ffa531cec16670c108357ef31b6b4202b602606ccbb9d97ef6c4a7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 308, + 482, + 343 + ], + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 343 + ], + "type": "text", + "content": "Fig.9: Boundary artifacts The most of frames didn't provide the information of edge. The existing method [29] suffers the boundary line artifacts, but BurstM represents clear images without the boundary line artifacts." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "content": "Computation complexity Despite advancements in computational system, SR field still requires shorter computation times as the image resolutions increase. Table 4 compares the computation time usage of BurstM with the other methods [11,12,29]. We implements " + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "content": " SR with " + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "inline_equation", + "content": "80 \\times 80" + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "content": " frame sizes of BurstSR dataset [3] on NVIDIA RTX3090 24GB. While BurstM has a large number of parameters, BurstM demonstrates the shortest computation time." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 133, + 418, + 330, + 525 + ], + "blocks": [ + { + "bbox": [ + 133, + 418, + 330, + 525 + ], + "lines": [ + { + "bbox": [ + 133, + 418, + 330, + 525 + ], + "spans": [ + { + "bbox": [ + 133, + 418, + 330, + 525 + ], + "type": "image", + "image_path": "bbea5824731e2b6578ed494a066e72cc7506d6a72027f753db81fd457d0b8e56.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 525, + 482, + 559 + ], + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 559 + ], + "type": "text", + "content": "Table 4: Computational complexity on one GPU (RTX3090). Despite a large number of parameters, BurstM shows the inference time (ms) and best PSNR (dB) performance. Red : best metrics. † indicates Transformer-based method." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 331, + 418, + 485, + 522 + ], + "blocks": [ + { + "bbox": [ + 331, + 418, + 485, + 522 + ], + "lines": [ + { + "bbox": [ + 331, + 418, + 485, + 522 + ], + "spans": [ + { + "bbox": [ + 331, + 418, + 485, + 522 + ], + "type": "table", + "html": "
SR scale: ×4Params.(M) / Mem.(GB)Inference time(ms)PSNR (dB)
BIPNet† [11]6.7 / 11.820.848.49
Burstormer† [12]2.5 / 12.020.848.82
BSRT-S† [29]4.92 / 1.0148.048.48
BSRT-L† [29]20.71 / 1.5266.148.57
BurstM (Ours)14.9 / 8.411.649.12
", + "image_path": "1d9a5ec120abbd17640dd932ecf6a8851a0eae27a1facf8961c8143513eca434.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Super-Resolution with floating scale factors BurstM can estimate the multi-scale SR with floating scale factors such as " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\times 3.5" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " without any architecture changes. However, the training of the floating scale factors (" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\times 2\\sim \\times 4" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ") shows lower PSNR (39.44dB) than the training of the integer scale factors (41.74dB). The performance drop is caused by the demosaic process when generating low-resolution frames. To convert from 'sRGB' to 'RGGB', the dimensions (H, W) of the images are changed to (H/2, W/2). If the image size of 'sRGB' is not divisible by two when a floating scale factor is applied, the network apply an additional" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "BurstM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "downsampling operation to make the dimension of " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "(\\mathrm{H} / 2, \\mathrm{W} / 2)" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": ". The double downsampling imposes a severe degradation of the generated low-resolution images, and it results in low PSNRs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 156, + 482, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 156, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 156, + 482, + 228 + ], + "type": "text", + "content": "**Corner case** The major challenge on optical flow is occlusions which are caused by a new object that is not present in the reference frame. BurstM also has the occlusion issue (Figure 10) even though BurstM demonstrate accurate alignment performance than DCN based methods. Fortunately, BurstM utilize inaccurate optical flow of an occlusion as meaningful prior and reconstruct the high-quality images in a high-resolution grid." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 149, + 228, + 465, + 413 + ], + "blocks": [ + { + "bbox": [ + 149, + 228, + 465, + 413 + ], + "lines": [ + { + "bbox": [ + 149, + 228, + 465, + 413 + ], + "spans": [ + { + "bbox": [ + 149, + 228, + 465, + 413 + ], + "type": "image", + "image_path": "04b3ada32e3a25a2b1902f784b7241a3a5860dc8209a631f673bb473c0f85768.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 415, + 482, + 437 + ], + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 437 + ], + "type": "text", + "content": "Fig. 10: Corner case. Inaccurate optical flow prediction due to occlusions which are caused by new objects (red arrow) not present in the reference image appear." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 444, + 220, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 444, + 220, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 444, + 220, + 456 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 605 + ], + "type": "text", + "content": "In this paper, we proposed Deep Burst Multi-scale SR using Fourier Space with Optical Flow (BurstM) to overcome the limitations of existing methods: SR of fixed scale factor " + }, + { + "bbox": [ + 130, + 462, + 482, + 605 + ], + "type": "inline_equation", + "content": "(\\times 4)" + }, + { + "bbox": [ + 130, + 462, + 482, + 605 + ], + "type": "text", + "content": " and low-quality of real-world dataset [3]. Our network performs the precise alignment using optical flow and the advanced warping process based on estimated Fourier information. We demonstrated their effectiveness in representing the high-frequency textures with " + }, + { + "bbox": [ + 130, + 462, + 482, + 605 + ], + "type": "inline_equation", + "content": "+0.17\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 462, + 482, + 605 + ], + "type": "text", + "content": " PSNR gain in the real-world dataset [3] compared to the previous state-of-the-art model. Furthermore, BurstM successfully implements multi-scale SR " + }, + { + "bbox": [ + 130, + 462, + 482, + 605 + ], + "type": "inline_equation", + "content": "(\\times 2,\\times 3,\\times 4)" + }, + { + "bbox": [ + 130, + 462, + 482, + 605 + ], + "type": "text", + "content": " of the unimodel by representing high-frequency details using the INR method. The function of multi-scale SR enhances our model's capacity compared to the other learning-based methods. In addition, we demonstrate our approach is faster inference time compared to the existing methods by the efficient computation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": "Acknowledgement This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (RS-2021-II212068, Artificial Intelligence Innovation Hub) and the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (RS-2024-00335741)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "EG Kang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 137, + 481, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 481, + 180 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 481, + 180 + ], + "type": "text", + "content": "1. Bascle, B., Blake, A., Zisserman, A.: Motion deblurring and super-resolution from an image sequence. In: Computer Vision—ECCV'96: 4th European Conference on Computer Vision Cambridge, UK, April 15–18, 1996 Proceedings Volume II 4. pp. 571–582. Springer (1996)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 181, + 481, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 181, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 138, + 181, + 481, + 213 + ], + "type": "text", + "content": "2. Bhat, G., Danelljan, M., Timofte, R.: Ntire 2021 challenge on burst superresolution: Methods and results. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 613-626 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 213, + 481, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 213, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 213, + 481, + 245 + ], + "type": "text", + "content": "3. Bhat, G., Danelljan, M., Van Gool, L., Timofte, R.: Deep burst super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9209-9218 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 481, + 277 + ], + "type": "text", + "content": "4. Bhat, G., Danelljan, M., Yu, F., Van Gool, L., Timofte, R.: Deep reparametrization of multi-frame super-resolution and denoising. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2460-2470 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 277, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 481, + 310 + ], + "type": "text", + "content": "5. Brooks, T., Mildenhall, B., Xue, T., Chen, J., Sharlet, D., Barron, J.T.: Unprocessing images for learned raw denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11036-11045 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 310, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 310, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 310, + 481, + 352 + ], + "type": "text", + "content": "6. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basicvsr: The search for essential components in video super-resolution and beyond. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4947-4956 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 353, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 481, + 396 + ], + "type": "text", + "content": "7. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-Trained Image Processing Transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12299-12310 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 396, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 481, + 429 + ], + "type": "text", + "content": "8. Chen, Y., Liu, S., Wang, X.: Learning continuous image representation with local implicit image function. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8628-8638 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 429, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 481, + 460 + ], + "type": "text", + "content": "9. Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., Wei, Y.: Deformable convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 764-773 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 461, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 461, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 138, + 461, + 481, + 493 + ], + "type": "text", + "content": "0. Dai, T., Cai, J., Zhang, Y., Xia, S.T., Zhang, L.: Second-order attention network for single image super-resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11065-11074 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 493, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 493, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 493, + 481, + 525 + ], + "type": "text", + "content": "1. Dudhane, A., Zamir, S.W., Khan, S., Khan, F.S., Yang, M.H.: Burst image restoration and enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5759-5768 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 481, + 557 + ], + "type": "text", + "content": "2. Dudhane, A., Zamir, S.W., Khan, S., Khan, F.S., Yang, M.H.: Burstomer: Burst image restoration and enhancement transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5703-5712 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 558, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 558, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 138, + 558, + 481, + 590 + ], + "type": "text", + "content": "3. Elad, M., Feuer, A.: Restoration of a single superresolution image from several blurred, noisy, and undersampled measured images. IEEE transactions on image processing 6(12), 1646-1658 (1997)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 590, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 481, + 632 + ], + "type": "text", + "content": "4. Hardie, R.C., Barnard, K.J., Bognar, J.G., Armstrong, E.E., Watson, E.A.: High-resolution image reconstruction from a sequence of rotated and translated frames and its application to an infrared imaging system. Optical Engineering 37(1), 247-260 (1998)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "text", + "content": "5. Hirschmuller, H.: Stereo processing by semiglobal matching and mutual information. IEEE Transactions on pattern analysis and machine intelligence 30(2), 328-341 (2007)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "BurstM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "16. Horn, B.K., Schunck, B.G.: Determining optical flow. Artificial intelligence 17(1-3), 185-203 (1981)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "text", + "content": "17. Hornik, K., Stinchcombe, M., White, H.: Multilayer feedforward networks are universal approximators. Neural networks 2(5), 359-366 (1989)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "18. Irani, M., Peleg, S.: Improving resolution by image registration. CVGIP: Graphical models and image processing 53(3), 231-239 (1991)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 134, + 183, + 480, + 215 + ], + "type": "text", + "content": "19. Janai, J., Güney, F., Behl, A., Geiger, A., et al.: Computer vision for autonomous vehicles: Problems, datasets and state of the art. Foundations and Trends in Computer Graphics and Vision 12(1-3), 1-308 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 480, + 248 + ], + "type": "text", + "content": "20. Jiang, C.M., Sud, A., Makadia, A., Huang, J., Niessner, M., Funkhouser, T.: Local Implicit Grid Representations for 3D Scenes. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "21. Kim, M., Lee, J., Lee, B., Im, S., Jin, K.H.: Implicit neural image stitching with enhanced and blended feature reconstruction (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "text", + "content": "22. Lai, W.S., Huang, J.B., Ahuja, N., Yang, M.H.: Deep laplacian pyramid networks for fast and accurate super-resolution. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 624-632 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "23. Lee, J., Jin, K.H.: Local texture estimator for implicit representation function. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1929-1938 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "24. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: SwinIR: Image Restoration Using Swin Transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 1833-1844 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "25. Lim, B., Son, S., Kim, H., Nah, S., Mu Lee, K.: Enhanced deep residual networks for single image super-resolution. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 136-144 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 403, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 403, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 133, + 403, + 480, + 445 + ], + "type": "text", + "content": "26. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10012-10022 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "text", + "content": "27. Loshchilov, I., Hutter, F.: Stochastic gradient descent with warm restarts. In: Proceedings of the 5th Int. Conf. Learning Representations. pp. 1-16" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 489 + ], + "type": "text", + "content": "28. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 490, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 490, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 133, + 490, + 480, + 533 + ], + "type": "text", + "content": "29. Luo, Z., Li, Y., Cheng, S., Yu, L., Wu, Q., Wen, Z., Fan, H., Sun, J., Liu, S.: BSRT: Improving burst super-resolution with swim transformer and flow-guided deformable alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 998-1008 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 534, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 534, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 133, + 534, + 480, + 577 + ], + "type": "text", + "content": "30. Mehta, N., Dudhane, A., Murala, S., Zamir, S.W., Khan, S., Khan, F.S.: Gated multi-resolution transfer network for burst restoration and enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22201-22210 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 133, + 578, + 480, + 610 + ], + "type": "text", + "content": "31. Mei, Y., Fan, Y., Zhou, Y.: Image super-resolution with non-local sparse attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3517-3526 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 611, + 480, + 643 + ], + "type": "text", + "content": "32. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "33. Peleg, S., Keren, D., Schweitzer, L.: Improving image resolution using subpixel motion. Pattern recognition letters 5(3), 223-226 (1987)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 102 + ], + "type": "text", + "content": "EG Kang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 578 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 172 + ], + "type": "text", + "content": "34. Rahaman, N., Baratin, A., Arpit, D., Draxler, F., Lin, M., Hamprecht, F., Bengio, Y., Courville, A.: On the Spectral Bias of Neural Networks. In: Chaudhuri, K., Salakhutdinov, R. (eds.) Proceedings of the 36th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 97, pp. 5301-5310. PMLR (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "35. Ranjan, A., Black, M.J.: Optical flow estimation using a spatial pyramid network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4161-4170 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "36. Sajjadi, M.S., Vemulapalli, R., Brown, M.: Frame-recurrent video super-resolution. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6626-6634 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 237, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 282 + ], + "type": "text", + "content": "37. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1874-1883 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "text", + "content": "38. Simonyan, K., Zisserman, A.: Two-stream convolutional networks for action recognition in videos. Advances in neural information processing systems 27 (2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 347 + ], + "type": "text", + "content": "39. Sitzmann, V., Martel, J., Bergman, A., Lindell, D., Wetzstein, G.: Implicit Neural Representations with Periodic Activation Functions. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M.F., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 7462-7473. Curran Associates, Inc. (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "40. Sun, D., Yang, X., Liu, M.Y., Kautz, J.: Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8934-8943 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "41. Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 402-419. Springer (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 413, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 457 + ], + "type": "text", + "content": "42. Triggs, B., McLauchlan, P.F., Hartley, R.I., Fitzgibbon, A.W.: Bundle adjustment—a modern synthesis. In: Vision Algorithms: Theory and Practice: International Workshop on Vision Algorithms Corfu, Greece, September 21–22, 1999 Proceedings. pp. 298–372. Springer (2000)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "text", + "content": "43. Tsai, R.Y., Huang, T.S.: Multiframe image restoration and registration. Multiframe image restoration and registration 1, 317-339 (1984)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "type": "text", + "content": "44. Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image super-resolution using very deep residual channel attention networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 286-301 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "45. Zhang, Y., Tian, Y., Kong, Y., Zhong, B., Fu, Y.: Residual dense network for image super-resolution. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2472-2481 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 544, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 578 + ], + "type": "text", + "content": "46. Zhang, Z., Wang, R., Zhang, H., Chen, Y., Zuo, W.: Self-supervised learning for real-world super-resolution from dual zoomed observations. In: European Conference on Computer Vision. pp. 610-627. Springer (2022)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 414, + 91, + 447, + 100 + ], + "type": "text", + "content": "BurstM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/ByteEdit_ Boost, Comply and Accelerate Generative Image Editing/f2b7631d-9e11-4a5b-9762-122914fe2a69_content_list.json b/2024/ByteEdit_ Boost, Comply and Accelerate Generative Image Editing/f2b7631d-9e11-4a5b-9762-122914fe2a69_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..35b886cb8bc44870dc774adc3bbec515da80d0ff --- /dev/null +++ b/2024/ByteEdit_ Boost, Comply and Accelerate Generative Image Editing/f2b7631d-9e11-4a5b-9762-122914fe2a69_content_list.json @@ -0,0 +1,1523 @@ +[ + { + "type": "text", + "text": "ByteEdit: Boost, Comply and Accelerate Generative Image Editing", + "text_level": 1, + "bbox": [ + 267, + 140, + 736, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuxi Ren*, Jie Wu*†, Yanzuo Lu*, Huafeng Kuang, Xin Xia, Xionghui Wang, Qianqian Wang, Yixing Zhu, Pan Xie, Shiyin Wang, Xuefeng Xiao, Yitong Wang, Min Zheng, and Lean Fu", + "bbox": [ + 251, + 212, + 750, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ByteDance \nProject Page: https://byte-edit.github.io", + "bbox": [ + 346, + 270, + 653, + 299 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/06b8ed23b91324ec96d0ab590a7a942aae563d5ec4e5f443e4a4d2eeb5a21559.jpg", + "image_caption": [ + "Fig. 1: We introduce ByteEdit, a novel framework that utilizes feedback learning to enhance generative image editing tasks, resulting in outstanding generation performance, improved consistency, enhanced instruction adherence, and accelerated generation speed. To the best of our knowledge, ByteEdit emerges as the most superior and the fastest solution currently in the field of generative editing." + ], + "image_footnote": [], + "bbox": [ + 264, + 342, + 740, + 700 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. †Corresponding author.", + "bbox": [ + 238, + 824, + 537, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recent advancements in diffusion-based generative image editing have sparked a profound revolution, reshaping the landscape of image outpainting and inpainting tasks. Despite these strides, the field grapples with inherent challenges, including: i) inferior quality; ii) poor consistency; iii) insufficient instruction adherence; iv) suboptimal generation efficiency. To address these obstacles, we present ByteEdit, an innovative feedback learning framework meticulously designed to Boost, Comply, and Accelerate Generative Image Editing tasks. ByteEdit seamlessly integrates image reward models dedicated to enhancing aesthetics and image-text alignment, while also introducing a dense, pixel-level reward model tailored to foster coherence in the output. Furthermore, we propose a pioneering adversarial and progressive feedback learning strategy to expedite the model's inference speed. Through extensive large-scale user evaluations, we demonstrate that ByteEdit surpasses leading generative image editing products, including Adobe, Canva, and MeiTu, in both generation quality and consistency. ByteEdit-Outpainting exhibits a remarkable enhancement of $388\\%$ and $135\\%$ in quality and consistency, respectively, when compared to the baseline model. Experiments also verified that our acceleration models maintain excellent performance results in terms of quality and consistency.", + "bbox": [ + 261, + 147, + 743, + 424 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keywords: Outpainting $\\cdot$ Intpainting $\\cdot$ Feedback Learning", + "bbox": [ + 261, + 440, + 660, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 489, + 375, + 506 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The field of generative image editing has experienced remarkable advancements in recent years [5, 6, 20, 21, 23, 34, 37, 38, 42, 44, 45], propelled by the development of diffusion models [12, 25, 27, 28, 47]. This progress has led to the emergence of influential products that have reshaped the landscape of image editing. A notable example is Adobe Firefly [1], which has revolutionized the creative process by enabling users to seamlessly incorporate, extend, or remove content from images through simple text prompts, thereby transcending the traditional boundaries of Photoshop. In our paper, we focus on the domain of generative image editing, with particular emphasis on two key aspects: 1) Outpainting, involving the expansion of the surrounding scene in an image based on provided input or even without explicit prompts, and 2) Inpainting, encompassing the random masking of specific image regions followed by the generation of corresponding content guided by given prompts (Inpainting-Editing) or erase certain objects (Inpainting-Erasing). Despite the notable advancements achieved through diffusion-based algorithms, several challenges persist within this field:", + "bbox": [ + 217, + 532, + 787, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inferior Quality: the quality of generated images frequently falls short in terms of realism, aesthetic appeal, and fidelity to minute details.", + "bbox": [ + 215, + 762, + 785, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Insufficient Instruction Adherence: The existing models grapple with the arduous task of faithfully adhering to provided instructions, resulting in a lack of alignment between the generated image and the input text;", + "bbox": [ + 215, + 794, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Ren et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Poor Consistency: The generated regions exhibit an unsatisfactory level of coherence with the original image, manifesting as a deficiency in terms of color, style, texture, and other salient visual attributes;", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Suboptimal Generation Efficiency: The generation process is characterized by sluggish speeds and inadequate efficiency, thereby imposing significant obstacles when confronted with large-scale image editing endeavors.", + "bbox": [ + 212, + 191, + 782, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, various efforts have been made to address the aforementioned challenges in the field. For instance, Imagen Editor [34] has employed an object detection approach to extract inpainting masks, while simultaneously capitalizing on original high-resolution images to faithfully capture intricate details. SmartBrush [37] has adopted a multi-task training strategy coupled with precision controls, encompassing both text and shape guidance, to enhance visual quality, mask controllability, and preserve the background. Additionally, RePaint [21] has utilized a pre-trained unconditional DDPM [12] prior and ingeniously modified the reverse diffusion iterations to generate high-quality and diverse output images. However, these approaches have primarily focused on addressing singular problems and have yet to achieve a more comprehensive solution. Large Language Models (LLMs) has made a notable surge in incorporating learning based on human feedback, and initial endeavors have been undertaken in the Text-to-Image (T2I) domain [9,17,39,43,46,48]. Inspired by these developments, we pose the question: Can we leverage human feedback to guide generative image editing to unleash the potential for superior generation outcomes?", + "bbox": [ + 212, + 237, + 785, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This paper introduces ByteEdit, an innovative framework for optimizing generative image editing through the incorporation of feedback learning. ByteEdit builds multiple reward models, namely the Aesthetic reward model, Alignment reward model, and Coherent reward model, to achieve exceptional generation effects, improved instruction adherence and enhanced consistency, respectively. These carefully designed reward models serve as the foundation for our proposed perceptual feedback learning (PeFL) mechanism, which provides task-aware and comprehensive supervision signals. Moreover, ByteEdit introduce an adversarial feedback learning strategy that employs the trainable reward model as the discriminator. This strategy ensures that the model benefits from the PeFL supervision and provide clear images even during high-noise stages, further improves both the performance and speed of our model. To expedite the sampling process, a progressive training strategy is employed to gradually reduce the optimization time steps and facilitate model inference in a low-steps regime.", + "bbox": [ + 212, + 479, + 785, + 690 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- New Insight: To the best of our knowledge, we offer the first attempt to incorporate human feedback into the field of generative image editing. By-Edit significantly enhances the overall performance of the model across various key aspects, opening new horizons in this field of study.", + "- Comprehensive Framework: By designing complementary global-level and pixel-level reward models, we effectively guide the model towards achieving improved beauty, enhanced consistency, and superior image-text alignment.", + "Efficiency and Pioneering: Progressive feedback and adversarial learning techniques are introduced to accomplish a remarkable acceleration in the" + ], + "bbox": [ + 225, + 703, + 782, + 838 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "ByteEdit", + "bbox": [ + 666, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "model's inference speed, all while maintaining a minimal compromise on output quality. Notably, ByteEdit stands as the first successful endeavor in accelerating generative editing models.", + "bbox": [ + 241, + 146, + 785, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Outstanding Performance: Extensive user studies show that ByteEdit exhibits obvious advantages in terms of quality, consistency, efficiency, and speed, compared to the most competitive products. ByteEdit emerges as the fastest and most superior solution currently available in image editing.", + "bbox": [ + 225, + 191, + 785, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 277, + 387, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generative Image Editing. Generative Image Editing is a research area focused on filling user-specified regions of interest with desired contents. GLIDE [23] is the pioneering work that introduced text-to-image diffusion for editing purposes, and Repaint [21], on the other hand, conditions an unconditionally trained model (e.g. DDPM [12]) and leverages visible pixels to fill in missing areas. To enable more precise editing control, Blended Diffusion [5] incorporates multimodal embeddings and enforces similarity between images and text using CLIP [26]. SmartBrush [37] pushes the boundaries of mask generation by utilizing instance and panoptic segmentation datasets instead of random generation. Further improvements include the introduction of the powerful Segment Anything (SAM) [15] model by [45], which achieves mask-free removal, filling, and replacing of multiple pipelines. Inst-Inpaint [44] specializes in text-driven object removal without the need for explicit masking. Additionally, this method proposes the GQA-Inpaint dataset, which comprises image pairs with and without the object of interest, facilitating effective object removal guided by textual input. In addition to comparing our proposed method with existing academic approaches, we also benchmark against industry methods like Adobe [1], Canva [2], and MeiTu [3], providing a comprehensive evaluation across different domains and highlighting the strengths of our approach.", + "bbox": [ + 212, + 311, + 787, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human Feedback Learning. Foundation models for text-to-image diffusion often rely on pre-training with large-scale web datasets, such as LAION-5B [31], which may result in generated content that deviates from human ethical values and legal compliance requirements. Previous approaches [9,17] attempted to address this issue by constructing preference datasets using hand-crafted prompts or expert generators. However, these methods suffered from over-fitting due to their limited real-world scenarios and generator capabilities. To overcome these limitations, researchers proposed various reward models trained with expert annotations [35,39] or feedback from web users [13,16] to enforce alignment with human preferences. Drawing inspiration from reinforcement learning with human feedback (RLHF) utilized in natural language processing (NLP), researchers explored the application of human feedback learning in text-to-image diffusion [39, 43, 46, 48] to achieve more realistic, faithful, and ethical outcomes. Among these efforts, ImageReward [39] primarily focused on overall image quality and overlooked the complexity of human perception. In our work, we extend the concept of human feedback learning by introducing three fine-grained independ-", + "bbox": [ + 212, + 598, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Ren et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "dent reward models tailored for generative image editing: aesthetics, image-text alignment, and pixel-level coherence.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 ByteEdit: Boost, Comply and Accelerate", + "text_level": 1, + "bbox": [ + 214, + 203, + 648, + 222 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ByteEdit, focuses on generative image editing tasks that enable users to manipulate image content within a specific region of interest using textual descriptions. With an input image $x$ , a region-of-interest mask $m$ , and a user-provided textual description $c$ , our primary objective is to generate an output image $y$ that preserves the unmasked region in the input image $x$ , while aligning the masked region well with both the description of $c$ and visual attributes in $x$ . In this study, we introduce two key functionalities within ByteEdit: ByteEdit-Outpainting and ByteEdit-Inpainting. ByteEdit-Outpainting extends the image by generating content beyond the boundaries of the input image, while ByteEdit-Inpainting fills or erases in arbitrary areas of the input image.", + "bbox": [ + 212, + 239, + 787, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The ByteEdit pipeline is presented in Fig 2, providing an overview of the system's workflow. In the subsequent subsections, we delve into the details of two crucial components: Boost (Sec. 3.1) and Comply (Sec. 3.2). Furthermore, we elaborate on the Accelerate scheme in Sec. 3.3, which illustrates an approach to expedite the processing time and improve the efficiency of the ByteEdit system.", + "bbox": [ + 212, + 391, + 787, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Boost: Perceptual Feedback Learning", + "text_level": 1, + "bbox": [ + 214, + 493, + 570, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the field of generative image editing, the persistent challenge of subpar quality has impelled us to propose a pioneering approach that introduces human feedback, hitherto unexplored in this domain. Our novel pipeline comprises three key components: feedback data collection, reward model training, and perceptual feedback learning.", + "bbox": [ + 212, + 522, + 785, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feedback Data Collection. We first randomly extract more than 1,500,000 text prompts from the Midjourney Discord [33] and MS-COCO Caption [7] datasets. To ensure the diversity, a clustering algorithm, namely K-Means, was employed, leveraging the similarities derived from state-of-the-art large language models [19]. Further, the features were visualized in lower dimensions using t-SNE [22], enabling the identification of data points possessing the largest average distance from their k-nearest neighbors. We also manually eliminate less informative and decorative-dominated prompts such as \"unbelievable\", \"fantastic\" and \"brilliant\" to improve the prompt quality. This meticulous procedure yielded approximately 400,000 candidate prompts, exhibiting diverse distributions, which were subsequently subjected to multiple text-to-image diffusion models, including SD1.5 [28] and SDXL [24]. Those images which excessively inferior quality or ambiguous characteristic are manually removed. Accompanying each prompt, a set of four generated images was presented to experts, who were tasked with selecting the best and worst images based on various aspects, encompassing aesthetic appeal, color composition, structural coherence, and brightness. The", + "bbox": [ + 212, + 598, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "ByteEdit", + "bbox": [ + 666, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9aed3886d29a05e2f603ec6aa56f481fdab35e7cb99183c0e2843c78d338734b.jpg", + "image_caption": [ + "Fig. 2: ByteEdit formulates a comprehensive feedback learning framework that facilitating aesthetics, image-text matching, consistency and inference speed." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 787, + 324 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "resulting dataset, herein referred to as the aesthetic preference dataset $\\mathcal{D}_{aes}$ , encompasses a collection of massive triplets $(c,x_p,x_n)$ , where $x_{p}$ and $x_{n}$ correspond to the preferred and non-preferred generated images of prompt $c$ , respectively.", + "bbox": [ + 212, + 391, + 787, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reward Model Training. Building upon this dataset, we follow the training techniques in [39] to learn an aesthetic reward model $R_{\\alpha}(\\cdot)$ of trainable parameters $\\alpha$ , which we briefly summarize here. The image and text features of the input are extracted from the BLIP [18] backbone, combined with cross attention, and fed into an MLP to obtain an aesthetic score. The training objective can be formulated as,", + "bbox": [ + 212, + 436, + 787, + 526 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\alpha) = - \\mathbb {E} _ {(c, x _ {p}, x _ {n}) \\sim \\mathcal {D} _ {a e s}} [ \\log \\sigma (R _ {\\alpha} (c, x _ {p}) - R _ {\\alpha} (c, x _ {n})) ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 305, + 541, + 785, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\sigma (\\cdot)$ represents the Sigmoid function used for normalization.", + "bbox": [ + 212, + 569, + 694, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Perceptual Feedback Learning. Leveraging the power of our crafted reward model, we specifically introduce Perceptual Feedback Learning (PeFL) to fine-tune diffusion models with human feedback for generative image editing. Departing from the conventional practice of sequentially refining the predictions from the final step $x_{T}$ to the initial step $x_0'$ ( $x_{T} \\to x_{T-1} \\to \\dots \\to x_0'$ ), we adopt an innovative perspective by performing optimization on the direct prediction outcomes $x_{t} \\to x_{0}'$ at various intermediate steps $t \\in [1,T]$ ( $T = 20$ in this case). Through this dynamic exploration across different stages of denoising, we uncover the following distinctive observations:", + "bbox": [ + 212, + 585, + 787, + 722 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- During the initial stages of denoising ( $t \\in [11, 20]$ ), the generative model (i.e. U-Net [29]) struggles to seamlessly complement the full image. Consequently, the reward model encounters challenges in accurately evaluating images that are hindered by obstacle masking.", + "- As the denoising process progresses ( $t \\in [1, 10]$ ), the reward model is able to identify and reward visually appealing images even in the presence of mild noise." + ], + "bbox": [ + 223, + 733, + 782, + 838 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Ren et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Drawing upon these insightful observations, we present an innovative stagewise approach, to amplify the potential of generative image editing. Our proposed method encompasses the following key stages: 1) In stage 1 ( $t \\in [16,20]$ ), we simply skip these steps with extremely heavy noise by diffusing the masked input $x \\odot (1 - m)$ into noisy latents at a fixed step $T_{1} = 15$ . This strategy is motivated by the fact that the generative model's ability to fill in intricate details within the masked region is limited in this timestep, thus rendering the training overhead unnecessary. This is the central difference between our approach and the ReFL proposed in [39], whose training paradigm relies solely on prompts and starts inference from pure noise. We aim to bring about a more pronounced correlation between the output and input image, thus facilitating the preservation of details in the unmasked areas; 2) In stage 2 ( $t \\in [t',15]$ ), we randomly select a denosing step $t' \\sim [1,T_2]$ ( $T_{2} = 10$ in this case) and perform inference without gradient starting from the noisy latent, i.e. $x_{T_1} \\to x_{T_1-1} \\to \\dots \\to x_{t'}$ . This method ensures that the complemented image generated at this stage exhibits a sufficiently high level of quality, rendering it more amenable to evaluation by the reward model; 3) In stage 3 ( $x_{t'} \\to x_0'$ ), the final stage entails the direct prediction of the complemented image $x_0'$ . We leverage the aesthetic score obtained from the reward model as invaluable human feedback to refine the generative model $G_{\\phi}(\\cdot)$ . This refinement process is achieved through the utilization of the following loss function:", + "bbox": [ + 212, + 146, + 787, + 464 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e w a r d}} (\\phi) = - \\mathbb {E} _ {(x, m, c) \\sim \\mathcal {D} _ {\\text {t r a i n}}, t ^ {\\prime} \\sim [ 1, T _ {2} ]} \\left[ \\log \\sigma \\left(R _ {\\alpha} (c, G _ {\\phi} (x, m, c, t ^ {\\prime}))\\right) \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 259, + 474, + 785, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{D}_{train}$ represents the fine-tuning dataset (i.e. LAION-5B [31]). The term $G\\phi(x, m, c, t')$ denotes the decoded output image $x_0'$ generated by the generative model at step $t'$ , given the masked input $x \\odot (1 - m)$ and the prompt $c$ . To further maintain the consistency and detail fidelity of the generated area and the original image area, we introduce pixel-level regularization (i.e., L1 loss) and a perceptual loss, which captures the discrepancy in VGG features [32]. Collectively, these regularization techniques can be formulated as follows:", + "bbox": [ + 212, + 502, + 787, + 609 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {r e g} (\\phi) = \\mathbb {E} _ {(x, m, c) \\sim \\mathcal {D} _ {t r a i n}, t ^ {\\prime} \\sim [ 1, T _ {2} ]} \\| x - G _ {\\phi} (x, m, c, t ^ {\\prime}) \\| _ {1}, \\\\ \\mathcal {L} _ {r e g} (\\cdot) = \\mathbb {E} _ {\\| V (x) - V (C _ {r e g}, t ^ {\\prime}) \\|} \\end{array} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 281, + 618, + 785, + 643 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v g g} (\\phi) = \\mathbb {E} _ {(x, m, c) \\sim \\mathcal {D} _ {t r a i n}, t ^ {\\prime} \\sim [ 1, T _ {2} ]} \\| V (x) - V \\left(G _ {\\phi} (x, m, c, t ^ {\\prime})\\right) \\| _ {1},\n$$\n", + "text_format": "latex", + "bbox": [ + 281, + 638, + 722, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $V(\\cdot)$ represents the VGG network. The overall training objective of our PeFL can be summarized as,", + "bbox": [ + 212, + 664, + 787, + 694 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p e f l} (\\phi) = \\mathcal {L} _ {r e w a r d} + \\eta \\left(\\mathcal {L} _ {r e g} + \\mathcal {L} _ {v g g}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 705, + 785, + 723 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\eta$ is a hyperparameter for balancing loss weights.", + "bbox": [ + 214, + 733, + 614, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Comply: Image-Text Alignment with Coherence", + "text_level": 1, + "bbox": [ + 214, + 768, + 658, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Diverging from the text-to-image synthesis focus of [39], our method encompasses an additional emphasis on assessing the alignment between the generated content of the masked area and the user-specified prompt, as well as", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "ByteEdit", + "bbox": [ + 666, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ensuring coherence with the unmasked region at the pixel level. To achieve this, we introduce two further components in this section, which complement the aesthetic reward model $R_{\\alpha}$ proposed earlier.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Image-Text Alignment. We note that the presence of numerous poorly matched image-text pairs within the LAION dataset [31]. Exploiting these pairs as non-preferred samples for training reward models allows us to reduce manual annotation costs significantly. Initially, we employ the CLIP model [26] to identify image-text pairs with lower CLIPScore [11] from the LAION dataset. Subsequently, we leverage advanced multi-modal large language models such as LLAVA [19] to generate more informative and descriptive captions for the input images. These captions are considered more accurate than the original prompts. This process yields approximately 40,000 triplets $(c_p, c_n, x)$ as alignment dataset $\\mathcal{D}_{\\text{align}}$ , where $c_p$ and $c_n$ correspond to the preferred and non-preferred textual descriptions of the image $x$ , respectively. These triplets are utilized for training the image-text alignment reward model, denoted as $R_{\\beta}(\\cdot)$ . The architecture of $R_{\\beta}$ mirrors that of $R_{\\alpha}$ , while the training objective is similar to Eq. 1:", + "bbox": [ + 212, + 191, + 792, + 388 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\beta) = - \\mathbb {E} _ {\\left(c _ {p}, c _ {n}, x\\right) \\sim \\mathcal {D} _ {\\text {a l i g n}}} \\left[ \\log \\sigma \\left(R _ {\\beta} \\left(c _ {p}, x\\right) - R _ {\\beta} \\left(c _ {n}, x\\right)\\right) \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 300, + 398, + 785, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pixel-Level Coherence. The issue of coherence arises from the presence of inconsistent content within and outside the regions of interest, characterized by subtle visual cues such as color variations, stylistic discrepancies, and textural differences. To tackle this challenge, a coherent reward model, denoted as $R_{\\gamma}(\\cdot)$ , is specifically designed for pixel-level discrimination, as opposed to the holistic evaluation performed by $R_{\\alpha}(\\cdot)$ and $R_{\\beta}(\\cdot)$ . Our approach entails training a ViT-based [10] backbone network, followed by a prediction MLP head, to assess the authenticity and assign a score to each pixel in the input image. By formulating the loss function as follows:", + "bbox": [ + 212, + 425, + 787, + 559 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\gamma) = - \\mathbb {E} _ {(x, m, c) \\sim \\mathcal {D} _ {\\text {t r a i n}, t ^ {\\prime}} \\sim [ 1, T _ {2} ]} \\left[ \\log \\sigma \\left(R _ {\\gamma} (z)\\right) + \\log \\left(1 - \\sigma \\left(R _ {\\gamma} \\left(z ^ {\\prime}\\right)\\right)\\right) \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 259, + 571, + 785, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $z \\sim x \\in \\mathbb{R}^{H \\times W \\times 3}, z' \\sim G_{\\phi}(x, m, c, t') \\in \\mathbb{R}^{H \\times W \\times 3}$ are pixels of the corresponding image and $H, W$ represent the height and weight respectively.", + "bbox": [ + 212, + 597, + 784, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Accelerate: Adversarial and Progressive Training", + "text_level": 1, + "bbox": [ + 214, + 650, + 665, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adversarial training. Concurrent works such as UFOGen [41] and SDXL-Turbo [30] proposed to introduce adversarial training objective into fine-tuning diffusion models, which dramatically speeds up the sampling process and allows for one-step generation. They supposed that the Gaussian assumption of diffusion process does not hold anymore when the inference steps are extremely low, and therefore enabling the generative model to output samples in a single forward step by adversarial objective [36, 40]. We note that the functionality of our coherent reward model $R_{\\gamma}(\\cdot)$ is very similar to that of the discriminator in adversarial training, except for the different granularity in prediction. To this end, unlike the aesthetic and alignment reward models, which necessitate offline learning prior to fine-tuning, the coherent reward model can be learned online", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Ren et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and seamlessly integrated into the fine-tuning process. The adversarial objective of generator that raises the score of output image is also in compliance with our feedback learning in Eq. (2), we can simply achieve adversarial training by incorporating the optimization of $R_{\\gamma}(\\cdot)$ into fine-tuning to serve as a discriminator. Thus the Eq. (2) can be reformulated as follows:", + "bbox": [ + 212, + 146, + 782, + 223 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e w a r d}} (\\phi) = - \\underset { \\begin{array}{c} (x, m, c) \\sim \\mathcal {D} _ {\\text {t r a i n}} \\\\ t ^ {\\prime} \\sim [ 1, T _ {2} ] \\end{array} } {\\mathbb {E}} \\sum_ {\\theta \\in \\{\\alpha , \\beta , \\gamma \\}} \\log \\sigma \\left(R _ {\\theta} \\left(c, G _ {\\phi} (x, m, c, t ^ {\\prime})\\right)\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 269, + 237, + 785, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For completeness, we also rewrite the overall training objective as,", + "bbox": [ + 214, + 281, + 692, + 297 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {G} (\\phi) = \\mathcal {L} _ {p e f l} (\\phi) = \\mathcal {L} _ {r e w a r d} + \\eta \\left(\\mathcal {L} _ {r e g} + \\mathcal {L} _ {v g g}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 246, + 309, + 784, + 335 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {D} (\\gamma) = - \\mathbb {E} _ {(x, m, c) \\sim \\mathcal {D} _ {t r a i n}, t ^ {\\prime} \\sim [ 1, T _ {2} ]} \\left[ \\log \\sigma \\left(R _ {\\gamma} (z)\\right) + \\log \\left(1 - \\sigma \\left(R _ {\\gamma} \\left(z ^ {\\prime}\\right)\\right)\\right) \\right].\n$$\n", + "text_format": "latex", + "bbox": [ + 246, + 330, + 736, + 347 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Progressive training. To expedite the sampling process, we employ a progressive training strategy where we gradually reduce the optimization time steps $T$ . Surprisingly, we find that the quality of the generated images does not significantly degrade under the supervisor of reward models. This approach strikes a fine balance between performance and speed, leading to compelling results. In our experiments, we adopt a two-phase progressive strategy. During phase 1, we set the optimization time steps as $T = 20$ , $T_{1} = 15$ , and $T_{2} = 10$ . In phase 2, we further decrease the time steps to $T = 8$ , $T_{1} = 6$ , and $T_{2} = 3$ . Notably, we achieve remarkable outcomes without employing any distillation operations, relying solely on the inheritance of model parameters.", + "bbox": [ + 212, + 359, + 784, + 510 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 536, + 374, + 553 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 570, + 455, + 585 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dataset. The fine-tuning dataset, denoted as $\\mathcal{D}_{train}$ , consists of a substantial collection of 7,562,283 images encompassing diverse domains such as real-world scenes, authentic portraits, and computer-generated (CG) images. To enhance the model's generalization ability and generation quality, we adopted a meticulous fine-grained masking strategy inspired by StableDiffusion [28]. Our masking strategy encompasses four distinct types of masks: global masks, irregular shape masks, square masks, and outward expansion masks. Each mask type corresponds to a specific probability value, which is randomly selected and applied to images during the training process. Moreover, we devised a specialized masking strategy tailored for Inpainting-Editing tasks. Leveraging instance-level data, which identifies specific objects within images, we introduced random dilation operations to coarsen the masks during training. These coarsened masks were then integrated with randomly generated masks surrounding the instances. This approach not only enhances the model's performance in instruction-based image editing tasks but also facilitates more accurate instance generation, ultimately leading to superior quality outputs.", + "bbox": [ + 212, + 598, + 787, + 843 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "ByteEdit", + "bbox": [ + 666, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To evaluate the performance of our approach, we conducted comprehensive qualitative and quantitative assessments using two distinct datasets. The first dataset, UserBench, was meticulously curated by gathering a vast amount of user-customized image-text matching data from various online sources. This dataset proved invaluable for evaluating image inpainting and outpainting tasks. From this extensive collection, we judiciously handpicked 100 high-quality image-text matching pairs to serve as our test data. We also leverage the experimental results from this dataset to collect and report human preferences. The second dataset, EditBench [34], presents a novel benchmark specifically tailored for text-guided image inpainting. Consisting of 240 images, each image within EditBench is paired with a corresponding mask that precisely delineates the region within the image to be modified through inpainting.", + "bbox": [ + 212, + 146, + 787, + 328 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training Setting. To facilitate the perceptual feedback learning stage, we employed a relatively small learning rate of 2e-06, complemented by a learning rate scheduling strategy that encompassed a warm-up phase consisting of 1000 iterations. Furthermore, to ensure stability in model parameter updates, we incorporated an exponential moving average (EMA) decay parameter set to 0.9999. Instead of employing $100\\%$ noise as in ReFL [39], we introduced a 50The weight assigned to the perceptual feedback loss was set to 0.01. During the adversarial acceleration stage, we maintained similar settings to the perceptual feedback learning stage, with an additional adversarial loss weighted 0.05.", + "bbox": [ + 212, + 332, + 787, + 470 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Evaluation Principles and Criteria", + "text_level": 1, + "bbox": [ + 215, + 513, + 545, + 529 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Subjective metrics. To assess the robustness of our proposed method, we conducted a comprehensive subjective evaluation involving both expert evaluations and a large number of volunteer participants. Expert evaluators were tasked with individually assessing each generated image and assigning scores based on three key aspects: coherence, structure, and aesthetics. These aspects were rated on a scale of 1 to 5, with higher scores indicating superior generation quality: 1) Coherence focused on evaluating the consistency of texture, style, and color between the generated region and the original image. 2) Structure emphasized the clarity, sharpness, and absence of deformations or mutilations, particularly in human body parts. 3) Aesthetics gauged the overall level of creativity and diversity exhibited by the generated images. In addition to expert evaluations, we also sought the opinions of a large number of volunteers, with over 36,000 samples collected. These volunteers were presented with pairs of generated images and asked to choose between \"Good\", \"Same\", or \"Bad\", representing their preference in terms of GSB (Good-Same-Bad).", + "bbox": [ + 212, + 563, + 787, + 790 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Objective metrics. In this study, we also incorporate objective text-image alignment metrics, specifically CLIPScore [11,26] and BLIPScore [18], to comprehensively evaluate the alignment of our models.", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Ren et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/833c2881cba17089f16db7badee976bb7fce3ae8b256cd9eb50ff739b39b2c33.jpg", + "table_caption": [ + "Table 1: Comparisons with state-of-the-art generative image editing systems in terms of coherence, structure and aesthetic scored by experts. More than 6000 image-text pairs are randomly sampled for each task and we report the average scores." + ], + "table_footnote": [], + "table_body": "
MethodOutpaintingInpainting-EditingInpainting-Erasing
coherencestructureaestheticcoherencestructureaestheticcoherencestructure
MeiTu [3]3.012.732.752.772.892.513.313.25
Canva [2]2.722.852.653.423.403.082.922.90
Adobe [1]3.523.073.143.463.603.223.854.28
ByteEdit3.543.253.263.733.393.253.994.03
", + "bbox": [ + 225, + 198, + 772, + 282 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/cf636bf8a617296e06d56109b66d0df90bbf6322a0792683ded90017dc4a4555.jpg", + "image_caption": [ + "Fig. 3: Comparisons with state-of-the-art generative image editing systems in terms of human preference (i.e. GSB). More than 12,000 samples are collected for each task. For simplicity and to minimize the difficulty of collecting a large number of user opinions, we only offer the generated images by Adobe and our ByteEdit to the volunteers. \"Good\" indicates the generated images by our ByteEdit is preferred and vice versa." + ], + "image_footnote": [], + "bbox": [ + 222, + 301, + 779, + 416 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Comparisons with State of the arts", + "text_level": 1, + "bbox": [ + 214, + 534, + 552, + 549 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We compare our method with concurrent state-of-the-art generative image editing systems such as Adobe [1], Canva [2] and MeiTu [3]. The comparisons cover three different tasks, including inpainting, inpainting-editing and inpainting-erasing. The inpainting editing will specify the content to be generated for the region of interest in the prompt. In contrast, inpainting-erasing requires the model to remove content within it and be as consistent as possible. Since the erased image has little change, experts were not asked to score aesthetics for user study.", + "bbox": [ + 212, + 566, + 787, + 686 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "User study. The average scores evaluated by experts are shown in Tab. 1. From the results, our ByteEdit significantly outperforms the state-of-the-art across different metrics in the outpainting task. It demonstrates that our method works well to expand images based on existing content and maintains superior consistency, structural integrity, and creativity. As for the inpainting tasks, our method also can provide the most coherent edited or erased images. To further investigate the gap between Adobe and our proposed ByteEdit, we solicited feedback from a large number of volunteers on the images generated by both, and the results are illustrated in Fig. 3. The results show that users generally found the images we generated to be more natural in overall perception. Our GSB", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "ByteEdit", + "bbox": [ + 666, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/a16ea7fb64e1e2b7bdcd8253f55cad09bba32266650de6ff64ce7cab64527a58.jpg", + "table_caption": [ + "Table 2: The quantitative results of ByteEdit and recent state-of-the-art approaches." + ], + "table_footnote": [], + "table_body": "
MetricsUserBenchEditBench
Meitu [3]Canva [2]Adobe [1]ByteEditDiffEdit [8]BLD [4]EMILIE [14]ByteEdit
CLIPScore0.2350.2410.2370.2550.2720.2800.3110.329
BLIPScore0.1740.4670.4500.6870.5820.5960.6200.691
", + "bbox": [ + 217, + 171, + 785, + 234 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2ed6c91839c8f57f22cd4e2f2b4db298eaa9e4af96f6a5de68881f7ac9ddd2c4.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparison in inpainting. We highlight key areas with red boxes." + ], + "image_footnote": [], + "bbox": [ + 251, + 263, + 756, + 422 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "superiority percentages (i.e. $(\\mathrm{G} + \\mathrm{S}) / (\\mathrm{S} + \\mathrm{B}) * 100\\%$ ) on three different tasks are $105\\%$ , $163\\%$ , and $112\\%$ , respectively.", + "bbox": [ + 212, + 491, + 784, + 523 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Quantitative comparison. To quantitatively evaluate the performance of our method compared to other approaches, we conduct a quantitative evaluation of how well the edited image can capture the semantics of the edit instruction successfully by measuring the CLIPScore and BLIPScore. We conduct the experiment in inpainting-editing task and the results are provided in Table 2. From the UserBench against state-of-the-art generative image editing systems, we noticed that the score results are not completely consistent with user opinion. Nevertheless, our method is still ahead of the second-place Canva by $5.8\\% (+0.014)$ and $47.1\\% (+0.22)$ in terms of CLIPScore and BLIPScore, respectively. As for the EditBench, we follow [14] to compare our method with several concurrent editing approaches, i.e. DiffEdit [8], BLD [4] and EMILIE [14]. It demonstrates that the ByteEdit consistently yields the state-of-the-art performance, which shows our superior quality, consistency and instruction adherence.", + "bbox": [ + 212, + 530, + 787, + 727 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative comparison. In Figure 4 and 5, we visualize samples produced by different systems under different tasks. It clearly shows that our method exhibits a superior performance for learning both coherence and aesthetic. For the inpainting task, the ByteEdit consistently follows the user-specified instructions and generates coherent images with better image-text alignment. It is worth noting that our system allows both prompt and prompt-free generation when it comes to outpainting, which has broader application scenarios in reality.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Ren et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2fd3228be46e22b2d5511fad86f39e758db184777a7121df7de83fe54e5c2fe0.jpg", + "image_caption": [ + "Fig. 5: Qualitative comparison in outpainting. We highlight key areas with red boxes." + ], + "image_footnote": [], + "bbox": [ + 243, + 146, + 759, + 328 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b47e8664a10faa04b98ea508f38751d4b719770e0df2d26b52bb680484e4e2ff.jpg", + "image_caption": [ + "(a) w/ vs w/o PeFL" + ], + "image_footnote": [], + "bbox": [ + 218, + 367, + 498, + 518 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5d5138c8965e926afcb2ceb4e1efac445ef22008f9c393404657f5769e94784c.jpg", + "image_caption": [ + "(b) w/ vs w/o Acceleration", + "Fig. 6: Human Perference Evaluation on our proposed PeFL and Acceleration." + ], + "image_footnote": [], + "bbox": [ + 501, + 366, + 782, + 520 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 590, + 400, + 603 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Figure 6, we conduct ablation experiments on both our proposed PeFL and acceleration strategy. The experts were asked to choose GSB preference and we report the human preference rates in the figure, which are calculated as $(\\mathrm{G} + \\mathrm{S}) / [((\\mathrm{G} + \\mathrm{S}) + (\\mathrm{S} + \\mathrm{B}))] * 100\\%$ for win and $(\\mathrm{S} + \\mathrm{B}) / [((\\mathrm{G} + \\mathrm{S}) + (\\mathrm{S} + \\mathrm{B}))] * 100\\%$ for lose, respectively. The evaluation is similar to the user study, except that we combine structure and aesthetics to reduce evaluation costs. More visualizations are also included in Figure 7.", + "bbox": [ + 212, + 613, + 787, + 718 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "PeFL preference. From the results in Figure 6(a), our proposed PeFL significantly improves the generation quality, outperforming the baseline on all different tasks. Especially in the outpainting task with PeFL, our method exceeds the baseline by about $60\\%$ in terms of structure and aesthetic, which is consistent with the edited image shown at the top of Figure 7 that is more realistic and conforms to the rules of realistic physics.", + "bbox": [ + 212, + 719, + 787, + 809 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Acceleration preference. In Figure 6(b), we demonstrate that our model has no significant loss in either consistency or structure and aesthetic with the pro", + "bbox": [ + 214, + 809, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "ByteEdit", + "bbox": [ + 668, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d098ace01f6fdeeab4ea8d6aa5108b57bd86e15abe9483cf188d30209631a3a5.jpg", + "image_caption": [ + "Fig. 7: Ablation Studies Visualization." + ], + "image_footnote": [], + "bbox": [ + 217, + 146, + 781, + 320 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "gressive training strategy. To our surprise, we have even achieved both increasing speed and quality in the outpainting and inpainting-editing tasks. Based on our experimental observations, this phenomenon can be attributed to two underlying factors: i) Stable Training: By considering the discriminator as a reward model, trainable reward model offers flexible and robust supervision for PeFL, alleviating issues related to model over-optimization; ii) Expand Supervision Scope: The incorporation of adversarial supervision enables us to extend the time step of PEFL optimization. Consequently, even at high-noise stages, such as step 999, the model can still benefit from PeFL supervision, further driving improvements in model performance. The visualization at the bottom of Figure 7 also verifies this, where the outputs become more realistic and natural after acceleration.", + "bbox": [ + 212, + 375, + 787, + 541 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Discussion", + "text_level": 1, + "bbox": [ + 215, + 561, + 354, + 578 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "ByteEdit has demonstrated remarkable performance across various editing tasks. However, several promising directions warrant further exploration:", + "bbox": [ + 212, + 592, + 784, + 625 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Performance: One avenue for improvement lies in developing more targeted reward models tailored to specific editing tasks. By refining the reward models, we can potentially unlock even higher levels of performance and generate more precise and desirable output.", + "- Acceleration: Another area of interest is investigating how ByteEdit can be further integrated with advanced techniques such as LCM and SDXL-turbo to achieve accelerated processing speeds.", + "- Task: Expanding the capabilities of ByteEdit beyond image editing to domains like video editing or instruction editing holds significant potential." + ], + "bbox": [ + 223, + 633, + 785, + 768 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "By incorporating human feedback to optimize generative image editing, ByteEdit can greatly enhance the practicality and usability in real-world scenarios. We hope that our work will provide valuable insights and inspire deeper reflections in this field, propelling its future development.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Ren et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Adobe firefly - free generative ai for creatives. https://www.adobe.com/products/firefly.html 2, 4, 11, 12", + "2. Free ai image generator: Online text to image app | canva. https://www.canva.com/ai-image-generator/ 4, 11, 12", + "3. Miraclevision. https://ai.meitu.com/index/ 4,11,12", + "4. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. ACM Transactions on Graphics (TOG) 42(4), 1-11 (2023) 12", + "5. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18208-18218 (2022) 2, 4", + "6. Chen, X., Huang, L., Liu, Y., Shen, Y., Zhao, D., Zhao, H.: Anydoor: Zero-shot object-level image customization. arXiv preprint arXiv:2307.09481 (2023) 2", + "7. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dollar, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015) 5", + "8. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427 (2022) 12", + "9. Dong, H., Xiong, W., Goyal, D., Pan, R., Diao, S., Zhang, J., Shum, K., Zhang, T.: Raft: Reward ranked finetuning for generative foundation model alignment. arXiv preprint arXiv:2304.06767 (2023) 3, 4", + "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 8", + "11. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021) 8, 10", + "12. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020) 2, 3, 4", + "13. Isajanyan, A., Shatveryan, A., Kocharyan, D., Wang, Z., Shi, H.: Social reward: Evaluating and enhancing generative ai through million-user feedback from an online creative community. arXiv preprint arXiv:2402.09872 (2024) 4", + "14. Joseph, K., Udhayanan, P., Shukla, T., Agarwal, A., Karanam, S., Goswami, K., Srinivasan, B.V.: Iterative multi-granular image editing using diffusion models. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 8107-8116 (2024) 12", + "15. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) 4", + "16. Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., Levy, O.: Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems 36 (2024) 4", + "17. Lee, K., Liu, H., Ryu, M., Watkins, O., Du, Y., Boutilier, C., Abbeel, P., Ghavamzadeh, M., Gu, S.S.: Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192 (2023) 3, 4", + "18. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022) 6, 10" + ], + "bbox": [ + 218, + 172, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "ByteEdit", + "bbox": [ + 666, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023) 5, 8", + "20. Lu, Y., Zhang, M., Ma, A.J., Xie, X., Lai, J.H.: Coarse-to-fine latent diffusion for pose-guided person image synthesis. arXiv preprint arXiv:2402.18078 (2024) 2", + "21. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11461-11471 (2022) 2, 3, 4", + "22. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008) 5", + "23. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021) 2, 4", + "24. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023) 5", + "25. Qin, J., Wu, J., Chen, W., Ren, Y., Li, H., Wu, H., Xiao, X., Wang, R., Wen, S.: Diffusionpt: Llm-driven text-to-image generation system. arXiv preprint arXiv:2401.10061 (2024) 2", + "26. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 4, 8, 10", + "27. Ren, Y., Wu, J., Zhang, P., Zhang, M., Xiao, X., He, Q., Wang, R., Zheng, M., Pan, X.: Ugc: Unified gan compression for efficient image-to-image translation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17281-17291 (2023) 2", + "28. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) 2, 5, 9", + "29. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) 6", + "30. Sauer, A., Lorenz, D., Blattmann, A., Rombach, R.: Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042 (2023) 8", + "31. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022) 4, 7, 8", + "32. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) 7", + "33. Turc, I., Nemade, G.: Midjourney user prompts & generated images (250k) (2022). https://doi.org/10.34740/KAGGLE/DS/2349267 5", + "34. Wang, S., Sahara, C., Montgomery, C., Pont-Tuset, J., Noy, S., Pellegrini, S., Onoe, Y., Laszlo, S., Fleet, D.J., Soricut, R., et al.: Imagen editor and editbench: Advancing and evaluating text-guided image inpainting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18359-18369 (2023) 2, 3, 10" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Ren et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Wu, X., Sun, K., Zhu, F., Zhao, R., Li, H.: Human preference score: Better aligning text-to-image models with human preference. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2096-2105 (2023) 4", + "36. Xiao, Z., Kreis, K., Vahdat, A.: Tackling the generative learning trilemma with denoising diffusion gans. arXiv preprint arXiv:2112.07804 (2021) 8", + "37. Xie, S., Zhang, Z., Lin, Z., Hinz, T., Zhang, K.: Smartbrush: Text and shape guided object inpainting with diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22428-22437 (2023) 2, 3, 4", + "38. Xie, S., Zhao, Y., Xiao, Z., Chan, K.C., Li, Y., Xu, Y., Zhang, K., Hou, T.: Dreaminpainter: Text-guided subject-driven image inpainting with diffusion models. arXiv preprint arXiv:2312.03771 (2023) 2", + "39. Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagereward: Learning and evaluating human preferences for text-to-image generation. arXiv preprint arXiv:2304.05977 (2023) 3, 4, 6, 7, 10", + "40. Xu, Y., Gong, M., Xie, S., Wei, W., Grundmann, M., Hou, T., et al.: Semi-implicit denoising diffusion models (siddms). arXiv preprint arXiv:2306.12511 (2023) 8", + "41. Xu, Y., Zhao, Y., Xiao, Z., Hou, T.: Ufogen: You forward once large scale text-to-image generation via diffusion gans. arXiv preprint arXiv:2311.09257 (2023) 8", + "42. Yang, B., Gu, S., Zhang, B., Zhang, T., Chen, X., Sun, X., Chen, D., Wen, F.: Paint by example: Exemplar-based image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18381-18391 (2023) 2", + "43. Yang, S., Chen, T., Zhou, M.: A dense reward view on aligning text-to-image diffusion with preference. arXiv preprint arXiv:2402.08265 (2024) 3, 4", + "44. Yildirim, A.B., Baday, V., Erdem, E., Erdem, A., Dundar, A.: Inst-inpaint: Instructing to remove objects with diffusion models. arXiv preprint arXiv:2304.03246 (2023) 2, 4", + "45. Yu, T., Feng, R., Feng, R., Liu, J., Jin, X., Zeng, W., Chen, Z.: Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790 (2023) 2, 4", + "46. Yuan, H., Chen, Z., Ji, K., Gu, Q.: Self-play fine-tuning of diffusion models for text-to-image generation. arXiv preprint arXiv:2402.10210 (2024) 3, 4", + "47. Zhang, M., Wu, J., Ren, Y., Li, M., Qin, J., Xiao, X., Liu, W., Wang, R., Zheng, M., Ma, A.J.: Diffusionengine: Diffusion model is scalable data engine for object detection. arXiv preprint arXiv:2309.03893 (2023) 2", + "48. Zhang, Z., Zhang, S., Zhan, Y., Luo, Y., Wen, Y., Tao, D.: Confronting reward overoptimization for diffusion models: A perspective of inductive and primacy biases. arXiv preprint arXiv:2402.08552 (2024) 3, 4" + ], + "bbox": [ + 212, + 146, + 787, + 686 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "ByteEdit", + "bbox": [ + 666, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/ByteEdit_ Boost, Comply and Accelerate Generative Image Editing/f2b7631d-9e11-4a5b-9762-122914fe2a69_model.json b/2024/ByteEdit_ Boost, Comply and Accelerate Generative Image Editing/f2b7631d-9e11-4a5b-9762-122914fe2a69_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f0e4deecc87bb04fd0b8a8013170dbc2787c18b0 --- /dev/null +++ b/2024/ByteEdit_ Boost, Comply and Accelerate Generative Image Editing/f2b7631d-9e11-4a5b-9762-122914fe2a69_model.json @@ -0,0 +1,2159 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.269, + 0.141, + 0.737, + 0.187 + ], + "angle": 0, + "content": "ByteEdit: Boost, Comply and Accelerate Generative Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.213, + 0.75, + 0.26 + ], + "angle": 0, + "content": "Yuxi Ren*, Jie Wu*†, Yanzuo Lu*, Huafeng Kuang, Xin Xia, Xionghui Wang, Qianqian Wang, Yixing Zhu, Pan Xie, Shiyin Wang, Xuefeng Xiao, Yitong Wang, Min Zheng, and Lean Fu" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.271, + 0.655, + 0.3 + ], + "angle": 0, + "content": "ByteDance \nProject Page: https://byte-edit.github.io" + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.343, + 0.741, + 0.702 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.715, + 0.789, + 0.788 + ], + "angle": 0, + "content": "Fig. 1: We introduce ByteEdit, a novel framework that utilizes feedback learning to enhance generative image editing tasks, resulting in outstanding generation performance, improved consistency, enhanced instruction adherence, and accelerated generation speed. To the best of our knowledge, ByteEdit emerges as the most superior and the fastest solution currently in the field of generative editing." + }, + { + "type": "page_footnote", + "bbox": [ + 0.239, + 0.825, + 0.538, + 0.841 + ], + "angle": 0, + "content": "*Equal contribution. †Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Y. Ren et al." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.148, + 0.745, + 0.425 + ], + "angle": 0, + "content": "Abstract. Recent advancements in diffusion-based generative image editing have sparked a profound revolution, reshaping the landscape of image outpainting and inpainting tasks. Despite these strides, the field grapples with inherent challenges, including: i) inferior quality; ii) poor consistency; iii) insufficient instruction adherence; iv) suboptimal generation efficiency. To address these obstacles, we present ByteEdit, an innovative feedback learning framework meticulously designed to Boost, Comply, and Accelerate Generative Image Editing tasks. ByteEdit seamlessly integrates image reward models dedicated to enhancing aesthetics and image-text alignment, while also introducing a dense, pixel-level reward model tailored to foster coherence in the output. Furthermore, we propose a pioneering adversarial and progressive feedback learning strategy to expedite the model's inference speed. Through extensive large-scale user evaluations, we demonstrate that ByteEdit surpasses leading generative image editing products, including Adobe, Canva, and MeiTu, in both generation quality and consistency. ByteEdit-Outpainting exhibits a remarkable enhancement of \\(388\\%\\) and \\(135\\%\\) in quality and consistency, respectively, when compared to the baseline model. Experiments also verified that our acceleration models maintain excellent performance results in terms of quality and consistency." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.441, + 0.661, + 0.456 + ], + "angle": 0, + "content": "Keywords: Outpainting \\(\\cdot\\) Intpainting \\(\\cdot\\) Feedback Learning" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.49, + 0.377, + 0.507 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.533, + 0.788, + 0.761 + ], + "angle": 0, + "content": "The field of generative image editing has experienced remarkable advancements in recent years [5, 6, 20, 21, 23, 34, 37, 38, 42, 44, 45], propelled by the development of diffusion models [12, 25, 27, 28, 47]. This progress has led to the emergence of influential products that have reshaped the landscape of image editing. A notable example is Adobe Firefly [1], which has revolutionized the creative process by enabling users to seamlessly incorporate, extend, or remove content from images through simple text prompts, thereby transcending the traditional boundaries of Photoshop. In our paper, we focus on the domain of generative image editing, with particular emphasis on two key aspects: 1) Outpainting, involving the expansion of the surrounding scene in an image based on provided input or even without explicit prompts, and 2) Inpainting, encompassing the random masking of specific image regions followed by the generation of corresponding content guided by given prompts (Inpainting-Editing) or erase certain objects (Inpainting-Erasing). Despite the notable advancements achieved through diffusion-based algorithms, several challenges persist within this field:" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.763, + 0.787, + 0.793 + ], + "angle": 0, + "content": "Inferior Quality: the quality of generated images frequently falls short in terms of realism, aesthetic appeal, and fidelity to minute details." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Insufficient Instruction Adherence: The existing models grapple with the arduous task of faithfully adhering to provided instructions, resulting in a lack of alignment between the generated image and the input text;" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ByteEdit" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "Poor Consistency: The generated regions exhibit an unsatisfactory level of coherence with the original image, manifesting as a deficiency in terms of color, style, texture, and other salient visual attributes;" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.784, + 0.237 + ], + "angle": 0, + "content": "Suboptimal Generation Efficiency: The generation process is characterized by sluggish speeds and inadequate efficiency, thereby imposing significant obstacles when confronted with large-scale image editing endeavors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.786, + 0.479 + ], + "angle": 0, + "content": "Recently, various efforts have been made to address the aforementioned challenges in the field. For instance, Imagen Editor [34] has employed an object detection approach to extract inpainting masks, while simultaneously capitalizing on original high-resolution images to faithfully capture intricate details. SmartBrush [37] has adopted a multi-task training strategy coupled with precision controls, encompassing both text and shape guidance, to enhance visual quality, mask controllability, and preserve the background. Additionally, RePaint [21] has utilized a pre-trained unconditional DDPM [12] prior and ingeniously modified the reverse diffusion iterations to generate high-quality and diverse output images. However, these approaches have primarily focused on addressing singular problems and have yet to achieve a more comprehensive solution. Large Language Models (LLMs) has made a notable surge in incorporating learning based on human feedback, and initial endeavors have been undertaken in the Text-to-Image (T2I) domain [9,17,39,43,46,48]. Inspired by these developments, we pose the question: Can we leverage human feedback to guide generative image editing to unleash the potential for superior generation outcomes?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.786, + 0.691 + ], + "angle": 0, + "content": "This paper introduces ByteEdit, an innovative framework for optimizing generative image editing through the incorporation of feedback learning. ByteEdit builds multiple reward models, namely the Aesthetic reward model, Alignment reward model, and Coherent reward model, to achieve exceptional generation effects, improved instruction adherence and enhanced consistency, respectively. These carefully designed reward models serve as the foundation for our proposed perceptual feedback learning (PeFL) mechanism, which provides task-aware and comprehensive supervision signals. Moreover, ByteEdit introduce an adversarial feedback learning strategy that employs the trainable reward model as the discriminator. This strategy ensures that the model benefits from the PeFL supervision and provide clear images even during high-noise stages, further improves both the performance and speed of our model. To expedite the sampling process, a progressive training strategy is employed to gradually reduce the optimization time steps and facilitate model inference in a low-steps regime." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.704, + 0.784, + 0.763 + ], + "angle": 0, + "content": "- New Insight: To the best of our knowledge, we offer the first attempt to incorporate human feedback into the field of generative image editing. By-Edit significantly enhances the overall performance of the model across various key aspects, opening new horizons in this field of study." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.765, + 0.784, + 0.809 + ], + "angle": 0, + "content": "- Comprehensive Framework: By designing complementary global-level and pixel-level reward models, we effectively guide the model towards achieving improved beauty, enhanced consistency, and superior image-text alignment." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.811, + 0.784, + 0.839 + ], + "angle": 0, + "content": "Efficiency and Pioneering: Progressive feedback and adversarial learning techniques are introduced to accomplish a remarkable acceleration in the" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.704, + 0.784, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "Y. Ren et al." + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.147, + 0.786, + 0.192 + ], + "angle": 0, + "content": "model's inference speed, all while maintaining a minimal compromise on output quality. Notably, ByteEdit stands as the first successful endeavor in accelerating generative editing models." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.193, + 0.787, + 0.253 + ], + "angle": 0, + "content": "Outstanding Performance: Extensive user studies show that ByteEdit exhibits obvious advantages in terms of quality, consistency, efficiency, and speed, compared to the most competitive products. ByteEdit emerges as the fastest and most superior solution currently available in image editing." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.278, + 0.388, + 0.294 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.788, + 0.599 + ], + "angle": 0, + "content": "Generative Image Editing. Generative Image Editing is a research area focused on filling user-specified regions of interest with desired contents. GLIDE [23] is the pioneering work that introduced text-to-image diffusion for editing purposes, and Repaint [21], on the other hand, conditions an unconditionally trained model (e.g. DDPM [12]) and leverages visible pixels to fill in missing areas. To enable more precise editing control, Blended Diffusion [5] incorporates multimodal embeddings and enforces similarity between images and text using CLIP [26]. SmartBrush [37] pushes the boundaries of mask generation by utilizing instance and panoptic segmentation datasets instead of random generation. Further improvements include the introduction of the powerful Segment Anything (SAM) [15] model by [45], which achieves mask-free removal, filling, and replacing of multiple pipelines. Inst-Inpaint [44] specializes in text-driven object removal without the need for explicit masking. Additionally, this method proposes the GQA-Inpaint dataset, which comprises image pairs with and without the object of interest, facilitating effective object removal guided by textual input. In addition to comparing our proposed method with existing academic approaches, we also benchmark against industry methods like Adobe [1], Canva [2], and MeiTu [3], providing a comprehensive evaluation across different domains and highlighting the strengths of our approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Human Feedback Learning. Foundation models for text-to-image diffusion often rely on pre-training with large-scale web datasets, such as LAION-5B [31], which may result in generated content that deviates from human ethical values and legal compliance requirements. Previous approaches [9,17] attempted to address this issue by constructing preference datasets using hand-crafted prompts or expert generators. However, these methods suffered from over-fitting due to their limited real-world scenarios and generator capabilities. To overcome these limitations, researchers proposed various reward models trained with expert annotations [35,39] or feedback from web users [13,16] to enforce alignment with human preferences. Drawing inspiration from reinforcement learning with human feedback (RLHF) utilized in natural language processing (NLP), researchers explored the application of human feedback learning in text-to-image diffusion [39, 43, 46, 48] to achieve more realistic, faithful, and ethical outcomes. Among these efforts, ImageReward [39] primarily focused on overall image quality and overlooked the complexity of human perception. In our work, we extend the concept of human feedback learning by introducing three fine-grained independ-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ByteEdit" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "dent reward models tailored for generative image editing: aesthetics, image-text alignment, and pixel-level coherence." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.204, + 0.649, + 0.223 + ], + "angle": 0, + "content": "3 ByteEdit: Boost, Comply and Accelerate" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.24, + 0.788, + 0.39 + ], + "angle": 0, + "content": "ByteEdit, focuses on generative image editing tasks that enable users to manipulate image content within a specific region of interest using textual descriptions. With an input image \\( x \\), a region-of-interest mask \\( m \\), and a user-provided textual description \\( c \\), our primary objective is to generate an output image \\( y \\) that preserves the unmasked region in the input image \\( x \\), while aligning the masked region well with both the description of \\( c \\) and visual attributes in \\( x \\). In this study, we introduce two key functionalities within ByteEdit: ByteEdit-Outpainting and ByteEdit-Inpainting. ByteEdit-Outpainting extends the image by generating content beyond the boundaries of the input image, while ByteEdit-Inpainting fills or erases in arbitrary areas of the input image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.392, + 0.788, + 0.468 + ], + "angle": 0, + "content": "The ByteEdit pipeline is presented in Fig 2, providing an overview of the system's workflow. In the subsequent subsections, we delve into the details of two crucial components: Boost (Sec. 3.1) and Comply (Sec. 3.2). Furthermore, we elaborate on the Accelerate scheme in Sec. 3.3, which illustrates an approach to expedite the processing time and improve the efficiency of the ByteEdit system." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.494, + 0.571, + 0.51 + ], + "angle": 0, + "content": "3.1 Boost: Perceptual Feedback Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.787, + 0.598 + ], + "angle": 0, + "content": "In the field of generative image editing, the persistent challenge of subpar quality has impelled us to propose a pioneering approach that introduces human feedback, hitherto unexplored in this domain. Our novel pipeline comprises three key components: feedback data collection, reward model training, and perceptual feedback learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Feedback Data Collection. We first randomly extract more than 1,500,000 text prompts from the Midjourney Discord [33] and MS-COCO Caption [7] datasets. To ensure the diversity, a clustering algorithm, namely K-Means, was employed, leveraging the similarities derived from state-of-the-art large language models [19]. Further, the features were visualized in lower dimensions using t-SNE [22], enabling the identification of data points possessing the largest average distance from their k-nearest neighbors. We also manually eliminate less informative and decorative-dominated prompts such as \"unbelievable\", \"fantastic\" and \"brilliant\" to improve the prompt quality. This meticulous procedure yielded approximately 400,000 candidate prompts, exhibiting diverse distributions, which were subsequently subjected to multiple text-to-image diffusion models, including SD1.5 [28] and SDXL [24]. Those images which excessively inferior quality or ambiguous characteristic are manually removed. Accompanying each prompt, a set of four generated images was presented to experts, who were tasked with selecting the best and worst images based on various aspects, encompassing aesthetic appeal, color composition, structural coherence, and brightness. The" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Y. Ren et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.789, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.332, + 0.788, + 0.361 + ], + "angle": 0, + "content": "Fig. 2: ByteEdit formulates a comprehensive feedback learning framework that facilitating aesthetics, image-text matching, consistency and inference speed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.392, + 0.789, + 0.437 + ], + "angle": 0, + "content": "resulting dataset, herein referred to as the aesthetic preference dataset \\(\\mathcal{D}_{aes}\\), encompasses a collection of massive triplets \\((c,x_p,x_n)\\), where \\(x_{p}\\) and \\(x_{n}\\) correspond to the preferred and non-preferred generated images of prompt \\(c\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.789, + 0.527 + ], + "angle": 0, + "content": "Reward Model Training. Building upon this dataset, we follow the training techniques in [39] to learn an aesthetic reward model \\( R_{\\alpha}(\\cdot) \\) of trainable parameters \\( \\alpha \\), which we briefly summarize here. The image and text features of the input are extracted from the BLIP [18] backbone, combined with cross attention, and fed into an MLP to obtain an aesthetic score. The training objective can be formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.306, + 0.542, + 0.787, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\alpha) = - \\mathbb {E} _ {(c, x _ {p}, x _ {n}) \\sim \\mathcal {D} _ {a e s}} [ \\log \\sigma (R _ {\\alpha} (c, x _ {p}) - R _ {\\alpha} (c, x _ {n})) ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.696, + 0.585 + ], + "angle": 0, + "content": "where \\(\\sigma (\\cdot)\\) represents the Sigmoid function used for normalization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.586, + 0.789, + 0.723 + ], + "angle": 0, + "content": "Perceptual Feedback Learning. Leveraging the power of our crafted reward model, we specifically introduce Perceptual Feedback Learning (PeFL) to fine-tune diffusion models with human feedback for generative image editing. Departing from the conventional practice of sequentially refining the predictions from the final step \\( x_{T} \\) to the initial step \\( x_0' \\) (\\( x_{T} \\to x_{T-1} \\to \\dots \\to x_0' \\)), we adopt an innovative perspective by performing optimization on the direct prediction outcomes \\( x_{t} \\to x_{0}' \\) at various intermediate steps \\( t \\in [1,T] \\) (\\( T = 20 \\) in this case). Through this dynamic exploration across different stages of denoising, we uncover the following distinctive observations:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.734, + 0.784, + 0.794 + ], + "angle": 0, + "content": "- During the initial stages of denoising (\\(t \\in [11, 20]\\)), the generative model (i.e. U-Net [29]) struggles to seamlessly complement the full image. Consequently, the reward model encounters challenges in accurately evaluating images that are hindered by obstacle masking." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.795, + 0.784, + 0.839 + ], + "angle": 0, + "content": "- As the denoising process progresses (\\(t \\in [1, 10]\\)), the reward model is able to identify and reward visually appealing images even in the presence of mild noise." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.734, + 0.784, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ByteEdit" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.465 + ], + "angle": 0, + "content": "Drawing upon these insightful observations, we present an innovative stagewise approach, to amplify the potential of generative image editing. Our proposed method encompasses the following key stages: 1) In stage 1 (\\(t \\in [16,20]\\)), we simply skip these steps with extremely heavy noise by diffusing the masked input \\(x \\odot (1 - m)\\) into noisy latents at a fixed step \\(T_{1} = 15\\). This strategy is motivated by the fact that the generative model's ability to fill in intricate details within the masked region is limited in this timestep, thus rendering the training overhead unnecessary. This is the central difference between our approach and the ReFL proposed in [39], whose training paradigm relies solely on prompts and starts inference from pure noise. We aim to bring about a more pronounced correlation between the output and input image, thus facilitating the preservation of details in the unmasked areas; 2) In stage 2 (\\(t \\in [t',15]\\)), we randomly select a denosing step \\(t' \\sim [1,T_2]\\) (\\(T_{2} = 10\\) in this case) and perform inference without gradient starting from the noisy latent, i.e. \\(x_{T_1} \\to x_{T_1-1} \\to \\dots \\to x_{t'}\\). This method ensures that the complemented image generated at this stage exhibits a sufficiently high level of quality, rendering it more amenable to evaluation by the reward model; 3) In stage 3 (\\(x_{t'} \\to x_0'\\)), the final stage entails the direct prediction of the complemented image \\(x_0'\\). We leverage the aesthetic score obtained from the reward model as invaluable human feedback to refine the generative model \\(G_{\\phi}(\\cdot)\\). This refinement process is achieved through the utilization of the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.261, + 0.476, + 0.786, + 0.493 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e w a r d}} (\\phi) = - \\mathbb {E} _ {(x, m, c) \\sim \\mathcal {D} _ {\\text {t r a i n}}, t ^ {\\prime} \\sim [ 1, T _ {2} ]} \\left[ \\log \\sigma \\left(R _ {\\alpha} (c, G _ {\\phi} (x, m, c, t ^ {\\prime}))\\right) \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.788, + 0.61 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}_{train}\\) represents the fine-tuning dataset (i.e. LAION-5B [31]). The term \\(G\\phi(x, m, c, t')\\) denotes the decoded output image \\(x_0'\\) generated by the generative model at step \\(t'\\), given the masked input \\(x \\odot (1 - m)\\) and the prompt \\(c\\). To further maintain the consistency and detail fidelity of the generated area and the original image area, we introduce pixel-level regularization (i.e., L1 loss) and a perceptual loss, which captures the discrepancy in VGG features [32]. Collectively, these regularization techniques can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.282, + 0.619, + 0.786, + 0.644 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {r e g} (\\phi) = \\mathbb {E} _ {(x, m, c) \\sim \\mathcal {D} _ {t r a i n}, t ^ {\\prime} \\sim [ 1, T _ {2} ]} \\| x - G _ {\\phi} (x, m, c, t ^ {\\prime}) \\| _ {1}, \\\\ \\mathcal {L} _ {r e g} (\\cdot) = \\mathbb {E} _ {\\| V (x) - V (C _ {r e g}, t ^ {\\prime}) \\|} \\end{array} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.282, + 0.639, + 0.723, + 0.655 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v g g} (\\phi) = \\mathbb {E} _ {(x, m, c) \\sim \\mathcal {D} _ {t r a i n}, t ^ {\\prime} \\sim [ 1, T _ {2} ]} \\| V (x) - V \\left(G _ {\\phi} (x, m, c, t ^ {\\prime})\\right) \\| _ {1},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.665, + 0.788, + 0.695 + ], + "angle": 0, + "content": "where \\( V(\\cdot) \\) represents the VGG network. The overall training objective of our PeFL can be summarized as," + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.707, + 0.786, + 0.724 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p e f l} (\\phi) = \\mathcal {L} _ {r e w a r d} + \\eta \\left(\\mathcal {L} _ {r e g} + \\mathcal {L} _ {v g g}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.734, + 0.616, + 0.75 + ], + "angle": 0, + "content": "where \\(\\eta\\) is a hyperparameter for balancing loss weights." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.77, + 0.66, + 0.787 + ], + "angle": 0, + "content": "3.2 Comply: Image-Text Alignment with Coherence" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Diverging from the text-to-image synthesis focus of [39], our method encompasses an additional emphasis on assessing the alignment between the generated content of the masked area and the user-specified prompt, as well as" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "Y. Ren et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "ensuring coherence with the unmasked region at the pixel level. To achieve this, we introduce two further components in this section, which complement the aesthetic reward model \\( R_{\\alpha} \\) proposed earlier." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.794, + 0.39 + ], + "angle": 0, + "content": "Image-Text Alignment. We note that the presence of numerous poorly matched image-text pairs within the LAION dataset [31]. Exploiting these pairs as non-preferred samples for training reward models allows us to reduce manual annotation costs significantly. Initially, we employ the CLIP model [26] to identify image-text pairs with lower CLIPScore [11] from the LAION dataset. Subsequently, we leverage advanced multi-modal large language models such as LLAVA [19] to generate more informative and descriptive captions for the input images. These captions are considered more accurate than the original prompts. This process yields approximately 40,000 triplets \\((c_p, c_n, x)\\) as alignment dataset \\(\\mathcal{D}_{\\text{align}}\\), where \\(c_p\\) and \\(c_n\\) correspond to the preferred and non-preferred textual descriptions of the image \\(x\\), respectively. These triplets are utilized for training the image-text alignment reward model, denoted as \\(R_{\\beta}(\\cdot)\\). The architecture of \\(R_{\\beta}\\) mirrors that of \\(R_{\\alpha}\\), while the training objective is similar to Eq. 1:" + }, + { + "type": "equation", + "bbox": [ + 0.302, + 0.399, + 0.786, + 0.418 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\beta) = - \\mathbb {E} _ {\\left(c _ {p}, c _ {n}, x\\right) \\sim \\mathcal {D} _ {\\text {a l i g n}}} \\left[ \\log \\sigma \\left(R _ {\\beta} \\left(c _ {p}, x\\right) - R _ {\\beta} \\left(c _ {n}, x\\right)\\right) \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.426, + 0.788, + 0.56 + ], + "angle": 0, + "content": "Pixel-Level Coherence. The issue of coherence arises from the presence of inconsistent content within and outside the regions of interest, characterized by subtle visual cues such as color variations, stylistic discrepancies, and textural differences. To tackle this challenge, a coherent reward model, denoted as \\( R_{\\gamma}(\\cdot) \\), is specifically designed for pixel-level discrimination, as opposed to the holistic evaluation performed by \\( R_{\\alpha}(\\cdot) \\) and \\( R_{\\beta}(\\cdot) \\). Our approach entails training a ViT-based [10] backbone network, followed by a prediction MLP head, to assess the authenticity and assign a score to each pixel in the input image. By formulating the loss function as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.26, + 0.572, + 0.786, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\gamma) = - \\mathbb {E} _ {(x, m, c) \\sim \\mathcal {D} _ {\\text {t r a i n}, t ^ {\\prime}} \\sim [ 1, T _ {2} ]} \\left[ \\log \\sigma \\left(R _ {\\gamma} (z)\\right) + \\log \\left(1 - \\sigma \\left(R _ {\\gamma} \\left(z ^ {\\prime}\\right)\\right)\\right) \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.785, + 0.63 + ], + "angle": 0, + "content": "where \\(z \\sim x \\in \\mathbb{R}^{H \\times W \\times 3}, z' \\sim G_{\\phi}(x, m, c, t') \\in \\mathbb{R}^{H \\times W \\times 3}\\) are pixels of the corresponding image and \\(H, W\\) represent the height and weight respectively." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.651, + 0.666, + 0.667 + ], + "angle": 0, + "content": "3.3 Accelerate: Adversarial and Progressive Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Adversarial training. Concurrent works such as UFOGen [41] and SDXL-Turbo [30] proposed to introduce adversarial training objective into fine-tuning diffusion models, which dramatically speeds up the sampling process and allows for one-step generation. They supposed that the Gaussian assumption of diffusion process does not hold anymore when the inference steps are extremely low, and therefore enabling the generative model to output samples in a single forward step by adversarial objective [36, 40]. We note that the functionality of our coherent reward model \\( R_{\\gamma}(\\cdot) \\) is very similar to that of the discriminator in adversarial training, except for the different granularity in prediction. To this end, unlike the aesthetic and alignment reward models, which necessitate offline learning prior to fine-tuning, the coherent reward model can be learned online" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "ByteEdit" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.224 + ], + "angle": 0, + "content": "and seamlessly integrated into the fine-tuning process. The adversarial objective of generator that raises the score of output image is also in compliance with our feedback learning in Eq. (2), we can simply achieve adversarial training by incorporating the optimization of \\( R_{\\gamma}(\\cdot) \\) into fine-tuning to serve as a discriminator. Thus the Eq. (2) can be reformulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.271, + 0.238, + 0.786, + 0.27 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e w a r d}} (\\phi) = - \\underset { \\begin{array}{c} (x, m, c) \\sim \\mathcal {D} _ {\\text {t r a i n}} \\\\ t ^ {\\prime} \\sim [ 1, T _ {2} ] \\end{array} } {\\mathbb {E}} \\sum_ {\\theta \\in \\{\\alpha , \\beta , \\gamma \\}} \\log \\sigma \\left(R _ {\\theta} \\left(c, G _ {\\phi} (x, m, c, t ^ {\\prime})\\right)\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.282, + 0.694, + 0.298 + ], + "angle": 0, + "content": "For completeness, we also rewrite the overall training objective as," + }, + { + "type": "equation", + "bbox": [ + 0.247, + 0.31, + 0.785, + 0.336 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {G} (\\phi) = \\mathcal {L} _ {p e f l} (\\phi) = \\mathcal {L} _ {r e w a r d} + \\eta \\left(\\mathcal {L} _ {r e g} + \\mathcal {L} _ {v g g}\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.247, + 0.331, + 0.737, + 0.348 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {D} (\\gamma) = - \\mathbb {E} _ {(x, m, c) \\sim \\mathcal {D} _ {t r a i n}, t ^ {\\prime} \\sim [ 1, T _ {2} ]} \\left[ \\log \\sigma \\left(R _ {\\gamma} (z)\\right) + \\log \\left(1 - \\sigma \\left(R _ {\\gamma} \\left(z ^ {\\prime}\\right)\\right)\\right) \\right].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.785, + 0.511 + ], + "angle": 0, + "content": "Progressive training. To expedite the sampling process, we employ a progressive training strategy where we gradually reduce the optimization time steps \\( T \\). Surprisingly, we find that the quality of the generated images does not significantly degrade under the supervisor of reward models. This approach strikes a fine balance between performance and speed, leading to compelling results. In our experiments, we adopt a two-phase progressive strategy. During phase 1, we set the optimization time steps as \\( T = 20 \\), \\( T_{1} = 15 \\), and \\( T_{2} = 10 \\). In phase 2, we further decrease the time steps to \\( T = 8 \\), \\( T_{1} = 6 \\), and \\( T_{2} = 3 \\). Notably, we achieve remarkable outcomes without employing any distillation operations, relying solely on the inheritance of model parameters." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.537, + 0.375, + 0.554 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.571, + 0.457, + 0.586 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Dataset. The fine-tuning dataset, denoted as \\( \\mathcal{D}_{train} \\), consists of a substantial collection of 7,562,283 images encompassing diverse domains such as real-world scenes, authentic portraits, and computer-generated (CG) images. To enhance the model's generalization ability and generation quality, we adopted a meticulous fine-grained masking strategy inspired by StableDiffusion [28]. Our masking strategy encompasses four distinct types of masks: global masks, irregular shape masks, square masks, and outward expansion masks. Each mask type corresponds to a specific probability value, which is randomly selected and applied to images during the training process. Moreover, we devised a specialized masking strategy tailored for Inpainting-Editing tasks. Leveraging instance-level data, which identifies specific objects within images, we introduced random dilation operations to coarsen the masks during training. These coarsened masks were then integrated with randomly generated masks surrounding the instances. This approach not only enhances the model's performance in instruction-based image editing tasks but also facilitates more accurate instance generation, ultimately leading to superior quality outputs." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "Y. Ren et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.329 + ], + "angle": 0, + "content": "To evaluate the performance of our approach, we conducted comprehensive qualitative and quantitative assessments using two distinct datasets. The first dataset, UserBench, was meticulously curated by gathering a vast amount of user-customized image-text matching data from various online sources. This dataset proved invaluable for evaluating image inpainting and outpainting tasks. From this extensive collection, we judiciously handpicked 100 high-quality image-text matching pairs to serve as our test data. We also leverage the experimental results from this dataset to collect and report human preferences. The second dataset, EditBench [34], presents a novel benchmark specifically tailored for text-guided image inpainting. Consisting of 240 images, each image within EditBench is paired with a corresponding mask that precisely delineates the region within the image to be modified through inpainting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.333, + 0.788, + 0.471 + ], + "angle": 0, + "content": "Training Setting. To facilitate the perceptual feedback learning stage, we employed a relatively small learning rate of 2e-06, complemented by a learning rate scheduling strategy that encompassed a warm-up phase consisting of 1000 iterations. Furthermore, to ensure stability in model parameter updates, we incorporated an exponential moving average (EMA) decay parameter set to 0.9999. Instead of employing \\(100\\%\\) noise as in ReFL [39], we introduced a 50The weight assigned to the perceptual feedback loss was set to 0.01. During the adversarial acceleration stage, we maintained similar settings to the perceptual feedback learning stage, with an additional adversarial loss weighted 0.05." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.515, + 0.546, + 0.53 + ], + "angle": 0, + "content": "4.2 Evaluation Principles and Criteria" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.791 + ], + "angle": 0, + "content": "Subjective metrics. To assess the robustness of our proposed method, we conducted a comprehensive subjective evaluation involving both expert evaluations and a large number of volunteer participants. Expert evaluators were tasked with individually assessing each generated image and assigning scores based on three key aspects: coherence, structure, and aesthetics. These aspects were rated on a scale of 1 to 5, with higher scores indicating superior generation quality: 1) Coherence focused on evaluating the consistency of texture, style, and color between the generated region and the original image. 2) Structure emphasized the clarity, sharpness, and absence of deformations or mutilations, particularly in human body parts. 3) Aesthetics gauged the overall level of creativity and diversity exhibited by the generated images. In addition to expert evaluations, we also sought the opinions of a large number of volunteers, with over 36,000 samples collected. These volunteers were presented with pairs of generated images and asked to choose between \"Good\", \"Same\", or \"Bad\", representing their preference in terms of GSB (Good-Same-Bad)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Objective metrics. In this study, we also incorporate objective text-image alignment metrics, specifically CLIPScore [11,26] and BLIPScore [18], to comprehensively evaluate the alignment of our models." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "ByteEdit" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.784, + 0.187 + ], + "angle": 0, + "content": "Table 1: Comparisons with state-of-the-art generative image editing systems in terms of coherence, structure and aesthetic scored by experts. More than 6000 image-text pairs are randomly sampled for each task and we report the average scores." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.199, + 0.773, + 0.284 + ], + "angle": 0, + "content": "
MethodOutpaintingInpainting-EditingInpainting-Erasing
coherencestructureaestheticcoherencestructureaestheticcoherencestructure
MeiTu [3]3.012.732.752.772.892.513.313.25
Canva [2]2.722.852.653.423.403.082.922.90
Adobe [1]3.523.073.143.463.603.223.854.28
ByteEdit3.543.253.263.733.393.253.994.03
" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.303, + 0.78, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.504 + ], + "angle": 0, + "content": "Fig. 3: Comparisons with state-of-the-art generative image editing systems in terms of human preference (i.e. GSB). More than 12,000 samples are collected for each task. For simplicity and to minimize the difficulty of collecting a large number of user opinions, we only offer the generated images by Adobe and our ByteEdit to the volunteers. \"Good\" indicates the generated images by our ByteEdit is preferred and vice versa." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.535, + 0.553, + 0.55 + ], + "angle": 0, + "content": "4.3 Comparisons with State of the arts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.688 + ], + "angle": 0, + "content": "We compare our method with concurrent state-of-the-art generative image editing systems such as Adobe [1], Canva [2] and MeiTu [3]. The comparisons cover three different tasks, including inpainting, inpainting-editing and inpainting-erasing. The inpainting editing will specify the content to be generated for the region of interest in the prompt. In contrast, inpainting-erasing requires the model to remove content within it and be as consistent as possible. Since the erased image has little change, experts were not asked to score aesthetics for user study." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "User study. The average scores evaluated by experts are shown in Tab. 1. From the results, our ByteEdit significantly outperforms the state-of-the-art across different metrics in the outpainting task. It demonstrates that our method works well to expand images based on existing content and maintains superior consistency, structural integrity, and creativity. As for the inpainting tasks, our method also can provide the most coherent edited or erased images. To further investigate the gap between Adobe and our proposed ByteEdit, we solicited feedback from a large number of volunteers on the images generated by both, and the results are illustrated in Fig. 3. The results show that users generally found the images we generated to be more natural in overall perception. Our GSB" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "Y. Ren et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.784, + 0.16 + ], + "angle": 0, + "content": "Table 2: The quantitative results of ByteEdit and recent state-of-the-art approaches." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.172, + 0.787, + 0.236 + ], + "angle": 0, + "content": "
MetricsUserBenchEditBench
Meitu [3]Canva [2]Adobe [1]ByteEditDiffEdit [8]BLD [4]EMILIE [14]ByteEdit
CLIPScore0.2350.2410.2370.2550.2720.2800.3110.329
BLIPScore0.1740.4670.4500.6870.5820.5960.6200.691
" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.264, + 0.758, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.433, + 0.782, + 0.447 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparison in inpainting. We highlight key areas with red boxes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.785, + 0.524 + ], + "angle": 0, + "content": "superiority percentages (i.e. \\((\\mathrm{G} + \\mathrm{S}) / (\\mathrm{S} + \\mathrm{B}) * 100\\%\\) ) on three different tasks are \\(105\\%\\), \\(163\\%\\), and \\(112\\%\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.788, + 0.728 + ], + "angle": 0, + "content": "Quantitative comparison. To quantitatively evaluate the performance of our method compared to other approaches, we conduct a quantitative evaluation of how well the edited image can capture the semantics of the edit instruction successfully by measuring the CLIPScore and BLIPScore. We conduct the experiment in inpainting-editing task and the results are provided in Table 2. From the UserBench against state-of-the-art generative image editing systems, we noticed that the score results are not completely consistent with user opinion. Nevertheless, our method is still ahead of the second-place Canva by \\(5.8\\% (+0.014)\\) and \\(47.1\\% (+0.22)\\) in terms of CLIPScore and BLIPScore, respectively. As for the EditBench, we follow [14] to compare our method with several concurrent editing approaches, i.e. DiffEdit [8], BLD [4] and EMILIE [14]. It demonstrates that the ByteEdit consistently yields the state-of-the-art performance, which shows our superior quality, consistency and instruction adherence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Qualitative comparison. In Figure 4 and 5, we visualize samples produced by different systems under different tasks. It clearly shows that our method exhibits a superior performance for learning both coherence and aesthetic. For the inpainting task, the ByteEdit consistently follows the user-specified instructions and generates coherent images with better image-text alignment. It is worth noting that our system allows both prompt and prompt-free generation when it comes to outpainting, which has broader application scenarios in reality." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.669, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "ByteEdit" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.147, + 0.761, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.338, + 0.785, + 0.352 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparison in outpainting. We highlight key areas with red boxes." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.368, + 0.499, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.525, + 0.421, + 0.536 + ], + "angle": 0, + "content": "(a) w/ vs w/o PeFL" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.367, + 0.783, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.525, + 0.724, + 0.536 + ], + "angle": 0, + "content": "(b) w/ vs w/o Acceleration" + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.548, + 0.763, + 0.563 + ], + "angle": 0, + "content": "Fig. 6: Human Perference Evaluation on our proposed PeFL and Acceleration." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.591, + 0.401, + 0.604 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.719 + ], + "angle": 0, + "content": "In Figure 6, we conduct ablation experiments on both our proposed PeFL and acceleration strategy. The experts were asked to choose GSB preference and we report the human preference rates in the figure, which are calculated as \\((\\mathrm{G} + \\mathrm{S}) / [((\\mathrm{G} + \\mathrm{S}) + (\\mathrm{S} + \\mathrm{B}))] * 100\\%\\) for win and \\((\\mathrm{S} + \\mathrm{B}) / [((\\mathrm{G} + \\mathrm{S}) + (\\mathrm{S} + \\mathrm{B}))] * 100\\%\\) for lose, respectively. The evaluation is similar to the user study, except that we combine structure and aesthetics to reduce evaluation costs. More visualizations are also included in Figure 7." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.81 + ], + "angle": 0, + "content": "PeFL preference. From the results in Figure 6(a), our proposed PeFL significantly improves the generation quality, outperforming the baseline on all different tasks. Especially in the outpainting task with PeFL, our method exceeds the baseline by about \\(60\\%\\) in terms of structure and aesthetic, which is consistent with the edited image shown at the top of Figure 7 that is more realistic and conforms to the rules of realistic physics." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Acceleration preference. In Figure 6(b), we demonstrate that our model has no significant loss in either consistency or structure and aesthetic with the pro" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Y. Ren et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.147, + 0.782, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.37, + 0.331, + 0.631, + 0.346 + ], + "angle": 0, + "content": "Fig. 7: Ablation Studies Visualization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.788, + 0.542 + ], + "angle": 0, + "content": "gressive training strategy. To our surprise, we have even achieved both increasing speed and quality in the outpainting and inpainting-editing tasks. Based on our experimental observations, this phenomenon can be attributed to two underlying factors: i) Stable Training: By considering the discriminator as a reward model, trainable reward model offers flexible and robust supervision for PeFL, alleviating issues related to model over-optimization; ii) Expand Supervision Scope: The incorporation of adversarial supervision enables us to extend the time step of PEFL optimization. Consequently, even at high-noise stages, such as step 999, the model can still benefit from PeFL supervision, further driving improvements in model performance. The visualization at the bottom of Figure 7 also verifies this, where the outputs become more realistic and natural after acceleration." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.563, + 0.355, + 0.579 + ], + "angle": 0, + "content": "5 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.593, + 0.785, + 0.625 + ], + "angle": 0, + "content": "ByteEdit has demonstrated remarkable performance across various editing tasks. However, several promising directions warrant further exploration:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.634, + 0.787, + 0.693 + ], + "angle": 0, + "content": "- Performance: One avenue for improvement lies in developing more targeted reward models tailored to specific editing tasks. By refining the reward models, we can potentially unlock even higher levels of performance and generate more precise and desirable output." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.694, + 0.787, + 0.739 + ], + "angle": 0, + "content": "- Acceleration: Another area of interest is investigating how ByteEdit can be further integrated with advanced techniques such as LCM and SDXL-turbo to achieve accelerated processing speeds." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.739, + 0.787, + 0.77 + ], + "angle": 0, + "content": "- Task: Expanding the capabilities of ByteEdit beyond image editing to domains like video editing or instruction editing holds significant potential." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.634, + 0.787, + 0.77 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "By incorporating human feedback to optimize generative image editing, ByteEdit can greatly enhance the practicality and usability in real-world scenarios. We hope that our work will provide valuable insights and inspire deeper reflections in this field, propelling its future development." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "ByteEdit" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.174, + 0.785, + 0.2 + ], + "angle": 0, + "content": "1. Adobe firefly - free generative ai for creatives. https://www.adobe.com/products/firefly.html 2, 4, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.201, + 0.785, + 0.227 + ], + "angle": 0, + "content": "2. Free ai image generator: Online text to image app | canva. https://www.canva.com/ai-image-generator/ 4, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.228, + 0.615, + 0.24 + ], + "angle": 0, + "content": "3. Miraclevision. https://ai.meitu.com/index/ 4,11,12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.241, + 0.785, + 0.268 + ], + "angle": 0, + "content": "4. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. ACM Transactions on Graphics (TOG) 42(4), 1-11 (2023) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.268, + 0.787, + 0.309 + ], + "angle": 0, + "content": "5. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18208-18218 (2022) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.785, + 0.336 + ], + "angle": 0, + "content": "6. Chen, X., Huang, L., Liu, Y., Shen, Y., Zhao, D., Zhao, H.: Anydoor: Zero-shot object-level image customization. arXiv preprint arXiv:2307.09481 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.337, + 0.785, + 0.377 + ], + "angle": 0, + "content": "7. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dollar, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.416 + ], + "angle": 0, + "content": "8. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427 (2022) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.418, + 0.785, + 0.459 + ], + "angle": 0, + "content": "9. Dong, H., Xiong, W., Goyal, D., Pan, R., Diao, S., Zhang, J., Shum, K., Zhang, T.: Raft: Reward ranked finetuning for generative foundation model alignment. arXiv preprint arXiv:2304.06767 (2023) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.459, + 0.785, + 0.513 + ], + "angle": 0, + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.514, + 0.785, + 0.553 + ], + "angle": 0, + "content": "11. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021) 8, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.554, + 0.785, + 0.581 + ], + "angle": 0, + "content": "12. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020) 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.582, + 0.785, + 0.622 + ], + "angle": 0, + "content": "13. Isajanyan, A., Shatveryan, A., Kocharyan, D., Wang, Z., Shi, H.: Social reward: Evaluating and enhancing generative ai through million-user feedback from an online creative community. arXiv preprint arXiv:2402.09872 (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.622, + 0.785, + 0.676 + ], + "angle": 0, + "content": "14. Joseph, K., Udhayanan, P., Shukla, T., Agarwal, A., Karanam, S., Goswami, K., Srinivasan, B.V.: Iterative multi-granular image editing using diffusion models. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 8107-8116 (2024) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.677, + 0.785, + 0.717 + ], + "angle": 0, + "content": "15. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.718, + 0.785, + 0.758 + ], + "angle": 0, + "content": "16. Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., Levy, O.: Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems 36 (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "17. Lee, K., Liu, H., Ryu, M., Watkins, O., Du, Y., Boutilier, C., Abbeel, P., Ghavamzadeh, M., Gu, S.S.: Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192 (2023) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "18. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022) 6, 10" + }, + { + "type": "list", + "bbox": [ + 0.219, + 0.174, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.36, + 0.127 + ], + "angle": 0, + "content": "Y. Ren et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "19. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023) 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.202 + ], + "angle": 0, + "content": "20. Lu, Y., Zhang, M., Ma, A.J., Xie, X., Lai, J.H.: Coarse-to-fine latent diffusion for pose-guided person image synthesis. arXiv preprint arXiv:2402.18078 (2024) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.257 + ], + "angle": 0, + "content": "21. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11461-11471 (2022) 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.283 + ], + "angle": 0, + "content": "22. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.283, + 0.785, + 0.336 + ], + "angle": 0, + "content": "23. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.337, + 0.785, + 0.378 + ], + "angle": 0, + "content": "24. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.418 + ], + "angle": 0, + "content": "25. Qin, J., Wu, J., Chen, W., Ren, Y., Li, H., Wu, H., Xiao, X., Wang, R., Wen, S.: Diffusionpt: Llm-driven text-to-image generation system. arXiv preprint arXiv:2401.10061 (2024) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.419, + 0.785, + 0.473 + ], + "angle": 0, + "content": "26. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 4, 8, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.528 + ], + "angle": 0, + "content": "27. Ren, Y., Wu, J., Zhang, P., Zhang, M., Xiao, X., He, Q., Wang, R., Zheng, M., Pan, X.: Ugc: Unified gan compression for efficient image-to-image translation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17281-17291 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.785, + 0.58 + ], + "angle": 0, + "content": "28. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) 2, 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.785, + 0.637 + ], + "angle": 0, + "content": "29. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.663 + ], + "angle": 0, + "content": "30. Sauer, A., Lorenz, D., Blattmann, A., Rombach, R.: Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042 (2023) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.663, + 0.785, + 0.718 + ], + "angle": 0, + "content": "31. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022) 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.745 + ], + "angle": 0, + "content": "32. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.785, + 0.771 + ], + "angle": 0, + "content": "33. Turc, I., Nemade, G.: Midjourney user prompts & generated images (250k) (2022). https://doi.org/10.34740/KAGGLE/DS/2349267 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.785, + 0.84 + ], + "angle": 0, + "content": "34. Wang, S., Sahara, C., Montgomery, C., Pont-Tuset, J., Noy, S., Pellegrini, S., Onoe, Y., Laszlo, S., Fleet, D.J., Soricut, R., et al.: Imagen editor and editbench: Advancing and evaluating text-guided image inpainting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18359-18369 (2023) 2, 3, 10" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.668, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "ByteEdit" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "35. Wu, X., Sun, K., Zhu, F., Zhao, R., Li, H.: Human preference score: Better aligning text-to-image models with human preference. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2096-2105 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "36. Xiao, Z., Kreis, K., Vahdat, A.: Tackling the generative learning trilemma with denoising diffusion gans. arXiv preprint arXiv:2112.07804 (2021) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.787, + 0.272 + ], + "angle": 0, + "content": "37. Xie, S., Zhang, Z., Lin, Z., Hinz, T., Zhang, K.: Smartbrush: Text and shape guided object inpainting with diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22428-22437 (2023) 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "38. Xie, S., Zhao, Y., Xiao, Z., Chan, K.C., Li, Y., Xu, Y., Zhang, K., Hou, T.: Dreaminpainter: Text-guided subject-driven image inpainting with diffusion models. arXiv preprint arXiv:2312.03771 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.315, + 0.787, + 0.357 + ], + "angle": 0, + "content": "39. Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagereward: Learning and evaluating human preferences for text-to-image generation. arXiv preprint arXiv:2304.05977 (2023) 3, 4, 6, 7, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.357, + 0.787, + 0.384 + ], + "angle": 0, + "content": "40. Xu, Y., Gong, M., Xie, S., Wei, W., Grundmann, M., Hou, T., et al.: Semi-implicit denoising diffusion models (siddms). arXiv preprint arXiv:2306.12511 (2023) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.384, + 0.787, + 0.411 + ], + "angle": 0, + "content": "41. Xu, Y., Zhao, Y., Xiao, Z., Hou, T.: Ufogen: You forward once large scale text-to-image generation via diffusion gans. arXiv preprint arXiv:2311.09257 (2023) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.411, + 0.787, + 0.466 + ], + "angle": 0, + "content": "42. Yang, B., Gu, S., Zhang, B., Zhang, T., Chen, X., Sun, X., Chen, D., Wen, F.: Paint by example: Exemplar-based image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18381-18391 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.467, + 0.787, + 0.494 + ], + "angle": 0, + "content": "43. Yang, S., Chen, T., Zhou, M.: A dense reward view on aligning text-to-image diffusion with preference. arXiv preprint arXiv:2402.08265 (2024) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.495, + 0.787, + 0.535 + ], + "angle": 0, + "content": "44. Yildirim, A.B., Baday, V., Erdem, E., Erdem, A., Dundar, A.: Inst-inpaint: Instructing to remove objects with diffusion models. arXiv preprint arXiv:2304.03246 (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.536, + 0.787, + 0.576 + ], + "angle": 0, + "content": "45. Yu, T., Feng, R., Feng, R., Liu, J., Jin, X., Zeng, W., Chen, Z.: Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790 (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.577, + 0.787, + 0.605 + ], + "angle": 0, + "content": "46. Yuan, H., Chen, Z., Ji, K., Gu, Q.: Self-play fine-tuning of diffusion models for text-to-image generation. arXiv preprint arXiv:2402.10210 (2024) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "47. Zhang, M., Wu, J., Ren, Y., Li, M., Qin, J., Xiao, X., Liu, W., Wang, R., Zheng, M., Ma, A.J.: Diffusionengine: Diffusion model is scalable data engine for object detection. arXiv preprint arXiv:2309.03893 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "48. Zhang, Z., Zhang, S., Zhan, Y., Luo, Y., Wen, Y., Tao, D.: Confronting reward overoptimization for diffusion models: A perspective of inductive and primacy biases. arXiv preprint arXiv:2402.08552 (2024) 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.688 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/ByteEdit_ Boost, Comply and Accelerate Generative Image Editing/f2b7631d-9e11-4a5b-9762-122914fe2a69_origin.pdf b/2024/ByteEdit_ Boost, Comply and Accelerate Generative Image Editing/f2b7631d-9e11-4a5b-9762-122914fe2a69_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e257626f2da0f2f121050f473bbcf612619be5a4 --- /dev/null +++ b/2024/ByteEdit_ Boost, Comply and Accelerate Generative Image Editing/f2b7631d-9e11-4a5b-9762-122914fe2a69_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13aa9d75d2a674ab85530ed034e4d0b939a0c2b909553d9e2e0cf0422cfb9c38 +size 11928106 diff --git a/2024/ByteEdit_ Boost, Comply and Accelerate Generative Image Editing/full.md b/2024/ByteEdit_ Boost, Comply and Accelerate Generative Image Editing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e68d63a94eb16f6bc03489a0257eaf9863868da0 --- /dev/null +++ b/2024/ByteEdit_ Boost, Comply and Accelerate Generative Image Editing/full.md @@ -0,0 +1,269 @@ +# ByteEdit: Boost, Comply and Accelerate Generative Image Editing + +Yuxi Ren*, Jie Wu*†, Yanzuo Lu*, Huafeng Kuang, Xin Xia, Xionghui Wang, Qianqian Wang, Yixing Zhu, Pan Xie, Shiyin Wang, Xuefeng Xiao, Yitong Wang, Min Zheng, and Lean Fu + +ByteDance +Project Page: https://byte-edit.github.io + +![](images/06b8ed23b91324ec96d0ab590a7a942aae563d5ec4e5f443e4a4d2eeb5a21559.jpg) +Fig. 1: We introduce ByteEdit, a novel framework that utilizes feedback learning to enhance generative image editing tasks, resulting in outstanding generation performance, improved consistency, enhanced instruction adherence, and accelerated generation speed. To the best of our knowledge, ByteEdit emerges as the most superior and the fastest solution currently in the field of generative editing. + +Abstract. Recent advancements in diffusion-based generative image editing have sparked a profound revolution, reshaping the landscape of image outpainting and inpainting tasks. Despite these strides, the field grapples with inherent challenges, including: i) inferior quality; ii) poor consistency; iii) insufficient instruction adherence; iv) suboptimal generation efficiency. To address these obstacles, we present ByteEdit, an innovative feedback learning framework meticulously designed to Boost, Comply, and Accelerate Generative Image Editing tasks. ByteEdit seamlessly integrates image reward models dedicated to enhancing aesthetics and image-text alignment, while also introducing a dense, pixel-level reward model tailored to foster coherence in the output. Furthermore, we propose a pioneering adversarial and progressive feedback learning strategy to expedite the model's inference speed. Through extensive large-scale user evaluations, we demonstrate that ByteEdit surpasses leading generative image editing products, including Adobe, Canva, and MeiTu, in both generation quality and consistency. ByteEdit-Outpainting exhibits a remarkable enhancement of $388\%$ and $135\%$ in quality and consistency, respectively, when compared to the baseline model. Experiments also verified that our acceleration models maintain excellent performance results in terms of quality and consistency. + +Keywords: Outpainting $\cdot$ Intpainting $\cdot$ Feedback Learning + +# 1 Introduction + +The field of generative image editing has experienced remarkable advancements in recent years [5, 6, 20, 21, 23, 34, 37, 38, 42, 44, 45], propelled by the development of diffusion models [12, 25, 27, 28, 47]. This progress has led to the emergence of influential products that have reshaped the landscape of image editing. A notable example is Adobe Firefly [1], which has revolutionized the creative process by enabling users to seamlessly incorporate, extend, or remove content from images through simple text prompts, thereby transcending the traditional boundaries of Photoshop. In our paper, we focus on the domain of generative image editing, with particular emphasis on two key aspects: 1) Outpainting, involving the expansion of the surrounding scene in an image based on provided input or even without explicit prompts, and 2) Inpainting, encompassing the random masking of specific image regions followed by the generation of corresponding content guided by given prompts (Inpainting-Editing) or erase certain objects (Inpainting-Erasing). Despite the notable advancements achieved through diffusion-based algorithms, several challenges persist within this field: + +Inferior Quality: the quality of generated images frequently falls short in terms of realism, aesthetic appeal, and fidelity to minute details. + +Insufficient Instruction Adherence: The existing models grapple with the arduous task of faithfully adhering to provided instructions, resulting in a lack of alignment between the generated image and the input text; + +Poor Consistency: The generated regions exhibit an unsatisfactory level of coherence with the original image, manifesting as a deficiency in terms of color, style, texture, and other salient visual attributes; + +Suboptimal Generation Efficiency: The generation process is characterized by sluggish speeds and inadequate efficiency, thereby imposing significant obstacles when confronted with large-scale image editing endeavors. + +Recently, various efforts have been made to address the aforementioned challenges in the field. For instance, Imagen Editor [34] has employed an object detection approach to extract inpainting masks, while simultaneously capitalizing on original high-resolution images to faithfully capture intricate details. SmartBrush [37] has adopted a multi-task training strategy coupled with precision controls, encompassing both text and shape guidance, to enhance visual quality, mask controllability, and preserve the background. Additionally, RePaint [21] has utilized a pre-trained unconditional DDPM [12] prior and ingeniously modified the reverse diffusion iterations to generate high-quality and diverse output images. However, these approaches have primarily focused on addressing singular problems and have yet to achieve a more comprehensive solution. Large Language Models (LLMs) has made a notable surge in incorporating learning based on human feedback, and initial endeavors have been undertaken in the Text-to-Image (T2I) domain [9,17,39,43,46,48]. Inspired by these developments, we pose the question: Can we leverage human feedback to guide generative image editing to unleash the potential for superior generation outcomes? + +This paper introduces ByteEdit, an innovative framework for optimizing generative image editing through the incorporation of feedback learning. ByteEdit builds multiple reward models, namely the Aesthetic reward model, Alignment reward model, and Coherent reward model, to achieve exceptional generation effects, improved instruction adherence and enhanced consistency, respectively. These carefully designed reward models serve as the foundation for our proposed perceptual feedback learning (PeFL) mechanism, which provides task-aware and comprehensive supervision signals. Moreover, ByteEdit introduce an adversarial feedback learning strategy that employs the trainable reward model as the discriminator. This strategy ensures that the model benefits from the PeFL supervision and provide clear images even during high-noise stages, further improves both the performance and speed of our model. To expedite the sampling process, a progressive training strategy is employed to gradually reduce the optimization time steps and facilitate model inference in a low-steps regime. + +- New Insight: To the best of our knowledge, we offer the first attempt to incorporate human feedback into the field of generative image editing. By-Edit significantly enhances the overall performance of the model across various key aspects, opening new horizons in this field of study. +- Comprehensive Framework: By designing complementary global-level and pixel-level reward models, we effectively guide the model towards achieving improved beauty, enhanced consistency, and superior image-text alignment. +Efficiency and Pioneering: Progressive feedback and adversarial learning techniques are introduced to accomplish a remarkable acceleration in the + +model's inference speed, all while maintaining a minimal compromise on output quality. Notably, ByteEdit stands as the first successful endeavor in accelerating generative editing models. + +Outstanding Performance: Extensive user studies show that ByteEdit exhibits obvious advantages in terms of quality, consistency, efficiency, and speed, compared to the most competitive products. ByteEdit emerges as the fastest and most superior solution currently available in image editing. + +# 2 Related Work + +Generative Image Editing. Generative Image Editing is a research area focused on filling user-specified regions of interest with desired contents. GLIDE [23] is the pioneering work that introduced text-to-image diffusion for editing purposes, and Repaint [21], on the other hand, conditions an unconditionally trained model (e.g. DDPM [12]) and leverages visible pixels to fill in missing areas. To enable more precise editing control, Blended Diffusion [5] incorporates multimodal embeddings and enforces similarity between images and text using CLIP [26]. SmartBrush [37] pushes the boundaries of mask generation by utilizing instance and panoptic segmentation datasets instead of random generation. Further improvements include the introduction of the powerful Segment Anything (SAM) [15] model by [45], which achieves mask-free removal, filling, and replacing of multiple pipelines. Inst-Inpaint [44] specializes in text-driven object removal without the need for explicit masking. Additionally, this method proposes the GQA-Inpaint dataset, which comprises image pairs with and without the object of interest, facilitating effective object removal guided by textual input. In addition to comparing our proposed method with existing academic approaches, we also benchmark against industry methods like Adobe [1], Canva [2], and MeiTu [3], providing a comprehensive evaluation across different domains and highlighting the strengths of our approach. + +Human Feedback Learning. Foundation models for text-to-image diffusion often rely on pre-training with large-scale web datasets, such as LAION-5B [31], which may result in generated content that deviates from human ethical values and legal compliance requirements. Previous approaches [9,17] attempted to address this issue by constructing preference datasets using hand-crafted prompts or expert generators. However, these methods suffered from over-fitting due to their limited real-world scenarios and generator capabilities. To overcome these limitations, researchers proposed various reward models trained with expert annotations [35,39] or feedback from web users [13,16] to enforce alignment with human preferences. Drawing inspiration from reinforcement learning with human feedback (RLHF) utilized in natural language processing (NLP), researchers explored the application of human feedback learning in text-to-image diffusion [39, 43, 46, 48] to achieve more realistic, faithful, and ethical outcomes. Among these efforts, ImageReward [39] primarily focused on overall image quality and overlooked the complexity of human perception. In our work, we extend the concept of human feedback learning by introducing three fine-grained independ- + +dent reward models tailored for generative image editing: aesthetics, image-text alignment, and pixel-level coherence. + +# 3 ByteEdit: Boost, Comply and Accelerate + +ByteEdit, focuses on generative image editing tasks that enable users to manipulate image content within a specific region of interest using textual descriptions. With an input image $x$ , a region-of-interest mask $m$ , and a user-provided textual description $c$ , our primary objective is to generate an output image $y$ that preserves the unmasked region in the input image $x$ , while aligning the masked region well with both the description of $c$ and visual attributes in $x$ . In this study, we introduce two key functionalities within ByteEdit: ByteEdit-Outpainting and ByteEdit-Inpainting. ByteEdit-Outpainting extends the image by generating content beyond the boundaries of the input image, while ByteEdit-Inpainting fills or erases in arbitrary areas of the input image. + +The ByteEdit pipeline is presented in Fig 2, providing an overview of the system's workflow. In the subsequent subsections, we delve into the details of two crucial components: Boost (Sec. 3.1) and Comply (Sec. 3.2). Furthermore, we elaborate on the Accelerate scheme in Sec. 3.3, which illustrates an approach to expedite the processing time and improve the efficiency of the ByteEdit system. + +# 3.1 Boost: Perceptual Feedback Learning + +In the field of generative image editing, the persistent challenge of subpar quality has impelled us to propose a pioneering approach that introduces human feedback, hitherto unexplored in this domain. Our novel pipeline comprises three key components: feedback data collection, reward model training, and perceptual feedback learning. + +Feedback Data Collection. We first randomly extract more than 1,500,000 text prompts from the Midjourney Discord [33] and MS-COCO Caption [7] datasets. To ensure the diversity, a clustering algorithm, namely K-Means, was employed, leveraging the similarities derived from state-of-the-art large language models [19]. Further, the features were visualized in lower dimensions using t-SNE [22], enabling the identification of data points possessing the largest average distance from their k-nearest neighbors. We also manually eliminate less informative and decorative-dominated prompts such as "unbelievable", "fantastic" and "brilliant" to improve the prompt quality. This meticulous procedure yielded approximately 400,000 candidate prompts, exhibiting diverse distributions, which were subsequently subjected to multiple text-to-image diffusion models, including SD1.5 [28] and SDXL [24]. Those images which excessively inferior quality or ambiguous characteristic are manually removed. Accompanying each prompt, a set of four generated images was presented to experts, who were tasked with selecting the best and worst images based on various aspects, encompassing aesthetic appeal, color composition, structural coherence, and brightness. The + +![](images/9aed3886d29a05e2f603ec6aa56f481fdab35e7cb99183c0e2843c78d338734b.jpg) +Fig. 2: ByteEdit formulates a comprehensive feedback learning framework that facilitating aesthetics, image-text matching, consistency and inference speed. + +resulting dataset, herein referred to as the aesthetic preference dataset $\mathcal{D}_{aes}$ , encompasses a collection of massive triplets $(c,x_p,x_n)$ , where $x_{p}$ and $x_{n}$ correspond to the preferred and non-preferred generated images of prompt $c$ , respectively. + +Reward Model Training. Building upon this dataset, we follow the training techniques in [39] to learn an aesthetic reward model $R_{\alpha}(\cdot)$ of trainable parameters $\alpha$ , which we briefly summarize here. The image and text features of the input are extracted from the BLIP [18] backbone, combined with cross attention, and fed into an MLP to obtain an aesthetic score. The training objective can be formulated as, + +$$ +\mathcal {L} (\alpha) = - \mathbb {E} _ {(c, x _ {p}, x _ {n}) \sim \mathcal {D} _ {a e s}} [ \log \sigma (R _ {\alpha} (c, x _ {p}) - R _ {\alpha} (c, x _ {n})) ], \tag {1} +$$ + +where $\sigma (\cdot)$ represents the Sigmoid function used for normalization. + +Perceptual Feedback Learning. Leveraging the power of our crafted reward model, we specifically introduce Perceptual Feedback Learning (PeFL) to fine-tune diffusion models with human feedback for generative image editing. Departing from the conventional practice of sequentially refining the predictions from the final step $x_{T}$ to the initial step $x_0'$ ( $x_{T} \to x_{T-1} \to \dots \to x_0'$ ), we adopt an innovative perspective by performing optimization on the direct prediction outcomes $x_{t} \to x_{0}'$ at various intermediate steps $t \in [1,T]$ ( $T = 20$ in this case). Through this dynamic exploration across different stages of denoising, we uncover the following distinctive observations: + +- During the initial stages of denoising ( $t \in [11, 20]$ ), the generative model (i.e. U-Net [29]) struggles to seamlessly complement the full image. Consequently, the reward model encounters challenges in accurately evaluating images that are hindered by obstacle masking. +- As the denoising process progresses ( $t \in [1, 10]$ ), the reward model is able to identify and reward visually appealing images even in the presence of mild noise. + +Drawing upon these insightful observations, we present an innovative stagewise approach, to amplify the potential of generative image editing. Our proposed method encompasses the following key stages: 1) In stage 1 ( $t \in [16,20]$ ), we simply skip these steps with extremely heavy noise by diffusing the masked input $x \odot (1 - m)$ into noisy latents at a fixed step $T_{1} = 15$ . This strategy is motivated by the fact that the generative model's ability to fill in intricate details within the masked region is limited in this timestep, thus rendering the training overhead unnecessary. This is the central difference between our approach and the ReFL proposed in [39], whose training paradigm relies solely on prompts and starts inference from pure noise. We aim to bring about a more pronounced correlation between the output and input image, thus facilitating the preservation of details in the unmasked areas; 2) In stage 2 ( $t \in [t',15]$ ), we randomly select a denosing step $t' \sim [1,T_2]$ ( $T_{2} = 10$ in this case) and perform inference without gradient starting from the noisy latent, i.e. $x_{T_1} \to x_{T_1-1} \to \dots \to x_{t'}$ . This method ensures that the complemented image generated at this stage exhibits a sufficiently high level of quality, rendering it more amenable to evaluation by the reward model; 3) In stage 3 ( $x_{t'} \to x_0'$ ), the final stage entails the direct prediction of the complemented image $x_0'$ . We leverage the aesthetic score obtained from the reward model as invaluable human feedback to refine the generative model $G_{\phi}(\cdot)$ . This refinement process is achieved through the utilization of the following loss function: + +$$ +\mathcal {L} _ {\text {r e w a r d}} (\phi) = - \mathbb {E} _ {(x, m, c) \sim \mathcal {D} _ {\text {t r a i n}}, t ^ {\prime} \sim [ 1, T _ {2} ]} \left[ \log \sigma \left(R _ {\alpha} (c, G _ {\phi} (x, m, c, t ^ {\prime}))\right) \right], \tag {2} +$$ + +where $\mathcal{D}_{train}$ represents the fine-tuning dataset (i.e. LAION-5B [31]). The term $G\phi(x, m, c, t')$ denotes the decoded output image $x_0'$ generated by the generative model at step $t'$ , given the masked input $x \odot (1 - m)$ and the prompt $c$ . To further maintain the consistency and detail fidelity of the generated area and the original image area, we introduce pixel-level regularization (i.e., L1 loss) and a perceptual loss, which captures the discrepancy in VGG features [32]. Collectively, these regularization techniques can be formulated as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {r e g} (\phi) = \mathbb {E} _ {(x, m, c) \sim \mathcal {D} _ {t r a i n}, t ^ {\prime} \sim [ 1, T _ {2} ]} \| x - G _ {\phi} (x, m, c, t ^ {\prime}) \| _ {1}, \\ \mathcal {L} _ {r e g} (\cdot) = \mathbb {E} _ {\| V (x) - V (C _ {r e g}, t ^ {\prime}) \|} \end{array} \tag {3} +$$ + +$$ +\mathcal {L} _ {v g g} (\phi) = \mathbb {E} _ {(x, m, c) \sim \mathcal {D} _ {t r a i n}, t ^ {\prime} \sim [ 1, T _ {2} ]} \| V (x) - V \left(G _ {\phi} (x, m, c, t ^ {\prime})\right) \| _ {1}, +$$ + +where $V(\cdot)$ represents the VGG network. The overall training objective of our PeFL can be summarized as, + +$$ +\mathcal {L} _ {p e f l} (\phi) = \mathcal {L} _ {r e w a r d} + \eta \left(\mathcal {L} _ {r e g} + \mathcal {L} _ {v g g}\right), \tag {4} +$$ + +where $\eta$ is a hyperparameter for balancing loss weights. + +# 3.2 Comply: Image-Text Alignment with Coherence + +Diverging from the text-to-image synthesis focus of [39], our method encompasses an additional emphasis on assessing the alignment between the generated content of the masked area and the user-specified prompt, as well as + +ensuring coherence with the unmasked region at the pixel level. To achieve this, we introduce two further components in this section, which complement the aesthetic reward model $R_{\alpha}$ proposed earlier. + +Image-Text Alignment. We note that the presence of numerous poorly matched image-text pairs within the LAION dataset [31]. Exploiting these pairs as non-preferred samples for training reward models allows us to reduce manual annotation costs significantly. Initially, we employ the CLIP model [26] to identify image-text pairs with lower CLIPScore [11] from the LAION dataset. Subsequently, we leverage advanced multi-modal large language models such as LLAVA [19] to generate more informative and descriptive captions for the input images. These captions are considered more accurate than the original prompts. This process yields approximately 40,000 triplets $(c_p, c_n, x)$ as alignment dataset $\mathcal{D}_{\text{align}}$ , where $c_p$ and $c_n$ correspond to the preferred and non-preferred textual descriptions of the image $x$ , respectively. These triplets are utilized for training the image-text alignment reward model, denoted as $R_{\beta}(\cdot)$ . The architecture of $R_{\beta}$ mirrors that of $R_{\alpha}$ , while the training objective is similar to Eq. 1: + +$$ +\mathcal {L} (\beta) = - \mathbb {E} _ {\left(c _ {p}, c _ {n}, x\right) \sim \mathcal {D} _ {\text {a l i g n}}} \left[ \log \sigma \left(R _ {\beta} \left(c _ {p}, x\right) - R _ {\beta} \left(c _ {n}, x\right)\right) \right], \tag {5} +$$ + +Pixel-Level Coherence. The issue of coherence arises from the presence of inconsistent content within and outside the regions of interest, characterized by subtle visual cues such as color variations, stylistic discrepancies, and textural differences. To tackle this challenge, a coherent reward model, denoted as $R_{\gamma}(\cdot)$ , is specifically designed for pixel-level discrimination, as opposed to the holistic evaluation performed by $R_{\alpha}(\cdot)$ and $R_{\beta}(\cdot)$ . Our approach entails training a ViT-based [10] backbone network, followed by a prediction MLP head, to assess the authenticity and assign a score to each pixel in the input image. By formulating the loss function as follows: + +$$ +\mathcal {L} (\gamma) = - \mathbb {E} _ {(x, m, c) \sim \mathcal {D} _ {\text {t r a i n}, t ^ {\prime}} \sim [ 1, T _ {2} ]} \left[ \log \sigma \left(R _ {\gamma} (z)\right) + \log \left(1 - \sigma \left(R _ {\gamma} \left(z ^ {\prime}\right)\right)\right) \right], \tag {6} +$$ + +where $z \sim x \in \mathbb{R}^{H \times W \times 3}, z' \sim G_{\phi}(x, m, c, t') \in \mathbb{R}^{H \times W \times 3}$ are pixels of the corresponding image and $H, W$ represent the height and weight respectively. + +# 3.3 Accelerate: Adversarial and Progressive Training + +Adversarial training. Concurrent works such as UFOGen [41] and SDXL-Turbo [30] proposed to introduce adversarial training objective into fine-tuning diffusion models, which dramatically speeds up the sampling process and allows for one-step generation. They supposed that the Gaussian assumption of diffusion process does not hold anymore when the inference steps are extremely low, and therefore enabling the generative model to output samples in a single forward step by adversarial objective [36, 40]. We note that the functionality of our coherent reward model $R_{\gamma}(\cdot)$ is very similar to that of the discriminator in adversarial training, except for the different granularity in prediction. To this end, unlike the aesthetic and alignment reward models, which necessitate offline learning prior to fine-tuning, the coherent reward model can be learned online + +and seamlessly integrated into the fine-tuning process. The adversarial objective of generator that raises the score of output image is also in compliance with our feedback learning in Eq. (2), we can simply achieve adversarial training by incorporating the optimization of $R_{\gamma}(\cdot)$ into fine-tuning to serve as a discriminator. Thus the Eq. (2) can be reformulated as follows: + +$$ +\mathcal {L} _ {\text {r e w a r d}} (\phi) = - \underset { \begin{array}{c} (x, m, c) \sim \mathcal {D} _ {\text {t r a i n}} \\ t ^ {\prime} \sim [ 1, T _ {2} ] \end{array} } {\mathbb {E}} \sum_ {\theta \in \{\alpha , \beta , \gamma \}} \log \sigma \left(R _ {\theta} \left(c, G _ {\phi} (x, m, c, t ^ {\prime})\right)\right). \tag {7} +$$ + +For completeness, we also rewrite the overall training objective as, + +$$ +\mathcal {L} ^ {G} (\phi) = \mathcal {L} _ {p e f l} (\phi) = \mathcal {L} _ {r e w a r d} + \eta \left(\mathcal {L} _ {r e g} + \mathcal {L} _ {v g g}\right), \tag {8} +$$ + +$$ +\mathcal {L} ^ {D} (\gamma) = - \mathbb {E} _ {(x, m, c) \sim \mathcal {D} _ {t r a i n}, t ^ {\prime} \sim [ 1, T _ {2} ]} \left[ \log \sigma \left(R _ {\gamma} (z)\right) + \log \left(1 - \sigma \left(R _ {\gamma} \left(z ^ {\prime}\right)\right)\right) \right]. +$$ + +Progressive training. To expedite the sampling process, we employ a progressive training strategy where we gradually reduce the optimization time steps $T$ . Surprisingly, we find that the quality of the generated images does not significantly degrade under the supervisor of reward models. This approach strikes a fine balance between performance and speed, leading to compelling results. In our experiments, we adopt a two-phase progressive strategy. During phase 1, we set the optimization time steps as $T = 20$ , $T_{1} = 15$ , and $T_{2} = 10$ . In phase 2, we further decrease the time steps to $T = 8$ , $T_{1} = 6$ , and $T_{2} = 3$ . Notably, we achieve remarkable outcomes without employing any distillation operations, relying solely on the inheritance of model parameters. + +# 4 Experiments + +# 4.1 Implementation Details + +Dataset. The fine-tuning dataset, denoted as $\mathcal{D}_{train}$ , consists of a substantial collection of 7,562,283 images encompassing diverse domains such as real-world scenes, authentic portraits, and computer-generated (CG) images. To enhance the model's generalization ability and generation quality, we adopted a meticulous fine-grained masking strategy inspired by StableDiffusion [28]. Our masking strategy encompasses four distinct types of masks: global masks, irregular shape masks, square masks, and outward expansion masks. Each mask type corresponds to a specific probability value, which is randomly selected and applied to images during the training process. Moreover, we devised a specialized masking strategy tailored for Inpainting-Editing tasks. Leveraging instance-level data, which identifies specific objects within images, we introduced random dilation operations to coarsen the masks during training. These coarsened masks were then integrated with randomly generated masks surrounding the instances. This approach not only enhances the model's performance in instruction-based image editing tasks but also facilitates more accurate instance generation, ultimately leading to superior quality outputs. + +To evaluate the performance of our approach, we conducted comprehensive qualitative and quantitative assessments using two distinct datasets. The first dataset, UserBench, was meticulously curated by gathering a vast amount of user-customized image-text matching data from various online sources. This dataset proved invaluable for evaluating image inpainting and outpainting tasks. From this extensive collection, we judiciously handpicked 100 high-quality image-text matching pairs to serve as our test data. We also leverage the experimental results from this dataset to collect and report human preferences. The second dataset, EditBench [34], presents a novel benchmark specifically tailored for text-guided image inpainting. Consisting of 240 images, each image within EditBench is paired with a corresponding mask that precisely delineates the region within the image to be modified through inpainting. + +Training Setting. To facilitate the perceptual feedback learning stage, we employed a relatively small learning rate of 2e-06, complemented by a learning rate scheduling strategy that encompassed a warm-up phase consisting of 1000 iterations. Furthermore, to ensure stability in model parameter updates, we incorporated an exponential moving average (EMA) decay parameter set to 0.9999. Instead of employing $100\%$ noise as in ReFL [39], we introduced a 50The weight assigned to the perceptual feedback loss was set to 0.01. During the adversarial acceleration stage, we maintained similar settings to the perceptual feedback learning stage, with an additional adversarial loss weighted 0.05. + +# 4.2 Evaluation Principles and Criteria + +Subjective metrics. To assess the robustness of our proposed method, we conducted a comprehensive subjective evaluation involving both expert evaluations and a large number of volunteer participants. Expert evaluators were tasked with individually assessing each generated image and assigning scores based on three key aspects: coherence, structure, and aesthetics. These aspects were rated on a scale of 1 to 5, with higher scores indicating superior generation quality: 1) Coherence focused on evaluating the consistency of texture, style, and color between the generated region and the original image. 2) Structure emphasized the clarity, sharpness, and absence of deformations or mutilations, particularly in human body parts. 3) Aesthetics gauged the overall level of creativity and diversity exhibited by the generated images. In addition to expert evaluations, we also sought the opinions of a large number of volunteers, with over 36,000 samples collected. These volunteers were presented with pairs of generated images and asked to choose between "Good", "Same", or "Bad", representing their preference in terms of GSB (Good-Same-Bad). + +Objective metrics. In this study, we also incorporate objective text-image alignment metrics, specifically CLIPScore [11,26] and BLIPScore [18], to comprehensively evaluate the alignment of our models. + +Table 1: Comparisons with state-of-the-art generative image editing systems in terms of coherence, structure and aesthetic scored by experts. More than 6000 image-text pairs are randomly sampled for each task and we report the average scores. + +
MethodOutpaintingInpainting-EditingInpainting-Erasing
coherencestructureaestheticcoherencestructureaestheticcoherencestructure
MeiTu [3]3.012.732.752.772.892.513.313.25
Canva [2]2.722.852.653.423.403.082.922.90
Adobe [1]3.523.073.143.463.603.223.854.28
ByteEdit3.543.253.263.733.393.253.994.03
+ +![](images/cf636bf8a617296e06d56109b66d0df90bbf6322a0792683ded90017dc4a4555.jpg) +Fig. 3: Comparisons with state-of-the-art generative image editing systems in terms of human preference (i.e. GSB). More than 12,000 samples are collected for each task. For simplicity and to minimize the difficulty of collecting a large number of user opinions, we only offer the generated images by Adobe and our ByteEdit to the volunteers. "Good" indicates the generated images by our ByteEdit is preferred and vice versa. + +# 4.3 Comparisons with State of the arts + +We compare our method with concurrent state-of-the-art generative image editing systems such as Adobe [1], Canva [2] and MeiTu [3]. The comparisons cover three different tasks, including inpainting, inpainting-editing and inpainting-erasing. The inpainting editing will specify the content to be generated for the region of interest in the prompt. In contrast, inpainting-erasing requires the model to remove content within it and be as consistent as possible. Since the erased image has little change, experts were not asked to score aesthetics for user study. + +User study. The average scores evaluated by experts are shown in Tab. 1. From the results, our ByteEdit significantly outperforms the state-of-the-art across different metrics in the outpainting task. It demonstrates that our method works well to expand images based on existing content and maintains superior consistency, structural integrity, and creativity. As for the inpainting tasks, our method also can provide the most coherent edited or erased images. To further investigate the gap between Adobe and our proposed ByteEdit, we solicited feedback from a large number of volunteers on the images generated by both, and the results are illustrated in Fig. 3. The results show that users generally found the images we generated to be more natural in overall perception. Our GSB + +Table 2: The quantitative results of ByteEdit and recent state-of-the-art approaches. + +
MetricsUserBenchEditBench
Meitu [3]Canva [2]Adobe [1]ByteEditDiffEdit [8]BLD [4]EMILIE [14]ByteEdit
CLIPScore0.2350.2410.2370.2550.2720.2800.3110.329
BLIPScore0.1740.4670.4500.6870.5820.5960.6200.691
+ +![](images/2ed6c91839c8f57f22cd4e2f2b4db298eaa9e4af96f6a5de68881f7ac9ddd2c4.jpg) +Fig. 4: Qualitative comparison in inpainting. We highlight key areas with red boxes. + +superiority percentages (i.e. $(\mathrm{G} + \mathrm{S}) / (\mathrm{S} + \mathrm{B}) * 100\%$ ) on three different tasks are $105\%$ , $163\%$ , and $112\%$ , respectively. + +Quantitative comparison. To quantitatively evaluate the performance of our method compared to other approaches, we conduct a quantitative evaluation of how well the edited image can capture the semantics of the edit instruction successfully by measuring the CLIPScore and BLIPScore. We conduct the experiment in inpainting-editing task and the results are provided in Table 2. From the UserBench against state-of-the-art generative image editing systems, we noticed that the score results are not completely consistent with user opinion. Nevertheless, our method is still ahead of the second-place Canva by $5.8\% (+0.014)$ and $47.1\% (+0.22)$ in terms of CLIPScore and BLIPScore, respectively. As for the EditBench, we follow [14] to compare our method with several concurrent editing approaches, i.e. DiffEdit [8], BLD [4] and EMILIE [14]. It demonstrates that the ByteEdit consistently yields the state-of-the-art performance, which shows our superior quality, consistency and instruction adherence. + +Qualitative comparison. In Figure 4 and 5, we visualize samples produced by different systems under different tasks. It clearly shows that our method exhibits a superior performance for learning both coherence and aesthetic. For the inpainting task, the ByteEdit consistently follows the user-specified instructions and generates coherent images with better image-text alignment. It is worth noting that our system allows both prompt and prompt-free generation when it comes to outpainting, which has broader application scenarios in reality. + +![](images/2fd3228be46e22b2d5511fad86f39e758db184777a7121df7de83fe54e5c2fe0.jpg) +Fig. 5: Qualitative comparison in outpainting. We highlight key areas with red boxes. + +![](images/b47e8664a10faa04b98ea508f38751d4b719770e0df2d26b52bb680484e4e2ff.jpg) +(a) w/ vs w/o PeFL + +![](images/5d5138c8965e926afcb2ceb4e1efac445ef22008f9c393404657f5769e94784c.jpg) +(b) w/ vs w/o Acceleration +Fig. 6: Human Perference Evaluation on our proposed PeFL and Acceleration. + +# 4.4 Ablation Studies + +In Figure 6, we conduct ablation experiments on both our proposed PeFL and acceleration strategy. The experts were asked to choose GSB preference and we report the human preference rates in the figure, which are calculated as $(\mathrm{G} + \mathrm{S}) / [((\mathrm{G} + \mathrm{S}) + (\mathrm{S} + \mathrm{B}))] * 100\%$ for win and $(\mathrm{S} + \mathrm{B}) / [((\mathrm{G} + \mathrm{S}) + (\mathrm{S} + \mathrm{B}))] * 100\%$ for lose, respectively. The evaluation is similar to the user study, except that we combine structure and aesthetics to reduce evaluation costs. More visualizations are also included in Figure 7. + +PeFL preference. From the results in Figure 6(a), our proposed PeFL significantly improves the generation quality, outperforming the baseline on all different tasks. Especially in the outpainting task with PeFL, our method exceeds the baseline by about $60\%$ in terms of structure and aesthetic, which is consistent with the edited image shown at the top of Figure 7 that is more realistic and conforms to the rules of realistic physics. + +Acceleration preference. In Figure 6(b), we demonstrate that our model has no significant loss in either consistency or structure and aesthetic with the pro + +![](images/d098ace01f6fdeeab4ea8d6aa5108b57bd86e15abe9483cf188d30209631a3a5.jpg) +Fig. 7: Ablation Studies Visualization. + +gressive training strategy. To our surprise, we have even achieved both increasing speed and quality in the outpainting and inpainting-editing tasks. Based on our experimental observations, this phenomenon can be attributed to two underlying factors: i) Stable Training: By considering the discriminator as a reward model, trainable reward model offers flexible and robust supervision for PeFL, alleviating issues related to model over-optimization; ii) Expand Supervision Scope: The incorporation of adversarial supervision enables us to extend the time step of PEFL optimization. Consequently, even at high-noise stages, such as step 999, the model can still benefit from PeFL supervision, further driving improvements in model performance. The visualization at the bottom of Figure 7 also verifies this, where the outputs become more realistic and natural after acceleration. + +# 5 Discussion + +ByteEdit has demonstrated remarkable performance across various editing tasks. However, several promising directions warrant further exploration: + +- Performance: One avenue for improvement lies in developing more targeted reward models tailored to specific editing tasks. By refining the reward models, we can potentially unlock even higher levels of performance and generate more precise and desirable output. +- Acceleration: Another area of interest is investigating how ByteEdit can be further integrated with advanced techniques such as LCM and SDXL-turbo to achieve accelerated processing speeds. +- Task: Expanding the capabilities of ByteEdit beyond image editing to domains like video editing or instruction editing holds significant potential. + +By incorporating human feedback to optimize generative image editing, ByteEdit can greatly enhance the practicality and usability in real-world scenarios. We hope that our work will provide valuable insights and inspire deeper reflections in this field, propelling its future development. + +# References + +1. Adobe firefly - free generative ai for creatives. https://www.adobe.com/products/firefly.html 2, 4, 11, 12 +2. Free ai image generator: Online text to image app | canva. https://www.canva.com/ai-image-generator/ 4, 11, 12 +3. Miraclevision. https://ai.meitu.com/index/ 4,11,12 +4. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. ACM Transactions on Graphics (TOG) 42(4), 1-11 (2023) 12 +5. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18208-18218 (2022) 2, 4 +6. Chen, X., Huang, L., Liu, Y., Shen, Y., Zhao, D., Zhao, H.: Anydoor: Zero-shot object-level image customization. arXiv preprint arXiv:2307.09481 (2023) 2 +7. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dollar, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015) 5 +8. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427 (2022) 12 +9. Dong, H., Xiong, W., Goyal, D., Pan, R., Diao, S., Zhang, J., Shum, K., Zhang, T.: Raft: Reward ranked finetuning for generative foundation model alignment. arXiv preprint arXiv:2304.06767 (2023) 3, 4 +10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 8 +11. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021) 8, 10 +12. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020) 2, 3, 4 +13. Isajanyan, A., Shatveryan, A., Kocharyan, D., Wang, Z., Shi, H.: Social reward: Evaluating and enhancing generative ai through million-user feedback from an online creative community. arXiv preprint arXiv:2402.09872 (2024) 4 +14. Joseph, K., Udhayanan, P., Shukla, T., Agarwal, A., Karanam, S., Goswami, K., Srinivasan, B.V.: Iterative multi-granular image editing using diffusion models. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 8107-8116 (2024) 12 +15. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) 4 +16. Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., Levy, O.: Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems 36 (2024) 4 +17. Lee, K., Liu, H., Ryu, M., Watkins, O., Du, Y., Boutilier, C., Abbeel, P., Ghavamzadeh, M., Gu, S.S.: Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192 (2023) 3, 4 +18. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022) 6, 10 + +19. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023) 5, 8 +20. Lu, Y., Zhang, M., Ma, A.J., Xie, X., Lai, J.H.: Coarse-to-fine latent diffusion for pose-guided person image synthesis. arXiv preprint arXiv:2402.18078 (2024) 2 +21. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11461-11471 (2022) 2, 3, 4 +22. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008) 5 +23. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021) 2, 4 +24. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023) 5 +25. Qin, J., Wu, J., Chen, W., Ren, Y., Li, H., Wu, H., Xiao, X., Wang, R., Wen, S.: Diffusionpt: Llm-driven text-to-image generation system. arXiv preprint arXiv:2401.10061 (2024) 2 +26. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 4, 8, 10 +27. Ren, Y., Wu, J., Zhang, P., Zhang, M., Xiao, X., He, Q., Wang, R., Zheng, M., Pan, X.: Ugc: Unified gan compression for efficient image-to-image translation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17281-17291 (2023) 2 +28. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) 2, 5, 9 +29. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) 6 +30. Sauer, A., Lorenz, D., Blattmann, A., Rombach, R.: Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042 (2023) 8 +31. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022) 4, 7, 8 +32. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) 7 +33. Turc, I., Nemade, G.: Midjourney user prompts & generated images (250k) (2022). https://doi.org/10.34740/KAGGLE/DS/2349267 5 +34. Wang, S., Sahara, C., Montgomery, C., Pont-Tuset, J., Noy, S., Pellegrini, S., Onoe, Y., Laszlo, S., Fleet, D.J., Soricut, R., et al.: Imagen editor and editbench: Advancing and evaluating text-guided image inpainting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18359-18369 (2023) 2, 3, 10 + +35. Wu, X., Sun, K., Zhu, F., Zhao, R., Li, H.: Human preference score: Better aligning text-to-image models with human preference. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2096-2105 (2023) 4 +36. Xiao, Z., Kreis, K., Vahdat, A.: Tackling the generative learning trilemma with denoising diffusion gans. arXiv preprint arXiv:2112.07804 (2021) 8 +37. Xie, S., Zhang, Z., Lin, Z., Hinz, T., Zhang, K.: Smartbrush: Text and shape guided object inpainting with diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22428-22437 (2023) 2, 3, 4 +38. Xie, S., Zhao, Y., Xiao, Z., Chan, K.C., Li, Y., Xu, Y., Zhang, K., Hou, T.: Dreaminpainter: Text-guided subject-driven image inpainting with diffusion models. arXiv preprint arXiv:2312.03771 (2023) 2 +39. Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagereward: Learning and evaluating human preferences for text-to-image generation. arXiv preprint arXiv:2304.05977 (2023) 3, 4, 6, 7, 10 +40. Xu, Y., Gong, M., Xie, S., Wei, W., Grundmann, M., Hou, T., et al.: Semi-implicit denoising diffusion models (siddms). arXiv preprint arXiv:2306.12511 (2023) 8 +41. Xu, Y., Zhao, Y., Xiao, Z., Hou, T.: Ufogen: You forward once large scale text-to-image generation via diffusion gans. arXiv preprint arXiv:2311.09257 (2023) 8 +42. Yang, B., Gu, S., Zhang, B., Zhang, T., Chen, X., Sun, X., Chen, D., Wen, F.: Paint by example: Exemplar-based image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18381-18391 (2023) 2 +43. Yang, S., Chen, T., Zhou, M.: A dense reward view on aligning text-to-image diffusion with preference. arXiv preprint arXiv:2402.08265 (2024) 3, 4 +44. Yildirim, A.B., Baday, V., Erdem, E., Erdem, A., Dundar, A.: Inst-inpaint: Instructing to remove objects with diffusion models. arXiv preprint arXiv:2304.03246 (2023) 2, 4 +45. Yu, T., Feng, R., Feng, R., Liu, J., Jin, X., Zeng, W., Chen, Z.: Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790 (2023) 2, 4 +46. Yuan, H., Chen, Z., Ji, K., Gu, Q.: Self-play fine-tuning of diffusion models for text-to-image generation. arXiv preprint arXiv:2402.10210 (2024) 3, 4 +47. Zhang, M., Wu, J., Ren, Y., Li, M., Qin, J., Xiao, X., Liu, W., Wang, R., Zheng, M., Ma, A.J.: Diffusionengine: Diffusion model is scalable data engine for object detection. arXiv preprint arXiv:2309.03893 (2023) 2 +48. Zhang, Z., Zhang, S., Zhan, Y., Luo, Y., Wen, Y., Tao, D.: Confronting reward overoptimization for diffusion models: A perspective of inductive and primacy biases. arXiv preprint arXiv:2402.08552 (2024) 3, 4 \ No newline at end of file diff --git a/2024/ByteEdit_ Boost, Comply and Accelerate Generative Image Editing/images.zip b/2024/ByteEdit_ Boost, Comply and Accelerate Generative Image Editing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1b7c030409ef3290696186ef2ee7974c397982b2 --- /dev/null +++ b/2024/ByteEdit_ Boost, Comply and Accelerate Generative Image Editing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:526efa5bde15f40ab466ecf2960d5abe1f22d2f342b9d3fe4dea5de0d11fea4f +size 570333 diff --git a/2024/ByteEdit_ Boost, Comply and Accelerate Generative Image Editing/layout.json b/2024/ByteEdit_ Boost, Comply and Accelerate Generative Image Editing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c370ee3fa156fbed587b22e2693a0a0496691e16 --- /dev/null +++ b/2024/ByteEdit_ Boost, Comply and Accelerate Generative Image Editing/layout.json @@ -0,0 +1,8160 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 164, + 111, + 451, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 111, + 451, + 148 + ], + "spans": [ + { + "bbox": [ + 164, + 111, + 451, + 148 + ], + "type": "text", + "content": "ByteEdit: Boost, Comply and Accelerate Generative Image Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 154, + 168, + 459, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 168, + 459, + 205 + ], + "spans": [ + { + "bbox": [ + 154, + 168, + 459, + 205 + ], + "type": "text", + "content": "Yuxi Ren*, Jie Wu*†, Yanzuo Lu*, Huafeng Kuang, Xin Xia, Xionghui Wang, Qianqian Wang, Yixing Zhu, Pan Xie, Shiyin Wang, Xuefeng Xiao, Yitong Wang, Min Zheng, and Lean Fu" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 212, + 214, + 400, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 214, + 400, + 237 + ], + "spans": [ + { + "bbox": [ + 212, + 214, + 400, + 237 + ], + "type": "text", + "content": "ByteDance \nProject Page: https://byte-edit.github.io" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 162, + 271, + 453, + 555 + ], + "blocks": [ + { + "bbox": [ + 162, + 271, + 453, + 555 + ], + "lines": [ + { + "bbox": [ + 162, + 271, + 453, + 555 + ], + "spans": [ + { + "bbox": [ + 162, + 271, + 453, + 555 + ], + "type": "image", + "image_path": "06b8ed23b91324ec96d0ab590a7a942aae563d5ec4e5f443e4a4d2eeb5a21559.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 566, + 482, + 624 + ], + "lines": [ + { + "bbox": [ + 130, + 566, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 566, + 482, + 624 + ], + "type": "text", + "content": "Fig. 1: We introduce ByteEdit, a novel framework that utilizes feedback learning to enhance generative image editing tasks, resulting in outstanding generation performance, improved consistency, enhanced instruction adherence, and accelerated generation speed. To the best of our knowledge, ByteEdit emerges as the most superior and the fastest solution currently in the field of generative editing." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 653, + 329, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 653, + 329, + 666 + ], + "spans": [ + { + "bbox": [ + 146, + 653, + 329, + 666 + ], + "type": "text", + "content": "*Equal contribution. †Corresponding author." + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 160, + 117, + 455, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 117, + 455, + 336 + ], + "spans": [ + { + "bbox": [ + 160, + 117, + 455, + 336 + ], + "type": "text", + "content": "Abstract. Recent advancements in diffusion-based generative image editing have sparked a profound revolution, reshaping the landscape of image outpainting and inpainting tasks. Despite these strides, the field grapples with inherent challenges, including: i) inferior quality; ii) poor consistency; iii) insufficient instruction adherence; iv) suboptimal generation efficiency. To address these obstacles, we present ByteEdit, an innovative feedback learning framework meticulously designed to Boost, Comply, and Accelerate Generative Image Editing tasks. ByteEdit seamlessly integrates image reward models dedicated to enhancing aesthetics and image-text alignment, while also introducing a dense, pixel-level reward model tailored to foster coherence in the output. Furthermore, we propose a pioneering adversarial and progressive feedback learning strategy to expedite the model's inference speed. Through extensive large-scale user evaluations, we demonstrate that ByteEdit surpasses leading generative image editing products, including Adobe, Canva, and MeiTu, in both generation quality and consistency. ByteEdit-Outpainting exhibits a remarkable enhancement of " + }, + { + "bbox": [ + 160, + 117, + 455, + 336 + ], + "type": "inline_equation", + "content": "388\\%" + }, + { + "bbox": [ + 160, + 117, + 455, + 336 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 160, + 117, + 455, + 336 + ], + "type": "inline_equation", + "content": "135\\%" + }, + { + "bbox": [ + 160, + 117, + 455, + 336 + ], + "type": "text", + "content": " in quality and consistency, respectively, when compared to the baseline model. Experiments also verified that our acceleration models maintain excellent performance results in terms of quality and consistency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 349, + 404, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 349, + 404, + 361 + ], + "spans": [ + { + "bbox": [ + 160, + 349, + 404, + 361 + ], + "type": "text", + "content": "Keywords: Outpainting " + }, + { + "bbox": [ + 160, + 349, + 404, + 361 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 349, + 404, + 361 + ], + "type": "text", + "content": " Intpainting " + }, + { + "bbox": [ + 160, + 349, + 404, + 361 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 349, + 404, + 361 + ], + "type": "text", + "content": " Feedback Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 388, + 230, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 388, + 230, + 401 + ], + "spans": [ + { + "bbox": [ + 133, + 388, + 230, + 401 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 422, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 422, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 133, + 422, + 482, + 602 + ], + "type": "text", + "content": "The field of generative image editing has experienced remarkable advancements in recent years [5, 6, 20, 21, 23, 34, 37, 38, 42, 44, 45], propelled by the development of diffusion models [12, 25, 27, 28, 47]. This progress has led to the emergence of influential products that have reshaped the landscape of image editing. A notable example is Adobe Firefly [1], which has revolutionized the creative process by enabling users to seamlessly incorporate, extend, or remove content from images through simple text prompts, thereby transcending the traditional boundaries of Photoshop. In our paper, we focus on the domain of generative image editing, with particular emphasis on two key aspects: 1) Outpainting, involving the expansion of the surrounding scene in an image based on provided input or even without explicit prompts, and 2) Inpainting, encompassing the random masking of specific image regions followed by the generation of corresponding content guided by given prompts (Inpainting-Editing) or erase certain objects (Inpainting-Erasing). Despite the notable advancements achieved through diffusion-based algorithms, several challenges persist within this field:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 604, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 604, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 604, + 481, + 628 + ], + "type": "text", + "content": "Inferior Quality: the quality of generated images frequently falls short in terms of realism, aesthetic appeal, and fidelity to minute details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "text", + "content": "Insufficient Instruction Adherence: The existing models grapple with the arduous task of faithfully adhering to provided instructions, resulting in a lack of alignment between the generated image and the input text;" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Y. Ren et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "Poor Consistency: The generated regions exhibit an unsatisfactory level of coherence with the original image, manifesting as a deficiency in terms of color, style, texture, and other salient visual attributes;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 479, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 479, + 187 + ], + "type": "text", + "content": "Suboptimal Generation Efficiency: The generation process is characterized by sluggish speeds and inadequate efficiency, thereby imposing significant obstacles when confronted with large-scale image editing endeavors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 188, + 481, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 481, + 379 + ], + "type": "text", + "content": "Recently, various efforts have been made to address the aforementioned challenges in the field. For instance, Imagen Editor [34] has employed an object detection approach to extract inpainting masks, while simultaneously capitalizing on original high-resolution images to faithfully capture intricate details. SmartBrush [37] has adopted a multi-task training strategy coupled with precision controls, encompassing both text and shape guidance, to enhance visual quality, mask controllability, and preserve the background. Additionally, RePaint [21] has utilized a pre-trained unconditional DDPM [12] prior and ingeniously modified the reverse diffusion iterations to generate high-quality and diverse output images. However, these approaches have primarily focused on addressing singular problems and have yet to achieve a more comprehensive solution. Large Language Models (LLMs) has made a notable surge in incorporating learning based on human feedback, and initial endeavors have been undertaken in the Text-to-Image (T2I) domain [9,17,39,43,46,48]. Inspired by these developments, we pose the question: Can we leverage human feedback to guide generative image editing to unleash the potential for superior generation outcomes?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 380, + 481, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 481, + 547 + ], + "type": "text", + "content": "This paper introduces ByteEdit, an innovative framework for optimizing generative image editing through the incorporation of feedback learning. ByteEdit builds multiple reward models, namely the Aesthetic reward model, Alignment reward model, and Coherent reward model, to achieve exceptional generation effects, improved instruction adherence and enhanced consistency, respectively. These carefully designed reward models serve as the foundation for our proposed perceptual feedback learning (PeFL) mechanism, which provides task-aware and comprehensive supervision signals. Moreover, ByteEdit introduce an adversarial feedback learning strategy that employs the trainable reward model as the discriminator. This strategy ensures that the model benefits from the PeFL supervision and provide clear images even during high-noise stages, further improves both the performance and speed of our model. To expedite the sampling process, a progressive training strategy is employed to gradually reduce the optimization time steps and facilitate model inference in a low-steps regime." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 557, + 479, + 664 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 557, + 479, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 557, + 479, + 604 + ], + "spans": [ + { + "bbox": [ + 138, + 557, + 479, + 604 + ], + "type": "text", + "content": "- New Insight: To the best of our knowledge, we offer the first attempt to incorporate human feedback into the field of generative image editing. By-Edit significantly enhances the overall performance of the model across various key aspects, opening new horizons in this field of study." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 605, + 479, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 605, + 479, + 640 + ], + "spans": [ + { + "bbox": [ + 138, + 605, + 479, + 640 + ], + "type": "text", + "content": "- Comprehensive Framework: By designing complementary global-level and pixel-level reward models, we effectively guide the model towards achieving improved beauty, enhanced consistency, and superior image-text alignment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 642, + 479, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 642, + 479, + 664 + ], + "spans": [ + { + "bbox": [ + 138, + 642, + 479, + 664 + ], + "type": "text", + "content": "Efficiency and Pioneering: Progressive feedback and adversarial learning techniques are introduced to accomplish a remarkable acceleration in the" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "text", + "content": "ByteEdit" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 148, + 116, + 481, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 116, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 148, + 116, + 481, + 152 + ], + "type": "text", + "content": "model's inference speed, all while maintaining a minimal compromise on output quality. Notably, ByteEdit stands as the first successful endeavor in accelerating generative editing models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 152, + 481, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 152, + 481, + 200 + ], + "spans": [ + { + "bbox": [ + 138, + 152, + 481, + 200 + ], + "type": "text", + "content": "Outstanding Performance: Extensive user studies show that ByteEdit exhibits obvious advantages in terms of quality, consistency, efficiency, and speed, compared to the most competitive products. ByteEdit emerges as the fastest and most superior solution currently available in image editing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 220, + 237, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 220, + 237, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 220, + 237, + 232 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 247, + 482, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 474 + ], + "type": "text", + "content": "Generative Image Editing. Generative Image Editing is a research area focused on filling user-specified regions of interest with desired contents. GLIDE [23] is the pioneering work that introduced text-to-image diffusion for editing purposes, and Repaint [21], on the other hand, conditions an unconditionally trained model (e.g. DDPM [12]) and leverages visible pixels to fill in missing areas. To enable more precise editing control, Blended Diffusion [5] incorporates multimodal embeddings and enforces similarity between images and text using CLIP [26]. SmartBrush [37] pushes the boundaries of mask generation by utilizing instance and panoptic segmentation datasets instead of random generation. Further improvements include the introduction of the powerful Segment Anything (SAM) [15] model by [45], which achieves mask-free removal, filling, and replacing of multiple pipelines. Inst-Inpaint [44] specializes in text-driven object removal without the need for explicit masking. Additionally, this method proposes the GQA-Inpaint dataset, which comprises image pairs with and without the object of interest, facilitating effective object removal guided by textual input. In addition to comparing our proposed method with existing academic approaches, we also benchmark against industry methods like Adobe [1], Canva [2], and MeiTu [3], providing a comprehensive evaluation across different domains and highlighting the strengths of our approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": "Human Feedback Learning. Foundation models for text-to-image diffusion often rely on pre-training with large-scale web datasets, such as LAION-5B [31], which may result in generated content that deviates from human ethical values and legal compliance requirements. Previous approaches [9,17] attempted to address this issue by constructing preference datasets using hand-crafted prompts or expert generators. However, these methods suffered from over-fitting due to their limited real-world scenarios and generator capabilities. To overcome these limitations, researchers proposed various reward models trained with expert annotations [35,39] or feedback from web users [13,16] to enforce alignment with human preferences. Drawing inspiration from reinforcement learning with human feedback (RLHF) utilized in natural language processing (NLP), researchers explored the application of human feedback learning in text-to-image diffusion [39, 43, 46, 48] to achieve more realistic, faithful, and ethical outcomes. Among these efforts, ImageReward [39] primarily focused on overall image quality and overlooked the complexity of human perception. In our work, we extend the concept of human feedback learning by introducing three fine-grained independ-" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "Y. Ren et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "dent reward models tailored for generative image editing: aesthetics, image-text alignment, and pixel-level coherence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 161, + 397, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 161, + 397, + 176 + ], + "spans": [ + { + "bbox": [ + 131, + 161, + 397, + 176 + ], + "type": "text", + "content": "3 ByteEdit: Boost, Comply and Accelerate" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 190, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 482, + 308 + ], + "type": "text", + "content": "ByteEdit, focuses on generative image editing tasks that enable users to manipulate image content within a specific region of interest using textual descriptions. With an input image " + }, + { + "bbox": [ + 130, + 190, + 482, + 308 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 190, + 482, + 308 + ], + "type": "text", + "content": ", a region-of-interest mask " + }, + { + "bbox": [ + 130, + 190, + 482, + 308 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 190, + 482, + 308 + ], + "type": "text", + "content": ", and a user-provided textual description " + }, + { + "bbox": [ + 130, + 190, + 482, + 308 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 190, + 482, + 308 + ], + "type": "text", + "content": ", our primary objective is to generate an output image " + }, + { + "bbox": [ + 130, + 190, + 482, + 308 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 190, + 482, + 308 + ], + "type": "text", + "content": " that preserves the unmasked region in the input image " + }, + { + "bbox": [ + 130, + 190, + 482, + 308 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 190, + 482, + 308 + ], + "type": "text", + "content": ", while aligning the masked region well with both the description of " + }, + { + "bbox": [ + 130, + 190, + 482, + 308 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 190, + 482, + 308 + ], + "type": "text", + "content": " and visual attributes in " + }, + { + "bbox": [ + 130, + 190, + 482, + 308 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 190, + 482, + 308 + ], + "type": "text", + "content": ". In this study, we introduce two key functionalities within ByteEdit: ByteEdit-Outpainting and ByteEdit-Inpainting. ByteEdit-Outpainting extends the image by generating content beyond the boundaries of the input image, while ByteEdit-Inpainting fills or erases in arbitrary areas of the input image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 310, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 482, + 370 + ], + "type": "text", + "content": "The ByteEdit pipeline is presented in Fig 2, providing an overview of the system's workflow. In the subsequent subsections, we delve into the details of two crucial components: Boost (Sec. 3.1) and Comply (Sec. 3.2). Furthermore, we elaborate on the Accelerate scheme in Sec. 3.3, which illustrates an approach to expedite the processing time and improve the efficiency of the ByteEdit system." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 391, + 349, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 391, + 349, + 403 + ], + "spans": [ + { + "bbox": [ + 131, + 391, + 349, + 403 + ], + "type": "text", + "content": "3.1 Boost: Perceptual Feedback Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 414, + 481, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 481, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 481, + 473 + ], + "type": "text", + "content": "In the field of generative image editing, the persistent challenge of subpar quality has impelled us to propose a pioneering approach that introduces human feedback, hitherto unexplored in this domain. Our novel pipeline comprises three key components: feedback data collection, reward model training, and perceptual feedback learning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": "Feedback Data Collection. We first randomly extract more than 1,500,000 text prompts from the Midjourney Discord [33] and MS-COCO Caption [7] datasets. To ensure the diversity, a clustering algorithm, namely K-Means, was employed, leveraging the similarities derived from state-of-the-art large language models [19]. Further, the features were visualized in lower dimensions using t-SNE [22], enabling the identification of data points possessing the largest average distance from their k-nearest neighbors. We also manually eliminate less informative and decorative-dominated prompts such as \"unbelievable\", \"fantastic\" and \"brilliant\" to improve the prompt quality. This meticulous procedure yielded approximately 400,000 candidate prompts, exhibiting diverse distributions, which were subsequently subjected to multiple text-to-image diffusion models, including SD1.5 [28] and SDXL [24]. Those images which excessively inferior quality or ambiguous characteristic are manually removed. Accompanying each prompt, a set of four generated images was presented to experts, who were tasked with selecting the best and worst images based on various aspects, encompassing aesthetic appeal, color composition, structural coherence, and brightness. The" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "text", + "content": "ByteEdit" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 482, + 257 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 482, + 257 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 482, + 257 + ], + "type": "image", + "image_path": "9aed3886d29a05e2f603ec6aa56f481fdab35e7cb99183c0e2843c78d338734b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 262, + 482, + 285 + ], + "lines": [ + { + "bbox": [ + 131, + 262, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 131, + 262, + 482, + 285 + ], + "type": "text", + "content": "Fig. 2: ByteEdit formulates a comprehensive feedback learning framework that facilitating aesthetics, image-text matching, consistency and inference speed." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 310, + 482, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 482, + 346 + ], + "type": "text", + "content": "resulting dataset, herein referred to as the aesthetic preference dataset " + }, + { + "bbox": [ + 130, + 310, + 482, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{aes}" + }, + { + "bbox": [ + 130, + 310, + 482, + 346 + ], + "type": "text", + "content": ", encompasses a collection of massive triplets " + }, + { + "bbox": [ + 130, + 310, + 482, + 346 + ], + "type": "inline_equation", + "content": "(c,x_p,x_n)" + }, + { + "bbox": [ + 130, + 310, + 482, + 346 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 310, + 482, + 346 + ], + "type": "inline_equation", + "content": "x_{p}" + }, + { + "bbox": [ + 130, + 310, + 482, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 310, + 482, + 346 + ], + "type": "inline_equation", + "content": "x_{n}" + }, + { + "bbox": [ + 130, + 310, + 482, + 346 + ], + "type": "text", + "content": " correspond to the preferred and non-preferred generated images of prompt " + }, + { + "bbox": [ + 130, + 310, + 482, + 346 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 310, + 482, + 346 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 346, + 482, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 417 + ], + "type": "text", + "content": "Reward Model Training. Building upon this dataset, we follow the training techniques in [39] to learn an aesthetic reward model " + }, + { + "bbox": [ + 130, + 346, + 482, + 417 + ], + "type": "inline_equation", + "content": "R_{\\alpha}(\\cdot)" + }, + { + "bbox": [ + 130, + 346, + 482, + 417 + ], + "type": "text", + "content": " of trainable parameters " + }, + { + "bbox": [ + 130, + 346, + 482, + 417 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 346, + 482, + 417 + ], + "type": "text", + "content": ", which we briefly summarize here. The image and text features of the input are extracted from the BLIP [18] backbone, combined with cross attention, and fed into an MLP to obtain an aesthetic score. The training objective can be formulated as," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 187, + 429, + 481, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 429, + 481, + 444 + ], + "spans": [ + { + "bbox": [ + 187, + 429, + 481, + 444 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\alpha) = - \\mathbb {E} _ {(c, x _ {p}, x _ {n}) \\sim \\mathcal {D} _ {a e s}} [ \\log \\sigma (R _ {\\alpha} (c, x _ {p}) - R _ {\\alpha} (c, x _ {n})) ], \\tag {1}", + "image_path": "de90ad80de9e6189105355026c699c41456f023bffe8848850cc4abeacb2e9c1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 451, + 425, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 425, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 425, + 463 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 451, + 425, + 463 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 130, + 451, + 425, + 463 + ], + "type": "text", + "content": " represents the Sigmoid function used for normalization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 464, + 482, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 572 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 572 + ], + "type": "text", + "content": "Perceptual Feedback Learning. Leveraging the power of our crafted reward model, we specifically introduce Perceptual Feedback Learning (PeFL) to fine-tune diffusion models with human feedback for generative image editing. Departing from the conventional practice of sequentially refining the predictions from the final step " + }, + { + "bbox": [ + 130, + 464, + 482, + 572 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 130, + 464, + 482, + 572 + ], + "type": "text", + "content": " to the initial step " + }, + { + "bbox": [ + 130, + 464, + 482, + 572 + ], + "type": "inline_equation", + "content": "x_0'" + }, + { + "bbox": [ + 130, + 464, + 482, + 572 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 464, + 482, + 572 + ], + "type": "inline_equation", + "content": "x_{T} \\to x_{T-1} \\to \\dots \\to x_0'" + }, + { + "bbox": [ + 130, + 464, + 482, + 572 + ], + "type": "text", + "content": "), we adopt an innovative perspective by performing optimization on the direct prediction outcomes " + }, + { + "bbox": [ + 130, + 464, + 482, + 572 + ], + "type": "inline_equation", + "content": "x_{t} \\to x_{0}'" + }, + { + "bbox": [ + 130, + 464, + 482, + 572 + ], + "type": "text", + "content": " at various intermediate steps " + }, + { + "bbox": [ + 130, + 464, + 482, + 572 + ], + "type": "inline_equation", + "content": "t \\in [1,T]" + }, + { + "bbox": [ + 130, + 464, + 482, + 572 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 464, + 482, + 572 + ], + "type": "inline_equation", + "content": "T = 20" + }, + { + "bbox": [ + 130, + 464, + 482, + 572 + ], + "type": "text", + "content": " in this case). Through this dynamic exploration across different stages of denoising, we uncover the following distinctive observations:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 581, + 479, + 664 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 137, + 581, + 479, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 581, + 479, + 628 + ], + "spans": [ + { + "bbox": [ + 137, + 581, + 479, + 628 + ], + "type": "text", + "content": "- During the initial stages of denoising (" + }, + { + "bbox": [ + 137, + 581, + 479, + 628 + ], + "type": "inline_equation", + "content": "t \\in [11, 20]" + }, + { + "bbox": [ + 137, + 581, + 479, + 628 + ], + "type": "text", + "content": "), the generative model (i.e. U-Net [29]) struggles to seamlessly complement the full image. Consequently, the reward model encounters challenges in accurately evaluating images that are hindered by obstacle masking." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 629, + 479, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 629, + 479, + 664 + ], + "spans": [ + { + "bbox": [ + 137, + 629, + 479, + 664 + ], + "type": "text", + "content": "- As the denoising process progresses (" + }, + { + "bbox": [ + 137, + 629, + 479, + 664 + ], + "type": "inline_equation", + "content": "t \\in [1, 10]" + }, + { + "bbox": [ + 137, + 629, + 479, + 664 + ], + "type": "text", + "content": "), the reward model is able to identify and reward visually appealing images even in the presence of mild noise." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Y. Ren et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 368 + ], + "type": "text", + "content": "Drawing upon these insightful observations, we present an innovative stagewise approach, to amplify the potential of generative image editing. Our proposed method encompasses the following key stages: 1) In stage 1 (" + }, + { + "bbox": [ + 130, + 116, + 482, + 368 + ], + "type": "inline_equation", + "content": "t \\in [16,20]" + }, + { + "bbox": [ + 130, + 116, + 482, + 368 + ], + "type": "text", + "content": "), we simply skip these steps with extremely heavy noise by diffusing the masked input " + }, + { + "bbox": [ + 130, + 116, + 482, + 368 + ], + "type": "inline_equation", + "content": "x \\odot (1 - m)" + }, + { + "bbox": [ + 130, + 116, + 482, + 368 + ], + "type": "text", + "content": " into noisy latents at a fixed step " + }, + { + "bbox": [ + 130, + 116, + 482, + 368 + ], + "type": "inline_equation", + "content": "T_{1} = 15" + }, + { + "bbox": [ + 130, + 116, + 482, + 368 + ], + "type": "text", + "content": ". This strategy is motivated by the fact that the generative model's ability to fill in intricate details within the masked region is limited in this timestep, thus rendering the training overhead unnecessary. This is the central difference between our approach and the ReFL proposed in [39], whose training paradigm relies solely on prompts and starts inference from pure noise. We aim to bring about a more pronounced correlation between the output and input image, thus facilitating the preservation of details in the unmasked areas; 2) In stage 2 (" + }, + { + "bbox": [ + 130, + 116, + 482, + 368 + ], + "type": "inline_equation", + "content": "t \\in [t',15]" + }, + { + "bbox": [ + 130, + 116, + 482, + 368 + ], + "type": "text", + "content": "), we randomly select a denosing step " + }, + { + "bbox": [ + 130, + 116, + 482, + 368 + ], + "type": "inline_equation", + "content": "t' \\sim [1,T_2]" + }, + { + "bbox": [ + 130, + 116, + 482, + 368 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 116, + 482, + 368 + ], + "type": "inline_equation", + "content": "T_{2} = 10" + }, + { + "bbox": [ + 130, + 116, + 482, + 368 + ], + "type": "text", + "content": " in this case) and perform inference without gradient starting from the noisy latent, i.e. " + }, + { + "bbox": [ + 130, + 116, + 482, + 368 + ], + "type": "inline_equation", + "content": "x_{T_1} \\to x_{T_1-1} \\to \\dots \\to x_{t'}" + }, + { + "bbox": [ + 130, + 116, + 482, + 368 + ], + "type": "text", + "content": ". This method ensures that the complemented image generated at this stage exhibits a sufficiently high level of quality, rendering it more amenable to evaluation by the reward model; 3) In stage 3 (" + }, + { + "bbox": [ + 130, + 116, + 482, + 368 + ], + "type": "inline_equation", + "content": "x_{t'} \\to x_0'" + }, + { + "bbox": [ + 130, + 116, + 482, + 368 + ], + "type": "text", + "content": "), the final stage entails the direct prediction of the complemented image " + }, + { + "bbox": [ + 130, + 116, + 482, + 368 + ], + "type": "inline_equation", + "content": "x_0'" + }, + { + "bbox": [ + 130, + 116, + 482, + 368 + ], + "type": "text", + "content": ". We leverage the aesthetic score obtained from the reward model as invaluable human feedback to refine the generative model " + }, + { + "bbox": [ + 130, + 116, + 482, + 368 + ], + "type": "inline_equation", + "content": "G_{\\phi}(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 482, + 368 + ], + "type": "text", + "content": ". This refinement process is achieved through the utilization of the following loss function:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 376, + 481, + 390 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 376, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 159, + 376, + 481, + 390 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e w a r d}} (\\phi) = - \\mathbb {E} _ {(x, m, c) \\sim \\mathcal {D} _ {\\text {t r a i n}}, t ^ {\\prime} \\sim [ 1, T _ {2} ]} \\left[ \\log \\sigma \\left(R _ {\\alpha} (c, G _ {\\phi} (x, m, c, t ^ {\\prime}))\\right) \\right], \\tag {2}", + "image_path": "69cb051f55bb84189c75d849cdce216ff1450f0decbaa9abd55846e2d68d9239.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 398, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 482, + 483 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 398, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{train}" + }, + { + "bbox": [ + 130, + 398, + 482, + 483 + ], + "type": "text", + "content": " represents the fine-tuning dataset (i.e. LAION-5B [31]). The term " + }, + { + "bbox": [ + 130, + 398, + 482, + 483 + ], + "type": "inline_equation", + "content": "G\\phi(x, m, c, t')" + }, + { + "bbox": [ + 130, + 398, + 482, + 483 + ], + "type": "text", + "content": " denotes the decoded output image " + }, + { + "bbox": [ + 130, + 398, + 482, + 483 + ], + "type": "inline_equation", + "content": "x_0'" + }, + { + "bbox": [ + 130, + 398, + 482, + 483 + ], + "type": "text", + "content": " generated by the generative model at step " + }, + { + "bbox": [ + 130, + 398, + 482, + 483 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 130, + 398, + 482, + 483 + ], + "type": "text", + "content": ", given the masked input " + }, + { + "bbox": [ + 130, + 398, + 482, + 483 + ], + "type": "inline_equation", + "content": "x \\odot (1 - m)" + }, + { + "bbox": [ + 130, + 398, + 482, + 483 + ], + "type": "text", + "content": " and the prompt " + }, + { + "bbox": [ + 130, + 398, + 482, + 483 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 398, + 482, + 483 + ], + "type": "text", + "content": ". To further maintain the consistency and detail fidelity of the generated area and the original image area, we introduce pixel-level regularization (i.e., L1 loss) and a perceptual loss, which captures the discrepancy in VGG features [32]. Collectively, these regularization techniques can be formulated as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 172, + 490, + 481, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 490, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 172, + 490, + 481, + 510 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {r e g} (\\phi) = \\mathbb {E} _ {(x, m, c) \\sim \\mathcal {D} _ {t r a i n}, t ^ {\\prime} \\sim [ 1, T _ {2} ]} \\| x - G _ {\\phi} (x, m, c, t ^ {\\prime}) \\| _ {1}, \\\\ \\mathcal {L} _ {r e g} (\\cdot) = \\mathbb {E} _ {\\| V (x) - V (C _ {r e g}, t ^ {\\prime}) \\|} \\end{array} \\tag {3}", + "image_path": "d000fe4c55fc8d94043a33d9d84600210bf37b922e9ac90cfd7b03bb5e77e548.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 172, + 506, + 442, + 518 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 506, + 442, + 518 + ], + "spans": [ + { + "bbox": [ + 172, + 506, + 442, + 518 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v g g} (\\phi) = \\mathbb {E} _ {(x, m, c) \\sim \\mathcal {D} _ {t r a i n}, t ^ {\\prime} \\sim [ 1, T _ {2} ]} \\| V (x) - V \\left(G _ {\\phi} (x, m, c, t ^ {\\prime})\\right) \\| _ {1},", + "image_path": "1ccb90d2064a756fa299c0c499fdd3ca3c5494c9d2da5f7fa97e92ed104ecc5c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 526, + 482, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 526, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 526, + 482, + 550 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 526, + 482, + 550 + ], + "type": "inline_equation", + "content": "V(\\cdot)" + }, + { + "bbox": [ + 130, + 526, + 482, + 550 + ], + "type": "text", + "content": " represents the VGG network. The overall training objective of our PeFL can be summarized as," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 225, + 559, + 481, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 559, + 481, + 573 + ], + "spans": [ + { + "bbox": [ + 225, + 559, + 481, + 573 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p e f l} (\\phi) = \\mathcal {L} _ {r e w a r d} + \\eta \\left(\\mathcal {L} _ {r e g} + \\mathcal {L} _ {v g g}\\right), \\tag {4}", + "image_path": "1a73299e8dc0d016cb991183eecff924fb78ff242323bdfa16a4db94189d20cd.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 581, + 376, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 581, + 376, + 594 + ], + "spans": [ + { + "bbox": [ + 131, + 581, + 376, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 581, + 376, + 594 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 131, + 581, + 376, + 594 + ], + "type": "text", + "content": " is a hyperparameter for balancing loss weights." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 609, + 403, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 403, + 623 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 403, + 623 + ], + "type": "text", + "content": "3.2 Comply: Image-Text Alignment with Coherence" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Diverging from the text-to-image synthesis focus of [39], our method encompasses an additional emphasis on assessing the alignment between the generated content of the masked area and the user-specified prompt, as well as" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "text", + "content": "ByteEdit" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "ensuring coherence with the unmasked region at the pixel level. To achieve this, we introduce two further components in this section, which complement the aesthetic reward model " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "R_{\\alpha}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " proposed earlier." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 485, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 485, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 485, + 308 + ], + "type": "text", + "content": "Image-Text Alignment. We note that the presence of numerous poorly matched image-text pairs within the LAION dataset [31]. Exploiting these pairs as non-preferred samples for training reward models allows us to reduce manual annotation costs significantly. Initially, we employ the CLIP model [26] to identify image-text pairs with lower CLIPScore [11] from the LAION dataset. Subsequently, we leverage advanced multi-modal large language models such as LLAVA [19] to generate more informative and descriptive captions for the input images. These captions are considered more accurate than the original prompts. This process yields approximately 40,000 triplets " + }, + { + "bbox": [ + 130, + 152, + 485, + 308 + ], + "type": "inline_equation", + "content": "(c_p, c_n, x)" + }, + { + "bbox": [ + 130, + 152, + 485, + 308 + ], + "type": "text", + "content": " as alignment dataset " + }, + { + "bbox": [ + 130, + 152, + 485, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\text{align}}" + }, + { + "bbox": [ + 130, + 152, + 485, + 308 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 152, + 485, + 308 + ], + "type": "inline_equation", + "content": "c_p" + }, + { + "bbox": [ + 130, + 152, + 485, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 152, + 485, + 308 + ], + "type": "inline_equation", + "content": "c_n" + }, + { + "bbox": [ + 130, + 152, + 485, + 308 + ], + "type": "text", + "content": " correspond to the preferred and non-preferred textual descriptions of the image " + }, + { + "bbox": [ + 130, + 152, + 485, + 308 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 152, + 485, + 308 + ], + "type": "text", + "content": ", respectively. These triplets are utilized for training the image-text alignment reward model, denoted as " + }, + { + "bbox": [ + 130, + 152, + 485, + 308 + ], + "type": "inline_equation", + "content": "R_{\\beta}(\\cdot)" + }, + { + "bbox": [ + 130, + 152, + 485, + 308 + ], + "type": "text", + "content": ". The architecture of " + }, + { + "bbox": [ + 130, + 152, + 485, + 308 + ], + "type": "inline_equation", + "content": "R_{\\beta}" + }, + { + "bbox": [ + 130, + 152, + 485, + 308 + ], + "type": "text", + "content": " mirrors that of " + }, + { + "bbox": [ + 130, + 152, + 485, + 308 + ], + "type": "inline_equation", + "content": "R_{\\alpha}" + }, + { + "bbox": [ + 130, + 152, + 485, + 308 + ], + "type": "text", + "content": ", while the training objective is similar to Eq. 1:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 184, + 316, + 481, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 316, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 184, + 316, + 481, + 331 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\beta) = - \\mathbb {E} _ {\\left(c _ {p}, c _ {n}, x\\right) \\sim \\mathcal {D} _ {\\text {a l i g n}}} \\left[ \\log \\sigma \\left(R _ {\\beta} \\left(c _ {p}, x\\right) - R _ {\\beta} \\left(c _ {n}, x\\right)\\right) \\right], \\tag {5}", + "image_path": "1ca278c8b68c3f919cd973573c1dbe254b85f3f1d6c8891c09b1c0532021bf66.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 337, + 482, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 337, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 482, + 443 + ], + "type": "text", + "content": "Pixel-Level Coherence. The issue of coherence arises from the presence of inconsistent content within and outside the regions of interest, characterized by subtle visual cues such as color variations, stylistic discrepancies, and textural differences. To tackle this challenge, a coherent reward model, denoted as " + }, + { + "bbox": [ + 130, + 337, + 482, + 443 + ], + "type": "inline_equation", + "content": "R_{\\gamma}(\\cdot)" + }, + { + "bbox": [ + 130, + 337, + 482, + 443 + ], + "type": "text", + "content": ", is specifically designed for pixel-level discrimination, as opposed to the holistic evaluation performed by " + }, + { + "bbox": [ + 130, + 337, + 482, + 443 + ], + "type": "inline_equation", + "content": "R_{\\alpha}(\\cdot)" + }, + { + "bbox": [ + 130, + 337, + 482, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 337, + 482, + 443 + ], + "type": "inline_equation", + "content": "R_{\\beta}(\\cdot)" + }, + { + "bbox": [ + 130, + 337, + 482, + 443 + ], + "type": "text", + "content": ". Our approach entails training a ViT-based [10] backbone network, followed by a prediction MLP head, to assess the authenticity and assign a score to each pixel in the input image. By formulating the loss function as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 453, + 481, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 453, + 481, + 467 + ], + "spans": [ + { + "bbox": [ + 159, + 453, + 481, + 467 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\gamma) = - \\mathbb {E} _ {(x, m, c) \\sim \\mathcal {D} _ {\\text {t r a i n}, t ^ {\\prime}} \\sim [ 1, T _ {2} ]} \\left[ \\log \\sigma \\left(R _ {\\gamma} (z)\\right) + \\log \\left(1 - \\sigma \\left(R _ {\\gamma} \\left(z ^ {\\prime}\\right)\\right)\\right) \\right], \\tag {6}", + "image_path": "69e2bcb0441d3241f9d0813b5eedaaf8e05f5c348b1f2d862b3eca463dea48c3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 473, + 480, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 480, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 473, + 480, + 498 + ], + "type": "inline_equation", + "content": "z \\sim x \\in \\mathbb{R}^{H \\times W \\times 3}, z' \\sim G_{\\phi}(x, m, c, t') \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 473, + 480, + 498 + ], + "type": "text", + "content": " are pixels of the corresponding image and " + }, + { + "bbox": [ + 130, + 473, + 480, + 498 + ], + "type": "inline_equation", + "content": "H, W" + }, + { + "bbox": [ + 130, + 473, + 480, + 498 + ], + "type": "text", + "content": " represent the height and weight respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 515, + 407, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 515, + 407, + 528 + ], + "spans": [ + { + "bbox": [ + 131, + 515, + 407, + 528 + ], + "type": "text", + "content": "3.3 Accelerate: Adversarial and Progressive Training" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Adversarial training. Concurrent works such as UFOGen [41] and SDXL-Turbo [30] proposed to introduce adversarial training objective into fine-tuning diffusion models, which dramatically speeds up the sampling process and allows for one-step generation. They supposed that the Gaussian assumption of diffusion process does not hold anymore when the inference steps are extremely low, and therefore enabling the generative model to output samples in a single forward step by adversarial objective [36, 40]. We note that the functionality of our coherent reward model " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "R_{\\gamma}(\\cdot)" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " is very similar to that of the discriminator in adversarial training, except for the different granularity in prediction. To this end, unlike the aesthetic and alignment reward models, which necessitate offline learning prior to fine-tuning, the coherent reward model can be learned online" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "Y. Ren et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 177 + ], + "type": "text", + "content": "and seamlessly integrated into the fine-tuning process. The adversarial objective of generator that raises the score of output image is also in compliance with our feedback learning in Eq. (2), we can simply achieve adversarial training by incorporating the optimization of " + }, + { + "bbox": [ + 130, + 116, + 479, + 177 + ], + "type": "inline_equation", + "content": "R_{\\gamma}(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 479, + 177 + ], + "type": "text", + "content": " into fine-tuning to serve as a discriminator. Thus the Eq. (2) can be reformulated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 165, + 188, + 481, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 188, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 165, + 188, + 481, + 213 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e w a r d}} (\\phi) = - \\underset { \\begin{array}{c} (x, m, c) \\sim \\mathcal {D} _ {\\text {t r a i n}} \\\\ t ^ {\\prime} \\sim [ 1, T _ {2} ] \\end{array} } {\\mathbb {E}} \\sum_ {\\theta \\in \\{\\alpha , \\beta , \\gamma \\}} \\log \\sigma \\left(R _ {\\theta} \\left(c, G _ {\\phi} (x, m, c, t ^ {\\prime})\\right)\\right). \\tag {7}", + "image_path": "a8473c8c1af1a42004b92d09509feca651d5436107dfa30161a6862c654070dd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 223, + 424, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 223, + 424, + 236 + ], + "spans": [ + { + "bbox": [ + 131, + 223, + 424, + 236 + ], + "type": "text", + "content": "For completeness, we also rewrite the overall training objective as," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 245, + 480, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 245, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 151, + 245, + 480, + 266 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {G} (\\phi) = \\mathcal {L} _ {p e f l} (\\phi) = \\mathcal {L} _ {r e w a r d} + \\eta \\left(\\mathcal {L} _ {r e g} + \\mathcal {L} _ {v g g}\\right), \\tag {8}", + "image_path": "ea3432a127b75fc5d568b3208bd19488f9adcdce190da7e2e578403471ff99f1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 262, + 451, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 262, + 451, + 275 + ], + "spans": [ + { + "bbox": [ + 151, + 262, + 451, + 275 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {D} (\\gamma) = - \\mathbb {E} _ {(x, m, c) \\sim \\mathcal {D} _ {t r a i n}, t ^ {\\prime} \\sim [ 1, T _ {2} ]} \\left[ \\log \\sigma \\left(R _ {\\gamma} (z)\\right) + \\log \\left(1 - \\sigma \\left(R _ {\\gamma} \\left(z ^ {\\prime}\\right)\\right)\\right) \\right].", + "image_path": "6a33f45cbb942fda5bebffa33e3d113f6c7c677bd5d1385c58c4dc593fdac5ee.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "text", + "content": "Progressive training. To expedite the sampling process, we employ a progressive training strategy where we gradually reduce the optimization time steps " + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "text", + "content": ". Surprisingly, we find that the quality of the generated images does not significantly degrade under the supervisor of reward models. This approach strikes a fine balance between performance and speed, leading to compelling results. In our experiments, we adopt a two-phase progressive strategy. During phase 1, we set the optimization time steps as " + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "inline_equation", + "content": "T = 20" + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "inline_equation", + "content": "T_{1} = 15" + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "inline_equation", + "content": "T_{2} = 10" + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "text", + "content": ". In phase 2, we further decrease the time steps to " + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "inline_equation", + "content": "T = 8" + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "inline_equation", + "content": "T_{1} = 6" + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "inline_equation", + "content": "T_{2} = 3" + }, + { + "bbox": [ + 130, + 285, + 480, + 404 + ], + "type": "text", + "content": ". Notably, we achieve remarkable outcomes without employing any distillation operations, relying solely on the inheritance of model parameters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 425, + 229, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 425, + 229, + 438 + ], + "spans": [ + { + "bbox": [ + 131, + 425, + 229, + 438 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 452, + 279, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 452, + 279, + 464 + ], + "spans": [ + { + "bbox": [ + 131, + 452, + 279, + 464 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": "Dataset. The fine-tuning dataset, denoted as " + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{train}" + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": ", consists of a substantial collection of 7,562,283 images encompassing diverse domains such as real-world scenes, authentic portraits, and computer-generated (CG) images. To enhance the model's generalization ability and generation quality, we adopted a meticulous fine-grained masking strategy inspired by StableDiffusion [28]. Our masking strategy encompasses four distinct types of masks: global masks, irregular shape masks, square masks, and outward expansion masks. Each mask type corresponds to a specific probability value, which is randomly selected and applied to images during the training process. Moreover, we devised a specialized masking strategy tailored for Inpainting-Editing tasks. Leveraging instance-level data, which identifies specific objects within images, we introduced random dilation operations to coarsen the masks during training. These coarsened masks were then integrated with randomly generated masks surrounding the instances. This approach not only enhances the model's performance in instruction-based image editing tasks but also facilitates more accurate instance generation, ultimately leading to superior quality outputs." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 102 + ], + "type": "text", + "content": "ByteEdit" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": "To evaluate the performance of our approach, we conducted comprehensive qualitative and quantitative assessments using two distinct datasets. The first dataset, UserBench, was meticulously curated by gathering a vast amount of user-customized image-text matching data from various online sources. This dataset proved invaluable for evaluating image inpainting and outpainting tasks. From this extensive collection, we judiciously handpicked 100 high-quality image-text matching pairs to serve as our test data. We also leverage the experimental results from this dataset to collect and report human preferences. The second dataset, EditBench [34], presents a novel benchmark specifically tailored for text-guided image inpainting. Consisting of 240 images, each image within EditBench is paired with a corresponding mask that precisely delineates the region within the image to be modified through inpainting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 263, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 263, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 482, + 373 + ], + "type": "text", + "content": "Training Setting. To facilitate the perceptual feedback learning stage, we employed a relatively small learning rate of 2e-06, complemented by a learning rate scheduling strategy that encompassed a warm-up phase consisting of 1000 iterations. Furthermore, to ensure stability in model parameter updates, we incorporated an exponential moving average (EMA) decay parameter set to 0.9999. Instead of employing " + }, + { + "bbox": [ + 130, + 263, + 482, + 373 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 263, + 482, + 373 + ], + "type": "text", + "content": " noise as in ReFL [39], we introduced a 50The weight assigned to the perceptual feedback loss was set to 0.01. During the adversarial acceleration stage, we maintained similar settings to the perceptual feedback learning stage, with an additional adversarial loss weighted 0.05." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 407, + 334, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 334, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 334, + 419 + ], + "type": "text", + "content": "4.2 Evaluation Principles and Criteria" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 446, + 482, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 626 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 626 + ], + "type": "text", + "content": "Subjective metrics. To assess the robustness of our proposed method, we conducted a comprehensive subjective evaluation involving both expert evaluations and a large number of volunteer participants. Expert evaluators were tasked with individually assessing each generated image and assigning scores based on three key aspects: coherence, structure, and aesthetics. These aspects were rated on a scale of 1 to 5, with higher scores indicating superior generation quality: 1) Coherence focused on evaluating the consistency of texture, style, and color between the generated region and the original image. 2) Structure emphasized the clarity, sharpness, and absence of deformations or mutilations, particularly in human body parts. 3) Aesthetics gauged the overall level of creativity and diversity exhibited by the generated images. In addition to expert evaluations, we also sought the opinions of a large number of volunteers, with over 36,000 samples collected. These volunteers were presented with pairs of generated images and asked to choose between \"Good\", \"Same\", or \"Bad\", representing their preference in terms of GSB (Good-Same-Bad)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Objective metrics. In this study, we also incorporate objective text-image alignment metrics, specifically CLIPScore [11,26] and BLIPScore [18], to comprehensively evaluate the alignment of our models." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "Y. Ren et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 157, + 473, + 224 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 479, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 479, + 148 + ], + "type": "text", + "content": "Table 1: Comparisons with state-of-the-art generative image editing systems in terms of coherence, structure and aesthetic scored by experts. More than 6000 image-text pairs are randomly sampled for each task and we report the average scores." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 157, + 473, + 224 + ], + "lines": [ + { + "bbox": [ + 138, + 157, + 473, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 157, + 473, + 224 + ], + "type": "table", + "html": "
MethodOutpaintingInpainting-EditingInpainting-Erasing
coherencestructureaestheticcoherencestructureaestheticcoherencestructure
MeiTu [3]3.012.732.752.772.892.513.313.25
Canva [2]2.722.852.653.423.403.082.922.90
Adobe [1]3.523.073.143.463.603.223.854.28
ByteEdit3.543.253.263.733.393.253.994.03
", + "image_path": "833c2881cba17089f16db7badee976bb7fce3ae8b256cd9eb50ff739b39b2c33.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 136, + 239, + 477, + 330 + ], + "blocks": [ + { + "bbox": [ + 136, + 239, + 477, + 330 + ], + "lines": [ + { + "bbox": [ + 136, + 239, + 477, + 330 + ], + "spans": [ + { + "bbox": [ + 136, + 239, + 477, + 330 + ], + "type": "image", + "image_path": "cf636bf8a617296e06d56109b66d0df90bbf6322a0792683ded90017dc4a4555.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 342, + 482, + 399 + ], + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 399 + ], + "type": "text", + "content": "Fig. 3: Comparisons with state-of-the-art generative image editing systems in terms of human preference (i.e. GSB). More than 12,000 samples are collected for each task. For simplicity and to minimize the difficulty of collecting a large number of user opinions, we only offer the generated images by Adobe and our ByteEdit to the volunteers. \"Good\" indicates the generated images by our ByteEdit is preferred and vice versa." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 423, + 338, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 423, + 338, + 435 + ], + "spans": [ + { + "bbox": [ + 131, + 423, + 338, + 435 + ], + "type": "text", + "content": "4.3 Comparisons with State of the arts" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "text", + "content": "We compare our method with concurrent state-of-the-art generative image editing systems such as Adobe [1], Canva [2] and MeiTu [3]. The comparisons cover three different tasks, including inpainting, inpainting-editing and inpainting-erasing. The inpainting editing will specify the content to be generated for the region of interest in the prompt. In contrast, inpainting-erasing requires the model to remove content within it and be as consistent as possible. Since the erased image has little change, experts were not asked to score aesthetics for user study." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "User study. The average scores evaluated by experts are shown in Tab. 1. From the results, our ByteEdit significantly outperforms the state-of-the-art across different metrics in the outpainting task. It demonstrates that our method works well to expand images based on existing content and maintains superior consistency, structural integrity, and creativity. As for the inpainting tasks, our method also can provide the most coherent edited or erased images. To further investigate the gap between Adobe and our proposed ByteEdit, we solicited feedback from a large number of volunteers on the images generated by both, and the results are illustrated in Fig. 3. The results show that users generally found the images we generated to be more natural in overall perception. Our GSB" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 101 + ], + "type": "text", + "content": "ByteEdit" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 136, + 481, + 186 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "type": "text", + "content": "Table 2: The quantitative results of ByteEdit and recent state-of-the-art approaches." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 136, + 481, + 186 + ], + "lines": [ + { + "bbox": [ + 133, + 136, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 133, + 136, + 481, + 186 + ], + "type": "table", + "html": "
MetricsUserBenchEditBench
Meitu [3]Canva [2]Adobe [1]ByteEditDiffEdit [8]BLD [4]EMILIE [14]ByteEdit
CLIPScore0.2350.2410.2370.2550.2720.2800.3110.329
BLIPScore0.1740.4670.4500.6870.5820.5960.6200.691
", + "image_path": "a16ea7fb64e1e2b7bdcd8253f55cad09bba32266650de6ff64ce7cab64527a58.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 154, + 209, + 463, + 335 + ], + "blocks": [ + { + "bbox": [ + 154, + 209, + 463, + 335 + ], + "lines": [ + { + "bbox": [ + 154, + 209, + 463, + 335 + ], + "spans": [ + { + "bbox": [ + 154, + 209, + 463, + 335 + ], + "type": "image", + "image_path": "2ed6c91839c8f57f22cd4e2f2b4db298eaa9e4af96f6a5de68881f7ac9ddd2c4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 342, + 478, + 354 + ], + "lines": [ + { + "bbox": [ + 133, + 342, + 478, + 354 + ], + "spans": [ + { + "bbox": [ + 133, + 342, + 478, + 354 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparison in inpainting. We highlight key areas with red boxes." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 389, + 480, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 480, + 415 + ], + "type": "text", + "content": "superiority percentages (i.e. " + }, + { + "bbox": [ + 130, + 389, + 480, + 415 + ], + "type": "inline_equation", + "content": "(\\mathrm{G} + \\mathrm{S}) / (\\mathrm{S} + \\mathrm{B}) * 100\\%" + }, + { + "bbox": [ + 130, + 389, + 480, + 415 + ], + "type": "text", + "content": " ) on three different tasks are " + }, + { + "bbox": [ + 130, + 389, + 480, + 415 + ], + "type": "inline_equation", + "content": "105\\%" + }, + { + "bbox": [ + 130, + 389, + 480, + 415 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 389, + 480, + 415 + ], + "type": "inline_equation", + "content": "163\\%" + }, + { + "bbox": [ + 130, + 389, + 480, + 415 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 389, + 480, + 415 + ], + "type": "inline_equation", + "content": "112\\%" + }, + { + "bbox": [ + 130, + 389, + 480, + 415 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 420, + 482, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 482, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 482, + 576 + ], + "type": "text", + "content": "Quantitative comparison. To quantitatively evaluate the performance of our method compared to other approaches, we conduct a quantitative evaluation of how well the edited image can capture the semantics of the edit instruction successfully by measuring the CLIPScore and BLIPScore. We conduct the experiment in inpainting-editing task and the results are provided in Table 2. From the UserBench against state-of-the-art generative image editing systems, we noticed that the score results are not completely consistent with user opinion. Nevertheless, our method is still ahead of the second-place Canva by " + }, + { + "bbox": [ + 130, + 420, + 482, + 576 + ], + "type": "inline_equation", + "content": "5.8\\% (+0.014)" + }, + { + "bbox": [ + 130, + 420, + 482, + 576 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 420, + 482, + 576 + ], + "type": "inline_equation", + "content": "47.1\\% (+0.22)" + }, + { + "bbox": [ + 130, + 420, + 482, + 576 + ], + "type": "text", + "content": " in terms of CLIPScore and BLIPScore, respectively. As for the EditBench, we follow [14] to compare our method with several concurrent editing approaches, i.e. DiffEdit [8], BLD [4] and EMILIE [14]. It demonstrates that the ByteEdit consistently yields the state-of-the-art performance, which shows our superior quality, consistency and instruction adherence." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Qualitative comparison. In Figure 4 and 5, we visualize samples produced by different systems under different tasks. It clearly shows that our method exhibits a superior performance for learning both coherence and aesthetic. For the inpainting task, the ByteEdit consistently follows the user-specified instructions and generates coherent images with better image-text alignment. It is worth noting that our system allows both prompt and prompt-free generation when it comes to outpainting, which has broader application scenarios in reality." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "Y. Ren et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 116, + 465, + 260 + ], + "blocks": [ + { + "bbox": [ + 149, + 116, + 465, + 260 + ], + "lines": [ + { + "bbox": [ + 149, + 116, + 465, + 260 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 465, + 260 + ], + "type": "image", + "image_path": "2fd3228be46e22b2d5511fad86f39e758db184777a7121df7de83fe54e5c2fe0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 267, + 480, + 278 + ], + "lines": [ + { + "bbox": [ + 132, + 267, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 480, + 278 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparison in outpainting. We highlight key areas with red boxes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 291, + 305, + 411 + ], + "blocks": [ + { + "bbox": [ + 134, + 291, + 305, + 411 + ], + "lines": [ + { + "bbox": [ + 134, + 291, + 305, + 411 + ], + "spans": [ + { + "bbox": [ + 134, + 291, + 305, + 411 + ], + "type": "image", + "image_path": "b47e8664a10faa04b98ea508f38751d4b719770e0df2d26b52bb680484e4e2ff.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 415, + 257, + 424 + ], + "lines": [ + { + "bbox": [ + 184, + 415, + 257, + 424 + ], + "spans": [ + { + "bbox": [ + 184, + 415, + 257, + 424 + ], + "type": "text", + "content": "(a) w/ vs w/o PeFL" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 290, + 479, + 412 + ], + "blocks": [ + { + "bbox": [ + 307, + 290, + 479, + 412 + ], + "lines": [ + { + "bbox": [ + 307, + 290, + 479, + 412 + ], + "spans": [ + { + "bbox": [ + 307, + 290, + 479, + 412 + ], + "type": "image", + "image_path": "5d5138c8965e926afcb2ceb4e1efac445ef22008f9c393404657f5769e94784c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 415, + 443, + 424 + ], + "lines": [ + { + "bbox": [ + 345, + 415, + 443, + 424 + ], + "spans": [ + { + "bbox": [ + 345, + 415, + 443, + 424 + ], + "type": "text", + "content": "(b) w/ vs w/o Acceleration" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 145, + 434, + 466, + 445 + ], + "lines": [ + { + "bbox": [ + 145, + 434, + 466, + 445 + ], + "spans": [ + { + "bbox": [ + 145, + 434, + 466, + 445 + ], + "type": "text", + "content": "Fig. 6: Human Perference Evaluation on our proposed PeFL and Acceleration." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 468, + 245, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 245, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 245, + 478 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 486, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 569 + ], + "type": "text", + "content": "In Figure 6, we conduct ablation experiments on both our proposed PeFL and acceleration strategy. The experts were asked to choose GSB preference and we report the human preference rates in the figure, which are calculated as " + }, + { + "bbox": [ + 130, + 486, + 482, + 569 + ], + "type": "inline_equation", + "content": "(\\mathrm{G} + \\mathrm{S}) / [((\\mathrm{G} + \\mathrm{S}) + (\\mathrm{S} + \\mathrm{B}))] * 100\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 569 + ], + "type": "text", + "content": " for win and " + }, + { + "bbox": [ + 130, + 486, + 482, + 569 + ], + "type": "inline_equation", + "content": "(\\mathrm{S} + \\mathrm{B}) / [((\\mathrm{G} + \\mathrm{S}) + (\\mathrm{S} + \\mathrm{B}))] * 100\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 569 + ], + "type": "text", + "content": " for lose, respectively. The evaluation is similar to the user study, except that we combine structure and aesthetics to reduce evaluation costs. More visualizations are also included in Figure 7." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "type": "text", + "content": "PeFL preference. From the results in Figure 6(a), our proposed PeFL significantly improves the generation quality, outperforming the baseline on all different tasks. Especially in the outpainting task with PeFL, our method exceeds the baseline by about " + }, + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 641 + ], + "type": "text", + "content": " in terms of structure and aesthetic, which is consistent with the edited image shown at the top of Figure 7 that is more realistic and conforms to the rules of realistic physics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "content": "Acceleration preference. In Figure 6(b), we demonstrate that our model has no significant loss in either consistency or structure and aesthetic with the pro" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 409, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 409, + 91, + 447, + 101 + ], + "type": "text", + "content": "ByteEdit" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 116, + 478, + 254 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 478, + 254 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 478, + 254 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 478, + 254 + ], + "type": "image", + "image_path": "d098ace01f6fdeeab4ea8d6aa5108b57bd86e15abe9483cf188d30209631a3a5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 226, + 262, + 386, + 274 + ], + "lines": [ + { + "bbox": [ + 226, + 262, + 386, + 274 + ], + "spans": [ + { + "bbox": [ + 226, + 262, + 386, + 274 + ], + "type": "text", + "content": "Fig. 7: Ablation Studies Visualization." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 297, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 429 + ], + "type": "text", + "content": "gressive training strategy. To our surprise, we have even achieved both increasing speed and quality in the outpainting and inpainting-editing tasks. Based on our experimental observations, this phenomenon can be attributed to two underlying factors: i) Stable Training: By considering the discriminator as a reward model, trainable reward model offers flexible and robust supervision for PeFL, alleviating issues related to model over-optimization; ii) Expand Supervision Scope: The incorporation of adversarial supervision enables us to extend the time step of PEFL optimization. Consequently, even at high-noise stages, such as step 999, the model can still benefit from PeFL supervision, further driving improvements in model performance. The visualization at the bottom of Figure 7 also verifies this, where the outputs become more realistic and natural after acceleration." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 445, + 217, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 445, + 217, + 458 + ], + "spans": [ + { + "bbox": [ + 132, + 445, + 217, + 458 + ], + "type": "text", + "content": "5 Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 469, + 480, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 469, + 480, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 469, + 480, + 495 + ], + "type": "text", + "content": "ByteEdit has demonstrated remarkable performance across various editing tasks. However, several promising directions warrant further exploration:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 502, + 481, + 609 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 137, + 502, + 481, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 502, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 137, + 502, + 481, + 548 + ], + "type": "text", + "content": "- Performance: One avenue for improvement lies in developing more targeted reward models tailored to specific editing tasks. By refining the reward models, we can potentially unlock even higher levels of performance and generate more precise and desirable output." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 549, + 481, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 549, + 481, + 585 + ], + "spans": [ + { + "bbox": [ + 137, + 549, + 481, + 585 + ], + "type": "text", + "content": "- Acceleration: Another area of interest is investigating how ByteEdit can be further integrated with advanced techniques such as LCM and SDXL-turbo to achieve accelerated processing speeds." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 585, + 481, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 585, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 137, + 585, + 481, + 609 + ], + "type": "text", + "content": "- Task: Expanding the capabilities of ByteEdit beyond image editing to domains like video editing or instruction editing holds significant potential." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "By incorporating human feedback to optimize generative image editing, ByteEdit can greatly enhance the practicality and usability in real-world scenarios. We hope that our work will provide valuable insights and inspire deeper reflections in this field, propelling its future development." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Y. Ren et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 137, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "text", + "content": "1. Adobe firefly - free generative ai for creatives. https://www.adobe.com/products/firefly.html 2, 4, 11, 12" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 159, + 480, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 159, + 480, + 179 + ], + "spans": [ + { + "bbox": [ + 138, + 159, + 480, + 179 + ], + "type": "text", + "content": "2. Free ai image generator: Online text to image app | canva. https://www.canva.com/ai-image-generator/ 4, 11, 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 180, + 376, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 180, + 376, + 190 + ], + "spans": [ + { + "bbox": [ + 138, + 180, + 376, + 190 + ], + "type": "text", + "content": "3. Miraclevision. https://ai.meitu.com/index/ 4,11,12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 190, + 480, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 190, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 138, + 190, + 480, + 212 + ], + "type": "text", + "content": "4. Avrahami, O., Fried, O., Lischinski, D.: Blended latent diffusion. ACM Transactions on Graphics (TOG) 42(4), 1-11 (2023) 12" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 212, + 481, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 212, + 481, + 244 + ], + "spans": [ + { + "bbox": [ + 138, + 212, + 481, + 244 + ], + "type": "text", + "content": "5. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18208-18218 (2022) 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "type": "text", + "content": "6. Chen, X., Huang, L., Liu, Y., Shen, Y., Zhao, D., Zhao, H.: Anydoor: Zero-shot object-level image customization. arXiv preprint arXiv:2307.09481 (2023) 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 266, + 480, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 480, + 298 + ], + "type": "text", + "content": "7. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dollar, P., Zitnick, C.L.: Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015) 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 299, + 480, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 329 + ], + "type": "text", + "content": "8. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427 (2022) 12" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 331, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 480, + 363 + ], + "type": "text", + "content": "9. Dong, H., Xiong, W., Goyal, D., Pan, R., Diao, S., Zhang, J., Shum, K., Zhang, T.: Raft: Reward ranked finetuning for generative foundation model alignment. arXiv preprint arXiv:2304.06767 (2023) 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 363, + 480, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 363, + 480, + 406 + ], + "spans": [ + { + "bbox": [ + 134, + 363, + 480, + 406 + ], + "type": "text", + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 407, + 480, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 407, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 134, + 407, + 480, + 437 + ], + "type": "text", + "content": "11. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021) 8, 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 438, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 438, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 134, + 438, + 480, + 460 + ], + "type": "text", + "content": "12. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020) 2, 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 460, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 460, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 134, + 460, + 480, + 492 + ], + "type": "text", + "content": "13. Isajanyan, A., Shatveryan, A., Kocharyan, D., Wang, Z., Shi, H.: Social reward: Evaluating and enhancing generative ai through million-user feedback from an online creative community. arXiv preprint arXiv:2402.09872 (2024) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 492, + 480, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 492, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 134, + 492, + 480, + 535 + ], + "type": "text", + "content": "14. Joseph, K., Udhayanan, P., Shukla, T., Agarwal, A., Karanam, S., Goswami, K., Srinivasan, B.V.: Iterative multi-granular image editing using diffusion models. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 8107-8116 (2024) 12" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 536, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 536, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 134, + 536, + 480, + 567 + ], + "type": "text", + "content": "15. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 568, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 568, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 134, + 568, + 480, + 600 + ], + "type": "text", + "content": "16. Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., Levy, O.: Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems 36 (2024) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "text", + "content": "17. Lee, K., Liu, H., Ryu, M., Watkins, O., Du, Y., Boutilier, C., Abbeel, P., Ghavamzadeh, M., Gu, S.S.: Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192 (2023) 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "18. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022) 6, 10" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 101 + ], + "type": "text", + "content": "ByteEdit" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "19. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023) 5, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 159 + ], + "type": "text", + "content": "20. Lu, Y., Zhang, M., Ma, A.J., Xie, X., Lai, J.H.: Coarse-to-fine latent diffusion for pose-guided person image synthesis. arXiv preprint arXiv:2402.18078 (2024) 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 203 + ], + "type": "text", + "content": "21. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11461-11471 (2022) 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "text", + "content": "22. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008) 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 266 + ], + "type": "text", + "content": "23. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021) 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 266, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 266, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 132, + 266, + 480, + 299 + ], + "type": "text", + "content": "24. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023) 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 331 + ], + "type": "text", + "content": "25. Qin, J., Wu, J., Chen, W., Ren, Y., Li, H., Wu, H., Xiao, X., Wang, R., Wen, S.: Diffusionpt: Llm-driven text-to-image generation system. arXiv preprint arXiv:2401.10061 (2024) 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 331, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 480, + 374 + ], + "type": "text", + "content": "26. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 4, 8, 10" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 418 + ], + "type": "text", + "content": "27. Ren, Y., Wu, J., Zhang, P., Zhang, M., Xiao, X., He, Q., Wang, R., Zheng, M., Pan, X.: Ugc: Unified gan compression for efficient image-to-image translation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17281-17291 (2023) 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 418, + 480, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 459 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 459 + ], + "type": "text", + "content": "28. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) 2, 5, 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 460, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 480, + 504 + ], + "type": "text", + "content": "29. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "text", + "content": "30. Sauer, A., Lorenz, D., Blattmann, A., Rombach, R.: Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042 (2023) 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 525, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 568 + ], + "type": "text", + "content": "31. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022) 4, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 568, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 590 + ], + "type": "text", + "content": "32. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 590, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 610 + ], + "type": "text", + "content": "33. Turc, I., Nemade, G.: Midjourney user prompts & generated images (250k) (2022). https://doi.org/10.34740/KAGGLE/DS/2349267 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 610, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 480, + 665 + ], + "type": "text", + "content": "34. Wang, S., Sahara, C., Montgomery, C., Pont-Tuset, J., Noy, S., Pellegrini, S., Onoe, Y., Laszlo, S., Fleet, D.J., Soricut, R., et al.: Imagen editor and editbench: Advancing and evaluating text-guided image inpainting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18359-18369 (2023) 2, 3, 10" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Y. Ren et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 544 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "35. Wu, X., Sun, K., Zhu, F., Zhao, R., Li, H.: Human preference score: Better aligning text-to-image models with human preference. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2096-2105 (2023) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "text", + "content": "36. Xiao, Z., Kreis, K., Vahdat, A.: Tackling the generative learning trilemma with denoising diffusion gans. arXiv preprint arXiv:2112.07804 (2021) 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "type": "text", + "content": "37. Xie, S., Zhang, Z., Lin, Z., Hinz, T., Zhang, K.: Smartbrush: Text and shape guided object inpainting with diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22428-22437 (2023) 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "38. Xie, S., Zhao, Y., Xiao, Z., Chan, K.C., Li, Y., Xu, Y., Zhang, K., Hou, T.: Dreaminpainter: Text-guided subject-driven image inpainting with diffusion models. arXiv preprint arXiv:2312.03771 (2023) 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "text", + "content": "39. Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagereward: Learning and evaluating human preferences for text-to-image generation. arXiv preprint arXiv:2304.05977 (2023) 3, 4, 6, 7, 10" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "text", + "content": "40. Xu, Y., Gong, M., Xie, S., Wei, W., Grundmann, M., Hou, T., et al.: Semi-implicit denoising diffusion models (siddms). arXiv preprint arXiv:2306.12511 (2023) 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "text", + "content": "41. Xu, Y., Zhao, Y., Xiao, Z., Hou, T.: Ufogen: You forward once large scale text-to-image generation via diffusion gans. arXiv preprint arXiv:2311.09257 (2023) 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 369 + ], + "type": "text", + "content": "42. Yang, B., Gu, S., Zhang, B., Zhang, T., Chen, X., Sun, X., Chen, D., Wen, F.: Paint by example: Exemplar-based image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18381-18391 (2023) 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 391 + ], + "type": "text", + "content": "43. Yang, S., Chen, T., Zhou, M.: A dense reward view on aligning text-to-image diffusion with preference. arXiv preprint arXiv:2402.08265 (2024) 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "text", + "content": "44. Yildirim, A.B., Baday, V., Erdem, E., Erdem, A., Dundar, A.: Inst-inpaint: Instructing to remove objects with diffusion models. arXiv preprint arXiv:2304.03246 (2023) 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "text", + "content": "45. Yu, T., Feng, R., Feng, R., Liu, J., Jin, X., Zeng, W., Chen, Z.: Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790 (2023) 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 456, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 481, + 479 + ], + "type": "text", + "content": "46. Yuan, H., Chen, Z., Ji, K., Gu, Q.: Self-play fine-tuning of diffusion models for text-to-image generation. arXiv preprint arXiv:2402.10210 (2024) 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "47. Zhang, M., Wu, J., Ren, Y., Li, M., Qin, J., Xiao, X., Liu, W., Wang, R., Zheng, M., Ma, A.J.: Diffusionengine: Diffusion model is scalable data engine for object detection. arXiv preprint arXiv:2309.03893 (2023) 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "48. Zhang, Z., Zhang, S., Zhan, Y., Luo, Y., Wen, Y., Tao, D.: Confronting reward overoptimization for diffusion models: A perspective of inductive and primacy biases. arXiv preprint arXiv:2402.08552 (2024) 3, 4" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 408, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 408, + 91, + 447, + 101 + ], + "type": "text", + "content": "ByteEdit" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/C2C_ Component-to-Composition Learning for Zero-Shot Compositional Action Recognition/c79e54f7-761f-47c3-bbd9-d80fc5c9ad18_content_list.json b/2024/C2C_ Component-to-Composition Learning for Zero-Shot Compositional Action Recognition/c79e54f7-761f-47c3-bbd9-d80fc5c9ad18_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..43b4374088c90c7580753153c26c138c7d9dd4ad --- /dev/null +++ b/2024/C2C_ Component-to-Composition Learning for Zero-Shot Compositional Action Recognition/c79e54f7-761f-47c3-bbd9-d80fc5c9ad18_content_list.json @@ -0,0 +1,1855 @@ +[ + { + "type": "text", + "text": "C2C: Component-to-Composition Learning for Zero-Shot Compositional Action Recognition", + "text_level": 1, + "bbox": [ + 223, + 141, + 776, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rongchang Li $^{1,2}$ , Zhenhua Feng $^{1,2,3,4}$ , Tianyang Xu $^{1}$ , Linze Li $^{1}$ , Xiao-Jun Wu $^{1*}$ , Muhammad Awais $^{2,4}$ , Sara Atito $^{2,4}$ , and Josef Kittler $^{2,3}$", + "bbox": [ + 215, + 210, + 785, + 243 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ School of Artificial Intelligence and Computer Science, Jiangnan University, China", + "2 Centre for Vision, Speech and Signal Processing (CVSSP), University of Surrey, UK", + "$^{3}$ School of Computer Science and Electronic Engineering, University of Surrey, UK" + ], + "bbox": [ + 215, + 253, + 784, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Surrey Institute of People-centred AI (SI-PAI), University of Surrey, UK {li_rongchang,linze.li}@stu.jiangnan.edu.cn;", + "bbox": [ + 243, + 296, + 759, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "feng-zhenhua@outlook.com; {wu_xiaojun, tianyang_xu}@jiangnan.edu.cn; {muhammad.awais, sara.atito, j.kittler}@surrey.ac.uk;", + "bbox": [ + 243, + 325, + 756, + 352 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Compositional actions consist of dynamic (verbs) and static (objects) concepts. Humans can easily recognize unseen compositions using the learned concepts. For machines, solving such a problem requires a model to recognize unseen actions composed of previously observed verbs and objects, thus requiring so-called compositional generalization ability. To facilitate this research, we propose a novel Zero-Shot Compositional Action Recognition (ZS-CAR) task. For evaluating the task, we construct a new benchmark, Something-composition (Sth-com), based on the widely used Something-Something V2 dataset. We also propose a novel Component-to-Composition (C2C) learning method to solve the new ZS-CAR task. C2C includes an independent component learning module and a composition inference module. Last, we devise an enhanced training strategy to address the challenges of component variations between seen and unseen compositions and to handle the subtle balance between learning seen and unseen actions. The experimental results demonstrate that the proposed framework significantly surpasses the existing compositional generalization methods and sets a new state-of-the-art. The new Sth-com benchmark and code are available at https://github.com/RongchangLi/ZSCAR_C2C.", + "bbox": [ + 259, + 383, + 740, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Zero-shot compositional action recognition $\\cdot$ Action recognition $\\cdot$ Compositional generalization", + "bbox": [ + 259, + 660, + 740, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 710, + 375, + 727 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans can understand basic dynamic and static concepts (i.e., verbs and objects) in seen compositional actions and generalise the concepts to unseen action categories. For example, by seeing close a box and open a bag, we can envision close a bag even if we have never seen this before. This compositional generalization capability is a fundamental function of human intelligence, by which we can", + "bbox": [ + 212, + 741, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding Author.", + "bbox": [ + 217, + 824, + 388, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aec7be6f90d7e549c916c8d9e90d6d3f7e59a6a1bbc2e7869d90efab3a5cba4b.jpg", + "image_caption": [ + "Fig. 1: Zero-Shot Compositional Action Recognition (ZS-CAR) requires models to recognize unseen actions composed of verbs and objects observed in seen actions." + ], + "image_footnote": [], + "bbox": [ + 346, + 143, + 660, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "understand an 'infinite' number of complex concepts through assimilating the learned 'finite' concepts [8,14]. Compositional generalization has recently gained attention in the computer vision community, giving rise to the corresponding image analysis task Compositional Zero-Shot Learning (CZSL) [43]. In video analysis, [42] proposed the compositional action recognition task. However, this task only requires recognizing the dynamic component. It makes the task essentially a regular verb classification task characterized by explicit sample selection bias [20,68] regarding verb-object compositions, rather than verifying compositional generalization ability. Exploring the compositional generalization problem in the video modality is valuable as it presents a challenging video understanding task that requires comprehending both dynamic and static components while also generalising to novel compositions. Also, it provides a reliable way to develop open-scenario-related applications in autonomous vehicles, robots, etc.", + "bbox": [ + 212, + 294, + 787, + 489 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To facilitate the related research, we propose a novel task, namely Zero-Shot Compositional Action Recognition (ZS-CAR). We define a compositional action as the composition of a dynamic component verb (e.g., open, pick up, ...) and a static component object (e.g., door, wallet, ...). As shown in Fig. 1, ZS-CAR aims to enable machines to identify unseen action categories consisting of seen verbs and objects. The new proposed task poses a significant challenge. First, a model has to understand the intricate spatial and temporal information to identify the verb and object components, which is more challenging than the existing action recognition task. Second, the task naturally contains component domain variation and component compatibility variation. The first one is similar to the challenges in domain generalization, caused by the contextuality difference (e.g., transferring knowledge of book in close a book to recognize bend a book). The second one stems from the action category difference in the training and test domains. Overly accepting or rejecting verb-object relationships observed during training can detrimentally impact the generalization capability to recognize unseen actions. Given these challenges, the existing visual compositional generalization solutions cannot handle the ZS-CAR task well, as shown in Tab. 2.", + "bbox": [ + 212, + 491, + 787, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To solve ZS-CAR, we present an innovative Component-to-Composition (C2C) method to serve as the baseline. The pipeline of C2C is shown in Fig. 2. C2C first independently learns the dynamic and static components (i.e., verbs and objects). Then it employs the component consensus to recognize compositions (i.e., actions) from both dynamic and static perspectives, ensuring that both these two components are well learned. The procedure is motivated by the human process", + "bbox": [ + 212, + 750, + 790, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "R. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of assembling compositions. It involves an independent observation of a component, assessing its compatibility with another part, and gradually assembling the units into the entire structure. Moreover, C2C enables explicitly addressing the component domain variation and component compatibility variation problems. Accordingly, we propose an enhanced training strategy that involves minimizing the dependence on spurious information within visual features and balancing the emphasis between fitting seen and imagining unseen actions.", + "bbox": [ + 212, + 146, + 782, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Last, to enable the training of C2C and to evaluate its performance for the ZS-CAR task, we create a new dataset, Something-composition (Sth-com), based on the popular Something-Something V2 (Sth-V2) dataset [16]. The experimental results demonstrate that the vanilla or enhanced version of C2C achieves state-of-the-art results. To summarize, the main contributions of the paper include:", + "bbox": [ + 212, + 252, + 784, + 328 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A novel zero-shot compositional action recognition (ZS-CAR) task that investigates compositional generalization in the video modality.", + "- A benchmarking dataset for evaluating ZS-CAR. We elaborate on the challenges in ZS-CAR and demonstrate that the existing visual compositional generalization methods cannot handle such problems well.", + "- A novel Component-to-Composition (C2C) method as the vanilla solution and an enhanced training strategy to address the identified challenges in ZS-CAR. The experimental results demonstrate that C2C solves these challenges effectively and attains SOTA performances." + ], + "bbox": [ + 225, + 335, + 784, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 492, + 382, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video action recognition aims to identify actions in videos. The main challenge is extracting discriminative action features from videos. Previous CNN-based methods address this challenge in two primary ways: using 2D CNNs with separate temporal operations [27,28,30-32,34,35,58,59,62], or jointly using 3D CNNs [5,12,13,19,23,51,57,61] to dig out spatiotemporal features. Recently, Transformer-based models transform a video into 3D tokens and model the spatiotemporal correlations through attention layers [1,2,11,29,37,46,49], or adapt the visual image encoder in large vision-language models to the video modality [47,48,66].", + "bbox": [ + 212, + 522, + 784, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Zero-shot action recognition (ZSAR) aims to recognize action categories not appearing in the training set. A widely used idea is to measure video features and action embeddings in a common semantic space, hoping this space can generalise to unseen actions. The semantic embeddings are obtained by side semantic information, e.g., word embeddings [4, 41, 63, 69], multiple attributes [15, 36], elaborative rehearsals [7] and action descriptions [9, 52]. Different from ZSAR, the proposed ZS-CAR aims to explore the visual compositionality of actions and generalize the learned composition knowledge to unseen actions.", + "bbox": [ + 212, + 659, + 784, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compositional action recognition: The compositional nature of actions has recently received increasing attention. Action genome [22] takes action dynamics as the relationship changes and decomposes actions into spatiotemporal scene graphs. Something-else [42] proposes compositional action recognition that", + "bbox": [ + 212, + 779, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "C2C: Component-to-Composition Learning", + "bbox": [ + 442, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "only requires recognising verbs in unseen verb-object compositions. Subsequent studies [39,56,64,65] follow this setting and focus on extracting robust dynamic information. However, [42] essentially is a supervised task with explicit sample selection bias, rather than a compositional generalization problem.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image-based compositional generalization: In the image modality, [43] first introduces the Compositional Zero-Shot Learning (CZSL) task that requires models trained on seen attribute-object pairs to recognize unseen ones. A common idea is to find the nearest holistic composition embeddings for input images in a common space [40, 43, 44, 50, 54]. Some studies disentangle visual features to align with component embeddings [18, 24, 26, 33, 54, 55]. Recent works propose to adapt the vision-language model [53] to CZSL by text prompting methods [38, 45] and enhancing the interactions between language and visual features [38].", + "bbox": [ + 212, + 207, + 784, + 328 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Problem Definition and Benchmark", + "text_level": 1, + "bbox": [ + 215, + 356, + 594, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 The Proposed ZS-CAR Task", + "text_level": 1, + "bbox": [ + 215, + 391, + 496, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In traditional action recognition, a model predicts the action label defined as either a verb or a verb-object composition. For example, the Sth-V2 dataset [16] defines verb-only actions, such as open sth and pick up sth. Kinetics-400 [25] defines actions as both verb-object compositions and verb-only actions, such as play basketball, draw, etc. The recently proposed compositional action recognition task considers the interaction between verbs and objects. It requires recognising verbs of which the paired objects in the test set differ from those in the training set. Though introducing action compositionally by this constraint, this task only takes the verb-object relation as a sample selection bias controller, making it not a thorough compositional generalization problem.", + "bbox": [ + 212, + 420, + 784, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast to previous tasks, we introduce the compositional generalization perspective to understand actions, by proposing a novel Zero-Shot Compositional Action Recognition (ZS-CAR) task. We explicitly define a compositional action as a concept of verb and object. To verify the compositional generalization ability, ZS-CAR requires models to recognize unseen action categories composed of verbs and objects observed in seen actions. Formally, we denote the training set as $D_{\\mathrm{train}} = \\{(\\mathcal{X}_i,a_i)\\}_{i = 1}^N$ consisting of video $\\mathcal{X}$ labelled as action $a\\in A_{\\mathrm{train}}$ , with $A_{\\mathrm{train}}\\subset V\\times O = \\{(v,o)|v\\in V,o\\in O\\}$ , where $V$ and $O$ are the verb and object sets. The validation and test sets are similarly denoted as $D_{\\mathrm{val}}$ and $D_{\\mathrm{test}}$ , with label sets as $A_{\\mathrm{val}}$ and $A_{\\mathrm{test}}$ . We follow the generalised zero-shot learning setting [6,50], i.e., both seen and unseen action categories appear in the validation/test set: $A_{\\mathrm{train}}\\cap A_{\\mathrm{val / test}}\\neq \\emptyset$ and $\\exists a(a\\in A_{\\mathrm{val / test}}\\oplus a\\in A_{\\mathrm{train}})$ .", + "bbox": [ + 212, + 573, + 784, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 The Something-composition Benchmark", + "text_level": 1, + "bbox": [ + 215, + 780, + 591, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To assess the proposed task, we build a benchmark Something-composition (Sth-com), based on the widely used Something-Something V2 (Sth-V2) dataset [16].", + "bbox": [ + 212, + 809, + 782, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "R. Li et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/2cf08a3b0f3c9b0ce92cc91f6b2570789b26175e03b5605e53fbebd4c98ace6e.jpg", + "table_caption": [ + "Table 1: Statistics of the proposed Sth-com benchmark." + ], + "table_footnote": [], + "table_body": "
# Objects# Verbs# Verb-Object Actions# Sample
Train248161345138034
Val225157733 Seen + 717 Unseen18774
Test233161976 Seen + 956 Unseen22657
All248161512479465
", + "bbox": [ + 341, + 171, + 656, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sth-V2 has 220,847 videos of humans performing pre-defined compositional actions with everyday objects. It is widely used by the action recognition community, as its verbs cannot be identified from a single frame.", + "bbox": [ + 212, + 257, + 782, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Annotation: We use the training and test split of Sth-else [42], a subset of Sth-V2, as the starting point. Then, we iteratively clean the subsets and ensure that: (1) Each composition (i.e., action) has more than five samples. (2) The components (i.e., verb & object) present in the test set are seen in the training set. This yields 248 objects, 161 verbs, 5124 compositions and 79465 video samples. Next, we select $1/3$ of compositions in the training and test sets respectively and interchange $50\\%$ of the selected compositions, thereby introducing seen action categories to the test set. Subsequently, following the ratio (3:4) in MIT-State [21], we divide the test set into the validation and test sets, respectively. The statistics of Sth-com are summarised in Tab. 1.", + "bbox": [ + 212, + 301, + 784, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Component-to-Composition Learning", + "text_level": 1, + "bbox": [ + 215, + 476, + 612, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address the proposed ZS-CAR task, we propose a novel Component-to-Composition (C2C) learning framework, as shown in Fig. 2, to serve as a vanilla solution. We simulate the process of human assembly to infer compositions. We commence with a specific component and scrutinise the compatibility of various categories of another component, leading to a reasonable composition proposal. These are correspondingly implemented in the proposed independent component learning module and composition to component learning module. Importantly, we analyse the challenges in ZS-CAR and devise an enhanced training strategy to facilitate generalising to unseen actions better.", + "bbox": [ + 212, + 507, + 787, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Independent Component Learning", + "text_level": 1, + "bbox": [ + 215, + 664, + 545, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this module, our goal is to identify the verb/object score of an input video and acquire their semantic prototypes. Specifically, we first encode the raw video sequence $\\mathcal{X}$ with shape $T\\times 224\\times 224\\times 3$ to a general representation $\\mathbf{F}_x\\in \\mathbb{R}^{T\\times D}$ , where $T$ and $D$ denote the frame and channel numbers, respectively. Then, we respectively adopt dynamic and static encoders to extract dynamic and static features, denoted as $\\mathbf{f}_v$ and $\\mathbf{f}_o\\in \\mathbb{R}^C$ , where $C$ is the channel number. The static encoder consists of a temporal average pooling layer and two MLP layers, while the dynamics encoder contains two temporal convolution layers and a temporal pooling layer. We encode the component labels as the general representations of verbs and objects, denoted as the component prototypes $\\mathbf{E}_v = \\{e_{v,i}\\}_{i = 1}^{N_v}$ and", + "bbox": [ + 212, + 688, + 787, + 843 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "C2C: Component-to-Composition Learning", + "bbox": [ + 442, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e913e1be1af537543f79c24fab0979c538f8cbe92cd32b37dd215fc40174eea9.jpg", + "image_caption": [ + "Fig. 2: The proposed Component-to-Composition (C2C) learning framework. C2C first aligns verb/object prototypes with corresponding visual features to obtain component scores in the Independent Component Learning module. Then, actions are inferred through two paths (dynamics and static) in the Component to Composition module. In the dynamics path, verb prototypes and visual features are used to compute conditional object scores. Then the independent verb scores and conditional object scores are multiplied to gain the action scores. The static path follows a similar procedure. The final output is a consensus of the results from both paths." + ], + "image_footnote": [], + "bbox": [ + 233, + 146, + 769, + 377 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathbf{E}_o = \\{e_{o,i}\\}_{i = 1}^{N_o}$ , where $N_{v}$ and $N_{o}$ are the verb and object numbers, respectively. The prototypes are learned by being aligned with the visual features $\\mathbf{f}_v$ and $\\mathbf{f}_o$ using cosine similarity as the gauge. Denoting the action label of the input video as $a_{i} = (v_{l},o_{k})$ , the optimization objective is defined as:", + "bbox": [ + 212, + 526, + 787, + 588 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {v e r b}} = - \\log \\frac {\\exp \\left(\\cos \\left(\\mathbf {f} _ {v} , \\mathbf {e} _ {v , l}\\right) / \\tau\\right)}{\\sum_ {v _ {j} \\in V} \\exp \\left(\\cos \\left(\\mathbf {f} _ {v} , \\mathbf {e} _ {v , j}\\right) / \\tau\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 595, + 785, + 631 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {o b j} = - \\log \\frac {\\exp \\left(\\cos \\left(\\mathbf {f} _ {o} , \\mathbf {e} _ {o , k}\\right) / \\tau\\right)}{\\sum_ {o _ {j} \\in O} \\exp \\left(\\cos \\left(\\mathbf {f} _ {o} , \\mathbf {e} _ {o , j}\\right) / \\tau\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 637, + 785, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{e}_{v,l}$ and $\\mathbf{e}_{o,k}$ are the corresponding prototypes. $\\cos(,)$ is the cosine similarity. $\\tau$ is the temperature coefficient for better optimization. We denote $\\mathcal{L}_{comp} = \\mathcal{L}_{verb} + \\mathcal{L}_{obj}$ as the component loss. By aligning with diverse component features in the training set, the prototypes inherently acquire semantic meanings. Accordingly, the verb and object scores, $\\mathbf{s}_v$ and $\\mathbf{s}_o$ , are the cosine similarities between the visual features and corresponding component prototypes.", + "bbox": [ + 212, + 678, + 787, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 From component to composition", + "text_level": 1, + "bbox": [ + 214, + 787, + 532, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this module, we infer the action (composition) score $\\mathbf{s}_a \\in \\mathbb{R}^{N_a}$ using the prototypes and the verbs and objects (components) scores. We define the conditional", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "R. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "scores $\\mathbf{S}_{o|v} \\in \\mathbb{R}^{N_v \\times N_o}$ and $\\mathbf{S}_{v|o} \\in \\mathbb{R}^{N_o \\times N_v}$ , that represents the confidence of one component given the visual input and another component as the conditions. Accordingly, for one action $a_i = (v_l, o_k)$ , its confidence score can be calculated by either $s_{a,i} = s_{v,l} \\cdot s_{o = o_k|v = v_l}$ or $s_{a,i} = s_{o,k} \\cdot s_{v = v_l|o = o_k}$ . The first one relies on the verbs to infer actions, called the dynamics path. Similarly, the other one is called the static path. As shown in Fig. 2, in the dynamics path, we first encode $\\mathbf{F}_x$ using a static encoder to provide the visual reference for the object component. Then we concatenate the output with the verb prototype $\\mathbf{e}_{v,l}$ and fuse them to the compound prototype-visual feature $\\mathbf{f}_{v - x,l}$ , using an MLP layer. Then we linearly map the cosine similarity $\\cos(\\mathbf{f}_{v - x,l}, \\mathbf{e}_{o,k})$ to the range of [0,1], obtaining the conditional object score $s_{o = o_k|v = v_l}$ . The static path follows a similar process except for encoding $\\mathbf{F}_x$ by the dynamics encoder. Last, we average the inference results from the two paths to obtain the final inference output. Taking $s_{a,i}$ as the score of the groundtruth action $a_i$ , the entire model is optimized using:", + "bbox": [ + 212, + 145, + 787, + 358 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c o m} = - \\log \\frac {\\exp \\left(s _ {a , i} / \\tau\\right)}{\\sum_ {a _ {j} \\in A _ {\\text {t r a i n}}} \\exp \\left(s _ {a , j} / \\tau\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 364, + 785, + 400 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Enhanced training strategy", + "text_level": 1, + "bbox": [ + 215, + 417, + 485, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Risk in generalizing components. Naturally, there is a high probability of exhibiting different component appearances between training and test domains, namely component domain variation. As shown in Fig. 3, the component 'book' exhibits a notable difference in appearance between 'close a book' and 'bend a book'. Besides, given the combination of 'close' and", + "bbox": [ + 212, + 439, + 431, + 633 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ba8a62b8eb8f5b84c1acf216c9484bdd1c383cb1c1d58fef493db1877242ca2.jpg", + "image_caption": [ + "Fig. 3: Component domain variations. For unseen actions, an object may exhibit different appearances (left). To deal with this (right), we reduce the spurious information and enlarge the independence between component-specific features." + ], + "image_footnote": [], + "bbox": [ + 447, + 468, + 767, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "'book' in the training set, there is a tendency that the model might align features associated with 'close' to 'book', or vice versa. These dependencies may confuse the model when encountering unseen action categories.", + "bbox": [ + 212, + 633, + 787, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To cope with it, in the independent component learning module, we posit that the ideal dynamic-focused feature $\\mathbf{f}_v$ and static-focused feature $\\mathbf{f}_o$ should encompass two types of information: common and specific. The former denotes features pertinent to both components and the latter denotes the invariant features, that are crucial for identifying the respective components. Notably, the invariant features of verbs are spurious for identifying objects, and vice versa. As shown in Fig. 3, our solution is to minimize the spurious information in $\\mathbf{f}_v$ and $\\mathbf{f}_o$ , and explicitly reduce the mutual associations between their component-specific parts. To achieve this, we propose the following auxiliary losses:", + "bbox": [ + 212, + 680, + 787, + 816 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\sup , v e r b} = h \\left(\\mathbf {f} _ {x}, \\mathbf {f} _ {v}\\right) - h \\left(\\mathbf {f} _ {v}, \\mathbf {y} _ {v}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 824, + 785, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "C2C: Component-to-Composition Learning", + "bbox": [ + 442, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s u p, o b j} = h \\left(\\mathbf {f} _ {x}, \\mathbf {f} _ {o}\\right) - h \\left(\\mathbf {f} _ {o}, \\mathbf {y} _ {o}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 161, + 785, + 179 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i n d} = \\mathcal {L} _ {s u p, v e r b} + \\mathcal {L} _ {s u p, o b j} + h \\left(\\mathbf {f} _ {v} ^ {\\prime}, \\mathbf {f} _ {o} ^ {\\prime}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 198, + 785, + 215 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $h(,.)$ is the Hilbert-Schmidt Independence Criterion (HSIC) [17] for independence measure. $h(\\mathbf{f},\\mathbf{h}) = 0$ if and only if $\\mathbf{f} \\perp \\mathbf{h}$ . $\\mathbf{f}_x$ is the temporal pooling result of $\\mathbf{F}_x$ . $\\mathbf{y}_v$ and $\\mathbf{y}_o$ are the one-hot label vectors. $\\mathcal{L}_{sup,verb}$ and $\\mathcal{L}_{sup,obj}$ force the model to compress information, but retain sufficient information to identify the components, thus reducing the general spurious information. $\\mathbf{f}_v'$ or $\\mathbf{f}_o'$ is part of $\\mathbf{f}_v$ or $\\mathbf{f}_o$ , defined as $\\mathbf{f}_v' = \\mathbf{f}_v[: \\rho C], \\mathbf{f}_o' = \\mathbf{f}_o[: \\rho C]$ ( $0 \\leq \\rho \\leq 1$ ). They are seen as component-specific features. So the term $h(\\mathbf{f}_v',\\mathbf{f}_o')$ ensures the independence of the invariant part of the verb and object features.", + "bbox": [ + 212, + 220, + 787, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Balanced learning of seen & unseen actions. In ZS-CAR, the verb-object connections in the training and test sets differ significantly, as shown in Fig. 4. While the seen connections help optimize the modules, an excessive emphasis on them may hinder generalizing to unseen actions. Our solution is simultaneously enhancing the learning of seen and unseen connections, enabling them to strike an optimal balance.", + "bbox": [ + 212, + 343, + 478, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We first assist the model in learning seen actions explicitly. We calculate the observed conditional", + "bbox": [ + 212, + 523, + 477, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "probabilities in the training set as the observed conditional scores $\\hat{\\mathbf{S}}_{v|o}$ and $\\hat{\\mathbf{S}}_{o|v}$ . Then, we force the conditional scores in a training batch closer to $\\hat{\\mathbf{S}}_{v|o}$ and $\\hat{\\mathbf{S}}_{o|v}$ , with the objectives:", + "bbox": [ + 212, + 568, + 787, + 617 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n}, v} = - \\frac {1}{N _ {v}} \\sum_ {v _ {i} \\in V} \\hat {\\mathbf {s}} _ {o | v = v _ {i}} \\log \\mathbf {s} _ {o | v = v _ {i}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 627, + 785, + 664 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n}, o} = - \\frac {1}{N _ {o}} \\sum_ {o _ {i} \\in O} \\hat {\\mathbf {s}} _ {v | o = o _ {i}} \\log \\mathbf {s} _ {v | o = o _ {i}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 676, + 785, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $N_v$ & $N_o$ denote the verb & object numbers, and the condition loss $\\mathcal{L}_{con}$ is defined as the sum of $\\mathcal{L}_{con,v}$ and $\\mathcal{L}_{con,o}$ . $\\hat{\\mathbf{s}}_{o|v = v_i}$ or $\\mathbf{s}_{o|v = v_i}$ denotes the conditional object score vector when verb is $v_i$ . $\\hat{\\mathbf{s}}_{v|o = o_i}$ and $\\mathbf{s}_{v|o = o_i}$ are defined similarly.", + "bbox": [ + 212, + 719, + 787, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Then, inspired by the human ability to create new things by imagining compositions of known components, we use CutMix [67] to assemble samples in a training batch as the imagined unseen actions, thereby reducing the risk of overemphasizing seen compositions. Specifically, for an input $\\mathcal{X}_i$ with label $a_i = (v_l,o_k)$ , we randomly select another sample $\\mathcal{X}_j$ with label $a_j = (v_m,o_n)$", + "bbox": [ + 212, + 766, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1c3d32af904095b36440c6f43eb8a1cd3b7ecbe0519a944234b9574ddda48a26.jpg", + "image_caption": [ + "Fig. 4: Component compatibility variations. The component relations are different between the training and test sets (left). To solve this (right), we use the observed conditional score to fit seen relations and encourage the model to imagine unseen actions to avoid being limited to seen relations." + ], + "image_footnote": [], + "bbox": [ + 488, + 372, + 787, + 444 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "R. Li et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c6773eb28dc02674d44c8c4153d4c4a9da32e11e1b5592e8722ad160c7eee254.jpg", + "table_caption": [ + "Table 2: A comparison with the SOTA methods on Sth-com, using the vision model with word embedding model paradigm. We report both vanilla (only $\\mathcal{L}_{com}$ and $\\mathcal{L}_{comp}$ ) and enhanced (using the advocated enhanced training strategy) versions of C2C. Since ADE is tailored for transformer architectures, we only adapt it to VideoSwin." + ], + "table_footnote": [], + "table_body": "
TSM-18VideoSwin-T
Methodsverbobj.seenunseenHMAUCverbobj.seenunseenHMAUC
TMN [50]41.435.931.628.122.07.333.445.631.030.223.07.9
Compcos [40]45.840.034.835.127.510.636.750.534.237.827.210.9
CGE [44]35.536.427.625.520.05.931.848.131.230.323.17.9
OADis [55]39.140.733.532.325.59.341.648.938.138.330.212.6
ADE [18]------35.349.133.836.226.810.3
CAN [60]40.039.232.431.624.58.732.649.533.033.325.09.2
C2C (Vanilla)47.742.738.139.130.813.247.950.942.143.634.916.4
C2C (Enhanced)50.144.339.441.033.014.548.852.843.444.136.717.4
", + "bbox": [ + 258, + 212, + 743, + 333 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "and crop it at a random position of random size. The cropped part replaces the same position of $\\mathcal{X}_i$ , generating the augmented input $\\bar{\\mathcal{X}}_i$ . Denoting the cropping ratio as $\\lambda$ , we modify the previous losses as: $\\bar{\\mathcal{L}}_{com} = (1 - \\lambda)\\mathcal{L}_{com}^{a_i} + \\lambda \\mathcal{L}_{com}^{a_j}$ and $\\bar{\\mathcal{L}}_{comp} = (1 - \\lambda)\\mathcal{L}_{comp}^{v_l,o_k} + \\lambda \\mathcal{L}_{comp}^{v_m,o_n}$ . $\\mathcal{L}_{comp}^{a_i}$ calculates $\\mathcal{L}_{com}$ using Eq. (3) with $a_i$ as groundtruth. $\\mathcal{L}_{comp}^{v_l,o_k}$ calculates $\\mathcal{L}_{comp}$ with $v_l$ and $o_k$ as groundtruth. $\\mathcal{L}_{comp}^{a_j}$ and $\\mathcal{L}_{comp}^{v_m,o_n}$ are defined similarly. Also, $\\mathcal{L}_{ind}$ is modified to $\\bar{\\mathcal{L}}_{ind}$ by replacing $\\mathbf{y}_i$ to $(1 - \\lambda)\\mathbf{y}_i + \\lambda \\mathbf{y}_j$ in Eq. (4) and Eq. (5). Particularly, we introduce new compositions $\\bar{a}_1 = (v_l,o_n)$ and $\\bar{a}_2 = (v_m,o_k)$ . $\\bar{a}_1$ and $\\bar{a}_2$ are not constrained in $A_{\\mathrm{test}}$ or $A_{\\mathrm{train}}$ . It gives a novel loss $\\mathcal{L}_{new} = \\mathcal{L}_{com}^{\\bar{a}_1} + \\mathcal{L}_{com}^{\\bar{a}_2}$ , which forces the model to consider all the possible component connections.", + "bbox": [ + 212, + 362, + 787, + 515 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A concern is that the new compositions $a_1$ or $a_2$ may be counterfactual. However, it only introduces noise to the component-to-composition module, as the component labels still match the visual input. In this way, we still reserve these counterfactual labels as they can provide regularization effects that avoid overfitting to seen verb-object relations. Besides, it is also consistent with human imagination which is not only constrained within reasonable compositions.", + "bbox": [ + 212, + 516, + 787, + 608 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4 More details", + "text_level": 1, + "bbox": [ + 215, + 631, + 366, + 645 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Network details. Following previous practice, we implement the video encoder and component label encoding with two paradigms:", + "bbox": [ + 212, + 657, + 785, + 688 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) Vision model with word embedding model. In this paradigm, we adopt an action recognition model (the CNN-based TSM-18 [35] or the transformer-based VideoSwin-T [37]) to extract the general video feature and a word embedding model (fasttext [3]) to encode the labels.", + "(2) Vision-language models. Recently, large pre-trained vision-language models, such as CLIP [53], have shown superior power. In this paradigm, we adopt the language branch of recent prompt-based compositional generalization methods (Coop [70], CSP [45] and SPM [38]) to encode the labels. As CLIP are trained using image-text pairs, we apply image-to-video adapters introduced in AIM [66] to adapt the CLIP visual encoder (ViT-B/32 [10]) to video modality." + ], + "bbox": [ + 212, + 688, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "C2C: Component-to-Composition Learning", + "bbox": [ + 442, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/4ce21c17755a8161a5fe4065a3b82344555080de4122dfba5f0553edc1a9af47.jpg", + "table_caption": [ + "Table 3: A comparison with the SOTA CLIP-based methods on Sth-com." + ], + "table_footnote": [], + "table_body": "
methodsverbobj.seenunseenHMAUC
Coop [70]48.455.143.946.436.618.1
+C2C(vanilla)56.256.948.354.642.724.0
+C2C(enhance)56.358.950.756.644.526.0
CSP [45]49.054.843.647.436.018.0
+C2C(vanilla)56.356.649.053.442.523.8
+C2C(enhance)57.158.649.956.244.325.5
SPM [38]48.954.544.046.335.517.7
+DFM [38]47.455.443.147.335.817.9
+C2C(vanilla)55.757.349.353.442.223.8
+C2C(enhance)56.558.449.955.743.825.2
", + "bbox": [ + 354, + 170, + 643, + 303 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training Objectives. When applying the proposed enhanced training strategy, we perform CutMix to generate new compositions with a probability $p$ . If apply CutMix, the final loss is calculated as:", + "bbox": [ + 212, + 330, + 787, + 377 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\mathcal {L}} _ {t o t a l} = \\bar {\\mathcal {L}} _ {c o m} + \\alpha \\bar {\\mathcal {L}} _ {c o m p} + \\beta \\bar {\\mathcal {L}} _ {i n d} + \\gamma \\mathcal {L} _ {n e w}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 386, + 785, + 405 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "If not, the final loss is:", + "bbox": [ + 214, + 414, + 382, + 429 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c o m}} + \\alpha \\mathcal {L} _ {\\text {c o m p}} + \\beta \\mathcal {L} _ {\\text {i n d}} + \\gamma \\mathcal {L} _ {\\text {c o n}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 441, + 784, + 458 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As $\\mathcal{L}_{new}$ and $\\mathcal{L}_{con}$ respectively guide the model focus on unseen and seen actions, we use the same coefficient $\\gamma$ .", + "bbox": [ + 212, + 468, + 785, + 498 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 214, + 521, + 375, + 539 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Metrics. The metrics include Verb, Object, Seen, Unseen, HM and AUC. Verb and Object respectively denote the accuracy of verbs and objects. To gain other metrics, we first add a constant to the seen actions scores. The constant ranges from negative infinity to positive infinity, making the model shift from completely biased to unseen actions to completely biased to seen actions and yielding the seen-unseen curve shown in Fig. 5. Seen and Unseen are the best accuracy of seen/unseen actions, i.e. the intersection points with x/y-axis in Fig. 5. HM is the best harmonic mean with formulation: $HM = 2(Seen * Unseen) / (Seen + Unseen)$ . AUC is the area under the curve.", + "bbox": [ + 212, + 551, + 787, + 686 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation details. $\\alpha$ , $\\beta$ $\\gamma$ in Eq. (9) and Eq. (10) are 0.2, 0.1 and 0.1, respectively. The probability $p$ of conducting CutMix is 0.7. More coefficient ablations and implementation details are reported in supplementary materials.", + "bbox": [ + 212, + 686, + 787, + 734 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Comparisons with State-of-the-Art", + "text_level": 1, + "bbox": [ + 214, + 753, + 550, + 771 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Note that the existing compositional action recognition methods cannot apply to the ZS-CAR task. To comprehensively demonstrate the proposed C2C, we first adapt the existing CZSL methods to the proposed Sth-com dataset. We replace their image encoders with video encoders and adopt the same word", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "R. Li et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/40f9618959d02742b46cbd8b4138ce65c779c72155adfca6ee31635cc3e3ec23.jpg", + "table_caption": [ + "Table 4: A comparison with the SOTA methods on C-GQA." + ], + "table_footnote": [], + "table_body": "
multiseenunseenHMAUC
CGE [44]×38.017.118.55.4
OADis [55]×38.319.820.17.0
CoT [26]✓39.222.722.17.4
C2C×39.123.023.07.7
C2C✓39.423.623.38.2
", + "bbox": [ + 377, + 170, + 620, + 247 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/69d4fc5f2d531e6f06b3119081459611c96e46cb952a2072fc11da6e4c36c460.jpg", + "table_caption": [ + "Table 5: Results of different hypotheses inferring compositions from components." + ], + "table_footnote": [], + "table_body": "
verbobj.seenunseenHMAUC
Compcos [40]45.840.034.835.127.510.6
Independent47.942.237.437.429.812.2
Knowledge-agnostic48.443.138.736.830.212.6
Dynamic centered46.742.537.438.029.912.5
Static centered46.641.137.036.728.911.7
Full47.742.738.139.130.813.2
", + "bbox": [ + 341, + 291, + 656, + 378 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "embedding model, making sure only the compositional generalization parts are different between the adapted methods and our proposed methods. Besides, we also modify C2C and compare it with the existing SOTA methods on the image-based CZSL task using the challenging C-GQA [44] dataset.", + "bbox": [ + 212, + 411, + 784, + 470 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparisons on Sth-com. We first compare our method with the adapted CZSL methods of the vision model with the word embedding paradigm, using both CNN-based and transformer-based video encoders. The results are reported in Tab. 2. The vanilla version of C2C uses two basic losses. The enhanced version uses the proposed enhanced training strategy.", + "bbox": [ + 212, + 473, + 784, + 549 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tab. 2 demonstrate that our method outperforms the existing methods by significant margins across all the metrics on ZS-CAR. Furthermore, the proposed enhanced training strategy provides consistent performance gains across different video encoders. Tab. 2 shows that the recently proposed methods based on Compcos are inferior to Compcos. Second, except for OADis, using the recently proposed VideoSwin-T backbone does not yield significant improvements. The first observation may stem from failing to generalise the embedding relations, not disentangling spatial and temporal information well, and not applying the conditional component relations in actions. Besides, jointly learning the compositions may lead to unbalanced learning of the components. These deficiencies degrade the effectiveness of video encoders, leading directly to the second observation. Tab. 2 also compares two video encoders. Though VideoSwin is an advanced model, its verb accuracy is consistently inferior or similar to TSM across all methods. Conversely, it outperforms TSM in object accuracy. This implies that our ZS-CAR task offers a new avenue for assessing biases regarding dynamic and static patterns in action recognition models.", + "bbox": [ + 212, + 551, + 785, + 792 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Then we apply our C2C method to recently proposed CLIP-based compositional zero-shot learning methods. The results are reported in Tab. 3. Coop, CSP and SPM are three text prompting methods. We can see that the proposed", + "bbox": [ + 212, + 794, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "C2C: Component-to-Composition Learning", + "bbox": [ + 442, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "C2C method can improve their performance significantly on the ZS-CAR task. DFM [38] was proposed along with SPM. It enables cross-modal interactions between language and visual features for better composition prediction. However, the improvements are slight on the ZS-CAR task. Compared to it, our C2C method provides more significant improvements.", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison on C-GQA. To further evaluate the proposed C2C method, we modify it for the image-based compositional zero-shot learning task by deleting all the temporal-related operations. We report the results on the recently proposed challenging benchmark C-GQA in Tab. 4, with fine-tuned ViT-B [10] as the backbone. In the table, the term 'multi' denotes the use of multi-level backbone outputs. When using multi-level outputs in C2C, we merely average them, much simpler than CoT.", + "bbox": [ + 212, + 222, + 488, + 433 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Though designed for video analysis, the proposed C2C exhibits superior performance over existing SOTA methods. Like CoT [26], we also try to utilize multi-level backbone outputs. But compared to the specially designed module in CoT, we merely average the multi-level outputs. This demonstrates that our C2C method can be a general solution to the compositional generalization problem.", + "bbox": [ + 212, + 433, + 787, + 510 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/51f290a908be2a3536cefab74079e55c2d0da19a2a292d7acb7e84d508057d11.jpg", + "image_caption": [ + "Fig. 5: Seen-unseen curves, drawn by gradually forcing the model biased from unseen to seen actions." + ], + "image_footnote": [], + "bbox": [ + 527, + 255, + 754, + 362 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 529, + 390, + 545 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Different composition generation methods. We first compare different methods that can be used to generate compositions from verb and object components, including: (1) Independent supposes components are completely independent, so the inference is straightforward: $s_{c,i} = s_{v,i} \\cdot s_{o,i}$ . (2) Knowledge-agnostic assumes components are not completely independent, but the component relations are agnostic: $s_{c,i} = s_{v,i} \\cdot 1 + s_{o,i} \\cdot 1$ . (3) One path inference considers the compatibility between compositions but only relies on either inference path, i.e., the dynamics or the static path result (Fig. 2). (4) Full is the final scheme, integrating both paths.", + "bbox": [ + 212, + 551, + 787, + 688 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For a fair comparison, the experiments only use the base losses $\\mathcal{L}_{\\text{com}}$ and $\\mathcal{L}_{\\text{comp}}$ . The results are reported in Tab. 5. We use Compcos as the baseline result. The simplest inference method, i.e., Independent, has surpassed Compcos on all the metrics. When using the knowledge-agnostic hypothesis, the seen accuracy and holistic metrics are better. But its unseen accuracy is lower. By considering the component relations, utilizing both inference paths substantially improves the unseen accuracy, yielding the best results for both HM and AUC. The results suggest that our proposed inference approach achieves a more balanced performance between seen and unseen actions. Fig. 5 reports the seen-unseen curves, implying that the proposed method is more robust.", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "R. Li et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/4c0d7ec66e06244cc3c3743541ec847ab774d262bfaa0b9e4db16fa7f333fc95.jpg", + "table_caption": [ + "Table 6: Effectiveness of the elements in the enhanced training strategy." + ], + "table_footnote": [], + "table_body": "
\\( \\mathcal{L}_{com} \\)\\( \\mathcal{L}_{comp} \\)\\( \\mathcal{L}_{ind} \\)CutMix\\( \\mathcal{L}_{con} \\)\\( \\mathcal{L}_{new} \\)HMAUC
✓29.812.5
✓✓30.813.2
✓✓✓31.413.6
✓✓✓✓31.513.9
✓✓✓✓✓32.114.0
✓✓✓✓✓✓33.014.5
", + "bbox": [ + 383, + 170, + 614, + 244 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/5e117a1b4ed0aa685ee372d5325c63306b63cae988b02898a91266e7f246aec1.jpg", + "table_caption": [ + "Table 7: The accuracy (%) on actions containing verbs that cause deformations." + ], + "table_footnote": [], + "table_body": "
methodsverbobjectcomposition
seenunseenseenunseenseenunseen
Vanilla70.160.268.734.856.927.5
Vanilla + Lind71.865.268.245.158.036.7
", + "bbox": [ + 354, + 291, + 643, + 339 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/09993f07d534ae0917c9076eea0e80e9a28afc71014ba1c5bd40a1d63dd92103.jpg", + "table_caption": [ + "Table 8: Unbiased accuracy (%) of seen and unseen actions and the corresponding harmonic mean (HM)." + ], + "table_footnote": [], + "table_body": "
methodsseenunseenHM
baseline36.817.223.5
+Lcon36.5(-0.3)16.4(-0.8)22.6(-0.9)
+cutmix35.7(-1.1)17.7(+0.5)23.6(+0.1)
+cutmix+Lcon36.1(+0.4)17.6(-0.1)23.7(+0.1)
+cutmix+Lnew37.0(+1.3)18.2(+0.5)24.4(+0.8)
+cutmix+Lcon+Lnew37.5(+0.5/+1.4)18.7(+0.5/+1.1)24.9(+0.5/+1.2)
", + "bbox": [ + 299, + 398, + 697, + 481 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The enhanced training strategy. In Tab. 6, we demonstrate the effectiveness of our enhanced training strategy. It shows that the results exhibit consistent improvements as gradually applying the proposed strategy.", + "bbox": [ + 212, + 516, + 784, + 561 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To further validate the proposed method, we verify the merits of the independent loss $L_{ind}$ on actions containing verbs that can cause object deformations (such as squeeze and twist). The results are shown in Tab. 7. They show that the unseen accuracy improves significantly after adding $L_{ind}$ . However, the object accuracy of seen actions slightly decreases. This might be because spurious features are useful for recognizing seen components.", + "bbox": [ + 212, + 565, + 787, + 655 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Next, we validate the proposed balanced learning of actions. With the vanilla version with $L_{ind}$ as the baseline, we report the accuracy of all seen and unseen actions and HM without adding bias in Tab. 8. It shows that adding only $L_{con}$ leads to performance degradation, but the degradation in seen accuracy is minor. Similarly, adding just CutMix results in a slight improvement, but with an obvious degradation in seen accuracy. When using CutMix, we add either $L_{con}$ or $L_{new}$ . The results indicate that $L_{new}$ yields a universal improvement, while $L_{con}$ contributes positively to seen accuracy. Last, when $L_{new}$ and $L_{con}$ are integrated, they both bring improvements on all the metrics, and the gains are more pronounced than adding either one alone. This indicates that our training strategy enables the model to achieve a better balance between seen and unseen actions.", + "bbox": [ + 212, + 657, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "C2C: Component-to-Composition Learning", + "bbox": [ + 442, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9525e3d9e569ed9ad921650b66893113037999dd8a211cc7da63a621d5b280e3.jpg", + "image_caption": [ + "Fig. 6: The top-3 compatible components Fig. 7: Examples of successes and failures with the same verb or object components. of our methods." + ], + "image_footnote": [], + "bbox": [ + 236, + 145, + 354, + 248 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/efd711524ac51cac2ad16c6e86d8a9e08f1cef4a5f48594eeb71095c35553d2c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 145, + 483, + 252 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/da02af2561f47b06a6167813210502738e9359cccf957c304621a965a3ef6227.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 143, + 635, + 252 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/acd696c7cf7c8d807d18bed514368e24a0a77aff00e7958fee93251f037926a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 148, + 738, + 252 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.3 Qualitative analysis", + "text_level": 1, + "bbox": [ + 215, + 324, + 424, + 339 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In Fig. 6, we show the top three compatible components for different samples, while keeping another component the same. When the verb is squeeze, for the first sample, the second and third compatible objects are cloth and paper, which are close to the visual contents and compatible with squeeze. For the second sample, the model identifies different objects. The second object basket is not included in the visual content but is compatible with squeeze. The third object bed cannot be combined with squeeze but is visually close to the sample. When the fixed component is an object, the situation is similar. It demonstrates that the proposed model can identify component compatibility according to both visual contents and component attributions.", + "bbox": [ + 212, + 358, + 787, + 508 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In Fig. 7, we visualize the top three recognized actions of some successful and failed samples. In the top row, our model accurately predicts the given actions. Besides, the second and third predictions relate to the visual content, such as fold paper for the first sample and tear paper just a little bit for the second sample. The bottom row shows some failures. Though wrong, the predictions still reflect the video content, such as move part of plant and tear leaf into two pieces for the second example. However, the failures also demonstrate some limitations of the proposed method. For example, it struggles with the task of distinguishing objects or verbs of similar visual appearances.", + "bbox": [ + 212, + 511, + 789, + 647 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 678, + 359, + 694 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper introduces a novel task, Zero-Shot Compositional Action Recognition (ZS-CAR), that investigates the compositional generalization in the video modality. A benchmarking dataset Sth-com and a baseline method C2C are proposed along with the novel task. Besides, an enhanced training strategy is presented to address two inherent challenges in ZS-CAR: component domain variations and component compatibility variations. The approach demonstrates superior performance compared to existing compositional generalization solutions in both image- and video-based zero-shot compositional learning settings.", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "R. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported in part by the National Key Research and Development Program of China (2023YFF1105102, 2023YFF1105105), the National Natural Science Foundation of China (62020106012, 62332008, 62106089, U1836218, 62336004), the 111 Project of Ministry of Education of China (B12018), the Postgraduate Research & Practice Innovation Program of Jiangsu Province (KYCX22_2307), and the UK EPSRC (EP/V002856/1, EP/T022205/1).", + "bbox": [ + 212, + 180, + 787, + 272 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 297, + 321, + 313 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Arnab, A., Dehghani, M., Heigold, G., Sun, C., Lučić, M., Schmid, C.: Vivit: A video vision transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6836-6846 (2021)", + "2. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: International Conference on Machine Learning. pp. 813-824 (2021)", + "3. Bojanowski, P., Grave, E., Joulin, A., Mikolov, T.: Enriching word vectors with subword information. Transactions of the association for computational linguistics 5, 135-146 (2017)", + "4. Brattoli, B., Tighe, J., Zhdanov, F., Perona, P., Chalupka, K.: Rethinking zero-shot video classification: End-to-end training for realistic applications. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4613-4623 (2020)", + "5. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)", + "6. Chao, W.L., Changpinyo, S., Gong, B., Sha, F.: An empirical study and analysis of generalized zero-shot learning for object recognition in the wild. In: Proceedings of the European Conference on Computer Vision. pp. 52-68 (2016)", + "7. Chen, S., Huang, D.: Elaborative rehearsal for zero-shot action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13638-13647 (2021)", + "8. Chomsky, N.: Aspects of the Theory of Syntax. No. 11, MIT press (2014)", + "9. Doshi, K., Yilmaz, Y.: Zero-shot action recognition with transformer-based video semantic embedding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4858-4867 (2023)", + "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020)", + "11. Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6824-6835 (2021)", + "12. Feichtenhofer, C.: X3d: Expanding architectures for efficient video recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 203-213 (2020)" + ], + "bbox": [ + 217, + 332, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "C2C: Component-to-Composition Learning", + "bbox": [ + 442, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6202-6211 (2019)", + "14. Fodor, J.A., Pylyshyn, Z.W.: Connectionism and cognitive architecture: A critical analysis. Cognition 28(1-2), 3-71 (1988)", + "15. Fu, Y., Hospedales, T.M., Xiang, T., Gong, S.: Attribute learning for understanding unstructured social activity. In: Proceedings of the European Conference on Computer Vision. pp. 530-543 (2012)", + "16. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The\" something something\" video database for learning and evaluating visual common sense. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5842-5850 (2017)", + "17. Gretton, A., Bousquet, O., Smola, A., Scholkopf, B.: Measuring statistical dependence with hilbert-schmidt norms. In: International Conference on Algorithmic Learning Theory. pp. 63-77 (2005)", + "18. Hao, S., Han, K., Wong, K.Y.K.: Learning attention as disentangler for compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15315-15324 (2023)", + "19. Hara, K., Kataoka, H., Satoh, Y.: Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet? In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6546-6555 (2018)", + "20. Heckman, J.J.: Sample selection bias as a specification error. Econometrica: Journal of the econometric society pp. 153-161 (1979)", + "21. Isola, P., Lim, J.J., Adelson, E.H.: Discovering states and transformations in image collections. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1383-1391 (2015)", + "22. Ji, J., Krishna, R., Fei-Fei, L., Niebles, J.C.: Action genome: Actions as compositions of spatio-temporal scene graphs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10236-10247 (2020)", + "23. Ji, S., Xu, W., Yang, M., Yu, K.: 3d convolutional neural networks for human action recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 35(1), 221-231 (2012)", + "24. Karthik, S., Mancini, M., Akata, Z.: Kg-sp: Knowledge guided simple primitives for open world compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9336-9345 (2022)", + "25. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017)", + "26. Kim, H., Lee, J., Park, S., Sohn, K.: Hierarchical visual primitive experts for compositional zero-shot learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5675-5685 (2023)", + "27. Kwon, H., Kim, M., Kwak, S., Cho, M.: Motionsqueeze: Neural motion feature learning for video understanding. In: Proceedings of the European Conference on Computer Vision. pp. 345-362 (2020)", + "28. Kwon, H., Kim, M., Kwak, S., Cho, M.: Learning self-similarity in space and time as generalized motion for video action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13065-13075 (2021)", + "29. Li, K., Wang, Y., Gao, P., Song, G., Liu, Y., Li, H., Qiao, Y.: Uniformer: Unified transformer for efficient spatiotemporal representation learning. arXiv preprint arXiv:2201.04676 (2022)" + ], + "bbox": [ + 217, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "R. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Li, R.C., Wu, X.J., Wu, C., Xu, T.Y., Kittler, J.: Dynamic information enhancement for video classification. Image and Vision Computing 114, 104244 (2021)", + "31. Li, R., Wu, X.J., Xu, T.: Video is graph: Structured graph module for video action recognition. arXiv preprint arXiv:2110.05904 (2021)", + "32. Li, R., Xu, T., Wu, X.J., Shen, Z., Kittler, J.: Perceiving actions via temporal video frame pairs. ACM Transactions on Intelligent Systems and Technology 15(3), 1-20 (2024)", + "33. Li, X., Yang, X., Wei, K., Deng, C., Yang, M.: Siamese contrastive embedding network for compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9326-9335 (2022)", + "34. Li, Y., Ji, B., Shi, X., Zhang, J., Kang, B., Wang, L.: Tea: Temporal excitation and aggregation for action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 909-918 (2020)", + "35. Lin, J., Gan, C., Han, S.: Tsm: Temporal shift module for efficient video understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7083-7093 (2019)", + "36. Liu, J., Kuipers, B., Savarese, S.: Recognizing human actions by attributes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3337-3344 (2011)", + "37. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3202-3211 (2022)", + "38. Lu, X., Guo, S., Liu, Z., Guo, J.: Decomposed soft prompt guided fusion enhancing for compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23560-23569 (2023)", + "39. Ma, L., Zheng, Y., Zhang, Z., Yao, Y., Fan, X., Ye, Q.: Motion stimulation for compositional action recognition. IEEE Transactions on Circuits and Systems for Video Technology (2022)", + "40. Mancini, M., Naeem, M.F., Xian, Y., Akata, Z.: Open world compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5222-5230 (2021)", + "41. Mandal, D., Narayan, S., Dwivedi, S.K., Gupta, V., Ahmed, S., Khan, F.S., Shao, L.: Out-of-distribution detection for generalized zero-shot action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9985-9993 (2019)", + "42. Materzynska, J., Xiao, T., Herzig, R., Xu, H., Wang, X., Darrell, T.: Something-else: Compositional action recognition with spatial-temporal interaction networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1049-1059 (2020)", + "43. Misra, I., Gupta, A., Hebert, M.: From red wine to red tomato: Composition with context. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1792-1801 (2017)", + "44. Naeem, M.F., Xian, Y., Tombari, F., Akata, Z.: Learning graph embeddings for compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 953-962 (2021)", + "45. Nayak, N.V., Yu, P., Bach, S.: Learning to compose soft prompts for compositional zero-shot learning. In: The Eleventh International Conference on Learning Representations (2022)", + "46. Neimark, D., Bar, O., Zohar, M., Asselmann, D.: Video transformer network. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3163-3172 (2021)" + ], + "bbox": [ + 212, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "C2C: Component-to-Composition Learning", + "bbox": [ + 442, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "47. Pan, J., Lin, Z., Zhu, X., Shao, J., Li, H.: St-adapter: Parameter-efficient image-to-video transfer learning. Advances in Neural Information Processing Systems 35, 26462-26477 (2022)", + "48. Park, J., Lee, J., Sohn, K.: Dual-path adaptation from image to video transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2203-2213 (2023)", + "49. Piergiovanni, A., Kuo, W., Angelova, A.: Rethinking video vits: Sparse video tubes for joint image and video learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2214-2224 (2023)", + "50. Purushwalkam, S., Nickel, M., Gupta, A., Ranzato, M.: Task-driven modular networks for zero-shot compositional learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3593-3602 (2019)", + "51. Qian, R., Lin, W., See, J., Li, D.: Controllable augmentations for video representation learning. Visual Intelligence 2(1), 1 (2024)", + "52. Qian, Y., Yu, L., Liu, W., Hauptmann, A.G.: Rethinking zero-shot action recognition: Learning from latent atomic actions. In: Proceedings of the European Conference on Computer Vision. pp. 104-120 (2022)", + "53. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "54. Ruis, F., Burghouts, G., Bucur, D.: Independent prototype propagation for zero-shot compositionality. Advances in Neural Information Processing Systems 34, 10641-10653 (2021)", + "55. Saini, N., Pham, K., Shrivastava, A.: Disentangling visual embeddings for attributes and objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13658-13667 (2022)", + "56. Sun, P., Wu, B., Li, X., Li, W., Duan, L., Gan, C.: Counterfactual debiasing inference for compositional action recognition. In: Proceedings of the ACM International Conference on Multimedia. pp. 3220-3228 (2021)", + "57. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4489-4497 (2015)", + "58. Wang, L., Tong, Z., Ji, B., Wu, G.: Tdn: Temporal difference networks for efficient action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1895-1904 (2021)", + "59. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Van Gool, L.: Temporal segment networks: Towards good practices for deep action recognition. In: Proceedings of the European Conference on Computer Vision. pp. 20-36 (2016)", + "60. Wang, Q., Liu, L., Jing, C., Chen, H., Liang, G., Wang, P., Shen, C.: Learning conditional attributes for compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11197-11206 (2023)", + "61. Xie, S., Sun, C., Huang, J., Tu, Z., Murphy, K.: Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In: Proceedings of the European Conference on Computer Vision. pp. 305–321 (2018)", + "62. Xu, T., Zhu, X.F., Wu, X.J.: Learning spatio-temporal discriminative model for affine subspace based visual object tracking. Visual Intelligence 1(1), 4 (2023)", + "63. Xu, X., Hospedales, T., Gong, S.: Transductive zero-shot action recognition by word-vector embedding. International Journal of Computer Vision 123, 309-333 (2017)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "R. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "64. Yan, R., Huang, P., Shu, X., Zhang, J., Pan, Y., Tang, J.: Look less think more: Rethinking compositional action recognition. In: Proceedings of the ACM International Conference on Multimedia. pp. 3666-3675 (2022)", + "65. Yan, R., Xie, L., Shu, X., Zhang, L., Tang, J.: Progressive instance-aware feature learning for compositional action recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)", + "66. Yang, T., Zhu, Y., Xie, Y., Zhang, A., Chen, C., Li, M.: Aim: Adapting image models for efficient video action recognition. arXiv preprint arXiv:2302.03024 (2023)", + "67. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6023-6032 (2019)", + "68. Zadrozny, B.: Learning and evaluating classifiers under sample selection bias. In: Proceedings of the International Conference on Machine Learning. p. 114 (2004)", + "69. Zellers, R., Choi, Y.: Zero-shot activity recognition with verb attribute induction. In: Proceedings of the Conference on Empirical Methods in Natural Language Processing. pp. 946-958 (2017)", + "70. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 397 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "C2C: Component-to-Composition Learning", + "bbox": [ + 442, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/C2C_ Component-to-Composition Learning for Zero-Shot Compositional Action Recognition/c79e54f7-761f-47c3-bbd9-d80fc5c9ad18_model.json b/2024/C2C_ Component-to-Composition Learning for Zero-Shot Compositional Action Recognition/c79e54f7-761f-47c3-bbd9-d80fc5c9ad18_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4c807ab7cfeb652a193f31637f3a1e2115012453 --- /dev/null +++ b/2024/C2C_ Component-to-Composition Learning for Zero-Shot Compositional Action Recognition/c79e54f7-761f-47c3-bbd9-d80fc5c9ad18_model.json @@ -0,0 +1,2713 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.225, + 0.142, + 0.777, + 0.187 + ], + "angle": 0, + "content": "C2C: Component-to-Composition Learning for Zero-Shot Compositional Action Recognition" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.212, + 0.786, + 0.244 + ], + "angle": 0, + "content": "Rongchang Li\\(^{1,2}\\), Zhenhua Feng\\(^{1,2,3,4}\\), Tianyang Xu\\(^{1}\\), Linze Li\\(^{1}\\), Xiao-Jun Wu\\(^{1*}\\), Muhammad Awais\\(^{2,4}\\), Sara Atito\\(^{2,4}\\), and Josef Kittler\\(^{2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.255, + 0.784, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Artificial Intelligence and Computer Science, Jiangnan University, China" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.269, + 0.785, + 0.283 + ], + "angle": 0, + "content": "2 Centre for Vision, Speech and Signal Processing (CVSSP), University of Surrey, UK" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.283, + 0.784, + 0.297 + ], + "angle": 0, + "content": "\\(^{3}\\) School of Computer Science and Electronic Engineering, University of Surrey, UK" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.255, + 0.785, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.297, + 0.76, + 0.326 + ], + "angle": 0, + "content": "\\(^{4}\\) Surrey Institute of People-centred AI (SI-PAI), University of Surrey, UK {li_rongchang,linze.li}@stu.jiangnan.edu.cn;" + }, + { + "type": "text", + "bbox": [ + 0.244, + 0.326, + 0.758, + 0.353 + ], + "angle": 0, + "content": "feng-zhenhua@outlook.com; {wu_xiaojun, tianyang_xu}@jiangnan.edu.cn; {muhammad.awais, sara.atito, j.kittler}@surrey.ac.uk;" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.385, + 0.741, + 0.649 + ], + "angle": 0, + "content": "Abstract. Compositional actions consist of dynamic (verbs) and static (objects) concepts. Humans can easily recognize unseen compositions using the learned concepts. For machines, solving such a problem requires a model to recognize unseen actions composed of previously observed verbs and objects, thus requiring so-called compositional generalization ability. To facilitate this research, we propose a novel Zero-Shot Compositional Action Recognition (ZS-CAR) task. For evaluating the task, we construct a new benchmark, Something-composition (Sth-com), based on the widely used Something-Something V2 dataset. We also propose a novel Component-to-Composition (C2C) learning method to solve the new ZS-CAR task. C2C includes an independent component learning module and a composition inference module. Last, we devise an enhanced training strategy to address the challenges of component variations between seen and unseen compositions and to handle the subtle balance between learning seen and unseen actions. The experimental results demonstrate that the proposed framework significantly surpasses the existing compositional generalization methods and sets a new state-of-the-art. The new Sth-com benchmark and code are available at https://github.com/RongchangLi/ZSCAR_C2C." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.661, + 0.741, + 0.689 + ], + "angle": 0, + "content": "Keywords: Zero-shot compositional action recognition \\(\\cdot\\) Action recognition \\(\\cdot\\) Compositional generalization" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.712, + 0.376, + 0.728 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.742, + 0.788, + 0.819 + ], + "angle": 0, + "content": "Humans can understand basic dynamic and static concepts (i.e., verbs and objects) in seen compositional actions and generalise the concepts to unseen action categories. For example, by seeing close a box and open a bag, we can envision close a bag even if we have never seen this before. This compositional generalization capability is a fundamental function of human intelligence, by which we can" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.39, + 0.84 + ], + "angle": 0, + "content": "* Corresponding Author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "R. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.144, + 0.661, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.235, + 0.788, + 0.264 + ], + "angle": 0, + "content": "Fig. 1: Zero-Shot Compositional Action Recognition (ZS-CAR) requires models to recognize unseen actions composed of verbs and objects observed in seen actions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.295, + 0.789, + 0.491 + ], + "angle": 0, + "content": "understand an 'infinite' number of complex concepts through assimilating the learned 'finite' concepts [8,14]. Compositional generalization has recently gained attention in the computer vision community, giving rise to the corresponding image analysis task Compositional Zero-Shot Learning (CZSL) [43]. In video analysis, [42] proposed the compositional action recognition task. However, this task only requires recognizing the dynamic component. It makes the task essentially a regular verb classification task characterized by explicit sample selection bias [20,68] regarding verb-object compositions, rather than verifying compositional generalization ability. Exploring the compositional generalization problem in the video modality is valuable as it presents a challenging video understanding task that requires comprehending both dynamic and static components while also generalising to novel compositions. Also, it provides a reliable way to develop open-scenario-related applications in autonomous vehicles, robots, etc." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.789, + 0.749 + ], + "angle": 0, + "content": "To facilitate the related research, we propose a novel task, namely Zero-Shot Compositional Action Recognition (ZS-CAR). We define a compositional action as the composition of a dynamic component verb (e.g., open, pick up, ...) and a static component object (e.g., door, wallet, ...). As shown in Fig. 1, ZS-CAR aims to enable machines to identify unseen action categories consisting of seen verbs and objects. The new proposed task poses a significant challenge. First, a model has to understand the intricate spatial and temporal information to identify the verb and object components, which is more challenging than the existing action recognition task. Second, the task naturally contains component domain variation and component compatibility variation. The first one is similar to the challenges in domain generalization, caused by the contextuality difference (e.g., transferring knowledge of book in close a book to recognize bend a book). The second one stems from the action category difference in the training and test domains. Overly accepting or rejecting verb-object relationships observed during training can detrimentally impact the generalization capability to recognize unseen actions. Given these challenges, the existing visual compositional generalization solutions cannot handle the ZS-CAR task well, as shown in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.792, + 0.841 + ], + "angle": 0, + "content": "To solve ZS-CAR, we present an innovative Component-to-Composition (C2C) method to serve as the baseline. The pipeline of C2C is shown in Fig. 2. C2C first independently learns the dynamic and static components (i.e., verbs and objects). Then it employs the component consensus to recognize compositions (i.e., actions) from both dynamic and static perspectives, ensuring that both these two components are well learned. The procedure is motivated by the human process" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.444, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "C2C: Component-to-Composition Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.253 + ], + "angle": 0, + "content": "of assembling compositions. It involves an independent observation of a component, assessing its compatibility with another part, and gradually assembling the units into the entire structure. Moreover, C2C enables explicitly addressing the component domain variation and component compatibility variation problems. Accordingly, we propose an enhanced training strategy that involves minimizing the dependence on spurious information within visual features and balancing the emphasis between fitting seen and imagining unseen actions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.785, + 0.329 + ], + "angle": 0, + "content": "Last, to enable the training of C2C and to evaluate its performance for the ZS-CAR task, we create a new dataset, Something-composition (Sth-com), based on the popular Something-Something V2 (Sth-V2) dataset [16]. The experimental results demonstrate that the vanilla or enhanced version of C2C achieves state-of-the-art results. To summarize, the main contributions of the paper include:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.336, + 0.784, + 0.365 + ], + "angle": 0, + "content": "- A novel zero-shot compositional action recognition (ZS-CAR) task that investigates compositional generalization in the video modality." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.366, + 0.785, + 0.41 + ], + "angle": 0, + "content": "- A benchmarking dataset for evaluating ZS-CAR. We elaborate on the challenges in ZS-CAR and demonstrate that the existing visual compositional generalization methods cannot handle such problems well." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.411, + 0.785, + 0.47 + ], + "angle": 0, + "content": "- A novel Component-to-Composition (C2C) method as the vanilla solution and an enhanced training strategy to address the identified challenges in ZS-CAR. The experimental results demonstrate that C2C solves these challenges effectively and attains SOTA performances." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.336, + 0.785, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.493, + 0.383, + 0.509 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.785, + 0.658 + ], + "angle": 0, + "content": "Video action recognition aims to identify actions in videos. The main challenge is extracting discriminative action features from videos. Previous CNN-based methods address this challenge in two primary ways: using 2D CNNs with separate temporal operations [27,28,30-32,34,35,58,59,62], or jointly using 3D CNNs [5,12,13,19,23,51,57,61] to dig out spatiotemporal features. Recently, Transformer-based models transform a video into 3D tokens and model the spatiotemporal correlations through attention layers [1,2,11,29,37,46,49], or adapt the visual image encoder in large vision-language models to the video modality [47,48,66]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.785, + 0.78 + ], + "angle": 0, + "content": "Zero-shot action recognition (ZSAR) aims to recognize action categories not appearing in the training set. A widely used idea is to measure video features and action embeddings in a common semantic space, hoping this space can generalise to unseen actions. The semantic embeddings are obtained by side semantic information, e.g., word embeddings [4, 41, 63, 69], multiple attributes [15, 36], elaborative rehearsals [7] and action descriptions [9, 52]. Different from ZSAR, the proposed ZS-CAR aims to explore the visual compositionality of actions and generalize the learned composition knowledge to unseen actions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Compositional action recognition: The compositional nature of actions has recently received increasing attention. Action genome [22] takes action dynamics as the relationship changes and decomposes actions into spatiotemporal scene graphs. Something-else [42] proposes compositional action recognition that" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.127 + ], + "angle": 0, + "content": "R. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "only requires recognising verbs in unseen verb-object compositions. Subsequent studies [39,56,64,65] follow this setting and focus on extracting robust dynamic information. However, [42] essentially is a supervised task with explicit sample selection bias, rather than a compositional generalization problem." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.329 + ], + "angle": 0, + "content": "Image-based compositional generalization: In the image modality, [43] first introduces the Compositional Zero-Shot Learning (CZSL) task that requires models trained on seen attribute-object pairs to recognize unseen ones. A common idea is to find the nearest holistic composition embeddings for input images in a common space [40, 43, 44, 50, 54]. Some studies disentangle visual features to align with component embeddings [18, 24, 26, 33, 54, 55]. Recent works propose to adapt the vision-language model [53] to CZSL by text prompting methods [38, 45] and enhancing the interactions between language and visual features [38]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.357, + 0.596, + 0.373 + ], + "angle": 0, + "content": "3 Problem Definition and Benchmark" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.392, + 0.498, + 0.407 + ], + "angle": 0, + "content": "3.1 The Proposed ZS-CAR Task" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.785, + 0.572 + ], + "angle": 0, + "content": "In traditional action recognition, a model predicts the action label defined as either a verb or a verb-object composition. For example, the Sth-V2 dataset [16] defines verb-only actions, such as open sth and pick up sth. Kinetics-400 [25] defines actions as both verb-object compositions and verb-only actions, such as play basketball, draw, etc. The recently proposed compositional action recognition task considers the interaction between verbs and objects. It requires recognising verbs of which the paired objects in the test set differ from those in the training set. Though introducing action compositionally by this constraint, this task only takes the verb-object relation as a sample selection bias controller, making it not a thorough compositional generalization problem." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.574, + 0.785, + 0.755 + ], + "angle": 0, + "content": "In contrast to previous tasks, we introduce the compositional generalization perspective to understand actions, by proposing a novel Zero-Shot Compositional Action Recognition (ZS-CAR) task. We explicitly define a compositional action as a concept of verb and object. To verify the compositional generalization ability, ZS-CAR requires models to recognize unseen action categories composed of verbs and objects observed in seen actions. Formally, we denote the training set as \\( D_{\\mathrm{train}} = \\{(\\mathcal{X}_i,a_i)\\}_{i = 1}^N \\) consisting of video \\( \\mathcal{X} \\) labelled as action \\( a\\in A_{\\mathrm{train}} \\), with \\( A_{\\mathrm{train}}\\subset V\\times O = \\{(v,o)|v\\in V,o\\in O\\} \\), where \\( V \\) and \\( O \\) are the verb and object sets. The validation and test sets are similarly denoted as \\( D_{\\mathrm{val}} \\) and \\( D_{\\mathrm{test}} \\), with label sets as \\( A_{\\mathrm{val}} \\) and \\( A_{\\mathrm{test}} \\). We follow the generalised zero-shot learning setting [6,50], i.e., both seen and unseen action categories appear in the validation/test set: \\( A_{\\mathrm{train}}\\cap A_{\\mathrm{val / test}}\\neq \\emptyset \\) and \\( \\exists a(a\\in A_{\\mathrm{val / test}}\\oplus a\\in A_{\\mathrm{train}}) \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.781, + 0.592, + 0.796 + ], + "angle": 0, + "content": "3.2 The Something-composition Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.841 + ], + "angle": 0, + "content": "To assess the proposed task, we build a benchmark Something-composition (Sth-com), based on the widely used Something-Something V2 (Sth-V2) dataset [16]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.444, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "C2C: Component-to-Composition Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "table_caption", + "bbox": [ + 0.312, + 0.145, + 0.692, + 0.16 + ], + "angle": 0, + "content": "Table 1: Statistics of the proposed Sth-com benchmark." + }, + { + "type": "table", + "bbox": [ + 0.342, + 0.172, + 0.658, + 0.23 + ], + "angle": 0, + "content": "
# Objects# Verbs# Verb-Object Actions# Sample
Train248161345138034
Val225157733 Seen + 717 Unseen18774
Test233161976 Seen + 956 Unseen22657
All248161512479465
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.258, + 0.784, + 0.302 + ], + "angle": 0, + "content": "Sth-V2 has 220,847 videos of humans performing pre-defined compositional actions with everyday objects. It is widely used by the action recognition community, as its verbs cannot be identified from a single frame." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.303, + 0.785, + 0.455 + ], + "angle": 0, + "content": "Annotation: We use the training and test split of Sth-else [42], a subset of Sth-V2, as the starting point. Then, we iteratively clean the subsets and ensure that: (1) Each composition (i.e., action) has more than five samples. (2) The components (i.e., verb & object) present in the test set are seen in the training set. This yields 248 objects, 161 verbs, 5124 compositions and 79465 video samples. Next, we select \\(1/3\\) of compositions in the training and test sets respectively and interchange \\(50\\%\\) of the selected compositions, thereby introducing seen action categories to the test set. Subsequently, following the ratio (3:4) in MIT-State [21], we divide the test set into the validation and test sets, respectively. The statistics of Sth-com are summarised in Tab. 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.477, + 0.614, + 0.496 + ], + "angle": 0, + "content": "4 Component-to-Composition Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.645 + ], + "angle": 0, + "content": "To address the proposed ZS-CAR task, we propose a novel Component-to-Composition (C2C) learning framework, as shown in Fig. 2, to serve as a vanilla solution. We simulate the process of human assembly to infer compositions. We commence with a specific component and scrutinise the compatibility of various categories of another component, leading to a reasonable composition proposal. These are correspondingly implemented in the proposed independent component learning module and composition to component learning module. Importantly, we analyse the challenges in ZS-CAR and devise an enhanced training strategy to facilitate generalising to unseen actions better." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.665, + 0.547, + 0.682 + ], + "angle": 0, + "content": "4.1 Independent Component Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.844 + ], + "angle": 0, + "content": "In this module, our goal is to identify the verb/object score of an input video and acquire their semantic prototypes. Specifically, we first encode the raw video sequence \\(\\mathcal{X}\\) with shape \\(T\\times 224\\times 224\\times 3\\) to a general representation \\(\\mathbf{F}_x\\in \\mathbb{R}^{T\\times D}\\), where \\(T\\) and \\(D\\) denote the frame and channel numbers, respectively. Then, we respectively adopt dynamic and static encoders to extract dynamic and static features, denoted as \\(\\mathbf{f}_v\\) and \\(\\mathbf{f}_o\\in \\mathbb{R}^C\\), where \\(C\\) is the channel number. The static encoder consists of a temporal average pooling layer and two MLP layers, while the dynamics encoder contains two temporal convolution layers and a temporal pooling layer. We encode the component labels as the general representations of verbs and objects, denoted as the component prototypes \\(\\mathbf{E}_v = \\{e_{v,i}\\}_{i = 1}^{N_v}\\) and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "R. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.147, + 0.77, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.39, + 0.788, + 0.502 + ], + "angle": 0, + "content": "Fig. 2: The proposed Component-to-Composition (C2C) learning framework. C2C first aligns verb/object prototypes with corresponding visual features to obtain component scores in the Independent Component Learning module. Then, actions are inferred through two paths (dynamics and static) in the Component to Composition module. In the dynamics path, verb prototypes and visual features are used to compute conditional object scores. Then the independent verb scores and conditional object scores are multiplied to gain the action scores. The static path follows a similar procedure. The final output is a consensus of the results from both paths." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.527, + 0.788, + 0.589 + ], + "angle": 0, + "content": "\\(\\mathbf{E}_o = \\{e_{o,i}\\}_{i = 1}^{N_o}\\), where \\(N_{v}\\) and \\(N_{o}\\) are the verb and object numbers, respectively. The prototypes are learned by being aligned with the visual features \\(\\mathbf{f}_v\\) and \\(\\mathbf{f}_o\\) using cosine similarity as the gauge. Denoting the action label of the input video as \\(a_{i} = (v_{l},o_{k})\\), the optimization objective is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.596, + 0.787, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {v e r b}} = - \\log \\frac {\\exp \\left(\\cos \\left(\\mathbf {f} _ {v} , \\mathbf {e} _ {v , l}\\right) / \\tau\\right)}{\\sum_ {v _ {j} \\in V} \\exp \\left(\\cos \\left(\\mathbf {f} _ {v} , \\mathbf {e} _ {v , j}\\right) / \\tau\\right)}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.638, + 0.787, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {o b j} = - \\log \\frac {\\exp \\left(\\cos \\left(\\mathbf {f} _ {o} , \\mathbf {e} _ {o , k}\\right) / \\tau\\right)}{\\sum_ {o _ {j} \\in O} \\exp \\left(\\cos \\left(\\mathbf {f} _ {o} , \\mathbf {e} _ {o , j}\\right) / \\tau\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.679, + 0.788, + 0.77 + ], + "angle": 0, + "content": "where \\(\\mathbf{e}_{v,l}\\) and \\(\\mathbf{e}_{o,k}\\) are the corresponding prototypes. \\(\\cos(,)\\) is the cosine similarity. \\(\\tau\\) is the temperature coefficient for better optimization. We denote \\(\\mathcal{L}_{comp} = \\mathcal{L}_{verb} + \\mathcal{L}_{obj}\\) as the component loss. By aligning with diverse component features in the training set, the prototypes inherently acquire semantic meanings. Accordingly, the verb and object scores, \\(\\mathbf{s}_v\\) and \\(\\mathbf{s}_o\\), are the cosine similarities between the visual features and corresponding component prototypes." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.789, + 0.533, + 0.804 + ], + "angle": 0, + "content": "4.2 From component to composition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "In this module, we infer the action (composition) score \\(\\mathbf{s}_a \\in \\mathbb{R}^{N_a}\\) using the prototypes and the verbs and objects (components) scores. We define the conditional" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.444, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "C2C: Component-to-Composition Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.359 + ], + "angle": 0, + "content": "scores \\(\\mathbf{S}_{o|v} \\in \\mathbb{R}^{N_v \\times N_o}\\) and \\(\\mathbf{S}_{v|o} \\in \\mathbb{R}^{N_o \\times N_v}\\), that represents the confidence of one component given the visual input and another component as the conditions. Accordingly, for one action \\(a_i = (v_l, o_k)\\), its confidence score can be calculated by either \\(s_{a,i} = s_{v,l} \\cdot s_{o = o_k|v = v_l}\\) or \\(s_{a,i} = s_{o,k} \\cdot s_{v = v_l|o = o_k}\\). The first one relies on the verbs to infer actions, called the dynamics path. Similarly, the other one is called the static path. As shown in Fig. 2, in the dynamics path, we first encode \\(\\mathbf{F}_x\\) using a static encoder to provide the visual reference for the object component. Then we concatenate the output with the verb prototype \\(\\mathbf{e}_{v,l}\\) and fuse them to the compound prototype-visual feature \\(\\mathbf{f}_{v - x,l}\\), using an MLP layer. Then we linearly map the cosine similarity \\(\\cos(\\mathbf{f}_{v - x,l}, \\mathbf{e}_{o,k})\\) to the range of [0,1], obtaining the conditional object score \\(s_{o = o_k|v = v_l}\\). The static path follows a similar process except for encoding \\(\\mathbf{F}_x\\) by the dynamics encoder. Last, we average the inference results from the two paths to obtain the final inference output. Taking \\(s_{a,i}\\) as the score of the groundtruth action \\(a_i\\), the entire model is optimized using:" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.365, + 0.786, + 0.401 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c o m} = - \\log \\frac {\\exp \\left(s _ {a , i} / \\tau\\right)}{\\sum_ {a _ {j} \\in A _ {\\text {t r a i n}}} \\exp \\left(s _ {a , j} / \\tau\\right)}, \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.418, + 0.486, + 0.433 + ], + "angle": 0, + "content": "4.3 Enhanced training strategy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.432, + 0.635 + ], + "angle": 0, + "content": "Risk in generalizing components. Naturally, there is a high probability of exhibiting different component appearances between training and test domains, namely component domain variation. As shown in Fig. 3, the component 'book' exhibits a notable difference in appearance between 'close a book' and 'bend a book'. Besides, given the combination of 'close' and" + }, + { + "type": "image", + "bbox": [ + 0.448, + 0.469, + 0.768, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.44, + 0.559, + 0.788, + 0.629 + ], + "angle": 0, + "content": "Fig. 3: Component domain variations. For unseen actions, an object may exhibit different appearances (left). To deal with this (right), we reduce the spurious information and enlarge the independence between component-specific features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.635, + 0.788, + 0.681 + ], + "angle": 0, + "content": "'book' in the training set, there is a tendency that the model might align features associated with 'close' to 'book', or vice versa. These dependencies may confuse the model when encountering unseen action categories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.681, + 0.788, + 0.817 + ], + "angle": 0, + "content": "To cope with it, in the independent component learning module, we posit that the ideal dynamic-focused feature \\(\\mathbf{f}_v\\) and static-focused feature \\(\\mathbf{f}_o\\) should encompass two types of information: common and specific. The former denotes features pertinent to both components and the latter denotes the invariant features, that are crucial for identifying the respective components. Notably, the invariant features of verbs are spurious for identifying objects, and vice versa. As shown in Fig. 3, our solution is to minimize the spurious information in \\(\\mathbf{f}_v\\) and \\(\\mathbf{f}_o\\), and explicitly reduce the mutual associations between their component-specific parts. To achieve this, we propose the following auxiliary losses:" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.825, + 0.787, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\sup , v e r b} = h \\left(\\mathbf {f} _ {x}, \\mathbf {f} _ {v}\\right) - h \\left(\\mathbf {f} _ {v}, \\mathbf {y} _ {v}\\right), \\tag {4}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.128 + ], + "angle": 0, + "content": "R. Li et al." + }, + { + "type": "equation", + "bbox": [ + 0.392, + 0.162, + 0.786, + 0.18 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s u p, o b j} = h \\left(\\mathbf {f} _ {x}, \\mathbf {f} _ {o}\\right) - h \\left(\\mathbf {f} _ {o}, \\mathbf {y} _ {o}\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.199, + 0.786, + 0.217 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i n d} = \\mathcal {L} _ {s u p, v e r b} + \\mathcal {L} _ {s u p, o b j} + h \\left(\\mathbf {f} _ {v} ^ {\\prime}, \\mathbf {f} _ {o} ^ {\\prime}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.222, + 0.788, + 0.344 + ], + "angle": 0, + "content": "where \\( h(,.) \\) is the Hilbert-Schmidt Independence Criterion (HSIC) [17] for independence measure. \\( h(\\mathbf{f},\\mathbf{h}) = 0 \\) if and only if \\( \\mathbf{f} \\perp \\mathbf{h} \\). \\( \\mathbf{f}_x \\) is the temporal pooling result of \\( \\mathbf{F}_x \\). \\( \\mathbf{y}_v \\) and \\( \\mathbf{y}_o \\) are the one-hot label vectors. \\( \\mathcal{L}_{sup,verb} \\) and \\( \\mathcal{L}_{sup,obj} \\) force the model to compress information, but retain sufficient information to identify the components, thus reducing the general spurious information. \\( \\mathbf{f}_v' \\) or \\( \\mathbf{f}_o' \\) is part of \\( \\mathbf{f}_v \\) or \\( \\mathbf{f}_o \\), defined as \\( \\mathbf{f}_v' = \\mathbf{f}_v[: \\rho C], \\mathbf{f}_o' = \\mathbf{f}_o[: \\rho C] \\) (\\( 0 \\leq \\rho \\leq 1 \\)). They are seen as component-specific features. So the term \\( h(\\mathbf{f}_v',\\mathbf{f}_o') \\) ensures the independence of the invariant part of the verb and object features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.479, + 0.524 + ], + "angle": 0, + "content": "Balanced learning of seen & unseen actions. In ZS-CAR, the verb-object connections in the training and test sets differ significantly, as shown in Fig. 4. While the seen connections help optimize the modules, an excessive emphasis on them may hinder generalizing to unseen actions. Our solution is simultaneously enhancing the learning of seen and unseen connections, enabling them to strike an optimal balance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.525, + 0.478, + 0.568 + ], + "angle": 0, + "content": "We first assist the model in learning seen actions explicitly. We calculate the observed conditional" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.618 + ], + "angle": 0, + "content": "probabilities in the training set as the observed conditional scores \\(\\hat{\\mathbf{S}}_{v|o}\\) and \\(\\hat{\\mathbf{S}}_{o|v}\\). Then, we force the conditional scores in a training batch closer to \\(\\hat{\\mathbf{S}}_{v|o}\\) and \\(\\hat{\\mathbf{S}}_{o|v}\\), with the objectives:" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.628, + 0.787, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n}, v} = - \\frac {1}{N _ {v}} \\sum_ {v _ {i} \\in V} \\hat {\\mathbf {s}} _ {o | v = v _ {i}} \\log \\mathbf {s} _ {o | v = v _ {i}}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.677, + 0.787, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n}, o} = - \\frac {1}{N _ {o}} \\sum_ {o _ {i} \\in O} \\hat {\\mathbf {s}} _ {v | o = o _ {i}} \\log \\mathbf {s} _ {v | o = o _ {i}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.766 + ], + "angle": 0, + "content": "where \\(N_v\\) & \\(N_o\\) denote the verb & object numbers, and the condition loss \\(\\mathcal{L}_{con}\\) is defined as the sum of \\(\\mathcal{L}_{con,v}\\) and \\(\\mathcal{L}_{con,o}\\). \\(\\hat{\\mathbf{s}}_{o|v = v_i}\\) or \\(\\mathbf{s}_{o|v = v_i}\\) denotes the conditional object score vector when verb is \\(v_i\\). \\(\\hat{\\mathbf{s}}_{v|o = o_i}\\) and \\(\\mathbf{s}_{v|o = o_i}\\) are defined similarly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.767, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Then, inspired by the human ability to create new things by imagining compositions of known components, we use CutMix [67] to assemble samples in a training batch as the imagined unseen actions, thereby reducing the risk of overemphasizing seen compositions. Specifically, for an input \\(\\mathcal{X}_i\\) with label \\(a_i = (v_l,o_k)\\), we randomly select another sample \\(\\mathcal{X}_j\\) with label \\(a_j = (v_m,o_n)\\)" + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.373, + 0.788, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.486, + 0.453, + 0.788, + 0.551 + ], + "angle": 0, + "content": "Fig. 4: Component compatibility variations. The component relations are different between the training and test sets (left). To solve this (right), we use the observed conditional score to fit seen relations and encourage the model to imagine unseen actions to avoid being limited to seen relations." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.444, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "C2C: Component-to-Composition Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 2: A comparison with the SOTA methods on Sth-com, using the vision model with word embedding model paradigm. We report both vanilla (only \\(\\mathcal{L}_{com}\\) and \\(\\mathcal{L}_{comp}\\)) and enhanced (using the advocated enhanced training strategy) versions of C2C. Since ADE is tailored for transformer architectures, we only adapt it to VideoSwin." + }, + { + "type": "table", + "bbox": [ + 0.259, + 0.213, + 0.744, + 0.334 + ], + "angle": 0, + "content": "
TSM-18VideoSwin-T
Methodsverbobj.seenunseenHMAUCverbobj.seenunseenHMAUC
TMN [50]41.435.931.628.122.07.333.445.631.030.223.07.9
Compcos [40]45.840.034.835.127.510.636.750.534.237.827.210.9
CGE [44]35.536.427.625.520.05.931.848.131.230.323.17.9
OADis [55]39.140.733.532.325.59.341.648.938.138.330.212.6
ADE [18]------35.349.133.836.226.810.3
CAN [60]40.039.232.431.624.58.732.649.533.033.325.09.2
C2C (Vanilla)47.742.738.139.130.813.247.950.942.143.634.916.4
C2C (Enhanced)50.144.339.441.033.014.548.852.843.444.136.717.4
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.363, + 0.788, + 0.516 + ], + "angle": 0, + "content": "and crop it at a random position of random size. The cropped part replaces the same position of \\(\\mathcal{X}_i\\), generating the augmented input \\(\\bar{\\mathcal{X}}_i\\). Denoting the cropping ratio as \\(\\lambda\\), we modify the previous losses as: \\(\\bar{\\mathcal{L}}_{com} = (1 - \\lambda)\\mathcal{L}_{com}^{a_i} + \\lambda \\mathcal{L}_{com}^{a_j}\\) and \\(\\bar{\\mathcal{L}}_{comp} = (1 - \\lambda)\\mathcal{L}_{comp}^{v_l,o_k} + \\lambda \\mathcal{L}_{comp}^{v_m,o_n}\\). \\(\\mathcal{L}_{comp}^{a_i}\\) calculates \\(\\mathcal{L}_{com}\\) using Eq. (3) with \\(a_i\\) as groundtruth. \\(\\mathcal{L}_{comp}^{v_l,o_k}\\) calculates \\(\\mathcal{L}_{comp}\\) with \\(v_l\\) and \\(o_k\\) as groundtruth. \\(\\mathcal{L}_{comp}^{a_j}\\) and \\(\\mathcal{L}_{comp}^{v_m,o_n}\\) are defined similarly. Also, \\(\\mathcal{L}_{ind}\\) is modified to \\(\\bar{\\mathcal{L}}_{ind}\\) by replacing \\(\\mathbf{y}_i\\) to \\((1 - \\lambda)\\mathbf{y}_i + \\lambda \\mathbf{y}_j\\) in Eq. (4) and Eq. (5). Particularly, we introduce new compositions \\(\\bar{a}_1 = (v_l,o_n)\\) and \\(\\bar{a}_2 = (v_m,o_k)\\). \\(\\bar{a}_1\\) and \\(\\bar{a}_2\\) are not constrained in \\(A_{\\mathrm{test}}\\) or \\(A_{\\mathrm{train}}\\). It gives a novel loss \\(\\mathcal{L}_{new} = \\mathcal{L}_{com}^{\\bar{a}_1} + \\mathcal{L}_{com}^{\\bar{a}_2}\\), which forces the model to consider all the possible component connections." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.517, + 0.788, + 0.609 + ], + "angle": 0, + "content": "A concern is that the new compositions \\( a_1 \\) or \\( a_2 \\) may be counterfactual. However, it only introduces noise to the component-to-composition module, as the component labels still match the visual input. In this way, we still reserve these counterfactual labels as they can provide regularization effects that avoid overfitting to seen verb-object relations. Besides, it is also consistent with human imagination which is not only constrained within reasonable compositions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.632, + 0.367, + 0.646 + ], + "angle": 0, + "content": "4.4 More details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.689 + ], + "angle": 0, + "content": "Network details. Following previous practice, we implement the video encoder and component label encoding with two paradigms:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.749 + ], + "angle": 0, + "content": "(1) Vision model with word embedding model. In this paradigm, we adopt an action recognition model (the CNN-based TSM-18 [35] or the transformer-based VideoSwin-T [37]) to extract the general video feature and a word embedding model (fasttext [3]) to encode the labels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.841 + ], + "angle": 0, + "content": "(2) Vision-language models. Recently, large pre-trained vision-language models, such as CLIP [53], have shown superior power. In this paradigm, we adopt the language branch of recent prompt-based compositional generalization methods (Coop [70], CSP [45] and SPM [38]) to encode the labels. As CLIP are trained using image-text pairs, we apply image-to-video adapters introduced in AIM [66] to adapt the CLIP visual encoder (ViT-B/32 [10]) to video modality." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.128 + ], + "angle": 0, + "content": "R. Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.253, + 0.145, + 0.748, + 0.159 + ], + "angle": 0, + "content": "Table 3: A comparison with the SOTA CLIP-based methods on Sth-com." + }, + { + "type": "table", + "bbox": [ + 0.356, + 0.171, + 0.645, + 0.304 + ], + "angle": 0, + "content": "
methodsverbobj.seenunseenHMAUC
Coop [70]48.455.143.946.436.618.1
+C2C(vanilla)56.256.948.354.642.724.0
+C2C(enhance)56.358.950.756.644.526.0
CSP [45]49.054.843.647.436.018.0
+C2C(vanilla)56.356.649.053.442.523.8
+C2C(enhance)57.158.649.956.244.325.5
SPM [38]48.954.544.046.335.517.7
+DFM [38]47.455.443.147.335.817.9
+C2C(vanilla)55.757.349.353.442.223.8
+C2C(enhance)56.558.449.955.743.825.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.332, + 0.789, + 0.378 + ], + "angle": 0, + "content": "Training Objectives. When applying the proposed enhanced training strategy, we perform CutMix to generate new compositions with a probability \\( p \\). If apply CutMix, the final loss is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.348, + 0.387, + 0.786, + 0.406 + ], + "angle": 0, + "content": "\\[\n\\bar {\\mathcal {L}} _ {t o t a l} = \\bar {\\mathcal {L}} _ {c o m} + \\alpha \\bar {\\mathcal {L}} _ {c o m p} + \\beta \\bar {\\mathcal {L}} _ {i n d} + \\gamma \\mathcal {L} _ {n e w}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.415, + 0.383, + 0.43 + ], + "angle": 0, + "content": "If not, the final loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.35, + 0.442, + 0.785, + 0.459 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c o m}} + \\alpha \\mathcal {L} _ {\\text {c o m p}} + \\beta \\mathcal {L} _ {\\text {i n d}} + \\gamma \\mathcal {L} _ {\\text {c o n}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.786, + 0.499 + ], + "angle": 0, + "content": "As \\(\\mathcal{L}_{new}\\) and \\(\\mathcal{L}_{con}\\) respectively guide the model focus on unseen and seen actions, we use the same coefficient \\(\\gamma\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.522, + 0.377, + 0.54 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.688 + ], + "angle": 0, + "content": "Evaluation Metrics. The metrics include Verb, Object, Seen, Unseen, HM and AUC. Verb and Object respectively denote the accuracy of verbs and objects. To gain other metrics, we first add a constant to the seen actions scores. The constant ranges from negative infinity to positive infinity, making the model shift from completely biased to unseen actions to completely biased to seen actions and yielding the seen-unseen curve shown in Fig. 5. Seen and Unseen are the best accuracy of seen/unseen actions, i.e. the intersection points with x/y-axis in Fig. 5. HM is the best harmonic mean with formulation: \\( HM = 2(Seen * Unseen) / (Seen + Unseen) \\). AUC is the area under the curve." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Implementation details. \\(\\alpha\\), \\(\\beta\\) \\(\\gamma\\) in Eq. (9) and Eq. (10) are 0.2, 0.1 and 0.1, respectively. The probability \\(p\\) of conducting CutMix is 0.7. More coefficient ablations and implementation details are reported in supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.755, + 0.551, + 0.772 + ], + "angle": 0, + "content": "5.1 Comparisons with State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Note that the existing compositional action recognition methods cannot apply to the ZS-CAR task. To comprehensively demonstrate the proposed C2C, we first adapt the existing CZSL methods to the proposed Sth-com dataset. We replace their image encoders with video encoders and adopt the same word" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.444, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "C2C: Component-to-Composition Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.297, + 0.145, + 0.704, + 0.159 + ], + "angle": 0, + "content": "Table 4: A comparison with the SOTA methods on C-GQA." + }, + { + "type": "table", + "bbox": [ + 0.378, + 0.171, + 0.621, + 0.248 + ], + "angle": 0, + "content": "
multiseenunseenHMAUC
CGE [44]×38.017.118.55.4
OADis [55]×38.319.820.17.0
CoT [26]✓39.222.722.17.4
C2C×39.123.023.07.7
C2C✓39.423.623.38.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.228, + 0.265, + 0.772, + 0.279 + ], + "angle": 0, + "content": "Table 5: Results of different hypotheses inferring compositions from components." + }, + { + "type": "table", + "bbox": [ + 0.342, + 0.292, + 0.658, + 0.38 + ], + "angle": 0, + "content": "
verbobj.seenunseenHMAUC
Compcos [40]45.840.034.835.127.510.6
Independent47.942.237.437.429.812.2
Knowledge-agnostic48.443.138.736.830.212.6
Dynamic centered46.742.537.438.029.912.5
Static centered46.641.137.036.728.911.7
Full47.742.738.139.130.813.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.785, + 0.471 + ], + "angle": 0, + "content": "embedding model, making sure only the compositional generalization parts are different between the adapted methods and our proposed methods. Besides, we also modify C2C and compare it with the existing SOTA methods on the image-based CZSL task using the challenging C-GQA [44] dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.785, + 0.55 + ], + "angle": 0, + "content": "Comparisons on Sth-com. We first compare our method with the adapted CZSL methods of the vision model with the word embedding paradigm, using both CNN-based and transformer-based video encoders. The results are reported in Tab. 2. The vanilla version of C2C uses two basic losses. The enhanced version uses the proposed enhanced training strategy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.787, + 0.793 + ], + "angle": 0, + "content": "Tab. 2 demonstrate that our method outperforms the existing methods by significant margins across all the metrics on ZS-CAR. Furthermore, the proposed enhanced training strategy provides consistent performance gains across different video encoders. Tab. 2 shows that the recently proposed methods based on Compcos are inferior to Compcos. Second, except for OADis, using the recently proposed VideoSwin-T backbone does not yield significant improvements. The first observation may stem from failing to generalise the embedding relations, not disentangling spatial and temporal information well, and not applying the conditional component relations in actions. Besides, jointly learning the compositions may lead to unbalanced learning of the components. These deficiencies degrade the effectiveness of video encoders, leading directly to the second observation. Tab. 2 also compares two video encoders. Though VideoSwin is an advanced model, its verb accuracy is consistently inferior or similar to TSM across all methods. Conversely, it outperforms TSM in object accuracy. This implies that our ZS-CAR task offers a new avenue for assessing biases regarding dynamic and static patterns in action recognition models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Then we apply our C2C method to recently proposed CLIP-based compositional zero-shot learning methods. The results are reported in Tab. 3. Coop, CSP and SPM are three text prompting methods. We can see that the proposed" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.128 + ], + "angle": 0, + "content": "R. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "C2C method can improve their performance significantly on the ZS-CAR task. DFM [38] was proposed along with SPM. It enables cross-modal interactions between language and visual features for better composition prediction. However, the improvements are slight on the ZS-CAR task. Compared to it, our C2C method provides more significant improvements." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.489, + 0.434 + ], + "angle": 0, + "content": "Comparison on C-GQA. To further evaluate the proposed C2C method, we modify it for the image-based compositional zero-shot learning task by deleting all the temporal-related operations. We report the results on the recently proposed challenging benchmark C-GQA in Tab. 4, with fine-tuned ViT-B [10] as the backbone. In the table, the term 'multi' denotes the use of multi-level backbone outputs. When using multi-level outputs in C2C, we merely average them, much simpler than CoT." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.511 + ], + "angle": 0, + "content": "Though designed for video analysis, the proposed C2C exhibits superior performance over existing SOTA methods. Like CoT [26], we also try to utilize multi-level backbone outputs. But compared to the specially designed module in CoT, we merely average the multi-level outputs. This demonstrates that our C2C method can be a general solution to the compositional generalization problem." + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.256, + 0.756, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.376, + 0.788, + 0.418 + ], + "angle": 0, + "content": "Fig. 5: Seen-unseen curves, drawn by gradually forcing the model biased from unseen to seen actions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.53, + 0.391, + 0.546 + ], + "angle": 0, + "content": "5.2 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Different composition generation methods. We first compare different methods that can be used to generate compositions from verb and object components, including: (1) Independent supposes components are completely independent, so the inference is straightforward: \\( s_{c,i} = s_{v,i} \\cdot s_{o,i} \\). (2) Knowledge-agnostic assumes components are not completely independent, but the component relations are agnostic: \\( s_{c,i} = s_{v,i} \\cdot 1 + s_{o,i} \\cdot 1 \\). (3) One path inference considers the compatibility between compositions but only relies on either inference path, i.e., the dynamics or the static path result (Fig. 2). (4) Full is the final scheme, integrating both paths." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "For a fair comparison, the experiments only use the base losses \\(\\mathcal{L}_{\\text{com}}\\) and \\(\\mathcal{L}_{\\text{comp}}\\). The results are reported in Tab. 5. We use Compcos as the baseline result. The simplest inference method, i.e., Independent, has surpassed Compcos on all the metrics. When using the knowledge-agnostic hypothesis, the seen accuracy and holistic metrics are better. But its unseen accuracy is lower. By considering the component relations, utilizing both inference paths substantially improves the unseen accuracy, yielding the best results for both HM and AUC. The results suggest that our proposed inference approach achieves a more balanced performance between seen and unseen actions. Fig. 5 reports the seen-unseen curves, implying that the proposed method is more robust." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.444, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "C2C: Component-to-Composition Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.258, + 0.145, + 0.744, + 0.159 + ], + "angle": 0, + "content": "Table 6: Effectiveness of the elements in the enhanced training strategy." + }, + { + "type": "table", + "bbox": [ + 0.384, + 0.171, + 0.615, + 0.246 + ], + "angle": 0, + "content": "
\\( \\mathcal{L}_{com} \\)\\( \\mathcal{L}_{comp} \\)\\( \\mathcal{L}_{ind} \\)CutMix\\( \\mathcal{L}_{con} \\)\\( \\mathcal{L}_{new} \\)HMAUC
✓29.812.5
✓✓30.813.2
✓✓✓31.413.6
✓✓✓✓31.513.9
✓✓✓✓✓32.114.0
✓✓✓✓✓✓33.014.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.23, + 0.265, + 0.771, + 0.279 + ], + "angle": 0, + "content": "Table 7: The accuracy (%) on actions containing verbs that cause deformations." + }, + { + "type": "table", + "bbox": [ + 0.356, + 0.292, + 0.644, + 0.34 + ], + "angle": 0, + "content": "
methodsverbobjectcomposition
seenunseenseenunseenseenunseen
Vanilla70.160.268.734.856.927.5
Vanilla + Lind71.865.268.245.158.036.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.359, + 0.785, + 0.388 + ], + "angle": 0, + "content": "Table 8: Unbiased accuracy (%) of seen and unseen actions and the corresponding harmonic mean (HM)." + }, + { + "type": "table", + "bbox": [ + 0.3, + 0.399, + 0.699, + 0.482 + ], + "angle": 0, + "content": "
methodsseenunseenHM
baseline36.817.223.5
+Lcon36.5(-0.3)16.4(-0.8)22.6(-0.9)
+cutmix35.7(-1.1)17.7(+0.5)23.6(+0.1)
+cutmix+Lcon36.1(+0.4)17.6(-0.1)23.7(+0.1)
+cutmix+Lnew37.0(+1.3)18.2(+0.5)24.4(+0.8)
+cutmix+Lcon+Lnew37.5(+0.5/+1.4)18.7(+0.5/+1.1)24.9(+0.5/+1.2)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.517, + 0.785, + 0.562 + ], + "angle": 0, + "content": "The enhanced training strategy. In Tab. 6, we demonstrate the effectiveness of our enhanced training strategy. It shows that the results exhibit consistent improvements as gradually applying the proposed strategy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.656 + ], + "angle": 0, + "content": "To further validate the proposed method, we verify the merits of the independent loss \\( L_{ind} \\) on actions containing verbs that can cause object deformations (such as squeeze and twist). The results are shown in Tab. 7. They show that the unseen accuracy improves significantly after adding \\( L_{ind} \\). However, the object accuracy of seen actions slightly decreases. This might be because spurious features are useful for recognizing seen components." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Next, we validate the proposed balanced learning of actions. With the vanilla version with \\( L_{ind} \\) as the baseline, we report the accuracy of all seen and unseen actions and HM without adding bias in Tab. 8. It shows that adding only \\( L_{con} \\) leads to performance degradation, but the degradation in seen accuracy is minor. Similarly, adding just CutMix results in a slight improvement, but with an obvious degradation in seen accuracy. When using CutMix, we add either \\( L_{con} \\) or \\( L_{new} \\). The results indicate that \\( L_{new} \\) yields a universal improvement, while \\( L_{con} \\) contributes positively to seen accuracy. Last, when \\( L_{new} \\) and \\( L_{con} \\) are integrated, they both bring improvements on all the metrics, and the gains are more pronounced than adding either one alone. This indicates that our training strategy enables the model to achieve a better balance between seen and unseen actions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "R. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.146, + 0.355, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.146, + 0.484, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.144, + 0.637, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.149, + 0.739, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.263, + 0.781, + 0.292 + ], + "angle": 0, + "content": "Fig. 6: The top-3 compatible components Fig. 7: Examples of successes and failures with the same verb or object components. of our methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.325, + 0.425, + 0.34 + ], + "angle": 0, + "content": "5.3 Qualitative analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.509 + ], + "angle": 0, + "content": "In Fig. 6, we show the top three compatible components for different samples, while keeping another component the same. When the verb is squeeze, for the first sample, the second and third compatible objects are cloth and paper, which are close to the visual contents and compatible with squeeze. For the second sample, the model identifies different objects. The second object basket is not included in the visual content but is compatible with squeeze. The third object bed cannot be combined with squeeze but is visually close to the sample. When the fixed component is an object, the situation is similar. It demonstrates that the proposed model can identify component compatibility according to both visual contents and component attributions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.79, + 0.648 + ], + "angle": 0, + "content": "In Fig. 7, we visualize the top three recognized actions of some successful and failed samples. In the top row, our model accurately predicts the given actions. Besides, the second and third predictions relate to the visual content, such as fold paper for the first sample and tear paper just a little bit for the second sample. The bottom row shows some failures. Though wrong, the predictions still reflect the video content, such as move part of plant and tear leaf into two pieces for the second example. However, the failures also demonstrate some limitations of the proposed method. For example, it struggles with the task of distinguishing objects or verbs of similar visual appearances." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.679, + 0.36, + 0.695 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This paper introduces a novel task, Zero-Shot Compositional Action Recognition (ZS-CAR), that investigates the compositional generalization in the video modality. A benchmarking dataset Sth-com and a baseline method C2C are proposed along with the novel task. Besides, an enhanced training strategy is presented to address two inherent challenges in ZS-CAR: component domain variations and component compatibility variations. The approach demonstrates superior performance compared to existing compositional generalization solutions in both image- and video-based zero-shot compositional learning settings." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.444, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "C2C: Component-to-Composition Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.788, + 0.273 + ], + "angle": 0, + "content": "This work is supported in part by the National Key Research and Development Program of China (2023YFF1105102, 2023YFF1105105), the National Natural Science Foundation of China (62020106012, 62332008, 62106089, U1836218, 62336004), the 111 Project of Ministry of Education of China (B12018), the Postgraduate Research & Practice Innovation Program of Jiangsu Province (KYCX22_2307), and the UK EPSRC (EP/V002856/1, EP/T022205/1)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.298, + 0.323, + 0.314 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.333, + 0.785, + 0.375 + ], + "angle": 0, + "content": "1. Arnab, A., Dehghani, M., Heigold, G., Sun, C., Lučić, M., Schmid, C.: Vivit: A video vision transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6836-6846 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.377, + 0.785, + 0.417 + ], + "angle": 0, + "content": "2. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: International Conference on Machine Learning. pp. 813-824 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.419, + 0.785, + 0.46 + ], + "angle": 0, + "content": "3. Bojanowski, P., Grave, E., Joulin, A., Mikolov, T.: Enriching word vectors with subword information. Transactions of the association for computational linguistics 5, 135-146 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.785, + 0.515 + ], + "angle": 0, + "content": "4. Brattoli, B., Tighe, J., Zhdanov, F., Perona, P., Chalupka, K.: Rethinking zero-shot video classification: End-to-end training for realistic applications. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4613-4623 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.517, + 0.785, + 0.559 + ], + "angle": 0, + "content": "5. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.56, + 0.785, + 0.601 + ], + "angle": 0, + "content": "6. Chao, W.L., Changpinyo, S., Gong, B., Sha, F.: An empirical study and analysis of generalized zero-shot learning for object recognition in the wild. In: Proceedings of the European Conference on Computer Vision. pp. 52-68 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.602, + 0.785, + 0.642 + ], + "angle": 0, + "content": "7. Chen, S., Huang, D.: Elaborative rehearsal for zero-shot action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13638-13647 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.644, + 0.733, + 0.658 + ], + "angle": 0, + "content": "8. Chomsky, N.: Aspects of the Theory of Syntax. No. 11, MIT press (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.785, + 0.699 + ], + "angle": 0, + "content": "9. Doshi, K., Yilmaz, Y.: Zero-shot action recognition with transformer-based video semantic embedding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4858-4867 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.701, + 0.785, + 0.756 + ], + "angle": 0, + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "11. Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6824-6835 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "12. Feichtenhofer, C.: X3d: Expanding architectures for efficient video recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 203-213 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.333, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "R. Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "13. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6202-6211 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.217 + ], + "angle": 0, + "content": "14. Fodor, J.A., Pylyshyn, Z.W.: Connectionism and cognitive architecture: A critical analysis. Cognition 28(1-2), 3-71 (1988)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.785, + 0.257 + ], + "angle": 0, + "content": "15. Fu, Y., Hospedales, T.M., Xiang, T., Gong, S.: Attribute learning for understanding unstructured social activity. In: Proceedings of the European Conference on Computer Vision. pp. 530-543 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.257, + 0.785, + 0.325 + ], + "angle": 0, + "content": "16. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The\" something something\" video database for learning and evaluating visual common sense. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5842-5850 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.325, + 0.785, + 0.367 + ], + "angle": 0, + "content": "17. Gretton, A., Bousquet, O., Smola, A., Scholkopf, B.: Measuring statistical dependence with hilbert-schmidt norms. In: International Conference on Algorithmic Learning Theory. pp. 63-77 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.367, + 0.785, + 0.407 + ], + "angle": 0, + "content": "18. Hao, S., Han, K., Wong, K.Y.K.: Learning attention as disentangler for compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15315-15324 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.407, + 0.785, + 0.447 + ], + "angle": 0, + "content": "19. Hara, K., Kataoka, H., Satoh, Y.: Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet? In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6546-6555 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.447, + 0.785, + 0.473 + ], + "angle": 0, + "content": "20. Heckman, J.J.: Sample selection bias as a specification error. Econometrica: Journal of the econometric society pp. 153-161 (1979)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "21. Isola, P., Lim, J.J., Adelson, E.H.: Discovering states and transformations in image collections. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1383-1391 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.514, + 0.785, + 0.556 + ], + "angle": 0, + "content": "22. Ji, J., Krishna, R., Fei-Fei, L., Niebles, J.C.: Action genome: Actions as compositions of spatio-temporal scene graphs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10236-10247 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.556, + 0.785, + 0.596 + ], + "angle": 0, + "content": "23. Ji, S., Xu, W., Yang, M., Yu, K.: 3d convolutional neural networks for human action recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 35(1), 221-231 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.596, + 0.785, + 0.637 + ], + "angle": 0, + "content": "24. Karthik, S., Mancini, M., Akata, Z.: Kg-sp: Knowledge guided simple primitives for open world compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9336-9345 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.637, + 0.785, + 0.678 + ], + "angle": 0, + "content": "25. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.678, + 0.785, + 0.718 + ], + "angle": 0, + "content": "26. Kim, H., Lee, J., Park, S., Sohn, K.: Hierarchical visual primitive experts for compositional zero-shot learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5675-5685 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "27. Kwon, H., Kim, M., Kwak, S., Cho, M.: Motionsqueeze: Neural motion feature learning for video understanding. In: Proceedings of the European Conference on Computer Vision. pp. 345-362 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "28. Kwon, H., Kim, M., Kwak, S., Cho, M.: Learning self-similarity in space and time as generalized motion for video action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13065-13075 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "29. Li, K., Wang, Y., Gao, P., Song, G., Liu, Y., Li, H., Qiao, Y.: Uniformer: Unified transformer for efficient spatiotemporal representation learning. arXiv preprint arXiv:2201.04676 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.444, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "C2C: Component-to-Composition Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "30. Li, R.C., Wu, X.J., Wu, C., Xu, T.Y., Kittler, J.: Dynamic information enhancement for video classification. Image and Vision Computing 114, 104244 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.785, + 0.202 + ], + "angle": 0, + "content": "31. Li, R., Wu, X.J., Xu, T.: Video is graph: Structured graph module for video action recognition. arXiv preprint arXiv:2110.05904 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.785, + 0.242 + ], + "angle": 0, + "content": "32. Li, R., Xu, T., Wu, X.J., Shen, Z., Kittler, J.: Perceiving actions via temporal video frame pairs. ACM Transactions on Intelligent Systems and Technology 15(3), 1-20 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.243, + 0.785, + 0.284 + ], + "angle": 0, + "content": "33. Li, X., Yang, X., Wei, K., Deng, C., Yang, M.: Siamese contrastive embedding network for compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9326-9335 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.285, + 0.785, + 0.324 + ], + "angle": 0, + "content": "34. Li, Y., Ji, B., Shi, X., Zhang, J., Kang, B., Wang, L.: Tea: Temporal excitation and aggregation for action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 909-918 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.325, + 0.785, + 0.365 + ], + "angle": 0, + "content": "35. Lin, J., Gan, C., Han, S.: Tsm: Temporal shift module for efficient video understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7083-7093 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.365, + 0.785, + 0.405 + ], + "angle": 0, + "content": "36. Liu, J., Kuipers, B., Savarese, S.: Recognizing human actions by attributes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3337-3344 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.406, + 0.785, + 0.446 + ], + "angle": 0, + "content": "37. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3202-3211 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.447, + 0.785, + 0.487 + ], + "angle": 0, + "content": "38. Lu, X., Guo, S., Liu, Z., Guo, J.: Decomposed soft prompt guided fusion enhancing for compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23560-23569 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.488, + 0.785, + 0.528 + ], + "angle": 0, + "content": "39. Ma, L., Zheng, Y., Zhang, Z., Yao, Y., Fan, X., Ye, Q.: Motion stimulation for compositional action recognition. IEEE Transactions on Circuits and Systems for Video Technology (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.529, + 0.785, + 0.568 + ], + "angle": 0, + "content": "40. Mancini, M., Naeem, M.F., Xian, Y., Akata, Z.: Open world compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5222-5230 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.569, + 0.785, + 0.622 + ], + "angle": 0, + "content": "41. Mandal, D., Narayan, S., Dwivedi, S.K., Gupta, V., Ahmed, S., Khan, F.S., Shao, L.: Out-of-distribution detection for generalized zero-shot action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9985-9993 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.623, + 0.785, + 0.677 + ], + "angle": 0, + "content": "42. Materzynska, J., Xiao, T., Herzig, R., Xu, H., Wang, X., Darrell, T.: Something-else: Compositional action recognition with spatial-temporal interaction networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1049-1059 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.678, + 0.785, + 0.718 + ], + "angle": 0, + "content": "43. Misra, I., Gupta, A., Hebert, M.: From red wine to red tomato: Composition with context. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1792-1801 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.719, + 0.785, + 0.759 + ], + "angle": 0, + "content": "44. Naeem, M.F., Xian, Y., Tombari, F., Akata, Z.: Learning graph embeddings for compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 953-962 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.76, + 0.785, + 0.799 + ], + "angle": 0, + "content": "45. Nayak, N.V., Yu, P., Bach, S.: Learning to compose soft prompts for compositional zero-shot learning. In: The Eleventh International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "46. Neimark, D., Bar, O., Zohar, M., Asselmann, D.: Video transformer network. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3163-3172 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "R. Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.188 + ], + "angle": 0, + "content": "47. Pan, J., Lin, Z., Zhu, X., Shao, J., Li, H.: St-adapter: Parameter-efficient image-to-video transfer learning. Advances in Neural Information Processing Systems 35, 26462-26477 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "48. Park, J., Lee, J., Sohn, K.: Dual-path adaptation from image to video transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2203-2213 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.785, + 0.271 + ], + "angle": 0, + "content": "49. Piergiovanni, A., Kuo, W., Angelova, A.: Rethinking video vits: Sparse video tubes for joint image and video learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2214-2224 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.311 + ], + "angle": 0, + "content": "50. Purushwalkam, S., Nickel, M., Gupta, A., Ranzato, M.: Task-driven modular networks for zero-shot compositional learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3593-3602 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.338 + ], + "angle": 0, + "content": "51. Qian, R., Lin, W., See, J., Li, D.: Controllable augmentations for video representation learning. Visual Intelligence 2(1), 1 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.785, + 0.379 + ], + "angle": 0, + "content": "52. Qian, Y., Yu, L., Liu, W., Hauptmann, A.G.: Rethinking zero-shot action recognition: Learning from latent atomic actions. In: Proceedings of the European Conference on Computer Vision. pp. 104-120 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.433 + ], + "angle": 0, + "content": "53. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.785, + 0.473 + ], + "angle": 0, + "content": "54. Ruis, F., Burghouts, G., Bucur, D.: Independent prototype propagation for zero-shot compositionality. Advances in Neural Information Processing Systems 34, 10641-10653 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "55. Saini, N., Pham, K., Shrivastava, A.: Disentangling visual embeddings for attributes and objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13658-13667 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.555 + ], + "angle": 0, + "content": "56. Sun, P., Wu, B., Li, X., Li, W., Duan, L., Gan, C.: Counterfactual debiasing inference for compositional action recognition. In: Proceedings of the ACM International Conference on Multimedia. pp. 3220-3228 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.785, + 0.596 + ], + "angle": 0, + "content": "57. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4489-4497 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.785, + 0.637 + ], + "angle": 0, + "content": "58. Wang, L., Tong, Z., Ji, B., Wu, G.: Tdn: Temporal difference networks for efficient action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1895-1904 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.678 + ], + "angle": 0, + "content": "59. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Van Gool, L.: Temporal segment networks: Towards good practices for deep action recognition. In: Proceedings of the European Conference on Computer Vision. pp. 20-36 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.732 + ], + "angle": 0, + "content": "60. Wang, Q., Liu, L., Jing, C., Chen, H., Liang, G., Wang, P., Shen, C.: Learning conditional attributes for compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11197-11206 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.785, + 0.772 + ], + "angle": 0, + "content": "61. Xie, S., Sun, C., Huang, J., Tu, Z., Murphy, K.: Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In: Proceedings of the European Conference on Computer Vision. pp. 305–321 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "62. Xu, T., Zhu, X.F., Wu, X.J.: Learning spatio-temporal discriminative model for affine subspace based visual object tracking. Visual Intelligence 1(1), 4 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "63. Xu, X., Hospedales, T., Gong, S.: Transductive zero-shot action recognition by word-vector embedding. International Journal of Computer Vision 123, 309-333 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.444, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "C2C: Component-to-Composition Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "64. Yan, R., Huang, P., Shu, X., Zhang, J., Pan, Y., Tang, J.: Look less think more: Rethinking compositional action recognition. In: Proceedings of the ACM International Conference on Multimedia. pp. 3666-3675 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "65. Yan, R., Xie, L., Shu, X., Zhang, L., Tang, J.: Progressive instance-aware feature learning for compositional action recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.259 + ], + "angle": 0, + "content": "66. Yang, T., Zhu, Y., Xie, Y., Zhang, A., Chen, C., Li, M.: Aim: Adapting image models for efficient video action recognition. arXiv preprint arXiv:2302.03024 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.3 + ], + "angle": 0, + "content": "67. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6023-6032 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.328 + ], + "angle": 0, + "content": "68. Zadrozny, B.: Learning and evaluating classifiers under sample selection bias. In: Proceedings of the International Conference on Machine Learning. p. 114 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.37 + ], + "angle": 0, + "content": "69. Zellers, R., Choi, Y.: Zero-shot activity recognition with verb attribute induction. In: Proceedings of the Conference on Empirical Methods in Natural Language Processing. pp. 946-958 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.398 + ], + "angle": 0, + "content": "70. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.398 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/C2C_ Component-to-Composition Learning for Zero-Shot Compositional Action Recognition/c79e54f7-761f-47c3-bbd9-d80fc5c9ad18_origin.pdf b/2024/C2C_ Component-to-Composition Learning for Zero-Shot Compositional Action Recognition/c79e54f7-761f-47c3-bbd9-d80fc5c9ad18_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..da5141898b7714ec435c2b45b7131ec70b07be2b --- /dev/null +++ b/2024/C2C_ Component-to-Composition Learning for Zero-Shot Compositional Action Recognition/c79e54f7-761f-47c3-bbd9-d80fc5c9ad18_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:33fdc083e440f192cd4940e35fe40db7df24383133fe761f10ed18a9c7297c50 +size 1941142 diff --git a/2024/C2C_ Component-to-Composition Learning for Zero-Shot Compositional Action Recognition/full.md b/2024/C2C_ Component-to-Composition Learning for Zero-Shot Compositional Action Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..83006205fdc7105717fe0fb549787238bd6a9285 --- /dev/null +++ b/2024/C2C_ Component-to-Composition Learning for Zero-Shot Compositional Action Recognition/full.md @@ -0,0 +1,339 @@ +# C2C: Component-to-Composition Learning for Zero-Shot Compositional Action Recognition + +Rongchang Li $^{1,2}$ , Zhenhua Feng $^{1,2,3,4}$ , Tianyang Xu $^{1}$ , Linze Li $^{1}$ , Xiao-Jun Wu $^{1*}$ , Muhammad Awais $^{2,4}$ , Sara Atito $^{2,4}$ , and Josef Kittler $^{2,3}$ + +$^{1}$ School of Artificial Intelligence and Computer Science, Jiangnan University, China +2 Centre for Vision, Speech and Signal Processing (CVSSP), University of Surrey, UK +$^{3}$ School of Computer Science and Electronic Engineering, University of Surrey, UK + +$^{4}$ Surrey Institute of People-centred AI (SI-PAI), University of Surrey, UK {li_rongchang,linze.li}@stu.jiangnan.edu.cn; + +feng-zhenhua@outlook.com; {wu_xiaojun, tianyang_xu}@jiangnan.edu.cn; {muhammad.awais, sara.atito, j.kittler}@surrey.ac.uk; + +Abstract. Compositional actions consist of dynamic (verbs) and static (objects) concepts. Humans can easily recognize unseen compositions using the learned concepts. For machines, solving such a problem requires a model to recognize unseen actions composed of previously observed verbs and objects, thus requiring so-called compositional generalization ability. To facilitate this research, we propose a novel Zero-Shot Compositional Action Recognition (ZS-CAR) task. For evaluating the task, we construct a new benchmark, Something-composition (Sth-com), based on the widely used Something-Something V2 dataset. We also propose a novel Component-to-Composition (C2C) learning method to solve the new ZS-CAR task. C2C includes an independent component learning module and a composition inference module. Last, we devise an enhanced training strategy to address the challenges of component variations between seen and unseen compositions and to handle the subtle balance between learning seen and unseen actions. The experimental results demonstrate that the proposed framework significantly surpasses the existing compositional generalization methods and sets a new state-of-the-art. The new Sth-com benchmark and code are available at https://github.com/RongchangLi/ZSCAR_C2C. + +Keywords: Zero-shot compositional action recognition $\cdot$ Action recognition $\cdot$ Compositional generalization + +# 1 Introduction + +Humans can understand basic dynamic and static concepts (i.e., verbs and objects) in seen compositional actions and generalise the concepts to unseen action categories. For example, by seeing close a box and open a bag, we can envision close a bag even if we have never seen this before. This compositional generalization capability is a fundamental function of human intelligence, by which we can + +![](images/aec7be6f90d7e549c916c8d9e90d6d3f7e59a6a1bbc2e7869d90efab3a5cba4b.jpg) +Fig. 1: Zero-Shot Compositional Action Recognition (ZS-CAR) requires models to recognize unseen actions composed of verbs and objects observed in seen actions. + +understand an 'infinite' number of complex concepts through assimilating the learned 'finite' concepts [8,14]. Compositional generalization has recently gained attention in the computer vision community, giving rise to the corresponding image analysis task Compositional Zero-Shot Learning (CZSL) [43]. In video analysis, [42] proposed the compositional action recognition task. However, this task only requires recognizing the dynamic component. It makes the task essentially a regular verb classification task characterized by explicit sample selection bias [20,68] regarding verb-object compositions, rather than verifying compositional generalization ability. Exploring the compositional generalization problem in the video modality is valuable as it presents a challenging video understanding task that requires comprehending both dynamic and static components while also generalising to novel compositions. Also, it provides a reliable way to develop open-scenario-related applications in autonomous vehicles, robots, etc. + +To facilitate the related research, we propose a novel task, namely Zero-Shot Compositional Action Recognition (ZS-CAR). We define a compositional action as the composition of a dynamic component verb (e.g., open, pick up, ...) and a static component object (e.g., door, wallet, ...). As shown in Fig. 1, ZS-CAR aims to enable machines to identify unseen action categories consisting of seen verbs and objects. The new proposed task poses a significant challenge. First, a model has to understand the intricate spatial and temporal information to identify the verb and object components, which is more challenging than the existing action recognition task. Second, the task naturally contains component domain variation and component compatibility variation. The first one is similar to the challenges in domain generalization, caused by the contextuality difference (e.g., transferring knowledge of book in close a book to recognize bend a book). The second one stems from the action category difference in the training and test domains. Overly accepting or rejecting verb-object relationships observed during training can detrimentally impact the generalization capability to recognize unseen actions. Given these challenges, the existing visual compositional generalization solutions cannot handle the ZS-CAR task well, as shown in Tab. 2. + +To solve ZS-CAR, we present an innovative Component-to-Composition (C2C) method to serve as the baseline. The pipeline of C2C is shown in Fig. 2. C2C first independently learns the dynamic and static components (i.e., verbs and objects). Then it employs the component consensus to recognize compositions (i.e., actions) from both dynamic and static perspectives, ensuring that both these two components are well learned. The procedure is motivated by the human process + +of assembling compositions. It involves an independent observation of a component, assessing its compatibility with another part, and gradually assembling the units into the entire structure. Moreover, C2C enables explicitly addressing the component domain variation and component compatibility variation problems. Accordingly, we propose an enhanced training strategy that involves minimizing the dependence on spurious information within visual features and balancing the emphasis between fitting seen and imagining unseen actions. + +Last, to enable the training of C2C and to evaluate its performance for the ZS-CAR task, we create a new dataset, Something-composition (Sth-com), based on the popular Something-Something V2 (Sth-V2) dataset [16]. The experimental results demonstrate that the vanilla or enhanced version of C2C achieves state-of-the-art results. To summarize, the main contributions of the paper include: + +- A novel zero-shot compositional action recognition (ZS-CAR) task that investigates compositional generalization in the video modality. +- A benchmarking dataset for evaluating ZS-CAR. We elaborate on the challenges in ZS-CAR and demonstrate that the existing visual compositional generalization methods cannot handle such problems well. +- A novel Component-to-Composition (C2C) method as the vanilla solution and an enhanced training strategy to address the identified challenges in ZS-CAR. The experimental results demonstrate that C2C solves these challenges effectively and attains SOTA performances. + +# 2 Related work + +Video action recognition aims to identify actions in videos. The main challenge is extracting discriminative action features from videos. Previous CNN-based methods address this challenge in two primary ways: using 2D CNNs with separate temporal operations [27,28,30-32,34,35,58,59,62], or jointly using 3D CNNs [5,12,13,19,23,51,57,61] to dig out spatiotemporal features. Recently, Transformer-based models transform a video into 3D tokens and model the spatiotemporal correlations through attention layers [1,2,11,29,37,46,49], or adapt the visual image encoder in large vision-language models to the video modality [47,48,66]. + +Zero-shot action recognition (ZSAR) aims to recognize action categories not appearing in the training set. A widely used idea is to measure video features and action embeddings in a common semantic space, hoping this space can generalise to unseen actions. The semantic embeddings are obtained by side semantic information, e.g., word embeddings [4, 41, 63, 69], multiple attributes [15, 36], elaborative rehearsals [7] and action descriptions [9, 52]. Different from ZSAR, the proposed ZS-CAR aims to explore the visual compositionality of actions and generalize the learned composition knowledge to unseen actions. + +Compositional action recognition: The compositional nature of actions has recently received increasing attention. Action genome [22] takes action dynamics as the relationship changes and decomposes actions into spatiotemporal scene graphs. Something-else [42] proposes compositional action recognition that + +only requires recognising verbs in unseen verb-object compositions. Subsequent studies [39,56,64,65] follow this setting and focus on extracting robust dynamic information. However, [42] essentially is a supervised task with explicit sample selection bias, rather than a compositional generalization problem. + +Image-based compositional generalization: In the image modality, [43] first introduces the Compositional Zero-Shot Learning (CZSL) task that requires models trained on seen attribute-object pairs to recognize unseen ones. A common idea is to find the nearest holistic composition embeddings for input images in a common space [40, 43, 44, 50, 54]. Some studies disentangle visual features to align with component embeddings [18, 24, 26, 33, 54, 55]. Recent works propose to adapt the vision-language model [53] to CZSL by text prompting methods [38, 45] and enhancing the interactions between language and visual features [38]. + +# 3 Problem Definition and Benchmark + +# 3.1 The Proposed ZS-CAR Task + +In traditional action recognition, a model predicts the action label defined as either a verb or a verb-object composition. For example, the Sth-V2 dataset [16] defines verb-only actions, such as open sth and pick up sth. Kinetics-400 [25] defines actions as both verb-object compositions and verb-only actions, such as play basketball, draw, etc. The recently proposed compositional action recognition task considers the interaction between verbs and objects. It requires recognising verbs of which the paired objects in the test set differ from those in the training set. Though introducing action compositionally by this constraint, this task only takes the verb-object relation as a sample selection bias controller, making it not a thorough compositional generalization problem. + +In contrast to previous tasks, we introduce the compositional generalization perspective to understand actions, by proposing a novel Zero-Shot Compositional Action Recognition (ZS-CAR) task. We explicitly define a compositional action as a concept of verb and object. To verify the compositional generalization ability, ZS-CAR requires models to recognize unseen action categories composed of verbs and objects observed in seen actions. Formally, we denote the training set as $D_{\mathrm{train}} = \{(\mathcal{X}_i,a_i)\}_{i = 1}^N$ consisting of video $\mathcal{X}$ labelled as action $a\in A_{\mathrm{train}}$ , with $A_{\mathrm{train}}\subset V\times O = \{(v,o)|v\in V,o\in O\}$ , where $V$ and $O$ are the verb and object sets. The validation and test sets are similarly denoted as $D_{\mathrm{val}}$ and $D_{\mathrm{test}}$ , with label sets as $A_{\mathrm{val}}$ and $A_{\mathrm{test}}$ . We follow the generalised zero-shot learning setting [6,50], i.e., both seen and unseen action categories appear in the validation/test set: $A_{\mathrm{train}}\cap A_{\mathrm{val / test}}\neq \emptyset$ and $\exists a(a\in A_{\mathrm{val / test}}\oplus a\in A_{\mathrm{train}})$ . + +# 3.2 The Something-composition Benchmark + +To assess the proposed task, we build a benchmark Something-composition (Sth-com), based on the widely used Something-Something V2 (Sth-V2) dataset [16]. + +Table 1: Statistics of the proposed Sth-com benchmark. + +
# Objects# Verbs# Verb-Object Actions# Sample
Train248161345138034
Val225157733 Seen + 717 Unseen18774
Test233161976 Seen + 956 Unseen22657
All248161512479465
+ +Sth-V2 has 220,847 videos of humans performing pre-defined compositional actions with everyday objects. It is widely used by the action recognition community, as its verbs cannot be identified from a single frame. + +Annotation: We use the training and test split of Sth-else [42], a subset of Sth-V2, as the starting point. Then, we iteratively clean the subsets and ensure that: (1) Each composition (i.e., action) has more than five samples. (2) The components (i.e., verb & object) present in the test set are seen in the training set. This yields 248 objects, 161 verbs, 5124 compositions and 79465 video samples. Next, we select $1/3$ of compositions in the training and test sets respectively and interchange $50\%$ of the selected compositions, thereby introducing seen action categories to the test set. Subsequently, following the ratio (3:4) in MIT-State [21], we divide the test set into the validation and test sets, respectively. The statistics of Sth-com are summarised in Tab. 1. + +# 4 Component-to-Composition Learning + +To address the proposed ZS-CAR task, we propose a novel Component-to-Composition (C2C) learning framework, as shown in Fig. 2, to serve as a vanilla solution. We simulate the process of human assembly to infer compositions. We commence with a specific component and scrutinise the compatibility of various categories of another component, leading to a reasonable composition proposal. These are correspondingly implemented in the proposed independent component learning module and composition to component learning module. Importantly, we analyse the challenges in ZS-CAR and devise an enhanced training strategy to facilitate generalising to unseen actions better. + +# 4.1 Independent Component Learning + +In this module, our goal is to identify the verb/object score of an input video and acquire their semantic prototypes. Specifically, we first encode the raw video sequence $\mathcal{X}$ with shape $T\times 224\times 224\times 3$ to a general representation $\mathbf{F}_x\in \mathbb{R}^{T\times D}$ , where $T$ and $D$ denote the frame and channel numbers, respectively. Then, we respectively adopt dynamic and static encoders to extract dynamic and static features, denoted as $\mathbf{f}_v$ and $\mathbf{f}_o\in \mathbb{R}^C$ , where $C$ is the channel number. The static encoder consists of a temporal average pooling layer and two MLP layers, while the dynamics encoder contains two temporal convolution layers and a temporal pooling layer. We encode the component labels as the general representations of verbs and objects, denoted as the component prototypes $\mathbf{E}_v = \{e_{v,i}\}_{i = 1}^{N_v}$ and + +![](images/e913e1be1af537543f79c24fab0979c538f8cbe92cd32b37dd215fc40174eea9.jpg) +Fig. 2: The proposed Component-to-Composition (C2C) learning framework. C2C first aligns verb/object prototypes with corresponding visual features to obtain component scores in the Independent Component Learning module. Then, actions are inferred through two paths (dynamics and static) in the Component to Composition module. In the dynamics path, verb prototypes and visual features are used to compute conditional object scores. Then the independent verb scores and conditional object scores are multiplied to gain the action scores. The static path follows a similar procedure. The final output is a consensus of the results from both paths. + +$\mathbf{E}_o = \{e_{o,i}\}_{i = 1}^{N_o}$ , where $N_{v}$ and $N_{o}$ are the verb and object numbers, respectively. The prototypes are learned by being aligned with the visual features $\mathbf{f}_v$ and $\mathbf{f}_o$ using cosine similarity as the gauge. Denoting the action label of the input video as $a_{i} = (v_{l},o_{k})$ , the optimization objective is defined as: + +$$ +\mathcal {L} _ {\text {v e r b}} = - \log \frac {\exp \left(\cos \left(\mathbf {f} _ {v} , \mathbf {e} _ {v , l}\right) / \tau\right)}{\sum_ {v _ {j} \in V} \exp \left(\cos \left(\mathbf {f} _ {v} , \mathbf {e} _ {v , j}\right) / \tau\right)}, \tag {1} +$$ + +$$ +\mathcal {L} _ {o b j} = - \log \frac {\exp \left(\cos \left(\mathbf {f} _ {o} , \mathbf {e} _ {o , k}\right) / \tau\right)}{\sum_ {o _ {j} \in O} \exp \left(\cos \left(\mathbf {f} _ {o} , \mathbf {e} _ {o , j}\right) / \tau\right)}, \tag {2} +$$ + +where $\mathbf{e}_{v,l}$ and $\mathbf{e}_{o,k}$ are the corresponding prototypes. $\cos(,)$ is the cosine similarity. $\tau$ is the temperature coefficient for better optimization. We denote $\mathcal{L}_{comp} = \mathcal{L}_{verb} + \mathcal{L}_{obj}$ as the component loss. By aligning with diverse component features in the training set, the prototypes inherently acquire semantic meanings. Accordingly, the verb and object scores, $\mathbf{s}_v$ and $\mathbf{s}_o$ , are the cosine similarities between the visual features and corresponding component prototypes. + +# 4.2 From component to composition + +In this module, we infer the action (composition) score $\mathbf{s}_a \in \mathbb{R}^{N_a}$ using the prototypes and the verbs and objects (components) scores. We define the conditional + +scores $\mathbf{S}_{o|v} \in \mathbb{R}^{N_v \times N_o}$ and $\mathbf{S}_{v|o} \in \mathbb{R}^{N_o \times N_v}$ , that represents the confidence of one component given the visual input and another component as the conditions. Accordingly, for one action $a_i = (v_l, o_k)$ , its confidence score can be calculated by either $s_{a,i} = s_{v,l} \cdot s_{o = o_k|v = v_l}$ or $s_{a,i} = s_{o,k} \cdot s_{v = v_l|o = o_k}$ . The first one relies on the verbs to infer actions, called the dynamics path. Similarly, the other one is called the static path. As shown in Fig. 2, in the dynamics path, we first encode $\mathbf{F}_x$ using a static encoder to provide the visual reference for the object component. Then we concatenate the output with the verb prototype $\mathbf{e}_{v,l}$ and fuse them to the compound prototype-visual feature $\mathbf{f}_{v - x,l}$ , using an MLP layer. Then we linearly map the cosine similarity $\cos(\mathbf{f}_{v - x,l}, \mathbf{e}_{o,k})$ to the range of [0,1], obtaining the conditional object score $s_{o = o_k|v = v_l}$ . The static path follows a similar process except for encoding $\mathbf{F}_x$ by the dynamics encoder. Last, we average the inference results from the two paths to obtain the final inference output. Taking $s_{a,i}$ as the score of the groundtruth action $a_i$ , the entire model is optimized using: + +$$ +\mathcal {L} _ {c o m} = - \log \frac {\exp \left(s _ {a , i} / \tau\right)}{\sum_ {a _ {j} \in A _ {\text {t r a i n}}} \exp \left(s _ {a , j} / \tau\right)}, \tag {3} +$$ + +# 4.3 Enhanced training strategy + +Risk in generalizing components. Naturally, there is a high probability of exhibiting different component appearances between training and test domains, namely component domain variation. As shown in Fig. 3, the component 'book' exhibits a notable difference in appearance between 'close a book' and 'bend a book'. Besides, given the combination of 'close' and + +![](images/7ba8a62b8eb8f5b84c1acf216c9484bdd1c383cb1c1d58fef493db1877242ca2.jpg) +Fig. 3: Component domain variations. For unseen actions, an object may exhibit different appearances (left). To deal with this (right), we reduce the spurious information and enlarge the independence between component-specific features. + +'book' in the training set, there is a tendency that the model might align features associated with 'close' to 'book', or vice versa. These dependencies may confuse the model when encountering unseen action categories. + +To cope with it, in the independent component learning module, we posit that the ideal dynamic-focused feature $\mathbf{f}_v$ and static-focused feature $\mathbf{f}_o$ should encompass two types of information: common and specific. The former denotes features pertinent to both components and the latter denotes the invariant features, that are crucial for identifying the respective components. Notably, the invariant features of verbs are spurious for identifying objects, and vice versa. As shown in Fig. 3, our solution is to minimize the spurious information in $\mathbf{f}_v$ and $\mathbf{f}_o$ , and explicitly reduce the mutual associations between their component-specific parts. To achieve this, we propose the following auxiliary losses: + +$$ +\mathcal {L} _ {\sup , v e r b} = h \left(\mathbf {f} _ {x}, \mathbf {f} _ {v}\right) - h \left(\mathbf {f} _ {v}, \mathbf {y} _ {v}\right), \tag {4} +$$ + +$$ +\mathcal {L} _ {s u p, o b j} = h \left(\mathbf {f} _ {x}, \mathbf {f} _ {o}\right) - h \left(\mathbf {f} _ {o}, \mathbf {y} _ {o}\right), \tag {5} +$$ + +$$ +\mathcal {L} _ {i n d} = \mathcal {L} _ {s u p, v e r b} + \mathcal {L} _ {s u p, o b j} + h \left(\mathbf {f} _ {v} ^ {\prime}, \mathbf {f} _ {o} ^ {\prime}\right), \tag {6} +$$ + +where $h(,.)$ is the Hilbert-Schmidt Independence Criterion (HSIC) [17] for independence measure. $h(\mathbf{f},\mathbf{h}) = 0$ if and only if $\mathbf{f} \perp \mathbf{h}$ . $\mathbf{f}_x$ is the temporal pooling result of $\mathbf{F}_x$ . $\mathbf{y}_v$ and $\mathbf{y}_o$ are the one-hot label vectors. $\mathcal{L}_{sup,verb}$ and $\mathcal{L}_{sup,obj}$ force the model to compress information, but retain sufficient information to identify the components, thus reducing the general spurious information. $\mathbf{f}_v'$ or $\mathbf{f}_o'$ is part of $\mathbf{f}_v$ or $\mathbf{f}_o$ , defined as $\mathbf{f}_v' = \mathbf{f}_v[: \rho C], \mathbf{f}_o' = \mathbf{f}_o[: \rho C]$ ( $0 \leq \rho \leq 1$ ). They are seen as component-specific features. So the term $h(\mathbf{f}_v',\mathbf{f}_o')$ ensures the independence of the invariant part of the verb and object features. + +Balanced learning of seen & unseen actions. In ZS-CAR, the verb-object connections in the training and test sets differ significantly, as shown in Fig. 4. While the seen connections help optimize the modules, an excessive emphasis on them may hinder generalizing to unseen actions. Our solution is simultaneously enhancing the learning of seen and unseen connections, enabling them to strike an optimal balance. + +We first assist the model in learning seen actions explicitly. We calculate the observed conditional + +probabilities in the training set as the observed conditional scores $\hat{\mathbf{S}}_{v|o}$ and $\hat{\mathbf{S}}_{o|v}$ . Then, we force the conditional scores in a training batch closer to $\hat{\mathbf{S}}_{v|o}$ and $\hat{\mathbf{S}}_{o|v}$ , with the objectives: + +$$ +\mathcal {L} _ {\text {c o n}, v} = - \frac {1}{N _ {v}} \sum_ {v _ {i} \in V} \hat {\mathbf {s}} _ {o | v = v _ {i}} \log \mathbf {s} _ {o | v = v _ {i}}, \tag {7} +$$ + +$$ +\mathcal {L} _ {\text {c o n}, o} = - \frac {1}{N _ {o}} \sum_ {o _ {i} \in O} \hat {\mathbf {s}} _ {v | o = o _ {i}} \log \mathbf {s} _ {v | o = o _ {i}}, \tag {8} +$$ + +where $N_v$ & $N_o$ denote the verb & object numbers, and the condition loss $\mathcal{L}_{con}$ is defined as the sum of $\mathcal{L}_{con,v}$ and $\mathcal{L}_{con,o}$ . $\hat{\mathbf{s}}_{o|v = v_i}$ or $\mathbf{s}_{o|v = v_i}$ denotes the conditional object score vector when verb is $v_i$ . $\hat{\mathbf{s}}_{v|o = o_i}$ and $\mathbf{s}_{v|o = o_i}$ are defined similarly. + +Then, inspired by the human ability to create new things by imagining compositions of known components, we use CutMix [67] to assemble samples in a training batch as the imagined unseen actions, thereby reducing the risk of overemphasizing seen compositions. Specifically, for an input $\mathcal{X}_i$ with label $a_i = (v_l,o_k)$ , we randomly select another sample $\mathcal{X}_j$ with label $a_j = (v_m,o_n)$ + +![](images/1c3d32af904095b36440c6f43eb8a1cd3b7ecbe0519a944234b9574ddda48a26.jpg) +Fig. 4: Component compatibility variations. The component relations are different between the training and test sets (left). To solve this (right), we use the observed conditional score to fit seen relations and encourage the model to imagine unseen actions to avoid being limited to seen relations. + +Table 2: A comparison with the SOTA methods on Sth-com, using the vision model with word embedding model paradigm. We report both vanilla (only $\mathcal{L}_{com}$ and $\mathcal{L}_{comp}$ ) and enhanced (using the advocated enhanced training strategy) versions of C2C. Since ADE is tailored for transformer architectures, we only adapt it to VideoSwin. + +
TSM-18VideoSwin-T
Methodsverbobj.seenunseenHMAUCverbobj.seenunseenHMAUC
TMN [50]41.435.931.628.122.07.333.445.631.030.223.07.9
Compcos [40]45.840.034.835.127.510.636.750.534.237.827.210.9
CGE [44]35.536.427.625.520.05.931.848.131.230.323.17.9
OADis [55]39.140.733.532.325.59.341.648.938.138.330.212.6
ADE [18]------35.349.133.836.226.810.3
CAN [60]40.039.232.431.624.58.732.649.533.033.325.09.2
C2C (Vanilla)47.742.738.139.130.813.247.950.942.143.634.916.4
C2C (Enhanced)50.144.339.441.033.014.548.852.843.444.136.717.4
+ +and crop it at a random position of random size. The cropped part replaces the same position of $\mathcal{X}_i$ , generating the augmented input $\bar{\mathcal{X}}_i$ . Denoting the cropping ratio as $\lambda$ , we modify the previous losses as: $\bar{\mathcal{L}}_{com} = (1 - \lambda)\mathcal{L}_{com}^{a_i} + \lambda \mathcal{L}_{com}^{a_j}$ and $\bar{\mathcal{L}}_{comp} = (1 - \lambda)\mathcal{L}_{comp}^{v_l,o_k} + \lambda \mathcal{L}_{comp}^{v_m,o_n}$ . $\mathcal{L}_{comp}^{a_i}$ calculates $\mathcal{L}_{com}$ using Eq. (3) with $a_i$ as groundtruth. $\mathcal{L}_{comp}^{v_l,o_k}$ calculates $\mathcal{L}_{comp}$ with $v_l$ and $o_k$ as groundtruth. $\mathcal{L}_{comp}^{a_j}$ and $\mathcal{L}_{comp}^{v_m,o_n}$ are defined similarly. Also, $\mathcal{L}_{ind}$ is modified to $\bar{\mathcal{L}}_{ind}$ by replacing $\mathbf{y}_i$ to $(1 - \lambda)\mathbf{y}_i + \lambda \mathbf{y}_j$ in Eq. (4) and Eq. (5). Particularly, we introduce new compositions $\bar{a}_1 = (v_l,o_n)$ and $\bar{a}_2 = (v_m,o_k)$ . $\bar{a}_1$ and $\bar{a}_2$ are not constrained in $A_{\mathrm{test}}$ or $A_{\mathrm{train}}$ . It gives a novel loss $\mathcal{L}_{new} = \mathcal{L}_{com}^{\bar{a}_1} + \mathcal{L}_{com}^{\bar{a}_2}$ , which forces the model to consider all the possible component connections. + +A concern is that the new compositions $a_1$ or $a_2$ may be counterfactual. However, it only introduces noise to the component-to-composition module, as the component labels still match the visual input. In this way, we still reserve these counterfactual labels as they can provide regularization effects that avoid overfitting to seen verb-object relations. Besides, it is also consistent with human imagination which is not only constrained within reasonable compositions. + +# 4.4 More details + +Network details. Following previous practice, we implement the video encoder and component label encoding with two paradigms: + +(1) Vision model with word embedding model. In this paradigm, we adopt an action recognition model (the CNN-based TSM-18 [35] or the transformer-based VideoSwin-T [37]) to extract the general video feature and a word embedding model (fasttext [3]) to encode the labels. +(2) Vision-language models. Recently, large pre-trained vision-language models, such as CLIP [53], have shown superior power. In this paradigm, we adopt the language branch of recent prompt-based compositional generalization methods (Coop [70], CSP [45] and SPM [38]) to encode the labels. As CLIP are trained using image-text pairs, we apply image-to-video adapters introduced in AIM [66] to adapt the CLIP visual encoder (ViT-B/32 [10]) to video modality. + +Table 3: A comparison with the SOTA CLIP-based methods on Sth-com. + +
methodsverbobj.seenunseenHMAUC
Coop [70]48.455.143.946.436.618.1
+C2C(vanilla)56.256.948.354.642.724.0
+C2C(enhance)56.358.950.756.644.526.0
CSP [45]49.054.843.647.436.018.0
+C2C(vanilla)56.356.649.053.442.523.8
+C2C(enhance)57.158.649.956.244.325.5
SPM [38]48.954.544.046.335.517.7
+DFM [38]47.455.443.147.335.817.9
+C2C(vanilla)55.757.349.353.442.223.8
+C2C(enhance)56.558.449.955.743.825.2
+ +Training Objectives. When applying the proposed enhanced training strategy, we perform CutMix to generate new compositions with a probability $p$ . If apply CutMix, the final loss is calculated as: + +$$ +\bar {\mathcal {L}} _ {t o t a l} = \bar {\mathcal {L}} _ {c o m} + \alpha \bar {\mathcal {L}} _ {c o m p} + \beta \bar {\mathcal {L}} _ {i n d} + \gamma \mathcal {L} _ {n e w}. \tag {9} +$$ + +If not, the final loss is: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {c o m}} + \alpha \mathcal {L} _ {\text {c o m p}} + \beta \mathcal {L} _ {\text {i n d}} + \gamma \mathcal {L} _ {\text {c o n}}. \tag {10} +$$ + +As $\mathcal{L}_{new}$ and $\mathcal{L}_{con}$ respectively guide the model focus on unseen and seen actions, we use the same coefficient $\gamma$ . + +# 5 Experiments + +Evaluation Metrics. The metrics include Verb, Object, Seen, Unseen, HM and AUC. Verb and Object respectively denote the accuracy of verbs and objects. To gain other metrics, we first add a constant to the seen actions scores. The constant ranges from negative infinity to positive infinity, making the model shift from completely biased to unseen actions to completely biased to seen actions and yielding the seen-unseen curve shown in Fig. 5. Seen and Unseen are the best accuracy of seen/unseen actions, i.e. the intersection points with x/y-axis in Fig. 5. HM is the best harmonic mean with formulation: $HM = 2(Seen * Unseen) / (Seen + Unseen)$ . AUC is the area under the curve. + +Implementation details. $\alpha$ , $\beta$ $\gamma$ in Eq. (9) and Eq. (10) are 0.2, 0.1 and 0.1, respectively. The probability $p$ of conducting CutMix is 0.7. More coefficient ablations and implementation details are reported in supplementary materials. + +# 5.1 Comparisons with State-of-the-Art + +Note that the existing compositional action recognition methods cannot apply to the ZS-CAR task. To comprehensively demonstrate the proposed C2C, we first adapt the existing CZSL methods to the proposed Sth-com dataset. We replace their image encoders with video encoders and adopt the same word + +Table 4: A comparison with the SOTA methods on C-GQA. + +
multiseenunseenHMAUC
CGE [44]×38.017.118.55.4
OADis [55]×38.319.820.17.0
CoT [26]✓39.222.722.17.4
C2C×39.123.023.07.7
C2C✓39.423.623.38.2
+ +Table 5: Results of different hypotheses inferring compositions from components. + +
verbobj.seenunseenHMAUC
Compcos [40]45.840.034.835.127.510.6
Independent47.942.237.437.429.812.2
Knowledge-agnostic48.443.138.736.830.212.6
Dynamic centered46.742.537.438.029.912.5
Static centered46.641.137.036.728.911.7
Full47.742.738.139.130.813.2
+ +embedding model, making sure only the compositional generalization parts are different between the adapted methods and our proposed methods. Besides, we also modify C2C and compare it with the existing SOTA methods on the image-based CZSL task using the challenging C-GQA [44] dataset. + +Comparisons on Sth-com. We first compare our method with the adapted CZSL methods of the vision model with the word embedding paradigm, using both CNN-based and transformer-based video encoders. The results are reported in Tab. 2. The vanilla version of C2C uses two basic losses. The enhanced version uses the proposed enhanced training strategy. + +Tab. 2 demonstrate that our method outperforms the existing methods by significant margins across all the metrics on ZS-CAR. Furthermore, the proposed enhanced training strategy provides consistent performance gains across different video encoders. Tab. 2 shows that the recently proposed methods based on Compcos are inferior to Compcos. Second, except for OADis, using the recently proposed VideoSwin-T backbone does not yield significant improvements. The first observation may stem from failing to generalise the embedding relations, not disentangling spatial and temporal information well, and not applying the conditional component relations in actions. Besides, jointly learning the compositions may lead to unbalanced learning of the components. These deficiencies degrade the effectiveness of video encoders, leading directly to the second observation. Tab. 2 also compares two video encoders. Though VideoSwin is an advanced model, its verb accuracy is consistently inferior or similar to TSM across all methods. Conversely, it outperforms TSM in object accuracy. This implies that our ZS-CAR task offers a new avenue for assessing biases regarding dynamic and static patterns in action recognition models. + +Then we apply our C2C method to recently proposed CLIP-based compositional zero-shot learning methods. The results are reported in Tab. 3. Coop, CSP and SPM are three text prompting methods. We can see that the proposed + +C2C method can improve their performance significantly on the ZS-CAR task. DFM [38] was proposed along with SPM. It enables cross-modal interactions between language and visual features for better composition prediction. However, the improvements are slight on the ZS-CAR task. Compared to it, our C2C method provides more significant improvements. + +Comparison on C-GQA. To further evaluate the proposed C2C method, we modify it for the image-based compositional zero-shot learning task by deleting all the temporal-related operations. We report the results on the recently proposed challenging benchmark C-GQA in Tab. 4, with fine-tuned ViT-B [10] as the backbone. In the table, the term 'multi' denotes the use of multi-level backbone outputs. When using multi-level outputs in C2C, we merely average them, much simpler than CoT. + +Though designed for video analysis, the proposed C2C exhibits superior performance over existing SOTA methods. Like CoT [26], we also try to utilize multi-level backbone outputs. But compared to the specially designed module in CoT, we merely average the multi-level outputs. This demonstrates that our C2C method can be a general solution to the compositional generalization problem. + +![](images/51f290a908be2a3536cefab74079e55c2d0da19a2a292d7acb7e84d508057d11.jpg) +Fig. 5: Seen-unseen curves, drawn by gradually forcing the model biased from unseen to seen actions. + +# 5.2 Ablation Study + +Different composition generation methods. We first compare different methods that can be used to generate compositions from verb and object components, including: (1) Independent supposes components are completely independent, so the inference is straightforward: $s_{c,i} = s_{v,i} \cdot s_{o,i}$ . (2) Knowledge-agnostic assumes components are not completely independent, but the component relations are agnostic: $s_{c,i} = s_{v,i} \cdot 1 + s_{o,i} \cdot 1$ . (3) One path inference considers the compatibility between compositions but only relies on either inference path, i.e., the dynamics or the static path result (Fig. 2). (4) Full is the final scheme, integrating both paths. + +For a fair comparison, the experiments only use the base losses $\mathcal{L}_{\text{com}}$ and $\mathcal{L}_{\text{comp}}$ . The results are reported in Tab. 5. We use Compcos as the baseline result. The simplest inference method, i.e., Independent, has surpassed Compcos on all the metrics. When using the knowledge-agnostic hypothesis, the seen accuracy and holistic metrics are better. But its unseen accuracy is lower. By considering the component relations, utilizing both inference paths substantially improves the unseen accuracy, yielding the best results for both HM and AUC. The results suggest that our proposed inference approach achieves a more balanced performance between seen and unseen actions. Fig. 5 reports the seen-unseen curves, implying that the proposed method is more robust. + +Table 6: Effectiveness of the elements in the enhanced training strategy. + +
\( \mathcal{L}_{com} \)\( \mathcal{L}_{comp} \)\( \mathcal{L}_{ind} \)CutMix\( \mathcal{L}_{con} \)\( \mathcal{L}_{new} \)HMAUC
✓29.812.5
✓✓30.813.2
✓✓✓31.413.6
✓✓✓✓31.513.9
✓✓✓✓✓32.114.0
✓✓✓✓✓✓33.014.5
+ +Table 7: The accuracy (%) on actions containing verbs that cause deformations. + +
methodsverbobjectcomposition
seenunseenseenunseenseenunseen
Vanilla70.160.268.734.856.927.5
Vanilla + Lind71.865.268.245.158.036.7
+ +Table 8: Unbiased accuracy (%) of seen and unseen actions and the corresponding harmonic mean (HM). + +
methodsseenunseenHM
baseline36.817.223.5
+Lcon36.5(-0.3)16.4(-0.8)22.6(-0.9)
+cutmix35.7(-1.1)17.7(+0.5)23.6(+0.1)
+cutmix+Lcon36.1(+0.4)17.6(-0.1)23.7(+0.1)
+cutmix+Lnew37.0(+1.3)18.2(+0.5)24.4(+0.8)
+cutmix+Lcon+Lnew37.5(+0.5/+1.4)18.7(+0.5/+1.1)24.9(+0.5/+1.2)
+ +The enhanced training strategy. In Tab. 6, we demonstrate the effectiveness of our enhanced training strategy. It shows that the results exhibit consistent improvements as gradually applying the proposed strategy. + +To further validate the proposed method, we verify the merits of the independent loss $L_{ind}$ on actions containing verbs that can cause object deformations (such as squeeze and twist). The results are shown in Tab. 7. They show that the unseen accuracy improves significantly after adding $L_{ind}$ . However, the object accuracy of seen actions slightly decreases. This might be because spurious features are useful for recognizing seen components. + +Next, we validate the proposed balanced learning of actions. With the vanilla version with $L_{ind}$ as the baseline, we report the accuracy of all seen and unseen actions and HM without adding bias in Tab. 8. It shows that adding only $L_{con}$ leads to performance degradation, but the degradation in seen accuracy is minor. Similarly, adding just CutMix results in a slight improvement, but with an obvious degradation in seen accuracy. When using CutMix, we add either $L_{con}$ or $L_{new}$ . The results indicate that $L_{new}$ yields a universal improvement, while $L_{con}$ contributes positively to seen accuracy. Last, when $L_{new}$ and $L_{con}$ are integrated, they both bring improvements on all the metrics, and the gains are more pronounced than adding either one alone. This indicates that our training strategy enables the model to achieve a better balance between seen and unseen actions. + +![](images/9525e3d9e569ed9ad921650b66893113037999dd8a211cc7da63a621d5b280e3.jpg) +Fig. 6: The top-3 compatible components Fig. 7: Examples of successes and failures with the same verb or object components. of our methods. + +![](images/efd711524ac51cac2ad16c6e86d8a9e08f1cef4a5f48594eeb71095c35553d2c.jpg) + +![](images/da02af2561f47b06a6167813210502738e9359cccf957c304621a965a3ef6227.jpg) + +![](images/acd696c7cf7c8d807d18bed514368e24a0a77aff00e7958fee93251f037926a8.jpg) + +# 5.3 Qualitative analysis + +In Fig. 6, we show the top three compatible components for different samples, while keeping another component the same. When the verb is squeeze, for the first sample, the second and third compatible objects are cloth and paper, which are close to the visual contents and compatible with squeeze. For the second sample, the model identifies different objects. The second object basket is not included in the visual content but is compatible with squeeze. The third object bed cannot be combined with squeeze but is visually close to the sample. When the fixed component is an object, the situation is similar. It demonstrates that the proposed model can identify component compatibility according to both visual contents and component attributions. + +In Fig. 7, we visualize the top three recognized actions of some successful and failed samples. In the top row, our model accurately predicts the given actions. Besides, the second and third predictions relate to the visual content, such as fold paper for the first sample and tear paper just a little bit for the second sample. The bottom row shows some failures. Though wrong, the predictions still reflect the video content, such as move part of plant and tear leaf into two pieces for the second example. However, the failures also demonstrate some limitations of the proposed method. For example, it struggles with the task of distinguishing objects or verbs of similar visual appearances. + +# 6 Conclusion + +This paper introduces a novel task, Zero-Shot Compositional Action Recognition (ZS-CAR), that investigates the compositional generalization in the video modality. A benchmarking dataset Sth-com and a baseline method C2C are proposed along with the novel task. Besides, an enhanced training strategy is presented to address two inherent challenges in ZS-CAR: component domain variations and component compatibility variations. The approach demonstrates superior performance compared to existing compositional generalization solutions in both image- and video-based zero-shot compositional learning settings. + +# Acknowledgements + +This work is supported in part by the National Key Research and Development Program of China (2023YFF1105102, 2023YFF1105105), the National Natural Science Foundation of China (62020106012, 62332008, 62106089, U1836218, 62336004), the 111 Project of Ministry of Education of China (B12018), the Postgraduate Research & Practice Innovation Program of Jiangsu Province (KYCX22_2307), and the UK EPSRC (EP/V002856/1, EP/T022205/1). + +# References + +1. Arnab, A., Dehghani, M., Heigold, G., Sun, C., Lučić, M., Schmid, C.: Vivit: A video vision transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6836-6846 (2021) +2. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: International Conference on Machine Learning. pp. 813-824 (2021) +3. Bojanowski, P., Grave, E., Joulin, A., Mikolov, T.: Enriching word vectors with subword information. Transactions of the association for computational linguistics 5, 135-146 (2017) +4. Brattoli, B., Tighe, J., Zhdanov, F., Perona, P., Chalupka, K.: Rethinking zero-shot video classification: End-to-end training for realistic applications. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4613-4623 (2020) +5. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017) +6. Chao, W.L., Changpinyo, S., Gong, B., Sha, F.: An empirical study and analysis of generalized zero-shot learning for object recognition in the wild. In: Proceedings of the European Conference on Computer Vision. pp. 52-68 (2016) +7. Chen, S., Huang, D.: Elaborative rehearsal for zero-shot action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13638-13647 (2021) +8. Chomsky, N.: Aspects of the Theory of Syntax. No. 11, MIT press (2014) +9. Doshi, K., Yilmaz, Y.: Zero-shot action recognition with transformer-based video semantic embedding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4858-4867 (2023) +10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020) +11. Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6824-6835 (2021) +12. Feichtenhofer, C.: X3d: Expanding architectures for efficient video recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 203-213 (2020) + +13. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6202-6211 (2019) +14. Fodor, J.A., Pylyshyn, Z.W.: Connectionism and cognitive architecture: A critical analysis. Cognition 28(1-2), 3-71 (1988) +15. Fu, Y., Hospedales, T.M., Xiang, T., Gong, S.: Attribute learning for understanding unstructured social activity. In: Proceedings of the European Conference on Computer Vision. pp. 530-543 (2012) +16. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The" something something" video database for learning and evaluating visual common sense. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5842-5850 (2017) +17. Gretton, A., Bousquet, O., Smola, A., Scholkopf, B.: Measuring statistical dependence with hilbert-schmidt norms. In: International Conference on Algorithmic Learning Theory. pp. 63-77 (2005) +18. Hao, S., Han, K., Wong, K.Y.K.: Learning attention as disentangler for compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15315-15324 (2023) +19. Hara, K., Kataoka, H., Satoh, Y.: Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet? In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6546-6555 (2018) +20. Heckman, J.J.: Sample selection bias as a specification error. Econometrica: Journal of the econometric society pp. 153-161 (1979) +21. Isola, P., Lim, J.J., Adelson, E.H.: Discovering states and transformations in image collections. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1383-1391 (2015) +22. Ji, J., Krishna, R., Fei-Fei, L., Niebles, J.C.: Action genome: Actions as compositions of spatio-temporal scene graphs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10236-10247 (2020) +23. Ji, S., Xu, W., Yang, M., Yu, K.: 3d convolutional neural networks for human action recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 35(1), 221-231 (2012) +24. Karthik, S., Mancini, M., Akata, Z.: Kg-sp: Knowledge guided simple primitives for open world compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9336-9345 (2022) +25. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017) +26. Kim, H., Lee, J., Park, S., Sohn, K.: Hierarchical visual primitive experts for compositional zero-shot learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5675-5685 (2023) +27. Kwon, H., Kim, M., Kwak, S., Cho, M.: Motionsqueeze: Neural motion feature learning for video understanding. In: Proceedings of the European Conference on Computer Vision. pp. 345-362 (2020) +28. Kwon, H., Kim, M., Kwak, S., Cho, M.: Learning self-similarity in space and time as generalized motion for video action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13065-13075 (2021) +29. Li, K., Wang, Y., Gao, P., Song, G., Liu, Y., Li, H., Qiao, Y.: Uniformer: Unified transformer for efficient spatiotemporal representation learning. arXiv preprint arXiv:2201.04676 (2022) + +30. Li, R.C., Wu, X.J., Wu, C., Xu, T.Y., Kittler, J.: Dynamic information enhancement for video classification. Image and Vision Computing 114, 104244 (2021) +31. Li, R., Wu, X.J., Xu, T.: Video is graph: Structured graph module for video action recognition. arXiv preprint arXiv:2110.05904 (2021) +32. Li, R., Xu, T., Wu, X.J., Shen, Z., Kittler, J.: Perceiving actions via temporal video frame pairs. ACM Transactions on Intelligent Systems and Technology 15(3), 1-20 (2024) +33. Li, X., Yang, X., Wei, K., Deng, C., Yang, M.: Siamese contrastive embedding network for compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9326-9335 (2022) +34. Li, Y., Ji, B., Shi, X., Zhang, J., Kang, B., Wang, L.: Tea: Temporal excitation and aggregation for action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 909-918 (2020) +35. Lin, J., Gan, C., Han, S.: Tsm: Temporal shift module for efficient video understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7083-7093 (2019) +36. Liu, J., Kuipers, B., Savarese, S.: Recognizing human actions by attributes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3337-3344 (2011) +37. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3202-3211 (2022) +38. Lu, X., Guo, S., Liu, Z., Guo, J.: Decomposed soft prompt guided fusion enhancing for compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23560-23569 (2023) +39. Ma, L., Zheng, Y., Zhang, Z., Yao, Y., Fan, X., Ye, Q.: Motion stimulation for compositional action recognition. IEEE Transactions on Circuits and Systems for Video Technology (2022) +40. Mancini, M., Naeem, M.F., Xian, Y., Akata, Z.: Open world compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5222-5230 (2021) +41. Mandal, D., Narayan, S., Dwivedi, S.K., Gupta, V., Ahmed, S., Khan, F.S., Shao, L.: Out-of-distribution detection for generalized zero-shot action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9985-9993 (2019) +42. Materzynska, J., Xiao, T., Herzig, R., Xu, H., Wang, X., Darrell, T.: Something-else: Compositional action recognition with spatial-temporal interaction networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1049-1059 (2020) +43. Misra, I., Gupta, A., Hebert, M.: From red wine to red tomato: Composition with context. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1792-1801 (2017) +44. Naeem, M.F., Xian, Y., Tombari, F., Akata, Z.: Learning graph embeddings for compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 953-962 (2021) +45. Nayak, N.V., Yu, P., Bach, S.: Learning to compose soft prompts for compositional zero-shot learning. In: The Eleventh International Conference on Learning Representations (2022) +46. Neimark, D., Bar, O., Zohar, M., Asselmann, D.: Video transformer network. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3163-3172 (2021) + +47. Pan, J., Lin, Z., Zhu, X., Shao, J., Li, H.: St-adapter: Parameter-efficient image-to-video transfer learning. Advances in Neural Information Processing Systems 35, 26462-26477 (2022) +48. Park, J., Lee, J., Sohn, K.: Dual-path adaptation from image to video transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2203-2213 (2023) +49. Piergiovanni, A., Kuo, W., Angelova, A.: Rethinking video vits: Sparse video tubes for joint image and video learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2214-2224 (2023) +50. Purushwalkam, S., Nickel, M., Gupta, A., Ranzato, M.: Task-driven modular networks for zero-shot compositional learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3593-3602 (2019) +51. Qian, R., Lin, W., See, J., Li, D.: Controllable augmentations for video representation learning. Visual Intelligence 2(1), 1 (2024) +52. Qian, Y., Yu, L., Liu, W., Hauptmann, A.G.: Rethinking zero-shot action recognition: Learning from latent atomic actions. In: Proceedings of the European Conference on Computer Vision. pp. 104-120 (2022) +53. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +54. Ruis, F., Burghouts, G., Bucur, D.: Independent prototype propagation for zero-shot compositionality. Advances in Neural Information Processing Systems 34, 10641-10653 (2021) +55. Saini, N., Pham, K., Shrivastava, A.: Disentangling visual embeddings for attributes and objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13658-13667 (2022) +56. Sun, P., Wu, B., Li, X., Li, W., Duan, L., Gan, C.: Counterfactual debiasing inference for compositional action recognition. In: Proceedings of the ACM International Conference on Multimedia. pp. 3220-3228 (2021) +57. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4489-4497 (2015) +58. Wang, L., Tong, Z., Ji, B., Wu, G.: Tdn: Temporal difference networks for efficient action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1895-1904 (2021) +59. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Van Gool, L.: Temporal segment networks: Towards good practices for deep action recognition. In: Proceedings of the European Conference on Computer Vision. pp. 20-36 (2016) +60. Wang, Q., Liu, L., Jing, C., Chen, H., Liang, G., Wang, P., Shen, C.: Learning conditional attributes for compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11197-11206 (2023) +61. Xie, S., Sun, C., Huang, J., Tu, Z., Murphy, K.: Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In: Proceedings of the European Conference on Computer Vision. pp. 305–321 (2018) +62. Xu, T., Zhu, X.F., Wu, X.J.: Learning spatio-temporal discriminative model for affine subspace based visual object tracking. Visual Intelligence 1(1), 4 (2023) +63. Xu, X., Hospedales, T., Gong, S.: Transductive zero-shot action recognition by word-vector embedding. International Journal of Computer Vision 123, 309-333 (2017) + +64. Yan, R., Huang, P., Shu, X., Zhang, J., Pan, Y., Tang, J.: Look less think more: Rethinking compositional action recognition. In: Proceedings of the ACM International Conference on Multimedia. pp. 3666-3675 (2022) +65. Yan, R., Xie, L., Shu, X., Zhang, L., Tang, J.: Progressive instance-aware feature learning for compositional action recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) +66. Yang, T., Zhu, Y., Xie, Y., Zhang, A., Chen, C., Li, M.: Aim: Adapting image models for efficient video action recognition. arXiv preprint arXiv:2302.03024 (2023) +67. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6023-6032 (2019) +68. Zadrozny, B.: Learning and evaluating classifiers under sample selection bias. In: Proceedings of the International Conference on Machine Learning. p. 114 (2004) +69. Zellers, R., Choi, Y.: Zero-shot activity recognition with verb attribute induction. In: Proceedings of the Conference on Empirical Methods in Natural Language Processing. pp. 946-958 (2017) +70. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022) \ No newline at end of file diff --git a/2024/C2C_ Component-to-Composition Learning for Zero-Shot Compositional Action Recognition/images.zip b/2024/C2C_ Component-to-Composition Learning for Zero-Shot Compositional Action Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c85d827cbe88ccb799d00922890f6b7330605b0e --- /dev/null +++ b/2024/C2C_ Component-to-Composition Learning for Zero-Shot Compositional Action Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ef5f3f282aa7a236fa7a587c91557f50930be68f91fee8c601a8966da3aeaa0 +size 482074 diff --git a/2024/C2C_ Component-to-Composition Learning for Zero-Shot Compositional Action Recognition/layout.json b/2024/C2C_ Component-to-Composition Learning for Zero-Shot Compositional Action Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5fd636d58c29fdb11fd2afdf3f93b52e2d72813a --- /dev/null +++ b/2024/C2C_ Component-to-Composition Learning for Zero-Shot Compositional Action Recognition/layout.json @@ -0,0 +1,11294 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 137, + 112, + 475, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 112, + 475, + 148 + ], + "spans": [ + { + "bbox": [ + 137, + 112, + 475, + 148 + ], + "type": "text", + "content": "C2C: Component-to-Composition Learning for Zero-Shot Compositional Action Recognition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "text", + "content": "Rongchang Li" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "text", + "content": ", Zhenhua Feng" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "inline_equation", + "content": "^{1,2,3,4}" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "text", + "content": ", Tianyang Xu" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "text", + "content": ", Linze Li" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "text", + "content": ", Xiao-Jun Wu" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "text", + "content": ", Muhammad Awais" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "text", + "content": ", Sara Atito" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "text", + "content": ", and Josef Kittler" + }, + { + "bbox": [ + 132, + 167, + 481, + 193 + ], + "type": "inline_equation", + "content": "^{2,3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 201, + 480, + 235 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 201, + 479, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 201, + 479, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 201, + 479, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 201, + 479, + 213 + ], + "type": "text", + "content": " School of Artificial Intelligence and Computer Science, Jiangnan University, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 213, + 480, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 480, + 224 + ], + "type": "text", + "content": "2 Centre for Vision, Speech and Signal Processing (CVSSP), University of Surrey, UK" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 224, + 479, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 224, + 479, + 235 + ], + "spans": [ + { + "bbox": [ + 134, + 224, + 479, + 235 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 134, + 224, + 479, + 235 + ], + "type": "text", + "content": " School of Computer Science and Electronic Engineering, University of Surrey, UK" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 149, + 235, + 465, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 235, + 465, + 258 + ], + "spans": [ + { + "bbox": [ + 149, + 235, + 465, + 258 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 149, + 235, + 465, + 258 + ], + "type": "text", + "content": " Surrey Institute of People-centred AI (SI-PAI), University of Surrey, UK {li_rongchang,linze.li}@stu.jiangnan.edu.cn;" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 149, + 258, + 463, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 258, + 463, + 279 + ], + "spans": [ + { + "bbox": [ + 149, + 258, + 463, + 279 + ], + "type": "text", + "content": "feng-zhenhua@outlook.com; {wu_xiaojun, tianyang_xu}@jiangnan.edu.cn; {muhammad.awais, sara.atito, j.kittler}@surrey.ac.uk;" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 304, + 453, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 304, + 453, + 514 + ], + "spans": [ + { + "bbox": [ + 159, + 304, + 453, + 514 + ], + "type": "text", + "content": "Abstract. Compositional actions consist of dynamic (verbs) and static (objects) concepts. Humans can easily recognize unseen compositions using the learned concepts. For machines, solving such a problem requires a model to recognize unseen actions composed of previously observed verbs and objects, thus requiring so-called compositional generalization ability. To facilitate this research, we propose a novel Zero-Shot Compositional Action Recognition (ZS-CAR) task. For evaluating the task, we construct a new benchmark, Something-composition (Sth-com), based on the widely used Something-Something V2 dataset. We also propose a novel Component-to-Composition (C2C) learning method to solve the new ZS-CAR task. C2C includes an independent component learning module and a composition inference module. Last, we devise an enhanced training strategy to address the challenges of component variations between seen and unseen compositions and to handle the subtle balance between learning seen and unseen actions. The experimental results demonstrate that the proposed framework significantly surpasses the existing compositional generalization methods and sets a new state-of-the-art. The new Sth-com benchmark and code are available at https://github.com/RongchangLi/ZSCAR_C2C." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 159, + 523, + 453, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 523, + 453, + 545 + ], + "spans": [ + { + "bbox": [ + 159, + 523, + 453, + 545 + ], + "type": "text", + "content": "Keywords: Zero-shot compositional action recognition " + }, + { + "bbox": [ + 159, + 523, + 453, + 545 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 523, + 453, + 545 + ], + "type": "text", + "content": " Action recognition " + }, + { + "bbox": [ + 159, + 523, + 453, + 545 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 523, + 453, + 545 + ], + "type": "text", + "content": " Compositional generalization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 563, + 230, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 563, + 230, + 576 + ], + "spans": [ + { + "bbox": [ + 132, + 563, + 230, + 576 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 587, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 587, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 587, + 482, + 648 + ], + "type": "text", + "content": "Humans can understand basic dynamic and static concepts (i.e., verbs and objects) in seen compositional actions and generalise the concepts to unseen action categories. For example, by seeing close a box and open a bag, we can envision close a bag even if we have never seen this before. This compositional generalization capability is a fundamental function of human intelligence, by which we can" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 238, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 238, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 238, + 665 + ], + "type": "text", + "content": "* Corresponding Author." + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 212, + 114, + 404, + 178 + ], + "blocks": [ + { + "bbox": [ + 212, + 114, + 404, + 178 + ], + "lines": [ + { + "bbox": [ + 212, + 114, + 404, + 178 + ], + "spans": [ + { + "bbox": [ + 212, + 114, + 404, + 178 + ], + "type": "image", + "image_path": "aec7be6f90d7e549c916c8d9e90d6d3f7e59a6a1bbc2e7869d90efab3a5cba4b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 186, + 482, + 209 + ], + "lines": [ + { + "bbox": [ + 131, + 186, + 482, + 209 + ], + "spans": [ + { + "bbox": [ + 131, + 186, + 482, + 209 + ], + "type": "text", + "content": "Fig. 1: Zero-Shot Compositional Action Recognition (ZS-CAR) requires models to recognize unseen actions composed of verbs and objects observed in seen actions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 233, + 482, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 233, + 482, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 482, + 388 + ], + "type": "text", + "content": "understand an 'infinite' number of complex concepts through assimilating the learned 'finite' concepts [8,14]. Compositional generalization has recently gained attention in the computer vision community, giving rise to the corresponding image analysis task Compositional Zero-Shot Learning (CZSL) [43]. In video analysis, [42] proposed the compositional action recognition task. However, this task only requires recognizing the dynamic component. It makes the task essentially a regular verb classification task characterized by explicit sample selection bias [20,68] regarding verb-object compositions, rather than verifying compositional generalization ability. Exploring the compositional generalization problem in the video modality is valuable as it presents a challenging video understanding task that requires comprehending both dynamic and static components while also generalising to novel compositions. Also, it provides a reliable way to develop open-scenario-related applications in autonomous vehicles, robots, etc." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 389, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 593 + ], + "type": "text", + "content": "To facilitate the related research, we propose a novel task, namely Zero-Shot Compositional Action Recognition (ZS-CAR). We define a compositional action as the composition of a dynamic component verb (e.g., open, pick up, ...) and a static component object (e.g., door, wallet, ...). As shown in Fig. 1, ZS-CAR aims to enable machines to identify unseen action categories consisting of seen verbs and objects. The new proposed task poses a significant challenge. First, a model has to understand the intricate spatial and temporal information to identify the verb and object components, which is more challenging than the existing action recognition task. Second, the task naturally contains component domain variation and component compatibility variation. The first one is similar to the challenges in domain generalization, caused by the contextuality difference (e.g., transferring knowledge of book in close a book to recognize bend a book). The second one stems from the action category difference in the training and test domains. Overly accepting or rejecting verb-object relationships observed during training can detrimentally impact the generalization capability to recognize unseen actions. Given these challenges, the existing visual compositional generalization solutions cannot handle the ZS-CAR task well, as shown in Tab. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 484, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 484, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 484, + 666 + ], + "type": "text", + "content": "To solve ZS-CAR, we present an innovative Component-to-Composition (C2C) method to serve as the baseline. The pipeline of C2C is shown in Fig. 2. C2C first independently learns the dynamic and static components (i.e., verbs and objects). Then it employs the component consensus to recognize compositions (i.e., actions) from both dynamic and static perspectives, ensuring that both these two components are well learned. The procedure is motivated by the human process" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "R. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": "of assembling compositions. It involves an independent observation of a component, assessing its compatibility with another part, and gradually assembling the units into the entire structure. Moreover, C2C enables explicitly addressing the component domain variation and component compatibility variation problems. Accordingly, we propose an enhanced training strategy that involves minimizing the dependence on spurious information within visual features and balancing the emphasis between fitting seen and imagining unseen actions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 480, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 480, + 260 + ], + "type": "text", + "content": "Last, to enable the training of C2C and to evaluate its performance for the ZS-CAR task, we create a new dataset, Something-composition (Sth-com), based on the popular Something-Something V2 (Sth-V2) dataset [16]. The experimental results demonstrate that the vanilla or enhanced version of C2C achieves state-of-the-art results. To summarize, the main contributions of the paper include:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 266, + 480, + 372 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 266, + 479, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 479, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 479, + 289 + ], + "type": "text", + "content": "- A novel zero-shot compositional action recognition (ZS-CAR) task that investigates compositional generalization in the video modality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 289, + 480, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 480, + 324 + ], + "type": "text", + "content": "- A benchmarking dataset for evaluating ZS-CAR. We elaborate on the challenges in ZS-CAR and demonstrate that the existing visual compositional generalization methods cannot handle such problems well." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 325, + 480, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 325, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 138, + 325, + 480, + 372 + ], + "type": "text", + "content": "- A novel Component-to-Composition (C2C) method as the vanilla solution and an enhanced training strategy to address the identified challenges in ZS-CAR. The experimental results demonstrate that C2C solves these challenges effectively and attains SOTA performances." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 390, + 234, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 390, + 234, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 390, + 234, + 403 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 414, + 480, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 480, + 521 + ], + "type": "text", + "content": "Video action recognition aims to identify actions in videos. The main challenge is extracting discriminative action features from videos. Previous CNN-based methods address this challenge in two primary ways: using 2D CNNs with separate temporal operations [27,28,30-32,34,35,58,59,62], or jointly using 3D CNNs [5,12,13,19,23,51,57,61] to dig out spatiotemporal features. Recently, Transformer-based models transform a video into 3D tokens and model the spatiotemporal correlations through attention layers [1,2,11,29,37,46,49], or adapt the visual image encoder in large vision-language models to the video modality [47,48,66]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 522, + 480, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 480, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 480, + 617 + ], + "type": "text", + "content": "Zero-shot action recognition (ZSAR) aims to recognize action categories not appearing in the training set. A widely used idea is to measure video features and action embeddings in a common semantic space, hoping this space can generalise to unseen actions. The semantic embeddings are obtained by side semantic information, e.g., word embeddings [4, 41, 63, 69], multiple attributes [15, 36], elaborative rehearsals [7] and action descriptions [9, 52]. Different from ZSAR, the proposed ZS-CAR aims to explore the visual compositionality of actions and generalize the learned composition knowledge to unseen actions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "content": "Compositional action recognition: The compositional nature of actions has recently received increasing attention. Action genome [22] takes action dynamics as the relationship changes and decomposes actions into spatiotemporal scene graphs. Something-else [42] proposes compositional action recognition that" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 271, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 271, + 91, + 448, + 102 + ], + "type": "text", + "content": "C2C: Component-to-Composition Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "only requires recognising verbs in unseen verb-object compositions. Subsequent studies [39,56,64,65] follow this setting and focus on extracting robust dynamic information. However, [42] essentially is a supervised task with explicit sample selection bias, rather than a compositional generalization problem." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 260 + ], + "type": "text", + "content": "Image-based compositional generalization: In the image modality, [43] first introduces the Compositional Zero-Shot Learning (CZSL) task that requires models trained on seen attribute-object pairs to recognize unseen ones. A common idea is to find the nearest holistic composition embeddings for input images in a common space [40, 43, 44, 50, 54]. Some studies disentangle visual features to align with component embeddings [18, 24, 26, 33, 54, 55]. Recent works propose to adapt the vision-language model [53] to CZSL by text prompting methods [38, 45] and enhancing the interactions between language and visual features [38]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 282, + 364, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 364, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 364, + 295 + ], + "type": "text", + "content": "3 Problem Definition and Benchmark" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 310, + 304, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 304, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 304, + 322 + ], + "type": "text", + "content": "3.1 The Proposed ZS-CAR Task" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 333, + 480, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 480, + 453 + ], + "type": "text", + "content": "In traditional action recognition, a model predicts the action label defined as either a verb or a verb-object composition. For example, the Sth-V2 dataset [16] defines verb-only actions, such as open sth and pick up sth. Kinetics-400 [25] defines actions as both verb-object compositions and verb-only actions, such as play basketball, draw, etc. The recently proposed compositional action recognition task considers the interaction between verbs and objects. It requires recognising verbs of which the paired objects in the test set differ from those in the training set. Though introducing action compositionally by this constraint, this task only takes the verb-object relation as a sample selection bias controller, making it not a thorough compositional generalization problem." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "text", + "content": "In contrast to previous tasks, we introduce the compositional generalization perspective to understand actions, by proposing a novel Zero-Shot Compositional Action Recognition (ZS-CAR) task. We explicitly define a compositional action as a concept of verb and object. To verify the compositional generalization ability, ZS-CAR requires models to recognize unseen action categories composed of verbs and objects observed in seen actions. Formally, we denote the training set as " + }, + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{train}} = \\{(\\mathcal{X}_i,a_i)\\}_{i = 1}^N" + }, + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "text", + "content": " consisting of video " + }, + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "text", + "content": " labelled as action " + }, + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "inline_equation", + "content": "a\\in A_{\\mathrm{train}}" + }, + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{train}}\\subset V\\times O = \\{(v,o)|v\\in V,o\\in O\\}" + }, + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "text", + "content": " are the verb and object sets. The validation and test sets are similarly denoted as " + }, + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{val}}" + }, + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{test}}" + }, + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "text", + "content": ", with label sets as " + }, + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{val}}" + }, + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{test}}" + }, + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "text", + "content": ". We follow the generalised zero-shot learning setting [6,50], i.e., both seen and unseen action categories appear in the validation/test set: " + }, + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{train}}\\cap A_{\\mathrm{val / test}}\\neq \\emptyset" + }, + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "inline_equation", + "content": "\\exists a(a\\in A_{\\mathrm{val / test}}\\oplus a\\in A_{\\mathrm{train}})" + }, + { + "bbox": [ + 130, + 454, + 480, + 597 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 618, + 362, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 618, + 362, + 630 + ], + "spans": [ + { + "bbox": [ + 132, + 618, + 362, + 630 + ], + "type": "text", + "content": "3.2 The Something-composition Benchmark" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "content": "To assess the proposed task, we build a benchmark Something-composition (Sth-com), based on the widely used Something-Something V2 (Sth-V2) dataset [16]." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "R. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 209, + 136, + 402, + 182 + ], + "blocks": [ + { + "bbox": [ + 190, + 114, + 423, + 126 + ], + "lines": [ + { + "bbox": [ + 190, + 114, + 423, + 126 + ], + "spans": [ + { + "bbox": [ + 190, + 114, + 423, + 126 + ], + "type": "text", + "content": "Table 1: Statistics of the proposed Sth-com benchmark." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 209, + 136, + 402, + 182 + ], + "lines": [ + { + "bbox": [ + 209, + 136, + 402, + 182 + ], + "spans": [ + { + "bbox": [ + 209, + 136, + 402, + 182 + ], + "type": "table", + "html": "
# Objects# Verbs# Verb-Object Actions# Sample
Train248161345138034
Val225157733 Seen + 717 Unseen18774
Test233161976 Seen + 956 Unseen22657
All248161512479465
", + "image_path": "2cf08a3b0f3c9b0ce92cc91f6b2570789b26175e03b5605e53fbebd4c98ace6e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 204, + 479, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 204, + 479, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 479, + 239 + ], + "type": "text", + "content": "Sth-V2 has 220,847 videos of humans performing pre-defined compositional actions with everyday objects. It is widely used by the action recognition community, as its verbs cannot be identified from a single frame." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 239, + 480, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 480, + 360 + ], + "type": "text", + "content": "Annotation: We use the training and test split of Sth-else [42], a subset of Sth-V2, as the starting point. Then, we iteratively clean the subsets and ensure that: (1) Each composition (i.e., action) has more than five samples. (2) The components (i.e., verb & object) present in the test set are seen in the training set. This yields 248 objects, 161 verbs, 5124 compositions and 79465 video samples. Next, we select " + }, + { + "bbox": [ + 130, + 239, + 480, + 360 + ], + "type": "inline_equation", + "content": "1/3" + }, + { + "bbox": [ + 130, + 239, + 480, + 360 + ], + "type": "text", + "content": " of compositions in the training and test sets respectively and interchange " + }, + { + "bbox": [ + 130, + 239, + 480, + 360 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 239, + 480, + 360 + ], + "type": "text", + "content": " of the selected compositions, thereby introducing seen action categories to the test set. Subsequently, following the ratio (3:4) in MIT-State [21], we divide the test set into the validation and test sets, respectively. The statistics of Sth-com are summarised in Tab. 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 377, + 375, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 377, + 375, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 377, + 375, + 392 + ], + "type": "text", + "content": "4 Component-to-Composition Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 402, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 510 + ], + "type": "text", + "content": "To address the proposed ZS-CAR task, we propose a novel Component-to-Composition (C2C) learning framework, as shown in Fig. 2, to serve as a vanilla solution. We simulate the process of human assembly to infer compositions. We commence with a specific component and scrutinise the compatibility of various categories of another component, leading to a reasonable composition proposal. These are correspondingly implemented in the proposed independent component learning module and composition to component learning module. Importantly, we analyse the challenges in ZS-CAR and devise an enhanced training strategy to facilitate generalising to unseen actions better." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 526, + 334, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 526, + 334, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 334, + 540 + ], + "type": "text", + "content": "4.1 Independent Component Learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": "In this module, our goal is to identify the verb/object score of an input video and acquire their semantic prototypes. Specifically, we first encode the raw video sequence " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " with shape " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "T\\times 224\\times 224\\times 3" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " to a general representation " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_x\\in \\mathbb{R}^{T\\times D}" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " denote the frame and channel numbers, respectively. Then, we respectively adopt dynamic and static encoders to extract dynamic and static features, denoted as " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_v" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_o\\in \\mathbb{R}^C" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " is the channel number. The static encoder consists of a temporal average pooling layer and two MLP layers, while the dynamics encoder contains two temporal convolution layers and a temporal pooling layer. We encode the component labels as the general representations of verbs and objects, denoted as the component prototypes " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_v = \\{e_{v,i}\\}_{i = 1}^{N_v}" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 271, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 271, + 91, + 448, + 103 + ], + "type": "text", + "content": "C2C: Component-to-Composition Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 116, + 471, + 299 + ], + "blocks": [ + { + "bbox": [ + 143, + 116, + 471, + 299 + ], + "lines": [ + { + "bbox": [ + 143, + 116, + 471, + 299 + ], + "spans": [ + { + "bbox": [ + 143, + 116, + 471, + 299 + ], + "type": "image", + "image_path": "e913e1be1af537543f79c24fab0979c538f8cbe92cd32b37dd215fc40174eea9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 308, + 482, + 397 + ], + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 397 + ], + "type": "text", + "content": "Fig. 2: The proposed Component-to-Composition (C2C) learning framework. C2C first aligns verb/object prototypes with corresponding visual features to obtain component scores in the Independent Component Learning module. Then, actions are inferred through two paths (dynamics and static) in the Component to Composition module. In the dynamics path, verb prototypes and visual features are used to compute conditional object scores. Then the independent verb scores and conditional object scores are multiplied to gain the action scores. The static path follows a similar procedure. The final output is a consensus of the results from both paths." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 417, + 482, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 482, + 466 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 482, + 466 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_o = \\{e_{o,i}\\}_{i = 1}^{N_o}" + }, + { + "bbox": [ + 130, + 417, + 482, + 466 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 417, + 482, + 466 + ], + "type": "inline_equation", + "content": "N_{v}" + }, + { + "bbox": [ + 130, + 417, + 482, + 466 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 417, + 482, + 466 + ], + "type": "inline_equation", + "content": "N_{o}" + }, + { + "bbox": [ + 130, + 417, + 482, + 466 + ], + "type": "text", + "content": " are the verb and object numbers, respectively. The prototypes are learned by being aligned with the visual features " + }, + { + "bbox": [ + 130, + 417, + 482, + 466 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_v" + }, + { + "bbox": [ + 130, + 417, + 482, + 466 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 417, + 482, + 466 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_o" + }, + { + "bbox": [ + 130, + 417, + 482, + 466 + ], + "type": "text", + "content": " using cosine similarity as the gauge. Denoting the action label of the input video as " + }, + { + "bbox": [ + 130, + 417, + 482, + 466 + ], + "type": "inline_equation", + "content": "a_{i} = (v_{l},o_{k})" + }, + { + "bbox": [ + 130, + 417, + 482, + 466 + ], + "type": "text", + "content": ", the optimization objective is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 213, + 472, + 481, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 472, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 213, + 472, + 481, + 500 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {v e r b}} = - \\log \\frac {\\exp \\left(\\cos \\left(\\mathbf {f} _ {v} , \\mathbf {e} _ {v , l}\\right) / \\tau\\right)}{\\sum_ {v _ {j} \\in V} \\exp \\left(\\cos \\left(\\mathbf {f} _ {v} , \\mathbf {e} _ {v , j}\\right) / \\tau\\right)}, \\tag {1}", + "image_path": "47dcc9f92ade7e30850e71aa78c3510387a4b6d65af0179830f8c938906df2fa.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 216, + 505, + 481, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 505, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 216, + 505, + 481, + 534 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {o b j} = - \\log \\frac {\\exp \\left(\\cos \\left(\\mathbf {f} _ {o} , \\mathbf {e} _ {o , k}\\right) / \\tau\\right)}{\\sum_ {o _ {j} \\in O} \\exp \\left(\\cos \\left(\\mathbf {f} _ {o} , \\mathbf {e} _ {o , j}\\right) / \\tau\\right)}, \\tag {2}", + "image_path": "08e6babf7c72ce62691c2951a5fad8f501c3a8923a82df2fc015377d2076c5ac.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 537, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 537, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 537, + 482, + 609 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 537, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{v,l}" + }, + { + "bbox": [ + 130, + 537, + 482, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 537, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{o,k}" + }, + { + "bbox": [ + 130, + 537, + 482, + 609 + ], + "type": "text", + "content": " are the corresponding prototypes. " + }, + { + "bbox": [ + 130, + 537, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\cos(,)" + }, + { + "bbox": [ + 130, + 537, + 482, + 609 + ], + "type": "text", + "content": " is the cosine similarity. " + }, + { + "bbox": [ + 130, + 537, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 537, + 482, + 609 + ], + "type": "text", + "content": " is the temperature coefficient for better optimization. We denote " + }, + { + "bbox": [ + 130, + 537, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{comp} = \\mathcal{L}_{verb} + \\mathcal{L}_{obj}" + }, + { + "bbox": [ + 130, + 537, + 482, + 609 + ], + "type": "text", + "content": " as the component loss. By aligning with diverse component features in the training set, the prototypes inherently acquire semantic meanings. Accordingly, the verb and object scores, " + }, + { + "bbox": [ + 130, + 537, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_v" + }, + { + "bbox": [ + 130, + 537, + 482, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 537, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_o" + }, + { + "bbox": [ + 130, + 537, + 482, + 609 + ], + "type": "text", + "content": ", are the cosine similarities between the visual features and corresponding component prototypes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 624, + 326, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 326, + 636 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 326, + 636 + ], + "type": "text", + "content": "4.2 From component to composition" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "In this module, we infer the action (composition) score " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_a \\in \\mathbb{R}^{N_a}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " using the prototypes and the verbs and objects (components) scores. We define the conditional" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "R. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "text", + "content": "scores " + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{o|v} \\in \\mathbb{R}^{N_v \\times N_o}" + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{v|o} \\in \\mathbb{R}^{N_o \\times N_v}" + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "text", + "content": ", that represents the confidence of one component given the visual input and another component as the conditions. Accordingly, for one action " + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "inline_equation", + "content": "a_i = (v_l, o_k)" + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "text", + "content": ", its confidence score can be calculated by either " + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "inline_equation", + "content": "s_{a,i} = s_{v,l} \\cdot s_{o = o_k|v = v_l}" + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "inline_equation", + "content": "s_{a,i} = s_{o,k} \\cdot s_{v = v_l|o = o_k}" + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "text", + "content": ". The first one relies on the verbs to infer actions, called the dynamics path. Similarly, the other one is called the static path. As shown in Fig. 2, in the dynamics path, we first encode " + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_x" + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "text", + "content": " using a static encoder to provide the visual reference for the object component. Then we concatenate the output with the verb prototype " + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{v,l}" + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "text", + "content": " and fuse them to the compound prototype-visual feature " + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{v - x,l}" + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "text", + "content": ", using an MLP layer. Then we linearly map the cosine similarity " + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\cos(\\mathbf{f}_{v - x,l}, \\mathbf{e}_{o,k})" + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "text", + "content": " to the range of [0,1], obtaining the conditional object score " + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "inline_equation", + "content": "s_{o = o_k|v = v_l}" + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "text", + "content": ". The static path follows a similar process except for encoding " + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_x" + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "text", + "content": " by the dynamics encoder. Last, we average the inference results from the two paths to obtain the final inference output. Taking " + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "inline_equation", + "content": "s_{a,i}" + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "text", + "content": " as the score of the groundtruth action " + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "inline_equation", + "content": "a_i" + }, + { + "bbox": [ + 130, + 115, + 482, + 284 + ], + "type": "text", + "content": ", the entire model is optimized using:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 223, + 289, + 481, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 289, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 223, + 289, + 481, + 317 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c o m} = - \\log \\frac {\\exp \\left(s _ {a , i} / \\tau\\right)}{\\sum_ {a _ {j} \\in A _ {\\text {t r a i n}}} \\exp \\left(s _ {a , j} / \\tau\\right)}, \\tag {3}", + "image_path": "5795ccf54dcee54e377280162add214c686d2df776b7e2d2519328519027a93c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 331, + 297, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 297, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 297, + 342 + ], + "type": "text", + "content": "4.3 Enhanced training strategy" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 348, + 264, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 264, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 264, + 502 + ], + "type": "text", + "content": "Risk in generalizing components. Naturally, there is a high probability of exhibiting different component appearances between training and test domains, namely component domain variation. As shown in Fig. 3, the component 'book' exhibits a notable difference in appearance between 'close a book' and 'bend a book'. Besides, given the combination of 'close' and" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 274, + 371, + 470, + 434 + ], + "blocks": [ + { + "bbox": [ + 274, + 371, + 470, + 434 + ], + "lines": [ + { + "bbox": [ + 274, + 371, + 470, + 434 + ], + "spans": [ + { + "bbox": [ + 274, + 371, + 470, + 434 + ], + "type": "image", + "image_path": "7ba8a62b8eb8f5b84c1acf216c9484bdd1c383cb1c1d58fef493db1877242ca2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 442, + 482, + 498 + ], + "lines": [ + { + "bbox": [ + 269, + 442, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 269, + 442, + 482, + 498 + ], + "type": "text", + "content": "Fig. 3: Component domain variations. For unseen actions, an object may exhibit different appearances (left). To deal with this (right), we reduce the spurious information and enlarge the independence between component-specific features." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 502, + 482, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 539 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 539 + ], + "type": "text", + "content": "'book' in the training set, there is a tendency that the model might align features associated with 'close' to 'book', or vice versa. These dependencies may confuse the model when encountering unseen action categories." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "type": "text", + "content": "To cope with it, in the independent component learning module, we posit that the ideal dynamic-focused feature " + }, + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_v" + }, + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "type": "text", + "content": " and static-focused feature " + }, + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_o" + }, + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "type": "text", + "content": " should encompass two types of information: common and specific. The former denotes features pertinent to both components and the latter denotes the invariant features, that are crucial for identifying the respective components. Notably, the invariant features of verbs are spurious for identifying objects, and vice versa. As shown in Fig. 3, our solution is to minimize the spurious information in " + }, + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_v" + }, + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_o" + }, + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "type": "text", + "content": ", and explicitly reduce the mutual associations between their component-specific parts. To achieve this, we propose the following auxiliary losses:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 235, + 653, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 653, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 235, + 653, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\sup , v e r b} = h \\left(\\mathbf {f} _ {x}, \\mathbf {f} _ {v}\\right) - h \\left(\\mathbf {f} _ {v}, \\mathbf {y} _ {v}\\right), \\tag {4}", + "image_path": "904643ae0ad12a0b74c4f974fb5b9047726ac1521204dc8a44398980cb5d0827.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 271, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 271, + 91, + 448, + 102 + ], + "type": "text", + "content": "C2C: Component-to-Composition Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 239, + 128, + 481, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 128, + 481, + 142 + ], + "spans": [ + { + "bbox": [ + 239, + 128, + 481, + 142 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s u p, o b j} = h \\left(\\mathbf {f} _ {x}, \\mathbf {f} _ {o}\\right) - h \\left(\\mathbf {f} _ {o}, \\mathbf {y} _ {o}\\right), \\tag {5}", + "image_path": "8eaf8faefd014e49ca1dde377708584722b8c76a02ee2d745843007b83ff5275.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 222, + 157, + 481, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 157, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 222, + 157, + 481, + 171 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i n d} = \\mathcal {L} _ {s u p, v e r b} + \\mathcal {L} _ {s u p, o b j} + h \\left(\\mathbf {f} _ {v} ^ {\\prime}, \\mathbf {f} _ {o} ^ {\\prime}\\right), \\tag {6}", + "image_path": "f4ba3314e6b3345b1591f25e19f1dddec095dc282463678ee0fbc66c0ee0adeb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "inline_equation", + "content": "h(,.)" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": " is the Hilbert-Schmidt Independence Criterion (HSIC) [17] for independence measure. " + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "inline_equation", + "content": "h(\\mathbf{f},\\mathbf{h}) = 0" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": " if and only if " + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{f} \\perp \\mathbf{h}" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_x" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": " is the temporal pooling result of " + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_x" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_v" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_o" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": " are the one-hot label vectors. " + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sup,verb}" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sup,obj}" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": " force the model to compress information, but retain sufficient information to identify the components, thus reducing the general spurious information. " + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_v'" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_o'" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": " is part of " + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_v" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_o" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": ", defined as " + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_v' = \\mathbf{f}_v[: \\rho C], \\mathbf{f}_o' = \\mathbf{f}_o[: \\rho C]" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "inline_equation", + "content": "0 \\leq \\rho \\leq 1" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": "). They are seen as component-specific features. So the term " + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "inline_equation", + "content": "h(\\mathbf{f}_v',\\mathbf{f}_o')" + }, + { + "bbox": [ + 130, + 175, + 482, + 272 + ], + "type": "text", + "content": " ensures the independence of the invariant part of the verb and object features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 272, + 293, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 293, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 293, + 415 + ], + "type": "text", + "content": "Balanced learning of seen & unseen actions. In ZS-CAR, the verb-object connections in the training and test sets differ significantly, as shown in Fig. 4. While the seen connections help optimize the modules, an excessive emphasis on them may hinder generalizing to unseen actions. Our solution is simultaneously enhancing the learning of seen and unseen connections, enabling them to strike an optimal balance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 415, + 292, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 292, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 292, + 449 + ], + "type": "text", + "content": "We first assist the model in learning seen actions explicitly. We calculate the observed conditional" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 450, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 489 + ], + "type": "text", + "content": "probabilities in the training set as the observed conditional scores " + }, + { + "bbox": [ + 130, + 450, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{S}}_{v|o}" + }, + { + "bbox": [ + 130, + 450, + 482, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 450, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{S}}_{o|v}" + }, + { + "bbox": [ + 130, + 450, + 482, + 489 + ], + "type": "text", + "content": ". Then, we force the conditional scores in a training batch closer to " + }, + { + "bbox": [ + 130, + 450, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{S}}_{v|o}" + }, + { + "bbox": [ + 130, + 450, + 482, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 450, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{S}}_{o|v}" + }, + { + "bbox": [ + 130, + 450, + 482, + 489 + ], + "type": "text", + "content": ", with the objectives:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 224, + 497, + 481, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 497, + 481, + 526 + ], + "spans": [ + { + "bbox": [ + 224, + 497, + 481, + 526 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n}, v} = - \\frac {1}{N _ {v}} \\sum_ {v _ {i} \\in V} \\hat {\\mathbf {s}} _ {o | v = v _ {i}} \\log \\mathbf {s} _ {o | v = v _ {i}}, \\tag {7}", + "image_path": "c5d3e6b6831ba8002944760c37f45b9b82f64c3367ce54101a8bb6350abaeff0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 224, + 536, + 481, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 536, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 224, + 536, + 481, + 564 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n}, o} = - \\frac {1}{N _ {o}} \\sum_ {o _ {i} \\in O} \\hat {\\mathbf {s}} _ {v | o = o _ {i}} \\log \\mathbf {s} _ {v | o = o _ {i}}, \\tag {8}", + "image_path": "649109f925043e6ae169fe3f1f6cdc89070b950840ddf7d49852d86136876e35.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 606 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 570, + 482, + 606 + ], + "type": "inline_equation", + "content": "N_v" + }, + { + "bbox": [ + 130, + 570, + 482, + 606 + ], + "type": "text", + "content": " & " + }, + { + "bbox": [ + 130, + 570, + 482, + 606 + ], + "type": "inline_equation", + "content": "N_o" + }, + { + "bbox": [ + 130, + 570, + 482, + 606 + ], + "type": "text", + "content": " denote the verb & object numbers, and the condition loss " + }, + { + "bbox": [ + 130, + 570, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con}" + }, + { + "bbox": [ + 130, + 570, + 482, + 606 + ], + "type": "text", + "content": " is defined as the sum of " + }, + { + "bbox": [ + 130, + 570, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con,v}" + }, + { + "bbox": [ + 130, + 570, + 482, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con,o}" + }, + { + "bbox": [ + 130, + 570, + 482, + 606 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 570, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{s}}_{o|v = v_i}" + }, + { + "bbox": [ + 130, + 570, + 482, + 606 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 570, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{o|v = v_i}" + }, + { + "bbox": [ + 130, + 570, + 482, + 606 + ], + "type": "text", + "content": " denotes the conditional object score vector when verb is " + }, + { + "bbox": [ + 130, + 570, + 482, + 606 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 130, + 570, + 482, + 606 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 570, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{s}}_{v|o = o_i}" + }, + { + "bbox": [ + 130, + 570, + 482, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{v|o = o_i}" + }, + { + "bbox": [ + 130, + 570, + 482, + 606 + ], + "type": "text", + "content": " are defined similarly." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 607, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 607, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 607, + 482, + 667 + ], + "type": "text", + "content": "Then, inspired by the human ability to create new things by imagining compositions of known components, we use CutMix [67] to assemble samples in a training batch as the imagined unseen actions, thereby reducing the risk of overemphasizing seen compositions. Specifically, for an input " + }, + { + "bbox": [ + 130, + 607, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i" + }, + { + "bbox": [ + 130, + 607, + 482, + 667 + ], + "type": "text", + "content": " with label " + }, + { + "bbox": [ + 130, + 607, + 482, + 667 + ], + "type": "inline_equation", + "content": "a_i = (v_l,o_k)" + }, + { + "bbox": [ + 130, + 607, + 482, + 667 + ], + "type": "text", + "content": ", we randomly select another sample " + }, + { + "bbox": [ + 130, + 607, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_j" + }, + { + "bbox": [ + 130, + 607, + 482, + 667 + ], + "type": "text", + "content": " with label " + }, + { + "bbox": [ + 130, + 607, + 482, + 667 + ], + "type": "inline_equation", + "content": "a_j = (v_m,o_n)" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 299, + 295, + 482, + 352 + ], + "blocks": [ + { + "bbox": [ + 299, + 295, + 482, + 352 + ], + "lines": [ + { + "bbox": [ + 299, + 295, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 299, + 295, + 482, + 352 + ], + "type": "image", + "image_path": "1c3d32af904095b36440c6f43eb8a1cd3b7ecbe0519a944234b9574ddda48a26.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 297, + 358, + 482, + 436 + ], + "lines": [ + { + "bbox": [ + 297, + 358, + 482, + 436 + ], + "spans": [ + { + "bbox": [ + 297, + 358, + 482, + 436 + ], + "type": "text", + "content": "Fig. 4: Component compatibility variations. The component relations are different between the training and test sets (left). To solve this (right), we use the observed conditional score to fit seen relations and encourage the model to imagine unseen actions to avoid being limited to seen relations." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "R. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 158, + 168, + 455, + 264 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 2: A comparison with the SOTA methods on Sth-com, using the vision model with word embedding model paradigm. We report both vanilla (only " + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{com}" + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{comp}" + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": ") and enhanced (using the advocated enhanced training strategy) versions of C2C. Since ADE is tailored for transformer architectures, we only adapt it to VideoSwin." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 158, + 168, + 455, + 264 + ], + "lines": [ + { + "bbox": [ + 158, + 168, + 455, + 264 + ], + "spans": [ + { + "bbox": [ + 158, + 168, + 455, + 264 + ], + "type": "table", + "html": "
TSM-18VideoSwin-T
Methodsverbobj.seenunseenHMAUCverbobj.seenunseenHMAUC
TMN [50]41.435.931.628.122.07.333.445.631.030.223.07.9
Compcos [40]45.840.034.835.127.510.636.750.534.237.827.210.9
CGE [44]35.536.427.625.520.05.931.848.131.230.323.17.9
OADis [55]39.140.733.532.325.59.341.648.938.138.330.212.6
ADE [18]------35.349.133.836.226.810.3
CAN [60]40.039.232.431.624.58.732.649.533.033.325.09.2
C2C (Vanilla)47.742.738.139.130.813.247.950.942.143.634.916.4
C2C (Enhanced)50.144.339.441.033.014.548.852.843.444.136.717.4
", + "image_path": "c6773eb28dc02674d44c8c4153d4c4a9da32e11e1b5592e8722ad160c7eee254.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "spans": [ + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": "and crop it at a random position of random size. The cropped part replaces the same position of " + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i" + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": ", generating the augmented input " + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{X}}_i" + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": ". Denoting the cropping ratio as " + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": ", we modify the previous losses as: " + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{L}}_{com} = (1 - \\lambda)\\mathcal{L}_{com}^{a_i} + \\lambda \\mathcal{L}_{com}^{a_j}" + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{L}}_{comp} = (1 - \\lambda)\\mathcal{L}_{comp}^{v_l,o_k} + \\lambda \\mathcal{L}_{comp}^{v_m,o_n}" + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{comp}^{a_i}" + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": " calculates " + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{com}" + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": " using Eq. (3) with " + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "inline_equation", + "content": "a_i" + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": " as groundtruth. " + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{comp}^{v_l,o_k}" + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": " calculates " + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{comp}" + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "inline_equation", + "content": "v_l" + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "inline_equation", + "content": "o_k" + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": " as groundtruth. " + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{comp}^{a_j}" + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{comp}^{v_m,o_n}" + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": " are defined similarly. Also, " + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ind}" + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": " is modified to " + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{L}}_{ind}" + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": " by replacing " + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i" + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "inline_equation", + "content": "(1 - \\lambda)\\mathbf{y}_i + \\lambda \\mathbf{y}_j" + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": " in Eq. (4) and Eq. (5). Particularly, we introduce new compositions " + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\bar{a}_1 = (v_l,o_n)" + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\bar{a}_2 = (v_m,o_k)" + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\bar{a}_1" + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\bar{a}_2" + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": " are not constrained in " + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{test}}" + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{train}}" + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": ". It gives a novel loss " + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{new} = \\mathcal{L}_{com}^{\\bar{a}_1} + \\mathcal{L}_{com}^{\\bar{a}_2}" + }, + { + "bbox": [ + 130, + 287, + 482, + 408 + ], + "type": "text", + "content": ", which forces the model to consider all the possible component connections." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 409, + 482, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 409, + 482, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 482, + 482 + ], + "type": "text", + "content": "A concern is that the new compositions " + }, + { + "bbox": [ + 130, + 409, + 482, + 482 + ], + "type": "inline_equation", + "content": "a_1" + }, + { + "bbox": [ + 130, + 409, + 482, + 482 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 409, + 482, + 482 + ], + "type": "inline_equation", + "content": "a_2" + }, + { + "bbox": [ + 130, + 409, + 482, + 482 + ], + "type": "text", + "content": " may be counterfactual. However, it only introduces noise to the component-to-composition module, as the component labels still match the visual input. In this way, we still reserve these counterfactual labels as they can provide regularization effects that avoid overfitting to seen verb-object relations. Besides, it is also consistent with human imagination which is not only constrained within reasonable compositions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 500, + 224, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 224, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 224, + 511 + ], + "type": "text", + "content": "4.4 More details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 481, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 545 + ], + "type": "text", + "content": "Network details. Following previous practice, we implement the video encoder and component label encoding with two paradigms:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 130, + 545, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 593 + ], + "type": "text", + "content": "(1) Vision model with word embedding model. In this paradigm, we adopt an action recognition model (the CNN-based TSM-18 [35] or the transformer-based VideoSwin-T [37]) to extract the general video feature and a word embedding model (fasttext [3]) to encode the labels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "content": "(2) Vision-language models. Recently, large pre-trained vision-language models, such as CLIP [53], have shown superior power. In this paradigm, we adopt the language branch of recent prompt-based compositional generalization methods (Coop [70], CSP [45] and SPM [38]) to encode the labels. As CLIP are trained using image-text pairs, we apply image-to-video adapters introduced in AIM [66] to adapt the CLIP visual encoder (ViT-B/32 [10]) to video modality." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 271, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 271, + 91, + 448, + 102 + ], + "type": "text", + "content": "C2C: Component-to-Composition Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 217, + 135, + 394, + 240 + ], + "blocks": [ + { + "bbox": [ + 154, + 114, + 457, + 125 + ], + "lines": [ + { + "bbox": [ + 154, + 114, + 457, + 125 + ], + "spans": [ + { + "bbox": [ + 154, + 114, + 457, + 125 + ], + "type": "text", + "content": "Table 3: A comparison with the SOTA CLIP-based methods on Sth-com." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 217, + 135, + 394, + 240 + ], + "lines": [ + { + "bbox": [ + 217, + 135, + 394, + 240 + ], + "spans": [ + { + "bbox": [ + 217, + 135, + 394, + 240 + ], + "type": "table", + "html": "
methodsverbobj.seenunseenHMAUC
Coop [70]48.455.143.946.436.618.1
+C2C(vanilla)56.256.948.354.642.724.0
+C2C(enhance)56.358.950.756.644.526.0
CSP [45]49.054.843.647.436.018.0
+C2C(vanilla)56.356.649.053.442.523.8
+C2C(enhance)57.158.649.956.244.325.5
SPM [38]48.954.544.046.335.517.7
+DFM [38]47.455.443.147.335.817.9
+C2C(vanilla)55.757.349.353.442.223.8
+C2C(enhance)56.558.449.955.743.825.2
", + "image_path": "4ce21c17755a8161a5fe4065a3b82344555080de4122dfba5f0553edc1a9af47.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 262, + 482, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 299 + ], + "type": "text", + "content": "Training Objectives. When applying the proposed enhanced training strategy, we perform CutMix to generate new compositions with a probability " + }, + { + "bbox": [ + 130, + 262, + 482, + 299 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 262, + 482, + 299 + ], + "type": "text", + "content": ". If apply CutMix, the final loss is calculated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 212, + 306, + 481, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 306, + 481, + 321 + ], + "spans": [ + { + "bbox": [ + 212, + 306, + 481, + 321 + ], + "type": "interline_equation", + "content": "\\bar {\\mathcal {L}} _ {t o t a l} = \\bar {\\mathcal {L}} _ {c o m} + \\alpha \\bar {\\mathcal {L}} _ {c o m p} + \\beta \\bar {\\mathcal {L}} _ {i n d} + \\gamma \\mathcal {L} _ {n e w}. \\tag {9}", + "image_path": "3f163864323895a80b81fdd0e1d39714666564349c8f0521809da8652ce393a7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 328, + 234, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 328, + 234, + 340 + ], + "spans": [ + { + "bbox": [ + 131, + 328, + 234, + 340 + ], + "type": "text", + "content": "If not, the final loss is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 214, + 350, + 480, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 350, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 214, + 350, + 480, + 363 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c o m}} + \\alpha \\mathcal {L} _ {\\text {c o m p}} + \\beta \\mathcal {L} _ {\\text {i n d}} + \\gamma \\mathcal {L} _ {\\text {c o n}}. \\tag {10}", + "image_path": "d1013c10a5babe17d1044d41d939828230e00c89b05556378f10e421ac147cc0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 371, + 481, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 481, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 481, + 395 + ], + "type": "text", + "content": "As " + }, + { + "bbox": [ + 130, + 371, + 481, + 395 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{new}" + }, + { + "bbox": [ + 130, + 371, + 481, + 395 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 371, + 481, + 395 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con}" + }, + { + "bbox": [ + 130, + 371, + 481, + 395 + ], + "type": "text", + "content": " respectively guide the model focus on unseen and seen actions, we use the same coefficient " + }, + { + "bbox": [ + 130, + 371, + 481, + 395 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 371, + 481, + 395 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 413, + 230, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 413, + 230, + 427 + ], + "spans": [ + { + "bbox": [ + 131, + 413, + 230, + 427 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 437, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 544 + ], + "type": "text", + "content": "Evaluation Metrics. The metrics include Verb, Object, Seen, Unseen, HM and AUC. Verb and Object respectively denote the accuracy of verbs and objects. To gain other metrics, we first add a constant to the seen actions scores. The constant ranges from negative infinity to positive infinity, making the model shift from completely biased to unseen actions to completely biased to seen actions and yielding the seen-unseen curve shown in Fig. 5. Seen and Unseen are the best accuracy of seen/unseen actions, i.e. the intersection points with x/y-axis in Fig. 5. HM is the best harmonic mean with formulation: " + }, + { + "bbox": [ + 130, + 437, + 482, + 544 + ], + "type": "inline_equation", + "content": "HM = 2(Seen * Unseen) / (Seen + Unseen)" + }, + { + "bbox": [ + 130, + 437, + 482, + 544 + ], + "type": "text", + "content": ". AUC is the area under the curve." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 544, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 582 + ], + "type": "text", + "content": "Implementation details. " + }, + { + "bbox": [ + 130, + 544, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 544, + 482, + 582 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 544, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 544, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 544, + 482, + 582 + ], + "type": "text", + "content": " in Eq. (9) and Eq. (10) are 0.2, 0.1 and 0.1, respectively. The probability " + }, + { + "bbox": [ + 130, + 544, + 482, + 582 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 544, + 482, + 582 + ], + "type": "text", + "content": " of conducting CutMix is 0.7. More coefficient ablations and implementation details are reported in supplementary materials." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 597, + 337, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 597, + 337, + 611 + ], + "spans": [ + { + "bbox": [ + 131, + 597, + 337, + 611 + ], + "type": "text", + "content": "5.1 Comparisons with State-of-the-Art" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Note that the existing compositional action recognition methods cannot apply to the ZS-CAR task. To comprehensively demonstrate the proposed C2C, we first adapt the existing CZSL methods to the proposed Sth-com dataset. We replace their image encoders with video encoders and adopt the same word" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "R. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 231, + 135, + 380, + 196 + ], + "blocks": [ + { + "bbox": [ + 181, + 114, + 430, + 125 + ], + "lines": [ + { + "bbox": [ + 181, + 114, + 430, + 125 + ], + "spans": [ + { + "bbox": [ + 181, + 114, + 430, + 125 + ], + "type": "text", + "content": "Table 4: A comparison with the SOTA methods on C-GQA." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 231, + 135, + 380, + 196 + ], + "lines": [ + { + "bbox": [ + 231, + 135, + 380, + 196 + ], + "spans": [ + { + "bbox": [ + 231, + 135, + 380, + 196 + ], + "type": "table", + "html": "
multiseenunseenHMAUC
CGE [44]×38.017.118.55.4
OADis [55]×38.319.820.17.0
CoT [26]✓39.222.722.17.4
C2C×39.123.023.07.7
C2C✓39.423.623.38.2
", + "image_path": "40f9618959d02742b46cbd8b4138ce65c779c72155adfca6ee31635cc3e3ec23.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 209, + 231, + 402, + 300 + ], + "blocks": [ + { + "bbox": [ + 139, + 209, + 472, + 220 + ], + "lines": [ + { + "bbox": [ + 139, + 209, + 472, + 220 + ], + "spans": [ + { + "bbox": [ + 139, + 209, + 472, + 220 + ], + "type": "text", + "content": "Table 5: Results of different hypotheses inferring compositions from components." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 209, + 231, + 402, + 300 + ], + "lines": [ + { + "bbox": [ + 209, + 231, + 402, + 300 + ], + "spans": [ + { + "bbox": [ + 209, + 231, + 402, + 300 + ], + "type": "table", + "html": "
verbobj.seenunseenHMAUC
Compcos [40]45.840.034.835.127.510.6
Independent47.942.237.437.429.812.2
Knowledge-agnostic48.443.138.736.830.212.6
Dynamic centered46.742.537.438.029.912.5
Static centered46.641.137.036.728.911.7
Full47.742.738.139.130.813.2
", + "image_path": "69d4fc5f2d531e6f06b3119081459611c96e46cb952a2072fc11da6e4c36c460.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 326, + 480, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 480, + 373 + ], + "type": "text", + "content": "embedding model, making sure only the compositional generalization parts are different between the adapted methods and our proposed methods. Besides, we also modify C2C and compare it with the existing SOTA methods on the image-based CZSL task using the challenging C-GQA [44] dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 375, + 480, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 480, + 435 + ], + "type": "text", + "content": "Comparisons on Sth-com. We first compare our method with the adapted CZSL methods of the vision model with the word embedding paradigm, using both CNN-based and transformer-based video encoders. The results are reported in Tab. 2. The vanilla version of C2C uses two basic losses. The enhanced version uses the proposed enhanced training strategy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 437, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 481, + 628 + ], + "type": "text", + "content": "Tab. 2 demonstrate that our method outperforms the existing methods by significant margins across all the metrics on ZS-CAR. Furthermore, the proposed enhanced training strategy provides consistent performance gains across different video encoders. Tab. 2 shows that the recently proposed methods based on Compcos are inferior to Compcos. Second, except for OADis, using the recently proposed VideoSwin-T backbone does not yield significant improvements. The first observation may stem from failing to generalise the embedding relations, not disentangling spatial and temporal information well, and not applying the conditional component relations in actions. Besides, jointly learning the compositions may lead to unbalanced learning of the components. These deficiencies degrade the effectiveness of video encoders, leading directly to the second observation. Tab. 2 also compares two video encoders. Though VideoSwin is an advanced model, its verb accuracy is consistently inferior or similar to TSM across all methods. Conversely, it outperforms TSM in object accuracy. This implies that our ZS-CAR task offers a new avenue for assessing biases regarding dynamic and static patterns in action recognition models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": "Then we apply our C2C method to recently proposed CLIP-based compositional zero-shot learning methods. The results are reported in Tab. 3. Coop, CSP and SPM are three text prompting methods. We can see that the proposed" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "type": "text", + "content": "C2C: Component-to-Composition Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "C2C method can improve their performance significantly on the ZS-CAR task. DFM [38] was proposed along with SPM. It enables cross-modal interactions between language and visual features for better composition prediction. However, the improvements are slight on the ZS-CAR task. Compared to it, our C2C method provides more significant improvements." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 299, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 299, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 299, + 343 + ], + "type": "text", + "content": "Comparison on C-GQA. To further evaluate the proposed C2C method, we modify it for the image-based compositional zero-shot learning task by deleting all the temporal-related operations. We report the results on the recently proposed challenging benchmark C-GQA in Tab. 4, with fine-tuned ViT-B [10] as the backbone. In the table, the term 'multi' denotes the use of multi-level backbone outputs. When using multi-level outputs in C2C, we merely average them, much simpler than CoT." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 343, + 482, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 404 + ], + "type": "text", + "content": "Though designed for video analysis, the proposed C2C exhibits superior performance over existing SOTA methods. Like CoT [26], we also try to utilize multi-level backbone outputs. But compared to the specially designed module in CoT, we merely average the multi-level outputs. This demonstrates that our C2C method can be a general solution to the compositional generalization problem." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 323, + 202, + 462, + 287 + ], + "blocks": [ + { + "bbox": [ + 323, + 202, + 462, + 287 + ], + "lines": [ + { + "bbox": [ + 323, + 202, + 462, + 287 + ], + "spans": [ + { + "bbox": [ + 323, + 202, + 462, + 287 + ], + "type": "image", + "image_path": "51f290a908be2a3536cefab74079e55c2d0da19a2a292d7acb7e84d508057d11.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 297, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 304, + 297, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 482, + 331 + ], + "type": "text", + "content": "Fig. 5: Seen-unseen curves, drawn by gradually forcing the model biased from unseen to seen actions." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 419, + 239, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 419, + 239, + 432 + ], + "spans": [ + { + "bbox": [ + 131, + 419, + 239, + 432 + ], + "type": "text", + "content": "5.2 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 437, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 545 + ], + "type": "text", + "content": "Different composition generation methods. We first compare different methods that can be used to generate compositions from verb and object components, including: (1) Independent supposes components are completely independent, so the inference is straightforward: " + }, + { + "bbox": [ + 130, + 437, + 482, + 545 + ], + "type": "inline_equation", + "content": "s_{c,i} = s_{v,i} \\cdot s_{o,i}" + }, + { + "bbox": [ + 130, + 437, + 482, + 545 + ], + "type": "text", + "content": ". (2) Knowledge-agnostic assumes components are not completely independent, but the component relations are agnostic: " + }, + { + "bbox": [ + 130, + 437, + 482, + 545 + ], + "type": "inline_equation", + "content": "s_{c,i} = s_{v,i} \\cdot 1 + s_{o,i} \\cdot 1" + }, + { + "bbox": [ + 130, + 437, + 482, + 545 + ], + "type": "text", + "content": ". (3) One path inference considers the compatibility between compositions but only relies on either inference path, i.e., the dynamics or the static path result (Fig. 2). (4) Full is the final scheme, integrating both paths." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "For a fair comparison, the experiments only use the base losses " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{com}}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{comp}}" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": ". The results are reported in Tab. 5. We use Compcos as the baseline result. The simplest inference method, i.e., Independent, has surpassed Compcos on all the metrics. When using the knowledge-agnostic hypothesis, the seen accuracy and holistic metrics are better. But its unseen accuracy is lower. By considering the component relations, utilizing both inference paths substantially improves the unseen accuracy, yielding the best results for both HM and AUC. The results suggest that our proposed inference approach achieves a more balanced performance between seen and unseen actions. Fig. 5 reports the seen-unseen curves, implying that the proposed method is more robust." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "R. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 235, + 135, + 376, + 194 + ], + "blocks": [ + { + "bbox": [ + 157, + 114, + 455, + 125 + ], + "lines": [ + { + "bbox": [ + 157, + 114, + 455, + 125 + ], + "spans": [ + { + "bbox": [ + 157, + 114, + 455, + 125 + ], + "type": "text", + "content": "Table 6: Effectiveness of the elements in the enhanced training strategy." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 235, + 135, + 376, + 194 + ], + "lines": [ + { + "bbox": [ + 235, + 135, + 376, + 194 + ], + "spans": [ + { + "bbox": [ + 235, + 135, + 376, + 194 + ], + "type": "table", + "html": "
\\( \\mathcal{L}_{com} \\)\\( \\mathcal{L}_{comp} \\)\\( \\mathcal{L}_{ind} \\)CutMix\\( \\mathcal{L}_{con} \\)\\( \\mathcal{L}_{new} \\)HMAUC
✓29.812.5
✓✓30.813.2
✓✓✓31.413.6
✓✓✓✓31.513.9
✓✓✓✓✓32.114.0
✓✓✓✓✓✓33.014.5
", + "image_path": "4c0d7ec66e06244cc3c3743541ec847ab774d262bfaa0b9e4db16fa7f333fc95.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 217, + 231, + 394, + 269 + ], + "blocks": [ + { + "bbox": [ + 140, + 209, + 471, + 220 + ], + "lines": [ + { + "bbox": [ + 140, + 209, + 471, + 220 + ], + "spans": [ + { + "bbox": [ + 140, + 209, + 471, + 220 + ], + "type": "text", + "content": "Table 7: The accuracy (%) on actions containing verbs that cause deformations." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 217, + 231, + 394, + 269 + ], + "lines": [ + { + "bbox": [ + 217, + 231, + 394, + 269 + ], + "spans": [ + { + "bbox": [ + 217, + 231, + 394, + 269 + ], + "type": "table", + "html": "
methodsverbobjectcomposition
seenunseenseenunseenseenunseen
Vanilla70.160.268.734.856.927.5
Vanilla + Lind71.865.268.245.158.036.7
", + "image_path": "5e117a1b4ed0aa685ee372d5325c63306b63cae988b02898a91266e7f246aec1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 183, + 316, + 427, + 381 + ], + "blocks": [ + { + "bbox": [ + 131, + 284, + 480, + 307 + ], + "lines": [ + { + "bbox": [ + 131, + 284, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 131, + 284, + 480, + 307 + ], + "type": "text", + "content": "Table 8: Unbiased accuracy (%) of seen and unseen actions and the corresponding harmonic mean (HM)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 183, + 316, + 427, + 381 + ], + "lines": [ + { + "bbox": [ + 183, + 316, + 427, + 381 + ], + "spans": [ + { + "bbox": [ + 183, + 316, + 427, + 381 + ], + "type": "table", + "html": "
methodsseenunseenHM
baseline36.817.223.5
+Lcon36.5(-0.3)16.4(-0.8)22.6(-0.9)
+cutmix35.7(-1.1)17.7(+0.5)23.6(+0.1)
+cutmix+Lcon36.1(+0.4)17.6(-0.1)23.7(+0.1)
+cutmix+Lnew37.0(+1.3)18.2(+0.5)24.4(+0.8)
+cutmix+Lcon+Lnew37.5(+0.5/+1.4)18.7(+0.5/+1.1)24.9(+0.5/+1.2)
", + "image_path": "09993f07d534ae0917c9076eea0e80e9a28afc71014ba1c5bd40a1d63dd92103.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 409, + 480, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 409, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 480, + 445 + ], + "type": "text", + "content": "The enhanced training strategy. In Tab. 6, we demonstrate the effectiveness of our enhanced training strategy. It shows that the results exhibit consistent improvements as gradually applying the proposed strategy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 448, + 482, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 519 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 519 + ], + "type": "text", + "content": "To further validate the proposed method, we verify the merits of the independent loss " + }, + { + "bbox": [ + 130, + 448, + 482, + 519 + ], + "type": "inline_equation", + "content": "L_{ind}" + }, + { + "bbox": [ + 130, + 448, + 482, + 519 + ], + "type": "text", + "content": " on actions containing verbs that can cause object deformations (such as squeeze and twist). The results are shown in Tab. 7. They show that the unseen accuracy improves significantly after adding " + }, + { + "bbox": [ + 130, + 448, + 482, + 519 + ], + "type": "inline_equation", + "content": "L_{ind}" + }, + { + "bbox": [ + 130, + 448, + 482, + 519 + ], + "type": "text", + "content": ". However, the object accuracy of seen actions slightly decreases. This might be because spurious features are useful for recognizing seen components." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": "Next, we validate the proposed balanced learning of actions. With the vanilla version with " + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "inline_equation", + "content": "L_{ind}" + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": " as the baseline, we report the accuracy of all seen and unseen actions and HM without adding bias in Tab. 8. It shows that adding only " + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "inline_equation", + "content": "L_{con}" + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": " leads to performance degradation, but the degradation in seen accuracy is minor. Similarly, adding just CutMix results in a slight improvement, but with an obvious degradation in seen accuracy. When using CutMix, we add either " + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "inline_equation", + "content": "L_{con}" + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "inline_equation", + "content": "L_{new}" + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": ". The results indicate that " + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "inline_equation", + "content": "L_{new}" + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": " yields a universal improvement, while " + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "inline_equation", + "content": "L_{con}" + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": " contributes positively to seen accuracy. Last, when " + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "inline_equation", + "content": "L_{new}" + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "inline_equation", + "content": "L_{con}" + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": " are integrated, they both bring improvements on all the metrics, and the gains are more pronounced than adding either one alone. This indicates that our training strategy enables the model to achieve a better balance between seen and unseen actions." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "type": "text", + "content": "C2C: Component-to-Composition Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 115, + 217, + 197 + ], + "blocks": [ + { + "bbox": [ + 145, + 115, + 217, + 197 + ], + "lines": [ + { + "bbox": [ + 145, + 115, + 217, + 197 + ], + "spans": [ + { + "bbox": [ + 145, + 115, + 217, + 197 + ], + "type": "image", + "image_path": "9525e3d9e569ed9ad921650b66893113037999dd8a211cc7da63a621d5b280e3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 208, + 477, + 231 + ], + "lines": [ + { + "bbox": [ + 135, + 208, + 477, + 231 + ], + "spans": [ + { + "bbox": [ + 135, + 208, + 477, + 231 + ], + "type": "text", + "content": "Fig. 6: The top-3 compatible components Fig. 7: Examples of successes and failures with the same verb or object components. of our methods." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 225, + 115, + 296, + 200 + ], + "blocks": [ + { + "bbox": [ + 225, + 115, + 296, + 200 + ], + "lines": [ + { + "bbox": [ + 225, + 115, + 296, + 200 + ], + "spans": [ + { + "bbox": [ + 225, + 115, + 296, + 200 + ], + "type": "image", + "image_path": "efd711524ac51cac2ad16c6e86d8a9e08f1cef4a5f48594eeb71095c35553d2c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 329, + 114, + 389, + 200 + ], + "blocks": [ + { + "bbox": [ + 329, + 114, + 389, + 200 + ], + "lines": [ + { + "bbox": [ + 329, + 114, + 389, + 200 + ], + "spans": [ + { + "bbox": [ + 329, + 114, + 389, + 200 + ], + "type": "image", + "image_path": "da02af2561f47b06a6167813210502738e9359cccf957c304621a965a3ef6227.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 392, + 118, + 452, + 200 + ], + "blocks": [ + { + "bbox": [ + 392, + 118, + 452, + 200 + ], + "lines": [ + { + "bbox": [ + 392, + 118, + 452, + 200 + ], + "spans": [ + { + "bbox": [ + 392, + 118, + 452, + 200 + ], + "type": "image", + "image_path": "acd696c7cf7c8d807d18bed514368e24a0a77aff00e7958fee93251f037926a8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 257, + 260, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 260, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 260, + 269 + ], + "type": "text", + "content": "5.3 Qualitative analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 284, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 403 + ], + "type": "text", + "content": "In Fig. 6, we show the top three compatible components for different samples, while keeping another component the same. When the verb is squeeze, for the first sample, the second and third compatible objects are cloth and paper, which are close to the visual contents and compatible with squeeze. For the second sample, the model identifies different objects. The second object basket is not included in the visual content but is compatible with squeeze. The third object bed cannot be combined with squeeze but is visually close to the sample. When the fixed component is an object, the situation is similar. It demonstrates that the proposed model can identify component compatibility according to both visual contents and component attributions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 405, + 483, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 483, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 483, + 513 + ], + "type": "text", + "content": "In Fig. 7, we visualize the top three recognized actions of some successful and failed samples. In the top row, our model accurately predicts the given actions. Besides, the second and third predictions relate to the visual content, such as fold paper for the first sample and tear paper just a little bit for the second sample. The bottom row shows some failures. Though wrong, the predictions still reflect the video content, such as move part of plant and tear leaf into two pieces for the second example. However, the failures also demonstrate some limitations of the proposed method. For example, it struggles with the task of distinguishing objects or verbs of similar visual appearances." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 537, + 220, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 537, + 220, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 537, + 220, + 550 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "This paper introduces a novel task, Zero-Shot Compositional Action Recognition (ZS-CAR), that investigates the compositional generalization in the video modality. A benchmarking dataset Sth-com and a baseline method C2C are proposed along with the novel task. Besides, an enhanced training strategy is presented to address two inherent challenges in ZS-CAR: component domain variations and component compatibility variations. The approach demonstrates superior performance compared to existing compositional generalization solutions in both image- and video-based zero-shot compositional learning settings." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "R. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 482, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 482, + 216 + ], + "type": "text", + "content": "This work is supported in part by the National Key Research and Development Program of China (2023YFF1105102, 2023YFF1105105), the National Natural Science Foundation of China (62020106012, 62332008, 62106089, U1836218, 62336004), the 111 Project of Ministry of Education of China (B12018), the Postgraduate Research & Practice Innovation Program of Jiangsu Province (KYCX22_2307), and the UK EPSRC (EP/V002856/1, EP/T022205/1)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 236, + 197, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 197, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 197, + 248 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 263, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 263, + 480, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 263, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 263, + 480, + 297 + ], + "type": "text", + "content": "1. Arnab, A., Dehghani, M., Heigold, G., Sun, C., Lučić, M., Schmid, C.: Vivit: A video vision transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6836-6846 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 298, + 480, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 298, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 138, + 298, + 480, + 330 + ], + "type": "text", + "content": "2. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: International Conference on Machine Learning. pp. 813-824 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 331, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 480, + 364 + ], + "type": "text", + "content": "3. Bojanowski, P., Grave, E., Joulin, A., Mikolov, T.: Enriching word vectors with subword information. Transactions of the association for computational linguistics 5, 135-146 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 365, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 407 + ], + "type": "text", + "content": "4. Brattoli, B., Tighe, J., Zhdanov, F., Perona, P., Chalupka, K.: Rethinking zero-shot video classification: End-to-end training for realistic applications. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4613-4623 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 409, + 480, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 409, + 480, + 442 + ], + "spans": [ + { + "bbox": [ + 138, + 409, + 480, + 442 + ], + "type": "text", + "content": "5. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 443, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 443, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 443, + 480, + 475 + ], + "type": "text", + "content": "6. Chao, W.L., Changpinyo, S., Gong, B., Sha, F.: An empirical study and analysis of generalized zero-shot learning for object recognition in the wild. In: Proceedings of the European Conference on Computer Vision. pp. 52-68 (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 476, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 476, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 138, + 476, + 480, + 508 + ], + "type": "text", + "content": "7. Chen, S., Huang, D.: Elaborative rehearsal for zero-shot action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13638-13647 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 510, + 448, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 448, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 448, + 521 + ], + "type": "text", + "content": "8. Chomsky, N.: Aspects of the Theory of Syntax. No. 11, MIT press (2014)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 521, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 480, + 553 + ], + "type": "text", + "content": "9. Doshi, K., Yilmaz, Y.: Zero-shot action recognition with transformer-based video semantic embedding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4858-4867 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 555, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 555, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 133, + 555, + 480, + 598 + ], + "type": "text", + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 599, + 480, + 632 + ], + "type": "text", + "content": "11. Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6824-6835 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "text", + "content": "12. Feichtenhofer, C.: X3d: Expanding architectures for efficient video recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 203-213 (2020)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "type": "text", + "content": "C2C: Component-to-Composition Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "13. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6202-6211 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "type": "text", + "content": "14. Fodor, J.A., Pylyshyn, Z.W.: Connectionism and cognitive architecture: A critical analysis. Cognition 28(1-2), 3-71 (1988)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 203 + ], + "type": "text", + "content": "15. Fu, Y., Hospedales, T.M., Xiang, T., Gong, S.: Attribute learning for understanding unstructured social activity. In: Proceedings of the European Conference on Computer Vision. pp. 530-543 (2012)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 203, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 203, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 134, + 203, + 480, + 257 + ], + "type": "text", + "content": "16. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The\" something something\" video database for learning and evaluating visual common sense. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5842-5850 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 257, + 480, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 257, + 480, + 290 + ], + "spans": [ + { + "bbox": [ + 134, + 257, + 480, + 290 + ], + "type": "text", + "content": "17. Gretton, A., Bousquet, O., Smola, A., Scholkopf, B.: Measuring statistical dependence with hilbert-schmidt norms. In: International Conference on Algorithmic Learning Theory. pp. 63-77 (2005)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 290, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 290, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 134, + 290, + 480, + 322 + ], + "type": "text", + "content": "18. Hao, S., Han, K., Wong, K.Y.K.: Learning attention as disentangler for compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15315-15324 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 322, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 322, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 134, + 322, + 480, + 354 + ], + "type": "text", + "content": "19. Hara, K., Kataoka, H., Satoh, Y.: Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet? In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6546-6555 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 354, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 354, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 134, + 354, + 480, + 374 + ], + "type": "text", + "content": "20. Heckman, J.J.: Sample selection bias as a specification error. Econometrica: Journal of the econometric society pp. 153-161 (1979)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 134, + 374, + 480, + 407 + ], + "type": "text", + "content": "21. Isola, P., Lim, J.J., Adelson, E.H.: Discovering states and transformations in image collections. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1383-1391 (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 407, + 480, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 407, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 134, + 407, + 480, + 440 + ], + "type": "text", + "content": "22. Ji, J., Krishna, R., Fei-Fei, L., Niebles, J.C.: Action genome: Actions as compositions of spatio-temporal scene graphs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10236-10247 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 440, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 440, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 134, + 440, + 480, + 472 + ], + "type": "text", + "content": "23. Ji, S., Xu, W., Yang, M., Yu, K.: 3d convolutional neural networks for human action recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 35(1), 221-231 (2012)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 472, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 472, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 134, + 472, + 480, + 504 + ], + "type": "text", + "content": "24. Karthik, S., Mancini, M., Akata, Z.: Kg-sp: Knowledge guided simple primitives for open world compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9336-9345 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 134, + 504, + 480, + 536 + ], + "type": "text", + "content": "25. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 134, + 536, + 480, + 568 + ], + "type": "text", + "content": "26. Kim, H., Lee, J., Park, S., Sohn, K.: Hierarchical visual primitive experts for compositional zero-shot learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5675-5685 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 134, + 568, + 480, + 601 + ], + "type": "text", + "content": "27. Kwon, H., Kim, M., Kwak, S., Cho, M.: Motionsqueeze: Neural motion feature learning for video understanding. In: Proceedings of the European Conference on Computer Vision. pp. 345-362 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 480, + 633 + ], + "type": "text", + "content": "28. Kwon, H., Kim, M., Kwak, S., Cho, M.: Learning self-similarity in space and time as generalized motion for video action recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13065-13075 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "29. Li, K., Wang, Y., Gao, P., Song, G., Liu, Y., Li, H., Qiao, Y.: Uniformer: Unified transformer for efficient spatiotemporal representation learning. arXiv preprint arXiv:2201.04676 (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "R. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "30. Li, R.C., Wu, X.J., Wu, C., Xu, T.Y., Kittler, J.: Dynamic information enhancement for video classification. Image and Vision Computing 114, 104244 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 480, + 159 + ], + "type": "text", + "content": "31. Li, R., Wu, X.J., Xu, T.: Video is graph: Structured graph module for video action recognition. arXiv preprint arXiv:2110.05904 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 480, + 191 + ], + "type": "text", + "content": "32. Li, R., Xu, T., Wu, X.J., Shen, Z., Kittler, J.: Perceiving actions via temporal video frame pairs. ACM Transactions on Intelligent Systems and Technology 15(3), 1-20 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 224 + ], + "type": "text", + "content": "33. Li, X., Yang, X., Wei, K., Deng, C., Yang, M.: Siamese contrastive embedding network for compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9326-9335 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 225, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 256 + ], + "type": "text", + "content": "34. Li, Y., Ji, B., Shi, X., Zhang, J., Kang, B., Wang, L.: Tea: Temporal excitation and aggregation for action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 909-918 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "type": "text", + "content": "35. Lin, J., Gan, C., Han, S.: Tsm: Temporal shift module for efficient video understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7083-7093 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 289, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 320 + ], + "type": "text", + "content": "36. Liu, J., Kuipers, B., Savarese, S.: Recognizing human actions by attributes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3337-3344 (2011)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 321, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 480, + 353 + ], + "type": "text", + "content": "37. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3202-3211 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 354, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 385 + ], + "type": "text", + "content": "38. Lu, X., Guo, S., Liu, Z., Guo, J.: Decomposed soft prompt guided fusion enhancing for compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23560-23569 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "text", + "content": "39. Ma, L., Zheng, Y., Zhang, Z., Yao, Y., Fan, X., Ye, Q.: Motion stimulation for compositional action recognition. IEEE Transactions on Circuits and Systems for Video Technology (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 418, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 449 + ], + "type": "text", + "content": "40. Mancini, M., Naeem, M.F., Xian, Y., Akata, Z.: Open world compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5222-5230 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 450, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 492 + ], + "type": "text", + "content": "41. Mandal, D., Narayan, S., Dwivedi, S.K., Gupta, V., Ahmed, S., Khan, F.S., Shao, L.: Out-of-distribution detection for generalized zero-shot action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9985-9993 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 493, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 480, + 536 + ], + "type": "text", + "content": "42. Materzynska, J., Xiao, T., Herzig, R., Xu, H., Wang, X., Darrell, T.: Something-else: Compositional action recognition with spatial-temporal interaction networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1049-1059 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "text", + "content": "43. Misra, I., Gupta, A., Hebert, M.: From red wine to red tomato: Composition with context. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1792-1801 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 569, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 480, + 601 + ], + "type": "text", + "content": "44. Naeem, M.F., Xian, Y., Tombari, F., Akata, Z.: Learning graph embeddings for compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 953-962 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "text", + "content": "45. Nayak, N.V., Yu, P., Bach, S.: Learning to compose soft prompts for compositional zero-shot learning. In: The Eleventh International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "46. Neimark, D., Bar, O., Zohar, M., Asselmann, D.: Video transformer network. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3163-3172 (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "type": "text", + "content": "C2C: Component-to-Composition Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 148 + ], + "type": "text", + "content": "47. Pan, J., Lin, Z., Zhu, X., Shao, J., Li, H.: St-adapter: Parameter-efficient image-to-video transfer learning. Advances in Neural Information Processing Systems 35, 26462-26477 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "text", + "content": "48. Park, J., Lee, J., Sohn, K.: Dual-path adaptation from image to video transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2203-2213 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "type": "text", + "content": "49. Piergiovanni, A., Kuo, W., Angelova, A.: Rethinking video vits: Sparse video tubes for joint image and video learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2214-2224 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 246 + ], + "type": "text", + "content": "50. Purushwalkam, S., Nickel, M., Gupta, A., Ranzato, M.: Task-driven modular networks for zero-shot compositional learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3593-3602 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "type": "text", + "content": "51. Qian, R., Lin, W., See, J., Li, D.: Controllable augmentations for video representation learning. Visual Intelligence 2(1), 1 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "type": "text", + "content": "52. Qian, Y., Yu, L., Liu, W., Hauptmann, A.G.: Rethinking zero-shot action recognition: Learning from latent atomic actions. In: Proceedings of the European Conference on Computer Vision. pp. 104-120 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 342 + ], + "type": "text", + "content": "53. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "text", + "content": "54. Ruis, F., Burghouts, G., Bucur, D.: Independent prototype propagation for zero-shot compositionality. Advances in Neural Information Processing Systems 34, 10641-10653 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "text", + "content": "55. Saini, N., Pham, K., Shrivastava, A.: Disentangling visual embeddings for attributes and objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13658-13667 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "type": "text", + "content": "56. Sun, P., Wu, B., Li, X., Li, W., Duan, L., Gan, C.: Counterfactual debiasing inference for compositional action recognition. In: Proceedings of the ACM International Conference on Multimedia. pp. 3220-3228 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 439, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 480, + 472 + ], + "type": "text", + "content": "57. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4489-4497 (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 504 + ], + "type": "text", + "content": "58. Wang, L., Tong, Z., Ji, B., Wu, G.: Tdn: Temporal difference networks for efficient action recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1895-1904 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "text", + "content": "59. Wang, L., Xiong, Y., Wang, Z., Qiao, Y., Lin, D., Tang, X., Van Gool, L.: Temporal segment networks: Towards good practices for deep action recognition. In: Proceedings of the European Conference on Computer Vision. pp. 20-36 (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 536, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 579 + ], + "type": "text", + "content": "60. Wang, Q., Liu, L., Jing, C., Chen, H., Liang, G., Wang, P., Shen, C.: Learning conditional attributes for compositional zero-shot learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11197-11206 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "type": "text", + "content": "61. Xie, S., Sun, C., Huang, J., Tu, Z., Murphy, K.: Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In: Proceedings of the European Conference on Computer Vision. pp. 305–321 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "62. Xu, T., Zhu, X.F., Wu, X.J.: Learning spatio-temporal discriminative model for affine subspace based visual object tracking. Visual Intelligence 1(1), 4 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "63. Xu, X., Hospedales, T., Gong, S.: Transductive zero-shot action recognition by word-vector embedding. International Journal of Computer Vision 123, 309-333 (2017)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "R. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 315 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "64. Yan, R., Huang, P., Shu, X., Zhang, J., Pan, Y., Tang, J.: Look less think more: Rethinking compositional action recognition. In: Proceedings of the ACM International Conference on Multimedia. pp. 3666-3675 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "65. Yan, R., Xie, L., Shu, X., Zhang, L., Tang, J.: Progressive instance-aware feature learning for compositional action recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "type": "text", + "content": "66. Yang, T., Zhu, Y., Xie, Y., Zhang, A., Chen, C., Li, M.: Aim: Adapting image models for efficient video action recognition. arXiv preprint arXiv:2302.03024 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "text", + "content": "67. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6023-6032 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "text", + "content": "68. Zadrozny, B.: Learning and evaluating classifiers under sample selection bias. In: Proceedings of the International Conference on Machine Learning. p. 114 (2004)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "text", + "content": "69. Zellers, R., Choi, Y.: Zero-shot activity recognition with verb attribute induction. In: Proceedings of the Conference on Empirical Methods in Natural Language Processing. pp. 946-958 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 293, + 482, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 315 + ], + "type": "text", + "content": "70. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 271, + 91, + 447, + 102 + ], + "type": "text", + "content": "C2C: Component-to-Composition Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CARB-Net_ Camera-Assisted Radar-Based Network for Vulnerable Road User Detection/ed9a2da4-1d84-4538-8283-1656bf597855_content_list.json b/2024/CARB-Net_ Camera-Assisted Radar-Based Network for Vulnerable Road User Detection/ed9a2da4-1d84-4538-8283-1656bf597855_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..72685ecc883469a276ee3b84b97453d4fbdeaf2e --- /dev/null +++ b/2024/CARB-Net_ Camera-Assisted Radar-Based Network for Vulnerable Road User Detection/ed9a2da4-1d84-4538-8283-1656bf597855_content_list.json @@ -0,0 +1,1385 @@ +[ + { + "type": "text", + "text": "CARB-Net: Camera-Assisted Radar-Based Network for Vulnerable Road User Detection", + "text_level": 1, + "bbox": [ + 243, + 140, + 759, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wei-Yu Lee $^{\\text{①}}$ , Martin Dimitrievski $^{\\text{②}}$ , David Van Hamme $^{\\text{③}}$ , Jan Aelterman $^{\\text{④}}$ , Ljubomir Jovanov $^{\\text{⑤}}$ , and Wilfried Philips $^{\\text{⑥}}$", + "bbox": [ + 227, + 212, + 774, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "TELIN-IPI, Ghent University-IMEC, Gent, Belgium", + "bbox": [ + 325, + 255, + 676, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Ensuring a reliable perception of vulnerable road users is crucial for safe autonomous driving. Radar stands out as an appealing sensor choice due to its resilience in adverse weather, cost-effectiveness, depth sensing capabilities, and established role in adaptive cruise control. Nevertheless, radar's limited angular resolution poses challenges in object recognition, especially in distinguishing targets in close proximity. To tackle this limitation, we present the Camera-Assisted Radar-Based Network (CARB-Net), a novel and efficient framework that merges the angular accuracy of a camera with the robustness and depth sensing capabilities of radar. We integrate camera detection information through a ground plane feed-forward array, entangling it with the early stages of a radar-based detection network. Furthermore, we introduce a unique context learning approach to ensure graceful degradation in situations of poor radar Doppler information or unfavorable camera viewing conditions. Experimental validations on public and our proposed datasets, along with benchmark comparisons, showcase CARB-Net's superiority, boasting up to a $12\\%$ improvement in mAP performance. A series of ablation studies further emphasize the efficacy of the CARB-Net architecture. Our proposed dataset is released at https://github.com/weiyulee/RadVRU.", + "bbox": [ + 261, + 300, + 754, + 564 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Micro-Doppler Signature $\\cdot$ Vulnerable Road User Detection", + "bbox": [ + 261, + 575, + 740, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 614, + 375, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Robust Vulnerable Road User (VRU) perception is vital for driving safety. Among the myriad of sensors used, millimeter-wave radar stands out due to its resilience in challenging light and weather conditions, such as rain, fog, and snow, and its depth and Doppler velocity capability. Its cost-effectiveness has also contributed to its widespread adoption, making it a fundamental component in many Advanced Driving Assistance Systems (ADAS) applications.", + "bbox": [ + 212, + 643, + 787, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, radar sensors have inherent limitations: Their angular resolution is limited by the size of the antenna array and operating wavelength. Compared to HD cameras, typical radars have two orders of magnitude lower angular resolution [30, 36, 40], making it challenging to distinguish between closely spaced objects. Beyond detection, this low signal resolution makes classifying between object categories even more difficult. In cluttered environments, multipath interference noise further complicates the interpretation of radar data.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7437572b73d31a9953fd02091fb67d4ed3dee7944e9df342f7a1c1c46dbb247c.jpg", + "image_caption": [ + "Fig. 1: An example of micro-Doppler patterns. left: The radar signal is projected and overlaid with the synchronized camera image. right: The radar Doppler signal is plotted over time. The micro-Doppler pattern of the cyclist is highlighted by a dotted line." + ], + "image_footnote": [], + "bbox": [ + 217, + 148, + 782, + 229 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the literature, radar is often used as a means to improve 3D object detection in monocular systems. By fusing imaging information from cameras with radar, researchers have proposed to exploit the strengths of both technologies [13,14,42]. In these prior studies, cameras play a key role in providing rich visual cues that can be used to train effective object detection and classification models. However, in low-light environments, cameras cannot provide the recognition of road users, and detection of targets by only using radar point clouds becomes difficult due to their sparsity. One of the possible alternative cues for VRU detection is using Doppler information from dense radar tensors, alternatively referred to as Radar Frequency (RF) images. This radar data representation provides the unique motion patterns in the time-frequency domain [5,16,17]. For example, in Figure 1, a cyclist exhibits an oscillatory motion of the limbs, which results in a RF signal with sufficient features to classify these road users. These periodically changing motion patterns are commonly referred to as micro-Doppler patterns, providing fundamental features for identifying road users. However, the low azimuth resolution of radar sensors still limits the accuracy of the predicted object positions in these studies.", + "bbox": [ + 212, + 306, + 787, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by these challenges, we propose a RF-based, camera-assisted fusion network with two significant novelties that enhance detection accuracy in nominal conditions (i.e., sensors are working as expected), but more importantly, under conditions of sensor failure. A new cooperative fusion approach introduces summarized camera observations into the early stages of a radar-based network, helping in better feature extraction over the Doppler signal in the RF images. Furthermore, we propose a robust way of entangling camera and radar data, helping the network enhance robustness against scenarios where a sensor fails to detect the targets. By applying these two improvements, we achieve detector robustness without the need for excessive training in unforeseen situations.", + "bbox": [ + 212, + 565, + 787, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We focus on the task of Bird's Eye View (BEV) vulnerable road user perception, where we train a radar-based network on radar frequency images, or more specifically, Range-Azimuth-Doppler (RAD) tensors. The RAD encodes the micro-Doppler pattern of targets at individual range-azimuth spatial cells and is used as the core feature for both detection and classification. Camera information is appended to the RAD tensor in the form of a summarized detection array in the BEV plane, which assists the radar-based network when it is available. We use an off-the-shelf camera-based object detection model to produce detec", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "W.Y Lee et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions of road users and populate the feed-forward array using back-projection. The input layer of the network then applies pixel-shuffle [37] to reorganize the high-resolution camera observations into a common representation, matching the low-resolution radar data. This allows a seamless cooperative fusion with radar tensors without loss of information.", + "bbox": [ + 212, + 146, + 782, + 219 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Robustness against scenarios where a sensor fails to detect the targets is achieved by augmenting the data and expanding the ground truth labels into sub-categories of possible fusion contexts of nominal and deteriorated sensor performance for each object. The novel aspect of training is to generate and encode these conditions as control variables, that augment the ground truths during training. This strategy encourages the model to learn more sophisticated and finer features for each modality without additional human annotation, enhancing its adaptability to diverse fusion contexts. Our experiments, conducted on both public and our proposed RadVRU datasets, demonstrate the effectiveness of our approach. Our method not only achieves state-of-the-art detection performance but also exhibits a remarkable up to $12\\%$ improvement in challenging sensor failure cases. Our contributions are outlined as follows:", + "bbox": [ + 212, + 222, + 784, + 401 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Novel camera-assisted radar-based network based on pixel-shuffling of high-resolution vision feed-forward and low-resolution micro-Doppler signatures.", + "2. Novel context learning strategy, utilizing context-specific control variables to train a model that is robust to conditions where a sensor is unable to detect the targets.", + "3. An extensive experimental evaluation on two datasets for BEV detection where we objectively compare to existing state-of-the-art methods." + ], + "bbox": [ + 220, + 411, + 784, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 537, + 387, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Radar-Based Object Detection Prior studies utilize radar as the primary sensor for object detection, employing various data representations: point clouds and dense radar tensors. For radar point clouds, the sparsity and lack of context pose detection challenges [33, 35, 41] for neural networks to learn features effectively. On the other hand, image-like radar tensors provide rich but noisy contextual data [8, 15]. Many prior studies utilize deep learning to recognize objects by their motion characteristics [3, 9]. For instance, radar's unique micro-Doppler signature, a time-varying Doppler frequency shift captured in RAD tensors, enables accurate recognition of road users like pedestrians and cyclists [5,6,16,17]. This distinctive motion pattern advantage motivates our use of RAD tensors as inputs, harnessing the micro-Doppler signature for robust object detection.", + "bbox": [ + 212, + 566, + 784, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "On the other hand, radar sensors, while effective at detecting moving road users, struggle to separate closely packed objects, such as groups of pedestrians, due to their low angular resolution [41]. Additionally, radar sensors fail to detect stationary objects due to interference from other reflections and a lack of motion information. As a result, the existing radar-only methods cannot accurately localize and recognize all objects. In this study, we utilize observations from the camera as assistance to improve the precision of azimuth localization.", + "bbox": [ + 212, + 734, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Camera-Assisted Radar-Based Network for Vulnerable Road User Detection", + "bbox": [ + 225, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Radar-Camera-Based Object Detection Recently, using BEV representation to merge images from multiple camera and radar sensors has sparked a rising wave of research in autonomous driving. Prior studies adopt large pretrained image networks, such as LSS [32], BEVFormer [20], or BEVDepth [19], to project 2D image features onto the BEV for further fusion with radar data. For instance, in [13], the authors use proposals from a camera-based network to associate and merge with radar point clouds. In [14], a radar-assisted transformation method is proposed to project 2D images onto BEV and aggregate image and radar feature maps for detection. However, the perspective transformation of camera views highly relies on accurate calibration parameters and good light conditions. Numerous objects visible to cameras may yield only a few radar points, posing significant challenges when cameras are unreliable and sparse point clouds cannot provide sufficient evidence for recognition.", + "bbox": [ + 212, + 146, + 787, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast, our method interfaces with the radar at a much earlier stage, utilizing 3D dense radar tensors as input to learn the micro-Doppler patterns for recognition instead of relying on camera features. Therefore, detection can be performed only on radar data in low-light conditions without cameras. Moreover, existing fusion methods often necessitate training a huge network on a large amount of both camera and radar data, which poses challenges due to the lack of publicly available datasets containing both modalities and the high computing effort. To address this, we utilize detections from an efficient off-the-shelf camera-based detection model, reducing the training and inference burden. Since our method's core is a radar-based detector using micro-Doppler patterns to recognize VRUs, only aided by the camera when it is available, we mainly consider the dataset providing 3D radar tensors for experiments, and the existing BEV radar-camera fusion methods cannot be directly applied to this task.", + "bbox": [ + 212, + 343, + 789, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Robustness of Multimodal Fusion In practice, it is quite common for sensors to become unreliable in specific contexts. This negatively affects the robustness of the perception system, such as occlusions or low-light environments for cameras. Prior studies introduce sensor dropout [4,11,38,39] during training to enhance the network's ability to handle these conditions with single modality features. However, simply removing the sensor data from the input constrains the fusion potential and usually degrades the overall detection performance. In this work, we introduce a novel context learning strategy that augments the data by creating challenging scenarios that render sensor observations unreliable, and we instruct the network to recognize these situations through auxiliary queries, enhancing model robustness in detection tasks.", + "bbox": [ + 212, + 560, + 789, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 748, + 380, + 766 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we introduce Camera-Assisted Radar-Based Network (CARB-Net), a novel neural network model for integrating low spatial resolution radar tensors and high-resolution camera images. As depicted in Fig. 2, the proposed framework operates in the spatio-temporal domain, taking sequential camera", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "W.Y Lee et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2af3859c66ae4e3d05f041cd77ccd6e1ec928bb0235bff8cc2a1f1c91c7e6128.jpg", + "image_caption": [ + "Fig.2: Overview of the proposed Camera-Assisted Radar-Based Network (CARB-Net)." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 787, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "images $I_{\\mathrm{c}} \\in \\mathbb{R}^{NC_{c} \\times H_{c} \\times W_{c}}$ and sequential RAD radar tensors $T_{\\mathrm{r}} \\in \\mathbb{R}^{NC_{r} \\times H_{r} \\times W_{r}}$ , where $N$ denotes the number of tensors and images in the sequence, $C_{c}, C_{r}$ represent channel sizes, and $(H_{c}, W_{c})$ and $(H_{r}, W_{r})$ are the spatial dimensions of the RAD tensors and camera images.", + "bbox": [ + 212, + 358, + 787, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unified Data Representation The data representations of radars and cameras are different, which poses challenges in sensor fusion as association requires relating objects in different coordinate systems. Therefore, it is essential to establish a common representation suitable for both fusion and detection. Front View and Bird's Eye View are the two mainstream representations in prior studies. For front view, due to the low resolution in the azimuth angle provided by the radar as well as camera calibration errors, projected radar signals may deviate from the object in the image, which makes the observations of the same object contradictory and hard to be fused by networks. On the other hand, BEV provides a view that accurately portrays the scale and positions of objects, making it a suitable representation for autonomous driving tasks. In this study, we use BEV as the unified data representation.", + "bbox": [ + 212, + 458, + 787, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the image coordinates $(u, v)$ , we rely on the intrinsic camera parameters and extrinsic parameters to perform transformation from camera reference frame into BEV coordinates. We assume the point where the target touches the ground plane (foot or wheel) is with zero height, then the image coordinates $(u, v)$ can be projected to BEV [10, 27]. However, this assumption can be inaccurate in cases when the camera cannot clearly capture the bottom-most point of an object, the target is jumping in the air, or the ground is not flat. This inaccuracy severely degrades the detection performance when using perspective transformation to project the RGB image to BEV for aligning radar and camera signals. Furthermore, perspective transformation also introduces image distortion because of the zero height assumption. The distortion can severely corrupt objects' visual appearance, such as changing the shape and structure, which causes difficulties for the network to learn from the camera images [18].", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Camera-Assisted Radar-Based Network for Vulnerable Road User Detection", + "bbox": [ + 225, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/544b31def2fdbc60b3f505c05c7dc9422526a27d724156d33dcd979d5bc8e971.jpg", + "image_caption": [ + "Fig.3: A camera image with predicted bounding boxes and the cooperative vision feed-forward image in BEV. Blue and green colors represent the pedestrian and bicycle classes. Yellow arrows indicate the mapping after perspective transformation." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cooperative Vision Feed-Forward In order to address the aforementioned problems, we propose Cooperative Vision Feed-Forward to merge radar data uncertain in azimuth but certain in range with RGB data uncertain in range but certain in azimuth. Despite the uncertainties of radar and camera observations, there is usually a small region on the BEV plane where the radar and camera observations exhibit a strong response, which we use for road user detection.", + "bbox": [ + 212, + 351, + 782, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As depicted in Figure 3, we introduce the detections made by using camera data in the form of a vision feed-forward array on the BEV plane. Specifically, we utilize the off-the-shelf camera-based detection network to produce 2D detections of road users based on the RGB images $I_{\\mathrm{c}}$ , then we project the center of the predicted bounding boxes of the road users to the BEV to make vision feedforward images $I_{\\mathrm{f}}$ . The projected values on the BEV are the confidence scores, and these projected centers can be seen as summarized information from the pretrained network using RGB images. By doing this, little to no spatial distortion will be caused by the back-projection.", + "bbox": [ + 212, + 441, + 787, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to the projection based on the zero height assumption, we estimate the depth of a target from its width and height in the image frame and its average size (i.e., averaged human heights and shoulder widths) in the real-world to estimate robust depth values. As a result, we use the median value of these estimations as the depth value. To compensate for the projection errors caused by inaccurate perspective transformation, we enlarge the projected positions from points to regions by applying a Gaussian kernel to cover the uncertainty of the depth estimation. Specifically, we never know the actual heights and widths of pedestrians in practice, but instead we use a statistical model of the mean height and width of humans, $1.2\\mathrm{m}$ to $2.0\\mathrm{m}$ height and $0.36\\mathrm{m}$ to $0.44\\mathrm{m}$ width, respectively. That means the projection ends up on a segment of a perspective line of $x$ meters, which becomes the sigma of our Gaussian kernel. As long as the sigma is not underestimated, the result will not critically depend on its value. Therefore, the inaccuracy problem caused by perspective transformation can be alleviated.", + "bbox": [ + 212, + 579, + 787, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Spatial Resolution Alignment Efficiently aligning low resolution radar data and high resolution camera signals on the ground plane is a challenge. A straight-", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "W.Y Lee et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "forward solution for aligning is to use a BEV space with the same low spatial resolution as the radar sensor [21]. However, the low spatial resolution of the radar causes this approach to discard high-resolution camera information, negatively affecting the fusion outcomes. On the other hand, upscaling low-resolution radar to align with high-resolution cameras is also not practical because of the high computing effort. Therefore, we introduce Pixel-shuffle [37] Downsampling (PD) to tackle this resolution misalignment problem. PD is to reorganize the high-resolution image data into several low resolution channels. This operation only rearranges the order of the elements to down-scale the image instead of encoding or pooling. Hence, all the information of the image can be preserved, and the network's time and space complexity will not be increased. Specifically, we make vision feed-forward images $I_{\\mathrm{f}} \\in \\mathbb{R}^{N \\times H_f \\times W_f}$ , where $(H_f \\times W_f)$ is an intermediate spatial resolution between RGB images and radar cubes. In this instance, we apply PD on the vision feed-forward images $I_{\\mathrm{f}}$ to obtain $I_{\\mathrm{f}}' \\in \\mathbb{R}^{ND^2 \\times H_r \\times W_r}$ , where $D$ is the down-scaling factor, such that the subsequent network may jointly extract features from the radar tensors and vision feed-forward images with varying resolutions using convolution kernels. The solution has two main advantages: (1) no additional network training effort on camera images is needed, and (2) resolution alignment without information loss.", + "bbox": [ + 212, + 146, + 787, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Camera-Assisted Radar-Based Network The aim of the proposed network is to identify road users by using sequential RAD tensors $T_{\\mathrm{r}}$ and vision feedforward images $I_{\\mathrm{f}}$ . The RAD tensors capture micro-Doppler patterns and precise range measurements but have limited azimuth resolution. Our network interprets these tensors to pinpoint and classify various target categories. Conversely, the vision feed-forward images offer detailed azimuth localization of VRUs from a pre-trained camera-based network but lack accuracy in range estimation. Despite the inherent uncertainties, there is a tiny region on the BEV plane where all sensors exhibit a strong response, which we use for road user detection.", + "bbox": [ + 212, + 481, + 787, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our CARB-Net is developed based on the principles of multi-resolution feature decomposition in U-Net [34] with several improvements. First, we use a small convolution network, which we call the Fusion Encoder, to encode the RF information from the radar and vision feed-forward images into a compact channel-representation as the input for U-Net. In addition, we apply the concept of pixel-shuffle [37] again to replace the max-pooling and up-scaling modules inside U-Net. The feature maps are easily up- or down-scaled by reorganizing the image channels without adding additional model parameters. Compared to deconvolution, max-pooling, or bilinear interpolation, pixel-shuffle is used as an effective way to prevent resolution loss during the encoding and decoding steps.", + "bbox": [ + 212, + 622, + 787, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To obtain high-resolution output, we apply ESPCN [37] and pixel-shuffle up-sampling by factor $D$ to reorganize the channels of the network output. Ultimately, the spatial resolution of the output tensor becomes $(H_{f} \\times W_{f})$ . With this architecture, we allow the networks to process input with different resolutions.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Camera-Assisted Radar-Based Network for Vulnerable Road User Detection", + "bbox": [ + 225, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Context Learning Radars or cameras might not be able to provide recognition of road users in specific contexts, which is unacceptable for real-world deployment. For instance, a stationary object is poorly visible to radars, and adverse weather conditions may make objects hard-to-spot by cameras. However, most existing methods do not focus on such cases during training but only emphasize average detection performance, causing significant performance drops under such contexts compared to favorable circumstances.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In order to improve robustness in these specific contexts, we make the network answer 2 additional auxiliary questions to distinguish the sensing capability of each sensor. For each position in the output confidence map, the network has to determine: (1) is there a VRU or not; (2) can a VRU be detected by the radar; and (3) can a VRU be detected by the camera. During training, we apply data augmentation and increase the variability of available labeled data in order to cover more combinations of the possible answers to these questions.", + "bbox": [ + 212, + 256, + 789, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We define $y \\in \\{y_0, y_1, y_2\\}$ to represent the 3 correct answers to the 3 questions. Specifically, $y_0$ indicates whether this position is occupied by a road user. The network should use all the sensor data as input to detect VRUs. $y_1$ describes if the road user's micro-Doppler pattern is detectable by the radar sensor. In this case, we focus on this specific context that makes the targets poorly visible to radar. Since the targets are stationary or moving without micro-Doppler patterns, the network cannot distinguish their radar signals from other obstacles' reflections. Finally, $y_2$ indicates whether the pre-trained camera-based network can detect the target or not. In this case, we focus on the conditions that make the camera sensor unable to see the target. By training the network to answer these 3 questions, we use the network's first answer as the final fusion decision to distinguish if the road user is present, and we use the others to describe the specific fusion contexts for each target.", + "bbox": [ + 212, + 364, + 789, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For instance, $[y_0, y_1, y_2] = [1, 1, 1]$ is the common case that both radar and camera can sense the targets, and $[y_0, y_1, y_2] = [1, 0, 1]$ means the road user is only detectable by the camera. This way, our network learns from the correct answers to distinguish these circumstances and understand the scene in more detail. During training, this approach additionally allows us to consider and penalize each class (e.g., pedestrian, cyclist, etc.) independently. This is achieved by tasking the network to perform this classification as well, thus penalizing solutions that fail to interpret their input data or context. That is, for each class, the network is required to output the 3 answers to the 3 questions, which are independent of the other classes' answers.", + "bbox": [ + 212, + 564, + 787, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In order to teach the network to distinguish these specific contexts, we control the input data to generate the corresponding contexts for training. For radar, the ideal way to simulate targets moving without micro-Doppler patterns or stationary targets is to remove their micro-Doppler patterns and retain only the mean velocities. However, this solution is not practical because the radar signals usually contain reflections from other objects and clutter. Moreover, micro-Doppler information from one object can spill over into nearby cells, depending on scene geometry. It is challenging to only remove a certain target's motion information", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "W.Y Lee et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "while not affecting the others. Instead, we propose an alternative way to simulate such cases: the latest RAD tensor is repeated inside the sequential RAD tensors $T_{\\mathrm{r}}$ to replace the other $N - 1$ radar tensors. That is, all the objects in the scene are preserved at their mean velocities, and most of their micro-Doppler patterns are eliminated. So, the objects' motion inside the repeated tensors is seen as an extension of the current movement without introducing any periodic motion. However, there are several downsides to this solution. First, if the target is moving very fast, the radar can still capture its micro-Doppler patterns in a single radar cube. Secondly, repeating the latest radar cube results in radar data where objects are static in RA plane but have non-zero velocity, a situation that is impossible in reality. However, this data augmentation is a close enough simulation of the most important radar limitations: its weakness in detecting slow motion and people with low micro-Doppler variation amplitude. Furthermore, it is easy to implement, and we do not need any additional computing effort to carefully remove the micro-Doppler patterns. For RGB cameras, we switch on and off the cooperative vision feed-forward to simulate the contexts in which the camera sensor cannot detect the target. In this case, the network should only use radar signals to recognize road users.", + "bbox": [ + 212, + 146, + 787, + 419 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Anchor Points and Offsets In the output layer, we incorporate anchor points and offsets to precisely refine VRU positions at any resolution. This technique increases the achievable resolution without increasing the data or grid dimensions by using an additional channel of output data to encode highly precise spatial offsets to a target from a low-dimensional grid, low-precision position. We predefine our anchor points over a uniformly spaced range and azimuth positions to allow for an equal capacity for detecting objects at all distances. Since there is a low variation in the size of the VRUs in BEV, we refrain from predicting object width and height. Specifically, we utilize range and azimuth bins from the radar configuration to pre-define the anchor points, representing certain positions on the ground plane. The network predicts 2 offsets for two directions relative to the default anchor point. To ensure that each VRU only be matched by a subset of anchor points, we restrict the offsets to being only inside the 4 surrounding anchor points. Finally, we perform non-maximum suppression to merge nearby anchor points using a greedy approach that favors the highest confidence scores within a pre-defined radius. This is a standard methodology employed in image object detection, which we found also works well in BEV detection using radar.", + "bbox": [ + 212, + 440, + 787, + 698 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Loss Functions For a certain position $\\mathbf{r}$ in the output tensor, it contains $(3k + 2)$ different output values: 2 offsets and $3k$ classes (original $k$ different categories from ground truths and 3 answers for context learning), which results in a total of $(3k + 2)H_{f}W_{f}$ outputs. We define $O_{c,y}(\\mathbf{r})$ that refers to the probability of category $c$ with answer $y$ , where $c \\in \\{c_0, c_1, \\dots, c_k\\}$ and $y \\in \\{y_0, y_1, y_2\\}$ . In order to reduce the inherent class imbalance between the road users and background in the sparsely populated training BEV maps, we apply Focal loss [22] to compute the classification loss $\\mathcal{L}_{\\mathrm{cls}}$ :", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Camera-Assisted Radar-Based Network for Vulnerable Road User Detection", + "bbox": [ + 225, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c l s}} (\\mathbf {r}) = - \\sum_ {c} \\sum_ {y} \\alpha_ {c} \\left[ (1 - O _ {c, y} (\\mathbf {r})) ^ {\\gamma} \\log \\left(O _ {c, y} (\\mathbf {r})\\right) + O _ {c, y} (\\mathbf {r}) ^ {\\gamma} (1 - y) \\log \\left(1 - O _ {c, y} (\\mathbf {r})\\right) \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 172, + 799, + 219 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\alpha_{c}$ represents the weighting factor for class $c$ , and $\\gamma \\geq 0$ is the hyperparameter to down-weigh easy samples. We use smooth L1-norm as our localization loss $\\mathcal{L}_{\\mathrm{loc}}$ . This is similar to the bounding box regression in SSD [23] but we remove the multiple boxes and only predict the offsets for every anchor point.", + "bbox": [ + 212, + 220, + 787, + 282 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 215, + 310, + 366, + 325 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Datasets and Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 345, + 570, + 361 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In most public datasets, data representations of radar signals are in the form of a processed point cloud containing targets [2] or projected RD or RA maps [1,28, 31,36,40]. Since our method is based on the micro-Doppler signature to recognize road users, RAD tensors with synchronized camera images are necessary. Hence, we identified the publicly available CARRADA dataset [30] to be best suited for conducting experimental evaluation. In addition, we present results from experimentation using our proposed RadVRU dataset.", + "bbox": [ + 212, + 377, + 787, + 483 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "CARRADA Dataset The CARRADA dataset [30] includes RAD tensors, RD, RA, and AD maps, all synchronized with camera images. Annotations comprise sparse points, dense masks, and bounding boxes for object detection, tracking, and radar semantic segmentation. Mounted on a stationary vehicle, the hardware captures data in a controlled environment. This dataset encompasses 12k annotated frames across three object categories: pedestrians, cyclists, and vehicles. Each RAD tensor has dimensions of $256 \\times 256 \\times 64$ , and the accompanying camera images boast a resolution of $1232 \\times 1028$ .", + "bbox": [ + 212, + 508, + 787, + 631 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Proposed RadVRU Dataset Our RadVRU dataset is captured in several cities in Western Europe. The acquisition hardware consists of a 77 GHz FMCW radar (TI AWR1243), a full HD RGB camera (Realsense D435), and a 3D LiDAR (Ouster OS1-128) mounted on the top of a vehicle. This dataset contains total 78k frames, including 317 sequences of duration from $10\\mathrm{~s}$ to $20\\mathrm{~s}$ . These frames are representative of the challenging driving scenarios, including different lighting conditions, complex traffic with cars, pedestrians, and cyclists. The annotations are performed by using a camera- and a LiDAR-based object detection model to pre-label the instances, and then these labels are manually corrected by human annotators. The pedestrians are annotated as single points on the ground plane. The dimensions of range, azimuth, and Doppler of every RAD tensor are $128 \\times 16 \\times 128$ , and the resolution of the camera images is $1920 \\times 1080$ .", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "W.Y Lee et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/a1430ac3776c1cffe0ce59b732bdcdc728aacf7d543b1fbf08ad3007e1e2bcd1.jpg", + "table_caption": [ + "Table 1: Comparisons with other methods on CARRADA dataset." + ], + "table_footnote": [], + "table_body": "
MethodTraining DataAPmAPmean Dist. Error
Ped.Cycll.Car
TMVA-Net [29]Radar0.440.430.370.370.67
PeakConv [25]Radar0.450.430.380.380.67
ARC [17]Radar0.620.570.500.540.67
RAMP-CNN [7]Radar0.750.640.500.610.64
CARB-Net-baselineRadar0.770.650.530.630.65
U-Net [34]Radar + Camera0.570.480.470.510.70
FCN [26]Radar + Camera0.570.430.470.460.69
Lim et al. [21]Radar + Camera0.610.530.510.530.66
CARB-NetRadar + Camera0.820.750.600.730.63
", + "bbox": [ + 248, + 170, + 750, + 311 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation Details We use the data from CARRADA as an example to explain our implementation details. First, we down-sample the original radar tensor to $128 \\times 128 \\times 64$ to reduce the high computing effort. Similar to [29], we use 5 sequential radar tensors and camera images as input, so the sizes become $[NC_{r}, H_{r}, W_{r}] = [512, 128, 128]$ and $[NC_{c}, H_{c}, W_{c}] = [15, 1232, 1028]$ . The cooperative vision feed-forward is made by a pre-trained \"yolov8m\" [12] model. We filter the detections from YOLOv8 with a confidence threshold of 0.1, and project the detections to the size of vision feed-forward images $[N, H_{f}, W_{f}] = [5, 256, 256]$ . After pixel-shuffle down-sampling by factor $D = 2$ , it becomes $[ND^{2}, H_{r}, W_{r}] = [20, 128, 128]$ . Similarly, after the ESPCN [37], we apply pixel-shuffle up-sampling by factor $D = 2$ to restore the resolution [256, 256]. For the proposed anchor offsets and context learning, we have $(3k + 2) = 11$ predictions. Hence, the size of the final output tensor is [11, 256, 256]. We use 0.8 meter as the radius of our non-maximum suppression algorithm. Please refer to the Supplementary Materials for more implementation details and network architecture.", + "bbox": [ + 212, + 340, + 787, + 568 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Experiment Results", + "text_level": 1, + "bbox": [ + 215, + 590, + 426, + 606 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation Metric Our experiments employ Average Precision (AP) and mean distance error as primary evaluation metrics. We define true positives as detections falling within a 1.5-meter boundary around the ground truth. We avoid multiple matches within the same boundary. Additionally, we measure localization precision by calculating the distance between true positives and ground truths in meters. Averaging these distance errors yields the mean distance error, providing an evaluation of localization accuracy.", + "bbox": [ + 212, + 618, + 785, + 726 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation on CARRADA Dataset We compare the object detection performance of our proposed method and the other state-of-the-art approaches on the CARRADA dataset. Since some of the previous methods output semantic segmentation masks instead of detections, we carefully find the centers of the predicted masks as their localization results. For FCN [26] and U-Net [34] baselines, we use inverse projection mapping to project the RGB images to BEV", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Camera-Assisted Radar-Based Network for Vulnerable Road User Detection", + "bbox": [ + 225, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/216e355a1e9d82e068b161c067667dd87fbca1053ee7b5632d93103c60cb47ec.jpg", + "table_caption": [ + "Table 2: Comparisons with other methods on our proposed RadVRU dataset." + ], + "table_footnote": [], + "table_body": "
MethodTraining DatamAPmean Dist. Error
ARC [17]Radar0.510.64
CARB-Net-baselineRadar0.550.62
U-Net [34]Radar + Camera0.520.87
FCN [26]Radar + Camera0.350.94
Lim et al. [21]Radar + Camera0.670.63
EchoFusion* [24]Radar + Camera0.690.54
CARB-NetRadar + Camera0.810.47
", + "bbox": [ + 274, + 171, + 722, + 276 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "down to a cell resolution of $128 \\times 128$ . Then, we concatenate the projected images and radar tensors along the channel and output the BEV map. For fair comparisons, we ensure the resolutions of the radar data and projected image are the same for all methods. Furthermore, we remove the cooperative vision feed-forward and train our CARB-Net-baseline model only on the radar tensors to compare with the radar-based methods. Although our method is designed for VRUs, we still involve cars for training and testing. We use the cars' center points in BEV space as the ground truth for detection. The vehicles' orientations are not considered during evaluation by all the methods. As observed in Tab. 1, compared to the others, our proposed framework can not only clearly improve the detection performance but also achieve a smaller mean distance error on the RA plane. The results demonstrate that the proposed method successfully integrates the detections from camera images to achieve better accuracy.", + "bbox": [ + 212, + 301, + 787, + 498 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Evaluation on Proposed RadVRU Dataset. Similarly, we conduct the same experiment on our RadVRU dataset. We evaluate our framework against the other four different models. For EchoFusion* [24], we use RA tensors as their input and re-implement their method with reduced network size due to our limited computing hardware. As illustrated in Tab. 2, our proposed approach also outperforms the other methods by a significant margin.", + "bbox": [ + 212, + 518, + 787, + 609 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Visualization In Fig. 4, we present camera images with bounding boxes predicted by YOLOv8 alongside the radar-based model ARC's [17] and our method's detections in BEV. Notably, our detections outperform ARC's, exhibiting fewer false positives and negatives, leading to superior detection performance. Even in low-light conditions, where YOLOv8 fails, our model successfully identifies pedestrians using radar data alone. Additionally, when road users are stationary, ARC fails due to the absence of micro-Doppler patterns, whereas our method utilizes the vision feed-forward to accurately identify pedestrians.", + "bbox": [ + 212, + 628, + 787, + 750 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 771, + 390, + 786 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effects of Proposed Context Learning To further investigate the impact of the proposed context learning, we conduct several experiments to show the robustness of our model when one of the sensors fail. In Tab. 3, the first row model", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "W.Y Lee et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2217ae1c1395838775ae22e640466c5183b987dd86da666175ea46ade1375742.jpg", + "image_caption": [ + "(a) Camera-based model detections" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 506, + 443 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1f2913b846689011fb5338d7e1f1b9ad9e1146e8ed0900a106dcb4f0ee16121b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 141, + 645, + 243 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7554deb7af81caef1e80d93ea79d5a08702a0aedd1ff142e199d193e82dfa73d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 141, + 779, + 243 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/64838ba3caf42518a0a4719efc9fcc9ec90b5d858b691af12269ef60b4d3bd38.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 513, + 244, + 645, + 343 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/545fb2f719a5e9b0f6db7c277e0a8af0da5f2684c4557d61d649791283632275.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 244, + 779, + 343 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a6a8064d4328a4d9d81b3a0b809d883fd6e85641afa8adfd4a24e09893f74165.jpg", + "image_caption": [ + "(b) Radar-based model confidence map" + ], + "image_footnote": [], + "bbox": [ + 513, + 343, + 645, + 443 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/eebc681ce55accea42f421c6085b93dae0e9ec490e160e96becbdb3706b67342.jpg", + "image_caption": [ + "(c) Our confidence map", + "Fig. 4: Visual comparisons of the detection results for three different scenarios in our proposed dataset. Green boxes in (a) represent the true positives detected by camera-based model. Red boxes/circles indicate false positives and negatives, and white boxes denote ground truth locations. Our model (c) exhibits enhanced detection performance compared to ARC [17] (b), integrating radar and camera data effectively." + ], + "image_footnote": [], + "bbox": [ + 651, + 343, + 779, + 443 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/5b28f9658d2e010068a31711f5948ba2f15f767dc958491859ca1672bd04e3be.jpg", + "table_caption": [ + "Table 3: Effects of proposed context learning." + ], + "table_footnote": [], + "table_body": "
MethodTraining DatamAP w/ Radar OnlymAP w/ Camera OnlymAP w/ Radar & Camera
CARB-Net-baselineRadar0.55--
CARB-Net wo/ Context learningRadar + Camera0.100.450.75
CARB-Net w/ Mix trainingRadar + Camera0.400.520.75
CARB-Net w/ Context learningRadar + Camera0.540.650.81
", + "bbox": [ + 217, + 584, + 781, + 655 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "is only retrained on radar data without cooperative vision feed-forward signals. The purpose of this model is to show the upper bound of the detection performance when there is only radar data available. The second row model is retrained without context learning. The third row model is retrained by a straightforward solution to consider sensor failure cases during training. We randomly dropout the radar or camera from the input to mix-train the model, which forces the model to learn how to detect in only one modality.", + "bbox": [ + 212, + 686, + 787, + 792 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The model without context learning shows a significant performance drop when the camera-based model cannot detect the targets. In other words, this model is sensitive to the input and is not able to perform detection when one", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Camera-Assisted Radar-Based Network for Vulnerable Road User Detection", + "bbox": [ + 225, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/f73eaab95bf7810c4ee0fbeae6264c72b05cbbca16ef5755bb7c383e96131a08.jpg", + "table_caption": [ + "Table 4: Effects of proposed components." + ], + "table_footnote": [], + "table_body": "
MethodVision Feed-ForwardImproved U-NetAnchor OffsetsContext LearningmAP
U-Net [34]----0.53
CARB-Net✓---0.65
✓✓--0.70
✓✓✓-0.75
✓✓✓✓0.81
", + "bbox": [ + 269, + 170, + 730, + 266 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "sensor is not working as expected. The third model shows marginally improved results when the one of the sensor is unable to detect, but the detection performance in nominal cases are affected. In contrast, our proposed method shows significant improvements, no matter in nominal or sensor failure cases.", + "bbox": [ + 212, + 292, + 785, + 354 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effects of Proposed Components To clarify the effect of every proposed component, we conduct several experiments to show the performance difference with or without these components. Tab. 4 demonstrates that all the proposed components contribute to performance improvements. The most notable improvement is brought about by the cooperative vision feed-forward. When it is considered, the model achieves a greater level of performance. The projection of the predicted bounding boxes provided by the off-the-shelf camera-based model not only reduces the training effort but also prevents the distortion problem. The results show that when all the components are incorporated, the performance of the model is considerably enhanced.", + "bbox": [ + 212, + 363, + 789, + 513 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 536, + 359, + 553 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce Camera-Assisted Radar-Based Network (CARB-Net), a novel radar-based framework aimed at addressing the low angular resolution of radar signals and enhancing robustness to conditions where a sensor is unable to detect the targets. CARB-Net effectively integrates detections from an off-the-shelf camera-based object detection model with radar data, enabling precise localization by learning object micro-Doppler patterns from radar tensors and incorporating vision feed-forward signals. Our context learning method ensures reliable performance even when one sensor fails. The paper contributes significantly by showcasing improved performance for vulnerable road users compared to state-of-the-art methods. Validation on our larger radar-camera dataset confirms these achievements. Additionally, our method demonstrates up to a $12\\%$ improvement in mAP performance across diverse datasets, with ablation studies highlighting the efficacy of each framework component. Future research directions could explore optimizing fusion mechanisms and incorporating additional sensor modalities like thermal or acoustic sensors for even more accurate object detection and localization. However, as more sensors are integrated, ensuring robustness in multi-sensor object detection remains critical for real-world applications focused on enhancing road user safety.", + "bbox": [ + 212, + 566, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "W.Y Lee et al.", + "bbox": [ + 271, + 114, + 372, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 145, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was funded by EU Horizon 2020 ECSEL JU research and innovation programme under grant agreement 876487 (NextPerception), by EU Horizon KDT JU research and innovation programme under grant agreement 101139769 (Distribuse), and by the Flemish Government (AI Research Program).", + "bbox": [ + 215, + 178, + 785, + 239 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 263, + 321, + 277 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Barnes, D., Gadd, M., Murcutt, P., Newman, P., Posner, I.: The oxford radar robotcar dataset: A radar extension to the oxford robotcar dataset. In: ICRA (2020)", + "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR (2020)", + "3. Chen, S., He, W., Ren, J., Jiang, X.: Attention-based dual-stream vision transformer for radar gait recognition. In: ICASSP. IEEE (2022)", + "4. Chen, X., Ma, H., Wan, J., Li, B., Xia, T.: Multi-view 3d object detection network for autonomous driving. In: CVPR (2017)", + "5. Dimitrievski, M., Shopovska, I., Van Hamme, D., Veelaert, P., Philips, W.: Weakly supervised deep learning method for vulnerable road user detection in fmcw radar. In: IEEE International Conference on Intelligent Transportation Systems (ITSC). IEEE (2020)", + "6. Gao, X., Xing, G., Roy, S., Liu, H.: Experiments with mmwave automotive radar test-bed. In: 2019 53rd Asilomar conference on signals, systems, and computers. pp. 1-6. IEEE (2019)", + "7. Gao, X., Xing, G., Roy, S., Liu, H.: Ramp-cnn: A novel neural network for enhanced automotive radar object recognition. IEEE Sensors Journal (2020)", + "8. Griffiths, H., Cohen, L., Watts, S., Mokole, E., Baker, C., Wicks, M., Blunt, S.: Radar spectrum engineering and management: Technical and regulatory issues. Proceedings of the IEEE (2014)", + "9. Gusland, D., Christiansen, J.M., Torvik, B., Fioranelli, F., Gurbuz, S.Z., Ritchie, M.: Open radar initiative: Large scale dataset for benchmarking of micro-doppler recognition algorithms. In: IEEE Radar Conference (RadarConf). pp. 1-6 (2021)", + "10. Hartley, R., Zisserman, A.: Multiple view geometry in computer vision. Cambridge university press (2003)", + "1. Hwang, J.J., Kretzschmar, H., Manela, J., Rafferty, S., Armstrong-Crews, N., Chen, T., Anguelov, D.: Cramnet: Camera-radar fusion with ray-constrained cross-attention for robust 3d object detection. In: ECCV (2022)", + "2. Jocher, G., Chaurasia, A., Qiu, J.: YOLO by Ultralytics (2023), https://github.com/ultralytics/ultralytics", + "3. Kim, Y., Kim, S., Choi, J.W., Kum, D.: Craft: Camera-radar 3d object detection with spatio-contextual fusion transformer. In: AAAI (2023)", + "4. Kim, Y., Shin, J., Kim, S., Lee, I.J., Choi, J.W., Kum, D.: Crn: Camera radar net for accurate, robust, efficient 3d perception. In: ICCV (2023)", + "5. Kopp, J., Kellner, D., Piroli, A., Dietmayer, K.: Tackling clutter in radar data-label generation and detection using pointnet++. arXiv preprint arXiv:2303.09530 (2023)" + ], + "bbox": [ + 225, + 295, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Camera-Assisted Radar-Based Network for Vulnerable Road User Detection", + "bbox": [ + 225, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Lee, W.Y., Dimitrievski, M., Jovanov, L., Philips, W.: Spatio-temporal consistency for semi-supervised learning using 3d radar cubes. In: IEEE Intelligent Vehicles Symposium (IV). IEEE (2021)", + "17. Lee, W.Y., Jovanov, L., Kumcu, A., Philips, W.: Arc: Automotive radar consistency regularization for semi-supervised learning. IEEE Transactions on Intelligent Vehicles (2023)", + "18. Lee, W.Y., Jovanov, L., Philips, W.: Multi-view target transformation for pedestrian detection. In: WACV Workshops (2023)", + "19. Li, Y., Ge, Z., Yu, G., Wang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: AAAI (2023)", + "20. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: ECCV (2022)", + "21. Lim, T.Y., Ansari, A., Major, B., Fontijne, D., Hamilton, M., Gowaikar, R., Subramanian, S.: Radar and camera early fusion for vehicle detection in advanced driver assistance systems. In: NIPS Workshops (2019)", + "22. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: ICCV (2017)", + "23. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.Y., Berg, A.C.: Ssd: Single shot multibox detector. In: ECCV (2016)", + "24. Liu, Y., Wang, F., Wang, N., ZHANG, Z.X.: Echoes beyond points: Unleashing the power of raw radar data in multi-modality fusion. NeurIPS (2024)", + "25. Liwen, Z., Xinyan, Z., Youcheng, Z., Yufei, G., Yuanpei, C., Xuhui, H., Zhe, M.: Peakconv: Learning peak receptive field for radar semantic segmentation. In: CVPR (2023)", + "26. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: CVPR (2015)", + "27. Mallot, H.A., Bulthoff, H.H., Little, J.J., Bohrer, S.: Inverse perspective mapping simplifies optical flow computation and obstacle detection. Biological cybernetics (1991)", + "28. Meyer, M., Kuschk, G.: Automotive radar dataset for deep learning based 3d object detection. In: European Radar Conference (EuRAD) (2019)", + "29. Ouaknine, A., Newson, A., Pérez, P., Tupin, F., Rebut, J.: Multi-view radar semantic segmentation. In: ICCV (2021)", + "30. Ouaknine, A., Newson, A., Rebut, J., Tupin, F., Pérez, P.: Carrada dataset: Camera and automotive radar with range- angle- doppler annotations. In: ICPR (2021)", + "31. Paek, D.H., Kong, S.H., Wijaya, K.T.: K-radar: 4d radar object detection for autonomous driving in various weather conditions. NeurIPS (2022)", + "32. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 194-210. Springer (2020)", + "33. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NIPS (2017)", + "34. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI. Springer (2015)", + "35. Schumann, O., Hahn, M., Dickmann, J., Wöhler, C.: Semantic segmentation on radar point clouds. In: International Conference on Information Fusion (FUSION) (2018)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "W.Y Lee et al.", + "bbox": [ + 271, + 114, + 370, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "36. Sheeny, M., De Pellegrin, E., Mukherjee, S., Ahrabian, A., Wang, S., Wallace, A.: Radiate: A radar dataset for automotive perception in bad weather. In: ICRA (2021)", + "37. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: CVPR (2016)", + "38. Vaizman, Y., Weibel, N., Lanckriet, G.: Context recognition in-the-wild: Unified model for multi-modal sensors and multi-label classification. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies (2018)", + "39. Wang, W., Tran, D., Feiszli, M.: What makes training multi-modal classification networks hard? In: CVPR (2020)", + "40. Wang, Y., Wang, G., Hsu, H.M., Liu, H., Hwang, J.N.: Rethinking of radar's role: A camera-radar dataset and systematic annotator via coordinate alignment. In: CVPR (2021)", + "41. Yao, S., Guan, R., Huang, X., Li, Z., Sha, X., Yue, Y., Lim, E.G., Seo, H., Man, K.L., Zhu, X., Yue, Y.: Radar-camera fusion for object detection and semantic segmentation in autonomous driving: A comprehensive review. IEEE Transactions on Intelligent Vehicles (2023)", + "42. Zhou, T., Chen, J., Shi, Y., Jiang, K., Yang, M., Yang, D.: Bridging the view disparity between radar and camera features for multi-modal fusion 3d object detection. IEEE Transactions on Intelligent Vehicles (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 438 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Camera-Assisted Radar-Based Network for Vulnerable Road User Detection", + "bbox": [ + 225, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/CARB-Net_ Camera-Assisted Radar-Based Network for Vulnerable Road User Detection/ed9a2da4-1d84-4538-8283-1656bf597855_model.json b/2024/CARB-Net_ Camera-Assisted Radar-Based Network for Vulnerable Road User Detection/ed9a2da4-1d84-4538-8283-1656bf597855_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0d387b6085804ad05fd97dafda355aac13046de1 --- /dev/null +++ b/2024/CARB-Net_ Camera-Assisted Radar-Based Network for Vulnerable Road User Detection/ed9a2da4-1d84-4538-8283-1656bf597855_model.json @@ -0,0 +1,1917 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.245, + 0.141, + 0.761, + 0.186 + ], + "angle": 0, + "content": "CARB-Net: Camera-Assisted Radar-Based Network for Vulnerable Road User Detection" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.213, + 0.776, + 0.244 + ], + "angle": 0, + "content": "Wei-Yu Lee\\(^{\\text{①}}\\), Martin Dimitrievski\\(^{\\text{②}}\\), David Van Hamme\\(^{\\text{③}}\\), Jan Aelterman\\(^{\\text{④}}\\), Ljubomir Jovanov\\(^{\\text{⑤}}\\), and Wilfried Philips\\(^{\\text{⑥}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.326, + 0.256, + 0.677, + 0.27 + ], + "angle": 0, + "content": "TELIN-IPI, Ghent University-IMEC, Gent, Belgium" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.301, + 0.756, + 0.565 + ], + "angle": 0, + "content": "Abstract. Ensuring a reliable perception of vulnerable road users is crucial for safe autonomous driving. Radar stands out as an appealing sensor choice due to its resilience in adverse weather, cost-effectiveness, depth sensing capabilities, and established role in adaptive cruise control. Nevertheless, radar's limited angular resolution poses challenges in object recognition, especially in distinguishing targets in close proximity. To tackle this limitation, we present the Camera-Assisted Radar-Based Network (CARB-Net), a novel and efficient framework that merges the angular accuracy of a camera with the robustness and depth sensing capabilities of radar. We integrate camera detection information through a ground plane feed-forward array, entangling it with the early stages of a radar-based detection network. Furthermore, we introduce a unique context learning approach to ensure graceful degradation in situations of poor radar Doppler information or unfavorable camera viewing conditions. Experimental validations on public and our proposed datasets, along with benchmark comparisons, showcase CARB-Net's superiority, boasting up to a \\(12\\%\\) improvement in mAP performance. A series of ablation studies further emphasize the efficacy of the CARB-Net architecture. Our proposed dataset is released at https://github.com/weiyulee/RadVRU." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.577, + 0.741, + 0.592 + ], + "angle": 0, + "content": "Keywords: Micro-Doppler Signature \\(\\cdot\\) Vulnerable Road User Detection" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.615, + 0.377, + 0.631 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Robust Vulnerable Road User (VRU) perception is vital for driving safety. Among the myriad of sensors used, millimeter-wave radar stands out due to its resilience in challenging light and weather conditions, such as rain, fog, and snow, and its depth and Doppler velocity capability. Its cost-effectiveness has also contributed to its widespread adoption, making it a fundamental component in many Advanced Driving Assistance Systems (ADAS) applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "However, radar sensors have inherent limitations: Their angular resolution is limited by the size of the antenna array and operating wavelength. Compared to HD cameras, typical radars have two orders of magnitude lower angular resolution [30, 36, 40], making it challenging to distinguish between closely spaced objects. Beyond detection, this low signal resolution makes classifying between object categories even more difficult. In cluttered environments, multipath interference noise further complicates the interpretation of radar data." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.128 + ], + "angle": 0, + "content": "W.Y Lee et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.149, + 0.784, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.242, + 0.788, + 0.285 + ], + "angle": 0, + "content": "Fig. 1: An example of micro-Doppler patterns. left: The radar signal is projected and overlaid with the synchronized camera image. right: The radar Doppler signal is plotted over time. The micro-Doppler pattern of the cyclist is highlighted by a dotted line." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.789, + 0.563 + ], + "angle": 0, + "content": "In the literature, radar is often used as a means to improve 3D object detection in monocular systems. By fusing imaging information from cameras with radar, researchers have proposed to exploit the strengths of both technologies [13,14,42]. In these prior studies, cameras play a key role in providing rich visual cues that can be used to train effective object detection and classification models. However, in low-light environments, cameras cannot provide the recognition of road users, and detection of targets by only using radar point clouds becomes difficult due to their sparsity. One of the possible alternative cues for VRU detection is using Doppler information from dense radar tensors, alternatively referred to as Radar Frequency (RF) images. This radar data representation provides the unique motion patterns in the time-frequency domain [5,16,17]. For example, in Figure 1, a cyclist exhibits an oscillatory motion of the limbs, which results in a RF signal with sufficient features to classify these road users. These periodically changing motion patterns are commonly referred to as micro-Doppler patterns, providing fundamental features for identifying road users. However, the low azimuth resolution of radar sensors still limits the accuracy of the predicted object positions in these studies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.789, + 0.717 + ], + "angle": 0, + "content": "Motivated by these challenges, we propose a RF-based, camera-assisted fusion network with two significant novelties that enhance detection accuracy in nominal conditions (i.e., sensors are working as expected), but more importantly, under conditions of sensor failure. A new cooperative fusion approach introduces summarized camera observations into the early stages of a radar-based network, helping in better feature extraction over the Doppler signal in the RF images. Furthermore, we propose a robust way of entangling camera and radar data, helping the network enhance robustness against scenarios where a sensor fails to detect the targets. By applying these two improvements, we achieve detector robustness without the need for excessive training in unforeseen situations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.84 + ], + "angle": 0, + "content": "We focus on the task of Bird's Eye View (BEV) vulnerable road user perception, where we train a radar-based network on radar frequency images, or more specifically, Range-Azimuth-Doppler (RAD) tensors. The RAD encodes the micro-Doppler pattern of targets at individual range-azimuth spatial cells and is used as the core feature for both detection and classification. Camera information is appended to the RAD tensor in the form of a summarized detection array in the BEV plane, which assists the radar-based network when it is available. We use an off-the-shelf camera-based object detection model to produce detec" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.731, + 0.128 + ], + "angle": 0, + "content": "Camera-Assisted Radar-Based Network for Vulnerable Road User Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.22 + ], + "angle": 0, + "content": "tions of road users and populate the feed-forward array using back-projection. The input layer of the network then applies pixel-shuffle [37] to reorganize the high-resolution camera observations into a common representation, matching the low-resolution radar data. This allows a seamless cooperative fusion with radar tensors without loss of information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.402 + ], + "angle": 0, + "content": "Robustness against scenarios where a sensor fails to detect the targets is achieved by augmenting the data and expanding the ground truth labels into sub-categories of possible fusion contexts of nominal and deteriorated sensor performance for each object. The novel aspect of training is to generate and encode these conditions as control variables, that augment the ground truths during training. This strategy encourages the model to learn more sophisticated and finer features for each modality without additional human annotation, enhancing its adaptability to diverse fusion contexts. Our experiments, conducted on both public and our proposed RadVRU datasets, demonstrate the effectiveness of our approach. Our method not only achieves state-of-the-art detection performance but also exhibits a remarkable up to \\(12\\%\\) improvement in challenging sensor failure cases. Our contributions are outlined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.412, + 0.784, + 0.441 + ], + "angle": 0, + "content": "1. Novel camera-assisted radar-based network based on pixel-shuffling of high-resolution vision feed-forward and low-resolution micro-Doppler signatures." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.442, + 0.785, + 0.486 + ], + "angle": 0, + "content": "2. Novel context learning strategy, utilizing context-specific control variables to train a model that is robust to conditions where a sensor is unable to detect the targets." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.486, + 0.785, + 0.516 + ], + "angle": 0, + "content": "3. An extensive experimental evaluation on two datasets for BEV detection where we objectively compare to existing state-of-the-art methods." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.412, + 0.785, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.538, + 0.388, + 0.554 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.785, + 0.734 + ], + "angle": 0, + "content": "Radar-Based Object Detection Prior studies utilize radar as the primary sensor for object detection, employing various data representations: point clouds and dense radar tensors. For radar point clouds, the sparsity and lack of context pose detection challenges [33, 35, 41] for neural networks to learn features effectively. On the other hand, image-like radar tensors provide rich but noisy contextual data [8, 15]. Many prior studies utilize deep learning to recognize objects by their motion characteristics [3, 9]. For instance, radar's unique micro-Doppler signature, a time-varying Doppler frequency shift captured in RAD tensors, enables accurate recognition of road users like pedestrians and cyclists [5,6,16,17]. This distinctive motion pattern advantage motivates our use of RAD tensors as inputs, harnessing the micro-Doppler signature for robust object detection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.785, + 0.84 + ], + "angle": 0, + "content": "On the other hand, radar sensors, while effective at detecting moving road users, struggle to separate closely packed objects, such as groups of pedestrians, due to their low angular resolution [41]. Additionally, radar sensors fail to detect stationary objects due to interference from other reflections and a lack of motion information. As a result, the existing radar-only methods cannot accurately localize and recognize all objects. In this study, we utilize observations from the camera as assistance to improve the precision of azimuth localization." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.128 + ], + "angle": 0, + "content": "W.Y Lee et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "Radar-Camera-Based Object Detection Recently, using BEV representation to merge images from multiple camera and radar sensors has sparked a rising wave of research in autonomous driving. Prior studies adopt large pretrained image networks, such as LSS [32], BEVFormer [20], or BEVDepth [19], to project 2D image features onto the BEV for further fusion with radar data. For instance, in [13], the authors use proposals from a camera-based network to associate and merge with radar point clouds. In [14], a radar-assisted transformation method is proposed to project 2D images onto BEV and aggregate image and radar feature maps for detection. However, the perspective transformation of camera views highly relies on accurate calibration parameters and good light conditions. Numerous objects visible to cameras may yield only a few radar points, posing significant challenges when cameras are unreliable and sparse point clouds cannot provide sufficient evidence for recognition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.79, + 0.542 + ], + "angle": 0, + "content": "In contrast, our method interfaces with the radar at a much earlier stage, utilizing 3D dense radar tensors as input to learn the micro-Doppler patterns for recognition instead of relying on camera features. Therefore, detection can be performed only on radar data in low-light conditions without cameras. Moreover, existing fusion methods often necessitate training a huge network on a large amount of both camera and radar data, which poses challenges due to the lack of publicly available datasets containing both modalities and the high computing effort. To address this, we utilize detections from an efficient off-the-shelf camera-based detection model, reducing the training and inference burden. Since our method's core is a radar-based detector using micro-Doppler patterns to recognize VRUs, only aided by the camera when it is available, we mainly consider the dataset providing 3D radar tensors for experiments, and the existing BEV radar-camera fusion methods cannot be directly applied to this task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.561, + 0.79, + 0.728 + ], + "angle": 0, + "content": "Robustness of Multimodal Fusion In practice, it is quite common for sensors to become unreliable in specific contexts. This negatively affects the robustness of the perception system, such as occlusions or low-light environments for cameras. Prior studies introduce sensor dropout [4,11,38,39] during training to enhance the network's ability to handle these conditions with single modality features. However, simply removing the sensor data from the input constrains the fusion potential and usually degrades the overall detection performance. In this work, we introduce a novel context learning strategy that augments the data by creating challenging scenarios that render sensor observations unreliable, and we instruct the network to recognize these situations through auxiliary queries, enhancing model robustness in detection tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.749, + 0.381, + 0.767 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In this paper, we introduce Camera-Assisted Radar-Based Network (CARB-Net), a novel neural network model for integrating low spatial resolution radar tensors and high-resolution camera images. As depicted in Fig. 2, the proposed framework operates in the spatio-temporal domain, taking sequential camera" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.731, + 0.128 + ], + "angle": 0, + "content": "Camera-Assisted Radar-Based Network for Vulnerable Road User Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.144, + 0.788, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.295, + 0.785, + 0.325 + ], + "angle": 0, + "content": "Fig.2: Overview of the proposed Camera-Assisted Radar-Based Network (CARB-Net)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.422 + ], + "angle": 0, + "content": "images \\( I_{\\mathrm{c}} \\in \\mathbb{R}^{NC_{c} \\times H_{c} \\times W_{c}} \\) and sequential RAD radar tensors \\( T_{\\mathrm{r}} \\in \\mathbb{R}^{NC_{r} \\times H_{r} \\times W_{r}} \\), where \\( N \\) denotes the number of tensors and images in the sequence, \\( C_{c}, C_{r} \\) represent channel sizes, and \\( (H_{c}, W_{c}) \\) and \\( (H_{r}, W_{r}) \\) are the spatial dimensions of the RAD tensors and camera images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.642 + ], + "angle": 0, + "content": "Unified Data Representation The data representations of radars and cameras are different, which poses challenges in sensor fusion as association requires relating objects in different coordinate systems. Therefore, it is essential to establish a common representation suitable for both fusion and detection. Front View and Bird's Eye View are the two mainstream representations in prior studies. For front view, due to the low resolution in the azimuth angle provided by the radar as well as camera calibration errors, projected radar signals may deviate from the object in the image, which makes the observations of the same object contradictory and hard to be fused by networks. On the other hand, BEV provides a view that accurately portrays the scale and positions of objects, making it a suitable representation for autonomous driving tasks. In this study, we use BEV as the unified data representation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Given the image coordinates \\((u, v)\\), we rely on the intrinsic camera parameters and extrinsic parameters to perform transformation from camera reference frame into BEV coordinates. We assume the point where the target touches the ground plane (foot or wheel) is with zero height, then the image coordinates \\((u, v)\\) can be projected to BEV [10, 27]. However, this assumption can be inaccurate in cases when the camera cannot clearly capture the bottom-most point of an object, the target is jumping in the air, or the ground is not flat. This inaccuracy severely degrades the detection performance when using perspective transformation to project the RGB image to BEV for aligning radar and camera signals. Furthermore, perspective transformation also introduces image distortion because of the zero height assumption. The distortion can severely corrupt objects' visual appearance, such as changing the shape and structure, which causes difficulties for the network to learn from the camera images [18]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.128 + ], + "angle": 0, + "content": "W.Y Lee et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.282, + 0.788, + 0.325 + ], + "angle": 0, + "content": "Fig.3: A camera image with predicted bounding boxes and the cooperative vision feed-forward image in BEV. Blue and green colors represent the pedestrian and bicycle classes. Yellow arrows indicate the mapping after perspective transformation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.352, + 0.784, + 0.442 + ], + "angle": 0, + "content": "Cooperative Vision Feed-Forward In order to address the aforementioned problems, we propose Cooperative Vision Feed-Forward to merge radar data uncertain in azimuth but certain in range with RGB data uncertain in range but certain in azimuth. Despite the uncertainties of radar and camera observations, there is usually a small region on the BEV plane where the radar and camera observations exhibit a strong response, which we use for road user detection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.789, + 0.579 + ], + "angle": 0, + "content": "As depicted in Figure 3, we introduce the detections made by using camera data in the form of a vision feed-forward array on the BEV plane. Specifically, we utilize the off-the-shelf camera-based detection network to produce 2D detections of road users based on the RGB images \\( I_{\\mathrm{c}} \\), then we project the center of the predicted bounding boxes of the road users to the BEV to make vision feedforward images \\( I_{\\mathrm{f}} \\). The projected values on the BEV are the confidence scores, and these projected centers can be seen as summarized information from the pretrained network using RGB images. By doing this, little to no spatial distortion will be caused by the back-projection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.789, + 0.791 + ], + "angle": 0, + "content": "In addition to the projection based on the zero height assumption, we estimate the depth of a target from its width and height in the image frame and its average size (i.e., averaged human heights and shoulder widths) in the real-world to estimate robust depth values. As a result, we use the median value of these estimations as the depth value. To compensate for the projection errors caused by inaccurate perspective transformation, we enlarge the projected positions from points to regions by applying a Gaussian kernel to cover the uncertainty of the depth estimation. Specifically, we never know the actual heights and widths of pedestrians in practice, but instead we use a statistical model of the mean height and width of humans, \\(1.2\\mathrm{m}\\) to \\(2.0\\mathrm{m}\\) height and \\(0.36\\mathrm{m}\\) to \\(0.44\\mathrm{m}\\) width, respectively. That means the projection ends up on a segment of a perspective line of \\(x\\) meters, which becomes the sigma of our Gaussian kernel. As long as the sigma is not underestimated, the result will not critically depend on its value. Therefore, the inaccuracy problem caused by perspective transformation can be alleviated." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Spatial Resolution Alignment Efficiently aligning low resolution radar data and high resolution camera signals on the ground plane is a challenge. A straight-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.731, + 0.128 + ], + "angle": 0, + "content": "Camera-Assisted Radar-Based Network for Vulnerable Road User Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.435 + ], + "angle": 0, + "content": "forward solution for aligning is to use a BEV space with the same low spatial resolution as the radar sensor [21]. However, the low spatial resolution of the radar causes this approach to discard high-resolution camera information, negatively affecting the fusion outcomes. On the other hand, upscaling low-resolution radar to align with high-resolution cameras is also not practical because of the high computing effort. Therefore, we introduce Pixel-shuffle [37] Downsampling (PD) to tackle this resolution misalignment problem. PD is to reorganize the high-resolution image data into several low resolution channels. This operation only rearranges the order of the elements to down-scale the image instead of encoding or pooling. Hence, all the information of the image can be preserved, and the network's time and space complexity will not be increased. Specifically, we make vision feed-forward images \\( I_{\\mathrm{f}} \\in \\mathbb{R}^{N \\times H_f \\times W_f} \\), where \\( (H_f \\times W_f) \\) is an intermediate spatial resolution between RGB images and radar cubes. In this instance, we apply PD on the vision feed-forward images \\( I_{\\mathrm{f}} \\) to obtain \\( I_{\\mathrm{f}}' \\in \\mathbb{R}^{ND^2 \\times H_r \\times W_r} \\), where \\( D \\) is the down-scaling factor, such that the subsequent network may jointly extract features from the radar tensors and vision feed-forward images with varying resolutions using convolution kernels. The solution has two main advantages: (1) no additional network training effort on camera images is needed, and (2) resolution alignment without information loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.788, + 0.619 + ], + "angle": 0, + "content": "Camera-Assisted Radar-Based Network The aim of the proposed network is to identify road users by using sequential RAD tensors \\( T_{\\mathrm{r}} \\) and vision feedforward images \\( I_{\\mathrm{f}} \\). The RAD tensors capture micro-Doppler patterns and precise range measurements but have limited azimuth resolution. Our network interprets these tensors to pinpoint and classify various target categories. Conversely, the vision feed-forward images offer detailed azimuth localization of VRUs from a pre-trained camera-based network but lack accuracy in range estimation. Despite the inherent uncertainties, there is a tiny region on the BEV plane where all sensors exhibit a strong response, which we use for road user detection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.623, + 0.788, + 0.777 + ], + "angle": 0, + "content": "Our CARB-Net is developed based on the principles of multi-resolution feature decomposition in U-Net [34] with several improvements. First, we use a small convolution network, which we call the Fusion Encoder, to encode the RF information from the radar and vision feed-forward images into a compact channel-representation as the input for U-Net. In addition, we apply the concept of pixel-shuffle [37] again to replace the max-pooling and up-scaling modules inside U-Net. The feature maps are easily up- or down-scaled by reorganizing the image channels without adding additional model parameters. Compared to deconvolution, max-pooling, or bilinear interpolation, pixel-shuffle is used as an effective way to prevent resolution loss during the encoding and decoding steps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To obtain high-resolution output, we apply ESPCN [37] and pixel-shuffle up-sampling by factor \\(D\\) to reorganize the channels of the network output. Ultimately, the spatial resolution of the output tensor becomes \\((H_{f} \\times W_{f})\\). With this architecture, we allow the networks to process input with different resolutions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.128 + ], + "angle": 0, + "content": "W.Y Lee et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "Context Learning Radars or cameras might not be able to provide recognition of road users in specific contexts, which is unacceptable for real-world deployment. For instance, a stationary object is poorly visible to radars, and adverse weather conditions may make objects hard-to-spot by cameras. However, most existing methods do not focus on such cases during training but only emphasize average detection performance, causing significant performance drops under such contexts compared to favorable circumstances." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.257, + 0.79, + 0.363 + ], + "angle": 0, + "content": "In order to improve robustness in these specific contexts, we make the network answer 2 additional auxiliary questions to distinguish the sensing capability of each sensor. For each position in the output confidence map, the network has to determine: (1) is there a VRU or not; (2) can a VRU be detected by the radar; and (3) can a VRU be detected by the camera. During training, we apply data augmentation and increase the variability of available labeled data in order to cover more combinations of the possible answers to these questions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.366, + 0.79, + 0.562 + ], + "angle": 0, + "content": "We define \\( y \\in \\{y_0, y_1, y_2\\} \\) to represent the 3 correct answers to the 3 questions. Specifically, \\( y_0 \\) indicates whether this position is occupied by a road user. The network should use all the sensor data as input to detect VRUs. \\( y_1 \\) describes if the road user's micro-Doppler pattern is detectable by the radar sensor. In this case, we focus on this specific context that makes the targets poorly visible to radar. Since the targets are stationary or moving without micro-Doppler patterns, the network cannot distinguish their radar signals from other obstacles' reflections. Finally, \\( y_2 \\) indicates whether the pre-trained camera-based network can detect the target or not. In this case, we focus on the conditions that make the camera sensor unable to see the target. By training the network to answer these 3 questions, we use the network's first answer as the final fusion decision to distinguish if the road user is present, and we use the others to describe the specific fusion contexts for each target." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.788, + 0.716 + ], + "angle": 0, + "content": "For instance, \\([y_0, y_1, y_2] = [1, 1, 1]\\) is the common case that both radar and camera can sense the targets, and \\([y_0, y_1, y_2] = [1, 0, 1]\\) means the road user is only detectable by the camera. This way, our network learns from the correct answers to distinguish these circumstances and understand the scene in more detail. During training, this approach additionally allows us to consider and penalize each class (e.g., pedestrian, cyclist, etc.) independently. This is achieved by tasking the network to perform this classification as well, thus penalizing solutions that fail to interpret their input data or context. That is, for each class, the network is required to output the 3 answers to the 3 questions, which are independent of the other classes' answers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In order to teach the network to distinguish these specific contexts, we control the input data to generate the corresponding contexts for training. For radar, the ideal way to simulate targets moving without micro-Doppler patterns or stationary targets is to remove their micro-Doppler patterns and retain only the mean velocities. However, this solution is not practical because the radar signals usually contain reflections from other objects and clutter. Moreover, micro-Doppler information from one object can spill over into nearby cells, depending on scene geometry. It is challenging to only remove a certain target's motion information" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.731, + 0.128 + ], + "angle": 0, + "content": "Camera-Assisted Radar-Based Network for Vulnerable Road User Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.42 + ], + "angle": 0, + "content": "while not affecting the others. Instead, we propose an alternative way to simulate such cases: the latest RAD tensor is repeated inside the sequential RAD tensors \\( T_{\\mathrm{r}} \\) to replace the other \\( N - 1 \\) radar tensors. That is, all the objects in the scene are preserved at their mean velocities, and most of their micro-Doppler patterns are eliminated. So, the objects' motion inside the repeated tensors is seen as an extension of the current movement without introducing any periodic motion. However, there are several downsides to this solution. First, if the target is moving very fast, the radar can still capture its micro-Doppler patterns in a single radar cube. Secondly, repeating the latest radar cube results in radar data where objects are static in RA plane but have non-zero velocity, a situation that is impossible in reality. However, this data augmentation is a close enough simulation of the most important radar limitations: its weakness in detecting slow motion and people with low micro-Doppler variation amplitude. Furthermore, it is easy to implement, and we do not need any additional computing effort to carefully remove the micro-Doppler patterns. For RGB cameras, we switch on and off the cooperative vision feed-forward to simulate the contexts in which the camera sensor cannot detect the target. In this case, the network should only use radar signals to recognize road users." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.788, + 0.699 + ], + "angle": 0, + "content": "Anchor Points and Offsets In the output layer, we incorporate anchor points and offsets to precisely refine VRU positions at any resolution. This technique increases the achievable resolution without increasing the data or grid dimensions by using an additional channel of output data to encode highly precise spatial offsets to a target from a low-dimensional grid, low-precision position. We predefine our anchor points over a uniformly spaced range and azimuth positions to allow for an equal capacity for detecting objects at all distances. Since there is a low variation in the size of the VRUs in BEV, we refrain from predicting object width and height. Specifically, we utilize range and azimuth bins from the radar configuration to pre-define the anchor points, representing certain positions on the ground plane. The network predicts 2 offsets for two directions relative to the default anchor point. To ensure that each VRU only be matched by a subset of anchor points, we restrict the offsets to being only inside the 4 surrounding anchor points. Finally, we perform non-maximum suppression to merge nearby anchor points using a greedy approach that favors the highest confidence scores within a pre-defined radius. This is a standard methodology employed in image object detection, which we found also works well in BEV detection using radar." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Loss Functions For a certain position \\(\\mathbf{r}\\) in the output tensor, it contains \\((3k + 2)\\) different output values: 2 offsets and \\(3k\\) classes (original \\(k\\) different categories from ground truths and 3 answers for context learning), which results in a total of \\((3k + 2)H_{f}W_{f}\\) outputs. We define \\(O_{c,y}(\\mathbf{r})\\) that refers to the probability of category \\(c\\) with answer \\(y\\), where \\(c \\in \\{c_0, c_1, \\dots, c_k\\}\\) and \\(y \\in \\{y_0, y_1, y_2\\}\\). In order to reduce the inherent class imbalance between the road users and background in the sparsely populated training BEV maps, we apply Focal loss [22] to compute the classification loss \\(\\mathcal{L}_{\\mathrm{cls}}\\):" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.128 + ], + "angle": 0, + "content": "W.Y Lee et al." + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.174, + 0.8, + 0.22 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c l s}} (\\mathbf {r}) = - \\sum_ {c} \\sum_ {y} \\alpha_ {c} \\left[ (1 - O _ {c, y} (\\mathbf {r})) ^ {\\gamma} \\log \\left(O _ {c, y} (\\mathbf {r})\\right) + O _ {c, y} (\\mathbf {r}) ^ {\\gamma} (1 - y) \\log \\left(1 - O _ {c, y} (\\mathbf {r})\\right) \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.222, + 0.788, + 0.283 + ], + "angle": 0, + "content": "where \\(\\alpha_{c}\\) represents the weighting factor for class \\(c\\), and \\(\\gamma \\geq 0\\) is the hyperparameter to down-weigh easy samples. We use smooth L1-norm as our localization loss \\(\\mathcal{L}_{\\mathrm{loc}}\\). This is similar to the bounding box regression in SSD [23] but we remove the multiple boxes and only predict the offsets for every anchor point." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.311, + 0.367, + 0.327 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.347, + 0.571, + 0.362 + ], + "angle": 0, + "content": "4.1 Datasets and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.378, + 0.788, + 0.484 + ], + "angle": 0, + "content": "In most public datasets, data representations of radar signals are in the form of a processed point cloud containing targets [2] or projected RD or RA maps [1,28, 31,36,40]. Since our method is based on the micro-Doppler signature to recognize road users, RAD tensors with synchronized camera images are necessary. Hence, we identified the publicly available CARRADA dataset [30] to be best suited for conducting experimental evaluation. In addition, we present results from experimentation using our proposed RadVRU dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.632 + ], + "angle": 0, + "content": "CARRADA Dataset The CARRADA dataset [30] includes RAD tensors, RD, RA, and AD maps, all synchronized with camera images. Annotations comprise sparse points, dense masks, and bounding boxes for object detection, tracking, and radar semantic segmentation. Mounted on a stationary vehicle, the hardware captures data in a controlled environment. This dataset encompasses 12k annotated frames across three object categories: pedestrians, cyclists, and vehicles. Each RAD tensor has dimensions of \\(256 \\times 256 \\times 64\\), and the accompanying camera images boast a resolution of \\(1232 \\times 1028\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Proposed RadVRU Dataset Our RadVRU dataset is captured in several cities in Western Europe. The acquisition hardware consists of a 77 GHz FMCW radar (TI AWR1243), a full HD RGB camera (Realsense D435), and a 3D LiDAR (Ouster OS1-128) mounted on the top of a vehicle. This dataset contains total 78k frames, including 317 sequences of duration from \\(10\\mathrm{~s}\\) to \\(20\\mathrm{~s}\\). These frames are representative of the challenging driving scenarios, including different lighting conditions, complex traffic with cars, pedestrians, and cyclists. The annotations are performed by using a camera- and a LiDAR-based object detection model to pre-label the instances, and then these labels are manually corrected by human annotators. The pedestrians are annotated as single points on the ground plane. The dimensions of range, azimuth, and Doppler of every RAD tensor are \\(128 \\times 16 \\times 128\\), and the resolution of the camera images is \\(1920 \\times 1080\\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Camera-Assisted Radar-Based Network for Vulnerable Road User Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.275, + 0.145, + 0.726, + 0.16 + ], + "angle": 0, + "content": "Table 1: Comparisons with other methods on CARRADA dataset." + }, + { + "type": "table", + "bbox": [ + 0.249, + 0.171, + 0.751, + 0.312 + ], + "angle": 0, + "content": "
MethodTraining DataAPmAPmean Dist. Error
Ped.Cycll.Car
TMVA-Net [29]Radar0.440.430.370.370.67
PeakConv [25]Radar0.450.430.380.380.67
ARC [17]Radar0.620.570.500.540.67
RAMP-CNN [7]Radar0.750.640.500.610.64
CARB-Net-baselineRadar0.770.650.530.630.65
U-Net [34]Radar + Camera0.570.480.470.510.70
FCN [26]Radar + Camera0.570.430.470.460.69
Lim et al. [21]Radar + Camera0.610.530.510.530.66
CARB-NetRadar + Camera0.820.750.600.730.63
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.788, + 0.569 + ], + "angle": 0, + "content": "Implementation Details We use the data from CARRADA as an example to explain our implementation details. First, we down-sample the original radar tensor to \\(128 \\times 128 \\times 64\\) to reduce the high computing effort. Similar to [29], we use 5 sequential radar tensors and camera images as input, so the sizes become \\([NC_{r}, H_{r}, W_{r}] = [512, 128, 128]\\) and \\([NC_{c}, H_{c}, W_{c}] = [15, 1232, 1028]\\). The cooperative vision feed-forward is made by a pre-trained \"yolov8m\" [12] model. We filter the detections from YOLOv8 with a confidence threshold of 0.1, and project the detections to the size of vision feed-forward images \\([N, H_{f}, W_{f}] = [5, 256, 256]\\). After pixel-shuffle down-sampling by factor \\(D = 2\\), it becomes \\([ND^{2}, H_{r}, W_{r}] = [20, 128, 128]\\). Similarly, after the ESPCN [37], we apply pixel-shuffle up-sampling by factor \\(D = 2\\) to restore the resolution [256, 256]. For the proposed anchor offsets and context learning, we have \\((3k + 2) = 11\\) predictions. Hence, the size of the final output tensor is [11, 256, 256]. We use 0.8 meter as the radius of our non-maximum suppression algorithm. Please refer to the Supplementary Materials for more implementation details and network architecture." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.592, + 0.427, + 0.607 + ], + "angle": 0, + "content": "4.2 Experiment Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.619, + 0.787, + 0.727 + ], + "angle": 0, + "content": "Evaluation Metric Our experiments employ Average Precision (AP) and mean distance error as primary evaluation metrics. We define true positives as detections falling within a 1.5-meter boundary around the ground truth. We avoid multiple matches within the same boundary. Additionally, we measure localization precision by calculating the distance between true positives and ground truths in meters. Averaging these distance errors yields the mean distance error, providing an evaluation of localization accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Evaluation on CARRADA Dataset We compare the object detection performance of our proposed method and the other state-of-the-art approaches on the CARRADA dataset. Since some of the previous methods output semantic segmentation masks instead of detections, we carefully find the centers of the predicted masks as their localization results. For FCN [26] and U-Net [34] baselines, we use inverse projection mapping to project the RGB images to BEV" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.128 + ], + "angle": 0, + "content": "W.Y Lee et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.239, + 0.145, + 0.761, + 0.159 + ], + "angle": 0, + "content": "Table 2: Comparisons with other methods on our proposed RadVRU dataset." + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.172, + 0.723, + 0.277 + ], + "angle": 0, + "content": "
MethodTraining DatamAPmean Dist. Error
ARC [17]Radar0.510.64
CARB-Net-baselineRadar0.550.62
U-Net [34]Radar + Camera0.520.87
FCN [26]Radar + Camera0.350.94
Lim et al. [21]Radar + Camera0.670.63
EchoFusion* [24]Radar + Camera0.690.54
CARB-NetRadar + Camera0.810.47
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.302, + 0.788, + 0.499 + ], + "angle": 0, + "content": "down to a cell resolution of \\(128 \\times 128\\). Then, we concatenate the projected images and radar tensors along the channel and output the BEV map. For fair comparisons, we ensure the resolutions of the radar data and projected image are the same for all methods. Furthermore, we remove the cooperative vision feed-forward and train our CARB-Net-baseline model only on the radar tensors to compare with the radar-based methods. Although our method is designed for VRUs, we still involve cars for training and testing. We use the cars' center points in BEV space as the ground truth for detection. The vehicles' orientations are not considered during evaluation by all the methods. As observed in Tab. 1, compared to the others, our proposed framework can not only clearly improve the detection performance but also achieve a smaller mean distance error on the RA plane. The results demonstrate that the proposed method successfully integrates the detections from camera images to achieve better accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.788, + 0.61 + ], + "angle": 0, + "content": "Evaluation on Proposed RadVRU Dataset. Similarly, we conduct the same experiment on our RadVRU dataset. We evaluate our framework against the other four different models. For EchoFusion* [24], we use RA tensors as their input and re-implement their method with reduced network size due to our limited computing hardware. As illustrated in Tab. 2, our proposed approach also outperforms the other methods by a significant margin." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.751 + ], + "angle": 0, + "content": "Visualization In Fig. 4, we present camera images with bounding boxes predicted by YOLOv8 alongside the radar-based model ARC's [17] and our method's detections in BEV. Notably, our detections outperform ARC's, exhibiting fewer false positives and negatives, leading to superior detection performance. Even in low-light conditions, where YOLOv8 fails, our model successfully identifies pedestrians using radar data alone. Additionally, when road users are stationary, ARC fails due to the absence of micro-Doppler patterns, whereas our method utilizes the vision feed-forward to accurately identify pedestrians." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.772, + 0.391, + 0.787 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Effects of Proposed Context Learning To further investigate the impact of the proposed context learning, we conduct several experiments to show the robustness of our model when one of the sensors fail. In Tab. 3, the first row model" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Camera-Assisted Radar-Based Network for Vulnerable Road User Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.508, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.446, + 0.436, + 0.455 + ], + "angle": 0, + "content": "(a) Camera-based model detections" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.142, + 0.646, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.142, + 0.781, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.245, + 0.646, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.245, + 0.781, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.344, + 0.646, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.344, + 0.781, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.446, + 0.663, + 0.457 + ], + "angle": 0, + "content": "(b) Radar-based model confidence map" + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.446, + 0.772, + 0.456 + ], + "angle": 0, + "content": "(c) Our confidence map" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.469, + 0.788, + 0.541 + ], + "angle": 0, + "content": "Fig. 4: Visual comparisons of the detection results for three different scenarios in our proposed dataset. Green boxes in (a) represent the true positives detected by camera-based model. Red boxes/circles indicate false positives and negatives, and white boxes denote ground truth locations. Our model (c) exhibits enhanced detection performance compared to ARC [17] (b), integrating radar and camera data effectively." + }, + { + "type": "table_caption", + "bbox": [ + 0.345, + 0.559, + 0.658, + 0.573 + ], + "angle": 0, + "content": "Table 3: Effects of proposed context learning." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.585, + 0.782, + 0.656 + ], + "angle": 0, + "content": "
MethodTraining DatamAP w/ Radar OnlymAP w/ Camera OnlymAP w/ Radar & Camera
CARB-Net-baselineRadar0.55--
CARB-Net wo/ Context learningRadar + Camera0.100.450.75
CARB-Net w/ Mix trainingRadar + Camera0.400.520.75
CARB-Net w/ Context learningRadar + Camera0.540.650.81
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.788, + 0.793 + ], + "angle": 0, + "content": "is only retrained on radar data without cooperative vision feed-forward signals. The purpose of this model is to show the upper bound of the detection performance when there is only radar data available. The second row model is retrained without context learning. The third row model is retrained by a straightforward solution to consider sensor failure cases during training. We randomly dropout the radar or camera from the input to mix-train the model, which forces the model to learn how to detect in only one modality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "The model without context learning shows a significant performance drop when the camera-based model cannot detect the targets. In other words, this model is sensitive to the input and is not able to perform detection when one" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.128 + ], + "angle": 0, + "content": "W.Y Lee et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.36, + 0.145, + 0.645, + 0.16 + ], + "angle": 0, + "content": "Table 4: Effects of proposed components." + }, + { + "type": "table", + "bbox": [ + 0.271, + 0.171, + 0.732, + 0.267 + ], + "angle": 0, + "content": "
MethodVision Feed-ForwardImproved U-NetAnchor OffsetsContext LearningmAP
U-Net [34]----0.53
CARB-Net✓---0.65
✓✓--0.70
✓✓✓-0.75
✓✓✓✓0.81
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.294, + 0.787, + 0.355 + ], + "angle": 0, + "content": "sensor is not working as expected. The third model shows marginally improved results when the one of the sensor is unable to detect, but the detection performance in nominal cases are affected. In contrast, our proposed method shows significant improvements, no matter in nominal or sensor failure cases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.364, + 0.79, + 0.515 + ], + "angle": 0, + "content": "Effects of Proposed Components To clarify the effect of every proposed component, we conduct several experiments to show the performance difference with or without these components. Tab. 4 demonstrates that all the proposed components contribute to performance improvements. The most notable improvement is brought about by the cooperative vision feed-forward. When it is considered, the model achieves a greater level of performance. The projection of the predicted bounding boxes provided by the off-the-shelf camera-based model not only reduces the training effort but also prevents the distortion problem. The results show that when all the components are incorporated, the performance of the model is considerably enhanced." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.537, + 0.36, + 0.554 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In this paper, we introduce Camera-Assisted Radar-Based Network (CARB-Net), a novel radar-based framework aimed at addressing the low angular resolution of radar signals and enhancing robustness to conditions where a sensor is unable to detect the targets. CARB-Net effectively integrates detections from an off-the-shelf camera-based object detection model with radar data, enabling precise localization by learning object micro-Doppler patterns from radar tensors and incorporating vision feed-forward signals. Our context learning method ensures reliable performance even when one sensor fails. The paper contributes significantly by showcasing improved performance for vulnerable road users compared to state-of-the-art methods. Validation on our larger radar-camera dataset confirms these achievements. Additionally, our method demonstrates up to a \\(12\\%\\) improvement in mAP performance across diverse datasets, with ablation studies highlighting the efficacy of each framework component. Future research directions could explore optimizing fusion mechanisms and incorporating additional sensor modalities like thermal or acoustic sensors for even more accurate object detection and localization. However, as more sensors are integrated, ensuring robustness in multi-sensor object detection remains critical for real-world applications focused on enhancing road user safety." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Camera-Assisted Radar-Based Network for Vulnerable Road User Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.146, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.179, + 0.787, + 0.24 + ], + "angle": 0, + "content": "This work was funded by EU Horizon 2020 ECSEL JU research and innovation programme under grant agreement 876487 (NextPerception), by EU Horizon KDT JU research and innovation programme under grant agreement 101139769 (Distribuse), and by the Flemish Government (AI Research Program)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.264, + 0.323, + 0.279 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.785, + 0.338 + ], + "angle": 0, + "content": "1. Barnes, D., Gadd, M., Murcutt, P., Newman, P., Posner, I.: The oxford radar robotcar dataset: A radar extension to the oxford robotcar dataset. In: ICRA (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.339, + 0.785, + 0.38 + ], + "angle": 0, + "content": "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.381, + 0.785, + 0.408 + ], + "angle": 0, + "content": "3. Chen, S., He, W., Ren, J., Jiang, X.: Attention-based dual-stream vision transformer for radar gait recognition. In: ICASSP. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.409, + 0.785, + 0.435 + ], + "angle": 0, + "content": "4. Chen, X., Ma, H., Wan, J., Li, B., Xia, T.: Multi-view 3d object detection network for autonomous driving. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.437, + 0.785, + 0.491 + ], + "angle": 0, + "content": "5. Dimitrievski, M., Shopovska, I., Van Hamme, D., Veelaert, P., Philips, W.: Weakly supervised deep learning method for vulnerable road user detection in fmcw radar. In: IEEE International Conference on Intelligent Transportation Systems (ITSC). IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.493, + 0.785, + 0.533 + ], + "angle": 0, + "content": "6. Gao, X., Xing, G., Roy, S., Liu, H.: Experiments with mmwave automotive radar test-bed. In: 2019 53rd Asilomar conference on signals, systems, and computers. pp. 1-6. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.534, + 0.785, + 0.561 + ], + "angle": 0, + "content": "7. Gao, X., Xing, G., Roy, S., Liu, H.: Ramp-cnn: A novel neural network for enhanced automotive radar object recognition. IEEE Sensors Journal (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.563, + 0.785, + 0.603 + ], + "angle": 0, + "content": "8. Griffiths, H., Cohen, L., Watts, S., Mokole, E., Baker, C., Wicks, M., Blunt, S.: Radar spectrum engineering and management: Technical and regulatory issues. Proceedings of the IEEE (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.604, + 0.785, + 0.645 + ], + "angle": 0, + "content": "9. Gusland, D., Christiansen, J.M., Torvik, B., Fioranelli, F., Gurbuz, S.Z., Ritchie, M.: Open radar initiative: Large scale dataset for benchmarking of micro-doppler recognition algorithms. In: IEEE Radar Conference (RadarConf). pp. 1-6 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.672 + ], + "angle": 0, + "content": "10. Hartley, R., Zisserman, A.: Multiple view geometry in computer vision. Cambridge university press (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.674, + 0.785, + 0.714 + ], + "angle": 0, + "content": "1. Hwang, J.J., Kretzschmar, H., Manela, J., Rafferty, S., Armstrong-Crews, N., Chen, T., Anguelov, D.: Cramnet: Camera-radar fusion with ray-constrained cross-attention for robust 3d object detection. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.715, + 0.785, + 0.742 + ], + "angle": 0, + "content": "2. Jocher, G., Chaurasia, A., Qiu, J.: YOLO by Ultralytics (2023), https://github.com/ultralytics/ultralytics" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.743, + 0.785, + 0.77 + ], + "angle": 0, + "content": "3. Kim, Y., Kim, S., Choi, J.W., Kum, D.: Craft: Camera-radar 3d object detection with spatio-contextual fusion transformer. In: AAAI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.785, + 0.798 + ], + "angle": 0, + "content": "4. Kim, Y., Shin, J., Kim, S., Lee, I.J., Choi, J.W., Kum, D.: Crn: Camera radar net for accurate, robust, efficient 3d perception. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Kopp, J., Kellner, D., Piroli, A., Dietmayer, K.: Tackling clutter in radar data-label generation and detection using pointnet++. arXiv preprint arXiv:2303.09530 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.296, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.127 + ], + "angle": 0, + "content": "W.Y Lee et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "16. Lee, W.Y., Dimitrievski, M., Jovanov, L., Philips, W.: Spatio-temporal consistency for semi-supervised learning using 3d radar cubes. In: IEEE Intelligent Vehicles Symposium (IV). IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.786, + 0.232 + ], + "angle": 0, + "content": "17. Lee, W.Y., Jovanov, L., Kumcu, A., Philips, W.: Arc: Automotive radar consistency regularization for semi-supervised learning. IEEE Transactions on Intelligent Vehicles (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.786, + 0.261 + ], + "angle": 0, + "content": "18. Lee, W.Y., Jovanov, L., Philips, W.: Multi-view target transformation for pedestrian detection. In: WACV Workshops (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.786, + 0.289 + ], + "angle": 0, + "content": "19. Li, Y., Ge, Z., Yu, G., Wang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: AAAI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.786, + 0.331 + ], + "angle": 0, + "content": "20. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.786, + 0.374 + ], + "angle": 0, + "content": "21. Lim, T.Y., Ansari, A., Major, B., Fontijne, D., Hamilton, M., Gowaikar, R., Subramanian, S.: Radar and camera early fusion for vehicle detection in advanced driver assistance systems. In: NIPS Workshops (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.786, + 0.402 + ], + "angle": 0, + "content": "22. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.786, + 0.43 + ], + "angle": 0, + "content": "23. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.Y., Berg, A.C.: Ssd: Single shot multibox detector. In: ECCV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.786, + 0.459 + ], + "angle": 0, + "content": "24. Liu, Y., Wang, F., Wang, N., ZHANG, Z.X.: Echoes beyond points: Unleashing the power of raw radar data in multi-modality fusion. NeurIPS (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.786, + 0.501 + ], + "angle": 0, + "content": "25. Liwen, Z., Xinyan, Z., Youcheng, Z., Yufei, G., Yuanpei, C., Xuhui, H., Zhe, M.: Peakconv: Learning peak receptive field for radar semantic segmentation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.786, + 0.529 + ], + "angle": 0, + "content": "26. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: CVPR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.786, + 0.572 + ], + "angle": 0, + "content": "27. Mallot, H.A., Bulthoff, H.H., Little, J.J., Bohrer, S.: Inverse perspective mapping simplifies optical flow computation and obstacle detection. Biological cybernetics (1991)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.786, + 0.6 + ], + "angle": 0, + "content": "28. Meyer, M., Kuschk, G.: Automotive radar dataset for deep learning based 3d object detection. In: European Radar Conference (EuRAD) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.786, + 0.628 + ], + "angle": 0, + "content": "29. Ouaknine, A., Newson, A., Pérez, P., Tupin, F., Rebut, J.: Multi-view radar semantic segmentation. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.786, + 0.657 + ], + "angle": 0, + "content": "30. Ouaknine, A., Newson, A., Rebut, J., Tupin, F., Pérez, P.: Carrada dataset: Camera and automotive radar with range- angle- doppler annotations. In: ICPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.786, + 0.685 + ], + "angle": 0, + "content": "31. Paek, D.H., Kong, S.H., Wijaya, K.T.: K-radar: 4d radar object detection for autonomous driving in various weather conditions. NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.786, + 0.741 + ], + "angle": 0, + "content": "32. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 194-210. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.786, + 0.77 + ], + "angle": 0, + "content": "33. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.786, + 0.799 + ], + "angle": 0, + "content": "34. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI. Springer (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "35. Schumann, O., Hahn, M., Dickmann, J., Wöhler, C.: Semantic segmentation on radar point clouds. In: International Conference on Information Fusion (FUSION) (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Camera-Assisted Radar-Based Network for Vulnerable Road User Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "36. Sheeny, M., De Pellegrin, E., Mukherjee, S., Ahrabian, A., Wang, S., Wallace, A.: Radiate: A radar dataset for automotive perception in bad weather. In: ICRA (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.231 + ], + "angle": 0, + "content": "37. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.785, + 0.272 + ], + "angle": 0, + "content": "38. Vaizman, Y., Weibel, N., Lanckriet, G.: Context recognition in-the-wild: Unified model for multi-modal sensors and multi-label classification. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "39. Wang, W., Tran, D., Feiszli, M.: What makes training multi-modal classification networks hard? In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "40. Wang, Y., Wang, G., Hsu, H.M., Liu, H., Hwang, J.N.: Rethinking of radar's role: A camera-radar dataset and systematic annotator via coordinate alignment. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.397 + ], + "angle": 0, + "content": "41. Yao, S., Guan, R., Huang, X., Li, Z., Sha, X., Yue, Y., Lim, E.G., Seo, H., Man, K.L., Zhu, X., Yue, Y.: Radar-camera fusion for object detection and semantic segmentation in autonomous driving: A comprehensive review. IEEE Transactions on Intelligent Vehicles (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.397, + 0.785, + 0.439 + ], + "angle": 0, + "content": "42. Zhou, T., Chen, J., Shi, Y., Jiang, K., Yang, M., Yang, D.: Bridging the view disparity between radar and camera features for multi-modal fusion 3d object detection. IEEE Transactions on Intelligent Vehicles (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.439 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CARB-Net_ Camera-Assisted Radar-Based Network for Vulnerable Road User Detection/ed9a2da4-1d84-4538-8283-1656bf597855_origin.pdf b/2024/CARB-Net_ Camera-Assisted Radar-Based Network for Vulnerable Road User Detection/ed9a2da4-1d84-4538-8283-1656bf597855_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b2051de818393abd3fba76e42fc0a27ae1c48dde --- /dev/null +++ b/2024/CARB-Net_ Camera-Assisted Radar-Based Network for Vulnerable Road User Detection/ed9a2da4-1d84-4538-8283-1656bf597855_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f332b0232265044c501c828fb103cbc37ad4d03fa28362dfb95b0cb27f512f13 +size 933112 diff --git a/2024/CARB-Net_ Camera-Assisted Radar-Based Network for Vulnerable Road User Detection/full.md b/2024/CARB-Net_ Camera-Assisted Radar-Based Network for Vulnerable Road User Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5806f00848cb9ca655da4ca258ca314d05455d87 --- /dev/null +++ b/2024/CARB-Net_ Camera-Assisted Radar-Based Network for Vulnerable Road User Detection/full.md @@ -0,0 +1,223 @@ +# CARB-Net: Camera-Assisted Radar-Based Network for Vulnerable Road User Detection + +Wei-Yu Lee $^{\text{①}}$ , Martin Dimitrievski $^{\text{②}}$ , David Van Hamme $^{\text{③}}$ , Jan Aelterman $^{\text{④}}$ , Ljubomir Jovanov $^{\text{⑤}}$ , and Wilfried Philips $^{\text{⑥}}$ + +TELIN-IPI, Ghent University-IMEC, Gent, Belgium + +Abstract. Ensuring a reliable perception of vulnerable road users is crucial for safe autonomous driving. Radar stands out as an appealing sensor choice due to its resilience in adverse weather, cost-effectiveness, depth sensing capabilities, and established role in adaptive cruise control. Nevertheless, radar's limited angular resolution poses challenges in object recognition, especially in distinguishing targets in close proximity. To tackle this limitation, we present the Camera-Assisted Radar-Based Network (CARB-Net), a novel and efficient framework that merges the angular accuracy of a camera with the robustness and depth sensing capabilities of radar. We integrate camera detection information through a ground plane feed-forward array, entangling it with the early stages of a radar-based detection network. Furthermore, we introduce a unique context learning approach to ensure graceful degradation in situations of poor radar Doppler information or unfavorable camera viewing conditions. Experimental validations on public and our proposed datasets, along with benchmark comparisons, showcase CARB-Net's superiority, boasting up to a $12\%$ improvement in mAP performance. A series of ablation studies further emphasize the efficacy of the CARB-Net architecture. Our proposed dataset is released at https://github.com/weiyulee/RadVRU. + +Keywords: Micro-Doppler Signature $\cdot$ Vulnerable Road User Detection + +# 1 Introduction + +Robust Vulnerable Road User (VRU) perception is vital for driving safety. Among the myriad of sensors used, millimeter-wave radar stands out due to its resilience in challenging light and weather conditions, such as rain, fog, and snow, and its depth and Doppler velocity capability. Its cost-effectiveness has also contributed to its widespread adoption, making it a fundamental component in many Advanced Driving Assistance Systems (ADAS) applications. + +However, radar sensors have inherent limitations: Their angular resolution is limited by the size of the antenna array and operating wavelength. Compared to HD cameras, typical radars have two orders of magnitude lower angular resolution [30, 36, 40], making it challenging to distinguish between closely spaced objects. Beyond detection, this low signal resolution makes classifying between object categories even more difficult. In cluttered environments, multipath interference noise further complicates the interpretation of radar data. + +![](images/7437572b73d31a9953fd02091fb67d4ed3dee7944e9df342f7a1c1c46dbb247c.jpg) +Fig. 1: An example of micro-Doppler patterns. left: The radar signal is projected and overlaid with the synchronized camera image. right: The radar Doppler signal is plotted over time. The micro-Doppler pattern of the cyclist is highlighted by a dotted line. + +In the literature, radar is often used as a means to improve 3D object detection in monocular systems. By fusing imaging information from cameras with radar, researchers have proposed to exploit the strengths of both technologies [13,14,42]. In these prior studies, cameras play a key role in providing rich visual cues that can be used to train effective object detection and classification models. However, in low-light environments, cameras cannot provide the recognition of road users, and detection of targets by only using radar point clouds becomes difficult due to their sparsity. One of the possible alternative cues for VRU detection is using Doppler information from dense radar tensors, alternatively referred to as Radar Frequency (RF) images. This radar data representation provides the unique motion patterns in the time-frequency domain [5,16,17]. For example, in Figure 1, a cyclist exhibits an oscillatory motion of the limbs, which results in a RF signal with sufficient features to classify these road users. These periodically changing motion patterns are commonly referred to as micro-Doppler patterns, providing fundamental features for identifying road users. However, the low azimuth resolution of radar sensors still limits the accuracy of the predicted object positions in these studies. + +Motivated by these challenges, we propose a RF-based, camera-assisted fusion network with two significant novelties that enhance detection accuracy in nominal conditions (i.e., sensors are working as expected), but more importantly, under conditions of sensor failure. A new cooperative fusion approach introduces summarized camera observations into the early stages of a radar-based network, helping in better feature extraction over the Doppler signal in the RF images. Furthermore, we propose a robust way of entangling camera and radar data, helping the network enhance robustness against scenarios where a sensor fails to detect the targets. By applying these two improvements, we achieve detector robustness without the need for excessive training in unforeseen situations. + +We focus on the task of Bird's Eye View (BEV) vulnerable road user perception, where we train a radar-based network on radar frequency images, or more specifically, Range-Azimuth-Doppler (RAD) tensors. The RAD encodes the micro-Doppler pattern of targets at individual range-azimuth spatial cells and is used as the core feature for both detection and classification. Camera information is appended to the RAD tensor in the form of a summarized detection array in the BEV plane, which assists the radar-based network when it is available. We use an off-the-shelf camera-based object detection model to produce detec + +tions of road users and populate the feed-forward array using back-projection. The input layer of the network then applies pixel-shuffle [37] to reorganize the high-resolution camera observations into a common representation, matching the low-resolution radar data. This allows a seamless cooperative fusion with radar tensors without loss of information. + +Robustness against scenarios where a sensor fails to detect the targets is achieved by augmenting the data and expanding the ground truth labels into sub-categories of possible fusion contexts of nominal and deteriorated sensor performance for each object. The novel aspect of training is to generate and encode these conditions as control variables, that augment the ground truths during training. This strategy encourages the model to learn more sophisticated and finer features for each modality without additional human annotation, enhancing its adaptability to diverse fusion contexts. Our experiments, conducted on both public and our proposed RadVRU datasets, demonstrate the effectiveness of our approach. Our method not only achieves state-of-the-art detection performance but also exhibits a remarkable up to $12\%$ improvement in challenging sensor failure cases. Our contributions are outlined as follows: + +1. Novel camera-assisted radar-based network based on pixel-shuffling of high-resolution vision feed-forward and low-resolution micro-Doppler signatures. +2. Novel context learning strategy, utilizing context-specific control variables to train a model that is robust to conditions where a sensor is unable to detect the targets. +3. An extensive experimental evaluation on two datasets for BEV detection where we objectively compare to existing state-of-the-art methods. + +# 2 Related Work + +Radar-Based Object Detection Prior studies utilize radar as the primary sensor for object detection, employing various data representations: point clouds and dense radar tensors. For radar point clouds, the sparsity and lack of context pose detection challenges [33, 35, 41] for neural networks to learn features effectively. On the other hand, image-like radar tensors provide rich but noisy contextual data [8, 15]. Many prior studies utilize deep learning to recognize objects by their motion characteristics [3, 9]. For instance, radar's unique micro-Doppler signature, a time-varying Doppler frequency shift captured in RAD tensors, enables accurate recognition of road users like pedestrians and cyclists [5,6,16,17]. This distinctive motion pattern advantage motivates our use of RAD tensors as inputs, harnessing the micro-Doppler signature for robust object detection. + +On the other hand, radar sensors, while effective at detecting moving road users, struggle to separate closely packed objects, such as groups of pedestrians, due to their low angular resolution [41]. Additionally, radar sensors fail to detect stationary objects due to interference from other reflections and a lack of motion information. As a result, the existing radar-only methods cannot accurately localize and recognize all objects. In this study, we utilize observations from the camera as assistance to improve the precision of azimuth localization. + +Radar-Camera-Based Object Detection Recently, using BEV representation to merge images from multiple camera and radar sensors has sparked a rising wave of research in autonomous driving. Prior studies adopt large pretrained image networks, such as LSS [32], BEVFormer [20], or BEVDepth [19], to project 2D image features onto the BEV for further fusion with radar data. For instance, in [13], the authors use proposals from a camera-based network to associate and merge with radar point clouds. In [14], a radar-assisted transformation method is proposed to project 2D images onto BEV and aggregate image and radar feature maps for detection. However, the perspective transformation of camera views highly relies on accurate calibration parameters and good light conditions. Numerous objects visible to cameras may yield only a few radar points, posing significant challenges when cameras are unreliable and sparse point clouds cannot provide sufficient evidence for recognition. + +In contrast, our method interfaces with the radar at a much earlier stage, utilizing 3D dense radar tensors as input to learn the micro-Doppler patterns for recognition instead of relying on camera features. Therefore, detection can be performed only on radar data in low-light conditions without cameras. Moreover, existing fusion methods often necessitate training a huge network on a large amount of both camera and radar data, which poses challenges due to the lack of publicly available datasets containing both modalities and the high computing effort. To address this, we utilize detections from an efficient off-the-shelf camera-based detection model, reducing the training and inference burden. Since our method's core is a radar-based detector using micro-Doppler patterns to recognize VRUs, only aided by the camera when it is available, we mainly consider the dataset providing 3D radar tensors for experiments, and the existing BEV radar-camera fusion methods cannot be directly applied to this task. + +Robustness of Multimodal Fusion In practice, it is quite common for sensors to become unreliable in specific contexts. This negatively affects the robustness of the perception system, such as occlusions or low-light environments for cameras. Prior studies introduce sensor dropout [4,11,38,39] during training to enhance the network's ability to handle these conditions with single modality features. However, simply removing the sensor data from the input constrains the fusion potential and usually degrades the overall detection performance. In this work, we introduce a novel context learning strategy that augments the data by creating challenging scenarios that render sensor observations unreliable, and we instruct the network to recognize these situations through auxiliary queries, enhancing model robustness in detection tasks. + +# 3 Methodology + +In this paper, we introduce Camera-Assisted Radar-Based Network (CARB-Net), a novel neural network model for integrating low spatial resolution radar tensors and high-resolution camera images. As depicted in Fig. 2, the proposed framework operates in the spatio-temporal domain, taking sequential camera + +![](images/2af3859c66ae4e3d05f041cd77ccd6e1ec928bb0235bff8cc2a1f1c91c7e6128.jpg) +Fig.2: Overview of the proposed Camera-Assisted Radar-Based Network (CARB-Net). + +images $I_{\mathrm{c}} \in \mathbb{R}^{NC_{c} \times H_{c} \times W_{c}}$ and sequential RAD radar tensors $T_{\mathrm{r}} \in \mathbb{R}^{NC_{r} \times H_{r} \times W_{r}}$ , where $N$ denotes the number of tensors and images in the sequence, $C_{c}, C_{r}$ represent channel sizes, and $(H_{c}, W_{c})$ and $(H_{r}, W_{r})$ are the spatial dimensions of the RAD tensors and camera images. + +Unified Data Representation The data representations of radars and cameras are different, which poses challenges in sensor fusion as association requires relating objects in different coordinate systems. Therefore, it is essential to establish a common representation suitable for both fusion and detection. Front View and Bird's Eye View are the two mainstream representations in prior studies. For front view, due to the low resolution in the azimuth angle provided by the radar as well as camera calibration errors, projected radar signals may deviate from the object in the image, which makes the observations of the same object contradictory and hard to be fused by networks. On the other hand, BEV provides a view that accurately portrays the scale and positions of objects, making it a suitable representation for autonomous driving tasks. In this study, we use BEV as the unified data representation. + +Given the image coordinates $(u, v)$ , we rely on the intrinsic camera parameters and extrinsic parameters to perform transformation from camera reference frame into BEV coordinates. We assume the point where the target touches the ground plane (foot or wheel) is with zero height, then the image coordinates $(u, v)$ can be projected to BEV [10, 27]. However, this assumption can be inaccurate in cases when the camera cannot clearly capture the bottom-most point of an object, the target is jumping in the air, or the ground is not flat. This inaccuracy severely degrades the detection performance when using perspective transformation to project the RGB image to BEV for aligning radar and camera signals. Furthermore, perspective transformation also introduces image distortion because of the zero height assumption. The distortion can severely corrupt objects' visual appearance, such as changing the shape and structure, which causes difficulties for the network to learn from the camera images [18]. + +![](images/544b31def2fdbc60b3f505c05c7dc9422526a27d724156d33dcd979d5bc8e971.jpg) +Fig.3: A camera image with predicted bounding boxes and the cooperative vision feed-forward image in BEV. Blue and green colors represent the pedestrian and bicycle classes. Yellow arrows indicate the mapping after perspective transformation. + +Cooperative Vision Feed-Forward In order to address the aforementioned problems, we propose Cooperative Vision Feed-Forward to merge radar data uncertain in azimuth but certain in range with RGB data uncertain in range but certain in azimuth. Despite the uncertainties of radar and camera observations, there is usually a small region on the BEV plane where the radar and camera observations exhibit a strong response, which we use for road user detection. + +As depicted in Figure 3, we introduce the detections made by using camera data in the form of a vision feed-forward array on the BEV plane. Specifically, we utilize the off-the-shelf camera-based detection network to produce 2D detections of road users based on the RGB images $I_{\mathrm{c}}$ , then we project the center of the predicted bounding boxes of the road users to the BEV to make vision feedforward images $I_{\mathrm{f}}$ . The projected values on the BEV are the confidence scores, and these projected centers can be seen as summarized information from the pretrained network using RGB images. By doing this, little to no spatial distortion will be caused by the back-projection. + +In addition to the projection based on the zero height assumption, we estimate the depth of a target from its width and height in the image frame and its average size (i.e., averaged human heights and shoulder widths) in the real-world to estimate robust depth values. As a result, we use the median value of these estimations as the depth value. To compensate for the projection errors caused by inaccurate perspective transformation, we enlarge the projected positions from points to regions by applying a Gaussian kernel to cover the uncertainty of the depth estimation. Specifically, we never know the actual heights and widths of pedestrians in practice, but instead we use a statistical model of the mean height and width of humans, $1.2\mathrm{m}$ to $2.0\mathrm{m}$ height and $0.36\mathrm{m}$ to $0.44\mathrm{m}$ width, respectively. That means the projection ends up on a segment of a perspective line of $x$ meters, which becomes the sigma of our Gaussian kernel. As long as the sigma is not underestimated, the result will not critically depend on its value. Therefore, the inaccuracy problem caused by perspective transformation can be alleviated. + +Spatial Resolution Alignment Efficiently aligning low resolution radar data and high resolution camera signals on the ground plane is a challenge. A straight- + +forward solution for aligning is to use a BEV space with the same low spatial resolution as the radar sensor [21]. However, the low spatial resolution of the radar causes this approach to discard high-resolution camera information, negatively affecting the fusion outcomes. On the other hand, upscaling low-resolution radar to align with high-resolution cameras is also not practical because of the high computing effort. Therefore, we introduce Pixel-shuffle [37] Downsampling (PD) to tackle this resolution misalignment problem. PD is to reorganize the high-resolution image data into several low resolution channels. This operation only rearranges the order of the elements to down-scale the image instead of encoding or pooling. Hence, all the information of the image can be preserved, and the network's time and space complexity will not be increased. Specifically, we make vision feed-forward images $I_{\mathrm{f}} \in \mathbb{R}^{N \times H_f \times W_f}$ , where $(H_f \times W_f)$ is an intermediate spatial resolution between RGB images and radar cubes. In this instance, we apply PD on the vision feed-forward images $I_{\mathrm{f}}$ to obtain $I_{\mathrm{f}}' \in \mathbb{R}^{ND^2 \times H_r \times W_r}$ , where $D$ is the down-scaling factor, such that the subsequent network may jointly extract features from the radar tensors and vision feed-forward images with varying resolutions using convolution kernels. The solution has two main advantages: (1) no additional network training effort on camera images is needed, and (2) resolution alignment without information loss. + +Camera-Assisted Radar-Based Network The aim of the proposed network is to identify road users by using sequential RAD tensors $T_{\mathrm{r}}$ and vision feedforward images $I_{\mathrm{f}}$ . The RAD tensors capture micro-Doppler patterns and precise range measurements but have limited azimuth resolution. Our network interprets these tensors to pinpoint and classify various target categories. Conversely, the vision feed-forward images offer detailed azimuth localization of VRUs from a pre-trained camera-based network but lack accuracy in range estimation. Despite the inherent uncertainties, there is a tiny region on the BEV plane where all sensors exhibit a strong response, which we use for road user detection. + +Our CARB-Net is developed based on the principles of multi-resolution feature decomposition in U-Net [34] with several improvements. First, we use a small convolution network, which we call the Fusion Encoder, to encode the RF information from the radar and vision feed-forward images into a compact channel-representation as the input for U-Net. In addition, we apply the concept of pixel-shuffle [37] again to replace the max-pooling and up-scaling modules inside U-Net. The feature maps are easily up- or down-scaled by reorganizing the image channels without adding additional model parameters. Compared to deconvolution, max-pooling, or bilinear interpolation, pixel-shuffle is used as an effective way to prevent resolution loss during the encoding and decoding steps. + +To obtain high-resolution output, we apply ESPCN [37] and pixel-shuffle up-sampling by factor $D$ to reorganize the channels of the network output. Ultimately, the spatial resolution of the output tensor becomes $(H_{f} \times W_{f})$ . With this architecture, we allow the networks to process input with different resolutions. + +Context Learning Radars or cameras might not be able to provide recognition of road users in specific contexts, which is unacceptable for real-world deployment. For instance, a stationary object is poorly visible to radars, and adverse weather conditions may make objects hard-to-spot by cameras. However, most existing methods do not focus on such cases during training but only emphasize average detection performance, causing significant performance drops under such contexts compared to favorable circumstances. + +In order to improve robustness in these specific contexts, we make the network answer 2 additional auxiliary questions to distinguish the sensing capability of each sensor. For each position in the output confidence map, the network has to determine: (1) is there a VRU or not; (2) can a VRU be detected by the radar; and (3) can a VRU be detected by the camera. During training, we apply data augmentation and increase the variability of available labeled data in order to cover more combinations of the possible answers to these questions. + +We define $y \in \{y_0, y_1, y_2\}$ to represent the 3 correct answers to the 3 questions. Specifically, $y_0$ indicates whether this position is occupied by a road user. The network should use all the sensor data as input to detect VRUs. $y_1$ describes if the road user's micro-Doppler pattern is detectable by the radar sensor. In this case, we focus on this specific context that makes the targets poorly visible to radar. Since the targets are stationary or moving without micro-Doppler patterns, the network cannot distinguish their radar signals from other obstacles' reflections. Finally, $y_2$ indicates whether the pre-trained camera-based network can detect the target or not. In this case, we focus on the conditions that make the camera sensor unable to see the target. By training the network to answer these 3 questions, we use the network's first answer as the final fusion decision to distinguish if the road user is present, and we use the others to describe the specific fusion contexts for each target. + +For instance, $[y_0, y_1, y_2] = [1, 1, 1]$ is the common case that both radar and camera can sense the targets, and $[y_0, y_1, y_2] = [1, 0, 1]$ means the road user is only detectable by the camera. This way, our network learns from the correct answers to distinguish these circumstances and understand the scene in more detail. During training, this approach additionally allows us to consider and penalize each class (e.g., pedestrian, cyclist, etc.) independently. This is achieved by tasking the network to perform this classification as well, thus penalizing solutions that fail to interpret their input data or context. That is, for each class, the network is required to output the 3 answers to the 3 questions, which are independent of the other classes' answers. + +In order to teach the network to distinguish these specific contexts, we control the input data to generate the corresponding contexts for training. For radar, the ideal way to simulate targets moving without micro-Doppler patterns or stationary targets is to remove their micro-Doppler patterns and retain only the mean velocities. However, this solution is not practical because the radar signals usually contain reflections from other objects and clutter. Moreover, micro-Doppler information from one object can spill over into nearby cells, depending on scene geometry. It is challenging to only remove a certain target's motion information + +while not affecting the others. Instead, we propose an alternative way to simulate such cases: the latest RAD tensor is repeated inside the sequential RAD tensors $T_{\mathrm{r}}$ to replace the other $N - 1$ radar tensors. That is, all the objects in the scene are preserved at their mean velocities, and most of their micro-Doppler patterns are eliminated. So, the objects' motion inside the repeated tensors is seen as an extension of the current movement without introducing any periodic motion. However, there are several downsides to this solution. First, if the target is moving very fast, the radar can still capture its micro-Doppler patterns in a single radar cube. Secondly, repeating the latest radar cube results in radar data where objects are static in RA plane but have non-zero velocity, a situation that is impossible in reality. However, this data augmentation is a close enough simulation of the most important radar limitations: its weakness in detecting slow motion and people with low micro-Doppler variation amplitude. Furthermore, it is easy to implement, and we do not need any additional computing effort to carefully remove the micro-Doppler patterns. For RGB cameras, we switch on and off the cooperative vision feed-forward to simulate the contexts in which the camera sensor cannot detect the target. In this case, the network should only use radar signals to recognize road users. + +Anchor Points and Offsets In the output layer, we incorporate anchor points and offsets to precisely refine VRU positions at any resolution. This technique increases the achievable resolution without increasing the data or grid dimensions by using an additional channel of output data to encode highly precise spatial offsets to a target from a low-dimensional grid, low-precision position. We predefine our anchor points over a uniformly spaced range and azimuth positions to allow for an equal capacity for detecting objects at all distances. Since there is a low variation in the size of the VRUs in BEV, we refrain from predicting object width and height. Specifically, we utilize range and azimuth bins from the radar configuration to pre-define the anchor points, representing certain positions on the ground plane. The network predicts 2 offsets for two directions relative to the default anchor point. To ensure that each VRU only be matched by a subset of anchor points, we restrict the offsets to being only inside the 4 surrounding anchor points. Finally, we perform non-maximum suppression to merge nearby anchor points using a greedy approach that favors the highest confidence scores within a pre-defined radius. This is a standard methodology employed in image object detection, which we found also works well in BEV detection using radar. + +Loss Functions For a certain position $\mathbf{r}$ in the output tensor, it contains $(3k + 2)$ different output values: 2 offsets and $3k$ classes (original $k$ different categories from ground truths and 3 answers for context learning), which results in a total of $(3k + 2)H_{f}W_{f}$ outputs. We define $O_{c,y}(\mathbf{r})$ that refers to the probability of category $c$ with answer $y$ , where $c \in \{c_0, c_1, \dots, c_k\}$ and $y \in \{y_0, y_1, y_2\}$ . In order to reduce the inherent class imbalance between the road users and background in the sparsely populated training BEV maps, we apply Focal loss [22] to compute the classification loss $\mathcal{L}_{\mathrm{cls}}$ : + +$$ +\mathcal {L} _ {\mathrm {c l s}} (\mathbf {r}) = - \sum_ {c} \sum_ {y} \alpha_ {c} \left[ (1 - O _ {c, y} (\mathbf {r})) ^ {\gamma} \log \left(O _ {c, y} (\mathbf {r})\right) + O _ {c, y} (\mathbf {r}) ^ {\gamma} (1 - y) \log \left(1 - O _ {c, y} (\mathbf {r})\right) \right], \tag {1} +$$ + +where $\alpha_{c}$ represents the weighting factor for class $c$ , and $\gamma \geq 0$ is the hyperparameter to down-weigh easy samples. We use smooth L1-norm as our localization loss $\mathcal{L}_{\mathrm{loc}}$ . This is similar to the bounding box regression in SSD [23] but we remove the multiple boxes and only predict the offsets for every anchor point. + +# 4 Experiment + +# 4.1 Datasets and Implementation Details + +In most public datasets, data representations of radar signals are in the form of a processed point cloud containing targets [2] or projected RD or RA maps [1,28, 31,36,40]. Since our method is based on the micro-Doppler signature to recognize road users, RAD tensors with synchronized camera images are necessary. Hence, we identified the publicly available CARRADA dataset [30] to be best suited for conducting experimental evaluation. In addition, we present results from experimentation using our proposed RadVRU dataset. + +CARRADA Dataset The CARRADA dataset [30] includes RAD tensors, RD, RA, and AD maps, all synchronized with camera images. Annotations comprise sparse points, dense masks, and bounding boxes for object detection, tracking, and radar semantic segmentation. Mounted on a stationary vehicle, the hardware captures data in a controlled environment. This dataset encompasses 12k annotated frames across three object categories: pedestrians, cyclists, and vehicles. Each RAD tensor has dimensions of $256 \times 256 \times 64$ , and the accompanying camera images boast a resolution of $1232 \times 1028$ . + +Proposed RadVRU Dataset Our RadVRU dataset is captured in several cities in Western Europe. The acquisition hardware consists of a 77 GHz FMCW radar (TI AWR1243), a full HD RGB camera (Realsense D435), and a 3D LiDAR (Ouster OS1-128) mounted on the top of a vehicle. This dataset contains total 78k frames, including 317 sequences of duration from $10\mathrm{~s}$ to $20\mathrm{~s}$ . These frames are representative of the challenging driving scenarios, including different lighting conditions, complex traffic with cars, pedestrians, and cyclists. The annotations are performed by using a camera- and a LiDAR-based object detection model to pre-label the instances, and then these labels are manually corrected by human annotators. The pedestrians are annotated as single points on the ground plane. The dimensions of range, azimuth, and Doppler of every RAD tensor are $128 \times 16 \times 128$ , and the resolution of the camera images is $1920 \times 1080$ . + +Table 1: Comparisons with other methods on CARRADA dataset. + +
MethodTraining DataAPmAPmean Dist. Error
Ped.Cycll.Car
TMVA-Net [29]Radar0.440.430.370.370.67
PeakConv [25]Radar0.450.430.380.380.67
ARC [17]Radar0.620.570.500.540.67
RAMP-CNN [7]Radar0.750.640.500.610.64
CARB-Net-baselineRadar0.770.650.530.630.65
U-Net [34]Radar + Camera0.570.480.470.510.70
FCN [26]Radar + Camera0.570.430.470.460.69
Lim et al. [21]Radar + Camera0.610.530.510.530.66
CARB-NetRadar + Camera0.820.750.600.730.63
+ +Implementation Details We use the data from CARRADA as an example to explain our implementation details. First, we down-sample the original radar tensor to $128 \times 128 \times 64$ to reduce the high computing effort. Similar to [29], we use 5 sequential radar tensors and camera images as input, so the sizes become $[NC_{r}, H_{r}, W_{r}] = [512, 128, 128]$ and $[NC_{c}, H_{c}, W_{c}] = [15, 1232, 1028]$ . The cooperative vision feed-forward is made by a pre-trained "yolov8m" [12] model. We filter the detections from YOLOv8 with a confidence threshold of 0.1, and project the detections to the size of vision feed-forward images $[N, H_{f}, W_{f}] = [5, 256, 256]$ . After pixel-shuffle down-sampling by factor $D = 2$ , it becomes $[ND^{2}, H_{r}, W_{r}] = [20, 128, 128]$ . Similarly, after the ESPCN [37], we apply pixel-shuffle up-sampling by factor $D = 2$ to restore the resolution [256, 256]. For the proposed anchor offsets and context learning, we have $(3k + 2) = 11$ predictions. Hence, the size of the final output tensor is [11, 256, 256]. We use 0.8 meter as the radius of our non-maximum suppression algorithm. Please refer to the Supplementary Materials for more implementation details and network architecture. + +# 4.2 Experiment Results + +Evaluation Metric Our experiments employ Average Precision (AP) and mean distance error as primary evaluation metrics. We define true positives as detections falling within a 1.5-meter boundary around the ground truth. We avoid multiple matches within the same boundary. Additionally, we measure localization precision by calculating the distance between true positives and ground truths in meters. Averaging these distance errors yields the mean distance error, providing an evaluation of localization accuracy. + +Evaluation on CARRADA Dataset We compare the object detection performance of our proposed method and the other state-of-the-art approaches on the CARRADA dataset. Since some of the previous methods output semantic segmentation masks instead of detections, we carefully find the centers of the predicted masks as their localization results. For FCN [26] and U-Net [34] baselines, we use inverse projection mapping to project the RGB images to BEV + +Table 2: Comparisons with other methods on our proposed RadVRU dataset. + +
MethodTraining DatamAPmean Dist. Error
ARC [17]Radar0.510.64
CARB-Net-baselineRadar0.550.62
U-Net [34]Radar + Camera0.520.87
FCN [26]Radar + Camera0.350.94
Lim et al. [21]Radar + Camera0.670.63
EchoFusion* [24]Radar + Camera0.690.54
CARB-NetRadar + Camera0.810.47
+ +down to a cell resolution of $128 \times 128$ . Then, we concatenate the projected images and radar tensors along the channel and output the BEV map. For fair comparisons, we ensure the resolutions of the radar data and projected image are the same for all methods. Furthermore, we remove the cooperative vision feed-forward and train our CARB-Net-baseline model only on the radar tensors to compare with the radar-based methods. Although our method is designed for VRUs, we still involve cars for training and testing. We use the cars' center points in BEV space as the ground truth for detection. The vehicles' orientations are not considered during evaluation by all the methods. As observed in Tab. 1, compared to the others, our proposed framework can not only clearly improve the detection performance but also achieve a smaller mean distance error on the RA plane. The results demonstrate that the proposed method successfully integrates the detections from camera images to achieve better accuracy. + +Evaluation on Proposed RadVRU Dataset. Similarly, we conduct the same experiment on our RadVRU dataset. We evaluate our framework against the other four different models. For EchoFusion* [24], we use RA tensors as their input and re-implement their method with reduced network size due to our limited computing hardware. As illustrated in Tab. 2, our proposed approach also outperforms the other methods by a significant margin. + +Visualization In Fig. 4, we present camera images with bounding boxes predicted by YOLOv8 alongside the radar-based model ARC's [17] and our method's detections in BEV. Notably, our detections outperform ARC's, exhibiting fewer false positives and negatives, leading to superior detection performance. Even in low-light conditions, where YOLOv8 fails, our model successfully identifies pedestrians using radar data alone. Additionally, when road users are stationary, ARC fails due to the absence of micro-Doppler patterns, whereas our method utilizes the vision feed-forward to accurately identify pedestrians. + +# 4.3 Ablation Study + +Effects of Proposed Context Learning To further investigate the impact of the proposed context learning, we conduct several experiments to show the robustness of our model when one of the sensors fail. In Tab. 3, the first row model + +![](images/2217ae1c1395838775ae22e640466c5183b987dd86da666175ea46ade1375742.jpg) +(a) Camera-based model detections + +![](images/1f2913b846689011fb5338d7e1f1b9ad9e1146e8ed0900a106dcb4f0ee16121b.jpg) + +![](images/7554deb7af81caef1e80d93ea79d5a08702a0aedd1ff142e199d193e82dfa73d.jpg) + +![](images/64838ba3caf42518a0a4719efc9fcc9ec90b5d858b691af12269ef60b4d3bd38.jpg) + +![](images/545fb2f719a5e9b0f6db7c277e0a8af0da5f2684c4557d61d649791283632275.jpg) + +![](images/a6a8064d4328a4d9d81b3a0b809d883fd6e85641afa8adfd4a24e09893f74165.jpg) +(b) Radar-based model confidence map + +![](images/eebc681ce55accea42f421c6085b93dae0e9ec490e160e96becbdb3706b67342.jpg) +(c) Our confidence map +Fig. 4: Visual comparisons of the detection results for three different scenarios in our proposed dataset. Green boxes in (a) represent the true positives detected by camera-based model. Red boxes/circles indicate false positives and negatives, and white boxes denote ground truth locations. Our model (c) exhibits enhanced detection performance compared to ARC [17] (b), integrating radar and camera data effectively. + +Table 3: Effects of proposed context learning. + +
MethodTraining DatamAP w/ Radar OnlymAP w/ Camera OnlymAP w/ Radar & Camera
CARB-Net-baselineRadar0.55--
CARB-Net wo/ Context learningRadar + Camera0.100.450.75
CARB-Net w/ Mix trainingRadar + Camera0.400.520.75
CARB-Net w/ Context learningRadar + Camera0.540.650.81
+ +is only retrained on radar data without cooperative vision feed-forward signals. The purpose of this model is to show the upper bound of the detection performance when there is only radar data available. The second row model is retrained without context learning. The third row model is retrained by a straightforward solution to consider sensor failure cases during training. We randomly dropout the radar or camera from the input to mix-train the model, which forces the model to learn how to detect in only one modality. + +The model without context learning shows a significant performance drop when the camera-based model cannot detect the targets. In other words, this model is sensitive to the input and is not able to perform detection when one + +Table 4: Effects of proposed components. + +
MethodVision Feed-ForwardImproved U-NetAnchor OffsetsContext LearningmAP
U-Net [34]----0.53
CARB-Net✓---0.65
✓✓--0.70
✓✓✓-0.75
✓✓✓✓0.81
+ +sensor is not working as expected. The third model shows marginally improved results when the one of the sensor is unable to detect, but the detection performance in nominal cases are affected. In contrast, our proposed method shows significant improvements, no matter in nominal or sensor failure cases. + +Effects of Proposed Components To clarify the effect of every proposed component, we conduct several experiments to show the performance difference with or without these components. Tab. 4 demonstrates that all the proposed components contribute to performance improvements. The most notable improvement is brought about by the cooperative vision feed-forward. When it is considered, the model achieves a greater level of performance. The projection of the predicted bounding boxes provided by the off-the-shelf camera-based model not only reduces the training effort but also prevents the distortion problem. The results show that when all the components are incorporated, the performance of the model is considerably enhanced. + +# 5 Conclusion + +In this paper, we introduce Camera-Assisted Radar-Based Network (CARB-Net), a novel radar-based framework aimed at addressing the low angular resolution of radar signals and enhancing robustness to conditions where a sensor is unable to detect the targets. CARB-Net effectively integrates detections from an off-the-shelf camera-based object detection model with radar data, enabling precise localization by learning object micro-Doppler patterns from radar tensors and incorporating vision feed-forward signals. Our context learning method ensures reliable performance even when one sensor fails. The paper contributes significantly by showcasing improved performance for vulnerable road users compared to state-of-the-art methods. Validation on our larger radar-camera dataset confirms these achievements. Additionally, our method demonstrates up to a $12\%$ improvement in mAP performance across diverse datasets, with ablation studies highlighting the efficacy of each framework component. Future research directions could explore optimizing fusion mechanisms and incorporating additional sensor modalities like thermal or acoustic sensors for even more accurate object detection and localization. However, as more sensors are integrated, ensuring robustness in multi-sensor object detection remains critical for real-world applications focused on enhancing road user safety. + +# Acknowledgements + +This work was funded by EU Horizon 2020 ECSEL JU research and innovation programme under grant agreement 876487 (NextPerception), by EU Horizon KDT JU research and innovation programme under grant agreement 101139769 (Distribuse), and by the Flemish Government (AI Research Program). + +# References + +1. Barnes, D., Gadd, M., Murcutt, P., Newman, P., Posner, I.: The oxford radar robotcar dataset: A radar extension to the oxford robotcar dataset. In: ICRA (2020) +2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR (2020) +3. Chen, S., He, W., Ren, J., Jiang, X.: Attention-based dual-stream vision transformer for radar gait recognition. In: ICASSP. IEEE (2022) +4. Chen, X., Ma, H., Wan, J., Li, B., Xia, T.: Multi-view 3d object detection network for autonomous driving. In: CVPR (2017) +5. Dimitrievski, M., Shopovska, I., Van Hamme, D., Veelaert, P., Philips, W.: Weakly supervised deep learning method for vulnerable road user detection in fmcw radar. In: IEEE International Conference on Intelligent Transportation Systems (ITSC). IEEE (2020) +6. Gao, X., Xing, G., Roy, S., Liu, H.: Experiments with mmwave automotive radar test-bed. In: 2019 53rd Asilomar conference on signals, systems, and computers. pp. 1-6. IEEE (2019) +7. Gao, X., Xing, G., Roy, S., Liu, H.: Ramp-cnn: A novel neural network for enhanced automotive radar object recognition. IEEE Sensors Journal (2020) +8. Griffiths, H., Cohen, L., Watts, S., Mokole, E., Baker, C., Wicks, M., Blunt, S.: Radar spectrum engineering and management: Technical and regulatory issues. Proceedings of the IEEE (2014) +9. Gusland, D., Christiansen, J.M., Torvik, B., Fioranelli, F., Gurbuz, S.Z., Ritchie, M.: Open radar initiative: Large scale dataset for benchmarking of micro-doppler recognition algorithms. In: IEEE Radar Conference (RadarConf). pp. 1-6 (2021) +10. Hartley, R., Zisserman, A.: Multiple view geometry in computer vision. Cambridge university press (2003) +1. Hwang, J.J., Kretzschmar, H., Manela, J., Rafferty, S., Armstrong-Crews, N., Chen, T., Anguelov, D.: Cramnet: Camera-radar fusion with ray-constrained cross-attention for robust 3d object detection. In: ECCV (2022) +2. Jocher, G., Chaurasia, A., Qiu, J.: YOLO by Ultralytics (2023), https://github.com/ultralytics/ultralytics +3. Kim, Y., Kim, S., Choi, J.W., Kum, D.: Craft: Camera-radar 3d object detection with spatio-contextual fusion transformer. In: AAAI (2023) +4. Kim, Y., Shin, J., Kim, S., Lee, I.J., Choi, J.W., Kum, D.: Crn: Camera radar net for accurate, robust, efficient 3d perception. In: ICCV (2023) +5. Kopp, J., Kellner, D., Piroli, A., Dietmayer, K.: Tackling clutter in radar data-label generation and detection using pointnet++. arXiv preprint arXiv:2303.09530 (2023) + +16. Lee, W.Y., Dimitrievski, M., Jovanov, L., Philips, W.: Spatio-temporal consistency for semi-supervised learning using 3d radar cubes. In: IEEE Intelligent Vehicles Symposium (IV). IEEE (2021) +17. Lee, W.Y., Jovanov, L., Kumcu, A., Philips, W.: Arc: Automotive radar consistency regularization for semi-supervised learning. IEEE Transactions on Intelligent Vehicles (2023) +18. Lee, W.Y., Jovanov, L., Philips, W.: Multi-view target transformation for pedestrian detection. In: WACV Workshops (2023) +19. Li, Y., Ge, Z., Yu, G., Wang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: AAAI (2023) +20. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: ECCV (2022) +21. Lim, T.Y., Ansari, A., Major, B., Fontijne, D., Hamilton, M., Gowaikar, R., Subramanian, S.: Radar and camera early fusion for vehicle detection in advanced driver assistance systems. In: NIPS Workshops (2019) +22. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: ICCV (2017) +23. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.Y., Berg, A.C.: Ssd: Single shot multibox detector. In: ECCV (2016) +24. Liu, Y., Wang, F., Wang, N., ZHANG, Z.X.: Echoes beyond points: Unleashing the power of raw radar data in multi-modality fusion. NeurIPS (2024) +25. Liwen, Z., Xinyan, Z., Youcheng, Z., Yufei, G., Yuanpei, C., Xuhui, H., Zhe, M.: Peakconv: Learning peak receptive field for radar semantic segmentation. In: CVPR (2023) +26. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: CVPR (2015) +27. Mallot, H.A., Bulthoff, H.H., Little, J.J., Bohrer, S.: Inverse perspective mapping simplifies optical flow computation and obstacle detection. Biological cybernetics (1991) +28. Meyer, M., Kuschk, G.: Automotive radar dataset for deep learning based 3d object detection. In: European Radar Conference (EuRAD) (2019) +29. Ouaknine, A., Newson, A., Pérez, P., Tupin, F., Rebut, J.: Multi-view radar semantic segmentation. In: ICCV (2021) +30. Ouaknine, A., Newson, A., Rebut, J., Tupin, F., Pérez, P.: Carrada dataset: Camera and automotive radar with range- angle- doppler annotations. In: ICPR (2021) +31. Paek, D.H., Kong, S.H., Wijaya, K.T.: K-radar: 4d radar object detection for autonomous driving in various weather conditions. NeurIPS (2022) +32. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 194-210. Springer (2020) +33. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NIPS (2017) +34. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI. Springer (2015) +35. Schumann, O., Hahn, M., Dickmann, J., Wöhler, C.: Semantic segmentation on radar point clouds. In: International Conference on Information Fusion (FUSION) (2018) + +36. Sheeny, M., De Pellegrin, E., Mukherjee, S., Ahrabian, A., Wang, S., Wallace, A.: Radiate: A radar dataset for automotive perception in bad weather. In: ICRA (2021) +37. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: CVPR (2016) +38. Vaizman, Y., Weibel, N., Lanckriet, G.: Context recognition in-the-wild: Unified model for multi-modal sensors and multi-label classification. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies (2018) +39. Wang, W., Tran, D., Feiszli, M.: What makes training multi-modal classification networks hard? In: CVPR (2020) +40. Wang, Y., Wang, G., Hsu, H.M., Liu, H., Hwang, J.N.: Rethinking of radar's role: A camera-radar dataset and systematic annotator via coordinate alignment. In: CVPR (2021) +41. Yao, S., Guan, R., Huang, X., Li, Z., Sha, X., Yue, Y., Lim, E.G., Seo, H., Man, K.L., Zhu, X., Yue, Y.: Radar-camera fusion for object detection and semantic segmentation in autonomous driving: A comprehensive review. IEEE Transactions on Intelligent Vehicles (2023) +42. Zhou, T., Chen, J., Shi, Y., Jiang, K., Yang, M., Yang, D.: Bridging the view disparity between radar and camera features for multi-modal fusion 3d object detection. IEEE Transactions on Intelligent Vehicles (2023) \ No newline at end of file diff --git a/2024/CARB-Net_ Camera-Assisted Radar-Based Network for Vulnerable Road User Detection/images.zip b/2024/CARB-Net_ Camera-Assisted Radar-Based Network for Vulnerable Road User Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b9927db0e6fad396066d3ab787d4d70392f591e1 --- /dev/null +++ b/2024/CARB-Net_ Camera-Assisted Radar-Based Network for Vulnerable Road User Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05e65d9f35221158b3f8b965e776d5bca0296c66d196c04aaa29ad55b7f18de6 +size 309506 diff --git a/2024/CARB-Net_ Camera-Assisted Radar-Based Network for Vulnerable Road User Detection/layout.json b/2024/CARB-Net_ Camera-Assisted Radar-Based Network for Vulnerable Road User Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a3b79008d0e375c6f4f067275b9bba07d622fa22 --- /dev/null +++ b/2024/CARB-Net_ Camera-Assisted Radar-Based Network for Vulnerable Road User Detection/layout.json @@ -0,0 +1,7362 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 111, + 465, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 111, + 465, + 147 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 465, + 147 + ], + "type": "text", + "content": "CARB-Net: Camera-Assisted Radar-Based Network for Vulnerable Road User Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 139, + 168, + 474, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 168, + 474, + 193 + ], + "spans": [ + { + "bbox": [ + 139, + 168, + 474, + 193 + ], + "type": "text", + "content": "Wei-Yu Lee" + }, + { + "bbox": [ + 139, + 168, + 474, + 193 + ], + "type": "inline_equation", + "content": "^{\\text{①}}" + }, + { + "bbox": [ + 139, + 168, + 474, + 193 + ], + "type": "text", + "content": ", Martin Dimitrievski" + }, + { + "bbox": [ + 139, + 168, + 474, + 193 + ], + "type": "inline_equation", + "content": "^{\\text{②}}" + }, + { + "bbox": [ + 139, + 168, + 474, + 193 + ], + "type": "text", + "content": ", David Van Hamme" + }, + { + "bbox": [ + 139, + 168, + 474, + 193 + ], + "type": "inline_equation", + "content": "^{\\text{③}}" + }, + { + "bbox": [ + 139, + 168, + 474, + 193 + ], + "type": "text", + "content": ", Jan Aelterman" + }, + { + "bbox": [ + 139, + 168, + 474, + 193 + ], + "type": "inline_equation", + "content": "^{\\text{④}}" + }, + { + "bbox": [ + 139, + 168, + 474, + 193 + ], + "type": "text", + "content": ", Ljubomir Jovanov" + }, + { + "bbox": [ + 139, + 168, + 474, + 193 + ], + "type": "inline_equation", + "content": "^{\\text{⑤}}" + }, + { + "bbox": [ + 139, + 168, + 474, + 193 + ], + "type": "text", + "content": ", and Wilfried Philips" + }, + { + "bbox": [ + 139, + 168, + 474, + 193 + ], + "type": "inline_equation", + "content": "^{\\text{⑥}}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 199, + 202, + 414, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 202, + 414, + 213 + ], + "spans": [ + { + "bbox": [ + 199, + 202, + 414, + 213 + ], + "type": "text", + "content": "TELIN-IPI, Ghent University-IMEC, Gent, Belgium" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 238, + 462, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 238, + 462, + 447 + ], + "spans": [ + { + "bbox": [ + 160, + 238, + 462, + 447 + ], + "type": "text", + "content": "Abstract. Ensuring a reliable perception of vulnerable road users is crucial for safe autonomous driving. Radar stands out as an appealing sensor choice due to its resilience in adverse weather, cost-effectiveness, depth sensing capabilities, and established role in adaptive cruise control. Nevertheless, radar's limited angular resolution poses challenges in object recognition, especially in distinguishing targets in close proximity. To tackle this limitation, we present the Camera-Assisted Radar-Based Network (CARB-Net), a novel and efficient framework that merges the angular accuracy of a camera with the robustness and depth sensing capabilities of radar. We integrate camera detection information through a ground plane feed-forward array, entangling it with the early stages of a radar-based detection network. Furthermore, we introduce a unique context learning approach to ensure graceful degradation in situations of poor radar Doppler information or unfavorable camera viewing conditions. Experimental validations on public and our proposed datasets, along with benchmark comparisons, showcase CARB-Net's superiority, boasting up to a " + }, + { + "bbox": [ + 160, + 238, + 462, + 447 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 160, + 238, + 462, + 447 + ], + "type": "text", + "content": " improvement in mAP performance. A series of ablation studies further emphasize the efficacy of the CARB-Net architecture. Our proposed dataset is released at https://github.com/weiyulee/RadVRU." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 456, + 453, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 456, + 453, + 468 + ], + "spans": [ + { + "bbox": [ + 160, + 456, + 453, + 468 + ], + "type": "text", + "content": "Keywords: Micro-Doppler Signature " + }, + { + "bbox": [ + 160, + 456, + 453, + 468 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 456, + 453, + 468 + ], + "type": "text", + "content": " Vulnerable Road User Detection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 487, + 230, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 230, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 230, + 499 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 582 + ], + "type": "text", + "content": "Robust Vulnerable Road User (VRU) perception is vital for driving safety. Among the myriad of sensors used, millimeter-wave radar stands out due to its resilience in challenging light and weather conditions, such as rain, fog, and snow, and its depth and Doppler velocity capability. Its cost-effectiveness has also contributed to its widespread adoption, making it a fundamental component in many Advanced Driving Assistance Systems (ADAS) applications." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "However, radar sensors have inherent limitations: Their angular resolution is limited by the size of the antenna array and operating wavelength. Compared to HD cameras, typical radars have two orders of magnitude lower angular resolution [30, 36, 40], making it challenging to distinguish between closely spaced objects. Beyond detection, this low signal resolution makes classifying between object categories even more difficult. In cluttered environments, multipath interference noise further complicates the interpretation of radar data." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 118, + 479, + 182 + ], + "blocks": [ + { + "bbox": [ + 133, + 118, + 479, + 182 + ], + "lines": [ + { + "bbox": [ + 133, + 118, + 479, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 118, + 479, + 182 + ], + "type": "image", + "image_path": "7437572b73d31a9953fd02091fb67d4ed3dee7944e9df342f7a1c1c46dbb247c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 191, + 482, + 225 + ], + "lines": [ + { + "bbox": [ + 130, + 191, + 482, + 225 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 482, + 225 + ], + "type": "text", + "content": "Fig. 1: An example of micro-Doppler patterns. left: The radar signal is projected and overlaid with the synchronized camera image. right: The radar Doppler signal is plotted over time. The micro-Doppler pattern of the cyclist is highlighted by a dotted line." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 243, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 445 + ], + "type": "text", + "content": "In the literature, radar is often used as a means to improve 3D object detection in monocular systems. By fusing imaging information from cameras with radar, researchers have proposed to exploit the strengths of both technologies [13,14,42]. In these prior studies, cameras play a key role in providing rich visual cues that can be used to train effective object detection and classification models. However, in low-light environments, cameras cannot provide the recognition of road users, and detection of targets by only using radar point clouds becomes difficult due to their sparsity. One of the possible alternative cues for VRU detection is using Doppler information from dense radar tensors, alternatively referred to as Radar Frequency (RF) images. This radar data representation provides the unique motion patterns in the time-frequency domain [5,16,17]. For example, in Figure 1, a cyclist exhibits an oscillatory motion of the limbs, which results in a RF signal with sufficient features to classify these road users. These periodically changing motion patterns are commonly referred to as micro-Doppler patterns, providing fundamental features for identifying road users. However, the low azimuth resolution of radar sensors still limits the accuracy of the predicted object positions in these studies." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 448, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 567 + ], + "type": "text", + "content": "Motivated by these challenges, we propose a RF-based, camera-assisted fusion network with two significant novelties that enhance detection accuracy in nominal conditions (i.e., sensors are working as expected), but more importantly, under conditions of sensor failure. A new cooperative fusion approach introduces summarized camera observations into the early stages of a radar-based network, helping in better feature extraction over the Doppler signal in the RF images. Furthermore, we propose a robust way of entangling camera and radar data, helping the network enhance robustness against scenarios where a sensor fails to detect the targets. By applying these two improvements, we achieve detector robustness without the need for excessive training in unforeseen situations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "We focus on the task of Bird's Eye View (BEV) vulnerable road user perception, where we train a radar-based network on radar frequency images, or more specifically, Range-Azimuth-Doppler (RAD) tensors. The RAD encodes the micro-Doppler pattern of targets at individual range-azimuth spatial cells and is used as the core feature for both detection and classification. Camera information is appended to the RAD tensor in the form of a summarized detection array in the BEV plane, which assists the radar-based network when it is available. We use an off-the-shelf camera-based object detection model to produce detec" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "W.Y Lee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 174 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 174 + ], + "type": "text", + "content": "tions of road users and populate the feed-forward array using back-projection. The input layer of the network then applies pixel-shuffle [37] to reorganize the high-resolution camera observations into a common representation, matching the low-resolution radar data. This allows a seamless cooperative fusion with radar tensors without loss of information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 318 + ], + "type": "text", + "content": "Robustness against scenarios where a sensor fails to detect the targets is achieved by augmenting the data and expanding the ground truth labels into sub-categories of possible fusion contexts of nominal and deteriorated sensor performance for each object. The novel aspect of training is to generate and encode these conditions as control variables, that augment the ground truths during training. This strategy encourages the model to learn more sophisticated and finer features for each modality without additional human annotation, enhancing its adaptability to diverse fusion contexts. Our experiments, conducted on both public and our proposed RadVRU datasets, demonstrate the effectiveness of our approach. Our method not only achieves state-of-the-art detection performance but also exhibits a remarkable up to " + }, + { + "bbox": [ + 130, + 176, + 480, + 318 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 130, + 176, + 480, + 318 + ], + "type": "text", + "content": " improvement in challenging sensor failure cases. Our contributions are outlined as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 326, + 480, + 408 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 136, + 326, + 479, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 326, + 479, + 349 + ], + "spans": [ + { + "bbox": [ + 136, + 326, + 479, + 349 + ], + "type": "text", + "content": "1. Novel camera-assisted radar-based network based on pixel-shuffling of high-resolution vision feed-forward and low-resolution micro-Doppler signatures." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 350, + 480, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 350, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 136, + 350, + 480, + 384 + ], + "type": "text", + "content": "2. Novel context learning strategy, utilizing context-specific control variables to train a model that is robust to conditions where a sensor is unable to detect the targets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 384, + 480, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 384, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 135, + 384, + 480, + 408 + ], + "type": "text", + "content": "3. An extensive experimental evaluation on two datasets for BEV detection where we objectively compare to existing state-of-the-art methods." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 426, + 237, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 426, + 237, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 237, + 438 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 449, + 480, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 480, + 581 + ], + "type": "text", + "content": "Radar-Based Object Detection Prior studies utilize radar as the primary sensor for object detection, employing various data representations: point clouds and dense radar tensors. For radar point clouds, the sparsity and lack of context pose detection challenges [33, 35, 41] for neural networks to learn features effectively. On the other hand, image-like radar tensors provide rich but noisy contextual data [8, 15]. Many prior studies utilize deep learning to recognize objects by their motion characteristics [3, 9]. For instance, radar's unique micro-Doppler signature, a time-varying Doppler frequency shift captured in RAD tensors, enables accurate recognition of road users like pedestrians and cyclists [5,6,16,17]. This distinctive motion pattern advantage motivates our use of RAD tensors as inputs, harnessing the micro-Doppler signature for robust object detection." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "type": "text", + "content": "On the other hand, radar sensors, while effective at detecting moving road users, struggle to separate closely packed objects, such as groups of pedestrians, due to their low angular resolution [41]. Additionally, radar sensors fail to detect stationary objects due to interference from other reflections and a lack of motion information. As a result, the existing radar-only methods cannot accurately localize and recognize all objects. In this study, we utilize observations from the camera as assistance to improve the precision of azimuth localization." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 101 + ], + "type": "text", + "content": "Camera-Assisted Radar-Based Network for Vulnerable Road User Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "Radar-Camera-Based Object Detection Recently, using BEV representation to merge images from multiple camera and radar sensors has sparked a rising wave of research in autonomous driving. Prior studies adopt large pretrained image networks, such as LSS [32], BEVFormer [20], or BEVDepth [19], to project 2D image features onto the BEV for further fusion with radar data. For instance, in [13], the authors use proposals from a camera-based network to associate and merge with radar point clouds. In [14], a radar-assisted transformation method is proposed to project 2D images onto BEV and aggregate image and radar feature maps for detection. However, the perspective transformation of camera views highly relies on accurate calibration parameters and good light conditions. Numerous objects visible to cameras may yield only a few radar points, posing significant challenges when cameras are unreliable and sparse point clouds cannot provide sufficient evidence for recognition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 272, + 483, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 483, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 483, + 429 + ], + "type": "text", + "content": "In contrast, our method interfaces with the radar at a much earlier stage, utilizing 3D dense radar tensors as input to learn the micro-Doppler patterns for recognition instead of relying on camera features. Therefore, detection can be performed only on radar data in low-light conditions without cameras. Moreover, existing fusion methods often necessitate training a huge network on a large amount of both camera and radar data, which poses challenges due to the lack of publicly available datasets containing both modalities and the high computing effort. To address this, we utilize detections from an efficient off-the-shelf camera-based detection model, reducing the training and inference burden. Since our method's core is a radar-based detector using micro-Doppler patterns to recognize VRUs, only aided by the camera when it is available, we mainly consider the dataset providing 3D radar tensors for experiments, and the existing BEV radar-camera fusion methods cannot be directly applied to this task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 444, + 483, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 444, + 483, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 483, + 576 + ], + "type": "text", + "content": "Robustness of Multimodal Fusion In practice, it is quite common for sensors to become unreliable in specific contexts. This negatively affects the robustness of the perception system, such as occlusions or low-light environments for cameras. Prior studies introduce sensor dropout [4,11,38,39] during training to enhance the network's ability to handle these conditions with single modality features. However, simply removing the sensor data from the input constrains the fusion potential and usually degrades the overall detection performance. In this work, we introduce a novel context learning strategy that augments the data by creating challenging scenarios that render sensor observations unreliable, and we instruct the network to recognize these situations through auxiliary queries, enhancing model robustness in detection tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 593, + 233, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 593, + 233, + 607 + ], + "spans": [ + { + "bbox": [ + 132, + 593, + 233, + 607 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we introduce Camera-Assisted Radar-Based Network (CARB-Net), a novel neural network model for integrating low spatial resolution radar tensors and high-resolution camera images. As depicted in Fig. 2, the proposed framework operates in the spatio-temporal domain, taking sequential camera" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "W.Y Lee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 482, + 224 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 224 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 224 + ], + "type": "image", + "image_path": "2af3859c66ae4e3d05f041cd77ccd6e1ec928bb0235bff8cc2a1f1c91c7e6128.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 233, + 480, + 257 + ], + "lines": [ + { + "bbox": [ + 130, + 233, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 480, + 257 + ], + "type": "text", + "content": "Fig.2: Overview of the proposed Camera-Assisted Radar-Based Network (CARB-Net)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 284, + 482, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 334 + ], + "type": "text", + "content": "images " + }, + { + "bbox": [ + 130, + 284, + 482, + 334 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{c}} \\in \\mathbb{R}^{NC_{c} \\times H_{c} \\times W_{c}}" + }, + { + "bbox": [ + 130, + 284, + 482, + 334 + ], + "type": "text", + "content": " and sequential RAD radar tensors " + }, + { + "bbox": [ + 130, + 284, + 482, + 334 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{r}} \\in \\mathbb{R}^{NC_{r} \\times H_{r} \\times W_{r}}" + }, + { + "bbox": [ + 130, + 284, + 482, + 334 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 284, + 482, + 334 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 284, + 482, + 334 + ], + "type": "text", + "content": " denotes the number of tensors and images in the sequence, " + }, + { + "bbox": [ + 130, + 284, + 482, + 334 + ], + "type": "inline_equation", + "content": "C_{c}, C_{r}" + }, + { + "bbox": [ + 130, + 284, + 482, + 334 + ], + "type": "text", + "content": " represent channel sizes, and " + }, + { + "bbox": [ + 130, + 284, + 482, + 334 + ], + "type": "inline_equation", + "content": "(H_{c}, W_{c})" + }, + { + "bbox": [ + 130, + 284, + 482, + 334 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 284, + 482, + 334 + ], + "type": "inline_equation", + "content": "(H_{r}, W_{r})" + }, + { + "bbox": [ + 130, + 284, + 482, + 334 + ], + "type": "text", + "content": " are the spatial dimensions of the RAD tensors and camera images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 363, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 508 + ], + "type": "text", + "content": "Unified Data Representation The data representations of radars and cameras are different, which poses challenges in sensor fusion as association requires relating objects in different coordinate systems. Therefore, it is essential to establish a common representation suitable for both fusion and detection. Front View and Bird's Eye View are the two mainstream representations in prior studies. For front view, due to the low resolution in the azimuth angle provided by the radar as well as camera calibration errors, projected radar signals may deviate from the object in the image, which makes the observations of the same object contradictory and hard to be fused by networks. On the other hand, BEV provides a view that accurately portrays the scale and positions of objects, making it a suitable representation for autonomous driving tasks. In this study, we use BEV as the unified data representation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Given the image coordinates " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": ", we rely on the intrinsic camera parameters and extrinsic parameters to perform transformation from camera reference frame into BEV coordinates. We assume the point where the target touches the ground plane (foot or wheel) is with zero height, then the image coordinates " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " can be projected to BEV [10, 27]. However, this assumption can be inaccurate in cases when the camera cannot clearly capture the bottom-most point of an object, the target is jumping in the air, or the ground is not flat. This inaccuracy severely degrades the detection performance when using perspective transformation to project the RGB image to BEV for aligning radar and camera signals. Furthermore, perspective transformation also introduces image distortion because of the zero height assumption. The distortion can severely corrupt objects' visual appearance, such as changing the shape and structure, which causes difficulties for the network to learn from the camera images [18]." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 101 + ], + "type": "text", + "content": "Camera-Assisted Radar-Based Network for Vulnerable Road User Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 216 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 216 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 216 + ], + "type": "image", + "image_path": "544b31def2fdbc60b3f505c05c7dc9422526a27d724156d33dcd979d5bc8e971.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 223, + 482, + 257 + ], + "lines": [ + { + "bbox": [ + 130, + 223, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 482, + 257 + ], + "type": "text", + "content": "Fig.3: A camera image with predicted bounding boxes and the cooperative vision feed-forward image in BEV. Blue and green colors represent the pedestrian and bicycle classes. Yellow arrows indicate the mapping after perspective transformation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 278, + 479, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 479, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 479, + 350 + ], + "type": "text", + "content": "Cooperative Vision Feed-Forward In order to address the aforementioned problems, we propose Cooperative Vision Feed-Forward to merge radar data uncertain in azimuth but certain in range with RGB data uncertain in range but certain in azimuth. Despite the uncertainties of radar and camera observations, there is usually a small region on the BEV plane where the radar and camera observations exhibit a strong response, which we use for road user detection." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 350, + 482, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 458 + ], + "type": "text", + "content": "As depicted in Figure 3, we introduce the detections made by using camera data in the form of a vision feed-forward array on the BEV plane. Specifically, we utilize the off-the-shelf camera-based detection network to produce 2D detections of road users based on the RGB images " + }, + { + "bbox": [ + 130, + 350, + 482, + 458 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{c}}" + }, + { + "bbox": [ + 130, + 350, + 482, + 458 + ], + "type": "text", + "content": ", then we project the center of the predicted bounding boxes of the road users to the BEV to make vision feedforward images " + }, + { + "bbox": [ + 130, + 350, + 482, + 458 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{f}}" + }, + { + "bbox": [ + 130, + 350, + 482, + 458 + ], + "type": "text", + "content": ". The projected values on the BEV are the confidence scores, and these projected centers can be seen as summarized information from the pretrained network using RGB images. By doing this, little to no spatial distortion will be caused by the back-projection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 459, + 482, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 482, + 626 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 482, + 626 + ], + "type": "text", + "content": "In addition to the projection based on the zero height assumption, we estimate the depth of a target from its width and height in the image frame and its average size (i.e., averaged human heights and shoulder widths) in the real-world to estimate robust depth values. As a result, we use the median value of these estimations as the depth value. To compensate for the projection errors caused by inaccurate perspective transformation, we enlarge the projected positions from points to regions by applying a Gaussian kernel to cover the uncertainty of the depth estimation. Specifically, we never know the actual heights and widths of pedestrians in practice, but instead we use a statistical model of the mean height and width of humans, " + }, + { + "bbox": [ + 130, + 459, + 482, + 626 + ], + "type": "inline_equation", + "content": "1.2\\mathrm{m}" + }, + { + "bbox": [ + 130, + 459, + 482, + 626 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 459, + 482, + 626 + ], + "type": "inline_equation", + "content": "2.0\\mathrm{m}" + }, + { + "bbox": [ + 130, + 459, + 482, + 626 + ], + "type": "text", + "content": " height and " + }, + { + "bbox": [ + 130, + 459, + 482, + 626 + ], + "type": "inline_equation", + "content": "0.36\\mathrm{m}" + }, + { + "bbox": [ + 130, + 459, + 482, + 626 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 459, + 482, + 626 + ], + "type": "inline_equation", + "content": "0.44\\mathrm{m}" + }, + { + "bbox": [ + 130, + 459, + 482, + 626 + ], + "type": "text", + "content": " width, respectively. That means the projection ends up on a segment of a perspective line of " + }, + { + "bbox": [ + 130, + 459, + 482, + 626 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 459, + 482, + 626 + ], + "type": "text", + "content": " meters, which becomes the sigma of our Gaussian kernel. As long as the sigma is not underestimated, the result will not critically depend on its value. Therefore, the inaccuracy problem caused by perspective transformation can be alleviated." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Spatial Resolution Alignment Efficiently aligning low resolution radar data and high resolution camera signals on the ground plane is a challenge. A straight-" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "W.Y Lee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "content": "forward solution for aligning is to use a BEV space with the same low spatial resolution as the radar sensor [21]. However, the low spatial resolution of the radar causes this approach to discard high-resolution camera information, negatively affecting the fusion outcomes. On the other hand, upscaling low-resolution radar to align with high-resolution cameras is also not practical because of the high computing effort. Therefore, we introduce Pixel-shuffle [37] Downsampling (PD) to tackle this resolution misalignment problem. PD is to reorganize the high-resolution image data into several low resolution channels. This operation only rearranges the order of the elements to down-scale the image instead of encoding or pooling. Hence, all the information of the image can be preserved, and the network's time and space complexity will not be increased. Specifically, we make vision feed-forward images " + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{f}} \\in \\mathbb{R}^{N \\times H_f \\times W_f}" + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "inline_equation", + "content": "(H_f \\times W_f)" + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "content": " is an intermediate spatial resolution between RGB images and radar cubes. In this instance, we apply PD on the vision feed-forward images " + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{f}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{f}}' \\in \\mathbb{R}^{ND^2 \\times H_r \\times W_r}" + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "content": " is the down-scaling factor, such that the subsequent network may jointly extract features from the radar tensors and vision feed-forward images with varying resolutions using convolution kernels. The solution has two main advantages: (1) no additional network training effort on camera images is needed, and (2) resolution alignment without information loss." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 381, + 482, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 482, + 490 + ], + "type": "text", + "content": "Camera-Assisted Radar-Based Network The aim of the proposed network is to identify road users by using sequential RAD tensors " + }, + { + "bbox": [ + 130, + 381, + 482, + 490 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{r}}" + }, + { + "bbox": [ + 130, + 381, + 482, + 490 + ], + "type": "text", + "content": " and vision feedforward images " + }, + { + "bbox": [ + 130, + 381, + 482, + 490 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{f}}" + }, + { + "bbox": [ + 130, + 381, + 482, + 490 + ], + "type": "text", + "content": ". The RAD tensors capture micro-Doppler patterns and precise range measurements but have limited azimuth resolution. Our network interprets these tensors to pinpoint and classify various target categories. Conversely, the vision feed-forward images offer detailed azimuth localization of VRUs from a pre-trained camera-based network but lack accuracy in range estimation. Despite the inherent uncertainties, there is a tiny region on the BEV plane where all sensors exhibit a strong response, which we use for road user detection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 493, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 493, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 493, + 482, + 615 + ], + "type": "text", + "content": "Our CARB-Net is developed based on the principles of multi-resolution feature decomposition in U-Net [34] with several improvements. First, we use a small convolution network, which we call the Fusion Encoder, to encode the RF information from the radar and vision feed-forward images into a compact channel-representation as the input for U-Net. In addition, we apply the concept of pixel-shuffle [37] again to replace the max-pooling and up-scaling modules inside U-Net. The feature maps are easily up- or down-scaled by reorganizing the image channels without adding additional model parameters. Compared to deconvolution, max-pooling, or bilinear interpolation, pixel-shuffle is used as an effective way to prevent resolution loss during the encoding and decoding steps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "To obtain high-resolution output, we apply ESPCN [37] and pixel-shuffle up-sampling by factor " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " to reorganize the channels of the network output. Ultimately, the spatial resolution of the output tensor becomes " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "(H_{f} \\times W_{f})" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ". With this architecture, we allow the networks to process input with different resolutions." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 101 + ], + "type": "text", + "content": "Camera-Assisted Radar-Based Network for Vulnerable Road User Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "Context Learning Radars or cameras might not be able to provide recognition of road users in specific contexts, which is unacceptable for real-world deployment. For instance, a stationary object is poorly visible to radars, and adverse weather conditions may make objects hard-to-spot by cameras. However, most existing methods do not focus on such cases during training but only emphasize average detection performance, causing significant performance drops under such contexts compared to favorable circumstances." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 203, + 483, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 203, + 483, + 287 + ], + "spans": [ + { + "bbox": [ + 130, + 203, + 483, + 287 + ], + "type": "text", + "content": "In order to improve robustness in these specific contexts, we make the network answer 2 additional auxiliary questions to distinguish the sensing capability of each sensor. For each position in the output confidence map, the network has to determine: (1) is there a VRU or not; (2) can a VRU be detected by the radar; and (3) can a VRU be detected by the camera. During training, we apply data augmentation and increase the variability of available labeled data in order to cover more combinations of the possible answers to these questions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 289, + 483, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 483, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 483, + 445 + ], + "type": "text", + "content": "We define " + }, + { + "bbox": [ + 130, + 289, + 483, + 445 + ], + "type": "inline_equation", + "content": "y \\in \\{y_0, y_1, y_2\\}" + }, + { + "bbox": [ + 130, + 289, + 483, + 445 + ], + "type": "text", + "content": " to represent the 3 correct answers to the 3 questions. Specifically, " + }, + { + "bbox": [ + 130, + 289, + 483, + 445 + ], + "type": "inline_equation", + "content": "y_0" + }, + { + "bbox": [ + 130, + 289, + 483, + 445 + ], + "type": "text", + "content": " indicates whether this position is occupied by a road user. The network should use all the sensor data as input to detect VRUs. " + }, + { + "bbox": [ + 130, + 289, + 483, + 445 + ], + "type": "inline_equation", + "content": "y_1" + }, + { + "bbox": [ + 130, + 289, + 483, + 445 + ], + "type": "text", + "content": " describes if the road user's micro-Doppler pattern is detectable by the radar sensor. In this case, we focus on this specific context that makes the targets poorly visible to radar. Since the targets are stationary or moving without micro-Doppler patterns, the network cannot distinguish their radar signals from other obstacles' reflections. Finally, " + }, + { + "bbox": [ + 130, + 289, + 483, + 445 + ], + "type": "inline_equation", + "content": "y_2" + }, + { + "bbox": [ + 130, + 289, + 483, + 445 + ], + "type": "text", + "content": " indicates whether the pre-trained camera-based network can detect the target or not. In this case, we focus on the conditions that make the camera sensor unable to see the target. By training the network to answer these 3 questions, we use the network's first answer as the final fusion decision to distinguish if the road user is present, and we use the others to describe the specific fusion contexts for each target." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 447, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 567 + ], + "type": "text", + "content": "For instance, " + }, + { + "bbox": [ + 130, + 447, + 482, + 567 + ], + "type": "inline_equation", + "content": "[y_0, y_1, y_2] = [1, 1, 1]" + }, + { + "bbox": [ + 130, + 447, + 482, + 567 + ], + "type": "text", + "content": " is the common case that both radar and camera can sense the targets, and " + }, + { + "bbox": [ + 130, + 447, + 482, + 567 + ], + "type": "inline_equation", + "content": "[y_0, y_1, y_2] = [1, 0, 1]" + }, + { + "bbox": [ + 130, + 447, + 482, + 567 + ], + "type": "text", + "content": " means the road user is only detectable by the camera. This way, our network learns from the correct answers to distinguish these circumstances and understand the scene in more detail. During training, this approach additionally allows us to consider and penalize each class (e.g., pedestrian, cyclist, etc.) independently. This is achieved by tasking the network to perform this classification as well, thus penalizing solutions that fail to interpret their input data or context. That is, for each class, the network is required to output the 3 answers to the 3 questions, which are independent of the other classes' answers." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "In order to teach the network to distinguish these specific contexts, we control the input data to generate the corresponding contexts for training. For radar, the ideal way to simulate targets moving without micro-Doppler patterns or stationary targets is to remove their micro-Doppler patterns and retain only the mean velocities. However, this solution is not practical because the radar signals usually contain reflections from other objects and clutter. Moreover, micro-Doppler information from one object can spill over into nearby cells, depending on scene geometry. It is challenging to only remove a certain target's motion information" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "W.Y Lee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": "while not affecting the others. Instead, we propose an alternative way to simulate such cases: the latest RAD tensor is repeated inside the sequential RAD tensors " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{r}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": " to replace the other " + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 130, + 116, + 482, + 332 + ], + "type": "text", + "content": " radar tensors. That is, all the objects in the scene are preserved at their mean velocities, and most of their micro-Doppler patterns are eliminated. So, the objects' motion inside the repeated tensors is seen as an extension of the current movement without introducing any periodic motion. However, there are several downsides to this solution. First, if the target is moving very fast, the radar can still capture its micro-Doppler patterns in a single radar cube. Secondly, repeating the latest radar cube results in radar data where objects are static in RA plane but have non-zero velocity, a situation that is impossible in reality. However, this data augmentation is a close enough simulation of the most important radar limitations: its weakness in detecting slow motion and people with low micro-Doppler variation amplitude. Furthermore, it is easy to implement, and we do not need any additional computing effort to carefully remove the micro-Doppler patterns. For RGB cameras, we switch on and off the cooperative vision feed-forward to simulate the contexts in which the camera sensor cannot detect the target. In this case, the network should only use radar signals to recognize road users." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 349, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 482, + 553 + ], + "type": "text", + "content": "Anchor Points and Offsets In the output layer, we incorporate anchor points and offsets to precisely refine VRU positions at any resolution. This technique increases the achievable resolution without increasing the data or grid dimensions by using an additional channel of output data to encode highly precise spatial offsets to a target from a low-dimensional grid, low-precision position. We predefine our anchor points over a uniformly spaced range and azimuth positions to allow for an equal capacity for detecting objects at all distances. Since there is a low variation in the size of the VRUs in BEV, we refrain from predicting object width and height. Specifically, we utilize range and azimuth bins from the radar configuration to pre-define the anchor points, representing certain positions on the ground plane. The network predicts 2 offsets for two directions relative to the default anchor point. To ensure that each VRU only be matched by a subset of anchor points, we restrict the offsets to being only inside the 4 surrounding anchor points. Finally, we perform non-maximum suppression to merge nearby anchor points using a greedy approach that favors the highest confidence scores within a pre-defined radius. This is a standard methodology employed in image object detection, which we found also works well in BEV detection using radar." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Loss Functions For a certain position " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " in the output tensor, it contains " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "(3k + 2)" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " different output values: 2 offsets and " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "3k" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " classes (original " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " different categories from ground truths and 3 answers for context learning), which results in a total of " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "(3k + 2)H_{f}W_{f}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " outputs. We define " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "O_{c,y}(\\mathbf{r})" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " that refers to the probability of category " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " with answer " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "c \\in \\{c_0, c_1, \\dots, c_k\\}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "y \\in \\{y_0, y_1, y_2\\}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ". In order to reduce the inherent class imbalance between the road users and background in the sparsely populated training BEV maps, we apply Focal loss [22] to compute the classification loss " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cls}}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 101 + ], + "type": "text", + "content": "Camera-Assisted Radar-Based Network for Vulnerable Road User Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 137, + 489, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 489, + 174 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 489, + 174 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c l s}} (\\mathbf {r}) = - \\sum_ {c} \\sum_ {y} \\alpha_ {c} \\left[ (1 - O _ {c, y} (\\mathbf {r})) ^ {\\gamma} \\log \\left(O _ {c, y} (\\mathbf {r})\\right) + O _ {c, y} (\\mathbf {r}) ^ {\\gamma} (1 - y) \\log \\left(1 - O _ {c, y} (\\mathbf {r})\\right) \\right], \\tag {1}", + "image_path": "926045149a0e569a31590568de642a55ffd1d625588913e6615850362887b8a2.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 175, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 482, + 224 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 175, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\alpha_{c}" + }, + { + "bbox": [ + 130, + 175, + 482, + 224 + ], + "type": "text", + "content": " represents the weighting factor for class " + }, + { + "bbox": [ + 130, + 175, + 482, + 224 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 175, + 482, + 224 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 175, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\gamma \\geq 0" + }, + { + "bbox": [ + 130, + 175, + 482, + 224 + ], + "type": "text", + "content": " is the hyperparameter to down-weigh easy samples. We use smooth L1-norm as our localization loss " + }, + { + "bbox": [ + 130, + 175, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{loc}}" + }, + { + "bbox": [ + 130, + 175, + 482, + 224 + ], + "type": "text", + "content": ". This is similar to the bounding box regression in SSD [23] but we remove the multiple boxes and only predict the offsets for every anchor point." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 246, + 224, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 224, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 224, + 258 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 274, + 349, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 349, + 286 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 349, + 286 + ], + "type": "text", + "content": "4.1 Datasets and Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 299, + 482, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 299, + 482, + 383 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 482, + 383 + ], + "type": "text", + "content": "In most public datasets, data representations of radar signals are in the form of a processed point cloud containing targets [2] or projected RD or RA maps [1,28, 31,36,40]. Since our method is based on the micro-Doppler signature to recognize road users, RAD tensors with synchronized camera images are necessary. Hence, we identified the publicly available CARRADA dataset [30] to be best suited for conducting experimental evaluation. In addition, we present results from experimentation using our proposed RadVRU dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 403, + 482, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 500 + ], + "type": "text", + "content": "CARRADA Dataset The CARRADA dataset [30] includes RAD tensors, RD, RA, and AD maps, all synchronized with camera images. Annotations comprise sparse points, dense masks, and bounding boxes for object detection, tracking, and radar semantic segmentation. Mounted on a stationary vehicle, the hardware captures data in a controlled environment. This dataset encompasses 12k annotated frames across three object categories: pedestrians, cyclists, and vehicles. Each RAD tensor has dimensions of " + }, + { + "bbox": [ + 130, + 403, + 482, + 500 + ], + "type": "inline_equation", + "content": "256 \\times 256 \\times 64" + }, + { + "bbox": [ + 130, + 403, + 482, + 500 + ], + "type": "text", + "content": ", and the accompanying camera images boast a resolution of " + }, + { + "bbox": [ + 130, + 403, + 482, + 500 + ], + "type": "inline_equation", + "content": "1232 \\times 1028" + }, + { + "bbox": [ + 130, + 403, + 482, + 500 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Proposed RadVRU Dataset Our RadVRU dataset is captured in several cities in Western Europe. The acquisition hardware consists of a 77 GHz FMCW radar (TI AWR1243), a full HD RGB camera (Realsense D435), and a 3D LiDAR (Ouster OS1-128) mounted on the top of a vehicle. This dataset contains total 78k frames, including 317 sequences of duration from " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "10\\mathrm{~s}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "20\\mathrm{~s}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ". These frames are representative of the challenging driving scenarios, including different lighting conditions, complex traffic with cars, pedestrians, and cyclists. The annotations are performed by using a camera- and a LiDAR-based object detection model to pre-label the instances, and then these labels are manually corrected by human annotators. The pedestrians are annotated as single points on the ground plane. The dimensions of range, azimuth, and Doppler of every RAD tensor are " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "128 \\times 16 \\times 128" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", and the resolution of the camera images is " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "1920 \\times 1080" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "W.Y Lee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 152, + 135, + 459, + 247 + ], + "blocks": [ + { + "bbox": [ + 168, + 114, + 444, + 126 + ], + "lines": [ + { + "bbox": [ + 168, + 114, + 444, + 126 + ], + "spans": [ + { + "bbox": [ + 168, + 114, + 444, + 126 + ], + "type": "text", + "content": "Table 1: Comparisons with other methods on CARRADA dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 152, + 135, + 459, + 247 + ], + "lines": [ + { + "bbox": [ + 152, + 135, + 459, + 247 + ], + "spans": [ + { + "bbox": [ + 152, + 135, + 459, + 247 + ], + "type": "table", + "html": "
MethodTraining DataAPmAPmean Dist. Error
Ped.Cycll.Car
TMVA-Net [29]Radar0.440.430.370.370.67
PeakConv [25]Radar0.450.430.380.380.67
ARC [17]Radar0.620.570.500.540.67
RAMP-CNN [7]Radar0.750.640.500.610.64
CARB-Net-baselineRadar0.770.650.530.630.65
U-Net [34]Radar + Camera0.570.480.470.510.70
FCN [26]Radar + Camera0.570.430.470.460.69
Lim et al. [21]Radar + Camera0.610.530.510.530.66
CARB-NetRadar + Camera0.820.750.600.730.63
", + "image_path": "a1430ac3776c1cffe0ce59b732bdcdc728aacf7d543b1fbf08ad3007e1e2bcd1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 270, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 450 + ], + "type": "text", + "content": "Implementation Details We use the data from CARRADA as an example to explain our implementation details. First, we down-sample the original radar tensor to " + }, + { + "bbox": [ + 130, + 270, + 482, + 450 + ], + "type": "inline_equation", + "content": "128 \\times 128 \\times 64" + }, + { + "bbox": [ + 130, + 270, + 482, + 450 + ], + "type": "text", + "content": " to reduce the high computing effort. Similar to [29], we use 5 sequential radar tensors and camera images as input, so the sizes become " + }, + { + "bbox": [ + 130, + 270, + 482, + 450 + ], + "type": "inline_equation", + "content": "[NC_{r}, H_{r}, W_{r}] = [512, 128, 128]" + }, + { + "bbox": [ + 130, + 270, + 482, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 270, + 482, + 450 + ], + "type": "inline_equation", + "content": "[NC_{c}, H_{c}, W_{c}] = [15, 1232, 1028]" + }, + { + "bbox": [ + 130, + 270, + 482, + 450 + ], + "type": "text", + "content": ". The cooperative vision feed-forward is made by a pre-trained \"yolov8m\" [12] model. We filter the detections from YOLOv8 with a confidence threshold of 0.1, and project the detections to the size of vision feed-forward images " + }, + { + "bbox": [ + 130, + 270, + 482, + 450 + ], + "type": "inline_equation", + "content": "[N, H_{f}, W_{f}] = [5, 256, 256]" + }, + { + "bbox": [ + 130, + 270, + 482, + 450 + ], + "type": "text", + "content": ". After pixel-shuffle down-sampling by factor " + }, + { + "bbox": [ + 130, + 270, + 482, + 450 + ], + "type": "inline_equation", + "content": "D = 2" + }, + { + "bbox": [ + 130, + 270, + 482, + 450 + ], + "type": "text", + "content": ", it becomes " + }, + { + "bbox": [ + 130, + 270, + 482, + 450 + ], + "type": "inline_equation", + "content": "[ND^{2}, H_{r}, W_{r}] = [20, 128, 128]" + }, + { + "bbox": [ + 130, + 270, + 482, + 450 + ], + "type": "text", + "content": ". Similarly, after the ESPCN [37], we apply pixel-shuffle up-sampling by factor " + }, + { + "bbox": [ + 130, + 270, + 482, + 450 + ], + "type": "inline_equation", + "content": "D = 2" + }, + { + "bbox": [ + 130, + 270, + 482, + 450 + ], + "type": "text", + "content": " to restore the resolution [256, 256]. For the proposed anchor offsets and context learning, we have " + }, + { + "bbox": [ + 130, + 270, + 482, + 450 + ], + "type": "inline_equation", + "content": "(3k + 2) = 11" + }, + { + "bbox": [ + 130, + 270, + 482, + 450 + ], + "type": "text", + "content": " predictions. Hence, the size of the final output tensor is [11, 256, 256]. We use 0.8 meter as the radius of our non-maximum suppression algorithm. Please refer to the Supplementary Materials for more implementation details and network architecture." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 468, + 261, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 261, + 480 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 261, + 480 + ], + "type": "text", + "content": "4.2 Experiment Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 490, + 481, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 481, + 575 + ], + "type": "text", + "content": "Evaluation Metric Our experiments employ Average Precision (AP) and mean distance error as primary evaluation metrics. We define true positives as detections falling within a 1.5-meter boundary around the ground truth. We avoid multiple matches within the same boundary. Additionally, we measure localization precision by calculating the distance between true positives and ground truths in meters. Averaging these distance errors yields the mean distance error, providing an evaluation of localization accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Evaluation on CARRADA Dataset We compare the object detection performance of our proposed method and the other state-of-the-art approaches on the CARRADA dataset. Since some of the previous methods output semantic segmentation masks instead of detections, we carefully find the centers of the predicted masks as their localization results. For FCN [26] and U-Net [34] baselines, we use inverse projection mapping to project the RGB images to BEV" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "text", + "content": "Camera-Assisted Radar-Based Network for Vulnerable Road User Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 168, + 136, + 442, + 219 + ], + "blocks": [ + { + "bbox": [ + 146, + 114, + 465, + 125 + ], + "lines": [ + { + "bbox": [ + 146, + 114, + 465, + 125 + ], + "spans": [ + { + "bbox": [ + 146, + 114, + 465, + 125 + ], + "type": "text", + "content": "Table 2: Comparisons with other methods on our proposed RadVRU dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 168, + 136, + 442, + 219 + ], + "lines": [ + { + "bbox": [ + 168, + 136, + 442, + 219 + ], + "spans": [ + { + "bbox": [ + 168, + 136, + 442, + 219 + ], + "type": "table", + "html": "
MethodTraining DatamAPmean Dist. Error
ARC [17]Radar0.510.64
CARB-Net-baselineRadar0.550.62
U-Net [34]Radar + Camera0.520.87
FCN [26]Radar + Camera0.350.94
Lim et al. [21]Radar + Camera0.670.63
EchoFusion* [24]Radar + Camera0.690.54
CARB-NetRadar + Camera0.810.47
", + "image_path": "216e355a1e9d82e068b161c067667dd87fbca1053ee7b5632d93103c60cb47ec.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 239, + 482, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 395 + ], + "type": "text", + "content": "down to a cell resolution of " + }, + { + "bbox": [ + 130, + 239, + 482, + 395 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 130, + 239, + 482, + 395 + ], + "type": "text", + "content": ". Then, we concatenate the projected images and radar tensors along the channel and output the BEV map. For fair comparisons, we ensure the resolutions of the radar data and projected image are the same for all methods. Furthermore, we remove the cooperative vision feed-forward and train our CARB-Net-baseline model only on the radar tensors to compare with the radar-based methods. Although our method is designed for VRUs, we still involve cars for training and testing. We use the cars' center points in BEV space as the ground truth for detection. The vehicles' orientations are not considered during evaluation by all the methods. As observed in Tab. 1, compared to the others, our proposed framework can not only clearly improve the detection performance but also achieve a smaller mean distance error on the RA plane. The results demonstrate that the proposed method successfully integrates the detections from camera images to achieve better accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 411, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 483 + ], + "type": "text", + "content": "Evaluation on Proposed RadVRU Dataset. Similarly, we conduct the same experiment on our RadVRU dataset. We evaluate our framework against the other four different models. For EchoFusion* [24], we use RA tensors as their input and re-implement their method with reduced network size due to our limited computing hardware. As illustrated in Tab. 2, our proposed approach also outperforms the other methods by a significant margin." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 594 + ], + "type": "text", + "content": "Visualization In Fig. 4, we present camera images with bounding boxes predicted by YOLOv8 alongside the radar-based model ARC's [17] and our method's detections in BEV. Notably, our detections outperform ARC's, exhibiting fewer false positives and negatives, leading to superior detection performance. Even in low-light conditions, where YOLOv8 fails, our model successfully identifies pedestrians using radar data alone. Additionally, when road users are stationary, ARC fails due to the absence of micro-Doppler patterns, whereas our method utilizes the vision feed-forward to accurately identify pedestrians." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 611, + 239, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 611, + 239, + 623 + ], + "spans": [ + { + "bbox": [ + 131, + 611, + 239, + 623 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Effects of Proposed Context Learning To further investigate the impact of the proposed context learning, we conduct several experiments to show the robustness of our model when one of the sensors fail. In Tab. 3, the first row model" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "W.Y Lee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 310, + 351 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 310, + 351 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 310, + 351 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 310, + 351 + ], + "type": "image", + "image_path": "2217ae1c1395838775ae22e640466c5183b987dd86da666175ea46ade1375742.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 353, + 266, + 360 + ], + "lines": [ + { + "bbox": [ + 176, + 353, + 266, + 360 + ], + "spans": [ + { + "bbox": [ + 176, + 353, + 266, + 360 + ], + "type": "text", + "content": "(a) Camera-based model detections" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 313, + 112, + 395, + 193 + ], + "blocks": [ + { + "bbox": [ + 313, + 112, + 395, + 193 + ], + "lines": [ + { + "bbox": [ + 313, + 112, + 395, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 112, + 395, + 193 + ], + "type": "image", + "image_path": "1f2913b846689011fb5338d7e1f1b9ad9e1146e8ed0900a106dcb4f0ee16121b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 399, + 112, + 477, + 193 + ], + "blocks": [ + { + "bbox": [ + 399, + 112, + 477, + 193 + ], + "lines": [ + { + "bbox": [ + 399, + 112, + 477, + 193 + ], + "spans": [ + { + "bbox": [ + 399, + 112, + 477, + 193 + ], + "type": "image", + "image_path": "7554deb7af81caef1e80d93ea79d5a08702a0aedd1ff142e199d193e82dfa73d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 314, + 194, + 395, + 272 + ], + "blocks": [ + { + "bbox": [ + 314, + 194, + 395, + 272 + ], + "lines": [ + { + "bbox": [ + 314, + 194, + 395, + 272 + ], + "spans": [ + { + "bbox": [ + 314, + 194, + 395, + 272 + ], + "type": "image", + "image_path": "64838ba3caf42518a0a4719efc9fcc9ec90b5d858b691af12269ef60b4d3bd38.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 399, + 194, + 477, + 272 + ], + "blocks": [ + { + "bbox": [ + 399, + 194, + 477, + 272 + ], + "lines": [ + { + "bbox": [ + 399, + 194, + 477, + 272 + ], + "spans": [ + { + "bbox": [ + 399, + 194, + 477, + 272 + ], + "type": "image", + "image_path": "545fb2f719a5e9b0f6db7c277e0a8af0da5f2684c4557d61d649791283632275.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 314, + 272, + 395, + 351 + ], + "blocks": [ + { + "bbox": [ + 314, + 272, + 395, + 351 + ], + "lines": [ + { + "bbox": [ + 314, + 272, + 395, + 351 + ], + "spans": [ + { + "bbox": [ + 314, + 272, + 395, + 351 + ], + "type": "image", + "image_path": "a6a8064d4328a4d9d81b3a0b809d883fd6e85641afa8adfd4a24e09893f74165.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 353, + 405, + 361 + ], + "lines": [ + { + "bbox": [ + 306, + 353, + 405, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 353, + 405, + 361 + ], + "type": "text", + "content": "(b) Radar-based model confidence map" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 399, + 272, + 477, + 351 + ], + "blocks": [ + { + "bbox": [ + 399, + 272, + 477, + 351 + ], + "lines": [ + { + "bbox": [ + 399, + 272, + 477, + 351 + ], + "spans": [ + { + "bbox": [ + 399, + 272, + 477, + 351 + ], + "type": "image", + "image_path": "eebc681ce55accea42f421c6085b93dae0e9ec490e160e96becbdb3706b67342.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 353, + 472, + 361 + ], + "lines": [ + { + "bbox": [ + 410, + 353, + 472, + 361 + ], + "spans": [ + { + "bbox": [ + 410, + 353, + 472, + 361 + ], + "type": "text", + "content": "(c) Our confidence map" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 371, + 482, + 428 + ], + "lines": [ + { + "bbox": [ + 130, + 371, + 482, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 482, + 428 + ], + "type": "text", + "content": "Fig. 4: Visual comparisons of the detection results for three different scenarios in our proposed dataset. Green boxes in (a) represent the true positives detected by camera-based model. Red boxes/circles indicate false positives and negatives, and white boxes denote ground truth locations. Our model (c) exhibits enhanced detection performance compared to ARC [17] (b), integrating radar and camera data effectively." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 133, + 463, + 478, + 519 + ], + "blocks": [ + { + "bbox": [ + 211, + 442, + 402, + 453 + ], + "lines": [ + { + "bbox": [ + 211, + 442, + 402, + 453 + ], + "spans": [ + { + "bbox": [ + 211, + 442, + 402, + 453 + ], + "type": "text", + "content": "Table 3: Effects of proposed context learning." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 463, + 478, + 519 + ], + "lines": [ + { + "bbox": [ + 133, + 463, + 478, + 519 + ], + "spans": [ + { + "bbox": [ + 133, + 463, + 478, + 519 + ], + "type": "table", + "html": "
MethodTraining DatamAP w/ Radar OnlymAP w/ Camera OnlymAP w/ Radar & Camera
CARB-Net-baselineRadar0.55--
CARB-Net wo/ Context learningRadar + Camera0.100.450.75
CARB-Net w/ Mix trainingRadar + Camera0.400.520.75
CARB-Net w/ Context learningRadar + Camera0.540.650.81
", + "image_path": "5b28f9658d2e010068a31711f5948ba2f15f767dc958491859ca1672bd04e3be.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 544, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 628 + ], + "type": "text", + "content": "is only retrained on radar data without cooperative vision feed-forward signals. The purpose of this model is to show the upper bound of the detection performance when there is only radar data available. The second row model is retrained without context learning. The third row model is retrained by a straightforward solution to consider sensor failure cases during training. We randomly dropout the radar or camera from the input to mix-train the model, which forces the model to learn how to detect in only one modality." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "The model without context learning shows a significant performance drop when the camera-based model cannot detect the targets. In other words, this model is sensitive to the input and is not able to perform detection when one" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 101 + ], + "type": "text", + "content": "Camera-Assisted Radar-Based Network for Vulnerable Road User Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 165, + 135, + 447, + 211 + ], + "blocks": [ + { + "bbox": [ + 220, + 114, + 394, + 126 + ], + "lines": [ + { + "bbox": [ + 220, + 114, + 394, + 126 + ], + "spans": [ + { + "bbox": [ + 220, + 114, + 394, + 126 + ], + "type": "text", + "content": "Table 4: Effects of proposed components." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 165, + 135, + 447, + 211 + ], + "lines": [ + { + "bbox": [ + 165, + 135, + 447, + 211 + ], + "spans": [ + { + "bbox": [ + 165, + 135, + 447, + 211 + ], + "type": "table", + "html": "
MethodVision Feed-ForwardImproved U-NetAnchor OffsetsContext LearningmAP
U-Net [34]----0.53
CARB-Net✓---0.65
✓✓--0.70
✓✓✓-0.75
✓✓✓✓0.81
", + "image_path": "f73eaab95bf7810c4ee0fbeae6264c72b05cbbca16ef5755bb7c383e96131a08.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 232, + 481, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 232, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 481, + 281 + ], + "type": "text", + "content": "sensor is not working as expected. The third model shows marginally improved results when the one of the sensor is unable to detect, but the detection performance in nominal cases are affected. In contrast, our proposed method shows significant improvements, no matter in nominal or sensor failure cases." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 288, + 483, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 288, + 483, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 483, + 407 + ], + "type": "text", + "content": "Effects of Proposed Components To clarify the effect of every proposed component, we conduct several experiments to show the performance difference with or without these components. Tab. 4 demonstrates that all the proposed components contribute to performance improvements. The most notable improvement is brought about by the cooperative vision feed-forward. When it is considered, the model achieves a greater level of performance. The projection of the predicted bounding boxes provided by the off-the-shelf camera-based model not only reduces the training effort but also prevents the distortion problem. The results show that when all the components are incorporated, the performance of the model is considerably enhanced." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 425, + 220, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 220, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 220, + 438 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we introduce Camera-Assisted Radar-Based Network (CARB-Net), a novel radar-based framework aimed at addressing the low angular resolution of radar signals and enhancing robustness to conditions where a sensor is unable to detect the targets. CARB-Net effectively integrates detections from an off-the-shelf camera-based object detection model with radar data, enabling precise localization by learning object micro-Doppler patterns from radar tensors and incorporating vision feed-forward signals. Our context learning method ensures reliable performance even when one sensor fails. The paper contributes significantly by showcasing improved performance for vulnerable road users compared to state-of-the-art methods. Validation on our larger radar-camera dataset confirms these achievements. Additionally, our method demonstrates up to a " + }, + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "type": "text", + "content": " improvement in mAP performance across diverse datasets, with ablation studies highlighting the efficacy of each framework component. Future research directions could explore optimizing fusion mechanisms and incorporating additional sensor modalities like thermal or acoustic sensors for even more accurate object detection and localization. However, as more sensors are integrated, ensuring robustness in multi-sensor object detection remains critical for real-world applications focused on enhancing road user safety." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 101 + ], + "type": "text", + "content": "W.Y Lee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 481, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 481, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 481, + 190 + ], + "type": "text", + "content": "This work was funded by EU Horizon 2020 ECSEL JU research and innovation programme under grant agreement 876487 (NextPerception), by EU Horizon KDT JU research and innovation programme under grant agreement 101139769 (Distribuse), and by the Flemish Government (AI Research Program)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 209, + 197, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 209, + 197, + 220 + ], + "spans": [ + { + "bbox": [ + 133, + 209, + 197, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 234, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 234, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 480, + 267 + ], + "type": "text", + "content": "1. Barnes, D., Gadd, M., Murcutt, P., Newman, P., Posner, I.: The oxford radar robotcar dataset: A radar extension to the oxford robotcar dataset. In: ICRA (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 268, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 268, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 268, + 480, + 300 + ], + "type": "text", + "content": "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 301, + 480, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 301, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 138, + 301, + 480, + 323 + ], + "type": "text", + "content": "3. Chen, S., He, W., Ren, J., Jiang, X.: Attention-based dual-stream vision transformer for radar gait recognition. In: ICASSP. IEEE (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 323, + 480, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 480, + 344 + ], + "type": "text", + "content": "4. Chen, X., Ma, H., Wan, J., Li, B., Xia, T.: Multi-view 3d object detection network for autonomous driving. In: CVPR (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 346, + 480, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 480, + 388 + ], + "type": "text", + "content": "5. Dimitrievski, M., Shopovska, I., Van Hamme, D., Veelaert, P., Philips, W.: Weakly supervised deep learning method for vulnerable road user detection in fmcw radar. In: IEEE International Conference on Intelligent Transportation Systems (ITSC). IEEE (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 390, + 480, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 390, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 390, + 480, + 422 + ], + "type": "text", + "content": "6. Gao, X., Xing, G., Roy, S., Liu, H.: Experiments with mmwave automotive radar test-bed. In: 2019 53rd Asilomar conference on signals, systems, and computers. pp. 1-6. IEEE (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 422, + 480, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 480, + 444 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 480, + 444 + ], + "type": "text", + "content": "7. Gao, X., Xing, G., Roy, S., Liu, H.: Ramp-cnn: A novel neural network for enhanced automotive radar object recognition. IEEE Sensors Journal (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 445, + 480, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 480, + 477 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 480, + 477 + ], + "type": "text", + "content": "8. Griffiths, H., Cohen, L., Watts, S., Mokole, E., Baker, C., Wicks, M., Blunt, S.: Radar spectrum engineering and management: Technical and regulatory issues. Proceedings of the IEEE (2014)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 478, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 478, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 478, + 480, + 510 + ], + "type": "text", + "content": "9. Gusland, D., Christiansen, J.M., Torvik, B., Fioranelli, F., Gurbuz, S.Z., Ritchie, M.: Open radar initiative: Large scale dataset for benchmarking of micro-doppler recognition algorithms. In: IEEE Radar Conference (RadarConf). pp. 1-6 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 512, + 480, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 532 + ], + "type": "text", + "content": "10. Hartley, R., Zisserman, A.: Multiple view geometry in computer vision. Cambridge university press (2003)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 533, + 480, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 480, + 565 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 480, + 565 + ], + "type": "text", + "content": "1. Hwang, J.J., Kretzschmar, H., Manela, J., Rafferty, S., Armstrong-Crews, N., Chen, T., Anguelov, D.: Cramnet: Camera-radar fusion with ray-constrained cross-attention for robust 3d object detection. In: ECCV (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 566, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 566, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 566, + 480, + 587 + ], + "type": "text", + "content": "2. Jocher, G., Chaurasia, A., Qiu, J.: YOLO by Ultralytics (2023), https://github.com/ultralytics/ultralytics" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 588, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 480, + 609 + ], + "type": "text", + "content": "3. Kim, Y., Kim, S., Choi, J.W., Kum, D.: Craft: Camera-radar 3d object detection with spatio-contextual fusion transformer. In: AAAI (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 610, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 480, + 632 + ], + "type": "text", + "content": "4. Kim, Y., Shin, J., Kim, S., Lee, I.J., Choi, J.W., Kum, D.: Crn: Camera radar net for accurate, robust, efficient 3d perception. In: ICCV (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "5. Kopp, J., Kellner, D., Piroli, A., Dietmayer, K.: Tackling clutter in radar data-label generation and detection using pointnet++. arXiv preprint arXiv:2303.09530 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 101 + ], + "type": "text", + "content": "Camera-Assisted Radar-Based Network for Vulnerable Road User Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "16. Lee, W.Y., Dimitrievski, M., Jovanov, L., Philips, W.: Spatio-temporal consistency for semi-supervised learning using 3d radar cubes. In: IEEE Intelligent Vehicles Symposium (IV). IEEE (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "17. Lee, W.Y., Jovanov, L., Kumcu, A., Philips, W.: Arc: Automotive radar consistency regularization for semi-supervised learning. IEEE Transactions on Intelligent Vehicles (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 206 + ], + "type": "text", + "content": "18. Lee, W.Y., Jovanov, L., Philips, W.: Multi-view target transformation for pedestrian detection. In: WACV Workshops (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 207, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 481, + 228 + ], + "type": "text", + "content": "19. Li, Y., Ge, Z., Yu, G., Wang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: AAAI (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "text", + "content": "20. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: ECCV (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "type": "text", + "content": "21. Lim, T.Y., Ansari, A., Major, B., Fontijne, D., Hamilton, M., Gowaikar, R., Subramanian, S.: Radar and camera early fusion for vehicle detection in advanced driver assistance systems. In: NIPS Workshops (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 297, + 481, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 318 + ], + "type": "text", + "content": "22. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: ICCV (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 319, + 481, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 481, + 340 + ], + "type": "text", + "content": "23. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.Y., Berg, A.C.: Ssd: Single shot multibox detector. In: ECCV (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 341, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 481, + 363 + ], + "type": "text", + "content": "24. Liu, Y., Wang, F., Wang, N., ZHANG, Z.X.: Echoes beyond points: Unleashing the power of raw radar data in multi-modality fusion. NeurIPS (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "type": "text", + "content": "25. Liwen, Z., Xinyan, Z., Youcheng, Z., Yufei, G., Yuanpei, C., Xuhui, H., Zhe, M.: Peakconv: Learning peak receptive field for radar semantic segmentation. In: CVPR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 397, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 418 + ], + "type": "text", + "content": "26. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: CVPR (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 481, + 453 + ], + "type": "text", + "content": "27. Mallot, H.A., Bulthoff, H.H., Little, J.J., Bohrer, S.: Inverse perspective mapping simplifies optical flow computation and obstacle detection. Biological cybernetics (1991)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 453, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 475 + ], + "type": "text", + "content": "28. Meyer, M., Kuschk, G.: Automotive radar dataset for deep learning based 3d object detection. In: European Radar Conference (EuRAD) (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 475, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 497 + ], + "type": "text", + "content": "29. Ouaknine, A., Newson, A., Pérez, P., Tupin, F., Rebut, J.: Multi-view radar semantic segmentation. In: ICCV (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 498, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 520 + ], + "type": "text", + "content": "30. Ouaknine, A., Newson, A., Rebut, J., Tupin, F., Pérez, P.: Carrada dataset: Camera and automotive radar with range- angle- doppler annotations. In: ICPR (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 521, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 481, + 542 + ], + "type": "text", + "content": "31. Paek, D.H., Kong, S.H., Wijaya, K.T.: K-radar: 4d radar object detection for autonomous driving in various weather conditions. NeurIPS (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 543, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 481, + 586 + ], + "type": "text", + "content": "32. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 194-210. Springer (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 587, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 481, + 609 + ], + "type": "text", + "content": "33. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NIPS (2017)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "text", + "content": "34. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI. Springer (2015)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "35. Schumann, O., Hahn, M., Dickmann, J., Wöhler, C.: Semantic segmentation on radar point clouds. In: International Conference on Information Fusion (FUSION) (2018)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 100 + ], + "type": "text", + "content": "W.Y Lee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 347 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "36. Sheeny, M., De Pellegrin, E., Mukherjee, S., Ahrabian, A., Wang, S., Wallace, A.: Radiate: A radar dataset for automotive perception in bad weather. In: ICRA (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "text", + "content": "37. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: CVPR (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 215 + ], + "type": "text", + "content": "38. Vaizman, Y., Weibel, N., Lanckriet, G.: Context recognition in-the-wild: Unified model for multi-modal sensors and multi-label classification. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "39. Wang, W., Tran, D., Feiszli, M.: What makes training multi-modal classification networks hard? In: CVPR (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "40. Wang, Y., Wang, G., Hsu, H.M., Liu, H., Hwang, J.N.: Rethinking of radar's role: A camera-radar dataset and systematic annotator via coordinate alignment. In: CVPR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 314 + ], + "type": "text", + "content": "41. Yao, S., Guan, R., Huang, X., Li, Z., Sha, X., Yue, Y., Lim, E.G., Seo, H., Man, K.L., Zhu, X., Yue, Y.: Radar-camera fusion for object detection and semantic segmentation in autonomous driving: A comprehensive review. IEEE Transactions on Intelligent Vehicles (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 314, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 314, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 480, + 347 + ], + "type": "text", + "content": "42. Zhou, T., Chen, J., Shi, Y., Jiang, K., Yang, M., Yang, D.: Bridging the view disparity between radar and camera features for multi-modal fusion 3d object detection. IEEE Transactions on Intelligent Vehicles (2023)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 101 + ], + "type": "text", + "content": "Camera-Assisted Radar-Based Network for Vulnerable Road User Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CARFF_ Conditional Auto-encoded Radiance Field for 3D Scene Forecasting/4d2110cb-f049-4580-b1d7-c4b728d00449_content_list.json b/2024/CARFF_ Conditional Auto-encoded Radiance Field for 3D Scene Forecasting/4d2110cb-f049-4580-b1d7-c4b728d00449_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6f8a8ac8a99b12a297ba56598ee084562b2e0312 --- /dev/null +++ b/2024/CARFF_ Conditional Auto-encoded Radiance Field for 3D Scene Forecasting/4d2110cb-f049-4580-b1d7-c4b728d00449_content_list.json @@ -0,0 +1,1679 @@ +[ + { + "type": "text", + "text": "CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting", + "text_level": 1, + "bbox": [ + 241, + 140, + 759, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiezhi \"Stephen\" Yang $^{1}$ , Khushi Desai $^{2}$ , Charles Packer $^{3}$ , Harshil Bhatia $^{4}$ , Nicholas Rhinehart $^{3}$ , Rowan McAllister $^{5}$ , and Joseph E. Gonzalez $^{3}$", + "bbox": [ + 243, + 218, + 759, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Harvard University MA 02138, USA", + "bbox": [ + 372, + 277, + 630, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Columbia University NY 10025, USA", + "bbox": [ + 370, + 291, + 632, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "UC Berkeley CA 94720, USA", + "bbox": [ + 397, + 306, + 606, + 318 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Avataar.ai KA 560103, India", + "bbox": [ + 398, + 319, + 604, + 332 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5 Toyota Research Institute CA 94022, USA", + "bbox": [ + 352, + 333, + 648, + 345 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We propose CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting, a method for predicting future 3D scenes given past observations. Our method maps 2D ego-centric images to a distribution over plausible 3D latent scene configurations and predicts the evolution of hypothesized scenes through time. Our latents condition a global Neural Radiance Field (NeRF) to represent a 3D scene model, enabling explainable predictions and straightforward downstream planning. This approach models the world as a POMDP and considers complex scenarios of uncertainty in environmental states and dynamics. Specifically, we employ a two-stage training of Pose-Conditional-VAE and NeRF to learn 3D representations, and auto-regressively predict latent scene representations utilizing a mixture density network. We demonstrate the utility of our method in scenarios using the CARLA driving simulator, where CARFF enables efficient trajectory and contingency planning in complex multi-agent autonomous driving scenarios involving occlusions. Video and code are available at: www.carff website.", + "bbox": [ + 259, + 385, + 743, + 608 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 633, + 375, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans often imagine what they cannot see given partial visual context. Consider a scenario where reasoning about the unobserved is critical to safe decision-making: for example, a driver navigating a blind intersection. An expert driver will plan according to what they believe may or may not exist in occluded regions of their vision. The driver's belief - defined as the understanding of the world modeled with consideration for inherent environment uncertainties - is informed by their partial observations (i.e., the presence of other vehicles on the road), as well as their prior knowledge (e.g., past experience navigating this intersection).", + "bbox": [ + 212, + 665, + 787, + 785 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "When reasoning about the unobserved, humans form complex beliefs about the existence, position, shapes, colors, and textures of occluded scene portions", + "bbox": [ + 214, + 786, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Core contributors", + "bbox": [ + 217, + 824, + 354, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(e.g., an oncoming car). Autonomous systems with high-dimensional sensor data, like video or LiDAR, traditionally reduce this data to low-dimensional state information (e.g., position and velocity of tracked objects) for prediction and planning.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition to tracking fully observed objects, this object-centric framework handles partially observed settings by considering potentially dangerous unobserved objects. These systems often plan for worst-case scenarios, such as a \"ghost car\" at the edge of the visible field of view (44).", + "bbox": [ + 212, + 207, + 782, + 267 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b84fe5410986cc3ec666ba1809f5163c18eadc7a63813e26b65137ad51372830.jpg", + "image_caption": [ + "Fig. 1: CARFF 3D planning application for driving. An input image containing a partially observable view of an intersection is processed by CARFF's encoder to establish 3D environment state beliefs, i.e. the predicted possible state of the world: whether or not there could be another vehicle approaching the intersection. These beliefs are used to forecast the future in 3D for planning, generating one among two possible actions for the vehicle to merge into the other lane." + ], + "image_footnote": [], + "bbox": [ + 333, + 297, + 674, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advances in neural rendering, particularly Neural Radiance Fields (NeRF), have significantly improved 3D scene representation learning. NeRF enables novel view synthesis, thus simplifying the process of viewing behind occlusions. NeRF decouples the dependency of scene representation from traditional object detection and tracking, allowing for the capture of vital visual information that might be missed by detectors, yet is crucial for safe decision-making. NeRF's implicit density representation of explicit geometry also facilitates its direct application in motion planning without the need for rendering. NeRF's ability to represent both visual and geometric information makes them a more general and intuitive 3D representation for autonomous systems.", + "bbox": [ + 212, + 537, + 787, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite NeRF's advantages, achieving probabilistic predictions in 3D based on reasoning from occluded views is challenging. For example, discriminative models that yield categorical predictions are unable to capture the underlying 3D structure, impeding their ability to model uncertainty. While prior work on 3D representation captures view-invariant structures, their application is primarily confined to simple scenarios (20). We present CARFF, which to our knowledge, is the first forecasting approach in scenarios with partial observations that uniquely facilitates stochastic predictions in a partially observable Markov decision process (POMDP) within a 3D representation, effectively integrating visual perception and geometry. Specifically, we make the following contributions:", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. We propose a novel architecture $PC$ -VAE: Pose-Conditioned Variational Autoencoder. The encoder maps potentially partially observable ego-centric images to pose-invariant latent scene representations, which hold state beliefs of the POMDP with implicit probability distributions (see Sec. 3.1).", + "bbox": [ + 222, + 146, + 785, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. We develop the two-stage training pipeline that uniquely enables complex scene modeling with a probabilistic objective. This involves separately training the PC-VAE and a latent conditioned neural radiance field that functions as a 3D decoder, enabling interpretable predictions (see Sec. 3.1).", + "bbox": [ + 222, + 207, + 785, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. We design a mixture density model to predict the evolution of 3D scenes over time stochastically and regressively in the encoder belief space (see Sec. 3.2). This allows for an effective sampling based-controller to output actions in the POMDP.", + "bbox": [ + 222, + 266, + 785, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We demonstrate how CARFF can be used to enable contingency planning in complex driving scenarios that require reasoning into visual occlusions on CARLA simulated datasets inspired by autonomous driving planning tasks (29; 54; 55; 56). A potential application of CARFF is illustrated in Fig. 1.", + "bbox": [ + 215, + 332, + 785, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 412, + 382, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 NeRF and 3D representations", + "text_level": 1, + "bbox": [ + 215, + 440, + 509, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural radiance fields. Neural Radiance Fields (NeRF) (2; 27; 43) for 3D representations generate high-resolution, photorealistic scenes. Instant Neural Graphics Primitive (Instant-NGP) (28) speeds up training and rendering time by introducing a multi-resolution hash encoding. Other works like Plenoxels (12) and DirectVoxGo (DVGO) (41) also provide similar speedups. Recent advancements in volumetric representations such as 3D Gaussian Splatting (18) enhance rendering efficiency while maintaining compatibility with traditional NeRF applications (11). We utilize Instant-NGP for its accessibility, although our approach is adaptable to alternative rendering methods. NeRFs have also been extended for several tasks such as modeling large-scale unbounded scenes (2; 42; 49), scene from sparse views (7; 39; 48) and multiple scenes (20; 50). For an in-depth survey on neural representation learning and its applications we refer the reader to (45).", + "bbox": [ + 212, + 460, + 787, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generalizable novel view synthesis models, like pixelNeRF and pixelSplat (5; 53), learn a scene prior to render novel views from sparse existing ones. In contrast, CARFF is based on a VAE, encoding a probabilistic objective and decoding to future 3D scenes. Dynamic NeRF models scenes with moving or deforming objects, within which a widely used approach is to construct a canonical space and predict a deformation field (22; 33; 34; 36). The canonical space is usually a static scene, and the model learns an implicitly represented flow field (33; 36). A recent line of work also models dynamic scenes via different representations and decomposition (3; 40). These approaches tend to perform better for spatially bounded and predictable scenes with relatively small variations (3; 23; 33; 53). Moreover, these methods only solve for changes in the environment but are limited in incorporating stochasticity in the environment.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting", + "bbox": [ + 218, + 113, + 730, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/911c9359ba538544b664309f060185792214c88863a23d0e2a3844cb4b919ec8.jpg", + "image_caption": [ + "Ego-centric Lane View", + "Fig. 2: Novel view planning application. CARFF allows reasoning behind occluded views from the ego car as simple as moving the camera to see the sampled belief predictions, allowing simple downstream planning using, for example, density probing or 2D segmentation models from arbitrary angles." + ], + "image_footnote": [], + "bbox": [ + 393, + 147, + 496, + 213 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7f904b0876f8b554d72dcdbc4a9168a495f5a5cb691242159c4c1ec33921f7dd.jpg", + "image_caption": [ + "Predicted 3D View" + ], + "image_footnote": [], + "bbox": [ + 501, + 147, + 614, + 213 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-scene NeRF: Our approach builds on multi-scene NeRF approaches (20; 47; 50; 51) that learn a global latent scene representation, which conditions the NeRF, allowing a single NeRF to effectively represent various scenes. A similar method, NeRF-VAE, was introduced by Kosiorek et al. (20) to create a geometrically consistent 3D generative model with generalization to out-of-distribution cameras. However, NeRF-VAE (20) is prone to mode collapse when faced with complex visual information (see Sec. 4.2).", + "bbox": [ + 212, + 323, + 787, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Scene Forecasting", + "text_level": 1, + "bbox": [ + 215, + 467, + 410, + 483 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Planning in 2D space: Planning in large, continuous state-action spaces is challenging due to exponentially large search spaces (32), leading to various approximation methods for tractability (26; 35). Model-free (13; 31; 46) and model-based (4) reinforcement learning frameworks, along with other learning-based methods (6; 29), have emerged as viable approaches. Additionally, methods forecast for downstream control (16), learn behavior models for contingency planning (38), or predict the existence and intentions of unobserved agents (30). While these methods operate in 2D, we reason under partial observations and account for these factors in 3D.", + "bbox": [ + 212, + 507, + 787, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "NeRF in robotics: Recent works have applied NeRFs in robotics for localization (52), navigation (1; 25), dynamics modeling (10; 22), and robotic grasping (15; 19). Adamkiewicz et al.(1) propose quadcopter motion planning in NeRF models by sampling the learned density function, useful for forecasting and planning. Driess et al.(10) employ a graph neural network to learn dynamics in a multi-object NeRF scene. Li et al. (21) focus on pushing tasks and address grasping and planning with NeRF and a separate latent dynamics model. Prior approaches work in simple, static scenes (1) or uses deterministic dynamics models (21). CARFF addresses complex, realistic environments with both state and dynamics uncertainty, considering potential object existence and unknown movements.", + "bbox": [ + 212, + 672, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 143, + 330, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recent advancements in 3D scene representation allow for modeling environments in a contextually rich and interactive 3D space. This offers analytical benefits, such as spatial analysis with soft occupancy grids and object detection through novel view synthesis. Given these advantages, our primary objective is to develop a model for probabilistic 3D scene forecasting in dynamic environments. However, direct integration of 3D scene representation via NeRF and probabilistic models like VAE often involves non-convex and inter-dependent optimization, which causes unstable training. For instance, NeRF's optimization may rely on the VAE's latent space being structured to provide informative gradients. To", + "bbox": [ + 212, + 174, + 787, + 310 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a431e18c5007cafac98de48641d9202701cb8965d45e1dc5703eb3ff32b6ada2.jpg", + "image_caption": [ + "Fig. 3: Visualizing CARFF's two stage training process. Left: The convolutional VIT-based encoder encodes each image $I$ at timestamps $t, t'$ and camera poses $c, c'$ into Gaussian latent distributions. Assuming two timestamps and an overparameterized latent, one Gaussian distribution will have a smaller $\\sigma^2$ , and different $\\mu$ across timestamps. Upper Right: The pose-conditional decoder stochastically decodes the sampled latent $z$ using the camera pose $c''$ into images $I_{c''}^t$ and $I_{c''}^{t'}$ . The decoded reconstruction and ground truth images are used for the loss $\\mathcal{L}_{\\mathrm{MSE}}$ , PC-VAE. Lower Right: A NeRF is trained by conditioning on the latent variables sampled from the optimized Gaussian parameters. These parameters characterize the distinct timestamp distributions derived from the PC-VAE. An MSE loss is calculated for NeRF as $\\mathcal{L}_{\\mathrm{MSE}}$ , NeRF." + ], + "image_footnote": [], + "bbox": [ + 246, + 334, + 759, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "navigate these complexities, our method bifurcates the training process into two stages (see Fig. 3). First, we train the PC-VAE to learn view-invariant scene representations. Next, we replace the decoder with a NeRF to learn a 3D scene from the latent representations. The latent scene representations capture the environmental states and dynamics over possible underlying scenes, while NeRF synthesizes novel views within the belief space, giving us the ability to see the unobserved (see Fig. 2 and Sec. 3.1). During prediction, uncertainties can be modeled by sampling latents auto-regressively from a predicted Gaussian mixture, allowing for effective decision-making. To this extent, we approach scene forecasting as a POMDP over latent distributions, which enables us to capture multi-modal beliefs for planning amidst perceptual uncertainty (see Sec. 3.2).", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting", + "bbox": [ + 218, + 113, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Pose-Conditional VAE (PC-VAE) and NeRF", + "text_level": 1, + "bbox": [ + 215, + 146, + 630, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Architecture: We assume that the model follows a Markovian process, and thus each belief state only depends on the previous. Given a scene $S_{t}$ at timestamp $t$ , we have an ego-centric observation image $I_{c}^{t}$ captured from camera pose $c$ . The objective is to formulate a 3D representation of the image that holds implicit probability distributions of the possible states, where we can perform a forecasting step that evolves the scene forward. Here, the POMDP can be seen as an MDP in belief space (17). To achieve this, we utilize a radiance field conditioned on latent variable $z$ sampled from the posterior distribution $q_{\\phi}(z|I_{c}^{t})$ . Now, to learn the posterior, we utilize PC-VAE. We construct an encoder using convolutional layers and a pre-trained ViT on ImageNet (8). The encoder learns a mapping from the image space to a Gaussian distributed latent space $q_{\\phi}(z|I_{c}^{t}) = \\mathcal{N}(\\mu, \\sigma^{2})$ parametrized by mean $\\mu$ and variance $\\sigma^2$ . The decoder, $p(I|z,c)$ , conditioned on camera pose $c$ , maps the latent $z \\sim \\mathcal{N}(\\mu, \\sigma^2)$ into the image space $I$ . This helps the encoder to generate latents that are invariant to the camera pose $c$ .", + "bbox": [ + 212, + 169, + 787, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To enable 3D scene modeling, we employ Instant-NGP (28), which incorporates a hash grid and an occupancy grid to enhance computation efficiency. Additionally, a smaller multilayer perceptron (MLP), $F_{\\theta}(z)$ can be utilized to model the density and appearance, given by:", + "bbox": [ + 212, + 395, + 787, + 455 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\theta} (z): (\\mathbf {x}, \\mathbf {d}, z) \\rightarrow ((r, g, b), \\sigma) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 465, + 785, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $\\mathbf{x} \\in \\mathbb{R}^3$ and $\\mathbf{d} \\in (\\theta, \\phi)$ represent the location vector and the viewing direction respectively. The MLP is conditioned on the sampled scene latents $z \\sim q_{\\phi}(z|I_c^t)$ (see Appendix ??).", + "bbox": [ + 214, + 489, + 785, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training methodology: The architecture alone does not enable us to model complex scenarios, as seen through a similar example in NeRF-VAE (20). A crucial contribution of our work is our two-stage training framework which stabilizes the training. First, we optimize the convolutional ViT based encoder and pose-conditional convolutional decoder in the pixel space for reconstruction. This enables our method to deal with more complex and realistic scenes as the encoding is learned in a semantically rich 2D space. By conditioning the decoder on camera poses, we achieve disentanglement between camera view angles and scene context, making the representation view-invariant and the encoder 3D-aware. Once rich latent representations are learned, we replace the decoder with a latent-conditioned NeRF over the latent space of the frozen encoder. The NeRF reconstructs encoder beliefs in 3D for novel view synthesis.", + "bbox": [ + 214, + 547, + 787, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Loss: PC-VAE is trained using standard VAE loss, with mean square error (MSE) and a Kullback-Leibler (KL) divergence given by evidence lower bound:", + "bbox": [ + 215, + 741, + 787, + 772 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {P C - V A E} = \\mathcal {L} _ {M S E, P C - V A E} + \\mathcal {L} _ {K L D, P C - V A E} = \\\\ \\left\\| p \\left(I | z, c ^ {\\prime \\prime}\\right) - I _ {c ^ {\\prime \\prime}} ^ {t} \\right\\| ^ {2} + \\mathbb {E} _ {q (z \\mid I _ {c} ^ {t})} [ \\log p (I | z) ] - w _ {K L} D _ {K L} \\left(q _ {\\phi} \\left(z \\mid I _ {c} ^ {t}\\right) \\mid p (I | z)\\right) \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 801, + 784, + 835 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $w_{KL}$ denotes the KL divergence loss weight and $z \\sim q_{\\phi}(z|I_c^t)$ . To make our representation 3D-aware, our posterior is encoded using camera $c$ while the decoder is conditioned on a randomly sampled pose $c''$ .", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "KL divergence regularizes the latent space to balance conditioned reconstruction and stochasticity under occlusion. An elevated KL divergence loss weight $w_{KL}$ pushes the latents closer to a standard normal distribution, $\\mathcal{N}(0,1)$ , thereby ensuring probabilistic sampling in scenarios under partial observation. However, excessive regularization causes the latents to be less separable, leading to mode collapse. To mitigate this, we adopt delayed linear KL divergence loss weight scheduling to strike a balanced $w_{KL}$ .", + "bbox": [ + 212, + 191, + 784, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we learn a NeRF decoder on the posterior of the VAE to model scenes. At any timestamp $t$ we use a standard photometric loss for training the NeRF, given by the following equation:", + "bbox": [ + 212, + 297, + 782, + 342 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {M S E, N e R F} = \\left\\| I _ {c} ^ {t} - r e n d e r \\left(F _ {\\theta} \\left(\\cdot \\mid q _ {\\phi} \\left(z \\mid I _ {c} ^ {t}\\right)\\right)\\right) \\right\\| ^ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 343, + 351, + 784, + 368 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use a standard rendering algorithm as proposed by Müller et al. (28). Next, we build a forecasting module over the learned latent space of our pose-conditional encoder.", + "bbox": [ + 212, + 375, + 787, + 419 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Scene Forecasting", + "text_level": 1, + "bbox": [ + 214, + 440, + 408, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Formulation: The current formulation allows us to model scenes with different configurations across timestamps. In order to forecast future configurations of a scene given an ego-centric view, we need to predict future latent distributions. We formulate the forecasting as a POMDP over the posterior distribution $q_{\\phi}(z|I_c^t)$ in the PC-VAE's latent space.", + "bbox": [ + 212, + 462, + 782, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "During inference, we observe stochastic behaviors under occlusion, which motivates us to learn a mixture of several Gaussian distributions that potentially denote different scene possibilities. Therefore, we model the POMDP using a Mixture Density Network (MDN), with multi-headed MLPs, that predicts a mixture of $K$ Gaussians. At any timestamp $t$ the distribution is given as:", + "bbox": [ + 212, + 537, + 782, + 613 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nq _ {\\phi} ^ {\\prime} \\left(z _ {t} \\mid I _ {c} ^ {t - 1}\\right) = M D N \\left(q _ {\\phi} \\left(z _ {t - 1} \\mid I _ {c} ^ {t - 1}\\right)\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 622, + 784, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The model is conditioned on the posterior distribution $q_{\\phi}(z_{t - 1})$ to learn a predicted posterior distribution $q_{\\phi}'(z_t|I_c^{t - 1})$ at each timestamp. The predicted posterior distribution is given by the mixture of Gaussian:", + "bbox": [ + 212, + 646, + 782, + 691 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nq _ {\\phi} ^ {\\prime} \\left(z _ {t}\\right) = \\sum_ {i = 1} ^ {K} \\pi_ {i} \\mathcal {N} \\left(\\mu_ {i}, \\sigma_ {i} ^ {2}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 700, + 784, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "here, $\\pi_i$ , $\\mu_i$ , and $\\sigma_i^2$ denote the mixture weight, mean, and variance of the $i^{th}$ Gaussian distribution within the posterior distribution. Here, $K$ is the total number of Gaussians. For brevity we remove their conditioning on the posterior $q_{\\phi}(z_{t-1})$ and sampled latent $z_{t-1}$ . We sample $z_t$ from the mixture of Gaussians $q_{\\phi}^{\\prime}(z_t)$ , where $z_t$ likely falls within one of the Gaussian modes. The configuration corresponding to the mode is reflected in the 3D scene rendered by NeRF.", + "bbox": [ + 214, + 748, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting", + "bbox": [ + 218, + 113, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7379d9d391663f7e3ed896bd624d5d8ee05057575b6903fed5329e19f50b314a.jpg", + "image_caption": [ + "Fig.4: Multi-scene CARLA datasets. Varying car configurations and scenes for the Multi-Scene Two Lane Merge dataset (left) and the Multi-Scene Approaching Intersection dataset (right)." + ], + "image_footnote": [], + "bbox": [ + 225, + 207, + 346, + 276 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/77088b311dcf40ce6bf05ce57903c749dd732695fd3c03cc0f7846dc544cbf3a.jpg", + "image_caption": [ + "Fig.5: Blender dataset. Blender dataset with a blue cube and a potential red cylinder exhibiting probabilistic temporal movement. The possible occlusions from different camera angles demonstrate how movement needs to be modeled probabilistically." + ], + "image_footnote": [], + "bbox": [ + 354, + 207, + 478, + 275 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/afbb542921bb5c6efec2691d08739e8079af0267e0e4a8378ee2eb6ae6b38538.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 143, + 581, + 275 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1f7ccc21d1f36ed77b383958701e42df86a5a7fa5bf0c0f06c76d122c56bc8c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 143, + 674, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/07cdc9ebb6f1228e46f4cba95d292c4f9ec9a600cbb7ad2d55a31179ce468385.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 143, + 774, + 275 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Loss: To optimize the MDN, we minimize a negative log-likelihood function, given by:", + "bbox": [ + 212, + 397, + 782, + 426 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {M D N} = - \\sum_ {j = 1} ^ {N} \\log \\left(\\sum_ {i = 1} ^ {K} \\pi_ {i} \\mathcal {N} \\left(y _ {j}; \\mu_ {i}, \\sigma_ {i} ^ {2}\\right)\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 426, + 784, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $y_{i} \\sim q_{\\phi}(z_{t})$ is sampled from the distribution of latent $z_{t}$ , learned by the encoder, and $N$ denotes the total number of samples.", + "bbox": [ + 212, + 474, + 784, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inference: We consider an unseen ego-centric image and retrieve its posterior $q_{\\phi}(z_t)$ through the encoder. Next, we predict the possible future posterior distribution $q_{\\phi}'(z_{t + 1})$ . From the predicted posterior, we sample a scene latent and perform localization. We achieve this via (a) density probing the NeRF or (b) segmenting the rendered novel views using off-the-shelf methods such as YOLO (37) (see Fig. 2). These allow us to retrieve a corresponding Gaussian distribution $q_{\\phi}(z_{t + 1})$ in encoder latent space. This is auto-regressively fed back into the MDN to predict the next timestamp. See Fig. 6 for an overview of the pipeline.", + "bbox": [ + 212, + 516, + 787, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Results", + "text_level": 1, + "bbox": [ + 215, + 659, + 323, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Decision-making under perceptual uncertainty is a pervasive challenge faced in robotics and autonomous driving, as the real environment is mostly likely partially observable, making it a POMDP. In a partially observable driving scenario, accurate inference regarding the presence of potentially obscured agents is pivotal. We evaluate the effectiveness of CARFF on common driving situations with partial observability and added complexity. We implemented several scenarios in the CARLA driving simulator (9) (see Fig. 4). A single NVIDIA RTX 3090 GPU is used to train PC-VAE, NeRF, and the MDN. All models, trained sequentially, tend to converge within a combined time frame of 24 hours. A detailed experimental setup can be found in Appendix ??. We show that, given", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "partially observable 2D inputs, CARFF performs well in predicting latent distributions that represent complete 3D scenes. Using these predictions we design a CARFF-based controller for performing downstream planning tasks.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Data Generation", + "text_level": 1, + "bbox": [ + 215, + 213, + 401, + 227 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conduct experiments on (a) synthetic blender dataset for principle experiments to test the probabilistic modeling capacities in isolation of the vision encoder (it is visually as simple as possible, but requires the full predictive model proposed in CARFF) and (b) CARLA-based driving datasets for more complex driving scenarios (9). To deliver convincing results, we model these driving scenarios off of related works (29; 54; 55; 56) that concern planning for driving under difficult situations. We generate the datasets in 3D by programming an ego object and varying actor objects in different configurations.", + "bbox": [ + 212, + 238, + 784, + 359 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/19902d354443b11142b0941a8bce70e38f2d59be8ee146263f795cf147edc0a9.jpg", + "image_caption": [ + "Fig. 6: Auto-regressive inference in scene prediction. The input image at timesamp $t$ , $I_{c}^{t}$ , is encoded using the pre-trained encoder from PC-VAE. The corresponding latent distribution is fed into the Mixture Density Network, which predicts a mixture of Gaussians. Each of the $K$ Gaussians is a latent distribution that may correspond to different beliefs at the next timestamp. The mixture of Gaussians is sampled repeatedly for the predicted latent beliefs, visualized as $I_{c',scni}^{t+1}$ , representing potentially the $i$ th possible outcome. This is used to condition the NeRF to generate 3D views of the scene. To accomplish autoregressive predictions, we probe the NeRF for the location of the car and feed this information back to the pre-trained encoder to predict the scene at the next timestamp." + ], + "image_footnote": [], + "bbox": [ + 232, + 387, + 772, + 526 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Blender synthetic dataset: This comprises of a stationary blue cube (ego) accompanied by a red cylinder (actor) that may or may not be present (see Fig. 5). If the actor is present, it exhibits lateral movement as depicted in Fig. 5. This simplistic setting provides an interpretable framework to evaluate our model.", + "bbox": [ + 212, + 719, + 784, + 780 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "CARLA dataset: Each dataset is simulated for $N$ timestamps and uses $C = 100$ predefined camera poses to capture images of the environment under full observation, partial observation, and no visibility. These datasets are modeled after", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting", + "bbox": [ + 218, + 114, + 730, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "common driving scenarios involving state uncertainty that have been proposed in related works such as Active Visual Planning (29).", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "a) Single-Scene Approaching Intersection: The ego vehicle is positioned at a T-intersection. An actor vehicle traverses the crossing along an evenly spaced, predefined trajectory. We simulate this for $N = 10$ timestamps. We mainly use this dataset to predict the evolution of timestamps under full observation.", + "b) Multi-Scene Approaching Intersection: We extend the previous scenario to a more complicated setting with state uncertainty, by making the existence of the actor vehicle probabilistic. A similar intersection crossing is simulated for $N = 3$ timestamps for both possibilities. The ego vehicle's view of the actor may be occluded as it approaches the T-intersection over the $N$ timestamps. The ego vehicle either moves forward or halts at the junction (see Fig. 4).", + "c) Multi-Scene Multi-actor Two Lane Merge: To add more environment dynamics uncertainty, we consider a multi-actor setting at an intersection of two merging lanes. We simulate the scenario at an intersection with partial occlusions, with the second approaching actor having variable speed. Here the ego vehicle can either merge into the left lane before the second actor or after all the actors pass, (see Fig. 4). Each branch is simulated for $N = 3$ timestamps." + ], + "bbox": [ + 212, + 176, + 787, + 420 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 CARFF Evaluation", + "text_level": 1, + "bbox": [ + 215, + 443, + 423, + 458 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/f1048e02c4f7ced63579f0768d4518624ef27639ce5017b236a3fed76387d037.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method3DComplex ScenariosState UncertaintyDynamics UncertaintyPredictionPlanningCode Released
CARFF✓✓✓✓✓✓✓
(20)✓✓
(21)✓✓✓✓
(1)✓✓✓✓
(29)✓✓✓✓✓
", + "bbox": [ + 215, + 494, + 787, + 606 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 1: Qualitative comparison of CARFF to related works. CARFF accomplishes all highlighted objectives as opposed to NeRF-VAE (20), NeRF for Visuomotor Control (21), Vision-only NeRF Navigation (1), and AVP (29). We compare whether methods reason in a 3D environment and perform novel view synthesis; work on complex scenarios; predict probabilistically under state and dynamics uncertainty; forecast into the future; and use model predictions for decision-making.", + "bbox": [ + 212, + 616, + 787, + 700 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A desirable behavior from our model is that it should predict a complete set of possible scenes consistent with the given ego-centric image, which could be partially observable. This is crucial for autonomous driving in unpredictable environments as it ensures strategic decision-making based on potential hazards. To achieve this we require a rich PC-VAE latent space, high-quality novel view synthesis, and auto-regressive probabilistic predictions of latents at future timestamps. We evaluate CARFF on a simple synthetic blender-based dataset", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "and each CARLA-based dataset. Additionally, we extend our model application to a hand-manipulation dataset in Appendix ??", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparisons with related work: We attempt to compare CARFF to existing approaches. NeRF-VAE has the most comparable objective, but during our experiments, it collapses to black using CARLA datasets. We make further qualitative comparisons to other most similar methods in Tab. 1, but none aligns with ours enough to make any possible quantitative comparisons.", + "bbox": [ + 212, + 191, + 784, + 267 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation on blender dataset: In Fig. 5, for both Scene 1a and 1b, our model correctly forecasts the lateral movement of the cylinder to be in either position approximately $50\\%$ of the time, considering a left viewing angle. In Scene 2, with the absence of the red cylinder in the input camera angle, the model predicts the potential existence of the red cylinder approximately $50\\%$ of the time, and predicts lateral movements with roughly equal probability. This validates PC-VAE's ability to predict and infer occlusions in the latent space, aligning with human intuitions. These intuitions, shown in the Blender dataset's simple scenes, can transfer to driving scenarios in our CARLA datasets.", + "bbox": [ + 212, + 282, + 785, + 417 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/688dd54c0d8771ce3865e20ca2164cf80a4e372f3027a55385a62270d4e89b8f.jpg", + "image_caption": [ + "Fig. 7: PC-VAE reconstructions. The encoder input, $I_{c}^{t}$ , among the other ground truth images $I_{c}$ viewed from camera pose $c$ at different timestamps, is reconstructed across a new set of poses $c''$ respecting timestamp $t$ , generating $I_{c''}^{t}$ . A complete grid is in Appendix ??" + ], + "image_footnote": [], + "bbox": [ + 287, + 455, + 717, + 517 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "PC-VAE performance and ablations: We evaluate the performance of PC-VAE on CARLA datasets with multiple encoder architectures. We show that PC-VAE effectively reconstructs complex environments involving variable scenes, actor configurations, and environmental noise given potentially partially observable inputs (see Fig. 7). We calculated an average Peak Signal-to-Noise Ratio (PSNR) over the training data, as well as novel view encoder inputs. To evaluate the quality of the latent space generated by the encoder, we utilize t-SNE (24) plots to visualize the distribution of latent samples for each image in a given dataset (see Appendix ??). We introduce a Support Vector Machine (SVM) (14) based metric to measure the visualized clustering quantitatively, where a higher value indicates better clustering based on timestamps. Most latent scene samples are separable by timestamps, which indicates that the latents are view-invariant. Samples that are misclassified or lie on the boundary usually represent partially or fully occluded regions. This is desirable for forecasting, as it enables", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting", + "bbox": [ + 218, + 113, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/d982bedbd0b3d0ee8c0515e595e4a074fdf7bfc74a5af62c539fc9058593886a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Ground Truth Prediction PairAvg. PSNR (Scene 1)Avg. PSNR (Scene 2)
Single-Scene Intersection
Matching Pairs29.06N.A
Un-matching P.24.01N.A
Multi-Scene Intersection
Matching Pairs28.0028.26
Un-matching P.23.2724.56
Multi-Scene Two Lane Merge
Matching Pairs28.1428.17
Un-matching P.22.7423.32
", + "bbox": [ + 223, + 150, + 485, + 366 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/4183dfeb7507b91a6da4e5c1276baa0819159b41dc44868512366ef263d09329.jpg", + "table_caption": [ + "Table 2: Averaged PSNR for fully observable 3D predictions. CARFF correctly predicts scene evolution across all timestamps for each dataset. The average PSNR is high for predictions $\\hat{I_{t_i}}$ and matching ground truths, $I_{t_i}$ . PSNR values for incorrect correspondences, $\\hat{I_{t_i}}, I_{t_j}$ , is a result of matching surroundings. See complete table in Appendix ??" + ], + "table_footnote": [], + "table_body": "
Multi-Scene Intersection
Controller TypeActorNo Actor
Underconfident30/300/30
Overconfident0/3030/30
CARFF (n=2)17/3030/30
CARFF (n=10)30/3030/30
CARFF (n=35)30/3019/30
Multi-Scene Two Lane Merge
Controller TypeFastSlow
Underconfident30/300/30
Overconfident0/3030/30
CARFF (n=2)21/3030/30
CARFF (n=10)30/3030/30
CARFF (n=35)30/3022/30
", + "bbox": [ + 513, + 143, + 782, + 364 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 3: Planning in 3D with controllers with varying sampling numbers $\\mathbf{n}$ . CARFF-based controllers outperform baselines in success rate over 30 trials. For $n = 10$ , the CARFF-based controller consistently chooses the optimal action in potential collision scenarios. To maintain consistency, we use one single image input across 30 trials.", + "bbox": [ + 504, + 375, + 787, + 501 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "us to model probabilistic behavior over these samples. In this process, balancing KL divergence weight scheduling maintains the quality of the PC-VAE's latent space and reconstructions (see Appendix ??). Additionally, we substantiate the benefits of our PC-VAE encoder architecture through our ablations (see Appendix ??).", + "bbox": [ + 212, + 532, + 787, + 608 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "3D novel view synthesis: Given an unseen ego-centric view with potentially partial observations, our method maintains all possible current state beliefs in 3D, and faithfully reconstructs novel views from arbitrary camera angles for each belief. Fig. 2 illustrates one of the possible 3D beliefs that CARFF holds. This demonstrates our method's ability to generate 3D beliefs that could be used for novel view synthesis in a view-consistent manner. Our model's ability to achieve accurate and complete 3D environmental understanding is important for applications like prediction-based planning.", + "bbox": [ + 212, + 625, + 787, + 748 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Inference under full and partial observations: Under full observation, we use MDN to predict the subsequent car positions in all three datasets. PSNR values are calculated based on bird-eye view NeRF renderings and ground truth bird-eye view images of the scene across different timestamps. In Tab. 2 we report the PSNR values for rendered images over the predicted posterior with the ground", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "truth images at each timestamp. We also evaluate the efficacy of our prediction model using the accuracy curve given in Fig. 8. This represents CARFF's ability to generate stable beliefs, without producing incorrect predictions, based on actor(s) localization results. For each number of samples between $n = 0$ to $n = 50$ , we choose a random subset of 3 fully observable ego images and take an average of the accuracies. In scenarios with partial observable ego-centric images where several plausible scenarios exist, we utilize recall instead of accuracy using a similar setup. This lets us evaluate the encoder's ability to avoid false negative predictions of potential danger.", + "bbox": [ + 212, + 146, + 782, + 282 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Fig. 8 shows that our model achieves high accuracy and recall in both datasets, demonstrating the ability to model state uncertainty (Approaching Intersection) and dynamic uncertainty (Two Lane Merge). The results indicate CARFF's resilience against randomness in resampling, and completeness in probabilistic modeling of the belief space. Given these observations, we now build a reliable controller to plan and navigate through complex scenarios.", + "bbox": [ + 212, + 282, + 784, + 375 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Planning", + "text_level": 1, + "bbox": [ + 215, + 393, + 336, + 409 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In all our experiments, the ego vehicle must make decisions to advance under certain observability. The scenarios are designed such that the ego views contain partial occlusion and the state of the actor(s) is uncertain in some scenarios.", + "bbox": [ + 212, + 417, + 782, + 460 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In order to facilitate decision-making using CARFF, we design a controller that takes ego-centric input images and outputs an action. Decisions are made incorporating sample consistency from the mixture density network. For instance, the controller infers occlusion and promotes the ego car to pause when scenes alternate between actor presence and absence in the samples. We use the two multi-scene datasets to assess the performance of the CARFF-based controller as they contain actors with potentially unknown behaviors.", + "bbox": [ + 212, + 462, + 784, + 566 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To design an effective controller, we need to find a balance between accuracy and recall (see Fig. 8). A lowered accuracy from excessive sampling means unwanted randomness in the predicted state. However, taking insufficient samples would generate low recall i.e., not recovering all plausible states. This would lead to incorrect predictions as we would be unable to account for the plausible uncertainty present in the environment. To achieve optimal balance, we designed an open-loop planning controller using a sampling strategy that generates $n = 2, 10, 35$ samples. The hyperparameter $n$ is tuned per scene for peak performance but is expected to remain relatively stable across scenes. We demonstrate that $n = 10$ performs well consistently in varying CARLA scenarios (Fig. 8) and do not anticipate this being very different for other experiments.", + "bbox": [ + 212, + 568, + 785, + 733 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "For sampling values that lie on the borders of the accuracy and recall margin, for example, $n = 2$ and 35, we see that the CARFF-based controller obtains lower success rates, whereas $n = 10$ produces the best result. For actor exists and fast-actor scenes in Tab. 3, we consider occluded ego-centric inputs to test the controller's ability to avoid collisions. For no-actor and slow-actor scenes, we consider state observability and test the controllers' ability to recognize the optimal action to advance. Across the two datasets, the overconfident controller", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting", + "bbox": [ + 218, + 113, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7e4fbc87d5e4654bd04e8bafc34f14bdc37e5f360d7db39e2aeb0852c4c0f8f8.jpg", + "image_caption": [ + "(a) Intersection" + ], + "image_footnote": [], + "bbox": [ + 349, + 141, + 500, + 239 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9d432054d0fdcabc8b1e68591090396a8dfca8f51b511df57b933c4cde202da2.jpg", + "image_caption": [ + "(b) Lane Merge", + "Fig. 8: Multi-Scene dataset accuracy and recall curves from predicted beliefs. We test our framework across $n = 1$ and $n = 50$ samples from MDN's predicted latent distributions from ego-centric image input. Across the number of samples $n$ , we achieve an ideal margin of belief state coverage generated under partial observation (recall), and the proportion of correct beliefs sampled under full observation (accuracy). As we significantly increase the number of samples, the accuracy starts to decrease due to randomness in latent distribution resampling." + ], + "image_footnote": [], + "bbox": [ + 501, + 141, + 651, + 239 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "will inevitably experience collisions in case of a truck approaching, since it does not cautiously account for occlusions. On the other hand, an overly cautious approach results in stasis, inhibiting the controller's ability to advance in the scene. This nuanced decision-making using CARFF-based controller is especially crucial in driving scenarios, as it enhances safety and efficiency by adapting to complex and unpredictable road environments, thereby fostering a more reliable and human-like response in autonomous vehicles.", + "bbox": [ + 212, + 396, + 787, + 502 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Discussion", + "text_level": 1, + "bbox": [ + 215, + 523, + 354, + 539 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations: Like other NeRF-based methods, CARFF currently relies on posed images of specific scenes such as road intersections, limiting its direct applicability to unseen environments. However, we anticipate enhanced generalizability with the increasing deployment of cameras around populated areas, such as traffic cameras at intersections. Additionally, handling very complex dynamics with an extremely large number of actors still poses a challenge for our method, requiring per-scene optimization to balance comprehensive dynamics modeling against accuracy. Potentially stronger models in the near future may offer a promising avenue for further enhancements in this regard.", + "bbox": [ + 212, + 554, + 787, + 690 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Conclusion: We present CARFF, a novel method for probabilistic 3D scene forecasting from partial observations. By employing a Pose-Conditional VAE, a NeRF conditioned on the learned posterior, and a mixture density network that forecasts future scenes, we effectively model, predict, and plan in complex environments with state and dynamics uncertainty in a POMDP. We further demonstrate the capabilities of our method in simulated autonomous driving scenarios. Overall, CARFF offers an intuitive framework to perceiving, forecasting, and acting under uncertainty that could prove invaluable for vision-based algorithms in unstructured environments.", + "bbox": [ + 212, + 703, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Bibliography", + "text_level": 1, + "bbox": [ + 426, + 141, + 576, + 162 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Adamkiewicz, M., Chen, T., Caccavale, A., Gardner, R., Culbertson, P., Bohg, J., Schwager, M.: Vision-only robot navigation in a neural radiance world. IEEE Robotics and Automation Letters 7(2), 4606-4613 (2022)", + "[2] Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Int. Conf. Comput. Vis. pp. 5855–5864 (2021)", + "[3] Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 130-141 (2023)", + "[4] Cao, J., Wang, X., Darrell, T., Yu, F.: Instance-aware predictive navigation in multi-agent environments. In: IEEE Int. Conf. on Robotics and Automation. pp. 5096-5102. IEEE (2021)", + "[5] Charatan, D., Li, S., Tagliasacchi, A., Sitzmann, V.: pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction (2023)", + "[6] Codevilla, F., Santana, E., López, A.M., Gaidon, A.: Exploring the limitations of behavior cloning for autonomous driving. In: Int. Conf. Comput. Vis. pp. 9329-9338 (2019)", + "[7] Deng, K., Liu, A., Zhu, J.Y., Ramanan, D.: Depth-supervised nerf: Fewer views and faster training for free. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12882-12891 (2022)", + "[8] Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Int. Conf. Learn. Represent. (2021)", + "[9] Dosovitskiy, A., Ros, G., Codevilla, F., Lopez, A., Koltun, V.: CARLA: An open urban driving simulator. In: Conf. on Robol Learning. pp. 1-16 (2017)", + "[10] Driess, D., Huang, Z., Li, Y., Tedrake, R., Toussaint, M.: Learning multi-object dynamics with compositional neural radiance fields. arXiv preprint arXiv:2202.11855 (2022)", + "[11] Fei, B., Xu, J., Zhang, R., Zhou, Q., Yang, W., He, Y.: 3d gaussian as a new vision era: A survey (2024)", + "[12] Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5501–5510 (2022)", + "[13] Hausknecht, M., Stone, P.: Deep recurrent q-learning for partially observable mdps. In: AAAI (2015)", + "[14] Hearst, M.A., Dumais, S.T., Osuna, E., Platt, J., Scholkopf, B.: Support vector machines. IEEE Intelligent Systems and their applications 13(4), 18-28 (1998)", + "[15] Ichnowski, J., Avigal, Y., Kerr, J., Goldberg, K.: Dex-nerf: Using a neural radiance field to grasp transparent objects. arXiv preprint arXiv:2110.14217 (2021)" + ], + "bbox": [ + 217, + 210, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Ivanovic, B., Elhafsi, A., Rosman, G., Gaidon, A., Pavone, M.: Mats: An interpretable trajectory forecasting representation for planning and control. In: Conf. on Robol Learning (2021)", + "[17] Kaelbling, L.P., Littman, M.L., Cassandra, A.R.: Planning and acting in partially observable stochastic domains. Artificial Intelligence 101(1), 99-134 (1998). https://doi.org/https://doi.org/10.1016/S0004-3702(98)00023-X, https://www.sciencedirect.com/science/article/pii/S000437029800023X", + "[18] Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/", + "[19] Kerr, J., Fu, L., Huang, H., Avigal, Y., Tancik, M., Ichnowski, J., Kanazawa, A., Goldberg, K.: Evo-nerf: Evolving nerf for sequential robot grasping of transparent objects. In: Conf. on Robol Learning (2022)", + "[20] Kosiorek, A.R., Strathmann, H., Zoran, D., Moreno, P., Schneider, R., Mokrá, S., Rezende, D.J.: NeRF-VAE: A geometry aware 3d scene generative model. In: ICML. pp. 5742–5752 (2021)", + "[21] Li, Y., Li, S., Sitzmann, V., Agrawal, P., Torralba, A.: 3d neural scene representations for visuomotor control. In: Conf. on Robol Learning. pp. 112-123 (2022)", + "[22] Liu, J.W., Cao, Y.P., Mao, W., Zhang, W., Zhang, D.J., Keppo, J., Shan, Y., Qie, X., Shou, M.Z.: Devrf: Fast deformable voxel radiance fields for dynamic scenes. In: Adv. Neural Inform. Process. Syst. vol. 35, pp. 36762-36775 (2022)", + "[23] Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis (2023)", + "[24] van der Maaten, L., Hinton, G.: Visualizing data using t-SNE. Journal of Machine Learning Research 9, 2579-2605 (2008), http://www.jmlr.org/papers/v9/vandermaaten08a.html", + "[25] Marza, P., Matignon, L., Simonin, O., Wolf, C.: Multi-object navigation with dynamically learned neural implicit representations. In: Int. Conf. Comput. Vis. pp. 11004-11015 (2023)", + "[26] McAllister, R., Rasmussen, C.E.: Data-efficient reinforcement learning in continuous state-action gaussian-pomdpds. In: Adv. Neural Inform. Process. Syst. (2017)", + "[27] Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: Eur. Conf. Comput. Vis. (2020)", + "[28] Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 1-15 (2022)", + "[29] Packer, C., Rhinehart, N., McAllister, R.T., Wright, M.A., Wang, X., He, J., Levine, S., Gonzalez, J.E.: Is anyone there? learning a planner contingent on perceptual uncertainty. In: Conf. on Robol Learning (2022)" + ], + "bbox": [ + 215, + 145, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Packer, C., Rhinehart, N., McAllister, R.T., Wright, M.A., Wang, X., He, J., Levine, S., Gonzalez, J.E.: Is anyone there? learning a planner contingent on perceptual uncertainty. In: Liu, K., Kulic, D., Ichnowski, J. (eds.) Conf. on Robol Learning. pp. 1607-1617 (2023)", + "[31] Pan, X., You, Y., Wang, Z., Lu, C.: Virtual to real reinforcement learning for autonomous driving. arXiv preprint arXiv:1704.03952 (2017)", + "[32] Papadimitriou, C.H., Tsitsiklis, J.N.: The complexity of markov decision processes. Mathematics of operations research 12(3), 441-450 (1987)", + "[33] Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: Int. Conf. Comput. Vis. pp. 5865–5874 (2021)", + "[34] Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph. (2021)", + "[35] Pineau, J., Gordon, G., Thrun, S., et al.: Point-based value iteration: An anytime algorithm for pomdpps. In: IJCAI. vol. 3, pp. 1025-1032 (2003)", + "[36] Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-NeRF: Neural Radiance Fields for Dynamic Scenes. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)", + "[37] Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2016)", + "[38] Rhinehart, N., He, J., Packer, C., Wright, M.A., McAllister, R., Gonzalez, J.E., Levine, S.: Contingencies from observations: Tractable contingency planning with learned behavior models. In: IEEE Int. Conf. on Robotics and Automation. pp. 13663-13669 (2021)", + "[39] Roessle, B., Barron, J.T., Mildenhall, B., Srinivasan, P.P., Nießner, M.: Dense depth priors for neural radiance fields from sparse input views. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12892–12901 (2022)", + "[40] Shao, R., Zheng, Z., Tu, H., Liu, B., Zhang, H., Liu, Y.: Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 16632-16642 (2023)", + "[41] Sun, C., Sun, M., Chen, H.T.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5459-5469 (2022)", + "[42] Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8248–8258 (2022)", + "[43] Tancik, M., Srinivasan, P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. In: Adv. Neural Inform. Process. Syst. pp. 7537–7547 (2020)" + ], + "bbox": [ + 215, + 145, + 784, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting", + "bbox": [ + 218, + 113, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Tas, O.S., Stiller, C.: Limited visibility and uncertainty aware motion planning for automated driving. In: IEEE Intelligent Vehicles Symposium (IV) (jun 2018)", + "[45] Tewari, A., Thies, J., Mildenhall, B., Srinivasan, P., Tretschk, E., Yifan, W., Lassner, C., Sitzmann, V., Martin-Brualla, R., Lombardi, S., Simon, T., Theobalt, C., Nießner, M., Barron, J.T., Wetzstein, G., Zollhöfer, M., Golyanik, V.: Advances in Neural Rendering. Comput. Graph. Forum (2022)", + "[46] Toromanoff, M., Wirbel, E., Moutarde, F.: End-to-end model-free reinforcement learning for urban driving using implicit affordances. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7153-7162 (2020)", + "[47] Tretschk, E., Golyanik, V., Zollhoefer, M., Bozic, A., Lassner, C., Theobalt, C.: Scenerflow: Time-consistent reconstruction of general dynamic scenes. In: International Conference on 3D Vision (3DV) (2023)", + "[48] Truong, P., Rakotosaona, M.J., Manhardt, F., Tombari, F.: Sparf: Neural radiance fields from sparse and noisy poses. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4190-4200 (2023)", + "[49] Turki, H., Ramanan, D., Satyanarayanan, M.: Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12922–12931 (2022)", + "[50] Wang, Q., Wang, Z., Genova, K., Srinivasan, P.P., Zhou, H., Barron, J.T., Martin-Brualla, R., Snavely, N., Funkhouser, T.: Ibrnet: Learning multi-view image-based rendering. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4690-4699 (2021)", + "[51] Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5438-5448 (2022)", + "[52] Yen-Chen, L., Florence, P., Barron, J.T., Rodriguez, A., Isola, P., Lin, T.Y.: inerf: Inverting neural radiance fields for pose estimation. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1323-1330. IEEE (2021)", + "[53] Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images (2021)", + "[54] Yu, M.Y., Vasudevan, R., Johnson-Roberson, M.: Occlusion-aware risk assessment for autonomous driving in urban environments. IEEE Robotics and Automation Letters 4(2), 2235-2241 (Apr 2019). https://doi.org/10.1109/lra.2019.2900453, http://dx.doi.org/10.1109/LRA.2019.2900453", + "[55] Zhang, C., Steinhauser, F., Hinz, G., Knoll, A.: Improved occlusion scenario coverage with a pomdp-based behavior planner for autonomous urban driving. In: 2021 IEEE International Intelligent Transportation Systems Conference (ITSC). pp. 593-600 (2021). https://doi.org/10.1109/ITSC48978.2021.9564424", + "[56] Zhang, C., Steinhauser, F., Hinz, G., Knoll, A.: Occlusion-aware planning for autonomous driving with vehicle-to-everything communication. IEEE Transactions on Intelligent Vehicles 9(1), 1229-1242 (2024). https://doi.org/10.1109/TIV.2023.3308098" + ], + "bbox": [ + 215, + 145, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "J. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/CARFF_ Conditional Auto-encoded Radiance Field for 3D Scene Forecasting/4d2110cb-f049-4580-b1d7-c4b728d00449_model.json b/2024/CARFF_ Conditional Auto-encoded Radiance Field for 3D Scene Forecasting/4d2110cb-f049-4580-b1d7-c4b728d00449_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a804bad92f9337291173bca268e8b65a03ecef34 --- /dev/null +++ b/2024/CARFF_ Conditional Auto-encoded Radiance Field for 3D Scene Forecasting/4d2110cb-f049-4580-b1d7-c4b728d00449_model.json @@ -0,0 +1,2359 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.243, + 0.141, + 0.761, + 0.187 + ], + "angle": 0, + "content": "CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting" + }, + { + "type": "text", + "bbox": [ + 0.244, + 0.219, + 0.761, + 0.266 + ], + "angle": 0, + "content": "Jiezhi \"Stephen\" Yang\\(^{1}\\), Khushi Desai\\(^{2}\\), Charles Packer\\(^{3}\\), Harshil Bhatia\\(^{4}\\), Nicholas Rhinehart\\(^{3}\\), Rowan McAllister\\(^{5}\\), and Joseph E. Gonzalez\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.374, + 0.278, + 0.631, + 0.291 + ], + "angle": 0, + "content": "1 Harvard University MA 02138, USA" + }, + { + "type": "text", + "bbox": [ + 0.371, + 0.292, + 0.633, + 0.306 + ], + "angle": 0, + "content": "2 Columbia University NY 10025, USA" + }, + { + "type": "text", + "bbox": [ + 0.398, + 0.307, + 0.607, + 0.319 + ], + "angle": 0, + "content": "UC Berkeley CA 94720, USA" + }, + { + "type": "text", + "bbox": [ + 0.399, + 0.32, + 0.605, + 0.333 + ], + "angle": 0, + "content": "4 Avataar.ai KA 560103, India" + }, + { + "type": "text", + "bbox": [ + 0.353, + 0.334, + 0.65, + 0.347 + ], + "angle": 0, + "content": "5 Toyota Research Institute CA 94022, USA" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.386, + 0.744, + 0.609 + ], + "angle": 0, + "content": "Abstract. We propose CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting, a method for predicting future 3D scenes given past observations. Our method maps 2D ego-centric images to a distribution over plausible 3D latent scene configurations and predicts the evolution of hypothesized scenes through time. Our latents condition a global Neural Radiance Field (NeRF) to represent a 3D scene model, enabling explainable predictions and straightforward downstream planning. This approach models the world as a POMDP and considers complex scenarios of uncertainty in environmental states and dynamics. Specifically, we employ a two-stage training of Pose-Conditional-VAE and NeRF to learn 3D representations, and auto-regressively predict latent scene representations utilizing a mixture density network. We demonstrate the utility of our method in scenarios using the CARLA driving simulator, where CARFF enables efficient trajectory and contingency planning in complex multi-agent autonomous driving scenarios involving occlusions. Video and code are available at: www.carff website." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.634, + 0.376, + 0.649 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.666, + 0.788, + 0.786 + ], + "angle": 0, + "content": "Humans often imagine what they cannot see given partial visual context. Consider a scenario where reasoning about the unobserved is critical to safe decision-making: for example, a driver navigating a blind intersection. An expert driver will plan according to what they believe may or may not exist in occluded regions of their vision. The driver's belief - defined as the understanding of the world modeled with consideration for inherent environment uncertainties - is informed by their partial observations (i.e., the presence of other vehicles on the road), as well as their prior knowledge (e.g., past experience navigating this intersection)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.787, + 0.788, + 0.817 + ], + "angle": 0, + "content": "When reasoning about the unobserved, humans form complex beliefs about the existence, position, shapes, colors, and textures of occluded scene portions" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.355, + 0.839 + ], + "angle": 0, + "content": "* Core contributors" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.207 + ], + "angle": 0, + "content": "(e.g., an oncoming car). Autonomous systems with high-dimensional sensor data, like video or LiDAR, traditionally reduce this data to low-dimensional state information (e.g., position and velocity of tracked objects) for prediction and planning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.784, + 0.268 + ], + "angle": 0, + "content": "In addition to tracking fully observed objects, this object-centric framework handles partially observed settings by considering potentially dangerous unobserved objects. These systems often plan for worst-case scenarios, such as a \"ghost car\" at the edge of the visible field of view (44)." + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.298, + 0.675, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.426, + 0.788, + 0.51 + ], + "angle": 0, + "content": "Fig. 1: CARFF 3D planning application for driving. An input image containing a partially observable view of an intersection is processed by CARFF's encoder to establish 3D environment state beliefs, i.e. the predicted possible state of the world: whether or not there could be another vehicle approaching the intersection. These beliefs are used to forecast the future in 3D for planning, generating one among two possible actions for the vehicle to merge into the other lane." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Recent advances in neural rendering, particularly Neural Radiance Fields (NeRF), have significantly improved 3D scene representation learning. NeRF enables novel view synthesis, thus simplifying the process of viewing behind occlusions. NeRF decouples the dependency of scene representation from traditional object detection and tracking, allowing for the capture of vital visual information that might be missed by detectors, yet is crucial for safe decision-making. NeRF's implicit density representation of explicit geometry also facilitates its direct application in motion planning without the need for rendering. NeRF's ability to represent both visual and geometric information makes them a more general and intuitive 3D representation for autonomous systems." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Despite NeRF's advantages, achieving probabilistic predictions in 3D based on reasoning from occluded views is challenging. For example, discriminative models that yield categorical predictions are unable to capture the underlying 3D structure, impeding their ability to model uncertainty. While prior work on 3D representation captures view-invariant structures, their application is primarily confined to simple scenarios (20). We present CARFF, which to our knowledge, is the first forecasting approach in scenarios with partial observations that uniquely facilitates stochastic predictions in a partially observable Markov decision process (POMDP) within a 3D representation, effectively integrating visual perception and geometry. Specifically, we make the following contributions:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.147, + 0.786, + 0.207 + ], + "angle": 0, + "content": "1. We propose a novel architecture \\(PC\\)-VAE: Pose-Conditioned Variational Autoencoder. The encoder maps potentially partially observable ego-centric images to pose-invariant latent scene representations, which hold state beliefs of the POMDP with implicit probability distributions (see Sec. 3.1)." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.208, + 0.787, + 0.265 + ], + "angle": 0, + "content": "2. We develop the two-stage training pipeline that uniquely enables complex scene modeling with a probabilistic objective. This involves separately training the PC-VAE and a latent conditioned neural radiance field that functions as a 3D decoder, enabling interpretable predictions (see Sec. 3.1)." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.267, + 0.786, + 0.325 + ], + "angle": 0, + "content": "3. We design a mixture density model to predict the evolution of 3D scenes over time stochastically and regressively in the encoder belief space (see Sec. 3.2). This allows for an effective sampling based-controller to output actions in the POMDP." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.333, + 0.786, + 0.394 + ], + "angle": 0, + "content": "We demonstrate how CARFF can be used to enable contingency planning in complex driving scenarios that require reasoning into visual occlusions on CARLA simulated datasets inspired by autonomous driving planning tasks (29; 54; 55; 56). A potential application of CARFF is illustrated in Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.413, + 0.383, + 0.429 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.441, + 0.51, + 0.457 + ], + "angle": 0, + "content": "2.1 NeRF and 3D representations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.646 + ], + "angle": 0, + "content": "Neural radiance fields. Neural Radiance Fields (NeRF) (2; 27; 43) for 3D representations generate high-resolution, photorealistic scenes. Instant Neural Graphics Primitive (Instant-NGP) (28) speeds up training and rendering time by introducing a multi-resolution hash encoding. Other works like Plenoxels (12) and DirectVoxGo (DVGO) (41) also provide similar speedups. Recent advancements in volumetric representations such as 3D Gaussian Splatting (18) enhance rendering efficiency while maintaining compatibility with traditional NeRF applications (11). We utilize Instant-NGP for its accessibility, although our approach is adaptable to alternative rendering methods. NeRFs have also been extended for several tasks such as modeling large-scale unbounded scenes (2; 42; 49), scene from sparse views (7; 39; 48) and multiple scenes (20; 50). For an in-depth survey on neural representation learning and its applications we refer the reader to (45)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Generalizable novel view synthesis models, like pixelNeRF and pixelSplat (5; 53), learn a scene prior to render novel views from sparse existing ones. In contrast, CARFF is based on a VAE, encoding a probabilistic objective and decoding to future 3D scenes. Dynamic NeRF models scenes with moving or deforming objects, within which a widely used approach is to construct a canonical space and predict a deformation field (22; 33; 34; 36). The canonical space is usually a static scene, and the model learns an implicitly represented flow field (33; 36). A recent line of work also models dynamic scenes via different representations and decomposition (3; 40). These approaches tend to perform better for spatially bounded and predictable scenes with relatively small variations (3; 23; 33; 53). Moreover, these methods only solve for changes in the environment but are limited in incorporating stochasticity in the environment." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "image", + "bbox": [ + 0.394, + 0.148, + 0.498, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.214, + 0.476, + 0.221 + ], + "angle": 0, + "content": "Ego-centric Lane View" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.148, + 0.616, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.214, + 0.578, + 0.221 + ], + "angle": 0, + "content": "Predicted 3D View" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.288 + ], + "angle": 0, + "content": "Fig. 2: Novel view planning application. CARFF allows reasoning behind occluded views from the ego car as simple as moving the camera to see the sampled belief predictions, allowing simple downstream planning using, for example, density probing or 2D segmentation models from arbitrary angles." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.324, + 0.788, + 0.43 + ], + "angle": 0, + "content": "Multi-scene NeRF: Our approach builds on multi-scene NeRF approaches (20; 47; 50; 51) that learn a global latent scene representation, which conditions the NeRF, allowing a single NeRF to effectively represent various scenes. A similar method, NeRF-VAE, was introduced by Kosiorek et al. (20) to create a geometrically consistent 3D generative model with generalization to out-of-distribution cameras. However, NeRF-VAE (20) is prone to mode collapse when faced with complex visual information (see Sec. 4.2)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.468, + 0.411, + 0.484 + ], + "angle": 0, + "content": "2.2 Scene Forecasting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.645 + ], + "angle": 0, + "content": "Planning in 2D space: Planning in large, continuous state-action spaces is challenging due to exponentially large search spaces (32), leading to various approximation methods for tractability (26; 35). Model-free (13; 31; 46) and model-based (4) reinforcement learning frameworks, along with other learning-based methods (6; 29), have emerged as viable approaches. Additionally, methods forecast for downstream control (16), learn behavior models for contingency planning (38), or predict the existence and intentions of unobserved agents (30). While these methods operate in 2D, we reason under partial observations and account for these factors in 3D." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.84 + ], + "angle": 0, + "content": "NeRF in robotics: Recent works have applied NeRFs in robotics for localization (52), navigation (1; 25), dynamics modeling (10; 22), and robotic grasping (15; 19). Adamkiewicz et al.(1) propose quadcopter motion planning in NeRF models by sampling the learned density function, useful for forecasting and planning. Driess et al.(10) employ a graph neural network to learn dynamics in a multi-object NeRF scene. Li et al. (21) focus on pushing tasks and address grasping and planning with NeRF and a separate latent dynamics model. Prior approaches work in simple, static scenes (1) or uses deterministic dynamics models (21). CARFF addresses complex, realistic environments with both state and dynamics uncertainty, considering potential object existence and unknown movements." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.331, + 0.161 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.311 + ], + "angle": 0, + "content": "Recent advancements in 3D scene representation allow for modeling environments in a contextually rich and interactive 3D space. This offers analytical benefits, such as spatial analysis with soft occupancy grids and object detection through novel view synthesis. Given these advantages, our primary objective is to develop a model for probabilistic 3D scene forecasting in dynamic environments. However, direct integration of 3D scene representation via NeRF and probabilistic models like VAE often involves non-convex and inter-dependent optimization, which causes unstable training. For instance, NeRF's optimization may rely on the VAE's latent space being structured to provide informative gradients. To" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.335, + 0.761, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.507, + 0.788, + 0.646 + ], + "angle": 0, + "content": "Fig. 3: Visualizing CARFF's two stage training process. Left: The convolutional VIT-based encoder encodes each image \\(I\\) at timestamps \\(t, t'\\) and camera poses \\(c, c'\\) into Gaussian latent distributions. Assuming two timestamps and an overparameterized latent, one Gaussian distribution will have a smaller \\(\\sigma^2\\), and different \\(\\mu\\) across timestamps. Upper Right: The pose-conditional decoder stochastically decodes the sampled latent \\(z\\) using the camera pose \\(c''\\) into images \\(I_{c''}^t\\) and \\(I_{c''}^{t'}\\). The decoded reconstruction and ground truth images are used for the loss \\(\\mathcal{L}_{\\mathrm{MSE}}\\), PC-VAE. Lower Right: A NeRF is trained by conditioning on the latent variables sampled from the optimized Gaussian parameters. These parameters characterize the distinct timestamp distributions derived from the PC-VAE. An MSE loss is calculated for NeRF as \\(\\mathcal{L}_{\\mathrm{MSE}}\\), NeRF." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "navigate these complexities, our method bifurcates the training process into two stages (see Fig. 3). First, we train the PC-VAE to learn view-invariant scene representations. Next, we replace the decoder with a NeRF to learn a 3D scene from the latent representations. The latent scene representations capture the environmental states and dynamics over possible underlying scenes, while NeRF synthesizes novel views within the belief space, giving us the ability to see the unobserved (see Fig. 2 and Sec. 3.1). During prediction, uncertainties can be modeled by sampling latents auto-regressively from a predicted Gaussian mixture, allowing for effective decision-making. To this extent, we approach scene forecasting as a POMDP over latent distributions, which enables us to capture multi-modal beliefs for planning amidst perceptual uncertainty (see Sec. 3.2)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.631, + 0.163 + ], + "angle": 0, + "content": "3.1 Pose-Conditional VAE (PC-VAE) and NeRF" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.395 + ], + "angle": 0, + "content": "Architecture: We assume that the model follows a Markovian process, and thus each belief state only depends on the previous. Given a scene \\( S_{t} \\) at timestamp \\( t \\), we have an ego-centric observation image \\( I_{c}^{t} \\) captured from camera pose \\( c \\). The objective is to formulate a 3D representation of the image that holds implicit probability distributions of the possible states, where we can perform a forecasting step that evolves the scene forward. Here, the POMDP can be seen as an MDP in belief space (17). To achieve this, we utilize a radiance field conditioned on latent variable \\( z \\) sampled from the posterior distribution \\( q_{\\phi}(z|I_{c}^{t}) \\). Now, to learn the posterior, we utilize PC-VAE. We construct an encoder using convolutional layers and a pre-trained ViT on ImageNet (8). The encoder learns a mapping from the image space to a Gaussian distributed latent space \\( q_{\\phi}(z|I_{c}^{t}) = \\mathcal{N}(\\mu, \\sigma^{2}) \\) parametrized by mean \\( \\mu \\) and variance \\( \\sigma^2 \\). The decoder, \\( p(I|z,c) \\), conditioned on camera pose \\( c \\), maps the latent \\( z \\sim \\mathcal{N}(\\mu, \\sigma^2) \\) into the image space \\( I \\). This helps the encoder to generate latents that are invariant to the camera pose \\( c \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.788, + 0.457 + ], + "angle": 0, + "content": "To enable 3D scene modeling, we employ Instant-NGP (28), which incorporates a hash grid and an occupancy grid to enhance computation efficiency. Additionally, a smaller multilayer perceptron (MLP), \\( F_{\\theta}(z) \\) can be utilized to model the density and appearance, given by:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.467, + 0.786, + 0.482 + ], + "angle": 0, + "content": "\\[\nF _ {\\theta} (z): (\\mathbf {x}, \\mathbf {d}, z) \\rightarrow ((r, g, b), \\sigma) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.49, + 0.787, + 0.538 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{x} \\in \\mathbb{R}^3\\) and \\(\\mathbf{d} \\in (\\theta, \\phi)\\) represent the location vector and the viewing direction respectively. The MLP is conditioned on the sampled scene latents \\(z \\sim q_{\\phi}(z|I_c^t)\\) (see Appendix ??)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.548, + 0.788, + 0.73 + ], + "angle": 0, + "content": "Training methodology: The architecture alone does not enable us to model complex scenarios, as seen through a similar example in NeRF-VAE (20). A crucial contribution of our work is our two-stage training framework which stabilizes the training. First, we optimize the convolutional ViT based encoder and pose-conditional convolutional decoder in the pixel space for reconstruction. This enables our method to deal with more complex and realistic scenes as the encoding is learned in a semantically rich 2D space. By conditioning the decoder on camera poses, we achieve disentanglement between camera view angles and scene context, making the representation view-invariant and the encoder 3D-aware. Once rich latent representations are learned, we replace the decoder with a latent-conditioned NeRF over the latent space of the frozen encoder. The NeRF reconstructs encoder beliefs in 3D for novel view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.742, + 0.788, + 0.773 + ], + "angle": 0, + "content": "Loss: PC-VAE is trained using standard VAE loss, with mean square error (MSE) and a Kullback-Leibler (KL) divergence given by evidence lower bound:" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.802, + 0.785, + 0.837 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {P C - V A E} = \\mathcal {L} _ {M S E, P C - V A E} + \\mathcal {L} _ {K L D, P C - V A E} = \\\\ \\left\\| p \\left(I | z, c ^ {\\prime \\prime}\\right) - I _ {c ^ {\\prime \\prime}} ^ {t} \\right\\| ^ {2} + \\mathbb {E} _ {q (z \\mid I _ {c} ^ {t})} [ \\log p (I | z) ] - w _ {K L} D _ {K L} \\left(q _ {\\phi} \\left(z \\mid I _ {c} ^ {t}\\right) \\mid p (I | z)\\right) \\tag {2} \\\\ \\end{array}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.219, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "where \\( w_{KL} \\) denotes the KL divergence loss weight and \\( z \\sim q_{\\phi}(z|I_c^t) \\). To make our representation 3D-aware, our posterior is encoded using camera \\( c \\) while the decoder is conditioned on a randomly sampled pose \\( c'' \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.298 + ], + "angle": 0, + "content": "KL divergence regularizes the latent space to balance conditioned reconstruction and stochasticity under occlusion. An elevated KL divergence loss weight \\( w_{KL} \\) pushes the latents closer to a standard normal distribution, \\( \\mathcal{N}(0,1) \\), thereby ensuring probabilistic sampling in scenarios under partial observation. However, excessive regularization causes the latents to be less separable, leading to mode collapse. To mitigate this, we adopt delayed linear KL divergence loss weight scheduling to strike a balanced \\( w_{KL} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.784, + 0.343 + ], + "angle": 0, + "content": "Next, we learn a NeRF decoder on the posterior of the VAE to model scenes. At any timestamp \\( t \\) we use a standard photometric loss for training the NeRF, given by the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.344, + 0.352, + 0.785, + 0.369 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {M S E, N e R F} = \\left\\| I _ {c} ^ {t} - r e n d e r \\left(F _ {\\theta} \\left(\\cdot \\mid q _ {\\phi} \\left(z \\mid I _ {c} ^ {t}\\right)\\right)\\right) \\right\\| ^ {2} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.788, + 0.42 + ], + "angle": 0, + "content": "We use a standard rendering algorithm as proposed by Müller et al. (28). Next, we build a forecasting module over the learned latent space of our pose-conditional encoder." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.441, + 0.41, + 0.456 + ], + "angle": 0, + "content": "3.2 Scene Forecasting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.784, + 0.538 + ], + "angle": 0, + "content": "Formulation: The current formulation allows us to model scenes with different configurations across timestamps. In order to forecast future configurations of a scene given an ego-centric view, we need to predict future latent distributions. We formulate the forecasting as a POMDP over the posterior distribution \\( q_{\\phi}(z|I_c^t) \\) in the PC-VAE's latent space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.784, + 0.614 + ], + "angle": 0, + "content": "During inference, we observe stochastic behaviors under occlusion, which motivates us to learn a mixture of several Gaussian distributions that potentially denote different scene possibilities. Therefore, we model the POMDP using a Mixture Density Network (MDN), with multi-headed MLPs, that predicts a mixture of \\( K \\) Gaussians. At any timestamp \\( t \\) the distribution is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.376, + 0.623, + 0.785, + 0.64 + ], + "angle": 0, + "content": "\\[\nq _ {\\phi} ^ {\\prime} \\left(z _ {t} \\mid I _ {c} ^ {t - 1}\\right) = M D N \\left(q _ {\\phi} \\left(z _ {t - 1} \\mid I _ {c} ^ {t - 1}\\right)\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.647, + 0.784, + 0.692 + ], + "angle": 0, + "content": "The model is conditioned on the posterior distribution \\( q_{\\phi}(z_{t - 1}) \\) to learn a predicted posterior distribution \\( q_{\\phi}'(z_t|I_c^{t - 1}) \\) at each timestamp. The predicted posterior distribution is given by the mixture of Gaussian:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.702, + 0.785, + 0.74 + ], + "angle": 0, + "content": "\\[\nq _ {\\phi} ^ {\\prime} \\left(z _ {t}\\right) = \\sum_ {i = 1} ^ {K} \\pi_ {i} \\mathcal {N} \\left(\\mu_ {i}, \\sigma_ {i} ^ {2}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.749, + 0.784, + 0.84 + ], + "angle": 0, + "content": "here, \\(\\pi_i\\), \\(\\mu_i\\), and \\(\\sigma_i^2\\) denote the mixture weight, mean, and variance of the \\(i^{th}\\) Gaussian distribution within the posterior distribution. Here, \\(K\\) is the total number of Gaussians. For brevity we remove their conditioning on the posterior \\(q_{\\phi}(z_{t-1})\\) and sampled latent \\(z_{t-1}\\). We sample \\(z_t\\) from the mixture of Gaussians \\(q_{\\phi}^{\\prime}(z_t)\\), where \\(z_t\\) likely falls within one of the Gaussian modes. The configuration corresponding to the mode is reflected in the 3D scene rendered by NeRF." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.208, + 0.348, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.208, + 0.48, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.144, + 0.582, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.592, + 0.144, + 0.675, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.145, + 0.775, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.286, + 0.482, + 0.371 + ], + "angle": 0, + "content": "Fig.4: Multi-scene CARLA datasets. Varying car configurations and scenes for the Multi-Scene Two Lane Merge dataset (left) and the Multi-Scene Approaching Intersection dataset (right)." + }, + { + "type": "image_caption", + "bbox": [ + 0.492, + 0.286, + 0.782, + 0.371 + ], + "angle": 0, + "content": "Fig.5: Blender dataset. Blender dataset with a blue cube and a potential red cylinder exhibiting probabilistic temporal movement. The possible occlusions from different camera angles demonstrate how movement needs to be modeled probabilistically." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.784, + 0.427 + ], + "angle": 0, + "content": "Loss: To optimize the MDN, we minimize a negative log-likelihood function, given by:" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.428, + 0.785, + 0.469 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {M D N} = - \\sum_ {j = 1} ^ {N} \\log \\left(\\sum_ {i = 1} ^ {K} \\pi_ {i} \\mathcal {N} \\left(y _ {j}; \\mu_ {i}, \\sigma_ {i} ^ {2}\\right)\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.785, + 0.505 + ], + "angle": 0, + "content": "where \\( y_{i} \\sim q_{\\phi}(z_{t}) \\) is sampled from the distribution of latent \\( z_{t} \\), learned by the encoder, and \\( N \\) denotes the total number of samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.517, + 0.788, + 0.64 + ], + "angle": 0, + "content": "Inference: We consider an unseen ego-centric image and retrieve its posterior \\( q_{\\phi}(z_t) \\) through the encoder. Next, we predict the possible future posterior distribution \\( q_{\\phi}'(z_{t + 1}) \\). From the predicted posterior, we sample a scene latent and perform localization. We achieve this via (a) density probing the NeRF or (b) segmenting the rendered novel views using off-the-shelf methods such as YOLO (37) (see Fig. 2). These allow us to retrieve a corresponding Gaussian distribution \\( q_{\\phi}(z_{t + 1}) \\) in encoder latent space. This is auto-regressively fed back into the MDN to predict the next timestamp. See Fig. 6 for an overview of the pipeline." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.66, + 0.325, + 0.675 + ], + "angle": 0, + "content": "4 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Decision-making under perceptual uncertainty is a pervasive challenge faced in robotics and autonomous driving, as the real environment is mostly likely partially observable, making it a POMDP. In a partially observable driving scenario, accurate inference regarding the presence of potentially obscured agents is pivotal. We evaluate the effectiveness of CARFF on common driving situations with partial observability and added complexity. We implemented several scenarios in the CARLA driving simulator (9) (see Fig. 4). A single NVIDIA RTX 3090 GPU is used to train PC-VAE, NeRF, and the MDN. All models, trained sequentially, tend to converge within a combined time frame of 24 hours. A detailed experimental setup can be found in Appendix ??. We show that, given" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "partially observable 2D inputs, CARFF performs well in predicting latent distributions that represent complete 3D scenes. Using these predictions we design a CARFF-based controller for performing downstream planning tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.214, + 0.403, + 0.228 + ], + "angle": 0, + "content": "4.1 Data Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.785, + 0.36 + ], + "angle": 0, + "content": "We conduct experiments on (a) synthetic blender dataset for principle experiments to test the probabilistic modeling capacities in isolation of the vision encoder (it is visually as simple as possible, but requires the full predictive model proposed in CARFF) and (b) CARLA-based driving datasets for more complex driving scenarios (9). To deliver convincing results, we model these driving scenarios off of related works (29; 54; 55; 56) that concern planning for driving under difficult situations. We generate the datasets in 3D by programming an ego object and varying actor objects in different configurations." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.388, + 0.773, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.678 + ], + "angle": 0, + "content": "Fig. 6: Auto-regressive inference in scene prediction. The input image at timesamp \\( t \\), \\( I_{c}^{t} \\), is encoded using the pre-trained encoder from PC-VAE. The corresponding latent distribution is fed into the Mixture Density Network, which predicts a mixture of Gaussians. Each of the \\( K \\) Gaussians is a latent distribution that may correspond to different beliefs at the next timestamp. The mixture of Gaussians is sampled repeatedly for the predicted latent beliefs, visualized as \\( I_{c',scni}^{t+1} \\), representing potentially the \\( i \\)th possible outcome. This is used to condition the NeRF to generate 3D views of the scene. To accomplish autoregressive predictions, we probe the NeRF for the location of the car and feed this information back to the pre-trained encoder to predict the scene at the next timestamp." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.785, + 0.781 + ], + "angle": 0, + "content": "Blender synthetic dataset: This comprises of a stationary blue cube (ego) accompanied by a red cylinder (actor) that may or may not be present (see Fig. 5). If the actor is present, it exhibits lateral movement as depicted in Fig. 5. This simplistic setting provides an interpretable framework to evaluate our model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "CARLA dataset: Each dataset is simulated for \\(N\\) timestamps and uses \\(C = 100\\) predefined camera poses to capture images of the environment under full observation, partial observation, and no visibility. These datasets are modeled after" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "common driving scenarios involving state uncertainty that have been proposed in related works such as Active Visual Planning (29)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.237 + ], + "angle": 0, + "content": "a) Single-Scene Approaching Intersection: The ego vehicle is positioned at a T-intersection. An actor vehicle traverses the crossing along an evenly spaced, predefined trajectory. We simulate this for \\( N = 10 \\) timestamps. We mainly use this dataset to predict the evolution of timestamps under full observation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.788, + 0.329 + ], + "angle": 0, + "content": "b) Multi-Scene Approaching Intersection: We extend the previous scenario to a more complicated setting with state uncertainty, by making the existence of the actor vehicle probabilistic. A similar intersection crossing is simulated for \\( N = 3 \\) timestamps for both possibilities. The ego vehicle's view of the actor may be occluded as it approaches the T-intersection over the \\( N \\) timestamps. The ego vehicle either moves forward or halts at the junction (see Fig. 4)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.421 + ], + "angle": 0, + "content": "c) Multi-Scene Multi-actor Two Lane Merge: To add more environment dynamics uncertainty, we consider a multi-actor setting at an intersection of two merging lanes. We simulate the scenario at an intersection with partial occlusions, with the second approaching actor having variable speed. Here the ego vehicle can either merge into the left lane before the second actor or after all the actors pass, (see Fig. 4). Each branch is simulated for \\( N = 3 \\) timestamps." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.444, + 0.424, + 0.459 + ], + "angle": 0, + "content": "4.2 CARFF Evaluation" + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.495, + 0.789, + 0.607 + ], + "angle": 0, + "content": "
Method3DComplex ScenariosState UncertaintyDynamics UncertaintyPredictionPlanningCode Released
CARFF✓✓✓✓✓✓✓
(20)✓✓
(21)✓✓✓✓
(1)✓✓✓✓
(29)✓✓✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.617, + 0.788, + 0.702 + ], + "angle": 0, + "content": "Table 1: Qualitative comparison of CARFF to related works. CARFF accomplishes all highlighted objectives as opposed to NeRF-VAE (20), NeRF for Visuomotor Control (21), Vision-only NeRF Navigation (1), and AVP (29). We compare whether methods reason in a 3D environment and perform novel view synthesis; work on complex scenarios; predict probabilistically under state and dynamics uncertainty; forecast into the future; and use model predictions for decision-making." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "A desirable behavior from our model is that it should predict a complete set of possible scenes consistent with the given ego-centric image, which could be partially observable. This is crucial for autonomous driving in unpredictable environments as it ensures strategic decision-making based on potential hazards. To achieve this we require a rich PC-VAE latent space, high-quality novel view synthesis, and auto-regressive probabilistic predictions of latents at future timestamps. We evaluate CARFF on a simple synthetic blender-based dataset" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "and each CARLA-based dataset. Additionally, we extend our model application to a hand-manipulation dataset in Appendix ??" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.785, + 0.268 + ], + "angle": 0, + "content": "Comparisons with related work: We attempt to compare CARFF to existing approaches. NeRF-VAE has the most comparable objective, but during our experiments, it collapses to black using CARLA datasets. We make further qualitative comparisons to other most similar methods in Tab. 1, but none aligns with ours enough to make any possible quantitative comparisons." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.787, + 0.418 + ], + "angle": 0, + "content": "Evaluation on blender dataset: In Fig. 5, for both Scene 1a and 1b, our model correctly forecasts the lateral movement of the cylinder to be in either position approximately \\(50\\%\\) of the time, considering a left viewing angle. In Scene 2, with the absence of the red cylinder in the input camera angle, the model predicts the potential existence of the red cylinder approximately \\(50\\%\\) of the time, and predicts lateral movements with roughly equal probability. This validates PC-VAE's ability to predict and infer occlusions in the latent space, aligning with human intuitions. These intuitions, shown in the Blender dataset's simple scenes, can transfer to driving scenarios in our CARLA datasets." + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.456, + 0.718, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.584 + ], + "angle": 0, + "content": "Fig. 7: PC-VAE reconstructions. The encoder input, \\( I_{c}^{t} \\), among the other ground truth images \\( I_{c} \\) viewed from camera pose \\( c \\) at different timestamps, is reconstructed across a new set of poses \\( c'' \\) respecting timestamp \\( t \\), generating \\( I_{c''}^{t} \\). A complete grid is in Appendix ??" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.841 + ], + "angle": 0, + "content": "PC-VAE performance and ablations: We evaluate the performance of PC-VAE on CARLA datasets with multiple encoder architectures. We show that PC-VAE effectively reconstructs complex environments involving variable scenes, actor configurations, and environmental noise given potentially partially observable inputs (see Fig. 7). We calculated an average Peak Signal-to-Noise Ratio (PSNR) over the training data, as well as novel view encoder inputs. To evaluate the quality of the latent space generated by the encoder, we utilize t-SNE (24) plots to visualize the distribution of latent samples for each image in a given dataset (see Appendix ??). We introduce a Support Vector Machine (SVM) (14) based metric to measure the visualized clustering quantitatively, where a higher value indicates better clustering based on timestamps. Most latent scene samples are separable by timestamps, which indicates that the latents are view-invariant. Samples that are misclassified or lie on the boundary usually represent partially or fully occluded regions. This is desirable for forecasting, as it enables" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.151, + 0.486, + 0.367 + ], + "angle": 0, + "content": "
Ground Truth Prediction PairAvg. PSNR (Scene 1)Avg. PSNR (Scene 2)
Single-Scene Intersection
Matching Pairs29.06N.A
Un-matching P.24.01N.A
Multi-Scene Intersection
Matching Pairs28.0028.26
Un-matching P.23.2724.56
Multi-Scene Two Lane Merge
Matching Pairs28.1428.17
Un-matching P.22.7423.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.376, + 0.496, + 0.501 + ], + "angle": 0, + "content": "Table 2: Averaged PSNR for fully observable 3D predictions. CARFF correctly predicts scene evolution across all timestamps for each dataset. The average PSNR is high for predictions \\(\\hat{I_{t_i}}\\) and matching ground truths, \\(I_{t_i}\\). PSNR values for incorrect correspondences, \\(\\hat{I_{t_i}}, I_{t_j}\\), is a result of matching surroundings. See complete table in Appendix ??" + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.144, + 0.783, + 0.366 + ], + "angle": 0, + "content": "
Multi-Scene Intersection
Controller TypeActorNo Actor
Underconfident30/300/30
Overconfident0/3030/30
CARFF (n=2)17/3030/30
CARFF (n=10)30/3030/30
CARFF (n=35)30/3019/30
Multi-Scene Two Lane Merge
Controller TypeFastSlow
Underconfident30/300/30
Overconfident0/3030/30
CARFF (n=2)21/3030/30
CARFF (n=10)30/3030/30
CARFF (n=35)30/3022/30
" + }, + { + "type": "table_caption", + "bbox": [ + 0.505, + 0.376, + 0.788, + 0.502 + ], + "angle": 0, + "content": "Table 3: Planning in 3D with controllers with varying sampling numbers \\( \\mathbf{n} \\). CARFF-based controllers outperform baselines in success rate over 30 trials. For \\( n = 10 \\), the CARFF-based controller consistently chooses the optimal action in potential collision scenarios. To maintain consistency, we use one single image input across 30 trials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.788, + 0.609 + ], + "angle": 0, + "content": "us to model probabilistic behavior over these samples. In this process, balancing KL divergence weight scheduling maintains the quality of the PC-VAE's latent space and reconstructions (see Appendix ??). Additionally, we substantiate the benefits of our PC-VAE encoder architecture through our ablations (see Appendix ??)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.788, + 0.749 + ], + "angle": 0, + "content": "3D novel view synthesis: Given an unseen ego-centric view with potentially partial observations, our method maintains all possible current state beliefs in 3D, and faithfully reconstructs novel views from arbitrary camera angles for each belief. Fig. 2 illustrates one of the possible 3D beliefs that CARFF holds. This demonstrates our method's ability to generate 3D beliefs that could be used for novel view synthesis in a view-consistent manner. Our model's ability to achieve accurate and complete 3D environmental understanding is important for applications like prediction-based planning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Inference under full and partial observations: Under full observation, we use MDN to predict the subsequent car positions in all three datasets. PSNR values are calculated based on bird-eye view NeRF renderings and ground truth bird-eye view images of the scene across different timestamps. In Tab. 2 we report the PSNR values for rendered images over the predicted posterior with the ground" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.283 + ], + "angle": 0, + "content": "truth images at each timestamp. We also evaluate the efficacy of our prediction model using the accuracy curve given in Fig. 8. This represents CARFF's ability to generate stable beliefs, without producing incorrect predictions, based on actor(s) localization results. For each number of samples between \\( n = 0 \\) to \\( n = 50 \\), we choose a random subset of 3 fully observable ego images and take an average of the accuracies. In scenarios with partial observable ego-centric images where several plausible scenarios exist, we utilize recall instead of accuracy using a similar setup. This lets us evaluate the encoder's ability to avoid false negative predictions of potential danger." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.785, + 0.375 + ], + "angle": 0, + "content": "Fig. 8 shows that our model achieves high accuracy and recall in both datasets, demonstrating the ability to model state uncertainty (Approaching Intersection) and dynamic uncertainty (Two Lane Merge). The results indicate CARFF's resilience against randomness in resampling, and completeness in probabilistic modeling of the belief space. Given these observations, we now build a reliable controller to plan and navigate through complex scenarios." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.394, + 0.338, + 0.41 + ], + "angle": 0, + "content": "4.3 Planning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.784, + 0.461 + ], + "angle": 0, + "content": "In all our experiments, the ego vehicle must make decisions to advance under certain observability. The scenarios are designed such that the ego views contain partial occlusion and the state of the actor(s) is uncertain in some scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.785, + 0.568 + ], + "angle": 0, + "content": "In order to facilitate decision-making using CARFF, we design a controller that takes ego-centric input images and outputs an action. Decisions are made incorporating sample consistency from the mixture density network. For instance, the controller infers occlusion and promotes the ego car to pause when scenes alternate between actor presence and absence in the samples. We use the two multi-scene datasets to assess the performance of the CARFF-based controller as they contain actors with potentially unknown behaviors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.786, + 0.734 + ], + "angle": 0, + "content": "To design an effective controller, we need to find a balance between accuracy and recall (see Fig. 8). A lowered accuracy from excessive sampling means unwanted randomness in the predicted state. However, taking insufficient samples would generate low recall i.e., not recovering all plausible states. This would lead to incorrect predictions as we would be unable to account for the plausible uncertainty present in the environment. To achieve optimal balance, we designed an open-loop planning controller using a sampling strategy that generates \\( n = 2, 10, 35 \\) samples. The hyperparameter \\( n \\) is tuned per scene for peak performance but is expected to remain relatively stable across scenes. We demonstrate that \\( n = 10 \\) performs well consistently in varying CARLA scenarios (Fig. 8) and do not anticipate this being very different for other experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "For sampling values that lie on the borders of the accuracy and recall margin, for example, \\( n = 2 \\) and 35, we see that the CARFF-based controller obtains lower success rates, whereas \\( n = 10 \\) produces the best result. For actor exists and fast-actor scenes in Tab. 3, we consider occluded ego-centric inputs to test the controller's ability to avoid collisions. For no-actor and slow-actor scenes, we consider state observability and test the controllers' ability to recognize the optimal action to advance. Across the two datasets, the overconfident controller" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "image", + "bbox": [ + 0.35, + 0.142, + 0.501, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.244, + 0.48, + 0.258 + ], + "angle": 0, + "content": "(a) Intersection" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.142, + 0.653, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.244, + 0.634, + 0.259 + ], + "angle": 0, + "content": "(b) Lane Merge" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.27, + 0.788, + 0.368 + ], + "angle": 0, + "content": "Fig. 8: Multi-Scene dataset accuracy and recall curves from predicted beliefs. We test our framework across \\( n = 1 \\) and \\( n = 50 \\) samples from MDN's predicted latent distributions from ego-centric image input. Across the number of samples \\( n \\), we achieve an ideal margin of belief state coverage generated under partial observation (recall), and the proportion of correct beliefs sampled under full observation (accuracy). As we significantly increase the number of samples, the accuracy starts to decrease due to randomness in latent distribution resampling." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.503 + ], + "angle": 0, + "content": "will inevitably experience collisions in case of a truck approaching, since it does not cautiously account for occlusions. On the other hand, an overly cautious approach results in stasis, inhibiting the controller's ability to advance in the scene. This nuanced decision-making using CARFF-based controller is especially crucial in driving scenarios, as it enhances safety and efficiency by adapting to complex and unpredictable road environments, thereby fostering a more reliable and human-like response in autonomous vehicles." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.524, + 0.355, + 0.54 + ], + "angle": 0, + "content": "5 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.788, + 0.691 + ], + "angle": 0, + "content": "Limitations: Like other NeRF-based methods, CARFF currently relies on posed images of specific scenes such as road intersections, limiting its direct applicability to unseen environments. However, we anticipate enhanced generalizability with the increasing deployment of cameras around populated areas, such as traffic cameras at intersections. Additionally, handling very complex dynamics with an extremely large number of actors still poses a challenge for our method, requiring per-scene optimization to balance comprehensive dynamics modeling against accuracy. Potentially stronger models in the near future may offer a promising avenue for further enhancements in this regard." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Conclusion: We present CARFF, a novel method for probabilistic 3D scene forecasting from partial observations. By employing a Pose-Conditional VAE, a NeRF conditioned on the learned posterior, and a mixture density network that forecasts future scenes, we effectively model, predict, and plan in complex environments with state and dynamics uncertainty in a POMDP. We further demonstrate the capabilities of our method in simulated autonomous driving scenarios. Overall, CARFF offers an intuitive framework to perceiving, forecasting, and acting under uncertainty that could prove invaluable for vision-based algorithms in unstructured environments." + } + ], + [ + { + "type": "title", + "bbox": [ + 0.427, + 0.142, + 0.577, + 0.164 + ], + "angle": 0, + "content": "Bibliography" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.211, + 0.787, + 0.258 + ], + "angle": 0, + "content": "[1] Adamkiewicz, M., Chen, T., Caccavale, A., Gardner, R., Culbertson, P., Bohg, J., Schwager, M.: Vision-only robot navigation in a neural radiance world. IEEE Robotics and Automation Letters 7(2), 4606-4613 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.259, + 0.787, + 0.304 + ], + "angle": 0, + "content": "[2] Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Int. Conf. Comput. Vis. pp. 5855–5864 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.305, + 0.787, + 0.335 + ], + "angle": 0, + "content": "[3] Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 130-141 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.336, + 0.787, + 0.38 + ], + "angle": 0, + "content": "[4] Cao, J., Wang, X., Darrell, T., Yu, F.: Instance-aware predictive navigation in multi-agent environments. In: IEEE Int. Conf. on Robotics and Automation. pp. 5096-5102. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.381, + 0.787, + 0.411 + ], + "angle": 0, + "content": "[5] Charatan, D., Li, S., Tagliasacchi, A., Sitzmann, V.: pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.787, + 0.457 + ], + "angle": 0, + "content": "[6] Codevilla, F., Santana, E., López, A.M., Gaidon, A.: Exploring the limitations of behavior cloning for autonomous driving. In: Int. Conf. Comput. Vis. pp. 9329-9338 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.458, + 0.787, + 0.503 + ], + "angle": 0, + "content": "[7] Deng, K., Liu, A., Zhu, J.Y., Ramanan, D.: Depth-supervised nerf: Fewer views and faster training for free. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12882-12891 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.504, + 0.787, + 0.564 + ], + "angle": 0, + "content": "[8] Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Int. Conf. Learn. Represent. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.787, + 0.595 + ], + "angle": 0, + "content": "[9] Dosovitskiy, A., Ros, G., Codevilla, F., Lopez, A., Koltun, V.: CARLA: An open urban driving simulator. In: Conf. on Robol Learning. pp. 1-16 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.596, + 0.787, + 0.641 + ], + "angle": 0, + "content": "[10] Driess, D., Huang, Z., Li, Y., Tedrake, R., Toussaint, M.: Learning multi-object dynamics with compositional neural radiance fields. arXiv preprint arXiv:2202.11855 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.642, + 0.787, + 0.672 + ], + "angle": 0, + "content": "[11] Fei, B., Xu, J., Zhang, R., Zhou, Q., Yang, W., He, Y.: 3d gaussian as a new vision era: A survey (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.673, + 0.787, + 0.718 + ], + "angle": 0, + "content": "[12] Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5501–5510 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.719, + 0.787, + 0.748 + ], + "angle": 0, + "content": "[13] Hausknecht, M., Stone, P.: Deep recurrent q-learning for partially observable mdps. In: AAAI (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.749, + 0.787, + 0.793 + ], + "angle": 0, + "content": "[14] Hearst, M.A., Dumais, S.T., Osuna, E., Platt, J., Scholkopf, B.: Support vector machines. IEEE Intelligent Systems and their applications 13(4), 18-28 (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "[15] Ichnowski, J., Avigal, Y., Kerr, J., Goldberg, K.: Dex-nerf: Using a neural radiance field to grasp transparent objects. arXiv preprint arXiv:2110.14217 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.211, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.146, + 0.785, + 0.193 + ], + "angle": 0, + "content": "[16] Ivanovic, B., Elhafsi, A., Rosman, G., Gaidon, A., Pavone, M.: Mats: An interpretable trajectory forecasting representation for planning and control. In: Conf. on Robol Learning (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.193, + 0.787, + 0.269 + ], + "angle": 0, + "content": "[17] Kaelbling, L.P., Littman, M.L., Cassandra, A.R.: Planning and acting in partially observable stochastic domains. Artificial Intelligence 101(1), 99-134 (1998). https://doi.org/https://doi.org/10.1016/S0004-3702(98)00023-X, https://www.sciencedirect.com/science/article/pii/S000437029800023X" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.27, + 0.787, + 0.331 + ], + "angle": 0, + "content": "[18] Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.787, + 0.377 + ], + "angle": 0, + "content": "[19] Kerr, J., Fu, L., Huang, H., Avigal, Y., Tancik, M., Ichnowski, J., Kanazawa, A., Goldberg, K.: Evo-nerf: Evolving nerf for sequential robot grasping of transparent objects. In: Conf. on Robol Learning (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.378, + 0.787, + 0.424 + ], + "angle": 0, + "content": "[20] Kosiorek, A.R., Strathmann, H., Zoran, D., Moreno, P., Schneider, R., Mokrá, S., Rezende, D.J.: NeRF-VAE: A geometry aware 3d scene generative model. In: ICML. pp. 5742–5752 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.469 + ], + "angle": 0, + "content": "[21] Li, Y., Li, S., Sitzmann, V., Agrawal, P., Torralba, A.: 3d neural scene representations for visuomotor control. In: Conf. on Robol Learning. pp. 112-123 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.47, + 0.787, + 0.532 + ], + "angle": 0, + "content": "[22] Liu, J.W., Cao, Y.P., Mao, W., Zhang, W., Zhang, D.J., Keppo, J., Shan, Y., Qie, X., Shou, M.Z.: Devrf: Fast deformable voxel radiance fields for dynamic scenes. In: Adv. Neural Inform. Process. Syst. vol. 35, pp. 36762-36775 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.533, + 0.787, + 0.563 + ], + "angle": 0, + "content": "[23] Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.787, + 0.609 + ], + "angle": 0, + "content": "[24] van der Maaten, L., Hinton, G.: Visualizing data using t-SNE. Journal of Machine Learning Research 9, 2579-2605 (2008), http://www.jmlr.org/papers/v9/vandermaaten08a.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.787, + 0.655 + ], + "angle": 0, + "content": "[25] Marza, P., Matignon, L., Simonin, O., Wolf, C.: Multi-object navigation with dynamically learned neural implicit representations. In: Int. Conf. Comput. Vis. pp. 11004-11015 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.787, + 0.702 + ], + "angle": 0, + "content": "[26] McAllister, R., Rasmussen, C.E.: Data-efficient reinforcement learning in continuous state-action gaussian-pomdpds. In: Adv. Neural Inform. Process. Syst. (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.787, + 0.748 + ], + "angle": 0, + "content": "[27] Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: Eur. Conf. Comput. Vis. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.749, + 0.787, + 0.794 + ], + "angle": 0, + "content": "[28] Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 1-15 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "[29] Packer, C., Rhinehart, N., McAllister, R.T., Wright, M.A., Wang, X., He, J., Levine, S., Gonzalez, J.E.: Is anyone there? learning a planner contingent on perceptual uncertainty. In: Conf. on Robol Learning (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.146, + 0.787, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.146, + 0.785, + 0.208 + ], + "angle": 0, + "content": "[30] Packer, C., Rhinehart, N., McAllister, R.T., Wright, M.A., Wang, X., He, J., Levine, S., Gonzalez, J.E.: Is anyone there? learning a planner contingent on perceptual uncertainty. In: Liu, K., Kulic, D., Ichnowski, J. (eds.) Conf. on Robol Learning. pp. 1607-1617 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.209, + 0.785, + 0.239 + ], + "angle": 0, + "content": "[31] Pan, X., You, Y., Wang, Z., Lu, C.: Virtual to real reinforcement learning for autonomous driving. arXiv preprint arXiv:1704.03952 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.239, + 0.785, + 0.271 + ], + "angle": 0, + "content": "[32] Papadimitriou, C.H., Tsitsiklis, J.N.: The complexity of markov decision processes. Mathematics of operations research 12(3), 441-450 (1987)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.317 + ], + "angle": 0, + "content": "[33] Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: Int. Conf. Comput. Vis. pp. 5865–5874 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.317, + 0.785, + 0.378 + ], + "angle": 0, + "content": "[34] Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.41 + ], + "angle": 0, + "content": "[35] Pineau, J., Gordon, G., Thrun, S., et al.: Point-based value iteration: An anytime algorithm for pomdpps. In: IJCAI. vol. 3, pp. 1025-1032 (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.41, + 0.785, + 0.456 + ], + "angle": 0, + "content": "[36] Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-NeRF: Neural Radiance Fields for Dynamic Scenes. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.457, + 0.785, + 0.502 + ], + "angle": 0, + "content": "[37] Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.564 + ], + "angle": 0, + "content": "[38] Rhinehart, N., He, J., Packer, C., Wright, M.A., McAllister, R., Gonzalez, J.E., Levine, S.: Contingencies from observations: Tractable contingency planning with learned behavior models. In: IEEE Int. Conf. on Robotics and Automation. pp. 13663-13669 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.61 + ], + "angle": 0, + "content": "[39] Roessle, B., Barron, J.T., Mildenhall, B., Srinivasan, P.P., Nießner, M.: Dense depth priors for neural radiance fields from sparse input views. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12892–12901 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.672 + ], + "angle": 0, + "content": "[40] Shao, R., Zheng, Z., Tu, H., Liu, B., Zhang, H., Liu, Y.: Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 16632-16642 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.785, + 0.718 + ], + "angle": 0, + "content": "[41] Sun, C., Sun, M., Chen, H.T.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5459-5469 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.779 + ], + "angle": 0, + "content": "[42] Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8248–8258 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.779, + 0.785, + 0.842 + ], + "angle": 0, + "content": "[43] Tancik, M., Srinivasan, P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. In: Adv. Neural Inform. Process. Syst. pp. 7537–7547 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.146, + 0.785, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Yang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.146, + 0.785, + 0.193 + ], + "angle": 0, + "content": "[44] Tas, O.S., Stiller, C.: Limited visibility and uncertainty aware motion planning for automated driving. In: IEEE Intelligent Vehicles Symposium (IV) (jun 2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.193, + 0.785, + 0.252 + ], + "angle": 0, + "content": "[45] Tewari, A., Thies, J., Mildenhall, B., Srinivasan, P., Tretschk, E., Yifan, W., Lassner, C., Sitzmann, V., Martin-Brualla, R., Lombardi, S., Simon, T., Theobalt, C., Nießner, M., Barron, J.T., Wetzstein, G., Zollhöfer, M., Golyanik, V.: Advances in Neural Rendering. Comput. Graph. Forum (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.253, + 0.785, + 0.298 + ], + "angle": 0, + "content": "[46] Toromanoff, M., Wirbel, E., Moutarde, F.: End-to-end model-free reinforcement learning for urban driving using implicit affordances. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7153-7162 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.299, + 0.785, + 0.343 + ], + "angle": 0, + "content": "[47] Tretschk, E., Golyanik, V., Zollhoefer, M., Bozic, A., Lassner, C., Theobalt, C.: Scenerflow: Time-consistent reconstruction of general dynamic scenes. In: International Conference on 3D Vision (3DV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.344, + 0.785, + 0.389 + ], + "angle": 0, + "content": "[48] Truong, P., Rakotosaona, M.J., Manhardt, F., Tombari, F.: Sparf: Neural radiance fields from sparse and noisy poses. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4190-4200 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.785, + 0.433 + ], + "angle": 0, + "content": "[49] Turki, H., Ramanan, D., Satyanarayanan, M.: Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12922–12931 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.493 + ], + "angle": 0, + "content": "[50] Wang, Q., Wang, Z., Genova, K., Srinivasan, P.P., Zhou, H., Barron, J.T., Martin-Brualla, R., Snavely, N., Funkhouser, T.: Ibrnet: Learning multi-view image-based rendering. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4690-4699 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.539 + ], + "angle": 0, + "content": "[51] Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5438-5448 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.539, + 0.785, + 0.599 + ], + "angle": 0, + "content": "[52] Yen-Chen, L., Florence, P., Barron, J.T., Rodriguez, A., Isola, P., Lin, T.Y.: inerf: Inverting neural radiance fields for pose estimation. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1323-1330. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.6, + 0.785, + 0.629 + ], + "angle": 0, + "content": "[53] Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.63, + 0.785, + 0.703 + ], + "angle": 0, + "content": "[54] Yu, M.Y., Vasudevan, R., Johnson-Roberson, M.: Occlusion-aware risk assessment for autonomous driving in urban environments. IEEE Robotics and Automation Letters 4(2), 2235-2241 (Apr 2019). https://doi.org/10.1109/lra.2019.2900453, http://dx.doi.org/10.1109/LRA.2019.2900453" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.785, + 0.779 + ], + "angle": 0, + "content": "[55] Zhang, C., Steinhauser, F., Hinz, G., Knoll, A.: Improved occlusion scenario coverage with a pomdp-based behavior planner for autonomous urban driving. In: 2021 IEEE International Intelligent Transportation Systems Conference (ITSC). pp. 593-600 (2021). https://doi.org/10.1109/ITSC48978.2021.9564424" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "[56] Zhang, C., Steinhauser, F., Hinz, G., Knoll, A.: Occlusion-aware planning for autonomous driving with vehicle-to-everything communication. IEEE Transactions on Intelligent Vehicles 9(1), 1229-1242 (2024). https://doi.org/10.1109/TIV.2023.3308098" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.146, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CARFF_ Conditional Auto-encoded Radiance Field for 3D Scene Forecasting/4d2110cb-f049-4580-b1d7-c4b728d00449_origin.pdf b/2024/CARFF_ Conditional Auto-encoded Radiance Field for 3D Scene Forecasting/4d2110cb-f049-4580-b1d7-c4b728d00449_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3eeb67a8dd3618e2ec6590077c6f8a0d868eb55b --- /dev/null +++ b/2024/CARFF_ Conditional Auto-encoded Radiance Field for 3D Scene Forecasting/4d2110cb-f049-4580-b1d7-c4b728d00449_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a87464cb0d982e8d1af416589a322fe37e4b4c3af668b956027231399335b2e3 +size 23574598 diff --git a/2024/CARFF_ Conditional Auto-encoded Radiance Field for 3D Scene Forecasting/full.md b/2024/CARFF_ Conditional Auto-encoded Radiance Field for 3D Scene Forecasting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a69495a030555e0e2a18f07ca9067b59ec433d75 --- /dev/null +++ b/2024/CARFF_ Conditional Auto-encoded Radiance Field for 3D Scene Forecasting/full.md @@ -0,0 +1,291 @@ +# CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting + +Jiezhi "Stephen" Yang $^{1}$ , Khushi Desai $^{2}$ , Charles Packer $^{3}$ , Harshil Bhatia $^{4}$ , Nicholas Rhinehart $^{3}$ , Rowan McAllister $^{5}$ , and Joseph E. Gonzalez $^{3}$ + +1 Harvard University MA 02138, USA + +2 Columbia University NY 10025, USA + +UC Berkeley CA 94720, USA + +4 Avataar.ai KA 560103, India + +5 Toyota Research Institute CA 94022, USA + +Abstract. We propose CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting, a method for predicting future 3D scenes given past observations. Our method maps 2D ego-centric images to a distribution over plausible 3D latent scene configurations and predicts the evolution of hypothesized scenes through time. Our latents condition a global Neural Radiance Field (NeRF) to represent a 3D scene model, enabling explainable predictions and straightforward downstream planning. This approach models the world as a POMDP and considers complex scenarios of uncertainty in environmental states and dynamics. Specifically, we employ a two-stage training of Pose-Conditional-VAE and NeRF to learn 3D representations, and auto-regressively predict latent scene representations utilizing a mixture density network. We demonstrate the utility of our method in scenarios using the CARLA driving simulator, where CARFF enables efficient trajectory and contingency planning in complex multi-agent autonomous driving scenarios involving occlusions. Video and code are available at: www.carff website. + +# 1 Introduction + +Humans often imagine what they cannot see given partial visual context. Consider a scenario where reasoning about the unobserved is critical to safe decision-making: for example, a driver navigating a blind intersection. An expert driver will plan according to what they believe may or may not exist in occluded regions of their vision. The driver's belief - defined as the understanding of the world modeled with consideration for inherent environment uncertainties - is informed by their partial observations (i.e., the presence of other vehicles on the road), as well as their prior knowledge (e.g., past experience navigating this intersection). + +When reasoning about the unobserved, humans form complex beliefs about the existence, position, shapes, colors, and textures of occluded scene portions + +(e.g., an oncoming car). Autonomous systems with high-dimensional sensor data, like video or LiDAR, traditionally reduce this data to low-dimensional state information (e.g., position and velocity of tracked objects) for prediction and planning. + +In addition to tracking fully observed objects, this object-centric framework handles partially observed settings by considering potentially dangerous unobserved objects. These systems often plan for worst-case scenarios, such as a "ghost car" at the edge of the visible field of view (44). + +![](images/b84fe5410986cc3ec666ba1809f5163c18eadc7a63813e26b65137ad51372830.jpg) +Fig. 1: CARFF 3D planning application for driving. An input image containing a partially observable view of an intersection is processed by CARFF's encoder to establish 3D environment state beliefs, i.e. the predicted possible state of the world: whether or not there could be another vehicle approaching the intersection. These beliefs are used to forecast the future in 3D for planning, generating one among two possible actions for the vehicle to merge into the other lane. + +Recent advances in neural rendering, particularly Neural Radiance Fields (NeRF), have significantly improved 3D scene representation learning. NeRF enables novel view synthesis, thus simplifying the process of viewing behind occlusions. NeRF decouples the dependency of scene representation from traditional object detection and tracking, allowing for the capture of vital visual information that might be missed by detectors, yet is crucial for safe decision-making. NeRF's implicit density representation of explicit geometry also facilitates its direct application in motion planning without the need for rendering. NeRF's ability to represent both visual and geometric information makes them a more general and intuitive 3D representation for autonomous systems. + +Despite NeRF's advantages, achieving probabilistic predictions in 3D based on reasoning from occluded views is challenging. For example, discriminative models that yield categorical predictions are unable to capture the underlying 3D structure, impeding their ability to model uncertainty. While prior work on 3D representation captures view-invariant structures, their application is primarily confined to simple scenarios (20). We present CARFF, which to our knowledge, is the first forecasting approach in scenarios with partial observations that uniquely facilitates stochastic predictions in a partially observable Markov decision process (POMDP) within a 3D representation, effectively integrating visual perception and geometry. Specifically, we make the following contributions: + +1. We propose a novel architecture $PC$ -VAE: Pose-Conditioned Variational Autoencoder. The encoder maps potentially partially observable ego-centric images to pose-invariant latent scene representations, which hold state beliefs of the POMDP with implicit probability distributions (see Sec. 3.1). + +2. We develop the two-stage training pipeline that uniquely enables complex scene modeling with a probabilistic objective. This involves separately training the PC-VAE and a latent conditioned neural radiance field that functions as a 3D decoder, enabling interpretable predictions (see Sec. 3.1). + +3. We design a mixture density model to predict the evolution of 3D scenes over time stochastically and regressively in the encoder belief space (see Sec. 3.2). This allows for an effective sampling based-controller to output actions in the POMDP. + +We demonstrate how CARFF can be used to enable contingency planning in complex driving scenarios that require reasoning into visual occlusions on CARLA simulated datasets inspired by autonomous driving planning tasks (29; 54; 55; 56). A potential application of CARFF is illustrated in Fig. 1. + +# 2 Related work + +# 2.1 NeRF and 3D representations + +Neural radiance fields. Neural Radiance Fields (NeRF) (2; 27; 43) for 3D representations generate high-resolution, photorealistic scenes. Instant Neural Graphics Primitive (Instant-NGP) (28) speeds up training and rendering time by introducing a multi-resolution hash encoding. Other works like Plenoxels (12) and DirectVoxGo (DVGO) (41) also provide similar speedups. Recent advancements in volumetric representations such as 3D Gaussian Splatting (18) enhance rendering efficiency while maintaining compatibility with traditional NeRF applications (11). We utilize Instant-NGP for its accessibility, although our approach is adaptable to alternative rendering methods. NeRFs have also been extended for several tasks such as modeling large-scale unbounded scenes (2; 42; 49), scene from sparse views (7; 39; 48) and multiple scenes (20; 50). For an in-depth survey on neural representation learning and its applications we refer the reader to (45). + +Generalizable novel view synthesis models, like pixelNeRF and pixelSplat (5; 53), learn a scene prior to render novel views from sparse existing ones. In contrast, CARFF is based on a VAE, encoding a probabilistic objective and decoding to future 3D scenes. Dynamic NeRF models scenes with moving or deforming objects, within which a widely used approach is to construct a canonical space and predict a deformation field (22; 33; 34; 36). The canonical space is usually a static scene, and the model learns an implicitly represented flow field (33; 36). A recent line of work also models dynamic scenes via different representations and decomposition (3; 40). These approaches tend to perform better for spatially bounded and predictable scenes with relatively small variations (3; 23; 33; 53). Moreover, these methods only solve for changes in the environment but are limited in incorporating stochasticity in the environment. + +![](images/911c9359ba538544b664309f060185792214c88863a23d0e2a3844cb4b919ec8.jpg) +Ego-centric Lane View +Fig. 2: Novel view planning application. CARFF allows reasoning behind occluded views from the ego car as simple as moving the camera to see the sampled belief predictions, allowing simple downstream planning using, for example, density probing or 2D segmentation models from arbitrary angles. + +![](images/7f904b0876f8b554d72dcdbc4a9168a495f5a5cb691242159c4c1ec33921f7dd.jpg) +Predicted 3D View + +Multi-scene NeRF: Our approach builds on multi-scene NeRF approaches (20; 47; 50; 51) that learn a global latent scene representation, which conditions the NeRF, allowing a single NeRF to effectively represent various scenes. A similar method, NeRF-VAE, was introduced by Kosiorek et al. (20) to create a geometrically consistent 3D generative model with generalization to out-of-distribution cameras. However, NeRF-VAE (20) is prone to mode collapse when faced with complex visual information (see Sec. 4.2). + +# 2.2 Scene Forecasting + +Planning in 2D space: Planning in large, continuous state-action spaces is challenging due to exponentially large search spaces (32), leading to various approximation methods for tractability (26; 35). Model-free (13; 31; 46) and model-based (4) reinforcement learning frameworks, along with other learning-based methods (6; 29), have emerged as viable approaches. Additionally, methods forecast for downstream control (16), learn behavior models for contingency planning (38), or predict the existence and intentions of unobserved agents (30). While these methods operate in 2D, we reason under partial observations and account for these factors in 3D. + +NeRF in robotics: Recent works have applied NeRFs in robotics for localization (52), navigation (1; 25), dynamics modeling (10; 22), and robotic grasping (15; 19). Adamkiewicz et al.(1) propose quadcopter motion planning in NeRF models by sampling the learned density function, useful for forecasting and planning. Driess et al.(10) employ a graph neural network to learn dynamics in a multi-object NeRF scene. Li et al. (21) focus on pushing tasks and address grasping and planning with NeRF and a separate latent dynamics model. Prior approaches work in simple, static scenes (1) or uses deterministic dynamics models (21). CARFF addresses complex, realistic environments with both state and dynamics uncertainty, considering potential object existence and unknown movements. + +# 3 Method + +Recent advancements in 3D scene representation allow for modeling environments in a contextually rich and interactive 3D space. This offers analytical benefits, such as spatial analysis with soft occupancy grids and object detection through novel view synthesis. Given these advantages, our primary objective is to develop a model for probabilistic 3D scene forecasting in dynamic environments. However, direct integration of 3D scene representation via NeRF and probabilistic models like VAE often involves non-convex and inter-dependent optimization, which causes unstable training. For instance, NeRF's optimization may rely on the VAE's latent space being structured to provide informative gradients. To + +![](images/a431e18c5007cafac98de48641d9202701cb8965d45e1dc5703eb3ff32b6ada2.jpg) +Fig. 3: Visualizing CARFF's two stage training process. Left: The convolutional VIT-based encoder encodes each image $I$ at timestamps $t, t'$ and camera poses $c, c'$ into Gaussian latent distributions. Assuming two timestamps and an overparameterized latent, one Gaussian distribution will have a smaller $\sigma^2$ , and different $\mu$ across timestamps. Upper Right: The pose-conditional decoder stochastically decodes the sampled latent $z$ using the camera pose $c''$ into images $I_{c''}^t$ and $I_{c''}^{t'}$ . The decoded reconstruction and ground truth images are used for the loss $\mathcal{L}_{\mathrm{MSE}}$ , PC-VAE. Lower Right: A NeRF is trained by conditioning on the latent variables sampled from the optimized Gaussian parameters. These parameters characterize the distinct timestamp distributions derived from the PC-VAE. An MSE loss is calculated for NeRF as $\mathcal{L}_{\mathrm{MSE}}$ , NeRF. + +navigate these complexities, our method bifurcates the training process into two stages (see Fig. 3). First, we train the PC-VAE to learn view-invariant scene representations. Next, we replace the decoder with a NeRF to learn a 3D scene from the latent representations. The latent scene representations capture the environmental states and dynamics over possible underlying scenes, while NeRF synthesizes novel views within the belief space, giving us the ability to see the unobserved (see Fig. 2 and Sec. 3.1). During prediction, uncertainties can be modeled by sampling latents auto-regressively from a predicted Gaussian mixture, allowing for effective decision-making. To this extent, we approach scene forecasting as a POMDP over latent distributions, which enables us to capture multi-modal beliefs for planning amidst perceptual uncertainty (see Sec. 3.2). + +# 3.1 Pose-Conditional VAE (PC-VAE) and NeRF + +Architecture: We assume that the model follows a Markovian process, and thus each belief state only depends on the previous. Given a scene $S_{t}$ at timestamp $t$ , we have an ego-centric observation image $I_{c}^{t}$ captured from camera pose $c$ . The objective is to formulate a 3D representation of the image that holds implicit probability distributions of the possible states, where we can perform a forecasting step that evolves the scene forward. Here, the POMDP can be seen as an MDP in belief space (17). To achieve this, we utilize a radiance field conditioned on latent variable $z$ sampled from the posterior distribution $q_{\phi}(z|I_{c}^{t})$ . Now, to learn the posterior, we utilize PC-VAE. We construct an encoder using convolutional layers and a pre-trained ViT on ImageNet (8). The encoder learns a mapping from the image space to a Gaussian distributed latent space $q_{\phi}(z|I_{c}^{t}) = \mathcal{N}(\mu, \sigma^{2})$ parametrized by mean $\mu$ and variance $\sigma^2$ . The decoder, $p(I|z,c)$ , conditioned on camera pose $c$ , maps the latent $z \sim \mathcal{N}(\mu, \sigma^2)$ into the image space $I$ . This helps the encoder to generate latents that are invariant to the camera pose $c$ . + +To enable 3D scene modeling, we employ Instant-NGP (28), which incorporates a hash grid and an occupancy grid to enhance computation efficiency. Additionally, a smaller multilayer perceptron (MLP), $F_{\theta}(z)$ can be utilized to model the density and appearance, given by: + +$$ +F _ {\theta} (z): (\mathbf {x}, \mathbf {d}, z) \rightarrow ((r, g, b), \sigma) \tag {1} +$$ + +Here, $\mathbf{x} \in \mathbb{R}^3$ and $\mathbf{d} \in (\theta, \phi)$ represent the location vector and the viewing direction respectively. The MLP is conditioned on the sampled scene latents $z \sim q_{\phi}(z|I_c^t)$ (see Appendix ??). + +Training methodology: The architecture alone does not enable us to model complex scenarios, as seen through a similar example in NeRF-VAE (20). A crucial contribution of our work is our two-stage training framework which stabilizes the training. First, we optimize the convolutional ViT based encoder and pose-conditional convolutional decoder in the pixel space for reconstruction. This enables our method to deal with more complex and realistic scenes as the encoding is learned in a semantically rich 2D space. By conditioning the decoder on camera poses, we achieve disentanglement between camera view angles and scene context, making the representation view-invariant and the encoder 3D-aware. Once rich latent representations are learned, we replace the decoder with a latent-conditioned NeRF over the latent space of the frozen encoder. The NeRF reconstructs encoder beliefs in 3D for novel view synthesis. + +Loss: PC-VAE is trained using standard VAE loss, with mean square error (MSE) and a Kullback-Leibler (KL) divergence given by evidence lower bound: + +$$ +\begin{array}{l} \mathcal {L} _ {P C - V A E} = \mathcal {L} _ {M S E, P C - V A E} + \mathcal {L} _ {K L D, P C - V A E} = \\ \left\| p \left(I | z, c ^ {\prime \prime}\right) - I _ {c ^ {\prime \prime}} ^ {t} \right\| ^ {2} + \mathbb {E} _ {q (z \mid I _ {c} ^ {t})} [ \log p (I | z) ] - w _ {K L} D _ {K L} \left(q _ {\phi} \left(z \mid I _ {c} ^ {t}\right) \mid p (I | z)\right) \tag {2} \\ \end{array} +$$ + +where $w_{KL}$ denotes the KL divergence loss weight and $z \sim q_{\phi}(z|I_c^t)$ . To make our representation 3D-aware, our posterior is encoded using camera $c$ while the decoder is conditioned on a randomly sampled pose $c''$ . + +KL divergence regularizes the latent space to balance conditioned reconstruction and stochasticity under occlusion. An elevated KL divergence loss weight $w_{KL}$ pushes the latents closer to a standard normal distribution, $\mathcal{N}(0,1)$ , thereby ensuring probabilistic sampling in scenarios under partial observation. However, excessive regularization causes the latents to be less separable, leading to mode collapse. To mitigate this, we adopt delayed linear KL divergence loss weight scheduling to strike a balanced $w_{KL}$ . + +Next, we learn a NeRF decoder on the posterior of the VAE to model scenes. At any timestamp $t$ we use a standard photometric loss for training the NeRF, given by the following equation: + +$$ +\mathcal {L} _ {M S E, N e R F} = \left\| I _ {c} ^ {t} - r e n d e r \left(F _ {\theta} \left(\cdot \mid q _ {\phi} \left(z \mid I _ {c} ^ {t}\right)\right)\right) \right\| ^ {2} \tag {3} +$$ + +We use a standard rendering algorithm as proposed by Müller et al. (28). Next, we build a forecasting module over the learned latent space of our pose-conditional encoder. + +# 3.2 Scene Forecasting + +Formulation: The current formulation allows us to model scenes with different configurations across timestamps. In order to forecast future configurations of a scene given an ego-centric view, we need to predict future latent distributions. We formulate the forecasting as a POMDP over the posterior distribution $q_{\phi}(z|I_c^t)$ in the PC-VAE's latent space. + +During inference, we observe stochastic behaviors under occlusion, which motivates us to learn a mixture of several Gaussian distributions that potentially denote different scene possibilities. Therefore, we model the POMDP using a Mixture Density Network (MDN), with multi-headed MLPs, that predicts a mixture of $K$ Gaussians. At any timestamp $t$ the distribution is given as: + +$$ +q _ {\phi} ^ {\prime} \left(z _ {t} \mid I _ {c} ^ {t - 1}\right) = M D N \left(q _ {\phi} \left(z _ {t - 1} \mid I _ {c} ^ {t - 1}\right)\right) \tag {4} +$$ + +The model is conditioned on the posterior distribution $q_{\phi}(z_{t - 1})$ to learn a predicted posterior distribution $q_{\phi}'(z_t|I_c^{t - 1})$ at each timestamp. The predicted posterior distribution is given by the mixture of Gaussian: + +$$ +q _ {\phi} ^ {\prime} \left(z _ {t}\right) = \sum_ {i = 1} ^ {K} \pi_ {i} \mathcal {N} \left(\mu_ {i}, \sigma_ {i} ^ {2}\right) \tag {5} +$$ + +here, $\pi_i$ , $\mu_i$ , and $\sigma_i^2$ denote the mixture weight, mean, and variance of the $i^{th}$ Gaussian distribution within the posterior distribution. Here, $K$ is the total number of Gaussians. For brevity we remove their conditioning on the posterior $q_{\phi}(z_{t-1})$ and sampled latent $z_{t-1}$ . We sample $z_t$ from the mixture of Gaussians $q_{\phi}^{\prime}(z_t)$ , where $z_t$ likely falls within one of the Gaussian modes. The configuration corresponding to the mode is reflected in the 3D scene rendered by NeRF. + +![](images/7379d9d391663f7e3ed896bd624d5d8ee05057575b6903fed5329e19f50b314a.jpg) +Fig.4: Multi-scene CARLA datasets. Varying car configurations and scenes for the Multi-Scene Two Lane Merge dataset (left) and the Multi-Scene Approaching Intersection dataset (right). + +![](images/77088b311dcf40ce6bf05ce57903c749dd732695fd3c03cc0f7846dc544cbf3a.jpg) +Fig.5: Blender dataset. Blender dataset with a blue cube and a potential red cylinder exhibiting probabilistic temporal movement. The possible occlusions from different camera angles demonstrate how movement needs to be modeled probabilistically. + +![](images/afbb542921bb5c6efec2691d08739e8079af0267e0e4a8378ee2eb6ae6b38538.jpg) + +![](images/1f7ccc21d1f36ed77b383958701e42df86a5a7fa5bf0c0f06c76d122c56bc8c2.jpg) + +![](images/07cdc9ebb6f1228e46f4cba95d292c4f9ec9a600cbb7ad2d55a31179ce468385.jpg) + +Loss: To optimize the MDN, we minimize a negative log-likelihood function, given by: + +$$ +\mathcal {L} _ {M D N} = - \sum_ {j = 1} ^ {N} \log \left(\sum_ {i = 1} ^ {K} \pi_ {i} \mathcal {N} \left(y _ {j}; \mu_ {i}, \sigma_ {i} ^ {2}\right)\right) \tag {6} +$$ + +where $y_{i} \sim q_{\phi}(z_{t})$ is sampled from the distribution of latent $z_{t}$ , learned by the encoder, and $N$ denotes the total number of samples. + +Inference: We consider an unseen ego-centric image and retrieve its posterior $q_{\phi}(z_t)$ through the encoder. Next, we predict the possible future posterior distribution $q_{\phi}'(z_{t + 1})$ . From the predicted posterior, we sample a scene latent and perform localization. We achieve this via (a) density probing the NeRF or (b) segmenting the rendered novel views using off-the-shelf methods such as YOLO (37) (see Fig. 2). These allow us to retrieve a corresponding Gaussian distribution $q_{\phi}(z_{t + 1})$ in encoder latent space. This is auto-regressively fed back into the MDN to predict the next timestamp. See Fig. 6 for an overview of the pipeline. + +# 4 Results + +Decision-making under perceptual uncertainty is a pervasive challenge faced in robotics and autonomous driving, as the real environment is mostly likely partially observable, making it a POMDP. In a partially observable driving scenario, accurate inference regarding the presence of potentially obscured agents is pivotal. We evaluate the effectiveness of CARFF on common driving situations with partial observability and added complexity. We implemented several scenarios in the CARLA driving simulator (9) (see Fig. 4). A single NVIDIA RTX 3090 GPU is used to train PC-VAE, NeRF, and the MDN. All models, trained sequentially, tend to converge within a combined time frame of 24 hours. A detailed experimental setup can be found in Appendix ??. We show that, given + +partially observable 2D inputs, CARFF performs well in predicting latent distributions that represent complete 3D scenes. Using these predictions we design a CARFF-based controller for performing downstream planning tasks. + +# 4.1 Data Generation + +We conduct experiments on (a) synthetic blender dataset for principle experiments to test the probabilistic modeling capacities in isolation of the vision encoder (it is visually as simple as possible, but requires the full predictive model proposed in CARFF) and (b) CARLA-based driving datasets for more complex driving scenarios (9). To deliver convincing results, we model these driving scenarios off of related works (29; 54; 55; 56) that concern planning for driving under difficult situations. We generate the datasets in 3D by programming an ego object and varying actor objects in different configurations. + +![](images/19902d354443b11142b0941a8bce70e38f2d59be8ee146263f795cf147edc0a9.jpg) +Fig. 6: Auto-regressive inference in scene prediction. The input image at timesamp $t$ , $I_{c}^{t}$ , is encoded using the pre-trained encoder from PC-VAE. The corresponding latent distribution is fed into the Mixture Density Network, which predicts a mixture of Gaussians. Each of the $K$ Gaussians is a latent distribution that may correspond to different beliefs at the next timestamp. The mixture of Gaussians is sampled repeatedly for the predicted latent beliefs, visualized as $I_{c',scni}^{t+1}$ , representing potentially the $i$ th possible outcome. This is used to condition the NeRF to generate 3D views of the scene. To accomplish autoregressive predictions, we probe the NeRF for the location of the car and feed this information back to the pre-trained encoder to predict the scene at the next timestamp. + +Blender synthetic dataset: This comprises of a stationary blue cube (ego) accompanied by a red cylinder (actor) that may or may not be present (see Fig. 5). If the actor is present, it exhibits lateral movement as depicted in Fig. 5. This simplistic setting provides an interpretable framework to evaluate our model. + +CARLA dataset: Each dataset is simulated for $N$ timestamps and uses $C = 100$ predefined camera poses to capture images of the environment under full observation, partial observation, and no visibility. These datasets are modeled after + +common driving scenarios involving state uncertainty that have been proposed in related works such as Active Visual Planning (29). + +a) Single-Scene Approaching Intersection: The ego vehicle is positioned at a T-intersection. An actor vehicle traverses the crossing along an evenly spaced, predefined trajectory. We simulate this for $N = 10$ timestamps. We mainly use this dataset to predict the evolution of timestamps under full observation. +b) Multi-Scene Approaching Intersection: We extend the previous scenario to a more complicated setting with state uncertainty, by making the existence of the actor vehicle probabilistic. A similar intersection crossing is simulated for $N = 3$ timestamps for both possibilities. The ego vehicle's view of the actor may be occluded as it approaches the T-intersection over the $N$ timestamps. The ego vehicle either moves forward or halts at the junction (see Fig. 4). +c) Multi-Scene Multi-actor Two Lane Merge: To add more environment dynamics uncertainty, we consider a multi-actor setting at an intersection of two merging lanes. We simulate the scenario at an intersection with partial occlusions, with the second approaching actor having variable speed. Here the ego vehicle can either merge into the left lane before the second actor or after all the actors pass, (see Fig. 4). Each branch is simulated for $N = 3$ timestamps. + +# 4.2 CARFF Evaluation + +
Method3DComplex ScenariosState UncertaintyDynamics UncertaintyPredictionPlanningCode Released
CARFF✓✓✓✓✓✓✓
(20)✓✓
(21)✓✓✓✓
(1)✓✓✓✓
(29)✓✓✓✓✓
+ +Table 1: Qualitative comparison of CARFF to related works. CARFF accomplishes all highlighted objectives as opposed to NeRF-VAE (20), NeRF for Visuomotor Control (21), Vision-only NeRF Navigation (1), and AVP (29). We compare whether methods reason in a 3D environment and perform novel view synthesis; work on complex scenarios; predict probabilistically under state and dynamics uncertainty; forecast into the future; and use model predictions for decision-making. + +A desirable behavior from our model is that it should predict a complete set of possible scenes consistent with the given ego-centric image, which could be partially observable. This is crucial for autonomous driving in unpredictable environments as it ensures strategic decision-making based on potential hazards. To achieve this we require a rich PC-VAE latent space, high-quality novel view synthesis, and auto-regressive probabilistic predictions of latents at future timestamps. We evaluate CARFF on a simple synthetic blender-based dataset + +and each CARLA-based dataset. Additionally, we extend our model application to a hand-manipulation dataset in Appendix ?? + +Comparisons with related work: We attempt to compare CARFF to existing approaches. NeRF-VAE has the most comparable objective, but during our experiments, it collapses to black using CARLA datasets. We make further qualitative comparisons to other most similar methods in Tab. 1, but none aligns with ours enough to make any possible quantitative comparisons. + +Evaluation on blender dataset: In Fig. 5, for both Scene 1a and 1b, our model correctly forecasts the lateral movement of the cylinder to be in either position approximately $50\%$ of the time, considering a left viewing angle. In Scene 2, with the absence of the red cylinder in the input camera angle, the model predicts the potential existence of the red cylinder approximately $50\%$ of the time, and predicts lateral movements with roughly equal probability. This validates PC-VAE's ability to predict and infer occlusions in the latent space, aligning with human intuitions. These intuitions, shown in the Blender dataset's simple scenes, can transfer to driving scenarios in our CARLA datasets. + +![](images/688dd54c0d8771ce3865e20ca2164cf80a4e372f3027a55385a62270d4e89b8f.jpg) +Fig. 7: PC-VAE reconstructions. The encoder input, $I_{c}^{t}$ , among the other ground truth images $I_{c}$ viewed from camera pose $c$ at different timestamps, is reconstructed across a new set of poses $c''$ respecting timestamp $t$ , generating $I_{c''}^{t}$ . A complete grid is in Appendix ?? + +PC-VAE performance and ablations: We evaluate the performance of PC-VAE on CARLA datasets with multiple encoder architectures. We show that PC-VAE effectively reconstructs complex environments involving variable scenes, actor configurations, and environmental noise given potentially partially observable inputs (see Fig. 7). We calculated an average Peak Signal-to-Noise Ratio (PSNR) over the training data, as well as novel view encoder inputs. To evaluate the quality of the latent space generated by the encoder, we utilize t-SNE (24) plots to visualize the distribution of latent samples for each image in a given dataset (see Appendix ??). We introduce a Support Vector Machine (SVM) (14) based metric to measure the visualized clustering quantitatively, where a higher value indicates better clustering based on timestamps. Most latent scene samples are separable by timestamps, which indicates that the latents are view-invariant. Samples that are misclassified or lie on the boundary usually represent partially or fully occluded regions. This is desirable for forecasting, as it enables + +
Ground Truth Prediction PairAvg. PSNR (Scene 1)Avg. PSNR (Scene 2)
Single-Scene Intersection
Matching Pairs29.06N.A
Un-matching P.24.01N.A
Multi-Scene Intersection
Matching Pairs28.0028.26
Un-matching P.23.2724.56
Multi-Scene Two Lane Merge
Matching Pairs28.1428.17
Un-matching P.22.7423.32
+ +Table 2: Averaged PSNR for fully observable 3D predictions. CARFF correctly predicts scene evolution across all timestamps for each dataset. The average PSNR is high for predictions $\hat{I_{t_i}}$ and matching ground truths, $I_{t_i}$ . PSNR values for incorrect correspondences, $\hat{I_{t_i}}, I_{t_j}$ , is a result of matching surroundings. See complete table in Appendix ?? + +
Multi-Scene Intersection
Controller TypeActorNo Actor
Underconfident30/300/30
Overconfident0/3030/30
CARFF (n=2)17/3030/30
CARFF (n=10)30/3030/30
CARFF (n=35)30/3019/30
Multi-Scene Two Lane Merge
Controller TypeFastSlow
Underconfident30/300/30
Overconfident0/3030/30
CARFF (n=2)21/3030/30
CARFF (n=10)30/3030/30
CARFF (n=35)30/3022/30
+ +Table 3: Planning in 3D with controllers with varying sampling numbers $\mathbf{n}$ . CARFF-based controllers outperform baselines in success rate over 30 trials. For $n = 10$ , the CARFF-based controller consistently chooses the optimal action in potential collision scenarios. To maintain consistency, we use one single image input across 30 trials. + +us to model probabilistic behavior over these samples. In this process, balancing KL divergence weight scheduling maintains the quality of the PC-VAE's latent space and reconstructions (see Appendix ??). Additionally, we substantiate the benefits of our PC-VAE encoder architecture through our ablations (see Appendix ??). + +3D novel view synthesis: Given an unseen ego-centric view with potentially partial observations, our method maintains all possible current state beliefs in 3D, and faithfully reconstructs novel views from arbitrary camera angles for each belief. Fig. 2 illustrates one of the possible 3D beliefs that CARFF holds. This demonstrates our method's ability to generate 3D beliefs that could be used for novel view synthesis in a view-consistent manner. Our model's ability to achieve accurate and complete 3D environmental understanding is important for applications like prediction-based planning. + +Inference under full and partial observations: Under full observation, we use MDN to predict the subsequent car positions in all three datasets. PSNR values are calculated based on bird-eye view NeRF renderings and ground truth bird-eye view images of the scene across different timestamps. In Tab. 2 we report the PSNR values for rendered images over the predicted posterior with the ground + +truth images at each timestamp. We also evaluate the efficacy of our prediction model using the accuracy curve given in Fig. 8. This represents CARFF's ability to generate stable beliefs, without producing incorrect predictions, based on actor(s) localization results. For each number of samples between $n = 0$ to $n = 50$ , we choose a random subset of 3 fully observable ego images and take an average of the accuracies. In scenarios with partial observable ego-centric images where several plausible scenarios exist, we utilize recall instead of accuracy using a similar setup. This lets us evaluate the encoder's ability to avoid false negative predictions of potential danger. + +Fig. 8 shows that our model achieves high accuracy and recall in both datasets, demonstrating the ability to model state uncertainty (Approaching Intersection) and dynamic uncertainty (Two Lane Merge). The results indicate CARFF's resilience against randomness in resampling, and completeness in probabilistic modeling of the belief space. Given these observations, we now build a reliable controller to plan and navigate through complex scenarios. + +# 4.3 Planning + +In all our experiments, the ego vehicle must make decisions to advance under certain observability. The scenarios are designed such that the ego views contain partial occlusion and the state of the actor(s) is uncertain in some scenarios. + +In order to facilitate decision-making using CARFF, we design a controller that takes ego-centric input images and outputs an action. Decisions are made incorporating sample consistency from the mixture density network. For instance, the controller infers occlusion and promotes the ego car to pause when scenes alternate between actor presence and absence in the samples. We use the two multi-scene datasets to assess the performance of the CARFF-based controller as they contain actors with potentially unknown behaviors. + +To design an effective controller, we need to find a balance between accuracy and recall (see Fig. 8). A lowered accuracy from excessive sampling means unwanted randomness in the predicted state. However, taking insufficient samples would generate low recall i.e., not recovering all plausible states. This would lead to incorrect predictions as we would be unable to account for the plausible uncertainty present in the environment. To achieve optimal balance, we designed an open-loop planning controller using a sampling strategy that generates $n = 2, 10, 35$ samples. The hyperparameter $n$ is tuned per scene for peak performance but is expected to remain relatively stable across scenes. We demonstrate that $n = 10$ performs well consistently in varying CARLA scenarios (Fig. 8) and do not anticipate this being very different for other experiments. + +For sampling values that lie on the borders of the accuracy and recall margin, for example, $n = 2$ and 35, we see that the CARFF-based controller obtains lower success rates, whereas $n = 10$ produces the best result. For actor exists and fast-actor scenes in Tab. 3, we consider occluded ego-centric inputs to test the controller's ability to avoid collisions. For no-actor and slow-actor scenes, we consider state observability and test the controllers' ability to recognize the optimal action to advance. Across the two datasets, the overconfident controller + +![](images/7e4fbc87d5e4654bd04e8bafc34f14bdc37e5f360d7db39e2aeb0852c4c0f8f8.jpg) +(a) Intersection + +![](images/9d432054d0fdcabc8b1e68591090396a8dfca8f51b511df57b933c4cde202da2.jpg) +(b) Lane Merge +Fig. 8: Multi-Scene dataset accuracy and recall curves from predicted beliefs. We test our framework across $n = 1$ and $n = 50$ samples from MDN's predicted latent distributions from ego-centric image input. Across the number of samples $n$ , we achieve an ideal margin of belief state coverage generated under partial observation (recall), and the proportion of correct beliefs sampled under full observation (accuracy). As we significantly increase the number of samples, the accuracy starts to decrease due to randomness in latent distribution resampling. + +will inevitably experience collisions in case of a truck approaching, since it does not cautiously account for occlusions. On the other hand, an overly cautious approach results in stasis, inhibiting the controller's ability to advance in the scene. This nuanced decision-making using CARFF-based controller is especially crucial in driving scenarios, as it enhances safety and efficiency by adapting to complex and unpredictable road environments, thereby fostering a more reliable and human-like response in autonomous vehicles. + +# 5 Discussion + +Limitations: Like other NeRF-based methods, CARFF currently relies on posed images of specific scenes such as road intersections, limiting its direct applicability to unseen environments. However, we anticipate enhanced generalizability with the increasing deployment of cameras around populated areas, such as traffic cameras at intersections. Additionally, handling very complex dynamics with an extremely large number of actors still poses a challenge for our method, requiring per-scene optimization to balance comprehensive dynamics modeling against accuracy. Potentially stronger models in the near future may offer a promising avenue for further enhancements in this regard. + +Conclusion: We present CARFF, a novel method for probabilistic 3D scene forecasting from partial observations. By employing a Pose-Conditional VAE, a NeRF conditioned on the learned posterior, and a mixture density network that forecasts future scenes, we effectively model, predict, and plan in complex environments with state and dynamics uncertainty in a POMDP. We further demonstrate the capabilities of our method in simulated autonomous driving scenarios. Overall, CARFF offers an intuitive framework to perceiving, forecasting, and acting under uncertainty that could prove invaluable for vision-based algorithms in unstructured environments. + +# Bibliography + +[1] Adamkiewicz, M., Chen, T., Caccavale, A., Gardner, R., Culbertson, P., Bohg, J., Schwager, M.: Vision-only robot navigation in a neural radiance world. IEEE Robotics and Automation Letters 7(2), 4606-4613 (2022) +[2] Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Int. Conf. Comput. Vis. pp. 5855–5864 (2021) +[3] Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 130-141 (2023) +[4] Cao, J., Wang, X., Darrell, T., Yu, F.: Instance-aware predictive navigation in multi-agent environments. In: IEEE Int. Conf. on Robotics and Automation. pp. 5096-5102. IEEE (2021) +[5] Charatan, D., Li, S., Tagliasacchi, A., Sitzmann, V.: pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction (2023) +[6] Codevilla, F., Santana, E., López, A.M., Gaidon, A.: Exploring the limitations of behavior cloning for autonomous driving. In: Int. Conf. Comput. Vis. pp. 9329-9338 (2019) +[7] Deng, K., Liu, A., Zhu, J.Y., Ramanan, D.: Depth-supervised nerf: Fewer views and faster training for free. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12882-12891 (2022) +[8] Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Int. Conf. Learn. Represent. (2021) +[9] Dosovitskiy, A., Ros, G., Codevilla, F., Lopez, A., Koltun, V.: CARLA: An open urban driving simulator. In: Conf. on Robol Learning. pp. 1-16 (2017) +[10] Driess, D., Huang, Z., Li, Y., Tedrake, R., Toussaint, M.: Learning multi-object dynamics with compositional neural radiance fields. arXiv preprint arXiv:2202.11855 (2022) +[11] Fei, B., Xu, J., Zhang, R., Zhou, Q., Yang, W., He, Y.: 3d gaussian as a new vision era: A survey (2024) +[12] Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5501–5510 (2022) +[13] Hausknecht, M., Stone, P.: Deep recurrent q-learning for partially observable mdps. In: AAAI (2015) +[14] Hearst, M.A., Dumais, S.T., Osuna, E., Platt, J., Scholkopf, B.: Support vector machines. IEEE Intelligent Systems and their applications 13(4), 18-28 (1998) +[15] Ichnowski, J., Avigal, Y., Kerr, J., Goldberg, K.: Dex-nerf: Using a neural radiance field to grasp transparent objects. arXiv preprint arXiv:2110.14217 (2021) + +[16] Ivanovic, B., Elhafsi, A., Rosman, G., Gaidon, A., Pavone, M.: Mats: An interpretable trajectory forecasting representation for planning and control. In: Conf. on Robol Learning (2021) +[17] Kaelbling, L.P., Littman, M.L., Cassandra, A.R.: Planning and acting in partially observable stochastic domains. Artificial Intelligence 101(1), 99-134 (1998). https://doi.org/https://doi.org/10.1016/S0004-3702(98)00023-X, https://www.sciencedirect.com/science/article/pii/S000437029800023X +[18] Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/ +[19] Kerr, J., Fu, L., Huang, H., Avigal, Y., Tancik, M., Ichnowski, J., Kanazawa, A., Goldberg, K.: Evo-nerf: Evolving nerf for sequential robot grasping of transparent objects. In: Conf. on Robol Learning (2022) +[20] Kosiorek, A.R., Strathmann, H., Zoran, D., Moreno, P., Schneider, R., Mokrá, S., Rezende, D.J.: NeRF-VAE: A geometry aware 3d scene generative model. In: ICML. pp. 5742–5752 (2021) +[21] Li, Y., Li, S., Sitzmann, V., Agrawal, P., Torralba, A.: 3d neural scene representations for visuomotor control. In: Conf. on Robol Learning. pp. 112-123 (2022) +[22] Liu, J.W., Cao, Y.P., Mao, W., Zhang, W., Zhang, D.J., Keppo, J., Shan, Y., Qie, X., Shou, M.Z.: Devrf: Fast deformable voxel radiance fields for dynamic scenes. In: Adv. Neural Inform. Process. Syst. vol. 35, pp. 36762-36775 (2022) +[23] Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis (2023) +[24] van der Maaten, L., Hinton, G.: Visualizing data using t-SNE. Journal of Machine Learning Research 9, 2579-2605 (2008), http://www.jmlr.org/papers/v9/vandermaaten08a.html +[25] Marza, P., Matignon, L., Simonin, O., Wolf, C.: Multi-object navigation with dynamically learned neural implicit representations. In: Int. Conf. Comput. Vis. pp. 11004-11015 (2023) +[26] McAllister, R., Rasmussen, C.E.: Data-efficient reinforcement learning in continuous state-action gaussian-pomdpds. In: Adv. Neural Inform. Process. Syst. (2017) +[27] Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: Eur. Conf. Comput. Vis. (2020) +[28] Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 1-15 (2022) +[29] Packer, C., Rhinehart, N., McAllister, R.T., Wright, M.A., Wang, X., He, J., Levine, S., Gonzalez, J.E.: Is anyone there? learning a planner contingent on perceptual uncertainty. In: Conf. on Robol Learning (2022) + +[30] Packer, C., Rhinehart, N., McAllister, R.T., Wright, M.A., Wang, X., He, J., Levine, S., Gonzalez, J.E.: Is anyone there? learning a planner contingent on perceptual uncertainty. In: Liu, K., Kulic, D., Ichnowski, J. (eds.) Conf. on Robol Learning. pp. 1607-1617 (2023) +[31] Pan, X., You, Y., Wang, Z., Lu, C.: Virtual to real reinforcement learning for autonomous driving. arXiv preprint arXiv:1704.03952 (2017) +[32] Papadimitriou, C.H., Tsitsiklis, J.N.: The complexity of markov decision processes. Mathematics of operations research 12(3), 441-450 (1987) +[33] Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: Int. Conf. Comput. Vis. pp. 5865–5874 (2021) +[34] Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph. (2021) +[35] Pineau, J., Gordon, G., Thrun, S., et al.: Point-based value iteration: An anytime algorithm for pomdpps. In: IJCAI. vol. 3, pp. 1025-1032 (2003) +[36] Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-NeRF: Neural Radiance Fields for Dynamic Scenes. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020) +[37] Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2016) +[38] Rhinehart, N., He, J., Packer, C., Wright, M.A., McAllister, R., Gonzalez, J.E., Levine, S.: Contingencies from observations: Tractable contingency planning with learned behavior models. In: IEEE Int. Conf. on Robotics and Automation. pp. 13663-13669 (2021) +[39] Roessle, B., Barron, J.T., Mildenhall, B., Srinivasan, P.P., Nießner, M.: Dense depth priors for neural radiance fields from sparse input views. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12892–12901 (2022) +[40] Shao, R., Zheng, Z., Tu, H., Liu, B., Zhang, H., Liu, Y.: Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 16632-16642 (2023) +[41] Sun, C., Sun, M., Chen, H.T.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5459-5469 (2022) +[42] Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8248–8258 (2022) +[43] Tancik, M., Srinivasan, P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. In: Adv. Neural Inform. Process. Syst. pp. 7537–7547 (2020) + +[44] Tas, O.S., Stiller, C.: Limited visibility and uncertainty aware motion planning for automated driving. In: IEEE Intelligent Vehicles Symposium (IV) (jun 2018) +[45] Tewari, A., Thies, J., Mildenhall, B., Srinivasan, P., Tretschk, E., Yifan, W., Lassner, C., Sitzmann, V., Martin-Brualla, R., Lombardi, S., Simon, T., Theobalt, C., Nießner, M., Barron, J.T., Wetzstein, G., Zollhöfer, M., Golyanik, V.: Advances in Neural Rendering. Comput. Graph. Forum (2022) +[46] Toromanoff, M., Wirbel, E., Moutarde, F.: End-to-end model-free reinforcement learning for urban driving using implicit affordances. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7153-7162 (2020) +[47] Tretschk, E., Golyanik, V., Zollhoefer, M., Bozic, A., Lassner, C., Theobalt, C.: Scenerflow: Time-consistent reconstruction of general dynamic scenes. In: International Conference on 3D Vision (3DV) (2023) +[48] Truong, P., Rakotosaona, M.J., Manhardt, F., Tombari, F.: Sparf: Neural radiance fields from sparse and noisy poses. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4190-4200 (2023) +[49] Turki, H., Ramanan, D., Satyanarayanan, M.: Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12922–12931 (2022) +[50] Wang, Q., Wang, Z., Genova, K., Srinivasan, P.P., Zhou, H., Barron, J.T., Martin-Brualla, R., Snavely, N., Funkhouser, T.: Ibrnet: Learning multi-view image-based rendering. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4690-4699 (2021) +[51] Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5438-5448 (2022) +[52] Yen-Chen, L., Florence, P., Barron, J.T., Rodriguez, A., Isola, P., Lin, T.Y.: inerf: Inverting neural radiance fields for pose estimation. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1323-1330. IEEE (2021) +[53] Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images (2021) +[54] Yu, M.Y., Vasudevan, R., Johnson-Roberson, M.: Occlusion-aware risk assessment for autonomous driving in urban environments. IEEE Robotics and Automation Letters 4(2), 2235-2241 (Apr 2019). https://doi.org/10.1109/lra.2019.2900453, http://dx.doi.org/10.1109/LRA.2019.2900453 +[55] Zhang, C., Steinhauser, F., Hinz, G., Knoll, A.: Improved occlusion scenario coverage with a pomdp-based behavior planner for autonomous urban driving. In: 2021 IEEE International Intelligent Transportation Systems Conference (ITSC). pp. 593-600 (2021). https://doi.org/10.1109/ITSC48978.2021.9564424 +[56] Zhang, C., Steinhauser, F., Hinz, G., Knoll, A.: Occlusion-aware planning for autonomous driving with vehicle-to-everything communication. IEEE Transactions on Intelligent Vehicles 9(1), 1229-1242 (2024). https://doi.org/10.1109/TIV.2023.3308098 \ No newline at end of file diff --git a/2024/CARFF_ Conditional Auto-encoded Radiance Field for 3D Scene Forecasting/images.zip b/2024/CARFF_ Conditional Auto-encoded Radiance Field for 3D Scene Forecasting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..eb1c309cf7cd268f7c2578a4df00b06fa71d614f --- /dev/null +++ b/2024/CARFF_ Conditional Auto-encoded Radiance Field for 3D Scene Forecasting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5cbfac1c9f4d9140aaeb3b0d2cd40c0e0bb8e4308ee31f658ea8656ce8b40a0a +size 329917 diff --git a/2024/CARFF_ Conditional Auto-encoded Radiance Field for 3D Scene Forecasting/layout.json b/2024/CARFF_ Conditional Auto-encoded Radiance Field for 3D Scene Forecasting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bdd831ca0135302d17c85383406c95441555b81c --- /dev/null +++ b/2024/CARFF_ Conditional Auto-encoded Radiance Field for 3D Scene Forecasting/layout.json @@ -0,0 +1,9262 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 111, + 465, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 111, + 465, + 148 + ], + "spans": [ + { + "bbox": [ + 148, + 111, + 465, + 148 + ], + "type": "text", + "content": "CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 149, + 173, + 465, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 173, + 465, + 210 + ], + "spans": [ + { + "bbox": [ + 149, + 173, + 465, + 210 + ], + "type": "text", + "content": "Jiezhi \"Stephen\" Yang" + }, + { + "bbox": [ + 149, + 173, + 465, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 173, + 465, + 210 + ], + "type": "text", + "content": ", Khushi Desai" + }, + { + "bbox": [ + 149, + 173, + 465, + 210 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 173, + 465, + 210 + ], + "type": "text", + "content": ", Charles Packer" + }, + { + "bbox": [ + 149, + 173, + 465, + 210 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 149, + 173, + 465, + 210 + ], + "type": "text", + "content": ", Harshil Bhatia" + }, + { + "bbox": [ + 149, + 173, + 465, + 210 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 149, + 173, + 465, + 210 + ], + "type": "text", + "content": ", Nicholas Rhinehart" + }, + { + "bbox": [ + 149, + 173, + 465, + 210 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 149, + 173, + 465, + 210 + ], + "type": "text", + "content": ", Rowan McAllister" + }, + { + "bbox": [ + 149, + 173, + 465, + 210 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 149, + 173, + 465, + 210 + ], + "type": "text", + "content": ", and Joseph E. Gonzalez" + }, + { + "bbox": [ + 149, + 173, + 465, + 210 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 228, + 220, + 386, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 220, + 386, + 230 + ], + "spans": [ + { + "bbox": [ + 228, + 220, + 386, + 230 + ], + "type": "text", + "content": "1 Harvard University MA 02138, USA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 227, + 231, + 387, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 231, + 387, + 242 + ], + "spans": [ + { + "bbox": [ + 227, + 231, + 387, + 242 + ], + "type": "text", + "content": "2 Columbia University NY 10025, USA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 243, + 243, + 371, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 243, + 371, + 252 + ], + "spans": [ + { + "bbox": [ + 243, + 243, + 371, + 252 + ], + "type": "text", + "content": "UC Berkeley CA 94720, USA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 244, + 253, + 370, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 253, + 370, + 263 + ], + "spans": [ + { + "bbox": [ + 244, + 253, + 370, + 263 + ], + "type": "text", + "content": "4 Avataar.ai KA 560103, India" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 216, + 264, + 397, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 264, + 397, + 274 + ], + "spans": [ + { + "bbox": [ + 216, + 264, + 397, + 274 + ], + "type": "text", + "content": "5 Toyota Research Institute CA 94022, USA" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 305, + 455, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 305, + 455, + 482 + ], + "spans": [ + { + "bbox": [ + 159, + 305, + 455, + 482 + ], + "type": "text", + "content": "Abstract. We propose CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting, a method for predicting future 3D scenes given past observations. Our method maps 2D ego-centric images to a distribution over plausible 3D latent scene configurations and predicts the evolution of hypothesized scenes through time. Our latents condition a global Neural Radiance Field (NeRF) to represent a 3D scene model, enabling explainable predictions and straightforward downstream planning. This approach models the world as a POMDP and considers complex scenarios of uncertainty in environmental states and dynamics. Specifically, we employ a two-stage training of Pose-Conditional-VAE and NeRF to learn 3D representations, and auto-regressively predict latent scene representations utilizing a mixture density network. We demonstrate the utility of our method in scenarios using the CARLA driving simulator, where CARFF enables efficient trajectory and contingency planning in complex multi-agent autonomous driving scenarios involving occlusions. Video and code are available at: www.carff website." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 502, + 230, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 230, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 230, + 514 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 527, + 482, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 527, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 527, + 482, + 622 + ], + "type": "text", + "content": "Humans often imagine what they cannot see given partial visual context. Consider a scenario where reasoning about the unobserved is critical to safe decision-making: for example, a driver navigating a blind intersection. An expert driver will plan according to what they believe may or may not exist in occluded regions of their vision. The driver's belief - defined as the understanding of the world modeled with consideration for inherent environment uncertainties - is informed by their partial observations (i.e., the presence of other vehicles on the road), as well as their prior knowledge (e.g., past experience navigating this intersection)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 623, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 623, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 131, + 623, + 482, + 647 + ], + "type": "text", + "content": "When reasoning about the unobserved, humans form complex beliefs about the existence, position, shapes, colors, and textures of occluded scene portions" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 217, + 664 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 217, + 664 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 217, + 664 + ], + "type": "text", + "content": "* Core contributors" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "(e.g., an oncoming car). Autonomous systems with high-dimensional sensor data, like video or LiDAR, traditionally reduce this data to low-dimensional state information (e.g., position and velocity of tracked objects) for prediction and planning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 479, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 479, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 479, + 212 + ], + "type": "text", + "content": "In addition to tracking fully observed objects, this object-centric framework handles partially observed settings by considering potentially dangerous unobserved objects. These systems often plan for worst-case scenarios, such as a \"ghost car\" at the edge of the visible field of view (44)." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 204, + 236, + 413, + 327 + ], + "blocks": [ + { + "bbox": [ + 204, + 236, + 413, + 327 + ], + "lines": [ + { + "bbox": [ + 204, + 236, + 413, + 327 + ], + "spans": [ + { + "bbox": [ + 204, + 236, + 413, + 327 + ], + "type": "image", + "image_path": "b84fe5410986cc3ec666ba1809f5163c18eadc7a63813e26b65137ad51372830.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 337, + 482, + 403 + ], + "lines": [ + { + "bbox": [ + 130, + 337, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 482, + 403 + ], + "type": "text", + "content": "Fig. 1: CARFF 3D planning application for driving. An input image containing a partially observable view of an intersection is processed by CARFF's encoder to establish 3D environment state beliefs, i.e. the predicted possible state of the world: whether or not there could be another vehicle approaching the intersection. These beliefs are used to forecast the future in 3D for planning, generating one among two possible actions for the vehicle to merge into the other lane." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": "Recent advances in neural rendering, particularly Neural Radiance Fields (NeRF), have significantly improved 3D scene representation learning. NeRF enables novel view synthesis, thus simplifying the process of viewing behind occlusions. NeRF decouples the dependency of scene representation from traditional object detection and tracking, allowing for the capture of vital visual information that might be missed by detectors, yet is crucial for safe decision-making. NeRF's implicit density representation of explicit geometry also facilitates its direct application in motion planning without the need for rendering. NeRF's ability to represent both visual and geometric information makes them a more general and intuitive 3D representation for autonomous systems." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Despite NeRF's advantages, achieving probabilistic predictions in 3D based on reasoning from occluded views is challenging. For example, discriminative models that yield categorical predictions are unable to capture the underlying 3D structure, impeding their ability to model uncertainty. While prior work on 3D representation captures view-invariant structures, their application is primarily confined to simple scenarios (20). We present CARFF, which to our knowledge, is the first forecasting approach in scenarios with partial observations that uniquely facilitates stochastic predictions in a partially observable Markov decision process (POMDP) within a 3D representation, effectively integrating visual perception and geometry. Specifically, we make the following contributions:" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 136, + 116, + 481, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 116, + 481, + 163 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 481, + 163 + ], + "type": "text", + "content": "1. We propose a novel architecture " + }, + { + "bbox": [ + 136, + 116, + 481, + 163 + ], + "type": "inline_equation", + "content": "PC" + }, + { + "bbox": [ + 136, + 116, + 481, + 163 + ], + "type": "text", + "content": "-VAE: Pose-Conditioned Variational Autoencoder. The encoder maps potentially partially observable ego-centric images to pose-invariant latent scene representations, which hold state beliefs of the POMDP with implicit probability distributions (see Sec. 3.1)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 164, + 481, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 164, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 136, + 164, + 481, + 209 + ], + "type": "text", + "content": "2. We develop the two-stage training pipeline that uniquely enables complex scene modeling with a probabilistic objective. This involves separately training the PC-VAE and a latent conditioned neural radiance field that functions as a 3D decoder, enabling interpretable predictions (see Sec. 3.1)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 211, + 481, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 211, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 136, + 211, + 481, + 257 + ], + "type": "text", + "content": "3. We design a mixture density model to predict the evolution of 3D scenes over time stochastically and regressively in the encoder belief space (see Sec. 3.2). This allows for an effective sampling based-controller to output actions in the POMDP." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 263, + 481, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 481, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 481, + 312 + ], + "type": "text", + "content": "We demonstrate how CARFF can be used to enable contingency planning in complex driving scenarios that require reasoning into visual occlusions on CARLA simulated datasets inspired by autonomous driving planning tasks (29; 54; 55; 56). A potential application of CARFF is illustrated in Fig. 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 327, + 234, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 234, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 234, + 339 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 349, + 312, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 349, + 312, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 349, + 312, + 361 + ], + "type": "text", + "content": "2.1 NeRF and 3D representations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 365, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 511 + ], + "type": "text", + "content": "Neural radiance fields. Neural Radiance Fields (NeRF) (2; 27; 43) for 3D representations generate high-resolution, photorealistic scenes. Instant Neural Graphics Primitive (Instant-NGP) (28) speeds up training and rendering time by introducing a multi-resolution hash encoding. Other works like Plenoxels (12) and DirectVoxGo (DVGO) (41) also provide similar speedups. Recent advancements in volumetric representations such as 3D Gaussian Splatting (18) enhance rendering efficiency while maintaining compatibility with traditional NeRF applications (11). We utilize Instant-NGP for its accessibility, although our approach is adaptable to alternative rendering methods. NeRFs have also been extended for several tasks such as modeling large-scale unbounded scenes (2; 42; 49), scene from sparse views (7; 39; 48) and multiple scenes (20; 50). For an in-depth survey on neural representation learning and its applications we refer the reader to (45)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Generalizable novel view synthesis models, like pixelNeRF and pixelSplat (5; 53), learn a scene prior to render novel views from sparse existing ones. In contrast, CARFF is based on a VAE, encoding a probabilistic objective and decoding to future 3D scenes. Dynamic NeRF models scenes with moving or deforming objects, within which a widely used approach is to construct a canonical space and predict a deformation field (22; 33; 34; 36). The canonical space is usually a static scene, and the model learns an implicitly represented flow field (33; 36). A recent line of work also models dynamic scenes via different representations and decomposition (3; 40). These approaches tend to perform better for spatially bounded and predictable scenes with relatively small variations (3; 23; 33; 53). Moreover, these methods only solve for changes in the environment but are limited in incorporating stochasticity in the environment." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 134, + 90, + 447, + 103 + ], + "type": "text", + "content": "CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 241, + 117, + 304, + 169 + ], + "blocks": [ + { + "bbox": [ + 241, + 117, + 304, + 169 + ], + "lines": [ + { + "bbox": [ + 241, + 117, + 304, + 169 + ], + "spans": [ + { + "bbox": [ + 241, + 117, + 304, + 169 + ], + "type": "image", + "image_path": "911c9359ba538544b664309f060185792214c88863a23d0e2a3844cb4b919ec8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 169, + 291, + 175 + ], + "lines": [ + { + "bbox": [ + 253, + 169, + 291, + 175 + ], + "spans": [ + { + "bbox": [ + 253, + 169, + 291, + 175 + ], + "type": "text", + "content": "Ego-centric Lane View" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 182, + 482, + 228 + ], + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 228 + ], + "type": "text", + "content": "Fig. 2: Novel view planning application. CARFF allows reasoning behind occluded views from the ego car as simple as moving the camera to see the sampled belief predictions, allowing simple downstream planning using, for example, density probing or 2D segmentation models from arbitrary angles." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 117, + 376, + 169 + ], + "blocks": [ + { + "bbox": [ + 307, + 117, + 376, + 169 + ], + "lines": [ + { + "bbox": [ + 307, + 117, + 376, + 169 + ], + "spans": [ + { + "bbox": [ + 307, + 117, + 376, + 169 + ], + "type": "image", + "image_path": "7f904b0876f8b554d72dcdbc4a9168a495f5a5cb691242159c4c1ec33921f7dd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 169, + 353, + 175 + ], + "lines": [ + { + "bbox": [ + 323, + 169, + 353, + 175 + ], + "spans": [ + { + "bbox": [ + 323, + 169, + 353, + 175 + ], + "type": "text", + "content": "Predicted 3D View" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 256, + 482, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 256, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 482, + 340 + ], + "type": "text", + "content": "Multi-scene NeRF: Our approach builds on multi-scene NeRF approaches (20; 47; 50; 51) that learn a global latent scene representation, which conditions the NeRF, allowing a single NeRF to effectively represent various scenes. A similar method, NeRF-VAE, was introduced by Kosiorek et al. (20) to create a geometrically consistent 3D generative model with generalization to out-of-distribution cameras. However, NeRF-VAE (20) is prone to mode collapse when faced with complex visual information (see Sec. 4.2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 370, + 251, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 251, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 251, + 383 + ], + "type": "text", + "content": "2.2 Scene Forecasting" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 402, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 510 + ], + "type": "text", + "content": "Planning in 2D space: Planning in large, continuous state-action spaces is challenging due to exponentially large search spaces (32), leading to various approximation methods for tractability (26; 35). Model-free (13; 31; 46) and model-based (4) reinforcement learning frameworks, along with other learning-based methods (6; 29), have emerged as viable approaches. Additionally, methods forecast for downstream control (16), learn behavior models for contingency planning (38), or predict the existence and intentions of unobserved agents (30). While these methods operate in 2D, we reason under partial observations and account for these factors in 3D." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": "NeRF in robotics: Recent works have applied NeRFs in robotics for localization (52), navigation (1; 25), dynamics modeling (10; 22), and robotic grasping (15; 19). Adamkiewicz et al.(1) propose quadcopter motion planning in NeRF models by sampling the learned density function, useful for forecasting and planning. Driess et al.(10) employ a graph neural network to learn dynamics in a multi-object NeRF scene. Li et al. (21) focus on pushing tasks and address grasping and planning with NeRF and a separate latent dynamics model. Prior approaches work in simple, static scenes (1) or uses deterministic dynamics models (21). CARFF addresses complex, realistic environments with both state and dynamics uncertainty, considering potential object existence and unknown movements." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 202, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 202, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 202, + 127 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 246 + ], + "type": "text", + "content": "Recent advancements in 3D scene representation allow for modeling environments in a contextually rich and interactive 3D space. This offers analytical benefits, such as spatial analysis with soft occupancy grids and object detection through novel view synthesis. Given these advantages, our primary objective is to develop a model for probabilistic 3D scene forecasting in dynamic environments. However, direct integration of 3D scene representation via NeRF and probabilistic models like VAE often involves non-convex and inter-dependent optimization, which causes unstable training. For instance, NeRF's optimization may rely on the VAE's latent space being structured to provide informative gradients. To" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 151, + 265, + 465, + 393 + ], + "blocks": [ + { + "bbox": [ + 151, + 265, + 465, + 393 + ], + "lines": [ + { + "bbox": [ + 151, + 265, + 465, + 393 + ], + "spans": [ + { + "bbox": [ + 151, + 265, + 465, + 393 + ], + "type": "image", + "image_path": "a431e18c5007cafac98de48641d9202701cb8965d45e1dc5703eb3ff32b6ada2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 401, + 482, + 511 + ], + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 511 + ], + "type": "text", + "content": "Fig. 3: Visualizing CARFF's two stage training process. Left: The convolutional VIT-based encoder encodes each image " + }, + { + "bbox": [ + 130, + 401, + 482, + 511 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 401, + 482, + 511 + ], + "type": "text", + "content": " at timestamps " + }, + { + "bbox": [ + 130, + 401, + 482, + 511 + ], + "type": "inline_equation", + "content": "t, t'" + }, + { + "bbox": [ + 130, + 401, + 482, + 511 + ], + "type": "text", + "content": " and camera poses " + }, + { + "bbox": [ + 130, + 401, + 482, + 511 + ], + "type": "inline_equation", + "content": "c, c'" + }, + { + "bbox": [ + 130, + 401, + 482, + 511 + ], + "type": "text", + "content": " into Gaussian latent distributions. Assuming two timestamps and an overparameterized latent, one Gaussian distribution will have a smaller " + }, + { + "bbox": [ + 130, + 401, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 130, + 401, + 482, + 511 + ], + "type": "text", + "content": ", and different " + }, + { + "bbox": [ + 130, + 401, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 401, + 482, + 511 + ], + "type": "text", + "content": " across timestamps. Upper Right: The pose-conditional decoder stochastically decodes the sampled latent " + }, + { + "bbox": [ + 130, + 401, + 482, + 511 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 401, + 482, + 511 + ], + "type": "text", + "content": " using the camera pose " + }, + { + "bbox": [ + 130, + 401, + 482, + 511 + ], + "type": "inline_equation", + "content": "c''" + }, + { + "bbox": [ + 130, + 401, + 482, + 511 + ], + "type": "text", + "content": " into images " + }, + { + "bbox": [ + 130, + 401, + 482, + 511 + ], + "type": "inline_equation", + "content": "I_{c''}^t" + }, + { + "bbox": [ + 130, + 401, + 482, + 511 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 401, + 482, + 511 + ], + "type": "inline_equation", + "content": "I_{c''}^{t'}" + }, + { + "bbox": [ + 130, + 401, + 482, + 511 + ], + "type": "text", + "content": ". The decoded reconstruction and ground truth images are used for the loss " + }, + { + "bbox": [ + 130, + 401, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MSE}}" + }, + { + "bbox": [ + 130, + 401, + 482, + 511 + ], + "type": "text", + "content": ", PC-VAE. Lower Right: A NeRF is trained by conditioning on the latent variables sampled from the optimized Gaussian parameters. These parameters characterize the distinct timestamp distributions derived from the PC-VAE. An MSE loss is calculated for NeRF as " + }, + { + "bbox": [ + 130, + 401, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MSE}}" + }, + { + "bbox": [ + 130, + 401, + 482, + 511 + ], + "type": "text", + "content": ", NeRF." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "navigate these complexities, our method bifurcates the training process into two stages (see Fig. 3). First, we train the PC-VAE to learn view-invariant scene representations. Next, we replace the decoder with a NeRF to learn a 3D scene from the latent representations. The latent scene representations capture the environmental states and dynamics over possible underlying scenes, while NeRF synthesizes novel views within the belief space, giving us the ability to see the unobserved (see Fig. 2 and Sec. 3.1). During prediction, uncertainties can be modeled by sampling latents auto-regressively from a predicted Gaussian mixture, allowing for effective decision-making. To this extent, we approach scene forecasting as a POMDP over latent distributions, which enables us to capture multi-modal beliefs for planning amidst perceptual uncertainty (see Sec. 3.2)." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 90, + 447, + 102 + ], + "type": "text", + "content": "CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 386, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 386, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 386, + 129 + ], + "type": "text", + "content": "3.1 Pose-Conditional VAE (PC-VAE) and NeRF" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "text", + "content": "Architecture: We assume that the model follows a Markovian process, and thus each belief state only depends on the previous. Given a scene " + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "text", + "content": " at timestamp " + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "text", + "content": ", we have an ego-centric observation image " + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "inline_equation", + "content": "I_{c}^{t}" + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "text", + "content": " captured from camera pose " + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "text", + "content": ". The objective is to formulate a 3D representation of the image that holds implicit probability distributions of the possible states, where we can perform a forecasting step that evolves the scene forward. Here, the POMDP can be seen as an MDP in belief space (17). To achieve this, we utilize a radiance field conditioned on latent variable " + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "text", + "content": " sampled from the posterior distribution " + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "inline_equation", + "content": "q_{\\phi}(z|I_{c}^{t})" + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "text", + "content": ". Now, to learn the posterior, we utilize PC-VAE. We construct an encoder using convolutional layers and a pre-trained ViT on ImageNet (8). The encoder learns a mapping from the image space to a Gaussian distributed latent space " + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "inline_equation", + "content": "q_{\\phi}(z|I_{c}^{t}) = \\mathcal{N}(\\mu, \\sigma^{2})" + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "text", + "content": " parametrized by mean " + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "text", + "content": ". The decoder, " + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "inline_equation", + "content": "p(I|z,c)" + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "text", + "content": ", conditioned on camera pose " + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "text", + "content": ", maps the latent " + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "inline_equation", + "content": "z \\sim \\mathcal{N}(\\mu, \\sigma^2)" + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "text", + "content": " into the image space " + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "text", + "content": ". This helps the encoder to generate latents that are invariant to the camera pose " + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 134, + 482, + 312 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 313, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 482, + 361 + ], + "type": "text", + "content": "To enable 3D scene modeling, we employ Instant-NGP (28), which incorporates a hash grid and an occupancy grid to enhance computation efficiency. Additionally, a smaller multilayer perceptron (MLP), " + }, + { + "bbox": [ + 130, + 313, + 482, + 361 + ], + "type": "inline_equation", + "content": "F_{\\theta}(z)" + }, + { + "bbox": [ + 130, + 313, + 482, + 361 + ], + "type": "text", + "content": " can be utilized to model the density and appearance, given by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 242, + 369, + 481, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 369, + 481, + 381 + ], + "spans": [ + { + "bbox": [ + 242, + 369, + 481, + 381 + ], + "type": "interline_equation", + "content": "F _ {\\theta} (z): (\\mathbf {x}, \\mathbf {d}, z) \\rightarrow ((r, g, b), \\sigma) \\tag {1}", + "image_path": "ce4a16e2f3e7cad9ad00e51decc460144b0bba59e79f31cd8cc576f06db327d9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 388, + 481, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 388, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 131, + 388, + 481, + 426 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 388, + 481, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 131, + 388, + 481, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 388, + 481, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{d} \\in (\\theta, \\phi)" + }, + { + "bbox": [ + 131, + 388, + 481, + 426 + ], + "type": "text", + "content": " represent the location vector and the viewing direction respectively. The MLP is conditioned on the sampled scene latents " + }, + { + "bbox": [ + 131, + 388, + 481, + 426 + ], + "type": "inline_equation", + "content": "z \\sim q_{\\phi}(z|I_c^t)" + }, + { + "bbox": [ + 131, + 388, + 481, + 426 + ], + "type": "text", + "content": " (see Appendix ??)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 434, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 434, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 131, + 434, + 482, + 578 + ], + "type": "text", + "content": "Training methodology: The architecture alone does not enable us to model complex scenarios, as seen through a similar example in NeRF-VAE (20). A crucial contribution of our work is our two-stage training framework which stabilizes the training. First, we optimize the convolutional ViT based encoder and pose-conditional convolutional decoder in the pixel space for reconstruction. This enables our method to deal with more complex and realistic scenes as the encoding is learned in a semantically rich 2D space. By conditioning the decoder on camera poses, we achieve disentanglement between camera view angles and scene context, making the representation view-invariant and the encoder 3D-aware. Once rich latent representations are learned, we replace the decoder with a latent-conditioned NeRF over the latent space of the frozen encoder. The NeRF reconstructs encoder beliefs in 3D for novel view synthesis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 587, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 482, + 612 + ], + "type": "text", + "content": "Loss: PC-VAE is trained using standard VAE loss, with mean square error (MSE) and a Kullback-Leibler (KL) divergence given by evidence lower bound:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 144, + 635, + 480, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 635, + 480, + 662 + ], + "spans": [ + { + "bbox": [ + 144, + 635, + 480, + 662 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {P C - V A E} = \\mathcal {L} _ {M S E, P C - V A E} + \\mathcal {L} _ {K L D, P C - V A E} = \\\\ \\left\\| p \\left(I | z, c ^ {\\prime \\prime}\\right) - I _ {c ^ {\\prime \\prime}} ^ {t} \\right\\| ^ {2} + \\mathbb {E} _ {q (z \\mid I _ {c} ^ {t})} [ \\log p (I | z) ] - w _ {K L} D _ {K L} \\left(q _ {\\phi} \\left(z \\mid I _ {c} ^ {t}\\right) \\mid p (I | z)\\right) \\tag {2} \\\\ \\end{array}", + "image_path": "7c81541fe7771db0e78f541bc47747ac2549398379def7dedf52400c4cd4c5bb.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "w_{KL}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " denotes the KL divergence loss weight and " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "z \\sim q_{\\phi}(z|I_c^t)" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": ". To make our representation 3D-aware, our posterior is encoded using camera " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " while the decoder is conditioned on a randomly sampled pose " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "c''" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "text", + "content": "KL divergence regularizes the latent space to balance conditioned reconstruction and stochasticity under occlusion. An elevated KL divergence loss weight " + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "inline_equation", + "content": "w_{KL}" + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "text", + "content": " pushes the latents closer to a standard normal distribution, " + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,1)" + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "text", + "content": ", thereby ensuring probabilistic sampling in scenarios under partial observation. However, excessive regularization causes the latents to be less separable, leading to mode collapse. To mitigate this, we adopt delayed linear KL divergence loss weight scheduling to strike a balanced " + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "inline_equation", + "content": "w_{KL}" + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 479, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 479, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 479, + 271 + ], + "type": "text", + "content": "Next, we learn a NeRF decoder on the posterior of the VAE to model scenes. At any timestamp " + }, + { + "bbox": [ + 130, + 236, + 479, + 271 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 236, + 479, + 271 + ], + "type": "text", + "content": " we use a standard photometric loss for training the NeRF, given by the following equation:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 210, + 278, + 480, + 292 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 278, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 210, + 278, + 480, + 292 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {M S E, N e R F} = \\left\\| I _ {c} ^ {t} - r e n d e r \\left(F _ {\\theta} \\left(\\cdot \\mid q _ {\\phi} \\left(z \\mid I _ {c} ^ {t}\\right)\\right)\\right) \\right\\| ^ {2} \\tag {3}", + "image_path": "9869eb5b716af49fa66458e57856ff0ba9609aff3c27e40e9007d7c506dbcca3.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 297, + 482, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 332 + ], + "type": "text", + "content": "We use a standard rendering algorithm as proposed by Müller et al. (28). Next, we build a forecasting module over the learned latent space of our pose-conditional encoder." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 349, + 250, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 349, + 250, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 349, + 250, + 361 + ], + "type": "text", + "content": "3.2 Scene Forecasting" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 366, + 479, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 479, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 479, + 426 + ], + "type": "text", + "content": "Formulation: The current formulation allows us to model scenes with different configurations across timestamps. In order to forecast future configurations of a scene given an ego-centric view, we need to predict future latent distributions. We formulate the forecasting as a POMDP over the posterior distribution " + }, + { + "bbox": [ + 130, + 366, + 479, + 426 + ], + "type": "inline_equation", + "content": "q_{\\phi}(z|I_c^t)" + }, + { + "bbox": [ + 130, + 366, + 479, + 426 + ], + "type": "text", + "content": " in the PC-VAE's latent space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 426, + 479, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 479, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 479, + 486 + ], + "type": "text", + "content": "During inference, we observe stochastic behaviors under occlusion, which motivates us to learn a mixture of several Gaussian distributions that potentially denote different scene possibilities. Therefore, we model the POMDP using a Mixture Density Network (MDN), with multi-headed MLPs, that predicts a mixture of " + }, + { + "bbox": [ + 130, + 426, + 479, + 486 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 426, + 479, + 486 + ], + "type": "text", + "content": " Gaussians. At any timestamp " + }, + { + "bbox": [ + 130, + 426, + 479, + 486 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 426, + 479, + 486 + ], + "type": "text", + "content": " the distribution is given as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 230, + 493, + 480, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 493, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 230, + 493, + 480, + 506 + ], + "type": "interline_equation", + "content": "q _ {\\phi} ^ {\\prime} \\left(z _ {t} \\mid I _ {c} ^ {t - 1}\\right) = M D N \\left(q _ {\\phi} \\left(z _ {t - 1} \\mid I _ {c} ^ {t - 1}\\right)\\right) \\tag {4}", + "image_path": "32f1b5d748dd382146d293f205db1c4c2350c17e3a76da88e918b0bd4ac41732.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 512, + 479, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 479, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 479, + 548 + ], + "type": "text", + "content": "The model is conditioned on the posterior distribution " + }, + { + "bbox": [ + 130, + 512, + 479, + 548 + ], + "type": "inline_equation", + "content": "q_{\\phi}(z_{t - 1})" + }, + { + "bbox": [ + 130, + 512, + 479, + 548 + ], + "type": "text", + "content": " to learn a predicted posterior distribution " + }, + { + "bbox": [ + 130, + 512, + 479, + 548 + ], + "type": "inline_equation", + "content": "q_{\\phi}'(z_t|I_c^{t - 1})" + }, + { + "bbox": [ + 130, + 512, + 479, + 548 + ], + "type": "text", + "content": " at each timestamp. The predicted posterior distribution is given by the mixture of Gaussian:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 252, + 555, + 480, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 555, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 252, + 555, + 480, + 586 + ], + "type": "interline_equation", + "content": "q _ {\\phi} ^ {\\prime} \\left(z _ {t}\\right) = \\sum_ {i = 1} ^ {K} \\pi_ {i} \\mathcal {N} \\left(\\mu_ {i}, \\sigma_ {i} ^ {2}\\right) \\tag {5}", + "image_path": "fba9dafa1ac820c3c663137ab8be340ffc92c5e93defb96fd9294e73a57e3f26.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 593, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 593, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 593, + 479, + 665 + ], + "type": "text", + "content": "here, " + }, + { + "bbox": [ + 131, + 593, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\pi_i" + }, + { + "bbox": [ + 131, + 593, + 479, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 593, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mu_i" + }, + { + "bbox": [ + 131, + 593, + 479, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 593, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\sigma_i^2" + }, + { + "bbox": [ + 131, + 593, + 479, + 665 + ], + "type": "text", + "content": " denote the mixture weight, mean, and variance of the " + }, + { + "bbox": [ + 131, + 593, + 479, + 665 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 131, + 593, + 479, + 665 + ], + "type": "text", + "content": " Gaussian distribution within the posterior distribution. Here, " + }, + { + "bbox": [ + 131, + 593, + 479, + 665 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 593, + 479, + 665 + ], + "type": "text", + "content": " is the total number of Gaussians. For brevity we remove their conditioning on the posterior " + }, + { + "bbox": [ + 131, + 593, + 479, + 665 + ], + "type": "inline_equation", + "content": "q_{\\phi}(z_{t-1})" + }, + { + "bbox": [ + 131, + 593, + 479, + 665 + ], + "type": "text", + "content": " and sampled latent " + }, + { + "bbox": [ + 131, + 593, + 479, + 665 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 131, + 593, + 479, + 665 + ], + "type": "text", + "content": ". We sample " + }, + { + "bbox": [ + 131, + 593, + 479, + 665 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 131, + 593, + 479, + 665 + ], + "type": "text", + "content": " from the mixture of Gaussians " + }, + { + "bbox": [ + 131, + 593, + 479, + 665 + ], + "type": "inline_equation", + "content": "q_{\\phi}^{\\prime}(z_t)" + }, + { + "bbox": [ + 131, + 593, + 479, + 665 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 593, + 479, + 665 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 131, + 593, + 479, + 665 + ], + "type": "text", + "content": " likely falls within one of the Gaussian modes. The configuration corresponding to the mode is reflected in the 3D scene rendered by NeRF." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 134, + 90, + 448, + 103 + ], + "type": "text", + "content": "CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 164, + 212, + 219 + ], + "blocks": [ + { + "bbox": [ + 138, + 164, + 212, + 219 + ], + "lines": [ + { + "bbox": [ + 138, + 164, + 212, + 219 + ], + "spans": [ + { + "bbox": [ + 138, + 164, + 212, + 219 + ], + "type": "image", + "image_path": "7379d9d391663f7e3ed896bd624d5d8ee05057575b6903fed5329e19f50b314a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 226, + 294, + 293 + ], + "lines": [ + { + "bbox": [ + 135, + 226, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 135, + 226, + 294, + 293 + ], + "type": "text", + "content": "Fig.4: Multi-scene CARLA datasets. Varying car configurations and scenes for the Multi-Scene Two Lane Merge dataset (left) and the Multi-Scene Approaching Intersection dataset (right)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 217, + 164, + 293, + 218 + ], + "blocks": [ + { + "bbox": [ + 217, + 164, + 293, + 218 + ], + "lines": [ + { + "bbox": [ + 217, + 164, + 293, + 218 + ], + "spans": [ + { + "bbox": [ + 217, + 164, + 293, + 218 + ], + "type": "image", + "image_path": "77088b311dcf40ce6bf05ce57903c749dd732695fd3c03cc0f7846dc544cbf3a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 226, + 478, + 293 + ], + "lines": [ + { + "bbox": [ + 301, + 226, + 478, + 293 + ], + "spans": [ + { + "bbox": [ + 301, + 226, + 478, + 293 + ], + "type": "text", + "content": "Fig.5: Blender dataset. Blender dataset with a blue cube and a potential red cylinder exhibiting probabilistic temporal movement. The possible occlusions from different camera angles demonstrate how movement needs to be modeled probabilistically." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 302, + 114, + 356, + 218 + ], + "blocks": [ + { + "bbox": [ + 302, + 114, + 356, + 218 + ], + "lines": [ + { + "bbox": [ + 302, + 114, + 356, + 218 + ], + "spans": [ + { + "bbox": [ + 302, + 114, + 356, + 218 + ], + "type": "image", + "image_path": "afbb542921bb5c6efec2691d08739e8079af0267e0e4a8378ee2eb6ae6b38538.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 362, + 114, + 413, + 217 + ], + "blocks": [ + { + "bbox": [ + 362, + 114, + 413, + 217 + ], + "lines": [ + { + "bbox": [ + 362, + 114, + 413, + 217 + ], + "spans": [ + { + "bbox": [ + 362, + 114, + 413, + 217 + ], + "type": "image", + "image_path": "1f7ccc21d1f36ed77b383958701e42df86a5a7fa5bf0c0f06c76d122c56bc8c2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 423, + 114, + 474, + 218 + ], + "blocks": [ + { + "bbox": [ + 423, + 114, + 474, + 218 + ], + "lines": [ + { + "bbox": [ + 423, + 114, + 474, + 218 + ], + "spans": [ + { + "bbox": [ + 423, + 114, + 474, + 218 + ], + "type": "image", + "image_path": "07cdc9ebb6f1228e46f4cba95d292c4f9ec9a600cbb7ad2d55a31179ce468385.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 315, + 479, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 479, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 479, + 338 + ], + "type": "text", + "content": "Loss: To optimize the MDN, we minimize a negative log-likelihood function, given by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 216, + 338, + 480, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 338, + 480, + 371 + ], + "spans": [ + { + "bbox": [ + 216, + 338, + 480, + 371 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {M D N} = - \\sum_ {j = 1} ^ {N} \\log \\left(\\sum_ {i = 1} ^ {K} \\pi_ {i} \\mathcal {N} \\left(y _ {j}; \\mu_ {i}, \\sigma_ {i} ^ {2}\\right)\\right) \\tag {6}", + "image_path": "01ce069e463bd2a1d3771e631b4e21db95e38920c335303bca232388acd5fee6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 376, + 480, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 480, + 399 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 376, + 480, + 399 + ], + "type": "inline_equation", + "content": "y_{i} \\sim q_{\\phi}(z_{t})" + }, + { + "bbox": [ + 130, + 376, + 480, + 399 + ], + "type": "text", + "content": " is sampled from the distribution of latent " + }, + { + "bbox": [ + 130, + 376, + 480, + 399 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 376, + 480, + 399 + ], + "type": "text", + "content": ", learned by the encoder, and " + }, + { + "bbox": [ + 130, + 376, + 480, + 399 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 376, + 480, + 399 + ], + "type": "text", + "content": " denotes the total number of samples." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 409, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 409, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 482, + 506 + ], + "type": "text", + "content": "Inference: We consider an unseen ego-centric image and retrieve its posterior " + }, + { + "bbox": [ + 130, + 409, + 482, + 506 + ], + "type": "inline_equation", + "content": "q_{\\phi}(z_t)" + }, + { + "bbox": [ + 130, + 409, + 482, + 506 + ], + "type": "text", + "content": " through the encoder. Next, we predict the possible future posterior distribution " + }, + { + "bbox": [ + 130, + 409, + 482, + 506 + ], + "type": "inline_equation", + "content": "q_{\\phi}'(z_{t + 1})" + }, + { + "bbox": [ + 130, + 409, + 482, + 506 + ], + "type": "text", + "content": ". From the predicted posterior, we sample a scene latent and perform localization. We achieve this via (a) density probing the NeRF or (b) segmenting the rendered novel views using off-the-shelf methods such as YOLO (37) (see Fig. 2). These allow us to retrieve a corresponding Gaussian distribution " + }, + { + "bbox": [ + 130, + 409, + 482, + 506 + ], + "type": "inline_equation", + "content": "q_{\\phi}(z_{t + 1})" + }, + { + "bbox": [ + 130, + 409, + 482, + 506 + ], + "type": "text", + "content": " in encoder latent space. This is auto-regressively fed back into the MDN to predict the next timestamp. See Fig. 6 for an overview of the pipeline." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 522, + 198, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 198, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 198, + 534 + ], + "type": "text", + "content": "4 Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Decision-making under perceptual uncertainty is a pervasive challenge faced in robotics and autonomous driving, as the real environment is mostly likely partially observable, making it a POMDP. In a partially observable driving scenario, accurate inference regarding the presence of potentially obscured agents is pivotal. We evaluate the effectiveness of CARFF on common driving situations with partial observability and added complexity. We implemented several scenarios in the CARLA driving simulator (9) (see Fig. 4). A single NVIDIA RTX 3090 GPU is used to train PC-VAE, NeRF, and the MDN. All models, trained sequentially, tend to converge within a combined time frame of 24 hours. A detailed experimental setup can be found in Appendix ??. We show that, given" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "partially observable 2D inputs, CARFF performs well in predicting latent distributions that represent complete 3D scenes. Using these predictions we design a CARFF-based controller for performing downstream planning tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 169, + 246, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 169, + 246, + 180 + ], + "spans": [ + { + "bbox": [ + 132, + 169, + 246, + 180 + ], + "type": "text", + "content": "4.1 Data Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 189, + 480, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 480, + 285 + ], + "type": "text", + "content": "We conduct experiments on (a) synthetic blender dataset for principle experiments to test the probabilistic modeling capacities in isolation of the vision encoder (it is visually as simple as possible, but requires the full predictive model proposed in CARFF) and (b) CARLA-based driving datasets for more complex driving scenarios (9). To deliver convincing results, we model these driving scenarios off of related works (29; 54; 55; 56) that concern planning for driving under difficult situations. We generate the datasets in 3D by programming an ego object and varying actor objects in different configurations." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 142, + 307, + 473, + 417 + ], + "blocks": [ + { + "bbox": [ + 142, + 307, + 473, + 417 + ], + "lines": [ + { + "bbox": [ + 142, + 307, + 473, + 417 + ], + "spans": [ + { + "bbox": [ + 142, + 307, + 473, + 417 + ], + "type": "image", + "image_path": "19902d354443b11142b0941a8bce70e38f2d59be8ee146263f795cf147edc0a9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 425, + 482, + 536 + ], + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 536 + ], + "type": "text", + "content": "Fig. 6: Auto-regressive inference in scene prediction. The input image at timesamp " + }, + { + "bbox": [ + 130, + 425, + 482, + 536 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 425, + 482, + 536 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 425, + 482, + 536 + ], + "type": "inline_equation", + "content": "I_{c}^{t}" + }, + { + "bbox": [ + 130, + 425, + 482, + 536 + ], + "type": "text", + "content": ", is encoded using the pre-trained encoder from PC-VAE. The corresponding latent distribution is fed into the Mixture Density Network, which predicts a mixture of Gaussians. Each of the " + }, + { + "bbox": [ + 130, + 425, + 482, + 536 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 425, + 482, + 536 + ], + "type": "text", + "content": " Gaussians is a latent distribution that may correspond to different beliefs at the next timestamp. The mixture of Gaussians is sampled repeatedly for the predicted latent beliefs, visualized as " + }, + { + "bbox": [ + 130, + 425, + 482, + 536 + ], + "type": "inline_equation", + "content": "I_{c',scni}^{t+1}" + }, + { + "bbox": [ + 130, + 425, + 482, + 536 + ], + "type": "text", + "content": ", representing potentially the " + }, + { + "bbox": [ + 130, + 425, + 482, + 536 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 425, + 482, + 536 + ], + "type": "text", + "content": "th possible outcome. This is used to condition the NeRF to generate 3D views of the scene. To accomplish autoregressive predictions, we probe the NeRF for the location of the car and feed this information back to the pre-trained encoder to predict the scene at the next timestamp." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 480, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 480, + 618 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 480, + 618 + ], + "type": "text", + "content": "Blender synthetic dataset: This comprises of a stationary blue cube (ego) accompanied by a red cylinder (actor) that may or may not be present (see Fig. 5). If the actor is present, it exhibits lateral movement as depicted in Fig. 5. This simplistic setting provides an interpretable framework to evaluate our model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "CARLA dataset: Each dataset is simulated for " + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": " timestamps and uses " + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "C = 100" + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": " predefined camera poses to capture images of the environment under full observation, partial observation, and no visibility. These datasets are modeled after" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 447, + 103 + ], + "type": "text", + "content": "CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "common driving scenarios involving state uncertainty that have been proposed in related works such as Active Visual Planning (29)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 333 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 130, + 140, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 187 + ], + "type": "text", + "content": "a) Single-Scene Approaching Intersection: The ego vehicle is positioned at a T-intersection. An actor vehicle traverses the crossing along an evenly spaced, predefined trajectory. We simulate this for " + }, + { + "bbox": [ + 130, + 140, + 482, + 187 + ], + "type": "inline_equation", + "content": "N = 10" + }, + { + "bbox": [ + 130, + 140, + 482, + 187 + ], + "type": "text", + "content": " timestamps. We mainly use this dataset to predict the evolution of timestamps under full observation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 189, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 482, + 260 + ], + "type": "text", + "content": "b) Multi-Scene Approaching Intersection: We extend the previous scenario to a more complicated setting with state uncertainty, by making the existence of the actor vehicle probabilistic. A similar intersection crossing is simulated for " + }, + { + "bbox": [ + 130, + 189, + 482, + 260 + ], + "type": "inline_equation", + "content": "N = 3" + }, + { + "bbox": [ + 130, + 189, + 482, + 260 + ], + "type": "text", + "content": " timestamps for both possibilities. The ego vehicle's view of the actor may be occluded as it approaches the T-intersection over the " + }, + { + "bbox": [ + 130, + 189, + 482, + 260 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 189, + 482, + 260 + ], + "type": "text", + "content": " timestamps. The ego vehicle either moves forward or halts at the junction (see Fig. 4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 261, + 482, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 333 + ], + "type": "text", + "content": "c) Multi-Scene Multi-actor Two Lane Merge: To add more environment dynamics uncertainty, we consider a multi-actor setting at an intersection of two merging lanes. We simulate the scenario at an intersection with partial occlusions, with the second approaching actor having variable speed. Here the ego vehicle can either merge into the left lane before the second actor or after all the actors pass, (see Fig. 4). Each branch is simulated for " + }, + { + "bbox": [ + 130, + 261, + 482, + 333 + ], + "type": "inline_equation", + "content": "N = 3" + }, + { + "bbox": [ + 130, + 261, + 482, + 333 + ], + "type": "text", + "content": " timestamps." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 351, + 259, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 351, + 259, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 351, + 259, + 363 + ], + "type": "text", + "content": "4.2 CARFF Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 132, + 392, + 482, + 480 + ], + "blocks": [ + { + "bbox": [ + 132, + 392, + 482, + 480 + ], + "lines": [ + { + "bbox": [ + 132, + 392, + 482, + 480 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 482, + 480 + ], + "type": "table", + "html": "
Method3DComplex ScenariosState UncertaintyDynamics UncertaintyPredictionPlanningCode Released
CARFF✓✓✓✓✓✓✓
(20)✓✓
(21)✓✓✓✓
(1)✓✓✓✓
(29)✓✓✓✓✓
", + "image_path": "f1048e02c4f7ced63579f0768d4518624ef27639ce5017b236a3fed76387d037.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 488, + 482, + 555 + ], + "lines": [ + { + "bbox": [ + 130, + 488, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 488, + 482, + 555 + ], + "type": "text", + "content": "Table 1: Qualitative comparison of CARFF to related works. CARFF accomplishes all highlighted objectives as opposed to NeRF-VAE (20), NeRF for Visuomotor Control (21), Vision-only NeRF Navigation (1), and AVP (29). We compare whether methods reason in a 3D environment and perform novel view synthesis; work on complex scenarios; predict probabilistically under state and dynamics uncertainty; forecast into the future; and use model predictions for decision-making." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "A desirable behavior from our model is that it should predict a complete set of possible scenes consistent with the given ego-centric image, which could be partially observable. This is crucial for autonomous driving in unpredictable environments as it ensures strategic decision-making based on potential hazards. To achieve this we require a rich PC-VAE latent space, high-quality novel view synthesis, and auto-regressive probabilistic predictions of latents at future timestamps. We evaluate CARFF on a simple synthetic blender-based dataset" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "and each CARLA-based dataset. Additionally, we extend our model application to a hand-manipulation dataset in Appendix ??" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 212 + ], + "type": "text", + "content": "Comparisons with related work: We attempt to compare CARFF to existing approaches. NeRF-VAE has the most comparable objective, but during our experiments, it collapses to black using CARLA datasets. We make further qualitative comparisons to other most similar methods in Tab. 1, but none aligns with ours enough to make any possible quantitative comparisons." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "type": "text", + "content": "Evaluation on blender dataset: In Fig. 5, for both Scene 1a and 1b, our model correctly forecasts the lateral movement of the cylinder to be in either position approximately " + }, + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "type": "text", + "content": " of the time, considering a left viewing angle. In Scene 2, with the absence of the red cylinder in the input camera angle, the model predicts the potential existence of the red cylinder approximately " + }, + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "type": "text", + "content": " of the time, and predicts lateral movements with roughly equal probability. This validates PC-VAE's ability to predict and infer occlusions in the latent space, aligning with human intuitions. These intuitions, shown in the Blender dataset's simple scenes, can transfer to driving scenarios in our CARLA datasets." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 176, + 361, + 439, + 410 + ], + "blocks": [ + { + "bbox": [ + 176, + 361, + 439, + 410 + ], + "lines": [ + { + "bbox": [ + 176, + 361, + 439, + 410 + ], + "spans": [ + { + "bbox": [ + 176, + 361, + 439, + 410 + ], + "type": "image", + "image_path": "688dd54c0d8771ce3865e20ca2164cf80a4e372f3027a55385a62270d4e89b8f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 418, + 482, + 462 + ], + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 462 + ], + "type": "text", + "content": "Fig. 7: PC-VAE reconstructions. The encoder input, " + }, + { + "bbox": [ + 130, + 418, + 482, + 462 + ], + "type": "inline_equation", + "content": "I_{c}^{t}" + }, + { + "bbox": [ + 130, + 418, + 482, + 462 + ], + "type": "text", + "content": ", among the other ground truth images " + }, + { + "bbox": [ + 130, + 418, + 482, + 462 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 130, + 418, + 482, + 462 + ], + "type": "text", + "content": " viewed from camera pose " + }, + { + "bbox": [ + 130, + 418, + 482, + 462 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 418, + 482, + 462 + ], + "type": "text", + "content": " at different timestamps, is reconstructed across a new set of poses " + }, + { + "bbox": [ + 130, + 418, + 482, + 462 + ], + "type": "inline_equation", + "content": "c''" + }, + { + "bbox": [ + 130, + 418, + 482, + 462 + ], + "type": "text", + "content": " respecting timestamp " + }, + { + "bbox": [ + 130, + 418, + 482, + 462 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 418, + 482, + 462 + ], + "type": "text", + "content": ", generating " + }, + { + "bbox": [ + 130, + 418, + 482, + 462 + ], + "type": "inline_equation", + "content": "I_{c''}^{t}" + }, + { + "bbox": [ + 130, + 418, + 482, + 462 + ], + "type": "text", + "content": ". A complete grid is in Appendix ??" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "PC-VAE performance and ablations: We evaluate the performance of PC-VAE on CARLA datasets with multiple encoder architectures. We show that PC-VAE effectively reconstructs complex environments involving variable scenes, actor configurations, and environmental noise given potentially partially observable inputs (see Fig. 7). We calculated an average Peak Signal-to-Noise Ratio (PSNR) over the training data, as well as novel view encoder inputs. To evaluate the quality of the latent space generated by the encoder, we utilize t-SNE (24) plots to visualize the distribution of latent samples for each image in a given dataset (see Appendix ??). We introduce a Support Vector Machine (SVM) (14) based metric to measure the visualized clustering quantitatively, where a higher value indicates better clustering based on timestamps. Most latent scene samples are separable by timestamps, which indicates that the latents are view-invariant. Samples that are misclassified or lie on the boundary usually represent partially or fully occluded regions. This is desirable for forecasting, as it enables" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 90, + 447, + 102 + ], + "type": "text", + "content": "CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 119, + 297, + 290 + ], + "blocks": [ + { + "bbox": [ + 137, + 119, + 297, + 290 + ], + "lines": [ + { + "bbox": [ + 137, + 119, + 297, + 290 + ], + "spans": [ + { + "bbox": [ + 137, + 119, + 297, + 290 + ], + "type": "table", + "html": "
Ground Truth Prediction PairAvg. PSNR (Scene 1)Avg. PSNR (Scene 2)
Single-Scene Intersection
Matching Pairs29.06N.A
Un-matching P.24.01N.A
Multi-Scene Intersection
Matching Pairs28.0028.26
Un-matching P.23.2724.56
Multi-Scene Two Lane Merge
Matching Pairs28.1428.17
Un-matching P.22.7423.32
", + "image_path": "d982bedbd0b3d0ee8c0515e595e4a074fdf7bfc74a5af62c539fc9058593886a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 314, + 114, + 479, + 289 + ], + "blocks": [ + { + "bbox": [ + 130, + 297, + 303, + 396 + ], + "lines": [ + { + "bbox": [ + 130, + 297, + 303, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 303, + 396 + ], + "type": "text", + "content": "Table 2: Averaged PSNR for fully observable 3D predictions. CARFF correctly predicts scene evolution across all timestamps for each dataset. The average PSNR is high for predictions " + }, + { + "bbox": [ + 130, + 297, + 303, + 396 + ], + "type": "inline_equation", + "content": "\\hat{I_{t_i}}" + }, + { + "bbox": [ + 130, + 297, + 303, + 396 + ], + "type": "text", + "content": " and matching ground truths, " + }, + { + "bbox": [ + 130, + 297, + 303, + 396 + ], + "type": "inline_equation", + "content": "I_{t_i}" + }, + { + "bbox": [ + 130, + 297, + 303, + 396 + ], + "type": "text", + "content": ". PSNR values for incorrect correspondences, " + }, + { + "bbox": [ + 130, + 297, + 303, + 396 + ], + "type": "inline_equation", + "content": "\\hat{I_{t_i}}, I_{t_j}" + }, + { + "bbox": [ + 130, + 297, + 303, + 396 + ], + "type": "text", + "content": ", is a result of matching surroundings. See complete table in Appendix ??" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 114, + 479, + 289 + ], + "lines": [ + { + "bbox": [ + 314, + 114, + 479, + 289 + ], + "spans": [ + { + "bbox": [ + 314, + 114, + 479, + 289 + ], + "type": "table", + "html": "
Multi-Scene Intersection
Controller TypeActorNo Actor
Underconfident30/300/30
Overconfident0/3030/30
CARFF (n=2)17/3030/30
CARFF (n=10)30/3030/30
CARFF (n=35)30/3019/30
Multi-Scene Two Lane Merge
Controller TypeFastSlow
Underconfident30/300/30
Overconfident0/3030/30
CARFF (n=2)21/3030/30
CARFF (n=10)30/3030/30
CARFF (n=35)30/3022/30
", + "image_path": "4183dfeb7507b91a6da4e5c1276baa0819159b41dc44868512366ef263d09329.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 309, + 297, + 482, + 397 + ], + "lines": [ + { + "bbox": [ + 309, + 297, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 309, + 297, + 482, + 397 + ], + "type": "text", + "content": "Table 3: Planning in 3D with controllers with varying sampling numbers " + }, + { + "bbox": [ + 309, + 297, + 482, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 309, + 297, + 482, + 397 + ], + "type": "text", + "content": ". CARFF-based controllers outperform baselines in success rate over 30 trials. For " + }, + { + "bbox": [ + 309, + 297, + 482, + 397 + ], + "type": "inline_equation", + "content": "n = 10" + }, + { + "bbox": [ + 309, + 297, + 482, + 397 + ], + "type": "text", + "content": ", the CARFF-based controller consistently chooses the optimal action in potential collision scenarios. To maintain consistency, we use one single image input across 30 trials." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 422, + 482, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 482 + ], + "type": "text", + "content": "us to model probabilistic behavior over these samples. In this process, balancing KL divergence weight scheduling maintains the quality of the PC-VAE's latent space and reconstructions (see Appendix ??). Additionally, we substantiate the benefits of our PC-VAE encoder architecture through our ablations (see Appendix ??)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 593 + ], + "type": "text", + "content": "3D novel view synthesis: Given an unseen ego-centric view with potentially partial observations, our method maintains all possible current state beliefs in 3D, and faithfully reconstructs novel views from arbitrary camera angles for each belief. Fig. 2 illustrates one of the possible 3D beliefs that CARFF holds. This demonstrates our method's ability to generate 3D beliefs that could be used for novel view synthesis in a view-consistent manner. Our model's ability to achieve accurate and complete 3D environmental understanding is important for applications like prediction-based planning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Inference under full and partial observations: Under full observation, we use MDN to predict the subsequent car positions in all three datasets. PSNR values are calculated based on bird-eye view NeRF renderings and ground truth bird-eye view images of the scene across different timestamps. In Tab. 2 we report the PSNR values for rendered images over the predicted posterior with the ground" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": "truth images at each timestamp. We also evaluate the efficacy of our prediction model using the accuracy curve given in Fig. 8. This represents CARFF's ability to generate stable beliefs, without producing incorrect predictions, based on actor(s) localization results. For each number of samples between " + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "inline_equation", + "content": "n = 0" + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "inline_equation", + "content": "n = 50" + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": ", we choose a random subset of 3 fully observable ego images and take an average of the accuracies. In scenarios with partial observable ego-centric images where several plausible scenarios exist, we utilize recall instead of accuracy using a similar setup. This lets us evaluate the encoder's ability to avoid false negative predictions of potential danger." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 480, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 480, + 297 + ], + "type": "text", + "content": "Fig. 8 shows that our model achieves high accuracy and recall in both datasets, demonstrating the ability to model state uncertainty (Approaching Intersection) and dynamic uncertainty (Two Lane Merge). The results indicate CARFF's resilience against randomness in resampling, and completeness in probabilistic modeling of the belief space. Given these observations, we now build a reliable controller to plan and navigate through complex scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 312, + 206, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 206, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 206, + 324 + ], + "type": "text", + "content": "4.3 Planning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 331, + 479, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 479, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 479, + 365 + ], + "type": "text", + "content": "In all our experiments, the ego vehicle must make decisions to advance under certain observability. The scenarios are designed such that the ego views contain partial occlusion and the state of the actor(s) is uncertain in some scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 366, + 480, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 480, + 449 + ], + "type": "text", + "content": "In order to facilitate decision-making using CARFF, we design a controller that takes ego-centric input images and outputs an action. Decisions are made incorporating sample consistency from the mixture density network. For instance, the controller infers occlusion and promotes the ego car to pause when scenes alternate between actor presence and absence in the samples. We use the two multi-scene datasets to assess the performance of the CARFF-based controller as they contain actors with potentially unknown behaviors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 450, + 481, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 481, + 581 + ], + "type": "text", + "content": "To design an effective controller, we need to find a balance between accuracy and recall (see Fig. 8). A lowered accuracy from excessive sampling means unwanted randomness in the predicted state. However, taking insufficient samples would generate low recall i.e., not recovering all plausible states. This would lead to incorrect predictions as we would be unable to account for the plausible uncertainty present in the environment. To achieve optimal balance, we designed an open-loop planning controller using a sampling strategy that generates " + }, + { + "bbox": [ + 130, + 450, + 481, + 581 + ], + "type": "inline_equation", + "content": "n = 2, 10, 35" + }, + { + "bbox": [ + 130, + 450, + 481, + 581 + ], + "type": "text", + "content": " samples. The hyperparameter " + }, + { + "bbox": [ + 130, + 450, + 481, + 581 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 450, + 481, + 581 + ], + "type": "text", + "content": " is tuned per scene for peak performance but is expected to remain relatively stable across scenes. We demonstrate that " + }, + { + "bbox": [ + 130, + 450, + 481, + 581 + ], + "type": "inline_equation", + "content": "n = 10" + }, + { + "bbox": [ + 130, + 450, + 481, + 581 + ], + "type": "text", + "content": " performs well consistently in varying CARLA scenarios (Fig. 8) and do not anticipate this being very different for other experiments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "For sampling values that lie on the borders of the accuracy and recall margin, for example, " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "n = 2" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": " and 35, we see that the CARFF-based controller obtains lower success rates, whereas " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "n = 10" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": " produces the best result. For actor exists and fast-actor scenes in Tab. 3, we consider occluded ego-centric inputs to test the controller's ability to avoid collisions. For no-actor and slow-actor scenes, we consider state observability and test the controllers' ability to recognize the optimal action to advance. Across the two datasets, the overconfident controller" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 90, + 447, + 102 + ], + "type": "text", + "content": "CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 214, + 112, + 306, + 190 + ], + "blocks": [ + { + "bbox": [ + 214, + 112, + 306, + 190 + ], + "lines": [ + { + "bbox": [ + 214, + 112, + 306, + 190 + ], + "spans": [ + { + "bbox": [ + 214, + 112, + 306, + 190 + ], + "type": "image", + "image_path": "7e4fbc87d5e4654bd04e8bafc34f14bdc37e5f360d7db39e2aeb0852c4c0f8f8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 193, + 293, + 204 + ], + "lines": [ + { + "bbox": [ + 227, + 193, + 293, + 204 + ], + "spans": [ + { + "bbox": [ + 227, + 193, + 293, + 204 + ], + "type": "text", + "content": "(a) Intersection" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 112, + 399, + 190 + ], + "blocks": [ + { + "bbox": [ + 307, + 112, + 399, + 190 + ], + "lines": [ + { + "bbox": [ + 307, + 112, + 399, + 190 + ], + "spans": [ + { + "bbox": [ + 307, + 112, + 399, + 190 + ], + "type": "image", + "image_path": "9d432054d0fdcabc8b1e68591090396a8dfca8f51b511df57b933c4cde202da2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 193, + 388, + 205 + ], + "lines": [ + { + "bbox": [ + 320, + 193, + 388, + 205 + ], + "spans": [ + { + "bbox": [ + 320, + 193, + 388, + 205 + ], + "type": "text", + "content": "(b) Lane Merge" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 213, + 482, + 291 + ], + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 291 + ], + "type": "text", + "content": "Fig. 8: Multi-Scene dataset accuracy and recall curves from predicted beliefs. We test our framework across " + }, + { + "bbox": [ + 130, + 213, + 482, + 291 + ], + "type": "inline_equation", + "content": "n = 1" + }, + { + "bbox": [ + 130, + 213, + 482, + 291 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 213, + 482, + 291 + ], + "type": "inline_equation", + "content": "n = 50" + }, + { + "bbox": [ + 130, + 213, + 482, + 291 + ], + "type": "text", + "content": " samples from MDN's predicted latent distributions from ego-centric image input. Across the number of samples " + }, + { + "bbox": [ + 130, + 213, + 482, + 291 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 213, + 482, + 291 + ], + "type": "text", + "content": ", we achieve an ideal margin of belief state coverage generated under partial observation (recall), and the proportion of correct beliefs sampled under full observation (accuracy). As we significantly increase the number of samples, the accuracy starts to decrease due to randomness in latent distribution resampling." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 314, + 482, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 398 + ], + "type": "text", + "content": "will inevitably experience collisions in case of a truck approaching, since it does not cautiously account for occlusions. On the other hand, an overly cautious approach results in stasis, inhibiting the controller's ability to advance in the scene. This nuanced decision-making using CARFF-based controller is especially crucial in driving scenarios, as it enhances safety and efficiency by adapting to complex and unpredictable road environments, thereby fostering a more reliable and human-like response in autonomous vehicles." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 415, + 217, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 415, + 217, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 415, + 217, + 427 + ], + "type": "text", + "content": "5 Discussion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 439, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 482, + 547 + ], + "type": "text", + "content": "Limitations: Like other NeRF-based methods, CARFF currently relies on posed images of specific scenes such as road intersections, limiting its direct applicability to unseen environments. However, we anticipate enhanced generalizability with the increasing deployment of cameras around populated areas, such as traffic cameras at intersections. Additionally, handling very complex dynamics with an extremely large number of actors still poses a challenge for our method, requiring per-scene optimization to balance comprehensive dynamics modeling against accuracy. Potentially stronger models in the near future may offer a promising avenue for further enhancements in this regard." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": "Conclusion: We present CARFF, a novel method for probabilistic 3D scene forecasting from partial observations. By employing a Pose-Conditional VAE, a NeRF conditioned on the learned posterior, and a mixture density network that forecasts future scenes, we effectively model, predict, and plan in complex environments with state and dynamics uncertainty in a POMDP. We further demonstrate the capabilities of our method in simulated autonomous driving scenarios. Overall, CARFF offers an intuitive framework to perceiving, forecasting, and acting under uncertainty that could prove invaluable for vision-based algorithms in unstructured environments." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 261, + 112, + 353, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 112, + 353, + 129 + ], + "spans": [ + { + "bbox": [ + 261, + 112, + 353, + 129 + ], + "type": "text", + "content": "Bibliography" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 133, + 167, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 167, + 481, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 167, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 138, + 167, + 481, + 204 + ], + "type": "text", + "content": "[1] Adamkiewicz, M., Chen, T., Caccavale, A., Gardner, R., Culbertson, P., Bohg, J., Schwager, M.: Vision-only robot navigation in a neural radiance world. IEEE Robotics and Automation Letters 7(2), 4606-4613 (2022)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 138, + 205, + 481, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 481, + 240 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 481, + 240 + ], + "type": "text", + "content": "[2] Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Int. Conf. Comput. Vis. pp. 5855–5864 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 241, + 481, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 241, + 481, + 265 + ], + "spans": [ + { + "bbox": [ + 138, + 241, + 481, + 265 + ], + "type": "text", + "content": "[3] Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 130-141 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 266, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 481, + 300 + ], + "type": "text", + "content": "[4] Cao, J., Wang, X., Darrell, T., Yu, F.: Instance-aware predictive navigation in multi-agent environments. In: IEEE Int. Conf. on Robotics and Automation. pp. 5096-5102. IEEE (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 301, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 301, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 138, + 301, + 481, + 325 + ], + "type": "text", + "content": "[5] Charatan, D., Li, S., Tagliasacchi, A., Sitzmann, V.: pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 326, + 481, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 481, + 361 + ], + "type": "text", + "content": "[6] Codevilla, F., Santana, E., López, A.M., Gaidon, A.: Exploring the limitations of behavior cloning for autonomous driving. In: Int. Conf. Comput. Vis. pp. 9329-9338 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 362, + 481, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 362, + 481, + 398 + ], + "spans": [ + { + "bbox": [ + 138, + 362, + 481, + 398 + ], + "type": "text", + "content": "[7] Deng, K., Liu, A., Zhu, J.Y., Ramanan, D.: Depth-supervised nerf: Fewer views and faster training for free. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12882-12891 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 399, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 481, + 446 + ], + "type": "text", + "content": "[8] Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Int. Conf. Learn. Represent. (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 447, + 481, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 481, + 471 + ], + "type": "text", + "content": "[9] Dosovitskiy, A., Ros, G., Codevilla, F., Lopez, A., Koltun, V.: CARLA: An open urban driving simulator. In: Conf. on Robol Learning. pp. 1-16 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 472, + 481, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 472, + 481, + 507 + ], + "spans": [ + { + "bbox": [ + 133, + 472, + 481, + 507 + ], + "type": "text", + "content": "[10] Driess, D., Huang, Z., Li, Y., Tedrake, R., Toussaint, M.: Learning multi-object dynamics with compositional neural radiance fields. arXiv preprint arXiv:2202.11855 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 508, + 481, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 508, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 133, + 508, + 481, + 532 + ], + "type": "text", + "content": "[11] Fei, B., Xu, J., Zhang, R., Zhou, Q., Yang, W., He, Y.: 3d gaussian as a new vision era: A survey (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 533, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 533, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 133, + 533, + 481, + 568 + ], + "type": "text", + "content": "[12] Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5501–5510 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 569, + 481, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 569, + 481, + 592 + ], + "spans": [ + { + "bbox": [ + 133, + 569, + 481, + 592 + ], + "type": "text", + "content": "[13] Hausknecht, M., Stone, P.: Deep recurrent q-learning for partially observable mdps. In: AAAI (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 593, + 481, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 593, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 133, + 593, + 481, + 628 + ], + "type": "text", + "content": "[14] Hearst, M.A., Dumais, S.T., Osuna, E., Platt, J., Scholkopf, B.: Support vector machines. IEEE Intelligent Systems and their applications 13(4), 18-28 (1998)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 629, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 629, + 481, + 665 + ], + "type": "text", + "content": "[15] Ichnowski, J., Avigal, Y., Kerr, J., Goldberg, K.: Dex-nerf: Using a neural radiance field to grasp transparent objects. arXiv preprint arXiv:2110.14217 (2021)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 133, + 115, + 480, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 480, + 152 + ], + "type": "text", + "content": "[16] Ivanovic, B., Elhafsi, A., Rosman, G., Gaidon, A., Pavone, M.: Mats: An interpretable trajectory forecasting representation for planning and control. In: Conf. on Robol Learning (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 152, + 481, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 152, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 152, + 481, + 213 + ], + "type": "text", + "content": "[17] Kaelbling, L.P., Littman, M.L., Cassandra, A.R.: Planning and acting in partially observable stochastic domains. Artificial Intelligence 101(1), 99-134 (1998). https://doi.org/https://doi.org/10.1016/S0004-3702(98)00023-X, https://www.sciencedirect.com/science/article/pii/S000437029800023X" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 213, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 481, + 262 + ], + "type": "text", + "content": "[18] Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 262, + 481, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 481, + 298 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 481, + 298 + ], + "type": "text", + "content": "[19] Kerr, J., Fu, L., Huang, H., Avigal, Y., Tancik, M., Ichnowski, J., Kanazawa, A., Goldberg, K.: Evo-nerf: Evolving nerf for sequential robot grasping of transparent objects. In: Conf. on Robol Learning (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 299, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 299, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 299, + 481, + 335 + ], + "type": "text", + "content": "[20] Kosiorek, A.R., Strathmann, H., Zoran, D., Moreno, P., Schneider, R., Mokrá, S., Rezende, D.J.: NeRF-VAE: A geometry aware 3d scene generative model. In: ICML. pp. 5742–5752 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 336, + 481, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 371 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 371 + ], + "type": "text", + "content": "[21] Li, Y., Li, S., Sitzmann, V., Agrawal, P., Torralba, A.: 3d neural scene representations for visuomotor control. In: Conf. on Robol Learning. pp. 112-123 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 372, + 481, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 372, + 481, + 421 + ], + "spans": [ + { + "bbox": [ + 132, + 372, + 481, + 421 + ], + "type": "text", + "content": "[22] Liu, J.W., Cao, Y.P., Mao, W., Zhang, W., Zhang, D.J., Keppo, J., Shan, Y., Qie, X., Shou, M.Z.: Devrf: Fast deformable voxel radiance fields for dynamic scenes. In: Adv. Neural Inform. Process. Syst. vol. 35, pp. 36762-36775 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 422, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 422, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 422, + 481, + 445 + ], + "type": "text", + "content": "[23] Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 446, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 482 + ], + "type": "text", + "content": "[24] van der Maaten, L., Hinton, G.: Visualizing data using t-SNE. Journal of Machine Learning Research 9, 2579-2605 (2008), http://www.jmlr.org/papers/v9/vandermaaten08a.html" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 483, + 481, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 481, + 518 + ], + "type": "text", + "content": "[25] Marza, P., Matignon, L., Simonin, O., Wolf, C.: Multi-object navigation with dynamically learned neural implicit representations. In: Int. Conf. Comput. Vis. pp. 11004-11015 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 520, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 481, + 555 + ], + "type": "text", + "content": "[26] McAllister, R., Rasmussen, C.E.: Data-efficient reinforcement learning in continuous state-action gaussian-pomdpds. In: Adv. Neural Inform. Process. Syst. (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 556, + 481, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 481, + 592 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 481, + 592 + ], + "type": "text", + "content": "[27] Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: Eur. Conf. Comput. Vis. (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 593, + 481, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 593, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 593, + 481, + 628 + ], + "type": "text", + "content": "[28] Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 1-15 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 629, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 481, + 666 + ], + "type": "text", + "content": "[29] Packer, C., Rhinehart, N., McAllister, R.T., Wright, M.A., Wang, X., He, J., Levine, S., Gonzalez, J.E.: Is anyone there? learning a planner contingent on perceptual uncertainty. In: Conf. on Robol Learning (2022)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 133, + 115, + 480, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 480, + 164 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 480, + 164 + ], + "type": "text", + "content": "[30] Packer, C., Rhinehart, N., McAllister, R.T., Wright, M.A., Wang, X., He, J., Levine, S., Gonzalez, J.E.: Is anyone there? learning a planner contingent on perceptual uncertainty. In: Liu, K., Kulic, D., Ichnowski, J. (eds.) Conf. on Robol Learning. pp. 1607-1617 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 165, + 480, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 165, + 480, + 189 + ], + "spans": [ + { + "bbox": [ + 133, + 165, + 480, + 189 + ], + "type": "text", + "content": "[31] Pan, X., You, Y., Wang, Z., Lu, C.: Virtual to real reinforcement learning for autonomous driving. arXiv preprint arXiv:1704.03952 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 189, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 189, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 189, + 480, + 214 + ], + "type": "text", + "content": "[32] Papadimitriou, C.H., Tsitsiklis, J.N.: The complexity of markov decision processes. Mathematics of operations research 12(3), 441-450 (1987)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 251 + ], + "type": "text", + "content": "[33] Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: Int. Conf. Comput. Vis. pp. 5865–5874 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 480, + 299 + ], + "type": "text", + "content": "[34] Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph. (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 300, + 480, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 324 + ], + "type": "text", + "content": "[35] Pineau, J., Gordon, G., Thrun, S., et al.: Point-based value iteration: An anytime algorithm for pomdpps. In: IJCAI. vol. 3, pp. 1025-1032 (2003)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 324, + 480, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 324, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 324, + 480, + 361 + ], + "type": "text", + "content": "[36] Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-NeRF: Neural Radiance Fields for Dynamic Scenes. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 361, + 480, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 480, + 397 + ], + "type": "text", + "content": "[37] Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 397, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 446 + ], + "type": "text", + "content": "[38] Rhinehart, N., He, J., Packer, C., Wright, M.A., McAllister, R., Gonzalez, J.E., Levine, S.: Contingencies from observations: Tractable contingency planning with learned behavior models. In: IEEE Int. Conf. on Robotics and Automation. pp. 13663-13669 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 446, + 480, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 483 + ], + "type": "text", + "content": "[39] Roessle, B., Barron, J.T., Mildenhall, B., Srinivasan, P.P., Nießner, M.: Dense depth priors for neural radiance fields from sparse input views. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12892–12901 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 483, + 480, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 532 + ], + "type": "text", + "content": "[40] Shao, R., Zheng, Z., Tu, H., Liu, B., Zhang, H., Liu, Y.: Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 16632-16642 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 532, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 480, + 568 + ], + "type": "text", + "content": "[41] Sun, C., Sun, M., Chen, H.T.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5459-5469 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 568, + 480, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 616 + ], + "type": "text", + "content": "[42] Tancik, M., Casser, V., Yan, X., Pradhan, S., Mildenhall, B., Srinivasan, P.P., Barron, J.T., Kretzschmar, H.: Block-nerf: Scalable large scene neural view synthesis. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8248–8258 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 616, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 616, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 616, + 480, + 666 + ], + "type": "text", + "content": "[43] Tancik, M., Srinivasan, P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. In: Adv. Neural Inform. Process. Syst. pp. 7537–7547 (2020)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 90, + 447, + 102 + ], + "type": "text", + "content": "CARFF: Conditional Auto-encoded Radiance Field for 3D Scene Forecasting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 133, + 115, + 480, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 480, + 152 + ], + "type": "text", + "content": "[44] Tas, O.S., Stiller, C.: Limited visibility and uncertainty aware motion planning for automated driving. In: IEEE Intelligent Vehicles Symposium (IV) (jun 2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 152, + 480, + 199 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 152, + 480, + 199 + ], + "spans": [ + { + "bbox": [ + 133, + 152, + 480, + 199 + ], + "type": "text", + "content": "[45] Tewari, A., Thies, J., Mildenhall, B., Srinivasan, P., Tretschk, E., Yifan, W., Lassner, C., Sitzmann, V., Martin-Brualla, R., Lombardi, S., Simon, T., Theobalt, C., Nießner, M., Barron, J.T., Wetzstein, G., Zollhöfer, M., Golyanik, V.: Advances in Neural Rendering. Comput. Graph. Forum (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 200, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 200, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 200, + 480, + 236 + ], + "type": "text", + "content": "[46] Toromanoff, M., Wirbel, E., Moutarde, F.: End-to-end model-free reinforcement learning for urban driving using implicit affordances. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7153-7162 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 236, + 480, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 271 + ], + "type": "text", + "content": "[47] Tretschk, E., Golyanik, V., Zollhoefer, M., Bozic, A., Lassner, C., Theobalt, C.: Scenerflow: Time-consistent reconstruction of general dynamic scenes. In: International Conference on 3D Vision (3DV) (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 272, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 272, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 272, + 480, + 308 + ], + "type": "text", + "content": "[48] Truong, P., Rakotosaona, M.J., Manhardt, F., Tombari, F.: Sparf: Neural radiance fields from sparse and noisy poses. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4190-4200 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 308, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 480, + 342 + ], + "type": "text", + "content": "[49] Turki, H., Ramanan, D., Satyanarayanan, M.: Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12922–12931 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 343, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 390 + ], + "type": "text", + "content": "[50] Wang, Q., Wang, Z., Genova, K., Srinivasan, P.P., Zhou, H., Barron, J.T., Martin-Brualla, R., Snavely, N., Funkhouser, T.: Ibrnet: Learning multi-view image-based rendering. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4690-4699 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 391, + 480, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 426 + ], + "type": "text", + "content": "[51] Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5438-5448 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 426, + 480, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 426, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 480, + 474 + ], + "type": "text", + "content": "[52] Yen-Chen, L., Florence, P., Barron, J.T., Rodriguez, A., Isola, P., Lin, T.Y.: inerf: Inverting neural radiance fields for pose estimation. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1323-1330. IEEE (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 475, + 480, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 480, + 498 + ], + "type": "text", + "content": "[53] Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 498, + 480, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 556 + ], + "type": "text", + "content": "[54] Yu, M.Y., Vasudevan, R., Johnson-Roberson, M.: Occlusion-aware risk assessment for autonomous driving in urban environments. IEEE Robotics and Automation Letters 4(2), 2235-2241 (Apr 2019). https://doi.org/10.1109/lra.2019.2900453, http://dx.doi.org/10.1109/LRA.2019.2900453" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 558, + 480, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 480, + 616 + ], + "type": "text", + "content": "[55] Zhang, C., Steinhauser, F., Hinz, G., Knoll, A.: Improved occlusion scenario coverage with a pomdp-based behavior planner for autonomous urban driving. In: 2021 IEEE International Intelligent Transportation Systems Conference (ITSC). pp. 593-600 (2021). https://doi.org/10.1109/ITSC48978.2021.9564424" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 617, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 480, + 665 + ], + "type": "text", + "content": "[56] Zhang, C., Steinhauser, F., Hinz, G., Knoll, A.: Occlusion-aware planning for autonomous driving with vehicle-to-everything communication. IEEE Transactions on Intelligent Vehicles 9(1), 1229-1242 (2024). https://doi.org/10.1109/TIV.2023.3308098" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CAT-SAM_ Conditional Tuning for Few-Shot Adaptation of Segment Anything Model/bd649bc1-acc8-4027-a33f-08ccf2414aa5_content_list.json b/2024/CAT-SAM_ Conditional Tuning for Few-Shot Adaptation of Segment Anything Model/bd649bc1-acc8-4027-a33f-08ccf2414aa5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f64318c4131e5324f34806ce0952224af200aae5 --- /dev/null +++ b/2024/CAT-SAM_ Conditional Tuning for Few-Shot Adaptation of Segment Anything Model/bd649bc1-acc8-4027-a33f-08ccf2414aa5_content_list.json @@ -0,0 +1,1630 @@ +[ + { + "type": "text", + "text": "CAT-SAM: Conditional Tuning for Few-Shot Adaptation of Segment Anything Model", + "text_level": 1, + "bbox": [ + 243, + 140, + 756, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aoran Xiao $^{1*}$ , Weihao Xuan $^{2,3*}$ , Heli Qi $^{4}$ , Yun Xing $^{1}$ , Ruijie Ren $^{5}$ , Xiaojin Zhang $^{6}$ , Ling Shao $^{7}$ , and Shijian Lu $^{1\\dagger}$", + "bbox": [ + 232, + 210, + 771, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Nanyang Technological University, Singapore \n2 The University of Tokyo, Japan \n3 RIKEN AIP, Japan", + "bbox": [ + 341, + 253, + 658, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Nara Institute of Science and Technology, Japan \n $^{5}$ Waseda University, Japan", + "bbox": [ + 331, + 296, + 669, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "6 Zhejiang University of Technology, China", + "bbox": [ + 354, + 324, + 645, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "UCAS-Terminus AI Lab, University of Chinese Academy of Sciences, China", + "bbox": [ + 240, + 338, + 759, + 351 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The Segment Anything Model (SAM) has demonstrated remarkable zero-shot capability and flexible geometric prompting in general image segmentation. However, it often struggles in domains that are either sparsely represented or lie outside its training distribution, such as aerial, medical, and non-RGB images. Recent efforts have predominantly focused on adapting SAM to these domains using fully supervised methods, which necessitate large amounts of annotated training data and pose practical challenges in data collection. This paper presents CAT-SAM, a ConditionAI Tuning network that explores few-shot adaptation of SAM toward various challenging downstream domains in a data-efficient manner. The core design is a prompt bridge structure that enables decoder-conditioned joint tuning of the heavyweight image encoder and the lightweight mask decoder. The bridging maps the domain-specific features of the mask decoder to the image encoder, fostering synergic adaptation of both components with mutual benefits with few-shot target samples only, ultimately leading to superior segmentation in various downstream tasks. We develop two CAT-SAM variants that adopt two tuning strategies for the image encoder: one injecting learnable prompt tokens in the input space and the other inserting lightweight adapter networks. Extensive experiments over 11 downstream tasks show that CAT-SAM achieves superior segmentation consistently even under the very challenging one-shot adaptation setup.", + "bbox": [ + 259, + 378, + 740, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Segment Anything Model $\\cdot$ Few-shot Learning $\\cdot$ Parameter Efficient Tuning", + "bbox": [ + 259, + 696, + 740, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 746, + 375, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Accurate image segmentation is a fundamental computer vision task that plays a pivotal role in various applications such as robotics, autonomous driving, health-", + "bbox": [ + 212, + 773, + 785, + 805 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Co-first authors with equal contributions. † Corresponding author.", + "bbox": [ + 230, + 811, + 687, + 825 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Project page: https://xiaooran.github.io/projects/CAT-SAM", + "bbox": [ + 230, + 825, + 668, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/069a80822340aee613fa1fe128d81549102ab37092b55803c0a0ff70fd76f468.jpg", + "image_caption": [ + "Fig. 1: The proposed CAT-SAM performs ConditionAl joint Tuning (CAT) to establish communication between SAM's heavyweight image encoder and lightweight mask decoder. This enables synergistic adaptation of the two network components, mitigating tuning imbalances and improving few-shot SAM adaptation." + ], + "image_footnote": [], + "bbox": [ + 320, + 150, + 684, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "care, earth observation, etc. The recently developed Segment Anything Model (SAM) [23], trained with 1.1 billion masks, has emerged as a momentous leap forward in image segmentation. By taking geometric prompts with points, boxes, or masks as input, SAM demonstrates remarkable zero-shot capability for general image segmentation, as well as great potential for fine-grained mask segmentation in different downstream tasks across a variety of contexts.", + "bbox": [ + 212, + 386, + 787, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, SAM often fails to generate high-quality predictions when dealing with domains that are either sparsely represented or lie outside its training distribution [5,19], such as challenging RGB images including aerial, medical, and intricate structural images, as well as non-RGB images like X-ray, Sonar, Synthetic Aperture Radar (SAR) images, etc. This greatly undermines the versatility and applicability of SAM as a foundational segmentation model while handling various downstream tasks. Several studies attempted to address this issue by fine-tuning SAM with a large number of target samples with mask annotations [21, 30]. However, they require extensive masked images for each specialized field, posing a formidable task with poor scalability. How to effectively adapt SAM toward various downstream tasks in a data-efficient manner has become an essential challenge in the segmentation field.", + "bbox": [ + 212, + 477, + 787, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We exploit few-shot target samples for effective and efficient adaptation of SAM toward various downstream tasks. Inspired by the recent advancements in parameter-efficient tuning for foundation models [1,13], we freeze the entire SAM and lightly expand its image encoder and mask decoder with a tiny amount of learnable parameters. This expansion preserves SAM's zero-shot capabilities and flexibility while capturing domain-specific features that are essential for the segmentation of downstream data. However, SAM's image encoder is much larger than its mask decoder, which directly leads to tuning imbalance and further suboptimal adaptation, especially under the limited supervision of few-shot target samples.", + "bbox": [ + 212, + 657, + 787, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this challenge, we design CAT-SAM, a novel ConditionAl Tuning network for effective and data-efficient adaptation of SAM. The key idea is to ex", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "A. Xiao and W. Xuan et al.", + "bbox": [ + 271, + 114, + 457, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ploit the lightweight decoder tuning to guide the heavyweight encoder tuning and formulate decoder-conditioned joint tuning of both as depicted in Fig. 1. To this end, we design a prompt bridge structure that maps the domain-specific features from the mask decoder to the image encoder. With the collaboration between the two partners, CAT-SAM enables synergistic adaptation of both which mitigates the tuning imbalance and improves the few-shot SAM adaptation effectively. In addition, the prompt bridge can be seamlessly embedded into two representative tuning approaches, including prompt tuning [20, 25, 27] that introduces learnable prompt tokens into the input space and adapters [15, 29, 37, 38] that leverage lightweight adaptive networks. This directly leads to two CAT-SAM variants, namely, CAT-SAM-T and CAT-SAM-A. Extensive evaluations over 11 diverse and challenging downstream tasks show that both variants achieve superior adaptation and segmentation with few-shot target samples only.", + "bbox": [ + 212, + 146, + 787, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The major contributions of this work can be summarized in three aspects. First, we propose CAT-SAM, a conditional tuning network that enables effective and data-efficient adaptation of SAM toward various challenging downstream domains. We design prompt bridge within CAT-SAM, a decoder-conditioned joint tuning structure that enables synergistic and data-efficient adaptation of the heavyweight image encoder and the lightweight mask decoder effectively. Second, we develop two CAT-SAM variants by embedding the prompt bridge into two representative tuning strategies, one introducing learnable prompt tokens in the input space and the other inserting lightweight adapter networks. Third, extensive experiments over 11 diverse segmentation datasets show that CAT-SAM achieves superior image segmentation consistently even under the challenging one-shot setup.", + "bbox": [ + 212, + 343, + 789, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 549, + 388, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Parameter-Efficient Tuning for foundation models has become increasingly crucial due to the over-size model parameters and costly one-task-one-model deployment. One popular approach involves updating only the newly added learnable parameters while keeping the model backbone frozen. These lightweight parameters are often introduced in two ways: 1) as prompt tuning [12,20,22,58] by adding learnable tokens to input tokens at transformer layers, 2) as adapters [11, 29,37,38,49] by integrating learnable lightweight sub-networks into transformer layers. However, prior visual studies primarily focus on adapting single backbone modules [3,20,29] or dual encoder backbones for vision-language tasks [22,57], which are suboptimal tuning solutions while applying to SAM due to its particular architecture of a heavyweight image encoder and a lightweight mask decoder. In contrast, we explore a novel conditional joint tuning strategy tailored to SAM, which facilitates its adaptation even with limited downstream data.", + "bbox": [ + 212, + 582, + 787, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SAM for Downstream Tasks. The release of SAM has spurred several subsequent studies. Some concentrate on expanding semantic recognition capabilities [24, 26, 43, 50, 54], while others aim to create more lightweight SAM variants [48, 52, 55] for faster computations. Several studies explore SAM's adapta", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "CAT-SAM", + "bbox": [ + 656, + 113, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion in underperforming downstream scenarios, primarily by directly fine-tuning SAM's mask decoder [4,8,30,41,46,53], which can severely degrade SAM's zero-shot capabilities. The recent HQ-SAM [21] freezes the entire pre-trained SAM and prompt-tunes the mask decoder with a new mask head for adaptive mask prediction. However, these adaptation studies rely on large target datasets with a considerable number of annotated images, which are often costly to collect. In contrast, our CAT-SAM achieves superior adaptation with only few-shot target samples, significantly reducing adaptation costs and expanding the applicability of SAM as a visual foundation model.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 304, + 330, + 320 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminaries of SAM", + "text_level": 1, + "bbox": [ + 215, + 333, + 442, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SAM [23] handles each image with three key modules: a heavyweight image encoder (i.e., ViT [9] as the backbone) that extracts image embeddings, a prompt encoder that encodes the geometric prompts (i.e., points, a box, or a coarse mask) to generate prompt embeddings, and a lightweight mask decoder that combines the two types of embeddings to predict segmentation masks. The released SAM model is trained on a super large-scale SA-1B dataset, which consists of over 11 million images as well as 1.1 billion automatically generated masks. SAM has demonstrated remarkable zero-shot capability while dealing with various conventional natural images, as well as superb flexibility in accepting various geometric prompt inputs. On the other hand, it often struggles while facing challenging downstream domains that are often sparse or falling outside of its training data distribution, such as RGB domains of aerial, medical, and intricate structural images, as well as non-RGB domains of X-ray, SAR, and Sonar images. More details about SAM are presented in [23].", + "bbox": [ + 212, + 357, + 787, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 CAT-SAM", + "text_level": 1, + "bbox": [ + 215, + 589, + 352, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This subsection introduces CAT-SAM, a ConditionAl Tuning network that is designed for adapting SAM toward various under-performed downstream data. The objective is to adapt SAM efficiently with only few-shot annotated target images, meanwhile preserving SAM's powerful zero-shot capability and geometric prompting flexibility.", + "bbox": [ + 212, + 612, + 787, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CAT-SAM freezes the entire SAM and facilitates the simultaneous tuning of SAM's image encoder and mask decoder to capture representative information from few-shot target samples for effective adaptation of SAM. However, a critical issue with simultaneous tuning arises from the clear imbalance between the heavyweight image encoder (308.3 M parameters for ViT-L [9]) and the lightweight mask decoder (4.1 M), which often leads to sub-optimal adaptation while only few-shot target samples are available. To tackle this challenge, we design decoder-conditioned joint tuning, a novel tuning strategy that mitigates the imbalance by establishing a linkage between the encoding tuning and the decoding tuning. Specifically, we design prompt bridge—a lightweight network that", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "A. Xiao and W. Xuan et al.", + "bbox": [ + 271, + 114, + 457, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fc430856e4daded764b4acbdea330eb06029422ea75a3c539c5fdffc38bcae2b.jpg", + "image_caption": [ + "(a) CAT-SAM-T" + ], + "image_footnote": [], + "bbox": [ + 220, + 148, + 380, + 280 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cd1359f00a47f4f792f48395e146e58919584e31308d1a27be1a36b51cea6189.jpg", + "image_caption": [ + "(b) CAT-SAM-A" + ], + "image_footnote": [], + "bbox": [ + 398, + 148, + 558, + 280 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/62da2b3ef65a27621f55538a957a9411f2ef56f2935339eafdca266b94a62478.jpg", + "image_caption": [ + "(c) Prompt Bridge", + "Fig. 2: Overview of CAT-SAM. CAT-SAM keeps the whole SAM frozen while simultaneously tuning the image encoder and mask decoder for downstream adaptation. To address the tuning imbalance between these two network components, we introduce decoder-conditioned joint tuning through the design of Prompt Bridge structure, enabling synergetic and enhanced adaptation. We present two CAT-SAM variants: CAT-SAM-T in (a) and CAT-SAM-A in (b), achieved by integrating the prompt bridge with prompt-based and adapter-based tuning strategies for the image encoder, respectively. (c) illustrates two tailored prompt bridge structures, PB-T and PB-A." + ], + "image_footnote": [], + "bbox": [ + 575, + 148, + 785, + 279 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "maps the domain-specific features from the mask decoder to the image encoder. The bridge design facilitates joint and balanced tuning of both network components, leading to synergistic and effective adaptation of SAM with few-shot and even a single target sample.", + "bbox": [ + 212, + 450, + 782, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We develop two CAT-SAM variants by integrating the proposed decoder-conditioned joint tuning into two prevalent tuning approaches, namely CAT-SAM-T and CAT-SAM-A, as illustrated in Fig. 2.", + "bbox": [ + 212, + 511, + 782, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tuning Image Encoder. We first brief SAM's image encoder and then introduce two tuning approaches in our CAT-SAM, including prompt tuning in CAT-SAM-T and lightweight adapters in CAT-SAM-A.", + "bbox": [ + 212, + 582, + 782, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- SAM's Image Encoder, denoted as $\\mathcal{V}$ with $K$ transformer layers $\\{\\mathcal{V}_i\\}_{i=1}^K$ , first splits the input image $I$ of the size $(H, W)$ into $M$ fixed-size patches, and then projects them into patch embeddings $E_0 \\in \\mathbb{R}^{M \\times d_v}$ . Patch embeddings $E_{i-1}$ are input to the $i^{th}$ transformer layer $(\\mathcal{V}_i)$ and sequentially processed through $K$ transformer layers:", + "bbox": [ + 225, + 630, + 784, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n[ E _ {i} ] = \\mathcal {V} _ {i} ([ E _ {i - 1} ]), i = 1, 2, \\dots , K.\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 720, + 627, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- CAT-SAM-T: As illustrated in Fig. 2 (a), we introduce a set of learnable tokens $\\{P_i \\in \\mathbb{R}^{d_v}\\}_{i=0}^{K-1}$ in the image encoder alongside the input patch embeddings. These new learnable tokens are introduced in each transformer layer of the image encoder $\\mathcal{V}$ and updated through the adaptation process:", + "bbox": [ + 225, + 750, + 782, + 810 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n[ E _ {i} ] = \\mathcal {V} _ {i} ([ E _ {i - 1}, P _ {i - 1} ]) ], i = 1, 2, \\dots , K.\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 823, + 650, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "CAT-SAM", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- CAT-SAM-A: As illustrated in Fig. 2 (b), we introduce Adapters [15] to tune SAM's image encoder in CAT-SAM-A. The adapter is a lightweight sub-network that is inserted into each transformer layer, comprising a linear down-projection, a nonlinear activation function, a linear up-projection, and a residual connection. High-frequency image information is also incorporated in CAT-SAM as the input of the adapters as in [29]. In this process, high-frequency components $I_{hfc}$ are first extracted from the input image $I$ and then reversed to the space domain using the Fast Fourier Transform and its inverse. These components are then partitioned into small patches $I_{hfc}^p$ similar to the image patches. Subsequently, convolutional layers and linear layers project $I_{hfc}$ and the original image embeddings $E_0$ into $c$ -dimensional features independently, yielding $F_{hfc}$ and $F_{pe}$ , respectively. The result of element-wise addition of $F_{pe}$ and $F_{hfc}$ is then forwarded (as denoted \"EVP\" in Fig. 2 (b)) to the adapter $\\mathrm{Adapt}_i$ for transformer layer $i$ . Finally, the output is added element-wise to the original input of the transformer layer $i$ :", + "bbox": [ + 225, + 145, + 787, + 373 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nE _ {i - 1} = E _ {i - 1} + \\mathrm {A d a p t} _ {i} (F _ {p e} + F _ {h f c}).\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 381, + 643, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tuning Mask Decoder. We tune the mask decoder of SAM similarly as [21]. Specifically, we freeze the entire pre-trained decoder of SAM and introduce a learnable CAT-Token that is concatenated with SAM's original output tokens and prompt tokens to form the input to SAM's mask decoder. The CAT-Token is updated during training but remains fixed during inference, the same as SAM's output tokens. After passing through two decoder layers, the updated CAT-Token is used to generate dynamic weights for a newly introduced three-layer MLP. Simultaneously, SAM's mask decoder features are fused with the image features decoded by the image encoder. Finally, the fused features and the three-layer MLP are combined through the dot product to generate the target mask. Please refer to the appendix for more details about the decoder structures in CAT-SAM.", + "bbox": [ + 212, + 415, + 787, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Decoder-Conditioned Joint Tuning. The proposed decoder-conditioned joint tuning can be seamlessly embedded via two implementations of the proposed prompt bridging ideas, i.e., PB-T and PB-A as illustrated in Fig. 2 (c).", + "bbox": [ + 212, + 616, + 790, + 660 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- CAT-SAM-T: The prompt bridge $\\mathrm{PB}_i^T$ is a two-layer MLP. It projects the CAT-Token $Q$ in the mask decoder to each transformer layer $i$ in the image encoder and outputs a single learnable token replacing $P_i$ by $\\tilde{P}_i^T = \\mathrm{PB}_i^T(Q)$ . These mapped tokens are directly applied for prompt tuning of the image encoder as follows:", + "bbox": [ + 225, + 661, + 787, + 736 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ E _ {i} \\right] = \\mathcal {V} _ {i} \\left(\\left[ E _ {i - 1}, \\tilde {P} _ {i - 1} ^ {T} \\right]\\right), i = 1, 2, \\dots , K.\n$$\n", + "text_format": "latex", + "bbox": [ + 379, + 743, + 651, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- CAT-SAM-A: The prompt bridge $\\mathrm{PB}^A$ maps the CAT-Token $Q$ into an embedding $\\tilde{P}^A = \\mathrm{PB}^A(Q)$ of the same size as $F_{pe}$ and $F_{hfc}$ . $\\tilde{P}^A$ is then element-wise added to the EVP before being fed to every adapter as follows:", + "bbox": [ + 227, + 770, + 787, + 816 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nE _ {i - 1} = E _ {i - 1} + \\mathrm {A d a p t} _ {i} (F _ {p e} + F _ {h f c} + \\tilde {P} ^ {A}).\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 821, + 666, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "A. Xiao and W. Xuan et al.", + "bbox": [ + 271, + 114, + 457, + 127 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/030ebc0460d94393032ec2222157b42fd9a533b1e1e22884b4c2cdfa62d20b11.jpg", + "table_caption": [ + "Table 1: Comparisons of trainable parameters between ViT-L [9] based SAM and two variants of the proposed CAT-SAM." + ], + "table_footnote": [], + "table_body": "
ModelTrainable Parameters (M)Trainable Params vs. SAM
SAM [23]312.4-
CAT-SAM-T3.31.1%
CAT-SAM-A1.90.6%
", + "bbox": [ + 220, + 184, + 777, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2 (c), PB-A employs $c$ different linear down-projection layers to generate embeddings independently. After GELU layers, these embeddings are up-projected to the dimension of $1 \\times M$ ( $M$ is the spatial size of the patch) with a shared linear layer to mitigate the computation cost.", + "bbox": [ + 240, + 285, + 785, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, the up-projected embeddings are concatenated to produce $\\tilde{P}^A$ .", + "bbox": [ + 240, + 345, + 751, + 361 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 1 compares the amount of trainable parameters between the SAM (implemented with the ViT-L [9]) and our CAT-SAM that is built upon the SAM. It can be observed that both CAT-SAM variants introduce very limited additional parameters, yet yield significant performance improvements in various downstream segmentation tasks as to be detailed in Section 4.", + "bbox": [ + 212, + 361, + 785, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Training and Inference", + "text_level": 1, + "bbox": [ + 215, + 458, + 450, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "During training, we feed target samples together with geometric prompts into CAT-SAM to generate CAT-SAM masks, under the supervision of ground-truth masks as well as a linear combination of BCE loss and dice loss [31]. We freeze the pre-trained SAM parameters and update solely the parameters that are introduced in the above-described tuning modules. Note we train CAT-SAM with a mixture of sampled geometric prompts, including bounding boxes, randomly selected points, and coarse masks. In addition, we create degraded coarse masks by introducing random Gaussian noise in the boundary regions of the ground-truth masks as in [21]. Further training details can be found in Section 4.1 and the appendix.", + "bbox": [ + 212, + 483, + 787, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "During the inference phase, the input image is fed into SAM's image encoder along with the integrated prompt tokens in CAT-SAM-T or adapters in CAT-SAM-A, to generate adaptive image embeddings. These embeddings, combined with the prompt tokens from SAM's prompt encoder, serve as input to the mask decoder. Subsequently, the updated CAT-token and associated MLP are utilized for the target mask prediction. Lastly, we up-sample the mask to the original resolution to produce the final output.", + "bbox": [ + 212, + 633, + 787, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 763, + 375, + 780 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate CAT-SAM's efficacy through comprehensive experiments involving eight segmentation tasks across 11 datasets, all hailing from challenging downstream fields that SAM struggles to address effectively. Our experiments encom", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "CAT-SAM", + "bbox": [ + 656, + 113, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f29f503d8b0287e3d3765e9780558716ff6899eac8abf3348b057bce8af41b66.jpg", + "table_caption": [ + "Table 2: We evaluate and benchmark CAT-SAM over eight challenging segmentation tasks across 11 datasets." + ], + "table_footnote": [], + "table_body": "
TasksDataset NamesImagery
Building segmentationWHU [18]Aerial images
Road segmentationMA. Roads [32]Aerial images
Polyp segmentationKvasir [34]Medical images
Chest X-ray segmentationJSRT [39]X-ray images
Marine debris segmentationFLS [40]Sonar images
Ship instance segmentationHRSID [45]SAR images
Shadow segmentationSBU-Shadow [42]Natural images
Intricate segmentationDIS [35],ThinObject [28],HRSOD [51],COIFT [28]Natural images
", + "bbox": [ + 217, + 185, + 810, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pass a wide array of scenarios and tasks, spanning from varying target shots to different imagery modalities.", + "bbox": [ + 212, + 367, + 784, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 416, + 426, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments on two collections of datasets listed in Tab. 2: 1) True-color images, including WHU [18] for building segmentation, Kvasir [34] for polyp segmentation, SBU-Shadow [42] for shadow segmentation, Massachusetts (MA.) Roads [32] for road segmentation, as well as fine-grained segmentation datasets DIS [35], ThinObject [28], HRSOD [51], and COIFT [28]. 2) Non-RGB domains, encompassing JSRT [39] for chest organ segmentation (X-ray images), HRSID [45] for ship instance segmentation (high-resolution SAR images), and FLS [40] for marine debris segmentation (Sonar images). We use official dataset splits for fair comparisons.", + "bbox": [ + 212, + 436, + 794, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation Metrics. For single-class datasets including WHU, Kvasir, SBU-Shadow, and MA. Roads, we employ standard mask IoU of foreground. JSRT and FLS, which have multiple classes, are evaluated using individual class IoU and their average. As for DIS, ThinObject, HRSOD, and COIFT, we follow [21] and report mIoU and the boundary metrics (mBIoU) [7] for fair comparisons. For instance segmentation of HRSID, we use standard AP, $\\mathrm{AP}_{50}$ , and $\\mathrm{AP}_{75}$ .", + "bbox": [ + 212, + 573, + 785, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation Details. Unless otherwise specified, CAT-SAM and its comparing models use ViT-L [9] as the image encoder backbone. We train on 1 NVIDIA RTX A6000 GPU for one-shot adaptation, while 4 GPUs for multiple-shots. Following [21], ground truth boxes serve as the default input geometric prompts during evaluation to ensure fair comparison and minimize randomness, with the exception of Fig. 3 where point prompts are evaluated. Please refer to the appendix for more implementation details.", + "bbox": [ + 212, + 664, + 785, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 787, + 401, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We first investigate the contributions of different tuning modules within both CAT-SAM variants to assess their impact on overall adaptation performance.", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "A. Xiao and W. Xuan et al.", + "bbox": [ + 271, + 114, + 457, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/47913fa1d14db37ea3662c230ca49d1529987a168de0244127baaeff6b17adac.jpg", + "table_caption": [ + "Table 3: Ablation study of CAT-SAM-T in (a) and CAT-SAM-A in (b) for 1-shot adaptation over datasets: \"W\" for WHU [18] (on building segmentation), \"K\" for Kvasir [34] (on polyp segmentation), and \"S\" for SBU-Shadow [42] (on shadow segmentation). \"Enc-T.\" and \"Enc-A.\" means tuning the image encoder with prompt tokens and adapters, respectively. \"Dec.\" means tuning the mask decoder. \"PB-T\" and \"PB-A\" denote two customized prompt bridges for conditional tuning. Model 1 is the baseline SAM without adaptation.", + "(a) CAT-SAM-T." + ], + "table_footnote": [], + "table_body": "
Enc-T. Dec. PB-TWKSAVG
143.579.062.461.6
2✓57.880.476.0
3✓71.273.863.5
4✓✓66.473.261.8
5✓✓✓86.883.4
", + "bbox": [ + 225, + 268, + 506, + 375 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/a4d92c61b23d4638a303b3fee5752a7d24ba44e57f01a0d34366c4871332a3b8.jpg", + "table_caption": [ + "(b) CAT-SAM-A." + ], + "table_footnote": [], + "table_body": "
Enc-A. Dec. PB-AWKSAVG
143.579.062.461.6
2✓86.573.776.3
3✓71.273.863.5
4✓✓83.479.876.3
5✓✓✓88.285.4
", + "bbox": [ + 511, + 268, + 792, + 375 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tab. 3 shows the experiments of one-shot adaptation across three true-color image segmentation benchmarks WHU [18], Kvasir [34], and SBU-Shadow [42]. For each CAM-SAM variant CAM-SAM-T/CAM-SAM-A, we compare 5 models including: 1. The original SAM without adaptation (baseline); 2. Tuning SAM's image encoder alone with prompt tokens or adapters; 3. Tuning SAM's mask decoder alone; 4. Tuning SAM's image encoder and mask decoder independently; and 5. Conditional tuning of SAM's image encoder and mask decoder with the proposed prompt bridging (i.e., the complete CAT-SAM-T and CAT-SAM-A).", + "bbox": [ + 212, + 404, + 787, + 525 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We can see that the original SAM struggles to produce high-quality masks in these challenging domains, indicating its limitations as a foundational segmentation model for many downstream applications. Tuning the image encoder or the mask decoder alone in 2 and 3 enhances target segmentation, underlining the efficacy and necessity of SAM adaptation. However, simultaneously tuning the image encoder and mask decoder in an independent way in model 4 does not show complementary and consistent improvement. This suggests instability and imbalance in tuning between the heavyweight image encoder and the lightweight mask decoder, which could lead to suboptimal adaptation solutions. Differently, including the proposed prompt bridging with either PB-T or PB-A (i.e., the complete CAT-SAM-T and CAT-SAM-A) mitigates the tuning imbalance and produces substantial performance enhancement consistently.", + "bbox": [ + 212, + 527, + 787, + 708 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Comparison with the State-of-the-Art", + "text_level": 1, + "bbox": [ + 214, + 734, + 576, + 750 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "CAT-SAM for One-Shot Adaptation. To assess CAT-SAM's effectiveness in reducing downstream training data and its adaptability with few-shot samples, we first evaluate it under the extremely challenging one-shot adaptation scenario. Given the lack of prior studies under such setup, we comprehensively benchmark it against closely related methods that can be broadly categorized into", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "CAT-SAM", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/1090a075e0542cb192d665fff8076f46685cf3e5d88d6541ed9623286673c818.jpg", + "table_caption": [ + "Table 4: Comparison of adaptive segmentation performance on challenging true-color datasets: WHU [18] (on building segmentation), Kvasir [34] (on Polyp segmentation), and SBU-Shadow [42] (on shadow segmentation). The baseline is SAM without adaptation. All compared methods utilize one-shot sample for adaptation, except for those denoted by * which utilize 20 shots with metric numbers sourced from [47]." + ], + "table_footnote": [], + "table_body": "
MethodsWHUKvasirSBU-ShadowAverage
SAM [23] (baseline)43.579.062.461.6
VPT-shallow [20]60.879.868.769.8
VPT-deep [20]57.880.476.071.4
AdaptFormer [3]83.276.877.279.1
LoRA [16]86.177.574.479.3
Med-SA [46]34.528.622.128.4
MedSAM [30]30.629.853.437.9
SimAda [41]48.618.749.338.9
SPFS-SAM* [47] (20-shots)-53.4--
SAMed* [53] (20-shots)-51.7--
HQ-SAM [21]71.273.863.569.5
CAT-SAM-T (Ours)86.883.478.082.7
CAT-SAM-A (Ours)88.285.481.985.2
", + "bbox": [ + 217, + 226, + 785, + 457 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "two groups. The first group comprises state-of-the-art tuning methods for foundation models of different modalities, including VPT [20] and AdaptFormer [3] for large vision foundation models, and LoRA [16] for large language models. The second group comprises recent SAM-based adaptation methods which were largely designed for full-shot adaptation [21,30,41,46,53] ( [47] tackles few-shot adaptation but requires 20 shots).", + "bbox": [ + 215, + 500, + 785, + 590 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tab. 4 presents results over WHU, Kvasir, and SBU-Shadow datasets. We can observe that all three tuning methods for foundation models exhibit clear adaptation effects. Notably, AdaptFormer and LoRA demonstrate significantly better performance than VPT-deep and VPT-shallow, highlighting the effectiveness of modifying the model's structure over prompt tuning in the input space during SAM adaptation. On the other hand, most SAM-based adaptation methods, including Med-SA, SimAda, SPFS-SAM, SAMed, and MedSAM, yield much lower performance, primarily due to their reliance on many more target images which leads to overfitting while learning from one-shot target sample. HQ-SAM achieves substantial segmentation improvements but it still cannot compete with the three tuning methods, demonstrating the significance of tuning the image encoder over the mask decoder. Finally, both CAT-SAM variants consistently outperform all compared methods, with CAT-SAM-A exhibiting a slight edge due to its network structure modifications. These experiments validate the superiority of the proposed decoder-conditioned joint tuning approach in extracting domain-specific features with few-shot samples.", + "bbox": [ + 215, + 598, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "A. Xiao and W. Xuan et al.", + "bbox": [ + 271, + 114, + 457, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e88927a2f36f45a1c40543bcab6081cffa0744a0c86cb9c31a0f354b62dd555a.jpg", + "table_caption": [ + "Table 5: Comparison of state-of-the-art segmentation achieved by fully supervised learning (FSL) methods trained on full-shot data (upper part) and mask segmentation by SAM and our CAT-SAM (lower part). Evaluation follows official criteria of each FSL benchmark, reporting IoU for WHU, mIoU for Kvasir, and BER for SBU-Shadow to ensure consistency. \"(1)\", \"(16)\", and \"(full)\" after CAT-SAM-T/CAT-SAM-A denote tuning with 1-shot, 16-shots, and full-shots samples." + ], + "table_footnote": [], + "table_body": "
WHUKvasirSBU-Shadow
ModelIoUModelmIoUModelBER(↓)
STT [2]90.5CASCADE [36]87.8BDRAR [59]3.64
BuildFormer [44]91.4PatchRefineNet [33]89.1DSDNet [56]3.45
CBR-Net [14]91.4DUCK-Net [10]90.5MTMT [6]3.15
SAM43.5SAM87.3SAM13.22
CAT-SAM-T (1)86.8CAT-SAM-T (1)90.0CAT-SAM-T (1)7.81
CAT-SAM-A (1)88.2CAT-SAM-A (1)91.3CAT-SAM-A (1)5.27
CAT-SAM-T (16)89.6CAT-SAM-T (16)93.1CAT-SAM-T (16)4.04
CAT-SAM-A (16)90.3CAT-SAM-A (16)93.6CAT-SAM-A (16)3.80
CAT-SAM-T (full)93.3CAT-SAM-T (full)94.5CAT-SAM-T (full)2.54
CAT-SAM-A (full)93.6CAT-SAM-A (full)94.3CAT-SAM-A (full)2.39
", + "bbox": [ + 215, + 239, + 792, + 438 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Full-Shot Learning. While CAT-SAM is primarily designed for few-shot adaptation, it also demonstrates proficiency in full-shot adaptation scenarios. In Tab. 5, we present the state-of-the-art performances of traditional fully-supervised learning (FSL) methods trained over full-shot data and compare them with SAM and our CAT-SAM (adapted with various target samples). We report official evaluation metrics for three public benchmarks, including IoU for WHU, mIoU for Kvasir, and BER [17] for SBU-Shadow. Note the distinction in setups between the FSL segmentation and the mask segmentation: FSL focuses on traditional semantic segmentation of specific classes without prompt input while SAM and CAT-SAM involve general mask segmentation with prompt input. Despite these differences, these performances can serve as valuable references for evaluating CAT-SAM's segmentation ability.", + "bbox": [ + 212, + 470, + 789, + 652 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "With full target training data, both CAT-SAM variants outperform state-of-the-art FSL methods consistently. Notably, the two CAT-SAM variants achieve impressive target segmentation under the challenging one-shot setup with a single target sample. With 16 shots of target samples, their performance is even on par with the FSL methods that are trained by using full target training data. These experiments underscore the superiority of effective transfer of broad knowledge from the foundation model SAM as compared with training from scratch with FSL. It also demonstrates the remarkable efficacy of CAT-SAM in substantially expanding SAM's applicability across various downstream tasks.", + "bbox": [ + 212, + 655, + 787, + 792 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We conducted further experiments to compare CAT-SAM against HQ-SAM [21], a state-of-the-art SAM-based adaptation solution for intricate structural image segmentation. For fair comparisons, we follow the evaluation protocol of", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "CAT-SAM", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/68682ba8097a04ead049c12213791d9b67dc23edd850f2eabf936913ec79ff54.jpg", + "table_caption": [ + "Table 6: Extremely fine-grained segmentation over four intricate structural image datasets. HQ-SAM and the two CAT-SAM variants are fine-tuned over full HQSeg-44K [21]. All adopt the boxes converted from their GT masks as the box prompt input for fair comparisons." + ], + "table_footnote": [], + "table_body": "
ModelsDIS [35]COIFT [28]HRSOD [51]ThinObject [28]Average
mIoUmBIoUmIoUmBIoUmIoUmBIoUmIoUmBIoUmIoUmBIoU
SAM [23]62.052.892.186.590.283.173.661.879.571.1
HQ-SAM [21]78.670.494.890.193.686.989.579.989.181.8
CAT-SAM-T84.078.195.692.093.487.694.087.991.786.4
CAT-SAM-A83.677.795.692.293.187.194.188.291.686.3
", + "bbox": [ + 217, + 212, + 785, + 313 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "HQ-SAM and tune CAT-SAM on the complete HQSeg-44K dataset, consisting of 44,320 meticulously annotated image masks. Tab. 6 shows experiments across four exceptionally fine-grained segmentation datasets including DIS [35], ThinObject [28], HRSOD [51], and COIFT [28] (in mIoU and the boundary metrics (mBIoU) [7]). We can see that CAT-SAM-T achieves slightly better performance than CAT-SAM-A, while both clearly outperform HQ-SAM and the original SAM. These experiments further demonstrate the efficacy of our proposed conditional tuning in adapting SAM toward challenging downstream tasks. Visual qualitative comparisons are provided in the appendix.", + "bbox": [ + 215, + 340, + 784, + 477 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 CAT-SAM for Non-RGB Domains", + "text_level": 1, + "bbox": [ + 217, + 497, + 545, + 512 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We also explore CAT-SAM's performance across more challenging image modalities characterized by larger domain discrepancies. Specifically, we assess its efficacy on non-RGB domains, including JSRT [39] for chest organ segmentation using X-ray images and FLS [40] for marine debris segmentation with Sonar images, and HRSID [45] for ship instance segmentation with SAR images. For adaptation, we randomly select either one or sixteen target images for 1-shot and 16-shot scenarios, respectively, while employing the entire training set for full-shot adaptation. These datasets represent novel image distributions as compared with SAM's training data. Note we train instances of all classes together while evaluating them by classes. Since no images in FLS contain objects of all 11 classes, we only report results of 16-shot and full-shot adaptation for FLS.", + "bbox": [ + 215, + 522, + 784, + 686 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Tab. 7 shows experimental results. Notably, CAT-SAM-T and CAT-SAM-A improve the mIoU to $93.0\\%$ and $92.6\\%$ , respectively, with only one shot of JSRT training image. The improvements saturate gradually as the number of training samples increases, with marginal gains of $1.4\\%$ and $2.0\\%$ , respectively, when adapted with full-shot target data. These findings underscore the superior few-shot capability of the proposed CAT-SAM and align well with those reported in Tab. 5. Experiments on FLS [40] show different results. While both CAT-SAM variants exhibit clear adaptation effects, the performance gains for both 16-shot and full-shot adaptation are notably lower as compared with previously evaluated datasets. The discrepancy is largely attributed to two primary", + "bbox": [ + 215, + 688, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "A. Xiao and W. Xuan et al.", + "bbox": [ + 271, + 114, + 457, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/7fe6d1c848573f55a82cf00700efdf7b4d336f9433ce39d0d0c1a0abce40395f.jpg", + "table_caption": [ + "Table 7: Few-shot adaptation with varying number of training samples for non-RGB datasets, including FLS [40] with Sonar images for marine debris segmentation, JSRT [39] for chest X-ray segmentation, and HRSID [45] for ship instance segmentation with SAR images. \"N/A\" denotes that 1-shot adaptation over FLS is unattainable due to the absence of images containing objects from all 11 classes." + ], + "table_footnote": [], + "table_body": "
#SamplesMethodsJSRT [39]FLS [40]HRSID [45]
LungsHeartmIoUmIoUAPAP50AP75
NoneSAM [23]85.071.978.569.738.286.626.8
1-shotCAT-SAM-T95.490.593.0N/A46.088.445.4
CAT-SAM-A94.790.592.6N/A44.988.041.4
16-shotCAT-SAM-T95.892.794.273.246.289.345.2
CAT-SAM-A95.591.593.571.445.789.443.7
Full-shotCAT-SAM-T96.392.694.481.751.493.055.0
CAT-SAM-A96.492.894.682.052.993.956.0
", + "bbox": [ + 222, + 226, + 776, + 393 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "factors: 1) the inherent difficulty of FLS with its 11 semantic classes (per-class performances are provided in appendix); and 2) the larger variation of FLS data in shapes and object sizes as compared with previously evaluated data. As for ship instance segmentation using SAR images in HRSID [45], this dataset presents different challenges as ships are very small in size. The results show that CAT-SAM consistently outperforms SAM in this diverse domain, highlighting its efficiency in domain adaptation with large domain discrepancy. These experiments further underscore the need for a more comprehensive adaptation of SAM for exceptionally challenging downstream tasks. We expect further studies along this research direction.", + "bbox": [ + 212, + 422, + 787, + 574 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Analysis & Discussion", + "text_level": 1, + "bbox": [ + 215, + 599, + 444, + 614 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Prompts with Single Point. SAM provides flexibility in geometric prompts including a single foreground point. However, prompting with a single foreground point can be challenging as the single point could correspond to multiple objects [23]. Nonetheless, adaptation with few-shot target samples can alleviate this ambiguity by guiding the model to focus on the specific foreground distribution within the annotation space. We evaluate this by using the same randomly-selected single point as the geometric prompt, and Fig. 3 shows", + "bbox": [ + 212, + 628, + 545, + 810 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "the segmentation results for images from WHU, Kvasir, SBU-Shadow, and MA. Roads [32] on road segmentation.", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/918a78860252361b79e735130a6795a93b496d59c3d6b858d0076014486d2089.jpg", + "image_caption": [ + "Fig. 3: Single point to mask evaluation for one-shot adaptation." + ], + "image_footnote": [], + "bbox": [ + 555, + 660, + 782, + 760 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "CAT-SAM", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/288cfba52fa1f43e027ad782621ea1f43be7e63611c172717c612dbb21c2cb98.jpg", + "image_caption": [ + "Fig. 4: Visual comparisons of SAM [23] (top row) and CAT-SAM (bottom row). We illustrate samples from WHU [18] for building segmentation, MA. Roads [32] for road segmentation, SBU-Shadow [42] for shadow segmentation, Kvasir [34] for polyp segmentation, JSRT [39] for chest organ segmentation (X-ray images), FLS [40] for marine debris segmentation (Sonar images), and HRSID [45] for ship instance segmentation (SAR images). CAT-SAM exhibits one-shot adaptation across most datasets, except for 16-shot over FLS. Red boxes and stars denote geometric prompts, colored regions are mask predictions, and lines show the boundary of ground truth segmentation." + ], + "image_footnote": [], + "bbox": [ + 218, + 147, + 782, + 273 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Similar to the experiments using boxes as prompts, CAT-SAM consistently demonstrates superior target segmentation while using a single point as prompt, underscoring its robustness in adaptation while preserving SAM's inherent geometric prompting flexibility. In addition, both CAT-SAM variants achieve relatively less improvement for the dataset Kvasir. This can be largely attributed to the high similarities in color and texture between the polyps and the local gastrointestinal regions in Kvasir, making it challenging for a single point to provide clear geometric guidance.", + "bbox": [ + 212, + 429, + 787, + 549 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Visual Illustrations. Fig. 4 shows qualitative comparisons between SAM and CAT-SAM across multiple segmentation tasks. These illustrations demonstrate how our proposed CAT-SAM remarkably enhances the mask quality of SAM under the presence of only one-shot target samples. Please refer to the appendix for more visual comparisons.", + "bbox": [ + 212, + 550, + 787, + 626 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 652, + 359, + 669 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduce CAT-SAM, a conditional tuning network tailored for few-shot adaptation of SAM. We propose decoder-conditional joint tuning to mitigate the tuning imbalance between SAM's heavyweight image encoder and lightweight mask decoder and facilitate efficient SAM adaptation. To achieve this, we design the prompt bridge structures, enabling interaction when tuning these two network modules. We develop two CAT-SAM variants by integrating the prompt bridge with two representative tuning strategies including prompt tuning and adapter. Our comprehensive evaluation across diverse scenarios and tasks on 11 segmentation datasets underscores the superior domain adaptive efficiency of both CAT-SAM variants, even in the extremely challenging one-shot scenario.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "A. Xiao and W. Xuan et al.", + "bbox": [ + 271, + 114, + 457, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This project is partly funded by the Ministry of Education Singapore, under the Tier-1 scheme with project number RG18/22 and the Tier-2 scheme with project number MOE-T2EP20123-0003. It is also supported in part by the National Natural Science Foundation of China [grant no. U2033210], in part by the Zhejiang Provincial Natural Science Foundation [grant no. LDT23F02024F02].", + "bbox": [ + 212, + 175, + 787, + 252 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 273, + 321, + 287 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., et al.: On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021)", + "2. Chen, K., Zou, Z., Shi, Z.: Building extraction from remote sensing images with sparse token transformers. Remote Sensing 13(21), 4441 (2021)", + "3. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022)", + "4. Chen, T., Zhu, L., Deng, C., Cao, R., Wang, Y., Zhang, S., Li, Z., Sun, L., Zang, Y., Mao, P.: Sam-adapter: Adapting segment anything in underperformed scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 3367-3375 (October 2023)", + "5. Chen, T., Zhu, L., Ding, C., Cao, R., Zhang, S., Wang, Y., Li, Z., Sun, L., Mao, P., Zang, Y.: Sam fails to segment anything?—sam-adapter: Adapting sam in underperformed scenes: Camouflage, shadow, and more. arXiv preprint arXiv:2304.09148 (2023)", + "6. Chen, Z., Zhu, L., Wan, L., Wang, S., Feng, W., Heng, P.A.: A multi-task mean teacher for semi-supervised shadow detection. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 5611-5620 (2020)", + "7. Cheng, B., Girshick, R., Dollar, P., Berg, A.C., Kirillov, A.: Boundary iou: Improving object-centric image segmentation evaluation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15334-15342 (2021)", + "8. Cheng, J., Ye, J., Deng, Z., Chen, J., Li, T., Wang, H., Su, Y., Huang, Z., Chen, J., Jiang, L., et al.: Sam-med2d. arXiv preprint arXiv:2308.16184 (2023)", + "9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "0. Dumitru, R.G., Peteleaza, D., Craciun, C.: Using duck-net for polyp image segmentation. Scientific Reports 13(1), 9803 (2023)", + "1. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision pp. 1-15 (2023)", + "2. Gong, Z., Li, F., Deng, Y., Bhattacharjee, D., Zhu, X., Ji, Z.: Coda: Instructive chain-of-domain adaptation with severity-aware visual prompt tuning (2024)", + "3. Gu, J., Han, Z., Chen, S., Beirami, A., He, B., Zhang, G., Liao, R., Qin, Y., Tresp, V., Torr, P.: A systematic survey of prompt engineering on vision-language foundation models. arXiv preprint arXiv:2307.12980 (2023)" + ], + "bbox": [ + 225, + 301, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "CAT-SAM", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Guo, H., Du, B., Zhang, L., Su, X.: A coarse-to-fine boundary refinement network for building footprint extraction from remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing 183, 240-252 (2022)", + "15. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019)", + "16. Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2021)", + "17. Hu, X., Zhu, L., Fu, C.W., Qin, J., Heng, P.A.: Direction-aware spatial context features for shadow detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7454-7462 (2018)", + "18. Ji, S., Wei, S., Lu, M.: Fully convolutional networks for multisource building extraction from an open aerial and satellite imagery data set. IEEE Transactions on geoscience and remote sensing 57(1), 574-586 (2018)", + "19. Ji, W., Li, J., Bi, Q., Li, W., Cheng, L.: Segment anything is not always perfect: An investigation of sam on different real-world applications. arXiv preprint arXiv:2304.05750 (2023)", + "20. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022)", + "21. Ke, L., Ye, M., Danelljan, M., Liu, Y., Tai, Y.W., Tang, C.K., Yu, F.: Segment anything in high quality. arXiv preprint arXiv:2306.01567 (2023)", + "22. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19113-19122 (2023)", + "23. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)", + "24. Lan, X., Lyu, J., Jiang, H., Dong, K., Niu, Z., Zhang, Y., Xue, J.: Foodsam: Any food segmentation. IEEE Transactions on Multimedia pp. 1-14 (2023). https://doi.org/10.1109/TMM.2023.3330047", + "25. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021)", + "26. Li, F., Zhang, H., Sun, P., Zou, X., Liu, S., Yang, J., Li, C., Zhang, L., Gao, J.: Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767 (2023)", + "27. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021)", + "28. Liew, J.H., Cohen, S., Price, B., Mai, L., Feng, J.: Deep interactive thin object selection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 305-314 (2021)", + "29. Liu, W., Shen, X., Pun, C.M., Cun, X.: Explicit visual prompting for low-level structure segmentations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19434-19445 (2023)", + "30. Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature Communications 15(1), 654 (2024)", + "31. Miletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks for volumetric medical image segmentation. In: 2016 fourth international conference on 3D vision (3DV). pp. 565-571. IEEE (2016)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "A. Xiao and W. Xuan et al.", + "bbox": [ + 271, + 114, + 457, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Mnih, V.: Machine Learning for Aerial Image Labeling. Ph.D. thesis, University of Toronto (2013)", + "33. Nagendra, S., Kifer, D.: Patchrefinenet: Improving binary segmentation by incorporating signals from optimal patch-wise binarization. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1361-1372 (2024)", + "34. Pogorelov, K., Randel, K.R., Griwodz, C., Eskeland, S.L., de Lange, T., Johansen, D., Spampinato, C., Dang-Nguyen, D.T., Lux, M., Schmidt, P.T., et al.: Kvasir: A multi-class image dataset for computer aided gastrointestinal disease detection. In: Proceedings of the 8th ACM on Multimedia Systems Conference. pp. 164-169 (2017)", + "35. Qin, X., Dai, H., Hu, X., Fan, D.P., Shao, L., Van Gool, L.: Highly accurate dichotomous image segmentation. In: European Conference on Computer Vision. pp. 38-56. Springer (2022)", + "36. Rahman, M.M., Marculescu, R.: Medical image segmentation via cascaded attention decoding. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6222-6231 (2023)", + "37. Rebuffi, S.A., Bilen, H., Vedaldi, A.: Learning multiple visual domains with residual adapters. Advances in neural information processing systems 30 (2017)", + "38. Rebuffi, S.A., Bilen, H., Vedaldi, A.: Efficient parametrization of multi-domain deep neural networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8119-8127 (2018)", + "39. Shiraishi, J., Katsuragawa, S., Ikezoe, J., Matsumoto, T., Kobayashi, T., Komatsu, K.i., Matsui, M., Fujita, H., Kodera, Y., Doi, K.: Development of a digital image database for chest radiographs with and without a lung nodule: receiver operating characteristic analysis of radiologists' detection of pulmonary nodules. American Journal of Roentgenology 174(1), 71-74 (2000)", + "40. Singh, D., Valdenegro-Toro, M.: The marine debris dataset for forward-looking sonar semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3741-3749 (2021)", + "41. Song, Y., Zhou, Q., Lu, X., Shao, Z., Ma, L.: Simada: A simple unified framework for adapting segment anything model in underperformed scenes. arXiv preprint arXiv:2401.17803 (2024)", + "42. Vicente, T.F.Y., Hou, L., Yu, C.P., Hoai, M., Samaras, D.: Large-scale training of shadow detectors with noisily-annotated shadow examples. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI 14. pp. 816-832. Springer (2016)", + "43. Wang, H., Vasu, P.K.A., Faghri, F., Vemulapalli, R., Farajtabar, M., Mehta, S., Rastegari, M., Tuzel, O., Pouransari, H.: Sam-clip: Merging vision foundation models towards semantic and spatial understanding. arXiv preprint arXiv:2310.15308 (2023)", + "44. Wang, L., Fang, S., Meng, X., Li, R.: Building extraction with vision transformer. IEEE Transactions on Geoscience and Remote Sensing 60, 1-11 (2022)", + "45. Wei, S., Zeng, X., Qu, Q., Wang, M., Su, H., Shi, J.: Hrsid: A high-resolution sar images dataset for ship detection and instance segmentation. IEEE Access 8, 120234-120254 (2020)", + "46. Wu, J., Fu, R., Fang, H., Liu, Y., Wang, Z., Xu, Y., Jin, Y., Arbel, T.: Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "CAT-SAM", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "47. Wu, Q., Zhang, Y., Elbatel, M.: Self-prompting large vision models for few-shot medical image segmentation. In: MICCAI Workshop on Domain Adaptation and Representation Transfer. pp. 156-167. Springer (2023)", + "48. Xiong, Y., Varadarajan, B., Wu, L., Xiang, X., Xiao, F., Zhu, C., Dai, X., Wang, D., Sun, F., Iandola, F., et al.: Efficientsam: Leveraged masked image pretraining for efficient segment anything. arXiv preprint arXiv:2312.00863 (2023)", + "49. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2945-2954 (2023)", + "50. Yang, H., Ma, C., Wen, B., Jiang, Y., Yuan, Z., Zhu, X.: Recognize any regions. arXiv preprint arXiv:2311.01373 (2023)", + "51. Zeng, Y., Zhang, P., Zhang, J., Lin, Z., Lu, H.: Towards high-resolution salient object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7234-7243 (2019)", + "52. Zhang, C., Han, D., Qiao, Y., Kim, J.U., Bae, S.H., Lee, S., Hong, C.S.: Faster segment anything: Towards lightweight sam for mobile applications. arXiv preprint arXiv:2306.14289 (2023)", + "53. Zhang, K., Liu, D.: Customized segment anything model for medical image segmentation. arXiv preprint arXiv:2304.13785 (2023)", + "54. Zhang, R., Jiang, Z., Guo, Z., Yan, S., Pan, J., Dong, H., Gao, P., Li, H.: Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048 (2023)", + "55. Zhao, X., Ding, W., An, Y., Du, Y., Yu, T., Li, M., Tang, M., Wang, J.: Fast segment anything. arXiv preprint arXiv:2306.12156 (2023)", + "56. Zheng, Q., Qiao, X., Cao, Y., Lau, R.W.: Distraction-aware shadow detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5167-5176 (2019)", + "57. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022)", + "58. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)", + "59. Zhu, L., Deng, Z., Hu, X., Fu, C.W., Xu, X., Qin, J., Heng, P.A.: Bidirectional feature pyramid network with recurrent attention residual modules for shadow detection. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 121-136 (2018)" + ], + "bbox": [ + 215, + 146, + 784, + 632 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "A. Xiao and W. Xuan et al.", + "bbox": [ + 271, + 114, + 457, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/CAT-SAM_ Conditional Tuning for Few-Shot Adaptation of Segment Anything Model/bd649bc1-acc8-4027-a33f-08ccf2414aa5_model.json b/2024/CAT-SAM_ Conditional Tuning for Few-Shot Adaptation of Segment Anything Model/bd649bc1-acc8-4027-a33f-08ccf2414aa5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ce7241a5c95044c62b42e9a741890003c2c02b96 --- /dev/null +++ b/2024/CAT-SAM_ Conditional Tuning for Few-Shot Adaptation of Segment Anything Model/bd649bc1-acc8-4027-a33f-08ccf2414aa5_model.json @@ -0,0 +1,2337 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.245, + 0.141, + 0.758, + 0.187 + ], + "angle": 0, + "content": "CAT-SAM: Conditional Tuning for Few-Shot Adaptation of Segment Anything Model" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.212, + 0.772, + 0.245 + ], + "angle": 0, + "content": "Aoran Xiao\\(^{1*}\\), Weihao Xuan\\(^{2,3*}\\), Heli Qi\\(^{4}\\), Yun Xing\\(^{1}\\), Ruijie Ren\\(^{5}\\), Xiaojin Zhang\\(^{6}\\), Ling Shao\\(^{7}\\), and Shijian Lu\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.255, + 0.66, + 0.297 + ], + "angle": 0, + "content": "1 Nanyang Technological University, Singapore \n2 The University of Tokyo, Japan \n3 RIKEN AIP, Japan" + }, + { + "type": "text", + "bbox": [ + 0.333, + 0.297, + 0.67, + 0.325 + ], + "angle": 0, + "content": "\\(^{4}\\) Nara Institute of Science and Technology, Japan \n\\(^{5}\\) Waseda University, Japan" + }, + { + "type": "text", + "bbox": [ + 0.355, + 0.325, + 0.646, + 0.339 + ], + "angle": 0, + "content": "6 Zhejiang University of Technology, China" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.339, + 0.76, + 0.352 + ], + "angle": 0, + "content": "UCAS-Terminus AI Lab, University of Chinese Academy of Sciences, China" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.38, + 0.741, + 0.685 + ], + "angle": 0, + "content": "Abstract. The Segment Anything Model (SAM) has demonstrated remarkable zero-shot capability and flexible geometric prompting in general image segmentation. However, it often struggles in domains that are either sparsely represented or lie outside its training distribution, such as aerial, medical, and non-RGB images. Recent efforts have predominantly focused on adapting SAM to these domains using fully supervised methods, which necessitate large amounts of annotated training data and pose practical challenges in data collection. This paper presents CAT-SAM, a ConditionAI Tuning network that explores few-shot adaptation of SAM toward various challenging downstream domains in a data-efficient manner. The core design is a prompt bridge structure that enables decoder-conditioned joint tuning of the heavyweight image encoder and the lightweight mask decoder. The bridging maps the domain-specific features of the mask decoder to the image encoder, fostering synergic adaptation of both components with mutual benefits with few-shot target samples only, ultimately leading to superior segmentation in various downstream tasks. We develop two CAT-SAM variants that adopt two tuning strategies for the image encoder: one injecting learnable prompt tokens in the input space and the other inserting lightweight adapter networks. Extensive experiments over 11 downstream tasks show that CAT-SAM achieves superior segmentation consistently even under the very challenging one-shot adaptation setup." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.697, + 0.741, + 0.725 + ], + "angle": 0, + "content": "Keywords: Segment Anything Model \\(\\cdot\\) Few-shot Learning \\(\\cdot\\) Parameter Efficient Tuning" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.747, + 0.376, + 0.763 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.775, + 0.787, + 0.806 + ], + "angle": 0, + "content": "Accurate image segmentation is a fundamental computer vision task that plays a pivotal role in various applications such as robotics, autonomous driving, health-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.231, + 0.812, + 0.688, + 0.826 + ], + "angle": 0, + "content": "* Co-first authors with equal contributions. † Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.231, + 0.826, + 0.669, + 0.84 + ], + "angle": 0, + "content": "Project page: https://xiaooran.github.io/projects/CAT-SAM" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.458, + 0.128 + ], + "angle": 0, + "content": "A. Xiao and W. Xuan et al." + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.151, + 0.685, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.358 + ], + "angle": 0, + "content": "Fig. 1: The proposed CAT-SAM performs ConditionAl joint Tuning (CAT) to establish communication between SAM's heavyweight image encoder and lightweight mask decoder. This enables synergistic adaptation of the two network components, mitigating tuning imbalances and improving few-shot SAM adaptation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.788, + 0.476 + ], + "angle": 0, + "content": "care, earth observation, etc. The recently developed Segment Anything Model (SAM) [23], trained with 1.1 billion masks, has emerged as a momentous leap forward in image segmentation. By taking geometric prompts with points, boxes, or masks as input, SAM demonstrates remarkable zero-shot capability for general image segmentation, as well as great potential for fine-grained mask segmentation in different downstream tasks across a variety of contexts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.788, + 0.658 + ], + "angle": 0, + "content": "On the other hand, SAM often fails to generate high-quality predictions when dealing with domains that are either sparsely represented or lie outside its training distribution [5,19], such as challenging RGB images including aerial, medical, and intricate structural images, as well as non-RGB images like X-ray, Sonar, Synthetic Aperture Radar (SAR) images, etc. This greatly undermines the versatility and applicability of SAM as a foundational segmentation model while handling various downstream tasks. Several studies attempted to address this issue by fine-tuning SAM with a large number of target samples with mask annotations [21, 30]. However, they require extensive masked images for each specialized field, posing a formidable task with poor scalability. How to effectively adapt SAM toward various downstream tasks in a data-efficient manner has become an essential challenge in the segmentation field." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.809 + ], + "angle": 0, + "content": "We exploit few-shot target samples for effective and efficient adaptation of SAM toward various downstream tasks. Inspired by the recent advancements in parameter-efficient tuning for foundation models [1,13], we freeze the entire SAM and lightly expand its image encoder and mask decoder with a tiny amount of learnable parameters. This expansion preserves SAM's zero-shot capabilities and flexibility while capturing domain-specific features that are essential for the segmentation of downstream data. However, SAM's image encoder is much larger than its mask decoder, which directly leads to tuning imbalance and further suboptimal adaptation, especially under the limited supervision of few-shot target samples." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To address this challenge, we design CAT-SAM, a novel ConditionAl Tuning network for effective and data-efficient adaptation of SAM. The key idea is to ex" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "CAT-SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "ploit the lightweight decoder tuning to guide the heavyweight encoder tuning and formulate decoder-conditioned joint tuning of both as depicted in Fig. 1. To this end, we design a prompt bridge structure that maps the domain-specific features from the mask decoder to the image encoder. With the collaboration between the two partners, CAT-SAM enables synergistic adaptation of both which mitigates the tuning imbalance and improves the few-shot SAM adaptation effectively. In addition, the prompt bridge can be seamlessly embedded into two representative tuning approaches, including prompt tuning [20, 25, 27] that introduces learnable prompt tokens into the input space and adapters [15, 29, 37, 38] that leverage lightweight adaptive networks. This directly leads to two CAT-SAM variants, namely, CAT-SAM-T and CAT-SAM-A. Extensive evaluations over 11 diverse and challenging downstream tasks show that both variants achieve superior adaptation and segmentation with few-shot target samples only." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.79, + 0.526 + ], + "angle": 0, + "content": "The major contributions of this work can be summarized in three aspects. First, we propose CAT-SAM, a conditional tuning network that enables effective and data-efficient adaptation of SAM toward various challenging downstream domains. We design prompt bridge within CAT-SAM, a decoder-conditioned joint tuning structure that enables synergistic and data-efficient adaptation of the heavyweight image encoder and the lightweight mask decoder effectively. Second, we develop two CAT-SAM variants by embedding the prompt bridge into two representative tuning strategies, one introducing learnable prompt tokens in the input space and the other inserting lightweight adapter networks. Third, extensive experiments over 11 diverse segmentation datasets show that CAT-SAM achieves superior image segmentation consistently even under the challenging one-shot setup." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.55, + 0.39, + 0.567 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Parameter-Efficient Tuning for foundation models has become increasingly crucial due to the over-size model parameters and costly one-task-one-model deployment. One popular approach involves updating only the newly added learnable parameters while keeping the model backbone frozen. These lightweight parameters are often introduced in two ways: 1) as prompt tuning [12,20,22,58] by adding learnable tokens to input tokens at transformer layers, 2) as adapters [11, 29,37,38,49] by integrating learnable lightweight sub-networks into transformer layers. However, prior visual studies primarily focus on adapting single backbone modules [3,20,29] or dual encoder backbones for vision-language tasks [22,57], which are suboptimal tuning solutions while applying to SAM due to its particular architecture of a heavyweight image encoder and a lightweight mask decoder. In contrast, we explore a novel conditional joint tuning strategy tailored to SAM, which facilitates its adaptation even with limited downstream data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "SAM for Downstream Tasks. The release of SAM has spurred several subsequent studies. Some concentrate on expanding semantic recognition capabilities [24, 26, 43, 50, 54], while others aim to create more lightweight SAM variants [48, 52, 55] for faster computations. Several studies explore SAM's adapta" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.458, + 0.128 + ], + "angle": 0, + "content": "A. Xiao and W. Xuan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "tion in underperforming downstream scenarios, primarily by directly fine-tuning SAM's mask decoder [4,8,30,41,46,53], which can severely degrade SAM's zero-shot capabilities. The recent HQ-SAM [21] freezes the entire pre-trained SAM and prompt-tunes the mask decoder with a new mask head for adaptive mask prediction. However, these adaptation studies rely on large target datasets with a considerable number of annotated images, which are often costly to collect. In contrast, our CAT-SAM achieves superior adaptation with only few-shot target samples, significantly reducing adaptation costs and expanding the applicability of SAM as a visual foundation model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.305, + 0.331, + 0.321 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.334, + 0.444, + 0.349 + ], + "angle": 0, + "content": "3.1 Preliminaries of SAM" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.789, + 0.571 + ], + "angle": 0, + "content": "SAM [23] handles each image with three key modules: a heavyweight image encoder (i.e., ViT [9] as the backbone) that extracts image embeddings, a prompt encoder that encodes the geometric prompts (i.e., points, a box, or a coarse mask) to generate prompt embeddings, and a lightweight mask decoder that combines the two types of embeddings to predict segmentation masks. The released SAM model is trained on a super large-scale SA-1B dataset, which consists of over 11 million images as well as 1.1 billion automatically generated masks. SAM has demonstrated remarkable zero-shot capability while dealing with various conventional natural images, as well as superb flexibility in accepting various geometric prompt inputs. On the other hand, it often struggles while facing challenging downstream domains that are often sparse or falling outside of its training data distribution, such as RGB domains of aerial, medical, and intricate structural images, as well as non-RGB domains of X-ray, SAR, and Sonar images. More details about SAM are presented in [23]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.59, + 0.353, + 0.605 + ], + "angle": 0, + "content": "3.2 CAT-SAM" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.689 + ], + "angle": 0, + "content": "This subsection introduces CAT-SAM, a ConditionAl Tuning network that is designed for adapting SAM toward various under-performed downstream data. The objective is to adapt SAM efficiently with only few-shot annotated target images, meanwhile preserving SAM's powerful zero-shot capability and geometric prompting flexibility." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "CAT-SAM freezes the entire SAM and facilitates the simultaneous tuning of SAM's image encoder and mask decoder to capture representative information from few-shot target samples for effective adaptation of SAM. However, a critical issue with simultaneous tuning arises from the clear imbalance between the heavyweight image encoder (308.3 M parameters for ViT-L [9]) and the lightweight mask decoder (4.1 M), which often leads to sub-optimal adaptation while only few-shot target samples are available. To tackle this challenge, we design decoder-conditioned joint tuning, a novel tuning strategy that mitigates the imbalance by establishing a linkage between the encoding tuning and the decoding tuning. Specifically, we design prompt bridge—a lightweight network that" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CAT-SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.149, + 0.382, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.273, + 0.283, + 0.326, + 0.291 + ], + "angle": 0, + "content": "(a) CAT-SAM-T" + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.149, + 0.56, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.283, + 0.504, + 0.291 + ], + "angle": 0, + "content": "(b) CAT-SAM-A" + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.15, + 0.787, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.283, + 0.708, + 0.291 + ], + "angle": 0, + "content": "(c) Prompt Bridge" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.308, + 0.785, + 0.419 + ], + "angle": 0, + "content": "Fig. 2: Overview of CAT-SAM. CAT-SAM keeps the whole SAM frozen while simultaneously tuning the image encoder and mask decoder for downstream adaptation. To address the tuning imbalance between these two network components, we introduce decoder-conditioned joint tuning through the design of Prompt Bridge structure, enabling synergetic and enhanced adaptation. We present two CAT-SAM variants: CAT-SAM-T in (a) and CAT-SAM-A in (b), achieved by integrating the prompt bridge with prompt-based and adapter-based tuning strategies for the image encoder, respectively. (c) illustrates two tailored prompt bridge structures, PB-T and PB-A." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.451, + 0.784, + 0.51 + ], + "angle": 0, + "content": "maps the domain-specific features from the mask decoder to the image encoder. The bridge design facilitates joint and balanced tuning of both network components, leading to synergistic and effective adaptation of SAM with few-shot and even a single target sample." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.784, + 0.556 + ], + "angle": 0, + "content": "We develop two CAT-SAM variants by integrating the proposed decoder-conditioned joint tuning into two prevalent tuning approaches, namely CAT-SAM-T and CAT-SAM-A, as illustrated in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.784, + 0.628 + ], + "angle": 0, + "content": "Tuning Image Encoder. We first brief SAM's image encoder and then introduce two tuning approaches in our CAT-SAM, including prompt tuning in CAT-SAM-T and lightweight adapters in CAT-SAM-A." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.631, + 0.785, + 0.707 + ], + "angle": 0, + "content": "- SAM's Image Encoder, denoted as \\(\\mathcal{V}\\) with \\(K\\) transformer layers \\(\\{\\mathcal{V}_i\\}_{i=1}^K\\), first splits the input image \\(I\\) of the size \\((H, W)\\) into \\(M\\) fixed-size patches, and then projects them into patch embeddings \\(E_0 \\in \\mathbb{R}^{M \\times d_v}\\). Patch embeddings \\(E_{i-1}\\) are input to the \\(i^{th}\\) transformer layer \\((\\mathcal{V}_i)\\) and sequentially processed through \\(K\\) transformer layers:" + }, + { + "type": "equation", + "bbox": [ + 0.4, + 0.721, + 0.629, + 0.739 + ], + "angle": 0, + "content": "\\[\n[ E _ {i} ] = \\mathcal {V} _ {i} ([ E _ {i - 1} ]), i = 1, 2, \\dots , K.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.751, + 0.784, + 0.811 + ], + "angle": 0, + "content": "- CAT-SAM-T: As illustrated in Fig. 2 (a), we introduce a set of learnable tokens \\(\\{P_i \\in \\mathbb{R}^{d_v}\\}_{i=0}^{K-1}\\) in the image encoder alongside the input patch embeddings. These new learnable tokens are introduced in each transformer layer of the image encoder \\(\\mathcal{V}\\) and updated through the adaptation process:" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.824, + 0.651, + 0.842 + ], + "angle": 0, + "content": "\\[\n[ E _ {i} ] = \\mathcal {V} _ {i} ([ E _ {i - 1}, P _ {i - 1} ]) ], i = 1, 2, \\dots , K.\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.458, + 0.128 + ], + "angle": 0, + "content": "A. Xiao and W. Xuan et al." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.146, + 0.789, + 0.374 + ], + "angle": 0, + "content": "- CAT-SAM-A: As illustrated in Fig. 2 (b), we introduce Adapters [15] to tune SAM's image encoder in CAT-SAM-A. The adapter is a lightweight sub-network that is inserted into each transformer layer, comprising a linear down-projection, a nonlinear activation function, a linear up-projection, and a residual connection. High-frequency image information is also incorporated in CAT-SAM as the input of the adapters as in [29]. In this process, high-frequency components \\( I_{hfc} \\) are first extracted from the input image \\( I \\) and then reversed to the space domain using the Fast Fourier Transform and its inverse. These components are then partitioned into small patches \\( I_{hfc}^p \\) similar to the image patches. Subsequently, convolutional layers and linear layers project \\( I_{hfc} \\) and the original image embeddings \\( E_0 \\) into \\( c \\)-dimensional features independently, yielding \\( F_{hfc} \\) and \\( F_{pe} \\), respectively. The result of element-wise addition of \\( F_{pe} \\) and \\( F_{hfc} \\) is then forwarded (as denoted \"EVP\" in Fig. 2 (b)) to the adapter \\( \\mathrm{Adapt}_i \\) for transformer layer \\( i \\). Finally, the output is added element-wise to the original input of the transformer layer \\( i \\):" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.382, + 0.644, + 0.399 + ], + "angle": 0, + "content": "\\[\nE _ {i - 1} = E _ {i - 1} + \\mathrm {A d a p t} _ {i} (F _ {p e} + F _ {h f c}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.789, + 0.597 + ], + "angle": 0, + "content": "Tuning Mask Decoder. We tune the mask decoder of SAM similarly as [21]. Specifically, we freeze the entire pre-trained decoder of SAM and introduce a learnable CAT-Token that is concatenated with SAM's original output tokens and prompt tokens to form the input to SAM's mask decoder. The CAT-Token is updated during training but remains fixed during inference, the same as SAM's output tokens. After passing through two decoder layers, the updated CAT-Token is used to generate dynamic weights for a newly introduced three-layer MLP. Simultaneously, SAM's mask decoder features are fused with the image features decoded by the image encoder. Finally, the fused features and the three-layer MLP are combined through the dot product to generate the target mask. Please refer to the appendix for more details about the decoder structures in CAT-SAM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.617, + 0.791, + 0.661 + ], + "angle": 0, + "content": "Decoder-Conditioned Joint Tuning. The proposed decoder-conditioned joint tuning can be seamlessly embedded via two implementations of the proposed prompt bridging ideas, i.e., PB-T and PB-A as illustrated in Fig. 2 (c)." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.662, + 0.788, + 0.737 + ], + "angle": 0, + "content": "- CAT-SAM-T: The prompt bridge \\(\\mathrm{PB}_i^T\\) is a two-layer MLP. It projects the CAT-Token \\(Q\\) in the mask decoder to each transformer layer \\(i\\) in the image encoder and outputs a single learnable token replacing \\(P_i\\) by \\(\\tilde{P}_i^T = \\mathrm{PB}_i^T(Q)\\). These mapped tokens are directly applied for prompt tuning of the image encoder as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.38, + 0.744, + 0.652, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\left[ E _ {i} \\right] = \\mathcal {V} _ {i} \\left(\\left[ E _ {i - 1}, \\tilde {P} _ {i - 1} ^ {T} \\right]\\right), i = 1, 2, \\dots , K.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.771, + 0.788, + 0.817 + ], + "angle": 0, + "content": "- CAT-SAM-A: The prompt bridge \\(\\mathrm{PB}^A\\) maps the CAT-Token \\(Q\\) into an embedding \\(\\tilde{P}^A = \\mathrm{PB}^A(Q)\\) of the same size as \\(F_{pe}\\) and \\(F_{hfc}\\). \\(\\tilde{P}^A\\) is then element-wise added to the EVP before being fed to every adapter as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.823, + 0.667, + 0.843 + ], + "angle": 0, + "content": "\\[\nE _ {i - 1} = E _ {i - 1} + \\mathrm {A d a p t} _ {i} (F _ {p e} + F _ {h f c} + \\tilde {P} ^ {A}).\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "CAT-SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.144, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 1: Comparisons of trainable parameters between ViT-L [9] based SAM and two variants of the proposed CAT-SAM." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.185, + 0.778, + 0.259 + ], + "angle": 0, + "content": "
ModelTrainable Parameters (M)Trainable Params vs. SAM
SAM [23]312.4-
CAT-SAM-T3.31.1%
CAT-SAM-A1.90.6%
" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.286, + 0.786, + 0.346 + ], + "angle": 0, + "content": "As illustrated in Fig. 2 (c), PB-A employs \\( c \\) different linear down-projection layers to generate embeddings independently. After GELU layers, these embeddings are up-projected to the dimension of \\( 1 \\times M \\) (\\( M \\) is the spatial size of the patch) with a shared linear layer to mitigate the computation cost." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.346, + 0.753, + 0.362 + ], + "angle": 0, + "content": "Finally, the up-projected embeddings are concatenated to produce \\(\\tilde{P}^A\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.362, + 0.787, + 0.437 + ], + "angle": 0, + "content": "Tab. 1 compares the amount of trainable parameters between the SAM (implemented with the ViT-L [9]) and our CAT-SAM that is built upon the SAM. It can be observed that both CAT-SAM variants introduce very limited additional parameters, yet yield significant performance improvements in various downstream segmentation tasks as to be detailed in Section 4." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.459, + 0.452, + 0.475 + ], + "angle": 0, + "content": "3.3 Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.484, + 0.789, + 0.635 + ], + "angle": 0, + "content": "During training, we feed target samples together with geometric prompts into CAT-SAM to generate CAT-SAM masks, under the supervision of ground-truth masks as well as a linear combination of BCE loss and dice loss [31]. We freeze the pre-trained SAM parameters and update solely the parameters that are introduced in the above-described tuning modules. Note we train CAT-SAM with a mixture of sampled geometric prompts, including bounding boxes, randomly selected points, and coarse masks. In addition, we create degraded coarse masks by introducing random Gaussian noise in the boundary regions of the ground-truth masks as in [21]. Further training details can be found in Section 4.1 and the appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.635, + 0.789, + 0.743 + ], + "angle": 0, + "content": "During the inference phase, the input image is fed into SAM's image encoder along with the integrated prompt tokens in CAT-SAM-T or adapters in CAT-SAM-A, to generate adaptive image embeddings. These embeddings, combined with the prompt tokens from SAM's prompt encoder, serve as input to the mask decoder. Subsequently, the updated CAT-token and associated MLP are utilized for the target mask prediction. Lastly, we up-sample the mask to the original resolution to produce the final output." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.764, + 0.377, + 0.781 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We evaluate CAT-SAM's efficacy through comprehensive experiments involving eight segmentation tasks across 11 datasets, all hailing from challenging downstream fields that SAM struggles to address effectively. Our experiments encom" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.459, + 0.128 + ], + "angle": 0, + "content": "A. Xiao and W. Xuan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 2: We evaluate and benchmark CAT-SAM over eight challenging segmentation tasks across 11 datasets." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.186, + 0.812, + 0.342 + ], + "angle": 0, + "content": "
TasksDataset NamesImagery
Building segmentationWHU [18]Aerial images
Road segmentationMA. Roads [32]Aerial images
Polyp segmentationKvasir [34]Medical images
Chest X-ray segmentationJSRT [39]X-ray images
Marine debris segmentationFLS [40]Sonar images
Ship instance segmentationHRSID [45]SAR images
Shadow segmentationSBU-Shadow [42]Natural images
Intricate segmentationDIS [35],ThinObject [28],HRSOD [51],COIFT [28]Natural images
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.368, + 0.785, + 0.399 + ], + "angle": 0, + "content": "pass a wide array of scenarios and tasks, spanning from varying target shots to different imagery modalities." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.417, + 0.428, + 0.433 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.795, + 0.573 + ], + "angle": 0, + "content": "Datasets. We conduct experiments on two collections of datasets listed in Tab. 2: 1) True-color images, including WHU [18] for building segmentation, Kvasir [34] for polyp segmentation, SBU-Shadow [42] for shadow segmentation, Massachusetts (MA.) Roads [32] for road segmentation, as well as fine-grained segmentation datasets DIS [35], ThinObject [28], HRSOD [51], and COIFT [28]. 2) Non-RGB domains, encompassing JSRT [39] for chest organ segmentation (X-ray images), HRSID [45] for ship instance segmentation (high-resolution SAR images), and FLS [40] for marine debris segmentation (Sonar images). We use official dataset splits for fair comparisons." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.574, + 0.787, + 0.664 + ], + "angle": 0, + "content": "Evaluation Metrics. For single-class datasets including WHU, Kvasir, SBU-Shadow, and MA. Roads, we employ standard mask IoU of foreground. JSRT and FLS, which have multiple classes, are evaluated using individual class IoU and their average. As for DIS, ThinObject, HRSOD, and COIFT, we follow [21] and report mIoU and the boundary metrics (mBIoU) [7] for fair comparisons. For instance segmentation of HRSID, we use standard AP, \\(\\mathrm{AP}_{50}\\), and \\(\\mathrm{AP}_{75}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.665, + 0.787, + 0.772 + ], + "angle": 0, + "content": "Implementation Details. Unless otherwise specified, CAT-SAM and its comparing models use ViT-L [9] as the image encoder backbone. We train on 1 NVIDIA RTX A6000 GPU for one-shot adaptation, while 4 GPUs for multiple-shots. Following [21], ground truth boxes serve as the default input geometric prompts during evaluation to ensure fair comparison and minimize randomness, with the exception of Fig. 3 where point prompts are evaluated. Please refer to the appendix for more implementation details." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.789, + 0.402, + 0.803 + ], + "angle": 0, + "content": "4.2 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "We first investigate the contributions of different tuning modules within both CAT-SAM variants to assess their impact on overall adaptation performance." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CAT-SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.243 + ], + "angle": 0, + "content": "Table 3: Ablation study of CAT-SAM-T in (a) and CAT-SAM-A in (b) for 1-shot adaptation over datasets: \"W\" for WHU [18] (on building segmentation), \"K\" for Kvasir [34] (on polyp segmentation), and \"S\" for SBU-Shadow [42] (on shadow segmentation). \"Enc-T.\" and \"Enc-A.\" means tuning the image encoder with prompt tokens and adapters, respectively. \"Dec.\" means tuning the mask decoder. \"PB-T\" and \"PB-A\" denote two customized prompt bridges for conditional tuning. Model 1 is the baseline SAM without adaptation." + }, + { + "type": "table_caption", + "bbox": [ + 0.31, + 0.256, + 0.425, + 0.269 + ], + "angle": 0, + "content": "(a) CAT-SAM-T." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.27, + 0.508, + 0.375 + ], + "angle": 0, + "content": "
Enc-T. Dec. PB-TWKSAVG
143.579.062.461.6
2✓57.880.476.0
3✓71.273.863.5
4✓✓66.473.261.8
5✓✓✓86.883.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.594, + 0.256, + 0.71, + 0.269 + ], + "angle": 0, + "content": "(b) CAT-SAM-A." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.27, + 0.793, + 0.375 + ], + "angle": 0, + "content": "
Enc-A. Dec. PB-AWKSAVG
143.579.062.461.6
2✓86.573.776.3
3✓71.273.863.5
4✓✓83.479.876.3
5✓✓✓88.285.4
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.788, + 0.526 + ], + "angle": 0, + "content": "Tab. 3 shows the experiments of one-shot adaptation across three true-color image segmentation benchmarks WHU [18], Kvasir [34], and SBU-Shadow [42]. For each CAM-SAM variant CAM-SAM-T/CAM-SAM-A, we compare 5 models including: 1. The original SAM without adaptation (baseline); 2. Tuning SAM's image encoder alone with prompt tokens or adapters; 3. Tuning SAM's mask decoder alone; 4. Tuning SAM's image encoder and mask decoder independently; and 5. Conditional tuning of SAM's image encoder and mask decoder with the proposed prompt bridging (i.e., the complete CAT-SAM-T and CAT-SAM-A)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.709 + ], + "angle": 0, + "content": "We can see that the original SAM struggles to produce high-quality masks in these challenging domains, indicating its limitations as a foundational segmentation model for many downstream applications. Tuning the image encoder or the mask decoder alone in 2 and 3 enhances target segmentation, underlining the efficacy and necessity of SAM adaptation. However, simultaneously tuning the image encoder and mask decoder in an independent way in model 4 does not show complementary and consistent improvement. This suggests instability and imbalance in tuning between the heavyweight image encoder and the lightweight mask decoder, which could lead to suboptimal adaptation solutions. Differently, including the proposed prompt bridging with either PB-T or PB-A (i.e., the complete CAT-SAM-T and CAT-SAM-A) mitigates the tuning imbalance and produces substantial performance enhancement consistently." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.735, + 0.577, + 0.751 + ], + "angle": 0, + "content": "4.3 Comparison with the State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "CAT-SAM for One-Shot Adaptation. To assess CAT-SAM's effectiveness in reducing downstream training data and its adaptability with few-shot samples, we first evaluate it under the extremely challenging one-shot adaptation scenario. Given the lack of prior studies under such setup, we comprehensively benchmark it against closely related methods that can be broadly categorized into" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.458, + 0.128 + ], + "angle": 0, + "content": "A. Xiao and W. Xuan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 4: Comparison of adaptive segmentation performance on challenging true-color datasets: WHU [18] (on building segmentation), Kvasir [34] (on Polyp segmentation), and SBU-Shadow [42] (on shadow segmentation). The baseline is SAM without adaptation. All compared methods utilize one-shot sample for adaptation, except for those denoted by * which utilize 20 shots with metric numbers sourced from [47]." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.227, + 0.787, + 0.458 + ], + "angle": 0, + "content": "
MethodsWHUKvasirSBU-ShadowAverage
SAM [23] (baseline)43.579.062.461.6
VPT-shallow [20]60.879.868.769.8
VPT-deep [20]57.880.476.071.4
AdaptFormer [3]83.276.877.279.1
LoRA [16]86.177.574.479.3
Med-SA [46]34.528.622.128.4
MedSAM [30]30.629.853.437.9
SimAda [41]48.618.749.338.9
SPFS-SAM* [47] (20-shots)-53.4--
SAMed* [53] (20-shots)-51.7--
HQ-SAM [21]71.273.863.569.5
CAT-SAM-T (Ours)86.883.478.082.7
CAT-SAM-A (Ours)88.285.481.985.2
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.501, + 0.787, + 0.592 + ], + "angle": 0, + "content": "two groups. The first group comprises state-of-the-art tuning methods for foundation models of different modalities, including VPT [20] and AdaptFormer [3] for large vision foundation models, and LoRA [16] for large language models. The second group comprises recent SAM-based adaptation methods which were largely designed for full-shot adaptation [21,30,41,46,53] ( [47] tackles few-shot adaptation but requires 20 shots)." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.599, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Tab. 4 presents results over WHU, Kvasir, and SBU-Shadow datasets. We can observe that all three tuning methods for foundation models exhibit clear adaptation effects. Notably, AdaptFormer and LoRA demonstrate significantly better performance than VPT-deep and VPT-shallow, highlighting the effectiveness of modifying the model's structure over prompt tuning in the input space during SAM adaptation. On the other hand, most SAM-based adaptation methods, including Med-SA, SimAda, SPFS-SAM, SAMed, and MedSAM, yield much lower performance, primarily due to their reliance on many more target images which leads to overfitting while learning from one-shot target sample. HQ-SAM achieves substantial segmentation improvements but it still cannot compete with the three tuning methods, demonstrating the significance of tuning the image encoder over the mask decoder. Finally, both CAT-SAM variants consistently outperform all compared methods, with CAT-SAM-A exhibiting a slight edge due to its network structure modifications. These experiments validate the superiority of the proposed decoder-conditioned joint tuning approach in extracting domain-specific features with few-shot samples." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CAT-SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 5: Comparison of state-of-the-art segmentation achieved by fully supervised learning (FSL) methods trained on full-shot data (upper part) and mask segmentation by SAM and our CAT-SAM (lower part). Evaluation follows official criteria of each FSL benchmark, reporting IoU for WHU, mIoU for Kvasir, and BER for SBU-Shadow to ensure consistency. \"(1)\", \"(16)\", and \"(full)\" after CAT-SAM-T/CAT-SAM-A denote tuning with 1-shot, 16-shots, and full-shots samples." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.24, + 0.794, + 0.439 + ], + "angle": 0, + "content": "
WHUKvasirSBU-Shadow
ModelIoUModelmIoUModelBER(↓)
STT [2]90.5CASCADE [36]87.8BDRAR [59]3.64
BuildFormer [44]91.4PatchRefineNet [33]89.1DSDNet [56]3.45
CBR-Net [14]91.4DUCK-Net [10]90.5MTMT [6]3.15
SAM43.5SAM87.3SAM13.22
CAT-SAM-T (1)86.8CAT-SAM-T (1)90.0CAT-SAM-T (1)7.81
CAT-SAM-A (1)88.2CAT-SAM-A (1)91.3CAT-SAM-A (1)5.27
CAT-SAM-T (16)89.6CAT-SAM-T (16)93.1CAT-SAM-T (16)4.04
CAT-SAM-A (16)90.3CAT-SAM-A (16)93.6CAT-SAM-A (16)3.80
CAT-SAM-T (full)93.3CAT-SAM-T (full)94.5CAT-SAM-T (full)2.54
CAT-SAM-A (full)93.6CAT-SAM-A (full)94.3CAT-SAM-A (full)2.39
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.79, + 0.653 + ], + "angle": 0, + "content": "Full-Shot Learning. While CAT-SAM is primarily designed for few-shot adaptation, it also demonstrates proficiency in full-shot adaptation scenarios. In Tab. 5, we present the state-of-the-art performances of traditional fully-supervised learning (FSL) methods trained over full-shot data and compare them with SAM and our CAT-SAM (adapted with various target samples). We report official evaluation metrics for three public benchmarks, including IoU for WHU, mIoU for Kvasir, and BER [17] for SBU-Shadow. Note the distinction in setups between the FSL segmentation and the mask segmentation: FSL focuses on traditional semantic segmentation of specific classes without prompt input while SAM and CAT-SAM involve general mask segmentation with prompt input. Despite these differences, these performances can serve as valuable references for evaluating CAT-SAM's segmentation ability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.656, + 0.788, + 0.793 + ], + "angle": 0, + "content": "With full target training data, both CAT-SAM variants outperform state-of-the-art FSL methods consistently. Notably, the two CAT-SAM variants achieve impressive target segmentation under the challenging one-shot setup with a single target sample. With 16 shots of target samples, their performance is even on par with the FSL methods that are trained by using full target training data. These experiments underscore the superiority of effective transfer of broad knowledge from the foundation model SAM as compared with training from scratch with FSL. It also demonstrates the remarkable efficacy of CAT-SAM in substantially expanding SAM's applicability across various downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We conducted further experiments to compare CAT-SAM against HQ-SAM [21], a state-of-the-art SAM-based adaptation solution for intricate structural image segmentation. For fair comparisons, we follow the evaluation protocol of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.458, + 0.128 + ], + "angle": 0, + "content": "A. Xiao and W. Xuan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.2 + ], + "angle": 0, + "content": "Table 6: Extremely fine-grained segmentation over four intricate structural image datasets. HQ-SAM and the two CAT-SAM variants are fine-tuned over full HQSeg-44K [21]. All adopt the boxes converted from their GT masks as the box prompt input for fair comparisons." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.213, + 0.787, + 0.314 + ], + "angle": 0, + "content": "
ModelsDIS [35]COIFT [28]HRSOD [51]ThinObject [28]Average
mIoUmBIoUmIoUmBIoUmIoUmBIoUmIoUmBIoUmIoUmBIoU
SAM [23]62.052.892.186.590.283.173.661.879.571.1
HQ-SAM [21]78.670.494.890.193.686.989.579.989.181.8
CAT-SAM-T84.078.195.692.093.487.694.087.991.786.4
CAT-SAM-A83.677.795.692.293.187.194.188.291.686.3
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.341, + 0.785, + 0.478 + ], + "angle": 0, + "content": "HQ-SAM and tune CAT-SAM on the complete HQSeg-44K dataset, consisting of 44,320 meticulously annotated image masks. Tab. 6 shows experiments across four exceptionally fine-grained segmentation datasets including DIS [35], ThinObject [28], HRSOD [51], and COIFT [28] (in mIoU and the boundary metrics (mBIoU) [7]). We can see that CAT-SAM-T achieves slightly better performance than CAT-SAM-A, while both clearly outperform HQ-SAM and the original SAM. These experiments further demonstrate the efficacy of our proposed conditional tuning in adapting SAM toward challenging downstream tasks. Visual qualitative comparisons are provided in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.498, + 0.546, + 0.513 + ], + "angle": 0, + "content": "4.4 CAT-SAM for Non-RGB Domains" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.688 + ], + "angle": 0, + "content": "We also explore CAT-SAM's performance across more challenging image modalities characterized by larger domain discrepancies. Specifically, we assess its efficacy on non-RGB domains, including JSRT [39] for chest organ segmentation using X-ray images and FLS [40] for marine debris segmentation with Sonar images, and HRSID [45] for ship instance segmentation with SAR images. For adaptation, we randomly select either one or sixteen target images for 1-shot and 16-shot scenarios, respectively, while employing the entire training set for full-shot adaptation. These datasets represent novel image distributions as compared with SAM's training data. Note we train instances of all classes together while evaluating them by classes. Since no images in FLS contain objects of all 11 classes, we only report results of 16-shot and full-shot adaptation for FLS." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.689, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Tab. 7 shows experimental results. Notably, CAT-SAM-T and CAT-SAM-A improve the mIoU to \\(93.0\\%\\) and \\(92.6\\%\\), respectively, with only one shot of JSRT training image. The improvements saturate gradually as the number of training samples increases, with marginal gains of \\(1.4\\%\\) and \\(2.0\\%\\), respectively, when adapted with full-shot target data. These findings underscore the superior few-shot capability of the proposed CAT-SAM and align well with those reported in Tab. 5. Experiments on FLS [40] show different results. While both CAT-SAM variants exhibit clear adaptation effects, the performance gains for both 16-shot and full-shot adaptation are notably lower as compared with previously evaluated datasets. The discrepancy is largely attributed to two primary" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CAT-SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 7: Few-shot adaptation with varying number of training samples for non-RGB datasets, including FLS [40] with Sonar images for marine debris segmentation, JSRT [39] for chest X-ray segmentation, and HRSID [45] for ship instance segmentation with SAR images. \"N/A\" denotes that 1-shot adaptation over FLS is unattainable due to the absence of images containing objects from all 11 classes." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.227, + 0.777, + 0.394 + ], + "angle": 0, + "content": "
#SamplesMethodsJSRT [39]FLS [40]HRSID [45]
LungsHeartmIoUmIoUAPAP50AP75
NoneSAM [23]85.071.978.569.738.286.626.8
1-shotCAT-SAM-T95.490.593.0N/A46.088.445.4
CAT-SAM-A94.790.592.6N/A44.988.041.4
16-shotCAT-SAM-T95.892.794.273.246.289.345.2
CAT-SAM-A95.591.593.571.445.789.443.7
Full-shotCAT-SAM-T96.392.694.481.751.493.055.0
CAT-SAM-A96.492.894.682.052.993.956.0
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.575 + ], + "angle": 0, + "content": "factors: 1) the inherent difficulty of FLS with its 11 semantic classes (per-class performances are provided in appendix); and 2) the larger variation of FLS data in shapes and object sizes as compared with previously evaluated data. As for ship instance segmentation using SAR images in HRSID [45], this dataset presents different challenges as ships are very small in size. The results show that CAT-SAM consistently outperforms SAM in this diverse domain, highlighting its efficiency in domain adaptation with large domain discrepancy. These experiments further underscore the need for a more comprehensive adaptation of SAM for exceptionally challenging downstream tasks. We expect further studies along this research direction." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.6, + 0.445, + 0.616 + ], + "angle": 0, + "content": "4.5 Analysis & Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.546, + 0.811 + ], + "angle": 0, + "content": "Prompts with Single Point. SAM provides flexibility in geometric prompts including a single foreground point. However, prompting with a single foreground point can be challenging as the single point could correspond to multiple objects [23]. Nonetheless, adaptation with few-shot target samples can alleviate this ambiguity by guiding the model to focus on the specific foreground distribution within the annotation space. We evaluate this by using the same randomly-selected single point as the geometric prompt, and Fig. 3 shows" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "the segmentation results for images from WHU, Kvasir, SBU-Shadow, and MA. Roads [32] on road segmentation." + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.661, + 0.784, + 0.761 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.771, + 0.785, + 0.799 + ], + "angle": 0, + "content": "Fig. 3: Single point to mask evaluation for one-shot adaptation." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.458, + 0.128 + ], + "angle": 0, + "content": "A. Xiao and W. Xuan et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.148, + 0.784, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.4 + ], + "angle": 0, + "content": "Fig. 4: Visual comparisons of SAM [23] (top row) and CAT-SAM (bottom row). We illustrate samples from WHU [18] for building segmentation, MA. Roads [32] for road segmentation, SBU-Shadow [42] for shadow segmentation, Kvasir [34] for polyp segmentation, JSRT [39] for chest organ segmentation (X-ray images), FLS [40] for marine debris segmentation (Sonar images), and HRSID [45] for ship instance segmentation (SAR images). CAT-SAM exhibits one-shot adaptation across most datasets, except for 16-shot over FLS. Red boxes and stars denote geometric prompts, colored regions are mask predictions, and lines show the boundary of ground truth segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.55 + ], + "angle": 0, + "content": "Similar to the experiments using boxes as prompts, CAT-SAM consistently demonstrates superior target segmentation while using a single point as prompt, underscoring its robustness in adaptation while preserving SAM's inherent geometric prompting flexibility. In addition, both CAT-SAM variants achieve relatively less improvement for the dataset Kvasir. This can be largely attributed to the high similarities in color and texture between the polyps and the local gastrointestinal regions in Kvasir, making it challenging for a single point to provide clear geometric guidance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.627 + ], + "angle": 0, + "content": "Visual Illustrations. Fig. 4 shows qualitative comparisons between SAM and CAT-SAM across multiple segmentation tasks. These illustrations demonstrate how our proposed CAT-SAM remarkably enhances the mask quality of SAM under the presence of only one-shot target samples. Please refer to the appendix for more visual comparisons." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.654, + 0.36, + 0.67 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We introduce CAT-SAM, a conditional tuning network tailored for few-shot adaptation of SAM. We propose decoder-conditional joint tuning to mitigate the tuning imbalance between SAM's heavyweight image encoder and lightweight mask decoder and facilitate efficient SAM adaptation. To achieve this, we design the prompt bridge structures, enabling interaction when tuning these two network modules. We develop two CAT-SAM variants by integrating the prompt bridge with two representative tuning strategies including prompt tuning and adapter. Our comprehensive evaluation across diverse scenarios and tasks on 11 segmentation datasets underscores the superior domain adaptive efficiency of both CAT-SAM variants, even in the extremely challenging one-shot scenario." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CAT-SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.253 + ], + "angle": 0, + "content": "This project is partly funded by the Ministry of Education Singapore, under the Tier-1 scheme with project number RG18/22 and the Tier-2 scheme with project number MOE-T2EP20123-0003. It is also supported in part by the National Natural Science Foundation of China [grant no. U2033210], in part by the Zhejiang Provincial Natural Science Foundation [grant no. LDT23F02024F02]." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.274, + 0.323, + 0.289 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.303, + 0.785, + 0.345 + ], + "angle": 0, + "content": "1. Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., et al.: On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.346, + 0.785, + 0.373 + ], + "angle": 0, + "content": "2. Chen, K., Zou, Z., Shi, Z.: Building extraction from remote sensing images with sparse token transformers. Remote Sensing 13(21), 4441 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.373, + 0.785, + 0.413 + ], + "angle": 0, + "content": "3. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.414, + 0.785, + 0.469 + ], + "angle": 0, + "content": "4. Chen, T., Zhu, L., Deng, C., Cao, R., Wang, Y., Zhang, S., Li, Z., Sun, L., Zang, Y., Mao, P.: Sam-adapter: Adapting segment anything in underperformed scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 3367-3375 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.469, + 0.785, + 0.524 + ], + "angle": 0, + "content": "5. Chen, T., Zhu, L., Ding, C., Cao, R., Zhang, S., Wang, Y., Li, Z., Sun, L., Mao, P., Zang, Y.: Sam fails to segment anything?—sam-adapter: Adapting sam in underperformed scenes: Camouflage, shadow, and more. arXiv preprint arXiv:2304.09148 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.525, + 0.785, + 0.566 + ], + "angle": 0, + "content": "6. Chen, Z., Zhu, L., Wan, L., Wang, S., Feng, W., Heng, P.A.: A multi-task mean teacher for semi-supervised shadow detection. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 5611-5620 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.566, + 0.785, + 0.62 + ], + "angle": 0, + "content": "7. Cheng, B., Girshick, R., Dollar, P., Berg, A.C., Kirillov, A.: Boundary iou: Improving object-centric image segmentation evaluation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15334-15342 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.621, + 0.785, + 0.648 + ], + "angle": 0, + "content": "8. Cheng, J., Ye, J., Deng, Z., Chen, J., Li, T., Wang, H., Su, Y., Huang, Z., Chen, J., Jiang, L., et al.: Sam-med2d. arXiv preprint arXiv:2308.16184 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.649, + 0.785, + 0.703 + ], + "angle": 0, + "content": "9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.704, + 0.785, + 0.731 + ], + "angle": 0, + "content": "0. Dumitru, R.G., Peteleaza, D., Craciun, C.: Using duck-net for polyp image segmentation. Scientific Reports 13(1), 9803 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.785, + 0.772 + ], + "angle": 0, + "content": "1. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision pp. 1-15 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "2. Gong, Z., Li, F., Deng, Y., Bhattacharjee, D., Zhu, X., Ji, Z.: Coda: Instructive chain-of-domain adaptation with severity-aware visual prompt tuning (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "3. Gu, J., Han, Z., Chen, S., Beirami, A., He, B., Zhang, G., Liao, R., Qin, Y., Tresp, V., Torr, P.: A systematic survey of prompt engineering on vision-language foundation models. arXiv preprint arXiv:2307.12980 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.303, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.458, + 0.128 + ], + "angle": 0, + "content": "A. Xiao and W. Xuan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "14. Guo, H., Du, B., Zhang, L., Su, X.: A coarse-to-fine boundary refinement network for building footprint extraction from remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing 183, 240-252 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "15. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.787, + 0.272 + ], + "angle": 0, + "content": "16. Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "17. Hu, X., Zhu, L., Fu, C.W., Qin, J., Heng, P.A.: Direction-aware spatial context features for shadow detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7454-7462 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "18. Ji, S., Wei, S., Lu, M.: Fully convolutional networks for multisource building extraction from an open aerial and satellite imagery data set. IEEE Transactions on geoscience and remote sensing 57(1), 574-586 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.356, + 0.787, + 0.397 + ], + "angle": 0, + "content": "19. Ji, W., Li, J., Bi, Q., Li, W., Cheng, L.: Segment anything is not always perfect: An investigation of sam on different real-world applications. arXiv preprint arXiv:2304.05750 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.438 + ], + "angle": 0, + "content": "20. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.466 + ], + "angle": 0, + "content": "21. Ke, L., Ye, M., Danelljan, M., Liu, Y., Tai, Y.W., Tang, C.K., Yu, F.: Segment anything in high quality. arXiv preprint arXiv:2306.01567 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.508 + ], + "angle": 0, + "content": "22. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19113-19122 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.787, + 0.55 + ], + "angle": 0, + "content": "23. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.591 + ], + "angle": 0, + "content": "24. Lan, X., Lyu, J., Jiang, H., Dong, K., Niu, Z., Zhang, Y., Xue, J.: Foodsam: Any food segmentation. IEEE Transactions on Multimedia pp. 1-14 (2023). https://doi.org/10.1109/TMM.2023.3330047" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.787, + 0.619 + ], + "angle": 0, + "content": "25. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.787, + 0.66 + ], + "angle": 0, + "content": "26. Li, F., Zhang, H., Sun, P., Zou, X., Liu, S., Yang, J., Li, C., Zhang, L., Gao, J.: Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.787, + 0.688 + ], + "angle": 0, + "content": "27. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.73 + ], + "angle": 0, + "content": "28. Liew, J.H., Cohen, S., Price, B., Mai, L., Feng, J.: Deep interactive thin object selection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 305-314 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.787, + 0.771 + ], + "angle": 0, + "content": "29. Liu, W., Shen, X., Pun, C.M., Cun, X.: Explicit visual prompting for low-level structure segmentations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19434-19445 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.787, + 0.799 + ], + "angle": 0, + "content": "30. Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature Communications 15(1), 654 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "31. Miletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks for volumetric medical image segmentation. In: 2016 fourth international conference on 3D vision (3DV). pp. 565-571. IEEE (2016)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CAT-SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.176 + ], + "angle": 0, + "content": "32. Mnih, V.: Machine Learning for Aerial Image Labeling. Ph.D. thesis, University of Toronto (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.232 + ], + "angle": 0, + "content": "33. Nagendra, S., Kifer, D.: Patchrefinenet: Improving binary segmentation by incorporating signals from optimal patch-wise binarization. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1361-1372 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.233, + 0.788, + 0.302 + ], + "angle": 0, + "content": "34. Pogorelov, K., Randel, K.R., Griwodz, C., Eskeland, S.L., de Lange, T., Johansen, D., Spampinato, C., Dang-Nguyen, D.T., Lux, M., Schmidt, P.T., et al.: Kvasir: A multi-class image dataset for computer aided gastrointestinal disease detection. In: Proceedings of the 8th ACM on Multimedia Systems Conference. pp. 164-169 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.303, + 0.788, + 0.346 + ], + "angle": 0, + "content": "35. Qin, X., Dai, H., Hu, X., Fan, D.P., Shao, L., Van Gool, L.: Highly accurate dichotomous image segmentation. In: European Conference on Computer Vision. pp. 38-56. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.389 + ], + "angle": 0, + "content": "36. Rahman, M.M., Marculescu, R.: Medical image segmentation via cascaded attention decoding. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6222-6231 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.389, + 0.788, + 0.416 + ], + "angle": 0, + "content": "37. Rebuffi, S.A., Bilen, H., Vedaldi, A.: Learning multiple visual domains with residual adapters. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.417, + 0.788, + 0.46 + ], + "angle": 0, + "content": "38. Rebuffi, S.A., Bilen, H., Vedaldi, A.: Efficient parametrization of multi-domain deep neural networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8119-8127 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.46, + 0.788, + 0.529 + ], + "angle": 0, + "content": "39. Shiraishi, J., Katsuragawa, S., Ikezoe, J., Matsumoto, T., Kobayashi, T., Komatsu, K.i., Matsui, M., Fujita, H., Kodera, Y., Doi, K.: Development of a digital image database for chest radiographs with and without a lung nodule: receiver operating characteristic analysis of radiologists' detection of pulmonary nodules. American Journal of Roentgenology 174(1), 71-74 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.53, + 0.788, + 0.572 + ], + "angle": 0, + "content": "40. Singh, D., Valdenegro-Toro, M.: The marine debris dataset for forward-looking sonar semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3741-3749 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.572, + 0.788, + 0.614 + ], + "angle": 0, + "content": "41. Song, Y., Zhou, Q., Lu, X., Shao, Z., Ma, L.: Simada: A simple unified framework for adapting segment anything model in underperformed scenes. arXiv preprint arXiv:2401.17803 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.615, + 0.788, + 0.671 + ], + "angle": 0, + "content": "42. Vicente, T.F.Y., Hou, L., Yu, C.P., Hoai, M., Samaras, D.: Large-scale training of shadow detectors with noisily-annotated shadow examples. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI 14. pp. 816-832. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.671, + 0.788, + 0.727 + ], + "angle": 0, + "content": "43. Wang, H., Vasu, P.K.A., Faghri, F., Vemulapalli, R., Farajtabar, M., Mehta, S., Rastegari, M., Tuzel, O., Pouransari, H.: Sam-clip: Merging vision foundation models towards semantic and spatial understanding. arXiv preprint arXiv:2310.15308 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.728, + 0.788, + 0.755 + ], + "angle": 0, + "content": "44. Wang, L., Fang, S., Meng, X., Li, R.: Building extraction with vision transformer. IEEE Transactions on Geoscience and Remote Sensing 60, 1-11 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.757, + 0.788, + 0.799 + ], + "angle": 0, + "content": "45. Wei, S., Zeng, X., Qu, Q., Wang, M., Su, H., Shi, J.: Hrsid: A high-resolution sar images dataset for ship detection and instance segmentation. IEEE Access 8, 120234-120254 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.842 + ], + "angle": 0, + "content": "46. Wu, J., Fu, R., Fang, H., Liu, Y., Wang, Z., Xu, Y., Jin, Y., Arbel, T.: Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.458, + 0.128 + ], + "angle": 0, + "content": "A. Xiao and W. Xuan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "47. Wu, Q., Zhang, Y., Elbatel, M.: Self-prompting large vision models for few-shot medical image segmentation. In: MICCAI Workshop on Domain Adaptation and Representation Transfer. pp. 156-167. Springer (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "48. Xiong, Y., Varadarajan, B., Wu, L., Xiang, X., Xiao, F., Zhu, C., Dai, X., Wang, D., Sun, F., Iandola, F., et al.: Efficientsam: Leveraged masked image pretraining for efficient segment anything. arXiv preprint arXiv:2312.00863 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "49. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2945-2954 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "50. Yang, H., Ma, C., Wen, B., Jiang, Y., Yuan, Z., Zhu, X.: Recognize any regions. arXiv preprint arXiv:2311.01373 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "51. Zeng, Y., Zhang, P., Zhang, J., Lin, Z., Lu, H.: Towards high-resolution salient object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7234-7243 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.383 + ], + "angle": 0, + "content": "52. Zhang, C., Han, D., Qiao, Y., Kim, J.U., Bae, S.H., Lee, S., Hong, C.S.: Faster segment anything: Towards lightweight sam for mobile applications. arXiv preprint arXiv:2306.14289 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.785, + 0.411 + ], + "angle": 0, + "content": "53. Zhang, K., Liu, D.: Customized segment anything model for medical image segmentation. arXiv preprint arXiv:2304.13785 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.439 + ], + "angle": 0, + "content": "54. Zhang, R., Jiang, Z., Guo, Z., Yan, S., Pan, J., Dong, H., Gao, P., Li, H.: Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.467 + ], + "angle": 0, + "content": "55. Zhao, X., Ding, W., An, Y., Du, Y., Yu, T., Li, M., Tang, M., Wang, J.: Fast segment anything. arXiv preprint arXiv:2306.12156 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "56. Zheng, Q., Qiao, X., Cao, Y., Lau, R.W.: Distraction-aware shadow detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5167-5176 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.785, + 0.55 + ], + "angle": 0, + "content": "57. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "58. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.577, + 0.785, + 0.633 + ], + "angle": 0, + "content": "59. Zhu, L., Deng, Z., Hu, X., Fu, C.W., Xu, X., Qin, J., Heng, P.A.: Bidirectional feature pyramid network with recurrent attention residual modules for shadow detection. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 121-136 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.633 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CAT-SAM_ Conditional Tuning for Few-Shot Adaptation of Segment Anything Model/bd649bc1-acc8-4027-a33f-08ccf2414aa5_origin.pdf b/2024/CAT-SAM_ Conditional Tuning for Few-Shot Adaptation of Segment Anything Model/bd649bc1-acc8-4027-a33f-08ccf2414aa5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8f5ca6ec27d1503402b75c0f6f6398d8b06636a2 --- /dev/null +++ b/2024/CAT-SAM_ Conditional Tuning for Few-Shot Adaptation of Segment Anything Model/bd649bc1-acc8-4027-a33f-08ccf2414aa5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e9a95a0a57e467d18bd7040431fc7da4cf209362f206f21f32efec14405720e +size 993936 diff --git a/2024/CAT-SAM_ Conditional Tuning for Few-Shot Adaptation of Segment Anything Model/full.md b/2024/CAT-SAM_ Conditional Tuning for Few-Shot Adaptation of Segment Anything Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..134758f5624e18691a7b81ba76b3b7a91f2f3506 --- /dev/null +++ b/2024/CAT-SAM_ Conditional Tuning for Few-Shot Adaptation of Segment Anything Model/full.md @@ -0,0 +1,287 @@ +# CAT-SAM: Conditional Tuning for Few-Shot Adaptation of Segment Anything Model + +Aoran Xiao $^{1*}$ , Weihao Xuan $^{2,3*}$ , Heli Qi $^{4}$ , Yun Xing $^{1}$ , Ruijie Ren $^{5}$ , Xiaojin Zhang $^{6}$ , Ling Shao $^{7}$ , and Shijian Lu $^{1\dagger}$ + +1 Nanyang Technological University, Singapore +2 The University of Tokyo, Japan +3 RIKEN AIP, Japan + +$^{4}$ Nara Institute of Science and Technology, Japan + $^{5}$ Waseda University, Japan + +6 Zhejiang University of Technology, China + +UCAS-Terminus AI Lab, University of Chinese Academy of Sciences, China + +Abstract. The Segment Anything Model (SAM) has demonstrated remarkable zero-shot capability and flexible geometric prompting in general image segmentation. However, it often struggles in domains that are either sparsely represented or lie outside its training distribution, such as aerial, medical, and non-RGB images. Recent efforts have predominantly focused on adapting SAM to these domains using fully supervised methods, which necessitate large amounts of annotated training data and pose practical challenges in data collection. This paper presents CAT-SAM, a ConditionAI Tuning network that explores few-shot adaptation of SAM toward various challenging downstream domains in a data-efficient manner. The core design is a prompt bridge structure that enables decoder-conditioned joint tuning of the heavyweight image encoder and the lightweight mask decoder. The bridging maps the domain-specific features of the mask decoder to the image encoder, fostering synergic adaptation of both components with mutual benefits with few-shot target samples only, ultimately leading to superior segmentation in various downstream tasks. We develop two CAT-SAM variants that adopt two tuning strategies for the image encoder: one injecting learnable prompt tokens in the input space and the other inserting lightweight adapter networks. Extensive experiments over 11 downstream tasks show that CAT-SAM achieves superior segmentation consistently even under the very challenging one-shot adaptation setup. + +Keywords: Segment Anything Model $\cdot$ Few-shot Learning $\cdot$ Parameter Efficient Tuning + +# 1 Introduction + +Accurate image segmentation is a fundamental computer vision task that plays a pivotal role in various applications such as robotics, autonomous driving, health- + +![](images/069a80822340aee613fa1fe128d81549102ab37092b55803c0a0ff70fd76f468.jpg) +Fig. 1: The proposed CAT-SAM performs ConditionAl joint Tuning (CAT) to establish communication between SAM's heavyweight image encoder and lightweight mask decoder. This enables synergistic adaptation of the two network components, mitigating tuning imbalances and improving few-shot SAM adaptation. + +care, earth observation, etc. The recently developed Segment Anything Model (SAM) [23], trained with 1.1 billion masks, has emerged as a momentous leap forward in image segmentation. By taking geometric prompts with points, boxes, or masks as input, SAM demonstrates remarkable zero-shot capability for general image segmentation, as well as great potential for fine-grained mask segmentation in different downstream tasks across a variety of contexts. + +On the other hand, SAM often fails to generate high-quality predictions when dealing with domains that are either sparsely represented or lie outside its training distribution [5,19], such as challenging RGB images including aerial, medical, and intricate structural images, as well as non-RGB images like X-ray, Sonar, Synthetic Aperture Radar (SAR) images, etc. This greatly undermines the versatility and applicability of SAM as a foundational segmentation model while handling various downstream tasks. Several studies attempted to address this issue by fine-tuning SAM with a large number of target samples with mask annotations [21, 30]. However, they require extensive masked images for each specialized field, posing a formidable task with poor scalability. How to effectively adapt SAM toward various downstream tasks in a data-efficient manner has become an essential challenge in the segmentation field. + +We exploit few-shot target samples for effective and efficient adaptation of SAM toward various downstream tasks. Inspired by the recent advancements in parameter-efficient tuning for foundation models [1,13], we freeze the entire SAM and lightly expand its image encoder and mask decoder with a tiny amount of learnable parameters. This expansion preserves SAM's zero-shot capabilities and flexibility while capturing domain-specific features that are essential for the segmentation of downstream data. However, SAM's image encoder is much larger than its mask decoder, which directly leads to tuning imbalance and further suboptimal adaptation, especially under the limited supervision of few-shot target samples. + +To address this challenge, we design CAT-SAM, a novel ConditionAl Tuning network for effective and data-efficient adaptation of SAM. The key idea is to ex + +ploit the lightweight decoder tuning to guide the heavyweight encoder tuning and formulate decoder-conditioned joint tuning of both as depicted in Fig. 1. To this end, we design a prompt bridge structure that maps the domain-specific features from the mask decoder to the image encoder. With the collaboration between the two partners, CAT-SAM enables synergistic adaptation of both which mitigates the tuning imbalance and improves the few-shot SAM adaptation effectively. In addition, the prompt bridge can be seamlessly embedded into two representative tuning approaches, including prompt tuning [20, 25, 27] that introduces learnable prompt tokens into the input space and adapters [15, 29, 37, 38] that leverage lightweight adaptive networks. This directly leads to two CAT-SAM variants, namely, CAT-SAM-T and CAT-SAM-A. Extensive evaluations over 11 diverse and challenging downstream tasks show that both variants achieve superior adaptation and segmentation with few-shot target samples only. + +The major contributions of this work can be summarized in three aspects. First, we propose CAT-SAM, a conditional tuning network that enables effective and data-efficient adaptation of SAM toward various challenging downstream domains. We design prompt bridge within CAT-SAM, a decoder-conditioned joint tuning structure that enables synergistic and data-efficient adaptation of the heavyweight image encoder and the lightweight mask decoder effectively. Second, we develop two CAT-SAM variants by embedding the prompt bridge into two representative tuning strategies, one introducing learnable prompt tokens in the input space and the other inserting lightweight adapter networks. Third, extensive experiments over 11 diverse segmentation datasets show that CAT-SAM achieves superior image segmentation consistently even under the challenging one-shot setup. + +# 2 Related Work + +Parameter-Efficient Tuning for foundation models has become increasingly crucial due to the over-size model parameters and costly one-task-one-model deployment. One popular approach involves updating only the newly added learnable parameters while keeping the model backbone frozen. These lightweight parameters are often introduced in two ways: 1) as prompt tuning [12,20,22,58] by adding learnable tokens to input tokens at transformer layers, 2) as adapters [11, 29,37,38,49] by integrating learnable lightweight sub-networks into transformer layers. However, prior visual studies primarily focus on adapting single backbone modules [3,20,29] or dual encoder backbones for vision-language tasks [22,57], which are suboptimal tuning solutions while applying to SAM due to its particular architecture of a heavyweight image encoder and a lightweight mask decoder. In contrast, we explore a novel conditional joint tuning strategy tailored to SAM, which facilitates its adaptation even with limited downstream data. + +SAM for Downstream Tasks. The release of SAM has spurred several subsequent studies. Some concentrate on expanding semantic recognition capabilities [24, 26, 43, 50, 54], while others aim to create more lightweight SAM variants [48, 52, 55] for faster computations. Several studies explore SAM's adapta + +tion in underperforming downstream scenarios, primarily by directly fine-tuning SAM's mask decoder [4,8,30,41,46,53], which can severely degrade SAM's zero-shot capabilities. The recent HQ-SAM [21] freezes the entire pre-trained SAM and prompt-tunes the mask decoder with a new mask head for adaptive mask prediction. However, these adaptation studies rely on large target datasets with a considerable number of annotated images, which are often costly to collect. In contrast, our CAT-SAM achieves superior adaptation with only few-shot target samples, significantly reducing adaptation costs and expanding the applicability of SAM as a visual foundation model. + +# 3 Method + +# 3.1 Preliminaries of SAM + +SAM [23] handles each image with three key modules: a heavyweight image encoder (i.e., ViT [9] as the backbone) that extracts image embeddings, a prompt encoder that encodes the geometric prompts (i.e., points, a box, or a coarse mask) to generate prompt embeddings, and a lightweight mask decoder that combines the two types of embeddings to predict segmentation masks. The released SAM model is trained on a super large-scale SA-1B dataset, which consists of over 11 million images as well as 1.1 billion automatically generated masks. SAM has demonstrated remarkable zero-shot capability while dealing with various conventional natural images, as well as superb flexibility in accepting various geometric prompt inputs. On the other hand, it often struggles while facing challenging downstream domains that are often sparse or falling outside of its training data distribution, such as RGB domains of aerial, medical, and intricate structural images, as well as non-RGB domains of X-ray, SAR, and Sonar images. More details about SAM are presented in [23]. + +# 3.2 CAT-SAM + +This subsection introduces CAT-SAM, a ConditionAl Tuning network that is designed for adapting SAM toward various under-performed downstream data. The objective is to adapt SAM efficiently with only few-shot annotated target images, meanwhile preserving SAM's powerful zero-shot capability and geometric prompting flexibility. + +CAT-SAM freezes the entire SAM and facilitates the simultaneous tuning of SAM's image encoder and mask decoder to capture representative information from few-shot target samples for effective adaptation of SAM. However, a critical issue with simultaneous tuning arises from the clear imbalance between the heavyweight image encoder (308.3 M parameters for ViT-L [9]) and the lightweight mask decoder (4.1 M), which often leads to sub-optimal adaptation while only few-shot target samples are available. To tackle this challenge, we design decoder-conditioned joint tuning, a novel tuning strategy that mitigates the imbalance by establishing a linkage between the encoding tuning and the decoding tuning. Specifically, we design prompt bridge—a lightweight network that + +![](images/fc430856e4daded764b4acbdea330eb06029422ea75a3c539c5fdffc38bcae2b.jpg) +(a) CAT-SAM-T + +![](images/cd1359f00a47f4f792f48395e146e58919584e31308d1a27be1a36b51cea6189.jpg) +(b) CAT-SAM-A + +![](images/62da2b3ef65a27621f55538a957a9411f2ef56f2935339eafdca266b94a62478.jpg) +(c) Prompt Bridge +Fig. 2: Overview of CAT-SAM. CAT-SAM keeps the whole SAM frozen while simultaneously tuning the image encoder and mask decoder for downstream adaptation. To address the tuning imbalance between these two network components, we introduce decoder-conditioned joint tuning through the design of Prompt Bridge structure, enabling synergetic and enhanced adaptation. We present two CAT-SAM variants: CAT-SAM-T in (a) and CAT-SAM-A in (b), achieved by integrating the prompt bridge with prompt-based and adapter-based tuning strategies for the image encoder, respectively. (c) illustrates two tailored prompt bridge structures, PB-T and PB-A. + +maps the domain-specific features from the mask decoder to the image encoder. The bridge design facilitates joint and balanced tuning of both network components, leading to synergistic and effective adaptation of SAM with few-shot and even a single target sample. + +We develop two CAT-SAM variants by integrating the proposed decoder-conditioned joint tuning into two prevalent tuning approaches, namely CAT-SAM-T and CAT-SAM-A, as illustrated in Fig. 2. + +Tuning Image Encoder. We first brief SAM's image encoder and then introduce two tuning approaches in our CAT-SAM, including prompt tuning in CAT-SAM-T and lightweight adapters in CAT-SAM-A. + +- SAM's Image Encoder, denoted as $\mathcal{V}$ with $K$ transformer layers $\{\mathcal{V}_i\}_{i=1}^K$ , first splits the input image $I$ of the size $(H, W)$ into $M$ fixed-size patches, and then projects them into patch embeddings $E_0 \in \mathbb{R}^{M \times d_v}$ . Patch embeddings $E_{i-1}$ are input to the $i^{th}$ transformer layer $(\mathcal{V}_i)$ and sequentially processed through $K$ transformer layers: + +$$ +[ E _ {i} ] = \mathcal {V} _ {i} ([ E _ {i - 1} ]), i = 1, 2, \dots , K. +$$ + +- CAT-SAM-T: As illustrated in Fig. 2 (a), we introduce a set of learnable tokens $\{P_i \in \mathbb{R}^{d_v}\}_{i=0}^{K-1}$ in the image encoder alongside the input patch embeddings. These new learnable tokens are introduced in each transformer layer of the image encoder $\mathcal{V}$ and updated through the adaptation process: + +$$ +[ E _ {i} ] = \mathcal {V} _ {i} ([ E _ {i - 1}, P _ {i - 1} ]) ], i = 1, 2, \dots , K. +$$ + +- CAT-SAM-A: As illustrated in Fig. 2 (b), we introduce Adapters [15] to tune SAM's image encoder in CAT-SAM-A. The adapter is a lightweight sub-network that is inserted into each transformer layer, comprising a linear down-projection, a nonlinear activation function, a linear up-projection, and a residual connection. High-frequency image information is also incorporated in CAT-SAM as the input of the adapters as in [29]. In this process, high-frequency components $I_{hfc}$ are first extracted from the input image $I$ and then reversed to the space domain using the Fast Fourier Transform and its inverse. These components are then partitioned into small patches $I_{hfc}^p$ similar to the image patches. Subsequently, convolutional layers and linear layers project $I_{hfc}$ and the original image embeddings $E_0$ into $c$ -dimensional features independently, yielding $F_{hfc}$ and $F_{pe}$ , respectively. The result of element-wise addition of $F_{pe}$ and $F_{hfc}$ is then forwarded (as denoted "EVP" in Fig. 2 (b)) to the adapter $\mathrm{Adapt}_i$ for transformer layer $i$ . Finally, the output is added element-wise to the original input of the transformer layer $i$ : + +$$ +E _ {i - 1} = E _ {i - 1} + \mathrm {A d a p t} _ {i} (F _ {p e} + F _ {h f c}). +$$ + +Tuning Mask Decoder. We tune the mask decoder of SAM similarly as [21]. Specifically, we freeze the entire pre-trained decoder of SAM and introduce a learnable CAT-Token that is concatenated with SAM's original output tokens and prompt tokens to form the input to SAM's mask decoder. The CAT-Token is updated during training but remains fixed during inference, the same as SAM's output tokens. After passing through two decoder layers, the updated CAT-Token is used to generate dynamic weights for a newly introduced three-layer MLP. Simultaneously, SAM's mask decoder features are fused with the image features decoded by the image encoder. Finally, the fused features and the three-layer MLP are combined through the dot product to generate the target mask. Please refer to the appendix for more details about the decoder structures in CAT-SAM. + +Decoder-Conditioned Joint Tuning. The proposed decoder-conditioned joint tuning can be seamlessly embedded via two implementations of the proposed prompt bridging ideas, i.e., PB-T and PB-A as illustrated in Fig. 2 (c). + +- CAT-SAM-T: The prompt bridge $\mathrm{PB}_i^T$ is a two-layer MLP. It projects the CAT-Token $Q$ in the mask decoder to each transformer layer $i$ in the image encoder and outputs a single learnable token replacing $P_i$ by $\tilde{P}_i^T = \mathrm{PB}_i^T(Q)$ . These mapped tokens are directly applied for prompt tuning of the image encoder as follows: + +$$ +\left[ E _ {i} \right] = \mathcal {V} _ {i} \left(\left[ E _ {i - 1}, \tilde {P} _ {i - 1} ^ {T} \right]\right), i = 1, 2, \dots , K. +$$ + +- CAT-SAM-A: The prompt bridge $\mathrm{PB}^A$ maps the CAT-Token $Q$ into an embedding $\tilde{P}^A = \mathrm{PB}^A(Q)$ of the same size as $F_{pe}$ and $F_{hfc}$ . $\tilde{P}^A$ is then element-wise added to the EVP before being fed to every adapter as follows: + +$$ +E _ {i - 1} = E _ {i - 1} + \mathrm {A d a p t} _ {i} (F _ {p e} + F _ {h f c} + \tilde {P} ^ {A}). +$$ + +Table 1: Comparisons of trainable parameters between ViT-L [9] based SAM and two variants of the proposed CAT-SAM. + +
ModelTrainable Parameters (M)Trainable Params vs. SAM
SAM [23]312.4-
CAT-SAM-T3.31.1%
CAT-SAM-A1.90.6%
+ +As illustrated in Fig. 2 (c), PB-A employs $c$ different linear down-projection layers to generate embeddings independently. After GELU layers, these embeddings are up-projected to the dimension of $1 \times M$ ( $M$ is the spatial size of the patch) with a shared linear layer to mitigate the computation cost. + +Finally, the up-projected embeddings are concatenated to produce $\tilde{P}^A$ . + +Tab. 1 compares the amount of trainable parameters between the SAM (implemented with the ViT-L [9]) and our CAT-SAM that is built upon the SAM. It can be observed that both CAT-SAM variants introduce very limited additional parameters, yet yield significant performance improvements in various downstream segmentation tasks as to be detailed in Section 4. + +# 3.3 Training and Inference + +During training, we feed target samples together with geometric prompts into CAT-SAM to generate CAT-SAM masks, under the supervision of ground-truth masks as well as a linear combination of BCE loss and dice loss [31]. We freeze the pre-trained SAM parameters and update solely the parameters that are introduced in the above-described tuning modules. Note we train CAT-SAM with a mixture of sampled geometric prompts, including bounding boxes, randomly selected points, and coarse masks. In addition, we create degraded coarse masks by introducing random Gaussian noise in the boundary regions of the ground-truth masks as in [21]. Further training details can be found in Section 4.1 and the appendix. + +During the inference phase, the input image is fed into SAM's image encoder along with the integrated prompt tokens in CAT-SAM-T or adapters in CAT-SAM-A, to generate adaptive image embeddings. These embeddings, combined with the prompt tokens from SAM's prompt encoder, serve as input to the mask decoder. Subsequently, the updated CAT-token and associated MLP are utilized for the target mask prediction. Lastly, we up-sample the mask to the original resolution to produce the final output. + +# 4 Experiments + +We evaluate CAT-SAM's efficacy through comprehensive experiments involving eight segmentation tasks across 11 datasets, all hailing from challenging downstream fields that SAM struggles to address effectively. Our experiments encom + +Table 2: We evaluate and benchmark CAT-SAM over eight challenging segmentation tasks across 11 datasets. + +
TasksDataset NamesImagery
Building segmentationWHU [18]Aerial images
Road segmentationMA. Roads [32]Aerial images
Polyp segmentationKvasir [34]Medical images
Chest X-ray segmentationJSRT [39]X-ray images
Marine debris segmentationFLS [40]Sonar images
Ship instance segmentationHRSID [45]SAR images
Shadow segmentationSBU-Shadow [42]Natural images
Intricate segmentationDIS [35],ThinObject [28],HRSOD [51],COIFT [28]Natural images
+ +pass a wide array of scenarios and tasks, spanning from varying target shots to different imagery modalities. + +# 4.1 Experimental Setup + +Datasets. We conduct experiments on two collections of datasets listed in Tab. 2: 1) True-color images, including WHU [18] for building segmentation, Kvasir [34] for polyp segmentation, SBU-Shadow [42] for shadow segmentation, Massachusetts (MA.) Roads [32] for road segmentation, as well as fine-grained segmentation datasets DIS [35], ThinObject [28], HRSOD [51], and COIFT [28]. 2) Non-RGB domains, encompassing JSRT [39] for chest organ segmentation (X-ray images), HRSID [45] for ship instance segmentation (high-resolution SAR images), and FLS [40] for marine debris segmentation (Sonar images). We use official dataset splits for fair comparisons. + +Evaluation Metrics. For single-class datasets including WHU, Kvasir, SBU-Shadow, and MA. Roads, we employ standard mask IoU of foreground. JSRT and FLS, which have multiple classes, are evaluated using individual class IoU and their average. As for DIS, ThinObject, HRSOD, and COIFT, we follow [21] and report mIoU and the boundary metrics (mBIoU) [7] for fair comparisons. For instance segmentation of HRSID, we use standard AP, $\mathrm{AP}_{50}$ , and $\mathrm{AP}_{75}$ . + +Implementation Details. Unless otherwise specified, CAT-SAM and its comparing models use ViT-L [9] as the image encoder backbone. We train on 1 NVIDIA RTX A6000 GPU for one-shot adaptation, while 4 GPUs for multiple-shots. Following [21], ground truth boxes serve as the default input geometric prompts during evaluation to ensure fair comparison and minimize randomness, with the exception of Fig. 3 where point prompts are evaluated. Please refer to the appendix for more implementation details. + +# 4.2 Ablation Studies + +We first investigate the contributions of different tuning modules within both CAT-SAM variants to assess their impact on overall adaptation performance. + +Table 3: Ablation study of CAT-SAM-T in (a) and CAT-SAM-A in (b) for 1-shot adaptation over datasets: "W" for WHU [18] (on building segmentation), "K" for Kvasir [34] (on polyp segmentation), and "S" for SBU-Shadow [42] (on shadow segmentation). "Enc-T." and "Enc-A." means tuning the image encoder with prompt tokens and adapters, respectively. "Dec." means tuning the mask decoder. "PB-T" and "PB-A" denote two customized prompt bridges for conditional tuning. Model 1 is the baseline SAM without adaptation. +(a) CAT-SAM-T. + +
Enc-T. Dec. PB-TWKSAVG
143.579.062.461.6
2✓57.880.476.0
3✓71.273.863.5
4✓✓66.473.261.8
5✓✓✓86.883.4
+ +(b) CAT-SAM-A. + +
Enc-A. Dec. PB-AWKSAVG
143.579.062.461.6
2✓86.573.776.3
3✓71.273.863.5
4✓✓83.479.876.3
5✓✓✓88.285.4
+ +Tab. 3 shows the experiments of one-shot adaptation across three true-color image segmentation benchmarks WHU [18], Kvasir [34], and SBU-Shadow [42]. For each CAM-SAM variant CAM-SAM-T/CAM-SAM-A, we compare 5 models including: 1. The original SAM without adaptation (baseline); 2. Tuning SAM's image encoder alone with prompt tokens or adapters; 3. Tuning SAM's mask decoder alone; 4. Tuning SAM's image encoder and mask decoder independently; and 5. Conditional tuning of SAM's image encoder and mask decoder with the proposed prompt bridging (i.e., the complete CAT-SAM-T and CAT-SAM-A). + +We can see that the original SAM struggles to produce high-quality masks in these challenging domains, indicating its limitations as a foundational segmentation model for many downstream applications. Tuning the image encoder or the mask decoder alone in 2 and 3 enhances target segmentation, underlining the efficacy and necessity of SAM adaptation. However, simultaneously tuning the image encoder and mask decoder in an independent way in model 4 does not show complementary and consistent improvement. This suggests instability and imbalance in tuning between the heavyweight image encoder and the lightweight mask decoder, which could lead to suboptimal adaptation solutions. Differently, including the proposed prompt bridging with either PB-T or PB-A (i.e., the complete CAT-SAM-T and CAT-SAM-A) mitigates the tuning imbalance and produces substantial performance enhancement consistently. + +# 4.3 Comparison with the State-of-the-Art + +CAT-SAM for One-Shot Adaptation. To assess CAT-SAM's effectiveness in reducing downstream training data and its adaptability with few-shot samples, we first evaluate it under the extremely challenging one-shot adaptation scenario. Given the lack of prior studies under such setup, we comprehensively benchmark it against closely related methods that can be broadly categorized into + +Table 4: Comparison of adaptive segmentation performance on challenging true-color datasets: WHU [18] (on building segmentation), Kvasir [34] (on Polyp segmentation), and SBU-Shadow [42] (on shadow segmentation). The baseline is SAM without adaptation. All compared methods utilize one-shot sample for adaptation, except for those denoted by * which utilize 20 shots with metric numbers sourced from [47]. + +
MethodsWHUKvasirSBU-ShadowAverage
SAM [23] (baseline)43.579.062.461.6
VPT-shallow [20]60.879.868.769.8
VPT-deep [20]57.880.476.071.4
AdaptFormer [3]83.276.877.279.1
LoRA [16]86.177.574.479.3
Med-SA [46]34.528.622.128.4
MedSAM [30]30.629.853.437.9
SimAda [41]48.618.749.338.9
SPFS-SAM* [47] (20-shots)-53.4--
SAMed* [53] (20-shots)-51.7--
HQ-SAM [21]71.273.863.569.5
CAT-SAM-T (Ours)86.883.478.082.7
CAT-SAM-A (Ours)88.285.481.985.2
+ +two groups. The first group comprises state-of-the-art tuning methods for foundation models of different modalities, including VPT [20] and AdaptFormer [3] for large vision foundation models, and LoRA [16] for large language models. The second group comprises recent SAM-based adaptation methods which were largely designed for full-shot adaptation [21,30,41,46,53] ( [47] tackles few-shot adaptation but requires 20 shots). + +Tab. 4 presents results over WHU, Kvasir, and SBU-Shadow datasets. We can observe that all three tuning methods for foundation models exhibit clear adaptation effects. Notably, AdaptFormer and LoRA demonstrate significantly better performance than VPT-deep and VPT-shallow, highlighting the effectiveness of modifying the model's structure over prompt tuning in the input space during SAM adaptation. On the other hand, most SAM-based adaptation methods, including Med-SA, SimAda, SPFS-SAM, SAMed, and MedSAM, yield much lower performance, primarily due to their reliance on many more target images which leads to overfitting while learning from one-shot target sample. HQ-SAM achieves substantial segmentation improvements but it still cannot compete with the three tuning methods, demonstrating the significance of tuning the image encoder over the mask decoder. Finally, both CAT-SAM variants consistently outperform all compared methods, with CAT-SAM-A exhibiting a slight edge due to its network structure modifications. These experiments validate the superiority of the proposed decoder-conditioned joint tuning approach in extracting domain-specific features with few-shot samples. + +Table 5: Comparison of state-of-the-art segmentation achieved by fully supervised learning (FSL) methods trained on full-shot data (upper part) and mask segmentation by SAM and our CAT-SAM (lower part). Evaluation follows official criteria of each FSL benchmark, reporting IoU for WHU, mIoU for Kvasir, and BER for SBU-Shadow to ensure consistency. "(1)", "(16)", and "(full)" after CAT-SAM-T/CAT-SAM-A denote tuning with 1-shot, 16-shots, and full-shots samples. + +
WHUKvasirSBU-Shadow
ModelIoUModelmIoUModelBER(↓)
STT [2]90.5CASCADE [36]87.8BDRAR [59]3.64
BuildFormer [44]91.4PatchRefineNet [33]89.1DSDNet [56]3.45
CBR-Net [14]91.4DUCK-Net [10]90.5MTMT [6]3.15
SAM43.5SAM87.3SAM13.22
CAT-SAM-T (1)86.8CAT-SAM-T (1)90.0CAT-SAM-T (1)7.81
CAT-SAM-A (1)88.2CAT-SAM-A (1)91.3CAT-SAM-A (1)5.27
CAT-SAM-T (16)89.6CAT-SAM-T (16)93.1CAT-SAM-T (16)4.04
CAT-SAM-A (16)90.3CAT-SAM-A (16)93.6CAT-SAM-A (16)3.80
CAT-SAM-T (full)93.3CAT-SAM-T (full)94.5CAT-SAM-T (full)2.54
CAT-SAM-A (full)93.6CAT-SAM-A (full)94.3CAT-SAM-A (full)2.39
+ +Full-Shot Learning. While CAT-SAM is primarily designed for few-shot adaptation, it also demonstrates proficiency in full-shot adaptation scenarios. In Tab. 5, we present the state-of-the-art performances of traditional fully-supervised learning (FSL) methods trained over full-shot data and compare them with SAM and our CAT-SAM (adapted with various target samples). We report official evaluation metrics for three public benchmarks, including IoU for WHU, mIoU for Kvasir, and BER [17] for SBU-Shadow. Note the distinction in setups between the FSL segmentation and the mask segmentation: FSL focuses on traditional semantic segmentation of specific classes without prompt input while SAM and CAT-SAM involve general mask segmentation with prompt input. Despite these differences, these performances can serve as valuable references for evaluating CAT-SAM's segmentation ability. + +With full target training data, both CAT-SAM variants outperform state-of-the-art FSL methods consistently. Notably, the two CAT-SAM variants achieve impressive target segmentation under the challenging one-shot setup with a single target sample. With 16 shots of target samples, their performance is even on par with the FSL methods that are trained by using full target training data. These experiments underscore the superiority of effective transfer of broad knowledge from the foundation model SAM as compared with training from scratch with FSL. It also demonstrates the remarkable efficacy of CAT-SAM in substantially expanding SAM's applicability across various downstream tasks. + +We conducted further experiments to compare CAT-SAM against HQ-SAM [21], a state-of-the-art SAM-based adaptation solution for intricate structural image segmentation. For fair comparisons, we follow the evaluation protocol of + +Table 6: Extremely fine-grained segmentation over four intricate structural image datasets. HQ-SAM and the two CAT-SAM variants are fine-tuned over full HQSeg-44K [21]. All adopt the boxes converted from their GT masks as the box prompt input for fair comparisons. + +
ModelsDIS [35]COIFT [28]HRSOD [51]ThinObject [28]Average
mIoUmBIoUmIoUmBIoUmIoUmBIoUmIoUmBIoUmIoUmBIoU
SAM [23]62.052.892.186.590.283.173.661.879.571.1
HQ-SAM [21]78.670.494.890.193.686.989.579.989.181.8
CAT-SAM-T84.078.195.692.093.487.694.087.991.786.4
CAT-SAM-A83.677.795.692.293.187.194.188.291.686.3
+ +HQ-SAM and tune CAT-SAM on the complete HQSeg-44K dataset, consisting of 44,320 meticulously annotated image masks. Tab. 6 shows experiments across four exceptionally fine-grained segmentation datasets including DIS [35], ThinObject [28], HRSOD [51], and COIFT [28] (in mIoU and the boundary metrics (mBIoU) [7]). We can see that CAT-SAM-T achieves slightly better performance than CAT-SAM-A, while both clearly outperform HQ-SAM and the original SAM. These experiments further demonstrate the efficacy of our proposed conditional tuning in adapting SAM toward challenging downstream tasks. Visual qualitative comparisons are provided in the appendix. + +# 4.4 CAT-SAM for Non-RGB Domains + +We also explore CAT-SAM's performance across more challenging image modalities characterized by larger domain discrepancies. Specifically, we assess its efficacy on non-RGB domains, including JSRT [39] for chest organ segmentation using X-ray images and FLS [40] for marine debris segmentation with Sonar images, and HRSID [45] for ship instance segmentation with SAR images. For adaptation, we randomly select either one or sixteen target images for 1-shot and 16-shot scenarios, respectively, while employing the entire training set for full-shot adaptation. These datasets represent novel image distributions as compared with SAM's training data. Note we train instances of all classes together while evaluating them by classes. Since no images in FLS contain objects of all 11 classes, we only report results of 16-shot and full-shot adaptation for FLS. + +Tab. 7 shows experimental results. Notably, CAT-SAM-T and CAT-SAM-A improve the mIoU to $93.0\%$ and $92.6\%$ , respectively, with only one shot of JSRT training image. The improvements saturate gradually as the number of training samples increases, with marginal gains of $1.4\%$ and $2.0\%$ , respectively, when adapted with full-shot target data. These findings underscore the superior few-shot capability of the proposed CAT-SAM and align well with those reported in Tab. 5. Experiments on FLS [40] show different results. While both CAT-SAM variants exhibit clear adaptation effects, the performance gains for both 16-shot and full-shot adaptation are notably lower as compared with previously evaluated datasets. The discrepancy is largely attributed to two primary + +Table 7: Few-shot adaptation with varying number of training samples for non-RGB datasets, including FLS [40] with Sonar images for marine debris segmentation, JSRT [39] for chest X-ray segmentation, and HRSID [45] for ship instance segmentation with SAR images. "N/A" denotes that 1-shot adaptation over FLS is unattainable due to the absence of images containing objects from all 11 classes. + +
#SamplesMethodsJSRT [39]FLS [40]HRSID [45]
LungsHeartmIoUmIoUAPAP50AP75
NoneSAM [23]85.071.978.569.738.286.626.8
1-shotCAT-SAM-T95.490.593.0N/A46.088.445.4
CAT-SAM-A94.790.592.6N/A44.988.041.4
16-shotCAT-SAM-T95.892.794.273.246.289.345.2
CAT-SAM-A95.591.593.571.445.789.443.7
Full-shotCAT-SAM-T96.392.694.481.751.493.055.0
CAT-SAM-A96.492.894.682.052.993.956.0
+ +factors: 1) the inherent difficulty of FLS with its 11 semantic classes (per-class performances are provided in appendix); and 2) the larger variation of FLS data in shapes and object sizes as compared with previously evaluated data. As for ship instance segmentation using SAR images in HRSID [45], this dataset presents different challenges as ships are very small in size. The results show that CAT-SAM consistently outperforms SAM in this diverse domain, highlighting its efficiency in domain adaptation with large domain discrepancy. These experiments further underscore the need for a more comprehensive adaptation of SAM for exceptionally challenging downstream tasks. We expect further studies along this research direction. + +# 4.5 Analysis & Discussion + +Prompts with Single Point. SAM provides flexibility in geometric prompts including a single foreground point. However, prompting with a single foreground point can be challenging as the single point could correspond to multiple objects [23]. Nonetheless, adaptation with few-shot target samples can alleviate this ambiguity by guiding the model to focus on the specific foreground distribution within the annotation space. We evaluate this by using the same randomly-selected single point as the geometric prompt, and Fig. 3 shows + +the segmentation results for images from WHU, Kvasir, SBU-Shadow, and MA. Roads [32] on road segmentation. + +![](images/918a78860252361b79e735130a6795a93b496d59c3d6b858d0076014486d2089.jpg) +Fig. 3: Single point to mask evaluation for one-shot adaptation. + +![](images/288cfba52fa1f43e027ad782621ea1f43be7e63611c172717c612dbb21c2cb98.jpg) +Fig. 4: Visual comparisons of SAM [23] (top row) and CAT-SAM (bottom row). We illustrate samples from WHU [18] for building segmentation, MA. Roads [32] for road segmentation, SBU-Shadow [42] for shadow segmentation, Kvasir [34] for polyp segmentation, JSRT [39] for chest organ segmentation (X-ray images), FLS [40] for marine debris segmentation (Sonar images), and HRSID [45] for ship instance segmentation (SAR images). CAT-SAM exhibits one-shot adaptation across most datasets, except for 16-shot over FLS. Red boxes and stars denote geometric prompts, colored regions are mask predictions, and lines show the boundary of ground truth segmentation. + +Similar to the experiments using boxes as prompts, CAT-SAM consistently demonstrates superior target segmentation while using a single point as prompt, underscoring its robustness in adaptation while preserving SAM's inherent geometric prompting flexibility. In addition, both CAT-SAM variants achieve relatively less improvement for the dataset Kvasir. This can be largely attributed to the high similarities in color and texture between the polyps and the local gastrointestinal regions in Kvasir, making it challenging for a single point to provide clear geometric guidance. + +Visual Illustrations. Fig. 4 shows qualitative comparisons between SAM and CAT-SAM across multiple segmentation tasks. These illustrations demonstrate how our proposed CAT-SAM remarkably enhances the mask quality of SAM under the presence of only one-shot target samples. Please refer to the appendix for more visual comparisons. + +# 5 Conclusion + +We introduce CAT-SAM, a conditional tuning network tailored for few-shot adaptation of SAM. We propose decoder-conditional joint tuning to mitigate the tuning imbalance between SAM's heavyweight image encoder and lightweight mask decoder and facilitate efficient SAM adaptation. To achieve this, we design the prompt bridge structures, enabling interaction when tuning these two network modules. We develop two CAT-SAM variants by integrating the prompt bridge with two representative tuning strategies including prompt tuning and adapter. Our comprehensive evaluation across diverse scenarios and tasks on 11 segmentation datasets underscores the superior domain adaptive efficiency of both CAT-SAM variants, even in the extremely challenging one-shot scenario. + +# Acknowledgements + +This project is partly funded by the Ministry of Education Singapore, under the Tier-1 scheme with project number RG18/22 and the Tier-2 scheme with project number MOE-T2EP20123-0003. It is also supported in part by the National Natural Science Foundation of China [grant no. U2033210], in part by the Zhejiang Provincial Natural Science Foundation [grant no. LDT23F02024F02]. + +# References + +1. Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., et al.: On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021) +2. Chen, K., Zou, Z., Shi, Z.: Building extraction from remote sensing images with sparse token transformers. Remote Sensing 13(21), 4441 (2021) +3. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022) +4. Chen, T., Zhu, L., Deng, C., Cao, R., Wang, Y., Zhang, S., Li, Z., Sun, L., Zang, Y., Mao, P.: Sam-adapter: Adapting segment anything in underperformed scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 3367-3375 (October 2023) +5. Chen, T., Zhu, L., Ding, C., Cao, R., Zhang, S., Wang, Y., Li, Z., Sun, L., Mao, P., Zang, Y.: Sam fails to segment anything?—sam-adapter: Adapting sam in underperformed scenes: Camouflage, shadow, and more. arXiv preprint arXiv:2304.09148 (2023) +6. Chen, Z., Zhu, L., Wan, L., Wang, S., Feng, W., Heng, P.A.: A multi-task mean teacher for semi-supervised shadow detection. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 5611-5620 (2020) +7. Cheng, B., Girshick, R., Dollar, P., Berg, A.C., Kirillov, A.: Boundary iou: Improving object-centric image segmentation evaluation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15334-15342 (2021) +8. Cheng, J., Ye, J., Deng, Z., Chen, J., Li, T., Wang, H., Su, Y., Huang, Z., Chen, J., Jiang, L., et al.: Sam-med2d. arXiv preprint arXiv:2308.16184 (2023) +9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +0. Dumitru, R.G., Peteleaza, D., Craciun, C.: Using duck-net for polyp image segmentation. Scientific Reports 13(1), 9803 (2023) +1. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision pp. 1-15 (2023) +2. Gong, Z., Li, F., Deng, Y., Bhattacharjee, D., Zhu, X., Ji, Z.: Coda: Instructive chain-of-domain adaptation with severity-aware visual prompt tuning (2024) +3. Gu, J., Han, Z., Chen, S., Beirami, A., He, B., Zhang, G., Liao, R., Qin, Y., Tresp, V., Torr, P.: A systematic survey of prompt engineering on vision-language foundation models. arXiv preprint arXiv:2307.12980 (2023) + +14. Guo, H., Du, B., Zhang, L., Su, X.: A coarse-to-fine boundary refinement network for building footprint extraction from remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing 183, 240-252 (2022) +15. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019) +16. Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2021) +17. Hu, X., Zhu, L., Fu, C.W., Qin, J., Heng, P.A.: Direction-aware spatial context features for shadow detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7454-7462 (2018) +18. Ji, S., Wei, S., Lu, M.: Fully convolutional networks for multisource building extraction from an open aerial and satellite imagery data set. IEEE Transactions on geoscience and remote sensing 57(1), 574-586 (2018) +19. Ji, W., Li, J., Bi, Q., Li, W., Cheng, L.: Segment anything is not always perfect: An investigation of sam on different real-world applications. arXiv preprint arXiv:2304.05750 (2023) +20. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022) +21. Ke, L., Ye, M., Danelljan, M., Liu, Y., Tai, Y.W., Tang, C.K., Yu, F.: Segment anything in high quality. arXiv preprint arXiv:2306.01567 (2023) +22. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19113-19122 (2023) +23. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) +24. Lan, X., Lyu, J., Jiang, H., Dong, K., Niu, Z., Zhang, Y., Xue, J.: Foodsam: Any food segmentation. IEEE Transactions on Multimedia pp. 1-14 (2023). https://doi.org/10.1109/TMM.2023.3330047 +25. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021) +26. Li, F., Zhang, H., Sun, P., Zou, X., Liu, S., Yang, J., Li, C., Zhang, L., Gao, J.: Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767 (2023) +27. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021) +28. Liew, J.H., Cohen, S., Price, B., Mai, L., Feng, J.: Deep interactive thin object selection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 305-314 (2021) +29. Liu, W., Shen, X., Pun, C.M., Cun, X.: Explicit visual prompting for low-level structure segmentations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19434-19445 (2023) +30. Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature Communications 15(1), 654 (2024) +31. Miletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks for volumetric medical image segmentation. In: 2016 fourth international conference on 3D vision (3DV). pp. 565-571. IEEE (2016) + +32. Mnih, V.: Machine Learning for Aerial Image Labeling. Ph.D. thesis, University of Toronto (2013) +33. Nagendra, S., Kifer, D.: Patchrefinenet: Improving binary segmentation by incorporating signals from optimal patch-wise binarization. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1361-1372 (2024) +34. Pogorelov, K., Randel, K.R., Griwodz, C., Eskeland, S.L., de Lange, T., Johansen, D., Spampinato, C., Dang-Nguyen, D.T., Lux, M., Schmidt, P.T., et al.: Kvasir: A multi-class image dataset for computer aided gastrointestinal disease detection. In: Proceedings of the 8th ACM on Multimedia Systems Conference. pp. 164-169 (2017) +35. Qin, X., Dai, H., Hu, X., Fan, D.P., Shao, L., Van Gool, L.: Highly accurate dichotomous image segmentation. In: European Conference on Computer Vision. pp. 38-56. Springer (2022) +36. Rahman, M.M., Marculescu, R.: Medical image segmentation via cascaded attention decoding. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6222-6231 (2023) +37. Rebuffi, S.A., Bilen, H., Vedaldi, A.: Learning multiple visual domains with residual adapters. Advances in neural information processing systems 30 (2017) +38. Rebuffi, S.A., Bilen, H., Vedaldi, A.: Efficient parametrization of multi-domain deep neural networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8119-8127 (2018) +39. Shiraishi, J., Katsuragawa, S., Ikezoe, J., Matsumoto, T., Kobayashi, T., Komatsu, K.i., Matsui, M., Fujita, H., Kodera, Y., Doi, K.: Development of a digital image database for chest radiographs with and without a lung nodule: receiver operating characteristic analysis of radiologists' detection of pulmonary nodules. American Journal of Roentgenology 174(1), 71-74 (2000) +40. Singh, D., Valdenegro-Toro, M.: The marine debris dataset for forward-looking sonar semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3741-3749 (2021) +41. Song, Y., Zhou, Q., Lu, X., Shao, Z., Ma, L.: Simada: A simple unified framework for adapting segment anything model in underperformed scenes. arXiv preprint arXiv:2401.17803 (2024) +42. Vicente, T.F.Y., Hou, L., Yu, C.P., Hoai, M., Samaras, D.: Large-scale training of shadow detectors with noisily-annotated shadow examples. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI 14. pp. 816-832. Springer (2016) +43. Wang, H., Vasu, P.K.A., Faghri, F., Vemulapalli, R., Farajtabar, M., Mehta, S., Rastegari, M., Tuzel, O., Pouransari, H.: Sam-clip: Merging vision foundation models towards semantic and spatial understanding. arXiv preprint arXiv:2310.15308 (2023) +44. Wang, L., Fang, S., Meng, X., Li, R.: Building extraction with vision transformer. IEEE Transactions on Geoscience and Remote Sensing 60, 1-11 (2022) +45. Wei, S., Zeng, X., Qu, Q., Wang, M., Su, H., Shi, J.: Hrsid: A high-resolution sar images dataset for ship detection and instance segmentation. IEEE Access 8, 120234-120254 (2020) +46. Wu, J., Fu, R., Fang, H., Liu, Y., Wang, Z., Xu, Y., Jin, Y., Arbel, T.: Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620 (2023) + +47. Wu, Q., Zhang, Y., Elbatel, M.: Self-prompting large vision models for few-shot medical image segmentation. In: MICCAI Workshop on Domain Adaptation and Representation Transfer. pp. 156-167. Springer (2023) +48. Xiong, Y., Varadarajan, B., Wu, L., Xiang, X., Xiao, F., Zhu, C., Dai, X., Wang, D., Sun, F., Iandola, F., et al.: Efficientsam: Leveraged masked image pretraining for efficient segment anything. arXiv preprint arXiv:2312.00863 (2023) +49. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2945-2954 (2023) +50. Yang, H., Ma, C., Wen, B., Jiang, Y., Yuan, Z., Zhu, X.: Recognize any regions. arXiv preprint arXiv:2311.01373 (2023) +51. Zeng, Y., Zhang, P., Zhang, J., Lin, Z., Lu, H.: Towards high-resolution salient object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7234-7243 (2019) +52. Zhang, C., Han, D., Qiao, Y., Kim, J.U., Bae, S.H., Lee, S., Hong, C.S.: Faster segment anything: Towards lightweight sam for mobile applications. arXiv preprint arXiv:2306.14289 (2023) +53. Zhang, K., Liu, D.: Customized segment anything model for medical image segmentation. arXiv preprint arXiv:2304.13785 (2023) +54. Zhang, R., Jiang, Z., Guo, Z., Yan, S., Pan, J., Dong, H., Gao, P., Li, H.: Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048 (2023) +55. Zhao, X., Ding, W., An, Y., Du, Y., Yu, T., Li, M., Tang, M., Wang, J.: Fast segment anything. arXiv preprint arXiv:2306.12156 (2023) +56. Zheng, Q., Qiao, X., Cao, Y., Lau, R.W.: Distraction-aware shadow detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5167-5176 (2019) +57. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022) +58. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022) +59. Zhu, L., Deng, Z., Hu, X., Fu, C.W., Xu, X., Qin, J., Heng, P.A.: Bidirectional feature pyramid network with recurrent attention residual modules for shadow detection. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 121-136 (2018) \ No newline at end of file diff --git a/2024/CAT-SAM_ Conditional Tuning for Few-Shot Adaptation of Segment Anything Model/images.zip b/2024/CAT-SAM_ Conditional Tuning for Few-Shot Adaptation of Segment Anything Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7e8cf4eb6fde88893dcb075a124b3e7cad62ebc5 --- /dev/null +++ b/2024/CAT-SAM_ Conditional Tuning for Few-Shot Adaptation of Segment Anything Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd0483104b414a0ab0d33d8d635aa4b850c0f98cc7c7a8a8d20e9bd98cf893df +size 517546 diff --git a/2024/CAT-SAM_ Conditional Tuning for Few-Shot Adaptation of Segment Anything Model/layout.json b/2024/CAT-SAM_ Conditional Tuning for Few-Shot Adaptation of Segment Anything Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ae84131021ba7744347e5176e131fb1078a90e09 --- /dev/null +++ b/2024/CAT-SAM_ Conditional Tuning for Few-Shot Adaptation of Segment Anything Model/layout.json @@ -0,0 +1,8366 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 111, + 463, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 111, + 463, + 148 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 463, + 148 + ], + "type": "text", + "content": "CAT-SAM: Conditional Tuning for Few-Shot Adaptation of Segment Anything Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "spans": [ + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "content": "Aoran Xiao" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "content": ", Weihao Xuan" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{2,3*}" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "content": ", Heli Qi" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "content": ", Yun Xing" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "content": ", Ruijie Ren" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "content": ", Xiaojin Zhang" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "content": ", Ling Shao" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "content": ", and Shijian Lu" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 209, + 201, + 403, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 201, + 403, + 235 + ], + "spans": [ + { + "bbox": [ + 209, + 201, + 403, + 235 + ], + "type": "text", + "content": "1 Nanyang Technological University, Singapore \n2 The University of Tokyo, Japan \n3 RIKEN AIP, Japan" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 203, + 235, + 410, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 235, + 410, + 257 + ], + "spans": [ + { + "bbox": [ + 203, + 235, + 410, + 257 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 203, + 235, + 410, + 257 + ], + "type": "text", + "content": " Nara Institute of Science and Technology, Japan \n" + }, + { + "bbox": [ + 203, + 235, + 410, + 257 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 203, + 235, + 410, + 257 + ], + "type": "text", + "content": " Waseda University, Japan" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 217, + 257, + 395, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 257, + 395, + 268 + ], + "spans": [ + { + "bbox": [ + 217, + 257, + 395, + 268 + ], + "type": "text", + "content": "6 Zhejiang University of Technology, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 147, + 268, + 465, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 268, + 465, + 278 + ], + "spans": [ + { + "bbox": [ + 147, + 268, + 465, + 278 + ], + "type": "text", + "content": "UCAS-Terminus AI Lab, University of Chinese Academy of Sciences, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 300, + 453, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 300, + 453, + 542 + ], + "spans": [ + { + "bbox": [ + 159, + 300, + 453, + 542 + ], + "type": "text", + "content": "Abstract. The Segment Anything Model (SAM) has demonstrated remarkable zero-shot capability and flexible geometric prompting in general image segmentation. However, it often struggles in domains that are either sparsely represented or lie outside its training distribution, such as aerial, medical, and non-RGB images. Recent efforts have predominantly focused on adapting SAM to these domains using fully supervised methods, which necessitate large amounts of annotated training data and pose practical challenges in data collection. This paper presents CAT-SAM, a ConditionAI Tuning network that explores few-shot adaptation of SAM toward various challenging downstream domains in a data-efficient manner. The core design is a prompt bridge structure that enables decoder-conditioned joint tuning of the heavyweight image encoder and the lightweight mask decoder. The bridging maps the domain-specific features of the mask decoder to the image encoder, fostering synergic adaptation of both components with mutual benefits with few-shot target samples only, ultimately leading to superior segmentation in various downstream tasks. We develop two CAT-SAM variants that adopt two tuning strategies for the image encoder: one injecting learnable prompt tokens in the input space and the other inserting lightweight adapter networks. Extensive experiments over 11 downstream tasks show that CAT-SAM achieves superior segmentation consistently even under the very challenging one-shot adaptation setup." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 552, + 453, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 552, + 453, + 574 + ], + "spans": [ + { + "bbox": [ + 159, + 552, + 453, + 574 + ], + "type": "text", + "content": "Keywords: Segment Anything Model " + }, + { + "bbox": [ + 159, + 552, + 453, + 574 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 552, + 453, + 574 + ], + "type": "text", + "content": " Few-shot Learning " + }, + { + "bbox": [ + 159, + 552, + 453, + 574 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 552, + 453, + 574 + ], + "type": "text", + "content": " Parameter Efficient Tuning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 591, + 230, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 591, + 230, + 604 + ], + "spans": [ + { + "bbox": [ + 132, + 591, + 230, + 604 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 613, + 481, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 613, + 481, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 613, + 481, + 638 + ], + "type": "text", + "content": "Accurate image segmentation is a fundamental computer vision task that plays a pivotal role in various applications such as robotics, autonomous driving, health-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 643, + 421, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 643, + 421, + 654 + ], + "spans": [ + { + "bbox": [ + 141, + 643, + 421, + 654 + ], + "type": "text", + "content": "* Co-first authors with equal contributions. † Corresponding author." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 654, + 409, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 409, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 409, + 665 + ], + "type": "text", + "content": "Project page: https://xiaooran.github.io/projects/CAT-SAM" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 196, + 119, + 419, + 224 + ], + "blocks": [ + { + "bbox": [ + 196, + 119, + 419, + 224 + ], + "lines": [ + { + "bbox": [ + 196, + 119, + 419, + 224 + ], + "spans": [ + { + "bbox": [ + 196, + 119, + 419, + 224 + ], + "type": "image", + "image_path": "069a80822340aee613fa1fe128d81549102ab37092b55803c0a0ff70fd76f468.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 237, + 482, + 283 + ], + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 283 + ], + "type": "text", + "content": "Fig. 1: The proposed CAT-SAM performs ConditionAl joint Tuning (CAT) to establish communication between SAM's heavyweight image encoder and lightweight mask decoder. This enables synergistic adaptation of the two network components, mitigating tuning imbalances and improving few-shot SAM adaptation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 306, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 482, + 376 + ], + "type": "text", + "content": "care, earth observation, etc. The recently developed Segment Anything Model (SAM) [23], trained with 1.1 billion masks, has emerged as a momentous leap forward in image segmentation. By taking geometric prompts with points, boxes, or masks as input, SAM demonstrates remarkable zero-shot capability for general image segmentation, as well as great potential for fine-grained mask segmentation in different downstream tasks across a variety of contexts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 378, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 521 + ], + "type": "text", + "content": "On the other hand, SAM often fails to generate high-quality predictions when dealing with domains that are either sparsely represented or lie outside its training distribution [5,19], such as challenging RGB images including aerial, medical, and intricate structural images, as well as non-RGB images like X-ray, Sonar, Synthetic Aperture Radar (SAR) images, etc. This greatly undermines the versatility and applicability of SAM as a foundational segmentation model while handling various downstream tasks. Several studies attempted to address this issue by fine-tuning SAM with a large number of target samples with mask annotations [21, 30]. However, they require extensive masked images for each specialized field, posing a formidable task with poor scalability. How to effectively adapt SAM toward various downstream tasks in a data-efficient manner has become an essential challenge in the segmentation field." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 640 + ], + "type": "text", + "content": "We exploit few-shot target samples for effective and efficient adaptation of SAM toward various downstream tasks. Inspired by the recent advancements in parameter-efficient tuning for foundation models [1,13], we freeze the entire SAM and lightly expand its image encoder and mask decoder with a tiny amount of learnable parameters. This expansion preserves SAM's zero-shot capabilities and flexibility while capturing domain-specific features that are essential for the segmentation of downstream data. However, SAM's image encoder is much larger than its mask decoder, which directly leads to tuning imbalance and further suboptimal adaptation, especially under the limited supervision of few-shot target samples." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "To address this challenge, we design CAT-SAM, a novel ConditionAl Tuning network for effective and data-efficient adaptation of SAM. The key idea is to ex" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "type": "text", + "content": "A. Xiao and W. Xuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "ploit the lightweight decoder tuning to guide the heavyweight encoder tuning and formulate decoder-conditioned joint tuning of both as depicted in Fig. 1. To this end, we design a prompt bridge structure that maps the domain-specific features from the mask decoder to the image encoder. With the collaboration between the two partners, CAT-SAM enables synergistic adaptation of both which mitigates the tuning imbalance and improves the few-shot SAM adaptation effectively. In addition, the prompt bridge can be seamlessly embedded into two representative tuning approaches, including prompt tuning [20, 25, 27] that introduces learnable prompt tokens into the input space and adapters [15, 29, 37, 38] that leverage lightweight adaptive networks. This directly leads to two CAT-SAM variants, namely, CAT-SAM-T and CAT-SAM-A. Extensive evaluations over 11 diverse and challenging downstream tasks show that both variants achieve superior adaptation and segmentation with few-shot target samples only." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 272, + 483, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 483, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 483, + 416 + ], + "type": "text", + "content": "The major contributions of this work can be summarized in three aspects. First, we propose CAT-SAM, a conditional tuning network that enables effective and data-efficient adaptation of SAM toward various challenging downstream domains. We design prompt bridge within CAT-SAM, a decoder-conditioned joint tuning structure that enables synergistic and data-efficient adaptation of the heavyweight image encoder and the lightweight mask decoder effectively. Second, we develop two CAT-SAM variants by embedding the prompt bridge into two representative tuning strategies, one introducing learnable prompt tokens in the input space and the other inserting lightweight adapter networks. Third, extensive experiments over 11 diverse segmentation datasets show that CAT-SAM achieves superior image segmentation consistently even under the challenging one-shot setup." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 435, + 238, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 238, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 238, + 449 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 461, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 616 + ], + "type": "text", + "content": "Parameter-Efficient Tuning for foundation models has become increasingly crucial due to the over-size model parameters and costly one-task-one-model deployment. One popular approach involves updating only the newly added learnable parameters while keeping the model backbone frozen. These lightweight parameters are often introduced in two ways: 1) as prompt tuning [12,20,22,58] by adding learnable tokens to input tokens at transformer layers, 2) as adapters [11, 29,37,38,49] by integrating learnable lightweight sub-networks into transformer layers. However, prior visual studies primarily focus on adapting single backbone modules [3,20,29] or dual encoder backbones for vision-language tasks [22,57], which are suboptimal tuning solutions while applying to SAM due to its particular architecture of a heavyweight image encoder and a lightweight mask decoder. In contrast, we explore a novel conditional joint tuning strategy tailored to SAM, which facilitates its adaptation even with limited downstream data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "SAM for Downstream Tasks. The release of SAM has spurred several subsequent studies. Some concentrate on expanding semantic recognition capabilities [24, 26, 43, 50, 54], while others aim to create more lightweight SAM variants [48, 52, 55] for faster computations. Several studies explore SAM's adapta" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 402, + 90, + 447, + 101 + ], + "type": "text", + "content": "CAT-SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "tion in underperforming downstream scenarios, primarily by directly fine-tuning SAM's mask decoder [4,8,30,41,46,53], which can severely degrade SAM's zero-shot capabilities. The recent HQ-SAM [21] freezes the entire pre-trained SAM and prompt-tunes the mask decoder with a new mask head for adaptive mask prediction. However, these adaptation studies rely on large target datasets with a considerable number of annotated images, which are often costly to collect. In contrast, our CAT-SAM achieves superior adaptation with only few-shot target samples, significantly reducing adaptation costs and expanding the applicability of SAM as a visual foundation model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 241, + 202, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 241, + 202, + 254 + ], + "spans": [ + { + "bbox": [ + 132, + 241, + 202, + 254 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 264, + 271, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 264, + 271, + 276 + ], + "spans": [ + { + "bbox": [ + 132, + 264, + 271, + 276 + ], + "type": "text", + "content": "3.1 Preliminaries of SAM" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 283, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 452 + ], + "type": "text", + "content": "SAM [23] handles each image with three key modules: a heavyweight image encoder (i.e., ViT [9] as the backbone) that extracts image embeddings, a prompt encoder that encodes the geometric prompts (i.e., points, a box, or a coarse mask) to generate prompt embeddings, and a lightweight mask decoder that combines the two types of embeddings to predict segmentation masks. The released SAM model is trained on a super large-scale SA-1B dataset, which consists of over 11 million images as well as 1.1 billion automatically generated masks. SAM has demonstrated remarkable zero-shot capability while dealing with various conventional natural images, as well as superb flexibility in accepting various geometric prompt inputs. On the other hand, it often struggles while facing challenging downstream domains that are often sparse or falling outside of its training data distribution, such as RGB domains of aerial, medical, and intricate structural images, as well as non-RGB domains of X-ray, SAR, and Sonar images. More details about SAM are presented in [23]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 467, + 216, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 467, + 216, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 467, + 216, + 479 + ], + "type": "text", + "content": "3.2 CAT-SAM" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 485, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 545 + ], + "type": "text", + "content": "This subsection introduces CAT-SAM, a ConditionAl Tuning network that is designed for adapting SAM toward various under-performed downstream data. The objective is to adapt SAM efficiently with only few-shot annotated target images, meanwhile preserving SAM's powerful zero-shot capability and geometric prompting flexibility." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "CAT-SAM freezes the entire SAM and facilitates the simultaneous tuning of SAM's image encoder and mask decoder to capture representative information from few-shot target samples for effective adaptation of SAM. However, a critical issue with simultaneous tuning arises from the clear imbalance between the heavyweight image encoder (308.3 M parameters for ViT-L [9]) and the lightweight mask decoder (4.1 M), which often leads to sub-optimal adaptation while only few-shot target samples are available. To tackle this challenge, we design decoder-conditioned joint tuning, a novel tuning strategy that mitigates the imbalance by establishing a linkage between the encoding tuning and the decoding tuning. Specifically, we design prompt bridge—a lightweight network that" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "type": "text", + "content": "A. Xiao and W. Xuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 118, + 233, + 222 + ], + "blocks": [ + { + "bbox": [ + 135, + 118, + 233, + 222 + ], + "lines": [ + { + "bbox": [ + 135, + 118, + 233, + 222 + ], + "spans": [ + { + "bbox": [ + 135, + 118, + 233, + 222 + ], + "type": "image", + "image_path": "fc430856e4daded764b4acbdea330eb06029422ea75a3c539c5fdffc38bcae2b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 224, + 199, + 230 + ], + "lines": [ + { + "bbox": [ + 167, + 224, + 199, + 230 + ], + "spans": [ + { + "bbox": [ + 167, + 224, + 199, + 230 + ], + "type": "text", + "content": "(a) CAT-SAM-T" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 244, + 118, + 342, + 222 + ], + "blocks": [ + { + "bbox": [ + 244, + 118, + 342, + 222 + ], + "lines": [ + { + "bbox": [ + 244, + 118, + 342, + 222 + ], + "spans": [ + { + "bbox": [ + 244, + 118, + 342, + 222 + ], + "type": "image", + "image_path": "cd1359f00a47f4f792f48395e146e58919584e31308d1a27be1a36b51cea6189.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 224, + 308, + 230 + ], + "lines": [ + { + "bbox": [ + 276, + 224, + 308, + 230 + ], + "spans": [ + { + "bbox": [ + 276, + 224, + 308, + 230 + ], + "type": "text", + "content": "(b) CAT-SAM-A" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 352, + 118, + 481, + 221 + ], + "blocks": [ + { + "bbox": [ + 352, + 118, + 481, + 221 + ], + "lines": [ + { + "bbox": [ + 352, + 118, + 481, + 221 + ], + "spans": [ + { + "bbox": [ + 352, + 118, + 481, + 221 + ], + "type": "image", + "image_path": "62da2b3ef65a27621f55538a957a9411f2ef56f2935339eafdca266b94a62478.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 224, + 433, + 230 + ], + "lines": [ + { + "bbox": [ + 399, + 224, + 433, + 230 + ], + "spans": [ + { + "bbox": [ + 399, + 224, + 433, + 230 + ], + "type": "text", + "content": "(c) Prompt Bridge" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 243, + 480, + 331 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 480, + 331 + ], + "type": "text", + "content": "Fig. 2: Overview of CAT-SAM. CAT-SAM keeps the whole SAM frozen while simultaneously tuning the image encoder and mask decoder for downstream adaptation. To address the tuning imbalance between these two network components, we introduce decoder-conditioned joint tuning through the design of Prompt Bridge structure, enabling synergetic and enhanced adaptation. We present two CAT-SAM variants: CAT-SAM-T in (a) and CAT-SAM-A in (b), achieved by integrating the prompt bridge with prompt-based and adapter-based tuning strategies for the image encoder, respectively. (c) illustrates two tailored prompt bridge structures, PB-T and PB-A." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 357, + 479, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 479, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 479, + 403 + ], + "type": "text", + "content": "maps the domain-specific features from the mask decoder to the image encoder. The bridge design facilitates joint and balanced tuning of both network components, leading to synergistic and effective adaptation of SAM with few-shot and even a single target sample." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 405, + 479, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 479, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 479, + 440 + ], + "type": "text", + "content": "We develop two CAT-SAM variants by integrating the proposed decoder-conditioned joint tuning into two prevalent tuning approaches, namely CAT-SAM-T and CAT-SAM-A, as illustrated in Fig. 2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 461, + 479, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 479, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 479, + 497 + ], + "type": "text", + "content": "Tuning Image Encoder. We first brief SAM's image encoder and then introduce two tuning approaches in our CAT-SAM, including prompt tuning in CAT-SAM-T and lightweight adapters in CAT-SAM-A." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 499, + 480, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 499, + 480, + 559 + ], + "spans": [ + { + "bbox": [ + 138, + 499, + 480, + 559 + ], + "type": "text", + "content": "- SAM's Image Encoder, denoted as " + }, + { + "bbox": [ + 138, + 499, + 480, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 138, + 499, + 480, + 559 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 138, + 499, + 480, + 559 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 138, + 499, + 480, + 559 + ], + "type": "text", + "content": " transformer layers " + }, + { + "bbox": [ + 138, + 499, + 480, + 559 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{V}_i\\}_{i=1}^K" + }, + { + "bbox": [ + 138, + 499, + 480, + 559 + ], + "type": "text", + "content": ", first splits the input image " + }, + { + "bbox": [ + 138, + 499, + 480, + 559 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 138, + 499, + 480, + 559 + ], + "type": "text", + "content": " of the size " + }, + { + "bbox": [ + 138, + 499, + 480, + 559 + ], + "type": "inline_equation", + "content": "(H, W)" + }, + { + "bbox": [ + 138, + 499, + 480, + 559 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 138, + 499, + 480, + 559 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 138, + 499, + 480, + 559 + ], + "type": "text", + "content": " fixed-size patches, and then projects them into patch embeddings " + }, + { + "bbox": [ + 138, + 499, + 480, + 559 + ], + "type": "inline_equation", + "content": "E_0 \\in \\mathbb{R}^{M \\times d_v}" + }, + { + "bbox": [ + 138, + 499, + 480, + 559 + ], + "type": "text", + "content": ". Patch embeddings " + }, + { + "bbox": [ + 138, + 499, + 480, + 559 + ], + "type": "inline_equation", + "content": "E_{i-1}" + }, + { + "bbox": [ + 138, + 499, + 480, + 559 + ], + "type": "text", + "content": " are input to the " + }, + { + "bbox": [ + 138, + 499, + 480, + 559 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 138, + 499, + 480, + 559 + ], + "type": "text", + "content": " transformer layer " + }, + { + "bbox": [ + 138, + 499, + 480, + 559 + ], + "type": "inline_equation", + "content": "(\\mathcal{V}_i)" + }, + { + "bbox": [ + 138, + 499, + 480, + 559 + ], + "type": "text", + "content": " and sequentially processed through " + }, + { + "bbox": [ + 138, + 499, + 480, + 559 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 138, + 499, + 480, + 559 + ], + "type": "text", + "content": " transformer layers:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 244, + 571, + 384, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 571, + 384, + 585 + ], + "spans": [ + { + "bbox": [ + 244, + 571, + 384, + 585 + ], + "type": "interline_equation", + "content": "[ E _ {i} ] = \\mathcal {V} _ {i} ([ E _ {i - 1} ]), i = 1, 2, \\dots , K.", + "image_path": "0a6ac82ba662a846f2c846b07dbff4b6b79191a2f0b810824cef09d54e3abbec.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 594, + 479, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 594, + 479, + 642 + ], + "spans": [ + { + "bbox": [ + 138, + 594, + 479, + 642 + ], + "type": "text", + "content": "- CAT-SAM-T: As illustrated in Fig. 2 (a), we introduce a set of learnable tokens " + }, + { + "bbox": [ + 138, + 594, + 479, + 642 + ], + "type": "inline_equation", + "content": "\\{P_i \\in \\mathbb{R}^{d_v}\\}_{i=0}^{K-1}" + }, + { + "bbox": [ + 138, + 594, + 479, + 642 + ], + "type": "text", + "content": " in the image encoder alongside the input patch embeddings. These new learnable tokens are introduced in each transformer layer of the image encoder " + }, + { + "bbox": [ + 138, + 594, + 479, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 138, + 594, + 479, + 642 + ], + "type": "text", + "content": " and updated through the adaptation process:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 231, + 652, + 398, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 652, + 398, + 666 + ], + "spans": [ + { + "bbox": [ + 231, + 652, + 398, + 666 + ], + "type": "interline_equation", + "content": "[ E _ {i} ] = \\mathcal {V} _ {i} ([ E _ {i - 1}, P _ {i - 1} ]) ], i = 1, 2, \\dots , K.", + "image_path": "909d83765796820e6adffa47836c0e2f20aededa4c6b36af9dc626b6fac73706.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "CAT-SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "text", + "content": "- CAT-SAM-A: As illustrated in Fig. 2 (b), we introduce Adapters [15] to tune SAM's image encoder in CAT-SAM-A. The adapter is a lightweight sub-network that is inserted into each transformer layer, comprising a linear down-projection, a nonlinear activation function, a linear up-projection, and a residual connection. High-frequency image information is also incorporated in CAT-SAM as the input of the adapters as in [29]. In this process, high-frequency components " + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "inline_equation", + "content": "I_{hfc}" + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "text", + "content": " are first extracted from the input image " + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "text", + "content": " and then reversed to the space domain using the Fast Fourier Transform and its inverse. These components are then partitioned into small patches " + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "inline_equation", + "content": "I_{hfc}^p" + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "text", + "content": " similar to the image patches. Subsequently, convolutional layers and linear layers project " + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "inline_equation", + "content": "I_{hfc}" + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "text", + "content": " and the original image embeddings " + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "inline_equation", + "content": "E_0" + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "text", + "content": "-dimensional features independently, yielding " + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "inline_equation", + "content": "F_{hfc}" + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "inline_equation", + "content": "F_{pe}" + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "text", + "content": ", respectively. The result of element-wise addition of " + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "inline_equation", + "content": "F_{pe}" + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "inline_equation", + "content": "F_{hfc}" + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "text", + "content": " is then forwarded (as denoted \"EVP\" in Fig. 2 (b)) to the adapter " + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "inline_equation", + "content": "\\mathrm{Adapt}_i" + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "text", + "content": " for transformer layer " + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "text", + "content": ". Finally, the output is added element-wise to the original input of the transformer layer " + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 138, + 115, + 482, + 296 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 236, + 302, + 394, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 302, + 394, + 316 + ], + "spans": [ + { + "bbox": [ + 236, + 302, + 394, + 316 + ], + "type": "interline_equation", + "content": "E _ {i - 1} = E _ {i - 1} + \\mathrm {A d a p t} _ {i} (F _ {p e} + F _ {h f c}).", + "image_path": "9dc31d9d283d1608140c4ab1164bc03f949db639337625aa7a67a2469f468c39.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 329, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 472 + ], + "type": "text", + "content": "Tuning Mask Decoder. We tune the mask decoder of SAM similarly as [21]. Specifically, we freeze the entire pre-trained decoder of SAM and introduce a learnable CAT-Token that is concatenated with SAM's original output tokens and prompt tokens to form the input to SAM's mask decoder. The CAT-Token is updated during training but remains fixed during inference, the same as SAM's output tokens. After passing through two decoder layers, the updated CAT-Token is used to generate dynamic weights for a newly introduced three-layer MLP. Simultaneously, SAM's mask decoder features are fused with the image features decoded by the image encoder. Finally, the fused features and the three-layer MLP are combined through the dot product to generate the target mask. Please refer to the appendix for more details about the decoder structures in CAT-SAM." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 488, + 484, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 488, + 484, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 488, + 484, + 523 + ], + "type": "text", + "content": "Decoder-Conditioned Joint Tuning. The proposed decoder-conditioned joint tuning can be seamlessly embedded via two implementations of the proposed prompt bridging ideas, i.e., PB-T and PB-A as illustrated in Fig. 2 (c)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 524, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 524, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 138, + 524, + 482, + 583 + ], + "type": "text", + "content": "- CAT-SAM-T: The prompt bridge " + }, + { + "bbox": [ + 138, + 524, + 482, + 583 + ], + "type": "inline_equation", + "content": "\\mathrm{PB}_i^T" + }, + { + "bbox": [ + 138, + 524, + 482, + 583 + ], + "type": "text", + "content": " is a two-layer MLP. It projects the CAT-Token " + }, + { + "bbox": [ + 138, + 524, + 482, + 583 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 138, + 524, + 482, + 583 + ], + "type": "text", + "content": " in the mask decoder to each transformer layer " + }, + { + "bbox": [ + 138, + 524, + 482, + 583 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 138, + 524, + 482, + 583 + ], + "type": "text", + "content": " in the image encoder and outputs a single learnable token replacing " + }, + { + "bbox": [ + 138, + 524, + 482, + 583 + ], + "type": "inline_equation", + "content": "P_i" + }, + { + "bbox": [ + 138, + 524, + 482, + 583 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 138, + 524, + 482, + 583 + ], + "type": "inline_equation", + "content": "\\tilde{P}_i^T = \\mathrm{PB}_i^T(Q)" + }, + { + "bbox": [ + 138, + 524, + 482, + 583 + ], + "type": "text", + "content": ". These mapped tokens are directly applied for prompt tuning of the image encoder as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 232, + 589, + 399, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 589, + 399, + 604 + ], + "spans": [ + { + "bbox": [ + 232, + 589, + 399, + 604 + ], + "type": "interline_equation", + "content": "\\left[ E _ {i} \\right] = \\mathcal {V} _ {i} \\left(\\left[ E _ {i - 1}, \\tilde {P} _ {i - 1} ^ {T} \\right]\\right), i = 1, 2, \\dots , K.", + "image_path": "1735c22a89afab9e0181a61fe5bd1a288eb5b8c758aca41d44a0a20fbc41d94e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 139, + 610, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 610, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 139, + 610, + 482, + 647 + ], + "type": "text", + "content": "- CAT-SAM-A: The prompt bridge " + }, + { + "bbox": [ + 139, + 610, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathrm{PB}^A" + }, + { + "bbox": [ + 139, + 610, + 482, + 647 + ], + "type": "text", + "content": " maps the CAT-Token " + }, + { + "bbox": [ + 139, + 610, + 482, + 647 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 139, + 610, + 482, + 647 + ], + "type": "text", + "content": " into an embedding " + }, + { + "bbox": [ + 139, + 610, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\tilde{P}^A = \\mathrm{PB}^A(Q)" + }, + { + "bbox": [ + 139, + 610, + 482, + 647 + ], + "type": "text", + "content": " of the same size as " + }, + { + "bbox": [ + 139, + 610, + 482, + 647 + ], + "type": "inline_equation", + "content": "F_{pe}" + }, + { + "bbox": [ + 139, + 610, + 482, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 139, + 610, + 482, + 647 + ], + "type": "inline_equation", + "content": "F_{hfc}" + }, + { + "bbox": [ + 139, + 610, + 482, + 647 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 139, + 610, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\tilde{P}^A" + }, + { + "bbox": [ + 139, + 610, + 482, + 647 + ], + "type": "text", + "content": " is then element-wise added to the EVP before being fed to every adapter as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 222, + 651, + 408, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 651, + 408, + 667 + ], + "spans": [ + { + "bbox": [ + 222, + 651, + 408, + 667 + ], + "type": "interline_equation", + "content": "E _ {i - 1} = E _ {i - 1} + \\mathrm {A d a p t} _ {i} (F _ {p e} + F _ {h f c} + \\tilde {P} ^ {A}).", + "image_path": "8ad4c7f95b6f0733c99c89d8c0835d3b81558dcdae16ee81be3ea79629e8f31d.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "type": "text", + "content": "A. Xiao and W. Xuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 146, + 476, + 205 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 1: Comparisons of trainable parameters between ViT-L [9] based SAM and two variants of the proposed CAT-SAM." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 146, + 476, + 205 + ], + "lines": [ + { + "bbox": [ + 135, + 146, + 476, + 205 + ], + "spans": [ + { + "bbox": [ + 135, + 146, + 476, + 205 + ], + "type": "table", + "html": "
ModelTrainable Parameters (M)Trainable Params vs. SAM
SAM [23]312.4-
CAT-SAM-T3.31.1%
CAT-SAM-A1.90.6%
", + "image_path": "030ebc0460d94393032ec2222157b42fd9a533b1e1e22884b4c2cdfa62d20b11.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 147, + 226, + 481, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 226, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 147, + 226, + 481, + 274 + ], + "type": "text", + "content": "As illustrated in Fig. 2 (c), PB-A employs " + }, + { + "bbox": [ + 147, + 226, + 481, + 274 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 147, + 226, + 481, + 274 + ], + "type": "text", + "content": " different linear down-projection layers to generate embeddings independently. After GELU layers, these embeddings are up-projected to the dimension of " + }, + { + "bbox": [ + 147, + 226, + 481, + 274 + ], + "type": "inline_equation", + "content": "1 \\times M" + }, + { + "bbox": [ + 147, + 226, + 481, + 274 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 147, + 226, + 481, + 274 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 147, + 226, + 481, + 274 + ], + "type": "text", + "content": " is the spatial size of the patch) with a shared linear layer to mitigate the computation cost." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 147, + 274, + 460, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 274, + 460, + 286 + ], + "spans": [ + { + "bbox": [ + 147, + 274, + 460, + 286 + ], + "type": "text", + "content": "Finally, the up-projected embeddings are concatenated to produce " + }, + { + "bbox": [ + 147, + 274, + 460, + 286 + ], + "type": "inline_equation", + "content": "\\tilde{P}^A" + }, + { + "bbox": [ + 147, + 274, + 460, + 286 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 286, + 481, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 286, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 481, + 346 + ], + "type": "text", + "content": "Tab. 1 compares the amount of trainable parameters between the SAM (implemented with the ViT-L [9]) and our CAT-SAM that is built upon the SAM. It can be observed that both CAT-SAM variants introduce very limited additional parameters, yet yield significant performance improvements in various downstream segmentation tasks as to be detailed in Section 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 363, + 276, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 276, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 276, + 376 + ], + "type": "text", + "content": "3.3 Training and Inference" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 383, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 383, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 383, + 482, + 502 + ], + "type": "text", + "content": "During training, we feed target samples together with geometric prompts into CAT-SAM to generate CAT-SAM masks, under the supervision of ground-truth masks as well as a linear combination of BCE loss and dice loss [31]. We freeze the pre-trained SAM parameters and update solely the parameters that are introduced in the above-described tuning modules. Note we train CAT-SAM with a mixture of sampled geometric prompts, including bounding boxes, randomly selected points, and coarse masks. In addition, we create degraded coarse masks by introducing random Gaussian noise in the boundary regions of the ground-truth masks as in [21]. Further training details can be found in Section 4.1 and the appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 502, + 482, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 588 + ], + "type": "text", + "content": "During the inference phase, the input image is fed into SAM's image encoder along with the integrated prompt tokens in CAT-SAM-T or adapters in CAT-SAM-A, to generate adaptive image embeddings. These embeddings, combined with the prompt tokens from SAM's prompt encoder, serve as input to the mask decoder. Subsequently, the updated CAT-token and associated MLP are utilized for the target mask prediction. Lastly, we up-sample the mask to the original resolution to produce the final output." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 605, + 230, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 230, + 618 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 230, + 618 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "We evaluate CAT-SAM's efficacy through comprehensive experiments involving eight segmentation tasks across 11 datasets, all hailing from challenging downstream fields that SAM struggles to address effectively. Our experiments encom" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 402, + 90, + 447, + 101 + ], + "type": "text", + "content": "CAT-SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 147, + 496, + 270 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 2: We evaluate and benchmark CAT-SAM over eight challenging segmentation tasks across 11 datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 147, + 496, + 270 + ], + "lines": [ + { + "bbox": [ + 133, + 147, + 496, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 147, + 496, + 270 + ], + "type": "table", + "html": "
TasksDataset NamesImagery
Building segmentationWHU [18]Aerial images
Road segmentationMA. Roads [32]Aerial images
Polyp segmentationKvasir [34]Medical images
Chest X-ray segmentationJSRT [39]X-ray images
Marine debris segmentationFLS [40]Sonar images
Ship instance segmentationHRSID [45]SAR images
Shadow segmentationSBU-Shadow [42]Natural images
Intricate segmentationDIS [35],ThinObject [28],HRSOD [51],COIFT [28]Natural images
", + "image_path": "f29f503d8b0287e3d3765e9780558716ff6899eac8abf3348b057bce8af41b66.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 291, + 480, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 480, + 316 + ], + "type": "text", + "content": "pass a wide array of scenarios and tasks, spanning from varying target shots to different imagery modalities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 330, + 261, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 261, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 261, + 342 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 346, + 486, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 486, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 486, + 453 + ], + "type": "text", + "content": "Datasets. We conduct experiments on two collections of datasets listed in Tab. 2: 1) True-color images, including WHU [18] for building segmentation, Kvasir [34] for polyp segmentation, SBU-Shadow [42] for shadow segmentation, Massachusetts (MA.) Roads [32] for road segmentation, as well as fine-grained segmentation datasets DIS [35], ThinObject [28], HRSOD [51], and COIFT [28]. 2) Non-RGB domains, encompassing JSRT [39] for chest organ segmentation (X-ray images), HRSID [45] for ship instance segmentation (high-resolution SAR images), and FLS [40] for marine debris segmentation (Sonar images). We use official dataset splits for fair comparisons." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 454, + 481, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 454, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 454, + 481, + 525 + ], + "type": "text", + "content": "Evaluation Metrics. For single-class datasets including WHU, Kvasir, SBU-Shadow, and MA. Roads, we employ standard mask IoU of foreground. JSRT and FLS, which have multiple classes, are evaluated using individual class IoU and their average. As for DIS, ThinObject, HRSOD, and COIFT, we follow [21] and report mIoU and the boundary metrics (mBIoU) [7] for fair comparisons. For instance segmentation of HRSID, we use standard AP, " + }, + { + "bbox": [ + 130, + 454, + 481, + 525 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 130, + 454, + 481, + 525 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 454, + 481, + 525 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{75}" + }, + { + "bbox": [ + 130, + 454, + 481, + 525 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 526, + 481, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 526, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 526, + 481, + 611 + ], + "type": "text", + "content": "Implementation Details. Unless otherwise specified, CAT-SAM and its comparing models use ViT-L [9] as the image encoder backbone. We train on 1 NVIDIA RTX A6000 GPU for one-shot adaptation, while 4 GPUs for multiple-shots. Following [21], ground truth boxes serve as the default input geometric prompts during evaluation to ensure fair comparison and minimize randomness, with the exception of Fig. 3 where point prompts are evaluated. Please refer to the appendix for more implementation details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 624, + 246, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 246, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 246, + 635 + ], + "type": "text", + "content": "4.2 Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "We first investigate the contributions of different tuning modules within both CAT-SAM variants to assess their impact on overall adaptation performance." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "type": "text", + "content": "A. Xiao and W. Xuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 213, + 310, + 297 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 192 + ], + "type": "text", + "content": "Table 3: Ablation study of CAT-SAM-T in (a) and CAT-SAM-A in (b) for 1-shot adaptation over datasets: \"W\" for WHU [18] (on building segmentation), \"K\" for Kvasir [34] (on polyp segmentation), and \"S\" for SBU-Shadow [42] (on shadow segmentation). \"Enc-T.\" and \"Enc-A.\" means tuning the image encoder with prompt tokens and adapters, respectively. \"Dec.\" means tuning the mask decoder. \"PB-T\" and \"PB-A\" denote two customized prompt bridges for conditional tuning. Model 1 is the baseline SAM without adaptation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 189, + 202, + 260, + 213 + ], + "lines": [ + { + "bbox": [ + 189, + 202, + 260, + 213 + ], + "spans": [ + { + "bbox": [ + 189, + 202, + 260, + 213 + ], + "type": "text", + "content": "(a) CAT-SAM-T." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 213, + 310, + 297 + ], + "lines": [ + { + "bbox": [ + 138, + 213, + 310, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 213, + 310, + 297 + ], + "type": "table", + "html": "
Enc-T. Dec. PB-TWKSAVG
143.579.062.461.6
2✓57.880.476.0
3✓71.273.863.5
4✓✓66.473.261.8
5✓✓✓86.883.4
", + "image_path": "47913fa1d14db37ea3662c230ca49d1529987a168de0244127baaeff6b17adac.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 313, + 213, + 485, + 297 + ], + "blocks": [ + { + "bbox": [ + 363, + 202, + 434, + 213 + ], + "lines": [ + { + "bbox": [ + 363, + 202, + 434, + 213 + ], + "spans": [ + { + "bbox": [ + 363, + 202, + 434, + 213 + ], + "type": "text", + "content": "(b) CAT-SAM-A." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 213, + 485, + 297 + ], + "lines": [ + { + "bbox": [ + 313, + 213, + 485, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 485, + 297 + ], + "type": "table", + "html": "
Enc-A. Dec. PB-AWKSAVG
143.579.062.461.6
2✓86.573.776.3
3✓71.273.863.5
4✓✓83.479.876.3
5✓✓✓88.285.4
", + "image_path": "a4d92c61b23d4638a303b3fee5752a7d24ba44e57f01a0d34366c4871332a3b8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 320, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 416 + ], + "type": "text", + "content": "Tab. 3 shows the experiments of one-shot adaptation across three true-color image segmentation benchmarks WHU [18], Kvasir [34], and SBU-Shadow [42]. For each CAM-SAM variant CAM-SAM-T/CAM-SAM-A, we compare 5 models including: 1. The original SAM without adaptation (baseline); 2. Tuning SAM's image encoder alone with prompt tokens or adapters; 3. Tuning SAM's mask decoder alone; 4. Tuning SAM's image encoder and mask decoder independently; and 5. Conditional tuning of SAM's image encoder and mask decoder with the proposed prompt bridging (i.e., the complete CAT-SAM-T and CAT-SAM-A)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 418, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 561 + ], + "type": "text", + "content": "We can see that the original SAM struggles to produce high-quality masks in these challenging domains, indicating its limitations as a foundational segmentation model for many downstream applications. Tuning the image encoder or the mask decoder alone in 2 and 3 enhances target segmentation, underlining the efficacy and necessity of SAM adaptation. However, simultaneously tuning the image encoder and mask decoder in an independent way in model 4 does not show complementary and consistent improvement. This suggests instability and imbalance in tuning between the heavyweight image encoder and the lightweight mask decoder, which could lead to suboptimal adaptation solutions. Differently, including the proposed prompt bridging with either PB-T or PB-A (i.e., the complete CAT-SAM-T and CAT-SAM-A) mitigates the tuning imbalance and produces substantial performance enhancement consistently." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 582, + 353, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 582, + 353, + 594 + ], + "spans": [ + { + "bbox": [ + 131, + 582, + 353, + 594 + ], + "type": "text", + "content": "4.3 Comparison with the State-of-the-Art" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "CAT-SAM for One-Shot Adaptation. To assess CAT-SAM's effectiveness in reducing downstream training data and its adaptability with few-shot samples, we first evaluate it under the extremely challenging one-shot adaptation scenario. Given the lack of prior studies under such setup, we comprehensively benchmark it against closely related methods that can be broadly categorized into" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "CAT-SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 179, + 481, + 362 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 4: Comparison of adaptive segmentation performance on challenging true-color datasets: WHU [18] (on building segmentation), Kvasir [34] (on Polyp segmentation), and SBU-Shadow [42] (on shadow segmentation). The baseline is SAM without adaptation. All compared methods utilize one-shot sample for adaptation, except for those denoted by * which utilize 20 shots with metric numbers sourced from [47]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 179, + 481, + 362 + ], + "lines": [ + { + "bbox": [ + 133, + 179, + 481, + 362 + ], + "spans": [ + { + "bbox": [ + 133, + 179, + 481, + 362 + ], + "type": "table", + "html": "
MethodsWHUKvasirSBU-ShadowAverage
SAM [23] (baseline)43.579.062.461.6
VPT-shallow [20]60.879.868.769.8
VPT-deep [20]57.880.476.071.4
AdaptFormer [3]83.276.877.279.1
LoRA [16]86.177.574.479.3
Med-SA [46]34.528.622.128.4
MedSAM [30]30.629.853.437.9
SimAda [41]48.618.749.338.9
SPFS-SAM* [47] (20-shots)-53.4--
SAMed* [53] (20-shots)-51.7--
HQ-SAM [21]71.273.863.569.5
CAT-SAM-T (Ours)86.883.478.082.7
CAT-SAM-A (Ours)88.285.481.985.2
", + "image_path": "1090a075e0542cb192d665fff8076f46685cf3e5d88d6541ed9623286673c818.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 396, + 481, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 468 + ], + "type": "text", + "content": "two groups. The first group comprises state-of-the-art tuning methods for foundation models of different modalities, including VPT [20] and AdaptFormer [3] for large vision foundation models, and LoRA [16] for large language models. The second group comprises recent SAM-based adaptation methods which were largely designed for full-shot adaptation [21,30,41,46,53] ( [47] tackles few-shot adaptation but requires 20 shots)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 474, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 474, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 474, + 482, + 665 + ], + "type": "text", + "content": "Tab. 4 presents results over WHU, Kvasir, and SBU-Shadow datasets. We can observe that all three tuning methods for foundation models exhibit clear adaptation effects. Notably, AdaptFormer and LoRA demonstrate significantly better performance than VPT-deep and VPT-shallow, highlighting the effectiveness of modifying the model's structure over prompt tuning in the input space during SAM adaptation. On the other hand, most SAM-based adaptation methods, including Med-SA, SimAda, SPFS-SAM, SAMed, and MedSAM, yield much lower performance, primarily due to their reliance on many more target images which leads to overfitting while learning from one-shot target sample. HQ-SAM achieves substantial segmentation improvements but it still cannot compete with the three tuning methods, demonstrating the significance of tuning the image encoder over the mask decoder. Finally, both CAT-SAM variants consistently outperform all compared methods, with CAT-SAM-A exhibiting a slight edge due to its network structure modifications. These experiments validate the superiority of the proposed decoder-conditioned joint tuning approach in extracting domain-specific features with few-shot samples." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "type": "text", + "content": "A. Xiao and W. Xuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 190, + 485, + 347 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 5: Comparison of state-of-the-art segmentation achieved by fully supervised learning (FSL) methods trained on full-shot data (upper part) and mask segmentation by SAM and our CAT-SAM (lower part). Evaluation follows official criteria of each FSL benchmark, reporting IoU for WHU, mIoU for Kvasir, and BER for SBU-Shadow to ensure consistency. \"(1)\", \"(16)\", and \"(full)\" after CAT-SAM-T/CAT-SAM-A denote tuning with 1-shot, 16-shots, and full-shots samples." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 190, + 485, + 347 + ], + "lines": [ + { + "bbox": [ + 132, + 190, + 485, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 190, + 485, + 347 + ], + "type": "table", + "html": "
WHUKvasirSBU-Shadow
ModelIoUModelmIoUModelBER(↓)
STT [2]90.5CASCADE [36]87.8BDRAR [59]3.64
BuildFormer [44]91.4PatchRefineNet [33]89.1DSDNet [56]3.45
CBR-Net [14]91.4DUCK-Net [10]90.5MTMT [6]3.15
SAM43.5SAM87.3SAM13.22
CAT-SAM-T (1)86.8CAT-SAM-T (1)90.0CAT-SAM-T (1)7.81
CAT-SAM-A (1)88.2CAT-SAM-A (1)91.3CAT-SAM-A (1)5.27
CAT-SAM-T (16)89.6CAT-SAM-T (16)93.1CAT-SAM-T (16)4.04
CAT-SAM-A (16)90.3CAT-SAM-A (16)93.6CAT-SAM-A (16)3.80
CAT-SAM-T (full)93.3CAT-SAM-T (full)94.5CAT-SAM-T (full)2.54
CAT-SAM-A (full)93.6CAT-SAM-A (full)94.3CAT-SAM-A (full)2.39
", + "image_path": "e88927a2f36f45a1c40543bcab6081cffa0744a0c86cb9c31a0f354b62dd555a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 373, + 483, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 483, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 483, + 517 + ], + "type": "text", + "content": "Full-Shot Learning. While CAT-SAM is primarily designed for few-shot adaptation, it also demonstrates proficiency in full-shot adaptation scenarios. In Tab. 5, we present the state-of-the-art performances of traditional fully-supervised learning (FSL) methods trained over full-shot data and compare them with SAM and our CAT-SAM (adapted with various target samples). We report official evaluation metrics for three public benchmarks, including IoU for WHU, mIoU for Kvasir, and BER [17] for SBU-Shadow. Note the distinction in setups between the FSL segmentation and the mask segmentation: FSL focuses on traditional semantic segmentation of specific classes without prompt input while SAM and CAT-SAM involve general mask segmentation with prompt input. Despite these differences, these performances can serve as valuable references for evaluating CAT-SAM's segmentation ability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 519, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 519, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 482, + 628 + ], + "type": "text", + "content": "With full target training data, both CAT-SAM variants outperform state-of-the-art FSL methods consistently. Notably, the two CAT-SAM variants achieve impressive target segmentation under the challenging one-shot setup with a single target sample. With 16 shots of target samples, their performance is even on par with the FSL methods that are trained by using full target training data. These experiments underscore the superiority of effective transfer of broad knowledge from the foundation model SAM as compared with training from scratch with FSL. It also demonstrates the remarkable efficacy of CAT-SAM in substantially expanding SAM's applicability across various downstream tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "We conducted further experiments to compare CAT-SAM against HQ-SAM [21], a state-of-the-art SAM-based adaptation solution for intricate structural image segmentation. For fair comparisons, we follow the evaluation protocol of" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "CAT-SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 168, + 481, + 248 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 158 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 158 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 158 + ], + "type": "text", + "content": "Table 6: Extremely fine-grained segmentation over four intricate structural image datasets. HQ-SAM and the two CAT-SAM variants are fine-tuned over full HQSeg-44K [21]. All adopt the boxes converted from their GT masks as the box prompt input for fair comparisons." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 168, + 481, + 248 + ], + "lines": [ + { + "bbox": [ + 133, + 168, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 168, + 481, + 248 + ], + "type": "table", + "html": "
ModelsDIS [35]COIFT [28]HRSOD [51]ThinObject [28]Average
mIoUmBIoUmIoUmBIoUmIoUmBIoUmIoUmBIoUmIoUmBIoU
SAM [23]62.052.892.186.590.283.173.661.879.571.1
HQ-SAM [21]78.670.494.890.193.686.989.579.989.181.8
CAT-SAM-T84.078.195.692.093.487.694.087.991.786.4
CAT-SAM-A83.677.795.692.293.187.194.188.291.686.3
", + "image_path": "68682ba8097a04ead049c12213791d9b67dc23edd850f2eabf936913ec79ff54.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 270, + 480, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 378 + ], + "type": "text", + "content": "HQ-SAM and tune CAT-SAM on the complete HQSeg-44K dataset, consisting of 44,320 meticulously annotated image masks. Tab. 6 shows experiments across four exceptionally fine-grained segmentation datasets including DIS [35], ThinObject [28], HRSOD [51], and COIFT [28] (in mIoU and the boundary metrics (mBIoU) [7]). We can see that CAT-SAM-T achieves slightly better performance than CAT-SAM-A, while both clearly outperform HQ-SAM and the original SAM. These experiments further demonstrate the efficacy of our proposed conditional tuning in adapting SAM toward challenging downstream tasks. Visual qualitative comparisons are provided in the appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 394, + 334, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 394, + 334, + 406 + ], + "spans": [ + { + "bbox": [ + 133, + 394, + 334, + 406 + ], + "type": "text", + "content": "4.4 CAT-SAM for Non-RGB Domains" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 414, + 480, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 544 + ], + "type": "text", + "content": "We also explore CAT-SAM's performance across more challenging image modalities characterized by larger domain discrepancies. Specifically, we assess its efficacy on non-RGB domains, including JSRT [39] for chest organ segmentation using X-ray images and FLS [40] for marine debris segmentation with Sonar images, and HRSID [45] for ship instance segmentation with SAR images. For adaptation, we randomly select either one or sixteen target images for 1-shot and 16-shot scenarios, respectively, while employing the entire training set for full-shot adaptation. These datasets represent novel image distributions as compared with SAM's training data. Note we train instances of all classes together while evaluating them by classes. Since no images in FLS contain objects of all 11 classes, we only report results of 16-shot and full-shot adaptation for FLS." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "type": "text", + "content": "Tab. 7 shows experimental results. Notably, CAT-SAM-T and CAT-SAM-A improve the mIoU to " + }, + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "type": "inline_equation", + "content": "93.0\\%" + }, + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "type": "inline_equation", + "content": "92.6\\%" + }, + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "type": "text", + "content": ", respectively, with only one shot of JSRT training image. The improvements saturate gradually as the number of training samples increases, with marginal gains of " + }, + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "type": "text", + "content": ", respectively, when adapted with full-shot target data. These findings underscore the superior few-shot capability of the proposed CAT-SAM and align well with those reported in Tab. 5. Experiments on FLS [40] show different results. While both CAT-SAM variants exhibit clear adaptation effects, the performance gains for both 16-shot and full-shot adaptation are notably lower as compared with previously evaluated datasets. The discrepancy is largely attributed to two primary" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "type": "text", + "content": "A. Xiao and W. Xuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 179, + 475, + 312 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 7: Few-shot adaptation with varying number of training samples for non-RGB datasets, including FLS [40] with Sonar images for marine debris segmentation, JSRT [39] for chest X-ray segmentation, and HRSID [45] for ship instance segmentation with SAR images. \"N/A\" denotes that 1-shot adaptation over FLS is unattainable due to the absence of images containing objects from all 11 classes." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 179, + 475, + 312 + ], + "lines": [ + { + "bbox": [ + 136, + 179, + 475, + 312 + ], + "spans": [ + { + "bbox": [ + 136, + 179, + 475, + 312 + ], + "type": "table", + "html": "
#SamplesMethodsJSRT [39]FLS [40]HRSID [45]
LungsHeartmIoUmIoUAPAP50AP75
NoneSAM [23]85.071.978.569.738.286.626.8
1-shotCAT-SAM-T95.490.593.0N/A46.088.445.4
CAT-SAM-A94.790.592.6N/A44.988.041.4
16-shotCAT-SAM-T95.892.794.273.246.289.345.2
CAT-SAM-A95.591.593.571.445.789.443.7
Full-shotCAT-SAM-T96.392.694.481.751.493.055.0
CAT-SAM-A96.492.894.682.052.993.956.0
", + "image_path": "7fe6d1c848573f55a82cf00700efdf7b4d336f9433ce39d0d0c1a0abce40395f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 335, + 482, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 455 + ], + "type": "text", + "content": "factors: 1) the inherent difficulty of FLS with its 11 semantic classes (per-class performances are provided in appendix); and 2) the larger variation of FLS data in shapes and object sizes as compared with previously evaluated data. As for ship instance segmentation using SAR images in HRSID [45], this dataset presents different challenges as ships are very small in size. The results show that CAT-SAM consistently outperforms SAM in this diverse domain, highlighting its efficiency in domain adaptation with large domain discrepancy. These experiments further underscore the need for a more comprehensive adaptation of SAM for exceptionally challenging downstream tasks. We expect further studies along this research direction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 475, + 272, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 272, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 272, + 487 + ], + "type": "text", + "content": "4.5 Analysis & Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 334, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 334, + 642 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 334, + 642 + ], + "type": "text", + "content": "Prompts with Single Point. SAM provides flexibility in geometric prompts including a single foreground point. However, prompting with a single foreground point can be challenging as the single point could correspond to multiple objects [23]. Nonetheless, adaptation with few-shot target samples can alleviate this ambiguity by guiding the model to focus on the specific foreground distribution within the annotation space. We evaluate this by using the same randomly-selected single point as the geometric prompt, and Fig. 3 shows" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "the segmentation results for images from WHU, Kvasir, SBU-Shadow, and MA. Roads [32] on road segmentation." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 340, + 523, + 479, + 602 + ], + "blocks": [ + { + "bbox": [ + 340, + 523, + 479, + 602 + ], + "lines": [ + { + "bbox": [ + 340, + 523, + 479, + 602 + ], + "spans": [ + { + "bbox": [ + 340, + 523, + 479, + 602 + ], + "type": "image", + "image_path": "918a78860252361b79e735130a6795a93b496d59c3d6b858d0076014486d2089.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 610, + 480, + 632 + ], + "lines": [ + { + "bbox": [ + 339, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 339, + 610, + 480, + 632 + ], + "type": "text", + "content": "Fig. 3: Single point to mask evaluation for one-shot adaptation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "CAT-SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 117, + 479, + 217 + ], + "blocks": [ + { + "bbox": [ + 134, + 117, + 479, + 217 + ], + "lines": [ + { + "bbox": [ + 134, + 117, + 479, + 217 + ], + "spans": [ + { + "bbox": [ + 134, + 117, + 479, + 217 + ], + "type": "image", + "image_path": "288cfba52fa1f43e027ad782621ea1f43be7e63611c172717c612dbb21c2cb98.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 226, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 316 + ], + "type": "text", + "content": "Fig. 4: Visual comparisons of SAM [23] (top row) and CAT-SAM (bottom row). We illustrate samples from WHU [18] for building segmentation, MA. Roads [32] for road segmentation, SBU-Shadow [42] for shadow segmentation, Kvasir [34] for polyp segmentation, JSRT [39] for chest organ segmentation (X-ray images), FLS [40] for marine debris segmentation (Sonar images), and HRSID [45] for ship instance segmentation (SAR images). CAT-SAM exhibits one-shot adaptation across most datasets, except for 16-shot over FLS. Red boxes and stars denote geometric prompts, colored regions are mask predictions, and lines show the boundary of ground truth segmentation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 340, + 482, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 435 + ], + "type": "text", + "content": "Similar to the experiments using boxes as prompts, CAT-SAM consistently demonstrates superior target segmentation while using a single point as prompt, underscoring its robustness in adaptation while preserving SAM's inherent geometric prompting flexibility. In addition, both CAT-SAM variants achieve relatively less improvement for the dataset Kvasir. This can be largely attributed to the high similarities in color and texture between the polyps and the local gastrointestinal regions in Kvasir, making it challenging for a single point to provide clear geometric guidance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 436, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 496 + ], + "type": "text", + "content": "Visual Illustrations. Fig. 4 shows qualitative comparisons between SAM and CAT-SAM across multiple segmentation tasks. These illustrations demonstrate how our proposed CAT-SAM remarkably enhances the mask quality of SAM under the presence of only one-shot target samples. Please refer to the appendix for more visual comparisons." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 517, + 220, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 517, + 220, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 220, + 530 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "We introduce CAT-SAM, a conditional tuning network tailored for few-shot adaptation of SAM. We propose decoder-conditional joint tuning to mitigate the tuning imbalance between SAM's heavyweight image encoder and lightweight mask decoder and facilitate efficient SAM adaptation. To achieve this, we design the prompt bridge structures, enabling interaction when tuning these two network modules. We develop two CAT-SAM variants by integrating the prompt bridge with two representative tuning strategies including prompt tuning and adapter. Our comprehensive evaluation across diverse scenarios and tasks on 11 segmentation datasets underscores the superior domain adaptive efficiency of both CAT-SAM variants, even in the extremely challenging one-shot scenario." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "type": "text", + "content": "A. Xiao and W. Xuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 200 + ], + "type": "text", + "content": "This project is partly funded by the Ministry of Education Singapore, under the Tier-1 scheme with project number RG18/22 and the Tier-2 scheme with project number MOE-T2EP20123-0003. It is also supported in part by the National Natural Science Foundation of China [grant no. U2033210], in part by the Zhejiang Provincial Natural Science Foundation [grant no. LDT23F02024F02]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 217, + 197, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 197, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 197, + 228 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 239, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 239, + 480, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 480, + 273 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 480, + 273 + ], + "type": "text", + "content": "1. Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., et al.: On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 274, + 480, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 274, + 480, + 295 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 480, + 295 + ], + "type": "text", + "content": "2. Chen, K., Zou, Z., Shi, Z.: Building extraction from remote sensing images with sparse token transformers. Remote Sensing 13(21), 4441 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 295, + 480, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 295, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 295, + 480, + 327 + ], + "type": "text", + "content": "3. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 327, + 480, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 327, + 480, + 371 + ], + "spans": [ + { + "bbox": [ + 138, + 327, + 480, + 371 + ], + "type": "text", + "content": "4. Chen, T., Zhu, L., Deng, C., Cao, R., Wang, Y., Zhang, S., Li, Z., Sun, L., Zang, Y., Mao, P.: Sam-adapter: Adapting segment anything in underperformed scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 3367-3375 (October 2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 371, + 480, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 371, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 138, + 371, + 480, + 415 + ], + "type": "text", + "content": "5. Chen, T., Zhu, L., Ding, C., Cao, R., Zhang, S., Wang, Y., Li, Z., Sun, L., Mao, P., Zang, Y.: Sam fails to segment anything?—sam-adapter: Adapting sam in underperformed scenes: Camouflage, shadow, and more. arXiv preprint arXiv:2304.09148 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 415, + 480, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 480, + 448 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 480, + 448 + ], + "type": "text", + "content": "6. Chen, Z., Zhu, L., Wan, L., Wang, S., Feng, W., Heng, P.A.: A multi-task mean teacher for semi-supervised shadow detection. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 5611-5620 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 448, + 480, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 448, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 448, + 480, + 491 + ], + "type": "text", + "content": "7. Cheng, B., Girshick, R., Dollar, P., Berg, A.C., Kirillov, A.: Boundary iou: Improving object-centric image segmentation evaluation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15334-15342 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 491, + 480, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 513 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 513 + ], + "type": "text", + "content": "8. Cheng, J., Ye, J., Deng, Z., Chen, J., Li, T., Wang, H., Su, Y., Huang, Z., Chen, J., Jiang, L., et al.: Sam-med2d. arXiv preprint arXiv:2308.16184 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 514, + 480, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 556 + ], + "type": "text", + "content": "9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 557, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 557, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 557, + 480, + 578 + ], + "type": "text", + "content": "0. Dumitru, R.G., Peteleaza, D., Craciun, C.: Using duck-net for polyp image segmentation. Scientific Reports 13(1), 9803 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 578, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 611 + ], + "type": "text", + "content": "1. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision pp. 1-15 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "text", + "content": "2. Gong, Z., Li, F., Deng, Y., Bhattacharjee, D., Zhu, X., Ji, Z.: Coda: Instructive chain-of-domain adaptation with severity-aware visual prompt tuning (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "3. Gu, J., Han, Z., Chen, S., Beirami, A., He, B., Zhang, G., Liao, R., Qin, Y., Tresp, V., Torr, P.: A systematic survey of prompt engineering on vision-language foundation models. arXiv preprint arXiv:2307.12980 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "CAT-SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "14. Guo, H., Du, B., Zhang, L., Su, X.: A coarse-to-fine boundary refinement network for building footprint extraction from remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing 183, 240-252 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "15. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 481, + 215 + ], + "type": "text", + "content": "16. Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 481, + 248 + ], + "type": "text", + "content": "17. Hu, X., Zhu, L., Fu, C.W., Qin, J., Heng, P.A.: Direction-aware spatial context features for shadow detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7454-7462 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 133, + 249, + 481, + 281 + ], + "type": "text", + "content": "18. Ji, S., Wei, S., Lu, M.: Fully convolutional networks for multisource building extraction from an open aerial and satellite imagery data set. IEEE Transactions on geoscience and remote sensing 57(1), 574-586 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 281, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 281, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 133, + 281, + 481, + 314 + ], + "type": "text", + "content": "19. Ji, W., Li, J., Bi, Q., Li, W., Cheng, L.: Segment anything is not always perfect: An investigation of sam on different real-world applications. arXiv preprint arXiv:2304.05750 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "text", + "content": "20. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision. pp. 709-727. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 369 + ], + "type": "text", + "content": "21. Ke, L., Ye, M., Danelljan, M., Liu, Y., Tai, Y.W., Tang, C.K., Yu, F.: Segment anything in high quality. arXiv preprint arXiv:2306.01567 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 402 + ], + "type": "text", + "content": "22. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19113-19122 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 435 + ], + "type": "text", + "content": "23. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "24. Lan, X., Lyu, J., Jiang, H., Dong, K., Niu, Z., Zhang, Y., Xue, J.: Foodsam: Any food segmentation. IEEE Transactions on Multimedia pp. 1-14 (2023). https://doi.org/10.1109/TMM.2023.3330047" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "text", + "content": "25. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "text", + "content": "26. Li, F., Zhang, H., Sun, P., Zou, X., Liu, S., Yang, J., Li, C., Zhang, L., Gao, J.: Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 522, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 544 + ], + "type": "text", + "content": "27. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "text", + "content": "28. Liew, J.H., Cohen, S., Price, B., Mai, L., Feng, J.: Deep interactive thin object selection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 305-314 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "text", + "content": "29. Liu, W., Shen, X., Pun, C.M., Cun, X.: Explicit visual prompting for low-level structure segmentations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19434-19445 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "text", + "content": "30. Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature Communications 15(1), 654 (2024)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "31. Miletari, F., Navab, N., Ahmadi, S.A.: V-net: Fully convolutional neural networks for volumetric medical image segmentation. In: 2016 fourth international conference on 3D vision (3DV). pp. 565-571. IEEE (2016)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "type": "text", + "content": "A. Xiao and W. Xuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 139 + ], + "type": "text", + "content": "32. Mnih, V.: Machine Learning for Aerial Image Labeling. Ph.D. thesis, University of Toronto (2013)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 183 + ], + "type": "text", + "content": "33. Nagendra, S., Kifer, D.: Patchrefinenet: Improving binary segmentation by incorporating signals from optimal patch-wise binarization. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1361-1372 (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 184, + 482, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 184, + 482, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 184, + 482, + 239 + ], + "type": "text", + "content": "34. Pogorelov, K., Randel, K.R., Griwodz, C., Eskeland, S.L., de Lange, T., Johansen, D., Spampinato, C., Dang-Nguyen, D.T., Lux, M., Schmidt, P.T., et al.: Kvasir: A multi-class image dataset for computer aided gastrointestinal disease detection. In: Proceedings of the 8th ACM on Multimedia Systems Conference. pp. 164-169 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 239, + 482, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 274 + ], + "type": "text", + "content": "35. Qin, X., Dai, H., Hu, X., Fan, D.P., Shao, L., Van Gool, L.: Highly accurate dichotomous image segmentation. In: European Conference on Computer Vision. pp. 38-56. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 274, + 482, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 308 + ], + "type": "text", + "content": "36. Rahman, M.M., Marculescu, R.: Medical image segmentation via cascaded attention decoding. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6222-6231 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 308, + 482, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 329 + ], + "type": "text", + "content": "37. Rebuffi, S.A., Bilen, H., Vedaldi, A.: Learning multiple visual domains with residual adapters. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 330, + 482, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 364 + ], + "type": "text", + "content": "38. Rebuffi, S.A., Bilen, H., Vedaldi, A.: Efficient parametrization of multi-domain deep neural networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8119-8127 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 364, + 482, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 418 + ], + "type": "text", + "content": "39. Shiraishi, J., Katsuragawa, S., Ikezoe, J., Matsumoto, T., Kobayashi, T., Komatsu, K.i., Matsui, M., Fujita, H., Kodera, Y., Doi, K.: Development of a digital image database for chest radiographs with and without a lung nodule: receiver operating characteristic analysis of radiologists' detection of pulmonary nodules. American Journal of Roentgenology 174(1), 71-74 (2000)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 419, + 482, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 482, + 453 + ], + "type": "text", + "content": "40. Singh, D., Valdenegro-Toro, M.: The marine debris dataset for forward-looking sonar semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3741-3749 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 453, + 482, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 486 + ], + "type": "text", + "content": "41. Song, Y., Zhou, Q., Lu, X., Shao, Z., Ma, L.: Simada: A simple unified framework for adapting segment anything model in underperformed scenes. arXiv preprint arXiv:2401.17803 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 487, + 482, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 531 + ], + "type": "text", + "content": "42. Vicente, T.F.Y., Hou, L., Yu, C.P., Hoai, M., Samaras, D.: Large-scale training of shadow detectors with noisily-annotated shadow examples. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI 14. pp. 816-832. Springer (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 531, + 482, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 482, + 575 + ], + "type": "text", + "content": "43. Wang, H., Vasu, P.K.A., Faghri, F., Vemulapalli, R., Farajtabar, M., Mehta, S., Rastegari, M., Tuzel, O., Pouransari, H.: Sam-clip: Merging vision foundation models towards semantic and spatial understanding. arXiv preprint arXiv:2310.15308 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 576, + 482, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 576, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 576, + 482, + 597 + ], + "type": "text", + "content": "44. Wang, L., Fang, S., Meng, X., Li, R.: Building extraction with vision transformer. IEEE Transactions on Geoscience and Remote Sensing 60, 1-11 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 599, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 599, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 599, + 482, + 632 + ], + "type": "text", + "content": "45. Wei, S., Zeng, X., Qu, Q., Wang, M., Su, H., Shi, J.: Hrsid: A high-resolution sar images dataset for ship detection and instance segmentation. IEEE Access 8, 120234-120254 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "text", + "content": "46. Wu, J., Fu, R., Fang, H., Liu, Y., Wang, Z., Xu, Y., Jin, Y., Arbel, T.: Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "CAT-SAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 501 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "47. Wu, Q., Zhang, Y., Elbatel, M.: Self-prompting large vision models for few-shot medical image segmentation. In: MICCAI Workshop on Domain Adaptation and Representation Transfer. pp. 156-167. Springer (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "48. Xiong, Y., Varadarajan, B., Wu, L., Xiang, X., Xiao, F., Zhu, C., Dai, X., Wang, D., Sun, F., Iandola, F., et al.: Efficientsam: Leveraged masked image pretraining for efficient segment anything. arXiv preprint arXiv:2312.00863 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "49. Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X.: Side adapter network for open-vocabulary semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2945-2954 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "50. Yang, H., Ma, C., Wen, B., Jiang, Y., Yuan, Z., Zhu, X.: Recognize any regions. arXiv preprint arXiv:2311.01373 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "51. Zeng, Y., Zhang, P., Zhang, J., Lin, Z., Lu, H.: Towards high-resolution salient object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7234-7243 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "type": "text", + "content": "52. Zhang, C., Han, D., Qiao, Y., Kim, J.U., Bae, S.H., Lee, S., Hong, C.S.: Faster segment anything: Towards lightweight sam for mobile applications. arXiv preprint arXiv:2306.14289 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 303, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 480, + 325 + ], + "type": "text", + "content": "53. Zhang, K., Liu, D.: Customized segment anything model for medical image segmentation. arXiv preprint arXiv:2304.13785 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 347 + ], + "type": "text", + "content": "54. Zhang, R., Jiang, Z., Guo, Z., Yan, S., Pan, J., Dong, H., Gao, P., Li, H.: Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "text", + "content": "55. Zhao, X., Ding, W., An, Y., Du, Y., Yu, T., Li, M., Tang, M., Wang, J.: Fast segment anything. arXiv preprint arXiv:2306.12156 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "56. Zheng, Q., Qiao, X., Cao, Y., Lau, R.W.: Distraction-aware shadow detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5167-5176 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 402, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 480, + 435 + ], + "type": "text", + "content": "57. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "text", + "content": "58. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 456, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 480, + 501 + ], + "type": "text", + "content": "59. Zhu, L., Deng, Z., Hu, X., Fu, C.W., Xu, X., Qin, J., Heng, P.A.: Bidirectional feature pyramid network with recurrent attention residual modules for shadow detection. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 121-136 (2018)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 280, + 101 + ], + "type": "text", + "content": "A. Xiao and W. Xuan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CAT_ Enhancing Multimodal Large Language Model to Answer Questions in Dynamic Audio-Visual Scenarios/825b02cb-2d21-47e3-b7b8-2d59e5594aa4_content_list.json b/2024/CAT_ Enhancing Multimodal Large Language Model to Answer Questions in Dynamic Audio-Visual Scenarios/825b02cb-2d21-47e3-b7b8-2d59e5594aa4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1e9fbf5fc9a989241dc3250a5eea74dcd09e554f --- /dev/null +++ b/2024/CAT_ Enhancing Multimodal Large Language Model to Answer Questions in Dynamic Audio-Visual Scenarios/825b02cb-2d21-47e3-b7b8-2d59e5594aa4_content_list.json @@ -0,0 +1,1807 @@ +[ + { + "type": "text", + "text": "CAT: Enhancing Multimodal Large Language Model to Answer Questions in Dynamic Audio-Visual Scenarios", + "text_level": 1, + "bbox": [ + 220, + 143, + 781, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qilang Ye $^{1}$ , Zitong Yu $^{1}$ *, Rui Shao $^{2}$ , Xinyu Xie $^{1}$ , Philip Torr $^{3}$ , and Xiaochun Cao $^{4}$", + "bbox": [ + 320, + 244, + 679, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Great Bay University", + "bbox": [ + 421, + 287, + 581, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Harbin Institute of Technology, Shenzhen", + "bbox": [ + 354, + 301, + 645, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 University of Oxford", + "bbox": [ + 423, + 316, + 576, + 329 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Shenzhen Campus of Sun Yat-sen University", + "bbox": [ + 344, + 330, + 656, + 344 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. This paper focuses on the challenge of answering questions in scenarios that are composed of rich and complex dynamic audio-visual components. Although existing Multimodal Large Language Models (MLLMs) can respond to audio-visual content, these responses are sometimes ambiguous and fail to describe specific audio-visual events. To overcome this limitation, we introduce the CAT, which enhances MLLM in three ways: 1) besides straightforwardly bridging audio and video, we design a clue aggregator that aggregates question-related clues in dynamic audio-visual scenarios to enrich the detailed knowledge required for large language models. 2) CAT is trained on a mixed multimodal dataset, allowing direct application in audio-visual scenarios. Notably, we collect an audio-visual joint instruction dataset named AVinstruct, to further enhance the capacity of CAT to model cross-semantic correlations. 3) we propose AI-assisted ambiguity-aware direct preference optimization, a strategy specialized in retraining the model to favor the non-ambiguity response and improve the ability to localize specific audio-visual objects. Extensive experimental results demonstrate that CAT outperforms existing methods on multimodal tasks, especially in Audio-Visual Question Answering (AVQA) tasks. The codes and the collected instructions will be released soon.", + "bbox": [ + 259, + 380, + 743, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Multimodal Large Language Model $\\cdot$ Audio-visual Question Answering", + "bbox": [ + 261, + 670, + 740, + 699 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 724, + 375, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The real world revolves around sound and visual information, and their combination enhances our ability to perceive the world. Similarly, the development of Multimodal Large Language Models (MLLMs) [14, 64, 67] are closely related to audio and visual. Supervised fine-tuning [35, 46, 66] with specific instruction", + "bbox": [ + 212, + 755, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author: zitong.yu@ieee.org", + "bbox": [ + 217, + 824, + 513, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/94ba68624e535d3e41b8ad7f78d5f5d088bc827f024ed4183a87c00d0b6df2b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 289, + 143, + 493, + 246 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/30c64ddfd88c267d6bebac2a77f61be0d21e375496b704fe99e8409911789191.jpg", + "image_caption": [ + "Fig. 1: Comparison between existing MLLMs and CAT Red words for incorrect response, Green words for correct response, and Gray words for useless response. Left: Most of the existing MLLMs straightforwardly bridge multimodal to large language models. Instead, CAT builds on this foundation by designing the clue aggregator to learn more detailed knowledge related to the question. Moreover, CAT constrains itself to learn a sharper response through AI-assisted ambiguity-aware direct preference optimization. Right: In comparison with audio-visual-language models Video-LLaMA [64] and ChatBridge [67], our method accurately recognizes the answers to questions with the most streamlined responses." + ], + "image_footnote": [], + "bbox": [ + 289, + 247, + 491, + 354 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c3206106d4c5640b37a232e86ee5b90c2156d52af42671be69cab6d971d2573c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 146, + 712, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "datasets empowers Large Language Models (LLMs) for multimodal understanding. Despite MLLMs' strong causal reasoning achieving impressive results in common-sense answering [53,63], academic answering [16], etc., predicting questions in dynamic audio-visual scenarios remains challenging. This is due to the difficulty of aligning LLMs with cross-domain data during training on large-scale multimodal corpora. It leads to particular ambiguity when describing specific objects in dynamic audio-visual scenarios. These interferences not only cause the model to make random guesses but also seriously affect the inference of detailed answers in Audio-Visual Question Answering (AVQA) [1,26,59,62] tasks.", + "bbox": [ + 212, + 532, + 787, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A series of bridging modules [12,20,31,32,39] for MLLMs have been designed. The simplest methods [15, 18, 28, 33] use projectors to directly align text with other modalities, but partially limit the ability to capture fine-grained information. In addition, using a cross-attention mechanism to query the audio-visual context [8,21,38,39] to solve multimodal alignment problems is effective but still occurs that a certain visual object or sound cannot be localized in practice.", + "bbox": [ + 212, + 670, + 787, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we explore possible factors contributing to the above failures: 1) Audio-visual insufficiently correlates with the question. As illustrated in the upper left of Fig. 1, most existing MLLMs [28,64,67] are mainly designed with multiple branches to handle multiple modalities individually, followed by concatenating the modality embeddings with prompts as inputs to the LLM.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Q. Ye et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paradigm cannot allow the network to learn paired text and audiovisuals, and the tandem embedding is not relevant. 2) Alignment of multimodal with text is challenging. The multimodal-text corpus [67] is difficult to be aligned, making the model generation of responses sometimes ambiguous. This ambiguity generally manifests in ambiguous words of the corresponding audio-visual content, as well as in generating useless text that overly responds. For example, the descriptions illustrated by Video-LLaMA [64] and ChatBridge [67] in Fig. 1 are all relatively ambiguous in describing the question-related video and audio content. These descriptions fail to identify the correct answer for \"bagpipes\" when asked for audio-visual reasoning. In addition, when the description is accompanied by many useless and controversy-prone words, it is unfavorable to evaluate the closed-ended AVQA task or open-ended AVQA task.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To overcome the two aforementioned issues, we introduce the CAT, enhancing MLLM in three ways: 1) Aggregation of question-related key clues. As illustrated in the lower left of Fig. 1, besides the global visual and audio information, we design a clue aggregator to dynamically capture question-aware visual and audio hidden features to enrich fine-grained clues. The aggregator receives sufficient low-level textual information to interact with audio-visual, it is capable of improving audio-visual grounding. 2) Mixed audio-visual training strategy. The training of CAT includes feature alignment using video-text pairs and audio-text pairs, and high-quality instructions to enhance audio-visual awareness. This strategy allows the CAT to be directly involved in real-world scenarios containing both visual and sound. Notably, we collect an audio-visual joint instruction dataset, named AVinstruct, to further empower CAT in AVQA tasks. 3) Retraining MLLMs to mitigate ambiguity. The DPO proposed by Rafailov et al. [52] has inspired us. Although MLLMs after training are equipped with multimodal understanding, they lack the flexibility to perceive ambiguity. Therefore, we reframe ambiguity elimination as a model preference optimization process and propose an AI-assisted Ambiguity-aware Direct Preference Optimization (ADPO) strategy. Specifically, we refer to ambiguous responses that express the lack of clarity of specific audio-visual objects as negative responses, and then we collect negative responses in the training set and utilize GPT to rewrite them into positive responses. After the multimodal training, we perform ADPO to retrain the model to bias towards the positive response, which is the precise description after the rewrite, and reject the negative response, which is the ambiguous description. Through this learning strategy, CAT can constrain itself to favor non-ambiguity responses. As shown in Fig. 1 on the right, CAT correctly responds to the question and excludes all useless information.", + "bbox": [ + 212, + 333, + 787, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 238, + 731, + 607, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We introduce a novel audio-visual-language model, dubbed as CAT, that is capable of learning question-related clues and engaging directly in dynamic audio-visual inference. Notably, we collect AVInstruct, an audio-visual joint instruction dataset to ensure the stability of CAT in AVQA tasks.", + "bbox": [ + 225, + 779, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "CAT", + "bbox": [ + 696, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- As a powerful learning strategy, we propose AI-assisted ambiguity-aware direct preference optimization to overcome the problem that MLLMs tend to ambiguously describe specific audio-visual objects.", + "bbox": [ + 223, + 146, + 785, + 190 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- We evaluate CAT on a wide range of multimodal tasks. Extensive experiments demonstrate the superiority of CAT (e.g., outperforms the state-of-the-art on a variety of AVQA tasks, and achieves remarkable results in the evaluation of video-based text generation tasks and zero-shot video question-answering tasks).", + "bbox": [ + 225, + 191, + 785, + 266 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 284, + 395, + 300 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Audio-visual question answering (AVQA). AVQA aims to produce the most accurate linguistic representation of a given video and audio based on the question content, which requires multimodal understanding and reasoning across different semantic levels. The earliest studies [57, 58, 60] emphasize the understanding of the whole video and return a simple word that is relatively correct to the question. Subsequently, the emergence of dynamic audio-visual datasets (e.g., music scenes [26, 59], and 360-degree panoramicas [62]) increases the challenge of Question-Answering (QA), which requires the mining of temporal and spatial information in different modalities. For instance, Music-AVQA [26] requires distinguishing how many instruments are present based on the audio content. AVQA [59] requires finding the most plausible of multiple options based on the audio-visual clues from dynamic scenarios.", + "bbox": [ + 212, + 311, + 787, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MLLMs for audio-visual question answering. Extending LLM to other multimodal tasks is a recently emerging field. Despite the ability of MLLMs to combine information between different modalities, performance on downstream tasks, especially AVQA, remains suboptimal. Many works [8, 20, 39] emphasize the design of elegant bridging methods to improve the performance of question answering. The simplest bridging modules [37,40,64] use one or more linear layers for feature alignment. Although such methods minimize parameter updates, they still have limitations in exploring fine-grained information. Others design more complex bridging networks to query visual information. For example, Lyu et al. [38] propose an alignment module for harmonizing different representations before entering the LLM. Ma et al. [39] propose to keep the distance between all visual tokens and any linguistic tokens consistent within the LLM. However, a large multimodal corpus is difficult to align during training, neither algorithms with low parameter counts [12] nor complex bridging networks [4] are prone to the problem of failing to accurately depict specific audio-visual events.", + "bbox": [ + 212, + 492, + 787, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human-preference learning. Reinforcement Learning from Human Feedback (RLHF) [41, 46] is the most classical instance of human preference learning [45, 55], it constructs a reward model to optimize the policy model to favor preference responses. Dai et al. [8] have demonstrated that such preference learning can enhance LLM to generate more accurate information. Recently, Rafailov et al. [52] propose a Direct Preference Optimization (DPO) strategy that learns preferences directly by bypassing learning reward models, this simple yet efficient approach inspires us to solve the ambiguity problem.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Q. Ye et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fd1da11cfd1466c864099619b84d1827ca664efd9239b548a73b98444c76f8ec.jpg", + "image_caption": [ + "Fig. 2: Illustration of the proposed CAT and its training strategy. (a) Overview of CAT. CAT first extracts overall audio-visual knowledge from video and audio and transforms them into visual tokens $x^{vid}$ and audio tokens $x^{aud}$ . We input question tagged with $<\\mathrm{Q}> $ in the prompt into the clue aggregator, aiming to aggregate question-aware audio-visual hidden features and yield clue tokens $x^{cue}$ . Finally, we merge multimodal tokens and language and feed into the frozen large language model with LoRA [17] to output the response. (b) The training paradigm of CAT involves pre-alignment of the audio-visual projectors and instruction tuning on the entire model." + ], + "image_footnote": [], + "bbox": [ + 215, + 146, + 776, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Our Approach", + "text_level": 1, + "bbox": [ + 215, + 450, + 392, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first present in detail the proposed CAT (Sec. 3.1 and 3.2). As shown on the left in Fig. 2, CAT consists of three branches that draw on visual knowledge, audio knowledge, and question-related clues to feed into the LLM, respectively. Second, we present the multimodel training strategy for CAT as shown on the right in Fig. 2, where a high-quality audio-visual joint instruction dataset is collected for further fine-tuning (Sec. 3.3). Lastly, we introduce the AI-assisted Ambiguity-aware Direct Preference Optimization (ADPO) strategy, which reinforces CAT to favor non-ambiguous descriptions (Sec. 3.4).", + "bbox": [ + 212, + 481, + 787, + 603 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Multimodal Inputs", + "text_level": 1, + "bbox": [ + 215, + 625, + 419, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ImageBind [13] with a single joint embedding space demonstrates strong modality learning ability. Therefore, we leverage the frozen ImageBind as a universal encoder for all modalities. Given a video $\\mathbf{V}$ and an audio $\\mathbf{A}$ , the encoded multimodal hidden features can be obtained by:", + "bbox": [ + 212, + 648, + 785, + 709 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh _ {v} = \\operatorname {I m a g e B i n d} (\\mathbf {V}), \\quad h _ {a} = \\operatorname {I m a g e B i n d} (\\mathbf {A}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 720, + 785, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $h_v \\in \\mathbb{R}^{T \\times d_h}$ , $h_a \\in \\mathbb{R}^{1 \\times d_h}$ are video and audio features, respectively. $T$ denotes the length of the given video and $d_h$ is the specific dimension. Notably, we temporally compress the frame-level features $h_v$ to address the computational burden. Further, we employ two linear projection layers to align the hidden dimensions of the inputs $h_v$ , $h_a$ to obtain visual tokens $x^{v_id}$ and audio tokens $x^{aud}$ , respectively.", + "bbox": [ + 212, + 748, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "CAT", + "bbox": [ + 696, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Aggregating Key Clues", + "text_level": 1, + "bbox": [ + 215, + 146, + 455, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "An important component of MLLMs is the design of efficient bridging modules. The simple approach is to apply linear layers to incorporate the full visual and audio. The advantage is that it does not require any reduction of spatio-temporal information. However, such bridge approaches sometimes make LLM fail to reflect certain content and consistently generate a chunk of description about the video [28, 64]. Therefore, we devise a Clue Aggregator (CA) that enriches LLM-acquired knowledge by mining multimodal clues related to the question. CA is divided into two steps, as illustrated in Fig. 3.", + "bbox": [ + 212, + 176, + 787, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Step1: perceiving question-related visual and audio clues. To extract", + "bbox": [ + 215, + 297, + 787, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/70516e33dc025a4fca401f3ae0947dcf8b2b7b330391af6ef41f6e05abc82d0d.jpg", + "image_caption": [ + "Fig. 3: Illustration of clue aggregator in a simple example." + ], + "image_footnote": [], + "bbox": [ + 217, + 343, + 501, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "more sufficient details from the input video and audio, we devise a clue perceiver, which consists of multiple different transformer layers. The perceiver can be viewed as a tiny transformer network that arranges the self-attention (SA), cross-attention (XA), and feed-forward network (FFN) in forward and reverse order. The forward order block $\\mathcal{B}_1$ is to perform attention-based question-aware localization. The reverse order block $\\mathcal{B}_2$ is to consolidate the original audio-visual representation at the attentional level.", + "bbox": [ + 508, + 313, + 787, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, given tokens representing \"Is there a dog around?\" and the corresponding text embedding $h_t$ :", + "bbox": [ + 212, + 523, + 785, + 555 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} _ {1} \\left(h _ {t}; X\\right) = \\mathrm {X A} \\left(h _ {t}, \\operatorname {F F N} (\\operatorname {S A} (X))\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 568, + 785, + 585 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} _ {2} (X; \\mathcal {B} _ {1} \\left(h _ {t}; X\\right)) = \\operatorname {S A} (\\text {F F N} (\\operatorname {X A} (\\operatorname {F F N} (\\operatorname {S A} (X)), \\mathcal {B} _ {1} \\left(h _ {t}, X\\right)))) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 279, + 616, + 785, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In $\\mathrm{XA}(\\cdot ,\\cdot)$ , the former represents the query and the latter represents the key and value. Based on Eqs. 2 and 3, we use $h_v^\\prime = \\mathcal{B}_2(h_v;\\mathcal{B}_1(h_t;h_v))$ and $h_a^\\prime = \\mathcal{B}_2(h_a;\\mathcal{B}_1(h_t;h_a))$ to obtain $h_v^\\prime$ , $h_a^\\prime$ , the question-aware visual and audio features, respectively. Notably, the perceiver for visual and the perceiver for audio implement shared parameters to learn potential associations.", + "bbox": [ + 212, + 642, + 787, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Step2: aggregating the perceived clues. A three-minute video undergoes a frozen encoder to get frame-level features of about 400 to 500 in length, leading to the fact that general machines cannot bear the burden of feeding all the frame-level features into the LLM. Thanks to the Q-former architecture proposed by Li et al. [27], which reduces the computational cost of end-to-end training of multimodal-language models. Specifically, we customize a learnable query vector $q$ in length $K$ to further extract useful information from the input question-aware features. Notably, we expand the $h_a'$ in the time dimension to be consistent with", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Q. Ye et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$h_v'$ and concatenate them. In practice, we set $K = 48$ to aggregate all question-aware features and obtain the clue tokens $x^{cue}$ via a clues projector to align the dimension with the LLM.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Multimodal Training Strategy", + "text_level": 1, + "bbox": [ + 214, + 217, + 509, + 233 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We follow the previous works [35] to promote comprehension over video and audio. As shown in Fig. 2 on the right, we pre-training CAT on the video-level and audio-level tasks in stage-I. In stage-II, we fine-tune CAT on high-quality audio-visual-level instructions.", + "bbox": [ + 212, + 244, + 782, + 304 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Stage-I: feature alignment for audio and visual projectors. First, we employ the video-text Webvid 2.5M dataset [3] to train the visual projector (audio projector not involved in training). Second, we employ the audio-text WavCap dataset [42] to train the audio projector (visual projector not involved in training). During the above two training periods, we freeze the LLM and the encoder to align the semantic information for vision and audio, respectively.", + "bbox": [ + 212, + 306, + 784, + 396 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Stage-II: audio-visual joint instruction tuning. To equip CAT with the ability to reason jointly based on visual and auditory, we collect an audio-visual joint instruction dataset, named AVinstruct, which emphasizes co-learning dynamic audio-visual pairs to solve diverse AVQA tasks. Specifically, we collect a large number of raw videos containing audio information from YouTube and VGGSound [5] as well as QA pairs from the training set in closed-ended AVQA tasks (i.e. music scene [26] and real-world scene [59]). Then different subtitle generators (i.e. BLIP2 [27] and Whisper [51]) are utilized to obtain video descriptions. Finally, we use GPT to generalize from human-written examples to synthesize question-guided audio-visual descriptions. We integrate all compositions after stage-I training and freeze the visual projector and audio projector, only fine-tuning clue aggregator and combined LoRA parameters on 100k video instruction [40] and AVinstruct. Notably, to highlight the question for applying in the clue aggregator, we reconstruct the form of the input prompt by adding two simple tokens $< Q >$ , $< /Q>$ , where $< Q>$ denotes the start position of the question and $< /Q>$ denotes the end position of the question. During the instruction fine-tuning phase, we use predefined prompts based on the template:", + "bbox": [ + 212, + 397, + 785, + 654 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\text {U S E R}: < \\text {s y s t e m} > < Q > < / Q > < \\text {v i d e o} > < \\text {a u d i o} > < \\text {c l u e s} > \\quad \\text {A s s i s t a n t}:\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 667, + 779, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this prompt, $<$ system $>$ denotes the official guidance message. $<$ video $>$ , $<$ audio $>$ , and $<$ clues $>$ are associated with visual tokens, audio tokens, and question-related clue tokens, respectively.", + "bbox": [ + 212, + 696, + 782, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 AI-assisted Ambiguity-aware Direct Preference Optimization", + "text_level": 1, + "bbox": [ + 214, + 766, + 763, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "At first, we attempt to design various prompts to allow MLLMs to generate the most concise descriptions. However, such a no-learning approach is challenging to achieve the goal of accurate responses in a variety of dynamic audio-visual", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "CAT", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/70c78657b54bb5127b4417b49c3264437e266ec442a7a749c20b0a55bc01a48b.jpg", + "image_caption": [ + "Fig. 4: Trained-CAT denotes CAT after feature alignment and instruction tuning. Our proposed ADPO strategy involves two steps. First, we collect the negative response generated by trained-CAT and correct it by GPT to obtain a positive response based on the original answer. Second, we perform ADPO training to skew CAT toward positive responses and reject negative responses." + ], + "image_footnote": [], + "bbox": [ + 225, + 143, + 781, + 209 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "scenarios. We find that MLLMs are easy to ambiguity when describing specific audio-visual objects. For example, an ambiguous response is illustrated by trained-CAT in Fig. 4, where trained-CAT denotes CAT after feature alignment and instruction tuning. ADPO is designed to allow MLLMs to review their mistakes and relearn with a new objective to better expression. Specifically, ADPO is divided into two steps to optimize CAT:", + "bbox": [ + 212, + 318, + 782, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Collecting ambiguous samples and rewriting. As illustrated on the left in Fig. 4, assume that a given training set $D_{train}$ and the corresponding trained-CAT output set $D_{des}$ . We first provide GPT with a detailed prompt template to review each output in $D_{des}$ for ambiguity. The rule of the review is to identify output ambiguities, i.e., responses that differ significantly from the original answer, which is referred to as a negative response. We let GPT correct this negative response, with the principle of generating a positive response that clearly describes the visual objects or sound without significant modification. Repeat the above steps to get the sample set $D_{dpo}$ to be optimized, where $D_{dpo} = \\{x^{vid}, x^{aud}, x^{cue}, x^{txt}, y_{pos}, y_{neg}\\}$ . $x$ denotes the input modal token and $x^{txt}$ denotes the text tokens, $y_{pos}$ , $y_{neg}$ denote the positive and negative responses, respectively.", + "bbox": [ + 212, + 409, + 784, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Encouraging CAT to favor positive responses. As illustrated on the right in Fig. 4, ADPO directly optimizes the low-rank adaptation matrix parameters [17] in the policy model. We assume CAT as a policy model $f^{pol}$ , given a reference model $f^{ref}$ , a positive response $y_{pos}$ , and a negative response $y_{neg}$ . The reference model $f^{ref}$ is a deep copy of $f^{pol}$ and during training we encourage $f^{pol}$ to favor $y_{pos}$ , while $f^{ref}$ does not update the weights. Specifically, we define DPO loss $\\mathcal{L}_{DPO}$ as:", + "bbox": [ + 212, + 590, + 787, + 696 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {D P O} \\left(f ^ {p o l}; f ^ {r e f}\\right) = - E _ {\\left(x ^ {v i d}, x ^ {a u d}, x ^ {c u e}, x ^ {t x t}, y _ {p o s}, y _ {n e g}\\right)} \\sim D _ {d p o} \\\\ \\left[ \\right. \\log \\sigma \\left(\\beta \\log \\frac {f ^ {p o l} \\left(y _ {p o s} \\mid [ x ^ {v i d}: x ^ {a u d}: x ^ {c u e}: x ^ {t x t} ]\\right)}{f ^ {r e f} \\left(y _ {p o s} \\mid [ x ^ {v i d}: x ^ {a u d}: x ^ {c u e}: x ^ {t x t} ]\\right)} \\right. \\quad (4) \\\\ \\left. - \\beta l o g \\frac {f ^ {p o l} \\left(y _ {n e g} \\mid \\left[ x ^ {v i d}: x ^ {a u d}: x ^ {c u e}: x ^ {t x t} \\right]\\right)}{f ^ {r e f} \\left(y _ {n e g} \\mid \\left[ x ^ {v i d}: x ^ {a u d}: x ^ {c u e}: x ^ {t x t} \\right]\\right)}\\right) \\Bigg ], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 259, + 705, + 785, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "[ ] denotes the concatenation. $\\sigma$ is the non-linear function, Sigmoid, and $\\beta$ is a hyperparameter. This objective function aims to directly optimize the", + "bbox": [ + 212, + 808, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Q. Ye et al.", + "bbox": [ + 271, + 114, + 351, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "model that favors the positive response $y_{pos}$ and rejects the negative response $y_{neg}$ . Assuming that $\\beta = 1$ , we simplify the output $\\hat{r}$ of the model based on Eq. 4 as:", + "bbox": [ + 214, + 146, + 782, + 189 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {r} \\left(x ^ {v i d}, x ^ {a u d}, x ^ {c u e}, x ^ {t x t}, y\\right) = \\log \\frac {f ^ {p o l} (y \\mid \\left[ x ^ {v i d}: x ^ {a u d}: x ^ {c u e}: x ^ {t x t} \\right])}{f ^ {r e f} (y \\mid \\left[ x ^ {v i d}: x ^ {a u d}: x ^ {c u e}: x ^ {t x t} \\right])}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 272, + 195, + 782, + 229 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The optimal goal is to make the variance of $\\hat{r}(x^{vid}, x^{aud}, x^{cue}, x^{txt}, y_{pos}) - \\hat{r}(x^{vid}, x^{aud}, x^{cue}, x^{txt}, y_{neg})$ larger and larger as a way to steer the model to generate descriptions devoid of ambiguous. However, when the difference between positive and negative responses is small, DPO loss supervision alone is not an effective facilitator. Therefore, we introduce an additional objective $\\mathcal{L}_{SFT}$ , which is similar to the process of supervised fine-tuning, to supervise the model to stabilize the bias towards the positive response. Specifically, we define $\\mathcal{L}_{SFT}$ as:", + "bbox": [ + 214, + 234, + 782, + 340 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {S F T} = - \\sum \\log P \\left(y _ {\\text {p o s}} \\mid \\left[ x ^ {\\text {v i d e o}}: x ^ {\\text {a u d i o}}: x ^ {\\text {c l u e s}}: x ^ {\\text {t x t}} \\right]; f ^ {\\text {p o l}}\\right), \\tag {6} \\\\ \\left\\{x ^ {v i d}, x ^ {a u d}, x ^ {c u e}, x ^ {t x t}, y _ {p o s} \\right\\} \\sim D _ {d p o}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 285, + 348, + 784, + 390 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "this loss function can be interpreted as using the positive output probability in the policy model to aid in optimization. During ADPO training, we sum up $\\mathcal{L}_{DPO}$ and $\\mathcal{L}_{SFT}$ with a bias $\\lambda$ to achieve direct preference optimization:", + "bbox": [ + 214, + 395, + 782, + 441 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {D P O} + \\lambda \\mathcal {L} _ {S F T}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 450, + 782, + 463 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where we set $\\lambda = 0.1$ by default.", + "bbox": [ + 214, + 470, + 449, + 484 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 503, + 374, + 522 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 214, + 532, + 333, + 546 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Video-based text generation tasks. We evaluate the proposed CAT on the video understanding benchmarks proposed by Video-ChatGPT [40]. Specifically, Video-ChatGPT proposes five metrics: Correctness of Information, Consistency, Detail Orientation, Contextual Understanding, and Temporal Understanding, which test the ability of MLLMs to describe videos.", + "bbox": [ + 214, + 551, + 782, + 627 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Zero-shot on video question answering tasks. To evaluate whether CAT has the basic ability to communicate regarding video, we conduct zero-shot tests on MSRVTT-QA [58] and ActivityNet-QA [60]. MSRVTT-QA and ActivityNet-QA consist of 10k and 5.8k videos containing audio information, respectively, where the QA pairs are mostly questions about daily life.", + "bbox": [ + 214, + 628, + 782, + 703 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Closed-ended AVQA tasks. We categorize the Music-AVQA [26] and AVQA [59] datasets as closed-ended AVQA tasks. These datasets consist of up to 42 candidate answers that require the selection of the most appropriate one based on visual and auditory content.", + "bbox": [ + 214, + 704, + 782, + 762 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Open-ended AVQA tasks. We select audio-visual dialogue (AVSD [1]), and audio-visual captioning (VALOR [6]) tasks for the evaluation of open-ended AVQA. These tasks require precise language to interpret, correlate, and reason about cross-modal information. We evaluate the zero-shot ability of CAT on these datasets.", + "bbox": [ + 214, + 763, + 782, + 838 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "CAT", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/e3ac27d5c6226cb3221f4e6b393d13fe7b3432431d64dd658fa03e132877c83e.jpg", + "table_caption": [ + "Table 1: GPT-based evaluation [40] for video-based text generation and zero-shot video question answering. For clarity, five scores are reported (\"Cr.\": Correctness of Information, \"Cs.\": Consistency, \"De.\": Detail Orientation, \"Ct\": Contextual Understanding, \"Te.\": Temporal Understanding). Score indicates the confidence." + ], + "table_footnote": [], + "table_body": "
MethodLLM SizeVideo-based text generationZero-shot on MSRVTT-QA Acc.ScoreZero-shot on ActivityNet-QA Acc.Score
Cr.De.Ct.Te.Cs.
LLaMA-VID [31]13B61.461.072.051.652.658.93.347.53.3
Video-LLaMA [64]7B39.243.643.236.435.829.61.812.41.1
LLaMA-Adapter [65]7B40.646.446.039.643.043.82.734.22.7
VideoChat [28]7B44.650.050.638.844.845.02.526.52.2
Video-ChatGPT [40]7B48.050.452.439.647.449.32.835.22.7
VISTA-LLaMA [39]7B48.852.863.645.246.260.53.348.33.3
VideoChat2 [29]7B60.457.870.253.256.254.13.349.13.3
Chat-UniVi [20]7B57.858.269.257.856.254.63.145.83.2
LLaMA-VID [31]7B59.260.070.649.250.257.73.247.13.3
CAT (Ours)7B61.659.069.856.257.862.13.550.23.5
", + "bbox": [ + 217, + 212, + 787, + 377 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 404, + 426, + 420 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation metrics. For video-based text generation and zero-shot on video question-answering tasks, we follow the evaluation pipeline proposed by VideoChatGPT [40], which is based on GPT-3.5 to evaluate predictive descriptions against correct descriptions, with a score from 0 to 5 indicating accuracy. In this paper, we standardize 0 to 5 as 0 to 100 to align common accuracy rubrics. For closed-ended AVQA tasks, we report the accuracy of the correct sample. For open-ended AVQA tasks, we report the CIDEr [56] that specializes in evaluating visually descriptive tasks.", + "bbox": [ + 212, + 429, + 787, + 549 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Architecture. We use frozen ImageBind [13] and LLaMA2-7B [55] as audiovisual encoders and LLM, respectively. The size of modality embeddings for each modality are $\\mathbb{R}^{T\\times 1024}$ . The outputs $x^{vid}$ , $x^{aud}$ , and $x^{cue}$ are $\\mathbb{R}^{1\\times 4096}$ , $\\mathbb{R}^{1\\times 4096}$ , and $\\mathbb{R}^{48\\times 4096}$ , respectively.", + "bbox": [ + 212, + 550, + 787, + 611 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training details. We complete feature alignment training, instruction tuning, and ADPO training with 1 NVIDIA A100 GPU. In detail, For the feature alignment training and instruction tuning, we use the AdamW optimizer with a cosine learning rate decay and a warm-up period. When LoRA is added, we set $r = 64$ and alpha $= 128$ for the LoRA parameters, and the total batch size is set to 128 for training 1 epoch with a learning rate of $2e^{-5}$ . For ADPO training, we only select the training set in the instruction tuning phase for optimization. We set $r = 64$ and alpha $= 16$ for the LoRA parameters, and the total batch size is set to 1 for training 1 epoch with a learning rate of $4e^{-6}$ . The hyperparameter $\\beta$ is set to 0.1.", + "bbox": [ + 212, + 611, + 787, + 762 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Comparison to State-of-the-Art", + "text_level": 1, + "bbox": [ + 215, + 784, + 522, + 800 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparison on video-based text generation tasks. We follow the benchmark proposed by Video-ChatGPT [40] to evaluate CAT. On the left of Table 1,", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Q. Ye et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/a1b2f5a580d113b7da80f10be4bf23d024d8a2a9cae5d7f39705adb99e1f92d0.jpg", + "table_caption": [ + "Table 2: Comparison with non-LLMs-based methods on fine-tuned Music-AVQA dataset." + ], + "table_footnote": [], + "table_body": "
MethodLanguage ModelTrainable Params (M)Audio avg.Visual avg.Audio-Visual avg.Overall avg.
FCNLSTM [11]CLIP [50]9.7968.956.260.460.8
GRU [2]CLIP-68.367.063.065.0
HCAattn [36]CLIP-64.965.360.362.5
MCAN [61]CLIP56.070.671.861.565.8
PSAC [30]CLIP-72.069.463.666.6
HME [10]CLIP-69.968.864.866.7
HCRN [24]CLIP-63.765.249.856.3
AVSD [54]CLIP8.3568.870.365.467.3
Panp-AVQA [62]CLIP-72.173.267.069.5
AVST [26]CLIP18.4873.974.469.571.6
PSTP-Net [25]BERT [9]4.29770.977.372.673.5
LAVISH [34]CLIP21.0977.177.377.0-
CAD [43]GLoVE [48]-78.179.776.978.2
VALOR [6]BERT----78.9
VAST [7]BERT----80.7
CAT-7B (Ours)LLaMA2 [55]5.81384.986.183.284.3
", + "bbox": [ + 264, + 184, + 736, + 363 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "we show that CAT achieves the state-of-the-art in terms of correctness of description information (Cr.), detailed description of the problem (De.), and coherence of the description (Cs.). In addition, CAT does not parse longer visual tokens, it still achieves competitive results when comparing time-related descriptions (e.g., contextual and temporal understanding).", + "bbox": [ + 212, + 396, + 782, + 470 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison on zero-shot video question answering tasks. On the right of Table 1, we show the zero-shot video question answering performance of CAT on several open-ended datasets. While recent MLLMs designed with bridging modules have produced substantial results, CAT is way ahead of them in recognizing accurate answers. We consistently outperform the state-of-the-art on the MSRVTT-QA [58] and ActivityNet-QA [60] benchmarks.", + "bbox": [ + 212, + 473, + 784, + 565 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison on closed-ended AVQA tasks. We choose to evaluate Music-AVQA [26] to demonstrate that CAT can perceive specific audio-visual objects to answer questions. In Table 2, under full supervision of the training set, CAT accurately retrieves specific objects in dynamic audio-visual scenarios and comprehensively outperforms all non-LLMs-based models. Thanks to LoRA [17], we can improve the evaluation quality with the help of the world knowledge inside LLaMA2 [55] and only 5.8M parameters are trainable. Also, we examine the ability of CAT for zero-shot on Music-AVQA [26]. In Table 3, we show the results comparing LLMs-based models. Notably, for a fair comparison, we remove the LoRA parameters of CAT that are fine-tuned on AVinstruct, which is derived from the Music-AVQA [26] and AVQA [59] training sets. Even though our model does not draw on the larger-scale LLM, it still achieves a small advantage over ChatBridge [67] with a 13B LLM size. Furthermore, we show the performance of CAT in multiple-choice scenarios [59] in Table 4, CAT continues outstanding.", + "bbox": [ + 212, + 566, + 785, + 777 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison on open-ended AVQA tasks. Open-ended AVQA tasks require responses to daily events based on audio-visual content. We conduct a comparative analysis of our model with multimodal-based LLMs: OneLLM [14], and ChatBridge [67]. The tests are divided into zero-shot and full supervision, and", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "CAT", + "bbox": [ + 696, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "we only choose to evaluate the zero-shot to demonstrate the level of practical application of the CAT. In Table 5, CAT surpasses OneLLM and ChatBridge on zero-shot complex text reasoning tasks AVSD [1] and VALOR [6]. Moreover, our approach is close to the results of FA [44] using fully supervised training, further demonstrating the strong multimodal understanding of CAT.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/db01d7819d9e9b14bdd22e00b42bf4057da977bff691df95979e5bfce87fd29f.jpg", + "table_caption": [ + "Table 3: Comparison with LLMs-based methods on zero-shot Music- AVQA dataset. Table 5: Evaluation on zero-shot open-ended" + ], + "table_footnote": [], + "table_body": "
Methodzero-shotAcc.
OneLLM-7B [14]✓43.0
ChatBridge-13B [67]✓47.6
CAT-7B (Ours)✓48.6
Table 4: Evaluation on fine-tuned AVQA dataset.
MethodAcc.
HGA + HAVF [19]87.7
HCRN + HAVF [59]89.0
PSTP-Net [25]90.2
CAT-7B (Ours)92.0
OneLLM-7B [14]✓74.5
ChatBridge-13B [67]✓75.4
CAT-7B (Ours)✓79.0
", + "bbox": [ + 215, + 297, + 792, + 448 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablations and Analyses", + "text_level": 1, + "bbox": [ + 215, + 498, + 455, + 513 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this subsection, we explore the effects of each component of CAT. We test the effect of input modal tokens on video-based text generation [40]. In addition, to avoid the huge expense associated with the GPT3.5 evaluation, we test each component of the clue aggregator on the test set of AVinstruct, as well as the generalizability of ADPO on AVSD [1].", + "bbox": [ + 212, + 522, + 785, + 598 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Does question-related clues matter? Visual tokens and audio tokens already have the ability to distinguish video content. To explore whether the added question-related clues are meaningful, we conduct ablation experiments on the video-based text generation task in Table 6. We first test visual knowledge separately from question-related clues knowledge on video comprehension. Observations reveal that the results achieved when inputting question-related clues alone have been able to outperform the joint input of audio-visual knowledge. Further, we add clues knowledge to the original audio-visual knowledge, and the results demonstrate that the fusion of the three can fully satisfy the information required for LLM reasoning.", + "bbox": [ + 212, + 598, + 785, + 750 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation on clue aggregator. We categorize a portion of the collected audiovisual joint instruction data into a test set and present the ablation results in Table 7. The evaluation indicators B, M, and C denote BLEU-4 [47], METEOR [22], and ROUGE-L, respectively. We split $\\mathcal{B}_1$ and $\\mathcal{B}_2$ and query subjects $h_v$ and $h_a$ to explore the impact of CA internals on reasoning about dynamic audiovisual pairs. We find that visual attention influences context much more than", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Q. Ye et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/220c0aad02a8222100a89d258f71f76f054e1ea83f7269b2776a5dbe2165d725.jpg", + "table_caption": [ + "Table 6: Ablation experiment Table 7: Ablation exper- Table 8: Ablation exper- about input modals on video- iments about clue aggre- iment about ADPO on based text generation. gator on audio-visual joint AVSD [1]. instruction dataset." + ], + "table_footnote": [], + "table_body": "
Input-modalCr.De.Te.
\\( x^{vid} \\)\\( x^{aud} \\)\\( x^{cue} \\)
✓38.640.434.0
✓✓40.644.835.2
✓58.657.455.8
✓✓✓61.659.056.2
", + "bbox": [ + 225, + 200, + 405, + 272 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/56aa1c4f1a2f6e940264da8c6c0d19fdbf83acbc0227ec96b9b65542256e9c0b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodB@4MR
B1 + B230.858.768.5
B1 + B2 w/o hv10.037.149.9
B1 + B2 w/o ha18.849.361.8
only B125.954.764.2
", + "bbox": [ + 424, + 212, + 589, + 273 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/e1c200741b48959cea744facf39911d952507115ca48f0cef59f360b09310b27.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelADPOAVSD
B@4MC
Video-LLaMA×18.440.163.7
Video-LLaMA✓22.248.269.2
CAT (Ours)×28.956.274.8
CAT (Ours)✓34.259.879.0
", + "bbox": [ + 599, + 200, + 792, + 271 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/bf80cdb6e84cc68f7e9a9ca3719a766e1ee21e457d7275b45cd8de341e8c2b21.jpg", + "image_caption": [ + "(a) Different number of visual tokens on video-based text generation." + ], + "image_footnote": [], + "bbox": [ + 230, + 290, + 473, + 405 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b48cafed2be87849d7aed939d4e897da04a302f8358d1a5abbeb552b4ee1f9bf.jpg", + "image_caption": [ + "(b) Different number of clue tokens on video-based text generation.", + "Fig. 5: The impacts of input modal tokens. Avg. represents the average accuracy of temporal (Te.), consistency (CS.), and detail (De.)." + ], + "image_footnote": [], + "bbox": [ + 516, + 290, + 764, + 406 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "auditory. This is because vision has more information that can be mined and is more relevant to the question. Next, we reduce parameters to investigate whether the same performance can be achieved with a single block. However, since the question acts as a query, without $\\mathcal{B}_2$ to recover the original modal length does not provide an advantage over two blocks.", + "bbox": [ + 212, + 503, + 784, + 579 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Number of input modal tokens. We evaluate the effect of the number of input modal tokens on video understanding [40]. As shown in Fig. 5, we first study the effect of the number of visual tokens on a time-related description task. Indeed, as the number of visual tokens increases, the accuracy in the evaluation of the various descriptions decreases, even if the change is not large, but it also confirms that the language model may not be able to reason based on longer visual tokens. Next, we examine the impact of the number of clue tokens set by the Q-former [27]. Clue aggregator uses Q-Former to transform frame-level features into specific tokens, and we find that it works best when the length of queries $K$ is set to 48.", + "bbox": [ + 212, + 580, + 787, + 733 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Impact of ADPO. As the paper explains, supervised fine-tuning alone still underperforms in audio-visual scenarios. We test the generalizability of ADPO on Video-LLaMA [64] and ADPO's effect on AVSD [1] in Table 8 to demonstrate the superiority of this learning strategy. C denotes the CiDER [56]. We find that for different MLLMs, ADPO does improve the descriptive power of the model without much learning cost, it brings different levels of enhancement to both Video-LLaMA and CAT.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "CAT", + "bbox": [ + 694, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c3c690717c48db9cc5fabcca04f16195d68ec28dcbd5ccd581d2d68393f5ba02.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 258, + 143, + 504, + 239 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e902a88381942ace40ae07de6ab88f2632e85273e0021b92a1e5e74b4d2a1ae4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 143, + 746, + 239 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/fab1f9f6e3f4f0cf1ad743983e8b0f19ce9a4b2ef089a32befdfca93a45a52f2.jpg", + "image_caption": [ + "Fig. 6: Qualitative results of different types of question answering." + ], + "image_footnote": [], + "bbox": [ + 258, + 241, + 504, + 335 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/1ca152d7e8057b5c6a16dc48545f02df74507b290251bff25a33d8a9fb4ad6ab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 241, + 746, + 335 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.5 Qualitative Analysis", + "text_level": 1, + "bbox": [ + 215, + 388, + 429, + 402 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In Fig. 6, we analyze the qualitative results with Video-LLaMA [64]. In the example of video-specific contextual inference, we show that CAT has a sharp perception of complex indoor scenes. Our descriptions can accurately represent what the person doing and what background sounds are in the video. Video-LLaMA is biased towards describing scene information and incorrect character information. In the specific audio question answering example, Video-LLaMA lost the ability to answer \"How many instruments are sounding in the video?\" In contrast, CAT can accurately answer the quantity question. In the audio-visual question answering example, Video-LLaMA answers incorrectly due to failed audio-visual grounding, while our CAT answers correctly due to its strong ability to capture specific objects in audio-visual scenarios.", + "bbox": [ + 212, + 411, + 787, + 578 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 598, + 359, + 614 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we introduce CAT to enhance LLMs' multimodal understanding in dynamic audio-visual scenarios. We propose a clue aggregator to aggregate the question-related clues for detailed reasoning. We mix datasets containing audio and video to empower LLM with multimodal understanding. To more consistently infer audio-visual scenarios, we collect an audio-visual joint instruction dataset to further fine-tune the CAT. Moreover, we propose an AI-assisted ambiguity-aware direct preference optimization strategy to retrain the model for more accurate responses. CAT has achieved comparable results when applied to a variety of complex audio-visual scenarios.", + "bbox": [ + 212, + 628, + 787, + 763 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by the National Natural Science Foundation of China under Grant 62306061, Guangdong Basic and Applied Basic Research Foundation (Grant No. 2023A1515140037), and the Open Fund of National Engineering Laboratory for Big Data System Computing Technology (Grant No. SZU-BDSC-OF2024-02).", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Q. Ye et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. AlAmri, H., Cartillier, V., Das, A., Wang, J., Cherian, A., Essa, I., Batra, D., Marks, T.K., Hori, C., Anderson, P., Lee, S., Parikh, D.: Audio visual scene-aware dialog. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR. pp. 7558-7567 (2019)", + "2. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: VQA: visual question answering. In: IEEE International Conference on Computer Vision, ICCV. pp. 2425-2433. IEEE Computer Society (2015)", + "3. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: IEEE/CVF International Conference on Computer Vision, ICCV. pp. 1708-1718. IEEE (2021)", + "4. Chen, G., Shen, L., Shao, R., Deng, X., Nie, L.: LION: Empowering multimodal large language model with dual-level visual knowledge. CoRR abs/2311.11860 (2023)", + "5. Chen, H., Xie, W., Vedaldi, A., Zisserman, A.: Vggsound: A large-scale audiovisual dataset. In: IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP. pp. 721-725. IEEE (2020)", + "6. Chen, S., He, X., Guo, L., Zhu, X., Wang, W., Tang, J., Liu, J.: VALOR: vision-audio-language omni-perception pretraining model and dataset. CoRR abs/2304.08345 (2023)", + "7. Chen, S., Li, H., Wang, Q., Zhao, Z., Sun, M., Zhu, X., Liu, J.: VAST: A vision-audio-subtitle-text omni-modality foundation model and dataset. CoRR abs/2305.18500 (2023)", + "8. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.C.H.: Instructclip: Towards general-purpose vision-language models with instruction tuning. CoRR abs/2305.06500 (2023)", + "9. Devlin, J., Chang, M., Lee, K., Toutanova, K.: BERT: pre-training of deep bidirectional transformers for language understanding. In: Burstein, J., Doran, C., Solorio, T. (eds.) Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT. pp. 4171-4186. Association for Computational Linguistics (2019)", + "10. Fan, C., Zhang, X., Zhang, S., Wang, W., Zhang, C., Huang, H.: Heterogeneous memory enhanced multimodal attention model for video question answering. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR. pp. 1999-2007. Computer Vision Foundation / IEEE (2019)", + "11. Fayek, H.M., Johnson, J.: Temporal reasoning via audio question answering. IEEE ACM Trans. Audio Speech Lang. Process. 28, 2283-2294 (2020)", + "12. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., Li, H., Qiao, Y.: Llama-adapter V2: parameter-efficient visual instruction model. CoRR abs/2304.15010 (2023)", + "13. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind one embedding space to bind them all. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR. pp. 15180-15190 (2023)", + "14. Han, J., Gong, K., Zhang, Y., Wang, J., Zhang, K., Lin, D., Qiao, Y., Gao, P., Yue, X.: Onellm: One framework to align all modalities with language. CoRR abs/2312.03700 (2023)", + "15. Han, J., Zhang, R., Shao, W., Gao, P., Xu, P., Xiao, H., Zhang, K., Liu, C., Wen, S., Guo, Z., Lu, X., Ren, S., Wen, Y., Chen, X., Yue, X., Li, H., Qiao, Y.: Imagebind-llm: Multi-modality instruction tuning. CoRR abs/2309.03905 (2023)" + ], + "bbox": [ + 218, + 175, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "CAT", + "bbox": [ + 696, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Hendrycks, D., Burns, C., Basart, S., Zou, A., Mazeika, M., Song, D., Steinhardt, J.: Measuring massive multitask language understanding. In: ICLR (2021)", + "17. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. In: The Tenth International Conference on Learning Representations, ICLR (2022)", + "18. Huang, B., Wang, X., Chen, H., Song, Z., Zhu, W.: Vtimellm: Empower LLM to grasp video moments. CoRR abs/2311.18445 (2023)", + "19. Jiang, P., Han, Y.: Reasoning with heterogeneous graph alignment for video question answering. In: AAAI. pp. 11109-11116. AAAI Press (2020)", + "20. Jin, P., Takanobu, R., Zhang, C., Cao, X., Yuan, L.: Chat-univi: Unified visual representation empowers large language models with image and video understanding. CoRR abs/2311.08046 (2023)", + "21. Korbar, B., Xian, Y., Tonioni, A., Zisserman, A., Tombari, F.: Text-conditioned resampler for long form video understanding. CoRR abs/2312.11897 (2023)", + "22. Lavie, A., Agarwal, A.: METEOR: an automatic metric for MT evaluation with high levels of correlation with human judgments. In: Callison-Burch, C., Koehn, P., Fordyce, C.S., Monz, C. (eds.) Proceedings of the Second Workshop on Statistical Machine Translation. pp. 228-231. Association for Computational Linguistics (2007)", + "23. Le, H., Sahoo, D., Chen, N.F., Hoi, S.C.H.: Multimodal transformer networks for end-to-end video-grounded dialogue systems. In: Korhonen, A., Traum, D.R., Marquez, L. (eds.) Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL. pp. 5612-5623. Association for Computational Linguistics (2019)", + "24. Le, T.M., Le, V., Venkatesh, S., Tran, T.: Hierarchical conditional relation networks for video question answering. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR. pp. 9969-9978. Computer Vision Foundation / IEEE (2020)", + "25. Li, G., Hou, W., Hu, D.: Progressive spatio-temporal perception for audio-visual question answering. In: El-Saddik, A., Mei, T., Cucchiara, R., Bertini, M., Vallejo, D.P.T., Atrey, P.K., Hossain, M.S. (eds.) ACM MM. pp. 7808-7816. ACM (2023)", + "26. Li, G., Wei, Y., Tian, Y., Xu, C., Wen, J.R., Hu, D.: Learning to answer questions in dynamic audio-visual scenarios. In: CVPR. p. 19086–19096 (2022)", + "27. Li, J., Li, D., Savarese, S., Hoi, S.C.H.: BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In: Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., Scarlett, J. (eds.) International Conference on Machine Learning, ICML. pp. 19730–19742. PMLR (2023)", + "28. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: Videochat: Chat-centric video understanding. CoRR abs/2305.06355 (2023)", + "29. Li, K., Wang, Y., He, Y., Li, Y., Wang, Y., Liu, Y., Wang, Z., Xu, J., Chen, G., Luo, P., Wang, L., Qiao, Y.: Mvbench: A comprehensive multi-modal video understanding benchmark. CoRR abs/2311.17005 (2023)", + "30. Li, X., Song, J., Gao, L., Liu, X., Huang, W., He, X., Gan, C.: Beyond rnns: Positional self-attention with co-attention for video question answering. In: AAAI. pp. 8658-8665. AAAI Press (2019)", + "31. Li, Y., Wang, C., Jia, J.: Llama-vid: An image is worth 2 tokens in large language models. CoRR abs/2311.17043 (2023)", + "32. Li, Z., Yang, B., Liu, Q., Ma, Z., Zhang, S., Yang, J., Sun, Y., Liu, Y., Bai, X.: Monkey: Image resolution and text label are important things for large multi-modal models. CoRR abs/2311.06607 (2023)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Q. Ye et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Lin, B., Ye, Y., Zhu, B., Cui, J., Ning, M., Jin, P., Yuan, L.: Video-llava: Learning united visual representation by alignment before projection. CoRR abs/2311.10122 (2023)", + "34. Lin, Y., Sung, Y., Lei, J., Bansal, M., Bertasius, G.: Vision transformers are parameter-efficient audio-visual learners. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR. pp. 2299-2309. IEEE (2023)", + "35. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. CoRR abs/2304.08485 (2023)", + "36. Lu, J., Yang, J., Batra, D., Parikh, D.: Hierarchical question-image co-attention for visual question answering. In: Lee, D.D., Sugiyama, M., von Luxburg, U., Guyon, I., Garnett, R. (eds.) Advances in Neural Information Processing Systems NIPs. pp. 289–297 (2016)", + "37. Luo, R., Zhao, Z., Yang, M., Dong, J., Qiu, M., Lu, P., Wang, T., Wei, Z.: Valley: Video assistant with large language model enhanced ability. CoRR abs/2306.07207 (2023)", + "38. Lyu, C., Wu, M., Wang, L., Huang, X., Liu, B., Du, Z., Shi, S., Tu, Z.: Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. CoRR abs/2306.09093 (2023)", + "39. Ma, F., Jin, X., Wang, H., Xian, Y., Feng, J., Yang, Y.: Vista-llama: Reliable video narrator via equal distance to visual tokens. CoRR abs/2312.08870 (2023)", + "40. Maaz, M., Rasheed, H.A., Khan, S.H., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models. CoRR abs/2306.05424 (2023)", + "41. MacGlashan, J., Ho, M.K., Loftin, R.T., Peng, B., Wang, G., Roberts, D.L., Taylor, M.E., Littman, M.L.: Interactive learning from policy-dependent human feedback. In: Precup, D., Teh, Y.W. (eds.) Proceedings of the 34th International Conference on Machine Learning, ICML. pp. 2285-2294 (2017)", + "42. Mei, X., Meng, C., Liu, H., Kong, Q., Ko, T., Zhao, C., Plumbley, M.D., Zou, Y., Wang, W.: Wavcaps: A chatgpt-assisted weakly-labelled audio captioning dataset for audio-language multimodal research. CoRR abs/2303.17395 (2023)", + "43. Nadeem, A., Hilton, A., Dawes, R., Thomas, G., Mustafa, A.: CAD - contextual multi-modal alignment for dynamic AVQA. CoRR abs/2310.16754 (2023)", + "44. Nguyen, D.T., Sharma, S., Schulz, H., Asri, L.E.: From film to video: Multi-turn question answering with multi-modal context. CoRR abs/1812.07023 (2018)", + "45. OpenAI: GPT-4 technical report. CoRR abs/2303.08774 (2023)", + "46. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P.F., Leike, J., Lowe, R.: Training language models to follow instructions with human feedback. In: Advances in Neural Information Processing Systems, NeurIPS (2022)", + "47. Papineni, K., Roukos, S., Ward, T., Zhu, W.: Bleu: a method for automatic evaluation of machine translation. In: Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics. pp. 311-318. ACL (2002)", + "48. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: Moschitti, A., Pang, B., Daelemans, W. (eds.) Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, EMNLP. pp. 1532-1543. ACL (2014)", + "49. Pham, H., Le, T.M., Le, V., Phuong, T.M., Tran, T.: Video dialog as conversation about objects living in space-time. In: Avidan, S., Brostow, G.J., Cissé, M., Farinella, G.M., Hassner, T. (eds.) ECCV. vol. 13699, pp. 710-726. Springer (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "CAT", + "bbox": [ + 696, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "50. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the International Conference on Machine Learning, ICML. vol. 139, pp. 8748-8763. PMLR (2021)", + "51. Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Robust speech recognition via large-scale weak supervision. In: Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., Scarlett, J. (eds.) International Conference on Machine Learning, ICML. pp. 28492-28518. PMLR (2023)", + "52. Rafailov, R., Sharma, A., Mitchell, E., Ermon, S., Manning, C.D., Finn, C.: Direct preference optimization: Your language model is secretly a reward model. CoRR abs/2305.18290 (2023)", + "53. Sakaguchi, K., Bras, R.L., Bhagavatula, C., Choi, Y.: Winogrande: An adversarial winograd schema challenge at scale. In: AAAI. pp. 8732-8740 (2020)", + "54. Schwartz, I., Schwing, A.G., Hazan, T.: A simple baseline for audio-visual scene-aware dialog. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR. pp. 12548-12558. Computer Vision Foundation / IEEE (2019)", + "55. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D., Blecher, L., Canton-Ferrer, C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I., Korenev, A., Koura, P.S., Lachaux, M., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A., Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X.E., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models. CoRR abs/2307.09288 (2023)", + "56. Vedantam, R., Zitnick, C.L., Parikh, D.: Cider: Consensus-based image description evaluation. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR. pp. 4566-4575 (2015)", + "57. Wang, X., Wu, J., Chen, J., Li, L., Wang, Y., Wang, W.Y.: Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In: IEEE/CVF International Conference on Computer Vision, ICCV. pp. 4580-4590. IEEE (2019)", + "58. Xu, D., Zhao, Z., Xiao, J., Wu, F., Zhang, H., He, X., Zhuang, Y.: Video question answering via gradually refined attention over appearance and motion. In: ACM MM. pp. 1645-1653 (2017)", + "59. Yang, P., Wang, X., Duan, X., Chen, H., Hou, R., Jin, C., Zhu, W.: AVQA: A dataset for audio-visual question answering on videos. In: ACM MM. pp. 3480-3491. ACM (2022)", + "60. Yu, Z., Xu, D., Yu, J., Yu, T., Zhao, Z., Zhuang, Y., Tao, D.: Activitynet-qa: A dataset for understanding complex web videos via question answering. In: AAAI. pp. 9127-9134 (2019)", + "61. Yu, Z., Yu, J., Cui, Y., Tao, D., Tian, Q.: Deep modular co-attention networks for visual question answering. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR. pp. 6281-6290. Computer Vision Foundation / IEEE (2019)", + "62. Yun, H., Yu, Y., Yang, W., Lee, K.I., Kim, G.H.: Pano-avqa: Grounded audiovisual question answering on $360^{\\circ}$ videos. In: CVPR. p. 2031-2041 (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Q. Ye et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "63. Zellers, R., Holtzman, A., Bisk, Y., Farhadi, A., Choi, Y.: Hellaswag: Can a machine really finish your sentence? In: Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL. pp. 4791-4800 (2019)", + "64. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. In: Proceedings of the Empirical Methods in Natural Language Processing, EMNLP. pp. 543-553 (2023)", + "65. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. CoRR abs/2303.16199 (2023)", + "66. Zhang, S., Dong, L., Li, X., Zhang, S., Sun, X., Wang, S., Li, J., Hu, R., Zhang, T., Wu, F., Wang, G.: Instruction tuning for large language models: A survey. CoRR abs/2308.10792 (2023)", + "67. Zhao, Z., Guo, L., Yue, T., Chen, S., Shao, S., Zhu, X., Yuan, Z., Liu, J.: Chatbridge: Bridging modalities with large language model as a language catalyst. CoRR abs/2305.16103 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 354 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "CAT", + "bbox": [ + 696, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/CAT_ Enhancing Multimodal Large Language Model to Answer Questions in Dynamic Audio-Visual Scenarios/825b02cb-2d21-47e3-b7b8-2d59e5594aa4_model.json b/2024/CAT_ Enhancing Multimodal Large Language Model to Answer Questions in Dynamic Audio-Visual Scenarios/825b02cb-2d21-47e3-b7b8-2d59e5594aa4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7b4fd14247d15ba9c82a47d4c12facbdaf472257 --- /dev/null +++ b/2024/CAT_ Enhancing Multimodal Large Language Model to Answer Questions in Dynamic Audio-Visual Scenarios/825b02cb-2d21-47e3-b7b8-2d59e5594aa4_model.json @@ -0,0 +1,2548 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.222, + 0.144, + 0.782, + 0.219 + ], + "angle": 0, + "content": "CAT: Enhancing Multimodal Large Language Model to Answer Questions in Dynamic Audio-Visual Scenarios" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.246, + 0.681, + 0.278 + ], + "angle": 0, + "content": "Qilang Ye\\(^{1}\\), Zitong Yu\\(^{1}\\)*, Rui Shao\\(^{2}\\), Xinyu Xie\\(^{1}\\), Philip Torr\\(^{3}\\), and Xiaochun Cao\\(^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.422, + 0.289, + 0.582, + 0.303 + ], + "angle": 0, + "content": "1 Great Bay University" + }, + { + "type": "text", + "bbox": [ + 0.355, + 0.303, + 0.646, + 0.317 + ], + "angle": 0, + "content": "\\(^{2}\\) Harbin Institute of Technology, Shenzhen" + }, + { + "type": "text", + "bbox": [ + 0.424, + 0.317, + 0.578, + 0.33 + ], + "angle": 0, + "content": "3 University of Oxford" + }, + { + "type": "text", + "bbox": [ + 0.346, + 0.331, + 0.657, + 0.345 + ], + "angle": 0, + "content": "4 Shenzhen Campus of Sun Yat-sen University" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.381, + 0.744, + 0.658 + ], + "angle": 0, + "content": "Abstract. This paper focuses on the challenge of answering questions in scenarios that are composed of rich and complex dynamic audio-visual components. Although existing Multimodal Large Language Models (MLLMs) can respond to audio-visual content, these responses are sometimes ambiguous and fail to describe specific audio-visual events. To overcome this limitation, we introduce the CAT, which enhances MLLM in three ways: 1) besides straightforwardly bridging audio and video, we design a clue aggregator that aggregates question-related clues in dynamic audio-visual scenarios to enrich the detailed knowledge required for large language models. 2) CAT is trained on a mixed multimodal dataset, allowing direct application in audio-visual scenarios. Notably, we collect an audio-visual joint instruction dataset named AVinstruct, to further enhance the capacity of CAT to model cross-semantic correlations. 3) we propose AI-assisted ambiguity-aware direct preference optimization, a strategy specialized in retraining the model to favor the non-ambiguity response and improve the ability to localize specific audio-visual objects. Extensive experimental results demonstrate that CAT outperforms existing methods on multimodal tasks, especially in Audio-Visual Question Answering (AVQA) tasks. The codes and the collected instructions will be released soon." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.671, + 0.741, + 0.7 + ], + "angle": 0, + "content": "Keywords: Multimodal Large Language Model \\(\\cdot\\) Audio-visual Question Answering" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.725, + 0.376, + 0.74 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.756, + 0.788, + 0.817 + ], + "angle": 0, + "content": "The real world revolves around sound and visual information, and their combination enhances our ability to perceive the world. Similarly, the development of Multimodal Large Language Models (MLLMs) [14, 64, 67] are closely related to audio and visual. Supervised fine-tuning [35, 46, 66] with specific instruction" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.514, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author: zitong.yu@ieee.org" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Q. Ye et al." + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.145, + 0.495, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.248, + 0.493, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.147, + 0.713, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.499 + ], + "angle": 0, + "content": "Fig. 1: Comparison between existing MLLMs and CAT Red words for incorrect response, Green words for correct response, and Gray words for useless response. Left: Most of the existing MLLMs straightforwardly bridge multimodal to large language models. Instead, CAT builds on this foundation by designing the clue aggregator to learn more detailed knowledge related to the question. Moreover, CAT constrains itself to learn a sharper response through AI-assisted ambiguity-aware direct preference optimization. Right: In comparison with audio-visual-language models Video-LLaMA [64] and ChatBridge [67], our method accurately recognizes the answers to questions with the most streamlined responses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.788, + 0.67 + ], + "angle": 0, + "content": "datasets empowers Large Language Models (LLMs) for multimodal understanding. Despite MLLMs' strong causal reasoning achieving impressive results in common-sense answering [53,63], academic answering [16], etc., predicting questions in dynamic audio-visual scenarios remains challenging. This is due to the difficulty of aligning LLMs with cross-domain data during training on large-scale multimodal corpora. It leads to particular ambiguity when describing specific objects in dynamic audio-visual scenarios. These interferences not only cause the model to make random guesses but also seriously affect the inference of detailed answers in Audio-Visual Question Answering (AVQA) [1,26,59,62] tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.788, + 0.763 + ], + "angle": 0, + "content": "A series of bridging modules [12,20,31,32,39] for MLLMs have been designed. The simplest methods [15, 18, 28, 33] use projectors to directly align text with other modalities, but partially limit the ability to capture fine-grained information. In addition, using a cross-attention mechanism to query the audio-visual context [8,21,38,39] to solve multimodal alignment problems is effective but still occurs that a certain visual object or sound cannot be localized in practice." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this work, we explore possible factors contributing to the above failures: 1) Audio-visual insufficiently correlates with the question. As illustrated in the upper left of Fig. 1, most existing MLLMs [28,64,67] are mainly designed with multiple branches to handle multiple modalities individually, followed by concatenating the modality embeddings with prompts as inputs to the LLM." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.697, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CAT" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "This paradigm cannot allow the network to learn paired text and audiovisuals, and the tandem embedding is not relevant. 2) Alignment of multimodal with text is challenging. The multimodal-text corpus [67] is difficult to be aligned, making the model generation of responses sometimes ambiguous. This ambiguity generally manifests in ambiguous words of the corresponding audio-visual content, as well as in generating useless text that overly responds. For example, the descriptions illustrated by Video-LLaMA [64] and ChatBridge [67] in Fig. 1 are all relatively ambiguous in describing the question-related video and audio content. These descriptions fail to identify the correct answer for \"bagpipes\" when asked for audio-visual reasoning. In addition, when the description is accompanied by many useless and controversy-prone words, it is unfavorable to evaluate the closed-ended AVQA task or open-ended AVQA task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.334, + 0.788, + 0.727 + ], + "angle": 0, + "content": "To overcome the two aforementioned issues, we introduce the CAT, enhancing MLLM in three ways: 1) Aggregation of question-related key clues. As illustrated in the lower left of Fig. 1, besides the global visual and audio information, we design a clue aggregator to dynamically capture question-aware visual and audio hidden features to enrich fine-grained clues. The aggregator receives sufficient low-level textual information to interact with audio-visual, it is capable of improving audio-visual grounding. 2) Mixed audio-visual training strategy. The training of CAT includes feature alignment using video-text pairs and audio-text pairs, and high-quality instructions to enhance audio-visual awareness. This strategy allows the CAT to be directly involved in real-world scenarios containing both visual and sound. Notably, we collect an audio-visual joint instruction dataset, named AVinstruct, to further empower CAT in AVQA tasks. 3) Retraining MLLMs to mitigate ambiguity. The DPO proposed by Rafailov et al. [52] has inspired us. Although MLLMs after training are equipped with multimodal understanding, they lack the flexibility to perceive ambiguity. Therefore, we reframe ambiguity elimination as a model preference optimization process and propose an AI-assisted Ambiguity-aware Direct Preference Optimization (ADPO) strategy. Specifically, we refer to ambiguous responses that express the lack of clarity of specific audio-visual objects as negative responses, and then we collect negative responses in the training set and utilize GPT to rewrite them into positive responses. After the multimodal training, we perform ADPO to retrain the model to bias towards the positive response, which is the precise description after the rewrite, and reject the negative response, which is the ambiguous description. Through this learning strategy, CAT can constrain itself to favor non-ambiguity responses. As shown in Fig. 1 on the right, CAT correctly responds to the question and excludes all useless information." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.732, + 0.608, + 0.746 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "- We introduce a novel audio-visual-language model, dubbed as CAT, that is capable of learning question-related clues and engaging directly in dynamic audio-visual inference. Notably, we collect AVInstruct, an audio-visual joint instruction dataset to ensure the stability of CAT in AVQA tasks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Q. Ye et al." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.147, + 0.786, + 0.191 + ], + "angle": 0, + "content": "- As a powerful learning strategy, we propose AI-assisted ambiguity-aware direct preference optimization to overcome the problem that MLLMs tend to ambiguously describe specific audio-visual objects." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.192, + 0.787, + 0.267 + ], + "angle": 0, + "content": "- We evaluate CAT on a wide range of multimodal tasks. Extensive experiments demonstrate the superiority of CAT (e.g., outperforms the state-of-the-art on a variety of AVQA tasks, and achieves remarkable results in the evaluation of video-based text generation tasks and zero-shot video question-answering tasks)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.285, + 0.396, + 0.301 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Audio-visual question answering (AVQA). AVQA aims to produce the most accurate linguistic representation of a given video and audio based on the question content, which requires multimodal understanding and reasoning across different semantic levels. The earliest studies [57, 58, 60] emphasize the understanding of the whole video and return a simple word that is relatively correct to the question. Subsequently, the emergence of dynamic audio-visual datasets (e.g., music scenes [26, 59], and 360-degree panoramicas [62]) increases the challenge of Question-Answering (QA), which requires the mining of temporal and spatial information in different modalities. For instance, Music-AVQA [26] requires distinguishing how many instruments are present based on the audio content. AVQA [59] requires finding the most plausible of multiple options based on the audio-visual clues from dynamic scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.719 + ], + "angle": 0, + "content": "MLLMs for audio-visual question answering. Extending LLM to other multimodal tasks is a recently emerging field. Despite the ability of MLLMs to combine information between different modalities, performance on downstream tasks, especially AVQA, remains suboptimal. Many works [8, 20, 39] emphasize the design of elegant bridging methods to improve the performance of question answering. The simplest bridging modules [37,40,64] use one or more linear layers for feature alignment. Although such methods minimize parameter updates, they still have limitations in exploring fine-grained information. Others design more complex bridging networks to query visual information. For example, Lyu et al. [38] propose an alignment module for harmonizing different representations before entering the LLM. Ma et al. [39] propose to keep the distance between all visual tokens and any linguistic tokens consistent within the LLM. However, a large multimodal corpus is difficult to align during training, neither algorithms with low parameter counts [12] nor complex bridging networks [4] are prone to the problem of failing to accurately depict specific audio-visual events." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Human-preference learning. Reinforcement Learning from Human Feedback (RLHF) [41, 46] is the most classical instance of human preference learning [45, 55], it constructs a reward model to optimize the policy model to favor preference responses. Dai et al. [8] have demonstrated that such preference learning can enhance LLM to generate more accurate information. Recently, Rafailov et al. [52] propose a Direct Preference Optimization (DPO) strategy that learns preferences directly by bypassing learning reward models, this simple yet efficient approach inspires us to solve the ambiguity problem." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.697, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CAT" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.147, + 0.777, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.308, + 0.788, + 0.424 + ], + "angle": 0, + "content": "Fig. 2: Illustration of the proposed CAT and its training strategy. (a) Overview of CAT. CAT first extracts overall audio-visual knowledge from video and audio and transforms them into visual tokens \\( x^{vid} \\) and audio tokens \\( x^{aud} \\). We input question tagged with \\( <\\mathrm{Q}> \\) in the prompt into the clue aggregator, aiming to aggregate question-aware audio-visual hidden features and yield clue tokens \\( x^{cue} \\). Finally, we merge multimodal tokens and language and feed into the frozen large language model with LoRA [17] to output the response. (b) The training paradigm of CAT involves pre-alignment of the audio-visual projectors and instruction tuning on the entire model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.451, + 0.393, + 0.468 + ], + "angle": 0, + "content": "3 Our Approach" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.788, + 0.604 + ], + "angle": 0, + "content": "In this section, we first present in detail the proposed CAT (Sec. 3.1 and 3.2). As shown on the left in Fig. 2, CAT consists of three branches that draw on visual knowledge, audio knowledge, and question-related clues to feed into the LLM, respectively. Second, we present the multimodel training strategy for CAT as shown on the right in Fig. 2, where a high-quality audio-visual joint instruction dataset is collected for further fine-tuning (Sec. 3.3). Lastly, we introduce the AI-assisted Ambiguity-aware Direct Preference Optimization (ADPO) strategy, which reinforces CAT to favor non-ambiguous descriptions (Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.625, + 0.42, + 0.641 + ], + "angle": 0, + "content": "3.1 Multimodal Inputs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.65, + 0.787, + 0.71 + ], + "angle": 0, + "content": "ImageBind [13] with a single joint embedding space demonstrates strong modality learning ability. Therefore, we leverage the frozen ImageBind as a universal encoder for all modalities. Given a video \\(\\mathbf{V}\\) and an audio \\(\\mathbf{A}\\), the encoded multimodal hidden features can be obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.341, + 0.722, + 0.786, + 0.739 + ], + "angle": 0, + "content": "\\[\nh _ {v} = \\operatorname {I m a g e B i n d} (\\mathbf {V}), \\quad h _ {a} = \\operatorname {I m a g e B i n d} (\\mathbf {A}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\( h_v \\in \\mathbb{R}^{T \\times d_h} \\), \\( h_a \\in \\mathbb{R}^{1 \\times d_h} \\) are video and audio features, respectively. \\( T \\) denotes the length of the given video and \\( d_h \\) is the specific dimension. Notably, we temporally compress the frame-level features \\( h_v \\) to address the computational burden. Further, we employ two linear projection layers to align the hidden dimensions of the inputs \\( h_v \\), \\( h_a \\) to obtain visual tokens \\( x^{v_id} \\) and audio tokens \\( x^{aud} \\), respectively." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Q. Ye et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.456, + 0.163 + ], + "angle": 0, + "content": "3.2 Aggregating Key Clues" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.298 + ], + "angle": 0, + "content": "An important component of MLLMs is the design of efficient bridging modules. The simple approach is to apply linear layers to incorporate the full visual and audio. The advantage is that it does not require any reduction of spatio-temporal information. However, such bridge approaches sometimes make LLM fail to reflect certain content and consistently generate a chunk of description about the video [28, 64]. Therefore, we devise a Clue Aggregator (CA) that enriches LLM-acquired knowledge by mining multimodal clues related to the question. CA is divided into two steps, as illustrated in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.299, + 0.788, + 0.313 + ], + "angle": 0, + "content": "Step1: perceiving question-related visual and audio clues. To extract" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.344, + 0.502, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.494, + 0.5, + 0.522 + ], + "angle": 0, + "content": "Fig. 3: Illustration of clue aggregator in a simple example." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.314, + 0.788, + 0.524 + ], + "angle": 0, + "content": "more sufficient details from the input video and audio, we devise a clue perceiver, which consists of multiple different transformer layers. The perceiver can be viewed as a tiny transformer network that arranges the self-attention (SA), cross-attention (XA), and feed-forward network (FFN) in forward and reverse order. The forward order block \\(\\mathcal{B}_1\\) is to perform attention-based question-aware localization. The reverse order block \\(\\mathcal{B}_2\\) is to consolidate the original audio-visual representation at the attentional level." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.525, + 0.787, + 0.556 + ], + "angle": 0, + "content": "Specifically, given tokens representing \"Is there a dog around?\" and the corresponding text embedding \\( h_t \\):" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.569, + 0.786, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} _ {1} \\left(h _ {t}; X\\right) = \\mathrm {X A} \\left(h _ {t}, \\operatorname {F F N} (\\operatorname {S A} (X))\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.281, + 0.617, + 0.787, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} _ {2} (X; \\mathcal {B} _ {1} \\left(h _ {t}; X\\right)) = \\operatorname {S A} (\\text {F F N} (\\operatorname {X A} (\\operatorname {F F N} (\\operatorname {S A} (X)), \\mathcal {B} _ {1} \\left(h _ {t}, X\\right)))) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.719 + ], + "angle": 0, + "content": "In \\(\\mathrm{XA}(\\cdot ,\\cdot)\\), the former represents the query and the latter represents the key and value. Based on Eqs. 2 and 3, we use \\(h_v^\\prime = \\mathcal{B}_2(h_v;\\mathcal{B}_1(h_t;h_v))\\) and \\(h_a^\\prime = \\mathcal{B}_2(h_a;\\mathcal{B}_1(h_t;h_a))\\) to obtain \\(h_v^\\prime\\), \\(h_a^\\prime\\), the question-aware visual and audio features, respectively. Notably, the perceiver for visual and the perceiver for audio implement shared parameters to learn potential associations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Step2: aggregating the perceived clues. A three-minute video undergoes a frozen encoder to get frame-level features of about 400 to 500 in length, leading to the fact that general machines cannot bear the burden of feeding all the frame-level features into the LLM. Thanks to the Q-former architecture proposed by Li et al. [27], which reduces the computational cost of end-to-end training of multimodal-language models. Specifically, we customize a learnable query vector \\( q \\) in length \\( K \\) to further extract useful information from the input question-aware features. Notably, we expand the \\( h_a' \\) in the time dimension to be consistent with" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.696, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CAT" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "\\(h_v'\\) and concatenate them. In practice, we set \\(K = 48\\) to aggregate all question-aware features and obtain the clue tokens \\(x^{cue}\\) via a clues projector to align the dimension with the LLM." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.218, + 0.51, + 0.234 + ], + "angle": 0, + "content": "3.3 Multimodal Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.245, + 0.784, + 0.305 + ], + "angle": 0, + "content": "We follow the previous works [35] to promote comprehension over video and audio. As shown in Fig. 2 on the right, we pre-training CAT on the video-level and audio-level tasks in stage-I. In stage-II, we fine-tune CAT on high-quality audio-visual-level instructions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.785, + 0.397 + ], + "angle": 0, + "content": "Stage-I: feature alignment for audio and visual projectors. First, we employ the video-text Webvid 2.5M dataset [3] to train the visual projector (audio projector not involved in training). Second, we employ the audio-text WavCap dataset [42] to train the audio projector (visual projector not involved in training). During the above two training periods, we freeze the LLM and the encoder to align the semantic information for vision and audio, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.786, + 0.655 + ], + "angle": 0, + "content": "Stage-II: audio-visual joint instruction tuning. To equip CAT with the ability to reason jointly based on visual and auditory, we collect an audio-visual joint instruction dataset, named AVinstruct, which emphasizes co-learning dynamic audio-visual pairs to solve diverse AVQA tasks. Specifically, we collect a large number of raw videos containing audio information from YouTube and VGGSound [5] as well as QA pairs from the training set in closed-ended AVQA tasks (i.e. music scene [26] and real-world scene [59]). Then different subtitle generators (i.e. BLIP2 [27] and Whisper [51]) are utilized to obtain video descriptions. Finally, we use GPT to generalize from human-written examples to synthesize question-guided audio-visual descriptions. We integrate all compositions after stage-I training and freeze the visual projector and audio projector, only fine-tuning clue aggregator and combined LoRA parameters on 100k video instruction [40] and AVinstruct. Notably, to highlight the question for applying in the clue aggregator, we reconstruct the form of the input prompt by adding two simple tokens \\(< Q >\\), \\(< /Q>\\), where \\(< Q>\\) denotes the start position of the question and \\(< /Q>\\) denotes the end position of the question. During the instruction fine-tuning phase, we use predefined prompts based on the template:" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.668, + 0.78, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\text {U S E R}: < \\text {s y s t e m} > < Q > < / Q > < \\text {v i d e o} > < \\text {a u d i o} > < \\text {c l u e s} > \\quad \\text {A s s i s t a n t}:\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.784, + 0.743 + ], + "angle": 0, + "content": "In this prompt, \\( < \\) system \\( > \\) denotes the official guidance message. \\( < \\) video \\( > \\), \\( < \\) audio \\( > \\), and \\( < \\) clues \\( > \\) are associated with visual tokens, audio tokens, and question-related clue tokens, respectively." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.767, + 0.764, + 0.783 + ], + "angle": 0, + "content": "3.4 AI-assisted Ambiguity-aware Direct Preference Optimization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "At first, we attempt to design various prompts to allow MLLMs to generate the most concise descriptions. However, such a no-learning approach is challenging to achieve the goal of accurate responses in a variety of dynamic audio-visual" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.129 + ], + "angle": 0, + "content": "Q. Ye et al." + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.144, + 0.782, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.22, + 0.788, + 0.292 + ], + "angle": 0, + "content": "Fig. 4: Trained-CAT denotes CAT after feature alignment and instruction tuning. Our proposed ADPO strategy involves two steps. First, we collect the negative response generated by trained-CAT and correct it by GPT to obtain a positive response based on the original answer. Second, we perform ADPO training to skew CAT toward positive responses and reject negative responses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.319, + 0.784, + 0.41 + ], + "angle": 0, + "content": "scenarios. We find that MLLMs are easy to ambiguity when describing specific audio-visual objects. For example, an ambiguous response is illustrated by trained-CAT in Fig. 4, where trained-CAT denotes CAT after feature alignment and instruction tuning. ADPO is designed to allow MLLMs to review their mistakes and relearn with a new objective to better expression. Specifically, ADPO is divided into two steps to optimize CAT:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.785, + 0.59 + ], + "angle": 0, + "content": "Collecting ambiguous samples and rewriting. As illustrated on the left in Fig. 4, assume that a given training set \\( D_{train} \\) and the corresponding trained-CAT output set \\( D_{des} \\). We first provide GPT with a detailed prompt template to review each output in \\( D_{des} \\) for ambiguity. The rule of the review is to identify output ambiguities, i.e., responses that differ significantly from the original answer, which is referred to as a negative response. We let GPT correct this negative response, with the principle of generating a positive response that clearly describes the visual objects or sound without significant modification. Repeat the above steps to get the sample set \\( D_{dpo} \\) to be optimized, where \\( D_{dpo} = \\{x^{vid}, x^{aud}, x^{cue}, x^{txt}, y_{pos}, y_{neg}\\} \\). \\( x \\) denotes the input modal token and \\( x^{txt} \\) denotes the text tokens, \\( y_{pos} \\), \\( y_{neg} \\) denote the positive and negative responses, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.591, + 0.788, + 0.698 + ], + "angle": 0, + "content": "Encouraging CAT to favor positive responses. As illustrated on the right in Fig. 4, ADPO directly optimizes the low-rank adaptation matrix parameters [17] in the policy model. We assume CAT as a policy model \\( f^{pol} \\), given a reference model \\( f^{ref} \\), a positive response \\( y_{pos} \\), and a negative response \\( y_{neg} \\). The reference model \\( f^{ref} \\) is a deep copy of \\( f^{pol} \\) and during training we encourage \\( f^{pol} \\) to favor \\( y_{pos} \\), while \\( f^{ref} \\) does not update the weights. Specifically, we define DPO loss \\( \\mathcal{L}_{DPO} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.261, + 0.706, + 0.786, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {D P O} \\left(f ^ {p o l}; f ^ {r e f}\\right) = - E _ {\\left(x ^ {v i d}, x ^ {a u d}, x ^ {c u e}, x ^ {t x t}, y _ {p o s}, y _ {n e g}\\right)} \\sim D _ {d p o} \\\\ \\left[ \\right. \\log \\sigma \\left(\\beta \\log \\frac {f ^ {p o l} \\left(y _ {p o s} \\mid [ x ^ {v i d}: x ^ {a u d}: x ^ {c u e}: x ^ {t x t} ]\\right)}{f ^ {r e f} \\left(y _ {p o s} \\mid [ x ^ {v i d}: x ^ {a u d}: x ^ {c u e}: x ^ {t x t} ]\\right)} \\right. \\quad (4) \\\\ \\left. - \\beta l o g \\frac {f ^ {p o l} \\left(y _ {n e g} \\mid \\left[ x ^ {v i d}: x ^ {a u d}: x ^ {c u e}: x ^ {t x t} \\right]\\right)}{f ^ {r e f} \\left(y _ {n e g} \\mid \\left[ x ^ {v i d}: x ^ {a u d}: x ^ {c u e}: x ^ {t x t} \\right]\\right)}\\right) \\Bigg ], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.809, + 0.788, + 0.842 + ], + "angle": 0, + "content": "[ ] denotes the concatenation. \\(\\sigma\\) is the non-linear function, Sigmoid, and \\(\\beta\\) is a hyperparameter. This objective function aims to directly optimize the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.696, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CAT" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.19 + ], + "angle": 0, + "content": "model that favors the positive response \\( y_{pos} \\) and rejects the negative response \\( y_{neg} \\). Assuming that \\( \\beta = 1 \\), we simplify the output \\( \\hat{r} \\) of the model based on Eq. 4 as:" + }, + { + "type": "equation", + "bbox": [ + 0.274, + 0.196, + 0.784, + 0.23 + ], + "angle": 0, + "content": "\\[\n\\hat {r} \\left(x ^ {v i d}, x ^ {a u d}, x ^ {c u e}, x ^ {t x t}, y\\right) = \\log \\frac {f ^ {p o l} (y \\mid \\left[ x ^ {v i d}: x ^ {a u d}: x ^ {c u e}: x ^ {t x t} \\right])}{f ^ {r e f} (y \\mid \\left[ x ^ {v i d}: x ^ {a u d}: x ^ {c u e}: x ^ {t x t} \\right])}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.235, + 0.784, + 0.342 + ], + "angle": 0, + "content": "The optimal goal is to make the variance of \\(\\hat{r}(x^{vid}, x^{aud}, x^{cue}, x^{txt}, y_{pos}) - \\hat{r}(x^{vid}, x^{aud}, x^{cue}, x^{txt}, y_{neg})\\) larger and larger as a way to steer the model to generate descriptions devoid of ambiguous. However, when the difference between positive and negative responses is small, DPO loss supervision alone is not an effective facilitator. Therefore, we introduce an additional objective \\(\\mathcal{L}_{SFT}\\), which is similar to the process of supervised fine-tuning, to supervise the model to stabilize the bias towards the positive response. Specifically, we define \\(\\mathcal{L}_{SFT}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.286, + 0.349, + 0.785, + 0.391 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {S F T} = - \\sum \\log P \\left(y _ {\\text {p o s}} \\mid \\left[ x ^ {\\text {v i d e o}}: x ^ {\\text {a u d i o}}: x ^ {\\text {c l u e s}}: x ^ {\\text {t x t}} \\right]; f ^ {\\text {p o l}}\\right), \\tag {6} \\\\ \\left\\{x ^ {v i d}, x ^ {a u d}, x ^ {c u e}, x ^ {t x t}, y _ {p o s} \\right\\} \\sim D _ {d p o}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.396, + 0.784, + 0.442 + ], + "angle": 0, + "content": "this loss function can be interpreted as using the positive output probability in the policy model to aid in optimization. During ADPO training, we sum up \\(\\mathcal{L}_{DPO}\\) and \\(\\mathcal{L}_{SFT}\\) with a bias \\(\\lambda\\) to achieve direct preference optimization:" + }, + { + "type": "equation", + "bbox": [ + 0.427, + 0.451, + 0.784, + 0.464 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {D P O} + \\lambda \\mathcal {L} _ {S F T}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.471, + 0.45, + 0.485 + ], + "angle": 0, + "content": "where we set \\(\\lambda = 0.1\\) by default." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.505, + 0.375, + 0.523 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.533, + 0.334, + 0.547 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.553, + 0.784, + 0.628 + ], + "angle": 0, + "content": "Video-based text generation tasks. We evaluate the proposed CAT on the video understanding benchmarks proposed by Video-ChatGPT [40]. Specifically, Video-ChatGPT proposes five metrics: Correctness of Information, Consistency, Detail Orientation, Contextual Understanding, and Temporal Understanding, which test the ability of MLLMs to describe videos." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.63, + 0.784, + 0.704 + ], + "angle": 0, + "content": "Zero-shot on video question answering tasks. To evaluate whether CAT has the basic ability to communicate regarding video, we conduct zero-shot tests on MSRVTT-QA [58] and ActivityNet-QA [60]. MSRVTT-QA and ActivityNet-QA consist of 10k and 5.8k videos containing audio information, respectively, where the QA pairs are mostly questions about daily life." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.705, + 0.784, + 0.763 + ], + "angle": 0, + "content": "Closed-ended AVQA tasks. We categorize the Music-AVQA [26] and AVQA [59] datasets as closed-ended AVQA tasks. These datasets consist of up to 42 candidate answers that require the selection of the most appropriate one based on visual and auditory content." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.765, + 0.784, + 0.839 + ], + "angle": 0, + "content": "Open-ended AVQA tasks. We select audio-visual dialogue (AVSD [1]), and audio-visual captioning (VALOR [6]) tasks for the evaluation of open-ended AVQA. These tasks require precise language to interpret, correlate, and reason about cross-modal information. We evaluate the zero-shot ability of CAT on these datasets." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Q. Ye et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: GPT-based evaluation [40] for video-based text generation and zero-shot video question answering. For clarity, five scores are reported (\"Cr.\": Correctness of Information, \"Cs.\": Consistency, \"De.\": Detail Orientation, \"Ct\": Contextual Understanding, \"Te.\": Temporal Understanding). Score indicates the confidence." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.213, + 0.789, + 0.378 + ], + "angle": 0, + "content": "
MethodLLM SizeVideo-based text generationZero-shot on MSRVTT-QA Acc.ScoreZero-shot on ActivityNet-QA Acc.Score
Cr.De.Ct.Te.Cs.
LLaMA-VID [31]13B61.461.072.051.652.658.93.347.53.3
Video-LLaMA [64]7B39.243.643.236.435.829.61.812.41.1
LLaMA-Adapter [65]7B40.646.446.039.643.043.82.734.22.7
VideoChat [28]7B44.650.050.638.844.845.02.526.52.2
Video-ChatGPT [40]7B48.050.452.439.647.449.32.835.22.7
VISTA-LLaMA [39]7B48.852.863.645.246.260.53.348.33.3
VideoChat2 [29]7B60.457.870.253.256.254.13.349.13.3
Chat-UniVi [20]7B57.858.269.257.856.254.63.145.83.2
LLaMA-VID [31]7B59.260.070.649.250.257.73.247.13.3
CAT (Ours)7B61.659.069.856.257.862.13.550.23.5
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.405, + 0.427, + 0.421 + ], + "angle": 0, + "content": "4.2 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.55 + ], + "angle": 0, + "content": "Evaluation metrics. For video-based text generation and zero-shot on video question-answering tasks, we follow the evaluation pipeline proposed by VideoChatGPT [40], which is based on GPT-3.5 to evaluate predictive descriptions against correct descriptions, with a score from 0 to 5 indicating accuracy. In this paper, we standardize 0 to 5 as 0 to 100 to align common accuracy rubrics. For closed-ended AVQA tasks, we report the accuracy of the correct sample. For open-ended AVQA tasks, we report the CIDEr [56] that specializes in evaluating visually descriptive tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.612 + ], + "angle": 0, + "content": "Architecture. We use frozen ImageBind [13] and LLaMA2-7B [55] as audiovisual encoders and LLM, respectively. The size of modality embeddings for each modality are \\(\\mathbb{R}^{T\\times 1024}\\). The outputs \\(x^{vid}\\), \\(x^{aud}\\), and \\(x^{cue}\\) are \\(\\mathbb{R}^{1\\times 4096}\\), \\(\\mathbb{R}^{1\\times 4096}\\), and \\(\\mathbb{R}^{48\\times 4096}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.789, + 0.763 + ], + "angle": 0, + "content": "Training details. We complete feature alignment training, instruction tuning, and ADPO training with 1 NVIDIA A100 GPU. In detail, For the feature alignment training and instruction tuning, we use the AdamW optimizer with a cosine learning rate decay and a warm-up period. When LoRA is added, we set \\( r = 64 \\) and alpha \\( = 128 \\) for the LoRA parameters, and the total batch size is set to 128 for training 1 epoch with a learning rate of \\( 2e^{-5} \\). For ADPO training, we only select the training set in the instruction tuning phase for optimization. We set \\( r = 64 \\) and alpha \\( = 16 \\) for the LoRA parameters, and the total batch size is set to 1 for training 1 epoch with a learning rate of \\( 4e^{-6} \\). The hyperparameter \\( \\beta \\) is set to 0.1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.785, + 0.524, + 0.801 + ], + "angle": 0, + "content": "4.3 Comparison to State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Comparison on video-based text generation tasks. We follow the benchmark proposed by Video-ChatGPT [40] to evaluate CAT. On the left of Table 1," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.697, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CAT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.172 + ], + "angle": 0, + "content": "Table 2: Comparison with non-LLMs-based methods on fine-tuned Music-AVQA dataset." + }, + { + "type": "table", + "bbox": [ + 0.266, + 0.185, + 0.737, + 0.364 + ], + "angle": 0, + "content": "
MethodLanguage ModelTrainable Params (M)Audio avg.Visual avg.Audio-Visual avg.Overall avg.
FCNLSTM [11]CLIP [50]9.7968.956.260.460.8
GRU [2]CLIP-68.367.063.065.0
HCAattn [36]CLIP-64.965.360.362.5
MCAN [61]CLIP56.070.671.861.565.8
PSAC [30]CLIP-72.069.463.666.6
HME [10]CLIP-69.968.864.866.7
HCRN [24]CLIP-63.765.249.856.3
AVSD [54]CLIP8.3568.870.365.467.3
Panp-AVQA [62]CLIP-72.173.267.069.5
AVST [26]CLIP18.4873.974.469.571.6
PSTP-Net [25]BERT [9]4.29770.977.372.673.5
LAVISH [34]CLIP21.0977.177.377.0-
CAD [43]GLoVE [48]-78.179.776.978.2
VALOR [6]BERT----78.9
VAST [7]BERT----80.7
CAT-7B (Ours)LLaMA2 [55]5.81384.986.183.284.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.784, + 0.472 + ], + "angle": 0, + "content": "we show that CAT achieves the state-of-the-art in terms of correctness of description information (Cr.), detailed description of the problem (De.), and coherence of the description (Cs.). In addition, CAT does not parse longer visual tokens, it still achieves competitive results when comparing time-related descriptions (e.g., contextual and temporal understanding)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.785, + 0.566 + ], + "angle": 0, + "content": "Comparison on zero-shot video question answering tasks. On the right of Table 1, we show the zero-shot video question answering performance of CAT on several open-ended datasets. While recent MLLMs designed with bridging modules have produced substantial results, CAT is way ahead of them in recognizing accurate answers. We consistently outperform the state-of-the-art on the MSRVTT-QA [58] and ActivityNet-QA [60] benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.787, + 0.779 + ], + "angle": 0, + "content": "Comparison on closed-ended AVQA tasks. We choose to evaluate Music-AVQA [26] to demonstrate that CAT can perceive specific audio-visual objects to answer questions. In Table 2, under full supervision of the training set, CAT accurately retrieves specific objects in dynamic audio-visual scenarios and comprehensively outperforms all non-LLMs-based models. Thanks to LoRA [17], we can improve the evaluation quality with the help of the world knowledge inside LLaMA2 [55] and only 5.8M parameters are trainable. Also, we examine the ability of CAT for zero-shot on Music-AVQA [26]. In Table 3, we show the results comparing LLMs-based models. Notably, for a fair comparison, we remove the LoRA parameters of CAT that are fine-tuned on AVinstruct, which is derived from the Music-AVQA [26] and AVQA [59] training sets. Even though our model does not draw on the larger-scale LLM, it still achieves a small advantage over ChatBridge [67] with a 13B LLM size. Furthermore, we show the performance of CAT in multiple-choice scenarios [59] in Table 4, CAT continues outstanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Comparison on open-ended AVQA tasks. Open-ended AVQA tasks require responses to daily events based on audio-visual content. We conduct a comparative analysis of our model with multimodal-based LLMs: OneLLM [14], and ChatBridge [67]. The tests are divided into zero-shot and full supervision, and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Q. Ye et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "we only choose to evaluate the zero-shot to demonstrate the level of practical application of the CAT. In Table 5, CAT surpasses OneLLM and ChatBridge on zero-shot complex text reasoning tasks AVSD [1] and VALOR [6]. Moreover, our approach is close to the results of FA [44] using fully supervised training, further demonstrating the strong multimodal understanding of CAT." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.249, + 0.79, + 0.286 + ], + "angle": 0, + "content": "Table 3: Comparison with LLMs-based methods on zero-shot Music- AVQA dataset. Table 5: Evaluation on zero-shot open-ended" + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.298, + 0.794, + 0.449 + ], + "angle": 0, + "content": "
Methodzero-shotAcc.
OneLLM-7B [14]✓43.0
ChatBridge-13B [67]✓47.6
CAT-7B (Ours)✓48.6
Table 4: Evaluation on fine-tuned AVQA dataset.
MethodAcc.
HGA + HAVF [19]87.7
HCRN + HAVF [59]89.0
PSTP-Net [25]90.2
CAT-7B (Ours)92.0
OneLLM-7B [14]✓74.5
ChatBridge-13B [67]✓75.4
CAT-7B (Ours)✓79.0
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.499, + 0.456, + 0.515 + ], + "angle": 0, + "content": "4.4 Ablations and Analyses" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.787, + 0.599 + ], + "angle": 0, + "content": "In this subsection, we explore the effects of each component of CAT. We test the effect of input modal tokens on video-based text generation [40]. In addition, to avoid the huge expense associated with the GPT3.5 evaluation, we test each component of the clue aggregator on the test set of AVinstruct, as well as the generalizability of ADPO on AVSD [1]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.75 + ], + "angle": 0, + "content": "Does question-related clues matter? Visual tokens and audio tokens already have the ability to distinguish video content. To explore whether the added question-related clues are meaningful, we conduct ablation experiments on the video-based text generation task in Table 6. We first test visual knowledge separately from question-related clues knowledge on video comprehension. Observations reveal that the results achieved when inputting question-related clues alone have been able to outperform the joint input of audio-visual knowledge. Further, we add clues knowledge to the original audio-visual knowledge, and the results demonstrate that the fusion of the three can fully satisfy the information required for LLM reasoning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Ablation on clue aggregator. We categorize a portion of the collected audiovisual joint instruction data into a test set and present the ablation results in Table 7. The evaluation indicators B, M, and C denote BLEU-4 [47], METEOR [22], and ROUGE-L, respectively. We split \\(\\mathcal{B}_1\\) and \\(\\mathcal{B}_2\\) and query subjects \\(h_v\\) and \\(h_a\\) to explore the impact of CA internals on reasoning about dynamic audiovisual pairs. We find that visual attention influences context much more than" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.696, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CAT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.146, + 0.769, + 0.199 + ], + "angle": 0, + "content": "Table 6: Ablation experiment Table 7: Ablation exper- Table 8: Ablation exper- about input modals on video- iments about clue aggre- iment about ADPO on based text generation. gator on audio-visual joint AVSD [1]. instruction dataset." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.201, + 0.406, + 0.273 + ], + "angle": 0, + "content": "
Input-modalCr.De.Te.
\\( x^{vid} \\)\\( x^{aud} \\)\\( x^{cue} \\)
✓38.640.434.0
✓✓40.644.835.2
✓58.657.455.8
✓✓✓61.659.056.2
" + }, + { + "type": "table", + "bbox": [ + 0.425, + 0.213, + 0.591, + 0.274 + ], + "angle": 0, + "content": "
MethodB@4MR
B1 + B230.858.768.5
B1 + B2 w/o hv10.037.149.9
B1 + B2 w/o ha18.849.361.8
only B125.954.764.2
" + }, + { + "type": "table", + "bbox": [ + 0.601, + 0.202, + 0.794, + 0.272 + ], + "angle": 0, + "content": "
ModelADPOAVSD
B@4MC
Video-LLaMA×18.440.163.7
Video-LLaMA✓22.248.269.2
CAT (Ours)×28.956.274.8
CAT (Ours)✓34.259.879.0
" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.291, + 0.474, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.409, + 0.515, + 0.431 + ], + "angle": 0, + "content": "(a) Different number of visual tokens on video-based text generation." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.291, + 0.765, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.409, + 0.774, + 0.431 + ], + "angle": 0, + "content": "(b) Different number of clue tokens on video-based text generation." + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.443, + 0.788, + 0.471 + ], + "angle": 0, + "content": "Fig. 5: The impacts of input modal tokens. Avg. represents the average accuracy of temporal (Te.), consistency (CS.), and detail (De.)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.785, + 0.58 + ], + "angle": 0, + "content": "auditory. This is because vision has more information that can be mined and is more relevant to the question. Next, we reduce parameters to investigate whether the same performance can be achieved with a single block. However, since the question acts as a query, without \\(\\mathcal{B}_2\\) to recover the original modal length does not provide an advantage over two blocks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Number of input modal tokens. We evaluate the effect of the number of input modal tokens on video understanding [40]. As shown in Fig. 5, we first study the effect of the number of visual tokens on a time-related description task. Indeed, as the number of visual tokens increases, the accuracy in the evaluation of the various descriptions decreases, even if the change is not large, but it also confirms that the language model may not be able to reason based on longer visual tokens. Next, we examine the impact of the number of clue tokens set by the Q-former [27]. Clue aggregator uses Q-Former to transform frame-level features into specific tokens, and we find that it works best when the length of queries \\( K \\) is set to 48." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Impact of ADPO. As the paper explains, supervised fine-tuning alone still underperforms in audio-visual scenarios. We test the generalizability of ADPO on Video-LLaMA [64] and ADPO's effect on AVSD [1] in Table 8 to demonstrate the superiority of this learning strategy. C denotes the CiDER [56]. We find that for different MLLMs, ADPO does improve the descriptive power of the model without much learning cost, it brings different levels of enhancement to both Video-LLaMA and CAT." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Q. Ye et al." + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.144, + 0.505, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.144, + 0.747, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.242, + 0.506, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.242, + 0.747, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.278, + 0.346, + 0.724, + 0.362 + ], + "angle": 0, + "content": "Fig. 6: Qualitative results of different types of question answering." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.389, + 0.431, + 0.404 + ], + "angle": 0, + "content": "4.5 Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.789, + 0.579 + ], + "angle": 0, + "content": "In Fig. 6, we analyze the qualitative results with Video-LLaMA [64]. In the example of video-specific contextual inference, we show that CAT has a sharp perception of complex indoor scenes. Our descriptions can accurately represent what the person doing and what background sounds are in the video. Video-LLaMA is biased towards describing scene information and incorrect character information. In the specific audio question answering example, Video-LLaMA lost the ability to answer \"How many instruments are sounding in the video?\" In contrast, CAT can accurately answer the quantity question. In the audio-visual question answering example, Video-LLaMA answers incorrectly due to failed audio-visual grounding, while our CAT answers correctly due to its strong ability to capture specific objects in audio-visual scenarios." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.599, + 0.36, + 0.615 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.764 + ], + "angle": 0, + "content": "In this work, we introduce CAT to enhance LLMs' multimodal understanding in dynamic audio-visual scenarios. We propose a clue aggregator to aggregate the question-related clues for detailed reasoning. We mix datasets containing audio and video to empower LLM with multimodal understanding. To more consistently infer audio-visual scenarios, we collect an audio-visual joint instruction dataset to further fine-tune the CAT. Moreover, we propose an AI-assisted ambiguity-aware direct preference optimization strategy to retrain the model for more accurate responses. CAT has achieved comparable results when applied to a variety of complex audio-visual scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by the National Natural Science Foundation of China under Grant 62306061, Guangdong Basic and Applied Basic Research Foundation (Grant No. 2023A1515140037), and the Open Fund of National Engineering Laboratory for Big Data System Computing Technology (Grant No. SZU-BDSC-OF2024-02)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.697, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CAT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.786, + 0.232 + ], + "angle": 0, + "content": "1. AlAmri, H., Cartillier, V., Das, A., Wang, J., Cherian, A., Essa, I., Batra, D., Marks, T.K., Hori, C., Anderson, P., Lee, S., Parikh, D.: Audio visual scene-aware dialog. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR. pp. 7558-7567 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.233, + 0.786, + 0.273 + ], + "angle": 0, + "content": "2. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: VQA: visual question answering. In: IEEE International Conference on Computer Vision, ICCV. pp. 2425-2433. IEEE Computer Society (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.786, + 0.315 + ], + "angle": 0, + "content": "3. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: IEEE/CVF International Conference on Computer Vision, ICCV. pp. 1708-1718. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.316, + 0.786, + 0.355 + ], + "angle": 0, + "content": "4. Chen, G., Shen, L., Shao, R., Deng, X., Nie, L.: LION: Empowering multimodal large language model with dual-level visual knowledge. CoRR abs/2311.11860 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.357, + 0.786, + 0.398 + ], + "angle": 0, + "content": "5. Chen, H., Xie, W., Vedaldi, A., Zisserman, A.: Vggsound: A large-scale audiovisual dataset. In: IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP. pp. 721-725. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.786, + 0.439 + ], + "angle": 0, + "content": "6. Chen, S., He, X., Guo, L., Zhu, X., Wang, W., Tang, J., Liu, J.: VALOR: vision-audio-language omni-perception pretraining model and dataset. CoRR abs/2304.08345 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.786, + 0.481 + ], + "angle": 0, + "content": "7. Chen, S., Li, H., Wang, Q., Zhao, Z., Sun, M., Zhu, X., Liu, J.: VAST: A vision-audio-subtitle-text omni-modality foundation model and dataset. CoRR abs/2305.18500 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.482, + 0.786, + 0.522 + ], + "angle": 0, + "content": "8. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.C.H.: Instructclip: Towards general-purpose vision-language models with instruction tuning. CoRR abs/2305.06500 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.786, + 0.591 + ], + "angle": 0, + "content": "9. Devlin, J., Chang, M., Lee, K., Toutanova, K.: BERT: pre-training of deep bidirectional transformers for language understanding. In: Burstein, J., Doran, C., Solorio, T. (eds.) Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT. pp. 4171-4186. Association for Computational Linguistics (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.592, + 0.786, + 0.647 + ], + "angle": 0, + "content": "10. Fan, C., Zhang, X., Zhang, S., Wang, W., Zhang, C., Huang, H.: Heterogeneous memory enhanced multimodal attention model for video question answering. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR. pp. 1999-2007. Computer Vision Foundation / IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.648, + 0.786, + 0.674 + ], + "angle": 0, + "content": "11. Fayek, H.M., Johnson, J.: Temporal reasoning via audio question answering. IEEE ACM Trans. Audio Speech Lang. Process. 28, 2283-2294 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.675, + 0.786, + 0.715 + ], + "angle": 0, + "content": "12. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., Li, H., Qiao, Y.: Llama-adapter V2: parameter-efficient visual instruction model. CoRR abs/2304.15010 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.716, + 0.786, + 0.757 + ], + "angle": 0, + "content": "13. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind one embedding space to bind them all. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR. pp. 15180-15190 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.786, + 0.799 + ], + "angle": 0, + "content": "14. Han, J., Gong, K., Zhang, Y., Wang, J., Zhang, K., Lin, D., Qiao, Y., Gao, P., Yue, X.: Onellm: One framework to align all modalities with language. CoRR abs/2312.03700 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "15. Han, J., Zhang, R., Shao, W., Gao, P., Xu, P., Xiao, H., Zhang, K., Liu, C., Wen, S., Guo, Z., Lu, X., Ren, S., Wen, Y., Chen, X., Yue, X., Li, H., Qiao, Y.: Imagebind-llm: Multi-modality instruction tuning. CoRR abs/2309.03905 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.176, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Q. Ye et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Hendrycks, D., Burns, C., Basart, S., Zou, A., Mazeika, M., Song, D., Steinhardt, J.: Measuring massive multitask language understanding. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "17. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. In: The Tenth International Conference on Learning Representations, ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "18. Huang, B., Wang, X., Chen, H., Song, Z., Zhu, W.: Vtimellm: Empower LLM to grasp video moments. CoRR abs/2311.18445 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.272 + ], + "angle": 0, + "content": "19. Jiang, P., Han, Y.: Reasoning with heterogeneous graph alignment for video question answering. In: AAAI. pp. 11109-11116. AAAI Press (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.313 + ], + "angle": 0, + "content": "20. Jin, P., Takanobu, R., Zhang, C., Cao, X., Yuan, L.: Chat-univi: Unified visual representation empowers large language models with image and video understanding. CoRR abs/2311.08046 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "21. Korbar, B., Xian, Y., Tonioni, A., Zisserman, A., Tombari, F.: Text-conditioned resampler for long form video understanding. CoRR abs/2312.11897 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.41 + ], + "angle": 0, + "content": "22. Lavie, A., Agarwal, A.: METEOR: an automatic metric for MT evaluation with high levels of correlation with human judgments. In: Callison-Burch, C., Koehn, P., Fordyce, C.S., Monz, C. (eds.) Proceedings of the Second Workshop on Statistical Machine Translation. pp. 228-231. Association for Computational Linguistics (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.48 + ], + "angle": 0, + "content": "23. Le, H., Sahoo, D., Chen, N.F., Hoi, S.C.H.: Multimodal transformer networks for end-to-end video-grounded dialogue systems. In: Korhonen, A., Traum, D.R., Marquez, L. (eds.) Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL. pp. 5612-5623. Association for Computational Linguistics (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.535 + ], + "angle": 0, + "content": "24. Le, T.M., Le, V., Venkatesh, S., Tran, T.: Hierarchical conditional relation networks for video question answering. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR. pp. 9969-9978. Computer Vision Foundation / IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "25. Li, G., Hou, W., Hu, D.: Progressive spatio-temporal perception for audio-visual question answering. In: El-Saddik, A., Mei, T., Cucchiara, R., Bertini, M., Vallejo, D.P.T., Atrey, P.K., Hossain, M.S. (eds.) ACM MM. pp. 7808-7816. ACM (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.604 + ], + "angle": 0, + "content": "26. Li, G., Wei, Y., Tian, Y., Xu, C., Wen, J.R., Hu, D.: Learning to answer questions in dynamic audio-visual scenarios. In: CVPR. p. 19086–19096 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.66 + ], + "angle": 0, + "content": "27. Li, J., Li, D., Savarese, S., Hoi, S.C.H.: BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In: Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., Scarlett, J. (eds.) International Conference on Machine Learning, ICML. pp. 19730–19742. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.687 + ], + "angle": 0, + "content": "28. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: Videochat: Chat-centric video understanding. CoRR abs/2305.06355 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.785, + 0.729 + ], + "angle": 0, + "content": "29. Li, K., Wang, Y., He, Y., Li, Y., Wang, Y., Liu, Y., Wang, Z., Xu, J., Chen, G., Luo, P., Wang, L., Qiao, Y.: Mvbench: A comprehensive multi-modal video understanding benchmark. CoRR abs/2311.17005 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.771 + ], + "angle": 0, + "content": "30. Li, X., Song, J., Gao, L., Liu, X., Huang, W., He, X., Gan, C.: Beyond rnns: Positional self-attention with co-attention for video question answering. In: AAAI. pp. 8658-8665. AAAI Press (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.798 + ], + "angle": 0, + "content": "31. Li, Y., Wang, C., Jia, J.: Llama-vid: An image is worth 2 tokens in large language models. CoRR abs/2311.17043 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "32. Li, Z., Yang, B., Liu, Q., Ma, Z., Zhang, S., Yang, J., Sun, Y., Liu, Y., Bai, X.: Monkey: Image resolution and text label are important things for large multi-modal models. CoRR abs/2311.06607 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.697, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CAT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "33. Lin, B., Ye, Y., Zhu, B., Cui, J., Ning, M., Jin, P., Yuan, L.: Video-llava: Learning united visual representation by alignment before projection. CoRR abs/2311.10122 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "34. Lin, Y., Sung, Y., Lei, J., Bansal, M., Bertasius, G.: Vision transformers are parameter-efficient audio-visual learners. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR. pp. 2299-2309. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.259 + ], + "angle": 0, + "content": "35. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. CoRR abs/2304.08485 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.26, + 0.788, + 0.314 + ], + "angle": 0, + "content": "36. Lu, J., Yang, J., Batra, D., Parikh, D.: Hierarchical question-image co-attention for visual question answering. In: Lee, D.D., Sugiyama, M., von Luxburg, U., Guyon, I., Garnett, R. (eds.) Advances in Neural Information Processing Systems NIPs. pp. 289–297 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.315, + 0.788, + 0.355 + ], + "angle": 0, + "content": "37. Luo, R., Zhao, Z., Yang, M., Dong, J., Qiu, M., Lu, P., Wang, T., Wei, Z.: Valley: Video assistant with large language model enhanced ability. CoRR abs/2306.07207 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.397 + ], + "angle": 0, + "content": "38. Lyu, C., Wu, M., Wang, L., Huang, X., Liu, B., Du, Z., Shi, S., Tu, Z.: Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. CoRR abs/2306.09093 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.424 + ], + "angle": 0, + "content": "39. Ma, F., Jin, X., Wang, H., Xian, Y., Feng, J., Yang, Y.: Vista-llama: Reliable video narrator via equal distance to visual tokens. CoRR abs/2312.08870 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.467 + ], + "angle": 0, + "content": "40. Maaz, M., Rasheed, H.A., Khan, S.H., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models. CoRR abs/2306.05424 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.468, + 0.788, + 0.522 + ], + "angle": 0, + "content": "41. MacGlashan, J., Ho, M.K., Loftin, R.T., Peng, B., Wang, G., Roberts, D.L., Taylor, M.E., Littman, M.L.: Interactive learning from policy-dependent human feedback. In: Precup, D., Teh, Y.W. (eds.) Proceedings of the 34th International Conference on Machine Learning, ICML. pp. 2285-2294 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.563 + ], + "angle": 0, + "content": "42. Mei, X., Meng, C., Liu, H., Kong, Q., Ko, T., Zhao, C., Plumbley, M.D., Zou, Y., Wang, W.: Wavcaps: A chatgpt-assisted weakly-labelled audio captioning dataset for audio-language multimodal research. CoRR abs/2303.17395 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.565, + 0.788, + 0.592 + ], + "angle": 0, + "content": "43. Nadeem, A., Hilton, A., Dawes, R., Thomas, G., Mustafa, A.: CAD - contextual multi-modal alignment for dynamic AVQA. CoRR abs/2310.16754 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.592, + 0.788, + 0.619 + ], + "angle": 0, + "content": "44. Nguyen, D.T., Sharma, S., Schulz, H., Asri, L.E.: From film to video: Multi-turn question answering with multi-modal context. CoRR abs/1812.07023 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.62, + 0.684, + 0.633 + ], + "angle": 0, + "content": "45. OpenAI: GPT-4 technical report. CoRR abs/2303.08774 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.634, + 0.788, + 0.702 + ], + "angle": 0, + "content": "46. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P.F., Leike, J., Lowe, R.: Training language models to follow instructions with human feedback. In: Advances in Neural Information Processing Systems, NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.702, + 0.788, + 0.743 + ], + "angle": 0, + "content": "47. Papineni, K., Roukos, S., Ward, T., Zhu, W.: Bleu: a method for automatic evaluation of machine translation. In: Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics. pp. 311-318. ACL (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.744, + 0.788, + 0.799 + ], + "angle": 0, + "content": "48. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: Moschitti, A., Pang, B., Daelemans, W. (eds.) Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, EMNLP. pp. 1532-1543. ACL (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.8, + 0.788, + 0.841 + ], + "angle": 0, + "content": "49. Pham, H., Le, T.M., Le, V., Phuong, T.M., Tran, T.: Video dialog as conversation about objects living in space-time. In: Avidan, S., Brostow, G.J., Cissé, M., Farinella, G.M., Hassner, T. (eds.) ECCV. vol. 13699, pp. 710-726. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Q. Ye et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.218 + ], + "angle": 0, + "content": "50. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the International Conference on Machine Learning, ICML. vol. 139, pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.219, + 0.787, + 0.274 + ], + "angle": 0, + "content": "51. Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Robust speech recognition via large-scale weak supervision. In: Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., Scarlett, J. (eds.) International Conference on Machine Learning, ICML. pp. 28492-28518. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.787, + 0.317 + ], + "angle": 0, + "content": "52. Rafailov, R., Sharma, A., Mitchell, E., Ermon, S., Manning, C.D., Finn, C.: Direct preference optimization: Your language model is secretly a reward model. CoRR abs/2305.18290 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.787, + 0.346 + ], + "angle": 0, + "content": "53. Sakaguchi, K., Bras, R.L., Bhagavatula, C., Choi, Y.: Winogrande: An adversarial winograd schema challenge at scale. In: AAAI. pp. 8732-8740 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.347, + 0.787, + 0.389 + ], + "angle": 0, + "content": "54. Schwartz, I., Schwing, A.G., Hazan, T.: A simple baseline for audio-visual scene-aware dialog. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR. pp. 12548-12558. Computer Vision Foundation / IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.39, + 0.787, + 0.555 + ], + "angle": 0, + "content": "55. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D., Blecher, L., Canton-Ferrer, C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I., Korenev, A., Koura, P.S., Lachaux, M., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A., Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X.E., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models. CoRR abs/2307.09288 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.557, + 0.787, + 0.599 + ], + "angle": 0, + "content": "56. Vedantam, R., Zitnick, C.L., Parikh, D.: Cider: Consensus-based image description evaluation. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR. pp. 4566-4575 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.6, + 0.787, + 0.641 + ], + "angle": 0, + "content": "57. Wang, X., Wu, J., Chen, J., Li, L., Wang, Y., Wang, W.Y.: Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In: IEEE/CVF International Conference on Computer Vision, ICCV. pp. 4580-4590. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.642, + 0.787, + 0.683 + ], + "angle": 0, + "content": "58. Xu, D., Zhao, Z., Xiao, J., Wu, F., Zhang, H., He, X., Zhuang, Y.: Video question answering via gradually refined attention over appearance and motion. In: ACM MM. pp. 1645-1653 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.684, + 0.787, + 0.726 + ], + "angle": 0, + "content": "59. Yang, P., Wang, X., Duan, X., Chen, H., Hou, R., Jin, C., Zhu, W.: AVQA: A dataset for audio-visual question answering on videos. In: ACM MM. pp. 3480-3491. ACM (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.727, + 0.787, + 0.769 + ], + "angle": 0, + "content": "60. Yu, Z., Xu, D., Yu, J., Yu, T., Zhao, Z., Zhuang, Y., Tao, D.: Activitynet-qa: A dataset for understanding complex web videos via question answering. In: AAAI. pp. 9127-9134 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.77, + 0.787, + 0.812 + ], + "angle": 0, + "content": "61. Yu, Z., Yu, J., Cui, Y., Tao, D., Tian, Q.: Deep modular co-attention networks for visual question answering. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR. pp. 6281-6290. Computer Vision Foundation / IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.841 + ], + "angle": 0, + "content": "62. Yun, H., Yu, Y., Yang, W., Lee, K.I., Kim, G.H.: Pano-avqa: Grounded audiovisual question answering on \\(360^{\\circ}\\) videos. In: CVPR. p. 2031-2041 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.697, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CAT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "63. Zellers, R., Holtzman, A., Bisk, Y., Farhadi, A., Choi, Y.: Hellaswag: Can a machine really finish your sentence? In: Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL. pp. 4791-4800 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "64. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. In: Proceedings of the Empirical Methods in Natural Language Processing, EMNLP. pp. 543-553 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.272 + ], + "angle": 0, + "content": "65. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. CoRR abs/2303.16199 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.313 + ], + "angle": 0, + "content": "66. Zhang, S., Dong, L., Li, X., Zhang, S., Sun, X., Wang, S., Li, J., Hu, R., Zhang, T., Wu, F., Wang, G.: Instruction tuning for large language models: A survey. CoRR abs/2308.10792 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.356 + ], + "angle": 0, + "content": "67. Zhao, Z., Guo, L., Yue, T., Chen, S., Shao, S., Zhu, X., Yuan, Z., Liu, J.: Chatbridge: Bridging modalities with large language model as a language catalyst. CoRR abs/2305.16103 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.356 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CAT_ Enhancing Multimodal Large Language Model to Answer Questions in Dynamic Audio-Visual Scenarios/825b02cb-2d21-47e3-b7b8-2d59e5594aa4_origin.pdf b/2024/CAT_ Enhancing Multimodal Large Language Model to Answer Questions in Dynamic Audio-Visual Scenarios/825b02cb-2d21-47e3-b7b8-2d59e5594aa4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7e25ee78f1f265c73488b393dea1c52ed146969a --- /dev/null +++ b/2024/CAT_ Enhancing Multimodal Large Language Model to Answer Questions in Dynamic Audio-Visual Scenarios/825b02cb-2d21-47e3-b7b8-2d59e5594aa4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5439169fcf9af97594fd7c55d3c32dfa14bcc05e3bec6d2f004a3f526bab5733 +size 2762002 diff --git a/2024/CAT_ Enhancing Multimodal Large Language Model to Answer Questions in Dynamic Audio-Visual Scenarios/full.md b/2024/CAT_ Enhancing Multimodal Large Language Model to Answer Questions in Dynamic Audio-Visual Scenarios/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a72543caf83379d3e0d254574f9b22cbca91e607 --- /dev/null +++ b/2024/CAT_ Enhancing Multimodal Large Language Model to Answer Questions in Dynamic Audio-Visual Scenarios/full.md @@ -0,0 +1,317 @@ +# CAT: Enhancing Multimodal Large Language Model to Answer Questions in Dynamic Audio-Visual Scenarios + +Qilang Ye $^{1}$ , Zitong Yu $^{1}$ *, Rui Shao $^{2}$ , Xinyu Xie $^{1}$ , Philip Torr $^{3}$ , and Xiaochun Cao $^{4}$ + +1 Great Bay University + +$^{2}$ Harbin Institute of Technology, Shenzhen + +3 University of Oxford + +4 Shenzhen Campus of Sun Yat-sen University + +Abstract. This paper focuses on the challenge of answering questions in scenarios that are composed of rich and complex dynamic audio-visual components. Although existing Multimodal Large Language Models (MLLMs) can respond to audio-visual content, these responses are sometimes ambiguous and fail to describe specific audio-visual events. To overcome this limitation, we introduce the CAT, which enhances MLLM in three ways: 1) besides straightforwardly bridging audio and video, we design a clue aggregator that aggregates question-related clues in dynamic audio-visual scenarios to enrich the detailed knowledge required for large language models. 2) CAT is trained on a mixed multimodal dataset, allowing direct application in audio-visual scenarios. Notably, we collect an audio-visual joint instruction dataset named AVinstruct, to further enhance the capacity of CAT to model cross-semantic correlations. 3) we propose AI-assisted ambiguity-aware direct preference optimization, a strategy specialized in retraining the model to favor the non-ambiguity response and improve the ability to localize specific audio-visual objects. Extensive experimental results demonstrate that CAT outperforms existing methods on multimodal tasks, especially in Audio-Visual Question Answering (AVQA) tasks. The codes and the collected instructions will be released soon. + +Keywords: Multimodal Large Language Model $\cdot$ Audio-visual Question Answering + +# 1 Introduction + +The real world revolves around sound and visual information, and their combination enhances our ability to perceive the world. Similarly, the development of Multimodal Large Language Models (MLLMs) [14, 64, 67] are closely related to audio and visual. Supervised fine-tuning [35, 46, 66] with specific instruction + +![](images/94ba68624e535d3e41b8ad7f78d5f5d088bc827f024ed4183a87c00d0b6df2b2.jpg) + +![](images/30c64ddfd88c267d6bebac2a77f61be0d21e375496b704fe99e8409911789191.jpg) +Fig. 1: Comparison between existing MLLMs and CAT Red words for incorrect response, Green words for correct response, and Gray words for useless response. Left: Most of the existing MLLMs straightforwardly bridge multimodal to large language models. Instead, CAT builds on this foundation by designing the clue aggregator to learn more detailed knowledge related to the question. Moreover, CAT constrains itself to learn a sharper response through AI-assisted ambiguity-aware direct preference optimization. Right: In comparison with audio-visual-language models Video-LLaMA [64] and ChatBridge [67], our method accurately recognizes the answers to questions with the most streamlined responses. + +![](images/c3206106d4c5640b37a232e86ee5b90c2156d52af42671be69cab6d971d2573c.jpg) + +datasets empowers Large Language Models (LLMs) for multimodal understanding. Despite MLLMs' strong causal reasoning achieving impressive results in common-sense answering [53,63], academic answering [16], etc., predicting questions in dynamic audio-visual scenarios remains challenging. This is due to the difficulty of aligning LLMs with cross-domain data during training on large-scale multimodal corpora. It leads to particular ambiguity when describing specific objects in dynamic audio-visual scenarios. These interferences not only cause the model to make random guesses but also seriously affect the inference of detailed answers in Audio-Visual Question Answering (AVQA) [1,26,59,62] tasks. + +A series of bridging modules [12,20,31,32,39] for MLLMs have been designed. The simplest methods [15, 18, 28, 33] use projectors to directly align text with other modalities, but partially limit the ability to capture fine-grained information. In addition, using a cross-attention mechanism to query the audio-visual context [8,21,38,39] to solve multimodal alignment problems is effective but still occurs that a certain visual object or sound cannot be localized in practice. + +In this work, we explore possible factors contributing to the above failures: 1) Audio-visual insufficiently correlates with the question. As illustrated in the upper left of Fig. 1, most existing MLLMs [28,64,67] are mainly designed with multiple branches to handle multiple modalities individually, followed by concatenating the modality embeddings with prompts as inputs to the LLM. + +This paradigm cannot allow the network to learn paired text and audiovisuals, and the tandem embedding is not relevant. 2) Alignment of multimodal with text is challenging. The multimodal-text corpus [67] is difficult to be aligned, making the model generation of responses sometimes ambiguous. This ambiguity generally manifests in ambiguous words of the corresponding audio-visual content, as well as in generating useless text that overly responds. For example, the descriptions illustrated by Video-LLaMA [64] and ChatBridge [67] in Fig. 1 are all relatively ambiguous in describing the question-related video and audio content. These descriptions fail to identify the correct answer for "bagpipes" when asked for audio-visual reasoning. In addition, when the description is accompanied by many useless and controversy-prone words, it is unfavorable to evaluate the closed-ended AVQA task or open-ended AVQA task. + +To overcome the two aforementioned issues, we introduce the CAT, enhancing MLLM in three ways: 1) Aggregation of question-related key clues. As illustrated in the lower left of Fig. 1, besides the global visual and audio information, we design a clue aggregator to dynamically capture question-aware visual and audio hidden features to enrich fine-grained clues. The aggregator receives sufficient low-level textual information to interact with audio-visual, it is capable of improving audio-visual grounding. 2) Mixed audio-visual training strategy. The training of CAT includes feature alignment using video-text pairs and audio-text pairs, and high-quality instructions to enhance audio-visual awareness. This strategy allows the CAT to be directly involved in real-world scenarios containing both visual and sound. Notably, we collect an audio-visual joint instruction dataset, named AVinstruct, to further empower CAT in AVQA tasks. 3) Retraining MLLMs to mitigate ambiguity. The DPO proposed by Rafailov et al. [52] has inspired us. Although MLLMs after training are equipped with multimodal understanding, they lack the flexibility to perceive ambiguity. Therefore, we reframe ambiguity elimination as a model preference optimization process and propose an AI-assisted Ambiguity-aware Direct Preference Optimization (ADPO) strategy. Specifically, we refer to ambiguous responses that express the lack of clarity of specific audio-visual objects as negative responses, and then we collect negative responses in the training set and utilize GPT to rewrite them into positive responses. After the multimodal training, we perform ADPO to retrain the model to bias towards the positive response, which is the precise description after the rewrite, and reject the negative response, which is the ambiguous description. Through this learning strategy, CAT can constrain itself to favor non-ambiguity responses. As shown in Fig. 1 on the right, CAT correctly responds to the question and excludes all useless information. + +Our main contributions are summarized as follows: + +- We introduce a novel audio-visual-language model, dubbed as CAT, that is capable of learning question-related clues and engaging directly in dynamic audio-visual inference. Notably, we collect AVInstruct, an audio-visual joint instruction dataset to ensure the stability of CAT in AVQA tasks. + +- As a powerful learning strategy, we propose AI-assisted ambiguity-aware direct preference optimization to overcome the problem that MLLMs tend to ambiguously describe specific audio-visual objects. + +- We evaluate CAT on a wide range of multimodal tasks. Extensive experiments demonstrate the superiority of CAT (e.g., outperforms the state-of-the-art on a variety of AVQA tasks, and achieves remarkable results in the evaluation of video-based text generation tasks and zero-shot video question-answering tasks). + +# 2 Related Works + +Audio-visual question answering (AVQA). AVQA aims to produce the most accurate linguistic representation of a given video and audio based on the question content, which requires multimodal understanding and reasoning across different semantic levels. The earliest studies [57, 58, 60] emphasize the understanding of the whole video and return a simple word that is relatively correct to the question. Subsequently, the emergence of dynamic audio-visual datasets (e.g., music scenes [26, 59], and 360-degree panoramicas [62]) increases the challenge of Question-Answering (QA), which requires the mining of temporal and spatial information in different modalities. For instance, Music-AVQA [26] requires distinguishing how many instruments are present based on the audio content. AVQA [59] requires finding the most plausible of multiple options based on the audio-visual clues from dynamic scenarios. + +MLLMs for audio-visual question answering. Extending LLM to other multimodal tasks is a recently emerging field. Despite the ability of MLLMs to combine information between different modalities, performance on downstream tasks, especially AVQA, remains suboptimal. Many works [8, 20, 39] emphasize the design of elegant bridging methods to improve the performance of question answering. The simplest bridging modules [37,40,64] use one or more linear layers for feature alignment. Although such methods minimize parameter updates, they still have limitations in exploring fine-grained information. Others design more complex bridging networks to query visual information. For example, Lyu et al. [38] propose an alignment module for harmonizing different representations before entering the LLM. Ma et al. [39] propose to keep the distance between all visual tokens and any linguistic tokens consistent within the LLM. However, a large multimodal corpus is difficult to align during training, neither algorithms with low parameter counts [12] nor complex bridging networks [4] are prone to the problem of failing to accurately depict specific audio-visual events. + +Human-preference learning. Reinforcement Learning from Human Feedback (RLHF) [41, 46] is the most classical instance of human preference learning [45, 55], it constructs a reward model to optimize the policy model to favor preference responses. Dai et al. [8] have demonstrated that such preference learning can enhance LLM to generate more accurate information. Recently, Rafailov et al. [52] propose a Direct Preference Optimization (DPO) strategy that learns preferences directly by bypassing learning reward models, this simple yet efficient approach inspires us to solve the ambiguity problem. + +![](images/fd1da11cfd1466c864099619b84d1827ca664efd9239b548a73b98444c76f8ec.jpg) +Fig. 2: Illustration of the proposed CAT and its training strategy. (a) Overview of CAT. CAT first extracts overall audio-visual knowledge from video and audio and transforms them into visual tokens $x^{vid}$ and audio tokens $x^{aud}$ . We input question tagged with $<\mathrm{Q}> $ in the prompt into the clue aggregator, aiming to aggregate question-aware audio-visual hidden features and yield clue tokens $x^{cue}$ . Finally, we merge multimodal tokens and language and feed into the frozen large language model with LoRA [17] to output the response. (b) The training paradigm of CAT involves pre-alignment of the audio-visual projectors and instruction tuning on the entire model. + +# 3 Our Approach + +In this section, we first present in detail the proposed CAT (Sec. 3.1 and 3.2). As shown on the left in Fig. 2, CAT consists of three branches that draw on visual knowledge, audio knowledge, and question-related clues to feed into the LLM, respectively. Second, we present the multimodel training strategy for CAT as shown on the right in Fig. 2, where a high-quality audio-visual joint instruction dataset is collected for further fine-tuning (Sec. 3.3). Lastly, we introduce the AI-assisted Ambiguity-aware Direct Preference Optimization (ADPO) strategy, which reinforces CAT to favor non-ambiguous descriptions (Sec. 3.4). + +# 3.1 Multimodal Inputs + +ImageBind [13] with a single joint embedding space demonstrates strong modality learning ability. Therefore, we leverage the frozen ImageBind as a universal encoder for all modalities. Given a video $\mathbf{V}$ and an audio $\mathbf{A}$ , the encoded multimodal hidden features can be obtained by: + +$$ +h _ {v} = \operatorname {I m a g e B i n d} (\mathbf {V}), \quad h _ {a} = \operatorname {I m a g e B i n d} (\mathbf {A}), \tag {1} +$$ + +where $h_v \in \mathbb{R}^{T \times d_h}$ , $h_a \in \mathbb{R}^{1 \times d_h}$ are video and audio features, respectively. $T$ denotes the length of the given video and $d_h$ is the specific dimension. Notably, we temporally compress the frame-level features $h_v$ to address the computational burden. Further, we employ two linear projection layers to align the hidden dimensions of the inputs $h_v$ , $h_a$ to obtain visual tokens $x^{v_id}$ and audio tokens $x^{aud}$ , respectively. + +# 3.2 Aggregating Key Clues + +An important component of MLLMs is the design of efficient bridging modules. The simple approach is to apply linear layers to incorporate the full visual and audio. The advantage is that it does not require any reduction of spatio-temporal information. However, such bridge approaches sometimes make LLM fail to reflect certain content and consistently generate a chunk of description about the video [28, 64]. Therefore, we devise a Clue Aggregator (CA) that enriches LLM-acquired knowledge by mining multimodal clues related to the question. CA is divided into two steps, as illustrated in Fig. 3. + +Step1: perceiving question-related visual and audio clues. To extract + +![](images/70516e33dc025a4fca401f3ae0947dcf8b2b7b330391af6ef41f6e05abc82d0d.jpg) +Fig. 3: Illustration of clue aggregator in a simple example. + +more sufficient details from the input video and audio, we devise a clue perceiver, which consists of multiple different transformer layers. The perceiver can be viewed as a tiny transformer network that arranges the self-attention (SA), cross-attention (XA), and feed-forward network (FFN) in forward and reverse order. The forward order block $\mathcal{B}_1$ is to perform attention-based question-aware localization. The reverse order block $\mathcal{B}_2$ is to consolidate the original audio-visual representation at the attentional level. + +Specifically, given tokens representing "Is there a dog around?" and the corresponding text embedding $h_t$ : + +$$ +\mathcal {B} _ {1} \left(h _ {t}; X\right) = \mathrm {X A} \left(h _ {t}, \operatorname {F F N} (\operatorname {S A} (X))\right), \tag {2} +$$ + +$$ +\mathcal {B} _ {2} (X; \mathcal {B} _ {1} \left(h _ {t}; X\right)) = \operatorname {S A} (\text {F F N} (\operatorname {X A} (\operatorname {F F N} (\operatorname {S A} (X)), \mathcal {B} _ {1} \left(h _ {t}, X\right)))) \tag {3} +$$ + +In $\mathrm{XA}(\cdot ,\cdot)$ , the former represents the query and the latter represents the key and value. Based on Eqs. 2 and 3, we use $h_v^\prime = \mathcal{B}_2(h_v;\mathcal{B}_1(h_t;h_v))$ and $h_a^\prime = \mathcal{B}_2(h_a;\mathcal{B}_1(h_t;h_a))$ to obtain $h_v^\prime$ , $h_a^\prime$ , the question-aware visual and audio features, respectively. Notably, the perceiver for visual and the perceiver for audio implement shared parameters to learn potential associations. + +Step2: aggregating the perceived clues. A three-minute video undergoes a frozen encoder to get frame-level features of about 400 to 500 in length, leading to the fact that general machines cannot bear the burden of feeding all the frame-level features into the LLM. Thanks to the Q-former architecture proposed by Li et al. [27], which reduces the computational cost of end-to-end training of multimodal-language models. Specifically, we customize a learnable query vector $q$ in length $K$ to further extract useful information from the input question-aware features. Notably, we expand the $h_a'$ in the time dimension to be consistent with + +$h_v'$ and concatenate them. In practice, we set $K = 48$ to aggregate all question-aware features and obtain the clue tokens $x^{cue}$ via a clues projector to align the dimension with the LLM. + +# 3.3 Multimodal Training Strategy + +We follow the previous works [35] to promote comprehension over video and audio. As shown in Fig. 2 on the right, we pre-training CAT on the video-level and audio-level tasks in stage-I. In stage-II, we fine-tune CAT on high-quality audio-visual-level instructions. + +Stage-I: feature alignment for audio and visual projectors. First, we employ the video-text Webvid 2.5M dataset [3] to train the visual projector (audio projector not involved in training). Second, we employ the audio-text WavCap dataset [42] to train the audio projector (visual projector not involved in training). During the above two training periods, we freeze the LLM and the encoder to align the semantic information for vision and audio, respectively. + +Stage-II: audio-visual joint instruction tuning. To equip CAT with the ability to reason jointly based on visual and auditory, we collect an audio-visual joint instruction dataset, named AVinstruct, which emphasizes co-learning dynamic audio-visual pairs to solve diverse AVQA tasks. Specifically, we collect a large number of raw videos containing audio information from YouTube and VGGSound [5] as well as QA pairs from the training set in closed-ended AVQA tasks (i.e. music scene [26] and real-world scene [59]). Then different subtitle generators (i.e. BLIP2 [27] and Whisper [51]) are utilized to obtain video descriptions. Finally, we use GPT to generalize from human-written examples to synthesize question-guided audio-visual descriptions. We integrate all compositions after stage-I training and freeze the visual projector and audio projector, only fine-tuning clue aggregator and combined LoRA parameters on 100k video instruction [40] and AVinstruct. Notably, to highlight the question for applying in the clue aggregator, we reconstruct the form of the input prompt by adding two simple tokens $< Q >$ , $< /Q>$ , where $< Q>$ denotes the start position of the question and $< /Q>$ denotes the end position of the question. During the instruction fine-tuning phase, we use predefined prompts based on the template: + +$$ +\text {U S E R}: < \text {s y s t e m} > < Q > < / Q > < \text {v i d e o} > < \text {a u d i o} > < \text {c l u e s} > \quad \text {A s s i s t a n t}: +$$ + +In this prompt, $<$ system $>$ denotes the official guidance message. $<$ video $>$ , $<$ audio $>$ , and $<$ clues $>$ are associated with visual tokens, audio tokens, and question-related clue tokens, respectively. + +# 3.4 AI-assisted Ambiguity-aware Direct Preference Optimization + +At first, we attempt to design various prompts to allow MLLMs to generate the most concise descriptions. However, such a no-learning approach is challenging to achieve the goal of accurate responses in a variety of dynamic audio-visual + +![](images/70c78657b54bb5127b4417b49c3264437e266ec442a7a749c20b0a55bc01a48b.jpg) +Fig. 4: Trained-CAT denotes CAT after feature alignment and instruction tuning. Our proposed ADPO strategy involves two steps. First, we collect the negative response generated by trained-CAT and correct it by GPT to obtain a positive response based on the original answer. Second, we perform ADPO training to skew CAT toward positive responses and reject negative responses. + +scenarios. We find that MLLMs are easy to ambiguity when describing specific audio-visual objects. For example, an ambiguous response is illustrated by trained-CAT in Fig. 4, where trained-CAT denotes CAT after feature alignment and instruction tuning. ADPO is designed to allow MLLMs to review their mistakes and relearn with a new objective to better expression. Specifically, ADPO is divided into two steps to optimize CAT: + +Collecting ambiguous samples and rewriting. As illustrated on the left in Fig. 4, assume that a given training set $D_{train}$ and the corresponding trained-CAT output set $D_{des}$ . We first provide GPT with a detailed prompt template to review each output in $D_{des}$ for ambiguity. The rule of the review is to identify output ambiguities, i.e., responses that differ significantly from the original answer, which is referred to as a negative response. We let GPT correct this negative response, with the principle of generating a positive response that clearly describes the visual objects or sound without significant modification. Repeat the above steps to get the sample set $D_{dpo}$ to be optimized, where $D_{dpo} = \{x^{vid}, x^{aud}, x^{cue}, x^{txt}, y_{pos}, y_{neg}\}$ . $x$ denotes the input modal token and $x^{txt}$ denotes the text tokens, $y_{pos}$ , $y_{neg}$ denote the positive and negative responses, respectively. + +Encouraging CAT to favor positive responses. As illustrated on the right in Fig. 4, ADPO directly optimizes the low-rank adaptation matrix parameters [17] in the policy model. We assume CAT as a policy model $f^{pol}$ , given a reference model $f^{ref}$ , a positive response $y_{pos}$ , and a negative response $y_{neg}$ . The reference model $f^{ref}$ is a deep copy of $f^{pol}$ and during training we encourage $f^{pol}$ to favor $y_{pos}$ , while $f^{ref}$ does not update the weights. Specifically, we define DPO loss $\mathcal{L}_{DPO}$ as: + +$$ +\begin{array}{l} \mathcal {L} _ {D P O} \left(f ^ {p o l}; f ^ {r e f}\right) = - E _ {\left(x ^ {v i d}, x ^ {a u d}, x ^ {c u e}, x ^ {t x t}, y _ {p o s}, y _ {n e g}\right)} \sim D _ {d p o} \\ \left[ \right. \log \sigma \left(\beta \log \frac {f ^ {p o l} \left(y _ {p o s} \mid [ x ^ {v i d}: x ^ {a u d}: x ^ {c u e}: x ^ {t x t} ]\right)}{f ^ {r e f} \left(y _ {p o s} \mid [ x ^ {v i d}: x ^ {a u d}: x ^ {c u e}: x ^ {t x t} ]\right)} \right. \quad (4) \\ \left. - \beta l o g \frac {f ^ {p o l} \left(y _ {n e g} \mid \left[ x ^ {v i d}: x ^ {a u d}: x ^ {c u e}: x ^ {t x t} \right]\right)}{f ^ {r e f} \left(y _ {n e g} \mid \left[ x ^ {v i d}: x ^ {a u d}: x ^ {c u e}: x ^ {t x t} \right]\right)}\right) \Bigg ], \\ \end{array} +$$ + +[ ] denotes the concatenation. $\sigma$ is the non-linear function, Sigmoid, and $\beta$ is a hyperparameter. This objective function aims to directly optimize the + +model that favors the positive response $y_{pos}$ and rejects the negative response $y_{neg}$ . Assuming that $\beta = 1$ , we simplify the output $\hat{r}$ of the model based on Eq. 4 as: + +$$ +\hat {r} \left(x ^ {v i d}, x ^ {a u d}, x ^ {c u e}, x ^ {t x t}, y\right) = \log \frac {f ^ {p o l} (y \mid \left[ x ^ {v i d}: x ^ {a u d}: x ^ {c u e}: x ^ {t x t} \right])}{f ^ {r e f} (y \mid \left[ x ^ {v i d}: x ^ {a u d}: x ^ {c u e}: x ^ {t x t} \right])}. \tag {5} +$$ + +The optimal goal is to make the variance of $\hat{r}(x^{vid}, x^{aud}, x^{cue}, x^{txt}, y_{pos}) - \hat{r}(x^{vid}, x^{aud}, x^{cue}, x^{txt}, y_{neg})$ larger and larger as a way to steer the model to generate descriptions devoid of ambiguous. However, when the difference between positive and negative responses is small, DPO loss supervision alone is not an effective facilitator. Therefore, we introduce an additional objective $\mathcal{L}_{SFT}$ , which is similar to the process of supervised fine-tuning, to supervise the model to stabilize the bias towards the positive response. Specifically, we define $\mathcal{L}_{SFT}$ as: + +$$ +\begin{array}{l} \mathcal {L} _ {S F T} = - \sum \log P \left(y _ {\text {p o s}} \mid \left[ x ^ {\text {v i d e o}}: x ^ {\text {a u d i o}}: x ^ {\text {c l u e s}}: x ^ {\text {t x t}} \right]; f ^ {\text {p o l}}\right), \tag {6} \\ \left\{x ^ {v i d}, x ^ {a u d}, x ^ {c u e}, x ^ {t x t}, y _ {p o s} \right\} \sim D _ {d p o}, \\ \end{array} +$$ + +this loss function can be interpreted as using the positive output probability in the policy model to aid in optimization. During ADPO training, we sum up $\mathcal{L}_{DPO}$ and $\mathcal{L}_{SFT}$ with a bias $\lambda$ to achieve direct preference optimization: + +$$ +\mathcal {L} = \mathcal {L} _ {D P O} + \lambda \mathcal {L} _ {S F T}, \tag {7} +$$ + +where we set $\lambda = 0.1$ by default. + +# 4 Experiments + +# 4.1 Datasets + +Video-based text generation tasks. We evaluate the proposed CAT on the video understanding benchmarks proposed by Video-ChatGPT [40]. Specifically, Video-ChatGPT proposes five metrics: Correctness of Information, Consistency, Detail Orientation, Contextual Understanding, and Temporal Understanding, which test the ability of MLLMs to describe videos. + +Zero-shot on video question answering tasks. To evaluate whether CAT has the basic ability to communicate regarding video, we conduct zero-shot tests on MSRVTT-QA [58] and ActivityNet-QA [60]. MSRVTT-QA and ActivityNet-QA consist of 10k and 5.8k videos containing audio information, respectively, where the QA pairs are mostly questions about daily life. + +Closed-ended AVQA tasks. We categorize the Music-AVQA [26] and AVQA [59] datasets as closed-ended AVQA tasks. These datasets consist of up to 42 candidate answers that require the selection of the most appropriate one based on visual and auditory content. + +Open-ended AVQA tasks. We select audio-visual dialogue (AVSD [1]), and audio-visual captioning (VALOR [6]) tasks for the evaluation of open-ended AVQA. These tasks require precise language to interpret, correlate, and reason about cross-modal information. We evaluate the zero-shot ability of CAT on these datasets. + +Table 1: GPT-based evaluation [40] for video-based text generation and zero-shot video question answering. For clarity, five scores are reported ("Cr.": Correctness of Information, "Cs.": Consistency, "De.": Detail Orientation, "Ct": Contextual Understanding, "Te.": Temporal Understanding). Score indicates the confidence. + +
MethodLLM SizeVideo-based text generationZero-shot on MSRVTT-QA Acc.ScoreZero-shot on ActivityNet-QA Acc.Score
Cr.De.Ct.Te.Cs.
LLaMA-VID [31]13B61.461.072.051.652.658.93.347.53.3
Video-LLaMA [64]7B39.243.643.236.435.829.61.812.41.1
LLaMA-Adapter [65]7B40.646.446.039.643.043.82.734.22.7
VideoChat [28]7B44.650.050.638.844.845.02.526.52.2
Video-ChatGPT [40]7B48.050.452.439.647.449.32.835.22.7
VISTA-LLaMA [39]7B48.852.863.645.246.260.53.348.33.3
VideoChat2 [29]7B60.457.870.253.256.254.13.349.13.3
Chat-UniVi [20]7B57.858.269.257.856.254.63.145.83.2
LLaMA-VID [31]7B59.260.070.649.250.257.73.247.13.3
CAT (Ours)7B61.659.069.856.257.862.13.550.23.5
+ +# 4.2 Experimental Setup + +Evaluation metrics. For video-based text generation and zero-shot on video question-answering tasks, we follow the evaluation pipeline proposed by VideoChatGPT [40], which is based on GPT-3.5 to evaluate predictive descriptions against correct descriptions, with a score from 0 to 5 indicating accuracy. In this paper, we standardize 0 to 5 as 0 to 100 to align common accuracy rubrics. For closed-ended AVQA tasks, we report the accuracy of the correct sample. For open-ended AVQA tasks, we report the CIDEr [56] that specializes in evaluating visually descriptive tasks. + +Architecture. We use frozen ImageBind [13] and LLaMA2-7B [55] as audiovisual encoders and LLM, respectively. The size of modality embeddings for each modality are $\mathbb{R}^{T\times 1024}$ . The outputs $x^{vid}$ , $x^{aud}$ , and $x^{cue}$ are $\mathbb{R}^{1\times 4096}$ , $\mathbb{R}^{1\times 4096}$ , and $\mathbb{R}^{48\times 4096}$ , respectively. + +Training details. We complete feature alignment training, instruction tuning, and ADPO training with 1 NVIDIA A100 GPU. In detail, For the feature alignment training and instruction tuning, we use the AdamW optimizer with a cosine learning rate decay and a warm-up period. When LoRA is added, we set $r = 64$ and alpha $= 128$ for the LoRA parameters, and the total batch size is set to 128 for training 1 epoch with a learning rate of $2e^{-5}$ . For ADPO training, we only select the training set in the instruction tuning phase for optimization. We set $r = 64$ and alpha $= 16$ for the LoRA parameters, and the total batch size is set to 1 for training 1 epoch with a learning rate of $4e^{-6}$ . The hyperparameter $\beta$ is set to 0.1. + +# 4.3 Comparison to State-of-the-Art + +Comparison on video-based text generation tasks. We follow the benchmark proposed by Video-ChatGPT [40] to evaluate CAT. On the left of Table 1, + +Table 2: Comparison with non-LLMs-based methods on fine-tuned Music-AVQA dataset. + +
MethodLanguage ModelTrainable Params (M)Audio avg.Visual avg.Audio-Visual avg.Overall avg.
FCNLSTM [11]CLIP [50]9.7968.956.260.460.8
GRU [2]CLIP-68.367.063.065.0
HCAattn [36]CLIP-64.965.360.362.5
MCAN [61]CLIP56.070.671.861.565.8
PSAC [30]CLIP-72.069.463.666.6
HME [10]CLIP-69.968.864.866.7
HCRN [24]CLIP-63.765.249.856.3
AVSD [54]CLIP8.3568.870.365.467.3
Panp-AVQA [62]CLIP-72.173.267.069.5
AVST [26]CLIP18.4873.974.469.571.6
PSTP-Net [25]BERT [9]4.29770.977.372.673.5
LAVISH [34]CLIP21.0977.177.377.0-
CAD [43]GLoVE [48]-78.179.776.978.2
VALOR [6]BERT----78.9
VAST [7]BERT----80.7
CAT-7B (Ours)LLaMA2 [55]5.81384.986.183.284.3
+ +we show that CAT achieves the state-of-the-art in terms of correctness of description information (Cr.), detailed description of the problem (De.), and coherence of the description (Cs.). In addition, CAT does not parse longer visual tokens, it still achieves competitive results when comparing time-related descriptions (e.g., contextual and temporal understanding). + +Comparison on zero-shot video question answering tasks. On the right of Table 1, we show the zero-shot video question answering performance of CAT on several open-ended datasets. While recent MLLMs designed with bridging modules have produced substantial results, CAT is way ahead of them in recognizing accurate answers. We consistently outperform the state-of-the-art on the MSRVTT-QA [58] and ActivityNet-QA [60] benchmarks. + +Comparison on closed-ended AVQA tasks. We choose to evaluate Music-AVQA [26] to demonstrate that CAT can perceive specific audio-visual objects to answer questions. In Table 2, under full supervision of the training set, CAT accurately retrieves specific objects in dynamic audio-visual scenarios and comprehensively outperforms all non-LLMs-based models. Thanks to LoRA [17], we can improve the evaluation quality with the help of the world knowledge inside LLaMA2 [55] and only 5.8M parameters are trainable. Also, we examine the ability of CAT for zero-shot on Music-AVQA [26]. In Table 3, we show the results comparing LLMs-based models. Notably, for a fair comparison, we remove the LoRA parameters of CAT that are fine-tuned on AVinstruct, which is derived from the Music-AVQA [26] and AVQA [59] training sets. Even though our model does not draw on the larger-scale LLM, it still achieves a small advantage over ChatBridge [67] with a 13B LLM size. Furthermore, we show the performance of CAT in multiple-choice scenarios [59] in Table 4, CAT continues outstanding. + +Comparison on open-ended AVQA tasks. Open-ended AVQA tasks require responses to daily events based on audio-visual content. We conduct a comparative analysis of our model with multimodal-based LLMs: OneLLM [14], and ChatBridge [67]. The tests are divided into zero-shot and full supervision, and + +we only choose to evaluate the zero-shot to demonstrate the level of practical application of the CAT. In Table 5, CAT surpasses OneLLM and ChatBridge on zero-shot complex text reasoning tasks AVSD [1] and VALOR [6]. Moreover, our approach is close to the results of FA [44] using fully supervised training, further demonstrating the strong multimodal understanding of CAT. + +Table 3: Comparison with LLMs-based methods on zero-shot Music- AVQA dataset. Table 5: Evaluation on zero-shot open-ended + +
Methodzero-shotAcc.
OneLLM-7B [14]✓43.0
ChatBridge-13B [67]✓47.6
CAT-7B (Ours)✓48.6
Table 4: Evaluation on fine-tuned AVQA dataset.
MethodAcc.
HGA + HAVF [19]87.7
HCRN + HAVF [59]89.0
PSTP-Net [25]90.2
CAT-7B (Ours)92.0
OneLLM-7B [14]✓74.5
ChatBridge-13B [67]✓75.4
CAT-7B (Ours)✓79.0
+ +# 4.4 Ablations and Analyses + +In this subsection, we explore the effects of each component of CAT. We test the effect of input modal tokens on video-based text generation [40]. In addition, to avoid the huge expense associated with the GPT3.5 evaluation, we test each component of the clue aggregator on the test set of AVinstruct, as well as the generalizability of ADPO on AVSD [1]. + +Does question-related clues matter? Visual tokens and audio tokens already have the ability to distinguish video content. To explore whether the added question-related clues are meaningful, we conduct ablation experiments on the video-based text generation task in Table 6. We first test visual knowledge separately from question-related clues knowledge on video comprehension. Observations reveal that the results achieved when inputting question-related clues alone have been able to outperform the joint input of audio-visual knowledge. Further, we add clues knowledge to the original audio-visual knowledge, and the results demonstrate that the fusion of the three can fully satisfy the information required for LLM reasoning. + +Ablation on clue aggregator. We categorize a portion of the collected audiovisual joint instruction data into a test set and present the ablation results in Table 7. The evaluation indicators B, M, and C denote BLEU-4 [47], METEOR [22], and ROUGE-L, respectively. We split $\mathcal{B}_1$ and $\mathcal{B}_2$ and query subjects $h_v$ and $h_a$ to explore the impact of CA internals on reasoning about dynamic audiovisual pairs. We find that visual attention influences context much more than + +Table 6: Ablation experiment Table 7: Ablation exper- Table 8: Ablation exper- about input modals on video- iments about clue aggre- iment about ADPO on based text generation. gator on audio-visual joint AVSD [1]. instruction dataset. + +
Input-modalCr.De.Te.
\( x^{vid} \)\( x^{aud} \)\( x^{cue} \)
✓38.640.434.0
✓✓40.644.835.2
✓58.657.455.8
✓✓✓61.659.056.2
+ +
MethodB@4MR
B1 + B230.858.768.5
B1 + B2 w/o hv10.037.149.9
B1 + B2 w/o ha18.849.361.8
only B125.954.764.2
+ +
ModelADPOAVSD
B@4MC
Video-LLaMA×18.440.163.7
Video-LLaMA✓22.248.269.2
CAT (Ours)×28.956.274.8
CAT (Ours)✓34.259.879.0
+ +![](images/bf80cdb6e84cc68f7e9a9ca3719a766e1ee21e457d7275b45cd8de341e8c2b21.jpg) +(a) Different number of visual tokens on video-based text generation. + +![](images/b48cafed2be87849d7aed939d4e897da04a302f8358d1a5abbeb552b4ee1f9bf.jpg) +(b) Different number of clue tokens on video-based text generation. +Fig. 5: The impacts of input modal tokens. Avg. represents the average accuracy of temporal (Te.), consistency (CS.), and detail (De.). + +auditory. This is because vision has more information that can be mined and is more relevant to the question. Next, we reduce parameters to investigate whether the same performance can be achieved with a single block. However, since the question acts as a query, without $\mathcal{B}_2$ to recover the original modal length does not provide an advantage over two blocks. + +Number of input modal tokens. We evaluate the effect of the number of input modal tokens on video understanding [40]. As shown in Fig. 5, we first study the effect of the number of visual tokens on a time-related description task. Indeed, as the number of visual tokens increases, the accuracy in the evaluation of the various descriptions decreases, even if the change is not large, but it also confirms that the language model may not be able to reason based on longer visual tokens. Next, we examine the impact of the number of clue tokens set by the Q-former [27]. Clue aggregator uses Q-Former to transform frame-level features into specific tokens, and we find that it works best when the length of queries $K$ is set to 48. + +Impact of ADPO. As the paper explains, supervised fine-tuning alone still underperforms in audio-visual scenarios. We test the generalizability of ADPO on Video-LLaMA [64] and ADPO's effect on AVSD [1] in Table 8 to demonstrate the superiority of this learning strategy. C denotes the CiDER [56]. We find that for different MLLMs, ADPO does improve the descriptive power of the model without much learning cost, it brings different levels of enhancement to both Video-LLaMA and CAT. + +![](images/c3c690717c48db9cc5fabcca04f16195d68ec28dcbd5ccd581d2d68393f5ba02.jpg) + +![](images/e902a88381942ace40ae07de6ab88f2632e85273e0021b92a1e5e74b4d2a1ae4.jpg) + +![](images/fab1f9f6e3f4f0cf1ad743983e8b0f19ce9a4b2ef089a32befdfca93a45a52f2.jpg) +Fig. 6: Qualitative results of different types of question answering. + +![](images/1ca152d7e8057b5c6a16dc48545f02df74507b290251bff25a33d8a9fb4ad6ab.jpg) + +# 4.5 Qualitative Analysis + +In Fig. 6, we analyze the qualitative results with Video-LLaMA [64]. In the example of video-specific contextual inference, we show that CAT has a sharp perception of complex indoor scenes. Our descriptions can accurately represent what the person doing and what background sounds are in the video. Video-LLaMA is biased towards describing scene information and incorrect character information. In the specific audio question answering example, Video-LLaMA lost the ability to answer "How many instruments are sounding in the video?" In contrast, CAT can accurately answer the quantity question. In the audio-visual question answering example, Video-LLaMA answers incorrectly due to failed audio-visual grounding, while our CAT answers correctly due to its strong ability to capture specific objects in audio-visual scenarios. + +# 5 Conclusion + +In this work, we introduce CAT to enhance LLMs' multimodal understanding in dynamic audio-visual scenarios. We propose a clue aggregator to aggregate the question-related clues for detailed reasoning. We mix datasets containing audio and video to empower LLM with multimodal understanding. To more consistently infer audio-visual scenarios, we collect an audio-visual joint instruction dataset to further fine-tune the CAT. Moreover, we propose an AI-assisted ambiguity-aware direct preference optimization strategy to retrain the model for more accurate responses. CAT has achieved comparable results when applied to a variety of complex audio-visual scenarios. + +Acknowledgements. This work was supported by the National Natural Science Foundation of China under Grant 62306061, Guangdong Basic and Applied Basic Research Foundation (Grant No. 2023A1515140037), and the Open Fund of National Engineering Laboratory for Big Data System Computing Technology (Grant No. SZU-BDSC-OF2024-02). + +# References + +1. AlAmri, H., Cartillier, V., Das, A., Wang, J., Cherian, A., Essa, I., Batra, D., Marks, T.K., Hori, C., Anderson, P., Lee, S., Parikh, D.: Audio visual scene-aware dialog. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR. pp. 7558-7567 (2019) +2. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: VQA: visual question answering. In: IEEE International Conference on Computer Vision, ICCV. pp. 2425-2433. IEEE Computer Society (2015) +3. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: IEEE/CVF International Conference on Computer Vision, ICCV. pp. 1708-1718. IEEE (2021) +4. Chen, G., Shen, L., Shao, R., Deng, X., Nie, L.: LION: Empowering multimodal large language model with dual-level visual knowledge. CoRR abs/2311.11860 (2023) +5. Chen, H., Xie, W., Vedaldi, A., Zisserman, A.: Vggsound: A large-scale audiovisual dataset. In: IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP. pp. 721-725. IEEE (2020) +6. Chen, S., He, X., Guo, L., Zhu, X., Wang, W., Tang, J., Liu, J.: VALOR: vision-audio-language omni-perception pretraining model and dataset. CoRR abs/2304.08345 (2023) +7. Chen, S., Li, H., Wang, Q., Zhao, Z., Sun, M., Zhu, X., Liu, J.: VAST: A vision-audio-subtitle-text omni-modality foundation model and dataset. CoRR abs/2305.18500 (2023) +8. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.C.H.: Instructclip: Towards general-purpose vision-language models with instruction tuning. CoRR abs/2305.06500 (2023) +9. Devlin, J., Chang, M., Lee, K., Toutanova, K.: BERT: pre-training of deep bidirectional transformers for language understanding. In: Burstein, J., Doran, C., Solorio, T. (eds.) Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT. pp. 4171-4186. Association for Computational Linguistics (2019) +10. Fan, C., Zhang, X., Zhang, S., Wang, W., Zhang, C., Huang, H.: Heterogeneous memory enhanced multimodal attention model for video question answering. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR. pp. 1999-2007. Computer Vision Foundation / IEEE (2019) +11. Fayek, H.M., Johnson, J.: Temporal reasoning via audio question answering. IEEE ACM Trans. Audio Speech Lang. Process. 28, 2283-2294 (2020) +12. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., Li, H., Qiao, Y.: Llama-adapter V2: parameter-efficient visual instruction model. CoRR abs/2304.15010 (2023) +13. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind one embedding space to bind them all. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR. pp. 15180-15190 (2023) +14. Han, J., Gong, K., Zhang, Y., Wang, J., Zhang, K., Lin, D., Qiao, Y., Gao, P., Yue, X.: Onellm: One framework to align all modalities with language. CoRR abs/2312.03700 (2023) +15. Han, J., Zhang, R., Shao, W., Gao, P., Xu, P., Xiao, H., Zhang, K., Liu, C., Wen, S., Guo, Z., Lu, X., Ren, S., Wen, Y., Chen, X., Yue, X., Li, H., Qiao, Y.: Imagebind-llm: Multi-modality instruction tuning. CoRR abs/2309.03905 (2023) + +16. Hendrycks, D., Burns, C., Basart, S., Zou, A., Mazeika, M., Song, D., Steinhardt, J.: Measuring massive multitask language understanding. In: ICLR (2021) +17. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. In: The Tenth International Conference on Learning Representations, ICLR (2022) +18. Huang, B., Wang, X., Chen, H., Song, Z., Zhu, W.: Vtimellm: Empower LLM to grasp video moments. CoRR abs/2311.18445 (2023) +19. Jiang, P., Han, Y.: Reasoning with heterogeneous graph alignment for video question answering. In: AAAI. pp. 11109-11116. AAAI Press (2020) +20. Jin, P., Takanobu, R., Zhang, C., Cao, X., Yuan, L.: Chat-univi: Unified visual representation empowers large language models with image and video understanding. CoRR abs/2311.08046 (2023) +21. Korbar, B., Xian, Y., Tonioni, A., Zisserman, A., Tombari, F.: Text-conditioned resampler for long form video understanding. CoRR abs/2312.11897 (2023) +22. Lavie, A., Agarwal, A.: METEOR: an automatic metric for MT evaluation with high levels of correlation with human judgments. In: Callison-Burch, C., Koehn, P., Fordyce, C.S., Monz, C. (eds.) Proceedings of the Second Workshop on Statistical Machine Translation. pp. 228-231. Association for Computational Linguistics (2007) +23. Le, H., Sahoo, D., Chen, N.F., Hoi, S.C.H.: Multimodal transformer networks for end-to-end video-grounded dialogue systems. In: Korhonen, A., Traum, D.R., Marquez, L. (eds.) Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL. pp. 5612-5623. Association for Computational Linguistics (2019) +24. Le, T.M., Le, V., Venkatesh, S., Tran, T.: Hierarchical conditional relation networks for video question answering. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR. pp. 9969-9978. Computer Vision Foundation / IEEE (2020) +25. Li, G., Hou, W., Hu, D.: Progressive spatio-temporal perception for audio-visual question answering. In: El-Saddik, A., Mei, T., Cucchiara, R., Bertini, M., Vallejo, D.P.T., Atrey, P.K., Hossain, M.S. (eds.) ACM MM. pp. 7808-7816. ACM (2023) +26. Li, G., Wei, Y., Tian, Y., Xu, C., Wen, J.R., Hu, D.: Learning to answer questions in dynamic audio-visual scenarios. In: CVPR. p. 19086–19096 (2022) +27. Li, J., Li, D., Savarese, S., Hoi, S.C.H.: BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In: Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., Scarlett, J. (eds.) International Conference on Machine Learning, ICML. pp. 19730–19742. PMLR (2023) +28. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: Videochat: Chat-centric video understanding. CoRR abs/2305.06355 (2023) +29. Li, K., Wang, Y., He, Y., Li, Y., Wang, Y., Liu, Y., Wang, Z., Xu, J., Chen, G., Luo, P., Wang, L., Qiao, Y.: Mvbench: A comprehensive multi-modal video understanding benchmark. CoRR abs/2311.17005 (2023) +30. Li, X., Song, J., Gao, L., Liu, X., Huang, W., He, X., Gan, C.: Beyond rnns: Positional self-attention with co-attention for video question answering. In: AAAI. pp. 8658-8665. AAAI Press (2019) +31. Li, Y., Wang, C., Jia, J.: Llama-vid: An image is worth 2 tokens in large language models. CoRR abs/2311.17043 (2023) +32. Li, Z., Yang, B., Liu, Q., Ma, Z., Zhang, S., Yang, J., Sun, Y., Liu, Y., Bai, X.: Monkey: Image resolution and text label are important things for large multi-modal models. CoRR abs/2311.06607 (2023) + +33. Lin, B., Ye, Y., Zhu, B., Cui, J., Ning, M., Jin, P., Yuan, L.: Video-llava: Learning united visual representation by alignment before projection. CoRR abs/2311.10122 (2023) +34. Lin, Y., Sung, Y., Lei, J., Bansal, M., Bertasius, G.: Vision transformers are parameter-efficient audio-visual learners. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR. pp. 2299-2309. IEEE (2023) +35. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. CoRR abs/2304.08485 (2023) +36. Lu, J., Yang, J., Batra, D., Parikh, D.: Hierarchical question-image co-attention for visual question answering. In: Lee, D.D., Sugiyama, M., von Luxburg, U., Guyon, I., Garnett, R. (eds.) Advances in Neural Information Processing Systems NIPs. pp. 289–297 (2016) +37. Luo, R., Zhao, Z., Yang, M., Dong, J., Qiu, M., Lu, P., Wang, T., Wei, Z.: Valley: Video assistant with large language model enhanced ability. CoRR abs/2306.07207 (2023) +38. Lyu, C., Wu, M., Wang, L., Huang, X., Liu, B., Du, Z., Shi, S., Tu, Z.: Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. CoRR abs/2306.09093 (2023) +39. Ma, F., Jin, X., Wang, H., Xian, Y., Feng, J., Yang, Y.: Vista-llama: Reliable video narrator via equal distance to visual tokens. CoRR abs/2312.08870 (2023) +40. Maaz, M., Rasheed, H.A., Khan, S.H., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models. CoRR abs/2306.05424 (2023) +41. MacGlashan, J., Ho, M.K., Loftin, R.T., Peng, B., Wang, G., Roberts, D.L., Taylor, M.E., Littman, M.L.: Interactive learning from policy-dependent human feedback. In: Precup, D., Teh, Y.W. (eds.) Proceedings of the 34th International Conference on Machine Learning, ICML. pp. 2285-2294 (2017) +42. Mei, X., Meng, C., Liu, H., Kong, Q., Ko, T., Zhao, C., Plumbley, M.D., Zou, Y., Wang, W.: Wavcaps: A chatgpt-assisted weakly-labelled audio captioning dataset for audio-language multimodal research. CoRR abs/2303.17395 (2023) +43. Nadeem, A., Hilton, A., Dawes, R., Thomas, G., Mustafa, A.: CAD - contextual multi-modal alignment for dynamic AVQA. CoRR abs/2310.16754 (2023) +44. Nguyen, D.T., Sharma, S., Schulz, H., Asri, L.E.: From film to video: Multi-turn question answering with multi-modal context. CoRR abs/1812.07023 (2018) +45. OpenAI: GPT-4 technical report. CoRR abs/2303.08774 (2023) +46. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P.F., Leike, J., Lowe, R.: Training language models to follow instructions with human feedback. In: Advances in Neural Information Processing Systems, NeurIPS (2022) +47. Papineni, K., Roukos, S., Ward, T., Zhu, W.: Bleu: a method for automatic evaluation of machine translation. In: Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics. pp. 311-318. ACL (2002) +48. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: Moschitti, A., Pang, B., Daelemans, W. (eds.) Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, EMNLP. pp. 1532-1543. ACL (2014) +49. Pham, H., Le, T.M., Le, V., Phuong, T.M., Tran, T.: Video dialog as conversation about objects living in space-time. In: Avidan, S., Brostow, G.J., Cissé, M., Farinella, G.M., Hassner, T. (eds.) ECCV. vol. 13699, pp. 710-726. Springer (2022) + +50. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the International Conference on Machine Learning, ICML. vol. 139, pp. 8748-8763. PMLR (2021) +51. Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Robust speech recognition via large-scale weak supervision. In: Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., Scarlett, J. (eds.) International Conference on Machine Learning, ICML. pp. 28492-28518. PMLR (2023) +52. Rafailov, R., Sharma, A., Mitchell, E., Ermon, S., Manning, C.D., Finn, C.: Direct preference optimization: Your language model is secretly a reward model. CoRR abs/2305.18290 (2023) +53. Sakaguchi, K., Bras, R.L., Bhagavatula, C., Choi, Y.: Winogrande: An adversarial winograd schema challenge at scale. In: AAAI. pp. 8732-8740 (2020) +54. Schwartz, I., Schwing, A.G., Hazan, T.: A simple baseline for audio-visual scene-aware dialog. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR. pp. 12548-12558. Computer Vision Foundation / IEEE (2019) +55. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D., Blecher, L., Canton-Ferrer, C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I., Korenev, A., Koura, P.S., Lachaux, M., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A., Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X.E., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models. CoRR abs/2307.09288 (2023) +56. Vedantam, R., Zitnick, C.L., Parikh, D.: Cider: Consensus-based image description evaluation. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR. pp. 4566-4575 (2015) +57. Wang, X., Wu, J., Chen, J., Li, L., Wang, Y., Wang, W.Y.: Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In: IEEE/CVF International Conference on Computer Vision, ICCV. pp. 4580-4590. IEEE (2019) +58. Xu, D., Zhao, Z., Xiao, J., Wu, F., Zhang, H., He, X., Zhuang, Y.: Video question answering via gradually refined attention over appearance and motion. In: ACM MM. pp. 1645-1653 (2017) +59. Yang, P., Wang, X., Duan, X., Chen, H., Hou, R., Jin, C., Zhu, W.: AVQA: A dataset for audio-visual question answering on videos. In: ACM MM. pp. 3480-3491. ACM (2022) +60. Yu, Z., Xu, D., Yu, J., Yu, T., Zhao, Z., Zhuang, Y., Tao, D.: Activitynet-qa: A dataset for understanding complex web videos via question answering. In: AAAI. pp. 9127-9134 (2019) +61. Yu, Z., Yu, J., Cui, Y., Tao, D., Tian, Q.: Deep modular co-attention networks for visual question answering. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR. pp. 6281-6290. Computer Vision Foundation / IEEE (2019) +62. Yun, H., Yu, Y., Yang, W., Lee, K.I., Kim, G.H.: Pano-avqa: Grounded audiovisual question answering on $360^{\circ}$ videos. In: CVPR. p. 2031-2041 (2021) + +63. Zellers, R., Holtzman, A., Bisk, Y., Farhadi, A., Choi, Y.: Hellaswag: Can a machine really finish your sentence? In: Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL. pp. 4791-4800 (2019) +64. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. In: Proceedings of the Empirical Methods in Natural Language Processing, EMNLP. pp. 543-553 (2023) +65. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. CoRR abs/2303.16199 (2023) +66. Zhang, S., Dong, L., Li, X., Zhang, S., Sun, X., Wang, S., Li, J., Hu, R., Zhang, T., Wu, F., Wang, G.: Instruction tuning for large language models: A survey. CoRR abs/2308.10792 (2023) +67. Zhao, Z., Guo, L., Yue, T., Chen, S., Shao, S., Zhu, X., Yuan, Z., Liu, J.: Chatbridge: Bridging modalities with large language model as a language catalyst. CoRR abs/2305.16103 (2023) \ No newline at end of file diff --git a/2024/CAT_ Enhancing Multimodal Large Language Model to Answer Questions in Dynamic Audio-Visual Scenarios/images.zip b/2024/CAT_ Enhancing Multimodal Large Language Model to Answer Questions in Dynamic Audio-Visual Scenarios/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..27bf4fe48f9c2843d8ddeb6098cf4aff781c0363 --- /dev/null +++ b/2024/CAT_ Enhancing Multimodal Large Language Model to Answer Questions in Dynamic Audio-Visual Scenarios/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97840172fb3c60bad449a26f32b9c745a30ebcd7328a8a0ee89632ee90790026 +size 544363 diff --git a/2024/CAT_ Enhancing Multimodal Large Language Model to Answer Questions in Dynamic Audio-Visual Scenarios/layout.json b/2024/CAT_ Enhancing Multimodal Large Language Model to Answer Questions in Dynamic Audio-Visual Scenarios/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e8543c1bc1e9dee7b5bbe5a4c6681a795c4b1fac --- /dev/null +++ b/2024/CAT_ Enhancing Multimodal Large Language Model to Answer Questions in Dynamic Audio-Visual Scenarios/layout.json @@ -0,0 +1,9884 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 114, + 478, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 114, + 478, + 173 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 478, + 173 + ], + "type": "text", + "content": "CAT: Enhancing Multimodal Large Language Model to Answer Questions in Dynamic Audio-Visual Scenarios" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 196, + 194, + 416, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 194, + 416, + 220 + ], + "spans": [ + { + "bbox": [ + 196, + 194, + 416, + 220 + ], + "type": "text", + "content": "Qilang Ye" + }, + { + "bbox": [ + 196, + 194, + 416, + 220 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 196, + 194, + 416, + 220 + ], + "type": "text", + "content": ", Zitong Yu" + }, + { + "bbox": [ + 196, + 194, + 416, + 220 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 196, + 194, + 416, + 220 + ], + "type": "text", + "content": "*, Rui Shao" + }, + { + "bbox": [ + 196, + 194, + 416, + 220 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 196, + 194, + 416, + 220 + ], + "type": "text", + "content": ", Xinyu Xie" + }, + { + "bbox": [ + 196, + 194, + 416, + 220 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 196, + 194, + 416, + 220 + ], + "type": "text", + "content": ", Philip Torr" + }, + { + "bbox": [ + 196, + 194, + 416, + 220 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 196, + 194, + 416, + 220 + ], + "type": "text", + "content": ", and Xiaochun Cao" + }, + { + "bbox": [ + 196, + 194, + 416, + 220 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 258, + 228, + 356, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 228, + 356, + 239 + ], + "spans": [ + { + "bbox": [ + 258, + 228, + 356, + 239 + ], + "type": "text", + "content": "1 Great Bay University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 217, + 239, + 395, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 239, + 395, + 251 + ], + "spans": [ + { + "bbox": [ + 217, + 239, + 395, + 251 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 217, + 239, + 395, + 251 + ], + "type": "text", + "content": " Harbin Institute of Technology, Shenzhen" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 259, + 251, + 353, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 251, + 353, + 261 + ], + "spans": [ + { + "bbox": [ + 259, + 251, + 353, + 261 + ], + "type": "text", + "content": "3 University of Oxford" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 211, + 262, + 402, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 262, + 402, + 273 + ], + "spans": [ + { + "bbox": [ + 211, + 262, + 402, + 273 + ], + "type": "text", + "content": "4 Shenzhen Campus of Sun Yat-sen University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 301, + 455, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 301, + 455, + 521 + ], + "spans": [ + { + "bbox": [ + 159, + 301, + 455, + 521 + ], + "type": "text", + "content": "Abstract. This paper focuses on the challenge of answering questions in scenarios that are composed of rich and complex dynamic audio-visual components. Although existing Multimodal Large Language Models (MLLMs) can respond to audio-visual content, these responses are sometimes ambiguous and fail to describe specific audio-visual events. To overcome this limitation, we introduce the CAT, which enhances MLLM in three ways: 1) besides straightforwardly bridging audio and video, we design a clue aggregator that aggregates question-related clues in dynamic audio-visual scenarios to enrich the detailed knowledge required for large language models. 2) CAT is trained on a mixed multimodal dataset, allowing direct application in audio-visual scenarios. Notably, we collect an audio-visual joint instruction dataset named AVinstruct, to further enhance the capacity of CAT to model cross-semantic correlations. 3) we propose AI-assisted ambiguity-aware direct preference optimization, a strategy specialized in retraining the model to favor the non-ambiguity response and improve the ability to localize specific audio-visual objects. Extensive experimental results demonstrate that CAT outperforms existing methods on multimodal tasks, especially in Audio-Visual Question Answering (AVQA) tasks. The codes and the collected instructions will be released soon." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 531, + 453, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 531, + 453, + 554 + ], + "spans": [ + { + "bbox": [ + 160, + 531, + 453, + 554 + ], + "type": "text", + "content": "Keywords: Multimodal Large Language Model " + }, + { + "bbox": [ + 160, + 531, + 453, + 554 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 531, + 453, + 554 + ], + "type": "text", + "content": " Audio-visual Question Answering" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 574, + 230, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 230, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 230, + 586 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": "The real world revolves around sound and visual information, and their combination enhances our ability to perceive the world. Similarly, the development of Multimodal Large Language Models (MLLMs) [14, 64, 67] are closely related to audio and visual. Supervised fine-tuning [35, 46, 66] with specific instruction" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 314, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 314, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 314, + 665 + ], + "type": "text", + "content": "* Corresponding author: zitong.yu@ieee.org" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 177, + 114, + 302, + 195 + ], + "blocks": [ + { + "bbox": [ + 177, + 114, + 302, + 195 + ], + "lines": [ + { + "bbox": [ + 177, + 114, + 302, + 195 + ], + "spans": [ + { + "bbox": [ + 177, + 114, + 302, + 195 + ], + "type": "image", + "image_path": "94ba68624e535d3e41b8ad7f78d5f5d088bc827f024ed4183a87c00d0b6df2b2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 177, + 196, + 301, + 281 + ], + "blocks": [ + { + "bbox": [ + 177, + 196, + 301, + 281 + ], + "lines": [ + { + "bbox": [ + 177, + 196, + 301, + 281 + ], + "spans": [ + { + "bbox": [ + 177, + 196, + 301, + 281 + ], + "type": "image", + "image_path": "30c64ddfd88c267d6bebac2a77f61be0d21e375496b704fe99e8409911789191.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 293, + 482, + 395 + ], + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 395 + ], + "type": "text", + "content": "Fig. 1: Comparison between existing MLLMs and CAT Red words for incorrect response, Green words for correct response, and Gray words for useless response. Left: Most of the existing MLLMs straightforwardly bridge multimodal to large language models. Instead, CAT builds on this foundation by designing the clue aggregator to learn more detailed knowledge related to the question. Moreover, CAT constrains itself to learn a sharper response through AI-assisted ambiguity-aware direct preference optimization. Right: In comparison with audio-visual-language models Video-LLaMA [64] and ChatBridge [67], our method accurately recognizes the answers to questions with the most streamlined responses." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 308, + 116, + 436, + 281 + ], + "blocks": [ + { + "bbox": [ + 308, + 116, + 436, + 281 + ], + "lines": [ + { + "bbox": [ + 308, + 116, + 436, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 116, + 436, + 281 + ], + "type": "image", + "image_path": "c3206106d4c5640b37a232e86ee5b90c2156d52af42671be69cab6d971d2573c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 422, + 482, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 530 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 530 + ], + "type": "text", + "content": "datasets empowers Large Language Models (LLMs) for multimodal understanding. Despite MLLMs' strong causal reasoning achieving impressive results in common-sense answering [53,63], academic answering [16], etc., predicting questions in dynamic audio-visual scenarios remains challenging. This is due to the difficulty of aligning LLMs with cross-domain data during training on large-scale multimodal corpora. It leads to particular ambiguity when describing specific objects in dynamic audio-visual scenarios. These interferences not only cause the model to make random guesses but also seriously affect the inference of detailed answers in Audio-Visual Question Answering (AVQA) [1,26,59,62] tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 531, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 482, + 604 + ], + "type": "text", + "content": "A series of bridging modules [12,20,31,32,39] for MLLMs have been designed. The simplest methods [15, 18, 28, 33] use projectors to directly align text with other modalities, but partially limit the ability to capture fine-grained information. In addition, using a cross-attention mechanism to query the audio-visual context [8,21,38,39] to solve multimodal alignment problems is effective but still occurs that a certain visual object or sound cannot be localized in practice." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "In this work, we explore possible factors contributing to the above failures: 1) Audio-visual insufficiently correlates with the question. As illustrated in the upper left of Fig. 1, most existing MLLMs [28,64,67] are mainly designed with multiple branches to handle multiple modalities individually, followed by concatenating the modality embeddings with prompts as inputs to the LLM." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Q. Ye et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "This paradigm cannot allow the network to learn paired text and audiovisuals, and the tandem embedding is not relevant. 2) Alignment of multimodal with text is challenging. The multimodal-text corpus [67] is difficult to be aligned, making the model generation of responses sometimes ambiguous. This ambiguity generally manifests in ambiguous words of the corresponding audio-visual content, as well as in generating useless text that overly responds. For example, the descriptions illustrated by Video-LLaMA [64] and ChatBridge [67] in Fig. 1 are all relatively ambiguous in describing the question-related video and audio content. These descriptions fail to identify the correct answer for \"bagpipes\" when asked for audio-visual reasoning. In addition, when the description is accompanied by many useless and controversy-prone words, it is unfavorable to evaluate the closed-ended AVQA task or open-ended AVQA task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 264, + 482, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 264, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 482, + 575 + ], + "type": "text", + "content": "To overcome the two aforementioned issues, we introduce the CAT, enhancing MLLM in three ways: 1) Aggregation of question-related key clues. As illustrated in the lower left of Fig. 1, besides the global visual and audio information, we design a clue aggregator to dynamically capture question-aware visual and audio hidden features to enrich fine-grained clues. The aggregator receives sufficient low-level textual information to interact with audio-visual, it is capable of improving audio-visual grounding. 2) Mixed audio-visual training strategy. The training of CAT includes feature alignment using video-text pairs and audio-text pairs, and high-quality instructions to enhance audio-visual awareness. This strategy allows the CAT to be directly involved in real-world scenarios containing both visual and sound. Notably, we collect an audio-visual joint instruction dataset, named AVinstruct, to further empower CAT in AVQA tasks. 3) Retraining MLLMs to mitigate ambiguity. The DPO proposed by Rafailov et al. [52] has inspired us. Although MLLMs after training are equipped with multimodal understanding, they lack the flexibility to perceive ambiguity. Therefore, we reframe ambiguity elimination as a model preference optimization process and propose an AI-assisted Ambiguity-aware Direct Preference Optimization (ADPO) strategy. Specifically, we refer to ambiguous responses that express the lack of clarity of specific audio-visual objects as negative responses, and then we collect negative responses in the training set and utilize GPT to rewrite them into positive responses. After the multimodal training, we perform ADPO to retrain the model to bias towards the positive response, which is the precise description after the rewrite, and reject the negative response, which is the ambiguous description. Through this learning strategy, CAT can constrain itself to favor non-ambiguity responses. As shown in Fig. 1 on the right, CAT correctly responds to the question and excludes all useless information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 579, + 372, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 579, + 372, + 590 + ], + "spans": [ + { + "bbox": [ + 146, + 579, + 372, + 590 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 617, + 482, + 665 + ], + "type": "text", + "content": "- We introduce a novel audio-visual-language model, dubbed as CAT, that is capable of learning question-related clues and engaging directly in dynamic audio-visual inference. Notably, we collect AVInstruct, an audio-visual joint instruction dataset to ensure the stability of CAT in AVQA tasks." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 426, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 426, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 426, + 91, + 447, + 100 + ], + "type": "text", + "content": "CAT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 481, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 481, + 151 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 481, + 151 + ], + "type": "text", + "content": "- As a powerful learning strategy, we propose AI-assisted ambiguity-aware direct preference optimization to overcome the problem that MLLMs tend to ambiguously describe specific audio-visual objects." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 152, + 481, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 152, + 481, + 211 + ], + "spans": [ + { + "bbox": [ + 138, + 152, + 481, + 211 + ], + "type": "text", + "content": "- We evaluate CAT on a wide range of multimodal tasks. Extensive experiments demonstrate the superiority of CAT (e.g., outperforms the state-of-the-art on a variety of AVQA tasks, and achieves remarkable results in the evaluation of video-based text generation tasks and zero-shot video question-answering tasks)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 225, + 242, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 242, + 238 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 242, + 238 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 247, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 390 + ], + "type": "text", + "content": "Audio-visual question answering (AVQA). AVQA aims to produce the most accurate linguistic representation of a given video and audio based on the question content, which requires multimodal understanding and reasoning across different semantic levels. The earliest studies [57, 58, 60] emphasize the understanding of the whole video and return a simple word that is relatively correct to the question. Subsequently, the emergence of dynamic audio-visual datasets (e.g., music scenes [26, 59], and 360-degree panoramicas [62]) increases the challenge of Question-Answering (QA), which requires the mining of temporal and spatial information in different modalities. For instance, Music-AVQA [26] requires distinguishing how many instruments are present based on the audio content. AVQA [59] requires finding the most plausible of multiple options based on the audio-visual clues from dynamic scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 569 + ], + "type": "text", + "content": "MLLMs for audio-visual question answering. Extending LLM to other multimodal tasks is a recently emerging field. Despite the ability of MLLMs to combine information between different modalities, performance on downstream tasks, especially AVQA, remains suboptimal. Many works [8, 20, 39] emphasize the design of elegant bridging methods to improve the performance of question answering. The simplest bridging modules [37,40,64] use one or more linear layers for feature alignment. Although such methods minimize parameter updates, they still have limitations in exploring fine-grained information. Others design more complex bridging networks to query visual information. For example, Lyu et al. [38] propose an alignment module for harmonizing different representations before entering the LLM. Ma et al. [39] propose to keep the distance between all visual tokens and any linguistic tokens consistent within the LLM. However, a large multimodal corpus is difficult to align during training, neither algorithms with low parameter counts [12] nor complex bridging networks [4] are prone to the problem of failing to accurately depict specific audio-visual events." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Human-preference learning. Reinforcement Learning from Human Feedback (RLHF) [41, 46] is the most classical instance of human preference learning [45, 55], it constructs a reward model to optimize the policy model to favor preference responses. Dai et al. [8] have demonstrated that such preference learning can enhance LLM to generate more accurate information. Recently, Rafailov et al. [52] propose a Direct Preference Optimization (DPO) strategy that learns preferences directly by bypassing learning reward models, this simple yet efficient approach inspires us to solve the ambiguity problem." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Q. Ye et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 116, + 475, + 232 + ], + "blocks": [ + { + "bbox": [ + 132, + 116, + 475, + 232 + ], + "lines": [ + { + "bbox": [ + 132, + 116, + 475, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 475, + 232 + ], + "type": "image", + "image_path": "fd1da11cfd1466c864099619b84d1827ca664efd9239b548a73b98444c76f8ec.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 243, + 482, + 335 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 335 + ], + "type": "text", + "content": "Fig. 2: Illustration of the proposed CAT and its training strategy. (a) Overview of CAT. CAT first extracts overall audio-visual knowledge from video and audio and transforms them into visual tokens " + }, + { + "bbox": [ + 130, + 243, + 482, + 335 + ], + "type": "inline_equation", + "content": "x^{vid}" + }, + { + "bbox": [ + 130, + 243, + 482, + 335 + ], + "type": "text", + "content": " and audio tokens " + }, + { + "bbox": [ + 130, + 243, + 482, + 335 + ], + "type": "inline_equation", + "content": "x^{aud}" + }, + { + "bbox": [ + 130, + 243, + 482, + 335 + ], + "type": "text", + "content": ". We input question tagged with " + }, + { + "bbox": [ + 130, + 243, + 482, + 335 + ], + "type": "inline_equation", + "content": "<\\mathrm{Q}> " + }, + { + "bbox": [ + 130, + 243, + 482, + 335 + ], + "type": "text", + "content": " in the prompt into the clue aggregator, aiming to aggregate question-aware audio-visual hidden features and yield clue tokens " + }, + { + "bbox": [ + 130, + 243, + 482, + 335 + ], + "type": "inline_equation", + "content": "x^{cue}" + }, + { + "bbox": [ + 130, + 243, + 482, + 335 + ], + "type": "text", + "content": ". Finally, we merge multimodal tokens and language and feed into the frozen large language model with LoRA [17] to output the response. (b) The training paradigm of CAT involves pre-alignment of the audio-visual projectors and instruction tuning on the entire model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 357, + 240, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 240, + 370 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 240, + 370 + ], + "type": "text", + "content": "3 Our Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 381, + 482, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 482, + 478 + ], + "type": "text", + "content": "In this section, we first present in detail the proposed CAT (Sec. 3.1 and 3.2). As shown on the left in Fig. 2, CAT consists of three branches that draw on visual knowledge, audio knowledge, and question-related clues to feed into the LLM, respectively. Second, we present the multimodel training strategy for CAT as shown on the right in Fig. 2, where a high-quality audio-visual joint instruction dataset is collected for further fine-tuning (Sec. 3.3). Lastly, we introduce the AI-assisted Ambiguity-aware Direct Preference Optimization (ADPO) strategy, which reinforces CAT to favor non-ambiguous descriptions (Sec. 3.4)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 495, + 257, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 257, + 507 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 257, + 507 + ], + "type": "text", + "content": "3.1 Multimodal Inputs" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 514, + 481, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 481, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 481, + 562 + ], + "type": "text", + "content": "ImageBind [13] with a single joint embedding space demonstrates strong modality learning ability. Therefore, we leverage the frozen ImageBind as a universal encoder for all modalities. Given a video " + }, + { + "bbox": [ + 130, + 514, + 481, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 130, + 514, + 481, + 562 + ], + "type": "text", + "content": " and an audio " + }, + { + "bbox": [ + 130, + 514, + 481, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 130, + 514, + 481, + 562 + ], + "type": "text", + "content": ", the encoded multimodal hidden features can be obtained by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 208, + 571, + 481, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 571, + 481, + 585 + ], + "spans": [ + { + "bbox": [ + 208, + 571, + 481, + 585 + ], + "type": "interline_equation", + "content": "h _ {v} = \\operatorname {I m a g e B i n d} (\\mathbf {V}), \\quad h _ {a} = \\operatorname {I m a g e B i n d} (\\mathbf {A}), \\tag {1}", + "image_path": "2a4754776adbfe00ada384136bc317359b536cfe40447cf473a487c524655e12.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "h_v \\in \\mathbb{R}^{T \\times d_h}" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "h_a \\in \\mathbb{R}^{1 \\times d_h}" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": " are video and audio features, respectively. " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": " denotes the length of the given video and " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "d_h" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": " is the specific dimension. Notably, we temporally compress the frame-level features " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "h_v" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": " to address the computational burden. Further, we employ two linear projection layers to align the hidden dimensions of the inputs " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "h_v" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "h_a" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": " to obtain visual tokens " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "x^{v_id}" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": " and audio tokens " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "x^{aud}" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 426, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 426, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 426, + 91, + 447, + 100 + ], + "type": "text", + "content": "CAT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 279, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 279, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 279, + 129 + ], + "type": "text", + "content": "3.2 Aggregating Key Clues" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "content": "An important component of MLLMs is the design of efficient bridging modules. The simple approach is to apply linear layers to incorporate the full visual and audio. The advantage is that it does not require any reduction of spatio-temporal information. However, such bridge approaches sometimes make LLM fail to reflect certain content and consistently generate a chunk of description about the video [28, 64]. Therefore, we devise a Clue Aggregator (CA) that enriches LLM-acquired knowledge by mining multimodal clues related to the question. CA is divided into two steps, as illustrated in Fig. 3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 236, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 482, + 247 + ], + "type": "text", + "content": "Step1: perceiving question-related visual and audio clues. To extract" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 133, + 272, + 307, + 381 + ], + "blocks": [ + { + "bbox": [ + 133, + 272, + 307, + 381 + ], + "lines": [ + { + "bbox": [ + 133, + 272, + 307, + 381 + ], + "spans": [ + { + "bbox": [ + 133, + 272, + 307, + 381 + ], + "type": "image", + "image_path": "70516e33dc025a4fca401f3ae0947dcf8b2b7b330391af6ef41f6e05abc82d0d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 391, + 306, + 413 + ], + "lines": [ + { + "bbox": [ + 131, + 391, + 306, + 413 + ], + "spans": [ + { + "bbox": [ + 131, + 391, + 306, + 413 + ], + "type": "text", + "content": "Fig. 3: Illustration of clue aggregator in a simple example." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 311, + 248, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 248, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 311, + 248, + 482, + 415 + ], + "type": "text", + "content": "more sufficient details from the input video and audio, we devise a clue perceiver, which consists of multiple different transformer layers. The perceiver can be viewed as a tiny transformer network that arranges the self-attention (SA), cross-attention (XA), and feed-forward network (FFN) in forward and reverse order. The forward order block " + }, + { + "bbox": [ + 311, + 248, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_1" + }, + { + "bbox": [ + 311, + 248, + 482, + 415 + ], + "type": "text", + "content": " is to perform attention-based question-aware localization. The reverse order block " + }, + { + "bbox": [ + 311, + 248, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_2" + }, + { + "bbox": [ + 311, + 248, + 482, + 415 + ], + "type": "text", + "content": " is to consolidate the original audio-visual representation at the attentional level." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 415, + 481, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 481, + 440 + ], + "type": "text", + "content": "Specifically, given tokens representing \"Is there a dog around?\" and the corresponding text embedding " + }, + { + "bbox": [ + 130, + 415, + 481, + 440 + ], + "type": "inline_equation", + "content": "h_t" + }, + { + "bbox": [ + 130, + 415, + 481, + 440 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 227, + 450, + 481, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 450, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 227, + 450, + 481, + 464 + ], + "type": "interline_equation", + "content": "\\mathcal {B} _ {1} \\left(h _ {t}; X\\right) = \\mathrm {X A} \\left(h _ {t}, \\operatorname {F F N} (\\operatorname {S A} (X))\\right), \\tag {2}", + "image_path": "6640df059cbda26caae1981973c663e273859999d02eb6601693f1c770651112.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 171, + 488, + 481, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 488, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 171, + 488, + 481, + 502 + ], + "type": "interline_equation", + "content": "\\mathcal {B} _ {2} (X; \\mathcal {B} _ {1} \\left(h _ {t}; X\\right)) = \\operatorname {S A} (\\text {F F N} (\\operatorname {X A} (\\operatorname {F F N} (\\operatorname {S A} (X)), \\mathcal {B} _ {1} \\left(h _ {t}, X\\right)))) \\tag {3}", + "image_path": "069c7c528e10700d85a3f61caa207e5d098d2a1d54436d205160bd3d946878fc.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 509, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 569 + ], + "type": "text", + "content": "In " + }, + { + "bbox": [ + 130, + 509, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\mathrm{XA}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 130, + 509, + 482, + 569 + ], + "type": "text", + "content": ", the former represents the query and the latter represents the key and value. Based on Eqs. 2 and 3, we use " + }, + { + "bbox": [ + 130, + 509, + 482, + 569 + ], + "type": "inline_equation", + "content": "h_v^\\prime = \\mathcal{B}_2(h_v;\\mathcal{B}_1(h_t;h_v))" + }, + { + "bbox": [ + 130, + 509, + 482, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 509, + 482, + 569 + ], + "type": "inline_equation", + "content": "h_a^\\prime = \\mathcal{B}_2(h_a;\\mathcal{B}_1(h_t;h_a))" + }, + { + "bbox": [ + 130, + 509, + 482, + 569 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 130, + 509, + 482, + 569 + ], + "type": "inline_equation", + "content": "h_v^\\prime" + }, + { + "bbox": [ + 130, + 509, + 482, + 569 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 509, + 482, + 569 + ], + "type": "inline_equation", + "content": "h_a^\\prime" + }, + { + "bbox": [ + 130, + 509, + 482, + 569 + ], + "type": "text", + "content": ", the question-aware visual and audio features, respectively. Notably, the perceiver for visual and the perceiver for audio implement shared parameters to learn potential associations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Step2: aggregating the perceived clues. A three-minute video undergoes a frozen encoder to get frame-level features of about 400 to 500 in length, leading to the fact that general machines cannot bear the burden of feeding all the frame-level features into the LLM. Thanks to the Q-former architecture proposed by Li et al. [27], which reduces the computational cost of end-to-end training of multimodal-language models. Specifically, we customize a learnable query vector " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " in length " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " to further extract useful information from the input question-aware features. Notably, we expand the " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "h_a'" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " in the time dimension to be consistent with" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Q. Ye et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "h_v'" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " and concatenate them. In practice, we set " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "K = 48" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " to aggregate all question-aware features and obtain the clue tokens " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "x^{cue}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " via a clues projector to align the dimension with the LLM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 172, + 312, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 172, + 312, + 185 + ], + "spans": [ + { + "bbox": [ + 131, + 172, + 312, + 185 + ], + "type": "text", + "content": "3.3 Multimodal Training Strategy" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 479, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 479, + 241 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 479, + 241 + ], + "type": "text", + "content": "We follow the previous works [35] to promote comprehension over video and audio. As shown in Fig. 2 on the right, we pre-training CAT on the video-level and audio-level tasks in stage-I. In stage-II, we fine-tune CAT on high-quality audio-visual-level instructions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 243, + 480, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 480, + 314 + ], + "type": "text", + "content": "Stage-I: feature alignment for audio and visual projectors. First, we employ the video-text Webvid 2.5M dataset [3] to train the visual projector (audio projector not involved in training). Second, we employ the audio-text WavCap dataset [42] to train the audio projector (visual projector not involved in training). During the above two training periods, we freeze the LLM and the encoder to align the semantic information for vision and audio, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 315, + 481, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 481, + 518 + ], + "type": "text", + "content": "Stage-II: audio-visual joint instruction tuning. To equip CAT with the ability to reason jointly based on visual and auditory, we collect an audio-visual joint instruction dataset, named AVinstruct, which emphasizes co-learning dynamic audio-visual pairs to solve diverse AVQA tasks. Specifically, we collect a large number of raw videos containing audio information from YouTube and VGGSound [5] as well as QA pairs from the training set in closed-ended AVQA tasks (i.e. music scene [26] and real-world scene [59]). Then different subtitle generators (i.e. BLIP2 [27] and Whisper [51]) are utilized to obtain video descriptions. Finally, we use GPT to generalize from human-written examples to synthesize question-guided audio-visual descriptions. We integrate all compositions after stage-I training and freeze the visual projector and audio projector, only fine-tuning clue aggregator and combined LoRA parameters on 100k video instruction [40] and AVinstruct. Notably, to highlight the question for applying in the clue aggregator, we reconstruct the form of the input prompt by adding two simple tokens " + }, + { + "bbox": [ + 130, + 315, + 481, + 518 + ], + "type": "inline_equation", + "content": "< Q >" + }, + { + "bbox": [ + 130, + 315, + 481, + 518 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 315, + 481, + 518 + ], + "type": "inline_equation", + "content": "< /Q>" + }, + { + "bbox": [ + 130, + 315, + 481, + 518 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 315, + 481, + 518 + ], + "type": "inline_equation", + "content": "< Q>" + }, + { + "bbox": [ + 130, + 315, + 481, + 518 + ], + "type": "text", + "content": " denotes the start position of the question and " + }, + { + "bbox": [ + 130, + 315, + 481, + 518 + ], + "type": "inline_equation", + "content": "< /Q>" + }, + { + "bbox": [ + 130, + 315, + 481, + 518 + ], + "type": "text", + "content": " denotes the end position of the question. During the instruction fine-tuning phase, we use predefined prompts based on the template:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 529, + 477, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 529, + 477, + 541 + ], + "spans": [ + { + "bbox": [ + 134, + 529, + 477, + 541 + ], + "type": "interline_equation", + "content": "\\text {U S E R}: < \\text {s y s t e m} > < Q > < / Q > < \\text {v i d e o} > < \\text {a u d i o} > < \\text {c l u e s} > \\quad \\text {A s s i s t a n t}:", + "image_path": "73405cc5252b971ed6de78547d2912b1292d698ed6832f5040def695cfa7263b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 552, + 479, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 479, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 479, + 588 + ], + "type": "text", + "content": "In this prompt, " + }, + { + "bbox": [ + 130, + 552, + 479, + 588 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 130, + 552, + 479, + 588 + ], + "type": "text", + "content": " system " + }, + { + "bbox": [ + 130, + 552, + 479, + 588 + ], + "type": "inline_equation", + "content": ">" + }, + { + "bbox": [ + 130, + 552, + 479, + 588 + ], + "type": "text", + "content": " denotes the official guidance message. " + }, + { + "bbox": [ + 130, + 552, + 479, + 588 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 130, + 552, + 479, + 588 + ], + "type": "text", + "content": " video " + }, + { + "bbox": [ + 130, + 552, + 479, + 588 + ], + "type": "inline_equation", + "content": ">" + }, + { + "bbox": [ + 130, + 552, + 479, + 588 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 552, + 479, + 588 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 130, + 552, + 479, + 588 + ], + "type": "text", + "content": " audio " + }, + { + "bbox": [ + 130, + 552, + 479, + 588 + ], + "type": "inline_equation", + "content": ">" + }, + { + "bbox": [ + 130, + 552, + 479, + 588 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 552, + 479, + 588 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 130, + 552, + 479, + 588 + ], + "type": "text", + "content": " clues " + }, + { + "bbox": [ + 130, + 552, + 479, + 588 + ], + "type": "inline_equation", + "content": ">" + }, + { + "bbox": [ + 130, + 552, + 479, + 588 + ], + "type": "text", + "content": " are associated with visual tokens, audio tokens, and question-related clue tokens, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 607, + 467, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 607, + 467, + 620 + ], + "spans": [ + { + "bbox": [ + 131, + 607, + 467, + 620 + ], + "type": "text", + "content": "3.4 AI-assisted Ambiguity-aware Direct Preference Optimization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "At first, we attempt to design various prompts to allow MLLMs to generate the most concise descriptions. However, such a no-learning approach is challenging to achieve the goal of accurate responses in a variety of dynamic audio-visual" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "CAT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 478, + 166 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 478, + 166 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 478, + 166 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 478, + 166 + ], + "type": "image", + "image_path": "70c78657b54bb5127b4417b49c3264437e266ec442a7a749c20b0a55bc01a48b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 174, + 482, + 231 + ], + "lines": [ + { + "bbox": [ + 130, + 174, + 482, + 231 + ], + "spans": [ + { + "bbox": [ + 130, + 174, + 482, + 231 + ], + "type": "text", + "content": "Fig. 4: Trained-CAT denotes CAT after feature alignment and instruction tuning. Our proposed ADPO strategy involves two steps. First, we collect the negative response generated by trained-CAT and correct it by GPT to obtain a positive response based on the original answer. Second, we perform ADPO training to skew CAT toward positive responses and reject negative responses." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 252, + 479, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 252, + 479, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 252, + 479, + 324 + ], + "type": "text", + "content": "scenarios. We find that MLLMs are easy to ambiguity when describing specific audio-visual objects. For example, an ambiguous response is illustrated by trained-CAT in Fig. 4, where trained-CAT denotes CAT after feature alignment and instruction tuning. ADPO is designed to allow MLLMs to review their mistakes and relearn with a new objective to better expression. Specifically, ADPO is divided into two steps to optimize CAT:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 324, + 480, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 480, + 467 + ], + "type": "text", + "content": "Collecting ambiguous samples and rewriting. As illustrated on the left in Fig. 4, assume that a given training set " + }, + { + "bbox": [ + 130, + 324, + 480, + 467 + ], + "type": "inline_equation", + "content": "D_{train}" + }, + { + "bbox": [ + 130, + 324, + 480, + 467 + ], + "type": "text", + "content": " and the corresponding trained-CAT output set " + }, + { + "bbox": [ + 130, + 324, + 480, + 467 + ], + "type": "inline_equation", + "content": "D_{des}" + }, + { + "bbox": [ + 130, + 324, + 480, + 467 + ], + "type": "text", + "content": ". We first provide GPT with a detailed prompt template to review each output in " + }, + { + "bbox": [ + 130, + 324, + 480, + 467 + ], + "type": "inline_equation", + "content": "D_{des}" + }, + { + "bbox": [ + 130, + 324, + 480, + 467 + ], + "type": "text", + "content": " for ambiguity. The rule of the review is to identify output ambiguities, i.e., responses that differ significantly from the original answer, which is referred to as a negative response. We let GPT correct this negative response, with the principle of generating a positive response that clearly describes the visual objects or sound without significant modification. Repeat the above steps to get the sample set " + }, + { + "bbox": [ + 130, + 324, + 480, + 467 + ], + "type": "inline_equation", + "content": "D_{dpo}" + }, + { + "bbox": [ + 130, + 324, + 480, + 467 + ], + "type": "text", + "content": " to be optimized, where " + }, + { + "bbox": [ + 130, + 324, + 480, + 467 + ], + "type": "inline_equation", + "content": "D_{dpo} = \\{x^{vid}, x^{aud}, x^{cue}, x^{txt}, y_{pos}, y_{neg}\\}" + }, + { + "bbox": [ + 130, + 324, + 480, + 467 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 324, + 480, + 467 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 324, + 480, + 467 + ], + "type": "text", + "content": " denotes the input modal token and " + }, + { + "bbox": [ + 130, + 324, + 480, + 467 + ], + "type": "inline_equation", + "content": "x^{txt}" + }, + { + "bbox": [ + 130, + 324, + 480, + 467 + ], + "type": "text", + "content": " denotes the text tokens, " + }, + { + "bbox": [ + 130, + 324, + 480, + 467 + ], + "type": "inline_equation", + "content": "y_{pos}" + }, + { + "bbox": [ + 130, + 324, + 480, + 467 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 324, + 480, + 467 + ], + "type": "inline_equation", + "content": "y_{neg}" + }, + { + "bbox": [ + 130, + 324, + 480, + 467 + ], + "type": "text", + "content": " denote the positive and negative responses, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "text", + "content": "Encouraging CAT to favor positive responses. As illustrated on the right in Fig. 4, ADPO directly optimizes the low-rank adaptation matrix parameters [17] in the policy model. We assume CAT as a policy model " + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "inline_equation", + "content": "f^{pol}" + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "text", + "content": ", given a reference model " + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "inline_equation", + "content": "f^{ref}" + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "text", + "content": ", a positive response " + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "inline_equation", + "content": "y_{pos}" + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "text", + "content": ", and a negative response " + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "inline_equation", + "content": "y_{neg}" + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "text", + "content": ". The reference model " + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "inline_equation", + "content": "f^{ref}" + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "text", + "content": " is a deep copy of " + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "inline_equation", + "content": "f^{pol}" + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "text", + "content": " and during training we encourage " + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "inline_equation", + "content": "f^{pol}" + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "text", + "content": " to favor " + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "inline_equation", + "content": "y_{pos}" + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "inline_equation", + "content": "f^{ref}" + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "text", + "content": " does not update the weights. Specifically, we define DPO loss " + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{DPO}" + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 559, + 481, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 559, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 159, + 559, + 481, + 634 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {D P O} \\left(f ^ {p o l}; f ^ {r e f}\\right) = - E _ {\\left(x ^ {v i d}, x ^ {a u d}, x ^ {c u e}, x ^ {t x t}, y _ {p o s}, y _ {n e g}\\right)} \\sim D _ {d p o} \\\\ \\left[ \\right. \\log \\sigma \\left(\\beta \\log \\frac {f ^ {p o l} \\left(y _ {p o s} \\mid [ x ^ {v i d}: x ^ {a u d}: x ^ {c u e}: x ^ {t x t} ]\\right)}{f ^ {r e f} \\left(y _ {p o s} \\mid [ x ^ {v i d}: x ^ {a u d}: x ^ {c u e}: x ^ {t x t} ]\\right)} \\right. \\quad (4) \\\\ \\left. - \\beta l o g \\frac {f ^ {p o l} \\left(y _ {n e g} \\mid \\left[ x ^ {v i d}: x ^ {a u d}: x ^ {c u e}: x ^ {t x t} \\right]\\right)}{f ^ {r e f} \\left(y _ {n e g} \\mid \\left[ x ^ {v i d}: x ^ {a u d}: x ^ {c u e}: x ^ {t x t} \\right]\\right)}\\right) \\Bigg ], \\\\ \\end{array}", + "image_path": "9725b1d6ef1d69d22c14cafc6a1c6bff452c26baed41bf6ae73f243749acdd81.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "text", + "content": "[ ] denotes the concatenation. " + }, + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "text", + "content": " is the non-linear function, Sigmoid, and " + }, + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "text", + "content": " is a hyperparameter. This objective function aims to directly optimize the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 102 + ], + "type": "text", + "content": "Q. Ye et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 150 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 150 + ], + "type": "text", + "content": "model that favors the positive response " + }, + { + "bbox": [ + 131, + 116, + 479, + 150 + ], + "type": "inline_equation", + "content": "y_{pos}" + }, + { + "bbox": [ + 131, + 116, + 479, + 150 + ], + "type": "text", + "content": " and rejects the negative response " + }, + { + "bbox": [ + 131, + 116, + 479, + 150 + ], + "type": "inline_equation", + "content": "y_{neg}" + }, + { + "bbox": [ + 131, + 116, + 479, + 150 + ], + "type": "text", + "content": ". Assuming that " + }, + { + "bbox": [ + 131, + 116, + 479, + 150 + ], + "type": "inline_equation", + "content": "\\beta = 1" + }, + { + "bbox": [ + 131, + 116, + 479, + 150 + ], + "type": "text", + "content": ", we simplify the output " + }, + { + "bbox": [ + 131, + 116, + 479, + 150 + ], + "type": "inline_equation", + "content": "\\hat{r}" + }, + { + "bbox": [ + 131, + 116, + 479, + 150 + ], + "type": "text", + "content": " of the model based on Eq. 4 as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 167, + 155, + 479, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 155, + 479, + 182 + ], + "spans": [ + { + "bbox": [ + 167, + 155, + 479, + 182 + ], + "type": "interline_equation", + "content": "\\hat {r} \\left(x ^ {v i d}, x ^ {a u d}, x ^ {c u e}, x ^ {t x t}, y\\right) = \\log \\frac {f ^ {p o l} (y \\mid \\left[ x ^ {v i d}: x ^ {a u d}: x ^ {c u e}: x ^ {t x t} \\right])}{f ^ {r e f} (y \\mid \\left[ x ^ {v i d}: x ^ {a u d}: x ^ {c u e}: x ^ {t x t} \\right])}. \\tag {5}", + "image_path": "236dae71e70003a2dbe5295bfd64838a29b952575731b288e4616c28601587dc.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 186, + 479, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 186, + 479, + 270 + ], + "spans": [ + { + "bbox": [ + 131, + 186, + 479, + 270 + ], + "type": "text", + "content": "The optimal goal is to make the variance of " + }, + { + "bbox": [ + 131, + 186, + 479, + 270 + ], + "type": "inline_equation", + "content": "\\hat{r}(x^{vid}, x^{aud}, x^{cue}, x^{txt}, y_{pos}) - \\hat{r}(x^{vid}, x^{aud}, x^{cue}, x^{txt}, y_{neg})" + }, + { + "bbox": [ + 131, + 186, + 479, + 270 + ], + "type": "text", + "content": " larger and larger as a way to steer the model to generate descriptions devoid of ambiguous. However, when the difference between positive and negative responses is small, DPO loss supervision alone is not an effective facilitator. Therefore, we introduce an additional objective " + }, + { + "bbox": [ + 131, + 186, + 479, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SFT}" + }, + { + "bbox": [ + 131, + 186, + 479, + 270 + ], + "type": "text", + "content": ", which is similar to the process of supervised fine-tuning, to supervise the model to stabilize the bias towards the positive response. Specifically, we define " + }, + { + "bbox": [ + 131, + 186, + 479, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SFT}" + }, + { + "bbox": [ + 131, + 186, + 479, + 270 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 175, + 276, + 480, + 309 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 276, + 480, + 309 + ], + "spans": [ + { + "bbox": [ + 175, + 276, + 480, + 309 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {S F T} = - \\sum \\log P \\left(y _ {\\text {p o s}} \\mid \\left[ x ^ {\\text {v i d e o}}: x ^ {\\text {a u d i o}}: x ^ {\\text {c l u e s}}: x ^ {\\text {t x t}} \\right]; f ^ {\\text {p o l}}\\right), \\tag {6} \\\\ \\left\\{x ^ {v i d}, x ^ {a u d}, x ^ {c u e}, x ^ {t x t}, y _ {p o s} \\right\\} \\sim D _ {d p o}, \\\\ \\end{array}", + "image_path": "cb2a5f3519be18cb82b16b56795ada175a04d599b2c9803a2789a2f6ee980f6a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 313, + 479, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 313, + 479, + 350 + ], + "spans": [ + { + "bbox": [ + 131, + 313, + 479, + 350 + ], + "type": "text", + "content": "this loss function can be interpreted as using the positive output probability in the policy model to aid in optimization. During ADPO training, we sum up " + }, + { + "bbox": [ + 131, + 313, + 479, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{DPO}" + }, + { + "bbox": [ + 131, + 313, + 479, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 313, + 479, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SFT}" + }, + { + "bbox": [ + 131, + 313, + 479, + 350 + ], + "type": "text", + "content": " with a bias " + }, + { + "bbox": [ + 131, + 313, + 479, + 350 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 131, + 313, + 479, + 350 + ], + "type": "text", + "content": " to achieve direct preference optimization:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 261, + 357, + 479, + 367 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 357, + 479, + 367 + ], + "spans": [ + { + "bbox": [ + 261, + 357, + 479, + 367 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {D P O} + \\lambda \\mathcal {L} _ {S F T}, \\tag {7}", + "image_path": "f58f0638715b14396ecd3e70f9d0aefc8112a5c5dd7bfef55f10932248d956fa.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 373, + 275, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 373, + 275, + 384 + ], + "spans": [ + { + "bbox": [ + 131, + 373, + 275, + 384 + ], + "type": "text", + "content": "where we set " + }, + { + "bbox": [ + 131, + 373, + 275, + 384 + ], + "type": "inline_equation", + "content": "\\lambda = 0.1" + }, + { + "bbox": [ + 131, + 373, + 275, + 384 + ], + "type": "text", + "content": " by default." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 399, + 229, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 399, + 229, + 414 + ], + "spans": [ + { + "bbox": [ + 131, + 399, + 229, + 414 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 422, + 204, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 422, + 204, + 433 + ], + "spans": [ + { + "bbox": [ + 131, + 422, + 204, + 433 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 437, + 479, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 437, + 479, + 497 + ], + "spans": [ + { + "bbox": [ + 131, + 437, + 479, + 497 + ], + "type": "text", + "content": "Video-based text generation tasks. We evaluate the proposed CAT on the video understanding benchmarks proposed by Video-ChatGPT [40]. Specifically, Video-ChatGPT proposes five metrics: Correctness of Information, Consistency, Detail Orientation, Contextual Understanding, and Temporal Understanding, which test the ability of MLLMs to describe videos." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 498, + 479, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 479, + 557 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 479, + 557 + ], + "type": "text", + "content": "Zero-shot on video question answering tasks. To evaluate whether CAT has the basic ability to communicate regarding video, we conduct zero-shot tests on MSRVTT-QA [58] and ActivityNet-QA [60]. MSRVTT-QA and ActivityNet-QA consist of 10k and 5.8k videos containing audio information, respectively, where the QA pairs are mostly questions about daily life." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 558, + 479, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 558, + 479, + 604 + ], + "spans": [ + { + "bbox": [ + 131, + 558, + 479, + 604 + ], + "type": "text", + "content": "Closed-ended AVQA tasks. We categorize the Music-AVQA [26] and AVQA [59] datasets as closed-ended AVQA tasks. These datasets consist of up to 42 candidate answers that require the selection of the most appropriate one based on visual and auditory content." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 605, + 479, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 479, + 664 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 479, + 664 + ], + "type": "text", + "content": "Open-ended AVQA tasks. We select audio-visual dialogue (AVSD [1]), and audio-visual captioning (VALOR [6]) tasks for the evaluation of open-ended AVQA. These tasks require precise language to interpret, correlate, and reason about cross-modal information. We evaluate the zero-shot ability of CAT on these datasets." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "CAT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 168, + 482, + 299 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: GPT-based evaluation [40] for video-based text generation and zero-shot video question answering. For clarity, five scores are reported (\"Cr.\": Correctness of Information, \"Cs.\": Consistency, \"De.\": Detail Orientation, \"Ct\": Contextual Understanding, \"Te.\": Temporal Understanding). Score indicates the confidence." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 168, + 482, + 299 + ], + "lines": [ + { + "bbox": [ + 133, + 168, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 133, + 168, + 482, + 299 + ], + "type": "table", + "html": "
MethodLLM SizeVideo-based text generationZero-shot on MSRVTT-QA Acc.ScoreZero-shot on ActivityNet-QA Acc.Score
Cr.De.Ct.Te.Cs.
LLaMA-VID [31]13B61.461.072.051.652.658.93.347.53.3
Video-LLaMA [64]7B39.243.643.236.435.829.61.812.41.1
LLaMA-Adapter [65]7B40.646.446.039.643.043.82.734.22.7
VideoChat [28]7B44.650.050.638.844.845.02.526.52.2
Video-ChatGPT [40]7B48.050.452.439.647.449.32.835.22.7
VISTA-LLaMA [39]7B48.852.863.645.246.260.53.348.33.3
VideoChat2 [29]7B60.457.870.253.256.254.13.349.13.3
Chat-UniVi [20]7B57.858.269.257.856.254.63.145.83.2
LLaMA-VID [31]7B59.260.070.649.250.257.73.247.13.3
CAT (Ours)7B61.659.069.856.257.862.13.550.23.5
", + "image_path": "e3ac27d5c6226cb3221f4e6b393d13fe7b3432431d64dd658fa03e132877c83e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 320, + 261, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 320, + 261, + 333 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 261, + 333 + ], + "type": "text", + "content": "4.2 Experimental Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 340, + 482, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 435 + ], + "type": "text", + "content": "Evaluation metrics. For video-based text generation and zero-shot on video question-answering tasks, we follow the evaluation pipeline proposed by VideoChatGPT [40], which is based on GPT-3.5 to evaluate predictive descriptions against correct descriptions, with a score from 0 to 5 indicating accuracy. In this paper, we standardize 0 to 5 as 0 to 100 to align common accuracy rubrics. For closed-ended AVQA tasks, we report the accuracy of the correct sample. For open-ended AVQA tasks, we report the CIDEr [56] that specializes in evaluating visually descriptive tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 436, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 484 + ], + "type": "text", + "content": "Architecture. We use frozen ImageBind [13] and LLaMA2-7B [55] as audiovisual encoders and LLM, respectively. The size of modality embeddings for each modality are " + }, + { + "bbox": [ + 130, + 436, + 482, + 484 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{T\\times 1024}" + }, + { + "bbox": [ + 130, + 436, + 482, + 484 + ], + "type": "text", + "content": ". The outputs " + }, + { + "bbox": [ + 130, + 436, + 482, + 484 + ], + "type": "inline_equation", + "content": "x^{vid}" + }, + { + "bbox": [ + 130, + 436, + 482, + 484 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 436, + 482, + 484 + ], + "type": "inline_equation", + "content": "x^{aud}" + }, + { + "bbox": [ + 130, + 436, + 482, + 484 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 436, + 482, + 484 + ], + "type": "inline_equation", + "content": "x^{cue}" + }, + { + "bbox": [ + 130, + 436, + 482, + 484 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 130, + 436, + 482, + 484 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{1\\times 4096}" + }, + { + "bbox": [ + 130, + 436, + 482, + 484 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 436, + 482, + 484 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{1\\times 4096}" + }, + { + "bbox": [ + 130, + 436, + 482, + 484 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 436, + 482, + 484 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{48\\times 4096}" + }, + { + "bbox": [ + 130, + 436, + 482, + 484 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 484, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 604 + ], + "type": "text", + "content": "Training details. We complete feature alignment training, instruction tuning, and ADPO training with 1 NVIDIA A100 GPU. In detail, For the feature alignment training and instruction tuning, we use the AdamW optimizer with a cosine learning rate decay and a warm-up period. When LoRA is added, we set " + }, + { + "bbox": [ + 130, + 484, + 482, + 604 + ], + "type": "inline_equation", + "content": "r = 64" + }, + { + "bbox": [ + 130, + 484, + 482, + 604 + ], + "type": "text", + "content": " and alpha " + }, + { + "bbox": [ + 130, + 484, + 482, + 604 + ], + "type": "inline_equation", + "content": "= 128" + }, + { + "bbox": [ + 130, + 484, + 482, + 604 + ], + "type": "text", + "content": " for the LoRA parameters, and the total batch size is set to 128 for training 1 epoch with a learning rate of " + }, + { + "bbox": [ + 130, + 484, + 482, + 604 + ], + "type": "inline_equation", + "content": "2e^{-5}" + }, + { + "bbox": [ + 130, + 484, + 482, + 604 + ], + "type": "text", + "content": ". For ADPO training, we only select the training set in the instruction tuning phase for optimization. We set " + }, + { + "bbox": [ + 130, + 484, + 482, + 604 + ], + "type": "inline_equation", + "content": "r = 64" + }, + { + "bbox": [ + 130, + 484, + 482, + 604 + ], + "type": "text", + "content": " and alpha " + }, + { + "bbox": [ + 130, + 484, + 482, + 604 + ], + "type": "inline_equation", + "content": "= 16" + }, + { + "bbox": [ + 130, + 484, + 482, + 604 + ], + "type": "text", + "content": " for the LoRA parameters, and the total batch size is set to 1 for training 1 epoch with a learning rate of " + }, + { + "bbox": [ + 130, + 484, + 482, + 604 + ], + "type": "inline_equation", + "content": "4e^{-6}" + }, + { + "bbox": [ + 130, + 484, + 482, + 604 + ], + "type": "text", + "content": ". The hyperparameter " + }, + { + "bbox": [ + 130, + 484, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 484, + 482, + 604 + ], + "type": "text", + "content": " is set to 0.1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 621, + 320, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 320, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 320, + 634 + ], + "type": "text", + "content": "4.3 Comparison to State-of-the-Art" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Comparison on video-based text generation tasks. We follow the benchmark proposed by Video-ChatGPT [40] to evaluate CAT. On the left of Table 1," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Q. Ye et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 162, + 146, + 451, + 288 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 136 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 136 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 136 + ], + "type": "text", + "content": "Table 2: Comparison with non-LLMs-based methods on fine-tuned Music-AVQA dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 162, + 146, + 451, + 288 + ], + "lines": [ + { + "bbox": [ + 162, + 146, + 451, + 288 + ], + "spans": [ + { + "bbox": [ + 162, + 146, + 451, + 288 + ], + "type": "table", + "html": "
MethodLanguage ModelTrainable Params (M)Audio avg.Visual avg.Audio-Visual avg.Overall avg.
FCNLSTM [11]CLIP [50]9.7968.956.260.460.8
GRU [2]CLIP-68.367.063.065.0
HCAattn [36]CLIP-64.965.360.362.5
MCAN [61]CLIP56.070.671.861.565.8
PSAC [30]CLIP-72.069.463.666.6
HME [10]CLIP-69.968.864.866.7
HCRN [24]CLIP-63.765.249.856.3
AVSD [54]CLIP8.3568.870.365.467.3
Panp-AVQA [62]CLIP-72.173.267.069.5
AVST [26]CLIP18.4873.974.469.571.6
PSTP-Net [25]BERT [9]4.29770.977.372.673.5
LAVISH [34]CLIP21.0977.177.377.0-
CAD [43]GLoVE [48]-78.179.776.978.2
VALOR [6]BERT----78.9
VAST [7]BERT----80.7
CAT-7B (Ours)LLaMA2 [55]5.81384.986.183.284.3
", + "image_path": "a1b2f5a580d113b7da80f10be4bf23d024d8a2a9cae5d7f39705adb99e1f92d0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 314, + 479, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 479, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 479, + 373 + ], + "type": "text", + "content": "we show that CAT achieves the state-of-the-art in terms of correctness of description information (Cr.), detailed description of the problem (De.), and coherence of the description (Cs.). In addition, CAT does not parse longer visual tokens, it still achieves competitive results when comparing time-related descriptions (e.g., contextual and temporal understanding)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 375, + 480, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 480, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 480, + 448 + ], + "type": "text", + "content": "Comparison on zero-shot video question answering tasks. On the right of Table 1, we show the zero-shot video question answering performance of CAT on several open-ended datasets. While recent MLLMs designed with bridging modules have produced substantial results, CAT is way ahead of them in recognizing accurate answers. We consistently outperform the state-of-the-art on the MSRVTT-QA [58] and ActivityNet-QA [60] benchmarks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 481, + 616 + ], + "type": "text", + "content": "Comparison on closed-ended AVQA tasks. We choose to evaluate Music-AVQA [26] to demonstrate that CAT can perceive specific audio-visual objects to answer questions. In Table 2, under full supervision of the training set, CAT accurately retrieves specific objects in dynamic audio-visual scenarios and comprehensively outperforms all non-LLMs-based models. Thanks to LoRA [17], we can improve the evaluation quality with the help of the world knowledge inside LLaMA2 [55] and only 5.8M parameters are trainable. Also, we examine the ability of CAT for zero-shot on Music-AVQA [26]. In Table 3, we show the results comparing LLMs-based models. Notably, for a fair comparison, we remove the LoRA parameters of CAT that are fine-tuned on AVinstruct, which is derived from the Music-AVQA [26] and AVQA [59] training sets. Even though our model does not draw on the larger-scale LLM, it still achieves a small advantage over ChatBridge [67] with a 13B LLM size. Furthermore, we show the performance of CAT in multiple-choice scenarios [59] in Table 4, CAT continues outstanding." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Comparison on open-ended AVQA tasks. Open-ended AVQA tasks require responses to daily events based on audio-visual content. We conduct a comparative analysis of our model with multimodal-based LLMs: OneLLM [14], and ChatBridge [67]. The tests are divided into zero-shot and full supervision, and" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 426, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 426, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 426, + 91, + 447, + 100 + ], + "type": "text", + "content": "CAT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "we only choose to evaluate the zero-shot to demonstrate the level of practical application of the CAT. In Table 5, CAT surpasses OneLLM and ChatBridge on zero-shot complex text reasoning tasks AVSD [1] and VALOR [6]. Moreover, our approach is close to the results of FA [44] using fully supervised training, further demonstrating the strong multimodal understanding of CAT." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 132, + 236, + 485, + 355 + ], + "blocks": [ + { + "bbox": [ + 132, + 197, + 483, + 226 + ], + "lines": [ + { + "bbox": [ + 132, + 197, + 483, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 197, + 483, + 226 + ], + "type": "text", + "content": "Table 3: Comparison with LLMs-based methods on zero-shot Music- AVQA dataset. Table 5: Evaluation on zero-shot open-ended" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 236, + 485, + 355 + ], + "lines": [ + { + "bbox": [ + 132, + 236, + 485, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 485, + 355 + ], + "type": "table", + "html": "
Methodzero-shotAcc.
OneLLM-7B [14]✓43.0
ChatBridge-13B [67]✓47.6
CAT-7B (Ours)✓48.6
Table 4: Evaluation on fine-tuned AVQA dataset.
MethodAcc.
HGA + HAVF [19]87.7
HCRN + HAVF [59]89.0
PSTP-Net [25]90.2
CAT-7B (Ours)92.0
OneLLM-7B [14]✓74.5
ChatBridge-13B [67]✓75.4
CAT-7B (Ours)✓79.0
", + "image_path": "db01d7819d9e9b14bdd22e00b42bf4057da977bff691df95979e5bfce87fd29f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 395, + 279, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 395, + 279, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 395, + 279, + 407 + ], + "type": "text", + "content": "4.4 Ablations and Analyses" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 414, + 481, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 481, + 474 + ], + "type": "text", + "content": "In this subsection, we explore the effects of each component of CAT. We test the effect of input modal tokens on video-based text generation [40]. In addition, to avoid the huge expense associated with the GPT3.5 evaluation, we test each component of the clue aggregator on the test set of AVinstruct, as well as the generalizability of ADPO on AVSD [1]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 481, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 594 + ], + "type": "text", + "content": "Does question-related clues matter? Visual tokens and audio tokens already have the ability to distinguish video content. To explore whether the added question-related clues are meaningful, we conduct ablation experiments on the video-based text generation task in Table 6. We first test visual knowledge separately from question-related clues knowledge on video comprehension. Observations reveal that the results achieved when inputting question-related clues alone have been able to outperform the joint input of audio-visual knowledge. Further, we add clues knowledge to the original audio-visual knowledge, and the results demonstrate that the fusion of the three can fully satisfy the information required for LLM reasoning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "Ablation on clue aggregator. We categorize a portion of the collected audiovisual joint instruction data into a test set and present the ablation results in Table 7. The evaluation indicators B, M, and C denote BLEU-4 [47], METEOR [22], and ROUGE-L, respectively. We split " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_1" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_2" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " and query subjects " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "h_v" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "h_a" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " to explore the impact of CA internals on reasoning about dynamic audiovisual pairs. We find that visual attention influences context much more than" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Q. Ye et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 159, + 248, + 216 + ], + "blocks": [ + { + "bbox": [ + 132, + 115, + 470, + 157 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 470, + 157 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 470, + 157 + ], + "type": "text", + "content": "Table 6: Ablation experiment Table 7: Ablation exper- Table 8: Ablation exper- about input modals on video- iments about clue aggre- iment about ADPO on based text generation. gator on audio-visual joint AVSD [1]. instruction dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 159, + 248, + 216 + ], + "lines": [ + { + "bbox": [ + 138, + 159, + 248, + 216 + ], + "spans": [ + { + "bbox": [ + 138, + 159, + 248, + 216 + ], + "type": "table", + "html": "
Input-modalCr.De.Te.
\\( x^{vid} \\)\\( x^{aud} \\)\\( x^{cue} \\)
✓38.640.434.0
✓✓40.644.835.2
✓58.657.455.8
✓✓✓61.659.056.2
", + "image_path": "220c0aad02a8222100a89d258f71f76f054e1ea83f7269b2776a5dbe2165d725.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 260, + 168, + 361, + 217 + ], + "blocks": [ + { + "bbox": [ + 260, + 168, + 361, + 217 + ], + "lines": [ + { + "bbox": [ + 260, + 168, + 361, + 217 + ], + "spans": [ + { + "bbox": [ + 260, + 168, + 361, + 217 + ], + "type": "table", + "html": "
MethodB@4MR
B1 + B230.858.768.5
B1 + B2 w/o hv10.037.149.9
B1 + B2 w/o ha18.849.361.8
only B125.954.764.2
", + "image_path": "56aa1c4f1a2f6e940264da8c6c0d19fdbf83acbc0227ec96b9b65542256e9c0b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 367, + 159, + 485, + 215 + ], + "blocks": [ + { + "bbox": [ + 367, + 159, + 485, + 215 + ], + "lines": [ + { + "bbox": [ + 367, + 159, + 485, + 215 + ], + "spans": [ + { + "bbox": [ + 367, + 159, + 485, + 215 + ], + "type": "table", + "html": "
ModelADPOAVSD
B@4MC
Video-LLaMA×18.440.163.7
Video-LLaMA✓22.248.269.2
CAT (Ours)×28.956.274.8
CAT (Ours)✓34.259.879.0
", + "image_path": "e1c200741b48959cea744facf39911d952507115ca48f0cef59f360b09310b27.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 141, + 230, + 290, + 321 + ], + "blocks": [ + { + "bbox": [ + 141, + 230, + 290, + 321 + ], + "lines": [ + { + "bbox": [ + 141, + 230, + 290, + 321 + ], + "spans": [ + { + "bbox": [ + 141, + 230, + 290, + 321 + ], + "type": "image", + "image_path": "bf80cdb6e84cc68f7e9a9ca3719a766e1ee21e457d7275b45cd8de341e8c2b21.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 139, + 323, + 315, + 341 + ], + "lines": [ + { + "bbox": [ + 139, + 323, + 315, + 341 + ], + "spans": [ + { + "bbox": [ + 139, + 323, + 315, + 341 + ], + "type": "text", + "content": "(a) Different number of visual tokens on video-based text generation." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 230, + 468, + 322 + ], + "blocks": [ + { + "bbox": [ + 316, + 230, + 468, + 322 + ], + "lines": [ + { + "bbox": [ + 316, + 230, + 468, + 322 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 468, + 322 + ], + "type": "image", + "image_path": "b48cafed2be87849d7aed939d4e897da04a302f8358d1a5abbeb552b4ee1f9bf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 323, + 473, + 341 + ], + "lines": [ + { + "bbox": [ + 315, + 323, + 473, + 341 + ], + "spans": [ + { + "bbox": [ + 315, + 323, + 473, + 341 + ], + "type": "text", + "content": "(b) Different number of clue tokens on video-based text generation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 350, + 482, + 373 + ], + "lines": [ + { + "bbox": [ + 132, + 350, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 350, + 482, + 373 + ], + "type": "text", + "content": "Fig. 5: The impacts of input modal tokens. Avg. represents the average accuracy of temporal (Te.), consistency (CS.), and detail (De.)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 399, + 480, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 480, + 459 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 480, + 459 + ], + "type": "text", + "content": "auditory. This is because vision has more information that can be mined and is more relevant to the question. Next, we reduce parameters to investigate whether the same performance can be achieved with a single block. However, since the question acts as a query, without " + }, + { + "bbox": [ + 130, + 399, + 480, + 459 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_2" + }, + { + "bbox": [ + 130, + 399, + 480, + 459 + ], + "type": "text", + "content": " to recover the original modal length does not provide an advantage over two blocks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 460, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 581 + ], + "type": "text", + "content": "Number of input modal tokens. We evaluate the effect of the number of input modal tokens on video understanding [40]. As shown in Fig. 5, we first study the effect of the number of visual tokens on a time-related description task. Indeed, as the number of visual tokens increases, the accuracy in the evaluation of the various descriptions decreases, even if the change is not large, but it also confirms that the language model may not be able to reason based on longer visual tokens. Next, we examine the impact of the number of clue tokens set by the Q-former [27]. Clue aggregator uses Q-Former to transform frame-level features into specific tokens, and we find that it works best when the length of queries " + }, + { + "bbox": [ + 130, + 460, + 482, + 581 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 460, + 482, + 581 + ], + "type": "text", + "content": " is set to 48." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Impact of ADPO. As the paper explains, supervised fine-tuning alone still underperforms in audio-visual scenarios. We test the generalizability of ADPO on Video-LLaMA [64] and ADPO's effect on AVSD [1] in Table 8 to demonstrate the superiority of this learning strategy. C denotes the CiDER [56]. We find that for different MLLMs, ADPO does improve the descriptive power of the model without much learning cost, it brings different levels of enhancement to both Video-LLaMA and CAT." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 425, + 91, + 447, + 100 + ], + "type": "text", + "content": "CAT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 158, + 114, + 309, + 190 + ], + "blocks": [ + { + "bbox": [ + 158, + 114, + 309, + 190 + ], + "lines": [ + { + "bbox": [ + 158, + 114, + 309, + 190 + ], + "spans": [ + { + "bbox": [ + 158, + 114, + 309, + 190 + ], + "type": "image", + "image_path": "c3c690717c48db9cc5fabcca04f16195d68ec28dcbd5ccd581d2d68393f5ba02.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 312, + 114, + 457, + 190 + ], + "blocks": [ + { + "bbox": [ + 312, + 114, + 457, + 190 + ], + "lines": [ + { + "bbox": [ + 312, + 114, + 457, + 190 + ], + "spans": [ + { + "bbox": [ + 312, + 114, + 457, + 190 + ], + "type": "image", + "image_path": "e902a88381942ace40ae07de6ab88f2632e85273e0021b92a1e5e74b4d2a1ae4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 158, + 191, + 309, + 266 + ], + "blocks": [ + { + "bbox": [ + 158, + 191, + 309, + 266 + ], + "lines": [ + { + "bbox": [ + 158, + 191, + 309, + 266 + ], + "spans": [ + { + "bbox": [ + 158, + 191, + 309, + 266 + ], + "type": "image", + "image_path": "fab1f9f6e3f4f0cf1ad743983e8b0f19ce9a4b2ef089a32befdfca93a45a52f2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 274, + 443, + 286 + ], + "lines": [ + { + "bbox": [ + 170, + 274, + 443, + 286 + ], + "spans": [ + { + "bbox": [ + 170, + 274, + 443, + 286 + ], + "type": "text", + "content": "Fig. 6: Qualitative results of different types of question answering." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 312, + 191, + 457, + 266 + ], + "blocks": [ + { + "bbox": [ + 312, + 191, + 457, + 266 + ], + "lines": [ + { + "bbox": [ + 312, + 191, + 457, + 266 + ], + "spans": [ + { + "bbox": [ + 312, + 191, + 457, + 266 + ], + "type": "image", + "image_path": "1ca152d7e8057b5c6a16dc48545f02df74507b290251bff25a33d8a9fb4ad6ab.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 308, + 263, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 263, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 263, + 319 + ], + "type": "text", + "content": "4.5 Qualitative Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 326, + 482, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 482, + 458 + ], + "type": "text", + "content": "In Fig. 6, we analyze the qualitative results with Video-LLaMA [64]. In the example of video-specific contextual inference, we show that CAT has a sharp perception of complex indoor scenes. Our descriptions can accurately represent what the person doing and what background sounds are in the video. Video-LLaMA is biased towards describing scene information and incorrect character information. In the specific audio question answering example, Video-LLaMA lost the ability to answer \"How many instruments are sounding in the video?\" In contrast, CAT can accurately answer the quantity question. In the audio-visual question answering example, Video-LLaMA answers incorrectly due to failed audio-visual grounding, while our CAT answers correctly due to its strong ability to capture specific objects in audio-visual scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 474, + 220, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 474, + 220, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 474, + 220, + 487 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": "In this work, we introduce CAT to enhance LLMs' multimodal understanding in dynamic audio-visual scenarios. We propose a clue aggregator to aggregate the question-related clues for detailed reasoning. We mix datasets containing audio and video to empower LLM with multimodal understanding. To more consistently infer audio-visual scenarios, we collect an audio-visual joint instruction dataset to further fine-tune the CAT. Moreover, we propose an AI-assisted ambiguity-aware direct preference optimization strategy to retrain the model for more accurate responses. CAT has achieved comparable results when applied to a variety of complex audio-visual scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by the National Natural Science Foundation of China under Grant 62306061, Guangdong Basic and Applied Basic Research Foundation (Grant No. 2023A1515140037), and the Open Fund of National Engineering Laboratory for Big Data System Computing Technology (Grant No. SZU-BDSC-OF2024-02)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Q. Ye et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 139, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 139, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 481, + 183 + ], + "type": "text", + "content": "1. AlAmri, H., Cartillier, V., Das, A., Wang, J., Cherian, A., Essa, I., Batra, D., Marks, T.K., Hori, C., Anderson, P., Lee, S., Parikh, D.: Audio visual scene-aware dialog. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR. pp. 7558-7567 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 184, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 481, + 216 + ], + "type": "text", + "content": "2. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: VQA: visual question answering. In: IEEE International Conference on Computer Vision, ICCV. pp. 2425-2433. IEEE Computer Society (2015)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 217, + 481, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 481, + 249 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 481, + 249 + ], + "type": "text", + "content": "3. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: IEEE/CVF International Conference on Computer Vision, ICCV. pp. 1708-1718. IEEE (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 250, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 250, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 250, + 481, + 281 + ], + "type": "text", + "content": "4. Chen, G., Shen, L., Shao, R., Deng, X., Nie, L.: LION: Empowering multimodal large language model with dual-level visual knowledge. CoRR abs/2311.11860 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 282, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 282, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 282, + 481, + 315 + ], + "type": "text", + "content": "5. Chen, H., Xie, W., Vedaldi, A., Zisserman, A.: Vggsound: A large-scale audiovisual dataset. In: IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP. pp. 721-725. IEEE (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 316, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 481, + 347 + ], + "type": "text", + "content": "6. Chen, S., He, X., Guo, L., Zhu, X., Wang, W., Tang, J., Liu, J.: VALOR: vision-audio-language omni-perception pretraining model and dataset. CoRR abs/2304.08345 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 348, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 481, + 380 + ], + "type": "text", + "content": "7. Chen, S., Li, H., Wang, Q., Zhao, Z., Sun, M., Zhu, X., Liu, J.: VAST: A vision-audio-subtitle-text omni-modality foundation model and dataset. CoRR abs/2305.18500 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 381, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 481, + 413 + ], + "type": "text", + "content": "8. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.C.H.: Instructclip: Towards general-purpose vision-language models with instruction tuning. CoRR abs/2305.06500 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 414, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 481, + 468 + ], + "type": "text", + "content": "9. Devlin, J., Chang, M., Lee, K., Toutanova, K.: BERT: pre-training of deep bidirectional transformers for language understanding. In: Burstein, J., Doran, C., Solorio, T. (eds.) Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT. pp. 4171-4186. Association for Computational Linguistics (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 468, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 468, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 134, + 468, + 481, + 512 + ], + "type": "text", + "content": "10. Fan, C., Zhang, X., Zhang, S., Wang, W., Zhang, C., Huang, H.: Heterogeneous memory enhanced multimodal attention model for video question answering. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR. pp. 1999-2007. Computer Vision Foundation / IEEE (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 513, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 513, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 134, + 513, + 481, + 533 + ], + "type": "text", + "content": "11. Fayek, H.M., Johnson, J.: Temporal reasoning via audio question answering. IEEE ACM Trans. Audio Speech Lang. Process. 28, 2283-2294 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 134, + 534, + 481, + 566 + ], + "type": "text", + "content": "12. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., Li, H., Qiao, Y.: Llama-adapter V2: parameter-efficient visual instruction model. CoRR abs/2304.15010 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 134, + 567, + 481, + 599 + ], + "type": "text", + "content": "13. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind one embedding space to bind them all. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR. pp. 15180-15190 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "type": "text", + "content": "14. Han, J., Gong, K., Zhang, Y., Wang, J., Zhang, K., Lin, D., Qiao, Y., Gao, P., Yue, X.: Onellm: One framework to align all modalities with language. CoRR abs/2312.03700 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "text", + "content": "15. Han, J., Zhang, R., Shao, W., Gao, P., Xu, P., Xiao, H., Zhang, K., Liu, C., Wen, S., Guo, Z., Lu, X., Ren, S., Wen, Y., Chen, X., Yue, X., Li, H., Qiao, Y.: Imagebind-llm: Multi-modality instruction tuning. CoRR abs/2309.03905 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 426, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 426, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 426, + 91, + 447, + 100 + ], + "type": "text", + "content": "CAT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "16. Hendrycks, D., Burns, C., Basart, S., Zou, A., Mazeika, M., Song, D., Steinhardt, J.: Measuring massive multitask language understanding. In: ICLR (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "type": "text", + "content": "17. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. In: The Tenth International Conference on Learning Representations, ICLR (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 193 + ], + "type": "text", + "content": "18. Huang, B., Wang, X., Chen, H., Song, Z., Zhu, W.: Vtimellm: Empower LLM to grasp video moments. CoRR abs/2311.18445 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 194, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 215 + ], + "type": "text", + "content": "19. Jiang, P., Han, Y.: Reasoning with heterogeneous graph alignment for video question answering. In: AAAI. pp. 11109-11116. AAAI Press (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "text", + "content": "20. Jin, P., Takanobu, R., Zhang, C., Cao, X., Yuan, L.: Chat-univi: Unified visual representation empowers large language models with image and video understanding. CoRR abs/2311.08046 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "21. Korbar, B., Xian, Y., Tonioni, A., Zisserman, A., Tombari, F.: Text-conditioned resampler for long form video understanding. CoRR abs/2312.11897 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 480, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 324 + ], + "type": "text", + "content": "22. Lavie, A., Agarwal, A.: METEOR: an automatic metric for MT evaluation with high levels of correlation with human judgments. In: Callison-Burch, C., Koehn, P., Fordyce, C.S., Monz, C. (eds.) Proceedings of the Second Workshop on Statistical Machine Translation. pp. 228-231. Association for Computational Linguistics (2007)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 380 + ], + "type": "text", + "content": "23. Le, H., Sahoo, D., Chen, N.F., Hoi, S.C.H.: Multimodal transformer networks for end-to-end video-grounded dialogue systems. In: Korhonen, A., Traum, D.R., Marquez, L. (eds.) Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL. pp. 5612-5623. Association for Computational Linguistics (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 423 + ], + "type": "text", + "content": "24. Le, T.M., Le, V., Venkatesh, S., Tran, T.: Hierarchical conditional relation networks for video question answering. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR. pp. 9969-9978. Computer Vision Foundation / IEEE (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "25. Li, G., Hou, W., Hu, D.: Progressive spatio-temporal perception for audio-visual question answering. In: El-Saddik, A., Mei, T., Cucchiara, R., Bertini, M., Vallejo, D.P.T., Atrey, P.K., Hossain, M.S. (eds.) ACM MM. pp. 7808-7816. ACM (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "type": "text", + "content": "26. Li, G., Wei, Y., Tian, Y., Xu, C., Wen, J.R., Hu, D.: Learning to answer questions in dynamic audio-visual scenarios. In: CVPR. p. 19086–19096 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 522 + ], + "type": "text", + "content": "27. Li, J., Li, D., Savarese, S., Hoi, S.C.H.: BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In: Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., Scarlett, J. (eds.) International Conference on Machine Learning, ICML. pp. 19730–19742. PMLR (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "type": "text", + "content": "28. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: Videochat: Chat-centric video understanding. CoRR abs/2305.06355 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 544, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 480, + 577 + ], + "type": "text", + "content": "29. Li, K., Wang, Y., He, Y., Li, Y., Wang, Y., Liu, Y., Wang, Z., Xu, J., Chen, G., Luo, P., Wang, L., Qiao, Y.: Mvbench: A comprehensive multi-modal video understanding benchmark. CoRR abs/2311.17005 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "text", + "content": "30. Li, X., Song, J., Gao, L., Liu, X., Huang, W., He, X., Gan, C.: Beyond rnns: Positional self-attention with co-attention for video question answering. In: AAAI. pp. 8658-8665. AAAI Press (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "31. Li, Y., Wang, C., Jia, J.: Llama-vid: An image is worth 2 tokens in large language models. CoRR abs/2311.17043 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "32. Li, Z., Yang, B., Liu, Q., Ma, Z., Zhang, S., Yang, J., Sun, Y., Liu, Y., Bai, X.: Monkey: Image resolution and text label are important things for large multi-modal models. CoRR abs/2311.06607 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Q. Ye et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "33. Lin, B., Ye, Y., Zhu, B., Cui, J., Ning, M., Jin, P., Yuan, L.: Video-llava: Learning united visual representation by alignment before projection. CoRR abs/2311.10122 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "34. Lin, Y., Sung, Y., Lei, J., Bansal, M., Bertasius, G.: Vision transformers are parameter-efficient audio-visual learners. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR. pp. 2299-2309. IEEE (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "type": "text", + "content": "35. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. CoRR abs/2304.08485 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 248 + ], + "type": "text", + "content": "36. Lu, J., Yang, J., Batra, D., Parikh, D.: Hierarchical question-image co-attention for visual question answering. In: Lee, D.D., Sugiyama, M., von Luxburg, U., Guyon, I., Garnett, R. (eds.) Advances in Neural Information Processing Systems NIPs. pp. 289–297 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 249, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 281 + ], + "type": "text", + "content": "37. Luo, R., Zhao, Z., Yang, M., Dong, J., Qiu, M., Lu, P., Wang, T., Wei, Z.: Valley: Video assistant with large language model enhanced ability. CoRR abs/2306.07207 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 314 + ], + "type": "text", + "content": "38. Lyu, C., Wu, M., Wang, L., Huang, X., Liu, B., Du, Z., Shi, S., Tu, Z.: Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. CoRR abs/2306.09093 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 315, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 335 + ], + "type": "text", + "content": "39. Ma, F., Jin, X., Wang, H., Xian, Y., Feng, J., Yang, Y.: Vista-llama: Reliable video narrator via equal distance to visual tokens. CoRR abs/2312.08870 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "type": "text", + "content": "40. Maaz, M., Rasheed, H.A., Khan, S.H., Khan, F.S.: Video-chatgpt: Towards detailed video understanding via large vision and language models. CoRR abs/2306.05424 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 370, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 370, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 482, + 413 + ], + "type": "text", + "content": "41. MacGlashan, J., Ho, M.K., Loftin, R.T., Peng, B., Wang, G., Roberts, D.L., Taylor, M.E., Littman, M.L.: Interactive learning from policy-dependent human feedback. In: Precup, D., Teh, Y.W. (eds.) Proceedings of the 34th International Conference on Machine Learning, ICML. pp. 2285-2294 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 414, + 482, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 445 + ], + "type": "text", + "content": "42. Mei, X., Meng, C., Liu, H., Kong, Q., Ko, T., Zhao, C., Plumbley, M.D., Zou, Y., Wang, W.: Wavcaps: A chatgpt-assisted weakly-labelled audio captioning dataset for audio-language multimodal research. CoRR abs/2303.17395 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 447, + 482, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 468 + ], + "type": "text", + "content": "43. Nadeem, A., Hilton, A., Dawes, R., Thomas, G., Mustafa, A.: CAD - contextual multi-modal alignment for dynamic AVQA. CoRR abs/2310.16754 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 468, + 482, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 490 + ], + "type": "text", + "content": "44. Nguyen, D.T., Sharma, S., Schulz, H., Asri, L.E.: From film to video: Multi-turn question answering with multi-modal context. CoRR abs/1812.07023 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 491, + 418, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 418, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 418, + 501 + ], + "type": "text", + "content": "45. OpenAI: GPT-4 technical report. CoRR abs/2303.08774 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 502, + 482, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 555 + ], + "type": "text", + "content": "46. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P.F., Leike, J., Lowe, R.: Training language models to follow instructions with human feedback. In: Advances in Neural Information Processing Systems, NeurIPS (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "type": "text", + "content": "47. Papineni, K., Roukos, S., Ward, T., Zhu, W.: Bleu: a method for automatic evaluation of machine translation. In: Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics. pp. 311-318. ACL (2002)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 589, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 589, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 589, + 482, + 632 + ], + "type": "text", + "content": "48. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: Moschitti, A., Pang, B., Daelemans, W. (eds.) Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, EMNLP. pp. 1532-1543. ACL (2014)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 633, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 633, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 633, + 482, + 666 + ], + "type": "text", + "content": "49. Pham, H., Le, T.M., Le, V., Phuong, T.M., Tran, T.: Video dialog as conversation about objects living in space-time. In: Avidan, S., Brostow, G.J., Cissé, M., Farinella, G.M., Hassner, T. (eds.) ECCV. vol. 13699, pp. 710-726. Springer (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 426, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 426, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 426, + 91, + 447, + 100 + ], + "type": "text", + "content": "CAT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 172 + ], + "type": "text", + "content": "50. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the International Conference on Machine Learning, ICML. vol. 139, pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 173, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 173, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 481, + 217 + ], + "type": "text", + "content": "51. Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Robust speech recognition via large-scale weak supervision. In: Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., Scarlett, J. (eds.) International Conference on Machine Learning, ICML. pp. 28492-28518. PMLR (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 251 + ], + "type": "text", + "content": "52. Rafailov, R., Sharma, A., Mitchell, E., Ermon, S., Manning, C.D., Finn, C.: Direct preference optimization: Your language model is secretly a reward model. CoRR abs/2305.18290 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 251, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 274 + ], + "type": "text", + "content": "53. Sakaguchi, K., Bras, R.L., Bhagavatula, C., Choi, Y.: Winogrande: An adversarial winograd schema challenge at scale. In: AAAI. pp. 8732-8740 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "type": "text", + "content": "54. Schwartz, I., Schwing, A.G., Hazan, T.: A simple baseline for audio-visual scene-aware dialog. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR. pp. 12548-12558. Computer Vision Foundation / IEEE (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 308, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 439 + ], + "type": "text", + "content": "55. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D., Blecher, L., Canton-Ferrer, C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I., Korenev, A., Koura, P.S., Lachaux, M., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A., Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X.E., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models. CoRR abs/2307.09288 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 441, + 481, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 481, + 474 + ], + "type": "text", + "content": "56. Vedantam, R., Zitnick, C.L., Parikh, D.: Cider: Consensus-based image description evaluation. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR. pp. 4566-4575 (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 475, + 481, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 507 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 507 + ], + "type": "text", + "content": "57. Wang, X., Wu, J., Chen, J., Li, L., Wang, Y., Wang, W.Y.: Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In: IEEE/CVF International Conference on Computer Vision, ICCV. pp. 4580-4590. IEEE (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 508, + 481, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 508, + 481, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 508, + 481, + 540 + ], + "type": "text", + "content": "58. Xu, D., Zhao, Z., Xiao, J., Wu, F., Zhang, H., He, X., Zhuang, Y.: Video question answering via gradually refined attention over appearance and motion. In: ACM MM. pp. 1645-1653 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 541, + 481, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 541, + 481, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 541, + 481, + 574 + ], + "type": "text", + "content": "59. Yang, P., Wang, X., Duan, X., Chen, H., Hou, R., Jin, C., Zhu, W.: AVQA: A dataset for audio-visual question answering on videos. In: ACM MM. pp. 3480-3491. ACM (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 575, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 575, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 575, + 481, + 609 + ], + "type": "text", + "content": "60. Yu, Z., Xu, D., Yu, J., Yu, T., Zhao, Z., Zhuang, Y., Tao, D.: Activitynet-qa: A dataset for understanding complex web videos via question answering. In: AAAI. pp. 9127-9134 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 609, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 481, + 643 + ], + "type": "text", + "content": "61. Yu, Z., Yu, J., Cui, Y., Tao, D., Tian, Q.: Deep modular co-attention networks for visual question answering. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR. pp. 6281-6290. Computer Vision Foundation / IEEE (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "62. Yun, H., Yu, Y., Yang, W., Lee, K.I., Kim, G.H.: Pano-avqa: Grounded audiovisual question answering on " + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": " videos. In: CVPR. p. 2031-2041 (2021)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Q. Ye et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 281 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "63. Zellers, R., Holtzman, A., Bisk, Y., Farhadi, A., Choi, Y.: Hellaswag: Can a machine really finish your sentence? In: Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL. pp. 4791-4800 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "64. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding. In: Proceedings of the Empirical Methods in Natural Language Processing, EMNLP. pp. 543-553 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "text", + "content": "65. Zhang, R., Han, J., Zhou, A., Hu, X., Yan, S., Lu, P., Li, H., Gao, P., Qiao, Y.: Llama-adapter: Efficient fine-tuning of language models with zero-init attention. CoRR abs/2303.16199 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 247 + ], + "type": "text", + "content": "66. Zhang, S., Dong, L., Li, X., Zhang, S., Sun, X., Wang, S., Li, J., Hu, R., Zhang, T., Wu, F., Wang, G.: Instruction tuning for large language models: A survey. CoRR abs/2308.10792 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "type": "text", + "content": "67. Zhao, Z., Guo, L., Yue, T., Chen, S., Shao, S., Zhu, X., Yuan, Z., Liu, J.: Chatbridge: Bridging modalities with large language model as a language catalyst. CoRR abs/2305.16103 (2023)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 426, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 426, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 426, + 91, + 447, + 100 + ], + "type": "text", + "content": "CAT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CC-SAM_ Enhancing SAM with Cross-feature Attention and Context for Ultrasound Image Segmentation/4f25c9c8-80da-4658-ab8a-cd035c32acd6_content_list.json b/2024/CC-SAM_ Enhancing SAM with Cross-feature Attention and Context for Ultrasound Image Segmentation/4f25c9c8-80da-4658-ab8a-cd035c32acd6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..331e04447a6ec1b1884d1b694228a1f82efa0196 --- /dev/null +++ b/2024/CC-SAM_ Enhancing SAM with Cross-feature Attention and Context for Ultrasound Image Segmentation/4f25c9c8-80da-4658-ab8a-cd035c32acd6_content_list.json @@ -0,0 +1,1727 @@ +[ + { + "type": "text", + "text": "CC-SAM: SAM with Cross-feature Attention and Context for Ultrasound Image Segmentation", + "text_level": 1, + "bbox": [ + 218, + 140, + 782, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shreyank N Gowda1 and David A. Clifton1,2", + "bbox": [ + 336, + 212, + 661, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Engineering Sciences, University of Oxford, OX3 7DQ Oxford, UK", + "bbox": [ + 220, + 239, + 781, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Oxford Suzhou Centre for Advanced Research, University of Oxford, Suzhou 215123, Jiangsu, China", + "bbox": [ + 238, + 253, + 761, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The Segment Anything Model (SAM) has achieved remarkable successes in the realm of natural image segmentation, but its deployment in the medical imaging sphere has encountered challenges. Specifically, the model struggles with medical images that feature low contrast, faint boundaries, intricate morphologies, and small-sized objects. To address these challenges and enhance SAM's performance in the medical domain, we introduce a comprehensive modification. Firstly, we incorporate a frozen Convolutional Neural Network (CNN) branch as an image encoder, which synergizes with SAM's original Vision Transformer (ViT) encoder through a novel variational attention fusion module. This integration bolsters the model's capability to capture local spatial information, which is often paramount in medical imagery. Moreover, to further optimize SAM for medical imaging, we introduce feature and position adapters within the ViT branch, refining the encoder's representations. We see that compared to current prompting strategies to fine-tune SAM for ultrasound medical segmentation, the use of text descriptions that serve as text prompts for SAM helps significantly improve the performance. Leveraging ChatGPT's natural language understanding capabilities, we generate prompts that offer contextual information and guidance to SAM, enabling it to better understand the nuances of ultrasound medical images and improve its segmentation accuracy. Our method, in its entirety, represents a significant stride towards making universal image segmentation models more adaptable and efficient in the medical domain.", + "bbox": [ + 259, + 323, + 738, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 684, + 375, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning has revolutionized the fields of image [19, 24, 27], video classification [15, 18] and medical vision [45, 50, 56], enabling unprecedented accuracy and efficiency in these areas. Medical image segmentation plays a pivotal role in delineating and emphasizing particular organs, tissues, and anomalies within medical scans, forming a cornerstone of computer-aided diagnostic systems [35]. The development of numerous deep learning models for autonomous medical image segmentation has signaled a transformative shift in the potential of this technology [7, 45, 50, 56]. Nevertheless, these models are typically specialized for", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9b1618edebc9859558c44ab5b09ccd1effda37631ef997a96953c9ec7d244f73.jpg", + "image_caption": [ + "Fig. 1: Comparison of methods using SAM for medical image segmentation." + ], + "image_footnote": [], + "bbox": [ + 236, + 143, + 769, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "distinct tasks and require recalibration when used for different tasks, a limitation which poses significant challenges in medical data.", + "bbox": [ + 212, + 382, + 784, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Segment anything model (SAM) [31] has emerged as a versatile foundational model for vision segmentation, gaining significant attention due to its unparalleled ability to segment a wide variety of objects and exhibit powerful zero-shot generalization. With user-driven prompts such as points, bounding boxes, and coarse masks, SAM dynamically segments the relevant objects. This adaptability positions SAM as an ideal tool for a myriad of segmentation use cases, facilitating the amalgamation of several discrete medical image segmentation tasks into a holistic model, enhancing its utility for clinical applications [29].", + "bbox": [ + 212, + 414, + 784, + 535 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, while SAM boasts an extensive dataset (SA1B), its performance diminishes in the medical realm due to the lack of comprehensive clinical annotations [29]. To enhance SAM's performance in medical image segmentation, many approaches have been proposed such as tuning the mask decoder [37], applying the LoRA strategy to the image encoder [53], introducing task-specific information using adapters in the ViT image encoder [51]. Previous endeavors to adapt SAM for medical image segmentation by optimizing it on medical datasets have encountered limitations, given their tokenization approach that can obfuscate essential local data.", + "bbox": [ + 212, + 536, + 785, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SAMUS [33] was subsequently introduced, aiming to transplant SAM's superior segmentation prowess and robust generalization to medical image segmentation, all while ensuring computational efficiency. Notable improvements within SAMUS encompass a redesigned ViT image encoder, integration of a parallel CNN-branch, and the introduction of a cross-branch attention mechanism, among others.", + "bbox": [ + 212, + 672, + 785, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce CC-SAM, a novel model built upon SAMUS. Notably, CC-SAM replaces the adjustable CNN with a fixed one, enhanced with adapters and employs a novel variational attention fusion instead of the cross-attention branch. Utilizing descriptions from Chat-GPT to prompt the model, we achieve a marked improvement in performance over existing prompting strategies", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Shreyank N Gowda and David A. Clifton", + "bbox": [ + 269, + 114, + 547, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tailored for medical-oriented SAM. This new approach significantly improves medical-oriented SAM performance. See Figure 1 for a comparison with other medical SAM adaptations.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The chief contributions of this paper are delineated as follows:", + "bbox": [ + 238, + 191, + 686, + 207 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The development of CC-SAM: a refined foundation model tailored for universal medical image segmentation, optimized for computational efficiency.", + "- The introduction of a static CNN with adapters, complementing the ViT encoder of SAMUS and further reducing computational costs.", + "- The innovative integration of variational attention fusion, enhancing the synergy between CNN and ViT branch features.", + "- The use of text-based prompts, sourced from Chat-GPT, to articulate the segmentation problem, which significantly amplifies segmentation performance in medical contexts." + ], + "bbox": [ + 225, + 220, + 782, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 382, + 385, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Medical Image Segmentation", + "text_level": 1, + "bbox": [ + 215, + 415, + 500, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Medical image segmentation is crucial for identifying and measuring structures in medical images. Before deep learning, methods like thresholding, clustering, and active contour models were common [43]. However, CNNs, particularly U-Net [45], have since dominated due to their quality results even with limited data.", + "bbox": [ + 212, + 441, + 782, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Extensions of U-Net, such as DRINet [7], tackled class imbalance and high-resolution outputs. U-Net++ [56] used nested and dense skip pathways to improve segmentation, while UNeXt [48] optimized for speed and efficiency in medical segmentation by using a Convolutional MLP-based network.", + "bbox": [ + 212, + 517, + 782, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The rise of image transformers [9] led to new methods. FAT-Net [50] introduced feature adaptive transformers for detailed segmentation. UNeTR [22] combined transformers with a U-shaped design for 3D medical imaging. Our approach merges CNNs and transformers to harness both local and global feature extraction, creating a hybrid framework that leverages the strengths of both paradigms.", + "bbox": [ + 212, + 578, + 782, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Adapting Foundational Models", + "text_level": 1, + "bbox": [ + 215, + 691, + 516, + 707 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Foundational models, often pre-trained on large datasets, provide a generic understanding of complex patterns in data. These models are crucial stepping stones in many modern machine learning workflows. However, leveraging them effectively for specific tasks necessitates a thoughtful adaptation strategy.", + "bbox": [ + 212, + 719, + 782, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adapter modules [26] offer a solution, allowing the main network weights to remain unchanged while adding learnable layers to each Transformer layer. This adds task-specific abilities with less overhead, ensuring efficient transfer learning. This approach keeps the base knowledge while adapting to new tasks.", + "bbox": [ + 212, + 780, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Title Suppressed Due to Excessive Length", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These modules have excelled in various language tasks [13, 36, 47]. For instance, they've achieved top results in question answering with $95\\%$ fewer parameters than standard finetuning [42]. In textual entailment, AdaMix [49] found adapters performed comparably to full finetuning but with only $3\\%$ added parameters. They've also been effective in computer vision tasks like image [10, 12] and video classification [16, 17, 40]. Essentially, they ensure efficient transfer learning across tasks by adding minimal layers to fixed networks.", + "bbox": [ + 212, + 146, + 787, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We build on the idea of adapters and freeze both our CNN and ViT backbones and only use adapters for efficient fine-tuning of both models.", + "bbox": [ + 212, + 252, + 785, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Adapting SAM for Medical Images", + "text_level": 1, + "bbox": [ + 214, + 300, + 552, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Segment Anything Model (SAM) [31] has demonstrated notable zero-shot segmentation using various input prompts. However, its performance on medical images is subpar [25, 46].", + "bbox": [ + 212, + 321, + 784, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Several strategies have been proposed to adapt SAM for medical imaging. MedSAM optimizes the mask decoder for efficient fine-tuning. SAMed [53] freezes the image encoder and employs a low-rank finetuning strategy while adjusting the prompt encoder and mask decoder. MSA [51] uses adapters in the ViT image encoder. To address transformer backbone limitations, SAMUS [33] incorporates a CNN-based backbone for improved local feature learning.", + "bbox": [ + 212, + 367, + 787, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Building on SAMUS [33], we optimize its framework. Instead of end-to-end CNN fine-tuning, we utilize a frozen pre-trained CNN and an adapter. We found that using text prompts, especially with Grounding DINO [34], significantly enhances SAMUS's performance. This method is not only more effective but also faster. We use GPT-4 [39] to generate text prompts from class labels.", + "bbox": [ + 212, + 457, + 787, + 534 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/211dd504b834998cb26db7e9be60fcef457b8d99aeddf97e748a2f50ec2815b1.jpg", + "image_caption": [ + "Fig. 2: Overview of CC-SAM. We use adapters (or in the case of the CNN a FC layer) to enhance local and global features for ultrasound segmentation. 'OPaE' refers to overlapping patch embeddings and 'PE' refers to positional embeddings." + ], + "image_footnote": [], + "bbox": [ + 230, + 558, + 774, + 758 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Shreyank N Gowda and David A. Clifton", + "bbox": [ + 271, + 114, + 547, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Proposed Method", + "text_level": 1, + "bbox": [ + 215, + 143, + 426, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 2 provides a detailed view of our method. Essentially, we process an image $I$ using ViT-B [9] (pre-trained image encoder of SAM) and ResNet50 [24], the ResNet50 is pre-trained on RadImageNet [38]. After feature extraction, adapters refine these representations. Features are then fused using a variational attention mechanism, forming a combined feature for the Mask Decoder of SAM.", + "bbox": [ + 212, + 183, + 782, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For prompts, we use a GPT-4 generated label description and embed it with Med-BERT [44]. The Grounding DINO model, untrained on medical data, creates a bounding box from this embedding, which is input to the Prompt Encoder. Together with the variational attention feature, the Prompt Encoder helps the Mask Decoder create the final segmentation mask.", + "bbox": [ + 212, + 258, + 784, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Frozen Backbones with Adapters", + "text_level": 1, + "bbox": [ + 214, + 364, + 534, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We've enhanced SAM's image encoder (the ViT branch) for better adaptability to smaller inputs and medical images by introducing a position adapter and five feature adapters, following SAMUS [33]. These adapters fine-tune the ViT branch efficiently with fewer parameters.", + "bbox": [ + 212, + 395, + 782, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, the position adapter modifies positional embeddings to match the embedded sequence resolution. It first downsamples these embeddings using max pooling and then refines them with a convolution operation, enabling the ViT to better manage smaller inputs. Each of the five feature adapters follows a consistent design with a downward linear projection, an activation, and an upward projection. This is mathematically expressed as:", + "bbox": [ + 212, + 457, + 784, + 547 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA (x) = G (x \\mathcal {M} _ {d}) \\mathcal {M} _ {u} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 428, + 568, + 784, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, Here, $G$ denotes the GELU activation function, while $\\mathcal{M}_d\\in \\mathbb{R}^{d\\times \\frac{d}{4}}$ and $\\mathcal{M}_u\\in \\mathbb{R}^{d\\times \\frac{d}{4}}$ are the projection matrices. In this context, $^{\\prime}d^{\\prime}$ represents the dimension of the feature embedding.", + "bbox": [ + 212, + 595, + 782, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike SAMUS's end-to-end CNN training, we use a static, RadImageNet [38] pre-trained ResNet-50 model. We add a trainable fully-connected layer, acting as an adapter, before its classification layer. The output features from both branches then go into our variational attention fusion block.", + "bbox": [ + 212, + 643, + 782, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Variational Attention Fusion Block", + "text_level": 1, + "bbox": [ + 214, + 732, + 549, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A major contribution we've made is the Variational Attention Fusion Block, which adeptly merges local CNN features and global ViT features. While SAMUS introduced a cross-branch attention module, our method, which models uncertainty around these features and applies variational attentional fusion, outperforms it (details in Sec 4.3). See Figure 3 for an overview.", + "bbox": [ + 212, + 763, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Title Suppressed Due to Excessive Length", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7a3215e96cf544e83a943dccf29de6aba509c0b8a7551bd11df0b21053c2921b.jpg", + "image_caption": [ + "Fig. 3: Overview of the proposed Variational Attention Fusion Block. Each 'mode' has an intra-modal uncertainty learning encoder (represented as $\\mathbf{E}$ or $\\mathbf{E}^{\\prime}$ in the figure), these obtain robust modality-specific features in the latent subspace. Subsequently, VAF combines these inputs and constructs a multimodal representation by estimating weights that are specific to each modality, effectively capturing their dependencies." + ], + "image_footnote": [], + "bbox": [ + 299, + 151, + 709, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Uncertainty Learning Incorporating uncertainty learning, for an input image $I$ , we derive features $s^v$ from the ViT and $s^c$ from the ResNet-50. Both $s^v$ and $s^c \\in \\mathbb{R}^{d^h}$ represent a fixed dimension in a latent space. We achieve a consistent size using latent encoders $E_v, E_c$ at each network's end. To address data uncertainty, we design a latent distribution for each sample's content, helping capture semantic relations. This distribution is shaped as a parameterized diagonal Gaussian. This is mathematically represented as:", + "bbox": [ + 214, + 503, + 787, + 612 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np \\left(z ^ {v} \\mid s ^ {v}\\right) = \\mathcal {N} \\left(z ^ {v}; \\mu^ {v}, \\left(\\sigma^ {v}\\right) ^ {2} \\mathbf {I}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 625, + 785, + 643 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np \\left(z ^ {c} \\mid s ^ {c}\\right) = \\mathcal {N} \\left(z ^ {c}; \\mu^ {c}, \\left(\\sigma^ {c}\\right) ^ {2} \\mathbf {I}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 400, + 662, + 785, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $z^v$ and $z^c$ are reconstructed vectors. The means, $\\mu^v, \\mu^c$ , capture each mode's core feature, while variances, $\\sigma^v, \\sigma^c$ , indicate the noise-induced uncertainty in predicting these means. A higher variance signifies more uncertainty about the content observed. Both Gaussian parameters depend on the input and are forecasted through MLPs. For instance, for ViT features: $\\mu^v = f\\theta_{v1}(sv)$ and $\\sigma^v = f\\theta_{v2}(sv)$ , with $\\theta_{v1}$ and $\\theta_{v2}$ being respective model parameters. The same applies to CNN features. Now, each sample's feature representation shifts from deterministic to a stochastic Gaussian-drawn embedding in the latent space.", + "bbox": [ + 214, + 688, + 787, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Due to the non-differentiable nature of sampling, we use a re-parameterization trick [30] to maintain gradient flow. We draw random noise $\\eta$ from a normal dis", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Shreyank N Gowda and David A. Clifton", + "bbox": [ + 271, + 114, + 547, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ribution, independent of model parameters, and produce $z^v$ using it, as detailed in Eq 4. We repeat this for $z^c$ .", + "bbox": [ + 214, + 146, + 785, + 176 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nz ^ {v} = \\mu^ {v} + \\eta \\sigma^ {v}, \\eta \\in \\mathcal {N} (0, 1) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 191, + 785, + 209 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Variational Attentional Fusion CNNs and ViTs capture different feature aspects from images (local and global). We see these as two modes: $v$ for ViTs and $c$ for CNNs. Combining them isn't straightforward due to different confidence levels. Standard methods, which get weights from the data, miss the unique traits of each mode. To tackle this, we've developed a Variational Attention Fusion (VAF) module. This module aims to capture the nuances between modalities by determining modality-specific weights for a seamless integration.", + "bbox": [ + 214, + 228, + 782, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In short, using features from each mode, a standard attention method creates a probability, $a^k$ (where $k$ is either $v$ or $c$ ), as shown in Eq 5. Here, $a^k$ shows how much each mode contributes, and $\\mathbf{W}_m$ and $b_m$ are trainable factors. By having set weights, it picks the right features without being affected by varied confidence levels between modes.", + "bbox": [ + 214, + 334, + 782, + 407 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {a} ^ {k} = \\mathbf {W} _ {m} \\mathbf {z} ^ {k} + b _ {m}\n$$\n", + "text_format": "latex", + "bbox": [ + 460, + 433, + 589, + 449 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\na ^ {k} = \\frac {\\exp \\left(\\hat {a} ^ {k}\\right)}{\\sum_ {k ^ {\\prime} \\in \\{v , c \\}} \\exp \\left(\\hat {a} ^ {k ^ {\\prime}}\\right)} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 453, + 784, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our VAF module uses variational attention weights instead of the point-estimated attention vector $\\hat{a}^k$ . This approach, grounded in a probability distribution (Eq. 6), better accounts for uncertainties across modalities.", + "bbox": [ + 214, + 498, + 782, + 544 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {a} ^ {k} \\sim q _ {\\theta} (\\hat {a} ^ {k} \\mid z ^ {k}) = \\mathcal {N} \\left(\\mu_ {a} ^ {k}, \\left(\\sigma_ {a} ^ {k}\\right) ^ {2} \\mathbf {I}\\right), k \\in \\{v, c \\} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 554, + 784, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We used an identity transformation to preserve modality traits in the variational attention. Mean and variance parameters are predicted by MLPs from the input. The $\\left(\\sigma_{a}^{k}\\right)$ value indicates confidence levels among modalities. With the VAF method, more confident modality features are enhanced, while less confident ones are reduced, optimizing multimodal data fusion and capturing complementary features. The final representation combines modality-specific outputs, $z_{c}$ and $z_{v}$ , using weighted aggregation as shown in Eq 7, where $\\mathbf{W}_h$ is a learnable weight matrix.", + "bbox": [ + 214, + 584, + 785, + 704 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nh = \\sum_ {k \\in \\{v, c \\}} a ^ {k} \\mathbf {W} _ {h} z ^ {k} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 429, + 717, + 785, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Guiding the Prompt Encoder", + "text_level": 1, + "bbox": [ + 214, + 770, + 504, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SAM's mask decoder [31] needs $h$ and an input from the prompt encoder. While past adaptations of SAM for medical imaging used point prompts, our results show that a reasonably good bounding box enhances performance. Creating this", + "bbox": [ + 214, + 794, + 782, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Title Suppressed Due to Excessive Length", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "bounding box requires a dedicated model. Using random bounding boxes may hinder results, necessitating specific prompts. We employ Grounding-DINO [34] for bounding box generation. Though not exclusively trained for medical images, Grounding-DINO is an effective object detector. We use GPT-4 [39] to craft text prompts for specific class labels, and then MedBERT [44] generates an embedding input for Grounding-DINO. Whilst this does improve our performance in comparison to using point prompts, we show in supplementary that using the same set of random point prompts, we outperform all other medical foundational models.", + "bbox": [ + 212, + 146, + 787, + 281 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Input to GPT-4 \"Create a concise description for a medical image analysis task. The task involves using ultrasound images to segment and identify specific anatomical structures or pathologies. Use [target] for to make a generic description that can be adapted for any segmentation target.\" Output from GPT-4 \"Segment and identify [target] within ultrasound images. This requires the precise delineation of [target] from surrounding tissues, enabling accurate diagnosis and assessment. The challenge encompasses dealing with the inherent variability of ultrasound image quality, including speckle noise, shadowing, and the wide range of anatomical variations among patients. Success in this task is measured by the algorithm's ability to consistently and accurately identify and outline the [target], providing crucial information for medical decision-making processes.\" Where [target] is modified for each dataset.", + "bbox": [ + 214, + 282, + 789, + 449 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Loss Function", + "text_level": 1, + "bbox": [ + 215, + 465, + 377, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our overall loss function can be seen in Eq 8. Here, $\\mathcal{L}_{BCE}$ and $\\mathcal{L}_D$ are segmentation losses: binary cross entropy and dice loss respectively.", + "bbox": [ + 214, + 487, + 787, + 518 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {B C E} + \\mathcal {L} _ {D} + \\lambda_ {c} \\mathcal {L} _ {c} + \\lambda_ {v} \\mathcal {L} _ {v} + \\lambda_ {a} \\mathcal {L} _ {a} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 532, + 784, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\mathcal{L}c$ and $\\mathcal{L}v$ are regularization terms for uncertainty learning, inspired by the variational information bottleneck [2]. To grasp uncertainty in a modality, a component is introduced during learning. It ensures the data's distribution, $\\mathcal{N}\\left(z^v;\\mu^v,(\\sigma^v)^2\\right)$ , resembles a standard bell curve, $\\mathcal{N}(\\epsilon ;0,\\mathbf{I})$ . This similarity is gauged using Kullback-Leibler divergence (KLD) between the distributions, promoting model diversity and reducing uncertainty, leading to robust features. Eq 9 shows how $\\mathcal{L}v$ is computed, with analogous methods for $\\mathcal{L}c$ and $\\mathcal{L}_a$ , replacing $z^{v}$ with $z^c$ and $\\hat{a}^k$ .", + "bbox": [ + 212, + 551, + 787, + 679 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {v} = K L \\left(\\mathcal {N} \\left(z ^ {v}; \\mu^ {v}, (\\sigma^ {v}) ^ {2}\\right) \\| \\mathcal {N} (\\epsilon ; 0, \\mathbf {I})\\right) \\\\ = - \\frac {1}{2} \\left(1 + \\log (\\sigma^ {v}) ^ {2} - (\\mu^ {v}) ^ {2} - (\\sigma^ {v}) ^ {2}\\right) \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 690, + 784, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experimental Analysis", + "text_level": 1, + "bbox": [ + 215, + 761, + 472, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 215, + 789, + 336, + 803 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To compare with SAMUS [33], we test CC-SAM on the same seven public datasets: TN3K [14], DDTI [41], TG3K [52], BUSI [1], UDIAT [3], CAMUS [32],", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Shreyank N Gowda and David A. Clifton", + "bbox": [ + 271, + 114, + 547, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and HMC-QU [8]. We also recreate the US30K dataset from SAMUS for broader comparison. Dataset details are in the supplementary material.", + "bbox": [ + 212, + 145, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "TN3K and TG3K datasets follow TRFE [14] for segmentation into training, validation, and testing. BUSI is randomly split in a 7:1:2 ratio for these purposes. CAMUS is initially divided into training and testing per the challenge [32], with $10\\%$ of training data later used for validation.", + "bbox": [ + 212, + 176, + 782, + 236 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For model generalizability, US30K datasets remain unseen during training and validation. In line with SAMUS, we split US30k into seen and unseen datasets. We then evaluate on unseen datasets (DDTI, UDIAT, HMC-QU) not used in training or validation. This enables direct comparison of generalization with other state-of-the-art models. Moreover, we contrast CC-SAM with other base models by training on the full US30K dataset and testing across different tasks.", + "bbox": [ + 212, + 238, + 784, + 342 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 369, + 457, + 386 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We follow SAMUS in setting hyper parameters and adapter sizes. We change the fully tuneable CNN in SAMUS to a static ResNet50. The model is trained using the Adam optimizer with an initial learning rate of 0.01 that reduces by a tenth every 50 epochs for a total of 200 epochs. The encoders are fully connected networks with the latent space having dimensions 1024. All fully connected layers has a Leaky RELU activation function and a dropout probability of 0.5.", + "bbox": [ + 212, + 397, + 782, + 488 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We use 1xA100 GPU with 40 GB capacity. But we also verified that a single 3090 Ti with 24 GB memory is enough to train the model. We set a batch size of 32.", + "bbox": [ + 212, + 489, + 784, + 532 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 560, + 388, + 575 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We consider SAMUS [33] as our baseline model and ablate our added components. Specifically, we show replacing the end-to-end CNN model with a fixed pre-trained model improves performance. Further adding the FC layer at the end slightly improves performance. Crucially, replacing the cross attention fusion with our variational attention fusion significantly improves performance. Further, replacing the point based prompting with our text-based prompting gives us the best results. We show these results on two of the datasets TN3K and BUSI in Table 1.", + "bbox": [ + 212, + 588, + 782, + 708 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We do a further analysis of how the prompting compares with different prompting strategies in the supplementary material.", + "bbox": [ + 212, + 710, + 782, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4 Comparison with state-of-the-art", + "text_level": 1, + "bbox": [ + 214, + 766, + 532, + 781 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We consider two scenarios for comparing with state-of-the-art models following SAMUS [33]. We compare CC-SAM with task-specific methods and foundational models.", + "bbox": [ + 212, + 794, + 782, + 838 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Title Suppressed Due to Excessive Length", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/d2a82fd8cf37d91e0e415517b30d202697ef65f0d2528b98c15818c02ba66559.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ComponentsTN3KBUSI
ResNet50VAFText PromptDiceHDDiceHD
×××84.4528.2285.7725.49
✓××84.9127.2586.2225.12
×✓×85.5526.8487.1524.42
××✓85.9526.2287.9223.92
✓×✓86.4225.8288.4523.20
×✓✓86.2525.9288.2223.51
✓✓×85.7226.2187.4224.61
✓✓✓87.1125.2589.5122.86
", + "bbox": [ + 313, + 143, + 687, + 284 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Task-specific methods Following SAMUS [33], thirteen state-of-the-art, task-specific techniques are chosen for evaluation, spanning CNN-based, transformer-based, and CNN-Transformer hybrid methods. Methods other than SAMUS include CNN-based techniques encompass U-Net [45], CPFNet [11], CA-Net [20], CE-Net [21], and AAU-Net [5]. The transformer-centric methods consist of SwinUnet [4], SETR [55], and MISSFormer [28]. The hybrids of CNN and Transformer are TransUNet [6], TransFuse [54], FAT-Net [50], and H2Former [23].", + "bbox": [ + 212, + 364, + 784, + 472 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/34116519db88cf85f8d4fd1a5d906eb8696653c5695140778d51743c930ec764.jpg", + "table_caption": [ + "Table 1: Ablation study of adding different components of our proposed approach. The first row corresponds to the baseline SAMUS [33]." + ], + "table_footnote": [], + "table_body": "
MethodCAMUS-LATN3KBUSICAMUS-LVCAMUS-MYO
DiceHDDiceHDDiceHDDiceHDDiceHD
U-Net91.0012.9179.0134.1278.1133.6093.569.9086.8616.87
CPFNet91.5112.2679.4333.0780.5627.9893.329.6386.6816.51
CA-Net91.2812.3280.5233.6581.8828.6793.599.7787.2116.24
CE-Net91.1412.2980.3732.7980.2130.1993.369.9186.4716.66
AAU-Net91.3312.1282.2830.5380.8130.3993.329.9786.9816.49
SwinUNet89.8014.7470.0844.1367.2347.0291.7212.8084.4620.25
SETR90.5213.9167.8044.1168.2240.3792.8211.3486.2018.27
MISSFormer91.1811.8279.4232.8578.4333.1093.259.9486.5716.50
TransUNet91.3712.4681.4430.9882.2227.5493.609.6087.2017.25
FAT-Net91.5512.0580.4532.7782.1628.5593.599.2087.1915.93
H2Former90.9811.9282.4830.5881.4827.8493.449.6087.3116.60
SAMUS91.5811.6084.4528.2285.7725.4993.739.7987.4616.74
CC-SAM92.0311.1185.2027.1087.0124.2293.959.1188.2516.11
", + "bbox": [ + 230, + 508, + 772, + 719 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 2: Quantitative comparison between our CC-SAM method and the state-of-the-art (SOTA) task-specific techniques for segmenting thyroid nodules (TN3K), breast cancer (BUSI), left ventricle (CAMUS-LV), myocardium (CAMUS-MYO), and left atrium (CAMUS-LA). We assessed the performance using the Dice score (\\%) and the Hausdorff distance (HD). The top-performing results are highlighted in bold.", + "bbox": [ + 212, + 720, + 784, + 789 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Shreyank N Gowda and David A. Clifton", + "bbox": [ + 271, + 114, + 547, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/c27111f42dfcdcd089e6bd03ffb33716acc56e2b4237d47255da2b8cce2cc3cd.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparison between our CC-SAM method and the state-of-the-art (SOTA) task-specific techniques." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 782, + 321 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative Analysis. Table 2 summarizes the numerical outcomes of various task-specific methods on TN3K, BUSI, CAMUS-LV, CAMUS-MYO, and CAMUS-LA. In comparison, CC-SAM consistently excels in performance across all these five tasks, setting a new benchmark for SOTA results. While CC-SAM surpasses recent SOTA in most benchmarks, it does fall short by a very small margin when evaluated with the Hausdorff distance (HD) on CAMUS-MYO.", + "bbox": [ + 212, + 388, + 784, + 479 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Qualitative Analysis. Figure 4 showcases the segmentation results from various methods, including U-Net [45], AAU-Net [5], MISSFormer [28], H2Former [23], SAMUS [33], and our proposed CC-SAM. Segmenting ultrasound images is notably challenging given their low contrast, inconsistent features, and indistinct object boundaries. Many existing techniques face difficulties in precisely differentiating the target from its background, often resulting in significant false negatives or false positives. In contrast, SAMUS stands out in maintaining the integrity of target areas and minimizing false positives. Our CC-SAM further refines the outcomes observed with SAMUS, a testament to the intrinsic strengths of SAM and the tailored modifications and innovations integrated into CC-SAM.", + "bbox": [ + 212, + 491, + 787, + 642 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Generalization Ability. Evaluating the generalization performance of various task-specific methods is crucial as it provides insights into how well these methods can adapt and function across unfamiliar or unseen datasets. Figure 5 offers a quantitative assessment of this aspect. When we assess performance discrepancies between known (seeable) and unknown (unseen) datasets, CC-SAM stands out by displaying the most minimal performance drop among all methods across the three segmentation tasks. This underlines the remarkable generalization prowess of CC-SAM, emphasizing its resilience and versatility in a variety of medical image segmentation situations.", + "bbox": [ + 212, + 654, + 787, + 789 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Foundational models We also compare with five state-of-the-art foundational models. These include the original SAM [31], SAMed [53], SAMUS [33], MSA [51]", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Title Suppressed Due to Excessive Length", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e6c2172cc496fc37ea44f667a21ddae0b7d5fa77b2d207aa37c1f60a37059481.jpg", + "image_caption": [ + "Fig. 5: Comparison of CC-SAM with task-specific techniques on see datasets (highlighted in blue) and unseen datasets not previously encountered (indicated in orange). Higher orange bars indicate stronger generalization ability." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 316 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "and MedSAM [37]. SAM is trained on SA-1B and kept frozen due to the cost involved in fine-tuning. All other foundational models are pre-trained on the US30k dataset and evaluated using Dice and HD across TN3K, BUSI, CAMUS-LV, CAMUS-MYO, and CAMUS-LA datasets.", + "bbox": [ + 212, + 402, + 787, + 464 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/00b9173c6e9e8d926afb22a3923af20df3bad588a07d885ae4b7a6336f94b188.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTN3KBUSICAMUS-LVCAMUS-MYOCAMUS-LA
DiceHDDiceHDDiceHDDiceHDDiceHD
SAM29.59134.8754.0182.3928.18196.6429.42184.1017.28193.70
MedSAM71.0942.9177.7534.2687.5215.2876.0725.7288.0615.70
SAMed80.4031.2974.8234.6087.6713.2482.6019.4890.9212.60
MSA82.6729.1581.6628.8790.9511.2982.4719.2891.8011.59
SAMUS83.0528.8284.5427.2491.1311.7683.1118.9992.0012.08
CC-SAM85.5927.7486.2225.8592.8510.8885.6117.1193.5111.06
", + "bbox": [ + 230, + 503, + 772, + 619 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 3: Quantitative comparison of our CC-SAM and other foundation models on seeable US30K data. The performance is evaluated by the Dice score (\\%) and Hausdorff distance (HD).", + "bbox": [ + 212, + 619, + 787, + 660 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Quantitative Analysis. Table 3 summarizes the numerical outcomes of various foundational methods on TN3K, BUSI, CAMUS-LV, CAMUS-MYO, and CAMUS-LA. In comparison, CC-SAM consistently excels in performance across all tasks, setting a new benchmark for foundational models.", + "bbox": [ + 212, + 703, + 787, + 763 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative Analysis. Figure 6 shows the qualitative segmentation results from various foundational models, including SAM, MedSAM, SAMed, MSA, SAMUS, and CC-SAM. SAM, without adjustments for medical images, fails at segmentation. However, MedSAM, SAMed, and MSA, through tuning, recover some", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Shreyank N Gowda and David A. Clifton", + "bbox": [ + 271, + 114, + 547, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "segmentation ability but struggle with accurately defining borders, resulting in many false positives and negatives. In contrast, SAMUS excels in identifying precise segmentation borders, even in low contrast, suggesting the benefit of integrating local data with the image encoder for better boundary and shape retention in medical image segmentation. CC-SAM goes further, improving upon SAMUS's performance and setting new benchmarks.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/413bb74e2354fc947cc7cceeecef1c37197b7be244f7afd8d3182c98c435eaba.jpg", + "image_caption": [ + "Fig. 6: Qualitative comparison between our CC-SAM method and the state-of-the-art (SOTA) foundational medical segmentation models." + ], + "image_footnote": [], + "bbox": [ + 218, + 270, + 782, + 444 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Generalization Ability. Generally, foundation models trained on US30K exhibit a much improved generalization performance in medical image segmentation tasks compared to the original SAM. When considering the three specific segmentation tasks: thyroid nodule segmentation, breast cancer segmentation, and myocardium segmentation, it's clear that all foundation models face significant challenges with myocardium segmentation, while they perform relatively well in the breast cancer segmentation task. CC-SAM outperforms all foundational models significantly. This highlights the remarkable generalization prowess of CC-SAM, as it consistently surpasses other foundation models, especially in unfamiliar domains. Figure 7 offers a quantitative assessment of this aspect. We see that the original SAM performs the worst, which is expected and the proposed CC-SAM obtains a new state-of-the-art on all datasets.", + "bbox": [ + 212, + 550, + 787, + 733 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Deployment Cost Comparison with Other Foundational Models We compare CCSAM with other foundational models in a more detailed manner. We look at 'Generalization Ability/Dice $(\\%)$ ,GPU Memory/G, Parameters/M,Computation/GFLOPs', Inference Speed/FPS' and Segmentation Performance/Dice $(\\%)$ .The spider plot of this can be seen in Figure 8.", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Title Suppressed Due to Excessive Length", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/73ae6c8b6854707c443c0576ab63864cad3cfcc815e2d268f2848e3cb3894f50.jpg", + "image_caption": [ + "Fig.7: Comparing segmentation and generalization ability comparison of our CC-SAMUS and other foundation models. Whilst SAM obtains the lowest Dice Score, CC-SAM obtains the highest across all datasets." + ], + "image_footnote": [], + "bbox": [ + 220, + 155, + 480, + 305 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/b0fc5e0214f75dcdc52bb87a7fa5db9da34ae8e7cda733d60ee273229c44cc0d.jpg", + "image_caption": [ + "Fig. 8: Spider plot comparing efficiency and segmentation performance of all foundational models." + ], + "image_footnote": [], + "bbox": [ + 511, + 143, + 781, + 345 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 446, + 359, + 463 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The Segment Anything Model (SAM), while proficient in natural image segmentation, faced hurdles in medical imaging due to issues with low contrast, faint boundaries, and other complexities. In this study, we identified and addressed the limitations of the Segment Anything Model (SAM) when applied to medical imaging, especially with images that have inherent challenges like low contrast and intricate morphologies. Our enhanced approach integrates a CNN branch with SAM's original Vision Transformer (ViT) encoder, utilizing a unique variational attention fusion module to capture vital local spatial information present in medical images. By introducing feature and position adapters within the ViT branch, the model's representations are further refined. Notably, we found that using text prompts, especially those generated leveraging ChatGPT's capabilities, significantly boosts SAM's performance in ultrasound medical segmentation. These innovations mark a pivotal step towards improving the adaptability and efficiency of universal image segmentation models in the medical field.", + "bbox": [ + 212, + 493, + 787, + 705 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements DAC was supported by the Pandemic Sciences Institute at the University of Oxford; the National Institute for Health Research (NIHR) Oxford Biomedical Research Centre (BRC); an NIHR Research Professorship; a Royal Academy of Engineering Research Chair; the Wellcome Trust funded VITAL project (grant 204904/Z/16/Z); the EPSRC (grant EP/W031744/1); and the InnoHK Hong Kong Centre for Cerebro-cardiovascular Engineering (COCHE).", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Shreyank N Gowda and David A. Clifton", + "bbox": [ + 271, + 114, + 547, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Al-Dhabyani, W., Gomaa, M., Khaled, H., Aly, F.: Deep learning approaches for data augmentation and classification of breast masses using ultrasound images. Int. J. Adv. Comput. Sci. Appl 10(5), 1-11 (2019)", + "2. Alemi, A.A., Fischer, I., Dillon, J.V., Murphy, K.: Deep variational information bottleneck. arXiv preprint arXiv:1612.00410 (2016)", + "3. Byra, M., Jarosik, P., Szubert, A., Galperin, M., Ojeda-Fournier, H., Olson, L., O'Boyle, M., Comstock, C., Andre, M.: Breast mass segmentation in ultrasound with selective kernel u-net convolutional neural network. Biomedical Signal Processing and Control 61, 102027 (2020)", + "4. Cao, H., Wang, Y., Chen, J., Jiang, D., Zhang, X., Tian, Q., Wang, M.: Swinunet: Unet-like pure transformer for medical image segmentation. In: European conference on computer vision. pp. 205-218. Springer (2022)", + "5. Chen, G., Li, L., Dai, Y., Zhang, J., Yap, M.H.: Aau-net: an adaptive attention u-net for breast lesions segmentation in ultrasound images. IEEE Transactions on Medical Imaging (2022)", + "6. Chen, J., Lu, Y., Yu, Q., Luo, X., Adeli, E., Wang, Y., Lu, L., Yuille, A.L., Zhou, Y.: Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306 (2021)", + "7. Chen, L., Bentley, P., Mori, K., Misawa, K., Fujiwara, M., Rueckert, D.: Drinet for medical image segmentation. IEEE transactions on medical imaging 37(11), 2453-2462 (2018)", + "8. Degerli, A., Kiranyaz, S., Hamid, T., Mazhar, R., Gabbouj, M.: Early myocardial infarction detection over multi-view echocardiography. Biomedical Signal Processing and Control 87, 105448 (2024). https://doi.org/https://doi.org/10.1016/j.bspc.2023.105448", + "9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "0. Ermis, B., Zappella, G., Wistuba, M., Rawal, A., Archambeau, C.: Continual learning with transformers for image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3774-3781 (2022)", + "1. Feng, S., Zhao, H., Shi, F., Cheng, X., Wang, M., Ma, Y., Xiang, D., Zhu, W., Chen, X.: Cpfnet: Context pyramid fusion network for medical image segmentation. IEEE transactions on medical imaging 39(10), 3008-3018 (2020)", + "2. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision pp. 1-15 (2023)", + "3. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023)", + "4. Gong, H., Chen, J., Chen, G., Li, H., Li, G., Chen, F.: Thyroid region prior guided attention for ultrasound segmentation of thyroid nodules. Computers in Biology and Medicine 155, 106389 (2023)", + "5. Gowda, S.N.: Human activity recognition using combinatorial deep belief networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 1-6 (2017)" + ], + "bbox": [ + 225, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Title Suppressed Due to Excessive Length", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Gowda, S.N., Arnab, A., Huang, J.: Optimizing vivit training: Time and memory reduction for action recognition. arXiv preprint arXiv:2306.04822 (2023)", + "17. Gowda, S.N., Gao, B., Clifton, D.: Fe-adapter: adapting image-based emotion classifiers to videos (2024)", + "18. Gowda, S.N., Rohrbach, M., Sevilla-Lara, L.: Smart frame selection for action recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1451-1459 (2021)", + "19. Gowda, S.N., Yuan, C.: Colornet: Investigating the importance of color spaces for image classification. In: Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part IV 14. pp. 581–596. Springer (2019)", + "20. Gu, R., Wang, G., Song, T., Huang, R., Aertsen, M., Deprest, J., Ourselin, S., Vercauteren, T., Zhang, S.: Ca-net: Comprehensive attention convolutional neural networks for explainable medical image segmentation. IEEE transactions on medical imaging 40(2), 699-711 (2020)", + "21. Gu, Z., Cheng, J., Fu, H., Zhou, K., Hao, H., Zhao, Y., Zhang, T., Gao, S., Liu, J.: Ce-net: Context encoder network for 2d medical image segmentation. IEEE transactions on medical imaging 38(10), 2281-2292 (2019)", + "22. Hatamizadeh, A., Tang, Y., Nath, V., Yang, D., Myronenko, A., Landman, B., Roth, H.R., Xu, D.: Unetr: Transformers for 3d medical image segmentation. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 574-584 (2022)", + "23. He, A., Wang, K., Li, T., Du, C., Xia, S., Fu, H.: H2former: An efficient hierarchical hybrid transformer for medical image segmentation. IEEE Transactions on Medical Imaging (2023)", + "24. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "25. He, S., Bao, R., Li, J., Grant, P.E., Ou, Y.: Accuracy of segment-anything model (sam) in medical image segmentation tasks. arXiv preprint arXiv:2304.09324 (2023)", + "26. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019)", + "27. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017)", + "28. Huang, X., Deng, Z., Li, D., Yuan, X.: Missformer: An effective medical image segmentation transformer. arXiv preprint arXiv:2109.07162 (2021)", + "29. Huang, Y., Yang, X., Liu, L., Zhou, H., Chang, A., Zhou, X., Chen, R., Yu, J., Chen, J., Chen, C., et al.: Segment anything model for medical images? arXiv preprint arXiv:2304.14660 (2023)", + "30. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)", + "31. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)", + "32. Leclerc, S., Smistad, E., Pedrosa, J., Østvik, A., Cervenansky, F., Espinosa, F., Espeland, T., Berg, E.A.R., Jodoin, P.M., Grenier, T., et al.: Deep learning for segmentation using an open large-scale dataset in 2d echocardiography. IEEE transactions on medical imaging 38(9), 2198-2210 (2019)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Shreyank N Gowda and David A. Clifton", + "bbox": [ + 271, + 114, + 545, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Lin, X., Xiang, Y., Zhang, L., Yang, X., Yan, Z., Yu, L.: Samus: Adapting segment anything model for clinically-friendly and generalizable ultrasound image segmentation. arXiv preprint arXiv:2309.06824 (2023)", + "34. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)", + "35. Liu, X., Song, L., Liu, S., Zhang, Y.: A review of deep-learning-based medical image segmentation methods. Sustainability 13(3), 1224 (2021)", + "36. Luo, Z., Hu, Z., Xi, Y., Zhang, R., Ma, J.: I-tuning: Tuning frozen language models with image for lightweight image captioning. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)", + "37. Ma, J., Wang, B.: Segment anything in medical images. arXiv preprint arXiv:2304.12306 (2023)", + "38. Mei, X., Liu, Z., Robson, P.M., Marinelli, B., Huang, M., Doshi, A., Jacobi, A., Cao, C., Link, K.E., Yang, T., et al.: Radimagenet: an open radiologic deep learning research dataset for effective transfer learning. Radiology: Artificial Intelligence 4(5), e210315 (2022)", + "39. OpenAI: Gpt-4 technical report (2023)", + "40. Pan, J., Lin, Z., Zhu, X., Shao, J., Li, H.: St-adapter: Parameter-efficient image-to-video transfer learning. Advances in Neural Information Processing Systems 35, 26462-26477 (2022)", + "41. Pedraza, L., Vargas, C., Narváez, F., Durán, O., Muñoz, E., Romero, E.: An open access thyroid ultrasound image database. In: 10th International symposium on medical information processing and analysis. vol. 9287, pp. 188-193. SPIE (2015)", + "42. Pfeiffer, J., Rückle, A., Poth, C., Kamath, A., Vulic, I., Ruder, S., Cho, K., Gurevych, I.: Adapterhub: A framework for adapting transformers. arXiv preprint arXiv:2007.07779 (2020)", + "43. Pham, D.L., Xu, C., Prince, J.L.: Current methods in medical image segmentation. Annual review of biomedical engineering 2(1), 315-337 (2000)", + "44. Rasmy, L., Xiang, Y., Xie, Z., Tao, C., Zhi, D.: Med-bert: pretrained contextualized embeddings on large-scale structured electronic health records for disease prediction. NPJ digital medicine 4(1), 86 (2021)", + "45. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)", + "46. Roy, S., Wald, T., Koehler, G., Rokuss, M.R., Disch, N., Holzschuh, J., Zimmerer, D., Maier-Hein, K.H.: Sam. md: Zero-shot medical image segmentation capabilities of the segment anything model. arXiv preprint arXiv:2304.05396 (2023)", + "47. Sung, Y.L., Cho, J., Bansal, M.: Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5227-5237 (2022)", + "48. Valanarasu, J.M.J., Patel, V.M.: Unext: Mlp-based rapid medical image segmentation network. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 23-33. Springer (2022)", + "49. Wang, Y., Mukherjee, S., Liu, X., Gao, J., Awadallah, A.H., Gao, J.: Adamix: Mixture-of-adapter for parameter-efficient tuning of large language models. arXiv preprint arXiv:2205.12410 1(2), 4 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Title Suppressed Due to Excessive Length", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "50. Wu, H., Chen, S., Chen, G., Wang, W., Lei, B., Wen, Z.: Fat-net: Feature adaptive transformers for automated skin lesion segmentation. Medical image analysis 76, 102327 (2022)", + "51. Wu, J., Fu, R., Fang, H., Liu, Y., Wang, Z., Xu, Y., Jin, Y., Arbel, T.: Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620 (2023)", + "52. Wunderling, T., Golla, B., Poudel, P., Arens, C., Friebe, M., Hansen, C.: Comparison of thyroid segmentation techniques for 3d ultrasound. In: Medical Imaging 2017: Image Processing. vol. 10133, pp. 346-352. SPIE (2017)", + "53. Zhang, K., Liu, D.: Customized segment anything model for medical image segmentation. arXiv preprint arXiv:2304.13785 (2023)", + "54. Zhang, Y., Liu, H., Hu, Q.: Transfuse: Fusing transformers and cnns for medical image segmentation. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part I 24. pp. 14-24. Springer (2021)", + "55. Zheng, S., Lu, J., Zhao, H., Zhu, X., Luo, Z., Wang, Y., Fu, Y., Feng, J., Xiang, T., Torr, P.H., et al.: Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6881-6890 (2021)", + "56. Zhou, Z., Rahman Siddiquee, M.M., Tajbakhsh, N., Liang, J.: Unet++: A nested u-net architecture for medical image segmentation. In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 20, 2018, Proceedings 4. pp. 3-11. Springer (2018)" + ], + "bbox": [ + 215, + 146, + 787, + 494 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Shreyank N Gowda and David A. Clifton", + "bbox": [ + 271, + 114, + 547, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/CC-SAM_ Enhancing SAM with Cross-feature Attention and Context for Ultrasound Image Segmentation/4f25c9c8-80da-4658-ab8a-cd035c32acd6_model.json b/2024/CC-SAM_ Enhancing SAM with Cross-feature Attention and Context for Ultrasound Image Segmentation/4f25c9c8-80da-4658-ab8a-cd035c32acd6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6978f34e269a19f18820e4f2e08996d5d45075ab --- /dev/null +++ b/2024/CC-SAM_ Enhancing SAM with Cross-feature Attention and Context for Ultrasound Image Segmentation/4f25c9c8-80da-4658-ab8a-cd035c32acd6_model.json @@ -0,0 +1,2381 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.219, + 0.141, + 0.784, + 0.186 + ], + "angle": 0, + "content": "CC-SAM: SAM with Cross-feature Attention and Context for Ultrasound Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.213, + 0.662, + 0.228 + ], + "angle": 0, + "content": "Shreyank N Gowda1 and David A. Clifton1,2" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.24, + 0.782, + 0.255 + ], + "angle": 0, + "content": "\\(^{1}\\) Department of Engineering Sciences, University of Oxford, OX3 7DQ Oxford, UK" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.255, + 0.763, + 0.282 + ], + "angle": 0, + "content": "\\(^{2}\\) Oxford Suzhou Centre for Advanced Research, University of Oxford, Suzhou 215123, Jiangsu, China" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.324, + 0.74, + 0.655 + ], + "angle": 0, + "content": "Abstract. The Segment Anything Model (SAM) has achieved remarkable successes in the realm of natural image segmentation, but its deployment in the medical imaging sphere has encountered challenges. Specifically, the model struggles with medical images that feature low contrast, faint boundaries, intricate morphologies, and small-sized objects. To address these challenges and enhance SAM's performance in the medical domain, we introduce a comprehensive modification. Firstly, we incorporate a frozen Convolutional Neural Network (CNN) branch as an image encoder, which synergizes with SAM's original Vision Transformer (ViT) encoder through a novel variational attention fusion module. This integration bolsters the model's capability to capture local spatial information, which is often paramount in medical imagery. Moreover, to further optimize SAM for medical imaging, we introduce feature and position adapters within the ViT branch, refining the encoder's representations. We see that compared to current prompting strategies to fine-tune SAM for ultrasound medical segmentation, the use of text descriptions that serve as text prompts for SAM helps significantly improve the performance. Leveraging ChatGPT's natural language understanding capabilities, we generate prompts that offer contextual information and guidance to SAM, enabling it to better understand the nuances of ultrasound medical images and improve its segmentation accuracy. Our method, in its entirety, represents a significant stride towards making universal image segmentation models more adaptable and efficient in the medical domain." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.685, + 0.377, + 0.701 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Deep learning has revolutionized the fields of image [19, 24, 27], video classification [15, 18] and medical vision [45, 50, 56], enabling unprecedented accuracy and efficiency in these areas. Medical image segmentation plays a pivotal role in delineating and emphasizing particular organs, tissues, and anomalies within medical scans, forming a cornerstone of computer-aided diagnostic systems [35]. The development of numerous deep learning models for autonomous medical image segmentation has signaled a transformative shift in the potential of this technology [7, 45, 50, 56]. Nevertheless, these models are typically specialized for" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.549, + 0.129 + ], + "angle": 0, + "content": "Shreyank N Gowda and David A. Clifton" + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.145, + 0.77, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.337, + 0.757, + 0.352 + ], + "angle": 0, + "content": "Fig. 1: Comparison of methods using SAM for medical image segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.785, + 0.413 + ], + "angle": 0, + "content": "distinct tasks and require recalibration when used for different tasks, a limitation which poses significant challenges in medical data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.785, + 0.536 + ], + "angle": 0, + "content": "Segment anything model (SAM) [31] has emerged as a versatile foundational model for vision segmentation, gaining significant attention due to its unparalleled ability to segment a wide variety of objects and exhibit powerful zero-shot generalization. With user-driven prompts such as points, bounding boxes, and coarse masks, SAM dynamically segments the relevant objects. This adaptability positions SAM as an ideal tool for a myriad of segmentation use cases, facilitating the amalgamation of several discrete medical image segmentation tasks into a holistic model, enhancing its utility for clinical applications [29]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.787, + 0.672 + ], + "angle": 0, + "content": "However, while SAM boasts an extensive dataset (SA1B), its performance diminishes in the medical realm due to the lack of comprehensive clinical annotations [29]. To enhance SAM's performance in medical image segmentation, many approaches have been proposed such as tuning the mask decoder [37], applying the LoRA strategy to the image encoder [53], introducing task-specific information using adapters in the ViT image encoder [51]. Previous endeavors to adapt SAM for medical image segmentation by optimizing it on medical datasets have encountered limitations, given their tokenization approach that can obfuscate essential local data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.787, + 0.763 + ], + "angle": 0, + "content": "SAMUS [33] was subsequently introduced, aiming to transplant SAM's superior segmentation prowess and robust generalization to medical image segmentation, all while ensuring computational efficiency. Notable improvements within SAMUS encompass a redesigned ViT image encoder, integration of a parallel CNN-branch, and the introduction of a cross-branch attention mechanism, among others." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.842 + ], + "angle": 0, + "content": "In this paper, we introduce CC-SAM, a novel model built upon SAMUS. Notably, CC-SAM replaces the adjustable CNN with a fixed one, enhanced with adapters and employs a novel variational attention fusion instead of the cross-attention branch. Utilizing descriptions from Chat-GPT to prompt the model, we achieve a marked improvement in performance over existing prompting strategies" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Title Suppressed Due to Excessive Length" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "tailored for medical-oriented SAM. This new approach significantly improves medical-oriented SAM performance. See Figure 1 for a comparison with other medical SAM adaptations." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.193, + 0.687, + 0.208 + ], + "angle": 0, + "content": "The chief contributions of this paper are delineated as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.221, + 0.784, + 0.251 + ], + "angle": 0, + "content": "- The development of CC-SAM: a refined foundation model tailored for universal medical image segmentation, optimized for computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.253, + 0.784, + 0.281 + ], + "angle": 0, + "content": "- The introduction of a static CNN with adapters, complementing the ViT encoder of SAMUS and further reducing computational costs." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.283, + 0.784, + 0.311 + ], + "angle": 0, + "content": "- The innovative integration of variational attention fusion, enhancing the synergy between CNN and ViT branch features." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.313, + 0.784, + 0.356 + ], + "angle": 0, + "content": "- The use of text-based prompts, sourced from Chat-GPT, to articulate the segmentation problem, which significantly amplifies segmentation performance in medical contexts." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.221, + 0.784, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.383, + 0.387, + 0.399 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.416, + 0.501, + 0.431 + ], + "angle": 0, + "content": "2.1 Medical Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.784, + 0.516 + ], + "angle": 0, + "content": "Medical image segmentation is crucial for identifying and measuring structures in medical images. Before deep learning, methods like thresholding, clustering, and active contour models were common [43]. However, CNNs, particularly U-Net [45], have since dominated due to their quality results even with limited data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.784, + 0.577 + ], + "angle": 0, + "content": "Extensions of U-Net, such as DRINet [7], tackled class imbalance and high-resolution outputs. U-Net++ [56] used nested and dense skip pathways to improve segmentation, while UNeXt [48] optimized for speed and efficiency in medical segmentation by using a Convolutional MLP-based network." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.784, + 0.669 + ], + "angle": 0, + "content": "The rise of image transformers [9] led to new methods. FAT-Net [50] introduced feature adaptive transformers for detailed segmentation. UNeTR [22] combined transformers with a U-shaped design for 3D medical imaging. Our approach merges CNNs and transformers to harness both local and global feature extraction, creating a hybrid framework that leverages the strengths of both paradigms." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.693, + 0.517, + 0.708 + ], + "angle": 0, + "content": "2.2 Adapting Foundational Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.784, + 0.779 + ], + "angle": 0, + "content": "Foundational models, often pre-trained on large datasets, provide a generic understanding of complex patterns in data. These models are crucial stepping stones in many modern machine learning workflows. However, leveraging them effectively for specific tasks necessitates a thoughtful adaptation strategy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.781, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Adapter modules [26] offer a solution, allowing the main network weights to remain unchanged while adding learnable layers to each Transformer layer. This adds task-specific abilities with less overhead, ensuring efficient transfer learning. This approach keeps the base knowledge while adapting to new tasks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.549, + 0.129 + ], + "angle": 0, + "content": "Shreyank N Gowda and David A. Clifton" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.252 + ], + "angle": 0, + "content": "These modules have excelled in various language tasks [13, 36, 47]. For instance, they've achieved top results in question answering with \\(95\\%\\) fewer parameters than standard finetuning [42]. In textual entailment, AdaMix [49] found adapters performed comparably to full finetuning but with only \\(3\\%\\) added parameters. They've also been effective in computer vision tasks like image [10, 12] and video classification [16, 17, 40]. Essentially, they ensure efficient transfer learning across tasks by adding minimal layers to fixed networks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.787, + 0.283 + ], + "angle": 0, + "content": "We build on the idea of adapters and freeze both our CNN and ViT backbones and only use adapters for efficient fine-tuning of both models." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.301, + 0.553, + 0.318 + ], + "angle": 0, + "content": "2.3 Adapting SAM for Medical Images" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.323, + 0.785, + 0.367 + ], + "angle": 0, + "content": "The Segment Anything Model (SAM) [31] has demonstrated notable zero-shot segmentation using various input prompts. However, its performance on medical images is subpar [25, 46]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.368, + 0.788, + 0.457 + ], + "angle": 0, + "content": "Several strategies have been proposed to adapt SAM for medical imaging. MedSAM optimizes the mask decoder for efficient fine-tuning. SAMed [53] freezes the image encoder and employs a low-rank finetuning strategy while adjusting the prompt encoder and mask decoder. MSA [51] uses adapters in the ViT image encoder. To address transformer backbone limitations, SAMUS [33] incorporates a CNN-based backbone for improved local feature learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.458, + 0.788, + 0.535 + ], + "angle": 0, + "content": "Building on SAMUS [33], we optimize its framework. Instead of end-to-end CNN fine-tuning, we utilize a frozen pre-trained CNN and an adapter. We found that using text prompts, especially with Grounding DINO [34], significantly enhances SAMUS's performance. This method is not only more effective but also faster. We use GPT-4 [39] to generate text prompts from class labels." + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.559, + 0.776, + 0.759 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.767, + 0.788, + 0.811 + ], + "angle": 0, + "content": "Fig. 2: Overview of CC-SAM. We use adapters (or in the case of the CNN a FC layer) to enhance local and global features for ultrasound segmentation. 'OPaE' refers to overlapping patch embeddings and 'PE' refers to positional embeddings." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Title Suppressed Due to Excessive Length" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.427, + 0.163 + ], + "angle": 0, + "content": "3 Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.184, + 0.784, + 0.258 + ], + "angle": 0, + "content": "Figure 2 provides a detailed view of our method. Essentially, we process an image \\(I\\) using ViT-B [9] (pre-trained image encoder of SAM) and ResNet50 [24], the ResNet50 is pre-trained on RadImageNet [38]. After feature extraction, adapters refine these representations. Features are then fused using a variational attention mechanism, forming a combined feature for the Mask Decoder of SAM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.26, + 0.785, + 0.336 + ], + "angle": 0, + "content": "For prompts, we use a GPT-4 generated label description and embed it with Med-BERT [44]. The Grounding DINO model, untrained on medical data, creates a bounding box from this embedding, which is input to the Prompt Encoder. Together with the variational attention feature, the Prompt Encoder helps the Mask Decoder create the final segmentation mask." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.365, + 0.535, + 0.38 + ], + "angle": 0, + "content": "3.1 Frozen Backbones with Adapters" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.784, + 0.457 + ], + "angle": 0, + "content": "We've enhanced SAM's image encoder (the ViT branch) for better adaptability to smaller inputs and medical images by introducing a position adapter and five feature adapters, following SAMUS [33]. These adapters fine-tune the ViT branch efficiently with fewer parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.458, + 0.785, + 0.549 + ], + "angle": 0, + "content": "Specifically, the position adapter modifies positional embeddings to match the embedded sequence resolution. It first downsamples these embeddings using max pooling and then refines them with a convolution operation, enabling the ViT to better manage smaller inputs. Each of the five feature adapters follows a consistent design with a downward linear projection, an activation, and an upward projection. This is mathematically expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.429, + 0.569, + 0.785, + 0.584 + ], + "angle": 0, + "content": "\\[\nA (x) = G (x \\mathcal {M} _ {d}) \\mathcal {M} _ {u} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.784, + 0.642 + ], + "angle": 0, + "content": "Here, Here, \\(G\\) denotes the GELU activation function, while \\(\\mathcal{M}_d\\in \\mathbb{R}^{d\\times \\frac{d}{4}}\\) and \\(\\mathcal{M}_u\\in \\mathbb{R}^{d\\times \\frac{d}{4}}\\) are the projection matrices. In this context, \\(^{\\prime}d^{\\prime}\\) represents the dimension of the feature embedding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.784, + 0.704 + ], + "angle": 0, + "content": "Unlike SAMUS's end-to-end CNN training, we use a static, RadImageNet [38] pre-trained ResNet-50 model. We add a trainable fully-connected layer, acting as an adapter, before its classification layer. The output features from both branches then go into our variational attention fusion block." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.733, + 0.55, + 0.748 + ], + "angle": 0, + "content": "3.2 Variational Attention Fusion Block" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.84 + ], + "angle": 0, + "content": "A major contribution we've made is the Variational Attention Fusion Block, which adeptly merges local CNN features and global ViT features. While SAMUS introduced a cross-branch attention module, our method, which models uncertainty around these features and applies variational attentional fusion, outperforms it (details in Sec 4.3). See Figure 3 for an overview." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.549, + 0.129 + ], + "angle": 0, + "content": "Shreyank N Gowda and David A. Clifton" + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.152, + 0.71, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.404, + 0.789, + 0.475 + ], + "angle": 0, + "content": "Fig. 3: Overview of the proposed Variational Attention Fusion Block. Each 'mode' has an intra-modal uncertainty learning encoder (represented as \\(\\mathbf{E}\\) or \\(\\mathbf{E}^{\\prime}\\) in the figure), these obtain robust modality-specific features in the latent subspace. Subsequently, VAF combines these inputs and constructs a multimodal representation by estimating weights that are specific to each modality, effectively capturing their dependencies." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.505, + 0.789, + 0.613 + ], + "angle": 0, + "content": "Uncertainty Learning Incorporating uncertainty learning, for an input image \\(I\\), we derive features \\(s^v\\) from the ViT and \\(s^c\\) from the ResNet-50. Both \\(s^v\\) and \\(s^c \\in \\mathbb{R}^{d^h}\\) represent a fixed dimension in a latent space. We achieve a consistent size using latent encoders \\(E_v, E_c\\) at each network's end. To address data uncertainty, we design a latent distribution for each sample's content, helping capture semantic relations. This distribution is shaped as a parameterized diagonal Gaussian. This is mathematically represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.398, + 0.626, + 0.787, + 0.645 + ], + "angle": 0, + "content": "\\[\np \\left(z ^ {v} \\mid s ^ {v}\\right) = \\mathcal {N} \\left(z ^ {v}; \\mu^ {v}, \\left(\\sigma^ {v}\\right) ^ {2} \\mathbf {I}\\right) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.401, + 0.664, + 0.787, + 0.683 + ], + "angle": 0, + "content": "\\[\np \\left(z ^ {c} \\mid s ^ {c}\\right) = \\mathcal {N} \\left(z ^ {c}; \\mu^ {c}, \\left(\\sigma^ {c}\\right) ^ {2} \\mathbf {I}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.689, + 0.789, + 0.81 + ], + "angle": 0, + "content": "Here, \\( z^v \\) and \\( z^c \\) are reconstructed vectors. The means, \\( \\mu^v, \\mu^c \\), capture each mode's core feature, while variances, \\( \\sigma^v, \\sigma^c \\), indicate the noise-induced uncertainty in predicting these means. A higher variance signifies more uncertainty about the content observed. Both Gaussian parameters depend on the input and are forecasted through MLPs. For instance, for ViT features: \\( \\mu^v = f\\theta_{v1}(sv) \\) and \\( \\sigma^v = f\\theta_{v2}(sv) \\), with \\( \\theta_{v1} \\) and \\( \\theta_{v2} \\) being respective model parameters. The same applies to CNN features. Now, each sample's feature representation shifts from deterministic to a stochastic Gaussian-drawn embedding in the latent space." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Due to the non-differentiable nature of sampling, we use a re-parameterization trick [30] to maintain gradient flow. We draw random noise \\(\\eta\\) from a normal dis" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Title Suppressed Due to Excessive Length" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.786, + 0.178 + ], + "angle": 0, + "content": "ribution, independent of model parameters, and produce \\( z^v \\) using it, as detailed in Eq 4. We repeat this for \\( z^c \\)." + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.192, + 0.786, + 0.21 + ], + "angle": 0, + "content": "\\[\nz ^ {v} = \\mu^ {v} + \\eta \\sigma^ {v}, \\eta \\in \\mathcal {N} (0, 1) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.229, + 0.784, + 0.334 + ], + "angle": 0, + "content": "Variational Attentional Fusion CNNs and ViTs capture different feature aspects from images (local and global). We see these as two modes: \\( v \\) for ViTs and \\( c \\) for CNNs. Combining them isn't straightforward due to different confidence levels. Standard methods, which get weights from the data, miss the unique traits of each mode. To tackle this, we've developed a Variational Attention Fusion (VAF) module. This module aims to capture the nuances between modalities by determining modality-specific weights for a seamless integration." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.335, + 0.784, + 0.409 + ], + "angle": 0, + "content": "In short, using features from each mode, a standard attention method creates a probability, \\(a^k\\) (where \\(k\\) is either \\(v\\) or \\(c\\)), as shown in Eq 5. Here, \\(a^k\\) shows how much each mode contributes, and \\(\\mathbf{W}_m\\) and \\(b_m\\) are trainable factors. By having set weights, it picks the right features without being affected by varied confidence levels between modes." + }, + { + "type": "equation", + "bbox": [ + 0.462, + 0.434, + 0.59, + 0.45 + ], + "angle": 0, + "content": "\\[\n\\hat {a} ^ {k} = \\mathbf {W} _ {m} \\mathbf {z} ^ {k} + b _ {m}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.454, + 0.785, + 0.49 + ], + "angle": 0, + "content": "\\[\na ^ {k} = \\frac {\\exp \\left(\\hat {a} ^ {k}\\right)}{\\sum_ {k ^ {\\prime} \\in \\{v , c \\}} \\exp \\left(\\hat {a} ^ {k ^ {\\prime}}\\right)} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.499, + 0.784, + 0.545 + ], + "angle": 0, + "content": "Our VAF module uses variational attention weights instead of the point-estimated attention vector \\(\\hat{a}^k\\). This approach, grounded in a probability distribution (Eq. 6), better accounts for uncertainties across modalities." + }, + { + "type": "equation", + "bbox": [ + 0.337, + 0.555, + 0.785, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\hat {a} ^ {k} \\sim q _ {\\theta} (\\hat {a} ^ {k} \\mid z ^ {k}) = \\mathcal {N} \\left(\\mu_ {a} ^ {k}, \\left(\\sigma_ {a} ^ {k}\\right) ^ {2} \\mathbf {I}\\right), k \\in \\{v, c \\} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.585, + 0.786, + 0.705 + ], + "angle": 0, + "content": "We used an identity transformation to preserve modality traits in the variational attention. Mean and variance parameters are predicted by MLPs from the input. The \\(\\left(\\sigma_{a}^{k}\\right)\\) value indicates confidence levels among modalities. With the VAF method, more confident modality features are enhanced, while less confident ones are reduced, optimizing multimodal data fusion and capturing complementary features. The final representation combines modality-specific outputs, \\(z_{c}\\) and \\(z_{v}\\), using weighted aggregation as shown in Eq 7, where \\(\\mathbf{W}_h\\) is a learnable weight matrix." + }, + { + "type": "equation", + "bbox": [ + 0.431, + 0.718, + 0.786, + 0.751 + ], + "angle": 0, + "content": "\\[\nh = \\sum_ {k \\in \\{v, c \\}} a ^ {k} \\mathbf {W} _ {h} z ^ {k} \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.771, + 0.505, + 0.787 + ], + "angle": 0, + "content": "3.3 Guiding the Prompt Encoder" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.784, + 0.841 + ], + "angle": 0, + "content": "SAM's mask decoder [31] needs \\( h \\) and an input from the prompt encoder. While past adaptations of SAM for medical imaging used point prompts, our results show that a reasonably good bounding box enhances performance. Creating this" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.549, + 0.129 + ], + "angle": 0, + "content": "Shreyank N Gowda and David A. Clifton" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.282 + ], + "angle": 0, + "content": "bounding box requires a dedicated model. Using random bounding boxes may hinder results, necessitating specific prompts. We employ Grounding-DINO [34] for bounding box generation. Though not exclusively trained for medical images, Grounding-DINO is an effective object detector. We use GPT-4 [39] to craft text prompts for specific class labels, and then MedBERT [44] generates an embedding input for Grounding-DINO. Whilst this does improve our performance in comparison to using point prompts, we show in supplementary that using the same set of random point prompts, we outperform all other medical foundational models." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.284, + 0.79, + 0.45 + ], + "angle": 0, + "content": "Input to GPT-4 \"Create a concise description for a medical image analysis task. The task involves using ultrasound images to segment and identify specific anatomical structures or pathologies. Use [target] for to make a generic description that can be adapted for any segmentation target.\" Output from GPT-4 \"Segment and identify [target] within ultrasound images. This requires the precise delineation of [target] from surrounding tissues, enabling accurate diagnosis and assessment. The challenge encompasses dealing with the inherent variability of ultrasound image quality, including speckle noise, shadowing, and the wide range of anatomical variations among patients. Success in this task is measured by the algorithm's ability to consistently and accurately identify and outline the [target], providing crucial information for medical decision-making processes.\" Where [target] is modified for each dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.467, + 0.379, + 0.48 + ], + "angle": 0, + "content": "3.4 Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.488, + 0.789, + 0.52 + ], + "angle": 0, + "content": "Our overall loss function can be seen in Eq 8. Here, \\(\\mathcal{L}_{BCE}\\) and \\(\\mathcal{L}_D\\) are segmentation losses: binary cross entropy and dice loss respectively." + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.533, + 0.785, + 0.549 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {B C E} + \\mathcal {L} _ {D} + \\lambda_ {c} \\mathcal {L} _ {c} + \\lambda_ {v} \\mathcal {L} _ {v} + \\lambda_ {a} \\mathcal {L} _ {a} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.68 + ], + "angle": 0, + "content": "\\(\\mathcal{L}c\\) and \\(\\mathcal{L}v\\) are regularization terms for uncertainty learning, inspired by the variational information bottleneck [2]. To grasp uncertainty in a modality, a component is introduced during learning. It ensures the data's distribution, \\(\\mathcal{N}\\left(z^v;\\mu^v,(\\sigma^v)^2\\right)\\), resembles a standard bell curve, \\(\\mathcal{N}(\\epsilon ;0,\\mathbf{I})\\). This similarity is gauged using Kullback-Leibler divergence (KLD) between the distributions, promoting model diversity and reducing uncertainty, leading to robust features. Eq 9 shows how \\(\\mathcal{L}v\\) is computed, with analogous methods for \\(\\mathcal{L}c\\) and \\(\\mathcal{L}_a\\), replacing \\(z^{v}\\) with \\(z^c\\) and \\(\\hat{a}^k\\)." + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.691, + 0.785, + 0.747 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {v} = K L \\left(\\mathcal {N} \\left(z ^ {v}; \\mu^ {v}, (\\sigma^ {v}) ^ {2}\\right) \\| \\mathcal {N} (\\epsilon ; 0, \\mathbf {I})\\right) \\\\ = - \\frac {1}{2} \\left(1 + \\log (\\sigma^ {v}) ^ {2} - (\\mu^ {v}) ^ {2} - (\\sigma^ {v}) ^ {2}\\right) \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.762, + 0.473, + 0.78 + ], + "angle": 0, + "content": "4 Experimental Analysis" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.79, + 0.337, + 0.804 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "To compare with SAMUS [33], we test CC-SAM on the same seven public datasets: TN3K [14], DDTI [41], TG3K [52], BUSI [1], UDIAT [3], CAMUS [32]," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Title Suppressed Due to Excessive Length" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.177 + ], + "angle": 0, + "content": "and HMC-QU [8]. We also recreate the US30K dataset from SAMUS for broader comparison. Dataset details are in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.784, + 0.237 + ], + "angle": 0, + "content": "TN3K and TG3K datasets follow TRFE [14] for segmentation into training, validation, and testing. BUSI is randomly split in a 7:1:2 ratio for these purposes. CAMUS is initially divided into training and testing per the challenge [32], with \\(10\\%\\) of training data later used for validation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.785, + 0.343 + ], + "angle": 0, + "content": "For model generalizability, US30K datasets remain unseen during training and validation. In line with SAMUS, we split US30k into seen and unseen datasets. We then evaluate on unseen datasets (DDTI, UDIAT, HMC-QU) not used in training or validation. This enables direct comparison of generalization with other state-of-the-art models. Moreover, we contrast CC-SAM with other base models by training on the full US30K dataset and testing across different tasks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.371, + 0.458, + 0.387 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.784, + 0.489 + ], + "angle": 0, + "content": "We follow SAMUS in setting hyper parameters and adapter sizes. We change the fully tuneable CNN in SAMUS to a static ResNet50. The model is trained using the Adam optimizer with an initial learning rate of 0.01 that reduces by a tenth every 50 epochs for a total of 200 epochs. The encoders are fully connected networks with the latent space having dimensions 1024. All fully connected layers has a Leaky RELU activation function and a dropout probability of 0.5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.785, + 0.534 + ], + "angle": 0, + "content": "We use 1xA100 GPU with 40 GB capacity. But we also verified that a single 3090 Ti with 24 GB memory is enough to train the model. We set a batch size of 32." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.561, + 0.39, + 0.577 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.589, + 0.784, + 0.709 + ], + "angle": 0, + "content": "We consider SAMUS [33] as our baseline model and ablate our added components. Specifically, we show replacing the end-to-end CNN model with a fixed pre-trained model improves performance. Further adding the FC layer at the end slightly improves performance. Crucially, replacing the cross attention fusion with our variational attention fusion significantly improves performance. Further, replacing the point based prompting with our text-based prompting gives us the best results. We show these results on two of the datasets TN3K and BUSI in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.711, + 0.784, + 0.741 + ], + "angle": 0, + "content": "We do a further analysis of how the prompting compares with different prompting strategies in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.767, + 0.534, + 0.782 + ], + "angle": 0, + "content": "4.4 Comparison with state-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.839 + ], + "angle": 0, + "content": "We consider two scenarios for comparing with state-of-the-art models following SAMUS [33]. We compare CC-SAM with task-specific methods and foundational models." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.549, + 0.129 + ], + "angle": 0, + "content": "Shreyank N Gowda and David A. Clifton" + }, + { + "type": "table", + "bbox": [ + 0.315, + 0.144, + 0.689, + 0.285 + ], + "angle": 0, + "content": "
ComponentsTN3KBUSI
ResNet50VAFText PromptDiceHDDiceHD
×××84.4528.2285.7725.49
✓××84.9127.2586.2225.12
×✓×85.5526.8487.1524.42
××✓85.9526.2287.9223.92
✓×✓86.4225.8288.4523.20
×✓✓86.2525.9288.2223.51
✓✓×85.7226.2187.4224.61
✓✓✓87.1125.2589.5122.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.286, + 0.784, + 0.313 + ], + "angle": 0, + "content": "Table 1: Ablation study of adding different components of our proposed approach. The first row corresponds to the baseline SAMUS [33]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.366, + 0.785, + 0.473 + ], + "angle": 0, + "content": "Task-specific methods Following SAMUS [33], thirteen state-of-the-art, task-specific techniques are chosen for evaluation, spanning CNN-based, transformer-based, and CNN-Transformer hybrid methods. Methods other than SAMUS include CNN-based techniques encompass U-Net [45], CPFNet [11], CA-Net [20], CE-Net [21], and AAU-Net [5]. The transformer-centric methods consist of SwinUnet [4], SETR [55], and MISSFormer [28]. The hybrids of CNN and Transformer are TransUNet [6], TransFuse [54], FAT-Net [50], and H2Former [23]." + }, + { + "type": "table", + "bbox": [ + 0.231, + 0.51, + 0.773, + 0.72 + ], + "angle": 0, + "content": "
MethodCAMUS-LATN3KBUSICAMUS-LVCAMUS-MYO
DiceHDDiceHDDiceHDDiceHDDiceHD
U-Net91.0012.9179.0134.1278.1133.6093.569.9086.8616.87
CPFNet91.5112.2679.4333.0780.5627.9893.329.6386.6816.51
CA-Net91.2812.3280.5233.6581.8828.6793.599.7787.2116.24
CE-Net91.1412.2980.3732.7980.2130.1993.369.9186.4716.66
AAU-Net91.3312.1282.2830.5380.8130.3993.329.9786.9816.49
SwinUNet89.8014.7470.0844.1367.2347.0291.7212.8084.4620.25
SETR90.5213.9167.8044.1168.2240.3792.8211.3486.2018.27
MISSFormer91.1811.8279.4232.8578.4333.1093.259.9486.5716.50
TransUNet91.3712.4681.4430.9882.2227.5493.609.6087.2017.25
FAT-Net91.5512.0580.4532.7782.1628.5593.599.2087.1915.93
H2Former90.9811.9282.4830.5881.4827.8493.449.6087.3116.60
SAMUS91.5811.6084.4528.2285.7725.4993.739.7987.4616.74
CC-SAM92.0311.1185.2027.1087.0124.2293.959.1188.2516.11
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.721, + 0.785, + 0.79 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison between our CC-SAM method and the state-of-the-art (SOTA) task-specific techniques for segmenting thyroid nodules (TN3K), breast cancer (BUSI), left ventricle (CAMUS-LV), myocardium (CAMUS-MYO), and left atrium (CAMUS-LA). We assessed the performance using the Dice score (\\%) and the Hausdorff distance (HD). The top-performing results are highlighted in bold." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Title Suppressed Due to Excessive Length" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.145, + 0.784, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.334, + 0.785, + 0.362 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparison between our CC-SAM method and the state-of-the-art (SOTA) task-specific techniques." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.785, + 0.48 + ], + "angle": 0, + "content": "Quantitative Analysis. Table 2 summarizes the numerical outcomes of various task-specific methods on TN3K, BUSI, CAMUS-LV, CAMUS-MYO, and CAMUS-LA. In comparison, CC-SAM consistently excels in performance across all these five tasks, setting a new benchmark for SOTA results. While CC-SAM surpasses recent SOTA in most benchmarks, it does fall short by a very small margin when evaluated with the Hausdorff distance (HD) on CAMUS-MYO." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.643 + ], + "angle": 0, + "content": "Qualitative Analysis. Figure 4 showcases the segmentation results from various methods, including U-Net [45], AAU-Net [5], MISSFormer [28], H2Former [23], SAMUS [33], and our proposed CC-SAM. Segmenting ultrasound images is notably challenging given their low contrast, inconsistent features, and indistinct object boundaries. Many existing techniques face difficulties in precisely differentiating the target from its background, often resulting in significant false negatives or false positives. In contrast, SAMUS stands out in maintaining the integrity of target areas and minimizing false positives. Our CC-SAM further refines the outcomes observed with SAMUS, a testament to the intrinsic strengths of SAM and the tailored modifications and innovations integrated into CC-SAM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.655, + 0.788, + 0.79 + ], + "angle": 0, + "content": "Generalization Ability. Evaluating the generalization performance of various task-specific methods is crucial as it provides insights into how well these methods can adapt and function across unfamiliar or unseen datasets. Figure 5 offers a quantitative assessment of this aspect. When we assess performance discrepancies between known (seeable) and unknown (unseen) datasets, CC-SAM stands out by displaying the most minimal performance drop among all methods across the three segmentation tasks. This underlines the remarkable generalization prowess of CC-SAM, emphasizing its resilience and versatility in a variety of medical image segmentation situations." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Foundational models We also compare with five state-of-the-art foundational models. These include the original SAM [31], SAMed [53], SAMUS [33], MSA [51]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.549, + 0.129 + ], + "angle": 0, + "content": "Shreyank N Gowda and David A. Clifton" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Fig. 5: Comparison of CC-SAM with task-specific techniques on see datasets (highlighted in blue) and unseen datasets not previously encountered (indicated in orange). Higher orange bars indicate stronger generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.789, + 0.465 + ], + "angle": 0, + "content": "and MedSAM [37]. SAM is trained on SA-1B and kept frozen due to the cost involved in fine-tuning. All other foundational models are pre-trained on the US30k dataset and evaluated using Dice and HD across TN3K, BUSI, CAMUS-LV, CAMUS-MYO, and CAMUS-LA datasets." + }, + { + "type": "table", + "bbox": [ + 0.232, + 0.505, + 0.773, + 0.62 + ], + "angle": 0, + "content": "
MethodTN3KBUSICAMUS-LVCAMUS-MYOCAMUS-LA
DiceHDDiceHDDiceHDDiceHDDiceHD
SAM29.59134.8754.0182.3928.18196.6429.42184.1017.28193.70
MedSAM71.0942.9177.7534.2687.5215.2876.0725.7288.0615.70
SAMed80.4031.2974.8234.6087.6713.2482.6019.4890.9212.60
MSA82.6729.1581.6628.8790.9511.2982.4719.2891.8011.59
SAMUS83.0528.8284.5427.2491.1311.7683.1118.9992.0012.08
CC-SAM85.5927.7486.2225.8592.8510.8885.6117.1193.5111.06
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.62, + 0.788, + 0.661 + ], + "angle": 0, + "content": "Table 3: Quantitative comparison of our CC-SAM and other foundation models on seeable US30K data. The performance is evaluated by the Dice score (\\%) and Hausdorff distance (HD)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Quantitative Analysis. Table 3 summarizes the numerical outcomes of various foundational methods on TN3K, BUSI, CAMUS-LV, CAMUS-MYO, and CAMUS-LA. In comparison, CC-SAM consistently excels in performance across all tasks, setting a new benchmark for foundational models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Qualitative Analysis. Figure 6 shows the qualitative segmentation results from various foundational models, including SAM, MedSAM, SAMed, MSA, SAMUS, and CC-SAM. SAM, without adjustments for medical images, fails at segmentation. However, MedSAM, SAMed, and MSA, through tuning, recover some" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Title Suppressed Due to Excessive Length" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "segmentation ability but struggle with accurately defining borders, resulting in many false positives and negatives. In contrast, SAMUS excels in identifying precise segmentation borders, even in low contrast, suggesting the benefit of integrating local data with the image encoder for better boundary and shape retention in medical image segmentation. CC-SAM goes further, improving upon SAMUS's performance and setting new benchmarks." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.271, + 0.784, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.458, + 0.785, + 0.486 + ], + "angle": 0, + "content": "Fig. 6: Qualitative comparison between our CC-SAM method and the state-of-the-art (SOTA) foundational medical segmentation models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Generalization Ability. Generally, foundation models trained on US30K exhibit a much improved generalization performance in medical image segmentation tasks compared to the original SAM. When considering the three specific segmentation tasks: thyroid nodule segmentation, breast cancer segmentation, and myocardium segmentation, it's clear that all foundation models face significant challenges with myocardium segmentation, while they perform relatively well in the breast cancer segmentation task. CC-SAM outperforms all foundational models significantly. This highlights the remarkable generalization prowess of CC-SAM, as it consistently surpasses other foundation models, especially in unfamiliar domains. Figure 7 offers a quantitative assessment of this aspect. We see that the original SAM performs the worst, which is expected and the proposed CC-SAM obtains a new state-of-the-art on all datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Deployment Cost Comparison with Other Foundational Models We compare CCSAM with other foundational models in a more detailed manner. We look at 'Generalization Ability/Dice \\((\\%)\\) ,GPU Memory/G, Parameters/M,Computation/GFLOPs', Inference Speed/FPS' and Segmentation Performance/Dice \\((\\%)\\) .The spider plot of this can be seen in Figure 8." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.549, + 0.129 + ], + "angle": 0, + "content": "Shreyank N Gowda and David A. Clifton" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.156, + 0.481, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.321, + 0.488, + 0.404 + ], + "angle": 0, + "content": "Fig.7: Comparing segmentation and generalization ability comparison of our CC-SAMUS and other foundation models. Whilst SAM obtains the lowest Dice Score, CC-SAM obtains the highest across all datasets." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.145, + 0.782, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.371, + 0.788, + 0.412 + ], + "angle": 0, + "content": "Fig. 8: Spider plot comparing efficiency and segmentation performance of all foundational models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.448, + 0.36, + 0.464 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.789, + 0.706 + ], + "angle": 0, + "content": "The Segment Anything Model (SAM), while proficient in natural image segmentation, faced hurdles in medical imaging due to issues with low contrast, faint boundaries, and other complexities. In this study, we identified and addressed the limitations of the Segment Anything Model (SAM) when applied to medical imaging, especially with images that have inherent challenges like low contrast and intricate morphologies. Our enhanced approach integrates a CNN branch with SAM's original Vision Transformer (ViT) encoder, utilizing a unique variational attention fusion module to capture vital local spatial information present in medical images. By introducing feature and position adapters within the ViT branch, the model's representations are further refined. Notably, we found that using text prompts, especially those generated leveraging ChatGPT's capabilities, significantly boosts SAM's performance in ultrasound medical segmentation. These innovations mark a pivotal step towards improving the adaptability and efficiency of universal image segmentation models in the medical field." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Acknowledgements DAC was supported by the Pandemic Sciences Institute at the University of Oxford; the National Institute for Health Research (NIHR) Oxford Biomedical Research Centre (BRC); an NIHR Research Professorship; a Royal Academy of Engineering Research Chair; the Wellcome Trust funded VITAL project (grant 204904/Z/16/Z); the EPSRC (grant EP/W031744/1); and the InnoHK Hong Kong Centre for Cerebro-cardiovascular Engineering (COCHE)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Title Suppressed Due to Excessive Length" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.221 + ], + "angle": 0, + "content": "1. Al-Dhabyani, W., Gomaa, M., Khaled, H., Aly, F.: Deep learning approaches for data augmentation and classification of breast masses using ultrasound images. Int. J. Adv. Comput. Sci. Appl 10(5), 1-11 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.223, + 0.785, + 0.25 + ], + "angle": 0, + "content": "2. Alemi, A.A., Fischer, I., Dillon, J.V., Murphy, K.: Deep variational information bottleneck. arXiv preprint arXiv:1612.00410 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.252, + 0.785, + 0.306 + ], + "angle": 0, + "content": "3. Byra, M., Jarosik, P., Szubert, A., Galperin, M., Ojeda-Fournier, H., Olson, L., O'Boyle, M., Comstock, C., Andre, M.: Breast mass segmentation in ultrasound with selective kernel u-net convolutional neural network. Biomedical Signal Processing and Control 61, 102027 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.308, + 0.785, + 0.348 + ], + "angle": 0, + "content": "4. Cao, H., Wang, Y., Chen, J., Jiang, D., Zhang, X., Tian, Q., Wang, M.: Swinunet: Unet-like pure transformer for medical image segmentation. In: European conference on computer vision. pp. 205-218. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.35, + 0.785, + 0.39 + ], + "angle": 0, + "content": "5. Chen, G., Li, L., Dai, Y., Zhang, J., Yap, M.H.: Aau-net: an adaptive attention u-net for breast lesions segmentation in ultrasound images. IEEE Transactions on Medical Imaging (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.392, + 0.785, + 0.433 + ], + "angle": 0, + "content": "6. Chen, J., Lu, Y., Yu, Q., Luo, X., Adeli, E., Wang, Y., Lu, L., Yuille, A.L., Zhou, Y.: Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.435, + 0.785, + 0.473 + ], + "angle": 0, + "content": "7. Chen, L., Bentley, P., Mori, K., Misawa, K., Fujiwara, M., Rueckert, D.: Drinet for medical image segmentation. IEEE transactions on medical imaging 37(11), 2453-2462 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.476, + 0.785, + 0.53 + ], + "angle": 0, + "content": "8. Degerli, A., Kiranyaz, S., Hamid, T., Mazhar, R., Gabbouj, M.: Early myocardial infarction detection over multi-view echocardiography. Biomedical Signal Processing and Control 87, 105448 (2024). https://doi.org/https://doi.org/10.1016/j.bspc.2023.105448" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.532, + 0.785, + 0.587 + ], + "angle": 0, + "content": "9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.589, + 0.785, + 0.629 + ], + "angle": 0, + "content": "0. Ermis, B., Zappella, G., Wistuba, M., Rawal, A., Archambeau, C.: Continual learning with transformers for image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3774-3781 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.631, + 0.785, + 0.671 + ], + "angle": 0, + "content": "1. Feng, S., Zhao, H., Shi, F., Cheng, X., Wang, M., Ma, Y., Xiang, D., Zhu, W., Chen, X.: Cpfnet: Context pyramid fusion network for medical image segmentation. IEEE transactions on medical imaging 39(10), 3008-3018 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.785, + 0.713 + ], + "angle": 0, + "content": "2. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision pp. 1-15 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.715, + 0.785, + 0.756 + ], + "angle": 0, + "content": "3. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.798 + ], + "angle": 0, + "content": "4. Gong, H., Chen, J., Chen, G., Li, H., Li, G., Chen, F.: Thyroid region prior guided attention for ultrasound segmentation of thyroid nodules. Computers in Biology and Medicine 155, 106389 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Gowda, S.N.: Human activity recognition using combinatorial deep belief networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 1-6 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.546, + 0.128 + ], + "angle": 0, + "content": "Shreyank N Gowda and David A. Clifton" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Gowda, S.N., Arnab, A., Huang, J.: Optimizing vivit training: Time and memory reduction for action recognition. arXiv preprint arXiv:2306.04822 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "17. Gowda, S.N., Gao, B., Clifton, D.: Fe-adapter: adapting image-based emotion classifiers to videos (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.785, + 0.242 + ], + "angle": 0, + "content": "18. Gowda, S.N., Rohrbach, M., Sevilla-Lara, L.: Smart frame selection for action recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1451-1459 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.785, + 0.297 + ], + "angle": 0, + "content": "19. Gowda, S.N., Yuan, C.: Colornet: Investigating the importance of color spaces for image classification. In: Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part IV 14. pp. 581–596. Springer (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.352 + ], + "angle": 0, + "content": "20. Gu, R., Wang, G., Song, T., Huang, R., Aertsen, M., Deprest, J., Ourselin, S., Vercauteren, T., Zhang, S.: Ca-net: Comprehensive attention convolutional neural networks for explainable medical image segmentation. IEEE transactions on medical imaging 40(2), 699-711 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.353, + 0.785, + 0.393 + ], + "angle": 0, + "content": "21. Gu, Z., Cheng, J., Fu, H., Zhou, K., Hao, H., Zhao, Y., Zhang, T., Gao, S., Liu, J.: Ce-net: Context encoder network for 2d medical image segmentation. IEEE transactions on medical imaging 38(10), 2281-2292 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.394, + 0.785, + 0.447 + ], + "angle": 0, + "content": "22. Hatamizadeh, A., Tang, Y., Nath, V., Yang, D., Myronenko, A., Landman, B., Roth, H.R., Xu, D.: Unetr: Transformers for 3d medical image segmentation. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 574-584 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.448, + 0.785, + 0.488 + ], + "angle": 0, + "content": "23. He, A., Wang, K., Li, T., Du, C., Xia, S., Fu, H.: H2former: An efficient hierarchical hybrid transformer for medical image segmentation. IEEE Transactions on Medical Imaging (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.489, + 0.785, + 0.529 + ], + "angle": 0, + "content": "24. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.785, + 0.569 + ], + "angle": 0, + "content": "25. He, S., Bao, R., Li, J., Grant, P.E., Ou, Y.: Accuracy of segment-anything model (sam) in medical image segmentation tasks. arXiv preprint arXiv:2304.09324 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.61 + ], + "angle": 0, + "content": "26. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.611, + 0.785, + 0.651 + ], + "angle": 0, + "content": "27. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.652, + 0.785, + 0.677 + ], + "angle": 0, + "content": "28. Huang, X., Deng, Z., Li, D., Yuan, X.: Missformer: An effective medical image segmentation transformer. arXiv preprint arXiv:2109.07162 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.718 + ], + "angle": 0, + "content": "29. Huang, Y., Yang, X., Liu, L., Zhou, H., Chang, A., Zhou, X., Chen, R., Yu, J., Chen, J., Chen, C., et al.: Segment anything model for medical images? arXiv preprint arXiv:2304.14660 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.719, + 0.785, + 0.745 + ], + "angle": 0, + "content": "30. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.785, + 0.786 + ], + "angle": 0, + "content": "31. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "32. Leclerc, S., Smistad, E., Pedrosa, J., Østvik, A., Cervenansky, F., Espinosa, F., Espeland, T., Berg, E.A.R., Jodoin, P.M., Grenier, T., et al.: Deep learning for segmentation using an open large-scale dataset in 2d echocardiography. IEEE transactions on medical imaging 38(9), 2198-2210 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Title Suppressed Due to Excessive Length" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "33. Lin, X., Xiang, Y., Zhang, L., Yang, X., Yan, Z., Yu, L.: Samus: Adapting segment anything model for clinically-friendly and generalizable ultrasound image segmentation. arXiv preprint arXiv:2309.06824 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.233 + ], + "angle": 0, + "content": "34. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.234, + 0.786, + 0.261 + ], + "angle": 0, + "content": "35. Liu, X., Song, L., Liu, S., Zhang, Y.: A review of deep-learning-based medical image segmentation methods. Sustainability 13(3), 1224 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.787, + 0.317 + ], + "angle": 0, + "content": "36. Luo, Z., Hu, Z., Xi, Y., Zhang, R., Ma, J.: I-tuning: Tuning frozen language models with image for lightweight image captioning. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.787, + 0.346 + ], + "angle": 0, + "content": "37. Ma, J., Wang, B.: Segment anything in medical images. arXiv preprint arXiv:2304.12306 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.787, + 0.403 + ], + "angle": 0, + "content": "38. Mei, X., Liu, Z., Robson, P.M., Marinelli, B., Huang, M., Doshi, A., Jacobi, A., Cao, C., Link, K.E., Yang, T., et al.: Radimagenet: an open radiologic deep learning research dataset for effective transfer learning. Radiology: Artificial Intelligence 4(5), e210315 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.503, + 0.417 + ], + "angle": 0, + "content": "39. OpenAI: Gpt-4 technical report (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.418, + 0.787, + 0.459 + ], + "angle": 0, + "content": "40. Pan, J., Lin, Z., Zhu, X., Shao, J., Li, H.: St-adapter: Parameter-efficient image-to-video transfer learning. Advances in Neural Information Processing Systems 35, 26462-26477 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.46, + 0.787, + 0.502 + ], + "angle": 0, + "content": "41. Pedraza, L., Vargas, C., Narváez, F., Durán, O., Muñoz, E., Romero, E.: An open access thyroid ultrasound image database. In: 10th International symposium on medical information processing and analysis. vol. 9287, pp. 188-193. SPIE (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.503, + 0.787, + 0.544 + ], + "angle": 0, + "content": "42. Pfeiffer, J., Rückle, A., Poth, C., Kamath, A., Vulic, I., Ruder, S., Cho, K., Gurevych, I.: Adapterhub: A framework for adapting transformers. arXiv preprint arXiv:2007.07779 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.545, + 0.787, + 0.573 + ], + "angle": 0, + "content": "43. Pham, D.L., Xu, C., Prince, J.L.: Current methods in medical image segmentation. Annual review of biomedical engineering 2(1), 315-337 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.574, + 0.787, + 0.615 + ], + "angle": 0, + "content": "44. Rasmy, L., Xiang, Y., Xie, Z., Tao, C., Zhi, D.: Med-bert: pretrained contextualized embeddings on large-scale structured electronic health records for disease prediction. NPJ digital medicine 4(1), 86 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.616, + 0.787, + 0.671 + ], + "angle": 0, + "content": "45. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.672, + 0.787, + 0.713 + ], + "angle": 0, + "content": "46. Roy, S., Wald, T., Koehler, G., Rokuss, M.R., Disch, N., Holzschuh, J., Zimmerer, D., Maier-Hein, K.H.: Sam. md: Zero-shot medical image segmentation capabilities of the segment anything model. arXiv preprint arXiv:2304.05396 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.714, + 0.787, + 0.755 + ], + "angle": 0, + "content": "47. Sung, Y.L., Cho, J., Bansal, M.: Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5227-5237 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.757, + 0.787, + 0.798 + ], + "angle": 0, + "content": "48. Valanarasu, J.M.J., Patel, V.M.: Unext: Mlp-based rapid medical image segmentation network. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 23-33. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "49. Wang, Y., Mukherjee, S., Liu, X., Gao, J., Awadallah, A.H., Gao, J.: Adamix: Mixture-of-adapter for parameter-efficient tuning of large language models. arXiv preprint arXiv:2205.12410 1(2), 4 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.549, + 0.129 + ], + "angle": 0, + "content": "Shreyank N Gowda and David A. Clifton" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "50. Wu, H., Chen, S., Chen, G., Wang, W., Lei, B., Wen, Z.: Fat-net: Feature adaptive transformers for automated skin lesion segmentation. Medical image analysis 76, 102327 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "51. Wu, J., Fu, R., Fang, H., Liu, Y., Wang, Z., Xu, Y., Jin, Y., Arbel, T.: Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.273 + ], + "angle": 0, + "content": "52. Wunderling, T., Golla, B., Poudel, P., Arens, C., Friebe, M., Hansen, C.: Comparison of thyroid segmentation techniques for 3d ultrasound. In: Medical Imaging 2017: Image Processing. vol. 10133, pp. 346-352. SPIE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "53. Zhang, K., Liu, D.: Customized segment anything model for medical image segmentation. arXiv preprint arXiv:2304.13785 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.357 + ], + "angle": 0, + "content": "54. Zhang, Y., Liu, H., Hu, Q.: Transfuse: Fusing transformers and cnns for medical image segmentation. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part I 24. pp. 14-24. Springer (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.412 + ], + "angle": 0, + "content": "55. Zheng, S., Lu, J., Zhao, H., Zhu, X., Luo, Z., Wang, Y., Fu, Y., Feng, J., Xiang, T., Torr, P.H., et al.: Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6881-6890 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.495 + ], + "angle": 0, + "content": "56. Zhou, Z., Rahman Siddiquee, M.M., Tajbakhsh, N., Liang, J.: Unet++: A nested u-net architecture for medical image segmentation. In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 20, 2018, Proceedings 4. pp. 3-11. Springer (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.495 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CC-SAM_ Enhancing SAM with Cross-feature Attention and Context for Ultrasound Image Segmentation/4f25c9c8-80da-4658-ab8a-cd035c32acd6_origin.pdf b/2024/CC-SAM_ Enhancing SAM with Cross-feature Attention and Context for Ultrasound Image Segmentation/4f25c9c8-80da-4658-ab8a-cd035c32acd6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5c2ee42e09a32a489b1da0a0605d1ad710ac8727 --- /dev/null +++ b/2024/CC-SAM_ Enhancing SAM with Cross-feature Attention and Context for Ultrasound Image Segmentation/4f25c9c8-80da-4658-ab8a-cd035c32acd6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a49c927bfdc1ff0bd899481d146b9795673b93a813e72e2e41247b379f36fa8 +size 3011261 diff --git a/2024/CC-SAM_ Enhancing SAM with Cross-feature Attention and Context for Ultrasound Image Segmentation/full.md b/2024/CC-SAM_ Enhancing SAM with Cross-feature Attention and Context for Ultrasound Image Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1bec97cc2f444c64920df9a61b6c9fbc3678e2e7 --- /dev/null +++ b/2024/CC-SAM_ Enhancing SAM with Cross-feature Attention and Context for Ultrasound Image Segmentation/full.md @@ -0,0 +1,304 @@ +# CC-SAM: SAM with Cross-feature Attention and Context for Ultrasound Image Segmentation + +Shreyank N Gowda1 and David A. Clifton1,2 + +$^{1}$ Department of Engineering Sciences, University of Oxford, OX3 7DQ Oxford, UK + +$^{2}$ Oxford Suzhou Centre for Advanced Research, University of Oxford, Suzhou 215123, Jiangsu, China + +Abstract. The Segment Anything Model (SAM) has achieved remarkable successes in the realm of natural image segmentation, but its deployment in the medical imaging sphere has encountered challenges. Specifically, the model struggles with medical images that feature low contrast, faint boundaries, intricate morphologies, and small-sized objects. To address these challenges and enhance SAM's performance in the medical domain, we introduce a comprehensive modification. Firstly, we incorporate a frozen Convolutional Neural Network (CNN) branch as an image encoder, which synergizes with SAM's original Vision Transformer (ViT) encoder through a novel variational attention fusion module. This integration bolsters the model's capability to capture local spatial information, which is often paramount in medical imagery. Moreover, to further optimize SAM for medical imaging, we introduce feature and position adapters within the ViT branch, refining the encoder's representations. We see that compared to current prompting strategies to fine-tune SAM for ultrasound medical segmentation, the use of text descriptions that serve as text prompts for SAM helps significantly improve the performance. Leveraging ChatGPT's natural language understanding capabilities, we generate prompts that offer contextual information and guidance to SAM, enabling it to better understand the nuances of ultrasound medical images and improve its segmentation accuracy. Our method, in its entirety, represents a significant stride towards making universal image segmentation models more adaptable and efficient in the medical domain. + +# 1 Introduction + +Deep learning has revolutionized the fields of image [19, 24, 27], video classification [15, 18] and medical vision [45, 50, 56], enabling unprecedented accuracy and efficiency in these areas. Medical image segmentation plays a pivotal role in delineating and emphasizing particular organs, tissues, and anomalies within medical scans, forming a cornerstone of computer-aided diagnostic systems [35]. The development of numerous deep learning models for autonomous medical image segmentation has signaled a transformative shift in the potential of this technology [7, 45, 50, 56]. Nevertheless, these models are typically specialized for + +![](images/9b1618edebc9859558c44ab5b09ccd1effda37631ef997a96953c9ec7d244f73.jpg) +Fig. 1: Comparison of methods using SAM for medical image segmentation. + +distinct tasks and require recalibration when used for different tasks, a limitation which poses significant challenges in medical data. + +Segment anything model (SAM) [31] has emerged as a versatile foundational model for vision segmentation, gaining significant attention due to its unparalleled ability to segment a wide variety of objects and exhibit powerful zero-shot generalization. With user-driven prompts such as points, bounding boxes, and coarse masks, SAM dynamically segments the relevant objects. This adaptability positions SAM as an ideal tool for a myriad of segmentation use cases, facilitating the amalgamation of several discrete medical image segmentation tasks into a holistic model, enhancing its utility for clinical applications [29]. + +However, while SAM boasts an extensive dataset (SA1B), its performance diminishes in the medical realm due to the lack of comprehensive clinical annotations [29]. To enhance SAM's performance in medical image segmentation, many approaches have been proposed such as tuning the mask decoder [37], applying the LoRA strategy to the image encoder [53], introducing task-specific information using adapters in the ViT image encoder [51]. Previous endeavors to adapt SAM for medical image segmentation by optimizing it on medical datasets have encountered limitations, given their tokenization approach that can obfuscate essential local data. + +SAMUS [33] was subsequently introduced, aiming to transplant SAM's superior segmentation prowess and robust generalization to medical image segmentation, all while ensuring computational efficiency. Notable improvements within SAMUS encompass a redesigned ViT image encoder, integration of a parallel CNN-branch, and the introduction of a cross-branch attention mechanism, among others. + +In this paper, we introduce CC-SAM, a novel model built upon SAMUS. Notably, CC-SAM replaces the adjustable CNN with a fixed one, enhanced with adapters and employs a novel variational attention fusion instead of the cross-attention branch. Utilizing descriptions from Chat-GPT to prompt the model, we achieve a marked improvement in performance over existing prompting strategies + +tailored for medical-oriented SAM. This new approach significantly improves medical-oriented SAM performance. See Figure 1 for a comparison with other medical SAM adaptations. + +The chief contributions of this paper are delineated as follows: + +- The development of CC-SAM: a refined foundation model tailored for universal medical image segmentation, optimized for computational efficiency. +- The introduction of a static CNN with adapters, complementing the ViT encoder of SAMUS and further reducing computational costs. +- The innovative integration of variational attention fusion, enhancing the synergy between CNN and ViT branch features. +- The use of text-based prompts, sourced from Chat-GPT, to articulate the segmentation problem, which significantly amplifies segmentation performance in medical contexts. + +# 2 Related Work + +# 2.1 Medical Image Segmentation + +Medical image segmentation is crucial for identifying and measuring structures in medical images. Before deep learning, methods like thresholding, clustering, and active contour models were common [43]. However, CNNs, particularly U-Net [45], have since dominated due to their quality results even with limited data. + +Extensions of U-Net, such as DRINet [7], tackled class imbalance and high-resolution outputs. U-Net++ [56] used nested and dense skip pathways to improve segmentation, while UNeXt [48] optimized for speed and efficiency in medical segmentation by using a Convolutional MLP-based network. + +The rise of image transformers [9] led to new methods. FAT-Net [50] introduced feature adaptive transformers for detailed segmentation. UNeTR [22] combined transformers with a U-shaped design for 3D medical imaging. Our approach merges CNNs and transformers to harness both local and global feature extraction, creating a hybrid framework that leverages the strengths of both paradigms. + +# 2.2 Adapting Foundational Models + +Foundational models, often pre-trained on large datasets, provide a generic understanding of complex patterns in data. These models are crucial stepping stones in many modern machine learning workflows. However, leveraging them effectively for specific tasks necessitates a thoughtful adaptation strategy. + +Adapter modules [26] offer a solution, allowing the main network weights to remain unchanged while adding learnable layers to each Transformer layer. This adds task-specific abilities with less overhead, ensuring efficient transfer learning. This approach keeps the base knowledge while adapting to new tasks. + +These modules have excelled in various language tasks [13, 36, 47]. For instance, they've achieved top results in question answering with $95\%$ fewer parameters than standard finetuning [42]. In textual entailment, AdaMix [49] found adapters performed comparably to full finetuning but with only $3\%$ added parameters. They've also been effective in computer vision tasks like image [10, 12] and video classification [16, 17, 40]. Essentially, they ensure efficient transfer learning across tasks by adding minimal layers to fixed networks. + +We build on the idea of adapters and freeze both our CNN and ViT backbones and only use adapters for efficient fine-tuning of both models. + +# 2.3 Adapting SAM for Medical Images + +The Segment Anything Model (SAM) [31] has demonstrated notable zero-shot segmentation using various input prompts. However, its performance on medical images is subpar [25, 46]. + +Several strategies have been proposed to adapt SAM for medical imaging. MedSAM optimizes the mask decoder for efficient fine-tuning. SAMed [53] freezes the image encoder and employs a low-rank finetuning strategy while adjusting the prompt encoder and mask decoder. MSA [51] uses adapters in the ViT image encoder. To address transformer backbone limitations, SAMUS [33] incorporates a CNN-based backbone for improved local feature learning. + +Building on SAMUS [33], we optimize its framework. Instead of end-to-end CNN fine-tuning, we utilize a frozen pre-trained CNN and an adapter. We found that using text prompts, especially with Grounding DINO [34], significantly enhances SAMUS's performance. This method is not only more effective but also faster. We use GPT-4 [39] to generate text prompts from class labels. + +![](images/211dd504b834998cb26db7e9be60fcef457b8d99aeddf97e748a2f50ec2815b1.jpg) +Fig. 2: Overview of CC-SAM. We use adapters (or in the case of the CNN a FC layer) to enhance local and global features for ultrasound segmentation. 'OPaE' refers to overlapping patch embeddings and 'PE' refers to positional embeddings. + +# 3 Proposed Method + +Figure 2 provides a detailed view of our method. Essentially, we process an image $I$ using ViT-B [9] (pre-trained image encoder of SAM) and ResNet50 [24], the ResNet50 is pre-trained on RadImageNet [38]. After feature extraction, adapters refine these representations. Features are then fused using a variational attention mechanism, forming a combined feature for the Mask Decoder of SAM. + +For prompts, we use a GPT-4 generated label description and embed it with Med-BERT [44]. The Grounding DINO model, untrained on medical data, creates a bounding box from this embedding, which is input to the Prompt Encoder. Together with the variational attention feature, the Prompt Encoder helps the Mask Decoder create the final segmentation mask. + +# 3.1 Frozen Backbones with Adapters + +We've enhanced SAM's image encoder (the ViT branch) for better adaptability to smaller inputs and medical images by introducing a position adapter and five feature adapters, following SAMUS [33]. These adapters fine-tune the ViT branch efficiently with fewer parameters. + +Specifically, the position adapter modifies positional embeddings to match the embedded sequence resolution. It first downsamples these embeddings using max pooling and then refines them with a convolution operation, enabling the ViT to better manage smaller inputs. Each of the five feature adapters follows a consistent design with a downward linear projection, an activation, and an upward projection. This is mathematically expressed as: + +$$ +A (x) = G (x \mathcal {M} _ {d}) \mathcal {M} _ {u} \tag {1} +$$ + +Here, Here, $G$ denotes the GELU activation function, while $\mathcal{M}_d\in \mathbb{R}^{d\times \frac{d}{4}}$ and $\mathcal{M}_u\in \mathbb{R}^{d\times \frac{d}{4}}$ are the projection matrices. In this context, $^{\prime}d^{\prime}$ represents the dimension of the feature embedding. + +Unlike SAMUS's end-to-end CNN training, we use a static, RadImageNet [38] pre-trained ResNet-50 model. We add a trainable fully-connected layer, acting as an adapter, before its classification layer. The output features from both branches then go into our variational attention fusion block. + +# 3.2 Variational Attention Fusion Block + +A major contribution we've made is the Variational Attention Fusion Block, which adeptly merges local CNN features and global ViT features. While SAMUS introduced a cross-branch attention module, our method, which models uncertainty around these features and applies variational attentional fusion, outperforms it (details in Sec 4.3). See Figure 3 for an overview. + +![](images/7a3215e96cf544e83a943dccf29de6aba509c0b8a7551bd11df0b21053c2921b.jpg) +Fig. 3: Overview of the proposed Variational Attention Fusion Block. Each 'mode' has an intra-modal uncertainty learning encoder (represented as $\mathbf{E}$ or $\mathbf{E}^{\prime}$ in the figure), these obtain robust modality-specific features in the latent subspace. Subsequently, VAF combines these inputs and constructs a multimodal representation by estimating weights that are specific to each modality, effectively capturing their dependencies. + +Uncertainty Learning Incorporating uncertainty learning, for an input image $I$ , we derive features $s^v$ from the ViT and $s^c$ from the ResNet-50. Both $s^v$ and $s^c \in \mathbb{R}^{d^h}$ represent a fixed dimension in a latent space. We achieve a consistent size using latent encoders $E_v, E_c$ at each network's end. To address data uncertainty, we design a latent distribution for each sample's content, helping capture semantic relations. This distribution is shaped as a parameterized diagonal Gaussian. This is mathematically represented as: + +$$ +p \left(z ^ {v} \mid s ^ {v}\right) = \mathcal {N} \left(z ^ {v}; \mu^ {v}, \left(\sigma^ {v}\right) ^ {2} \mathbf {I}\right) \tag {2} +$$ + +$$ +p \left(z ^ {c} \mid s ^ {c}\right) = \mathcal {N} \left(z ^ {c}; \mu^ {c}, \left(\sigma^ {c}\right) ^ {2} \mathbf {I}\right) \tag {3} +$$ + +Here, $z^v$ and $z^c$ are reconstructed vectors. The means, $\mu^v, \mu^c$ , capture each mode's core feature, while variances, $\sigma^v, \sigma^c$ , indicate the noise-induced uncertainty in predicting these means. A higher variance signifies more uncertainty about the content observed. Both Gaussian parameters depend on the input and are forecasted through MLPs. For instance, for ViT features: $\mu^v = f\theta_{v1}(sv)$ and $\sigma^v = f\theta_{v2}(sv)$ , with $\theta_{v1}$ and $\theta_{v2}$ being respective model parameters. The same applies to CNN features. Now, each sample's feature representation shifts from deterministic to a stochastic Gaussian-drawn embedding in the latent space. + +Due to the non-differentiable nature of sampling, we use a re-parameterization trick [30] to maintain gradient flow. We draw random noise $\eta$ from a normal dis + +ribution, independent of model parameters, and produce $z^v$ using it, as detailed in Eq 4. We repeat this for $z^c$ . + +$$ +z ^ {v} = \mu^ {v} + \eta \sigma^ {v}, \eta \in \mathcal {N} (0, 1) \tag {4} +$$ + +Variational Attentional Fusion CNNs and ViTs capture different feature aspects from images (local and global). We see these as two modes: $v$ for ViTs and $c$ for CNNs. Combining them isn't straightforward due to different confidence levels. Standard methods, which get weights from the data, miss the unique traits of each mode. To tackle this, we've developed a Variational Attention Fusion (VAF) module. This module aims to capture the nuances between modalities by determining modality-specific weights for a seamless integration. + +In short, using features from each mode, a standard attention method creates a probability, $a^k$ (where $k$ is either $v$ or $c$ ), as shown in Eq 5. Here, $a^k$ shows how much each mode contributes, and $\mathbf{W}_m$ and $b_m$ are trainable factors. By having set weights, it picks the right features without being affected by varied confidence levels between modes. + +$$ +\hat {a} ^ {k} = \mathbf {W} _ {m} \mathbf {z} ^ {k} + b _ {m} +$$ + +$$ +a ^ {k} = \frac {\exp \left(\hat {a} ^ {k}\right)}{\sum_ {k ^ {\prime} \in \{v , c \}} \exp \left(\hat {a} ^ {k ^ {\prime}}\right)} \tag {5} +$$ + +Our VAF module uses variational attention weights instead of the point-estimated attention vector $\hat{a}^k$ . This approach, grounded in a probability distribution (Eq. 6), better accounts for uncertainties across modalities. + +$$ +\hat {a} ^ {k} \sim q _ {\theta} (\hat {a} ^ {k} \mid z ^ {k}) = \mathcal {N} \left(\mu_ {a} ^ {k}, \left(\sigma_ {a} ^ {k}\right) ^ {2} \mathbf {I}\right), k \in \{v, c \} \tag {6} +$$ + +We used an identity transformation to preserve modality traits in the variational attention. Mean and variance parameters are predicted by MLPs from the input. The $\left(\sigma_{a}^{k}\right)$ value indicates confidence levels among modalities. With the VAF method, more confident modality features are enhanced, while less confident ones are reduced, optimizing multimodal data fusion and capturing complementary features. The final representation combines modality-specific outputs, $z_{c}$ and $z_{v}$ , using weighted aggregation as shown in Eq 7, where $\mathbf{W}_h$ is a learnable weight matrix. + +$$ +h = \sum_ {k \in \{v, c \}} a ^ {k} \mathbf {W} _ {h} z ^ {k} \tag {7} +$$ + +# 3.3 Guiding the Prompt Encoder + +SAM's mask decoder [31] needs $h$ and an input from the prompt encoder. While past adaptations of SAM for medical imaging used point prompts, our results show that a reasonably good bounding box enhances performance. Creating this + +bounding box requires a dedicated model. Using random bounding boxes may hinder results, necessitating specific prompts. We employ Grounding-DINO [34] for bounding box generation. Though not exclusively trained for medical images, Grounding-DINO is an effective object detector. We use GPT-4 [39] to craft text prompts for specific class labels, and then MedBERT [44] generates an embedding input for Grounding-DINO. Whilst this does improve our performance in comparison to using point prompts, we show in supplementary that using the same set of random point prompts, we outperform all other medical foundational models. + +Input to GPT-4 "Create a concise description for a medical image analysis task. The task involves using ultrasound images to segment and identify specific anatomical structures or pathologies. Use [target] for to make a generic description that can be adapted for any segmentation target." Output from GPT-4 "Segment and identify [target] within ultrasound images. This requires the precise delineation of [target] from surrounding tissues, enabling accurate diagnosis and assessment. The challenge encompasses dealing with the inherent variability of ultrasound image quality, including speckle noise, shadowing, and the wide range of anatomical variations among patients. Success in this task is measured by the algorithm's ability to consistently and accurately identify and outline the [target], providing crucial information for medical decision-making processes." Where [target] is modified for each dataset. + +# 3.4 Loss Function + +Our overall loss function can be seen in Eq 8. Here, $\mathcal{L}_{BCE}$ and $\mathcal{L}_D$ are segmentation losses: binary cross entropy and dice loss respectively. + +$$ +\mathcal {L} = \mathcal {L} _ {B C E} + \mathcal {L} _ {D} + \lambda_ {c} \mathcal {L} _ {c} + \lambda_ {v} \mathcal {L} _ {v} + \lambda_ {a} \mathcal {L} _ {a} \tag {8} +$$ + +$\mathcal{L}c$ and $\mathcal{L}v$ are regularization terms for uncertainty learning, inspired by the variational information bottleneck [2]. To grasp uncertainty in a modality, a component is introduced during learning. It ensures the data's distribution, $\mathcal{N}\left(z^v;\mu^v,(\sigma^v)^2\right)$ , resembles a standard bell curve, $\mathcal{N}(\epsilon ;0,\mathbf{I})$ . This similarity is gauged using Kullback-Leibler divergence (KLD) between the distributions, promoting model diversity and reducing uncertainty, leading to robust features. Eq 9 shows how $\mathcal{L}v$ is computed, with analogous methods for $\mathcal{L}c$ and $\mathcal{L}_a$ , replacing $z^{v}$ with $z^c$ and $\hat{a}^k$ . + +$$ +\begin{array}{l} \mathcal {L} _ {v} = K L \left(\mathcal {N} \left(z ^ {v}; \mu^ {v}, (\sigma^ {v}) ^ {2}\right) \| \mathcal {N} (\epsilon ; 0, \mathbf {I})\right) \\ = - \frac {1}{2} \left(1 + \log (\sigma^ {v}) ^ {2} - (\mu^ {v}) ^ {2} - (\sigma^ {v}) ^ {2}\right) \tag {9} \\ \end{array} +$$ + +# 4 Experimental Analysis + +# 4.1 Datasets + +To compare with SAMUS [33], we test CC-SAM on the same seven public datasets: TN3K [14], DDTI [41], TG3K [52], BUSI [1], UDIAT [3], CAMUS [32], + +and HMC-QU [8]. We also recreate the US30K dataset from SAMUS for broader comparison. Dataset details are in the supplementary material. + +TN3K and TG3K datasets follow TRFE [14] for segmentation into training, validation, and testing. BUSI is randomly split in a 7:1:2 ratio for these purposes. CAMUS is initially divided into training and testing per the challenge [32], with $10\%$ of training data later used for validation. + +For model generalizability, US30K datasets remain unseen during training and validation. In line with SAMUS, we split US30k into seen and unseen datasets. We then evaluate on unseen datasets (DDTI, UDIAT, HMC-QU) not used in training or validation. This enables direct comparison of generalization with other state-of-the-art models. Moreover, we contrast CC-SAM with other base models by training on the full US30K dataset and testing across different tasks. + +# 4.2 Implementation Details + +We follow SAMUS in setting hyper parameters and adapter sizes. We change the fully tuneable CNN in SAMUS to a static ResNet50. The model is trained using the Adam optimizer with an initial learning rate of 0.01 that reduces by a tenth every 50 epochs for a total of 200 epochs. The encoders are fully connected networks with the latent space having dimensions 1024. All fully connected layers has a Leaky RELU activation function and a dropout probability of 0.5. + +We use 1xA100 GPU with 40 GB capacity. But we also verified that a single 3090 Ti with 24 GB memory is enough to train the model. We set a batch size of 32. + +# 4.3 Ablation Study + +We consider SAMUS [33] as our baseline model and ablate our added components. Specifically, we show replacing the end-to-end CNN model with a fixed pre-trained model improves performance. Further adding the FC layer at the end slightly improves performance. Crucially, replacing the cross attention fusion with our variational attention fusion significantly improves performance. Further, replacing the point based prompting with our text-based prompting gives us the best results. We show these results on two of the datasets TN3K and BUSI in Table 1. + +We do a further analysis of how the prompting compares with different prompting strategies in the supplementary material. + +# 4.4 Comparison with state-of-the-art + +We consider two scenarios for comparing with state-of-the-art models following SAMUS [33]. We compare CC-SAM with task-specific methods and foundational models. + +
ComponentsTN3KBUSI
ResNet50VAFText PromptDiceHDDiceHD
×××84.4528.2285.7725.49
✓××84.9127.2586.2225.12
×✓×85.5526.8487.1524.42
××✓85.9526.2287.9223.92
✓×✓86.4225.8288.4523.20
×✓✓86.2525.9288.2223.51
✓✓×85.7226.2187.4224.61
✓✓✓87.1125.2589.5122.86
+ +Task-specific methods Following SAMUS [33], thirteen state-of-the-art, task-specific techniques are chosen for evaluation, spanning CNN-based, transformer-based, and CNN-Transformer hybrid methods. Methods other than SAMUS include CNN-based techniques encompass U-Net [45], CPFNet [11], CA-Net [20], CE-Net [21], and AAU-Net [5]. The transformer-centric methods consist of SwinUnet [4], SETR [55], and MISSFormer [28]. The hybrids of CNN and Transformer are TransUNet [6], TransFuse [54], FAT-Net [50], and H2Former [23]. + +Table 1: Ablation study of adding different components of our proposed approach. The first row corresponds to the baseline SAMUS [33]. + +
MethodCAMUS-LATN3KBUSICAMUS-LVCAMUS-MYO
DiceHDDiceHDDiceHDDiceHDDiceHD
U-Net91.0012.9179.0134.1278.1133.6093.569.9086.8616.87
CPFNet91.5112.2679.4333.0780.5627.9893.329.6386.6816.51
CA-Net91.2812.3280.5233.6581.8828.6793.599.7787.2116.24
CE-Net91.1412.2980.3732.7980.2130.1993.369.9186.4716.66
AAU-Net91.3312.1282.2830.5380.8130.3993.329.9786.9816.49
SwinUNet89.8014.7470.0844.1367.2347.0291.7212.8084.4620.25
SETR90.5213.9167.8044.1168.2240.3792.8211.3486.2018.27
MISSFormer91.1811.8279.4232.8578.4333.1093.259.9486.5716.50
TransUNet91.3712.4681.4430.9882.2227.5493.609.6087.2017.25
FAT-Net91.5512.0580.4532.7782.1628.5593.599.2087.1915.93
H2Former90.9811.9282.4830.5881.4827.8493.449.6087.3116.60
SAMUS91.5811.6084.4528.2285.7725.4993.739.7987.4616.74
CC-SAM92.0311.1185.2027.1087.0124.2293.959.1188.2516.11
+ +Table 2: Quantitative comparison between our CC-SAM method and the state-of-the-art (SOTA) task-specific techniques for segmenting thyroid nodules (TN3K), breast cancer (BUSI), left ventricle (CAMUS-LV), myocardium (CAMUS-MYO), and left atrium (CAMUS-LA). We assessed the performance using the Dice score (\%) and the Hausdorff distance (HD). The top-performing results are highlighted in bold. + +![](images/c27111f42dfcdcd089e6bd03ffb33716acc56e2b4237d47255da2b8cce2cc3cd.jpg) +Fig. 4: Qualitative comparison between our CC-SAM method and the state-of-the-art (SOTA) task-specific techniques. + +Quantitative Analysis. Table 2 summarizes the numerical outcomes of various task-specific methods on TN3K, BUSI, CAMUS-LV, CAMUS-MYO, and CAMUS-LA. In comparison, CC-SAM consistently excels in performance across all these five tasks, setting a new benchmark for SOTA results. While CC-SAM surpasses recent SOTA in most benchmarks, it does fall short by a very small margin when evaluated with the Hausdorff distance (HD) on CAMUS-MYO. + +Qualitative Analysis. Figure 4 showcases the segmentation results from various methods, including U-Net [45], AAU-Net [5], MISSFormer [28], H2Former [23], SAMUS [33], and our proposed CC-SAM. Segmenting ultrasound images is notably challenging given their low contrast, inconsistent features, and indistinct object boundaries. Many existing techniques face difficulties in precisely differentiating the target from its background, often resulting in significant false negatives or false positives. In contrast, SAMUS stands out in maintaining the integrity of target areas and minimizing false positives. Our CC-SAM further refines the outcomes observed with SAMUS, a testament to the intrinsic strengths of SAM and the tailored modifications and innovations integrated into CC-SAM. + +Generalization Ability. Evaluating the generalization performance of various task-specific methods is crucial as it provides insights into how well these methods can adapt and function across unfamiliar or unseen datasets. Figure 5 offers a quantitative assessment of this aspect. When we assess performance discrepancies between known (seeable) and unknown (unseen) datasets, CC-SAM stands out by displaying the most minimal performance drop among all methods across the three segmentation tasks. This underlines the remarkable generalization prowess of CC-SAM, emphasizing its resilience and versatility in a variety of medical image segmentation situations. + +Foundational models We also compare with five state-of-the-art foundational models. These include the original SAM [31], SAMed [53], SAMUS [33], MSA [51] + +![](images/e6c2172cc496fc37ea44f667a21ddae0b7d5fa77b2d207aa37c1f60a37059481.jpg) +Fig. 5: Comparison of CC-SAM with task-specific techniques on see datasets (highlighted in blue) and unseen datasets not previously encountered (indicated in orange). Higher orange bars indicate stronger generalization ability. + +and MedSAM [37]. SAM is trained on SA-1B and kept frozen due to the cost involved in fine-tuning. All other foundational models are pre-trained on the US30k dataset and evaluated using Dice and HD across TN3K, BUSI, CAMUS-LV, CAMUS-MYO, and CAMUS-LA datasets. + +
MethodTN3KBUSICAMUS-LVCAMUS-MYOCAMUS-LA
DiceHDDiceHDDiceHDDiceHDDiceHD
SAM29.59134.8754.0182.3928.18196.6429.42184.1017.28193.70
MedSAM71.0942.9177.7534.2687.5215.2876.0725.7288.0615.70
SAMed80.4031.2974.8234.6087.6713.2482.6019.4890.9212.60
MSA82.6729.1581.6628.8790.9511.2982.4719.2891.8011.59
SAMUS83.0528.8284.5427.2491.1311.7683.1118.9992.0012.08
CC-SAM85.5927.7486.2225.8592.8510.8885.6117.1193.5111.06
+ +Table 3: Quantitative comparison of our CC-SAM and other foundation models on seeable US30K data. The performance is evaluated by the Dice score (\%) and Hausdorff distance (HD). + +Quantitative Analysis. Table 3 summarizes the numerical outcomes of various foundational methods on TN3K, BUSI, CAMUS-LV, CAMUS-MYO, and CAMUS-LA. In comparison, CC-SAM consistently excels in performance across all tasks, setting a new benchmark for foundational models. + +Qualitative Analysis. Figure 6 shows the qualitative segmentation results from various foundational models, including SAM, MedSAM, SAMed, MSA, SAMUS, and CC-SAM. SAM, without adjustments for medical images, fails at segmentation. However, MedSAM, SAMed, and MSA, through tuning, recover some + +segmentation ability but struggle with accurately defining borders, resulting in many false positives and negatives. In contrast, SAMUS excels in identifying precise segmentation borders, even in low contrast, suggesting the benefit of integrating local data with the image encoder for better boundary and shape retention in medical image segmentation. CC-SAM goes further, improving upon SAMUS's performance and setting new benchmarks. + +![](images/413bb74e2354fc947cc7cceeecef1c37197b7be244f7afd8d3182c98c435eaba.jpg) +Fig. 6: Qualitative comparison between our CC-SAM method and the state-of-the-art (SOTA) foundational medical segmentation models. + +Generalization Ability. Generally, foundation models trained on US30K exhibit a much improved generalization performance in medical image segmentation tasks compared to the original SAM. When considering the three specific segmentation tasks: thyroid nodule segmentation, breast cancer segmentation, and myocardium segmentation, it's clear that all foundation models face significant challenges with myocardium segmentation, while they perform relatively well in the breast cancer segmentation task. CC-SAM outperforms all foundational models significantly. This highlights the remarkable generalization prowess of CC-SAM, as it consistently surpasses other foundation models, especially in unfamiliar domains. Figure 7 offers a quantitative assessment of this aspect. We see that the original SAM performs the worst, which is expected and the proposed CC-SAM obtains a new state-of-the-art on all datasets. + +Deployment Cost Comparison with Other Foundational Models We compare CCSAM with other foundational models in a more detailed manner. We look at 'Generalization Ability/Dice $(\%)$ ,GPU Memory/G, Parameters/M,Computation/GFLOPs', Inference Speed/FPS' and Segmentation Performance/Dice $(\%)$ .The spider plot of this can be seen in Figure 8. + +![](images/73ae6c8b6854707c443c0576ab63864cad3cfcc815e2d268f2848e3cb3894f50.jpg) +Fig.7: Comparing segmentation and generalization ability comparison of our CC-SAMUS and other foundation models. Whilst SAM obtains the lowest Dice Score, CC-SAM obtains the highest across all datasets. + +![](images/b0fc5e0214f75dcdc52bb87a7fa5db9da34ae8e7cda733d60ee273229c44cc0d.jpg) +Fig. 8: Spider plot comparing efficiency and segmentation performance of all foundational models. + +# 5 Conclusion + +The Segment Anything Model (SAM), while proficient in natural image segmentation, faced hurdles in medical imaging due to issues with low contrast, faint boundaries, and other complexities. In this study, we identified and addressed the limitations of the Segment Anything Model (SAM) when applied to medical imaging, especially with images that have inherent challenges like low contrast and intricate morphologies. Our enhanced approach integrates a CNN branch with SAM's original Vision Transformer (ViT) encoder, utilizing a unique variational attention fusion module to capture vital local spatial information present in medical images. By introducing feature and position adapters within the ViT branch, the model's representations are further refined. Notably, we found that using text prompts, especially those generated leveraging ChatGPT's capabilities, significantly boosts SAM's performance in ultrasound medical segmentation. These innovations mark a pivotal step towards improving the adaptability and efficiency of universal image segmentation models in the medical field. + +Acknowledgements DAC was supported by the Pandemic Sciences Institute at the University of Oxford; the National Institute for Health Research (NIHR) Oxford Biomedical Research Centre (BRC); an NIHR Research Professorship; a Royal Academy of Engineering Research Chair; the Wellcome Trust funded VITAL project (grant 204904/Z/16/Z); the EPSRC (grant EP/W031744/1); and the InnoHK Hong Kong Centre for Cerebro-cardiovascular Engineering (COCHE). + +# References + +1. Al-Dhabyani, W., Gomaa, M., Khaled, H., Aly, F.: Deep learning approaches for data augmentation and classification of breast masses using ultrasound images. Int. J. Adv. Comput. Sci. Appl 10(5), 1-11 (2019) +2. Alemi, A.A., Fischer, I., Dillon, J.V., Murphy, K.: Deep variational information bottleneck. arXiv preprint arXiv:1612.00410 (2016) +3. Byra, M., Jarosik, P., Szubert, A., Galperin, M., Ojeda-Fournier, H., Olson, L., O'Boyle, M., Comstock, C., Andre, M.: Breast mass segmentation in ultrasound with selective kernel u-net convolutional neural network. Biomedical Signal Processing and Control 61, 102027 (2020) +4. Cao, H., Wang, Y., Chen, J., Jiang, D., Zhang, X., Tian, Q., Wang, M.: Swinunet: Unet-like pure transformer for medical image segmentation. In: European conference on computer vision. pp. 205-218. Springer (2022) +5. Chen, G., Li, L., Dai, Y., Zhang, J., Yap, M.H.: Aau-net: an adaptive attention u-net for breast lesions segmentation in ultrasound images. IEEE Transactions on Medical Imaging (2022) +6. Chen, J., Lu, Y., Yu, Q., Luo, X., Adeli, E., Wang, Y., Lu, L., Yuille, A.L., Zhou, Y.: Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306 (2021) +7. Chen, L., Bentley, P., Mori, K., Misawa, K., Fujiwara, M., Rueckert, D.: Drinet for medical image segmentation. IEEE transactions on medical imaging 37(11), 2453-2462 (2018) +8. Degerli, A., Kiranyaz, S., Hamid, T., Mazhar, R., Gabbouj, M.: Early myocardial infarction detection over multi-view echocardiography. Biomedical Signal Processing and Control 87, 105448 (2024). https://doi.org/https://doi.org/10.1016/j.bspc.2023.105448 +9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +0. Ermis, B., Zappella, G., Wistuba, M., Rawal, A., Archambeau, C.: Continual learning with transformers for image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3774-3781 (2022) +1. Feng, S., Zhao, H., Shi, F., Cheng, X., Wang, M., Ma, Y., Xiang, D., Zhu, W., Chen, X.: Cpfnet: Context pyramid fusion network for medical image segmentation. IEEE transactions on medical imaging 39(10), 3008-3018 (2020) +2. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision pp. 1-15 (2023) +3. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023) +4. Gong, H., Chen, J., Chen, G., Li, H., Li, G., Chen, F.: Thyroid region prior guided attention for ultrasound segmentation of thyroid nodules. Computers in Biology and Medicine 155, 106389 (2023) +5. Gowda, S.N.: Human activity recognition using combinatorial deep belief networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 1-6 (2017) + +16. Gowda, S.N., Arnab, A., Huang, J.: Optimizing vivit training: Time and memory reduction for action recognition. arXiv preprint arXiv:2306.04822 (2023) +17. Gowda, S.N., Gao, B., Clifton, D.: Fe-adapter: adapting image-based emotion classifiers to videos (2024) +18. Gowda, S.N., Rohrbach, M., Sevilla-Lara, L.: Smart frame selection for action recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1451-1459 (2021) +19. Gowda, S.N., Yuan, C.: Colornet: Investigating the importance of color spaces for image classification. In: Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part IV 14. pp. 581–596. Springer (2019) +20. Gu, R., Wang, G., Song, T., Huang, R., Aertsen, M., Deprest, J., Ourselin, S., Vercauteren, T., Zhang, S.: Ca-net: Comprehensive attention convolutional neural networks for explainable medical image segmentation. IEEE transactions on medical imaging 40(2), 699-711 (2020) +21. Gu, Z., Cheng, J., Fu, H., Zhou, K., Hao, H., Zhao, Y., Zhang, T., Gao, S., Liu, J.: Ce-net: Context encoder network for 2d medical image segmentation. IEEE transactions on medical imaging 38(10), 2281-2292 (2019) +22. Hatamizadeh, A., Tang, Y., Nath, V., Yang, D., Myronenko, A., Landman, B., Roth, H.R., Xu, D.: Unetr: Transformers for 3d medical image segmentation. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 574-584 (2022) +23. He, A., Wang, K., Li, T., Du, C., Xia, S., Fu, H.: H2former: An efficient hierarchical hybrid transformer for medical image segmentation. IEEE Transactions on Medical Imaging (2023) +24. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +25. He, S., Bao, R., Li, J., Grant, P.E., Ou, Y.: Accuracy of segment-anything model (sam) in medical image segmentation tasks. arXiv preprint arXiv:2304.09324 (2023) +26. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019) +27. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017) +28. Huang, X., Deng, Z., Li, D., Yuan, X.: Missformer: An effective medical image segmentation transformer. arXiv preprint arXiv:2109.07162 (2021) +29. Huang, Y., Yang, X., Liu, L., Zhou, H., Chang, A., Zhou, X., Chen, R., Yu, J., Chen, J., Chen, C., et al.: Segment anything model for medical images? arXiv preprint arXiv:2304.14660 (2023) +30. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013) +31. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) +32. Leclerc, S., Smistad, E., Pedrosa, J., Østvik, A., Cervenansky, F., Espinosa, F., Espeland, T., Berg, E.A.R., Jodoin, P.M., Grenier, T., et al.: Deep learning for segmentation using an open large-scale dataset in 2d echocardiography. IEEE transactions on medical imaging 38(9), 2198-2210 (2019) + +33. Lin, X., Xiang, Y., Zhang, L., Yang, X., Yan, Z., Yu, L.: Samus: Adapting segment anything model for clinically-friendly and generalizable ultrasound image segmentation. arXiv preprint arXiv:2309.06824 (2023) +34. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023) +35. Liu, X., Song, L., Liu, S., Zhang, Y.: A review of deep-learning-based medical image segmentation methods. Sustainability 13(3), 1224 (2021) +36. Luo, Z., Hu, Z., Xi, Y., Zhang, R., Ma, J.: I-tuning: Tuning frozen language models with image for lightweight image captioning. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023) +37. Ma, J., Wang, B.: Segment anything in medical images. arXiv preprint arXiv:2304.12306 (2023) +38. Mei, X., Liu, Z., Robson, P.M., Marinelli, B., Huang, M., Doshi, A., Jacobi, A., Cao, C., Link, K.E., Yang, T., et al.: Radimagenet: an open radiologic deep learning research dataset for effective transfer learning. Radiology: Artificial Intelligence 4(5), e210315 (2022) +39. OpenAI: Gpt-4 technical report (2023) +40. Pan, J., Lin, Z., Zhu, X., Shao, J., Li, H.: St-adapter: Parameter-efficient image-to-video transfer learning. Advances in Neural Information Processing Systems 35, 26462-26477 (2022) +41. Pedraza, L., Vargas, C., Narváez, F., Durán, O., Muñoz, E., Romero, E.: An open access thyroid ultrasound image database. In: 10th International symposium on medical information processing and analysis. vol. 9287, pp. 188-193. SPIE (2015) +42. Pfeiffer, J., Rückle, A., Poth, C., Kamath, A., Vulic, I., Ruder, S., Cho, K., Gurevych, I.: Adapterhub: A framework for adapting transformers. arXiv preprint arXiv:2007.07779 (2020) +43. Pham, D.L., Xu, C., Prince, J.L.: Current methods in medical image segmentation. Annual review of biomedical engineering 2(1), 315-337 (2000) +44. Rasmy, L., Xiang, Y., Xie, Z., Tao, C., Zhi, D.: Med-bert: pretrained contextualized embeddings on large-scale structured electronic health records for disease prediction. NPJ digital medicine 4(1), 86 (2021) +45. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) +46. Roy, S., Wald, T., Koehler, G., Rokuss, M.R., Disch, N., Holzschuh, J., Zimmerer, D., Maier-Hein, K.H.: Sam. md: Zero-shot medical image segmentation capabilities of the segment anything model. arXiv preprint arXiv:2304.05396 (2023) +47. Sung, Y.L., Cho, J., Bansal, M.: Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5227-5237 (2022) +48. Valanarasu, J.M.J., Patel, V.M.: Unext: Mlp-based rapid medical image segmentation network. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 23-33. Springer (2022) +49. Wang, Y., Mukherjee, S., Liu, X., Gao, J., Awadallah, A.H., Gao, J.: Adamix: Mixture-of-adapter for parameter-efficient tuning of large language models. arXiv preprint arXiv:2205.12410 1(2), 4 (2022) + +50. Wu, H., Chen, S., Chen, G., Wang, W., Lei, B., Wen, Z.: Fat-net: Feature adaptive transformers for automated skin lesion segmentation. Medical image analysis 76, 102327 (2022) +51. Wu, J., Fu, R., Fang, H., Liu, Y., Wang, Z., Xu, Y., Jin, Y., Arbel, T.: Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620 (2023) +52. Wunderling, T., Golla, B., Poudel, P., Arens, C., Friebe, M., Hansen, C.: Comparison of thyroid segmentation techniques for 3d ultrasound. In: Medical Imaging 2017: Image Processing. vol. 10133, pp. 346-352. SPIE (2017) +53. Zhang, K., Liu, D.: Customized segment anything model for medical image segmentation. arXiv preprint arXiv:2304.13785 (2023) +54. Zhang, Y., Liu, H., Hu, Q.: Transfuse: Fusing transformers and cnns for medical image segmentation. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part I 24. pp. 14-24. Springer (2021) +55. Zheng, S., Lu, J., Zhao, H., Zhu, X., Luo, Z., Wang, Y., Fu, Y., Feng, J., Xiang, T., Torr, P.H., et al.: Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6881-6890 (2021) +56. Zhou, Z., Rahman Siddiquee, M.M., Tajbakhsh, N., Liang, J.: Unet++: A nested u-net architecture for medical image segmentation. In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 20, 2018, Proceedings 4. pp. 3-11. Springer (2018) \ No newline at end of file diff --git a/2024/CC-SAM_ Enhancing SAM with Cross-feature Attention and Context for Ultrasound Image Segmentation/images.zip b/2024/CC-SAM_ Enhancing SAM with Cross-feature Attention and Context for Ultrasound Image Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..faa0e2df4f2dd93511200841a1ab5e7fd61de5fe --- /dev/null +++ b/2024/CC-SAM_ Enhancing SAM with Cross-feature Attention and Context for Ultrasound Image Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea1b642d00ae9001b9a35339e61cfda40ca10bd06c98145649bd662fa320d740 +size 617228 diff --git a/2024/CC-SAM_ Enhancing SAM with Cross-feature Attention and Context for Ultrasound Image Segmentation/layout.json b/2024/CC-SAM_ Enhancing SAM with Cross-feature Attention and Context for Ultrasound Image Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..06bf3d18a93dfcc542cf64dd0618841989d8e65f --- /dev/null +++ b/2024/CC-SAM_ Enhancing SAM with Cross-feature Attention and Context for Ultrasound Image Segmentation/layout.json @@ -0,0 +1,8412 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 479, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 479, + 147 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 479, + 147 + ], + "type": "text", + "content": "CC-SAM: SAM with Cross-feature Attention and Context for Ultrasound Image Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 206, + 168, + 405, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 168, + 405, + 180 + ], + "spans": [ + { + "bbox": [ + 206, + 168, + 405, + 180 + ], + "type": "text", + "content": "Shreyank N Gowda1 and David A. Clifton1,2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 135, + 190, + 478, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 190, + 478, + 201 + ], + "spans": [ + { + "bbox": [ + 135, + 190, + 478, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 190, + 478, + 201 + ], + "type": "text", + "content": " Department of Engineering Sciences, University of Oxford, OX3 7DQ Oxford, UK" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 201, + 466, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 201, + 466, + 223 + ], + "spans": [ + { + "bbox": [ + 146, + 201, + 466, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 201, + 466, + 223 + ], + "type": "text", + "content": " Oxford Suzhou Centre for Advanced Research, University of Oxford, Suzhou 215123, Jiangsu, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 256, + 452, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 256, + 452, + 518 + ], + "spans": [ + { + "bbox": [ + 159, + 256, + 452, + 518 + ], + "type": "text", + "content": "Abstract. The Segment Anything Model (SAM) has achieved remarkable successes in the realm of natural image segmentation, but its deployment in the medical imaging sphere has encountered challenges. Specifically, the model struggles with medical images that feature low contrast, faint boundaries, intricate morphologies, and small-sized objects. To address these challenges and enhance SAM's performance in the medical domain, we introduce a comprehensive modification. Firstly, we incorporate a frozen Convolutional Neural Network (CNN) branch as an image encoder, which synergizes with SAM's original Vision Transformer (ViT) encoder through a novel variational attention fusion module. This integration bolsters the model's capability to capture local spatial information, which is often paramount in medical imagery. Moreover, to further optimize SAM for medical imaging, we introduce feature and position adapters within the ViT branch, refining the encoder's representations. We see that compared to current prompting strategies to fine-tune SAM for ultrasound medical segmentation, the use of text descriptions that serve as text prompts for SAM helps significantly improve the performance. Leveraging ChatGPT's natural language understanding capabilities, we generate prompts that offer contextual information and guidance to SAM, enabling it to better understand the nuances of ultrasound medical images and improve its segmentation accuracy. Our method, in its entirety, represents a significant stride towards making universal image segmentation models more adaptable and efficient in the medical domain." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 542, + 230, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 542, + 230, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 542, + 230, + 555 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Deep learning has revolutionized the fields of image [19, 24, 27], video classification [15, 18] and medical vision [45, 50, 56], enabling unprecedented accuracy and efficiency in these areas. Medical image segmentation plays a pivotal role in delineating and emphasizing particular organs, tissues, and anomalies within medical scans, forming a cornerstone of computer-aided diagnostic systems [35]. The development of numerous deep learning models for autonomous medical image segmentation has signaled a transformative shift in the potential of this technology [7, 45, 50, 56]. Nevertheless, these models are typically specialized for" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 114, + 471, + 258 + ], + "blocks": [ + { + "bbox": [ + 145, + 114, + 471, + 258 + ], + "lines": [ + { + "bbox": [ + 145, + 114, + 471, + 258 + ], + "spans": [ + { + "bbox": [ + 145, + 114, + 471, + 258 + ], + "type": "image", + "image_path": "9b1618edebc9859558c44ab5b09ccd1effda37631ef997a96953c9ec7d244f73.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 150, + 266, + 463, + 278 + ], + "lines": [ + { + "bbox": [ + 150, + 266, + 463, + 278 + ], + "spans": [ + { + "bbox": [ + 150, + 266, + 463, + 278 + ], + "type": "text", + "content": "Fig. 1: Comparison of methods using SAM for medical image segmentation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 303, + 480, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 480, + 327 + ], + "type": "text", + "content": "distinct tasks and require recalibration when used for different tasks, a limitation which poses significant challenges in medical data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 328, + 480, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 480, + 424 + ], + "type": "text", + "content": "Segment anything model (SAM) [31] has emerged as a versatile foundational model for vision segmentation, gaining significant attention due to its unparalleled ability to segment a wide variety of objects and exhibit powerful zero-shot generalization. With user-driven prompts such as points, bounding boxes, and coarse masks, SAM dynamically segments the relevant objects. This adaptability positions SAM as an ideal tool for a myriad of segmentation use cases, facilitating the amalgamation of several discrete medical image segmentation tasks into a holistic model, enhancing its utility for clinical applications [29]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 425, + 481, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 481, + 532 + ], + "type": "text", + "content": "However, while SAM boasts an extensive dataset (SA1B), its performance diminishes in the medical realm due to the lack of comprehensive clinical annotations [29]. To enhance SAM's performance in medical image segmentation, many approaches have been proposed such as tuning the mask decoder [37], applying the LoRA strategy to the image encoder [53], introducing task-specific information using adapters in the ViT image encoder [51]. Previous endeavors to adapt SAM for medical image segmentation by optimizing it on medical datasets have encountered limitations, given their tokenization approach that can obfuscate essential local data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 481, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 604 + ], + "type": "text", + "content": "SAMUS [33] was subsequently introduced, aiming to transplant SAM's superior segmentation prowess and robust generalization to medical image segmentation, all while ensuring computational efficiency. Notable improvements within SAMUS encompass a redesigned ViT image encoder, integration of a parallel CNN-branch, and the introduction of a cross-branch attention mechanism, among others." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "In this paper, we introduce CC-SAM, a novel model built upon SAMUS. Notably, CC-SAM replaces the adjustable CNN with a fixed one, enhanced with adapters and employs a novel variational attention fusion instead of the cross-attention branch. Utilizing descriptions from Chat-GPT to prompt the model, we achieve a marked improvement in performance over existing prompting strategies" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 335, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 335, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 335, + 102 + ], + "type": "text", + "content": "Shreyank N Gowda and David A. Clifton" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "tailored for medical-oriented SAM. This new approach significantly improves medical-oriented SAM performance. See Figure 1 for a comparison with other medical SAM adaptations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 152, + 420, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 152, + 420, + 164 + ], + "spans": [ + { + "bbox": [ + 146, + 152, + 420, + 164 + ], + "type": "text", + "content": "The chief contributions of this paper are delineated as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 175, + 479, + 281 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 175, + 479, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 175, + 479, + 198 + ], + "spans": [ + { + "bbox": [ + 138, + 175, + 479, + 198 + ], + "type": "text", + "content": "- The development of CC-SAM: a refined foundation model tailored for universal medical image segmentation, optimized for computational efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 200, + 479, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 200, + 479, + 222 + ], + "spans": [ + { + "bbox": [ + 138, + 200, + 479, + 222 + ], + "type": "text", + "content": "- The introduction of a static CNN with adapters, complementing the ViT encoder of SAMUS and further reducing computational costs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 224, + 479, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 479, + 246 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 479, + 246 + ], + "type": "text", + "content": "- The innovative integration of variational attention fusion, enhancing the synergy between CNN and ViT branch features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 247, + 479, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 247, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 247, + 479, + 281 + ], + "type": "text", + "content": "- The use of text-based prompts, sourced from Chat-GPT, to articulate the segmentation problem, which significantly amplifies segmentation performance in medical contexts." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 303, + 236, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 236, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 236, + 316 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 329, + 306, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 329, + 306, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 329, + 306, + 341 + ], + "type": "text", + "content": "2.1 Medical Image Segmentation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 350, + 479, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 479, + 408 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 479, + 408 + ], + "type": "text", + "content": "Medical image segmentation is crucial for identifying and measuring structures in medical images. Before deep learning, methods like thresholding, clustering, and active contour models were common [43]. However, CNNs, particularly U-Net [45], have since dominated due to their quality results even with limited data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 410, + 479, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 479, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 479, + 456 + ], + "type": "text", + "content": "Extensions of U-Net, such as DRINet [7], tackled class imbalance and high-resolution outputs. U-Net++ [56] used nested and dense skip pathways to improve segmentation, while UNeXt [48] optimized for speed and efficiency in medical segmentation by using a Convolutional MLP-based network." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 458, + 479, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 479, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 479, + 529 + ], + "type": "text", + "content": "The rise of image transformers [9] led to new methods. FAT-Net [50] introduced feature adaptive transformers for detailed segmentation. UNeTR [22] combined transformers with a U-shaped design for 3D medical imaging. Our approach merges CNNs and transformers to harness both local and global feature extraction, creating a hybrid framework that leverages the strengths of both paradigms." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 548, + 316, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 316, + 560 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 316, + 560 + ], + "type": "text", + "content": "2.2 Adapting Foundational Models" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 570, + 479, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 479, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 479, + 616 + ], + "type": "text", + "content": "Foundational models, often pre-trained on large datasets, provide a generic understanding of complex patterns in data. These models are crucial stepping stones in many modern machine learning workflows. However, leveraging them effectively for specific tasks necessitates a thoughtful adaptation strategy." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 618, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 618, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 618, + 479, + 665 + ], + "type": "text", + "content": "Adapter modules [26] offer a solution, allowing the main network weights to remain unchanged while adding learnable layers to each Transformer layer. This adds task-specific abilities with less overhead, ensuring efficient transfer learning. This approach keeps the base knowledge while adapting to new tasks." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Title Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "content": "These modules have excelled in various language tasks [13, 36, 47]. For instance, they've achieved top results in question answering with " + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "content": " fewer parameters than standard finetuning [42]. In textual entailment, AdaMix [49] found adapters performed comparably to full finetuning but with only " + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "content": " added parameters. They've also been effective in computer vision tasks like image [10, 12] and video classification [16, 17, 40]. Essentially, they ensure efficient transfer learning across tasks by adding minimal layers to fixed networks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 481, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 481, + 224 + ], + "type": "text", + "content": "We build on the idea of adapters and freeze both our CNN and ViT backbones and only use adapters for efficient fine-tuning of both models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 238, + 338, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 238, + 338, + 251 + ], + "spans": [ + { + "bbox": [ + 131, + 238, + 338, + 251 + ], + "type": "text", + "content": "2.3 Adapting SAM for Medical Images" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 255, + 480, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 480, + 290 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 480, + 290 + ], + "type": "text", + "content": "The Segment Anything Model (SAM) [31] has demonstrated notable zero-shot segmentation using various input prompts. However, its performance on medical images is subpar [25, 46]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 291, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 482, + 361 + ], + "type": "text", + "content": "Several strategies have been proposed to adapt SAM for medical imaging. MedSAM optimizes the mask decoder for efficient fine-tuning. SAMed [53] freezes the image encoder and employs a low-rank finetuning strategy while adjusting the prompt encoder and mask decoder. MSA [51] uses adapters in the ViT image encoder. To address transformer backbone limitations, SAMUS [33] incorporates a CNN-based backbone for improved local feature learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 362, + 482, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 362, + 482, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 362, + 482, + 423 + ], + "type": "text", + "content": "Building on SAMUS [33], we optimize its framework. Instead of end-to-end CNN fine-tuning, we utilize a frozen pre-trained CNN and an adapter. We found that using text prompts, especially with Grounding DINO [34], significantly enhances SAMUS's performance. This method is not only more effective but also faster. We use GPT-4 [39] to generate text prompts from class labels." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 141, + 442, + 474, + 601 + ], + "blocks": [ + { + "bbox": [ + 141, + 442, + 474, + 601 + ], + "lines": [ + { + "bbox": [ + 141, + 442, + 474, + 601 + ], + "spans": [ + { + "bbox": [ + 141, + 442, + 474, + 601 + ], + "type": "image", + "image_path": "211dd504b834998cb26db7e9be60fcef457b8d99aeddf97e748a2f50ec2815b1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 607, + 482, + 642 + ], + "lines": [ + { + "bbox": [ + 130, + 607, + 482, + 642 + ], + "spans": [ + { + "bbox": [ + 130, + 607, + 482, + 642 + ], + "type": "text", + "content": "Fig. 2: Overview of CC-SAM. We use adapters (or in the case of the CNN a FC layer) to enhance local and global features for ultrasound segmentation. 'OPaE' refers to overlapping patch embeddings and 'PE' refers to positional embeddings." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 335, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 335, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 335, + 102 + ], + "type": "text", + "content": "Shreyank N Gowda and David A. Clifton" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 261, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 261, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 261, + 129 + ], + "type": "text", + "content": "3 Proposed Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 145, + 479, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 145, + 479, + 204 + ], + "spans": [ + { + "bbox": [ + 130, + 145, + 479, + 204 + ], + "type": "text", + "content": "Figure 2 provides a detailed view of our method. Essentially, we process an image " + }, + { + "bbox": [ + 130, + 145, + 479, + 204 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 145, + 479, + 204 + ], + "type": "text", + "content": " using ViT-B [9] (pre-trained image encoder of SAM) and ResNet50 [24], the ResNet50 is pre-trained on RadImageNet [38]. After feature extraction, adapters refine these representations. Features are then fused using a variational attention mechanism, forming a combined feature for the Mask Decoder of SAM." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 205, + 480, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 480, + 266 + ], + "type": "text", + "content": "For prompts, we use a GPT-4 generated label description and embed it with Med-BERT [44]. The Grounding DINO model, untrained on medical data, creates a bounding box from this embedding, which is input to the Prompt Encoder. Together with the variational attention feature, the Prompt Encoder helps the Mask Decoder create the final segmentation mask." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 289, + 327, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 289, + 327, + 300 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 327, + 300 + ], + "type": "text", + "content": "3.1 Frozen Backbones with Adapters" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 313, + 479, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 479, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 479, + 361 + ], + "type": "text", + "content": "We've enhanced SAM's image encoder (the ViT branch) for better adaptability to smaller inputs and medical images by introducing a position adapter and five feature adapters, following SAMUS [33]. These adapters fine-tune the ViT branch efficiently with fewer parameters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 362, + 480, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 362, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 362, + 480, + 434 + ], + "type": "text", + "content": "Specifically, the position adapter modifies positional embeddings to match the embedded sequence resolution. It first downsamples these embeddings using max pooling and then refines them with a convolution operation, enabling the ViT to better manage smaller inputs. Each of the five feature adapters follows a consistent design with a downward linear projection, an activation, and an upward projection. This is mathematically expressed as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 262, + 450, + 480, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 450, + 480, + 462 + ], + "spans": [ + { + "bbox": [ + 262, + 450, + 480, + 462 + ], + "type": "interline_equation", + "content": "A (x) = G (x \\mathcal {M} _ {d}) \\mathcal {M} _ {u} \\tag {1}", + "image_path": "1c6063ca07f33abe78c21c014332d65b9aaaa0f86e9a18844978491e0850cef5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 472, + 479, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 479, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 479, + 508 + ], + "type": "text", + "content": "Here, Here, " + }, + { + "bbox": [ + 130, + 472, + 479, + 508 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 472, + 479, + 508 + ], + "type": "text", + "content": " denotes the GELU activation function, while " + }, + { + "bbox": [ + 130, + 472, + 479, + 508 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_d\\in \\mathbb{R}^{d\\times \\frac{d}{4}}" + }, + { + "bbox": [ + 130, + 472, + 479, + 508 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 472, + 479, + 508 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_u\\in \\mathbb{R}^{d\\times \\frac{d}{4}}" + }, + { + "bbox": [ + 130, + 472, + 479, + 508 + ], + "type": "text", + "content": " are the projection matrices. In this context, " + }, + { + "bbox": [ + 130, + 472, + 479, + 508 + ], + "type": "inline_equation", + "content": "^{\\prime}d^{\\prime}" + }, + { + "bbox": [ + 130, + 472, + 479, + 508 + ], + "type": "text", + "content": " represents the dimension of the feature embedding." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 479, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 479, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 479, + 557 + ], + "type": "text", + "content": "Unlike SAMUS's end-to-end CNN training, we use a static, RadImageNet [38] pre-trained ResNet-50 model. We add a trainable fully-connected layer, acting as an adapter, before its classification layer. The output features from both branches then go into our variational attention fusion block." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 580, + 336, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 580, + 336, + 592 + ], + "spans": [ + { + "bbox": [ + 131, + 580, + 336, + 592 + ], + "type": "text", + "content": "3.2 Variational Attention Fusion Block" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": "A major contribution we've made is the Variational Attention Fusion Block, which adeptly merges local CNN features and global ViT features. While SAMUS introduced a cross-branch attention module, our method, which models uncertainty around these features and applies variational attentional fusion, outperforms it (details in Sec 4.3). See Figure 3 for an overview." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Title Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 183, + 120, + 434, + 304 + ], + "blocks": [ + { + "bbox": [ + 183, + 120, + 434, + 304 + ], + "lines": [ + { + "bbox": [ + 183, + 120, + 434, + 304 + ], + "spans": [ + { + "bbox": [ + 183, + 120, + 434, + 304 + ], + "type": "image", + "image_path": "7a3215e96cf544e83a943dccf29de6aba509c0b8a7551bd11df0b21053c2921b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 319, + 482, + 376 + ], + "lines": [ + { + "bbox": [ + 131, + 319, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 131, + 319, + 482, + 376 + ], + "type": "text", + "content": "Fig. 3: Overview of the proposed Variational Attention Fusion Block. Each 'mode' has an intra-modal uncertainty learning encoder (represented as " + }, + { + "bbox": [ + 131, + 319, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 131, + 319, + 482, + 376 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 131, + 319, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^{\\prime}" + }, + { + "bbox": [ + 131, + 319, + 482, + 376 + ], + "type": "text", + "content": " in the figure), these obtain robust modality-specific features in the latent subspace. Subsequently, VAF combines these inputs and constructs a multimodal representation by estimating weights that are specific to each modality, effectively capturing their dependencies." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 399, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 399, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 131, + 399, + 482, + 485 + ], + "type": "text", + "content": "Uncertainty Learning Incorporating uncertainty learning, for an input image " + }, + { + "bbox": [ + 131, + 399, + 482, + 485 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 131, + 399, + 482, + 485 + ], + "type": "text", + "content": ", we derive features " + }, + { + "bbox": [ + 131, + 399, + 482, + 485 + ], + "type": "inline_equation", + "content": "s^v" + }, + { + "bbox": [ + 131, + 399, + 482, + 485 + ], + "type": "text", + "content": " from the ViT and " + }, + { + "bbox": [ + 131, + 399, + 482, + 485 + ], + "type": "inline_equation", + "content": "s^c" + }, + { + "bbox": [ + 131, + 399, + 482, + 485 + ], + "type": "text", + "content": " from the ResNet-50. Both " + }, + { + "bbox": [ + 131, + 399, + 482, + 485 + ], + "type": "inline_equation", + "content": "s^v" + }, + { + "bbox": [ + 131, + 399, + 482, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 399, + 482, + 485 + ], + "type": "inline_equation", + "content": "s^c \\in \\mathbb{R}^{d^h}" + }, + { + "bbox": [ + 131, + 399, + 482, + 485 + ], + "type": "text", + "content": " represent a fixed dimension in a latent space. We achieve a consistent size using latent encoders " + }, + { + "bbox": [ + 131, + 399, + 482, + 485 + ], + "type": "inline_equation", + "content": "E_v, E_c" + }, + { + "bbox": [ + 131, + 399, + 482, + 485 + ], + "type": "text", + "content": " at each network's end. To address data uncertainty, we design a latent distribution for each sample's content, helping capture semantic relations. This distribution is shaped as a parameterized diagonal Gaussian. This is mathematically represented as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 243, + 495, + 481, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 495, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 243, + 495, + 481, + 510 + ], + "type": "interline_equation", + "content": "p \\left(z ^ {v} \\mid s ^ {v}\\right) = \\mathcal {N} \\left(z ^ {v}; \\mu^ {v}, \\left(\\sigma^ {v}\\right) ^ {2} \\mathbf {I}\\right) \\tag {2}", + "image_path": "8193cfd8185d2ffa5df79dc3606d65b1457917add6c59254904aaa2937f4101b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 245, + 525, + 481, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 525, + 481, + 540 + ], + "spans": [ + { + "bbox": [ + 245, + 525, + 481, + 540 + ], + "type": "interline_equation", + "content": "p \\left(z ^ {c} \\mid s ^ {c}\\right) = \\mathcal {N} \\left(z ^ {c}; \\mu^ {c}, \\left(\\sigma^ {c}\\right) ^ {2} \\mathbf {I}\\right) \\tag {3}", + "image_path": "2cff435378496f65ece2702e524e1aabe58212205022edc0559b06115955f990.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 545, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 545, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 131, + 545, + 482, + 641 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 545, + 482, + 641 + ], + "type": "inline_equation", + "content": "z^v" + }, + { + "bbox": [ + 131, + 545, + 482, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 545, + 482, + 641 + ], + "type": "inline_equation", + "content": "z^c" + }, + { + "bbox": [ + 131, + 545, + 482, + 641 + ], + "type": "text", + "content": " are reconstructed vectors. The means, " + }, + { + "bbox": [ + 131, + 545, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mu^v, \\mu^c" + }, + { + "bbox": [ + 131, + 545, + 482, + 641 + ], + "type": "text", + "content": ", capture each mode's core feature, while variances, " + }, + { + "bbox": [ + 131, + 545, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\sigma^v, \\sigma^c" + }, + { + "bbox": [ + 131, + 545, + 482, + 641 + ], + "type": "text", + "content": ", indicate the noise-induced uncertainty in predicting these means. A higher variance signifies more uncertainty about the content observed. Both Gaussian parameters depend on the input and are forecasted through MLPs. For instance, for ViT features: " + }, + { + "bbox": [ + 131, + 545, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mu^v = f\\theta_{v1}(sv)" + }, + { + "bbox": [ + 131, + 545, + 482, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 545, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\sigma^v = f\\theta_{v2}(sv)" + }, + { + "bbox": [ + 131, + 545, + 482, + 641 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 131, + 545, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\theta_{v1}" + }, + { + "bbox": [ + 131, + 545, + 482, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 545, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\theta_{v2}" + }, + { + "bbox": [ + 131, + 545, + 482, + 641 + ], + "type": "text", + "content": " being respective model parameters. The same applies to CNN features. Now, each sample's feature representation shifts from deterministic to a stochastic Gaussian-drawn embedding in the latent space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "Due to the non-differentiable nature of sampling, we use a re-parameterization trick [30] to maintain gradient flow. We draw random noise " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": " from a normal dis" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 335, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 335, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 335, + 102 + ], + "type": "text", + "content": "Shreyank N Gowda and David A. Clifton" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 481, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 481, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 481, + 140 + ], + "type": "text", + "content": "ribution, independent of model parameters, and produce " + }, + { + "bbox": [ + 131, + 116, + 481, + 140 + ], + "type": "inline_equation", + "content": "z^v" + }, + { + "bbox": [ + 131, + 116, + 481, + 140 + ], + "type": "text", + "content": " using it, as detailed in Eq 4. We repeat this for " + }, + { + "bbox": [ + 131, + 116, + 481, + 140 + ], + "type": "inline_equation", + "content": "z^c" + }, + { + "bbox": [ + 131, + 116, + 481, + 140 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 249, + 152, + 481, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 152, + 481, + 166 + ], + "spans": [ + { + "bbox": [ + 249, + 152, + 481, + 166 + ], + "type": "interline_equation", + "content": "z ^ {v} = \\mu^ {v} + \\eta \\sigma^ {v}, \\eta \\in \\mathcal {N} (0, 1) \\tag {4}", + "image_path": "dd7267756f5fc3d4f176540ea20f262c2d1e432e14b103f6e31c9d63c0afbe3c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 181, + 479, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 181, + 479, + 264 + ], + "spans": [ + { + "bbox": [ + 131, + 181, + 479, + 264 + ], + "type": "text", + "content": "Variational Attentional Fusion CNNs and ViTs capture different feature aspects from images (local and global). We see these as two modes: " + }, + { + "bbox": [ + 131, + 181, + 479, + 264 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 131, + 181, + 479, + 264 + ], + "type": "text", + "content": " for ViTs and " + }, + { + "bbox": [ + 131, + 181, + 479, + 264 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 131, + 181, + 479, + 264 + ], + "type": "text", + "content": " for CNNs. Combining them isn't straightforward due to different confidence levels. Standard methods, which get weights from the data, miss the unique traits of each mode. To tackle this, we've developed a Variational Attention Fusion (VAF) module. This module aims to capture the nuances between modalities by determining modality-specific weights for a seamless integration." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 265, + 479, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 265, + 479, + 323 + ], + "spans": [ + { + "bbox": [ + 131, + 265, + 479, + 323 + ], + "type": "text", + "content": "In short, using features from each mode, a standard attention method creates a probability, " + }, + { + "bbox": [ + 131, + 265, + 479, + 323 + ], + "type": "inline_equation", + "content": "a^k" + }, + { + "bbox": [ + 131, + 265, + 479, + 323 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 131, + 265, + 479, + 323 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 265, + 479, + 323 + ], + "type": "text", + "content": " is either " + }, + { + "bbox": [ + 131, + 265, + 479, + 323 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 131, + 265, + 479, + 323 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 131, + 265, + 479, + 323 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 131, + 265, + 479, + 323 + ], + "type": "text", + "content": "), as shown in Eq 5. Here, " + }, + { + "bbox": [ + 131, + 265, + 479, + 323 + ], + "type": "inline_equation", + "content": "a^k" + }, + { + "bbox": [ + 131, + 265, + 479, + 323 + ], + "type": "text", + "content": " shows how much each mode contributes, and " + }, + { + "bbox": [ + 131, + 265, + 479, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_m" + }, + { + "bbox": [ + 131, + 265, + 479, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 265, + 479, + 323 + ], + "type": "inline_equation", + "content": "b_m" + }, + { + "bbox": [ + 131, + 265, + 479, + 323 + ], + "type": "text", + "content": " are trainable factors. By having set weights, it picks the right features without being affected by varied confidence levels between modes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 282, + 343, + 361, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 343, + 361, + 356 + ], + "spans": [ + { + "bbox": [ + 282, + 343, + 361, + 356 + ], + "type": "interline_equation", + "content": "\\hat {a} ^ {k} = \\mathbf {W} _ {m} \\mathbf {z} ^ {k} + b _ {m}", + "image_path": "ee4f2c02229e040a8f671f548f6831d1016cb01e6ff19ec56026ff2ba5a90dd9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 254, + 359, + 480, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 359, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 254, + 359, + 480, + 388 + ], + "type": "interline_equation", + "content": "a ^ {k} = \\frac {\\exp \\left(\\hat {a} ^ {k}\\right)}{\\sum_ {k ^ {\\prime} \\in \\{v , c \\}} \\exp \\left(\\hat {a} ^ {k ^ {\\prime}}\\right)} \\tag {5}", + "image_path": "031a1796dc36e693260a47648a9b5ca3e7fedd4e29a63855a368104b73272723.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 395, + 479, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 395, + 479, + 431 + ], + "spans": [ + { + "bbox": [ + 131, + 395, + 479, + 431 + ], + "type": "text", + "content": "Our VAF module uses variational attention weights instead of the point-estimated attention vector " + }, + { + "bbox": [ + 131, + 395, + 479, + 431 + ], + "type": "inline_equation", + "content": "\\hat{a}^k" + }, + { + "bbox": [ + 131, + 395, + 479, + 431 + ], + "type": "text", + "content": ". This approach, grounded in a probability distribution (Eq. 6), better accounts for uncertainties across modalities." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 206, + 439, + 480, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 439, + 480, + 459 + ], + "spans": [ + { + "bbox": [ + 206, + 439, + 480, + 459 + ], + "type": "interline_equation", + "content": "\\hat {a} ^ {k} \\sim q _ {\\theta} (\\hat {a} ^ {k} \\mid z ^ {k}) = \\mathcal {N} \\left(\\mu_ {a} ^ {k}, \\left(\\sigma_ {a} ^ {k}\\right) ^ {2} \\mathbf {I}\\right), k \\in \\{v, c \\} \\tag {6}", + "image_path": "dcce2406ebec86bb6a9875de7d21b8759c4378fbdf6305ac9f900d267217dd16.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 463, + 481, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 463, + 481, + 558 + ], + "spans": [ + { + "bbox": [ + 131, + 463, + 481, + 558 + ], + "type": "text", + "content": "We used an identity transformation to preserve modality traits in the variational attention. Mean and variance parameters are predicted by MLPs from the input. The " + }, + { + "bbox": [ + 131, + 463, + 481, + 558 + ], + "type": "inline_equation", + "content": "\\left(\\sigma_{a}^{k}\\right)" + }, + { + "bbox": [ + 131, + 463, + 481, + 558 + ], + "type": "text", + "content": " value indicates confidence levels among modalities. With the VAF method, more confident modality features are enhanced, while less confident ones are reduced, optimizing multimodal data fusion and capturing complementary features. The final representation combines modality-specific outputs, " + }, + { + "bbox": [ + 131, + 463, + 481, + 558 + ], + "type": "inline_equation", + "content": "z_{c}" + }, + { + "bbox": [ + 131, + 463, + 481, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 463, + 481, + 558 + ], + "type": "inline_equation", + "content": "z_{v}" + }, + { + "bbox": [ + 131, + 463, + 481, + 558 + ], + "type": "text", + "content": ", using weighted aggregation as shown in Eq 7, where " + }, + { + "bbox": [ + 131, + 463, + 481, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_h" + }, + { + "bbox": [ + 131, + 463, + 481, + 558 + ], + "type": "text", + "content": " is a learnable weight matrix." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 263, + 568, + 481, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 568, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 263, + 568, + 481, + 594 + ], + "type": "interline_equation", + "content": "h = \\sum_ {k \\in \\{v, c \\}} a ^ {k} \\mathbf {W} _ {h} z ^ {k} \\tag {7}", + "image_path": "33fd41db84ad09fa64481dea937049b8bba988494ed08f8056cb2639eb3b732f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 610, + 309, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 610, + 309, + 623 + ], + "spans": [ + { + "bbox": [ + 131, + 610, + 309, + 623 + ], + "type": "text", + "content": "3.3 Guiding the Prompt Encoder" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 629, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 479, + 666 + ], + "type": "text", + "content": "SAM's mask decoder [31] needs " + }, + { + "bbox": [ + 131, + 629, + 479, + 666 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 131, + 629, + 479, + 666 + ], + "type": "text", + "content": " and an input from the prompt encoder. While past adaptations of SAM for medical imaging used point prompts, our results show that a reasonably good bounding box enhances performance. Creating this" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Title Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "text", + "content": "bounding box requires a dedicated model. Using random bounding boxes may hinder results, necessitating specific prompts. We employ Grounding-DINO [34] for bounding box generation. Though not exclusively trained for medical images, Grounding-DINO is an effective object detector. We use GPT-4 [39] to craft text prompts for specific class labels, and then MedBERT [44] generates an embedding input for Grounding-DINO. Whilst this does improve our performance in comparison to using point prompts, we show in supplementary that using the same set of random point prompts, we outperform all other medical foundational models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 224, + 483, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 224, + 483, + 356 + ], + "spans": [ + { + "bbox": [ + 131, + 224, + 483, + 356 + ], + "type": "text", + "content": "Input to GPT-4 \"Create a concise description for a medical image analysis task. The task involves using ultrasound images to segment and identify specific anatomical structures or pathologies. Use [target] for to make a generic description that can be adapted for any segmentation target.\" Output from GPT-4 \"Segment and identify [target] within ultrasound images. This requires the precise delineation of [target] from surrounding tissues, enabling accurate diagnosis and assessment. The challenge encompasses dealing with the inherent variability of ultrasound image quality, including speckle noise, shadowing, and the wide range of anatomical variations among patients. Success in this task is measured by the algorithm's ability to consistently and accurately identify and outline the [target], providing crucial information for medical decision-making processes.\" Where [target] is modified for each dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 369, + 231, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 231, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 231, + 380 + ], + "type": "text", + "content": "3.4 Loss Function" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 386, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 386, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 131, + 386, + 482, + 411 + ], + "type": "text", + "content": "Our overall loss function can be seen in Eq 8. Here, " + }, + { + "bbox": [ + 131, + 386, + 482, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{BCE}" + }, + { + "bbox": [ + 131, + 386, + 482, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 386, + 482, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 131, + 386, + 482, + 411 + ], + "type": "text", + "content": " are segmentation losses: binary cross entropy and dice loss respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 219, + 422, + 480, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 422, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 219, + 422, + 480, + 434 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {B C E} + \\mathcal {L} _ {D} + \\lambda_ {c} \\mathcal {L} _ {c} + \\lambda_ {v} \\mathcal {L} _ {v} + \\lambda_ {a} \\mathcal {L} _ {a} \\tag {8}", + "image_path": "221d01a3a9a05613e1dce6bdbc2a5a001507e4b8eef6f3b87e6c79af034c0683.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 437, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{L}c" + }, + { + "bbox": [ + 130, + 437, + 482, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 437, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{L}v" + }, + { + "bbox": [ + 130, + 437, + 482, + 538 + ], + "type": "text", + "content": " are regularization terms for uncertainty learning, inspired by the variational information bottleneck [2]. To grasp uncertainty in a modality, a component is introduced during learning. It ensures the data's distribution, " + }, + { + "bbox": [ + 130, + 437, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{N}\\left(z^v;\\mu^v,(\\sigma^v)^2\\right)" + }, + { + "bbox": [ + 130, + 437, + 482, + 538 + ], + "type": "text", + "content": ", resembles a standard bell curve, " + }, + { + "bbox": [ + 130, + 437, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\epsilon ;0,\\mathbf{I})" + }, + { + "bbox": [ + 130, + 437, + 482, + 538 + ], + "type": "text", + "content": ". This similarity is gauged using Kullback-Leibler divergence (KLD) between the distributions, promoting model diversity and reducing uncertainty, leading to robust features. Eq 9 shows how " + }, + { + "bbox": [ + 130, + 437, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{L}v" + }, + { + "bbox": [ + 130, + 437, + 482, + 538 + ], + "type": "text", + "content": " is computed, with analogous methods for " + }, + { + "bbox": [ + 130, + 437, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{L}c" + }, + { + "bbox": [ + 130, + 437, + 482, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 437, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_a" + }, + { + "bbox": [ + 130, + 437, + 482, + 538 + ], + "type": "text", + "content": ", replacing " + }, + { + "bbox": [ + 130, + 437, + 482, + 538 + ], + "type": "inline_equation", + "content": "z^{v}" + }, + { + "bbox": [ + 130, + 437, + 482, + 538 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 437, + 482, + 538 + ], + "type": "inline_equation", + "content": "z^c" + }, + { + "bbox": [ + 130, + 437, + 482, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 437, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\hat{a}^k" + }, + { + "bbox": [ + 130, + 437, + 482, + 538 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 217, + 547, + 480, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 547, + 480, + 591 + ], + "spans": [ + { + "bbox": [ + 217, + 547, + 480, + 591 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {v} = K L \\left(\\mathcal {N} \\left(z ^ {v}; \\mu^ {v}, (\\sigma^ {v}) ^ {2}\\right) \\| \\mathcal {N} (\\epsilon ; 0, \\mathbf {I})\\right) \\\\ = - \\frac {1}{2} \\left(1 + \\log (\\sigma^ {v}) ^ {2} - (\\mu^ {v}) ^ {2} - (\\sigma^ {v}) ^ {2}\\right) \\tag {9} \\\\ \\end{array}", + "image_path": "0f17a33a65375962eeb42513fa63259a9b6300533c670ea02d610d513d9c1987.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 603, + 289, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 603, + 289, + 617 + ], + "spans": [ + { + "bbox": [ + 132, + 603, + 289, + 617 + ], + "type": "text", + "content": "4 Experimental Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 625, + 206, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 625, + 206, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 625, + 206, + 636 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "To compare with SAMUS [33], we test CC-SAM on the same seven public datasets: TN3K [14], DDTI [41], TG3K [52], BUSI [1], UDIAT [3], CAMUS [32]," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 335, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 335, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 335, + 102 + ], + "type": "text", + "content": "Shreyank N Gowda and David A. Clifton" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": "and HMC-QU [8]. We also recreate the US30K dataset from SAMUS for broader comparison. Dataset details are in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 479, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 479, + 187 + ], + "type": "text", + "content": "TN3K and TG3K datasets follow TRFE [14] for segmentation into training, validation, and testing. BUSI is randomly split in a 7:1:2 ratio for these purposes. CAMUS is initially divided into training and testing per the challenge [32], with " + }, + { + "bbox": [ + 130, + 140, + 479, + 187 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 140, + 479, + 187 + ], + "type": "text", + "content": " of training data later used for validation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 189, + 480, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 480, + 271 + ], + "type": "text", + "content": "For model generalizability, US30K datasets remain unseen during training and validation. In line with SAMUS, we split US30k into seen and unseen datasets. We then evaluate on unseen datasets (DDTI, UDIAT, HMC-QU) not used in training or validation. This enables direct comparison of generalization with other state-of-the-art models. Moreover, we contrast CC-SAM with other base models by training on the full US30K dataset and testing across different tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 293, + 280, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 293, + 280, + 306 + ], + "spans": [ + { + "bbox": [ + 131, + 293, + 280, + 306 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 315, + 479, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 479, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 479, + 387 + ], + "type": "text", + "content": "We follow SAMUS in setting hyper parameters and adapter sizes. We change the fully tuneable CNN in SAMUS to a static ResNet50. The model is trained using the Adam optimizer with an initial learning rate of 0.01 that reduces by a tenth every 50 epochs for a total of 200 epochs. The encoders are fully connected networks with the latent space having dimensions 1024. All fully connected layers has a Leaky RELU activation function and a dropout probability of 0.5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 388, + 480, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 480, + 422 + ], + "type": "text", + "content": "We use 1xA100 GPU with 40 GB capacity. But we also verified that a single 3090 Ti with 24 GB memory is enough to train the model. We set a batch size of 32." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 444, + 238, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 444, + 238, + 456 + ], + "spans": [ + { + "bbox": [ + 131, + 444, + 238, + 456 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 466, + 479, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 466, + 479, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 466, + 479, + 561 + ], + "type": "text", + "content": "We consider SAMUS [33] as our baseline model and ablate our added components. Specifically, we show replacing the end-to-end CNN model with a fixed pre-trained model improves performance. Further adding the FC layer at the end slightly improves performance. Crucially, replacing the cross attention fusion with our variational attention fusion significantly improves performance. Further, replacing the point based prompting with our text-based prompting gives us the best results. We show these results on two of the datasets TN3K and BUSI in Table 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 563, + 479, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 479, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 479, + 586 + ], + "type": "text", + "content": "We do a further analysis of how the prompting compares with different prompting strategies in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 607, + 326, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 607, + 326, + 619 + ], + "spans": [ + { + "bbox": [ + 131, + 607, + 326, + 619 + ], + "type": "text", + "content": "4.4 Comparison with state-of-the-art" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 479, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 664 + ], + "type": "text", + "content": "We consider two scenarios for comparing with state-of-the-art models following SAMUS [33]. We compare CC-SAM with task-specific methods and foundational models." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Title Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 192, + 114, + 421, + 225 + ], + "blocks": [ + { + "bbox": [ + 192, + 114, + 421, + 225 + ], + "lines": [ + { + "bbox": [ + 192, + 114, + 421, + 225 + ], + "spans": [ + { + "bbox": [ + 192, + 114, + 421, + 225 + ], + "type": "table", + "html": "
ComponentsTN3KBUSI
ResNet50VAFText PromptDiceHDDiceHD
×××84.4528.2285.7725.49
✓××84.9127.2586.2225.12
×✓×85.5526.8487.1524.42
××✓85.9526.2287.9223.92
✓×✓86.4225.8288.4523.20
×✓✓86.2525.9288.2223.51
✓✓×85.7226.2187.4224.61
✓✓✓87.1125.2589.5122.86
", + "image_path": "d2a82fd8cf37d91e0e415517b30d202697ef65f0d2528b98c15818c02ba66559.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 289, + 480, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 480, + 374 + ], + "type": "text", + "content": "Task-specific methods Following SAMUS [33], thirteen state-of-the-art, task-specific techniques are chosen for evaluation, spanning CNN-based, transformer-based, and CNN-Transformer hybrid methods. Methods other than SAMUS include CNN-based techniques encompass U-Net [45], CPFNet [11], CA-Net [20], CE-Net [21], and AAU-Net [5]. The transformer-centric methods consist of SwinUnet [4], SETR [55], and MISSFormer [28]. The hybrids of CNN and Transformer are TransUNet [6], TransFuse [54], FAT-Net [50], and H2Former [23]." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 141, + 403, + 473, + 570 + ], + "blocks": [ + { + "bbox": [ + 132, + 226, + 479, + 247 + ], + "lines": [ + { + "bbox": [ + 132, + 226, + 479, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 226, + 479, + 247 + ], + "type": "text", + "content": "Table 1: Ablation study of adding different components of our proposed approach. The first row corresponds to the baseline SAMUS [33]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 403, + 473, + 570 + ], + "lines": [ + { + "bbox": [ + 141, + 403, + 473, + 570 + ], + "spans": [ + { + "bbox": [ + 141, + 403, + 473, + 570 + ], + "type": "table", + "html": "
MethodCAMUS-LATN3KBUSICAMUS-LVCAMUS-MYO
DiceHDDiceHDDiceHDDiceHDDiceHD
U-Net91.0012.9179.0134.1278.1133.6093.569.9086.8616.87
CPFNet91.5112.2679.4333.0780.5627.9893.329.6386.6816.51
CA-Net91.2812.3280.5233.6581.8828.6793.599.7787.2116.24
CE-Net91.1412.2980.3732.7980.2130.1993.369.9186.4716.66
AAU-Net91.3312.1282.2830.5380.8130.3993.329.9786.9816.49
SwinUNet89.8014.7470.0844.1367.2347.0291.7212.8084.4620.25
SETR90.5213.9167.8044.1168.2240.3792.8211.3486.2018.27
MISSFormer91.1811.8279.4232.8578.4333.1093.259.9486.5716.50
TransUNet91.3712.4681.4430.9882.2227.5493.609.6087.2017.25
FAT-Net91.5512.0580.4532.7782.1628.5593.599.2087.1915.93
H2Former90.9811.9282.4830.5881.4827.8493.449.6087.3116.60
SAMUS91.5811.6084.4528.2285.7725.4993.739.7987.4616.74
CC-SAM92.0311.1185.2027.1087.0124.2293.959.1188.2516.11
", + "image_path": "34116519db88cf85f8d4fd1a5d906eb8696653c5695140778d51743c930ec764.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 571, + 480, + 625 + ], + "lines": [ + { + "bbox": [ + 130, + 571, + 480, + 625 + ], + "spans": [ + { + "bbox": [ + 130, + 571, + 480, + 625 + ], + "type": "text", + "content": "Table 2: Quantitative comparison between our CC-SAM method and the state-of-the-art (SOTA) task-specific techniques for segmenting thyroid nodules (TN3K), breast cancer (BUSI), left ventricle (CAMUS-LV), myocardium (CAMUS-MYO), and left atrium (CAMUS-LA). We assessed the performance using the Dice score (\\%) and the Hausdorff distance (HD). The top-performing results are highlighted in bold." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 335, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 335, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 335, + 102 + ], + "type": "text", + "content": "Shreyank N Gowda and David A. Clifton" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 479, + 255 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 479, + 255 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 479, + 255 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 479, + 255 + ], + "type": "image", + "image_path": "c27111f42dfcdcd089e6bd03ffb33716acc56e2b4237d47255da2b8cce2cc3cd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 264, + 480, + 286 + ], + "lines": [ + { + "bbox": [ + 132, + 264, + 480, + 286 + ], + "spans": [ + { + "bbox": [ + 132, + 264, + 480, + 286 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparison between our CC-SAM method and the state-of-the-art (SOTA) task-specific techniques." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 308, + 480, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 480, + 380 + ], + "type": "text", + "content": "Quantitative Analysis. Table 2 summarizes the numerical outcomes of various task-specific methods on TN3K, BUSI, CAMUS-LV, CAMUS-MYO, and CAMUS-LA. In comparison, CC-SAM consistently excels in performance across all these five tasks, setting a new benchmark for SOTA results. While CC-SAM surpasses recent SOTA in most benchmarks, it does fall short by a very small margin when evaluated with the Hausdorff distance (HD) on CAMUS-MYO." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "text", + "content": "Qualitative Analysis. Figure 4 showcases the segmentation results from various methods, including U-Net [45], AAU-Net [5], MISSFormer [28], H2Former [23], SAMUS [33], and our proposed CC-SAM. Segmenting ultrasound images is notably challenging given their low contrast, inconsistent features, and indistinct object boundaries. Many existing techniques face difficulties in precisely differentiating the target from its background, often resulting in significant false negatives or false positives. In contrast, SAMUS stands out in maintaining the integrity of target areas and minimizing false positives. Our CC-SAM further refines the outcomes observed with SAMUS, a testament to the intrinsic strengths of SAM and the tailored modifications and innovations integrated into CC-SAM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 518, + 482, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 518, + 482, + 625 + ], + "spans": [ + { + "bbox": [ + 130, + 518, + 482, + 625 + ], + "type": "text", + "content": "Generalization Ability. Evaluating the generalization performance of various task-specific methods is crucial as it provides insights into how well these methods can adapt and function across unfamiliar or unseen datasets. Figure 5 offers a quantitative assessment of this aspect. When we assess performance discrepancies between known (seeable) and unknown (unseen) datasets, CC-SAM stands out by displaying the most minimal performance drop among all methods across the three segmentation tasks. This underlines the remarkable generalization prowess of CC-SAM, emphasizing its resilience and versatility in a variety of medical image segmentation situations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "Foundational models We also compare with five state-of-the-art foundational models. These include the original SAM [31], SAMed [53], SAMUS [33], MSA [51]" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Title Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 251 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 251 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 251 + ], + "type": "image", + "image_path": "e6c2172cc496fc37ea44f667a21ddae0b7d5fa77b2d207aa37c1f60a37059481.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 261, + 482, + 297 + ], + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 297 + ], + "type": "text", + "content": "Fig. 5: Comparison of CC-SAM with task-specific techniques on see datasets (highlighted in blue) and unseen datasets not previously encountered (indicated in orange). Higher orange bars indicate stronger generalization ability." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 319, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 368 + ], + "type": "text", + "content": "and MedSAM [37]. SAM is trained on SA-1B and kept frozen due to the cost involved in fine-tuning. All other foundational models are pre-trained on the US30k dataset and evaluated using Dice and HD across TN3K, BUSI, CAMUS-LV, CAMUS-MYO, and CAMUS-LA datasets." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 141, + 399, + 473, + 491 + ], + "blocks": [ + { + "bbox": [ + 141, + 399, + 473, + 491 + ], + "lines": [ + { + "bbox": [ + 141, + 399, + 473, + 491 + ], + "spans": [ + { + "bbox": [ + 141, + 399, + 473, + 491 + ], + "type": "table", + "html": "
MethodTN3KBUSICAMUS-LVCAMUS-MYOCAMUS-LA
DiceHDDiceHDDiceHDDiceHDDiceHD
SAM29.59134.8754.0182.3928.18196.6429.42184.1017.28193.70
MedSAM71.0942.9177.7534.2687.5215.2876.0725.7288.0615.70
SAMed80.4031.2974.8234.6087.6713.2482.6019.4890.9212.60
MSA82.6729.1581.6628.8790.9511.2982.4719.2891.8011.59
SAMUS83.0528.8284.5427.2491.1311.7683.1118.9992.0012.08
CC-SAM85.5927.7486.2225.8592.8510.8885.6117.1193.5111.06
", + "image_path": "00b9173c6e9e8d926afb22a3923af20df3bad588a07d885ae4b7a6336f94b188.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 491, + 482, + 523 + ], + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 523 + ], + "type": "text", + "content": "Table 3: Quantitative comparison of our CC-SAM and other foundation models on seeable US30K data. The performance is evaluated by the Dice score (\\%) and Hausdorff distance (HD)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 557, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 605 + ], + "type": "text", + "content": "Quantitative Analysis. Table 3 summarizes the numerical outcomes of various foundational methods on TN3K, BUSI, CAMUS-LV, CAMUS-MYO, and CAMUS-LA. In comparison, CC-SAM consistently excels in performance across all tasks, setting a new benchmark for foundational models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Qualitative Analysis. Figure 6 shows the qualitative segmentation results from various foundational models, including SAM, MedSAM, SAMed, MSA, SAMUS, and CC-SAM. SAM, without adjustments for medical images, fails at segmentation. However, MedSAM, SAMed, and MSA, through tuning, recover some" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 335, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 335, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 335, + 102 + ], + "type": "text", + "content": "Shreyank N Gowda and David A. Clifton" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "segmentation ability but struggle with accurately defining borders, resulting in many false positives and negatives. In contrast, SAMUS excels in identifying precise segmentation borders, even in low contrast, suggesting the benefit of integrating local data with the image encoder for better boundary and shape retention in medical image segmentation. CC-SAM goes further, improving upon SAMUS's performance and setting new benchmarks." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 214, + 479, + 352 + ], + "blocks": [ + { + "bbox": [ + 134, + 214, + 479, + 352 + ], + "lines": [ + { + "bbox": [ + 134, + 214, + 479, + 352 + ], + "spans": [ + { + "bbox": [ + 134, + 214, + 479, + 352 + ], + "type": "image", + "image_path": "413bb74e2354fc947cc7cceeecef1c37197b7be244f7afd8d3182c98c435eaba.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 362, + 480, + 384 + ], + "lines": [ + { + "bbox": [ + 132, + 362, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 362, + 480, + 384 + ], + "type": "text", + "content": "Fig. 6: Qualitative comparison between our CC-SAM method and the state-of-the-art (SOTA) foundational medical segmentation models." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 436, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 581 + ], + "type": "text", + "content": "Generalization Ability. Generally, foundation models trained on US30K exhibit a much improved generalization performance in medical image segmentation tasks compared to the original SAM. When considering the three specific segmentation tasks: thyroid nodule segmentation, breast cancer segmentation, and myocardium segmentation, it's clear that all foundation models face significant challenges with myocardium segmentation, while they perform relatively well in the breast cancer segmentation task. CC-SAM outperforms all foundational models significantly. This highlights the remarkable generalization prowess of CC-SAM, as it consistently surpasses other foundation models, especially in unfamiliar domains. Figure 7 offers a quantitative assessment of this aspect. We see that the original SAM performs the worst, which is expected and the proposed CC-SAM obtains a new state-of-the-art on all datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Deployment Cost Comparison with Other Foundational Models We compare CCSAM with other foundational models in a more detailed manner. We look at 'Generalization Ability/Dice " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " ,GPU Memory/G, Parameters/M,Computation/GFLOPs', Inference Speed/FPS' and Segmentation Performance/Dice " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " .The spider plot of this can be seen in Figure 8." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Title Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 123, + 294, + 242 + ], + "blocks": [ + { + "bbox": [ + 135, + 123, + 294, + 242 + ], + "lines": [ + { + "bbox": [ + 135, + 123, + 294, + 242 + ], + "spans": [ + { + "bbox": [ + 135, + 123, + 294, + 242 + ], + "type": "image", + "image_path": "73ae6c8b6854707c443c0576ab63864cad3cfcc815e2d268f2848e3cb3894f50.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 254, + 298, + 319 + ], + "lines": [ + { + "bbox": [ + 130, + 254, + 298, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 298, + 319 + ], + "type": "text", + "content": "Fig.7: Comparing segmentation and generalization ability comparison of our CC-SAMUS and other foundation models. Whilst SAM obtains the lowest Dice Score, CC-SAM obtains the highest across all datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 313, + 114, + 478, + 274 + ], + "blocks": [ + { + "bbox": [ + 313, + 114, + 478, + 274 + ], + "lines": [ + { + "bbox": [ + 313, + 114, + 478, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 478, + 274 + ], + "type": "image", + "image_path": "b0fc5e0214f75dcdc52bb87a7fa5db9da34ae8e7cda733d60ee273229c44cc0d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 293, + 482, + 326 + ], + "lines": [ + { + "bbox": [ + 307, + 293, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 293, + 482, + 326 + ], + "type": "text", + "content": "Fig. 8: Spider plot comparing efficiency and segmentation performance of all foundational models." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 354, + 220, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 220, + 367 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 220, + 367 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 391, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 559 + ], + "type": "text", + "content": "The Segment Anything Model (SAM), while proficient in natural image segmentation, faced hurdles in medical imaging due to issues with low contrast, faint boundaries, and other complexities. In this study, we identified and addressed the limitations of the Segment Anything Model (SAM) when applied to medical imaging, especially with images that have inherent challenges like low contrast and intricate morphologies. Our enhanced approach integrates a CNN branch with SAM's original Vision Transformer (ViT) encoder, utilizing a unique variational attention fusion module to capture vital local spatial information present in medical images. By introducing feature and position adapters within the ViT branch, the model's representations are further refined. Notably, we found that using text prompts, especially those generated leveraging ChatGPT's capabilities, significantly boosts SAM's performance in ultrasound medical segmentation. These innovations mark a pivotal step towards improving the adaptability and efficiency of universal image segmentation models in the medical field." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Acknowledgements DAC was supported by the Pandemic Sciences Institute at the University of Oxford; the National Institute for Health Research (NIHR) Oxford Biomedical Research Centre (BRC); an NIHR Research Professorship; a Royal Academy of Engineering Research Chair; the Wellcome Trust funded VITAL project (grant 204904/Z/16/Z); the EPSRC (grant EP/W031744/1); and the InnoHK Hong Kong Centre for Cerebro-cardiovascular Engineering (COCHE)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 335, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 335, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 335, + 102 + ], + "type": "text", + "content": "Shreyank N Gowda and David A. Clifton" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 175 + ], + "type": "text", + "content": "1. Al-Dhabyani, W., Gomaa, M., Khaled, H., Aly, F.: Deep learning approaches for data augmentation and classification of breast masses using ultrasound images. Int. J. Adv. Comput. Sci. Appl 10(5), 1-11 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 176, + 480, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 480, + 198 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 480, + 198 + ], + "type": "text", + "content": "2. Alemi, A.A., Fischer, I., Dillon, J.V., Murphy, K.: Deep variational information bottleneck. arXiv preprint arXiv:1612.00410 (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 199, + 480, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 199, + 480, + 242 + ], + "spans": [ + { + "bbox": [ + 138, + 199, + 480, + 242 + ], + "type": "text", + "content": "3. Byra, M., Jarosik, P., Szubert, A., Galperin, M., Ojeda-Fournier, H., Olson, L., O'Boyle, M., Comstock, C., Andre, M.: Breast mass segmentation in ultrasound with selective kernel u-net convolutional neural network. Biomedical Signal Processing and Control 61, 102027 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 243, + 480, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 480, + 275 + ], + "type": "text", + "content": "4. Cao, H., Wang, Y., Chen, J., Jiang, D., Zhang, X., Tian, Q., Wang, M.: Swinunet: Unet-like pure transformer for medical image segmentation. In: European conference on computer vision. pp. 205-218. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 277, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 480, + 308 + ], + "type": "text", + "content": "5. Chen, G., Li, L., Dai, Y., Zhang, J., Yap, M.H.: Aau-net: an adaptive attention u-net for breast lesions segmentation in ultrasound images. IEEE Transactions on Medical Imaging (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 310, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 310, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 310, + 480, + 342 + ], + "type": "text", + "content": "6. Chen, J., Lu, Y., Yu, Q., Luo, X., Adeli, E., Wang, Y., Lu, L., Yuille, A.L., Zhou, Y.: Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 344, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 344, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 344, + 480, + 374 + ], + "type": "text", + "content": "7. Chen, L., Bentley, P., Mori, K., Misawa, K., Fujiwara, M., Rueckert, D.: Drinet for medical image segmentation. IEEE transactions on medical imaging 37(11), 2453-2462 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 376, + 480, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 480, + 419 + ], + "type": "text", + "content": "8. Degerli, A., Kiranyaz, S., Hamid, T., Mazhar, R., Gabbouj, M.: Early myocardial infarction detection over multi-view echocardiography. Biomedical Signal Processing and Control 87, 105448 (2024). https://doi.org/https://doi.org/10.1016/j.bspc.2023.105448" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 421, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 421, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 421, + 480, + 464 + ], + "type": "text", + "content": "9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 466, + 480, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 466, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 466, + 480, + 498 + ], + "type": "text", + "content": "0. Ermis, B., Zappella, G., Wistuba, M., Rawal, A., Archambeau, C.: Continual learning with transformers for image classification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3774-3781 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 499, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 499, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 138, + 499, + 480, + 531 + ], + "type": "text", + "content": "1. Feng, S., Zhao, H., Shi, F., Cheng, X., Wang, M., Ma, Y., Xiang, D., Zhu, W., Chen, X.: Cpfnet: Context pyramid fusion network for medical image segmentation. IEEE transactions on medical imaging 39(10), 3008-3018 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 533, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 480, + 564 + ], + "type": "text", + "content": "2. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision pp. 1-15 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 566, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 566, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 138, + 566, + 480, + 598 + ], + "type": "text", + "content": "3. Gao, P., Han, J., Zhang, R., Lin, Z., Geng, S., Zhou, A., Zhang, W., Lu, P., He, C., Yue, X., et al.: Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "text", + "content": "4. Gong, H., Chen, J., Chen, G., Li, H., Li, G., Chen, F.: Thyroid region prior guided attention for ultrasound segmentation of thyroid nodules. Computers in Biology and Medicine 155, 106389 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "5. Gowda, S.N.: Human activity recognition using combinatorial deep belief networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 1-6 (2017)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Title Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "16. Gowda, S.N., Arnab, A., Huang, J.: Optimizing vivit training: Time and memory reduction for action recognition. arXiv preprint arXiv:2306.04822 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "text", + "content": "17. Gowda, S.N., Gao, B., Clifton, D.: Fe-adapter: adapting image-based emotion classifiers to videos (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 160, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 480, + 191 + ], + "type": "text", + "content": "18. Gowda, S.N., Rohrbach, M., Sevilla-Lara, L.: Smart frame selection for action recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 1451-1459 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 192, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 480, + 235 + ], + "type": "text", + "content": "19. Gowda, S.N., Yuan, C.: Colornet: Investigating the importance of color spaces for image classification. In: Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part IV 14. pp. 581–596. Springer (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 278 + ], + "type": "text", + "content": "20. Gu, R., Wang, G., Song, T., Huang, R., Aertsen, M., Deprest, J., Ourselin, S., Vercauteren, T., Zhang, S.: Ca-net: Comprehensive attention convolutional neural networks for explainable medical image segmentation. IEEE transactions on medical imaging 40(2), 699-711 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 279, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 480, + 311 + ], + "type": "text", + "content": "21. Gu, Z., Cheng, J., Fu, H., Zhou, K., Hao, H., Zhao, Y., Zhang, T., Gao, S., Liu, J.: Ce-net: Context encoder network for 2d medical image segmentation. IEEE transactions on medical imaging 38(10), 2281-2292 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 312, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 480, + 354 + ], + "type": "text", + "content": "22. Hatamizadeh, A., Tang, Y., Nath, V., Yang, D., Myronenko, A., Landman, B., Roth, H.R., Xu, D.: Unetr: Transformers for 3d medical image segmentation. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 574-584 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 354, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 386 + ], + "type": "text", + "content": "23. He, A., Wang, K., Li, T., Du, C., Xia, S., Fu, H.: H2former: An efficient hierarchical hybrid transformer for medical image segmentation. IEEE Transactions on Medical Imaging (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 387, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 480, + 418 + ], + "type": "text", + "content": "24. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 419, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 480, + 450 + ], + "type": "text", + "content": "25. He, S., Bao, R., Li, J., Grant, P.E., Ou, Y.: Accuracy of segment-anything model (sam) in medical image segmentation tasks. arXiv preprint arXiv:2304.09324 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "type": "text", + "content": "26. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 483, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 515 + ], + "type": "text", + "content": "27. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 516, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 516, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 516, + 480, + 536 + ], + "type": "text", + "content": "28. Huang, X., Deng, Z., Li, D., Yuan, X.: Missformer: An effective medical image segmentation transformer. arXiv preprint arXiv:2109.07162 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "text", + "content": "29. Huang, Y., Yang, X., Liu, L., Zhou, H., Chang, A., Zhou, X., Chen, R., Yu, J., Chen, J., Chen, C., et al.: Segment anything model for medical images? arXiv preprint arXiv:2304.14660 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 480, + 590 + ], + "type": "text", + "content": "30. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "type": "text", + "content": "31. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "text", + "content": "32. Leclerc, S., Smistad, E., Pedrosa, J., Østvik, A., Cervenansky, F., Espinosa, F., Espeland, T., Berg, E.A.R., Jodoin, P.M., Grenier, T., et al.: Deep learning for segmentation using an open large-scale dataset in 2d echocardiography. IEEE transactions on medical imaging 38(9), 2198-2210 (2019)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 334, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 334, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 334, + 101 + ], + "type": "text", + "content": "Shreyank N Gowda and David A. Clifton" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "33. Lin, X., Xiang, Y., Zhang, L., Yang, X., Yan, Z., Yu, L.: Samus: Adapting segment anything model for clinically-friendly and generalizable ultrasound image segmentation. arXiv preprint arXiv:2309.06824 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 184 + ], + "type": "text", + "content": "34. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 481, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 481, + 206 + ], + "type": "text", + "content": "35. Liu, X., Song, L., Liu, S., Zhang, Y.: A review of deep-learning-based medical image segmentation methods. Sustainability 13(3), 1224 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 207, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 481, + 251 + ], + "type": "text", + "content": "36. Luo, Z., Hu, Z., Xi, Y., Zhang, R., Ma, J.: I-tuning: Tuning frozen language models with image for lightweight image captioning. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 274 + ], + "type": "text", + "content": "37. Ma, J., Wang, B.: Segment anything in medical images. arXiv preprint arXiv:2304.12306 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 319 + ], + "type": "text", + "content": "38. Mei, X., Liu, Z., Robson, P.M., Marinelli, B., Huang, M., Doshi, A., Jacobi, A., Cao, C., Link, K.E., Yang, T., et al.: Radimagenet: an open radiologic deep learning research dataset for effective transfer learning. Radiology: Artificial Intelligence 4(5), e210315 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 319, + 307, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 307, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 307, + 330 + ], + "type": "text", + "content": "39. OpenAI: Gpt-4 technical report (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 331, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 331, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 133, + 331, + 481, + 363 + ], + "type": "text", + "content": "40. Pan, J., Lin, Z., Zhu, X., Shao, J., Li, H.: St-adapter: Parameter-efficient image-to-video transfer learning. Advances in Neural Information Processing Systems 35, 26462-26477 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 364, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 364, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 133, + 364, + 481, + 397 + ], + "type": "text", + "content": "41. Pedraza, L., Vargas, C., Narváez, F., Durán, O., Muñoz, E., Romero, E.: An open access thyroid ultrasound image database. In: 10th International symposium on medical information processing and analysis. vol. 9287, pp. 188-193. SPIE (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 398, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 398, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 133, + 398, + 481, + 430 + ], + "type": "text", + "content": "42. Pfeiffer, J., Rückle, A., Poth, C., Kamath, A., Vulic, I., Ruder, S., Cho, K., Gurevych, I.: Adapterhub: A framework for adapting transformers. arXiv preprint arXiv:2007.07779 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 431, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 431, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 133, + 431, + 481, + 453 + ], + "type": "text", + "content": "43. Pham, D.L., Xu, C., Prince, J.L.: Current methods in medical image segmentation. Annual review of biomedical engineering 2(1), 315-337 (2000)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 454, + 481, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 454, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 133, + 454, + 481, + 487 + ], + "type": "text", + "content": "44. Rasmy, L., Xiang, Y., Xie, Z., Tao, C., Zhi, D.: Med-bert: pretrained contextualized embeddings on large-scale structured electronic health records for disease prediction. NPJ digital medicine 4(1), 86 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 487, + 481, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 487, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 133, + 487, + 481, + 531 + ], + "type": "text", + "content": "45. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 532, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 532, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 133, + 532, + 481, + 564 + ], + "type": "text", + "content": "46. Roy, S., Wald, T., Koehler, G., Rokuss, M.R., Disch, N., Holzschuh, J., Zimmerer, D., Maier-Hein, K.H.: Sam. md: Zero-shot medical image segmentation capabilities of the segment anything model. arXiv preprint arXiv:2304.05396 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 565, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 565, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 133, + 565, + 481, + 597 + ], + "type": "text", + "content": "47. Sung, Y.L., Cho, J., Bansal, M.: Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5227-5237 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 599, + 481, + 632 + ], + "type": "text", + "content": "48. Valanarasu, J.M.J., Patel, V.M.: Unext: Mlp-based rapid medical image segmentation network. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 23-33. Springer (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "type": "text", + "content": "49. Wang, Y., Mukherjee, S., Liu, X., Gao, J., Awadallah, A.H., Gao, J.: Adamix: Mixture-of-adapter for parameter-efficient tuning of large language models. arXiv preprint arXiv:2205.12410 1(2), 4 (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Title Suppressed Due to Excessive Length" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 392 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "50. Wu, H., Chen, S., Chen, G., Wang, W., Lei, B., Wen, Z.: Fat-net: Feature adaptive transformers for automated skin lesion segmentation. Medical image analysis 76, 102327 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "text", + "content": "51. Wu, J., Fu, R., Fang, H., Liu, Y., Wang, Z., Xu, Y., Jin, Y., Arbel, T.: Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "text", + "content": "52. Wunderling, T., Golla, B., Poudel, P., Arens, C., Friebe, M., Hansen, C.: Comparison of thyroid segmentation techniques for 3d ultrasound. In: Medical Imaging 2017: Image Processing. vol. 10133, pp. 346-352. SPIE (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "53. Zhang, K., Liu, D.: Customized segment anything model for medical image segmentation. arXiv preprint arXiv:2304.13785 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 282 + ], + "type": "text", + "content": "54. Zhang, Y., Liu, H., Hu, Q.: Transfuse: Fusing transformers and cnns for medical image segmentation. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part I 24. pp. 14-24. Springer (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 326 + ], + "type": "text", + "content": "55. Zheng, S., Lu, J., Zhao, H., Zhu, X., Luo, Z., Wang, Y., Fu, Y., Feng, J., Xiang, T., Torr, P.H., et al.: Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6881-6890 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 392 + ], + "type": "text", + "content": "56. Zhou, Z., Rahman Siddiquee, M.M., Tajbakhsh, N., Liang, J.: Unet++: A nested u-net architecture for medical image segmentation. In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 20, 2018, Proceedings 4. pp. 3-11. Springer (2018)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 335, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 335, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 335, + 102 + ], + "type": "text", + "content": "Shreyank N Gowda and David A. Clifton" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CG-SLAM_ Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field/4bc60e08-fd5b-49a8-aee4-fd600735bd69_content_list.json b/2024/CG-SLAM_ Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field/4bc60e08-fd5b-49a8-aee4-fd600735bd69_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..710f21296e33bf8b143c1a95b9e75494dab8ef06 --- /dev/null +++ b/2024/CG-SLAM_ Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field/4bc60e08-fd5b-49a8-aee4-fd600735bd69_content_list.json @@ -0,0 +1,1916 @@ +[ + { + "type": "text", + "text": "CG-SLAM: Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field", + "text_level": 1, + "bbox": [ + 222, + 140, + 781, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiarui Hu $^{1*}$ , Xianhao Chen $^{2*}$ , Boyin Feng $^{1}$ , Guanglin Li $^{1}$ , Liangjing Yang $^{2}$ , Hujun Bao $^{1}$ , Guofeng Zhang $^{1}$ , and Zhaopeng Cui $^{1\\dagger}$", + "bbox": [ + 232, + 210, + 771, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State Key Lab of CAD&CG, Zhejiang University $^{2}$ ZJU-UIUC Institute, International Campus, Zhejiang University", + "bbox": [ + 277, + 253, + 723, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recently neural radiance fields (NeRF) have been widely exploited as 3D representations for dense simultaneous localization and mapping (SLAM). Despite their notable successes in surface modeling and novel view synthesis, existing NeRF-based methods are hindered by their computationally intensive and time-consuming volume rendering pipeline. This paper presents an efficient dense RGB-D SLAM system, i.e., CG-SLAM, based on a novel uncertainty-aware 3D Gaussian field with high consistency and geometric stability. Through an in-depth analysis of Gaussian Splatting, we propose several techniques to construct a consistent and stable 3D Gaussian field suitable for tracking and mapping. Additionally, a novel depth uncertainty model is proposed to ensure the selection of valuable Gaussian primitives during optimization, thereby improving tracking efficiency and accuracy. Experiments on various datasets demonstrate that CG-SLAM achieves superior tracking and mapping performance with a notable tracking speed of around $15\\mathrm{Hz}$ . We will make our source code publicly available. Project page: https://zju3dv.github.io/cg-slam.", + "bbox": [ + 261, + 316, + 743, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Dense Visual SLAM $\\cdot$ Neural Rendering $\\cdot$ 3D Gaussian field", + "bbox": [ + 261, + 565, + 740, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 604, + 375, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dense visual Localization and Mapping (Visual SLAM) is a long-standing problem in 3D computer vision over recent decades, which targets performing pose tracking and scene mapping simultaneously with a variety of downstream applications such as virtual/augmented reality (VR/AR), robot navigation, and autonomous driving. Traditional visual SLAM systems [24] have shown accurate tracking performance across various scenes, while the underlying 3D representations (e.g., point cloud, mesh, and surfel) demonstrate limitations in facilitating highly free scene exploration, such as photorealistic scene touring, fine-grained map updating, etc.", + "bbox": [ + 212, + 635, + 787, + 771 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inspired by the Neural Radiance Field (NeRF) [29] in surface reconstruction and view rendering, some novel NeRF-based SLAM methods [17,28,37,61] have", + "bbox": [ + 212, + 771, + 787, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Jiarui Hu and Xianhao Chen contributed equally to this work.", + "bbox": [ + 236, + 810, + 661, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding authors.", + "bbox": [ + 240, + 824, + 405, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c9250fa538b056f7b1c254309eae5937950543481d07a2031d8b3afd8a5b5e9d.jpg", + "image_caption": [ + "Fig. 1: CG-SLAM, which adopts a well-designed 3D Gaussian field, can simultaneously achieve state-of-the-art performance in localization, reconstruction and rendering. Benefiting from 3D Gaussian representation and a new GPU-accelerated framework that is developed from a thorough derivative analysis of camera pose in 3D Gaussian Splating [23], CG-SLAM can perform extremely fast rendering and solve the long-standing efficiency bottleneck suffered by previous rendering-based SLAM methods." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 305, + 234 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0a37d550eece540360ccfb3cf5fa13631ecb7fbadd4e083d0ff23e1d5f692175.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 145, + 513, + 234 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/93e5e1fd8a503283efa7c6ed73fbd240ba80cb21da472b44bb22ce0fef11e8fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 145, + 694, + 234 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0bcb09f1f3f85d8e34eb29dbb90fa7931babfa73ad1f9b7562c289e028c4d8ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 146, + 784, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "been proposed recently and demonstrated promising performance in tracking, surface modeling, and novel view synthesis. Nevertheless, existing NeRF-based methods follow the ray-tracing rendering pipeline that is computation-intensive and time-consuming. Therefore they can only perform tracking and mapping by sampling a limited number of camera rays, ignoring the natural structural information in images. To avoid local optima in tracking and artifacts, they normally require many optimization steps, which makes it struggle to bring the best of both worlds concerning accuracy and efficiency.", + "bbox": [ + 212, + 366, + 787, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Very recently the 3D Gaussian Splatting [23] method has been introduced for efficient novel view synthesis, and its rasterization-based rendering pipeline allows much faster image-level rendering, showing great potential in solving the inherent challenge of NeRF-based SLAM. However, it is nontrivial to reasonably incorporate the 3D Gaussian field in the SLAM setting. As a photorealistic view synthesis technique, the 3D Gaussian field is prone to overfitting the input images due to strong anisotropy and the lack of explicit multi-view constraints. As a result, on one hand, the 3D Gaussian splatting can not guarantee accurate modeling of 3D surfaces; on the other hand, since the Gaussians may not align with the environment's surfaces, this will lead to poor extrapolation capability and further degrade the camera tracking. Moreover, the increase of Gaussians in the mapping process will inevitably slow down the tracking efficiency.", + "bbox": [ + 212, + 489, + 787, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce a real-time Gaussian splatting SLAM system, i.e., CG-SLAM, based on a novel uncertainty-aware 3D Gaussian field with high consistency and geometric stability. To this end, we first conduct a comprehensive mathematical analysis regarding the derivatives of camera poses in the EWA (Elliptical Weighted Average) splatting process [63], and develop a CUDA framework tailored for the SLAM task that effectively decouples the tracking and mapping components. Second, in order to reduce the inherent overfitting problem, we utilize a scale regularization term that appropriately encourages the Gaussian ellipsoids to approximate Gaussian spheres, to reduce anisotropy and achieve a good trade-off between tracking accuracy and rendering realism. At the same time, we observed that solely employing alpha-blending depth cannot", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Jiarui Hu and Xianhao Chen et al.", + "bbox": [ + 271, + 114, + 504, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "impose effective constraints on the positions of Gaussian primitives. Thus, towards high-quality mapping, we further align median depth and alpha-blending depth to encourage Gaussian primitives to be well distributed over the scene surfaces, facilitating a consistent Gaussian field with more concentrated geometry density. Furthermore, in order to further improve the system's accuracy and efficiency, we design a novel depth uncertainty model to guide our Gaussian-based SLAM to focus on those stable and informative ones. We evaluate our system on a wide variety of RGB-D datasets, and the experimental results demonstrate that our CG-SLAM has superior performance in terms of tracking accuracy, reconstruction quality, and runtime efficiency.", + "bbox": [ + 212, + 146, + 782, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall, our contributions can be summarized as follows:", + "bbox": [ + 238, + 297, + 650, + 311 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a new GPU-accelerated framework for real-time dense RGB-D SLAM based on a thorough theoretical analysis of camera pose derivatives in 3D Gaussian Splatting.", + "- We design multiple loss terms to build up a consistent and stable 3D Gaussian field suitable for tracking and mapping.", + "- We propose a novel depth uncertainty model, which assists our system in selecting more valuable Gaussian primitives during optimization, thereby improving tracking efficiency and accuracy.", + "- Experiments on various datasets demonstrate that our method can achieve competitive or better tracking and mapping results compared to baselines." + ], + "bbox": [ + 228, + 323, + 782, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 496, + 387, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Dense Visual SLAM", + "text_level": 1, + "bbox": [ + 215, + 527, + 429, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following the seminal contributions of DTAM [32] and KinectFusion [31] to dense visual SLAM systems, there has been significant progress in developing efficient scene representation models. TSDF [31, 32], surfels [22, 39, 52], voxel hashing [6, 33, 34], and octrees [42, 49] have been introduced to tackle the challenges of scalability within the SLAM task. Some more advanced technologies, including bundle adjustment [39], loop closure, and learning-based algorithms [25, 26, 46, 47, 55], were subsequently integrated into the SLAM framework to further improve system performance. These enhancements have significantly brought better accuracy and robustness for localization and reconstruction capabilities, thus pushing the frontiers of what is achievable in dense SLAM systems. Compared with traditional methods, our CG-SLAM can reconstruct a fully dense 3D map for rich applications.", + "bbox": [ + 212, + 551, + 784, + 732 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Neural Implict Radiance Field based SLAM", + "text_level": 1, + "bbox": [ + 215, + 753, + 624, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural radiance fields [29] have shown promising potential in many 3D computer vision applications, such as novel view synthesis [3, 4, 14, 30], dynamic scene modeling [13, 15, 41, 48], and generalization [18, 40, 51, 53, 58]. Recent research works [27, 45] attempted to replace traditional maps, in tasks such as structure", + "bbox": [ + 212, + 779, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "CG-SLAM", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4f21ae97300a0cbe06ac5ba9adb44588eb90b1d2c5c04998bf84d17ba31c9562.jpg", + "image_caption": [ + "Fig. 2: System Overview. In a 3D Gaussian field constructed from an RGB-D sequence, we can render color, depth, opacity, and uncertainty maps through a GPU-accelerated rasterizer. Additionally, we attach a new uncertainty property to each Gaussian primitive to filter informative primitives. In the mapping process, we utilize multiple rendering results to design effective loss functions towards a consistent and stable Gaussian field. Subsequently, we employ appearance and geometry cues to perform accurate and efficient tracking." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from motion (SFM) and SLAM, with the neural implicit field to jointly optimize scene representation and camera poses. Different kinds of neural fields brought insights into NeRF-SLAM works. NICE-SLAM [61] chose a fully covered voxel grid to store neural features, while Vox-Fusion [56] further improved this grid to an adaptive size. Besides, Point-SLAM [37] attached feature embeddings to the point cloud on object surfaces. The neural point-based method is more flexible and can encode more concentrated volume density. Co-SLAM [50] adopted a hybrid representation that includes coordinate encoding and hash grids to achieve smoother reconstruction and faster convergence. In addition to the aforementioned works, some methods [8, 20, 36] only used the neural field as a map and still performed tracking based on a traditional feature point-based visual odometry. Our CG-SLAM system can achieve better and more efficient performance in tracking, mapping, and rendering than NeRF-based methods.", + "bbox": [ + 212, + 430, + 787, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 3D Gaussian Splitting Field", + "text_level": 1, + "bbox": [ + 215, + 648, + 496, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D Gaussian splatting is a revolutionary novel-view synthesis approach in 3D computer vision. This approach does not contain any neural network and allows photorealistic real-time(≥100 FPS) rendering at 1080p resolution. 3D Gaussian splatting has influenced a wide range of research topics, such as the 3D avatar [1, 19, 35, 59, 62], scene editing [7, 12, 57], image segment [60], and surface reconstruction [5, 11, 16]. These studies have demonstrated its generalization and effectiveness. Similar to the neural field, some concurrent manuscripts [21, 54] have been made to reversely infer camera poses within a built 3D Gaussian field. However, these works straightforwardly apply a raw 3D Gaussian field in the SLAM framework without specialized designs, such as anisotropy regularization and uninformative Gaussian primitive filtering. Additionally, they overlook", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Jiarui Hu and Xianhao Chen et al.", + "bbox": [ + 271, + 114, + 503, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the design on efficiency, which is the most important improvement that the Gaussian splatting technique should bring to a SLAM system. We believe that the advantages of the 3D Gaussian field in pose optimization have not been fully explored in these works, especially in terms of efficiency, and expect to further develop an advanced Gaussian-based visual SLAM system.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 244, + 330, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overview of our proposed rasterization-based Gaussian SLAM system is shown in Fig. 2. Given a set of RGB-D sequences, our system incrementally generates a stable, consistent, and uncertainty-aware Gaussian field, serving camera pose optimization and geometry reconstruction. In Sec. 3.1, we briefly introduce the 3D Gaussian splatting model and rasterization principles. We incorporated an uncertainty model that utilizes the geometry prior to attach the uncertainty property on rendered images and Gaussian primitives. This strategy helps remove outliers in mapping and makes full use of informative Gaussians in tracking (Sec. 3.2). Moreover, in Sec. 3.3, we detail the Gaussian primitive management strategy and some innovative loss terms that ensure geometry stability and accuracy. Finally, by minimizing the re-rendering loss from low-uncertainty primitives, we can build a real-time and accurate tracking module (Sec. 3.4).", + "bbox": [ + 212, + 277, + 787, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminary", + "text_level": 1, + "bbox": [ + 215, + 479, + 361, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scene Representation. 3D Gaussian Splatting defines a 3D scene as a set of anisotropic Gaussian distributions, which are associated with means $\\mathbf{X} \\in \\mathbb{R}^3$ and covariances $\\boldsymbol{\\Sigma} \\in \\mathbb{R}^{3 \\times 3}$ . To ensure that the covariance matrix remains positive semi-definite throughout the gradient descent, it is endowed with a more intuitive and comprehensible physical meaning, that is, the configuration of an ellipsoid. Specifically, $\\boldsymbol{\\Sigma}$ is simplified and decomposed into:", + "bbox": [ + 212, + 505, + 787, + 597 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {T} \\mathbf {R} ^ {T}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 444, + 608, + 784, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where scaling matrix $\\mathbf{S} = \\text{diag}([\\mathbf{s}])$ is derived from the scale factor $\\mathbf{s} \\in \\mathbb{R}^3$ and rotation matrix $\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}$ is derived from the quaternion $\\mathbf{q} \\in \\mathbb{R}^4$ . Each Gaussian ellipsoid is also assigned an opacity $\\sigma \\in \\mathbb{R}$ and spherical harmonics $(SH)$ coefficients, which respectively represent the volume density and view-dependent radiance within a nearby local region.", + "bbox": [ + 212, + 635, + 787, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following Zwicker et al. [63], given a world-to-camera rotation $\\mathbf{W}$ and the Jacobian $\\mathbf{J}$ of the affine approximation of the projective transformation, the EWA splatting algorithm illustrates how to approximately project a 3D Gaussian ellipsoid onto the image plane to determine its effective range and per-pixel opacity values on this image plane. We can obtain the corresponding 2D Gaussian distribution $\\mathcal{N}(\\hat{\\mu},\\hat{\\Sigma})$ as:", + "bbox": [ + 212, + 710, + 787, + 801 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {\\Sigma}} = \\mathbf {J} \\mathbf {W} \\boldsymbol {\\Sigma} \\mathbf {W} ^ {T} \\mathbf {J} ^ {T}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 800, + 784, + 818 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and $\\hat{\\mu}$ is the 2D pixel location of a 3D gaussian primitive center.", + "bbox": [ + 214, + 824, + 679, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "CG-SLAM", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fast Rasterization-based Rendering. Fast Gaussian splatting rasterizer enables efficient pixel-by-pixel parallel rendering, and is fully differentiable, which provides a useful GPU-accelerated framework. For an incoming frame, the rasterizer can pre-sort all visible Gaussian primitives in order of depth from near to far. In terms of color rendering, the Gaussian splatting rasterizer adopts an $\\alpha$ -blending solution, which accumulates radiance $c$ and opacity values $\\sigma$ on a given pixel by traversing the above depth queue as follows:", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} = \\sum_ {i = 1} ^ {N} \\alpha_ {i} T _ {i} c _ {i}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 446, + 262, + 785, + 303 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nT _ {i} = \\prod_ {k = 1} ^ {i - 1} \\left(1 - \\alpha_ {k}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 316, + 785, + 357 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} = \\mathcal {N} \\left(\\hat {\\mu} _ {\\mathbf {i}}, \\hat {\\Sigma} _ {\\mathbf {i}}\\right) \\sigma_ {i}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 366, + 785, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{I}$ is the rendered color, $T_{i}$ is the accumulated transmittance, $\\alpha_{i}$ is the opacity contributed to a pixel, and $c_{i}$ is the color of a Gaussian primitive computed from its SH coefficients. $N$ is the number of Gaussian primitives involved in the splatting process of a pixel. In terms of depth rendering, considering the loss term designed for geometry consistency, our rasterizer provides not only $\\alpha$ -blending depth $\\hat{D}_{\\text{alpha}}$ but also the median depth $\\hat{D}_{\\text{median}}$ :", + "bbox": [ + 212, + 388, + 787, + 481 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {D} _ {\\text {a l p h a}} = \\sum_ {i = 1} ^ {N} \\alpha_ {i} T _ {i} d _ {i}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 493, + 785, + 532 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {D} _ {\\text {m e d i a n}} = d _ {\\text {m e d i a n}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 429, + 547, + 784, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $d_{i}$ is the depth of a Gaussian primitive. For a pixel, in its splatting process, we regard a Gaussian at which the cumulative transmittance $T$ falls below 0.5 for the first time as the \"median Gaussian\". Its depth is recorded as $d_{\\text{median}}$ . $T_{\\text{median}}$ is the cumulative transmittance at this median Gaussian.", + "bbox": [ + 214, + 570, + 787, + 631 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left(T _ {\\text {m e d i a n}} \\geq 0. 5\\right) \\quad \\text {a n d} \\quad \\left(T _ {\\text {m e d i a n} + 1} < 0. 5\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 643, + 784, + 659 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Besides, the accumulated opacity value $\\hat{O}$ is similarly required to distinguish unobserved areas for spawning Gaussians:", + "bbox": [ + 214, + 667, + 787, + 700 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {O} = \\sum_ {i = 1} ^ {N} \\alpha_ {i} T _ {i}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 450, + 710, + 785, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Uncertainty Modeling", + "text_level": 1, + "bbox": [ + 214, + 770, + 446, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Uncertainty model remains a trending topic in multi-view 3D reconstruction in recent decades. Inspired by [38], we believe that explicitly modeling uncertainty in our 3D Gaussian field has a positive effect on increasing the ratio of informative", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Jiarui Hu and Xianhao Chen et al.", + "bbox": [ + 271, + 114, + 504, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/47de0d07bf1559dc85804f5ff7596f9311bca4715c2774f7aec5bc06aaf64477.jpg", + "image_caption": [ + "Fig. 3: Uncertainty of the Gaussian Primitives. Uncertainty of a Gaussian primitive is derived from its dominated pixels and corresponding depth biases, reflecting the geometric value and confidence of this primitive." + ], + "image_footnote": [], + "bbox": [ + 289, + 143, + 712, + 277 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Gaussians, which is crucial for the robustness and conciseness of a SLAM system. Hence, we propose an uncertainty model suitable for RGB-D observations from two perspectives: rendering images and Gaussian primitives.", + "bbox": [ + 214, + 362, + 782, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Uncertainty Map. $\\alpha$ -blending depth is essentially an expected value calculated by sampling Gaussian ellipsoids along a pixel ray. Under the reasonable assumption of a normal distribution, we think that the uncertainty map is highly related to its variance. We can render an uncertainty value from the 3D Gaussian field as in Eq. (10).", + "bbox": [ + 214, + 407, + 782, + 482 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nU = \\sum_ {i = 1} ^ {N} \\alpha_ {i} T _ {i} \\left(d _ {i} - D\\right) ^ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 483, + 784, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $D$ represents depth observations from the camera sensor. To mitigate drastic changes in positions of Gaussian primitives during optimization, we proposed a geometry variance loss term (Eq. (11)) based on the $H \\times W$ uncertainty map to force them to be near the observed depth.", + "bbox": [ + 214, + 530, + 782, + 590 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v a r} = \\frac {1}{H W} \\sum_ {n = 1} ^ {H W} | U _ {n} |. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 602, + 784, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Uncertainty of Gaussian primitives. From the perspective of geometric consistency, we design a loss term $\\mathcal{L}_{\\text{align}}$ as in Eq. (12) to align the $\\alpha$ -blending depth and median depth.", + "bbox": [ + 214, + 655, + 782, + 700 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l i g n}} = \\frac {1}{H W} \\sum_ {n = 1} ^ {H W} \\left| \\hat {D} _ {\\text {a l p h a}} ^ {n} - \\hat {D} _ {\\text {m e d i a n}} ^ {n} \\right|. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 712, + 784, + 752 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The $\\alpha$ -blending depth of a pixel is dominated by the Gaussian primitive with the largest weight. We call this pixel a \"dominated pixel\" of this maximum weight Gaussian primitive. Our alignment loss forces $\\hat{D}_{\\text{alpha}}$ and $\\hat{D}_{\\text{median}}$ to be similar, which makes this maximum weight Gaussian primitive always occur at $d_{\\text{median}}$ . Thus, as shown in Eq. (13), the uncertainty $v_i$ of the $i$ th Gaussian primitive", + "bbox": [ + 214, + 763, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "CG-SLAM", + "bbox": [ + 656, + 113, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "is determined by the difference between its depth and depth observations from all its dominated pixels within a keyframe window $F = \\{f_1, f_2, \\dots, f_k\\}$ . Fig. 3 further visualizes that a Gaussian primitive receives depth difference values from its dominated pixels in multiple viewpoints.", + "bbox": [ + 212, + 146, + 787, + 208 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\nu_ {i} = \\frac {1}{M _ {1} + \\dots + M _ {k}} \\sum_ {f _ {k} \\in F} \\sum_ {p = 1} ^ {M _ {k}} \\alpha_ {i} ^ {k, p} T _ {i} ^ {k, p} \\left(D _ {p} ^ {k} - d _ {i} ^ {k}\\right) ^ {2}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 215, + 785, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In a keyframe $f_{k}$ , $\\alpha_{i}^{k,p}$ and $T_{i}^{k,p}$ represent the opacity and transmittance of the $i$ th Gaussian primitive on a pixel $p$ . $D_{p}^{k}$ represents the depth observation on a pixel $p$ in $f_{k}$ . $d_{i}^{k}$ is the depth value of the $i$ -th Gaussian primitive at $f_{k}$ . $\\{M_{1},\\dots,M_{k}\\}$ are the number of dominated pixels of the $i$ th Gaussian primitive in different keyframes. Benefiting from uncertainty modeling, we can regularly detect and remove unreliable Gaussian primitives with high uncertainty exceeding a threshold $\\tau = 0.025$ . Specifically, during the mapping optimization, primitives with $v_{i} > \\tau$ will be manually reduced to a low-opacity level. These low-opacity Gaussian primitives can be optimized again to remove truly irreversible ones, which is a more adaptive and reasonable strategy.", + "bbox": [ + 214, + 268, + 789, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Mapping", + "text_level": 1, + "bbox": [ + 215, + 443, + 338, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We employ various loss functions to update Gaussian properties, aiming for a consistent and stable Gaussian field. In addition to color and $SSIM$ loss in the original 3D Gaussian splatting, previous experience from NeRF-SLAM works suggests that geometry loss is a necessary part. To overcome anisotropic interference (Arrow-shaped Gaussian primitives), we add a soft scale regularization loss in the mapping process. Note that our system performs the initialization at a slightly higher cost, i.e., more optimization iters.", + "bbox": [ + 212, + 465, + 787, + 571 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o l o r}} = \\frac {1}{H W} \\sum_ {n = 1} ^ {H W} \\left| \\hat {I} _ {n} - I _ {n} \\right|, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 400, + 580, + 785, + 621 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s s i m} = S S I M (\\hat {I}, I), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 423, + 633, + 784, + 651 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {g e o}} = \\frac {1}{H W} \\sum_ {n = 1} ^ {H W} \\left| \\hat {D} _ {\\text {a l p h a}} ^ {n} - D _ {n} \\right|, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 667, + 784, + 708 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i s o} = \\frac {1}{G} \\sum_ {i \\in G} \\max \\left(\\frac {\\operatorname* {m a x} \\left(\\left\\{s _ {i} ^ {x} , s _ {i} ^ {y} , s _ {i} ^ {z} \\right\\}\\right)}{\\operatorname* {m i n} \\left(\\left\\{s _ {i} ^ {x} , s _ {i} ^ {y} , s _ {i} ^ {z} \\right\\}\\right)}, \\epsilon\\right) - \\epsilon , \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 715, + 784, + 751 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $I_{n}$ and $D_{n}$ represent ground-truth color and depth, $\\epsilon = 1.0$ is a hyperparameter that controls the level of anisotropy, and $G$ represents all visible Gaussians under the current view.", + "bbox": [ + 212, + 755, + 787, + 800 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {m a p p i n g}} = \\omega_ {1} \\mathcal {L} _ {\\text {c o l o r}} + \\omega_ {2} \\mathcal {L} _ {\\text {s s i m}} + \\omega_ {3} \\mathcal {L} _ {\\text {g e o}} \\tag {18} \\\\ + \\omega_ {4} \\mathcal {L} _ {\\text {a l i g n}} + \\omega_ {5} \\mathcal {L} _ {\\text {i s o}} + \\omega_ {6} \\mathcal {L} _ {\\text {v a r}}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 809, + 785, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Jiarui Hu and Xianhao Chen et al.", + "bbox": [ + 271, + 114, + 504, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "These loss functions customized for the SLAM task facilitate faster convergence in mapping and lay a solid foundation for subsequent tracking.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Gaussian Management. In initialization, we densely project Gaussian primitives into 3D space based on depth observations of the first frame. In subsequent mapping, we set an empirical threshold $\\psi = 0.5$ to extract unobserved or under-constructed pixels where $\\hat{O} < \\psi$ . Then, we utilize color and depth information on these pixels to spawn fresh Gaussian primitives. In addition, we inherited the original splitting, cloning, and removing strategy for Gaussian densification.", + "bbox": [ + 212, + 176, + 784, + 267 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Tracking", + "text_level": 1, + "bbox": [ + 214, + 289, + 334, + 304 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In our system, we have proposed the first comprehensive mathematical theory on derivatives w.r.t. pose in 3D Gaussian splatting framework (refer to supplementary), and empirically discovered that the Lie algebraic representation is more advantageous for camera tracking, especially for the rotation, in a Gaussian field. The camera pose optimization, i.e., rotation and translation $\\{\\mathfrak{so}(3)|T\\}$ , mainly includes two parts: sequential tracking and sliding bundle adjustment.", + "bbox": [ + 212, + 313, + 782, + 404 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Sequential Tracking. Given the fixed scene representation, the camera pose is initially guessed via the constant speed assumption where the last pose is transformed by the last relative transformation, and then we refine this rough pose using similar photometric and geometric losses weighted by $\\lambda = \\{\\lambda_1,\\lambda_2\\}$ .", + "bbox": [ + 212, + 405, + 784, + 465 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t r a c k i n g}} = \\lambda_ {1} \\mathcal {L} _ {\\text {c o l o r}} + \\lambda_ {2} \\mathcal {L} _ {\\text {g e o}}, \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 477, + 784, + 492 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\{\\mathfrak {s o} (3) | T \\} = \\underset {\\{\\mathfrak {s o} (3) | T \\}} {\\operatorname {a r g m i n}} \\left(\\mathcal {L} _ {\\text {t r a c k i n g}}\\right). \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 518, + 784, + 545 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Sliding Bundle Adjustment. Cumulative error is a typical problem in SLAM, also in Gaussian-based SLAM systems. To ease it, we set up a sliding window $F$ containing $k$ co-visible keyframes and jointly optimize camera extrinsic and scene representation in this window. Due to the efficiency, we encode keyframes into a descriptor pool with a pre-trained NetVLAD [2] model to determine co-visibility through the cosine similarity scores, instead of the view frustum overlap method in previous works. In addition to keyframes from NetVLAD [2], we also added the current frame and the most recent keyframes in the sliding window considering temporal associations. We still employ $\\mathcal{L}_{\\text{mapping}}$ in sliding bundle adjustment, where $\\varPsi$ is a set of all estimizable Gaussian properties.", + "bbox": [ + 212, + 550, + 785, + 702 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\varPsi , \\{\\mathfrak {s o} (3) | T \\} = \\underset {\\varPsi , \\{\\mathfrak {s o} (3) | T \\}} {\\operatorname {a r g m i n}} \\left(\\mathcal {L} _ {\\text {m a p p i n g}}\\right). \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 369, + 717, + 784, + 742 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 763, + 375, + 780 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we describe our experimental setup and validate that the proposed system can achieve improvement in both accuracy (Sec. 4.1 and Sec. 4.2) and efficiency (Sec. 4.3). We also confirmed the effectiveness of our design choices", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "CG-SLAM", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "(Sec. 4.4). Additionally, we demonstrate our advantages in image rendering and capability for online third-person view rendering in supplementary. We color each cell as best, second best, and third best.", + "bbox": [ + 212, + 146, + 782, + 190 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. To evaluate our system in various scenarios, we use three standard benchmarks: Replica [43], TUM [44], and ScanNet [10]. The Replica dataset contains 8 available synthetic RGB-D sequences generated by Sucar et al. [45]. We examined the generalization of our method on real-world TUM [44] and ScanNet [10] datasets, which contain 5 and 6 challenging scenes respectively.", + "bbox": [ + 212, + 191, + 784, + 268 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details. We run our system on a desktop equipped with an Intel i9-14900K and an NVIDIA RTX 4090 GPU. We set the learning rate of $\\{\\mathfrak{so}(3)|T\\}$ to $\\{0.0015, 0.00215\\}$ in sequential tracking in all experiments. For the Replica [43] dataset, we perform 60-iteration mapping with weights $\\omega = \\{0.7, 0.1, 0.25, 0.25, 0.1, 0.15\\}$ in a sliding window with $k = 4$ keyframes and 15-iteration sequential pose optimization weighted by $\\lambda = \\{0.2, 1.0\\}$ . We select keyframes at an interval of 30. For TUM [44] and ScanNet [10] datasets, we use $\\omega = \\{1.0, 0.1, 0.8, 0.5, 0.1, 0.5\\}$ , $k = 4$ , $\\lambda = \\{1.0, 0.6\\}$ . Faced with challenging real-world scenes, we need to extract more keyframes at an interval of 15, perform mapping at $40 \\sim 50$ iterations, and increase tracking iterations to 25. For further implementation details, please refer to our supplementary.", + "bbox": [ + 212, + 268, + 785, + 434 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Metrics. We quantitatively evaluate reconstruction quality using different 3D metrics. Given 3D triangle meshes, we compute mapping Accuracy [cm], Completion [cm], and Completion Ratio $[<5\\mathrm{cm}\\%]$ . Following NICE-SLAM [61], we discard unobserved regions that are not in any viewpoints. As for tracking performance, we measure ATE RMSE [44] for estimated trajectories.", + "bbox": [ + 212, + 436, + 784, + 511 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines. We primarily consider state-of-the-art NeRF-SLAM works, including NICE-SLAM [61], Co-SLAM [50], Point-SLAM [37], and Vox-Fusion [56], as baselines. For a fair comparison, we reproduced all results from these baselines and reported their reconstruction performance with the same evaluation mechanism. We also add some concurrent manuscripts such as GS-SLAM [54] and SplaTAM [21] for reference, and we directly report the results in their papers.", + "bbox": [ + 212, + 512, + 784, + 603 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Localization Evaluation", + "text_level": 1, + "bbox": [ + 214, + 628, + 455, + 643 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We report the localization accuracy of our system in 8 Replica [43] scenes in Tab. 1, where we surpass all other methods by a notable margin of around $25\\% \\sim 75\\%$ . In our Gaussian-based system, image-level pose optimization and the well-designed Gaussian field promote fast and stable convergence to an optimal solution. This is the reason why we have lower variances and higher accuracy. As shown in Tab. 2, despite noisy and sparse depth information in the real-world TUM-RGBD dataset [44], our method still achieves better or competitive performance in 5 selected scenarios. We also benchmark our method and baselines on the similarly challenging ScanNet [10] to compare their tracking performance in Tab. 3. Sensor data from multiple large-scale ScanNet scenes suffers from severe motion blur and specular reflections. Our method further demonstrates its effectiveness and superiority on this dataset, excelling or maintaining com", + "bbox": [ + 212, + 657, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Jiarui Hu and Xianhao Chen et al.", + "bbox": [ + 271, + 114, + 501, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/70ce5871dbf69a5a31486015285ee17b31f175c19c0f5d815eed8daf65657d58.jpg", + "table_caption": [ + "Table 1: Tracking Results on the Replica Dataset [43] (ATE RMSE [cm] $\\downarrow$ ). Our system consistently achieved the best performance in this dataset, both for 8 individual scenes and for the average. It is worth noting that GS-SLAM [54] and SplaTAM [21] are concurrent with ours." + ], + "table_footnote": [], + "table_body": "
Methodrm-0rm-1rm-2off-0off-1off-2off-3off-4Avg.
NICE-SLAM0.971.311.070.881.001.061.101.131.06
Co-SLAM0.771.041.090.580.532.051.490.840.99
Point-SLAM0.560.470.300.350.620.550.720.730.54
Vox-Fusion0.400.540.540.500.460.750.500.600.54
GS-SLAM0.480.530.330.520.410.590.460.700.50
SplaTAM0.310.400.290.470.270.290.320.550.36
Ours0.290.270.250.330.140.280.310.290.27
Ours-light0.440.360.330.290.270.430.520.580.40
", + "bbox": [ + 308, + 210, + 692, + 323 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/99afe811309c84c92dd987bc93fb170820f4e72d726942e1753f67639bb8cd2d.jpg", + "table_caption": [ + "Table 2: Tracking Results on the TUM-RGBD Dataset [44] (ATE RMSE [cm] ↓). Our system achieves better tracking accuracy and lower variance in different scenarios. \"-\" indicates unavailable results because the related work is not open source." + ], + "table_footnote": [], + "table_body": "
Methodfr1/deskfr1/desk2fr1/roomfr2.xyzfr3/officeAvg.
NICE-SLAM4.264.9934.4931.733.8715.87
Co-SLAM2.74.5730.161.92.68.38
Point-SLAM4.344.5430.921.313.488.92
Vox-Fusion3.526.0019.531.4926.0111.31
GS-SLAM3.3--1.36.6-
SplaTAM3.356.5411.131.245.165.48
Ours2.434.549.391.202.454.0
Ours-light3.144.7310.671.282.604.48
", + "bbox": [ + 308, + 388, + 692, + 497 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/f3be7be962146be29eca923b53cc5bb81e455ce5a2765d550df5b9f4996ccf8c.jpg", + "table_caption": [ + "Table 3: Tracking Results on the ScanNet Dataset [10] (ATE RMSE [cm] ↓). Our method achieves state-of-the-art tracking results in 6 scenes and exceeds other methods on average. \"-\" indicates failure results in Vox-Fusion [56]." + ], + "table_footnote": [], + "table_body": "
MethodSc.0000Sc.0059Sc.0106Sc.0169Sc.0181Sc.0207Avg.
NICE-SLAM12.0014.007.9010.9013.406.2010.70
Co-SLAM7.1812.2910.96.6213.437.139.37
Point-SLAM10.248.2911.8622.1614.779.5412.19
Vox-Fusion8.398.95-9.5012.206.43-
SplaTAM12.8310.1017.7212.0811.107.4611.88
Ours7.097.468.888.1611.605.348.08
Ours-light5.628.739.787.9312.025.458.17
", + "bbox": [ + 308, + 561, + 692, + 660 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "petitiveness in various scenes. Extensive experiments showcase the remarkable ability of our proposed system to track and handle complex situations.", + "bbox": [ + 212, + 686, + 785, + 717 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Reconstruction Evaluation", + "text_level": 1, + "bbox": [ + 214, + 724, + 482, + 739 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In Tab. 4, we quantitatively measure the mapping performance of our proposed system, in comparison to NICE-SLAM [61], Co-SLAM [50], Point-SLAM [37], and Vox-Fusion [56]. It can be observed that our method outperforms all baselines on mapping accuracy. We use TSDF-Fusion [9] to construct our mesh map. We achieve state-of-the-art reconstruction in observed areas. It is worth noting that the Gaussian-based method neither has a global MLP nor a fully covered", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "CG-SLAM", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/f996adce620ebd39aee9f00a97d2935c567cb541b0c3cfacb36b57a4a8d9cc93.jpg", + "table_caption": [ + "Table 4: Reconstruction Results on the Replica [43] Dataset. In terms of mapping accuracy, our method can outperform all existing methods. Due to the inherent limitation of 3D Gaussian representation, our method is slightly worse in completion." + ], + "table_footnote": [], + "table_body": "
MethodMetricrm-0rm-1rm-2off-0off-1off-2off-3off-4Avg.
NICE-SLAMAcc.[cm]↓3.533.603.035.563.354.713.843.353.87
Comp.[cm]↓3.403.623.274.554.033.943.994.153.87
Comp.Ratio<5cm%|↑86.0580.7587.2379.3482.1380.3580.5582.8882.41
Co-SLAMAcc.[cm]↓2.111.681.991.571.312.843.062.232.10
Comp.[cm]↓2.021.811.961.561.592.432.722.522.08
Comp.Ratio<5cm%|↑95.2695.1993.5896.0994.6591.6390.7290.4493.44
Point-SLAMAcc.[cm]↓1.451.141.191.050.861.311.571.511.26
Comp.[cm]↓3.463.022.651.652.213.623.473.903.00
Comp.Ratio<5cm%|↑88.4889.4490.1393.3990.5186.1786.0085.7488.73
Vox-FusionAcc.[cm]↓1.771.512.231.631.442.092.332.021.88
Comp.[cm]↓2.692.312.581.871.663.032.813.512.56
Comp.Ratio<5cm%|↑92.0392.4790.1393.8694.4088.9489.1086.5390.93
OursAcc.[cm]↓1.100.970.960.850.671.101.281.161.01
Comp.[cm]↓3.262.772.791.492.153.343.233.662.84
Comp.Ratio<5cm%|↑88.2689.4889.1093.6090.1486.0485.7885.6688.51
Ours-lightAcc.[cm]↓1.170.980.990.870.711.21.361.261.06
Comp.[cm]↓3.322.652.811.512.203.443.273.782.87
Comp.Ratio<5cm%|↑88.2089.3389.1093.4190.1485.6485.4785.3088.43
", + "bbox": [ + 245, + 196, + 756, + 420 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "feature grid, as in Co-SLAM [50]. Consequently, such a system exhibits a slightly weaker hole-filling ability compared to the NeRF-based method, which explains why our system is slightly worse in the Completion metric. As shown in Fig. 4, we qualitatively present the ground truth mesh and mesh reconstructions from ours and other baselines. Evidently, our system achieves more detailed geometry and less noisy outliers.", + "bbox": [ + 212, + 446, + 787, + 539 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Runtime and Memory Analysis", + "text_level": 1, + "bbox": [ + 215, + 559, + 519, + 574 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We evaluate the runtime and memory footprint of our system compared to other works in Tab. 5. We reported the tracking and mapping efficiency in terms of per-iteration time consumption and the total number of optimization iterations. With the support of the GPU-accelerated rasterizer, our system can operate at around $8.5\\mathrm{Hz}$ . Meanwhile, our carefully designed pipeline allows this system to expand to a lightweight version and to work with half-resolution images. For tracking, this lightweight version can work twice as fast as the original one, at the cost of a slight decrease in accuracy, as shown in Tabs. 1 to 3. For mapping, our lightweight version demonstrates similar superiority in Tab. 4. It can be clearly seen that both our versions achieve better performance than Co-SLAM [50]. Further efficiency analysis on TUM-RGBD [44] and ScanNet [10] is shown in supplementary. Also, our customized Gaussian field allows us to outperform the concurrent Gaussian-based works with less computational burden.", + "bbox": [ + 212, + 583, + 785, + 777 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "However, as a non-MLP scene representation, the 3D Gaussian field inevitably requires much memory consumption to store different properties. This reason results in a considerable memory footprint in the Gaussian-based SLAM system, which is a common limitation in other Gaussian-based research topics.", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Jiarui Hu and Xianhao Chen et al.", + "bbox": [ + 271, + 114, + 501, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/207df06401ea72020225d7e679de9303ab1b0908010c35b9b70d443fe8999c4b.jpg", + "image_caption": [ + "Fig. 4: Reconstruction Performance on Replica [43] Dataset. We qualitatively compared the mesh reconstruction results from CG-SLAM and other baselines, where CG-SLAM can produce more detailed geometry at a lower computation cost." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 787, + 310 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/eac535597150c1a9d33d940981e9854522881cbca37257271332a88b9a337e02.jpg", + "table_caption": [ + "Table 5: Runtime and Memory Usage. We comprehensively compared the runtime and memory usage on Replica [43] Office 0. Our proposed CG-SLAM can perform more efficient tracking and mapping than existing works, actually reaching a real-time level. \"-\" indicates unavailable results in related works." + ], + "table_footnote": [], + "table_body": "
MethodTracking [ms× it]↓Mapping [ms× it]↓Mapping IntervalSystem FPS↑Decoder Param↓Scene Embedding↓
Vox-Fusion23.61 × 3086.55 × 10101.250.98 MB0.162 MB
NICE-SLAM6.19 × 1091.59 × 6050.860.43 MB89.56 MB
Co-SLAM4.45 × 1011.9 × 10514.646.37 MB-
Point-SLAM6.14 × 4022.25 × 30050.630.54 MB28.11 MB
GS-SLAM11.9 × 1012.8 × 100-8.34--
SplaTAM41.7 × 4050.1 × 6010.21--
Ours6.02 × 1028.9 × 60308.47-231.66 MB
Ours-light3.80 × 1010.70 × 603016.83-56.50 MB
", + "bbox": [ + 274, + 446, + 730, + 579 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 609, + 388, + 625 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To verify the rationality of our designs, we investigate the effectiveness of the anisotropy regularization, alignment and variance losses, and uncertainty model.", + "bbox": [ + 212, + 641, + 785, + 671 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of Isotropy Loss. Ablation results in Tab. 6 reveal how the anisotropy regularization term affects tracking metrics. We can notice that anisotropy $(\\mathrm{w} / \\mathrm{o}\\mathcal{L}_{iso})$ will reduce the tracking accuracy by disrupting the view extrapolation of a 3D Gaussian field. Oppositely, excessive regularization (Isotropic Gaussians) will degrade rendering and subsequently affect the photometric loss, leading to poor tracking performance. Overall, a soft controllable constraint like $\\mathcal{L}_{iso}$ is a reasonable choice. To more intuitively illustrate this phenomenon, we display opacity maps with and without anisotropy regularization in Fig. 5 (b), and apparent arrow-shaped artifacts occur in the latter.", + "bbox": [ + 212, + 672, + 795, + 808 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of Alignment and Variance Losses. Alignment and variance losses push primitives closer to object surfaces, facilitating novel view syn-", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "CG-SLAM", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ab1376c0784003867633c6e892e7229ea1d8aa03f6fa3cd7bffa2929676bf490.jpg", + "image_caption": [ + "Fig. 5: Uncertainty Model Ablation and Anisotropy Interference. (a) Uncertainty Model Ablation. This plot illustrates that the uncertainty model helps improve tracking accuracy while avoiding some extreme errors. (b) Anisotropy Interference. It can be clearly seen that in the case of w/o $\\mathcal{L}_{iso}$ , serious arrow-shaped artifacts occur on the edges of the image." + ], + "image_footnote": [], + "bbox": [ + 259, + 143, + 743, + 244 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/7f4940df4dc69c59a720e483b835e5fc333fb13f61bdb627dbfff43594734eb2.jpg", + "table_caption": [ + "Table 6: Isotropy Loss Ablation Results(ATE RMSE [cm] ↓). The experimental results demonstrate the effectiveness of our anisotropy regularization term. \"--\" indicates a failure situation." + ], + "table_footnote": [], + "table_body": "
Settingrm-0rm-1rm-2off-0off-1off-2off-3off-4Avg.
w/ Liso0.290.270.250.330.140.280.310.290.27
w/o Liso0.320.310.540.360.240.310.72--
Isotropic Gaussians0.300.321.040.297.620.381.541.361.61
", + "bbox": [ + 303, + 392, + 702, + 445 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "thesis and SLAM. The effectiveness of these two losses is demonstrated in the supplementary.", + "bbox": [ + 212, + 473, + 784, + 503 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effectiveness of Uncertainty Model. As shown in Fig. 5 (a), tracking error curves show that our uncertainty model can improve tracking accuracy while avoiding some extreme errors. For further quantitative ablation results, refer to the supplementary material.", + "bbox": [ + 212, + 505, + 785, + 565 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 592, + 359, + 607 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have proposed CG-SLAM, a dense RGB-D SLAM based on a consistent and uncertainty-aware 3D Gaussian field. Our targeted loss functions strengthen the 3D Gaussian field in terms of consistency and stability. The uncertainty model further distills highly informative primitives in this field to reduce interference from outliers. Besides, a customized GPU-accelerated rasterization pipeline enables our system to achieve state-of-the-art accuracy and efficiency in various scenes. Through extensive experiments, it can be concluded that our method outperforms previous works regarding tracking, mapping, and efficiency.", + "bbox": [ + 212, + 627, + 787, + 750 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. Considerable memory usage is one limitation of the Gaussian-based system. We expect that a more compact Gaussian field can be adopted in the SLAM task. In addition, the Gaussian-based method has a weak prediction ability for unobserved areas. Moreover, our system is not capable of handling dynamic objects in the environment. We believe it is a very interesting direction for future work.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Jiarui Hu and Xianhao Chen et al.", + "bbox": [ + 271, + 114, + 504, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was partially supported by the NSFC (No. 62102356).", + "bbox": [ + 215, + 178, + 679, + 194 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 219, + 321, + 234 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abdal, R., Yifan, W., Shi, Z., Xu, Y., Po, R., Kuang, Z., Chen, Q., Yeung, D.Y., Wetzstein, G.: Gaussian shell maps for efficient 3d human generation (2023)", + "2. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn architecture for weakly supervised place recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5297-5307 (2016)", + "3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields (2021)", + "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR (2022)", + "5. Chen, H., Li, C., Lee, G.H.: Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance (2023)", + "6. Chen, J., Bautembach, D., Izadi, S.: Scalable real-time volumetric surface reconstruction. ACM Trans. Graph. 32(4), 113:1-113:16 (2013). https://doi.org/10.1145/2461912.2461940, https://doi.org/10.1145/2461912.2461940", + "7. Chen, Y., Chen, Z., Zhang, C., Wang, F., Yang, X., Wang, Y., Cai, Z., Yang, L., Liu, H., Lin, G.: Gaussianeditor: Swift and controllable 3d editing with gaussian splatting (2023)", + "8. Chung, C., Tseng, Y., Hsu, Y., Shi, X.Q., Hua, Y., Yeh, J., Chen, W., Chen, Y., Hsu, W.H.: Orbeez-slam: A real-time monocular visual SLAM with ORB features and nerf-realized mapping. In: IEEE International Conference on Robotics and Automation, ICRA 2023, London, UK, May 29 - June 2, 2023. pp. 9400-9406. IEEE (2023). https://doi.org/10.1109/ICRA48891.2023.10160950, https://doi.org/10.1109/ICRA48891.2023.10160950", + "9. Curless, B., Levoy, M.: A volumetric method for building complex models from range images. In: Proceedings of the 23rd annual conference on Computer graphics and interactive techniques. pp. 303-312 (1996)", + "0. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017)", + "1. Duisterhof, B.P., Mandi, Z., Yao, Y., Liu, J.W., Shou, M.Z., Song, S., Ichnowski, J.: Md-splatting: Learning metric deformation from 4d gaussians in highly deformable scenes (2023)", + "2. Fang, J., Wang, J., Zhang, X., Xie, L., Tian, Q.: Gaussianeditor: Editing 3d gaussians delicately with text instructions. arXiv preprint arXiv:2311.16037 (2023)", + "3. Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Niešner, M., Tian, Q.: Fast dynamic radiance fields with time-aware neural voxels. In: SIGGRAPH Asia 2022 Conference Papers (2022)", + "4. Fridovich-Keil and Yu, Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: CVPR (2022)", + "5. Geng, C., Peng, S., Xu, Z., Bao, H., Zhou, X.: Learning neural volumetric representations of dynamic humans in minutes. In: CVPR (2023)" + ], + "bbox": [ + 225, + 252, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "CG-SLAM", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Guédon, A., Lepetit, V.: Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. arXiv preprint arXiv:2311.12775 (2023)", + "17. Hu, J., Mao, M., Bao, H., Zhang, G., Cui, Z.: Cp-slam: Collaborative neural point-based slam system. Advances in Neural Information Processing Systems 36 (2024)", + "18. Irshad, M.Z., Zakharov, S., Liu, K., Guizilini, V., Kollar, T., Gaidon, A., Kira, Z., Ambrus, R.: Neo 360: Neural fields for sparse view synthesis of outdoor scenes (2023), https://arxiv.org/abs/2308.12967", + "19. Jena, R., Iyer, G.S., Choudhary, S., Smith, B., Chaudhari, P., Gee, J.: Splatarmor: Articulated gaussian splatting for animatable humans from monocular rgb videos. arXiv preprint arXiv:2311.10812 (2023)", + "20. Katragadda, S., Lee, W., Peng, Y., Geneva, P., Chen, C., Guo, C., Li, M., Huang, G.: Nerf-vins: A real-time neural radiance field map-based visual-inertial navigation system. CoRR abs/2309.09295 (2023). https://doi.org/10.48550/ARXIV.2309.09295, https://doi.org/10.48550/arXiv.2309.09295", + "21. Keetha, N., Karhade, J., Jatavallabhula, K.M., Yang, G., Scherer, S., Ramanan, D., Luiten, J.: Splatam: Splat, track & map 3d gaussians for dense rgb-d slam. arXiv preprint arXiv:2312.02126 (2023)", + "22. Keller, M., Lefloch, D., Lambers, M., Izadi, S., Weyrich, T., Kolb, A.: Real-time 3d reconstruction in dynamic scenes using point-based fusion. In: 2013 International Conference on 3D Vision, 3DV 2013, Seattle, Washington, USA, June 29 - July 1, 2013. pp. 1-8. IEEE Computer Society (2013). https://doi.org/10.1109/3DV.2013.9, https://doi.org/10.1109/3DV.2013.9", + "23. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)", + "24. Kerl, C., Sturm, J., Cremers, D.: Dense visual slam for rgb-d cameras. In: 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems. pp. 2100-2106. IEEE (2013)", + "25. Koestler, L., Yang, N., Zeller, N., Cremers, D.: Tandem: Tracking and dense mapping in real-time using deep multi-view stereo. In: Conference on Robot Learning (CoRL) (2021)", + "26. Li, Y., Liu, X., Dong, W., Zhou, H., Bao, H., Zhang, G., Zhang, Y., Cui, Z.: Deltar: Depth estimation from a light-weight tof sensor and rgb image. In: European conference on computer vision. pp. 619-636. Springer (2022)", + "27. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: Barf: Bundle-adjusting neural radiance fields. In: IEEE International Conference on Computer Vision (ICCV) (2021)", + "28. Liu, X., Li, Y., Teng, Y., Bao, H., Zhang, G., Zhang, Y., Cui, Z.: Multi-modal neural radiance field for monocular dense slam with a light-weight tof sensor. In: Proceedings of the iee/cvf international conference on computer vision. pp. 1-11 (2023)", + "29. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: The European Conference on Computer Vision (ECCV) (2020)", + "30. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022). https://doi.org/10.1145/3528223.3530127, https://doi.org/10.1145/3528223.3530127", + "31. Newcombe, R.A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohli, P., Shotton, J., Hodges, S., Fitzgibbon, A.W.: Kinectfusion: Real-time dense surface mapping and tracking. In: 10th IEEE International Symposium on Mixed" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Jiarui Hu and Xianhao Chen et al.", + "bbox": [ + 271, + 114, + 501, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Augmented Reality, ISMAR 2011, Basel, Switzerland, October 26-29, 2011. pp. 127-136. IEEE Computer Society (2011). https://doi.org/10.1109/ISMAR.2011.6092378, https://doi.org/10.1109/ISMAR.2011.6092378", + "32. Newcombe, R.A., Lovegrove, S., Davison, A.J.: DTAM: dense tracking and mapping in real-time. In: Metaxas, D.N., Quan, L., Sanfeliu, A., Gool, L.V. (eds.) IEEE International Conference on Computer Vision, ICCV 2011, Barcelona, Spain, November 6-13, 2011. pp. 2320-2327. IEEE Computer Society (2011). https://doi.org/10.1109/ICCV.2011.6126513, https://doi.org/10.1109/ICCV.2011.6126513", + "33. Nießner, M., Zollhöfer, M., Izadi, S., Stamminger, M.: Real-time 3d reconstruction at scale using voxel hashing. ACM Trans. Graph. 32(6), 169:1-169:11 (2013). https://doi.org/10.1145/2508363.2508374, https://doi.org/10.1145/2508363.2508374", + "34. Oleynikova, H., Taylor, Z., Fehr, M., Siegwart, R., Nieto, J.I.: Voxblox: Incremental 3d euclidean signed distance fields for on-board MAV planning. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2017, Vancouver, BC, Canada, September 24-28, 2017. pp. 1366-1373. IEEE (2017). https://doi.org/10.1109/IROS.2017.8202315, https://doi.org/10.1109/IROS.2017.8202315", + "35. Pang, H., Zhu, H., Kortylewski, A., Theobalt, C., Habermann, M.: Ash: Animatable gaussian splats for efficient and photoreal human rendering (2023)", + "36. Rosinol, A., Leonard, J.J., Carlone, L.: Nerf-slam: Real-time dense monocular SLAM with neural radiance fields. In: IROS. pp. 3437-3444 (2023). https://doi.org/10.1109/IROS55552.2023.10341922, https://doi.org/10.1109/IROS55552.2023.10341922", + "37. Sandström, E., Li, Y., Van Gool, L., Oswald, M.R.: Point-slam: Dense neural point cloud-based slam. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18433-18444 (2023)", + "38. Sandström, E., Ta, K., Van Gool, L., Oswald, M.R.: Uncle-slam: Uncertainty learning for dense neural slam. arXiv preprint arXiv:2306.11048 (2023)", + "39. Schops, T., Sattler, T., Pollefeys, M.: Bad slam: Bundle adjusted direct rgb-d slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)", + "40. Schwarz, K., Liao, Y., Niemeyer, M., Geiger, A.: Graf: Generative radiance fields for 3d-aware image synthesis. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)", + "41. Shao, R., Zheng, Z., Tu, H., Liu, B., Zhang, H., Liu, Y.: Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2023)", + "42. Steinbrücker, F., Kerl, C., Cremers, D.: Large-scale multi-resolution surface reconstruction from RGB-D sequences. In: IEEE International Conference on Computer Vision, ICCV 2013, Sydney, Australia, December 1-8, 2013. pp. 3264-3271. IEEE Computer Society (2013). https://doi.org/10.1109/ICCV.2013.405, https://doi.org/10.1109/ICCV.2013.405", + "43. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., Clarkson, A., Yan, M., Budge, B., Yan, Y., Pan, X., Yon, J., Zou, Y., Leon, K., Carter, N., Briales, J., Gillingham, T., Mueggler, E., Pesqueira, L., Savva, M., Batra, D., Strasdat, H.M., Nardi, R.D., Goesele, M., Lovegrove, S., Newcombe, R.: The replica dataset: A digital replica of indoor spaces (2019)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "CG-SLAM", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "44. Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for the evaluation of rgb-d slam systems. In: 2012 IEEE/RSJ international conference on intelligent robots and systems. pp. 573-580. IEEE (2012)", + "45. Sucar, E., Liu, S., Ortiz, J., Davison, A.J.:imap: Implicit mapping and positioning in real-time. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6229-6238 (2021)", + "46. Tateno, K., Tombari, F., Laina, I., Navab, N.: CNN-SLAM: real-time dense monocular SLAM with learned depth prediction. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017. pp. 6565-6574. IEEE Computer Society (2017). https://doi.org/10.1109/CVPR.2017.695, https://doi.org/10.1109/CVPR.2017.695", + "47. Teed, Z., Deng, J.: DROID-SLAM: deep visual SLAM for monocular, stereo, and RGB-D cameras. In: Ranzato, M., Beygelzimer, A., Dauphin, Y.N., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual. pp. 16558-16569 (2021), https://proceedings.neurips.cc/paper/2021/hash/89fcd07f20b6785b92134bd6c1d0fa42-AAbstract.html", + "48. Tretschk, E., Golyanik, V., Zollhöfer, M., Bozic, A., Lassner, C., Theobalt, C.: Scenerflow: Time-consistent reconstruction of general dynamic scenes. In: International Conference on 3D Vision (3DV) (2024)", + "49. Vespa, E., Nikolov, N., Grimm, M., Nardi, L., Kelly, P.H.J., Leutenegger, S.: Efficient octree-based volumetric SLAM supporting signed-distance and occupancy mapping. IEEE Robotics Autom. Lett. 3(2), 1144-1151 (2018). https://doi.org/10.1109/LRA.2018.2792537, https://doi.org/10.1109/LRA.2018.2792537", + "50. Wang, H., Wang, J., Agapito, L.: Co-slam: Joint coordinate and sparse parametric encodings for neural real-time slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13293-13302 (2023)", + "51. Wang, T., Zhang, B., Zhang, T., Gu, S., Bao, J., Baltrusaitis, T., Shen, J., Chen, D., Wen, F., Chen, Q., et al.: Rodin: A generative model for sculpting 3d digital avatars using diffusion. arXiv preprint arXiv:2212.06135 (2022)", + "52. Whelan, T., Salas-Moreno, R.F., Glocker, B., Davison, A.J., Leutenegger, S.: Elastic fusion: Real-time dense SLAM and light source estimation. Int. J. Robotics Res. 35(14), 1697-1716 (2016). https://doi.org/10.1177/0278364916669237, https://doi.org/10.1177/0278364916669237", + "53. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. arXiv preprint arXiv:2201.08845 (2022)", + "54. Yan, C., Qu, D., Xu, D., Zhao, B., Wang, Z., Wang, D., Li, X.: Gs-slam: Dense visual slam with 3d gaussian splatting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19595-19604 (2024)", + "55. Yang, N., von Stumberg, L., Wang, R., Cremers, D.: D3VO: deep depth, deep pose and deep uncertainty for monocular visual odometry. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020. pp. 1278-1289. Computer Vision Foundation / IEEE (2020). https://doi.org/10.1109/CVPR42600.2020.00136, https://openaccess.thecvf.com/content_CVPR_2020/html/Yang_D3VO_Dep_depth_Decel_Pose_and_Decel_Uncertainty_for_Monocular_CVPR_2020_paper.html", + "56. Yang, X., Li, H., Zhai, H., Ming, Y., Liu, Y., Zhang, G.: Vox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. In: 2022 IEEE International Symposium on Mixed and Augmented Reality (ISMAR). pp. 499-507. IEEE (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Jiarui Hu and Xianhao Chen et al.", + "bbox": [ + 271, + 114, + 501, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "57. Ye, M., Danelljan, M., Yu, F., Ke, L.: Gaussian grouping: Segment and edit anything in 3d scenes. arXiv preprint arXiv:2312.00732 (2023)", + "58. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelNeRF: Neural radiance fields from one or few images. In: CVPR (2021)", + "59. Yuan, Y., Li, X., Huang, Y., De Mello, S., Nagano, K., Kautz, J., Iqbal, U.: Gavatar: Animatable 3d gaussian avatars with implicit mesh learning. arXiv preprint arXiv:2312.11461 (2023)", + "60. Zhou, S., Chang, H., Jiang, S., Fan, Z., Zhu, Z., Xu, D., Chari, P., You, S., Wang, Z., Kadambi, A.: Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. arXiv preprint arXiv:2312.03203 (2023)", + "61. Zhu, Z., Peng, S., Larsson, V., Xu, W., Bao, H., Cui, Z., Oswald, M.R., Pollefeys, M.: Nice-slam: Neural implicit scalable encoding for slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12786-12796 (2022)", + "62. Zielonka, W., Bagautdinov, T., Saito, S., Zollhöfer, M., Thies, J., Romero, J.: Drivable 3d gaussian avatars (2023)", + "63. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Ewa volume splatting. In: Proceedings Visualization, 2001. VIS'01. pp. 29-538. IEEE (2001)" + ], + "bbox": [ + 212, + 146, + 787, + 397 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "CG-SLAM", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/CG-SLAM_ Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field/4bc60e08-fd5b-49a8-aee4-fd600735bd69_model.json b/2024/CG-SLAM_ Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field/4bc60e08-fd5b-49a8-aee4-fd600735bd69_model.json new file mode 100644 index 0000000000000000000000000000000000000000..42caf41e254bfd87b52e23cf5e28a7ba1eddb842 --- /dev/null +++ b/2024/CG-SLAM_ Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field/4bc60e08-fd5b-49a8-aee4-fd600735bd69_model.json @@ -0,0 +1,2647 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.223, + 0.141, + 0.782, + 0.187 + ], + "angle": 0, + "content": "CG-SLAM: Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.212, + 0.772, + 0.245 + ], + "angle": 0, + "content": "Jiarui Hu\\(^{1*}\\), Xianhao Chen\\(^{2*}\\), Boyin Feng\\(^{1}\\), Guanglin Li\\(^{1}\\), Liangjing Yang\\(^{2}\\), Hujun Bao\\(^{1}\\), Guofeng Zhang\\(^{1}\\), and Zhaopeng Cui\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.255, + 0.724, + 0.285 + ], + "angle": 0, + "content": "\\(^{1}\\) State Key Lab of CAD&CG, Zhejiang University \\(^{2}\\) ZJU-UIUC Institute, International Campus, Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.318, + 0.744, + 0.555 + ], + "angle": 0, + "content": "Abstract. Recently neural radiance fields (NeRF) have been widely exploited as 3D representations for dense simultaneous localization and mapping (SLAM). Despite their notable successes in surface modeling and novel view synthesis, existing NeRF-based methods are hindered by their computationally intensive and time-consuming volume rendering pipeline. This paper presents an efficient dense RGB-D SLAM system, i.e., CG-SLAM, based on a novel uncertainty-aware 3D Gaussian field with high consistency and geometric stability. Through an in-depth analysis of Gaussian Splatting, we propose several techniques to construct a consistent and stable 3D Gaussian field suitable for tracking and mapping. Additionally, a novel depth uncertainty model is proposed to ensure the selection of valuable Gaussian primitives during optimization, thereby improving tracking efficiency and accuracy. Experiments on various datasets demonstrate that CG-SLAM achieves superior tracking and mapping performance with a notable tracking speed of around \\(15\\mathrm{Hz}\\). We will make our source code publicly available. Project page: https://zju3dv.github.io/cg-slam." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.566, + 0.741, + 0.58 + ], + "angle": 0, + "content": "Keywords: Dense Visual SLAM \\(\\cdot\\) Neural Rendering \\(\\cdot\\) 3D Gaussian field" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.606, + 0.377, + 0.621 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.788, + 0.772 + ], + "angle": 0, + "content": "Dense visual Localization and Mapping (Visual SLAM) is a long-standing problem in 3D computer vision over recent decades, which targets performing pose tracking and scene mapping simultaneously with a variety of downstream applications such as virtual/augmented reality (VR/AR), robot navigation, and autonomous driving. Traditional visual SLAM systems [24] have shown accurate tracking performance across various scenes, while the underlying 3D representations (e.g., point cloud, mesh, and surfel) demonstrate limitations in facilitating highly free scene exploration, such as photorealistic scene touring, fine-grained map updating, etc." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.772, + 0.788, + 0.803 + ], + "angle": 0, + "content": "Inspired by the Neural Radiance Field (NeRF) [29] in surface reconstruction and view rendering, some novel NeRF-based SLAM methods [17,28,37,61] have" + }, + { + "type": "page_footnote", + "bbox": [ + 0.238, + 0.811, + 0.663, + 0.825 + ], + "angle": 0, + "content": "*Jiarui Hu and Xianhao Chen contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.241, + 0.825, + 0.406, + 0.84 + ], + "angle": 0, + "content": "†Corresponding authors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.505, + 0.128 + ], + "angle": 0, + "content": "Jiarui Hu and Xianhao Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.307, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.146, + 0.514, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.146, + 0.695, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.147, + 0.785, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.247, + 0.788, + 0.332 + ], + "angle": 0, + "content": "Fig. 1: CG-SLAM, which adopts a well-designed 3D Gaussian field, can simultaneously achieve state-of-the-art performance in localization, reconstruction and rendering. Benefiting from 3D Gaussian representation and a new GPU-accelerated framework that is developed from a thorough derivative analysis of camera pose in 3D Gaussian Splating [23], CG-SLAM can perform extremely fast rendering and solve the long-standing efficiency bottleneck suffered by previous rendering-based SLAM methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.488 + ], + "angle": 0, + "content": "been proposed recently and demonstrated promising performance in tracking, surface modeling, and novel view synthesis. Nevertheless, existing NeRF-based methods follow the ray-tracing rendering pipeline that is computation-intensive and time-consuming. Therefore they can only perform tracking and mapping by sampling a limited number of camera rays, ignoring the natural structural information in images. To avoid local optima in tracking and artifacts, they normally require many optimization steps, which makes it struggle to bring the best of both worlds concerning accuracy and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.788, + 0.672 + ], + "angle": 0, + "content": "Very recently the 3D Gaussian Splatting [23] method has been introduced for efficient novel view synthesis, and its rasterization-based rendering pipeline allows much faster image-level rendering, showing great potential in solving the inherent challenge of NeRF-based SLAM. However, it is nontrivial to reasonably incorporate the 3D Gaussian field in the SLAM setting. As a photorealistic view synthesis technique, the 3D Gaussian field is prone to overfitting the input images due to strong anisotropy and the lack of explicit multi-view constraints. As a result, on one hand, the 3D Gaussian splatting can not guarantee accurate modeling of 3D surfaces; on the other hand, since the Gaussians may not align with the environment's surfaces, this will lead to poor extrapolation capability and further degrade the camera tracking. Moreover, the increase of Gaussians in the mapping process will inevitably slow down the tracking efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we introduce a real-time Gaussian splatting SLAM system, i.e., CG-SLAM, based on a novel uncertainty-aware 3D Gaussian field with high consistency and geometric stability. To this end, we first conduct a comprehensive mathematical analysis regarding the derivatives of camera poses in the EWA (Elliptical Weighted Average) splatting process [63], and develop a CUDA framework tailored for the SLAM task that effectively decouples the tracking and mapping components. Second, in order to reduce the inherent overfitting problem, we utilize a scale regularization term that appropriately encourages the Gaussian ellipsoids to approximate Gaussian spheres, to reduce anisotropy and achieve a good trade-off between tracking accuracy and rendering realism. At the same time, we observed that solely employing alpha-blending depth cannot" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CG-SLAM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.297 + ], + "angle": 0, + "content": "impose effective constraints on the positions of Gaussian primitives. Thus, towards high-quality mapping, we further align median depth and alpha-blending depth to encourage Gaussian primitives to be well distributed over the scene surfaces, facilitating a consistent Gaussian field with more concentrated geometry density. Furthermore, in order to further improve the system's accuracy and efficiency, we design a novel depth uncertainty model to guide our Gaussian-based SLAM to focus on those stable and informative ones. We evaluate our system on a wide variety of RGB-D datasets, and the experimental results demonstrate that our CG-SLAM has superior performance in terms of tracking accuracy, reconstruction quality, and runtime efficiency." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.299, + 0.651, + 0.312 + ], + "angle": 0, + "content": "Overall, our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.324, + 0.784, + 0.368 + ], + "angle": 0, + "content": "- We present a new GPU-accelerated framework for real-time dense RGB-D SLAM based on a thorough theoretical analysis of camera pose derivatives in 3D Gaussian Splatting." + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.37, + 0.784, + 0.398 + ], + "angle": 0, + "content": "- We design multiple loss terms to build up a consistent and stable 3D Gaussian field suitable for tracking and mapping." + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.4, + 0.784, + 0.441 + ], + "angle": 0, + "content": "- We propose a novel depth uncertainty model, which assists our system in selecting more valuable Gaussian primitives during optimization, thereby improving tracking efficiency and accuracy." + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.445, + 0.784, + 0.472 + ], + "angle": 0, + "content": "- Experiments on various datasets demonstrate that our method can achieve competitive or better tracking and mapping results compared to baselines." + }, + { + "type": "list", + "bbox": [ + 0.23, + 0.324, + 0.784, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.497, + 0.388, + 0.513 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.528, + 0.431, + 0.542 + ], + "angle": 0, + "content": "2.1 Dense Visual SLAM" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.785, + 0.733 + ], + "angle": 0, + "content": "Following the seminal contributions of DTAM [32] and KinectFusion [31] to dense visual SLAM systems, there has been significant progress in developing efficient scene representation models. TSDF [31, 32], surfels [22, 39, 52], voxel hashing [6, 33, 34], and octrees [42, 49] have been introduced to tackle the challenges of scalability within the SLAM task. Some more advanced technologies, including bundle adjustment [39], loop closure, and learning-based algorithms [25, 26, 46, 47, 55], were subsequently integrated into the SLAM framework to further improve system performance. These enhancements have significantly brought better accuracy and robustness for localization and reconstruction capabilities, thus pushing the frontiers of what is achievable in dense SLAM systems. Compared with traditional methods, our CG-SLAM can reconstruct a fully dense 3D map for rich applications." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.755, + 0.625, + 0.77 + ], + "angle": 0, + "content": "2.2 Neural Implict Radiance Field based SLAM" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Neural radiance fields [29] have shown promising potential in many 3D computer vision applications, such as novel view synthesis [3, 4, 14, 30], dynamic scene modeling [13, 15, 41, 48], and generalization [18, 40, 51, 53, 58]. Recent research works [27, 45] attempted to replace traditional maps, in tasks such as structure" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.504, + 0.128 + ], + "angle": 0, + "content": "Jiarui Hu and Xianhao Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.788, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.304, + 0.789, + 0.404 + ], + "angle": 0, + "content": "Fig. 2: System Overview. In a 3D Gaussian field constructed from an RGB-D sequence, we can render color, depth, opacity, and uncertainty maps through a GPU-accelerated rasterizer. Additionally, we attach a new uncertainty property to each Gaussian primitive to filter informative primitives. In the mapping process, we utilize multiple rendering results to design effective loss functions towards a consistent and stable Gaussian field. Subsequently, we employ appearance and geometry cues to perform accurate and efficient tracking." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.789, + 0.629 + ], + "angle": 0, + "content": "from motion (SFM) and SLAM, with the neural implicit field to jointly optimize scene representation and camera poses. Different kinds of neural fields brought insights into NeRF-SLAM works. NICE-SLAM [61] chose a fully covered voxel grid to store neural features, while Vox-Fusion [56] further improved this grid to an adaptive size. Besides, Point-SLAM [37] attached feature embeddings to the point cloud on object surfaces. The neural point-based method is more flexible and can encode more concentrated volume density. Co-SLAM [50] adopted a hybrid representation that includes coordinate encoding and hash grids to achieve smoother reconstruction and faster convergence. In addition to the aforementioned works, some methods [8, 20, 36] only used the neural field as a map and still performed tracking based on a traditional feature point-based visual odometry. Our CG-SLAM system can achieve better and more efficient performance in tracking, mapping, and rendering than NeRF-based methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.65, + 0.497, + 0.665 + ], + "angle": 0, + "content": "2.3 3D Gaussian Splitting Field" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.842 + ], + "angle": 0, + "content": "3D Gaussian splatting is a revolutionary novel-view synthesis approach in 3D computer vision. This approach does not contain any neural network and allows photorealistic real-time(≥100 FPS) rendering at 1080p resolution. 3D Gaussian splatting has influenced a wide range of research topics, such as the 3D avatar [1, 19, 35, 59, 62], scene editing [7, 12, 57], image segment [60], and surface reconstruction [5, 11, 16]. These studies have demonstrated its generalization and effectiveness. Similar to the neural field, some concurrent manuscripts [21, 54] have been made to reversely infer camera poses within a built 3D Gaussian field. However, these works straightforwardly apply a raw 3D Gaussian field in the SLAM framework without specialized designs, such as anisotropy regularization and uninformative Gaussian primitive filtering. Additionally, they overlook" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CG-SLAM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.222 + ], + "angle": 0, + "content": "the design on efficiency, which is the most important improvement that the Gaussian splatting technique should bring to a SLAM system. We believe that the advantages of the 3D Gaussian field in pose optimization have not been fully explored in these works, especially in terms of efficiency, and expect to further develop an advanced Gaussian-based visual SLAM system." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.246, + 0.331, + 0.263 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.278, + 0.788, + 0.46 + ], + "angle": 0, + "content": "The overview of our proposed rasterization-based Gaussian SLAM system is shown in Fig. 2. Given a set of RGB-D sequences, our system incrementally generates a stable, consistent, and uncertainty-aware Gaussian field, serving camera pose optimization and geometry reconstruction. In Sec. 3.1, we briefly introduce the 3D Gaussian splatting model and rasterization principles. We incorporated an uncertainty model that utilizes the geometry prior to attach the uncertainty property on rendered images and Gaussian primitives. This strategy helps remove outliers in mapping and makes full use of informative Gaussians in tracking (Sec. 3.2). Moreover, in Sec. 3.3, we detail the Gaussian primitive management strategy and some innovative loss terms that ensure geometry stability and accuracy. Finally, by minimizing the re-rendering loss from low-uncertainty primitives, we can build a real-time and accurate tracking module (Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.481, + 0.362, + 0.497 + ], + "angle": 0, + "content": "3.1 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.788, + 0.598 + ], + "angle": 0, + "content": "Scene Representation. 3D Gaussian Splatting defines a 3D scene as a set of anisotropic Gaussian distributions, which are associated with means \\(\\mathbf{X} \\in \\mathbb{R}^3\\) and covariances \\(\\boldsymbol{\\Sigma} \\in \\mathbb{R}^{3 \\times 3}\\). To ensure that the covariance matrix remains positive semi-definite throughout the gradient descent, it is endowed with a more intuitive and comprehensible physical meaning, that is, the configuration of an ellipsoid. Specifically, \\(\\boldsymbol{\\Sigma}\\) is simplified and decomposed into:" + }, + { + "type": "equation", + "bbox": [ + 0.445, + 0.609, + 0.785, + 0.625 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {T} \\mathbf {R} ^ {T}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.788, + 0.711 + ], + "angle": 0, + "content": "where scaling matrix \\(\\mathbf{S} = \\text{diag}([\\mathbf{s}])\\) is derived from the scale factor \\(\\mathbf{s} \\in \\mathbb{R}^3\\) and rotation matrix \\(\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}\\) is derived from the quaternion \\(\\mathbf{q} \\in \\mathbb{R}^4\\). Each Gaussian ellipsoid is also assigned an opacity \\(\\sigma \\in \\mathbb{R}\\) and spherical harmonics \\((SH)\\) coefficients, which respectively represent the volume density and view-dependent radiance within a nearby local region." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.712, + 0.788, + 0.802 + ], + "angle": 0, + "content": "Following Zwicker et al. [63], given a world-to-camera rotation \\(\\mathbf{W}\\) and the Jacobian \\(\\mathbf{J}\\) of the affine approximation of the projective transformation, the EWA splatting algorithm illustrates how to approximately project a 3D Gaussian ellipsoid onto the image plane to determine its effective range and per-pixel opacity values on this image plane. We can obtain the corresponding 2D Gaussian distribution \\(\\mathcal{N}(\\hat{\\mu},\\hat{\\Sigma})\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.433, + 0.801, + 0.785, + 0.819 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {\\Sigma}} = \\mathbf {J} \\mathbf {W} \\boldsymbol {\\Sigma} \\mathbf {W} ^ {T} \\mathbf {J} ^ {T}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.825, + 0.681, + 0.841 + ], + "angle": 0, + "content": "and \\(\\hat{\\mu}\\) is the 2D pixel location of a 3D gaussian primitive center." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.505, + 0.128 + ], + "angle": 0, + "content": "Jiarui Hu and Xianhao Chen et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "Fast Rasterization-based Rendering. Fast Gaussian splatting rasterizer enables efficient pixel-by-pixel parallel rendering, and is fully differentiable, which provides a useful GPU-accelerated framework. For an incoming frame, the rasterizer can pre-sort all visible Gaussian primitives in order of depth from near to far. In terms of color rendering, the Gaussian splatting rasterizer adopts an \\(\\alpha\\)-blending solution, which accumulates radiance \\(c\\) and opacity values \\(\\sigma\\) on a given pixel by traversing the above depth queue as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.447, + 0.263, + 0.786, + 0.304 + ], + "angle": 0, + "content": "\\[\n\\hat {I} = \\sum_ {i = 1} ^ {N} \\alpha_ {i} T _ {i} c _ {i}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.436, + 0.317, + 0.786, + 0.358 + ], + "angle": 0, + "content": "\\[\nT _ {i} = \\prod_ {k = 1} ^ {i - 1} \\left(1 - \\alpha_ {k}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.436, + 0.367, + 0.786, + 0.385 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} = \\mathcal {N} \\left(\\hat {\\mu} _ {\\mathbf {i}}, \\hat {\\Sigma} _ {\\mathbf {i}}\\right) \\sigma_ {i}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.788, + 0.482 + ], + "angle": 0, + "content": "where \\(\\hat{I}\\) is the rendered color, \\(T_{i}\\) is the accumulated transmittance, \\(\\alpha_{i}\\) is the opacity contributed to a pixel, and \\(c_{i}\\) is the color of a Gaussian primitive computed from its SH coefficients. \\(N\\) is the number of Gaussian primitives involved in the splatting process of a pixel. In terms of depth rendering, considering the loss term designed for geometry consistency, our rasterizer provides not only \\(\\alpha\\)-blending depth \\(\\hat{D}_{\\text{alpha}}\\) but also the median depth \\(\\hat{D}_{\\text{median}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.494, + 0.786, + 0.533 + ], + "angle": 0, + "content": "\\[\n\\hat {D} _ {\\text {a l p h a}} = \\sum_ {i = 1} ^ {N} \\alpha_ {i} T _ {i} d _ {i}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.431, + 0.548, + 0.785, + 0.566 + ], + "angle": 0, + "content": "\\[\n\\hat {D} _ {\\text {m e d i a n}} = d _ {\\text {m e d i a n}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.571, + 0.788, + 0.632 + ], + "angle": 0, + "content": "where \\( d_{i} \\) is the depth of a Gaussian primitive. For a pixel, in its splatting process, we regard a Gaussian at which the cumulative transmittance \\( T \\) falls below 0.5 for the first time as the \"median Gaussian\". Its depth is recorded as \\( d_{\\text{median}} \\). \\( T_{\\text{median}} \\) is the cumulative transmittance at this median Gaussian." + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.644, + 0.785, + 0.66 + ], + "angle": 0, + "content": "\\[\n\\left(T _ {\\text {m e d i a n}} \\geq 0. 5\\right) \\quad \\text {a n d} \\quad \\left(T _ {\\text {m e d i a n} + 1} < 0. 5\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.669, + 0.788, + 0.701 + ], + "angle": 0, + "content": "Besides, the accumulated opacity value \\(\\hat{O}\\) is similarly required to distinguish unobserved areas for spawning Gaussians:" + }, + { + "type": "equation", + "bbox": [ + 0.451, + 0.711, + 0.786, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\hat {O} = \\sum_ {i = 1} ^ {N} \\alpha_ {i} T _ {i}. \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.771, + 0.447, + 0.787 + ], + "angle": 0, + "content": "3.2 Uncertainty Modeling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Uncertainty model remains a trending topic in multi-view 3D reconstruction in recent decades. Inspired by [38], we believe that explicitly modeling uncertainty in our 3D Gaussian field has a positive effect on increasing the ratio of informative" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CG-SLAM" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.145, + 0.713, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.29, + 0.784, + 0.332 + ], + "angle": 0, + "content": "Fig. 3: Uncertainty of the Gaussian Primitives. Uncertainty of a Gaussian primitive is derived from its dominated pixels and corresponding depth biases, reflecting the geometric value and confidence of this primitive." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.363, + 0.784, + 0.408 + ], + "angle": 0, + "content": "Gaussians, which is crucial for the robustness and conciseness of a SLAM system. Hence, we propose an uncertainty model suitable for RGB-D observations from two perspectives: rendering images and Gaussian primitives." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.408, + 0.784, + 0.483 + ], + "angle": 0, + "content": "Uncertainty Map. \\(\\alpha\\)-blending depth is essentially an expected value calculated by sampling Gaussian ellipsoids along a pixel ray. Under the reasonable assumption of a normal distribution, we think that the uncertainty map is highly related to its variance. We can render an uncertainty value from the 3D Gaussian field as in Eq. (10)." + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.484, + 0.785, + 0.525 + ], + "angle": 0, + "content": "\\[\nU = \\sum_ {i = 1} ^ {N} \\alpha_ {i} T _ {i} \\left(d _ {i} - D\\right) ^ {2}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.531, + 0.784, + 0.591 + ], + "angle": 0, + "content": "where \\(D\\) represents depth observations from the camera sensor. To mitigate drastic changes in positions of Gaussian primitives during optimization, we proposed a geometry variance loss term (Eq. (11)) based on the \\(H \\times W\\) uncertainty map to force them to be near the observed depth." + }, + { + "type": "equation", + "bbox": [ + 0.418, + 0.603, + 0.785, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v a r} = \\frac {1}{H W} \\sum_ {n = 1} ^ {H W} | U _ {n} |. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.656, + 0.784, + 0.701 + ], + "angle": 0, + "content": "Uncertainty of Gaussian primitives. From the perspective of geometric consistency, we design a loss term \\(\\mathcal{L}_{\\text{align}}\\) as in Eq. (12) to align the \\(\\alpha\\)-blending depth and median depth." + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.713, + 0.785, + 0.753 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l i g n}} = \\frac {1}{H W} \\sum_ {n = 1} ^ {H W} \\left| \\hat {D} _ {\\text {a l p h a}} ^ {n} - \\hat {D} _ {\\text {m e d i a n}} ^ {n} \\right|. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.765, + 0.784, + 0.84 + ], + "angle": 0, + "content": "The \\(\\alpha\\)-blending depth of a pixel is dominated by the Gaussian primitive with the largest weight. We call this pixel a \"dominated pixel\" of this maximum weight Gaussian primitive. Our alignment loss forces \\(\\hat{D}_{\\text{alpha}}\\) and \\(\\hat{D}_{\\text{median}}\\) to be similar, which makes this maximum weight Gaussian primitive always occur at \\(d_{\\text{median}}\\). Thus, as shown in Eq. (13), the uncertainty \\(v_i\\) of the \\(i\\)th Gaussian primitive" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.505, + 0.129 + ], + "angle": 0, + "content": "Jiarui Hu and Xianhao Chen et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.209 + ], + "angle": 0, + "content": "is determined by the difference between its depth and depth observations from all its dominated pixels within a keyframe window \\( F = \\{f_1, f_2, \\dots, f_k\\} \\). Fig. 3 further visualizes that a Gaussian primitive receives depth difference values from its dominated pixels in multiple viewpoints." + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.217, + 0.786, + 0.259 + ], + "angle": 0, + "content": "\\[\n\\nu_ {i} = \\frac {1}{M _ {1} + \\dots + M _ {k}} \\sum_ {f _ {k} \\in F} \\sum_ {p = 1} ^ {M _ {k}} \\alpha_ {i} ^ {k, p} T _ {i} ^ {k, p} \\left(D _ {p} ^ {k} - d _ {i} ^ {k}\\right) ^ {2}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.269, + 0.79, + 0.424 + ], + "angle": 0, + "content": "In a keyframe \\( f_{k} \\), \\( \\alpha_{i}^{k,p} \\) and \\( T_{i}^{k,p} \\) represent the opacity and transmittance of the \\( i \\)th Gaussian primitive on a pixel \\( p \\). \\( D_{p}^{k} \\) represents the depth observation on a pixel \\( p \\) in \\( f_{k} \\). \\( d_{i}^{k} \\) is the depth value of the \\( i \\)-th Gaussian primitive at \\( f_{k} \\). \\( \\{M_{1},\\dots,M_{k}\\} \\) are the number of dominated pixels of the \\( i \\)th Gaussian primitive in different keyframes. Benefiting from uncertainty modeling, we can regularly detect and remove unreliable Gaussian primitives with high uncertainty exceeding a threshold \\( \\tau = 0.025 \\). Specifically, during the mapping optimization, primitives with \\( v_{i} > \\tau \\) will be manually reduced to a low-opacity level. These low-opacity Gaussian primitives can be optimized again to remove truly irreversible ones, which is a more adaptive and reasonable strategy." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.444, + 0.339, + 0.459 + ], + "angle": 0, + "content": "3.3 Mapping" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.788, + 0.572 + ], + "angle": 0, + "content": "We employ various loss functions to update Gaussian properties, aiming for a consistent and stable Gaussian field. In addition to color and \\(SSIM\\) loss in the original 3D Gaussian splatting, previous experience from NeRF-SLAM works suggests that geometry loss is a necessary part. To overcome anisotropic interference (Arrow-shaped Gaussian primitives), we add a soft scale regularization loss in the mapping process. Note that our system performs the initialization at a slightly higher cost, i.e., more optimization iters." + }, + { + "type": "equation", + "bbox": [ + 0.401, + 0.582, + 0.786, + 0.622 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o l o r}} = \\frac {1}{H W} \\sum_ {n = 1} ^ {H W} \\left| \\hat {I} _ {n} - I _ {n} \\right|, \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.424, + 0.635, + 0.785, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s s i m} = S S I M (\\hat {I}, I), \\tag {15}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.669, + 0.785, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {g e o}} = \\frac {1}{H W} \\sum_ {n = 1} ^ {H W} \\left| \\hat {D} _ {\\text {a l p h a}} ^ {n} - D _ {n} \\right|, \\tag {16}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.343, + 0.716, + 0.785, + 0.752 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i s o} = \\frac {1}{G} \\sum_ {i \\in G} \\max \\left(\\frac {\\operatorname* {m a x} \\left(\\left\\{s _ {i} ^ {x} , s _ {i} ^ {y} , s _ {i} ^ {z} \\right\\}\\right)}{\\operatorname* {m i n} \\left(\\left\\{s _ {i} ^ {x} , s _ {i} ^ {y} , s _ {i} ^ {z} \\right\\}\\right)}, \\epsilon\\right) - \\epsilon , \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.756, + 0.788, + 0.801 + ], + "angle": 0, + "content": "where \\( I_{n} \\) and \\( D_{n} \\) represent ground-truth color and depth, \\( \\epsilon = 1.0 \\) is a hyperparameter that controls the level of anisotropy, and \\( G \\) represents all visible Gaussians under the current view." + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.81, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {m a p p i n g}} = \\omega_ {1} \\mathcal {L} _ {\\text {c o l o r}} + \\omega_ {2} \\mathcal {L} _ {\\text {s s i m}} + \\omega_ {3} \\mathcal {L} _ {\\text {g e o}} \\tag {18} \\\\ + \\omega_ {4} \\mathcal {L} _ {\\text {a l i g n}} + \\omega_ {5} \\mathcal {L} _ {\\text {i s o}} + \\omega_ {6} \\mathcal {L} _ {\\text {v a r}}. \\\\ \\end{array}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CG-SLAM" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "These loss functions customized for the SLAM task facilitate faster convergence in mapping and lay a solid foundation for subsequent tracking." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.268 + ], + "angle": 0, + "content": "Gaussian Management. In initialization, we densely project Gaussian primitives into 3D space based on depth observations of the first frame. In subsequent mapping, we set an empirical threshold \\(\\psi = 0.5\\) to extract unobserved or under-constructed pixels where \\(\\hat{O} < \\psi\\). Then, we utilize color and depth information on these pixels to spawn fresh Gaussian primitives. In addition, we inherited the original splitting, cloning, and removing strategy for Gaussian densification." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.29, + 0.335, + 0.305 + ], + "angle": 0, + "content": "3.4 Tracking" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.784, + 0.405 + ], + "angle": 0, + "content": "In our system, we have proposed the first comprehensive mathematical theory on derivatives w.r.t. pose in 3D Gaussian splatting framework (refer to supplementary), and empirically discovered that the Lie algebraic representation is more advantageous for camera tracking, especially for the rotation, in a Gaussian field. The camera pose optimization, i.e., rotation and translation \\(\\{\\mathfrak{so}(3)|T\\}\\), mainly includes two parts: sequential tracking and sliding bundle adjustment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.785, + 0.466 + ], + "angle": 0, + "content": "Sequential Tracking. Given the fixed scene representation, the camera pose is initially guessed via the constant speed assumption where the last pose is transformed by the last relative transformation, and then we refine this rough pose using similar photometric and geometric losses weighted by \\(\\lambda = \\{\\lambda_1,\\lambda_2\\}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.391, + 0.478, + 0.785, + 0.493 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t r a c k i n g}} = \\lambda_ {1} \\mathcal {L} _ {\\text {c o l o r}} + \\lambda_ {2} \\mathcal {L} _ {\\text {g e o}}, \\tag {19}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.52, + 0.785, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\{\\mathfrak {s o} (3) | T \\} = \\underset {\\{\\mathfrak {s o} (3) | T \\}} {\\operatorname {a r g m i n}} \\left(\\mathcal {L} _ {\\text {t r a c k i n g}}\\right). \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.786, + 0.703 + ], + "angle": 0, + "content": "Sliding Bundle Adjustment. Cumulative error is a typical problem in SLAM, also in Gaussian-based SLAM systems. To ease it, we set up a sliding window \\( F \\) containing \\( k \\) co-visible keyframes and jointly optimize camera extrinsic and scene representation in this window. Due to the efficiency, we encode keyframes into a descriptor pool with a pre-trained NetVLAD [2] model to determine co-visibility through the cosine similarity scores, instead of the view frustum overlap method in previous works. In addition to keyframes from NetVLAD [2], we also added the current frame and the most recent keyframes in the sliding window considering temporal associations. We still employ \\( \\mathcal{L}_{\\text{mapping}} \\) in sliding bundle adjustment, where \\( \\varPsi \\) is a set of all estimizable Gaussian properties." + }, + { + "type": "equation", + "bbox": [ + 0.37, + 0.718, + 0.785, + 0.743 + ], + "angle": 0, + "content": "\\[\n\\varPsi , \\{\\mathfrak {s o} (3) | T \\} = \\underset {\\varPsi , \\{\\mathfrak {s o} (3) | T \\}} {\\operatorname {a r g m i n}} \\left(\\mathcal {L} _ {\\text {m a p p i n g}}\\right). \\tag {21}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.764, + 0.376, + 0.781 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "In this section, we describe our experimental setup and validate that the proposed system can achieve improvement in both accuracy (Sec. 4.1 and Sec. 4.2) and efficiency (Sec. 4.3). We also confirmed the effectiveness of our design choices" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.503, + 0.128 + ], + "angle": 0, + "content": "Jiarui Hu and Xianhao Chen et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "(Sec. 4.4). Additionally, we demonstrate our advantages in image rendering and capability for online third-person view rendering in supplementary. We color each cell as best, second best, and third best." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.269 + ], + "angle": 0, + "content": "Datasets. To evaluate our system in various scenarios, we use three standard benchmarks: Replica [43], TUM [44], and ScanNet [10]. The Replica dataset contains 8 available synthetic RGB-D sequences generated by Sucar et al. [45]. We examined the generalization of our method on real-world TUM [44] and ScanNet [10] datasets, which contain 5 and 6 challenging scenes respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.27, + 0.787, + 0.435 + ], + "angle": 0, + "content": "Implementation Details. We run our system on a desktop equipped with an Intel i9-14900K and an NVIDIA RTX 4090 GPU. We set the learning rate of \\(\\{\\mathfrak{so}(3)|T\\}\\) to \\(\\{0.0015, 0.00215\\}\\) in sequential tracking in all experiments. For the Replica [43] dataset, we perform 60-iteration mapping with weights \\(\\omega = \\{0.7, 0.1, 0.25, 0.25, 0.1, 0.15\\}\\) in a sliding window with \\(k = 4\\) keyframes and 15-iteration sequential pose optimization weighted by \\(\\lambda = \\{0.2, 1.0\\}\\). We select keyframes at an interval of 30. For TUM [44] and ScanNet [10] datasets, we use \\(\\omega = \\{1.0, 0.1, 0.8, 0.5, 0.1, 0.5\\}\\), \\(k = 4\\), \\(\\lambda = \\{1.0, 0.6\\}\\). Faced with challenging real-world scenes, we need to extract more keyframes at an interval of 15, perform mapping at \\(40 \\sim 50\\) iterations, and increase tracking iterations to 25. For further implementation details, please refer to our supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.437, + 0.785, + 0.512 + ], + "angle": 0, + "content": "Metrics. We quantitatively evaluate reconstruction quality using different 3D metrics. Given 3D triangle meshes, we compute mapping Accuracy [cm], Completion [cm], and Completion Ratio \\([<5\\mathrm{cm}\\%]\\). Following NICE-SLAM [61], we discard unobserved regions that are not in any viewpoints. As for tracking performance, we measure ATE RMSE [44] for estimated trajectories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.513, + 0.785, + 0.604 + ], + "angle": 0, + "content": "Baselines. We primarily consider state-of-the-art NeRF-SLAM works, including NICE-SLAM [61], Co-SLAM [50], Point-SLAM [37], and Vox-Fusion [56], as baselines. For a fair comparison, we reproduced all results from these baselines and reported their reconstruction performance with the same evaluation mechanism. We also add some concurrent manuscripts such as GS-SLAM [54] and SplaTAM [21] for reference, and we directly report the results in their papers." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.63, + 0.456, + 0.644 + ], + "angle": 0, + "content": "4.1 Localization Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We report the localization accuracy of our system in 8 Replica [43] scenes in Tab. 1, where we surpass all other methods by a notable margin of around \\(25\\% \\sim 75\\%\\). In our Gaussian-based system, image-level pose optimization and the well-designed Gaussian field promote fast and stable convergence to an optimal solution. This is the reason why we have lower variances and higher accuracy. As shown in Tab. 2, despite noisy and sparse depth information in the real-world TUM-RGBD dataset [44], our method still achieves better or competitive performance in 5 selected scenarios. We also benchmark our method and baselines on the similarly challenging ScanNet [10] to compare their tracking performance in Tab. 3. Sensor data from multiple large-scale ScanNet scenes suffers from severe motion blur and specular reflections. Our method further demonstrates its effectiveness and superiority on this dataset, excelling or maintaining com" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CG-SLAM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: Tracking Results on the Replica Dataset [43] (ATE RMSE [cm] \\(\\downarrow\\)). Our system consistently achieved the best performance in this dataset, both for 8 individual scenes and for the average. It is worth noting that GS-SLAM [54] and SplaTAM [21] are concurrent with ours." + }, + { + "type": "table", + "bbox": [ + 0.309, + 0.212, + 0.694, + 0.324 + ], + "angle": 0, + "content": "
Methodrm-0rm-1rm-2off-0off-1off-2off-3off-4Avg.
NICE-SLAM0.971.311.070.881.001.061.101.131.06
Co-SLAM0.771.041.090.580.532.051.490.840.99
Point-SLAM0.560.470.300.350.620.550.720.730.54
Vox-Fusion0.400.540.540.500.460.750.500.600.54
GS-SLAM0.480.530.330.520.410.590.460.700.50
SplaTAM0.310.400.290.470.270.290.320.550.36
Ours0.290.270.250.330.140.280.310.290.27
Ours-light0.440.360.330.290.270.430.520.580.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.336, + 0.788, + 0.379 + ], + "angle": 0, + "content": "Table 2: Tracking Results on the TUM-RGBD Dataset [44] (ATE RMSE [cm] ↓). Our system achieves better tracking accuracy and lower variance in different scenarios. \"-\" indicates unavailable results because the related work is not open source." + }, + { + "type": "table", + "bbox": [ + 0.309, + 0.389, + 0.694, + 0.498 + ], + "angle": 0, + "content": "
Methodfr1/deskfr1/desk2fr1/roomfr2.xyzfr3/officeAvg.
NICE-SLAM4.264.9934.4931.733.8715.87
Co-SLAM2.74.5730.161.92.68.38
Point-SLAM4.344.5430.921.313.488.92
Vox-Fusion3.526.0019.531.4926.0111.31
GS-SLAM3.3--1.36.6-
SplaTAM3.356.5411.131.245.165.48
Ours2.434.549.391.202.454.0
Ours-light3.144.7310.671.282.604.48
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.553 + ], + "angle": 0, + "content": "Table 3: Tracking Results on the ScanNet Dataset [10] (ATE RMSE [cm] ↓). Our method achieves state-of-the-art tracking results in 6 scenes and exceeds other methods on average. \"-\" indicates failure results in Vox-Fusion [56]." + }, + { + "type": "table", + "bbox": [ + 0.309, + 0.563, + 0.694, + 0.661 + ], + "angle": 0, + "content": "
MethodSc.0000Sc.0059Sc.0106Sc.0169Sc.0181Sc.0207Avg.
NICE-SLAM12.0014.007.9010.9013.406.2010.70
Co-SLAM7.1812.2910.96.6213.437.139.37
Point-SLAM10.248.2911.8622.1614.779.5412.19
Vox-Fusion8.398.95-9.5012.206.43-
SplaTAM12.8310.1017.7212.0811.107.4611.88
Ours7.097.468.888.1611.605.348.08
Ours-light5.628.739.787.9312.025.458.17
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.787, + 0.718 + ], + "angle": 0, + "content": "petitiveness in various scenes. Extensive experiments showcase the remarkable ability of our proposed system to track and handle complex situations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.726, + 0.483, + 0.74 + ], + "angle": 0, + "content": "4.2 Reconstruction Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In Tab. 4, we quantitatively measure the mapping performance of our proposed system, in comparison to NICE-SLAM [61], Co-SLAM [50], Point-SLAM [37], and Vox-Fusion [56]. It can be observed that our method outperforms all baselines on mapping accuracy. We use TSDF-Fusion [9] to construct our mesh map. We achieve state-of-the-art reconstruction in observed areas. It is worth noting that the Gaussian-based method neither has a global MLP nor a fully covered" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.503, + 0.128 + ], + "angle": 0, + "content": "Jiarui Hu and Xianhao Chen et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.187 + ], + "angle": 0, + "content": "Table 4: Reconstruction Results on the Replica [43] Dataset. In terms of mapping accuracy, our method can outperform all existing methods. Due to the inherent limitation of 3D Gaussian representation, our method is slightly worse in completion." + }, + { + "type": "table", + "bbox": [ + 0.246, + 0.198, + 0.758, + 0.421 + ], + "angle": 0, + "content": "
MethodMetricrm-0rm-1rm-2off-0off-1off-2off-3off-4Avg.
NICE-SLAMAcc.[cm]↓3.533.603.035.563.354.713.843.353.87
Comp.[cm]↓3.403.623.274.554.033.943.994.153.87
Comp.Ratio<5cm%|↑86.0580.7587.2379.3482.1380.3580.5582.8882.41
Co-SLAMAcc.[cm]↓2.111.681.991.571.312.843.062.232.10
Comp.[cm]↓2.021.811.961.561.592.432.722.522.08
Comp.Ratio<5cm%|↑95.2695.1993.5896.0994.6591.6390.7290.4493.44
Point-SLAMAcc.[cm]↓1.451.141.191.050.861.311.571.511.26
Comp.[cm]↓3.463.022.651.652.213.623.473.903.00
Comp.Ratio<5cm%|↑88.4889.4490.1393.3990.5186.1786.0085.7488.73
Vox-FusionAcc.[cm]↓1.771.512.231.631.442.092.332.021.88
Comp.[cm]↓2.692.312.581.871.663.032.813.512.56
Comp.Ratio<5cm%|↑92.0392.4790.1393.8694.4088.9489.1086.5390.93
OursAcc.[cm]↓1.100.970.960.850.671.101.281.161.01
Comp.[cm]↓3.262.772.791.492.153.343.233.662.84
Comp.Ratio<5cm%|↑88.2689.4889.1093.6090.1486.0485.7885.6688.51
Ours-lightAcc.[cm]↓1.170.980.990.870.711.21.361.261.06
Comp.[cm]↓3.322.652.811.512.203.443.273.782.87
Comp.Ratio<5cm%|↑88.2089.3389.1093.4190.1485.6485.4785.3088.43
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.54 + ], + "angle": 0, + "content": "feature grid, as in Co-SLAM [50]. Consequently, such a system exhibits a slightly weaker hole-filling ability compared to the NeRF-based method, which explains why our system is slightly worse in the Completion metric. As shown in Fig. 4, we qualitatively present the ground truth mesh and mesh reconstructions from ours and other baselines. Evidently, our system achieves more detailed geometry and less noisy outliers." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.56, + 0.521, + 0.575 + ], + "angle": 0, + "content": "4.3 Runtime and Memory Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.787, + 0.779 + ], + "angle": 0, + "content": "We evaluate the runtime and memory footprint of our system compared to other works in Tab. 5. We reported the tracking and mapping efficiency in terms of per-iteration time consumption and the total number of optimization iterations. With the support of the GPU-accelerated rasterizer, our system can operate at around \\(8.5\\mathrm{Hz}\\). Meanwhile, our carefully designed pipeline allows this system to expand to a lightweight version and to work with half-resolution images. For tracking, this lightweight version can work twice as fast as the original one, at the cost of a slight decrease in accuracy, as shown in Tabs. 1 to 3. For mapping, our lightweight version demonstrates similar superiority in Tab. 4. It can be clearly seen that both our versions achieve better performance than Co-SLAM [50]. Further efficiency analysis on TUM-RGBD [44] and ScanNet [10] is shown in supplementary. Also, our customized Gaussian field allows us to outperform the concurrent Gaussian-based works with less computational burden." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "However, as a non-MLP scene representation, the 3D Gaussian field inevitably requires much memory consumption to store different properties. This reason results in a considerable memory footprint in the Gaussian-based SLAM system, which is a common limitation in other Gaussian-based research topics." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CG-SLAM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.788, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.321, + 0.788, + 0.363 + ], + "angle": 0, + "content": "Fig. 4: Reconstruction Performance on Replica [43] Dataset. We qualitatively compared the mesh reconstruction results from CG-SLAM and other baselines, where CG-SLAM can produce more detailed geometry at a lower computation cost." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.38, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Table 5: Runtime and Memory Usage. We comprehensively compared the runtime and memory usage on Replica [43] Office 0. Our proposed CG-SLAM can perform more efficient tracking and mapping than existing works, actually reaching a real-time level. \"-\" indicates unavailable results in related works." + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.447, + 0.731, + 0.58 + ], + "angle": 0, + "content": "
MethodTracking [ms× it]↓Mapping [ms× it]↓Mapping IntervalSystem FPS↑Decoder Param↓Scene Embedding↓
Vox-Fusion23.61 × 3086.55 × 10101.250.98 MB0.162 MB
NICE-SLAM6.19 × 1091.59 × 6050.860.43 MB89.56 MB
Co-SLAM4.45 × 1011.9 × 10514.646.37 MB-
Point-SLAM6.14 × 4022.25 × 30050.630.54 MB28.11 MB
GS-SLAM11.9 × 1012.8 × 100-8.34--
SplaTAM41.7 × 4050.1 × 6010.21--
Ours6.02 × 1028.9 × 60308.47-231.66 MB
Ours-light3.80 × 1010.70 × 603016.83-56.50 MB
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.61, + 0.389, + 0.626 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.787, + 0.672 + ], + "angle": 0, + "content": "To verify the rationality of our designs, we investigate the effectiveness of the anisotropy regularization, alignment and variance losses, and uncertainty model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.797, + 0.809 + ], + "angle": 0, + "content": "Effectiveness of Isotropy Loss. Ablation results in Tab. 6 reveal how the anisotropy regularization term affects tracking metrics. We can notice that anisotropy \\((\\mathrm{w} / \\mathrm{o}\\mathcal{L}_{iso})\\) will reduce the tracking accuracy by disrupting the view extrapolation of a 3D Gaussian field. Oppositely, excessive regularization (Isotropic Gaussians) will degrade rendering and subsequently affect the photometric loss, leading to poor tracking performance. Overall, a soft controllable constraint like \\(\\mathcal{L}_{iso}\\) is a reasonable choice. To more intuitively illustrate this phenomenon, we display opacity maps with and without anisotropy regularization in Fig. 5 (b), and apparent arrow-shaped artifacts occur in the latter." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Effectiveness of Alignment and Variance Losses. Alignment and variance losses push primitives closer to object surfaces, facilitating novel view syn-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.505, + 0.128 + ], + "angle": 0, + "content": "Jiarui Hu and Xianhao Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.144, + 0.744, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.255, + 0.788, + 0.325 + ], + "angle": 0, + "content": "Fig. 5: Uncertainty Model Ablation and Anisotropy Interference. (a) Uncertainty Model Ablation. This plot illustrates that the uncertainty model helps improve tracking accuracy while avoiding some extreme errors. (b) Anisotropy Interference. It can be clearly seen that in the case of w/o \\(\\mathcal{L}_{iso}\\), serious arrow-shaped artifacts occur on the edges of the image." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.341, + 0.788, + 0.381 + ], + "angle": 0, + "content": "Table 6: Isotropy Loss Ablation Results(ATE RMSE [cm] ↓). The experimental results demonstrate the effectiveness of our anisotropy regularization term. \"--\" indicates a failure situation." + }, + { + "type": "table", + "bbox": [ + 0.304, + 0.393, + 0.703, + 0.446 + ], + "angle": 0, + "content": "
Settingrm-0rm-1rm-2off-0off-1off-2off-3off-4Avg.
w/ Liso0.290.270.250.330.140.280.310.290.27
w/o Liso0.320.310.540.360.240.310.72--
Isotropic Gaussians0.300.321.040.297.620.381.541.361.61
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.785, + 0.505 + ], + "angle": 0, + "content": "thesis and SLAM. The effectiveness of these two losses is demonstrated in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.787, + 0.566 + ], + "angle": 0, + "content": "Effectiveness of Uncertainty Model. As shown in Fig. 5 (a), tracking error curves show that our uncertainty model can improve tracking accuracy while avoiding some extreme errors. For further quantitative ablation results, refer to the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.593, + 0.36, + 0.608 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.75 + ], + "angle": 0, + "content": "We have proposed CG-SLAM, a dense RGB-D SLAM based on a consistent and uncertainty-aware 3D Gaussian field. Our targeted loss functions strengthen the 3D Gaussian field in terms of consistency and stability. The uncertainty model further distills highly informative primitives in this field to reduce interference from outliers. Besides, a customized GPU-accelerated rasterization pipeline enables our system to achieve state-of-the-art accuracy and efficiency in various scenes. Through extensive experiments, it can be concluded that our method outperforms previous works regarding tracking, mapping, and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Limitations. Considerable memory usage is one limitation of the Gaussian-based system. We expect that a more compact Gaussian field can be adopted in the SLAM task. In addition, the Gaussian-based method has a weak prediction ability for unobserved areas. Moreover, our system is not capable of handling dynamic objects in the environment. We believe it is a very interesting direction for future work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CG-SLAM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.179, + 0.681, + 0.195 + ], + "angle": 0, + "content": "This work was partially supported by the NSFC (No. 62102356)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.22, + 0.323, + 0.235 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.253, + 0.785, + 0.281 + ], + "angle": 0, + "content": "1. Abdal, R., Yifan, W., Shi, Z., Xu, Y., Po, R., Kuang, Z., Chen, Q., Yeung, D.Y., Wetzstein, G.: Gaussian shell maps for efficient 3d human generation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.283, + 0.785, + 0.322 + ], + "angle": 0, + "content": "2. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn architecture for weakly supervised place recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5297-5307 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.785, + 0.363 + ], + "angle": 0, + "content": "3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.365, + 0.785, + 0.392 + ], + "angle": 0, + "content": "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.394, + 0.785, + 0.42 + ], + "angle": 0, + "content": "5. Chen, H., Li, C., Lee, G.H.: Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.422, + 0.785, + 0.462 + ], + "angle": 0, + "content": "6. Chen, J., Bautembach, D., Izadi, S.: Scalable real-time volumetric surface reconstruction. ACM Trans. Graph. 32(4), 113:1-113:16 (2013). https://doi.org/10.1145/2461912.2461940, https://doi.org/10.1145/2461912.2461940" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.464, + 0.785, + 0.504 + ], + "angle": 0, + "content": "7. Chen, Y., Chen, Z., Zhang, C., Wang, F., Yang, X., Wang, Y., Cai, Z., Yang, L., Liu, H., Lin, G.: Gaussianeditor: Swift and controllable 3d editing with gaussian splatting (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.506, + 0.785, + 0.587 + ], + "angle": 0, + "content": "8. Chung, C., Tseng, Y., Hsu, Y., Shi, X.Q., Hua, Y., Yeh, J., Chen, W., Chen, Y., Hsu, W.H.: Orbeez-slam: A real-time monocular visual SLAM with ORB features and nerf-realized mapping. In: IEEE International Conference on Robotics and Automation, ICRA 2023, London, UK, May 29 - June 2, 2023. pp. 9400-9406. IEEE (2023). https://doi.org/10.1109/ICRA48891.2023.10160950, https://doi.org/10.1109/ICRA48891.2023.10160950" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.589, + 0.785, + 0.63 + ], + "angle": 0, + "content": "9. Curless, B., Levoy, M.: A volumetric method for building complex models from range images. In: Proceedings of the 23rd annual conference on Computer graphics and interactive techniques. pp. 303-312 (1996)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.632, + 0.785, + 0.671 + ], + "angle": 0, + "content": "0. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.785, + 0.713 + ], + "angle": 0, + "content": "1. Duisterhof, B.P., Mandi, Z., Yao, Y., Liu, J.W., Shou, M.Z., Song, S., Ichnowski, J.: Md-splatting: Learning metric deformation from 4d gaussians in highly deformable scenes (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.715, + 0.785, + 0.742 + ], + "angle": 0, + "content": "2. Fang, J., Wang, J., Zhang, X., Xie, L., Tian, Q.: Gaussianeditor: Editing 3d gaussians delicately with text instructions. arXiv preprint arXiv:2311.16037 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.743, + 0.785, + 0.783 + ], + "angle": 0, + "content": "3. Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Niešner, M., Tian, Q.: Fast dynamic radiance fields with time-aware neural voxels. In: SIGGRAPH Asia 2022 Conference Papers (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.811 + ], + "angle": 0, + "content": "4. Fridovich-Keil and Yu, Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Geng, C., Peng, S., Xu, Z., Bao, H., Zhou, X.: Learning neural volumetric representations of dynamic humans in minutes. In: CVPR (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.253, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.503, + 0.128 + ], + "angle": 0, + "content": "Jiarui Hu and Xianhao Chen et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "16. Guédon, A., Lepetit, V.: Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. arXiv preprint arXiv:2311.12775 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "17. Hu, J., Mao, M., Bao, H., Zhang, G., Cui, Z.: Cp-slam: Collaborative neural point-based slam system. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "18. Irshad, M.Z., Zakharov, S., Liu, K., Guizilini, V., Kollar, T., Gaidon, A., Kira, Z., Ambrus, R.: Neo 360: Neural fields for sparse view synthesis of outdoor scenes (2023), https://arxiv.org/abs/2308.12967" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.259, + 0.785, + 0.3 + ], + "angle": 0, + "content": "19. Jena, R., Iyer, G.S., Choudhary, S., Smith, B., Chaudhari, P., Gee, J.: Splatarmor: Articulated gaussian splatting for animatable humans from monocular rgb videos. arXiv preprint arXiv:2311.10812 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.357 + ], + "angle": 0, + "content": "20. Katragadda, S., Lee, W., Peng, Y., Geneva, P., Chen, C., Guo, C., Li, M., Huang, G.: Nerf-vins: A real-time neural radiance field map-based visual-inertial navigation system. CoRR abs/2309.09295 (2023). https://doi.org/10.48550/ARXIV.2309.09295, https://doi.org/10.48550/arXiv.2309.09295" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.398 + ], + "angle": 0, + "content": "21. Keetha, N., Karhade, J., Jatavallabhula, K.M., Yang, G., Scherer, S., Ramanan, D., Luiten, J.: Splatam: Splat, track & map 3d gaussians for dense rgb-d slam. arXiv preprint arXiv:2312.02126 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.467 + ], + "angle": 0, + "content": "22. Keller, M., Lefloch, D., Lambers, M., Izadi, S., Weyrich, T., Kolb, A.: Real-time 3d reconstruction in dynamic scenes using point-based fusion. In: 2013 International Conference on 3D Vision, 3DV 2013, Seattle, Washington, USA, June 29 - July 1, 2013. pp. 1-8. IEEE Computer Society (2013). https://doi.org/10.1109/3DV.2013.9, https://doi.org/10.1109/3DV.2013.9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.495 + ], + "angle": 0, + "content": "23. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.536 + ], + "angle": 0, + "content": "24. Kerl, C., Sturm, J., Cremers, D.: Dense visual slam for rgb-d cameras. In: 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems. pp. 2100-2106. IEEE (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "25. Koestler, L., Yang, N., Zeller, N., Cremers, D.: Tandem: Tracking and dense mapping in real-time using deep multi-view stereo. In: Conference on Robot Learning (CoRL) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.577, + 0.785, + 0.619 + ], + "angle": 0, + "content": "26. Li, Y., Liu, X., Dong, W., Zhou, H., Bao, H., Zhang, G., Zhang, Y., Cui, Z.: Deltar: Depth estimation from a light-weight tof sensor and rgb image. In: European conference on computer vision. pp. 619-636. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.647 + ], + "angle": 0, + "content": "27. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: Barf: Bundle-adjusting neural radiance fields. In: IEEE International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.702 + ], + "angle": 0, + "content": "28. Liu, X., Li, Y., Teng, Y., Bao, H., Zhang, G., Zhang, Y., Cui, Z.: Multi-modal neural radiance field for monocular dense slam with a light-weight tof sensor. In: Proceedings of the iee/cvf international conference on computer vision. pp. 1-11 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.743 + ], + "angle": 0, + "content": "29. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: The European Conference on Computer Vision (ECCV) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.798 + ], + "angle": 0, + "content": "30. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022). https://doi.org/10.1145/3528223.3530127, https://doi.org/10.1145/3528223.3530127" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "31. Newcombe, R.A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohli, P., Shotton, J., Hodges, S., Fitzgibbon, A.W.: Kinectfusion: Real-time dense surface mapping and tracking. In: 10th IEEE International Symposium on Mixed" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CG-SLAM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "and Augmented Reality, ISMAR 2011, Basel, Switzerland, October 26-29, 2011. pp. 127-136. IEEE Computer Society (2011). https://doi.org/10.1109/ISMAR.2011.6092378, https://doi.org/10.1109/ISMAR.2011.6092378" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.27 + ], + "angle": 0, + "content": "32. Newcombe, R.A., Lovegrove, S., Davison, A.J.: DTAM: dense tracking and mapping in real-time. In: Metaxas, D.N., Quan, L., Sanfeliu, A., Gool, L.V. (eds.) IEEE International Conference on Computer Vision, ICCV 2011, Barcelona, Spain, November 6-13, 2011. pp. 2320-2327. IEEE Computer Society (2011). https://doi.org/10.1109/ICCV.2011.6126513, https://doi.org/10.1109/ICCV.2011.6126513" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.787, + 0.324 + ], + "angle": 0, + "content": "33. Nießner, M., Zollhöfer, M., Izadi, S., Stamminger, M.: Real-time 3d reconstruction at scale using voxel hashing. ACM Trans. Graph. 32(6), 169:1-169:11 (2013). https://doi.org/10.1145/2508363.2508374, https://doi.org/10.1145/2508363.2508374" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.787, + 0.407 + ], + "angle": 0, + "content": "34. Oleynikova, H., Taylor, Z., Fehr, M., Siegwart, R., Nieto, J.I.: Voxblox: Incremental 3d euclidean signed distance fields for on-board MAV planning. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2017, Vancouver, BC, Canada, September 24-28, 2017. pp. 1366-1373. IEEE (2017). https://doi.org/10.1109/IROS.2017.8202315, https://doi.org/10.1109/IROS.2017.8202315" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.787, + 0.434 + ], + "angle": 0, + "content": "35. Pang, H., Zhu, H., Kortylewski, A., Theobalt, C., Habermann, M.: Ash: Animatable gaussian splats for efficient and photoreal human rendering (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.787, + 0.487 + ], + "angle": 0, + "content": "36. Rosinol, A., Leonard, J.J., Carlone, L.: Nerf-slam: Real-time dense monocular SLAM with neural radiance fields. In: IROS. pp. 3437-3444 (2023). https://doi.org/10.1109/IROS55552.2023.10341922, https://doi.org/10.1109/IROS55552.2023.10341922" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.528 + ], + "angle": 0, + "content": "37. Sandström, E., Li, Y., Van Gool, L., Oswald, M.R.: Point-slam: Dense neural point cloud-based slam. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18433-18444 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.787, + 0.555 + ], + "angle": 0, + "content": "38. Sandström, E., Ta, K., Van Gool, L., Oswald, M.R.: Uncle-slam: Uncertainty learning for dense neural slam. arXiv preprint arXiv:2306.11048 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.787, + 0.595 + ], + "angle": 0, + "content": "39. Schops, T., Sattler, T., Pollefeys, M.: Bad slam: Bundle adjusted direct rgb-d slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.595, + 0.787, + 0.635 + ], + "angle": 0, + "content": "40. Schwarz, K., Liao, Y., Niemeyer, M., Geiger, A.: Graf: Generative radiance fields for 3d-aware image synthesis. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.635, + 0.787, + 0.69 + ], + "angle": 0, + "content": "41. Shao, R., Zheng, Z., Tu, H., Liu, B., Zhang, H., Liu, Y.: Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.69, + 0.787, + 0.758 + ], + "angle": 0, + "content": "42. Steinbrücker, F., Kerl, C., Cremers, D.: Large-scale multi-resolution surface reconstruction from RGB-D sequences. In: IEEE International Conference on Computer Vision, ICCV 2013, Sydney, Australia, December 1-8, 2013. pp. 3264-3271. IEEE Computer Society (2013). https://doi.org/10.1109/ICCV.2013.405, https://doi.org/10.1109/ICCV.2013.405" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.84 + ], + "angle": 0, + "content": "43. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., Clarkson, A., Yan, M., Budge, B., Yan, Y., Pan, X., Yon, J., Zou, Y., Leon, K., Carter, N., Briales, J., Gillingham, T., Mueggler, E., Pesqueira, L., Savva, M., Batra, D., Strasdat, H.M., Nardi, R.D., Goesele, M., Lovegrove, S., Newcombe, R.: The replica dataset: A digital replica of indoor spaces (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.503, + 0.128 + ], + "angle": 0, + "content": "Jiarui Hu and Xianhao Chen et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "44. Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for the evaluation of rgb-d slam systems. In: 2012 IEEE/RSJ international conference on intelligent robots and systems. pp. 573-580. IEEE (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.229 + ], + "angle": 0, + "content": "45. Sucar, E., Liu, S., Ortiz, J., Davison, A.J.:imap: Implicit mapping and positioning in real-time. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6229-6238 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.229, + 0.787, + 0.298 + ], + "angle": 0, + "content": "46. Tateno, K., Tombari, F., Laina, I., Navab, N.: CNN-SLAM: real-time dense monocular SLAM with learned depth prediction. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017. pp. 6565-6574. IEEE Computer Society (2017). https://doi.org/10.1109/CVPR.2017.695, https://doi.org/10.1109/CVPR.2017.695" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.787, + 0.392 + ], + "angle": 0, + "content": "47. Teed, Z., Deng, J.: DROID-SLAM: deep visual SLAM for monocular, stereo, and RGB-D cameras. In: Ranzato, M., Beygelzimer, A., Dauphin, Y.N., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual. pp. 16558-16569 (2021), https://proceedings.neurips.cc/paper/2021/hash/89fcd07f20b6785b92134bd6c1d0fa42-AAbstract.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.392, + 0.787, + 0.434 + ], + "angle": 0, + "content": "48. Tretschk, E., Golyanik, V., Zollhöfer, M., Bozic, A., Lassner, C., Theobalt, C.: Scenerflow: Time-consistent reconstruction of general dynamic scenes. In: International Conference on 3D Vision (3DV) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.787, + 0.489 + ], + "angle": 0, + "content": "49. Vespa, E., Nikolov, N., Grimm, M., Nardi, L., Kelly, P.H.J., Leutenegger, S.: Efficient octree-based volumetric SLAM supporting signed-distance and occupancy mapping. IEEE Robotics Autom. Lett. 3(2), 1144-1151 (2018). https://doi.org/10.1109/LRA.2018.2792537, https://doi.org/10.1109/LRA.2018.2792537" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.528 + ], + "angle": 0, + "content": "50. Wang, H., Wang, J., Agapito, L.: Co-slam: Joint coordinate and sparse parametric encodings for neural real-time slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13293-13302 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.787, + 0.569 + ], + "angle": 0, + "content": "51. Wang, T., Zhang, B., Zhang, T., Gu, S., Bao, J., Baltrusaitis, T., Shen, J., Chen, D., Wen, F., Chen, Q., et al.: Rodin: A generative model for sculpting 3d digital avatars using diffusion. arXiv preprint arXiv:2212.06135 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.787, + 0.623 + ], + "angle": 0, + "content": "52. Whelan, T., Salas-Moreno, R.F., Glocker, B., Davison, A.J., Leutenegger, S.: Elastic fusion: Real-time dense SLAM and light source estimation. Int. J. Robotics Res. 35(14), 1697-1716 (2016). https://doi.org/10.1177/0278364916669237, https://doi.org/10.1177/0278364916669237" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.622, + 0.787, + 0.651 + ], + "angle": 0, + "content": "53. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. arXiv preprint arXiv:2201.08845 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.787, + 0.69 + ], + "angle": 0, + "content": "54. Yan, C., Qu, D., Xu, D., Zhao, B., Wang, Z., Wang, D., Li, X.: Gs-slam: Dense visual slam with 3d gaussian splatting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19595-19604 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.69, + 0.787, + 0.786 + ], + "angle": 0, + "content": "55. Yang, N., von Stumberg, L., Wang, R., Cremers, D.: D3VO: deep depth, deep pose and deep uncertainty for monocular visual odometry. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020. pp. 1278-1289. Computer Vision Foundation / IEEE (2020). https://doi.org/10.1109/CVPR42600.2020.00136, https://openaccess.thecvf.com/content_CVPR_2020/html/Yang_D3VO_Dep_depth_Decel_Pose_and_Decel_Uncertainty_for_Monocular_CVPR_2020_paper.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "56. Yang, X., Li, H., Zhai, H., Ming, Y., Liu, Y., Zhang, G.: Vox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. In: 2022 IEEE International Symposium on Mixed and Augmented Reality (ISMAR). pp. 499-507. IEEE (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.658, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CG-SLAM" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "57. Ye, M., Danelljan, M., Yu, F., Ke, L.: Gaussian grouping: Segment and edit anything in 3d scenes. arXiv preprint arXiv:2312.00732 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.203 + ], + "angle": 0, + "content": "58. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelNeRF: Neural radiance fields from one or few images. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.788, + 0.245 + ], + "angle": 0, + "content": "59. Yuan, Y., Li, X., Huang, Y., De Mello, S., Nagano, K., Kautz, J., Iqbal, U.: Gavatar: Animatable 3d gaussian avatars with implicit mesh learning. arXiv preprint arXiv:2312.11461 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.245, + 0.788, + 0.286 + ], + "angle": 0, + "content": "60. Zhou, S., Chang, H., Jiang, S., Fan, Z., Zhu, Z., Xu, D., Chari, P., You, S., Wang, Z., Kadambi, A.: Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. arXiv preprint arXiv:2312.03203 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.342 + ], + "angle": 0, + "content": "61. Zhu, Z., Peng, S., Larsson, V., Xu, W., Bao, H., Cui, Z., Oswald, M.R., Pollefeys, M.: Nice-slam: Neural implicit scalable encoding for slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12786-12796 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.37 + ], + "angle": 0, + "content": "62. Zielonka, W., Bagautdinov, T., Saito, S., Zollhöfer, M., Thies, J., Romero, J.: Drivable 3d gaussian avatars (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.398 + ], + "angle": 0, + "content": "63. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Ewa volume splatting. In: Proceedings Visualization, 2001. VIS'01. pp. 29-538. IEEE (2001)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.398 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CG-SLAM_ Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field/4bc60e08-fd5b-49a8-aee4-fd600735bd69_origin.pdf b/2024/CG-SLAM_ Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field/4bc60e08-fd5b-49a8-aee4-fd600735bd69_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..73c684d085cb90a1667eabd8e91525c799299640 --- /dev/null +++ b/2024/CG-SLAM_ Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field/4bc60e08-fd5b-49a8-aee4-fd600735bd69_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d90a1d08afc55852c5f23e4248b3db559e87bed69fd94c28b7bc492ec247cb7 +size 9402023 diff --git a/2024/CG-SLAM_ Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field/full.md b/2024/CG-SLAM_ Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e0a453b5e94172170c570c1971da11c13957b8f7 --- /dev/null +++ b/2024/CG-SLAM_ Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field/full.md @@ -0,0 +1,360 @@ +# CG-SLAM: Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field + +Jiarui Hu $^{1*}$ , Xianhao Chen $^{2*}$ , Boyin Feng $^{1}$ , Guanglin Li $^{1}$ , Liangjing Yang $^{2}$ , Hujun Bao $^{1}$ , Guofeng Zhang $^{1}$ , and Zhaopeng Cui $^{1\dagger}$ + +$^{1}$ State Key Lab of CAD&CG, Zhejiang University $^{2}$ ZJU-UIUC Institute, International Campus, Zhejiang University + +Abstract. Recently neural radiance fields (NeRF) have been widely exploited as 3D representations for dense simultaneous localization and mapping (SLAM). Despite their notable successes in surface modeling and novel view synthesis, existing NeRF-based methods are hindered by their computationally intensive and time-consuming volume rendering pipeline. This paper presents an efficient dense RGB-D SLAM system, i.e., CG-SLAM, based on a novel uncertainty-aware 3D Gaussian field with high consistency and geometric stability. Through an in-depth analysis of Gaussian Splatting, we propose several techniques to construct a consistent and stable 3D Gaussian field suitable for tracking and mapping. Additionally, a novel depth uncertainty model is proposed to ensure the selection of valuable Gaussian primitives during optimization, thereby improving tracking efficiency and accuracy. Experiments on various datasets demonstrate that CG-SLAM achieves superior tracking and mapping performance with a notable tracking speed of around $15\mathrm{Hz}$ . We will make our source code publicly available. Project page: https://zju3dv.github.io/cg-slam. + +Keywords: Dense Visual SLAM $\cdot$ Neural Rendering $\cdot$ 3D Gaussian field + +# 1 Introduction + +Dense visual Localization and Mapping (Visual SLAM) is a long-standing problem in 3D computer vision over recent decades, which targets performing pose tracking and scene mapping simultaneously with a variety of downstream applications such as virtual/augmented reality (VR/AR), robot navigation, and autonomous driving. Traditional visual SLAM systems [24] have shown accurate tracking performance across various scenes, while the underlying 3D representations (e.g., point cloud, mesh, and surfel) demonstrate limitations in facilitating highly free scene exploration, such as photorealistic scene touring, fine-grained map updating, etc. + +Inspired by the Neural Radiance Field (NeRF) [29] in surface reconstruction and view rendering, some novel NeRF-based SLAM methods [17,28,37,61] have + +![](images/c9250fa538b056f7b1c254309eae5937950543481d07a2031d8b3afd8a5b5e9d.jpg) +Fig. 1: CG-SLAM, which adopts a well-designed 3D Gaussian field, can simultaneously achieve state-of-the-art performance in localization, reconstruction and rendering. Benefiting from 3D Gaussian representation and a new GPU-accelerated framework that is developed from a thorough derivative analysis of camera pose in 3D Gaussian Splating [23], CG-SLAM can perform extremely fast rendering and solve the long-standing efficiency bottleneck suffered by previous rendering-based SLAM methods. + +![](images/0a37d550eece540360ccfb3cf5fa13631ecb7fbadd4e083d0ff23e1d5f692175.jpg) + +![](images/93e5e1fd8a503283efa7c6ed73fbd240ba80cb21da472b44bb22ce0fef11e8fc.jpg) + +![](images/0bcb09f1f3f85d8e34eb29dbb90fa7931babfa73ad1f9b7562c289e028c4d8ba.jpg) + +been proposed recently and demonstrated promising performance in tracking, surface modeling, and novel view synthesis. Nevertheless, existing NeRF-based methods follow the ray-tracing rendering pipeline that is computation-intensive and time-consuming. Therefore they can only perform tracking and mapping by sampling a limited number of camera rays, ignoring the natural structural information in images. To avoid local optima in tracking and artifacts, they normally require many optimization steps, which makes it struggle to bring the best of both worlds concerning accuracy and efficiency. + +Very recently the 3D Gaussian Splatting [23] method has been introduced for efficient novel view synthesis, and its rasterization-based rendering pipeline allows much faster image-level rendering, showing great potential in solving the inherent challenge of NeRF-based SLAM. However, it is nontrivial to reasonably incorporate the 3D Gaussian field in the SLAM setting. As a photorealistic view synthesis technique, the 3D Gaussian field is prone to overfitting the input images due to strong anisotropy and the lack of explicit multi-view constraints. As a result, on one hand, the 3D Gaussian splatting can not guarantee accurate modeling of 3D surfaces; on the other hand, since the Gaussians may not align with the environment's surfaces, this will lead to poor extrapolation capability and further degrade the camera tracking. Moreover, the increase of Gaussians in the mapping process will inevitably slow down the tracking efficiency. + +In this paper, we introduce a real-time Gaussian splatting SLAM system, i.e., CG-SLAM, based on a novel uncertainty-aware 3D Gaussian field with high consistency and geometric stability. To this end, we first conduct a comprehensive mathematical analysis regarding the derivatives of camera poses in the EWA (Elliptical Weighted Average) splatting process [63], and develop a CUDA framework tailored for the SLAM task that effectively decouples the tracking and mapping components. Second, in order to reduce the inherent overfitting problem, we utilize a scale regularization term that appropriately encourages the Gaussian ellipsoids to approximate Gaussian spheres, to reduce anisotropy and achieve a good trade-off between tracking accuracy and rendering realism. At the same time, we observed that solely employing alpha-blending depth cannot + +impose effective constraints on the positions of Gaussian primitives. Thus, towards high-quality mapping, we further align median depth and alpha-blending depth to encourage Gaussian primitives to be well distributed over the scene surfaces, facilitating a consistent Gaussian field with more concentrated geometry density. Furthermore, in order to further improve the system's accuracy and efficiency, we design a novel depth uncertainty model to guide our Gaussian-based SLAM to focus on those stable and informative ones. We evaluate our system on a wide variety of RGB-D datasets, and the experimental results demonstrate that our CG-SLAM has superior performance in terms of tracking accuracy, reconstruction quality, and runtime efficiency. + +Overall, our contributions can be summarized as follows: + +- We present a new GPU-accelerated framework for real-time dense RGB-D SLAM based on a thorough theoretical analysis of camera pose derivatives in 3D Gaussian Splatting. +- We design multiple loss terms to build up a consistent and stable 3D Gaussian field suitable for tracking and mapping. +- We propose a novel depth uncertainty model, which assists our system in selecting more valuable Gaussian primitives during optimization, thereby improving tracking efficiency and accuracy. +- Experiments on various datasets demonstrate that our method can achieve competitive or better tracking and mapping results compared to baselines. + +# 2 Related Work + +# 2.1 Dense Visual SLAM + +Following the seminal contributions of DTAM [32] and KinectFusion [31] to dense visual SLAM systems, there has been significant progress in developing efficient scene representation models. TSDF [31, 32], surfels [22, 39, 52], voxel hashing [6, 33, 34], and octrees [42, 49] have been introduced to tackle the challenges of scalability within the SLAM task. Some more advanced technologies, including bundle adjustment [39], loop closure, and learning-based algorithms [25, 26, 46, 47, 55], were subsequently integrated into the SLAM framework to further improve system performance. These enhancements have significantly brought better accuracy and robustness for localization and reconstruction capabilities, thus pushing the frontiers of what is achievable in dense SLAM systems. Compared with traditional methods, our CG-SLAM can reconstruct a fully dense 3D map for rich applications. + +# 2.2 Neural Implict Radiance Field based SLAM + +Neural radiance fields [29] have shown promising potential in many 3D computer vision applications, such as novel view synthesis [3, 4, 14, 30], dynamic scene modeling [13, 15, 41, 48], and generalization [18, 40, 51, 53, 58]. Recent research works [27, 45] attempted to replace traditional maps, in tasks such as structure + +![](images/4f21ae97300a0cbe06ac5ba9adb44588eb90b1d2c5c04998bf84d17ba31c9562.jpg) +Fig. 2: System Overview. In a 3D Gaussian field constructed from an RGB-D sequence, we can render color, depth, opacity, and uncertainty maps through a GPU-accelerated rasterizer. Additionally, we attach a new uncertainty property to each Gaussian primitive to filter informative primitives. In the mapping process, we utilize multiple rendering results to design effective loss functions towards a consistent and stable Gaussian field. Subsequently, we employ appearance and geometry cues to perform accurate and efficient tracking. + +from motion (SFM) and SLAM, with the neural implicit field to jointly optimize scene representation and camera poses. Different kinds of neural fields brought insights into NeRF-SLAM works. NICE-SLAM [61] chose a fully covered voxel grid to store neural features, while Vox-Fusion [56] further improved this grid to an adaptive size. Besides, Point-SLAM [37] attached feature embeddings to the point cloud on object surfaces. The neural point-based method is more flexible and can encode more concentrated volume density. Co-SLAM [50] adopted a hybrid representation that includes coordinate encoding and hash grids to achieve smoother reconstruction and faster convergence. In addition to the aforementioned works, some methods [8, 20, 36] only used the neural field as a map and still performed tracking based on a traditional feature point-based visual odometry. Our CG-SLAM system can achieve better and more efficient performance in tracking, mapping, and rendering than NeRF-based methods. + +# 2.3 3D Gaussian Splitting Field + +3D Gaussian splatting is a revolutionary novel-view synthesis approach in 3D computer vision. This approach does not contain any neural network and allows photorealistic real-time(≥100 FPS) rendering at 1080p resolution. 3D Gaussian splatting has influenced a wide range of research topics, such as the 3D avatar [1, 19, 35, 59, 62], scene editing [7, 12, 57], image segment [60], and surface reconstruction [5, 11, 16]. These studies have demonstrated its generalization and effectiveness. Similar to the neural field, some concurrent manuscripts [21, 54] have been made to reversely infer camera poses within a built 3D Gaussian field. However, these works straightforwardly apply a raw 3D Gaussian field in the SLAM framework without specialized designs, such as anisotropy regularization and uninformative Gaussian primitive filtering. Additionally, they overlook + +the design on efficiency, which is the most important improvement that the Gaussian splatting technique should bring to a SLAM system. We believe that the advantages of the 3D Gaussian field in pose optimization have not been fully explored in these works, especially in terms of efficiency, and expect to further develop an advanced Gaussian-based visual SLAM system. + +# 3 Method + +The overview of our proposed rasterization-based Gaussian SLAM system is shown in Fig. 2. Given a set of RGB-D sequences, our system incrementally generates a stable, consistent, and uncertainty-aware Gaussian field, serving camera pose optimization and geometry reconstruction. In Sec. 3.1, we briefly introduce the 3D Gaussian splatting model and rasterization principles. We incorporated an uncertainty model that utilizes the geometry prior to attach the uncertainty property on rendered images and Gaussian primitives. This strategy helps remove outliers in mapping and makes full use of informative Gaussians in tracking (Sec. 3.2). Moreover, in Sec. 3.3, we detail the Gaussian primitive management strategy and some innovative loss terms that ensure geometry stability and accuracy. Finally, by minimizing the re-rendering loss from low-uncertainty primitives, we can build a real-time and accurate tracking module (Sec. 3.4). + +# 3.1 Preliminary + +Scene Representation. 3D Gaussian Splatting defines a 3D scene as a set of anisotropic Gaussian distributions, which are associated with means $\mathbf{X} \in \mathbb{R}^3$ and covariances $\boldsymbol{\Sigma} \in \mathbb{R}^{3 \times 3}$ . To ensure that the covariance matrix remains positive semi-definite throughout the gradient descent, it is endowed with a more intuitive and comprehensible physical meaning, that is, the configuration of an ellipsoid. Specifically, $\boldsymbol{\Sigma}$ is simplified and decomposed into: + +$$ +\boldsymbol {\Sigma} = \mathbf {R} \mathbf {S} \mathbf {S} ^ {T} \mathbf {R} ^ {T}, \tag {1} +$$ + +where scaling matrix $\mathbf{S} = \text{diag}([\mathbf{s}])$ is derived from the scale factor $\mathbf{s} \in \mathbb{R}^3$ and rotation matrix $\mathbf{R} \in \mathbb{R}^{3 \times 3}$ is derived from the quaternion $\mathbf{q} \in \mathbb{R}^4$ . Each Gaussian ellipsoid is also assigned an opacity $\sigma \in \mathbb{R}$ and spherical harmonics $(SH)$ coefficients, which respectively represent the volume density and view-dependent radiance within a nearby local region. + +Following Zwicker et al. [63], given a world-to-camera rotation $\mathbf{W}$ and the Jacobian $\mathbf{J}$ of the affine approximation of the projective transformation, the EWA splatting algorithm illustrates how to approximately project a 3D Gaussian ellipsoid onto the image plane to determine its effective range and per-pixel opacity values on this image plane. We can obtain the corresponding 2D Gaussian distribution $\mathcal{N}(\hat{\mu},\hat{\Sigma})$ as: + +$$ +\hat {\boldsymbol {\Sigma}} = \mathbf {J} \mathbf {W} \boldsymbol {\Sigma} \mathbf {W} ^ {T} \mathbf {J} ^ {T}, \tag {2} +$$ + +and $\hat{\mu}$ is the 2D pixel location of a 3D gaussian primitive center. + +Fast Rasterization-based Rendering. Fast Gaussian splatting rasterizer enables efficient pixel-by-pixel parallel rendering, and is fully differentiable, which provides a useful GPU-accelerated framework. For an incoming frame, the rasterizer can pre-sort all visible Gaussian primitives in order of depth from near to far. In terms of color rendering, the Gaussian splatting rasterizer adopts an $\alpha$ -blending solution, which accumulates radiance $c$ and opacity values $\sigma$ on a given pixel by traversing the above depth queue as follows: + +$$ +\hat {I} = \sum_ {i = 1} ^ {N} \alpha_ {i} T _ {i} c _ {i}, \tag {3} +$$ + +$$ +T _ {i} = \prod_ {k = 1} ^ {i - 1} \left(1 - \alpha_ {k}\right), \tag {4} +$$ + +$$ +\alpha_ {i} = \mathcal {N} \left(\hat {\mu} _ {\mathbf {i}}, \hat {\Sigma} _ {\mathbf {i}}\right) \sigma_ {i}, \tag {5} +$$ + +where $\hat{I}$ is the rendered color, $T_{i}$ is the accumulated transmittance, $\alpha_{i}$ is the opacity contributed to a pixel, and $c_{i}$ is the color of a Gaussian primitive computed from its SH coefficients. $N$ is the number of Gaussian primitives involved in the splatting process of a pixel. In terms of depth rendering, considering the loss term designed for geometry consistency, our rasterizer provides not only $\alpha$ -blending depth $\hat{D}_{\text{alpha}}$ but also the median depth $\hat{D}_{\text{median}}$ : + +$$ +\hat {D} _ {\text {a l p h a}} = \sum_ {i = 1} ^ {N} \alpha_ {i} T _ {i} d _ {i}, \tag {6} +$$ + +$$ +\hat {D} _ {\text {m e d i a n}} = d _ {\text {m e d i a n}}, \tag {7} +$$ + +where $d_{i}$ is the depth of a Gaussian primitive. For a pixel, in its splatting process, we regard a Gaussian at which the cumulative transmittance $T$ falls below 0.5 for the first time as the "median Gaussian". Its depth is recorded as $d_{\text{median}}$ . $T_{\text{median}}$ is the cumulative transmittance at this median Gaussian. + +$$ +\left(T _ {\text {m e d i a n}} \geq 0. 5\right) \quad \text {a n d} \quad \left(T _ {\text {m e d i a n} + 1} < 0. 5\right). \tag {8} +$$ + +Besides, the accumulated opacity value $\hat{O}$ is similarly required to distinguish unobserved areas for spawning Gaussians: + +$$ +\hat {O} = \sum_ {i = 1} ^ {N} \alpha_ {i} T _ {i}. \tag {9} +$$ + +# 3.2 Uncertainty Modeling + +Uncertainty model remains a trending topic in multi-view 3D reconstruction in recent decades. Inspired by [38], we believe that explicitly modeling uncertainty in our 3D Gaussian field has a positive effect on increasing the ratio of informative + +![](images/47de0d07bf1559dc85804f5ff7596f9311bca4715c2774f7aec5bc06aaf64477.jpg) +Fig. 3: Uncertainty of the Gaussian Primitives. Uncertainty of a Gaussian primitive is derived from its dominated pixels and corresponding depth biases, reflecting the geometric value and confidence of this primitive. + +Gaussians, which is crucial for the robustness and conciseness of a SLAM system. Hence, we propose an uncertainty model suitable for RGB-D observations from two perspectives: rendering images and Gaussian primitives. + +Uncertainty Map. $\alpha$ -blending depth is essentially an expected value calculated by sampling Gaussian ellipsoids along a pixel ray. Under the reasonable assumption of a normal distribution, we think that the uncertainty map is highly related to its variance. We can render an uncertainty value from the 3D Gaussian field as in Eq. (10). + +$$ +U = \sum_ {i = 1} ^ {N} \alpha_ {i} T _ {i} \left(d _ {i} - D\right) ^ {2}, \tag {10} +$$ + +where $D$ represents depth observations from the camera sensor. To mitigate drastic changes in positions of Gaussian primitives during optimization, we proposed a geometry variance loss term (Eq. (11)) based on the $H \times W$ uncertainty map to force them to be near the observed depth. + +$$ +\mathcal {L} _ {v a r} = \frac {1}{H W} \sum_ {n = 1} ^ {H W} | U _ {n} |. \tag {11} +$$ + +Uncertainty of Gaussian primitives. From the perspective of geometric consistency, we design a loss term $\mathcal{L}_{\text{align}}$ as in Eq. (12) to align the $\alpha$ -blending depth and median depth. + +$$ +\mathcal {L} _ {\text {a l i g n}} = \frac {1}{H W} \sum_ {n = 1} ^ {H W} \left| \hat {D} _ {\text {a l p h a}} ^ {n} - \hat {D} _ {\text {m e d i a n}} ^ {n} \right|. \tag {12} +$$ + +The $\alpha$ -blending depth of a pixel is dominated by the Gaussian primitive with the largest weight. We call this pixel a "dominated pixel" of this maximum weight Gaussian primitive. Our alignment loss forces $\hat{D}_{\text{alpha}}$ and $\hat{D}_{\text{median}}$ to be similar, which makes this maximum weight Gaussian primitive always occur at $d_{\text{median}}$ . Thus, as shown in Eq. (13), the uncertainty $v_i$ of the $i$ th Gaussian primitive + +is determined by the difference between its depth and depth observations from all its dominated pixels within a keyframe window $F = \{f_1, f_2, \dots, f_k\}$ . Fig. 3 further visualizes that a Gaussian primitive receives depth difference values from its dominated pixels in multiple viewpoints. + +$$ +\nu_ {i} = \frac {1}{M _ {1} + \dots + M _ {k}} \sum_ {f _ {k} \in F} \sum_ {p = 1} ^ {M _ {k}} \alpha_ {i} ^ {k, p} T _ {i} ^ {k, p} \left(D _ {p} ^ {k} - d _ {i} ^ {k}\right) ^ {2}. \tag {13} +$$ + +In a keyframe $f_{k}$ , $\alpha_{i}^{k,p}$ and $T_{i}^{k,p}$ represent the opacity and transmittance of the $i$ th Gaussian primitive on a pixel $p$ . $D_{p}^{k}$ represents the depth observation on a pixel $p$ in $f_{k}$ . $d_{i}^{k}$ is the depth value of the $i$ -th Gaussian primitive at $f_{k}$ . $\{M_{1},\dots,M_{k}\}$ are the number of dominated pixels of the $i$ th Gaussian primitive in different keyframes. Benefiting from uncertainty modeling, we can regularly detect and remove unreliable Gaussian primitives with high uncertainty exceeding a threshold $\tau = 0.025$ . Specifically, during the mapping optimization, primitives with $v_{i} > \tau$ will be manually reduced to a low-opacity level. These low-opacity Gaussian primitives can be optimized again to remove truly irreversible ones, which is a more adaptive and reasonable strategy. + +# 3.3 Mapping + +We employ various loss functions to update Gaussian properties, aiming for a consistent and stable Gaussian field. In addition to color and $SSIM$ loss in the original 3D Gaussian splatting, previous experience from NeRF-SLAM works suggests that geometry loss is a necessary part. To overcome anisotropic interference (Arrow-shaped Gaussian primitives), we add a soft scale regularization loss in the mapping process. Note that our system performs the initialization at a slightly higher cost, i.e., more optimization iters. + +$$ +\mathcal {L} _ {\text {c o l o r}} = \frac {1}{H W} \sum_ {n = 1} ^ {H W} \left| \hat {I} _ {n} - I _ {n} \right|, \tag {14} +$$ + +$$ +\mathcal {L} _ {s s i m} = S S I M (\hat {I}, I), \tag {15} +$$ + +$$ +\mathcal {L} _ {\text {g e o}} = \frac {1}{H W} \sum_ {n = 1} ^ {H W} \left| \hat {D} _ {\text {a l p h a}} ^ {n} - D _ {n} \right|, \tag {16} +$$ + +$$ +\mathcal {L} _ {i s o} = \frac {1}{G} \sum_ {i \in G} \max \left(\frac {\operatorname* {m a x} \left(\left\{s _ {i} ^ {x} , s _ {i} ^ {y} , s _ {i} ^ {z} \right\}\right)}{\operatorname* {m i n} \left(\left\{s _ {i} ^ {x} , s _ {i} ^ {y} , s _ {i} ^ {z} \right\}\right)}, \epsilon\right) - \epsilon , \tag {17} +$$ + +where $I_{n}$ and $D_{n}$ represent ground-truth color and depth, $\epsilon = 1.0$ is a hyperparameter that controls the level of anisotropy, and $G$ represents all visible Gaussians under the current view. + +$$ +\begin{array}{l} \mathcal {L} _ {\text {m a p p i n g}} = \omega_ {1} \mathcal {L} _ {\text {c o l o r}} + \omega_ {2} \mathcal {L} _ {\text {s s i m}} + \omega_ {3} \mathcal {L} _ {\text {g e o}} \tag {18} \\ + \omega_ {4} \mathcal {L} _ {\text {a l i g n}} + \omega_ {5} \mathcal {L} _ {\text {i s o}} + \omega_ {6} \mathcal {L} _ {\text {v a r}}. \\ \end{array} +$$ + +These loss functions customized for the SLAM task facilitate faster convergence in mapping and lay a solid foundation for subsequent tracking. + +Gaussian Management. In initialization, we densely project Gaussian primitives into 3D space based on depth observations of the first frame. In subsequent mapping, we set an empirical threshold $\psi = 0.5$ to extract unobserved or under-constructed pixels where $\hat{O} < \psi$ . Then, we utilize color and depth information on these pixels to spawn fresh Gaussian primitives. In addition, we inherited the original splitting, cloning, and removing strategy for Gaussian densification. + +# 3.4 Tracking + +In our system, we have proposed the first comprehensive mathematical theory on derivatives w.r.t. pose in 3D Gaussian splatting framework (refer to supplementary), and empirically discovered that the Lie algebraic representation is more advantageous for camera tracking, especially for the rotation, in a Gaussian field. The camera pose optimization, i.e., rotation and translation $\{\mathfrak{so}(3)|T\}$ , mainly includes two parts: sequential tracking and sliding bundle adjustment. + +Sequential Tracking. Given the fixed scene representation, the camera pose is initially guessed via the constant speed assumption where the last pose is transformed by the last relative transformation, and then we refine this rough pose using similar photometric and geometric losses weighted by $\lambda = \{\lambda_1,\lambda_2\}$ . + +$$ +\mathcal {L} _ {\text {t r a c k i n g}} = \lambda_ {1} \mathcal {L} _ {\text {c o l o r}} + \lambda_ {2} \mathcal {L} _ {\text {g e o}}, \tag {19} +$$ + +$$ +\{\mathfrak {s o} (3) | T \} = \underset {\{\mathfrak {s o} (3) | T \}} {\operatorname {a r g m i n}} \left(\mathcal {L} _ {\text {t r a c k i n g}}\right). \tag {20} +$$ + +Sliding Bundle Adjustment. Cumulative error is a typical problem in SLAM, also in Gaussian-based SLAM systems. To ease it, we set up a sliding window $F$ containing $k$ co-visible keyframes and jointly optimize camera extrinsic and scene representation in this window. Due to the efficiency, we encode keyframes into a descriptor pool with a pre-trained NetVLAD [2] model to determine co-visibility through the cosine similarity scores, instead of the view frustum overlap method in previous works. In addition to keyframes from NetVLAD [2], we also added the current frame and the most recent keyframes in the sliding window considering temporal associations. We still employ $\mathcal{L}_{\text{mapping}}$ in sliding bundle adjustment, where $\varPsi$ is a set of all estimizable Gaussian properties. + +$$ +\varPsi , \{\mathfrak {s o} (3) | T \} = \underset {\varPsi , \{\mathfrak {s o} (3) | T \}} {\operatorname {a r g m i n}} \left(\mathcal {L} _ {\text {m a p p i n g}}\right). \tag {21} +$$ + +# 4 Experiments + +In this section, we describe our experimental setup and validate that the proposed system can achieve improvement in both accuracy (Sec. 4.1 and Sec. 4.2) and efficiency (Sec. 4.3). We also confirmed the effectiveness of our design choices + +(Sec. 4.4). Additionally, we demonstrate our advantages in image rendering and capability for online third-person view rendering in supplementary. We color each cell as best, second best, and third best. + +Datasets. To evaluate our system in various scenarios, we use three standard benchmarks: Replica [43], TUM [44], and ScanNet [10]. The Replica dataset contains 8 available synthetic RGB-D sequences generated by Sucar et al. [45]. We examined the generalization of our method on real-world TUM [44] and ScanNet [10] datasets, which contain 5 and 6 challenging scenes respectively. + +Implementation Details. We run our system on a desktop equipped with an Intel i9-14900K and an NVIDIA RTX 4090 GPU. We set the learning rate of $\{\mathfrak{so}(3)|T\}$ to $\{0.0015, 0.00215\}$ in sequential tracking in all experiments. For the Replica [43] dataset, we perform 60-iteration mapping with weights $\omega = \{0.7, 0.1, 0.25, 0.25, 0.1, 0.15\}$ in a sliding window with $k = 4$ keyframes and 15-iteration sequential pose optimization weighted by $\lambda = \{0.2, 1.0\}$ . We select keyframes at an interval of 30. For TUM [44] and ScanNet [10] datasets, we use $\omega = \{1.0, 0.1, 0.8, 0.5, 0.1, 0.5\}$ , $k = 4$ , $\lambda = \{1.0, 0.6\}$ . Faced with challenging real-world scenes, we need to extract more keyframes at an interval of 15, perform mapping at $40 \sim 50$ iterations, and increase tracking iterations to 25. For further implementation details, please refer to our supplementary. + +Metrics. We quantitatively evaluate reconstruction quality using different 3D metrics. Given 3D triangle meshes, we compute mapping Accuracy [cm], Completion [cm], and Completion Ratio $[<5\mathrm{cm}\%]$ . Following NICE-SLAM [61], we discard unobserved regions that are not in any viewpoints. As for tracking performance, we measure ATE RMSE [44] for estimated trajectories. + +Baselines. We primarily consider state-of-the-art NeRF-SLAM works, including NICE-SLAM [61], Co-SLAM [50], Point-SLAM [37], and Vox-Fusion [56], as baselines. For a fair comparison, we reproduced all results from these baselines and reported their reconstruction performance with the same evaluation mechanism. We also add some concurrent manuscripts such as GS-SLAM [54] and SplaTAM [21] for reference, and we directly report the results in their papers. + +# 4.1 Localization Evaluation + +We report the localization accuracy of our system in 8 Replica [43] scenes in Tab. 1, where we surpass all other methods by a notable margin of around $25\% \sim 75\%$ . In our Gaussian-based system, image-level pose optimization and the well-designed Gaussian field promote fast and stable convergence to an optimal solution. This is the reason why we have lower variances and higher accuracy. As shown in Tab. 2, despite noisy and sparse depth information in the real-world TUM-RGBD dataset [44], our method still achieves better or competitive performance in 5 selected scenarios. We also benchmark our method and baselines on the similarly challenging ScanNet [10] to compare their tracking performance in Tab. 3. Sensor data from multiple large-scale ScanNet scenes suffers from severe motion blur and specular reflections. Our method further demonstrates its effectiveness and superiority on this dataset, excelling or maintaining com + +Table 1: Tracking Results on the Replica Dataset [43] (ATE RMSE [cm] $\downarrow$ ). Our system consistently achieved the best performance in this dataset, both for 8 individual scenes and for the average. It is worth noting that GS-SLAM [54] and SplaTAM [21] are concurrent with ours. + +
Methodrm-0rm-1rm-2off-0off-1off-2off-3off-4Avg.
NICE-SLAM0.971.311.070.881.001.061.101.131.06
Co-SLAM0.771.041.090.580.532.051.490.840.99
Point-SLAM0.560.470.300.350.620.550.720.730.54
Vox-Fusion0.400.540.540.500.460.750.500.600.54
GS-SLAM0.480.530.330.520.410.590.460.700.50
SplaTAM0.310.400.290.470.270.290.320.550.36
Ours0.290.270.250.330.140.280.310.290.27
Ours-light0.440.360.330.290.270.430.520.580.40
+ +Table 2: Tracking Results on the TUM-RGBD Dataset [44] (ATE RMSE [cm] ↓). Our system achieves better tracking accuracy and lower variance in different scenarios. "-" indicates unavailable results because the related work is not open source. + +
Methodfr1/deskfr1/desk2fr1/roomfr2.xyzfr3/officeAvg.
NICE-SLAM4.264.9934.4931.733.8715.87
Co-SLAM2.74.5730.161.92.68.38
Point-SLAM4.344.5430.921.313.488.92
Vox-Fusion3.526.0019.531.4926.0111.31
GS-SLAM3.3--1.36.6-
SplaTAM3.356.5411.131.245.165.48
Ours2.434.549.391.202.454.0
Ours-light3.144.7310.671.282.604.48
+ +Table 3: Tracking Results on the ScanNet Dataset [10] (ATE RMSE [cm] ↓). Our method achieves state-of-the-art tracking results in 6 scenes and exceeds other methods on average. "-" indicates failure results in Vox-Fusion [56]. + +
MethodSc.0000Sc.0059Sc.0106Sc.0169Sc.0181Sc.0207Avg.
NICE-SLAM12.0014.007.9010.9013.406.2010.70
Co-SLAM7.1812.2910.96.6213.437.139.37
Point-SLAM10.248.2911.8622.1614.779.5412.19
Vox-Fusion8.398.95-9.5012.206.43-
SplaTAM12.8310.1017.7212.0811.107.4611.88
Ours7.097.468.888.1611.605.348.08
Ours-light5.628.739.787.9312.025.458.17
+ +petitiveness in various scenes. Extensive experiments showcase the remarkable ability of our proposed system to track and handle complex situations. + +# 4.2 Reconstruction Evaluation + +In Tab. 4, we quantitatively measure the mapping performance of our proposed system, in comparison to NICE-SLAM [61], Co-SLAM [50], Point-SLAM [37], and Vox-Fusion [56]. It can be observed that our method outperforms all baselines on mapping accuracy. We use TSDF-Fusion [9] to construct our mesh map. We achieve state-of-the-art reconstruction in observed areas. It is worth noting that the Gaussian-based method neither has a global MLP nor a fully covered + +Table 4: Reconstruction Results on the Replica [43] Dataset. In terms of mapping accuracy, our method can outperform all existing methods. Due to the inherent limitation of 3D Gaussian representation, our method is slightly worse in completion. + +
MethodMetricrm-0rm-1rm-2off-0off-1off-2off-3off-4Avg.
NICE-SLAMAcc.[cm]↓3.533.603.035.563.354.713.843.353.87
Comp.[cm]↓3.403.623.274.554.033.943.994.153.87
Comp.Ratio<5cm%|↑86.0580.7587.2379.3482.1380.3580.5582.8882.41
Co-SLAMAcc.[cm]↓2.111.681.991.571.312.843.062.232.10
Comp.[cm]↓2.021.811.961.561.592.432.722.522.08
Comp.Ratio<5cm%|↑95.2695.1993.5896.0994.6591.6390.7290.4493.44
Point-SLAMAcc.[cm]↓1.451.141.191.050.861.311.571.511.26
Comp.[cm]↓3.463.022.651.652.213.623.473.903.00
Comp.Ratio<5cm%|↑88.4889.4490.1393.3990.5186.1786.0085.7488.73
Vox-FusionAcc.[cm]↓1.771.512.231.631.442.092.332.021.88
Comp.[cm]↓2.692.312.581.871.663.032.813.512.56
Comp.Ratio<5cm%|↑92.0392.4790.1393.8694.4088.9489.1086.5390.93
OursAcc.[cm]↓1.100.970.960.850.671.101.281.161.01
Comp.[cm]↓3.262.772.791.492.153.343.233.662.84
Comp.Ratio<5cm%|↑88.2689.4889.1093.6090.1486.0485.7885.6688.51
Ours-lightAcc.[cm]↓1.170.980.990.870.711.21.361.261.06
Comp.[cm]↓3.322.652.811.512.203.443.273.782.87
Comp.Ratio<5cm%|↑88.2089.3389.1093.4190.1485.6485.4785.3088.43
+ +feature grid, as in Co-SLAM [50]. Consequently, such a system exhibits a slightly weaker hole-filling ability compared to the NeRF-based method, which explains why our system is slightly worse in the Completion metric. As shown in Fig. 4, we qualitatively present the ground truth mesh and mesh reconstructions from ours and other baselines. Evidently, our system achieves more detailed geometry and less noisy outliers. + +# 4.3 Runtime and Memory Analysis + +We evaluate the runtime and memory footprint of our system compared to other works in Tab. 5. We reported the tracking and mapping efficiency in terms of per-iteration time consumption and the total number of optimization iterations. With the support of the GPU-accelerated rasterizer, our system can operate at around $8.5\mathrm{Hz}$ . Meanwhile, our carefully designed pipeline allows this system to expand to a lightweight version and to work with half-resolution images. For tracking, this lightweight version can work twice as fast as the original one, at the cost of a slight decrease in accuracy, as shown in Tabs. 1 to 3. For mapping, our lightweight version demonstrates similar superiority in Tab. 4. It can be clearly seen that both our versions achieve better performance than Co-SLAM [50]. Further efficiency analysis on TUM-RGBD [44] and ScanNet [10] is shown in supplementary. Also, our customized Gaussian field allows us to outperform the concurrent Gaussian-based works with less computational burden. + +However, as a non-MLP scene representation, the 3D Gaussian field inevitably requires much memory consumption to store different properties. This reason results in a considerable memory footprint in the Gaussian-based SLAM system, which is a common limitation in other Gaussian-based research topics. + +![](images/207df06401ea72020225d7e679de9303ab1b0908010c35b9b70d443fe8999c4b.jpg) +Fig. 4: Reconstruction Performance on Replica [43] Dataset. We qualitatively compared the mesh reconstruction results from CG-SLAM and other baselines, where CG-SLAM can produce more detailed geometry at a lower computation cost. + +Table 5: Runtime and Memory Usage. We comprehensively compared the runtime and memory usage on Replica [43] Office 0. Our proposed CG-SLAM can perform more efficient tracking and mapping than existing works, actually reaching a real-time level. "-" indicates unavailable results in related works. + +
MethodTracking [ms× it]↓Mapping [ms× it]↓Mapping IntervalSystem FPS↑Decoder Param↓Scene Embedding↓
Vox-Fusion23.61 × 3086.55 × 10101.250.98 MB0.162 MB
NICE-SLAM6.19 × 1091.59 × 6050.860.43 MB89.56 MB
Co-SLAM4.45 × 1011.9 × 10514.646.37 MB-
Point-SLAM6.14 × 4022.25 × 30050.630.54 MB28.11 MB
GS-SLAM11.9 × 1012.8 × 100-8.34--
SplaTAM41.7 × 4050.1 × 6010.21--
Ours6.02 × 1028.9 × 60308.47-231.66 MB
Ours-light3.80 × 1010.70 × 603016.83-56.50 MB
+ +# 4.4 Ablation Study + +To verify the rationality of our designs, we investigate the effectiveness of the anisotropy regularization, alignment and variance losses, and uncertainty model. + +Effectiveness of Isotropy Loss. Ablation results in Tab. 6 reveal how the anisotropy regularization term affects tracking metrics. We can notice that anisotropy $(\mathrm{w} / \mathrm{o}\mathcal{L}_{iso})$ will reduce the tracking accuracy by disrupting the view extrapolation of a 3D Gaussian field. Oppositely, excessive regularization (Isotropic Gaussians) will degrade rendering and subsequently affect the photometric loss, leading to poor tracking performance. Overall, a soft controllable constraint like $\mathcal{L}_{iso}$ is a reasonable choice. To more intuitively illustrate this phenomenon, we display opacity maps with and without anisotropy regularization in Fig. 5 (b), and apparent arrow-shaped artifacts occur in the latter. + +Effectiveness of Alignment and Variance Losses. Alignment and variance losses push primitives closer to object surfaces, facilitating novel view syn- + +![](images/ab1376c0784003867633c6e892e7229ea1d8aa03f6fa3cd7bffa2929676bf490.jpg) +Fig. 5: Uncertainty Model Ablation and Anisotropy Interference. (a) Uncertainty Model Ablation. This plot illustrates that the uncertainty model helps improve tracking accuracy while avoiding some extreme errors. (b) Anisotropy Interference. It can be clearly seen that in the case of w/o $\mathcal{L}_{iso}$ , serious arrow-shaped artifacts occur on the edges of the image. + +Table 6: Isotropy Loss Ablation Results(ATE RMSE [cm] ↓). The experimental results demonstrate the effectiveness of our anisotropy regularization term. "--" indicates a failure situation. + +
Settingrm-0rm-1rm-2off-0off-1off-2off-3off-4Avg.
w/ Liso0.290.270.250.330.140.280.310.290.27
w/o Liso0.320.310.540.360.240.310.72--
Isotropic Gaussians0.300.321.040.297.620.381.541.361.61
+ +thesis and SLAM. The effectiveness of these two losses is demonstrated in the supplementary. + +Effectiveness of Uncertainty Model. As shown in Fig. 5 (a), tracking error curves show that our uncertainty model can improve tracking accuracy while avoiding some extreme errors. For further quantitative ablation results, refer to the supplementary material. + +# 5 Conclusion + +We have proposed CG-SLAM, a dense RGB-D SLAM based on a consistent and uncertainty-aware 3D Gaussian field. Our targeted loss functions strengthen the 3D Gaussian field in terms of consistency and stability. The uncertainty model further distills highly informative primitives in this field to reduce interference from outliers. Besides, a customized GPU-accelerated rasterization pipeline enables our system to achieve state-of-the-art accuracy and efficiency in various scenes. Through extensive experiments, it can be concluded that our method outperforms previous works regarding tracking, mapping, and efficiency. + +Limitations. Considerable memory usage is one limitation of the Gaussian-based system. We expect that a more compact Gaussian field can be adopted in the SLAM task. In addition, the Gaussian-based method has a weak prediction ability for unobserved areas. Moreover, our system is not capable of handling dynamic objects in the environment. We believe it is a very interesting direction for future work. + +# Acknowledgements + +This work was partially supported by the NSFC (No. 62102356). + +# References + +1. Abdal, R., Yifan, W., Shi, Z., Xu, Y., Po, R., Kuang, Z., Chen, Q., Yeung, D.Y., Wetzstein, G.: Gaussian shell maps for efficient 3d human generation (2023) +2. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn architecture for weakly supervised place recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5297-5307 (2016) +3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields (2021) +4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR (2022) +5. Chen, H., Li, C., Lee, G.H.: Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance (2023) +6. Chen, J., Bautembach, D., Izadi, S.: Scalable real-time volumetric surface reconstruction. ACM Trans. Graph. 32(4), 113:1-113:16 (2013). https://doi.org/10.1145/2461912.2461940, https://doi.org/10.1145/2461912.2461940 +7. Chen, Y., Chen, Z., Zhang, C., Wang, F., Yang, X., Wang, Y., Cai, Z., Yang, L., Liu, H., Lin, G.: Gaussianeditor: Swift and controllable 3d editing with gaussian splatting (2023) +8. Chung, C., Tseng, Y., Hsu, Y., Shi, X.Q., Hua, Y., Yeh, J., Chen, W., Chen, Y., Hsu, W.H.: Orbeez-slam: A real-time monocular visual SLAM with ORB features and nerf-realized mapping. In: IEEE International Conference on Robotics and Automation, ICRA 2023, London, UK, May 29 - June 2, 2023. pp. 9400-9406. IEEE (2023). https://doi.org/10.1109/ICRA48891.2023.10160950, https://doi.org/10.1109/ICRA48891.2023.10160950 +9. Curless, B., Levoy, M.: A volumetric method for building complex models from range images. In: Proceedings of the 23rd annual conference on Computer graphics and interactive techniques. pp. 303-312 (1996) +0. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017) +1. Duisterhof, B.P., Mandi, Z., Yao, Y., Liu, J.W., Shou, M.Z., Song, S., Ichnowski, J.: Md-splatting: Learning metric deformation from 4d gaussians in highly deformable scenes (2023) +2. Fang, J., Wang, J., Zhang, X., Xie, L., Tian, Q.: Gaussianeditor: Editing 3d gaussians delicately with text instructions. arXiv preprint arXiv:2311.16037 (2023) +3. Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Niešner, M., Tian, Q.: Fast dynamic radiance fields with time-aware neural voxels. In: SIGGRAPH Asia 2022 Conference Papers (2022) +4. Fridovich-Keil and Yu, Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: CVPR (2022) +5. Geng, C., Peng, S., Xu, Z., Bao, H., Zhou, X.: Learning neural volumetric representations of dynamic humans in minutes. In: CVPR (2023) + +16. Guédon, A., Lepetit, V.: Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. arXiv preprint arXiv:2311.12775 (2023) +17. Hu, J., Mao, M., Bao, H., Zhang, G., Cui, Z.: Cp-slam: Collaborative neural point-based slam system. Advances in Neural Information Processing Systems 36 (2024) +18. Irshad, M.Z., Zakharov, S., Liu, K., Guizilini, V., Kollar, T., Gaidon, A., Kira, Z., Ambrus, R.: Neo 360: Neural fields for sparse view synthesis of outdoor scenes (2023), https://arxiv.org/abs/2308.12967 +19. Jena, R., Iyer, G.S., Choudhary, S., Smith, B., Chaudhari, P., Gee, J.: Splatarmor: Articulated gaussian splatting for animatable humans from monocular rgb videos. arXiv preprint arXiv:2311.10812 (2023) +20. Katragadda, S., Lee, W., Peng, Y., Geneva, P., Chen, C., Guo, C., Li, M., Huang, G.: Nerf-vins: A real-time neural radiance field map-based visual-inertial navigation system. CoRR abs/2309.09295 (2023). https://doi.org/10.48550/ARXIV.2309.09295, https://doi.org/10.48550/arXiv.2309.09295 +21. Keetha, N., Karhade, J., Jatavallabhula, K.M., Yang, G., Scherer, S., Ramanan, D., Luiten, J.: Splatam: Splat, track & map 3d gaussians for dense rgb-d slam. arXiv preprint arXiv:2312.02126 (2023) +22. Keller, M., Lefloch, D., Lambers, M., Izadi, S., Weyrich, T., Kolb, A.: Real-time 3d reconstruction in dynamic scenes using point-based fusion. In: 2013 International Conference on 3D Vision, 3DV 2013, Seattle, Washington, USA, June 29 - July 1, 2013. pp. 1-8. IEEE Computer Society (2013). https://doi.org/10.1109/3DV.2013.9, https://doi.org/10.1109/3DV.2013.9 +23. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023) +24. Kerl, C., Sturm, J., Cremers, D.: Dense visual slam for rgb-d cameras. In: 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems. pp. 2100-2106. IEEE (2013) +25. Koestler, L., Yang, N., Zeller, N., Cremers, D.: Tandem: Tracking and dense mapping in real-time using deep multi-view stereo. In: Conference on Robot Learning (CoRL) (2021) +26. Li, Y., Liu, X., Dong, W., Zhou, H., Bao, H., Zhang, G., Zhang, Y., Cui, Z.: Deltar: Depth estimation from a light-weight tof sensor and rgb image. In: European conference on computer vision. pp. 619-636. Springer (2022) +27. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: Barf: Bundle-adjusting neural radiance fields. In: IEEE International Conference on Computer Vision (ICCV) (2021) +28. Liu, X., Li, Y., Teng, Y., Bao, H., Zhang, G., Zhang, Y., Cui, Z.: Multi-modal neural radiance field for monocular dense slam with a light-weight tof sensor. In: Proceedings of the iee/cvf international conference on computer vision. pp. 1-11 (2023) +29. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: The European Conference on Computer Vision (ECCV) (2020) +30. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022). https://doi.org/10.1145/3528223.3530127, https://doi.org/10.1145/3528223.3530127 +31. Newcombe, R.A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohli, P., Shotton, J., Hodges, S., Fitzgibbon, A.W.: Kinectfusion: Real-time dense surface mapping and tracking. In: 10th IEEE International Symposium on Mixed + +and Augmented Reality, ISMAR 2011, Basel, Switzerland, October 26-29, 2011. pp. 127-136. IEEE Computer Society (2011). https://doi.org/10.1109/ISMAR.2011.6092378, https://doi.org/10.1109/ISMAR.2011.6092378 +32. Newcombe, R.A., Lovegrove, S., Davison, A.J.: DTAM: dense tracking and mapping in real-time. In: Metaxas, D.N., Quan, L., Sanfeliu, A., Gool, L.V. (eds.) IEEE International Conference on Computer Vision, ICCV 2011, Barcelona, Spain, November 6-13, 2011. pp. 2320-2327. IEEE Computer Society (2011). https://doi.org/10.1109/ICCV.2011.6126513, https://doi.org/10.1109/ICCV.2011.6126513 +33. Nießner, M., Zollhöfer, M., Izadi, S., Stamminger, M.: Real-time 3d reconstruction at scale using voxel hashing. ACM Trans. Graph. 32(6), 169:1-169:11 (2013). https://doi.org/10.1145/2508363.2508374, https://doi.org/10.1145/2508363.2508374 +34. Oleynikova, H., Taylor, Z., Fehr, M., Siegwart, R., Nieto, J.I.: Voxblox: Incremental 3d euclidean signed distance fields for on-board MAV planning. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2017, Vancouver, BC, Canada, September 24-28, 2017. pp. 1366-1373. IEEE (2017). https://doi.org/10.1109/IROS.2017.8202315, https://doi.org/10.1109/IROS.2017.8202315 +35. Pang, H., Zhu, H., Kortylewski, A., Theobalt, C., Habermann, M.: Ash: Animatable gaussian splats for efficient and photoreal human rendering (2023) +36. Rosinol, A., Leonard, J.J., Carlone, L.: Nerf-slam: Real-time dense monocular SLAM with neural radiance fields. In: IROS. pp. 3437-3444 (2023). https://doi.org/10.1109/IROS55552.2023.10341922, https://doi.org/10.1109/IROS55552.2023.10341922 +37. Sandström, E., Li, Y., Van Gool, L., Oswald, M.R.: Point-slam: Dense neural point cloud-based slam. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18433-18444 (2023) +38. Sandström, E., Ta, K., Van Gool, L., Oswald, M.R.: Uncle-slam: Uncertainty learning for dense neural slam. arXiv preprint arXiv:2306.11048 (2023) +39. Schops, T., Sattler, T., Pollefeys, M.: Bad slam: Bundle adjusted direct rgb-d slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019) +40. Schwarz, K., Liao, Y., Niemeyer, M., Geiger, A.: Graf: Generative radiance fields for 3d-aware image synthesis. In: Advances in Neural Information Processing Systems (NeurIPS) (2020) +41. Shao, R., Zheng, Z., Tu, H., Liu, B., Zhang, H., Liu, Y.: Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2023) +42. Steinbrücker, F., Kerl, C., Cremers, D.: Large-scale multi-resolution surface reconstruction from RGB-D sequences. In: IEEE International Conference on Computer Vision, ICCV 2013, Sydney, Australia, December 1-8, 2013. pp. 3264-3271. IEEE Computer Society (2013). https://doi.org/10.1109/ICCV.2013.405, https://doi.org/10.1109/ICCV.2013.405 +43. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., Clarkson, A., Yan, M., Budge, B., Yan, Y., Pan, X., Yon, J., Zou, Y., Leon, K., Carter, N., Briales, J., Gillingham, T., Mueggler, E., Pesqueira, L., Savva, M., Batra, D., Strasdat, H.M., Nardi, R.D., Goesele, M., Lovegrove, S., Newcombe, R.: The replica dataset: A digital replica of indoor spaces (2019) + +44. Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for the evaluation of rgb-d slam systems. In: 2012 IEEE/RSJ international conference on intelligent robots and systems. pp. 573-580. IEEE (2012) +45. Sucar, E., Liu, S., Ortiz, J., Davison, A.J.:imap: Implicit mapping and positioning in real-time. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6229-6238 (2021) +46. Tateno, K., Tombari, F., Laina, I., Navab, N.: CNN-SLAM: real-time dense monocular SLAM with learned depth prediction. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017. pp. 6565-6574. IEEE Computer Society (2017). https://doi.org/10.1109/CVPR.2017.695, https://doi.org/10.1109/CVPR.2017.695 +47. Teed, Z., Deng, J.: DROID-SLAM: deep visual SLAM for monocular, stereo, and RGB-D cameras. In: Ranzato, M., Beygelzimer, A., Dauphin, Y.N., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual. pp. 16558-16569 (2021), https://proceedings.neurips.cc/paper/2021/hash/89fcd07f20b6785b92134bd6c1d0fa42-AAbstract.html +48. Tretschk, E., Golyanik, V., Zollhöfer, M., Bozic, A., Lassner, C., Theobalt, C.: Scenerflow: Time-consistent reconstruction of general dynamic scenes. In: International Conference on 3D Vision (3DV) (2024) +49. Vespa, E., Nikolov, N., Grimm, M., Nardi, L., Kelly, P.H.J., Leutenegger, S.: Efficient octree-based volumetric SLAM supporting signed-distance and occupancy mapping. IEEE Robotics Autom. Lett. 3(2), 1144-1151 (2018). https://doi.org/10.1109/LRA.2018.2792537, https://doi.org/10.1109/LRA.2018.2792537 +50. Wang, H., Wang, J., Agapito, L.: Co-slam: Joint coordinate and sparse parametric encodings for neural real-time slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13293-13302 (2023) +51. Wang, T., Zhang, B., Zhang, T., Gu, S., Bao, J., Baltrusaitis, T., Shen, J., Chen, D., Wen, F., Chen, Q., et al.: Rodin: A generative model for sculpting 3d digital avatars using diffusion. arXiv preprint arXiv:2212.06135 (2022) +52. Whelan, T., Salas-Moreno, R.F., Glocker, B., Davison, A.J., Leutenegger, S.: Elastic fusion: Real-time dense SLAM and light source estimation. Int. J. Robotics Res. 35(14), 1697-1716 (2016). https://doi.org/10.1177/0278364916669237, https://doi.org/10.1177/0278364916669237 +53. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. arXiv preprint arXiv:2201.08845 (2022) +54. Yan, C., Qu, D., Xu, D., Zhao, B., Wang, Z., Wang, D., Li, X.: Gs-slam: Dense visual slam with 3d gaussian splatting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19595-19604 (2024) +55. Yang, N., von Stumberg, L., Wang, R., Cremers, D.: D3VO: deep depth, deep pose and deep uncertainty for monocular visual odometry. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020. pp. 1278-1289. Computer Vision Foundation / IEEE (2020). https://doi.org/10.1109/CVPR42600.2020.00136, https://openaccess.thecvf.com/content_CVPR_2020/html/Yang_D3VO_Dep_depth_Decel_Pose_and_Decel_Uncertainty_for_Monocular_CVPR_2020_paper.html +56. Yang, X., Li, H., Zhai, H., Ming, Y., Liu, Y., Zhang, G.: Vox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. In: 2022 IEEE International Symposium on Mixed and Augmented Reality (ISMAR). pp. 499-507. IEEE (2022) + +57. Ye, M., Danelljan, M., Yu, F., Ke, L.: Gaussian grouping: Segment and edit anything in 3d scenes. arXiv preprint arXiv:2312.00732 (2023) +58. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelNeRF: Neural radiance fields from one or few images. In: CVPR (2021) +59. Yuan, Y., Li, X., Huang, Y., De Mello, S., Nagano, K., Kautz, J., Iqbal, U.: Gavatar: Animatable 3d gaussian avatars with implicit mesh learning. arXiv preprint arXiv:2312.11461 (2023) +60. Zhou, S., Chang, H., Jiang, S., Fan, Z., Zhu, Z., Xu, D., Chari, P., You, S., Wang, Z., Kadambi, A.: Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. arXiv preprint arXiv:2312.03203 (2023) +61. Zhu, Z., Peng, S., Larsson, V., Xu, W., Bao, H., Cui, Z., Oswald, M.R., Pollefeys, M.: Nice-slam: Neural implicit scalable encoding for slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12786-12796 (2022) +62. Zielonka, W., Bagautdinov, T., Saito, S., Zollhöfer, M., Thies, J., Romero, J.: Drivable 3d gaussian avatars (2023) +63. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Ewa volume splatting. In: Proceedings Visualization, 2001. VIS'01. pp. 29-538. IEEE (2001) \ No newline at end of file diff --git a/2024/CG-SLAM_ Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field/images.zip b/2024/CG-SLAM_ Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..35169925c0353d0cda0f947e4d076ba1cd72ac01 --- /dev/null +++ b/2024/CG-SLAM_ Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0962cbbe4a8379131e4e88e765b4d43bab5468a9de488f64778b0ddd4f7316a8 +size 563014 diff --git a/2024/CG-SLAM_ Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field/layout.json b/2024/CG-SLAM_ Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e195425da776cc107f7f346acafa782d7528466d --- /dev/null +++ b/2024/CG-SLAM_ Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field/layout.json @@ -0,0 +1,10057 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 111, + 478, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 111, + 478, + 148 + ], + "spans": [ + { + "bbox": [ + 136, + 111, + 478, + 148 + ], + "type": "text", + "content": "CG-SLAM: Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "spans": [ + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "content": "Jiarui Hu" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "content": ", Xianhao Chen" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "content": ", Boyin Feng" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "content": ", Guanglin Li" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "content": ", Liangjing Yang" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "content": ", Hujun Bao" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "content": ", Guofeng Zhang" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "content": ", and Zhaopeng Cui" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 170, + 201, + 443, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 201, + 443, + 225 + ], + "spans": [ + { + "bbox": [ + 170, + 201, + 443, + 225 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 170, + 201, + 443, + 225 + ], + "type": "text", + "content": " State Key Lab of CAD&CG, Zhejiang University " + }, + { + "bbox": [ + 170, + 201, + 443, + 225 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 170, + 201, + 443, + 225 + ], + "type": "text", + "content": " ZJU-UIUC Institute, International Campus, Zhejiang University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 251, + 455, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 251, + 455, + 439 + ], + "spans": [ + { + "bbox": [ + 160, + 251, + 455, + 439 + ], + "type": "text", + "content": "Abstract. Recently neural radiance fields (NeRF) have been widely exploited as 3D representations for dense simultaneous localization and mapping (SLAM). Despite their notable successes in surface modeling and novel view synthesis, existing NeRF-based methods are hindered by their computationally intensive and time-consuming volume rendering pipeline. This paper presents an efficient dense RGB-D SLAM system, i.e., CG-SLAM, based on a novel uncertainty-aware 3D Gaussian field with high consistency and geometric stability. Through an in-depth analysis of Gaussian Splatting, we propose several techniques to construct a consistent and stable 3D Gaussian field suitable for tracking and mapping. Additionally, a novel depth uncertainty model is proposed to ensure the selection of valuable Gaussian primitives during optimization, thereby improving tracking efficiency and accuracy. Experiments on various datasets demonstrate that CG-SLAM achieves superior tracking and mapping performance with a notable tracking speed of around " + }, + { + "bbox": [ + 160, + 251, + 455, + 439 + ], + "type": "inline_equation", + "content": "15\\mathrm{Hz}" + }, + { + "bbox": [ + 160, + 251, + 455, + 439 + ], + "type": "text", + "content": ". We will make our source code publicly available. Project page: https://zju3dv.github.io/cg-slam." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 448, + 453, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 448, + 453, + 459 + ], + "spans": [ + { + "bbox": [ + 160, + 448, + 453, + 459 + ], + "type": "text", + "content": "Keywords: Dense Visual SLAM " + }, + { + "bbox": [ + 160, + 448, + 453, + 459 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 448, + 453, + 459 + ], + "type": "text", + "content": " Neural Rendering " + }, + { + "bbox": [ + 160, + 448, + 453, + 459 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 448, + 453, + 459 + ], + "type": "text", + "content": " 3D Gaussian field" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 479, + 230, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 230, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 230, + 491 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 503, + 482, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 482, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 482, + 611 + ], + "type": "text", + "content": "Dense visual Localization and Mapping (Visual SLAM) is a long-standing problem in 3D computer vision over recent decades, which targets performing pose tracking and scene mapping simultaneously with a variety of downstream applications such as virtual/augmented reality (VR/AR), robot navigation, and autonomous driving. Traditional visual SLAM systems [24] have shown accurate tracking performance across various scenes, while the underlying 3D representations (e.g., point cloud, mesh, and surfel) demonstrate limitations in facilitating highly free scene exploration, such as photorealistic scene touring, fine-grained map updating, etc." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 611, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 482, + 635 + ], + "type": "text", + "content": "Inspired by the Neural Radiance Field (NeRF) [29] in surface reconstruction and view rendering, some novel NeRF-based SLAM methods [17,28,37,61] have" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 642, + 405, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 642, + 405, + 653 + ], + "spans": [ + { + "bbox": [ + 145, + 642, + 405, + 653 + ], + "type": "text", + "content": "*Jiarui Hu and Xianhao Chen contributed equally to this work." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 147, + 653, + 248, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 653, + 248, + 665 + ], + "spans": [ + { + "bbox": [ + 147, + 653, + 248, + 665 + ], + "type": "text", + "content": "†Corresponding authors." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 187, + 186 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 187, + 186 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 187, + 186 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 187, + 186 + ], + "type": "image", + "image_path": "c9250fa538b056f7b1c254309eae5937950543481d07a2031d8b3afd8a5b5e9d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 195, + 482, + 262 + ], + "lines": [ + { + "bbox": [ + 130, + 195, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 195, + 482, + 262 + ], + "type": "text", + "content": "Fig. 1: CG-SLAM, which adopts a well-designed 3D Gaussian field, can simultaneously achieve state-of-the-art performance in localization, reconstruction and rendering. Benefiting from 3D Gaussian representation and a new GPU-accelerated framework that is developed from a thorough derivative analysis of camera pose in 3D Gaussian Splating [23], CG-SLAM can perform extremely fast rendering and solve the long-standing efficiency bottleneck suffered by previous rendering-based SLAM methods." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 194, + 115, + 314, + 186 + ], + "blocks": [ + { + "bbox": [ + 194, + 115, + 314, + 186 + ], + "lines": [ + { + "bbox": [ + 194, + 115, + 314, + 186 + ], + "spans": [ + { + "bbox": [ + 194, + 115, + 314, + 186 + ], + "type": "image", + "image_path": "0a37d550eece540360ccfb3cf5fa13631ecb7fbadd4e083d0ff23e1d5f692175.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 316, + 115, + 425, + 186 + ], + "blocks": [ + { + "bbox": [ + 316, + 115, + 425, + 186 + ], + "lines": [ + { + "bbox": [ + 316, + 115, + 425, + 186 + ], + "spans": [ + { + "bbox": [ + 316, + 115, + 425, + 186 + ], + "type": "image", + "image_path": "93e5e1fd8a503283efa7c6ed73fbd240ba80cb21da472b44bb22ce0fef11e8fc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 427, + 116, + 480, + 186 + ], + "blocks": [ + { + "bbox": [ + 427, + 116, + 480, + 186 + ], + "lines": [ + { + "bbox": [ + 427, + 116, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 427, + 116, + 480, + 186 + ], + "type": "image", + "image_path": "0bcb09f1f3f85d8e34eb29dbb90fa7931babfa73ad1f9b7562c289e028c4d8ba.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 290, + 482, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 386 + ], + "type": "text", + "content": "been proposed recently and demonstrated promising performance in tracking, surface modeling, and novel view synthesis. Nevertheless, existing NeRF-based methods follow the ray-tracing rendering pipeline that is computation-intensive and time-consuming. Therefore they can only perform tracking and mapping by sampling a limited number of camera rays, ignoring the natural structural information in images. To avoid local optima in tracking and artifacts, they normally require many optimization steps, which makes it struggle to bring the best of both worlds concerning accuracy and efficiency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 388, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 532 + ], + "type": "text", + "content": "Very recently the 3D Gaussian Splatting [23] method has been introduced for efficient novel view synthesis, and its rasterization-based rendering pipeline allows much faster image-level rendering, showing great potential in solving the inherent challenge of NeRF-based SLAM. However, it is nontrivial to reasonably incorporate the 3D Gaussian field in the SLAM setting. As a photorealistic view synthesis technique, the 3D Gaussian field is prone to overfitting the input images due to strong anisotropy and the lack of explicit multi-view constraints. As a result, on one hand, the 3D Gaussian splatting can not guarantee accurate modeling of 3D surfaces; on the other hand, since the Gaussians may not align with the environment's surfaces, this will lead to poor extrapolation capability and further degrade the camera tracking. Moreover, the increase of Gaussians in the mapping process will inevitably slow down the tracking efficiency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we introduce a real-time Gaussian splatting SLAM system, i.e., CG-SLAM, based on a novel uncertainty-aware 3D Gaussian field with high consistency and geometric stability. To this end, we first conduct a comprehensive mathematical analysis regarding the derivatives of camera poses in the EWA (Elliptical Weighted Average) splatting process [63], and develop a CUDA framework tailored for the SLAM task that effectively decouples the tracking and mapping components. Second, in order to reduce the inherent overfitting problem, we utilize a scale regularization term that appropriately encourages the Gaussian ellipsoids to approximate Gaussian spheres, to reduce anisotropy and achieve a good trade-off between tracking accuracy and rendering realism. At the same time, we observed that solely employing alpha-blending depth cannot" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 309, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 309, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 309, + 101 + ], + "type": "text", + "content": "Jiarui Hu and Xianhao Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 235 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 235 + ], + "type": "text", + "content": "impose effective constraints on the positions of Gaussian primitives. Thus, towards high-quality mapping, we further align median depth and alpha-blending depth to encourage Gaussian primitives to be well distributed over the scene surfaces, facilitating a consistent Gaussian field with more concentrated geometry density. Furthermore, in order to further improve the system's accuracy and efficiency, we design a novel depth uncertainty model to guide our Gaussian-based SLAM to focus on those stable and informative ones. We evaluate our system on a wide variety of RGB-D datasets, and the experimental results demonstrate that our CG-SLAM has superior performance in terms of tracking accuracy, reconstruction quality, and runtime efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 236, + 398, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 236, + 398, + 247 + ], + "spans": [ + { + "bbox": [ + 146, + 236, + 398, + 247 + ], + "type": "text", + "content": "Overall, our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 256, + 479, + 373 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 140, + 256, + 479, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 256, + 479, + 291 + ], + "spans": [ + { + "bbox": [ + 140, + 256, + 479, + 291 + ], + "type": "text", + "content": "- We present a new GPU-accelerated framework for real-time dense RGB-D SLAM based on a thorough theoretical analysis of camera pose derivatives in 3D Gaussian Splatting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 140, + 293, + 479, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 293, + 479, + 315 + ], + "spans": [ + { + "bbox": [ + 140, + 293, + 479, + 315 + ], + "type": "text", + "content": "- We design multiple loss terms to build up a consistent and stable 3D Gaussian field suitable for tracking and mapping." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 140, + 316, + 479, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 316, + 479, + 349 + ], + "spans": [ + { + "bbox": [ + 140, + 316, + 479, + 349 + ], + "type": "text", + "content": "- We propose a novel depth uncertainty model, which assists our system in selecting more valuable Gaussian primitives during optimization, thereby improving tracking efficiency and accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 140, + 352, + 479, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 352, + 479, + 373 + ], + "spans": [ + { + "bbox": [ + 140, + 352, + 479, + 373 + ], + "type": "text", + "content": "- Experiments on various datasets demonstrate that our method can achieve competitive or better tracking and mapping results compared to baselines." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 393, + 237, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 393, + 237, + 406 + ], + "spans": [ + { + "bbox": [ + 132, + 393, + 237, + 406 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 418, + 263, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 263, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 263, + 429 + ], + "type": "text", + "content": "2.1 Dense Visual SLAM" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 437, + 480, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 480, + 580 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 480, + 580 + ], + "type": "text", + "content": "Following the seminal contributions of DTAM [32] and KinectFusion [31] to dense visual SLAM systems, there has been significant progress in developing efficient scene representation models. TSDF [31, 32], surfels [22, 39, 52], voxel hashing [6, 33, 34], and octrees [42, 49] have been introduced to tackle the challenges of scalability within the SLAM task. Some more advanced technologies, including bundle adjustment [39], loop closure, and learning-based algorithms [25, 26, 46, 47, 55], were subsequently integrated into the SLAM framework to further improve system performance. These enhancements have significantly brought better accuracy and robustness for localization and reconstruction capabilities, thus pushing the frontiers of what is achievable in dense SLAM systems. Compared with traditional methods, our CG-SLAM can reconstruct a fully dense 3D map for rich applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 597, + 382, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 382, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 382, + 609 + ], + "type": "text", + "content": "2.2 Neural Implict Radiance Field based SLAM" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": "Neural radiance fields [29] have shown promising potential in many 3D computer vision applications, such as novel view synthesis [3, 4, 14, 30], dynamic scene modeling [13, 15, 41, 48], and generalization [18, 40, 51, 53, 58]. Recent research works [27, 45] attempted to replace traditional maps, in tasks such as structure" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "CG-SLAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 232 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 232 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 232 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 232 + ], + "type": "image", + "image_path": "4f21ae97300a0cbe06ac5ba9adb44588eb90b1d2c5c04998bf84d17ba31c9562.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 240, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 130, + 240, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 482, + 319 + ], + "type": "text", + "content": "Fig. 2: System Overview. In a 3D Gaussian field constructed from an RGB-D sequence, we can render color, depth, opacity, and uncertainty maps through a GPU-accelerated rasterizer. Additionally, we attach a new uncertainty property to each Gaussian primitive to filter informative primitives. In the mapping process, we utilize multiple rendering results to design effective loss functions towards a consistent and stable Gaussian field. Subsequently, we employ appearance and geometry cues to perform accurate and efficient tracking." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 341, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 498 + ], + "type": "text", + "content": "from motion (SFM) and SLAM, with the neural implicit field to jointly optimize scene representation and camera poses. Different kinds of neural fields brought insights into NeRF-SLAM works. NICE-SLAM [61] chose a fully covered voxel grid to store neural features, while Vox-Fusion [56] further improved this grid to an adaptive size. Besides, Point-SLAM [37] attached feature embeddings to the point cloud on object surfaces. The neural point-based method is more flexible and can encode more concentrated volume density. Co-SLAM [50] adopted a hybrid representation that includes coordinate encoding and hash grids to achieve smoother reconstruction and faster convergence. In addition to the aforementioned works, some methods [8, 20, 36] only used the neural field as a map and still performed tracking based on a traditional feature point-based visual odometry. Our CG-SLAM system can achieve better and more efficient performance in tracking, mapping, and rendering than NeRF-based methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 514, + 304, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 304, + 526 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 304, + 526 + ], + "type": "text", + "content": "2.3 3D Gaussian Splitting Field" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "3D Gaussian splatting is a revolutionary novel-view synthesis approach in 3D computer vision. This approach does not contain any neural network and allows photorealistic real-time(≥100 FPS) rendering at 1080p resolution. 3D Gaussian splatting has influenced a wide range of research topics, such as the 3D avatar [1, 19, 35, 59, 62], scene editing [7, 12, 57], image segment [60], and surface reconstruction [5, 11, 16]. These studies have demonstrated its generalization and effectiveness. Similar to the neural field, some concurrent manuscripts [21, 54] have been made to reversely infer camera poses within a built 3D Gaussian field. However, these works straightforwardly apply a raw 3D Gaussian field in the SLAM framework without specialized designs, such as anisotropy regularization and uninformative Gaussian primitive filtering. Additionally, they overlook" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 308, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 308, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 308, + 101 + ], + "type": "text", + "content": "Jiarui Hu and Xianhao Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "the design on efficiency, which is the most important improvement that the Gaussian splatting technique should bring to a SLAM system. We believe that the advantages of the 3D Gaussian field in pose optimization have not been fully explored in these works, especially in terms of efficiency, and expect to further develop an advanced Gaussian-based visual SLAM system." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 194, + 202, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 202, + 208 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 202, + 208 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 220, + 482, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 482, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 482, + 364 + ], + "type": "text", + "content": "The overview of our proposed rasterization-based Gaussian SLAM system is shown in Fig. 2. Given a set of RGB-D sequences, our system incrementally generates a stable, consistent, and uncertainty-aware Gaussian field, serving camera pose optimization and geometry reconstruction. In Sec. 3.1, we briefly introduce the 3D Gaussian splatting model and rasterization principles. We incorporated an uncertainty model that utilizes the geometry prior to attach the uncertainty property on rendered images and Gaussian primitives. This strategy helps remove outliers in mapping and makes full use of informative Gaussians in tracking (Sec. 3.2). Moreover, in Sec. 3.3, we detail the Gaussian primitive management strategy and some innovative loss terms that ensure geometry stability and accuracy. Finally, by minimizing the re-rendering loss from low-uncertainty primitives, we can build a real-time and accurate tracking module (Sec. 3.4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 380, + 221, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 221, + 393 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 221, + 393 + ], + "type": "text", + "content": "3.1 Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 400, + 482, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 473 + ], + "type": "text", + "content": "Scene Representation. 3D Gaussian Splatting defines a 3D scene as a set of anisotropic Gaussian distributions, which are associated with means " + }, + { + "bbox": [ + 130, + 400, + 482, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 400, + 482, + 473 + ], + "type": "text", + "content": " and covariances " + }, + { + "bbox": [ + 130, + 400, + 482, + 473 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 130, + 400, + 482, + 473 + ], + "type": "text", + "content": ". To ensure that the covariance matrix remains positive semi-definite throughout the gradient descent, it is endowed with a more intuitive and comprehensible physical meaning, that is, the configuration of an ellipsoid. Specifically, " + }, + { + "bbox": [ + 130, + 400, + 482, + 473 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}" + }, + { + "bbox": [ + 130, + 400, + 482, + 473 + ], + "type": "text", + "content": " is simplified and decomposed into:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 272, + 482, + 480, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 482, + 480, + 495 + ], + "spans": [ + { + "bbox": [ + 272, + 482, + 480, + 495 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {T} \\mathbf {R} ^ {T}, \\tag {1}", + "image_path": "04134b18c671ad25fd76cf0d96fe29902dc3fdf100ed702c3a27170d378a0e5e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 503, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 482, + 563 + ], + "type": "text", + "content": "where scaling matrix " + }, + { + "bbox": [ + 130, + 503, + 482, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{S} = \\text{diag}([\\mathbf{s}])" + }, + { + "bbox": [ + 130, + 503, + 482, + 563 + ], + "type": "text", + "content": " is derived from the scale factor " + }, + { + "bbox": [ + 130, + 503, + 482, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 503, + 482, + 563 + ], + "type": "text", + "content": " and rotation matrix " + }, + { + "bbox": [ + 130, + 503, + 482, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 130, + 503, + 482, + 563 + ], + "type": "text", + "content": " is derived from the quaternion " + }, + { + "bbox": [ + 130, + 503, + 482, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{q} \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 130, + 503, + 482, + 563 + ], + "type": "text", + "content": ". Each Gaussian ellipsoid is also assigned an opacity " + }, + { + "bbox": [ + 130, + 503, + 482, + 563 + ], + "type": "inline_equation", + "content": "\\sigma \\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 503, + 482, + 563 + ], + "type": "text", + "content": " and spherical harmonics " + }, + { + "bbox": [ + 130, + 503, + 482, + 563 + ], + "type": "inline_equation", + "content": "(SH)" + }, + { + "bbox": [ + 130, + 503, + 482, + 563 + ], + "type": "text", + "content": " coefficients, which respectively represent the volume density and view-dependent radiance within a nearby local region." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "text", + "content": "Following Zwicker et al. [63], given a world-to-camera rotation " + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "text", + "content": " and the Jacobian " + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{J}" + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "text", + "content": " of the affine approximation of the projective transformation, the EWA splatting algorithm illustrates how to approximately project a 3D Gaussian ellipsoid onto the image plane to determine its effective range and per-pixel opacity values on this image plane. We can obtain the corresponding 2D Gaussian distribution " + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\hat{\\mu},\\hat{\\Sigma})" + }, + { + "bbox": [ + 130, + 563, + 482, + 635 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 264, + 634, + 480, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 634, + 480, + 648 + ], + "spans": [ + { + "bbox": [ + 264, + 634, + 480, + 648 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {\\Sigma}} = \\mathbf {J} \\mathbf {W} \\boldsymbol {\\Sigma} \\mathbf {W} ^ {T} \\mathbf {J} ^ {T}, \\tag {2}", + "image_path": "3c2831f4839b9e1af13c8dc48a029d1314fcaf37683aa31bdd28c44846522733.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 653, + 416, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 653, + 416, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 653, + 416, + 666 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 131, + 653, + 416, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mu}" + }, + { + "bbox": [ + 131, + 653, + 416, + 666 + ], + "type": "text", + "content": " is the 2D pixel location of a 3D gaussian primitive center." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "CG-SLAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "Fast Rasterization-based Rendering. Fast Gaussian splatting rasterizer enables efficient pixel-by-pixel parallel rendering, and is fully differentiable, which provides a useful GPU-accelerated framework. For an incoming frame, the rasterizer can pre-sort all visible Gaussian primitives in order of depth from near to far. In terms of color rendering, the Gaussian splatting rasterizer adopts an " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "-blending solution, which accumulates radiance " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " and opacity values " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " on a given pixel by traversing the above depth queue as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 273, + 208, + 481, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 208, + 481, + 240 + ], + "spans": [ + { + "bbox": [ + 273, + 208, + 481, + 240 + ], + "type": "interline_equation", + "content": "\\hat {I} = \\sum_ {i = 1} ^ {N} \\alpha_ {i} T _ {i} c _ {i}, \\tag {3}", + "image_path": "0573b52fb17abe10b7a3ee6eff9f681127b24bb6d2381f7cce7838a7a0a7daf4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 266, + 251, + 481, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 251, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 266, + 251, + 481, + 283 + ], + "type": "interline_equation", + "content": "T _ {i} = \\prod_ {k = 1} ^ {i - 1} \\left(1 - \\alpha_ {k}\\right), \\tag {4}", + "image_path": "823e7ce3faaa606b0d605cf3fb706b01efe72b864db6e6f6038b9d1b8765eb5c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 266, + 290, + 481, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 290, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 266, + 290, + 481, + 304 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} = \\mathcal {N} \\left(\\hat {\\mu} _ {\\mathbf {i}}, \\hat {\\Sigma} _ {\\mathbf {i}}\\right) \\sigma_ {i}, \\tag {5}", + "image_path": "8953ef911c0bb412d0026f714f7aaa54b897636feae9f18b35d17698a6143e71.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 308, + 482, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 381 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 308, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 130, + 308, + 482, + 381 + ], + "type": "text", + "content": " is the rendered color, " + }, + { + "bbox": [ + 130, + 308, + 482, + 381 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 130, + 308, + 482, + 381 + ], + "type": "text", + "content": " is the accumulated transmittance, " + }, + { + "bbox": [ + 130, + 308, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 130, + 308, + 482, + 381 + ], + "type": "text", + "content": " is the opacity contributed to a pixel, and " + }, + { + "bbox": [ + 130, + 308, + 482, + 381 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 308, + 482, + 381 + ], + "type": "text", + "content": " is the color of a Gaussian primitive computed from its SH coefficients. " + }, + { + "bbox": [ + 130, + 308, + 482, + 381 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 308, + 482, + 381 + ], + "type": "text", + "content": " is the number of Gaussian primitives involved in the splatting process of a pixel. In terms of depth rendering, considering the loss term designed for geometry consistency, our rasterizer provides not only " + }, + { + "bbox": [ + 130, + 308, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 308, + 482, + 381 + ], + "type": "text", + "content": "-blending depth " + }, + { + "bbox": [ + 130, + 308, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\hat{D}_{\\text{alpha}}" + }, + { + "bbox": [ + 130, + 308, + 482, + 381 + ], + "type": "text", + "content": " but also the median depth " + }, + { + "bbox": [ + 130, + 308, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\hat{D}_{\\text{median}}" + }, + { + "bbox": [ + 130, + 308, + 482, + 381 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 261, + 391, + 481, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 391, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 261, + 391, + 481, + 422 + ], + "type": "interline_equation", + "content": "\\hat {D} _ {\\text {a l p h a}} = \\sum_ {i = 1} ^ {N} \\alpha_ {i} T _ {i} d _ {i}, \\tag {6}", + "image_path": "f9cfa6aff609df90c3d3fa7ac9c91ac59d17e2ba2f9f064a9505907de0480f2e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 263, + 434, + 480, + 448 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 434, + 480, + 448 + ], + "spans": [ + { + "bbox": [ + 263, + 434, + 480, + 448 + ], + "type": "interline_equation", + "content": "\\hat {D} _ {\\text {m e d i a n}} = d _ {\\text {m e d i a n}}, \\tag {7}", + "image_path": "ad5a9c2032250d900b749d4f37a88dcf594049b2daee2c7c40f7dacae54f5728.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 452, + 482, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 452, + 482, + 500 + ], + "spans": [ + { + "bbox": [ + 131, + 452, + 482, + 500 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 452, + 482, + 500 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 131, + 452, + 482, + 500 + ], + "type": "text", + "content": " is the depth of a Gaussian primitive. For a pixel, in its splatting process, we regard a Gaussian at which the cumulative transmittance " + }, + { + "bbox": [ + 131, + 452, + 482, + 500 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 452, + 482, + 500 + ], + "type": "text", + "content": " falls below 0.5 for the first time as the \"median Gaussian\". Its depth is recorded as " + }, + { + "bbox": [ + 131, + 452, + 482, + 500 + ], + "type": "inline_equation", + "content": "d_{\\text{median}}" + }, + { + "bbox": [ + 131, + 452, + 482, + 500 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 452, + 482, + 500 + ], + "type": "inline_equation", + "content": "T_{\\text{median}}" + }, + { + "bbox": [ + 131, + 452, + 482, + 500 + ], + "type": "text", + "content": " is the cumulative transmittance at this median Gaussian." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 217, + 510, + 480, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 510, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 217, + 510, + 480, + 522 + ], + "type": "interline_equation", + "content": "\\left(T _ {\\text {m e d i a n}} \\geq 0. 5\\right) \\quad \\text {a n d} \\quad \\left(T _ {\\text {m e d i a n} + 1} < 0. 5\\right). \\tag {8}", + "image_path": "1a87de899d67430072041e46438242fb031c0a66f45df57abbd1fc5c746ca3d1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 529, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 529, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 131, + 529, + 482, + 555 + ], + "type": "text", + "content": "Besides, the accumulated opacity value " + }, + { + "bbox": [ + 131, + 529, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\hat{O}" + }, + { + "bbox": [ + 131, + 529, + 482, + 555 + ], + "type": "text", + "content": " is similarly required to distinguish unobserved areas for spawning Gaussians:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 276, + 563, + 481, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 563, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 276, + 563, + 481, + 594 + ], + "type": "interline_equation", + "content": "\\hat {O} = \\sum_ {i = 1} ^ {N} \\alpha_ {i} T _ {i}. \\tag {9}", + "image_path": "ebbaa9daf88f5dc66c3db0487239e3ddbd35c01e392b93d8ba21662ad29bff01.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 610, + 273, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 610, + 273, + 623 + ], + "spans": [ + { + "bbox": [ + 131, + 610, + 273, + 623 + ], + "type": "text", + "content": "3.2 Uncertainty Modeling" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Uncertainty model remains a trending topic in multi-view 3D reconstruction in recent decades. Inspired by [38], we believe that explicitly modeling uncertainty in our 3D Gaussian field has a positive effect on increasing the ratio of informative" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 309, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 309, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 309, + 101 + ], + "type": "text", + "content": "Jiarui Hu and Xianhao Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 177, + 114, + 436, + 220 + ], + "blocks": [ + { + "bbox": [ + 177, + 114, + 436, + 220 + ], + "lines": [ + { + "bbox": [ + 177, + 114, + 436, + 220 + ], + "spans": [ + { + "bbox": [ + 177, + 114, + 436, + 220 + ], + "type": "image", + "image_path": "47de0d07bf1559dc85804f5ff7596f9311bca4715c2774f7aec5bc06aaf64477.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 229, + 479, + 262 + ], + "lines": [ + { + "bbox": [ + 131, + 229, + 479, + 262 + ], + "spans": [ + { + "bbox": [ + 131, + 229, + 479, + 262 + ], + "type": "text", + "content": "Fig. 3: Uncertainty of the Gaussian Primitives. Uncertainty of a Gaussian primitive is derived from its dominated pixels and corresponding depth biases, reflecting the geometric value and confidence of this primitive." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 287, + 479, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 287, + 479, + 323 + ], + "spans": [ + { + "bbox": [ + 131, + 287, + 479, + 323 + ], + "type": "text", + "content": "Gaussians, which is crucial for the robustness and conciseness of a SLAM system. Hence, we propose an uncertainty model suitable for RGB-D observations from two perspectives: rendering images and Gaussian primitives." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 323, + 479, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 323, + 479, + 382 + ], + "spans": [ + { + "bbox": [ + 131, + 323, + 479, + 382 + ], + "type": "text", + "content": "Uncertainty Map. " + }, + { + "bbox": [ + 131, + 323, + 479, + 382 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 323, + 479, + 382 + ], + "type": "text", + "content": "-blending depth is essentially an expected value calculated by sampling Gaussian ellipsoids along a pixel ray. Under the reasonable assumption of a normal distribution, we think that the uncertainty map is highly related to its variance. We can render an uncertainty value from the 3D Gaussian field as in Eq. (10)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 249, + 383, + 480, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 383, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 249, + 383, + 480, + 415 + ], + "type": "interline_equation", + "content": "U = \\sum_ {i = 1} ^ {N} \\alpha_ {i} T _ {i} \\left(d _ {i} - D\\right) ^ {2}, \\tag {10}", + "image_path": "db1814aa942b3123e8b6b97759a0e29d8852e7f505c51418ef531770c829bc09.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 420, + 479, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 420, + 479, + 468 + ], + "spans": [ + { + "bbox": [ + 131, + 420, + 479, + 468 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 420, + 479, + 468 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 131, + 420, + 479, + 468 + ], + "type": "text", + "content": " represents depth observations from the camera sensor. To mitigate drastic changes in positions of Gaussian primitives during optimization, we proposed a geometry variance loss term (Eq. (11)) based on the " + }, + { + "bbox": [ + 131, + 420, + 479, + 468 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 131, + 420, + 479, + 468 + ], + "type": "text", + "content": " uncertainty map to force them to be near the observed depth." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 255, + 477, + 480, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 477, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 255, + 477, + 480, + 510 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v a r} = \\frac {1}{H W} \\sum_ {n = 1} ^ {H W} | U _ {n} |. \\tag {11}", + "image_path": "0c027fba4554c26541659e314dead123e222f55b2989154ca9aa7758743a28c2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 519, + 479, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 519, + 479, + 555 + ], + "spans": [ + { + "bbox": [ + 131, + 519, + 479, + 555 + ], + "type": "text", + "content": "Uncertainty of Gaussian primitives. From the perspective of geometric consistency, we design a loss term " + }, + { + "bbox": [ + 131, + 519, + 479, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{align}}" + }, + { + "bbox": [ + 131, + 519, + 479, + 555 + ], + "type": "text", + "content": " as in Eq. (12) to align the " + }, + { + "bbox": [ + 131, + 519, + 479, + 555 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 519, + 479, + 555 + ], + "type": "text", + "content": "-blending depth and median depth." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 220, + 564, + 480, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 564, + 480, + 596 + ], + "spans": [ + { + "bbox": [ + 220, + 564, + 480, + 596 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l i g n}} = \\frac {1}{H W} \\sum_ {n = 1} ^ {H W} \\left| \\hat {D} _ {\\text {a l p h a}} ^ {n} - \\hat {D} _ {\\text {m e d i a n}} ^ {n} \\right|. \\tag {12}", + "image_path": "547ce628b954d0efab6c34c0b08d823cae624790c66137549badf302eda8100b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "type": "text", + "content": "-blending depth of a pixel is dominated by the Gaussian primitive with the largest weight. We call this pixel a \"dominated pixel\" of this maximum weight Gaussian primitive. Our alignment loss forces " + }, + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\hat{D}_{\\text{alpha}}" + }, + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\hat{D}_{\\text{median}}" + }, + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "type": "text", + "content": " to be similar, which makes this maximum weight Gaussian primitive always occur at " + }, + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "type": "inline_equation", + "content": "d_{\\text{median}}" + }, + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "type": "text", + "content": ". Thus, as shown in Eq. (13), the uncertainty " + }, + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 605, + 479, + 665 + ], + "type": "text", + "content": "th Gaussian primitive" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 90, + 447, + 100 + ], + "type": "text", + "content": "CG-SLAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": "is determined by the difference between its depth and depth observations from all its dominated pixels within a keyframe window " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "F = \\{f_1, f_2, \\dots, f_k\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": ". Fig. 3 further visualizes that a Gaussian primitive receives depth difference values from its dominated pixels in multiple viewpoints." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 195, + 171, + 481, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 171, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 195, + 171, + 481, + 205 + ], + "type": "interline_equation", + "content": "\\nu_ {i} = \\frac {1}{M _ {1} + \\dots + M _ {k}} \\sum_ {f _ {k} \\in F} \\sum_ {p = 1} ^ {M _ {k}} \\alpha_ {i} ^ {k, p} T _ {i} ^ {k, p} \\left(D _ {p} ^ {k} - d _ {i} ^ {k}\\right) ^ {2}. \\tag {13}", + "image_path": "00587ab9d3d4640bbc560539580eed1f1b81c104770a95c40b03c64c3128e775.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "spans": [ + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "text", + "content": "In a keyframe " + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "inline_equation", + "content": "f_{k}" + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "inline_equation", + "content": "\\alpha_{i}^{k,p}" + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "inline_equation", + "content": "T_{i}^{k,p}" + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "text", + "content": " represent the opacity and transmittance of the " + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "text", + "content": "th Gaussian primitive on a pixel " + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "inline_equation", + "content": "D_{p}^{k}" + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "text", + "content": " represents the depth observation on a pixel " + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "inline_equation", + "content": "f_{k}" + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "inline_equation", + "content": "d_{i}^{k}" + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "text", + "content": " is the depth value of the " + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "text", + "content": "-th Gaussian primitive at " + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "inline_equation", + "content": "f_{k}" + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "inline_equation", + "content": "\\{M_{1},\\dots,M_{k}\\}" + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "text", + "content": " are the number of dominated pixels of the " + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "text", + "content": "th Gaussian primitive in different keyframes. Benefiting from uncertainty modeling, we can regularly detect and remove unreliable Gaussian primitives with high uncertainty exceeding a threshold " + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "inline_equation", + "content": "\\tau = 0.025" + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "text", + "content": ". Specifically, during the mapping optimization, primitives with " + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "inline_equation", + "content": "v_{i} > \\tau" + }, + { + "bbox": [ + 131, + 213, + 483, + 335 + ], + "type": "text", + "content": " will be manually reduced to a low-opacity level. These low-opacity Gaussian primitives can be optimized again to remove truly irreversible ones, which is a more adaptive and reasonable strategy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 351, + 207, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 351, + 207, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 351, + 207, + 363 + ], + "type": "text", + "content": "3.3 Mapping" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 369, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 453 + ], + "type": "text", + "content": "We employ various loss functions to update Gaussian properties, aiming for a consistent and stable Gaussian field. In addition to color and " + }, + { + "bbox": [ + 130, + 369, + 482, + 453 + ], + "type": "inline_equation", + "content": "SSIM" + }, + { + "bbox": [ + 130, + 369, + 482, + 453 + ], + "type": "text", + "content": " loss in the original 3D Gaussian splatting, previous experience from NeRF-SLAM works suggests that geometry loss is a necessary part. To overcome anisotropic interference (Arrow-shaped Gaussian primitives), we add a soft scale regularization loss in the mapping process. Note that our system performs the initialization at a slightly higher cost, i.e., more optimization iters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 245, + 460, + 481, + 492 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 460, + 481, + 492 + ], + "spans": [ + { + "bbox": [ + 245, + 460, + 481, + 492 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o l o r}} = \\frac {1}{H W} \\sum_ {n = 1} ^ {H W} \\left| \\hat {I} _ {n} - I _ {n} \\right|, \\tag {14}", + "image_path": "2a41e396d46cc767a150a0b2b64f17b5dad91051428e7ac173108cb6e8a3f106.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 259, + 502, + 480, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 502, + 480, + 516 + ], + "spans": [ + { + "bbox": [ + 259, + 502, + 480, + 516 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s s i m} = S S I M (\\hat {I}, I), \\tag {15}", + "image_path": "8c3e296fe21cff936747bbfc026a8614280f5f3e694e17bab7a43431daef0078.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 237, + 529, + 480, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 529, + 480, + 561 + ], + "spans": [ + { + "bbox": [ + 237, + 529, + 480, + 561 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {g e o}} = \\frac {1}{H W} \\sum_ {n = 1} ^ {H W} \\left| \\hat {D} _ {\\text {a l p h a}} ^ {n} - D _ {n} \\right|, \\tag {16}", + "image_path": "54fcbe032050a19e66ca04de4c667cf670451c91233bc4cd835ffaf228c1f330.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 209, + 567, + 480, + 595 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 567, + 480, + 595 + ], + "spans": [ + { + "bbox": [ + 209, + 567, + 480, + 595 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i s o} = \\frac {1}{G} \\sum_ {i \\in G} \\max \\left(\\frac {\\operatorname* {m a x} \\left(\\left\\{s _ {i} ^ {x} , s _ {i} ^ {y} , s _ {i} ^ {z} \\right\\}\\right)}{\\operatorname* {m i n} \\left(\\left\\{s _ {i} ^ {x} , s _ {i} ^ {y} , s _ {i} ^ {z} \\right\\}\\right)}, \\epsilon\\right) - \\epsilon , \\tag {17}", + "image_path": "5c545165d8ecb8153083709e2725002c4e772c016bb4ffde5910f3b203318e67.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 598, + 482, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 598, + 482, + 634 + ], + "spans": [ + { + "bbox": [ + 130, + 598, + 482, + 634 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 598, + 482, + 634 + ], + "type": "inline_equation", + "content": "I_{n}" + }, + { + "bbox": [ + 130, + 598, + 482, + 634 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 598, + 482, + 634 + ], + "type": "inline_equation", + "content": "D_{n}" + }, + { + "bbox": [ + 130, + 598, + 482, + 634 + ], + "type": "text", + "content": " represent ground-truth color and depth, " + }, + { + "bbox": [ + 130, + 598, + 482, + 634 + ], + "type": "inline_equation", + "content": "\\epsilon = 1.0" + }, + { + "bbox": [ + 130, + 598, + 482, + 634 + ], + "type": "text", + "content": " is a hyperparameter that controls the level of anisotropy, and " + }, + { + "bbox": [ + 130, + 598, + 482, + 634 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 598, + 482, + 634 + ], + "type": "text", + "content": " represents all visible Gaussians under the current view." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 212, + 641, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 641, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 212, + 641, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {m a p p i n g}} = \\omega_ {1} \\mathcal {L} _ {\\text {c o l o r}} + \\omega_ {2} \\mathcal {L} _ {\\text {s s i m}} + \\omega_ {3} \\mathcal {L} _ {\\text {g e o}} \\tag {18} \\\\ + \\omega_ {4} \\mathcal {L} _ {\\text {a l i g n}} + \\omega_ {5} \\mathcal {L} _ {\\text {i s o}} + \\omega_ {6} \\mathcal {L} _ {\\text {v a r}}. \\\\ \\end{array}", + "image_path": "806410b6c3aff9563185e1b66ffbc3d4f3c4e7150198c9cb8477b18b1e1c0654.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 309, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 309, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 309, + 102 + ], + "type": "text", + "content": "Jiarui Hu and Xianhao Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "These loss functions customized for the SLAM task facilitate faster convergence in mapping and lay a solid foundation for subsequent tracking." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": "Gaussian Management. In initialization, we densely project Gaussian primitives into 3D space based on depth observations of the first frame. In subsequent mapping, we set an empirical threshold " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "\\psi = 0.5" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": " to extract unobserved or under-constructed pixels where " + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "inline_equation", + "content": "\\hat{O} < \\psi" + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": ". Then, we utilize color and depth information on these pixels to spawn fresh Gaussian primitives. In addition, we inherited the original splitting, cloning, and removing strategy for Gaussian densification." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 229, + 205, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 229, + 205, + 241 + ], + "spans": [ + { + "bbox": [ + 131, + 229, + 205, + 241 + ], + "type": "text", + "content": "3.4 Tracking" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 248, + 479, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 479, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 479, + 320 + ], + "type": "text", + "content": "In our system, we have proposed the first comprehensive mathematical theory on derivatives w.r.t. pose in 3D Gaussian splatting framework (refer to supplementary), and empirically discovered that the Lie algebraic representation is more advantageous for camera tracking, especially for the rotation, in a Gaussian field. The camera pose optimization, i.e., rotation and translation " + }, + { + "bbox": [ + 130, + 248, + 479, + 320 + ], + "type": "inline_equation", + "content": "\\{\\mathfrak{so}(3)|T\\}" + }, + { + "bbox": [ + 130, + 248, + 479, + 320 + ], + "type": "text", + "content": ", mainly includes two parts: sequential tracking and sliding bundle adjustment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 321, + 480, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 480, + 369 + ], + "type": "text", + "content": "Sequential Tracking. Given the fixed scene representation, the camera pose is initially guessed via the constant speed assumption where the last pose is transformed by the last relative transformation, and then we refine this rough pose using similar photometric and geometric losses weighted by " + }, + { + "bbox": [ + 130, + 321, + 480, + 369 + ], + "type": "inline_equation", + "content": "\\lambda = \\{\\lambda_1,\\lambda_2\\}" + }, + { + "bbox": [ + 130, + 321, + 480, + 369 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 239, + 378, + 480, + 390 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 378, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 239, + 378, + 480, + 390 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t r a c k i n g}} = \\lambda_ {1} \\mathcal {L} _ {\\text {c o l o r}} + \\lambda_ {2} \\mathcal {L} _ {\\text {g e o}}, \\tag {19}", + "image_path": "cc950d50a4d4aa64df64896515bca0e117884cfdffe862250610270817269edf.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 237, + 411, + 480, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 411, + 480, + 432 + ], + "spans": [ + { + "bbox": [ + 237, + 411, + 480, + 432 + ], + "type": "interline_equation", + "content": "\\{\\mathfrak {s o} (3) | T \\} = \\underset {\\{\\mathfrak {s o} (3) | T \\}} {\\operatorname {a r g m i n}} \\left(\\mathcal {L} _ {\\text {t r a c k i n g}}\\right). \\tag {20}", + "image_path": "78e15f709f09575ab3caa1bc4a4cb81f4cd6338ddfb1c4d3bcff820ea2a726d7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 436, + 481, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 481, + 556 + ], + "type": "text", + "content": "Sliding Bundle Adjustment. Cumulative error is a typical problem in SLAM, also in Gaussian-based SLAM systems. To ease it, we set up a sliding window " + }, + { + "bbox": [ + 130, + 436, + 481, + 556 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 436, + 481, + 556 + ], + "type": "text", + "content": " containing " + }, + { + "bbox": [ + 130, + 436, + 481, + 556 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 436, + 481, + 556 + ], + "type": "text", + "content": " co-visible keyframes and jointly optimize camera extrinsic and scene representation in this window. Due to the efficiency, we encode keyframes into a descriptor pool with a pre-trained NetVLAD [2] model to determine co-visibility through the cosine similarity scores, instead of the view frustum overlap method in previous works. In addition to keyframes from NetVLAD [2], we also added the current frame and the most recent keyframes in the sliding window considering temporal associations. We still employ " + }, + { + "bbox": [ + 130, + 436, + 481, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{mapping}}" + }, + { + "bbox": [ + 130, + 436, + 481, + 556 + ], + "type": "text", + "content": " in sliding bundle adjustment, where " + }, + { + "bbox": [ + 130, + 436, + 481, + 556 + ], + "type": "inline_equation", + "content": "\\varPsi" + }, + { + "bbox": [ + 130, + 436, + 481, + 556 + ], + "type": "text", + "content": " is a set of all estimizable Gaussian properties." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 226, + 568, + 480, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 568, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 226, + 568, + 480, + 588 + ], + "type": "interline_equation", + "content": "\\varPsi , \\{\\mathfrak {s o} (3) | T \\} = \\underset {\\varPsi , \\{\\mathfrak {s o} (3) | T \\}} {\\operatorname {a r g m i n}} \\left(\\mathcal {L} _ {\\text {m a p p i n g}}\\right). \\tag {21}", + "image_path": "297d7b0b1c5a0bce1d8a123b28af997729139762e42d756b0ab52f688c01ec59.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 605, + 230, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 230, + 618 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 230, + 618 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "In this section, we describe our experimental setup and validate that the proposed system can achieve improvement in both accuracy (Sec. 4.1 and Sec. 4.2) and efficiency (Sec. 4.3). We also confirmed the effectiveness of our design choices" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "CG-SLAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": "(Sec. 4.4). Additionally, we demonstrate our advantages in image rendering and capability for online third-person view rendering in supplementary. We color each cell as best, second best, and third best." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 213 + ], + "type": "text", + "content": "Datasets. To evaluate our system in various scenarios, we use three standard benchmarks: Replica [43], TUM [44], and ScanNet [10]. The Replica dataset contains 8 available synthetic RGB-D sequences generated by Sucar et al. [45]. We examined the generalization of our method on real-world TUM [44] and ScanNet [10] datasets, which contain 5 and 6 challenging scenes respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 213, + 481, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 481, + 344 + ], + "type": "text", + "content": "Implementation Details. We run our system on a desktop equipped with an Intel i9-14900K and an NVIDIA RTX 4090 GPU. We set the learning rate of " + }, + { + "bbox": [ + 130, + 213, + 481, + 344 + ], + "type": "inline_equation", + "content": "\\{\\mathfrak{so}(3)|T\\}" + }, + { + "bbox": [ + 130, + 213, + 481, + 344 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 213, + 481, + 344 + ], + "type": "inline_equation", + "content": "\\{0.0015, 0.00215\\}" + }, + { + "bbox": [ + 130, + 213, + 481, + 344 + ], + "type": "text", + "content": " in sequential tracking in all experiments. For the Replica [43] dataset, we perform 60-iteration mapping with weights " + }, + { + "bbox": [ + 130, + 213, + 481, + 344 + ], + "type": "inline_equation", + "content": "\\omega = \\{0.7, 0.1, 0.25, 0.25, 0.1, 0.15\\}" + }, + { + "bbox": [ + 130, + 213, + 481, + 344 + ], + "type": "text", + "content": " in a sliding window with " + }, + { + "bbox": [ + 130, + 213, + 481, + 344 + ], + "type": "inline_equation", + "content": "k = 4" + }, + { + "bbox": [ + 130, + 213, + 481, + 344 + ], + "type": "text", + "content": " keyframes and 15-iteration sequential pose optimization weighted by " + }, + { + "bbox": [ + 130, + 213, + 481, + 344 + ], + "type": "inline_equation", + "content": "\\lambda = \\{0.2, 1.0\\}" + }, + { + "bbox": [ + 130, + 213, + 481, + 344 + ], + "type": "text", + "content": ". We select keyframes at an interval of 30. For TUM [44] and ScanNet [10] datasets, we use " + }, + { + "bbox": [ + 130, + 213, + 481, + 344 + ], + "type": "inline_equation", + "content": "\\omega = \\{1.0, 0.1, 0.8, 0.5, 0.1, 0.5\\}" + }, + { + "bbox": [ + 130, + 213, + 481, + 344 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 213, + 481, + 344 + ], + "type": "inline_equation", + "content": "k = 4" + }, + { + "bbox": [ + 130, + 213, + 481, + 344 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 213, + 481, + 344 + ], + "type": "inline_equation", + "content": "\\lambda = \\{1.0, 0.6\\}" + }, + { + "bbox": [ + 130, + 213, + 481, + 344 + ], + "type": "text", + "content": ". Faced with challenging real-world scenes, we need to extract more keyframes at an interval of 15, perform mapping at " + }, + { + "bbox": [ + 130, + 213, + 481, + 344 + ], + "type": "inline_equation", + "content": "40 \\sim 50" + }, + { + "bbox": [ + 130, + 213, + 481, + 344 + ], + "type": "text", + "content": " iterations, and increase tracking iterations to 25. For further implementation details, please refer to our supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 346, + 480, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 480, + 405 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 480, + 405 + ], + "type": "text", + "content": "Metrics. We quantitatively evaluate reconstruction quality using different 3D metrics. Given 3D triangle meshes, we compute mapping Accuracy [cm], Completion [cm], and Completion Ratio " + }, + { + "bbox": [ + 130, + 346, + 480, + 405 + ], + "type": "inline_equation", + "content": "[<5\\mathrm{cm}\\%]" + }, + { + "bbox": [ + 130, + 346, + 480, + 405 + ], + "type": "text", + "content": ". Following NICE-SLAM [61], we discard unobserved regions that are not in any viewpoints. As for tracking performance, we measure ATE RMSE [44] for estimated trajectories." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 406, + 480, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 406, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 406, + 480, + 478 + ], + "type": "text", + "content": "Baselines. We primarily consider state-of-the-art NeRF-SLAM works, including NICE-SLAM [61], Co-SLAM [50], Point-SLAM [37], and Vox-Fusion [56], as baselines. For a fair comparison, we reproduced all results from these baselines and reported their reconstruction performance with the same evaluation mechanism. We also add some concurrent manuscripts such as GS-SLAM [54] and SplaTAM [21] for reference, and we directly report the results in their papers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 498, + 279, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 279, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 279, + 510 + ], + "type": "text", + "content": "4.1 Localization Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": "We report the localization accuracy of our system in 8 Replica [43] scenes in Tab. 1, where we surpass all other methods by a notable margin of around " + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "inline_equation", + "content": "25\\% \\sim 75\\%" + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": ". In our Gaussian-based system, image-level pose optimization and the well-designed Gaussian field promote fast and stable convergence to an optimal solution. This is the reason why we have lower variances and higher accuracy. As shown in Tab. 2, despite noisy and sparse depth information in the real-world TUM-RGBD dataset [44], our method still achieves better or competitive performance in 5 selected scenarios. We also benchmark our method and baselines on the similarly challenging ScanNet [10] to compare their tracking performance in Tab. 3. Sensor data from multiple large-scale ScanNet scenes suffers from severe motion blur and specular reflections. Our method further demonstrates its effectiveness and superiority on this dataset, excelling or maintaining com" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 307, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 307, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 307, + 101 + ], + "type": "text", + "content": "Jiarui Hu and Xianhao Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 189, + 167, + 424, + 256 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: Tracking Results on the Replica Dataset [43] (ATE RMSE [cm] " + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "). Our system consistently achieved the best performance in this dataset, both for 8 individual scenes and for the average. It is worth noting that GS-SLAM [54] and SplaTAM [21] are concurrent with ours." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 189, + 167, + 424, + 256 + ], + "lines": [ + { + "bbox": [ + 189, + 167, + 424, + 256 + ], + "spans": [ + { + "bbox": [ + 189, + 167, + 424, + 256 + ], + "type": "table", + "html": "
Methodrm-0rm-1rm-2off-0off-1off-2off-3off-4Avg.
NICE-SLAM0.971.311.070.881.001.061.101.131.06
Co-SLAM0.771.041.090.580.532.051.490.840.99
Point-SLAM0.560.470.300.350.620.550.720.730.54
Vox-Fusion0.400.540.540.500.460.750.500.600.54
GS-SLAM0.480.530.330.520.410.590.460.700.50
SplaTAM0.310.400.290.470.270.290.320.550.36
Ours0.290.270.250.330.140.280.310.290.27
Ours-light0.440.360.330.290.270.430.520.580.40
", + "image_path": "70ce5871dbf69a5a31486015285ee17b31f175c19c0f5d815eed8daf65657d58.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 189, + 308, + 424, + 394 + ], + "blocks": [ + { + "bbox": [ + 130, + 266, + 482, + 300 + ], + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 300 + ], + "type": "text", + "content": "Table 2: Tracking Results on the TUM-RGBD Dataset [44] (ATE RMSE [cm] ↓). Our system achieves better tracking accuracy and lower variance in different scenarios. \"-\" indicates unavailable results because the related work is not open source." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 189, + 308, + 424, + 394 + ], + "lines": [ + { + "bbox": [ + 189, + 308, + 424, + 394 + ], + "spans": [ + { + "bbox": [ + 189, + 308, + 424, + 394 + ], + "type": "table", + "html": "
Methodfr1/deskfr1/desk2fr1/roomfr2.xyzfr3/officeAvg.
NICE-SLAM4.264.9934.4931.733.8715.87
Co-SLAM2.74.5730.161.92.68.38
Point-SLAM4.344.5430.921.313.488.92
Vox-Fusion3.526.0019.531.4926.0111.31
GS-SLAM3.3--1.36.6-
SplaTAM3.356.5411.131.245.165.48
Ours2.434.549.391.202.454.0
Ours-light3.144.7310.671.282.604.48
", + "image_path": "99afe811309c84c92dd987bc93fb170820f4e72d726942e1753f67639bb8cd2d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 189, + 445, + 424, + 523 + ], + "blocks": [ + { + "bbox": [ + 130, + 403, + 482, + 437 + ], + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 437 + ], + "type": "text", + "content": "Table 3: Tracking Results on the ScanNet Dataset [10] (ATE RMSE [cm] ↓). Our method achieves state-of-the-art tracking results in 6 scenes and exceeds other methods on average. \"-\" indicates failure results in Vox-Fusion [56]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 189, + 445, + 424, + 523 + ], + "lines": [ + { + "bbox": [ + 189, + 445, + 424, + 523 + ], + "spans": [ + { + "bbox": [ + 189, + 445, + 424, + 523 + ], + "type": "table", + "html": "
MethodSc.0000Sc.0059Sc.0106Sc.0169Sc.0181Sc.0207Avg.
NICE-SLAM12.0014.007.9010.9013.406.2010.70
Co-SLAM7.1812.2910.96.6213.437.139.37
Point-SLAM10.248.2911.8622.1614.779.5412.19
Vox-Fusion8.398.95-9.5012.206.43-
SplaTAM12.8310.1017.7212.0811.107.4611.88
Ours7.097.468.888.1611.605.348.08
Ours-light5.628.739.787.9312.025.458.17
", + "image_path": "f3be7be962146be29eca923b53cc5bb81e455ce5a2765d550df5b9f4996ccf8c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 544, + 481, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 481, + 568 + ], + "type": "text", + "content": "petitiveness in various scenes. Extensive experiments showcase the remarkable ability of our proposed system to track and handle complex situations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 574, + 295, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 574, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 574, + 295, + 586 + ], + "type": "text", + "content": "4.2 Reconstruction Evaluation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "In Tab. 4, we quantitatively measure the mapping performance of our proposed system, in comparison to NICE-SLAM [61], Co-SLAM [50], Point-SLAM [37], and Vox-Fusion [56]. It can be observed that our method outperforms all baselines on mapping accuracy. We use TSDF-Fusion [9] to construct our mesh map. We achieve state-of-the-art reconstruction in observed areas. It is worth noting that the Gaussian-based method neither has a global MLP nor a fully covered" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "CG-SLAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 150, + 156, + 463, + 333 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 148 + ], + "type": "text", + "content": "Table 4: Reconstruction Results on the Replica [43] Dataset. In terms of mapping accuracy, our method can outperform all existing methods. Due to the inherent limitation of 3D Gaussian representation, our method is slightly worse in completion." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 150, + 156, + 463, + 333 + ], + "lines": [ + { + "bbox": [ + 150, + 156, + 463, + 333 + ], + "spans": [ + { + "bbox": [ + 150, + 156, + 463, + 333 + ], + "type": "table", + "html": "
MethodMetricrm-0rm-1rm-2off-0off-1off-2off-3off-4Avg.
NICE-SLAMAcc.[cm]↓3.533.603.035.563.354.713.843.353.87
Comp.[cm]↓3.403.623.274.554.033.943.994.153.87
Comp.Ratio<5cm%|↑86.0580.7587.2379.3482.1380.3580.5582.8882.41
Co-SLAMAcc.[cm]↓2.111.681.991.571.312.843.062.232.10
Comp.[cm]↓2.021.811.961.561.592.432.722.522.08
Comp.Ratio<5cm%|↑95.2695.1993.5896.0994.6591.6390.7290.4493.44
Point-SLAMAcc.[cm]↓1.451.141.191.050.861.311.571.511.26
Comp.[cm]↓3.463.022.651.652.213.623.473.903.00
Comp.Ratio<5cm%|↑88.4889.4490.1393.3990.5186.1786.0085.7488.73
Vox-FusionAcc.[cm]↓1.771.512.231.631.442.092.332.021.88
Comp.[cm]↓2.692.312.581.871.663.032.813.512.56
Comp.Ratio<5cm%|↑92.0392.4790.1393.8694.4088.9489.1086.5390.93
OursAcc.[cm]↓1.100.970.960.850.671.101.281.161.01
Comp.[cm]↓3.262.772.791.492.153.343.233.662.84
Comp.Ratio<5cm%|↑88.2689.4889.1093.6090.1486.0485.7885.6688.51
Ours-lightAcc.[cm]↓1.170.980.990.870.711.21.361.261.06
Comp.[cm]↓3.322.652.811.512.203.443.273.782.87
Comp.Ratio<5cm%|↑88.2089.3389.1093.4190.1485.6485.4785.3088.43
", + "image_path": "f996adce620ebd39aee9f00a97d2935c567cb541b0c3cfacb36b57a4a8d9cc93.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 354, + 482, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 427 + ], + "type": "text", + "content": "feature grid, as in Co-SLAM [50]. Consequently, such a system exhibits a slightly weaker hole-filling ability compared to the NeRF-based method, which explains why our system is slightly worse in the Completion metric. As shown in Fig. 4, we qualitatively present the ground truth mesh and mesh reconstructions from ours and other baselines. Evidently, our system achieves more detailed geometry and less noisy outliers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 443, + 318, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 443, + 318, + 455 + ], + "spans": [ + { + "bbox": [ + 132, + 443, + 318, + 455 + ], + "type": "text", + "content": "4.3 Runtime and Memory Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 481, + 616 + ], + "type": "text", + "content": "We evaluate the runtime and memory footprint of our system compared to other works in Tab. 5. We reported the tracking and mapping efficiency in terms of per-iteration time consumption and the total number of optimization iterations. With the support of the GPU-accelerated rasterizer, our system can operate at around " + }, + { + "bbox": [ + 130, + 462, + 481, + 616 + ], + "type": "inline_equation", + "content": "8.5\\mathrm{Hz}" + }, + { + "bbox": [ + 130, + 462, + 481, + 616 + ], + "type": "text", + "content": ". Meanwhile, our carefully designed pipeline allows this system to expand to a lightweight version and to work with half-resolution images. For tracking, this lightweight version can work twice as fast as the original one, at the cost of a slight decrease in accuracy, as shown in Tabs. 1 to 3. For mapping, our lightweight version demonstrates similar superiority in Tab. 4. It can be clearly seen that both our versions achieve better performance than Co-SLAM [50]. Further efficiency analysis on TUM-RGBD [44] and ScanNet [10] is shown in supplementary. Also, our customized Gaussian field allows us to outperform the concurrent Gaussian-based works with less computational burden." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "However, as a non-MLP scene representation, the 3D Gaussian field inevitably requires much memory consumption to store different properties. This reason results in a considerable memory footprint in the Gaussian-based SLAM system, which is a common limitation in other Gaussian-based research topics." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 307, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 307, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 307, + 101 + ], + "type": "text", + "content": "Jiarui Hu and Xianhao Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 482, + 246 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 482, + 246 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 482, + 246 + ], + "type": "image", + "image_path": "207df06401ea72020225d7e679de9303ab1b0908010c35b9b70d443fe8999c4b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 254, + 482, + 287 + ], + "lines": [ + { + "bbox": [ + 130, + 254, + 482, + 287 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 482, + 287 + ], + "type": "text", + "content": "Fig. 4: Reconstruction Performance on Replica [43] Dataset. We qualitatively compared the mesh reconstruction results from CG-SLAM and other baselines, where CG-SLAM can produce more detailed geometry at a lower computation cost." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 168, + 354, + 447, + 459 + ], + "blocks": [ + { + "bbox": [ + 130, + 300, + 482, + 344 + ], + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 344 + ], + "type": "text", + "content": "Table 5: Runtime and Memory Usage. We comprehensively compared the runtime and memory usage on Replica [43] Office 0. Our proposed CG-SLAM can perform more efficient tracking and mapping than existing works, actually reaching a real-time level. \"-\" indicates unavailable results in related works." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 168, + 354, + 447, + 459 + ], + "lines": [ + { + "bbox": [ + 168, + 354, + 447, + 459 + ], + "spans": [ + { + "bbox": [ + 168, + 354, + 447, + 459 + ], + "type": "table", + "html": "
MethodTracking [ms× it]↓Mapping [ms× it]↓Mapping IntervalSystem FPS↑Decoder Param↓Scene Embedding↓
Vox-Fusion23.61 × 3086.55 × 10101.250.98 MB0.162 MB
NICE-SLAM6.19 × 1091.59 × 6050.860.43 MB89.56 MB
Co-SLAM4.45 × 1011.9 × 10514.646.37 MB-
Point-SLAM6.14 × 4022.25 × 30050.630.54 MB28.11 MB
GS-SLAM11.9 × 1012.8 × 100-8.34--
SplaTAM41.7 × 4050.1 × 6010.21--
Ours6.02 × 1028.9 × 60308.47-231.66 MB
Ours-light3.80 × 1010.70 × 603016.83-56.50 MB
", + "image_path": "eac535597150c1a9d33d940981e9854522881cbca37257271332a88b9a337e02.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 483, + 238, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 238, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 238, + 495 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 508, + 481, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 481, + 532 + ], + "type": "text", + "content": "To verify the rationality of our designs, we investigate the effectiveness of the anisotropy regularization, alignment and variance losses, and uncertainty model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 487, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 487, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 487, + 640 + ], + "type": "text", + "content": "Effectiveness of Isotropy Loss. Ablation results in Tab. 6 reveal how the anisotropy regularization term affects tracking metrics. We can notice that anisotropy " + }, + { + "bbox": [ + 130, + 533, + 487, + 640 + ], + "type": "inline_equation", + "content": "(\\mathrm{w} / \\mathrm{o}\\mathcal{L}_{iso})" + }, + { + "bbox": [ + 130, + 533, + 487, + 640 + ], + "type": "text", + "content": " will reduce the tracking accuracy by disrupting the view extrapolation of a 3D Gaussian field. Oppositely, excessive regularization (Isotropic Gaussians) will degrade rendering and subsequently affect the photometric loss, leading to poor tracking performance. Overall, a soft controllable constraint like " + }, + { + "bbox": [ + 130, + 533, + 487, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{iso}" + }, + { + "bbox": [ + 130, + 533, + 487, + 640 + ], + "type": "text", + "content": " is a reasonable choice. To more intuitively illustrate this phenomenon, we display opacity maps with and without anisotropy regularization in Fig. 5 (b), and apparent arrow-shaped artifacts occur in the latter." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Effectiveness of Alignment and Variance Losses. Alignment and variance losses push primitives closer to object surfaces, facilitating novel view syn-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "CG-SLAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 159, + 114, + 455, + 194 + ], + "blocks": [ + { + "bbox": [ + 159, + 114, + 455, + 194 + ], + "lines": [ + { + "bbox": [ + 159, + 114, + 455, + 194 + ], + "spans": [ + { + "bbox": [ + 159, + 114, + 455, + 194 + ], + "type": "image", + "image_path": "ab1376c0784003867633c6e892e7229ea1d8aa03f6fa3cd7bffa2929676bf490.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 201, + 482, + 257 + ], + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 257 + ], + "type": "text", + "content": "Fig. 5: Uncertainty Model Ablation and Anisotropy Interference. (a) Uncertainty Model Ablation. This plot illustrates that the uncertainty model helps improve tracking accuracy while avoiding some extreme errors. (b) Anisotropy Interference. It can be clearly seen that in the case of w/o " + }, + { + "bbox": [ + 130, + 201, + 482, + 257 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{iso}" + }, + { + "bbox": [ + 130, + 201, + 482, + 257 + ], + "type": "text", + "content": ", serious arrow-shaped artifacts occur on the edges of the image." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 186, + 311, + 430, + 353 + ], + "blocks": [ + { + "bbox": [ + 130, + 270, + 482, + 301 + ], + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 301 + ], + "type": "text", + "content": "Table 6: Isotropy Loss Ablation Results(ATE RMSE [cm] ↓). The experimental results demonstrate the effectiveness of our anisotropy regularization term. \"--\" indicates a failure situation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 186, + 311, + 430, + 353 + ], + "lines": [ + { + "bbox": [ + 186, + 311, + 430, + 353 + ], + "spans": [ + { + "bbox": [ + 186, + 311, + 430, + 353 + ], + "type": "table", + "html": "
Settingrm-0rm-1rm-2off-0off-1off-2off-3off-4Avg.
w/ Liso0.290.270.250.330.140.280.310.290.27
w/o Liso0.320.310.540.360.240.310.72--
Isotropic Gaussians0.300.321.040.297.620.381.541.361.61
", + "image_path": "7f4940df4dc69c59a720e483b835e5fc333fb13f61bdb627dbfff43594734eb2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 375, + 480, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 480, + 399 + ], + "type": "text", + "content": "thesis and SLAM. The effectiveness of these two losses is demonstrated in the supplementary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 400, + 481, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 481, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 481, + 448 + ], + "type": "text", + "content": "Effectiveness of Uncertainty Model. As shown in Fig. 5 (a), tracking error curves show that our uncertainty model can improve tracking accuracy while avoiding some extreme errors. For further quantitative ablation results, refer to the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 469, + 220, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 469, + 220, + 481 + ], + "spans": [ + { + "bbox": [ + 132, + 469, + 220, + 481 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 497, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 594 + ], + "type": "text", + "content": "We have proposed CG-SLAM, a dense RGB-D SLAM based on a consistent and uncertainty-aware 3D Gaussian field. Our targeted loss functions strengthen the 3D Gaussian field in terms of consistency and stability. The uncertainty model further distills highly informative primitives in this field to reduce interference from outliers. Besides, a customized GPU-accelerated rasterization pipeline enables our system to achieve state-of-the-art accuracy and efficiency in various scenes. Through extensive experiments, it can be concluded that our method outperforms previous works regarding tracking, mapping, and efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Limitations. Considerable memory usage is one limitation of the Gaussian-based system. We expect that a more compact Gaussian field can be adopted in the SLAM task. In addition, the Gaussian-based method has a weak prediction ability for unobserved areas. Moreover, our system is not capable of handling dynamic objects in the environment. We believe it is a very interesting direction for future work." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 309, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 309, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 309, + 101 + ], + "type": "text", + "content": "Jiarui Hu and Xianhao Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 416, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 416, + 154 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 416, + 154 + ], + "type": "text", + "content": "This work was partially supported by the NSFC (No. 62102356)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 174, + 197, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 174, + 197, + 186 + ], + "spans": [ + { + "bbox": [ + 132, + 174, + 197, + 186 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 200, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 200, + 480, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 200, + 480, + 222 + ], + "spans": [ + { + "bbox": [ + 138, + 200, + 480, + 222 + ], + "type": "text", + "content": "1. Abdal, R., Yifan, W., Shi, Z., Xu, Y., Po, R., Kuang, Z., Chen, Q., Yeung, D.Y., Wetzstein, G.: Gaussian shell maps for efficient 3d human generation (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 224, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 480, + 255 + ], + "type": "text", + "content": "2. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn architecture for weakly supervised place recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5297-5307 (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 256, + 480, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 480, + 287 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 480, + 287 + ], + "type": "text", + "content": "3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 289, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 480, + 310 + ], + "type": "text", + "content": "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 312, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 480, + 332 + ], + "type": "text", + "content": "5. Chen, H., Li, C., Lee, G.H.: Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 334, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 334, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 334, + 480, + 365 + ], + "type": "text", + "content": "6. Chen, J., Bautembach, D., Izadi, S.: Scalable real-time volumetric surface reconstruction. ACM Trans. Graph. 32(4), 113:1-113:16 (2013). https://doi.org/10.1145/2461912.2461940, https://doi.org/10.1145/2461912.2461940" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 367, + 480, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 367, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 367, + 480, + 399 + ], + "type": "text", + "content": "7. Chen, Y., Chen, Z., Zhang, C., Wang, F., Yang, X., Wang, Y., Cai, Z., Yang, L., Liu, H., Lin, G.: Gaussianeditor: Swift and controllable 3d editing with gaussian splatting (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 400, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 400, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 400, + 480, + 464 + ], + "type": "text", + "content": "8. Chung, C., Tseng, Y., Hsu, Y., Shi, X.Q., Hua, Y., Yeh, J., Chen, W., Chen, Y., Hsu, W.H.: Orbeez-slam: A real-time monocular visual SLAM with ORB features and nerf-realized mapping. In: IEEE International Conference on Robotics and Automation, ICRA 2023, London, UK, May 29 - June 2, 2023. pp. 9400-9406. IEEE (2023). https://doi.org/10.1109/ICRA48891.2023.10160950, https://doi.org/10.1109/ICRA48891.2023.10160950" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 466, + 480, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 466, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 466, + 480, + 498 + ], + "type": "text", + "content": "9. Curless, B., Levoy, M.: A volumetric method for building complex models from range images. In: Proceedings of the 23rd annual conference on Computer graphics and interactive techniques. pp. 303-312 (1996)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 500, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 500, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 138, + 500, + 480, + 531 + ], + "type": "text", + "content": "0. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828-5839 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 533, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 480, + 564 + ], + "type": "text", + "content": "1. Duisterhof, B.P., Mandi, Z., Yao, Y., Liu, J.W., Shou, M.Z., Song, S., Ichnowski, J.: Md-splatting: Learning metric deformation from 4d gaussians in highly deformable scenes (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 566, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 566, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 566, + 480, + 587 + ], + "type": "text", + "content": "2. Fang, J., Wang, J., Zhang, X., Xie, L., Tian, Q.: Gaussianeditor: Editing 3d gaussians delicately with text instructions. arXiv preprint arXiv:2311.16037 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 588, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 480, + 620 + ], + "type": "text", + "content": "3. Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Niešner, M., Tian, Q.: Fast dynamic radiance fields with time-aware neural voxels. In: SIGGRAPH Asia 2022 Conference Papers (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 621, + 480, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 642 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 642 + ], + "type": "text", + "content": "4. Fridovich-Keil and Yu, Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: CVPR (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "5. Geng, C., Peng, S., Xu, Z., Bao, H., Zhou, X.: Learning neural volumetric representations of dynamic humans in minutes. In: CVPR (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "CG-SLAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "16. Guédon, A., Lepetit, V.: Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. arXiv preprint arXiv:2311.12775 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "text", + "content": "17. Hu, J., Mao, M., Bao, H., Zhang, G., Cui, Z.: Cp-slam: Collaborative neural point-based slam system. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 205 + ], + "type": "text", + "content": "18. Irshad, M.Z., Zakharov, S., Liu, K., Guizilini, V., Kollar, T., Gaidon, A., Kira, Z., Ambrus, R.: Neo 360: Neural fields for sparse view synthesis of outdoor scenes (2023), https://arxiv.org/abs/2308.12967" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 480, + 237 + ], + "type": "text", + "content": "19. Jena, R., Iyer, G.S., Choudhary, S., Smith, B., Chaudhari, P., Gee, J.: Splatarmor: Articulated gaussian splatting for animatable humans from monocular rgb videos. arXiv preprint arXiv:2311.10812 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 282 + ], + "type": "text", + "content": "20. Katragadda, S., Lee, W., Peng, Y., Geneva, P., Chen, C., Guo, C., Li, M., Huang, G.: Nerf-vins: A real-time neural radiance field map-based visual-inertial navigation system. CoRR abs/2309.09295 (2023). https://doi.org/10.48550/ARXIV.2309.09295, https://doi.org/10.48550/arXiv.2309.09295" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 315 + ], + "type": "text", + "content": "21. Keetha, N., Karhade, J., Jatavallabhula, K.M., Yang, G., Scherer, S., Ramanan, D., Luiten, J.: Splatam: Splat, track & map 3d gaussians for dense rgb-d slam. arXiv preprint arXiv:2312.02126 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 369 + ], + "type": "text", + "content": "22. Keller, M., Lefloch, D., Lambers, M., Izadi, S., Weyrich, T., Kolb, A.: Real-time 3d reconstruction in dynamic scenes using point-based fusion. In: 2013 International Conference on 3D Vision, 3DV 2013, Seattle, Washington, USA, June 29 - July 1, 2013. pp. 1-8. IEEE Computer Society (2013). https://doi.org/10.1109/3DV.2013.9, https://doi.org/10.1109/3DV.2013.9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 369, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 392 + ], + "type": "text", + "content": "23. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "text", + "content": "24. Kerl, C., Sturm, J., Cremers, D.: Dense visual slam for rgb-d cameras. In: 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems. pp. 2100-2106. IEEE (2013)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "25. Koestler, L., Yang, N., Zeller, N., Cremers, D.: Tandem: Tracking and dense mapping in real-time using deep multi-view stereo. In: Conference on Robot Learning (CoRL) (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 456, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 480, + 490 + ], + "type": "text", + "content": "26. Li, Y., Liu, X., Dong, W., Zhou, H., Bao, H., Zhang, G., Zhang, Y., Cui, Z.: Deltar: Depth estimation from a light-weight tof sensor and rgb image. In: European conference on computer vision. pp. 619-636. Springer (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 490, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 512 + ], + "type": "text", + "content": "27. Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: Barf: Bundle-adjusting neural radiance fields. In: IEEE International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 555 + ], + "type": "text", + "content": "28. Liu, X., Li, Y., Teng, Y., Bao, H., Zhang, G., Zhang, Y., Cui, Z.: Multi-modal neural radiance field for monocular dense slam with a light-weight tof sensor. In: Proceedings of the iee/cvf international conference on computer vision. pp. 1-11 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "text", + "content": "29. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: The European Conference on Computer Vision (ECCV) (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 588, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 632 + ], + "type": "text", + "content": "30. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022). https://doi.org/10.1145/3528223.3530127, https://doi.org/10.1145/3528223.3530127" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "31. Newcombe, R.A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohli, P., Shotton, J., Hodges, S., Fitzgibbon, A.W.: Kinectfusion: Real-time dense surface mapping and tracking. In: 10th IEEE International Symposium on Mixed" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 307, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 307, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 307, + 101 + ], + "type": "text", + "content": "Jiarui Hu and Xianhao Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 149, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 480, + 149 + ], + "type": "text", + "content": "and Augmented Reality, ISMAR 2011, Basel, Switzerland, October 26-29, 2011. pp. 127-136. IEEE Computer Society (2011). https://doi.org/10.1109/ISMAR.2011.6092378, https://doi.org/10.1109/ISMAR.2011.6092378" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 213 + ], + "type": "text", + "content": "32. Newcombe, R.A., Lovegrove, S., Davison, A.J.: DTAM: dense tracking and mapping in real-time. In: Metaxas, D.N., Quan, L., Sanfeliu, A., Gool, L.V. (eds.) IEEE International Conference on Computer Vision, ICCV 2011, Barcelona, Spain, November 6-13, 2011. pp. 2320-2327. IEEE Computer Society (2011). https://doi.org/10.1109/ICCV.2011.6126513, https://doi.org/10.1109/ICCV.2011.6126513" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 214, + 481, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 481, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 481, + 256 + ], + "type": "text", + "content": "33. Nießner, M., Zollhöfer, M., Izadi, S., Stamminger, M.: Real-time 3d reconstruction at scale using voxel hashing. ACM Trans. Graph. 32(6), 169:1-169:11 (2013). https://doi.org/10.1145/2508363.2508374, https://doi.org/10.1145/2508363.2508374" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 257, + 481, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 481, + 322 + ], + "type": "text", + "content": "34. Oleynikova, H., Taylor, Z., Fehr, M., Siegwart, R., Nieto, J.I.: Voxblox: Incremental 3d euclidean signed distance fields for on-board MAV planning. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2017, Vancouver, BC, Canada, September 24-28, 2017. pp. 1366-1373. IEEE (2017). https://doi.org/10.1109/IROS.2017.8202315, https://doi.org/10.1109/IROS.2017.8202315" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 322, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 481, + 343 + ], + "type": "text", + "content": "35. Pang, H., Zhu, H., Kortylewski, A., Theobalt, C., Habermann, M.: Ash: Animatable gaussian splats for efficient and photoreal human rendering (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 343, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 481, + 385 + ], + "type": "text", + "content": "36. Rosinol, A., Leonard, J.J., Carlone, L.: Nerf-slam: Real-time dense monocular SLAM with neural radiance fields. In: IROS. pp. 3437-3444 (2023). https://doi.org/10.1109/IROS55552.2023.10341922, https://doi.org/10.1109/IROS55552.2023.10341922" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "text", + "content": "37. Sandström, E., Li, Y., Van Gool, L., Oswald, M.R.: Point-slam: Dense neural point cloud-based slam. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18433-18444 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 418, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 439 + ], + "type": "text", + "content": "38. Sandström, E., Ta, K., Van Gool, L., Oswald, M.R.: Uncle-slam: Uncertainty learning for dense neural slam. arXiv preprint arXiv:2306.11048 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 439, + 481, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 471 + ], + "type": "text", + "content": "39. Schops, T., Sattler, T., Pollefeys, M.: Bad slam: Bundle adjusted direct rgb-d slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 471, + 481, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 481, + 502 + ], + "type": "text", + "content": "40. Schwarz, K., Liao, Y., Niemeyer, M., Geiger, A.: Graf: Generative radiance fields for 3d-aware image synthesis. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 502, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 481, + 546 + ], + "type": "text", + "content": "41. Shao, R., Zheng, Z., Tu, H., Liu, B., Zhang, H., Liu, Y.: Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 546, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 481, + 600 + ], + "type": "text", + "content": "42. Steinbrücker, F., Kerl, C., Cremers, D.: Large-scale multi-resolution surface reconstruction from RGB-D sequences. In: IEEE International Conference on Computer Vision, ICCV 2013, Sydney, Australia, December 1-8, 2013. pp. 3264-3271. IEEE Computer Society (2013). https://doi.org/10.1109/ICCV.2013.405, https://doi.org/10.1109/ICCV.2013.405" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 600, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 665 + ], + "type": "text", + "content": "43. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., Clarkson, A., Yan, M., Budge, B., Yan, Y., Pan, X., Yon, J., Zou, Y., Leon, K., Carter, N., Briales, J., Gillingham, T., Mueggler, E., Pesqueira, L., Savva, M., Batra, D., Strasdat, H.M., Nardi, R.D., Goesele, M., Lovegrove, S., Newcombe, R.: The replica dataset: A digital replica of indoor spaces (2019)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "CG-SLAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "44. Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for the evaluation of rgb-d slam systems. In: 2012 IEEE/RSJ international conference on intelligent robots and systems. pp. 573-580. IEEE (2012)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 181 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 181 + ], + "type": "text", + "content": "45. Sucar, E., Liu, S., Ortiz, J., Davison, A.J.:imap: Implicit mapping and positioning in real-time. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6229-6238 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 181, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 181, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 181, + 481, + 236 + ], + "type": "text", + "content": "46. Tateno, K., Tombari, F., Laina, I., Navab, N.: CNN-SLAM: real-time dense monocular SLAM with learned depth prediction. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017. pp. 6565-6574. IEEE Computer Society (2017). https://doi.org/10.1109/CVPR.2017.695, https://doi.org/10.1109/CVPR.2017.695" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 236, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 481, + 310 + ], + "type": "text", + "content": "47. Teed, Z., Deng, J.: DROID-SLAM: deep visual SLAM for monocular, stereo, and RGB-D cameras. In: Ranzato, M., Beygelzimer, A., Dauphin, Y.N., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual. pp. 16558-16569 (2021), https://proceedings.neurips.cc/paper/2021/hash/89fcd07f20b6785b92134bd6c1d0fa42-AAbstract.html" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 310, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 481, + 343 + ], + "type": "text", + "content": "48. Tretschk, E., Golyanik, V., Zollhöfer, M., Bozic, A., Lassner, C., Theobalt, C.: Scenerflow: Time-consistent reconstruction of general dynamic scenes. In: International Conference on 3D Vision (3DV) (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 343, + 481, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 481, + 387 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 481, + 387 + ], + "type": "text", + "content": "49. Vespa, E., Nikolov, N., Grimm, M., Nardi, L., Kelly, P.H.J., Leutenegger, S.: Efficient octree-based volumetric SLAM supporting signed-distance and occupancy mapping. IEEE Robotics Autom. Lett. 3(2), 1144-1151 (2018). https://doi.org/10.1109/LRA.2018.2792537, https://doi.org/10.1109/LRA.2018.2792537" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "text", + "content": "50. Wang, H., Wang, J., Agapito, L.: Co-slam: Joint coordinate and sparse parametric encodings for neural real-time slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13293-13302 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "type": "text", + "content": "51. Wang, T., Zhang, B., Zhang, T., Gu, S., Bao, J., Baltrusaitis, T., Shen, J., Chen, D., Wen, F., Chen, Q., et al.: Rodin: A generative model for sculpting 3d digital avatars using diffusion. arXiv preprint arXiv:2212.06135 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 450, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 481, + 493 + ], + "type": "text", + "content": "52. Whelan, T., Salas-Moreno, R.F., Glocker, B., Davison, A.J., Leutenegger, S.: Elastic fusion: Real-time dense SLAM and light source estimation. Int. J. Robotics Res. 35(14), 1697-1716 (2016). https://doi.org/10.1177/0278364916669237, https://doi.org/10.1177/0278364916669237" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 492, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 492, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 492, + 481, + 515 + ], + "type": "text", + "content": "53. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. arXiv preprint arXiv:2201.08845 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 515, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 481, + 546 + ], + "type": "text", + "content": "54. Yan, C., Qu, D., Xu, D., Zhao, B., Wang, Z., Wang, D., Li, X.: Gs-slam: Dense visual slam with 3d gaussian splatting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19595-19604 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 546, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 481, + 622 + ], + "type": "text", + "content": "55. Yang, N., von Stumberg, L., Wang, R., Cremers, D.: D3VO: deep depth, deep pose and deep uncertainty for monocular visual odometry. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020. pp. 1278-1289. Computer Vision Foundation / IEEE (2020). https://doi.org/10.1109/CVPR42600.2020.00136, https://openaccess.thecvf.com/content_CVPR_2020/html/Yang_D3VO_Dep_depth_Decel_Pose_and_Decel_Uncertainty_for_Monocular_CVPR_2020_paper.html" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "text", + "content": "56. Yang, X., Li, H., Zhai, H., Ming, Y., Liu, Y., Zhang, G.: Vox-fusion: Dense tracking and mapping with voxel-based neural implicit representation. In: 2022 IEEE International Symposium on Mixed and Augmented Reality (ISMAR). pp. 499-507. IEEE (2022)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 307, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 307, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 307, + 101 + ], + "type": "text", + "content": "Jiarui Hu and Xianhao Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 315 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "57. Ye, M., Danelljan, M., Yu, F., Ke, L.: Gaussian grouping: Segment and edit anything in 3d scenes. arXiv preprint arXiv:2312.00732 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 160 + ], + "type": "text", + "content": "58. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelNeRF: Neural radiance fields from one or few images. In: CVPR (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 482, + 194 + ], + "type": "text", + "content": "59. Yuan, Y., Li, X., Huang, Y., De Mello, S., Nagano, K., Kautz, J., Iqbal, U.: Gavatar: Animatable 3d gaussian avatars with implicit mesh learning. arXiv preprint arXiv:2312.11461 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "text", + "content": "60. Zhou, S., Chang, H., Jiang, S., Fan, Z., Zhu, Z., Xu, D., Chari, P., You, S., Wang, Z., Kadambi, A.: Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. arXiv preprint arXiv:2312.03203 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 226, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 270 + ], + "type": "text", + "content": "61. Zhu, Z., Peng, S., Larsson, V., Xu, W., Bao, H., Cui, Z., Oswald, M.R., Pollefeys, M.: Nice-slam: Neural implicit scalable encoding for slam. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12786-12796 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "type": "text", + "content": "62. Zielonka, W., Bagautdinov, T., Saito, S., Zollhöfer, M., Thies, J., Romero, J.: Drivable 3d gaussian avatars (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 293, + 482, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 315 + ], + "type": "text", + "content": "63. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Ewa volume splatting. In: Proceedings Visualization, 2001. VIS'01. pp. 29-538. IEEE (2001)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "CG-SLAM" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CIC-BART-SSA_ _ Controllable Image Captioning with Structured Semantic Augmentation/87b30ae0-4f07-48fc-b46b-fbcdc143e18d_content_list.json b/2024/CIC-BART-SSA_ _ Controllable Image Captioning with Structured Semantic Augmentation/87b30ae0-4f07-48fc-b46b-fbcdc143e18d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bc330c7122146624e8f693f23ea779021e89e951 --- /dev/null +++ b/2024/CIC-BART-SSA_ _ Controllable Image Captioning with Structured Semantic Augmentation/87b30ae0-4f07-48fc-b46b-fbcdc143e18d_content_list.json @@ -0,0 +1,1573 @@ +[ + { + "type": "text", + "text": "CIC-BART-SSA: Controllable Image Captioning with Structured Semantic Augmentation", + "text_level": 1, + "bbox": [ + 223, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kalliopi Basioti $^{1,2*}$ , Mohamed A. Abdelsalam $^{2}$ , Federico Fancellu $^{3\\dagger}$ , Vladimir Pavlovic $^{1\\dagger}$ , and Afsaneh Fazly $^{2}$", + "bbox": [ + 220, + 210, + 781, + 244 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1 Rutgers University, New Jersey, USA {kalliopi.basioti, vladimir}@rutgers.edu", + "$^{2}$ Samsung AI Centre - Toronto, Toronto, Canada {m.abdelsalam, a.fazly}@samsung.com", + "3 Solventum ffancellu@solventum.com" + ], + "bbox": [ + 276, + 253, + 723, + 323 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d0c67294711deb8149e35a775c37ebb01dd7267933b85d9ec5c0d8c478830803.jpg", + "image_caption": [ + "Original Captions A couple of elephants standing next to trees. A couple of elephants standing by themselves. A couple elephants walking by a tree after sunset. Two elephants are standing in the grass near a tree. Two elephants are standing by the trees in the wild." + ], + "image_footnote": [], + "bbox": [ + 217, + 343, + 359, + 415 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Samples Generated from CIC-BART-SSA", + "text_level": 1, + "bbox": [ + 383, + 343, + 527, + 351 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. There is grass near a tree.", + "2. There is a tree near the grass.", + "3. The hillside is covered in", + "4. A field next to a big tree is open.", + "5. The grass under the trees is lush", + "6. A large-elephant standing in a field near a tree.", + "7. Two elephants walking under a tree in the sunset", + "8. Two elephants standing in the grass next to a tree", + "9. Two elephants standing next to a tree and a blue sky.", + "10. A big elephant standing next to a tree in a field of grass", + "11. An elephant standing next to a tree with a blue sky behind it.", + "12. Two elephants standing in the water near a and a blue sky", + "12. Two elephants standing in the grass near a tree and a blue sky 13. Two elephants standing next to each other on a lush green hill", + "13. Two elephants standing next to each other on a rush green hillside next to a forest.", + "14. Two elephants standing next to a tree in a field with a bluesky and trees behind them.", + "15. Two elephants standing next to each other on a grass covered field next to a lush green forest." + ], + "bbox": [ + 383, + 352, + 689, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/43248957ba4e19f716628cd282245a6048b2429162cb83424e188afead66d5b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 683, + 343, + 782, + 400 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/acb148dad3c5cf0c7a8f9daeaca9b14bb0a2e178027ab9a42c85f2e5549a92f1.jpg", + "image_caption": [ + "Fig. 1: Existing captioning datasets contain captions that describe the entirety of an image. This is reflected in the narrow distributions of the entities that appear in those captions and the caption lengths (the red-colored histograms). CIC aims to generate diverse descriptions by controllably re-focusing on different spatiosemantic aspects of an image, such as the semantically coherent subsets of image objects. Our proposed CIC-BART-SSA is designed to produce diverse, controlled captions ranging from brief and concise to detailed and comprehensive. Sentences 1-15 are example outputs of our approach where the highlighted text indicates the focus of a controllable caption. The histograms demonstrate that our approach generates high-quality descriptions for a wider range of scene focus (number of visual entities) and caption length compared to the original captions. Image is licensed under CC BY-SA 2.0." + ], + "image_footnote": [], + "bbox": [ + 684, + 404, + 779, + 457 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Controllable Image Captioning (CIC) aims at generating natural language descriptions for an image, conditioned on information provided by end users, e.g., regions, entities or events of interest. However, available image-language datasets mainly contain captions that describe the entirety of an image, making them ineffective for training CIC models that can potentially attend to any subset of regions or relationships. To tackle this challenge, we propose a novel, fully automatic method to sample additional focused and visually grounded captions using a unified structured semantic representation built on top of the existing set of captions associated with an image. We leverage Abstract Meaning Representation (AMR), a cross-lingual graph-based semantic formal", + "bbox": [ + 259, + 648, + 743, + 803 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during an internship at Samsung AI Centre - Toronto", + "bbox": [ + 230, + 811, + 676, + 825 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Work done while at Samsung AI Centre - Toronto", + "bbox": [ + 230, + 825, + 575, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ism, to encode all possible spatio-semantic relations between entities, beyond the typical spatial-relations-only focus of current methods. We use this Structured Semantic Augmentation (SSA) framework to augment existing image-caption datasets with the grounded controlled captions, increasing their spatial and semantic diversity and focal coverage. We then develop a new model, CIC-BART-SSA, specifically tailored for the CIC task, that sources its control signals from SSA-diversified datasets. We empirically show that, compared to SOTA CIC models, CIC-BART-SSA generates captions that are superior in diversity and text quality, are competitive in controllability, and, importantly, minimize the gap between broad and highly focused controlled captioning performance by efficiently generalizing to the challenging highly focused scenarios. Code is available at https://github.com/SamsungLabs/CIC-BART-SSA.", + "bbox": [ + 259, + 147, + 743, + 329 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 354, + 375, + 371 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image captioning refers to the task of providing an AI system with an input image, and asking the system to describe the visual content in natural language. This process requires the captioning system to understand what objects are present, in what context (e.g., event or scene), and how they relate. Recent deep learning approaches to this task [14,26,29,31,37,39,40,49,50,57] surpass human performance in standard image captioning metrics. However, these models tend to generate general captions that describe the entirety of an image, and are often of limited diversity; see Original Captions in Fig. 1.", + "bbox": [ + 212, + 386, + 787, + 507 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Controllable image captioning (CIC) overcomes these challenges by generating different descriptions for the same image in a user-controlled fashion. That is, a CIC model receives as input an image paired with a user-specified control signal (e.g., entities or regions of interest), and generates a caption conditioned on the control signal. CIC models are thus capable of generating a diverse set of captions by varying the control signal for the same image; see CIC generated captions 1-15 in Fig. 1.", + "bbox": [ + 212, + 507, + 787, + 612 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In realistic applications, the easiest way for the user to control the generation of captions is to limit the focus of the desired captions by selecting different entities (objects) using their bounding boxes, as shown in Figs. 1 and 2. Most previous work focuses on such spatial control signals [18, 23, 28, 47, 55, 56]. To improve performance, more recent studies supplement this spatial signal with additional information on the desired length, style, or syntactic and semantic structure of the generated text [12, 13], increasing the richness and complexity of control signals. However, for the CIC approach to succeed, the CIC models need to be trained on equally rich datasets that incorporate, explicitly or implicitly, those control signals. Unfortunately, most image captioning datasets today, such as Flickr30k [36] or MS-COCO [18], lack this necessary diversity of controls and corresponding captions.", + "bbox": [ + 212, + 613, + 787, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our goal is to achieve SOTA performance in CIC without the need for new, increasingly rich, yet also costly, and impractical-to-collect datasets, where human workers would face the burden of having to provide multitudes of control", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "K. Basioti et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/90e3134b7c2d3a31a77a16c47dddb9977be796c23db20863276cbc3ef1646074.jpg", + "image_caption": [ + "Fig. 2: An example of our Structured Semantic Augmentation (SSA) approach. Visually-grounded captions (1)-(5) are used to create a meta-vgAMR graph, which includes all available image information in one representation. Sub-graphs of meta-vgAMR are then sampled to generate a new and diverse set of captions, such as the sentences (a)-(e). Image is licensed under CC BY-SA 2.0." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 787, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "signals and corresponding descriptive captions. To achieve this goal, we propose a novel Structured Semantic Augmentation (SSA) method, which automatically generates an augmented set of captions and the corresponding control signals with diverse spatiosemantic focus starting from only the core set of \"original\" uncontrolled captions. The method takes advantage of a detailed visual-linguistic semantic graph (illustrated in Fig. 2) constructed from the original captions and their image groundings. To build these semantic graphs, we use Abstract Meaning Representation (AMR) [6], a semantic formalism that can capture fine-grained linguistic relations beyond the exclusively spatial relationships present in the common scene graphs [24]. The availability of robust AMR parsers [5,9] allows us to generate semantic graphs for individual captions automatically, which we then merge into a rich meta-AMR graph for the joint image-language pair. From this meta-graph, we sample diverse connected subgraphs that represent semantically coherent combinations of image-anchored entities, events, and their relations, which we then turn into controlled captions automatically via existing AMR-to-text models [9]. Fig. 2 depicts an example of our meta-graph inferred from the original uncontrolled captions associated with an image. Filled nodes in the meta-graph indicate image entity groundings. Five semantically coherent subgraphs (a)-(e) of variable complexity are then sampled from the meta-graph, which are subsequently used to generate novel captions, shown below each subgraph. These new captions augment the original caption set by providing both image focus, through node groundings, and increased semantic diversity induced by the sampled subgraphs. Building upon SSA, we introduce a new CIC model, CIC-BART, suitable for generating focused controlled captions. Alongside the regions of interest, CIC-BART also makes use of the length of the desired cap", + "bbox": [ + 212, + 462, + 787, + 843 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "CIC-BART-SSA", + "bbox": [ + 620, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion as a control signal proxy for the verosity of the caption. CIC-BART can be trained on SSA-augmented versions of standard VL datasets such as MS-COCO or Flicker30k to accommodate the CIC task. Our experiments show that, compared to several SOTA models, the captions generated by our model have superior text quality and diversity, while being comparable in terms of faithfulness to control signals.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, our contributions are:", + "bbox": [ + 214, + 237, + 470, + 251 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose a novel data augmentation technique, SSA, that draws on a structured semantic formalism (AMR) to automatically generate focused captions suitable for training of CIC models. We empirically show that our SSA technique enables CIC models to generate captions with high controllability, diversity, and text quality.", + "2. We propose CIC-BART, a model designed for CIC, that does not require overly descriptive and complex control signals that SOTA models often require to achieve high performance. We show a superior overall performance, compared to SOTA, while relying on simple control signals (i.e., regions of interest and preferred caption length).", + "3. We present an extensive evaluation of our model, compared with existing SOTA. Specifically, we report results on different aspects of generated captions, including controllability (faithfulness to control signal), diversity, and text quality (linguistic well-formedness). To account for the trade-off among these metrics, we propose an overall performance score based on their harmonic mean. This metric helps us identify models that perform well in all these aspects." + ], + "bbox": [ + 212, + 252, + 787, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 518, + 387, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Controllable Image Captioning (CIC). Various types of control have been used for CIC, including visual entities, a type of region-based control [18, 23, 28, 47, 55, 56], where generated captions should learn to focus on the regions of interest. Others draw on complex control signals where additional knowledge about the generated caption structure is provided. For example, some recent work provides the complete skeleton of the desired sentence in the form of a number of objects or attributes or object-relation-object templates [12, 13]. Additional control signals that CIC draws on include different caption styles, e.g., positive, negative, humorous, or romantic tone [20, 21, 32, 33, 45, 47, 53, 54], user personality [17, 41], or the length of the generated captions [19, 22, 45, 47, 48]. The use of complex control signals aims at improving the diversity of captions and the quality of the text in CIC models. However, it requires the users to provide a detailed description of the control signal, which is not realistic in practical settings where such models are to be deployed (e.g., a self-driving car or personal assistant). We instead draw on two simple control signals (regions of interest and desired caption length) and show that we can achieve competitive performance on CIC, while keeping the control signals simple and practical.", + "bbox": [ + 212, + 551, + 787, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent SOTA models that draw on spatial control include the SCT model [18] that also uses the Faster R-CNN feature vectors and object tags (corresponding", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "K. Basioti et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "GloVe vectors [35]) of the entities of interest, as well as models that include skeleton-based control, namely ASG2Caption [13] and VSR [12]. ASG2Caption uses an abstract scene graph (ASG) to express the desired structure of a caption. ASG contains three types of unlabeled abstract nodes (object, attribute, relationship) that are grounded in the image by extracting features from the corresponding bounding boxes (for objects and attributes) or from the union of bounding box pairs (for a relationship node). ASG2Caption shows improved controllability (by conditioning on ASGs), and diversity (by automatically sampling diverse ASGs as control signals). The VSR model [12] draws on GloVe embeddings of Faster R-CNN object tags for visual entities (as in SCT). It also uses a skeleton control signal (like ASG2Caption), but one that includes more detailed information and richer semantics. Specifically, the VSR control signal follows the form of a fine-grained PropBank entry $^4$ — i.e., specifying the exact verb(s) expressing action(s) depicted in the image, and their visually grounded arguments (e.g., subject, object, location, manner). Thus, VSR uses the most descriptive control signal among the SOTA models. Refer to our supplementary material for an illustrative example of the control signal used for each method.", + "bbox": [ + 217, + 145, + 785, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared to ASG2Caption and VSR, our control signal is kept minimal and only specifies the bounding boxes and desired caption lengths. To improve the diversity of captions, we draw on a structured semantic graph (AMR) that expresses the semantics of a sentence based on PropBank semantics. Notably, we do not use these rich graphs to express detailed and overly descriptive control signals (as in VSR), but we use these semantic structures to augment our training data with richer and diverse captions, which will result in the model learning to generate more diverse captions. Additionally, we include a length control signal to further increase diversity without needing to specify detailed information about the structure of the output (e.g., number of attributes per object, etc.). This way, we can generate a variety of captions for a fixed image sub-region by simply controlling the desired length of the output.", + "bbox": [ + 217, + 404, + 785, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Abstract Meaning Representation (AMR). AMR [7] is a rich semantic formalism for expressing the meaning of natural language sentences as a formal graph. AMR draws on PropBank, which is a rich lexical semantic resource encoding predicates expressing an action or state, as well as the number and nature of the participating entities (arguments and other semantic roles, such as location, manner, etc.). AMR is a widely researched semantic formalism for which highly accurate automatic Text-to-AMR and AMR-to-Text models are developed [5,9]. We rely on these models to augment original image-caption datasets with newly generated captions (as explained in Sect. 4).", + "bbox": [ + 217, + 585, + 785, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "AMRs vs. Scene Graphs. Recent studies [1,15,16,52] have shown that AMRs better capture the semantic relations of an image as compared to the scene graphs [11]. Existing scene graph annotations mainly capture geometric or possessive relations, which account for more than $90\\%$ of the relations captured, whereas more than $1/3$ of the captured entities refer to clothing, object, or body parts information [1,52]. This difference is crucial for high-quality image caption-", + "bbox": [ + 217, + 720, + 785, + 811 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "CIC-BART-SSA", + "bbox": [ + 620, + 113, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "4 https://propbank.github.io", + "bbox": [ + 218, + 824, + 434, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ing, as we use higher-level semantic relations in our everyday language rather than geometric ones. For instance, during a soccer game, we would probably describe a goal save as 'the player kicks the ball away from the goal' or 'the goalkeeper defends his team by saving a goal' and not by using mainly geometric and possessive relations like 'a person wearing a white shirt, standing with his right leg lifted, close to a ball which is above the ground'. In the supplementary, we provide a detailed comparison of AMR and scene graph representations, particularly focusing on their applications in data augmentation.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3 Model", + "text_level": 1, + "bbox": [ + 215, + 287, + 316, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We propose CIC-BART, specifically designed to generate controlled image captions. Specifically, it can generate descriptions of particular areas within a scene with a desired level of detail. Our model, based on VL-BART [14], utilizes a transformer-based encoder-decoder architecture, as shown in Fig. 3. CIC-BART extends VL-BART encoder to the CIC task by modifying the encoder input to include: a) a global image embedding that provides the context of the", + "bbox": [ + 212, + 319, + 486, + 515 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5e17183f1d38b7c2a689d07aecff75c9c2478c6b41ce55d737590e0c60c794d6.jpg", + "image_caption": [ + "Fig.3: The architecture diagram of our model, CIC-BART." + ], + "image_footnote": [], + "bbox": [ + 506, + 363, + 781, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "full image to the model; b) the visual control signal, including the visual embedding of the regions that contain the entities of interest; c) the text control signal, containing length control (indicating the desired length range of the output caption) and an optional verb signal that indicates the action we want the generated caption to concentrate on.", + "bbox": [ + 212, + 516, + 787, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The visual embeddings of the regions are position-aware embeddings from a Faster R-CNN model [38] trained for visual object and attribute classification [3] on Visual Genome. The global image feature vector is extracted as well from Faster R-CNN. For the length control signal, we add to our vocabulary $L$ tokens for the $L$ different caption length levels; for instance, level one represents sentences between one and nine words, and level two, ten to nineteen. These tokens describe our coarse levels, for a finer sentence size accuracy, we accompany the tokens with the desired number of words. This choice gives our model the capacity to generate diverse captions for a particular length level. Finally, the output of the decoder generates the desired, controlled image caption.", + "bbox": [ + 212, + 590, + 787, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Structured Semantic Augmentation (SSA)", + "text_level": 1, + "bbox": [ + 214, + 763, + 660, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The goal of our SSA method is to augment existing image captioning datasets with new focused captions along with their control signals (i.e., regions corresponding to entities). We rely on datasets where visual entities in the captions", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "K. Basioti et al.", + "bbox": [ + 271, + 114, + 380, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "are annotated with their corresponding regions (see Sect. 5 for details on the datasets). The SSA process consists of four main steps, as described below. For more details, refer to our supplementary material, which also includes a step-by-step example of our SSA methodology.", + "bbox": [ + 212, + 146, + 784, + 205 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Step 1: Image-level AMR graph generation. Our objective in this stage is to enclose all the information available from the visually grounded captions into a single representation. To accomplish this, we create a visually grounded AMR graph (vgAMR) for each caption of an image and then merge them into a single image-level graph, the meta-vgAMR. To create the vgAMRs of an image, we first convert each of its $N$ captions to their AMR representation, using the Neural transition-based Text-to-AMR parser [5] which also aligns words in a caption with their respective nodes in the AMR graph. We utilize the alignment information of 1) caption words and AMR nodes (from Text-to-AMR parser) and 2) caption words to image bounding boxes (from existing dataset annotations) to visually ground the AMR nodes. After this step, we get the collection of nodes referring to visual entities, where each grounded meta-AMR node is linked with a non-empty set of bounding boxes. This extended representation, 'AMR + visual grounded nodes', is our vgaMR.", + "bbox": [ + 212, + 207, + 785, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our next step is to combine the N vgAMRs to form a single meta-vgAMR. To achieve this, we employ a pairwise strategy to merge the most similar vgAMRs first (we measure similarity with Smatch score [10]). We use the UPGMA hierarchical clustering algorithm [30,34] to find the optimal merge ordering starting from the most similar graphs. UPGMA creates a hierarchy where the bottom level consists of the N individual vgAMRs. By merging all vgAMRs using the UPGMA ordering, we obtain a single structure called meta-vgAMR.", + "bbox": [ + 212, + 417, + 785, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "When merging two vgAMRs, the main challenge is identifying which nodes correspond to the same concepts, such as entities, attributes, actions, and relations. We use three node properties to accomplish this: a) visual grounding information, b) semantic similarity of node labels, and c) node neighborhood semantic similarity. We derive two node-merging criteria from there: 1) visually grounded entity nodes are merged if they point to the same image-bounding boxes. When 1) does not hold, we check the second criterion: 2) for the remaining non-grounded nodes, including amr-specific, predicates, adjectives, and adverbs, we use a combination of node label semantic similarity (cosine similarity of the labels using their GloVe embeddings) and neighborhood similarity. Neighborhood similarity examines the similarity of parents for adjectives/adverbs nodes and children for predicate nodes, along with the similarity of connecting edge roles. When two nodes satisfy criterion 1) or 2), we merge them into a single node. Moreover, if they have different labels, we maintain both names by keeping a list of synonyms to increase representation diversity. In the special case when the two vgAMRs describe two totally different concepts, and hence they have no common nodes, we add an amr-specific node called 'multi-sentence' as the root with the two independent vgAMRs as its children. The final graph, meta", + "bbox": [ + 212, + 523, + 785, + 796 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "CIC-BART-SSA", + "bbox": [ + 620, + 113, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "vgAMR, includes all non-redundant5 elements of the original $N$ captions while preserving the visual grounding between the meta nodes and their respective image regions.", + "bbox": [ + 212, + 145, + 785, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Remark: Meta-vgAMR efficiently compresses all available image information into a single structure. Following our approach, we can easily scale when new scene information becomes available by applying our pairwise merge procedure.", + "bbox": [ + 212, + 191, + 785, + 236 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Step 2: Event-based graph sampling from image-level AMRs. We start from the predicate nodes, which mainly correspond to verbs, to sample subgraphs in meta-vgAMR graphs. Predicate nodes are identified by their label and the edges connected to them. The label of a predicate node typically follows the format 'predicate_name-xx,' where 'xx' represents the different senses a word can have regarding the concept it is used for. Predicate nodes have outward ARGy edges, where 'y' can take values from 0 to 5, connecting them to their arguments. We sample subgraphs from these nodes by following the outgoing argument edges, which are labeled as ARGn in an AMR graph, each defining a particular semantic role (e.g., ARG0 points to the agent, ARG1 to the patient, etc.). Finally, we add one more subgraph containing the remaining children branches of other non-ARG optional predicate edges (e.g., 'location', 'time'). We repeat this process until the leaves of the graphs are reached. During sampling, we randomly select one of the synonyms if a node is a list of synonym labels, as mentioned in the previous step. The output of this step is our more focused event-focused sub-graphs. In Fig. 2, we can see some instances of our event-based sampling (SSA samples), where the predicate nodes include z0/sit-01, z7/dock-01, z13/sit-01, and so on6. Although we cannot show all the sampled event-based subgraphs in the figure, we included five of them and used colored roots and edges for visualization purposes.", + "bbox": [ + 212, + 237, + 787, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Step 3: New caption generation from sampled AMRs We use the SPRING AMR-to-Text model [8] to generate new event-focused captions from the sampled vgAMR subgraphs. Because both vgAMR merging and sampling steps introduce noise, the output captions are not always of good quality. We automatically filter low-quality captions by using a linguistic well-formedness measure, GRUEN [58], which is a reference-free metric based on BERT contextual embeddings.", + "bbox": [ + 212, + 539, + 789, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Step 4: Control signal generation. The last step is to create the control signal for the generated captions. The spatial control signal for a specific caption is extracted from the corresponding sampled vgAMR, by pulling the bounding boxes of the visual entity linked AMR nodes.", + "bbox": [ + 212, + 628, + 785, + 690 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "K. Basioti et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "5 A node may have different names for the same bounding box in different meta-vgAMRs, such as 'A male' and 'A person'. According to criterion 1), we merge the corresponding AMR nodes and keep both 'male' and 'person' in the names list to avoid redundancy. Therefore, criteria 1) and 2) ensure that multiple nodes don't describe the same concept in the meta-vgAMR.", + "bbox": [ + 217, + 700, + 785, + 768 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "Note that in Fig. 2, the node z17/green-03 is also categorized as a predicate. This may seem an error because we usually think of 'green' as an attribute node rather than a predicate. However, in AMRs, when 'green' is paired with its argument, in this case, z16/lawn, it encapsulates a predicate/verb that can be expressed in natural language as 'the lawn is green.'", + "bbox": [ + 218, + 770, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Mixing Strategies of Original and SSA Data", + "text_level": 1, + "bbox": [ + 215, + 146, + 627, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To analyze the impact of our SSA data, we explore various mixing strategies with the original training set. Assume $\\mathcal{D}$ represents the training control-caption pairs in the original dataset, containing $N_{\\mathcal{D}}$ samples, and SSA represents our SSA samples, containing $N_{SSA}$ instances. The augmented dataset $\\mathcal{D}_{SSA}$ is defined by combining $\\mathcal{D}$ and SSA: $\\mathcal{D}_{SSA} = \\mathrm{sam}_{\\mathcal{D}}(\\tau_{\\mathcal{D}}, p_{\\mathcal{D}}) \\cup \\mathrm{sam}_{SSA}(\\tau_{SSA}, p_{SSA})$ , where the functions $\\mathrm{sam}_{\\mathcal{D}}$ samples a subset of the original dataset, and $\\mathrm{sam}_{SSA}$ a subset of our SSA data. Since we are interested in the effect of our SSA, we assume that $\\mathrm{sam}_{\\mathcal{D}}(\\tau_{\\mathcal{D}}, p_{\\mathcal{D}}) = \\mathcal{D}$ , with $\\tau_{\\mathcal{D}} = \\text{'Random Sampling Strategy'}$ and $p_{\\mathcal{D}} = 100\\%$ , meaning that all original data are included in the mixed dataset. Depending on the $\\mathrm{sam}_{SSA}$ parameter $\\tau_{SSA}$ we have the cases:", + "bbox": [ + 212, + 175, + 787, + 325 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Random Sampling Strategy. In this case, we randomly select a pre-specified number of examples from $SSA$ . The parameter $p_{SSA}$ expresses the percentage of SSA samples included in $\\mathcal{D}_{SSA}$ . With boundary cases $p_{SSA} = 100\\%$ (all $N_{SSA}$ samples are included), and $p_{SSA} = 0\\%$ (no SSA data are added).", + "bbox": [ + 212, + 325, + 787, + 385 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Uniform-Coverage Sampling Strategy. To mitigate the original dataset's bias (having mainly samples describing the entire image), we aim to create a new focus-unbiased dataset. By modeling the control signal focus as the image area percentage covered by the bounding boxes of the control signal, we split the original data into $B$ coverage bins. Then, we will randomly add in each bin SSA samples, aiming to create a new uniform, coverage-unbiased $\\mathcal{D}_{SSA}$ dataset. Here, $p_{SSA}$ contains the range of each bin for the coverage histogram. For example, in the case where we choose ten uniform coverage bins, we have $p_{SSA} = \\{[0\\%, 10\\%), [10\\%, 20\\%), \\ldots, [90\\%, 100\\%]\\}$ .", + "bbox": [ + 212, + 386, + 787, + 523 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Due to space limitations, we present results from the Random Sampling Strategy for $p_{SSA} = 0\\%$ and $p_{SSA} = 100\\%$ in the main paper. Results from other scenarios can be found in the supplementary material.", + "bbox": [ + 212, + 523, + 787, + 568 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 595, + 444, + 612 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Data", + "text_level": 1, + "bbox": [ + 215, + 630, + 305, + 643 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We use Flickr30k Entities (Flickr-Ent) [36] and MS-COCO Entities (COCO-Ent) [18] for training and evaluation. Flickr-Ent augments the original captions of Flickr30k [51] with manually-annotated region-phrase groundings. Flickr-Ent contains the original $31K$ images annotated with five captions each. COCO-Ent augments the original MS-COCO [27] ( $120K$ images each annotated with around five captions) with semi-automatically collected grounding annotations; see [18] for details on the annotation process. For both datasets, we follow previous work and use the training and test splits by Karpathy et al. [25]. We apply our SSA algorithm on the aforementioned datasets to create their augmented variations, COCO-Ent-SSA and Flickr-Ent-SSA, containing about $800K$ and $250K$ training captions, respectively, of which $33\\%$ and $37\\%$ , are generated by our SSA algorithm.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "CIC-BART-SSA", + "bbox": [ + 620, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For all four training sets, we automatically generate image-control-caption triplets to train our model on. For spatial control, we extract from the grounded captions the bounding boxes of the entities of interest using the annotations from COCO-Ent and Flickr-Ent. Note that since these datasets do not contain the text control signal, we use each ground-truth caption as a proxy for a controlled caption, from which we first generate the coarse- and fine-length control levels and then extract their verbs using part-of-speech tagging for the optional action control.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Models and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 214, + 287, + 521, + 303 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We compare two variations of our model (with and without SSA augmentations) with SOTA models as our baselines: Show Control & Tell (SCT) [18] that uses region-based control (bounding boxes of visual entities of interest); ASG2Caption (ASG) [13] that draws on visually grounded abstract scene graphs as control signal; and VSR [12] that uses overly descriptive control signals that express verb(s) and fine-grained verb-specific semantic roles of the desired captions; ComPro [48] that learns a mapping from the bounding boxes of the entities of interest and caption length level to GPT-2 Large prompts aiming to retrieve controlled captions; and the LaBERT length-control-only model [19].", + "bbox": [ + 212, + 311, + 785, + 446 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We report the performance of our models and baselines using a comprehensive set of metrics that evaluate different aspects of caption controllability, diversity, and quality. We also propose and report an overall performance metric that summarizes these different aspects in a meaningful way. To measure diversity, we compute n-gram diversity, D-n for $n = 1$ , 2 [4], as well as self-CIDEr-based diversity (sC) [46]. For a fair comparison of the different CIC models, we measure diversity for the five generated captions for each test image (in COCO-Ent and Flickr-Ent), and report their average. To measure content controllability we design an extended version of the IoU metric of [18] that calculates the degree-of-match (faithfulness) between a control signal and the corresponding generated caption. For our control signal, we use the set of nouns $\\mathcal{E}$ that represent the entities of interest, which are the names of the visual objects in the control. To extract nouns from the predicted sentences, we use the Stanford part-of-speech tagger [42,43]. We then find the semantic intersection of the two sets using Hungarian Matching, as in [18]. Finally, we calculate the semantic intersection over union of the control nouns and the nouns extracted from the controllable caption, which gives us our content controllability IoU. We further analyze IoU by introducing the Hallucinating Nouns (Hal) metric. Details can be found in the supplemental material. For length controllability (L), we measure the Mean Absolute Error (MAE) between the fine length control (number of words) and the size of the resulting $M = 10$ controlled captions, which are generated from $M$ randomly created control signals. We also calculate the length precision (LP) [19] by determining the percentage of generated captions that match the desired coarse length level. We assess text quality of generated captions using GRUEN (G) [58], a reference free metric based on BERT contextual embeddings that measures the syntactic and semantic well-formedness of a text segment. Finally, we measure", + "bbox": [ + 212, + 448, + 787, + 843 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "K. Basioti et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/c025e3ee772d999ecb30177ed6ba071d34823eef49f89cfa508ea69526047013.jpg", + "table_caption": [ + "Table 1: Performance of CIC models for the original test sets. *ASG-type dataset is not released for Flickr-Ent; therefore, we could not reproduce the ASG results." + ], + "table_footnote": [], + "table_body": "
ModelH↑IoU↑G↑sC↑D-1↑D-2↑L↓H↑IoU↑G↑sC↑D-1↑D-2↑L↓
SCT [18]55.867.364.442.827.035.5-54.650.779.844.029.336.5-
ASG* [13]74.272.672.078.337.856.6--------
VSR [12]56.277.639.067.430.042.2-62.560.254.077.933.349.3-
CIC-BART75.976.273.078.738.056.2.4969.854.085.078.643.658.21.24
CIC-BART-SSA78.377.274.882.544.663.2.1171.355.086.081.747.062.61.05
", + "bbox": [ + 218, + 184, + 779, + 272 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/a3a790e519776ac2eb7921d99b905a9c3163112d17015e5be2bcdef583b76290.jpg", + "table_caption": [ + "Table 2: Performance of CIC models, for the SSA-only samples." + ], + "table_footnote": [], + "table_body": "
ModelH↑IoU↑G↑sC↑D-1↑D-2↑H↑IoU↑G↑sC↑D-1↑D-2↑
COCO-Ent (SSA only)Flickr-Ent (SSA only)
SCT [18]51.762.164.837.823.731.043.929.977.345.731.036.7
CIC-BART69.261.473.974.044.257.068.553.080.579.852.962.9
CIC-BART-SSA75.665.280.783.753.867.872.055.682.986.156.569.3
", + "bbox": [ + 225, + 310, + 771, + 378 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "the overall performance of each model based on its ability to balance content controllability, diversity, and text quality. To calculate this, we use the harmonic mean of IoU, G, and sC. All of these metrics range between 0 and 1, with a higher value indicating better performance. The harmonic mean $(H)$ helps us determine the model with the best overall performance. It prioritizes models that perform well across all metrics while penalizing those with poor performance, even in one metric. In our supplementary, we include comparisons of the CIC models on standard captioning metrics (like CIDEr [44] and Spice [2]).", + "bbox": [ + 212, + 404, + 787, + 526 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6 Results", + "text_level": 1, + "bbox": [ + 215, + 546, + 326, + 561 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6.1 Overall Performance", + "text_level": 1, + "bbox": [ + 215, + 575, + 433, + 589 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We first compare the overall performance of our models with the SCT, ASG, and VSR baselines with respect to controllable captioning metrics. We do not include ComPro in this comparison due to the unavailability of the codebase. We also exclude LaBERT since this model solely focuses on length controllability.", + "bbox": [ + 212, + 598, + 785, + 657 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tab. 1 presents results for content and length controllability (IoU and L), text quality (G), and diversity (sC, D-1, D-2), as well as the harmonic mean (H). For both datasets, CIC-BART-SSA has the best performance in all metrics, except IoU, where it is the second best. Specifically, CIC-BART-SSA is superior to all other models with respect to diversity (sC, D-1, D-2) and text quality (G), but comparable to VSR in terms of content controllability (IoU). The length controllability (L) scores show that our SSA augmentation helps the model learn to generate high-quality output at the desirable length (compare CIC-BART and CIC-BART-SSA). This is due to the increased diversity in caption length provided by our SSA augmentations.", + "bbox": [ + 212, + 659, + 787, + 809 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Importantly, we can see that model performance can vary depending on the metric. E.g., whereas VSR has the highest IoU, it falls behind in text quality and", + "bbox": [ + 214, + 810, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "CIC-BART-SSA", + "bbox": [ + 620, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "diversity. In our qualitative analysis, we observe the poor quality of the captions generated by VSR. The best-performing model should be identified based on the $H$ score that summarizes content controllability, text quality, and diversity into a single score. Based on this score, CIC-BART-SSA is better than SCT and VSR baselines by a large margin, and notably better than ASG. Nevertheless, ASG requires complex control signals in the form of scene graphs, in contrast to the simple control signal requirements of CIC-BART.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Next, we conduct a further evaluation of the CIC performance on our SSA samples from the test set images of COCO-Ent and Flickr-Ent. We present the results in Tab. 2 for our models CIC-BART, CIC-BART-SSA, and SCT. ASG and VSR are excluded since they need complex control signals (grounded abstract scene graphs for ASG and grounded verb semantic roles for VSR), which are only available for COCO-Ent and Flickr-Ent. We observe a significant improvement in overall performance for our CIC-BART-SSA model. We notice that the models (SCT and CIC-BART) trained on the original datasets, which describe the entire image, had difficulties generalizing to cases where they had to focus on a specific sub-region of an image. However, our model CIC-BART-SSA was able to generate focused and diverse descriptions of the challenging, highly focused examples present in our SSA data.", + "bbox": [ + 212, + 252, + 787, + 433 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We compare the length precision of our model with the baselines utilizing length control in Table 3. LaBERT uses only length-control signals without spatial control, while our model employs both spatial and length-control signals to generate focused captions. This makes the LaBERT task much easier since it only focuses on generating specific length descriptions of an image. On the other hand, our model focuses on generating captions that describe only a specific sub-region of the scene while maintaining a desired description length level. Although our task is", + "bbox": [ + 212, + 433, + 539, + 628 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/863a3c32efcf233ccfb84cde7a9d1a51fb702e297d4f1a2ad386763d4acd31a5.jpg", + "table_caption": [ + "Table 3: Length Precision (LP) for CIC models on COCO-Ent and Flickr-Ent original test sets." + ], + "table_footnote": [], + "table_body": "
ModelLP↑LP↑
COCO-EntFlickr-Ent
ComPro [48]94.781.4
LaBERT [19]99.798.4
CIC-BART99.988.0
CIC-BART-SSA99.991.3
", + "bbox": [ + 550, + 518, + 784, + 595 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "more challenging than LaBERT, we achieve competitive length precision performance. Lastly, we want to emphasize that the remarkable improvement in length controllability (L) and length precision (LP) from CIC-BART to CIC-BART-SSA stems from the increased length diversity found in our SSA augmentations, which enriches the original COCO-Ent and Flickr-Ent datasets. In the supplementary, we provide an analysis of the caption length statistics in the original datasets and our SSA-derived captions.", + "bbox": [ + 212, + 630, + 787, + 736 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6.2 Effect of SSA on Content Controllability", + "text_level": 1, + "bbox": [ + 214, + 755, + 596, + 771 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To analyze the impact of our SSA augmentations, we measure the content controllability (IoU) performance of CIC-BART at different levels of focus of the control signals and report it in Fig. 4. We use coverage, defined as the area of the image enclosed by the bounding boxes of the entities of interest in the control", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "K. Basioti et al.", + "bbox": [ + 271, + 114, + 380, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/df739c11ae44e09b08c33caa8eeb951bd62d3267c34302cd30000cb3e0134f4a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 212, + 143, + 514, + 236 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d2aebe78818bc5dbd0a8fec51790e1acc47c64a6dde0b9922686aa031aaf33ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 142, + 785, + 236 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7de535f4d8e4e743029c56f016ee5af8bc6d829a7c4b8659a989fc65da71bdd5.jpg", + "image_caption": [ + "Fig. 4: IoU histograms for CIC-BART (green) and CIC-BART-SSA (blue). The first column depicts the IoU on the original test set and the second on the original test set images' SSA (only) data. The Samples curve (orange) represents the distribution of test captions in each coverage (image area covered by the control signal) interval." + ], + "image_footnote": [], + "bbox": [ + 212, + 243, + 514, + 320 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2a1a95078e415f9689cb7116e18e8c9ccedce9b3f373eac72de65623bbb6e1b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 243, + 785, + 320 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3fd65594618cb87496e6f6382f1a7b5649d323fe7e3788e11c718565d96629ea.jpg", + "image_caption": [ + "Fig. 5: Qualitative examples for the original test sets. Strikethrough marks hallucinations and redundancies. a, b licensed under CC BY-SA 2.0; c, d under CC BY 2.0." + ], + "image_footnote": [], + "bbox": [ + 215, + 417, + 460, + 556 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/fa718370c4bfe476d621f99eb49475122f4ae6f87f22bbbc60f76206f450ed1c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 462, + 419, + 785, + 555 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "signal, to quantify that focus. For example, highly focused control signals cover a small area, yielding low coverage, while broader signals cover a larger area and have high coverage. We 'break down' the IoU performance into 10 coverage bands and report the average IoU over control signals in those bands. In addition, the 'Samples' curve shows the distribution of test captions over the same bands. The results in Fig. 4 indicate that by training with SSA (blue bars), the spatial controllability improves significantly in the low-coverage regime, where the control signals are highly focused. Interestingly, these are also the most underrepresented (data deprived) parts of the original dataset Flickr-Ent. Therefore, SSA which enriches the original datasets with highly focused examples (please refer to our supplementary for % Samples per coverage bands for the training sets), is effective in improving generalization performance in CIC. We include the performance of hallucinating nouns metric in the supplementary material.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "CIC-BART-SSA", + "bbox": [ + 620, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.3 Qualitative Analysis", + "text_level": 1, + "bbox": [ + 215, + 146, + 429, + 162 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In Fig. 5, we present qualitative examples from the original test sets, and in Fig. 6 examples from our SSA (only) test set control signals. In the two figures, each highlighted word found in the generated controlled captions corresponds to the control entity of the same color. This shows the match between the captions produced and the control signal. We also strike through the parts where the model hallucinates or generates redundant references to the entities of interest. Our models have been observed to outperform the previous state-of-the-art models by substantially enhancing the quality of the generated controlled captions. This behavior was expected from our quantitative analysis, which showed that our models have signifi", + "bbox": [ + 212, + 174, + 488, + 474 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "cantly higher text quality (G). More importantly, our CIC-BART-SSA model is capable of generating captions that are faithful to the control signal and better understand the relationships that connect the entities of interest. We include additional qualitative samples in the Supplementary Material.", + "bbox": [ + 212, + 476, + 787, + 537 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/6eaa26b9b592b75b60512fc99bcec7e07ee58b015f5d94ade8de858079ea21d2.jpg", + "image_caption": [ + "Fig. 6: Examples for SSA-only test set. Strikethrough marks hallucinations. Up to Down: a licensed under CC BY 2.0; b under CC PDM 1.0, for last two see Fig. 5." + ], + "image_footnote": [], + "bbox": [ + 504, + 218, + 781, + 367 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 563, + 369, + 579 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We address two main challenges faced by the controllable image captioning (CIC) models. First, standard image-caption datasets lack the controllability and diversity needed for proper training and evaluation of CIC. Second, most recent SOTA models require complex and overly descriptive control signals as input (including, e.g., the main action/verb to appear in the generated caption). To address the first challenge, we propose a novel technique that draws on a structured semantic augmentation (SSA) formalism to generate focused captions and the corresponding control signals for images. For the second challenge, we propose a transformer-based vision-language model attuned to the CIC task. We show that this model performs competitively with SOTA models without requiring complex and explicit control signals. Importantly, when combined with our SSA approach, our model generates highly diverse captions and significantly reduces the content controllability performance gap between the different levels of focus of the generated controlled captions. Finally, when provided with the commonly used verb guidance of other SOTA approaches, our model shows a substantial improvement in performance.", + "bbox": [ + 212, + 597, + 789, + 843 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "K. Basioti et al.", + "bbox": [ + 271, + 114, + 380, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abdelsalam, M.A., Shi, Z., Fancellu, F., Basioti, K., Bhatt, D., Pavlovic, V., Fazly, A.: Visual semantic parsing: From images to abstract meaning representation. In: Proceedings of the 26th Conference on Computational Natural Language Learning (CoNLL). pp. 282-300 (2022)", + "2. Anderson, P., Fernando, B., Johnson, M., Gould, S.: Spice: Semantic propositional image caption evaluation. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 382-398. Springer (2016)", + "3. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: CVPR (2018)", + "4. Aneja, J., Agrawal, H., Batra, D., Schwing, A.: Sequential latent spaces for modeling the intention during diverse image captioning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4261-4270 (2019)", + "5. Astudillo, R.F., Ballesteros, M., Naseem, T., Blodgett, A., Florian, R.: Transition-based parsing with stack-transformers. arXiv preprint arXiv:2010.10669 (2020)", + "6. Banarescu, L., Bonial, C., Cai, S., Georgescu, M., Griffitht, K., Hermjakob, U., Knight, K., Koehn, P., Palmer, M., Schneider, N.: Abstract meaning representation (amr) 1.0 specification. In: Parsing on Freebase from Question-Answer Pairs. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing. Seattle: ACL. pp. 1533–1544 (2012)", + "7. Banarescu, L., Bonial, C., Cai, S., Georgescu, M., Griffitht, K., Hermjakob, U., Knight, K., Koehn, P., Palmer, M., Schneider, N.: Abstract meaning representation for sembanking. In: Proceedings of the 7th linguistic annotation workshop and interoperability with discourse. pp. 178-186 (2013)", + "8. Bevilacqua, M., Błoshmi, R., Navigli, R.: One spring to rule them both: Symmetric amr semantic parsing and generation without a complex pipeline. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 12564-12573 (2021)", + "9. Bloshmi, R., Bevilacqua, M., Fabiano, E., Caruso, V., Navigli, R.: Spring goes online: end-to-end AMR parsing and generation. In: Proceedings of the 2021 conference on empirical methods in natural language processing: system demonstrations. pp. 134-142 (2021)", + "10. Cai, S., Knight, K.: Smatch: an evaluation metric for semantic feature structures. In: Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). pp. 748-752 (2013)", + "1. Chang, X., Ren, P., Xu, P., Li, Z., Chen, X., Hauptmann, A.: A comprehensive survey of scene graphs: Generation and application. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(1), 1-26 (2021)", + "2. Chen, L., Jiang, Z., Xiao, J., Liu, W.: Human-like controllable image captioning with verb-specific semantic roles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16846-16856 (2021)", + "3. Chen, S., Jin, Q., Wang, P., Wu, Q.: Say as you wish: Fine-grained control of image caption generation with abstract scene graphs. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9962-9971 (2020)", + "4. Cho, J., Lei, J., Tan, H., Bansal, M.: Unifying vision-and-language tasks via text generation. In: International Conference on Machine Learning. pp. 1931-1942. PMLR (2021)" + ], + "bbox": [ + 225, + 179, + 785, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "CIC-BART-SSA", + "bbox": [ + 620, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Choi, W.S., Heo, Y.J., Punithan, D., Zhang, B.T.: Scene graph parsing via abstract meaning representation in pre-trained language models. In: NAACL 2022 Workshop on Deep Learning on Graphs for Natural Language Processing (2022)", + "16. Choi, W.S., Heo, Y.J., Zhang, B.T.: Sgram: Improving scene graph parsing via abstract meaning representation. arXiv preprint arXiv:2210.08675 (2022)", + "17. Chunseong Park, C., Kim, B., Kim, G.: Attend to you: Personalized image captioning with context sequence memory networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 895-903 (2017)", + "18. Cornia, M., Baraldi, L., Cucchiara, R.: Show, control and tell: A framework for generating controllable and grounded captions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8307-8316 (2019)", + "19. Deng, C., Ding, N., Tan, M., Wu, Q.: Length-controllable image captioning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16. pp. 712-729. Springer (2020)", + "20. Gan, C., Gan, Z., He, X., Gao, J., Deng, L.: Stylenet: Generating attractive visual captions with styles. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3137-3146 (2017)", + "21. Guo, L., Liu, J., Yao, P., Li, J., Lu, H.: Mscap: Multi-style image captioning with unpaired stylized text. 2019 IEEE. In: CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4199-4208 (2019)", + "22. Hirsch, E., Tal, A.: Clid: Controlled-length image descriptions with limited data. arXiv preprint arXiv:2211.14835 (2022)", + "23. Johnson, J., Karpathy, A., Fei-Fei, L.: Densecap: Fully convolutional localization networks for dense captioning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4565-4574 (2016)", + "24. Johnson, J., Krishna, R., Stark, M., Li, L.J., Shamma, D., Bernstein, M., Fei-Fei, L.: Image retrieval using scene graphs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3668-3678 (2015)", + "25. Karpathy, A., Fei-Fei, L.: Deep visual-semantic alignments for generating image descriptions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3128-3137 (2015)", + "26. Li, X., Yin, X., Li, C., Zhang, P., Hu, X., Zhang, L., Wang, L., Hu, H., Dong, L., Wei, F., et al.: Oscar: Object-semantics aligned pre-training for vision-language tasks. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16. pp. 121-137. Springer (2020)", + "27. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "28. Lindh, A., Ross, R.J., Kelleher, J.D.: Language-driven region pointer advancement for controllable image captioning. arXiv preprint arXiv:2011.14901 (2020)", + "29. Lu, J., Yang, J., Batra, D., Parikh, D.: Neural baby talk. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7219-7228 (2018)", + "30. Lukasova, A.: Hierarchical agglomerative clustering procedure. Pattern Recognition 11(5-6), 365-381 (1979)", + "31. Luo, J., Li, Y., Pan, Y., Yao, T., Feng, J., Chao, H., Mei, T.: Semantic-conditional diffusion networks for image captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23359-23368 (2023)", + "32. Mathews, A., Xie, L., He, X.: Senticap: Generating image descriptions with sentiments. In: Proceedings of the AAAI conference on artificial intelligence. vol. 30 (2016)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "K. Basioti et al.", + "bbox": [ + 271, + 114, + 379, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Mathews, A., Xie, L., He, X.: Semstyle: Learning to generate stylised image captions using unaligned text. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8591-8600 (2018)", + "34. Müller, D.: Modern hierarchical, agglomerative clustering algorithms. arXiv preprint arXiv:1109.2378 (2011)", + "35. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP). pp. 1532-1543 (2014)", + "36. Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In: Proceedings of the IEEE international conference on computer vision. pp. 2641-2649 (2015)", + "37. Ramos, R., Martins, B., Elliott, D., Kementchedjheva, Y.: Smallcap: lightweight image captioning prompted with retrieval augmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2840-2849 (2023)", + "38. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems 28 (2015)", + "39. Ren, Y., Mao, Z., Fang, S., Lu, Y., He, T., Du, H., Zhang, Y., Ouyang, W.: Cross-ing the gap: Domain generalization for image captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2871-2880 (2023)", + "40. Rennie, S.J., Marcheret, E., Mroueh, Y., Ross, J., Goel, V.: Self-critical sequence training for image captioning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7008-7024 (2017)", + "41. Shuster, K., Humeau, S., Hu, H., Bordes, A., Weston, J.: Engaging image captioning via personality. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12516-12526 (2019)", + "42. Toutanova, K., Klein, D., Manning, C.D., Singer, Y.: Feature-rich part-of-speech tagging with a cyclic dependency network. In: Proceedings of the 2003 human language technology conference of the north american chapter of the association for computational linguistics. pp. 252-259 (2003)", + "43. Toutanvoa, K., Manning, C.D.: Enriching the knowledge sources used in a maximum entropy part-of-speech tagger. In: 2000 Joint SIGDAT conference on Empirical methods in natural language processing and very large corpora. pp. 63-70 (2000)", + "44. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4566-4575 (2015)", + "45. Wang, N., Xie, J., Wu, J., Jia, M., Li, L.: Controllable image captioning via prompting. In: AAAI (2023)", + "46. Wang, Q., Chan, A.B.: Describing like humans: on diversity in image captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4195-4203 (2019)", + "47. Wang, T., Zhang, J., Fei, J., Ge, Y., Zheng, H., Tang, Y., Li, Z., Gao, M., Zhao, S., Shan, Y., et al.: Caption anything: Interactive image description with diverse multimodal controls. arXiv preprint arXiv:2305.02677 (2023)", + "48. Wang, Z., Xiao, J., Chen, L., Gao, F., Shao, J., Chen, L.: Learning combinatorial prompts for universal controllable image captioning. arXiv preprint arXiv:2303.06338 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "CIC-BART-SSA", + "bbox": [ + 620, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. Xia, Q., Huang, H., Duan, N., Zhang, D., Ji, L., Sui, Z., Cui, E., Bharti, T., Zhou, M.: Xgpt: Cross-modal generative pre-training for image captioning. In: Natural Language Processing and Chinese Computing: 10th CCF International Conference, NLPCC 2021, Qingdao, China, October 13-17, 2021, Proceedings, Part I 10. pp. 786-797. Springer (2021)", + "50. Xu, K., Ba, J., Kiros, R., Cho, K., Courville, A., Salakhudinov, R., Zemel, R., Bengio, Y.: Show, attend and tell: Neural image caption generation with visual attention. In: International conference on machine learning. pp. 2048-2057. PMLR (2015)", + "51. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics 2, 67-78 (2014)", + "52. Zellers, R., Yatskar, M., Thomson, S., Choi, Y.: Neural motifs: Scene graph parsing with global context. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5831-5840 (2018)", + "53. Zeng, Z., Zhang, H., Lu, R., Wang, D., Chen, B., Wang, Z.: Conzic: Controllable zero-shot image captioning by sampling-based polishing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23465-23476 (2023)", + "54. Zhao, W., Wu, X., Zhang, X.: Memcap: Memorizing style knowledge for image captioning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 12984-12992 (2020)", + "55. Zhao, Y., Wei, J., Lin, Z., Sun, Y., Zhang, M., Zhang, M.: Visual spatial description: Controlled spatial-oriented image-to-text generation. arXiv preprint arXiv:2210.11109 (2022)", + "56. Zheng, Y., Li, Y., Wang, S.: Intention oriented image captions with guiding objects. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8395-8404 (2019)", + "57. Zhong, Y., Wang, L., Chen, J., Yu, D., Li, Y.: Comprehensive image captioning via scene graph decomposition. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 211-229. Springer (2020)", + "58. Zhu, W., Bhat, S.: Gruen for evaluating linguistic quality of generated text. arXiv preprint arXiv:2010.02498 (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 619 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "K. Basioti et al.", + "bbox": [ + 271, + 114, + 380, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/CIC-BART-SSA_ _ Controllable Image Captioning with Structured Semantic Augmentation/87b30ae0-4f07-48fc-b46b-fbcdc143e18d_model.json b/2024/CIC-BART-SSA_ _ Controllable Image Captioning with Structured Semantic Augmentation/87b30ae0-4f07-48fc-b46b-fbcdc143e18d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ae5374bb9287d3c44e9bdbd88aaebe1d56437f1f --- /dev/null +++ b/2024/CIC-BART-SSA_ _ Controllable Image Captioning with Structured Semantic Augmentation/87b30ae0-4f07-48fc-b46b-fbcdc143e18d_model.json @@ -0,0 +1,2447 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.225, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "CIC-BART-SSA: Controllable Image Captioning with Structured Semantic Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.212, + 0.782, + 0.245 + ], + "angle": 0, + "content": "Kalliopi Basioti\\(^{1,2*}\\), Mohamed A. Abdelsalam\\(^{2}\\), Federico Fancellu\\(^{3\\dagger}\\), Vladimir Pavlovic\\(^{1\\dagger}\\), and Afsaneh Fazly\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.255, + 0.704, + 0.283 + ], + "angle": 0, + "content": "1 Rutgers University, New Jersey, USA {kalliopi.basioti, vladimir}@rutgers.edu" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.283, + 0.724, + 0.31 + ], + "angle": 0, + "content": "\\(^{2}\\) Samsung AI Centre - Toronto, Toronto, Canada {m.abdelsalam, a.fazly}@samsung.com" + }, + { + "type": "text", + "bbox": [ + 0.367, + 0.31, + 0.636, + 0.324 + ], + "angle": 0, + "content": "3 Solventum ffancellu@solventum.com" + }, + { + "type": "list", + "bbox": [ + 0.277, + 0.255, + 0.724, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.344, + 0.361, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.417, + 0.38, + 0.461 + ], + "angle": 0, + "content": "Original Captions A couple of elephants standing next to trees. A couple of elephants standing by themselves. A couple elephants walking by a tree after sunset. Two elephants are standing in the grass near a tree. Two elephants are standing by the trees in the wild." + }, + { + "type": "title", + "bbox": [ + 0.384, + 0.344, + 0.528, + 0.352 + ], + "angle": 0, + "content": "Samples Generated from CIC-BART-SSA" + }, + { + "type": "text", + "bbox": [ + 0.385, + 0.353, + 0.478, + 0.36 + ], + "angle": 0, + "content": "1. There is grass near a tree." + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.36, + 0.49, + 0.366 + ], + "angle": 0, + "content": "2. There is a tree near the grass." + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.367, + 0.475, + 0.373 + ], + "angle": 0, + "content": "3. The hillside is covered in" + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.373, + 0.501, + 0.38 + ], + "angle": 0, + "content": "4. A field next to a big tree is open." + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.38, + 0.5, + 0.387 + ], + "angle": 0, + "content": "5. The grass under the trees is lush" + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.387, + 0.547, + 0.394 + ], + "angle": 0, + "content": "6. A large-elephant standing in a field near a tree." + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.394, + 0.552, + 0.401 + ], + "angle": 0, + "content": "7. Two elephants walking under a tree in the sunset" + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.401, + 0.553, + 0.408 + ], + "angle": 0, + "content": "8. Two elephants standing in the grass next to a tree" + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.408, + 0.563, + 0.416 + ], + "angle": 0, + "content": "9. Two elephants standing next to a tree and a blue sky." + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.416, + 0.571, + 0.423 + ], + "angle": 0, + "content": "10. A big elephant standing next to a tree in a field of grass" + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.423, + 0.587, + 0.429 + ], + "angle": 0, + "content": "11. An elephant standing next to a tree with a blue sky behind it." + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.43, + 0.59, + 0.436 + ], + "angle": 0, + "content": "12. Two elephants standing in the water near a and a blue sky" + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.436, + 0.59, + 0.443 + ], + "angle": 0, + "content": "12. Two elephants standing in the grass near a tree and a blue sky 13. Two elephants standing next to each other on a lush green hill" + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.443, + 0.661, + 0.449 + ], + "angle": 0, + "content": "13. Two elephants standing next to each other on a rush green hillside next to a forest." + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.449, + 0.661, + 0.456 + ], + "angle": 0, + "content": "14. Two elephants standing next to a tree in a field with a bluesky and trees behind them." + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.456, + 0.691, + 0.462 + ], + "angle": 0, + "content": "15. Two elephants standing next to each other on a grass covered field next to a lush green forest." + }, + { + "type": "list", + "bbox": [ + 0.385, + 0.353, + 0.691, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.344, + 0.783, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.405, + 0.781, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.475, + 0.789, + 0.628 + ], + "angle": 0, + "content": "Fig. 1: Existing captioning datasets contain captions that describe the entirety of an image. This is reflected in the narrow distributions of the entities that appear in those captions and the caption lengths (the red-colored histograms). CIC aims to generate diverse descriptions by controllably re-focusing on different spatiosemantic aspects of an image, such as the semantically coherent subsets of image objects. Our proposed CIC-BART-SSA is designed to produce diverse, controlled captions ranging from brief and concise to detailed and comprehensive. Sentences 1-15 are example outputs of our approach where the highlighted text indicates the focus of a controllable caption. The histograms demonstrate that our approach generates high-quality descriptions for a wider range of scene focus (number of visual entities) and caption length compared to the original captions. Image is licensed under CC BY-SA 2.0." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.65, + 0.744, + 0.804 + ], + "angle": 0, + "content": "Abstract. Controllable Image Captioning (CIC) aims at generating natural language descriptions for an image, conditioned on information provided by end users, e.g., regions, entities or events of interest. However, available image-language datasets mainly contain captions that describe the entirety of an image, making them ineffective for training CIC models that can potentially attend to any subset of regions or relationships. To tackle this challenge, we propose a novel, fully automatic method to sample additional focused and visually grounded captions using a unified structured semantic representation built on top of the existing set of captions associated with an image. We leverage Abstract Meaning Representation (AMR), a cross-lingual graph-based semantic formal" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.812, + 0.677, + 0.826 + ], + "angle": 0, + "content": "*Work done during an internship at Samsung AI Centre - Toronto" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.576, + 0.84 + ], + "angle": 0, + "content": "†Work done while at Samsung AI Centre - Toronto" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "K. Basioti et al." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.148, + 0.744, + 0.33 + ], + "angle": 0, + "content": "ism, to encode all possible spatio-semantic relations between entities, beyond the typical spatial-relations-only focus of current methods. We use this Structured Semantic Augmentation (SSA) framework to augment existing image-caption datasets with the grounded controlled captions, increasing their spatial and semantic diversity and focal coverage. We then develop a new model, CIC-BART-SSA, specifically tailored for the CIC task, that sources its control signals from SSA-diversified datasets. We empirically show that, compared to SOTA CIC models, CIC-BART-SSA generates captions that are superior in diversity and text quality, are competitive in controllability, and, importantly, minimize the gap between broad and highly focused controlled captioning performance by efficiently generalizing to the challenging highly focused scenarios. Code is available at https://github.com/SamsungLabs/CIC-BART-SSA." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.355, + 0.377, + 0.372 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.788, + 0.508 + ], + "angle": 0, + "content": "Image captioning refers to the task of providing an AI system with an input image, and asking the system to describe the visual content in natural language. This process requires the captioning system to understand what objects are present, in what context (e.g., event or scene), and how they relate. Recent deep learning approaches to this task [14,26,29,31,37,39,40,49,50,57] surpass human performance in standard image captioning metrics. However, these models tend to generate general captions that describe the entirety of an image, and are often of limited diversity; see Original Captions in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.789, + 0.613 + ], + "angle": 0, + "content": "Controllable image captioning (CIC) overcomes these challenges by generating different descriptions for the same image in a user-controlled fashion. That is, a CIC model receives as input an image paired with a user-specified control signal (e.g., entities or regions of interest), and generates a caption conditioned on the control signal. CIC models are thus capable of generating a diverse set of captions by varying the control signal for the same image; see CIC generated captions 1-15 in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.795 + ], + "angle": 0, + "content": "In realistic applications, the easiest way for the user to control the generation of captions is to limit the focus of the desired captions by selecting different entities (objects) using their bounding boxes, as shown in Figs. 1 and 2. Most previous work focuses on such spatial control signals [18, 23, 28, 47, 55, 56]. To improve performance, more recent studies supplement this spatial signal with additional information on the desired length, style, or syntactic and semantic structure of the generated text [12, 13], increasing the richness and complexity of control signals. However, for the CIC approach to succeed, the CIC models need to be trained on equally rich datasets that incorporate, explicitly or implicitly, those control signals. Unfortunately, most image captioning datasets today, such as Flickr30k [36] or MS-COCO [18], lack this necessary diversity of controls and corresponding captions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Our goal is to achieve SOTA performance in CIC without the need for new, increasingly rich, yet also costly, and impractical-to-collect datasets, where human workers would face the burden of having to provide multitudes of control" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.621, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CIC-BART-SSA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.144, + 0.789, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.351, + 0.788, + 0.422 + ], + "angle": 0, + "content": "Fig. 2: An example of our Structured Semantic Augmentation (SSA) approach. Visually-grounded captions (1)-(5) are used to create a meta-vgAMR graph, which includes all available image information in one representation. Sub-graphs of meta-vgAMR are then sampled to generate a new and diverse set of captions, such as the sentences (a)-(e). Image is licensed under CC BY-SA 2.0." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.789, + 0.844 + ], + "angle": 0, + "content": "signals and corresponding descriptive captions. To achieve this goal, we propose a novel Structured Semantic Augmentation (SSA) method, which automatically generates an augmented set of captions and the corresponding control signals with diverse spatiosemantic focus starting from only the core set of \"original\" uncontrolled captions. The method takes advantage of a detailed visual-linguistic semantic graph (illustrated in Fig. 2) constructed from the original captions and their image groundings. To build these semantic graphs, we use Abstract Meaning Representation (AMR) [6], a semantic formalism that can capture fine-grained linguistic relations beyond the exclusively spatial relationships present in the common scene graphs [24]. The availability of robust AMR parsers [5,9] allows us to generate semantic graphs for individual captions automatically, which we then merge into a rich meta-AMR graph for the joint image-language pair. From this meta-graph, we sample diverse connected subgraphs that represent semantically coherent combinations of image-anchored entities, events, and their relations, which we then turn into controlled captions automatically via existing AMR-to-text models [9]. Fig. 2 depicts an example of our meta-graph inferred from the original uncontrolled captions associated with an image. Filled nodes in the meta-graph indicate image entity groundings. Five semantically coherent subgraphs (a)-(e) of variable complexity are then sampled from the meta-graph, which are subsequently used to generate novel captions, shown below each subgraph. These new captions augment the original caption set by providing both image focus, through node groundings, and increased semantic diversity induced by the sampled subgraphs. Building upon SSA, we introduce a new CIC model, CIC-BART, suitable for generating focused controlled captions. Alongside the regions of interest, CIC-BART also makes use of the length of the desired cap" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "K. Basioti et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "tion as a control signal proxy for the verosity of the caption. CIC-BART can be trained on SSA-augmented versions of standard VL datasets such as MS-COCO or Flicker30k to accommodate the CIC task. Our experiments show that, compared to several SOTA models, the captions generated by our model have superior text quality and diversity, while being comparable in terms of faithfulness to control signals." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.238, + 0.472, + 0.252 + ], + "angle": 0, + "content": "In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.329 + ], + "angle": 0, + "content": "1. We propose a novel data augmentation technique, SSA, that draws on a structured semantic formalism (AMR) to automatically generate focused captions suitable for training of CIC models. We empirically show that our SSA technique enables CIC models to generate captions with high controllability, diversity, and text quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.788, + 0.404 + ], + "angle": 0, + "content": "2. We propose CIC-BART, a model designed for CIC, that does not require overly descriptive and complex control signals that SOTA models often require to achieve high performance. We show a superior overall performance, compared to SOTA, while relying on simple control signals (i.e., regions of interest and preferred caption length)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.788, + 0.495 + ], + "angle": 0, + "content": "3. We present an extensive evaluation of our model, compared with existing SOTA. Specifically, we report results on different aspects of generated captions, including controllability (faithfulness to control signal), diversity, and text quality (linguistic well-formedness). To account for the trade-off among these metrics, we propose an overall performance score based on their harmonic mean. This metric helps us identify models that perform well in all these aspects." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.519, + 0.388, + 0.536 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Controllable Image Captioning (CIC). Various types of control have been used for CIC, including visual entities, a type of region-based control [18, 23, 28, 47, 55, 56], where generated captions should learn to focus on the regions of interest. Others draw on complex control signals where additional knowledge about the generated caption structure is provided. For example, some recent work provides the complete skeleton of the desired sentence in the form of a number of objects or attributes or object-relation-object templates [12, 13]. Additional control signals that CIC draws on include different caption styles, e.g., positive, negative, humorous, or romantic tone [20, 21, 32, 33, 45, 47, 53, 54], user personality [17, 41], or the length of the generated captions [19, 22, 45, 47, 48]. The use of complex control signals aims at improving the diversity of captions and the quality of the text in CIC models. However, it requires the users to provide a detailed description of the control signal, which is not realistic in practical settings where such models are to be deployed (e.g., a self-driving car or personal assistant). We instead draw on two simple control signals (regions of interest and desired caption length) and show that we can achieve competitive performance on CIC, while keeping the control signals simple and practical." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Recent SOTA models that draw on spatial control include the SCT model [18] that also uses the Faster R-CNN feature vectors and object tags (corresponding" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.621, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "CIC-BART-SSA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.146, + 0.786, + 0.403 + ], + "angle": 0, + "content": "GloVe vectors [35]) of the entities of interest, as well as models that include skeleton-based control, namely ASG2Caption [13] and VSR [12]. ASG2Caption uses an abstract scene graph (ASG) to express the desired structure of a caption. ASG contains three types of unlabeled abstract nodes (object, attribute, relationship) that are grounded in the image by extracting features from the corresponding bounding boxes (for objects and attributes) or from the union of bounding box pairs (for a relationship node). ASG2Caption shows improved controllability (by conditioning on ASGs), and diversity (by automatically sampling diverse ASGs as control signals). The VSR model [12] draws on GloVe embeddings of Faster R-CNN object tags for visual entities (as in SCT). It also uses a skeleton control signal (like ASG2Caption), but one that includes more detailed information and richer semantics. Specifically, the VSR control signal follows the form of a fine-grained PropBank entry\\(^4\\) — i.e., specifying the exact verb(s) expressing action(s) depicted in the image, and their visually grounded arguments (e.g., subject, object, location, manner). Thus, VSR uses the most descriptive control signal among the SOTA models. Refer to our supplementary material for an illustrative example of the control signal used for each method." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.405, + 0.786, + 0.585 + ], + "angle": 0, + "content": "Compared to ASG2Caption and VSR, our control signal is kept minimal and only specifies the bounding boxes and desired caption lengths. To improve the diversity of captions, we draw on a structured semantic graph (AMR) that expresses the semantics of a sentence based on PropBank semantics. Notably, we do not use these rich graphs to express detailed and overly descriptive control signals (as in VSR), but we use these semantic structures to augment our training data with richer and diverse captions, which will result in the model learning to generate more diverse captions. Additionally, we include a length control signal to further increase diversity without needing to specify detailed information about the structure of the output (e.g., number of attributes per object, etc.). This way, we can generate a variety of captions for a fixed image sub-region by simply controlling the desired length of the output." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.586, + 0.786, + 0.72 + ], + "angle": 0, + "content": "Abstract Meaning Representation (AMR). AMR [7] is a rich semantic formalism for expressing the meaning of natural language sentences as a formal graph. AMR draws on PropBank, which is a rich lexical semantic resource encoding predicates expressing an action or state, as well as the number and nature of the participating entities (arguments and other semantic roles, such as location, manner, etc.). AMR is a widely researched semantic formalism for which highly accurate automatic Text-to-AMR and AMR-to-Text models are developed [5,9]. We rely on these models to augment original image-caption datasets with newly generated captions (as explained in Sect. 4)." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.722, + 0.786, + 0.812 + ], + "angle": 0, + "content": "AMRs vs. Scene Graphs. Recent studies [1,15,16,52] have shown that AMRs better capture the semantic relations of an image as compared to the scene graphs [11]. Existing scene graph annotations mainly capture geometric or possessive relations, which account for more than \\(90\\%\\) of the relations captured, whereas more than \\(1/3\\) of the captured entities refer to clothing, object, or body parts information [1,52]. This difference is crucial for high-quality image caption-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.825, + 0.436, + 0.84 + ], + "angle": 0, + "content": "4 https://propbank.github.io" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.128 + ], + "angle": 0, + "content": "K. Basioti et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "ing, as we use higher-level semantic relations in our everyday language rather than geometric ones. For instance, during a soccer game, we would probably describe a goal save as 'the player kicks the ball away from the goal' or 'the goalkeeper defends his team by saving a goal' and not by using mainly geometric and possessive relations like 'a person wearing a white shirt, standing with his right leg lifted, close to a ball which is above the ground'. In the supplementary, we provide a detailed comparison of AMR and scene graph representations, particularly focusing on their applications in data augmentation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.289, + 0.317, + 0.306 + ], + "angle": 0, + "content": "3 Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.32, + 0.488, + 0.516 + ], + "angle": 0, + "content": "We propose CIC-BART, specifically designed to generate controlled image captions. Specifically, it can generate descriptions of particular areas within a scene with a desired level of detail. Our model, based on VL-BART [14], utilizes a transformer-based encoder-decoder architecture, as shown in Fig. 3. CIC-BART extends VL-BART encoder to the CIC task by modifying the encoder input to include: a) a global image embedding that provides the context of the" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.364, + 0.782, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.455, + 0.788, + 0.484 + ], + "angle": 0, + "content": "Fig.3: The architecture diagram of our model, CIC-BART." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.517, + 0.788, + 0.591 + ], + "angle": 0, + "content": "full image to the model; b) the visual control signal, including the visual embedding of the regions that contain the entities of interest; c) the text control signal, containing length control (indicating the desired length range of the output caption) and an optional verb signal that indicates the action we want the generated caption to concentrate on." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.788, + 0.743 + ], + "angle": 0, + "content": "The visual embeddings of the regions are position-aware embeddings from a Faster R-CNN model [38] trained for visual object and attribute classification [3] on Visual Genome. The global image feature vector is extracted as well from Faster R-CNN. For the length control signal, we add to our vocabulary \\(L\\) tokens for the \\(L\\) different caption length levels; for instance, level one represents sentences between one and nine words, and level two, ten to nineteen. These tokens describe our coarse levels, for a finer sentence size accuracy, we accompany the tokens with the desired number of words. This choice gives our model the capacity to generate diverse captions for a particular length level. Finally, the output of the decoder generates the desired, controlled image caption." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.764, + 0.661, + 0.783 + ], + "angle": 0, + "content": "4 Structured Semantic Augmentation (SSA)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "The goal of our SSA method is to augment existing image captioning datasets with new focused captions along with their control signals (i.e., regions corresponding to entities). We rely on datasets where visual entities in the captions" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.621, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "CIC-BART-SSA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.207 + ], + "angle": 0, + "content": "are annotated with their corresponding regions (see Sect. 5 for details on the datasets). The SSA process consists of four main steps, as described below. For more details, refer to our supplementary material, which also includes a step-by-step example of our SSA methodology." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.787, + 0.418 + ], + "angle": 0, + "content": "Step 1: Image-level AMR graph generation. Our objective in this stage is to enclose all the information available from the visually grounded captions into a single representation. To accomplish this, we create a visually grounded AMR graph (vgAMR) for each caption of an image and then merge them into a single image-level graph, the meta-vgAMR. To create the vgAMRs of an image, we first convert each of its \\( N \\) captions to their AMR representation, using the Neural transition-based Text-to-AMR parser [5] which also aligns words in a caption with their respective nodes in the AMR graph. We utilize the alignment information of 1) caption words and AMR nodes (from Text-to-AMR parser) and 2) caption words to image bounding boxes (from existing dataset annotations) to visually ground the AMR nodes. After this step, we get the collection of nodes referring to visual entities, where each grounded meta-AMR node is linked with a non-empty set of bounding boxes. This extended representation, 'AMR + visual grounded nodes', is our vgaMR." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.787, + 0.524 + ], + "angle": 0, + "content": "Our next step is to combine the N vgAMRs to form a single meta-vgAMR. To achieve this, we employ a pairwise strategy to merge the most similar vgAMRs first (we measure similarity with Smatch score [10]). We use the UPGMA hierarchical clustering algorithm [30,34] to find the optimal merge ordering starting from the most similar graphs. UPGMA creates a hierarchy where the bottom level consists of the N individual vgAMRs. By merging all vgAMRs using the UPGMA ordering, we obtain a single structure called meta-vgAMR." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.525, + 0.787, + 0.797 + ], + "angle": 0, + "content": "When merging two vgAMRs, the main challenge is identifying which nodes correspond to the same concepts, such as entities, attributes, actions, and relations. We use three node properties to accomplish this: a) visual grounding information, b) semantic similarity of node labels, and c) node neighborhood semantic similarity. We derive two node-merging criteria from there: 1) visually grounded entity nodes are merged if they point to the same image-bounding boxes. When 1) does not hold, we check the second criterion: 2) for the remaining non-grounded nodes, including amr-specific, predicates, adjectives, and adverbs, we use a combination of node label semantic similarity (cosine similarity of the labels using their GloVe embeddings) and neighborhood similarity. Neighborhood similarity examines the similarity of parents for adjectives/adverbs nodes and children for predicate nodes, along with the similarity of connecting edge roles. When two nodes satisfy criterion 1) or 2), we merge them into a single node. Moreover, if they have different labels, we maintain both names by keeping a list of synonyms to increase representation diversity. In the special case when the two vgAMRs describe two totally different concepts, and hence they have no common nodes, we add an amr-specific node called 'multi-sentence' as the root with the two independent vgAMRs as its children. The final graph, meta" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "K. Basioti et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.787, + 0.192 + ], + "angle": 0, + "content": "vgAMR, includes all non-redundant5 elements of the original \\(N\\) captions while preserving the visual grounding between the meta nodes and their respective image regions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.787, + 0.237 + ], + "angle": 0, + "content": "Remark: Meta-vgAMR efficiently compresses all available image information into a single structure. Following our approach, we can easily scale when new scene information becomes available by applying our pairwise merge procedure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.789, + 0.539 + ], + "angle": 0, + "content": "Step 2: Event-based graph sampling from image-level AMRs. We start from the predicate nodes, which mainly correspond to verbs, to sample subgraphs in meta-vgAMR graphs. Predicate nodes are identified by their label and the edges connected to them. The label of a predicate node typically follows the format 'predicate_name-xx,' where 'xx' represents the different senses a word can have regarding the concept it is used for. Predicate nodes have outward ARGy edges, where 'y' can take values from 0 to 5, connecting them to their arguments. We sample subgraphs from these nodes by following the outgoing argument edges, which are labeled as ARGn in an AMR graph, each defining a particular semantic role (e.g., ARG0 points to the agent, ARG1 to the patient, etc.). Finally, we add one more subgraph containing the remaining children branches of other non-ARG optional predicate edges (e.g., 'location', 'time'). We repeat this process until the leaves of the graphs are reached. During sampling, we randomly select one of the synonyms if a node is a list of synonym labels, as mentioned in the previous step. The output of this step is our more focused event-focused sub-graphs. In Fig. 2, we can see some instances of our event-based sampling (SSA samples), where the predicate nodes include z0/sit-01, z7/dock-01, z13/sit-01, and so on6. Although we cannot show all the sampled event-based subgraphs in the figure, we included five of them and used colored roots and edges for visualization purposes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.79, + 0.63 + ], + "angle": 0, + "content": "Step 3: New caption generation from sampled AMRs We use the SPRING AMR-to-Text model [8] to generate new event-focused captions from the sampled vgAMR subgraphs. Because both vgAMR merging and sampling steps introduce noise, the output captions are not always of good quality. We automatically filter low-quality captions by using a linguistic well-formedness measure, GRUEN [58], which is a reference-free metric based on BERT contextual embeddings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.787, + 0.691 + ], + "angle": 0, + "content": "Step 4: Control signal generation. The last step is to create the control signal for the generated captions. The spatial control signal for a specific caption is extracted from the corresponding sampled vgAMR, by pulling the bounding boxes of the visual entity linked AMR nodes." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.701, + 0.787, + 0.77 + ], + "angle": 0, + "content": "5 A node may have different names for the same bounding box in different meta-vgAMRs, such as 'A male' and 'A person'. According to criterion 1), we merge the corresponding AMR nodes and keep both 'male' and 'person' in the names list to avoid redundancy. Therefore, criteria 1) and 2) ensure that multiple nodes don't describe the same concept in the meta-vgAMR." + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.771, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Note that in Fig. 2, the node z17/green-03 is also categorized as a predicate. This may seem an error because we usually think of 'green' as an attribute node rather than a predicate. However, in AMRs, when 'green' is paired with its argument, in this case, z16/lawn, it encapsulates a predicate/verb that can be expressed in natural language as 'the lawn is green.'" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.701, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.621, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CIC-BART-SSA" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.628, + 0.163 + ], + "angle": 0, + "content": "4.1 Mixing Strategies of Original and SSA Data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.326 + ], + "angle": 0, + "content": "To analyze the impact of our SSA data, we explore various mixing strategies with the original training set. Assume \\(\\mathcal{D}\\) represents the training control-caption pairs in the original dataset, containing \\(N_{\\mathcal{D}}\\) samples, and SSA represents our SSA samples, containing \\(N_{SSA}\\) instances. The augmented dataset \\(\\mathcal{D}_{SSA}\\) is defined by combining \\(\\mathcal{D}\\) and SSA: \\(\\mathcal{D}_{SSA} = \\mathrm{sam}_{\\mathcal{D}}(\\tau_{\\mathcal{D}}, p_{\\mathcal{D}}) \\cup \\mathrm{sam}_{SSA}(\\tau_{SSA}, p_{SSA})\\), where the functions \\(\\mathrm{sam}_{\\mathcal{D}}\\) samples a subset of the original dataset, and \\(\\mathrm{sam}_{SSA}\\) a subset of our SSA data. Since we are interested in the effect of our SSA, we assume that \\(\\mathrm{sam}_{\\mathcal{D}}(\\tau_{\\mathcal{D}}, p_{\\mathcal{D}}) = \\mathcal{D}\\), with \\(\\tau_{\\mathcal{D}} = \\text{'Random Sampling Strategy'}\\) and \\(p_{\\mathcal{D}} = 100\\%\\), meaning that all original data are included in the mixed dataset. Depending on the \\(\\mathrm{sam}_{SSA}\\) parameter \\(\\tau_{SSA}\\) we have the cases:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.788, + 0.386 + ], + "angle": 0, + "content": "Random Sampling Strategy. In this case, we randomly select a pre-specified number of examples from \\( SSA \\). The parameter \\( p_{SSA} \\) expresses the percentage of SSA samples included in \\( \\mathcal{D}_{SSA} \\). With boundary cases \\( p_{SSA} = 100\\% \\) (all \\( N_{SSA} \\) samples are included), and \\( p_{SSA} = 0\\% \\) (no SSA data are added)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.788, + 0.524 + ], + "angle": 0, + "content": "Uniform-Coverage Sampling Strategy. To mitigate the original dataset's bias (having mainly samples describing the entire image), we aim to create a new focus-unbiased dataset. By modeling the control signal focus as the image area percentage covered by the bounding boxes of the control signal, we split the original data into \\( B \\) coverage bins. Then, we will randomly add in each bin SSA samples, aiming to create a new uniform, coverage-unbiased \\( \\mathcal{D}_{SSA} \\) dataset. Here, \\( p_{SSA} \\) contains the range of each bin for the coverage histogram. For example, in the case where we choose ten uniform coverage bins, we have \\( p_{SSA} = \\{[0\\%, 10\\%), [10\\%, 20\\%), \\ldots, [90\\%, 100\\%]\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.569 + ], + "angle": 0, + "content": "Due to space limitations, we present results from the Random Sampling Strategy for \\( p_{SSA} = 0\\% \\) and \\( p_{SSA} = 100\\% \\) in the main paper. Results from other scenarios can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.596, + 0.446, + 0.613 + ], + "angle": 0, + "content": "5 Experimental Setup" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.631, + 0.307, + 0.645 + ], + "angle": 0, + "content": "5.1 Data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We use Flickr30k Entities (Flickr-Ent) [36] and MS-COCO Entities (COCO-Ent) [18] for training and evaluation. Flickr-Ent augments the original captions of Flickr30k [51] with manually-annotated region-phrase groundings. Flickr-Ent contains the original \\(31K\\) images annotated with five captions each. COCO-Ent augments the original MS-COCO [27] (\\(120K\\) images each annotated with around five captions) with semi-automatically collected grounding annotations; see [18] for details on the annotation process. For both datasets, we follow previous work and use the training and test splits by Karpathy et al. [25]. We apply our SSA algorithm on the aforementioned datasets to create their augmented variations, COCO-Ent-SSA and Flickr-Ent-SSA, containing about \\(800K\\) and \\(250K\\) training captions, respectively, of which \\(33\\%\\) and \\(37\\%\\), are generated by our SSA algorithm." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "K. Basioti et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "For all four training sets, we automatically generate image-control-caption triplets to train our model on. For spatial control, we extract from the grounded captions the bounding boxes of the entities of interest using the annotations from COCO-Ent and Flickr-Ent. Note that since these datasets do not contain the text control signal, we use each ground-truth caption as a proxy for a controlled caption, from which we first generate the coarse- and fine-length control levels and then extract their verbs using part-of-speech tagging for the optional action control." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.289, + 0.522, + 0.304 + ], + "angle": 0, + "content": "5.2 Models and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.787, + 0.447 + ], + "angle": 0, + "content": "We compare two variations of our model (with and without SSA augmentations) with SOTA models as our baselines: Show Control & Tell (SCT) [18] that uses region-based control (bounding boxes of visual entities of interest); ASG2Caption (ASG) [13] that draws on visually grounded abstract scene graphs as control signal; and VSR [12] that uses overly descriptive control signals that express verb(s) and fine-grained verb-specific semantic roles of the desired captions; ComPro [48] that learns a mapping from the bounding boxes of the entities of interest and caption length level to GPT-2 Large prompts aiming to retrieve controlled captions; and the LaBERT length-control-only model [19]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.449, + 0.789, + 0.844 + ], + "angle": 0, + "content": "We report the performance of our models and baselines using a comprehensive set of metrics that evaluate different aspects of caption controllability, diversity, and quality. We also propose and report an overall performance metric that summarizes these different aspects in a meaningful way. To measure diversity, we compute n-gram diversity, D-n for \\(n = 1\\), 2 [4], as well as self-CIDEr-based diversity (sC) [46]. For a fair comparison of the different CIC models, we measure diversity for the five generated captions for each test image (in COCO-Ent and Flickr-Ent), and report their average. To measure content controllability we design an extended version of the IoU metric of [18] that calculates the degree-of-match (faithfulness) between a control signal and the corresponding generated caption. For our control signal, we use the set of nouns \\(\\mathcal{E}\\) that represent the entities of interest, which are the names of the visual objects in the control. To extract nouns from the predicted sentences, we use the Stanford part-of-speech tagger [42,43]. We then find the semantic intersection of the two sets using Hungarian Matching, as in [18]. Finally, we calculate the semantic intersection over union of the control nouns and the nouns extracted from the controllable caption, which gives us our content controllability IoU. We further analyze IoU by introducing the Hallucinating Nouns (Hal) metric. Details can be found in the supplemental material. For length controllability (L), we measure the Mean Absolute Error (MAE) between the fine length control (number of words) and the size of the resulting \\(M = 10\\) controlled captions, which are generated from \\(M\\) randomly created control signals. We also calculate the length precision (LP) [19] by determining the percentage of generated captions that match the desired coarse length level. We assess text quality of generated captions using GRUEN (G) [58], a reference free metric based on BERT contextual embeddings that measures the syntactic and semantic well-formedness of a text segment. Finally, we measure" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.621, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CIC-BART-SSA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 1: Performance of CIC models for the original test sets. *ASG-type dataset is not released for Flickr-Ent; therefore, we could not reproduce the ASG results." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.185, + 0.78, + 0.273 + ], + "angle": 0, + "content": "
ModelH↑IoU↑G↑sC↑D-1↑D-2↑L↓H↑IoU↑G↑sC↑D-1↑D-2↑L↓
SCT [18]55.867.364.442.827.035.5-54.650.779.844.029.336.5-
ASG* [13]74.272.672.078.337.856.6--------
VSR [12]56.277.639.067.430.042.2-62.560.254.077.933.349.3-
CIC-BART75.976.273.078.738.056.2.4969.854.085.078.643.658.21.24
CIC-BART-SSA78.377.274.882.544.663.2.1171.355.086.081.747.062.61.05
" + }, + { + "type": "table_caption", + "bbox": [ + 0.284, + 0.285, + 0.719, + 0.3 + ], + "angle": 0, + "content": "Table 2: Performance of CIC models, for the SSA-only samples." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.311, + 0.772, + 0.38 + ], + "angle": 0, + "content": "
ModelH↑IoU↑G↑sC↑D-1↑D-2↑H↑IoU↑G↑sC↑D-1↑D-2↑
COCO-Ent (SSA only)Flickr-Ent (SSA only)
SCT [18]51.762.164.837.823.731.043.929.977.345.731.036.7
CIC-BART69.261.473.974.044.257.068.553.080.579.852.962.9
CIC-BART-SSA75.665.280.783.753.867.872.055.682.986.156.569.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.788, + 0.527 + ], + "angle": 0, + "content": "the overall performance of each model based on its ability to balance content controllability, diversity, and text quality. To calculate this, we use the harmonic mean of IoU, G, and sC. All of these metrics range between 0 and 1, with a higher value indicating better performance. The harmonic mean \\((H)\\) helps us determine the model with the best overall performance. It prioritizes models that perform well across all metrics while penalizing those with poor performance, even in one metric. In our supplementary, we include comparisons of the CIC models on standard captioning metrics (like CIDEr [44] and Spice [2])." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.547, + 0.327, + 0.563 + ], + "angle": 0, + "content": "6 Results" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.576, + 0.434, + 0.59 + ], + "angle": 0, + "content": "6.1 Overall Performance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.659 + ], + "angle": 0, + "content": "We first compare the overall performance of our models with the SCT, ASG, and VSR baselines with respect to controllable captioning metrics. We do not include ComPro in this comparison due to the unavailability of the codebase. We also exclude LaBERT since this model solely focuses on length controllability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Tab. 1 presents results for content and length controllability (IoU and L), text quality (G), and diversity (sC, D-1, D-2), as well as the harmonic mean (H). For both datasets, CIC-BART-SSA has the best performance in all metrics, except IoU, where it is the second best. Specifically, CIC-BART-SSA is superior to all other models with respect to diversity (sC, D-1, D-2) and text quality (G), but comparable to VSR in terms of content controllability (IoU). The length controllability (L) scores show that our SSA augmentation helps the model learn to generate high-quality output at the desirable length (compare CIC-BART and CIC-BART-SSA). This is due to the increased diversity in caption length provided by our SSA augmentations." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.811, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Importantly, we can see that model performance can vary depending on the metric. E.g., whereas VSR has the highest IoU, it falls behind in text quality and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.128 + ], + "angle": 0, + "content": "K. Basioti et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "diversity. In our qualitative analysis, we observe the poor quality of the captions generated by VSR. The best-performing model should be identified based on the \\( H \\) score that summarizes content controllability, text quality, and diversity into a single score. Based on this score, CIC-BART-SSA is better than SCT and VSR baselines by a large margin, and notably better than ASG. Nevertheless, ASG requires complex control signals in the form of scene graphs, in contrast to the simple control signal requirements of CIC-BART." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.434 + ], + "angle": 0, + "content": "Next, we conduct a further evaluation of the CIC performance on our SSA samples from the test set images of COCO-Ent and Flickr-Ent. We present the results in Tab. 2 for our models CIC-BART, CIC-BART-SSA, and SCT. ASG and VSR are excluded since they need complex control signals (grounded abstract scene graphs for ASG and grounded verb semantic roles for VSR), which are only available for COCO-Ent and Flickr-Ent. We observe a significant improvement in overall performance for our CIC-BART-SSA model. We notice that the models (SCT and CIC-BART) trained on the original datasets, which describe the entire image, had difficulties generalizing to cases where they had to focus on a specific sub-region of an image. However, our model CIC-BART-SSA was able to generate focused and diverse descriptions of the challenging, highly focused examples present in our SSA data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.54, + 0.63 + ], + "angle": 0, + "content": "We compare the length precision of our model with the baselines utilizing length control in Table 3. LaBERT uses only length-control signals without spatial control, while our model employs both spatial and length-control signals to generate focused captions. This makes the LaBERT task much easier since it only focuses on generating specific length descriptions of an image. On the other hand, our model focuses on generating captions that describe only a specific sub-region of the scene while maintaining a desired description length level. Although our task is" + }, + { + "type": "table_caption", + "bbox": [ + 0.55, + 0.478, + 0.788, + 0.52 + ], + "angle": 0, + "content": "Table 3: Length Precision (LP) for CIC models on COCO-Ent and Flickr-Ent original test sets." + }, + { + "type": "table", + "bbox": [ + 0.552, + 0.52, + 0.785, + 0.597 + ], + "angle": 0, + "content": "
ModelLP↑LP↑
COCO-EntFlickr-Ent
ComPro [48]94.781.4
LaBERT [19]99.798.4
CIC-BART99.988.0
CIC-BART-SSA99.991.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.631, + 0.788, + 0.737 + ], + "angle": 0, + "content": "more challenging than LaBERT, we achieve competitive length precision performance. Lastly, we want to emphasize that the remarkable improvement in length controllability (L) and length precision (LP) from CIC-BART to CIC-BART-SSA stems from the increased length diversity found in our SSA augmentations, which enriches the original COCO-Ent and Flickr-Ent datasets. In the supplementary, we provide an analysis of the caption length statistics in the original datasets and our SSA-derived captions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.756, + 0.598, + 0.772 + ], + "angle": 0, + "content": "6.2 Effect of SSA on Content Controllability" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To analyze the impact of our SSA augmentations, we measure the content controllability (IoU) performance of CIC-BART at different levels of focus of the control signals and report it in Fig. 4. We use coverage, defined as the area of the image enclosed by the bounding boxes of the entities of interest in the control" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.621, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CIC-BART-SSA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.144, + 0.515, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.143, + 0.787, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.244, + 0.516, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.244, + 0.787, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.331, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Fig. 4: IoU histograms for CIC-BART (green) and CIC-BART-SSA (blue). The first column depicts the IoU on the original test set and the second on the original test set images' SSA (only) data. The Samples curve (orange) represents the distribution of test captions in each coverage (image area covered by the control signal) interval." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.419, + 0.461, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.464, + 0.42, + 0.787, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.567, + 0.787, + 0.595 + ], + "angle": 0, + "content": "Fig. 5: Qualitative examples for the original test sets. Strikethrough marks hallucinations and redundancies. a, b licensed under CC BY-SA 2.0; c, d under CC BY 2.0." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "signal, to quantify that focus. For example, highly focused control signals cover a small area, yielding low coverage, while broader signals cover a larger area and have high coverage. We 'break down' the IoU performance into 10 coverage bands and report the average IoU over control signals in those bands. In addition, the 'Samples' curve shows the distribution of test captions over the same bands. The results in Fig. 4 indicate that by training with SSA (blue bars), the spatial controllability improves significantly in the low-coverage regime, where the control signals are highly focused. Interestingly, these are also the most underrepresented (data deprived) parts of the original dataset Flickr-Ent. Therefore, SSA which enriches the original datasets with highly focused examples (please refer to our supplementary for % Samples per coverage bands for the training sets), is effective in improving generalization performance in CIC. We include the performance of hallucinating nouns metric in the supplementary material." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.128 + ], + "angle": 0, + "content": "K. Basioti et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.43, + 0.163 + ], + "angle": 0, + "content": "6.3 Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.489, + 0.476 + ], + "angle": 0, + "content": "In Fig. 5, we present qualitative examples from the original test sets, and in Fig. 6 examples from our SSA (only) test set control signals. In the two figures, each highlighted word found in the generated controlled captions corresponds to the control entity of the same color. This shows the match between the captions produced and the control signal. We also strike through the parts where the model hallucinates or generates redundant references to the entities of interest. Our models have been observed to outperform the previous state-of-the-art models by substantially enhancing the quality of the generated controlled captions. This behavior was expected from our quantitative analysis, which showed that our models have signifi" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.789, + 0.539 + ], + "angle": 0, + "content": "cantly higher text quality (G). More importantly, our CIC-BART-SSA model is capable of generating captions that are faithful to the control signal and better understand the relationships that connect the entities of interest. We include additional qualitative samples in the Supplementary Material." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.219, + 0.782, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.389, + 0.789, + 0.446 + ], + "angle": 0, + "content": "Fig. 6: Examples for SSA-only test set. Strikethrough marks hallucinations. Up to Down: a licensed under CC BY 2.0; b under CC PDM 1.0, for last two see Fig. 5." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.564, + 0.37, + 0.58 + ], + "angle": 0, + "content": "7 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.79, + 0.844 + ], + "angle": 0, + "content": "We address two main challenges faced by the controllable image captioning (CIC) models. First, standard image-caption datasets lack the controllability and diversity needed for proper training and evaluation of CIC. Second, most recent SOTA models require complex and overly descriptive control signals as input (including, e.g., the main action/verb to appear in the generated caption). To address the first challenge, we propose a novel technique that draws on a structured semantic augmentation (SSA) formalism to generate focused captions and the corresponding control signals for images. For the second challenge, we propose a transformer-based vision-language model attuned to the CIC task. We show that this model performs competitively with SOTA models without requiring complex and explicit control signals. Importantly, when combined with our SSA approach, our model generates highly diverse captions and significantly reduces the content controllability performance gap between the different levels of focus of the generated controlled captions. Finally, when provided with the commonly used verb guidance of other SOTA approaches, our model shows a substantial improvement in performance." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.621, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CIC-BART-SSA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.18, + 0.786, + 0.236 + ], + "angle": 0, + "content": "1. Abdelsalam, M.A., Shi, Z., Fancellu, F., Basioti, K., Bhatt, D., Pavlovic, V., Fazly, A.: Visual semantic parsing: From images to abstract meaning representation. In: Proceedings of the 26th Conference on Computational Natural Language Learning (CoNLL). pp. 282-300 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.237, + 0.787, + 0.292 + ], + "angle": 0, + "content": "2. Anderson, P., Fernando, B., Johnson, M., Gould, S.: Spice: Semantic propositional image caption evaluation. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 382-398. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.293, + 0.787, + 0.334 + ], + "angle": 0, + "content": "3. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.334, + 0.787, + 0.377 + ], + "angle": 0, + "content": "4. Aneja, J., Agrawal, H., Batra, D., Schwing, A.: Sequential latent spaces for modeling the intention during diverse image captioning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4261-4270 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.787, + 0.405 + ], + "angle": 0, + "content": "5. Astudillo, R.F., Ballesteros, M., Naseem, T., Blodgett, A., Florian, R.: Transition-based parsing with stack-transformers. arXiv preprint arXiv:2010.10669 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.787, + 0.475 + ], + "angle": 0, + "content": "6. Banarescu, L., Bonial, C., Cai, S., Georgescu, M., Griffitht, K., Hermjakob, U., Knight, K., Koehn, P., Palmer, M., Schneider, N.: Abstract meaning representation (amr) 1.0 specification. In: Parsing on Freebase from Question-Answer Pairs. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing. Seattle: ACL. pp. 1533–1544 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.476, + 0.787, + 0.532 + ], + "angle": 0, + "content": "7. Banarescu, L., Bonial, C., Cai, S., Georgescu, M., Griffitht, K., Hermjakob, U., Knight, K., Koehn, P., Palmer, M., Schneider, N.: Abstract meaning representation for sembanking. In: Proceedings of the 7th linguistic annotation workshop and interoperability with discourse. pp. 178-186 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.532, + 0.787, + 0.573 + ], + "angle": 0, + "content": "8. Bevilacqua, M., Błoshmi, R., Navigli, R.: One spring to rule them both: Symmetric amr semantic parsing and generation without a complex pipeline. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 12564-12573 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.574, + 0.787, + 0.629 + ], + "angle": 0, + "content": "9. Bloshmi, R., Bevilacqua, M., Fabiano, E., Caruso, V., Navigli, R.: Spring goes online: end-to-end AMR parsing and generation. In: Proceedings of the 2021 conference on empirical methods in natural language processing: system demonstrations. pp. 134-142 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.63, + 0.787, + 0.672 + ], + "angle": 0, + "content": "10. Cai, S., Knight, K.: Smatch: an evaluation metric for semantic feature structures. In: Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). pp. 748-752 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.672, + 0.787, + 0.714 + ], + "angle": 0, + "content": "1. Chang, X., Ren, P., Xu, P., Li, Z., Chen, X., Hauptmann, A.: A comprehensive survey of scene graphs: Generation and application. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(1), 1-26 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.714, + 0.787, + 0.757 + ], + "angle": 0, + "content": "2. Chen, L., Jiang, Z., Xiao, J., Liu, W.: Human-like controllable image captioning with verb-specific semantic roles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16846-16856 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.787, + 0.799 + ], + "angle": 0, + "content": "3. Chen, S., Jin, Q., Wang, P., Wu, Q.: Say as you wish: Fine-grained control of image caption generation with abstract scene graphs. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9962-9971 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "4. Cho, J., Lei, J., Tan, H., Bansal, M.: Unifying vision-and-language tasks via text generation. In: International Conference on Machine Learning. pp. 1931-1942. PMLR (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.18, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.38, + 0.127 + ], + "angle": 0, + "content": "K. Basioti et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "15. Choi, W.S., Heo, Y.J., Punithan, D., Zhang, B.T.: Scene graph parsing via abstract meaning representation in pre-trained language models. In: NAACL 2022 Workshop on Deep Learning on Graphs for Natural Language Processing (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.217 + ], + "angle": 0, + "content": "16. Choi, W.S., Heo, Y.J., Zhang, B.T.: Sgram: Improving scene graph parsing via abstract meaning representation. arXiv preprint arXiv:2210.08675 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.787, + 0.257 + ], + "angle": 0, + "content": "17. Chunseong Park, C., Kim, B., Kim, G.: Attend to you: Personalized image captioning with context sequence memory networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 895-903 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.257, + 0.787, + 0.298 + ], + "angle": 0, + "content": "18. Cornia, M., Baraldi, L., Cucchiara, R.: Show, control and tell: A framework for generating controllable and grounded captions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8307-8316 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.298, + 0.787, + 0.338 + ], + "angle": 0, + "content": "19. Deng, C., Ding, N., Tan, M., Wu, Q.: Length-controllable image captioning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16. pp. 712-729. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.338, + 0.787, + 0.379 + ], + "angle": 0, + "content": "20. Gan, C., Gan, Z., He, X., Gao, J., Deng, L.: Stylenet: Generating attractive visual captions with styles. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3137-3146 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.787, + 0.42 + ], + "angle": 0, + "content": "21. Guo, L., Liu, J., Yao, P., Li, J., Lu, H.: Mscap: Multi-style image captioning with unpaired stylized text. 2019 IEEE. In: CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4199-4208 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.787, + 0.447 + ], + "angle": 0, + "content": "22. Hirsch, E., Tal, A.: Clid: Controlled-length image descriptions with limited data. arXiv preprint arXiv:2211.14835 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.787, + 0.488 + ], + "angle": 0, + "content": "23. Johnson, J., Karpathy, A., Fei-Fei, L.: Densecap: Fully convolutional localization networks for dense captioning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4565-4574 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.528 + ], + "angle": 0, + "content": "24. Johnson, J., Krishna, R., Stark, M., Li, L.J., Shamma, D., Bernstein, M., Fei-Fei, L.: Image retrieval using scene graphs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3668-3678 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.787, + 0.569 + ], + "angle": 0, + "content": "25. Karpathy, A., Fei-Fei, L.: Deep visual-semantic alignments for generating image descriptions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3128-3137 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.787, + 0.624 + ], + "angle": 0, + "content": "26. Li, X., Yin, X., Li, C., Zhang, P., Hu, X., Zhang, L., Wang, L., Hu, H., Dong, L., Wei, F., et al.: Oscar: Object-semantics aligned pre-training for vision-language tasks. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16. pp. 121-137. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.624, + 0.787, + 0.679 + ], + "angle": 0, + "content": "27. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.679, + 0.787, + 0.705 + ], + "angle": 0, + "content": "28. Lindh, A., Ross, R.J., Kelleher, J.D.: Language-driven region pointer advancement for controllable image captioning. arXiv preprint arXiv:2011.14901 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.787, + 0.732 + ], + "angle": 0, + "content": "29. Lu, J., Yang, J., Batra, D., Parikh, D.: Neural baby talk. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7219-7228 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.787, + 0.758 + ], + "angle": 0, + "content": "30. Lukasova, A.: Hierarchical agglomerative clustering procedure. Pattern Recognition 11(5-6), 365-381 (1979)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.8 + ], + "angle": 0, + "content": "31. Luo, J., Li, Y., Pan, Y., Yao, T., Feng, J., Chao, H., Mei, T.: Semantic-conditional diffusion networks for image captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23359-23368 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.841 + ], + "angle": 0, + "content": "32. Mathews, A., Xie, L., He, X.: Senticap: Generating image descriptions with sentiments. In: Proceedings of the AAAI conference on artificial intelligence. vol. 30 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.621, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CIC-BART-SSA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "33. Mathews, A., Xie, L., He, X.: Semstyle: Learning to generate stylised image captions using unaligned text. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8591-8600 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.217 + ], + "angle": 0, + "content": "34. Müller, D.: Modern hierarchical, agglomerative clustering algorithms. arXiv preprint arXiv:1109.2378 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.217, + 0.788, + 0.257 + ], + "angle": 0, + "content": "35. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP). pp. 1532-1543 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.257, + 0.788, + 0.311 + ], + "angle": 0, + "content": "36. Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In: Proceedings of the IEEE international conference on computer vision. pp. 2641-2649 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.311, + 0.788, + 0.365 + ], + "angle": 0, + "content": "37. Ramos, R., Martins, B., Elliott, D., Kementchedjheva, Y.: Smallcap: lightweight image captioning prompted with retrieval augmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2840-2849 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.405 + ], + "angle": 0, + "content": "38. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems 28 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.405, + 0.788, + 0.461 + ], + "angle": 0, + "content": "39. Ren, Y., Mao, Z., Fang, S., Lu, Y., He, T., Du, H., Zhang, Y., Ouyang, W.: Cross-ing the gap: Domain generalization for image captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2871-2880 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.461, + 0.788, + 0.502 + ], + "angle": 0, + "content": "40. Rennie, S.J., Marcheret, E., Mroueh, Y., Ross, J., Goel, V.: Self-critical sequence training for image captioning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7008-7024 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.502, + 0.788, + 0.542 + ], + "angle": 0, + "content": "41. Shuster, K., Humeau, S., Hu, H., Bordes, A., Weston, J.: Engaging image captioning via personality. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12516-12526 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.542, + 0.788, + 0.597 + ], + "angle": 0, + "content": "42. Toutanova, K., Klein, D., Manning, C.D., Singer, Y.: Feature-rich part-of-speech tagging with a cyclic dependency network. In: Proceedings of the 2003 human language technology conference of the north american chapter of the association for computational linguistics. pp. 252-259 (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.597, + 0.788, + 0.651 + ], + "angle": 0, + "content": "43. Toutanvoa, K., Manning, C.D.: Enriching the knowledge sources used in a maximum entropy part-of-speech tagger. In: 2000 Joint SIGDAT conference on Empirical methods in natural language processing and very large corpora. pp. 63-70 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.651, + 0.788, + 0.692 + ], + "angle": 0, + "content": "44. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4566-4575 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.692, + 0.788, + 0.718 + ], + "angle": 0, + "content": "45. Wang, N., Xie, J., Wu, J., Jia, M., Li, L.: Controllable image captioning via prompting. In: AAAI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.759 + ], + "angle": 0, + "content": "46. Wang, Q., Chan, A.B.: Describing like humans: on diversity in image captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4195-4203 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.759, + 0.788, + 0.8 + ], + "angle": 0, + "content": "47. Wang, T., Zhang, J., Fei, J., Ge, Y., Zheng, H., Tang, Y., Li, Z., Gao, M., Zhao, S., Shan, Y., et al.: Caption anything: Interactive image description with diverse multimodal controls. arXiv preprint arXiv:2305.02677 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.8, + 0.788, + 0.84 + ], + "angle": 0, + "content": "48. Wang, Z., Xiao, J., Chen, L., Gao, F., Shao, J., Chen, L.: Learning combinatorial prompts for universal controllable image captioning. arXiv preprint arXiv:2303.06338 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.128 + ], + "angle": 0, + "content": "K. Basioti et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.218 + ], + "angle": 0, + "content": "49. Xia, Q., Huang, H., Duan, N., Zhang, D., Ji, L., Sui, Z., Cui, E., Bharti, T., Zhou, M.: Xgpt: Cross-modal generative pre-training for image captioning. In: Natural Language Processing and Chinese Computing: 10th CCF International Conference, NLPCC 2021, Qingdao, China, October 13-17, 2021, Proceedings, Part I 10. pp. 786-797. Springer (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.272 + ], + "angle": 0, + "content": "50. Xu, K., Ba, J., Kiros, R., Cho, K., Courville, A., Salakhudinov, R., Zemel, R., Bengio, Y.: Show, attend and tell: Neural image caption generation with visual attention. In: International conference on machine learning. pp. 2048-2057. PMLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "51. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics 2, 67-78 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.357 + ], + "angle": 0, + "content": "52. Zellers, R., Yatskar, M., Thomson, S., Choi, Y.: Neural motifs: Scene graph parsing with global context. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5831-5840 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.41 + ], + "angle": 0, + "content": "53. Zeng, Z., Zhang, H., Lu, R., Wang, D., Chen, B., Wang, Z.: Conzic: Controllable zero-shot image captioning by sampling-based polishing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23465-23476 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.787, + 0.453 + ], + "angle": 0, + "content": "54. Zhao, W., Wu, X., Zhang, X.: Memcap: Memorizing style knowledge for image captioning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 12984-12992 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.495 + ], + "angle": 0, + "content": "55. Zhao, Y., Wei, J., Lin, Z., Sun, Y., Zhang, M., Zhang, M.: Visual spatial description: Controlled spatial-oriented image-to-text generation. arXiv preprint arXiv:2210.11109 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.536 + ], + "angle": 0, + "content": "56. Zheng, Y., Li, Y., Wang, S.: Intention oriented image captions with guiding objects. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8395-8404 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.787, + 0.592 + ], + "angle": 0, + "content": "57. Zhong, Y., Wang, L., Chen, J., Yu, D., Li, Y.: Comprehensive image captioning via scene graph decomposition. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 211-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.621 + ], + "angle": 0, + "content": "58. Zhu, W., Bhat, S.: Gruen for evaluating linguistic quality of generated text. arXiv preprint arXiv:2010.02498 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.621 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CIC-BART-SSA_ _ Controllable Image Captioning with Structured Semantic Augmentation/87b30ae0-4f07-48fc-b46b-fbcdc143e18d_origin.pdf b/2024/CIC-BART-SSA_ _ Controllable Image Captioning with Structured Semantic Augmentation/87b30ae0-4f07-48fc-b46b-fbcdc143e18d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..037ec111fc7de14b00618fc6041d321b3e538ac7 --- /dev/null +++ b/2024/CIC-BART-SSA_ _ Controllable Image Captioning with Structured Semantic Augmentation/87b30ae0-4f07-48fc-b46b-fbcdc143e18d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd7bc256e583c566de0345b41d299a068dcd9fd02936ed48ae828123364f1e72 +size 5511910 diff --git a/2024/CIC-BART-SSA_ _ Controllable Image Captioning with Structured Semantic Augmentation/full.md b/2024/CIC-BART-SSA_ _ Controllable Image Captioning with Structured Semantic Augmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4c6d331de44b305ea13ca41c9d1c94c970440d0b --- /dev/null +++ b/2024/CIC-BART-SSA_ _ Controllable Image Captioning with Structured Semantic Augmentation/full.md @@ -0,0 +1,263 @@ +# CIC-BART-SSA: Controllable Image Captioning with Structured Semantic Augmentation + +Kalliopi Basioti $^{1,2*}$ , Mohamed A. Abdelsalam $^{2}$ , Federico Fancellu $^{3\dagger}$ , Vladimir Pavlovic $^{1\dagger}$ , and Afsaneh Fazly $^{2}$ + +1 Rutgers University, New Jersey, USA {kalliopi.basioti, vladimir}@rutgers.edu +$^{2}$ Samsung AI Centre - Toronto, Toronto, Canada {m.abdelsalam, a.fazly}@samsung.com +3 Solventum ffancellu@solventum.com + +![](images/d0c67294711deb8149e35a775c37ebb01dd7267933b85d9ec5c0d8c478830803.jpg) +Original Captions A couple of elephants standing next to trees. A couple of elephants standing by themselves. A couple elephants walking by a tree after sunset. Two elephants are standing in the grass near a tree. Two elephants are standing by the trees in the wild. + +# Samples Generated from CIC-BART-SSA + +1. There is grass near a tree. +2. There is a tree near the grass. +3. The hillside is covered in +4. A field next to a big tree is open. +5. The grass under the trees is lush +6. A large-elephant standing in a field near a tree. +7. Two elephants walking under a tree in the sunset +8. Two elephants standing in the grass next to a tree +9. Two elephants standing next to a tree and a blue sky. +10. A big elephant standing next to a tree in a field of grass +11. An elephant standing next to a tree with a blue sky behind it. +12. Two elephants standing in the water near a and a blue sky +12. Two elephants standing in the grass near a tree and a blue sky 13. Two elephants standing next to each other on a lush green hill +13. Two elephants standing next to each other on a rush green hillside next to a forest. +14. Two elephants standing next to a tree in a field with a bluesky and trees behind them. +15. Two elephants standing next to each other on a grass covered field next to a lush green forest. + +![](images/43248957ba4e19f716628cd282245a6048b2429162cb83424e188afead66d5b1.jpg) + +![](images/acb148dad3c5cf0c7a8f9daeaca9b14bb0a2e178027ab9a42c85f2e5549a92f1.jpg) +Fig. 1: Existing captioning datasets contain captions that describe the entirety of an image. This is reflected in the narrow distributions of the entities that appear in those captions and the caption lengths (the red-colored histograms). CIC aims to generate diverse descriptions by controllably re-focusing on different spatiosemantic aspects of an image, such as the semantically coherent subsets of image objects. Our proposed CIC-BART-SSA is designed to produce diverse, controlled captions ranging from brief and concise to detailed and comprehensive. Sentences 1-15 are example outputs of our approach where the highlighted text indicates the focus of a controllable caption. The histograms demonstrate that our approach generates high-quality descriptions for a wider range of scene focus (number of visual entities) and caption length compared to the original captions. Image is licensed under CC BY-SA 2.0. + +Abstract. Controllable Image Captioning (CIC) aims at generating natural language descriptions for an image, conditioned on information provided by end users, e.g., regions, entities or events of interest. However, available image-language datasets mainly contain captions that describe the entirety of an image, making them ineffective for training CIC models that can potentially attend to any subset of regions or relationships. To tackle this challenge, we propose a novel, fully automatic method to sample additional focused and visually grounded captions using a unified structured semantic representation built on top of the existing set of captions associated with an image. We leverage Abstract Meaning Representation (AMR), a cross-lingual graph-based semantic formal + +ism, to encode all possible spatio-semantic relations between entities, beyond the typical spatial-relations-only focus of current methods. We use this Structured Semantic Augmentation (SSA) framework to augment existing image-caption datasets with the grounded controlled captions, increasing their spatial and semantic diversity and focal coverage. We then develop a new model, CIC-BART-SSA, specifically tailored for the CIC task, that sources its control signals from SSA-diversified datasets. We empirically show that, compared to SOTA CIC models, CIC-BART-SSA generates captions that are superior in diversity and text quality, are competitive in controllability, and, importantly, minimize the gap between broad and highly focused controlled captioning performance by efficiently generalizing to the challenging highly focused scenarios. Code is available at https://github.com/SamsungLabs/CIC-BART-SSA. + +# 1 Introduction + +Image captioning refers to the task of providing an AI system with an input image, and asking the system to describe the visual content in natural language. This process requires the captioning system to understand what objects are present, in what context (e.g., event or scene), and how they relate. Recent deep learning approaches to this task [14,26,29,31,37,39,40,49,50,57] surpass human performance in standard image captioning metrics. However, these models tend to generate general captions that describe the entirety of an image, and are often of limited diversity; see Original Captions in Fig. 1. + +Controllable image captioning (CIC) overcomes these challenges by generating different descriptions for the same image in a user-controlled fashion. That is, a CIC model receives as input an image paired with a user-specified control signal (e.g., entities or regions of interest), and generates a caption conditioned on the control signal. CIC models are thus capable of generating a diverse set of captions by varying the control signal for the same image; see CIC generated captions 1-15 in Fig. 1. + +In realistic applications, the easiest way for the user to control the generation of captions is to limit the focus of the desired captions by selecting different entities (objects) using their bounding boxes, as shown in Figs. 1 and 2. Most previous work focuses on such spatial control signals [18, 23, 28, 47, 55, 56]. To improve performance, more recent studies supplement this spatial signal with additional information on the desired length, style, or syntactic and semantic structure of the generated text [12, 13], increasing the richness and complexity of control signals. However, for the CIC approach to succeed, the CIC models need to be trained on equally rich datasets that incorporate, explicitly or implicitly, those control signals. Unfortunately, most image captioning datasets today, such as Flickr30k [36] or MS-COCO [18], lack this necessary diversity of controls and corresponding captions. + +Our goal is to achieve SOTA performance in CIC without the need for new, increasingly rich, yet also costly, and impractical-to-collect datasets, where human workers would face the burden of having to provide multitudes of control + +![](images/90e3134b7c2d3a31a77a16c47dddb9977be796c23db20863276cbc3ef1646074.jpg) +Fig. 2: An example of our Structured Semantic Augmentation (SSA) approach. Visually-grounded captions (1)-(5) are used to create a meta-vgAMR graph, which includes all available image information in one representation. Sub-graphs of meta-vgAMR are then sampled to generate a new and diverse set of captions, such as the sentences (a)-(e). Image is licensed under CC BY-SA 2.0. + +signals and corresponding descriptive captions. To achieve this goal, we propose a novel Structured Semantic Augmentation (SSA) method, which automatically generates an augmented set of captions and the corresponding control signals with diverse spatiosemantic focus starting from only the core set of "original" uncontrolled captions. The method takes advantage of a detailed visual-linguistic semantic graph (illustrated in Fig. 2) constructed from the original captions and their image groundings. To build these semantic graphs, we use Abstract Meaning Representation (AMR) [6], a semantic formalism that can capture fine-grained linguistic relations beyond the exclusively spatial relationships present in the common scene graphs [24]. The availability of robust AMR parsers [5,9] allows us to generate semantic graphs for individual captions automatically, which we then merge into a rich meta-AMR graph for the joint image-language pair. From this meta-graph, we sample diverse connected subgraphs that represent semantically coherent combinations of image-anchored entities, events, and their relations, which we then turn into controlled captions automatically via existing AMR-to-text models [9]. Fig. 2 depicts an example of our meta-graph inferred from the original uncontrolled captions associated with an image. Filled nodes in the meta-graph indicate image entity groundings. Five semantically coherent subgraphs (a)-(e) of variable complexity are then sampled from the meta-graph, which are subsequently used to generate novel captions, shown below each subgraph. These new captions augment the original caption set by providing both image focus, through node groundings, and increased semantic diversity induced by the sampled subgraphs. Building upon SSA, we introduce a new CIC model, CIC-BART, suitable for generating focused controlled captions. Alongside the regions of interest, CIC-BART also makes use of the length of the desired cap + +tion as a control signal proxy for the verosity of the caption. CIC-BART can be trained on SSA-augmented versions of standard VL datasets such as MS-COCO or Flicker30k to accommodate the CIC task. Our experiments show that, compared to several SOTA models, the captions generated by our model have superior text quality and diversity, while being comparable in terms of faithfulness to control signals. + +In summary, our contributions are: + +1. We propose a novel data augmentation technique, SSA, that draws on a structured semantic formalism (AMR) to automatically generate focused captions suitable for training of CIC models. We empirically show that our SSA technique enables CIC models to generate captions with high controllability, diversity, and text quality. +2. We propose CIC-BART, a model designed for CIC, that does not require overly descriptive and complex control signals that SOTA models often require to achieve high performance. We show a superior overall performance, compared to SOTA, while relying on simple control signals (i.e., regions of interest and preferred caption length). +3. We present an extensive evaluation of our model, compared with existing SOTA. Specifically, we report results on different aspects of generated captions, including controllability (faithfulness to control signal), diversity, and text quality (linguistic well-formedness). To account for the trade-off among these metrics, we propose an overall performance score based on their harmonic mean. This metric helps us identify models that perform well in all these aspects. + +# 2 Related Work + +Controllable Image Captioning (CIC). Various types of control have been used for CIC, including visual entities, a type of region-based control [18, 23, 28, 47, 55, 56], where generated captions should learn to focus on the regions of interest. Others draw on complex control signals where additional knowledge about the generated caption structure is provided. For example, some recent work provides the complete skeleton of the desired sentence in the form of a number of objects or attributes or object-relation-object templates [12, 13]. Additional control signals that CIC draws on include different caption styles, e.g., positive, negative, humorous, or romantic tone [20, 21, 32, 33, 45, 47, 53, 54], user personality [17, 41], or the length of the generated captions [19, 22, 45, 47, 48]. The use of complex control signals aims at improving the diversity of captions and the quality of the text in CIC models. However, it requires the users to provide a detailed description of the control signal, which is not realistic in practical settings where such models are to be deployed (e.g., a self-driving car or personal assistant). We instead draw on two simple control signals (regions of interest and desired caption length) and show that we can achieve competitive performance on CIC, while keeping the control signals simple and practical. + +Recent SOTA models that draw on spatial control include the SCT model [18] that also uses the Faster R-CNN feature vectors and object tags (corresponding + +GloVe vectors [35]) of the entities of interest, as well as models that include skeleton-based control, namely ASG2Caption [13] and VSR [12]. ASG2Caption uses an abstract scene graph (ASG) to express the desired structure of a caption. ASG contains three types of unlabeled abstract nodes (object, attribute, relationship) that are grounded in the image by extracting features from the corresponding bounding boxes (for objects and attributes) or from the union of bounding box pairs (for a relationship node). ASG2Caption shows improved controllability (by conditioning on ASGs), and diversity (by automatically sampling diverse ASGs as control signals). The VSR model [12] draws on GloVe embeddings of Faster R-CNN object tags for visual entities (as in SCT). It also uses a skeleton control signal (like ASG2Caption), but one that includes more detailed information and richer semantics. Specifically, the VSR control signal follows the form of a fine-grained PropBank entry $^4$ — i.e., specifying the exact verb(s) expressing action(s) depicted in the image, and their visually grounded arguments (e.g., subject, object, location, manner). Thus, VSR uses the most descriptive control signal among the SOTA models. Refer to our supplementary material for an illustrative example of the control signal used for each method. + +Compared to ASG2Caption and VSR, our control signal is kept minimal and only specifies the bounding boxes and desired caption lengths. To improve the diversity of captions, we draw on a structured semantic graph (AMR) that expresses the semantics of a sentence based on PropBank semantics. Notably, we do not use these rich graphs to express detailed and overly descriptive control signals (as in VSR), but we use these semantic structures to augment our training data with richer and diverse captions, which will result in the model learning to generate more diverse captions. Additionally, we include a length control signal to further increase diversity without needing to specify detailed information about the structure of the output (e.g., number of attributes per object, etc.). This way, we can generate a variety of captions for a fixed image sub-region by simply controlling the desired length of the output. + +Abstract Meaning Representation (AMR). AMR [7] is a rich semantic formalism for expressing the meaning of natural language sentences as a formal graph. AMR draws on PropBank, which is a rich lexical semantic resource encoding predicates expressing an action or state, as well as the number and nature of the participating entities (arguments and other semantic roles, such as location, manner, etc.). AMR is a widely researched semantic formalism for which highly accurate automatic Text-to-AMR and AMR-to-Text models are developed [5,9]. We rely on these models to augment original image-caption datasets with newly generated captions (as explained in Sect. 4). + +AMRs vs. Scene Graphs. Recent studies [1,15,16,52] have shown that AMRs better capture the semantic relations of an image as compared to the scene graphs [11]. Existing scene graph annotations mainly capture geometric or possessive relations, which account for more than $90\%$ of the relations captured, whereas more than $1/3$ of the captured entities refer to clothing, object, or body parts information [1,52]. This difference is crucial for high-quality image caption- + +ing, as we use higher-level semantic relations in our everyday language rather than geometric ones. For instance, during a soccer game, we would probably describe a goal save as 'the player kicks the ball away from the goal' or 'the goalkeeper defends his team by saving a goal' and not by using mainly geometric and possessive relations like 'a person wearing a white shirt, standing with his right leg lifted, close to a ball which is above the ground'. In the supplementary, we provide a detailed comparison of AMR and scene graph representations, particularly focusing on their applications in data augmentation. + +# 3 Model + +We propose CIC-BART, specifically designed to generate controlled image captions. Specifically, it can generate descriptions of particular areas within a scene with a desired level of detail. Our model, based on VL-BART [14], utilizes a transformer-based encoder-decoder architecture, as shown in Fig. 3. CIC-BART extends VL-BART encoder to the CIC task by modifying the encoder input to include: a) a global image embedding that provides the context of the + +![](images/5e17183f1d38b7c2a689d07aecff75c9c2478c6b41ce55d737590e0c60c794d6.jpg) +Fig.3: The architecture diagram of our model, CIC-BART. + +full image to the model; b) the visual control signal, including the visual embedding of the regions that contain the entities of interest; c) the text control signal, containing length control (indicating the desired length range of the output caption) and an optional verb signal that indicates the action we want the generated caption to concentrate on. + +The visual embeddings of the regions are position-aware embeddings from a Faster R-CNN model [38] trained for visual object and attribute classification [3] on Visual Genome. The global image feature vector is extracted as well from Faster R-CNN. For the length control signal, we add to our vocabulary $L$ tokens for the $L$ different caption length levels; for instance, level one represents sentences between one and nine words, and level two, ten to nineteen. These tokens describe our coarse levels, for a finer sentence size accuracy, we accompany the tokens with the desired number of words. This choice gives our model the capacity to generate diverse captions for a particular length level. Finally, the output of the decoder generates the desired, controlled image caption. + +# 4 Structured Semantic Augmentation (SSA) + +The goal of our SSA method is to augment existing image captioning datasets with new focused captions along with their control signals (i.e., regions corresponding to entities). We rely on datasets where visual entities in the captions + +are annotated with their corresponding regions (see Sect. 5 for details on the datasets). The SSA process consists of four main steps, as described below. For more details, refer to our supplementary material, which also includes a step-by-step example of our SSA methodology. + +Step 1: Image-level AMR graph generation. Our objective in this stage is to enclose all the information available from the visually grounded captions into a single representation. To accomplish this, we create a visually grounded AMR graph (vgAMR) for each caption of an image and then merge them into a single image-level graph, the meta-vgAMR. To create the vgAMRs of an image, we first convert each of its $N$ captions to their AMR representation, using the Neural transition-based Text-to-AMR parser [5] which also aligns words in a caption with their respective nodes in the AMR graph. We utilize the alignment information of 1) caption words and AMR nodes (from Text-to-AMR parser) and 2) caption words to image bounding boxes (from existing dataset annotations) to visually ground the AMR nodes. After this step, we get the collection of nodes referring to visual entities, where each grounded meta-AMR node is linked with a non-empty set of bounding boxes. This extended representation, 'AMR + visual grounded nodes', is our vgaMR. + +Our next step is to combine the N vgAMRs to form a single meta-vgAMR. To achieve this, we employ a pairwise strategy to merge the most similar vgAMRs first (we measure similarity with Smatch score [10]). We use the UPGMA hierarchical clustering algorithm [30,34] to find the optimal merge ordering starting from the most similar graphs. UPGMA creates a hierarchy where the bottom level consists of the N individual vgAMRs. By merging all vgAMRs using the UPGMA ordering, we obtain a single structure called meta-vgAMR. + +When merging two vgAMRs, the main challenge is identifying which nodes correspond to the same concepts, such as entities, attributes, actions, and relations. We use three node properties to accomplish this: a) visual grounding information, b) semantic similarity of node labels, and c) node neighborhood semantic similarity. We derive two node-merging criteria from there: 1) visually grounded entity nodes are merged if they point to the same image-bounding boxes. When 1) does not hold, we check the second criterion: 2) for the remaining non-grounded nodes, including amr-specific, predicates, adjectives, and adverbs, we use a combination of node label semantic similarity (cosine similarity of the labels using their GloVe embeddings) and neighborhood similarity. Neighborhood similarity examines the similarity of parents for adjectives/adverbs nodes and children for predicate nodes, along with the similarity of connecting edge roles. When two nodes satisfy criterion 1) or 2), we merge them into a single node. Moreover, if they have different labels, we maintain both names by keeping a list of synonyms to increase representation diversity. In the special case when the two vgAMRs describe two totally different concepts, and hence they have no common nodes, we add an amr-specific node called 'multi-sentence' as the root with the two independent vgAMRs as its children. The final graph, meta + +vgAMR, includes all non-redundant5 elements of the original $N$ captions while preserving the visual grounding between the meta nodes and their respective image regions. + +Remark: Meta-vgAMR efficiently compresses all available image information into a single structure. Following our approach, we can easily scale when new scene information becomes available by applying our pairwise merge procedure. + +Step 2: Event-based graph sampling from image-level AMRs. We start from the predicate nodes, which mainly correspond to verbs, to sample subgraphs in meta-vgAMR graphs. Predicate nodes are identified by their label and the edges connected to them. The label of a predicate node typically follows the format 'predicate_name-xx,' where 'xx' represents the different senses a word can have regarding the concept it is used for. Predicate nodes have outward ARGy edges, where 'y' can take values from 0 to 5, connecting them to their arguments. We sample subgraphs from these nodes by following the outgoing argument edges, which are labeled as ARGn in an AMR graph, each defining a particular semantic role (e.g., ARG0 points to the agent, ARG1 to the patient, etc.). Finally, we add one more subgraph containing the remaining children branches of other non-ARG optional predicate edges (e.g., 'location', 'time'). We repeat this process until the leaves of the graphs are reached. During sampling, we randomly select one of the synonyms if a node is a list of synonym labels, as mentioned in the previous step. The output of this step is our more focused event-focused sub-graphs. In Fig. 2, we can see some instances of our event-based sampling (SSA samples), where the predicate nodes include z0/sit-01, z7/dock-01, z13/sit-01, and so on6. Although we cannot show all the sampled event-based subgraphs in the figure, we included five of them and used colored roots and edges for visualization purposes. + +Step 3: New caption generation from sampled AMRs We use the SPRING AMR-to-Text model [8] to generate new event-focused captions from the sampled vgAMR subgraphs. Because both vgAMR merging and sampling steps introduce noise, the output captions are not always of good quality. We automatically filter low-quality captions by using a linguistic well-formedness measure, GRUEN [58], which is a reference-free metric based on BERT contextual embeddings. + +Step 4: Control signal generation. The last step is to create the control signal for the generated captions. The spatial control signal for a specific caption is extracted from the corresponding sampled vgAMR, by pulling the bounding boxes of the visual entity linked AMR nodes. + +# 4.1 Mixing Strategies of Original and SSA Data + +To analyze the impact of our SSA data, we explore various mixing strategies with the original training set. Assume $\mathcal{D}$ represents the training control-caption pairs in the original dataset, containing $N_{\mathcal{D}}$ samples, and SSA represents our SSA samples, containing $N_{SSA}$ instances. The augmented dataset $\mathcal{D}_{SSA}$ is defined by combining $\mathcal{D}$ and SSA: $\mathcal{D}_{SSA} = \mathrm{sam}_{\mathcal{D}}(\tau_{\mathcal{D}}, p_{\mathcal{D}}) \cup \mathrm{sam}_{SSA}(\tau_{SSA}, p_{SSA})$ , where the functions $\mathrm{sam}_{\mathcal{D}}$ samples a subset of the original dataset, and $\mathrm{sam}_{SSA}$ a subset of our SSA data. Since we are interested in the effect of our SSA, we assume that $\mathrm{sam}_{\mathcal{D}}(\tau_{\mathcal{D}}, p_{\mathcal{D}}) = \mathcal{D}$ , with $\tau_{\mathcal{D}} = \text{'Random Sampling Strategy'}$ and $p_{\mathcal{D}} = 100\%$ , meaning that all original data are included in the mixed dataset. Depending on the $\mathrm{sam}_{SSA}$ parameter $\tau_{SSA}$ we have the cases: + +Random Sampling Strategy. In this case, we randomly select a pre-specified number of examples from $SSA$ . The parameter $p_{SSA}$ expresses the percentage of SSA samples included in $\mathcal{D}_{SSA}$ . With boundary cases $p_{SSA} = 100\%$ (all $N_{SSA}$ samples are included), and $p_{SSA} = 0\%$ (no SSA data are added). + +Uniform-Coverage Sampling Strategy. To mitigate the original dataset's bias (having mainly samples describing the entire image), we aim to create a new focus-unbiased dataset. By modeling the control signal focus as the image area percentage covered by the bounding boxes of the control signal, we split the original data into $B$ coverage bins. Then, we will randomly add in each bin SSA samples, aiming to create a new uniform, coverage-unbiased $\mathcal{D}_{SSA}$ dataset. Here, $p_{SSA}$ contains the range of each bin for the coverage histogram. For example, in the case where we choose ten uniform coverage bins, we have $p_{SSA} = \{[0\%, 10\%), [10\%, 20\%), \ldots, [90\%, 100\%]\}$ . + +Due to space limitations, we present results from the Random Sampling Strategy for $p_{SSA} = 0\%$ and $p_{SSA} = 100\%$ in the main paper. Results from other scenarios can be found in the supplementary material. + +# 5 Experimental Setup + +# 5.1 Data + +We use Flickr30k Entities (Flickr-Ent) [36] and MS-COCO Entities (COCO-Ent) [18] for training and evaluation. Flickr-Ent augments the original captions of Flickr30k [51] with manually-annotated region-phrase groundings. Flickr-Ent contains the original $31K$ images annotated with five captions each. COCO-Ent augments the original MS-COCO [27] ( $120K$ images each annotated with around five captions) with semi-automatically collected grounding annotations; see [18] for details on the annotation process. For both datasets, we follow previous work and use the training and test splits by Karpathy et al. [25]. We apply our SSA algorithm on the aforementioned datasets to create their augmented variations, COCO-Ent-SSA and Flickr-Ent-SSA, containing about $800K$ and $250K$ training captions, respectively, of which $33\%$ and $37\%$ , are generated by our SSA algorithm. + +For all four training sets, we automatically generate image-control-caption triplets to train our model on. For spatial control, we extract from the grounded captions the bounding boxes of the entities of interest using the annotations from COCO-Ent and Flickr-Ent. Note that since these datasets do not contain the text control signal, we use each ground-truth caption as a proxy for a controlled caption, from which we first generate the coarse- and fine-length control levels and then extract their verbs using part-of-speech tagging for the optional action control. + +# 5.2 Models and Evaluation Metrics + +We compare two variations of our model (with and without SSA augmentations) with SOTA models as our baselines: Show Control & Tell (SCT) [18] that uses region-based control (bounding boxes of visual entities of interest); ASG2Caption (ASG) [13] that draws on visually grounded abstract scene graphs as control signal; and VSR [12] that uses overly descriptive control signals that express verb(s) and fine-grained verb-specific semantic roles of the desired captions; ComPro [48] that learns a mapping from the bounding boxes of the entities of interest and caption length level to GPT-2 Large prompts aiming to retrieve controlled captions; and the LaBERT length-control-only model [19]. + +We report the performance of our models and baselines using a comprehensive set of metrics that evaluate different aspects of caption controllability, diversity, and quality. We also propose and report an overall performance metric that summarizes these different aspects in a meaningful way. To measure diversity, we compute n-gram diversity, D-n for $n = 1$ , 2 [4], as well as self-CIDEr-based diversity (sC) [46]. For a fair comparison of the different CIC models, we measure diversity for the five generated captions for each test image (in COCO-Ent and Flickr-Ent), and report their average. To measure content controllability we design an extended version of the IoU metric of [18] that calculates the degree-of-match (faithfulness) between a control signal and the corresponding generated caption. For our control signal, we use the set of nouns $\mathcal{E}$ that represent the entities of interest, which are the names of the visual objects in the control. To extract nouns from the predicted sentences, we use the Stanford part-of-speech tagger [42,43]. We then find the semantic intersection of the two sets using Hungarian Matching, as in [18]. Finally, we calculate the semantic intersection over union of the control nouns and the nouns extracted from the controllable caption, which gives us our content controllability IoU. We further analyze IoU by introducing the Hallucinating Nouns (Hal) metric. Details can be found in the supplemental material. For length controllability (L), we measure the Mean Absolute Error (MAE) between the fine length control (number of words) and the size of the resulting $M = 10$ controlled captions, which are generated from $M$ randomly created control signals. We also calculate the length precision (LP) [19] by determining the percentage of generated captions that match the desired coarse length level. We assess text quality of generated captions using GRUEN (G) [58], a reference free metric based on BERT contextual embeddings that measures the syntactic and semantic well-formedness of a text segment. Finally, we measure + +Table 1: Performance of CIC models for the original test sets. *ASG-type dataset is not released for Flickr-Ent; therefore, we could not reproduce the ASG results. + +
ModelH↑IoU↑G↑sC↑D-1↑D-2↑L↓H↑IoU↑G↑sC↑D-1↑D-2↑L↓
SCT [18]55.867.364.442.827.035.5-54.650.779.844.029.336.5-
ASG* [13]74.272.672.078.337.856.6--------
VSR [12]56.277.639.067.430.042.2-62.560.254.077.933.349.3-
CIC-BART75.976.273.078.738.056.2.4969.854.085.078.643.658.21.24
CIC-BART-SSA78.377.274.882.544.663.2.1171.355.086.081.747.062.61.05
+ +Table 2: Performance of CIC models, for the SSA-only samples. + +
ModelH↑IoU↑G↑sC↑D-1↑D-2↑H↑IoU↑G↑sC↑D-1↑D-2↑
COCO-Ent (SSA only)Flickr-Ent (SSA only)
SCT [18]51.762.164.837.823.731.043.929.977.345.731.036.7
CIC-BART69.261.473.974.044.257.068.553.080.579.852.962.9
CIC-BART-SSA75.665.280.783.753.867.872.055.682.986.156.569.3
+ +the overall performance of each model based on its ability to balance content controllability, diversity, and text quality. To calculate this, we use the harmonic mean of IoU, G, and sC. All of these metrics range between 0 and 1, with a higher value indicating better performance. The harmonic mean $(H)$ helps us determine the model with the best overall performance. It prioritizes models that perform well across all metrics while penalizing those with poor performance, even in one metric. In our supplementary, we include comparisons of the CIC models on standard captioning metrics (like CIDEr [44] and Spice [2]). + +# 6 Results + +# 6.1 Overall Performance + +We first compare the overall performance of our models with the SCT, ASG, and VSR baselines with respect to controllable captioning metrics. We do not include ComPro in this comparison due to the unavailability of the codebase. We also exclude LaBERT since this model solely focuses on length controllability. + +Tab. 1 presents results for content and length controllability (IoU and L), text quality (G), and diversity (sC, D-1, D-2), as well as the harmonic mean (H). For both datasets, CIC-BART-SSA has the best performance in all metrics, except IoU, where it is the second best. Specifically, CIC-BART-SSA is superior to all other models with respect to diversity (sC, D-1, D-2) and text quality (G), but comparable to VSR in terms of content controllability (IoU). The length controllability (L) scores show that our SSA augmentation helps the model learn to generate high-quality output at the desirable length (compare CIC-BART and CIC-BART-SSA). This is due to the increased diversity in caption length provided by our SSA augmentations. + +Importantly, we can see that model performance can vary depending on the metric. E.g., whereas VSR has the highest IoU, it falls behind in text quality and + +diversity. In our qualitative analysis, we observe the poor quality of the captions generated by VSR. The best-performing model should be identified based on the $H$ score that summarizes content controllability, text quality, and diversity into a single score. Based on this score, CIC-BART-SSA is better than SCT and VSR baselines by a large margin, and notably better than ASG. Nevertheless, ASG requires complex control signals in the form of scene graphs, in contrast to the simple control signal requirements of CIC-BART. + +Next, we conduct a further evaluation of the CIC performance on our SSA samples from the test set images of COCO-Ent and Flickr-Ent. We present the results in Tab. 2 for our models CIC-BART, CIC-BART-SSA, and SCT. ASG and VSR are excluded since they need complex control signals (grounded abstract scene graphs for ASG and grounded verb semantic roles for VSR), which are only available for COCO-Ent and Flickr-Ent. We observe a significant improvement in overall performance for our CIC-BART-SSA model. We notice that the models (SCT and CIC-BART) trained on the original datasets, which describe the entire image, had difficulties generalizing to cases where they had to focus on a specific sub-region of an image. However, our model CIC-BART-SSA was able to generate focused and diverse descriptions of the challenging, highly focused examples present in our SSA data. + +We compare the length precision of our model with the baselines utilizing length control in Table 3. LaBERT uses only length-control signals without spatial control, while our model employs both spatial and length-control signals to generate focused captions. This makes the LaBERT task much easier since it only focuses on generating specific length descriptions of an image. On the other hand, our model focuses on generating captions that describe only a specific sub-region of the scene while maintaining a desired description length level. Although our task is + +Table 3: Length Precision (LP) for CIC models on COCO-Ent and Flickr-Ent original test sets. + +
ModelLP↑LP↑
COCO-EntFlickr-Ent
ComPro [48]94.781.4
LaBERT [19]99.798.4
CIC-BART99.988.0
CIC-BART-SSA99.991.3
+ +more challenging than LaBERT, we achieve competitive length precision performance. Lastly, we want to emphasize that the remarkable improvement in length controllability (L) and length precision (LP) from CIC-BART to CIC-BART-SSA stems from the increased length diversity found in our SSA augmentations, which enriches the original COCO-Ent and Flickr-Ent datasets. In the supplementary, we provide an analysis of the caption length statistics in the original datasets and our SSA-derived captions. + +# 6.2 Effect of SSA on Content Controllability + +To analyze the impact of our SSA augmentations, we measure the content controllability (IoU) performance of CIC-BART at different levels of focus of the control signals and report it in Fig. 4. We use coverage, defined as the area of the image enclosed by the bounding boxes of the entities of interest in the control + +![](images/df739c11ae44e09b08c33caa8eeb951bd62d3267c34302cd30000cb3e0134f4a.jpg) + +![](images/d2aebe78818bc5dbd0a8fec51790e1acc47c64a6dde0b9922686aa031aaf33ef.jpg) + +![](images/7de535f4d8e4e743029c56f016ee5af8bc6d829a7c4b8659a989fc65da71bdd5.jpg) +Fig. 4: IoU histograms for CIC-BART (green) and CIC-BART-SSA (blue). The first column depicts the IoU on the original test set and the second on the original test set images' SSA (only) data. The Samples curve (orange) represents the distribution of test captions in each coverage (image area covered by the control signal) interval. + +![](images/2a1a95078e415f9689cb7116e18e8c9ccedce9b3f373eac72de65623bbb6e1b4.jpg) + +![](images/3fd65594618cb87496e6f6382f1a7b5649d323fe7e3788e11c718565d96629ea.jpg) +Fig. 5: Qualitative examples for the original test sets. Strikethrough marks hallucinations and redundancies. a, b licensed under CC BY-SA 2.0; c, d under CC BY 2.0. + +![](images/fa718370c4bfe476d621f99eb49475122f4ae6f87f22bbbc60f76206f450ed1c.jpg) + +signal, to quantify that focus. For example, highly focused control signals cover a small area, yielding low coverage, while broader signals cover a larger area and have high coverage. We 'break down' the IoU performance into 10 coverage bands and report the average IoU over control signals in those bands. In addition, the 'Samples' curve shows the distribution of test captions over the same bands. The results in Fig. 4 indicate that by training with SSA (blue bars), the spatial controllability improves significantly in the low-coverage regime, where the control signals are highly focused. Interestingly, these are also the most underrepresented (data deprived) parts of the original dataset Flickr-Ent. Therefore, SSA which enriches the original datasets with highly focused examples (please refer to our supplementary for % Samples per coverage bands for the training sets), is effective in improving generalization performance in CIC. We include the performance of hallucinating nouns metric in the supplementary material. + +# 6.3 Qualitative Analysis + +In Fig. 5, we present qualitative examples from the original test sets, and in Fig. 6 examples from our SSA (only) test set control signals. In the two figures, each highlighted word found in the generated controlled captions corresponds to the control entity of the same color. This shows the match between the captions produced and the control signal. We also strike through the parts where the model hallucinates or generates redundant references to the entities of interest. Our models have been observed to outperform the previous state-of-the-art models by substantially enhancing the quality of the generated controlled captions. This behavior was expected from our quantitative analysis, which showed that our models have signifi + +cantly higher text quality (G). More importantly, our CIC-BART-SSA model is capable of generating captions that are faithful to the control signal and better understand the relationships that connect the entities of interest. We include additional qualitative samples in the Supplementary Material. + +![](images/6eaa26b9b592b75b60512fc99bcec7e07ee58b015f5d94ade8de858079ea21d2.jpg) +Fig. 6: Examples for SSA-only test set. Strikethrough marks hallucinations. Up to Down: a licensed under CC BY 2.0; b under CC PDM 1.0, for last two see Fig. 5. + +# 7 Conclusions + +We address two main challenges faced by the controllable image captioning (CIC) models. First, standard image-caption datasets lack the controllability and diversity needed for proper training and evaluation of CIC. Second, most recent SOTA models require complex and overly descriptive control signals as input (including, e.g., the main action/verb to appear in the generated caption). To address the first challenge, we propose a novel technique that draws on a structured semantic augmentation (SSA) formalism to generate focused captions and the corresponding control signals for images. For the second challenge, we propose a transformer-based vision-language model attuned to the CIC task. We show that this model performs competitively with SOTA models without requiring complex and explicit control signals. Importantly, when combined with our SSA approach, our model generates highly diverse captions and significantly reduces the content controllability performance gap between the different levels of focus of the generated controlled captions. Finally, when provided with the commonly used verb guidance of other SOTA approaches, our model shows a substantial improvement in performance. + +# References + +1. Abdelsalam, M.A., Shi, Z., Fancellu, F., Basioti, K., Bhatt, D., Pavlovic, V., Fazly, A.: Visual semantic parsing: From images to abstract meaning representation. In: Proceedings of the 26th Conference on Computational Natural Language Learning (CoNLL). pp. 282-300 (2022) +2. Anderson, P., Fernando, B., Johnson, M., Gould, S.: Spice: Semantic propositional image caption evaluation. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 382-398. Springer (2016) +3. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: CVPR (2018) +4. Aneja, J., Agrawal, H., Batra, D., Schwing, A.: Sequential latent spaces for modeling the intention during diverse image captioning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4261-4270 (2019) +5. Astudillo, R.F., Ballesteros, M., Naseem, T., Blodgett, A., Florian, R.: Transition-based parsing with stack-transformers. arXiv preprint arXiv:2010.10669 (2020) +6. Banarescu, L., Bonial, C., Cai, S., Georgescu, M., Griffitht, K., Hermjakob, U., Knight, K., Koehn, P., Palmer, M., Schneider, N.: Abstract meaning representation (amr) 1.0 specification. In: Parsing on Freebase from Question-Answer Pairs. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing. Seattle: ACL. pp. 1533–1544 (2012) +7. Banarescu, L., Bonial, C., Cai, S., Georgescu, M., Griffitht, K., Hermjakob, U., Knight, K., Koehn, P., Palmer, M., Schneider, N.: Abstract meaning representation for sembanking. In: Proceedings of the 7th linguistic annotation workshop and interoperability with discourse. pp. 178-186 (2013) +8. Bevilacqua, M., Błoshmi, R., Navigli, R.: One spring to rule them both: Symmetric amr semantic parsing and generation without a complex pipeline. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 12564-12573 (2021) +9. Bloshmi, R., Bevilacqua, M., Fabiano, E., Caruso, V., Navigli, R.: Spring goes online: end-to-end AMR parsing and generation. In: Proceedings of the 2021 conference on empirical methods in natural language processing: system demonstrations. pp. 134-142 (2021) +10. Cai, S., Knight, K.: Smatch: an evaluation metric for semantic feature structures. In: Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). pp. 748-752 (2013) +1. Chang, X., Ren, P., Xu, P., Li, Z., Chen, X., Hauptmann, A.: A comprehensive survey of scene graphs: Generation and application. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(1), 1-26 (2021) +2. Chen, L., Jiang, Z., Xiao, J., Liu, W.: Human-like controllable image captioning with verb-specific semantic roles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16846-16856 (2021) +3. Chen, S., Jin, Q., Wang, P., Wu, Q.: Say as you wish: Fine-grained control of image caption generation with abstract scene graphs. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9962-9971 (2020) +4. Cho, J., Lei, J., Tan, H., Bansal, M.: Unifying vision-and-language tasks via text generation. In: International Conference on Machine Learning. pp. 1931-1942. PMLR (2021) + +15. Choi, W.S., Heo, Y.J., Punithan, D., Zhang, B.T.: Scene graph parsing via abstract meaning representation in pre-trained language models. In: NAACL 2022 Workshop on Deep Learning on Graphs for Natural Language Processing (2022) +16. Choi, W.S., Heo, Y.J., Zhang, B.T.: Sgram: Improving scene graph parsing via abstract meaning representation. arXiv preprint arXiv:2210.08675 (2022) +17. Chunseong Park, C., Kim, B., Kim, G.: Attend to you: Personalized image captioning with context sequence memory networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 895-903 (2017) +18. Cornia, M., Baraldi, L., Cucchiara, R.: Show, control and tell: A framework for generating controllable and grounded captions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8307-8316 (2019) +19. Deng, C., Ding, N., Tan, M., Wu, Q.: Length-controllable image captioning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16. pp. 712-729. Springer (2020) +20. Gan, C., Gan, Z., He, X., Gao, J., Deng, L.: Stylenet: Generating attractive visual captions with styles. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3137-3146 (2017) +21. Guo, L., Liu, J., Yao, P., Li, J., Lu, H.: Mscap: Multi-style image captioning with unpaired stylized text. 2019 IEEE. In: CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4199-4208 (2019) +22. Hirsch, E., Tal, A.: Clid: Controlled-length image descriptions with limited data. arXiv preprint arXiv:2211.14835 (2022) +23. Johnson, J., Karpathy, A., Fei-Fei, L.: Densecap: Fully convolutional localization networks for dense captioning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4565-4574 (2016) +24. Johnson, J., Krishna, R., Stark, M., Li, L.J., Shamma, D., Bernstein, M., Fei-Fei, L.: Image retrieval using scene graphs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3668-3678 (2015) +25. Karpathy, A., Fei-Fei, L.: Deep visual-semantic alignments for generating image descriptions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3128-3137 (2015) +26. Li, X., Yin, X., Li, C., Zhang, P., Hu, X., Zhang, L., Wang, L., Hu, H., Dong, L., Wei, F., et al.: Oscar: Object-semantics aligned pre-training for vision-language tasks. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16. pp. 121-137. Springer (2020) +27. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +28. Lindh, A., Ross, R.J., Kelleher, J.D.: Language-driven region pointer advancement for controllable image captioning. arXiv preprint arXiv:2011.14901 (2020) +29. Lu, J., Yang, J., Batra, D., Parikh, D.: Neural baby talk. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7219-7228 (2018) +30. Lukasova, A.: Hierarchical agglomerative clustering procedure. Pattern Recognition 11(5-6), 365-381 (1979) +31. Luo, J., Li, Y., Pan, Y., Yao, T., Feng, J., Chao, H., Mei, T.: Semantic-conditional diffusion networks for image captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23359-23368 (2023) +32. Mathews, A., Xie, L., He, X.: Senticap: Generating image descriptions with sentiments. In: Proceedings of the AAAI conference on artificial intelligence. vol. 30 (2016) + +33. Mathews, A., Xie, L., He, X.: Semstyle: Learning to generate stylised image captions using unaligned text. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8591-8600 (2018) +34. Müller, D.: Modern hierarchical, agglomerative clustering algorithms. arXiv preprint arXiv:1109.2378 (2011) +35. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP). pp. 1532-1543 (2014) +36. Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In: Proceedings of the IEEE international conference on computer vision. pp. 2641-2649 (2015) +37. Ramos, R., Martins, B., Elliott, D., Kementchedjheva, Y.: Smallcap: lightweight image captioning prompted with retrieval augmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2840-2849 (2023) +38. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems 28 (2015) +39. Ren, Y., Mao, Z., Fang, S., Lu, Y., He, T., Du, H., Zhang, Y., Ouyang, W.: Cross-ing the gap: Domain generalization for image captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2871-2880 (2023) +40. Rennie, S.J., Marcheret, E., Mroueh, Y., Ross, J., Goel, V.: Self-critical sequence training for image captioning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7008-7024 (2017) +41. Shuster, K., Humeau, S., Hu, H., Bordes, A., Weston, J.: Engaging image captioning via personality. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12516-12526 (2019) +42. Toutanova, K., Klein, D., Manning, C.D., Singer, Y.: Feature-rich part-of-speech tagging with a cyclic dependency network. In: Proceedings of the 2003 human language technology conference of the north american chapter of the association for computational linguistics. pp. 252-259 (2003) +43. Toutanvoa, K., Manning, C.D.: Enriching the knowledge sources used in a maximum entropy part-of-speech tagger. In: 2000 Joint SIGDAT conference on Empirical methods in natural language processing and very large corpora. pp. 63-70 (2000) +44. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4566-4575 (2015) +45. Wang, N., Xie, J., Wu, J., Jia, M., Li, L.: Controllable image captioning via prompting. In: AAAI (2023) +46. Wang, Q., Chan, A.B.: Describing like humans: on diversity in image captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4195-4203 (2019) +47. Wang, T., Zhang, J., Fei, J., Ge, Y., Zheng, H., Tang, Y., Li, Z., Gao, M., Zhao, S., Shan, Y., et al.: Caption anything: Interactive image description with diverse multimodal controls. arXiv preprint arXiv:2305.02677 (2023) +48. Wang, Z., Xiao, J., Chen, L., Gao, F., Shao, J., Chen, L.: Learning combinatorial prompts for universal controllable image captioning. arXiv preprint arXiv:2303.06338 (2023) + +49. Xia, Q., Huang, H., Duan, N., Zhang, D., Ji, L., Sui, Z., Cui, E., Bharti, T., Zhou, M.: Xgpt: Cross-modal generative pre-training for image captioning. In: Natural Language Processing and Chinese Computing: 10th CCF International Conference, NLPCC 2021, Qingdao, China, October 13-17, 2021, Proceedings, Part I 10. pp. 786-797. Springer (2021) +50. Xu, K., Ba, J., Kiros, R., Cho, K., Courville, A., Salakhudinov, R., Zemel, R., Bengio, Y.: Show, attend and tell: Neural image caption generation with visual attention. In: International conference on machine learning. pp. 2048-2057. PMLR (2015) +51. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics 2, 67-78 (2014) +52. Zellers, R., Yatskar, M., Thomson, S., Choi, Y.: Neural motifs: Scene graph parsing with global context. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5831-5840 (2018) +53. Zeng, Z., Zhang, H., Lu, R., Wang, D., Chen, B., Wang, Z.: Conzic: Controllable zero-shot image captioning by sampling-based polishing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23465-23476 (2023) +54. Zhao, W., Wu, X., Zhang, X.: Memcap: Memorizing style knowledge for image captioning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 12984-12992 (2020) +55. Zhao, Y., Wei, J., Lin, Z., Sun, Y., Zhang, M., Zhang, M.: Visual spatial description: Controlled spatial-oriented image-to-text generation. arXiv preprint arXiv:2210.11109 (2022) +56. Zheng, Y., Li, Y., Wang, S.: Intention oriented image captions with guiding objects. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8395-8404 (2019) +57. Zhong, Y., Wang, L., Chen, J., Yu, D., Li, Y.: Comprehensive image captioning via scene graph decomposition. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 211-229. Springer (2020) +58. Zhu, W., Bhat, S.: Gruen for evaluating linguistic quality of generated text. arXiv preprint arXiv:2010.02498 (2020) \ No newline at end of file diff --git a/2024/CIC-BART-SSA_ _ Controllable Image Captioning with Structured Semantic Augmentation/images.zip b/2024/CIC-BART-SSA_ _ Controllable Image Captioning with Structured Semantic Augmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b7cd14b0bc88776cf54f96ef81d34cd08dc72a27 --- /dev/null +++ b/2024/CIC-BART-SSA_ _ Controllable Image Captioning with Structured Semantic Augmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dfafb6843fcf2122f409dbd94e58d6b3f9e74ce5bf2a8489807b5e6e51145958 +size 351115 diff --git a/2024/CIC-BART-SSA_ _ Controllable Image Captioning with Structured Semantic Augmentation/layout.json b/2024/CIC-BART-SSA_ _ Controllable Image Captioning with Structured Semantic Augmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c13f0695160e072f266c6be3b7995488f83ff93a --- /dev/null +++ b/2024/CIC-BART-SSA_ _ Controllable Image Captioning with Structured Semantic Augmentation/layout.json @@ -0,0 +1,8428 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 137, + 111, + 477, + 148 + ], + "type": "text", + "content": "CIC-BART-SSA: Controllable Image Captioning with Structured Semantic Augmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "spans": [ + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "text", + "content": "Kalliopi Basioti" + }, + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "text", + "content": ", Mohamed A. Abdelsalam" + }, + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "text", + "content": ", Federico Fancellu" + }, + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + }, + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "text", + "content": ", Vladimir Pavlovic" + }, + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "text", + "content": ", and Afsaneh Fazly" + }, + { + "bbox": [ + 135, + 167, + 478, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 169, + 201, + 443, + 256 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 181, + 201, + 430, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 201, + 430, + 224 + ], + "spans": [ + { + "bbox": [ + 181, + 201, + 430, + 224 + ], + "type": "text", + "content": "1 Rutgers University, New Jersey, USA {kalliopi.basioti, vladimir}@rutgers.edu" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 169, + 224, + 443, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 224, + 443, + 245 + ], + "spans": [ + { + "bbox": [ + 169, + 224, + 443, + 245 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 169, + 224, + 443, + 245 + ], + "type": "text", + "content": " Samsung AI Centre - Toronto, Toronto, Canada {m.abdelsalam, a.fazly}@samsung.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 224, + 245, + 389, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 245, + 389, + 256 + ], + "spans": [ + { + "bbox": [ + 224, + 245, + 389, + 256 + ], + "type": "text", + "content": "3 Solventum ffancellu@solventum.com" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 133, + 272, + 220, + 329 + ], + "blocks": [ + { + "bbox": [ + 133, + 272, + 220, + 329 + ], + "lines": [ + { + "bbox": [ + 133, + 272, + 220, + 329 + ], + "spans": [ + { + "bbox": [ + 133, + 272, + 220, + 329 + ], + "type": "image", + "image_path": "d0c67294711deb8149e35a775c37ebb01dd7267933b85d9ec5c0d8c478830803.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 330, + 232, + 365 + ], + "lines": [ + { + "bbox": [ + 133, + 330, + 232, + 365 + ], + "spans": [ + { + "bbox": [ + 133, + 330, + 232, + 365 + ], + "type": "text", + "content": "Original Captions A couple of elephants standing next to trees. A couple of elephants standing by themselves. A couple elephants walking by a tree after sunset. Two elephants are standing in the grass near a tree. Two elephants are standing by the trees in the wild." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 235, + 272, + 323, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 272, + 323, + 278 + ], + "spans": [ + { + "bbox": [ + 235, + 272, + 323, + 278 + ], + "type": "text", + "content": "Samples Generated from CIC-BART-SSA" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 235, + 279, + 422, + 365 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 235, + 279, + 292, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 279, + 292, + 285 + ], + "spans": [ + { + "bbox": [ + 235, + 279, + 292, + 285 + ], + "type": "text", + "content": "1. There is grass near a tree." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 236, + 285, + 299, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 285, + 299, + 289 + ], + "spans": [ + { + "bbox": [ + 236, + 285, + 299, + 289 + ], + "type": "text", + "content": "2. There is a tree near the grass." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 236, + 290, + 290, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 290, + 290, + 295 + ], + "spans": [ + { + "bbox": [ + 236, + 290, + 290, + 295 + ], + "type": "text", + "content": "3. The hillside is covered in" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 236, + 295, + 306, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 295, + 306, + 300 + ], + "spans": [ + { + "bbox": [ + 236, + 295, + 306, + 300 + ], + "type": "text", + "content": "4. A field next to a big tree is open." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 236, + 300, + 306, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 300, + 306, + 306 + ], + "spans": [ + { + "bbox": [ + 236, + 300, + 306, + 306 + ], + "type": "text", + "content": "5. The grass under the trees is lush" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 236, + 306, + 334, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 306, + 334, + 312 + ], + "spans": [ + { + "bbox": [ + 236, + 306, + 334, + 312 + ], + "type": "text", + "content": "6. A large-elephant standing in a field near a tree." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 236, + 312, + 337, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 312, + 337, + 317 + ], + "spans": [ + { + "bbox": [ + 236, + 312, + 337, + 317 + ], + "type": "text", + "content": "7. Two elephants walking under a tree in the sunset" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 236, + 317, + 338, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 317, + 338, + 323 + ], + "spans": [ + { + "bbox": [ + 236, + 317, + 338, + 323 + ], + "type": "text", + "content": "8. Two elephants standing in the grass next to a tree" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 236, + 323, + 344, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 323, + 344, + 329 + ], + "spans": [ + { + "bbox": [ + 236, + 323, + 344, + 329 + ], + "type": "text", + "content": "9. Two elephants standing next to a tree and a blue sky." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 236, + 329, + 349, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 329, + 349, + 335 + ], + "spans": [ + { + "bbox": [ + 236, + 329, + 349, + 335 + ], + "type": "text", + "content": "10. A big elephant standing next to a tree in a field of grass" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 236, + 335, + 359, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 335, + 359, + 339 + ], + "spans": [ + { + "bbox": [ + 236, + 335, + 359, + 339 + ], + "type": "text", + "content": "11. An elephant standing next to a tree with a blue sky behind it." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 236, + 340, + 361, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 340, + 361, + 345 + ], + "spans": [ + { + "bbox": [ + 236, + 340, + 361, + 345 + ], + "type": "text", + "content": "12. Two elephants standing in the water near a and a blue sky" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 236, + 345, + 361, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 345, + 361, + 350 + ], + "spans": [ + { + "bbox": [ + 236, + 345, + 361, + 350 + ], + "type": "text", + "content": "12. Two elephants standing in the grass near a tree and a blue sky 13. Two elephants standing next to each other on a lush green hill" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 236, + 350, + 404, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 350, + 404, + 355 + ], + "spans": [ + { + "bbox": [ + 236, + 350, + 404, + 355 + ], + "type": "text", + "content": "13. Two elephants standing next to each other on a rush green hillside next to a forest." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 236, + 355, + 404, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 355, + 404, + 361 + ], + "spans": [ + { + "bbox": [ + 236, + 355, + 404, + 361 + ], + "type": "text", + "content": "14. Two elephants standing next to a tree in a field with a bluesky and trees behind them." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 236, + 361, + 422, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 361, + 422, + 365 + ], + "spans": [ + { + "bbox": [ + 236, + 361, + 422, + 365 + ], + "type": "text", + "content": "15. Two elephants standing next to each other on a grass covered field next to a lush green forest." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 418, + 272, + 479, + 317 + ], + "blocks": [ + { + "bbox": [ + 418, + 272, + 479, + 317 + ], + "lines": [ + { + "bbox": [ + 418, + 272, + 479, + 317 + ], + "spans": [ + { + "bbox": [ + 418, + 272, + 479, + 317 + ], + "type": "image", + "image_path": "43248957ba4e19f716628cd282245a6048b2429162cb83424e188afead66d5b1.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 419, + 320, + 477, + 362 + ], + "blocks": [ + { + "bbox": [ + 419, + 320, + 477, + 362 + ], + "lines": [ + { + "bbox": [ + 419, + 320, + 477, + 362 + ], + "spans": [ + { + "bbox": [ + 419, + 320, + 477, + 362 + ], + "type": "image", + "image_path": "acb148dad3c5cf0c7a8f9daeaca9b14bb0a2e178027ab9a42c85f2e5549a92f1.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 376, + 482, + 497 + ], + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 497 + ], + "type": "text", + "content": "Fig. 1: Existing captioning datasets contain captions that describe the entirety of an image. This is reflected in the narrow distributions of the entities that appear in those captions and the caption lengths (the red-colored histograms). CIC aims to generate diverse descriptions by controllably re-focusing on different spatiosemantic aspects of an image, such as the semantically coherent subsets of image objects. Our proposed CIC-BART-SSA is designed to produce diverse, controlled captions ranging from brief and concise to detailed and comprehensive. Sentences 1-15 are example outputs of our approach where the highlighted text indicates the focus of a controllable caption. The histograms demonstrate that our approach generates high-quality descriptions for a wider range of scene focus (number of visual entities) and caption length compared to the original captions. Image is licensed under CC BY-SA 2.0." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "bbox": [ + 159, + 514, + 455, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 514, + 455, + 636 + ], + "spans": [ + { + "bbox": [ + 159, + 514, + 455, + 636 + ], + "type": "text", + "content": "Abstract. Controllable Image Captioning (CIC) aims at generating natural language descriptions for an image, conditioned on information provided by end users, e.g., regions, entities or events of interest. However, available image-language datasets mainly contain captions that describe the entirety of an image, making them ineffective for training CIC models that can potentially attend to any subset of regions or relationships. To tackle this challenge, we propose a novel, fully automatic method to sample additional focused and visually grounded captions using a unified structured semantic representation built on top of the existing set of captions associated with an image. We leverage Abstract Meaning Representation (AMR), a cross-lingual graph-based semantic formal" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 643, + 414, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 643, + 414, + 654 + ], + "spans": [ + { + "bbox": [ + 141, + 643, + 414, + 654 + ], + "type": "text", + "content": "*Work done during an internship at Samsung AI Centre - Toronto" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 141, + 654, + 352, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 352, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 352, + 665 + ], + "type": "text", + "content": "†Work done while at Samsung AI Centre - Toronto" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 117, + 455, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 117, + 455, + 261 + ], + "spans": [ + { + "bbox": [ + 159, + 117, + 455, + 261 + ], + "type": "text", + "content": "ism, to encode all possible spatio-semantic relations between entities, beyond the typical spatial-relations-only focus of current methods. We use this Structured Semantic Augmentation (SSA) framework to augment existing image-caption datasets with the grounded controlled captions, increasing their spatial and semantic diversity and focal coverage. We then develop a new model, CIC-BART-SSA, specifically tailored for the CIC task, that sources its control signals from SSA-diversified datasets. We empirically show that, compared to SOTA CIC models, CIC-BART-SSA generates captions that are superior in diversity and text quality, are competitive in controllability, and, importantly, minimize the gap between broad and highly focused controlled captioning performance by efficiently generalizing to the challenging highly focused scenarios. Code is available at https://github.com/SamsungLabs/CIC-BART-SSA." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 281, + 230, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 281, + 230, + 294 + ], + "spans": [ + { + "bbox": [ + 133, + 281, + 230, + 294 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 306, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 482, + 402 + ], + "type": "text", + "content": "Image captioning refers to the task of providing an AI system with an input image, and asking the system to describe the visual content in natural language. This process requires the captioning system to understand what objects are present, in what context (e.g., event or scene), and how they relate. Recent deep learning approaches to this task [14,26,29,31,37,39,40,49,50,57] surpass human performance in standard image captioning metrics. However, these models tend to generate general captions that describe the entirety of an image, and are often of limited diversity; see Original Captions in Fig. 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 402, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 485 + ], + "type": "text", + "content": "Controllable image captioning (CIC) overcomes these challenges by generating different descriptions for the same image in a user-controlled fashion. That is, a CIC model receives as input an image paired with a user-specified control signal (e.g., entities or regions of interest), and generates a caption conditioned on the control signal. CIC models are thus capable of generating a diverse set of captions by varying the control signal for the same image; see CIC generated captions 1-15 in Fig. 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 629 + ], + "type": "text", + "content": "In realistic applications, the easiest way for the user to control the generation of captions is to limit the focus of the desired captions by selecting different entities (objects) using their bounding boxes, as shown in Figs. 1 and 2. Most previous work focuses on such spatial control signals [18, 23, 28, 47, 55, 56]. To improve performance, more recent studies supplement this spatial signal with additional information on the desired length, style, or syntactic and semantic structure of the generated text [12, 13], increasing the richness and complexity of control signals. However, for the CIC approach to succeed, the CIC models need to be trained on equally rich datasets that incorporate, explicitly or implicitly, those control signals. Unfortunately, most image captioning datasets today, such as Flickr30k [36] or MS-COCO [18], lack this necessary diversity of controls and corresponding captions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Our goal is to achieve SOTA performance in CIC without the need for new, increasingly rich, yet also costly, and impractical-to-collect datasets, where human workers would face the burden of having to provide multitudes of control" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "K. Basioti et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 482, + 270 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 270 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 270 + ], + "type": "image", + "image_path": "90e3134b7c2d3a31a77a16c47dddb9977be796c23db20863276cbc3ef1646074.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 277, + 482, + 334 + ], + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 334 + ], + "type": "text", + "content": "Fig. 2: An example of our Structured Semantic Augmentation (SSA) approach. Visually-grounded captions (1)-(5) are used to create a meta-vgAMR graph, which includes all available image information in one representation. Sub-graphs of meta-vgAMR are then sampled to generate a new and diverse set of captions, such as the sentences (a)-(e). Image is licensed under CC BY-SA 2.0." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 366, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 668 + ], + "type": "text", + "content": "signals and corresponding descriptive captions. To achieve this goal, we propose a novel Structured Semantic Augmentation (SSA) method, which automatically generates an augmented set of captions and the corresponding control signals with diverse spatiosemantic focus starting from only the core set of \"original\" uncontrolled captions. The method takes advantage of a detailed visual-linguistic semantic graph (illustrated in Fig. 2) constructed from the original captions and their image groundings. To build these semantic graphs, we use Abstract Meaning Representation (AMR) [6], a semantic formalism that can capture fine-grained linguistic relations beyond the exclusively spatial relationships present in the common scene graphs [24]. The availability of robust AMR parsers [5,9] allows us to generate semantic graphs for individual captions automatically, which we then merge into a rich meta-AMR graph for the joint image-language pair. From this meta-graph, we sample diverse connected subgraphs that represent semantically coherent combinations of image-anchored entities, events, and their relations, which we then turn into controlled captions automatically via existing AMR-to-text models [9]. Fig. 2 depicts an example of our meta-graph inferred from the original uncontrolled captions associated with an image. Filled nodes in the meta-graph indicate image entity groundings. Five semantically coherent subgraphs (a)-(e) of variable complexity are then sampled from the meta-graph, which are subsequently used to generate novel captions, shown below each subgraph. These new captions augment the original caption set by providing both image focus, through node groundings, and increased semantic diversity induced by the sampled subgraphs. Building upon SSA, we introduce a new CIC model, CIC-BART, suitable for generating focused controlled captions. Alongside the regions of interest, CIC-BART also makes use of the length of the desired cap" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "type": "text", + "content": "CIC-BART-SSA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "tion as a control signal proxy for the verosity of the caption. CIC-BART can be trained on SSA-augmented versions of standard VL datasets such as MS-COCO or Flicker30k to accommodate the CIC task. Our experiments show that, compared to several SOTA models, the captions generated by our model have superior text quality and diversity, while being comparable in terms of faithfulness to control signals." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 188, + 288, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 188, + 288, + 199 + ], + "spans": [ + { + "bbox": [ + 131, + 188, + 288, + 199 + ], + "type": "text", + "content": "In summary, our contributions are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 200, + 482, + 392 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 130, + 200, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 260 + ], + "type": "text", + "content": "1. We propose a novel data augmentation technique, SSA, that draws on a structured semantic formalism (AMR) to automatically generate focused captions suitable for training of CIC models. We empirically show that our SSA technique enables CIC models to generate captions with high controllability, diversity, and text quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 260, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 319 + ], + "type": "text", + "content": "2. We propose CIC-BART, a model designed for CIC, that does not require overly descriptive and complex control signals that SOTA models often require to achieve high performance. We show a superior overall performance, compared to SOTA, while relying on simple control signals (i.e., regions of interest and preferred caption length)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 319, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 392 + ], + "type": "text", + "content": "3. We present an extensive evaluation of our model, compared with existing SOTA. Specifically, we report results on different aspects of generated captions, including controllability (faithfulness to control signal), diversity, and text quality (linguistic well-formedness). To account for the trade-off among these metrics, we propose an overall performance score based on their harmonic mean. This metric helps us identify models that perform well in all these aspects." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 411, + 237, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 411, + 237, + 424 + ], + "spans": [ + { + "bbox": [ + 131, + 411, + 237, + 424 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 437, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 641 + ], + "type": "text", + "content": "Controllable Image Captioning (CIC). Various types of control have been used for CIC, including visual entities, a type of region-based control [18, 23, 28, 47, 55, 56], where generated captions should learn to focus on the regions of interest. Others draw on complex control signals where additional knowledge about the generated caption structure is provided. For example, some recent work provides the complete skeleton of the desired sentence in the form of a number of objects or attributes or object-relation-object templates [12, 13]. Additional control signals that CIC draws on include different caption styles, e.g., positive, negative, humorous, or romantic tone [20, 21, 32, 33, 45, 47, 53, 54], user personality [17, 41], or the length of the generated captions [19, 22, 45, 47, 48]. The use of complex control signals aims at improving the diversity of captions and the quality of the text in CIC models. However, it requires the users to provide a detailed description of the control signal, which is not realistic in practical settings where such models are to be deployed (e.g., a self-driving car or personal assistant). We instead draw on two simple control signals (regions of interest and desired caption length) and show that we can achieve competitive performance on CIC, while keeping the control signals simple and practical." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "Recent SOTA models that draw on spatial control include the SCT model [18] that also uses the Faster R-CNN feature vectors and object tags (corresponding" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "K. Basioti et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 481, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 481, + 319 + ], + "type": "text", + "content": "GloVe vectors [35]) of the entities of interest, as well as models that include skeleton-based control, namely ASG2Caption [13] and VSR [12]. ASG2Caption uses an abstract scene graph (ASG) to express the desired structure of a caption. ASG contains three types of unlabeled abstract nodes (object, attribute, relationship) that are grounded in the image by extracting features from the corresponding bounding boxes (for objects and attributes) or from the union of bounding box pairs (for a relationship node). ASG2Caption shows improved controllability (by conditioning on ASGs), and diversity (by automatically sampling diverse ASGs as control signals). The VSR model [12] draws on GloVe embeddings of Faster R-CNN object tags for visual entities (as in SCT). It also uses a skeleton control signal (like ASG2Caption), but one that includes more detailed information and richer semantics. Specifically, the VSR control signal follows the form of a fine-grained PropBank entry" + }, + { + "bbox": [ + 133, + 115, + 481, + 319 + ], + "type": "inline_equation", + "content": "^4" + }, + { + "bbox": [ + 133, + 115, + 481, + 319 + ], + "type": "text", + "content": " — i.e., specifying the exact verb(s) expressing action(s) depicted in the image, and their visually grounded arguments (e.g., subject, object, location, manner). Thus, VSR uses the most descriptive control signal among the SOTA models. Refer to our supplementary material for an illustrative example of the control signal used for each method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 320, + 481, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 320, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 133, + 320, + 481, + 463 + ], + "type": "text", + "content": "Compared to ASG2Caption and VSR, our control signal is kept minimal and only specifies the bounding boxes and desired caption lengths. To improve the diversity of captions, we draw on a structured semantic graph (AMR) that expresses the semantics of a sentence based on PropBank semantics. Notably, we do not use these rich graphs to express detailed and overly descriptive control signals (as in VSR), but we use these semantic structures to augment our training data with richer and diverse captions, which will result in the model learning to generate more diverse captions. Additionally, we include a length control signal to further increase diversity without needing to specify detailed information about the structure of the output (e.g., number of attributes per object, etc.). This way, we can generate a variety of captions for a fixed image sub-region by simply controlling the desired length of the output." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 464, + 481, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 464, + 481, + 570 + ], + "spans": [ + { + "bbox": [ + 133, + 464, + 481, + 570 + ], + "type": "text", + "content": "Abstract Meaning Representation (AMR). AMR [7] is a rich semantic formalism for expressing the meaning of natural language sentences as a formal graph. AMR draws on PropBank, which is a rich lexical semantic resource encoding predicates expressing an action or state, as well as the number and nature of the participating entities (arguments and other semantic roles, such as location, manner, etc.). AMR is a widely researched semantic formalism for which highly accurate automatic Text-to-AMR and AMR-to-Text models are developed [5,9]. We rely on these models to augment original image-caption datasets with newly generated captions (as explained in Sect. 4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 571, + 481, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 571, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 571, + 481, + 643 + ], + "type": "text", + "content": "AMRs vs. Scene Graphs. Recent studies [1,15,16,52] have shown that AMRs better capture the semantic relations of an image as compared to the scene graphs [11]. Existing scene graph annotations mainly capture geometric or possessive relations, which account for more than " + }, + { + "bbox": [ + 133, + 571, + 481, + 643 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 133, + 571, + 481, + 643 + ], + "type": "text", + "content": " of the relations captured, whereas more than " + }, + { + "bbox": [ + 133, + 571, + 481, + 643 + ], + "type": "inline_equation", + "content": "1/3" + }, + { + "bbox": [ + 133, + 571, + 481, + 643 + ], + "type": "text", + "content": " of the captured entities refer to clothing, object, or body parts information [1,52]. This difference is crucial for high-quality image caption-" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 380, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 380, + 90, + 447, + 101 + ], + "type": "text", + "content": "CIC-BART-SSA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 134, + 653, + 266, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 653, + 266, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 653, + 266, + 665 + ], + "type": "text", + "content": "4 https://propbank.github.io" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "ing, as we use higher-level semantic relations in our everyday language rather than geometric ones. For instance, during a soccer game, we would probably describe a goal save as 'the player kicks the ball away from the goal' or 'the goalkeeper defends his team by saving a goal' and not by using mainly geometric and possessive relations like 'a person wearing a white shirt, standing with his right leg lifted, close to a ball which is above the ground'. In the supplementary, we provide a detailed comparison of AMR and scene graph representations, particularly focusing on their applications in data augmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 228, + 194, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 194, + 242 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 194, + 242 + ], + "type": "text", + "content": "3 Model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 253, + 298, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 253, + 298, + 408 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 298, + 408 + ], + "type": "text", + "content": "We propose CIC-BART, specifically designed to generate controlled image captions. Specifically, it can generate descriptions of particular areas within a scene with a desired level of detail. Our model, based on VL-BART [14], utilizes a transformer-based encoder-decoder architecture, as shown in Fig. 3. CIC-BART extends VL-BART encoder to the CIC task by modifying the encoder input to include: a) a global image embedding that provides the context of the" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 288, + 478, + 342 + ], + "blocks": [ + { + "bbox": [ + 310, + 288, + 478, + 342 + ], + "lines": [ + { + "bbox": [ + 310, + 288, + 478, + 342 + ], + "spans": [ + { + "bbox": [ + 310, + 288, + 478, + 342 + ], + "type": "image", + "image_path": "5e17183f1d38b7c2a689d07aecff75c9c2478c6b41ce55d737590e0c60c794d6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 360, + 482, + 383 + ], + "lines": [ + { + "bbox": [ + 304, + 360, + 482, + 383 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 482, + 383 + ], + "type": "text", + "content": "Fig.3: The architecture diagram of our model, CIC-BART." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 409, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 409, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 482, + 468 + ], + "type": "text", + "content": "full image to the model; b) the visual control signal, including the visual embedding of the regions that contain the entities of interest; c) the text control signal, containing length control (indicating the desired length range of the output caption) and an optional verb signal that indicates the action we want the generated caption to concentrate on." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 468, + 482, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 588 + ], + "type": "text", + "content": "The visual embeddings of the regions are position-aware embeddings from a Faster R-CNN model [38] trained for visual object and attribute classification [3] on Visual Genome. The global image feature vector is extracted as well from Faster R-CNN. For the length control signal, we add to our vocabulary " + }, + { + "bbox": [ + 130, + 468, + 482, + 588 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 468, + 482, + 588 + ], + "type": "text", + "content": " tokens for the " + }, + { + "bbox": [ + 130, + 468, + 482, + 588 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 468, + 482, + 588 + ], + "type": "text", + "content": " different caption length levels; for instance, level one represents sentences between one and nine words, and level two, ten to nineteen. These tokens describe our coarse levels, for a finer sentence size accuracy, we accompany the tokens with the desired number of words. This choice gives our model the capacity to generate diverse captions for a particular length level. Finally, the output of the decoder generates the desired, controlled image caption." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 605, + 404, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 404, + 620 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 404, + 620 + ], + "type": "text", + "content": "4 Structured Semantic Augmentation (SSA)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "The goal of our SSA method is to augment existing image captioning datasets with new focused captions along with their control signals (i.e., regions corresponding to entities). We rely on datasets where visual entities in the captions" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "text", + "content": "K. Basioti et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "content": "are annotated with their corresponding regions (see Sect. 5 for details on the datasets). The SSA process consists of four main steps, as described below. For more details, refer to our supplementary material, which also includes a step-by-step example of our SSA methodology." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 481, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 481, + 331 + ], + "type": "text", + "content": "Step 1: Image-level AMR graph generation. Our objective in this stage is to enclose all the information available from the visually grounded captions into a single representation. To accomplish this, we create a visually grounded AMR graph (vgAMR) for each caption of an image and then merge them into a single image-level graph, the meta-vgAMR. To create the vgAMRs of an image, we first convert each of its " + }, + { + "bbox": [ + 130, + 164, + 481, + 331 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 164, + 481, + 331 + ], + "type": "text", + "content": " captions to their AMR representation, using the Neural transition-based Text-to-AMR parser [5] which also aligns words in a caption with their respective nodes in the AMR graph. We utilize the alignment information of 1) caption words and AMR nodes (from Text-to-AMR parser) and 2) caption words to image bounding boxes (from existing dataset annotations) to visually ground the AMR nodes. After this step, we get the collection of nodes referring to visual entities, where each grounded meta-AMR node is linked with a non-empty set of bounding boxes. This extended representation, 'AMR + visual grounded nodes', is our vgaMR." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 331, + 481, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 481, + 415 + ], + "type": "text", + "content": "Our next step is to combine the N vgAMRs to form a single meta-vgAMR. To achieve this, we employ a pairwise strategy to merge the most similar vgAMRs first (we measure similarity with Smatch score [10]). We use the UPGMA hierarchical clustering algorithm [30,34] to find the optimal merge ordering starting from the most similar graphs. UPGMA creates a hierarchy where the bottom level consists of the N individual vgAMRs. By merging all vgAMRs using the UPGMA ordering, we obtain a single structure called meta-vgAMR." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 415, + 481, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 481, + 631 + ], + "type": "text", + "content": "When merging two vgAMRs, the main challenge is identifying which nodes correspond to the same concepts, such as entities, attributes, actions, and relations. We use three node properties to accomplish this: a) visual grounding information, b) semantic similarity of node labels, and c) node neighborhood semantic similarity. We derive two node-merging criteria from there: 1) visually grounded entity nodes are merged if they point to the same image-bounding boxes. When 1) does not hold, we check the second criterion: 2) for the remaining non-grounded nodes, including amr-specific, predicates, adjectives, and adverbs, we use a combination of node label semantic similarity (cosine similarity of the labels using their GloVe embeddings) and neighborhood similarity. Neighborhood similarity examines the similarity of parents for adjectives/adverbs nodes and children for predicate nodes, along with the similarity of connecting edge roles. When two nodes satisfy criterion 1) or 2), we merge them into a single node. Moreover, if they have different labels, we maintain both names by keeping a list of synonyms to increase representation diversity. In the special case when the two vgAMRs describe two totally different concepts, and hence they have no common nodes, we add an amr-specific node called 'multi-sentence' as the root with the two independent vgAMRs as its children. The final graph, meta" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 380, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 380, + 90, + 447, + 101 + ], + "type": "text", + "content": "CIC-BART-SSA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 481, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 481, + 152 + ], + "type": "text", + "content": "vgAMR, includes all non-redundant5 elements of the original " + }, + { + "bbox": [ + 130, + 115, + 481, + 152 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 115, + 481, + 152 + ], + "type": "text", + "content": " captions while preserving the visual grounding between the meta nodes and their respective image regions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 481, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 481, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 481, + 187 + ], + "type": "text", + "content": "Remark: Meta-vgAMR efficiently compresses all available image information into a single structure. Following our approach, we can easily scale when new scene information becomes available by applying our pairwise merge procedure." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 188, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 426 + ], + "type": "text", + "content": "Step 2: Event-based graph sampling from image-level AMRs. We start from the predicate nodes, which mainly correspond to verbs, to sample subgraphs in meta-vgAMR graphs. Predicate nodes are identified by their label and the edges connected to them. The label of a predicate node typically follows the format 'predicate_name-xx,' where 'xx' represents the different senses a word can have regarding the concept it is used for. Predicate nodes have outward ARGy edges, where 'y' can take values from 0 to 5, connecting them to their arguments. We sample subgraphs from these nodes by following the outgoing argument edges, which are labeled as ARGn in an AMR graph, each defining a particular semantic role (e.g., ARG0 points to the agent, ARG1 to the patient, etc.). Finally, we add one more subgraph containing the remaining children branches of other non-ARG optional predicate edges (e.g., 'location', 'time'). We repeat this process until the leaves of the graphs are reached. During sampling, we randomly select one of the synonyms if a node is a list of synonym labels, as mentioned in the previous step. The output of this step is our more focused event-focused sub-graphs. In Fig. 2, we can see some instances of our event-based sampling (SSA samples), where the predicate nodes include z0/sit-01, z7/dock-01, z13/sit-01, and so on6. Although we cannot show all the sampled event-based subgraphs in the figure, we included five of them and used colored roots and edges for visualization purposes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 427, + 483, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 483, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 483, + 498 + ], + "type": "text", + "content": "Step 3: New caption generation from sampled AMRs We use the SPRING AMR-to-Text model [8] to generate new event-focused captions from the sampled vgAMR subgraphs. Because both vgAMR merging and sampling steps introduce noise, the output captions are not always of good quality. We automatically filter low-quality captions by using a linguistic well-formedness measure, GRUEN [58], which is a reference-free metric based on BERT contextual embeddings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 481, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 547 + ], + "type": "text", + "content": "Step 4: Control signal generation. The last step is to create the control signal for the generated captions. The spatial control signal for a specific caption is extracted from the corresponding sampled vgAMR, by pulling the bounding boxes of the visual entity linked AMR nodes." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "K. Basioti et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 555, + 481, + 609 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 555, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 555, + 481, + 609 + ], + "type": "text", + "content": "5 A node may have different names for the same bounding box in different meta-vgAMRs, such as 'A male' and 'A person'. According to criterion 1), we merge the corresponding AMR nodes and keep both 'male' and 'person' in the names list to avoid redundancy. Therefore, criteria 1) and 2) ensure that multiple nodes don't describe the same concept in the meta-vgAMR." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 610, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 610, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 610, + 481, + 665 + ], + "type": "text", + "content": "Note that in Fig. 2, the node z17/green-03 is also categorized as a predicate. This may seem an error because we usually think of 'green' as an attribute node rather than a predicate. However, in AMRs, when 'green' is paired with its argument, in this case, z16/lawn, it encapsulates a predicate/verb that can be expressed in natural language as 'the lawn is green.'" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 384, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 384, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 384, + 129 + ], + "type": "text", + "content": "4.1 Mixing Strategies of Original and SSA Data" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "text", + "content": "To analyze the impact of our SSA data, we explore various mixing strategies with the original training set. Assume " + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "text", + "content": " represents the training control-caption pairs in the original dataset, containing " + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{D}}" + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "text", + "content": " samples, and SSA represents our SSA samples, containing " + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "inline_equation", + "content": "N_{SSA}" + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "text", + "content": " instances. The augmented dataset " + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{SSA}" + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "text", + "content": " is defined by combining " + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "text", + "content": " and SSA: " + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{SSA} = \\mathrm{sam}_{\\mathcal{D}}(\\tau_{\\mathcal{D}}, p_{\\mathcal{D}}) \\cup \\mathrm{sam}_{SSA}(\\tau_{SSA}, p_{SSA})" + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "text", + "content": ", where the functions " + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\mathrm{sam}_{\\mathcal{D}}" + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "text", + "content": " samples a subset of the original dataset, and " + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\mathrm{sam}_{SSA}" + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "text", + "content": " a subset of our SSA data. Since we are interested in the effect of our SSA, we assume that " + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\mathrm{sam}_{\\mathcal{D}}(\\tau_{\\mathcal{D}}, p_{\\mathcal{D}}) = \\mathcal{D}" + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathcal{D}} = \\text{'Random Sampling Strategy'}" + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "inline_equation", + "content": "p_{\\mathcal{D}} = 100\\%" + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "text", + "content": ", meaning that all original data are included in the mixed dataset. Depending on the " + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\mathrm{sam}_{SSA}" + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "text", + "content": " parameter " + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\tau_{SSA}" + }, + { + "bbox": [ + 130, + 139, + 482, + 258 + ], + "type": "text", + "content": " we have the cases:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 258, + 482, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 305 + ], + "type": "text", + "content": "Random Sampling Strategy. In this case, we randomly select a pre-specified number of examples from " + }, + { + "bbox": [ + 130, + 258, + 482, + 305 + ], + "type": "inline_equation", + "content": "SSA" + }, + { + "bbox": [ + 130, + 258, + 482, + 305 + ], + "type": "text", + "content": ". The parameter " + }, + { + "bbox": [ + 130, + 258, + 482, + 305 + ], + "type": "inline_equation", + "content": "p_{SSA}" + }, + { + "bbox": [ + 130, + 258, + 482, + 305 + ], + "type": "text", + "content": " expresses the percentage of SSA samples included in " + }, + { + "bbox": [ + 130, + 258, + 482, + 305 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{SSA}" + }, + { + "bbox": [ + 130, + 258, + 482, + 305 + ], + "type": "text", + "content": ". With boundary cases " + }, + { + "bbox": [ + 130, + 258, + 482, + 305 + ], + "type": "inline_equation", + "content": "p_{SSA} = 100\\%" + }, + { + "bbox": [ + 130, + 258, + 482, + 305 + ], + "type": "text", + "content": " (all " + }, + { + "bbox": [ + 130, + 258, + 482, + 305 + ], + "type": "inline_equation", + "content": "N_{SSA}" + }, + { + "bbox": [ + 130, + 258, + 482, + 305 + ], + "type": "text", + "content": " samples are included), and " + }, + { + "bbox": [ + 130, + 258, + 482, + 305 + ], + "type": "inline_equation", + "content": "p_{SSA} = 0\\%" + }, + { + "bbox": [ + 130, + 258, + 482, + 305 + ], + "type": "text", + "content": " (no SSA data are added)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 306, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 482, + 415 + ], + "type": "text", + "content": "Uniform-Coverage Sampling Strategy. To mitigate the original dataset's bias (having mainly samples describing the entire image), we aim to create a new focus-unbiased dataset. By modeling the control signal focus as the image area percentage covered by the bounding boxes of the control signal, we split the original data into " + }, + { + "bbox": [ + 130, + 306, + 482, + 415 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 306, + 482, + 415 + ], + "type": "text", + "content": " coverage bins. Then, we will randomly add in each bin SSA samples, aiming to create a new uniform, coverage-unbiased " + }, + { + "bbox": [ + 130, + 306, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{SSA}" + }, + { + "bbox": [ + 130, + 306, + 482, + 415 + ], + "type": "text", + "content": " dataset. Here, " + }, + { + "bbox": [ + 130, + 306, + 482, + 415 + ], + "type": "inline_equation", + "content": "p_{SSA}" + }, + { + "bbox": [ + 130, + 306, + 482, + 415 + ], + "type": "text", + "content": " contains the range of each bin for the coverage histogram. For example, in the case where we choose ten uniform coverage bins, we have " + }, + { + "bbox": [ + 130, + 306, + 482, + 415 + ], + "type": "inline_equation", + "content": "p_{SSA} = \\{[0\\%, 10\\%), [10\\%, 20\\%), \\ldots, [90\\%, 100\\%]\\}" + }, + { + "bbox": [ + 130, + 306, + 482, + 415 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 415, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 450 + ], + "type": "text", + "content": "Due to space limitations, we present results from the Random Sampling Strategy for " + }, + { + "bbox": [ + 130, + 415, + 482, + 450 + ], + "type": "inline_equation", + "content": "p_{SSA} = 0\\%" + }, + { + "bbox": [ + 130, + 415, + 482, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 415, + 482, + 450 + ], + "type": "inline_equation", + "content": "p_{SSA} = 100\\%" + }, + { + "bbox": [ + 130, + 415, + 482, + 450 + ], + "type": "text", + "content": " in the main paper. Results from other scenarios can be found in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 472, + 272, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 272, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 272, + 485 + ], + "type": "text", + "content": "5 Experimental Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 499, + 187, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 499, + 187, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 499, + 187, + 510 + ], + "type": "text", + "content": "5.1 Data" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "We use Flickr30k Entities (Flickr-Ent) [36] and MS-COCO Entities (COCO-Ent) [18] for training and evaluation. Flickr-Ent augments the original captions of Flickr30k [51] with manually-annotated region-phrase groundings. Flickr-Ent contains the original " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "31K" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " images annotated with five captions each. COCO-Ent augments the original MS-COCO [27] (" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "120K" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " images each annotated with around five captions) with semi-automatically collected grounding annotations; see [18] for details on the annotation process. For both datasets, we follow previous work and use the training and test splits by Karpathy et al. [25]. We apply our SSA algorithm on the aforementioned datasets to create their augmented variations, COCO-Ent-SSA and Flickr-Ent-SSA, containing about " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "800K" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "250K" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " training captions, respectively, of which " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "33\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "37\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", are generated by our SSA algorithm." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "type": "text", + "content": "CIC-BART-SSA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "For all four training sets, we automatically generate image-control-caption triplets to train our model on. For spatial control, we extract from the grounded captions the bounding boxes of the entities of interest using the annotations from COCO-Ent and Flickr-Ent. Note that since these datasets do not contain the text control signal, we use each ground-truth caption as a proxy for a controlled caption, from which we first generate the coarse- and fine-length control levels and then extract their verbs using part-of-speech tagging for the optional action control." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 228, + 319, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 228, + 319, + 240 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 319, + 240 + ], + "type": "text", + "content": "5.2 Models and Evaluation Metrics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 481, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 481, + 354 + ], + "type": "text", + "content": "We compare two variations of our model (with and without SSA augmentations) with SOTA models as our baselines: Show Control & Tell (SCT) [18] that uses region-based control (bounding boxes of visual entities of interest); ASG2Caption (ASG) [13] that draws on visually grounded abstract scene graphs as control signal; and VSR [12] that uses overly descriptive control signals that express verb(s) and fine-grained verb-specific semantic roles of the desired captions; ComPro [48] that learns a mapping from the bounding boxes of the entities of interest and caption length level to GPT-2 Large prompts aiming to retrieve controlled captions; and the LaBERT length-control-only model [19]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 355, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 355, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 355, + 482, + 668 + ], + "type": "text", + "content": "We report the performance of our models and baselines using a comprehensive set of metrics that evaluate different aspects of caption controllability, diversity, and quality. We also propose and report an overall performance metric that summarizes these different aspects in a meaningful way. To measure diversity, we compute n-gram diversity, D-n for " + }, + { + "bbox": [ + 130, + 355, + 482, + 668 + ], + "type": "inline_equation", + "content": "n = 1" + }, + { + "bbox": [ + 130, + 355, + 482, + 668 + ], + "type": "text", + "content": ", 2 [4], as well as self-CIDEr-based diversity (sC) [46]. For a fair comparison of the different CIC models, we measure diversity for the five generated captions for each test image (in COCO-Ent and Flickr-Ent), and report their average. To measure content controllability we design an extended version of the IoU metric of [18] that calculates the degree-of-match (faithfulness) between a control signal and the corresponding generated caption. For our control signal, we use the set of nouns " + }, + { + "bbox": [ + 130, + 355, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 355, + 482, + 668 + ], + "type": "text", + "content": " that represent the entities of interest, which are the names of the visual objects in the control. To extract nouns from the predicted sentences, we use the Stanford part-of-speech tagger [42,43]. We then find the semantic intersection of the two sets using Hungarian Matching, as in [18]. Finally, we calculate the semantic intersection over union of the control nouns and the nouns extracted from the controllable caption, which gives us our content controllability IoU. We further analyze IoU by introducing the Hallucinating Nouns (Hal) metric. Details can be found in the supplemental material. For length controllability (L), we measure the Mean Absolute Error (MAE) between the fine length control (number of words) and the size of the resulting " + }, + { + "bbox": [ + 130, + 355, + 482, + 668 + ], + "type": "inline_equation", + "content": "M = 10" + }, + { + "bbox": [ + 130, + 355, + 482, + 668 + ], + "type": "text", + "content": " controlled captions, which are generated from " + }, + { + "bbox": [ + 130, + 355, + 482, + 668 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 355, + 482, + 668 + ], + "type": "text", + "content": " randomly created control signals. We also calculate the length precision (LP) [19] by determining the percentage of generated captions that match the desired coarse length level. We assess text quality of generated captions using GRUEN (G) [58], a reference free metric based on BERT contextual embeddings that measures the syntactic and semantic well-formedness of a text segment. Finally, we measure" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "K. Basioti et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 146, + 477, + 216 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 1: Performance of CIC models for the original test sets. *ASG-type dataset is not released for Flickr-Ent; therefore, we could not reproduce the ASG results." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 146, + 477, + 216 + ], + "lines": [ + { + "bbox": [ + 134, + 146, + 477, + 216 + ], + "spans": [ + { + "bbox": [ + 134, + 146, + 477, + 216 + ], + "type": "table", + "html": "
ModelH↑IoU↑G↑sC↑D-1↑D-2↑L↓H↑IoU↑G↑sC↑D-1↑D-2↑L↓
SCT [18]55.867.364.442.827.035.5-54.650.779.844.029.336.5-
ASG* [13]74.272.672.078.337.856.6--------
VSR [12]56.277.639.067.430.042.2-62.560.254.077.933.349.3-
CIC-BART75.976.273.078.738.056.2.4969.854.085.078.643.658.21.24
CIC-BART-SSA78.377.274.882.544.663.2.1171.355.086.081.747.062.61.05
", + "image_path": "c025e3ee772d999ecb30177ed6ba071d34823eef49f89cfa508ea69526047013.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 138, + 246, + 472, + 300 + ], + "blocks": [ + { + "bbox": [ + 173, + 225, + 440, + 237 + ], + "lines": [ + { + "bbox": [ + 173, + 225, + 440, + 237 + ], + "spans": [ + { + "bbox": [ + 173, + 225, + 440, + 237 + ], + "type": "text", + "content": "Table 2: Performance of CIC models, for the SSA-only samples." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 246, + 472, + 300 + ], + "lines": [ + { + "bbox": [ + 138, + 246, + 472, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 472, + 300 + ], + "type": "table", + "html": "
ModelH↑IoU↑G↑sC↑D-1↑D-2↑H↑IoU↑G↑sC↑D-1↑D-2↑
COCO-Ent (SSA only)Flickr-Ent (SSA only)
SCT [18]51.762.164.837.823.731.043.929.977.345.731.036.7
CIC-BART69.261.473.974.044.257.068.553.080.579.852.962.9
CIC-BART-SSA75.665.280.783.753.867.872.055.682.986.156.569.3
", + "image_path": "a3a790e519776ac2eb7921d99b905a9c3163112d17015e5be2bcdef583b76290.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 320, + 482, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 417 + ], + "type": "text", + "content": "the overall performance of each model based on its ability to balance content controllability, diversity, and text quality. To calculate this, we use the harmonic mean of IoU, G, and sC. All of these metrics range between 0 and 1, with a higher value indicating better performance. The harmonic mean " + }, + { + "bbox": [ + 130, + 320, + 482, + 417 + ], + "type": "inline_equation", + "content": "(H)" + }, + { + "bbox": [ + 130, + 320, + 482, + 417 + ], + "type": "text", + "content": " helps us determine the model with the best overall performance. It prioritizes models that perform well across all metrics while penalizing those with poor performance, even in one metric. In our supplementary, we include comparisons of the CIC models on standard captioning metrics (like CIDEr [44] and Spice [2])." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 433, + 200, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 433, + 200, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 433, + 200, + 445 + ], + "type": "text", + "content": "6 Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 456, + 265, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 265, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 265, + 467 + ], + "type": "text", + "content": "6.1 Overall Performance" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 474, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 521 + ], + "type": "text", + "content": "We first compare the overall performance of our models with the SCT, ASG, and VSR baselines with respect to controllable captioning metrics. We do not include ComPro in this comparison due to the unavailability of the codebase. We also exclude LaBERT since this model solely focuses on length controllability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 522, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 641 + ], + "type": "text", + "content": "Tab. 1 presents results for content and length controllability (IoU and L), text quality (G), and diversity (sC, D-1, D-2), as well as the harmonic mean (H). For both datasets, CIC-BART-SSA has the best performance in all metrics, except IoU, where it is the second best. Specifically, CIC-BART-SSA is superior to all other models with respect to diversity (sC, D-1, D-2) and text quality (G), but comparable to VSR in terms of content controllability (IoU). The length controllability (L) scores show that our SSA augmentation helps the model learn to generate high-quality output at the desirable length (compare CIC-BART and CIC-BART-SSA). This is due to the increased diversity in caption length provided by our SSA augmentations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 642, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 642, + 481, + 666 + ], + "type": "text", + "content": "Importantly, we can see that model performance can vary depending on the metric. E.g., whereas VSR has the highest IoU, it falls behind in text quality and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "type": "text", + "content": "CIC-BART-SSA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "diversity. In our qualitative analysis, we observe the poor quality of the captions generated by VSR. The best-performing model should be identified based on the " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " score that summarizes content controllability, text quality, and diversity into a single score. Based on this score, CIC-BART-SSA is better than SCT and VSR baselines by a large margin, and notably better than ASG. Nevertheless, ASG requires complex control signals in the form of scene graphs, in contrast to the simple control signal requirements of CIC-BART." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 343 + ], + "type": "text", + "content": "Next, we conduct a further evaluation of the CIC performance on our SSA samples from the test set images of COCO-Ent and Flickr-Ent. We present the results in Tab. 2 for our models CIC-BART, CIC-BART-SSA, and SCT. ASG and VSR are excluded since they need complex control signals (grounded abstract scene graphs for ASG and grounded verb semantic roles for VSR), which are only available for COCO-Ent and Flickr-Ent. We observe a significant improvement in overall performance for our CIC-BART-SSA model. We notice that the models (SCT and CIC-BART) trained on the original datasets, which describe the entire image, had difficulties generalizing to cases where they had to focus on a specific sub-region of an image. However, our model CIC-BART-SSA was able to generate focused and diverse descriptions of the challenging, highly focused examples present in our SSA data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 343, + 330, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 330, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 330, + 498 + ], + "type": "text", + "content": "We compare the length precision of our model with the baselines utilizing length control in Table 3. LaBERT uses only length-control signals without spatial control, while our model employs both spatial and length-control signals to generate focused captions. This makes the LaBERT task much easier since it only focuses on generating specific length descriptions of an image. On the other hand, our model focuses on generating captions that describe only a specific sub-region of the scene while maintaining a desired description length level. Although our task is" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 337, + 411, + 480, + 472 + ], + "blocks": [ + { + "bbox": [ + 336, + 378, + 482, + 411 + ], + "lines": [ + { + "bbox": [ + 336, + 378, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 336, + 378, + 482, + 411 + ], + "type": "text", + "content": "Table 3: Length Precision (LP) for CIC models on COCO-Ent and Flickr-Ent original test sets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 411, + 480, + 472 + ], + "lines": [ + { + "bbox": [ + 337, + 411, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 337, + 411, + 480, + 472 + ], + "type": "table", + "html": "
ModelLP↑LP↑
COCO-EntFlickr-Ent
ComPro [48]94.781.4
LaBERT [19]99.798.4
CIC-BART99.988.0
CIC-BART-SSA99.991.3
", + "image_path": "863a3c32efcf233ccfb84cde7a9d1a51fb702e297d4f1a2ad386763d4acd31a5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 499, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 499, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 499, + 482, + 583 + ], + "type": "text", + "content": "more challenging than LaBERT, we achieve competitive length precision performance. Lastly, we want to emphasize that the remarkable improvement in length controllability (L) and length precision (LP) from CIC-BART to CIC-BART-SSA stems from the increased length diversity found in our SSA augmentations, which enriches the original COCO-Ent and Flickr-Ent datasets. In the supplementary, we provide an analysis of the caption length statistics in the original datasets and our SSA-derived captions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 598, + 365, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 598, + 365, + 611 + ], + "spans": [ + { + "bbox": [ + 131, + 598, + 365, + 611 + ], + "type": "text", + "content": "6.2 Effect of SSA on Content Controllability" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "To analyze the impact of our SSA augmentations, we measure the content controllability (IoU) performance of CIC-BART at different levels of focus of the control signals and report it in Fig. 4. We use coverage, defined as the area of the image enclosed by the bounding boxes of the entities of interest in the control" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "text", + "content": "K. Basioti et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 130, + 114, + 315, + 187 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 315, + 187 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 315, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 315, + 187 + ], + "type": "image", + "image_path": "df739c11ae44e09b08c33caa8eeb951bd62d3267c34302cd30000cb3e0134f4a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 323, + 113, + 481, + 187 + ], + "blocks": [ + { + "bbox": [ + 323, + 113, + 481, + 187 + ], + "lines": [ + { + "bbox": [ + 323, + 113, + 481, + 187 + ], + "spans": [ + { + "bbox": [ + 323, + 113, + 481, + 187 + ], + "type": "image", + "image_path": "d2aebe78818bc5dbd0a8fec51790e1acc47c64a6dde0b9922686aa031aaf33ef.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 130, + 193, + 315, + 254 + ], + "blocks": [ + { + "bbox": [ + 130, + 193, + 315, + 254 + ], + "lines": [ + { + "bbox": [ + 130, + 193, + 315, + 254 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 315, + 254 + ], + "type": "image", + "image_path": "7de535f4d8e4e743029c56f016ee5af8bc6d829a7c4b8659a989fc65da71bdd5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 262, + 482, + 307 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 307 + ], + "type": "text", + "content": "Fig. 4: IoU histograms for CIC-BART (green) and CIC-BART-SSA (blue). The first column depicts the IoU on the original test set and the second on the original test set images' SSA (only) data. The Samples curve (orange) represents the distribution of test captions in each coverage (image area covered by the control signal) interval." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 323, + 193, + 481, + 254 + ], + "blocks": [ + { + "bbox": [ + 323, + 193, + 481, + 254 + ], + "lines": [ + { + "bbox": [ + 323, + 193, + 481, + 254 + ], + "spans": [ + { + "bbox": [ + 323, + 193, + 481, + 254 + ], + "type": "image", + "image_path": "2a1a95078e415f9689cb7116e18e8c9ccedce9b3f373eac72de65623bbb6e1b4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 132, + 331, + 282, + 441 + ], + "blocks": [ + { + "bbox": [ + 132, + 331, + 282, + 441 + ], + "lines": [ + { + "bbox": [ + 132, + 331, + 282, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 282, + 441 + ], + "type": "image", + "image_path": "3fd65594618cb87496e6f6382f1a7b5649d323fe7e3788e11c718565d96629ea.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 449, + 481, + 471 + ], + "lines": [ + { + "bbox": [ + 131, + 449, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 131, + 449, + 481, + 471 + ], + "type": "text", + "content": "Fig. 5: Qualitative examples for the original test sets. Strikethrough marks hallucinations and redundancies. a, b licensed under CC BY-SA 2.0; c, d under CC BY 2.0." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 283, + 332, + 481, + 440 + ], + "blocks": [ + { + "bbox": [ + 283, + 332, + 481, + 440 + ], + "lines": [ + { + "bbox": [ + 283, + 332, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 283, + 332, + 481, + 440 + ], + "type": "image", + "image_path": "fa718370c4bfe476d621f99eb49475122f4ae6f87f22bbbc60f76206f450ed1c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "signal, to quantify that focus. For example, highly focused control signals cover a small area, yielding low coverage, while broader signals cover a larger area and have high coverage. We 'break down' the IoU performance into 10 coverage bands and report the average IoU over control signals in those bands. In addition, the 'Samples' curve shows the distribution of test captions over the same bands. The results in Fig. 4 indicate that by training with SSA (blue bars), the spatial controllability improves significantly in the low-coverage regime, where the control signals are highly focused. Interestingly, these are also the most underrepresented (data deprived) parts of the original dataset Flickr-Ent. Therefore, SSA which enriches the original datasets with highly focused examples (please refer to our supplementary for % Samples per coverage bands for the training sets), is effective in improving generalization performance in CIC. We include the performance of hallucinating nouns metric in the supplementary material." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "type": "text", + "content": "CIC-BART-SSA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 263, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 263, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 263, + 129 + ], + "type": "text", + "content": "6.3 Qualitative Analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 299, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 299, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 299, + 376 + ], + "type": "text", + "content": "In Fig. 5, we present qualitative examples from the original test sets, and in Fig. 6 examples from our SSA (only) test set control signals. In the two figures, each highlighted word found in the generated controlled captions corresponds to the control entity of the same color. This shows the match between the captions produced and the control signal. We also strike through the parts where the model hallucinates or generates redundant references to the entities of interest. Our models have been observed to outperform the previous state-of-the-art models by substantially enhancing the quality of the generated controlled captions. This behavior was expected from our quantitative analysis, which showed that our models have signifi" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 377, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 426 + ], + "type": "text", + "content": "cantly higher text quality (G). More importantly, our CIC-BART-SSA model is capable of generating captions that are faithful to the control signal and better understand the relationships that connect the entities of interest. We include additional qualitative samples in the Supplementary Material." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 173, + 478, + 291 + ], + "blocks": [ + { + "bbox": [ + 309, + 173, + 478, + 291 + ], + "lines": [ + { + "bbox": [ + 309, + 173, + 478, + 291 + ], + "spans": [ + { + "bbox": [ + 309, + 173, + 478, + 291 + ], + "type": "image", + "image_path": "6eaa26b9b592b75b60512fc99bcec7e07ee58b015f5d94ade8de858079ea21d2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 308, + 482, + 353 + ], + "lines": [ + { + "bbox": [ + 304, + 308, + 482, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 482, + 353 + ], + "type": "text", + "content": "Fig. 6: Examples for SSA-only test set. Strikethrough marks hallucinations. Up to Down: a licensed under CC BY 2.0; b under CC PDM 1.0, for last two see Fig. 5." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 446, + 226, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 226, + 459 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 226, + 459 + ], + "type": "text", + "content": "7 Conclusions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 473, + 483, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 483, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 483, + 668 + ], + "type": "text", + "content": "We address two main challenges faced by the controllable image captioning (CIC) models. First, standard image-caption datasets lack the controllability and diversity needed for proper training and evaluation of CIC. Second, most recent SOTA models require complex and overly descriptive control signals as input (including, e.g., the main action/verb to appear in the generated caption). To address the first challenge, we propose a novel technique that draws on a structured semantic augmentation (SSA) formalism to generate focused captions and the corresponding control signals for images. For the second challenge, we propose a transformer-based vision-language model attuned to the CIC task. We show that this model performs competitively with SOTA models without requiring complex and explicit control signals. Importantly, when combined with our SSA approach, our model generates highly diverse captions and significantly reduces the content controllability performance gap between the different levels of focus of the generated controlled captions. Finally, when provided with the commonly used verb guidance of other SOTA approaches, our model shows a substantial improvement in performance." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "text", + "content": "K. Basioti et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 142, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 142, + 481, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 481, + 186 + ], + "type": "text", + "content": "1. Abdelsalam, M.A., Shi, Z., Fancellu, F., Basioti, K., Bhatt, D., Pavlovic, V., Fazly, A.: Visual semantic parsing: From images to abstract meaning representation. In: Proceedings of the 26th Conference on Computational Natural Language Learning (CoNLL). pp. 282-300 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 187, + 481, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 187, + 481, + 231 + ], + "spans": [ + { + "bbox": [ + 138, + 187, + 481, + 231 + ], + "type": "text", + "content": "2. Anderson, P., Fernando, B., Johnson, M., Gould, S.: Spice: Semantic propositional image caption evaluation. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14. pp. 382-398. Springer (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 232, + 481, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 481, + 264 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 481, + 264 + ], + "type": "text", + "content": "3. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: CVPR (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 264, + 481, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 264, + 481, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 264, + 481, + 298 + ], + "type": "text", + "content": "4. Aneja, J., Agrawal, H., Batra, D., Schwing, A.: Sequential latent spaces for modeling the intention during diverse image captioning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4261-4270 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 299, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 481, + 320 + ], + "type": "text", + "content": "5. Astudillo, R.F., Ballesteros, M., Naseem, T., Blodgett, A., Florian, R.: Transition-based parsing with stack-transformers. arXiv preprint arXiv:2010.10669 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 321, + 481, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 481, + 376 + ], + "type": "text", + "content": "6. Banarescu, L., Bonial, C., Cai, S., Georgescu, M., Griffitht, K., Hermjakob, U., Knight, K., Koehn, P., Palmer, M., Schneider, N.: Abstract meaning representation (amr) 1.0 specification. In: Parsing on Freebase from Question-Answer Pairs. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing. Seattle: ACL. pp. 1533–1544 (2012)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 376, + 481, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 481, + 421 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 481, + 421 + ], + "type": "text", + "content": "7. Banarescu, L., Bonial, C., Cai, S., Georgescu, M., Griffitht, K., Hermjakob, U., Knight, K., Koehn, P., Palmer, M., Schneider, N.: Abstract meaning representation for sembanking. In: Proceedings of the 7th linguistic annotation workshop and interoperability with discourse. pp. 178-186 (2013)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 421, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 421, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 421, + 481, + 453 + ], + "type": "text", + "content": "8. Bevilacqua, M., Błoshmi, R., Navigli, R.: One spring to rule them both: Symmetric amr semantic parsing and generation without a complex pipeline. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 12564-12573 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 454, + 481, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 454, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 454, + 481, + 498 + ], + "type": "text", + "content": "9. Bloshmi, R., Bevilacqua, M., Fabiano, E., Caruso, V., Navigli, R.: Spring goes online: end-to-end AMR parsing and generation. In: Proceedings of the 2021 conference on empirical methods in natural language processing: system demonstrations. pp. 134-142 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 498, + 481, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 481, + 532 + ], + "type": "text", + "content": "10. Cai, S., Knight, K.: Smatch: an evaluation metric for semantic feature structures. In: Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). pp. 748-752 (2013)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 532, + 481, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 532, + 481, + 565 + ], + "spans": [ + { + "bbox": [ + 138, + 532, + 481, + 565 + ], + "type": "text", + "content": "1. Chang, X., Ren, P., Xu, P., Li, Z., Chen, X., Hauptmann, A.: A comprehensive survey of scene graphs: Generation and application. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(1), 1-26 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 565, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 565, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 565, + 481, + 599 + ], + "type": "text", + "content": "2. Chen, L., Jiang, Z., Xiao, J., Liu, W.: Human-like controllable image captioning with verb-specific semantic roles. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16846-16856 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 481, + 632 + ], + "type": "text", + "content": "3. Chen, S., Jin, Q., Wang, P., Wu, Q.: Say as you wish: Fine-grained control of image caption generation with abstract scene graphs. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9962-9971 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "type": "text", + "content": "4. Cho, J., Lei, J., Tan, H., Bansal, M.: Unifying vision-and-language tasks via text generation. In: International Conference on Machine Learning. pp. 1931-1942. PMLR (2021)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "type": "text", + "content": "CIC-BART-SSA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "15. Choi, W.S., Heo, Y.J., Punithan, D., Zhang, B.T.: Scene graph parsing via abstract meaning representation in pre-trained language models. In: NAACL 2022 Workshop on Deep Learning on Graphs for Natural Language Processing (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 171 + ], + "type": "text", + "content": "16. Choi, W.S., Heo, Y.J., Zhang, B.T.: Sgram: Improving scene graph parsing via abstract meaning representation. arXiv preprint arXiv:2210.08675 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 481, + 203 + ], + "type": "text", + "content": "17. Chunseong Park, C., Kim, B., Kim, G.: Attend to you: Personalized image captioning with context sequence memory networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 895-903 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 203, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 203, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 203, + 481, + 236 + ], + "type": "text", + "content": "18. Cornia, M., Baraldi, L., Cucchiara, R.: Show, control and tell: A framework for generating controllable and grounded captions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8307-8316 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 236, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 481, + 267 + ], + "type": "text", + "content": "19. Deng, C., Ding, N., Tan, M., Wu, Q.: Length-controllable image captioning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16. pp. 712-729. Springer (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 267, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 267, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 133, + 267, + 481, + 300 + ], + "type": "text", + "content": "20. Gan, C., Gan, Z., He, X., Gao, J., Deng, L.: Stylenet: Generating attractive visual captions with styles. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3137-3146 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "text", + "content": "21. Guo, L., Liu, J., Yao, P., Li, J., Lu, H.: Mscap: Multi-style image captioning with unpaired stylized text. 2019 IEEE. In: CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4199-4208 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 332, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 481, + 354 + ], + "type": "text", + "content": "22. Hirsch, E., Tal, A.: Clid: Controlled-length image descriptions with limited data. arXiv preprint arXiv:2211.14835 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "text", + "content": "23. Johnson, J., Karpathy, A., Fei-Fei, L.: Densecap: Fully convolutional localization networks for dense captioning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4565-4574 (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "text", + "content": "24. Johnson, J., Krishna, R., Stark, M., Li, L.J., Shamma, D., Bernstein, M., Fei-Fei, L.: Image retrieval using scene graphs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3668-3678 (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "type": "text", + "content": "25. Karpathy, A., Fei-Fei, L.: Deep visual-semantic alignments for generating image descriptions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3128-3137 (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 450, + 481, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 481, + 494 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 481, + 494 + ], + "type": "text", + "content": "26. Li, X., Yin, X., Li, C., Zhang, P., Hu, X., Zhang, L., Wang, L., Hu, H., Dong, L., Wei, F., et al.: Oscar: Object-semantics aligned pre-training for vision-language tasks. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16. pp. 121-137. Springer (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 494, + 481, + 537 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 481, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 481, + 537 + ], + "type": "text", + "content": "27. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 537, + 481, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 537, + 481, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 537, + 481, + 558 + ], + "type": "text", + "content": "28. Lindh, A., Ross, R.J., Kelleher, J.D.: Language-driven region pointer advancement for controllable image captioning. arXiv preprint arXiv:2011.14901 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 558, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 481, + 579 + ], + "type": "text", + "content": "29. Lu, J., Yang, J., Batra, D., Parikh, D.: Neural baby talk. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7219-7228 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 579, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 481, + 600 + ], + "type": "text", + "content": "30. Lukasova, A.: Hierarchical agglomerative clustering procedure. Pattern Recognition 11(5-6), 365-381 (1979)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 600, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 633 + ], + "type": "text", + "content": "31. Luo, J., Li, Y., Pan, Y., Yao, T., Feng, J., Chao, H., Mei, T.: Semantic-conditional diffusion networks for image captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23359-23368 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "text", + "content": "32. Mathews, A., Xie, L., He, X.: Senticap: Generating image descriptions with sentiments. In: Proceedings of the AAAI conference on artificial intelligence. vol. 30 (2016)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 232, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 232, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 232, + 100 + ], + "type": "text", + "content": "K. Basioti et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "33. Mathews, A., Xie, L., He, X.: Semstyle: Learning to generate stylised image captions using unaligned text. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8591-8600 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 171 + ], + "type": "text", + "content": "34. Müller, D.: Modern hierarchical, agglomerative clustering algorithms. arXiv preprint arXiv:1109.2378 (2011)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 171, + 482, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 171, + 482, + 203 + ], + "spans": [ + { + "bbox": [ + 130, + 171, + 482, + 203 + ], + "type": "text", + "content": "35. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP). pp. 1532-1543 (2014)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 203, + 482, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 203, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 203, + 482, + 246 + ], + "type": "text", + "content": "36. Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazebnik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In: Proceedings of the IEEE international conference on computer vision. pp. 2641-2649 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 246, + 482, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 289 + ], + "type": "text", + "content": "37. Ramos, R., Martins, B., Elliott, D., Kementchedjheva, Y.: Smallcap: lightweight image captioning prompted with retrieval augmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2840-2849 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 289, + 482, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 320 + ], + "type": "text", + "content": "38. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems 28 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 320, + 482, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 365 + ], + "type": "text", + "content": "39. Ren, Y., Mao, Z., Fang, S., Lu, Y., He, T., Du, H., Zhang, Y., Ouyang, W.: Cross-ing the gap: Domain generalization for image captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2871-2880 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 365, + 482, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 397 + ], + "type": "text", + "content": "40. Rennie, S.J., Marcheret, E., Mroueh, Y., Ross, J., Goel, V.: Self-critical sequence training for image captioning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7008-7024 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 397, + 482, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 482, + 429 + ], + "type": "text", + "content": "41. Shuster, K., Humeau, S., Hu, H., Bordes, A., Weston, J.: Engaging image captioning via personality. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12516-12526 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 429, + 482, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 472 + ], + "type": "text", + "content": "42. Toutanova, K., Klein, D., Manning, C.D., Singer, Y.: Feature-rich part-of-speech tagging with a cyclic dependency network. In: Proceedings of the 2003 human language technology conference of the north american chapter of the association for computational linguistics. pp. 252-259 (2003)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 472, + 482, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 515 + ], + "type": "text", + "content": "43. Toutanvoa, K., Manning, C.D.: Enriching the knowledge sources used in a maximum entropy part-of-speech tagger. In: 2000 Joint SIGDAT conference on Empirical methods in natural language processing and very large corpora. pp. 63-70 (2000)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 515, + 482, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 482, + 548 + ], + "type": "text", + "content": "44. Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4566-4575 (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 548, + 482, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 568 + ], + "type": "text", + "content": "45. Wang, N., Xie, J., Wu, J., Jia, M., Li, L.: Controllable image captioning via prompting. In: AAAI (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "type": "text", + "content": "46. Wang, Q., Chan, A.B.: Describing like humans: on diversity in image captioning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4195-4203 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 482, + 633 + ], + "type": "text", + "content": "47. Wang, T., Zhang, J., Fei, J., Ge, Y., Zheng, H., Tang, Y., Li, Z., Gao, M., Zhao, S., Shan, Y., et al.: Caption anything: Interactive image description with diverse multimodal controls. arXiv preprint arXiv:2305.02677 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 633, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 633, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 633, + 482, + 665 + ], + "type": "text", + "content": "48. Wang, Z., Xiao, J., Chen, L., Gao, F., Shao, J., Chen, L.: Learning combinatorial prompts for universal controllable image captioning. arXiv preprint arXiv:2303.06338 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 380, + 91, + 447, + 100 + ], + "type": "text", + "content": "CIC-BART-SSA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 491 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 172 + ], + "type": "text", + "content": "49. Xia, Q., Huang, H., Duan, N., Zhang, D., Ji, L., Sui, Z., Cui, E., Bharti, T., Zhou, M.: Xgpt: Cross-modal generative pre-training for image captioning. In: Natural Language Processing and Chinese Computing: 10th CCF International Conference, NLPCC 2021, Qingdao, China, October 13-17, 2021, Proceedings, Part I 10. pp. 786-797. Springer (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "type": "text", + "content": "50. Xu, K., Ba, J., Kiros, R., Cho, K., Courville, A., Salakhudinov, R., Zemel, R., Bengio, Y.: Show, attend and tell: Neural image caption generation with visual attention. In: International conference on machine learning. pp. 2048-2057. PMLR (2015)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "51. Young, P., Lai, A., Hodosh, M., Hockenmaier, J.: From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics 2, 67-78 (2014)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "text", + "content": "52. Zellers, R., Yatskar, M., Thomson, S., Choi, Y.: Neural motifs: Scene graph parsing with global context. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5831-5840 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 282, + 481, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 324 + ], + "type": "text", + "content": "53. Zeng, Z., Zhang, H., Lu, R., Wang, D., Chen, B., Wang, Z.: Conzic: Controllable zero-shot image captioning by sampling-based polishing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23465-23476 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "type": "text", + "content": "54. Zhao, W., Wu, X., Zhang, X.: Memcap: Memorizing style knowledge for image captioning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 12984-12992 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "text", + "content": "55. Zhao, Y., Wei, J., Lin, Z., Sun, Y., Zhang, M., Zhang, M.: Visual spatial description: Controlled spatial-oriented image-to-text generation. arXiv preprint arXiv:2210.11109 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 392, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 424 + ], + "type": "text", + "content": "56. Zheng, Y., Li, Y., Wang, S.: Intention oriented image captions with guiding objects. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8395-8404 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 424, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 468 + ], + "type": "text", + "content": "57. Zhong, Y., Wang, L., Chen, J., Yu, D., Li, Y.: Comprehensive image captioning via scene graph decomposition. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 211-229. Springer (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 468, + 481, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 491 + ], + "type": "text", + "content": "58. Zhu, W., Bhat, S.: Gruen for evaluating linguistic quality of generated text. arXiv preprint arXiv:2010.02498 (2020)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "text", + "content": "K. Basioti et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CLAMP-ViT_ Contrastive Data-Free Learning for Adaptive Post-Training Quantization of ViTs/9ac2383c-dda4-4497-a269-a04b9fe79777_content_list.json b/2024/CLAMP-ViT_ Contrastive Data-Free Learning for Adaptive Post-Training Quantization of ViTs/9ac2383c-dda4-4497-a269-a04b9fe79777_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8d29eccc777b803230c3ad3005cc32c870e94b8a --- /dev/null +++ b/2024/CLAMP-ViT_ Contrastive Data-Free Learning for Adaptive Post-Training Quantization of ViTs/9ac2383c-dda4-4497-a269-a04b9fe79777_content_list.json @@ -0,0 +1,1981 @@ +[ + { + "type": "text", + "text": "CLAMP-ViT: Contrastive Data-Free Learning for Adaptive Post-Training Quantization of ViTs", + "text_level": 1, + "bbox": [ + 218, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Akshat Ramachandran1, Souvik Kundu2, and Tushar Krishna1", + "bbox": [ + 248, + 212, + 754, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Georgia Institute of Technology, Atlanta, GA, USA", + "bbox": [ + 321, + 239, + 679, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "akshat.r@gatech.edu, tushar@ece.gatech.edu", + "bbox": [ + 336, + 255, + 663, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Intel Labs, San Diego, CA, USA", + "bbox": [ + 385, + 268, + 614, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "souvikk.kundu@intel.com", + "bbox": [ + 411, + 282, + 591, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We present CLAMP-ViT, a data-free post-training quantization method for vision transformers (ViTs). We identify the limitations of recent techniques, notably their inability to leverage meaningful inter-patch relationships, leading to the generation of simplistic and semantically vague data, impacting quantization accuracy. CLAMP-ViT employs a two-stage approach, cyclically adapting between data generation and model quantization. Specifically, we incorporate a patch-level contrastive learning scheme to generate richer, semantically meaningful data. Furthermore, we leverage contrastive learning in layer-wise evolutionary search for fixed- and mixed-precision quantization to identify optimal quantization parameters while mitigating the effects of a non-smooth loss landscape. Extensive evaluations across various vision tasks demonstrate the superiority of CLAMP-ViT, with performance improvements of up to $3\\%$ in top-1 accuracy for classification, 0.6 mAP for object detection, and 1.5 mIoU for segmentation at similar or better compression ratio over existing alternatives. Code is available at https://github.com/georgia-tech-synergy-lab/CLAMP-ViT.git", + "bbox": [ + 261, + 335, + 743, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Data-free quantization $\\cdot$ PTQ $\\cdot$ Contrastive learning $\\cdot$ Vision transformer", + "bbox": [ + 259, + 585, + 740, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 640, + 375, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision transformers [12] (ViTs) have recently gained a lot of traction due to their state-of-the-art (SoTA) performance across various computer vision (CV) tasks [34, 37, 48, 50]. Concurrently, the growing need to deploy these parameter-heavy models at the resource-limited edge [14], has inspired research on various model compression techniques. Model quantization [14, 30, 36, 45] has emerged as a popular technique to achieve memory and compute efficient deployment. Quantization reduces memory footprint and improves computation speed of a model by mapping full-precision (FP) weights to reduced precision formats (e.g., $\\leq$ 8-bit INT) [13, 17, 19, 46]. In particular, quantization-aware training (QAT) allows a model to train by taking quantization approximation in to account, enabling ease of quantization. Post-training quantization (PTQ), in contrast, acts", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "as a plug-and-play quantization applied on a pre-trained model. PTQ has become popular as it can leverage the pre-trained model and does not add training overhead [21, 23, 26]. We thus consider the PTQ setting in this work. However, PTQ requires access to a calibration set that is often drawn from the training data [18, 45]. This may be infeasible in situations involving privacy and security concerns [20, 49], making these techniques ill-suited to yield optimal performance.", + "bbox": [ + 212, + 146, + 787, + 237 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent works [25, 27] propose data-free PTQ (DFQ), generating synthetic calibration data $T_{\\mathrm{syn}}$ from Gaussian noise [7, 27], embedding information from the original dataset $T_{\\mathrm{orig}}$ , where $T_{\\mathrm{syn}} << T_{\\mathrm{orig}}$ .", + "bbox": [ + 212, + 241, + 486, + 334 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing DFQ methods for CNNs [2,51] exploit the batch-normalization (BN) layer statistics [33,44] to generate synthetic samples mimicking the original data distribution. For ViTs, absence of the BN layer makes these", + "bbox": [ + 212, + 337, + 486, + 428 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ce2b04350dd7881254d169e2d1cfb47df28261f16d54cc0e8a0cf747b0d1f79d.jpg", + "image_caption": [ + "Fig.1: Visualization of loss landscape of (a) PSAQ-ViT v2 and (b) CLAMP-ViT on DeiT-S with perturbations to quantized model weights and synthetic data [22]." + ], + "image_footnote": [], + "bbox": [ + 509, + 241, + 640, + 344 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4e6a46ce38d22d877f12978647ce2c999787c547daa070c464e7583ab7402f08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 241, + 781, + 343 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "techniques obsolete. Recent efforts to extend DFQ to ViTs include PSAQ-ViT v1 [27] and PSAQ-ViT v2 [25]. They rely on information embedded in the attention score output of the multi-head self attention (MHSA) layer. PSAQ-ViT v1 and v2 [25,27] introduce a relative value metric namely, patch similarity, to optimize Gaussian noise towards synthetic data by maximizing the entropy of patch similarity in a global manner. However, the metric considered in PSAQ-ViT v1 and v2 assumes all patches1 to be equally important, without considering spatial sensitivity [47]. This may fail to capture semantically meaningful inter-patch relations, potentially affecting robustness of the synthetic data. As we can see in Fig. 1(a) even insignificant perturbations in the generated images (augmenting pixels) or weights (to simulate quantization process) may cause significant jaggedness in the loss landscape of PSAQ-ViT v2. This also implies that the predictions may have large discrepancy even for small input/weight perturbation. Moreover, the synthetic data generation stage in these methods does not consider the informativeness of the generated samples towards the quantization process, nor do they establish countermeasures to ameliorate the non-smooth loss landscape during quantization, resulting in sub-optimal parameter search and poor generalization to test set [5].", + "bbox": [ + 212, + 428, + 787, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions. The discussion above hints at the potential limitation of [25,27] in capturing semantically meaningful and robust inter-patch relationships to generate synthetic data that is well-suited to quantization. Towards solving these limitations, we present contrastive data-free learning for adaptive post-training quantization of ViTs (CLAMP-ViT), a general DFQ method applicable to a wide range of vision tasks. To the best of our knowledge, CLAMP-", + "bbox": [ + 212, + 705, + 787, + 797 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna", + "bbox": [ + 271, + 114, + 671, + 128 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1 Patch (subset of image): group of neighboring pixels in an image.", + "bbox": [ + 217, + 824, + 666, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ViT is the first work to support both fixed-² and mixed-precision³ DFQ of ViTs, starting from a pre-trained FP model.", + "bbox": [ + 212, + 145, + 785, + 175 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, CLAMP-ViT utilizes the architectural characteristics of ViTs and inherent properties of real-images to generate semantically rich and meaningful data while ensuring spatial sensitivity. Here we leverage a novel patch-level contrastive learning scheme, where for each patch (anchor) in the MHSA layer output, we treat semantically similar patches in a neighborhood around the patch as positive patches (evaluated using cosine similarity) and the remaining patches in the neighborhood as negative patches (see Fig. 3 for an intuitive visualization). We then leverage a patch-level contrastive loss that drives the representation of the anchor patch closer to the positive patches and away from the negative patches, enabling exploration of semantically meaningful relations.", + "bbox": [ + 212, + 176, + 785, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, Evol-Q [14] identified a limitation of gradient based methods [25] to search over quantization parameters of ViTs having non-smooth loss landscape, leading to poor accuracy and generalization. We take inspiration from this and a recent work [35] to design a layer-wise evolutionary search to identify the optimal bit-width and scale factors for each layer. Additionally, we propose a novel local contrastive objective to capture distributional variance in intermediate layer outputs (crucial for layer-wise quantization) and improve search convergence. This loss sufficiently captures the representational divergence of intermediate layers outputs to identify optimal quantization parameters and yields a smooth loss landscape, as demonstrated in Fig. 1(b), enabling generalizability and better performance on test data [5]. Furthermore, to ensure the generated data is adaptive to the quantization process, CLAMP-ViT performs a cyclically adaptive strategy alternating between data-generation and quantization.", + "bbox": [ + 212, + 327, + 785, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To evaluate the merits of CLAMP-ViT we conduct extensive experiments on image classification, detection, and segmentation with different ViT variants and quantization scenarios and observe superior performance over SoTA.", + "bbox": [ + 212, + 523, + 785, + 569 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 214, + 590, + 397, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data-Driven PTQ for ViTs. PTQ offers an efficient alternative to QAT [23,26] by directly quantizing pre-trained models without the need for compute-heavy retraining. In specific, PTQ4ViT [45] employs twin uniform quantization and Hessian-guided scale optimization. FQ-ViT [30] uses a power-of-two factor quantization to handle inter-channel variation in LayerNorm and log-INT-Softmax. RepQ-ViT [28] separates quantization and inference, optimizing accuracy and efficiency by employing scale-reparameterized quantizers. These methods apply fixed-precision quantization, assuming all layers support similar approximations, potentially leading to sub-optimal accuracy [36]. On the other hand, techniques like, VT-PTQ [32] adopt mixed precision for specific modules based on sensitivity metrics, while PMQ [41] and LRP-QViT [36] allocate bit-widths by assessing both layer sensitivity and contribution to the output,", + "bbox": [ + 212, + 621, + 787, + 803 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "CLAMP-ViT", + "bbox": [ + 640, + 113, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2 weights/activations quantized to same precision for all layers.", + "bbox": [ + 217, + 810, + 643, + 825 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "3 weights/activations quantized to different precision for different layers.", + "bbox": [ + 218, + 825, + 700, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "respectively. However, all these methods assume a part of training set to be available for calibration which may be infeasible due to privacy [20].", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data-Free PTQ for ViTs. Recent data-free PTQ efforts (DFQ) for ViTs including PSAQ-ViT v1 [27] and v2 [25] utilize a \"patch similarity\" metric to refine Gaussian noise to synthetic data resembling real images. In specific, both these approaches leverage the fact that the self-attention has different response to real image and noise. PSAQ-ViT v1 [27] uses a two-stage cascaded framework to generate synthetic data for model quantization on image classification tasks. PSAQ-ViT v2 [25] expanded this to a broader range of tasks, employing a Min-Max game between full precision and quantized models for data generation and quantization. Despite these innovations, both the versions face several limitations as highlighted in Sec. 1, leading to poor DFQ performance, particularly at reduced precision. Additionally, these methods support only fixed-precision quantization. CLAMP-ViT, in contrast, introduces a DFQ method that addresses their shortcomings in yielding SoTA performance, while supporting both fixed and mixed-precision quantization.", + "bbox": [ + 212, + 176, + 785, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Contrastive Learning. Contrastive learning as a technique has been widely adopted in self-supervised settings [4,15] and is proven to combat overfitting via regularization against negative samples. Recently, Evol-Q [14] demonstrated the benefits of a global contrastive objective [8,43] coupled with evolutionary search in smoothening the loss landscape while calibrating the scaling factors of a prequantized model to enhance accuracy. Unlike [14] that only aims to adjust the scale factors of each layer of a pre-quantized model, we conduct complete quantization starting from a FP model. On evaluation of the global contrastive objective used in Evol-Q, we find it to be sub-optimal, causing premature convergence while quantizing an FP model. Instead, we present a local contrastive loss that captures the distributional variance in intermediate layer outputs, drastically improving convergence and hence, identification of quantization parameters.", + "bbox": [ + 212, + 388, + 787, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 CLAMP-ViT Framework", + "text_level": 1, + "bbox": [ + 215, + 593, + 495, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present an overview of CLAMP-ViT in Fig. 2. In this section, we first go through notations, computational process of ViTs and quantization strategy in the preliminaries, followed by a detailed description of the proposed contrastive loss and the two stages of CLAMP-ViT. Finally, the overall DFQ pipeline is summarized and presented (refer to Supplementary for the detailed Algorithm).", + "bbox": [ + 212, + 625, + 787, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 723, + 372, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notations. Let $X \\in \\mathcal{R}^{H \\times W \\times C}$ be the input image to an $L$ -layer ViT, where $(H, W, C)$ are the height, width, and channels, respectively (we ignore the batch dimension for simplicity). The input is partitioned into $N$ non-overlapping patches that are then linearly transformed to $d$ -dimensional patch embeddings, $(\\mathcal{R}^{N \\times d})$ that is passed through an encoder consisting of a series of transformer layers each composed of an MHSA and an MLP module. Each MHSA module is composed", + "bbox": [ + 212, + 748, + 792, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna", + "bbox": [ + 271, + 114, + 671, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f7653a12afd9e5305c763845f5a65525f8c086f2639047166dcdf38e838984f4.jpg", + "image_caption": [ + "Fig.2: Overview of the cyclically evolving two-stage CLAMP-ViT framework. In stage 1 (① - ②), $\\mathcal{L}^{SG}$ is minimized to update Gaussian noise towards synthesizing data. Stage 2 (③ - ⑤) conducts layer-wise evolutionary search to identify optimal quantization parameters while minimizes $\\mathcal{L}^F$ . Illustrated with multiple instances of models for clarity, only one instance of each model is used in the framework." + ], + "image_footnote": [], + "bbox": [ + 230, + 143, + 767, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of $h$ heads, to capture long-range patch correlations [24]. At head $j$ , $X$ is transformed to query $(Q_j)$ , key $(K_j)$ , and value $(V_j)$ tensors to perform self-attention $\\varPhi_j$ , that is linearly projected after concatenation over the heads,", + "bbox": [ + 212, + 440, + 787, + 486 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi_ {j} \\left(Q _ {j}, K _ {j}, V _ {j}\\right) = \\operatorname {s o f t m a x} \\left(Q _ {j} K _ {j} ^ {T} / \\sqrt {d}\\right) V _ {j}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 491, + 785, + 510 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {M H S A} (Q, K, V) = \\mathbf {c o n c a t} \\left(\\Phi_ {0}, \\Phi_ {1}, \\dots \\Phi_ {h - 1}\\right) W ^ {0} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 333, + 520, + 784, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The output of the $j^{th}$ head at $i^{th}$ MHSA layer is given by $O_{i,j} \\in \\mathcal{R}^{N \\times d}$ . The series of transformer layers are succeeded by task-specific heads for classification, detection, or segmentation, depending on the nature of the vision task.", + "bbox": [ + 212, + 542, + 787, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantization. In this paper, we perform uniform symmetric quantization (fixed/mixed-precision) of both weights and activations for ViTs, mapping full precision values into a uniform scale determined by bit-width (b), given as,", + "bbox": [ + 212, + 589, + 787, + 635 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ (X, \\gamma , b) = \\operatorname {c l i p} \\left(\\left\\lfloor \\frac {X}{\\gamma} \\right\\rfloor , - 2 ^ {b - 1} + 1, 2 ^ {b - 1} - 1\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 645, + 785, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma$ is the scale factor and $X$ represents the FP tensor.", + "bbox": [ + 212, + 686, + 643, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Contrastive Objective", + "text_level": 1, + "bbox": [ + 214, + 724, + 444, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Contrastive learning based on the infoNCE loss [39] helps learn an anchor sample from both the similar (positive) and dissimilar (negative) samples, typically using a softmax function to normalize the similarities into probabilities. However, infoNCE loss suffers from a disproportionate impact of a single positive and many negative samples [14]. This can affect learning the synthetic data as well as the quantized model parameters. Inspired by [39], we present a modified", + "bbox": [ + 212, + 750, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "CLAMP-ViT", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "infoNCE loss $(\\mathcal{L}_{i,j}^{C}$ in Eq. 4) that addresses this imbalance yielding latent impact of positive and negative samples pairs equally.", + "bbox": [ + 212, + 145, + 787, + 176 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i, j} ^ {C} = - \\log \\frac {\\sum_ {p +} \\exp \\left(\\lambda_ {i , j} ^ {p} \\cdot \\lambda_ {i , j} ^ {p +} / \\tau\\right)}{\\sum_ {p +} \\exp \\left(\\lambda_ {i , j} ^ {p} \\cdot \\lambda_ {i , j} ^ {p +} / \\tau\\right) + \\sum_ {p -} \\exp \\left(\\lambda_ {i , j} ^ {p} \\cdot \\lambda_ {i , j} ^ {p -} / \\tau\\right)} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 285, + 186, + 785, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "here superscript $p$ , $p+$ , and $p-$ correspond to the anchor prediction (drawn from quantized model), positive, and negative prediction (drawn from FP model), respectively. $\\tau$ controls the concentration level [39]. Subscript $(i,j)$ represents (# layer id, # head id), and (# layer id, # batch id) for the data generation and quantization stage, respectively. $\\lambda$ represents the variable taken to evaluate the log likelihood—the intermediate layer patch embeddings (stage 1) or layer output activations (stage 2). The final stage $s$ loss is given as $\\mathcal{L}^{C_s} = \\sum_i\\sum_j\\mathcal{L}_{i,j}^{C_s}$ .", + "bbox": [ + 214, + 239, + 787, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Stage 1: Synthetic Data Generation", + "text_level": 1, + "bbox": [ + 215, + 367, + 555, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our goal in Stage 1 is to generate semantically rich and meaningful images that can reliably exploit inter-patch relations while ensuring informativeness to the quantization process. For each \"anchor patch\" of an MHSA output in the quantized model, we first locate positive and", + "bbox": [ + 214, + 393, + 470, + 513 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bb55ab2dde3d6e5e061692689f734e3846f5a261fb1c2b3d95ee9b778952e1fd.jpg", + "image_caption": [ + "Fig. 3: Intuitive visualization of positive and negative patch selection in Stage 1." + ], + "image_footnote": [], + "bbox": [ + 480, + 397, + 767, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "negative patches (see Fig. 3) from the same layer id in the FP model, from a neighborhood of the anchor's spatial position. We then leverage a contrastive learning objective (see Sec. 3.2) that operates to maximize the similarity between the anchor and positive patches while minimizing similarity with the negative patches. Selection of the anchor patch from the quantized model for guiding data generations helps the model to recognize the semantic context within the generated data, during the subsequent quantization stage (informativeness).", + "bbox": [ + 212, + 513, + 787, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Patch Neighborhood. For every $k^{th}$ anchor patch $\\mathcal{P}_{ijk}$ corresponding to the $j^{\\text{th}}$ head in $i^{\\text{th}}$ MHSA layer of the quantized model, we identify a neighborhood of size $\\mathcal{N}_{ijk}$ with the same spatial location of $\\mathcal{P}_{ijk}$ as the center, but located in the MHSA layer output of the FP model. Within $\\mathcal{N}_{ijk}$ , to identify the most semantically correlated patches $(\\mathcal{P}_{\\mathcal{N}} \\in \\mathcal{N}_{ijk})$ we use cosine similarity as follows,", + "bbox": [ + 212, + 619, + 787, + 696 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\rho (i, j, k) = \\frac {\\mathcal {P} _ {i j k} ^ {T} \\cdot \\mathcal {P} _ {\\mathcal {N}}}{\\left| \\left| \\mathcal {P} _ {i j k} \\right| \\right| _ {2} \\cdot \\left| \\left| \\mathcal {P} _ {\\mathcal {N}} \\right| \\right| _ {2}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 705, + 785, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The cosine similarity $\\rho$ is estimated $\\forall P_{\\mathcal{N}} \\in \\mathcal{N}_{ijk}$ . We then select the positive patches to be the top- $n$ patches that have the highest $\\rho(i,j,k)$ with the anchor patch and the rest of the patches in the neighborhood correspond to negative patches. We empirically set $n = 4$ in our experiments for a neighborhood of size $3 \\times 3$ . We then compute $\\mathcal{L}_{i,j}^{C_1}$ for all anchor patches for each attention head output over all layers to get the contrastive loss $(\\mathcal{L}^{C_1})$ . We then compute the sample", + "bbox": [ + 214, + 747, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna", + "bbox": [ + 271, + 114, + 673, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "generation loss $\\mathcal{L}^{SG}$ by combining $\\mathcal{L}^{C_1}$ and the mean absolute error (MAE) (see Sec. 3.5) output loss $(\\mathcal{L}^O)$ and minimize it for data generation.", + "bbox": [ + 212, + 145, + 787, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Stage 2: Quantization", + "text_level": 1, + "bbox": [ + 214, + 199, + 441, + 214 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We now present our layer-wise evolutionary search with a local contrastive loss-based fitness function to rank suitable quantization parameters from a large search space. We detail the fitness function and the search steps as follows.", + "bbox": [ + 212, + 224, + 782, + 268 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fitness Function. To evaluate the quantization parameters explored by the evolutionary search algorithm, we introduce a fitness function $\\mathcal{L}^F$ that combines the contrastive loss $\\mathcal{L}^{C_2}$ (Eq. (4)) and the MAE loss $(\\mathcal{L}^O)$ computed with respect to the targets at the output (explained in Sec. 3.5) i.e., $\\mathcal{L}^F = \\mathcal{L}^{C_2} + \\mathcal{L}^O$ . Unlike Stage 1, we use the intermediate activations after each transformer layer to compute the contrastive loss $\\mathcal{L}^{C_2}$ . For a ViT, let the set of intermediate representations of FP and quantized model be denoted as, $\\mathcal{O}^{fp} = \\{\\mathcal{O}_0^{fp}, \\mathcal{O}_1^{fp}, \\dots, \\mathcal{O}_{L-1}^{fp}\\}$ and $\\mathcal{O}^q = \\{\\mathcal{O}_0^q, \\mathcal{O}_1^q, \\dots, \\mathcal{O}_{L-1}^q\\}$ , respectively. However, directly using high dimensional $\\mathcal{O}^{fp}$ and $\\mathcal{O}^q$ can result in high compute overhead. Therefore for each layer $i$ , at the intermediate output, we perform mean pooling along the patch dimension to obtain a low-dimensional representation ( $\\mathcal{O}_i \\in \\mathcal{R}^N$ ), reducing it by a factor of $h \\times d$ . We then apply the contrastive loss (Eq. (4)) sampled within the batch dimension on the low-dimensional $\\mathcal{O}_i^{fp}$ and $\\mathcal{O}_i^q$ . For the layer output activation generated from each constituent of a batch of synthetic data, the anchor ( $\\lambda^p$ ) corresponds to the intermediate layer output of the quantized model, positive ( $\\lambda^{p+}$ ) and negative ( $\\lambda^{p-}$ ) samples correspond to set of intermediate layer outputs of FP model that have the same and different targets respectively, relative to the anchor within the batch.", + "bbox": [ + 212, + 268, + 787, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Step 1: Candidate Initialization. A candidate quantization solution is encoded as a set $\\Lambda$ of $L$ tuples, such that for layer $i$ , tuple $\\Lambda[i]$ represents the two quantization parameters $\\langle b, \\gamma \\rangle$ . $b$ can take any integer value between 2 and 8, while $\\gamma$ is constrained to a uniform ball of radius $10^{-3}$ centered around, $\\gamma[i]^{init4}$ . The candidate scale factors are sampled as $\\gamma[i] = \\gamma[i]^{init} + f(-10^{-3}, +10^{+3})$ , where $f$ is a random sampling function. We begin the evolutionary search by creating a population via randomly sampling $\\mathcal{K}$ candidate $\\Lambda$ s, each consisting of layer-wise quantization parameters. We then evaluate the fitness function $\\mathcal{L}^F$ for each candidate $\\Lambda$ . We create $\\mathcal{K}$ tuples each with $(\\Lambda, \\mathcal{L}^F)$ to form the initial population. $\\mathcal{L}^F$ of each initial candidate with corresponding set $\\Lambda$ is pre-computed and stored to avoid recomputation.", + "bbox": [ + 212, + 544, + 787, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Step 2: Re-generation (Crossover and Mutation). Each candidate in the population is ranked based on corresponding $\\mathcal{L}^F$ s (lower the better) of which the top two serve as the parents for the next candidate generation (child). When evolving candidates, perturbing too many layer parameters based on parents can lead to search instability. To mitigate this, at each evolution step we employ a layer-wise regeneration approach, evolving a single transformer layer at a time based on chosen parents, keeping all other layer parameters to that of the top-1", + "bbox": [ + 212, + 710, + 787, + 816 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "CLAMP-ViT", + "bbox": [ + 640, + 113, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$^4$ $(\\max (\\theta_i) - \\min (\\theta_i)) / (2^b -1)$ , where $\\theta$ is the weight tensor.", + "bbox": [ + 217, + 824, + 632, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "parent (ranked via $\\mathcal{L}^F$ ). The child's parameter regeneration for a layer based on that of the chosen parents ( $p1, p2$ ) is formulated as,", + "bbox": [ + 212, + 145, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nb _ {\\text {c h i l d}} = \\mathbf {r} \\text {a n d o m} (\\mathbf {m i n} (b _ {p 1}, b _ {p 2}) - 1, \\mathbf {m a x} (b _ {p 1}, b _ {p 2}) + 1) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 185, + 785, + 200 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma_ {c h i l d} = \\text {m e a n} \\left(\\gamma_ {p 1}, \\gamma_ {p 2}\\right) + \\eta \\left(- 1 0 ^ {- 3}, 1 0 ^ {3}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 204, + 785, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Step 3: Diversity Promoting Selection. To avoid overfitting during search we follow [11] and introduce diversity into the population. In specific, we create $P = 5$ random parents and use each of them to crossover with the child generated in Step 2 and create $P'$ diverse children by following Eq. (6), Eq. (7).", + "bbox": [ + 212, + 227, + 782, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Step 4: Evaluation and Population Update. We evaluate all generated children in the steps above and acquire $\\mathcal{L}^F$ s. The child generated in Step 2 and the corresponding fitness function value is added to the population. We then rank the diversity promoting children from Step 3 and select the best child to be added to the population for the next iteration. In our layer-wise evolutionary search strategy, we employ $\\mathcal{P}$ passes over all layers of a ViT, and each layer is iterated over $\\mathcal{C}$ cycles in each pass. So, the population is updated $\\mathcal{P} \\times \\mathcal{C} \\times L$ times, i.e., the Step 2, 3, and 4 are iteratively executed $\\mathcal{P} \\times \\mathcal{C} \\times L$ times.", + "bbox": [ + 212, + 287, + 784, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Activation Quantization. We note that the sensitivity to quantization for activations is closely correlated to the sensitivity of weight parameters. Therefore for layer $i$ , we determine the output activation quantization parameters as follows, $b_{act}[i] = \\min (8,b[i]\\times 2)$ and $\\gamma_{act}[i] = \\gamma_{act}[i - 1] + \\gamma [i]$ .", + "bbox": [ + 212, + 409, + 782, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Overall Pipeline", + "text_level": 1, + "bbox": [ + 214, + 487, + 395, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 2, we illustrate the complete CLAMP-ViT framework. To ensure adaptive data-generation and informativeness for quantization, we use a cyclic strategy between the two stages, updating generated data based on the quantized model's needs for optimal parameter search. The framework requires an input batch of $\\mathcal{B}$ random Gaussian images $X_{\\mathcal{B}}$ , and corresponding task-specific targets $T_{G_{\\mathcal{B}}}$ ( $T_{G_{\\mathcal{B}}}$ for each task is detailed in Sec. 4). The targets direct the synthetic image generation towards task-specific goals as well as penalize inaccurate predictions of quantized model through the output loss $\\mathcal{L}^O$ ,", + "bbox": [ + 212, + 508, + 784, + 630 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {O} = \\frac {1}{n _ {c}} \\left(\\left| \\left| \\mathcal {Q} \\left(X _ {\\mathcal {B}}\\right) - T _ {G _ {\\mathcal {B}}} \\right| \\right| _ {1} + \\left| \\left| \\mathcal {F P} \\left(X _ {\\mathcal {B}}\\right) - T _ {G _ {\\mathcal {B}}} \\right| \\right| _ {1}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 641, + 785, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $n_c$ is the number of output classes (classification) or prediction map size (segmentation/detection). The quantized model is initialized to the best candidate from $\\mathcal{K}$ tuples. The framework assumes a range of bit-widths and a single bit-width for mixed- and fixed-precision search, respectively. In Stage 1, $X_{\\mathcal{B}}$ is fed to the frozen quantized $(\\mathcal{Q})$ and full-precision model $(\\mathcal{FP})$ to minimize the sample generation loss $\\mathcal{L}^{SG} = \\mathcal{L}^{C_1} + \\mathcal{L}^O$ , updating $X_{\\mathcal{B}}$ via backpropagation for $G$ iterations. In Stage 2, we use the generated data to quantize $\\mathcal{Q}$ for $\\mathcal{P} \\times \\mathcal{C} \\times L$ iterations by minimizing $\\mathcal{L}^F$ . We cyclically update the generated data every $\\mathcal{C}/2$ iterations. In every subsequent execution of Stage 1, we do not restart from Gaussian noise but use $X_{\\mathcal{B}}$ from the previous Stage 1 execution and update it for $G/2$ iterations. In this manner, the two stages are executed alternately.", + "bbox": [ + 212, + 672, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna", + "bbox": [ + 271, + 114, + 673, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/269bb678dd8d0a8846a6a3d970e222449ab8d44a6f01a6e8006992caf430d05c.jpg", + "table_caption": [ + "Table 1: Fixed-precision quantization accuracy comparison with SoTA on image classification tasks with ImageNet-1k testset. 'R', 'S' signifies real and synthetic calibration data and W/A indicates weight/activation bit-width. The values in bold and underline signifies, the best performance overall, and with synthetic data, respectively." + ], + "table_footnote": [], + "table_body": "
ModelMethodData#ImagesW/ATop-1W/ATop-1
ViT-BBaseline--32/3284.5332/3284.53
PSAQ-ViT v1S328/837.364/825.34
PTQ4ViTR328/884.254/867.16
FQ-ViTR10008/883.314/878.73
RepQ-ViTR328/881.454/876.29
CLAMP-ViT (Ours)S328/884.194/878.73
DeiT-TBaseline--32/3272.2132/3272.21
PSAQ-ViT v1S328/871.564/865.57
PSAQ-ViT v2S328/872.174/868.61
FQ-ViTR10008/871.614/866.91
RepQ-ViTR328/872.054/868.75
CLAMP-ViT (Ours)S328/872.174/869.93
DeiT-SBaseline--32/3279.8532/3279.85
PSAQ-ViT v1S328/876.924/873.23
PSAQ-ViT v2S328/879.564/876.36
PTQ4ViTR328/879.474/8-
FQ-ViTR10008/879.174/876.93
RepQ-ViTR328/879.554/876.75
CLAMP-ViT (Ours)S328/879.554/877.03
Swin-TBaseline--32/3281.3532/3281.35
PSAQ-ViT v1 [27]S328/875.354/871.79
PSAQ-ViT v2 [25]S328/880.214/876.28
FQ-ViT [30]R10008/881.294/880.73
RepQ-ViT [28]R328/881.284/880.51
CLAMP-ViT (Ours)S328/881.174/880.28
Swin-SBaseline--32/3283.2032/3283.20
PSAQ-ViT v1 [27]S328/876.644/875.14
PSAQ-ViT v2 [25]S328/882.134/878.86
FQ-ViT [30]R10008/882.134/881.67
RepQ-ViT [28]R328/882.344/882.14
CLAMP-ViT (Ours)S328/882.574/882.51
", + "bbox": [ + 297, + 200, + 696, + 554 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experimental Results", + "text_level": 1, + "bbox": [ + 215, + 568, + 460, + 585 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 601, + 426, + 617 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Models and Datasets. We evaluate CLAMP-ViT on various ViT model families (pre-trained FP models sourced from timm [40]) for image classification, object detection and semantic segmentation detailed as follows.", + "bbox": [ + 212, + 627, + 782, + 672 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Image Classification. We use ImageNet-1K [10] having 50K testset, with DeiT-B/T/S [38], Swin-T/S [31], and ViT-B/S [12] to evaluate accuracy.", + "bbox": [ + 212, + 672, + 784, + 702 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Object detection. We use the COCO 2017 dataset [29] having approximately 20K test data. Following [25,28,30], we use the Cascade Mask R-CNN [3] framework from MMdetection library [6] with DeiT-S and Swin-S as the backbone.", + "bbox": [ + 212, + 703, + 784, + 748 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Semantic Segmentation. We use the ADE20K dataset [52] with 3K test data encompassing 150 categories with DeiT-S and Swin-S as the backbone. We adopt the UperNet framework [42] in the MMsegmentation library [9] similar to [25]. Baselines. CLAMP-ViT is evaluated against SoTA PTQ (real data) and DFQ (synthetic data) methods for quantizing models from FP in various vision tasks. For image classification, it's compared with fixed-precision methods like PSAQ-", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "CLAMP-ViT", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/81f121cbd7fda4f4c5cf8331e07502b14ca18991eb8b6446df6a960fd242ed35.jpg", + "table_caption": [ + "Table 2: Mixed-Precision Quantization accuracy comparison on image classification tasks with ImageNet-1k testset. The values in bold indicate best performance overall." + ], + "table_footnote": [], + "table_body": "
DeiT-TDeiT-SSwin-TSwin-S
MethodDataW/ATop-1W/ATop-1W/ATop-1W/ATop-1
LRP-QViT [36]RMP6/MP671.03MP6/MP679.03--MP6/MP682.86
VT-PTQ [32]RMP6/MP669.46MP6/MP675.10MP6/MP679.61MP6/MP678.43
CLAMP-ViT (Ours)SMP4.9/MP6.271.69MP4.7/MP5.979.43MP5.5/MP6.981.78MP5.1/MP6.382.86
", + "bbox": [ + 228, + 172, + 767, + 224 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "ViT v1 [27], v2 [25], FQ-ViT [30], RepQ-ViT [28], and PTQ4ViT [45], and mixed-precision methods such as LRP-QViT [36] and VT-PTQ [32]. In object detection (Mask R-CNN), we use fixed-precision baselines FQ-ViT [30], PSAQ-ViT v2 [25], RepQ-ViT [28], and LRP-QViT [36] for mixed-precision comparison.", + "bbox": [ + 212, + 237, + 782, + 299 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For semantic segmentation (UperNet), PSAQ-ViT v2 [25] serves as the baseline. Evol-Q is excluded from the main comparison as it does not fully quantize a model starting from FP, and is presented for ablations in the supplementary.", + "bbox": [ + 212, + 299, + 509, + 388 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/bc3de0682dd7d11fa5f6ac5236957b9200d3b42368c3b52f47e6a003a864ec89.jpg", + "table_caption": [ + "Table 3: Hyperparameters list." + ], + "table_footnote": [], + "table_body": "
ParameterDescriptionValue
GGeneration Iterations500
NNeighborhood Size3 × 3
n# Positive Patches4
PPasses10
CCycles6
K# Initial Candidates15
BBatch Size32
bBit-width[2,8]
", + "bbox": [ + 552, + 314, + 751, + 387 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Experimental Setup. The CLAMP-ViT framework is implemented in PyTorch, and evaluated on a single NVIDIA Titan GPU. It features multiple hyperparameters, detailed in Tab. 3.", + "bbox": [ + 212, + 388, + 782, + 434 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Quantization Results for Image Classification", + "text_level": 1, + "bbox": [ + 215, + 454, + 633, + 469 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As highlighted in Sec. 3.5, CLAMP-ViT requires a batch $\\mathcal{B}$ of task-specific targets $T_{G_B}$ . For image classification on the ImageNet-1K, we create $T_{G_B} \\in \\mathcal{R}^{\\mathcal{B} \\times 1000}$ , where the class-wise probabilities are", + "bbox": [ + 212, + 477, + 509, + 551 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/34892a34dfa724d3a091bb6794e222db828a209a5ef91c47dffdac2e1574c3da.jpg", + "table_caption": [ + "Table 4: Comparison of quantized DeiT-S size (MB) and BOPS (G) [1]." + ], + "table_footnote": [], + "table_body": "
MethodW/ABOPSSize
Baseline32/32471088
PSAQ-ViT v24/829422
CLAMP-ViT (Ours)MP4.7/MP5.926720
", + "bbox": [ + 524, + 507, + 777, + 542 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "randomly determined and assigned. We discuss and compare the performance of CLAMP-ViT for two settings, fixed-precision (Tab. 1) and mixed-precision (Tab. 2). In specific, as shown in Tab. 1 CLAMP-ViT consistently provides similar or better accuracy at W8/A8, while for lower precision W4/A8 CLAMP-ViT shows significant performance boost over all the existing alternatives. We yield $\\sim 2.2\\%$ and $\\sim 1\\%$ average accuracy improvement compared to DFQ methods [25, 27] and data-driven methods, respectively. The superior performance of CLAMP-ViT can be attributed to the cyclically adaptive data-generation process, which ensures the generated data matches the requirements and representational capabilities of the quantized model and effective traversal of the search space through evolutionary search and contrastive learning. Whereas, PSAQ-ViT v2 [25], generates increasingly difficult samples which is less beneficial for aggressive 4-bit quantization. Surprisingly, PSAQ-ViT v1 [27] achieves poor accuracy of $25.34\\%$ (W4/A8) on ViT-B despite achieving reasonable accuracy on other ViTs. This result potentially supports our initial intuition that PSAQ-ViT [25, 27] does not consider the informativeness of the generated data to the quantization process.", + "bbox": [ + 212, + 553, + 787, + 809 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evident from Tab. 2, CLAMP-ViT consistently outperforms all baselines across models for the mixed-precision setting, maintaining accuracy close to", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna", + "bbox": [ + 271, + 114, + 671, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/64b7ca794a80f9f571454ee9312799a27f96ad95d864b5d99d2b59d4a04b5675.jpg", + "table_caption": [ + "Table 5: Mixed and fixed-precision quantization performance comparison against SoTA techniques for object detection on COCO 2017. The values in bold and underline signifies, the best performance overall, and with synthetic data, respectively." + ], + "table_footnote": [], + "table_body": "
DeiT-SSwin-S
MethodDataW/A\\( AP^{box} \\)\\( AP^{mask} \\)W/A\\( AP^{box} \\)\\( AP^{mask} \\)
Baseline-32/3248.041.432/3251.844.7
FQ-ViT [30]R8/847.440.98/850.844.1
PSAQ-ViT v2 [25]S8/847.340.88/850.944.1
RepQ-ViT [28]R8/847.941.18/851.644.6
CLAMP-ViT (Ours)S8/847.941.18/851.444.6
FQ-ViT [30]R4/845.140.24/848.241.3
PSAQ-ViT v2 [25]S4/844.838.84/847.941.4
RepQ-ViT [28]R4/845.639.54/849.242.8
CLAMP-ViT (Ours)S4/845.438.94/848.542.2
LRP-QViT [36]R---\\( MP_{6}/MP_{6} \\)51.444.6
CLAMP-ViT (Ours)S\\( MP_{5.5}/MP_{6.8} \\)\\( 47.9 \\)\\( 41.0 \\)\\( MP_{5.1}/MP_{6.4} \\)51.744.6
", + "bbox": [ + 299, + 186, + 696, + 308 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/7d227a6644140003ce2123698799e042cb2ea94f845acda3ee35b8e60211693e.jpg", + "table_caption": [ + "Table 6: Mixed-precision quantization performance comparison against PSAQ-ViT v2 for semantic segmentation. The values in bold indicate best performance overall." + ], + "table_footnote": [], + "table_body": "
DeiT-SSwin-S
MethodDataW/AmIoUW/AmIoU
Baseline-32/3244.032/3249.3
PSAQ-ViT v2 [25]S4/839.94/844.6
CLAMP-ViT (Ours)SMP4.8/MP6.242.4MP5.1/MP6.445.9
", + "bbox": [ + 326, + 342, + 668, + 397 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "the FP baseline despite having a significantly lower average W/A (for mixed-precision W/A is calculated by averaging bit-widths over all the layers for weights and activations). Furthermore, in Tab. 4, we report the reduction in model size and bit-operations-per-second (BOPS) [1] of mixed-precision quantized DeiT-S compared with W4/A8 PSAQ-ViT v2. CLAMP-ViT achieves $\\sim 10\\%$ lower model size in MB and BOPS when compared to W4/A8 PSAQ-ViT v2 while yielding $3.07\\%$ improved accuracy.", + "bbox": [ + 212, + 409, + 787, + 518 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Quantization Results for Object Detection", + "text_level": 1, + "bbox": [ + 214, + 534, + 612, + 550 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The target for object detection is $T_{G_{\\mathcal{B}}} \\in \\mathcal{R}^{\\mathcal{B} \\times bb \\times 5}$ where $bb$ is the number of bounding boxes in the image that is randomly selected from the integer set [1, 3]. $T_{G_{\\mathcal{B}}}[\\mathcal{B}, :, 0]$ corresponds to the bounding box category and $T_{G_{\\mathcal{B}}}[\\mathcal{B}, :, 1:4]$ is the bounding box coordinates $x, y, w, h$ [16]. Tab. 5 presents the fixed- and mixed-precision performance of CLAMP-ViT with respect to the baselines. Across different settings and models, CLAMP-ViT consistently outperforms DFQ method PSAQ-ViT v2 [25] by 0.6 box AP and 0.4 mask AP on average while closely matching performance to the SoTA data-driven method, RepQ-ViT [28]. Similar to Sec. 4.2, we observe improved performance with mixed-precision quantization, achieving near FP baseline performance. The average W/A for mixed-precision quantization for object detection is found to be higher than that for image classification due to the higher complexity of the task demanding larger bit-widths to maintain accuracy.", + "bbox": [ + 212, + 556, + 787, + 752 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Quantization Results for Semantic Segmentation", + "text_level": 1, + "bbox": [ + 214, + 772, + 663, + 787 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The target $T_{G_{\\mathcal{B}}}$ for this task is a pixel-wise classification map of the same size as $X_{\\mathcal{B}}$ i.e., $T_{G_{\\mathcal{B}}} \\in \\mathcal{R}^{\\mathcal{B} \\times 150 \\times H \\times W}$ . In Tab. 6, we show the quantization performance comparison, where CLAMP-ViT achieves average weight bit-width close", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "CLAMP-ViT", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6c6290827723187680d48497f20b367849fb261adbfe0f9be3ec406328d8cd56.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 272, + 143, + 424, + 220 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/83c64bc1cce77a1ff65fd6cd38dbc3d6843f22a3ba31a4f5dbee23c678ec9a02.jpg", + "image_caption": [ + "(b)", + "Fig. 4: Comparison of synthetic data generated by (a) PSAQ-ViT v1 [27], (b) PSAQ-ViT v2 [25] and (c) CLAMP-ViT (Ours). CLAMP-ViT generates detailed objects within contextually suitable backgrounds, boosting realism and informativeness." + ], + "image_footnote": [], + "bbox": [ + 426, + 143, + 576, + 220 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/28505fc328248d6aa44a7da75a5079a15d7c0f4f9bf5ca833877f92a5808616f.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 578, + 143, + 728, + 220 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "to 4-bit while also significantly reducing activation bit-width to $\\sim$ 6-bits. At a higher compression ratio of $15\\%$ , CLAMP-ViT achieves $1.5\\mathrm{mIoU}$ improvement on average over PSAQ-ViT v2.", + "bbox": [ + 212, + 294, + 787, + 339 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Analysis of Generated Samples", + "text_level": 1, + "bbox": [ + 214, + 364, + 517, + 380 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Fig. 4 visualizes generated samples from PSAQ-ViT v1 [27], v2 [25], and CLAMP-ViT (after $1^{st}$ round of stage 1 execution). PSAQ-ViT v1 (Fig. 4(a)) creates images with clear class-specific foregrounds but with overly simplistic and uniform backgrounds, resulting in a lack of realism potentially affecting model accuracy. PSAQ-ViT v2 (Fig. 4(b)) introduces more complex details but fails to convey meaningful semantic information, generating images with intricate but semantically vague structures due to its unguided, difficulty-increasing data-generation strategy. In contrast, CLAMP-ViT (Fig. 4(c)) excels by synthesizing data that mirrors real-world imagery, showcasing a sophisticated understanding of semantic relationships between patches. It ensures objects are detailed and in contextually fitting backgrounds, boosting realism and informativeness. For example, CLAMP-ViT places boats on water and zebras in grasslands (Fig. 4(c), row 2), showing its capability for creating semantically relevant and visually consistent synthetic data. We believe our patch semantics exploration with a contrastive objective, makes image generation informative that mimic real-world scenes.", + "bbox": [ + 212, + 392, + 787, + 619 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.6 Ablations and Discussions", + "text_level": 1, + "bbox": [ + 214, + 645, + 478, + 659 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Evolutionary Search Parameters. In Fig. 5(a), we detail an experiment to determine the ideal number of passes $\\mathcal{P}$ and cycles $\\mathcal{C}$ for the evolutionary search process by studying the variation in Top-1 accuracy of DeiT-S with different passes and cycles keeping the other fixed at their optimal value (Tab. 3). It is evident that a cycle count of $\\mathcal{C} = 6$ is optimal, as accuracy tends to decline with more cycles. Conversely, passes show a modest yet consistent improvement beyond 10, but due to the substantial rise in computational complexity, $\\mathcal{P} = 10$ is deemed the most suitable choice.", + "bbox": [ + 212, + 672, + 787, + 792 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effect of Batch Size $\\mathcal{B}$ . We also show the accuracy comparison with different batch sizes ranging from 8 to 64 in Fig. 5(c). It is apparent that there is minimal increase in accuracy beyond 32 for CLAMP-ViT, justifying the choice of batch", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna", + "bbox": [ + 271, + 114, + 673, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/30b03d79aa08403d1b930a36bd5f2bbb9debe5b573002b26a38fbcdf2617b0f7.jpg", + "image_caption": [ + "(a)", + "Fig. 5: CLAMP-ViT ablations for (a) Selecting evolutionary search parameters, (b) Mixed-precision quantization accuracy with different fitness functions, (c) Effect of batch size $\\mathcal{B}$ and (d) Effect of neighborhood size $\\mathcal{N}$ and top- $n$ positive patches." + ], + "image_footnote": [], + "bbox": [ + 217, + 148, + 349, + 224 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/08b76e5c90f863a817852fd9c607d2fb91aac6c95c053bae1fd7aad3e817ac2d.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 351, + 148, + 496, + 224 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a6f227ecbec8a773d1df01ea21fb7d3ffd9c03d43ddb6706140b2154902ebba8.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 500, + 147, + 633, + 224 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3cb97402d4e23d02da780da7b3e7c54cc37420712caa688ad4481792763f2b57.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 638, + 146, + 781, + 224 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "size. Interestingly, CLAMP-ViT achieves a decent accuracy of $\\sim 70\\%$ even with a batch size of 8, while PSAQ-ViT v1 and v2 achieve only close to $60\\%$ .", + "bbox": [ + 212, + 296, + 784, + 327 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effect of Neighborhood Size $\\mathcal{N}$ and Top- $n$ Patches. We explore $3 \\times 3$ , $5 \\times 5$ , and all patches for neighborhood sizes and $n \\in [1,4]$ for DeiT-S as shown in Fig. 5(d). A $3 \\times 3$ neighborhood with top-4 patches as positives yields the highest accuracy while being computationally least demanding. In the $5 \\times 5$ scenario, as the balance between positive and negative samples improves with increasing $n$ , accuracy rises but at higher computational cost. Employing all patches is computationally unviable and leads to the lowest accuracy for small $n$ values due to a positive-negative imbalance [39]. Moreover, identifying distant positive patches from the anchor neglects significant semantic patch relationships.", + "bbox": [ + 212, + 330, + 784, + 468 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Choice of Objective Function. The study in Tab. 7 examines how different loss function components of $\\mathcal{L}^{SG}$ i.e., $\\mathcal{L}^{C_1}$ and $\\mathcal{L}^O$ affect synthetic data generation effectiveness and its impact on top-1 accuracy of W4/A8 quantization of DeiT-S.We chose fixed-precision quantization", + "bbox": [ + 212, + 472, + 509, + 577 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/7390af3a079be8ff21e74e551cff9693112d06194bd40bada3fbafdac9ea6ae1.jpg", + "table_caption": [ + "Table 7: Impact of different loss components in synthetic data generation." + ], + "table_footnote": [], + "table_body": "
LPSELC1LOW/ATop-1Accuracy
---32/3279.85
✓✗✗4/874.18
✗✗✓4/833.93
✓✗✓4/876.47
✗✓✗4/876.59
✗✓✓4/877.03
✓✓✓4/876.98
", + "bbox": [ + 549, + 502, + 754, + 570 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "to avoid any bias from mixed-precision quantization, which might typically favor higher bitwidths to lessen accuracy loss due to low-quality images. Results show that a linear mix of $\\mathcal{L}^{C_1}$ and $\\mathcal{L}^O$ ( $\\mathcal{L}^{SG}$ ) achieves highest accuracy, while using $\\mathcal{L}^O$ alone leads to the lowest, indicating its limited utility in leveraging ViTs for synthetic data. In Tab. 7 $\\mathcal{L}^{PSE}$ corresponds to the patch similarity metric employed in [25]. While combining $\\mathcal{L}^{PSE}$ with $\\mathcal{L}^O$ [25] does offer a moderate accuracy boost, it falls short of the performance with $\\mathcal{L}^{SG}$ , due to inherent limitations of the patch similarity metric highlighted in Sec. 1. Furthermore, using all three loss functions simultaneously closely matches the performance of $\\mathcal{L}^{SG}$ further demonstrating that our proposed $\\mathcal{L}^{C_1}$ has the major contribution towards final accuracy.", + "bbox": [ + 212, + 578, + 787, + 743 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "For mixed-precision quantization of DeiT-S, the fitness function's accuracy $(\\mathcal{L}^F = \\mathcal{L}^{C_2} + \\mathcal{L}^O)$ is compared against global contrastive loss [14], MSE, and KL-divergence in Fig. 5(b). The accuracy analysis shows MSE and KL-divergence tend to overfit to synthetic data, evidenced by plateauing accuracy. Meanwhile, global contrastive loss initially matches but then accuracy gap widens from CLAMP-ViT's performance which is due to premature convergence, implying", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "CLAMP-ViT", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "that quantifying intermediate layer distributional divergence is crucial to find optimal quantization parameters.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effect of Adaptivity. We investigate the effectiveness of the cyclic evolution every $\\mathcal{C} / 2$ iterations to ensure the data generation adapts to the requirements of", + "bbox": [ + 212, + 176, + 509, + 236 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/98bb99737e72f4ac98ff87aeec01979247365bf7cfd3c20d6268a848f979e273.jpg", + "table_caption": [ + "Table 8: Effect of adaptability on quantized model accuracy." + ], + "table_footnote": [], + "table_body": "
AdaptivityW/ATop-1 Accuracy
XMP4.9/MP6.178.06
✓MP4.7/MP5.978.97
", + "bbox": [ + 542, + 205, + 761, + 234 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "the quantized model in Tab. 8. It can be observed that with adaptivity, we are not only able to achieve lower average W/A but also have improved accuracy.", + "bbox": [ + 212, + 237, + 782, + 267 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effect of Informativeness. To show CLAMP-ViT's effectiveness in generating informative data, we compare misprediction rates in PSAQ-ViT v1, v2, and CLAMP-ViT during quantization. PSAQ-ViT has higher misprediction rates (34% for v1, 41% for v2) than CLAMP-ViT's 22%, indicating PSAQ-ViT's data is less informative, leading to erratic predictions, sub-optimal performance.", + "bbox": [ + 212, + 267, + 542, + 402 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitation Discussion: Runtime Comparison. In Fig. 6, we show the runtime and", + "bbox": [ + 212, + 402, + 542, + 433 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "top-1 accuracy of different techniques on an CLAMP-ViT (mixed-precision) achieves signi to the other methods (fixed-precision) with o $(5\\% \\uparrow)$ compared to the best DFQ method. N calibrate a pre-quantized model on real data.", + "bbox": [ + 212, + 433, + 542, + 508 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/638b425e46e9de1f6e5fed0bbf8681ee55f6d7baf05e22d6b6f1915501b64ef5.jpg", + "image_caption": [ + "Fig.6: Runtime of different PTQ and DFQ methods where, FS: FP model as starting point. QS: Prequantized model as starting point." + ], + "image_footnote": [], + "bbox": [ + 586, + 273, + 750, + 369 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusions", + "text_level": 1, + "bbox": [ + 214, + 530, + 367, + 546 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper presents CLAMP-ViT, a novel mixed-precision DFQ technique using cyclic adaptation and contrastive learning. It employs patch-level contrastive learning that leverages properties of the MHSA modules for data generation. A local contrastive objective and layer-wise evolutionary search identify optimal quantization parameters while ensuring a smooth loss landscape. Experiments across CV tasks show superior performance of CLAMP-ViT, achieving up to $3\\%$ top-1 accuracy for classification, 0.6 mAP for detection, and 1.5 mIoU for segmentation. Future work aims to focus on extending its application to a wider range of architectures, like VLMs. While this study focuses on a useful impact and beneficial application of synthetic data generation for optimized and carbon efficient models for deployment, it is important to also be cognizant of the potential adverse effects of synthetic data such as deepfakes or racial biases.", + "bbox": [ + 212, + 560, + 787, + 743 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 763, + 403, + 781 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was supported in part by CoCoSys, one of the seven centers in JUMP 2.0, a Semiconductor Research Corporation (SRC) program sponsored by DARPA.", + "bbox": [ + 212, + 794, + 784, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna", + "bbox": [ + 271, + 114, + 671, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Baskin, C., Liss, N., Schwartz, E., Zheltonozhskii, E., Giryes, R., Bronstein, A.M., Mendelson, A.: Uniq: Uniform noise injection for non-uniform quantization of neural networks. ACM Transactions on Computer Systems (TOCS) 37(1-4), 1-15 (2021)", + "2. Cai, Y., Yao, Z., Dong, Z., Gholami, A., Mahoney, M.W., Keutzer, K.: Zeroq: A novel zero shot quantization framework. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13169-13178 (2020)", + "3. Cai, Z., Vasconcelos, N.: Cascade r-cnn: Delving into high quality object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6154-6162 (2018)", + "4. Cao, Y.H., Sun, P., Huang, Y., Wu, J., Zhou, S.: Synergistic self-supervised and quantization learning. In: European Conference on Computer Vision. pp. 587-604. Springer (2022)", + "5. Chen, H., Shao, S., Wang, Z., Shang, Z., Chen, J., Ji, X., Wu, X.: Bootstrap generalization ability from loss landscape perspective. In: European Conference on Computer Vision. pp. 500-517. Springer (2022)", + "6. Chen, K., Wang, J., Pang, J., Cao, Y., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Xu, J., et al.: Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155 (2019)", + "7. Choi, K., Hong, D., Park, N., Kim, Y., Lee, J.: Qimera: Data-free quantization with synthetic boundary supporting samples. Advances in Neural Information Processing Systems 34, 14835-14847 (2021)", + "8. Chuang, C.Y., Hjelm, R.D., Wang, X., Vineet, V., Joshi, N., Torralba, A., Jegelka, S., Song, Y.: Robust contrastive learning against noisy views. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16670-16681 (2022)", + "9. Contributors, M.: MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation (2020)", + "0. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)", + "1. Dong, P., Li, L., Wei, Z., Niu, X., Tian, Z., Pan, H.: Emq: Evolving training-free proxies for automated mixed precision quantization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17076-17086 (2023)", + "2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "3. Fei, W., Dai, W., Li, C., Zou, J., Xiong, H.: General bitwidth assignment for efficient deep convolutional neural network quantization. IEEE Transactions on Neural Networks and Learning Systems 33(10), 5253-5267 (2021)", + "4. Frumkin, N., Gope, D., Marculescu, D.: Jumping through local minima: Quantization in the loss landscape of vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16978-16988 (2023)", + "5. Fu, Y., Yu, Q., Li, M., Ouyang, X., Chandra, V., Lin, Y.: Contrastive quant: quantization makes stronger contrastive learning. In: Proceedings of the 59th ACM/IEEE Design Automation Conference. pp. 205-210 (2022)" + ], + "bbox": [ + 225, + 178, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "CLAMP-ViT", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Huang, H., Yu, P.S., Wang, C.: An introduction to image synthesis with generative adversarial nets. arXiv preprint arXiv:1803.04469 (2018)", + "17. Hubara, I., Courbariaux, M., Soudry, D., El-Yaniv, R., Bengio, Y.: Binarized neural networks. Advances in neural information processing systems 29 (2016)", + "18. Hubara, I., Nahshan, Y., Hanani, Y., Banner, R., Soudry, D.: Accurate post training quantization with small calibration sets. In: International Conference on Machine Learning. pp. 4466-4475. PMLR (2021)", + "19. Kim, N., Shin, D., Choi, W., Kim, G., Park, J.: Exploiting retraining-based mixed-precision quantization for low-cost dnn accelerator design. IEEE Transactions on Neural Networks and Learning Systems 32(7), 2925–2938 (2020)", + "20. Kundu, S., Sun, Q., Fu, Y., Pedram, M., Beerel, P.: Analyzing the confidentiality of undistillable teachers in knowledge distillation. Advances in Neural Information Processing Systems 34, 9181-9192 (2021)", + "21. Kundu, S., Wang, S., Sun, Q., Beerel, P.A., Pedram, M.: Bmpq: bit-gradient sensitivity-driven mixed-precision quantization of dnns from scratch. In: 2022 Design, Automation & Test in Europe Conference & Exhibition (DATE). pp. 588-591. IEEE (2022)", + "22. Li, H., Xu, Z., Taylor, G., Studer, C., Goldstein, T.: Visualizing the loss landscape of neural nets. Advances in neural information processing systems 31 (2018)", + "23. Li, Y., Xu, S., Zhang, B., Cao, X., Gao, P., Guo, G.: Q-vit: Accurate and fully quantized low-bit vision transformer. Advances in Neural Information Processing Systems 35, 34451-34463 (2022)", + "24. Li, Y., Yuan, G., Wen, Y., Hu, J., Evangelidis, G., Tulyakov, S., Wang, Y., Ren, J.: Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems 35, 12934-12949 (2022)", + "25. Li, Z., Chen, M., Xiao, J., Gu, Q.: Psaq-vit v2: Toward accurate and general data-free quantization for vision transformers. IEEE Transactions on Neural Networks and Learning Systems (2023)", + "26. Li, Z., Gu, Q.: I-vit: integer-only quantization for efficient vision transformer inference. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17065-17075 (2023)", + "27. Li, Z., Ma, L., Chen, M., Xiao, J., Gu, Q.: Patch similarity aware data-free quantization for vision transformers. In: European Conference on Computer Vision. pp. 154-170. Springer (2022)", + "28. Li, Z., Xiao, J., Yang, L., Gu, Q.: Repq-vit: Scale reparameterization for posttraining quantization of vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17227-17236 (2023)", + "29. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "30. Lin, Y., Zhang, T., Sun, P., Li, Z., Zhou, S.: Fq-vit: Post-training quantization for fully quantized vision transformer. arXiv preprint arXiv:2111.13824 (2021)", + "31. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)", + "32. Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., Gao, W.: Post-training quantization for vision transformer. Advances in Neural Information Processing Systems 34, 28092-28103 (2021)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna", + "bbox": [ + 271, + 114, + 671, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Peters, J.W., Welling, M.: Probabilistic binary neural networks. arXiv preprint arXiv:1809.03368 (2018)", + "34. Ramachandran, A., Dhiman, A., Vandrotti, B.S., Kim, J.: Ntrans-net: A multiscale neutrosophic-uncertainty guided transformer network for indoor depth completion. In: 2023 IEEE International Conference on Image Processing (ICIP). pp. 905–909. IEEE (2023)", + "35. Ramachandran, A., Wan, Z., Jeong, G., Gustafson, J., Krishna, T.: Algorithm-hardware co-design of distribution-aware logarithmic-posit encodings for efficient dnn inference. arXiv preprint arXiv:2403.05465 (2024)", + "36. Ranjan, N., Savakis, A.: Lrp-qvit: Mixed-precision vision transformer quantization via layer-wise relevance propagation. arXiv preprint arXiv:2401.11243 (2024)", + "37. Strudel, R., Garcia, R., Laptev, I., Schmid, C.: Segmenter: Transformer for semantic segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7262-7272 (2021)", + "38. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: International conference on machine learning. pp. 10347-10357. PMLR (2021)", + "39. Wang, J., Li, J., Li, W., Xuan, L., Zhang, T., Wang, W.: Positive-negative equal contrastive loss for semantic segmentation. Neurocomputing 535, 13-24 (2023)", + "40. Wightman, R.: Pytorch image models. https://github.com/rwrightman/pytorch-image-models (2019). https://doi.org/10.5281/zenodo.4414861", + "41. Xiao, J., Li, Z., Yang, L., Gu, Q.: Patch-wise mixed-precision quantization of vision transformer. arXiv preprint arXiv:2305.06559 (2023)", + "42. Xiao, T., Liu, Y., Zhou, B., Jiang, Y., Sun, J.: Unified perceptual parsing for scene understanding. In: Proceedings of the European conference on computer vision (ECCV). pp. 418-434 (2018)", + "43. Yeh, C.H., Hong, C.Y., Hsu, Y.C., Liu, T.L., Chen, Y., LeCun, Y.: Decoupled contrastive learning. In: European Conference on Computer Vision. pp. 668-684. Springer (2022)", + "44. Yin, H., Molchanov, P., Alvarez, J.M., Li, Z., Mallya, A., Hoiem, D., Jha, N.K., Kautz, J.: Dreaming to distill: Data-free knowledge transfer via deepinversion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8715-8724 (2020)", + "45. Yuan, Z., Xue, C., Chen, Y., Wu, Q., Sun, G.: Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization. In: European Conference on Computer Vision. pp. 191-207. Springer (2022)", + "46. Zhang, D., Yang, J., Ye, D., Hua, G.: Lq-nets: Learned quantization for highly accurate and compact deep neural networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 365-382 (2018)", + "47. Zhang, S., Zhou, Q., Wang, Z., Wang, F., Yan, J.: Patch-level contrastive learning via positional query for visual pre-training (2023)", + "48. Zhang, Y., Guo, X., Poggi, M., Zhu, Z., Huang, G., Mattoccia, S.: Completionformer: Depth completion with convolutions and vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18527-18536 (2023)", + "49. Zhang, Y., Chen, D., Kundu, S., Li, C., Beerel, P.A.: Sal-vit: Towards latency efficient private inference on vit using selective attention search with a learnable softmax approximation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5116-5125 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "CLAMP-ViT", + "bbox": [ + 640, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "50. Zhang, Z., Lu, X., Cao, G., Yang, Y., Jiao, L., Liu, F.: Vit-yolo: Transformer-based yolo for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2799-2808 (2021)", + "51. Zhong, Y., Lin, M., Nan, G., Liu, J., Zhang, B., Tian, Y., Ji, R.: Intraq: Learning synthetic images with intra-class heterogeneity for zero-shot network quantization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12339-12348 (2022)", + "52. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127, 302-321 (2019)" + ], + "bbox": [ + 215, + 146, + 787, + 285 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna", + "bbox": [ + 271, + 114, + 671, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/CLAMP-ViT_ Contrastive Data-Free Learning for Adaptive Post-Training Quantization of ViTs/9ac2383c-dda4-4497-a269-a04b9fe79777_model.json b/2024/CLAMP-ViT_ Contrastive Data-Free Learning for Adaptive Post-Training Quantization of ViTs/9ac2383c-dda4-4497-a269-a04b9fe79777_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b4373e77b4456ce07b6d07dd9f427b478e1c0bb2 --- /dev/null +++ b/2024/CLAMP-ViT_ Contrastive Data-Free Learning for Adaptive Post-Training Quantization of ViTs/9ac2383c-dda4-4497-a269-a04b9fe79777_model.json @@ -0,0 +1,2656 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.219, + 0.141, + 0.784, + 0.187 + ], + "angle": 0, + "content": "CLAMP-ViT: Contrastive Data-Free Learning for Adaptive Post-Training Quantization of ViTs" + }, + { + "type": "text", + "bbox": [ + 0.249, + 0.213, + 0.756, + 0.229 + ], + "angle": 0, + "content": "Akshat Ramachandran1, Souvik Kundu2, and Tushar Krishna1" + }, + { + "type": "text", + "bbox": [ + 0.322, + 0.24, + 0.681, + 0.255 + ], + "angle": 0, + "content": "1 Georgia Institute of Technology, Atlanta, GA, USA" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.256, + 0.665, + 0.269 + ], + "angle": 0, + "content": "akshat.r@gatech.edu, tushar@ece.gatech.edu" + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.269, + 0.616, + 0.282 + ], + "angle": 0, + "content": "\\(^{2}\\) Intel Labs, San Diego, CA, USA" + }, + { + "type": "text", + "bbox": [ + 0.412, + 0.284, + 0.593, + 0.295 + ], + "angle": 0, + "content": "souvikk.kundu@intel.com" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.336, + 0.744, + 0.573 + ], + "angle": 0, + "content": "Abstract. We present CLAMP-ViT, a data-free post-training quantization method for vision transformers (ViTs). We identify the limitations of recent techniques, notably their inability to leverage meaningful inter-patch relationships, leading to the generation of simplistic and semantically vague data, impacting quantization accuracy. CLAMP-ViT employs a two-stage approach, cyclically adapting between data generation and model quantization. Specifically, we incorporate a patch-level contrastive learning scheme to generate richer, semantically meaningful data. Furthermore, we leverage contrastive learning in layer-wise evolutionary search for fixed- and mixed-precision quantization to identify optimal quantization parameters while mitigating the effects of a non-smooth loss landscape. Extensive evaluations across various vision tasks demonstrate the superiority of CLAMP-ViT, with performance improvements of up to \\(3\\%\\) in top-1 accuracy for classification, 0.6 mAP for object detection, and 1.5 mIoU for segmentation at similar or better compression ratio over existing alternatives. Code is available at https://github.com/georgia-tech-synergy-lab/CLAMP-ViT.git" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.586, + 0.741, + 0.613 + ], + "angle": 0, + "content": "Keywords: Data-free quantization \\(\\cdot\\) PTQ \\(\\cdot\\) Contrastive learning \\(\\cdot\\) Vision transformer" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.641, + 0.377, + 0.657 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Vision transformers [12] (ViTs) have recently gained a lot of traction due to their state-of-the-art (SoTA) performance across various computer vision (CV) tasks [34, 37, 48, 50]. Concurrently, the growing need to deploy these parameter-heavy models at the resource-limited edge [14], has inspired research on various model compression techniques. Model quantization [14, 30, 36, 45] has emerged as a popular technique to achieve memory and compute efficient deployment. Quantization reduces memory footprint and improves computation speed of a model by mapping full-precision (FP) weights to reduced precision formats (e.g., \\(\\leq\\) 8-bit INT) [13, 17, 19, 46]. In particular, quantization-aware training (QAT) allows a model to train by taking quantization approximation in to account, enabling ease of quantization. Post-training quantization (PTQ), in contrast, acts" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.673, + 0.129 + ], + "angle": 0, + "content": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.238 + ], + "angle": 0, + "content": "as a plug-and-play quantization applied on a pre-trained model. PTQ has become popular as it can leverage the pre-trained model and does not add training overhead [21, 23, 26]. We thus consider the PTQ setting in this work. However, PTQ requires access to a calibration set that is often drawn from the training data [18, 45]. This may be infeasible in situations involving privacy and security concerns [20, 49], making these techniques ill-suited to yield optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.242, + 0.488, + 0.335 + ], + "angle": 0, + "content": "Recent works [25, 27] propose data-free PTQ (DFQ), generating synthetic calibration data \\( T_{\\mathrm{syn}} \\) from Gaussian noise [7, 27], embedding information from the original dataset \\( T_{\\mathrm{orig}} \\), where \\( T_{\\mathrm{syn}} << T_{\\mathrm{orig}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.338, + 0.487, + 0.429 + ], + "angle": 0, + "content": "Existing DFQ methods for CNNs [2,51] exploit the batch-normalization (BN) layer statistics [33,44] to generate synthetic samples mimicking the original data distribution. For ViTs, absence of the BN layer makes these" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.242, + 0.641, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.242, + 0.782, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.355, + 0.788, + 0.412 + ], + "angle": 0, + "content": "Fig.1: Visualization of loss landscape of (a) PSAQ-ViT v2 and (b) CLAMP-ViT on DeiT-S with perturbations to quantized model weights and synthetic data [22]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.789, + 0.702 + ], + "angle": 0, + "content": "techniques obsolete. Recent efforts to extend DFQ to ViTs include PSAQ-ViT v1 [27] and PSAQ-ViT v2 [25]. They rely on information embedded in the attention score output of the multi-head self attention (MHSA) layer. PSAQ-ViT v1 and v2 [25,27] introduce a relative value metric namely, patch similarity, to optimize Gaussian noise towards synthetic data by maximizing the entropy of patch similarity in a global manner. However, the metric considered in PSAQ-ViT v1 and v2 assumes all patches1 to be equally important, without considering spatial sensitivity [47]. This may fail to capture semantically meaningful inter-patch relations, potentially affecting robustness of the synthetic data. As we can see in Fig. 1(a) even insignificant perturbations in the generated images (augmenting pixels) or weights (to simulate quantization process) may cause significant jaggedness in the loss landscape of PSAQ-ViT v2. This also implies that the predictions may have large discrepancy even for small input/weight perturbation. Moreover, the synthetic data generation stage in these methods does not consider the informativeness of the generated samples towards the quantization process, nor do they establish countermeasures to ameliorate the non-smooth loss landscape during quantization, resulting in sub-optimal parameter search and poor generalization to test set [5]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.706, + 0.789, + 0.798 + ], + "angle": 0, + "content": "Our contributions. The discussion above hints at the potential limitation of [25,27] in capturing semantically meaningful and robust inter-patch relationships to generate synthetic data that is well-suited to quantization. Towards solving these limitations, we present contrastive data-free learning for adaptive post-training quantization of ViTs (CLAMP-ViT), a general DFQ method applicable to a wide range of vision tasks. To the best of our knowledge, CLAMP-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.668, + 0.841 + ], + "angle": 0, + "content": "1 Patch (subset of image): group of neighboring pixels in an image." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.641, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLAMP-ViT" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.786, + 0.176 + ], + "angle": 0, + "content": "ViT is the first work to support both fixed-² and mixed-precision³ DFQ of ViTs, starting from a pre-trained FP model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.787, + 0.328 + ], + "angle": 0, + "content": "Specifically, CLAMP-ViT utilizes the architectural characteristics of ViTs and inherent properties of real-images to generate semantically rich and meaningful data while ensuring spatial sensitivity. Here we leverage a novel patch-level contrastive learning scheme, where for each patch (anchor) in the MHSA layer output, we treat semantically similar patches in a neighborhood around the patch as positive patches (evaluated using cosine similarity) and the remaining patches in the neighborhood as negative patches (see Fig. 3 for an intuitive visualization). We then leverage a patch-level contrastive loss that drives the representation of the anchor patch closer to the positive patches and away from the negative patches, enabling exploration of semantically meaningful relations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.787, + 0.525 + ], + "angle": 0, + "content": "Recently, Evol-Q [14] identified a limitation of gradient based methods [25] to search over quantization parameters of ViTs having non-smooth loss landscape, leading to poor accuracy and generalization. We take inspiration from this and a recent work [35] to design a layer-wise evolutionary search to identify the optimal bit-width and scale factors for each layer. Additionally, we propose a novel local contrastive objective to capture distributional variance in intermediate layer outputs (crucial for layer-wise quantization) and improve search convergence. This loss sufficiently captures the representational divergence of intermediate layers outputs to identify optimal quantization parameters and yields a smooth loss landscape, as demonstrated in Fig. 1(b), enabling generalizability and better performance on test data [5]. Furthermore, to ensure the generated data is adaptive to the quantization process, CLAMP-ViT performs a cyclically adaptive strategy alternating between data-generation and quantization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.525, + 0.787, + 0.57 + ], + "angle": 0, + "content": "To evaluate the merits of CLAMP-ViT we conduct extensive experiments on image classification, detection, and segmentation with different ViT variants and quantization scenarios and observe superior performance over SoTA." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.592, + 0.398, + 0.608 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.622, + 0.788, + 0.804 + ], + "angle": 0, + "content": "Data-Driven PTQ for ViTs. PTQ offers an efficient alternative to QAT [23,26] by directly quantizing pre-trained models without the need for compute-heavy retraining. In specific, PTQ4ViT [45] employs twin uniform quantization and Hessian-guided scale optimization. FQ-ViT [30] uses a power-of-two factor quantization to handle inter-channel variation in LayerNorm and log-INT-Softmax. RepQ-ViT [28] separates quantization and inference, optimizing accuracy and efficiency by employing scale-reparameterized quantizers. These methods apply fixed-precision quantization, assuming all layers support similar approximations, potentially leading to sub-optimal accuracy [36]. On the other hand, techniques like, VT-PTQ [32] adopt mixed precision for specific modules based on sensitivity metrics, while PMQ [41] and LRP-QViT [36] allocate bit-widths by assessing both layer sensitivity and contribution to the output," + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.644, + 0.826 + ], + "angle": 0, + "content": "2 weights/activations quantized to same precision for all layers." + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.702, + 0.84 + ], + "angle": 0, + "content": "3 weights/activations quantized to different precision for different layers." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.811, + 0.702, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.673, + 0.129 + ], + "angle": 0, + "content": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "respectively. However, all these methods assume a part of training set to be available for calibration which may be infeasible due to privacy [20]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.388 + ], + "angle": 0, + "content": "Data-Free PTQ for ViTs. Recent data-free PTQ efforts (DFQ) for ViTs including PSAQ-ViT v1 [27] and v2 [25] utilize a \"patch similarity\" metric to refine Gaussian noise to synthetic data resembling real images. In specific, both these approaches leverage the fact that the self-attention has different response to real image and noise. PSAQ-ViT v1 [27] uses a two-stage cascaded framework to generate synthetic data for model quantization on image classification tasks. PSAQ-ViT v2 [25] expanded this to a broader range of tasks, employing a Min-Max game between full precision and quantized models for data generation and quantization. Despite these innovations, both the versions face several limitations as highlighted in Sec. 1, leading to poor DFQ performance, particularly at reduced precision. Additionally, these methods support only fixed-precision quantization. CLAMP-ViT, in contrast, introduces a DFQ method that addresses their shortcomings in yielding SoTA performance, while supporting both fixed and mixed-precision quantization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.788, + 0.571 + ], + "angle": 0, + "content": "Contrastive Learning. Contrastive learning as a technique has been widely adopted in self-supervised settings [4,15] and is proven to combat overfitting via regularization against negative samples. Recently, Evol-Q [14] demonstrated the benefits of a global contrastive objective [8,43] coupled with evolutionary search in smoothening the loss landscape while calibrating the scaling factors of a prequantized model to enhance accuracy. Unlike [14] that only aims to adjust the scale factors of each layer of a pre-quantized model, we conduct complete quantization starting from a FP model. On evaluation of the global contrastive objective used in Evol-Q, we find it to be sub-optimal, causing premature convergence while quantizing an FP model. Instead, we present a local contrastive loss that captures the distributional variance in intermediate layer outputs, drastically improving convergence and hence, identification of quantization parameters." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.594, + 0.496, + 0.61 + ], + "angle": 0, + "content": "3 CLAMP-ViT Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.788, + 0.703 + ], + "angle": 0, + "content": "We present an overview of CLAMP-ViT in Fig. 2. In this section, we first go through notations, computational process of ViTs and quantization strategy in the preliminaries, followed by a detailed description of the proposed contrastive loss and the two stages of CLAMP-ViT. Finally, the overall DFQ pipeline is summarized and presented (refer to Supplementary for the detailed Algorithm)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.724, + 0.373, + 0.739 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.794, + 0.841 + ], + "angle": 0, + "content": "Notations. Let \\( X \\in \\mathcal{R}^{H \\times W \\times C} \\) be the input image to an \\( L \\)-layer ViT, where \\( (H, W, C) \\) are the height, width, and channels, respectively (we ignore the batch dimension for simplicity). The input is partitioned into \\( N \\) non-overlapping patches that are then linearly transformed to \\( d \\)-dimensional patch embeddings, \\( (\\mathcal{R}^{N \\times d}) \\) that is passed through an encoder consisting of a series of transformer layers each composed of an MHSA and an MLP module. Each MHSA module is composed" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.641, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLAMP-ViT" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.144, + 0.768, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.43 + ], + "angle": 0, + "content": "Fig.2: Overview of the cyclically evolving two-stage CLAMP-ViT framework. In stage 1 (① - ②), \\(\\mathcal{L}^{SG}\\) is minimized to update Gaussian noise towards synthesizing data. Stage 2 (③ - ⑤) conducts layer-wise evolutionary search to identify optimal quantization parameters while minimizes \\(\\mathcal{L}^F\\). Illustrated with multiple instances of models for clarity, only one instance of each model is used in the framework." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.788, + 0.487 + ], + "angle": 0, + "content": "of \\( h \\) heads, to capture long-range patch correlations [24]. At head \\( j \\), \\( X \\) is transformed to query \\( (Q_j) \\), key \\( (K_j) \\), and value \\( (V_j) \\) tensors to perform self-attention \\( \\varPhi_j \\), that is linearly projected after concatenation over the heads," + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.492, + 0.786, + 0.511 + ], + "angle": 0, + "content": "\\[\n\\Phi_ {j} \\left(Q _ {j}, K _ {j}, V _ {j}\\right) = \\operatorname {s o f t m a x} \\left(Q _ {j} K _ {j} ^ {T} / \\sqrt {d}\\right) V _ {j}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.334, + 0.521, + 0.785, + 0.539 + ], + "angle": 0, + "content": "\\[\n\\operatorname {M H S A} (Q, K, V) = \\mathbf {c o n c a t} \\left(\\Phi_ {0}, \\Phi_ {1}, \\dots \\Phi_ {h - 1}\\right) W ^ {0} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.788, + 0.589 + ], + "angle": 0, + "content": "The output of the \\(j^{th}\\) head at \\(i^{th}\\) MHSA layer is given by \\(O_{i,j} \\in \\mathcal{R}^{N \\times d}\\). The series of transformer layers are succeeded by task-specific heads for classification, detection, or segmentation, depending on the nature of the vision task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.59, + 0.789, + 0.636 + ], + "angle": 0, + "content": "Quantization. In this paper, we perform uniform symmetric quantization (fixed/mixed-precision) of both weights and activations for ViTs, mapping full precision values into a uniform scale determined by bit-width (b), given as," + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.646, + 0.787, + 0.678 + ], + "angle": 0, + "content": "\\[\nQ (X, \\gamma , b) = \\operatorname {c l i p} \\left(\\left\\lfloor \\frac {X}{\\gamma} \\right\\rfloor , - 2 ^ {b - 1} + 1, 2 ^ {b - 1} - 1\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.644, + 0.703 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is the scale factor and \\(X\\) represents the FP tensor." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.725, + 0.445, + 0.74 + ], + "angle": 0, + "content": "3.2 Contrastive Objective" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Contrastive learning based on the infoNCE loss [39] helps learn an anchor sample from both the similar (positive) and dissimilar (negative) samples, typically using a softmax function to normalize the similarities into probabilities. However, infoNCE loss suffers from a disproportionate impact of a single positive and many negative samples [14]. This can affect learning the synthetic data as well as the quantized model parameters. Inspired by [39], we present a modified" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.674, + 0.129 + ], + "angle": 0, + "content": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.178 + ], + "angle": 0, + "content": "infoNCE loss \\((\\mathcal{L}_{i,j}^{C}\\) in Eq. 4) that addresses this imbalance yielding latent impact of positive and negative samples pairs equally." + }, + { + "type": "equation", + "bbox": [ + 0.287, + 0.188, + 0.787, + 0.23 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i, j} ^ {C} = - \\log \\frac {\\sum_ {p +} \\exp \\left(\\lambda_ {i , j} ^ {p} \\cdot \\lambda_ {i , j} ^ {p +} / \\tau\\right)}{\\sum_ {p +} \\exp \\left(\\lambda_ {i , j} ^ {p} \\cdot \\lambda_ {i , j} ^ {p +} / \\tau\\right) + \\sum_ {p -} \\exp \\left(\\lambda_ {i , j} ^ {p} \\cdot \\lambda_ {i , j} ^ {p -} / \\tau\\right)} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.24, + 0.789, + 0.349 + ], + "angle": 0, + "content": "here superscript \\( p \\), \\( p+ \\), and \\( p- \\) correspond to the anchor prediction (drawn from quantized model), positive, and negative prediction (drawn from FP model), respectively. \\( \\tau \\) controls the concentration level [39]. Subscript \\( (i,j) \\) represents (# layer id, # head id), and (# layer id, # batch id) for the data generation and quantization stage, respectively. \\( \\lambda \\) represents the variable taken to evaluate the log likelihood—the intermediate layer patch embeddings (stage 1) or layer output activations (stage 2). The final stage \\( s \\) loss is given as \\( \\mathcal{L}^{C_s} = \\sum_i\\sum_j\\mathcal{L}_{i,j}^{C_s} \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.368, + 0.557, + 0.384 + ], + "angle": 0, + "content": "3.3 Stage 1: Synthetic Data Generation" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.394, + 0.472, + 0.514 + ], + "angle": 0, + "content": "Our goal in Stage 1 is to generate semantically rich and meaningful images that can reliably exploit inter-patch relations while ensuring informativeness to the quantization process. For each \"anchor patch\" of an MHSA output in the quantized model, we first locate positive and" + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.398, + 0.769, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.481, + 0.481, + 0.788, + 0.51 + ], + "angle": 0, + "content": "Fig. 3: Intuitive visualization of positive and negative patch selection in Stage 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.514, + 0.788, + 0.621 + ], + "angle": 0, + "content": "negative patches (see Fig. 3) from the same layer id in the FP model, from a neighborhood of the anchor's spatial position. We then leverage a contrastive learning objective (see Sec. 3.2) that operates to maximize the similarity between the anchor and positive patches while minimizing similarity with the negative patches. Selection of the anchor patch from the quantized model for guiding data generations helps the model to recognize the semantic context within the generated data, during the subsequent quantization stage (informativeness)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.62, + 0.788, + 0.697 + ], + "angle": 0, + "content": "Patch Neighborhood. For every \\( k^{th} \\) anchor patch \\( \\mathcal{P}_{ijk} \\) corresponding to the \\( j^{\\text{th}} \\) head in \\( i^{\\text{th}} \\) MHSA layer of the quantized model, we identify a neighborhood of size \\( \\mathcal{N}_{ijk} \\) with the same spatial location of \\( \\mathcal{P}_{ijk} \\) as the center, but located in the MHSA layer output of the FP model. Within \\( \\mathcal{N}_{ijk} \\), to identify the most semantically correlated patches \\( (\\mathcal{P}_{\\mathcal{N}} \\in \\mathcal{N}_{ijk}) \\) we use cosine similarity as follows," + }, + { + "type": "equation", + "bbox": [ + 0.398, + 0.706, + 0.787, + 0.743 + ], + "angle": 0, + "content": "\\[\n\\rho (i, j, k) = \\frac {\\mathcal {P} _ {i j k} ^ {T} \\cdot \\mathcal {P} _ {\\mathcal {N}}}{\\left| \\left| \\mathcal {P} _ {i j k} \\right| \\right| _ {2} \\cdot \\left| \\left| \\mathcal {P} _ {\\mathcal {N}} \\right| \\right| _ {2}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.748, + 0.789, + 0.843 + ], + "angle": 0, + "content": "The cosine similarity \\(\\rho\\) is estimated \\(\\forall P_{\\mathcal{N}} \\in \\mathcal{N}_{ijk}\\). We then select the positive patches to be the top-\\(n\\) patches that have the highest \\(\\rho(i,j,k)\\) with the anchor patch and the rest of the patches in the neighborhood correspond to negative patches. We empirically set \\(n = 4\\) in our experiments for a neighborhood of size \\(3 \\times 3\\). We then compute \\(\\mathcal{L}_{i,j}^{C_1}\\) for all anchor patches for each attention head output over all layers to get the contrastive loss \\((\\mathcal{L}^{C_1})\\). We then compute the sample" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.641, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLAMP-ViT" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.178 + ], + "angle": 0, + "content": "generation loss \\(\\mathcal{L}^{SG}\\) by combining \\(\\mathcal{L}^{C_1}\\) and the mean absolute error (MAE) (see Sec. 3.5) output loss \\((\\mathcal{L}^O)\\) and minimize it for data generation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.2, + 0.442, + 0.215 + ], + "angle": 0, + "content": "3.4 Stage 2: Quantization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.225, + 0.784, + 0.269 + ], + "angle": 0, + "content": "We now present our layer-wise evolutionary search with a local contrastive loss-based fitness function to rank suitable quantization parameters from a large search space. We detail the fitness function and the search steps as follows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.27, + 0.788, + 0.543 + ], + "angle": 0, + "content": "Fitness Function. To evaluate the quantization parameters explored by the evolutionary search algorithm, we introduce a fitness function \\(\\mathcal{L}^F\\) that combines the contrastive loss \\(\\mathcal{L}^{C_2}\\) (Eq. (4)) and the MAE loss \\((\\mathcal{L}^O)\\) computed with respect to the targets at the output (explained in Sec. 3.5) i.e., \\(\\mathcal{L}^F = \\mathcal{L}^{C_2} + \\mathcal{L}^O\\). Unlike Stage 1, we use the intermediate activations after each transformer layer to compute the contrastive loss \\(\\mathcal{L}^{C_2}\\). For a ViT, let the set of intermediate representations of FP and quantized model be denoted as, \\(\\mathcal{O}^{fp} = \\{\\mathcal{O}_0^{fp}, \\mathcal{O}_1^{fp}, \\dots, \\mathcal{O}_{L-1}^{fp}\\}\\) and \\(\\mathcal{O}^q = \\{\\mathcal{O}_0^q, \\mathcal{O}_1^q, \\dots, \\mathcal{O}_{L-1}^q\\}\\), respectively. However, directly using high dimensional \\(\\mathcal{O}^{fp}\\) and \\(\\mathcal{O}^q\\) can result in high compute overhead. Therefore for each layer \\(i\\), at the intermediate output, we perform mean pooling along the patch dimension to obtain a low-dimensional representation (\\(\\mathcal{O}_i \\in \\mathcal{R}^N\\)), reducing it by a factor of \\(h \\times d\\). We then apply the contrastive loss (Eq. (4)) sampled within the batch dimension on the low-dimensional \\(\\mathcal{O}_i^{fp}\\) and \\(\\mathcal{O}_i^q\\). For the layer output activation generated from each constituent of a batch of synthetic data, the anchor (\\(\\lambda^p\\)) corresponds to the intermediate layer output of the quantized model, positive (\\(\\lambda^{p+}\\)) and negative (\\(\\lambda^{p-}\\)) samples correspond to set of intermediate layer outputs of FP model that have the same and different targets respectively, relative to the anchor within the batch." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.545, + 0.788, + 0.71 + ], + "angle": 0, + "content": "Step 1: Candidate Initialization. A candidate quantization solution is encoded as a set \\(\\Lambda\\) of \\(L\\) tuples, such that for layer \\(i\\), tuple \\(\\Lambda[i]\\) represents the two quantization parameters \\(\\langle b, \\gamma \\rangle\\). \\(b\\) can take any integer value between 2 and 8, while \\(\\gamma\\) is constrained to a uniform ball of radius \\(10^{-3}\\) centered around, \\(\\gamma[i]^{init4}\\). The candidate scale factors are sampled as \\(\\gamma[i] = \\gamma[i]^{init} + f(-10^{-3}, +10^{+3})\\), where \\(f\\) is a random sampling function. We begin the evolutionary search by creating a population via randomly sampling \\(\\mathcal{K}\\) candidate \\(\\Lambda\\)s, each consisting of layer-wise quantization parameters. We then evaluate the fitness function \\(\\mathcal{L}^F\\) for each candidate \\(\\Lambda\\). We create \\(\\mathcal{K}\\) tuples each with \\((\\Lambda, \\mathcal{L}^F)\\) to form the initial population. \\(\\mathcal{L}^F\\) of each initial candidate with corresponding set \\(\\Lambda\\) is pre-computed and stored to avoid recomputation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.711, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Step 2: Re-generation (Crossover and Mutation). Each candidate in the population is ranked based on corresponding \\(\\mathcal{L}^F\\)s (lower the better) of which the top two serve as the parents for the next candidate generation (child). When evolving candidates, perturbing too many layer parameters based on parents can lead to search instability. To mitigate this, at each evolution step we employ a layer-wise regeneration approach, evolving a single transformer layer at a time based on chosen parents, keeping all other layer parameters to that of the top-1" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.633, + 0.841 + ], + "angle": 0, + "content": "\\(^4\\) \\((\\max (\\theta_i) - \\min (\\theta_i)) / (2^b -1)\\), where \\(\\theta\\) is the weight tensor." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.674, + 0.129 + ], + "angle": 0, + "content": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.178 + ], + "angle": 0, + "content": "parent (ranked via \\(\\mathcal{L}^F\\)). The child's parameter regeneration for a layer based on that of the chosen parents (\\(p1, p2\\)) is formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.186, + 0.786, + 0.202 + ], + "angle": 0, + "content": "\\[\nb _ {\\text {c h i l d}} = \\mathbf {r} \\text {a n d o m} (\\mathbf {m i n} (b _ {p 1}, b _ {p 2}) - 1, \\mathbf {m a x} (b _ {p 1}, b _ {p 2}) + 1) \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.205, + 0.786, + 0.222 + ], + "angle": 0, + "content": "\\[\n\\gamma_ {c h i l d} = \\text {m e a n} \\left(\\gamma_ {p 1}, \\gamma_ {p 2}\\right) + \\eta \\left(- 1 0 ^ {- 3}, 1 0 ^ {3}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.228, + 0.784, + 0.289 + ], + "angle": 0, + "content": "Step 3: Diversity Promoting Selection. To avoid overfitting during search we follow [11] and introduce diversity into the population. In specific, we create \\( P = 5 \\) random parents and use each of them to crossover with the child generated in Step 2 and create \\( P' \\) diverse children by following Eq. (6), Eq. (7)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.289, + 0.785, + 0.409 + ], + "angle": 0, + "content": "Step 4: Evaluation and Population Update. We evaluate all generated children in the steps above and acquire \\(\\mathcal{L}^F\\)s. The child generated in Step 2 and the corresponding fitness function value is added to the population. We then rank the diversity promoting children from Step 3 and select the best child to be added to the population for the next iteration. In our layer-wise evolutionary search strategy, we employ \\(\\mathcal{P}\\) passes over all layers of a ViT, and each layer is iterated over \\(\\mathcal{C}\\) cycles in each pass. So, the population is updated \\(\\mathcal{P} \\times \\mathcal{C} \\times L\\) times, i.e., the Step 2, 3, and 4 are iteratively executed \\(\\mathcal{P} \\times \\mathcal{C} \\times L\\) times." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.784, + 0.471 + ], + "angle": 0, + "content": "Activation Quantization. We note that the sensitivity to quantization for activations is closely correlated to the sensitivity of weight parameters. Therefore for layer \\(i\\), we determine the output activation quantization parameters as follows, \\(b_{act}[i] = \\min (8,b[i]\\times 2)\\) and \\(\\gamma_{act}[i] = \\gamma_{act}[i - 1] + \\gamma [i]\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.488, + 0.396, + 0.504 + ], + "angle": 0, + "content": "3.5 Overall Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.785, + 0.631 + ], + "angle": 0, + "content": "In Fig. 2, we illustrate the complete CLAMP-ViT framework. To ensure adaptive data-generation and informativeness for quantization, we use a cyclic strategy between the two stages, updating generated data based on the quantized model's needs for optimal parameter search. The framework requires an input batch of \\(\\mathcal{B}\\) random Gaussian images \\(X_{\\mathcal{B}}\\), and corresponding task-specific targets \\(T_{G_{\\mathcal{B}}}\\) (\\(T_{G_{\\mathcal{B}}}\\) for each task is detailed in Sec. 4). The targets direct the synthetic image generation towards task-specific goals as well as penalize inaccurate predictions of quantized model through the output loss \\(\\mathcal{L}^O\\)," + }, + { + "type": "equation", + "bbox": [ + 0.324, + 0.642, + 0.786, + 0.671 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {O} = \\frac {1}{n _ {c}} \\left(\\left| \\left| \\mathcal {Q} \\left(X _ {\\mathcal {B}}\\right) - T _ {G _ {\\mathcal {B}}} \\right| \\right| _ {1} + \\left| \\left| \\mathcal {F P} \\left(X _ {\\mathcal {B}}\\right) - T _ {G _ {\\mathcal {B}}} \\right| \\right| _ {1}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.786, + 0.841 + ], + "angle": 0, + "content": "where \\( n_c \\) is the number of output classes (classification) or prediction map size (segmentation/detection). The quantized model is initialized to the best candidate from \\( \\mathcal{K} \\) tuples. The framework assumes a range of bit-widths and a single bit-width for mixed- and fixed-precision search, respectively. In Stage 1, \\( X_{\\mathcal{B}} \\) is fed to the frozen quantized \\( (\\mathcal{Q}) \\) and full-precision model \\( (\\mathcal{FP}) \\) to minimize the sample generation loss \\( \\mathcal{L}^{SG} = \\mathcal{L}^{C_1} + \\mathcal{L}^O \\), updating \\( X_{\\mathcal{B}} \\) via backpropagation for \\( G \\) iterations. In Stage 2, we use the generated data to quantize \\( \\mathcal{Q} \\) for \\( \\mathcal{P} \\times \\mathcal{C} \\times L \\) iterations by minimizing \\( \\mathcal{L}^F \\). We cyclically update the generated data every \\( \\mathcal{C}/2 \\) iterations. In every subsequent execution of Stage 1, we do not restart from Gaussian noise but use \\( X_{\\mathcal{B}} \\) from the previous Stage 1 execution and update it for \\( G/2 \\) iterations. In this manner, the two stages are executed alternately." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.641, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLAMP-ViT" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.2 + ], + "angle": 0, + "content": "Table 1: Fixed-precision quantization accuracy comparison with SoTA on image classification tasks with ImageNet-1k testset. 'R', 'S' signifies real and synthetic calibration data and W/A indicates weight/activation bit-width. The values in bold and underline signifies, the best performance overall, and with synthetic data, respectively." + }, + { + "type": "table", + "bbox": [ + 0.299, + 0.201, + 0.697, + 0.555 + ], + "angle": 0, + "content": "
ModelMethodData#ImagesW/ATop-1W/ATop-1
ViT-BBaseline--32/3284.5332/3284.53
PSAQ-ViT v1S328/837.364/825.34
PTQ4ViTR328/884.254/867.16
FQ-ViTR10008/883.314/878.73
RepQ-ViTR328/881.454/876.29
CLAMP-ViT (Ours)S328/884.194/878.73
DeiT-TBaseline--32/3272.2132/3272.21
PSAQ-ViT v1S328/871.564/865.57
PSAQ-ViT v2S328/872.174/868.61
FQ-ViTR10008/871.614/866.91
RepQ-ViTR328/872.054/868.75
CLAMP-ViT (Ours)S328/872.174/869.93
DeiT-SBaseline--32/3279.8532/3279.85
PSAQ-ViT v1S328/876.924/873.23
PSAQ-ViT v2S328/879.564/876.36
PTQ4ViTR328/879.474/8-
FQ-ViTR10008/879.174/876.93
RepQ-ViTR328/879.554/876.75
CLAMP-ViT (Ours)S328/879.554/877.03
Swin-TBaseline--32/3281.3532/3281.35
PSAQ-ViT v1 [27]S328/875.354/871.79
PSAQ-ViT v2 [25]S328/880.214/876.28
FQ-ViT [30]R10008/881.294/880.73
RepQ-ViT [28]R328/881.284/880.51
CLAMP-ViT (Ours)S328/881.174/880.28
Swin-SBaseline--32/3283.2032/3283.20
PSAQ-ViT v1 [27]S328/876.644/875.14
PSAQ-ViT v2 [25]S328/882.134/878.86
FQ-ViT [30]R10008/882.134/881.67
RepQ-ViT [28]R328/882.344/882.14
CLAMP-ViT (Ours)S328/882.574/882.51
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.569, + 0.461, + 0.586 + ], + "angle": 0, + "content": "4 Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.602, + 0.427, + 0.618 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.784, + 0.673 + ], + "angle": 0, + "content": "Models and Datasets. We evaluate CLAMP-ViT on various ViT model families (pre-trained FP models sourced from timm [40]) for image classification, object detection and semantic segmentation detailed as follows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.785, + 0.703 + ], + "angle": 0, + "content": "Image Classification. We use ImageNet-1K [10] having 50K testset, with DeiT-B/T/S [38], Swin-T/S [31], and ViT-B/S [12] to evaluate accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.785, + 0.749 + ], + "angle": 0, + "content": "Object detection. We use the COCO 2017 dataset [29] having approximately 20K test data. Following [25,28,30], we use the Cascade Mask R-CNN [3] framework from MMdetection library [6] with DeiT-S and Swin-S as the backbone." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Semantic Segmentation. We use the ADE20K dataset [52] with 3K test data encompassing 150 categories with DeiT-S and Swin-S as the backbone. We adopt the UperNet framework [42] in the MMsegmentation library [9] similar to [25]. Baselines. CLAMP-ViT is evaluated against SoTA PTQ (real data) and DFQ (synthetic data) methods for quantizing models from FP in various vision tasks. For image classification, it's compared with fixed-precision methods like PSAQ-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.672, + 0.129 + ], + "angle": 0, + "content": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 2: Mixed-Precision Quantization accuracy comparison on image classification tasks with ImageNet-1k testset. The values in bold indicate best performance overall." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.174, + 0.768, + 0.226 + ], + "angle": 0, + "content": "
DeiT-TDeiT-SSwin-TSwin-S
MethodDataW/ATop-1W/ATop-1W/ATop-1W/ATop-1
LRP-QViT [36]RMP6/MP671.03MP6/MP679.03--MP6/MP682.86
VT-PTQ [32]RMP6/MP669.46MP6/MP675.10MP6/MP679.61MP6/MP678.43
CLAMP-ViT (Ours)SMP4.9/MP6.271.69MP4.7/MP5.979.43MP5.5/MP6.981.78MP5.1/MP6.382.86
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.784, + 0.3 + ], + "angle": 0, + "content": "ViT v1 [27], v2 [25], FQ-ViT [30], RepQ-ViT [28], and PTQ4ViT [45], and mixed-precision methods such as LRP-QViT [36] and VT-PTQ [32]. In object detection (Mask R-CNN), we use fixed-precision baselines FQ-ViT [30], PSAQ-ViT v2 [25], RepQ-ViT [28], and LRP-QViT [36] for mixed-precision comparison." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.3, + 0.51, + 0.389 + ], + "angle": 0, + "content": "For semantic segmentation (UperNet), PSAQ-ViT v2 [25] serves as the baseline. Evol-Q is excluded from the main comparison as it does not fully quantize a model starting from FP, and is presented for ablations in the supplementary." + }, + { + "type": "table_caption", + "bbox": [ + 0.547, + 0.301, + 0.76, + 0.314 + ], + "angle": 0, + "content": "Table 3: Hyperparameters list." + }, + { + "type": "table", + "bbox": [ + 0.553, + 0.315, + 0.753, + 0.388 + ], + "angle": 0, + "content": "
ParameterDescriptionValue
GGeneration Iterations500
NNeighborhood Size3 × 3
n# Positive Patches4
PPasses10
CCycles6
K# Initial Candidates15
BBatch Size32
bBit-width[2,8]
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.784, + 0.435 + ], + "angle": 0, + "content": "Experimental Setup. The CLAMP-ViT framework is implemented in PyTorch, and evaluated on a single NVIDIA Titan GPU. It features multiple hyperparameters, detailed in Tab. 3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.455, + 0.635, + 0.47 + ], + "angle": 0, + "content": "4.2 Quantization Results for Image Classification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.51, + 0.553 + ], + "angle": 0, + "content": "As highlighted in Sec. 3.5, CLAMP-ViT requires a batch \\(\\mathcal{B}\\) of task-specific targets \\(T_{G_B}\\). For image classification on the ImageNet-1K, we create \\(T_{G_B} \\in \\mathcal{R}^{\\mathcal{B} \\times 1000}\\), where the class-wise probabilities are" + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.479, + 0.785, + 0.507 + ], + "angle": 0, + "content": "Table 4: Comparison of quantized DeiT-S size (MB) and BOPS (G) [1]." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.508, + 0.779, + 0.544 + ], + "angle": 0, + "content": "
MethodW/ABOPSSize
Baseline32/32471088
PSAQ-ViT v24/829422
CLAMP-ViT (Ours)MP4.7/MP5.926720
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.81 + ], + "angle": 0, + "content": "randomly determined and assigned. We discuss and compare the performance of CLAMP-ViT for two settings, fixed-precision (Tab. 1) and mixed-precision (Tab. 2). In specific, as shown in Tab. 1 CLAMP-ViT consistently provides similar or better accuracy at W8/A8, while for lower precision W4/A8 CLAMP-ViT shows significant performance boost over all the existing alternatives. We yield \\(\\sim 2.2\\%\\) and \\(\\sim 1\\%\\) average accuracy improvement compared to DFQ methods [25, 27] and data-driven methods, respectively. The superior performance of CLAMP-ViT can be attributed to the cyclically adaptive data-generation process, which ensures the generated data matches the requirements and representational capabilities of the quantized model and effective traversal of the search space through evolutionary search and contrastive learning. Whereas, PSAQ-ViT v2 [25], generates increasingly difficult samples which is less beneficial for aggressive 4-bit quantization. Surprisingly, PSAQ-ViT v1 [27] achieves poor accuracy of \\(25.34\\%\\) (W4/A8) on ViT-B despite achieving reasonable accuracy on other ViTs. This result potentially supports our initial intuition that PSAQ-ViT [25, 27] does not consider the informativeness of the generated data to the quantization process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Evident from Tab. 2, CLAMP-ViT consistently outperforms all baselines across models for the mixed-precision setting, maintaining accuracy close to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.641, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLAMP-ViT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 5: Mixed and fixed-precision quantization performance comparison against SoTA techniques for object detection on COCO 2017. The values in bold and underline signifies, the best performance overall, and with synthetic data, respectively." + }, + { + "type": "table", + "bbox": [ + 0.3, + 0.187, + 0.697, + 0.309 + ], + "angle": 0, + "content": "
DeiT-SSwin-S
MethodDataW/A\\( AP^{box} \\)\\( AP^{mask} \\)W/A\\( AP^{box} \\)\\( AP^{mask} \\)
Baseline-32/3248.041.432/3251.844.7
FQ-ViT [30]R8/847.440.98/850.844.1
PSAQ-ViT v2 [25]S8/847.340.88/850.944.1
RepQ-ViT [28]R8/847.941.18/851.644.6
CLAMP-ViT (Ours)S8/847.941.18/851.444.6
FQ-ViT [30]R4/845.140.24/848.241.3
PSAQ-ViT v2 [25]S4/844.838.84/847.941.4
RepQ-ViT [28]R4/845.639.54/849.242.8
CLAMP-ViT (Ours)S4/845.438.94/848.542.2
LRP-QViT [36]R---\\( MP_{6}/MP_{6} \\)51.444.6
CLAMP-ViT (Ours)S\\( MP_{5.5}/MP_{6.8} \\)\\( 47.9 \\)\\( 41.0 \\)\\( MP_{5.1}/MP_{6.4} \\)51.744.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.31, + 0.786, + 0.339 + ], + "angle": 0, + "content": "Table 6: Mixed-precision quantization performance comparison against PSAQ-ViT v2 for semantic segmentation. The values in bold indicate best performance overall." + }, + { + "type": "table", + "bbox": [ + 0.328, + 0.343, + 0.669, + 0.398 + ], + "angle": 0, + "content": "
DeiT-SSwin-S
MethodDataW/AmIoUW/AmIoU
Baseline-32/3244.032/3249.3
PSAQ-ViT v2 [25]S4/839.94/844.6
CLAMP-ViT (Ours)SMP4.8/MP6.242.4MP5.1/MP6.445.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.788, + 0.519 + ], + "angle": 0, + "content": "the FP baseline despite having a significantly lower average W/A (for mixed-precision W/A is calculated by averaging bit-widths over all the layers for weights and activations). Furthermore, in Tab. 4, we report the reduction in model size and bit-operations-per-second (BOPS) [1] of mixed-precision quantized DeiT-S compared with W4/A8 PSAQ-ViT v2. CLAMP-ViT achieves \\(\\sim 10\\%\\) lower model size in MB and BOPS when compared to W4/A8 PSAQ-ViT v2 while yielding \\(3.07\\%\\) improved accuracy." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.535, + 0.613, + 0.551 + ], + "angle": 0, + "content": "4.3 Quantization Results for Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.557, + 0.788, + 0.753 + ], + "angle": 0, + "content": "The target for object detection is \\( T_{G_{\\mathcal{B}}} \\in \\mathcal{R}^{\\mathcal{B} \\times bb \\times 5} \\) where \\( bb \\) is the number of bounding boxes in the image that is randomly selected from the integer set [1, 3]. \\( T_{G_{\\mathcal{B}}}[\\mathcal{B}, :, 0] \\) corresponds to the bounding box category and \\( T_{G_{\\mathcal{B}}}[\\mathcal{B}, :, 1:4] \\) is the bounding box coordinates \\( x, y, w, h \\) [16]. Tab. 5 presents the fixed- and mixed-precision performance of CLAMP-ViT with respect to the baselines. Across different settings and models, CLAMP-ViT consistently outperforms DFQ method PSAQ-ViT v2 [25] by 0.6 box AP and 0.4 mask AP on average while closely matching performance to the SoTA data-driven method, RepQ-ViT [28]. Similar to Sec. 4.2, we observe improved performance with mixed-precision quantization, achieving near FP baseline performance. The average W/A for mixed-precision quantization for object detection is found to be higher than that for image classification due to the higher complexity of the task demanding larger bit-widths to maintain accuracy." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.773, + 0.664, + 0.789 + ], + "angle": 0, + "content": "4.4 Quantization Results for Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "The target \\( T_{G_{\\mathcal{B}}} \\) for this task is a pixel-wise classification map of the same size as \\( X_{\\mathcal{B}} \\) i.e., \\( T_{G_{\\mathcal{B}}} \\in \\mathcal{R}^{\\mathcal{B} \\times 150 \\times H \\times W} \\). In Tab. 6, we show the quantization performance comparison, where CLAMP-ViT achieves average weight bit-width close" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.674, + 0.129 + ], + "angle": 0, + "content": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.144, + 0.425, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.221, + 0.355, + 0.227 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.144, + 0.577, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.221, + 0.508, + 0.227 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.144, + 0.73, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.222, + 0.662, + 0.227 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.239, + 0.785, + 0.281 + ], + "angle": 0, + "content": "Fig. 4: Comparison of synthetic data generated by (a) PSAQ-ViT v1 [27], (b) PSAQ-ViT v2 [25] and (c) CLAMP-ViT (Ours). CLAMP-ViT generates detailed objects within contextually suitable backgrounds, boosting realism and informativeness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.295, + 0.788, + 0.34 + ], + "angle": 0, + "content": "to 4-bit while also significantly reducing activation bit-width to \\(\\sim\\) 6-bits. At a higher compression ratio of \\(15\\%\\), CLAMP-ViT achieves \\(1.5\\mathrm{mIoU}\\) improvement on average over PSAQ-ViT v2." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.366, + 0.519, + 0.381 + ], + "angle": 0, + "content": "4.5 Analysis of Generated Samples" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.789, + 0.621 + ], + "angle": 0, + "content": "Fig. 4 visualizes generated samples from PSAQ-ViT v1 [27], v2 [25], and CLAMP-ViT (after \\(1^{st}\\) round of stage 1 execution). PSAQ-ViT v1 (Fig. 4(a)) creates images with clear class-specific foregrounds but with overly simplistic and uniform backgrounds, resulting in a lack of realism potentially affecting model accuracy. PSAQ-ViT v2 (Fig. 4(b)) introduces more complex details but fails to convey meaningful semantic information, generating images with intricate but semantically vague structures due to its unguided, difficulty-increasing data-generation strategy. In contrast, CLAMP-ViT (Fig. 4(c)) excels by synthesizing data that mirrors real-world imagery, showcasing a sophisticated understanding of semantic relationships between patches. It ensures objects are detailed and in contextually fitting backgrounds, boosting realism and informativeness. For example, CLAMP-ViT places boats on water and zebras in grasslands (Fig. 4(c), row 2), showing its capability for creating semantically relevant and visually consistent synthetic data. We believe our patch semantics exploration with a contrastive objective, makes image generation informative that mimic real-world scenes." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.646, + 0.479, + 0.66 + ], + "angle": 0, + "content": "4.6 Ablations and Discussions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Evolutionary Search Parameters. In Fig. 5(a), we detail an experiment to determine the ideal number of passes \\(\\mathcal{P}\\) and cycles \\(\\mathcal{C}\\) for the evolutionary search process by studying the variation in Top-1 accuracy of DeiT-S with different passes and cycles keeping the other fixed at their optimal value (Tab. 3). It is evident that a cycle count of \\(\\mathcal{C} = 6\\) is optimal, as accuracy tends to decline with more cycles. Conversely, passes show a modest yet consistent improvement beyond 10, but due to the substantial rise in computational complexity, \\(\\mathcal{P} = 10\\) is deemed the most suitable choice." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Effect of Batch Size \\(\\mathcal{B}\\). We also show the accuracy comparison with different batch sizes ranging from 8 to 64 in Fig. 5(c). It is apparent that there is minimal increase in accuracy beyond 32 for CLAMP-ViT, justifying the choice of batch" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.641, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLAMP-ViT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.149, + 0.351, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.287, + 0.226, + 0.3, + 0.235 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.149, + 0.498, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.426, + 0.226, + 0.44, + 0.235 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.148, + 0.635, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.226, + 0.588, + 0.235 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.147, + 0.782, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.226, + 0.714, + 0.235 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.244, + 0.787, + 0.287 + ], + "angle": 0, + "content": "Fig. 5: CLAMP-ViT ablations for (a) Selecting evolutionary search parameters, (b) Mixed-precision quantization accuracy with different fitness functions, (c) Effect of batch size \\(\\mathcal{B}\\) and (d) Effect of neighborhood size \\(\\mathcal{N}\\) and top-\\(n\\) positive patches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.297, + 0.785, + 0.328 + ], + "angle": 0, + "content": "size. Interestingly, CLAMP-ViT achieves a decent accuracy of \\(\\sim 70\\%\\) even with a batch size of 8, while PSAQ-ViT v1 and v2 achieve only close to \\(60\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.332, + 0.785, + 0.469 + ], + "angle": 0, + "content": "Effect of Neighborhood Size \\(\\mathcal{N}\\) and Top-\\(n\\) Patches. We explore \\(3 \\times 3\\), \\(5 \\times 5\\), and all patches for neighborhood sizes and \\(n \\in [1,4]\\) for DeiT-S as shown in Fig. 5(d). A \\(3 \\times 3\\) neighborhood with top-4 patches as positives yields the highest accuracy while being computationally least demanding. In the \\(5 \\times 5\\) scenario, as the balance between positive and negative samples improves with increasing \\(n\\), accuracy rises but at higher computational cost. Employing all patches is computationally unviable and leads to the lowest accuracy for small \\(n\\) values due to a positive-negative imbalance [39]. Moreover, identifying distant positive patches from the anchor neglects significant semantic patch relationships." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.511, + 0.578 + ], + "angle": 0, + "content": "Choice of Objective Function. The study in Tab. 7 examines how different loss function components of \\(\\mathcal{L}^{SG}\\) i.e., \\(\\mathcal{L}^{C_1}\\) and \\(\\mathcal{L}^O\\) affect synthetic data generation effectiveness and its impact on top-1 accuracy of W4/A8 quantization of DeiT-S.We chose fixed-precision quantization" + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.474, + 0.785, + 0.502 + ], + "angle": 0, + "content": "Table 7: Impact of different loss components in synthetic data generation." + }, + { + "type": "table", + "bbox": [ + 0.55, + 0.503, + 0.756, + 0.571 + ], + "angle": 0, + "content": "
LPSELC1LOW/ATop-1Accuracy
---32/3279.85
✓✗✗4/874.18
✗✗✓4/833.93
✓✗✓4/876.47
✗✓✗4/876.59
✗✓✓4/877.03
✓✓✓4/876.98
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.788, + 0.744 + ], + "angle": 0, + "content": "to avoid any bias from mixed-precision quantization, which might typically favor higher bitwidths to lessen accuracy loss due to low-quality images. Results show that a linear mix of \\(\\mathcal{L}^{C_1}\\) and \\(\\mathcal{L}^O\\) (\\(\\mathcal{L}^{SG}\\)) achieves highest accuracy, while using \\(\\mathcal{L}^O\\) alone leads to the lowest, indicating its limited utility in leveraging ViTs for synthetic data. In Tab. 7 \\(\\mathcal{L}^{PSE}\\) corresponds to the patch similarity metric employed in [25]. While combining \\(\\mathcal{L}^{PSE}\\) with \\(\\mathcal{L}^O\\) [25] does offer a moderate accuracy boost, it falls short of the performance with \\(\\mathcal{L}^{SG}\\), due to inherent limitations of the patch similarity metric highlighted in Sec. 1. Furthermore, using all three loss functions simultaneously closely matches the performance of \\(\\mathcal{L}^{SG}\\) further demonstrating that our proposed \\(\\mathcal{L}^{C_1}\\) has the major contribution towards final accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "For mixed-precision quantization of DeiT-S, the fitness function's accuracy \\((\\mathcal{L}^F = \\mathcal{L}^{C_2} + \\mathcal{L}^O)\\) is compared against global contrastive loss [14], MSE, and KL-divergence in Fig. 5(b). The accuracy analysis shows MSE and KL-divergence tend to overfit to synthetic data, evidenced by plateauing accuracy. Meanwhile, global contrastive loss initially matches but then accuracy gap widens from CLAMP-ViT's performance which is due to premature convergence, implying" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.672, + 0.128 + ], + "angle": 0, + "content": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "that quantifying intermediate layer distributional divergence is crucial to find optimal quantization parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.511, + 0.237 + ], + "angle": 0, + "content": "Effect of Adaptivity. We investigate the effectiveness of the cyclic evolution every \\(\\mathcal{C} / 2\\) iterations to ensure the data generation adapts to the requirements of" + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.179, + 0.786, + 0.206 + ], + "angle": 0, + "content": "Table 8: Effect of adaptability on quantized model accuracy." + }, + { + "type": "table", + "bbox": [ + 0.544, + 0.207, + 0.763, + 0.235 + ], + "angle": 0, + "content": "
AdaptivityW/ATop-1 Accuracy
XMP4.9/MP6.178.06
✓MP4.7/MP5.978.97
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.784, + 0.268 + ], + "angle": 0, + "content": "the quantized model in Tab. 8. It can be observed that with adaptivity, we are not only able to achieve lower average W/A but also have improved accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.543, + 0.403 + ], + "angle": 0, + "content": "Effect of Informativeness. To show CLAMP-ViT's effectiveness in generating informative data, we compare misprediction rates in PSAQ-ViT v1, v2, and CLAMP-ViT during quantization. PSAQ-ViT has higher misprediction rates (34% for v1, 41% for v2) than CLAMP-ViT's 22%, indicating PSAQ-ViT's data is less informative, leading to erratic predictions, sub-optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.543, + 0.434 + ], + "angle": 0, + "content": "Limitation Discussion: Runtime Comparison. In Fig. 6, we show the runtime and" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.543, + 0.509 + ], + "angle": 0, + "content": "top-1 accuracy of different techniques on an CLAMP-ViT (mixed-precision) achieves signi to the other methods (fixed-precision) with o \\((5\\% \\uparrow)\\) compared to the best DFQ method. N calibrate a pre-quantized model on real data." + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.275, + 0.75, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.374, + 0.787, + 0.43 + ], + "angle": 0, + "content": "Fig.6: Runtime of different PTQ and DFQ methods where, FS: FP model as starting point. QS: Prequantized model as starting point." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.531, + 0.368, + 0.547 + ], + "angle": 0, + "content": "5 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.561, + 0.788, + 0.744 + ], + "angle": 0, + "content": "This paper presents CLAMP-ViT, a novel mixed-precision DFQ technique using cyclic adaptation and contrastive learning. It employs patch-level contrastive learning that leverages properties of the MHSA modules for data generation. A local contrastive objective and layer-wise evolutionary search identify optimal quantization parameters while ensuring a smooth loss landscape. Experiments across CV tasks show superior performance of CLAMP-ViT, achieving up to \\(3\\%\\) top-1 accuracy for classification, 0.6 mAP for detection, and 1.5 mIoU for segmentation. Future work aims to focus on extending its application to a wider range of architectures, like VLMs. While this study focuses on a useful impact and beneficial application of synthetic data generation for optimized and carbon efficient models for deployment, it is important to also be cognizant of the potential adverse effects of synthetic data such as deepfakes or racial biases." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.765, + 0.404, + 0.782 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "This work was supported in part by CoCoSys, one of the seven centers in JUMP 2.0, a Semiconductor Research Corporation (SRC) program sponsored by DARPA." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.641, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLAMP-ViT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.786, + 0.236 + ], + "angle": 0, + "content": "1. Baskin, C., Liss, N., Schwartz, E., Zheltonozhskii, E., Giryes, R., Bronstein, A.M., Mendelson, A.: Uniq: Uniform noise injection for non-uniform quantization of neural networks. ACM Transactions on Computer Systems (TOCS) 37(1-4), 1-15 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.237, + 0.786, + 0.278 + ], + "angle": 0, + "content": "2. Cai, Y., Yao, Z., Dong, Z., Gholami, A., Mahoney, M.W., Keutzer, K.: Zeroq: A novel zero shot quantization framework. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13169-13178 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.279, + 0.786, + 0.32 + ], + "angle": 0, + "content": "3. Cai, Z., Vasconcelos, N.: Cascade r-cnn: Delving into high quality object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6154-6162 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.321, + 0.786, + 0.362 + ], + "angle": 0, + "content": "4. Cao, Y.H., Sun, P., Huang, Y., Wu, J., Zhou, S.: Synergistic self-supervised and quantization learning. In: European Conference on Computer Vision. pp. 587-604. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.363, + 0.786, + 0.404 + ], + "angle": 0, + "content": "5. Chen, H., Shao, S., Wang, Z., Shang, Z., Chen, J., Ji, X., Wu, X.: Bootstrap generalization ability from loss landscape perspective. In: European Conference on Computer Vision. pp. 500-517. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.786, + 0.447 + ], + "angle": 0, + "content": "6. Chen, K., Wang, J., Pang, J., Cao, Y., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Xu, J., et al.: Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.448, + 0.786, + 0.489 + ], + "angle": 0, + "content": "7. Choi, K., Hong, D., Park, N., Kim, Y., Lee, J.: Qimera: Data-free quantization with synthetic boundary supporting samples. Advances in Neural Information Processing Systems 34, 14835-14847 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.49, + 0.786, + 0.545 + ], + "angle": 0, + "content": "8. Chuang, C.Y., Hjelm, R.D., Wang, X., Vineet, V., Joshi, N., Torralba, A., Jegelka, S., Song, Y.: Robust contrastive learning against noisy views. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16670-16681 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.546, + 0.786, + 0.573 + ], + "angle": 0, + "content": "9. Contributors, M.: MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.574, + 0.786, + 0.615 + ], + "angle": 0, + "content": "0. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.616, + 0.786, + 0.658 + ], + "angle": 0, + "content": "1. Dong, P., Li, L., Wei, Z., Niu, X., Tian, Z., Pan, H.: Emq: Evolving training-free proxies for automated mixed precision quantization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17076-17086 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.786, + 0.713 + ], + "angle": 0, + "content": "2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.714, + 0.786, + 0.756 + ], + "angle": 0, + "content": "3. Fei, W., Dai, W., Li, C., Zou, J., Xiong, H.: General bitwidth assignment for efficient deep convolutional neural network quantization. IEEE Transactions on Neural Networks and Learning Systems 33(10), 5253-5267 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.786, + 0.798 + ], + "angle": 0, + "content": "4. Frumkin, N., Gope, D., Marculescu, D.: Jumping through local minima: Quantization in the loss landscape of vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16978-16988 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "5. Fu, Y., Yu, Q., Li, M., Ouyang, X., Chandra, V., Lin, Y.: Contrastive quant: quantization makes stronger contrastive learning. In: Proceedings of the 59th ACM/IEEE Design Automation Conference. pp. 205-210 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.672, + 0.128 + ], + "angle": 0, + "content": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Huang, H., Yu, P.S., Wang, C.: An introduction to image synthesis with generative adversarial nets. arXiv preprint arXiv:1803.04469 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "17. Hubara, I., Courbariaux, M., Soudry, D., El-Yaniv, R., Bengio, Y.: Binarized neural networks. Advances in neural information processing systems 29 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.245 + ], + "angle": 0, + "content": "18. Hubara, I., Nahshan, Y., Hanani, Y., Banner, R., Soudry, D.: Accurate post training quantization with small calibration sets. In: International Conference on Machine Learning. pp. 4466-4475. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "19. Kim, N., Shin, D., Choi, W., Kim, G., Park, J.: Exploiting retraining-based mixed-precision quantization for low-cost dnn accelerator design. IEEE Transactions on Neural Networks and Learning Systems 32(7), 2925–2938 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "20. Kundu, S., Sun, Q., Fu, Y., Pedram, M., Beerel, P.: Analyzing the confidentiality of undistillable teachers in knowledge distillation. Advances in Neural Information Processing Systems 34, 9181-9192 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.383 + ], + "angle": 0, + "content": "21. Kundu, S., Wang, S., Sun, Q., Beerel, P.A., Pedram, M.: Bmpq: bit-gradient sensitivity-driven mixed-precision quantization of dnns from scratch. In: 2022 Design, Automation & Test in Europe Conference & Exhibition (DATE). pp. 588-591. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "22. Li, H., Xu, Z., Taylor, G., Studer, C., Goldstein, T.: Visualizing the loss landscape of neural nets. Advances in neural information processing systems 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "23. Li, Y., Xu, S., Zhang, B., Cao, X., Gao, P., Guo, G.: Q-vit: Accurate and fully quantized low-bit vision transformer. Advances in Neural Information Processing Systems 35, 34451-34463 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.493 + ], + "angle": 0, + "content": "24. Li, Y., Yuan, G., Wen, Y., Hu, J., Evangelidis, G., Tulyakov, S., Wang, Y., Ren, J.: Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems 35, 12934-12949 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.535 + ], + "angle": 0, + "content": "25. Li, Z., Chen, M., Xiao, J., Gu, Q.: Psaq-vit v2: Toward accurate and general data-free quantization for vision transformers. IEEE Transactions on Neural Networks and Learning Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "26. Li, Z., Gu, Q.: I-vit: integer-only quantization for efficient vision transformer inference. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17065-17075 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.618 + ], + "angle": 0, + "content": "27. Li, Z., Ma, L., Chen, M., Xiao, J., Gu, Q.: Patch similarity aware data-free quantization for vision transformers. In: European Conference on Computer Vision. pp. 154-170. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.66 + ], + "angle": 0, + "content": "28. Li, Z., Xiao, J., Yang, L., Gu, Q.: Repq-vit: Scale reparameterization for posttraining quantization of vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17227-17236 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.715 + ], + "angle": 0, + "content": "29. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.743 + ], + "angle": 0, + "content": "30. Lin, Y., Zhang, T., Sun, P., Li, Z., Zhou, S.: Fq-vit: Post-training quantization for fully quantized vision transformer. arXiv preprint arXiv:2111.13824 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.799 + ], + "angle": 0, + "content": "31. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "32. Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., Gao, W.: Post-training quantization for vision transformer. Advances in Neural Information Processing Systems 34, 28092-28103 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.641, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLAMP-ViT" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.176 + ], + "angle": 0, + "content": "33. Peters, J.W., Welling, M.: Probabilistic binary neural networks. arXiv preprint arXiv:1809.03368 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.788, + 0.232 + ], + "angle": 0, + "content": "34. Ramachandran, A., Dhiman, A., Vandrotti, B.S., Kim, J.: Ntrans-net: A multiscale neutrosophic-uncertainty guided transformer network for indoor depth completion. In: 2023 IEEE International Conference on Image Processing (ICIP). pp. 905–909. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.787, + 0.275 + ], + "angle": 0, + "content": "35. Ramachandran, A., Wan, Z., Jeong, G., Gustafson, J., Krishna, T.: Algorithm-hardware co-design of distribution-aware logarithmic-posit encodings for efficient dnn inference. arXiv preprint arXiv:2403.05465 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.787, + 0.303 + ], + "angle": 0, + "content": "36. Ranjan, N., Savakis, A.: Lrp-qvit: Mixed-precision vision transformer quantization via layer-wise relevance propagation. arXiv preprint arXiv:2401.11243 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.787, + 0.346 + ], + "angle": 0, + "content": "37. Strudel, R., Garcia, R., Laptev, I., Schmid, C.: Segmenter: Transformer for semantic segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7262-7272 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.787, + 0.389 + ], + "angle": 0, + "content": "38. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: International conference on machine learning. pp. 10347-10357. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.39, + 0.787, + 0.416 + ], + "angle": 0, + "content": "39. Wang, J., Li, J., Li, W., Xuan, L., Zhang, T., Wang, W.: Positive-negative equal contrastive loss for semantic segmentation. Neurocomputing 535, 13-24 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.418, + 0.787, + 0.445 + ], + "angle": 0, + "content": "40. Wightman, R.: Pytorch image models. https://github.com/rwrightman/pytorch-image-models (2019). https://doi.org/10.5281/zenodo.4414861" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.787, + 0.473 + ], + "angle": 0, + "content": "41. Xiao, J., Li, Z., Yang, L., Gu, Q.: Patch-wise mixed-precision quantization of vision transformer. arXiv preprint arXiv:2305.06559 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.475, + 0.787, + 0.516 + ], + "angle": 0, + "content": "42. Xiao, T., Liu, Y., Zhou, B., Jiang, Y., Sun, J.: Unified perceptual parsing for scene understanding. In: Proceedings of the European conference on computer vision (ECCV). pp. 418-434 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.517, + 0.787, + 0.558 + ], + "angle": 0, + "content": "43. Yeh, C.H., Hong, C.Y., Hsu, Y.C., Liu, T.L., Chen, Y., LeCun, Y.: Decoupled contrastive learning. In: European Conference on Computer Vision. pp. 668-684. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.787, + 0.614 + ], + "angle": 0, + "content": "44. Yin, H., Molchanov, P., Alvarez, J.M., Li, Z., Mallya, A., Hoiem, D., Jha, N.K., Kautz, J.: Dreaming to distill: Data-free knowledge transfer via deepinversion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8715-8724 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.787, + 0.657 + ], + "angle": 0, + "content": "45. Yuan, Z., Xue, C., Chen, Y., Wu, Q., Sun, G.: Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization. In: European Conference on Computer Vision. pp. 191-207. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.787, + 0.699 + ], + "angle": 0, + "content": "46. Zhang, D., Yang, J., Ye, D., Hua, G.: Lq-nets: Learned quantization for highly accurate and compact deep neural networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 365-382 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.787, + 0.728 + ], + "angle": 0, + "content": "47. Zhang, S., Zhou, Q., Wang, Z., Wang, F., Yan, J.: Patch-level contrastive learning via positional query for visual pre-training (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.787, + 0.784 + ], + "angle": 0, + "content": "48. Zhang, Y., Guo, X., Poggi, M., Zhu, Z., Huang, G., Mattoccia, S.: Completionformer: Depth completion with convolutions and vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18527-18536 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "49. Zhang, Y., Chen, D., Kundu, S., Li, C., Beerel, P.A.: Sal-vit: Towards latency efficient private inference on vit using selective attention search with a learnable softmax approximation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5116-5125 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.673, + 0.129 + ], + "angle": 0, + "content": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "50. Zhang, Z., Lu, X., Cao, G., Yang, Y., Jiao, L., Liu, F.: Vit-yolo: Transformer-based yolo for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2799-2808 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.245 + ], + "angle": 0, + "content": "51. Zhong, Y., Lin, M., Nan, G., Liu, J., Zhang, B., Tian, Y., Ji, R.: Intraq: Learning synthetic images with intra-class heterogeneity for zero-shot network quantization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12339-12348 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.788, + 0.286 + ], + "angle": 0, + "content": "52. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127, 302-321 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.286 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CLAMP-ViT_ Contrastive Data-Free Learning for Adaptive Post-Training Quantization of ViTs/9ac2383c-dda4-4497-a269-a04b9fe79777_origin.pdf b/2024/CLAMP-ViT_ Contrastive Data-Free Learning for Adaptive Post-Training Quantization of ViTs/9ac2383c-dda4-4497-a269-a04b9fe79777_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4bd5c7108b4207c5e98e06d5584263a25c1228aa --- /dev/null +++ b/2024/CLAMP-ViT_ Contrastive Data-Free Learning for Adaptive Post-Training Quantization of ViTs/9ac2383c-dda4-4497-a269-a04b9fe79777_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ed82636de0b7feb62dea89090ee62f7a41dcbb0cfe42abb7c431da59efe207f +size 6568953 diff --git a/2024/CLAMP-ViT_ Contrastive Data-Free Learning for Adaptive Post-Training Quantization of ViTs/full.md b/2024/CLAMP-ViT_ Contrastive Data-Free Learning for Adaptive Post-Training Quantization of ViTs/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c4cb1964881e66bd56669c82ceb698611131494c --- /dev/null +++ b/2024/CLAMP-ViT_ Contrastive Data-Free Learning for Adaptive Post-Training Quantization of ViTs/full.md @@ -0,0 +1,344 @@ +# CLAMP-ViT: Contrastive Data-Free Learning for Adaptive Post-Training Quantization of ViTs + +Akshat Ramachandran1, Souvik Kundu2, and Tushar Krishna1 + +1 Georgia Institute of Technology, Atlanta, GA, USA + +akshat.r@gatech.edu, tushar@ece.gatech.edu + +$^{2}$ Intel Labs, San Diego, CA, USA + +souvikk.kundu@intel.com + +Abstract. We present CLAMP-ViT, a data-free post-training quantization method for vision transformers (ViTs). We identify the limitations of recent techniques, notably their inability to leverage meaningful inter-patch relationships, leading to the generation of simplistic and semantically vague data, impacting quantization accuracy. CLAMP-ViT employs a two-stage approach, cyclically adapting between data generation and model quantization. Specifically, we incorporate a patch-level contrastive learning scheme to generate richer, semantically meaningful data. Furthermore, we leverage contrastive learning in layer-wise evolutionary search for fixed- and mixed-precision quantization to identify optimal quantization parameters while mitigating the effects of a non-smooth loss landscape. Extensive evaluations across various vision tasks demonstrate the superiority of CLAMP-ViT, with performance improvements of up to $3\%$ in top-1 accuracy for classification, 0.6 mAP for object detection, and 1.5 mIoU for segmentation at similar or better compression ratio over existing alternatives. Code is available at https://github.com/georgia-tech-synergy-lab/CLAMP-ViT.git + +Keywords: Data-free quantization $\cdot$ PTQ $\cdot$ Contrastive learning $\cdot$ Vision transformer + +# 1 Introduction + +Vision transformers [12] (ViTs) have recently gained a lot of traction due to their state-of-the-art (SoTA) performance across various computer vision (CV) tasks [34, 37, 48, 50]. Concurrently, the growing need to deploy these parameter-heavy models at the resource-limited edge [14], has inspired research on various model compression techniques. Model quantization [14, 30, 36, 45] has emerged as a popular technique to achieve memory and compute efficient deployment. Quantization reduces memory footprint and improves computation speed of a model by mapping full-precision (FP) weights to reduced precision formats (e.g., $\leq$ 8-bit INT) [13, 17, 19, 46]. In particular, quantization-aware training (QAT) allows a model to train by taking quantization approximation in to account, enabling ease of quantization. Post-training quantization (PTQ), in contrast, acts + +as a plug-and-play quantization applied on a pre-trained model. PTQ has become popular as it can leverage the pre-trained model and does not add training overhead [21, 23, 26]. We thus consider the PTQ setting in this work. However, PTQ requires access to a calibration set that is often drawn from the training data [18, 45]. This may be infeasible in situations involving privacy and security concerns [20, 49], making these techniques ill-suited to yield optimal performance. + +Recent works [25, 27] propose data-free PTQ (DFQ), generating synthetic calibration data $T_{\mathrm{syn}}$ from Gaussian noise [7, 27], embedding information from the original dataset $T_{\mathrm{orig}}$ , where $T_{\mathrm{syn}} << T_{\mathrm{orig}}$ . + +Existing DFQ methods for CNNs [2,51] exploit the batch-normalization (BN) layer statistics [33,44] to generate synthetic samples mimicking the original data distribution. For ViTs, absence of the BN layer makes these + +![](images/ce2b04350dd7881254d169e2d1cfb47df28261f16d54cc0e8a0cf747b0d1f79d.jpg) +Fig.1: Visualization of loss landscape of (a) PSAQ-ViT v2 and (b) CLAMP-ViT on DeiT-S with perturbations to quantized model weights and synthetic data [22]. + +![](images/4e6a46ce38d22d877f12978647ce2c999787c547daa070c464e7583ab7402f08.jpg) + +techniques obsolete. Recent efforts to extend DFQ to ViTs include PSAQ-ViT v1 [27] and PSAQ-ViT v2 [25]. They rely on information embedded in the attention score output of the multi-head self attention (MHSA) layer. PSAQ-ViT v1 and v2 [25,27] introduce a relative value metric namely, patch similarity, to optimize Gaussian noise towards synthetic data by maximizing the entropy of patch similarity in a global manner. However, the metric considered in PSAQ-ViT v1 and v2 assumes all patches1 to be equally important, without considering spatial sensitivity [47]. This may fail to capture semantically meaningful inter-patch relations, potentially affecting robustness of the synthetic data. As we can see in Fig. 1(a) even insignificant perturbations in the generated images (augmenting pixels) or weights (to simulate quantization process) may cause significant jaggedness in the loss landscape of PSAQ-ViT v2. This also implies that the predictions may have large discrepancy even for small input/weight perturbation. Moreover, the synthetic data generation stage in these methods does not consider the informativeness of the generated samples towards the quantization process, nor do they establish countermeasures to ameliorate the non-smooth loss landscape during quantization, resulting in sub-optimal parameter search and poor generalization to test set [5]. + +Our contributions. The discussion above hints at the potential limitation of [25,27] in capturing semantically meaningful and robust inter-patch relationships to generate synthetic data that is well-suited to quantization. Towards solving these limitations, we present contrastive data-free learning for adaptive post-training quantization of ViTs (CLAMP-ViT), a general DFQ method applicable to a wide range of vision tasks. To the best of our knowledge, CLAMP- + +ViT is the first work to support both fixed-² and mixed-precision³ DFQ of ViTs, starting from a pre-trained FP model. + +Specifically, CLAMP-ViT utilizes the architectural characteristics of ViTs and inherent properties of real-images to generate semantically rich and meaningful data while ensuring spatial sensitivity. Here we leverage a novel patch-level contrastive learning scheme, where for each patch (anchor) in the MHSA layer output, we treat semantically similar patches in a neighborhood around the patch as positive patches (evaluated using cosine similarity) and the remaining patches in the neighborhood as negative patches (see Fig. 3 for an intuitive visualization). We then leverage a patch-level contrastive loss that drives the representation of the anchor patch closer to the positive patches and away from the negative patches, enabling exploration of semantically meaningful relations. + +Recently, Evol-Q [14] identified a limitation of gradient based methods [25] to search over quantization parameters of ViTs having non-smooth loss landscape, leading to poor accuracy and generalization. We take inspiration from this and a recent work [35] to design a layer-wise evolutionary search to identify the optimal bit-width and scale factors for each layer. Additionally, we propose a novel local contrastive objective to capture distributional variance in intermediate layer outputs (crucial for layer-wise quantization) and improve search convergence. This loss sufficiently captures the representational divergence of intermediate layers outputs to identify optimal quantization parameters and yields a smooth loss landscape, as demonstrated in Fig. 1(b), enabling generalizability and better performance on test data [5]. Furthermore, to ensure the generated data is adaptive to the quantization process, CLAMP-ViT performs a cyclically adaptive strategy alternating between data-generation and quantization. + +To evaluate the merits of CLAMP-ViT we conduct extensive experiments on image classification, detection, and segmentation with different ViT variants and quantization scenarios and observe superior performance over SoTA. + +# 2 Related Works + +Data-Driven PTQ for ViTs. PTQ offers an efficient alternative to QAT [23,26] by directly quantizing pre-trained models without the need for compute-heavy retraining. In specific, PTQ4ViT [45] employs twin uniform quantization and Hessian-guided scale optimization. FQ-ViT [30] uses a power-of-two factor quantization to handle inter-channel variation in LayerNorm and log-INT-Softmax. RepQ-ViT [28] separates quantization and inference, optimizing accuracy and efficiency by employing scale-reparameterized quantizers. These methods apply fixed-precision quantization, assuming all layers support similar approximations, potentially leading to sub-optimal accuracy [36]. On the other hand, techniques like, VT-PTQ [32] adopt mixed precision for specific modules based on sensitivity metrics, while PMQ [41] and LRP-QViT [36] allocate bit-widths by assessing both layer sensitivity and contribution to the output, + +respectively. However, all these methods assume a part of training set to be available for calibration which may be infeasible due to privacy [20]. + +Data-Free PTQ for ViTs. Recent data-free PTQ efforts (DFQ) for ViTs including PSAQ-ViT v1 [27] and v2 [25] utilize a "patch similarity" metric to refine Gaussian noise to synthetic data resembling real images. In specific, both these approaches leverage the fact that the self-attention has different response to real image and noise. PSAQ-ViT v1 [27] uses a two-stage cascaded framework to generate synthetic data for model quantization on image classification tasks. PSAQ-ViT v2 [25] expanded this to a broader range of tasks, employing a Min-Max game between full precision and quantized models for data generation and quantization. Despite these innovations, both the versions face several limitations as highlighted in Sec. 1, leading to poor DFQ performance, particularly at reduced precision. Additionally, these methods support only fixed-precision quantization. CLAMP-ViT, in contrast, introduces a DFQ method that addresses their shortcomings in yielding SoTA performance, while supporting both fixed and mixed-precision quantization. + +Contrastive Learning. Contrastive learning as a technique has been widely adopted in self-supervised settings [4,15] and is proven to combat overfitting via regularization against negative samples. Recently, Evol-Q [14] demonstrated the benefits of a global contrastive objective [8,43] coupled with evolutionary search in smoothening the loss landscape while calibrating the scaling factors of a prequantized model to enhance accuracy. Unlike [14] that only aims to adjust the scale factors of each layer of a pre-quantized model, we conduct complete quantization starting from a FP model. On evaluation of the global contrastive objective used in Evol-Q, we find it to be sub-optimal, causing premature convergence while quantizing an FP model. Instead, we present a local contrastive loss that captures the distributional variance in intermediate layer outputs, drastically improving convergence and hence, identification of quantization parameters. + +# 3 CLAMP-ViT Framework + +We present an overview of CLAMP-ViT in Fig. 2. In this section, we first go through notations, computational process of ViTs and quantization strategy in the preliminaries, followed by a detailed description of the proposed contrastive loss and the two stages of CLAMP-ViT. Finally, the overall DFQ pipeline is summarized and presented (refer to Supplementary for the detailed Algorithm). + +# 3.1 Preliminaries + +Notations. Let $X \in \mathcal{R}^{H \times W \times C}$ be the input image to an $L$ -layer ViT, where $(H, W, C)$ are the height, width, and channels, respectively (we ignore the batch dimension for simplicity). The input is partitioned into $N$ non-overlapping patches that are then linearly transformed to $d$ -dimensional patch embeddings, $(\mathcal{R}^{N \times d})$ that is passed through an encoder consisting of a series of transformer layers each composed of an MHSA and an MLP module. Each MHSA module is composed + +![](images/f7653a12afd9e5305c763845f5a65525f8c086f2639047166dcdf38e838984f4.jpg) +Fig.2: Overview of the cyclically evolving two-stage CLAMP-ViT framework. In stage 1 (① - ②), $\mathcal{L}^{SG}$ is minimized to update Gaussian noise towards synthesizing data. Stage 2 (③ - ⑤) conducts layer-wise evolutionary search to identify optimal quantization parameters while minimizes $\mathcal{L}^F$ . Illustrated with multiple instances of models for clarity, only one instance of each model is used in the framework. + +of $h$ heads, to capture long-range patch correlations [24]. At head $j$ , $X$ is transformed to query $(Q_j)$ , key $(K_j)$ , and value $(V_j)$ tensors to perform self-attention $\varPhi_j$ , that is linearly projected after concatenation over the heads, + +$$ +\Phi_ {j} \left(Q _ {j}, K _ {j}, V _ {j}\right) = \operatorname {s o f t m a x} \left(Q _ {j} K _ {j} ^ {T} / \sqrt {d}\right) V _ {j}, \tag {1} +$$ + +$$ +\operatorname {M H S A} (Q, K, V) = \mathbf {c o n c a t} \left(\Phi_ {0}, \Phi_ {1}, \dots \Phi_ {h - 1}\right) W ^ {0} \tag {2} +$$ + +The output of the $j^{th}$ head at $i^{th}$ MHSA layer is given by $O_{i,j} \in \mathcal{R}^{N \times d}$ . The series of transformer layers are succeeded by task-specific heads for classification, detection, or segmentation, depending on the nature of the vision task. + +Quantization. In this paper, we perform uniform symmetric quantization (fixed/mixed-precision) of both weights and activations for ViTs, mapping full precision values into a uniform scale determined by bit-width (b), given as, + +$$ +Q (X, \gamma , b) = \operatorname {c l i p} \left(\left\lfloor \frac {X}{\gamma} \right\rfloor , - 2 ^ {b - 1} + 1, 2 ^ {b - 1} - 1\right) \tag {3} +$$ + +where $\gamma$ is the scale factor and $X$ represents the FP tensor. + +# 3.2 Contrastive Objective + +Contrastive learning based on the infoNCE loss [39] helps learn an anchor sample from both the similar (positive) and dissimilar (negative) samples, typically using a softmax function to normalize the similarities into probabilities. However, infoNCE loss suffers from a disproportionate impact of a single positive and many negative samples [14]. This can affect learning the synthetic data as well as the quantized model parameters. Inspired by [39], we present a modified + +infoNCE loss $(\mathcal{L}_{i,j}^{C}$ in Eq. 4) that addresses this imbalance yielding latent impact of positive and negative samples pairs equally. + +$$ +\mathcal {L} _ {i, j} ^ {C} = - \log \frac {\sum_ {p +} \exp \left(\lambda_ {i , j} ^ {p} \cdot \lambda_ {i , j} ^ {p +} / \tau\right)}{\sum_ {p +} \exp \left(\lambda_ {i , j} ^ {p} \cdot \lambda_ {i , j} ^ {p +} / \tau\right) + \sum_ {p -} \exp \left(\lambda_ {i , j} ^ {p} \cdot \lambda_ {i , j} ^ {p -} / \tau\right)} \tag {4} +$$ + +here superscript $p$ , $p+$ , and $p-$ correspond to the anchor prediction (drawn from quantized model), positive, and negative prediction (drawn from FP model), respectively. $\tau$ controls the concentration level [39]. Subscript $(i,j)$ represents (# layer id, # head id), and (# layer id, # batch id) for the data generation and quantization stage, respectively. $\lambda$ represents the variable taken to evaluate the log likelihood—the intermediate layer patch embeddings (stage 1) or layer output activations (stage 2). The final stage $s$ loss is given as $\mathcal{L}^{C_s} = \sum_i\sum_j\mathcal{L}_{i,j}^{C_s}$ . + +# 3.3 Stage 1: Synthetic Data Generation + +Our goal in Stage 1 is to generate semantically rich and meaningful images that can reliably exploit inter-patch relations while ensuring informativeness to the quantization process. For each "anchor patch" of an MHSA output in the quantized model, we first locate positive and + +![](images/bb55ab2dde3d6e5e061692689f734e3846f5a261fb1c2b3d95ee9b778952e1fd.jpg) +Fig. 3: Intuitive visualization of positive and negative patch selection in Stage 1. + +negative patches (see Fig. 3) from the same layer id in the FP model, from a neighborhood of the anchor's spatial position. We then leverage a contrastive learning objective (see Sec. 3.2) that operates to maximize the similarity between the anchor and positive patches while minimizing similarity with the negative patches. Selection of the anchor patch from the quantized model for guiding data generations helps the model to recognize the semantic context within the generated data, during the subsequent quantization stage (informativeness). + +Patch Neighborhood. For every $k^{th}$ anchor patch $\mathcal{P}_{ijk}$ corresponding to the $j^{\text{th}}$ head in $i^{\text{th}}$ MHSA layer of the quantized model, we identify a neighborhood of size $\mathcal{N}_{ijk}$ with the same spatial location of $\mathcal{P}_{ijk}$ as the center, but located in the MHSA layer output of the FP model. Within $\mathcal{N}_{ijk}$ , to identify the most semantically correlated patches $(\mathcal{P}_{\mathcal{N}} \in \mathcal{N}_{ijk})$ we use cosine similarity as follows, + +$$ +\rho (i, j, k) = \frac {\mathcal {P} _ {i j k} ^ {T} \cdot \mathcal {P} _ {\mathcal {N}}}{\left| \left| \mathcal {P} _ {i j k} \right| \right| _ {2} \cdot \left| \left| \mathcal {P} _ {\mathcal {N}} \right| \right| _ {2}} \tag {5} +$$ + +The cosine similarity $\rho$ is estimated $\forall P_{\mathcal{N}} \in \mathcal{N}_{ijk}$ . We then select the positive patches to be the top- $n$ patches that have the highest $\rho(i,j,k)$ with the anchor patch and the rest of the patches in the neighborhood correspond to negative patches. We empirically set $n = 4$ in our experiments for a neighborhood of size $3 \times 3$ . We then compute $\mathcal{L}_{i,j}^{C_1}$ for all anchor patches for each attention head output over all layers to get the contrastive loss $(\mathcal{L}^{C_1})$ . We then compute the sample + +generation loss $\mathcal{L}^{SG}$ by combining $\mathcal{L}^{C_1}$ and the mean absolute error (MAE) (see Sec. 3.5) output loss $(\mathcal{L}^O)$ and minimize it for data generation. + +# 3.4 Stage 2: Quantization + +We now present our layer-wise evolutionary search with a local contrastive loss-based fitness function to rank suitable quantization parameters from a large search space. We detail the fitness function and the search steps as follows. + +Fitness Function. To evaluate the quantization parameters explored by the evolutionary search algorithm, we introduce a fitness function $\mathcal{L}^F$ that combines the contrastive loss $\mathcal{L}^{C_2}$ (Eq. (4)) and the MAE loss $(\mathcal{L}^O)$ computed with respect to the targets at the output (explained in Sec. 3.5) i.e., $\mathcal{L}^F = \mathcal{L}^{C_2} + \mathcal{L}^O$ . Unlike Stage 1, we use the intermediate activations after each transformer layer to compute the contrastive loss $\mathcal{L}^{C_2}$ . For a ViT, let the set of intermediate representations of FP and quantized model be denoted as, $\mathcal{O}^{fp} = \{\mathcal{O}_0^{fp}, \mathcal{O}_1^{fp}, \dots, \mathcal{O}_{L-1}^{fp}\}$ and $\mathcal{O}^q = \{\mathcal{O}_0^q, \mathcal{O}_1^q, \dots, \mathcal{O}_{L-1}^q\}$ , respectively. However, directly using high dimensional $\mathcal{O}^{fp}$ and $\mathcal{O}^q$ can result in high compute overhead. Therefore for each layer $i$ , at the intermediate output, we perform mean pooling along the patch dimension to obtain a low-dimensional representation ( $\mathcal{O}_i \in \mathcal{R}^N$ ), reducing it by a factor of $h \times d$ . We then apply the contrastive loss (Eq. (4)) sampled within the batch dimension on the low-dimensional $\mathcal{O}_i^{fp}$ and $\mathcal{O}_i^q$ . For the layer output activation generated from each constituent of a batch of synthetic data, the anchor ( $\lambda^p$ ) corresponds to the intermediate layer output of the quantized model, positive ( $\lambda^{p+}$ ) and negative ( $\lambda^{p-}$ ) samples correspond to set of intermediate layer outputs of FP model that have the same and different targets respectively, relative to the anchor within the batch. + +Step 1: Candidate Initialization. A candidate quantization solution is encoded as a set $\Lambda$ of $L$ tuples, such that for layer $i$ , tuple $\Lambda[i]$ represents the two quantization parameters $\langle b, \gamma \rangle$ . $b$ can take any integer value between 2 and 8, while $\gamma$ is constrained to a uniform ball of radius $10^{-3}$ centered around, $\gamma[i]^{init4}$ . The candidate scale factors are sampled as $\gamma[i] = \gamma[i]^{init} + f(-10^{-3}, +10^{+3})$ , where $f$ is a random sampling function. We begin the evolutionary search by creating a population via randomly sampling $\mathcal{K}$ candidate $\Lambda$ s, each consisting of layer-wise quantization parameters. We then evaluate the fitness function $\mathcal{L}^F$ for each candidate $\Lambda$ . We create $\mathcal{K}$ tuples each with $(\Lambda, \mathcal{L}^F)$ to form the initial population. $\mathcal{L}^F$ of each initial candidate with corresponding set $\Lambda$ is pre-computed and stored to avoid recomputation. + +Step 2: Re-generation (Crossover and Mutation). Each candidate in the population is ranked based on corresponding $\mathcal{L}^F$ s (lower the better) of which the top two serve as the parents for the next candidate generation (child). When evolving candidates, perturbing too many layer parameters based on parents can lead to search instability. To mitigate this, at each evolution step we employ a layer-wise regeneration approach, evolving a single transformer layer at a time based on chosen parents, keeping all other layer parameters to that of the top-1 + +parent (ranked via $\mathcal{L}^F$ ). The child's parameter regeneration for a layer based on that of the chosen parents ( $p1, p2$ ) is formulated as, + +$$ +b _ {\text {c h i l d}} = \mathbf {r} \text {a n d o m} (\mathbf {m i n} (b _ {p 1}, b _ {p 2}) - 1, \mathbf {m a x} (b _ {p 1}, b _ {p 2}) + 1) \tag {6} +$$ + +$$ +\gamma_ {c h i l d} = \text {m e a n} \left(\gamma_ {p 1}, \gamma_ {p 2}\right) + \eta \left(- 1 0 ^ {- 3}, 1 0 ^ {3}\right) \tag {7} +$$ + +Step 3: Diversity Promoting Selection. To avoid overfitting during search we follow [11] and introduce diversity into the population. In specific, we create $P = 5$ random parents and use each of them to crossover with the child generated in Step 2 and create $P'$ diverse children by following Eq. (6), Eq. (7). + +Step 4: Evaluation and Population Update. We evaluate all generated children in the steps above and acquire $\mathcal{L}^F$ s. The child generated in Step 2 and the corresponding fitness function value is added to the population. We then rank the diversity promoting children from Step 3 and select the best child to be added to the population for the next iteration. In our layer-wise evolutionary search strategy, we employ $\mathcal{P}$ passes over all layers of a ViT, and each layer is iterated over $\mathcal{C}$ cycles in each pass. So, the population is updated $\mathcal{P} \times \mathcal{C} \times L$ times, i.e., the Step 2, 3, and 4 are iteratively executed $\mathcal{P} \times \mathcal{C} \times L$ times. + +Activation Quantization. We note that the sensitivity to quantization for activations is closely correlated to the sensitivity of weight parameters. Therefore for layer $i$ , we determine the output activation quantization parameters as follows, $b_{act}[i] = \min (8,b[i]\times 2)$ and $\gamma_{act}[i] = \gamma_{act}[i - 1] + \gamma [i]$ . + +# 3.5 Overall Pipeline + +In Fig. 2, we illustrate the complete CLAMP-ViT framework. To ensure adaptive data-generation and informativeness for quantization, we use a cyclic strategy between the two stages, updating generated data based on the quantized model's needs for optimal parameter search. The framework requires an input batch of $\mathcal{B}$ random Gaussian images $X_{\mathcal{B}}$ , and corresponding task-specific targets $T_{G_{\mathcal{B}}}$ ( $T_{G_{\mathcal{B}}}$ for each task is detailed in Sec. 4). The targets direct the synthetic image generation towards task-specific goals as well as penalize inaccurate predictions of quantized model through the output loss $\mathcal{L}^O$ , + +$$ +\mathcal {L} ^ {O} = \frac {1}{n _ {c}} \left(\left| \left| \mathcal {Q} \left(X _ {\mathcal {B}}\right) - T _ {G _ {\mathcal {B}}} \right| \right| _ {1} + \left| \left| \mathcal {F P} \left(X _ {\mathcal {B}}\right) - T _ {G _ {\mathcal {B}}} \right| \right| _ {1}\right) \tag {8} +$$ + +where $n_c$ is the number of output classes (classification) or prediction map size (segmentation/detection). The quantized model is initialized to the best candidate from $\mathcal{K}$ tuples. The framework assumes a range of bit-widths and a single bit-width for mixed- and fixed-precision search, respectively. In Stage 1, $X_{\mathcal{B}}$ is fed to the frozen quantized $(\mathcal{Q})$ and full-precision model $(\mathcal{FP})$ to minimize the sample generation loss $\mathcal{L}^{SG} = \mathcal{L}^{C_1} + \mathcal{L}^O$ , updating $X_{\mathcal{B}}$ via backpropagation for $G$ iterations. In Stage 2, we use the generated data to quantize $\mathcal{Q}$ for $\mathcal{P} \times \mathcal{C} \times L$ iterations by minimizing $\mathcal{L}^F$ . We cyclically update the generated data every $\mathcal{C}/2$ iterations. In every subsequent execution of Stage 1, we do not restart from Gaussian noise but use $X_{\mathcal{B}}$ from the previous Stage 1 execution and update it for $G/2$ iterations. In this manner, the two stages are executed alternately. + +Table 1: Fixed-precision quantization accuracy comparison with SoTA on image classification tasks with ImageNet-1k testset. 'R', 'S' signifies real and synthetic calibration data and W/A indicates weight/activation bit-width. The values in bold and underline signifies, the best performance overall, and with synthetic data, respectively. + +
ModelMethodData#ImagesW/ATop-1W/ATop-1
ViT-BBaseline--32/3284.5332/3284.53
PSAQ-ViT v1S328/837.364/825.34
PTQ4ViTR328/884.254/867.16
FQ-ViTR10008/883.314/878.73
RepQ-ViTR328/881.454/876.29
CLAMP-ViT (Ours)S328/884.194/878.73
DeiT-TBaseline--32/3272.2132/3272.21
PSAQ-ViT v1S328/871.564/865.57
PSAQ-ViT v2S328/872.174/868.61
FQ-ViTR10008/871.614/866.91
RepQ-ViTR328/872.054/868.75
CLAMP-ViT (Ours)S328/872.174/869.93
DeiT-SBaseline--32/3279.8532/3279.85
PSAQ-ViT v1S328/876.924/873.23
PSAQ-ViT v2S328/879.564/876.36
PTQ4ViTR328/879.474/8-
FQ-ViTR10008/879.174/876.93
RepQ-ViTR328/879.554/876.75
CLAMP-ViT (Ours)S328/879.554/877.03
Swin-TBaseline--32/3281.3532/3281.35
PSAQ-ViT v1 [27]S328/875.354/871.79
PSAQ-ViT v2 [25]S328/880.214/876.28
FQ-ViT [30]R10008/881.294/880.73
RepQ-ViT [28]R328/881.284/880.51
CLAMP-ViT (Ours)S328/881.174/880.28
Swin-SBaseline--32/3283.2032/3283.20
PSAQ-ViT v1 [27]S328/876.644/875.14
PSAQ-ViT v2 [25]S328/882.134/878.86
FQ-ViT [30]R10008/882.134/881.67
RepQ-ViT [28]R328/882.344/882.14
CLAMP-ViT (Ours)S328/882.574/882.51
+ +# 4 Experimental Results + +# 4.1 Experimental Setup + +Models and Datasets. We evaluate CLAMP-ViT on various ViT model families (pre-trained FP models sourced from timm [40]) for image classification, object detection and semantic segmentation detailed as follows. + +Image Classification. We use ImageNet-1K [10] having 50K testset, with DeiT-B/T/S [38], Swin-T/S [31], and ViT-B/S [12] to evaluate accuracy. + +Object detection. We use the COCO 2017 dataset [29] having approximately 20K test data. Following [25,28,30], we use the Cascade Mask R-CNN [3] framework from MMdetection library [6] with DeiT-S and Swin-S as the backbone. + +Semantic Segmentation. We use the ADE20K dataset [52] with 3K test data encompassing 150 categories with DeiT-S and Swin-S as the backbone. We adopt the UperNet framework [42] in the MMsegmentation library [9] similar to [25]. Baselines. CLAMP-ViT is evaluated against SoTA PTQ (real data) and DFQ (synthetic data) methods for quantizing models from FP in various vision tasks. For image classification, it's compared with fixed-precision methods like PSAQ- + +Table 2: Mixed-Precision Quantization accuracy comparison on image classification tasks with ImageNet-1k testset. The values in bold indicate best performance overall. + +
DeiT-TDeiT-SSwin-TSwin-S
MethodDataW/ATop-1W/ATop-1W/ATop-1W/ATop-1
LRP-QViT [36]RMP6/MP671.03MP6/MP679.03--MP6/MP682.86
VT-PTQ [32]RMP6/MP669.46MP6/MP675.10MP6/MP679.61MP6/MP678.43
CLAMP-ViT (Ours)SMP4.9/MP6.271.69MP4.7/MP5.979.43MP5.5/MP6.981.78MP5.1/MP6.382.86
+ +ViT v1 [27], v2 [25], FQ-ViT [30], RepQ-ViT [28], and PTQ4ViT [45], and mixed-precision methods such as LRP-QViT [36] and VT-PTQ [32]. In object detection (Mask R-CNN), we use fixed-precision baselines FQ-ViT [30], PSAQ-ViT v2 [25], RepQ-ViT [28], and LRP-QViT [36] for mixed-precision comparison. + +For semantic segmentation (UperNet), PSAQ-ViT v2 [25] serves as the baseline. Evol-Q is excluded from the main comparison as it does not fully quantize a model starting from FP, and is presented for ablations in the supplementary. + +Table 3: Hyperparameters list. + +
ParameterDescriptionValue
GGeneration Iterations500
NNeighborhood Size3 × 3
n# Positive Patches4
PPasses10
CCycles6
K# Initial Candidates15
BBatch Size32
bBit-width[2,8]
+ +Experimental Setup. The CLAMP-ViT framework is implemented in PyTorch, and evaluated on a single NVIDIA Titan GPU. It features multiple hyperparameters, detailed in Tab. 3. + +# 4.2 Quantization Results for Image Classification + +As highlighted in Sec. 3.5, CLAMP-ViT requires a batch $\mathcal{B}$ of task-specific targets $T_{G_B}$ . For image classification on the ImageNet-1K, we create $T_{G_B} \in \mathcal{R}^{\mathcal{B} \times 1000}$ , where the class-wise probabilities are + +Table 4: Comparison of quantized DeiT-S size (MB) and BOPS (G) [1]. + +
MethodW/ABOPSSize
Baseline32/32471088
PSAQ-ViT v24/829422
CLAMP-ViT (Ours)MP4.7/MP5.926720
+ +randomly determined and assigned. We discuss and compare the performance of CLAMP-ViT for two settings, fixed-precision (Tab. 1) and mixed-precision (Tab. 2). In specific, as shown in Tab. 1 CLAMP-ViT consistently provides similar or better accuracy at W8/A8, while for lower precision W4/A8 CLAMP-ViT shows significant performance boost over all the existing alternatives. We yield $\sim 2.2\%$ and $\sim 1\%$ average accuracy improvement compared to DFQ methods [25, 27] and data-driven methods, respectively. The superior performance of CLAMP-ViT can be attributed to the cyclically adaptive data-generation process, which ensures the generated data matches the requirements and representational capabilities of the quantized model and effective traversal of the search space through evolutionary search and contrastive learning. Whereas, PSAQ-ViT v2 [25], generates increasingly difficult samples which is less beneficial for aggressive 4-bit quantization. Surprisingly, PSAQ-ViT v1 [27] achieves poor accuracy of $25.34\%$ (W4/A8) on ViT-B despite achieving reasonable accuracy on other ViTs. This result potentially supports our initial intuition that PSAQ-ViT [25, 27] does not consider the informativeness of the generated data to the quantization process. + +Evident from Tab. 2, CLAMP-ViT consistently outperforms all baselines across models for the mixed-precision setting, maintaining accuracy close to + +Table 5: Mixed and fixed-precision quantization performance comparison against SoTA techniques for object detection on COCO 2017. The values in bold and underline signifies, the best performance overall, and with synthetic data, respectively. + +
DeiT-SSwin-S
MethodDataW/A\( AP^{box} \)\( AP^{mask} \)W/A\( AP^{box} \)\( AP^{mask} \)
Baseline-32/3248.041.432/3251.844.7
FQ-ViT [30]R8/847.440.98/850.844.1
PSAQ-ViT v2 [25]S8/847.340.88/850.944.1
RepQ-ViT [28]R8/847.941.18/851.644.6
CLAMP-ViT (Ours)S8/847.941.18/851.444.6
FQ-ViT [30]R4/845.140.24/848.241.3
PSAQ-ViT v2 [25]S4/844.838.84/847.941.4
RepQ-ViT [28]R4/845.639.54/849.242.8
CLAMP-ViT (Ours)S4/845.438.94/848.542.2
LRP-QViT [36]R---\( MP_{6}/MP_{6} \)51.444.6
CLAMP-ViT (Ours)S\( MP_{5.5}/MP_{6.8} \)\( 47.9 \)\( 41.0 \)\( MP_{5.1}/MP_{6.4} \)51.744.6
+ +Table 6: Mixed-precision quantization performance comparison against PSAQ-ViT v2 for semantic segmentation. The values in bold indicate best performance overall. + +
DeiT-SSwin-S
MethodDataW/AmIoUW/AmIoU
Baseline-32/3244.032/3249.3
PSAQ-ViT v2 [25]S4/839.94/844.6
CLAMP-ViT (Ours)SMP4.8/MP6.242.4MP5.1/MP6.445.9
+ +the FP baseline despite having a significantly lower average W/A (for mixed-precision W/A is calculated by averaging bit-widths over all the layers for weights and activations). Furthermore, in Tab. 4, we report the reduction in model size and bit-operations-per-second (BOPS) [1] of mixed-precision quantized DeiT-S compared with W4/A8 PSAQ-ViT v2. CLAMP-ViT achieves $\sim 10\%$ lower model size in MB and BOPS when compared to W4/A8 PSAQ-ViT v2 while yielding $3.07\%$ improved accuracy. + +# 4.3 Quantization Results for Object Detection + +The target for object detection is $T_{G_{\mathcal{B}}} \in \mathcal{R}^{\mathcal{B} \times bb \times 5}$ where $bb$ is the number of bounding boxes in the image that is randomly selected from the integer set [1, 3]. $T_{G_{\mathcal{B}}}[\mathcal{B}, :, 0]$ corresponds to the bounding box category and $T_{G_{\mathcal{B}}}[\mathcal{B}, :, 1:4]$ is the bounding box coordinates $x, y, w, h$ [16]. Tab. 5 presents the fixed- and mixed-precision performance of CLAMP-ViT with respect to the baselines. Across different settings and models, CLAMP-ViT consistently outperforms DFQ method PSAQ-ViT v2 [25] by 0.6 box AP and 0.4 mask AP on average while closely matching performance to the SoTA data-driven method, RepQ-ViT [28]. Similar to Sec. 4.2, we observe improved performance with mixed-precision quantization, achieving near FP baseline performance. The average W/A for mixed-precision quantization for object detection is found to be higher than that for image classification due to the higher complexity of the task demanding larger bit-widths to maintain accuracy. + +# 4.4 Quantization Results for Semantic Segmentation + +The target $T_{G_{\mathcal{B}}}$ for this task is a pixel-wise classification map of the same size as $X_{\mathcal{B}}$ i.e., $T_{G_{\mathcal{B}}} \in \mathcal{R}^{\mathcal{B} \times 150 \times H \times W}$ . In Tab. 6, we show the quantization performance comparison, where CLAMP-ViT achieves average weight bit-width close + +![](images/6c6290827723187680d48497f20b367849fb261adbfe0f9be3ec406328d8cd56.jpg) +(a) + +![](images/83c64bc1cce77a1ff65fd6cd38dbc3d6843f22a3ba31a4f5dbee23c678ec9a02.jpg) +(b) +Fig. 4: Comparison of synthetic data generated by (a) PSAQ-ViT v1 [27], (b) PSAQ-ViT v2 [25] and (c) CLAMP-ViT (Ours). CLAMP-ViT generates detailed objects within contextually suitable backgrounds, boosting realism and informativeness. + +![](images/28505fc328248d6aa44a7da75a5079a15d7c0f4f9bf5ca833877f92a5808616f.jpg) +(c) + +to 4-bit while also significantly reducing activation bit-width to $\sim$ 6-bits. At a higher compression ratio of $15\%$ , CLAMP-ViT achieves $1.5\mathrm{mIoU}$ improvement on average over PSAQ-ViT v2. + +# 4.5 Analysis of Generated Samples + +Fig. 4 visualizes generated samples from PSAQ-ViT v1 [27], v2 [25], and CLAMP-ViT (after $1^{st}$ round of stage 1 execution). PSAQ-ViT v1 (Fig. 4(a)) creates images with clear class-specific foregrounds but with overly simplistic and uniform backgrounds, resulting in a lack of realism potentially affecting model accuracy. PSAQ-ViT v2 (Fig. 4(b)) introduces more complex details but fails to convey meaningful semantic information, generating images with intricate but semantically vague structures due to its unguided, difficulty-increasing data-generation strategy. In contrast, CLAMP-ViT (Fig. 4(c)) excels by synthesizing data that mirrors real-world imagery, showcasing a sophisticated understanding of semantic relationships between patches. It ensures objects are detailed and in contextually fitting backgrounds, boosting realism and informativeness. For example, CLAMP-ViT places boats on water and zebras in grasslands (Fig. 4(c), row 2), showing its capability for creating semantically relevant and visually consistent synthetic data. We believe our patch semantics exploration with a contrastive objective, makes image generation informative that mimic real-world scenes. + +# 4.6 Ablations and Discussions + +Evolutionary Search Parameters. In Fig. 5(a), we detail an experiment to determine the ideal number of passes $\mathcal{P}$ and cycles $\mathcal{C}$ for the evolutionary search process by studying the variation in Top-1 accuracy of DeiT-S with different passes and cycles keeping the other fixed at their optimal value (Tab. 3). It is evident that a cycle count of $\mathcal{C} = 6$ is optimal, as accuracy tends to decline with more cycles. Conversely, passes show a modest yet consistent improvement beyond 10, but due to the substantial rise in computational complexity, $\mathcal{P} = 10$ is deemed the most suitable choice. + +Effect of Batch Size $\mathcal{B}$ . We also show the accuracy comparison with different batch sizes ranging from 8 to 64 in Fig. 5(c). It is apparent that there is minimal increase in accuracy beyond 32 for CLAMP-ViT, justifying the choice of batch + +![](images/30b03d79aa08403d1b930a36bd5f2bbb9debe5b573002b26a38fbcdf2617b0f7.jpg) +(a) +Fig. 5: CLAMP-ViT ablations for (a) Selecting evolutionary search parameters, (b) Mixed-precision quantization accuracy with different fitness functions, (c) Effect of batch size $\mathcal{B}$ and (d) Effect of neighborhood size $\mathcal{N}$ and top- $n$ positive patches. + +![](images/08b76e5c90f863a817852fd9c607d2fb91aac6c95c053bae1fd7aad3e817ac2d.jpg) +(b) + +![](images/a6f227ecbec8a773d1df01ea21fb7d3ffd9c03d43ddb6706140b2154902ebba8.jpg) +(c) + +![](images/3cb97402d4e23d02da780da7b3e7c54cc37420712caa688ad4481792763f2b57.jpg) +(d) + +size. Interestingly, CLAMP-ViT achieves a decent accuracy of $\sim 70\%$ even with a batch size of 8, while PSAQ-ViT v1 and v2 achieve only close to $60\%$ . + +Effect of Neighborhood Size $\mathcal{N}$ and Top- $n$ Patches. We explore $3 \times 3$ , $5 \times 5$ , and all patches for neighborhood sizes and $n \in [1,4]$ for DeiT-S as shown in Fig. 5(d). A $3 \times 3$ neighborhood with top-4 patches as positives yields the highest accuracy while being computationally least demanding. In the $5 \times 5$ scenario, as the balance between positive and negative samples improves with increasing $n$ , accuracy rises but at higher computational cost. Employing all patches is computationally unviable and leads to the lowest accuracy for small $n$ values due to a positive-negative imbalance [39]. Moreover, identifying distant positive patches from the anchor neglects significant semantic patch relationships. + +Choice of Objective Function. The study in Tab. 7 examines how different loss function components of $\mathcal{L}^{SG}$ i.e., $\mathcal{L}^{C_1}$ and $\mathcal{L}^O$ affect synthetic data generation effectiveness and its impact on top-1 accuracy of W4/A8 quantization of DeiT-S.We chose fixed-precision quantization + +Table 7: Impact of different loss components in synthetic data generation. + +
LPSELC1LOW/ATop-1Accuracy
---32/3279.85
✓✗✗4/874.18
✗✗✓4/833.93
✓✗✓4/876.47
✗✓✗4/876.59
✗✓✓4/877.03
✓✓✓4/876.98
+ +to avoid any bias from mixed-precision quantization, which might typically favor higher bitwidths to lessen accuracy loss due to low-quality images. Results show that a linear mix of $\mathcal{L}^{C_1}$ and $\mathcal{L}^O$ ( $\mathcal{L}^{SG}$ ) achieves highest accuracy, while using $\mathcal{L}^O$ alone leads to the lowest, indicating its limited utility in leveraging ViTs for synthetic data. In Tab. 7 $\mathcal{L}^{PSE}$ corresponds to the patch similarity metric employed in [25]. While combining $\mathcal{L}^{PSE}$ with $\mathcal{L}^O$ [25] does offer a moderate accuracy boost, it falls short of the performance with $\mathcal{L}^{SG}$ , due to inherent limitations of the patch similarity metric highlighted in Sec. 1. Furthermore, using all three loss functions simultaneously closely matches the performance of $\mathcal{L}^{SG}$ further demonstrating that our proposed $\mathcal{L}^{C_1}$ has the major contribution towards final accuracy. + +For mixed-precision quantization of DeiT-S, the fitness function's accuracy $(\mathcal{L}^F = \mathcal{L}^{C_2} + \mathcal{L}^O)$ is compared against global contrastive loss [14], MSE, and KL-divergence in Fig. 5(b). The accuracy analysis shows MSE and KL-divergence tend to overfit to synthetic data, evidenced by plateauing accuracy. Meanwhile, global contrastive loss initially matches but then accuracy gap widens from CLAMP-ViT's performance which is due to premature convergence, implying + +that quantifying intermediate layer distributional divergence is crucial to find optimal quantization parameters. + +Effect of Adaptivity. We investigate the effectiveness of the cyclic evolution every $\mathcal{C} / 2$ iterations to ensure the data generation adapts to the requirements of + +Table 8: Effect of adaptability on quantized model accuracy. + +
AdaptivityW/ATop-1 Accuracy
XMP4.9/MP6.178.06
✓MP4.7/MP5.978.97
+ +the quantized model in Tab. 8. It can be observed that with adaptivity, we are not only able to achieve lower average W/A but also have improved accuracy. + +Effect of Informativeness. To show CLAMP-ViT's effectiveness in generating informative data, we compare misprediction rates in PSAQ-ViT v1, v2, and CLAMP-ViT during quantization. PSAQ-ViT has higher misprediction rates (34% for v1, 41% for v2) than CLAMP-ViT's 22%, indicating PSAQ-ViT's data is less informative, leading to erratic predictions, sub-optimal performance. + +Limitation Discussion: Runtime Comparison. In Fig. 6, we show the runtime and + +top-1 accuracy of different techniques on an CLAMP-ViT (mixed-precision) achieves signi to the other methods (fixed-precision) with o $(5\% \uparrow)$ compared to the best DFQ method. N calibrate a pre-quantized model on real data. + +![](images/638b425e46e9de1f6e5fed0bbf8681ee55f6d7baf05e22d6b6f1915501b64ef5.jpg) +Fig.6: Runtime of different PTQ and DFQ methods where, FS: FP model as starting point. QS: Prequantized model as starting point. + +# 5 Conclusions + +This paper presents CLAMP-ViT, a novel mixed-precision DFQ technique using cyclic adaptation and contrastive learning. It employs patch-level contrastive learning that leverages properties of the MHSA modules for data generation. A local contrastive objective and layer-wise evolutionary search identify optimal quantization parameters while ensuring a smooth loss landscape. Experiments across CV tasks show superior performance of CLAMP-ViT, achieving up to $3\%$ top-1 accuracy for classification, 0.6 mAP for detection, and 1.5 mIoU for segmentation. Future work aims to focus on extending its application to a wider range of architectures, like VLMs. While this study focuses on a useful impact and beneficial application of synthetic data generation for optimized and carbon efficient models for deployment, it is important to also be cognizant of the potential adverse effects of synthetic data such as deepfakes or racial biases. + +# Acknowledgements + +This work was supported in part by CoCoSys, one of the seven centers in JUMP 2.0, a Semiconductor Research Corporation (SRC) program sponsored by DARPA. + +# References + +1. Baskin, C., Liss, N., Schwartz, E., Zheltonozhskii, E., Giryes, R., Bronstein, A.M., Mendelson, A.: Uniq: Uniform noise injection for non-uniform quantization of neural networks. ACM Transactions on Computer Systems (TOCS) 37(1-4), 1-15 (2021) +2. Cai, Y., Yao, Z., Dong, Z., Gholami, A., Mahoney, M.W., Keutzer, K.: Zeroq: A novel zero shot quantization framework. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13169-13178 (2020) +3. Cai, Z., Vasconcelos, N.: Cascade r-cnn: Delving into high quality object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6154-6162 (2018) +4. Cao, Y.H., Sun, P., Huang, Y., Wu, J., Zhou, S.: Synergistic self-supervised and quantization learning. In: European Conference on Computer Vision. pp. 587-604. Springer (2022) +5. Chen, H., Shao, S., Wang, Z., Shang, Z., Chen, J., Ji, X., Wu, X.: Bootstrap generalization ability from loss landscape perspective. In: European Conference on Computer Vision. pp. 500-517. Springer (2022) +6. Chen, K., Wang, J., Pang, J., Cao, Y., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Xu, J., et al.: Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155 (2019) +7. Choi, K., Hong, D., Park, N., Kim, Y., Lee, J.: Qimera: Data-free quantization with synthetic boundary supporting samples. Advances in Neural Information Processing Systems 34, 14835-14847 (2021) +8. Chuang, C.Y., Hjelm, R.D., Wang, X., Vineet, V., Joshi, N., Torralba, A., Jegelka, S., Song, Y.: Robust contrastive learning against noisy views. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16670-16681 (2022) +9. Contributors, M.: MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation (2020) +0. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) +1. Dong, P., Li, L., Wei, Z., Niu, X., Tian, Z., Pan, H.: Emq: Evolving training-free proxies for automated mixed precision quantization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17076-17086 (2023) +2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +3. Fei, W., Dai, W., Li, C., Zou, J., Xiong, H.: General bitwidth assignment for efficient deep convolutional neural network quantization. IEEE Transactions on Neural Networks and Learning Systems 33(10), 5253-5267 (2021) +4. Frumkin, N., Gope, D., Marculescu, D.: Jumping through local minima: Quantization in the loss landscape of vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16978-16988 (2023) +5. Fu, Y., Yu, Q., Li, M., Ouyang, X., Chandra, V., Lin, Y.: Contrastive quant: quantization makes stronger contrastive learning. In: Proceedings of the 59th ACM/IEEE Design Automation Conference. pp. 205-210 (2022) + +16. Huang, H., Yu, P.S., Wang, C.: An introduction to image synthesis with generative adversarial nets. arXiv preprint arXiv:1803.04469 (2018) +17. Hubara, I., Courbariaux, M., Soudry, D., El-Yaniv, R., Bengio, Y.: Binarized neural networks. Advances in neural information processing systems 29 (2016) +18. Hubara, I., Nahshan, Y., Hanani, Y., Banner, R., Soudry, D.: Accurate post training quantization with small calibration sets. In: International Conference on Machine Learning. pp. 4466-4475. PMLR (2021) +19. Kim, N., Shin, D., Choi, W., Kim, G., Park, J.: Exploiting retraining-based mixed-precision quantization for low-cost dnn accelerator design. IEEE Transactions on Neural Networks and Learning Systems 32(7), 2925–2938 (2020) +20. Kundu, S., Sun, Q., Fu, Y., Pedram, M., Beerel, P.: Analyzing the confidentiality of undistillable teachers in knowledge distillation. Advances in Neural Information Processing Systems 34, 9181-9192 (2021) +21. Kundu, S., Wang, S., Sun, Q., Beerel, P.A., Pedram, M.: Bmpq: bit-gradient sensitivity-driven mixed-precision quantization of dnns from scratch. In: 2022 Design, Automation & Test in Europe Conference & Exhibition (DATE). pp. 588-591. IEEE (2022) +22. Li, H., Xu, Z., Taylor, G., Studer, C., Goldstein, T.: Visualizing the loss landscape of neural nets. Advances in neural information processing systems 31 (2018) +23. Li, Y., Xu, S., Zhang, B., Cao, X., Gao, P., Guo, G.: Q-vit: Accurate and fully quantized low-bit vision transformer. Advances in Neural Information Processing Systems 35, 34451-34463 (2022) +24. Li, Y., Yuan, G., Wen, Y., Hu, J., Evangelidis, G., Tulyakov, S., Wang, Y., Ren, J.: Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems 35, 12934-12949 (2022) +25. Li, Z., Chen, M., Xiao, J., Gu, Q.: Psaq-vit v2: Toward accurate and general data-free quantization for vision transformers. IEEE Transactions on Neural Networks and Learning Systems (2023) +26. Li, Z., Gu, Q.: I-vit: integer-only quantization for efficient vision transformer inference. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17065-17075 (2023) +27. Li, Z., Ma, L., Chen, M., Xiao, J., Gu, Q.: Patch similarity aware data-free quantization for vision transformers. In: European Conference on Computer Vision. pp. 154-170. Springer (2022) +28. Li, Z., Xiao, J., Yang, L., Gu, Q.: Repq-vit: Scale reparameterization for posttraining quantization of vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17227-17236 (2023) +29. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +30. Lin, Y., Zhang, T., Sun, P., Li, Z., Zhou, S.: Fq-vit: Post-training quantization for fully quantized vision transformer. arXiv preprint arXiv:2111.13824 (2021) +31. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) +32. Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., Gao, W.: Post-training quantization for vision transformer. Advances in Neural Information Processing Systems 34, 28092-28103 (2021) + +33. Peters, J.W., Welling, M.: Probabilistic binary neural networks. arXiv preprint arXiv:1809.03368 (2018) +34. Ramachandran, A., Dhiman, A., Vandrotti, B.S., Kim, J.: Ntrans-net: A multiscale neutrosophic-uncertainty guided transformer network for indoor depth completion. In: 2023 IEEE International Conference on Image Processing (ICIP). pp. 905–909. IEEE (2023) +35. Ramachandran, A., Wan, Z., Jeong, G., Gustafson, J., Krishna, T.: Algorithm-hardware co-design of distribution-aware logarithmic-posit encodings for efficient dnn inference. arXiv preprint arXiv:2403.05465 (2024) +36. Ranjan, N., Savakis, A.: Lrp-qvit: Mixed-precision vision transformer quantization via layer-wise relevance propagation. arXiv preprint arXiv:2401.11243 (2024) +37. Strudel, R., Garcia, R., Laptev, I., Schmid, C.: Segmenter: Transformer for semantic segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7262-7272 (2021) +38. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: International conference on machine learning. pp. 10347-10357. PMLR (2021) +39. Wang, J., Li, J., Li, W., Xuan, L., Zhang, T., Wang, W.: Positive-negative equal contrastive loss for semantic segmentation. Neurocomputing 535, 13-24 (2023) +40. Wightman, R.: Pytorch image models. https://github.com/rwrightman/pytorch-image-models (2019). https://doi.org/10.5281/zenodo.4414861 +41. Xiao, J., Li, Z., Yang, L., Gu, Q.: Patch-wise mixed-precision quantization of vision transformer. arXiv preprint arXiv:2305.06559 (2023) +42. Xiao, T., Liu, Y., Zhou, B., Jiang, Y., Sun, J.: Unified perceptual parsing for scene understanding. In: Proceedings of the European conference on computer vision (ECCV). pp. 418-434 (2018) +43. Yeh, C.H., Hong, C.Y., Hsu, Y.C., Liu, T.L., Chen, Y., LeCun, Y.: Decoupled contrastive learning. In: European Conference on Computer Vision. pp. 668-684. Springer (2022) +44. Yin, H., Molchanov, P., Alvarez, J.M., Li, Z., Mallya, A., Hoiem, D., Jha, N.K., Kautz, J.: Dreaming to distill: Data-free knowledge transfer via deepinversion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8715-8724 (2020) +45. Yuan, Z., Xue, C., Chen, Y., Wu, Q., Sun, G.: Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization. In: European Conference on Computer Vision. pp. 191-207. Springer (2022) +46. Zhang, D., Yang, J., Ye, D., Hua, G.: Lq-nets: Learned quantization for highly accurate and compact deep neural networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 365-382 (2018) +47. Zhang, S., Zhou, Q., Wang, Z., Wang, F., Yan, J.: Patch-level contrastive learning via positional query for visual pre-training (2023) +48. Zhang, Y., Guo, X., Poggi, M., Zhu, Z., Huang, G., Mattoccia, S.: Completionformer: Depth completion with convolutions and vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18527-18536 (2023) +49. Zhang, Y., Chen, D., Kundu, S., Li, C., Beerel, P.A.: Sal-vit: Towards latency efficient private inference on vit using selective attention search with a learnable softmax approximation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5116-5125 (2023) + +50. Zhang, Z., Lu, X., Cao, G., Yang, Y., Jiao, L., Liu, F.: Vit-yolo: Transformer-based yolo for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2799-2808 (2021) +51. Zhong, Y., Lin, M., Nan, G., Liu, J., Zhang, B., Tian, Y., Ji, R.: Intraq: Learning synthetic images with intra-class heterogeneity for zero-shot network quantization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12339-12348 (2022) +52. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127, 302-321 (2019) \ No newline at end of file diff --git a/2024/CLAMP-ViT_ Contrastive Data-Free Learning for Adaptive Post-Training Quantization of ViTs/images.zip b/2024/CLAMP-ViT_ Contrastive Data-Free Learning for Adaptive Post-Training Quantization of ViTs/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..68a5c31bcb31195c767617695f772d43e83aa6c3 --- /dev/null +++ b/2024/CLAMP-ViT_ Contrastive Data-Free Learning for Adaptive Post-Training Quantization of ViTs/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bec1a3fa73337e3241fe568ac3a8209d7d1cf1dc8a57e7215a0a5f46dab1fb24 +size 456894 diff --git a/2024/CLAMP-ViT_ Contrastive Data-Free Learning for Adaptive Post-Training Quantization of ViTs/layout.json b/2024/CLAMP-ViT_ Contrastive Data-Free Learning for Adaptive Post-Training Quantization of ViTs/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..40b595a38d0bab77eb2da721ab7d3a26f549fd28 --- /dev/null +++ b/2024/CLAMP-ViT_ Contrastive Data-Free Learning for Adaptive Post-Training Quantization of ViTs/layout.json @@ -0,0 +1,11854 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "text", + "content": "CLAMP-ViT: Contrastive Data-Free Learning for Adaptive Post-Training Quantization of ViTs" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 152, + 168, + 462, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 168, + 462, + 181 + ], + "spans": [ + { + "bbox": [ + 152, + 168, + 462, + 181 + ], + "type": "text", + "content": "Akshat Ramachandran1, Souvik Kundu2, and Tushar Krishna1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 197, + 190, + 416, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 190, + 416, + 201 + ], + "spans": [ + { + "bbox": [ + 197, + 190, + 416, + 201 + ], + "type": "text", + "content": "1 Georgia Institute of Technology, Atlanta, GA, USA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 206, + 202, + 406, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 202, + 406, + 213 + ], + "spans": [ + { + "bbox": [ + 206, + 202, + 406, + 213 + ], + "type": "text", + "content": "akshat.r@gatech.edu, tushar@ece.gatech.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 236, + 213, + 376, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 213, + 376, + 223 + ], + "spans": [ + { + "bbox": [ + 236, + 213, + 376, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 236, + 213, + 376, + 223 + ], + "type": "text", + "content": " Intel Labs, San Diego, CA, USA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 252, + 224, + 362, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 224, + 362, + 233 + ], + "spans": [ + { + "bbox": [ + 252, + 224, + 362, + 233 + ], + "type": "text", + "content": "souvikk.kundu@intel.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 266, + 455, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 266, + 455, + 453 + ], + "spans": [ + { + "bbox": [ + 160, + 266, + 455, + 453 + ], + "type": "text", + "content": "Abstract. We present CLAMP-ViT, a data-free post-training quantization method for vision transformers (ViTs). We identify the limitations of recent techniques, notably their inability to leverage meaningful inter-patch relationships, leading to the generation of simplistic and semantically vague data, impacting quantization accuracy. CLAMP-ViT employs a two-stage approach, cyclically adapting between data generation and model quantization. Specifically, we incorporate a patch-level contrastive learning scheme to generate richer, semantically meaningful data. Furthermore, we leverage contrastive learning in layer-wise evolutionary search for fixed- and mixed-precision quantization to identify optimal quantization parameters while mitigating the effects of a non-smooth loss landscape. Extensive evaluations across various vision tasks demonstrate the superiority of CLAMP-ViT, with performance improvements of up to " + }, + { + "bbox": [ + 160, + 266, + 455, + 453 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 160, + 266, + 455, + 453 + ], + "type": "text", + "content": " in top-1 accuracy for classification, 0.6 mAP for object detection, and 1.5 mIoU for segmentation at similar or better compression ratio over existing alternatives. Code is available at https://github.com/georgia-tech-synergy-lab/CLAMP-ViT.git" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 464, + 453, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 464, + 453, + 485 + ], + "spans": [ + { + "bbox": [ + 159, + 464, + 453, + 485 + ], + "type": "text", + "content": "Keywords: Data-free quantization " + }, + { + "bbox": [ + 159, + 464, + 453, + 485 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 464, + 453, + 485 + ], + "type": "text", + "content": " PTQ " + }, + { + "bbox": [ + 159, + 464, + 453, + 485 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 464, + 453, + 485 + ], + "type": "text", + "content": " Contrastive learning " + }, + { + "bbox": [ + 159, + 464, + 453, + 485 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 464, + 453, + 485 + ], + "type": "text", + "content": " Vision transformer" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 507, + 230, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 507, + 230, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 507, + 230, + 520 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Vision transformers [12] (ViTs) have recently gained a lot of traction due to their state-of-the-art (SoTA) performance across various computer vision (CV) tasks [34, 37, 48, 50]. Concurrently, the growing need to deploy these parameter-heavy models at the resource-limited edge [14], has inspired research on various model compression techniques. Model quantization [14, 30, 36, 45] has emerged as a popular technique to achieve memory and compute efficient deployment. Quantization reduces memory footprint and improves computation speed of a model by mapping full-precision (FP) weights to reduced precision formats (e.g., " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\leq" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " 8-bit INT) [13, 17, 19, 46]. In particular, quantization-aware training (QAT) allows a model to train by taking quantization approximation in to account, enabling ease of quantization. Post-training quantization (PTQ), in contrast, acts" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": "as a plug-and-play quantization applied on a pre-trained model. PTQ has become popular as it can leverage the pre-trained model and does not add training overhead [21, 23, 26]. We thus consider the PTQ setting in this work. However, PTQ requires access to a calibration set that is often drawn from the training data [18, 45]. This may be infeasible in situations involving privacy and security concerns [20, 49], making these techniques ill-suited to yield optimal performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 191, + 298, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 191, + 298, + 265 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 298, + 265 + ], + "type": "text", + "content": "Recent works [25, 27] propose data-free PTQ (DFQ), generating synthetic calibration data " + }, + { + "bbox": [ + 130, + 191, + 298, + 265 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{syn}}" + }, + { + "bbox": [ + 130, + 191, + 298, + 265 + ], + "type": "text", + "content": " from Gaussian noise [7, 27], embedding information from the original dataset " + }, + { + "bbox": [ + 130, + 191, + 298, + 265 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{orig}}" + }, + { + "bbox": [ + 130, + 191, + 298, + 265 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 191, + 298, + 265 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{syn}} << T_{\\mathrm{orig}}" + }, + { + "bbox": [ + 130, + 191, + 298, + 265 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 267, + 298, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 298, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 298, + 339 + ], + "type": "text", + "content": "Existing DFQ methods for CNNs [2,51] exploit the batch-normalization (BN) layer statistics [33,44] to generate synthetic samples mimicking the original data distribution. For ViTs, absence of the BN layer makes these" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 312, + 191, + 392, + 273 + ], + "blocks": [ + { + "bbox": [ + 312, + 191, + 392, + 273 + ], + "lines": [ + { + "bbox": [ + 312, + 191, + 392, + 273 + ], + "spans": [ + { + "bbox": [ + 312, + 191, + 392, + 273 + ], + "type": "image", + "image_path": "ce2b04350dd7881254d169e2d1cfb47df28261f16d54cc0e8a0cf747b0d1f79d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 281, + 482, + 326 + ], + "lines": [ + { + "bbox": [ + 304, + 281, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 482, + 326 + ], + "type": "text", + "content": "Fig.1: Visualization of loss landscape of (a) PSAQ-ViT v2 and (b) CLAMP-ViT on DeiT-S with perturbations to quantized model weights and synthetic data [22]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 394, + 191, + 478, + 272 + ], + "blocks": [ + { + "bbox": [ + 394, + 191, + 478, + 272 + ], + "lines": [ + { + "bbox": [ + 394, + 191, + 478, + 272 + ], + "spans": [ + { + "bbox": [ + 394, + 191, + 478, + 272 + ], + "type": "image", + "image_path": "4e6a46ce38d22d877f12978647ce2c999787c547daa070c464e7583ab7402f08.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 339, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 482, + 555 + ], + "type": "text", + "content": "techniques obsolete. Recent efforts to extend DFQ to ViTs include PSAQ-ViT v1 [27] and PSAQ-ViT v2 [25]. They rely on information embedded in the attention score output of the multi-head self attention (MHSA) layer. PSAQ-ViT v1 and v2 [25,27] introduce a relative value metric namely, patch similarity, to optimize Gaussian noise towards synthetic data by maximizing the entropy of patch similarity in a global manner. However, the metric considered in PSAQ-ViT v1 and v2 assumes all patches1 to be equally important, without considering spatial sensitivity [47]. This may fail to capture semantically meaningful inter-patch relations, potentially affecting robustness of the synthetic data. As we can see in Fig. 1(a) even insignificant perturbations in the generated images (augmenting pixels) or weights (to simulate quantization process) may cause significant jaggedness in the loss landscape of PSAQ-ViT v2. This also implies that the predictions may have large discrepancy even for small input/weight perturbation. Moreover, the synthetic data generation stage in these methods does not consider the informativeness of the generated samples towards the quantization process, nor do they establish countermeasures to ameliorate the non-smooth loss landscape during quantization, resulting in sub-optimal parameter search and poor generalization to test set [5]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 559, + 482, + 632 + ], + "type": "text", + "content": "Our contributions. The discussion above hints at the potential limitation of [25,27] in capturing semantically meaningful and robust inter-patch relationships to generate synthetic data that is well-suited to quantization. Towards solving these limitations, we present contrastive data-free learning for adaptive post-training quantization of ViTs (CLAMP-ViT), a general DFQ method applicable to a wide range of vision tasks. To the best of our knowledge, CLAMP-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 411, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 411, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 411, + 102 + ], + "type": "text", + "content": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 408, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 408, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 408, + 666 + ], + "type": "text", + "content": "1 Patch (subset of image): group of neighboring pixels in an image." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 481, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 481, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 481, + 139 + ], + "type": "text", + "content": "ViT is the first work to support both fixed-² and mixed-precision³ DFQ of ViTs, starting from a pre-trained FP model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 259 + ], + "type": "text", + "content": "Specifically, CLAMP-ViT utilizes the architectural characteristics of ViTs and inherent properties of real-images to generate semantically rich and meaningful data while ensuring spatial sensitivity. Here we leverage a novel patch-level contrastive learning scheme, where for each patch (anchor) in the MHSA layer output, we treat semantically similar patches in a neighborhood around the patch as positive patches (evaluated using cosine similarity) and the remaining patches in the neighborhood as negative patches (see Fig. 3 for an intuitive visualization). We then leverage a patch-level contrastive loss that drives the representation of the anchor patch closer to the positive patches and away from the negative patches, enabling exploration of semantically meaningful relations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 259, + 481, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 481, + 415 + ], + "type": "text", + "content": "Recently, Evol-Q [14] identified a limitation of gradient based methods [25] to search over quantization parameters of ViTs having non-smooth loss landscape, leading to poor accuracy and generalization. We take inspiration from this and a recent work [35] to design a layer-wise evolutionary search to identify the optimal bit-width and scale factors for each layer. Additionally, we propose a novel local contrastive objective to capture distributional variance in intermediate layer outputs (crucial for layer-wise quantization) and improve search convergence. This loss sufficiently captures the representational divergence of intermediate layers outputs to identify optimal quantization parameters and yields a smooth loss landscape, as demonstrated in Fig. 1(b), enabling generalizability and better performance on test data [5]. Furthermore, to ensure the generated data is adaptive to the quantization process, CLAMP-ViT performs a cyclically adaptive strategy alternating between data-generation and quantization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 415, + 481, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 481, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 481, + 451 + ], + "type": "text", + "content": "To evaluate the merits of CLAMP-ViT we conduct extensive experiments on image classification, detection, and segmentation with different ViT variants and quantization scenarios and observe superior performance over SoTA." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 468, + 243, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 468, + 243, + 481 + ], + "spans": [ + { + "bbox": [ + 131, + 468, + 243, + 481 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 492, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 482, + 636 + ], + "type": "text", + "content": "Data-Driven PTQ for ViTs. PTQ offers an efficient alternative to QAT [23,26] by directly quantizing pre-trained models without the need for compute-heavy retraining. In specific, PTQ4ViT [45] employs twin uniform quantization and Hessian-guided scale optimization. FQ-ViT [30] uses a power-of-two factor quantization to handle inter-channel variation in LayerNorm and log-INT-Softmax. RepQ-ViT [28] separates quantization and inference, optimizing accuracy and efficiency by employing scale-reparameterized quantizers. These methods apply fixed-precision quantization, assuming all layers support similar approximations, potentially leading to sub-optimal accuracy [36]. On the other hand, techniques like, VT-PTQ [32] adopt mixed precision for specific modules based on sensitivity metrics, while PMQ [41] and LRP-QViT [36] allocate bit-widths by assessing both layer sensitivity and contribution to the output," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 90, + 447, + 100 + ], + "type": "text", + "content": "CLAMP-ViT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 394, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 394, + 654 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 394, + 654 + ], + "type": "text", + "content": "2 weights/activations quantized to same precision for all layers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 654, + 429, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 429, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 429, + 665 + ], + "type": "text", + "content": "3 weights/activations quantized to different precision for different layers." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "respectively. However, all these methods assume a part of training set to be available for calibration which may be infeasible due to privacy [20]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 307 + ], + "type": "text", + "content": "Data-Free PTQ for ViTs. Recent data-free PTQ efforts (DFQ) for ViTs including PSAQ-ViT v1 [27] and v2 [25] utilize a \"patch similarity\" metric to refine Gaussian noise to synthetic data resembling real images. In specific, both these approaches leverage the fact that the self-attention has different response to real image and noise. PSAQ-ViT v1 [27] uses a two-stage cascaded framework to generate synthetic data for model quantization on image classification tasks. PSAQ-ViT v2 [25] expanded this to a broader range of tasks, employing a Min-Max game between full precision and quantized models for data generation and quantization. Despite these innovations, both the versions face several limitations as highlighted in Sec. 1, leading to poor DFQ performance, particularly at reduced precision. Additionally, these methods support only fixed-precision quantization. CLAMP-ViT, in contrast, introduces a DFQ method that addresses their shortcomings in yielding SoTA performance, while supporting both fixed and mixed-precision quantization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 452 + ], + "type": "text", + "content": "Contrastive Learning. Contrastive learning as a technique has been widely adopted in self-supervised settings [4,15] and is proven to combat overfitting via regularization against negative samples. Recently, Evol-Q [14] demonstrated the benefits of a global contrastive objective [8,43] coupled with evolutionary search in smoothening the loss landscape while calibrating the scaling factors of a prequantized model to enhance accuracy. Unlike [14] that only aims to adjust the scale factors of each layer of a pre-quantized model, we conduct complete quantization starting from a FP model. On evaluation of the global contrastive objective used in Evol-Q, we find it to be sub-optimal, causing premature convergence while quantizing an FP model. Instead, we present a local contrastive loss that captures the distributional variance in intermediate layer outputs, drastically improving convergence and hence, identification of quantization parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 470, + 303, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 470, + 303, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 470, + 303, + 483 + ], + "type": "text", + "content": "3 CLAMP-ViT Framework" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 495, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 556 + ], + "type": "text", + "content": "We present an overview of CLAMP-ViT in Fig. 2. In this section, we first go through notations, computational process of ViTs and quantization strategy in the preliminaries, followed by a detailed description of the proposed contrastive loss and the two stages of CLAMP-ViT. Finally, the overall DFQ pipeline is summarized and presented (refer to Supplementary for the detailed Algorithm)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 573, + 228, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 573, + 228, + 585 + ], + "spans": [ + { + "bbox": [ + 132, + 573, + 228, + 585 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 593, + 485, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 485, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 485, + 666 + ], + "type": "text", + "content": "Notations. Let " + }, + { + "bbox": [ + 130, + 593, + 485, + 666 + ], + "type": "inline_equation", + "content": "X \\in \\mathcal{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 130, + 593, + 485, + 666 + ], + "type": "text", + "content": " be the input image to an " + }, + { + "bbox": [ + 130, + 593, + 485, + 666 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 593, + 485, + 666 + ], + "type": "text", + "content": "-layer ViT, where " + }, + { + "bbox": [ + 130, + 593, + 485, + 666 + ], + "type": "inline_equation", + "content": "(H, W, C)" + }, + { + "bbox": [ + 130, + 593, + 485, + 666 + ], + "type": "text", + "content": " are the height, width, and channels, respectively (we ignore the batch dimension for simplicity). The input is partitioned into " + }, + { + "bbox": [ + 130, + 593, + 485, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 593, + 485, + 666 + ], + "type": "text", + "content": " non-overlapping patches that are then linearly transformed to " + }, + { + "bbox": [ + 130, + 593, + 485, + 666 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 593, + 485, + 666 + ], + "type": "text", + "content": "-dimensional patch embeddings, " + }, + { + "bbox": [ + 130, + 593, + 485, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{R}^{N \\times d})" + }, + { + "bbox": [ + 130, + 593, + 485, + 666 + ], + "type": "text", + "content": " that is passed through an encoder consisting of a series of transformer layers each composed of an MHSA and an MLP module. Each MHSA module is composed" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 411, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 411, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 411, + 102 + ], + "type": "text", + "content": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 114, + 470, + 277 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 470, + 277 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 470, + 277 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 470, + 277 + ], + "type": "image", + "image_path": "f7653a12afd9e5305c763845f5a65525f8c086f2639047166dcdf38e838984f4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 284, + 482, + 340 + ], + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 340 + ], + "type": "text", + "content": "Fig.2: Overview of the cyclically evolving two-stage CLAMP-ViT framework. In stage 1 (① - ②), " + }, + { + "bbox": [ + 130, + 284, + 482, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{SG}" + }, + { + "bbox": [ + 130, + 284, + 482, + 340 + ], + "type": "text", + "content": " is minimized to update Gaussian noise towards synthesizing data. Stage 2 (③ - ⑤) conducts layer-wise evolutionary search to identify optimal quantization parameters while minimizes " + }, + { + "bbox": [ + 130, + 284, + 482, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^F" + }, + { + "bbox": [ + 130, + 284, + 482, + 340 + ], + "type": "text", + "content": ". Illustrated with multiple instances of models for clarity, only one instance of each model is used in the framework." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 349, + 482, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 482, + 385 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 130, + 349, + 482, + 385 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 349, + 482, + 385 + ], + "type": "text", + "content": " heads, to capture long-range patch correlations [24]. At head " + }, + { + "bbox": [ + 130, + 349, + 482, + 385 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 349, + 482, + 385 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 349, + 482, + 385 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 349, + 482, + 385 + ], + "type": "text", + "content": " is transformed to query " + }, + { + "bbox": [ + 130, + 349, + 482, + 385 + ], + "type": "inline_equation", + "content": "(Q_j)" + }, + { + "bbox": [ + 130, + 349, + 482, + 385 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 130, + 349, + 482, + 385 + ], + "type": "inline_equation", + "content": "(K_j)" + }, + { + "bbox": [ + 130, + 349, + 482, + 385 + ], + "type": "text", + "content": ", and value " + }, + { + "bbox": [ + 130, + 349, + 482, + 385 + ], + "type": "inline_equation", + "content": "(V_j)" + }, + { + "bbox": [ + 130, + 349, + 482, + 385 + ], + "type": "text", + "content": " tensors to perform self-attention " + }, + { + "bbox": [ + 130, + 349, + 482, + 385 + ], + "type": "inline_equation", + "content": "\\varPhi_j" + }, + { + "bbox": [ + 130, + 349, + 482, + 385 + ], + "type": "text", + "content": ", that is linearly projected after concatenation over the heads," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 213, + 389, + 481, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 389, + 481, + 404 + ], + "spans": [ + { + "bbox": [ + 213, + 389, + 481, + 404 + ], + "type": "interline_equation", + "content": "\\Phi_ {j} \\left(Q _ {j}, K _ {j}, V _ {j}\\right) = \\operatorname {s o f t m a x} \\left(Q _ {j} K _ {j} ^ {T} / \\sqrt {d}\\right) V _ {j}, \\tag {1}", + "image_path": "2061929b967cc1ce69bfe47cff284bac530611f8ef2992c8e63c520f123da232.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 204, + 412, + 480, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 412, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 204, + 412, + 480, + 426 + ], + "type": "interline_equation", + "content": "\\operatorname {M H S A} (Q, K, V) = \\mathbf {c o n c a t} \\left(\\Phi_ {0}, \\Phi_ {1}, \\dots \\Phi_ {h - 1}\\right) W ^ {0} \\tag {2}", + "image_path": "36ecca06639afc743ee35158bbbd871b423b94ad97a5be57c611aef7f4b57294.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 430, + 482, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 466 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 466 + ], + "type": "text", + "content": "The output of the " + }, + { + "bbox": [ + 130, + 430, + 482, + 466 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 130, + 430, + 482, + 466 + ], + "type": "text", + "content": " head at " + }, + { + "bbox": [ + 130, + 430, + 482, + 466 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 430, + 482, + 466 + ], + "type": "text", + "content": " MHSA layer is given by " + }, + { + "bbox": [ + 130, + 430, + 482, + 466 + ], + "type": "inline_equation", + "content": "O_{i,j} \\in \\mathcal{R}^{N \\times d}" + }, + { + "bbox": [ + 130, + 430, + 482, + 466 + ], + "type": "text", + "content": ". The series of transformer layers are succeeded by task-specific heads for classification, detection, or segmentation, depending on the nature of the vision task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 467, + 482, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 467, + 482, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 467, + 482, + 503 + ], + "type": "text", + "content": "Quantization. In this paper, we perform uniform symmetric quantization (fixed/mixed-precision) of both weights and activations for ViTs, mapping full precision values into a uniform scale determined by bit-width (b), given as," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 209, + 511, + 481, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 511, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 209, + 511, + 481, + 536 + ], + "type": "interline_equation", + "content": "Q (X, \\gamma , b) = \\operatorname {c l i p} \\left(\\left\\lfloor \\frac {X}{\\gamma} \\right\\rfloor , - 2 ^ {b - 1} + 1, 2 ^ {b - 1} - 1\\right) \\tag {3}", + "image_path": "7586a8a6f864b2e4ac7ebd3ce4da3036be4a8c40314b4890d11d6ee20c8f47a6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 544, + 394, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 394, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 394, + 556 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 544, + 394, + 556 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 544, + 394, + 556 + ], + "type": "text", + "content": " is the scale factor and " + }, + { + "bbox": [ + 130, + 544, + 394, + 556 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 544, + 394, + 556 + ], + "type": "text", + "content": " represents the FP tensor." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 574, + 272, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 574, + 272, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 574, + 272, + 586 + ], + "type": "text", + "content": "3.2 Contrastive Objective" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": "Contrastive learning based on the infoNCE loss [39] helps learn an anchor sample from both the similar (positive) and dissimilar (negative) samples, typically using a softmax function to normalize the similarities into probabilities. However, infoNCE loss suffers from a disproportionate impact of a single positive and many negative samples [14]. This can affect learning the synthetic data as well as the quantized model parameters. Inspired by [39], we present a modified" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLAMP-ViT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 140 + ], + "type": "text", + "content": "infoNCE loss " + }, + { + "bbox": [ + 130, + 115, + 482, + 140 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{i,j}^{C}" + }, + { + "bbox": [ + 130, + 115, + 482, + 140 + ], + "type": "text", + "content": " in Eq. 4) that addresses this imbalance yielding latent impact of positive and negative samples pairs equally." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 175, + 148, + 481, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 148, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 175, + 148, + 481, + 182 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i, j} ^ {C} = - \\log \\frac {\\sum_ {p +} \\exp \\left(\\lambda_ {i , j} ^ {p} \\cdot \\lambda_ {i , j} ^ {p +} / \\tau\\right)}{\\sum_ {p +} \\exp \\left(\\lambda_ {i , j} ^ {p} \\cdot \\lambda_ {i , j} ^ {p +} / \\tau\\right) + \\sum_ {p -} \\exp \\left(\\lambda_ {i , j} ^ {p} \\cdot \\lambda_ {i , j} ^ {p -} / \\tau\\right)} \\tag {4}", + "image_path": "12eaaea8a3f1f9ce7c6ecaa835174a3a9ef674f6e0e03a89dbff6bf73e0e4b53.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 190, + 482, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 190, + 482, + 276 + ], + "spans": [ + { + "bbox": [ + 131, + 190, + 482, + 276 + ], + "type": "text", + "content": "here superscript " + }, + { + "bbox": [ + 131, + 190, + 482, + 276 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 131, + 190, + 482, + 276 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 190, + 482, + 276 + ], + "type": "inline_equation", + "content": "p+" + }, + { + "bbox": [ + 131, + 190, + 482, + 276 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 190, + 482, + 276 + ], + "type": "inline_equation", + "content": "p-" + }, + { + "bbox": [ + 131, + 190, + 482, + 276 + ], + "type": "text", + "content": " correspond to the anchor prediction (drawn from quantized model), positive, and negative prediction (drawn from FP model), respectively. " + }, + { + "bbox": [ + 131, + 190, + 482, + 276 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 131, + 190, + 482, + 276 + ], + "type": "text", + "content": " controls the concentration level [39]. Subscript " + }, + { + "bbox": [ + 131, + 190, + 482, + 276 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 131, + 190, + 482, + 276 + ], + "type": "text", + "content": " represents (# layer id, # head id), and (# layer id, # batch id) for the data generation and quantization stage, respectively. " + }, + { + "bbox": [ + 131, + 190, + 482, + 276 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 131, + 190, + 482, + 276 + ], + "type": "text", + "content": " represents the variable taken to evaluate the log likelihood—the intermediate layer patch embeddings (stage 1) or layer output activations (stage 2). The final stage " + }, + { + "bbox": [ + 131, + 190, + 482, + 276 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 131, + 190, + 482, + 276 + ], + "type": "text", + "content": " loss is given as " + }, + { + "bbox": [ + 131, + 190, + 482, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{C_s} = \\sum_i\\sum_j\\mathcal{L}_{i,j}^{C_s}" + }, + { + "bbox": [ + 131, + 190, + 482, + 276 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 291, + 340, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 291, + 340, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 291, + 340, + 304 + ], + "type": "text", + "content": "3.3 Stage 1: Synthetic Data Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 312, + 288, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 312, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 312, + 288, + 407 + ], + "type": "text", + "content": "Our goal in Stage 1 is to generate semantically rich and meaningful images that can reliably exploit inter-patch relations while ensuring informativeness to the quantization process. For each \"anchor patch\" of an MHSA output in the quantized model, we first locate positive and" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 294, + 315, + 470, + 373 + ], + "blocks": [ + { + "bbox": [ + 294, + 315, + 470, + 373 + ], + "lines": [ + { + "bbox": [ + 294, + 315, + 470, + 373 + ], + "spans": [ + { + "bbox": [ + 294, + 315, + 470, + 373 + ], + "type": "image", + "image_path": "bb55ab2dde3d6e5e061692689f734e3846f5a261fb1c2b3d95ee9b778952e1fd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 294, + 380, + 482, + 403 + ], + "lines": [ + { + "bbox": [ + 294, + 380, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 294, + 380, + 482, + 403 + ], + "type": "text", + "content": "Fig. 3: Intuitive visualization of positive and negative patch selection in Stage 1." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 407, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 491 + ], + "type": "text", + "content": "negative patches (see Fig. 3) from the same layer id in the FP model, from a neighborhood of the anchor's spatial position. We then leverage a contrastive learning objective (see Sec. 3.2) that operates to maximize the similarity between the anchor and positive patches while minimizing similarity with the negative patches. Selection of the anchor patch from the quantized model for guiding data generations helps the model to recognize the semantic context within the generated data, during the subsequent quantization stage (informativeness)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 491, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 552 + ], + "type": "text", + "content": "Patch Neighborhood. For every " + }, + { + "bbox": [ + 130, + 491, + 482, + 552 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 130, + 491, + 482, + 552 + ], + "type": "text", + "content": " anchor patch " + }, + { + "bbox": [ + 130, + 491, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{ijk}" + }, + { + "bbox": [ + 130, + 491, + 482, + 552 + ], + "type": "text", + "content": " corresponding to the " + }, + { + "bbox": [ + 130, + 491, + 482, + 552 + ], + "type": "inline_equation", + "content": "j^{\\text{th}}" + }, + { + "bbox": [ + 130, + 491, + 482, + 552 + ], + "type": "text", + "content": " head in " + }, + { + "bbox": [ + 130, + 491, + 482, + 552 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 130, + 491, + 482, + 552 + ], + "type": "text", + "content": " MHSA layer of the quantized model, we identify a neighborhood of size " + }, + { + "bbox": [ + 130, + 491, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{ijk}" + }, + { + "bbox": [ + 130, + 491, + 482, + 552 + ], + "type": "text", + "content": " with the same spatial location of " + }, + { + "bbox": [ + 130, + 491, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{ijk}" + }, + { + "bbox": [ + 130, + 491, + 482, + 552 + ], + "type": "text", + "content": " as the center, but located in the MHSA layer output of the FP model. Within " + }, + { + "bbox": [ + 130, + 491, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{ijk}" + }, + { + "bbox": [ + 130, + 491, + 482, + 552 + ], + "type": "text", + "content": ", to identify the most semantically correlated patches " + }, + { + "bbox": [ + 130, + 491, + 482, + 552 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}_{\\mathcal{N}} \\in \\mathcal{N}_{ijk})" + }, + { + "bbox": [ + 130, + 491, + 482, + 552 + ], + "type": "text", + "content": " we use cosine similarity as follows," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 243, + 559, + 481, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 559, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 243, + 559, + 481, + 588 + ], + "type": "interline_equation", + "content": "\\rho (i, j, k) = \\frac {\\mathcal {P} _ {i j k} ^ {T} \\cdot \\mathcal {P} _ {\\mathcal {N}}}{\\left| \\left| \\mathcal {P} _ {i j k} \\right| \\right| _ {2} \\cdot \\left| \\left| \\mathcal {P} _ {\\mathcal {N}} \\right| \\right| _ {2}} \\tag {5}", + "image_path": "3d6ccbf94ee694e60c5c5e97142c830e3b0a9f3da5491e9210fbbd050dff5539.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 592, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 592, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 592, + 482, + 667 + ], + "type": "text", + "content": "The cosine similarity " + }, + { + "bbox": [ + 131, + 592, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 131, + 592, + 482, + 667 + ], + "type": "text", + "content": " is estimated " + }, + { + "bbox": [ + 131, + 592, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\forall P_{\\mathcal{N}} \\in \\mathcal{N}_{ijk}" + }, + { + "bbox": [ + 131, + 592, + 482, + 667 + ], + "type": "text", + "content": ". We then select the positive patches to be the top-" + }, + { + "bbox": [ + 131, + 592, + 482, + 667 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 131, + 592, + 482, + 667 + ], + "type": "text", + "content": " patches that have the highest " + }, + { + "bbox": [ + 131, + 592, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\rho(i,j,k)" + }, + { + "bbox": [ + 131, + 592, + 482, + 667 + ], + "type": "text", + "content": " with the anchor patch and the rest of the patches in the neighborhood correspond to negative patches. We empirically set " + }, + { + "bbox": [ + 131, + 592, + 482, + 667 + ], + "type": "inline_equation", + "content": "n = 4" + }, + { + "bbox": [ + 131, + 592, + 482, + 667 + ], + "type": "text", + "content": " in our experiments for a neighborhood of size " + }, + { + "bbox": [ + 131, + 592, + 482, + 667 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 131, + 592, + 482, + 667 + ], + "type": "text", + "content": ". We then compute " + }, + { + "bbox": [ + 131, + 592, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{i,j}^{C_1}" + }, + { + "bbox": [ + 131, + 592, + 482, + 667 + ], + "type": "text", + "content": " for all anchor patches for each attention head output over all layers to get the contrastive loss " + }, + { + "bbox": [ + 131, + 592, + 482, + 667 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}^{C_1})" + }, + { + "bbox": [ + 131, + 592, + 482, + 667 + ], + "type": "text", + "content": ". We then compute the sample" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "type": "text", + "content": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 140 + ], + "type": "text", + "content": "generation loss " + }, + { + "bbox": [ + 130, + 115, + 482, + 140 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{SG}" + }, + { + "bbox": [ + 130, + 115, + 482, + 140 + ], + "type": "text", + "content": " by combining " + }, + { + "bbox": [ + 130, + 115, + 482, + 140 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{C_1}" + }, + { + "bbox": [ + 130, + 115, + 482, + 140 + ], + "type": "text", + "content": " and the mean absolute error (MAE) (see Sec. 3.5) output loss " + }, + { + "bbox": [ + 130, + 115, + 482, + 140 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}^O)" + }, + { + "bbox": [ + 130, + 115, + 482, + 140 + ], + "type": "text", + "content": " and minimize it for data generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 158, + 270, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 158, + 270, + 170 + ], + "spans": [ + { + "bbox": [ + 131, + 158, + 270, + 170 + ], + "type": "text", + "content": "3.4 Stage 2: Quantization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 178, + 479, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 178, + 479, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 479, + 213 + ], + "type": "text", + "content": "We now present our layer-wise evolutionary search with a local contrastive loss-based fitness function to rank suitable quantization parameters from a large search space. We detail the fitness function and the search steps as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "text", + "content": "Fitness Function. To evaluate the quantization parameters explored by the evolutionary search algorithm, we introduce a fitness function " + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^F" + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "text", + "content": " that combines the contrastive loss " + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{C_2}" + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "text", + "content": " (Eq. (4)) and the MAE loss " + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}^O)" + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "text", + "content": " computed with respect to the targets at the output (explained in Sec. 3.5) i.e., " + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^F = \\mathcal{L}^{C_2} + \\mathcal{L}^O" + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "text", + "content": ". Unlike Stage 1, we use the intermediate activations after each transformer layer to compute the contrastive loss " + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{C_2}" + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "text", + "content": ". For a ViT, let the set of intermediate representations of FP and quantized model be denoted as, " + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{O}^{fp} = \\{\\mathcal{O}_0^{fp}, \\mathcal{O}_1^{fp}, \\dots, \\mathcal{O}_{L-1}^{fp}\\}" + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{O}^q = \\{\\mathcal{O}_0^q, \\mathcal{O}_1^q, \\dots, \\mathcal{O}_{L-1}^q\\}" + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "text", + "content": ", respectively. However, directly using high dimensional " + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{O}^{fp}" + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{O}^q" + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "text", + "content": " can result in high compute overhead. Therefore for each layer " + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "text", + "content": ", at the intermediate output, we perform mean pooling along the patch dimension to obtain a low-dimensional representation (" + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_i \\in \\mathcal{R}^N" + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "text", + "content": "), reducing it by a factor of " + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "inline_equation", + "content": "h \\times d" + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "text", + "content": ". We then apply the contrastive loss (Eq. (4)) sampled within the batch dimension on the low-dimensional " + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_i^{fp}" + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_i^q" + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "text", + "content": ". For the layer output activation generated from each constituent of a batch of synthetic data, the anchor (" + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\lambda^p" + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "text", + "content": ") corresponds to the intermediate layer output of the quantized model, positive (" + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\lambda^{p+}" + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "text", + "content": ") and negative (" + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\lambda^{p-}" + }, + { + "bbox": [ + 130, + 213, + 482, + 430 + ], + "type": "text", + "content": ") samples correspond to set of intermediate layer outputs of FP model that have the same and different targets respectively, relative to the anchor within the batch." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "text", + "content": "Step 1: Candidate Initialization. A candidate quantization solution is encoded as a set " + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "inline_equation", + "content": "\\Lambda" + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "text", + "content": " tuples, such that for layer " + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "text", + "content": ", tuple " + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "inline_equation", + "content": "\\Lambda[i]" + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "text", + "content": " represents the two quantization parameters " + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "inline_equation", + "content": "\\langle b, \\gamma \\rangle" + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "text", + "content": " can take any integer value between 2 and 8, while " + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "text", + "content": " is constrained to a uniform ball of radius " + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "text", + "content": " centered around, " + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "inline_equation", + "content": "\\gamma[i]^{init4}" + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "text", + "content": ". The candidate scale factors are sampled as " + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "inline_equation", + "content": "\\gamma[i] = \\gamma[i]^{init} + f(-10^{-3}, +10^{+3})" + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "text", + "content": " is a random sampling function. We begin the evolutionary search by creating a population via randomly sampling " + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "inline_equation", + "content": "\\mathcal{K}" + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "text", + "content": " candidate " + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "inline_equation", + "content": "\\Lambda" + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "text", + "content": "s, each consisting of layer-wise quantization parameters. We then evaluate the fitness function " + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^F" + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "text", + "content": " for each candidate " + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "inline_equation", + "content": "\\Lambda" + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "text", + "content": ". We create " + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "inline_equation", + "content": "\\mathcal{K}" + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "text", + "content": " tuples each with " + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "inline_equation", + "content": "(\\Lambda, \\mathcal{L}^F)" + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "text", + "content": " to form the initial population. " + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^F" + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "text", + "content": " of each initial candidate with corresponding set " + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "inline_equation", + "content": "\\Lambda" + }, + { + "bbox": [ + 130, + 431, + 482, + 562 + ], + "type": "text", + "content": " is pre-computed and stored to avoid recomputation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "content": "Step 2: Re-generation (Crossover and Mutation). Each candidate in the population is ranked based on corresponding " + }, + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^F" + }, + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "content": "s (lower the better) of which the top two serve as the parents for the next candidate generation (child). When evolving candidates, perturbing too many layer parameters based on parents can lead to search instability. To mitigate this, at each evolution step we employ a layer-wise regeneration approach, evolving a single transformer layer at a time based on chosen parents, keeping all other layer parameters to that of the top-1" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 90, + 447, + 100 + ], + "type": "text", + "content": "CLAMP-ViT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 387, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 387, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 387, + 666 + ], + "type": "inline_equation", + "content": "^4" + }, + { + "bbox": [ + 133, + 653, + 387, + 666 + ], + "type": "inline_equation", + "content": "(\\max (\\theta_i) - \\min (\\theta_i)) / (2^b -1)" + }, + { + "bbox": [ + 133, + 653, + 387, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 133, + 653, + 387, + 666 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 133, + 653, + 387, + 666 + ], + "type": "text", + "content": " is the weight tensor." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": "parent (ranked via " + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^F" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": "). The child's parameter regeneration for a layer based on that of the chosen parents (" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "p1, p2" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": ") is formulated as," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 188, + 147, + 481, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 147, + 481, + 159 + ], + "spans": [ + { + "bbox": [ + 188, + 147, + 481, + 159 + ], + "type": "interline_equation", + "content": "b _ {\\text {c h i l d}} = \\mathbf {r} \\text {a n d o m} (\\mathbf {m i n} (b _ {p 1}, b _ {p 2}) - 1, \\mathbf {m a x} (b _ {p 1}, b _ {p 2}) + 1) \\tag {6}", + "image_path": "48f2f39fb2d88990bbe1c55b36906ca770e6c28470eac5e59fc84e28cf82d2cb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 219, + 162, + 481, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 162, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 219, + 162, + 481, + 175 + ], + "type": "interline_equation", + "content": "\\gamma_ {c h i l d} = \\text {m e a n} \\left(\\gamma_ {p 1}, \\gamma_ {p 2}\\right) + \\eta \\left(- 1 0 ^ {- 3}, 1 0 ^ {3}\\right) \\tag {7}", + "image_path": "c2bac7e3618cafacc52a2a8f8183ca304b552147f560b6f9b495c6c14aeacae3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 180, + 479, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 180, + 479, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 180, + 479, + 228 + ], + "type": "text", + "content": "Step 3: Diversity Promoting Selection. To avoid overfitting during search we follow [11] and introduce diversity into the population. In specific, we create " + }, + { + "bbox": [ + 130, + 180, + 479, + 228 + ], + "type": "inline_equation", + "content": "P = 5" + }, + { + "bbox": [ + 130, + 180, + 479, + 228 + ], + "type": "text", + "content": " random parents and use each of them to crossover with the child generated in Step 2 and create " + }, + { + "bbox": [ + 130, + 180, + 479, + 228 + ], + "type": "inline_equation", + "content": "P'" + }, + { + "bbox": [ + 130, + 180, + 479, + 228 + ], + "type": "text", + "content": " diverse children by following Eq. (6), Eq. (7)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 228, + 480, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 480, + 323 + ], + "type": "text", + "content": "Step 4: Evaluation and Population Update. We evaluate all generated children in the steps above and acquire " + }, + { + "bbox": [ + 130, + 228, + 480, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^F" + }, + { + "bbox": [ + 130, + 228, + 480, + 323 + ], + "type": "text", + "content": "s. The child generated in Step 2 and the corresponding fitness function value is added to the population. We then rank the diversity promoting children from Step 3 and select the best child to be added to the population for the next iteration. In our layer-wise evolutionary search strategy, we employ " + }, + { + "bbox": [ + 130, + 228, + 480, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 228, + 480, + 323 + ], + "type": "text", + "content": " passes over all layers of a ViT, and each layer is iterated over " + }, + { + "bbox": [ + 130, + 228, + 480, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 228, + 480, + 323 + ], + "type": "text", + "content": " cycles in each pass. So, the population is updated " + }, + { + "bbox": [ + 130, + 228, + 480, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\times \\mathcal{C} \\times L" + }, + { + "bbox": [ + 130, + 228, + 480, + 323 + ], + "type": "text", + "content": " times, i.e., the Step 2, 3, and 4 are iteratively executed " + }, + { + "bbox": [ + 130, + 228, + 480, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\times \\mathcal{C} \\times L" + }, + { + "bbox": [ + 130, + 228, + 480, + 323 + ], + "type": "text", + "content": " times." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 324, + 479, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 479, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 479, + 373 + ], + "type": "text", + "content": "Activation Quantization. We note that the sensitivity to quantization for activations is closely correlated to the sensitivity of weight parameters. Therefore for layer " + }, + { + "bbox": [ + 130, + 324, + 479, + 373 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 324, + 479, + 373 + ], + "type": "text", + "content": ", we determine the output activation quantization parameters as follows, " + }, + { + "bbox": [ + 130, + 324, + 479, + 373 + ], + "type": "inline_equation", + "content": "b_{act}[i] = \\min (8,b[i]\\times 2)" + }, + { + "bbox": [ + 130, + 324, + 479, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 324, + 479, + 373 + ], + "type": "inline_equation", + "content": "\\gamma_{act}[i] = \\gamma_{act}[i - 1] + \\gamma [i]" + }, + { + "bbox": [ + 130, + 324, + 479, + 373 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 386, + 242, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 386, + 242, + 399 + ], + "spans": [ + { + "bbox": [ + 131, + 386, + 242, + 399 + ], + "type": "text", + "content": "3.5 Overall Pipeline" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 403, + 480, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 480, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 480, + 499 + ], + "type": "text", + "content": "In Fig. 2, we illustrate the complete CLAMP-ViT framework. To ensure adaptive data-generation and informativeness for quantization, we use a cyclic strategy between the two stages, updating generated data based on the quantized model's needs for optimal parameter search. The framework requires an input batch of " + }, + { + "bbox": [ + 130, + 403, + 480, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 130, + 403, + 480, + 499 + ], + "type": "text", + "content": " random Gaussian images " + }, + { + "bbox": [ + 130, + 403, + 480, + 499 + ], + "type": "inline_equation", + "content": "X_{\\mathcal{B}}" + }, + { + "bbox": [ + 130, + 403, + 480, + 499 + ], + "type": "text", + "content": ", and corresponding task-specific targets " + }, + { + "bbox": [ + 130, + 403, + 480, + 499 + ], + "type": "inline_equation", + "content": "T_{G_{\\mathcal{B}}}" + }, + { + "bbox": [ + 130, + 403, + 480, + 499 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 403, + 480, + 499 + ], + "type": "inline_equation", + "content": "T_{G_{\\mathcal{B}}}" + }, + { + "bbox": [ + 130, + 403, + 480, + 499 + ], + "type": "text", + "content": " for each task is detailed in Sec. 4). The targets direct the synthetic image generation towards task-specific goals as well as penalize inaccurate predictions of quantized model through the output loss " + }, + { + "bbox": [ + 130, + 403, + 480, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^O" + }, + { + "bbox": [ + 130, + 403, + 480, + 499 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 198, + 508, + 481, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 508, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 198, + 508, + 481, + 531 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {O} = \\frac {1}{n _ {c}} \\left(\\left| \\left| \\mathcal {Q} \\left(X _ {\\mathcal {B}}\\right) - T _ {G _ {\\mathcal {B}}} \\right| \\right| _ {1} + \\left| \\left| \\mathcal {F P} \\left(X _ {\\mathcal {B}}\\right) - T _ {G _ {\\mathcal {B}}} \\right| \\right| _ {1}\\right) \\tag {8}", + "image_path": "ab8bbfe71a3a67d0ee4949a2843e14c8b04e80e40545dd7af65741c033ab0a3f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "n_c" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": " is the number of output classes (classification) or prediction map size (segmentation/detection). The quantized model is initialized to the best candidate from " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{K}" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": " tuples. The framework assumes a range of bit-widths and a single bit-width for mixed- and fixed-precision search, respectively. In Stage 1, " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "X_{\\mathcal{B}}" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": " is fed to the frozen quantized " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{Q})" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": " and full-precision model " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{FP})" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": " to minimize the sample generation loss " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{SG} = \\mathcal{L}^{C_1} + \\mathcal{L}^O" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": ", updating " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "X_{\\mathcal{B}}" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": " via backpropagation for " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": " iterations. In Stage 2, we use the generated data to quantize " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\times \\mathcal{C} \\times L" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": " iterations by minimizing " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^F" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": ". We cyclically update the generated data every " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{C}/2" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": " iterations. In every subsequent execution of Stage 1, we do not restart from Gaussian noise but use " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "X_{\\mathcal{B}}" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": " from the previous Stage 1 execution and update it for " + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "G/2" + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": " iterations. In this manner, the two stages are executed alternately." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "type": "text", + "content": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 182, + 159, + 426, + 439 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "text", + "content": "Table 1: Fixed-precision quantization accuracy comparison with SoTA on image classification tasks with ImageNet-1k testset. 'R', 'S' signifies real and synthetic calibration data and W/A indicates weight/activation bit-width. The values in bold and underline signifies, the best performance overall, and with synthetic data, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 182, + 159, + 426, + 439 + ], + "lines": [ + { + "bbox": [ + 182, + 159, + 426, + 439 + ], + "spans": [ + { + "bbox": [ + 182, + 159, + 426, + 439 + ], + "type": "table", + "html": "
ModelMethodData#ImagesW/ATop-1W/ATop-1
ViT-BBaseline--32/3284.5332/3284.53
PSAQ-ViT v1S328/837.364/825.34
PTQ4ViTR328/884.254/867.16
FQ-ViTR10008/883.314/878.73
RepQ-ViTR328/881.454/876.29
CLAMP-ViT (Ours)S328/884.194/878.73
DeiT-TBaseline--32/3272.2132/3272.21
PSAQ-ViT v1S328/871.564/865.57
PSAQ-ViT v2S328/872.174/868.61
FQ-ViTR10008/871.614/866.91
RepQ-ViTR328/872.054/868.75
CLAMP-ViT (Ours)S328/872.174/869.93
DeiT-SBaseline--32/3279.8532/3279.85
PSAQ-ViT v1S328/876.924/873.23
PSAQ-ViT v2S328/879.564/876.36
PTQ4ViTR328/879.474/8-
FQ-ViTR10008/879.174/876.93
RepQ-ViTR328/879.554/876.75
CLAMP-ViT (Ours)S328/879.554/877.03
Swin-TBaseline--32/3281.3532/3281.35
PSAQ-ViT v1 [27]S328/875.354/871.79
PSAQ-ViT v2 [25]S328/880.214/876.28
FQ-ViT [30]R10008/881.294/880.73
RepQ-ViT [28]R328/881.284/880.51
CLAMP-ViT (Ours)S328/881.174/880.28
Swin-SBaseline--32/3283.2032/3283.20
PSAQ-ViT v1 [27]S328/876.644/875.14
PSAQ-ViT v2 [25]S328/882.134/878.86
FQ-ViT [30]R10008/882.134/881.67
RepQ-ViT [28]R328/882.344/882.14
CLAMP-ViT (Ours)S328/882.574/882.51
", + "image_path": "269bb678dd8d0a8846a6a3d970e222449ab8d44a6f01a6e8006992caf430d05c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 450, + 282, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 282, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 282, + 464 + ], + "type": "text", + "content": "4 Experimental Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 476, + 261, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 261, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 261, + 489 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 497, + 479, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 479, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 479, + 533 + ], + "type": "text", + "content": "Models and Datasets. We evaluate CLAMP-ViT on various ViT model families (pre-trained FP models sourced from timm [40]) for image classification, object detection and semantic segmentation detailed as follows." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 480, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 480, + 556 + ], + "type": "text", + "content": "Image Classification. We use ImageNet-1K [10] having 50K testset, with DeiT-B/T/S [38], Swin-T/S [31], and ViT-B/S [12] to evaluate accuracy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 480, + 593 + ], + "type": "text", + "content": "Object detection. We use the COCO 2017 dataset [29] having approximately 20K test data. Following [25,28,30], we use the Cascade Mask R-CNN [3] framework from MMdetection library [6] with DeiT-S and Swin-S as the backbone." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "Semantic Segmentation. We use the ADE20K dataset [52] with 3K test data encompassing 150 categories with DeiT-S and Swin-S as the backbone. We adopt the UperNet framework [42] in the MMsegmentation library [9] similar to [25]. Baselines. CLAMP-ViT is evaluated against SoTA PTQ (real data) and DFQ (synthetic data) methods for quantizing models from FP in various vision tasks. For image classification, it's compared with fixed-precision methods like PSAQ-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLAMP-ViT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 137, + 470, + 178 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 2: Mixed-Precision Quantization accuracy comparison on image classification tasks with ImageNet-1k testset. The values in bold indicate best performance overall." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 137, + 470, + 178 + ], + "lines": [ + { + "bbox": [ + 140, + 137, + 470, + 178 + ], + "spans": [ + { + "bbox": [ + 140, + 137, + 470, + 178 + ], + "type": "table", + "html": "
DeiT-TDeiT-SSwin-TSwin-S
MethodDataW/ATop-1W/ATop-1W/ATop-1W/ATop-1
LRP-QViT [36]RMP6/MP671.03MP6/MP679.03--MP6/MP682.86
VT-PTQ [32]RMP6/MP669.46MP6/MP675.10MP6/MP679.61MP6/MP678.43
CLAMP-ViT (Ours)SMP4.9/MP6.271.69MP4.7/MP5.979.43MP5.5/MP6.981.78MP5.1/MP6.382.86
", + "image_path": "81f121cbd7fda4f4c5cf8331e07502b14ca18991eb8b6446df6a960fd242ed35.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 188, + 479, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 479, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 479, + 237 + ], + "type": "text", + "content": "ViT v1 [27], v2 [25], FQ-ViT [30], RepQ-ViT [28], and PTQ4ViT [45], and mixed-precision methods such as LRP-QViT [36] and VT-PTQ [32]. In object detection (Mask R-CNN), we use fixed-precision baselines FQ-ViT [30], PSAQ-ViT v2 [25], RepQ-ViT [28], and LRP-QViT [36] for mixed-precision comparison." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 237, + 312, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 312, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 312, + 308 + ], + "type": "text", + "content": "For semantic segmentation (UperNet), PSAQ-ViT v2 [25] serves as the baseline. Evol-Q is excluded from the main comparison as it does not fully quantize a model starting from FP, and is presented for ablations in the supplementary." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 338, + 249, + 460, + 307 + ], + "blocks": [ + { + "bbox": [ + 334, + 238, + 465, + 248 + ], + "lines": [ + { + "bbox": [ + 334, + 238, + 465, + 248 + ], + "spans": [ + { + "bbox": [ + 334, + 238, + 465, + 248 + ], + "type": "text", + "content": "Table 3: Hyperparameters list." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 338, + 249, + 460, + 307 + ], + "lines": [ + { + "bbox": [ + 338, + 249, + 460, + 307 + ], + "spans": [ + { + "bbox": [ + 338, + 249, + 460, + 307 + ], + "type": "table", + "html": "
ParameterDescriptionValue
GGeneration Iterations500
NNeighborhood Size3 × 3
n# Positive Patches4
PPasses10
CCycles6
K# Initial Candidates15
BBatch Size32
bBit-width[2,8]
", + "image_path": "bc3de0682dd7d11fa5f6ac5236957b9200d3b42368c3b52f47e6a003a864ec89.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 308, + 479, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 479, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 479, + 344 + ], + "type": "text", + "content": "Experimental Setup. The CLAMP-ViT framework is implemented in PyTorch, and evaluated on a single NVIDIA Titan GPU. It features multiple hyperparameters, detailed in Tab. 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 360, + 388, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 360, + 388, + 372 + ], + "spans": [ + { + "bbox": [ + 132, + 360, + 388, + 372 + ], + "type": "text", + "content": "4.2 Quantization Results for Image Classification" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 378, + 312, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 312, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 312, + 437 + ], + "type": "text", + "content": "As highlighted in Sec. 3.5, CLAMP-ViT requires a batch " + }, + { + "bbox": [ + 130, + 378, + 312, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 130, + 378, + 312, + 437 + ], + "type": "text", + "content": " of task-specific targets " + }, + { + "bbox": [ + 130, + 378, + 312, + 437 + ], + "type": "inline_equation", + "content": "T_{G_B}" + }, + { + "bbox": [ + 130, + 378, + 312, + 437 + ], + "type": "text", + "content": ". For image classification on the ImageNet-1K, we create " + }, + { + "bbox": [ + 130, + 378, + 312, + 437 + ], + "type": "inline_equation", + "content": "T_{G_B} \\in \\mathcal{R}^{\\mathcal{B} \\times 1000}" + }, + { + "bbox": [ + 130, + 378, + 312, + 437 + ], + "type": "text", + "content": ", where the class-wise probabilities are" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 321, + 402, + 476, + 430 + ], + "blocks": [ + { + "bbox": [ + 318, + 379, + 480, + 401 + ], + "lines": [ + { + "bbox": [ + 318, + 379, + 480, + 401 + ], + "spans": [ + { + "bbox": [ + 318, + 379, + 480, + 401 + ], + "type": "text", + "content": "Table 4: Comparison of quantized DeiT-S size (MB) and BOPS (G) [1]." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 402, + 476, + 430 + ], + "lines": [ + { + "bbox": [ + 321, + 402, + 476, + 430 + ], + "spans": [ + { + "bbox": [ + 321, + 402, + 476, + 430 + ], + "type": "table", + "html": "
MethodW/ABOPSSize
Baseline32/32471088
PSAQ-ViT v24/829422
CLAMP-ViT (Ours)MP4.7/MP5.926720
", + "image_path": "34892a34dfa724d3a091bb6794e222db828a209a5ef91c47dffdac2e1574c3da.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 438, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 641 + ], + "type": "text", + "content": "randomly determined and assigned. We discuss and compare the performance of CLAMP-ViT for two settings, fixed-precision (Tab. 1) and mixed-precision (Tab. 2). In specific, as shown in Tab. 1 CLAMP-ViT consistently provides similar or better accuracy at W8/A8, while for lower precision W4/A8 CLAMP-ViT shows significant performance boost over all the existing alternatives. We yield " + }, + { + "bbox": [ + 130, + 438, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\sim 2.2\\%" + }, + { + "bbox": [ + 130, + 438, + 482, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 438, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\sim 1\\%" + }, + { + "bbox": [ + 130, + 438, + 482, + 641 + ], + "type": "text", + "content": " average accuracy improvement compared to DFQ methods [25, 27] and data-driven methods, respectively. The superior performance of CLAMP-ViT can be attributed to the cyclically adaptive data-generation process, which ensures the generated data matches the requirements and representational capabilities of the quantized model and effective traversal of the search space through evolutionary search and contrastive learning. Whereas, PSAQ-ViT v2 [25], generates increasingly difficult samples which is less beneficial for aggressive 4-bit quantization. Surprisingly, PSAQ-ViT v1 [27] achieves poor accuracy of " + }, + { + "bbox": [ + 130, + 438, + 482, + 641 + ], + "type": "inline_equation", + "content": "25.34\\%" + }, + { + "bbox": [ + 130, + 438, + 482, + 641 + ], + "type": "text", + "content": " (W4/A8) on ViT-B despite achieving reasonable accuracy on other ViTs. This result potentially supports our initial intuition that PSAQ-ViT [25, 27] does not consider the informativeness of the generated data to the quantization process." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Evident from Tab. 2, CLAMP-ViT consistently outperforms all baselines across models for the mixed-precision setting, maintaining accuracy close to" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 411, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 411, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 411, + 102 + ], + "type": "text", + "content": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 183, + 148, + 426, + 244 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 5: Mixed and fixed-precision quantization performance comparison against SoTA techniques for object detection on COCO 2017. The values in bold and underline signifies, the best performance overall, and with synthetic data, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 183, + 148, + 426, + 244 + ], + "lines": [ + { + "bbox": [ + 183, + 148, + 426, + 244 + ], + "spans": [ + { + "bbox": [ + 183, + 148, + 426, + 244 + ], + "type": "table", + "html": "
DeiT-SSwin-S
MethodDataW/A\\( AP^{box} \\)\\( AP^{mask} \\)W/A\\( AP^{box} \\)\\( AP^{mask} \\)
Baseline-32/3248.041.432/3251.844.7
FQ-ViT [30]R8/847.440.98/850.844.1
PSAQ-ViT v2 [25]S8/847.340.88/850.944.1
RepQ-ViT [28]R8/847.941.18/851.644.6
CLAMP-ViT (Ours)S8/847.941.18/851.444.6
FQ-ViT [30]R4/845.140.24/848.241.3
PSAQ-ViT v2 [25]S4/844.838.84/847.941.4
RepQ-ViT [28]R4/845.639.54/849.242.8
CLAMP-ViT (Ours)S4/845.438.94/848.542.2
LRP-QViT [36]R---\\( MP_{6}/MP_{6} \\)51.444.6
CLAMP-ViT (Ours)S\\( MP_{5.5}/MP_{6.8} \\)\\( 47.9 \\)\\( 41.0 \\)\\( MP_{5.1}/MP_{6.4} \\)51.744.6
", + "image_path": "64b7ca794a80f9f571454ee9312799a27f96ad95d864b5d99d2b59d4a04b5675.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 200, + 271, + 409, + 315 + ], + "blocks": [ + { + "bbox": [ + 130, + 245, + 481, + 268 + ], + "lines": [ + { + "bbox": [ + 130, + 245, + 481, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 481, + 268 + ], + "type": "text", + "content": "Table 6: Mixed-precision quantization performance comparison against PSAQ-ViT v2 for semantic segmentation. The values in bold indicate best performance overall." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 200, + 271, + 409, + 315 + ], + "lines": [ + { + "bbox": [ + 200, + 271, + 409, + 315 + ], + "spans": [ + { + "bbox": [ + 200, + 271, + 409, + 315 + ], + "type": "table", + "html": "
DeiT-SSwin-S
MethodDataW/AmIoUW/AmIoU
Baseline-32/3244.032/3249.3
PSAQ-ViT v2 [25]S4/839.94/844.6
CLAMP-ViT (Ours)SMP4.8/MP6.242.4MP5.1/MP6.445.9
", + "image_path": "7d227a6644140003ce2123698799e042cb2ea94f845acda3ee35b8e60211693e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 324, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 411 + ], + "type": "text", + "content": "the FP baseline despite having a significantly lower average W/A (for mixed-precision W/A is calculated by averaging bit-widths over all the layers for weights and activations). Furthermore, in Tab. 4, we report the reduction in model size and bit-operations-per-second (BOPS) [1] of mixed-precision quantized DeiT-S compared with W4/A8 PSAQ-ViT v2. CLAMP-ViT achieves " + }, + { + "bbox": [ + 130, + 324, + 482, + 411 + ], + "type": "inline_equation", + "content": "\\sim 10\\%" + }, + { + "bbox": [ + 130, + 324, + 482, + 411 + ], + "type": "text", + "content": " lower model size in MB and BOPS when compared to W4/A8 PSAQ-ViT v2 while yielding " + }, + { + "bbox": [ + 130, + 324, + 482, + 411 + ], + "type": "inline_equation", + "content": "3.07\\%" + }, + { + "bbox": [ + 130, + 324, + 482, + 411 + ], + "type": "text", + "content": " improved accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 423, + 375, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 423, + 375, + 436 + ], + "spans": [ + { + "bbox": [ + 131, + 423, + 375, + 436 + ], + "type": "text", + "content": "4.3 Quantization Results for Object Detection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 441, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 596 + ], + "type": "text", + "content": "The target for object detection is " + }, + { + "bbox": [ + 130, + 441, + 482, + 596 + ], + "type": "inline_equation", + "content": "T_{G_{\\mathcal{B}}} \\in \\mathcal{R}^{\\mathcal{B} \\times bb \\times 5}" + }, + { + "bbox": [ + 130, + 441, + 482, + 596 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 441, + 482, + 596 + ], + "type": "inline_equation", + "content": "bb" + }, + { + "bbox": [ + 130, + 441, + 482, + 596 + ], + "type": "text", + "content": " is the number of bounding boxes in the image that is randomly selected from the integer set [1, 3]. " + }, + { + "bbox": [ + 130, + 441, + 482, + 596 + ], + "type": "inline_equation", + "content": "T_{G_{\\mathcal{B}}}[\\mathcal{B}, :, 0]" + }, + { + "bbox": [ + 130, + 441, + 482, + 596 + ], + "type": "text", + "content": " corresponds to the bounding box category and " + }, + { + "bbox": [ + 130, + 441, + 482, + 596 + ], + "type": "inline_equation", + "content": "T_{G_{\\mathcal{B}}}[\\mathcal{B}, :, 1:4]" + }, + { + "bbox": [ + 130, + 441, + 482, + 596 + ], + "type": "text", + "content": " is the bounding box coordinates " + }, + { + "bbox": [ + 130, + 441, + 482, + 596 + ], + "type": "inline_equation", + "content": "x, y, w, h" + }, + { + "bbox": [ + 130, + 441, + 482, + 596 + ], + "type": "text", + "content": " [16]. Tab. 5 presents the fixed- and mixed-precision performance of CLAMP-ViT with respect to the baselines. Across different settings and models, CLAMP-ViT consistently outperforms DFQ method PSAQ-ViT v2 [25] by 0.6 box AP and 0.4 mask AP on average while closely matching performance to the SoTA data-driven method, RepQ-ViT [28]. Similar to Sec. 4.2, we observe improved performance with mixed-precision quantization, achieving near FP baseline performance. The average W/A for mixed-precision quantization for object detection is found to be higher than that for image classification due to the higher complexity of the task demanding larger bit-widths to maintain accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 612, + 406, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 612, + 406, + 624 + ], + "spans": [ + { + "bbox": [ + 131, + 612, + 406, + 624 + ], + "type": "text", + "content": "4.4 Quantization Results for Semantic Segmentation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "The target " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "T_{G_{\\mathcal{B}}}" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": " for this task is a pixel-wise classification map of the same size as " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "X_{\\mathcal{B}}" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": " i.e., " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "T_{G_{\\mathcal{B}}} \\in \\mathcal{R}^{\\mathcal{B} \\times 150 \\times H \\times W}" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": ". In Tab. 6, we show the quantization performance comparison, where CLAMP-ViT achieves average weight bit-width close" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLAMP-ViT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 114, + 260, + 175 + ], + "blocks": [ + { + "bbox": [ + 167, + 114, + 260, + 175 + ], + "lines": [ + { + "bbox": [ + 167, + 114, + 260, + 175 + ], + "spans": [ + { + "bbox": [ + 167, + 114, + 260, + 175 + ], + "type": "image", + "image_path": "6c6290827723187680d48497f20b367849fb261adbfe0f9be3ec406328d8cd56.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 175, + 217, + 179 + ], + "lines": [ + { + "bbox": [ + 209, + 175, + 217, + 179 + ], + "spans": [ + { + "bbox": [ + 209, + 175, + 217, + 179 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 261, + 114, + 353, + 175 + ], + "blocks": [ + { + "bbox": [ + 261, + 114, + 353, + 175 + ], + "lines": [ + { + "bbox": [ + 261, + 114, + 353, + 175 + ], + "spans": [ + { + "bbox": [ + 261, + 114, + 353, + 175 + ], + "type": "image", + "image_path": "83c64bc1cce77a1ff65fd6cd38dbc3d6843f22a3ba31a4f5dbee23c678ec9a02.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 303, + 175, + 310, + 179 + ], + "lines": [ + { + "bbox": [ + 303, + 175, + 310, + 179 + ], + "spans": [ + { + "bbox": [ + 303, + 175, + 310, + 179 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 189, + 480, + 222 + ], + "lines": [ + { + "bbox": [ + 130, + 189, + 480, + 222 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 480, + 222 + ], + "type": "text", + "content": "Fig. 4: Comparison of synthetic data generated by (a) PSAQ-ViT v1 [27], (b) PSAQ-ViT v2 [25] and (c) CLAMP-ViT (Ours). CLAMP-ViT generates detailed objects within contextually suitable backgrounds, boosting realism and informativeness." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 354, + 114, + 446, + 175 + ], + "blocks": [ + { + "bbox": [ + 354, + 114, + 446, + 175 + ], + "lines": [ + { + "bbox": [ + 354, + 114, + 446, + 175 + ], + "spans": [ + { + "bbox": [ + 354, + 114, + 446, + 175 + ], + "type": "image", + "image_path": "28505fc328248d6aa44a7da75a5079a15d7c0f4f9bf5ca833877f92a5808616f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 175, + 405, + 179 + ], + "lines": [ + { + "bbox": [ + 396, + 175, + 405, + 179 + ], + "spans": [ + { + "bbox": [ + 396, + 175, + 405, + 179 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 233, + 482, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 233, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 482, + 269 + ], + "type": "text", + "content": "to 4-bit while also significantly reducing activation bit-width to " + }, + { + "bbox": [ + 130, + 233, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 130, + 233, + 482, + 269 + ], + "type": "text", + "content": " 6-bits. At a higher compression ratio of " + }, + { + "bbox": [ + 130, + 233, + 482, + 269 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 130, + 233, + 482, + 269 + ], + "type": "text", + "content": ", CLAMP-ViT achieves " + }, + { + "bbox": [ + 130, + 233, + 482, + 269 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{mIoU}" + }, + { + "bbox": [ + 130, + 233, + 482, + 269 + ], + "type": "text", + "content": " improvement on average over PSAQ-ViT v2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 289, + 317, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 289, + 317, + 301 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 317, + 301 + ], + "type": "text", + "content": "4.5 Analysis of Generated Samples" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 311, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 491 + ], + "type": "text", + "content": "Fig. 4 visualizes generated samples from PSAQ-ViT v1 [27], v2 [25], and CLAMP-ViT (after " + }, + { + "bbox": [ + 130, + 311, + 482, + 491 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 130, + 311, + 482, + 491 + ], + "type": "text", + "content": " round of stage 1 execution). PSAQ-ViT v1 (Fig. 4(a)) creates images with clear class-specific foregrounds but with overly simplistic and uniform backgrounds, resulting in a lack of realism potentially affecting model accuracy. PSAQ-ViT v2 (Fig. 4(b)) introduces more complex details but fails to convey meaningful semantic information, generating images with intricate but semantically vague structures due to its unguided, difficulty-increasing data-generation strategy. In contrast, CLAMP-ViT (Fig. 4(c)) excels by synthesizing data that mirrors real-world imagery, showcasing a sophisticated understanding of semantic relationships between patches. It ensures objects are detailed and in contextually fitting backgrounds, boosting realism and informativeness. For example, CLAMP-ViT places boats on water and zebras in grasslands (Fig. 4(c), row 2), showing its capability for creating semantically relevant and visually consistent synthetic data. We believe our patch semantics exploration with a contrastive objective, makes image generation informative that mimic real-world scenes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 511, + 293, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 511, + 293, + 522 + ], + "spans": [ + { + "bbox": [ + 131, + 511, + 293, + 522 + ], + "type": "text", + "content": "4.6 Ablations and Discussions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "content": "Evolutionary Search Parameters. In Fig. 5(a), we detail an experiment to determine the ideal number of passes " + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "content": " and cycles " + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "content": " for the evolutionary search process by studying the variation in Top-1 accuracy of DeiT-S with different passes and cycles keeping the other fixed at their optimal value (Tab. 3). It is evident that a cycle count of " + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = 6" + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "content": " is optimal, as accuracy tends to decline with more cycles. Conversely, passes show a modest yet consistent improvement beyond 10, but due to the substantial rise in computational complexity, " + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = 10" + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "content": " is deemed the most suitable choice." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Effect of Batch Size " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": ". We also show the accuracy comparison with different batch sizes ranging from 8 to 64 in Fig. 5(c). It is apparent that there is minimal increase in accuracy beyond 32 for CLAMP-ViT, justifying the choice of batch" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 412, + 102 + ], + "type": "text", + "content": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 118, + 214, + 178 + ], + "blocks": [ + { + "bbox": [ + 133, + 118, + 214, + 178 + ], + "lines": [ + { + "bbox": [ + 133, + 118, + 214, + 178 + ], + "spans": [ + { + "bbox": [ + 133, + 118, + 214, + 178 + ], + "type": "image", + "image_path": "30b03d79aa08403d1b930a36bd5f2bbb9debe5b573002b26a38fbcdf2617b0f7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 178, + 183, + 186 + ], + "lines": [ + { + "bbox": [ + 175, + 178, + 183, + 186 + ], + "spans": [ + { + "bbox": [ + 175, + 178, + 183, + 186 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 193, + 481, + 227 + ], + "lines": [ + { + "bbox": [ + 130, + 193, + 481, + 227 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 481, + 227 + ], + "type": "text", + "content": "Fig. 5: CLAMP-ViT ablations for (a) Selecting evolutionary search parameters, (b) Mixed-precision quantization accuracy with different fitness functions, (c) Effect of batch size " + }, + { + "bbox": [ + 130, + 193, + 481, + 227 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 130, + 193, + 481, + 227 + ], + "type": "text", + "content": " and (d) Effect of neighborhood size " + }, + { + "bbox": [ + 130, + 193, + 481, + 227 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 130, + 193, + 481, + 227 + ], + "type": "text", + "content": " and top-" + }, + { + "bbox": [ + 130, + 193, + 481, + 227 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 193, + 481, + 227 + ], + "type": "text", + "content": " positive patches." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 215, + 118, + 304, + 178 + ], + "blocks": [ + { + "bbox": [ + 215, + 118, + 304, + 178 + ], + "lines": [ + { + "bbox": [ + 215, + 118, + 304, + 178 + ], + "spans": [ + { + "bbox": [ + 215, + 118, + 304, + 178 + ], + "type": "image", + "image_path": "08b76e5c90f863a817852fd9c607d2fb91aac6c95c053bae1fd7aad3e817ac2d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 178, + 269, + 186 + ], + "lines": [ + { + "bbox": [ + 260, + 178, + 269, + 186 + ], + "spans": [ + { + "bbox": [ + 260, + 178, + 269, + 186 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 117, + 388, + 178 + ], + "blocks": [ + { + "bbox": [ + 306, + 117, + 388, + 178 + ], + "lines": [ + { + "bbox": [ + 306, + 117, + 388, + 178 + ], + "spans": [ + { + "bbox": [ + 306, + 117, + 388, + 178 + ], + "type": "image", + "image_path": "a6f227ecbec8a773d1df01ea21fb7d3ffd9c03d43ddb6706140b2154902ebba8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 178, + 359, + 186 + ], + "lines": [ + { + "bbox": [ + 351, + 178, + 359, + 186 + ], + "spans": [ + { + "bbox": [ + 351, + 178, + 359, + 186 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 391, + 116, + 478, + 178 + ], + "blocks": [ + { + "bbox": [ + 391, + 116, + 478, + 178 + ], + "lines": [ + { + "bbox": [ + 391, + 116, + 478, + 178 + ], + "spans": [ + { + "bbox": [ + 391, + 116, + 478, + 178 + ], + "type": "image", + "image_path": "3cb97402d4e23d02da780da7b3e7c54cc37420712caa688ad4481792763f2b57.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 178, + 436, + 186 + ], + "lines": [ + { + "bbox": [ + 429, + 178, + 436, + 186 + ], + "spans": [ + { + "bbox": [ + 429, + 178, + 436, + 186 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 235, + 480, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 480, + 259 + ], + "type": "text", + "content": "size. Interestingly, CLAMP-ViT achieves a decent accuracy of " + }, + { + "bbox": [ + 130, + 235, + 480, + 259 + ], + "type": "inline_equation", + "content": "\\sim 70\\%" + }, + { + "bbox": [ + 130, + 235, + 480, + 259 + ], + "type": "text", + "content": " even with a batch size of 8, while PSAQ-ViT v1 and v2 achieve only close to " + }, + { + "bbox": [ + 130, + 235, + 480, + 259 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 235, + 480, + 259 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 262, + 480, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 480, + 371 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 480, + 371 + ], + "type": "text", + "content": "Effect of Neighborhood Size " + }, + { + "bbox": [ + 130, + 262, + 480, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 130, + 262, + 480, + 371 + ], + "type": "text", + "content": " and Top-" + }, + { + "bbox": [ + 130, + 262, + 480, + 371 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 262, + 480, + 371 + ], + "type": "text", + "content": " Patches. We explore " + }, + { + "bbox": [ + 130, + 262, + 480, + 371 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 262, + 480, + 371 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 262, + 480, + 371 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 130, + 262, + 480, + 371 + ], + "type": "text", + "content": ", and all patches for neighborhood sizes and " + }, + { + "bbox": [ + 130, + 262, + 480, + 371 + ], + "type": "inline_equation", + "content": "n \\in [1,4]" + }, + { + "bbox": [ + 130, + 262, + 480, + 371 + ], + "type": "text", + "content": " for DeiT-S as shown in Fig. 5(d). A " + }, + { + "bbox": [ + 130, + 262, + 480, + 371 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 262, + 480, + 371 + ], + "type": "text", + "content": " neighborhood with top-4 patches as positives yields the highest accuracy while being computationally least demanding. In the " + }, + { + "bbox": [ + 130, + 262, + 480, + 371 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 130, + 262, + 480, + 371 + ], + "type": "text", + "content": " scenario, as the balance between positive and negative samples improves with increasing " + }, + { + "bbox": [ + 130, + 262, + 480, + 371 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 262, + 480, + 371 + ], + "type": "text", + "content": ", accuracy rises but at higher computational cost. Employing all patches is computationally unviable and leads to the lowest accuracy for small " + }, + { + "bbox": [ + 130, + 262, + 480, + 371 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 262, + 480, + 371 + ], + "type": "text", + "content": " values due to a positive-negative imbalance [39]. Moreover, identifying distant positive patches from the anchor neglects significant semantic patch relationships." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 374, + 312, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 312, + 457 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 312, + 457 + ], + "type": "text", + "content": "Choice of Objective Function. The study in Tab. 7 examines how different loss function components of " + }, + { + "bbox": [ + 130, + 374, + 312, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{SG}" + }, + { + "bbox": [ + 130, + 374, + 312, + 457 + ], + "type": "text", + "content": " i.e., " + }, + { + "bbox": [ + 130, + 374, + 312, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{C_1}" + }, + { + "bbox": [ + 130, + 374, + 312, + 457 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 374, + 312, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^O" + }, + { + "bbox": [ + 130, + 374, + 312, + 457 + ], + "type": "text", + "content": " affect synthetic data generation effectiveness and its impact on top-1 accuracy of W4/A8 quantization of DeiT-S.We chose fixed-precision quantization" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 336, + 398, + 462, + 452 + ], + "blocks": [ + { + "bbox": [ + 318, + 375, + 480, + 397 + ], + "lines": [ + { + "bbox": [ + 318, + 375, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 318, + 375, + 480, + 397 + ], + "type": "text", + "content": "Table 7: Impact of different loss components in synthetic data generation." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 336, + 398, + 462, + 452 + ], + "lines": [ + { + "bbox": [ + 336, + 398, + 462, + 452 + ], + "spans": [ + { + "bbox": [ + 336, + 398, + 462, + 452 + ], + "type": "table", + "html": "
LPSELC1LOW/ATop-1Accuracy
---32/3279.85
✓✗✗4/874.18
✗✗✓4/833.93
✓✗✓4/876.47
✗✓✗4/876.59
✗✓✓4/877.03
✓✓✓4/876.98
", + "image_path": "7390af3a079be8ff21e74e551cff9693112d06194bd40bada3fbafdac9ea6ae1.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 458, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 482, + 589 + ], + "type": "text", + "content": "to avoid any bias from mixed-precision quantization, which might typically favor higher bitwidths to lessen accuracy loss due to low-quality images. Results show that a linear mix of " + }, + { + "bbox": [ + 130, + 458, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{C_1}" + }, + { + "bbox": [ + 130, + 458, + 482, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 458, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^O" + }, + { + "bbox": [ + 130, + 458, + 482, + 589 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 458, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{SG}" + }, + { + "bbox": [ + 130, + 458, + 482, + 589 + ], + "type": "text", + "content": ") achieves highest accuracy, while using " + }, + { + "bbox": [ + 130, + 458, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^O" + }, + { + "bbox": [ + 130, + 458, + 482, + 589 + ], + "type": "text", + "content": " alone leads to the lowest, indicating its limited utility in leveraging ViTs for synthetic data. In Tab. 7 " + }, + { + "bbox": [ + 130, + 458, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{PSE}" + }, + { + "bbox": [ + 130, + 458, + 482, + 589 + ], + "type": "text", + "content": " corresponds to the patch similarity metric employed in [25]. While combining " + }, + { + "bbox": [ + 130, + 458, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{PSE}" + }, + { + "bbox": [ + 130, + 458, + 482, + 589 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 458, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^O" + }, + { + "bbox": [ + 130, + 458, + 482, + 589 + ], + "type": "text", + "content": " [25] does offer a moderate accuracy boost, it falls short of the performance with " + }, + { + "bbox": [ + 130, + 458, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{SG}" + }, + { + "bbox": [ + 130, + 458, + 482, + 589 + ], + "type": "text", + "content": ", due to inherent limitations of the patch similarity metric highlighted in Sec. 1. Furthermore, using all three loss functions simultaneously closely matches the performance of " + }, + { + "bbox": [ + 130, + 458, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{SG}" + }, + { + "bbox": [ + 130, + 458, + 482, + 589 + ], + "type": "text", + "content": " further demonstrating that our proposed " + }, + { + "bbox": [ + 130, + 458, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{C_1}" + }, + { + "bbox": [ + 130, + 458, + 482, + 589 + ], + "type": "text", + "content": " has the major contribution towards final accuracy." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "For mixed-precision quantization of DeiT-S, the fitness function's accuracy " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}^F = \\mathcal{L}^{C_2} + \\mathcal{L}^O)" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " is compared against global contrastive loss [14], MSE, and KL-divergence in Fig. 5(b). The accuracy analysis shows MSE and KL-divergence tend to overfit to synthetic data, evidenced by plateauing accuracy. Meanwhile, global contrastive loss initially matches but then accuracy gap widens from CLAMP-ViT's performance which is due to premature convergence, implying" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLAMP-ViT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "that quantifying intermediate layer distributional divergence is crucial to find optimal quantization parameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 312, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 312, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 312, + 187 + ], + "type": "text", + "content": "Effect of Adaptivity. We investigate the effectiveness of the cyclic evolution every " + }, + { + "bbox": [ + 130, + 140, + 312, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{C} / 2" + }, + { + "bbox": [ + 130, + 140, + 312, + 187 + ], + "type": "text", + "content": " iterations to ensure the data generation adapts to the requirements of" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 332, + 163, + 466, + 186 + ], + "blocks": [ + { + "bbox": [ + 318, + 141, + 481, + 163 + ], + "lines": [ + { + "bbox": [ + 318, + 141, + 481, + 163 + ], + "spans": [ + { + "bbox": [ + 318, + 141, + 481, + 163 + ], + "type": "text", + "content": "Table 8: Effect of adaptability on quantized model accuracy." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 163, + 466, + 186 + ], + "lines": [ + { + "bbox": [ + 332, + 163, + 466, + 186 + ], + "spans": [ + { + "bbox": [ + 332, + 163, + 466, + 186 + ], + "type": "table", + "html": "
AdaptivityW/ATop-1 Accuracy
XMP4.9/MP6.178.06
✓MP4.7/MP5.978.97
", + "image_path": "98bb99737e72f4ac98ff87aeec01979247365bf7cfd3c20d6268a848f979e273.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 188, + 479, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 479, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 479, + 212 + ], + "type": "text", + "content": "the quantized model in Tab. 8. It can be observed that with adaptivity, we are not only able to achieve lower average W/A but also have improved accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 212, + 332, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 332, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 332, + 319 + ], + "type": "text", + "content": "Effect of Informativeness. To show CLAMP-ViT's effectiveness in generating informative data, we compare misprediction rates in PSAQ-ViT v1, v2, and CLAMP-ViT during quantization. PSAQ-ViT has higher misprediction rates (34% for v1, 41% for v2) than CLAMP-ViT's 22%, indicating PSAQ-ViT's data is less informative, leading to erratic predictions, sub-optimal performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 319, + 332, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 332, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 332, + 343 + ], + "type": "text", + "content": "Limitation Discussion: Runtime Comparison. In Fig. 6, we show the runtime and" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 343, + 332, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 332, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 332, + 403 + ], + "type": "text", + "content": "top-1 accuracy of different techniques on an CLAMP-ViT (mixed-precision) achieves signi to the other methods (fixed-precision) with o " + }, + { + "bbox": [ + 130, + 343, + 332, + 403 + ], + "type": "inline_equation", + "content": "(5\\% \\uparrow)" + }, + { + "bbox": [ + 130, + 343, + 332, + 403 + ], + "type": "text", + "content": " compared to the best DFQ method. N calibrate a pre-quantized model on real data." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 359, + 217, + 459, + 293 + ], + "blocks": [ + { + "bbox": [ + 359, + 217, + 459, + 293 + ], + "lines": [ + { + "bbox": [ + 359, + 217, + 459, + 293 + ], + "spans": [ + { + "bbox": [ + 359, + 217, + 459, + 293 + ], + "type": "image", + "image_path": "638b425e46e9de1f6e5fed0bbf8681ee55f6d7baf05e22d6b6f1915501b64ef5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 338, + 296, + 481, + 340 + ], + "lines": [ + { + "bbox": [ + 338, + 296, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 338, + 296, + 481, + 340 + ], + "type": "text", + "content": "Fig.6: Runtime of different PTQ and DFQ methods where, FS: FP model as starting point. QS: Prequantized model as starting point." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 420, + 225, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 420, + 225, + 433 + ], + "spans": [ + { + "bbox": [ + 131, + 420, + 225, + 433 + ], + "type": "text", + "content": "5 Conclusions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 444, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 444, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 482, + 589 + ], + "type": "text", + "content": "This paper presents CLAMP-ViT, a novel mixed-precision DFQ technique using cyclic adaptation and contrastive learning. It employs patch-level contrastive learning that leverages properties of the MHSA modules for data generation. A local contrastive objective and layer-wise evolutionary search identify optimal quantization parameters while ensuring a smooth loss landscape. Experiments across CV tasks show superior performance of CLAMP-ViT, achieving up to " + }, + { + "bbox": [ + 130, + 444, + 482, + 589 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 130, + 444, + 482, + 589 + ], + "type": "text", + "content": " top-1 accuracy for classification, 0.6 mAP for detection, and 1.5 mIoU for segmentation. Future work aims to focus on extending its application to a wider range of architectures, like VLMs. While this study focuses on a useful impact and beneficial application of synthetic data generation for optimized and carbon efficient models for deployment, it is important to also be cognizant of the potential adverse effects of synthetic data such as deepfakes or racial biases." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 605, + 247, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 247, + 619 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 247, + 619 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": "This work was supported in part by CoCoSys, one of the seven centers in JUMP 2.0, a Semiconductor Research Corporation (SRC) program sponsored by DARPA." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 411, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 411, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 411, + 101 + ], + "type": "text", + "content": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 141, + 481, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 481, + 186 + ], + "type": "text", + "content": "1. Baskin, C., Liss, N., Schwartz, E., Zheltonozhskii, E., Giryes, R., Bronstein, A.M., Mendelson, A.: Uniq: Uniform noise injection for non-uniform quantization of neural networks. ACM Transactions on Computer Systems (TOCS) 37(1-4), 1-15 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 187, + 481, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 187, + 481, + 220 + ], + "spans": [ + { + "bbox": [ + 138, + 187, + 481, + 220 + ], + "type": "text", + "content": "2. Cai, Y., Yao, Z., Dong, Z., Gholami, A., Mahoney, M.W., Keutzer, K.: Zeroq: A novel zero shot quantization framework. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13169-13178 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 220, + 481, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 220, + 481, + 253 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 481, + 253 + ], + "type": "text", + "content": "3. Cai, Z., Vasconcelos, N.: Cascade r-cnn: Delving into high quality object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6154-6162 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 254, + 481, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 254, + 481, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 254, + 481, + 286 + ], + "type": "text", + "content": "4. Cao, Y.H., Sun, P., Huang, Y., Wu, J., Zhou, S.: Synergistic self-supervised and quantization learning. In: European Conference on Computer Vision. pp. 587-604. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 287, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 287, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 287, + 481, + 319 + ], + "type": "text", + "content": "5. Chen, H., Shao, S., Wang, Z., Shang, Z., Chen, J., Ji, X., Wu, X.: Bootstrap generalization ability from loss landscape perspective. In: European Conference on Computer Vision. pp. 500-517. Springer (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 320, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 481, + 354 + ], + "type": "text", + "content": "6. Chen, K., Wang, J., Pang, J., Cao, Y., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Xu, J., et al.: Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 354, + 481, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 481, + 387 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 481, + 387 + ], + "type": "text", + "content": "7. Choi, K., Hong, D., Park, N., Kim, Y., Lee, J.: Qimera: Data-free quantization with synthetic boundary supporting samples. Advances in Neural Information Processing Systems 34, 14835-14847 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 388, + 481, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 481, + 431 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 481, + 431 + ], + "type": "text", + "content": "8. Chuang, C.Y., Hjelm, R.D., Wang, X., Vineet, V., Joshi, N., Torralba, A., Jegelka, S., Song, Y.: Robust contrastive learning against noisy views. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16670-16681 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 432, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 432, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 432, + 481, + 453 + ], + "type": "text", + "content": "9. Contributors, M.: MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 454, + 481, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 454, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 454, + 481, + 487 + ], + "type": "text", + "content": "0. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 487, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 481, + 521 + ], + "type": "text", + "content": "1. Dong, P., Li, L., Wei, Z., Niu, X., Tian, Z., Pan, H.: Emq: Evolving training-free proxies for automated mixed precision quantization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17076-17086 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 521, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 481, + 564 + ], + "type": "text", + "content": "2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 565, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 565, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 138, + 565, + 481, + 598 + ], + "type": "text", + "content": "3. Fei, W., Dai, W., Li, C., Zou, J., Xiong, H.: General bitwidth assignment for efficient deep convolutional neural network quantization. IEEE Transactions on Neural Networks and Learning Systems 33(10), 5253-5267 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 481, + 632 + ], + "type": "text", + "content": "4. Frumkin, N., Gope, D., Marculescu, D.: Jumping through local minima: Quantization in the loss landscape of vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16978-16988 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "5. Fu, Y., Yu, Q., Li, M., Ouyang, X., Chandra, V., Lin, Y.: Contrastive quant: quantization makes stronger contrastive learning. In: Proceedings of the 59th ACM/IEEE Design Automation Conference. pp. 205-210 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLAMP-ViT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "16. Huang, H., Yu, P.S., Wang, C.: An introduction to image synthesis with generative adversarial nets. arXiv preprint arXiv:1803.04469 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "text", + "content": "17. Hubara, I., Courbariaux, M., Soudry, D., El-Yaniv, R., Bengio, Y.: Binarized neural networks. Advances in neural information processing systems 29 (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "type": "text", + "content": "18. Hubara, I., Nahshan, Y., Hanani, Y., Banner, R., Soudry, D.: Accurate post training quantization with small calibration sets. In: International Conference on Machine Learning. pp. 4466-4475. PMLR (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 226 + ], + "type": "text", + "content": "19. Kim, N., Shin, D., Choi, W., Kim, G., Park, J.: Exploiting retraining-based mixed-precision quantization for low-cost dnn accelerator design. IEEE Transactions on Neural Networks and Learning Systems 32(7), 2925–2938 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "20. Kundu, S., Sun, Q., Fu, Y., Pedram, M., Beerel, P.: Analyzing the confidentiality of undistillable teachers in knowledge distillation. Advances in Neural Information Processing Systems 34, 9181-9192 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 303 + ], + "type": "text", + "content": "21. Kundu, S., Wang, S., Sun, Q., Beerel, P.A., Pedram, M.: Bmpq: bit-gradient sensitivity-driven mixed-precision quantization of dnns from scratch. In: 2022 Design, Automation & Test in Europe Conference & Exhibition (DATE). pp. 588-591. IEEE (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "22. Li, H., Xu, Z., Taylor, G., Studer, C., Goldstein, T.: Visualizing the loss landscape of neural nets. Advances in neural information processing systems 31 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "text", + "content": "23. Li, Y., Xu, S., Zhang, B., Cao, X., Gao, P., Guo, G.: Q-vit: Accurate and fully quantized low-bit vision transformer. Advances in Neural Information Processing Systems 35, 34451-34463 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "text", + "content": "24. Li, Y., Yuan, G., Wen, Y., Hu, J., Evangelidis, G., Tulyakov, S., Wang, Y., Ren, J.: Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems 35, 12934-12949 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 391, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 423 + ], + "type": "text", + "content": "25. Li, Z., Chen, M., Xiao, J., Gu, Q.: Psaq-vit v2: Toward accurate and general data-free quantization for vision transformers. IEEE Transactions on Neural Networks and Learning Systems (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "26. Li, Z., Gu, Q.: I-vit: integer-only quantization for efficient vision transformer inference. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17065-17075 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "text", + "content": "27. Li, Z., Ma, L., Chen, M., Xiao, J., Gu, Q.: Patch similarity aware data-free quantization for vision transformers. In: European Conference on Computer Vision. pp. 154-170. Springer (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "text", + "content": "28. Li, Z., Xiao, J., Yang, L., Gu, Q.: Repq-vit: Scale reparameterization for posttraining quantization of vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17227-17236 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 523, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 566 + ], + "type": "text", + "content": "29. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "text", + "content": "30. Lin, Y., Zhang, T., Sun, P., Li, Z., Zhou, S.: Fq-vit: Post-training quantization for fully quantized vision transformer. arXiv preprint arXiv:2111.13824 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 589, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 632 + ], + "type": "text", + "content": "31. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "32. Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., Gao, W.: Post-training quantization for vision transformer. Advances in Neural Information Processing Systems 34, 28092-28103 (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 411, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 411, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 411, + 101 + ], + "type": "text", + "content": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 139 + ], + "type": "text", + "content": "33. Peters, J.W., Welling, M.: Probabilistic binary neural networks. arXiv preprint arXiv:1809.03368 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 482, + 183 + ], + "type": "text", + "content": "34. Ramachandran, A., Dhiman, A., Vandrotti, B.S., Kim, J.: Ntrans-net: A multiscale neutrosophic-uncertainty guided transformer network for indoor depth completion. In: 2023 IEEE International Conference on Image Processing (ICIP). pp. 905–909. IEEE (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 217 + ], + "type": "text", + "content": "35. Ramachandran, A., Wan, Z., Jeong, G., Gustafson, J., Krishna, T.: Algorithm-hardware co-design of distribution-aware logarithmic-posit encodings for efficient dnn inference. arXiv preprint arXiv:2403.05465 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "text", + "content": "36. Ranjan, N., Savakis, A.: Lrp-qvit: Mixed-precision vision transformer quantization via layer-wise relevance propagation. arXiv preprint arXiv:2401.11243 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "text", + "content": "37. Strudel, R., Garcia, R., Laptev, I., Schmid, C.: Segmenter: Transformer for semantic segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7262-7272 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 308 + ], + "type": "text", + "content": "38. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: International conference on machine learning. pp. 10347-10357. PMLR (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 329 + ], + "type": "text", + "content": "39. Wang, J., Li, J., Li, W., Xuan, L., Zhang, T., Wang, W.: Positive-negative equal contrastive loss for semantic segmentation. Neurocomputing 535, 13-24 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 331, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 481, + 352 + ], + "type": "text", + "content": "40. Wightman, R.: Pytorch image models. https://github.com/rwrightman/pytorch-image-models (2019). https://doi.org/10.5281/zenodo.4414861" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 353, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 481, + 374 + ], + "type": "text", + "content": "41. Xiao, J., Li, Z., Yang, L., Gu, Q.: Patch-wise mixed-precision quantization of vision transformer. arXiv preprint arXiv:2305.06559 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 376, + 481, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 481, + 408 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 481, + 408 + ], + "type": "text", + "content": "42. Xiao, T., Liu, Y., Zhou, B., Jiang, Y., Sun, J.: Unified perceptual parsing for scene understanding. In: Proceedings of the European conference on computer vision (ECCV). pp. 418-434 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 409, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 481, + 441 + ], + "type": "text", + "content": "43. Yeh, C.H., Hong, C.Y., Hsu, Y.C., Liu, T.L., Chen, Y., LeCun, Y.: Decoupled contrastive learning. In: European Conference on Computer Vision. pp. 668-684. Springer (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 442, + 481, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 486 + ], + "type": "text", + "content": "44. Yin, H., Molchanov, P., Alvarez, J.M., Li, Z., Mallya, A., Hoiem, D., Jha, N.K., Kautz, J.: Dreaming to distill: Data-free knowledge transfer via deepinversion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8715-8724 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 487, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 481, + 520 + ], + "type": "text", + "content": "45. Yuan, Z., Xue, C., Chen, Y., Wu, Q., Sun, G.: Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization. In: European Conference on Computer Vision. pp. 191-207. Springer (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 481, + 553 + ], + "type": "text", + "content": "46. Zhang, D., Yang, J., Ye, D., Hua, G.: Lq-nets: Learned quantization for highly accurate and compact deep neural networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 365-382 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 554, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 481, + 576 + ], + "type": "text", + "content": "47. Zhang, S., Zhou, Q., Wang, Z., Wang, F., Yan, J.: Patch-level contrastive learning via positional query for visual pre-training (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 577, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 481, + 620 + ], + "type": "text", + "content": "48. Zhang, Y., Guo, X., Poggi, M., Zhu, Z., Huang, G., Mattoccia, S.: Completionformer: Depth completion with convolutions and vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18527-18536 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "49. Zhang, Y., Chen, D., Kundu, S., Li, C., Beerel, P.A.: Sal-vit: Towards latency efficient private inference on vit using selective attention search with a learnable softmax approximation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5116-5125 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 392, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLAMP-ViT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 226 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "50. Zhang, Z., Lu, X., Cao, G., Yang, Y., Jiao, L., Liu, F.: Vit-yolo: Transformer-based yolo for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2799-2808 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 194 + ], + "type": "text", + "content": "51. Zhong, Y., Lin, M., Nan, G., Liu, J., Zhang, B., Tian, Y., Ji, R.: Intraq: Learning synthetic images with intra-class heterogeneity for zero-shot network quantization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12339-12348 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "type": "text", + "content": "52. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127, 302-321 (2019)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 411, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 411, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 411, + 102 + ], + "type": "text", + "content": "Akshat Ramachandran, Souvik Kundu, and Tushar Krishna" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CLAP_ Isolating Content from Style through Contrastive Learning with Augmented Prompts/f676b4a9-3170-48e9-b0be-002172d2221e_content_list.json b/2024/CLAP_ Isolating Content from Style through Contrastive Learning with Augmented Prompts/f676b4a9-3170-48e9-b0be-002172d2221e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3155f732fed30eb3f1ca94c2153d2577af85cae3 --- /dev/null +++ b/2024/CLAP_ Isolating Content from Style through Contrastive Learning with Augmented Prompts/f676b4a9-3170-48e9-b0be-002172d2221e_content_list.json @@ -0,0 +1,1669 @@ +[ + { + "type": "text", + "text": "CLAP: Isolating Content from Style through Contrastive Learning with Augmented Prompts", + "text_level": 1, + "bbox": [ + 228, + 140, + 774, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yichao Cai, Yuhang Liu, Zhen Zhang, and Javen Qinfeng Shi", + "bbox": [ + 256, + 212, + 746, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Australian Institute for Machine Learning, University of Adelaide, SA 5000, Australia {yichao.cai,yuhang.liu01,zhen.zhang02,javen.shi}@adelaide.edu.au", + "bbox": [ + 217, + 239, + 785, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Contrastive vision-language models, such as CLIP, have garnered considerable attention for various downstream tasks, mainly due to the remarkable ability of the learned features for generalization. However, the features they learned often blend content and style information, which somewhat limits their generalization capabilities under distribution shifts. To address this limitation, we adopt a causal generative perspective for multimodal data and propose contrastive learning with data augmentation to disentangle content features from the original representations. To achieve this, we begin with exploring image augmentation techniques and develop a method to seamlessly integrate them into pretrained CLIP-like models to extract pure content features. Taking a step further, recognizing the inherent semantic richness and logical structure of text data, we explore the use of text augmentation to isolate latent content from style features. This enables CLIP-like model's encoders to concentrate on latent content information, refining the learned representations by pre-trained CLIP-like models. Our extensive experiments across diverse datasets demonstrate significant improvements in zero-shot and few-shot classification tasks, alongside enhanced robustness to various perturbations. These results underscore the effectiveness of our proposed methods in refining vision-language representations and advancing the state-of-the-art in multimodal learning.", + "bbox": [ + 261, + 304, + 743, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Data Augmentation $\\cdot$ Latent Variables $\\cdot$ Disentanglement", + "bbox": [ + 261, + 609, + 732, + 623 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 648, + 375, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-language models, exemplified by CLIP [36], have garnered substantial attention due to their exceptional generalization capabilities, achieved through the learned features, obtained by utilizing a cross-modal contrastive loss [20,25,36]. However, despite being pre-trained on extensive datasets, CLIP-like models fall short in disentangling latent content information and latent style information. Consequently, they are not immune to spurious correlations, i.e., style-related information is erroneously utilized to predict task-related labels. These limitations become evident in the presence of distribution shifts or adversarial attacks where spurious correlations often change across different environments. For examples,", + "bbox": [ + 212, + 680, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1 Our code is available at https://github.com/YichaoCai1/CLAP.", + "bbox": [ + 217, + 824, + 663, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6e66cdac6ce80f08ff08ab337f4f8aa2c98ad0f5669134c6c15d54971900f099.jpg", + "image_caption": [ + "(a) Image augmentation" + ], + "image_footnote": [], + "bbox": [ + 228, + 145, + 370, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dde32168462b738ed9a57d9dc1dd9a0e30491f46efda143c8b1e5327deb7f068.jpg", + "image_caption": [ + "(c) CAM visualization examples" + ], + "image_footnote": [], + "bbox": [ + 377, + 150, + 784, + 303 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a5e203a1edbabb418ef483009400f31d58cf878cdd3d26c87a69176a88eb5137.jpg", + "image_caption": [ + "(b) Text augmentation", + "Fig. 1: Causal generative models of vision-language data. Image and text data are generated through distinct underlying deterministic processes, $\\mathbf{g}_{\\mathbf{x}}$ for images and $\\mathbf{g}_{\\mathbf{t}}$ for texts, derived from a unified latent space with latent content variables $\\mathbf{c}$ and latent style variables $\\mathbf{s}$ . Latent content $\\mathbf{c}$ exclusively determines the sample label $\\mathbf{y}$ . (a) Soft interventions on latent style variables $\\mathbf{s}$ result in $\\tilde{\\mathbf{s}}$ , subsequently generating augmented images $\\tilde{\\mathbf{x}}$ . (b) Due to the same latent space, soft interventions on latent style variables $\\mathbf{s}$ can also result in $\\tilde{\\mathbf{s}}$ , producing augmented text $\\tilde{\\mathbf{t}}$ . (c) A qualitative comparison of image features for zero-shot classification using \"a photo of a [class]\" prompts, visualized using class activation map (CAM) [32], demonstrates that while image augmentation can enhance CLIP features, the features obtained through text augmentation methods predominantly focus on the content." + ], + "image_footnote": [], + "bbox": [ + 227, + 234, + 370, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(1) a notable dependence on specific input text prompts has been reported for zero-shot capabilities [21, 47, 48]; (2) performance decline in few-shot scenarios has been observed in few-shot learning scenarios [13, 36]; and (3) susceptibility to adversarial attacks has been explored [33, 43, 45].", + "bbox": [ + 212, + 517, + 784, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Taking a causal perspective, this work begins with a simple yet effective method, image augmentation, to disentangle content and style information within the learned representations of CLIP-like models. This approach is inspired by recent advancements in theoretical development in causal representation learning [41], which demonstrate that augmented image can be interpreted as a result of soft interventions on latent style variables, as depicted in Fig. 1a. Such augmentation results in a natural data pair where content information remains unchanged while style information changes. Consequently, using contrastive learning, it becomes feasible to isolate the invariant content information from the variant style information. Motivated by this theoretical advancement, we propose a practical method to incorporate image augmentation into CLIP-like models to extract content information from the original learned features. Specifically, a disentangled network is designed to fine-tune the pre-trained CLIP model by using a contrastive loss with image augmentation.", + "bbox": [ + 212, + 580, + 787, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the advancements made in disentangling content and style information from the original features learned by CLIP-like models through image augmentation, we recognize an inherent limitation: it is generally challenging to", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Cai et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "design adequate image augmentations to ensure all style factors change in an image. Theoretically, disentangling content and style information necessitates changes in all style factors [41]. However, inducing sufficient changes in latent style through image augmentation poses challenges due to the high dimensionality and complexity of style information in image data. Achieving significant style variation via artificially designed image augmentation techniques, such as transforming a photograph of a dog into a sketch while preserving content but dramatically altering style, is notably difficult.", + "bbox": [ + 212, + 146, + 782, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Taking a further step, rather than relying on image augmentation, we explore the use of text augmentation to disentangle latent content and style factors. This shift is motivated by two key observations: 1) Vision and language data share the same latent space. Therefore, text augmentation can also be utilized to induce changes in latent style factors instead of image augmentation. 2) Text data inherently possesses high semanticity and logical structure, making it more amenable to property-wise manipulation compared to image data. Consequently, implementing sufficient style changes through text augmentation is more feasible than image augmentation, contributing to isolating content from style information, see Fig. 1c for visual comparison. For instance, transforming text from \"a photo of a dog\" to \"a sketch of a dog\" is straightforward in the language modality, whereas achieving a similar transformation in image data is challenging. Inspired by these observations, we posit that introducing style variations through text augmentation, as illustrated Fig. 1b, provides a more effective approach for learning vision-language content features than relying on image augmentation.", + "bbox": [ + 212, + 267, + 785, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our contributions include: (1) Aimed at disentangling latent content and style factors to refine vision-language features of pre-trained CLIP-like models, we propose constractive learning with data augmentation to fine tune the original features of pre-trained CLIP-like models from a causal perspective. (2) We present a novel method customized for pre-trained CLIP-like models. This method leverages a disentangled network, which is trained using contrastive learning with image augmentation, to extract latent content features from the learned features provided by image encoder of CLIP-like models. (3) We propose Contrastive Learning with Augmented Prompts (CLAP), to extract latent content features from representations of CLIP-like models. It begins by training a disentangled network using the pre-trained text encoder of CLIP-like models and text augmentation. Subsequently, the trained disentangled network is transferred to the image encoder of CLIP-like models. (4) Experiments conducted on a large real dataset demonstrate the effectiveness of the proposed image augmentation and text augmentation in terms of zero-shot and few-shot performance, as well as robustness against perturbations.", + "bbox": [ + 212, + 508, + 787, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 775, + 387, + 791 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrastive Vision-Language Models Using a cross-modal contrastive loss, CLIP [36] revolutionarily introduced a scalable contrastive vision-language model", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "CLAP: Contrastive Learning with Augmented Prompts", + "bbox": [ + 361, + 114, + 730, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "by leveraging a large corpus of internet-sourced image-text pairs, demonstrating unprecedented zero-shot learning capabilities and exceptional generalization ability across datasets and supporting numerous downstream tasks [38]. ALIGN [20] expanded the scale of contrastive vision-language modeling, training on up to one billion image-text pairs while integrating the vision transformer's self-attention mechanism [11], which further enhanced performance. Despite their successes, CLIP-like models exhibit sensitivity to input text prompts [21, 48], leading to variable performance across different prompts. Efforts to mitigate this prompt sensitivity through prompt learning and engineering [9, 14, 21, 47, 48] focus on customizing prompts for specific tasks but do not fundamentally enhance CLIP's representations. Furthermore, CLIP-like models are vulnerable to adversarial attacks [4, 12], with current strategies [33, 45] involving adversarial-natural image pairs to improve resilience. Our work diverges from task-specific approaches by aiming to enhance CLIP's representations from a disentanglement perspective, addressing the aforementioned issues inherent in CLIP-like models.", + "bbox": [ + 212, + 146, + 787, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Disentangled Representation Learning Aimed at segregating intrinsic latent factors in data into distinct, controllable representations, disentangled representation learning benefits various applications [24,40,44]. Specifically, in classification tasks, it's shown that enhancing the model's performance and robustness against data distribution perturbations can be achieved by more effectively disentangling invariant content variables, without needing to identify all intrinsic latent variables completely [22,26-28]. Within single modalities, studies such as [49] have illustrated that contrastive learning [7,16,18] can potentially reverse the data generative process, aiding in the separation of representations. Furthermore, [41] suggest that image augmentation can help isolate content variables from the latent space through significant stylistic changes. [19] employs mixture techniques for data augmentation, enabling more abundant cross-modal matches. Diverging from these methods, our approach focuses on employing text augmentation to disentangle latent content variables, introducing a unique approach to learn refined vision-language representations.", + "bbox": [ + 212, + 393, + 789, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 A Causal Generative Model for Multi-Modal Data", + "text_level": 1, + "bbox": [ + 214, + 643, + 741, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To understand pretrained CLIP-like models, we investigate the underlying causal generative process for vision-language data. We consider the following causal generative model as depicted in Fig. 1. In the proposed model, the shared latent space ruling vision and language data is divided into two distinct sub-spaces: one corresponding to the latent content variables $\\mathbf{c}$ and the other to the latent style variables $\\mathbf{s}$ . The latent content variables are posited to determine the object label $\\mathbf{y}$ , a relationship corroborated by prior studies [22, 29, 31]. Furthermore, to elucidate the correlation between the latent style variable $\\mathbf{s}$ and the object variable $\\mathbf{y}$ , our model incorporates the premise that the latent content variable $\\mathbf{c}$ causally influences the latent style variable $\\mathbf{s}$ , in concordance with the principles of causal representation learning highlighted in recent literature [10, 29, 41]. Additionally,", + "bbox": [ + 212, + 672, + 787, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Cai et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "considering the diversity between image data and text data, where information in image data is typically much more details while information in text data tends to be more logically structured nature, we posit distinct causal mechanisms for the generation processes. Our causal generative model is formulated as following structural causal models [2]:", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} := \\mathbf {g} _ {\\mathbf {s}} (\\mathbf {c}), \\mathbf {x} := \\mathbf {g} _ {\\mathbf {x}} (\\mathbf {c}, \\mathbf {s}), \\mathbf {t} := \\mathbf {g} _ {\\mathbf {t}} (\\mathbf {c}, \\mathbf {s}), \\mathbf {y} := \\mathbf {g} _ {\\mathbf {y}} (\\mathbf {c}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 229, + 785, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Eq. (1), the style variable $\\mathbf{s}$ is causally influenced by the content via $\\mathbf{g_s}$ ; $\\mathbf{x}$ and $\\mathbf{t}$ denote visual and textual data, respectively. Both visual and textual data are causally produced by the shared latent variables $\\mathbf{c}$ and $\\mathbf{s}$ through distinct, reversible generative processes: $\\mathbf{g_x}$ for images and $\\mathbf{g_t}$ for text data, respectively. The label $\\mathbf{y}$ of a sample is exclusively determined by the content variable $\\mathbf{c}$ via $\\mathbf{g_y}$ . For simplicity, exogenous noises are implicitly assumed but not explicitly represented in the causal generative model's formulation, aligning with the common understanding that each latent variable is influenced by exogenous noise.", + "bbox": [ + 212, + 253, + 787, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recent seminal work in [41] has demonstrated that the latent content variable $\\mathbf{c}$ can be identified up to block identifiability (i.e., $\\mathbf{c}$ can be isolated from style variable $\\mathbf{s}$ ), by requiring all latent style variables to change (e.g., soft interventions on all latent style variables). This change can be achieved through image augmentation, i.e., the augmented image $\\tilde{\\mathbf{x}}$ can be interpreted as a generative result of $\\tilde{\\mathbf{s}}$ , which is produced through soft interventions on original latent style variables $\\mathbf{s}$ . Despite such theoretical advancement, the practical implementation of this theoretical result within CLIP-like models remains unclear. In this study, we propose a practical method to disentangle content and style information within CLIP-like models by employing image augmentation, as detailed in Sec. 4.1. Moreover, we recognize that implementing sufficient changes on all latent style variables $\\mathbf{s}$ through text augmentation is more feasible than image augmentation, due to high semanticity and logical structure in text data, we delve into the use of text augmentation to separate content information from style information, as discussed in Sec. 4.2.", + "bbox": [ + 212, + 375, + 787, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Isolating Content from Style with Data Augmentation", + "text_level": 1, + "bbox": [ + 214, + 619, + 777, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we propose the employment of data augmentation to extract content information from the learned features in pre-trained CLIP-like models. Essentially, data augmentation facilitates the alteration of style factors while preserving content factors. Consequently, leveraging contrastive learning enables the segregation of content information from style information. We delve into two distinct forms of data augmentation, namely image augmentation (Sec. 4.1) and text augmentation (Sec. 4.2).", + "bbox": [ + 212, + 648, + 787, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Isolating Content from Style with Augmented Images", + "text_level": 1, + "bbox": [ + 214, + 773, + 705, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While recent studies (von et al., 2021) have offered assurance regarding the disentanglement of content and style through contrastive learning with data augmentation, it remains unclear how these theoretical findings can be applied to", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "CLAP: Contrastive Learning with Augmented Prompts", + "bbox": [ + 361, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d51d2e0024c09fbe32179e48985340220767f6226601d8137f87eed6017a0136.jpg", + "image_caption": [ + "(a) Image augmentation" + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 406, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/db85c17d17159a5651a4daf6640b7096bd24c40f6d9f8ce12b44872b3fa4b80f.jpg", + "image_caption": [ + "(b) Text augmentation", + "Fig. 2: Refining CLIP through data augmentation. (a) Training involves a disentangled network $\\mathbf{f_c}$ , utilizing contrastive loss on original and augmented image pairs $\\mathbf{x}$ and $\\tilde{\\mathbf{x}}$ , with CLIP's image encoder $\\mathbf{f_x^*}$ holding frozen gradients. (b) More efficient content feature learning is achieved through contrastive learning with augmented text prompts $\\mathbf{t}$ and $\\tilde{\\mathbf{t}}$ , using the fixed text encoder $\\mathbf{f_t^*}$ of CLIP. (c) Inference stage: The trained disentangled network $\\mathbf{f_c^*}$ integrates with CLIP's text and image encoders, $\\mathbf{f_t^*}$ and $\\mathbf{f_x^*}$ , to enable zero-shot inference for an input image $\\mathbf{x}$ and class names $\\mathbf{t_1}$ to $\\mathbf{t_n}$ ." + ], + "image_footnote": [], + "bbox": [ + 410, + 143, + 591, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/24c7f87d886d6e99a42e77d789d6bd111c1314f0329c5c597871cecd865e977f.jpg", + "image_caption": [ + "(c) Zero-shot inference" + ], + "image_footnote": [], + "bbox": [ + 596, + 143, + 779, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the realm of vision-language models. We convert the theoretical findings into CLIP-like models in the following. The theoretical findings suggest using InfoNCE loss [34] to extract content information, as outlined below:", + "bbox": [ + 212, + 376, + 787, + 422 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\mathbf {f}; \\{\\mathbf {x} _ {i}, \\tilde {\\mathbf {x}} _ {i} \\} _ {i = 1} ^ {b}, \\tau) = - \\frac {1}{b} \\sum_ {i = 1} ^ {b} \\log \\frac {\\exp [ \\langle \\mathbf {f} (\\mathbf {x} _ {i}) , \\mathbf {f} (\\tilde {\\mathbf {x}} _ {i}) \\rangle / \\tau ]}{\\sum_ {j = 1} ^ {b} \\exp [ \\langle \\mathbf {f} (\\mathbf {x} _ {i}) , \\mathbf {f} (\\tilde {\\mathbf {x}} _ {j}) \\rangle / \\tau ]}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 272, + 429, + 785, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\{\\mathbf{x}_i\\}_{i=1}^b$ represents a batch of $b$ samples from the training dataset, $\\mathbf{f}(\\mathbf{x}_i)$ denotes sample $\\mathbf{x}_i$ 's features through model $\\mathbf{f}$ , $\\tilde{\\mathbf{x}}_i$ is the augmented counterpart of $\\mathbf{x}_i$ , and $\\langle \\mathbf{z}_1, \\mathbf{z}_2 \\rangle$ represents the cosine similarity between two feature vectors, $\\mathbf{z}_1$ and $\\mathbf{z}_2$ , and $\\tau$ represents the temperature parameter influencing the loss.", + "bbox": [ + 212, + 474, + 784, + 535 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We extend it to refine pre-trained vision-language models, utilizing contrastive learning with augmented images (hereinafter referred to as \"Im.Aug\"). As illustrated in Fig. 2a, we train a disentangled network on top of CLIP's pretrained image encoder. To enhance training efficiency and the usability of the proposed method, we freeze the pre-trained image encoder. Based on an InfoNCE loss, the learning objective of Im.Aug is formulated as follows:", + "bbox": [ + 212, + 536, + 787, + 626 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {\\mathbf {c}} ^ {*} = \\underset {\\mathbf {f} _ {\\mathbf {c}}} {\\operatorname {a r g m i n}} \\underset {\\{\\mathbf {x} _ {i} \\} _ {i = 1} ^ {b} \\in \\mathcal {D} _ {\\mathbf {x}}} {\\mathbb {E}} \\mathcal {L} \\left(\\mathbf {f} _ {\\mathbf {c}} \\circ \\mathbf {f} _ {\\mathbf {x}} ^ {*}; \\left\\{\\mathbf {x} _ {i}, \\tilde {\\mathbf {x}} _ {i} \\right\\} _ {i = 1} ^ {b}, \\tau\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 633, + 785, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{D}_{\\mathbf{x}}$ denotes the training image dataset and $b$ represents the batch size, $\\mathbf{f_c}$ is the disentangled network undergoing training. The pre-trained CLIP image encoder is represented by $\\mathbf{f_x^*}$ , with the asterisk \"***\" signifying that the model weights remain fixed. The variable $\\mathbf{x}_i$ refers to an image sampled from $\\mathcal{D}_{\\mathbf{x}}$ , and $\\tilde{\\mathbf{x}}_i$ is its augmented view.", + "bbox": [ + 212, + 667, + 787, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The composition of the training dataset $\\mathcal{D}_{\\mathbf{x}}$ , the image augmentation techniques used, the structure of the disentangled network $\\mathbf{f_c}$ , and the utilization of $\\mathbf{f_c^*}$ post-training are detailed in the following subsections.", + "bbox": [ + 212, + 744, + 787, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data Synthesis and Image Augmentation To generate training image data, we combine class names with various image and object attributes to create text", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Cai et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7ce62a6755332c85eb9f5ce5e1dbaf24af7127ed305e060f329684fec451f01b.jpg", + "table_caption": [ + "Table 1: Template-based prompts. Attributes used to generate text prompts follow the structured format \"a [art style] [image type] of a [object size] [object color] [class]\", where \"[class]\" represents the class names." + ], + "table_footnote": [], + "table_body": "
Object ColorObject SizeImage TypeArt Style
yellow, green, black, blue, multicolored, orange, red, white, brown, purplelarge, small, normal sizedpainting, cartoon, infographic, sketch, photograph, clipart, mosaic art, sculpturerealistic, impressionistic
", + "bbox": [ + 248, + 198, + 754, + 263 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "prompts for each class. Using a stable diffusion model [39], we produce synthetic images that comprise our training dataset $\\mathcal{D}\\mathbf{x}$ . The creation of template prompts for stable diffusion is based on attributes such as object size, color, image type, and art style. As detailed in Tab. 1, the attributes include 10 colors and 3 sizes for objects, and 8 types and 2 art styles for images. By assembling these attributes into prompts like \"a [art style] [image type] of a [object size] [object color] [class]\", we generate 480 unique texts for each class, from which one image per prompt is synthesized. Further details on image synthesis and examples are available in Appendix B.1. For the image augmentation procedures, we adopt techniques commonly used in contrastive learning practice [7,8,41], specifically random cropping and color distortion.", + "bbox": [ + 212, + 291, + 787, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Disentangled Network Structure Since the training process is based on CLIP's pre-trained lower-dimensional features, our disentangled network adopts a multi-layer perceptron (MLP) architecture. To fully benefit from the pretrained CLIP text encoder, we construct a residual MLP featuring a zero-initialized projection, acting as the disentangled network, as depicted in Fig. 3. This design enables learning directly from the pre-trained representation space, avoiding a random starting point, inspired by ControlNet's zero-conv operation [46], which we adapt to a zero-linear operation within our residual MLP.", + "bbox": [ + 212, + 482, + 787, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Within this architecture, the main branch includes a zero-initialized, bias-free linear layer positioned subsequent to the combination of a SiLU activation and a normally initialized linear layer. Conventionally, the dimensions of features before the initial linear layer, situated between the first and second linear layers, and following the second linear layer, are named as the input $d_{in}$ , latent $d_{mid}$ , and output $d_{out}$ dimensions, respectively. To rectify any mismatches between the input and output dimensions, the network employs nearest-neighbor downsampling within the shortcut path, thereby ensuring both alignment and the preservation of sharpness for the input features. During the inference stage, a weighting parameter $\\alpha > 0$ is introduced to modulate the portion of features emanating from the main branch before their integration with the input features, whereas this parameter remains constant at 1 throughout the training phase.", + "bbox": [ + 212, + 604, + 787, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inference After training, the disentangled network $\\mathbf{f}_{\\mathbf{c}}^{*}$ is utilized following CLIP's image encoder to extract visual content features. Moreover, given that", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "CLAP: Contrastive Learning with Augmented Prompts", + "bbox": [ + 361, + 113, + 730, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/26bd3fc57531fc0688d8e6ba852c119d57c9e15fd1dd84bdc1a7e609acd3bb5e.jpg", + "image_caption": [ + "Fig. 3: Structure of the disentangled network. The architecture encompasses a residual block featuring a zero-initialized, bias-free linear layer to commence optimization from the input feature space. When the input and output dimension differ, a downsampling operation is utilized to achieve alignment. During inference, a scalar parameter $\\alpha$ balance the main branch and input features before combination." + ], + "image_footnote": [], + "bbox": [ + 341, + 143, + 663, + 246 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "vision-language data generation is rooted in a unified latent space, as depicted in Sec. 3, $\\mathbf{f}_{\\mathbf{c}}^{*}$ can be seamlessly integrated with CLIP's image and text encoders to enhance zero-shot capabilities. As shown in Fig. 2c, for an image $\\mathbf{x}$ , the operation is formulated as the composition function $\\mathbf{f}_{\\mathbf{c}}^{*} \\circ \\mathbf{f}_{\\mathbf{x}}^{*}(\\mathbf{x})$ , and similarly, for a text $\\mathbf{t}$ , as $\\mathbf{f}_{\\mathbf{c}}^{*} \\circ \\mathbf{f}_{\\mathbf{t}}^{*}(\\mathbf{t})$ . This integration preserves CLIP's zero-shot functionality while achieving refined features through the improved disentanglement of content.", + "bbox": [ + 212, + 356, + 787, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Isolating Content from Style with Augmented Prompts", + "text_level": 1, + "bbox": [ + 214, + 467, + 718, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Despite progress in disentangling content and style via image augmentation, adequately altering all style factors in an image remains challenging due to the high dimensionality and complexity of style information in images. Achieving substantial style changes through augmentation, essential for complete disentanglement [41], is difficult with existing image augmentation techniques. On the contrary, text data inherently possesses high semanticity and logical structure, making it more amenable to property-wise manipulation compared to image data. To further exploring the disentanglement of content, we propose Contrastive Learning with Augmented Prompts (CLAP).", + "bbox": [ + 212, + 491, + 787, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As depicted in Fig. 2b, CLAP employs an InfoNCE loss to train a disentangled network atop CLIP's pre-trained text encoder, keeping the encoder's gradients fixed, similar to Im.Aug. Leveraging the simpler structure of text, the template-based prompts previously utilized for synthesizing images now serve as the training text dataset, denoted by $\\mathcal{D}_{\\mathbf{t}}$ . Utilizing the same disentangled network as in Im.Aug, the learning objective of CLAP is outlined as follows:", + "bbox": [ + 212, + 627, + 787, + 718 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {\\mathbf {c}} ^ {*} = \\underset {\\mathbf {f} _ {\\mathbf {c}}} {\\operatorname {a r g m i n}} \\underset {\\left\\{\\mathbf {t} _ {i} \\right\\} _ {i = 1} ^ {b} \\in \\mathcal {D} _ {\\mathbf {t}}} {\\mathbb {E}} \\mathcal {L} \\left(\\mathbf {f} _ {\\mathbf {c}} \\circ \\mathbf {f} _ {\\mathbf {t}} ^ {*}; \\left\\{\\mathbf {t} _ {i}, \\tilde {\\mathbf {t}} _ {i} \\right\\} _ {i = 1} ^ {b}, \\tau\\right) + \\lambda \\mathcal {L} \\left(\\mathbf {f} _ {\\mathbf {c}} \\circ \\mathbf {f} _ {\\mathbf {t}} ^ {*}; \\left\\{\\mathbf {t} _ {i} ^ {c}, \\tilde {\\mathbf {t}} _ {i} \\right\\} _ {i = 1} ^ {b}, 1\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 227, + 726, + 785, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{f}_{\\mathbf{t}}^{*}$ denotes the pre-trained CLIP text encoder. The term $\\mathbf{t}_i$ references a text prompt from $\\mathcal{D}\\mathbf{t}$ , and $\\tilde{\\mathbf{t}}_i$ represents its augmented view, produced via prompt augmentation techniques. On the equation's right side, $\\mathbf{t}_i^c$ specifies the class name associated with the text prompt $\\mathbf{t}_i$ . This strategy aims to enhance variations between prompt pairs, especially in cases where the text dataset $\\mathcal{D}_{\\mathbf{t}}$", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Cai et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3c3fdca5fe64023cf19ba35b40d419860bcf3850227aed5dada119ea7a70f4c8.jpg", + "table_caption": [ + "Table 2: Prompt augmentation techniques. Various augmented views are generated from an original text prompt using specific augmentation techniques: OSD (Object Size Deletion), OCD (Object Color Deletion), ITD (Image Type Deletion), ASD (Art Style Deletion), and SPO (Swapping Prompt Order)." + ], + "table_footnote": [], + "table_body": "
OriginalOSDOCDITDASDSPO
a realistic painting of a large red cara realistic painting of a red cara realistic painting of a large cara realistic of a large red cara painting of a large red cara large red car in a realistic painting
", + "bbox": [ + 248, + 212, + 754, + 275 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "has a very limited number of samples. Here, $\\lambda$ serves for adjusting the second term's importance in the total loss function. All other symbols in Eq. (4) maintain their definitions as described earlier.", + "bbox": [ + 212, + 301, + 782, + 347 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "After training, the learned disentangled network is seamlessly integrated with both of CLIP's encoders to extract content representations, as depicted in Fig. 2c.", + "bbox": [ + 212, + 348, + 784, + 378 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Prompt Augmentation To ensure text prompts undergo stylistic changes without compromising their content, we have developed specific augmentation techniques for synthetic text prompts. Drawing inspiration from Easy Data Augmentation (EDA) techniques [42], we adapted the Random Deletion (RD) and Random Swap (RS) techniques from EDA, customizing them to suit our prompt structure. To avoid inadvertently altering the content by introducing new object names or changing the core idea of a text prompt, our augmentation methods do not include random word insertions or replacements. Our primary augmentation techniques are Object Size Deletion (OSD), Object Color Deletion (OCD), Image Type Deletion (ITD), Art Style Deletion (ASD), and Swapping Prompt Order (SPO), each applied with a certain probability, as detailed in Tab. 2. Additionally, for down-stream datasets with few categories, to rich the population of training samples, we use an additional augmentation, named IGN (Inserting Gaussian Noise). Following the initializing protocol of prompt learning methods [47,48], we insert a zero-mean Gaussian noise with 0.02 standard deviation with a noise length equals to 4, to the tokenized prompts.", + "bbox": [ + 212, + 402, + 787, + 645 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Intuitively, these prompt augmentation methods parallel random masking techniques used in image augmentation [6,17]. However, prompt augmentations are more effective and precise than their image counterparts. This effectiveness arises because prompt augmentations can specifically target and eliminate a particular style element without impacting the content, whereas image masking, operating at the pixel or patch level, might inadvertently damage content information or lead to insufficient style changes.", + "bbox": [ + 212, + 646, + 787, + 752 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 776, + 375, + 792 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conduct three primary experiments to assess our method: (1) zero-shot evaluation with diverse prompts to gauge zero-shot performance and its robustness", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "CLAP: Contrastive Learning with Augmented Prompts", + "bbox": [ + 361, + 114, + 730, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "to prompt perturbations; (2) linear probe tests on few-shot samples to evaluate the efficacy of the learned representations in few-shot settings; and (3) adversarial attack assessments on zero-shot and one-shot classifiers to determine their resistance to adversarial threats. We further conduct an ablative study on hyperparameters, explore the impact of different prompt augmentation combinations and various sources of training prompts on CLAP's performance, and replicate experiments across different CLIP model sizes.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 273, + 426, + 287 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation. Im.Aug and CLAP are implemented using the ViT-B/16 CLIP model and executed on an NVIDIA RTX 3090 GPU. To ensure reproducibility, the random seed for all stochastic processes is fixed at 2024. More information on implementation details is provided in Appendix A.1.", + "bbox": [ + 212, + 296, + 785, + 357 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. CLAP is assessed across four multi-domain datasets to examine its performance in varied environments: PACS [23] (4 domains, 7 categories), VLCS [1] (4 domains, 5 categories), OfficeHome [37] (4 domains, 65 categories), and DomainNet [35] (6 domains, 345 categories). For conciseness, we present average results across the domains for each dataset. Detailed experimental outcomes for each domain within these datasets are provided in Appendix A.4.", + "bbox": [ + 212, + 357, + 785, + 446 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Compute efficiency. CLAP demonstrates faster convergence and shorter training times compared to Im.Aug. For CLAP, training on the PACS and VLCS datasets is completed in roughly 11 minutes, OfficeHome in approximately 14 minutes, and DomainNet in about 47 minutes. In contrast, Im.Aug requires around 16 minutes for PACS and VLCS, 50 minutes for OfficeHome, and 3.3 hours for DomainNet. Both Im.Aug and CLAP maintain CLIP's inference efficiency due to the disentangled network's efficient two-layer MLP structure.", + "bbox": [ + 212, + 448, + 787, + 553 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Main Results", + "text_level": 1, + "bbox": [ + 215, + 574, + 372, + 588 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Zero-Shot Performance To assess zero-shot capabilities, CLAP undergoes evaluation using three specific fixed prompts: ZS(C), utilizing only the class name within \"[class]\"; ZS(PC), with the format \"a photo of a [class]\"; and ZS(CP), structured as \"a [class] in a photo\". To thoroughly examine zero-shot proficiency, a dynamic prompt, ZS(NC), formatted as \"[noise][class]\", is also used, where \"[noise]\" signifies the introduction of Gaussian noise characterized by a mean of 0 and a standard deviation of 0.02.", + "bbox": [ + 212, + 598, + 785, + 703 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As presented in Tab. 3, CLAP surpasses both CLIP and Im.Aug across all evaluated prompts for every dataset. Unlike the uniform enhancement in zero-shot performance CLAP achieves over CLIP, Im.Aug displays inconsistent results. A closer examination reveals CLAP's superiority over CLIP is especially significant for the dynamic ZS(NC) prompt. This demonstrates CLAP's effectiveness in significantly improving zero-shot performance compared to the original CLIP representations.", + "bbox": [ + 212, + 704, + 785, + 808 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In assessing the model's robustness to prompt perturbations, we examine the variances in zero-shot performance across different prompts by analyzing the", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Cai et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e50fe19e798722be31494a487a218550962a81b6dc1dcf2afb6d16c916ae9cd7.jpg", + "table_caption": [ + "Table 3: Zero-shot results across three distinct prompts: \"C\" for \"[class]\", \"CP\" for \"a [class] in a photo\", \"PC\" for \"a photo of a [class]\", and a dynamic prompt \"NC\" for \"[noise][class]\" showcase that CLAP consistently outperforms CLIP's zero-shot performance across all datasets, whereas image augmentation exhibits mixed outcomes." + ], + "table_footnote": [], + "table_body": "
PromptMethodZero-shot performance, avg. top-1 acc. (%) (↑)
PACSVLCSOff.HomeDom.NetOverall
ZS(C)CLIP95.776.479.857.877.4
Im.Aug96.579.577.051.576.1
CLAP97.282.681.058.779.9
ZS(CP)CLIP95.282.079.557.078.4
Im.Aug96.382.975.850.776.4
CLAP97.383.480.558.079.8
ZS(PC)CLIP96.182.482.557.779.7
Im.Aug96.583.078.651.677.4
CLAP97.283.483.059.080.6
ZS(NC)CLIP90.868.371.551.070.4
Im.Aug94.873.167.544.069.9
CLAP97.281.073.552.676.1
", + "bbox": [ + 310, + 212, + 694, + 422 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "range $(R)$ and standard deviation $(\\delta)$ of results derived from ZS(C), ZS(CP), and ZS(PC). Additionally, we investigate the decline $(\\varDelta_{(NC)})$ in performance from ZS(C) to ZS(NC) as a broad indicator of resilience to noised prompts.", + "bbox": [ + 212, + 449, + 782, + 494 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As presented in Tab. 4, CLAP significantly reduces the variance in zero-shot performance across various testing prompts, evidenced by markedly lower values of $\\delta$ and $R$ , and a less pronounced decrease in performance with a noised prompt, in contrast to Im.Aug and the baseline representations of CLIP. Although Im.Aug aids in reducing performance variance to some extent, its efficacy is notably inferior to that of CLAP. These findings highlight CLAP's enhanced robustness in maintaining consistent zero-shot performance across a diverse array of prompts.", + "bbox": [ + 212, + 496, + 787, + 617 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Few-Shot Performance We conduct evaluations of 1-shot, 4-shot, 8-shot, and 16-shot linear probes across each domain within the four datasets. As illustrated in Fig. 4, CLAP significantly outperforms both CLIP and Im.Aug in few-shot learning scenarios. Notably, in the 1-shot setting CLAP achieves performance improvements over the linear-probe CLIP model by margins of $+10\\%$ , $+3.5\\%$ , $+2.5\\%$ , and $+1.5\\%$ on the PACS, VLCS, OfficeHome, and DomainNet datasets, respectively. These improvements are especially significant in comparison to the gains observed with Im.Aug counterparts, underpinning CLAP's efficacy in few-shot scenarios. For detailed quantitative results, please refer to Appendix A.4.", + "bbox": [ + 212, + 637, + 787, + 773 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Adversarial Performance To assess adversarial robustness, zero-shot (ZS(C)) and one-shot classifiers are evaluated against prominent adversarial attack methods, such as FGSM [15], PGD [30], and CW [5], by generating adversarial sam", + "bbox": [ + 212, + 794, + 785, + 842 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "CLAP: Contrastive Learning with Augmented Prompts", + "bbox": [ + 361, + 113, + 730, + 130 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/fc590fac5daf7b2072e9461ed34e99a83d6854c831016282108eaf95b6fc9f0d.jpg", + "table_caption": [ + "Table 4: CLAP more effectively reduces zero-shot performance variance across prompts than image augmentation, with $R$ and $\\delta$ indicating the range and standard deviation for ZS(C), ZS(PC), and ZS(PC). The decrease $\\Delta_{(NC)}$ from ZS(C) to ZS(NC) highlights CLAP's enhanced robustness against prompt perturbations." + ], + "table_footnote": [], + "table_body": "
MetricMethodPerformance variance, avg. top-1 acc. (%) (↓)
PACSVLCSOff.HomeDom.NetOverall
RCLIP0.96.13.10.82.7
Im.Aug0.13.62.80.91.9
CLAP0.10.82.51.01.1
δCLIP0.42.81.40.41.2
Im.Aug0.11.71.20.40.8
CLAP0.00.41.10.40.5
Δ(NC)CLIP4.98.18.36.87.0
Im.Aug1.66.49.57.56.3
CLAP0.01.67.56.13.8
", + "bbox": [ + 318, + 212, + 684, + 378 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ef84eab710a70b77efa730ef360f20d3f62c4b3f3dfced28ba3bb86ea520a6a1.jpg", + "image_caption": [ + "Fig. 4: Few-shot linear probe comparisons of image-encoder features show that CLAP enhances CLIP's few-shot performance more effectively than Im.Aug. In the accompanying figure, \"ZS\" indicates the zero-shot performance using a \"[class]\" prompt." + ], + "image_footnote": [], + "bbox": [ + 222, + 391, + 367, + 479 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/83e8ef217dfe60fa8be48905b67747d5aa0757bec0b8f4ab5486c1f7f0431d96.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 391, + 501, + 478 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b93b693869ec8d910eceebeb10f6de19c4ffdbadeb5824ceb7156ea32d94d066.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 391, + 640, + 478 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/68af5ad4d6da7adb0c2a6088c6da0ff56934239eb540773eb0c03fd2fdab6acd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 391, + 779, + 478 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "plies for testing. For FGSM, 1 adversarial iteration is employed, whereas for PGD and CW, 20 iterations are used, all with an epsilon of 0.031. As indicated in Tab. 5, classifiers utilizing CLAP representations demonstrate superior resilience to these adversarial attacks compared to those based on CLIP representations. Across the four datasets, CLAP's zero-shot and 1-shot classifiers surpass CLIP by margins of $+7.6\\%$ and $+8.5\\%$ against FGSM, $+1.0\\%$ and $+11.7\\%$ against PGD-20, and $+1.1\\%$ and $+2.3\\%$ against CW-20, respectively. These figures notably exceed the performance improvements of $+4.4\\%$ and $+4.6\\%$ against FGSM, $+0.3\\%$ and $+6.2\\%$ against PGD-20, and $0\\%$ and $+1.3\\%$ against CW-20 achieved by Im.Aug. The result suggests that CLAP efficiently enhances robustness against adversarial attacks in both zero-shot and one-shot scenarios.", + "bbox": [ + 212, + 561, + 787, + 728 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 More Analysis", + "text_level": 1, + "bbox": [ + 214, + 752, + 382, + 767 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "t-SNE Visualization In our t-SNE visualizations, we examine the representations of CLIP, Im.Aug, and CLAP for all images within the Art Painting domain of the PACS dataset. Fig. 5 shows that CLAP's image representations display a marked inter-class separation and tighter intra-class clustering than those of", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Cai et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e90635379d70b989e96ffbb6c23cedc3058a96090de3b74e92a3fe030027d55f.jpg", + "table_caption": [ + "Table 5: Image augmentation and CLAP both enhance CLIP's zero-shot with the \"[class]\" prompt and 1-shot robustness against adversarial attacks, with CLAP showing greater improvements." + ], + "table_footnote": [], + "table_body": "
Setting MethodAvg. top-1 acc. (%) under adversarial attacks(↑)
FGSMPGD-20CW-20
PACSVLCSO.H.D.N.PACSVLCSO.H.D.N.PACSVLCSO.H.D.N.
ZS(C)CLIP86.865.657.922.529.12.010.110.727.41.57.47.6
Im.Aug88.069.655.137.931.32.110.49.029.41.77.05.8
CLAP88.771.958.544.230.83.210.611.229.82.38.18.0
1-shotCLIP66.745.234.322.534.816.05.611.318.90.74.53.2
Im.Aug79.447.137.123.555.216.18.512.523.20.95.13.4
CLAP89.652.237.123.973.421.27.412.527.01.15.03.5
", + "bbox": [ + 233, + 198, + 769, + 311 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/bde6823173755b0d7136ddf77d7cc069d97ecb91807e9627482ca5d65737a5f9.jpg", + "image_caption": [ + "(a) CLIP" + ], + "image_footnote": [], + "bbox": [ + 243, + 333, + 390, + 419 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/baeaa687e75002e296a8c7a2f3c3461f0b160ba5e229cf058e61fddf7816bd28.jpg", + "image_caption": [ + "(b) Im.Aug", + "Fig. 5: t-SNE visualizations of all images in the Art Painting of PACS dataset show CLAP outperforms the original CLIP and Im.Aug, with clearer inter-class distinctions and tighter intra-class clusters." + ], + "image_footnote": [], + "bbox": [ + 429, + 333, + 576, + 419 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b6856d7e0f9ca1e0d49d20f38781276f54f0eb1aab929c728b026ca9b8728e4f.jpg", + "image_caption": [ + "(c) CLAP" + ], + "image_footnote": [], + "bbox": [ + 616, + 333, + 764, + 419 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "CLIP and Im.Aug. This observation suggests that CLAP's representations are more closely tied to content information and less influenced by style information, in contrast to the other two.", + "bbox": [ + 212, + 526, + 787, + 571 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablations In Fig. 6, we assess the zero-shot capabilities of our model using two distinct prompts, ZS(C) and ZS(PC), on the VLCS dataset. This analysis forms part of an ablative study aimed at understanding the influence of various hyperparameters on model performance. Specifically, we examine: the dimensions of the latent layer within the MLP of the disentangled network, as illustrated in Fig. 6a; the temperature parameter $(\\tau)$ in the loss function, as depicted in Fig. 6b; and the weight coefficient $(\\alpha)$ during the inference stage, as shown in Fig. 6c. Our findings indicate that CLAP consistently enhances zero-shot performance across all tested configurations for both prompts, while also significantly reducing the gap between the performances elicited by each prompt. These results underscore the efficacy of CLAP in accommodating a wide range of hyper-parameters.", + "bbox": [ + 212, + 592, + 787, + 758 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Prompt Augmentation Combinations We explore diverse combinations of our tailored prompt augmentation methods and examine Easy Data Augmentation (EDA) techniques [42] on the VLCS dataset. Each tested technique showcases CLAP's enhancements over CLIP, with details available in Appendix A.2.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "CLAP: Contrastive Learning with Augmented Prompts", + "bbox": [ + 361, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/63de70ec9fd2118f1ca643f09b4a9c745e0844cabf7c0ffc7c53fd7b175694b7.jpg", + "image_caption": [ + "Fig. 6: We conduct ablative study on hyper-parameter choices on the VLCS dataset, including latent dimensions, $\\tau$ values, and $\\alpha$ values during the inference stage. CLAP continuously enhance CLIP's performance throughout the tested values." + ], + "image_footnote": [], + "bbox": [ + 222, + 150, + 405, + 255 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0803b3e90ceebc3f158b3216f06cf6745df399def7712334b047c53718ff8c82.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 148, + 591, + 255 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5ffbb7fdae4d2b46c073f4e7d612fbd06af3fdce918f9a750b01810b40981b46.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 148, + 781, + 255 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Prompt Sources We assess the impact of different training prompt formats, originating from various synthetic sources, on the performance of the VLCS dataset, incorporating EDA techniques. Our evaluation includes our template-based prompts, LLM-generated prompts by ChatGPT-3.5 [3] (with the generation process detailed in Appendix B.2), prompts structured as \"a [random] style of [class],\" where \"[random]\" is filled with terms from a random word generator $^2$ , and prompts produced using the PromptStyler method [9]. The findings indicate that the training prompts with simpler forms tend to yield better performance, with detailed quantitative results presented in Appendix A.3.", + "bbox": [ + 212, + 335, + 787, + 472 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Experiments on Different Model Scales In our repeated experiments assessing zero-shot performance on the ViT-L/14 and ResNet50x16 pre-trained with CLIP, we consistently find that CLAP improves zero-shot performance while also reducing performance variances. This consistent observation underscores CLAP's effectiveness in enhancing the quality of CLIP representations. For quantitative details supporting these findings, please see the Appendix C.", + "bbox": [ + 212, + 494, + 787, + 585 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 607, + 359, + 623 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "To enhance pre-trained CLIP-like models, this study delves into disentangling latent content variables. Through a causal analysis of the underlying generative processes of vision-language data, we discover that training a disentangled network in one modality can effectively disentangle content across both modalities. Given the high semantic nature of text data, we identify that disentanglement is more achievable within the language modality through text augmentation interventions. Building on these insights, we introduce CLAP (Contrastive Learning with Augmented Prompts) to acquire disentangled vision-language content features. Comprehensive experiments validate CLAP's effectiveness, demonstrating significant improvements in zero-shot and few-shot performance, and enhancing robustness against perturbations. We anticipate that our work will inspire further exploration into disentangling latent variables within vision-language models.", + "bbox": [ + 212, + 638, + 787, + 820 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Cai et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "page_footnote", + "text": "2 https://github.com/vaibhavsingh97/random-word", + "bbox": [ + 217, + 824, + 584, + 839 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Albuquerque, I., Monteiro, J., Darvishi, M., Falk, T.H., Mitliagkas, I.: Generalizing to unseen domains via distribution matching. arXiv preprint arXiv:1911.00804 (2019). https://doi.org/10.48550/arXiv.1911.00804", + "2. Bollen, K.A.: Structural equations with latent variables, vol. 210. John Wiley & Sons (1989). https://doi.org/10.1002/9781118619179", + "3. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020), https://dl.acm.org/doi/abs/10.5555/3495724.3495883", + "4. Carlini, N., Terzis, A.: Poisoning and backdooring contrastive learning. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=iC4UHbQ01Mp", + "5. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: 2017 IEEE Symposium on Security and Privacy (SP). pp. 39-57. IEEE Computer Society (2017). https://doi.org/10.1109/SP.2017.49", + "6. Chen, P., Liu, S., Zhao, H., Jia, J.: Gridmask data augmentation. arXiv preprint arXiv:2001.04086 (2020). https://doi.org/10.48550/arXiv.2001.04086", + "7. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020), https://dl.acm.org/doi/abs/10.5555/3524938.3525087", + "8. Chen, X., He, K.: Exploring simple siamese representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15750-15758 (2021). https://doi.org/10.1109/CVPR46437.2021.01549", + "9. Cho, J., Nam, G., Kim, S., Yang, H., Kwak, S.: Promptstyler: Prompt-driven style generation for source-free domain generalization. In: Proceedings of the IEEE international conference on computer vision. pp. 15702-15712 (2023). https://doi.org/10.1109/ICCV51070.2023.01439", + "0. Daunhawer, I., Bizeul, A., Palumbo, E., Marx, A., Vogt, J.E.: Identifiability results for multimodal contrastive learning. ICLR (2023), https://openreview.net/forum?id=U_2kuqoTcB", + "11. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020), https://openreview.net/forum?id=YicbFdNTTy", + "12. Fort, S.: Adversarial vulnerability of powerful near out-of-distribution detection. arXiv preprint arXiv:2201.07012 (2022). https://doi.org/10.48550/arXiv.2201.07012", + "13. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision 132(2), 581-595 (2024). https://doi.org/10.1007/s11263-023-01891-x", + "14. Ge, C., Huang, R., Xie, M., Lai, Z., Song, S., Li, S., Huang, G.: Domain adaptation via prompt learning. arXiv preprint arXiv:2202.06687 (2022). https://doi.org/10.48550/arXiv.2202.06687", + "15. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. In: International Conference on Learning Representations (2015). https://doi.org/10.48550/arXiv.1412.6572" + ], + "bbox": [ + 218, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "CLAP: Contrastive Learning with Augmented Prompts", + "bbox": [ + 361, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020), https://dl.acm.org/doi/abs/10.5555/3495724.3497510", + "17. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022). https://doi.org/10.1109/CVPR52688.2022.01553", + "18. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020). https://doi.org/10.1109/CVPR42600.2020.00975", + "19. Hong, T., Guo, X., Ma, J.: Itmix: Image-text mix augmentation for transferring clip to image classification. In: 2022 16th IEEE International Conference on Signal Processing (ICSP). vol. 1, pp. 129-133. IEEE (2022). https://doi.org/10.1109/ICSP56322.2022.9965292", + "20. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021), https://proceedings.mlr.press/v139/jia21b.html", + "21. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19113-19122 (2023). https://doi.org/10.1109/CVPR52729.2023.01832", + "22. Kong, L., Xie, S., Yao, W., Zheng, Y., Chen, G., Stojanov, P., Akinwande, V., Zhang, K.: Partial disentanglement for domain adaptation. In: International Conference on Machine Learning. pp. 11455-11472. PMLR (2022), https://proceedings.mlr.press/v162/kong22a.htm1", + "23. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.M.: Deeper, broader and artier domain generalization. In: Proceedings of the IEEE international conference on computer vision. pp. 5542-5550 (2017). https://doi.org/10.1109/ICCV.2017.591", + "24. Li, H., Wang, X., Zhang, Z., Yuan, Z., Li, H., Zhu, W.: Disentangled contrastive learning on graphs. Advances in Neural Information Processing Systems 34, 21872-21884 (2021), https://dl.acm.org/doi/10.5555/3540261.3541935", + "25. Li, Y., Liang, F., Zhao, L., Cui, Y., Ouyang, W., Shao, J., Yu, F., Yan, J.: Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In: International Conference on Learning Representations (2021)", + "26. Liu, Y., Zhang, Z., Gong, D., Gong, M., Huang, B., van den Hengel, A., Zhang, K., Shi, J.Q.: Identifiable latent polynomial causal models through the lens of change. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=ia9fK01Vjq", + "27. Liu, Y., Zhang, Z., Gong, D., Gong, M., Huang, B., Hengel, A.v.d., Zhang, K., Shi, J.Q.: Identifying weight-variant latent causal models. arXiv preprint arXiv:2208.14153 (2022). https://doi.org/10.48550/arXiv.2208.14153", + "28. Liu, Y., Zhang, Z., Gong, D., Gong, M., Huang, B., Hengel, A.v.d., Zhang, K., Shi, J.Q.: Identifiable latent neural causal models. arXiv preprint arXiv:2403.15711 (2024). https://doi.org/10.48550/arXiv.2403.15711", + "29. Liu, Y., Zhang, Z., Gong, D., Gong, M., Huang, B., Zhang, K., Shi, J.Q.: Identifying latent causal content for multi-source domain adaptation. arXiv preprint arXiv:2208.14161 (2022). https://doi.org/10.48550/arXiv.2208.14161" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Cai et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. In: International Conference on Learning Representations (2018), https://openreview.net/forum?id=rJzIBfZAb", + "31. Mahajan, D., Tople, S., Sharma, A.: Domain generalization using causal matching. In: International Conference on Machine Learning. pp. 7313-7324. PMLR (2021), https://proceedings.mlr.press/v139/mahajan21b.html", + "32. Mamooler, S.: Clip explainability. https://github.com/sMamooler/CLIP_Explainability, accessed: 2024-03-06", + "33. Mao, C., Geng, S., Yang, J., Wang, X., Vondrick, C.: Understanding zero-shot adversarial robustness for large-scale models. In: The Eleventh International Conference on Learning Representations (2022), https://openreview.net/forum?id=P4bXcawRi5J", + "34. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018). https://doi.org/10.48550/arXiv.1807.03748", + "35. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE international conference on computer vision. pp. 1406-1415 (2019). https://doi.org/10.1109/ICCV.2019.00149", + "36. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021), https://proceedings.mlr.press/v139/radford21a.html", + "37. Rahman, M.M., Fookes, C., Baktashmotlagh, M., Sridharan, S.: Multi-component image translation for deep domain generalization. In: 2019 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 579-588. IEEE (2019). https://doi.org/10.1109/WACV.2019.00067", + "38. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021), https://proceedings.mlrpress/v139/ramesh21a.html", + "39. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022). https://doi.org/10.1109/CVPR52729.2023.01389", + "40. Sanchez, E.H., Serrurier, M., Ortner, M.: Learning disentangled representations via mutual information estimation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 205-221. Springer (2020). https://doi.org/10.1007/978-3-030-58542-6_13", + "41. Von Kügelgen, J., Sharma, Y., Gresele, L., Brendel, W., Schölkopf, B., Besserve, M., Locatello, F.: Self-supervised learning with data augmentations provably isolates content from style. Advances in neural information processing systems 34, 16451-16467 (2021), https://d1.acm.org/doi/10.5555/3540261.3541519", + "42. Wei, J., Zou, K.: Eda: Easy data augmentation techniques for boosting performance on text classification tasks. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). pp. 6382-6388 (2019). https://doi.org/10.18653/v1/D19-1670" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "CLAP: Contrastive Learning with Augmented Prompts", + "bbox": [ + 361, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Wortsman, M., Ilharco, G., Kim, J.W., Li, M., Kornblith, S., Roelofs, R., Lopes, R.G., Hajishirzi, H., Farhadi, A., Namkoong, H., et al.: Robust fine-tuning of zero-shot models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7959-7971 (2022). https://doi.org/10.1109/CVPR52688.2022.00780", + "44. Yang, M., Liu, F., Chen, Z., Shen, X., Hao, J., Wang, J.: Causalvae: Disentangled representation learning via neural structural causal models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9593-9602 (2021). https://doi.org/10.1109/CVPR46437.2021.00947", + "45. Yang, W., Mirzasoleiman, B.: Robust contrastive language-image pretraining against adversarial attacks. arXiv preprint arXiv:2303.06854 (2023). https://doi.org/10.48550/arXiv.2303.06854", + "46. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: IEEE International Conference on Computer Vision (ICCV) (2023). https://doi.org/10.1109/ICCV51070.2023.00355", + "47. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022). https://doi.org/10.1109/CVPR52688.2022.01631", + "48. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022). https://doi.org/10.1007/s11263-022-01653-1", + "49. Zimmermann, R.S., Sharma, Y., Schneider, S., Bethge, M., Brendel, W.: Contrastive learning inverts the data generating process. In: International Conference on Machine Learning. pp. 12979-12990. PMLR (2021), https://proceedings.mlr.press/v139/zimmermann21a.html" + ], + "bbox": [ + 215, + 146, + 785, + 506 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Cai et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/CLAP_ Isolating Content from Style through Contrastive Learning with Augmented Prompts/f676b4a9-3170-48e9-b0be-002172d2221e_model.json b/2024/CLAP_ Isolating Content from Style through Contrastive Learning with Augmented Prompts/f676b4a9-3170-48e9-b0be-002172d2221e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..46b2064bc6f83afd139495685c4976194b76ab41 --- /dev/null +++ b/2024/CLAP_ Isolating Content from Style through Contrastive Learning with Augmented Prompts/f676b4a9-3170-48e9-b0be-002172d2221e_model.json @@ -0,0 +1,2304 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.229, + 0.141, + 0.776, + 0.187 + ], + "angle": 0, + "content": "CLAP: Isolating Content from Style through Contrastive Learning with Augmented Prompts" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.213, + 0.747, + 0.229 + ], + "angle": 0, + "content": "Yichao Cai, Yuhang Liu, Zhen Zhang, and Javen Qinfeng Shi" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.24, + 0.787, + 0.27 + ], + "angle": 0, + "content": "Australian Institute for Machine Learning, University of Adelaide, SA 5000, Australia {yichao.cai,yuhang.liu01,zhen.zhang02,javen.shi}@adelaide.edu.au" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.305, + 0.744, + 0.597 + ], + "angle": 0, + "content": "Abstract. Contrastive vision-language models, such as CLIP, have garnered considerable attention for various downstream tasks, mainly due to the remarkable ability of the learned features for generalization. However, the features they learned often blend content and style information, which somewhat limits their generalization capabilities under distribution shifts. To address this limitation, we adopt a causal generative perspective for multimodal data and propose contrastive learning with data augmentation to disentangle content features from the original representations. To achieve this, we begin with exploring image augmentation techniques and develop a method to seamlessly integrate them into pretrained CLIP-like models to extract pure content features. Taking a step further, recognizing the inherent semantic richness and logical structure of text data, we explore the use of text augmentation to isolate latent content from style features. This enables CLIP-like model's encoders to concentrate on latent content information, refining the learned representations by pre-trained CLIP-like models. Our extensive experiments across diverse datasets demonstrate significant improvements in zero-shot and few-shot classification tasks, alongside enhanced robustness to various perturbations. These results underscore the effectiveness of our proposed methods in refining vision-language representations and advancing the state-of-the-art in multimodal learning." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.61, + 0.733, + 0.624 + ], + "angle": 0, + "content": "Keywords: Data Augmentation \\(\\cdot\\) Latent Variables \\(\\cdot\\) Disentanglement" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.65, + 0.377, + 0.667 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.681, + 0.788, + 0.818 + ], + "angle": 0, + "content": "Vision-language models, exemplified by CLIP [36], have garnered substantial attention due to their exceptional generalization capabilities, achieved through the learned features, obtained by utilizing a cross-modal contrastive loss [20,25,36]. However, despite being pre-trained on extensive datasets, CLIP-like models fall short in disentangling latent content information and latent style information. Consequently, they are not immune to spurious correlations, i.e., style-related information is erroneously utilized to predict task-related labels. These limitations become evident in the presence of distribution shifts or adversarial attacks where spurious correlations often change across different environments. For examples," + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.665, + 0.84 + ], + "angle": 0, + "content": "1 Our code is available at https://github.com/YichaoCai1/CLAP." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Cai et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.146, + 0.371, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.218, + 0.355, + 0.229 + ], + "angle": 0, + "content": "(a) Image augmentation" + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.151, + 0.785, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.307, + 0.671, + 0.318 + ], + "angle": 0, + "content": "(c) CAM visualization examples" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.236, + 0.371, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.307, + 0.351, + 0.318 + ], + "angle": 0, + "content": "(b) Text augmentation" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.331, + 0.788, + 0.484 + ], + "angle": 0, + "content": "Fig. 1: Causal generative models of vision-language data. Image and text data are generated through distinct underlying deterministic processes, \\(\\mathbf{g}_{\\mathbf{x}}\\) for images and \\(\\mathbf{g}_{\\mathbf{t}}\\) for texts, derived from a unified latent space with latent content variables \\(\\mathbf{c}\\) and latent style variables \\(\\mathbf{s}\\). Latent content \\(\\mathbf{c}\\) exclusively determines the sample label \\(\\mathbf{y}\\). (a) Soft interventions on latent style variables \\(\\mathbf{s}\\) result in \\(\\tilde{\\mathbf{s}}\\), subsequently generating augmented images \\(\\tilde{\\mathbf{x}}\\). (b) Due to the same latent space, soft interventions on latent style variables \\(\\mathbf{s}\\) can also result in \\(\\tilde{\\mathbf{s}}\\), producing augmented text \\(\\tilde{\\mathbf{t}}\\). (c) A qualitative comparison of image features for zero-shot classification using \"a photo of a [class]\" prompts, visualized using class activation map (CAM) [32], demonstrates that while image augmentation can enhance CLIP features, the features obtained through text augmentation methods predominantly focus on the content." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.785, + 0.579 + ], + "angle": 0, + "content": "(1) a notable dependence on specific input text prompts has been reported for zero-shot capabilities [21, 47, 48]; (2) performance decline in few-shot scenarios has been observed in few-shot learning scenarios [13, 36]; and (3) susceptibility to adversarial attacks has been explored [33, 43, 45]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.788, + 0.793 + ], + "angle": 0, + "content": "Taking a causal perspective, this work begins with a simple yet effective method, image augmentation, to disentangle content and style information within the learned representations of CLIP-like models. This approach is inspired by recent advancements in theoretical development in causal representation learning [41], which demonstrate that augmented image can be interpreted as a result of soft interventions on latent style variables, as depicted in Fig. 1a. Such augmentation results in a natural data pair where content information remains unchanged while style information changes. Consequently, using contrastive learning, it becomes feasible to isolate the invariant content information from the variant style information. Motivated by this theoretical advancement, we propose a practical method to incorporate image augmentation into CLIP-like models to extract content information from the original learned features. Specifically, a disentangled network is designed to fine-tune the pre-trained CLIP model by using a contrastive loss with image augmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Despite the advancements made in disentangling content and style information from the original features learned by CLIP-like models through image augmentation, we recognize an inherent limitation: it is generally challenging to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.362, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "CLAP: Contrastive Learning with Augmented Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.265 + ], + "angle": 0, + "content": "design adequate image augmentations to ensure all style factors change in an image. Theoretically, disentangling content and style information necessitates changes in all style factors [41]. However, inducing sufficient changes in latent style through image augmentation poses challenges due to the high dimensionality and complexity of style information in image data. Achieving significant style variation via artificially designed image augmentation techniques, such as transforming a photograph of a dog into a sketch while preserving content but dramatically altering style, is notably difficult." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.786, + 0.508 + ], + "angle": 0, + "content": "Taking a further step, rather than relying on image augmentation, we explore the use of text augmentation to disentangle latent content and style factors. This shift is motivated by two key observations: 1) Vision and language data share the same latent space. Therefore, text augmentation can also be utilized to induce changes in latent style factors instead of image augmentation. 2) Text data inherently possesses high semanticity and logical structure, making it more amenable to property-wise manipulation compared to image data. Consequently, implementing sufficient style changes through text augmentation is more feasible than image augmentation, contributing to isolating content from style information, see Fig. 1c for visual comparison. For instance, transforming text from \"a photo of a dog\" to \"a sketch of a dog\" is straightforward in the language modality, whereas achieving a similar transformation in image data is challenging. Inspired by these observations, we posit that introducing style variations through text augmentation, as illustrated Fig. 1b, provides a more effective approach for learning vision-language content features than relying on image augmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.751 + ], + "angle": 0, + "content": "In summary, our contributions include: (1) Aimed at disentangling latent content and style factors to refine vision-language features of pre-trained CLIP-like models, we propose constractive learning with data augmentation to fine tune the original features of pre-trained CLIP-like models from a causal perspective. (2) We present a novel method customized for pre-trained CLIP-like models. This method leverages a disentangled network, which is trained using contrastive learning with image augmentation, to extract latent content features from the learned features provided by image encoder of CLIP-like models. (3) We propose Contrastive Learning with Augmented Prompts (CLAP), to extract latent content features from representations of CLIP-like models. It begins by training a disentangled network using the pre-trained text encoder of CLIP-like models and text augmentation. Subsequently, the trained disentangled network is transferred to the image encoder of CLIP-like models. (4) Experiments conducted on a large real dataset demonstrate the effectiveness of the proposed image augmentation and text augmentation in terms of zero-shot and few-shot performance, as well as robustness against perturbations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.776, + 0.388, + 0.792 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Contrastive Vision-Language Models Using a cross-modal contrastive loss, CLIP [36] revolutionarily introduced a scalable contrastive vision-language model" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Y. Cai et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.375 + ], + "angle": 0, + "content": "by leveraging a large corpus of internet-sourced image-text pairs, demonstrating unprecedented zero-shot learning capabilities and exceptional generalization ability across datasets and supporting numerous downstream tasks [38]. ALIGN [20] expanded the scale of contrastive vision-language modeling, training on up to one billion image-text pairs while integrating the vision transformer's self-attention mechanism [11], which further enhanced performance. Despite their successes, CLIP-like models exhibit sensitivity to input text prompts [21, 48], leading to variable performance across different prompts. Efforts to mitigate this prompt sensitivity through prompt learning and engineering [9, 14, 21, 47, 48] focus on customizing prompts for specific tasks but do not fundamentally enhance CLIP's representations. Furthermore, CLIP-like models are vulnerable to adversarial attacks [4, 12], with current strategies [33, 45] involving adversarial-natural image pairs to improve resilience. Our work diverges from task-specific approaches by aiming to enhance CLIP's representations from a disentanglement perspective, addressing the aforementioned issues inherent in CLIP-like models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.79, + 0.624 + ], + "angle": 0, + "content": "Disentangled Representation Learning Aimed at segregating intrinsic latent factors in data into distinct, controllable representations, disentangled representation learning benefits various applications [24,40,44]. Specifically, in classification tasks, it's shown that enhancing the model's performance and robustness against data distribution perturbations can be achieved by more effectively disentangling invariant content variables, without needing to identify all intrinsic latent variables completely [22,26-28]. Within single modalities, studies such as [49] have illustrated that contrastive learning [7,16,18] can potentially reverse the data generative process, aiding in the separation of representations. Furthermore, [41] suggest that image augmentation can help isolate content variables from the latent space through significant stylistic changes. [19] employs mixture techniques for data augmentation, enabling more abundant cross-modal matches. Diverging from these methods, our approach focuses on employing text augmentation to disentangle latent content variables, introducing a unique approach to learn refined vision-language representations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.644, + 0.742, + 0.661 + ], + "angle": 0, + "content": "3 A Causal Generative Model for Multi-Modal Data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.844 + ], + "angle": 0, + "content": "To understand pretrained CLIP-like models, we investigate the underlying causal generative process for vision-language data. We consider the following causal generative model as depicted in Fig. 1. In the proposed model, the shared latent space ruling vision and language data is divided into two distinct sub-spaces: one corresponding to the latent content variables \\(\\mathbf{c}\\) and the other to the latent style variables \\(\\mathbf{s}\\). The latent content variables are posited to determine the object label \\(\\mathbf{y}\\), a relationship corroborated by prior studies [22, 29, 31]. Furthermore, to elucidate the correlation between the latent style variable \\(\\mathbf{s}\\) and the object variable \\(\\mathbf{y}\\), our model incorporates the premise that the latent content variable \\(\\mathbf{c}\\) causally influences the latent style variable \\(\\mathbf{s}\\), in concordance with the principles of causal representation learning highlighted in recent literature [10, 29, 41]. Additionally," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.362, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "CLAP: Contrastive Learning with Augmented Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "considering the diversity between image data and text data, where information in image data is typically much more details while information in text data tends to be more logically structured nature, we posit distinct causal mechanisms for the generation processes. Our causal generative model is formulated as following structural causal models [2]:" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.231, + 0.786, + 0.248 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} := \\mathbf {g} _ {\\mathbf {s}} (\\mathbf {c}), \\mathbf {x} := \\mathbf {g} _ {\\mathbf {x}} (\\mathbf {c}, \\mathbf {s}), \\mathbf {t} := \\mathbf {g} _ {\\mathbf {t}} (\\mathbf {c}, \\mathbf {s}), \\mathbf {y} := \\mathbf {g} _ {\\mathbf {y}} (\\mathbf {c}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.254, + 0.788, + 0.375 + ], + "angle": 0, + "content": "In Eq. (1), the style variable \\(\\mathbf{s}\\) is causally influenced by the content via \\(\\mathbf{g_s}\\); \\(\\mathbf{x}\\) and \\(\\mathbf{t}\\) denote visual and textual data, respectively. Both visual and textual data are causally produced by the shared latent variables \\(\\mathbf{c}\\) and \\(\\mathbf{s}\\) through distinct, reversible generative processes: \\(\\mathbf{g_x}\\) for images and \\(\\mathbf{g_t}\\) for text data, respectively. The label \\(\\mathbf{y}\\) of a sample is exclusively determined by the content variable \\(\\mathbf{c}\\) via \\(\\mathbf{g_y}\\). For simplicity, exogenous noises are implicitly assumed but not explicitly represented in the causal generative model's formulation, aligning with the common understanding that each latent variable is influenced by exogenous noise." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.789, + 0.603 + ], + "angle": 0, + "content": "Recent seminal work in [41] has demonstrated that the latent content variable \\(\\mathbf{c}\\) can be identified up to block identifiability (i.e., \\(\\mathbf{c}\\) can be isolated from style variable \\(\\mathbf{s}\\)), by requiring all latent style variables to change (e.g., soft interventions on all latent style variables). This change can be achieved through image augmentation, i.e., the augmented image \\(\\tilde{\\mathbf{x}}\\) can be interpreted as a generative result of \\(\\tilde{\\mathbf{s}}\\), which is produced through soft interventions on original latent style variables \\(\\mathbf{s}\\). Despite such theoretical advancement, the practical implementation of this theoretical result within CLIP-like models remains unclear. In this study, we propose a practical method to disentangle content and style information within CLIP-like models by employing image augmentation, as detailed in Sec. 4.1. Moreover, we recognize that implementing sufficient changes on all latent style variables \\(\\mathbf{s}\\) through text augmentation is more feasible than image augmentation, due to high semanticity and logical structure in text data, we delve into the use of text augmentation to separate content information from style information, as discussed in Sec. 4.2." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.621, + 0.779, + 0.64 + ], + "angle": 0, + "content": "4 Isolating Content from Style with Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.649, + 0.788, + 0.756 + ], + "angle": 0, + "content": "In this section, we propose the employment of data augmentation to extract content information from the learned features in pre-trained CLIP-like models. Essentially, data augmentation facilitates the alteration of style factors while preserving content factors. Consequently, leveraging contrastive learning enables the segregation of content information from style information. We delve into two distinct forms of data augmentation, namely image augmentation (Sec. 4.1) and text augmentation (Sec. 4.2)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.774, + 0.707, + 0.79 + ], + "angle": 0, + "content": "4.1 Isolating Content from Style with Augmented Images" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "While recent studies (von et al., 2021) have offered assurance regarding the disentanglement of content and style through contrastive learning with data augmentation, it remains unclear how these theoretical findings can be applied to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Cai et al." + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.145, + 0.407, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.228, + 0.392, + 0.241 + ], + "angle": 0, + "content": "(a) Image augmentation" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.145, + 0.593, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.228, + 0.574, + 0.241 + ], + "angle": 0, + "content": "(b) Text augmentation" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.145, + 0.781, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.618, + 0.228, + 0.76, + 0.24 + ], + "angle": 0, + "content": "(c) Zero-shot inference" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.252, + 0.788, + 0.351 + ], + "angle": 0, + "content": "Fig. 2: Refining CLIP through data augmentation. (a) Training involves a disentangled network \\(\\mathbf{f_c}\\), utilizing contrastive loss on original and augmented image pairs \\(\\mathbf{x}\\) and \\(\\tilde{\\mathbf{x}}\\), with CLIP's image encoder \\(\\mathbf{f_x^*}\\) holding frozen gradients. (b) More efficient content feature learning is achieved through contrastive learning with augmented text prompts \\(\\mathbf{t}\\) and \\(\\tilde{\\mathbf{t}}\\), using the fixed text encoder \\(\\mathbf{f_t^*}\\) of CLIP. (c) Inference stage: The trained disentangled network \\(\\mathbf{f_c^*}\\) integrates with CLIP's text and image encoders, \\(\\mathbf{f_t^*}\\) and \\(\\mathbf{f_x^*}\\), to enable zero-shot inference for an input image \\(\\mathbf{x}\\) and class names \\(\\mathbf{t_1}\\) to \\(\\mathbf{t_n}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.788, + 0.423 + ], + "angle": 0, + "content": "the realm of vision-language models. We convert the theoretical findings into CLIP-like models in the following. The theoretical findings suggest using InfoNCE loss [34] to extract content information, as outlined below:" + }, + { + "type": "equation", + "bbox": [ + 0.274, + 0.43, + 0.786, + 0.468 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\mathbf {f}; \\{\\mathbf {x} _ {i}, \\tilde {\\mathbf {x}} _ {i} \\} _ {i = 1} ^ {b}, \\tau) = - \\frac {1}{b} \\sum_ {i = 1} ^ {b} \\log \\frac {\\exp [ \\langle \\mathbf {f} (\\mathbf {x} _ {i}) , \\mathbf {f} (\\tilde {\\mathbf {x}} _ {i}) \\rangle / \\tau ]}{\\sum_ {j = 1} ^ {b} \\exp [ \\langle \\mathbf {f} (\\mathbf {x} _ {i}) , \\mathbf {f} (\\tilde {\\mathbf {x}} _ {j}) \\rangle / \\tau ]}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.785, + 0.536 + ], + "angle": 0, + "content": "where \\(\\{\\mathbf{x}_i\\}_{i=1}^b\\) represents a batch of \\(b\\) samples from the training dataset, \\(\\mathbf{f}(\\mathbf{x}_i)\\) denotes sample \\(\\mathbf{x}_i\\)'s features through model \\(\\mathbf{f}\\), \\(\\tilde{\\mathbf{x}}_i\\) is the augmented counterpart of \\(\\mathbf{x}_i\\), and \\(\\langle \\mathbf{z}_1, \\mathbf{z}_2 \\rangle\\) represents the cosine similarity between two feature vectors, \\(\\mathbf{z}_1\\) and \\(\\mathbf{z}_2\\), and \\(\\tau\\) represents the temperature parameter influencing the loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.627 + ], + "angle": 0, + "content": "We extend it to refine pre-trained vision-language models, utilizing contrastive learning with augmented images (hereinafter referred to as \"Im.Aug\"). As illustrated in Fig. 2a, we train a disentangled network on top of CLIP's pretrained image encoder. To enhance training efficiency and the usability of the proposed method, we freeze the pre-trained image encoder. Based on an InfoNCE loss, the learning objective of Im.Aug is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.33, + 0.634, + 0.787, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {\\mathbf {c}} ^ {*} = \\underset {\\mathbf {f} _ {\\mathbf {c}}} {\\operatorname {a r g m i n}} \\underset {\\{\\mathbf {x} _ {i} \\} _ {i = 1} ^ {b} \\in \\mathcal {D} _ {\\mathbf {x}}} {\\mathbb {E}} \\mathcal {L} \\left(\\mathbf {f} _ {\\mathbf {c}} \\circ \\mathbf {f} _ {\\mathbf {x}} ^ {*}; \\left\\{\\mathbf {x} _ {i}, \\tilde {\\mathbf {x}} _ {i} \\right\\} _ {i = 1} ^ {b}, \\tau\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.669, + 0.788, + 0.744 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}_{\\mathbf{x}}\\) denotes the training image dataset and \\(b\\) represents the batch size, \\(\\mathbf{f_c}\\) is the disentangled network undergoing training. The pre-trained CLIP image encoder is represented by \\(\\mathbf{f_x^*}\\), with the asterisk \"***\" signifying that the model weights remain fixed. The variable \\(\\mathbf{x}_i\\) refers to an image sampled from \\(\\mathcal{D}_{\\mathbf{x}}\\), and \\(\\tilde{\\mathbf{x}}_i\\) is its augmented view." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.746, + 0.788, + 0.791 + ], + "angle": 0, + "content": "The composition of the training dataset \\( \\mathcal{D}_{\\mathbf{x}} \\), the image augmentation techniques used, the structure of the disentangled network \\( \\mathbf{f_c} \\), and the utilization of \\( \\mathbf{f_c^*} \\) post-training are detailed in the following subsections." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Data Synthesis and Image Augmentation To generate training image data, we combine class names with various image and object attributes to create text" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.362, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "CLAP: Contrastive Learning with Augmented Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 1: Template-based prompts. Attributes used to generate text prompts follow the structured format \"a [art style] [image type] of a [object size] [object color] [class]\", where \"[class]\" represents the class names." + }, + { + "type": "table", + "bbox": [ + 0.249, + 0.199, + 0.756, + 0.264 + ], + "angle": 0, + "content": "
Object ColorObject SizeImage TypeArt Style
yellow, green, black, blue, multicolored, orange, red, white, brown, purplelarge, small, normal sizedpainting, cartoon, infographic, sketch, photograph, clipart, mosaic art, sculpturerealistic, impressionistic
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.292, + 0.789, + 0.459 + ], + "angle": 0, + "content": "prompts for each class. Using a stable diffusion model [39], we produce synthetic images that comprise our training dataset \\(\\mathcal{D}\\mathbf{x}\\). The creation of template prompts for stable diffusion is based on attributes such as object size, color, image type, and art style. As detailed in Tab. 1, the attributes include 10 colors and 3 sizes for objects, and 8 types and 2 art styles for images. By assembling these attributes into prompts like \"a [art style] [image type] of a [object size] [object color] [class]\", we generate 480 unique texts for each class, from which one image per prompt is synthesized. Further details on image synthesis and examples are available in Appendix B.1. For the image augmentation procedures, we adopt techniques commonly used in contrastive learning practice [7,8,41], specifically random cropping and color distortion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.789, + 0.604 + ], + "angle": 0, + "content": "Disentangled Network Structure Since the training process is based on CLIP's pre-trained lower-dimensional features, our disentangled network adopts a multi-layer perceptron (MLP) architecture. To fully benefit from the pretrained CLIP text encoder, we construct a residual MLP featuring a zero-initialized projection, acting as the disentangled network, as depicted in Fig. 3. This design enables learning directly from the pre-trained representation space, avoiding a random starting point, inspired by ControlNet's zero-conv operation [46], which we adapt to a zero-linear operation within our residual MLP." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.605, + 0.789, + 0.788 + ], + "angle": 0, + "content": "Within this architecture, the main branch includes a zero-initialized, bias-free linear layer positioned subsequent to the combination of a SiLU activation and a normally initialized linear layer. Conventionally, the dimensions of features before the initial linear layer, situated between the first and second linear layers, and following the second linear layer, are named as the input \\( d_{in} \\), latent \\( d_{mid} \\), and output \\( d_{out} \\) dimensions, respectively. To rectify any mismatches between the input and output dimensions, the network employs nearest-neighbor downsampling within the shortcut path, thereby ensuring both alignment and the preservation of sharpness for the input features. During the inference stage, a weighting parameter \\( \\alpha > 0 \\) is introduced to modulate the portion of features emanating from the main branch before their integration with the input features, whereas this parameter remains constant at 1 throughout the training phase." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Inference After training, the disentangled network \\(\\mathbf{f}_{\\mathbf{c}}^{*}\\) is utilized following CLIP's image encoder to extract visual content features. Moreover, given that" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Cai et al." + }, + { + "type": "image", + "bbox": [ + 0.343, + 0.145, + 0.664, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.258, + 0.788, + 0.327 + ], + "angle": 0, + "content": "Fig. 3: Structure of the disentangled network. The architecture encompasses a residual block featuring a zero-initialized, bias-free linear layer to commence optimization from the input feature space. When the input and output dimension differ, a downsampling operation is utilized to achieve alignment. During inference, a scalar parameter \\(\\alpha\\) balance the main branch and input features before combination." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.448 + ], + "angle": 0, + "content": "vision-language data generation is rooted in a unified latent space, as depicted in Sec. 3, \\(\\mathbf{f}_{\\mathbf{c}}^{*}\\) can be seamlessly integrated with CLIP's image and text encoders to enhance zero-shot capabilities. As shown in Fig. 2c, for an image \\(\\mathbf{x}\\), the operation is formulated as the composition function \\(\\mathbf{f}_{\\mathbf{c}}^{*} \\circ \\mathbf{f}_{\\mathbf{x}}^{*}(\\mathbf{x})\\), and similarly, for a text \\(\\mathbf{t}\\), as \\(\\mathbf{f}_{\\mathbf{c}}^{*} \\circ \\mathbf{f}_{\\mathbf{t}}^{*}(\\mathbf{t})\\). This integration preserves CLIP's zero-shot functionality while achieving refined features through the improved disentanglement of content." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.468, + 0.719, + 0.485 + ], + "angle": 0, + "content": "4.2 Isolating Content from Style with Augmented Prompts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.627 + ], + "angle": 0, + "content": "Despite progress in disentangling content and style via image augmentation, adequately altering all style factors in an image remains challenging due to the high dimensionality and complexity of style information in images. Achieving substantial style changes through augmentation, essential for complete disentanglement [41], is difficult with existing image augmentation techniques. On the contrary, text data inherently possesses high semanticity and logical structure, making it more amenable to property-wise manipulation compared to image data. To further exploring the disentanglement of content, we propose Contrastive Learning with Augmented Prompts (CLAP)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.719 + ], + "angle": 0, + "content": "As depicted in Fig. 2b, CLAP employs an InfoNCE loss to train a disentangled network atop CLIP's pre-trained text encoder, keeping the encoder's gradients fixed, similar to Im.Aug. Leveraging the simpler structure of text, the template-based prompts previously utilized for synthesizing images now serve as the training text dataset, denoted by \\(\\mathcal{D}_{\\mathbf{t}}\\). Utilizing the same disentangled network as in Im.Aug, the learning objective of CLAP is outlined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.228, + 0.727, + 0.787, + 0.756 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {\\mathbf {c}} ^ {*} = \\underset {\\mathbf {f} _ {\\mathbf {c}}} {\\operatorname {a r g m i n}} \\underset {\\left\\{\\mathbf {t} _ {i} \\right\\} _ {i = 1} ^ {b} \\in \\mathcal {D} _ {\\mathbf {t}}} {\\mathbb {E}} \\mathcal {L} \\left(\\mathbf {f} _ {\\mathbf {c}} \\circ \\mathbf {f} _ {\\mathbf {t}} ^ {*}; \\left\\{\\mathbf {t} _ {i}, \\tilde {\\mathbf {t}} _ {i} \\right\\} _ {i = 1} ^ {b}, \\tau\\right) + \\lambda \\mathcal {L} \\left(\\mathbf {f} _ {\\mathbf {c}} \\circ \\mathbf {f} _ {\\mathbf {t}} ^ {*}; \\left\\{\\mathbf {t} _ {i} ^ {c}, \\tilde {\\mathbf {t}} _ {i} \\right\\} _ {i = 1} ^ {b}, 1\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}_{\\mathbf{t}}^{*}\\) denotes the pre-trained CLIP text encoder. The term \\(\\mathbf{t}_i\\) references a text prompt from \\(\\mathcal{D}\\mathbf{t}\\), and \\(\\tilde{\\mathbf{t}}_i\\) represents its augmented view, produced via prompt augmentation techniques. On the equation's right side, \\(\\mathbf{t}_i^c\\) specifies the class name associated with the text prompt \\(\\mathbf{t}_i\\). This strategy aims to enhance variations between prompt pairs, especially in cases where the text dataset \\(\\mathcal{D}_{\\mathbf{t}}\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.362, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "CLAP: Contrastive Learning with Augmented Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 2: Prompt augmentation techniques. Various augmented views are generated from an original text prompt using specific augmentation techniques: OSD (Object Size Deletion), OCD (Object Color Deletion), ITD (Image Type Deletion), ASD (Art Style Deletion), and SPO (Swapping Prompt Order)." + }, + { + "type": "table", + "bbox": [ + 0.249, + 0.213, + 0.756, + 0.276 + ], + "angle": 0, + "content": "
OriginalOSDOCDITDASDSPO
a realistic painting of a large red cara realistic painting of a red cara realistic painting of a large cara realistic of a large red cara painting of a large red cara large red car in a realistic painting
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.303, + 0.784, + 0.348 + ], + "angle": 0, + "content": "has a very limited number of samples. Here, \\(\\lambda\\) serves for adjusting the second term's importance in the total loss function. All other symbols in Eq. (4) maintain their definitions as described earlier." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.349, + 0.785, + 0.38 + ], + "angle": 0, + "content": "After training, the learned disentangled network is seamlessly integrated with both of CLIP's encoders to extract content representations, as depicted in Fig. 2c." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.789, + 0.646 + ], + "angle": 0, + "content": "Prompt Augmentation To ensure text prompts undergo stylistic changes without compromising their content, we have developed specific augmentation techniques for synthetic text prompts. Drawing inspiration from Easy Data Augmentation (EDA) techniques [42], we adapted the Random Deletion (RD) and Random Swap (RS) techniques from EDA, customizing them to suit our prompt structure. To avoid inadvertently altering the content by introducing new object names or changing the core idea of a text prompt, our augmentation methods do not include random word insertions or replacements. Our primary augmentation techniques are Object Size Deletion (OSD), Object Color Deletion (OCD), Image Type Deletion (ITD), Art Style Deletion (ASD), and Swapping Prompt Order (SPO), each applied with a certain probability, as detailed in Tab. 2. Additionally, for down-stream datasets with few categories, to rich the population of training samples, we use an additional augmentation, named IGN (Inserting Gaussian Noise). Following the initializing protocol of prompt learning methods [47,48], we insert a zero-mean Gaussian noise with 0.02 standard deviation with a noise length equals to 4, to the tokenized prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.647, + 0.789, + 0.753 + ], + "angle": 0, + "content": "Intuitively, these prompt augmentation methods parallel random masking techniques used in image augmentation [6,17]. However, prompt augmentations are more effective and precise than their image counterparts. This effectiveness arises because prompt augmentations can specifically target and eliminate a particular style element without impacting the content, whereas image masking, operating at the pixel or patch level, might inadvertently damage content information or lead to insufficient style changes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.777, + 0.377, + 0.794 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We conduct three primary experiments to assess our method: (1) zero-shot evaluation with diverse prompts to gauge zero-shot performance and its robustness" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Cai et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "to prompt perturbations; (2) linear probe tests on few-shot samples to evaluate the efficacy of the learned representations in few-shot settings; and (3) adversarial attack assessments on zero-shot and one-shot classifiers to determine their resistance to adversarial threats. We further conduct an ablative study on hyperparameters, explore the impact of different prompt augmentation combinations and various sources of training prompts on CLAP's performance, and replicate experiments across different CLIP model sizes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.274, + 0.428, + 0.289 + ], + "angle": 0, + "content": "5.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.297, + 0.787, + 0.358 + ], + "angle": 0, + "content": "Implementation. Im.Aug and CLAP are implemented using the ViT-B/16 CLIP model and executed on an NVIDIA RTX 3090 GPU. To ensure reproducibility, the random seed for all stochastic processes is fixed at 2024. More information on implementation details is provided in Appendix A.1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.787, + 0.448 + ], + "angle": 0, + "content": "Datasets. CLAP is assessed across four multi-domain datasets to examine its performance in varied environments: PACS [23] (4 domains, 7 categories), VLCS [1] (4 domains, 5 categories), OfficeHome [37] (4 domains, 65 categories), and DomainNet [35] (6 domains, 345 categories). For conciseness, we present average results across the domains for each dataset. Detailed experimental outcomes for each domain within these datasets are provided in Appendix A.4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.449, + 0.789, + 0.554 + ], + "angle": 0, + "content": "Compute efficiency. CLAP demonstrates faster convergence and shorter training times compared to Im.Aug. For CLAP, training on the PACS and VLCS datasets is completed in roughly 11 minutes, OfficeHome in approximately 14 minutes, and DomainNet in about 47 minutes. In contrast, Im.Aug requires around 16 minutes for PACS and VLCS, 50 minutes for OfficeHome, and 3.3 hours for DomainNet. Both Im.Aug and CLAP maintain CLIP's inference efficiency due to the disentangled network's efficient two-layer MLP structure." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.575, + 0.373, + 0.589 + ], + "angle": 0, + "content": "5.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.704 + ], + "angle": 0, + "content": "Zero-Shot Performance To assess zero-shot capabilities, CLAP undergoes evaluation using three specific fixed prompts: ZS(C), utilizing only the class name within \"[class]\"; ZS(PC), with the format \"a photo of a [class]\"; and ZS(CP), structured as \"a [class] in a photo\". To thoroughly examine zero-shot proficiency, a dynamic prompt, ZS(NC), formatted as \"[noise][class]\", is also used, where \"[noise]\" signifies the introduction of Gaussian noise characterized by a mean of 0 and a standard deviation of 0.02." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.809 + ], + "angle": 0, + "content": "As presented in Tab. 3, CLAP surpasses both CLIP and Im.Aug across all evaluated prompts for every dataset. Unlike the uniform enhancement in zero-shot performance CLAP achieves over CLIP, Im.Aug displays inconsistent results. A closer examination reveals CLAP's superiority over CLIP is especially significant for the dynamic ZS(NC) prompt. This demonstrates CLAP's effectiveness in significantly improving zero-shot performance compared to the original CLIP representations." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In assessing the model's robustness to prompt perturbations, we examine the variances in zero-shot performance across different prompts by analyzing the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.362, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "CLAP: Contrastive Learning with Augmented Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.144, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 3: Zero-shot results across three distinct prompts: \"C\" for \"[class]\", \"CP\" for \"a [class] in a photo\", \"PC\" for \"a photo of a [class]\", and a dynamic prompt \"NC\" for \"[noise][class]\" showcase that CLAP consistently outperforms CLIP's zero-shot performance across all datasets, whereas image augmentation exhibits mixed outcomes." + }, + { + "type": "table", + "bbox": [ + 0.312, + 0.213, + 0.695, + 0.424 + ], + "angle": 0, + "content": "
PromptMethodZero-shot performance, avg. top-1 acc. (%) (↑)
PACSVLCSOff.HomeDom.NetOverall
ZS(C)CLIP95.776.479.857.877.4
Im.Aug96.579.577.051.576.1
CLAP97.282.681.058.779.9
ZS(CP)CLIP95.282.079.557.078.4
Im.Aug96.382.975.850.776.4
CLAP97.383.480.558.079.8
ZS(PC)CLIP96.182.482.557.779.7
Im.Aug96.583.078.651.677.4
CLAP97.283.483.059.080.6
ZS(NC)CLIP90.868.371.551.070.4
Im.Aug94.873.167.544.069.9
CLAP97.281.073.552.676.1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.784, + 0.496 + ], + "angle": 0, + "content": "range \\((R)\\) and standard deviation \\((\\delta)\\) of results derived from ZS(C), ZS(CP), and ZS(PC). Additionally, we investigate the decline \\((\\varDelta_{(NC)})\\) in performance from ZS(C) to ZS(NC) as a broad indicator of resilience to noised prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.497, + 0.788, + 0.618 + ], + "angle": 0, + "content": "As presented in Tab. 4, CLAP significantly reduces the variance in zero-shot performance across various testing prompts, evidenced by markedly lower values of \\(\\delta\\) and \\(R\\), and a less pronounced decrease in performance with a noised prompt, in contrast to Im.Aug and the baseline representations of CLIP. Although Im.Aug aids in reducing performance variance to some extent, its efficacy is notably inferior to that of CLAP. These findings highlight CLAP's enhanced robustness in maintaining consistent zero-shot performance across a diverse array of prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.638, + 0.788, + 0.775 + ], + "angle": 0, + "content": "Few-Shot Performance We conduct evaluations of 1-shot, 4-shot, 8-shot, and 16-shot linear probes across each domain within the four datasets. As illustrated in Fig. 4, CLAP significantly outperforms both CLIP and Im.Aug in few-shot learning scenarios. Notably, in the 1-shot setting CLAP achieves performance improvements over the linear-probe CLIP model by margins of \\(+10\\%\\), \\(+3.5\\%\\), \\(+2.5\\%\\), and \\(+1.5\\%\\) on the PACS, VLCS, OfficeHome, and DomainNet datasets, respectively. These improvements are especially significant in comparison to the gains observed with Im.Aug counterparts, underpinning CLAP's efficacy in few-shot scenarios. For detailed quantitative results, please refer to Appendix A.4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.843 + ], + "angle": 0, + "content": "Adversarial Performance To assess adversarial robustness, zero-shot (ZS(C)) and one-shot classifiers are evaluated against prominent adversarial attack methods, such as FGSM [15], PGD [30], and CW [5], by generating adversarial sam" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Cai et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 4: CLAP more effectively reduces zero-shot performance variance across prompts than image augmentation, with \\( R \\) and \\( \\delta \\) indicating the range and standard deviation for ZS(C), ZS(PC), and ZS(PC). The decrease \\( \\Delta_{(NC)} \\) from ZS(C) to ZS(NC) highlights CLAP's enhanced robustness against prompt perturbations." + }, + { + "type": "table", + "bbox": [ + 0.32, + 0.213, + 0.685, + 0.38 + ], + "angle": 0, + "content": "
MetricMethodPerformance variance, avg. top-1 acc. (%) (↓)
PACSVLCSOff.HomeDom.NetOverall
RCLIP0.96.13.10.82.7
Im.Aug0.13.62.80.91.9
CLAP0.10.82.51.01.1
δCLIP0.42.81.40.41.2
Im.Aug0.11.71.20.40.8
CLAP0.00.41.10.40.5
Δ(NC)CLIP4.98.18.36.87.0
Im.Aug1.66.49.57.56.3
CLAP0.01.67.56.13.8
" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.392, + 0.368, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.392, + 0.503, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.392, + 0.641, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.392, + 0.78, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.489, + 0.788, + 0.534 + ], + "angle": 0, + "content": "Fig. 4: Few-shot linear probe comparisons of image-encoder features show that CLAP enhances CLIP's few-shot performance more effectively than Im.Aug. In the accompanying figure, \"ZS\" indicates the zero-shot performance using a \"[class]\" prompt." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.562, + 0.788, + 0.729 + ], + "angle": 0, + "content": "plies for testing. For FGSM, 1 adversarial iteration is employed, whereas for PGD and CW, 20 iterations are used, all with an epsilon of 0.031. As indicated in Tab. 5, classifiers utilizing CLAP representations demonstrate superior resilience to these adversarial attacks compared to those based on CLIP representations. Across the four datasets, CLAP's zero-shot and 1-shot classifiers surpass CLIP by margins of \\(+7.6\\%\\) and \\(+8.5\\%\\) against FGSM, \\(+1.0\\%\\) and \\(+11.7\\%\\) against PGD-20, and \\(+1.1\\%\\) and \\(+2.3\\%\\) against CW-20, respectively. These figures notably exceed the performance improvements of \\(+4.4\\%\\) and \\(+4.6\\%\\) against FGSM, \\(+0.3\\%\\) and \\(+6.2\\%\\) against PGD-20, and \\(0\\%\\) and \\(+1.3\\%\\) against CW-20 achieved by Im.Aug. The result suggests that CLAP efficiently enhances robustness against adversarial attacks in both zero-shot and one-shot scenarios." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.753, + 0.383, + 0.768 + ], + "angle": 0, + "content": "5.3 More Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "t-SNE Visualization In our t-SNE visualizations, we examine the representations of CLIP, Im.Aug, and CLAP for all images within the Art Painting domain of the PACS dataset. Fig. 5 shows that CLAP's image representations display a marked inter-class separation and tighter intra-class clustering than those of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.362, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "CLAP: Contrastive Learning with Augmented Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 5: Image augmentation and CLAP both enhance CLIP's zero-shot with the \"[class]\" prompt and 1-shot robustness against adversarial attacks, with CLAP showing greater improvements." + }, + { + "type": "table", + "bbox": [ + 0.234, + 0.199, + 0.77, + 0.313 + ], + "angle": 0, + "content": "
Setting MethodAvg. top-1 acc. (%) under adversarial attacks(↑)
FGSMPGD-20CW-20
PACSVLCSO.H.D.N.PACSVLCSO.H.D.N.PACSVLCSO.H.D.N.
ZS(C)CLIP86.865.657.922.529.12.010.110.727.41.57.47.6
Im.Aug88.069.655.137.931.32.110.49.029.41.77.05.8
CLAP88.771.958.544.230.83.210.611.229.82.38.18.0
1-shotCLIP66.745.234.322.534.816.05.611.318.90.74.53.2
Im.Aug79.447.137.123.555.216.18.512.523.20.95.13.4
CLAP89.652.237.123.973.421.27.412.527.01.15.03.5
" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.334, + 0.391, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.432, + 0.346, + 0.445 + ], + "angle": 0, + "content": "(a) CLIP" + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.334, + 0.578, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.432, + 0.54, + 0.446 + ], + "angle": 0, + "content": "(b) Im.Aug" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.334, + 0.765, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.432, + 0.722, + 0.445 + ], + "angle": 0, + "content": "(c) CLAP" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.456, + 0.788, + 0.499 + ], + "angle": 0, + "content": "Fig. 5: t-SNE visualizations of all images in the Art Painting of PACS dataset show CLAP outperforms the original CLIP and Im.Aug, with clearer inter-class distinctions and tighter intra-class clusters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.527, + 0.788, + 0.573 + ], + "angle": 0, + "content": "CLIP and Im.Aug. This observation suggests that CLAP's representations are more closely tied to content information and less influenced by style information, in contrast to the other two." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.593, + 0.789, + 0.76 + ], + "angle": 0, + "content": "Ablations In Fig. 6, we assess the zero-shot capabilities of our model using two distinct prompts, ZS(C) and ZS(PC), on the VLCS dataset. This analysis forms part of an ablative study aimed at understanding the influence of various hyperparameters on model performance. Specifically, we examine: the dimensions of the latent layer within the MLP of the disentangled network, as illustrated in Fig. 6a; the temperature parameter \\((\\tau)\\) in the loss function, as depicted in Fig. 6b; and the weight coefficient \\((\\alpha)\\) during the inference stage, as shown in Fig. 6c. Our findings indicate that CLAP consistently enhances zero-shot performance across all tested configurations for both prompts, while also significantly reducing the gap between the performances elicited by each prompt. These results underscore the efficacy of CLAP in accommodating a wide range of hyper-parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Prompt Augmentation Combinations We explore diverse combinations of our tailored prompt augmentation methods and examine Easy Data Augmentation (EDA) techniques [42] on the VLCS dataset. Each tested technique showcases CLAP's enhancements over CLIP, with details available in Appendix A.2." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Cai et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.151, + 0.406, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.15, + 0.593, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.15, + 0.782, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.266, + 0.787, + 0.308 + ], + "angle": 0, + "content": "Fig. 6: We conduct ablative study on hyper-parameter choices on the VLCS dataset, including latent dimensions, \\(\\tau\\) values, and \\(\\alpha\\) values during the inference stage. CLAP continuously enhance CLIP's performance throughout the tested values." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.788, + 0.473 + ], + "angle": 0, + "content": "Prompt Sources We assess the impact of different training prompt formats, originating from various synthetic sources, on the performance of the VLCS dataset, incorporating EDA techniques. Our evaluation includes our template-based prompts, LLM-generated prompts by ChatGPT-3.5 [3] (with the generation process detailed in Appendix B.2), prompts structured as \"a [random] style of [class],\" where \"[random]\" is filled with terms from a random word generator\\(^2\\), and prompts produced using the PromptStyler method [9]. The findings indicate that the training prompts with simpler forms tend to yield better performance, with detailed quantitative results presented in Appendix A.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.788, + 0.587 + ], + "angle": 0, + "content": "Experiments on Different Model Scales In our repeated experiments assessing zero-shot performance on the ViT-L/14 and ResNet50x16 pre-trained with CLIP, we consistently find that CLAP improves zero-shot performance while also reducing performance variances. This consistent observation underscores CLAP's effectiveness in enhancing the quality of CLIP representations. For quantitative details supporting these findings, please see the Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.608, + 0.36, + 0.624 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.639, + 0.788, + 0.821 + ], + "angle": 0, + "content": "To enhance pre-trained CLIP-like models, this study delves into disentangling latent content variables. Through a causal analysis of the underlying generative processes of vision-language data, we discover that training a disentangled network in one modality can effectively disentangle content across both modalities. Given the high semantic nature of text data, we identify that disentanglement is more achievable within the language modality through text augmentation interventions. Building on these insights, we introduce CLAP (Contrastive Learning with Augmented Prompts) to acquire disentangled vision-language content features. Comprehensive experiments validate CLAP's effectiveness, demonstrating significant improvements in zero-shot and few-shot performance, and enhancing robustness against perturbations. We anticipate that our work will inspire further exploration into disentangling latent variables within vision-language models." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.585, + 0.84 + ], + "angle": 0, + "content": "2 https://github.com/vaibhavsingh97/random-word" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.362, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "CLAP: Contrastive Learning with Augmented Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.214 + ], + "angle": 0, + "content": "1. Albuquerque, I., Monteiro, J., Darvishi, M., Falk, T.H., Mitliagkas, I.: Generalizing to unseen domains via distribution matching. arXiv preprint arXiv:1911.00804 (2019). https://doi.org/10.48550/arXiv.1911.00804" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.215, + 0.785, + 0.241 + ], + "angle": 0, + "content": "2. Bollen, K.A.: Structural equations with latent variables, vol. 210. John Wiley & Sons (1989). https://doi.org/10.1002/9781118619179" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.242, + 0.785, + 0.295 + ], + "angle": 0, + "content": "3. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020), https://dl.acm.org/doi/abs/10.5555/3495724.3495883" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.785, + 0.337 + ], + "angle": 0, + "content": "4. Carlini, N., Terzis, A.: Poisoning and backdooring contrastive learning. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=iC4UHbQ01Mp" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.785, + 0.378 + ], + "angle": 0, + "content": "5. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: 2017 IEEE Symposium on Security and Privacy (SP). pp. 39-57. IEEE Computer Society (2017). https://doi.org/10.1109/SP.2017.49" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.379, + 0.785, + 0.404 + ], + "angle": 0, + "content": "6. Chen, P., Liu, S., Zhao, H., Jia, J.: Gridmask data augmentation. arXiv preprint arXiv:2001.04086 (2020). https://doi.org/10.48550/arXiv.2001.04086" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.785, + 0.457 + ], + "angle": 0, + "content": "7. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020), https://dl.acm.org/doi/abs/10.5555/3524938.3525087" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.459, + 0.785, + 0.499 + ], + "angle": 0, + "content": "8. Chen, X., He, K.: Exploring simple siamese representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15750-15758 (2021). https://doi.org/10.1109/CVPR46437.2021.01549" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.5, + 0.785, + 0.553 + ], + "angle": 0, + "content": "9. Cho, J., Nam, G., Kim, S., Yang, H., Kwak, S.: Promptstyler: Prompt-driven style generation for source-free domain generalization. In: Proceedings of the IEEE international conference on computer vision. pp. 15702-15712 (2023). https://doi.org/10.1109/ICCV51070.2023.01439" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.554, + 0.785, + 0.595 + ], + "angle": 0, + "content": "0. Daunhawer, I., Bizeul, A., Palumbo, E., Marx, A., Vogt, J.E.: Identifiability results for multimodal contrastive learning. ICLR (2023), https://openreview.net/forum?id=U_2kuqoTcB" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.596, + 0.785, + 0.663 + ], + "angle": 0, + "content": "11. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020), https://openreview.net/forum?id=YicbFdNTTy" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.664, + 0.785, + 0.702 + ], + "angle": 0, + "content": "12. Fort, S.: Adversarial vulnerability of powerful near out-of-distribution detection. arXiv preprint arXiv:2201.07012 (2022). https://doi.org/10.48550/arXiv.2201.07012" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.704, + 0.785, + 0.757 + ], + "angle": 0, + "content": "13. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision 132(2), 581-595 (2024). https://doi.org/10.1007/s11263-023-01891-x" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.759, + 0.785, + 0.798 + ], + "angle": 0, + "content": "14. Ge, C., Huang, R., Xie, M., Lai, Z., Song, S., Li, S., Huang, G.: Domain adaptation via prompt learning. arXiv preprint arXiv:2202.06687 (2022). https://doi.org/10.48550/arXiv.2202.06687" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "15. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. In: International Conference on Learning Representations (2015). https://doi.org/10.48550/arXiv.1412.6572" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Cai et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.216 + ], + "angle": 0, + "content": "16. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020), https://dl.acm.org/doi/abs/10.5555/3495724.3497510" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.787, + 0.271 + ], + "angle": 0, + "content": "17. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022). https://doi.org/10.1109/CVPR52688.2022.01553" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.787, + 0.325 + ], + "angle": 0, + "content": "18. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020). https://doi.org/10.1109/CVPR42600.2020.00975" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.325, + 0.787, + 0.379 + ], + "angle": 0, + "content": "19. Hong, T., Guo, X., Ma, J.: Itmix: Image-text mix augmentation for transferring clip to image classification. In: 2022 16th IEEE International Conference on Signal Processing (ICSP). vol. 1, pp. 129-133. IEEE (2022). https://doi.org/10.1109/ICSP56322.2022.9965292" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.787, + 0.435 + ], + "angle": 0, + "content": "20. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021), https://proceedings.mlr.press/v139/jia21b.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.435, + 0.787, + 0.488 + ], + "angle": 0, + "content": "21. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19113-19122 (2023). https://doi.org/10.1109/CVPR52729.2023.01832" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.543 + ], + "angle": 0, + "content": "22. Kong, L., Xie, S., Yao, W., Zheng, Y., Chen, G., Stojanov, P., Akinwande, V., Zhang, K.: Partial disentanglement for domain adaptation. In: International Conference on Machine Learning. pp. 11455-11472. PMLR (2022), https://proceedings.mlr.press/v162/kong22a.htm1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.543, + 0.787, + 0.583 + ], + "angle": 0, + "content": "23. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.M.: Deeper, broader and artier domain generalization. In: Proceedings of the IEEE international conference on computer vision. pp. 5542-5550 (2017). https://doi.org/10.1109/ICCV.2017.591" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.787, + 0.624 + ], + "angle": 0, + "content": "24. Li, H., Wang, X., Zhang, Z., Yuan, Z., Li, H., Zhu, W.: Disentangled contrastive learning on graphs. Advances in Neural Information Processing Systems 34, 21872-21884 (2021), https://dl.acm.org/doi/10.5555/3540261.3541935" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.624, + 0.787, + 0.665 + ], + "angle": 0, + "content": "25. Li, Y., Liang, F., Zhao, L., Cui, Y., Ouyang, W., Shao, J., Yu, F., Yan, J.: Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In: International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.787, + 0.718 + ], + "angle": 0, + "content": "26. Liu, Y., Zhang, Z., Gong, D., Gong, M., Huang, B., van den Hengel, A., Zhang, K., Shi, J.Q.: Identifiable latent polynomial causal models through the lens of change. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=ia9fK01Vjq" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.787, + 0.759 + ], + "angle": 0, + "content": "27. Liu, Y., Zhang, Z., Gong, D., Gong, M., Huang, B., Hengel, A.v.d., Zhang, K., Shi, J.Q.: Identifying weight-variant latent causal models. arXiv preprint arXiv:2208.14153 (2022). https://doi.org/10.48550/arXiv.2208.14153" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.799 + ], + "angle": 0, + "content": "28. Liu, Y., Zhang, Z., Gong, D., Gong, M., Huang, B., Hengel, A.v.d., Zhang, K., Shi, J.Q.: Identifiable latent neural causal models. arXiv preprint arXiv:2403.15711 (2024). https://doi.org/10.48550/arXiv.2403.15711" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "29. Liu, Y., Zhang, Z., Gong, D., Gong, M., Huang, B., Zhang, K., Shi, J.Q.: Identifying latent causal content for multi-source domain adaptation. arXiv preprint arXiv:2208.14161 (2022). https://doi.org/10.48550/arXiv.2208.14161" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.362, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "CLAP: Contrastive Learning with Augmented Prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.19 + ], + "angle": 0, + "content": "30. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. In: International Conference on Learning Representations (2018), https://openreview.net/forum?id=rJzIBfZAb" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.233 + ], + "angle": 0, + "content": "31. Mahajan, D., Tople, S., Sharma, A.: Domain generalization using causal matching. In: International Conference on Machine Learning. pp. 7313-7324. PMLR (2021), https://proceedings.mlr.press/v139/mahajan21b.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.234, + 0.788, + 0.262 + ], + "angle": 0, + "content": "32. Mamooler, S.: Clip explainability. https://github.com/sMamooler/CLIP_Explainability, accessed: 2024-03-06" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.262, + 0.788, + 0.316 + ], + "angle": 0, + "content": "33. Mao, C., Geng, S., Yang, J., Wang, X., Vondrick, C.: Understanding zero-shot adversarial robustness for large-scale models. In: The Eleventh International Conference on Learning Representations (2022), https://openreview.net/forum?id=P4bXcawRi5J" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.318, + 0.788, + 0.358 + ], + "angle": 0, + "content": "34. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018). https://doi.org/10.48550/arXiv.1807.03748" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.361, + 0.788, + 0.416 + ], + "angle": 0, + "content": "35. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE international conference on computer vision. pp. 1406-1415 (2019). https://doi.org/10.1109/ICCV.2019.00149" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.418, + 0.788, + 0.486 + ], + "angle": 0, + "content": "36. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021), https://proceedings.mlr.press/v139/radford21a.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.543 + ], + "angle": 0, + "content": "37. Rahman, M.M., Fookes, C., Baktashmotlagh, M., Sridharan, S.: Multi-component image translation for deep domain generalization. In: 2019 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 579-588. IEEE (2019). https://doi.org/10.1109/WACV.2019.00067" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.545, + 0.788, + 0.601 + ], + "angle": 0, + "content": "38. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021), https://proceedings.mlrpress/v139/ramesh21a.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.602, + 0.788, + 0.657 + ], + "angle": 0, + "content": "39. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022). https://doi.org/10.1109/CVPR52729.2023.01389" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.714 + ], + "angle": 0, + "content": "40. Sanchez, E.H., Serrurier, M., Ortner, M.: Learning disentangled representations via mutual information estimation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 205-221. Springer (2020). https://doi.org/10.1007/978-3-030-58542-6_13" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.715, + 0.788, + 0.77 + ], + "angle": 0, + "content": "41. Von Kügelgen, J., Sharma, Y., Gresele, L., Brendel, W., Schölkopf, B., Besserve, M., Locatello, F.: Self-supervised learning with data augmentations provably isolates content from style. Advances in neural information processing systems 34, 16451-16467 (2021), https://d1.acm.org/doi/10.5555/3540261.3541519" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.772, + 0.788, + 0.84 + ], + "angle": 0, + "content": "42. Wei, J., Zou, K.: Eda: Easy data augmentation techniques for boosting performance on text classification tasks. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). pp. 6382-6388 (2019). https://doi.org/10.18653/v1/D19-1670" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Cai et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.216 + ], + "angle": 0, + "content": "43. Wortsman, M., Ilharco, G., Kim, J.W., Li, M., Kornblith, S., Roelofs, R., Lopes, R.G., Hajishirzi, H., Farhadi, A., Namkoong, H., et al.: Robust fine-tuning of zero-shot models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7959-7971 (2022). https://doi.org/10.1109/CVPR52688.2022.00780" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.272 + ], + "angle": 0, + "content": "44. Yang, M., Liu, F., Chen, Z., Shen, X., Hao, J., Wang, J.: Causalvae: Disentangled representation learning via neural structural causal models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9593-9602 (2021). https://doi.org/10.1109/CVPR46437.2021.00947" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.313 + ], + "angle": 0, + "content": "45. Yang, W., Mirzasoleiman, B.: Robust contrastive language-image pretraining against adversarial attacks. arXiv preprint arXiv:2303.06854 (2023). https://doi.org/10.48550/arXiv.2303.06854" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "46. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: IEEE International Conference on Computer Vision (ICCV) (2023). https://doi.org/10.1109/ICCV51070.2023.00355" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.41 + ], + "angle": 0, + "content": "47. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022). https://doi.org/10.1109/CVPR52688.2022.01631" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.452 + ], + "angle": 0, + "content": "48. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022). https://doi.org/10.1007/s11263-022-01653-1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.507 + ], + "angle": 0, + "content": "49. Zimmermann, R.S., Sharma, Y., Schneider, S., Bethge, M., Brendel, W.: Contrastive learning inverts the data generating process. In: International Conference on Machine Learning. pp. 12979-12990. PMLR (2021), https://proceedings.mlr.press/v139/zimmermann21a.html" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.507 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CLAP_ Isolating Content from Style through Contrastive Learning with Augmented Prompts/f676b4a9-3170-48e9-b0be-002172d2221e_origin.pdf b/2024/CLAP_ Isolating Content from Style through Contrastive Learning with Augmented Prompts/f676b4a9-3170-48e9-b0be-002172d2221e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..29a40a47cd7be21e4e3f6f29a945067227092f78 --- /dev/null +++ b/2024/CLAP_ Isolating Content from Style through Contrastive Learning with Augmented Prompts/f676b4a9-3170-48e9-b0be-002172d2221e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d28d7cca8700bf2e5beb4a7e2329a1362e07aca584a0a87df4549520992d8cf +size 3837615 diff --git a/2024/CLAP_ Isolating Content from Style through Contrastive Learning with Augmented Prompts/full.md b/2024/CLAP_ Isolating Content from Style through Contrastive Learning with Augmented Prompts/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b6062c04c52c1d277b6c1b436c93f8b491587bd7 --- /dev/null +++ b/2024/CLAP_ Isolating Content from Style through Contrastive Learning with Augmented Prompts/full.md @@ -0,0 +1,279 @@ +# CLAP: Isolating Content from Style through Contrastive Learning with Augmented Prompts + +Yichao Cai, Yuhang Liu, Zhen Zhang, and Javen Qinfeng Shi + +Australian Institute for Machine Learning, University of Adelaide, SA 5000, Australia {yichao.cai,yuhang.liu01,zhen.zhang02,javen.shi}@adelaide.edu.au + +Abstract. Contrastive vision-language models, such as CLIP, have garnered considerable attention for various downstream tasks, mainly due to the remarkable ability of the learned features for generalization. However, the features they learned often blend content and style information, which somewhat limits their generalization capabilities under distribution shifts. To address this limitation, we adopt a causal generative perspective for multimodal data and propose contrastive learning with data augmentation to disentangle content features from the original representations. To achieve this, we begin with exploring image augmentation techniques and develop a method to seamlessly integrate them into pretrained CLIP-like models to extract pure content features. Taking a step further, recognizing the inherent semantic richness and logical structure of text data, we explore the use of text augmentation to isolate latent content from style features. This enables CLIP-like model's encoders to concentrate on latent content information, refining the learned representations by pre-trained CLIP-like models. Our extensive experiments across diverse datasets demonstrate significant improvements in zero-shot and few-shot classification tasks, alongside enhanced robustness to various perturbations. These results underscore the effectiveness of our proposed methods in refining vision-language representations and advancing the state-of-the-art in multimodal learning. + +Keywords: Data Augmentation $\cdot$ Latent Variables $\cdot$ Disentanglement + +# 1 Introduction + +Vision-language models, exemplified by CLIP [36], have garnered substantial attention due to their exceptional generalization capabilities, achieved through the learned features, obtained by utilizing a cross-modal contrastive loss [20,25,36]. However, despite being pre-trained on extensive datasets, CLIP-like models fall short in disentangling latent content information and latent style information. Consequently, they are not immune to spurious correlations, i.e., style-related information is erroneously utilized to predict task-related labels. These limitations become evident in the presence of distribution shifts or adversarial attacks where spurious correlations often change across different environments. For examples, + +![](images/6e66cdac6ce80f08ff08ab337f4f8aa2c98ad0f5669134c6c15d54971900f099.jpg) +(a) Image augmentation + +![](images/dde32168462b738ed9a57d9dc1dd9a0e30491f46efda143c8b1e5327deb7f068.jpg) +(c) CAM visualization examples + +![](images/a5e203a1edbabb418ef483009400f31d58cf878cdd3d26c87a69176a88eb5137.jpg) +(b) Text augmentation +Fig. 1: Causal generative models of vision-language data. Image and text data are generated through distinct underlying deterministic processes, $\mathbf{g}_{\mathbf{x}}$ for images and $\mathbf{g}_{\mathbf{t}}$ for texts, derived from a unified latent space with latent content variables $\mathbf{c}$ and latent style variables $\mathbf{s}$ . Latent content $\mathbf{c}$ exclusively determines the sample label $\mathbf{y}$ . (a) Soft interventions on latent style variables $\mathbf{s}$ result in $\tilde{\mathbf{s}}$ , subsequently generating augmented images $\tilde{\mathbf{x}}$ . (b) Due to the same latent space, soft interventions on latent style variables $\mathbf{s}$ can also result in $\tilde{\mathbf{s}}$ , producing augmented text $\tilde{\mathbf{t}}$ . (c) A qualitative comparison of image features for zero-shot classification using "a photo of a [class]" prompts, visualized using class activation map (CAM) [32], demonstrates that while image augmentation can enhance CLIP features, the features obtained through text augmentation methods predominantly focus on the content. + +(1) a notable dependence on specific input text prompts has been reported for zero-shot capabilities [21, 47, 48]; (2) performance decline in few-shot scenarios has been observed in few-shot learning scenarios [13, 36]; and (3) susceptibility to adversarial attacks has been explored [33, 43, 45]. + +Taking a causal perspective, this work begins with a simple yet effective method, image augmentation, to disentangle content and style information within the learned representations of CLIP-like models. This approach is inspired by recent advancements in theoretical development in causal representation learning [41], which demonstrate that augmented image can be interpreted as a result of soft interventions on latent style variables, as depicted in Fig. 1a. Such augmentation results in a natural data pair where content information remains unchanged while style information changes. Consequently, using contrastive learning, it becomes feasible to isolate the invariant content information from the variant style information. Motivated by this theoretical advancement, we propose a practical method to incorporate image augmentation into CLIP-like models to extract content information from the original learned features. Specifically, a disentangled network is designed to fine-tune the pre-trained CLIP model by using a contrastive loss with image augmentation. + +Despite the advancements made in disentangling content and style information from the original features learned by CLIP-like models through image augmentation, we recognize an inherent limitation: it is generally challenging to + +design adequate image augmentations to ensure all style factors change in an image. Theoretically, disentangling content and style information necessitates changes in all style factors [41]. However, inducing sufficient changes in latent style through image augmentation poses challenges due to the high dimensionality and complexity of style information in image data. Achieving significant style variation via artificially designed image augmentation techniques, such as transforming a photograph of a dog into a sketch while preserving content but dramatically altering style, is notably difficult. + +Taking a further step, rather than relying on image augmentation, we explore the use of text augmentation to disentangle latent content and style factors. This shift is motivated by two key observations: 1) Vision and language data share the same latent space. Therefore, text augmentation can also be utilized to induce changes in latent style factors instead of image augmentation. 2) Text data inherently possesses high semanticity and logical structure, making it more amenable to property-wise manipulation compared to image data. Consequently, implementing sufficient style changes through text augmentation is more feasible than image augmentation, contributing to isolating content from style information, see Fig. 1c for visual comparison. For instance, transforming text from "a photo of a dog" to "a sketch of a dog" is straightforward in the language modality, whereas achieving a similar transformation in image data is challenging. Inspired by these observations, we posit that introducing style variations through text augmentation, as illustrated Fig. 1b, provides a more effective approach for learning vision-language content features than relying on image augmentation. + +In summary, our contributions include: (1) Aimed at disentangling latent content and style factors to refine vision-language features of pre-trained CLIP-like models, we propose constractive learning with data augmentation to fine tune the original features of pre-trained CLIP-like models from a causal perspective. (2) We present a novel method customized for pre-trained CLIP-like models. This method leverages a disentangled network, which is trained using contrastive learning with image augmentation, to extract latent content features from the learned features provided by image encoder of CLIP-like models. (3) We propose Contrastive Learning with Augmented Prompts (CLAP), to extract latent content features from representations of CLIP-like models. It begins by training a disentangled network using the pre-trained text encoder of CLIP-like models and text augmentation. Subsequently, the trained disentangled network is transferred to the image encoder of CLIP-like models. (4) Experiments conducted on a large real dataset demonstrate the effectiveness of the proposed image augmentation and text augmentation in terms of zero-shot and few-shot performance, as well as robustness against perturbations. + +# 2 Related Work + +Contrastive Vision-Language Models Using a cross-modal contrastive loss, CLIP [36] revolutionarily introduced a scalable contrastive vision-language model + +by leveraging a large corpus of internet-sourced image-text pairs, demonstrating unprecedented zero-shot learning capabilities and exceptional generalization ability across datasets and supporting numerous downstream tasks [38]. ALIGN [20] expanded the scale of contrastive vision-language modeling, training on up to one billion image-text pairs while integrating the vision transformer's self-attention mechanism [11], which further enhanced performance. Despite their successes, CLIP-like models exhibit sensitivity to input text prompts [21, 48], leading to variable performance across different prompts. Efforts to mitigate this prompt sensitivity through prompt learning and engineering [9, 14, 21, 47, 48] focus on customizing prompts for specific tasks but do not fundamentally enhance CLIP's representations. Furthermore, CLIP-like models are vulnerable to adversarial attacks [4, 12], with current strategies [33, 45] involving adversarial-natural image pairs to improve resilience. Our work diverges from task-specific approaches by aiming to enhance CLIP's representations from a disentanglement perspective, addressing the aforementioned issues inherent in CLIP-like models. + +Disentangled Representation Learning Aimed at segregating intrinsic latent factors in data into distinct, controllable representations, disentangled representation learning benefits various applications [24,40,44]. Specifically, in classification tasks, it's shown that enhancing the model's performance and robustness against data distribution perturbations can be achieved by more effectively disentangling invariant content variables, without needing to identify all intrinsic latent variables completely [22,26-28]. Within single modalities, studies such as [49] have illustrated that contrastive learning [7,16,18] can potentially reverse the data generative process, aiding in the separation of representations. Furthermore, [41] suggest that image augmentation can help isolate content variables from the latent space through significant stylistic changes. [19] employs mixture techniques for data augmentation, enabling more abundant cross-modal matches. Diverging from these methods, our approach focuses on employing text augmentation to disentangle latent content variables, introducing a unique approach to learn refined vision-language representations. + +# 3 A Causal Generative Model for Multi-Modal Data + +To understand pretrained CLIP-like models, we investigate the underlying causal generative process for vision-language data. We consider the following causal generative model as depicted in Fig. 1. In the proposed model, the shared latent space ruling vision and language data is divided into two distinct sub-spaces: one corresponding to the latent content variables $\mathbf{c}$ and the other to the latent style variables $\mathbf{s}$ . The latent content variables are posited to determine the object label $\mathbf{y}$ , a relationship corroborated by prior studies [22, 29, 31]. Furthermore, to elucidate the correlation between the latent style variable $\mathbf{s}$ and the object variable $\mathbf{y}$ , our model incorporates the premise that the latent content variable $\mathbf{c}$ causally influences the latent style variable $\mathbf{s}$ , in concordance with the principles of causal representation learning highlighted in recent literature [10, 29, 41]. Additionally, + +considering the diversity between image data and text data, where information in image data is typically much more details while information in text data tends to be more logically structured nature, we posit distinct causal mechanisms for the generation processes. Our causal generative model is formulated as following structural causal models [2]: + +$$ +\mathbf {s} := \mathbf {g} _ {\mathbf {s}} (\mathbf {c}), \mathbf {x} := \mathbf {g} _ {\mathbf {x}} (\mathbf {c}, \mathbf {s}), \mathbf {t} := \mathbf {g} _ {\mathbf {t}} (\mathbf {c}, \mathbf {s}), \mathbf {y} := \mathbf {g} _ {\mathbf {y}} (\mathbf {c}). \tag {1} +$$ + +In Eq. (1), the style variable $\mathbf{s}$ is causally influenced by the content via $\mathbf{g_s}$ ; $\mathbf{x}$ and $\mathbf{t}$ denote visual and textual data, respectively. Both visual and textual data are causally produced by the shared latent variables $\mathbf{c}$ and $\mathbf{s}$ through distinct, reversible generative processes: $\mathbf{g_x}$ for images and $\mathbf{g_t}$ for text data, respectively. The label $\mathbf{y}$ of a sample is exclusively determined by the content variable $\mathbf{c}$ via $\mathbf{g_y}$ . For simplicity, exogenous noises are implicitly assumed but not explicitly represented in the causal generative model's formulation, aligning with the common understanding that each latent variable is influenced by exogenous noise. + +Recent seminal work in [41] has demonstrated that the latent content variable $\mathbf{c}$ can be identified up to block identifiability (i.e., $\mathbf{c}$ can be isolated from style variable $\mathbf{s}$ ), by requiring all latent style variables to change (e.g., soft interventions on all latent style variables). This change can be achieved through image augmentation, i.e., the augmented image $\tilde{\mathbf{x}}$ can be interpreted as a generative result of $\tilde{\mathbf{s}}$ , which is produced through soft interventions on original latent style variables $\mathbf{s}$ . Despite such theoretical advancement, the practical implementation of this theoretical result within CLIP-like models remains unclear. In this study, we propose a practical method to disentangle content and style information within CLIP-like models by employing image augmentation, as detailed in Sec. 4.1. Moreover, we recognize that implementing sufficient changes on all latent style variables $\mathbf{s}$ through text augmentation is more feasible than image augmentation, due to high semanticity and logical structure in text data, we delve into the use of text augmentation to separate content information from style information, as discussed in Sec. 4.2. + +# 4 Isolating Content from Style with Data Augmentation + +In this section, we propose the employment of data augmentation to extract content information from the learned features in pre-trained CLIP-like models. Essentially, data augmentation facilitates the alteration of style factors while preserving content factors. Consequently, leveraging contrastive learning enables the segregation of content information from style information. We delve into two distinct forms of data augmentation, namely image augmentation (Sec. 4.1) and text augmentation (Sec. 4.2). + +# 4.1 Isolating Content from Style with Augmented Images + +While recent studies (von et al., 2021) have offered assurance regarding the disentanglement of content and style through contrastive learning with data augmentation, it remains unclear how these theoretical findings can be applied to + +![](images/d51d2e0024c09fbe32179e48985340220767f6226601d8137f87eed6017a0136.jpg) +(a) Image augmentation + +![](images/db85c17d17159a5651a4daf6640b7096bd24c40f6d9f8ce12b44872b3fa4b80f.jpg) +(b) Text augmentation +Fig. 2: Refining CLIP through data augmentation. (a) Training involves a disentangled network $\mathbf{f_c}$ , utilizing contrastive loss on original and augmented image pairs $\mathbf{x}$ and $\tilde{\mathbf{x}}$ , with CLIP's image encoder $\mathbf{f_x^*}$ holding frozen gradients. (b) More efficient content feature learning is achieved through contrastive learning with augmented text prompts $\mathbf{t}$ and $\tilde{\mathbf{t}}$ , using the fixed text encoder $\mathbf{f_t^*}$ of CLIP. (c) Inference stage: The trained disentangled network $\mathbf{f_c^*}$ integrates with CLIP's text and image encoders, $\mathbf{f_t^*}$ and $\mathbf{f_x^*}$ , to enable zero-shot inference for an input image $\mathbf{x}$ and class names $\mathbf{t_1}$ to $\mathbf{t_n}$ . + +![](images/24c7f87d886d6e99a42e77d789d6bd111c1314f0329c5c597871cecd865e977f.jpg) +(c) Zero-shot inference + +the realm of vision-language models. We convert the theoretical findings into CLIP-like models in the following. The theoretical findings suggest using InfoNCE loss [34] to extract content information, as outlined below: + +$$ +\mathcal {L} (\mathbf {f}; \{\mathbf {x} _ {i}, \tilde {\mathbf {x}} _ {i} \} _ {i = 1} ^ {b}, \tau) = - \frac {1}{b} \sum_ {i = 1} ^ {b} \log \frac {\exp [ \langle \mathbf {f} (\mathbf {x} _ {i}) , \mathbf {f} (\tilde {\mathbf {x}} _ {i}) \rangle / \tau ]}{\sum_ {j = 1} ^ {b} \exp [ \langle \mathbf {f} (\mathbf {x} _ {i}) , \mathbf {f} (\tilde {\mathbf {x}} _ {j}) \rangle / \tau ]}, \tag {2} +$$ + +where $\{\mathbf{x}_i\}_{i=1}^b$ represents a batch of $b$ samples from the training dataset, $\mathbf{f}(\mathbf{x}_i)$ denotes sample $\mathbf{x}_i$ 's features through model $\mathbf{f}$ , $\tilde{\mathbf{x}}_i$ is the augmented counterpart of $\mathbf{x}_i$ , and $\langle \mathbf{z}_1, \mathbf{z}_2 \rangle$ represents the cosine similarity between two feature vectors, $\mathbf{z}_1$ and $\mathbf{z}_2$ , and $\tau$ represents the temperature parameter influencing the loss. + +We extend it to refine pre-trained vision-language models, utilizing contrastive learning with augmented images (hereinafter referred to as "Im.Aug"). As illustrated in Fig. 2a, we train a disentangled network on top of CLIP's pretrained image encoder. To enhance training efficiency and the usability of the proposed method, we freeze the pre-trained image encoder. Based on an InfoNCE loss, the learning objective of Im.Aug is formulated as follows: + +$$ +\mathbf {f} _ {\mathbf {c}} ^ {*} = \underset {\mathbf {f} _ {\mathbf {c}}} {\operatorname {a r g m i n}} \underset {\{\mathbf {x} _ {i} \} _ {i = 1} ^ {b} \in \mathcal {D} _ {\mathbf {x}}} {\mathbb {E}} \mathcal {L} \left(\mathbf {f} _ {\mathbf {c}} \circ \mathbf {f} _ {\mathbf {x}} ^ {*}; \left\{\mathbf {x} _ {i}, \tilde {\mathbf {x}} _ {i} \right\} _ {i = 1} ^ {b}, \tau\right), \tag {3} +$$ + +where $\mathcal{D}_{\mathbf{x}}$ denotes the training image dataset and $b$ represents the batch size, $\mathbf{f_c}$ is the disentangled network undergoing training. The pre-trained CLIP image encoder is represented by $\mathbf{f_x^*}$ , with the asterisk "***" signifying that the model weights remain fixed. The variable $\mathbf{x}_i$ refers to an image sampled from $\mathcal{D}_{\mathbf{x}}$ , and $\tilde{\mathbf{x}}_i$ is its augmented view. + +The composition of the training dataset $\mathcal{D}_{\mathbf{x}}$ , the image augmentation techniques used, the structure of the disentangled network $\mathbf{f_c}$ , and the utilization of $\mathbf{f_c^*}$ post-training are detailed in the following subsections. + +Data Synthesis and Image Augmentation To generate training image data, we combine class names with various image and object attributes to create text + +Table 1: Template-based prompts. Attributes used to generate text prompts follow the structured format "a [art style] [image type] of a [object size] [object color] [class]", where "[class]" represents the class names. + +
Object ColorObject SizeImage TypeArt Style
yellow, green, black, blue, multicolored, orange, red, white, brown, purplelarge, small, normal sizedpainting, cartoon, infographic, sketch, photograph, clipart, mosaic art, sculpturerealistic, impressionistic
+ +prompts for each class. Using a stable diffusion model [39], we produce synthetic images that comprise our training dataset $\mathcal{D}\mathbf{x}$ . The creation of template prompts for stable diffusion is based on attributes such as object size, color, image type, and art style. As detailed in Tab. 1, the attributes include 10 colors and 3 sizes for objects, and 8 types and 2 art styles for images. By assembling these attributes into prompts like "a [art style] [image type] of a [object size] [object color] [class]", we generate 480 unique texts for each class, from which one image per prompt is synthesized. Further details on image synthesis and examples are available in Appendix B.1. For the image augmentation procedures, we adopt techniques commonly used in contrastive learning practice [7,8,41], specifically random cropping and color distortion. + +Disentangled Network Structure Since the training process is based on CLIP's pre-trained lower-dimensional features, our disentangled network adopts a multi-layer perceptron (MLP) architecture. To fully benefit from the pretrained CLIP text encoder, we construct a residual MLP featuring a zero-initialized projection, acting as the disentangled network, as depicted in Fig. 3. This design enables learning directly from the pre-trained representation space, avoiding a random starting point, inspired by ControlNet's zero-conv operation [46], which we adapt to a zero-linear operation within our residual MLP. + +Within this architecture, the main branch includes a zero-initialized, bias-free linear layer positioned subsequent to the combination of a SiLU activation and a normally initialized linear layer. Conventionally, the dimensions of features before the initial linear layer, situated between the first and second linear layers, and following the second linear layer, are named as the input $d_{in}$ , latent $d_{mid}$ , and output $d_{out}$ dimensions, respectively. To rectify any mismatches between the input and output dimensions, the network employs nearest-neighbor downsampling within the shortcut path, thereby ensuring both alignment and the preservation of sharpness for the input features. During the inference stage, a weighting parameter $\alpha > 0$ is introduced to modulate the portion of features emanating from the main branch before their integration with the input features, whereas this parameter remains constant at 1 throughout the training phase. + +Inference After training, the disentangled network $\mathbf{f}_{\mathbf{c}}^{*}$ is utilized following CLIP's image encoder to extract visual content features. Moreover, given that + +![](images/26bd3fc57531fc0688d8e6ba852c119d57c9e15fd1dd84bdc1a7e609acd3bb5e.jpg) +Fig. 3: Structure of the disentangled network. The architecture encompasses a residual block featuring a zero-initialized, bias-free linear layer to commence optimization from the input feature space. When the input and output dimension differ, a downsampling operation is utilized to achieve alignment. During inference, a scalar parameter $\alpha$ balance the main branch and input features before combination. + +vision-language data generation is rooted in a unified latent space, as depicted in Sec. 3, $\mathbf{f}_{\mathbf{c}}^{*}$ can be seamlessly integrated with CLIP's image and text encoders to enhance zero-shot capabilities. As shown in Fig. 2c, for an image $\mathbf{x}$ , the operation is formulated as the composition function $\mathbf{f}_{\mathbf{c}}^{*} \circ \mathbf{f}_{\mathbf{x}}^{*}(\mathbf{x})$ , and similarly, for a text $\mathbf{t}$ , as $\mathbf{f}_{\mathbf{c}}^{*} \circ \mathbf{f}_{\mathbf{t}}^{*}(\mathbf{t})$ . This integration preserves CLIP's zero-shot functionality while achieving refined features through the improved disentanglement of content. + +# 4.2 Isolating Content from Style with Augmented Prompts + +Despite progress in disentangling content and style via image augmentation, adequately altering all style factors in an image remains challenging due to the high dimensionality and complexity of style information in images. Achieving substantial style changes through augmentation, essential for complete disentanglement [41], is difficult with existing image augmentation techniques. On the contrary, text data inherently possesses high semanticity and logical structure, making it more amenable to property-wise manipulation compared to image data. To further exploring the disentanglement of content, we propose Contrastive Learning with Augmented Prompts (CLAP). + +As depicted in Fig. 2b, CLAP employs an InfoNCE loss to train a disentangled network atop CLIP's pre-trained text encoder, keeping the encoder's gradients fixed, similar to Im.Aug. Leveraging the simpler structure of text, the template-based prompts previously utilized for synthesizing images now serve as the training text dataset, denoted by $\mathcal{D}_{\mathbf{t}}$ . Utilizing the same disentangled network as in Im.Aug, the learning objective of CLAP is outlined as follows: + +$$ +\mathbf {f} _ {\mathbf {c}} ^ {*} = \underset {\mathbf {f} _ {\mathbf {c}}} {\operatorname {a r g m i n}} \underset {\left\{\mathbf {t} _ {i} \right\} _ {i = 1} ^ {b} \in \mathcal {D} _ {\mathbf {t}}} {\mathbb {E}} \mathcal {L} \left(\mathbf {f} _ {\mathbf {c}} \circ \mathbf {f} _ {\mathbf {t}} ^ {*}; \left\{\mathbf {t} _ {i}, \tilde {\mathbf {t}} _ {i} \right\} _ {i = 1} ^ {b}, \tau\right) + \lambda \mathcal {L} \left(\mathbf {f} _ {\mathbf {c}} \circ \mathbf {f} _ {\mathbf {t}} ^ {*}; \left\{\mathbf {t} _ {i} ^ {c}, \tilde {\mathbf {t}} _ {i} \right\} _ {i = 1} ^ {b}, 1\right), \tag {4} +$$ + +where $\mathbf{f}_{\mathbf{t}}^{*}$ denotes the pre-trained CLIP text encoder. The term $\mathbf{t}_i$ references a text prompt from $\mathcal{D}\mathbf{t}$ , and $\tilde{\mathbf{t}}_i$ represents its augmented view, produced via prompt augmentation techniques. On the equation's right side, $\mathbf{t}_i^c$ specifies the class name associated with the text prompt $\mathbf{t}_i$ . This strategy aims to enhance variations between prompt pairs, especially in cases where the text dataset $\mathcal{D}_{\mathbf{t}}$ + +Table 2: Prompt augmentation techniques. Various augmented views are generated from an original text prompt using specific augmentation techniques: OSD (Object Size Deletion), OCD (Object Color Deletion), ITD (Image Type Deletion), ASD (Art Style Deletion), and SPO (Swapping Prompt Order). + +
OriginalOSDOCDITDASDSPO
a realistic painting of a large red cara realistic painting of a red cara realistic painting of a large cara realistic of a large red cara painting of a large red cara large red car in a realistic painting
+ +has a very limited number of samples. Here, $\lambda$ serves for adjusting the second term's importance in the total loss function. All other symbols in Eq. (4) maintain their definitions as described earlier. + +After training, the learned disentangled network is seamlessly integrated with both of CLIP's encoders to extract content representations, as depicted in Fig. 2c. + +Prompt Augmentation To ensure text prompts undergo stylistic changes without compromising their content, we have developed specific augmentation techniques for synthetic text prompts. Drawing inspiration from Easy Data Augmentation (EDA) techniques [42], we adapted the Random Deletion (RD) and Random Swap (RS) techniques from EDA, customizing them to suit our prompt structure. To avoid inadvertently altering the content by introducing new object names or changing the core idea of a text prompt, our augmentation methods do not include random word insertions or replacements. Our primary augmentation techniques are Object Size Deletion (OSD), Object Color Deletion (OCD), Image Type Deletion (ITD), Art Style Deletion (ASD), and Swapping Prompt Order (SPO), each applied with a certain probability, as detailed in Tab. 2. Additionally, for down-stream datasets with few categories, to rich the population of training samples, we use an additional augmentation, named IGN (Inserting Gaussian Noise). Following the initializing protocol of prompt learning methods [47,48], we insert a zero-mean Gaussian noise with 0.02 standard deviation with a noise length equals to 4, to the tokenized prompts. + +Intuitively, these prompt augmentation methods parallel random masking techniques used in image augmentation [6,17]. However, prompt augmentations are more effective and precise than their image counterparts. This effectiveness arises because prompt augmentations can specifically target and eliminate a particular style element without impacting the content, whereas image masking, operating at the pixel or patch level, might inadvertently damage content information or lead to insufficient style changes. + +# 5 Experiments + +We conduct three primary experiments to assess our method: (1) zero-shot evaluation with diverse prompts to gauge zero-shot performance and its robustness + +to prompt perturbations; (2) linear probe tests on few-shot samples to evaluate the efficacy of the learned representations in few-shot settings; and (3) adversarial attack assessments on zero-shot and one-shot classifiers to determine their resistance to adversarial threats. We further conduct an ablative study on hyperparameters, explore the impact of different prompt augmentation combinations and various sources of training prompts on CLAP's performance, and replicate experiments across different CLIP model sizes. + +# 5.1 Experimental Setup + +Implementation. Im.Aug and CLAP are implemented using the ViT-B/16 CLIP model and executed on an NVIDIA RTX 3090 GPU. To ensure reproducibility, the random seed for all stochastic processes is fixed at 2024. More information on implementation details is provided in Appendix A.1. + +Datasets. CLAP is assessed across four multi-domain datasets to examine its performance in varied environments: PACS [23] (4 domains, 7 categories), VLCS [1] (4 domains, 5 categories), OfficeHome [37] (4 domains, 65 categories), and DomainNet [35] (6 domains, 345 categories). For conciseness, we present average results across the domains for each dataset. Detailed experimental outcomes for each domain within these datasets are provided in Appendix A.4. + +Compute efficiency. CLAP demonstrates faster convergence and shorter training times compared to Im.Aug. For CLAP, training on the PACS and VLCS datasets is completed in roughly 11 minutes, OfficeHome in approximately 14 minutes, and DomainNet in about 47 minutes. In contrast, Im.Aug requires around 16 minutes for PACS and VLCS, 50 minutes for OfficeHome, and 3.3 hours for DomainNet. Both Im.Aug and CLAP maintain CLIP's inference efficiency due to the disentangled network's efficient two-layer MLP structure. + +# 5.2 Main Results + +Zero-Shot Performance To assess zero-shot capabilities, CLAP undergoes evaluation using three specific fixed prompts: ZS(C), utilizing only the class name within "[class]"; ZS(PC), with the format "a photo of a [class]"; and ZS(CP), structured as "a [class] in a photo". To thoroughly examine zero-shot proficiency, a dynamic prompt, ZS(NC), formatted as "[noise][class]", is also used, where "[noise]" signifies the introduction of Gaussian noise characterized by a mean of 0 and a standard deviation of 0.02. + +As presented in Tab. 3, CLAP surpasses both CLIP and Im.Aug across all evaluated prompts for every dataset. Unlike the uniform enhancement in zero-shot performance CLAP achieves over CLIP, Im.Aug displays inconsistent results. A closer examination reveals CLAP's superiority over CLIP is especially significant for the dynamic ZS(NC) prompt. This demonstrates CLAP's effectiveness in significantly improving zero-shot performance compared to the original CLIP representations. + +In assessing the model's robustness to prompt perturbations, we examine the variances in zero-shot performance across different prompts by analyzing the + +Table 3: Zero-shot results across three distinct prompts: "C" for "[class]", "CP" for "a [class] in a photo", "PC" for "a photo of a [class]", and a dynamic prompt "NC" for "[noise][class]" showcase that CLAP consistently outperforms CLIP's zero-shot performance across all datasets, whereas image augmentation exhibits mixed outcomes. + +
PromptMethodZero-shot performance, avg. top-1 acc. (%) (↑)
PACSVLCSOff.HomeDom.NetOverall
ZS(C)CLIP95.776.479.857.877.4
Im.Aug96.579.577.051.576.1
CLAP97.282.681.058.779.9
ZS(CP)CLIP95.282.079.557.078.4
Im.Aug96.382.975.850.776.4
CLAP97.383.480.558.079.8
ZS(PC)CLIP96.182.482.557.779.7
Im.Aug96.583.078.651.677.4
CLAP97.283.483.059.080.6
ZS(NC)CLIP90.868.371.551.070.4
Im.Aug94.873.167.544.069.9
CLAP97.281.073.552.676.1
+ +range $(R)$ and standard deviation $(\delta)$ of results derived from ZS(C), ZS(CP), and ZS(PC). Additionally, we investigate the decline $(\varDelta_{(NC)})$ in performance from ZS(C) to ZS(NC) as a broad indicator of resilience to noised prompts. + +As presented in Tab. 4, CLAP significantly reduces the variance in zero-shot performance across various testing prompts, evidenced by markedly lower values of $\delta$ and $R$ , and a less pronounced decrease in performance with a noised prompt, in contrast to Im.Aug and the baseline representations of CLIP. Although Im.Aug aids in reducing performance variance to some extent, its efficacy is notably inferior to that of CLAP. These findings highlight CLAP's enhanced robustness in maintaining consistent zero-shot performance across a diverse array of prompts. + +Few-Shot Performance We conduct evaluations of 1-shot, 4-shot, 8-shot, and 16-shot linear probes across each domain within the four datasets. As illustrated in Fig. 4, CLAP significantly outperforms both CLIP and Im.Aug in few-shot learning scenarios. Notably, in the 1-shot setting CLAP achieves performance improvements over the linear-probe CLIP model by margins of $+10\%$ , $+3.5\%$ , $+2.5\%$ , and $+1.5\%$ on the PACS, VLCS, OfficeHome, and DomainNet datasets, respectively. These improvements are especially significant in comparison to the gains observed with Im.Aug counterparts, underpinning CLAP's efficacy in few-shot scenarios. For detailed quantitative results, please refer to Appendix A.4. + +Adversarial Performance To assess adversarial robustness, zero-shot (ZS(C)) and one-shot classifiers are evaluated against prominent adversarial attack methods, such as FGSM [15], PGD [30], and CW [5], by generating adversarial sam + +Table 4: CLAP more effectively reduces zero-shot performance variance across prompts than image augmentation, with $R$ and $\delta$ indicating the range and standard deviation for ZS(C), ZS(PC), and ZS(PC). The decrease $\Delta_{(NC)}$ from ZS(C) to ZS(NC) highlights CLAP's enhanced robustness against prompt perturbations. + +
MetricMethodPerformance variance, avg. top-1 acc. (%) (↓)
PACSVLCSOff.HomeDom.NetOverall
RCLIP0.96.13.10.82.7
Im.Aug0.13.62.80.91.9
CLAP0.10.82.51.01.1
δCLIP0.42.81.40.41.2
Im.Aug0.11.71.20.40.8
CLAP0.00.41.10.40.5
Δ(NC)CLIP4.98.18.36.87.0
Im.Aug1.66.49.57.56.3
CLAP0.01.67.56.13.8
+ +![](images/ef84eab710a70b77efa730ef360f20d3f62c4b3f3dfced28ba3bb86ea520a6a1.jpg) +Fig. 4: Few-shot linear probe comparisons of image-encoder features show that CLAP enhances CLIP's few-shot performance more effectively than Im.Aug. In the accompanying figure, "ZS" indicates the zero-shot performance using a "[class]" prompt. + +![](images/83e8ef217dfe60fa8be48905b67747d5aa0757bec0b8f4ab5486c1f7f0431d96.jpg) + +![](images/b93b693869ec8d910eceebeb10f6de19c4ffdbadeb5824ceb7156ea32d94d066.jpg) + +![](images/68af5ad4d6da7adb0c2a6088c6da0ff56934239eb540773eb0c03fd2fdab6acd.jpg) + +plies for testing. For FGSM, 1 adversarial iteration is employed, whereas for PGD and CW, 20 iterations are used, all with an epsilon of 0.031. As indicated in Tab. 5, classifiers utilizing CLAP representations demonstrate superior resilience to these adversarial attacks compared to those based on CLIP representations. Across the four datasets, CLAP's zero-shot and 1-shot classifiers surpass CLIP by margins of $+7.6\%$ and $+8.5\%$ against FGSM, $+1.0\%$ and $+11.7\%$ against PGD-20, and $+1.1\%$ and $+2.3\%$ against CW-20, respectively. These figures notably exceed the performance improvements of $+4.4\%$ and $+4.6\%$ against FGSM, $+0.3\%$ and $+6.2\%$ against PGD-20, and $0\%$ and $+1.3\%$ against CW-20 achieved by Im.Aug. The result suggests that CLAP efficiently enhances robustness against adversarial attacks in both zero-shot and one-shot scenarios. + +# 5.3 More Analysis + +t-SNE Visualization In our t-SNE visualizations, we examine the representations of CLIP, Im.Aug, and CLAP for all images within the Art Painting domain of the PACS dataset. Fig. 5 shows that CLAP's image representations display a marked inter-class separation and tighter intra-class clustering than those of + +Table 5: Image augmentation and CLAP both enhance CLIP's zero-shot with the "[class]" prompt and 1-shot robustness against adversarial attacks, with CLAP showing greater improvements. + +
Setting MethodAvg. top-1 acc. (%) under adversarial attacks(↑)
FGSMPGD-20CW-20
PACSVLCSO.H.D.N.PACSVLCSO.H.D.N.PACSVLCSO.H.D.N.
ZS(C)CLIP86.865.657.922.529.12.010.110.727.41.57.47.6
Im.Aug88.069.655.137.931.32.110.49.029.41.77.05.8
CLAP88.771.958.544.230.83.210.611.229.82.38.18.0
1-shotCLIP66.745.234.322.534.816.05.611.318.90.74.53.2
Im.Aug79.447.137.123.555.216.18.512.523.20.95.13.4
CLAP89.652.237.123.973.421.27.412.527.01.15.03.5
+ +![](images/bde6823173755b0d7136ddf77d7cc069d97ecb91807e9627482ca5d65737a5f9.jpg) +(a) CLIP + +![](images/baeaa687e75002e296a8c7a2f3c3461f0b160ba5e229cf058e61fddf7816bd28.jpg) +(b) Im.Aug +Fig. 5: t-SNE visualizations of all images in the Art Painting of PACS dataset show CLAP outperforms the original CLIP and Im.Aug, with clearer inter-class distinctions and tighter intra-class clusters. + +![](images/b6856d7e0f9ca1e0d49d20f38781276f54f0eb1aab929c728b026ca9b8728e4f.jpg) +(c) CLAP + +CLIP and Im.Aug. This observation suggests that CLAP's representations are more closely tied to content information and less influenced by style information, in contrast to the other two. + +Ablations In Fig. 6, we assess the zero-shot capabilities of our model using two distinct prompts, ZS(C) and ZS(PC), on the VLCS dataset. This analysis forms part of an ablative study aimed at understanding the influence of various hyperparameters on model performance. Specifically, we examine: the dimensions of the latent layer within the MLP of the disentangled network, as illustrated in Fig. 6a; the temperature parameter $(\tau)$ in the loss function, as depicted in Fig. 6b; and the weight coefficient $(\alpha)$ during the inference stage, as shown in Fig. 6c. Our findings indicate that CLAP consistently enhances zero-shot performance across all tested configurations for both prompts, while also significantly reducing the gap between the performances elicited by each prompt. These results underscore the efficacy of CLAP in accommodating a wide range of hyper-parameters. + +Prompt Augmentation Combinations We explore diverse combinations of our tailored prompt augmentation methods and examine Easy Data Augmentation (EDA) techniques [42] on the VLCS dataset. Each tested technique showcases CLAP's enhancements over CLIP, with details available in Appendix A.2. + +![](images/63de70ec9fd2118f1ca643f09b4a9c745e0844cabf7c0ffc7c53fd7b175694b7.jpg) +Fig. 6: We conduct ablative study on hyper-parameter choices on the VLCS dataset, including latent dimensions, $\tau$ values, and $\alpha$ values during the inference stage. CLAP continuously enhance CLIP's performance throughout the tested values. + +![](images/0803b3e90ceebc3f158b3216f06cf6745df399def7712334b047c53718ff8c82.jpg) + +![](images/5ffbb7fdae4d2b46c073f4e7d612fbd06af3fdce918f9a750b01810b40981b46.jpg) + +Prompt Sources We assess the impact of different training prompt formats, originating from various synthetic sources, on the performance of the VLCS dataset, incorporating EDA techniques. Our evaluation includes our template-based prompts, LLM-generated prompts by ChatGPT-3.5 [3] (with the generation process detailed in Appendix B.2), prompts structured as "a [random] style of [class]," where "[random]" is filled with terms from a random word generator $^2$ , and prompts produced using the PromptStyler method [9]. The findings indicate that the training prompts with simpler forms tend to yield better performance, with detailed quantitative results presented in Appendix A.3. + +Experiments on Different Model Scales In our repeated experiments assessing zero-shot performance on the ViT-L/14 and ResNet50x16 pre-trained with CLIP, we consistently find that CLAP improves zero-shot performance while also reducing performance variances. This consistent observation underscores CLAP's effectiveness in enhancing the quality of CLIP representations. For quantitative details supporting these findings, please see the Appendix C. + +# 6 Conclusion + +To enhance pre-trained CLIP-like models, this study delves into disentangling latent content variables. Through a causal analysis of the underlying generative processes of vision-language data, we discover that training a disentangled network in one modality can effectively disentangle content across both modalities. Given the high semantic nature of text data, we identify that disentanglement is more achievable within the language modality through text augmentation interventions. Building on these insights, we introduce CLAP (Contrastive Learning with Augmented Prompts) to acquire disentangled vision-language content features. Comprehensive experiments validate CLAP's effectiveness, demonstrating significant improvements in zero-shot and few-shot performance, and enhancing robustness against perturbations. We anticipate that our work will inspire further exploration into disentangling latent variables within vision-language models. + +# References + +1. Albuquerque, I., Monteiro, J., Darvishi, M., Falk, T.H., Mitliagkas, I.: Generalizing to unseen domains via distribution matching. arXiv preprint arXiv:1911.00804 (2019). https://doi.org/10.48550/arXiv.1911.00804 +2. Bollen, K.A.: Structural equations with latent variables, vol. 210. John Wiley & Sons (1989). https://doi.org/10.1002/9781118619179 +3. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020), https://dl.acm.org/doi/abs/10.5555/3495724.3495883 +4. Carlini, N., Terzis, A.: Poisoning and backdooring contrastive learning. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=iC4UHbQ01Mp +5. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: 2017 IEEE Symposium on Security and Privacy (SP). pp. 39-57. IEEE Computer Society (2017). https://doi.org/10.1109/SP.2017.49 +6. Chen, P., Liu, S., Zhao, H., Jia, J.: Gridmask data augmentation. arXiv preprint arXiv:2001.04086 (2020). https://doi.org/10.48550/arXiv.2001.04086 +7. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020), https://dl.acm.org/doi/abs/10.5555/3524938.3525087 +8. Chen, X., He, K.: Exploring simple siamese representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15750-15758 (2021). https://doi.org/10.1109/CVPR46437.2021.01549 +9. Cho, J., Nam, G., Kim, S., Yang, H., Kwak, S.: Promptstyler: Prompt-driven style generation for source-free domain generalization. In: Proceedings of the IEEE international conference on computer vision. pp. 15702-15712 (2023). https://doi.org/10.1109/ICCV51070.2023.01439 +0. Daunhawer, I., Bizeul, A., Palumbo, E., Marx, A., Vogt, J.E.: Identifiability results for multimodal contrastive learning. ICLR (2023), https://openreview.net/forum?id=U_2kuqoTcB +11. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020), https://openreview.net/forum?id=YicbFdNTTy +12. Fort, S.: Adversarial vulnerability of powerful near out-of-distribution detection. arXiv preprint arXiv:2201.07012 (2022). https://doi.org/10.48550/arXiv.2201.07012 +13. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision 132(2), 581-595 (2024). https://doi.org/10.1007/s11263-023-01891-x +14. Ge, C., Huang, R., Xie, M., Lai, Z., Song, S., Li, S., Huang, G.: Domain adaptation via prompt learning. arXiv preprint arXiv:2202.06687 (2022). https://doi.org/10.48550/arXiv.2202.06687 +15. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. In: International Conference on Learning Representations (2015). https://doi.org/10.48550/arXiv.1412.6572 + +16. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020), https://dl.acm.org/doi/abs/10.5555/3495724.3497510 +17. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022). https://doi.org/10.1109/CVPR52688.2022.01553 +18. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020). https://doi.org/10.1109/CVPR42600.2020.00975 +19. Hong, T., Guo, X., Ma, J.: Itmix: Image-text mix augmentation for transferring clip to image classification. In: 2022 16th IEEE International Conference on Signal Processing (ICSP). vol. 1, pp. 129-133. IEEE (2022). https://doi.org/10.1109/ICSP56322.2022.9965292 +20. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021), https://proceedings.mlr.press/v139/jia21b.html +21. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19113-19122 (2023). https://doi.org/10.1109/CVPR52729.2023.01832 +22. Kong, L., Xie, S., Yao, W., Zheng, Y., Chen, G., Stojanov, P., Akinwande, V., Zhang, K.: Partial disentanglement for domain adaptation. In: International Conference on Machine Learning. pp. 11455-11472. PMLR (2022), https://proceedings.mlr.press/v162/kong22a.htm1 +23. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.M.: Deeper, broader and artier domain generalization. In: Proceedings of the IEEE international conference on computer vision. pp. 5542-5550 (2017). https://doi.org/10.1109/ICCV.2017.591 +24. Li, H., Wang, X., Zhang, Z., Yuan, Z., Li, H., Zhu, W.: Disentangled contrastive learning on graphs. Advances in Neural Information Processing Systems 34, 21872-21884 (2021), https://dl.acm.org/doi/10.5555/3540261.3541935 +25. Li, Y., Liang, F., Zhao, L., Cui, Y., Ouyang, W., Shao, J., Yu, F., Yan, J.: Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In: International Conference on Learning Representations (2021) +26. Liu, Y., Zhang, Z., Gong, D., Gong, M., Huang, B., van den Hengel, A., Zhang, K., Shi, J.Q.: Identifiable latent polynomial causal models through the lens of change. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=ia9fK01Vjq +27. Liu, Y., Zhang, Z., Gong, D., Gong, M., Huang, B., Hengel, A.v.d., Zhang, K., Shi, J.Q.: Identifying weight-variant latent causal models. arXiv preprint arXiv:2208.14153 (2022). https://doi.org/10.48550/arXiv.2208.14153 +28. Liu, Y., Zhang, Z., Gong, D., Gong, M., Huang, B., Hengel, A.v.d., Zhang, K., Shi, J.Q.: Identifiable latent neural causal models. arXiv preprint arXiv:2403.15711 (2024). https://doi.org/10.48550/arXiv.2403.15711 +29. Liu, Y., Zhang, Z., Gong, D., Gong, M., Huang, B., Zhang, K., Shi, J.Q.: Identifying latent causal content for multi-source domain adaptation. arXiv preprint arXiv:2208.14161 (2022). https://doi.org/10.48550/arXiv.2208.14161 + +30. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. In: International Conference on Learning Representations (2018), https://openreview.net/forum?id=rJzIBfZAb +31. Mahajan, D., Tople, S., Sharma, A.: Domain generalization using causal matching. In: International Conference on Machine Learning. pp. 7313-7324. PMLR (2021), https://proceedings.mlr.press/v139/mahajan21b.html +32. Mamooler, S.: Clip explainability. https://github.com/sMamooler/CLIP_Explainability, accessed: 2024-03-06 +33. Mao, C., Geng, S., Yang, J., Wang, X., Vondrick, C.: Understanding zero-shot adversarial robustness for large-scale models. In: The Eleventh International Conference on Learning Representations (2022), https://openreview.net/forum?id=P4bXcawRi5J +34. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018). https://doi.org/10.48550/arXiv.1807.03748 +35. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE international conference on computer vision. pp. 1406-1415 (2019). https://doi.org/10.1109/ICCV.2019.00149 +36. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021), https://proceedings.mlr.press/v139/radford21a.html +37. Rahman, M.M., Fookes, C., Baktashmotlagh, M., Sridharan, S.: Multi-component image translation for deep domain generalization. In: 2019 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 579-588. IEEE (2019). https://doi.org/10.1109/WACV.2019.00067 +38. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021), https://proceedings.mlrpress/v139/ramesh21a.html +39. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022). https://doi.org/10.1109/CVPR52729.2023.01389 +40. Sanchez, E.H., Serrurier, M., Ortner, M.: Learning disentangled representations via mutual information estimation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 205-221. Springer (2020). https://doi.org/10.1007/978-3-030-58542-6_13 +41. Von Kügelgen, J., Sharma, Y., Gresele, L., Brendel, W., Schölkopf, B., Besserve, M., Locatello, F.: Self-supervised learning with data augmentations provably isolates content from style. Advances in neural information processing systems 34, 16451-16467 (2021), https://d1.acm.org/doi/10.5555/3540261.3541519 +42. Wei, J., Zou, K.: Eda: Easy data augmentation techniques for boosting performance on text classification tasks. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). pp. 6382-6388 (2019). https://doi.org/10.18653/v1/D19-1670 + +43. Wortsman, M., Ilharco, G., Kim, J.W., Li, M., Kornblith, S., Roelofs, R., Lopes, R.G., Hajishirzi, H., Farhadi, A., Namkoong, H., et al.: Robust fine-tuning of zero-shot models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7959-7971 (2022). https://doi.org/10.1109/CVPR52688.2022.00780 +44. Yang, M., Liu, F., Chen, Z., Shen, X., Hao, J., Wang, J.: Causalvae: Disentangled representation learning via neural structural causal models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9593-9602 (2021). https://doi.org/10.1109/CVPR46437.2021.00947 +45. Yang, W., Mirzasoleiman, B.: Robust contrastive language-image pretraining against adversarial attacks. arXiv preprint arXiv:2303.06854 (2023). https://doi.org/10.48550/arXiv.2303.06854 +46. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: IEEE International Conference on Computer Vision (ICCV) (2023). https://doi.org/10.1109/ICCV51070.2023.00355 +47. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022). https://doi.org/10.1109/CVPR52688.2022.01631 +48. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022). https://doi.org/10.1007/s11263-022-01653-1 +49. Zimmermann, R.S., Sharma, Y., Schneider, S., Bethge, M., Brendel, W.: Contrastive learning inverts the data generating process. In: International Conference on Machine Learning. pp. 12979-12990. PMLR (2021), https://proceedings.mlr.press/v139/zimmermann21a.html \ No newline at end of file diff --git a/2024/CLAP_ Isolating Content from Style through Contrastive Learning with Augmented Prompts/images.zip b/2024/CLAP_ Isolating Content from Style through Contrastive Learning with Augmented Prompts/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7fe7595250cbc9000bec7ea5b5a13806c7fe2679 --- /dev/null +++ b/2024/CLAP_ Isolating Content from Style through Contrastive Learning with Augmented Prompts/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e7636d54d30f9ce458266f8ecc264eaa94509307c876c678955f05954601b6e +size 402817 diff --git a/2024/CLAP_ Isolating Content from Style through Contrastive Learning with Augmented Prompts/layout.json b/2024/CLAP_ Isolating Content from Style through Contrastive Learning with Augmented Prompts/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9b64dd818e457915fd86c51e2757133f4dc3be96 --- /dev/null +++ b/2024/CLAP_ Isolating Content from Style through Contrastive Learning with Augmented Prompts/layout.json @@ -0,0 +1,9649 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "text", + "content": "CLAP: Isolating Content from Style through Contrastive Learning with Augmented Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 157, + 168, + 457, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 168, + 457, + 181 + ], + "spans": [ + { + "bbox": [ + 157, + 168, + 457, + 181 + ], + "type": "text", + "content": "Yichao Cai, Yuhang Liu, Zhen Zhang, and Javen Qinfeng Shi" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 190, + 481, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 190, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 190, + 481, + 213 + ], + "type": "text", + "content": "Australian Institute for Machine Learning, University of Adelaide, SA 5000, Australia {yichao.cai,yuhang.liu01,zhen.zhang02,javen.shi}@adelaide.edu.au" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 241, + 455, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 241, + 455, + 472 + ], + "spans": [ + { + "bbox": [ + 160, + 241, + 455, + 472 + ], + "type": "text", + "content": "Abstract. Contrastive vision-language models, such as CLIP, have garnered considerable attention for various downstream tasks, mainly due to the remarkable ability of the learned features for generalization. However, the features they learned often blend content and style information, which somewhat limits their generalization capabilities under distribution shifts. To address this limitation, we adopt a causal generative perspective for multimodal data and propose contrastive learning with data augmentation to disentangle content features from the original representations. To achieve this, we begin with exploring image augmentation techniques and develop a method to seamlessly integrate them into pretrained CLIP-like models to extract pure content features. Taking a step further, recognizing the inherent semantic richness and logical structure of text data, we explore the use of text augmentation to isolate latent content from style features. This enables CLIP-like model's encoders to concentrate on latent content information, refining the learned representations by pre-trained CLIP-like models. Our extensive experiments across diverse datasets demonstrate significant improvements in zero-shot and few-shot classification tasks, alongside enhanced robustness to various perturbations. These results underscore the effectiveness of our proposed methods in refining vision-language representations and advancing the state-of-the-art in multimodal learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 483, + 448, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 483, + 448, + 494 + ], + "spans": [ + { + "bbox": [ + 160, + 483, + 448, + 494 + ], + "type": "text", + "content": "Keywords: Data Augmentation " + }, + { + "bbox": [ + 160, + 483, + 448, + 494 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 483, + 448, + 494 + ], + "type": "text", + "content": " Latent Variables " + }, + { + "bbox": [ + 160, + 483, + 448, + 494 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 483, + 448, + 494 + ], + "type": "text", + "content": " Disentanglement" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "type": "text", + "content": "Vision-language models, exemplified by CLIP [36], have garnered substantial attention due to their exceptional generalization capabilities, achieved through the learned features, obtained by utilizing a cross-modal contrastive loss [20,25,36]. However, despite being pre-trained on extensive datasets, CLIP-like models fall short in disentangling latent content information and latent style information. Consequently, they are not immune to spurious correlations, i.e., style-related information is erroneously utilized to predict task-related labels. These limitations become evident in the presence of distribution shifts or adversarial attacks where spurious correlations often change across different environments. For examples," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 406, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 406, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 406, + 665 + ], + "type": "text", + "content": "1 Our code is available at https://github.com/YichaoCai1/CLAP." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 115, + 227, + 171 + ], + "blocks": [ + { + "bbox": [ + 140, + 115, + 227, + 171 + ], + "lines": [ + { + "bbox": [ + 140, + 115, + 227, + 171 + ], + "spans": [ + { + "bbox": [ + 140, + 115, + 227, + 171 + ], + "type": "image", + "image_path": "6e66cdac6ce80f08ff08ab337f4f8aa2c98ad0f5669134c6c15d54971900f099.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 172, + 217, + 181 + ], + "lines": [ + { + "bbox": [ + 149, + 172, + 217, + 181 + ], + "spans": [ + { + "bbox": [ + 149, + 172, + 217, + 181 + ], + "type": "text", + "content": "(a) Image augmentation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 231, + 119, + 480, + 240 + ], + "blocks": [ + { + "bbox": [ + 231, + 119, + 480, + 240 + ], + "lines": [ + { + "bbox": [ + 231, + 119, + 480, + 240 + ], + "spans": [ + { + "bbox": [ + 231, + 119, + 480, + 240 + ], + "type": "image", + "image_path": "dde32168462b738ed9a57d9dc1dd9a0e30491f46efda143c8b1e5327deb7f068.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 243, + 410, + 251 + ], + "lines": [ + { + "bbox": [ + 318, + 243, + 410, + 251 + ], + "spans": [ + { + "bbox": [ + 318, + 243, + 410, + 251 + ], + "type": "text", + "content": "(c) CAM visualization examples" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 139, + 186, + 227, + 242 + ], + "blocks": [ + { + "bbox": [ + 139, + 186, + 227, + 242 + ], + "lines": [ + { + "bbox": [ + 139, + 186, + 227, + 242 + ], + "spans": [ + { + "bbox": [ + 139, + 186, + 227, + 242 + ], + "type": "image", + "image_path": "a5e203a1edbabb418ef483009400f31d58cf878cdd3d26c87a69176a88eb5137.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 243, + 214, + 251 + ], + "lines": [ + { + "bbox": [ + 151, + 243, + 214, + 251 + ], + "spans": [ + { + "bbox": [ + 151, + 243, + 214, + 251 + ], + "type": "text", + "content": "(b) Text augmentation" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "type": "text", + "content": "Fig. 1: Causal generative models of vision-language data. Image and text data are generated through distinct underlying deterministic processes, " + }, + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "type": "text", + "content": " for images and " + }, + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_{\\mathbf{t}}" + }, + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "type": "text", + "content": " for texts, derived from a unified latent space with latent content variables " + }, + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "type": "text", + "content": " and latent style variables " + }, + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "type": "text", + "content": ". Latent content " + }, + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "type": "text", + "content": " exclusively determines the sample label " + }, + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "type": "text", + "content": ". (a) Soft interventions on latent style variables " + }, + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "type": "text", + "content": " result in " + }, + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{s}}" + }, + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "type": "text", + "content": ", subsequently generating augmented images " + }, + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "type": "text", + "content": ". (b) Due to the same latent space, soft interventions on latent style variables " + }, + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "type": "text", + "content": " can also result in " + }, + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{s}}" + }, + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "type": "text", + "content": ", producing augmented text " + }, + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{t}}" + }, + { + "bbox": [ + 130, + 262, + 482, + 383 + ], + "type": "text", + "content": ". (c) A qualitative comparison of image features for zero-shot classification using \"a photo of a [class]\" prompts, visualized using class activation map (CAM) [32], demonstrates that while image augmentation can enhance CLIP features, the features obtained through text augmentation methods predominantly focus on the content." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 410, + 480, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 480, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 480, + 458 + ], + "type": "text", + "content": "(1) a notable dependence on specific input text prompts has been reported for zero-shot capabilities [21, 47, 48]; (2) performance decline in few-shot scenarios has been observed in few-shot learning scenarios [13, 36]; and (3) susceptibility to adversarial attacks has been explored [33, 43, 45]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 460, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 628 + ], + "type": "text", + "content": "Taking a causal perspective, this work begins with a simple yet effective method, image augmentation, to disentangle content and style information within the learned representations of CLIP-like models. This approach is inspired by recent advancements in theoretical development in causal representation learning [41], which demonstrate that augmented image can be interpreted as a result of soft interventions on latent style variables, as depicted in Fig. 1a. Such augmentation results in a natural data pair where content information remains unchanged while style information changes. Consequently, using contrastive learning, it becomes feasible to isolate the invariant content information from the variant style information. Motivated by this theoretical advancement, we propose a practical method to incorporate image augmentation into CLIP-like models to extract content information from the original learned features. Specifically, a disentangled network is designed to fine-tune the pre-trained CLIP model by using a contrastive loss with image augmentation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Despite the advancements made in disentangling content and style information from the original features learned by CLIP-like models through image augmentation, we recognize an inherent limitation: it is generally challenging to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Cai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": "design adequate image augmentations to ensure all style factors change in an image. Theoretically, disentangling content and style information necessitates changes in all style factors [41]. However, inducing sufficient changes in latent style through image augmentation poses challenges due to the high dimensionality and complexity of style information in image data. Achieving significant style variation via artificially designed image augmentation techniques, such as transforming a photograph of a dog into a sketch while preserving content but dramatically altering style, is notably difficult." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 481, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 481, + 402 + ], + "type": "text", + "content": "Taking a further step, rather than relying on image augmentation, we explore the use of text augmentation to disentangle latent content and style factors. This shift is motivated by two key observations: 1) Vision and language data share the same latent space. Therefore, text augmentation can also be utilized to induce changes in latent style factors instead of image augmentation. 2) Text data inherently possesses high semanticity and logical structure, making it more amenable to property-wise manipulation compared to image data. Consequently, implementing sufficient style changes through text augmentation is more feasible than image augmentation, contributing to isolating content from style information, see Fig. 1c for visual comparison. For instance, transforming text from \"a photo of a dog\" to \"a sketch of a dog\" is straightforward in the language modality, whereas achieving a similar transformation in image data is challenging. Inspired by these observations, we posit that introducing style variations through text augmentation, as illustrated Fig. 1b, provides a more effective approach for learning vision-language content features than relying on image augmentation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 403, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 594 + ], + "type": "text", + "content": "In summary, our contributions include: (1) Aimed at disentangling latent content and style factors to refine vision-language features of pre-trained CLIP-like models, we propose constractive learning with data augmentation to fine tune the original features of pre-trained CLIP-like models from a causal perspective. (2) We present a novel method customized for pre-trained CLIP-like models. This method leverages a disentangled network, which is trained using contrastive learning with image augmentation, to extract latent content features from the learned features provided by image encoder of CLIP-like models. (3) We propose Contrastive Learning with Augmented Prompts (CLAP), to extract latent content features from representations of CLIP-like models. It begins by training a disentangled network using the pre-trained text encoder of CLIP-like models and text augmentation. Subsequently, the trained disentangled network is transferred to the image encoder of CLIP-like models. (4) Experiments conducted on a large real dataset demonstrate the effectiveness of the proposed image augmentation and text augmentation in terms of zero-shot and few-shot performance, as well as robustness against perturbations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 614, + 237, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 614, + 237, + 627 + ], + "spans": [ + { + "bbox": [ + 132, + 614, + 237, + 627 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "Contrastive Vision-Language Models Using a cross-modal contrastive loss, CLIP [36] revolutionarily introduced a scalable contrastive vision-language model" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 221, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 221, + 91, + 447, + 103 + ], + "type": "text", + "content": "CLAP: Contrastive Learning with Augmented Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": "by leveraging a large corpus of internet-sourced image-text pairs, demonstrating unprecedented zero-shot learning capabilities and exceptional generalization ability across datasets and supporting numerous downstream tasks [38]. ALIGN [20] expanded the scale of contrastive vision-language modeling, training on up to one billion image-text pairs while integrating the vision transformer's self-attention mechanism [11], which further enhanced performance. Despite their successes, CLIP-like models exhibit sensitivity to input text prompts [21, 48], leading to variable performance across different prompts. Efforts to mitigate this prompt sensitivity through prompt learning and engineering [9, 14, 21, 47, 48] focus on customizing prompts for specific tasks but do not fundamentally enhance CLIP's representations. Furthermore, CLIP-like models are vulnerable to adversarial attacks [4, 12], with current strategies [33, 45] involving adversarial-natural image pairs to improve resilience. Our work diverges from task-specific approaches by aiming to enhance CLIP's representations from a disentanglement perspective, addressing the aforementioned issues inherent in CLIP-like models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 312, + 483, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 483, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 483, + 494 + ], + "type": "text", + "content": "Disentangled Representation Learning Aimed at segregating intrinsic latent factors in data into distinct, controllable representations, disentangled representation learning benefits various applications [24,40,44]. Specifically, in classification tasks, it's shown that enhancing the model's performance and robustness against data distribution perturbations can be achieved by more effectively disentangling invariant content variables, without needing to identify all intrinsic latent variables completely [22,26-28]. Within single modalities, studies such as [49] have illustrated that contrastive learning [7,16,18] can potentially reverse the data generative process, aiding in the separation of representations. Furthermore, [41] suggest that image augmentation can help isolate content variables from the latent space through significant stylistic changes. [19] employs mixture techniques for data augmentation, enabling more abundant cross-modal matches. Diverging from these methods, our approach focuses on employing text augmentation to disentangle latent content variables, introducing a unique approach to learn refined vision-language representations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 510, + 454, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 510, + 454, + 523 + ], + "spans": [ + { + "bbox": [ + 131, + 510, + 454, + 523 + ], + "type": "text", + "content": "3 A Causal Generative Model for Multi-Modal Data" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": "To understand pretrained CLIP-like models, we investigate the underlying causal generative process for vision-language data. We consider the following causal generative model as depicted in Fig. 1. In the proposed model, the shared latent space ruling vision and language data is divided into two distinct sub-spaces: one corresponding to the latent content variables " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " and the other to the latent style variables " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": ". The latent content variables are posited to determine the object label " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": ", a relationship corroborated by prior studies [22, 29, 31]. Furthermore, to elucidate the correlation between the latent style variable " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " and the object variable " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": ", our model incorporates the premise that the latent content variable " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " causally influences the latent style variable " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": ", in concordance with the principles of causal representation learning highlighted in recent literature [10, 29, 41]. Additionally," + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Y. Cai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "considering the diversity between image data and text data, where information in image data is typically much more details while information in text data tends to be more logically structured nature, we posit distinct causal mechanisms for the generation processes. Our causal generative model is formulated as following structural causal models [2]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 194, + 182, + 481, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 182, + 481, + 196 + ], + "spans": [ + { + "bbox": [ + 194, + 182, + 481, + 196 + ], + "type": "interline_equation", + "content": "\\mathbf {s} := \\mathbf {g} _ {\\mathbf {s}} (\\mathbf {c}), \\mathbf {x} := \\mathbf {g} _ {\\mathbf {x}} (\\mathbf {c}, \\mathbf {s}), \\mathbf {t} := \\mathbf {g} _ {\\mathbf {t}} (\\mathbf {c}, \\mathbf {s}), \\mathbf {y} := \\mathbf {g} _ {\\mathbf {y}} (\\mathbf {c}). \\tag {1}", + "image_path": "1c45fab443ddc64947af325f51a508df14a2efa58017543d22e246d7b3f0ff32.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "text", + "content": "In Eq. (1), the style variable " + }, + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "text", + "content": " is causally influenced by the content via " + }, + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{g_s}" + }, + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "text", + "content": " denote visual and textual data, respectively. Both visual and textual data are causally produced by the shared latent variables " + }, + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "text", + "content": " through distinct, reversible generative processes: " + }, + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{g_x}" + }, + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "text", + "content": " for images and " + }, + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{g_t}" + }, + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "text", + "content": " for text data, respectively. The label " + }, + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "text", + "content": " of a sample is exclusively determined by the content variable " + }, + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "text", + "content": " via " + }, + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{g_y}" + }, + { + "bbox": [ + 130, + 201, + 482, + 297 + ], + "type": "text", + "content": ". For simplicity, exogenous noises are implicitly assumed but not explicitly represented in the causal generative model's formulation, aligning with the common understanding that each latent variable is influenced by exogenous noise." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 297, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 477 + ], + "type": "text", + "content": "Recent seminal work in [41] has demonstrated that the latent content variable " + }, + { + "bbox": [ + 130, + 297, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 297, + 482, + 477 + ], + "type": "text", + "content": " can be identified up to block identifiability (i.e., " + }, + { + "bbox": [ + 130, + 297, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 297, + 482, + 477 + ], + "type": "text", + "content": " can be isolated from style variable " + }, + { + "bbox": [ + 130, + 297, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 297, + 482, + 477 + ], + "type": "text", + "content": "), by requiring all latent style variables to change (e.g., soft interventions on all latent style variables). This change can be achieved through image augmentation, i.e., the augmented image " + }, + { + "bbox": [ + 130, + 297, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 297, + 482, + 477 + ], + "type": "text", + "content": " can be interpreted as a generative result of " + }, + { + "bbox": [ + 130, + 297, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{s}}" + }, + { + "bbox": [ + 130, + 297, + 482, + 477 + ], + "type": "text", + "content": ", which is produced through soft interventions on original latent style variables " + }, + { + "bbox": [ + 130, + 297, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 297, + 482, + 477 + ], + "type": "text", + "content": ". Despite such theoretical advancement, the practical implementation of this theoretical result within CLIP-like models remains unclear. In this study, we propose a practical method to disentangle content and style information within CLIP-like models by employing image augmentation, as detailed in Sec. 4.1. Moreover, we recognize that implementing sufficient changes on all latent style variables " + }, + { + "bbox": [ + 130, + 297, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 297, + 482, + 477 + ], + "type": "text", + "content": " through text augmentation is more feasible than image augmentation, due to high semanticity and logical structure in text data, we delve into the use of text augmentation to separate content information from style information, as discussed in Sec. 4.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 491, + 476, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 491, + 476, + 506 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 476, + 506 + ], + "type": "text", + "content": "4 Isolating Content from Style with Data Augmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 514, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 598 + ], + "type": "text", + "content": "In this section, we propose the employment of data augmentation to extract content information from the learned features in pre-trained CLIP-like models. Essentially, data augmentation facilitates the alteration of style factors while preserving content factors. Consequently, leveraging contrastive learning enables the segregation of content information from style information. We delve into two distinct forms of data augmentation, namely image augmentation (Sec. 4.1) and text augmentation (Sec. 4.2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 613, + 432, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 613, + 432, + 625 + ], + "spans": [ + { + "bbox": [ + 131, + 613, + 432, + 625 + ], + "type": "text", + "content": "4.1 Isolating Content from Style with Augmented Images" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "While recent studies (von et al., 2021) have offered assurance regarding the disentanglement of content and style through contrastive learning with data augmentation, it remains unclear how these theoretical findings can be applied to" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 221, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 221, + 91, + 448, + 103 + ], + "type": "text", + "content": "CLAP: Contrastive Learning with Augmented Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 249, + 178 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 249, + 178 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 249, + 178 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 249, + 178 + ], + "type": "image", + "image_path": "d51d2e0024c09fbe32179e48985340220767f6226601d8137f87eed6017a0136.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 180, + 239, + 190 + ], + "lines": [ + { + "bbox": [ + 146, + 180, + 239, + 190 + ], + "spans": [ + { + "bbox": [ + 146, + 180, + 239, + 190 + ], + "type": "text", + "content": "(a) Image augmentation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 251, + 114, + 362, + 178 + ], + "blocks": [ + { + "bbox": [ + 251, + 114, + 362, + 178 + ], + "lines": [ + { + "bbox": [ + 251, + 114, + 362, + 178 + ], + "spans": [ + { + "bbox": [ + 251, + 114, + 362, + 178 + ], + "type": "image", + "image_path": "db85c17d17159a5651a4daf6640b7096bd24c40f6d9f8ce12b44872b3fa4b80f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 180, + 351, + 190 + ], + "lines": [ + { + "bbox": [ + 263, + 180, + 351, + 190 + ], + "spans": [ + { + "bbox": [ + 263, + 180, + 351, + 190 + ], + "type": "text", + "content": "(b) Text augmentation" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "lines": [ + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "text", + "content": "Fig. 2: Refining CLIP through data augmentation. (a) Training involves a disentangled network " + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{f_c}" + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "text", + "content": ", utilizing contrastive loss on original and augmented image pairs " + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "text", + "content": ", with CLIP's image encoder " + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{f_x^*}" + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "text", + "content": " holding frozen gradients. (b) More efficient content feature learning is achieved through contrastive learning with augmented text prompts " + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{t}}" + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "text", + "content": ", using the fixed text encoder " + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{f_t^*}" + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "text", + "content": " of CLIP. (c) Inference stage: The trained disentangled network " + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{f_c^*}" + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "text", + "content": " integrates with CLIP's text and image encoders, " + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{f_t^*}" + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{f_x^*}" + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "text", + "content": ", to enable zero-shot inference for an input image " + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "text", + "content": " and class names " + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{t_1}" + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{t_n}" + }, + { + "bbox": [ + 130, + 199, + 482, + 277 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 365, + 114, + 477, + 178 + ], + "blocks": [ + { + "bbox": [ + 365, + 114, + 477, + 178 + ], + "lines": [ + { + "bbox": [ + 365, + 114, + 477, + 178 + ], + "spans": [ + { + "bbox": [ + 365, + 114, + 477, + 178 + ], + "type": "image", + "image_path": "24c7f87d886d6e99a42e77d789d6bd111c1314f0329c5c597871cecd865e977f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 180, + 465, + 190 + ], + "lines": [ + { + "bbox": [ + 378, + 180, + 465, + 190 + ], + "spans": [ + { + "bbox": [ + 378, + 180, + 465, + 190 + ], + "type": "text", + "content": "(c) Zero-shot inference" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 298, + 482, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 482, + 335 + ], + "type": "text", + "content": "the realm of vision-language models. We convert the theoretical findings into CLIP-like models in the following. The theoretical findings suggest using InfoNCE loss [34] to extract content information, as outlined below:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 167, + 340, + 481, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 340, + 481, + 370 + ], + "spans": [ + { + "bbox": [ + 167, + 340, + 481, + 370 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\mathbf {f}; \\{\\mathbf {x} _ {i}, \\tilde {\\mathbf {x}} _ {i} \\} _ {i = 1} ^ {b}, \\tau) = - \\frac {1}{b} \\sum_ {i = 1} ^ {b} \\log \\frac {\\exp [ \\langle \\mathbf {f} (\\mathbf {x} _ {i}) , \\mathbf {f} (\\tilde {\\mathbf {x}} _ {i}) \\rangle / \\tau ]}{\\sum_ {j = 1} ^ {b} \\exp [ \\langle \\mathbf {f} (\\mathbf {x} _ {i}) , \\mathbf {f} (\\tilde {\\mathbf {x}} _ {j}) \\rangle / \\tau ]}, \\tag {2}", + "image_path": "7159173bc7d1cfe744d4799f54900c34a5c15d371bf36fa78562def08a458fc5.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 376, + 480, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 480, + 424 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 376, + 480, + 424 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i\\}_{i=1}^b" + }, + { + "bbox": [ + 130, + 376, + 480, + 424 + ], + "type": "text", + "content": " represents a batch of " + }, + { + "bbox": [ + 130, + 376, + 480, + 424 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 376, + 480, + 424 + ], + "type": "text", + "content": " samples from the training dataset, " + }, + { + "bbox": [ + 130, + 376, + 480, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{f}(\\mathbf{x}_i)" + }, + { + "bbox": [ + 130, + 376, + 480, + 424 + ], + "type": "text", + "content": " denotes sample " + }, + { + "bbox": [ + 130, + 376, + 480, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 130, + 376, + 480, + 424 + ], + "type": "text", + "content": "'s features through model " + }, + { + "bbox": [ + 130, + 376, + 480, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 130, + 376, + 480, + 424 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 376, + 480, + 424 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_i" + }, + { + "bbox": [ + 130, + 376, + 480, + 424 + ], + "type": "text", + "content": " is the augmented counterpart of " + }, + { + "bbox": [ + 130, + 376, + 480, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 130, + 376, + 480, + 424 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 376, + 480, + 424 + ], + "type": "inline_equation", + "content": "\\langle \\mathbf{z}_1, \\mathbf{z}_2 \\rangle" + }, + { + "bbox": [ + 130, + 376, + 480, + 424 + ], + "type": "text", + "content": " represents the cosine similarity between two feature vectors, " + }, + { + "bbox": [ + 130, + 376, + 480, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_1" + }, + { + "bbox": [ + 130, + 376, + 480, + 424 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 376, + 480, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_2" + }, + { + "bbox": [ + 130, + 376, + 480, + 424 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 376, + 480, + 424 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 376, + 480, + 424 + ], + "type": "text", + "content": " represents the temperature parameter influencing the loss." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 425, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 496 + ], + "type": "text", + "content": "We extend it to refine pre-trained vision-language models, utilizing contrastive learning with augmented images (hereinafter referred to as \"Im.Aug\"). As illustrated in Fig. 2a, we train a disentangled network on top of CLIP's pretrained image encoder. To enhance training efficiency and the usability of the proposed method, we freeze the pre-trained image encoder. Based on an InfoNCE loss, the learning objective of Im.Aug is formulated as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 201, + 502, + 481, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 502, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 201, + 502, + 481, + 525 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {\\mathbf {c}} ^ {*} = \\underset {\\mathbf {f} _ {\\mathbf {c}}} {\\operatorname {a r g m i n}} \\underset {\\{\\mathbf {x} _ {i} \\} _ {i = 1} ^ {b} \\in \\mathcal {D} _ {\\mathbf {x}}} {\\mathbb {E}} \\mathcal {L} \\left(\\mathbf {f} _ {\\mathbf {c}} \\circ \\mathbf {f} _ {\\mathbf {x}} ^ {*}; \\left\\{\\mathbf {x} _ {i}, \\tilde {\\mathbf {x}} _ {i} \\right\\} _ {i = 1} ^ {b}, \\tau\\right), \\tag {3}", + "image_path": "313bed34a3891325ba7958a5223843cde34447f34e38be8b65b8bfc0a961f90d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 529, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 482, + 589 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 529, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 529, + 482, + 589 + ], + "type": "text", + "content": " denotes the training image dataset and " + }, + { + "bbox": [ + 130, + 529, + 482, + 589 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 529, + 482, + 589 + ], + "type": "text", + "content": " represents the batch size, " + }, + { + "bbox": [ + 130, + 529, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{f_c}" + }, + { + "bbox": [ + 130, + 529, + 482, + 589 + ], + "type": "text", + "content": " is the disentangled network undergoing training. The pre-trained CLIP image encoder is represented by " + }, + { + "bbox": [ + 130, + 529, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{f_x^*}" + }, + { + "bbox": [ + 130, + 529, + 482, + 589 + ], + "type": "text", + "content": ", with the asterisk \"***\" signifying that the model weights remain fixed. The variable " + }, + { + "bbox": [ + 130, + 529, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 130, + 529, + 482, + 589 + ], + "type": "text", + "content": " refers to an image sampled from " + }, + { + "bbox": [ + 130, + 529, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 529, + 482, + 589 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 529, + 482, + 589 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_i" + }, + { + "bbox": [ + 130, + 529, + 482, + 589 + ], + "type": "text", + "content": " is its augmented view." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 590, + 482, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 482, + 626 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 482, + 626 + ], + "type": "text", + "content": "The composition of the training dataset " + }, + { + "bbox": [ + 130, + 590, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 590, + 482, + 626 + ], + "type": "text", + "content": ", the image augmentation techniques used, the structure of the disentangled network " + }, + { + "bbox": [ + 130, + 590, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{f_c}" + }, + { + "bbox": [ + 130, + 590, + 482, + 626 + ], + "type": "text", + "content": ", and the utilization of " + }, + { + "bbox": [ + 130, + 590, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{f_c^*}" + }, + { + "bbox": [ + 130, + 590, + 482, + 626 + ], + "type": "text", + "content": " post-training are detailed in the following subsections." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Data Synthesis and Image Augmentation To generate training image data, we combine class names with various image and object attributes to create text" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Cai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 152, + 157, + 462, + 209 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 1: Template-based prompts. Attributes used to generate text prompts follow the structured format \"a [art style] [image type] of a [object size] [object color] [class]\", where \"[class]\" represents the class names." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 152, + 157, + 462, + 209 + ], + "lines": [ + { + "bbox": [ + 152, + 157, + 462, + 209 + ], + "spans": [ + { + "bbox": [ + 152, + 157, + 462, + 209 + ], + "type": "table", + "html": "
Object ColorObject SizeImage TypeArt Style
yellow, green, black, blue, multicolored, orange, red, white, brown, purplelarge, small, normal sizedpainting, cartoon, infographic, sketch, photograph, clipart, mosaic art, sculpturerealistic, impressionistic
", + "image_path": "7ce62a6755332c85eb9f5ce5e1dbaf24af7127ed305e060f329684fec451f01b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 231, + 482, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 231, + 482, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 482, + 363 + ], + "type": "text", + "content": "prompts for each class. Using a stable diffusion model [39], we produce synthetic images that comprise our training dataset " + }, + { + "bbox": [ + 130, + 231, + 482, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{D}\\mathbf{x}" + }, + { + "bbox": [ + 130, + 231, + 482, + 363 + ], + "type": "text", + "content": ". The creation of template prompts for stable diffusion is based on attributes such as object size, color, image type, and art style. As detailed in Tab. 1, the attributes include 10 colors and 3 sizes for objects, and 8 types and 2 art styles for images. By assembling these attributes into prompts like \"a [art style] [image type] of a [object size] [object color] [class]\", we generate 480 unique texts for each class, from which one image per prompt is synthesized. Further details on image synthesis and examples are available in Appendix B.1. For the image augmentation procedures, we adopt techniques commonly used in contrastive learning practice [7,8,41], specifically random cropping and color distortion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 382, + 482, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 482, + 478 + ], + "type": "text", + "content": "Disentangled Network Structure Since the training process is based on CLIP's pre-trained lower-dimensional features, our disentangled network adopts a multi-layer perceptron (MLP) architecture. To fully benefit from the pretrained CLIP text encoder, we construct a residual MLP featuring a zero-initialized projection, acting as the disentangled network, as depicted in Fig. 3. This design enables learning directly from the pre-trained representation space, avoiding a random starting point, inspired by ControlNet's zero-conv operation [46], which we adapt to a zero-linear operation within our residual MLP." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 479, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 624 + ], + "type": "text", + "content": "Within this architecture, the main branch includes a zero-initialized, bias-free linear layer positioned subsequent to the combination of a SiLU activation and a normally initialized linear layer. Conventionally, the dimensions of features before the initial linear layer, situated between the first and second linear layers, and following the second linear layer, are named as the input " + }, + { + "bbox": [ + 130, + 479, + 482, + 624 + ], + "type": "inline_equation", + "content": "d_{in}" + }, + { + "bbox": [ + 130, + 479, + 482, + 624 + ], + "type": "text", + "content": ", latent " + }, + { + "bbox": [ + 130, + 479, + 482, + 624 + ], + "type": "inline_equation", + "content": "d_{mid}" + }, + { + "bbox": [ + 130, + 479, + 482, + 624 + ], + "type": "text", + "content": ", and output " + }, + { + "bbox": [ + 130, + 479, + 482, + 624 + ], + "type": "inline_equation", + "content": "d_{out}" + }, + { + "bbox": [ + 130, + 479, + 482, + 624 + ], + "type": "text", + "content": " dimensions, respectively. To rectify any mismatches between the input and output dimensions, the network employs nearest-neighbor downsampling within the shortcut path, thereby ensuring both alignment and the preservation of sharpness for the input features. During the inference stage, a weighting parameter " + }, + { + "bbox": [ + 130, + 479, + 482, + 624 + ], + "type": "inline_equation", + "content": "\\alpha > 0" + }, + { + "bbox": [ + 130, + 479, + 482, + 624 + ], + "type": "text", + "content": " is introduced to modulate the portion of features emanating from the main branch before their integration with the input features, whereas this parameter remains constant at 1 throughout the training phase." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "Inference After training, the disentangled network " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{c}}^{*}" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": " is utilized following CLIP's image encoder to extract visual content features. Moreover, given that" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 221, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 221, + 90, + 447, + 103 + ], + "type": "text", + "content": "CLAP: Contrastive Learning with Augmented Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 209, + 114, + 406, + 195 + ], + "blocks": [ + { + "bbox": [ + 209, + 114, + 406, + 195 + ], + "lines": [ + { + "bbox": [ + 209, + 114, + 406, + 195 + ], + "spans": [ + { + "bbox": [ + 209, + 114, + 406, + 195 + ], + "type": "image", + "image_path": "26bd3fc57531fc0688d8e6ba852c119d57c9e15fd1dd84bdc1a7e609acd3bb5e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 204, + 482, + 258 + ], + "lines": [ + { + "bbox": [ + 130, + 204, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 482, + 258 + ], + "type": "text", + "content": "Fig. 3: Structure of the disentangled network. The architecture encompasses a residual block featuring a zero-initialized, bias-free linear layer to commence optimization from the input feature space. When the input and output dimension differ, a downsampling operation is utilized to achieve alignment. During inference, a scalar parameter " + }, + { + "bbox": [ + 130, + 204, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 204, + 482, + 258 + ], + "type": "text", + "content": " balance the main branch and input features before combination." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 282, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 354 + ], + "type": "text", + "content": "vision-language data generation is rooted in a unified latent space, as depicted in Sec. 3, " + }, + { + "bbox": [ + 130, + 282, + 482, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{c}}^{*}" + }, + { + "bbox": [ + 130, + 282, + 482, + 354 + ], + "type": "text", + "content": " can be seamlessly integrated with CLIP's image and text encoders to enhance zero-shot capabilities. As shown in Fig. 2c, for an image " + }, + { + "bbox": [ + 130, + 282, + 482, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 282, + 482, + 354 + ], + "type": "text", + "content": ", the operation is formulated as the composition function " + }, + { + "bbox": [ + 130, + 282, + 482, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{c}}^{*} \\circ \\mathbf{f}_{\\mathbf{x}}^{*}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 282, + 482, + 354 + ], + "type": "text", + "content": ", and similarly, for a text " + }, + { + "bbox": [ + 130, + 282, + 482, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 130, + 282, + 482, + 354 + ], + "type": "text", + "content": ", as " + }, + { + "bbox": [ + 130, + 282, + 482, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{c}}^{*} \\circ \\mathbf{f}_{\\mathbf{t}}^{*}(\\mathbf{t})" + }, + { + "bbox": [ + 130, + 282, + 482, + 354 + ], + "type": "text", + "content": ". This integration preserves CLIP's zero-shot functionality while achieving refined features through the improved disentanglement of content." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 370, + 440, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 370, + 440, + 384 + ], + "spans": [ + { + "bbox": [ + 131, + 370, + 440, + 384 + ], + "type": "text", + "content": "4.2 Isolating Content from Style with Augmented Prompts" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 389, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 496 + ], + "type": "text", + "content": "Despite progress in disentangling content and style via image augmentation, adequately altering all style factors in an image remains challenging due to the high dimensionality and complexity of style information in images. Achieving substantial style changes through augmentation, essential for complete disentanglement [41], is difficult with existing image augmentation techniques. On the contrary, text data inherently possesses high semanticity and logical structure, making it more amenable to property-wise manipulation compared to image data. To further exploring the disentanglement of content, we propose Contrastive Learning with Augmented Prompts (CLAP)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 497, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 569 + ], + "type": "text", + "content": "As depicted in Fig. 2b, CLAP employs an InfoNCE loss to train a disentangled network atop CLIP's pre-trained text encoder, keeping the encoder's gradients fixed, similar to Im.Aug. Leveraging the simpler structure of text, the template-based prompts previously utilized for synthesizing images now serve as the training text dataset, denoted by " + }, + { + "bbox": [ + 130, + 497, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathbf{t}}" + }, + { + "bbox": [ + 130, + 497, + 482, + 569 + ], + "type": "text", + "content": ". Utilizing the same disentangled network as in Im.Aug, the learning objective of CLAP is outlined as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 139, + 575, + 481, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 575, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 139, + 575, + 481, + 598 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {\\mathbf {c}} ^ {*} = \\underset {\\mathbf {f} _ {\\mathbf {c}}} {\\operatorname {a r g m i n}} \\underset {\\left\\{\\mathbf {t} _ {i} \\right\\} _ {i = 1} ^ {b} \\in \\mathcal {D} _ {\\mathbf {t}}} {\\mathbb {E}} \\mathcal {L} \\left(\\mathbf {f} _ {\\mathbf {c}} \\circ \\mathbf {f} _ {\\mathbf {t}} ^ {*}; \\left\\{\\mathbf {t} _ {i}, \\tilde {\\mathbf {t}} _ {i} \\right\\} _ {i = 1} ^ {b}, \\tau\\right) + \\lambda \\mathcal {L} \\left(\\mathbf {f} _ {\\mathbf {c}} \\circ \\mathbf {f} _ {\\mathbf {t}} ^ {*}; \\left\\{\\mathbf {t} _ {i} ^ {c}, \\tilde {\\mathbf {t}} _ {i} \\right\\} _ {i = 1} ^ {b}, 1\\right), \\tag {4}", + "image_path": "802e8aa45eb77c81a8a041e4a6ba3f04f4a82af1fa878800cddab0b5da3b3987.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{t}}^{*}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " denotes the pre-trained CLIP text encoder. The term " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " references a text prompt from " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{D}\\mathbf{t}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{t}}_i" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " represents its augmented view, produced via prompt augmentation techniques. On the equation's right side, " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i^c" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " specifies the class name associated with the text prompt " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ". This strategy aims to enhance variations between prompt pairs, especially in cases where the text dataset " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathbf{t}}" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Cai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 152, + 168, + 462, + 218 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 2: Prompt augmentation techniques. Various augmented views are generated from an original text prompt using specific augmentation techniques: OSD (Object Size Deletion), OCD (Object Color Deletion), ITD (Image Type Deletion), ASD (Art Style Deletion), and SPO (Swapping Prompt Order)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 152, + 168, + 462, + 218 + ], + "lines": [ + { + "bbox": [ + 152, + 168, + 462, + 218 + ], + "spans": [ + { + "bbox": [ + 152, + 168, + 462, + 218 + ], + "type": "table", + "html": "
OriginalOSDOCDITDASDSPO
a realistic painting of a large red cara realistic painting of a red cara realistic painting of a large cara realistic of a large red cara painting of a large red cara large red car in a realistic painting
", + "image_path": "3c3fdca5fe64023cf19ba35b40d419860bcf3850227aed5dada119ea7a70f4c8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 239, + 479, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 479, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 479, + 275 + ], + "type": "text", + "content": "has a very limited number of samples. Here, " + }, + { + "bbox": [ + 130, + 239, + 479, + 275 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 239, + 479, + 275 + ], + "type": "text", + "content": " serves for adjusting the second term's importance in the total loss function. All other symbols in Eq. (4) maintain their definitions as described earlier." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 276, + 480, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 276, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 480, + 300 + ], + "type": "text", + "content": "After training, the learned disentangled network is seamlessly integrated with both of CLIP's encoders to extract content representations, as depicted in Fig. 2c." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 319, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 511 + ], + "type": "text", + "content": "Prompt Augmentation To ensure text prompts undergo stylistic changes without compromising their content, we have developed specific augmentation techniques for synthetic text prompts. Drawing inspiration from Easy Data Augmentation (EDA) techniques [42], we adapted the Random Deletion (RD) and Random Swap (RS) techniques from EDA, customizing them to suit our prompt structure. To avoid inadvertently altering the content by introducing new object names or changing the core idea of a text prompt, our augmentation methods do not include random word insertions or replacements. Our primary augmentation techniques are Object Size Deletion (OSD), Object Color Deletion (OCD), Image Type Deletion (ITD), Art Style Deletion (ASD), and Swapping Prompt Order (SPO), each applied with a certain probability, as detailed in Tab. 2. Additionally, for down-stream datasets with few categories, to rich the population of training samples, we use an additional augmentation, named IGN (Inserting Gaussian Noise). Following the initializing protocol of prompt learning methods [47,48], we insert a zero-mean Gaussian noise with 0.02 standard deviation with a noise length equals to 4, to the tokenized prompts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 512, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 482, + 596 + ], + "type": "text", + "content": "Intuitively, these prompt augmentation methods parallel random masking techniques used in image augmentation [6,17]. However, prompt augmentations are more effective and precise than their image counterparts. This effectiveness arises because prompt augmentations can specifically target and eliminate a particular style element without impacting the content, whereas image masking, operating at the pixel or patch level, might inadvertently damage content information or lead to insufficient style changes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 615, + 230, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 615, + 230, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 615, + 230, + 628 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "We conduct three primary experiments to assess our method: (1) zero-shot evaluation with diverse prompts to gauge zero-shot performance and its robustness" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 221, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 221, + 91, + 447, + 103 + ], + "type": "text", + "content": "CLAP: Contrastive Learning with Augmented Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "to prompt perturbations; (2) linear probe tests on few-shot samples to evaluate the efficacy of the learned representations in few-shot settings; and (3) adversarial attack assessments on zero-shot and one-shot classifiers to determine their resistance to adversarial threats. We further conduct an ablative study on hyperparameters, explore the impact of different prompt augmentation combinations and various sources of training prompts on CLAP's performance, and replicate experiments across different CLIP model sizes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 217, + 261, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 261, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 261, + 228 + ], + "type": "text", + "content": "5.1 Experimental Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 235, + 481, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 481, + 283 + ], + "type": "text", + "content": "Implementation. Im.Aug and CLAP are implemented using the ViT-B/16 CLIP model and executed on an NVIDIA RTX 3090 GPU. To ensure reproducibility, the random seed for all stochastic processes is fixed at 2024. More information on implementation details is provided in Appendix A.1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 283, + 481, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 481, + 354 + ], + "type": "text", + "content": "Datasets. CLAP is assessed across four multi-domain datasets to examine its performance in varied environments: PACS [23] (4 domains, 7 categories), VLCS [1] (4 domains, 5 categories), OfficeHome [37] (4 domains, 65 categories), and DomainNet [35] (6 domains, 345 categories). For conciseness, we present average results across the domains for each dataset. Detailed experimental outcomes for each domain within these datasets are provided in Appendix A.4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 355, + 482, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 355, + 482, + 438 + ], + "spans": [ + { + "bbox": [ + 130, + 355, + 482, + 438 + ], + "type": "text", + "content": "Compute efficiency. CLAP demonstrates faster convergence and shorter training times compared to Im.Aug. For CLAP, training on the PACS and VLCS datasets is completed in roughly 11 minutes, OfficeHome in approximately 14 minutes, and DomainNet in about 47 minutes. In contrast, Im.Aug requires around 16 minutes for PACS and VLCS, 50 minutes for OfficeHome, and 3.3 hours for DomainNet. Both Im.Aug and CLAP maintain CLIP's inference efficiency due to the disentangled network's efficient two-layer MLP structure." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 455, + 228, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 455, + 228, + 466 + ], + "spans": [ + { + "bbox": [ + 132, + 455, + 228, + 466 + ], + "type": "text", + "content": "5.2 Main Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 474, + 481, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 557 + ], + "type": "text", + "content": "Zero-Shot Performance To assess zero-shot capabilities, CLAP undergoes evaluation using three specific fixed prompts: ZS(C), utilizing only the class name within \"[class]\"; ZS(PC), with the format \"a photo of a [class]\"; and ZS(CP), structured as \"a [class] in a photo\". To thoroughly examine zero-shot proficiency, a dynamic prompt, ZS(NC), formatted as \"[noise][class]\", is also used, where \"[noise]\" signifies the introduction of Gaussian noise characterized by a mean of 0 and a standard deviation of 0.02." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 640 + ], + "type": "text", + "content": "As presented in Tab. 3, CLAP surpasses both CLIP and Im.Aug across all evaluated prompts for every dataset. Unlike the uniform enhancement in zero-shot performance CLAP achieves over CLIP, Im.Aug displays inconsistent results. A closer examination reveals CLAP's superiority over CLIP is especially significant for the dynamic ZS(NC) prompt. This demonstrates CLAP's effectiveness in significantly improving zero-shot performance compared to the original CLIP representations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "In assessing the model's robustness to prompt perturbations, we examine the variances in zero-shot performance across different prompts by analyzing the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Cai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 190, + 168, + 425, + 335 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 3: Zero-shot results across three distinct prompts: \"C\" for \"[class]\", \"CP\" for \"a [class] in a photo\", \"PC\" for \"a photo of a [class]\", and a dynamic prompt \"NC\" for \"[noise][class]\" showcase that CLAP consistently outperforms CLIP's zero-shot performance across all datasets, whereas image augmentation exhibits mixed outcomes." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 190, + 168, + 425, + 335 + ], + "lines": [ + { + "bbox": [ + 190, + 168, + 425, + 335 + ], + "spans": [ + { + "bbox": [ + 190, + 168, + 425, + 335 + ], + "type": "table", + "html": "
PromptMethodZero-shot performance, avg. top-1 acc. (%) (↑)
PACSVLCSOff.HomeDom.NetOverall
ZS(C)CLIP95.776.479.857.877.4
Im.Aug96.579.577.051.576.1
CLAP97.282.681.058.779.9
ZS(CP)CLIP95.282.079.557.078.4
Im.Aug96.382.975.850.776.4
CLAP97.383.480.558.079.8
ZS(PC)CLIP96.182.482.557.779.7
Im.Aug96.583.078.651.677.4
CLAP97.283.483.059.080.6
ZS(NC)CLIP90.868.371.551.070.4
Im.Aug94.873.167.544.069.9
CLAP97.281.073.552.676.1
", + "image_path": "e50fe19e798722be31494a487a218550962a81b6dc1dcf2afb6d16c916ae9cd7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 356, + 479, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 479, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 479, + 392 + ], + "type": "text", + "content": "range " + }, + { + "bbox": [ + 130, + 356, + 479, + 392 + ], + "type": "inline_equation", + "content": "(R)" + }, + { + "bbox": [ + 130, + 356, + 479, + 392 + ], + "type": "text", + "content": " and standard deviation " + }, + { + "bbox": [ + 130, + 356, + 479, + 392 + ], + "type": "inline_equation", + "content": "(\\delta)" + }, + { + "bbox": [ + 130, + 356, + 479, + 392 + ], + "type": "text", + "content": " of results derived from ZS(C), ZS(CP), and ZS(PC). Additionally, we investigate the decline " + }, + { + "bbox": [ + 130, + 356, + 479, + 392 + ], + "type": "inline_equation", + "content": "(\\varDelta_{(NC)})" + }, + { + "bbox": [ + 130, + 356, + 479, + 392 + ], + "type": "text", + "content": " in performance from ZS(C) to ZS(NC) as a broad indicator of resilience to noised prompts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 393, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 393, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 393, + 482, + 489 + ], + "type": "text", + "content": "As presented in Tab. 4, CLAP significantly reduces the variance in zero-shot performance across various testing prompts, evidenced by markedly lower values of " + }, + { + "bbox": [ + 130, + 393, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 393, + 482, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 393, + 482, + 489 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 393, + 482, + 489 + ], + "type": "text", + "content": ", and a less pronounced decrease in performance with a noised prompt, in contrast to Im.Aug and the baseline representations of CLIP. Although Im.Aug aids in reducing performance variance to some extent, its efficacy is notably inferior to that of CLAP. These findings highlight CLAP's enhanced robustness in maintaining consistent zero-shot performance across a diverse array of prompts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 505, + 482, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 505, + 482, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 505, + 482, + 613 + ], + "type": "text", + "content": "Few-Shot Performance We conduct evaluations of 1-shot, 4-shot, 8-shot, and 16-shot linear probes across each domain within the four datasets. As illustrated in Fig. 4, CLAP significantly outperforms both CLIP and Im.Aug in few-shot learning scenarios. Notably, in the 1-shot setting CLAP achieves performance improvements over the linear-probe CLIP model by margins of " + }, + { + "bbox": [ + 130, + 505, + 482, + 613 + ], + "type": "inline_equation", + "content": "+10\\%" + }, + { + "bbox": [ + 130, + 505, + 482, + 613 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 505, + 482, + 613 + ], + "type": "inline_equation", + "content": "+3.5\\%" + }, + { + "bbox": [ + 130, + 505, + 482, + 613 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 505, + 482, + 613 + ], + "type": "inline_equation", + "content": "+2.5\\%" + }, + { + "bbox": [ + 130, + 505, + 482, + 613 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 505, + 482, + 613 + ], + "type": "inline_equation", + "content": "+1.5\\%" + }, + { + "bbox": [ + 130, + 505, + 482, + 613 + ], + "type": "text", + "content": " on the PACS, VLCS, OfficeHome, and DomainNet datasets, respectively. These improvements are especially significant in comparison to the gains observed with Im.Aug counterparts, underpinning CLAP's efficacy in few-shot scenarios. For detailed quantitative results, please refer to Appendix A.4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": "Adversarial Performance To assess adversarial robustness, zero-shot (ZS(C)) and one-shot classifiers are evaluated against prominent adversarial attack methods, such as FGSM [15], PGD [30], and CW [5], by generating adversarial sam" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 221, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 221, + 90, + 447, + 103 + ], + "type": "text", + "content": "CLAP: Contrastive Learning with Augmented Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 195, + 168, + 419, + 300 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 4: CLAP more effectively reduces zero-shot performance variance across prompts than image augmentation, with " + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": " indicating the range and standard deviation for ZS(C), ZS(PC), and ZS(PC). The decrease " + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "\\Delta_{(NC)}" + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": " from ZS(C) to ZS(NC) highlights CLAP's enhanced robustness against prompt perturbations." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 195, + 168, + 419, + 300 + ], + "lines": [ + { + "bbox": [ + 195, + 168, + 419, + 300 + ], + "spans": [ + { + "bbox": [ + 195, + 168, + 419, + 300 + ], + "type": "table", + "html": "
MetricMethodPerformance variance, avg. top-1 acc. (%) (↓)
PACSVLCSOff.HomeDom.NetOverall
RCLIP0.96.13.10.82.7
Im.Aug0.13.62.80.91.9
CLAP0.10.82.51.01.1
δCLIP0.42.81.40.41.2
Im.Aug0.11.71.20.40.8
CLAP0.00.41.10.40.5
Δ(NC)CLIP4.98.18.36.87.0
Im.Aug1.66.49.57.56.3
CLAP0.01.67.56.13.8
", + "image_path": "fc590fac5daf7b2072e9461ed34e99a83d6854c831016282108eaf95b6fc9f0d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 136, + 310, + 225, + 380 + ], + "blocks": [ + { + "bbox": [ + 136, + 310, + 225, + 380 + ], + "lines": [ + { + "bbox": [ + 136, + 310, + 225, + 380 + ], + "spans": [ + { + "bbox": [ + 136, + 310, + 225, + 380 + ], + "type": "image", + "image_path": "ef84eab710a70b77efa730ef360f20d3f62c4b3f3dfced28ba3bb86ea520a6a1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 387, + 482, + 422 + ], + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 422 + ], + "type": "text", + "content": "Fig. 4: Few-shot linear probe comparisons of image-encoder features show that CLAP enhances CLIP's few-shot performance more effectively than Im.Aug. In the accompanying figure, \"ZS\" indicates the zero-shot performance using a \"[class]\" prompt." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 225, + 310, + 307, + 379 + ], + "blocks": [ + { + "bbox": [ + 225, + 310, + 307, + 379 + ], + "lines": [ + { + "bbox": [ + 225, + 310, + 307, + 379 + ], + "spans": [ + { + "bbox": [ + 225, + 310, + 307, + 379 + ], + "type": "image", + "image_path": "83e8ef217dfe60fa8be48905b67747d5aa0757bec0b8f4ab5486c1f7f0431d96.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 308, + 310, + 392, + 379 + ], + "blocks": [ + { + "bbox": [ + 308, + 310, + 392, + 379 + ], + "lines": [ + { + "bbox": [ + 308, + 310, + 392, + 379 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 392, + 379 + ], + "type": "image", + "image_path": "b93b693869ec8d910eceebeb10f6de19c4ffdbadeb5824ceb7156ea32d94d066.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 392, + 310, + 477, + 379 + ], + "blocks": [ + { + "bbox": [ + 392, + 310, + 477, + 379 + ], + "lines": [ + { + "bbox": [ + 392, + 310, + 477, + 379 + ], + "spans": [ + { + "bbox": [ + 392, + 310, + 477, + 379 + ], + "type": "image", + "image_path": "68af5ad4d6da7adb0c2a6088c6da0ff56934239eb540773eb0c03fd2fdab6acd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "text", + "content": "plies for testing. For FGSM, 1 adversarial iteration is employed, whereas for PGD and CW, 20 iterations are used, all with an epsilon of 0.031. As indicated in Tab. 5, classifiers utilizing CLAP representations demonstrate superior resilience to these adversarial attacks compared to those based on CLIP representations. Across the four datasets, CLAP's zero-shot and 1-shot classifiers surpass CLIP by margins of " + }, + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "inline_equation", + "content": "+7.6\\%" + }, + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "inline_equation", + "content": "+8.5\\%" + }, + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "text", + "content": " against FGSM, " + }, + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "inline_equation", + "content": "+1.0\\%" + }, + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "inline_equation", + "content": "+11.7\\%" + }, + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "text", + "content": " against PGD-20, and " + }, + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "inline_equation", + "content": "+1.1\\%" + }, + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "inline_equation", + "content": "+2.3\\%" + }, + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "text", + "content": " against CW-20, respectively. These figures notably exceed the performance improvements of " + }, + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "inline_equation", + "content": "+4.4\\%" + }, + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "inline_equation", + "content": "+4.6\\%" + }, + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "text", + "content": " against FGSM, " + }, + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "inline_equation", + "content": "+0.3\\%" + }, + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "inline_equation", + "content": "+6.2\\%" + }, + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "text", + "content": " against PGD-20, and " + }, + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "inline_equation", + "content": "+1.3\\%" + }, + { + "bbox": [ + 130, + 445, + 482, + 577 + ], + "type": "text", + "content": " against CW-20 achieved by Im.Aug. The result suggests that CLAP efficiently enhances robustness against adversarial attacks in both zero-shot and one-shot scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 596, + 234, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 596, + 234, + 608 + ], + "spans": [ + { + "bbox": [ + 131, + 596, + 234, + 608 + ], + "type": "text", + "content": "5.3 More Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "t-SNE Visualization In our t-SNE visualizations, we examine the representations of CLIP, Im.Aug, and CLAP for all images within the Art Painting domain of the PACS dataset. Fig. 5 shows that CLAP's image representations display a marked inter-class separation and tighter intra-class clustering than those of" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Cai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 157, + 471, + 247 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 5: Image augmentation and CLAP both enhance CLIP's zero-shot with the \"[class]\" prompt and 1-shot robustness against adversarial attacks, with CLAP showing greater improvements." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 157, + 471, + 247 + ], + "lines": [ + { + "bbox": [ + 143, + 157, + 471, + 247 + ], + "spans": [ + { + "bbox": [ + 143, + 157, + 471, + 247 + ], + "type": "table", + "html": "
Setting MethodAvg. top-1 acc. (%) under adversarial attacks(↑)
FGSMPGD-20CW-20
PACSVLCSO.H.D.N.PACSVLCSO.H.D.N.PACSVLCSO.H.D.N.
ZS(C)CLIP86.865.657.922.529.12.010.110.727.41.57.47.6
Im.Aug88.069.655.137.931.32.110.49.029.41.77.05.8
CLAP88.771.958.544.230.83.210.611.229.82.38.18.0
1-shotCLIP66.745.234.322.534.816.05.611.318.90.74.53.2
Im.Aug79.447.137.123.555.216.18.512.523.20.95.13.4
CLAP89.652.237.123.973.421.27.412.527.01.15.03.5
", + "image_path": "e90635379d70b989e96ffbb6c23cedc3058a96090de3b74e92a3fe030027d55f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 149, + 264, + 239, + 332 + ], + "blocks": [ + { + "bbox": [ + 149, + 264, + 239, + 332 + ], + "lines": [ + { + "bbox": [ + 149, + 264, + 239, + 332 + ], + "spans": [ + { + "bbox": [ + 149, + 264, + 239, + 332 + ], + "type": "image", + "image_path": "bde6823173755b0d7136ddf77d7cc069d97ecb91807e9627482ca5d65737a5f9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 342, + 211, + 352 + ], + "lines": [ + { + "bbox": [ + 174, + 342, + 211, + 352 + ], + "spans": [ + { + "bbox": [ + 174, + 342, + 211, + 352 + ], + "type": "text", + "content": "(a) CLIP" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 263, + 264, + 353, + 332 + ], + "blocks": [ + { + "bbox": [ + 263, + 264, + 353, + 332 + ], + "lines": [ + { + "bbox": [ + 263, + 264, + 353, + 332 + ], + "spans": [ + { + "bbox": [ + 263, + 264, + 353, + 332 + ], + "type": "image", + "image_path": "baeaa687e75002e296a8c7a2f3c3461f0b160ba5e229cf058e61fddf7816bd28.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 284, + 342, + 330, + 353 + ], + "lines": [ + { + "bbox": [ + 284, + 342, + 330, + 353 + ], + "spans": [ + { + "bbox": [ + 284, + 342, + 330, + 353 + ], + "type": "text", + "content": "(b) Im.Aug" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 361, + 482, + 395 + ], + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 395 + ], + "type": "text", + "content": "Fig. 5: t-SNE visualizations of all images in the Art Painting of PACS dataset show CLAP outperforms the original CLIP and Im.Aug, with clearer inter-class distinctions and tighter intra-class clusters." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 377, + 264, + 468, + 332 + ], + "blocks": [ + { + "bbox": [ + 377, + 264, + 468, + 332 + ], + "lines": [ + { + "bbox": [ + 377, + 264, + 468, + 332 + ], + "spans": [ + { + "bbox": [ + 377, + 264, + 468, + 332 + ], + "type": "image", + "image_path": "b6856d7e0f9ca1e0d49d20f38781276f54f0eb1aab929c728b026ca9b8728e4f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 401, + 342, + 441, + 352 + ], + "lines": [ + { + "bbox": [ + 401, + 342, + 441, + 352 + ], + "spans": [ + { + "bbox": [ + 401, + 342, + 441, + 352 + ], + "type": "text", + "content": "(c) CLAP" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 417, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 482, + 453 + ], + "type": "text", + "content": "CLIP and Im.Aug. This observation suggests that CLAP's representations are more closely tied to content information and less influenced by style information, in contrast to the other two." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 469, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 469, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 469, + 482, + 601 + ], + "type": "text", + "content": "Ablations In Fig. 6, we assess the zero-shot capabilities of our model using two distinct prompts, ZS(C) and ZS(PC), on the VLCS dataset. This analysis forms part of an ablative study aimed at understanding the influence of various hyperparameters on model performance. Specifically, we examine: the dimensions of the latent layer within the MLP of the disentangled network, as illustrated in Fig. 6a; the temperature parameter " + }, + { + "bbox": [ + 130, + 469, + 482, + 601 + ], + "type": "inline_equation", + "content": "(\\tau)" + }, + { + "bbox": [ + 130, + 469, + 482, + 601 + ], + "type": "text", + "content": " in the loss function, as depicted in Fig. 6b; and the weight coefficient " + }, + { + "bbox": [ + 130, + 469, + 482, + 601 + ], + "type": "inline_equation", + "content": "(\\alpha)" + }, + { + "bbox": [ + 130, + 469, + 482, + 601 + ], + "type": "text", + "content": " during the inference stage, as shown in Fig. 6c. Our findings indicate that CLAP consistently enhances zero-shot performance across all tested configurations for both prompts, while also significantly reducing the gap between the performances elicited by each prompt. These results underscore the efficacy of CLAP in accommodating a wide range of hyper-parameters." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Prompt Augmentation Combinations We explore diverse combinations of our tailored prompt augmentation methods and examine Easy Data Augmentation (EDA) techniques [42] on the VLCS dataset. Each tested technique showcases CLAP's enhancements over CLIP, with details available in Appendix A.2." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 221, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 221, + 91, + 447, + 102 + ], + "type": "text", + "content": "CLAP: Contrastive Learning with Augmented Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 119, + 248, + 202 + ], + "blocks": [ + { + "bbox": [ + 136, + 119, + 248, + 202 + ], + "lines": [ + { + "bbox": [ + 136, + 119, + 248, + 202 + ], + "spans": [ + { + "bbox": [ + 136, + 119, + 248, + 202 + ], + "type": "image", + "image_path": "63de70ec9fd2118f1ca643f09b4a9c745e0844cabf7c0ffc7c53fd7b175694b7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 210, + 481, + 243 + ], + "lines": [ + { + "bbox": [ + 130, + 210, + 481, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 210, + 481, + 243 + ], + "type": "text", + "content": "Fig. 6: We conduct ablative study on hyper-parameter choices on the VLCS dataset, including latent dimensions, " + }, + { + "bbox": [ + 130, + 210, + 481, + 243 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 210, + 481, + 243 + ], + "type": "text", + "content": " values, and " + }, + { + "bbox": [ + 130, + 210, + 481, + 243 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 210, + 481, + 243 + ], + "type": "text", + "content": " values during the inference stage. CLAP continuously enhance CLIP's performance throughout the tested values." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 251, + 118, + 362, + 202 + ], + "blocks": [ + { + "bbox": [ + 251, + 118, + 362, + 202 + ], + "lines": [ + { + "bbox": [ + 251, + 118, + 362, + 202 + ], + "spans": [ + { + "bbox": [ + 251, + 118, + 362, + 202 + ], + "type": "image", + "image_path": "0803b3e90ceebc3f158b3216f06cf6745df399def7712334b047c53718ff8c82.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 365, + 118, + 478, + 202 + ], + "blocks": [ + { + "bbox": [ + 365, + 118, + 478, + 202 + ], + "lines": [ + { + "bbox": [ + 365, + 118, + 478, + 202 + ], + "spans": [ + { + "bbox": [ + 365, + 118, + 478, + 202 + ], + "type": "image", + "image_path": "5ffbb7fdae4d2b46c073f4e7d612fbd06af3fdce918f9a750b01810b40981b46.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 266, + 482, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 374 + ], + "type": "text", + "content": "Prompt Sources We assess the impact of different training prompt formats, originating from various synthetic sources, on the performance of the VLCS dataset, incorporating EDA techniques. Our evaluation includes our template-based prompts, LLM-generated prompts by ChatGPT-3.5 [3] (with the generation process detailed in Appendix B.2), prompts structured as \"a [random] style of [class],\" where \"[random]\" is filled with terms from a random word generator" + }, + { + "bbox": [ + 130, + 266, + 482, + 374 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 130, + 266, + 482, + 374 + ], + "type": "text", + "content": ", and prompts produced using the PromptStyler method [9]. The findings indicate that the training prompts with simpler forms tend to yield better performance, with detailed quantitative results presented in Appendix A.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 392, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 464 + ], + "type": "text", + "content": "Experiments on Different Model Scales In our repeated experiments assessing zero-shot performance on the ViT-L/14 and ResNet50x16 pre-trained with CLIP, we consistently find that CLAP improves zero-shot performance while also reducing performance variances. This consistent observation underscores CLAP's effectiveness in enhancing the quality of CLIP representations. For quantitative details supporting these findings, please see the Appendix C." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 481, + 220, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 481, + 220, + 494 + ], + "spans": [ + { + "bbox": [ + 131, + 481, + 220, + 494 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 506, + 482, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 650 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 650 + ], + "type": "text", + "content": "To enhance pre-trained CLIP-like models, this study delves into disentangling latent content variables. Through a causal analysis of the underlying generative processes of vision-language data, we discover that training a disentangled network in one modality can effectively disentangle content across both modalities. Given the high semantic nature of text data, we identify that disentanglement is more achievable within the language modality through text augmentation interventions. Building on these insights, we introduce CLAP (Contrastive Learning with Augmented Prompts) to acquire disentangled vision-language content features. Comprehensive experiments validate CLAP's effectiveness, demonstrating significant improvements in zero-shot and few-shot performance, and enhancing robustness against perturbations. We anticipate that our work will inspire further exploration into disentangling latent variables within vision-language models." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Cai et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 358, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 358, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 358, + 665 + ], + "type": "text", + "content": "2 https://github.com/vaibhavsingh97/random-word" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "text", + "content": "1. Albuquerque, I., Monteiro, J., Darvishi, M., Falk, T.H., Mitliagkas, I.: Generalizing to unseen domains via distribution matching. arXiv preprint arXiv:1911.00804 (2019). https://doi.org/10.48550/arXiv.1911.00804" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 170, + 480, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 170, + 480, + 190 + ], + "spans": [ + { + "bbox": [ + 138, + 170, + 480, + 190 + ], + "type": "text", + "content": "2. Bollen, K.A.: Structural equations with latent variables, vol. 210. John Wiley & Sons (1989). https://doi.org/10.1002/9781118619179" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 191, + 480, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 191, + 480, + 233 + ], + "spans": [ + { + "bbox": [ + 138, + 191, + 480, + 233 + ], + "type": "text", + "content": "3. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020), https://dl.acm.org/doi/abs/10.5555/3495724.3495883" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 234, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 480, + 266 + ], + "type": "text", + "content": "4. Carlini, N., Terzis, A.: Poisoning and backdooring contrastive learning. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=iC4UHbQ01Mp" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 267, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 480, + 299 + ], + "type": "text", + "content": "5. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: 2017 IEEE Symposium on Security and Privacy (SP). pp. 39-57. IEEE Computer Society (2017). https://doi.org/10.1109/SP.2017.49" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 300, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 480, + 319 + ], + "type": "text", + "content": "6. Chen, P., Liu, S., Zhao, H., Jia, J.: Gridmask data augmentation. arXiv preprint arXiv:2001.04086 (2020). https://doi.org/10.48550/arXiv.2001.04086" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 320, + 480, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 480, + 361 + ], + "type": "text", + "content": "7. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020), https://dl.acm.org/doi/abs/10.5555/3524938.3525087" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 363, + 480, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 363, + 480, + 395 + ], + "spans": [ + { + "bbox": [ + 138, + 363, + 480, + 395 + ], + "type": "text", + "content": "8. Chen, X., He, K.: Exploring simple siamese representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15750-15758 (2021). https://doi.org/10.1109/CVPR46437.2021.01549" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 396, + 480, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 480, + 437 + ], + "type": "text", + "content": "9. Cho, J., Nam, G., Kim, S., Yang, H., Kwak, S.: Promptstyler: Prompt-driven style generation for source-free domain generalization. In: Proceedings of the IEEE international conference on computer vision. pp. 15702-15712 (2023). https://doi.org/10.1109/ICCV51070.2023.01439" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 438, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 438, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 438, + 480, + 471 + ], + "type": "text", + "content": "0. Daunhawer, I., Bizeul, A., Palumbo, E., Marx, A., Vogt, J.E.: Identifiability results for multimodal contrastive learning. ICLR (2023), https://openreview.net/forum?id=U_2kuqoTcB" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 472, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 472, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 134, + 472, + 480, + 525 + ], + "type": "text", + "content": "11. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2020), https://openreview.net/forum?id=YicbFdNTTy" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 525, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 525, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 134, + 525, + 480, + 555 + ], + "type": "text", + "content": "12. Fort, S.: Adversarial vulnerability of powerful near out-of-distribution detection. arXiv preprint arXiv:2201.07012 (2022). https://doi.org/10.48550/arXiv.2201.07012" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 557, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 557, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 134, + 557, + 480, + 599 + ], + "type": "text", + "content": "13. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision 132(2), 581-595 (2024). https://doi.org/10.1007/s11263-023-01891-x" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "text", + "content": "14. Ge, C., Huang, R., Xie, M., Lai, Z., Song, S., Li, S., Huang, G.: Domain adaptation via prompt learning. arXiv preprint arXiv:2202.06687 (2022). https://doi.org/10.48550/arXiv.2202.06687" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "15. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. In: International Conference on Learning Representations (2015). https://doi.org/10.48550/arXiv.1412.6572" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 221, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 221, + 91, + 447, + 102 + ], + "type": "text", + "content": "CLAP: Contrastive Learning with Augmented Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 171 + ], + "type": "text", + "content": "16. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020), https://dl.acm.org/doi/abs/10.5555/3495724.3497510" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 171, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 481, + 214 + ], + "type": "text", + "content": "17. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022). https://doi.org/10.1109/CVPR52688.2022.01553" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 214, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 481, + 257 + ], + "type": "text", + "content": "18. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020). https://doi.org/10.1109/CVPR42600.2020.00975" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 257, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 257, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 133, + 257, + 481, + 300 + ], + "type": "text", + "content": "19. Hong, T., Guo, X., Ma, J.: Itmix: Image-text mix augmentation for transferring clip to image classification. In: 2022 16th IEEE International Conference on Signal Processing (ICSP). vol. 1, pp. 129-133. IEEE (2022). https://doi.org/10.1109/ICSP56322.2022.9965292" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 300, + 481, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 344 + ], + "type": "text", + "content": "20. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904-4916. PMLR (2021), https://proceedings.mlr.press/v139/jia21b.html" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 344, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 344, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 344, + 481, + 386 + ], + "type": "text", + "content": "21. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19113-19122 (2023). https://doi.org/10.1109/CVPR52729.2023.01832" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 386, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 430 + ], + "type": "text", + "content": "22. Kong, L., Xie, S., Yao, W., Zheng, Y., Chen, G., Stojanov, P., Akinwande, V., Zhang, K.: Partial disentanglement for domain adaptation. In: International Conference on Machine Learning. pp. 11455-11472. PMLR (2022), https://proceedings.mlr.press/v162/kong22a.htm1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 430, + 481, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 461 + ], + "type": "text", + "content": "23. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.M.: Deeper, broader and artier domain generalization. In: Proceedings of the IEEE international conference on computer vision. pp. 5542-5550 (2017). https://doi.org/10.1109/ICCV.2017.591" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 461, + 481, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 481, + 494 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 481, + 494 + ], + "type": "text", + "content": "24. Li, H., Wang, X., Zhang, Z., Yuan, Z., Li, H., Zhu, W.: Disentangled contrastive learning on graphs. Advances in Neural Information Processing Systems 34, 21872-21884 (2021), https://dl.acm.org/doi/10.5555/3540261.3541935" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 494, + 481, + 526 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 481, + 526 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 481, + 526 + ], + "type": "text", + "content": "25. Li, Y., Liang, F., Zhao, L., Cui, Y., Ouyang, W., Shao, J., Yu, F., Yan, J.: Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 525, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 568 + ], + "type": "text", + "content": "26. Liu, Y., Zhang, Z., Gong, D., Gong, M., Huang, B., van den Hengel, A., Zhang, K., Shi, J.Q.: Identifiable latent polynomial causal models through the lens of change. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=ia9fK01Vjq" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "text", + "content": "27. Liu, Y., Zhang, Z., Gong, D., Gong, M., Huang, B., Hengel, A.v.d., Zhang, K., Shi, J.Q.: Identifying weight-variant latent causal models. arXiv preprint arXiv:2208.14153 (2022). https://doi.org/10.48550/arXiv.2208.14153" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 601, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 632 + ], + "type": "text", + "content": "28. Liu, Y., Zhang, Z., Gong, D., Gong, M., Huang, B., Hengel, A.v.d., Zhang, K., Shi, J.Q.: Identifiable latent neural causal models. arXiv preprint arXiv:2403.15711 (2024). https://doi.org/10.48550/arXiv.2403.15711" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "29. Liu, Y., Zhang, Z., Gong, D., Gong, M., Huang, B., Zhang, K., Shi, J.Q.: Identifying latent causal content for multi-source domain adaptation. arXiv preprint arXiv:2208.14161 (2022). https://doi.org/10.48550/arXiv.2208.14161" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Cai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "text", + "content": "30. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. In: International Conference on Learning Representations (2018), https://openreview.net/forum?id=rJzIBfZAb" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 184 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 184 + ], + "type": "text", + "content": "31. Mahajan, D., Tople, S., Sharma, A.: Domain generalization using causal matching. In: International Conference on Machine Learning. pp. 7313-7324. PMLR (2021), https://proceedings.mlr.press/v139/mahajan21b.html" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 185, + 482, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 185, + 482, + 207 + ], + "spans": [ + { + "bbox": [ + 130, + 185, + 482, + 207 + ], + "type": "text", + "content": "32. Mamooler, S.: Clip explainability. https://github.com/sMamooler/CLIP_Explainability, accessed: 2024-03-06" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 207, + 482, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 207, + 482, + 250 + ], + "spans": [ + { + "bbox": [ + 130, + 207, + 482, + 250 + ], + "type": "text", + "content": "33. Mao, C., Geng, S., Yang, J., Wang, X., Vondrick, C.: Understanding zero-shot adversarial robustness for large-scale models. In: The Eleventh International Conference on Learning Representations (2022), https://openreview.net/forum?id=P4bXcawRi5J" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 251, + 482, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 283 + ], + "type": "text", + "content": "34. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018). https://doi.org/10.48550/arXiv.1807.03748" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 285, + 482, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 329 + ], + "type": "text", + "content": "35. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE international conference on computer vision. pp. 1406-1415 (2019). https://doi.org/10.1109/ICCV.2019.00149" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 331, + 482, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 384 + ], + "type": "text", + "content": "36. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021), https://proceedings.mlr.press/v139/radford21a.html" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 386, + 482, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 430 + ], + "type": "text", + "content": "37. Rahman, M.M., Fookes, C., Baktashmotlagh, M., Sridharan, S.: Multi-component image translation for deep domain generalization. In: 2019 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 579-588. IEEE (2019). https://doi.org/10.1109/WACV.2019.00067" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 431, + 482, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 431, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 431, + 482, + 475 + ], + "type": "text", + "content": "38. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021), https://proceedings.mlrpress/v139/ramesh21a.html" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 476, + 482, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 476, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 476, + 482, + 520 + ], + "type": "text", + "content": "39. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022). https://doi.org/10.1109/CVPR52729.2023.01389" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 521, + 482, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 565 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 565 + ], + "type": "text", + "content": "40. Sanchez, E.H., Serrurier, M., Ortner, M.: Learning disentangled representations via mutual information estimation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 205-221. Springer (2020). https://doi.org/10.1007/978-3-030-58542-6_13" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 566, + 482, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 566, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 566, + 482, + 609 + ], + "type": "text", + "content": "41. Von Kügelgen, J., Sharma, Y., Gresele, L., Brendel, W., Schölkopf, B., Besserve, M., Locatello, F.: Self-supervised learning with data augmentations provably isolates content from style. Advances in neural information processing systems 34, 16451-16467 (2021), https://d1.acm.org/doi/10.5555/3540261.3541519" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 611, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 482, + 665 + ], + "type": "text", + "content": "42. Wei, J., Zou, K.: Eda: Easy data augmentation techniques for boosting performance on text classification tasks. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). pp. 6382-6388 (2019). https://doi.org/10.18653/v1/D19-1670" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 221, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 221, + 91, + 447, + 102 + ], + "type": "text", + "content": "CLAP: Contrastive Learning with Augmented Prompts" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 401 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 171 + ], + "type": "text", + "content": "43. Wortsman, M., Ilharco, G., Kim, J.W., Li, M., Kornblith, S., Roelofs, R., Lopes, R.G., Hajishirzi, H., Farhadi, A., Namkoong, H., et al.: Robust fine-tuning of zero-shot models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7959-7971 (2022). https://doi.org/10.1109/CVPR52688.2022.00780" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 215 + ], + "type": "text", + "content": "44. Yang, M., Liu, F., Chen, Z., Shen, X., Hao, J., Wang, J.: Causalvae: Disentangled representation learning via neural structural causal models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9593-9602 (2021). https://doi.org/10.1109/CVPR46437.2021.00947" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "type": "text", + "content": "45. Yang, W., Mirzasoleiman, B.: Robust contrastive language-image pretraining against adversarial attacks. arXiv preprint arXiv:2303.06854 (2023). https://doi.org/10.48550/arXiv.2303.06854" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "46. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: IEEE International Conference on Computer Vision (ICCV) (2023). https://doi.org/10.1109/ICCV51070.2023.00355" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 282, + 481, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 324 + ], + "type": "text", + "content": "47. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022). https://doi.org/10.1109/CVPR52688.2022.01631" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "type": "text", + "content": "48. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022). https://doi.org/10.1007/s11263-022-01653-1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 358, + 481, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 401 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 401 + ], + "type": "text", + "content": "49. Zimmermann, R.S., Sharma, Y., Schneider, S., Bethge, M., Brendel, W.: Contrastive learning inverts the data generating process. In: International Conference on Machine Learning. pp. 12979-12990. PMLR (2021), https://proceedings.mlr.press/v139/zimmermann21a.html" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Cai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CLEO_ Continual Learning of Evolving Ontologies/7399fe90-4063-4ee4-8949-4afe680e9068_content_list.json b/2024/CLEO_ Continual Learning of Evolving Ontologies/7399fe90-4063-4ee4-8949-4afe680e9068_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d44155efcdc6588d186b6bca57cc325c7afb2bd3 --- /dev/null +++ b/2024/CLEO_ Continual Learning of Evolving Ontologies/7399fe90-4063-4ee4-8949-4afe680e9068_content_list.json @@ -0,0 +1,1527 @@ +[ + { + "type": "text", + "text": "CLEO: Continual Learning of Evolving Ontologies", + "text_level": 1, + "bbox": [ + 215, + 140, + 785, + 165 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shishir Muralidhara1, Saqib Bukhari2, Georg Schneider2, Didier Stricker1,3, and René Schuster1,3", + "bbox": [ + 215, + 188, + 789, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ DFKI - German Research Center for Artificial Intelligence", + "bbox": [ + 295, + 231, + 702, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 ZF Friedrichshafen AG", + "bbox": [ + 416, + 246, + 584, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ RPTU - University of Kaiserslautern-Landau", + "bbox": [ + 341, + 258, + 660, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{firstname这个名字}@{dfki.de，zf.com}", + "bbox": [ + 354, + 273, + 647, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Continual learning (CL) addresses the problem of catastrophic forgetting in neural networks, which occurs when a trained model tends to overwrite previously learned information, when presented with a new task. CL aims to instill the lifelong learning characteristic of humans in intelligent systems, making them capable of learning continuously while retaining what was already learned. Current CL problems involve either learning new domains (domain-incremental) or new and previously unseen classes (class-incremental). However, general learning processes are not just limited to learning information, but also refinement of existing information. In this paper, we define CLEO - Continual Learning of Evolving Ontologies, as a new incremental learning setting under CL to tackle evolving classes. CLEO is motivated by the need for intelligent systems to adapt to real-world ontologies that change over time, such as those in autonomous driving. We use Cityscapes, PASCAL VOC, and Mapillary Vistas to define the task settings and demonstrate the applicability of CLEO. We highlight the shortcomings of existing CIL methods in adapting to CLEO and propose a baseline solution, called Modelling Ontologies (MoOn). CLEO is a promising new approach to CL that addresses the challenge of evolving ontologies in real-world applications. MoOn surpasses previous CL approaches in the context of CLEO.", + "bbox": [ + 259, + 325, + 743, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Continual Learning $\\cdot$ Continual Semantic Segmentation $\\cdot$ Class-Incremental Learning", + "bbox": [ + 259, + 617, + 740, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 671, + 375, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Advancements in AI have led to the emergence of systems that can match or even outperform human abilities in various domains. However, these systems are often confined to specific datasets and tasks and their efficacy is affected when faced with new tasks or data. This presents a significant hurdle in practical real-world AI applications where the circumstances are constantly evolving. For example, a self-driving car needs to adapt to new traffic regulations, signage, and weather conditions. When a trained model encounters new information or tasks, it experiences a phenomenon known as catastrophic forgetting [26]. This issue arises from the rigidity of neural networks, where previously learned knowledge is", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6a67cab48edb88c9a6449a043885b11c7bcf63cb7b9690cb17be6611109678e2.jpg", + "image_caption": [ + "Fig. 1: CLEO extends class-incremental learning (CIL) by removing the constraint that incremental sets of classes must be non-overlapping. This allows for more flexible and nuanced learning, as existing classes can be refined into more specific classes." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 247 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "overwritten when learning something new. Consequently, the model's performance significantly deteriorates, which can have severe repercussions in scenarios where the model has to acclimate to dynamic environments or critical applications where performance is imperative. Therefore, tackling forgetting is crucial when developing AI systems with lifelong learning.", + "bbox": [ + 212, + 325, + 782, + 400 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, learning in humans is continuous and dynamic through which we continually assimilate new information while preserving and enriching previously learned knowledge. The goal of CL is to instill neural networks with this learning characteristic and make them capable of extending when objectives change. In incremental learning, which involves learning a sequence of tasks, the term task can represent changes either in the distribution of data, referred to as domain-incremental learning (DIL), or changes in the output space, which is known as class-incremental learning (CIL). However, learning is not limited only to the acquisition of new knowledge; it also involves refining and deepening our grasp of existing information. To illustrate, our initial understanding of animals might have been confined to recognizing a broad categorization, like distinguishing between birds and other animals. However, over time, our learning process matures to incorporate finer, more nuanced insights about individual animals. This highlights the necessity of broadening the scope of CL to accommodate both the introduction of novel classes, and also the enhancement of existing knowledge.", + "bbox": [ + 212, + 402, + 787, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose a novel incremental setting termed CLEO - Continual Learning of Evolving Ontologies. CLEO represents a progressive extension of CIL, overcoming the confining assumption that all newly encountered classes are entirely novel and previously unseen. CLEO acknowledges the dynamic nature of learning, where classes may not only emerge but also evolve over time. CLEO combines elements of novelty and refinement, paving the way for more resilient AI systems. The real-world relevance of CLEO is exemplified by the Mapillary Vistas dataset [30], where an initial set of classes has evolved into more fine-grained and specific classes. An illustration of CLEO is presented in Fig. 1 using the CIFAR-10 dataset [17] as an example. In the initial task, the system learns two broad categories, namely animals and vehicles. Subsequently, more specific classes, such as bird, frog, plane, and ship, are split from the original classes. CLEO presents a challenge where the system must not only learn the evolved", + "bbox": [ + 212, + 643, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. Muralidhara et al.", + "bbox": [ + 271, + 114, + 411, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "classes, which it had previously encountered within a different semantic context, but also retain the remaining information. In this example, after bird and frog were separated from the animal class, cat and dog are left as the remaining domestic animal class. As in CIL, CLEO also covers the case that entirely new classes are learned, like e.g. the classes horse and deer in the last task of Fig. 1. CLEO embodies the fundamental idea that AI systems must possess the ability to adapt to dynamic shifts in which classes emerge, evolve, and split over time. By integrating this adaptability into AI systems, we enhance their robustness in real-world applications. This paper makes several significant contributions:", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Formalizing Continual Learning of Evolving Ontologies: A primary contribution is the introduction and formalization of the concept of learning evolving ontologies, from a continual learning perspective.", + "- Defining Task Settings: To provide a comprehensive understanding of CLEO, we define several task settings for semantic segmentation, using three diverse datasets: Cityscapes [8], PASCAL VOC [11], and Mapillary Vistas [30].", + "- Evaluation Metrics for CLEO: By identifying class groups and measuring their performance, we define a more nuanced evaluation of how effectively the system learns refined classes and retains information.", + "- Baseline Solution - MoOn: To operationalize the concept for semantic segmentation, we propose a baseline solution called \"Modelling Ontologies\" (MoOn), to handle the challenges inherent in CLEO." + ], + "bbox": [ + 225, + 292, + 784, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 511, + 387, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Continual Learning", + "text_level": 1, + "bbox": [ + 215, + 542, + 421, + 558 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Continual learning (CL), also known as lifelong learning, and its sub-field of incremental learning, are critical areas of machine learning that address a fundamental limitation of traditional machine learning models: Their rigidity in learning new tasks. CL is a learning paradigm that aims to train models that can learn new tasks and information without forgetting what was previously learned. This contrasts traditional machine learning, where models are typically trained on a single dataset and then remain fixed. Continual learning presents a challenge of balancing between learning new information and retaining the old [27]. Overemphasizing learning new information leads to forgetting the old while focusing too much on retaining old information hinders learning the new task. CL approaches can be grouped into three categories: Architecture-based, replay-based methods, and regularization-based methods.", + "bbox": [ + 212, + 566, + 787, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Architecture-based approaches modify the architecture of neural networks to accommodate new tasks without overwriting previous information. These include methods such as parameter isolation [12, 24, 25], dynamically growing networks [1, 32, 39], or incorporation of additional components such as memory modules [2, 34]. These methods can face scalability challenges as network architecture grows, increasing memory and computational demands.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "CLEO: Continual Learning of Evolving Ontologies", + "bbox": [ + 395, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Replay-based approaches are of two types: Experience replay or rehearsal, involves storing a few instances from the previous task [3,6,23], which are used during training on a new task. Generative replay substitutes the memory buffer with a generative model to sample previous task instances when learning on a new task [15,33,35]. Challenges include privacy, data availability, memory and computational constraints in training generative models.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Regularization-based approaches [16, 20, 41] strike a balance between stability and plasticity when learning new tasks by introducing a regularization term into the loss function. One of the approaches to regularization is to determine the importance of parameters to the previous tasks and penalize overwriting them. Distillation is another approach which preserves previously learned knowledge by distilling information from the old model to the new.", + "bbox": [ + 212, + 237, + 789, + 328 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Continual Semantic Segmentation", + "text_level": 1, + "bbox": [ + 214, + 348, + 540, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Continual semantic segmentation (CSS) [5,37,38,40] is a specialized domain within CL that focuses on the task of semantic segmentation. Incremental Learning Techniques [28] (ILT) is a regularization-based method that distils knowledge from previously trained task models and additionally freezes the encoder for the current task, which acts as an additional regularization constraint. Modeling the Background [4] (MiB) addresses the problem of background shift in incremental segmentation. It proposes a novel distillation loss that accounts for the changing background between tasks. AWT [13] addresses the background shift in classifier initialization, and transfers the most relevant weights from the previous classifiers to the new classifier. In our baseline solution for CLEO, we will build on the ideas of MiB and extend it to model the semantic shift for all classes. PLOP [10] argues that treating the background class as a single class as done in MiB, which includes both previously encountered and novel classes can lead to catastrophic forgetting of the old classes as it learns new classes. PLOP addresses this problem by pseudolabeling old classes in the background. EWF [36] combines the weights of the old and new model trained with regularization to give a more balanced model. SDR [29] preserves latent representations of old classes through prototype matching and frees space for new classes through feature sparsity and contrastive learning. The decoder preserves previous knowledge via output-level distillation. RCIL [42] uses a representation compensation module to decouple the representation learning of old and new classes. This helps to prevent the model from forgetting old classes when learning new classes. REMINDER [31] proposes a class-similarity weighted distillation that distills knowledge from a previous model on old classes that are similar to the new classes. This reduces forgetting of old classes and improves learning of new classes.", + "bbox": [ + 212, + 371, + 789, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Evolving Ontologies", + "text_level": 1, + "bbox": [ + 214, + 770, + 426, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An emerging area within incremental learning focuses on learning evolving classes, which involves the refinement of existing knowledge, where a broad initial grouping of classes evolves over time into more specific and nuanced classes. In contrast,", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. Muralidhara et al.", + "bbox": [ + 271, + 114, + 411, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in CIL, each task consists of a non-overlapping subsets of novel classes.", + "bbox": [ + 215, + 146, + 725, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hierarchical Label Expansion [18] (HLE) for classification involves a model initially learning coarse classes, which then evolve into more fine-grained categories during each task across varying hierarchical levels. The expansion is restrictive, where the entire subset of classes at the current level expands in the next level during each task. CLEO differs from HLE in the following aspects:", + "bbox": [ + 215, + 176, + 785, + 252 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- CLEO allows a more flexible evolution at any point within the sequence of tasks, in which certain classes can remain (partially) unexpanded or expand into sub-groups.", + "- The baseline solution for HLE is a rehearsal-based method that uses pseudolabeling from previous classifiers - one for each level of the hierarchy - for determining the importance of data. We present a regularization-based solution that uses knowledge distillation.", + "- HLE considers image classification as application, while CLEO additionally deals with the more challenging problem of semantic segmentation." + ], + "bbox": [ + 223, + 262, + 785, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Learning with Evolving Class Ontologies [22] (LECO) has been the first attempt to address evolving ontologies in semantic segmentation. However, despite its intended framing as a continual learning problem, LECO deviates from the defining principles of continual learning. We address these shortcomings and define this problem strictly within the bounds of continual learning. The primary distinctions between our framework and LECO are outlined below:", + "bbox": [ + 215, + 407, + 785, + 498 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- LECO requires and stores all historical data, a constraint that contradicts the definition of CL [19], that all data is not available at once.", + "- LECO primarily revolves around the dilemma of whether to label new data or relabel the old data with the current label space. In contrast, CLEO confines itself to utilizing the input data and annotations within the dataset.", + "- Due to the availability of historical data, LECO employs naive strategies such as fine-tuning, freezing, or joint training. We highlight the limitations of existing CIL methods and propose a baseline specifically for CLEO.", + "- CLEO explicitly addresses catastrophic forgetting, which is one of the core problems of continual learning. Since LECO retrans entirely using all the data, the problem of catastrophic forgetting does not even arise." + ], + "bbox": [ + 223, + 512, + 782, + 675 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 CLEO", + "text_level": 1, + "bbox": [ + 217, + 702, + 313, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Deep learning has drawn inspiration from biological systems at various levels, from the way neurons form the fundamental building blocks of artificial neural networks to learning paradigms such as curriculum learning and reinforcement learning. Continual learning is another learning paradigm that aims to instill lifelong learning in intelligent systems. As discussed previously, current problems in continual learning involve adaptation to different domains, such as handling adverse conditions in autonomous driving or learning new classes over time. A", + "bbox": [ + 215, + 734, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "CLEO: Continual Learning of Evolving Ontologies", + "bbox": [ + 395, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "common characteristic in both these scenarios is that the task always represents new and previously unseen data. However, it is our contention that the scope of continual learning should not be limited solely to the discovery of novel classes. It should also include the enhancement and refinement of existing knowledge. This aligns closely with the way humans continually adapt and refine knowledge through neuroplasticity.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In light of this perspective, we introduce CLEO - Continual Learning of Evolving Ontologies. CLEO is centered around the idea of removing the constraints of class incremental learning and allows classes that have a hierarchical relation (partially) overlap between tasks, thereby facilitating the refinement of already learned information. CLEO is a real and open-world problem and this is exemplified in the Mapillary Vistas dataset [30]. The initial version of the dataset defined 66 classes, which later evolved into 124 classes in the latest version. It is noteworthy to observe the transformation within class categorization. On one hand, classes that were previously categorized under unlabeled, such as traffic island emerge as novel classes. Conversely, several pre-existing classes undergo refinement, leading to the creation of new, more specific categories. In Fig. 2, we can observe a tangible example of this evolution, where the previous billboard class has evolved into more specialized classes, including signage classes for advertisements, information, stores, and differentiating between the front and back of the signs. In critical applications like autonomous driving, it becomes imperative for AI systems to swiftly adapt to these nuanced changes, perpetually learn, and enhance scene understanding precision.", + "bbox": [ + 212, + 237, + 789, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Let $T = \\{0,1,2,\\dots,n\\}$ denote a sequence of $n + 1$ tasks. Each task is associated with a specific subset of classes denoted as $C_t$ . The union of all classes across tasks is $C = \\bigcup_{t=0}^{n} C_t$ , representing the entire set of classes encountered during the learning process. At each task $t$ , a subset of classes $C_t$ is introduced incrementally similar to CIL. $C_{0:t} = C_0 \\cup \\ldots \\cup C_{t-1} \\cup C_t$ is the set of all classes seen until step $t$ . CLEO acknowledges the possibility of class splitting, where an existing class $c_{t-1}$ from $C_{0:t-1}$ can evolve into multiple new classes in $C_t$ . Formally this can be expressed as, $C_t \\subset C$ , $C_i \\cap C_j = \\phi$ for $i \\neq j$ and the class splitting is expressed as $c_{t-1} \\in C_{0:t-1} \\to \\{c_{k+1}, c_{k+2}, \\ldots, c_{k+m}\\} \\subseteq C_t$ . Note that the splitting of a class is not necessarily exhaustive, i.e., $c_{t-1} \\in C_{0:t-1}$ may persist. An example for such an evolution of classes is given in Fig. 1. The learning objective of a task remains consistent with CIL: Train a model to recognize and differentiate the current subset of classes $C_t$ , while ensuring that the model retains knowledge from the previously encountered classes in $C_{0:t-1}$ . As in CIL, the model is not provided access to data from previous tasks.", + "bbox": [ + 212, + 508, + 789, + 737 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 MoOn: Modelling Ontologies for CLEO", + "text_level": 1, + "bbox": [ + 215, + 760, + 633, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Class-incremental learning involves the process of training a model to incrementally learn new classes. At each increment, the model is presented with images and the corresponding ground truth which contains annotations for the new classes", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. Muralidhara et al.", + "bbox": [ + 271, + 114, + 411, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4b75010f392e681e944f2d94b7bc5d60b8963568403800a73d85ac0878de5d81.jpg", + "image_caption": [ + "Fig. 2: The Mapillary Vistas dataset [30] provides a real-world example for an evolving ontology. This figure compares ground truth annotations for the same image in two dataset versions. It illustrates the introduction of new classes, such as traffic island and the finer categorization of existing classes such as billboard." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 401, + 251 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cc4e37b629060ca9560efa75e50c6895866c1e867dede7c8351b0bb16a904033.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 143, + 593, + 251 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/251045c1e490a729fb916f4266cdfdeadf7dc8be29e52a44850f2132129c3aea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 143, + 784, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "being learned. In class-incremental semantic segmentation (CISS), at each step, all unlabeled areas are marked as background. Thus in CISS, the phenomenon of semantic background shift occurs, because in each task, the meaning of the background class $(bg)$ changes. Distillation loss is a technique of transferring knowledge from an old model $f_{t-1}$ to a new model $f_t$ , and this is achieved by comparing the old model's prediction to the new model's predictions. The distillation loss enforces the new model's predictions to resemble those of the old model by penalizing any deviations and this helps mitigate forgetting. Consider an image $x$ with $N$ pixels and the corresponding ground truth $y$ at the current task $t$ with the class set $C_t$ . The standard distillation loss in [4] is defined as follows:", + "bbox": [ + 212, + 347, + 787, + 510 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {k d} ^ {t} (x, y) = - \\frac {1}{N} \\sum_ {i \\in x} \\sum_ {c \\in C _ {0: t - 1}} p _ {x} ^ {t - 1} (i, c) \\log \\hat {p} _ {x} ^ {t} (i, c) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 510, + 785, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $p_x^{t-1}(i, c)$ is the probability for class $c$ in pixel $i$ given by the model $f_{t-1}$ and $\\hat{p}_x^t(i, c)$ is for class $c$ in pixel $i$ given by the model $f_t$ but re-normalized across classes $C_{0:t-1}$ :", + "bbox": [ + 212, + 551, + 787, + 599 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {p} _ {x} ^ {t} (i, c) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} c \\in C _ {t} \\backslash \\{b g \\} \\\\ \\frac {p _ {x} ^ {t} (i , c)}{\\sum_ {k \\in C _ {0 : t - 1}} p _ {x} ^ {t} (i , k)} & \\text {i f} c \\in C _ {0: t - 1} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 608, + 785, + 664 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Standard distillation ignores the dynamic nature of the background in incremental semantic segmentation. The background class in segmentation is a catch-all class that includes all previously seen classes and potential future classes. Furthermore, the background class is present in all tasks and changes as new classes are added, leading to background shift. A class $c \\in C_t$ added during the current task was previously part of the background in $C_{t-1}$ . For a pixel representing a new class at the current task $t$ , the old model $f_{t-1}$ would predict the background class, and if standard distillation loss was applied naively, it would penalize the new model $f_t$ for predicting the correct class. Modeling the Background (MiB) [4] proposes a novel distillation loss that explicitly addresses the background shift in incremental segmentation. MiB does not directly compare the prediction of the", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "CLEO: Continual Learning of Evolving Ontologies", + "bbox": [ + 395, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/28a6aea6c00ffa2273b88ce4ef87e7ca1ce375a812a52d46eb31df18b2c89f4a.jpg", + "image_caption": [ + "Fig. 3: Modelling Ontologies (MoOn) generalizes the idea of MiB [4] for all classes. During distillation, forgetting is mitigated, by grouping class-logits according to the evolution of the class ontology. In this example, the teacher network of the previous task predicts the vehicle class, while the model tries to learn bicycle. To avoid penalizing this, the predictions for vehicle and bicycle are combined before the distillation loss." + ], + "image_footnote": [], + "bbox": [ + 215, + 147, + 782, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "new model $f_{t}$ with the corresponding prediction of the old model $f_{t-1}$ . Instead, the background prediction by $f_{t-1}$ is compared with the $f_{t}$ prediction of either a new class or the background. Eq. (3) modifies Eq. (2) by additionally considering the whole label space instead of limiting it to $C_{0:t-1}$ . MiB acknowledges that pixels that were thought to be part of the background by the old model might actually be the new classes the current model is trying to learn, thus avoiding penalizing the probabilities of new classes being learned.", + "bbox": [ + 212, + 446, + 787, + 551 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {p} _ {x} ^ {t} (i, c) = \\left\\{ \\begin{array}{l l} p _ {x} ^ {t} (i, c) & \\text {i f} c \\neq b g \\\\ \\sum_ {k \\in C _ {t}} p _ {x} ^ {t} (i, k) & \\text {i f} c = b g \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 561, + 787, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MiB is designed for CIL, where the new classes being learned at the current step, were previously unseen and are always a part of the background class. Therefore, it would suffice to take into consideration the background class, and hence modeling the background. Whereas CLEO is designed for learning of evolving classes and the new classes being learned may have been previously unseen and part of the background or been part of a broader class. As seen in Fig. 3, the class bike being learned at the current step may have been previously learned as a vehicle class grouped with other classes like truck and bus. Therefore we need to take into consideration that within CLEO, new classes may not always have been a part of the background class previously.", + "bbox": [ + 212, + 613, + 787, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To this end, we propose Modelling Ontologies (MoOn) as a baseline solution for CLEO, which can be seen as a generalized version of MiB that accounts for the fact that new classes may emerge from any existing class, and not just the background class. Similar to how MiB considered the possibility of the new class", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. Muralidhara et al.", + "bbox": [ + 271, + 114, + 411, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pixel belonging to the background class previously, MoOn additionally includes the possibility of the pixel belonging to an existing class. In MiB, the old model $f_{t-1}$ may predict the background class for the new class being learned, whereas the old model in MoOn may predict an existing class $c \\in C_{0:t-1}$ . Therefore, the correct distillation would be to distill the knowledge from the old model $f_{t-1}$ for the existing class $c \\in C_{0:t-1}$ that is being split into new classes in $C_t$ . In the illustrated scenario depicted in Fig. 3, distillation would involve transferring knowledge from the old model regarding the vehicle class, while training the new model to predict the bike class. To achieve this, at each step $t$ , MoOn uses a set $S_t$ of class maps $(c_{t-1} \\to \\{c_{k+1}, c_{k+2}, \\dots, c_{k+m}\\} = C_{new})$ to identify the class $c_{t-1} \\in C_{0:t-1}$ from which the current classes in $C_t$ are split. If the new class $c_t$ is completely new, then the superclass $c_{t-1}$ in the class mapping would be the background class $bg$ . We define $C_{e,t}$ as the set of all evolving classes $c_{t-1}$ in $S_t$ . The modified distillation loss for learning evolving classes in CLEO is given as:", + "bbox": [ + 212, + 146, + 787, + 358 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {p} _ {x} ^ {t} (i, c) = \\left\\{ \\begin{array}{l l} p _ {x} ^ {t} (i, c) & \\text {i f} c \\notin C _ {e, t} \\\\ p _ {x} ^ {t} (i, c _ {t - 1}) + \\sum_ {c _ {n} \\in C _ {\\text {n e w}}} p _ {x} ^ {t} (i, c _ {n}) & \\text {i f} c \\in C _ {e, t} \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 371, + 787, + 426 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 215, + 448, + 334, + 462 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Towards the introduction of CLEO for semantic segmentation, we consider the following three diverse datasets:", + "bbox": [ + 212, + 474, + 785, + 505 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Cityscapes: Cityscapes [8] is a large-scale dataset of urban street scenes captured in 50 different cities under daytime and good weather conditions. The dataset provides annotations for 30 classes out of which 19 are used for evaluation and grouped into 7 categories. An overview of the hierarchy of the 19 semantic classes grouped by their parent categories is presented in the supplementary.", + "- PASCAL VOC: PASCAL VOC 2012 [11] is part of the Visual Object Classes Challenge, which ran from 2005 to 2012. PASCAL VOC is a more general dataset like COCO [21] consisting of 20 categories, including vehicles, household objects, animals, and persons. The categorization by [43] of these classes along with intermediate groupings is shown in the supplementary.", + "- Mapillary Vistas: Mapillary Vistas [30] is a large-scale street-level image dataset for understanding street scenes around the world. It covers a wider range of geographic locations and diverse weather and illumination conditions. The initial version of the dataset provided annotations for 66 classes, which later nearly doubled into 124 classes in the current version v2.0." + ], + "bbox": [ + 225, + 518, + 787, + 760 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Baselines and Implementation", + "text_level": 1, + "bbox": [ + 215, + 782, + 511, + 799 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluate our method's performance against standard CL baselines, namely fine-tuning and joint training. Fine-tuning involves using the previous step model", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "CLEO: Continual Learning of Evolving Ontologies", + "bbox": [ + 395, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "as the starting point when learning on the current task. This is similar to transfer learning and even results in positive forward transfer. However, since there is no explicit attempt to prevent catastrophic forgetting, it involves the highest amount of forgetting and forms the lower baseline. The joint training model involves learning all of the tasks together in a single step, and as there is no incremental learning there is zero forgetting and this forms the theoretical upper baseline. Since in CLEO a redefinition of classes may occur because of the evolving ontology, we define the set of classes for the joint training as the final set of classes after the last task. Additionally, we compare against the state-of-the-art (SOTA) methods such as MiB [4], PLOP [10] and RCIL [42].", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For MiB, and our proposed approach, we adopt the framework proposed by MiB [4]. In the case of PLOP and RCIL, which involve modifications specific to their approaches, we adopt the RCIL framework [42]. All these approaches use the DeepLabV3 [7] architecture as segmentation network, utilizing a ResNet-101 [14] backbone with an output stride of 16. The backbone's initialization uses a pre-trained model from ImageNet [9]. The training methodology utilizes SGD, with a learning rate of 0.01 for the initial learning step and 0.001 for subsequent steps. Across all learning steps, we train the model using a batch size of 24, for 50 epochs on Cityscapes and for 30 epochs on PASCAL and Mapillary Vistas.", + "bbox": [ + 212, + 311, + 789, + 450 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Evaluation Metrics", + "text_level": 1, + "bbox": [ + 215, + 486, + 419, + 500 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Mean Intersection-over-Union (mIoU) is a commonly used metric for evaluating semantic segmentation models. It measures the average overlap between the predicted segmentation and the ground truth segmentation for all classes. To calculate the mIoU, the Intersection-over-Union (IoU) is first calculated for each class, and then the IoU scores for all classes are averaged. The performance of CL models are often evaluated by measuring the amount of information forgotten by comparing the model's performance on old tasks against the joint training baseline. Within the context of CLEO, while learning evolving classes, forgetting can occur at multiple levels. We identify three types of class groups and analyze the mIoU results for each of them:", + "bbox": [ + 212, + 525, + 787, + 676 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Unsplit Classes: These classes are learned in the first task and are not further split. Their performance is expected to be largely unaffected, as they have not been split nor learned incrementally.", + "- Split Classes: These classes have been split either from the background or a previously learned superclass. The mIoU results for these split classes will provide insight into the model's ability to learn new classes.", + "- Retained Classes: These are classes remaining in the parent class after it has been split partially to yield new classes. Analyzing the mIoU scores for retained classes sheds light on how well the model retains knowledge." + ], + "bbox": [ + 223, + 696, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. Muralidhara et al.", + "bbox": [ + 271, + 114, + 411, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/cd5075e0ec434b0ffc4b35acb4bc27934b859af1298b57cbe58f7e8ca0bb3012.jpg", + "table_caption": [ + "Table 1: Results for our two CLEO settings on Cityscapes [8] after learning all tasks." + ], + "table_footnote": [], + "table_body": "
MethodCS-Ex1CS-Ex2
UnsplitSplitRetainedAllUnsplitSplitRetainedAll
Fine-Tuning00.0000.0000.0000.0000.0014.4600.0008.68
Joint Training90.2661.1548.0455.6290.2661.1548.0455.62
MiB [4]90.2403.6607.9909.1190.2318.2511.3918.88
PLOP [10]87.5834.3613.0228.9186.0131.9115.5128.10
RCIL [42]89.8200.0007.9706.8889.0000.0007.7006.76
MoOn (Ours)90.2139.5037.0039.3190.2342.2835.7840.62
", + "bbox": [ + 246, + 171, + 754, + 297 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/389f3fe45d9b9b393b519fcf16ad3b36c663948fe88c49cf8f68e1ca0824d621.jpg", + "table_caption": [ + "Table 2: Task-wise results for CS-Ex2 after learning the final task." + ], + "table_footnote": [], + "table_body": "
MethodTask 0Task 1Task 2Task 3Task 4Task 5Task 6All
Fine-Tuning00.0000.0000.0000.0000.0000.0034.7008.68
Joint Training47.5097.5763.0930.2088.3164.3157.9355.62
MiB [4]19.8202.8109.2600.1974.0045.6215.5418.88
PLOP [10]22.3895.0841.5900.0084.8153.8413.2128.10
RCIL [42]16.9000.0000.0000.0000.0000.0000.0006.76
MoOn (Ours)38.1196.8956.2609.1988.1660.9726.1040.62
", + "bbox": [ + 232, + 339, + 769, + 452 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Experimental Settings and Results", + "text_level": 1, + "bbox": [ + 215, + 481, + 547, + 496 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We introduce seven initial experimental settings, that are designed for the previously introduced datasets and resemble the complexity and diversity of CLEO beyond classical CIL. For completeness and better understanding, we present the detailed semantic evolution of all settings by listing $C_t$ for every task in the supplementary, including the class groups as introduced in Sec. 4.3.", + "bbox": [ + 212, + 508, + 787, + 585 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 609, + 375, + 627 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Cityscapes From the Cityscapes class hierarchy, we derive two experimental settings. Both settings start in task 0 by learning the seven parent classes which together consist of the 19 grouped classes. The settings involve the incremental splitting of individual classes from parent classes. We do not always completely split a parent class, thus emphasizing not only learning the split class but also retaining the remaining information on the superclass.", + "bbox": [ + 212, + 643, + 787, + 736 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- CS-Ex1: Each step involves splitting out a single class from each parent class. Since there is an unequal number of classes in the parent classes, we repeat this process until we have 19 individual classes, resulting in six tasks.", + "- CS-Ex2: At each step, we split all but one of the classes from a single parent class to form individual classes. This results in 19 classes in six evolutionary increments (seven tasks in total) for six parent classes excluding sky." + ], + "bbox": [ + 223, + 750, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "CLEO: Continual Learning of Evolving Ontologies", + "bbox": [ + 395, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/b3e401994a93a7c47cb67daafa8e7a74827b0b3404a7ceb5ea7b5014362f9ca3.jpg", + "table_caption": [ + "Table 3: Results for our three CLEO settings on PASCAL [11] after learning all tasks." + ], + "table_footnote": [], + "table_body": "
MethodVOC-Ex1VOC-Ex2VOC-Ex3
UnsplitSplitRetainedAllUnsplitSplitRetainedAllUnsplitSplitRetainedAll
Fine-Tuning37.7736.3100.0029.6837.7405.3800.0007.8136.9125.6800.0023.08
Joint Training90.1972.5384.1276.9190.4873.7889.1477.7589.8475.2787.3678.38
MiB [4]88.9137.2854.9947.0586.9911.2837.8722.8487.0135.7533.2840.28
PLOP [10]86.8538.6743.5645.6285.7716.1803.6721.2683.4136.7406.6136.88
RCIL [42]88.0800.8441.1119.2985.2901.4323.8513.1887.4501.9918.6312.51
MoOn (Ours)88.9555.9868.8462.5284.4522.0825.4228.8283.3347.5142.3650.19
", + "bbox": [ + 215, + 171, + 785, + 273 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The results from these settings are presented in Tab. 1 and the task-wise results of $CS - Ex2$ in Tab. 2. We can observe the shortcomings of existing CIL methods in handling evolving ontologies, whereas our baseline solution, MoOn, outperforms these methods. All methods achieve similar results on the unsplit classes, which is sky in both settings. However, when learning split classes, only PLOP comes close to our approach. Notably, MoOn surpasses PLOP by a significant margin in preserving information within the existing class. In the second task setting, the last step involves learning five vehicle classes, which are well-represented in the dataset. FT and MIB methods benefit from this, whereas PLOP and RCIL achieve similar results across both settings. We provide task-wise results of $CS - Ex1$ and visualizations for MoOn in the supplementary.", + "bbox": [ + 212, + 304, + 787, + 470 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "PASCAL VOC From the PASCAL class hierarchy, we derive three experimental settings that represent the challenges of CLEO. In the first step (task 0), the model learns all the 20 classes grouped into the four parent classes namely animals, household, person, and vehicle. The person class does not contain any subclasses and therefore does not undergo splitting.", + "bbox": [ + 212, + 494, + 787, + 571 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- VOC-Ex1: This setting includes splitting into intermediate sub-groups. Task 1 includes learning the sub-groups Farmyard, furniture, and 2-wheeler which in turn contain individual classes. In the next and final step, these sub-groups are split into their individual classes.", + "- VOC-Ex2: One class is split sequentially from each parent class resulting in five increments. At the end of the learning, we retain the individual classes cat, bottle, and the subgroup 4-wheeler in their respective parent classes.", + "- VOC-Ex3: This setting involves splitting all but one of the classes from each parent class sequentially, resulting in four tasks, with three evolutionary steps for the three parent classes excluding person which remains unsplit." + ], + "bbox": [ + 225, + 583, + 785, + 736 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The results on the three settings are shown in Tab. 3, while the corresponding visualizations for VOC-Ex3 are presented in Fig. 4. The performance on the unsplit classes, person class in this dataset, remains consistent across the CL methods. We observe the biggest difference in performance between our approach and others for VOC-Ex1. This is likely due to the more challenging multiple splitting of classes into intermediate subclasses. The lowest results for all methods", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. Muralidhara et al.", + "bbox": [ + 271, + 114, + 411, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5aa9443b3e7ff40950ea72b79650161cff2c77f6119fb371201679b544c4e85d.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparison on PASCAL VOC [11] VOC-Ex3 after the final task." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 784, + 292 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/e5611b330ac910f39715e94885e1417843df583ab01355c7dca6aaa2e24e3aed.jpg", + "table_caption": [ + "Table 4: Results for CLEO settings on Mapillary Vistas [30] after learning all tasks." + ], + "table_footnote": [], + "table_body": "
MethodMV-Ex1MV-Ex2
UnsplitSplitRetainedAllUnsplitSplitRetainedAll
Fine-Tuning00.0018.5900.0008.6900.0001.4700.0000.69
Joint Training28.3722.6825.7925.5128.3722.6825.7925.51
MiB [4]27.0216.3317.7421.2709.8304.3100.0706.45
PLOP [10]27.0713.1418.4419.8603.5901.2600.0002.22
RCIL [42]28.5401.5712.7814.6506.3900.5500.1703.16
MoOn (Ours)26.7216.6719.4721.4409.8104.4000.0106.49
", + "bbox": [ + 246, + 364, + 754, + 489 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "are for VOC-Ex2, which involves a longer sequence of tasks, and learning sets of mixed classes one from each parent class at each step.", + "bbox": [ + 212, + 520, + 787, + 551 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Mapillary Vistas Using the two dataset versions of Mapillary Vistas [30], we have defined two CLEO experiments. Both settings begin by learning the original 66 classes in task 0. The newly introduced classes are grouped into their ten parent classes from the previous version resulting in the original classes. The class mapping between the new classes and their parent classes is provided in the supplementary. The subsequent steps involve learning these new classes.", + "bbox": [ + 212, + 571, + 787, + 664 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- MV-Ex1: After learning the original classes in task 0, all of the new classes are learned in task 1, resulting in two tasks. This exemplifies a real-world application of CLEO, where the model is specifically trained to learn the new classes without a complete retraining of all the classes.", + "- MV-Ex2: This setting is more challenging with a longer sequence of tasks, which can exacerbate forgetting. In each step, one of the ten parent classes is split into its new classes, resulting in a total of eleven tasks." + ], + "bbox": [ + 225, + 676, + 787, + 781 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Mapillary Vistas is challenging due to the large number of classes. As shown in Tab. 4, all methods exhibit suboptimal performance, including joint training which does not suffer from forgetting. The two settings are similar and have", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "CLEO: Continual Learning of Evolving Ontologies", + "bbox": [ + 395, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "learned the same classes after the final task. They differ in the number of steps, highlighting the effect of a longer sequence of tasks. $MV - Ex2$ with a longer sequence significantly impacts the performance of all methods.", + "bbox": [ + 215, + 146, + 787, + 191 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Limitations and Future Work", + "text_level": 1, + "bbox": [ + 215, + 214, + 535, + 231 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In its most general form, CLEO allows for merging of two or more known classes as a valid form of evolution. However, we note that the here presented formalism of CLEO only considers the separation of known parent classes. Yet, in some rare cases, it might be required to combine previous information to simplify previous tasks and free representational capacity of a model. This, however, as a version of intended forgetting opens up a mostly unexplored field of research in continual learning and is left for future work. Besides, semantic context can be achieved through post-processing by combining predictions of existing and new classes into a single class, preventing interference and avoiding further forgetting.", + "bbox": [ + 215, + 246, + 785, + 382 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We further highlight that our baseline solution MoOn requires the information of class relations $S_{t}$ about the evolution of classes. While we believe that this information is usually available to a sequentially trained model, there might be cases in which explicit knowledge about the ontology is unavailable. In such cases, $S_{t}$ could be inferred from the existing model's predictions on the current task data. If the model predicts an existing class, the class being learned was previously seen and is split from this parent class. If it predicts the background, the new class is entirely new and unseen. In our experimental evaluation, we found that MoOn and other approaches notably underperform with longer sequences in CLEO, highlighting the need for future attention to long-term learning settings.", + "bbox": [ + 215, + 383, + 785, + 534 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 556, + 357, + 571 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduce a novel framework for addressing the challenge of adapting to evolving classes, termed Continual Learning of Evolving Ontologies (CLEO). The motivation for CLEO stems from a practical scenario wherein pre-existing classes undergo transformations over time, resulting in more specialized classes. CLEO extends the functionality of CIL to not just learn new information but also refine existing information. The applicability of CLEO in the real world is exemplified through the Mapillary Vistas dataset, where an existing set of classes evolves into more detailed classes. Furthermore, we highlight the shortcomings of existing CIL methods and propose MoOn as an initial solution for CLEO. The effectiveness of MoOn for CLEO is demonstrated on seven experimental settings across three datasets, surpassing previous CIL approaches.", + "bbox": [ + 215, + 588, + 785, + 753 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 215, + 777, + 392, + 795 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was partially funded by the Federal Ministry of Education and Research Germany under the project DECODE (01IW21001).", + "bbox": [ + 215, + 809, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. Muralidhara et al.", + "bbox": [ + 271, + 114, + 411, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aljundi, R., Chakravarty, P., Tuytelaars, T.: Expert gate: Lifelong learning with a network of experts. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2017)", + "2. Asghar, N., Mou, L., Selby, K.A., Pantasdo, K.D., Poupart, P., Jiang, X.: Progressive memory banks for incremental domain adaptation. arXiv (2018)", + "3. Bang, J., Kim, H., Yoo, Y., Ha, J.W., Choi, J.: Rainbow memory: Continual learning with a memory of diverse samples. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "4. Cermelli, F., Mancini, M., Bulo, S.R., Ricci, E., Caputo, B.: Modeling the background for incremental learning in semantic segmentation. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2020)", + "5. Cha, S., kim, b., Yoo, Y., Moon, T.: Ssul: Semantic segmentation with unknown label for exemplar-based class-incremental learning. In: Advances in Neural Information Processing Systems (2021)", + "6. Chaudhry, A., Rohrbach, M., Elhoseiny, M., Ajanthan, T., Dokania, P.K., Torr, P.H., Ranzato, M.: On tiny episodic memories in continual learning. arXiv (2019)", + "7. Chen, L.C., Papandreou, G., Schroff, F., Adam, H.: Rethinking atrous convolution for semantic image segmentation. arXiv (2017)", + "8. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016)", + "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2009)", + "10. Douillard, A., Chen, Y., Dapogny, A., Cord, M.: Plop: Learning without forgetting for continual semantic segmentation. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "11. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International Journal of Computer Vision (IJCV) (2010)", + "12. Fernando, C., Banarse, D., Blundell, C., Zwols, Y., Ha, D., Rusu, A.A., Pritzel, A., Wierstra, D.: Pathnet: Evolution channels gradient descent in super neural networks. arXiv (2017)", + "13. Goswami, D., Schuster, R., van de Weijer, J., Stricker, D.: Attribution-aware weight transfer: A warm-start initialization for class-incremental semantic segmentation In: Winter Conference on Applications of Computer Vision (2023)", + "14. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016)", + "15. Kemker, R., Kanan, C.: Fearnet: Brain-inspired model for incremental learning. In: International Conference on Learning Representations (2018)", + "16. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences (2017)", + "17. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + ], + "bbox": [ + 218, + 179, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "CLEO: Continual Learning of Evolving Ontologies", + "bbox": [ + 395, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Lee, B.H., Jung, O., Choi, J., Chun, S.Y.: Online continual learning on hierarchical label expansion. In: International Conference on Computer Vision (ICCV) (2023)", + "19. Lesort, T., Lomonaco, V., Stoian, A., Maltoni, D., Filliat, D., Diaz-Rodriguez, N.: Continual learning for robotics: Definition, framework, learning strategies, opportunities and challenges. Information fusion (2020)", + "20. Li, Z., Hoiem, D.: Learning without forgetting. Transactions on Pattern Analysis and Machine Intelligence (T-PAMI) (2017)", + "21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dóllar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision (ECCV) (2014)", + "22. Lin, Z., Pathak, D., Wang, Y.X., Ramanan, D., Kong, S.: Continual learning with evolving class ontologies. In: Conference on Neural Information Processing Systems (NeurIPS) (2022)", + "23. Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. In: Conference on Neural Information Processing Systems (NeurIPS) (2017)", + "24. Mallya, A., Davis, D., Lazebnik, S.: Piggyback: Adapting a single network to multiple tasks by learning to mask weights. In: European Conference on Computer Vision (ECCV) (2018)", + "25. Mallya, A., Lazebnik, S.: Packnet: Adding multiple tasks to a single network by iterative pruning. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2018)", + "26. McCloskey, M., Cohen, N.J.: Catastrophic interference in connectionist networks: The sequential learning problem. Psychology of Learning and Motivation (1989)", + "27. Mermillod, M., Bugaiska, A., Bonin, P.: The stability-plasticity dilemma: Investigating the continuum from catastrophic forgetting to age-limited learning effects (2013)", + "28. Michieli, U., Zanuttigh, P.: Incremental learning techniques for semantic segmentation. In: Conference on Computer Vision and Pattern Recognition Workshops (CVPR-W) (2019)", + "29. Michieli, U., Zanuttigh, P.: Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "30. Neuhold, G., Ollmann, T., Rota Bulo, S., Kontschieder, P.: The mapillary vistas dataset for semantic understanding of street scenes. In: International Conference on Computer Vision (ICCV) (2017)", + "31. Phan, M.H., Ta, T.A., Phung, S.L., Tran-Thanh, L., Bouzerdoum, A.: Class similarity weighted knowledge distillation for continual semantic segmentation. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "32. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv (2016)", + "33. Shin, H., Lee, J.K., Kim, J., Kim, J.: Continual learning with deep generative replay. In: Conference on Neural Information Processing Systems (NeurIPS) (2017)", + "34. Sprechmann, P., Jayakumar, S.M., Rae, J.W., Pritzel, A., Badia, A.P., Uria, B., Vinyals, O., Hassabis, D., Pascanu, R., Blundell, C.: Memory-based parameter adaptation. In: International Conference on Learning Representations (ICLR) (2018)", + "35. Van de Ven, G.M., Siegelmann, H.T., Tolias, A.S.: Brain-inspired replay for continual learning with artificial neural networks. Nature communications (2020)", + "36. Xiao, J.W., Zhang, C.B., Feng, J., Liu, X., van de Weijer, J., Cheng, M.M.: Endpoints weight fusion for class incremental semantic segmentation. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. Muralidhara et al.", + "bbox": [ + 271, + 114, + 410, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Yang, G., Fini, E., Xu, D., Rota, P., Ding, M., Nabi, M., Alameda-Pineda, X., Ricci, E.: Uncertainty-aware contrastive distillation for incremental semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)", + "38. Yang, G., Fini, E., Xu, D., Rota, P., Ding, M., Tang, H., Alameda-Pineda, X., Ricci, E.: Continual attentive fusion for incremental learning in semantic segmentation. IEEE Transactions on Multimedia (2022)", + "39. Yoon, J., Yang, E., Lee, J., Hwang, S.J.: Lifelong learning with dynamically expandable networks. arXiv (2017)", + "40. Yu, L., Liu, X., Van de Weijer, J.: Self-training for class-incremental semantic segmentation. IEEE Transactions on Neural Networks and Learning Systems (2022)", + "41. Zenke, F., Poole, B., Ganguli, S.: Continual learning through synaptic intelligence. In: International Conference on Machine Learning (ICML) (2017)", + "42. Zhang, C.B., Xiao, J.W., Liu, X., Chen, Y.C., Cheng, M.M.: Representation compensation networks for continual semantic segmentation. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "43. Zhao, H., Hu, Q., Zhu, P., Wang, Y., Wang, P.: A recursive regularization based feature selection framework for hierarchical classification. Transactions on Knowledge and Data Engineering (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 397 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "CLEO: Continual Learning of Evolving Ontologies", + "bbox": [ + 395, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/CLEO_ Continual Learning of Evolving Ontologies/7399fe90-4063-4ee4-8949-4afe680e9068_model.json b/2024/CLEO_ Continual Learning of Evolving Ontologies/7399fe90-4063-4ee4-8949-4afe680e9068_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d2174900353df18aff0f1572cef12f7a979c7977 --- /dev/null +++ b/2024/CLEO_ Continual Learning of Evolving Ontologies/7399fe90-4063-4ee4-8949-4afe680e9068_model.json @@ -0,0 +1,2236 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.217, + 0.141, + 0.787, + 0.166 + ], + "angle": 0, + "content": "CLEO: Continual Learning of Evolving Ontologies" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.189, + 0.79, + 0.22 + ], + "angle": 0, + "content": "Shishir Muralidhara1, Saqib Bukhari2, Georg Schneider2, Didier Stricker1,3, and René Schuster1,3" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.232, + 0.703, + 0.246 + ], + "angle": 0, + "content": "\\(^{1}\\) DFKI - German Research Center for Artificial Intelligence" + }, + { + "type": "text", + "bbox": [ + 0.418, + 0.247, + 0.586, + 0.26 + ], + "angle": 0, + "content": "2 ZF Friedrichshafen AG" + }, + { + "type": "text", + "bbox": [ + 0.343, + 0.26, + 0.661, + 0.274 + ], + "angle": 0, + "content": "\\(^{3}\\) RPTU - University of Kaiserslautern-Landau" + }, + { + "type": "text", + "bbox": [ + 0.355, + 0.275, + 0.648, + 0.288 + ], + "angle": 0, + "content": "{firstname这个名字}@{dfki.de，zf.com}" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.327, + 0.744, + 0.604 + ], + "angle": 0, + "content": "Abstract. Continual learning (CL) addresses the problem of catastrophic forgetting in neural networks, which occurs when a trained model tends to overwrite previously learned information, when presented with a new task. CL aims to instill the lifelong learning characteristic of humans in intelligent systems, making them capable of learning continuously while retaining what was already learned. Current CL problems involve either learning new domains (domain-incremental) or new and previously unseen classes (class-incremental). However, general learning processes are not just limited to learning information, but also refinement of existing information. In this paper, we define CLEO - Continual Learning of Evolving Ontologies, as a new incremental learning setting under CL to tackle evolving classes. CLEO is motivated by the need for intelligent systems to adapt to real-world ontologies that change over time, such as those in autonomous driving. We use Cityscapes, PASCAL VOC, and Mapillary Vistas to define the task settings and demonstrate the applicability of CLEO. We highlight the shortcomings of existing CIL methods in adapting to CLEO and propose a baseline solution, called Modelling Ontologies (MoOn). CLEO is a promising new approach to CL that addresses the challenge of evolving ontologies in real-world applications. MoOn surpasses previous CL approaches in the context of CLEO." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.618, + 0.741, + 0.646 + ], + "angle": 0, + "content": "Keywords: Continual Learning \\(\\cdot\\) Continual Semantic Segmentation \\(\\cdot\\) Class-Incremental Learning" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.672, + 0.377, + 0.687 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Advancements in AI have led to the emergence of systems that can match or even outperform human abilities in various domains. However, these systems are often confined to specific datasets and tasks and their efficacy is affected when faced with new tasks or data. This presents a significant hurdle in practical real-world AI applications where the circumstances are constantly evolving. For example, a self-driving car needs to adapt to new traffic regulations, signage, and weather conditions. When a trained model encounters new information or tasks, it experiences a phenomenon known as catastrophic forgetting [26]. This issue arises from the rigidity of neural networks, where previously learned knowledge is" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.412, + 0.128 + ], + "angle": 0, + "content": "S. Muralidhara et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.302 + ], + "angle": 0, + "content": "Fig. 1: CLEO extends class-incremental learning (CIL) by removing the constraint that incremental sets of classes must be non-overlapping. This allows for more flexible and nuanced learning, as existing classes can be refined into more specific classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.784, + 0.401 + ], + "angle": 0, + "content": "overwritten when learning something new. Consequently, the model's performance significantly deteriorates, which can have severe repercussions in scenarios where the model has to acclimate to dynamic environments or critical applications where performance is imperative. Therefore, tackling forgetting is crucial when developing AI systems with lifelong learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.789, + 0.63 + ], + "angle": 0, + "content": "In contrast, learning in humans is continuous and dynamic through which we continually assimilate new information while preserving and enriching previously learned knowledge. The goal of CL is to instill neural networks with this learning characteristic and make them capable of extending when objectives change. In incremental learning, which involves learning a sequence of tasks, the term task can represent changes either in the distribution of data, referred to as domain-incremental learning (DIL), or changes in the output space, which is known as class-incremental learning (CIL). However, learning is not limited only to the acquisition of new knowledge; it also involves refining and deepening our grasp of existing information. To illustrate, our initial understanding of animals might have been confined to recognizing a broad categorization, like distinguishing between birds and other animals. However, over time, our learning process matures to incorporate finer, more nuanced insights about individual animals. This highlights the necessity of broadening the scope of CL to accommodate both the introduction of novel classes, and also the enhancement of existing knowledge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.79, + 0.841 + ], + "angle": 0, + "content": "We propose a novel incremental setting termed CLEO - Continual Learning of Evolving Ontologies. CLEO represents a progressive extension of CIL, overcoming the confining assumption that all newly encountered classes are entirely novel and previously unseen. CLEO acknowledges the dynamic nature of learning, where classes may not only emerge but also evolve over time. CLEO combines elements of novelty and refinement, paving the way for more resilient AI systems. The real-world relevance of CLEO is exemplified by the Mapillary Vistas dataset [30], where an initial set of classes has evolved into more fine-grained and specific classes. An illustration of CLEO is presented in Fig. 1 using the CIFAR-10 dataset [17] as an example. In the initial task, the system learns two broad categories, namely animals and vehicles. Subsequently, more specific classes, such as bird, frog, plane, and ship, are split from the original classes. CLEO presents a challenge where the system must not only learn the evolved" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.396, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "CLEO: Continual Learning of Evolving Ontologies" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "classes, which it had previously encountered within a different semantic context, but also retain the remaining information. In this example, after bird and frog were separated from the animal class, cat and dog are left as the remaining domestic animal class. As in CIL, CLEO also covers the case that entirely new classes are learned, like e.g. the classes horse and deer in the last task of Fig. 1. CLEO embodies the fundamental idea that AI systems must possess the ability to adapt to dynamic shifts in which classes emerge, evolve, and split over time. By integrating this adaptability into AI systems, we enhance their robustness in real-world applications. This paper makes several significant contributions:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.293, + 0.785, + 0.337 + ], + "angle": 0, + "content": "- Formalizing Continual Learning of Evolving Ontologies: A primary contribution is the introduction and formalization of the concept of learning evolving ontologies, from a continual learning perspective." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.338, + 0.785, + 0.397 + ], + "angle": 0, + "content": "- Defining Task Settings: To provide a comprehensive understanding of CLEO, we define several task settings for semantic segmentation, using three diverse datasets: Cityscapes [8], PASCAL VOC [11], and Mapillary Vistas [30]." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.399, + 0.785, + 0.441 + ], + "angle": 0, + "content": "- Evaluation Metrics for CLEO: By identifying class groups and measuring their performance, we define a more nuanced evaluation of how effectively the system learns refined classes and retains information." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.444, + 0.785, + 0.487 + ], + "angle": 0, + "content": "- Baseline Solution - MoOn: To operationalize the concept for semantic segmentation, we propose a baseline solution called \"Modelling Ontologies\" (MoOn), to handle the challenges inherent in CLEO." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.293, + 0.785, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.512, + 0.388, + 0.527 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.544, + 0.422, + 0.559 + ], + "angle": 0, + "content": "2.1 Continual Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Continual learning (CL), also known as lifelong learning, and its sub-field of incremental learning, are critical areas of machine learning that address a fundamental limitation of traditional machine learning models: Their rigidity in learning new tasks. CL is a learning paradigm that aims to train models that can learn new tasks and information without forgetting what was previously learned. This contrasts traditional machine learning, where models are typically trained on a single dataset and then remain fixed. Continual learning presents a challenge of balancing between learning new information and retaining the old [27]. Overemphasizing learning new information leads to forgetting the old while focusing too much on retaining old information hinders learning the new task. CL approaches can be grouped into three categories: Architecture-based, replay-based methods, and regularization-based methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Architecture-based approaches modify the architecture of neural networks to accommodate new tasks without overwriting previous information. These include methods such as parameter isolation [12, 24, 25], dynamically growing networks [1, 32, 39], or incorporation of additional components such as memory modules [2, 34]. These methods can face scalability challenges as network architecture grows, increasing memory and computational demands." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.412, + 0.128 + ], + "angle": 0, + "content": "S. Muralidhara et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "Replay-based approaches are of two types: Experience replay or rehearsal, involves storing a few instances from the previous task [3,6,23], which are used during training on a new task. Generative replay substitutes the memory buffer with a generative model to sample previous task instances when learning on a new task [15,33,35]. Challenges include privacy, data availability, memory and computational constraints in training generative models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.79, + 0.329 + ], + "angle": 0, + "content": "Regularization-based approaches [16, 20, 41] strike a balance between stability and plasticity when learning new tasks by introducing a regularization term into the loss function. One of the approaches to regularization is to determine the importance of parameters to the previous tasks and penalize overwriting them. Distillation is another approach which preserves previously learned knowledge by distilling information from the old model to the new." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.349, + 0.542, + 0.365 + ], + "angle": 0, + "content": "2.2 Continual Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.79, + 0.751 + ], + "angle": 0, + "content": "Continual semantic segmentation (CSS) [5,37,38,40] is a specialized domain within CL that focuses on the task of semantic segmentation. Incremental Learning Techniques [28] (ILT) is a regularization-based method that distils knowledge from previously trained task models and additionally freezes the encoder for the current task, which acts as an additional regularization constraint. Modeling the Background [4] (MiB) addresses the problem of background shift in incremental segmentation. It proposes a novel distillation loss that accounts for the changing background between tasks. AWT [13] addresses the background shift in classifier initialization, and transfers the most relevant weights from the previous classifiers to the new classifier. In our baseline solution for CLEO, we will build on the ideas of MiB and extend it to model the semantic shift for all classes. PLOP [10] argues that treating the background class as a single class as done in MiB, which includes both previously encountered and novel classes can lead to catastrophic forgetting of the old classes as it learns new classes. PLOP addresses this problem by pseudolabeling old classes in the background. EWF [36] combines the weights of the old and new model trained with regularization to give a more balanced model. SDR [29] preserves latent representations of old classes through prototype matching and frees space for new classes through feature sparsity and contrastive learning. The decoder preserves previous knowledge via output-level distillation. RCIL [42] uses a representation compensation module to decouple the representation learning of old and new classes. This helps to prevent the model from forgetting old classes when learning new classes. REMINDER [31] proposes a class-similarity weighted distillation that distills knowledge from a previous model on old classes that are similar to the new classes. This reduces forgetting of old classes and improves learning of new classes." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.771, + 0.427, + 0.787 + ], + "angle": 0, + "content": "2.3 Evolving Ontologies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.842 + ], + "angle": 0, + "content": "An emerging area within incremental learning focuses on learning evolving classes, which involves the refinement of existing knowledge, where a broad initial grouping of classes evolves over time into more specific and nuanced classes. In contrast," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.396, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "CLEO: Continual Learning of Evolving Ontologies" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.726, + 0.162 + ], + "angle": 0, + "content": "in CIL, each task consists of a non-overlapping subsets of novel classes." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.177, + 0.787, + 0.253 + ], + "angle": 0, + "content": "Hierarchical Label Expansion [18] (HLE) for classification involves a model initially learning coarse classes, which then evolve into more fine-grained categories during each task across varying hierarchical levels. The expansion is restrictive, where the entire subset of classes at the current level expands in the next level during each task. CLEO differs from HLE in the following aspects:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.263, + 0.786, + 0.306 + ], + "angle": 0, + "content": "- CLEO allows a more flexible evolution at any point within the sequence of tasks, in which certain classes can remain (partially) unexpanded or expand into sub-groups." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.309, + 0.786, + 0.365 + ], + "angle": 0, + "content": "- The baseline solution for HLE is a rehearsal-based method that uses pseudolabeling from previous classifiers - one for each level of the hierarchy - for determining the importance of data. We present a regularization-based solution that uses knowledge distillation." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.369, + 0.785, + 0.397 + ], + "angle": 0, + "content": "- HLE considers image classification as application, while CLEO additionally deals with the more challenging problem of semantic segmentation." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.263, + 0.786, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.409, + 0.787, + 0.499 + ], + "angle": 0, + "content": "Learning with Evolving Class Ontologies [22] (LECO) has been the first attempt to address evolving ontologies in semantic segmentation. However, despite its intended framing as a continual learning problem, LECO deviates from the defining principles of continual learning. We address these shortcomings and define this problem strictly within the bounds of continual learning. The primary distinctions between our framework and LECO are outlined below:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.513, + 0.784, + 0.54 + ], + "angle": 0, + "content": "- LECO requires and stores all historical data, a constraint that contradicts the definition of CL [19], that all data is not available at once." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.543, + 0.784, + 0.585 + ], + "angle": 0, + "content": "- LECO primarily revolves around the dilemma of whether to label new data or relabel the old data with the current label space. In contrast, CLEO confines itself to utilizing the input data and annotations within the dataset." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.589, + 0.784, + 0.631 + ], + "angle": 0, + "content": "- Due to the availability of historical data, LECO employs naive strategies such as fine-tuning, freezing, or joint training. We highlight the limitations of existing CIL methods and propose a baseline specifically for CLEO." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.634, + 0.784, + 0.676 + ], + "angle": 0, + "content": "- CLEO explicitly addresses catastrophic forgetting, which is one of the core problems of continual learning. Since LECO retrans entirely using all the data, the problem of catastrophic forgetting does not even arise." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.513, + 0.784, + 0.676 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.703, + 0.314, + 0.718 + ], + "angle": 0, + "content": "3 CLEO" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Deep learning has drawn inspiration from biological systems at various levels, from the way neurons form the fundamental building blocks of artificial neural networks to learning paradigms such as curriculum learning and reinforcement learning. Continual learning is another learning paradigm that aims to instill lifelong learning in intelligent systems. As discussed previously, current problems in continual learning involve adaptation to different domains, such as handling adverse conditions in autonomous driving or learning new classes over time. A" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.412, + 0.128 + ], + "angle": 0, + "content": "S. Muralidhara et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "common characteristic in both these scenarios is that the task always represents new and previously unseen data. However, it is our contention that the scope of continual learning should not be limited solely to the discovery of novel classes. It should also include the enhancement and refinement of existing knowledge. This aligns closely with the way humans continually adapt and refine knowledge through neuroplasticity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.79, + 0.495 + ], + "angle": 0, + "content": "In light of this perspective, we introduce CLEO - Continual Learning of Evolving Ontologies. CLEO is centered around the idea of removing the constraints of class incremental learning and allows classes that have a hierarchical relation (partially) overlap between tasks, thereby facilitating the refinement of already learned information. CLEO is a real and open-world problem and this is exemplified in the Mapillary Vistas dataset [30]. The initial version of the dataset defined 66 classes, which later evolved into 124 classes in the latest version. It is noteworthy to observe the transformation within class categorization. On one hand, classes that were previously categorized under unlabeled, such as traffic island emerge as novel classes. Conversely, several pre-existing classes undergo refinement, leading to the creation of new, more specific categories. In Fig. 2, we can observe a tangible example of this evolution, where the previous billboard class has evolved into more specialized classes, including signage classes for advertisements, information, stores, and differentiating between the front and back of the signs. In critical applications like autonomous driving, it becomes imperative for AI systems to swiftly adapt to these nuanced changes, perpetually learn, and enhance scene understanding precision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.79, + 0.738 + ], + "angle": 0, + "content": "Let \\( T = \\{0,1,2,\\dots,n\\} \\) denote a sequence of \\( n + 1 \\) tasks. Each task is associated with a specific subset of classes denoted as \\( C_t \\). The union of all classes across tasks is \\( C = \\bigcup_{t=0}^{n} C_t \\), representing the entire set of classes encountered during the learning process. At each task \\( t \\), a subset of classes \\( C_t \\) is introduced incrementally similar to CIL. \\( C_{0:t} = C_0 \\cup \\ldots \\cup C_{t-1} \\cup C_t \\) is the set of all classes seen until step \\( t \\). CLEO acknowledges the possibility of class splitting, where an existing class \\( c_{t-1} \\) from \\( C_{0:t-1} \\) can evolve into multiple new classes in \\( C_t \\). Formally this can be expressed as, \\( C_t \\subset C \\), \\( C_i \\cap C_j = \\phi \\) for \\( i \\neq j \\) and the class splitting is expressed as \\( c_{t-1} \\in C_{0:t-1} \\to \\{c_{k+1}, c_{k+2}, \\ldots, c_{k+m}\\} \\subseteq C_t \\). Note that the splitting of a class is not necessarily exhaustive, i.e., \\( c_{t-1} \\in C_{0:t-1} \\) may persist. An example for such an evolution of classes is given in Fig. 1. The learning objective of a task remains consistent with CIL: Train a model to recognize and differentiate the current subset of classes \\( C_t \\), while ensuring that the model retains knowledge from the previously encountered classes in \\( C_{0:t-1} \\). As in CIL, the model is not provided access to data from previous tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.761, + 0.634, + 0.779 + ], + "angle": 0, + "content": "4 MoOn: Modelling Ontologies for CLEO" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Class-incremental learning involves the process of training a model to incrementally learn new classes. At each increment, the model is presented with images and the corresponding ground truth which contains annotations for the new classes" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.396, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "CLEO: Continual Learning of Evolving Ontologies" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.402, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.144, + 0.594, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.145, + 0.785, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.262, + 0.788, + 0.318 + ], + "angle": 0, + "content": "Fig. 2: The Mapillary Vistas dataset [30] provides a real-world example for an evolving ontology. This figure compares ground truth annotations for the same image in two dataset versions. It illustrates the introduction of new classes, such as traffic island and the finer categorization of existing classes such as billboard." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.789, + 0.511 + ], + "angle": 0, + "content": "being learned. In class-incremental semantic segmentation (CISS), at each step, all unlabeled areas are marked as background. Thus in CISS, the phenomenon of semantic background shift occurs, because in each task, the meaning of the background class \\((bg)\\) changes. Distillation loss is a technique of transferring knowledge from an old model \\(f_{t-1}\\) to a new model \\(f_t\\), and this is achieved by comparing the old model's prediction to the new model's predictions. The distillation loss enforces the new model's predictions to resemble those of the old model by penalizing any deviations and this helps mitigate forgetting. Consider an image \\(x\\) with \\(N\\) pixels and the corresponding ground truth \\(y\\) at the current task \\(t\\) with the class set \\(C_t\\). The standard distillation loss in [4] is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.511, + 0.787, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\ell_ {k d} ^ {t} (x, y) = - \\frac {1}{N} \\sum_ {i \\in x} \\sum_ {c \\in C _ {0: t - 1}} p _ {x} ^ {t - 1} (i, c) \\log \\hat {p} _ {x} ^ {t} (i, c) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.6 + ], + "angle": 0, + "content": "where \\( p_x^{t-1}(i, c) \\) is the probability for class \\( c \\) in pixel \\( i \\) given by the model \\( f_{t-1} \\) and \\( \\hat{p}_x^t(i, c) \\) is for class \\( c \\) in pixel \\( i \\) given by the model \\( f_t \\) but re-normalized across classes \\( C_{0:t-1} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.328, + 0.609, + 0.787, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\hat {p} _ {x} ^ {t} (i, c) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} c \\in C _ {t} \\backslash \\{b g \\} \\\\ \\frac {p _ {x} ^ {t} (i , c)}{\\sum_ {k \\in C _ {0 : t - 1}} p _ {x} ^ {t} (i , k)} & \\text {i f} c \\in C _ {0: t - 1} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Standard distillation ignores the dynamic nature of the background in incremental semantic segmentation. The background class in segmentation is a catch-all class that includes all previously seen classes and potential future classes. Furthermore, the background class is present in all tasks and changes as new classes are added, leading to background shift. A class \\( c \\in C_t \\) added during the current task was previously part of the background in \\( C_{t-1} \\). For a pixel representing a new class at the current task \\( t \\), the old model \\( f_{t-1} \\) would predict the background class, and if standard distillation loss was applied naively, it would penalize the new model \\( f_t \\) for predicting the correct class. Modeling the Background (MiB) [4] proposes a novel distillation loss that explicitly addresses the background shift in incremental segmentation. MiB does not directly compare the prediction of the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.412, + 0.128 + ], + "angle": 0, + "content": "S. Muralidhara et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.148, + 0.784, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.417 + ], + "angle": 0, + "content": "Fig. 3: Modelling Ontologies (MoOn) generalizes the idea of MiB [4] for all classes. During distillation, forgetting is mitigated, by grouping class-logits according to the evolution of the class ontology. In this example, the teacher network of the previous task predicts the vehicle class, while the model tries to learn bicycle. To avoid penalizing this, the predictions for vehicle and bicycle are combined before the distillation loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.553 + ], + "angle": 0, + "content": "new model \\( f_{t} \\) with the corresponding prediction of the old model \\( f_{t-1} \\). Instead, the background prediction by \\( f_{t-1} \\) is compared with the \\( f_{t} \\) prediction of either a new class or the background. Eq. (3) modifies Eq. (2) by additionally considering the whole label space instead of limiting it to \\( C_{0:t-1} \\). MiB acknowledges that pixels that were thought to be part of the background by the old model might actually be the new classes the current model is trying to learn, thus avoiding penalizing the probabilities of new classes being learned." + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.563, + 0.788, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\hat {p} _ {x} ^ {t} (i, c) = \\left\\{ \\begin{array}{l l} p _ {x} ^ {t} (i, c) & \\text {i f} c \\neq b g \\\\ \\sum_ {k \\in C _ {t}} p _ {x} ^ {t} (i, k) & \\text {i f} c = b g \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.766 + ], + "angle": 0, + "content": "MiB is designed for CIL, where the new classes being learned at the current step, were previously unseen and are always a part of the background class. Therefore, it would suffice to take into consideration the background class, and hence modeling the background. Whereas CLEO is designed for learning of evolving classes and the new classes being learned may have been previously unseen and part of the background or been part of a broader class. As seen in Fig. 3, the class bike being learned at the current step may have been previously learned as a vehicle class grouped with other classes like truck and bus. Therefore we need to take into consideration that within CLEO, new classes may not always have been a part of the background class previously." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To this end, we propose Modelling Ontologies (MoOn) as a baseline solution for CLEO, which can be seen as a generalized version of MiB that accounts for the fact that new classes may emerge from any existing class, and not just the background class. Similar to how MiB considered the possibility of the new class" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.396, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "CLEO: Continual Learning of Evolving Ontologies" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.359 + ], + "angle": 0, + "content": "pixel belonging to the background class previously, MoOn additionally includes the possibility of the pixel belonging to an existing class. In MiB, the old model \\( f_{t-1} \\) may predict the background class for the new class being learned, whereas the old model in MoOn may predict an existing class \\( c \\in C_{0:t-1} \\). Therefore, the correct distillation would be to distill the knowledge from the old model \\( f_{t-1} \\) for the existing class \\( c \\in C_{0:t-1} \\) that is being split into new classes in \\( C_t \\). In the illustrated scenario depicted in Fig. 3, distillation would involve transferring knowledge from the old model regarding the vehicle class, while training the new model to predict the bike class. To achieve this, at each step \\( t \\), MoOn uses a set \\( S_t \\) of class maps \\( (c_{t-1} \\to \\{c_{k+1}, c_{k+2}, \\dots, c_{k+m}\\} = C_{new}) \\) to identify the class \\( c_{t-1} \\in C_{0:t-1} \\) from which the current classes in \\( C_t \\) are split. If the new class \\( c_t \\) is completely new, then the superclass \\( c_{t-1} \\) in the class mapping would be the background class \\( bg \\). We define \\( C_{e,t} \\) as the set of all evolving classes \\( c_{t-1} \\) in \\( S_t \\). The modified distillation loss for learning evolving classes in CLEO is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.31, + 0.372, + 0.788, + 0.428 + ], + "angle": 0, + "content": "\\[\n\\hat {p} _ {x} ^ {t} (i, c) = \\left\\{ \\begin{array}{l l} p _ {x} ^ {t} (i, c) & \\text {i f} c \\notin C _ {e, t} \\\\ p _ {x} ^ {t} (i, c _ {t - 1}) + \\sum_ {c _ {n} \\in C _ {\\text {n e w}}} p _ {x} ^ {t} (i, c _ {n}) & \\text {i f} c \\in C _ {e, t} \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.449, + 0.335, + 0.463 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.787, + 0.506 + ], + "angle": 0, + "content": "Towards the introduction of CLEO for semantic segmentation, we consider the following three diverse datasets:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.519, + 0.788, + 0.609 + ], + "angle": 0, + "content": "- Cityscapes: Cityscapes [8] is a large-scale dataset of urban street scenes captured in 50 different cities under daytime and good weather conditions. The dataset provides annotations for 30 classes out of which 19 are used for evaluation and grouped into 7 categories. An overview of the hierarchy of the 19 semantic classes grouped by their parent categories is presented in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.61, + 0.788, + 0.685 + ], + "angle": 0, + "content": "- PASCAL VOC: PASCAL VOC 2012 [11] is part of the Visual Object Classes Challenge, which ran from 2005 to 2012. PASCAL VOC is a more general dataset like COCO [21] consisting of 20 categories, including vehicles, household objects, animals, and persons. The categorization by [43] of these classes along with intermediate groupings is shown in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.686, + 0.788, + 0.761 + ], + "angle": 0, + "content": "- Mapillary Vistas: Mapillary Vistas [30] is a large-scale street-level image dataset for understanding street scenes around the world. It covers a wider range of geographic locations and diverse weather and illumination conditions. The initial version of the dataset provided annotations for 66 classes, which later nearly doubled into 124 classes in the current version v2.0." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.519, + 0.788, + 0.761 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.784, + 0.512, + 0.8 + ], + "angle": 0, + "content": "4.2 Baselines and Implementation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We evaluate our method's performance against standard CL baselines, namely fine-tuning and joint training. Fine-tuning involves using the previous step model" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.412, + 0.128 + ], + "angle": 0, + "content": "S. Muralidhara et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "as the starting point when learning on the current task. This is similar to transfer learning and even results in positive forward transfer. However, since there is no explicit attempt to prevent catastrophic forgetting, it involves the highest amount of forgetting and forms the lower baseline. The joint training model involves learning all of the tasks together in a single step, and as there is no incremental learning there is zero forgetting and this forms the theoretical upper baseline. Since in CLEO a redefinition of classes may occur because of the evolving ontology, we define the set of classes for the joint training as the final set of classes after the last task. Additionally, we compare against the state-of-the-art (SOTA) methods such as MiB [4], PLOP [10] and RCIL [42]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.79, + 0.451 + ], + "angle": 0, + "content": "For MiB, and our proposed approach, we adopt the framework proposed by MiB [4]. In the case of PLOP and RCIL, which involve modifications specific to their approaches, we adopt the RCIL framework [42]. All these approaches use the DeepLabV3 [7] architecture as segmentation network, utilizing a ResNet-101 [14] backbone with an output stride of 16. The backbone's initialization uses a pre-trained model from ImageNet [9]. The training methodology utilizes SGD, with a learning rate of 0.01 for the initial learning step and 0.001 for subsequent steps. Across all learning steps, we train the model using a batch size of 24, for 50 epochs on Cityscapes and for 30 epochs on PASCAL and Mapillary Vistas." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.487, + 0.421, + 0.501 + ], + "angle": 0, + "content": "4.3 Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.789, + 0.678 + ], + "angle": 0, + "content": "Mean Intersection-over-Union (mIoU) is a commonly used metric for evaluating semantic segmentation models. It measures the average overlap between the predicted segmentation and the ground truth segmentation for all classes. To calculate the mIoU, the Intersection-over-Union (IoU) is first calculated for each class, and then the IoU scores for all classes are averaged. The performance of CL models are often evaluated by measuring the amount of information forgotten by comparing the model's performance on old tasks against the joint training baseline. Within the context of CLEO, while learning evolving classes, forgetting can occur at multiple levels. We identify three types of class groups and analyze the mIoU results for each of them:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.698, + 0.787, + 0.744 + ], + "angle": 0, + "content": "- Unsplit Classes: These classes are learned in the first task and are not further split. Their performance is expected to be largely unaffected, as they have not been split nor learned incrementally." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.747, + 0.787, + 0.792 + ], + "angle": 0, + "content": "- Split Classes: These classes have been split either from the background or a previously learned superclass. The mIoU results for these split classes will provide insight into the model's ability to learn new classes." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- Retained Classes: These are classes remaining in the parent class after it has been split partially to yield new classes. Analyzing the mIoU scores for retained classes sheds light on how well the model retains knowledge." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.698, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.396, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "CLEO: Continual Learning of Evolving Ontologies" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.159 + ], + "angle": 0, + "content": "Table 1: Results for our two CLEO settings on Cityscapes [8] after learning all tasks." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.172, + 0.756, + 0.298 + ], + "angle": 0, + "content": "
MethodCS-Ex1CS-Ex2
UnsplitSplitRetainedAllUnsplitSplitRetainedAll
Fine-Tuning00.0000.0000.0000.0000.0014.4600.0008.68
Joint Training90.2661.1548.0455.6290.2661.1548.0455.62
MiB [4]90.2403.6607.9909.1190.2318.2511.3918.88
PLOP [10]87.5834.3613.0228.9186.0131.9115.5128.10
RCIL [42]89.8200.0007.9706.8889.0000.0007.7006.76
MoOn (Ours)90.2139.5037.0039.3190.2342.2835.7840.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.275, + 0.313, + 0.725, + 0.327 + ], + "angle": 0, + "content": "Table 2: Task-wise results for CS-Ex2 after learning the final task." + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.34, + 0.771, + 0.453 + ], + "angle": 0, + "content": "
MethodTask 0Task 1Task 2Task 3Task 4Task 5Task 6All
Fine-Tuning00.0000.0000.0000.0000.0000.0034.7008.68
Joint Training47.5097.5763.0930.2088.3164.3157.9355.62
MiB [4]19.8202.8109.2600.1974.0045.6215.5418.88
PLOP [10]22.3895.0841.5900.0084.8153.8413.2128.10
RCIL [42]16.9000.0000.0000.0000.0000.0000.0006.76
MoOn (Ours)38.1196.8956.2609.1988.1660.9726.1040.62
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.482, + 0.549, + 0.497 + ], + "angle": 0, + "content": "4.4 Experimental Settings and Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.587 + ], + "angle": 0, + "content": "We introduce seven initial experimental settings, that are designed for the previously introduced datasets and resemble the complexity and diversity of CLEO beyond classical CIL. For completeness and better understanding, we present the detailed semantic evolution of all settings by listing \\(C_t\\) for every task in the supplementary, including the class groups as introduced in Sec. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.611, + 0.377, + 0.628 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.737 + ], + "angle": 0, + "content": "Cityscapes From the Cityscapes class hierarchy, we derive two experimental settings. Both settings start in task 0 by learning the seven parent classes which together consist of the 19 grouped classes. The settings involve the incremental splitting of individual classes from parent classes. We do not always completely split a parent class, thus emphasizing not only learning the split class but also retaining the remaining information on the superclass." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.75, + 0.786, + 0.794 + ], + "angle": 0, + "content": "- CS-Ex1: Each step involves splitting out a single class from each parent class. Since there is an unequal number of classes in the parent classes, we repeat this process until we have 19 individual classes, resulting in six tasks." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.796, + 0.786, + 0.84 + ], + "angle": 0, + "content": "- CS-Ex2: At each step, we split all but one of the classes from a single parent class to form individual classes. This results in 19 classes in six evolutionary increments (seven tasks in total) for six parent classes excluding sky." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.75, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.412, + 0.128 + ], + "angle": 0, + "content": "S. Muralidhara et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.16 + ], + "angle": 0, + "content": "Table 3: Results for our three CLEO settings on PASCAL [11] after learning all tasks." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.172, + 0.787, + 0.274 + ], + "angle": 0, + "content": "
MethodVOC-Ex1VOC-Ex2VOC-Ex3
UnsplitSplitRetainedAllUnsplitSplitRetainedAllUnsplitSplitRetainedAll
Fine-Tuning37.7736.3100.0029.6837.7405.3800.0007.8136.9125.6800.0023.08
Joint Training90.1972.5384.1276.9190.4873.7889.1477.7589.8475.2787.3678.38
MiB [4]88.9137.2854.9947.0586.9911.2837.8722.8487.0135.7533.2840.28
PLOP [10]86.8538.6743.5645.6285.7716.1803.6721.2683.4136.7406.6136.88
RCIL [42]88.0800.8441.1119.2985.2901.4323.8513.1887.4501.9918.6312.51
MoOn (Ours)88.9555.9868.8462.5284.4522.0825.4228.8283.3347.5142.3650.19
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.305, + 0.789, + 0.471 + ], + "angle": 0, + "content": "The results from these settings are presented in Tab. 1 and the task-wise results of \\( CS - Ex2 \\) in Tab. 2. We can observe the shortcomings of existing CIL methods in handling evolving ontologies, whereas our baseline solution, MoOn, outperforms these methods. All methods achieve similar results on the unsplit classes, which is sky in both settings. However, when learning split classes, only PLOP comes close to our approach. Notably, MoOn surpasses PLOP by a significant margin in preserving information within the existing class. In the second task setting, the last step involves learning five vehicle classes, which are well-represented in the dataset. FT and MIB methods benefit from this, whereas PLOP and RCIL achieve similar results across both settings. We provide task-wise results of \\( CS - Ex1 \\) and visualizations for MoOn in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.789, + 0.572 + ], + "angle": 0, + "content": "PASCAL VOC From the PASCAL class hierarchy, we derive three experimental settings that represent the challenges of CLEO. In the first step (task 0), the model learns all the 20 classes grouped into the four parent classes namely animals, household, person, and vehicle. The person class does not contain any subclasses and therefore does not undergo splitting." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.584, + 0.787, + 0.645 + ], + "angle": 0, + "content": "- VOC-Ex1: This setting includes splitting into intermediate sub-groups. Task 1 includes learning the sub-groups Farmyard, furniture, and 2-wheeler which in turn contain individual classes. In the next and final step, these sub-groups are split into their individual classes." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.645, + 0.787, + 0.691 + ], + "angle": 0, + "content": "- VOC-Ex2: One class is split sequentially from each parent class resulting in five increments. At the end of the learning, we retain the individual classes cat, bottle, and the subgroup 4-wheeler in their respective parent classes." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.691, + 0.787, + 0.737 + ], + "angle": 0, + "content": "- VOC-Ex3: This setting involves splitting all but one of the classes from each parent class sequentially, resulting in four tasks, with three evolutionary steps for the three parent classes excluding person which remains unsplit." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.584, + 0.787, + 0.737 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The results on the three settings are shown in Tab. 3, while the corresponding visualizations for VOC-Ex3 are presented in Fig. 4. The performance on the unsplit classes, person class in this dataset, remains consistent across the CL methods. We observe the biggest difference in performance between our approach and others for VOC-Ex1. This is likely due to the more challenging multiple splitting of classes into intermediate subclasses. The lowest results for all methods" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.396, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "CLEO: Continual Learning of Evolving Ontologies" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.785, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.309, + 0.782, + 0.325 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparison on PASCAL VOC [11] VOC-Ex3 after the final task." + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.338, + 0.782, + 0.353 + ], + "angle": 0, + "content": "Table 4: Results for CLEO settings on Mapillary Vistas [30] after learning all tasks." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.365, + 0.756, + 0.491 + ], + "angle": 0, + "content": "
MethodMV-Ex1MV-Ex2
UnsplitSplitRetainedAllUnsplitSplitRetainedAll
Fine-Tuning00.0018.5900.0008.6900.0001.4700.0000.69
Joint Training28.3722.6825.7925.5128.3722.6825.7925.51
MiB [4]27.0216.3317.7421.2709.8304.3100.0706.45
PLOP [10]27.0713.1418.4419.8603.5901.2600.0002.22
RCIL [42]28.5401.5712.7814.6506.3900.5500.1703.16
MoOn (Ours)26.7216.6719.4721.4409.8104.4000.0106.49
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.788, + 0.552 + ], + "angle": 0, + "content": "are for VOC-Ex2, which involves a longer sequence of tasks, and learning sets of mixed classes one from each parent class at each step." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.573, + 0.788, + 0.665 + ], + "angle": 0, + "content": "Mapillary Vistas Using the two dataset versions of Mapillary Vistas [30], we have defined two CLEO experiments. Both settings begin by learning the original 66 classes in task 0. The newly introduced classes are grouped into their ten parent classes from the previous version resulting in the original classes. The class mapping between the new classes and their parent classes is provided in the supplementary. The subsequent steps involve learning these new classes." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.677, + 0.788, + 0.737 + ], + "angle": 0, + "content": "- MV-Ex1: After learning the original classes in task 0, all of the new classes are learned in task 1, resulting in two tasks. This exemplifies a real-world application of CLEO, where the model is specifically trained to learn the new classes without a complete retraining of all the classes." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.738, + 0.788, + 0.782 + ], + "angle": 0, + "content": "- MV-Ex2: This setting is more challenging with a longer sequence of tasks, which can exacerbate forgetting. In each step, one of the ten parent classes is split into its new classes, resulting in a total of eleven tasks." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.677, + 0.788, + 0.782 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Mapillary Vistas is challenging due to the large number of classes. As shown in Tab. 4, all methods exhibit suboptimal performance, including joint training which does not suffer from forgetting. The two settings are similar and have" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.412, + 0.128 + ], + "angle": 0, + "content": "S. Muralidhara et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "learned the same classes after the final task. They differ in the number of steps, highlighting the effect of a longer sequence of tasks. \\(MV - Ex2\\) with a longer sequence significantly impacts the performance of all methods." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.215, + 0.536, + 0.232 + ], + "angle": 0, + "content": "6 Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.247, + 0.787, + 0.383 + ], + "angle": 0, + "content": "In its most general form, CLEO allows for merging of two or more known classes as a valid form of evolution. However, we note that the here presented formalism of CLEO only considers the separation of known parent classes. Yet, in some rare cases, it might be required to combine previous information to simplify previous tasks and free representational capacity of a model. This, however, as a version of intended forgetting opens up a mostly unexplored field of research in continual learning and is left for future work. Besides, semantic context can be achieved through post-processing by combining predictions of existing and new classes into a single class, preventing interference and avoiding further forgetting." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.535 + ], + "angle": 0, + "content": "We further highlight that our baseline solution MoOn requires the information of class relations \\( S_{t} \\) about the evolution of classes. While we believe that this information is usually available to a sequentially trained model, there might be cases in which explicit knowledge about the ontology is unavailable. In such cases, \\( S_{t} \\) could be inferred from the existing model's predictions on the current task data. If the model predicts an existing class, the class being learned was previously seen and is split from this parent class. If it predicts the background, the new class is entirely new and unseen. In our experimental evaluation, we found that MoOn and other approaches notably underperform with longer sequences in CLEO, highlighting the need for future attention to long-term learning settings." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.557, + 0.358, + 0.573 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.589, + 0.787, + 0.755 + ], + "angle": 0, + "content": "We introduce a novel framework for addressing the challenge of adapting to evolving classes, termed Continual Learning of Evolving Ontologies (CLEO). The motivation for CLEO stems from a practical scenario wherein pre-existing classes undergo transformations over time, resulting in more specialized classes. CLEO extends the functionality of CIL to not just learn new information but also refine existing information. The applicability of CLEO in the real world is exemplified through the Mapillary Vistas dataset, where an existing set of classes evolves into more detailed classes. Furthermore, we highlight the shortcomings of existing CIL methods and propose MoOn as an initial solution for CLEO. The effectiveness of MoOn for CLEO is demonstrated on seven experimental settings across three datasets, surpassing previous CIL approaches." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.778, + 0.393, + 0.796 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "This work was partially funded by the Federal Ministry of Education and Research Germany under the project DECODE (01IW21001)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.396, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "CLEO: Continual Learning of Evolving Ontologies" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.18, + 0.785, + 0.221 + ], + "angle": 0, + "content": "1. Aljundi, R., Chakravarty, P., Tuytelaars, T.: Expert gate: Lifelong learning with a network of experts. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.223, + 0.785, + 0.249 + ], + "angle": 0, + "content": "2. Asghar, N., Mou, L., Selby, K.A., Pantasdo, K.D., Poupart, P., Jiang, X.: Progressive memory banks for incremental domain adaptation. arXiv (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.251, + 0.785, + 0.292 + ], + "angle": 0, + "content": "3. Bang, J., Kim, H., Yoo, Y., Ha, J.W., Choi, J.: Rainbow memory: Continual learning with a memory of diverse samples. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.294, + 0.787, + 0.334 + ], + "angle": 0, + "content": "4. Cermelli, F., Mancini, M., Bulo, S.R., Ricci, E., Caputo, B.: Modeling the background for incremental learning in semantic segmentation. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.336, + 0.785, + 0.376 + ], + "angle": 0, + "content": "5. Cha, S., kim, b., Yoo, Y., Moon, T.: Ssul: Semantic segmentation with unknown label for exemplar-based class-incremental learning. In: Advances in Neural Information Processing Systems (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.787, + 0.404 + ], + "angle": 0, + "content": "6. Chaudhry, A., Rohrbach, M., Elhoseiny, M., Ajanthan, T., Dokania, P.K., Torr, P.H., Ranzato, M.: On tiny episodic memories in continual learning. arXiv (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.407, + 0.785, + 0.432 + ], + "angle": 0, + "content": "7. Chen, L.C., Papandreou, G., Schroff, F., Adam, H.: Rethinking atrous convolution for semantic image segmentation. arXiv (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.488 + ], + "angle": 0, + "content": "8. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.49, + 0.785, + 0.531 + ], + "angle": 0, + "content": "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.533, + 0.785, + 0.573 + ], + "angle": 0, + "content": "10. Douillard, A., Chen, Y., Dapogny, A., Cord, M.: Plop: Learning without forgetting for continual semantic segmentation. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.575, + 0.785, + 0.614 + ], + "angle": 0, + "content": "11. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International Journal of Computer Vision (IJCV) (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.616, + 0.785, + 0.657 + ], + "angle": 0, + "content": "12. Fernando, C., Banarse, D., Blundell, C., Zwols, Y., Ha, D., Rusu, A.A., Pritzel, A., Wierstra, D.: Pathnet: Evolution channels gradient descent in super neural networks. arXiv (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.66, + 0.785, + 0.699 + ], + "angle": 0, + "content": "13. Goswami, D., Schuster, R., van de Weijer, J., Stricker, D.: Attribution-aware weight transfer: A warm-start initialization for class-incremental semantic segmentation In: Winter Conference on Applications of Computer Vision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.701, + 0.785, + 0.728 + ], + "angle": 0, + "content": "14. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.73, + 0.785, + 0.755 + ], + "angle": 0, + "content": "15. Kemker, R., Kanan, C.: Fearnet: Brain-inspired model for incremental learning. In: International Conference on Learning Representations (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.785, + 0.811 + ], + "angle": 0, + "content": "16. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "17. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.18, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.411, + 0.127 + ], + "angle": 0, + "content": "S. Muralidhara et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.175 + ], + "angle": 0, + "content": "18. Lee, B.H., Jung, O., Choi, J., Chun, S.Y.: Online continual learning on hierarchical label expansion. In: International Conference on Computer Vision (ICCV) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.787, + 0.217 + ], + "angle": 0, + "content": "19. Lesort, T., Lomonaco, V., Stoian, A., Maltoni, D., Filliat, D., Diaz-Rodriguez, N.: Continual learning for robotics: Definition, framework, learning strategies, opportunities and challenges. Information fusion (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.787, + 0.243 + ], + "angle": 0, + "content": "20. Li, Z., Hoiem, D.: Learning without forgetting. Transactions on Pattern Analysis and Machine Intelligence (T-PAMI) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.787, + 0.284 + ], + "angle": 0, + "content": "21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dóllar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision (ECCV) (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.787, + 0.325 + ], + "angle": 0, + "content": "22. Lin, Z., Pathak, D., Wang, Y.X., Ramanan, D., Kong, S.: Continual learning with evolving class ontologies. In: Conference on Neural Information Processing Systems (NeurIPS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.787, + 0.352 + ], + "angle": 0, + "content": "23. Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. In: Conference on Neural Information Processing Systems (NeurIPS) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.787, + 0.393 + ], + "angle": 0, + "content": "24. Mallya, A., Davis, D., Lazebnik, S.: Piggyback: Adapting a single network to multiple tasks by learning to mask weights. In: European Conference on Computer Vision (ECCV) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.787, + 0.434 + ], + "angle": 0, + "content": "25. Mallya, A., Lazebnik, S.: Packnet: Adding multiple tasks to a single network by iterative pruning. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.787, + 0.46 + ], + "angle": 0, + "content": "26. McCloskey, M., Cohen, N.J.: Catastrophic interference in connectionist networks: The sequential learning problem. Psychology of Learning and Motivation (1989)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.787, + 0.5 + ], + "angle": 0, + "content": "27. Mermillod, M., Bugaiska, A., Bonin, P.: The stability-plasticity dilemma: Investigating the continuum from catastrophic forgetting to age-limited learning effects (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.787, + 0.542 + ], + "angle": 0, + "content": "28. Michieli, U., Zanuttigh, P.: Incremental learning techniques for semantic segmentation. In: Conference on Computer Vision and Pattern Recognition Workshops (CVPR-W) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.787, + 0.583 + ], + "angle": 0, + "content": "29. Michieli, U., Zanuttigh, P.: Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.787, + 0.623 + ], + "angle": 0, + "content": "30. Neuhold, G., Ollmann, T., Rota Bulo, S., Kontschieder, P.: The mapillary vistas dataset for semantic understanding of street scenes. In: International Conference on Computer Vision (ICCV) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.787, + 0.664 + ], + "angle": 0, + "content": "31. Phan, M.H., Ta, T.A., Phung, S.L., Tran-Thanh, L., Bouzerdoum, A.: Class similarity weighted knowledge distillation for continual semantic segmentation. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.787, + 0.704 + ], + "angle": 0, + "content": "32. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.787, + 0.732 + ], + "angle": 0, + "content": "33. Shin, H., Lee, J.K., Kim, J., Kim, J.: Continual learning with deep generative replay. In: Conference on Neural Information Processing Systems (NeurIPS) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.787, + 0.772 + ], + "angle": 0, + "content": "34. Sprechmann, P., Jayakumar, S.M., Rae, J.W., Pritzel, A., Badia, A.P., Uria, B., Vinyals, O., Hassabis, D., Pascanu, R., Blundell, C.: Memory-based parameter adaptation. In: International Conference on Learning Representations (ICLR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "35. Van de Ven, G.M., Siegelmann, H.T., Tolias, A.S.: Brain-inspired replay for continual learning with artificial neural networks. Nature communications (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "36. Xiao, J.W., Zhang, C.B., Feng, J., Liu, X., van de Weijer, J., Cheng, M.M.: Endpoints weight fusion for class incremental semantic segmentation. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.396, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "CLEO: Continual Learning of Evolving Ontologies" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "37. Yang, G., Fini, E., Xu, D., Rota, P., Ding, M., Nabi, M., Alameda-Pineda, X., Ricci, E.: Uncertainty-aware contrastive distillation for incremental semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "38. Yang, G., Fini, E., Xu, D., Rota, P., Ding, M., Tang, H., Alameda-Pineda, X., Ricci, E.: Continual attentive fusion for incremental learning in semantic segmentation. IEEE Transactions on Multimedia (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.259 + ], + "angle": 0, + "content": "39. Yoon, J., Yang, E., Lee, J., Hwang, S.J.: Lifelong learning with dynamically expandable networks. arXiv (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.286 + ], + "angle": 0, + "content": "40. Yu, L., Liu, X., Van de Weijer, J.: Self-training for class-incremental semantic segmentation. IEEE Transactions on Neural Networks and Learning Systems (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.287, + 0.788, + 0.314 + ], + "angle": 0, + "content": "41. Zenke, F., Poole, B., Ganguli, S.: Continual learning through synaptic intelligence. In: International Conference on Machine Learning (ICML) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.315, + 0.788, + 0.357 + ], + "angle": 0, + "content": "42. Zhang, C.B., Xiao, J.W., Liu, X., Chen, Y.C., Cheng, M.M.: Representation compensation networks for continual semantic segmentation. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.398 + ], + "angle": 0, + "content": "43. Zhao, H., Hu, Q., Zhu, P., Wang, Y., Wang, P.: A recursive regularization based feature selection framework for hierarchical classification. Transactions on Knowledge and Data Engineering (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.398 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CLEO_ Continual Learning of Evolving Ontologies/7399fe90-4063-4ee4-8949-4afe680e9068_origin.pdf b/2024/CLEO_ Continual Learning of Evolving Ontologies/7399fe90-4063-4ee4-8949-4afe680e9068_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c16ecc77a8553a02108fb058a1883934647afcd1 --- /dev/null +++ b/2024/CLEO_ Continual Learning of Evolving Ontologies/7399fe90-4063-4ee4-8949-4afe680e9068_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76861a37d070bd0df144bb3574e9f1a07b047083d6efcac0e8bc758901072833 +size 2413925 diff --git a/2024/CLEO_ Continual Learning of Evolving Ontologies/full.md b/2024/CLEO_ Continual Learning of Evolving Ontologies/full.md new file mode 100644 index 0000000000000000000000000000000000000000..02529985723580e4ab32e1de52f993a75647d697 --- /dev/null +++ b/2024/CLEO_ Continual Learning of Evolving Ontologies/full.md @@ -0,0 +1,261 @@ +# CLEO: Continual Learning of Evolving Ontologies + +Shishir Muralidhara1, Saqib Bukhari2, Georg Schneider2, Didier Stricker1,3, and René Schuster1,3 + +$^{1}$ DFKI - German Research Center for Artificial Intelligence + +2 ZF Friedrichshafen AG + +$^{3}$ RPTU - University of Kaiserslautern-Landau + +{firstname这个名字}@{dfki.de，zf.com} + +Abstract. Continual learning (CL) addresses the problem of catastrophic forgetting in neural networks, which occurs when a trained model tends to overwrite previously learned information, when presented with a new task. CL aims to instill the lifelong learning characteristic of humans in intelligent systems, making them capable of learning continuously while retaining what was already learned. Current CL problems involve either learning new domains (domain-incremental) or new and previously unseen classes (class-incremental). However, general learning processes are not just limited to learning information, but also refinement of existing information. In this paper, we define CLEO - Continual Learning of Evolving Ontologies, as a new incremental learning setting under CL to tackle evolving classes. CLEO is motivated by the need for intelligent systems to adapt to real-world ontologies that change over time, such as those in autonomous driving. We use Cityscapes, PASCAL VOC, and Mapillary Vistas to define the task settings and demonstrate the applicability of CLEO. We highlight the shortcomings of existing CIL methods in adapting to CLEO and propose a baseline solution, called Modelling Ontologies (MoOn). CLEO is a promising new approach to CL that addresses the challenge of evolving ontologies in real-world applications. MoOn surpasses previous CL approaches in the context of CLEO. + +Keywords: Continual Learning $\cdot$ Continual Semantic Segmentation $\cdot$ Class-Incremental Learning + +# 1 Introduction + +Advancements in AI have led to the emergence of systems that can match or even outperform human abilities in various domains. However, these systems are often confined to specific datasets and tasks and their efficacy is affected when faced with new tasks or data. This presents a significant hurdle in practical real-world AI applications where the circumstances are constantly evolving. For example, a self-driving car needs to adapt to new traffic regulations, signage, and weather conditions. When a trained model encounters new information or tasks, it experiences a phenomenon known as catastrophic forgetting [26]. This issue arises from the rigidity of neural networks, where previously learned knowledge is + +![](images/6a67cab48edb88c9a6449a043885b11c7bcf63cb7b9690cb17be6611109678e2.jpg) +Fig. 1: CLEO extends class-incremental learning (CIL) by removing the constraint that incremental sets of classes must be non-overlapping. This allows for more flexible and nuanced learning, as existing classes can be refined into more specific classes. + +overwritten when learning something new. Consequently, the model's performance significantly deteriorates, which can have severe repercussions in scenarios where the model has to acclimate to dynamic environments or critical applications where performance is imperative. Therefore, tackling forgetting is crucial when developing AI systems with lifelong learning. + +In contrast, learning in humans is continuous and dynamic through which we continually assimilate new information while preserving and enriching previously learned knowledge. The goal of CL is to instill neural networks with this learning characteristic and make them capable of extending when objectives change. In incremental learning, which involves learning a sequence of tasks, the term task can represent changes either in the distribution of data, referred to as domain-incremental learning (DIL), or changes in the output space, which is known as class-incremental learning (CIL). However, learning is not limited only to the acquisition of new knowledge; it also involves refining and deepening our grasp of existing information. To illustrate, our initial understanding of animals might have been confined to recognizing a broad categorization, like distinguishing between birds and other animals. However, over time, our learning process matures to incorporate finer, more nuanced insights about individual animals. This highlights the necessity of broadening the scope of CL to accommodate both the introduction of novel classes, and also the enhancement of existing knowledge. + +We propose a novel incremental setting termed CLEO - Continual Learning of Evolving Ontologies. CLEO represents a progressive extension of CIL, overcoming the confining assumption that all newly encountered classes are entirely novel and previously unseen. CLEO acknowledges the dynamic nature of learning, where classes may not only emerge but also evolve over time. CLEO combines elements of novelty and refinement, paving the way for more resilient AI systems. The real-world relevance of CLEO is exemplified by the Mapillary Vistas dataset [30], where an initial set of classes has evolved into more fine-grained and specific classes. An illustration of CLEO is presented in Fig. 1 using the CIFAR-10 dataset [17] as an example. In the initial task, the system learns two broad categories, namely animals and vehicles. Subsequently, more specific classes, such as bird, frog, plane, and ship, are split from the original classes. CLEO presents a challenge where the system must not only learn the evolved + +classes, which it had previously encountered within a different semantic context, but also retain the remaining information. In this example, after bird and frog were separated from the animal class, cat and dog are left as the remaining domestic animal class. As in CIL, CLEO also covers the case that entirely new classes are learned, like e.g. the classes horse and deer in the last task of Fig. 1. CLEO embodies the fundamental idea that AI systems must possess the ability to adapt to dynamic shifts in which classes emerge, evolve, and split over time. By integrating this adaptability into AI systems, we enhance their robustness in real-world applications. This paper makes several significant contributions: + +- Formalizing Continual Learning of Evolving Ontologies: A primary contribution is the introduction and formalization of the concept of learning evolving ontologies, from a continual learning perspective. +- Defining Task Settings: To provide a comprehensive understanding of CLEO, we define several task settings for semantic segmentation, using three diverse datasets: Cityscapes [8], PASCAL VOC [11], and Mapillary Vistas [30]. +- Evaluation Metrics for CLEO: By identifying class groups and measuring their performance, we define a more nuanced evaluation of how effectively the system learns refined classes and retains information. +- Baseline Solution - MoOn: To operationalize the concept for semantic segmentation, we propose a baseline solution called "Modelling Ontologies" (MoOn), to handle the challenges inherent in CLEO. + +# 2 Related Work + +# 2.1 Continual Learning + +Continual learning (CL), also known as lifelong learning, and its sub-field of incremental learning, are critical areas of machine learning that address a fundamental limitation of traditional machine learning models: Their rigidity in learning new tasks. CL is a learning paradigm that aims to train models that can learn new tasks and information without forgetting what was previously learned. This contrasts traditional machine learning, where models are typically trained on a single dataset and then remain fixed. Continual learning presents a challenge of balancing between learning new information and retaining the old [27]. Overemphasizing learning new information leads to forgetting the old while focusing too much on retaining old information hinders learning the new task. CL approaches can be grouped into three categories: Architecture-based, replay-based methods, and regularization-based methods. + +Architecture-based approaches modify the architecture of neural networks to accommodate new tasks without overwriting previous information. These include methods such as parameter isolation [12, 24, 25], dynamically growing networks [1, 32, 39], or incorporation of additional components such as memory modules [2, 34]. These methods can face scalability challenges as network architecture grows, increasing memory and computational demands. + +Replay-based approaches are of two types: Experience replay or rehearsal, involves storing a few instances from the previous task [3,6,23], which are used during training on a new task. Generative replay substitutes the memory buffer with a generative model to sample previous task instances when learning on a new task [15,33,35]. Challenges include privacy, data availability, memory and computational constraints in training generative models. + +Regularization-based approaches [16, 20, 41] strike a balance between stability and plasticity when learning new tasks by introducing a regularization term into the loss function. One of the approaches to regularization is to determine the importance of parameters to the previous tasks and penalize overwriting them. Distillation is another approach which preserves previously learned knowledge by distilling information from the old model to the new. + +# 2.2 Continual Semantic Segmentation + +Continual semantic segmentation (CSS) [5,37,38,40] is a specialized domain within CL that focuses on the task of semantic segmentation. Incremental Learning Techniques [28] (ILT) is a regularization-based method that distils knowledge from previously trained task models and additionally freezes the encoder for the current task, which acts as an additional regularization constraint. Modeling the Background [4] (MiB) addresses the problem of background shift in incremental segmentation. It proposes a novel distillation loss that accounts for the changing background between tasks. AWT [13] addresses the background shift in classifier initialization, and transfers the most relevant weights from the previous classifiers to the new classifier. In our baseline solution for CLEO, we will build on the ideas of MiB and extend it to model the semantic shift for all classes. PLOP [10] argues that treating the background class as a single class as done in MiB, which includes both previously encountered and novel classes can lead to catastrophic forgetting of the old classes as it learns new classes. PLOP addresses this problem by pseudolabeling old classes in the background. EWF [36] combines the weights of the old and new model trained with regularization to give a more balanced model. SDR [29] preserves latent representations of old classes through prototype matching and frees space for new classes through feature sparsity and contrastive learning. The decoder preserves previous knowledge via output-level distillation. RCIL [42] uses a representation compensation module to decouple the representation learning of old and new classes. This helps to prevent the model from forgetting old classes when learning new classes. REMINDER [31] proposes a class-similarity weighted distillation that distills knowledge from a previous model on old classes that are similar to the new classes. This reduces forgetting of old classes and improves learning of new classes. + +# 2.3 Evolving Ontologies + +An emerging area within incremental learning focuses on learning evolving classes, which involves the refinement of existing knowledge, where a broad initial grouping of classes evolves over time into more specific and nuanced classes. In contrast, + +in CIL, each task consists of a non-overlapping subsets of novel classes. + +Hierarchical Label Expansion [18] (HLE) for classification involves a model initially learning coarse classes, which then evolve into more fine-grained categories during each task across varying hierarchical levels. The expansion is restrictive, where the entire subset of classes at the current level expands in the next level during each task. CLEO differs from HLE in the following aspects: + +- CLEO allows a more flexible evolution at any point within the sequence of tasks, in which certain classes can remain (partially) unexpanded or expand into sub-groups. +- The baseline solution for HLE is a rehearsal-based method that uses pseudolabeling from previous classifiers - one for each level of the hierarchy - for determining the importance of data. We present a regularization-based solution that uses knowledge distillation. +- HLE considers image classification as application, while CLEO additionally deals with the more challenging problem of semantic segmentation. + +Learning with Evolving Class Ontologies [22] (LECO) has been the first attempt to address evolving ontologies in semantic segmentation. However, despite its intended framing as a continual learning problem, LECO deviates from the defining principles of continual learning. We address these shortcomings and define this problem strictly within the bounds of continual learning. The primary distinctions between our framework and LECO are outlined below: + +- LECO requires and stores all historical data, a constraint that contradicts the definition of CL [19], that all data is not available at once. +- LECO primarily revolves around the dilemma of whether to label new data or relabel the old data with the current label space. In contrast, CLEO confines itself to utilizing the input data and annotations within the dataset. +- Due to the availability of historical data, LECO employs naive strategies such as fine-tuning, freezing, or joint training. We highlight the limitations of existing CIL methods and propose a baseline specifically for CLEO. +- CLEO explicitly addresses catastrophic forgetting, which is one of the core problems of continual learning. Since LECO retrans entirely using all the data, the problem of catastrophic forgetting does not even arise. + +# 3 CLEO + +Deep learning has drawn inspiration from biological systems at various levels, from the way neurons form the fundamental building blocks of artificial neural networks to learning paradigms such as curriculum learning and reinforcement learning. Continual learning is another learning paradigm that aims to instill lifelong learning in intelligent systems. As discussed previously, current problems in continual learning involve adaptation to different domains, such as handling adverse conditions in autonomous driving or learning new classes over time. A + +common characteristic in both these scenarios is that the task always represents new and previously unseen data. However, it is our contention that the scope of continual learning should not be limited solely to the discovery of novel classes. It should also include the enhancement and refinement of existing knowledge. This aligns closely with the way humans continually adapt and refine knowledge through neuroplasticity. + +In light of this perspective, we introduce CLEO - Continual Learning of Evolving Ontologies. CLEO is centered around the idea of removing the constraints of class incremental learning and allows classes that have a hierarchical relation (partially) overlap between tasks, thereby facilitating the refinement of already learned information. CLEO is a real and open-world problem and this is exemplified in the Mapillary Vistas dataset [30]. The initial version of the dataset defined 66 classes, which later evolved into 124 classes in the latest version. It is noteworthy to observe the transformation within class categorization. On one hand, classes that were previously categorized under unlabeled, such as traffic island emerge as novel classes. Conversely, several pre-existing classes undergo refinement, leading to the creation of new, more specific categories. In Fig. 2, we can observe a tangible example of this evolution, where the previous billboard class has evolved into more specialized classes, including signage classes for advertisements, information, stores, and differentiating between the front and back of the signs. In critical applications like autonomous driving, it becomes imperative for AI systems to swiftly adapt to these nuanced changes, perpetually learn, and enhance scene understanding precision. + +Let $T = \{0,1,2,\dots,n\}$ denote a sequence of $n + 1$ tasks. Each task is associated with a specific subset of classes denoted as $C_t$ . The union of all classes across tasks is $C = \bigcup_{t=0}^{n} C_t$ , representing the entire set of classes encountered during the learning process. At each task $t$ , a subset of classes $C_t$ is introduced incrementally similar to CIL. $C_{0:t} = C_0 \cup \ldots \cup C_{t-1} \cup C_t$ is the set of all classes seen until step $t$ . CLEO acknowledges the possibility of class splitting, where an existing class $c_{t-1}$ from $C_{0:t-1}$ can evolve into multiple new classes in $C_t$ . Formally this can be expressed as, $C_t \subset C$ , $C_i \cap C_j = \phi$ for $i \neq j$ and the class splitting is expressed as $c_{t-1} \in C_{0:t-1} \to \{c_{k+1}, c_{k+2}, \ldots, c_{k+m}\} \subseteq C_t$ . Note that the splitting of a class is not necessarily exhaustive, i.e., $c_{t-1} \in C_{0:t-1}$ may persist. An example for such an evolution of classes is given in Fig. 1. The learning objective of a task remains consistent with CIL: Train a model to recognize and differentiate the current subset of classes $C_t$ , while ensuring that the model retains knowledge from the previously encountered classes in $C_{0:t-1}$ . As in CIL, the model is not provided access to data from previous tasks. + +# 4 MoOn: Modelling Ontologies for CLEO + +Class-incremental learning involves the process of training a model to incrementally learn new classes. At each increment, the model is presented with images and the corresponding ground truth which contains annotations for the new classes + +![](images/4b75010f392e681e944f2d94b7bc5d60b8963568403800a73d85ac0878de5d81.jpg) +Fig. 2: The Mapillary Vistas dataset [30] provides a real-world example for an evolving ontology. This figure compares ground truth annotations for the same image in two dataset versions. It illustrates the introduction of new classes, such as traffic island and the finer categorization of existing classes such as billboard. + +![](images/cc4e37b629060ca9560efa75e50c6895866c1e867dede7c8351b0bb16a904033.jpg) + +![](images/251045c1e490a729fb916f4266cdfdeadf7dc8be29e52a44850f2132129c3aea.jpg) + +being learned. In class-incremental semantic segmentation (CISS), at each step, all unlabeled areas are marked as background. Thus in CISS, the phenomenon of semantic background shift occurs, because in each task, the meaning of the background class $(bg)$ changes. Distillation loss is a technique of transferring knowledge from an old model $f_{t-1}$ to a new model $f_t$ , and this is achieved by comparing the old model's prediction to the new model's predictions. The distillation loss enforces the new model's predictions to resemble those of the old model by penalizing any deviations and this helps mitigate forgetting. Consider an image $x$ with $N$ pixels and the corresponding ground truth $y$ at the current task $t$ with the class set $C_t$ . The standard distillation loss in [4] is defined as follows: + +$$ +\ell_ {k d} ^ {t} (x, y) = - \frac {1}{N} \sum_ {i \in x} \sum_ {c \in C _ {0: t - 1}} p _ {x} ^ {t - 1} (i, c) \log \hat {p} _ {x} ^ {t} (i, c) \tag {1} +$$ + +where $p_x^{t-1}(i, c)$ is the probability for class $c$ in pixel $i$ given by the model $f_{t-1}$ and $\hat{p}_x^t(i, c)$ is for class $c$ in pixel $i$ given by the model $f_t$ but re-normalized across classes $C_{0:t-1}$ : + +$$ +\hat {p} _ {x} ^ {t} (i, c) = \left\{ \begin{array}{l l} 0 & \text {i f} c \in C _ {t} \backslash \{b g \} \\ \frac {p _ {x} ^ {t} (i , c)}{\sum_ {k \in C _ {0 : t - 1}} p _ {x} ^ {t} (i , k)} & \text {i f} c \in C _ {0: t - 1} \end{array} \right. \tag {2} +$$ + +Standard distillation ignores the dynamic nature of the background in incremental semantic segmentation. The background class in segmentation is a catch-all class that includes all previously seen classes and potential future classes. Furthermore, the background class is present in all tasks and changes as new classes are added, leading to background shift. A class $c \in C_t$ added during the current task was previously part of the background in $C_{t-1}$ . For a pixel representing a new class at the current task $t$ , the old model $f_{t-1}$ would predict the background class, and if standard distillation loss was applied naively, it would penalize the new model $f_t$ for predicting the correct class. Modeling the Background (MiB) [4] proposes a novel distillation loss that explicitly addresses the background shift in incremental segmentation. MiB does not directly compare the prediction of the + +![](images/28a6aea6c00ffa2273b88ce4ef87e7ca1ce375a812a52d46eb31df18b2c89f4a.jpg) +Fig. 3: Modelling Ontologies (MoOn) generalizes the idea of MiB [4] for all classes. During distillation, forgetting is mitigated, by grouping class-logits according to the evolution of the class ontology. In this example, the teacher network of the previous task predicts the vehicle class, while the model tries to learn bicycle. To avoid penalizing this, the predictions for vehicle and bicycle are combined before the distillation loss. + +new model $f_{t}$ with the corresponding prediction of the old model $f_{t-1}$ . Instead, the background prediction by $f_{t-1}$ is compared with the $f_{t}$ prediction of either a new class or the background. Eq. (3) modifies Eq. (2) by additionally considering the whole label space instead of limiting it to $C_{0:t-1}$ . MiB acknowledges that pixels that were thought to be part of the background by the old model might actually be the new classes the current model is trying to learn, thus avoiding penalizing the probabilities of new classes being learned. + +$$ +\hat {p} _ {x} ^ {t} (i, c) = \left\{ \begin{array}{l l} p _ {x} ^ {t} (i, c) & \text {i f} c \neq b g \\ \sum_ {k \in C _ {t}} p _ {x} ^ {t} (i, k) & \text {i f} c = b g \end{array} \right. \tag {3} +$$ + +MiB is designed for CIL, where the new classes being learned at the current step, were previously unseen and are always a part of the background class. Therefore, it would suffice to take into consideration the background class, and hence modeling the background. Whereas CLEO is designed for learning of evolving classes and the new classes being learned may have been previously unseen and part of the background or been part of a broader class. As seen in Fig. 3, the class bike being learned at the current step may have been previously learned as a vehicle class grouped with other classes like truck and bus. Therefore we need to take into consideration that within CLEO, new classes may not always have been a part of the background class previously. + +To this end, we propose Modelling Ontologies (MoOn) as a baseline solution for CLEO, which can be seen as a generalized version of MiB that accounts for the fact that new classes may emerge from any existing class, and not just the background class. Similar to how MiB considered the possibility of the new class + +pixel belonging to the background class previously, MoOn additionally includes the possibility of the pixel belonging to an existing class. In MiB, the old model $f_{t-1}$ may predict the background class for the new class being learned, whereas the old model in MoOn may predict an existing class $c \in C_{0:t-1}$ . Therefore, the correct distillation would be to distill the knowledge from the old model $f_{t-1}$ for the existing class $c \in C_{0:t-1}$ that is being split into new classes in $C_t$ . In the illustrated scenario depicted in Fig. 3, distillation would involve transferring knowledge from the old model regarding the vehicle class, while training the new model to predict the bike class. To achieve this, at each step $t$ , MoOn uses a set $S_t$ of class maps $(c_{t-1} \to \{c_{k+1}, c_{k+2}, \dots, c_{k+m}\} = C_{new})$ to identify the class $c_{t-1} \in C_{0:t-1}$ from which the current classes in $C_t$ are split. If the new class $c_t$ is completely new, then the superclass $c_{t-1}$ in the class mapping would be the background class $bg$ . We define $C_{e,t}$ as the set of all evolving classes $c_{t-1}$ in $S_t$ . The modified distillation loss for learning evolving classes in CLEO is given as: + +$$ +\hat {p} _ {x} ^ {t} (i, c) = \left\{ \begin{array}{l l} p _ {x} ^ {t} (i, c) & \text {i f} c \notin C _ {e, t} \\ p _ {x} ^ {t} (i, c _ {t - 1}) + \sum_ {c _ {n} \in C _ {\text {n e w}}} p _ {x} ^ {t} (i, c _ {n}) & \text {i f} c \in C _ {e, t} \end{array} \right. \tag {4} +$$ + +# 4.1 Datasets + +Towards the introduction of CLEO for semantic segmentation, we consider the following three diverse datasets: + +- Cityscapes: Cityscapes [8] is a large-scale dataset of urban street scenes captured in 50 different cities under daytime and good weather conditions. The dataset provides annotations for 30 classes out of which 19 are used for evaluation and grouped into 7 categories. An overview of the hierarchy of the 19 semantic classes grouped by their parent categories is presented in the supplementary. +- PASCAL VOC: PASCAL VOC 2012 [11] is part of the Visual Object Classes Challenge, which ran from 2005 to 2012. PASCAL VOC is a more general dataset like COCO [21] consisting of 20 categories, including vehicles, household objects, animals, and persons. The categorization by [43] of these classes along with intermediate groupings is shown in the supplementary. +- Mapillary Vistas: Mapillary Vistas [30] is a large-scale street-level image dataset for understanding street scenes around the world. It covers a wider range of geographic locations and diverse weather and illumination conditions. The initial version of the dataset provided annotations for 66 classes, which later nearly doubled into 124 classes in the current version v2.0. + +# 4.2 Baselines and Implementation + +We evaluate our method's performance against standard CL baselines, namely fine-tuning and joint training. Fine-tuning involves using the previous step model + +as the starting point when learning on the current task. This is similar to transfer learning and even results in positive forward transfer. However, since there is no explicit attempt to prevent catastrophic forgetting, it involves the highest amount of forgetting and forms the lower baseline. The joint training model involves learning all of the tasks together in a single step, and as there is no incremental learning there is zero forgetting and this forms the theoretical upper baseline. Since in CLEO a redefinition of classes may occur because of the evolving ontology, we define the set of classes for the joint training as the final set of classes after the last task. Additionally, we compare against the state-of-the-art (SOTA) methods such as MiB [4], PLOP [10] and RCIL [42]. + +For MiB, and our proposed approach, we adopt the framework proposed by MiB [4]. In the case of PLOP and RCIL, which involve modifications specific to their approaches, we adopt the RCIL framework [42]. All these approaches use the DeepLabV3 [7] architecture as segmentation network, utilizing a ResNet-101 [14] backbone with an output stride of 16. The backbone's initialization uses a pre-trained model from ImageNet [9]. The training methodology utilizes SGD, with a learning rate of 0.01 for the initial learning step and 0.001 for subsequent steps. Across all learning steps, we train the model using a batch size of 24, for 50 epochs on Cityscapes and for 30 epochs on PASCAL and Mapillary Vistas. + +# 4.3 Evaluation Metrics + +Mean Intersection-over-Union (mIoU) is a commonly used metric for evaluating semantic segmentation models. It measures the average overlap between the predicted segmentation and the ground truth segmentation for all classes. To calculate the mIoU, the Intersection-over-Union (IoU) is first calculated for each class, and then the IoU scores for all classes are averaged. The performance of CL models are often evaluated by measuring the amount of information forgotten by comparing the model's performance on old tasks against the joint training baseline. Within the context of CLEO, while learning evolving classes, forgetting can occur at multiple levels. We identify three types of class groups and analyze the mIoU results for each of them: + +- Unsplit Classes: These classes are learned in the first task and are not further split. Their performance is expected to be largely unaffected, as they have not been split nor learned incrementally. +- Split Classes: These classes have been split either from the background or a previously learned superclass. The mIoU results for these split classes will provide insight into the model's ability to learn new classes. +- Retained Classes: These are classes remaining in the parent class after it has been split partially to yield new classes. Analyzing the mIoU scores for retained classes sheds light on how well the model retains knowledge. + +Table 1: Results for our two CLEO settings on Cityscapes [8] after learning all tasks. + +
MethodCS-Ex1CS-Ex2
UnsplitSplitRetainedAllUnsplitSplitRetainedAll
Fine-Tuning00.0000.0000.0000.0000.0014.4600.0008.68
Joint Training90.2661.1548.0455.6290.2661.1548.0455.62
MiB [4]90.2403.6607.9909.1190.2318.2511.3918.88
PLOP [10]87.5834.3613.0228.9186.0131.9115.5128.10
RCIL [42]89.8200.0007.9706.8889.0000.0007.7006.76
MoOn (Ours)90.2139.5037.0039.3190.2342.2835.7840.62
+ +Table 2: Task-wise results for CS-Ex2 after learning the final task. + +
MethodTask 0Task 1Task 2Task 3Task 4Task 5Task 6All
Fine-Tuning00.0000.0000.0000.0000.0000.0034.7008.68
Joint Training47.5097.5763.0930.2088.3164.3157.9355.62
MiB [4]19.8202.8109.2600.1974.0045.6215.5418.88
PLOP [10]22.3895.0841.5900.0084.8153.8413.2128.10
RCIL [42]16.9000.0000.0000.0000.0000.0000.0006.76
MoOn (Ours)38.1196.8956.2609.1988.1660.9726.1040.62
+ +# 4.4 Experimental Settings and Results + +We introduce seven initial experimental settings, that are designed for the previously introduced datasets and resemble the complexity and diversity of CLEO beyond classical CIL. For completeness and better understanding, we present the detailed semantic evolution of all settings by listing $C_t$ for every task in the supplementary, including the class groups as introduced in Sec. 4.3. + +# 5 Experiments + +Cityscapes From the Cityscapes class hierarchy, we derive two experimental settings. Both settings start in task 0 by learning the seven parent classes which together consist of the 19 grouped classes. The settings involve the incremental splitting of individual classes from parent classes. We do not always completely split a parent class, thus emphasizing not only learning the split class but also retaining the remaining information on the superclass. + +- CS-Ex1: Each step involves splitting out a single class from each parent class. Since there is an unequal number of classes in the parent classes, we repeat this process until we have 19 individual classes, resulting in six tasks. +- CS-Ex2: At each step, we split all but one of the classes from a single parent class to form individual classes. This results in 19 classes in six evolutionary increments (seven tasks in total) for six parent classes excluding sky. + +Table 3: Results for our three CLEO settings on PASCAL [11] after learning all tasks. + +
MethodVOC-Ex1VOC-Ex2VOC-Ex3
UnsplitSplitRetainedAllUnsplitSplitRetainedAllUnsplitSplitRetainedAll
Fine-Tuning37.7736.3100.0029.6837.7405.3800.0007.8136.9125.6800.0023.08
Joint Training90.1972.5384.1276.9190.4873.7889.1477.7589.8475.2787.3678.38
MiB [4]88.9137.2854.9947.0586.9911.2837.8722.8487.0135.7533.2840.28
PLOP [10]86.8538.6743.5645.6285.7716.1803.6721.2683.4136.7406.6136.88
RCIL [42]88.0800.8441.1119.2985.2901.4323.8513.1887.4501.9918.6312.51
MoOn (Ours)88.9555.9868.8462.5284.4522.0825.4228.8283.3347.5142.3650.19
+ +The results from these settings are presented in Tab. 1 and the task-wise results of $CS - Ex2$ in Tab. 2. We can observe the shortcomings of existing CIL methods in handling evolving ontologies, whereas our baseline solution, MoOn, outperforms these methods. All methods achieve similar results on the unsplit classes, which is sky in both settings. However, when learning split classes, only PLOP comes close to our approach. Notably, MoOn surpasses PLOP by a significant margin in preserving information within the existing class. In the second task setting, the last step involves learning five vehicle classes, which are well-represented in the dataset. FT and MIB methods benefit from this, whereas PLOP and RCIL achieve similar results across both settings. We provide task-wise results of $CS - Ex1$ and visualizations for MoOn in the supplementary. + +PASCAL VOC From the PASCAL class hierarchy, we derive three experimental settings that represent the challenges of CLEO. In the first step (task 0), the model learns all the 20 classes grouped into the four parent classes namely animals, household, person, and vehicle. The person class does not contain any subclasses and therefore does not undergo splitting. + +- VOC-Ex1: This setting includes splitting into intermediate sub-groups. Task 1 includes learning the sub-groups Farmyard, furniture, and 2-wheeler which in turn contain individual classes. In the next and final step, these sub-groups are split into their individual classes. +- VOC-Ex2: One class is split sequentially from each parent class resulting in five increments. At the end of the learning, we retain the individual classes cat, bottle, and the subgroup 4-wheeler in their respective parent classes. +- VOC-Ex3: This setting involves splitting all but one of the classes from each parent class sequentially, resulting in four tasks, with three evolutionary steps for the three parent classes excluding person which remains unsplit. + +The results on the three settings are shown in Tab. 3, while the corresponding visualizations for VOC-Ex3 are presented in Fig. 4. The performance on the unsplit classes, person class in this dataset, remains consistent across the CL methods. We observe the biggest difference in performance between our approach and others for VOC-Ex1. This is likely due to the more challenging multiple splitting of classes into intermediate subclasses. The lowest results for all methods + +![](images/5aa9443b3e7ff40950ea72b79650161cff2c77f6119fb371201679b544c4e85d.jpg) +Fig. 4: Qualitative comparison on PASCAL VOC [11] VOC-Ex3 after the final task. + +Table 4: Results for CLEO settings on Mapillary Vistas [30] after learning all tasks. + +
MethodMV-Ex1MV-Ex2
UnsplitSplitRetainedAllUnsplitSplitRetainedAll
Fine-Tuning00.0018.5900.0008.6900.0001.4700.0000.69
Joint Training28.3722.6825.7925.5128.3722.6825.7925.51
MiB [4]27.0216.3317.7421.2709.8304.3100.0706.45
PLOP [10]27.0713.1418.4419.8603.5901.2600.0002.22
RCIL [42]28.5401.5712.7814.6506.3900.5500.1703.16
MoOn (Ours)26.7216.6719.4721.4409.8104.4000.0106.49
+ +are for VOC-Ex2, which involves a longer sequence of tasks, and learning sets of mixed classes one from each parent class at each step. + +Mapillary Vistas Using the two dataset versions of Mapillary Vistas [30], we have defined two CLEO experiments. Both settings begin by learning the original 66 classes in task 0. The newly introduced classes are grouped into their ten parent classes from the previous version resulting in the original classes. The class mapping between the new classes and their parent classes is provided in the supplementary. The subsequent steps involve learning these new classes. + +- MV-Ex1: After learning the original classes in task 0, all of the new classes are learned in task 1, resulting in two tasks. This exemplifies a real-world application of CLEO, where the model is specifically trained to learn the new classes without a complete retraining of all the classes. +- MV-Ex2: This setting is more challenging with a longer sequence of tasks, which can exacerbate forgetting. In each step, one of the ten parent classes is split into its new classes, resulting in a total of eleven tasks. + +Mapillary Vistas is challenging due to the large number of classes. As shown in Tab. 4, all methods exhibit suboptimal performance, including joint training which does not suffer from forgetting. The two settings are similar and have + +learned the same classes after the final task. They differ in the number of steps, highlighting the effect of a longer sequence of tasks. $MV - Ex2$ with a longer sequence significantly impacts the performance of all methods. + +# 6 Limitations and Future Work + +In its most general form, CLEO allows for merging of two or more known classes as a valid form of evolution. However, we note that the here presented formalism of CLEO only considers the separation of known parent classes. Yet, in some rare cases, it might be required to combine previous information to simplify previous tasks and free representational capacity of a model. This, however, as a version of intended forgetting opens up a mostly unexplored field of research in continual learning and is left for future work. Besides, semantic context can be achieved through post-processing by combining predictions of existing and new classes into a single class, preventing interference and avoiding further forgetting. + +We further highlight that our baseline solution MoOn requires the information of class relations $S_{t}$ about the evolution of classes. While we believe that this information is usually available to a sequentially trained model, there might be cases in which explicit knowledge about the ontology is unavailable. In such cases, $S_{t}$ could be inferred from the existing model's predictions on the current task data. If the model predicts an existing class, the class being learned was previously seen and is split from this parent class. If it predicts the background, the new class is entirely new and unseen. In our experimental evaluation, we found that MoOn and other approaches notably underperform with longer sequences in CLEO, highlighting the need for future attention to long-term learning settings. + +# 7 Conclusion + +We introduce a novel framework for addressing the challenge of adapting to evolving classes, termed Continual Learning of Evolving Ontologies (CLEO). The motivation for CLEO stems from a practical scenario wherein pre-existing classes undergo transformations over time, resulting in more specialized classes. CLEO extends the functionality of CIL to not just learn new information but also refine existing information. The applicability of CLEO in the real world is exemplified through the Mapillary Vistas dataset, where an existing set of classes evolves into more detailed classes. Furthermore, we highlight the shortcomings of existing CIL methods and propose MoOn as an initial solution for CLEO. The effectiveness of MoOn for CLEO is demonstrated on seven experimental settings across three datasets, surpassing previous CIL approaches. + +# Acknowledgments + +This work was partially funded by the Federal Ministry of Education and Research Germany under the project DECODE (01IW21001). + +# References + +1. Aljundi, R., Chakravarty, P., Tuytelaars, T.: Expert gate: Lifelong learning with a network of experts. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2017) +2. Asghar, N., Mou, L., Selby, K.A., Pantasdo, K.D., Poupart, P., Jiang, X.: Progressive memory banks for incremental domain adaptation. arXiv (2018) +3. Bang, J., Kim, H., Yoo, Y., Ha, J.W., Choi, J.: Rainbow memory: Continual learning with a memory of diverse samples. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +4. Cermelli, F., Mancini, M., Bulo, S.R., Ricci, E., Caputo, B.: Modeling the background for incremental learning in semantic segmentation. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2020) +5. Cha, S., kim, b., Yoo, Y., Moon, T.: Ssul: Semantic segmentation with unknown label for exemplar-based class-incremental learning. In: Advances in Neural Information Processing Systems (2021) +6. Chaudhry, A., Rohrbach, M., Elhoseiny, M., Ajanthan, T., Dokania, P.K., Torr, P.H., Ranzato, M.: On tiny episodic memories in continual learning. arXiv (2019) +7. Chen, L.C., Papandreou, G., Schroff, F., Adam, H.: Rethinking atrous convolution for semantic image segmentation. arXiv (2017) +8. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016) +9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2009) +10. Douillard, A., Chen, Y., Dapogny, A., Cord, M.: Plop: Learning without forgetting for continual semantic segmentation. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +11. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International Journal of Computer Vision (IJCV) (2010) +12. Fernando, C., Banarse, D., Blundell, C., Zwols, Y., Ha, D., Rusu, A.A., Pritzel, A., Wierstra, D.: Pathnet: Evolution channels gradient descent in super neural networks. arXiv (2017) +13. Goswami, D., Schuster, R., van de Weijer, J., Stricker, D.: Attribution-aware weight transfer: A warm-start initialization for class-incremental semantic segmentation In: Winter Conference on Applications of Computer Vision (2023) +14. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016) +15. Kemker, R., Kanan, C.: Fearnet: Brain-inspired model for incremental learning. In: International Conference on Learning Representations (2018) +16. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences (2017) +17. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) + +18. Lee, B.H., Jung, O., Choi, J., Chun, S.Y.: Online continual learning on hierarchical label expansion. In: International Conference on Computer Vision (ICCV) (2023) +19. Lesort, T., Lomonaco, V., Stoian, A., Maltoni, D., Filliat, D., Diaz-Rodriguez, N.: Continual learning for robotics: Definition, framework, learning strategies, opportunities and challenges. Information fusion (2020) +20. Li, Z., Hoiem, D.: Learning without forgetting. Transactions on Pattern Analysis and Machine Intelligence (T-PAMI) (2017) +21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dóllar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision (ECCV) (2014) +22. Lin, Z., Pathak, D., Wang, Y.X., Ramanan, D., Kong, S.: Continual learning with evolving class ontologies. In: Conference on Neural Information Processing Systems (NeurIPS) (2022) +23. Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. In: Conference on Neural Information Processing Systems (NeurIPS) (2017) +24. Mallya, A., Davis, D., Lazebnik, S.: Piggyback: Adapting a single network to multiple tasks by learning to mask weights. In: European Conference on Computer Vision (ECCV) (2018) +25. Mallya, A., Lazebnik, S.: Packnet: Adding multiple tasks to a single network by iterative pruning. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2018) +26. McCloskey, M., Cohen, N.J.: Catastrophic interference in connectionist networks: The sequential learning problem. Psychology of Learning and Motivation (1989) +27. Mermillod, M., Bugaiska, A., Bonin, P.: The stability-plasticity dilemma: Investigating the continuum from catastrophic forgetting to age-limited learning effects (2013) +28. Michieli, U., Zanuttigh, P.: Incremental learning techniques for semantic segmentation. In: Conference on Computer Vision and Pattern Recognition Workshops (CVPR-W) (2019) +29. Michieli, U., Zanuttigh, P.: Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2021) +30. Neuhold, G., Ollmann, T., Rota Bulo, S., Kontschieder, P.: The mapillary vistas dataset for semantic understanding of street scenes. In: International Conference on Computer Vision (ICCV) (2017) +31. Phan, M.H., Ta, T.A., Phung, S.L., Tran-Thanh, L., Bouzerdoum, A.: Class similarity weighted knowledge distillation for continual semantic segmentation. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +32. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv (2016) +33. Shin, H., Lee, J.K., Kim, J., Kim, J.: Continual learning with deep generative replay. In: Conference on Neural Information Processing Systems (NeurIPS) (2017) +34. Sprechmann, P., Jayakumar, S.M., Rae, J.W., Pritzel, A., Badia, A.P., Uria, B., Vinyals, O., Hassabis, D., Pascanu, R., Blundell, C.: Memory-based parameter adaptation. In: International Conference on Learning Representations (ICLR) (2018) +35. Van de Ven, G.M., Siegelmann, H.T., Tolias, A.S.: Brain-inspired replay for continual learning with artificial neural networks. Nature communications (2020) +36. Xiao, J.W., Zhang, C.B., Feng, J., Liu, X., van de Weijer, J., Cheng, M.M.: Endpoints weight fusion for class incremental semantic segmentation. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023) + +37. Yang, G., Fini, E., Xu, D., Rota, P., Ding, M., Nabi, M., Alameda-Pineda, X., Ricci, E.: Uncertainty-aware contrastive distillation for incremental semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022) +38. Yang, G., Fini, E., Xu, D., Rota, P., Ding, M., Tang, H., Alameda-Pineda, X., Ricci, E.: Continual attentive fusion for incremental learning in semantic segmentation. IEEE Transactions on Multimedia (2022) +39. Yoon, J., Yang, E., Lee, J., Hwang, S.J.: Lifelong learning with dynamically expandable networks. arXiv (2017) +40. Yu, L., Liu, X., Van de Weijer, J.: Self-training for class-incremental semantic segmentation. IEEE Transactions on Neural Networks and Learning Systems (2022) +41. Zenke, F., Poole, B., Ganguli, S.: Continual learning through synaptic intelligence. In: International Conference on Machine Learning (ICML) (2017) +42. Zhang, C.B., Xiao, J.W., Liu, X., Chen, Y.C., Cheng, M.M.: Representation compensation networks for continual semantic segmentation. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +43. Zhao, H., Hu, Q., Zhu, P., Wang, Y., Wang, P.: A recursive regularization based feature selection framework for hierarchical classification. Transactions on Knowledge and Data Engineering (2021) \ No newline at end of file diff --git a/2024/CLEO_ Continual Learning of Evolving Ontologies/images.zip b/2024/CLEO_ Continual Learning of Evolving Ontologies/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fec1d467fef4de2b35ea9663fdd2c5a36ff661c3 --- /dev/null +++ b/2024/CLEO_ Continual Learning of Evolving Ontologies/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cda57272fdefb29722b6ce10672a8b8b6257f2865ef7a5945f64dc563ffb6f65 +size 422740 diff --git a/2024/CLEO_ Continual Learning of Evolving Ontologies/layout.json b/2024/CLEO_ Continual Learning of Evolving Ontologies/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bf5f198b5d7830f4f21f9df37d2840d973ac2606 --- /dev/null +++ b/2024/CLEO_ Continual Learning of Evolving Ontologies/layout.json @@ -0,0 +1,8111 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 132, + 111, + 481, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 111, + 481, + 131 + ], + "spans": [ + { + "bbox": [ + 132, + 111, + 481, + 131 + ], + "type": "text", + "content": "CLEO: Continual Learning of Evolving Ontologies" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 149, + 483, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 483, + 174 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 483, + 174 + ], + "type": "text", + "content": "Shishir Muralidhara1, Saqib Bukhari2, Georg Schneider2, Didier Stricker1,3, and René Schuster1,3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 181, + 183, + 430, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 183, + 430, + 194 + ], + "spans": [ + { + "bbox": [ + 181, + 183, + 430, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 181, + 183, + 430, + 194 + ], + "type": "text", + "content": " DFKI - German Research Center for Artificial Intelligence" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 255, + 195, + 358, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 195, + 358, + 205 + ], + "spans": [ + { + "bbox": [ + 255, + 195, + 358, + 205 + ], + "type": "text", + "content": "2 ZF Friedrichshafen AG" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 209, + 205, + 404, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 205, + 404, + 217 + ], + "spans": [ + { + "bbox": [ + 209, + 205, + 404, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 209, + 205, + 404, + 217 + ], + "type": "text", + "content": " RPTU - University of Kaiserslautern-Landau" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 217, + 217, + 396, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 217, + 396, + 228 + ], + "spans": [ + { + "bbox": [ + 217, + 217, + 396, + 228 + ], + "type": "text", + "content": "{firstname这个名字}@{dfki.de，zf.com}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 258, + 455, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 258, + 455, + 478 + ], + "spans": [ + { + "bbox": [ + 159, + 258, + 455, + 478 + ], + "type": "text", + "content": "Abstract. Continual learning (CL) addresses the problem of catastrophic forgetting in neural networks, which occurs when a trained model tends to overwrite previously learned information, when presented with a new task. CL aims to instill the lifelong learning characteristic of humans in intelligent systems, making them capable of learning continuously while retaining what was already learned. Current CL problems involve either learning new domains (domain-incremental) or new and previously unseen classes (class-incremental). However, general learning processes are not just limited to learning information, but also refinement of existing information. In this paper, we define CLEO - Continual Learning of Evolving Ontologies, as a new incremental learning setting under CL to tackle evolving classes. CLEO is motivated by the need for intelligent systems to adapt to real-world ontologies that change over time, such as those in autonomous driving. We use Cityscapes, PASCAL VOC, and Mapillary Vistas to define the task settings and demonstrate the applicability of CLEO. We highlight the shortcomings of existing CIL methods in adapting to CLEO and propose a baseline solution, called Modelling Ontologies (MoOn). CLEO is a promising new approach to CL that addresses the challenge of evolving ontologies in real-world applications. MoOn surpasses previous CL approaches in the context of CLEO." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 489, + 453, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 489, + 453, + 511 + ], + "spans": [ + { + "bbox": [ + 159, + 489, + 453, + 511 + ], + "type": "text", + "content": "Keywords: Continual Learning " + }, + { + "bbox": [ + 159, + 489, + 453, + 511 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 489, + 453, + 511 + ], + "type": "text", + "content": " Continual Semantic Segmentation " + }, + { + "bbox": [ + 159, + 489, + 453, + 511 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 489, + 453, + 511 + ], + "type": "text", + "content": " Class-Incremental Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 532, + 230, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 230, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 230, + 544 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Advancements in AI have led to the emergence of systems that can match or even outperform human abilities in various domains. However, these systems are often confined to specific datasets and tasks and their efficacy is affected when faced with new tasks or data. This presents a significant hurdle in practical real-world AI applications where the circumstances are constantly evolving. For example, a self-driving car needs to adapt to new traffic regulations, signage, and weather conditions. When a trained model encounters new information or tasks, it experiences a phenomenon known as catastrophic forgetting [26]. This issue arises from the rigidity of neural networks, where previously learned knowledge is" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 196 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 196 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 196 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 196 + ], + "type": "image", + "image_path": "6a67cab48edb88c9a6449a043885b11c7bcf63cb7b9690cb17be6611109678e2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 205, + 482, + 239 + ], + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 239 + ], + "type": "text", + "content": "Fig. 1: CLEO extends class-incremental learning (CIL) by removing the constraint that incremental sets of classes must be non-overlapping. This allows for more flexible and nuanced learning, as existing classes can be refined into more specific classes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 258, + 479, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 479, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 479, + 317 + ], + "type": "text", + "content": "overwritten when learning something new. Consequently, the model's performance significantly deteriorates, which can have severe repercussions in scenarios where the model has to acclimate to dynamic environments or critical applications where performance is imperative. Therefore, tackling forgetting is crucial when developing AI systems with lifelong learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 498 + ], + "type": "text", + "content": "In contrast, learning in humans is continuous and dynamic through which we continually assimilate new information while preserving and enriching previously learned knowledge. The goal of CL is to instill neural networks with this learning characteristic and make them capable of extending when objectives change. In incremental learning, which involves learning a sequence of tasks, the term task can represent changes either in the distribution of data, referred to as domain-incremental learning (DIL), or changes in the output space, which is known as class-incremental learning (CIL). However, learning is not limited only to the acquisition of new knowledge; it also involves refining and deepening our grasp of existing information. To illustrate, our initial understanding of animals might have been confined to recognizing a broad categorization, like distinguishing between birds and other animals. However, over time, our learning process matures to incorporate finer, more nuanced insights about individual animals. This highlights the necessity of broadening the scope of CL to accommodate both the introduction of novel classes, and also the enhancement of existing knowledge." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "type": "text", + "content": "We propose a novel incremental setting termed CLEO - Continual Learning of Evolving Ontologies. CLEO represents a progressive extension of CIL, overcoming the confining assumption that all newly encountered classes are entirely novel and previously unseen. CLEO acknowledges the dynamic nature of learning, where classes may not only emerge but also evolve over time. CLEO combines elements of novelty and refinement, paving the way for more resilient AI systems. The real-world relevance of CLEO is exemplified by the Mapillary Vistas dataset [30], where an initial set of classes has evolved into more fine-grained and specific classes. An illustration of CLEO is presented in Fig. 1 using the CIFAR-10 dataset [17] as an example. In the initial task, the system learns two broad categories, namely animals and vehicles. Subsequently, more specific classes, such as bird, frog, plane, and ship, are split from the original classes. CLEO presents a challenge where the system must not only learn the evolved" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "type": "text", + "content": "S. Muralidhara et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "classes, which it had previously encountered within a different semantic context, but also retain the remaining information. In this example, after bird and frog were separated from the animal class, cat and dog are left as the remaining domestic animal class. As in CIL, CLEO also covers the case that entirely new classes are learned, like e.g. the classes horse and deer in the last task of Fig. 1. CLEO embodies the fundamental idea that AI systems must possess the ability to adapt to dynamic shifts in which classes emerge, evolve, and split over time. By integrating this adaptability into AI systems, we enhance their robustness in real-world applications. This paper makes several significant contributions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 232, + 480, + 385 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 232, + 480, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 480, + 266 + ], + "type": "text", + "content": "- Formalizing Continual Learning of Evolving Ontologies: A primary contribution is the introduction and formalization of the concept of learning evolving ontologies, from a continual learning perspective." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 267, + 480, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 480, + 314 + ], + "type": "text", + "content": "- Defining Task Settings: To provide a comprehensive understanding of CLEO, we define several task settings for semantic segmentation, using three diverse datasets: Cityscapes [8], PASCAL VOC [11], and Mapillary Vistas [30]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 316, + 480, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 480, + 349 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 480, + 349 + ], + "type": "text", + "content": "- Evaluation Metrics for CLEO: By identifying class groups and measuring their performance, we define a more nuanced evaluation of how effectively the system learns refined classes and retains information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 351, + 480, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 351, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 351, + 480, + 385 + ], + "type": "text", + "content": "- Baseline Solution - MoOn: To operationalize the concept for semantic segmentation, we propose a baseline solution called \"Modelling Ontologies\" (MoOn), to handle the challenges inherent in CLEO." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 405, + 237, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 405, + 237, + 417 + ], + "spans": [ + { + "bbox": [ + 132, + 405, + 237, + 417 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 430, + 258, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 258, + 442 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 258, + 442 + ], + "type": "text", + "content": "2.1 Continual Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 449, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 593 + ], + "type": "text", + "content": "Continual learning (CL), also known as lifelong learning, and its sub-field of incremental learning, are critical areas of machine learning that address a fundamental limitation of traditional machine learning models: Their rigidity in learning new tasks. CL is a learning paradigm that aims to train models that can learn new tasks and information without forgetting what was previously learned. This contrasts traditional machine learning, where models are typically trained on a single dataset and then remain fixed. Continual learning presents a challenge of balancing between learning new information and retaining the old [27]. Overemphasizing learning new information leads to forgetting the old while focusing too much on retaining old information hinders learning the new task. CL approaches can be grouped into three categories: Architecture-based, replay-based methods, and regularization-based methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Architecture-based approaches modify the architecture of neural networks to accommodate new tasks without overwriting previous information. These include methods such as parameter isolation [12, 24, 25], dynamically growing networks [1, 32, 39], or incorporation of additional components such as memory modules [2, 34]. These methods can face scalability challenges as network architecture grows, increasing memory and computational demands." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 242, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 242, + 91, + 448, + 102 + ], + "type": "text", + "content": "CLEO: Continual Learning of Evolving Ontologies" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "Replay-based approaches are of two types: Experience replay or rehearsal, involves storing a few instances from the previous task [3,6,23], which are used during training on a new task. Generative replay substitutes the memory buffer with a generative model to sample previous task instances when learning on a new task [15,33,35]. Challenges include privacy, data availability, memory and computational constraints in training generative models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 483, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 483, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 483, + 260 + ], + "type": "text", + "content": "Regularization-based approaches [16, 20, 41] strike a balance between stability and plasticity when learning new tasks by introducing a regularization term into the loss function. One of the approaches to regularization is to determine the importance of parameters to the previous tasks and penalize overwriting them. Distillation is another approach which preserves previously learned knowledge by distilling information from the old model to the new." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 276, + 331, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 276, + 331, + 289 + ], + "spans": [ + { + "bbox": [ + 131, + 276, + 331, + 289 + ], + "type": "text", + "content": "2.2 Continual Semantic Segmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 294, + 483, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 483, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 483, + 594 + ], + "type": "text", + "content": "Continual semantic segmentation (CSS) [5,37,38,40] is a specialized domain within CL that focuses on the task of semantic segmentation. Incremental Learning Techniques [28] (ILT) is a regularization-based method that distils knowledge from previously trained task models and additionally freezes the encoder for the current task, which acts as an additional regularization constraint. Modeling the Background [4] (MiB) addresses the problem of background shift in incremental segmentation. It proposes a novel distillation loss that accounts for the changing background between tasks. AWT [13] addresses the background shift in classifier initialization, and transfers the most relevant weights from the previous classifiers to the new classifier. In our baseline solution for CLEO, we will build on the ideas of MiB and extend it to model the semantic shift for all classes. PLOP [10] argues that treating the background class as a single class as done in MiB, which includes both previously encountered and novel classes can lead to catastrophic forgetting of the old classes as it learns new classes. PLOP addresses this problem by pseudolabeling old classes in the background. EWF [36] combines the weights of the old and new model trained with regularization to give a more balanced model. SDR [29] preserves latent representations of old classes through prototype matching and frees space for new classes through feature sparsity and contrastive learning. The decoder preserves previous knowledge via output-level distillation. RCIL [42] uses a representation compensation module to decouple the representation learning of old and new classes. This helps to prevent the model from forgetting old classes when learning new classes. REMINDER [31] proposes a class-similarity weighted distillation that distills knowledge from a previous model on old classes that are similar to the new classes. This reduces forgetting of old classes and improves learning of new classes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 610, + 261, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 610, + 261, + 623 + ], + "spans": [ + { + "bbox": [ + 131, + 610, + 261, + 623 + ], + "type": "text", + "content": "2.3 Evolving Ontologies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "An emerging area within incremental learning focuses on learning evolving classes, which involves the refinement of existing knowledge, where a broad initial grouping of classes evolves over time into more specific and nuanced classes. In contrast," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "type": "text", + "content": "S. Muralidhara et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 444, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 444, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 444, + 128 + ], + "type": "text", + "content": "in CIL, each task consists of a non-overlapping subsets of novel classes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 481, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 481, + 200 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 481, + 200 + ], + "type": "text", + "content": "Hierarchical Label Expansion [18] (HLE) for classification involves a model initially learning coarse classes, which then evolve into more fine-grained categories during each task across varying hierarchical levels. The expansion is restrictive, where the entire subset of classes at the current level expands in the next level during each task. CLEO differs from HLE in the following aspects:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 208, + 481, + 314 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 208, + 481, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 208, + 481, + 242 + ], + "spans": [ + { + "bbox": [ + 138, + 208, + 481, + 242 + ], + "type": "text", + "content": "- CLEO allows a more flexible evolution at any point within the sequence of tasks, in which certain classes can remain (partially) unexpanded or expand into sub-groups." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 244, + 481, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 244, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 137, + 244, + 481, + 289 + ], + "type": "text", + "content": "- The baseline solution for HLE is a rehearsal-based method that uses pseudolabeling from previous classifiers - one for each level of the hierarchy - for determining the importance of data. We present a regularization-based solution that uses knowledge distillation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 292, + 480, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 292, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 137, + 292, + 480, + 314 + ], + "type": "text", + "content": "- HLE considers image classification as application, while CLEO additionally deals with the more challenging problem of semantic segmentation." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 323, + 481, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 323, + 481, + 395 + ], + "spans": [ + { + "bbox": [ + 132, + 323, + 481, + 395 + ], + "type": "text", + "content": "Learning with Evolving Class Ontologies [22] (LECO) has been the first attempt to address evolving ontologies in semantic segmentation. However, despite its intended framing as a continual learning problem, LECO deviates from the defining principles of continual learning. We address these shortcomings and define this problem strictly within the bounds of continual learning. The primary distinctions between our framework and LECO are outlined below:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 406, + 479, + 535 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 137, + 406, + 479, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 406, + 479, + 427 + ], + "spans": [ + { + "bbox": [ + 137, + 406, + 479, + 427 + ], + "type": "text", + "content": "- LECO requires and stores all historical data, a constraint that contradicts the definition of CL [19], that all data is not available at once." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 430, + 479, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 430, + 479, + 463 + ], + "spans": [ + { + "bbox": [ + 137, + 430, + 479, + 463 + ], + "type": "text", + "content": "- LECO primarily revolves around the dilemma of whether to label new data or relabel the old data with the current label space. In contrast, CLEO confines itself to utilizing the input data and annotations within the dataset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 137, + 466, + 479, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 466, + 479, + 499 + ], + "spans": [ + { + "bbox": [ + 137, + 466, + 479, + 499 + ], + "type": "text", + "content": "- Due to the availability of historical data, LECO employs naive strategies such as fine-tuning, freezing, or joint training. We highlight the limitations of existing CIL methods and propose a baseline specifically for CLEO." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 137, + 502, + 479, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 502, + 479, + 535 + ], + "spans": [ + { + "bbox": [ + 137, + 502, + 479, + 535 + ], + "type": "text", + "content": "- CLEO explicitly addresses catastrophic forgetting, which is one of the core problems of continual learning. Since LECO retrans entirely using all the data, the problem of catastrophic forgetting does not even arise." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 133, + 556, + 192, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 556, + 192, + 568 + ], + "spans": [ + { + "bbox": [ + 133, + 556, + 192, + 568 + ], + "type": "text", + "content": "3 CLEO" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 582, + 481, + 665 + ], + "type": "text", + "content": "Deep learning has drawn inspiration from biological systems at various levels, from the way neurons form the fundamental building blocks of artificial neural networks to learning paradigms such as curriculum learning and reinforcement learning. Continual learning is another learning paradigm that aims to instill lifelong learning in intelligent systems. As discussed previously, current problems in continual learning involve adaptation to different domains, such as handling adverse conditions in autonomous driving or learning new classes over time. A" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "type": "text", + "content": "CLEO: Continual Learning of Evolving Ontologies" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "common characteristic in both these scenarios is that the task always represents new and previously unseen data. However, it is our contention that the scope of continual learning should not be limited solely to the discovery of novel classes. It should also include the enhancement and refinement of existing knowledge. This aligns closely with the way humans continually adapt and refine knowledge through neuroplasticity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 483, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 483, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 483, + 392 + ], + "type": "text", + "content": "In light of this perspective, we introduce CLEO - Continual Learning of Evolving Ontologies. CLEO is centered around the idea of removing the constraints of class incremental learning and allows classes that have a hierarchical relation (partially) overlap between tasks, thereby facilitating the refinement of already learned information. CLEO is a real and open-world problem and this is exemplified in the Mapillary Vistas dataset [30]. The initial version of the dataset defined 66 classes, which later evolved into 124 classes in the latest version. It is noteworthy to observe the transformation within class categorization. On one hand, classes that were previously categorized under unlabeled, such as traffic island emerge as novel classes. Conversely, several pre-existing classes undergo refinement, leading to the creation of new, more specific categories. In Fig. 2, we can observe a tangible example of this evolution, where the previous billboard class has evolved into more specialized classes, including signage classes for advertisements, information, stores, and differentiating between the front and back of the signs. In critical applications like autonomous driving, it becomes imperative for AI systems to swiftly adapt to these nuanced changes, perpetually learn, and enhance scene understanding precision." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "inline_equation", + "content": "T = \\{0,1,2,\\dots,n\\}" + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "text", + "content": " denote a sequence of " + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "inline_equation", + "content": "n + 1" + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "text", + "content": " tasks. Each task is associated with a specific subset of classes denoted as " + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "text", + "content": ". The union of all classes across tasks is " + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "inline_equation", + "content": "C = \\bigcup_{t=0}^{n} C_t" + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "text", + "content": ", representing the entire set of classes encountered during the learning process. At each task " + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "text", + "content": ", a subset of classes " + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "text", + "content": " is introduced incrementally similar to CIL. " + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "inline_equation", + "content": "C_{0:t} = C_0 \\cup \\ldots \\cup C_{t-1} \\cup C_t" + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "text", + "content": " is the set of all classes seen until step " + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "text", + "content": ". CLEO acknowledges the possibility of class splitting, where an existing class " + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "inline_equation", + "content": "c_{t-1}" + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "inline_equation", + "content": "C_{0:t-1}" + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "text", + "content": " can evolve into multiple new classes in " + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "text", + "content": ". Formally this can be expressed as, " + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "inline_equation", + "content": "C_t \\subset C" + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "inline_equation", + "content": "C_i \\cap C_j = \\phi" + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "text", + "content": " and the class splitting is expressed as " + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "inline_equation", + "content": "c_{t-1} \\in C_{0:t-1} \\to \\{c_{k+1}, c_{k+2}, \\ldots, c_{k+m}\\} \\subseteq C_t" + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "text", + "content": ". Note that the splitting of a class is not necessarily exhaustive, i.e., " + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "inline_equation", + "content": "c_{t-1} \\in C_{0:t-1}" + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "text", + "content": " may persist. An example for such an evolution of classes is given in Fig. 1. The learning objective of a task remains consistent with CIL: Train a model to recognize and differentiate the current subset of classes " + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "text", + "content": ", while ensuring that the model retains knowledge from the previously encountered classes in " + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "inline_equation", + "content": "C_{0:t-1}" + }, + { + "bbox": [ + 130, + 403, + 483, + 584 + ], + "type": "text", + "content": ". As in CIL, the model is not provided access to data from previous tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 602, + 388, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 602, + 388, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 602, + 388, + 616 + ], + "type": "text", + "content": "4 MoOn: Modelling Ontologies for CLEO" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Class-incremental learning involves the process of training a model to incrementally learn new classes. At each increment, the model is presented with images and the corresponding ground truth which contains annotations for the new classes" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "type": "text", + "content": "S. Muralidhara et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 246, + 199 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 199 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 199 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 199 + ], + "type": "image", + "image_path": "4b75010f392e681e944f2d94b7bc5d60b8963568403800a73d85ac0878de5d81.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 207, + 482, + 251 + ], + "lines": [ + { + "bbox": [ + 130, + 207, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 207, + 482, + 251 + ], + "type": "text", + "content": "Fig. 2: The Mapillary Vistas dataset [30] provides a real-world example for an evolving ontology. This figure compares ground truth annotations for the same image in two dataset versions. It illustrates the introduction of new classes, such as traffic island and the finer categorization of existing classes such as billboard." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 250, + 114, + 363, + 199 + ], + "blocks": [ + { + "bbox": [ + 250, + 114, + 363, + 199 + ], + "lines": [ + { + "bbox": [ + 250, + 114, + 363, + 199 + ], + "spans": [ + { + "bbox": [ + 250, + 114, + 363, + 199 + ], + "type": "image", + "image_path": "cc4e37b629060ca9560efa75e50c6895866c1e867dede7c8351b0bb16a904033.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 369, + 114, + 480, + 198 + ], + "blocks": [ + { + "bbox": [ + 369, + 114, + 480, + 198 + ], + "lines": [ + { + "bbox": [ + 369, + 114, + 480, + 198 + ], + "spans": [ + { + "bbox": [ + 369, + 114, + 480, + 198 + ], + "type": "image", + "image_path": "251045c1e490a729fb916f4266cdfdeadf7dc8be29e52a44850f2132129c3aea.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 275, + 482, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 482, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 482, + 404 + ], + "type": "text", + "content": "being learned. In class-incremental semantic segmentation (CISS), at each step, all unlabeled areas are marked as background. Thus in CISS, the phenomenon of semantic background shift occurs, because in each task, the meaning of the background class " + }, + { + "bbox": [ + 130, + 275, + 482, + 404 + ], + "type": "inline_equation", + "content": "(bg)" + }, + { + "bbox": [ + 130, + 275, + 482, + 404 + ], + "type": "text", + "content": " changes. Distillation loss is a technique of transferring knowledge from an old model " + }, + { + "bbox": [ + 130, + 275, + 482, + 404 + ], + "type": "inline_equation", + "content": "f_{t-1}" + }, + { + "bbox": [ + 130, + 275, + 482, + 404 + ], + "type": "text", + "content": " to a new model " + }, + { + "bbox": [ + 130, + 275, + 482, + 404 + ], + "type": "inline_equation", + "content": "f_t" + }, + { + "bbox": [ + 130, + 275, + 482, + 404 + ], + "type": "text", + "content": ", and this is achieved by comparing the old model's prediction to the new model's predictions. The distillation loss enforces the new model's predictions to resemble those of the old model by penalizing any deviations and this helps mitigate forgetting. Consider an image " + }, + { + "bbox": [ + 130, + 275, + 482, + 404 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 275, + 482, + 404 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 275, + 482, + 404 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 275, + 482, + 404 + ], + "type": "text", + "content": " pixels and the corresponding ground truth " + }, + { + "bbox": [ + 130, + 275, + 482, + 404 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 275, + 482, + 404 + ], + "type": "text", + "content": " at the current task " + }, + { + "bbox": [ + 130, + 275, + 482, + 404 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 275, + 482, + 404 + ], + "type": "text", + "content": " with the class set " + }, + { + "bbox": [ + 130, + 275, + 482, + 404 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 130, + 275, + 482, + 404 + ], + "type": "text", + "content": ". The standard distillation loss in [4] is defined as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 203, + 404, + 481, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 404, + 481, + 432 + ], + "spans": [ + { + "bbox": [ + 203, + 404, + 481, + 432 + ], + "type": "interline_equation", + "content": "\\ell_ {k d} ^ {t} (x, y) = - \\frac {1}{N} \\sum_ {i \\in x} \\sum_ {c \\in C _ {0: t - 1}} p _ {x} ^ {t - 1} (i, c) \\log \\hat {p} _ {x} ^ {t} (i, c) \\tag {1}", + "image_path": "e2f661c197c0d8a09798a36d7cd303eae626c31b9bf1f4ef5f1e6b5fc8e83fed.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 437, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 437, + 482, + 475 + ], + "type": "inline_equation", + "content": "p_x^{t-1}(i, c)" + }, + { + "bbox": [ + 130, + 437, + 482, + 475 + ], + "type": "text", + "content": " is the probability for class " + }, + { + "bbox": [ + 130, + 437, + 482, + 475 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 437, + 482, + 475 + ], + "type": "text", + "content": " in pixel " + }, + { + "bbox": [ + 130, + 437, + 482, + 475 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 437, + 482, + 475 + ], + "type": "text", + "content": " given by the model " + }, + { + "bbox": [ + 130, + 437, + 482, + 475 + ], + "type": "inline_equation", + "content": "f_{t-1}" + }, + { + "bbox": [ + 130, + 437, + 482, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 437, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\hat{p}_x^t(i, c)" + }, + { + "bbox": [ + 130, + 437, + 482, + 475 + ], + "type": "text", + "content": " is for class " + }, + { + "bbox": [ + 130, + 437, + 482, + 475 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 437, + 482, + 475 + ], + "type": "text", + "content": " in pixel " + }, + { + "bbox": [ + 130, + 437, + 482, + 475 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 437, + 482, + 475 + ], + "type": "text", + "content": " given by the model " + }, + { + "bbox": [ + 130, + 437, + 482, + 475 + ], + "type": "inline_equation", + "content": "f_t" + }, + { + "bbox": [ + 130, + 437, + 482, + 475 + ], + "type": "text", + "content": " but re-normalized across classes " + }, + { + "bbox": [ + 130, + 437, + 482, + 475 + ], + "type": "inline_equation", + "content": "C_{0:t-1}" + }, + { + "bbox": [ + 130, + 437, + 482, + 475 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 200, + 482, + 481, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 482, + 481, + 526 + ], + "spans": [ + { + "bbox": [ + 200, + 482, + 481, + 526 + ], + "type": "interline_equation", + "content": "\\hat {p} _ {x} ^ {t} (i, c) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} c \\in C _ {t} \\backslash \\{b g \\} \\\\ \\frac {p _ {x} ^ {t} (i , c)}{\\sum_ {k \\in C _ {0 : t - 1}} p _ {x} ^ {t} (i , k)} & \\text {i f} c \\in C _ {0: t - 1} \\end{array} \\right. \\tag {2}", + "image_path": "fb4f55cd96e49eeb6864c69c38c42efd50504400adc8010b0f556bf023b3a81e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Standard distillation ignores the dynamic nature of the background in incremental semantic segmentation. The background class in segmentation is a catch-all class that includes all previously seen classes and potential future classes. Furthermore, the background class is present in all tasks and changes as new classes are added, leading to background shift. A class " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "c \\in C_t" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " added during the current task was previously part of the background in " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "C_{t-1}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ". For a pixel representing a new class at the current task " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", the old model " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "f_{t-1}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " would predict the background class, and if standard distillation loss was applied naively, it would penalize the new model " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "f_t" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " for predicting the correct class. Modeling the Background (MiB) [4] proposes a novel distillation loss that explicitly addresses the background shift in incremental segmentation. MiB does not directly compare the prediction of the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 242, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 242, + 91, + 448, + 102 + ], + "type": "text", + "content": "CLEO: Continual Learning of Evolving Ontologies" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 117, + 479, + 266 + ], + "blocks": [ + { + "bbox": [ + 132, + 117, + 479, + 266 + ], + "lines": [ + { + "bbox": [ + 132, + 117, + 479, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 479, + 266 + ], + "type": "image", + "image_path": "28a6aea6c00ffa2273b88ce4ef87e7ca1ce375a812a52d46eb31df18b2c89f4a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 274, + 482, + 330 + ], + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 330 + ], + "type": "text", + "content": "Fig. 3: Modelling Ontologies (MoOn) generalizes the idea of MiB [4] for all classes. During distillation, forgetting is mitigated, by grouping class-logits according to the evolution of the class ontology. In this example, the teacher network of the previous task predicts the vehicle class, while the model tries to learn bicycle. To avoid penalizing this, the predictions for vehicle and bicycle are combined before the distillation loss." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "text", + "content": "new model " + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "text", + "content": " with the corresponding prediction of the old model " + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "inline_equation", + "content": "f_{t-1}" + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "text", + "content": ". Instead, the background prediction by " + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "inline_equation", + "content": "f_{t-1}" + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "text", + "content": " is compared with the " + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "text", + "content": " prediction of either a new class or the background. Eq. (3) modifies Eq. (2) by additionally considering the whole label space instead of limiting it to " + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "inline_equation", + "content": "C_{0:t-1}" + }, + { + "bbox": [ + 130, + 354, + 482, + 437 + ], + "type": "text", + "content": ". MiB acknowledges that pixels that were thought to be part of the background by the old model might actually be the new classes the current model is trying to learn, thus avoiding penalizing the probabilities of new classes being learned." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 223, + 445, + 482, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 445, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 223, + 445, + 482, + 478 + ], + "type": "interline_equation", + "content": "\\hat {p} _ {x} ^ {t} (i, c) = \\left\\{ \\begin{array}{l l} p _ {x} ^ {t} (i, c) & \\text {i f} c \\neq b g \\\\ \\sum_ {k \\in C _ {t}} p _ {x} ^ {t} (i, k) & \\text {i f} c = b g \\end{array} \\right. \\tag {3}", + "image_path": "b848438b4804cebdcf8ce0ca6379b8a47d0ab9d7962062061896bbdafe8b4a7a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 606 + ], + "type": "text", + "content": "MiB is designed for CIL, where the new classes being learned at the current step, were previously unseen and are always a part of the background class. Therefore, it would suffice to take into consideration the background class, and hence modeling the background. Whereas CLEO is designed for learning of evolving classes and the new classes being learned may have been previously unseen and part of the background or been part of a broader class. As seen in Fig. 3, the class bike being learned at the current step may have been previously learned as a vehicle class grouped with other classes like truck and bus. Therefore we need to take into consideration that within CLEO, new classes may not always have been a part of the background class previously." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "To this end, we propose Modelling Ontologies (MoOn) as a baseline solution for CLEO, which can be seen as a generalized version of MiB that accounts for the fact that new classes may emerge from any existing class, and not just the background class. Similar to how MiB considered the possibility of the new class" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "type": "text", + "content": "S. Muralidhara et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": "pixel belonging to the background class previously, MoOn additionally includes the possibility of the pixel belonging to an existing class. In MiB, the old model " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "f_{t-1}" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " may predict the background class for the new class being learned, whereas the old model in MoOn may predict an existing class " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "c \\in C_{0:t-1}" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": ". Therefore, the correct distillation would be to distill the knowledge from the old model " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "f_{t-1}" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " for the existing class " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "c \\in C_{0:t-1}" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " that is being split into new classes in " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": ". In the illustrated scenario depicted in Fig. 3, distillation would involve transferring knowledge from the old model regarding the vehicle class, while training the new model to predict the bike class. To achieve this, at each step " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": ", MoOn uses a set " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "S_t" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " of class maps " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "(c_{t-1} \\to \\{c_{k+1}, c_{k+2}, \\dots, c_{k+m}\\} = C_{new})" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " to identify the class " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "c_{t-1} \\in C_{0:t-1}" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " from which the current classes in " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " are split. If the new class " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "c_t" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " is completely new, then the superclass " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "c_{t-1}" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " in the class mapping would be the background class " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "bg" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": ". We define " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "C_{e,t}" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " as the set of all evolving classes " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "c_{t-1}" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "S_t" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": ". The modified distillation loss for learning evolving classes in CLEO is given as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 189, + 294, + 482, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 294, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 189, + 294, + 482, + 338 + ], + "type": "interline_equation", + "content": "\\hat {p} _ {x} ^ {t} (i, c) = \\left\\{ \\begin{array}{l l} p _ {x} ^ {t} (i, c) & \\text {i f} c \\notin C _ {e, t} \\\\ p _ {x} ^ {t} (i, c _ {t - 1}) + \\sum_ {c _ {n} \\in C _ {\\text {n e w}}} p _ {x} ^ {t} (i, c _ {n}) & \\text {i f} c \\in C _ {e, t} \\end{array} \\right. \\tag {4}", + "image_path": "ca3999016e6861cb4cdba2bfd8cdd5458dada5e19cf6f418fe74e0bb3542a061.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 355, + 205, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 355, + 205, + 366 + ], + "spans": [ + { + "bbox": [ + 132, + 355, + 205, + 366 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 376, + 481, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 481, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 481, + 400 + ], + "type": "text", + "content": "Towards the introduction of CLEO for semantic segmentation, we consider the following three diverse datasets:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 411, + 482, + 602 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 411, + 482, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 482, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 482, + 482 + ], + "type": "text", + "content": "- Cityscapes: Cityscapes [8] is a large-scale dataset of urban street scenes captured in 50 different cities under daytime and good weather conditions. The dataset provides annotations for 30 classes out of which 19 are used for evaluation and grouped into 7 categories. An overview of the hierarchy of the 19 semantic classes grouped by their parent categories is presented in the supplementary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 483, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 483, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 138, + 483, + 482, + 542 + ], + "type": "text", + "content": "- PASCAL VOC: PASCAL VOC 2012 [11] is part of the Visual Object Classes Challenge, which ran from 2005 to 2012. PASCAL VOC is a more general dataset like COCO [21] consisting of 20 categories, including vehicles, household objects, animals, and persons. The categorization by [43] of these classes along with intermediate groupings is shown in the supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 543, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 543, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 138, + 543, + 482, + 602 + ], + "type": "text", + "content": "- Mapillary Vistas: Mapillary Vistas [30] is a large-scale street-level image dataset for understanding street scenes around the world. It covers a wider range of geographic locations and diverse weather and illumination conditions. The initial version of the dataset provided annotations for 66 classes, which later nearly doubled into 124 classes in the current version v2.0." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 620, + 313, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 313, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 313, + 633 + ], + "type": "text", + "content": "4.2 Baselines and Implementation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "We evaluate our method's performance against standard CL baselines, namely fine-tuning and joint training. Fine-tuning involves using the previous step model" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 242, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 242, + 91, + 448, + 102 + ], + "type": "text", + "content": "CLEO: Continual Learning of Evolving Ontologies" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "as the starting point when learning on the current task. This is similar to transfer learning and even results in positive forward transfer. However, since there is no explicit attempt to prevent catastrophic forgetting, it involves the highest amount of forgetting and forms the lower baseline. The joint training model involves learning all of the tasks together in a single step, and as there is no incremental learning there is zero forgetting and this forms the theoretical upper baseline. Since in CLEO a redefinition of classes may occur because of the evolving ontology, we define the set of classes for the joint training as the final set of classes after the last task. Additionally, we compare against the state-of-the-art (SOTA) methods such as MiB [4], PLOP [10] and RCIL [42]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 247, + 483, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 483, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 483, + 357 + ], + "type": "text", + "content": "For MiB, and our proposed approach, we adopt the framework proposed by MiB [4]. In the case of PLOP and RCIL, which involve modifications specific to their approaches, we adopt the RCIL framework [42]. All these approaches use the DeepLabV3 [7] architecture as segmentation network, utilizing a ResNet-101 [14] backbone with an output stride of 16. The backbone's initialization uses a pre-trained model from ImageNet [9]. The training methodology utilizes SGD, with a learning rate of 0.01 for the initial learning step and 0.001 for subsequent steps. Across all learning steps, we train the model using a batch size of 24, for 50 epochs on Cityscapes and for 30 epochs on PASCAL and Mapillary Vistas." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 385, + 257, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 257, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 257, + 396 + ], + "type": "text", + "content": "4.3 Evaluation Metrics" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 416, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 482, + 536 + ], + "type": "text", + "content": "Mean Intersection-over-Union (mIoU) is a commonly used metric for evaluating semantic segmentation models. It measures the average overlap between the predicted segmentation and the ground truth segmentation for all classes. To calculate the mIoU, the Intersection-over-Union (IoU) is first calculated for each class, and then the IoU scores for all classes are averaged. The performance of CL models are often evaluated by measuring the amount of information forgotten by comparing the model's performance on old tasks against the joint training baseline. Within the context of CLEO, while learning evolving classes, forgetting can occur at multiple levels. We identify three types of class groups and analyze the mIoU results for each of them:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 552, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 552, + 481, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 552, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 552, + 481, + 589 + ], + "type": "text", + "content": "- Unsplit Classes: These classes are learned in the first task and are not further split. Their performance is expected to be largely unaffected, as they have not been split nor learned incrementally." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 591, + 481, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 591, + 481, + 627 + ], + "spans": [ + { + "bbox": [ + 138, + 591, + 481, + 627 + ], + "type": "text", + "content": "- Split Classes: These classes have been split either from the background or a previously learned superclass. The mIoU results for these split classes will provide insight into the model's ability to learn new classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 629, + 481, + 665 + ], + "type": "text", + "content": "- Retained Classes: These are classes remaining in the parent class after it has been split partially to yield new classes. Analyzing the mIoU scores for retained classes sheds light on how well the model retains knowledge." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "type": "text", + "content": "S. Muralidhara et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 136, + 462, + 236 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 125 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 125 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 125 + ], + "type": "text", + "content": "Table 1: Results for our two CLEO settings on Cityscapes [8] after learning all tasks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 136, + 462, + 236 + ], + "lines": [ + { + "bbox": [ + 151, + 136, + 462, + 236 + ], + "spans": [ + { + "bbox": [ + 151, + 136, + 462, + 236 + ], + "type": "table", + "html": "
MethodCS-Ex1CS-Ex2
UnsplitSplitRetainedAllUnsplitSplitRetainedAll
Fine-Tuning00.0000.0000.0000.0000.0014.4600.0008.68
Joint Training90.2661.1548.0455.6290.2661.1548.0455.62
MiB [4]90.2403.6607.9909.1190.2318.2511.3918.88
PLOP [10]87.5834.3613.0228.9186.0131.9115.5128.10
RCIL [42]89.8200.0007.9706.8889.0000.0007.7006.76
MoOn (Ours)90.2139.5037.0039.3190.2342.2835.7840.62
", + "image_path": "cd5075e0ec434b0ffc4b35acb4bc27934b859af1298b57cbe58f7e8ca0bb3012.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 142, + 269, + 471, + 358 + ], + "blocks": [ + { + "bbox": [ + 168, + 247, + 443, + 258 + ], + "lines": [ + { + "bbox": [ + 168, + 247, + 443, + 258 + ], + "spans": [ + { + "bbox": [ + 168, + 247, + 443, + 258 + ], + "type": "text", + "content": "Table 2: Task-wise results for CS-Ex2 after learning the final task." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 142, + 269, + 471, + 358 + ], + "lines": [ + { + "bbox": [ + 142, + 269, + 471, + 358 + ], + "spans": [ + { + "bbox": [ + 142, + 269, + 471, + 358 + ], + "type": "table", + "html": "
MethodTask 0Task 1Task 2Task 3Task 4Task 5Task 6All
Fine-Tuning00.0000.0000.0000.0000.0000.0034.7008.68
Joint Training47.5097.5763.0930.2088.3164.3157.9355.62
MiB [4]19.8202.8109.2600.1974.0045.6215.5418.88
PLOP [10]22.3895.0841.5900.0084.8153.8413.2128.10
RCIL [42]16.9000.0000.0000.0000.0000.0000.0006.76
MoOn (Ours)38.1196.8956.2609.1988.1660.9726.1040.62
", + "image_path": "389f3fe45d9b9b393b519fcf16ad3b36c663948fe88c49cf8f68e1ca0824d621.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 381, + 335, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 381, + 335, + 393 + ], + "spans": [ + { + "bbox": [ + 132, + 381, + 335, + 393 + ], + "type": "text", + "content": "4.4 Experimental Settings and Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 403, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 464 + ], + "type": "text", + "content": "We introduce seven initial experimental settings, that are designed for the previously introduced datasets and resemble the complexity and diversity of CLEO beyond classical CIL. For completeness and better understanding, we present the detailed semantic evolution of all settings by listing " + }, + { + "bbox": [ + 130, + 403, + 482, + 464 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 130, + 403, + 482, + 464 + ], + "type": "text", + "content": " for every task in the supplementary, including the class groups as introduced in Sec. 4.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 483, + 230, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 230, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 230, + 497 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 583 + ], + "type": "text", + "content": "Cityscapes From the Cityscapes class hierarchy, we derive two experimental settings. Both settings start in task 0 by learning the seven parent classes which together consist of the 19 grouped classes. The settings involve the incremental splitting of individual classes from parent classes. We do not always completely split a parent class, thus emphasizing not only learning the split class but also retaining the remaining information on the superclass." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 594, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 137, + 594, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 594, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 137, + 594, + 481, + 628 + ], + "type": "text", + "content": "- CS-Ex1: Each step involves splitting out a single class from each parent class. Since there is an unequal number of classes in the parent classes, we repeat this process until we have 19 individual classes, resulting in six tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 137, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 630, + 481, + 665 + ], + "type": "text", + "content": "- CS-Ex2: At each step, we split all but one of the classes from a single parent class to form individual classes. This results in 19 classes in six evolutionary increments (seven tasks in total) for six parent classes excluding sky." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 242, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 242, + 91, + 448, + 102 + ], + "type": "text", + "content": "CLEO: Continual Learning of Evolving Ontologies" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 136, + 481, + 217 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "type": "text", + "content": "Table 3: Results for our three CLEO settings on PASCAL [11] after learning all tasks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 136, + 481, + 217 + ], + "lines": [ + { + "bbox": [ + 132, + 136, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 136, + 481, + 217 + ], + "type": "table", + "html": "
MethodVOC-Ex1VOC-Ex2VOC-Ex3
UnsplitSplitRetainedAllUnsplitSplitRetainedAllUnsplitSplitRetainedAll
Fine-Tuning37.7736.3100.0029.6837.7405.3800.0007.8136.9125.6800.0023.08
Joint Training90.1972.5384.1276.9190.4873.7889.1477.7589.8475.2787.3678.38
MiB [4]88.9137.2854.9947.0586.9911.2837.8722.8487.0135.7533.2840.28
PLOP [10]86.8538.6743.5645.6285.7716.1803.6721.2683.4136.7406.6136.88
RCIL [42]88.0800.8441.1119.2985.2901.4323.8513.1887.4501.9918.6312.51
MoOn (Ours)88.9555.9868.8462.5284.4522.0825.4228.8283.3347.5142.3650.19
", + "image_path": "b3e401994a93a7c47cb67daafa8e7a74827b0b3404a7ceb5ea7b5014362f9ca3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "type": "text", + "content": "The results from these settings are presented in Tab. 1 and the task-wise results of " + }, + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "type": "inline_equation", + "content": "CS - Ex2" + }, + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "type": "text", + "content": " in Tab. 2. We can observe the shortcomings of existing CIL methods in handling evolving ontologies, whereas our baseline solution, MoOn, outperforms these methods. All methods achieve similar results on the unsplit classes, which is sky in both settings. However, when learning split classes, only PLOP comes close to our approach. Notably, MoOn surpasses PLOP by a significant margin in preserving information within the existing class. In the second task setting, the last step involves learning five vehicle classes, which are well-represented in the dataset. FT and MIB methods benefit from this, whereas PLOP and RCIL achieve similar results across both settings. We provide task-wise results of " + }, + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "type": "inline_equation", + "content": "CS - Ex1" + }, + { + "bbox": [ + 130, + 241, + 482, + 373 + ], + "type": "text", + "content": " and visualizations for MoOn in the supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 392, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 453 + ], + "type": "text", + "content": "PASCAL VOC From the PASCAL class hierarchy, we derive three experimental settings that represent the challenges of CLEO. In the first step (task 0), the model learns all the 20 classes grouped into the four parent classes namely animals, household, person, and vehicle. The person class does not contain any subclasses and therefore does not undergo splitting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 462, + 481, + 583 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 462, + 481, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 462, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 462, + 481, + 510 + ], + "type": "text", + "content": "- VOC-Ex1: This setting includes splitting into intermediate sub-groups. Task 1 includes learning the sub-groups Farmyard, furniture, and 2-wheeler which in turn contain individual classes. In the next and final step, these sub-groups are split into their individual classes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 510, + 481, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 481, + 547 + ], + "type": "text", + "content": "- VOC-Ex2: One class is split sequentially from each parent class resulting in five increments. At the end of the learning, we retain the individual classes cat, bottle, and the subgroup 4-wheeler in their respective parent classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 547, + 481, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 481, + 583 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 481, + 583 + ], + "type": "text", + "content": "- VOC-Ex3: This setting involves splitting all but one of the classes from each parent class sequentially, resulting in four tasks, with three evolutionary steps for the three parent classes excluding person which remains unsplit." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "The results on the three settings are shown in Tab. 3, while the corresponding visualizations for VOC-Ex3 are presented in Fig. 4. The performance on the unsplit classes, person class in this dataset, remains consistent across the CL methods. We observe the biggest difference in performance between our approach and others for VOC-Ex1. This is likely due to the more challenging multiple splitting of classes into intermediate subclasses. The lowest results for all methods" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "type": "text", + "content": "S. Muralidhara et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 480, + 232 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 480, + 232 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 480, + 232 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 480, + 232 + ], + "type": "image", + "image_path": "5aa9443b3e7ff40950ea72b79650161cff2c77f6119fb371201679b544c4e85d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 244, + 478, + 257 + ], + "lines": [ + { + "bbox": [ + 133, + 244, + 478, + 257 + ], + "spans": [ + { + "bbox": [ + 133, + 244, + 478, + 257 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparison on PASCAL VOC [11] VOC-Ex3 after the final task." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 151, + 289, + 462, + 388 + ], + "blocks": [ + { + "bbox": [ + 133, + 267, + 478, + 279 + ], + "lines": [ + { + "bbox": [ + 133, + 267, + 478, + 279 + ], + "spans": [ + { + "bbox": [ + 133, + 267, + 478, + 279 + ], + "type": "text", + "content": "Table 4: Results for CLEO settings on Mapillary Vistas [30] after learning all tasks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 289, + 462, + 388 + ], + "lines": [ + { + "bbox": [ + 151, + 289, + 462, + 388 + ], + "spans": [ + { + "bbox": [ + 151, + 289, + 462, + 388 + ], + "type": "table", + "html": "
MethodMV-Ex1MV-Ex2
UnsplitSplitRetainedAllUnsplitSplitRetainedAll
Fine-Tuning00.0018.5900.0008.6900.0001.4700.0000.69
Joint Training28.3722.6825.7925.5128.3722.6825.7925.51
MiB [4]27.0216.3317.7421.2709.8304.3100.0706.45
PLOP [10]27.0713.1418.4419.8603.5901.2600.0002.22
RCIL [42]28.5401.5712.7814.6506.3900.5500.1703.16
MoOn (Ours)26.7216.6719.4721.4409.8104.4000.0106.49
", + "image_path": "e5611b330ac910f39715e94885e1417843df583ab01355c7dca6aaa2e24e3aed.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 412, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 437 + ], + "type": "text", + "content": "are for VOC-Ex2, which involves a longer sequence of tasks, and learning sets of mixed classes one from each parent class at each step." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 453, + 482, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 526 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 526 + ], + "type": "text", + "content": "Mapillary Vistas Using the two dataset versions of Mapillary Vistas [30], we have defined two CLEO experiments. Both settings begin by learning the original 66 classes in task 0. The newly introduced classes are grouped into their ten parent classes from the previous version resulting in the original classes. The class mapping between the new classes and their parent classes is provided in the supplementary. The subsequent steps involve learning these new classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 536, + 482, + 619 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 138, + 536, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 482, + 583 + ], + "type": "text", + "content": "- MV-Ex1: After learning the original classes in task 0, all of the new classes are learned in task 1, resulting in two tasks. This exemplifies a real-world application of CLEO, where the model is specifically trained to learn the new classes without a complete retraining of all the classes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 584, + 482, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 584, + 482, + 619 + ], + "spans": [ + { + "bbox": [ + 138, + 584, + 482, + 619 + ], + "type": "text", + "content": "- MV-Ex2: This setting is more challenging with a longer sequence of tasks, which can exacerbate forgetting. In each step, one of the ten parent classes is split into its new classes, resulting in a total of eleven tasks." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Mapillary Vistas is challenging due to the large number of classes. As shown in Tab. 4, all methods exhibit suboptimal performance, including joint training which does not suffer from forgetting. The two settings are similar and have" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 242, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 242, + 91, + 448, + 102 + ], + "type": "text", + "content": "CLEO: Continual Learning of Evolving Ontologies" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": "learned the same classes after the final task. They differ in the number of steps, highlighting the effect of a longer sequence of tasks. " + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "MV - Ex2" + }, + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": " with a longer sequence significantly impacts the performance of all methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 170, + 328, + 183 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 170, + 328, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 170, + 328, + 183 + ], + "type": "text", + "content": "6 Limitations and Future Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 195, + 481, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 481, + 303 + ], + "type": "text", + "content": "In its most general form, CLEO allows for merging of two or more known classes as a valid form of evolution. However, we note that the here presented formalism of CLEO only considers the separation of known parent classes. Yet, in some rare cases, it might be required to combine previous information to simplify previous tasks and free representational capacity of a model. This, however, as a version of intended forgetting opens up a mostly unexplored field of research in continual learning and is left for future work. Besides, semantic context can be achieved through post-processing by combining predictions of existing and new classes into a single class, preventing interference and avoiding further forgetting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 304, + 481, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 423 + ], + "type": "text", + "content": "We further highlight that our baseline solution MoOn requires the information of class relations " + }, + { + "bbox": [ + 132, + 304, + 481, + 423 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 132, + 304, + 481, + 423 + ], + "type": "text", + "content": " about the evolution of classes. While we believe that this information is usually available to a sequentially trained model, there might be cases in which explicit knowledge about the ontology is unavailable. In such cases, " + }, + { + "bbox": [ + 132, + 304, + 481, + 423 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 132, + 304, + 481, + 423 + ], + "type": "text", + "content": " could be inferred from the existing model's predictions on the current task data. If the model predicts an existing class, the class being learned was previously seen and is split from this parent class. If it predicts the background, the new class is entirely new and unseen. In our experimental evaluation, we found that MoOn and other approaches notably underperform with longer sequences in CLEO, highlighting the need for future attention to long-term learning settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 441, + 219, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 219, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 219, + 453 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 466, + 481, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 466, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 466, + 481, + 597 + ], + "type": "text", + "content": "We introduce a novel framework for addressing the challenge of adapting to evolving classes, termed Continual Learning of Evolving Ontologies (CLEO). The motivation for CLEO stems from a practical scenario wherein pre-existing classes undergo transformations over time, resulting in more specialized classes. CLEO extends the functionality of CIL to not just learn new information but also refine existing information. The applicability of CLEO in the real world is exemplified through the Mapillary Vistas dataset, where an existing set of classes evolves into more detailed classes. Furthermore, we highlight the shortcomings of existing CIL methods and propose MoOn as an initial solution for CLEO. The effectiveness of MoOn for CLEO is demonstrated on seven experimental settings across three datasets, surpassing previous CIL approaches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 616, + 240, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 616, + 240, + 630 + ], + "spans": [ + { + "bbox": [ + 132, + 616, + 240, + 630 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "content": "This work was partially funded by the Federal Ministry of Education and Research Germany under the project DECODE (01IW21001)." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 252, + 101 + ], + "type": "text", + "content": "S. Muralidhara et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 142, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 142, + 480, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 480, + 175 + ], + "type": "text", + "content": "1. Aljundi, R., Chakravarty, P., Tuytelaars, T.: Expert gate: Lifelong learning with a network of experts. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 176, + 480, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 480, + 197 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 480, + 197 + ], + "type": "text", + "content": "2. Asghar, N., Mou, L., Selby, K.A., Pantasdo, K.D., Poupart, P., Jiang, X.: Progressive memory banks for incremental domain adaptation. arXiv (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 198, + 480, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 198, + 480, + 231 + ], + "spans": [ + { + "bbox": [ + 138, + 198, + 480, + 231 + ], + "type": "text", + "content": "3. Bang, J., Kim, H., Yoo, Y., Ha, J.W., Choi, J.: Rainbow memory: Continual learning with a memory of diverse samples. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 232, + 481, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 481, + 264 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 481, + 264 + ], + "type": "text", + "content": "4. Cermelli, F., Mancini, M., Bulo, S.R., Ricci, E., Caputo, B.: Modeling the background for incremental learning in semantic segmentation. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 266, + 480, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 480, + 297 + ], + "type": "text", + "content": "5. Cha, S., kim, b., Yoo, Y., Moon, T.: Ssul: Semantic segmentation with unknown label for exemplar-based class-incremental learning. In: Advances in Neural Information Processing Systems (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 299, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 481, + 319 + ], + "type": "text", + "content": "6. Chaudhry, A., Rohrbach, M., Elhoseiny, M., Ajanthan, T., Dokania, P.K., Torr, P.H., Ranzato, M.: On tiny episodic memories in continual learning. arXiv (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 322, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 322, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 322, + 480, + 342 + ], + "type": "text", + "content": "7. Chen, L.C., Papandreou, G., Schroff, F., Adam, H.: Rethinking atrous convolution for semantic image segmentation. arXiv (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 343, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 386 + ], + "type": "text", + "content": "8. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 388, + 480, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 480, + 420 + ], + "type": "text", + "content": "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2009)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 422, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 422, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 134, + 422, + 480, + 453 + ], + "type": "text", + "content": "10. Douillard, A., Chen, Y., Dapogny, A., Cord, M.: Plop: Learning without forgetting for continual semantic segmentation. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 455, + 480, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 455, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 134, + 455, + 480, + 486 + ], + "type": "text", + "content": "11. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International Journal of Computer Vision (IJCV) (2010)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 487, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 487, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 134, + 487, + 480, + 520 + ], + "type": "text", + "content": "12. Fernando, C., Banarse, D., Blundell, C., Zwols, Y., Ha, D., Rusu, A.A., Pritzel, A., Wierstra, D.: Pathnet: Evolution channels gradient descent in super neural networks. arXiv (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 522, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 522, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 134, + 522, + 480, + 553 + ], + "type": "text", + "content": "13. Goswami, D., Schuster, R., van de Weijer, J., Stricker, D.: Attribution-aware weight transfer: A warm-start initialization for class-incremental semantic segmentation In: Winter Conference on Applications of Computer Vision (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 555, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 555, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 134, + 555, + 480, + 576 + ], + "type": "text", + "content": "14. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition In: Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 578, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 578, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 134, + 578, + 480, + 597 + ], + "type": "text", + "content": "15. Kemker, R., Kanan, C.: Fearnet: Brain-inspired model for incremental learning. In: International Conference on Learning Representations (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 600, + 480, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 480, + 642 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 480, + 642 + ], + "type": "text", + "content": "16. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "text", + "content": "17. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "type": "text", + "content": "CLEO: Continual Learning of Evolving Ontologies" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 138 + ], + "type": "text", + "content": "18. Lee, B.H., Jung, O., Choi, J., Chun, S.Y.: Online continual learning on hierarchical label expansion. In: International Conference on Computer Vision (ICCV) (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 481, + 171 + ], + "type": "text", + "content": "19. Lesort, T., Lomonaco, V., Stoian, A., Maltoni, D., Filliat, D., Diaz-Rodriguez, N.: Continual learning for robotics: Definition, framework, learning strategies, opportunities and challenges. Information fusion (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 481, + 192 + ], + "type": "text", + "content": "20. Li, Z., Hoiem, D.: Learning without forgetting. Transactions on Pattern Analysis and Machine Intelligence (T-PAMI) (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 481, + 224 + ], + "type": "text", + "content": "21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dóllar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision (ECCV) (2014)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 481, + 257 + ], + "type": "text", + "content": "22. Lin, Z., Pathak, D., Wang, Y.X., Ramanan, D., Kong, S.: Continual learning with evolving class ontologies. In: Conference on Neural Information Processing Systems (NeurIPS) (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 257, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 481, + 278 + ], + "type": "text", + "content": "23. Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. In: Conference on Neural Information Processing Systems (NeurIPS) (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 278, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 481, + 311 + ], + "type": "text", + "content": "24. Mallya, A., Davis, D., Lazebnik, S.: Piggyback: Adapting a single network to multiple tasks by learning to mask weights. In: European Conference on Computer Vision (ECCV) (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 311, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 481, + 343 + ], + "type": "text", + "content": "25. Mallya, A., Lazebnik, S.: Packnet: Adding multiple tasks to a single network by iterative pruning. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 343, + 481, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 481, + 364 + ], + "type": "text", + "content": "26. McCloskey, M., Cohen, N.J.: Catastrophic interference in connectionist networks: The sequential learning problem. Psychology of Learning and Motivation (1989)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "type": "text", + "content": "27. Mermillod, M., Bugaiska, A., Bonin, P.: The stability-plasticity dilemma: Investigating the continuum from catastrophic forgetting to age-limited learning effects (2013)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 429 + ], + "type": "text", + "content": "28. Michieli, U., Zanuttigh, P.: Incremental learning techniques for semantic segmentation. In: Conference on Computer Vision and Pattern Recognition Workshops (CVPR-W) (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 429, + 481, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 481, + 461 + ], + "type": "text", + "content": "29. Michieli, U., Zanuttigh, P.: Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 461, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 481, + 493 + ], + "type": "text", + "content": "30. Neuhold, G., Ollmann, T., Rota Bulo, S., Kontschieder, P.: The mapillary vistas dataset for semantic understanding of street scenes. In: International Conference on Computer Vision (ICCV) (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "type": "text", + "content": "31. Phan, M.H., Ta, T.A., Phung, S.L., Tran-Thanh, L., Bouzerdoum, A.: Class similarity weighted knowledge distillation for continual semantic segmentation. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "text", + "content": "32. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv (2016)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 558, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 481, + 579 + ], + "type": "text", + "content": "33. Shin, H., Lee, J.K., Kim, J., Kim, J.: Continual learning with deep generative replay. In: Conference on Neural Information Processing Systems (NeurIPS) (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 579, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 481, + 611 + ], + "type": "text", + "content": "34. Sprechmann, P., Jayakumar, S.M., Rae, J.W., Pritzel, A., Badia, A.P., Uria, B., Vinyals, O., Hassabis, D., Pascanu, R., Blundell, C.: Memory-based parameter adaptation. In: International Conference on Learning Representations (ICLR) (2018)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "text", + "content": "35. Van de Ven, G.M., Siegelmann, H.T., Tolias, A.S.: Brain-inspired replay for continual learning with artificial neural networks. Nature communications (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "36. Xiao, J.W., Zhang, C.B., Feng, J., Liu, X., van de Weijer, J., Cheng, M.M.: Endpoints weight fusion for class incremental semantic segmentation. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 251, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 251, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 251, + 100 + ], + "type": "text", + "content": "S. Muralidhara et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 315 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "37. Yang, G., Fini, E., Xu, D., Rota, P., Ding, M., Nabi, M., Alameda-Pineda, X., Ricci, E.: Uncertainty-aware contrastive distillation for incremental semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "38. Yang, G., Fini, E., Xu, D., Rota, P., Ding, M., Tang, H., Alameda-Pineda, X., Ricci, E.: Continual attentive fusion for incremental learning in semantic segmentation. IEEE Transactions on Multimedia (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "type": "text", + "content": "39. Yoon, J., Yang, E., Lee, J., Hwang, S.J.: Lifelong learning with dynamically expandable networks. arXiv (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 226 + ], + "type": "text", + "content": "40. Yu, L., Liu, X., Van de Weijer, J.: Self-training for class-incremental semantic segmentation. IEEE Transactions on Neural Networks and Learning Systems (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 227, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 227, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 482, + 248 + ], + "type": "text", + "content": "41. Zenke, F., Poole, B., Ganguli, S.: Continual learning through synaptic intelligence. In: International Conference on Machine Learning (ICML) (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 249, + 482, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 282 + ], + "type": "text", + "content": "42. Zhang, C.B., Xiao, J.W., Liu, X., Chen, Y.C., Cheng, M.M.: Representation compensation networks for continual semantic segmentation. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 282, + 482, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 315 + ], + "type": "text", + "content": "43. Zhao, H., Hu, Q., Zhu, P., Wang, Y., Wang, P.: A recursive regularization based feature selection framework for hierarchical classification. Transactions on Knowledge and Data Engineering (2021)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 242, + 91, + 447, + 102 + ], + "type": "text", + "content": "CLEO: Continual Learning of Evolving Ontologies" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CLIFF_ Continual Latent Diffusion for Open-Vocabulary Object Detection/51812c3e-19bd-4e17-9855-dbf65e0d5287_content_list.json b/2024/CLIFF_ Continual Latent Diffusion for Open-Vocabulary Object Detection/51812c3e-19bd-4e17-9855-dbf65e0d5287_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..be72ae1a25484b1ee159625f76674b9d583cbc00 --- /dev/null +++ b/2024/CLIFF_ Continual Latent Diffusion for Open-Vocabulary Object Detection/51812c3e-19bd-4e17-9855-dbf65e0d5287_content_list.json @@ -0,0 +1,1875 @@ +[ + { + "type": "text", + "text": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection", + "text_level": 1, + "bbox": [ + 282, + 140, + 722, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wuyang $\\mathrm{Li}^{1}$ , Xinyu Liu, Jiayi Ma2, and Yixuan Yuan", + "bbox": [ + 267, + 210, + 735, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The Chinese University of Hong Kong \n2 Wuhan University \nwymanbest@outlook.com, yxyuan@ee.cuhk.edu.hk", + "bbox": [ + 331, + 238, + 669, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Open-vocabulary object detection (OVD) utilizes image-level cues to expand the linguistic space of region proposals, thereby facilitating the detection of diverse novel classes. Recent works adapt CLIP embedding by minimizing the object-image and object-text discrepancy combinatorially in a discriminative paradigm. However, they ignore the underlying distribution and the disagreement between the image and text objective, leading to the misaligned distribution between the vision and language sub-space. To address the deficiency, we explore the advanced generative paradigm with distribution perception and propose a novel framework based on the diffusion model, coined Continual Latent Diffusion (CLIFF), which formulates a continual distribution transfer among the object, image, and text latent space probabilistically. CLIFF consists of a Variational Latent Sampler (VLS) enabling the probabilistic modeling and a Continual Diffusion Module (CDM) for the distribution transfer. Specifically, in VLS, we first establish a probabilistic object space with region proposals by estimating distribution parameters. Then, the object-centric noise is sampled from the estimated distribution to generate text embedding for OVD. To achieve this generation process, CDM conducts a short-distance object-to-image diffusion from the sampled noise to generate image embedding as the medium, which guides the long-distance diffusion to generate text embedding. Extensive experiments verify that CLIFF can significantly surpass state-of-the-art methods on benchmarks. The code is available at https://github.com/CUHK-AIM-Group/CLIFF.", + "bbox": [ + 261, + 316, + 743, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Open-vocabulary object detection $\\cdot$ Diffusion model $\\cdot$ Continual distribution transfer", + "bbox": [ + 261, + 648, + 743, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 703, + 375, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object detection [20, 37, 69] has made significant strides in recent years, as demonstrated by the impressive results on benchmarks [17, 47]. Nonetheless, due to the unsatisfactory closed-set constraint [77, 82], object detectors are limited to discovering the novel-class objects [38, 39] unseen during training, impeding a reliable scene understanding in the real world. Further, it hinders the journey toward achieving true artificial intelligence since it requires the human-like perceptual ability to recognize unexpected objects appearing everywhere.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3e1a0e4c75f05381ad71649270fa6c534b2bcd960881737f6641d1433feecc9a.jpg", + "image_caption": [ + "(a) Existing OVD works", + "(b) The proposed CLIFF framework modeling a continual distribution transfer", + "Fig. 1: Illustration of (a) existing discriminative paradigm [15,68,75] with a combinatorial objective (object→image and object→text in parallel), and (b) the proposed method conducting a continual diffusion (object→image→text) with generative nature." + ], + "image_footnote": [], + "bbox": [ + 218, + 142, + 784, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To break through this barrier, open-vocabulary object detection (OVD) has been studied that leverages auxiliary linguistic cues to generalize the detection of novel classes [82]. Recent advances adapt CLIP [67], a vision-language (VL) foundation model with strong zero-shot capacity, to boost the novel-class generalization of the object detector. Some works [13, 14] utilize the CLIP prediction (image-level) to generate pseudo labels at the object level, enabling the self-training of the object detector. To fully explore the hidden knowledge, some other works strive to align the object-level representation with CLIP embeddings in the hidden space. During training the object detector, they minimize the object-image [15, 68, 75] and object-text [46, 68, 76] discrepancy to align the sub-space among the object, image, and text modalities, thereby unleashing the novel-class generalization capacity from the pre-trained CLIP model.", + "bbox": [ + 212, + 328, + 787, + 508 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the significant success, the existing OVD paradigm, as depicted in Fig. 1(a), has two limitations regarding the sub-space alignment. First, existing works utilize text embedding as the classifier weight to explicitly optimize the discriminative class separation, which is agnostic to the underlying distribution and may lead to class bias [18]. This distribution-agnostic property results in the distribution misalignment [40, 42, 46] and severely hinders the novel-class generalization [78]. Besides, due to the uncalibrated latent space in discriminative models [45], e.g., numerical high-variance, the object distribution is prone to overfit to the labeled classes, worsening the misalignment with the unbiased CLIP distribution. With the rise of generative paradigm [2], this barrier has been gradually broken via diverse probabilistic modeling [18, 45, 59]. Hence, we aim to incorporate probabilistic advantages into discriminative detection pipelines, which can facilitate the distribution transfer across the object and text sub-spaces, thereby achieving an unbiased novel-class perception [45, 59].", + "bbox": [ + 212, + 510, + 787, + 720 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, some existing works minimize the object-image and object-text discrepancy simultaneously in a combinatorial manner, ignoring the disagreement between the image and text two objectives3. With this combinatorial nature, optimizing object-image alignment as the same role as the object-text counterpart is suboptimal due to the potential conflict between the two objectives [25].", + "bbox": [ + 212, + 722, + 789, + 799 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 326, + 126 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "3 As the first attempt, we discover the ever-overlooked disagreement issue and justify it in Appendix via experiments.", + "bbox": [ + 217, + 810, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Intuitively, the image and object share visual representation, while the image and text are coupled with CLIP pre-training, which motivates us to introduce the image as an interpolated sub-space [61] between the object and text counterpart, yielding a continual distribution transfer among three modalities.", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the above issues, we propose Continual Latent Diffusion (CLIFF) for OVD, as shown in Fig. 1(b). CLIFF breaks through the discriminative bottleneck via a probabilistic Variational Latent Sampler (VLS) and then formulates a diffusion-based distribution transfer [21] in a Continual Diffusion Module (CDM) to avoid the combinatorial limitation. Specifically, VLS leverages variational approximation [24] to establish a probabilistic object space with learned distribution parameters, and then samples object-centric noise to boost the diffusion process. CDM formulates the continual diffusion process to transfer the object distribution to the image and text counterpart, avoiding disagreement in the existing combinatorial objective. Besides, different from existing diffusion-based works [27, 78] with unsatisfactory efficiency, CLIFF achieves comparable and even better model efficiency compared with state-of-the-art OVD works, which is verified in Appendix. To be summarized, the contributions are four-fold:", + "bbox": [ + 212, + 207, + 789, + 402 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Continual Latent Diffusion (CLIFF) for OVD, which achieves a continual distribution transfer with DDPM-based generation. To the best of our knowledge, this work represents the first attempt to study latent diffusion in VL discriminative tasks and push forward benchmark results.", + "- Variational Latent Sampler is proposed to generate a probabilistic object space with variational modeling, in which the object-centric noise can be sampled with reparameterization to boost the generation.", + "- We design an efficient Continual Diffusion Module to facilitate the distribution transfer among three cross-modality sub-spaces, namely the object, image, and text space, using a new diffusion module and unified DDPM formulation.", + "- CLIFF is extensively validated through different benchmarks, which verify its state-of-the-art role and show great adaptability to various detectors." + ], + "bbox": [ + 225, + 414, + 785, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 214, + 618, + 397, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Open-Vocabulary Object Detection", + "text_level": 1, + "bbox": [ + 214, + 648, + 552, + 665 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Open-vocabulary object detection (OVD) leverages auxiliary image-level labels with linguistic cues, e.g., classification [68] and caption datasets [80], to enlarge the semantic space of the region proposals [5, 69], thereby facilitating the novel-class generalization for the real-world applications, such as automatic driving [41, 43, 50, 52], medical imaging analysis [31, 35, 36, 49] and diagnostic decision-making [1, 7, 44, 53, 79]. Zareian et al. [82] formulate the standard OVD problem setting and propose OVR-CNN to project the visual space to the linguistic counterpart on Faster-RCNN [69]. Some other works [22, 23, 60] delve into OVD with Vision Transformer (ViT) by designing large-scale region-aware pretraining [23], feature masking [22], which fully unleash the potential of ViT in fine-grained detection tasks. Recent works adapt CLIP [67] to objects detectors for the novel-class", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection", + "bbox": [ + 233, + 113, + 730, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "generalization. The other works adapt the CLIP-encoded latent embedding and aim to align three critical sub-spaces: the object space, the CLIP image space, and the CLIP text space. They bridge the gap between the object and text sub-space via region-text contrastive pre-training [85], region-word bipartite matching [46], CLIP-based conditional matching [81], and learnable detection-friendly prompts [9, 11, 12]. PromptDet [12] and DetPro [9] use learnable text prompts to learn with noisy uncurated web images, thereby enhancing the object-text alignment. Additionally, recent advances [15, 68, 75] adapt the image sub-space with the embedding distillation [15], weight transfer [68] and bags of regions [75], to fully use CLIP embedding, serving as a combinatorial objective with the object-text alignment. Existing works follow a similar discriminative paradigm for OVD, which cannot model the underlying distribution with a suboptimal distribution alignment. Differently, we formulate a continual transfer in a unified denoising diffusion module, enabling the distribution-level perception and adaptation among the object, image, and text sub-spaces.", + "bbox": [ + 212, + 146, + 787, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Denoising Diffusion Probabilistic Model", + "text_level": 1, + "bbox": [ + 214, + 395, + 591, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As the advanced generative paradigm, Denoising Diffusion Probabilistic Model (DDPM) [21] (\"diffusion model\" for short) generates samples by formulating a gradual denoising process. This process can be treated as a distribution transfer [51] from the normal distribution to the data in a flow-like manner [51, 54, 84]. Instead of diffusing in the data space, LDM [70] and VDM [66] deploy the denoising diffusion in the latent space to enable the high-resolution synthesis. Recent works can be generally categorized into faster sampling [55, 73], noise corruption designs [3, 62], and diffusion model architecture [33, 34, 65], etc. The diffusion model has achieved remarkable generative performance in the image, text, video, audio, graph, and motion generation [21, 28, 30, 32, 74] with different diffusion conditions for downstream applications.", + "bbox": [ + 212, + 420, + 787, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to generative tasks, recent works [6,16,27,56,78,83] have discovered that the diffusion feature has adequate semantic discriminability with richer distribution-level perception and data-efficient property [4,64] than conventional discriminative models [67]. Without fine-tuning, an off-the-shelled diffusion model can achieve satisfactory zero-shot classification [29]. Along this trend, some bottlenecks in discriminative models are gradually broken by involving the generative paradigm [78]. CARD [19] uses the diffusion model to estimate the uncertainty to improve image classification. Unlike existing works considering the noise $\\rightarrow$ data diffusion, we formulate a continual distribution transfer among three modalities, yielding the object $\\rightarrow$ image $\\rightarrow$ text transfer in the latent space.", + "bbox": [ + 212, + 587, + 787, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 214, + 761, + 382, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Problem Formulation. Assuming a set of base classes $\\mathcal{C}_B$ and novel classes $\\mathcal{C}_N$ , we have a detection dataset $\\mathcal{D}_{det} = \\{I_i,O_i\\}_{i = 1}^n$ with object-level labels $\\{O_i\\}_{i = 1}^n$ in $\\mathcal{C}_B$ and another auxiliary dataset $\\mathcal{D}_{aux} = \\{I_i,Y_i\\}_{i = 1}^{n'}$ with image-level", + "bbox": [ + 214, + 792, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 325, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Reverse Diffusion Process (Denoising): RevDif $(\\mathbf{x}_T,\\mathbf{c},L)$", + "bbox": [ + 215, + 146, + 723, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Require: $\\mathbf{x}_T$ : initial noise; c: diffusion condition; $L$ : the number reverse steps.", + "bbox": [ + 215, + 166, + 740, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ensure: $\\mathbf{x}_{T - L}$ generated samples with $L$ steps.", + "bbox": [ + 215, + 181, + 542, + 194 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: for $t = T, \\dots, T - L + 1$ do", + "2: $\\mathbf{z} \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$", + "3: $\\mathbf{x}_{t - 1} = \\frac{1}{\\sqrt{\\alpha_t}}\\left(\\mathbf{x}_t - \\frac{1 - \\alpha_t}{\\sqrt{1 - \\bar{\\alpha}_t}}\\pmb {\\epsilon}_\\theta (\\mathbf{x}_t,\\mathbf{c},t)\\right) + \\sigma_t\\mathbf{z}$", + "4: end for", + "5: return $\\mathbf{x}_{T - L}$" + ], + "bbox": [ + 222, + 195, + 545, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "labels, e.g., the image caption [21] and classification dataset [68], which has the weakly supervised labels in $\\mathcal{C}_B \\cup \\mathcal{C}_N$ . In $\\mathcal{D}_{det}$ , each image $I_i$ is associated with a set of objects $O_i = \\{(x,y,w,h), \\hat{y}\\}$ , including the bounding-box coordinates $\\{(x,y,w,h)\\}$ and class labels $\\{\\hat{y}\\} \\subset \\mathcal{C}_B$ . OVD aims to use $\\mathcal{D}_{det}$ and $\\mathcal{D}_{aux}$ for training and detect base and novel objects in $\\mathcal{C}_B \\cup \\mathcal{C}_N$ during the test stage.", + "bbox": [ + 212, + 297, + 785, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Diffusion Model. Formulating a generative paradigm with the Markov chain [21, 63], DDPM is able to generate data $\\mathbf{x}_0$ from the noise $\\mathbf{x}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ via a $T$ -step denoising process. Concretely, DDPM begins with the clean data $\\mathbf{x}_0$ and gradually adds Gaussian noise using the fixed forward process $q(\\mathbf{x}_t | \\mathbf{x}_{t-1})$ , where $t = 1, 2, \\dots, T$ . This process can be denoted as follows,", + "bbox": [ + 212, + 375, + 787, + 450 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 455, + 787, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_t$ is noise schedule [62]. To recover data $\\mathbf{x}_0$ from noise $\\mathbf{x}_T$ , the reverse process parameterized with $\\theta$ is learned to denoise inputs with condition $\\mathbf{c}$ :", + "bbox": [ + 212, + 479, + 785, + 510 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\mathbf {x} _ {0: T} \\mid \\mathbf {c}\\right) = p \\left(\\mathbf {x} _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}, \\mathbf {c}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 516, + 785, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p(\\mathbf{x}_T)$ is fixed to Gaussian. In this reverse step, the diffusion model typically learns a neural network $\\epsilon_{\\theta}(\\cdot)$ to generate the values for computing the mean of $p_{\\theta}(\\mathbf{x}_{t - 1}\\mid \\mathbf{x}_t,\\mathbf{c})$ given by the noisy input $\\mathbf{x}_t$ , which is optimized to minimize the variational lower bound (ELBO). As proved in [21, 29], this training objective can be re-written as:", + "bbox": [ + 212, + 563, + 787, + 637 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n- \\mathbb {E} _ {t, \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, \\mathbf {c}, t) \\| ^ {2} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 379, + 641, + 785, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the diffusion model $\\epsilon_{\\theta}(\\cdot)$ can estimate the added noise of each noisy input $\\mathbf{x}_t$ . After optimizing with Eq. 3, the diffusion model can generate a data sample $\\mathbf{x}_{T - L}$ at arbitrary time-step $T - L$ from a random noise $\\mathbf{x}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ with $L$ steps of the reverse diffusion, optionally guided by the condition $\\mathbf{c}$ . This process [21] is represented in Algorithm 1. In the following sections, we use RevDif( $\\cdot$ ) to represent the reverse diffusion (noise→data) for simplicity.", + "bbox": [ + 212, + 672, + 787, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Methodology", + "text_level": 1, + "bbox": [ + 214, + 781, + 380, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall workflow is illustrated in Fig. 2. We feed input images $\\{I_i\\}_{i=1}^B$ to the backbone and RPN [69] to extract $N$ region proposals, which are sent to the", + "bbox": [ + 212, + 808, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection", + "bbox": [ + 233, + 113, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fc393f73c102897fa4a24c4fe58915cd200cc74ce2ebe8e250eff65f1273dbba.jpg", + "image_caption": [ + "Fig. 2: Illustration of the proposed CLIFF. VLS establishes a probabilistic embedding space with region proposals, where we sampled object-centric noise $\\mathbf{x}_T$ . Then, CDM achieves a continual distribution transfer from the object to the image and text." + ], + "image_footnote": [], + "bbox": [ + 217, + 148, + 784, + 262 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "feature extractor to generate object embedding $\\mathbf{x}_{\\mathrm{obj}}$ . Then, VLS, parameterized with $\\phi$ , establishes and estimates the mean and variance of a probabilistic object space $p_{\\phi}(\\mathbf{x}_T|\\mathbf{x}_{\\mathrm{obj}})$ , where the reparameterization is conducted to sample object-centric noise $\\mathbf{x}_T$ , as shown in Fig. 2(a). After that, the sampled noise $\\mathbf{x}_T$ is sent to CDM, denoted as $\\epsilon_{\\theta}(\\cdot)$ , to generate text embedding $\\mathbf{x}_{\\mathrm{txt}}$ with diffusion, as depicted in Fig. 2(b). Conditioned on $\\mathbf{c}_{o2i}$ , we first conduct object-to-image diffusion to generate the image embedding $\\mathbf{x}_{\\mathrm{img}}$ with $L$ steps (top sub-figure), which is used to generate refined diffusion condition $\\mathbf{c}_{o2t}$ . Conditioned on $\\mathbf{c}_{o2t}$ , we further perform a $T$ -step object-to-text diffusion to generate text embedding $\\mathbf{x}_{\\mathrm{txt}}$ (bottom sub-figure), whose similarity with the CLIP text embedding $\\hat{\\mathbf{x}}_{\\mathrm{txt}}$ are measured to serve as the final class prediction.", + "bbox": [ + 212, + 349, + 787, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Formally, by extending Eq. 2, CLIFF consists of two diffusion processes among the object, image, and text embedding sub-space. Guided by the condition $\\mathbf{c}_{o2i}$ , the object-to-image diffusion, written as $p_{\\theta,\\phi}(\\mathbf{x}_{\\mathrm{img}}|\\mathbf{c}_{o2i},\\mathbf{x}_{\\mathrm{obj}})$ , aims to generate the associated CLIP image embedding from the object $\\mathbf{x}_{\\mathrm{obj}}$ with $L$ steps:", + "bbox": [ + 212, + 516, + 789, + 579 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta , \\phi} (\\mathbf {x} _ {\\mathrm {i m g}} | \\mathbf {c} _ {o 2 i}, \\mathbf {x} _ {\\mathrm {o b j}}) = p _ {\\phi} (\\mathbf {x} _ {T} | \\mathbf {x} _ {\\mathrm {o b j}}) \\prod_ {t = T - L + 1} ^ {T} p _ {\\theta} (\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}, \\mathbf {c} _ {o 2 i}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 277, + 593, + 787, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similarly, with condition $\\mathbf{c}_{o2t}$ , the object-to-text diffusion is designed to generate text embedding with $T$ steps for the final prediction, denoted as follows,", + "bbox": [ + 212, + 648, + 787, + 680 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta , \\phi} (\\mathbf {x} _ {\\mathrm {t x t}} | \\mathbf {c} _ {o 2 t}, \\mathbf {x} _ {\\mathrm {o b j}}) = p _ {\\phi} (\\mathbf {x} _ {T} | \\mathbf {x} _ {\\mathrm {o b j}}) \\prod_ {t = 1} ^ {T} p _ {\\theta} (\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}, \\mathbf {c} _ {o 2 t}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 297, + 694, + 787, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proposed CLIFF offers two significant improvements compared to the standard diffusion process (Eq. 2). First, instead of beginning from a random noise space $p(\\mathbf{x}_T)$ , we start from a differential probabilistic space $p_{\\phi}(\\mathbf{x}_T|\\mathbf{x}_{\\mathrm{obj}})$ with objects to generate the text embedding. Second, going beyond a single diffusion between two sub-spaces, we formulate the effective diffusion process among the object, image, and text in different modalities.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 326, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Variational Latent Sampler", + "text_level": 1, + "bbox": [ + 215, + 146, + 485, + 162 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With input images $\\{I_i\\}_{i=1}^B$ , the backbone, RPN4, and the feature extractor are adopted to obtain $N$ object embeddings $\\mathbf{x}_{\\mathrm{obj}} \\in \\mathbb{R}^{N \\times D}$ . Then, we send $\\mathbf{x}_{\\mathrm{obj}}$ to VLS to establish a probabilistic embedding space $p_{\\phi}(\\mathbf{x}_T | \\mathbf{x}_{\\mathrm{obj}})$ satisfying the DDPM w.r.t the normally distributed assumption for the diffusion-based transfer. This probabilistic space $p_{\\phi}(\\mathbf{x}_T | \\mathbf{x}_{\\mathrm{obj}})$ can break through the discriminative limitation of [14,15,46,75] without class bias [18]. Besides, it introduces object-level semantics to enable object-to-text transfer with a differential nature, different from the non-optimizable random space $p(\\mathbf{x}_T)$ of noise-to-data transfer [21].", + "bbox": [ + 212, + 166, + 787, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Formally, inspired by VAE [24], we establish the probabilistic space with centered isotropic multivariate Gaussian $p_{\\phi}(\\mathbf{x}_T|\\mathbf{x}_{\\mathrm{obj}}) = \\mathcal{N}(\\mathbf{x}_T;\\mathbf{0},\\mathbf{I})$ , whose distribution parameters are computed from the observed object embedding $\\mathbf{x}_{\\mathrm{obj}}$ with non-linear layers [18]. To this end, two fully connected layers are introduced on the shared region feature extractor of the basic object detector, denoted as $f_{\\mathrm{obj}}$ e.g., the ResNet blocks in Faster RCNN [69] and the MLP in Mask RCNN [20], to learn to estimate the distribution parameters. Specifically, given the object embedding $\\mathbf{x}_{\\mathrm{obj}}$ , the variational posterior $p_{\\phi}(\\mathbf{x}_T|\\mathbf{x}_{\\mathrm{obj}})$ is written as $\\log p_{\\phi}(\\mathbf{x}_T|\\mathbf{x}_{\\mathrm{obj}}^{(i)}) = \\log \\mathcal{N}(\\mathbf{x}_T;\\pmb{\\mu}^{(i)},\\pmb{\\sigma}^{2(i)}\\mathbf{I})$ . Without losing the differential property [24], the mean and s.d. of the probabilistic object space can be estimated,", + "bbox": [ + 212, + 287, + 789, + 443 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} _ {(i)} ^ {(i)} = \\mathbf {W} _ {\\mu} f _ {\\mathrm {o b j}} \\left(\\mathbf {x} _ {\\mathrm {o b j}} ^ {(i)}\\right) + \\mathbf {b} _ {\\mu}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 460, + 785, + 489 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\pmb {\\sigma} ^ {(i)} = \\mathbf {W} _ {\\sigma} f _ {\\mathrm {o b j}} (\\mathbf {x} _ {\\mathrm {o b j}} ^ {(i)}) + \\mathbf {b} _ {\\sigma},\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 484, + 596, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_{(\\cdot)}$ and $\\mathbf{b}_{(\\cdot)}$ are learned weight and bias. Then, with the probabilistic modeling, we sample object-centric noise from the latent object space $\\mathbf{x}_T \\sim p_{\\phi}(\\mathbf{x}_T | \\mathbf{x}_{\\mathrm{obj}})$ with reparameterization [24]: $\\mathbf{x}_T^{(i)} = \\boldsymbol{\\mu}^{(i)} + \\boldsymbol{\\sigma}^{(i)} \\odot \\boldsymbol{\\epsilon}$ , where $\\odot$ is the element-wise product and $\\boldsymbol{\\epsilon} \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ .", + "bbox": [ + 212, + 508, + 787, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To optimize the probabilistic space, we deploy Kullback-Leibler (KL) divergence $D_{\\mathrm{KL}} \\left(p_{\\phi}(\\mathbf{x}_T|\\mathbf{x}_{\\mathrm{obj}})||p(\\mathbf{x}_T)\\right)$ with $\\mathbf{x}_T \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ that encourages the Gaussian distributed [24], which has two advantages. First, the probabilistic nature of normal Gaussian removes the class bias within the object space [18] and promotes the generalization capacity towards discovering novel classes [45, 59]. Second, it naturally satisfies the DDPM constraint (Eq. 1) with the Gaussian formulation [21] for the following diffusion process. The implemented KL optimization objective, as proved in the Appendix, can be re-written in the following format,", + "bbox": [ + 212, + 571, + 787, + 691 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {K L} = - \\frac {1}{2 N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {D} \\left(1 + \\log \\left(\\boldsymbol {\\sigma} _ {j} ^ {2 (i)}\\right) - \\boldsymbol {\\mu} _ {j} ^ {2 (i)} - \\boldsymbol {\\sigma} _ {j} ^ {2 (i)}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 312, + 698, + 785, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VLS can establish a probabilistic space with differential property and sample object-centric noise $\\mathbf{x}_T$ for the diffusion process, serving to generate probabilistic object-text pairs in the diffusion-based distribution transfer [48,51].", + "bbox": [ + 212, + 744, + 787, + 792 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection", + "bbox": [ + 233, + 113, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "4 Besides $N_{rpn}$ RPN proposals, we follow [68, 81] to add additional proposals for training, including $N_{cs}$ class-specific proposals (from $\\mathcal{D}_{aux}$ ) and $N_{ca}$ class-agnostic proposals (from $\\mathcal{D}_{det}$ ) from [58]. The number of proposals is $N = N_{rpn} + N_{cs} + N_{ca}$ .", + "bbox": [ + 217, + 796, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Continual Diffusion Module", + "text_level": 1, + "bbox": [ + 215, + 146, + 491, + 161 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With object-centric noise $\\mathbf{x}_T$ , we further deploy the object-to-image (o2i) and object-to-text (o2t) diffusion to achieve a continual distribution transfer for OVD. To achieve this, an efficient diffusion module $\\epsilon_{\\theta}(\\cdot)$ with MLP architecture is designed, the details of which are presented in the Appendix.", + "bbox": [ + 212, + 171, + 787, + 231 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Object-to-image Diffusion. Given class-agnostic region proposals, we send them to the object detector to generate the object embedding $\\mathbf{x}_{\\mathrm{obj}}$ and feed them into frozen CLIP image encoder [67] to obtain static image embedding $\\hat{\\mathbf{x}}_{\\mathrm{img}}$ . Then, with a randomly sampled object-centric noise $\\mathbf{x}_T \\sim p_\\phi(\\mathbf{x}_T | \\mathbf{x}_{\\mathrm{obj}})$ from VLS, we encourage the diffusion module to generate the image embedding $\\mathbf{x}_{\\mathrm{img}}$ with object-level condition $\\mathbf{c}_{o2i} = \\mathbf{x}_{\\mathrm{obj}}$ using a $L$ -step reverse diffusion:", + "bbox": [ + 212, + 232, + 789, + 323 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {\\mathrm {i m g}} = \\operatorname {R e v D i f} \\left(\\mathbf {x} _ {T}, \\mathbf {c} _ {o 2 i}, L\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 335, + 787, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathrm{RevDif}(\\cdot)$ (Algorithm 1) recovers image embedding from object-centric noise $\\mathbf{x}_T$ . To this end, the object-to-image diffusion, denoted as $p_{\\theta,\\phi}(\\mathbf{x}_{\\mathrm{img}}|\\mathbf{c}_{o2i},\\mathbf{x}_{\\mathrm{obj}})$ , uses object embedding as the diffusion condition $\\mathbf{c}_{o2i} = \\mathbf{x}_{\\mathrm{obj}}$ , and is optimized by minimizing the distance between the generated embedding $\\mathbf{x}_{\\mathrm{img}}$ and CLIP image embedding $\\hat{\\mathbf{x}}_{\\mathrm{img}}$ , which is achieved via the following loss function,", + "bbox": [ + 214, + 361, + 795, + 439 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {o 2 i} = \\frac {1}{N _ {c a}} \\sum_ {i = 1} ^ {N _ {c a}} | | \\mathbf {x} _ {\\mathrm {i m g}} ^ {(i)} - \\hat {\\mathbf {x}} _ {\\mathrm {i m g}} ^ {(i)} | | _ {1}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 450, + 787, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $N_{ca}$ is the number of class-agnostic proposals. With the optimized diffusion process, the object distribution could be transferred to the CLIP image counterpart, facilitating the adaptation to the CLIP latent space.", + "bbox": [ + 212, + 500, + 787, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Object-to-text Diffusion. With diffusion-based formulation, we further encourage the model to reconstruct the associated text embedding $\\mathbf{x}_{\\mathrm{txt}}$ for region proposals $\\mathbf{x}_{\\mathrm{obj}}$ , which conducts a generative procedure in latent space. Note that this generative objective is different from existing OVD works that directly optimize the discriminative class separation. Specifically, we design the object-to-text diffusion with a $T$ -step diffusion $p_{\\theta,\\phi}(\\mathbf{x}_{\\mathrm{txt}}|\\mathbf{c}_{o2t},\\mathbf{x}_{\\mathrm{obj}})$ guided by condition $\\mathbf{c}_{o2t}$ . Considering the dominant role of condition [21] in the generated content, we propose a simple yet effective attentive mechanism to obtain the refined condition $\\mathbf{c}_{o2t} = \\bar{\\mathbf{x}}_{\\mathrm{obj}}^{\\mathrm{img}}$ with the generated image image-level context:", + "bbox": [ + 212, + 545, + 787, + 681 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {\\mathrm {o b j}} ^ {\\mathrm {i m g}} = \\operatorname {S i g m o i d} \\left(\\operatorname {F C} \\left(\\mathbf {x} _ {\\mathrm {i m g}}\\right)\\right) \\odot \\mathbf {x} _ {\\mathrm {o b j}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 695, + 787, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_{\\mathrm{img}}$ is generated via Eq. 8, $\\odot$ is the element-wise product, and FC is a liner projection. Then, we use the refined condition $\\mathbf{c}_{o2t} = \\mathbf{x}_{\\mathrm{obj}}^{\\mathrm{img}}$ to guide the text embedding generation for $\\mathbf{x}_{\\mathrm{txt}}$ with a $T$ -step reverse diffusion:", + "bbox": [ + 212, + 723, + 787, + 771 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {\\mathrm {t x t}} = \\operatorname {R e v D i f} \\left(\\mathbf {x} _ {T}, \\mathbf {c} _ {o 2 t}, T\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 405, + 784, + 787, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Considering the mutual benefits of generative and discriminative objectives for model learning as explored in [45], the loss function for the object-to-text", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 326, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "diffusion process is designed with these two distinct components. To this end, the optimization objective can be expressed as follows,", + "bbox": [ + 214, + 146, + 787, + 176 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {o 2 t} = \\underbrace {\\frac {1}{N _ {l}} \\sum_ {i} ^ {N _ {l}} | | \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\mathbf {x} _ {t} ^ {(i)} , \\mathbf {c} _ {o 2 t} ^ {(i)} , t\\right) - \\boldsymbol {\\epsilon} | | _ {1}} _ {\\text {G e n e r a t i v e}} + \\underbrace {\\mathcal {L} _ {C E} \\left(\\operatorname {S i g m o i d} \\left(\\frac {1}{\\tau} \\cos \\left(\\mathbf {x} _ {\\mathrm {t x t}}, \\hat {\\mathbf {x}} _ {\\mathrm {t x t}}\\right)\\right), \\hat {y}\\right)} _ {\\text {D i s c r i m i n a t i v e}} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 191, + 785, + 263 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ , $\\mathbf{x}_t$ is the noisy CLIP text embedding at a randomly sampled time-step $t \\sim \\mathcal{U}[1,T]$ , $\\tau = 0.02$ is a scaling factor as [15, 68], $\\hat{y}$ is the class label, $\\mathcal{L}_{CE}$ is the Cross-Entropy loss, and $N_{l} = N_{cs} + N_{rpn}$ counts the labeled proposals. The first term optimizes the diffusion module $\\epsilon_{\\theta}(\\cdot)$ to reconstruct the text embedding in a generative manner [21], serving for the objective of Eq. 3. The second term can give an interpretable score for NMS [69] in detectors and ensure the generated embedding with sufficient discriminability.", + "bbox": [ + 212, + 265, + 787, + 369 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Remark. We use a shared diffusion module with a $T$ -step diffusion formulation to optimize two processes, where the first $L$ steps are assigned for object-to-image, and the whole $T$ steps are for object-to-text diffusion. Note that even though we do not explicitly optimize image-to-text diffusion, this transition can be intrinsically achieved between the $L$ -th and $T$ -th steps, as illustrated by the gray-colored part in Fig. 1(b). The reason is based on the continual property of the Markov chain in DDPM, i.e., formulating a random process that undergoes transitions from one state to another within a finite number of possible states [21, 63].", + "bbox": [ + 212, + 372, + 787, + 494 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Training and Inference", + "text_level": 1, + "bbox": [ + 215, + 525, + 450, + 540 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training Stage. We implement the overall optimization objective on the basic object detector, Faster RCNN [69], to train the whole framework:", + "bbox": [ + 214, + 560, + 787, + 592 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C L I F F} = \\mathcal {L} _ {r p n} + \\mathcal {L} _ {r e g} + \\alpha \\mathcal {L} _ {K L} + \\beta_ {1} \\mathcal {L} _ {o 2 i} + \\beta_ {2} \\mathcal {L} _ {o 2 t}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 607, + 787, + 625 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{rpn}$ and $\\mathcal{L}_{reg}$ are RPN loss and the bounding box regression loss in the RCNN head [69] borrowed from object detectors. $\\alpha, \\beta_1, \\beta_2$ are the loss weight.", + "bbox": [ + 214, + 638, + 787, + 667 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Inference Stage. We first extract the region proposals $\\mathbf{x}_{\\mathrm{obj}}$ and then conduct reverse diffusion to generate the associated text embedding $\\mathbf{x}_{\\mathrm{txt}}$ for region proposals, whose similarity with $\\hat{\\mathbf{x}}_{\\mathrm{txt}}$ is measured for class prediction $\\tilde{y}$ :", + "bbox": [ + 214, + 670, + 789, + 717 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {y} = \\operatorname {S i g m o i d} \\left(\\frac {1}{\\tau} c o s \\left(\\mathbf {x} _ {\\mathrm {t x t}}, \\hat {\\mathbf {x}} _ {\\mathrm {t x t}}\\right)\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 731, + 787, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\tilde{y}$ is the per-class confidence, $\\mathbf{x}_{\\mathrm{txt}}$ is the generated text embedding via object-to-text diffusion (Eq. 11), and $\\hat{\\mathbf{x}}_{\\mathrm{txt}}$ is CLIP text embedding for all classes.", + "bbox": [ + 214, + 777, + 787, + 808 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection", + "bbox": [ + 233, + 113, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "5 Following Eq. 1, $\\mathbf{x}_t = \\sqrt{\\bar{\\alpha}_t}\\hat{\\mathbf{x}}_{\\mathrm{xt}} + (1 - \\bar{\\alpha}_t)\\epsilon$ as [21]", + "bbox": [ + 217, + 823, + 566, + 840 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/b8b82f1567a1cc447292effb74fad6a3abcac1a68de2da1c3a3db7dd8b1b88f2.jpg", + "table_caption": [ + "Table 1: Experimental results (%) on COCO benchmark. † indicates the reproduced results with the consistent baseline setting as [68], including ImageNet pre-training and CLIP prompting. ‡ indicates using the longer training schedule [68]." + ], + "table_footnote": [], + "table_body": "
MethodObject DetectorSub-space Alignment\\( \\mathrm{mAP}_{\\mathrm{n}} \\)\\( \\mathrm{mAP}_{\\mathrm{b}} \\)mAPReference
OVR-CNN [82]Faster RCNN C4Obj → Txt22.846.039.9CVPR-2021
RegionCLIP [85]Faster RCNN C4Obj → Txt & Img26.854.847.5CVPR-2022
ViLD [15]Faster RCNN C4Obj → Txt & Img27.659.551.3ICLR-2022
OV-DETR [81]Deformable DETRObj → Txt & Img29.461.052.7ECCV-2022
Detic [86]Faster RCNN C4Obj → Txt27.847.145.0ECCV-2022
Detic‡ [86]Faster RCNN C4Obj → Txt28.453.847.2ECCV-2022
PB-OVD [14]Faster RCNN C4Obj → Txt30.846.142.1ECCV-2022
MAVL-OVD [58]Faster RCNN C4Obj → Txt30.452.646.8ECCV-2022
\\( OCD^‡ \\)[68]Faster RCNN C4Obj → Txt & Img36.654.049.4NeurIPS-2022
F-VLM [26]Faster RCNN FPNObj → Txt & Img28.0-39.6ICLR-2023
VLDet [46]Faster RCNN C4Obj → Txt32.050.645.8ICLR-2023
BARON†,‡ [75]Faster RCNN C4Obj → Txt & Img41.053.350.1CVPR-2023
EdaDet [71]Faster RCNN C4Obj → Txt37.857.752.5ICCV-2023
GOAT [72]Faster RCNN C4Obj → Txt & Img36.453.048.6ICCV-2023
CLIFF (ours)Faster RCNN C4Obj → Image → Txt41.354.150.6-
CLIFF‡ (ours)Faster RCNN C4Obj → Image → Txt43.254.051.2-
", + "bbox": [ + 228, + 199, + 774, + 386 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 412, + 375, + 430 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 215, + 445, + 532, + 460 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We follow existing works [68, 75] to compare with state-of-the-art methods. COCO Setting. The COCO-2017 dataset [47] is used for training and evaluation, containing 80 object-level classes. As for the auxiliary dataset $\\mathcal{D}_{aux}$ , we follow [82] to use COCO caption dataset in $\\mathcal{C}_B \\cup \\mathcal{C}_N$ . The core evaluation metric is the novel-class Average Precision (AP) with an IoU threshold of 0.5, denoted as $\\mathrm{mAP_n}$ , for the novel-class generalization. The results are shown in Table 1.", + "bbox": [ + 212, + 470, + 787, + 561 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "LVIS Setting. LVIS v1.0 [17] contains 1203 classes and is split into frequent, common, and rare types. For the auxiliary datasets $\\mathcal{D}_{aux}$ , we follow [68] to use the subset of ImageNet-21K. The core evaluation metric is mask AP of rare classes, denoted as $\\mathrm{AP_r^m}$ . We report the comparison results in Table 2.", + "bbox": [ + 212, + 563, + 787, + 622 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Pascal VOC Setting. We follow [72] to evaluate the generalization to unseen datasets with Pascal VOC [10] and LVIS [17] using COCO-trained model. We use $\\mathrm{AP}_{50}$ to evaluate the performance as [72]. The results are given in Table 7.", + "bbox": [ + 212, + 623, + 787, + 669 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 691, + 455, + 708 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Object Detector. Considering that previous works use inconsistent detectors, we strictly follow the setting of [68,82] in different benchmarks. Concretely, for the COCO setting, we use Faster-RCNN C4 [69] with the ResNet-50 backbone. The backbone is pre-trained on ImageNet [8] without bells and whistles. For the LVIS setting, we follow the two main streams of existing works by using Mask-RCNN-FPN [20] and CenterNet v2 [87] as the basic object detectors for a thorough comparison, which is based on ReseNet-50 backbone. For the experiments of all benchmarks, we use the SGD optimizer with a weight decay of $1e^{-4}$ and a", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 325, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/ca8f4b76e4eb43f6cc73aee5c8bc490df2c2e3c3aaa3dd40bf8a1bac4b9c2b6b.jpg", + "table_caption": [ + "Table 2: Experimental results (\\%) on LVIS benchmark with the consistent object detector setting. Ens indicates using the ensemble trick [11]." + ], + "table_footnote": [], + "table_body": "
MethodObject DetectorAPmRAPmcAPmfAPReference
ViLD-Ens [15]Mask RCNN16.624.630.325.5ICLR-2022
MVAL-OVD [58]Mask RCNN17.021.226.122.4ECCV-2022
OCD [68]Mask RCNN21.125.029.125.9NeurIPS-2022
DetPro [9]Mask RCNN19.825.628.925.9CVPR-2022
RegionCLIP [85]Mask RCNN17.127.434.028.2CVPR-2022
F-VLM [26]Mask RCNN18.624.026.924.2ICLR-2023
BARON [75]Mask RCNN18.024.428.925.1CVPR-2023
BARON-Ens [75]Mask RCNN19.226.829.426.5CVPR-2023
CLIFF (ours)Mask RCNN22.426.929.826.8-
VLDet [46]CenterNet v221.729.834.330.1ICLR-2023
CoDet [57]CenterNet v223.430.034.630.7NeirIPS-2023
GOAT [72]CenterNet v223.329.734.330.4ICCV-2023
EdaDet [71]CenterNet v223.727.529.127.5ICCV-2023
CLIFF (ours)CenterNet v224.527.729.828.2-
", + "bbox": [ + 279, + 184, + 723, + 354 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "momentum of 0.9 following [68]. We use 4 NVIDIA V100 for training with a batch size of 16. We pre-train the base object detector with a $1 \\times$ schedule as [68], and then train the whole model with a $2 \\times$ schedule. We also explore the longer $8 \\times$ training schedule to make fair comparisons with specific methods.", + "bbox": [ + 212, + 381, + 787, + 441 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "CLIP Model. As the previous work [68], we use the single prompt, 'a photo of a {category}', to compute CLIP text embedding $\\hat{\\mathbf{x}}_{\\mathrm{txt}}$ , and use $224\\times 224$ input-scale for image embedding $\\hat{\\mathbf{x}}_{\\mathrm{img}}$ via ViT-B/32 model [67].", + "bbox": [ + 212, + 443, + 787, + 488 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "CLIFF. The number of the time-step in object-to-text $(T)$ and object-to-image $(L)$ diffusion are set $T = 10, L = 3$ , which are justified in Table 4. $\\alpha$ , $\\beta_{1}$ , and $\\beta_{2}$ are empirically set to 2.0, 15.0, and 1.0, respectively, which are verified by the experiments in Appendix. $L_{2}$ normalization is deployed on CLIP embedding as [15,68]. Class-specific and RPN proposals are used to optimize object-to-text diffusion [68,75], while class-agnostic ones are for object-to-image diffusion.", + "bbox": [ + 212, + 488, + 787, + 580 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.3 Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 214, + 602, + 620, + 618 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results on COCO Benchmark. As shown in Table 1, we compare the proposed method with state-of-the-art counterparts on the COCO benchmark. CLIFF achieves $41.3\\% \\mathrm{mAP_n}$ , which outperforms OCD [68] ( $36.6\\% \\mathrm{mAP_n}$ ), VLDet [46] ( $32.0\\% \\mathrm{mAP_n}$ ) and BARON [75] ( $41.0\\% \\mathrm{mAP_n}$ ) with $4.7\\%$ , $9.3\\%$ and $0.3\\% \\mathrm{mAP_n}$ , comprehensively verifying the effectiveness of the proposed CLIFF. The proposed method also surpasses EdaDet [71] and GOAT [72] by a noticeable margin. Besides, after using fair multi-scale training [68, 75], our method ( $43.2\\% \\mathrm{mAP_n}$ ) outperforms existing works by a large margin. Unlike existing works aligning object-text and object-image separately, we formulate a continual diffusion process among the three sub-spaces in a different learning paradigm.", + "bbox": [ + 212, + 627, + 787, + 779 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results on LVIS Benchmark. We present a thorough comparison with corresponding state-of-the-art methods using Mask RCNN and CenterNet v2 detectors, which is shown in Table 2. For Mask RCNN, it can be observed that CLIFF achieves the best $22.4\\%$ $\\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{m}}$ , which surpasses state-of-the-art counterparts", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection", + "bbox": [ + 233, + 113, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/49aff2b8c3083ee2e2ffeddf0479d98f9ace8f04889dcfa8f7c2551ceb7c9321.jpg", + "table_caption": [ + "Table 3: Ablation study results on COCO and LVIS benchmarks with $1 \\times$ schedule." + ], + "table_footnote": [], + "table_body": "
CDM o2t o2i VLSVLSCOCOLVIS-BoxLVIS-Mask
mAPnmAPbmAPAPbAPcAPbAPbAPmAPmAPcAPb
1(baseline)32.249.945.315.820.126.221.415.219.725.020.5
2✓36.649.546.116.220.927.422.615.820.125.721.6
3✓39.552.449.216.721.727.723.116.520.925.922.1
4✓38.853.349.317.321.927.823.316.721.126.122.2
5✓40.052.949.518.922.128.023.818.121.426.222.7
", + "bbox": [ + 223, + 171, + 779, + 258 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/929e0a7405a7a8b89078fea8488fb20f3c09c38bcf83345d8e821a6038747dc2.jpg", + "table_caption": [ + "Table 4: Analysis on the diffusion step. o2i and o2t indicates the object-to-image and object-to-text diffusion process respectively. * indicates sharing the same diffusion procedure and optimizing o2i and o2t in a combinatorial style [15,68]." + ], + "table_footnote": [], + "table_body": "
o2io2t\\( \\mathrm{mAP}_{\\mathrm{n}} \\)\\( \\mathrm{mAP}_{\\mathrm{b}} \\)mAP
111040.253.450.0
231041.354.150.6
351041.353.750.3
471041.253.950.5
591041.453.550.4
63540.953.850.3
731041.354.150.6
832039.753.449.8
9101041.453.650.4
10*10*1039.952.849.5
", + "bbox": [ + 222, + 367, + 486, + 494 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/804981751749b03a58764a57e35568bd9da68e8170d16a7e472406d26ecde3e6.jpg", + "table_caption": [ + "Table 5: Analysis of the discriminative and generative terms in Eq. 12 by replacing the random noise (Rand) with object-centric noise (OC) in different time steps $(2\\times$ schedule). rows 7 and 8 indicate removing individual terms." + ], + "table_footnote": [], + "table_body": "
DiscriminativeGenerative\\( \\mathrm{mAP}_{\\mathrm{n}} \\)
TT-1: 0TT-1: 0
1RandRandRandRand39.2
2OCRandOCRand41.0
3RandOCRandOC40.4
4OCOCOCRand41.1
5OCRandOCOC41.3
6OCOCOCOC41.3
7w/o DiscriminativeOCOC38.2
8OCOCw/o Generative40.1
", + "bbox": [ + 511, + 372, + 785, + 489 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "BARON [75] $(19.2\\% \\mathrm{AP_r^m})$ , OCD [68] $(21.3\\% \\mathrm{AP_r^m})$ with $3.1\\% \\mathrm{AP_r^m}$ and $1.3\\% \\mathrm{AP_r^m}$ , respectively. Besides, our method surpasses the baseline model MAVL-OVD [58] with a significant $5.4\\% \\mathrm{AP_r^m}$ , verifying the effectiveness of our method. For the comparison with CenterNet v2, we can find that our method gives $24.5\\% \\mathrm{AP_r^m}$ and also surpasses the latest published works noticeably, outperforming CoDet [57] and GOAT [72] with $1.1\\%$ and $1.2\\%$ gains. This can reveal that CLIFF is a generic method in terms of multi-modal alignment and can be deployed in different object detectors with great transferability and adaptability.", + "bbox": [ + 212, + 522, + 787, + 645 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.4 Quantitative Analysis", + "text_level": 1, + "bbox": [ + 214, + 664, + 441, + 680 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation Study. The ablative results are shown in Table 3 with a $1 \\times$ training schedule. The first line indicates the model trained with extra MAVL proposals [58, 68], which is a fair baseline model as [68, 75]. Compared with the baseline model $(32.2\\% \\mathrm{mAP}_{\\mathrm{n}}, 15.8\\% \\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{b}}, 15.2\\% \\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{m}})$ , introducing the object-to-text (row 2) diffusion gives satisfactory improvements with $36.6\\% \\mathrm{mAP}_{\\mathrm{n}}, 16.2\\% \\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{b}}$ and $15.8\\% \\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{m}}$ . After integrating object-to-image diffusion (row 3), i.e., the full CDM, we can observe a further improvement with $7.3\\% \\mathrm{mAP}_{\\mathrm{n}}, 0.9\\% \\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{b}}$ and $1.3\\% \\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{m}}$ gains, compared with the baseline. This verifies the superior effectiveness in modeling a continual distribution transfer than the existing discriminative paradigm. Besides, after replacing the randomly sampled noise by", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 325, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/7e4ed113fb356aee53bcd1898aeaa96c14386ab1ed9506dcbe856a9af8b980c4.jpg", + "table_caption": [ + "Table 6: Analysis of the refined condition ${\\mathbf{c}}_{o2t}$ in object-to-image diffusion (Eq. 5) with $2 \\times$ training schedule." + ], + "table_footnote": [], + "table_body": "
Setting of c02tmAPnmAPbmAP
1Object Embed: xobj40.453.449.9
2Image Embed: ximg39.652.949.4
3Additive: xobj + ximg40.853.550.2
4Concated: [xobj : ximg]41.053.650.4
5Attentive: Eq. 1041.354.150.6
", + "bbox": [ + 220, + 199, + 488, + 277 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/7bd8dd669fe3b043c403169de7d0285a02f526cc98974a6c903b1df84edc006b.jpg", + "table_caption": [ + "Table 7: Cross-dataset evaluation with $\\mathrm{AP}_{50}$ (\\%) transferring COCO-trained models to the PASCAL VOC and LVIS." + ], + "table_footnote": [], + "table_body": "
MethodPASCALLVIS
OVR-CNN [82]52.95.2
PB-OVD [14]59.28.0
VLDet [46]61.710.0
GOAT [72]63.614.0
CLIFF (Ours)66.116.7
", + "bbox": [ + 522, + 199, + 774, + 277 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "object-centric noise with VLS (row 5), it can be observed that the CDM (row 3) can be improved with $40.0\\% \\mathrm{mAP}_{\\mathrm{n}}$ , $18.9\\% \\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{b}}$ , and $18.1\\% \\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{m}}$ , verifying the effectiveness of each component in the proposed CLIFF.", + "bbox": [ + 212, + 306, + 785, + 353 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The Number of Diffusion Steps. We study different diffusion steps for o2i and o2t in Table 4, and have the following three observations. First, in rows 1-5, we observe that the results are stable when $L > 3$ (around $41.3\\% \\mathrm{mAP_n}$ ), but give notable performance drops when the step is insufficient $L = 1$ . Considering more computation requirements when enlarging the diffusion step, we select $L = 3$ with the best trade-off. Second, in rows 6-8, setting too small and large o2t steps both give significant performance drops, which may be caused by unsatisfactory diffusion optimization. Third, we set consistent diffusion steps and compare the proposed continual diffusion (row 9) with existing combinatorial style, i.e., sharing the diffusion process and generated embedding between the o2i and o2t (row 10). With a $1.5\\%$ performance gain over the combinatorial objective, the effectiveness of our continual transfer can be justified.", + "bbox": [ + 212, + 354, + 787, + 535 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Exploring Object-centric Noise. We further give a detailed analysis of the object-centric (OC) noise by replacing the random noise for the discriminative and generative terms in object-to-text objective (the core process for OVD) in different time steps, as shown in Table 5. Compared with the random-noise baseline with $39.2\\% \\mathrm{mAP}_{\\mathrm{n}}$ , controlling the initial stage (row 2) and middle stages (row 3) gives $41.0\\%$ are $40.4\\% \\mathrm{mAP}_{\\mathrm{n}}$ , respectively. Introducing more OC noise in the diffusion (rows 4-6) gives stable performance improvements, indicating the effectiveness of the VLS. Further, we conduct an ablative study on o2t diffusion (rows 7-8), i.e., the first and second terms in Eq. 12, verifying their critical role in optimizing the diffusion process with both aspects.", + "bbox": [ + 212, + 536, + 787, + 686 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Analyzing the Condition Design. The strategy of refining the object-to-text (Eq. 5) condition $\\mathbf{c}_{o2t} = \\mathbf{x}_{\\mathrm{obj}}^{\\mathrm{img}}$ is analyzed in Table 6. Compared with individually using the object and image (rows 1-2) condition with $40.4\\%$ and $39.6\\%$ $\\mathrm{mAP_n}$ , combining the two aspects for the diffusion condition generally works better (rows 3-5), e.g., addition with $40.8\\%$ $\\mathrm{mAP_n}$ , concatenation with $41.0\\%$ . Besides, we observe that the proposed attentive refinement on the condition (Eq. 10) achieves optimal performance and gives the best $41.3\\%$ $\\mathrm{mAP_n}$ , due to the adaptive fusion of image-level context. This can verify the effectiveness of our design.", + "bbox": [ + 212, + 688, + 787, + 808 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Generalization to Unseen Datasets. We further conduct the cross-dataset comparison following [72] by considering two different settings. Compared with", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection", + "bbox": [ + 233, + 113, + 730, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 127 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2cdba09ea8db7d68a01774b1fca21647f413b2a958854e216770d5fef9d3f8d9.jpg", + "image_caption": [ + "Fig. 3: Qualitative comparison among (a) the baseline [68], (b) our CLIFF, and (c) ground-truth on the COCO benchmark. We only visualize the novel-class predictions for a clearer evaluation of open-vocabulary generalization." + ], + "image_footnote": [], + "bbox": [ + 228, + 143, + 779, + 329 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "state-of-the-art counterpart GOAT [72], the proposed CLIFF gives $66.1\\%$ and $16.7\\% \\mathrm{AP}_{50}$ on the PASCAL VOC and LVIS datasets using the COCO-trained model, which surpasses GOAT with significant $2.5\\%$ and $2.7\\%$ gains, showing the superior generalization capacity of our method.", + "bbox": [ + 212, + 419, + 787, + 481 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.5 Qualitative Analysis", + "text_level": 1, + "bbox": [ + 214, + 498, + 429, + 513 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "As shown in Fig. 3, we visualize the novel-class prediction and conduct a qualitative comparison among (a) the baseline [68,75] trained with MAVL proposals [68], (b) the proposed CLIFF, and (c) ground-truth. It can be observed that the proposed method gives more accurate novel-class predictions, especially for occlusion situations and complex scenarios. Our method reduces the confusion among novel classes, e.g., correcting the wrong prediction (couch) given by the baseline in col. 1. Besides, the proposed CLIFF can relieve some missing errors, as illustrated in col. 2-5, revealing its superior novel-class generalization capacity. Hence, the proposed CLIFF is able to relieve the bias on the base classes and encourage the model to discover the novel-class objects with probabilistic modeling.", + "bbox": [ + 212, + 520, + 787, + 672 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 690, + 359, + 707 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper proposes a novel diffusion-based framework for OVD, coined CLIFF, which breaks through the bottlenecks of the combinatorial and discriminative limitation in existing OVD pipelines. The proposed CLIFF consists of a Variational Latent Sampler (VLS) to sample object-centric noise in the probabilistic object space, and leverages a Continual diffusion Module (CDM) to formulate a continual distribution transfer among three sub-spaces (object, image, and text) with cross-modal nature. Extensive experiments validate that the proposed method significantly outperforms existing approaches on large-scale benchmarks.", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 325, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by Innovation and Technology Commission- Innovation and Technology Fund ITS/229/22 and Hong Kong Research Grants Council (RGC) General Research Fund 11211221.", + "bbox": [ + 215, + 146, + 787, + 193 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 215, + 321, + 233 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ali, S., Ghatwary, N., Jha, D., Isik-Polat, E., Polat, G., Yang, C., Li, W., Galdran, A., Ballester, M.A.G., Thambawita, V., et al.: Assessing generalisability of deep learning-based polyp detection and segmentation methods through a computer vision challenge. Sci. Rep. (2024)", + "2. Amit, T., Nachmani, E., Shaharbany, T., Wolf, L.: Segdiff: Image segmentation with diffusion probabilistic models. ArXiv:2112.00390 (2021)", + "3. Bansal, A., Borgnia, E., Chu, H.M., Li, J.S., Kazemi, H., Huang, F., Goldblum, M., Geiping, J., Goldstein, T.: Cold diffusion: Inverting arbitrary image transforms without noise. ArXiv:2208.09392 (2022)", + "4. Baranchuk, D., Voynov, A., Rubachev, I., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. In: ICLR (2022)", + "5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV. pp. 213-229 (2020)", + "6. Chen, S., Sun, P., Song, Y., Luo, P.: Diffusiondet: Diffusion model for object detection. ArXiv:2211.09788 (2022)", + "7. Chen, Z., Li, W., Xing, X., Yuan, Y.: Medical federated learning with joint graph purification for noisy label learning. MedIA (2023)", + "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009)", + "9. Du, Y., Wei, F., Zhang, Z., Shi, M., Gao, Y., Li, G.: Learning to prompt for open-vocabulary object detection with vision-language model. ArXiv:2203.14940 (2022)", + "0. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. IJCV 88, 303-338 (2010)", + "1. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Expand your detector vocabulary with uncurated images. ArXiv:2203.16513 (2022)", + "2. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Towards open-vocabulary detection using uncurated images. In: ECCV. pp. 701-717. Springer (2022)", + "3. Gao, M., Xing, C., Niebles, J.C., Li, J., Xu, R., Liu, W., Xiong, C.: Towards open vocabulary object detection without human-provided bounding boxes. ArXiv:2111.09452 (2021)", + "4. Gao, M., Xing, C., Niebles, J.C., Li, J., Xu, R., Liu, W., Xiong, C.: Open vocabulary object detection with pseudo bounding-box labels. In: ECCV (2022)", + "5. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: ICLR (2022)", + "6. Gu, Z., Chen, H., Xu, Z., Lan, J., Meng, C., Wang, W.: Diffusioninst: Diffusion model for instance segmentation. ArXiv:2212.02773 (2022)", + "7. Gupta, A., Dollar, P., Girshick, R.: LVIS: A dataset for large vocabulary instance segmentation. In: CVPR (2019)", + "8. Han, J., Ren, Y., Ding, J., Yan, K., Xia, G.S.: Few-shot object detection via variational feature aggregation. ArXiv:2301.13411 (2023)" + ], + "bbox": [ + 225, + 250, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection", + "bbox": [ + 233, + 113, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Han, X., Zheng, H., Zhou, M.: Card: Classification and regression diffusion models. In: NeurIPS (2022)", + "20. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask R-CNN. In: CVPR (2017)", + "21. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)", + "22. Kim, D., Angelova, A., Kuo, W.: Contrastive feature masking open-vocabulary vision transformer. In: ICCV. pp. 15602-15612 (2023)", + "23. Kim, D., Angelova, A., Kuo, W.: Region-aware pretraining for open-vocabulary object detection with vision transformers. In: CVPR. pp. 11144-11154 (2023)", + "24. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. ArXiv:1312.6114 (2013)", + "25. Korte, B.H., Vygen, J., Korte, B., Vygen, J.: Combinatorial optimization, vol. 1. Springer (2011)", + "26. Kuo, W., Cui, Y., Gu, X., Piergiovanni, A., Angelova, A.: Open-vocabulary object detection upon frozen vision and language models. In: ICLR (2023)", + "27. Kwon, M., Jeong, J., Uh, Y.: Diffusion models already have a semantic latent space. In: ICLR (2023)", + "28. Leng, Y., Chen, Z., Guo, J., Liu, H., Chen, J., Tan, X., Mandic, D., He, L., Li, X.Y., Qin, T., et al.: Binauralgrad: A two-stage conditional diffusion probabilistic model for binaural audio synthesis. ArXiv:2205.14807 (2022)", + "29. Li, A.C., Prabhudesai, M., Duggal, S., Brown, E., Pathak, D.: Your diffusion model is secretly a zero-shot classifier. ArXiv:2303.16203 (2023)", + "30. Li, C., Feng, B.Y., Fan, Z., Pan, P., Wang, Z.: Steganerf: Embedding invisible information within neural radiance fields. In: ICCV. pp. 441-453 (2023)", + "31. Li, C., Lin, X., Mao, Y., Lin, W., Qi, Q., Ding, X., Huang, Y., Liang, D., Yu, Y.: Domain generalization on medical imaging classification using episodic training with task augmentation. Computers in biology and medicine 141, 105144 (2022)", + "32. Li, C., Liu, H., Fan, Z., Li, W., Liu, Y., Pan, P., Yuan, Y.: Gaussianstego: A generalizable stenography pipeline for generative 3d gaussians splatting. ArXiv:2407.01301 (2024)", + "33. Li, C., Liu, H., Liu, Y., Feng, B.Y., Li, W., Liu, X., Chen, Z., Shao, J., Yuan, Y.: Endora: Video generation models as endoscopy simulators. ArXiv:2403.11050 (2024)", + "34. Li, C., Liu, X., Li, W., Wang, C., Liu, H., Yuan, Y.: U-kan makes strong backbone for medical image segmentation and generation. ArXiv:2406.02918 (2024)", + "35. Li, C., Zhang, Y., Li, J., Huang, Y., Ding, X.: Unsupervised anomaly segmentation using image-semantic cycle translation. ArXiv:2103.09094 (2021)", + "36. Li, C., Zhang, Y., Liang, Z., Ma, W., Huang, Y., Ding, X.: Consistent posterior distributions under vessel-mixing: a regularization for cross-domain retinal artery/vein classification. In: ICIP. pp. 61-65 (2021)", + "37. Li, W., Chen, Z., Li, B., Zhang, D., Yuan, Y.: Htd: Heterogeneous task decoupling for two-stage object detection. TIP (2021)", + "38. Li, W., Guo, X., Yuan, Y.: Novel scenes & classes: Towards adaptive open-set object detection. In: ICCV. pp. 15780-15790 (2023)", + "39. Li, W., Liu, J., Han, B., Yuan, Y.: Adjustment and alignment for unbiased open set domain adaptation. In: CVPR. pp. 24110-24119 (2023)", + "40. Li, W., Liu, X., Yao, X., Yuan, Y.: Scan: Cross domain object detection with semantic conditioned adaptation. In: AAAI. pp. 1421-1428 (2022)", + "41. Li, W., Liu, X., Yuan, Y.: Scan++: Enhanced semantic conditioned adaptation for domain adaptive object detection. TMM (2022)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 325, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "42. Li, W., Liu, X., Yuan, Y.: Sigma: Semantic-complete graph matching for domain adaptive object detection. In: CVPR (2022)", + "43. Li, W., Liu, X., Yuan, Y.: Sigma++: Improved semantic-complete graph matching for domain adaptive object detection. TPAMI (2023)", + "44. Li, W., Yang, C., Liu, J., Liu, X., Guo, X., Yuan, Y.: Joint polyp detection and segmentation with heterogeneous endoscopic data. In: ISBI Workshop: EndoCV 2021. pp. 69-79. CEUR-WS Team (2021)", + "45. Liang, C., Wang, W., Miao, J., Yang, Y.: Gmmseg: Gaussian mixture based generative semantic segmentation models. ArXiv:2210.02025 (2022)", + "46. Lin, C., Sun, P., Jiang, Y., Luo, P., Qu, L., Haffari, G., Yuan, Z., Cai, J.: Learning object-language alignments for open-vocabulary object detection. In: ICLR (2023)", + "47. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014)", + "48. Lipman, Y., Chen, R.T.Q., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. In: ICLR (2023)", + "49. Liu, H., Liu, Y., Li, C., Li, W., Yuan, Y.: Lgs: A light-weight 4d gaussian splatting for efficient surgical scene reconstruction. ArXiv:2406.16073 (2024)", + "50. Liu, X., Li, W., Yamaguchi, T., Geng, Z., Tanaka, T., Tsai, D.P., Chen, M.K.: Stereo vision meta-lens-assisted driving vision. ACS Photonics (2024)", + "51. Liu, X., Gong, C., Liu, Q.: Flow straight and fast: Learning to generate and transfer data with rectified flow. ArXiv:2209.03003 (2022)", + "52. Liu, X., Li, W., Yang, Q., Li, B., Yuan, Y.: Towards robust adaptive object detection under noisy annotations. In: CVPR. pp. 14207-14216 (2022)", + "53. Liu, X., Li, W., Yuan, Y.: Decoupled unbiased teacher for source-free domain adaptive medical object detection. TNNLS (2023)", + "54. Liu, Y., Li, C., Yang, C., Yuan, Y.: Endogaussian: Gaussian splatting for deformable surgical scene reconstruction. ArXiv:2401.12561 (2024)", + "55. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. ArXiv:2211.01095 (2022)", + "56. Luo, J., Li, Y., Pan, Y., Yao, T., Feng, J., Chao, H., Mei, T.: Semantic-conditional diffusion networks for image captioning. ArXiv:2212.03099 (2022)", + "57. Ma, C., Jiang, Y., Wen, X., Yuan, Z., Qi, X.: Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. NeurIPS 36 (2024)", + "58. Maaz, M., Rasheed, H., Khan, S., Khan, F.S., Anwer, R.M., Yang, M.H.: Class-agnostic object detection with multi-modal transformer. In: ECCV (2022)", + "59. Miller, D., Sünderhauf, N., Milford, M., Dayoub, F.: Uncertainty for identifying open-set errors in visual object detection. Robot. Lett. 7(1), 215-222 (2021)", + "60. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., et al.: Simple open-vocabulary object detection. In: ECCV. pp. 728-755 (2022)", + "61. Ni, J., Qiu, Q., Chellappa, R.: Subspace interpolation via dictionary learning for unsupervised domain adaptation. In: CVPR (June 2013)", + "62. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: ICML (2021)", + "63. Norris, J.R.: Markov chains. No. 2, Cambridge university press (1998)", + "64. Pan, P., Fan, Z., Feng, B.Y., Wang, P., Li, C., Wang, Z.: Learning to estimate 6dof pose from limited data: A few-shot, generalizable approach using rgb images. In: 2024 International Conference on 3D Vision (3DV). pp. 1059-1071. IEEE (2024)", + "65. Peebles, W., Xie, S.: Scalable diffusion models with transformers. ArXiv:2212.09748 (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection", + "bbox": [ + 233, + 113, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "66. Preechakul, K., Chathee, N., Wizadwongsa, S., Suwajanakorn, S.: Diffusion autoencoders: Toward a meaningful and decodable representation. In: CVPR. pp. 10619-10629 (2022)", + "67. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)", + "68. Rasheed, H.A., Maaz, M., Khattak, M.U., Khan, S., Khan, F.: Bridging the gap between object and image-level representations for open-vocabulary detection. In: NeurIPS (2022)", + "69. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. In: NeurIPS (2015)", + "70. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)", + "71. Shi, C., Yang, S.: Edadet: Open-vocabulary object detection using early dense alignment. In: ICCV. pp. 15724-15734 (2023)", + "72. Wang, J., Zhang, H., Hong, H., Jin, X., He, Y., Xue, H., Zhao, Z.: Open-vocabulary object detection with an open corpus. In: ICCV. pp. 6759-6769 (2023)", + "73. Wizadwongsa, S., Suwajanakorn, S.: Accelerating guided diffusion sampling with splitting numerical methods. In: ICLR (2023)", + "74. Wu, L., Gong, C., Liu, X., Ye, M., Liu, Q.: Diffusion-based molecule generation with informative prior bridges. ArXiv:2209.00865 (2022)", + "75. Wu, S., Zhang, W., Jin, S., Liu, W., Loy, C.C.: Aligning bag of regions for open-vocabulary object detection. ArXiv:2302.13996 (2023)", + "76. Wu, X., Zhu, F., Zhao, R., Li, H.: Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. ArXiv:2303.13076 (2023)", + "77. Xu, H., Zhang, Y., Sun, L., Li, C., Huang, Y., Ding, X.: Afsc: Adaptive fourier space compression for anomaly detection. ArXiv:2204.07963 (2022)", + "78. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. ArXiv:2303.04803 (2023)", + "79. Yang, Q., Li, W., Li, B., Yuan, Y.: Mrm: Masked relation modeling for medical image pre-training with genetics. In: ICCV. pp. 21452-21462 (2023)", + "80. Ye, K., Zhang, M., Kovashka, A., Li, W., Qin, D., Berent, J.: Cap2det: Learning to amplify weak caption supervision for object detection. In: ICCV (2019)", + "81. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary detr with conditional matching. In: ECCV. pp. 106-122 (2022)", + "82. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: CVPR (2021)", + "83. Zbinden, L., Doorenbos, L., Pissas, T., Huber, A.T., Sznitman, R., Márquez-Neila, P.: Stochastic segmentation with conditional categorical diffusion models. In: ICCV (2023)", + "84. Zhang, Y., Li, C., Lin, X., Sun, L., Zhuang, Y., Huang, Y., Ding, X., Liu, X., Yu, Y.: Generator versus segmentor: Pseudo-healthy synthesis. In: MICCAI. pp. 150–160 (2021)", + "85. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: CVPR. pp. 16793-16803 (2022)", + "86. Zhou, X., Girdhar, R., Joulin, A., Krahenbuhl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: ECCV. pp. 350-368 (2022)", + "87. Zhou, X., Koltun, V., Krahenbuhl, P.: Probabilistic two-stage detection. In: ArXiv:2103.07461 (2021)" + ], + "bbox": [ + 215, + 147, + 787, + 832 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Li et al.", + "bbox": [ + 271, + 114, + 325, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/CLIFF_ Continual Latent Diffusion for Open-Vocabulary Object Detection/51812c3e-19bd-4e17-9855-dbf65e0d5287_model.json b/2024/CLIFF_ Continual Latent Diffusion for Open-Vocabulary Object Detection/51812c3e-19bd-4e17-9855-dbf65e0d5287_model.json new file mode 100644 index 0000000000000000000000000000000000000000..aaf0d8ed792304c286df0e6b3db4e4c08e3933af --- /dev/null +++ b/2024/CLIFF_ Continual Latent Diffusion for Open-Vocabulary Object Detection/51812c3e-19bd-4e17-9855-dbf65e0d5287_model.json @@ -0,0 +1,2909 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.284, + 0.141, + 0.723, + 0.187 + ], + "angle": 0, + "content": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.268, + 0.212, + 0.736, + 0.228 + ], + "angle": 0, + "content": "Wuyang \\(\\mathrm{Li}^{1}\\) , Xinyu Liu, Jiayi Ma2, and Yixuan Yuan" + }, + { + "type": "text", + "bbox": [ + 0.333, + 0.239, + 0.671, + 0.283 + ], + "angle": 0, + "content": "1 The Chinese University of Hong Kong \n2 Wuhan University \nwymanbest@outlook.com, yxyuan@ee.cuhk.edu.hk" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.318, + 0.744, + 0.637 + ], + "angle": 0, + "content": "Abstract. Open-vocabulary object detection (OVD) utilizes image-level cues to expand the linguistic space of region proposals, thereby facilitating the detection of diverse novel classes. Recent works adapt CLIP embedding by minimizing the object-image and object-text discrepancy combinatorially in a discriminative paradigm. However, they ignore the underlying distribution and the disagreement between the image and text objective, leading to the misaligned distribution between the vision and language sub-space. To address the deficiency, we explore the advanced generative paradigm with distribution perception and propose a novel framework based on the diffusion model, coined Continual Latent Diffusion (CLIFF), which formulates a continual distribution transfer among the object, image, and text latent space probabilistically. CLIFF consists of a Variational Latent Sampler (VLS) enabling the probabilistic modeling and a Continual Diffusion Module (CDM) for the distribution transfer. Specifically, in VLS, we first establish a probabilistic object space with region proposals by estimating distribution parameters. Then, the object-centric noise is sampled from the estimated distribution to generate text embedding for OVD. To achieve this generation process, CDM conducts a short-distance object-to-image diffusion from the sampled noise to generate image embedding as the medium, which guides the long-distance diffusion to generate text embedding. Extensive experiments verify that CLIFF can significantly surpass state-of-the-art methods on benchmarks. The code is available at https://github.com/CUHK-AIM-Group/CLIFF." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.65, + 0.744, + 0.678 + ], + "angle": 0, + "content": "Keywords: Open-vocabulary object detection \\(\\cdot\\) Diffusion model \\(\\cdot\\) Continual distribution transfer" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.704, + 0.377, + 0.719 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Object detection [20, 37, 69] has made significant strides in recent years, as demonstrated by the impressive results on benchmarks [17, 47]. Nonetheless, due to the unsatisfactory closed-set constraint [77, 82], object detectors are limited to discovering the novel-class objects [38, 39] unseen during training, impeding a reliable scene understanding in the real world. Further, it hinders the journey toward achieving true artificial intelligence since it requires the human-like perceptual ability to recognize unexpected objects appearing everywhere." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.327, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.143, + 0.785, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.23, + 0.361, + 0.24 + ], + "angle": 0, + "content": "(a) Existing OVD works" + }, + { + "type": "image_caption", + "bbox": [ + 0.437, + 0.23, + 0.753, + 0.24 + ], + "angle": 0, + "content": "(b) The proposed CLIFF framework modeling a continual distribution transfer" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.255, + 0.788, + 0.298 + ], + "angle": 0, + "content": "Fig. 1: Illustration of (a) existing discriminative paradigm [15,68,75] with a combinatorial objective (object→image and object→text in parallel), and (b) the proposed method conducting a continual diffusion (object→image→text) with generative nature." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.789, + 0.51 + ], + "angle": 0, + "content": "To break through this barrier, open-vocabulary object detection (OVD) has been studied that leverages auxiliary linguistic cues to generalize the detection of novel classes [82]. Recent advances adapt CLIP [67], a vision-language (VL) foundation model with strong zero-shot capacity, to boost the novel-class generalization of the object detector. Some works [13, 14] utilize the CLIP prediction (image-level) to generate pseudo labels at the object level, enabling the self-training of the object detector. To fully explore the hidden knowledge, some other works strive to align the object-level representation with CLIP embeddings in the hidden space. During training the object detector, they minimize the object-image [15, 68, 75] and object-text [46, 68, 76] discrepancy to align the sub-space among the object, image, and text modalities, thereby unleashing the novel-class generalization capacity from the pre-trained CLIP model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.789, + 0.722 + ], + "angle": 0, + "content": "Despite the significant success, the existing OVD paradigm, as depicted in Fig. 1(a), has two limitations regarding the sub-space alignment. First, existing works utilize text embedding as the classifier weight to explicitly optimize the discriminative class separation, which is agnostic to the underlying distribution and may lead to class bias [18]. This distribution-agnostic property results in the distribution misalignment [40, 42, 46] and severely hinders the novel-class generalization [78]. Besides, due to the uncalibrated latent space in discriminative models [45], e.g., numerical high-variance, the object distribution is prone to overfit to the labeled classes, worsening the misalignment with the unbiased CLIP distribution. With the rise of generative paradigm [2], this barrier has been gradually broken via diverse probabilistic modeling [18, 45, 59]. Hence, we aim to incorporate probabilistic advantages into discriminative detection pipelines, which can facilitate the distribution transfer across the object and text sub-spaces, thereby achieving an unbiased novel-class perception [45, 59]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.723, + 0.79, + 0.8 + ], + "angle": 0, + "content": "Second, some existing works minimize the object-image and object-text discrepancy simultaneously in a combinatorial manner, ignoring the disagreement between the image and text two objectives3. With this combinatorial nature, optimizing object-image alignment as the same role as the object-text counterpart is suboptimal due to the potential conflict between the two objectives [25]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.841 + ], + "angle": 0, + "content": "3 As the first attempt, we discover the ever-overlooked disagreement issue and justify it in Appendix via experiments." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.235, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "Intuitively, the image and object share visual representation, while the image and text are coupled with CLIP pre-training, which motivates us to introduce the image as an interpolated sub-space [61] between the object and text counterpart, yielding a continual distribution transfer among three modalities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.79, + 0.404 + ], + "angle": 0, + "content": "To address the above issues, we propose Continual Latent Diffusion (CLIFF) for OVD, as shown in Fig. 1(b). CLIFF breaks through the discriminative bottleneck via a probabilistic Variational Latent Sampler (VLS) and then formulates a diffusion-based distribution transfer [21] in a Continual Diffusion Module (CDM) to avoid the combinatorial limitation. Specifically, VLS leverages variational approximation [24] to establish a probabilistic object space with learned distribution parameters, and then samples object-centric noise to boost the diffusion process. CDM formulates the continual diffusion process to transfer the object distribution to the image and text counterpart, avoiding disagreement in the existing combinatorial objective. Besides, different from existing diffusion-based works [27, 78] with unsatisfactory efficiency, CLIFF achieves comparable and even better model efficiency compared with state-of-the-art OVD works, which is verified in Appendix. To be summarized, the contributions are four-fold:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.415, + 0.787, + 0.475 + ], + "angle": 0, + "content": "- We propose Continual Latent Diffusion (CLIFF) for OVD, which achieves a continual distribution transfer with DDPM-based generation. To the best of our knowledge, this work represents the first attempt to study latent diffusion in VL discriminative tasks and push forward benchmark results." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.476, + 0.787, + 0.52 + ], + "angle": 0, + "content": "- Variational Latent Sampler is proposed to generate a probabilistic object space with variational modeling, in which the object-centric noise can be sampled with reparameterization to boost the generation." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.521, + 0.787, + 0.566 + ], + "angle": 0, + "content": "- We design an efficient Continual Diffusion Module to facilitate the distribution transfer among three cross-modality sub-spaces, namely the object, image, and text space, using a new diffusion module and unified DDPM formulation." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.566, + 0.787, + 0.596 + ], + "angle": 0, + "content": "- CLIFF is extensively validated through different benchmarks, which verify its state-of-the-art role and show great adaptability to various detectors." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.415, + 0.787, + 0.596 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.619, + 0.398, + 0.635 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.65, + 0.553, + 0.666 + ], + "angle": 0, + "content": "2.1 Open-Vocabulary Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Open-vocabulary object detection (OVD) leverages auxiliary image-level labels with linguistic cues, e.g., classification [68] and caption datasets [80], to enlarge the semantic space of the region proposals [5, 69], thereby facilitating the novel-class generalization for the real-world applications, such as automatic driving [41, 43, 50, 52], medical imaging analysis [31, 35, 36, 49] and diagnostic decision-making [1, 7, 44, 53, 79]. Zareian et al. [82] formulate the standard OVD problem setting and propose OVR-CNN to project the visual space to the linguistic counterpart on Faster-RCNN [69]. Some other works [22, 23, 60] delve into OVD with Vision Transformer (ViT) by designing large-scale region-aware pretraining [23], feature masking [22], which fully unleash the potential of ViT in fine-grained detection tasks. Recent works adapt CLIP [67] to objects detectors for the novel-class" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.326, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.375 + ], + "angle": 0, + "content": "generalization. The other works adapt the CLIP-encoded latent embedding and aim to align three critical sub-spaces: the object space, the CLIP image space, and the CLIP text space. They bridge the gap between the object and text sub-space via region-text contrastive pre-training [85], region-word bipartite matching [46], CLIP-based conditional matching [81], and learnable detection-friendly prompts [9, 11, 12]. PromptDet [12] and DetPro [9] use learnable text prompts to learn with noisy uncurated web images, thereby enhancing the object-text alignment. Additionally, recent advances [15, 68, 75] adapt the image sub-space with the embedding distillation [15], weight transfer [68] and bags of regions [75], to fully use CLIP embedding, serving as a combinatorial objective with the object-text alignment. Existing works follow a similar discriminative paradigm for OVD, which cannot model the underlying distribution with a suboptimal distribution alignment. Differently, we formulate a continual transfer in a unified denoising diffusion module, enabling the distribution-level perception and adaptation among the object, image, and text sub-spaces." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.396, + 0.593, + 0.412 + ], + "angle": 0, + "content": "2.2 Denoising Diffusion Probabilistic Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.789, + 0.587 + ], + "angle": 0, + "content": "As the advanced generative paradigm, Denoising Diffusion Probabilistic Model (DDPM) [21] (\"diffusion model\" for short) generates samples by formulating a gradual denoising process. This process can be treated as a distribution transfer [51] from the normal distribution to the data in a flow-like manner [51, 54, 84]. Instead of diffusing in the data space, LDM [70] and VDM [66] deploy the denoising diffusion in the latent space to enable the high-resolution synthesis. Recent works can be generally categorized into faster sampling [55, 73], noise corruption designs [3, 62], and diffusion model architecture [33, 34, 65], etc. The diffusion model has achieved remarkable generative performance in the image, text, video, audio, graph, and motion generation [21, 28, 30, 32, 74] with different diffusion conditions for downstream applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.588, + 0.789, + 0.739 + ], + "angle": 0, + "content": "In addition to generative tasks, recent works [6,16,27,56,78,83] have discovered that the diffusion feature has adequate semantic discriminability with richer distribution-level perception and data-efficient property [4,64] than conventional discriminative models [67]. Without fine-tuning, an off-the-shelled diffusion model can achieve satisfactory zero-shot classification [29]. Along this trend, some bottlenecks in discriminative models are gradually broken by involving the generative paradigm [78]. CARD [19] uses the diffusion model to estimate the uncertainty to improve image classification. Unlike existing works considering the noise \\(\\rightarrow\\) data diffusion, we formulate a continual distribution transfer among three modalities, yielding the object \\(\\rightarrow\\) image \\(\\rightarrow\\) text transfer in the latent space." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.762, + 0.383, + 0.779 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.793, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Problem Formulation. Assuming a set of base classes \\(\\mathcal{C}_B\\) and novel classes \\(\\mathcal{C}_N\\), we have a detection dataset \\(\\mathcal{D}_{det} = \\{I_i,O_i\\}_{i = 1}^n\\) with object-level labels \\(\\{O_i\\}_{i = 1}^n\\) in \\(\\mathcal{C}_B\\) and another auxiliary dataset \\(\\mathcal{D}_{aux} = \\{I_i,Y_i\\}_{i = 1}^{n'}\\) with image-level" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.235, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "code_caption", + "bbox": [ + 0.217, + 0.147, + 0.724, + 0.163 + ], + "angle": 0, + "content": "Algorithm 1 Reverse Diffusion Process (Denoising): RevDif \\((\\mathbf{x}_T,\\mathbf{c},L)\\)" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.167, + 0.741, + 0.182 + ], + "angle": 0, + "content": "Require: \\(\\mathbf{x}_T\\): initial noise; c: diffusion condition; \\(L\\): the number reverse steps." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.182, + 0.543, + 0.195 + ], + "angle": 0, + "content": "Ensure: \\(\\mathbf{x}_{T - L}\\) generated samples with \\(L\\) steps." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.196, + 0.424, + 0.208 + ], + "angle": 0, + "content": "1: for \\( t = T, \\dots, T - L + 1 \\) do" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.21, + 0.339, + 0.222 + ], + "angle": 0, + "content": "2: \\(\\mathbf{z} \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\)" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.223, + 0.546, + 0.244 + ], + "angle": 0, + "content": "3: \\(\\mathbf{x}_{t - 1} = \\frac{1}{\\sqrt{\\alpha_t}}\\left(\\mathbf{x}_t - \\frac{1 - \\alpha_t}{\\sqrt{1 - \\bar{\\alpha}_t}}\\pmb {\\epsilon}_\\theta (\\mathbf{x}_t,\\mathbf{c},t)\\right) + \\sigma_t\\mathbf{z}\\)" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.244, + 0.3, + 0.255 + ], + "angle": 0, + "content": "4: end for" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.258, + 0.34, + 0.27 + ], + "angle": 0, + "content": "5: return \\(\\mathbf{x}_{T - L}\\)" + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.196, + 0.546, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.787, + 0.375 + ], + "angle": 0, + "content": "labels, e.g., the image caption [21] and classification dataset [68], which has the weakly supervised labels in \\(\\mathcal{C}_B \\cup \\mathcal{C}_N\\). In \\(\\mathcal{D}_{det}\\), each image \\(I_i\\) is associated with a set of objects \\(O_i = \\{(x,y,w,h), \\hat{y}\\}\\), including the bounding-box coordinates \\(\\{(x,y,w,h)\\}\\) and class labels \\(\\{\\hat{y}\\} \\subset \\mathcal{C}_B\\). OVD aims to use \\(\\mathcal{D}_{det}\\) and \\(\\mathcal{D}_{aux}\\) for training and detect base and novel objects in \\(\\mathcal{C}_B \\cup \\mathcal{C}_N\\) during the test stage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.788, + 0.451 + ], + "angle": 0, + "content": "Diffusion Model. Formulating a generative paradigm with the Markov chain [21, 63], DDPM is able to generate data \\(\\mathbf{x}_0\\) from the noise \\(\\mathbf{x}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\) via a \\(T\\)-step denoising process. Concretely, DDPM begins with the clean data \\(\\mathbf{x}_0\\) and gradually adds Gaussian noise using the fixed forward process \\(q(\\mathbf{x}_t | \\mathbf{x}_{t-1})\\), where \\(t = 1, 2, \\dots, T\\). This process can be denoted as follows," + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.457, + 0.788, + 0.475 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.787, + 0.511 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_t\\) is noise schedule [62]. To recover data \\(\\mathbf{x}_0\\) from noise \\(\\mathbf{x}_T\\), the reverse process parameterized with \\(\\theta\\) is learned to denoise inputs with condition \\(\\mathbf{c}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.517, + 0.787, + 0.557 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\mathbf {x} _ {0: T} \\mid \\mathbf {c}\\right) = p \\left(\\mathbf {x} _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}, \\mathbf {c}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.638 + ], + "angle": 0, + "content": "where \\(p(\\mathbf{x}_T)\\) is fixed to Gaussian. In this reverse step, the diffusion model typically learns a neural network \\(\\epsilon_{\\theta}(\\cdot)\\) to generate the values for computing the mean of \\(p_{\\theta}(\\mathbf{x}_{t - 1}\\mid \\mathbf{x}_t,\\mathbf{c})\\) given by the noisy input \\(\\mathbf{x}_t\\), which is optimized to minimize the variational lower bound (ELBO). As proved in [21, 29], this training objective can be re-written as:" + }, + { + "type": "equation", + "bbox": [ + 0.38, + 0.642, + 0.787, + 0.668 + ], + "angle": 0, + "content": "\\[\n- \\mathbb {E} _ {t, \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, \\mathbf {c}, t) \\| ^ {2} \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.764 + ], + "angle": 0, + "content": "where the diffusion model \\(\\epsilon_{\\theta}(\\cdot)\\) can estimate the added noise of each noisy input \\(\\mathbf{x}_t\\). After optimizing with Eq. 3, the diffusion model can generate a data sample \\(\\mathbf{x}_{T - L}\\) at arbitrary time-step \\(T - L\\) from a random noise \\(\\mathbf{x}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\) with \\(L\\) steps of the reverse diffusion, optionally guided by the condition \\(\\mathbf{c}\\). This process [21] is represented in Algorithm 1. In the following sections, we use RevDif(\\(\\cdot\\)) to represent the reverse diffusion (noise→data) for simplicity." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.782, + 0.381, + 0.801 + ], + "angle": 0, + "content": "4 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.809, + 0.787, + 0.841 + ], + "angle": 0, + "content": "The overall workflow is illustrated in Fig. 2. We feed input images \\(\\{I_i\\}_{i=1}^B\\) to the backbone and RPN [69] to extract \\(N\\) region proposals, which are sent to the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.327, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.15, + 0.785, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.274, + 0.788, + 0.318 + ], + "angle": 0, + "content": "Fig. 2: Illustration of the proposed CLIFF. VLS establishes a probabilistic embedding space with region proposals, where we sampled object-centric noise \\(\\mathbf{x}_T\\). Then, CDM achieves a continual distribution transfer from the object to the image and text." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.789, + 0.516 + ], + "angle": 0, + "content": "feature extractor to generate object embedding \\(\\mathbf{x}_{\\mathrm{obj}}\\). Then, VLS, parameterized with \\(\\phi\\), establishes and estimates the mean and variance of a probabilistic object space \\(p_{\\phi}(\\mathbf{x}_T|\\mathbf{x}_{\\mathrm{obj}})\\), where the reparameterization is conducted to sample object-centric noise \\(\\mathbf{x}_T\\), as shown in Fig. 2(a). After that, the sampled noise \\(\\mathbf{x}_T\\) is sent to CDM, denoted as \\(\\epsilon_{\\theta}(\\cdot)\\), to generate text embedding \\(\\mathbf{x}_{\\mathrm{txt}}\\) with diffusion, as depicted in Fig. 2(b). Conditioned on \\(\\mathbf{c}_{o2i}\\), we first conduct object-to-image diffusion to generate the image embedding \\(\\mathbf{x}_{\\mathrm{img}}\\) with \\(L\\) steps (top sub-figure), which is used to generate refined diffusion condition \\(\\mathbf{c}_{o2t}\\). Conditioned on \\(\\mathbf{c}_{o2t}\\), we further perform a \\(T\\)-step object-to-text diffusion to generate text embedding \\(\\mathbf{x}_{\\mathrm{txt}}\\) (bottom sub-figure), whose similarity with the CLIP text embedding \\(\\hat{\\mathbf{x}}_{\\mathrm{txt}}\\) are measured to serve as the final class prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.517, + 0.79, + 0.58 + ], + "angle": 0, + "content": "Formally, by extending Eq. 2, CLIFF consists of two diffusion processes among the object, image, and text embedding sub-space. Guided by the condition \\(\\mathbf{c}_{o2i}\\), the object-to-image diffusion, written as \\(p_{\\theta,\\phi}(\\mathbf{x}_{\\mathrm{img}}|\\mathbf{c}_{o2i},\\mathbf{x}_{\\mathrm{obj}})\\), aims to generate the associated CLIP image embedding from the object \\(\\mathbf{x}_{\\mathrm{obj}}\\) with \\(L\\) steps:" + }, + { + "type": "equation", + "bbox": [ + 0.278, + 0.594, + 0.788, + 0.637 + ], + "angle": 0, + "content": "\\[\np _ {\\theta , \\phi} (\\mathbf {x} _ {\\mathrm {i m g}} | \\mathbf {c} _ {o 2 i}, \\mathbf {x} _ {\\mathrm {o b j}}) = p _ {\\phi} (\\mathbf {x} _ {T} | \\mathbf {x} _ {\\mathrm {o b j}}) \\prod_ {t = T - L + 1} ^ {T} p _ {\\theta} (\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}, \\mathbf {c} _ {o 2 i}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.65, + 0.788, + 0.681 + ], + "angle": 0, + "content": "Similarly, with condition \\(\\mathbf{c}_{o2t}\\), the object-to-text diffusion is designed to generate text embedding with \\(T\\) steps for the final prediction, denoted as follows," + }, + { + "type": "equation", + "bbox": [ + 0.299, + 0.695, + 0.788, + 0.737 + ], + "angle": 0, + "content": "\\[\np _ {\\theta , \\phi} (\\mathbf {x} _ {\\mathrm {t x t}} | \\mathbf {c} _ {o 2 t}, \\mathbf {x} _ {\\mathrm {o b j}}) = p _ {\\phi} (\\mathbf {x} _ {T} | \\mathbf {x} _ {\\mathrm {o b j}}) \\prod_ {t = 1} ^ {T} p _ {\\theta} (\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}, \\mathbf {c} _ {o 2 t}). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.841 + ], + "angle": 0, + "content": "The proposed CLIFF offers two significant improvements compared to the standard diffusion process (Eq. 2). First, instead of beginning from a random noise space \\( p(\\mathbf{x}_T) \\), we start from a differential probabilistic space \\( p_{\\phi}(\\mathbf{x}_T|\\mathbf{x}_{\\mathrm{obj}}) \\) with objects to generate the text embedding. Second, going beyond a single diffusion between two sub-spaces, we formulate the effective diffusion process among the object, image, and text in different modalities." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.235, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.486, + 0.163 + ], + "angle": 0, + "content": "4.1 Variational Latent Sampler" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.167, + 0.788, + 0.288 + ], + "angle": 0, + "content": "With input images \\(\\{I_i\\}_{i=1}^B\\), the backbone, RPN4, and the feature extractor are adopted to obtain \\(N\\) object embeddings \\(\\mathbf{x}_{\\mathrm{obj}} \\in \\mathbb{R}^{N \\times D}\\). Then, we send \\(\\mathbf{x}_{\\mathrm{obj}}\\) to VLS to establish a probabilistic embedding space \\(p_{\\phi}(\\mathbf{x}_T | \\mathbf{x}_{\\mathrm{obj}})\\) satisfying the DDPM w.r.t the normally distributed assumption for the diffusion-based transfer. This probabilistic space \\(p_{\\phi}(\\mathbf{x}_T | \\mathbf{x}_{\\mathrm{obj}})\\) can break through the discriminative limitation of [14,15,46,75] without class bias [18]. Besides, it introduces object-level semantics to enable object-to-text transfer with a differential nature, different from the non-optimizable random space \\(p(\\mathbf{x}_T)\\) of noise-to-data transfer [21]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.288, + 0.79, + 0.444 + ], + "angle": 0, + "content": "Formally, inspired by VAE [24], we establish the probabilistic space with centered isotropic multivariate Gaussian \\( p_{\\phi}(\\mathbf{x}_T|\\mathbf{x}_{\\mathrm{obj}}) = \\mathcal{N}(\\mathbf{x}_T;\\mathbf{0},\\mathbf{I}) \\), whose distribution parameters are computed from the observed object embedding \\( \\mathbf{x}_{\\mathrm{obj}} \\) with non-linear layers [18]. To this end, two fully connected layers are introduced on the shared region feature extractor of the basic object detector, denoted as \\( f_{\\mathrm{obj}} \\) e.g., the ResNet blocks in Faster RCNN [69] and the MLP in Mask RCNN [20], to learn to estimate the distribution parameters. Specifically, given the object embedding \\( \\mathbf{x}_{\\mathrm{obj}} \\), the variational posterior \\( p_{\\phi}(\\mathbf{x}_T|\\mathbf{x}_{\\mathrm{obj}}) \\) is written as \\( \\log p_{\\phi}(\\mathbf{x}_T|\\mathbf{x}_{\\mathrm{obj}}^{(i)}) = \\log \\mathcal{N}(\\mathbf{x}_T;\\pmb{\\mu}^{(i)},\\pmb{\\sigma}^{2(i)}\\mathbf{I}) \\). Without losing the differential property [24], the mean and s.d. of the probabilistic object space can be estimated," + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.461, + 0.786, + 0.49 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} _ {(i)} ^ {(i)} = \\mathbf {W} _ {\\mu} f _ {\\mathrm {o b j}} \\left(\\mathbf {x} _ {\\mathrm {o b j}} ^ {(i)}\\right) + \\mathbf {b} _ {\\mu}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.486, + 0.597, + 0.505 + ], + "angle": 0, + "content": "\\[\n\\pmb {\\sigma} ^ {(i)} = \\mathbf {W} _ {\\sigma} f _ {\\mathrm {o b j}} (\\mathbf {x} _ {\\mathrm {o b j}} ^ {(i)}) + \\mathbf {b} _ {\\sigma},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.572 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_{(\\cdot)}\\) and \\(\\mathbf{b}_{(\\cdot)}\\) are learned weight and bias. Then, with the probabilistic modeling, we sample object-centric noise from the latent object space \\(\\mathbf{x}_T \\sim p_{\\phi}(\\mathbf{x}_T | \\mathbf{x}_{\\mathrm{obj}})\\) with reparameterization [24]: \\(\\mathbf{x}_T^{(i)} = \\boldsymbol{\\mu}^{(i)} + \\boldsymbol{\\sigma}^{(i)} \\odot \\boldsymbol{\\epsilon}\\), where \\(\\odot\\) is the element-wise product and \\(\\boldsymbol{\\epsilon} \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.788, + 0.693 + ], + "angle": 0, + "content": "To optimize the probabilistic space, we deploy Kullback-Leibler (KL) divergence \\( D_{\\mathrm{KL}} \\left(p_{\\phi}(\\mathbf{x}_T|\\mathbf{x}_{\\mathrm{obj}})||p(\\mathbf{x}_T)\\right) \\) with \\( \\mathbf{x}_T \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}) \\) that encourages the Gaussian distributed [24], which has two advantages. First, the probabilistic nature of normal Gaussian removes the class bias within the object space [18] and promotes the generalization capacity towards discovering novel classes [45, 59]. Second, it naturally satisfies the DDPM constraint (Eq. 1) with the Gaussian formulation [21] for the following diffusion process. The implemented KL optimization objective, as proved in the Appendix, can be re-written in the following format," + }, + { + "type": "equation", + "bbox": [ + 0.313, + 0.699, + 0.787, + 0.741 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {K L} = - \\frac {1}{2 N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {D} \\left(1 + \\log \\left(\\boldsymbol {\\sigma} _ {j} ^ {2 (i)}\\right) - \\boldsymbol {\\mu} _ {j} ^ {2 (i)} - \\boldsymbol {\\sigma} _ {j} ^ {2 (i)}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.745, + 0.788, + 0.793 + ], + "angle": 0, + "content": "VLS can establish a probabilistic space with differential property and sample object-centric noise \\(\\mathbf{x}_T\\) for the diffusion process, serving to generate probabilistic object-text pairs in the diffusion-based distribution transfer [48,51]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.797, + 0.788, + 0.842 + ], + "angle": 0, + "content": "4 Besides \\(N_{rpn}\\) RPN proposals, we follow [68, 81] to add additional proposals for training, including \\(N_{cs}\\) class-specific proposals (from \\(\\mathcal{D}_{aux}\\)) and \\(N_{ca}\\) class-agnostic proposals (from \\(\\mathcal{D}_{det}\\)) from [58]. The number of proposals is \\(N = N_{rpn} + N_{cs} + N_{ca}\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.327, + 0.128 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.492, + 0.162 + ], + "angle": 0, + "content": "4.2 Continual Diffusion Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.232 + ], + "angle": 0, + "content": "With object-centric noise \\(\\mathbf{x}_T\\), we further deploy the object-to-image (o2i) and object-to-text (o2t) diffusion to achieve a continual distribution transfer for OVD. To achieve this, an efficient diffusion module \\(\\epsilon_{\\theta}(\\cdot)\\) with MLP architecture is designed, the details of which are presented in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.233, + 0.79, + 0.324 + ], + "angle": 0, + "content": "Object-to-image Diffusion. Given class-agnostic region proposals, we send them to the object detector to generate the object embedding \\(\\mathbf{x}_{\\mathrm{obj}}\\) and feed them into frozen CLIP image encoder [67] to obtain static image embedding \\(\\hat{\\mathbf{x}}_{\\mathrm{img}}\\). Then, with a randomly sampled object-centric noise \\(\\mathbf{x}_T \\sim p_\\phi(\\mathbf{x}_T | \\mathbf{x}_{\\mathrm{obj}})\\) from VLS, we encourage the diffusion module to generate the image embedding \\(\\mathbf{x}_{\\mathrm{img}}\\) with object-level condition \\(\\mathbf{c}_{o2i} = \\mathbf{x}_{\\mathrm{obj}}\\) using a \\(L\\)-step reverse diffusion:" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.336, + 0.788, + 0.351 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {\\mathrm {i m g}} = \\operatorname {R e v D i f} \\left(\\mathbf {x} _ {T}, \\mathbf {c} _ {o 2 i}, L\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.362, + 0.796, + 0.44 + ], + "angle": 0, + "content": "where \\(\\mathrm{RevDif}(\\cdot)\\) (Algorithm 1) recovers image embedding from object-centric noise \\(\\mathbf{x}_T\\). To this end, the object-to-image diffusion, denoted as \\(p_{\\theta,\\phi}(\\mathbf{x}_{\\mathrm{img}}|\\mathbf{c}_{o2i},\\mathbf{x}_{\\mathrm{obj}})\\), uses object embedding as the diffusion condition \\(\\mathbf{c}_{o2i} = \\mathbf{x}_{\\mathrm{obj}}\\), and is optimized by minimizing the distance between the generated embedding \\(\\mathbf{x}_{\\mathrm{img}}\\) and CLIP image embedding \\(\\hat{\\mathbf{x}}_{\\mathrm{img}}\\), which is achieved via the following loss function," + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.451, + 0.788, + 0.49 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {o 2 i} = \\frac {1}{N _ {c a}} \\sum_ {i = 1} ^ {N _ {c a}} | | \\mathbf {x} _ {\\mathrm {i m g}} ^ {(i)} - \\hat {\\mathbf {x}} _ {\\mathrm {i m g}} ^ {(i)} | | _ {1}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.5, + 0.789, + 0.546 + ], + "angle": 0, + "content": "where \\(N_{ca}\\) is the number of class-agnostic proposals. With the optimized diffusion process, the object distribution could be transferred to the CLIP image counterpart, facilitating the adaptation to the CLIP latent space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.789, + 0.683 + ], + "angle": 0, + "content": "Object-to-text Diffusion. With diffusion-based formulation, we further encourage the model to reconstruct the associated text embedding \\(\\mathbf{x}_{\\mathrm{txt}}\\) for region proposals \\(\\mathbf{x}_{\\mathrm{obj}}\\), which conducts a generative procedure in latent space. Note that this generative objective is different from existing OVD works that directly optimize the discriminative class separation. Specifically, we design the object-to-text diffusion with a \\(T\\)-step diffusion \\(p_{\\theta,\\phi}(\\mathbf{x}_{\\mathrm{txt}}|\\mathbf{c}_{o2t},\\mathbf{x}_{\\mathrm{obj}})\\) guided by condition \\(\\mathbf{c}_{o2t}\\). Considering the dominant role of condition [21] in the generated content, we propose a simple yet effective attentive mechanism to obtain the refined condition \\(\\mathbf{c}_{o2t} = \\bar{\\mathbf{x}}_{\\mathrm{obj}}^{\\mathrm{img}}\\) with the generated image image-level context:" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.696, + 0.788, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {\\mathrm {o b j}} ^ {\\mathrm {i m g}} = \\operatorname {S i g m o i d} \\left(\\operatorname {F C} \\left(\\mathbf {x} _ {\\mathrm {i m g}}\\right)\\right) \\odot \\mathbf {x} _ {\\mathrm {o b j}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.724, + 0.788, + 0.772 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_{\\mathrm{img}}\\) is generated via Eq. 8, \\(\\odot\\) is the element-wise product, and FC is a liner projection. Then, we use the refined condition \\(\\mathbf{c}_{o2t} = \\mathbf{x}_{\\mathrm{obj}}^{\\mathrm{img}}\\) to guide the text embedding generation for \\(\\mathbf{x}_{\\mathrm{txt}}\\) with a \\(T\\)-step reverse diffusion:" + }, + { + "type": "equation", + "bbox": [ + 0.406, + 0.785, + 0.788, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {\\mathrm {t x t}} = \\operatorname {R e v D i f} \\left(\\mathbf {x} _ {T}, \\mathbf {c} _ {o 2 t}, T\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Considering the mutual benefits of generative and discriminative objectives for model learning as explored in [45], the loss function for the object-to-text" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.235, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.178 + ], + "angle": 0, + "content": "diffusion process is designed with these two distinct components. To this end, the optimization objective can be expressed as follows," + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.192, + 0.787, + 0.265 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {o 2 t} = \\underbrace {\\frac {1}{N _ {l}} \\sum_ {i} ^ {N _ {l}} | | \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\mathbf {x} _ {t} ^ {(i)} , \\mathbf {c} _ {o 2 t} ^ {(i)} , t\\right) - \\boldsymbol {\\epsilon} | | _ {1}} _ {\\text {G e n e r a t i v e}} + \\underbrace {\\mathcal {L} _ {C E} \\left(\\operatorname {S i g m o i d} \\left(\\frac {1}{\\tau} \\cos \\left(\\mathbf {x} _ {\\mathrm {t x t}}, \\hat {\\mathbf {x}} _ {\\mathrm {t x t}}\\right)\\right), \\hat {y}\\right)} _ {\\text {D i s c r i m i n a t i v e}} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.266, + 0.789, + 0.371 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\), \\(\\mathbf{x}_t\\) is the noisy CLIP text embedding at a randomly sampled time-step \\(t \\sim \\mathcal{U}[1,T]\\), \\(\\tau = 0.02\\) is a scaling factor as [15, 68], \\(\\hat{y}\\) is the class label, \\(\\mathcal{L}_{CE}\\) is the Cross-Entropy loss, and \\(N_{l} = N_{cs} + N_{rpn}\\) counts the labeled proposals. The first term optimizes the diffusion module \\(\\epsilon_{\\theta}(\\cdot)\\) to reconstruct the text embedding in a generative manner [21], serving for the objective of Eq. 3. The second term can give an interpretable score for NMS [69] in detectors and ensure the generated embedding with sufficient discriminability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.789, + 0.495 + ], + "angle": 0, + "content": "Remark. We use a shared diffusion module with a \\(T\\)-step diffusion formulation to optimize two processes, where the first \\(L\\) steps are assigned for object-to-image, and the whole \\(T\\) steps are for object-to-text diffusion. Note that even though we do not explicitly optimize image-to-text diffusion, this transition can be intrinsically achieved between the \\(L\\)-th and \\(T\\)-th steps, as illustrated by the gray-colored part in Fig. 1(b). The reason is based on the continual property of the Markov chain in DDPM, i.e., formulating a random process that undergoes transitions from one state to another within a finite number of possible states [21, 63]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.526, + 0.452, + 0.541 + ], + "angle": 0, + "content": "4.3 Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.561, + 0.788, + 0.593 + ], + "angle": 0, + "content": "Training Stage. We implement the overall optimization objective on the basic object detector, Faster RCNN [69], to train the whole framework:" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.608, + 0.788, + 0.625 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C L I F F} = \\mathcal {L} _ {r p n} + \\mathcal {L} _ {r e g} + \\alpha \\mathcal {L} _ {K L} + \\beta_ {1} \\mathcal {L} _ {o 2 i} + \\beta_ {2} \\mathcal {L} _ {o 2 t}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.639, + 0.788, + 0.669 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{rpn}\\) and \\(\\mathcal{L}_{reg}\\) are RPN loss and the bounding box regression loss in the RCNN head [69] borrowed from object detectors. \\(\\alpha, \\beta_1, \\beta_2\\) are the loss weight." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.671, + 0.79, + 0.718 + ], + "angle": 0, + "content": "Inference Stage. We first extract the region proposals \\(\\mathbf{x}_{\\mathrm{obj}}\\) and then conduct reverse diffusion to generate the associated text embedding \\(\\mathbf{x}_{\\mathrm{txt}}\\) for region proposals, whose similarity with \\(\\hat{\\mathbf{x}}_{\\mathrm{txt}}\\) is measured for class prediction \\(\\tilde{y}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.732, + 0.788, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\tilde {y} = \\operatorname {S i g m o i d} \\left(\\frac {1}{\\tau} c o s \\left(\\mathbf {x} _ {\\mathrm {t x t}}, \\hat {\\mathbf {x}} _ {\\mathrm {t x t}}\\right)\\right), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.778, + 0.788, + 0.809 + ], + "angle": 0, + "content": "where \\(\\tilde{y}\\) is the per-class confidence, \\(\\mathbf{x}_{\\mathrm{txt}}\\) is the generated text embedding via object-to-text diffusion (Eq. 11), and \\(\\hat{\\mathbf{x}}_{\\mathrm{txt}}\\) is CLIP text embedding for all classes." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.824, + 0.567, + 0.842 + ], + "angle": 0, + "content": "5 Following Eq. 1, \\(\\mathbf{x}_t = \\sqrt{\\bar{\\alpha}_t}\\hat{\\mathbf{x}}_{\\mathrm{xt}} + (1 - \\bar{\\alpha}_t)\\epsilon\\) as [21]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.326, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 1: Experimental results (%) on COCO benchmark. † indicates the reproduced results with the consistent baseline setting as [68], including ImageNet pre-training and CLIP prompting. ‡ indicates using the longer training schedule [68]." + }, + { + "type": "table", + "bbox": [ + 0.23, + 0.2, + 0.776, + 0.387 + ], + "angle": 0, + "content": "
MethodObject DetectorSub-space Alignment\\( \\mathrm{mAP}_{\\mathrm{n}} \\)\\( \\mathrm{mAP}_{\\mathrm{b}} \\)mAPReference
OVR-CNN [82]Faster RCNN C4Obj → Txt22.846.039.9CVPR-2021
RegionCLIP [85]Faster RCNN C4Obj → Txt & Img26.854.847.5CVPR-2022
ViLD [15]Faster RCNN C4Obj → Txt & Img27.659.551.3ICLR-2022
OV-DETR [81]Deformable DETRObj → Txt & Img29.461.052.7ECCV-2022
Detic [86]Faster RCNN C4Obj → Txt27.847.145.0ECCV-2022
Detic‡ [86]Faster RCNN C4Obj → Txt28.453.847.2ECCV-2022
PB-OVD [14]Faster RCNN C4Obj → Txt30.846.142.1ECCV-2022
MAVL-OVD [58]Faster RCNN C4Obj → Txt30.452.646.8ECCV-2022
\\( OCD^‡ \\)[68]Faster RCNN C4Obj → Txt & Img36.654.049.4NeurIPS-2022
F-VLM [26]Faster RCNN FPNObj → Txt & Img28.0-39.6ICLR-2023
VLDet [46]Faster RCNN C4Obj → Txt32.050.645.8ICLR-2023
BARON†,‡ [75]Faster RCNN C4Obj → Txt & Img41.053.350.1CVPR-2023
EdaDet [71]Faster RCNN C4Obj → Txt37.857.752.5ICCV-2023
GOAT [72]Faster RCNN C4Obj → Txt & Img36.453.048.6ICCV-2023
CLIFF (ours)Faster RCNN C4Obj → Image → Txt41.354.150.6-
CLIFF‡ (ours)Faster RCNN C4Obj → Image → Txt43.254.051.2-
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.413, + 0.377, + 0.431 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.446, + 0.533, + 0.461 + ], + "angle": 0, + "content": "5.1 Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.788, + 0.563 + ], + "angle": 0, + "content": "We follow existing works [68, 75] to compare with state-of-the-art methods. COCO Setting. The COCO-2017 dataset [47] is used for training and evaluation, containing 80 object-level classes. As for the auxiliary dataset \\( \\mathcal{D}_{aux} \\), we follow [82] to use COCO caption dataset in \\( \\mathcal{C}_B \\cup \\mathcal{C}_N \\). The core evaluation metric is the novel-class Average Precision (AP) with an IoU threshold of 0.5, denoted as \\( \\mathrm{mAP_n} \\), for the novel-class generalization. The results are shown in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.623 + ], + "angle": 0, + "content": "LVIS Setting. LVIS v1.0 [17] contains 1203 classes and is split into frequent, common, and rare types. For the auxiliary datasets \\(\\mathcal{D}_{aux}\\), we follow [68] to use the subset of ImageNet-21K. The core evaluation metric is mask AP of rare classes, denoted as \\(\\mathrm{AP_r^m}\\). We report the comparison results in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.788, + 0.67 + ], + "angle": 0, + "content": "Pascal VOC Setting. We follow [72] to evaluate the generalization to unseen datasets with Pascal VOC [10] and LVIS [17] using COCO-trained model. We use \\(\\mathrm{AP}_{50}\\) to evaluate the performance as [72]. The results are given in Table 7." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.693, + 0.457, + 0.709 + ], + "angle": 0, + "content": "5.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Object Detector. Considering that previous works use inconsistent detectors, we strictly follow the setting of [68,82] in different benchmarks. Concretely, for the COCO setting, we use Faster-RCNN C4 [69] with the ResNet-50 backbone. The backbone is pre-trained on ImageNet [8] without bells and whistles. For the LVIS setting, we follow the two main streams of existing works by using Mask-RCNN-FPN [20] and CenterNet v2 [87] as the basic object detectors for a thorough comparison, which is based on ReseNet-50 backbone. For the experiments of all benchmarks, we use the SGD optimizer with a weight decay of \\(1e^{-4}\\) and a" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.235, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 2: Experimental results (\\%) on LVIS benchmark with the consistent object detector setting. Ens indicates using the ensemble trick [11]." + }, + { + "type": "table", + "bbox": [ + 0.28, + 0.185, + 0.725, + 0.355 + ], + "angle": 0, + "content": "
MethodObject DetectorAPmRAPmcAPmfAPReference
ViLD-Ens [15]Mask RCNN16.624.630.325.5ICLR-2022
MVAL-OVD [58]Mask RCNN17.021.226.122.4ECCV-2022
OCD [68]Mask RCNN21.125.029.125.9NeurIPS-2022
DetPro [9]Mask RCNN19.825.628.925.9CVPR-2022
RegionCLIP [85]Mask RCNN17.127.434.028.2CVPR-2022
F-VLM [26]Mask RCNN18.624.026.924.2ICLR-2023
BARON [75]Mask RCNN18.024.428.925.1CVPR-2023
BARON-Ens [75]Mask RCNN19.226.829.426.5CVPR-2023
CLIFF (ours)Mask RCNN22.426.929.826.8-
VLDet [46]CenterNet v221.729.834.330.1ICLR-2023
CoDet [57]CenterNet v223.430.034.630.7NeirIPS-2023
GOAT [72]CenterNet v223.329.734.330.4ICCV-2023
EdaDet [71]CenterNet v223.727.529.127.5ICCV-2023
CLIFF (ours)CenterNet v224.527.729.828.2-
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.788, + 0.442 + ], + "angle": 0, + "content": "momentum of 0.9 following [68]. We use 4 NVIDIA V100 for training with a batch size of 16. We pre-train the base object detector with a \\(1 \\times\\) schedule as [68], and then train the whole model with a \\(2 \\times\\) schedule. We also explore the longer \\(8 \\times\\) training schedule to make fair comparisons with specific methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.788, + 0.489 + ], + "angle": 0, + "content": "CLIP Model. As the previous work [68], we use the single prompt, 'a photo of a {category}', to compute CLIP text embedding \\(\\hat{\\mathbf{x}}_{\\mathrm{txt}}\\), and use \\(224\\times 224\\) input-scale for image embedding \\(\\hat{\\mathbf{x}}_{\\mathrm{img}}\\) via ViT-B/32 model [67]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.788, + 0.581 + ], + "angle": 0, + "content": "CLIFF. The number of the time-step in object-to-text \\((T)\\) and object-to-image \\((L)\\) diffusion are set \\(T = 10, L = 3\\), which are justified in Table 4. \\(\\alpha\\), \\(\\beta_{1}\\), and \\(\\beta_{2}\\) are empirically set to 2.0, 15.0, and 1.0, respectively, which are verified by the experiments in Appendix. \\(L_{2}\\) normalization is deployed on CLIP embedding as [15,68]. Class-specific and RPN proposals are used to optimize object-to-text diffusion [68,75], while class-agnostic ones are for object-to-image diffusion." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.603, + 0.622, + 0.619 + ], + "angle": 0, + "content": "5.3 Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Results on COCO Benchmark. As shown in Table 1, we compare the proposed method with state-of-the-art counterparts on the COCO benchmark. CLIFF achieves \\(41.3\\% \\mathrm{mAP_n}\\), which outperforms OCD [68] (\\(36.6\\% \\mathrm{mAP_n}\\)), VLDet [46] (\\(32.0\\% \\mathrm{mAP_n}\\)) and BARON [75] (\\(41.0\\% \\mathrm{mAP_n}\\)) with \\(4.7\\%\\), \\(9.3\\%\\) and \\(0.3\\% \\mathrm{mAP_n}\\), comprehensively verifying the effectiveness of the proposed CLIFF. The proposed method also surpasses EdaDet [71] and GOAT [72] by a noticeable margin. Besides, after using fair multi-scale training [68, 75], our method (\\(43.2\\% \\mathrm{mAP_n}\\)) outperforms existing works by a large margin. Unlike existing works aligning object-text and object-image separately, we formulate a continual diffusion process among the three sub-spaces in a different learning paradigm." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Results on LVIS Benchmark. We present a thorough comparison with corresponding state-of-the-art methods using Mask RCNN and CenterNet v2 detectors, which is shown in Table 2. For Mask RCNN, it can be observed that CLIFF achieves the best \\(22.4\\%\\) \\(\\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{m}}\\), which surpasses state-of-the-art counterparts" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.326, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.22, + 0.145, + 0.782, + 0.159 + ], + "angle": 0, + "content": "Table 3: Ablation study results on COCO and LVIS benchmarks with \\(1 \\times\\) schedule." + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.172, + 0.78, + 0.259 + ], + "angle": 0, + "content": "
CDM o2t o2i VLSVLSCOCOLVIS-BoxLVIS-Mask
mAPnmAPbmAPAPbAPcAPbAPbAPmAPmAPcAPb
1(baseline)32.249.945.315.820.126.221.415.219.725.020.5
2✓36.649.546.116.220.927.422.615.820.125.721.6
3✓39.552.449.216.721.727.723.116.520.925.922.1
4✓38.853.349.317.321.927.823.316.721.126.122.2
5✓40.052.949.518.922.128.023.818.121.426.222.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.272, + 0.495, + 0.356 + ], + "angle": 0, + "content": "Table 4: Analysis on the diffusion step. o2i and o2t indicates the object-to-image and object-to-text diffusion process respectively. * indicates sharing the same diffusion procedure and optimizing o2i and o2t in a combinatorial style [15,68]." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.368, + 0.487, + 0.496 + ], + "angle": 0, + "content": "
o2io2t\\( \\mathrm{mAP}_{\\mathrm{n}} \\)\\( \\mathrm{mAP}_{\\mathrm{b}} \\)mAP
111040.253.450.0
231041.354.150.6
351041.353.750.3
471041.253.950.5
591041.453.550.4
63540.953.850.3
731041.354.150.6
832039.753.449.8
9101041.453.650.4
10*10*1039.952.849.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.276, + 0.789, + 0.359 + ], + "angle": 0, + "content": "Table 5: Analysis of the discriminative and generative terms in Eq. 12 by replacing the random noise (Rand) with object-centric noise (OC) in different time steps \\((2\\times\\) schedule). rows 7 and 8 indicate removing individual terms." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.373, + 0.787, + 0.491 + ], + "angle": 0, + "content": "
DiscriminativeGenerative\\( \\mathrm{mAP}_{\\mathrm{n}} \\)
TT-1: 0TT-1: 0
1RandRandRandRand39.2
2OCRandOCRand41.0
3RandOCRandOC40.4
4OCOCOCRand41.1
5OCRandOCOC41.3
6OCOCOCOC41.3
7w/o DiscriminativeOCOC38.2
8OCOCw/o Generative40.1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.789, + 0.646 + ], + "angle": 0, + "content": "BARON [75] \\((19.2\\% \\mathrm{AP_r^m})\\), OCD [68] \\((21.3\\% \\mathrm{AP_r^m})\\) with \\(3.1\\% \\mathrm{AP_r^m}\\) and \\(1.3\\% \\mathrm{AP_r^m}\\), respectively. Besides, our method surpasses the baseline model MAVL-OVD [58] with a significant \\(5.4\\% \\mathrm{AP_r^m}\\), verifying the effectiveness of our method. For the comparison with CenterNet v2, we can find that our method gives \\(24.5\\% \\mathrm{AP_r^m}\\) and also surpasses the latest published works noticeably, outperforming CoDet [57] and GOAT [72] with \\(1.1\\%\\) and \\(1.2\\%\\) gains. This can reveal that CLIFF is a generic method in terms of multi-modal alignment and can be deployed in different object detectors with great transferability and adaptability." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.665, + 0.442, + 0.681 + ], + "angle": 0, + "content": "5.4 Quantitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Ablation Study. The ablative results are shown in Table 3 with a \\(1 \\times\\) training schedule. The first line indicates the model trained with extra MAVL proposals [58, 68], which is a fair baseline model as [68, 75]. Compared with the baseline model \\((32.2\\% \\mathrm{mAP}_{\\mathrm{n}}, 15.8\\% \\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{b}}, 15.2\\% \\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{m}})\\), introducing the object-to-text (row 2) diffusion gives satisfactory improvements with \\(36.6\\% \\mathrm{mAP}_{\\mathrm{n}}, 16.2\\% \\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{b}}\\) and \\(15.8\\% \\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{m}}\\). After integrating object-to-image diffusion (row 3), i.e., the full CDM, we can observe a further improvement with \\(7.3\\% \\mathrm{mAP}_{\\mathrm{n}}, 0.9\\% \\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{b}}\\) and \\(1.3\\% \\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{m}}\\) gains, compared with the baseline. This verifies the superior effectiveness in modeling a continual distribution transfer than the existing discriminative paradigm. Besides, after replacing the randomly sampled noise by" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.235, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.128 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.495, + 0.189 + ], + "angle": 0, + "content": "Table 6: Analysis of the refined condition \\( {\\mathbf{c}}_{o2t} \\) in object-to-image diffusion (Eq. 5) with \\( 2 \\times \\) training schedule." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.2, + 0.489, + 0.279 + ], + "angle": 0, + "content": "
Setting of c02tmAPnmAPbmAP
1Object Embed: xobj40.453.449.9
2Image Embed: ximg39.652.949.4
3Additive: xobj + ximg40.853.550.2
4Concated: [xobj : ximg]41.053.650.4
5Attentive: Eq. 1041.354.150.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.145, + 0.788, + 0.188 + ], + "angle": 0, + "content": "Table 7: Cross-dataset evaluation with \\(\\mathrm{AP}_{50}\\) (\\%) transferring COCO-trained models to the PASCAL VOC and LVIS." + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.2, + 0.776, + 0.279 + ], + "angle": 0, + "content": "
MethodPASCALLVIS
OVR-CNN [82]52.95.2
PB-OVD [14]59.28.0
VLDet [46]61.710.0
GOAT [72]63.614.0
CLIFF (Ours)66.116.7
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.308, + 0.787, + 0.354 + ], + "angle": 0, + "content": "object-centric noise with VLS (row 5), it can be observed that the CDM (row 3) can be improved with \\(40.0\\% \\mathrm{mAP}_{\\mathrm{n}}\\), \\(18.9\\% \\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{b}}\\), and \\(18.1\\% \\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{m}}\\), verifying the effectiveness of each component in the proposed CLIFF." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.789, + 0.536 + ], + "angle": 0, + "content": "The Number of Diffusion Steps. We study different diffusion steps for o2i and o2t in Table 4, and have the following three observations. First, in rows 1-5, we observe that the results are stable when \\( L > 3 \\) (around \\( 41.3\\% \\mathrm{mAP_n} \\)), but give notable performance drops when the step is insufficient \\( L = 1 \\). Considering more computation requirements when enlarging the diffusion step, we select \\( L = 3 \\) with the best trade-off. Second, in rows 6-8, setting too small and large o2t steps both give significant performance drops, which may be caused by unsatisfactory diffusion optimization. Third, we set consistent diffusion steps and compare the proposed continual diffusion (row 9) with existing combinatorial style, i.e., sharing the diffusion process and generated embedding between the o2i and o2t (row 10). With a \\( 1.5\\% \\) performance gain over the combinatorial objective, the effectiveness of our continual transfer can be justified." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.789, + 0.688 + ], + "angle": 0, + "content": "Exploring Object-centric Noise. We further give a detailed analysis of the object-centric (OC) noise by replacing the random noise for the discriminative and generative terms in object-to-text objective (the core process for OVD) in different time steps, as shown in Table 5. Compared with the random-noise baseline with \\(39.2\\% \\mathrm{mAP}_{\\mathrm{n}}\\), controlling the initial stage (row 2) and middle stages (row 3) gives \\(41.0\\%\\) are \\(40.4\\% \\mathrm{mAP}_{\\mathrm{n}}\\), respectively. Introducing more OC noise in the diffusion (rows 4-6) gives stable performance improvements, indicating the effectiveness of the VLS. Further, we conduct an ablative study on o2t diffusion (rows 7-8), i.e., the first and second terms in Eq. 12, verifying their critical role in optimizing the diffusion process with both aspects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.809 + ], + "angle": 0, + "content": "Analyzing the Condition Design. The strategy of refining the object-to-text (Eq. 5) condition \\(\\mathbf{c}_{o2t} = \\mathbf{x}_{\\mathrm{obj}}^{\\mathrm{img}}\\) is analyzed in Table 6. Compared with individually using the object and image (rows 1-2) condition with \\(40.4\\%\\) and \\(39.6\\%\\) \\(\\mathrm{mAP_n}\\), combining the two aspects for the diffusion condition generally works better (rows 3-5), e.g., addition with \\(40.8\\%\\) \\(\\mathrm{mAP_n}\\), concatenation with \\(41.0\\%\\). Besides, we observe that the proposed attentive refinement on the condition (Eq. 10) achieves optimal performance and gives the best \\(41.3\\%\\) \\(\\mathrm{mAP_n}\\), due to the adaptive fusion of image-level context. This can verify the effectiveness of our design." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Generalization to Unseen Datasets. We further conduct the cross-dataset comparison following [72] by considering two different settings. Compared with" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.326, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.144, + 0.78, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.351, + 0.788, + 0.396 + ], + "angle": 0, + "content": "Fig. 3: Qualitative comparison among (a) the baseline [68], (b) our CLIFF, and (c) ground-truth on the COCO benchmark. We only visualize the novel-class predictions for a clearer evaluation of open-vocabulary generalization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.788, + 0.482 + ], + "angle": 0, + "content": "state-of-the-art counterpart GOAT [72], the proposed CLIFF gives \\(66.1\\%\\) and \\(16.7\\% \\mathrm{AP}_{50}\\) on the PASCAL VOC and LVIS datasets using the COCO-trained model, which surpasses GOAT with significant \\(2.5\\%\\) and \\(2.7\\%\\) gains, showing the superior generalization capacity of our method." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.499, + 0.431, + 0.515 + ], + "angle": 0, + "content": "5.5 Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.789, + 0.673 + ], + "angle": 0, + "content": "As shown in Fig. 3, we visualize the novel-class prediction and conduct a qualitative comparison among (a) the baseline [68,75] trained with MAVL proposals [68], (b) the proposed CLIFF, and (c) ground-truth. It can be observed that the proposed method gives more accurate novel-class predictions, especially for occlusion situations and complex scenarios. Our method reduces the confusion among novel classes, e.g., correcting the wrong prediction (couch) given by the baseline in col. 1. Besides, the proposed CLIFF can relieve some missing errors, as illustrated in col. 2-5, revealing its superior novel-class generalization capacity. Hence, the proposed CLIFF is able to relieve the bias on the base classes and encourage the model to discover the novel-class objects with probabilistic modeling." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.691, + 0.36, + 0.708 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.789, + 0.841 + ], + "angle": 0, + "content": "This paper proposes a novel diffusion-based framework for OVD, coined CLIFF, which breaks through the bottlenecks of the combinatorial and discriminative limitation in existing OVD pipelines. The proposed CLIFF consists of a Variational Latent Sampler (VLS) to sample object-centric noise in the probabilistic object space, and leverages a Continual diffusion Module (CDM) to formulate a continual distribution transfer among three sub-spaces (object, image, and text) with cross-modal nature. Extensive experiments validate that the proposed method significantly outperforms existing approaches on large-scale benchmarks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.235, + 0.114, + 0.732, + 0.13 + ], + "angle": 0, + "content": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.194 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by Innovation and Technology Commission- Innovation and Technology Fund ITS/229/22 and Hong Kong Research Grants Council (RGC) General Research Fund 11211221." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.217, + 0.323, + 0.234 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.251, + 0.787, + 0.307 + ], + "angle": 0, + "content": "1. Ali, S., Ghatwary, N., Jha, D., Isik-Polat, E., Polat, G., Yang, C., Li, W., Galdran, A., Ballester, M.A.G., Thambawita, V., et al.: Assessing generalisability of deep learning-based polyp detection and segmentation methods through a computer vision challenge. Sci. Rep. (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.308, + 0.786, + 0.336 + ], + "angle": 0, + "content": "2. Amit, T., Nachmani, E., Shaharbany, T., Wolf, L.: Segdiff: Image segmentation with diffusion probabilistic models. ArXiv:2112.00390 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.337, + 0.787, + 0.377 + ], + "angle": 0, + "content": "3. Bansal, A., Borgnia, E., Chu, H.M., Li, J.S., Kazemi, H., Huang, F., Goldblum, M., Geiping, J., Goldstein, T.: Cold diffusion: Inverting arbitrary image transforms without noise. ArXiv:2208.09392 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.379, + 0.786, + 0.405 + ], + "angle": 0, + "content": "4. Baranchuk, D., Voynov, A., Rubachev, I., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.407, + 0.787, + 0.433 + ], + "angle": 0, + "content": "5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV. pp. 213-229 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.435, + 0.786, + 0.462 + ], + "angle": 0, + "content": "6. Chen, S., Sun, P., Song, Y., Luo, P.: Diffusiondet: Diffusion model for object detection. ArXiv:2211.09788 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.463, + 0.786, + 0.49 + ], + "angle": 0, + "content": "7. Chen, Z., Li, W., Xing, X., Yuan, Y.: Medical federated learning with joint graph purification for noisy label learning. MedIA (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.491, + 0.786, + 0.518 + ], + "angle": 0, + "content": "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.519, + 0.786, + 0.559 + ], + "angle": 0, + "content": "9. Du, Y., Wei, F., Zhang, Z., Shi, M., Gao, Y., Li, G.: Learning to prompt for open-vocabulary object detection with vision-language model. ArXiv:2203.14940 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.561, + 0.786, + 0.587 + ], + "angle": 0, + "content": "0. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. IJCV 88, 303-338 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.589, + 0.787, + 0.616 + ], + "angle": 0, + "content": "1. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Expand your detector vocabulary with uncurated images. ArXiv:2203.16513 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.617, + 0.787, + 0.657 + ], + "angle": 0, + "content": "2. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Towards open-vocabulary detection using uncurated images. In: ECCV. pp. 701-717. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.787, + 0.699 + ], + "angle": 0, + "content": "3. Gao, M., Xing, C., Niebles, J.C., Li, J., Xu, R., Liu, W., Xiong, C.: Towards open vocabulary object detection without human-provided bounding boxes. ArXiv:2111.09452 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.701, + 0.787, + 0.728 + ], + "angle": 0, + "content": "4. Gao, M., Xing, C., Niebles, J.C., Li, J., Xu, R., Liu, W., Xiong, C.: Open vocabulary object detection with pseudo bounding-box labels. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.787, + 0.756 + ], + "angle": 0, + "content": "5. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.787, + 0.784 + ], + "angle": 0, + "content": "6. Gu, Z., Chen, H., Xu, Z., Lan, J., Meng, C., Wang, W.: Diffusioninst: Diffusion model for instance segmentation. ArXiv:2212.02773 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.787, + 0.812 + ], + "angle": 0, + "content": "7. Gupta, A., Dollar, P., Girshick, R.: LVIS: A dataset for large vocabulary instance segmentation. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "8. Han, J., Ren, Y., Ding, J., Yan, K., Xia, G.S.: Few-shot object detection via variational feature aggregation. ArXiv:2301.13411 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.251, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.326, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.175 + ], + "angle": 0, + "content": "19. Han, X., Zheng, H., Zhou, M.: Card: Classification and regression diffusion models. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.764, + 0.189 + ], + "angle": 0, + "content": "20. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask R-CNN. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.786, + 0.217 + ], + "angle": 0, + "content": "21. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.786, + 0.244 + ], + "angle": 0, + "content": "22. Kim, D., Angelova, A., Kuo, W.: Contrastive feature masking open-vocabulary vision transformer. In: ICCV. pp. 15602-15612 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.786, + 0.273 + ], + "angle": 0, + "content": "23. Kim, D., Angelova, A., Kuo, W.: Region-aware pretraining for open-vocabulary object detection with vision transformers. In: CVPR. pp. 11144-11154 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.786, + 0.299 + ], + "angle": 0, + "content": "24. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. ArXiv:1312.6114 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.327 + ], + "angle": 0, + "content": "25. Korte, B.H., Vygen, J., Korte, B., Vygen, J.: Combinatorial optimization, vol. 1. Springer (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.786, + 0.356 + ], + "angle": 0, + "content": "26. Kuo, W., Cui, Y., Gu, X., Piergiovanni, A., Angelova, A.: Open-vocabulary object detection upon frozen vision and language models. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.383 + ], + "angle": 0, + "content": "27. Kwon, M., Jeong, J., Uh, Y.: Diffusion models already have a semantic latent space. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.424 + ], + "angle": 0, + "content": "28. Leng, Y., Chen, Z., Guo, J., Liu, H., Chen, J., Tan, X., Mandic, D., He, L., Li, X.Y., Qin, T., et al.: Binauralgrad: A two-stage conditional diffusion probabilistic model for binaural audio synthesis. ArXiv:2205.14807 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.786, + 0.452 + ], + "angle": 0, + "content": "29. Li, A.C., Prabhudesai, M., Duggal, S., Brown, E., Pathak, D.: Your diffusion model is secretly a zero-shot classifier. ArXiv:2303.16203 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.786, + 0.479 + ], + "angle": 0, + "content": "30. Li, C., Feng, B.Y., Fan, Z., Pan, P., Wang, Z.: Steganerf: Embedding invisible information within neural radiance fields. In: ICCV. pp. 441-453 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.787, + 0.522 + ], + "angle": 0, + "content": "31. Li, C., Lin, X., Mao, Y., Lin, W., Qi, Q., Ding, X., Huang, Y., Liang, D., Yu, Y.: Domain generalization on medical imaging classification using episodic training with task augmentation. Computers in biology and medicine 141, 105144 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.787, + 0.562 + ], + "angle": 0, + "content": "32. Li, C., Liu, H., Fan, Z., Li, W., Liu, Y., Pan, P., Yuan, Y.: Gaussianstego: A generalizable stenography pipeline for generative 3d gaussians splatting. ArXiv:2407.01301 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.563, + 0.787, + 0.604 + ], + "angle": 0, + "content": "33. Li, C., Liu, H., Liu, Y., Feng, B.Y., Li, W., Liu, X., Chen, Z., Shao, J., Yuan, Y.: Endora: Video generation models as endoscopy simulators. ArXiv:2403.11050 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.787, + 0.633 + ], + "angle": 0, + "content": "34. Li, C., Liu, X., Li, W., Wang, C., Liu, H., Yuan, Y.: U-kan makes strong backbone for medical image segmentation and generation. ArXiv:2406.02918 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.66 + ], + "angle": 0, + "content": "35. Li, C., Zhang, Y., Li, J., Huang, Y., Ding, X.: Unsupervised anomaly segmentation using image-semantic cycle translation. ArXiv:2103.09094 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.787, + 0.702 + ], + "angle": 0, + "content": "36. Li, C., Zhang, Y., Liang, Z., Ma, W., Huang, Y., Ding, X.: Consistent posterior distributions under vessel-mixing: a regularization for cross-domain retinal artery/vein classification. In: ICIP. pp. 61-65 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.787, + 0.729 + ], + "angle": 0, + "content": "37. Li, W., Chen, Z., Li, B., Zhang, D., Yuan, Y.: Htd: Heterogeneous task decoupling for two-stage object detection. TIP (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.787, + 0.757 + ], + "angle": 0, + "content": "38. Li, W., Guo, X., Yuan, Y.: Novel scenes & classes: Towards adaptive open-set object detection. In: ICCV. pp. 15780-15790 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.785 + ], + "angle": 0, + "content": "39. Li, W., Liu, J., Han, B., Yuan, Y.: Adjustment and alignment for unbiased open set domain adaptation. In: CVPR. pp. 24110-24119 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.813 + ], + "angle": 0, + "content": "40. Li, W., Liu, X., Yao, X., Yuan, Y.: Scan: Cross domain object detection with semantic conditioned adaptation. In: AAAI. pp. 1421-1428 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "41. Li, W., Liu, X., Yuan, Y.: Scan++: Enhanced semantic conditioned adaptation for domain adaptive object detection. TMM (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.235, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "42. Li, W., Liu, X., Yuan, Y.: Sigma: Semantic-complete graph matching for domain adaptive object detection. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.203 + ], + "angle": 0, + "content": "43. Li, W., Liu, X., Yuan, Y.: Sigma++: Improved semantic-complete graph matching for domain adaptive object detection. TPAMI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.785, + 0.245 + ], + "angle": 0, + "content": "44. Li, W., Yang, C., Liu, J., Liu, X., Guo, X., Yuan, Y.: Joint polyp detection and segmentation with heterogeneous endoscopic data. In: ISBI Workshop: EndoCV 2021. pp. 69-79. CEUR-WS Team (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.273 + ], + "angle": 0, + "content": "45. Liang, C., Wang, W., Miao, J., Yang, Y.: Gmmseg: Gaussian mixture based generative semantic segmentation models. ArXiv:2210.02025 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "46. Lin, C., Sun, P., Jiang, Y., Luo, P., Qu, L., Haffari, G., Yuan, Z., Cai, J.: Learning object-language alignments for open-vocabulary object detection. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.328 + ], + "angle": 0, + "content": "47. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.356 + ], + "angle": 0, + "content": "48. Lipman, Y., Chen, R.T.Q., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.383 + ], + "angle": 0, + "content": "49. Liu, H., Liu, Y., Li, C., Li, W., Yuan, Y.: Lgs: A light-weight 4d gaussian splatting for efficient surgical scene reconstruction. ArXiv:2406.16073 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.785, + 0.411 + ], + "angle": 0, + "content": "50. Liu, X., Li, W., Yamaguchi, T., Geng, Z., Tanaka, T., Tsai, D.P., Chen, M.K.: Stereo vision meta-lens-assisted driving vision. ACS Photonics (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.438 + ], + "angle": 0, + "content": "51. Liu, X., Gong, C., Liu, Q.: Flow straight and fast: Learning to generate and transfer data with rectified flow. ArXiv:2209.03003 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "52. Liu, X., Li, W., Yang, Q., Li, B., Yuan, Y.: Towards robust adaptive object detection under noisy annotations. In: CVPR. pp. 14207-14216 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "53. Liu, X., Li, W., Yuan, Y.: Decoupled unbiased teacher for source-free domain adaptive medical object detection. TNNLS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "54. Liu, Y., Li, C., Yang, C., Yuan, Y.: Endogaussian: Gaussian splatting for deformable surgical scene reconstruction. ArXiv:2401.12561 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.55 + ], + "angle": 0, + "content": "55. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. ArXiv:2211.01095 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "56. Luo, J., Li, Y., Pan, Y., Yao, T., Feng, J., Chao, H., Mei, T.: Semantic-conditional diffusion networks for image captioning. ArXiv:2212.03099 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.605 + ], + "angle": 0, + "content": "57. Ma, C., Jiang, Y., Wen, X., Yuan, Z., Qi, X.: Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. NeurIPS 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.633 + ], + "angle": 0, + "content": "58. Maaz, M., Rasheed, H., Khan, S., Khan, F.S., Anwer, R.M., Yang, M.H.: Class-agnostic object detection with multi-modal transformer. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.661 + ], + "angle": 0, + "content": "59. Miller, D., Sünderhauf, N., Milford, M., Dayoub, F.: Uncertainty for identifying open-set errors in visual object detection. Robot. Lett. 7(1), 215-222 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "60. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., et al.: Simple open-vocabulary object detection. In: ECCV. pp. 728-755 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.73 + ], + "angle": 0, + "content": "61. Ni, J., Qiu, Q., Chellappa, R.: Subspace interpolation via dictionary learning for unsupervised domain adaptation. In: CVPR (June 2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.757 + ], + "angle": 0, + "content": "62. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.71, + 0.771 + ], + "angle": 0, + "content": "63. Norris, J.R.: Markov chains. No. 2, Cambridge university press (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.813 + ], + "angle": 0, + "content": "64. Pan, P., Fan, Z., Feng, B.Y., Wang, P., Li, C., Wang, Z.: Learning to estimate 6dof pose from limited data: A few-shot, generalizable approach using rgb images. In: 2024 International Conference on 3D Vision (3DV). pp. 1059-1071. IEEE (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "65. Peebles, W., Xie, S.: Scalable diffusion models with transformers. ArXiv:2212.09748 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.326, + 0.127 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.788, + 0.189 + ], + "angle": 0, + "content": "66. Preechakul, K., Chathee, N., Wizadwongsa, S., Suwajanakorn, S.: Diffusion autoencoders: Toward a meaningful and decodable representation. In: CVPR. pp. 10619-10629 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.23 + ], + "angle": 0, + "content": "67. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.787, + 0.271 + ], + "angle": 0, + "content": "68. Rasheed, H.A., Maaz, M., Khattak, M.U., Khan, S., Khan, F.: Bridging the gap between object and image-level representations for open-vocabulary detection. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.787, + 0.297 + ], + "angle": 0, + "content": "69. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. In: NeurIPS (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.297, + 0.787, + 0.324 + ], + "angle": 0, + "content": "70. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.324, + 0.787, + 0.351 + ], + "angle": 0, + "content": "71. Shi, C., Yang, S.: Edadet: Open-vocabulary object detection using early dense alignment. In: ICCV. pp. 15724-15734 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.351, + 0.787, + 0.377 + ], + "angle": 0, + "content": "72. Wang, J., Zhang, H., Hong, H., Jin, X., He, Y., Xue, H., Zhao, Z.: Open-vocabulary object detection with an open corpus. In: ICCV. pp. 6759-6769 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.377, + 0.787, + 0.404 + ], + "angle": 0, + "content": "73. Wizadwongsa, S., Suwajanakorn, S.: Accelerating guided diffusion sampling with splitting numerical methods. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.404, + 0.787, + 0.431 + ], + "angle": 0, + "content": "74. Wu, L., Gong, C., Liu, X., Ye, M., Liu, Q.: Diffusion-based molecule generation with informative prior bridges. ArXiv:2209.00865 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.787, + 0.457 + ], + "angle": 0, + "content": "75. Wu, S., Zhang, W., Jin, S., Liu, W., Loy, C.C.: Aligning bag of regions for open-vocabulary object detection. ArXiv:2302.13996 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.457, + 0.787, + 0.484 + ], + "angle": 0, + "content": "76. Wu, X., Zhu, F., Zhao, R., Li, H.: Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. ArXiv:2303.13076 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.484, + 0.787, + 0.511 + ], + "angle": 0, + "content": "77. Xu, H., Zhang, Y., Sun, L., Li, C., Huang, Y., Ding, X.: Afsc: Adaptive fourier space compression for anomaly detection. ArXiv:2204.07963 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.511, + 0.787, + 0.551 + ], + "angle": 0, + "content": "78. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. ArXiv:2303.04803 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.551, + 0.787, + 0.578 + ], + "angle": 0, + "content": "79. Yang, Q., Li, W., Li, B., Yuan, Y.: Mrm: Masked relation modeling for medical image pre-training with genetics. In: ICCV. pp. 21452-21462 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.604 + ], + "angle": 0, + "content": "80. Ye, K., Zhang, M., Kovashka, A., Li, W., Qin, D., Berent, J.: Cap2det: Learning to amplify weak caption supervision for object detection. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.604, + 0.787, + 0.631 + ], + "angle": 0, + "content": "81. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary detr with conditional matching. In: ECCV. pp. 106-122 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.631, + 0.787, + 0.657 + ], + "angle": 0, + "content": "82. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.787, + 0.697 + ], + "angle": 0, + "content": "83. Zbinden, L., Doorenbos, L., Pissas, T., Huber, A.T., Sznitman, R., Márquez-Neila, P.: Stochastic segmentation with conditional categorical diffusion models. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.697, + 0.787, + 0.738 + ], + "angle": 0, + "content": "84. Zhang, Y., Li, C., Lin, X., Sun, L., Zhuang, Y., Huang, Y., Ding, X., Liu, X., Yu, Y.: Generator versus segmentor: Pseudo-healthy synthesis. In: MICCAI. pp. 150–160 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.738, + 0.787, + 0.779 + ], + "angle": 0, + "content": "85. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: CVPR. pp. 16793-16803 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.779, + 0.787, + 0.806 + ], + "angle": 0, + "content": "86. Zhou, X., Girdhar, R., Joulin, A., Krahenbuhl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: ECCV. pp. 350-368 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.806, + 0.787, + 0.833 + ], + "angle": 0, + "content": "87. Zhou, X., Koltun, V., Krahenbuhl, P.: Probabilistic two-stage detection. In: ArXiv:2103.07461 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.788, + 0.833 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CLIFF_ Continual Latent Diffusion for Open-Vocabulary Object Detection/51812c3e-19bd-4e17-9855-dbf65e0d5287_origin.pdf b/2024/CLIFF_ Continual Latent Diffusion for Open-Vocabulary Object Detection/51812c3e-19bd-4e17-9855-dbf65e0d5287_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9f49c2e486c82606b4deeeae7a2766b69b8fc8e5 --- /dev/null +++ b/2024/CLIFF_ Continual Latent Diffusion for Open-Vocabulary Object Detection/51812c3e-19bd-4e17-9855-dbf65e0d5287_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd2eddffa762c2e0896c78197771c53579f2cdf72498033c33f0937f0f09ef02 +size 4518756 diff --git a/2024/CLIFF_ Continual Latent Diffusion for Open-Vocabulary Object Detection/full.md b/2024/CLIFF_ Continual Latent Diffusion for Open-Vocabulary Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..aaff69b2a2f0a3665634aadc15669744ecc43e0c --- /dev/null +++ b/2024/CLIFF_ Continual Latent Diffusion for Open-Vocabulary Object Detection/full.md @@ -0,0 +1,372 @@ +# CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection + +Wuyang $\mathrm{Li}^{1}$ , Xinyu Liu, Jiayi Ma2, and Yixuan Yuan + +1 The Chinese University of Hong Kong +2 Wuhan University +wymanbest@outlook.com, yxyuan@ee.cuhk.edu.hk + +Abstract. Open-vocabulary object detection (OVD) utilizes image-level cues to expand the linguistic space of region proposals, thereby facilitating the detection of diverse novel classes. Recent works adapt CLIP embedding by minimizing the object-image and object-text discrepancy combinatorially in a discriminative paradigm. However, they ignore the underlying distribution and the disagreement between the image and text objective, leading to the misaligned distribution between the vision and language sub-space. To address the deficiency, we explore the advanced generative paradigm with distribution perception and propose a novel framework based on the diffusion model, coined Continual Latent Diffusion (CLIFF), which formulates a continual distribution transfer among the object, image, and text latent space probabilistically. CLIFF consists of a Variational Latent Sampler (VLS) enabling the probabilistic modeling and a Continual Diffusion Module (CDM) for the distribution transfer. Specifically, in VLS, we first establish a probabilistic object space with region proposals by estimating distribution parameters. Then, the object-centric noise is sampled from the estimated distribution to generate text embedding for OVD. To achieve this generation process, CDM conducts a short-distance object-to-image diffusion from the sampled noise to generate image embedding as the medium, which guides the long-distance diffusion to generate text embedding. Extensive experiments verify that CLIFF can significantly surpass state-of-the-art methods on benchmarks. The code is available at https://github.com/CUHK-AIM-Group/CLIFF. + +Keywords: Open-vocabulary object detection $\cdot$ Diffusion model $\cdot$ Continual distribution transfer + +# 1 Introduction + +Object detection [20, 37, 69] has made significant strides in recent years, as demonstrated by the impressive results on benchmarks [17, 47]. Nonetheless, due to the unsatisfactory closed-set constraint [77, 82], object detectors are limited to discovering the novel-class objects [38, 39] unseen during training, impeding a reliable scene understanding in the real world. Further, it hinders the journey toward achieving true artificial intelligence since it requires the human-like perceptual ability to recognize unexpected objects appearing everywhere. + +![](images/3e1a0e4c75f05381ad71649270fa6c534b2bcd960881737f6641d1433feecc9a.jpg) +(a) Existing OVD works +(b) The proposed CLIFF framework modeling a continual distribution transfer +Fig. 1: Illustration of (a) existing discriminative paradigm [15,68,75] with a combinatorial objective (object→image and object→text in parallel), and (b) the proposed method conducting a continual diffusion (object→image→text) with generative nature. + +To break through this barrier, open-vocabulary object detection (OVD) has been studied that leverages auxiliary linguistic cues to generalize the detection of novel classes [82]. Recent advances adapt CLIP [67], a vision-language (VL) foundation model with strong zero-shot capacity, to boost the novel-class generalization of the object detector. Some works [13, 14] utilize the CLIP prediction (image-level) to generate pseudo labels at the object level, enabling the self-training of the object detector. To fully explore the hidden knowledge, some other works strive to align the object-level representation with CLIP embeddings in the hidden space. During training the object detector, they minimize the object-image [15, 68, 75] and object-text [46, 68, 76] discrepancy to align the sub-space among the object, image, and text modalities, thereby unleashing the novel-class generalization capacity from the pre-trained CLIP model. + +Despite the significant success, the existing OVD paradigm, as depicted in Fig. 1(a), has two limitations regarding the sub-space alignment. First, existing works utilize text embedding as the classifier weight to explicitly optimize the discriminative class separation, which is agnostic to the underlying distribution and may lead to class bias [18]. This distribution-agnostic property results in the distribution misalignment [40, 42, 46] and severely hinders the novel-class generalization [78]. Besides, due to the uncalibrated latent space in discriminative models [45], e.g., numerical high-variance, the object distribution is prone to overfit to the labeled classes, worsening the misalignment with the unbiased CLIP distribution. With the rise of generative paradigm [2], this barrier has been gradually broken via diverse probabilistic modeling [18, 45, 59]. Hence, we aim to incorporate probabilistic advantages into discriminative detection pipelines, which can facilitate the distribution transfer across the object and text sub-spaces, thereby achieving an unbiased novel-class perception [45, 59]. + +Second, some existing works minimize the object-image and object-text discrepancy simultaneously in a combinatorial manner, ignoring the disagreement between the image and text two objectives3. With this combinatorial nature, optimizing object-image alignment as the same role as the object-text counterpart is suboptimal due to the potential conflict between the two objectives [25]. + +Intuitively, the image and object share visual representation, while the image and text are coupled with CLIP pre-training, which motivates us to introduce the image as an interpolated sub-space [61] between the object and text counterpart, yielding a continual distribution transfer among three modalities. + +To address the above issues, we propose Continual Latent Diffusion (CLIFF) for OVD, as shown in Fig. 1(b). CLIFF breaks through the discriminative bottleneck via a probabilistic Variational Latent Sampler (VLS) and then formulates a diffusion-based distribution transfer [21] in a Continual Diffusion Module (CDM) to avoid the combinatorial limitation. Specifically, VLS leverages variational approximation [24] to establish a probabilistic object space with learned distribution parameters, and then samples object-centric noise to boost the diffusion process. CDM formulates the continual diffusion process to transfer the object distribution to the image and text counterpart, avoiding disagreement in the existing combinatorial objective. Besides, different from existing diffusion-based works [27, 78] with unsatisfactory efficiency, CLIFF achieves comparable and even better model efficiency compared with state-of-the-art OVD works, which is verified in Appendix. To be summarized, the contributions are four-fold: + +- We propose Continual Latent Diffusion (CLIFF) for OVD, which achieves a continual distribution transfer with DDPM-based generation. To the best of our knowledge, this work represents the first attempt to study latent diffusion in VL discriminative tasks and push forward benchmark results. +- Variational Latent Sampler is proposed to generate a probabilistic object space with variational modeling, in which the object-centric noise can be sampled with reparameterization to boost the generation. +- We design an efficient Continual Diffusion Module to facilitate the distribution transfer among three cross-modality sub-spaces, namely the object, image, and text space, using a new diffusion module and unified DDPM formulation. +- CLIFF is extensively validated through different benchmarks, which verify its state-of-the-art role and show great adaptability to various detectors. + +# 2 Related Works + +# 2.1 Open-Vocabulary Object Detection + +Open-vocabulary object detection (OVD) leverages auxiliary image-level labels with linguistic cues, e.g., classification [68] and caption datasets [80], to enlarge the semantic space of the region proposals [5, 69], thereby facilitating the novel-class generalization for the real-world applications, such as automatic driving [41, 43, 50, 52], medical imaging analysis [31, 35, 36, 49] and diagnostic decision-making [1, 7, 44, 53, 79]. Zareian et al. [82] formulate the standard OVD problem setting and propose OVR-CNN to project the visual space to the linguistic counterpart on Faster-RCNN [69]. Some other works [22, 23, 60] delve into OVD with Vision Transformer (ViT) by designing large-scale region-aware pretraining [23], feature masking [22], which fully unleash the potential of ViT in fine-grained detection tasks. Recent works adapt CLIP [67] to objects detectors for the novel-class + +generalization. The other works adapt the CLIP-encoded latent embedding and aim to align three critical sub-spaces: the object space, the CLIP image space, and the CLIP text space. They bridge the gap between the object and text sub-space via region-text contrastive pre-training [85], region-word bipartite matching [46], CLIP-based conditional matching [81], and learnable detection-friendly prompts [9, 11, 12]. PromptDet [12] and DetPro [9] use learnable text prompts to learn with noisy uncurated web images, thereby enhancing the object-text alignment. Additionally, recent advances [15, 68, 75] adapt the image sub-space with the embedding distillation [15], weight transfer [68] and bags of regions [75], to fully use CLIP embedding, serving as a combinatorial objective with the object-text alignment. Existing works follow a similar discriminative paradigm for OVD, which cannot model the underlying distribution with a suboptimal distribution alignment. Differently, we formulate a continual transfer in a unified denoising diffusion module, enabling the distribution-level perception and adaptation among the object, image, and text sub-spaces. + +# 2.2 Denoising Diffusion Probabilistic Model + +As the advanced generative paradigm, Denoising Diffusion Probabilistic Model (DDPM) [21] ("diffusion model" for short) generates samples by formulating a gradual denoising process. This process can be treated as a distribution transfer [51] from the normal distribution to the data in a flow-like manner [51, 54, 84]. Instead of diffusing in the data space, LDM [70] and VDM [66] deploy the denoising diffusion in the latent space to enable the high-resolution synthesis. Recent works can be generally categorized into faster sampling [55, 73], noise corruption designs [3, 62], and diffusion model architecture [33, 34, 65], etc. The diffusion model has achieved remarkable generative performance in the image, text, video, audio, graph, and motion generation [21, 28, 30, 32, 74] with different diffusion conditions for downstream applications. + +In addition to generative tasks, recent works [6,16,27,56,78,83] have discovered that the diffusion feature has adequate semantic discriminability with richer distribution-level perception and data-efficient property [4,64] than conventional discriminative models [67]. Without fine-tuning, an off-the-shelled diffusion model can achieve satisfactory zero-shot classification [29]. Along this trend, some bottlenecks in discriminative models are gradually broken by involving the generative paradigm [78]. CARD [19] uses the diffusion model to estimate the uncertainty to improve image classification. Unlike existing works considering the noise $\rightarrow$ data diffusion, we formulate a continual distribution transfer among three modalities, yielding the object $\rightarrow$ image $\rightarrow$ text transfer in the latent space. + +# 3 Preliminaries + +Problem Formulation. Assuming a set of base classes $\mathcal{C}_B$ and novel classes $\mathcal{C}_N$ , we have a detection dataset $\mathcal{D}_{det} = \{I_i,O_i\}_{i = 1}^n$ with object-level labels $\{O_i\}_{i = 1}^n$ in $\mathcal{C}_B$ and another auxiliary dataset $\mathcal{D}_{aux} = \{I_i,Y_i\}_{i = 1}^{n'}$ with image-level + +Algorithm 1 Reverse Diffusion Process (Denoising): RevDif $(\mathbf{x}_T,\mathbf{c},L)$ + +Require: $\mathbf{x}_T$ : initial noise; c: diffusion condition; $L$ : the number reverse steps. + +Ensure: $\mathbf{x}_{T - L}$ generated samples with $L$ steps. + +1: for $t = T, \dots, T - L + 1$ do +2: $\mathbf{z} \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ +3: $\mathbf{x}_{t - 1} = \frac{1}{\sqrt{\alpha_t}}\left(\mathbf{x}_t - \frac{1 - \alpha_t}{\sqrt{1 - \bar{\alpha}_t}}\pmb {\epsilon}_\theta (\mathbf{x}_t,\mathbf{c},t)\right) + \sigma_t\mathbf{z}$ +4: end for +5: return $\mathbf{x}_{T - L}$ + +labels, e.g., the image caption [21] and classification dataset [68], which has the weakly supervised labels in $\mathcal{C}_B \cup \mathcal{C}_N$ . In $\mathcal{D}_{det}$ , each image $I_i$ is associated with a set of objects $O_i = \{(x,y,w,h), \hat{y}\}$ , including the bounding-box coordinates $\{(x,y,w,h)\}$ and class labels $\{\hat{y}\} \subset \mathcal{C}_B$ . OVD aims to use $\mathcal{D}_{det}$ and $\mathcal{D}_{aux}$ for training and detect base and novel objects in $\mathcal{C}_B \cup \mathcal{C}_N$ during the test stage. + +Diffusion Model. Formulating a generative paradigm with the Markov chain [21, 63], DDPM is able to generate data $\mathbf{x}_0$ from the noise $\mathbf{x}_T \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ via a $T$ -step denoising process. Concretely, DDPM begins with the clean data $\mathbf{x}_0$ and gradually adds Gaussian noise using the fixed forward process $q(\mathbf{x}_t | \mathbf{x}_{t-1})$ , where $t = 1, 2, \dots, T$ . This process can be denoted as follows, + +$$ +q \left(\mathbf {x} _ {t} \mid \mathbf {x} _ {0}\right) = \mathcal {N} \left(\mathbf {x} _ {t}; \sqrt {\bar {\alpha} _ {t}} \mathbf {x} _ {0}, (1 - \bar {\alpha} _ {t}) \mathbf {I}\right), \tag {1} +$$ + +where $\bar{\alpha}_t$ is noise schedule [62]. To recover data $\mathbf{x}_0$ from noise $\mathbf{x}_T$ , the reverse process parameterized with $\theta$ is learned to denoise inputs with condition $\mathbf{c}$ : + +$$ +p _ {\theta} \left(\mathbf {x} _ {0: T} \mid \mathbf {c}\right) = p \left(\mathbf {x} _ {T}\right) \prod_ {t = 1} ^ {T} p _ {\theta} \left(\mathbf {x} _ {t - 1} \mid \mathbf {x} _ {t}, \mathbf {c}\right), \tag {2} +$$ + +where $p(\mathbf{x}_T)$ is fixed to Gaussian. In this reverse step, the diffusion model typically learns a neural network $\epsilon_{\theta}(\cdot)$ to generate the values for computing the mean of $p_{\theta}(\mathbf{x}_{t - 1}\mid \mathbf{x}_t,\mathbf{c})$ given by the noisy input $\mathbf{x}_t$ , which is optimized to minimize the variational lower bound (ELBO). As proved in [21, 29], this training objective can be re-written as: + +$$ +- \mathbb {E} _ {t, \epsilon \sim \mathcal {N} (\mathbf {0}, \mathbf {I})} \left[ \| \epsilon - \epsilon_ {\theta} (\mathbf {x} _ {t}, \mathbf {c}, t) \| ^ {2} \right], \tag {3} +$$ + +where the diffusion model $\epsilon_{\theta}(\cdot)$ can estimate the added noise of each noisy input $\mathbf{x}_t$ . After optimizing with Eq. 3, the diffusion model can generate a data sample $\mathbf{x}_{T - L}$ at arbitrary time-step $T - L$ from a random noise $\mathbf{x}_T \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ with $L$ steps of the reverse diffusion, optionally guided by the condition $\mathbf{c}$ . This process [21] is represented in Algorithm 1. In the following sections, we use RevDif( $\cdot$ ) to represent the reverse diffusion (noise→data) for simplicity. + +# 4 Methodology + +The overall workflow is illustrated in Fig. 2. We feed input images $\{I_i\}_{i=1}^B$ to the backbone and RPN [69] to extract $N$ region proposals, which are sent to the + +![](images/fc393f73c102897fa4a24c4fe58915cd200cc74ce2ebe8e250eff65f1273dbba.jpg) +Fig. 2: Illustration of the proposed CLIFF. VLS establishes a probabilistic embedding space with region proposals, where we sampled object-centric noise $\mathbf{x}_T$ . Then, CDM achieves a continual distribution transfer from the object to the image and text. + +feature extractor to generate object embedding $\mathbf{x}_{\mathrm{obj}}$ . Then, VLS, parameterized with $\phi$ , establishes and estimates the mean and variance of a probabilistic object space $p_{\phi}(\mathbf{x}_T|\mathbf{x}_{\mathrm{obj}})$ , where the reparameterization is conducted to sample object-centric noise $\mathbf{x}_T$ , as shown in Fig. 2(a). After that, the sampled noise $\mathbf{x}_T$ is sent to CDM, denoted as $\epsilon_{\theta}(\cdot)$ , to generate text embedding $\mathbf{x}_{\mathrm{txt}}$ with diffusion, as depicted in Fig. 2(b). Conditioned on $\mathbf{c}_{o2i}$ , we first conduct object-to-image diffusion to generate the image embedding $\mathbf{x}_{\mathrm{img}}$ with $L$ steps (top sub-figure), which is used to generate refined diffusion condition $\mathbf{c}_{o2t}$ . Conditioned on $\mathbf{c}_{o2t}$ , we further perform a $T$ -step object-to-text diffusion to generate text embedding $\mathbf{x}_{\mathrm{txt}}$ (bottom sub-figure), whose similarity with the CLIP text embedding $\hat{\mathbf{x}}_{\mathrm{txt}}$ are measured to serve as the final class prediction. + +Formally, by extending Eq. 2, CLIFF consists of two diffusion processes among the object, image, and text embedding sub-space. Guided by the condition $\mathbf{c}_{o2i}$ , the object-to-image diffusion, written as $p_{\theta,\phi}(\mathbf{x}_{\mathrm{img}}|\mathbf{c}_{o2i},\mathbf{x}_{\mathrm{obj}})$ , aims to generate the associated CLIP image embedding from the object $\mathbf{x}_{\mathrm{obj}}$ with $L$ steps: + +$$ +p _ {\theta , \phi} (\mathbf {x} _ {\mathrm {i m g}} | \mathbf {c} _ {o 2 i}, \mathbf {x} _ {\mathrm {o b j}}) = p _ {\phi} (\mathbf {x} _ {T} | \mathbf {x} _ {\mathrm {o b j}}) \prod_ {t = T - L + 1} ^ {T} p _ {\theta} (\mathbf {x} _ {t - 1} \mid \mathbf {x} _ {t}, \mathbf {c} _ {o 2 i}). \tag {4} +$$ + +Similarly, with condition $\mathbf{c}_{o2t}$ , the object-to-text diffusion is designed to generate text embedding with $T$ steps for the final prediction, denoted as follows, + +$$ +p _ {\theta , \phi} (\mathbf {x} _ {\mathrm {t x t}} | \mathbf {c} _ {o 2 t}, \mathbf {x} _ {\mathrm {o b j}}) = p _ {\phi} (\mathbf {x} _ {T} | \mathbf {x} _ {\mathrm {o b j}}) \prod_ {t = 1} ^ {T} p _ {\theta} (\mathbf {x} _ {t - 1} \mid \mathbf {x} _ {t}, \mathbf {c} _ {o 2 t}). \tag {5} +$$ + +The proposed CLIFF offers two significant improvements compared to the standard diffusion process (Eq. 2). First, instead of beginning from a random noise space $p(\mathbf{x}_T)$ , we start from a differential probabilistic space $p_{\phi}(\mathbf{x}_T|\mathbf{x}_{\mathrm{obj}})$ with objects to generate the text embedding. Second, going beyond a single diffusion between two sub-spaces, we formulate the effective diffusion process among the object, image, and text in different modalities. + +# 4.1 Variational Latent Sampler + +With input images $\{I_i\}_{i=1}^B$ , the backbone, RPN4, and the feature extractor are adopted to obtain $N$ object embeddings $\mathbf{x}_{\mathrm{obj}} \in \mathbb{R}^{N \times D}$ . Then, we send $\mathbf{x}_{\mathrm{obj}}$ to VLS to establish a probabilistic embedding space $p_{\phi}(\mathbf{x}_T | \mathbf{x}_{\mathrm{obj}})$ satisfying the DDPM w.r.t the normally distributed assumption for the diffusion-based transfer. This probabilistic space $p_{\phi}(\mathbf{x}_T | \mathbf{x}_{\mathrm{obj}})$ can break through the discriminative limitation of [14,15,46,75] without class bias [18]. Besides, it introduces object-level semantics to enable object-to-text transfer with a differential nature, different from the non-optimizable random space $p(\mathbf{x}_T)$ of noise-to-data transfer [21]. + +Formally, inspired by VAE [24], we establish the probabilistic space with centered isotropic multivariate Gaussian $p_{\phi}(\mathbf{x}_T|\mathbf{x}_{\mathrm{obj}}) = \mathcal{N}(\mathbf{x}_T;\mathbf{0},\mathbf{I})$ , whose distribution parameters are computed from the observed object embedding $\mathbf{x}_{\mathrm{obj}}$ with non-linear layers [18]. To this end, two fully connected layers are introduced on the shared region feature extractor of the basic object detector, denoted as $f_{\mathrm{obj}}$ e.g., the ResNet blocks in Faster RCNN [69] and the MLP in Mask RCNN [20], to learn to estimate the distribution parameters. Specifically, given the object embedding $\mathbf{x}_{\mathrm{obj}}$ , the variational posterior $p_{\phi}(\mathbf{x}_T|\mathbf{x}_{\mathrm{obj}})$ is written as $\log p_{\phi}(\mathbf{x}_T|\mathbf{x}_{\mathrm{obj}}^{(i)}) = \log \mathcal{N}(\mathbf{x}_T;\pmb{\mu}^{(i)},\pmb{\sigma}^{2(i)}\mathbf{I})$ . Without losing the differential property [24], the mean and s.d. of the probabilistic object space can be estimated, + +$$ +\boldsymbol {\mu} _ {(i)} ^ {(i)} = \mathbf {W} _ {\mu} f _ {\mathrm {o b j}} \left(\mathbf {x} _ {\mathrm {o b j}} ^ {(i)}\right) + \mathbf {b} _ {\mu}, \tag {6} +$$ + +$$ +\pmb {\sigma} ^ {(i)} = \mathbf {W} _ {\sigma} f _ {\mathrm {o b j}} (\mathbf {x} _ {\mathrm {o b j}} ^ {(i)}) + \mathbf {b} _ {\sigma}, +$$ + +where $\mathbf{W}_{(\cdot)}$ and $\mathbf{b}_{(\cdot)}$ are learned weight and bias. Then, with the probabilistic modeling, we sample object-centric noise from the latent object space $\mathbf{x}_T \sim p_{\phi}(\mathbf{x}_T | \mathbf{x}_{\mathrm{obj}})$ with reparameterization [24]: $\mathbf{x}_T^{(i)} = \boldsymbol{\mu}^{(i)} + \boldsymbol{\sigma}^{(i)} \odot \boldsymbol{\epsilon}$ , where $\odot$ is the element-wise product and $\boldsymbol{\epsilon} \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ . + +To optimize the probabilistic space, we deploy Kullback-Leibler (KL) divergence $D_{\mathrm{KL}} \left(p_{\phi}(\mathbf{x}_T|\mathbf{x}_{\mathrm{obj}})||p(\mathbf{x}_T)\right)$ with $\mathbf{x}_T \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ that encourages the Gaussian distributed [24], which has two advantages. First, the probabilistic nature of normal Gaussian removes the class bias within the object space [18] and promotes the generalization capacity towards discovering novel classes [45, 59]. Second, it naturally satisfies the DDPM constraint (Eq. 1) with the Gaussian formulation [21] for the following diffusion process. The implemented KL optimization objective, as proved in the Appendix, can be re-written in the following format, + +$$ +\mathcal {L} _ {K L} = - \frac {1}{2 N} \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {D} \left(1 + \log \left(\boldsymbol {\sigma} _ {j} ^ {2 (i)}\right) - \boldsymbol {\mu} _ {j} ^ {2 (i)} - \boldsymbol {\sigma} _ {j} ^ {2 (i)}\right). \tag {7} +$$ + +VLS can establish a probabilistic space with differential property and sample object-centric noise $\mathbf{x}_T$ for the diffusion process, serving to generate probabilistic object-text pairs in the diffusion-based distribution transfer [48,51]. + +# 4.2 Continual Diffusion Module + +With object-centric noise $\mathbf{x}_T$ , we further deploy the object-to-image (o2i) and object-to-text (o2t) diffusion to achieve a continual distribution transfer for OVD. To achieve this, an efficient diffusion module $\epsilon_{\theta}(\cdot)$ with MLP architecture is designed, the details of which are presented in the Appendix. + +Object-to-image Diffusion. Given class-agnostic region proposals, we send them to the object detector to generate the object embedding $\mathbf{x}_{\mathrm{obj}}$ and feed them into frozen CLIP image encoder [67] to obtain static image embedding $\hat{\mathbf{x}}_{\mathrm{img}}$ . Then, with a randomly sampled object-centric noise $\mathbf{x}_T \sim p_\phi(\mathbf{x}_T | \mathbf{x}_{\mathrm{obj}})$ from VLS, we encourage the diffusion module to generate the image embedding $\mathbf{x}_{\mathrm{img}}$ with object-level condition $\mathbf{c}_{o2i} = \mathbf{x}_{\mathrm{obj}}$ using a $L$ -step reverse diffusion: + +$$ +\mathbf {x} _ {\mathrm {i m g}} = \operatorname {R e v D i f} \left(\mathbf {x} _ {T}, \mathbf {c} _ {o 2 i}, L\right), \tag {8} +$$ + +where $\mathrm{RevDif}(\cdot)$ (Algorithm 1) recovers image embedding from object-centric noise $\mathbf{x}_T$ . To this end, the object-to-image diffusion, denoted as $p_{\theta,\phi}(\mathbf{x}_{\mathrm{img}}|\mathbf{c}_{o2i},\mathbf{x}_{\mathrm{obj}})$ , uses object embedding as the diffusion condition $\mathbf{c}_{o2i} = \mathbf{x}_{\mathrm{obj}}$ , and is optimized by minimizing the distance between the generated embedding $\mathbf{x}_{\mathrm{img}}$ and CLIP image embedding $\hat{\mathbf{x}}_{\mathrm{img}}$ , which is achieved via the following loss function, + +$$ +\mathcal {L} _ {o 2 i} = \frac {1}{N _ {c a}} \sum_ {i = 1} ^ {N _ {c a}} | | \mathbf {x} _ {\mathrm {i m g}} ^ {(i)} - \hat {\mathbf {x}} _ {\mathrm {i m g}} ^ {(i)} | | _ {1}, \tag {9} +$$ + +where $N_{ca}$ is the number of class-agnostic proposals. With the optimized diffusion process, the object distribution could be transferred to the CLIP image counterpart, facilitating the adaptation to the CLIP latent space. + +Object-to-text Diffusion. With diffusion-based formulation, we further encourage the model to reconstruct the associated text embedding $\mathbf{x}_{\mathrm{txt}}$ for region proposals $\mathbf{x}_{\mathrm{obj}}$ , which conducts a generative procedure in latent space. Note that this generative objective is different from existing OVD works that directly optimize the discriminative class separation. Specifically, we design the object-to-text diffusion with a $T$ -step diffusion $p_{\theta,\phi}(\mathbf{x}_{\mathrm{txt}}|\mathbf{c}_{o2t},\mathbf{x}_{\mathrm{obj}})$ guided by condition $\mathbf{c}_{o2t}$ . Considering the dominant role of condition [21] in the generated content, we propose a simple yet effective attentive mechanism to obtain the refined condition $\mathbf{c}_{o2t} = \bar{\mathbf{x}}_{\mathrm{obj}}^{\mathrm{img}}$ with the generated image image-level context: + +$$ +\mathbf {x} _ {\mathrm {o b j}} ^ {\mathrm {i m g}} = \operatorname {S i g m o i d} \left(\operatorname {F C} \left(\mathbf {x} _ {\mathrm {i m g}}\right)\right) \odot \mathbf {x} _ {\mathrm {o b j}}, \tag {10} +$$ + +where $\mathbf{x}_{\mathrm{img}}$ is generated via Eq. 8, $\odot$ is the element-wise product, and FC is a liner projection. Then, we use the refined condition $\mathbf{c}_{o2t} = \mathbf{x}_{\mathrm{obj}}^{\mathrm{img}}$ to guide the text embedding generation for $\mathbf{x}_{\mathrm{txt}}$ with a $T$ -step reverse diffusion: + +$$ +\mathbf {x} _ {\mathrm {t x t}} = \operatorname {R e v D i f} \left(\mathbf {x} _ {T}, \mathbf {c} _ {o 2 t}, T\right). \tag {11} +$$ + +Considering the mutual benefits of generative and discriminative objectives for model learning as explored in [45], the loss function for the object-to-text + +diffusion process is designed with these two distinct components. To this end, the optimization objective can be expressed as follows, + +$$ +\mathcal {L} _ {o 2 t} = \underbrace {\frac {1}{N _ {l}} \sum_ {i} ^ {N _ {l}} | | \boldsymbol {\epsilon} _ {\theta} \left(\mathbf {x} _ {t} ^ {(i)} , \mathbf {c} _ {o 2 t} ^ {(i)} , t\right) - \boldsymbol {\epsilon} | | _ {1}} _ {\text {G e n e r a t i v e}} + \underbrace {\mathcal {L} _ {C E} \left(\operatorname {S i g m o i d} \left(\frac {1}{\tau} \cos \left(\mathbf {x} _ {\mathrm {t x t}}, \hat {\mathbf {x}} _ {\mathrm {t x t}}\right)\right), \hat {y}\right)} _ {\text {D i s c r i m i n a t i v e}} \tag {12} +$$ + +where $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ , $\mathbf{x}_t$ is the noisy CLIP text embedding at a randomly sampled time-step $t \sim \mathcal{U}[1,T]$ , $\tau = 0.02$ is a scaling factor as [15, 68], $\hat{y}$ is the class label, $\mathcal{L}_{CE}$ is the Cross-Entropy loss, and $N_{l} = N_{cs} + N_{rpn}$ counts the labeled proposals. The first term optimizes the diffusion module $\epsilon_{\theta}(\cdot)$ to reconstruct the text embedding in a generative manner [21], serving for the objective of Eq. 3. The second term can give an interpretable score for NMS [69] in detectors and ensure the generated embedding with sufficient discriminability. + +Remark. We use a shared diffusion module with a $T$ -step diffusion formulation to optimize two processes, where the first $L$ steps are assigned for object-to-image, and the whole $T$ steps are for object-to-text diffusion. Note that even though we do not explicitly optimize image-to-text diffusion, this transition can be intrinsically achieved between the $L$ -th and $T$ -th steps, as illustrated by the gray-colored part in Fig. 1(b). The reason is based on the continual property of the Markov chain in DDPM, i.e., formulating a random process that undergoes transitions from one state to another within a finite number of possible states [21, 63]. + +# 4.3 Training and Inference + +Training Stage. We implement the overall optimization objective on the basic object detector, Faster RCNN [69], to train the whole framework: + +$$ +\mathcal {L} _ {C L I F F} = \mathcal {L} _ {r p n} + \mathcal {L} _ {r e g} + \alpha \mathcal {L} _ {K L} + \beta_ {1} \mathcal {L} _ {o 2 i} + \beta_ {2} \mathcal {L} _ {o 2 t}, \tag {13} +$$ + +where $\mathcal{L}_{rpn}$ and $\mathcal{L}_{reg}$ are RPN loss and the bounding box regression loss in the RCNN head [69] borrowed from object detectors. $\alpha, \beta_1, \beta_2$ are the loss weight. + +Inference Stage. We first extract the region proposals $\mathbf{x}_{\mathrm{obj}}$ and then conduct reverse diffusion to generate the associated text embedding $\mathbf{x}_{\mathrm{txt}}$ for region proposals, whose similarity with $\hat{\mathbf{x}}_{\mathrm{txt}}$ is measured for class prediction $\tilde{y}$ : + +$$ +\tilde {y} = \operatorname {S i g m o i d} \left(\frac {1}{\tau} c o s \left(\mathbf {x} _ {\mathrm {t x t}}, \hat {\mathbf {x}} _ {\mathrm {t x t}}\right)\right), \tag {14} +$$ + +where $\tilde{y}$ is the per-class confidence, $\mathbf{x}_{\mathrm{txt}}$ is the generated text embedding via object-to-text diffusion (Eq. 11), and $\hat{\mathbf{x}}_{\mathrm{txt}}$ is CLIP text embedding for all classes. + +Table 1: Experimental results (%) on COCO benchmark. † indicates the reproduced results with the consistent baseline setting as [68], including ImageNet pre-training and CLIP prompting. ‡ indicates using the longer training schedule [68]. + +
MethodObject DetectorSub-space Alignment\( \mathrm{mAP}_{\mathrm{n}} \)\( \mathrm{mAP}_{\mathrm{b}} \)mAPReference
OVR-CNN [82]Faster RCNN C4Obj → Txt22.846.039.9CVPR-2021
RegionCLIP [85]Faster RCNN C4Obj → Txt & Img26.854.847.5CVPR-2022
ViLD [15]Faster RCNN C4Obj → Txt & Img27.659.551.3ICLR-2022
OV-DETR [81]Deformable DETRObj → Txt & Img29.461.052.7ECCV-2022
Detic [86]Faster RCNN C4Obj → Txt27.847.145.0ECCV-2022
Detic‡ [86]Faster RCNN C4Obj → Txt28.453.847.2ECCV-2022
PB-OVD [14]Faster RCNN C4Obj → Txt30.846.142.1ECCV-2022
MAVL-OVD [58]Faster RCNN C4Obj → Txt30.452.646.8ECCV-2022
\( OCD^‡ \)[68]Faster RCNN C4Obj → Txt & Img36.654.049.4NeurIPS-2022
F-VLM [26]Faster RCNN FPNObj → Txt & Img28.0-39.6ICLR-2023
VLDet [46]Faster RCNN C4Obj → Txt32.050.645.8ICLR-2023
BARON†,‡ [75]Faster RCNN C4Obj → Txt & Img41.053.350.1CVPR-2023
EdaDet [71]Faster RCNN C4Obj → Txt37.857.752.5ICCV-2023
GOAT [72]Faster RCNN C4Obj → Txt & Img36.453.048.6ICCV-2023
CLIFF (ours)Faster RCNN C4Obj → Image → Txt41.354.150.6-
CLIFF‡ (ours)Faster RCNN C4Obj → Image → Txt43.254.051.2-
+ +# 5 Experiments + +# 5.1 Datasets and Evaluation Metrics + +We follow existing works [68, 75] to compare with state-of-the-art methods. COCO Setting. The COCO-2017 dataset [47] is used for training and evaluation, containing 80 object-level classes. As for the auxiliary dataset $\mathcal{D}_{aux}$ , we follow [82] to use COCO caption dataset in $\mathcal{C}_B \cup \mathcal{C}_N$ . The core evaluation metric is the novel-class Average Precision (AP) with an IoU threshold of 0.5, denoted as $\mathrm{mAP_n}$ , for the novel-class generalization. The results are shown in Table 1. + +LVIS Setting. LVIS v1.0 [17] contains 1203 classes and is split into frequent, common, and rare types. For the auxiliary datasets $\mathcal{D}_{aux}$ , we follow [68] to use the subset of ImageNet-21K. The core evaluation metric is mask AP of rare classes, denoted as $\mathrm{AP_r^m}$ . We report the comparison results in Table 2. + +Pascal VOC Setting. We follow [72] to evaluate the generalization to unseen datasets with Pascal VOC [10] and LVIS [17] using COCO-trained model. We use $\mathrm{AP}_{50}$ to evaluate the performance as [72]. The results are given in Table 7. + +# 5.2 Implementation Details + +Object Detector. Considering that previous works use inconsistent detectors, we strictly follow the setting of [68,82] in different benchmarks. Concretely, for the COCO setting, we use Faster-RCNN C4 [69] with the ResNet-50 backbone. The backbone is pre-trained on ImageNet [8] without bells and whistles. For the LVIS setting, we follow the two main streams of existing works by using Mask-RCNN-FPN [20] and CenterNet v2 [87] as the basic object detectors for a thorough comparison, which is based on ReseNet-50 backbone. For the experiments of all benchmarks, we use the SGD optimizer with a weight decay of $1e^{-4}$ and a + +Table 2: Experimental results (\%) on LVIS benchmark with the consistent object detector setting. Ens indicates using the ensemble trick [11]. + +
MethodObject DetectorAPmRAPmcAPmfAPReference
ViLD-Ens [15]Mask RCNN16.624.630.325.5ICLR-2022
MVAL-OVD [58]Mask RCNN17.021.226.122.4ECCV-2022
OCD [68]Mask RCNN21.125.029.125.9NeurIPS-2022
DetPro [9]Mask RCNN19.825.628.925.9CVPR-2022
RegionCLIP [85]Mask RCNN17.127.434.028.2CVPR-2022
F-VLM [26]Mask RCNN18.624.026.924.2ICLR-2023
BARON [75]Mask RCNN18.024.428.925.1CVPR-2023
BARON-Ens [75]Mask RCNN19.226.829.426.5CVPR-2023
CLIFF (ours)Mask RCNN22.426.929.826.8-
VLDet [46]CenterNet v221.729.834.330.1ICLR-2023
CoDet [57]CenterNet v223.430.034.630.7NeirIPS-2023
GOAT [72]CenterNet v223.329.734.330.4ICCV-2023
EdaDet [71]CenterNet v223.727.529.127.5ICCV-2023
CLIFF (ours)CenterNet v224.527.729.828.2-
+ +momentum of 0.9 following [68]. We use 4 NVIDIA V100 for training with a batch size of 16. We pre-train the base object detector with a $1 \times$ schedule as [68], and then train the whole model with a $2 \times$ schedule. We also explore the longer $8 \times$ training schedule to make fair comparisons with specific methods. + +CLIP Model. As the previous work [68], we use the single prompt, 'a photo of a {category}', to compute CLIP text embedding $\hat{\mathbf{x}}_{\mathrm{txt}}$ , and use $224\times 224$ input-scale for image embedding $\hat{\mathbf{x}}_{\mathrm{img}}$ via ViT-B/32 model [67]. + +CLIFF. The number of the time-step in object-to-text $(T)$ and object-to-image $(L)$ diffusion are set $T = 10, L = 3$ , which are justified in Table 4. $\alpha$ , $\beta_{1}$ , and $\beta_{2}$ are empirically set to 2.0, 15.0, and 1.0, respectively, which are verified by the experiments in Appendix. $L_{2}$ normalization is deployed on CLIP embedding as [15,68]. Class-specific and RPN proposals are used to optimize object-to-text diffusion [68,75], while class-agnostic ones are for object-to-image diffusion. + +# 5.3 Comparison with State-of-the-Art Methods + +Results on COCO Benchmark. As shown in Table 1, we compare the proposed method with state-of-the-art counterparts on the COCO benchmark. CLIFF achieves $41.3\% \mathrm{mAP_n}$ , which outperforms OCD [68] ( $36.6\% \mathrm{mAP_n}$ ), VLDet [46] ( $32.0\% \mathrm{mAP_n}$ ) and BARON [75] ( $41.0\% \mathrm{mAP_n}$ ) with $4.7\%$ , $9.3\%$ and $0.3\% \mathrm{mAP_n}$ , comprehensively verifying the effectiveness of the proposed CLIFF. The proposed method also surpasses EdaDet [71] and GOAT [72] by a noticeable margin. Besides, after using fair multi-scale training [68, 75], our method ( $43.2\% \mathrm{mAP_n}$ ) outperforms existing works by a large margin. Unlike existing works aligning object-text and object-image separately, we formulate a continual diffusion process among the three sub-spaces in a different learning paradigm. + +Results on LVIS Benchmark. We present a thorough comparison with corresponding state-of-the-art methods using Mask RCNN and CenterNet v2 detectors, which is shown in Table 2. For Mask RCNN, it can be observed that CLIFF achieves the best $22.4\%$ $\mathrm{AP}_{\mathrm{r}}^{\mathrm{m}}$ , which surpasses state-of-the-art counterparts + +Table 3: Ablation study results on COCO and LVIS benchmarks with $1 \times$ schedule. + +
CDM o2t o2i VLSVLSCOCOLVIS-BoxLVIS-Mask
mAPnmAPbmAPAPbAPcAPbAPbAPmAPmAPcAPb
1(baseline)32.249.945.315.820.126.221.415.219.725.020.5
2✓36.649.546.116.220.927.422.615.820.125.721.6
3✓39.552.449.216.721.727.723.116.520.925.922.1
4✓38.853.349.317.321.927.823.316.721.126.122.2
5✓40.052.949.518.922.128.023.818.121.426.222.7
+ +Table 4: Analysis on the diffusion step. o2i and o2t indicates the object-to-image and object-to-text diffusion process respectively. * indicates sharing the same diffusion procedure and optimizing o2i and o2t in a combinatorial style [15,68]. + +
o2io2t\( \mathrm{mAP}_{\mathrm{n}} \)\( \mathrm{mAP}_{\mathrm{b}} \)mAP
111040.253.450.0
231041.354.150.6
351041.353.750.3
471041.253.950.5
591041.453.550.4
63540.953.850.3
731041.354.150.6
832039.753.449.8
9101041.453.650.4
10*10*1039.952.849.5
+ +Table 5: Analysis of the discriminative and generative terms in Eq. 12 by replacing the random noise (Rand) with object-centric noise (OC) in different time steps $(2\times$ schedule). rows 7 and 8 indicate removing individual terms. + +
DiscriminativeGenerative\( \mathrm{mAP}_{\mathrm{n}} \)
TT-1: 0TT-1: 0
1RandRandRandRand39.2
2OCRandOCRand41.0
3RandOCRandOC40.4
4OCOCOCRand41.1
5OCRandOCOC41.3
6OCOCOCOC41.3
7w/o DiscriminativeOCOC38.2
8OCOCw/o Generative40.1
+ +BARON [75] $(19.2\% \mathrm{AP_r^m})$ , OCD [68] $(21.3\% \mathrm{AP_r^m})$ with $3.1\% \mathrm{AP_r^m}$ and $1.3\% \mathrm{AP_r^m}$ , respectively. Besides, our method surpasses the baseline model MAVL-OVD [58] with a significant $5.4\% \mathrm{AP_r^m}$ , verifying the effectiveness of our method. For the comparison with CenterNet v2, we can find that our method gives $24.5\% \mathrm{AP_r^m}$ and also surpasses the latest published works noticeably, outperforming CoDet [57] and GOAT [72] with $1.1\%$ and $1.2\%$ gains. This can reveal that CLIFF is a generic method in terms of multi-modal alignment and can be deployed in different object detectors with great transferability and adaptability. + +# 5.4 Quantitative Analysis + +Ablation Study. The ablative results are shown in Table 3 with a $1 \times$ training schedule. The first line indicates the model trained with extra MAVL proposals [58, 68], which is a fair baseline model as [68, 75]. Compared with the baseline model $(32.2\% \mathrm{mAP}_{\mathrm{n}}, 15.8\% \mathrm{AP}_{\mathrm{r}}^{\mathrm{b}}, 15.2\% \mathrm{AP}_{\mathrm{r}}^{\mathrm{m}})$ , introducing the object-to-text (row 2) diffusion gives satisfactory improvements with $36.6\% \mathrm{mAP}_{\mathrm{n}}, 16.2\% \mathrm{AP}_{\mathrm{r}}^{\mathrm{b}}$ and $15.8\% \mathrm{AP}_{\mathrm{r}}^{\mathrm{m}}$ . After integrating object-to-image diffusion (row 3), i.e., the full CDM, we can observe a further improvement with $7.3\% \mathrm{mAP}_{\mathrm{n}}, 0.9\% \mathrm{AP}_{\mathrm{r}}^{\mathrm{b}}$ and $1.3\% \mathrm{AP}_{\mathrm{r}}^{\mathrm{m}}$ gains, compared with the baseline. This verifies the superior effectiveness in modeling a continual distribution transfer than the existing discriminative paradigm. Besides, after replacing the randomly sampled noise by + +Table 6: Analysis of the refined condition ${\mathbf{c}}_{o2t}$ in object-to-image diffusion (Eq. 5) with $2 \times$ training schedule. + +
Setting of c02tmAPnmAPbmAP
1Object Embed: xobj40.453.449.9
2Image Embed: ximg39.652.949.4
3Additive: xobj + ximg40.853.550.2
4Concated: [xobj : ximg]41.053.650.4
5Attentive: Eq. 1041.354.150.6
+ +Table 7: Cross-dataset evaluation with $\mathrm{AP}_{50}$ (\%) transferring COCO-trained models to the PASCAL VOC and LVIS. + +
MethodPASCALLVIS
OVR-CNN [82]52.95.2
PB-OVD [14]59.28.0
VLDet [46]61.710.0
GOAT [72]63.614.0
CLIFF (Ours)66.116.7
+ +object-centric noise with VLS (row 5), it can be observed that the CDM (row 3) can be improved with $40.0\% \mathrm{mAP}_{\mathrm{n}}$ , $18.9\% \mathrm{AP}_{\mathrm{r}}^{\mathrm{b}}$ , and $18.1\% \mathrm{AP}_{\mathrm{r}}^{\mathrm{m}}$ , verifying the effectiveness of each component in the proposed CLIFF. + +The Number of Diffusion Steps. We study different diffusion steps for o2i and o2t in Table 4, and have the following three observations. First, in rows 1-5, we observe that the results are stable when $L > 3$ (around $41.3\% \mathrm{mAP_n}$ ), but give notable performance drops when the step is insufficient $L = 1$ . Considering more computation requirements when enlarging the diffusion step, we select $L = 3$ with the best trade-off. Second, in rows 6-8, setting too small and large o2t steps both give significant performance drops, which may be caused by unsatisfactory diffusion optimization. Third, we set consistent diffusion steps and compare the proposed continual diffusion (row 9) with existing combinatorial style, i.e., sharing the diffusion process and generated embedding between the o2i and o2t (row 10). With a $1.5\%$ performance gain over the combinatorial objective, the effectiveness of our continual transfer can be justified. + +Exploring Object-centric Noise. We further give a detailed analysis of the object-centric (OC) noise by replacing the random noise for the discriminative and generative terms in object-to-text objective (the core process for OVD) in different time steps, as shown in Table 5. Compared with the random-noise baseline with $39.2\% \mathrm{mAP}_{\mathrm{n}}$ , controlling the initial stage (row 2) and middle stages (row 3) gives $41.0\%$ are $40.4\% \mathrm{mAP}_{\mathrm{n}}$ , respectively. Introducing more OC noise in the diffusion (rows 4-6) gives stable performance improvements, indicating the effectiveness of the VLS. Further, we conduct an ablative study on o2t diffusion (rows 7-8), i.e., the first and second terms in Eq. 12, verifying their critical role in optimizing the diffusion process with both aspects. + +Analyzing the Condition Design. The strategy of refining the object-to-text (Eq. 5) condition $\mathbf{c}_{o2t} = \mathbf{x}_{\mathrm{obj}}^{\mathrm{img}}$ is analyzed in Table 6. Compared with individually using the object and image (rows 1-2) condition with $40.4\%$ and $39.6\%$ $\mathrm{mAP_n}$ , combining the two aspects for the diffusion condition generally works better (rows 3-5), e.g., addition with $40.8\%$ $\mathrm{mAP_n}$ , concatenation with $41.0\%$ . Besides, we observe that the proposed attentive refinement on the condition (Eq. 10) achieves optimal performance and gives the best $41.3\%$ $\mathrm{mAP_n}$ , due to the adaptive fusion of image-level context. This can verify the effectiveness of our design. + +Generalization to Unseen Datasets. We further conduct the cross-dataset comparison following [72] by considering two different settings. Compared with + +![](images/2cdba09ea8db7d68a01774b1fca21647f413b2a958854e216770d5fef9d3f8d9.jpg) +Fig. 3: Qualitative comparison among (a) the baseline [68], (b) our CLIFF, and (c) ground-truth on the COCO benchmark. We only visualize the novel-class predictions for a clearer evaluation of open-vocabulary generalization. + +state-of-the-art counterpart GOAT [72], the proposed CLIFF gives $66.1\%$ and $16.7\% \mathrm{AP}_{50}$ on the PASCAL VOC and LVIS datasets using the COCO-trained model, which surpasses GOAT with significant $2.5\%$ and $2.7\%$ gains, showing the superior generalization capacity of our method. + +# 5.5 Qualitative Analysis + +As shown in Fig. 3, we visualize the novel-class prediction and conduct a qualitative comparison among (a) the baseline [68,75] trained with MAVL proposals [68], (b) the proposed CLIFF, and (c) ground-truth. It can be observed that the proposed method gives more accurate novel-class predictions, especially for occlusion situations and complex scenarios. Our method reduces the confusion among novel classes, e.g., correcting the wrong prediction (couch) given by the baseline in col. 1. Besides, the proposed CLIFF can relieve some missing errors, as illustrated in col. 2-5, revealing its superior novel-class generalization capacity. Hence, the proposed CLIFF is able to relieve the bias on the base classes and encourage the model to discover the novel-class objects with probabilistic modeling. + +# 6 Conclusion + +This paper proposes a novel diffusion-based framework for OVD, coined CLIFF, which breaks through the bottlenecks of the combinatorial and discriminative limitation in existing OVD pipelines. The proposed CLIFF consists of a Variational Latent Sampler (VLS) to sample object-centric noise in the probabilistic object space, and leverages a Continual diffusion Module (CDM) to formulate a continual distribution transfer among three sub-spaces (object, image, and text) with cross-modal nature. Extensive experiments validate that the proposed method significantly outperforms existing approaches on large-scale benchmarks. + +Acknowledgement. This work was supported by Innovation and Technology Commission- Innovation and Technology Fund ITS/229/22 and Hong Kong Research Grants Council (RGC) General Research Fund 11211221. + +# References + +1. Ali, S., Ghatwary, N., Jha, D., Isik-Polat, E., Polat, G., Yang, C., Li, W., Galdran, A., Ballester, M.A.G., Thambawita, V., et al.: Assessing generalisability of deep learning-based polyp detection and segmentation methods through a computer vision challenge. Sci. Rep. (2024) +2. Amit, T., Nachmani, E., Shaharbany, T., Wolf, L.: Segdiff: Image segmentation with diffusion probabilistic models. ArXiv:2112.00390 (2021) +3. Bansal, A., Borgnia, E., Chu, H.M., Li, J.S., Kazemi, H., Huang, F., Goldblum, M., Geiping, J., Goldstein, T.: Cold diffusion: Inverting arbitrary image transforms without noise. ArXiv:2208.09392 (2022) +4. Baranchuk, D., Voynov, A., Rubachev, I., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. In: ICLR (2022) +5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV. pp. 213-229 (2020) +6. Chen, S., Sun, P., Song, Y., Luo, P.: Diffusiondet: Diffusion model for object detection. ArXiv:2211.09788 (2022) +7. Chen, Z., Li, W., Xing, X., Yuan, Y.: Medical federated learning with joint graph purification for noisy label learning. MedIA (2023) +8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009) +9. Du, Y., Wei, F., Zhang, Z., Shi, M., Gao, Y., Li, G.: Learning to prompt for open-vocabulary object detection with vision-language model. ArXiv:2203.14940 (2022) +0. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. IJCV 88, 303-338 (2010) +1. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Expand your detector vocabulary with uncurated images. ArXiv:2203.16513 (2022) +2. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Towards open-vocabulary detection using uncurated images. In: ECCV. pp. 701-717. Springer (2022) +3. Gao, M., Xing, C., Niebles, J.C., Li, J., Xu, R., Liu, W., Xiong, C.: Towards open vocabulary object detection without human-provided bounding boxes. ArXiv:2111.09452 (2021) +4. Gao, M., Xing, C., Niebles, J.C., Li, J., Xu, R., Liu, W., Xiong, C.: Open vocabulary object detection with pseudo bounding-box labels. In: ECCV (2022) +5. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: ICLR (2022) +6. Gu, Z., Chen, H., Xu, Z., Lan, J., Meng, C., Wang, W.: Diffusioninst: Diffusion model for instance segmentation. ArXiv:2212.02773 (2022) +7. Gupta, A., Dollar, P., Girshick, R.: LVIS: A dataset for large vocabulary instance segmentation. In: CVPR (2019) +8. Han, J., Ren, Y., Ding, J., Yan, K., Xia, G.S.: Few-shot object detection via variational feature aggregation. ArXiv:2301.13411 (2023) + +19. Han, X., Zheng, H., Zhou, M.: Card: Classification and regression diffusion models. In: NeurIPS (2022) +20. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask R-CNN. In: CVPR (2017) +21. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020) +22. Kim, D., Angelova, A., Kuo, W.: Contrastive feature masking open-vocabulary vision transformer. In: ICCV. pp. 15602-15612 (2023) +23. Kim, D., Angelova, A., Kuo, W.: Region-aware pretraining for open-vocabulary object detection with vision transformers. In: CVPR. pp. 11144-11154 (2023) +24. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. ArXiv:1312.6114 (2013) +25. Korte, B.H., Vygen, J., Korte, B., Vygen, J.: Combinatorial optimization, vol. 1. Springer (2011) +26. Kuo, W., Cui, Y., Gu, X., Piergiovanni, A., Angelova, A.: Open-vocabulary object detection upon frozen vision and language models. In: ICLR (2023) +27. Kwon, M., Jeong, J., Uh, Y.: Diffusion models already have a semantic latent space. In: ICLR (2023) +28. Leng, Y., Chen, Z., Guo, J., Liu, H., Chen, J., Tan, X., Mandic, D., He, L., Li, X.Y., Qin, T., et al.: Binauralgrad: A two-stage conditional diffusion probabilistic model for binaural audio synthesis. ArXiv:2205.14807 (2022) +29. Li, A.C., Prabhudesai, M., Duggal, S., Brown, E., Pathak, D.: Your diffusion model is secretly a zero-shot classifier. ArXiv:2303.16203 (2023) +30. Li, C., Feng, B.Y., Fan, Z., Pan, P., Wang, Z.: Steganerf: Embedding invisible information within neural radiance fields. In: ICCV. pp. 441-453 (2023) +31. Li, C., Lin, X., Mao, Y., Lin, W., Qi, Q., Ding, X., Huang, Y., Liang, D., Yu, Y.: Domain generalization on medical imaging classification using episodic training with task augmentation. Computers in biology and medicine 141, 105144 (2022) +32. Li, C., Liu, H., Fan, Z., Li, W., Liu, Y., Pan, P., Yuan, Y.: Gaussianstego: A generalizable stenography pipeline for generative 3d gaussians splatting. ArXiv:2407.01301 (2024) +33. Li, C., Liu, H., Liu, Y., Feng, B.Y., Li, W., Liu, X., Chen, Z., Shao, J., Yuan, Y.: Endora: Video generation models as endoscopy simulators. ArXiv:2403.11050 (2024) +34. Li, C., Liu, X., Li, W., Wang, C., Liu, H., Yuan, Y.: U-kan makes strong backbone for medical image segmentation and generation. ArXiv:2406.02918 (2024) +35. Li, C., Zhang, Y., Li, J., Huang, Y., Ding, X.: Unsupervised anomaly segmentation using image-semantic cycle translation. ArXiv:2103.09094 (2021) +36. Li, C., Zhang, Y., Liang, Z., Ma, W., Huang, Y., Ding, X.: Consistent posterior distributions under vessel-mixing: a regularization for cross-domain retinal artery/vein classification. In: ICIP. pp. 61-65 (2021) +37. Li, W., Chen, Z., Li, B., Zhang, D., Yuan, Y.: Htd: Heterogeneous task decoupling for two-stage object detection. TIP (2021) +38. Li, W., Guo, X., Yuan, Y.: Novel scenes & classes: Towards adaptive open-set object detection. In: ICCV. pp. 15780-15790 (2023) +39. Li, W., Liu, J., Han, B., Yuan, Y.: Adjustment and alignment for unbiased open set domain adaptation. In: CVPR. pp. 24110-24119 (2023) +40. Li, W., Liu, X., Yao, X., Yuan, Y.: Scan: Cross domain object detection with semantic conditioned adaptation. In: AAAI. pp. 1421-1428 (2022) +41. Li, W., Liu, X., Yuan, Y.: Scan++: Enhanced semantic conditioned adaptation for domain adaptive object detection. TMM (2022) + +42. Li, W., Liu, X., Yuan, Y.: Sigma: Semantic-complete graph matching for domain adaptive object detection. In: CVPR (2022) +43. Li, W., Liu, X., Yuan, Y.: Sigma++: Improved semantic-complete graph matching for domain adaptive object detection. TPAMI (2023) +44. Li, W., Yang, C., Liu, J., Liu, X., Guo, X., Yuan, Y.: Joint polyp detection and segmentation with heterogeneous endoscopic data. In: ISBI Workshop: EndoCV 2021. pp. 69-79. CEUR-WS Team (2021) +45. Liang, C., Wang, W., Miao, J., Yang, Y.: Gmmseg: Gaussian mixture based generative semantic segmentation models. ArXiv:2210.02025 (2022) +46. Lin, C., Sun, P., Jiang, Y., Luo, P., Qu, L., Haffari, G., Yuan, Z., Cai, J.: Learning object-language alignments for open-vocabulary object detection. In: ICLR (2023) +47. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014) +48. Lipman, Y., Chen, R.T.Q., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. In: ICLR (2023) +49. Liu, H., Liu, Y., Li, C., Li, W., Yuan, Y.: Lgs: A light-weight 4d gaussian splatting for efficient surgical scene reconstruction. ArXiv:2406.16073 (2024) +50. Liu, X., Li, W., Yamaguchi, T., Geng, Z., Tanaka, T., Tsai, D.P., Chen, M.K.: Stereo vision meta-lens-assisted driving vision. ACS Photonics (2024) +51. Liu, X., Gong, C., Liu, Q.: Flow straight and fast: Learning to generate and transfer data with rectified flow. ArXiv:2209.03003 (2022) +52. Liu, X., Li, W., Yang, Q., Li, B., Yuan, Y.: Towards robust adaptive object detection under noisy annotations. In: CVPR. pp. 14207-14216 (2022) +53. Liu, X., Li, W., Yuan, Y.: Decoupled unbiased teacher for source-free domain adaptive medical object detection. TNNLS (2023) +54. Liu, Y., Li, C., Yang, C., Yuan, Y.: Endogaussian: Gaussian splatting for deformable surgical scene reconstruction. ArXiv:2401.12561 (2024) +55. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. ArXiv:2211.01095 (2022) +56. Luo, J., Li, Y., Pan, Y., Yao, T., Feng, J., Chao, H., Mei, T.: Semantic-conditional diffusion networks for image captioning. ArXiv:2212.03099 (2022) +57. Ma, C., Jiang, Y., Wen, X., Yuan, Z., Qi, X.: Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. NeurIPS 36 (2024) +58. Maaz, M., Rasheed, H., Khan, S., Khan, F.S., Anwer, R.M., Yang, M.H.: Class-agnostic object detection with multi-modal transformer. In: ECCV (2022) +59. Miller, D., Sünderhauf, N., Milford, M., Dayoub, F.: Uncertainty for identifying open-set errors in visual object detection. Robot. Lett. 7(1), 215-222 (2021) +60. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., et al.: Simple open-vocabulary object detection. In: ECCV. pp. 728-755 (2022) +61. Ni, J., Qiu, Q., Chellappa, R.: Subspace interpolation via dictionary learning for unsupervised domain adaptation. In: CVPR (June 2013) +62. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: ICML (2021) +63. Norris, J.R.: Markov chains. No. 2, Cambridge university press (1998) +64. Pan, P., Fan, Z., Feng, B.Y., Wang, P., Li, C., Wang, Z.: Learning to estimate 6dof pose from limited data: A few-shot, generalizable approach using rgb images. In: 2024 International Conference on 3D Vision (3DV). pp. 1059-1071. IEEE (2024) +65. Peebles, W., Xie, S.: Scalable diffusion models with transformers. ArXiv:2212.09748 (2022) + +66. Preechakul, K., Chathee, N., Wizadwongsa, S., Suwajanakorn, S.: Diffusion autoencoders: Toward a meaningful and decodable representation. In: CVPR. pp. 10619-10629 (2022) +67. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) +68. Rasheed, H.A., Maaz, M., Khattak, M.U., Khan, S., Khan, F.: Bridging the gap between object and image-level representations for open-vocabulary detection. In: NeurIPS (2022) +69. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. In: NeurIPS (2015) +70. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) +71. Shi, C., Yang, S.: Edadet: Open-vocabulary object detection using early dense alignment. In: ICCV. pp. 15724-15734 (2023) +72. Wang, J., Zhang, H., Hong, H., Jin, X., He, Y., Xue, H., Zhao, Z.: Open-vocabulary object detection with an open corpus. In: ICCV. pp. 6759-6769 (2023) +73. Wizadwongsa, S., Suwajanakorn, S.: Accelerating guided diffusion sampling with splitting numerical methods. In: ICLR (2023) +74. Wu, L., Gong, C., Liu, X., Ye, M., Liu, Q.: Diffusion-based molecule generation with informative prior bridges. ArXiv:2209.00865 (2022) +75. Wu, S., Zhang, W., Jin, S., Liu, W., Loy, C.C.: Aligning bag of regions for open-vocabulary object detection. ArXiv:2302.13996 (2023) +76. Wu, X., Zhu, F., Zhao, R., Li, H.: Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. ArXiv:2303.13076 (2023) +77. Xu, H., Zhang, Y., Sun, L., Li, C., Huang, Y., Ding, X.: Afsc: Adaptive fourier space compression for anomaly detection. ArXiv:2204.07963 (2022) +78. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. ArXiv:2303.04803 (2023) +79. Yang, Q., Li, W., Li, B., Yuan, Y.: Mrm: Masked relation modeling for medical image pre-training with genetics. In: ICCV. pp. 21452-21462 (2023) +80. Ye, K., Zhang, M., Kovashka, A., Li, W., Qin, D., Berent, J.: Cap2det: Learning to amplify weak caption supervision for object detection. In: ICCV (2019) +81. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary detr with conditional matching. In: ECCV. pp. 106-122 (2022) +82. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: CVPR (2021) +83. Zbinden, L., Doorenbos, L., Pissas, T., Huber, A.T., Sznitman, R., Márquez-Neila, P.: Stochastic segmentation with conditional categorical diffusion models. In: ICCV (2023) +84. Zhang, Y., Li, C., Lin, X., Sun, L., Zhuang, Y., Huang, Y., Ding, X., Liu, X., Yu, Y.: Generator versus segmentor: Pseudo-healthy synthesis. In: MICCAI. pp. 150–160 (2021) +85. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: CVPR. pp. 16793-16803 (2022) +86. Zhou, X., Girdhar, R., Joulin, A., Krahenbuhl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: ECCV. pp. 350-368 (2022) +87. Zhou, X., Koltun, V., Krahenbuhl, P.: Probabilistic two-stage detection. In: ArXiv:2103.07461 (2021) \ No newline at end of file diff --git a/2024/CLIFF_ Continual Latent Diffusion for Open-Vocabulary Object Detection/images.zip b/2024/CLIFF_ Continual Latent Diffusion for Open-Vocabulary Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4625311fd9391b7c40ec02cba144d0a19269ee06 --- /dev/null +++ b/2024/CLIFF_ Continual Latent Diffusion for Open-Vocabulary Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6803ae52dd3899f993f2a64fc9c80736aa9d034628865726e844eb2b30e09158 +size 545484 diff --git a/2024/CLIFF_ Continual Latent Diffusion for Open-Vocabulary Object Detection/layout.json b/2024/CLIFF_ Continual Latent Diffusion for Open-Vocabulary Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5f88b8ec9645278fa882c6a69482a46e520bb30e --- /dev/null +++ b/2024/CLIFF_ Continual Latent Diffusion for Open-Vocabulary Object Detection/layout.json @@ -0,0 +1,13364 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 173, + 111, + 442, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 111, + 442, + 148 + ], + "spans": [ + { + "bbox": [ + 173, + 111, + 442, + 148 + ], + "type": "text", + "content": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 164, + 167, + 450, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 167, + 450, + 180 + ], + "spans": [ + { + "bbox": [ + 164, + 167, + 450, + 180 + ], + "type": "text", + "content": "Wuyang " + }, + { + "bbox": [ + 164, + 167, + 450, + 180 + ], + "type": "inline_equation", + "content": "\\mathrm{Li}^{1}" + }, + { + "bbox": [ + 164, + 167, + 450, + 180 + ], + "type": "text", + "content": " , Xinyu Liu, Jiayi Ma2, and Yixuan Yuan" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 203, + 189, + 410, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 189, + 410, + 224 + ], + "spans": [ + { + "bbox": [ + 203, + 189, + 410, + 224 + ], + "type": "text", + "content": "1 The Chinese University of Hong Kong \n2 Wuhan University \nwymanbest@outlook.com, yxyuan@ee.cuhk.edu.hk" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 251, + 455, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 251, + 455, + 504 + ], + "spans": [ + { + "bbox": [ + 160, + 251, + 455, + 504 + ], + "type": "text", + "content": "Abstract. Open-vocabulary object detection (OVD) utilizes image-level cues to expand the linguistic space of region proposals, thereby facilitating the detection of diverse novel classes. Recent works adapt CLIP embedding by minimizing the object-image and object-text discrepancy combinatorially in a discriminative paradigm. However, they ignore the underlying distribution and the disagreement between the image and text objective, leading to the misaligned distribution between the vision and language sub-space. To address the deficiency, we explore the advanced generative paradigm with distribution perception and propose a novel framework based on the diffusion model, coined Continual Latent Diffusion (CLIFF), which formulates a continual distribution transfer among the object, image, and text latent space probabilistically. CLIFF consists of a Variational Latent Sampler (VLS) enabling the probabilistic modeling and a Continual Diffusion Module (CDM) for the distribution transfer. Specifically, in VLS, we first establish a probabilistic object space with region proposals by estimating distribution parameters. Then, the object-centric noise is sampled from the estimated distribution to generate text embedding for OVD. To achieve this generation process, CDM conducts a short-distance object-to-image diffusion from the sampled noise to generate image embedding as the medium, which guides the long-distance diffusion to generate text embedding. Extensive experiments verify that CLIFF can significantly surpass state-of-the-art methods on benchmarks. The code is available at https://github.com/CUHK-AIM-Group/CLIFF." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 514, + 455, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 514, + 455, + 536 + ], + "spans": [ + { + "bbox": [ + 160, + 514, + 455, + 536 + ], + "type": "text", + "content": "Keywords: Open-vocabulary object detection " + }, + { + "bbox": [ + 160, + 514, + 455, + 536 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 514, + 455, + 536 + ], + "type": "text", + "content": " Diffusion model " + }, + { + "bbox": [ + 160, + 514, + 455, + 536 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 514, + 455, + 536 + ], + "type": "text", + "content": " Continual distribution transfer" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 557, + 230, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 557, + 230, + 569 + ], + "spans": [ + { + "bbox": [ + 133, + 557, + 230, + 569 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Object detection [20, 37, 69] has made significant strides in recent years, as demonstrated by the impressive results on benchmarks [17, 47]. Nonetheless, due to the unsatisfactory closed-set constraint [77, 82], object detectors are limited to discovering the novel-class objects [38, 39] unseen during training, impeding a reliable scene understanding in the real world. Further, it hinders the journey toward achieving true artificial intelligence since it requires the human-like perceptual ability to recognize unexpected objects appearing everywhere." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 113, + 480, + 178 + ], + "blocks": [ + { + "bbox": [ + 134, + 113, + 480, + 178 + ], + "lines": [ + { + "bbox": [ + 134, + 113, + 480, + 178 + ], + "spans": [ + { + "bbox": [ + 134, + 113, + 480, + 178 + ], + "type": "image", + "image_path": "3e1a0e4c75f05381ad71649270fa6c534b2bcd960881737f6641d1433feecc9a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 182, + 220, + 190 + ], + "lines": [ + { + "bbox": [ + 157, + 182, + 220, + 190 + ], + "spans": [ + { + "bbox": [ + 157, + 182, + 220, + 190 + ], + "type": "text", + "content": "(a) Existing OVD works" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 267, + 182, + 460, + 190 + ], + "lines": [ + { + "bbox": [ + 267, + 182, + 460, + 190 + ], + "spans": [ + { + "bbox": [ + 267, + 182, + 460, + 190 + ], + "type": "text", + "content": "(b) The proposed CLIFF framework modeling a continual distribution transfer" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 201, + 482, + 236 + ], + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 236 + ], + "type": "text", + "content": "Fig. 1: Illustration of (a) existing discriminative paradigm [15,68,75] with a combinatorial objective (object→image and object→text in parallel), and (b) the proposed method conducting a continual diffusion (object→image→text) with generative nature." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 260, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 403 + ], + "type": "text", + "content": "To break through this barrier, open-vocabulary object detection (OVD) has been studied that leverages auxiliary linguistic cues to generalize the detection of novel classes [82]. Recent advances adapt CLIP [67], a vision-language (VL) foundation model with strong zero-shot capacity, to boost the novel-class generalization of the object detector. Some works [13, 14] utilize the CLIP prediction (image-level) to generate pseudo labels at the object level, enabling the self-training of the object detector. To fully explore the hidden knowledge, some other works strive to align the object-level representation with CLIP embeddings in the hidden space. During training the object detector, they minimize the object-image [15, 68, 75] and object-text [46, 68, 76] discrepancy to align the sub-space among the object, image, and text modalities, thereby unleashing the novel-class generalization capacity from the pre-trained CLIP model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 404, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 482, + 571 + ], + "type": "text", + "content": "Despite the significant success, the existing OVD paradigm, as depicted in Fig. 1(a), has two limitations regarding the sub-space alignment. First, existing works utilize text embedding as the classifier weight to explicitly optimize the discriminative class separation, which is agnostic to the underlying distribution and may lead to class bias [18]. This distribution-agnostic property results in the distribution misalignment [40, 42, 46] and severely hinders the novel-class generalization [78]. Besides, due to the uncalibrated latent space in discriminative models [45], e.g., numerical high-variance, the object distribution is prone to overfit to the labeled classes, worsening the misalignment with the unbiased CLIP distribution. With the rise of generative paradigm [2], this barrier has been gradually broken via diverse probabilistic modeling [18, 45, 59]. Hence, we aim to incorporate probabilistic advantages into discriminative detection pipelines, which can facilitate the distribution transfer across the object and text sub-spaces, thereby achieving an unbiased novel-class perception [45, 59]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 572, + 483, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 572, + 483, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 572, + 483, + 633 + ], + "type": "text", + "content": "Second, some existing works minimize the object-image and object-text discrepancy simultaneously in a combinatorial manner, ignoring the disagreement between the image and text two objectives3. With this combinatorial nature, optimizing object-image alignment as the same role as the object-text counterpart is suboptimal due to the potential conflict between the two objectives [25]." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 200, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 200, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 200, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "text", + "content": "3 As the first attempt, we discover the ever-overlooked disagreement issue and justify it in Appendix via experiments." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "Intuitively, the image and object share visual representation, while the image and text are coupled with CLIP pre-training, which motivates us to introduce the image as an interpolated sub-space [61] between the object and text counterpart, yielding a continual distribution transfer among three modalities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 483, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 483, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 483, + 319 + ], + "type": "text", + "content": "To address the above issues, we propose Continual Latent Diffusion (CLIFF) for OVD, as shown in Fig. 1(b). CLIFF breaks through the discriminative bottleneck via a probabilistic Variational Latent Sampler (VLS) and then formulates a diffusion-based distribution transfer [21] in a Continual Diffusion Module (CDM) to avoid the combinatorial limitation. Specifically, VLS leverages variational approximation [24] to establish a probabilistic object space with learned distribution parameters, and then samples object-centric noise to boost the diffusion process. CDM formulates the continual diffusion process to transfer the object distribution to the image and text counterpart, avoiding disagreement in the existing combinatorial objective. Besides, different from existing diffusion-based works [27, 78] with unsatisfactory efficiency, CLIFF achieves comparable and even better model efficiency compared with state-of-the-art OVD works, which is verified in Appendix. To be summarized, the contributions are four-fold:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 328, + 481, + 472 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 328, + 481, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 328, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 328, + 481, + 376 + ], + "type": "text", + "content": "- We propose Continual Latent Diffusion (CLIFF) for OVD, which achieves a continual distribution transfer with DDPM-based generation. To the best of our knowledge, this work represents the first attempt to study latent diffusion in VL discriminative tasks and push forward benchmark results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 376, + 481, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 481, + 411 + ], + "type": "text", + "content": "- Variational Latent Sampler is proposed to generate a probabilistic object space with variational modeling, in which the object-centric noise can be sampled with reparameterization to boost the generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 412, + 481, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 412, + 481, + 448 + ], + "spans": [ + { + "bbox": [ + 138, + 412, + 481, + 448 + ], + "type": "text", + "content": "- We design an efficient Continual Diffusion Module to facilitate the distribution transfer among three cross-modality sub-spaces, namely the object, image, and text space, using a new diffusion module and unified DDPM formulation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 448, + 481, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 448, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 138, + 448, + 481, + 472 + ], + "type": "text", + "content": "- CLIFF is extensively validated through different benchmarks, which verify its state-of-the-art role and show great adaptability to various detectors." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 490, + 243, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 490, + 243, + 502 + ], + "spans": [ + { + "bbox": [ + 131, + 490, + 243, + 502 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 514, + 338, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 514, + 338, + 527 + ], + "spans": [ + { + "bbox": [ + 131, + 514, + 338, + 527 + ], + "type": "text", + "content": "2.1 Open-Vocabulary Object Detection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Open-vocabulary object detection (OVD) leverages auxiliary image-level labels with linguistic cues, e.g., classification [68] and caption datasets [80], to enlarge the semantic space of the region proposals [5, 69], thereby facilitating the novel-class generalization for the real-world applications, such as automatic driving [41, 43, 50, 52], medical imaging analysis [31, 35, 36, 49] and diagnostic decision-making [1, 7, 44, 53, 79]. Zareian et al. [82] formulate the standard OVD problem setting and propose OVR-CNN to project the visual space to the linguistic counterpart on Faster-RCNN [69]. Some other works [22, 23, 60] delve into OVD with Vision Transformer (ViT) by designing large-scale region-aware pretraining [23], feature masking [22], which fully unleash the potential of ViT in fine-grained detection tasks. Recent works adapt CLIP [67] to objects detectors for the novel-class" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 143, + 90, + 447, + 103 + ], + "type": "text", + "content": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": "generalization. The other works adapt the CLIP-encoded latent embedding and aim to align three critical sub-spaces: the object space, the CLIP image space, and the CLIP text space. They bridge the gap between the object and text sub-space via region-text contrastive pre-training [85], region-word bipartite matching [46], CLIP-based conditional matching [81], and learnable detection-friendly prompts [9, 11, 12]. PromptDet [12] and DetPro [9] use learnable text prompts to learn with noisy uncurated web images, thereby enhancing the object-text alignment. Additionally, recent advances [15, 68, 75] adapt the image sub-space with the embedding distillation [15], weight transfer [68] and bags of regions [75], to fully use CLIP embedding, serving as a combinatorial objective with the object-text alignment. Existing works follow a similar discriminative paradigm for OVD, which cannot model the underlying distribution with a suboptimal distribution alignment. Differently, we formulate a continual transfer in a unified denoising diffusion module, enabling the distribution-level perception and adaptation among the object, image, and text sub-spaces." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 313, + 362, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 313, + 362, + 326 + ], + "spans": [ + { + "bbox": [ + 131, + 313, + 362, + 326 + ], + "type": "text", + "content": "2.2 Denoising Diffusion Probabilistic Model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 333, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 482, + 464 + ], + "type": "text", + "content": "As the advanced generative paradigm, Denoising Diffusion Probabilistic Model (DDPM) [21] (\"diffusion model\" for short) generates samples by formulating a gradual denoising process. This process can be treated as a distribution transfer [51] from the normal distribution to the data in a flow-like manner [51, 54, 84]. Instead of diffusing in the data space, LDM [70] and VDM [66] deploy the denoising diffusion in the latent space to enable the high-resolution synthesis. Recent works can be generally categorized into faster sampling [55, 73], noise corruption designs [3, 62], and diffusion model architecture [33, 34, 65], etc. The diffusion model has achieved remarkable generative performance in the image, text, video, audio, graph, and motion generation [21, 28, 30, 32, 74] with different diffusion conditions for downstream applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 465, + 482, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 465, + 482, + 585 + ], + "spans": [ + { + "bbox": [ + 130, + 465, + 482, + 585 + ], + "type": "text", + "content": "In addition to generative tasks, recent works [6,16,27,56,78,83] have discovered that the diffusion feature has adequate semantic discriminability with richer distribution-level perception and data-efficient property [4,64] than conventional discriminative models [67]. Without fine-tuning, an off-the-shelled diffusion model can achieve satisfactory zero-shot classification [29]. Along this trend, some bottlenecks in discriminative models are gradually broken by involving the generative paradigm [78]. CARD [19] uses the diffusion model to estimate the uncertainty to improve image classification. Unlike existing works considering the noise " + }, + { + "bbox": [ + 130, + 465, + 482, + 585 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 465, + 482, + 585 + ], + "type": "text", + "content": " data diffusion, we formulate a continual distribution transfer among three modalities, yielding the object " + }, + { + "bbox": [ + 130, + 465, + 482, + 585 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 465, + 482, + 585 + ], + "type": "text", + "content": " image " + }, + { + "bbox": [ + 130, + 465, + 482, + 585 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 465, + 482, + 585 + ], + "type": "text", + "content": " text transfer in the latent space." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 603, + 234, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 603, + 234, + 616 + ], + "spans": [ + { + "bbox": [ + 131, + 603, + 234, + 616 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 628, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 628, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 628, + 482, + 667 + ], + "type": "text", + "content": "Problem Formulation. Assuming a set of base classes " + }, + { + "bbox": [ + 131, + 628, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_B" + }, + { + "bbox": [ + 131, + 628, + 482, + 667 + ], + "type": "text", + "content": " and novel classes " + }, + { + "bbox": [ + 131, + 628, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_N" + }, + { + "bbox": [ + 131, + 628, + 482, + 667 + ], + "type": "text", + "content": ", we have a detection dataset " + }, + { + "bbox": [ + 131, + 628, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{det} = \\{I_i,O_i\\}_{i = 1}^n" + }, + { + "bbox": [ + 131, + 628, + 482, + 667 + ], + "type": "text", + "content": " with object-level labels " + }, + { + "bbox": [ + 131, + 628, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\{O_i\\}_{i = 1}^n" + }, + { + "bbox": [ + 131, + 628, + 482, + 667 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 131, + 628, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_B" + }, + { + "bbox": [ + 131, + 628, + 482, + 667 + ], + "type": "text", + "content": " and another auxiliary dataset " + }, + { + "bbox": [ + 131, + 628, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{aux} = \\{I_i,Y_i\\}_{i = 1}^{n'}" + }, + { + "bbox": [ + 131, + 628, + 482, + 667 + ], + "type": "text", + "content": " with image-level" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 443, + 129 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 443, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 443, + 129 + ], + "type": "text", + "content": "Algorithm 1 Reverse Diffusion Process (Denoising): RevDif " + }, + { + "bbox": [ + 132, + 116, + 443, + 129 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_T,\\mathbf{c},L)" + } + ] + } + ], + "index": 2, + "type": "text" + }, + { + "bbox": [ + 132, + 132, + 453, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 132, + 453, + 144 + ], + "spans": [ + { + "bbox": [ + 132, + 132, + 453, + 144 + ], + "type": "text", + "content": "Require: " + }, + { + "bbox": [ + 132, + 132, + 453, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 132, + 132, + 453, + 144 + ], + "type": "text", + "content": ": initial noise; c: diffusion condition; " + }, + { + "bbox": [ + 132, + 132, + 453, + 144 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 132, + 132, + 453, + 144 + ], + "type": "text", + "content": ": the number reverse steps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 144, + 332, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 144, + 332, + 154 + ], + "spans": [ + { + "bbox": [ + 132, + 144, + 332, + 154 + ], + "type": "text", + "content": "Ensure: " + }, + { + "bbox": [ + 132, + 144, + 332, + 154 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{T - L}" + }, + { + "bbox": [ + 132, + 144, + 332, + 154 + ], + "type": "text", + "content": " generated samples with " + }, + { + "bbox": [ + 132, + 144, + 332, + 154 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 132, + 144, + 332, + 154 + ], + "type": "text", + "content": " steps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 155, + 334, + 213 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 136, + 155, + 259, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 155, + 259, + 164 + ], + "spans": [ + { + "bbox": [ + 136, + 155, + 259, + 164 + ], + "type": "text", + "content": "1: for " + }, + { + "bbox": [ + 136, + 155, + 259, + 164 + ], + "type": "inline_equation", + "content": "t = T, \\dots, T - L + 1" + }, + { + "bbox": [ + 136, + 155, + 259, + 164 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 166, + 207, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 166, + 207, + 175 + ], + "spans": [ + { + "bbox": [ + 137, + 166, + 207, + 175 + ], + "type": "text", + "content": "2: " + }, + { + "bbox": [ + 137, + 166, + 207, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 176, + 334, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 176, + 334, + 193 + ], + "spans": [ + { + "bbox": [ + 137, + 176, + 334, + 193 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 137, + 176, + 334, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t - 1} = \\frac{1}{\\sqrt{\\alpha_t}}\\left(\\mathbf{x}_t - \\frac{1 - \\alpha_t}{\\sqrt{1 - \\bar{\\alpha}_t}}\\pmb {\\epsilon}_\\theta (\\mathbf{x}_t,\\mathbf{c},t)\\right) + \\sigma_t\\mathbf{z}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 193, + 183, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 193, + 183, + 201 + ], + "spans": [ + { + "bbox": [ + 137, + 193, + 183, + 201 + ], + "type": "text", + "content": "4: end for" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 204, + 208, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 204, + 208, + 213 + ], + "spans": [ + { + "bbox": [ + 137, + 204, + 208, + 213 + ], + "type": "text", + "content": "5: return " + }, + { + "bbox": [ + 137, + 204, + 208, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{T - L}" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 236, + 481, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 481, + 297 + ], + "type": "text", + "content": "labels, e.g., the image caption [21] and classification dataset [68], which has the weakly supervised labels in " + }, + { + "bbox": [ + 130, + 236, + 481, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_B \\cup \\mathcal{C}_N" + }, + { + "bbox": [ + 130, + 236, + 481, + 297 + ], + "type": "text", + "content": ". In " + }, + { + "bbox": [ + 130, + 236, + 481, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{det}" + }, + { + "bbox": [ + 130, + 236, + 481, + 297 + ], + "type": "text", + "content": ", each image " + }, + { + "bbox": [ + 130, + 236, + 481, + 297 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 130, + 236, + 481, + 297 + ], + "type": "text", + "content": " is associated with a set of objects " + }, + { + "bbox": [ + 130, + 236, + 481, + 297 + ], + "type": "inline_equation", + "content": "O_i = \\{(x,y,w,h), \\hat{y}\\}" + }, + { + "bbox": [ + 130, + 236, + 481, + 297 + ], + "type": "text", + "content": ", including the bounding-box coordinates " + }, + { + "bbox": [ + 130, + 236, + 481, + 297 + ], + "type": "inline_equation", + "content": "\\{(x,y,w,h)\\}" + }, + { + "bbox": [ + 130, + 236, + 481, + 297 + ], + "type": "text", + "content": " and class labels " + }, + { + "bbox": [ + 130, + 236, + 481, + 297 + ], + "type": "inline_equation", + "content": "\\{\\hat{y}\\} \\subset \\mathcal{C}_B" + }, + { + "bbox": [ + 130, + 236, + 481, + 297 + ], + "type": "text", + "content": ". OVD aims to use " + }, + { + "bbox": [ + 130, + 236, + 481, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{det}" + }, + { + "bbox": [ + 130, + 236, + 481, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 236, + 481, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{aux}" + }, + { + "bbox": [ + 130, + 236, + 481, + 297 + ], + "type": "text", + "content": " for training and detect base and novel objects in " + }, + { + "bbox": [ + 130, + 236, + 481, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_B \\cup \\mathcal{C}_N" + }, + { + "bbox": [ + 130, + 236, + 481, + 297 + ], + "type": "text", + "content": " during the test stage." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "text", + "content": "Diffusion Model. Formulating a generative paradigm with the Markov chain [21, 63], DDPM is able to generate data " + }, + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "text", + "content": " from the noise " + }, + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "text", + "content": " via a " + }, + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "text", + "content": "-step denoising process. Concretely, DDPM begins with the clean data " + }, + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "text", + "content": " and gradually adds Gaussian noise using the fixed forward process " + }, + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_t | \\mathbf{x}_{t-1})" + }, + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "inline_equation", + "content": "t = 1, 2, \\dots, T" + }, + { + "bbox": [ + 130, + 297, + 482, + 357 + ], + "type": "text", + "content": ". This process can be denoted as follows," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 222, + 361, + 482, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 361, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 222, + 361, + 482, + 376 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right), \\tag {1}", + "image_path": "716df1000662dc180f5957f96b9a5916b656ef04d5c5487b5a6eb4f58b05c843.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 380, + 481, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 481, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 481, + 404 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 380, + 481, + 404 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 130, + 380, + 481, + 404 + ], + "type": "text", + "content": " is noise schedule [62]. To recover data " + }, + { + "bbox": [ + 130, + 380, + 481, + 404 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 130, + 380, + 481, + 404 + ], + "type": "text", + "content": " from noise " + }, + { + "bbox": [ + 130, + 380, + 481, + 404 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 130, + 380, + 481, + 404 + ], + "type": "text", + "content": ", the reverse process parameterized with " + }, + { + "bbox": [ + 130, + 380, + 481, + 404 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 380, + 481, + 404 + ], + "type": "text", + "content": " is learned to denoise inputs with condition " + }, + { + "bbox": [ + 130, + 380, + 481, + 404 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 380, + 481, + 404 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 220, + 409, + 481, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 409, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 220, + 409, + 481, + 441 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\mathbf {x} _ {0: T} \\mid \\mathbf {c}\\right) = p \\left(\\mathbf {x} _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}, \\mathbf {c}\\right), \\tag {2}", + "image_path": "7f2e7dbe796c9f5bd38b29f81082abaa9e2e854570ed3ca0c60e646ae12188bc.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 446, + 482, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 505 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 505 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 446, + 482, + 505 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_T)" + }, + { + "bbox": [ + 130, + 446, + 482, + 505 + ], + "type": "text", + "content": " is fixed to Gaussian. In this reverse step, the diffusion model typically learns a neural network " + }, + { + "bbox": [ + 130, + 446, + 482, + 505 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 130, + 446, + 482, + 505 + ], + "type": "text", + "content": " to generate the values for computing the mean of " + }, + { + "bbox": [ + 130, + 446, + 482, + 505 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_{t - 1}\\mid \\mathbf{x}_t,\\mathbf{c})" + }, + { + "bbox": [ + 130, + 446, + 482, + 505 + ], + "type": "text", + "content": " given by the noisy input " + }, + { + "bbox": [ + 130, + 446, + 482, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 130, + 446, + 482, + 505 + ], + "type": "text", + "content": ", which is optimized to minimize the variational lower bound (ELBO). As proved in [21, 29], this training objective can be re-written as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 232, + 508, + 481, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 508, + 481, + 529 + ], + "spans": [ + { + "bbox": [ + 232, + 508, + 481, + 529 + ], + "type": "interline_equation", + "content": "- \\mathbb {E} _ {t, \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, \\mathbf {c}, t) \\| ^ {2} \\right], \\tag {3}", + "image_path": "126a0e1df8c380a17731bf661a8450f02498c5a154bbdc74dd65c5cd2e399d8f.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "text", + "content": "where the diffusion model " + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "text", + "content": " can estimate the added noise of each noisy input " + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "text", + "content": ". After optimizing with Eq. 3, the diffusion model can generate a data sample " + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{T - L}" + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "text", + "content": " at arbitrary time-step " + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "inline_equation", + "content": "T - L" + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "text", + "content": " from a random noise " + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "text", + "content": " steps of the reverse diffusion, optionally guided by the condition " + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "text", + "content": ". This process [21] is represented in Algorithm 1. In the following sections, we use RevDif(" + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 130, + 533, + 482, + 605 + ], + "type": "text", + "content": ") to represent the reverse diffusion (noise→data) for simplicity." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 131, + 619, + 233, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 619, + 233, + 634 + ], + "spans": [ + { + "bbox": [ + 131, + 619, + 233, + 634 + ], + "type": "text", + "content": "4 Methodology" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "text", + "content": "The overall workflow is illustrated in Fig. 2. We feed input images " + }, + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\{I_i\\}_{i=1}^B" + }, + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "text", + "content": " to the backbone and RPN [69] to extract " + }, + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 640, + 481, + 666 + ], + "type": "text", + "content": " region proposals, which are sent to the" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 143, + 90, + 447, + 102 + ], + "type": "text", + "content": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 118, + 480, + 208 + ], + "blocks": [ + { + "bbox": [ + 133, + 118, + 480, + 208 + ], + "lines": [ + { + "bbox": [ + 133, + 118, + 480, + 208 + ], + "spans": [ + { + "bbox": [ + 133, + 118, + 480, + 208 + ], + "type": "image", + "image_path": "fc393f73c102897fa4a24c4fe58915cd200cc74ce2ebe8e250eff65f1273dbba.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 217, + 482, + 251 + ], + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 251 + ], + "type": "text", + "content": "Fig. 2: Illustration of the proposed CLIFF. VLS establishes a probabilistic embedding space with region proposals, where we sampled object-centric noise " + }, + { + "bbox": [ + 130, + 217, + 482, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 130, + 217, + 482, + 251 + ], + "type": "text", + "content": ". Then, CDM achieves a continual distribution transfer from the object to the image and text." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "text", + "content": "feature extractor to generate object embedding " + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{obj}}" + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "text", + "content": ". Then, VLS, parameterized with " + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "text", + "content": ", establishes and estimates the mean and variance of a probabilistic object space " + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "inline_equation", + "content": "p_{\\phi}(\\mathbf{x}_T|\\mathbf{x}_{\\mathrm{obj}})" + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "text", + "content": ", where the reparameterization is conducted to sample object-centric noise " + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "text", + "content": ", as shown in Fig. 2(a). After that, the sampled noise " + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "text", + "content": " is sent to CDM, denoted as " + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "text", + "content": ", to generate text embedding " + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{txt}}" + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "text", + "content": " with diffusion, as depicted in Fig. 2(b). Conditioned on " + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{o2i}" + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "text", + "content": ", we first conduct object-to-image diffusion to generate the image embedding " + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{img}}" + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "text", + "content": " steps (top sub-figure), which is used to generate refined diffusion condition " + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{o2t}" + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "text", + "content": ". Conditioned on " + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{o2t}" + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "text", + "content": ", we further perform a " + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "text", + "content": "-step object-to-text diffusion to generate text embedding " + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{txt}}" + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "text", + "content": " (bottom sub-figure), whose similarity with the CLIP text embedding " + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{\\mathrm{txt}}" + }, + { + "bbox": [ + 130, + 277, + 482, + 408 + ], + "type": "text", + "content": " are measured to serve as the final class prediction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 409, + 483, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 409, + 483, + 459 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 483, + 459 + ], + "type": "text", + "content": "Formally, by extending Eq. 2, CLIFF consists of two diffusion processes among the object, image, and text embedding sub-space. Guided by the condition " + }, + { + "bbox": [ + 130, + 409, + 483, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{o2i}" + }, + { + "bbox": [ + 130, + 409, + 483, + 459 + ], + "type": "text", + "content": ", the object-to-image diffusion, written as " + }, + { + "bbox": [ + 130, + 409, + 483, + 459 + ], + "type": "inline_equation", + "content": "p_{\\theta,\\phi}(\\mathbf{x}_{\\mathrm{img}}|\\mathbf{c}_{o2i},\\mathbf{x}_{\\mathrm{obj}})" + }, + { + "bbox": [ + 130, + 409, + 483, + 459 + ], + "type": "text", + "content": ", aims to generate the associated CLIP image embedding from the object " + }, + { + "bbox": [ + 130, + 409, + 483, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{obj}}" + }, + { + "bbox": [ + 130, + 409, + 483, + 459 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 409, + 483, + 459 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 409, + 483, + 459 + ], + "type": "text", + "content": " steps:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 170, + 470, + 482, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 470, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 170, + 470, + 482, + 504 + ], + "type": "interline_equation", + "content": "p _ {\\theta , \\phi} (\\mathbf {x} _ {\\mathrm {i m g}} | \\mathbf {c} _ {o 2 i}, \\mathbf {x} _ {\\mathrm {o b j}}) = p _ {\\phi} (\\mathbf {x} _ {T} | \\mathbf {x} _ {\\mathrm {o b j}}) \\prod_ {t = T - L + 1} ^ {T} p _ {\\theta} (\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}, \\mathbf {c} _ {o 2 i}). \\tag {4}", + "image_path": "a551fd2419fd16ec6ff67fe19d71b14e81d2f4eee52f4af3e0d88966c48c176b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 514, + 482, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 539 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 539 + ], + "type": "text", + "content": "Similarly, with condition " + }, + { + "bbox": [ + 130, + 514, + 482, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{o2t}" + }, + { + "bbox": [ + 130, + 514, + 482, + 539 + ], + "type": "text", + "content": ", the object-to-text diffusion is designed to generate text embedding with " + }, + { + "bbox": [ + 130, + 514, + 482, + 539 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 514, + 482, + 539 + ], + "type": "text", + "content": " steps for the final prediction, denoted as follows," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 182, + 550, + 482, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 550, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 182, + 550, + 482, + 583 + ], + "type": "interline_equation", + "content": "p _ {\\theta , \\phi} (\\mathbf {x} _ {\\mathrm {t x t}} | \\mathbf {c} _ {o 2 t}, \\mathbf {x} _ {\\mathrm {o b j}}) = p _ {\\phi} (\\mathbf {x} _ {T} | \\mathbf {x} _ {\\mathrm {o b j}}) \\prod_ {t = 1} ^ {T} p _ {\\theta} (\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}, \\mathbf {c} _ {o 2 t}). \\tag {5}", + "image_path": "b92f9b281d20d7fc6b9a5b9b9b267b055070b83f5044f873d34721bed3aaab4b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "The proposed CLIFF offers two significant improvements compared to the standard diffusion process (Eq. 2). First, instead of beginning from a random noise space " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_T)" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", we start from a differential probabilistic space " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "p_{\\phi}(\\mathbf{x}_T|\\mathbf{x}_{\\mathrm{obj}})" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " with objects to generate the text embedding. Second, going beyond a single diffusion between two sub-spaces, we formulate the effective diffusion process among the object, image, and text in different modalities." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 200, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 200, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 200, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 297, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 297, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 297, + 129 + ], + "type": "text", + "content": "4.1 Variational Latent Sampler" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 132, + 482, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 132, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 132, + 482, + 228 + ], + "type": "text", + "content": "With input images " + }, + { + "bbox": [ + 130, + 132, + 482, + 228 + ], + "type": "inline_equation", + "content": "\\{I_i\\}_{i=1}^B" + }, + { + "bbox": [ + 130, + 132, + 482, + 228 + ], + "type": "text", + "content": ", the backbone, RPN4, and the feature extractor are adopted to obtain " + }, + { + "bbox": [ + 130, + 132, + 482, + 228 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 132, + 482, + 228 + ], + "type": "text", + "content": " object embeddings " + }, + { + "bbox": [ + 130, + 132, + 482, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{obj}} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 130, + 132, + 482, + 228 + ], + "type": "text", + "content": ". Then, we send " + }, + { + "bbox": [ + 130, + 132, + 482, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{obj}}" + }, + { + "bbox": [ + 130, + 132, + 482, + 228 + ], + "type": "text", + "content": " to VLS to establish a probabilistic embedding space " + }, + { + "bbox": [ + 130, + 132, + 482, + 228 + ], + "type": "inline_equation", + "content": "p_{\\phi}(\\mathbf{x}_T | \\mathbf{x}_{\\mathrm{obj}})" + }, + { + "bbox": [ + 130, + 132, + 482, + 228 + ], + "type": "text", + "content": " satisfying the DDPM w.r.t the normally distributed assumption for the diffusion-based transfer. This probabilistic space " + }, + { + "bbox": [ + 130, + 132, + 482, + 228 + ], + "type": "inline_equation", + "content": "p_{\\phi}(\\mathbf{x}_T | \\mathbf{x}_{\\mathrm{obj}})" + }, + { + "bbox": [ + 130, + 132, + 482, + 228 + ], + "type": "text", + "content": " can break through the discriminative limitation of [14,15,46,75] without class bias [18]. Besides, it introduces object-level semantics to enable object-to-text transfer with a differential nature, different from the non-optimizable random space " + }, + { + "bbox": [ + 130, + 132, + 482, + 228 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_T)" + }, + { + "bbox": [ + 130, + 132, + 482, + 228 + ], + "type": "text", + "content": " of noise-to-data transfer [21]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 228, + 483, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 483, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 483, + 351 + ], + "type": "text", + "content": "Formally, inspired by VAE [24], we establish the probabilistic space with centered isotropic multivariate Gaussian " + }, + { + "bbox": [ + 130, + 228, + 483, + 351 + ], + "type": "inline_equation", + "content": "p_{\\phi}(\\mathbf{x}_T|\\mathbf{x}_{\\mathrm{obj}}) = \\mathcal{N}(\\mathbf{x}_T;\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 130, + 228, + 483, + 351 + ], + "type": "text", + "content": ", whose distribution parameters are computed from the observed object embedding " + }, + { + "bbox": [ + 130, + 228, + 483, + 351 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{obj}}" + }, + { + "bbox": [ + 130, + 228, + 483, + 351 + ], + "type": "text", + "content": " with non-linear layers [18]. To this end, two fully connected layers are introduced on the shared region feature extractor of the basic object detector, denoted as " + }, + { + "bbox": [ + 130, + 228, + 483, + 351 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{obj}}" + }, + { + "bbox": [ + 130, + 228, + 483, + 351 + ], + "type": "text", + "content": " e.g., the ResNet blocks in Faster RCNN [69] and the MLP in Mask RCNN [20], to learn to estimate the distribution parameters. Specifically, given the object embedding " + }, + { + "bbox": [ + 130, + 228, + 483, + 351 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{obj}}" + }, + { + "bbox": [ + 130, + 228, + 483, + 351 + ], + "type": "text", + "content": ", the variational posterior " + }, + { + "bbox": [ + 130, + 228, + 483, + 351 + ], + "type": "inline_equation", + "content": "p_{\\phi}(\\mathbf{x}_T|\\mathbf{x}_{\\mathrm{obj}})" + }, + { + "bbox": [ + 130, + 228, + 483, + 351 + ], + "type": "text", + "content": " is written as " + }, + { + "bbox": [ + 130, + 228, + 483, + 351 + ], + "type": "inline_equation", + "content": "\\log p_{\\phi}(\\mathbf{x}_T|\\mathbf{x}_{\\mathrm{obj}}^{(i)}) = \\log \\mathcal{N}(\\mathbf{x}_T;\\pmb{\\mu}^{(i)},\\pmb{\\sigma}^{2(i)}\\mathbf{I})" + }, + { + "bbox": [ + 130, + 228, + 483, + 351 + ], + "type": "text", + "content": ". Without losing the differential property [24], the mean and s.d. of the probabilistic object space can be estimated," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 249, + 365, + 481, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 365, + 481, + 388 + ], + "spans": [ + { + "bbox": [ + 249, + 365, + 481, + 388 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} _ {(i)} ^ {(i)} = \\mathbf {W} _ {\\mu} f _ {\\mathrm {o b j}} \\left(\\mathbf {x} _ {\\mathrm {o b j}} ^ {(i)}\\right) + \\mathbf {b} _ {\\mu}, \\tag {6}", + "image_path": "0528ab08ad57246e929970d7b6f09d613094bd5453a33a0882d74fa3650d2915.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 249, + 384, + 365, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 384, + 365, + 399 + ], + "spans": [ + { + "bbox": [ + 249, + 384, + 365, + 399 + ], + "type": "interline_equation", + "content": "\\pmb {\\sigma} ^ {(i)} = \\mathbf {W} _ {\\sigma} f _ {\\mathrm {o b j}} (\\mathbf {x} _ {\\mathrm {o b j}} ^ {(i)}) + \\mathbf {b} _ {\\sigma},", + "image_path": "f47cd199456fff9fe9225ef88eca854fcb4685a032e36fd1e78d5fd42bc99ba0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{(\\cdot)}" + }, + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_{(\\cdot)}" + }, + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "text", + "content": " are learned weight and bias. Then, with the probabilistic modeling, we sample object-centric noise from the latent object space " + }, + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T \\sim p_{\\phi}(\\mathbf{x}_T | \\mathbf{x}_{\\mathrm{obj}})" + }, + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "text", + "content": " with reparameterization [24]: " + }, + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T^{(i)} = \\boldsymbol{\\mu}^{(i)} + \\boldsymbol{\\sigma}^{(i)} \\odot \\boldsymbol{\\epsilon}" + }, + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "text", + "content": " is the element-wise product and " + }, + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\epsilon} \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 130, + 403, + 482, + 453 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "type": "text", + "content": "To optimize the probabilistic space, we deploy Kullback-Leibler (KL) divergence " + }, + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{KL}} \\left(p_{\\phi}(\\mathbf{x}_T|\\mathbf{x}_{\\mathrm{obj}})||p(\\mathbf{x}_T)\\right)" + }, + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 130, + 453, + 482, + 548 + ], + "type": "text", + "content": " that encourages the Gaussian distributed [24], which has two advantages. First, the probabilistic nature of normal Gaussian removes the class bias within the object space [18] and promotes the generalization capacity towards discovering novel classes [45, 59]. Second, it naturally satisfies the DDPM constraint (Eq. 1) with the Gaussian formulation [21] for the following diffusion process. The implemented KL optimization objective, as proved in the Appendix, can be re-written in the following format," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 191, + 553, + 481, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 191, + 553, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 191, + 553, + 481, + 586 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {K L} = - \\frac {1}{2 N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {D} \\left(1 + \\log \\left(\\boldsymbol {\\sigma} _ {j} ^ {2 (i)}\\right) - \\boldsymbol {\\mu} _ {j} ^ {2 (i)} - \\boldsymbol {\\sigma} _ {j} ^ {2 (i)}\\right). \\tag {7}", + "image_path": "e0fb26df437bc9cde23be4a28b0b84f198b59a32a66e5042defaaa124721fa1f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 590, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 482, + 628 + ], + "type": "text", + "content": "VLS can establish a probabilistic space with differential property and sample object-centric noise " + }, + { + "bbox": [ + 130, + 590, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 130, + 590, + 482, + 628 + ], + "type": "text", + "content": " for the diffusion process, serving to generate probabilistic object-text pairs in the diffusion-based distribution transfer [48,51]." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 143, + 90, + 448, + 103 + ], + "type": "text", + "content": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 631, + 482, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 631, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 631, + 482, + 666 + ], + "type": "text", + "content": "4 Besides " + }, + { + "bbox": [ + 133, + 631, + 482, + 666 + ], + "type": "inline_equation", + "content": "N_{rpn}" + }, + { + "bbox": [ + 133, + 631, + 482, + 666 + ], + "type": "text", + "content": " RPN proposals, we follow [68, 81] to add additional proposals for training, including " + }, + { + "bbox": [ + 133, + 631, + 482, + 666 + ], + "type": "inline_equation", + "content": "N_{cs}" + }, + { + "bbox": [ + 133, + 631, + 482, + 666 + ], + "type": "text", + "content": " class-specific proposals (from " + }, + { + "bbox": [ + 133, + 631, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{aux}" + }, + { + "bbox": [ + 133, + 631, + 482, + 666 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 133, + 631, + 482, + 666 + ], + "type": "inline_equation", + "content": "N_{ca}" + }, + { + "bbox": [ + 133, + 631, + 482, + 666 + ], + "type": "text", + "content": " class-agnostic proposals (from " + }, + { + "bbox": [ + 133, + 631, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{det}" + }, + { + "bbox": [ + 133, + 631, + 482, + 666 + ], + "type": "text", + "content": ") from [58]. The number of proposals is " + }, + { + "bbox": [ + 133, + 631, + 482, + 666 + ], + "type": "inline_equation", + "content": "N = N_{rpn} + N_{cs} + N_{ca}" + }, + { + "bbox": [ + 133, + 631, + 482, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 301, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 301, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 301, + 128 + ], + "type": "text", + "content": "4.2 Continual Diffusion Module" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 183 + ], + "type": "text", + "content": "With object-centric noise " + }, + { + "bbox": [ + 130, + 136, + 482, + 183 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 130, + 136, + 482, + 183 + ], + "type": "text", + "content": ", we further deploy the object-to-image (o2i) and object-to-text (o2t) diffusion to achieve a continual distribution transfer for OVD. To achieve this, an efficient diffusion module " + }, + { + "bbox": [ + 130, + 136, + 482, + 183 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 130, + 136, + 482, + 183 + ], + "type": "text", + "content": " with MLP architecture is designed, the details of which are presented in the Appendix." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 184, + 483, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 184, + 483, + 256 + ], + "spans": [ + { + "bbox": [ + 130, + 184, + 483, + 256 + ], + "type": "text", + "content": "Object-to-image Diffusion. Given class-agnostic region proposals, we send them to the object detector to generate the object embedding " + }, + { + "bbox": [ + 130, + 184, + 483, + 256 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{obj}}" + }, + { + "bbox": [ + 130, + 184, + 483, + 256 + ], + "type": "text", + "content": " and feed them into frozen CLIP image encoder [67] to obtain static image embedding " + }, + { + "bbox": [ + 130, + 184, + 483, + 256 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{\\mathrm{img}}" + }, + { + "bbox": [ + 130, + 184, + 483, + 256 + ], + "type": "text", + "content": ". Then, with a randomly sampled object-centric noise " + }, + { + "bbox": [ + 130, + 184, + 483, + 256 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T \\sim p_\\phi(\\mathbf{x}_T | \\mathbf{x}_{\\mathrm{obj}})" + }, + { + "bbox": [ + 130, + 184, + 483, + 256 + ], + "type": "text", + "content": " from VLS, we encourage the diffusion module to generate the image embedding " + }, + { + "bbox": [ + 130, + 184, + 483, + 256 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{img}}" + }, + { + "bbox": [ + 130, + 184, + 483, + 256 + ], + "type": "text", + "content": " with object-level condition " + }, + { + "bbox": [ + 130, + 184, + 483, + 256 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{o2i} = \\mathbf{x}_{\\mathrm{obj}}" + }, + { + "bbox": [ + 130, + 184, + 483, + 256 + ], + "type": "text", + "content": " using a " + }, + { + "bbox": [ + 130, + 184, + 483, + 256 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 184, + 483, + 256 + ], + "type": "text", + "content": "-step reverse diffusion:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 247, + 266, + 482, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 266, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 247, + 266, + 482, + 277 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {\\mathrm {i m g}} = \\operatorname {R e v D i f} \\left(\\mathbf {x} _ {T}, \\mathbf {c} _ {o 2 i}, L\\right), \\tag {8}", + "image_path": "87388332cff774eebec25fc914083b7d2b3e1d0c804c9bfc08030c1fd085e66f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 286, + 487, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 286, + 487, + 348 + ], + "spans": [ + { + "bbox": [ + 131, + 286, + 487, + 348 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 286, + 487, + 348 + ], + "type": "inline_equation", + "content": "\\mathrm{RevDif}(\\cdot)" + }, + { + "bbox": [ + 131, + 286, + 487, + 348 + ], + "type": "text", + "content": " (Algorithm 1) recovers image embedding from object-centric noise " + }, + { + "bbox": [ + 131, + 286, + 487, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 131, + 286, + 487, + 348 + ], + "type": "text", + "content": ". To this end, the object-to-image diffusion, denoted as " + }, + { + "bbox": [ + 131, + 286, + 487, + 348 + ], + "type": "inline_equation", + "content": "p_{\\theta,\\phi}(\\mathbf{x}_{\\mathrm{img}}|\\mathbf{c}_{o2i},\\mathbf{x}_{\\mathrm{obj}})" + }, + { + "bbox": [ + 131, + 286, + 487, + 348 + ], + "type": "text", + "content": ", uses object embedding as the diffusion condition " + }, + { + "bbox": [ + 131, + 286, + 487, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{o2i} = \\mathbf{x}_{\\mathrm{obj}}" + }, + { + "bbox": [ + 131, + 286, + 487, + 348 + ], + "type": "text", + "content": ", and is optimized by minimizing the distance between the generated embedding " + }, + { + "bbox": [ + 131, + 286, + 487, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{img}}" + }, + { + "bbox": [ + 131, + 286, + 487, + 348 + ], + "type": "text", + "content": " and CLIP image embedding " + }, + { + "bbox": [ + 131, + 286, + 487, + 348 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{\\mathrm{img}}" + }, + { + "bbox": [ + 131, + 286, + 487, + 348 + ], + "type": "text", + "content": ", which is achieved via the following loss function," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 238, + 357, + 482, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 357, + 482, + 388 + ], + "spans": [ + { + "bbox": [ + 238, + 357, + 482, + 388 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {o 2 i} = \\frac {1}{N _ {c a}} \\sum_ {i = 1} ^ {N _ {c a}} | | \\mathbf {x} _ {\\mathrm {i m g}} ^ {(i)} - \\hat {\\mathbf {x}} _ {\\mathrm {i m g}} ^ {(i)} | | _ {1}, \\tag {9}", + "image_path": "041504f214b2f4b1d8125301923708d6f5aa5e467b2c6b874aa96fa6fb501e11.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 396, + 482, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 432 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 396, + 482, + 432 + ], + "type": "inline_equation", + "content": "N_{ca}" + }, + { + "bbox": [ + 130, + 396, + 482, + 432 + ], + "type": "text", + "content": " is the number of class-agnostic proposals. With the optimized diffusion process, the object distribution could be transferred to the CLIP image counterpart, facilitating the adaptation to the CLIP latent space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "text", + "content": "Object-to-text Diffusion. With diffusion-based formulation, we further encourage the model to reconstruct the associated text embedding " + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{txt}}" + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "text", + "content": " for region proposals " + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{obj}}" + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "text", + "content": ", which conducts a generative procedure in latent space. Note that this generative objective is different from existing OVD works that directly optimize the discriminative class separation. Specifically, we design the object-to-text diffusion with a " + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "text", + "content": "-step diffusion " + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "inline_equation", + "content": "p_{\\theta,\\phi}(\\mathbf{x}_{\\mathrm{txt}}|\\mathbf{c}_{o2t},\\mathbf{x}_{\\mathrm{obj}})" + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "text", + "content": " guided by condition " + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{o2t}" + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "text", + "content": ". Considering the dominant role of condition [21] in the generated content, we propose a simple yet effective attentive mechanism to obtain the refined condition " + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{o2t} = \\bar{\\mathbf{x}}_{\\mathrm{obj}}^{\\mathrm{img}}" + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "text", + "content": " with the generated image image-level context:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 231, + 551, + 482, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 551, + 482, + 566 + ], + "spans": [ + { + "bbox": [ + 231, + 551, + 482, + 566 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {\\mathrm {o b j}} ^ {\\mathrm {i m g}} = \\operatorname {S i g m o i d} \\left(\\operatorname {F C} \\left(\\mathbf {x} _ {\\mathrm {i m g}}\\right)\\right) \\odot \\mathbf {x} _ {\\mathrm {o b j}}, \\tag {10}", + "image_path": "bdae5a607edb152e8a4c383e5bb623c01954180136e490609eade700f8036d6a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 573, + 482, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 573, + 482, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 573, + 482, + 611 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 573, + 482, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{img}}" + }, + { + "bbox": [ + 130, + 573, + 482, + 611 + ], + "type": "text", + "content": " is generated via Eq. 8, " + }, + { + "bbox": [ + 130, + 573, + 482, + 611 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 573, + 482, + 611 + ], + "type": "text", + "content": " is the element-wise product, and FC is a liner projection. Then, we use the refined condition " + }, + { + "bbox": [ + 130, + 573, + 482, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{o2t} = \\mathbf{x}_{\\mathrm{obj}}^{\\mathrm{img}}" + }, + { + "bbox": [ + 130, + 573, + 482, + 611 + ], + "type": "text", + "content": " to guide the text embedding generation for " + }, + { + "bbox": [ + 130, + 573, + 482, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{txt}}" + }, + { + "bbox": [ + 130, + 573, + 482, + 611 + ], + "type": "text", + "content": " with a " + }, + { + "bbox": [ + 130, + 573, + 482, + 611 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 573, + 482, + 611 + ], + "type": "text", + "content": "-step reverse diffusion:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 248, + 621, + 482, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 621, + 482, + 633 + ], + "spans": [ + { + "bbox": [ + 248, + 621, + 482, + 633 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {\\mathrm {t x t}} = \\operatorname {R e v D i f} \\left(\\mathbf {x} _ {T}, \\mathbf {c} _ {o 2 t}, T\\right). \\tag {11}", + "image_path": "8e8445c86313ceb9151be8db6bf978b5da6e63774cb49277f8912047e88c3b12.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Considering the mutual benefits of generative and discriminative objectives for model learning as explored in [45], the loss function for the object-to-text" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 200, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 200, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 200, + 101 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 140 + ], + "type": "text", + "content": "diffusion process is designed with these two distinct components. To this end, the optimization objective can be expressed as follows," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 152, + 481, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 152, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 140, + 152, + 481, + 209 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {o 2 t} = \\underbrace {\\frac {1}{N _ {l}} \\sum_ {i} ^ {N _ {l}} | | \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\mathbf {x} _ {t} ^ {(i)} , \\mathbf {c} _ {o 2 t} ^ {(i)} , t\\right) - \\boldsymbol {\\epsilon} | | _ {1}} _ {\\text {G e n e r a t i v e}} + \\underbrace {\\mathcal {L} _ {C E} \\left(\\operatorname {S i g m o i d} \\left(\\frac {1}{\\tau} \\cos \\left(\\mathbf {x} _ {\\mathrm {t x t}}, \\hat {\\mathbf {x}} _ {\\mathrm {t x t}}\\right)\\right), \\hat {y}\\right)} _ {\\text {D i s c r i m i n a t i v e}} \\tag {12}", + "image_path": "dea80f79f2ef17eb90b9ddba1015c3ab8c85cfd355a46a052993413dab273d5b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 210, + 482, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 210, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 210, + 482, + 293 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 210, + 482, + 293 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 130, + 210, + 482, + 293 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 210, + 482, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 130, + 210, + 482, + 293 + ], + "type": "text", + "content": " is the noisy CLIP text embedding at a randomly sampled time-step " + }, + { + "bbox": [ + 130, + 210, + 482, + 293 + ], + "type": "inline_equation", + "content": "t \\sim \\mathcal{U}[1,T]" + }, + { + "bbox": [ + 130, + 210, + 482, + 293 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 210, + 482, + 293 + ], + "type": "inline_equation", + "content": "\\tau = 0.02" + }, + { + "bbox": [ + 130, + 210, + 482, + 293 + ], + "type": "text", + "content": " is a scaling factor as [15, 68], " + }, + { + "bbox": [ + 130, + 210, + 482, + 293 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 130, + 210, + 482, + 293 + ], + "type": "text", + "content": " is the class label, " + }, + { + "bbox": [ + 130, + 210, + 482, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 130, + 210, + 482, + 293 + ], + "type": "text", + "content": " is the Cross-Entropy loss, and " + }, + { + "bbox": [ + 130, + 210, + 482, + 293 + ], + "type": "inline_equation", + "content": "N_{l} = N_{cs} + N_{rpn}" + }, + { + "bbox": [ + 130, + 210, + 482, + 293 + ], + "type": "text", + "content": " counts the labeled proposals. The first term optimizes the diffusion module " + }, + { + "bbox": [ + 130, + 210, + 482, + 293 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 130, + 210, + 482, + 293 + ], + "type": "text", + "content": " to reconstruct the text embedding in a generative manner [21], serving for the objective of Eq. 3. The second term can give an interpretable score for NMS [69] in detectors and ensure the generated embedding with sufficient discriminability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 295, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 482, + 392 + ], + "type": "text", + "content": "Remark. We use a shared diffusion module with a " + }, + { + "bbox": [ + 130, + 295, + 482, + 392 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 295, + 482, + 392 + ], + "type": "text", + "content": "-step diffusion formulation to optimize two processes, where the first " + }, + { + "bbox": [ + 130, + 295, + 482, + 392 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 295, + 482, + 392 + ], + "type": "text", + "content": " steps are assigned for object-to-image, and the whole " + }, + { + "bbox": [ + 130, + 295, + 482, + 392 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 295, + 482, + 392 + ], + "type": "text", + "content": " steps are for object-to-text diffusion. Note that even though we do not explicitly optimize image-to-text diffusion, this transition can be intrinsically achieved between the " + }, + { + "bbox": [ + 130, + 295, + 482, + 392 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 295, + 482, + 392 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 130, + 295, + 482, + 392 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 295, + 482, + 392 + ], + "type": "text", + "content": "-th steps, as illustrated by the gray-colored part in Fig. 1(b). The reason is based on the continual property of the Markov chain in DDPM, i.e., formulating a random process that undergoes transitions from one state to another within a finite number of possible states [21, 63]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 416, + 276, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 416, + 276, + 428 + ], + "spans": [ + { + "bbox": [ + 132, + 416, + 276, + 428 + ], + "type": "text", + "content": "4.3 Training and Inference" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 444, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 444, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 131, + 444, + 482, + 469 + ], + "type": "text", + "content": "Training Stage. We implement the overall optimization objective on the basic object detector, Faster RCNN [69], to train the whole framework:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 194, + 481, + 482, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 481, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 194, + 481, + 482, + 495 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C L I F F} = \\mathcal {L} _ {r p n} + \\mathcal {L} _ {r e g} + \\alpha \\mathcal {L} _ {K L} + \\beta_ {1} \\mathcal {L} _ {o 2 i} + \\beta_ {2} \\mathcal {L} _ {o 2 t}, \\tag {13}", + "image_path": "56a38fcbdd2de0384d67c7cd8e5ba15399a6445d404fd90bf11c64d64c64257f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 506, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 506, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 131, + 506, + 482, + 529 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 506, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rpn}" + }, + { + "bbox": [ + 131, + 506, + 482, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 506, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 131, + 506, + 482, + 529 + ], + "type": "text", + "content": " are RPN loss and the bounding box regression loss in the RCNN head [69] borrowed from object detectors. " + }, + { + "bbox": [ + 131, + 506, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta_1, \\beta_2" + }, + { + "bbox": [ + 131, + 506, + 482, + 529 + ], + "type": "text", + "content": " are the loss weight." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 531, + 483, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 531, + 483, + 568 + ], + "spans": [ + { + "bbox": [ + 131, + 531, + 483, + 568 + ], + "type": "text", + "content": "Inference Stage. We first extract the region proposals " + }, + { + "bbox": [ + 131, + 531, + 483, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{obj}}" + }, + { + "bbox": [ + 131, + 531, + 483, + 568 + ], + "type": "text", + "content": " and then conduct reverse diffusion to generate the associated text embedding " + }, + { + "bbox": [ + 131, + 531, + 483, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{txt}}" + }, + { + "bbox": [ + 131, + 531, + 483, + 568 + ], + "type": "text", + "content": " for region proposals, whose similarity with " + }, + { + "bbox": [ + 131, + 531, + 483, + 568 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{\\mathrm{txt}}" + }, + { + "bbox": [ + 131, + 531, + 483, + 568 + ], + "type": "text", + "content": " is measured for class prediction " + }, + { + "bbox": [ + 131, + 531, + 483, + 568 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 131, + 531, + 483, + 568 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 235, + 579, + 482, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 579, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 235, + 579, + 482, + 605 + ], + "type": "interline_equation", + "content": "\\tilde {y} = \\operatorname {S i g m o i d} \\left(\\frac {1}{\\tau} c o s \\left(\\mathbf {x} _ {\\mathrm {t x t}}, \\hat {\\mathbf {x}} _ {\\mathrm {t x t}}\\right)\\right), \\tag {14}", + "image_path": "99aec0de66802656ad7bd748087af0d8de63240c165472dd47c2c3923c0e1a70.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 616, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 616, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 131, + 616, + 482, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 616, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 131, + 616, + 482, + 640 + ], + "type": "text", + "content": " is the per-class confidence, " + }, + { + "bbox": [ + 131, + 616, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{txt}}" + }, + { + "bbox": [ + 131, + 616, + 482, + 640 + ], + "type": "text", + "content": " is the generated text embedding via object-to-text diffusion (Eq. 11), and " + }, + { + "bbox": [ + 131, + 616, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{\\mathrm{txt}}" + }, + { + "bbox": [ + 131, + 616, + 482, + 640 + ], + "type": "text", + "content": " is CLIP text embedding for all classes." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 143, + 90, + 448, + 103 + ], + "type": "text", + "content": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 652, + 347, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 652, + 347, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 652, + 347, + 666 + ], + "type": "text", + "content": "5 Following Eq. 1, " + }, + { + "bbox": [ + 133, + 652, + 347, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t = \\sqrt{\\bar{\\alpha}_t}\\hat{\\mathbf{x}}_{\\mathrm{xt}} + (1 - \\bar{\\alpha}_t)\\epsilon" + }, + { + "bbox": [ + 133, + 652, + 347, + 666 + ], + "type": "text", + "content": " as [21]" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 158, + 474, + 306 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 1: Experimental results (%) on COCO benchmark. † indicates the reproduced results with the consistent baseline setting as [68], including ImageNet pre-training and CLIP prompting. ‡ indicates using the longer training schedule [68]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 158, + 474, + 306 + ], + "lines": [ + { + "bbox": [ + 140, + 158, + 474, + 306 + ], + "spans": [ + { + "bbox": [ + 140, + 158, + 474, + 306 + ], + "type": "table", + "html": "
MethodObject DetectorSub-space Alignment\\( \\mathrm{mAP}_{\\mathrm{n}} \\)\\( \\mathrm{mAP}_{\\mathrm{b}} \\)mAPReference
OVR-CNN [82]Faster RCNN C4Obj → Txt22.846.039.9CVPR-2021
RegionCLIP [85]Faster RCNN C4Obj → Txt & Img26.854.847.5CVPR-2022
ViLD [15]Faster RCNN C4Obj → Txt & Img27.659.551.3ICLR-2022
OV-DETR [81]Deformable DETRObj → Txt & Img29.461.052.7ECCV-2022
Detic [86]Faster RCNN C4Obj → Txt27.847.145.0ECCV-2022
Detic‡ [86]Faster RCNN C4Obj → Txt28.453.847.2ECCV-2022
PB-OVD [14]Faster RCNN C4Obj → Txt30.846.142.1ECCV-2022
MAVL-OVD [58]Faster RCNN C4Obj → Txt30.452.646.8ECCV-2022
\\( OCD^‡ \\)[68]Faster RCNN C4Obj → Txt & Img36.654.049.4NeurIPS-2022
F-VLM [26]Faster RCNN FPNObj → Txt & Img28.0-39.6ICLR-2023
VLDet [46]Faster RCNN C4Obj → Txt32.050.645.8ICLR-2023
BARON†,‡ [75]Faster RCNN C4Obj → Txt & Img41.053.350.1CVPR-2023
EdaDet [71]Faster RCNN C4Obj → Txt37.857.752.5ICCV-2023
GOAT [72]Faster RCNN C4Obj → Txt & Img36.453.048.6ICCV-2023
CLIFF (ours)Faster RCNN C4Obj → Image → Txt41.354.150.6-
CLIFF‡ (ours)Faster RCNN C4Obj → Image → Txt43.254.051.2-
", + "image_path": "b8b82f1567a1cc447292effb74fad6a3abcac1a68de2da1c3a3db7dd8b1b88f2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 327, + 230, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 230, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 230, + 341 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 353, + 326, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 326, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 326, + 365 + ], + "type": "text", + "content": "5.1 Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 373, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 445 + ], + "type": "text", + "content": "We follow existing works [68, 75] to compare with state-of-the-art methods. COCO Setting. The COCO-2017 dataset [47] is used for training and evaluation, containing 80 object-level classes. As for the auxiliary dataset " + }, + { + "bbox": [ + 130, + 373, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{aux}" + }, + { + "bbox": [ + 130, + 373, + 482, + 445 + ], + "type": "text", + "content": ", we follow [82] to use COCO caption dataset in " + }, + { + "bbox": [ + 130, + 373, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_B \\cup \\mathcal{C}_N" + }, + { + "bbox": [ + 130, + 373, + 482, + 445 + ], + "type": "text", + "content": ". The core evaluation metric is the novel-class Average Precision (AP) with an IoU threshold of 0.5, denoted as " + }, + { + "bbox": [ + 130, + 373, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP_n}" + }, + { + "bbox": [ + 130, + 373, + 482, + 445 + ], + "type": "text", + "content": ", for the novel-class generalization. The results are shown in Table 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 446, + 482, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 493 + ], + "type": "text", + "content": "LVIS Setting. LVIS v1.0 [17] contains 1203 classes and is split into frequent, common, and rare types. For the auxiliary datasets " + }, + { + "bbox": [ + 130, + 446, + 482, + 493 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{aux}" + }, + { + "bbox": [ + 130, + 446, + 482, + 493 + ], + "type": "text", + "content": ", we follow [68] to use the subset of ImageNet-21K. The core evaluation metric is mask AP of rare classes, denoted as " + }, + { + "bbox": [ + 130, + 446, + 482, + 493 + ], + "type": "inline_equation", + "content": "\\mathrm{AP_r^m}" + }, + { + "bbox": [ + 130, + 446, + 482, + 493 + ], + "type": "text", + "content": ". We report the comparison results in Table 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 494, + 482, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 482, + 530 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 482, + 530 + ], + "type": "text", + "content": "Pascal VOC Setting. We follow [72] to evaluate the generalization to unseen datasets with Pascal VOC [10] and LVIS [17] using COCO-trained model. We use " + }, + { + "bbox": [ + 130, + 494, + 482, + 530 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 130, + 494, + 482, + 530 + ], + "type": "text", + "content": " to evaluate the performance as [72]. The results are given in Table 7." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 548, + 279, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 279, + 561 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 279, + 561 + ], + "type": "text", + "content": "5.2 Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Object Detector. Considering that previous works use inconsistent detectors, we strictly follow the setting of [68,82] in different benchmarks. Concretely, for the COCO setting, we use Faster-RCNN C4 [69] with the ResNet-50 backbone. The backbone is pre-trained on ImageNet [8] without bells and whistles. For the LVIS setting, we follow the two main streams of existing works by using Mask-RCNN-FPN [20] and CenterNet v2 [87] as the basic object detectors for a thorough comparison, which is based on ReseNet-50 backbone. For the experiments of all benchmarks, we use the SGD optimizer with a weight decay of " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " and a" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 171, + 146, + 443, + 281 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 2: Experimental results (\\%) on LVIS benchmark with the consistent object detector setting. Ens indicates using the ensemble trick [11]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 171, + 146, + 443, + 281 + ], + "lines": [ + { + "bbox": [ + 171, + 146, + 443, + 281 + ], + "spans": [ + { + "bbox": [ + 171, + 146, + 443, + 281 + ], + "type": "table", + "html": "
MethodObject DetectorAPmRAPmcAPmfAPReference
ViLD-Ens [15]Mask RCNN16.624.630.325.5ICLR-2022
MVAL-OVD [58]Mask RCNN17.021.226.122.4ECCV-2022
OCD [68]Mask RCNN21.125.029.125.9NeurIPS-2022
DetPro [9]Mask RCNN19.825.628.925.9CVPR-2022
RegionCLIP [85]Mask RCNN17.127.434.028.2CVPR-2022
F-VLM [26]Mask RCNN18.624.026.924.2ICLR-2023
BARON [75]Mask RCNN18.024.428.925.1CVPR-2023
BARON-Ens [75]Mask RCNN19.226.829.426.5CVPR-2023
CLIFF (ours)Mask RCNN22.426.929.826.8-
VLDet [46]CenterNet v221.729.834.330.1ICLR-2023
CoDet [57]CenterNet v223.430.034.630.7NeirIPS-2023
GOAT [72]CenterNet v223.329.734.330.4ICCV-2023
EdaDet [71]CenterNet v223.727.529.127.5ICCV-2023
CLIFF (ours)CenterNet v224.527.729.828.2-
", + "image_path": "ca8f4b76e4eb43f6cc73aee5c8bc490df2c2e3c3aaa3dd40bf8a1bac4b9c2b6b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 302, + 482, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 482, + 350 + ], + "type": "text", + "content": "momentum of 0.9 following [68]. We use 4 NVIDIA V100 for training with a batch size of 16. We pre-train the base object detector with a " + }, + { + "bbox": [ + 130, + 302, + 482, + 350 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 130, + 302, + 482, + 350 + ], + "type": "text", + "content": " schedule as [68], and then train the whole model with a " + }, + { + "bbox": [ + 130, + 302, + 482, + 350 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 130, + 302, + 482, + 350 + ], + "type": "text", + "content": " schedule. We also explore the longer " + }, + { + "bbox": [ + 130, + 302, + 482, + 350 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 130, + 302, + 482, + 350 + ], + "type": "text", + "content": " training schedule to make fair comparisons with specific methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 351, + 482, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 387 + ], + "type": "text", + "content": "CLIP Model. As the previous work [68], we use the single prompt, 'a photo of a {category}', to compute CLIP text embedding " + }, + { + "bbox": [ + 130, + 351, + 482, + 387 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{\\mathrm{txt}}" + }, + { + "bbox": [ + 130, + 351, + 482, + 387 + ], + "type": "text", + "content": ", and use " + }, + { + "bbox": [ + 130, + 351, + 482, + 387 + ], + "type": "inline_equation", + "content": "224\\times 224" + }, + { + "bbox": [ + 130, + 351, + 482, + 387 + ], + "type": "text", + "content": " input-scale for image embedding " + }, + { + "bbox": [ + 130, + 351, + 482, + 387 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{\\mathrm{img}}" + }, + { + "bbox": [ + 130, + 351, + 482, + 387 + ], + "type": "text", + "content": " via ViT-B/32 model [67]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "text", + "content": "CLIFF. The number of the time-step in object-to-text " + }, + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "inline_equation", + "content": "(T)" + }, + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "text", + "content": " and object-to-image " + }, + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "inline_equation", + "content": "(L)" + }, + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "text", + "content": " diffusion are set " + }, + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "inline_equation", + "content": "T = 10, L = 3" + }, + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "text", + "content": ", which are justified in Table 4. " + }, + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\beta_{1}" + }, + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\beta_{2}" + }, + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "text", + "content": " are empirically set to 2.0, 15.0, and 1.0, respectively, which are verified by the experiments in Appendix. " + }, + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 387, + 482, + 460 + ], + "type": "text", + "content": " normalization is deployed on CLIP embedding as [15,68]. Class-specific and RPN proposals are used to optimize object-to-text diffusion [68,75], while class-agnostic ones are for object-to-image diffusion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 477, + 380, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 477, + 380, + 490 + ], + "spans": [ + { + "bbox": [ + 131, + 477, + 380, + 490 + ], + "type": "text", + "content": "5.3 Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": "Results on COCO Benchmark. As shown in Table 1, we compare the proposed method with state-of-the-art counterparts on the COCO benchmark. CLIFF achieves " + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "inline_equation", + "content": "41.3\\% \\mathrm{mAP_n}" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": ", which outperforms OCD [68] (" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "inline_equation", + "content": "36.6\\% \\mathrm{mAP_n}" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": "), VLDet [46] (" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "inline_equation", + "content": "32.0\\% \\mathrm{mAP_n}" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": ") and BARON [75] (" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "inline_equation", + "content": "41.0\\% \\mathrm{mAP_n}" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": ") with " + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "inline_equation", + "content": "4.7\\%" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "inline_equation", + "content": "9.3\\%" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "inline_equation", + "content": "0.3\\% \\mathrm{mAP_n}" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": ", comprehensively verifying the effectiveness of the proposed CLIFF. The proposed method also surpasses EdaDet [71] and GOAT [72] by a noticeable margin. Besides, after using fair multi-scale training [68, 75], our method (" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "inline_equation", + "content": "43.2\\% \\mathrm{mAP_n}" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": ") outperforms existing works by a large margin. Unlike existing works aligning object-text and object-image separately, we formulate a continual diffusion process among the three sub-spaces in a different learning paradigm." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Results on LVIS Benchmark. We present a thorough comparison with corresponding state-of-the-art methods using Mask RCNN and CenterNet v2 detectors, which is shown in Table 2. For Mask RCNN, it can be observed that CLIFF achieves the best " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "22.4\\%" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", which surpasses state-of-the-art counterparts" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 143, + 90, + 447, + 102 + ], + "type": "text", + "content": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 136, + 477, + 205 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 478, + 125 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 478, + 125 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 478, + 125 + ], + "type": "text", + "content": "Table 3: Ablation study results on COCO and LVIS benchmarks with " + }, + { + "bbox": [ + 134, + 114, + 478, + 125 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 134, + 114, + 478, + 125 + ], + "type": "text", + "content": " schedule." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 136, + 477, + 205 + ], + "lines": [ + { + "bbox": [ + 137, + 136, + 477, + 205 + ], + "spans": [ + { + "bbox": [ + 137, + 136, + 477, + 205 + ], + "type": "table", + "html": "
CDM o2t o2i VLSVLSCOCOLVIS-BoxLVIS-Mask
mAPnmAPbmAPAPbAPcAPbAPbAPmAPmAPcAPb
1(baseline)32.249.945.315.820.126.221.415.219.725.020.5
2✓36.649.546.116.220.927.422.615.820.125.721.6
3✓39.552.449.216.721.727.723.116.520.925.922.1
4✓38.853.349.317.321.927.823.316.721.126.122.2
5✓40.052.949.518.922.128.023.818.121.426.222.7
", + "image_path": "49aff2b8c3083ee2e2ffeddf0479d98f9ace8f04889dcfa8f7c2551ceb7c9321.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 136, + 291, + 298, + 392 + ], + "blocks": [ + { + "bbox": [ + 130, + 215, + 302, + 281 + ], + "lines": [ + { + "bbox": [ + 130, + 215, + 302, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 302, + 281 + ], + "type": "text", + "content": "Table 4: Analysis on the diffusion step. o2i and o2t indicates the object-to-image and object-to-text diffusion process respectively. * indicates sharing the same diffusion procedure and optimizing o2i and o2t in a combinatorial style [15,68]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 291, + 298, + 392 + ], + "lines": [ + { + "bbox": [ + 136, + 291, + 298, + 392 + ], + "spans": [ + { + "bbox": [ + 136, + 291, + 298, + 392 + ], + "type": "table", + "html": "
o2io2t\\( \\mathrm{mAP}_{\\mathrm{n}} \\)\\( \\mathrm{mAP}_{\\mathrm{b}} \\)mAP
111040.253.450.0
231041.354.150.6
351041.353.750.3
471041.253.950.5
591041.453.550.4
63540.953.850.3
731041.354.150.6
832039.753.449.8
9101041.453.650.4
10*10*1039.952.849.5
", + "image_path": "929e0a7405a7a8b89078fea8488fb20f3c09c38bcf83345d8e821a6038747dc2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 313, + 295, + 481, + 388 + ], + "blocks": [ + { + "bbox": [ + 311, + 218, + 482, + 284 + ], + "lines": [ + { + "bbox": [ + 311, + 218, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 311, + 218, + 482, + 284 + ], + "type": "text", + "content": "Table 5: Analysis of the discriminative and generative terms in Eq. 12 by replacing the random noise (Rand) with object-centric noise (OC) in different time steps " + }, + { + "bbox": [ + 311, + 218, + 482, + 284 + ], + "type": "inline_equation", + "content": "(2\\times" + }, + { + "bbox": [ + 311, + 218, + 482, + 284 + ], + "type": "text", + "content": " schedule). rows 7 and 8 indicate removing individual terms." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 295, + 481, + 388 + ], + "lines": [ + { + "bbox": [ + 313, + 295, + 481, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 481, + 388 + ], + "type": "table", + "html": "
DiscriminativeGenerative\\( \\mathrm{mAP}_{\\mathrm{n}} \\)
TT-1: 0TT-1: 0
1RandRandRandRand39.2
2OCRandOCRand41.0
3RandOCRandOC40.4
4OCOCOCRand41.1
5OCRandOCOC41.3
6OCOCOCOC41.3
7w/o DiscriminativeOCOC38.2
8OCOCw/o Generative40.1
", + "image_path": "804981751749b03a58764a57e35568bd9da68e8170d16a7e472406d26ecde3e6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "content": "BARON [75] " + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "inline_equation", + "content": "(19.2\\% \\mathrm{AP_r^m})" + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "content": ", OCD [68] " + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "inline_equation", + "content": "(21.3\\% \\mathrm{AP_r^m})" + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "inline_equation", + "content": "3.1\\% \\mathrm{AP_r^m}" + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "inline_equation", + "content": "1.3\\% \\mathrm{AP_r^m}" + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "content": ", respectively. Besides, our method surpasses the baseline model MAVL-OVD [58] with a significant " + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "inline_equation", + "content": "5.4\\% \\mathrm{AP_r^m}" + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "content": ", verifying the effectiveness of our method. For the comparison with CenterNet v2, we can find that our method gives " + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "inline_equation", + "content": "24.5\\% \\mathrm{AP_r^m}" + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "content": " and also surpasses the latest published works noticeably, outperforming CoDet [57] and GOAT [72] with " + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 130, + 414, + 482, + 511 + ], + "type": "text", + "content": " gains. This can reveal that CLIFF is a generic method in terms of multi-modal alignment and can be deployed in different object detectors with great transferability and adaptability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 526, + 270, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 526, + 270, + 539 + ], + "spans": [ + { + "bbox": [ + 131, + 526, + 270, + 539 + ], + "type": "text", + "content": "5.4 Quantitative Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Ablation Study. The ablative results are shown in Table 3 with a " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " training schedule. The first line indicates the model trained with extra MAVL proposals [58, 68], which is a fair baseline model as [68, 75]. Compared with the baseline model " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "(32.2\\% \\mathrm{mAP}_{\\mathrm{n}}, 15.8\\% \\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{b}}, 15.2\\% \\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{m}})" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ", introducing the object-to-text (row 2) diffusion gives satisfactory improvements with " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "36.6\\% \\mathrm{mAP}_{\\mathrm{n}}, 16.2\\% \\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{b}}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "15.8\\% \\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ". After integrating object-to-image diffusion (row 3), i.e., the full CDM, we can observe a further improvement with " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "7.3\\% \\mathrm{mAP}_{\\mathrm{n}}, 0.9\\% \\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{b}}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "1.3\\% \\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " gains, compared with the baseline. This verifies the superior effectiveness in modeling a continual distribution transfer than the existing discriminative paradigm. Besides, after replacing the randomly sampled noise by" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 158, + 299, + 220 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 302, + 149 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 302, + 149 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 302, + 149 + ], + "type": "text", + "content": "Table 6: Analysis of the refined condition " + }, + { + "bbox": [ + 131, + 114, + 302, + 149 + ], + "type": "inline_equation", + "content": "{\\mathbf{c}}_{o2t}" + }, + { + "bbox": [ + 131, + 114, + 302, + 149 + ], + "type": "text", + "content": " in object-to-image diffusion (Eq. 5) with " + }, + { + "bbox": [ + 131, + 114, + 302, + 149 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 131, + 114, + 302, + 149 + ], + "type": "text", + "content": " training schedule." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 158, + 299, + 220 + ], + "lines": [ + { + "bbox": [ + 135, + 158, + 299, + 220 + ], + "spans": [ + { + "bbox": [ + 135, + 158, + 299, + 220 + ], + "type": "table", + "html": "
Setting of c02tmAPnmAPbmAP
1Object Embed: xobj40.453.449.9
2Image Embed: ximg39.652.949.4
3Additive: xobj + ximg40.853.550.2
4Concated: [xobj : ximg]41.053.650.4
5Attentive: Eq. 1041.354.150.6
", + "image_path": "7e4ed113fb356aee53bcd1898aeaa96c14386ab1ed9506dcbe856a9af8b980c4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 320, + 158, + 474, + 220 + ], + "blocks": [ + { + "bbox": [ + 312, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 312, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 312, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 7: Cross-dataset evaluation with " + }, + { + "bbox": [ + 312, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 312, + 114, + 482, + 148 + ], + "type": "text", + "content": " (\\%) transferring COCO-trained models to the PASCAL VOC and LVIS." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 158, + 474, + 220 + ], + "lines": [ + { + "bbox": [ + 320, + 158, + 474, + 220 + ], + "spans": [ + { + "bbox": [ + 320, + 158, + 474, + 220 + ], + "type": "table", + "html": "
MethodPASCALLVIS
OVR-CNN [82]52.95.2
PB-OVD [14]59.28.0
VLDet [46]61.710.0
GOAT [72]63.614.0
CLIFF (Ours)66.116.7
", + "image_path": "7bd8dd669fe3b043c403169de7d0285a02f526cc98974a6c903b1df84edc006b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 243, + 481, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 481, + 280 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 481, + 280 + ], + "type": "text", + "content": "object-centric noise with VLS (row 5), it can be observed that the CDM (row 3) can be improved with " + }, + { + "bbox": [ + 130, + 243, + 481, + 280 + ], + "type": "inline_equation", + "content": "40.0\\% \\mathrm{mAP}_{\\mathrm{n}}" + }, + { + "bbox": [ + 130, + 243, + 481, + 280 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 243, + 481, + 280 + ], + "type": "inline_equation", + "content": "18.9\\% \\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{b}}" + }, + { + "bbox": [ + 130, + 243, + 481, + 280 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 243, + 481, + 280 + ], + "type": "inline_equation", + "content": "18.1\\% \\mathrm{AP}_{\\mathrm{r}}^{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 243, + 481, + 280 + ], + "type": "text", + "content": ", verifying the effectiveness of each component in the proposed CLIFF." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 281, + 482, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 424 + ], + "type": "text", + "content": "The Number of Diffusion Steps. We study different diffusion steps for o2i and o2t in Table 4, and have the following three observations. First, in rows 1-5, we observe that the results are stable when " + }, + { + "bbox": [ + 130, + 281, + 482, + 424 + ], + "type": "inline_equation", + "content": "L > 3" + }, + { + "bbox": [ + 130, + 281, + 482, + 424 + ], + "type": "text", + "content": " (around " + }, + { + "bbox": [ + 130, + 281, + 482, + 424 + ], + "type": "inline_equation", + "content": "41.3\\% \\mathrm{mAP_n}" + }, + { + "bbox": [ + 130, + 281, + 482, + 424 + ], + "type": "text", + "content": "), but give notable performance drops when the step is insufficient " + }, + { + "bbox": [ + 130, + 281, + 482, + 424 + ], + "type": "inline_equation", + "content": "L = 1" + }, + { + "bbox": [ + 130, + 281, + 482, + 424 + ], + "type": "text", + "content": ". Considering more computation requirements when enlarging the diffusion step, we select " + }, + { + "bbox": [ + 130, + 281, + 482, + 424 + ], + "type": "inline_equation", + "content": "L = 3" + }, + { + "bbox": [ + 130, + 281, + 482, + 424 + ], + "type": "text", + "content": " with the best trade-off. Second, in rows 6-8, setting too small and large o2t steps both give significant performance drops, which may be caused by unsatisfactory diffusion optimization. Third, we set consistent diffusion steps and compare the proposed continual diffusion (row 9) with existing combinatorial style, i.e., sharing the diffusion process and generated embedding between the o2i and o2t (row 10). With a " + }, + { + "bbox": [ + 130, + 281, + 482, + 424 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 130, + 281, + 482, + 424 + ], + "type": "text", + "content": " performance gain over the combinatorial objective, the effectiveness of our continual transfer can be justified." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": "Exploring Object-centric Noise. We further give a detailed analysis of the object-centric (OC) noise by replacing the random noise for the discriminative and generative terms in object-to-text objective (the core process for OVD) in different time steps, as shown in Table 5. Compared with the random-noise baseline with " + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "inline_equation", + "content": "39.2\\% \\mathrm{mAP}_{\\mathrm{n}}" + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": ", controlling the initial stage (row 2) and middle stages (row 3) gives " + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "inline_equation", + "content": "41.0\\%" + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "inline_equation", + "content": "40.4\\% \\mathrm{mAP}_{\\mathrm{n}}" + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": ", respectively. Introducing more OC noise in the diffusion (rows 4-6) gives stable performance improvements, indicating the effectiveness of the VLS. Further, we conduct an ablative study on o2t diffusion (rows 7-8), i.e., the first and second terms in Eq. 12, verifying their critical role in optimizing the diffusion process with both aspects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "text", + "content": "Analyzing the Condition Design. The strategy of refining the object-to-text (Eq. 5) condition " + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{o2t} = \\mathbf{x}_{\\mathrm{obj}}^{\\mathrm{img}}" + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "text", + "content": " is analyzed in Table 6. Compared with individually using the object and image (rows 1-2) condition with " + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "inline_equation", + "content": "40.4\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "inline_equation", + "content": "39.6\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP_n}" + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "text", + "content": ", combining the two aspects for the diffusion condition generally works better (rows 3-5), e.g., addition with " + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "inline_equation", + "content": "40.8\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP_n}" + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "text", + "content": ", concatenation with " + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "inline_equation", + "content": "41.0\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "text", + "content": ". Besides, we observe that the proposed attentive refinement on the condition (Eq. 10) achieves optimal performance and gives the best " + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "inline_equation", + "content": "41.3\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP_n}" + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "text", + "content": ", due to the adaptive fusion of image-level context. This can verify the effectiveness of our design." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "Generalization to Unseen Datasets. We further conduct the cross-dataset comparison following [72] by considering two different settings. Compared with" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 143, + 90, + 447, + 103 + ], + "type": "text", + "content": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 114, + 477, + 261 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 477, + 261 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 477, + 261 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 477, + 261 + ], + "type": "image", + "image_path": "2cdba09ea8db7d68a01774b1fca21647f413b2a958854e216770d5fef9d3f8d9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 277, + 482, + 313 + ], + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 313 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 313 + ], + "type": "text", + "content": "Fig. 3: Qualitative comparison among (a) the baseline [68], (b) our CLIFF, and (c) ground-truth on the COCO benchmark. We only visualize the novel-class predictions for a clearer evaluation of open-vocabulary generalization." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 332, + 482, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 381 + ], + "type": "text", + "content": "state-of-the-art counterpart GOAT [72], the proposed CLIFF gives " + }, + { + "bbox": [ + 130, + 332, + 482, + 381 + ], + "type": "inline_equation", + "content": "66.1\\%" + }, + { + "bbox": [ + 130, + 332, + 482, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 332, + 482, + 381 + ], + "type": "inline_equation", + "content": "16.7\\% \\mathrm{AP}_{50}" + }, + { + "bbox": [ + 130, + 332, + 482, + 381 + ], + "type": "text", + "content": " on the PASCAL VOC and LVIS datasets using the COCO-trained model, which surpasses GOAT with significant " + }, + { + "bbox": [ + 130, + 332, + 482, + 381 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 130, + 332, + 482, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 332, + 482, + 381 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 130, + 332, + 482, + 381 + ], + "type": "text", + "content": " gains, showing the superior generalization capacity of our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 395, + 263, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 395, + 263, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 395, + 263, + 407 + ], + "type": "text", + "content": "5.5 Qualitative Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 412, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 533 + ], + "type": "text", + "content": "As shown in Fig. 3, we visualize the novel-class prediction and conduct a qualitative comparison among (a) the baseline [68,75] trained with MAVL proposals [68], (b) the proposed CLIFF, and (c) ground-truth. It can be observed that the proposed method gives more accurate novel-class predictions, especially for occlusion situations and complex scenarios. Our method reduces the confusion among novel classes, e.g., correcting the wrong prediction (couch) given by the baseline in col. 1. Besides, the proposed CLIFF can relieve some missing errors, as illustrated in col. 2-5, revealing its superior novel-class generalization capacity. Hence, the proposed CLIFF is able to relieve the bias on the base classes and encourage the model to discover the novel-class objects with probabilistic modeling." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 547, + 220, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 547, + 220, + 560 + ], + "spans": [ + { + "bbox": [ + 131, + 547, + 220, + 560 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "This paper proposes a novel diffusion-based framework for OVD, coined CLIFF, which breaks through the bottlenecks of the combinatorial and discriminative limitation in existing OVD pipelines. The proposed CLIFF consists of a Variational Latent Sampler (VLS) to sample object-centric noise in the probabilistic object space, and leverages a Continual diffusion Module (CDM) to formulate a continual distribution transfer among three sub-spaces (object, image, and text) with cross-modal nature. Extensive experiments validate that the proposed method significantly outperforms existing approaches on large-scale benchmarks." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 153 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by Innovation and Technology Commission- Innovation and Technology Fund ITS/229/22 and Hong Kong Research Grants Council (RGC) General Research Fund 11211221." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 171, + 197, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 197, + 185 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 197, + 185 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 198, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 198, + 481, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 198, + 481, + 243 + ], + "spans": [ + { + "bbox": [ + 138, + 198, + 481, + 243 + ], + "type": "text", + "content": "1. Ali, S., Ghatwary, N., Jha, D., Isik-Polat, E., Polat, G., Yang, C., Li, W., Galdran, A., Ballester, M.A.G., Thambawita, V., et al.: Assessing generalisability of deep learning-based polyp detection and segmentation methods through a computer vision challenge. Sci. Rep. (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 243, + 481, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 481, + 266 + ], + "type": "text", + "content": "2. Amit, T., Nachmani, E., Shaharbany, T., Wolf, L.: Segdiff: Image segmentation with diffusion probabilistic models. ArXiv:2112.00390 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 266, + 481, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 481, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 481, + 298 + ], + "type": "text", + "content": "3. Bansal, A., Borgnia, E., Chu, H.M., Li, J.S., Kazemi, H., Huang, F., Goldblum, M., Geiping, J., Goldstein, T.: Cold diffusion: Inverting arbitrary image transforms without noise. ArXiv:2208.09392 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 300, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 481, + 320 + ], + "type": "text", + "content": "4. Baranchuk, D., Voynov, A., Rubachev, I., Khrulkov, V., Babenko, A.: Label-efficient semantic segmentation with diffusion models. In: ICLR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 322, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 322, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 322, + 481, + 342 + ], + "type": "text", + "content": "5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV. pp. 213-229 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 344, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 344, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 344, + 481, + 365 + ], + "type": "text", + "content": "6. Chen, S., Sun, P., Song, Y., Luo, P.: Diffusiondet: Diffusion model for object detection. ArXiv:2211.09788 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 366, + 481, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 366, + 481, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 366, + 481, + 388 + ], + "type": "text", + "content": "7. Chen, Z., Li, W., Xing, X., Yuan, Y.: Medical federated learning with joint graph purification for noisy label learning. MedIA (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 388, + 481, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 481, + 410 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 481, + 410 + ], + "type": "text", + "content": "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 411, + 481, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 481, + 442 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 481, + 442 + ], + "type": "text", + "content": "9. Du, Y., Wei, F., Zhang, Z., Shi, M., Gao, Y., Li, G.: Learning to prompt for open-vocabulary object detection with vision-language model. ArXiv:2203.14940 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 444, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 444, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 444, + 481, + 464 + ], + "type": "text", + "content": "0. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. IJCV 88, 303-338 (2010)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 466, + 481, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 466, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 466, + 481, + 487 + ], + "type": "text", + "content": "1. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Expand your detector vocabulary with uncurated images. ArXiv:2203.16513 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 488, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 488, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 488, + 481, + 520 + ], + "type": "text", + "content": "2. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Towards open-vocabulary detection using uncurated images. In: ECCV. pp. 701-717. Springer (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 521, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 481, + 553 + ], + "type": "text", + "content": "3. Gao, M., Xing, C., Niebles, J.C., Li, J., Xu, R., Liu, W., Xiong, C.: Towards open vocabulary object detection without human-provided bounding boxes. ArXiv:2111.09452 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 555, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 576 + ], + "type": "text", + "content": "4. Gao, M., Xing, C., Niebles, J.C., Li, J., Xu, R., Liu, W., Xiong, C.: Open vocabulary object detection with pseudo bounding-box labels. In: ECCV (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 577, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 481, + 598 + ], + "type": "text", + "content": "5. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: ICLR (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 599, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 481, + 620 + ], + "type": "text", + "content": "6. Gu, Z., Chen, H., Xu, Z., Lan, J., Meng, C., Wang, W.: Diffusioninst: Diffusion model for instance segmentation. ArXiv:2212.02773 (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 621, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 481, + 643 + ], + "type": "text", + "content": "7. Gupta, A., Dollar, P., Girshick, R.: LVIS: A dataset for large vocabulary instance segmentation. In: CVPR (2019)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "8. Han, J., Ren, Y., Ding, J., Yan, K., Xia, G.S.: Few-shot object detection via variational feature aggregation. ArXiv:2301.13411 (2023)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 143, + 90, + 447, + 102 + ], + "type": "text", + "content": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 138 + ], + "type": "text", + "content": "19. Han, X., Zheng, H., Zhou, M.: Card: Classification and regression diffusion models. In: NeurIPS (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 467, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 467, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 467, + 149 + ], + "type": "text", + "content": "20. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask R-CNN. In: CVPR (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "type": "text", + "content": "21. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "text", + "content": "22. Kim, D., Angelova, A., Kuo, W.: Contrastive feature masking open-vocabulary vision transformer. In: ICCV. pp. 15602-15612 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "23. Kim, D., Angelova, A., Kuo, W.: Region-aware pretraining for open-vocabulary object detection with vision transformers. In: CVPR. pp. 11144-11154 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 216, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 236 + ], + "type": "text", + "content": "24. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. ArXiv:1312.6114 (2013)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 237, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 258 + ], + "type": "text", + "content": "25. Korte, B.H., Vygen, J., Korte, B., Vygen, J.: Combinatorial optimization, vol. 1. Springer (2011)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "text", + "content": "26. Kuo, W., Cui, Y., Gu, X., Piergiovanni, A., Angelova, A.: Open-vocabulary object detection upon frozen vision and language models. In: ICLR (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "type": "text", + "content": "27. Kwon, M., Jeong, J., Uh, Y.: Diffusion models already have a semantic latent space. In: ICLR (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "text", + "content": "28. Leng, Y., Chen, Z., Guo, J., Liu, H., Chen, J., Tan, X., Mandic, D., He, L., Li, X.Y., Qin, T., et al.: Binauralgrad: A two-stage conditional diffusion probabilistic model for binaural audio synthesis. ArXiv:2205.14807 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "text", + "content": "29. Li, A.C., Prabhudesai, M., Duggal, S., Brown, E., Pathak, D.: Your diffusion model is secretly a zero-shot classifier. ArXiv:2303.16203 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "text", + "content": "30. Li, C., Feng, B.Y., Fan, Z., Pan, P., Wang, Z.: Steganerf: Embedding invisible information within neural radiance fields. In: ICCV. pp. 441-453 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "31. Li, C., Lin, X., Mao, Y., Lin, W., Qi, Q., Ding, X., Huang, Y., Liang, D., Yu, Y.: Domain generalization on medical imaging classification using episodic training with task augmentation. Computers in biology and medicine 141, 105144 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "text", + "content": "32. Li, C., Liu, H., Fan, Z., Li, W., Liu, Y., Pan, P., Yuan, Y.: Gaussianstego: A generalizable stenography pipeline for generative 3d gaussians splatting. ArXiv:2407.01301 (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 445, + 481, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 445, + 481, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 445, + 481, + 478 + ], + "type": "text", + "content": "33. Li, C., Liu, H., Liu, Y., Feng, B.Y., Li, W., Liu, X., Chen, Z., Shao, J., Yuan, Y.: Endora: Video generation models as endoscopy simulators. ArXiv:2403.11050 (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "text", + "content": "34. Li, C., Liu, X., Li, W., Wang, C., Liu, H., Yuan, Y.: U-kan makes strong backbone for medical image segmentation and generation. ArXiv:2406.02918 (2024)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 501, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 522 + ], + "type": "text", + "content": "35. Li, C., Zhang, Y., Li, J., Huang, Y., Ding, X.: Unsupervised anomaly segmentation using image-semantic cycle translation. ArXiv:2103.09094 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 522, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 555 + ], + "type": "text", + "content": "36. Li, C., Zhang, Y., Liang, Z., Ma, W., Huang, Y., Ding, X.: Consistent posterior distributions under vessel-mixing: a regularization for cross-domain retinal artery/vein classification. In: ICIP. pp. 61-65 (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 555, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 577 + ], + "type": "text", + "content": "37. Li, W., Chen, Z., Li, B., Zhang, D., Yuan, Y.: Htd: Heterogeneous task decoupling for two-stage object detection. TIP (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 577, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 481, + 599 + ], + "type": "text", + "content": "38. Li, W., Guo, X., Yuan, Y.: Novel scenes & classes: Towards adaptive open-set object detection. In: ICCV. pp. 15780-15790 (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "type": "text", + "content": "39. Li, W., Liu, J., Han, B., Yuan, Y.: Adjustment and alignment for unbiased open set domain adaptation. In: CVPR. pp. 24110-24119 (2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 621, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 643 + ], + "type": "text", + "content": "40. Li, W., Liu, X., Yao, X., Yuan, Y.: Scan: Cross domain object detection with semantic conditioned adaptation. In: AAAI. pp. 1421-1428 (2022)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "41. Li, W., Liu, X., Yuan, Y.: Scan++: Enhanced semantic conditioned adaptation for domain adaptive object detection. TMM (2022)" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "42. Li, W., Liu, X., Yuan, Y.: Sigma: Semantic-complete graph matching for domain adaptive object detection. In: CVPR (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "text", + "content": "43. Li, W., Liu, X., Yuan, Y.: Sigma++: Improved semantic-complete graph matching for domain adaptive object detection. TPAMI (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 480, + 194 + ], + "type": "text", + "content": "44. Li, W., Yang, C., Liu, J., Liu, X., Guo, X., Yuan, Y.: Joint polyp detection and segmentation with heterogeneous endoscopic data. In: ISBI Workshop: EndoCV 2021. pp. 69-79. CEUR-WS Team (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "text", + "content": "45. Liang, C., Wang, W., Miao, J., Yang, Y.: Gmmseg: Gaussian mixture based generative semantic segmentation models. ArXiv:2210.02025 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "46. Lin, C., Sun, P., Jiang, Y., Luo, P., Qu, L., Haffari, G., Yuan, Z., Cai, J.: Learning object-language alignments for open-vocabulary object detection. In: ICLR (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "text", + "content": "47. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "48. Lipman, Y., Chen, R.T.Q., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. In: ICLR (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "text", + "content": "49. Liu, H., Liu, Y., Li, C., Li, W., Yuan, Y.: Lgs: A light-weight 4d gaussian splatting for efficient surgical scene reconstruction. ArXiv:2406.16073 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 303, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 480, + 325 + ], + "type": "text", + "content": "50. Liu, X., Li, W., Yamaguchi, T., Geng, Z., Tanaka, T., Tsai, D.P., Chen, M.K.: Stereo vision meta-lens-assisted driving vision. ACS Photonics (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "type": "text", + "content": "51. Liu, X., Gong, C., Liu, Q.: Flow straight and fast: Learning to generate and transfer data with rectified flow. ArXiv:2209.03003 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "text", + "content": "52. Liu, X., Li, W., Yang, Q., Li, B., Yuan, Y.: Towards robust adaptive object detection under noisy annotations. In: CVPR. pp. 14207-14216 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "text", + "content": "53. Liu, X., Li, W., Yuan, Y.: Decoupled unbiased teacher for source-free domain adaptive medical object detection. TNNLS (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "text", + "content": "54. Liu, Y., Li, C., Yang, C., Yuan, Y.: Endogaussian: Gaussian splatting for deformable surgical scene reconstruction. ArXiv:2401.12561 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "type": "text", + "content": "55. Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., Zhu, J.: Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. ArXiv:2211.01095 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "text", + "content": "56. Luo, J., Li, Y., Pan, Y., Yao, T., Feng, J., Chao, H., Mei, T.: Semantic-conditional diffusion networks for image captioning. ArXiv:2212.03099 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "text", + "content": "57. Ma, C., Jiang, Y., Wen, X., Yuan, Z., Qi, X.: Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. NeurIPS 36 (2024)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "text", + "content": "58. Maaz, M., Rasheed, H., Khan, S., Khan, F.S., Anwer, R.M., Yang, M.H.: Class-agnostic object detection with multi-modal transformer. In: ECCV (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 501, + 480, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 523 + ], + "type": "text", + "content": "59. Miller, D., Sünderhauf, N., Milford, M., Dayoub, F.: Uncertainty for identifying open-set errors in visual object detection. Robot. Lett. 7(1), 215-222 (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "text", + "content": "60. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., et al.: Simple open-vocabulary object detection. In: ECCV. pp. 728-755 (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 555, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 578 + ], + "type": "text", + "content": "61. Ni, J., Qiu, Q., Chellappa, R.: Subspace interpolation via dictionary learning for unsupervised domain adaptation. In: CVPR (June 2013)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "text", + "content": "62. Nichol, A.Q., Dhariwal, P.: Improved denoising diffusion probabilistic models. In: ICML (2021)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 600, + 434, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 434, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 434, + 610 + ], + "type": "text", + "content": "63. Norris, J.R.: Markov chains. No. 2, Cambridge university press (1998)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "64. Pan, P., Fan, Z., Feng, B.Y., Wang, P., Li, C., Wang, Z.: Learning to estimate 6dof pose from limited data: A few-shot, generalizable approach using rgb images. In: 2024 International Conference on 3D Vision (3DV). pp. 1059-1071. IEEE (2024)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "65. Peebles, W., Xie, S.: Scalable diffusion models with transformers. ArXiv:2212.09748 (2022)" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 143, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 143, + 90, + 447, + 102 + ], + "type": "text", + "content": "CLIFF: Continual Latent Diffusion for Open-Vocabulary Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 482, + 659 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 133, + 117, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 482, + 149 + ], + "type": "text", + "content": "66. Preechakul, K., Chathee, N., Wizadwongsa, S., Suwajanakorn, S.: Diffusion autoencoders: Toward a meaningful and decodable representation. In: CVPR. pp. 10619-10629 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "type": "text", + "content": "67. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 214 + ], + "type": "text", + "content": "68. Rasheed, H.A., Maaz, M., Khattak, M.U., Khan, S., Khan, F.: Bridging the gap between object and image-level representations for open-vocabulary detection. In: NeurIPS (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 481, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 481, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 481, + 235 + ], + "type": "text", + "content": "69. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. In: NeurIPS (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 235, + 481, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 235, + 481, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 235, + 481, + 256 + ], + "type": "text", + "content": "70. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 256, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 256, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 256, + 481, + 277 + ], + "type": "text", + "content": "71. Shi, C., Yang, S.: Edadet: Open-vocabulary object detection using early dense alignment. In: ICCV. pp. 15724-15734 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 277, + 481, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 481, + 298 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 481, + 298 + ], + "type": "text", + "content": "72. Wang, J., Zhang, H., Hong, H., Jin, X., He, Y., Xue, H., Zhao, Z.: Open-vocabulary object detection with an open corpus. In: ICCV. pp. 6759-6769 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 298, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 298, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 298, + 481, + 319 + ], + "type": "text", + "content": "73. Wizadwongsa, S., Suwajanakorn, S.: Accelerating guided diffusion sampling with splitting numerical methods. In: ICLR (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 319, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 481, + 341 + ], + "type": "text", + "content": "74. Wu, L., Gong, C., Liu, X., Ye, M., Liu, Q.: Diffusion-based molecule generation with informative prior bridges. ArXiv:2209.00865 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 341, + 481, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 481, + 361 + ], + "type": "text", + "content": "75. Wu, S., Zhang, W., Jin, S., Liu, W., Loy, C.C.: Aligning bag of regions for open-vocabulary object detection. ArXiv:2302.13996 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 361, + 481, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 481, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 481, + 383 + ], + "type": "text", + "content": "76. Wu, X., Zhu, F., Zhao, R., Li, H.: Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. ArXiv:2303.13076 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 383, + 481, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 383, + 481, + 404 + ], + "spans": [ + { + "bbox": [ + 132, + 383, + 481, + 404 + ], + "type": "text", + "content": "77. Xu, H., Zhang, Y., Sun, L., Li, C., Huang, Y., Ding, X.: Afsc: Adaptive fourier space compression for anomaly detection. ArXiv:2204.07963 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 404, + 481, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 404, + 481, + 436 + ], + "spans": [ + { + "bbox": [ + 132, + 404, + 481, + 436 + ], + "type": "text", + "content": "78. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. ArXiv:2303.04803 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 436, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 481, + 457 + ], + "type": "text", + "content": "79. Yang, Q., Li, W., Li, B., Yuan, Y.: Mrm: Masked relation modeling for medical image pre-training with genetics. In: ICCV. pp. 21452-21462 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 457, + 481, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 478 + ], + "type": "text", + "content": "80. Ye, K., Zhang, M., Kovashka, A., Li, W., Qin, D., Berent, J.: Cap2det: Learning to amplify weak caption supervision for object detection. In: ICCV (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 478, + 481, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 478, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 478, + 481, + 499 + ], + "type": "text", + "content": "81. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary detr with conditional matching. In: ECCV. pp. 106-122 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 499, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 499, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 499, + 481, + 520 + ], + "type": "text", + "content": "82. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: CVPR (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 520, + 481, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 481, + 552 + ], + "type": "text", + "content": "83. Zbinden, L., Doorenbos, L., Pissas, T., Huber, A.T., Sznitman, R., Márquez-Neila, P.: Stochastic segmentation with conditional categorical diffusion models. In: ICCV (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 552, + 481, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 552, + 481, + 584 + ], + "spans": [ + { + "bbox": [ + 132, + 552, + 481, + 584 + ], + "type": "text", + "content": "84. Zhang, Y., Li, C., Lin, X., Sun, L., Zhuang, Y., Huang, Y., Ding, X., Liu, X., Yu, Y.: Generator versus segmentor: Pseudo-healthy synthesis. In: MICCAI. pp. 150–160 (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 584, + 481, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 584, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 584, + 481, + 616 + ], + "type": "text", + "content": "85. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: CVPR. pp. 16793-16803 (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 616, + 481, + 638 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 616, + 481, + 638 + ], + "spans": [ + { + "bbox": [ + 132, + 616, + 481, + 638 + ], + "type": "text", + "content": "86. Zhou, X., Girdhar, R., Joulin, A., Krahenbuhl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: ECCV. pp. 350-368 (2022)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 638, + 481, + 659 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 638, + 481, + 659 + ], + "spans": [ + { + "bbox": [ + 132, + 638, + 481, + 659 + ], + "type": "text", + "content": "87. Zhou, X., Koltun, V., Krahenbuhl, P.: Probabilistic two-stage detection. In: ArXiv:2103.07461 (2021)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 199, + 100 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CLIP-DINOiser_ Teaching CLIP a few DINO tricks for open-vocabulary semantic segmentation/ab181842-dc49-431d-ab8a-9cea7505e7b5_content_list.json b/2024/CLIP-DINOiser_ Teaching CLIP a few DINO tricks for open-vocabulary semantic segmentation/ab181842-dc49-431d-ab8a-9cea7505e7b5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4981b8390daa13f3458cc5767f8136410020fe4a --- /dev/null +++ b/2024/CLIP-DINOiser_ Teaching CLIP a few DINO tricks for open-vocabulary semantic segmentation/ab181842-dc49-431d-ab8a-9cea7505e7b5_content_list.json @@ -0,0 +1,1529 @@ +[ + { + "type": "text", + "text": "CLIP-DINOiser: Teaching CLIP a few DINO tricks for open-vocabulary semantic segmentation", + "text_level": 1, + "bbox": [ + 220, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Monika Wysoczanska1, Oriane Simeoni2, Michael Ramamonjisoa3*, Andrei Bursuc2, Tomasz Trzciński1,4,5, and Patrick Pérez2", + "bbox": [ + 246, + 212, + 754, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Warsaw University of Technology, 2valeo.ai, 3Meta AI, 4Tooploox, 5IDEAS NCBR", + "bbox": [ + 223, + 253, + 779, + 270 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The popular CLIP model displays impressive zero-shot capabilities thanks to its seamless interaction with arbitrary text prompts. However, its lack of spatial awareness makes it unsuitable for dense computer vision tasks, e.g., semantic segmentation, without an additional fine-tuning step that often uses annotations and can potentially suppress its original open-vocabulary properties. Meanwhile, self-supervised representation methods have demonstrated good localization properties without human-made annotations nor explicit supervision. In this work, we take the best of both worlds and propose an open-vocabulary semantic segmentation method, which does not require any annotations. We propose to locally improve dense MaskCLIP features, which are computed with a simple modification of CLIP's last pooling layer, by integrating localization priors extracted from self-supervised features. By doing so, we greatly improve the performance of MaskCLIP and produce smooth outputs. Moreover, we show that the used self-supervised feature properties can directly be learnt from CLIP features. Our method CLIP-DINOiser needs only a single forward pass of CLIP and two light convolutional layers at inference, no extra supervision nor extra memory and reaches state-of-the-art results on challenging and fine-grained benchmarks such as COCO, Pascal Context, Cityscapes and ADE20k. The code to reproduce our results is available at https://github.com/wysoczanska/clip_dinoiser.", + "bbox": [ + 261, + 310, + 743, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: open-vocabulary semantic segmentation $\\cdot$ self-supervised features $\\cdot$ annotation-free segmentation", + "bbox": [ + 259, + 628, + 746, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 690, + 377, + 707 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic segmentation is a key visual perception task for many real-world systems, e.g., self-driving cars, and industrial robots. Typically tackled in a dataset-oriented manner, best methods require a training dataset which is manually annotated for a specific and finite set of classes. The advent of powerful Vision-Language Models (VLM) [24,43,63] is stimulating a shift from a closed-vocabulary", + "bbox": [ + 212, + 724, + 792, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Work done outside of Meta and Meta was not involved in the research discussed here.", + "bbox": [ + 217, + 811, + 785, + 838 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3c26a0e78ac5a532f064d0e9558cf8d9b1b7183e8f2abf666a5549b63b5f5285.jpg", + "image_caption": [ + "Fig. 1: Examples of open-vocabulary semantic segmentation results obtained with our method CLIP-DINOiser on 'in-the-wild' images vs. those of MaskCLIP [67]. Our method improves MaskCLIP features with a smart pooling strategy which does not alter the original open-vocabulary properties. We use self-supervised DINO [5] as a guide to teach CLIP [22] to produce DINO-like localization features through two light convolutional layers. Our method, which achieves state-of-the-art results, only requires a single forward pass through CLIP model and our two layers. In addition to the correct prompts (light grey row) we list the irrelevant prompts predicted (in yellow) that we query in all images shown here." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 789, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "paradigm to an open-world one. Such models are trained with a simple but scalable objective: to align pairs of image and coarse text captions that can be obtained in large amounts with limited manual supervision. VLMs excel at associating global image content with arbitrary text inputs with remarkable generalization capabilities [17, 34], but struggle to provide dense open-vocabulary features [18, 67]. Obtaining such an alignment between pixels and language can lead to open-vocabulary extensions for multiple other modalities, such as point clouds [8, 23, 38, 41], 3D scenes [54], 3D shapes [1], radiance fields [27], intermodality alignment [19, 23], with multiple potential applications for which the construction of training datasets is even more challenging and where CLIP-derived models show promising results.", + "bbox": [ + 212, + 532, + 787, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Different strategies have been recently proposed towards improving CLIP's patch-level feature extraction abilities by modifying the original CLIP architecture for dense pooling and retraining [6, 37, 44, 61, 62] or finetuning on an annotated segmentation dataset with pre-defined classes [32, 67]. The former requires long training and/or large collections of annotated data, while the latter leads to an alteration of the vision-language associations of the CLIP features. An alternative line of approaches freezes the CLIP encoder and directly densifies its features with different heuristics, often with multiple forward passes [1, 23, 27, 49, 50, 59], but are less practical due to the extensive computa", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "M.Wysoczanska et al.", + "bbox": [ + 271, + 114, + 416, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tional overhead. MaskCLIP [67] arises as a computationally efficient dense CLIP extractor. It converts CLIP's global self-attention layer into a convolutional one to produce patch features with original vision-language qualities. If such features are local, they appear to be too noisy for high-quality segmentation mask extraction (see Fig. 2b middle column).", + "bbox": [ + 212, + 145, + 782, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Meanwhile, recent self-supervised learning (SSL) approaches [4, 5, 9, 68] produce strong visual representations displaying object localization properties, and such without requiring any manual annotation. DINO [5] stands out with its semantically meaningful features which have been exploited for unsupervised object discovery [51, 52, 56, 57]. DINO features prove useful also for zero-shot semantic segmentation [25, 27, 59], but require expensive sliding window sampling [27, 59] or building concept-specific prototypes and ensemble strategies [25].", + "bbox": [ + 212, + 222, + 784, + 328 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we aim for unaltered patch-level CLIP features with minimal runtime overhead. To this end, we re-examine the localization properties of MaskCLIP features and observe that it is possible to easily refine them with guidance from SSL models. In detail, we train a simple convolutional layer on unlabeled data to produce pooling weights to perform correlation-guided dense feature pooling from CLIP without distorting the vision-language alignment. This layer is optimized to mimic the patch correlations of DINO [5] that indicate likely layouts of visual concepts in the images. Furthermore, we show that the unsupervised objectness information given by FOUND [52] from DINO features can be also directly learned from CLIP features again in a fully-unsupervised fashion with a single convolutional layer and helps improve the segmentation of the ill-defined 'background' prompt. With CLIP-DINOiser, we obtain high-quality masks in a single forward pass on CLIP (see Fig. 1). CLIP-DINOiser is amenable to producing dense semantic maps.", + "bbox": [ + 212, + 329, + 785, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, our contributions are: (1) We propose a light pooling mechanism to refine MaskCLIP features by leveraging guidance from SSL features without degrading its original open-vocabulary properties. CLIP-DINOiser does not require any annotations, nor retraining CLIP from scratch, but only a single CLIP forward pass. (2) We show that CLIP already contains good localization properties which can be exploited. We leverage simple convolutional layers to emphasize visual concept layouts from dense CLIP features. We train them without any annotation on only 1k of raw images randomly sampled in ImageNet [13]. We believe that this finding could be further exploited in different contexts. (3) Our method achieves state-of-the-art results on complex semantic segmentation datasets such as COCO [3], Pascal Context [16], Cityscapes [11] and ADE20K [66].", + "bbox": [ + 212, + 540, + 787, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 744, + 387, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Zero-shot semantic segmentation. This task has been typically approached by methods which aim at generalizing from seen classes to unseen ones [2, 20, 21, 26, 30, 40, 60, 64]. Such strategies train models with full supervision on the set of seen classes and propose different solutions to extend them to unseen ones with", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "CLIP-DINOiser", + "bbox": [ + 627, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "out new images (labeled or unlabeled), e.g., by exploiting class information and relationships encapsulated in popular word embeddings [35,42]. While they produce fine segmentations without computational overhead, these methods require pixel-level annotations for the seen classes.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From CLIP to open-vocabulary segmentation. The surge of VLMs with aligned image-language representations [22, 24, 43] brought back into the spotlight the zero-shot classification task. However, the extension to zero-shot segmentation is not obvious as the CLIP architecture is not equipped to yield dense vision-language features [18, 67]. To produce dense CLIP features, several approaches fine-tune or train from scratch pixel-aligned CLIP-like models with additional modules, mechanisms or supervision objectives [6, 37, 44, 61, 62] on datasets with annotations of varying granularity and quality: dense annotations [29, 31], class-agnostic object masks [14, 18, 45], coarse captions [6, 18, 31–33, 37, 44, 61, 62, 65] or pseudo-labels [67]. Recent works leverage image-level captions to align text to regions (obtained without supervision): PACL [37] trains an embedder module to learn patch-to-text affinity, TCL [6] proposes a local contrastive objective to align well-selected patches to the text and ViewCO [46] leverages multi-view consistency. On the downside, such models require long training on millions of images or specific types of very costly annotations. Also, fine-tuning CLIP with a defined vocabulary is more computationally appealing [29, 31, 67], but alters the open-vocabulary properties of the features [23].", + "bbox": [ + 212, + 220, + 787, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Most related to us is a line of works that investigate how to directly densify CLIP features [1, 23, 27, 59, 67] to obtain per-patch CLIP features. Such densification can be performed by aggregating features from multiple views [1, 27] or from sliding windows [23, 59] at the extra-cost of multiple forward passes. MaskCLIP [67] drops the global pooling layer of CLIP and matches the projected features directly to text via a $1 \\times 1$ convolution layer. By doing so they achieve dense predictions, however noisy.", + "bbox": [ + 212, + 479, + 789, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With a concept-driven perspective, some methods [25,49,50] build codebooks of visual prototypes per concept, including negative prototypes [25], and then perform co-segmentation [49]. While such an approach yields good results, it is however at the cost of building expensive class-specific prototypes, therefore diverging from open-vocabulary scenarios. Instead, we aim to remain open to avoid retraining a model or building new expensive prototypes whenever a new concept is considered. To that end, we devise a dense CLIP-feature extraction method that preserves the open-vocabulary quality.", + "bbox": [ + 212, + 584, + 789, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Leveraging self-supervised models & CLIP. Recent self-supervised ViTs [4,5,9,12, 68] have demonstrated features with good localization properties [51,52,56,57]. Such features have also been exploited in the context of open-vocabulary segmentation methods, e.g. for pre-training for the visual backbone [7,44,62], cosegmentation [49], clustering patches into masks [47], representing object prototypes [25]. Related to us is the recent CLIP-DIY [59] which computes patch-level representations from CLIP features from different image crops with guidance from an unsupervised saliency segmenter [52] FOUND. While we also leverage", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "M.Wysoczanska et al.", + "bbox": [ + 271, + 114, + 415, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the latter, in contrast with CLIP-DIY which runs multiple forward passes to build their dense CLIP features, our method requires only a single forward pass of CLIP. Furthermore, our method mitigates the limits of FOUND in cluttered scenarios by integrating an uncertainty constraint. Finally, we leverage the informative patch correlation properties of DINO [5] and show that it is possible to teach CLIP to produce DINO-like features through light convolutional layers.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 261, + 331, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We present in this section CLIP-DINOiser, a simple and efficient strategy to improve MaskCLIP using localization information extracted from CLIP—with a lightweight model trained to mimic some of DINO's properties. We first set the goal in Sec. 3.1 and present MaskCLIP [67] in Sec. 3.2. We then introduce our strategy which leverages self-supervised features localization information to consolidate MaskCLIP features in Sec. 3.3 and discuss how such localization information can directly be learnt from CLIP in Sec. 3.4 (we visualize both steps in Fig. 3). We also propose a way to improve the 'background' filtering in Sec. 3.5.", + "bbox": [ + 212, + 294, + 787, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Problem statement", + "text_level": 1, + "bbox": [ + 214, + 453, + 421, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this work, we aim to produce open-vocabulary1 semantic segmentation of an image. We consider an image $X \\in \\mathbb{R}^{H \\times W \\times 3}$ which we split into a sequence of $N$ patches of dimensions $P \\times P \\times 3$ with $P \\times P$ the patch size and $N = \\lceil \\frac{H}{P} \\rceil \\cdot \\lceil \\frac{W}{P} \\rceil$ . A class token, noted CLS, is added to the input sequence and we feed the $N + 1$ patches to a ViT [15] model. We aim at producing dense visual features $F \\in \\mathbb{R}^{N \\times d}$ , with $d$ the feature dimension, that can later be matched to any set of text inputs embedded in the same space. In particular, the goal is to produce a segmentation map per textual query.", + "bbox": [ + 212, + 479, + 789, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Preliminaries on MaskCLIP", + "text_level": 1, + "bbox": [ + 214, + 625, + 493, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Extracting dense open-vocabulary features. The popular CLIP [22] model pretrained on image/caption pairs produces good global image features, but was not trained to generate high-quality 2D feature maps. In order to extract such dense feature maps relevant to semantic segmentation, Zhou et al. [67] revisit the global attention pooling layer of the last attention layer of the model. The authors discard the query and key embeddings of the layer and transform both the value projection and the last linear layer into a conv $1 \\times 1$ layer. With this new model, named MaskCLIP and denoted $\\phi(\\cdot)$ , we extract $d$ -dimensional features $\\phi^L(X) \\in \\mathbb{R}^{N \\times d}$ from the last layer $L$ which retains most of the open-vocabulary properties of CLIP [67].", + "bbox": [ + 212, + 648, + 787, + 801 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "CLIP-DINOiser", + "bbox": [ + 627, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1 We adopt the taxonomy defined in the recent survey [58] and define our method as 'open-vocabulary', with capabilities to generalize to unseen datasets.", + "bbox": [ + 217, + 810, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic segmentation given textual queries. We also extract CLIP textual features $\\phi_T(t_j)$ for each text query $t_j \\in \\mathcal{T}$ with $j \\in \\{1, \\dots, |\\mathcal{T}|\\}$ . Segmentation maps are then generated by computing the cosine similarity between each of the visual patch features and of the textual prompts, after L2-normalization. The most similar prompt is assigned to each patch. Note that a query 'background' can be added in order to obtain negative patches. Using MaskCLIP allows us to produce dense segmentation maps with a single forward pass of the classic CLIP model, but its outputs are noisy, as visible in Fig. 2b (middle column).", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 DINOising open-vocabulary features", + "text_level": 1, + "bbox": [ + 214, + 280, + 565, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this work, we aim to improve MaskCLIP's open-vocabulary features described above. To do so, we propose to leverage the known good localization properties of self-supervised features [5,39,51-53,57].", + "bbox": [ + 212, + 301, + 785, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Extracting self-supervised correlation information. Recent works [51, 57] have shown that the patch correlation information of the embeddings from the last attention layer of the self-supervised model, DINO [5] can help highlight objects in images. We use here the value embeddings which we observe have finer correlation than those of key and query (more discussion in supplementary material). We extract such self-supervised features $\\xi(X) \\in \\mathbb{R}^{N \\times d_{\\xi}}$ and discard the CLS token. We then compute the per-patch cosine-similarity and produce the affinity map $A^{\\xi} \\in [-1,1]^{N \\times N}$ . We compare in Fig. 4 the patch-similarities obtained for a patch seed with MaskCLIP and DINO features and observe that the self-supervised features are more densely and accurately correlated than those of CLIP.", + "bbox": [ + 212, + 358, + 787, + 525 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0729da00fee285f570f825a5d25997b27a4bcc703e4cc49a29d518b57c54c55d.jpg", + "image_caption": [ + "(a) Our guided pooling" + ], + "image_footnote": [], + "bbox": [ + 225, + 544, + 506, + 696 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9b6982e578b1649eb765a836d24013c91ea8b704e9ba97f8b52304b57d6020ee.jpg", + "image_caption": [ + "(b) Impact of the pooling", + "Fig. 2: We present in (a) our guided pooling strategy defined in Eq. (1). The $N \\times N$ affinity matrix is computed from patch features and is used to refine MaskCLIP features (bottom left). In (b) we compare our results with $F^{+}$ (right) versus those obtained with MaskCLIP features (middle)." + ], + "image_footnote": [], + "bbox": [ + 521, + 551, + 781, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Strengthening features with guided pooling. In order to locally consolidate MaskCLIP features $\\phi^L (X)$ , now noted $F$ , we propose to perform a concept-aware linear combination of the features per patch with guidance from the patch affinity $A^{\\xi}$ . The", + "bbox": [ + 212, + 794, + 808, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "M.Wysoczanska et al.", + "bbox": [ + 271, + 114, + 416, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "feature combination strategy can be seen as a form of voting mechanism that enforces similar patches to have similar CLIP features (and prediction) while attenuating noisy features. Specifically, we compute the new features $F^{+} \\in \\mathbb{R}^{N \\times d}$ as an average of MaskCLIP features $F$ weighted by $A^{\\xi}$ , presented in Fig. 2a. We zero-out $A^{\\xi}$ correlations below a threshold $\\gamma$ , following [51, 57], and compute the new features for patch $p \\in \\{1, \\dots, N\\}$ :", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nF _ {p} ^ {+} = \\frac {1}{\\sum_ {q = 1} ^ {N} A _ {p , q} ^ {\\xi}} \\sum_ {q = 1} ^ {N} A _ {p, q} ^ {\\xi} \\cdot F _ {q}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 244, + 785, + 286 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We then produce the segmentation maps $S \\in [-1,1]^{N \\times |\\mathcal{T}|}$ , by comparing the new features $F^+$ to each textual queries in $\\mathcal{T}$ . As shown in Fig. 2b, when using such consolidated features, we obtain more accurate outputs and the high-frequency predictions observed in MaskCLIP are smoothed out, showing the benefit of the pooling.", + "bbox": [ + 212, + 295, + 787, + 371 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Teaching CLIP a first DINO trick: object correlations", + "text_level": 1, + "bbox": [ + 214, + 390, + 705, + 406 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/716ea1effc3d6881eac5b1b729f0220ba496b5e336bfe34c39d3b5c2bb8fcf30.jpg", + "image_caption": [ + "Fig. 3: Overview of CLIP-DINOiser which leverages the quality of self-supervised features to improve the notoriously noisy MaskCLIP feature maps. We use DINO as a teacher which 'teaches' CLIP how to extract localization information. We train (left) a conv $3 \\times 3$ layer to reproduce the patch correlations obtained with DINO. At inference (right), an input image is forwarded through the frozen CLIP image backbone and MaskCLIP projection. The produced features are then improved with our pooling strategy which is guided by correlations predicted with the trained convolutional layer applied on CLIP. With this light 'DINOising' process, we obtain 'DINOised' features which are matched against the prompts features to produce CLIP-DINOiser outputs." + ], + "image_footnote": [], + "bbox": [ + 222, + 438, + 784, + 640 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We have shown in the previous section that self-supervised correlation information can successfully be used to improve the dense quality of open-vocabulary", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "CLIP-DINOiser", + "bbox": [ + 627, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "features. If the difficulty of densifying CLIP is well-known, we show here that CLIP features already contain good localization information which can be extracted with a light model. We indeed predict DINO correlations $A^{\\xi}$ from CLIP with a single convolutional layer.", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In order to predict the DINO affinity map $A^{\\xi}$ from CLIP features, we train a single $3 \\times 3$ convolutional layer $g(\\cdot): \\mathbb{R}^{d} \\to \\mathbb{R}^{d_{g}}$ which projects intermediate features $\\phi^{l}(X)$ -extracted from layer $l$ -into a smaller space of dimension $d_{g} < d$ . We enforce the patch correlations of the generated features $A^{\\phi} \\in [-1,1]^{N \\times N}$ :", + "bbox": [ + 212, + 207, + 787, + 268 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nA ^ {\\phi} = \\frac {g \\left(\\phi^ {l} (X)\\right)}{\\| g \\left(\\phi^ {l} (X)\\right) \\|} \\otimes \\left(\\frac {g \\left(\\phi^ {l} (X)\\right)}{\\| g \\left(\\phi^ {l} (X)\\right) \\|}\\right) ^ {\\top}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 361, + 277, + 785, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with $\\otimes$ denoting the outer product, to be close to the binarized correlations $D = A^{\\xi} > \\gamma$ (we use here the same $\\gamma$ as defined above), using the binary cross-entropy loss $\\mathcal{L}^c$ :", + "bbox": [ + 214, + 325, + 785, + 369 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {c} = \\sum_ {p = 1} ^ {N} \\left[ D _ {p} \\log A _ {p} ^ {\\phi} + (1 - D _ {p}) \\log \\left(1 - A _ {p} ^ {\\phi}\\right) \\right]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 381, + 785, + 424 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present our layer training in Fig. 3 (left part) and observe the quality of CLIP-predicted affinity matrix $A^{\\phi}$ . We also show in Fig. 4 another example of obtained $A^{\\phi}$ and observe their similarity to DINO-based correlations. We use the CLIP-produced correlations $A^{\\phi}$ to replace $A^{\\xi}$ in Eq. (1) to weight the pooling and observe a similar boost over MaskCLIP, thus showing that good patch correlations can indeed be extracted directly from CLIP. We can now discard DINO and we name CLIP-DINOiser the guided-pooling strategy which uses CLIP-based correlation. As shown in Fig. 3 (inference step), our method runs with a single forward pass of CLIP model and a small extra layer.", + "bbox": [ + 212, + 435, + 787, + 571 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f70992381d5ab9a8d6f8a4a3b46d5082dbda506ab7a5d0d2bfc451c3c6c1ec6f.jpg", + "image_caption": [ + "Fig. 4: Comparison of the affinity maps between a seed (one on the 'plant' and the other on a 'pillow') and the other patch features when using features of MaskCLIP, DINO and ours after training." + ], + "image_footnote": [], + "bbox": [ + 222, + 590, + 781, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Teaching CLIP a second DINO trick: background filtering", + "text_level": 1, + "bbox": [ + 214, + 739, + 741, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Moreover, as discussed earlier, a 'background' query may be added to the set of textual queries $\\mathcal{T}$ in order to help filter out patches falling in the background and not corresponding to any objects. We do not assume here any prior knowledge about classes of interest and focus rather on the foreground/background paradigm [52]. We argue that relying solely on the textual prompt 'background'", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "M.Wysoczanska et al.", + "bbox": [ + 271, + 114, + 416, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e6f1ef85e944166ce6e14abc8f30d5fa2df07c5a28b8733a85ea850f7e14f0af.jpg", + "image_caption": [ + "(a) Comparison of background filtering" + ], + "image_footnote": [], + "bbox": [ + 228, + 170, + 467, + 310 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/838aa25cde5bf5f63a9a16f11c3540528c48854aef5d5863b23e3cc6e50cc42c.jpg", + "image_caption": [ + "(b) Background filtering", + "Fig. 5: We present in (a) a comparison of objectness mask generated by FOUND [52] and with our layer using CLIP features. We carefully define the fusion operation and the simple training strategy of the conv $1 \\times 1$ again using DINO as a teacher in Sec. 3.5. In (b) is an overview of our background filtering which is applied when a 'background' prompt is provided and helps reduce hallucinations." + ], + "image_footnote": [], + "bbox": [ + 480, + 143, + 767, + 310 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "to catch all non-salient patches is underperforming and, similarly to [59], we propose to use a very light-weight unsupervised foreground/background segmentation method, namely FOUND [52] which also relies on DINO self-supervised features. We run FOUND on the entire image and extract a prediction mask $M \\in \\{0,1\\}^N$ in which a patch is assigned the value 1 if falling into the foreground and 0 otherwise. We also observe that saliencies produced by FOUND can be too restrictive and discard objects which are partially visible or in a clutter. To mitigate this behaviour, we propose to relax the background selection by integrating an additional uncertainty constraint. To this end, we fuse the background information from both modalities by assigning the 'background' prompt to patches $p$ which are both uncertain, e.g. have low confidence score $\\sigma(S)_p < \\delta$ , with $\\sigma(\\cdot)$ the softmax operation, and which fall in the background in $M$ .", + "bbox": [ + 212, + 433, + 787, + 617 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Learning FOUND objectness. Moreover, we are also able to learn the predictions of FOUND [52] directly from CLIP features. To do so, we train a single $1 \\times 1$ convolutional layer $h(\\cdot): \\mathbb{R}^d \\to \\mathbb{R}$ which predicts from the features $\\phi^l(X)$ an objectness map $M^\\phi = h(\\phi^l(X)) \\in \\mathbb{R}^N$ . We train the model to predict the FOUND binary mask $M$ with the binary cross-entropy loss $\\mathcal{L}^m$ :", + "bbox": [ + 212, + 628, + 787, + 705 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {m} = \\sum_ {p = 1} ^ {N} \\left[ M _ {p} \\log \\left(M _ {p} ^ {\\phi}\\right) + \\left(1 - M _ {p}\\right) \\log \\left(1 - M _ {p} ^ {\\phi}\\right) \\right].\n$$\n", + "text_format": "latex", + "bbox": [ + 320, + 713, + 679, + 755 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We show examples of predicted CLIP-based objectness in Fig. 5a and observe their very high similarity to those produced with DINO. Moreover, we can now replace $M$ defined above with the binarized CLIP-based scores $\\zeta(M^{\\phi}) > 0.5$ with $\\zeta(\\cdot)$ the sigmoid operation, and observe a minimal drop in performances. We show an example of the background filtering with trained objectness in Fig. 5b.", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "CLIP-DINOiser", + "bbox": [ + 627, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 162 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We detail in Sec. 4.1 the experimental setup used in our evaluation. We produce state-of-the-art results on the task of open-vocabulary semantic segmentation in Sec. 4.2 and ablation studies in Sec. 4.3.", + "bbox": [ + 212, + 178, + 787, + 223 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experimental setup", + "text_level": 1, + "bbox": [ + 215, + 247, + 424, + 263 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Technical details. We use in all experiments a frozen CLIP ViT-B/16 pre-trained following OpenCLIP [22]. Our method CLIP-DINOiser uses two convolutional layers to extract DINO-like information from CLIP layer $l = 10$ (the 3rd before the last which was shown to provide the best results [55]). The first layer $g(\\cdot)$ has a kernel $3 \\times 3$ and output dimension $d_{g} = 256$ and $h(\\cdot)$ a kernel $1 \\times 1$ with $d_{h} = 1$ . The first is trained to match the correlation information extracted from the value embeddings of the last layer of a ViT-B/16 model trained following DINO [5]. The second layer is trained to replicate the unsupervised object localization predictions of FOUND [52]–which also uses DINO model. We train both layers with a binary cross-entropy loss on only 1k raw images randomly sampled from ImageNet [13] dataset without any annotation. We report average scores over 3 runs with different sampling seeds and provide standard deviations in the supplementary material. We follow [57] and binarize the correlations with $\\gamma = 0.2$ . In the background filtering step, we use a high confidence score, i.e., $\\delta = 0.99$ . We train our model for 6k iterations with a batch size of 16 images using Adam optimizer [28], which takes approximately 3 hours on a single NVIDIA RTX A5000 GPU. We decrease the learning rate for both heads by a factor of 0.1 after 5k iterations. We apply data augmentations during training (random scale and cropping, flipping and photometric distortions).", + "bbox": [ + 212, + 272, + 789, + 561 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets and metric. We evaluate our method on eight benchmarks typically used for zero-shot semantic segmentation [6]. Following [6], we split them into two groups. The first consists in datasets with a 'background' query: PASCAL VOC [16] (noted 'VOC'), PASCAL Context [36] (noted 'Context'), and COCO Object [3] (noted 'Object') and the second without: PASCAL VOC20 [16] (noted 'VOC20'), PASCAL Context59 [36] (noted 'C59'), COCO-Stuff [3] (noted 'Stuff'), Cityscapes [11] (noted 'City'), and ADE20K [66] (noted 'ADE'). We evaluate results with the standard mIoU metric. We also follow the evaluation protocol of [6], use the implementations provided by MMSegmentation [10], employ a sliding window strategy, resize the input image to have a shorter side of 448. We also do not perform text expansions of the class names and use only the standard ImageNet prompts following [22, 61, 67].", + "bbox": [ + 212, + 570, + 789, + 752 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines. We compare our method against state-of-the-art methods on open-vocabulary zero-shot semantic segmentation. For a fair comparison between methods, we report results without any post-processing step. In our evaluations, we follow the taxonomy presented in [58] and compare our model with the methods relying on language-image pretraining, also called open-vocabulary.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "M.Wysoczanska et al.", + "bbox": [ + 271, + 114, + 416, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We split the compared baselines into four categories: (1) dataset specific which employ pseudo-labeling and supervised training of a segmentation model on target dataset: NamedMask [50], MaskCLIP + [67]); (2) construct prototypes: ReCO [49], OVDiff [25]; (3) train with text supervision including GroupViT [61], ZeroSeg [47], SegCLIP [33], TCL [6], CLIPpy [44], OVSegmentor [62], which all require access to additional datasets of millions of image/caption pairs (we note in the table the exact datasets used for the training); and finally use frozen CLIP i.e. CLIP-DIY [59] and MaskCLIP [67], which use pre-trained CLIP. Our method falls into the last category as we do not modify CLIP, and do not need access to additional caption annotations as we use only 1k unannotated images.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/9afa16e34166ebaead84cea2ede83a357d4435d11cb091eea252e9d6299a4282.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsConcept spec.Extra Inference data backboneNo background promptW/ bkg prompt Cont. Object VOC
VOC20C59StuffCityADE
Dataset specific
MaskCLIP+ [67]✓XIDLv2-31.118.0----
NamedMask [50]✓XIDLv3+-----27.759.2
Build prototypes per visual concept
ReCo [49]✓✓ICLIP57.822.314.821.111.219.915.7
OVDiff [25]✓✓XCLIP+ DINO+SD81.733.7--14.930.134.8
Text/image alignment training with captions
GroupViT [61] [6]XXITCLIP79.723.415.311.19.218.727.5
ZeroSeg [7]XXITCLIP-----21.822.1
SegCLIP [33]XXITCLIP---11.08.724.726.5
TCL [6]XXITCLIP77.530.319.623.114.924.330.4
CLIPpy [44]XXITCLIP----13.5-32.0
OVSegmentor [62]XXITCLIP----5.620.425.1
Frozen CLIP
CLIP-DIY [59]*X✓XCLIP+ DINO79.719.813.311.69.919.731.0
MaskCLIP [67] [6]X✓XCLIP53.723.314.721.610.821.115.5
MaskCLIP*X✓XCLIP61.825.617.625.014.322.916.4
MaskCLIP* †X✓XCLIP71.927.418.623.014.924.021.6
CLIP-DINOiserX✓I(1k)CLIP80.935.924.631.720.032.434.8
", + "bbox": [ + 218, + 328, + 782, + 625 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 1: Open-vocabulary semantic segmentation quantitative comparison using the mIoU metric. We separate in two columns the evaluation datasets: those without a 'background' prompt and those with (noted 'W/ bkg prompt'), as discussed in Sec. 4.1. We report all methods without post-processing. We note with * methods for which we computed scores; we obtained MaskCLIP* scores with OpenCLIP [22] and mark with † the use of MaskCLIP refinement. The first and second best methods are respectively bold and underlined. We specify if a method assumes prior access to names of concepts ('Concept spec.)' and if it employs a frozen backbone (''). We specify what additional data is used at training ('Extra data') ('I' stands for images and 'IT' for image/text aligned data). Our CLIP-DINOiser only needs 1k images from ImageNet to be trained. 'SD' stands for Stable Diffusion [48]. We refer to Sec. 4.1 for more details on baselines and we detail the datasets used for training by each method in the supplementary material.", + "bbox": [ + 212, + 636, + 787, + 816 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "CLIP-DINOiser", + "bbox": [ + 627, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Open-vocabulary semantic segmentation", + "text_level": 1, + "bbox": [ + 215, + 146, + 593, + 161 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We discuss in this section state-of-the-art results on the task of open-vocabulary semantic segmentation.", + "bbox": [ + 215, + 176, + 784, + 205 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Evaluation with no 'background' class. We first compare in Tab. 1 ('No background prompt' column) the results on datasets which aim at the segmentation of most of the pixels in an image and do not consider a 'background' class. We observe that our method CLIP-DINOiser achieves the best results on four datasets yielding $+2.2$ , $+5.0$ , $+6.7$ and $+5.1$ mIoU over the second best performing method. Interestingly, we outperform methods which build expensive prototypes per visual concept on fine-grained datasets, showing the benefit of our lightweight and generalizable method. The only drop $(-0.8$ mIoU) is seen on VOC20 with respect to OVDiff; we believe it is due to the benefit of generating per-concept negative prototypes which likely benefits this object-centric dataset. An adaptive granularity of feature correlation could help mitigate this drop, which we leave for future work.", + "bbox": [ + 215, + 226, + 785, + 407 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Evaluation with 'background' class. We now compare our method on datasets which include a 'background' query in Tab. 1 ('W/ bkg prompt' column). In this setup, we also apply our background detection mechanism (detailed in Sec. 3.5) on VOC and Object in order to improve the stuff-like background detection. We observe that CLIP-DINOiser significantly outperforms all methods which do not construct prototypes. Moreover, we surpass OVDiff (which uses an ensemble of three models) on Context dataset by $+2.3\\mathrm{mIoU}$ and are on par on Object. It is to be noted that with a single feature extractor, the performance of OVDiff drops by $-10\\mathrm{mIoU}$ and the method requires the construction of a 'background' prototype per concept, otherwise losing another $-10\\mathrm{mIoU}$ on VOC. On the other hand, CLIP-DINOiser produces segmentation masks in a single pass of CLIP with the light addition of two convolutional layers while remaining fully open-vocabulary as it does not require any concept-specific constructs.", + "bbox": [ + 215, + 428, + 785, + 623 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative results. We qualitatively compare in Fig. 6 CLIP-DINOiser with high-performing TCL [6], CLIP-DIY [59] (two recent methods which provide code) and our baseline method MaskCLIP [67] on images taken from the datasets considered in the evaluation. We observe that our method generates predictions accurate both in terms of localization and assignment. Indeed we obtain fined-grained results on the challenging datasets, e.g. in the Cityscapes example the text query 'car' and in the ADE20k example 'fountain' are accurately located when CLIP-DIY and TCL produce coarser results. Versus MaskCLIP, we can see the denoising capabilities of CLIP-DINOiser as MaskCLIP hallucinations grow with the number of text queries prompted at evaluation. Finally, in Fig. 1 we present 'in the wild' examples, beyond the evaluation benchmarks, and show that CLIP-DINOiser produces accurate segmentation masks for arbitrary and very specific prompts, such as 'wooden table' or 'leather bag'.", + "bbox": [ + 215, + 643, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "M.Wysoczanska et al.", + "bbox": [ + 271, + 114, + 416, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1ce59c04146c54e151e6a345fa5d585a1022ce5194e561eec165314b08037cbf.jpg", + "image_caption": [ + "Fig.6: Qualitative open-vocabulary segmentation results. We compare ours against CLIP-DIY [59], TCL [6] and MaskCLIP [67]. For a fair comparison, we do not apply post-processing. All pixels annotated in black are from the background class. We observe that our method achieves more accurate results both in terms of localization and class assignment." + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 792, + 510 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Ablation study", + "text_level": 1, + "bbox": [ + 215, + 619, + 387, + 635 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We now conduct an ablation study of the different components of CLIP-DINOiser and investigate the impact of both our feature pooling strategy and background detection.", + "bbox": [ + 212, + 651, + 787, + 696 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The impact of the pooling mechanism. We propose with CLIP-DINOiser to combine MaskCLIP features with a well-defined linear combination and compare different solutions in Tab. 2a. In [67], the authors proposed to refine the predictions with a combination weighted by CLIP key embeddings (noted 'CLIP keys (preds.)' in the table) and boost MaskCLIP results by more than $+8$ mIoU on VOC and VOC20, $+1.8$ and $+1.0$ and $+0.6$ mIoU on the other datasets. However, we show that working directly at the feature level allows us to achieve better results; we obtain consistent improvements ranging from $+6$ to $+19$ mIoU on", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "CLIP-DINOiser", + "bbox": [ + 627, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "all datasets when using DINO-based weight $A^{\\xi}$ and further improve when using trained CLIP-based weights $A^{\\phi}$ .", + "bbox": [ + 215, + 145, + 787, + 176 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/ed5c28b969d3133db29603925b8e41c3627b72e40a0f765f0b45961e6333668f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Pooling strategyVOCVOC20C59StuffADE
MaskCLIP [67] - baseline32.961.825.617.614.3
CLIP keys (preds.) [67]41.371.927.418.614.9
ours w. CLIP keys39.273.223.012.67.7
ours w. DINO Aξ53.779.135.524.720.4
ours w. trained Aφ54.080.935.924.620.0
", + "bbox": [ + 218, + 205, + 522, + 271 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/dc7051ededb486a1ab8910ff4db35a1ced00037aa8bb333e9e030e5a35f6a254.jpg", + "table_caption": [ + "(a) Pooling strategy" + ], + "table_footnote": [], + "table_body": "
PoolingBkg det.ObjectVOC
MaskCLIP [67] - baseline16.432.9
ours w. DINO Aξ29.953.7
ours w. DINO AξFOUND32.160.1
ours w. DINO Aξours w. M34.162.1
ours w. DINO Aξours w. Mφ34.261.9
ours w. trained Aφours w. Mφ34.862.1
", + "bbox": [ + 531, + 203, + 784, + 276 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "(b) Background detection", + "bbox": [ + 573, + 277, + 741, + 289 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 2: Impact of the pooling strategy (a) and background detection (b) on diverse datasets reported with the mIoU metric.", + "bbox": [ + 215, + 300, + 785, + 330 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The impact of the background detection. We now discuss the improvement provided by our background refinement strategy, which is applied when stuff-like background patches need to be detected. We report such results in Tab. 2b when employing our pooling strategy (either using DINO features, noted 'w. DINO $A^{\\xi}$ or those extracted from CLIP, noted 'w. trained $A^{\\phi}$ ). When using solely 'FOUND' for background detection, as in [59], we improve by $+6.4$ mIoU on VOC (achieving $60.1\\mathrm{mIoU}$ ), but when relaxing FOUND (see Sec. 3.5) with an uncertainty condition, we boost scores up to 62.1 on VOC, showing the limitation of using FOUND alone. We also achieve similar results when using CLIP-based predictions $M^{\\phi}$ both with DINO-based $A^{\\xi}$ and trained CLIP-based $A^{\\phi}$ correlations, although we observe that best results are overall obtained with trained $A^{\\phi}$ . We visualize CLIP-based mask $M^{\\phi}$ in Fig. 5a and see high similarity to DINO-based predictions, therefore showing the localization quality of CLIP.", + "bbox": [ + 215, + 357, + 787, + 554 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 583, + 369, + 599 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we propose to make the most out of CLIP features and show that the features already contain useful localization information. Indeed with light convolutional layers, we are able to learn both good patch-correlation and objectness information by using DINO self-supervised model as a guide. With such information, our method CLIP-DINOiser performs zero-shot open-vocabulary semantic segmentation in a single pass of CLIP model and with two light extra convolutional layers. CLIP-DINOiser reaches state-of-the-art results on complex semantic segmentation datasets.", + "bbox": [ + 212, + 621, + 787, + 743 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. Despite yielding strong results on open-vocabulary semantic segmentation, CLIP-DINOiser is still bounded by the capability of the CLIP model to separate classes, as it inherits its granularity. We believe that better prompt engineering paired with better image-text models could further boost the performance of CLIP-DINOiser.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "M.Wysoczanska et al.", + "bbox": [ + 271, + 114, + 416, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the National Centre of Science (Poland) Grant No. 2022/45/B/ST6/02817 and by the grant from NVIDIA providing one RTX A5000 24GB used for this project.", + "bbox": [ + 215, + 175, + 784, + 220 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 243, + 321, + 258 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abdelreheem, A., Skorokhodov, I., Ovsjanikov, M., Wonka, P.: Satr: Zero-shot semantic segmentation of 3d shapes. In: ICCV (2023)", + "2. Bucher, M., Vu, T.H., Cord, M., Pérez, P.: Zero-shot semantic segmentation. In: NeurIPS (2019)", + "3. Caesar, H., Uijlings, J., Ferrari, V.: Coco-stuff: Thing and stuff classes in context. In: CVPR (2018)", + "4. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. In: NeurIPS (2020)", + "5. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021)", + "6. Cha, J., Mun, J., Roh, B.: Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs. In: CVPR (2023)", + "7. Chen, J., Zhu, D., Qian, G., Ghanem, B., Yan, Z., Zhu, C., Xiao, F., Culatana, S.C., Elhoseiny, M.: Exploring open-vocabulary semantic segmentation from clip vision encoder distillation only. In: ICCV (2023)", + "8. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: Clip2scene: Towards label-efficient 3d scene understanding by clip. In: CVPR (2023)", + "9. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)", + "10. Contributors, M.: MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark (2020)", + "1. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: CVPR (2016)", + "2. Darcet, T., Oquab, M., Mairal, J., Bojanowski, P.: Vision transformers need registers. arXiv preprint arXiv:2309.16588 (2023)", + "3. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009)", + "4. Ding, J., Xue, N., Xia, G.S., Dai, D.: Decoupling zero-shot semantic segmentation. In: CVPR (2022)", + "5. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)", + "6. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results", + "7. Fang, A., Ilharco, G., Wortsman, M., Wan, Y., Shankar, V., Dave, A., Schmidt, L.: Data determines distributional robustness in contrastive language image pretraining (CLIP). In: ICML (2022)", + "8. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: ECCV (2022)" + ], + "bbox": [ + 225, + 273, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "CLIP-DINOiser", + "bbox": [ + 627, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: CVPR (2023)", + "20. Gu, Z., Zhou, S., Niu, L., Zhao, Z., Zhang, L.: Context-aware feature generation for zero-shot semantic segmentation. In: ACM MM (2020)", + "21. Hu, P., Sclaroff, S., Saenko, K.: Uncertainty-aware learning for zero-shot semantic segmentation. In: NeurIPS (2020)", + "22. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021). https://doi.org/10.5281/zenodo.5143773, https://doi.org/10.5281/zenodo.5143773", + "23. Jatavallabhula, K.M., Kuwajerwala, A., Gu, Q., Omama, M., Chen, T., Li, S., Iyer, G., Saryazdi, S., Keetha, N., Tewari, A., et al.: Conceptfusion: Open-set multimodal 3d mapping. In: RSS (2023)", + "24. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021)", + "25. Karazija, L., Laina, I., Vedaldi, A., Rupprecht, C.: Diffusion models for zero-shot open-vocabulary segmentation. arXiv preprint arXiv:2306.09316 (2023)", + "26. Kato, N., Yamasaki, T., Aizawa, K.: Zero-shot semantic segmentation via variational mapping. In: ICCVW (2019)", + "27. Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: Lerf: Language embedded radiance fields. In: ICCV (2023)", + "28. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015)", + "29. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: ICLR (2022)", + "30. Li, P., Wei, Y., Yang, Y.: Consistent structural relation learning for zero-shot segmentation. NeurIPS (2020)", + "31. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: CVPR (2023)", + "32. Liu, Q., Wen, Y., Han, J., Xu, C., Xu, H., Liang, X.: Open-world semantic segmentation via contrasting and clustering vision-language embedding. In: ECCV (2022)", + "33. Luo, H., Bao, J., Wu, Y., He, X., Li, T.: SegCLIP: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. In: ICML (2023)", + "34. Mayilvahanan, P., Wiedemer, T., Rusak, E., Bethge, M., Brendel, W.: Does CLIP's generalization performance mainly stem from high train-test similarity? arXiv preprint arXiv:2310.09562 (2023)", + "35. Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word representations in vector space. In: ICLR (2013)", + "36. Mottaghi, R., Chen, X., Liu, X., Cho, N.G., Lee, S.W., Fidler, S., Urtasun, R., Yuille, A.: The role of context for object detection and semantic segmentation in the wild. In: CVPR (2014)", + "37. Mukhoti, J., Lin, T.Y., Poursaeed, O., Wang, R., Shah, A., Torr, P.H., Lim, S.N.: Open vocabulary semantic segmentation with patch aligned contrastive learning. In: CVPR (2023)", + "38. Najibi, M., Ji, J., Zhou, Y., Qi, C.R., Yan, X., Ettinger, S., Anguelov, D.: Unsupervised 3d perception with 2d vision-language distillation for autonomous driving. In: ICCV (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "M.Wysoczanska et al.", + "bbox": [ + 271, + 114, + 415, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "39. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., HAZIZA, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: DINOv2: Learning robust visual features without supervision. TMLR (2024)", + "40. Pastore, G., Cermelli, F., Xian, Y., Mancini, M., Akata, Z., Caputo, B.: A closer look at self-training for zero-label semantic segmentation. In: CVPR (2021)", + "41. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: Openscene: 3d scene understanding with open vocabularies. In: CVPR (2023)", + "42. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: EMNLP (2014)", + "43. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)", + "44. Ranasinghe, K., McKinzie, B., Ravi, S., Yang, Y., Toshev, A., Shlens, J.: Perceptual grouping in contrastive vision-language models. In: ICCV (2023)", + "45. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: Denseclip: Language-guided dense prediction with context-aware prompting. In: CVPR (2022)", + "46. Ren, P., Li, C., Xu, H., Zhu, Y., Wang, G., Liu, J., Chang, X., Liang, X.: Viewco: Discovering text-supervised segmentation masks via multi-view semantic consistency. In: ICLR (2023)", + "47. Rewatbowornwong, P., Chatthee, N., Chuangsuwanich, E., Suwajanakorn, S.: Zero-guidance segmentation using zero segment labels. In: ICCV (2023)", + "48. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)", + "49. Shin, G., Xie, W., Albanie, S.: Reco: Retrieve and co-segment for zero-shot transfer. In: NeurIPS (2022)", + "50. Shin, G., Xie, W., Albanie, S.: Namedmask: Distilling segmenters from complementary foundation models. In: CVPRW (2023)", + "51. Simeoni, O., Puy, G., Vo, H.V., Roburin, S., Gidaris, S., Bursuc, A., Pérez, P., Marlet, R., Ponce, J.: Localizing objects with self-supervised transformers and no labels. In: BMVC (2021)", + "52. Simeoni, O., Sekkat, C., Puy, G., Vobecky, A., Zablocki, É., Pérez, P.: Unsupervised object localization: Observing the background to discover objects. In: CVPR (2023)", + "53. Simeoni, O., Zablocki, E., Gidaris, S., Puy, G., Pérez, P.: Unsupervised object localization in the era of self-supervised vits: A survey. arXiv preprint arXiv:2310.12904 (2023)", + "54. Vobecký, A., Simeoni, O., Hurych, D., Gidaris, S., Bursuc, A., Perez, P., Sivic, J.: Pop-3d: Open-vocabulary 3d occupancy prediction from images. In: NeurIPS (2023)", + "55. Walmer, M., Suri, S., Gupta, K., Shrivastava, A.: Teaching matters: Investigating the role of supervision in vision transformers. In: CVPR (2023)", + "56. Wang, X., Girdhar, R., Yu, S.X., Misra, I.: Cut and learn for unsupervised object detection and instance segmentation. In: CVPR (2023)", + "57. Wang, Y., Shen, X., Hu, S.X., Yuan, Y., Crowley, J.L., Vaufreydaz, D.: Self-supervised transformers for unsupervised object discovery using normalized cut. In: CVPR (2022)", + "58. Wu, J., Li, X., Xu, S., Yuan, H., Ding, H., Yang, Y., Li, X., Zhang, J., Tong, Y., Jiang, X., et al.: Towards open vocabulary learning: A survey. T-PAMI (2024)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "CLIP-DINOiser", + "bbox": [ + 627, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "59. Wysoczanska, M., Ramamonjisoa, M., Trzcinski, T., Simeoni, O.: Clip-diy: Clip dense inference yields open-vocabulary semantic segmentation for-free. In: WACV (2024)", + "60. Xian, Y., Choudhury, S., He, Y., Schiele, B., Akata, Z.: Semantic projection network for zero-and few-label semantic segmentation. In: CVPR (2019)", + "61. Xu, J., De Mello, S., Liu, S., Byeon, W., Breuel, T., Kautz, J., Wang, X.: Groupvit: Semantic segmentation emerges from text supervision. arXiv preprint arXiv:2202.11094 (2022)", + "62. Xu, J., Hou, J., Zhang, Y., Feng, R., Wang, Y., Qiao, Y., Xie, W.: Learning open-vocabulary semantic segmentation models from natural language supervision. In: CVPR (2023)", + "63. Zhai, X., Mustafa, B., Kolesnikov, A., Beyer, L.: Sigmoid loss for language image pre-training. In: ICCV (2023)", + "64. Zhao, H., Puig, X., Zhou, B., Fidler, S., Torralba, A.: Open vocabulary scene parsing. In: ICCV (2017)", + "65. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: CVPR (2022)", + "66. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. IJCV (2019)", + "67. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: ECCV (2022)", + "68. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: iBOT: Image bert pre-training with online tokenizer. In: ICLR (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 465 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "M.Wysoczanska et al.", + "bbox": [ + 271, + 114, + 415, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/CLIP-DINOiser_ Teaching CLIP a few DINO tricks for open-vocabulary semantic segmentation/ab181842-dc49-431d-ab8a-9cea7505e7b5_model.json b/2024/CLIP-DINOiser_ Teaching CLIP a few DINO tricks for open-vocabulary semantic segmentation/ab181842-dc49-431d-ab8a-9cea7505e7b5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3aab40cf69e6444365bfc2da12777f2cff23311e --- /dev/null +++ b/2024/CLIP-DINOiser_ Teaching CLIP a few DINO tricks for open-vocabulary semantic segmentation/ab181842-dc49-431d-ab8a-9cea7505e7b5_model.json @@ -0,0 +1,2293 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.221, + 0.141, + 0.783, + 0.187 + ], + "angle": 0, + "content": "CLIP-DINOiser: Teaching CLIP a few DINO tricks for open-vocabulary semantic segmentation" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.213, + 0.756, + 0.245 + ], + "angle": 0, + "content": "Monika Wysoczanska1, Oriane Simeoni2, Michael Ramamonjisoa3*, Andrei Bursuc2, Tomasz Trzciński1,4,5, and Patrick Pérez2" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.255, + 0.78, + 0.271 + ], + "angle": 0, + "content": "1Warsaw University of Technology, 2valeo.ai, 3Meta AI, 4Tooploox, 5IDEAS NCBR" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.311, + 0.744, + 0.617 + ], + "angle": 0, + "content": "Abstract. The popular CLIP model displays impressive zero-shot capabilities thanks to its seamless interaction with arbitrary text prompts. However, its lack of spatial awareness makes it unsuitable for dense computer vision tasks, e.g., semantic segmentation, without an additional fine-tuning step that often uses annotations and can potentially suppress its original open-vocabulary properties. Meanwhile, self-supervised representation methods have demonstrated good localization properties without human-made annotations nor explicit supervision. In this work, we take the best of both worlds and propose an open-vocabulary semantic segmentation method, which does not require any annotations. We propose to locally improve dense MaskCLIP features, which are computed with a simple modification of CLIP's last pooling layer, by integrating localization priors extracted from self-supervised features. By doing so, we greatly improve the performance of MaskCLIP and produce smooth outputs. Moreover, we show that the used self-supervised feature properties can directly be learnt from CLIP features. Our method CLIP-DINOiser needs only a single forward pass of CLIP and two light convolutional layers at inference, no extra supervision nor extra memory and reaches state-of-the-art results on challenging and fine-grained benchmarks such as COCO, Pascal Context, Cityscapes and ADE20k. The code to reproduce our results is available at https://github.com/wysoczanska/clip_dinoiser." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.63, + 0.748, + 0.658 + ], + "angle": 0, + "content": "Keywords: open-vocabulary semantic segmentation \\(\\cdot\\) self-supervised features \\(\\cdot\\) annotation-free segmentation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.691, + 0.378, + 0.708 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.725, + 0.793, + 0.803 + ], + "angle": 0, + "content": "Semantic segmentation is a key visual perception task for many real-world systems, e.g., self-driving cars, and industrial robots. Typically tackled in a dataset-oriented manner, best methods require a training dataset which is manually annotated for a specific and finite set of classes. The advent of powerful Vision-Language Models (VLM) [24,43,63] is stimulating a shift from a closed-vocabulary" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.812, + 0.787, + 0.839 + ], + "angle": 0, + "content": "* Work done outside of Meta and Meta was not involved in the research discussed here." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.417, + 0.128 + ], + "angle": 0, + "content": "M.Wysoczanska et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.79, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.37, + 0.789, + 0.496 + ], + "angle": 0, + "content": "Fig. 1: Examples of open-vocabulary semantic segmentation results obtained with our method CLIP-DINOiser on 'in-the-wild' images vs. those of MaskCLIP [67]. Our method improves MaskCLIP features with a smart pooling strategy which does not alter the original open-vocabulary properties. We use self-supervised DINO [5] as a guide to teach CLIP [22] to produce DINO-like localization features through two light convolutional layers. Our method, which achieves state-of-the-art results, only requires a single forward pass through CLIP model and our two layers. In addition to the correct prompts (light grey row) we list the irrelevant prompts predicted (in yellow) that we query in all images shown here." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.789, + 0.701 + ], + "angle": 0, + "content": "paradigm to an open-world one. Such models are trained with a simple but scalable objective: to align pairs of image and coarse text captions that can be obtained in large amounts with limited manual supervision. VLMs excel at associating global image content with arbitrary text inputs with remarkable generalization capabilities [17, 34], but struggle to provide dense open-vocabulary features [18, 67]. Obtaining such an alignment between pixels and language can lead to open-vocabulary extensions for multiple other modalities, such as point clouds [8, 23, 38, 41], 3D scenes [54], 3D shapes [1], radiance fields [27], intermodality alignment [19, 23], with multiple potential applications for which the construction of training datasets is even more challenging and where CLIP-derived models show promising results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Different strategies have been recently proposed towards improving CLIP's patch-level feature extraction abilities by modifying the original CLIP architecture for dense pooling and retraining [6, 37, 44, 61, 62] or finetuning on an annotated segmentation dataset with pre-defined classes [32, 67]. The former requires long training and/or large collections of annotated data, while the latter leads to an alteration of the vision-language associations of the CLIP features. An alternative line of approaches freezes the CLIP encoder and directly densifies its features with different heuristics, often with multiple forward passes [1, 23, 27, 49, 50, 59], but are less practical due to the extensive computa" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.628, + 0.116, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLIP-DINOiser" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.221 + ], + "angle": 0, + "content": "tional overhead. MaskCLIP [67] arises as a computationally efficient dense CLIP extractor. It converts CLIP's global self-attention layer into a convolutional one to produce patch features with original vision-language qualities. If such features are local, they appear to be too noisy for high-quality segmentation mask extraction (see Fig. 2b middle column)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.329 + ], + "angle": 0, + "content": "Meanwhile, recent self-supervised learning (SSL) approaches [4, 5, 9, 68] produce strong visual representations displaying object localization properties, and such without requiring any manual annotation. DINO [5] stands out with its semantically meaningful features which have been exploited for unsupervised object discovery [51, 52, 56, 57]. DINO features prove useful also for zero-shot semantic segmentation [25, 27, 59], but require expensive sliding window sampling [27, 59] or building concept-specific prototypes and ensemble strategies [25]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.787, + 0.54 + ], + "angle": 0, + "content": "In this work, we aim for unaltered patch-level CLIP features with minimal runtime overhead. To this end, we re-examine the localization properties of MaskCLIP features and observe that it is possible to easily refine them with guidance from SSL models. In detail, we train a simple convolutional layer on unlabeled data to produce pooling weights to perform correlation-guided dense feature pooling from CLIP without distorting the vision-language alignment. This layer is optimized to mimic the patch correlations of DINO [5] that indicate likely layouts of visual concepts in the images. Furthermore, we show that the unsupervised objectness information given by FOUND [52] from DINO features can be also directly learned from CLIP features again in a fully-unsupervised fashion with a single convolutional layer and helps improve the segmentation of the ill-defined 'background' prompt. With CLIP-DINOiser, we obtain high-quality masks in a single forward pass on CLIP (see Fig. 1). CLIP-DINOiser is amenable to producing dense semantic maps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.541, + 0.788, + 0.722 + ], + "angle": 0, + "content": "To summarize, our contributions are: (1) We propose a light pooling mechanism to refine MaskCLIP features by leveraging guidance from SSL features without degrading its original open-vocabulary properties. CLIP-DINOiser does not require any annotations, nor retraining CLIP from scratch, but only a single CLIP forward pass. (2) We show that CLIP already contains good localization properties which can be exploited. We leverage simple convolutional layers to emphasize visual concept layouts from dense CLIP features. We train them without any annotation on only 1k of raw images randomly sampled in ImageNet [13]. We believe that this finding could be further exploited in different contexts. (3) Our method achieves state-of-the-art results on complex semantic segmentation datasets such as COCO [3], Pascal Context [16], Cityscapes [11] and ADE20K [66]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.746, + 0.388, + 0.763 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Zero-shot semantic segmentation. This task has been typically approached by methods which aim at generalizing from seen classes to unseen ones [2, 20, 21, 26, 30, 40, 60, 64]. Such strategies train models with full supervision on the set of seen classes and propose different solutions to extend them to unseen ones with" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.416, + 0.128 + ], + "angle": 0, + "content": "M.Wysoczanska et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "out new images (labeled or unlabeled), e.g., by exploiting class information and relationships encapsulated in popular word embeddings [35,42]. While they produce fine segmentations without computational overhead, these methods require pixel-level annotations for the seen classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.221, + 0.789, + 0.479 + ], + "angle": 0, + "content": "From CLIP to open-vocabulary segmentation. The surge of VLMs with aligned image-language representations [22, 24, 43] brought back into the spotlight the zero-shot classification task. However, the extension to zero-shot segmentation is not obvious as the CLIP architecture is not equipped to yield dense vision-language features [18, 67]. To produce dense CLIP features, several approaches fine-tune or train from scratch pixel-aligned CLIP-like models with additional modules, mechanisms or supervision objectives [6, 37, 44, 61, 62] on datasets with annotations of varying granularity and quality: dense annotations [29, 31], class-agnostic object masks [14, 18, 45], coarse captions [6, 18, 31–33, 37, 44, 61, 62, 65] or pseudo-labels [67]. Recent works leverage image-level captions to align text to regions (obtained without supervision): PACL [37] trains an embedder module to learn patch-to-text affinity, TCL [6] proposes a local contrastive objective to align well-selected patches to the text and ViewCO [46] leverages multi-view consistency. On the downside, such models require long training on millions of images or specific types of very costly annotations. Also, fine-tuning CLIP with a defined vocabulary is more computationally appealing [29, 31, 67], but alters the open-vocabulary properties of the features [23]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.79, + 0.584 + ], + "angle": 0, + "content": "Most related to us is a line of works that investigate how to directly densify CLIP features [1, 23, 27, 59, 67] to obtain per-patch CLIP features. Such densification can be performed by aggregating features from multiple views [1, 27] or from sliding windows [23, 59] at the extra-cost of multiple forward passes. MaskCLIP [67] drops the global pooling layer of CLIP and matches the projected features directly to text via a \\(1 \\times 1\\) convolution layer. By doing so they achieve dense predictions, however noisy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.79, + 0.706 + ], + "angle": 0, + "content": "With a concept-driven perspective, some methods [25,49,50] build codebooks of visual prototypes per concept, including negative prototypes [25], and then perform co-segmentation [49]. While such an approach yields good results, it is however at the cost of building expensive class-specific prototypes, therefore diverging from open-vocabulary scenarios. Instead, we aim to remain open to avoid retraining a model or building new expensive prototypes whenever a new concept is considered. To that end, we devise a dense CLIP-feature extraction method that preserves the open-vocabulary quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Leveraging self-supervised models & CLIP. Recent self-supervised ViTs [4,5,9,12, 68] have demonstrated features with good localization properties [51,52,56,57]. Such features have also been exploited in the context of open-vocabulary segmentation methods, e.g. for pre-training for the visual backbone [7,44,62], cosegmentation [49], clustering patches into masks [47], representing object prototypes [25]. Related to us is the recent CLIP-DIY [59] which computes patch-level representations from CLIP features from different image crops with guidance from an unsupervised saliency segmenter [52] FOUND. While we also leverage" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.628, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLIP-DINOiser" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "the latter, in contrast with CLIP-DIY which runs multiple forward passes to build their dense CLIP features, our method requires only a single forward pass of CLIP. Furthermore, our method mitigates the limits of FOUND in cluttered scenarios by integrating an uncertainty constraint. Finally, we leverage the informative patch correlation properties of DINO [5] and show that it is possible to teach CLIP to produce DINO-like features through light convolutional layers." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.262, + 0.332, + 0.279 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.295, + 0.789, + 0.43 + ], + "angle": 0, + "content": "We present in this section CLIP-DINOiser, a simple and efficient strategy to improve MaskCLIP using localization information extracted from CLIP—with a lightweight model trained to mimic some of DINO's properties. We first set the goal in Sec. 3.1 and present MaskCLIP [67] in Sec. 3.2. We then introduce our strategy which leverages self-supervised features localization information to consolidate MaskCLIP features in Sec. 3.3 and discuss how such localization information can directly be learnt from CLIP in Sec. 3.4 (we visualize both steps in Fig. 3). We also propose a way to improve the 'background' filtering in Sec. 3.5." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.454, + 0.422, + 0.468 + ], + "angle": 0, + "content": "3.1 Problem statement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.79, + 0.603 + ], + "angle": 0, + "content": "In this work, we aim to produce open-vocabulary1 semantic segmentation of an image. We consider an image \\( X \\in \\mathbb{R}^{H \\times W \\times 3} \\) which we split into a sequence of \\( N \\) patches of dimensions \\( P \\times P \\times 3 \\) with \\( P \\times P \\) the patch size and \\( N = \\lceil \\frac{H}{P} \\rceil \\cdot \\lceil \\frac{W}{P} \\rceil \\). A class token, noted CLS, is added to the input sequence and we feed the \\( N + 1 \\) patches to a ViT [15] model. We aim at producing dense visual features \\( F \\in \\mathbb{R}^{N \\times d} \\), with \\( d \\) the feature dimension, that can later be matched to any set of text inputs embedded in the same space. In particular, the goal is to produce a segmentation map per textual query." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.625, + 0.494, + 0.64 + ], + "angle": 0, + "content": "3.2 Preliminaries on MaskCLIP" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.65, + 0.789, + 0.803 + ], + "angle": 0, + "content": "Extracting dense open-vocabulary features. The popular CLIP [22] model pretrained on image/caption pairs produces good global image features, but was not trained to generate high-quality 2D feature maps. In order to extract such dense feature maps relevant to semantic segmentation, Zhou et al. [67] revisit the global attention pooling layer of the last attention layer of the model. The authors discard the query and key embeddings of the layer and transform both the value projection and the last linear layer into a conv \\(1 \\times 1\\) layer. With this new model, named MaskCLIP and denoted \\(\\phi(\\cdot)\\), we extract \\(d\\)-dimensional features \\(\\phi^L(X) \\in \\mathbb{R}^{N \\times d}\\) from the last layer \\(L\\) which retains most of the open-vocabulary properties of CLIP [67]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.842 + ], + "angle": 0, + "content": "1 We adopt the taxonomy defined in the recent survey [58] and define our method as 'open-vocabulary', with capabilities to generalize to unseen datasets." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.418, + 0.128 + ], + "angle": 0, + "content": "M.Wysoczanska et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "Semantic segmentation given textual queries. We also extract CLIP textual features \\(\\phi_T(t_j)\\) for each text query \\(t_j \\in \\mathcal{T}\\) with \\(j \\in \\{1, \\dots, |\\mathcal{T}|\\}\\). Segmentation maps are then generated by computing the cosine similarity between each of the visual patch features and of the textual prompts, after L2-normalization. The most similar prompt is assigned to each patch. Note that a query 'background' can be added in order to obtain negative patches. Using MaskCLIP allows us to produce dense segmentation maps with a single forward pass of the classic CLIP model, but its outputs are noisy, as visible in Fig. 2b (middle column)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.281, + 0.566, + 0.297 + ], + "angle": 0, + "content": "3.3 DINOising open-vocabulary features" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.302, + 0.787, + 0.349 + ], + "angle": 0, + "content": "In this work, we aim to improve MaskCLIP's open-vocabulary features described above. To do so, we propose to leverage the known good localization properties of self-supervised features [5,39,51-53,57]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.789, + 0.526 + ], + "angle": 0, + "content": "Extracting self-supervised correlation information. Recent works [51, 57] have shown that the patch correlation information of the embeddings from the last attention layer of the self-supervised model, DINO [5] can help highlight objects in images. We use here the value embeddings which we observe have finer correlation than those of key and query (more discussion in supplementary material). We extract such self-supervised features \\(\\xi(X) \\in \\mathbb{R}^{N \\times d_{\\xi}}\\) and discard the CLS token. We then compute the per-patch cosine-similarity and produce the affinity map \\(A^{\\xi} \\in [-1,1]^{N \\times N}\\). We compare in Fig. 4 the patch-similarities obtained for a patch seed with MaskCLIP and DINO features and observe that the self-supervised features are more densely and accurately correlated than those of CLIP." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.545, + 0.507, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.699, + 0.436, + 0.711 + ], + "angle": 0, + "content": "(a) Our guided pooling" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.553, + 0.782, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.699, + 0.727, + 0.711 + ], + "angle": 0, + "content": "(b) Impact of the pooling" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.722, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Fig. 2: We present in (a) our guided pooling strategy defined in Eq. (1). The \\( N \\times N \\) affinity matrix is computed from patch features and is used to refine MaskCLIP features (bottom left). In (b) we compare our results with \\( F^{+} \\) (right) versus those obtained with MaskCLIP features (middle)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.81, + 0.842 + ], + "angle": 0, + "content": "Strengthening features with guided pooling. In order to locally consolidate MaskCLIP features \\(\\phi^L (X)\\), now noted \\(F\\), we propose to perform a concept-aware linear combination of the features per patch with guidance from the patch affinity \\(A^{\\xi}\\). The" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.628, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLIP-DINOiser" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "feature combination strategy can be seen as a form of voting mechanism that enforces similar patches to have similar CLIP features (and prediction) while attenuating noisy features. Specifically, we compute the new features \\( F^{+} \\in \\mathbb{R}^{N \\times d} \\) as an average of MaskCLIP features \\( F \\) weighted by \\( A^{\\xi} \\), presented in Fig. 2a. We zero-out \\( A^{\\xi} \\) correlations below a threshold \\( \\gamma \\), following [51, 57], and compute the new features for patch \\( p \\in \\{1, \\dots, N\\} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.39, + 0.245, + 0.786, + 0.287 + ], + "angle": 0, + "content": "\\[\nF _ {p} ^ {+} = \\frac {1}{\\sum_ {q = 1} ^ {N} A _ {p , q} ^ {\\xi}} \\sum_ {q = 1} ^ {N} A _ {p, q} ^ {\\xi} \\cdot F _ {q}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.296, + 0.788, + 0.372 + ], + "angle": 0, + "content": "We then produce the segmentation maps \\( S \\in [-1,1]^{N \\times |\\mathcal{T}|} \\), by comparing the new features \\( F^+ \\) to each textual queries in \\( \\mathcal{T} \\). As shown in Fig. 2b, when using such consolidated features, we obtain more accurate outputs and the high-frequency predictions observed in MaskCLIP are smoothed out, showing the benefit of the pooling." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.391, + 0.706, + 0.407 + ], + "angle": 0, + "content": "3.4 Teaching CLIP a first DINO trick: object correlations" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.439, + 0.785, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.657, + 0.788, + 0.782 + ], + "angle": 0, + "content": "Fig. 3: Overview of CLIP-DINOiser which leverages the quality of self-supervised features to improve the notoriously noisy MaskCLIP feature maps. We use DINO as a teacher which 'teaches' CLIP how to extract localization information. We train (left) a conv \\(3 \\times 3\\) layer to reproduce the patch correlations obtained with DINO. At inference (right), an input image is forwarded through the frozen CLIP image backbone and MaskCLIP projection. The produced features are then improved with our pooling strategy which is guided by correlations predicted with the trained convolutional layer applied on CLIP. With this light 'DINOising' process, we obtain 'DINOised' features which are matched against the prompts features to produce CLIP-DINOiser outputs." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We have shown in the previous section that self-supervised correlation information can successfully be used to improve the dense quality of open-vocabulary" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.418, + 0.128 + ], + "angle": 0, + "content": "M.Wysoczanska et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "features. If the difficulty of densifying CLIP is well-known, we show here that CLIP features already contain good localization information which can be extracted with a light model. We indeed predict DINO correlations \\( A^{\\xi} \\) from CLIP with a single convolutional layer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.788, + 0.269 + ], + "angle": 0, + "content": "In order to predict the DINO affinity map \\( A^{\\xi} \\) from CLIP features, we train a single \\( 3 \\times 3 \\) convolutional layer \\( g(\\cdot): \\mathbb{R}^{d} \\to \\mathbb{R}^{d_{g}} \\) which projects intermediate features \\( \\phi^{l}(X) \\)-extracted from layer \\( l \\)-into a smaller space of dimension \\( d_{g} < d \\). We enforce the patch correlations of the generated features \\( A^{\\phi} \\in [-1,1]^{N \\times N} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.362, + 0.279, + 0.786, + 0.318 + ], + "angle": 0, + "content": "\\[\nA ^ {\\phi} = \\frac {g \\left(\\phi^ {l} (X)\\right)}{\\| g \\left(\\phi^ {l} (X)\\right) \\|} \\otimes \\left(\\frac {g \\left(\\phi^ {l} (X)\\right)}{\\| g \\left(\\phi^ {l} (X)\\right) \\|}\\right) ^ {\\top}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.326, + 0.787, + 0.371 + ], + "angle": 0, + "content": "with \\(\\otimes\\) denoting the outer product, to be close to the binarized correlations \\(D = A^{\\xi} > \\gamma\\) (we use here the same \\(\\gamma\\) as defined above), using the binary cross-entropy loss \\(\\mathcal{L}^c\\):" + }, + { + "type": "equation", + "bbox": [ + 0.337, + 0.382, + 0.786, + 0.425 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {c} = \\sum_ {p = 1} ^ {N} \\left[ D _ {p} \\log A _ {p} ^ {\\phi} + (1 - D _ {p}) \\log \\left(1 - A _ {p} ^ {\\phi}\\right) \\right]. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.788, + 0.572 + ], + "angle": 0, + "content": "We present our layer training in Fig. 3 (left part) and observe the quality of CLIP-predicted affinity matrix \\( A^{\\phi} \\). We also show in Fig. 4 another example of obtained \\( A^{\\phi} \\) and observe their similarity to DINO-based correlations. We use the CLIP-produced correlations \\( A^{\\phi} \\) to replace \\( A^{\\xi} \\) in Eq. (1) to weight the pooling and observe a similar boost over MaskCLIP, thus showing that good patch correlations can indeed be extracted directly from CLIP. We can now discard DINO and we name CLIP-DINOiser the guided-pooling strategy which uses CLIP-based correlation. As shown in Fig. 3 (inference step), our method runs with a single forward pass of CLIP model and a small extra layer." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.591, + 0.782, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.664, + 0.788, + 0.707 + ], + "angle": 0, + "content": "Fig. 4: Comparison of the affinity maps between a seed (one on the 'plant' and the other on a 'pillow') and the other patch features when using features of MaskCLIP, DINO and ours after training." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.74, + 0.742, + 0.756 + ], + "angle": 0, + "content": "3.5 Teaching CLIP a second DINO trick: background filtering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Moreover, as discussed earlier, a 'background' query may be added to the set of textual queries \\(\\mathcal{T}\\) in order to help filter out patches falling in the background and not corresponding to any objects. We do not assume here any prior knowledge about classes of interest and focus rather on the foreground/background paradigm [52]. We argue that relying solely on the textual prompt 'background'" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.628, + 0.116, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLIP-DINOiser" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.171, + 0.468, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.313, + 0.462, + 0.325 + ], + "angle": 0, + "content": "(a) Comparison of background filtering" + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.145, + 0.768, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.313, + 0.706, + 0.325 + ], + "angle": 0, + "content": "(b) Background filtering" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.336, + 0.788, + 0.406 + ], + "angle": 0, + "content": "Fig. 5: We present in (a) a comparison of objectness mask generated by FOUND [52] and with our layer using CLIP features. We carefully define the fusion operation and the simple training strategy of the conv \\(1 \\times 1\\) again using DINO as a teacher in Sec. 3.5. In (b) is an overview of our background filtering which is applied when a 'background' prompt is provided and helps reduce hallucinations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.618 + ], + "angle": 0, + "content": "to catch all non-salient patches is underperforming and, similarly to [59], we propose to use a very light-weight unsupervised foreground/background segmentation method, namely FOUND [52] which also relies on DINO self-supervised features. We run FOUND on the entire image and extract a prediction mask \\( M \\in \\{0,1\\}^N \\) in which a patch is assigned the value 1 if falling into the foreground and 0 otherwise. We also observe that saliencies produced by FOUND can be too restrictive and discard objects which are partially visible or in a clutter. To mitigate this behaviour, we propose to relax the background selection by integrating an additional uncertainty constraint. To this end, we fuse the background information from both modalities by assigning the 'background' prompt to patches \\( p \\) which are both uncertain, e.g. have low confidence score \\( \\sigma(S)_p < \\delta \\), with \\( \\sigma(\\cdot) \\) the softmax operation, and which fall in the background in \\( M \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.706 + ], + "angle": 0, + "content": "Learning FOUND objectness. Moreover, we are also able to learn the predictions of FOUND [52] directly from CLIP features. To do so, we train a single \\(1 \\times 1\\) convolutional layer \\(h(\\cdot): \\mathbb{R}^d \\to \\mathbb{R}\\) which predicts from the features \\(\\phi^l(X)\\) an objectness map \\(M^\\phi = h(\\phi^l(X)) \\in \\mathbb{R}^N\\). We train the model to predict the FOUND binary mask \\(M\\) with the binary cross-entropy loss \\(\\mathcal{L}^m\\):" + }, + { + "type": "equation", + "bbox": [ + 0.321, + 0.714, + 0.68, + 0.756 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {m} = \\sum_ {p = 1} ^ {N} \\left[ M _ {p} \\log \\left(M _ {p} ^ {\\phi}\\right) + \\left(1 - M _ {p}\\right) \\log \\left(1 - M _ {p} ^ {\\phi}\\right) \\right].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "We show examples of predicted CLIP-based objectness in Fig. 5a and observe their very high similarity to those produced with DINO. Moreover, we can now replace \\(M\\) defined above with the binarized CLIP-based scores \\(\\zeta(M^{\\phi}) > 0.5\\) with \\(\\zeta(\\cdot)\\) the sigmoid operation, and observe a minimal drop in performances. We show an example of the background filtering with trained objectness in Fig. 5b." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.417, + 0.129 + ], + "angle": 0, + "content": "M.Wysoczanska et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.377, + 0.163 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.224 + ], + "angle": 0, + "content": "We detail in Sec. 4.1 the experimental setup used in our evaluation. We produce state-of-the-art results on the task of open-vocabulary semantic segmentation in Sec. 4.2 and ablation studies in Sec. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.248, + 0.426, + 0.264 + ], + "angle": 0, + "content": "4.1 Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.273, + 0.79, + 0.562 + ], + "angle": 0, + "content": "Technical details. We use in all experiments a frozen CLIP ViT-B/16 pre-trained following OpenCLIP [22]. Our method CLIP-DINOiser uses two convolutional layers to extract DINO-like information from CLIP layer \\( l = 10 \\) (the 3rd before the last which was shown to provide the best results [55]). The first layer \\( g(\\cdot) \\) has a kernel \\( 3 \\times 3 \\) and output dimension \\( d_{g} = 256 \\) and \\( h(\\cdot) \\) a kernel \\( 1 \\times 1 \\) with \\( d_{h} = 1 \\). The first is trained to match the correlation information extracted from the value embeddings of the last layer of a ViT-B/16 model trained following DINO [5]. The second layer is trained to replicate the unsupervised object localization predictions of FOUND [52]–which also uses DINO model. We train both layers with a binary cross-entropy loss on only 1k raw images randomly sampled from ImageNet [13] dataset without any annotation. We report average scores over 3 runs with different sampling seeds and provide standard deviations in the supplementary material. We follow [57] and binarize the correlations with \\( \\gamma = 0.2 \\). In the background filtering step, we use a high confidence score, i.e., \\( \\delta = 0.99 \\). We train our model for 6k iterations with a batch size of 16 images using Adam optimizer [28], which takes approximately 3 hours on a single NVIDIA RTX A5000 GPU. We decrease the learning rate for both heads by a factor of 0.1 after 5k iterations. We apply data augmentations during training (random scale and cropping, flipping and photometric distortions)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.571, + 0.79, + 0.753 + ], + "angle": 0, + "content": "Datasets and metric. We evaluate our method on eight benchmarks typically used for zero-shot semantic segmentation [6]. Following [6], we split them into two groups. The first consists in datasets with a 'background' query: PASCAL VOC [16] (noted 'VOC'), PASCAL Context [36] (noted 'Context'), and COCO Object [3] (noted 'Object') and the second without: PASCAL VOC20 [16] (noted 'VOC20'), PASCAL Context59 [36] (noted 'C59'), COCO-Stuff [3] (noted 'Stuff'), Cityscapes [11] (noted 'City'), and ADE20K [66] (noted 'ADE'). We evaluate results with the standard mIoU metric. We also follow the evaluation protocol of [6], use the implementations provided by MMSegmentation [10], employ a sliding window strategy, resize the input image to have a shorter side of 448. We also do not perform text expansions of the class names and use only the standard ImageNet prompts following [22, 61, 67]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Baselines. We compare our method against state-of-the-art methods on open-vocabulary zero-shot semantic segmentation. For a fair comparison between methods, we report results without any post-processing step. In our evaluations, we follow the taxonomy presented in [58] and compare our model with the methods relying on language-image pretraining, also called open-vocabulary." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.628, + 0.116, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLIP-DINOiser" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "We split the compared baselines into four categories: (1) dataset specific which employ pseudo-labeling and supervised training of a segmentation model on target dataset: NamedMask [50], MaskCLIP + [67]); (2) construct prototypes: ReCO [49], OVDiff [25]; (3) train with text supervision including GroupViT [61], ZeroSeg [47], SegCLIP [33], TCL [6], CLIPpy [44], OVSegmentor [62], which all require access to additional datasets of millions of image/caption pairs (we note in the table the exact datasets used for the training); and finally use frozen CLIP i.e. CLIP-DIY [59] and MaskCLIP [67], which use pre-trained CLIP. Our method falls into the last category as we do not modify CLIP, and do not need access to additional caption annotations as we use only 1k unannotated images." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.329, + 0.784, + 0.626 + ], + "angle": 0, + "content": "
MethodsConcept spec.Extra Inference data backboneNo background promptW/ bkg prompt Cont. Object VOC
VOC20C59StuffCityADE
Dataset specific
MaskCLIP+ [67]✓XIDLv2-31.118.0----
NamedMask [50]✓XIDLv3+-----27.759.2
Build prototypes per visual concept
ReCo [49]✓✓ICLIP57.822.314.821.111.219.915.7
OVDiff [25]✓✓XCLIP+ DINO+SD81.733.7--14.930.134.8
Text/image alignment training with captions
GroupViT [61] [6]XXITCLIP79.723.415.311.19.218.727.5
ZeroSeg [7]XXITCLIP-----21.822.1
SegCLIP [33]XXITCLIP---11.08.724.726.5
TCL [6]XXITCLIP77.530.319.623.114.924.330.4
CLIPpy [44]XXITCLIP----13.5-32.0
OVSegmentor [62]XXITCLIP----5.620.425.1
Frozen CLIP
CLIP-DIY [59]*X✓XCLIP+ DINO79.719.813.311.69.919.731.0
MaskCLIP [67] [6]X✓XCLIP53.723.314.721.610.821.115.5
MaskCLIP*X✓XCLIP61.825.617.625.014.322.916.4
MaskCLIP* †X✓XCLIP71.927.418.623.014.924.021.6
CLIP-DINOiserX✓I(1k)CLIP80.935.924.631.720.032.434.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.637, + 0.788, + 0.818 + ], + "angle": 0, + "content": "Table 1: Open-vocabulary semantic segmentation quantitative comparison using the mIoU metric. We separate in two columns the evaluation datasets: those without a 'background' prompt and those with (noted 'W/ bkg prompt'), as discussed in Sec. 4.1. We report all methods without post-processing. We note with * methods for which we computed scores; we obtained MaskCLIP* scores with OpenCLIP [22] and mark with † the use of MaskCLIP refinement. The first and second best methods are respectively bold and underlined. We specify if a method assumes prior access to names of concepts ('Concept spec.)' and if it employs a frozen backbone (''). We specify what additional data is used at training ('Extra data') ('I' stands for images and 'IT' for image/text aligned data). Our CLIP-DINOiser only needs 1k images from ImageNet to be trained. 'SD' stands for Stable Diffusion [48]. We refer to Sec. 4.1 for more details on baselines and we detail the datasets used for training by each method in the supplementary material." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.417, + 0.129 + ], + "angle": 0, + "content": "M.Wysoczanska et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.594, + 0.162 + ], + "angle": 0, + "content": "4.2 Open-vocabulary semantic segmentation" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.207 + ], + "angle": 0, + "content": "We discuss in this section state-of-the-art results on the task of open-vocabulary semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.227, + 0.787, + 0.408 + ], + "angle": 0, + "content": "Evaluation with no 'background' class. We first compare in Tab. 1 ('No background prompt' column) the results on datasets which aim at the segmentation of most of the pixels in an image and do not consider a 'background' class. We observe that our method CLIP-DINOiser achieves the best results on four datasets yielding \\(+2.2\\), \\(+5.0\\), \\(+6.7\\) and \\(+5.1\\) mIoU over the second best performing method. Interestingly, we outperform methods which build expensive prototypes per visual concept on fine-grained datasets, showing the benefit of our lightweight and generalizable method. The only drop \\((-0.8\\) mIoU) is seen on VOC20 with respect to OVDiff; we believe it is due to the benefit of generating per-concept negative prototypes which likely benefits this object-centric dataset. An adaptive granularity of feature correlation could help mitigate this drop, which we leave for future work." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.429, + 0.787, + 0.624 + ], + "angle": 0, + "content": "Evaluation with 'background' class. We now compare our method on datasets which include a 'background' query in Tab. 1 ('W/ bkg prompt' column). In this setup, we also apply our background detection mechanism (detailed in Sec. 3.5) on VOC and Object in order to improve the stuff-like background detection. We observe that CLIP-DINOiser significantly outperforms all methods which do not construct prototypes. Moreover, we surpass OVDiff (which uses an ensemble of three models) on Context dataset by \\(+2.3\\mathrm{mIoU}\\) and are on par on Object. It is to be noted that with a single feature extractor, the performance of OVDiff drops by \\(-10\\mathrm{mIoU}\\) and the method requires the construction of a 'background' prototype per concept, otherwise losing another \\(-10\\mathrm{mIoU}\\) on VOC. On the other hand, CLIP-DINOiser produces segmentation masks in a single pass of CLIP with the light addition of two convolutional layers while remaining fully open-vocabulary as it does not require any concept-specific constructs." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.644, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Qualitative results. We qualitatively compare in Fig. 6 CLIP-DINOiser with high-performing TCL [6], CLIP-DIY [59] (two recent methods which provide code) and our baseline method MaskCLIP [67] on images taken from the datasets considered in the evaluation. We observe that our method generates predictions accurate both in terms of localization and assignment. Indeed we obtain fined-grained results on the challenging datasets, e.g. in the Cityscapes example the text query 'car' and in the ADE20k example 'fountain' are accurately located when CLIP-DIY and TCL produce coarser results. Versus MaskCLIP, we can see the denoising capabilities of CLIP-DINOiser as MaskCLIP hallucinations grow with the number of text queries prompted at evaluation. Finally, in Fig. 1 we present 'in the wild' examples, beyond the evaluation benchmarks, and show that CLIP-DINOiser produces accurate segmentation masks for arbitrary and very specific prompts, such as 'wooden table' or 'leather bag'." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.628, + 0.116, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLIP-DINOiser" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.144, + 0.794, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.518, + 0.788, + 0.587 + ], + "angle": 0, + "content": "Fig.6: Qualitative open-vocabulary segmentation results. We compare ours against CLIP-DIY [59], TCL [6] and MaskCLIP [67]. For a fair comparison, we do not apply post-processing. All pixels annotated in black are from the background class. We observe that our method achieves more accurate results both in terms of localization and class assignment." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.621, + 0.388, + 0.636 + ], + "angle": 0, + "content": "4.3 Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.788, + 0.698 + ], + "angle": 0, + "content": "We now conduct an ablation study of the different components of CLIP-DINOiser and investigate the impact of both our feature pooling strategy and background detection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The impact of the pooling mechanism. We propose with CLIP-DINOiser to combine MaskCLIP features with a well-defined linear combination and compare different solutions in Tab. 2a. In [67], the authors proposed to refine the predictions with a combination weighted by CLIP key embeddings (noted 'CLIP keys (preds.)' in the table) and boost MaskCLIP results by more than \\(+8\\) mIoU on VOC and VOC20, \\(+1.8\\) and \\(+1.0\\) and \\(+0.6\\) mIoU on the other datasets. However, we show that working directly at the feature level allows us to achieve better results; we obtain consistent improvements ranging from \\(+6\\) to \\(+19\\) mIoU on" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.418, + 0.128 + ], + "angle": 0, + "content": "M.Wysoczanska et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.146, + 0.788, + 0.178 + ], + "angle": 0, + "content": "all datasets when using DINO-based weight \\( A^{\\xi} \\) and further improve when using trained CLIP-based weights \\( A^{\\phi} \\)." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.207, + 0.523, + 0.272 + ], + "angle": 0, + "content": "
Pooling strategyVOCVOC20C59StuffADE
MaskCLIP [67] - baseline32.961.825.617.614.3
CLIP keys (preds.) [67]41.371.927.418.614.9
ours w. CLIP keys39.273.223.012.67.7
ours w. DINO Aξ53.779.135.524.720.4
ours w. trained Aφ54.080.935.924.620.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.306, + 0.273, + 0.438, + 0.285 + ], + "angle": 0, + "content": "(a) Pooling strategy" + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.204, + 0.785, + 0.277 + ], + "angle": 0, + "content": "
PoolingBkg det.ObjectVOC
MaskCLIP [67] - baseline16.432.9
ours w. DINO Aξ29.953.7
ours w. DINO AξFOUND32.160.1
ours w. DINO Aξours w. M34.162.1
ours w. DINO Aξours w. Mφ34.261.9
ours w. trained Aφours w. Mφ34.862.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.575, + 0.278, + 0.743, + 0.29 + ], + "angle": 0, + "content": "(b) Background detection" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.301, + 0.787, + 0.331 + ], + "angle": 0, + "content": "Table 2: Impact of the pooling strategy (a) and background detection (b) on diverse datasets reported with the mIoU metric." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.358, + 0.789, + 0.555 + ], + "angle": 0, + "content": "The impact of the background detection. We now discuss the improvement provided by our background refinement strategy, which is applied when stuff-like background patches need to be detected. We report such results in Tab. 2b when employing our pooling strategy (either using DINO features, noted 'w. DINO \\(A^{\\xi}\\) or those extracted from CLIP, noted 'w. trained \\(A^{\\phi}\\) ). When using solely 'FOUND' for background detection, as in [59], we improve by \\(+6.4\\) mIoU on VOC (achieving \\(60.1\\mathrm{mIoU}\\)), but when relaxing FOUND (see Sec. 3.5) with an uncertainty condition, we boost scores up to 62.1 on VOC, showing the limitation of using FOUND alone. We also achieve similar results when using CLIP-based predictions \\(M^{\\phi}\\) both with DINO-based \\(A^{\\xi}\\) and trained CLIP-based \\(A^{\\phi}\\) correlations, although we observe that best results are overall obtained with trained \\(A^{\\phi}\\). We visualize CLIP-based mask \\(M^{\\phi}\\) in Fig. 5a and see high similarity to DINO-based predictions, therefore showing the localization quality of CLIP." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.584, + 0.37, + 0.6 + ], + "angle": 0, + "content": "5 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.622, + 0.788, + 0.744 + ], + "angle": 0, + "content": "In this work, we propose to make the most out of CLIP features and show that the features already contain useful localization information. Indeed with light convolutional layers, we are able to learn both good patch-correlation and objectness information by using DINO self-supervised model as a guide. With such information, our method CLIP-DINOiser performs zero-shot open-vocabulary semantic segmentation in a single pass of CLIP model and with two light extra convolutional layers. CLIP-DINOiser reaches state-of-the-art results on complex semantic segmentation datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Limitations. Despite yielding strong results on open-vocabulary semantic segmentation, CLIP-DINOiser is still bounded by the capability of the CLIP model to separate classes, as it inherits its granularity. We believe that better prompt engineering paired with better image-text models could further boost the performance of CLIP-DINOiser." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.628, + 0.116, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLIP-DINOiser" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.222 + ], + "angle": 0, + "content": "This work was supported by the National Centre of Science (Poland) Grant No. 2022/45/B/ST6/02817 and by the grant from NVIDIA providing one RTX A5000 24GB used for this project." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.244, + 0.323, + 0.26 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.785, + 0.304 + ], + "angle": 0, + "content": "1. Abdelreheem, A., Skorokhodov, I., Ovsjanikov, M., Wonka, P.: Satr: Zero-shot semantic segmentation of 3d shapes. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.304, + 0.785, + 0.33 + ], + "angle": 0, + "content": "2. Bucher, M., Vu, T.H., Cord, M., Pérez, P.: Zero-shot semantic segmentation. In: NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.785, + 0.358 + ], + "angle": 0, + "content": "3. Caesar, H., Uijlings, J., Ferrari, V.: Coco-stuff: Thing and stuff classes in context. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.359, + 0.785, + 0.385 + ], + "angle": 0, + "content": "4. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.386, + 0.785, + 0.413 + ], + "angle": 0, + "content": "5. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.413, + 0.785, + 0.441 + ], + "angle": 0, + "content": "6. Cha, J., Mun, J., Roh, B.: Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.441, + 0.785, + 0.482 + ], + "angle": 0, + "content": "7. Chen, J., Zhu, D., Qian, G., Ghanem, B., Yan, Z., Zhu, C., Xiao, F., Culatana, S.C., Elhoseiny, M.: Exploring open-vocabulary semantic segmentation from clip vision encoder distillation only. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.482, + 0.785, + 0.523 + ], + "angle": 0, + "content": "8. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: Clip2scene: Towards label-efficient 3d scene understanding by clip. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.524, + 0.785, + 0.551 + ], + "angle": 0, + "content": "9. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.551, + 0.785, + 0.578 + ], + "angle": 0, + "content": "10. Contributors, M.: MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.579, + 0.785, + 0.62 + ], + "angle": 0, + "content": "1. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.785, + 0.647 + ], + "angle": 0, + "content": "2. Darcet, T., Oquab, M., Mairal, J., Bojanowski, P.: Vision transformers need registers. arXiv preprint arXiv:2309.16588 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.675 + ], + "angle": 0, + "content": "3. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "4. Ding, J., Xue, N., Xia, G.S., Dai, D.: Decoupling zero-shot semantic segmentation. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.744 + ], + "angle": 0, + "content": "5. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "6. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.785, + 0.813 + ], + "angle": 0, + "content": "7. Fang, A., Ilharco, G., Wortsman, M., Wan, Y., Shankar, V., Dave, A., Schmidt, L.: Data determines distributional robustness in contrastive language image pretraining (CLIP). In: ICML (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "8. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: ECCV (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.274, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.416, + 0.128 + ], + "angle": 0, + "content": "M.Wysoczanska et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "19. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "20. Gu, Z., Zhou, S., Niu, L., Zhao, Z., Zhang, L.: Context-aware feature generation for zero-shot semantic segmentation. In: ACM MM (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "21. Hu, P., Sclaroff, S., Saenko, K.: Uncertainty-aware learning for zero-shot semantic segmentation. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.285 + ], + "angle": 0, + "content": "22. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021). https://doi.org/10.5281/zenodo.5143773, https://doi.org/10.5281/zenodo.5143773" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "23. Jatavallabhula, K.M., Kuwajerwala, A., Gu, Q., Omama, M., Chen, T., Li, S., Iyer, G., Saryazdi, S., Keetha, N., Tewari, A., et al.: Conceptfusion: Open-set multimodal 3d mapping. In: RSS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "24. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "25. Karazija, L., Laina, I., Vedaldi, A., Rupprecht, C.: Diffusion models for zero-shot open-vocabulary segmentation. arXiv preprint arXiv:2306.09316 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "26. Kato, N., Yamasaki, T., Aizawa, K.: Zero-shot semantic segmentation via variational mapping. In: ICCVW (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "27. Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: Lerf: Language embedded radiance fields. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "28. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.785, + 0.508 + ], + "angle": 0, + "content": "29. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "30. Li, P., Wei, Y., Yang, Y.: Consistent structural relation learning for zero-shot segmentation. NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.577 + ], + "angle": 0, + "content": "31. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.618 + ], + "angle": 0, + "content": "32. Liu, Q., Wen, Y., Han, J., Xu, C., Xu, H., Liang, X.: Open-world semantic segmentation via contrasting and clustering vision-language embedding. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.646 + ], + "angle": 0, + "content": "33. Luo, H., Bao, J., Wu, Y., He, X., Li, T.: SegCLIP: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "34. Mayilvahanan, P., Wiedemer, T., Rusak, E., Bethge, M., Brendel, W.: Does CLIP's generalization performance mainly stem from high train-test similarity? arXiv preprint arXiv:2310.09562 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.715 + ], + "angle": 0, + "content": "35. Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word representations in vector space. In: ICLR (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "36. Mottaghi, R., Chen, X., Liu, X., Cho, N.G., Lee, S.W., Fidler, S., Urtasun, R., Yuille, A.: The role of context for object detection and semantic segmentation in the wild. In: CVPR (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.798 + ], + "angle": 0, + "content": "37. Mukhoti, J., Lin, T.Y., Poursaeed, O., Wang, R., Shah, A., Torr, P.H., Lim, S.N.: Open vocabulary semantic segmentation with patch aligned contrastive learning. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "38. Najibi, M., Ji, J., Zhou, Y., Qi, C.R., Yan, X., Ettinger, S., Anguelov, D.: Unsupervised 3d perception with 2d vision-language distillation for autonomous driving. In: ICCV (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.628, + 0.116, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLIP-DINOiser" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.218 + ], + "angle": 0, + "content": "39. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., HAZIZA, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: DINOv2: Learning robust visual features without supervision. TMLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.245 + ], + "angle": 0, + "content": "40. Pastore, G., Cermelli, F., Xian, Y., Mancini, M., Akata, Z., Caputo, B.: A closer look at self-training for zero-label semantic segmentation. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.786, + 0.273 + ], + "angle": 0, + "content": "41. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: Openscene: 3d scene understanding with open vocabularies. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.786, + 0.3 + ], + "angle": 0, + "content": "42. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: EMNLP (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.786, + 0.342 + ], + "angle": 0, + "content": "43. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.786, + 0.37 + ], + "angle": 0, + "content": "44. Ranasinghe, K., McKinzie, B., Ravi, S., Yang, Y., Toshev, A., Shlens, J.: Perceptual grouping in contrastive vision-language models. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.786, + 0.411 + ], + "angle": 0, + "content": "45. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: Denseclip: Language-guided dense prediction with context-aware prompting. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.786, + 0.453 + ], + "angle": 0, + "content": "46. Ren, P., Li, C., Xu, H., Zhu, Y., Wang, G., Liu, J., Chang, X., Liang, X.: Viewco: Discovering text-supervised segmentation masks via multi-view semantic consistency. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.786, + 0.481 + ], + "angle": 0, + "content": "47. Rewatbowornwong, P., Chatthee, N., Chuangsuwanich, E., Suwajanakorn, S.: Zero-guidance segmentation using zero segment labels. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.786, + 0.509 + ], + "angle": 0, + "content": "48. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.786, + 0.536 + ], + "angle": 0, + "content": "49. Shin, G., Xie, W., Albanie, S.: Reco: Retrieve and co-segment for zero-shot transfer. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.786, + 0.563 + ], + "angle": 0, + "content": "50. Shin, G., Xie, W., Albanie, S.: Namedmask: Distilling segmenters from complementary foundation models. In: CVPRW (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.786, + 0.605 + ], + "angle": 0, + "content": "51. Simeoni, O., Puy, G., Vo, H.V., Roburin, S., Gidaris, S., Bursuc, A., Pérez, P., Marlet, R., Ponce, J.: Localizing objects with self-supervised transformers and no labels. In: BMVC (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.786, + 0.633 + ], + "angle": 0, + "content": "52. Simeoni, O., Sekkat, C., Puy, G., Vobecky, A., Zablocki, É., Pérez, P.: Unsupervised object localization: Observing the background to discover objects. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.786, + 0.674 + ], + "angle": 0, + "content": "53. Simeoni, O., Zablocki, E., Gidaris, S., Puy, G., Pérez, P.: Unsupervised object localization in the era of self-supervised vits: A survey. arXiv preprint arXiv:2310.12904 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.786, + 0.715 + ], + "angle": 0, + "content": "54. Vobecký, A., Simeoni, O., Hurych, D., Gidaris, S., Bursuc, A., Perez, P., Sivic, J.: Pop-3d: Open-vocabulary 3d occupancy prediction from images. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.715, + 0.786, + 0.744 + ], + "angle": 0, + "content": "55. Walmer, M., Suri, S., Gupta, K., Shrivastava, A.: Teaching matters: Investigating the role of supervision in vision transformers. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.786, + 0.771 + ], + "angle": 0, + "content": "56. Wang, X., Girdhar, R., Yu, S.X., Misra, I.: Cut and learn for unsupervised object detection and instance segmentation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.786, + 0.813 + ], + "angle": 0, + "content": "57. Wang, Y., Shen, X., Hu, S.X., Yuan, Y., Crowley, J.L., Vaufreydaz, D.: Self-supervised transformers for unsupervised object discovery using normalized cut. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.786, + 0.841 + ], + "angle": 0, + "content": "58. Wu, J., Li, X., Xu, S., Yuan, H., Ding, H., Yang, Y., Li, X., Zhang, J., Tong, Y., Jiang, X., et al.: Towards open vocabulary learning: A survey. T-PAMI (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.416, + 0.128 + ], + "angle": 0, + "content": "M.Wysoczanska et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.188 + ], + "angle": 0, + "content": "59. Wysoczanska, M., Ramamonjisoa, M., Trzcinski, T., Simeoni, O.: Clip-diy: Clip dense inference yields open-vocabulary semantic segmentation for-free. In: WACV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.217 + ], + "angle": 0, + "content": "60. Xian, Y., Choudhury, S., He, Y., Schiele, B., Akata, Z.: Semantic projection network for zero-and few-label semantic segmentation. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.258 + ], + "angle": 0, + "content": "61. Xu, J., De Mello, S., Liu, S., Byeon, W., Breuel, T., Kautz, J., Wang, X.: Groupvit: Semantic segmentation emerges from text supervision. arXiv preprint arXiv:2202.11094 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.299 + ], + "angle": 0, + "content": "62. Xu, J., Hou, J., Zhang, Y., Feng, R., Wang, Y., Qiao, Y., Xie, W.: Learning open-vocabulary semantic segmentation models from natural language supervision. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.327 + ], + "angle": 0, + "content": "63. Zhai, X., Mustafa, B., Kolesnikov, A., Beyer, L.: Sigmoid loss for language image pre-training. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.355 + ], + "angle": 0, + "content": "64. Zhao, H., Puig, X., Zhou, B., Fidler, S., Torralba, A.: Open vocabulary scene parsing. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.396 + ], + "angle": 0, + "content": "65. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.397, + 0.785, + 0.424 + ], + "angle": 0, + "content": "66. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. IJCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.438 + ], + "angle": 0, + "content": "67. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "68. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: iBOT: Image bert pre-training with online tokenizer. In: ICLR (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.466 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CLIP-DINOiser_ Teaching CLIP a few DINO tricks for open-vocabulary semantic segmentation/ab181842-dc49-431d-ab8a-9cea7505e7b5_origin.pdf b/2024/CLIP-DINOiser_ Teaching CLIP a few DINO tricks for open-vocabulary semantic segmentation/ab181842-dc49-431d-ab8a-9cea7505e7b5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5b306e2578a83867b9af0a047b7505bae24f08c1 --- /dev/null +++ b/2024/CLIP-DINOiser_ Teaching CLIP a few DINO tricks for open-vocabulary semantic segmentation/ab181842-dc49-431d-ab8a-9cea7505e7b5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd5370ff60196186e6d6dbb69d2eb40253468ccd0aef93a86833f2cc8f02476f +size 11781583 diff --git a/2024/CLIP-DINOiser_ Teaching CLIP a few DINO tricks for open-vocabulary semantic segmentation/full.md b/2024/CLIP-DINOiser_ Teaching CLIP a few DINO tricks for open-vocabulary semantic segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6082a588f095e0a2b53ad9ac55d2008b36dd0484 --- /dev/null +++ b/2024/CLIP-DINOiser_ Teaching CLIP a few DINO tricks for open-vocabulary semantic segmentation/full.md @@ -0,0 +1,263 @@ +# CLIP-DINOiser: Teaching CLIP a few DINO tricks for open-vocabulary semantic segmentation + +Monika Wysoczanska1, Oriane Simeoni2, Michael Ramamonjisoa3*, Andrei Bursuc2, Tomasz Trzciński1,4,5, and Patrick Pérez2 + +1Warsaw University of Technology, 2valeo.ai, 3Meta AI, 4Tooploox, 5IDEAS NCBR + +Abstract. The popular CLIP model displays impressive zero-shot capabilities thanks to its seamless interaction with arbitrary text prompts. However, its lack of spatial awareness makes it unsuitable for dense computer vision tasks, e.g., semantic segmentation, without an additional fine-tuning step that often uses annotations and can potentially suppress its original open-vocabulary properties. Meanwhile, self-supervised representation methods have demonstrated good localization properties without human-made annotations nor explicit supervision. In this work, we take the best of both worlds and propose an open-vocabulary semantic segmentation method, which does not require any annotations. We propose to locally improve dense MaskCLIP features, which are computed with a simple modification of CLIP's last pooling layer, by integrating localization priors extracted from self-supervised features. By doing so, we greatly improve the performance of MaskCLIP and produce smooth outputs. Moreover, we show that the used self-supervised feature properties can directly be learnt from CLIP features. Our method CLIP-DINOiser needs only a single forward pass of CLIP and two light convolutional layers at inference, no extra supervision nor extra memory and reaches state-of-the-art results on challenging and fine-grained benchmarks such as COCO, Pascal Context, Cityscapes and ADE20k. The code to reproduce our results is available at https://github.com/wysoczanska/clip_dinoiser. + +Keywords: open-vocabulary semantic segmentation $\cdot$ self-supervised features $\cdot$ annotation-free segmentation + +# 1 Introduction + +Semantic segmentation is a key visual perception task for many real-world systems, e.g., self-driving cars, and industrial robots. Typically tackled in a dataset-oriented manner, best methods require a training dataset which is manually annotated for a specific and finite set of classes. The advent of powerful Vision-Language Models (VLM) [24,43,63] is stimulating a shift from a closed-vocabulary + +![](images/3c26a0e78ac5a532f064d0e9558cf8d9b1b7183e8f2abf666a5549b63b5f5285.jpg) +Fig. 1: Examples of open-vocabulary semantic segmentation results obtained with our method CLIP-DINOiser on 'in-the-wild' images vs. those of MaskCLIP [67]. Our method improves MaskCLIP features with a smart pooling strategy which does not alter the original open-vocabulary properties. We use self-supervised DINO [5] as a guide to teach CLIP [22] to produce DINO-like localization features through two light convolutional layers. Our method, which achieves state-of-the-art results, only requires a single forward pass through CLIP model and our two layers. In addition to the correct prompts (light grey row) we list the irrelevant prompts predicted (in yellow) that we query in all images shown here. + +paradigm to an open-world one. Such models are trained with a simple but scalable objective: to align pairs of image and coarse text captions that can be obtained in large amounts with limited manual supervision. VLMs excel at associating global image content with arbitrary text inputs with remarkable generalization capabilities [17, 34], but struggle to provide dense open-vocabulary features [18, 67]. Obtaining such an alignment between pixels and language can lead to open-vocabulary extensions for multiple other modalities, such as point clouds [8, 23, 38, 41], 3D scenes [54], 3D shapes [1], radiance fields [27], intermodality alignment [19, 23], with multiple potential applications for which the construction of training datasets is even more challenging and where CLIP-derived models show promising results. + +Different strategies have been recently proposed towards improving CLIP's patch-level feature extraction abilities by modifying the original CLIP architecture for dense pooling and retraining [6, 37, 44, 61, 62] or finetuning on an annotated segmentation dataset with pre-defined classes [32, 67]. The former requires long training and/or large collections of annotated data, while the latter leads to an alteration of the vision-language associations of the CLIP features. An alternative line of approaches freezes the CLIP encoder and directly densifies its features with different heuristics, often with multiple forward passes [1, 23, 27, 49, 50, 59], but are less practical due to the extensive computa + +tional overhead. MaskCLIP [67] arises as a computationally efficient dense CLIP extractor. It converts CLIP's global self-attention layer into a convolutional one to produce patch features with original vision-language qualities. If such features are local, they appear to be too noisy for high-quality segmentation mask extraction (see Fig. 2b middle column). + +Meanwhile, recent self-supervised learning (SSL) approaches [4, 5, 9, 68] produce strong visual representations displaying object localization properties, and such without requiring any manual annotation. DINO [5] stands out with its semantically meaningful features which have been exploited for unsupervised object discovery [51, 52, 56, 57]. DINO features prove useful also for zero-shot semantic segmentation [25, 27, 59], but require expensive sliding window sampling [27, 59] or building concept-specific prototypes and ensemble strategies [25]. + +In this work, we aim for unaltered patch-level CLIP features with minimal runtime overhead. To this end, we re-examine the localization properties of MaskCLIP features and observe that it is possible to easily refine them with guidance from SSL models. In detail, we train a simple convolutional layer on unlabeled data to produce pooling weights to perform correlation-guided dense feature pooling from CLIP without distorting the vision-language alignment. This layer is optimized to mimic the patch correlations of DINO [5] that indicate likely layouts of visual concepts in the images. Furthermore, we show that the unsupervised objectness information given by FOUND [52] from DINO features can be also directly learned from CLIP features again in a fully-unsupervised fashion with a single convolutional layer and helps improve the segmentation of the ill-defined 'background' prompt. With CLIP-DINOiser, we obtain high-quality masks in a single forward pass on CLIP (see Fig. 1). CLIP-DINOiser is amenable to producing dense semantic maps. + +To summarize, our contributions are: (1) We propose a light pooling mechanism to refine MaskCLIP features by leveraging guidance from SSL features without degrading its original open-vocabulary properties. CLIP-DINOiser does not require any annotations, nor retraining CLIP from scratch, but only a single CLIP forward pass. (2) We show that CLIP already contains good localization properties which can be exploited. We leverage simple convolutional layers to emphasize visual concept layouts from dense CLIP features. We train them without any annotation on only 1k of raw images randomly sampled in ImageNet [13]. We believe that this finding could be further exploited in different contexts. (3) Our method achieves state-of-the-art results on complex semantic segmentation datasets such as COCO [3], Pascal Context [16], Cityscapes [11] and ADE20K [66]. + +# 2 Related Work + +Zero-shot semantic segmentation. This task has been typically approached by methods which aim at generalizing from seen classes to unseen ones [2, 20, 21, 26, 30, 40, 60, 64]. Such strategies train models with full supervision on the set of seen classes and propose different solutions to extend them to unseen ones with + +out new images (labeled or unlabeled), e.g., by exploiting class information and relationships encapsulated in popular word embeddings [35,42]. While they produce fine segmentations without computational overhead, these methods require pixel-level annotations for the seen classes. + +From CLIP to open-vocabulary segmentation. The surge of VLMs with aligned image-language representations [22, 24, 43] brought back into the spotlight the zero-shot classification task. However, the extension to zero-shot segmentation is not obvious as the CLIP architecture is not equipped to yield dense vision-language features [18, 67]. To produce dense CLIP features, several approaches fine-tune or train from scratch pixel-aligned CLIP-like models with additional modules, mechanisms or supervision objectives [6, 37, 44, 61, 62] on datasets with annotations of varying granularity and quality: dense annotations [29, 31], class-agnostic object masks [14, 18, 45], coarse captions [6, 18, 31–33, 37, 44, 61, 62, 65] or pseudo-labels [67]. Recent works leverage image-level captions to align text to regions (obtained without supervision): PACL [37] trains an embedder module to learn patch-to-text affinity, TCL [6] proposes a local contrastive objective to align well-selected patches to the text and ViewCO [46] leverages multi-view consistency. On the downside, such models require long training on millions of images or specific types of very costly annotations. Also, fine-tuning CLIP with a defined vocabulary is more computationally appealing [29, 31, 67], but alters the open-vocabulary properties of the features [23]. + +Most related to us is a line of works that investigate how to directly densify CLIP features [1, 23, 27, 59, 67] to obtain per-patch CLIP features. Such densification can be performed by aggregating features from multiple views [1, 27] or from sliding windows [23, 59] at the extra-cost of multiple forward passes. MaskCLIP [67] drops the global pooling layer of CLIP and matches the projected features directly to text via a $1 \times 1$ convolution layer. By doing so they achieve dense predictions, however noisy. + +With a concept-driven perspective, some methods [25,49,50] build codebooks of visual prototypes per concept, including negative prototypes [25], and then perform co-segmentation [49]. While such an approach yields good results, it is however at the cost of building expensive class-specific prototypes, therefore diverging from open-vocabulary scenarios. Instead, we aim to remain open to avoid retraining a model or building new expensive prototypes whenever a new concept is considered. To that end, we devise a dense CLIP-feature extraction method that preserves the open-vocabulary quality. + +Leveraging self-supervised models & CLIP. Recent self-supervised ViTs [4,5,9,12, 68] have demonstrated features with good localization properties [51,52,56,57]. Such features have also been exploited in the context of open-vocabulary segmentation methods, e.g. for pre-training for the visual backbone [7,44,62], cosegmentation [49], clustering patches into masks [47], representing object prototypes [25]. Related to us is the recent CLIP-DIY [59] which computes patch-level representations from CLIP features from different image crops with guidance from an unsupervised saliency segmenter [52] FOUND. While we also leverage + +the latter, in contrast with CLIP-DIY which runs multiple forward passes to build their dense CLIP features, our method requires only a single forward pass of CLIP. Furthermore, our method mitigates the limits of FOUND in cluttered scenarios by integrating an uncertainty constraint. Finally, we leverage the informative patch correlation properties of DINO [5] and show that it is possible to teach CLIP to produce DINO-like features through light convolutional layers. + +# 3 Method + +We present in this section CLIP-DINOiser, a simple and efficient strategy to improve MaskCLIP using localization information extracted from CLIP—with a lightweight model trained to mimic some of DINO's properties. We first set the goal in Sec. 3.1 and present MaskCLIP [67] in Sec. 3.2. We then introduce our strategy which leverages self-supervised features localization information to consolidate MaskCLIP features in Sec. 3.3 and discuss how such localization information can directly be learnt from CLIP in Sec. 3.4 (we visualize both steps in Fig. 3). We also propose a way to improve the 'background' filtering in Sec. 3.5. + +# 3.1 Problem statement + +In this work, we aim to produce open-vocabulary1 semantic segmentation of an image. We consider an image $X \in \mathbb{R}^{H \times W \times 3}$ which we split into a sequence of $N$ patches of dimensions $P \times P \times 3$ with $P \times P$ the patch size and $N = \lceil \frac{H}{P} \rceil \cdot \lceil \frac{W}{P} \rceil$ . A class token, noted CLS, is added to the input sequence and we feed the $N + 1$ patches to a ViT [15] model. We aim at producing dense visual features $F \in \mathbb{R}^{N \times d}$ , with $d$ the feature dimension, that can later be matched to any set of text inputs embedded in the same space. In particular, the goal is to produce a segmentation map per textual query. + +# 3.2 Preliminaries on MaskCLIP + +Extracting dense open-vocabulary features. The popular CLIP [22] model pretrained on image/caption pairs produces good global image features, but was not trained to generate high-quality 2D feature maps. In order to extract such dense feature maps relevant to semantic segmentation, Zhou et al. [67] revisit the global attention pooling layer of the last attention layer of the model. The authors discard the query and key embeddings of the layer and transform both the value projection and the last linear layer into a conv $1 \times 1$ layer. With this new model, named MaskCLIP and denoted $\phi(\cdot)$ , we extract $d$ -dimensional features $\phi^L(X) \in \mathbb{R}^{N \times d}$ from the last layer $L$ which retains most of the open-vocabulary properties of CLIP [67]. + +Semantic segmentation given textual queries. We also extract CLIP textual features $\phi_T(t_j)$ for each text query $t_j \in \mathcal{T}$ with $j \in \{1, \dots, |\mathcal{T}|\}$ . Segmentation maps are then generated by computing the cosine similarity between each of the visual patch features and of the textual prompts, after L2-normalization. The most similar prompt is assigned to each patch. Note that a query 'background' can be added in order to obtain negative patches. Using MaskCLIP allows us to produce dense segmentation maps with a single forward pass of the classic CLIP model, but its outputs are noisy, as visible in Fig. 2b (middle column). + +# 3.3 DINOising open-vocabulary features + +In this work, we aim to improve MaskCLIP's open-vocabulary features described above. To do so, we propose to leverage the known good localization properties of self-supervised features [5,39,51-53,57]. + +Extracting self-supervised correlation information. Recent works [51, 57] have shown that the patch correlation information of the embeddings from the last attention layer of the self-supervised model, DINO [5] can help highlight objects in images. We use here the value embeddings which we observe have finer correlation than those of key and query (more discussion in supplementary material). We extract such self-supervised features $\xi(X) \in \mathbb{R}^{N \times d_{\xi}}$ and discard the CLS token. We then compute the per-patch cosine-similarity and produce the affinity map $A^{\xi} \in [-1,1]^{N \times N}$ . We compare in Fig. 4 the patch-similarities obtained for a patch seed with MaskCLIP and DINO features and observe that the self-supervised features are more densely and accurately correlated than those of CLIP. + +![](images/0729da00fee285f570f825a5d25997b27a4bcc703e4cc49a29d518b57c54c55d.jpg) +(a) Our guided pooling + +![](images/9b6982e578b1649eb765a836d24013c91ea8b704e9ba97f8b52304b57d6020ee.jpg) +(b) Impact of the pooling +Fig. 2: We present in (a) our guided pooling strategy defined in Eq. (1). The $N \times N$ affinity matrix is computed from patch features and is used to refine MaskCLIP features (bottom left). In (b) we compare our results with $F^{+}$ (right) versus those obtained with MaskCLIP features (middle). + +Strengthening features with guided pooling. In order to locally consolidate MaskCLIP features $\phi^L (X)$ , now noted $F$ , we propose to perform a concept-aware linear combination of the features per patch with guidance from the patch affinity $A^{\xi}$ . The + +feature combination strategy can be seen as a form of voting mechanism that enforces similar patches to have similar CLIP features (and prediction) while attenuating noisy features. Specifically, we compute the new features $F^{+} \in \mathbb{R}^{N \times d}$ as an average of MaskCLIP features $F$ weighted by $A^{\xi}$ , presented in Fig. 2a. We zero-out $A^{\xi}$ correlations below a threshold $\gamma$ , following [51, 57], and compute the new features for patch $p \in \{1, \dots, N\}$ : + +$$ +F _ {p} ^ {+} = \frac {1}{\sum_ {q = 1} ^ {N} A _ {p , q} ^ {\xi}} \sum_ {q = 1} ^ {N} A _ {p, q} ^ {\xi} \cdot F _ {q}. \tag {1} +$$ + +We then produce the segmentation maps $S \in [-1,1]^{N \times |\mathcal{T}|}$ , by comparing the new features $F^+$ to each textual queries in $\mathcal{T}$ . As shown in Fig. 2b, when using such consolidated features, we obtain more accurate outputs and the high-frequency predictions observed in MaskCLIP are smoothed out, showing the benefit of the pooling. + +# 3.4 Teaching CLIP a first DINO trick: object correlations + +![](images/716ea1effc3d6881eac5b1b729f0220ba496b5e336bfe34c39d3b5c2bb8fcf30.jpg) +Fig. 3: Overview of CLIP-DINOiser which leverages the quality of self-supervised features to improve the notoriously noisy MaskCLIP feature maps. We use DINO as a teacher which 'teaches' CLIP how to extract localization information. We train (left) a conv $3 \times 3$ layer to reproduce the patch correlations obtained with DINO. At inference (right), an input image is forwarded through the frozen CLIP image backbone and MaskCLIP projection. The produced features are then improved with our pooling strategy which is guided by correlations predicted with the trained convolutional layer applied on CLIP. With this light 'DINOising' process, we obtain 'DINOised' features which are matched against the prompts features to produce CLIP-DINOiser outputs. + +We have shown in the previous section that self-supervised correlation information can successfully be used to improve the dense quality of open-vocabulary + +features. If the difficulty of densifying CLIP is well-known, we show here that CLIP features already contain good localization information which can be extracted with a light model. We indeed predict DINO correlations $A^{\xi}$ from CLIP with a single convolutional layer. + +In order to predict the DINO affinity map $A^{\xi}$ from CLIP features, we train a single $3 \times 3$ convolutional layer $g(\cdot): \mathbb{R}^{d} \to \mathbb{R}^{d_{g}}$ which projects intermediate features $\phi^{l}(X)$ -extracted from layer $l$ -into a smaller space of dimension $d_{g} < d$ . We enforce the patch correlations of the generated features $A^{\phi} \in [-1,1]^{N \times N}$ : + +$$ +A ^ {\phi} = \frac {g \left(\phi^ {l} (X)\right)}{\| g \left(\phi^ {l} (X)\right) \|} \otimes \left(\frac {g \left(\phi^ {l} (X)\right)}{\| g \left(\phi^ {l} (X)\right) \|}\right) ^ {\top}, \tag {2} +$$ + +with $\otimes$ denoting the outer product, to be close to the binarized correlations $D = A^{\xi} > \gamma$ (we use here the same $\gamma$ as defined above), using the binary cross-entropy loss $\mathcal{L}^c$ : + +$$ +\mathcal {L} ^ {c} = \sum_ {p = 1} ^ {N} \left[ D _ {p} \log A _ {p} ^ {\phi} + (1 - D _ {p}) \log \left(1 - A _ {p} ^ {\phi}\right) \right]. \tag {3} +$$ + +We present our layer training in Fig. 3 (left part) and observe the quality of CLIP-predicted affinity matrix $A^{\phi}$ . We also show in Fig. 4 another example of obtained $A^{\phi}$ and observe their similarity to DINO-based correlations. We use the CLIP-produced correlations $A^{\phi}$ to replace $A^{\xi}$ in Eq. (1) to weight the pooling and observe a similar boost over MaskCLIP, thus showing that good patch correlations can indeed be extracted directly from CLIP. We can now discard DINO and we name CLIP-DINOiser the guided-pooling strategy which uses CLIP-based correlation. As shown in Fig. 3 (inference step), our method runs with a single forward pass of CLIP model and a small extra layer. + +![](images/f70992381d5ab9a8d6f8a4a3b46d5082dbda506ab7a5d0d2bfc451c3c6c1ec6f.jpg) +Fig. 4: Comparison of the affinity maps between a seed (one on the 'plant' and the other on a 'pillow') and the other patch features when using features of MaskCLIP, DINO and ours after training. + +# 3.5 Teaching CLIP a second DINO trick: background filtering + +Moreover, as discussed earlier, a 'background' query may be added to the set of textual queries $\mathcal{T}$ in order to help filter out patches falling in the background and not corresponding to any objects. We do not assume here any prior knowledge about classes of interest and focus rather on the foreground/background paradigm [52]. We argue that relying solely on the textual prompt 'background' + +![](images/e6f1ef85e944166ce6e14abc8f30d5fa2df07c5a28b8733a85ea850f7e14f0af.jpg) +(a) Comparison of background filtering + +![](images/838aa25cde5bf5f63a9a16f11c3540528c48854aef5d5863b23e3cc6e50cc42c.jpg) +(b) Background filtering +Fig. 5: We present in (a) a comparison of objectness mask generated by FOUND [52] and with our layer using CLIP features. We carefully define the fusion operation and the simple training strategy of the conv $1 \times 1$ again using DINO as a teacher in Sec. 3.5. In (b) is an overview of our background filtering which is applied when a 'background' prompt is provided and helps reduce hallucinations. + +to catch all non-salient patches is underperforming and, similarly to [59], we propose to use a very light-weight unsupervised foreground/background segmentation method, namely FOUND [52] which also relies on DINO self-supervised features. We run FOUND on the entire image and extract a prediction mask $M \in \{0,1\}^N$ in which a patch is assigned the value 1 if falling into the foreground and 0 otherwise. We also observe that saliencies produced by FOUND can be too restrictive and discard objects which are partially visible or in a clutter. To mitigate this behaviour, we propose to relax the background selection by integrating an additional uncertainty constraint. To this end, we fuse the background information from both modalities by assigning the 'background' prompt to patches $p$ which are both uncertain, e.g. have low confidence score $\sigma(S)_p < \delta$ , with $\sigma(\cdot)$ the softmax operation, and which fall in the background in $M$ . + +Learning FOUND objectness. Moreover, we are also able to learn the predictions of FOUND [52] directly from CLIP features. To do so, we train a single $1 \times 1$ convolutional layer $h(\cdot): \mathbb{R}^d \to \mathbb{R}$ which predicts from the features $\phi^l(X)$ an objectness map $M^\phi = h(\phi^l(X)) \in \mathbb{R}^N$ . We train the model to predict the FOUND binary mask $M$ with the binary cross-entropy loss $\mathcal{L}^m$ : + +$$ +\mathcal {L} ^ {m} = \sum_ {p = 1} ^ {N} \left[ M _ {p} \log \left(M _ {p} ^ {\phi}\right) + \left(1 - M _ {p}\right) \log \left(1 - M _ {p} ^ {\phi}\right) \right]. +$$ + +We show examples of predicted CLIP-based objectness in Fig. 5a and observe their very high similarity to those produced with DINO. Moreover, we can now replace $M$ defined above with the binarized CLIP-based scores $\zeta(M^{\phi}) > 0.5$ with $\zeta(\cdot)$ the sigmoid operation, and observe a minimal drop in performances. We show an example of the background filtering with trained objectness in Fig. 5b. + +# 4 Experiments + +We detail in Sec. 4.1 the experimental setup used in our evaluation. We produce state-of-the-art results on the task of open-vocabulary semantic segmentation in Sec. 4.2 and ablation studies in Sec. 4.3. + +# 4.1 Experimental setup + +Technical details. We use in all experiments a frozen CLIP ViT-B/16 pre-trained following OpenCLIP [22]. Our method CLIP-DINOiser uses two convolutional layers to extract DINO-like information from CLIP layer $l = 10$ (the 3rd before the last which was shown to provide the best results [55]). The first layer $g(\cdot)$ has a kernel $3 \times 3$ and output dimension $d_{g} = 256$ and $h(\cdot)$ a kernel $1 \times 1$ with $d_{h} = 1$ . The first is trained to match the correlation information extracted from the value embeddings of the last layer of a ViT-B/16 model trained following DINO [5]. The second layer is trained to replicate the unsupervised object localization predictions of FOUND [52]–which also uses DINO model. We train both layers with a binary cross-entropy loss on only 1k raw images randomly sampled from ImageNet [13] dataset without any annotation. We report average scores over 3 runs with different sampling seeds and provide standard deviations in the supplementary material. We follow [57] and binarize the correlations with $\gamma = 0.2$ . In the background filtering step, we use a high confidence score, i.e., $\delta = 0.99$ . We train our model for 6k iterations with a batch size of 16 images using Adam optimizer [28], which takes approximately 3 hours on a single NVIDIA RTX A5000 GPU. We decrease the learning rate for both heads by a factor of 0.1 after 5k iterations. We apply data augmentations during training (random scale and cropping, flipping and photometric distortions). + +Datasets and metric. We evaluate our method on eight benchmarks typically used for zero-shot semantic segmentation [6]. Following [6], we split them into two groups. The first consists in datasets with a 'background' query: PASCAL VOC [16] (noted 'VOC'), PASCAL Context [36] (noted 'Context'), and COCO Object [3] (noted 'Object') and the second without: PASCAL VOC20 [16] (noted 'VOC20'), PASCAL Context59 [36] (noted 'C59'), COCO-Stuff [3] (noted 'Stuff'), Cityscapes [11] (noted 'City'), and ADE20K [66] (noted 'ADE'). We evaluate results with the standard mIoU metric. We also follow the evaluation protocol of [6], use the implementations provided by MMSegmentation [10], employ a sliding window strategy, resize the input image to have a shorter side of 448. We also do not perform text expansions of the class names and use only the standard ImageNet prompts following [22, 61, 67]. + +Baselines. We compare our method against state-of-the-art methods on open-vocabulary zero-shot semantic segmentation. For a fair comparison between methods, we report results without any post-processing step. In our evaluations, we follow the taxonomy presented in [58] and compare our model with the methods relying on language-image pretraining, also called open-vocabulary. + +We split the compared baselines into four categories: (1) dataset specific which employ pseudo-labeling and supervised training of a segmentation model on target dataset: NamedMask [50], MaskCLIP + [67]); (2) construct prototypes: ReCO [49], OVDiff [25]; (3) train with text supervision including GroupViT [61], ZeroSeg [47], SegCLIP [33], TCL [6], CLIPpy [44], OVSegmentor [62], which all require access to additional datasets of millions of image/caption pairs (we note in the table the exact datasets used for the training); and finally use frozen CLIP i.e. CLIP-DIY [59] and MaskCLIP [67], which use pre-trained CLIP. Our method falls into the last category as we do not modify CLIP, and do not need access to additional caption annotations as we use only 1k unannotated images. + +
MethodsConcept spec.Extra Inference data backboneNo background promptW/ bkg prompt Cont. Object VOC
VOC20C59StuffCityADE
Dataset specific
MaskCLIP+ [67]✓XIDLv2-31.118.0----
NamedMask [50]✓XIDLv3+-----27.759.2
Build prototypes per visual concept
ReCo [49]✓✓ICLIP57.822.314.821.111.219.915.7
OVDiff [25]✓✓XCLIP+ DINO+SD81.733.7--14.930.134.8
Text/image alignment training with captions
GroupViT [61] [6]XXITCLIP79.723.415.311.19.218.727.5
ZeroSeg [7]XXITCLIP-----21.822.1
SegCLIP [33]XXITCLIP---11.08.724.726.5
TCL [6]XXITCLIP77.530.319.623.114.924.330.4
CLIPpy [44]XXITCLIP----13.5-32.0
OVSegmentor [62]XXITCLIP----5.620.425.1
Frozen CLIP
CLIP-DIY [59]*X✓XCLIP+ DINO79.719.813.311.69.919.731.0
MaskCLIP [67] [6]X✓XCLIP53.723.314.721.610.821.115.5
MaskCLIP*X✓XCLIP61.825.617.625.014.322.916.4
MaskCLIP* †X✓XCLIP71.927.418.623.014.924.021.6
CLIP-DINOiserX✓I(1k)CLIP80.935.924.631.720.032.434.8
+ +Table 1: Open-vocabulary semantic segmentation quantitative comparison using the mIoU metric. We separate in two columns the evaluation datasets: those without a 'background' prompt and those with (noted 'W/ bkg prompt'), as discussed in Sec. 4.1. We report all methods without post-processing. We note with * methods for which we computed scores; we obtained MaskCLIP* scores with OpenCLIP [22] and mark with † the use of MaskCLIP refinement. The first and second best methods are respectively bold and underlined. We specify if a method assumes prior access to names of concepts ('Concept spec.)' and if it employs a frozen backbone (''). We specify what additional data is used at training ('Extra data') ('I' stands for images and 'IT' for image/text aligned data). Our CLIP-DINOiser only needs 1k images from ImageNet to be trained. 'SD' stands for Stable Diffusion [48]. We refer to Sec. 4.1 for more details on baselines and we detail the datasets used for training by each method in the supplementary material. + +# 4.2 Open-vocabulary semantic segmentation + +We discuss in this section state-of-the-art results on the task of open-vocabulary semantic segmentation. + +Evaluation with no 'background' class. We first compare in Tab. 1 ('No background prompt' column) the results on datasets which aim at the segmentation of most of the pixels in an image and do not consider a 'background' class. We observe that our method CLIP-DINOiser achieves the best results on four datasets yielding $+2.2$ , $+5.0$ , $+6.7$ and $+5.1$ mIoU over the second best performing method. Interestingly, we outperform methods which build expensive prototypes per visual concept on fine-grained datasets, showing the benefit of our lightweight and generalizable method. The only drop $(-0.8$ mIoU) is seen on VOC20 with respect to OVDiff; we believe it is due to the benefit of generating per-concept negative prototypes which likely benefits this object-centric dataset. An adaptive granularity of feature correlation could help mitigate this drop, which we leave for future work. + +Evaluation with 'background' class. We now compare our method on datasets which include a 'background' query in Tab. 1 ('W/ bkg prompt' column). In this setup, we also apply our background detection mechanism (detailed in Sec. 3.5) on VOC and Object in order to improve the stuff-like background detection. We observe that CLIP-DINOiser significantly outperforms all methods which do not construct prototypes. Moreover, we surpass OVDiff (which uses an ensemble of three models) on Context dataset by $+2.3\mathrm{mIoU}$ and are on par on Object. It is to be noted that with a single feature extractor, the performance of OVDiff drops by $-10\mathrm{mIoU}$ and the method requires the construction of a 'background' prototype per concept, otherwise losing another $-10\mathrm{mIoU}$ on VOC. On the other hand, CLIP-DINOiser produces segmentation masks in a single pass of CLIP with the light addition of two convolutional layers while remaining fully open-vocabulary as it does not require any concept-specific constructs. + +Qualitative results. We qualitatively compare in Fig. 6 CLIP-DINOiser with high-performing TCL [6], CLIP-DIY [59] (two recent methods which provide code) and our baseline method MaskCLIP [67] on images taken from the datasets considered in the evaluation. We observe that our method generates predictions accurate both in terms of localization and assignment. Indeed we obtain fined-grained results on the challenging datasets, e.g. in the Cityscapes example the text query 'car' and in the ADE20k example 'fountain' are accurately located when CLIP-DIY and TCL produce coarser results. Versus MaskCLIP, we can see the denoising capabilities of CLIP-DINOiser as MaskCLIP hallucinations grow with the number of text queries prompted at evaluation. Finally, in Fig. 1 we present 'in the wild' examples, beyond the evaluation benchmarks, and show that CLIP-DINOiser produces accurate segmentation masks for arbitrary and very specific prompts, such as 'wooden table' or 'leather bag'. + +![](images/1ce59c04146c54e151e6a345fa5d585a1022ce5194e561eec165314b08037cbf.jpg) +Fig.6: Qualitative open-vocabulary segmentation results. We compare ours against CLIP-DIY [59], TCL [6] and MaskCLIP [67]. For a fair comparison, we do not apply post-processing. All pixels annotated in black are from the background class. We observe that our method achieves more accurate results both in terms of localization and class assignment. + +# 4.3 Ablation study + +We now conduct an ablation study of the different components of CLIP-DINOiser and investigate the impact of both our feature pooling strategy and background detection. + +The impact of the pooling mechanism. We propose with CLIP-DINOiser to combine MaskCLIP features with a well-defined linear combination and compare different solutions in Tab. 2a. In [67], the authors proposed to refine the predictions with a combination weighted by CLIP key embeddings (noted 'CLIP keys (preds.)' in the table) and boost MaskCLIP results by more than $+8$ mIoU on VOC and VOC20, $+1.8$ and $+1.0$ and $+0.6$ mIoU on the other datasets. However, we show that working directly at the feature level allows us to achieve better results; we obtain consistent improvements ranging from $+6$ to $+19$ mIoU on + +all datasets when using DINO-based weight $A^{\xi}$ and further improve when using trained CLIP-based weights $A^{\phi}$ . + +
Pooling strategyVOCVOC20C59StuffADE
MaskCLIP [67] - baseline32.961.825.617.614.3
CLIP keys (preds.) [67]41.371.927.418.614.9
ours w. CLIP keys39.273.223.012.67.7
ours w. DINO Aξ53.779.135.524.720.4
ours w. trained Aφ54.080.935.924.620.0
+ +(a) Pooling strategy + +
PoolingBkg det.ObjectVOC
MaskCLIP [67] - baseline16.432.9
ours w. DINO Aξ29.953.7
ours w. DINO AξFOUND32.160.1
ours w. DINO Aξours w. M34.162.1
ours w. DINO Aξours w. Mφ34.261.9
ours w. trained Aφours w. Mφ34.862.1
+ +(b) Background detection + +Table 2: Impact of the pooling strategy (a) and background detection (b) on diverse datasets reported with the mIoU metric. + +The impact of the background detection. We now discuss the improvement provided by our background refinement strategy, which is applied when stuff-like background patches need to be detected. We report such results in Tab. 2b when employing our pooling strategy (either using DINO features, noted 'w. DINO $A^{\xi}$ or those extracted from CLIP, noted 'w. trained $A^{\phi}$ ). When using solely 'FOUND' for background detection, as in [59], we improve by $+6.4$ mIoU on VOC (achieving $60.1\mathrm{mIoU}$ ), but when relaxing FOUND (see Sec. 3.5) with an uncertainty condition, we boost scores up to 62.1 on VOC, showing the limitation of using FOUND alone. We also achieve similar results when using CLIP-based predictions $M^{\phi}$ both with DINO-based $A^{\xi}$ and trained CLIP-based $A^{\phi}$ correlations, although we observe that best results are overall obtained with trained $A^{\phi}$ . We visualize CLIP-based mask $M^{\phi}$ in Fig. 5a and see high similarity to DINO-based predictions, therefore showing the localization quality of CLIP. + +# 5 Conclusions + +In this work, we propose to make the most out of CLIP features and show that the features already contain useful localization information. Indeed with light convolutional layers, we are able to learn both good patch-correlation and objectness information by using DINO self-supervised model as a guide. With such information, our method CLIP-DINOiser performs zero-shot open-vocabulary semantic segmentation in a single pass of CLIP model and with two light extra convolutional layers. CLIP-DINOiser reaches state-of-the-art results on complex semantic segmentation datasets. + +Limitations. Despite yielding strong results on open-vocabulary semantic segmentation, CLIP-DINOiser is still bounded by the capability of the CLIP model to separate classes, as it inherits its granularity. We believe that better prompt engineering paired with better image-text models could further boost the performance of CLIP-DINOiser. + +# Acknowledgments + +This work was supported by the National Centre of Science (Poland) Grant No. 2022/45/B/ST6/02817 and by the grant from NVIDIA providing one RTX A5000 24GB used for this project. + +# References + +1. Abdelreheem, A., Skorokhodov, I., Ovsjanikov, M., Wonka, P.: Satr: Zero-shot semantic segmentation of 3d shapes. In: ICCV (2023) +2. Bucher, M., Vu, T.H., Cord, M., Pérez, P.: Zero-shot semantic segmentation. In: NeurIPS (2019) +3. Caesar, H., Uijlings, J., Ferrari, V.: Coco-stuff: Thing and stuff classes in context. In: CVPR (2018) +4. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. In: NeurIPS (2020) +5. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021) +6. Cha, J., Mun, J., Roh, B.: Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs. In: CVPR (2023) +7. Chen, J., Zhu, D., Qian, G., Ghanem, B., Yan, Z., Zhu, C., Xiao, F., Culatana, S.C., Elhoseiny, M.: Exploring open-vocabulary semantic segmentation from clip vision encoder distillation only. In: ICCV (2023) +8. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: Clip2scene: Towards label-efficient 3d scene understanding by clip. In: CVPR (2023) +9. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020) +10. Contributors, M.: MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark (2020) +1. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: CVPR (2016) +2. Darcet, T., Oquab, M., Mairal, J., Bojanowski, P.: Vision transformers need registers. arXiv preprint arXiv:2309.16588 (2023) +3. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009) +4. Ding, J., Xue, N., Xia, G.S., Dai, D.: Decoupling zero-shot semantic segmentation. In: CVPR (2022) +5. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) +6. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results +7. Fang, A., Ilharco, G., Wortsman, M., Wan, Y., Shankar, V., Dave, A., Schmidt, L.: Data determines distributional robustness in contrastive language image pretraining (CLIP). In: ICML (2022) +8. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: ECCV (2022) + +19. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: CVPR (2023) +20. Gu, Z., Zhou, S., Niu, L., Zhao, Z., Zhang, L.: Context-aware feature generation for zero-shot semantic segmentation. In: ACM MM (2020) +21. Hu, P., Sclaroff, S., Saenko, K.: Uncertainty-aware learning for zero-shot semantic segmentation. In: NeurIPS (2020) +22. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021). https://doi.org/10.5281/zenodo.5143773, https://doi.org/10.5281/zenodo.5143773 +23. Jatavallabhula, K.M., Kuwajerwala, A., Gu, Q., Omama, M., Chen, T., Li, S., Iyer, G., Saryazdi, S., Keetha, N., Tewari, A., et al.: Conceptfusion: Open-set multimodal 3d mapping. In: RSS (2023) +24. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021) +25. Karazija, L., Laina, I., Vedaldi, A., Rupprecht, C.: Diffusion models for zero-shot open-vocabulary segmentation. arXiv preprint arXiv:2306.09316 (2023) +26. Kato, N., Yamasaki, T., Aizawa, K.: Zero-shot semantic segmentation via variational mapping. In: ICCVW (2019) +27. Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: Lerf: Language embedded radiance fields. In: ICCV (2023) +28. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015) +29. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: ICLR (2022) +30. Li, P., Wei, Y., Yang, Y.: Consistent structural relation learning for zero-shot segmentation. NeurIPS (2020) +31. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: CVPR (2023) +32. Liu, Q., Wen, Y., Han, J., Xu, C., Xu, H., Liang, X.: Open-world semantic segmentation via contrasting and clustering vision-language embedding. In: ECCV (2022) +33. Luo, H., Bao, J., Wu, Y., He, X., Li, T.: SegCLIP: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. In: ICML (2023) +34. Mayilvahanan, P., Wiedemer, T., Rusak, E., Bethge, M., Brendel, W.: Does CLIP's generalization performance mainly stem from high train-test similarity? arXiv preprint arXiv:2310.09562 (2023) +35. Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word representations in vector space. In: ICLR (2013) +36. Mottaghi, R., Chen, X., Liu, X., Cho, N.G., Lee, S.W., Fidler, S., Urtasun, R., Yuille, A.: The role of context for object detection and semantic segmentation in the wild. In: CVPR (2014) +37. Mukhoti, J., Lin, T.Y., Poursaeed, O., Wang, R., Shah, A., Torr, P.H., Lim, S.N.: Open vocabulary semantic segmentation with patch aligned contrastive learning. In: CVPR (2023) +38. Najibi, M., Ji, J., Zhou, Y., Qi, C.R., Yan, X., Ettinger, S., Anguelov, D.: Unsupervised 3d perception with 2d vision-language distillation for autonomous driving. In: ICCV (2023) + +39. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., HAZIZA, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: DINOv2: Learning robust visual features without supervision. TMLR (2024) +40. Pastore, G., Cermelli, F., Xian, Y., Mancini, M., Akata, Z., Caputo, B.: A closer look at self-training for zero-label semantic segmentation. In: CVPR (2021) +41. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: Openscene: 3d scene understanding with open vocabularies. In: CVPR (2023) +42. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: EMNLP (2014) +43. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021) +44. Ranasinghe, K., McKinzie, B., Ravi, S., Yang, Y., Toshev, A., Shlens, J.: Perceptual grouping in contrastive vision-language models. In: ICCV (2023) +45. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: Denseclip: Language-guided dense prediction with context-aware prompting. In: CVPR (2022) +46. Ren, P., Li, C., Xu, H., Zhu, Y., Wang, G., Liu, J., Chang, X., Liang, X.: Viewco: Discovering text-supervised segmentation masks via multi-view semantic consistency. In: ICLR (2023) +47. Rewatbowornwong, P., Chatthee, N., Chuangsuwanich, E., Suwajanakorn, S.: Zero-guidance segmentation using zero segment labels. In: ICCV (2023) +48. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) +49. Shin, G., Xie, W., Albanie, S.: Reco: Retrieve and co-segment for zero-shot transfer. In: NeurIPS (2022) +50. Shin, G., Xie, W., Albanie, S.: Namedmask: Distilling segmenters from complementary foundation models. In: CVPRW (2023) +51. Simeoni, O., Puy, G., Vo, H.V., Roburin, S., Gidaris, S., Bursuc, A., Pérez, P., Marlet, R., Ponce, J.: Localizing objects with self-supervised transformers and no labels. In: BMVC (2021) +52. Simeoni, O., Sekkat, C., Puy, G., Vobecky, A., Zablocki, É., Pérez, P.: Unsupervised object localization: Observing the background to discover objects. In: CVPR (2023) +53. Simeoni, O., Zablocki, E., Gidaris, S., Puy, G., Pérez, P.: Unsupervised object localization in the era of self-supervised vits: A survey. arXiv preprint arXiv:2310.12904 (2023) +54. Vobecký, A., Simeoni, O., Hurych, D., Gidaris, S., Bursuc, A., Perez, P., Sivic, J.: Pop-3d: Open-vocabulary 3d occupancy prediction from images. In: NeurIPS (2023) +55. Walmer, M., Suri, S., Gupta, K., Shrivastava, A.: Teaching matters: Investigating the role of supervision in vision transformers. In: CVPR (2023) +56. Wang, X., Girdhar, R., Yu, S.X., Misra, I.: Cut and learn for unsupervised object detection and instance segmentation. In: CVPR (2023) +57. Wang, Y., Shen, X., Hu, S.X., Yuan, Y., Crowley, J.L., Vaufreydaz, D.: Self-supervised transformers for unsupervised object discovery using normalized cut. In: CVPR (2022) +58. Wu, J., Li, X., Xu, S., Yuan, H., Ding, H., Yang, Y., Li, X., Zhang, J., Tong, Y., Jiang, X., et al.: Towards open vocabulary learning: A survey. T-PAMI (2024) + +59. Wysoczanska, M., Ramamonjisoa, M., Trzcinski, T., Simeoni, O.: Clip-diy: Clip dense inference yields open-vocabulary semantic segmentation for-free. In: WACV (2024) +60. Xian, Y., Choudhury, S., He, Y., Schiele, B., Akata, Z.: Semantic projection network for zero-and few-label semantic segmentation. In: CVPR (2019) +61. Xu, J., De Mello, S., Liu, S., Byeon, W., Breuel, T., Kautz, J., Wang, X.: Groupvit: Semantic segmentation emerges from text supervision. arXiv preprint arXiv:2202.11094 (2022) +62. Xu, J., Hou, J., Zhang, Y., Feng, R., Wang, Y., Qiao, Y., Xie, W.: Learning open-vocabulary semantic segmentation models from natural language supervision. In: CVPR (2023) +63. Zhai, X., Mustafa, B., Kolesnikov, A., Beyer, L.: Sigmoid loss for language image pre-training. In: ICCV (2023) +64. Zhao, H., Puig, X., Zhou, B., Fidler, S., Torralba, A.: Open vocabulary scene parsing. In: ICCV (2017) +65. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: CVPR (2022) +66. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. IJCV (2019) +67. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: ECCV (2022) +68. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: iBOT: Image bert pre-training with online tokenizer. In: ICLR (2022) \ No newline at end of file diff --git a/2024/CLIP-DINOiser_ Teaching CLIP a few DINO tricks for open-vocabulary semantic segmentation/images.zip b/2024/CLIP-DINOiser_ Teaching CLIP a few DINO tricks for open-vocabulary semantic segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3c51724724a13b91201781a4403b1f77424f1fd1 --- /dev/null +++ b/2024/CLIP-DINOiser_ Teaching CLIP a few DINO tricks for open-vocabulary semantic segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03fd955f6b9b5e3a8071310e8be6bb0199e16beb159018fd05584198d266f3d7 +size 551696 diff --git a/2024/CLIP-DINOiser_ Teaching CLIP a few DINO tricks for open-vocabulary semantic segmentation/layout.json b/2024/CLIP-DINOiser_ Teaching CLIP a few DINO tricks for open-vocabulary semantic segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..891a92cd2a5c85303a8611fafde829117bb30a70 --- /dev/null +++ b/2024/CLIP-DINOiser_ Teaching CLIP a few DINO tricks for open-vocabulary semantic segmentation/layout.json @@ -0,0 +1,9072 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "type": "text", + "content": "CLIP-DINOiser: Teaching CLIP a few DINO tricks for open-vocabulary semantic segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 151, + 168, + 462, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 168, + 462, + 194 + ], + "spans": [ + { + "bbox": [ + 151, + 168, + 462, + 194 + ], + "type": "text", + "content": "Monika Wysoczanska1, Oriane Simeoni2, Michael Ramamonjisoa3*, Andrei Bursuc2, Tomasz Trzciński1,4,5, and Patrick Pérez2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 137, + 201, + 477, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 201, + 477, + 214 + ], + "spans": [ + { + "bbox": [ + 137, + 201, + 477, + 214 + ], + "type": "text", + "content": "1Warsaw University of Technology, 2valeo.ai, 3Meta AI, 4Tooploox, 5IDEAS NCBR" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 246, + 455, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 246, + 455, + 488 + ], + "spans": [ + { + "bbox": [ + 160, + 246, + 455, + 488 + ], + "type": "text", + "content": "Abstract. The popular CLIP model displays impressive zero-shot capabilities thanks to its seamless interaction with arbitrary text prompts. However, its lack of spatial awareness makes it unsuitable for dense computer vision tasks, e.g., semantic segmentation, without an additional fine-tuning step that often uses annotations and can potentially suppress its original open-vocabulary properties. Meanwhile, self-supervised representation methods have demonstrated good localization properties without human-made annotations nor explicit supervision. In this work, we take the best of both worlds and propose an open-vocabulary semantic segmentation method, which does not require any annotations. We propose to locally improve dense MaskCLIP features, which are computed with a simple modification of CLIP's last pooling layer, by integrating localization priors extracted from self-supervised features. By doing so, we greatly improve the performance of MaskCLIP and produce smooth outputs. Moreover, we show that the used self-supervised feature properties can directly be learnt from CLIP features. Our method CLIP-DINOiser needs only a single forward pass of CLIP and two light convolutional layers at inference, no extra supervision nor extra memory and reaches state-of-the-art results on challenging and fine-grained benchmarks such as COCO, Pascal Context, Cityscapes and ADE20k. The code to reproduce our results is available at https://github.com/wysoczanska/clip_dinoiser." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 498, + 457, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 498, + 457, + 521 + ], + "spans": [ + { + "bbox": [ + 159, + 498, + 457, + 521 + ], + "type": "text", + "content": "Keywords: open-vocabulary semantic segmentation " + }, + { + "bbox": [ + 159, + 498, + 457, + 521 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 498, + 457, + 521 + ], + "type": "text", + "content": " self-supervised features " + }, + { + "bbox": [ + 159, + 498, + 457, + 521 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 498, + 457, + 521 + ], + "type": "text", + "content": " annotation-free segmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 547, + 231, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 231, + 560 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 231, + 560 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 574, + 485, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 485, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 485, + 635 + ], + "type": "text", + "content": "Semantic segmentation is a key visual perception task for many real-world systems, e.g., self-driving cars, and industrial robots. Typically tackled in a dataset-oriented manner, best methods require a training dataset which is manually annotated for a specific and finite set of classes. The advent of powerful Vision-Language Models (VLM) [24,43,63] is stimulating a shift from a closed-vocabulary" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 643, + 481, + 664 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 481, + 664 + ], + "type": "text", + "content": "* Work done outside of Meta and Meta was not involved in the research discussed here." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 483, + 286 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 483, + 286 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 483, + 286 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 483, + 286 + ], + "type": "image", + "image_path": "3c26a0e78ac5a532f064d0e9558cf8d9b1b7183e8f2abf666a5549b63b5f5285.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 293, + 482, + 392 + ], + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 392 + ], + "type": "text", + "content": "Fig. 1: Examples of open-vocabulary semantic segmentation results obtained with our method CLIP-DINOiser on 'in-the-wild' images vs. those of MaskCLIP [67]. Our method improves MaskCLIP features with a smart pooling strategy which does not alter the original open-vocabulary properties. We use self-supervised DINO [5] as a guide to teach CLIP [22] to produce DINO-like localization features through two light convolutional layers. Our method, which achieves state-of-the-art results, only requires a single forward pass through CLIP model and our two layers. In addition to the correct prompts (light grey row) we list the irrelevant prompts predicted (in yellow) that we query in all images shown here." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 422, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 555 + ], + "type": "text", + "content": "paradigm to an open-world one. Such models are trained with a simple but scalable objective: to align pairs of image and coarse text captions that can be obtained in large amounts with limited manual supervision. VLMs excel at associating global image content with arbitrary text inputs with remarkable generalization capabilities [17, 34], but struggle to provide dense open-vocabulary features [18, 67]. Obtaining such an alignment between pixels and language can lead to open-vocabulary extensions for multiple other modalities, such as point clouds [8, 23, 38, 41], 3D scenes [54], 3D shapes [1], radiance fields [27], intermodality alignment [19, 23], with multiple potential applications for which the construction of training datasets is even more challenging and where CLIP-derived models show promising results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Different strategies have been recently proposed towards improving CLIP's patch-level feature extraction abilities by modifying the original CLIP architecture for dense pooling and retraining [6, 37, 44, 61, 62] or finetuning on an annotated segmentation dataset with pre-defined classes [32, 67]. The former requires long training and/or large collections of annotated data, while the latter leads to an alteration of the vision-language associations of the CLIP features. An alternative line of approaches freezes the CLIP encoder and directly densifies its features with different heuristics, often with multiple forward passes [1, 23, 27, 49, 50, 59], but are less practical due to the extensive computa" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 255, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 255, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 255, + 101 + ], + "type": "text", + "content": "M.Wysoczanska et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "text", + "content": "tional overhead. MaskCLIP [67] arises as a computationally efficient dense CLIP extractor. It converts CLIP's global self-attention layer into a convolutional one to produce patch features with original vision-language qualities. If such features are local, they appear to be too noisy for high-quality segmentation mask extraction (see Fig. 2b middle column)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 260 + ], + "type": "text", + "content": "Meanwhile, recent self-supervised learning (SSL) approaches [4, 5, 9, 68] produce strong visual representations displaying object localization properties, and such without requiring any manual annotation. DINO [5] stands out with its semantically meaningful features which have been exploited for unsupervised object discovery [51, 52, 56, 57]. DINO features prove useful also for zero-shot semantic segmentation [25, 27, 59], but require expensive sliding window sampling [27, 59] or building concept-specific prototypes and ensemble strategies [25]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 261, + 481, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 481, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 481, + 427 + ], + "type": "text", + "content": "In this work, we aim for unaltered patch-level CLIP features with minimal runtime overhead. To this end, we re-examine the localization properties of MaskCLIP features and observe that it is possible to easily refine them with guidance from SSL models. In detail, we train a simple convolutional layer on unlabeled data to produce pooling weights to perform correlation-guided dense feature pooling from CLIP without distorting the vision-language alignment. This layer is optimized to mimic the patch correlations of DINO [5] that indicate likely layouts of visual concepts in the images. Furthermore, we show that the unsupervised objectness information given by FOUND [52] from DINO features can be also directly learned from CLIP features again in a fully-unsupervised fashion with a single convolutional layer and helps improve the segmentation of the ill-defined 'background' prompt. With CLIP-DINOiser, we obtain high-quality masks in a single forward pass on CLIP (see Fig. 1). CLIP-DINOiser is amenable to producing dense semantic maps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 428, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 428, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 428, + 482, + 571 + ], + "type": "text", + "content": "To summarize, our contributions are: (1) We propose a light pooling mechanism to refine MaskCLIP features by leveraging guidance from SSL features without degrading its original open-vocabulary properties. CLIP-DINOiser does not require any annotations, nor retraining CLIP from scratch, but only a single CLIP forward pass. (2) We show that CLIP already contains good localization properties which can be exploited. We leverage simple convolutional layers to emphasize visual concept layouts from dense CLIP features. We train them without any annotation on only 1k of raw images randomly sampled in ImageNet [13]. We believe that this finding could be further exploited in different contexts. (3) Our method achieves state-of-the-art results on complex semantic segmentation datasets such as COCO [3], Pascal Context [16], Cityscapes [11] and ADE20K [66]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 590, + 237, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 237, + 604 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 237, + 604 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "Zero-shot semantic segmentation. This task has been typically approached by methods which aim at generalizing from seen classes to unseen ones [2, 20, 21, 26, 30, 40, 60, 64]. Such strategies train models with full supervision on the set of seen classes and propose different solutions to extend them to unseen ones with" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 384, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 384, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLIP-DINOiser" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "out new images (labeled or unlabeled), e.g., by exploiting class information and relationships encapsulated in popular word embeddings [35,42]. While they produce fine segmentations without computational overhead, these methods require pixel-level annotations for the seen classes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 175, + 482, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 482, + 379 + ], + "type": "text", + "content": "From CLIP to open-vocabulary segmentation. The surge of VLMs with aligned image-language representations [22, 24, 43] brought back into the spotlight the zero-shot classification task. However, the extension to zero-shot segmentation is not obvious as the CLIP architecture is not equipped to yield dense vision-language features [18, 67]. To produce dense CLIP features, several approaches fine-tune or train from scratch pixel-aligned CLIP-like models with additional modules, mechanisms or supervision objectives [6, 37, 44, 61, 62] on datasets with annotations of varying granularity and quality: dense annotations [29, 31], class-agnostic object masks [14, 18, 45], coarse captions [6, 18, 31–33, 37, 44, 61, 62, 65] or pseudo-labels [67]. Recent works leverage image-level captions to align text to regions (obtained without supervision): PACL [37] trains an embedder module to learn patch-to-text affinity, TCL [6] proposes a local contrastive objective to align well-selected patches to the text and ViewCO [46] leverages multi-view consistency. On the downside, such models require long training on millions of images or specific types of very costly annotations. Also, fine-tuning CLIP with a defined vocabulary is more computationally appealing [29, 31, 67], but alters the open-vocabulary properties of the features [23]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 380, + 483, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 483, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 483, + 462 + ], + "type": "text", + "content": "Most related to us is a line of works that investigate how to directly densify CLIP features [1, 23, 27, 59, 67] to obtain per-patch CLIP features. Such densification can be performed by aggregating features from multiple views [1, 27] or from sliding windows [23, 59] at the extra-cost of multiple forward passes. MaskCLIP [67] drops the global pooling layer of CLIP and matches the projected features directly to text via a " + }, + { + "bbox": [ + 130, + 380, + 483, + 462 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 380, + 483, + 462 + ], + "type": "text", + "content": " convolution layer. By doing so they achieve dense predictions, however noisy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 463, + 483, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 483, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 483, + 559 + ], + "type": "text", + "content": "With a concept-driven perspective, some methods [25,49,50] build codebooks of visual prototypes per concept, including negative prototypes [25], and then perform co-segmentation [49]. While such an approach yields good results, it is however at the cost of building expensive class-specific prototypes, therefore diverging from open-vocabulary scenarios. Instead, we aim to remain open to avoid retraining a model or building new expensive prototypes whenever a new concept is considered. To that end, we devise a dense CLIP-feature extraction method that preserves the open-vocabulary quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Leveraging self-supervised models & CLIP. Recent self-supervised ViTs [4,5,9,12, 68] have demonstrated features with good localization properties [51,52,56,57]. Such features have also been exploited in the context of open-vocabulary segmentation methods, e.g. for pre-training for the visual backbone [7,44,62], cosegmentation [49], clustering patches into masks [47], representing object prototypes [25]. Related to us is the recent CLIP-DIY [59] which computes patch-level representations from CLIP features from different image crops with guidance from an unsupervised saliency segmenter [52] FOUND. While we also leverage" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "type": "text", + "content": "M.Wysoczanska et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "the latter, in contrast with CLIP-DIY which runs multiple forward passes to build their dense CLIP features, our method requires only a single forward pass of CLIP. Furthermore, our method mitigates the limits of FOUND in cluttered scenarios by integrating an uncertainty constraint. Finally, we leverage the informative patch correlation properties of DINO [5] and show that it is possible to teach CLIP to produce DINO-like features through light convolutional layers." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 207, + 203, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 207, + 203, + 220 + ], + "spans": [ + { + "bbox": [ + 131, + 207, + 203, + 220 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 233, + 482, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 233, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 482, + 340 + ], + "type": "text", + "content": "We present in this section CLIP-DINOiser, a simple and efficient strategy to improve MaskCLIP using localization information extracted from CLIP—with a lightweight model trained to mimic some of DINO's properties. We first set the goal in Sec. 3.1 and present MaskCLIP [67] in Sec. 3.2. We then introduce our strategy which leverages self-supervised features localization information to consolidate MaskCLIP features in Sec. 3.3 and discuss how such localization information can directly be learnt from CLIP in Sec. 3.4 (we visualize both steps in Fig. 3). We also propose a way to improve the 'background' filtering in Sec. 3.5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 359, + 258, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 359, + 258, + 370 + ], + "spans": [ + { + "bbox": [ + 131, + 359, + 258, + 370 + ], + "type": "text", + "content": "3.1 Problem statement" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 380, + 483, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 483, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 483, + 477 + ], + "type": "text", + "content": "In this work, we aim to produce open-vocabulary1 semantic segmentation of an image. We consider an image " + }, + { + "bbox": [ + 130, + 380, + 483, + 477 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 380, + 483, + 477 + ], + "type": "text", + "content": " which we split into a sequence of " + }, + { + "bbox": [ + 130, + 380, + 483, + 477 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 380, + 483, + 477 + ], + "type": "text", + "content": " patches of dimensions " + }, + { + "bbox": [ + 130, + 380, + 483, + 477 + ], + "type": "inline_equation", + "content": "P \\times P \\times 3" + }, + { + "bbox": [ + 130, + 380, + 483, + 477 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 380, + 483, + 477 + ], + "type": "inline_equation", + "content": "P \\times P" + }, + { + "bbox": [ + 130, + 380, + 483, + 477 + ], + "type": "text", + "content": " the patch size and " + }, + { + "bbox": [ + 130, + 380, + 483, + 477 + ], + "type": "inline_equation", + "content": "N = \\lceil \\frac{H}{P} \\rceil \\cdot \\lceil \\frac{W}{P} \\rceil" + }, + { + "bbox": [ + 130, + 380, + 483, + 477 + ], + "type": "text", + "content": ". A class token, noted CLS, is added to the input sequence and we feed the " + }, + { + "bbox": [ + 130, + 380, + 483, + 477 + ], + "type": "inline_equation", + "content": "N + 1" + }, + { + "bbox": [ + 130, + 380, + 483, + 477 + ], + "type": "text", + "content": " patches to a ViT [15] model. We aim at producing dense visual features " + }, + { + "bbox": [ + 130, + 380, + 483, + 477 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 130, + 380, + 483, + 477 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 380, + 483, + 477 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 380, + 483, + 477 + ], + "type": "text", + "content": " the feature dimension, that can later be matched to any set of text inputs embedded in the same space. In particular, the goal is to produce a segmentation map per textual query." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 495, + 302, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 495, + 302, + 506 + ], + "spans": [ + { + "bbox": [ + 131, + 495, + 302, + 506 + ], + "type": "text", + "content": "3.2 Preliminaries on MaskCLIP" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 514, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 635 + ], + "type": "text", + "content": "Extracting dense open-vocabulary features. The popular CLIP [22] model pretrained on image/caption pairs produces good global image features, but was not trained to generate high-quality 2D feature maps. In order to extract such dense feature maps relevant to semantic segmentation, Zhou et al. [67] revisit the global attention pooling layer of the last attention layer of the model. The authors discard the query and key embeddings of the layer and transform both the value projection and the last linear layer into a conv " + }, + { + "bbox": [ + 130, + 514, + 482, + 635 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 514, + 482, + 635 + ], + "type": "text", + "content": " layer. With this new model, named MaskCLIP and denoted " + }, + { + "bbox": [ + 130, + 514, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot)" + }, + { + "bbox": [ + 130, + 514, + 482, + 635 + ], + "type": "text", + "content": ", we extract " + }, + { + "bbox": [ + 130, + 514, + 482, + 635 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 514, + 482, + 635 + ], + "type": "text", + "content": "-dimensional features " + }, + { + "bbox": [ + 130, + 514, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\phi^L(X) \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 130, + 514, + 482, + 635 + ], + "type": "text", + "content": " from the last layer " + }, + { + "bbox": [ + 130, + 514, + 482, + 635 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 514, + 482, + 635 + ], + "type": "text", + "content": " which retains most of the open-vocabulary properties of CLIP [67]." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 384, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 384, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLIP-DINOiser" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "text", + "content": "1 We adopt the taxonomy defined in the recent survey [58] and define our method as 'open-vocabulary', with capabilities to generalize to unseen datasets." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "Semantic segmentation given textual queries. We also extract CLIP textual features " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\phi_T(t_j)" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " for each text query " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "t_j \\in \\mathcal{T}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "j \\in \\{1, \\dots, |\\mathcal{T}|\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ". Segmentation maps are then generated by computing the cosine similarity between each of the visual patch features and of the textual prompts, after L2-normalization. The most similar prompt is assigned to each patch. Note that a query 'background' can be added in order to obtain negative patches. Using MaskCLIP allows us to produce dense segmentation maps with a single forward pass of the classic CLIP model, but its outputs are noisy, as visible in Fig. 2b (middle column)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 222, + 346, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 222, + 346, + 235 + ], + "spans": [ + { + "bbox": [ + 131, + 222, + 346, + 235 + ], + "type": "text", + "content": "3.3 DINOising open-vocabulary features" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 239, + 481, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 481, + 276 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 481, + 276 + ], + "type": "text", + "content": "In this work, we aim to improve MaskCLIP's open-vocabulary features described above. To do so, we propose to leverage the known good localization properties of self-supervised features [5,39,51-53,57]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "text", + "content": "Extracting self-supervised correlation information. Recent works [51, 57] have shown that the patch correlation information of the embeddings from the last attention layer of the self-supervised model, DINO [5] can help highlight objects in images. We use here the value embeddings which we observe have finer correlation than those of key and query (more discussion in supplementary material). We extract such self-supervised features " + }, + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\xi(X) \\in \\mathbb{R}^{N \\times d_{\\xi}}" + }, + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "text", + "content": " and discard the CLS token. We then compute the per-patch cosine-similarity and produce the affinity map " + }, + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "inline_equation", + "content": "A^{\\xi} \\in [-1,1]^{N \\times N}" + }, + { + "bbox": [ + 130, + 284, + 482, + 416 + ], + "type": "text", + "content": ". We compare in Fig. 4 the patch-similarities obtained for a patch seed with MaskCLIP and DINO features and observe that the self-supervised features are more densely and accurately correlated than those of CLIP." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 138, + 431, + 310, + 552 + ], + "blocks": [ + { + "bbox": [ + 138, + 431, + 310, + 552 + ], + "lines": [ + { + "bbox": [ + 138, + 431, + 310, + 552 + ], + "spans": [ + { + "bbox": [ + 138, + 431, + 310, + 552 + ], + "type": "image", + "image_path": "0729da00fee285f570f825a5d25997b27a4bcc703e4cc49a29d518b57c54c55d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 553, + 266, + 563 + ], + "lines": [ + { + "bbox": [ + 182, + 553, + 266, + 563 + ], + "spans": [ + { + "bbox": [ + 182, + 553, + 266, + 563 + ], + "type": "text", + "content": "(a) Our guided pooling" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 319, + 437, + 478, + 552 + ], + "blocks": [ + { + "bbox": [ + 319, + 437, + 478, + 552 + ], + "lines": [ + { + "bbox": [ + 319, + 437, + 478, + 552 + ], + "spans": [ + { + "bbox": [ + 319, + 437, + 478, + 552 + ], + "type": "image", + "image_path": "9b6982e578b1649eb765a836d24013c91ea8b704e9ba97f8b52304b57d6020ee.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 553, + 444, + 563 + ], + "lines": [ + { + "bbox": [ + 352, + 553, + 444, + 563 + ], + "spans": [ + { + "bbox": [ + 352, + 553, + 444, + 563 + ], + "type": "text", + "content": "(b) Impact of the pooling" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 571, + 482, + 616 + ], + "lines": [ + { + "bbox": [ + 130, + 571, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 571, + 482, + 616 + ], + "type": "text", + "content": "Fig. 2: We present in (a) our guided pooling strategy defined in Eq. (1). The " + }, + { + "bbox": [ + 130, + 571, + 482, + 616 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 130, + 571, + 482, + 616 + ], + "type": "text", + "content": " affinity matrix is computed from patch features and is used to refine MaskCLIP features (bottom left). In (b) we compare our results with " + }, + { + "bbox": [ + 130, + 571, + 482, + 616 + ], + "type": "inline_equation", + "content": "F^{+}" + }, + { + "bbox": [ + 130, + 571, + 482, + 616 + ], + "type": "text", + "content": " (right) versus those obtained with MaskCLIP features (middle)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 495, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 495, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 495, + 666 + ], + "type": "text", + "content": "Strengthening features with guided pooling. In order to locally consolidate MaskCLIP features " + }, + { + "bbox": [ + 130, + 629, + 495, + 666 + ], + "type": "inline_equation", + "content": "\\phi^L (X)" + }, + { + "bbox": [ + 130, + 629, + 495, + 666 + ], + "type": "text", + "content": ", now noted " + }, + { + "bbox": [ + 130, + 629, + 495, + 666 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 629, + 495, + 666 + ], + "type": "text", + "content": ", we propose to perform a concept-aware linear combination of the features per patch with guidance from the patch affinity " + }, + { + "bbox": [ + 130, + 629, + 495, + 666 + ], + "type": "inline_equation", + "content": "A^{\\xi}" + }, + { + "bbox": [ + 130, + 629, + 495, + 666 + ], + "type": "text", + "content": ". The" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 255, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 255, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 255, + 101 + ], + "type": "text", + "content": "M.Wysoczanska et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "feature combination strategy can be seen as a form of voting mechanism that enforces similar patches to have similar CLIP features (and prediction) while attenuating noisy features. Specifically, we compute the new features " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "F^{+} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " as an average of MaskCLIP features " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " weighted by " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "A^{\\xi}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ", presented in Fig. 2a. We zero-out " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "A^{\\xi}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " correlations below a threshold " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ", following [51, 57], and compute the new features for patch " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "p \\in \\{1, \\dots, N\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 238, + 194, + 481, + 227 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 194, + 481, + 227 + ], + "spans": [ + { + "bbox": [ + 238, + 194, + 481, + 227 + ], + "type": "interline_equation", + "content": "F _ {p} ^ {+} = \\frac {1}{\\sum_ {q = 1} ^ {N} A _ {p , q} ^ {\\xi}} \\sum_ {q = 1} ^ {N} A _ {p, q} ^ {\\xi} \\cdot F _ {q}. \\tag {1}", + "image_path": "c875c7f4affe2195de408ff424c159a5b7389652b53fc18dd8e5d26ae960264c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 234, + 482, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 234, + 482, + 294 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 482, + 294 + ], + "type": "text", + "content": "We then produce the segmentation maps " + }, + { + "bbox": [ + 130, + 234, + 482, + 294 + ], + "type": "inline_equation", + "content": "S \\in [-1,1]^{N \\times |\\mathcal{T}|}" + }, + { + "bbox": [ + 130, + 234, + 482, + 294 + ], + "type": "text", + "content": ", by comparing the new features " + }, + { + "bbox": [ + 130, + 234, + 482, + 294 + ], + "type": "inline_equation", + "content": "F^+" + }, + { + "bbox": [ + 130, + 234, + 482, + 294 + ], + "type": "text", + "content": " to each textual queries in " + }, + { + "bbox": [ + 130, + 234, + 482, + 294 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 234, + 482, + 294 + ], + "type": "text", + "content": ". As shown in Fig. 2b, when using such consolidated features, we obtain more accurate outputs and the high-frequency predictions observed in MaskCLIP are smoothed out, showing the benefit of the pooling." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 309, + 432, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 309, + 432, + 322 + ], + "spans": [ + { + "bbox": [ + 131, + 309, + 432, + 322 + ], + "type": "text", + "content": "3.4 Teaching CLIP a first DINO trick: object correlations" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 136, + 347, + 480, + 507 + ], + "blocks": [ + { + "bbox": [ + 136, + 347, + 480, + 507 + ], + "lines": [ + { + "bbox": [ + 136, + 347, + 480, + 507 + ], + "spans": [ + { + "bbox": [ + 136, + 347, + 480, + 507 + ], + "type": "image", + "image_path": "716ea1effc3d6881eac5b1b729f0220ba496b5e336bfe34c39d3b5c2bb8fcf30.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 520, + 482, + 619 + ], + "lines": [ + { + "bbox": [ + 130, + 520, + 482, + 619 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 482, + 619 + ], + "type": "text", + "content": "Fig. 3: Overview of CLIP-DINOiser which leverages the quality of self-supervised features to improve the notoriously noisy MaskCLIP feature maps. We use DINO as a teacher which 'teaches' CLIP how to extract localization information. We train (left) a conv " + }, + { + "bbox": [ + 130, + 520, + 482, + 619 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 520, + 482, + 619 + ], + "type": "text", + "content": " layer to reproduce the patch correlations obtained with DINO. At inference (right), an input image is forwarded through the frozen CLIP image backbone and MaskCLIP projection. The produced features are then improved with our pooling strategy which is guided by correlations predicted with the trained convolutional layer applied on CLIP. With this light 'DINOising' process, we obtain 'DINOised' features which are matched against the prompts features to produce CLIP-DINOiser outputs." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "We have shown in the previous section that self-supervised correlation information can successfully be used to improve the dense quality of open-vocabulary" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 384, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 384, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLIP-DINOiser" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "features. If the difficulty of densifying CLIP is well-known, we show here that CLIP features already contain good localization information which can be extracted with a light model. We indeed predict DINO correlations " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "A^{\\xi}" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " from CLIP with a single convolutional layer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 482, + 213 + ], + "type": "text", + "content": "In order to predict the DINO affinity map " + }, + { + "bbox": [ + 130, + 164, + 482, + 213 + ], + "type": "inline_equation", + "content": "A^{\\xi}" + }, + { + "bbox": [ + 130, + 164, + 482, + 213 + ], + "type": "text", + "content": " from CLIP features, we train a single " + }, + { + "bbox": [ + 130, + 164, + 482, + 213 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 164, + 482, + 213 + ], + "type": "text", + "content": " convolutional layer " + }, + { + "bbox": [ + 130, + 164, + 482, + 213 + ], + "type": "inline_equation", + "content": "g(\\cdot): \\mathbb{R}^{d} \\to \\mathbb{R}^{d_{g}}" + }, + { + "bbox": [ + 130, + 164, + 482, + 213 + ], + "type": "text", + "content": " which projects intermediate features " + }, + { + "bbox": [ + 130, + 164, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\phi^{l}(X)" + }, + { + "bbox": [ + 130, + 164, + 482, + 213 + ], + "type": "text", + "content": "-extracted from layer " + }, + { + "bbox": [ + 130, + 164, + 482, + 213 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 164, + 482, + 213 + ], + "type": "text", + "content": "-into a smaller space of dimension " + }, + { + "bbox": [ + 130, + 164, + 482, + 213 + ], + "type": "inline_equation", + "content": "d_{g} < d" + }, + { + "bbox": [ + 130, + 164, + 482, + 213 + ], + "type": "text", + "content": ". We enforce the patch correlations of the generated features " + }, + { + "bbox": [ + 130, + 164, + 482, + 213 + ], + "type": "inline_equation", + "content": "A^{\\phi} \\in [-1,1]^{N \\times N}" + }, + { + "bbox": [ + 130, + 164, + 482, + 213 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 221, + 220, + 481, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 220, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 221, + 220, + 481, + 251 + ], + "type": "interline_equation", + "content": "A ^ {\\phi} = \\frac {g \\left(\\phi^ {l} (X)\\right)}{\\| g \\left(\\phi^ {l} (X)\\right) \\|} \\otimes \\left(\\frac {g \\left(\\phi^ {l} (X)\\right)}{\\| g \\left(\\phi^ {l} (X)\\right) \\|}\\right) ^ {\\top}, \\tag {2}", + "image_path": "658ba00d9a4f0e95c61b128f58e6f2a9f4778deb70f1ca556ddb893fe428269f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 258, + 481, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 258, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 131, + 258, + 481, + 293 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 131, + 258, + 481, + 293 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 131, + 258, + 481, + 293 + ], + "type": "text", + "content": " denoting the outer product, to be close to the binarized correlations " + }, + { + "bbox": [ + 131, + 258, + 481, + 293 + ], + "type": "inline_equation", + "content": "D = A^{\\xi} > \\gamma" + }, + { + "bbox": [ + 131, + 258, + 481, + 293 + ], + "type": "text", + "content": " (we use here the same " + }, + { + "bbox": [ + 131, + 258, + 481, + 293 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 258, + 481, + 293 + ], + "type": "text", + "content": " as defined above), using the binary cross-entropy loss " + }, + { + "bbox": [ + 131, + 258, + 481, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^c" + }, + { + "bbox": [ + 131, + 258, + 481, + 293 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 206, + 302, + 481, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 302, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 206, + 302, + 481, + 336 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {c} = \\sum_ {p = 1} ^ {N} \\left[ D _ {p} \\log A _ {p} ^ {\\phi} + (1 - D _ {p}) \\log \\left(1 - A _ {p} ^ {\\phi}\\right) \\right]. \\tag {3}", + "image_path": "789235dee31017216997b155f8af7012269dd12c38c06d3b821aaa0b8168c1e1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 345, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 482, + 453 + ], + "type": "text", + "content": "We present our layer training in Fig. 3 (left part) and observe the quality of CLIP-predicted affinity matrix " + }, + { + "bbox": [ + 130, + 345, + 482, + 453 + ], + "type": "inline_equation", + "content": "A^{\\phi}" + }, + { + "bbox": [ + 130, + 345, + 482, + 453 + ], + "type": "text", + "content": ". We also show in Fig. 4 another example of obtained " + }, + { + "bbox": [ + 130, + 345, + 482, + 453 + ], + "type": "inline_equation", + "content": "A^{\\phi}" + }, + { + "bbox": [ + 130, + 345, + 482, + 453 + ], + "type": "text", + "content": " and observe their similarity to DINO-based correlations. We use the CLIP-produced correlations " + }, + { + "bbox": [ + 130, + 345, + 482, + 453 + ], + "type": "inline_equation", + "content": "A^{\\phi}" + }, + { + "bbox": [ + 130, + 345, + 482, + 453 + ], + "type": "text", + "content": " to replace " + }, + { + "bbox": [ + 130, + 345, + 482, + 453 + ], + "type": "inline_equation", + "content": "A^{\\xi}" + }, + { + "bbox": [ + 130, + 345, + 482, + 453 + ], + "type": "text", + "content": " in Eq. (1) to weight the pooling and observe a similar boost over MaskCLIP, thus showing that good patch correlations can indeed be extracted directly from CLIP. We can now discard DINO and we name CLIP-DINOiser the guided-pooling strategy which uses CLIP-based correlation. As shown in Fig. 3 (inference step), our method runs with a single forward pass of CLIP model and a small extra layer." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 136, + 468, + 478, + 524 + ], + "blocks": [ + { + "bbox": [ + 136, + 468, + 478, + 524 + ], + "lines": [ + { + "bbox": [ + 136, + 468, + 478, + 524 + ], + "spans": [ + { + "bbox": [ + 136, + 468, + 478, + 524 + ], + "type": "image", + "image_path": "f70992381d5ab9a8d6f8a4a3b46d5082dbda506ab7a5d0d2bfc451c3c6c1ec6f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 525, + 482, + 559 + ], + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 559 + ], + "type": "text", + "content": "Fig. 4: Comparison of the affinity maps between a seed (one on the 'plant' and the other on a 'pillow') and the other patch features when using features of MaskCLIP, DINO and ours after training." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 586, + 454, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 454, + 598 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 454, + 598 + ], + "type": "text", + "content": "3.5 Teaching CLIP a second DINO trick: background filtering" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Moreover, as discussed earlier, a 'background' query may be added to the set of textual queries " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " in order to help filter out patches falling in the background and not corresponding to any objects. We do not assume here any prior knowledge about classes of interest and focus rather on the foreground/background paradigm [52]. We argue that relying solely on the textual prompt 'background'" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 255, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 255, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 255, + 101 + ], + "type": "text", + "content": "M.Wysoczanska et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 135, + 286, + 246 + ], + "blocks": [ + { + "bbox": [ + 140, + 135, + 286, + 246 + ], + "lines": [ + { + "bbox": [ + 140, + 135, + 286, + 246 + ], + "spans": [ + { + "bbox": [ + 140, + 135, + 286, + 246 + ], + "type": "image", + "image_path": "e6f1ef85e944166ce6e14abc8f30d5fa2df07c5a28b8733a85ea850f7e14f0af.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 247, + 282, + 257 + ], + "lines": [ + { + "bbox": [ + 143, + 247, + 282, + 257 + ], + "spans": [ + { + "bbox": [ + 143, + 247, + 282, + 257 + ], + "type": "text", + "content": "(a) Comparison of background filtering" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 294, + 114, + 470, + 246 + ], + "blocks": [ + { + "bbox": [ + 294, + 114, + 470, + 246 + ], + "lines": [ + { + "bbox": [ + 294, + 114, + 470, + 246 + ], + "spans": [ + { + "bbox": [ + 294, + 114, + 470, + 246 + ], + "type": "image", + "image_path": "838aa25cde5bf5f63a9a16f11c3540528c48854aef5d5863b23e3cc6e50cc42c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 247, + 432, + 257 + ], + "lines": [ + { + "bbox": [ + 343, + 247, + 432, + 257 + ], + "spans": [ + { + "bbox": [ + 343, + 247, + 432, + 257 + ], + "type": "text", + "content": "(b) Background filtering" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 266, + 482, + 321 + ], + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 321 + ], + "type": "text", + "content": "Fig. 5: We present in (a) a comparison of objectness mask generated by FOUND [52] and with our layer using CLIP features. We carefully define the fusion operation and the simple training strategy of the conv " + }, + { + "bbox": [ + 130, + 266, + 482, + 321 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 266, + 482, + 321 + ], + "type": "text", + "content": " again using DINO as a teacher in Sec. 3.5. In (b) is an overview of our background filtering which is applied when a 'background' prompt is provided and helps reduce hallucinations." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 343, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 489 + ], + "type": "text", + "content": "to catch all non-salient patches is underperforming and, similarly to [59], we propose to use a very light-weight unsupervised foreground/background segmentation method, namely FOUND [52] which also relies on DINO self-supervised features. We run FOUND on the entire image and extract a prediction mask " + }, + { + "bbox": [ + 130, + 343, + 482, + 489 + ], + "type": "inline_equation", + "content": "M \\in \\{0,1\\}^N" + }, + { + "bbox": [ + 130, + 343, + 482, + 489 + ], + "type": "text", + "content": " in which a patch is assigned the value 1 if falling into the foreground and 0 otherwise. We also observe that saliencies produced by FOUND can be too restrictive and discard objects which are partially visible or in a clutter. To mitigate this behaviour, we propose to relax the background selection by integrating an additional uncertainty constraint. To this end, we fuse the background information from both modalities by assigning the 'background' prompt to patches " + }, + { + "bbox": [ + 130, + 343, + 482, + 489 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 343, + 482, + 489 + ], + "type": "text", + "content": " which are both uncertain, e.g. have low confidence score " + }, + { + "bbox": [ + 130, + 343, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\sigma(S)_p < \\delta" + }, + { + "bbox": [ + 130, + 343, + 482, + 489 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 343, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 130, + 343, + 482, + 489 + ], + "type": "text", + "content": " the softmax operation, and which fall in the background in " + }, + { + "bbox": [ + 130, + 343, + 482, + 489 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 343, + 482, + 489 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "text", + "content": "Learning FOUND objectness. Moreover, we are also able to learn the predictions of FOUND [52] directly from CLIP features. To do so, we train a single " + }, + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "text", + "content": " convolutional layer " + }, + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "inline_equation", + "content": "h(\\cdot): \\mathbb{R}^d \\to \\mathbb{R}" + }, + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "text", + "content": " which predicts from the features " + }, + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\phi^l(X)" + }, + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "text", + "content": " an objectness map " + }, + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "inline_equation", + "content": "M^\\phi = h(\\phi^l(X)) \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "text", + "content": ". We train the model to predict the FOUND binary mask " + }, + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "text", + "content": " with the binary cross-entropy loss " + }, + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^m" + }, + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 196, + 565, + 416, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 565, + 416, + 598 + ], + "spans": [ + { + "bbox": [ + 196, + 565, + 416, + 598 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {m} = \\sum_ {p = 1} ^ {N} \\left[ M _ {p} \\log \\left(M _ {p} ^ {\\phi}\\right) + \\left(1 - M _ {p}\\right) \\log \\left(1 - M _ {p} ^ {\\phi}\\right) \\right].", + "image_path": "86c88fac6c97415e5e01033a760dc969d698fa0858555afec2171a79e3cd4702.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "We show examples of predicted CLIP-based objectness in Fig. 5a and observe their very high similarity to those produced with DINO. Moreover, we can now replace " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " defined above with the binarized CLIP-based scores " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\zeta(M^{\\phi}) > 0.5" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\zeta(\\cdot)" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " the sigmoid operation, and observe a minimal drop in performances. We show an example of the background filtering with trained objectness in Fig. 5b." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 384, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 384, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLIP-DINOiser" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 177 + ], + "type": "text", + "content": "We detail in Sec. 4.1 the experimental setup used in our evaluation. We produce state-of-the-art results on the task of open-vocabulary semantic segmentation in Sec. 4.2 and ablation studies in Sec. 4.3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 196, + 260, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 260, + 209 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 260, + 209 + ], + "type": "text", + "content": "4.1 Experimental setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 483, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 483, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 483, + 445 + ], + "type": "text", + "content": "Technical details. We use in all experiments a frozen CLIP ViT-B/16 pre-trained following OpenCLIP [22]. Our method CLIP-DINOiser uses two convolutional layers to extract DINO-like information from CLIP layer " + }, + { + "bbox": [ + 130, + 216, + 483, + 445 + ], + "type": "inline_equation", + "content": "l = 10" + }, + { + "bbox": [ + 130, + 216, + 483, + 445 + ], + "type": "text", + "content": " (the 3rd before the last which was shown to provide the best results [55]). The first layer " + }, + { + "bbox": [ + 130, + 216, + 483, + 445 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 130, + 216, + 483, + 445 + ], + "type": "text", + "content": " has a kernel " + }, + { + "bbox": [ + 130, + 216, + 483, + 445 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 216, + 483, + 445 + ], + "type": "text", + "content": " and output dimension " + }, + { + "bbox": [ + 130, + 216, + 483, + 445 + ], + "type": "inline_equation", + "content": "d_{g} = 256" + }, + { + "bbox": [ + 130, + 216, + 483, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 216, + 483, + 445 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 130, + 216, + 483, + 445 + ], + "type": "text", + "content": " a kernel " + }, + { + "bbox": [ + 130, + 216, + 483, + 445 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 216, + 483, + 445 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 216, + 483, + 445 + ], + "type": "inline_equation", + "content": "d_{h} = 1" + }, + { + "bbox": [ + 130, + 216, + 483, + 445 + ], + "type": "text", + "content": ". The first is trained to match the correlation information extracted from the value embeddings of the last layer of a ViT-B/16 model trained following DINO [5]. The second layer is trained to replicate the unsupervised object localization predictions of FOUND [52]–which also uses DINO model. We train both layers with a binary cross-entropy loss on only 1k raw images randomly sampled from ImageNet [13] dataset without any annotation. We report average scores over 3 runs with different sampling seeds and provide standard deviations in the supplementary material. We follow [57] and binarize the correlations with " + }, + { + "bbox": [ + 130, + 216, + 483, + 445 + ], + "type": "inline_equation", + "content": "\\gamma = 0.2" + }, + { + "bbox": [ + 130, + 216, + 483, + 445 + ], + "type": "text", + "content": ". In the background filtering step, we use a high confidence score, i.e., " + }, + { + "bbox": [ + 130, + 216, + 483, + 445 + ], + "type": "inline_equation", + "content": "\\delta = 0.99" + }, + { + "bbox": [ + 130, + 216, + 483, + 445 + ], + "type": "text", + "content": ". We train our model for 6k iterations with a batch size of 16 images using Adam optimizer [28], which takes approximately 3 hours on a single NVIDIA RTX A5000 GPU. We decrease the learning rate for both heads by a factor of 0.1 after 5k iterations. We apply data augmentations during training (random scale and cropping, flipping and photometric distortions)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 452, + 483, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 452, + 483, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 452, + 483, + 596 + ], + "type": "text", + "content": "Datasets and metric. We evaluate our method on eight benchmarks typically used for zero-shot semantic segmentation [6]. Following [6], we split them into two groups. The first consists in datasets with a 'background' query: PASCAL VOC [16] (noted 'VOC'), PASCAL Context [36] (noted 'Context'), and COCO Object [3] (noted 'Object') and the second without: PASCAL VOC20 [16] (noted 'VOC20'), PASCAL Context59 [36] (noted 'C59'), COCO-Stuff [3] (noted 'Stuff'), Cityscapes [11] (noted 'City'), and ADE20K [66] (noted 'ADE'). We evaluate results with the standard mIoU metric. We also follow the evaluation protocol of [6], use the implementations provided by MMSegmentation [10], employ a sliding window strategy, resize the input image to have a shorter side of 448. We also do not perform text expansions of the class names and use only the standard ImageNet prompts following [22, 61, 67]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Baselines. We compare our method against state-of-the-art methods on open-vocabulary zero-shot semantic segmentation. For a fair comparison between methods, we report results without any post-processing step. In our evaluations, we follow the taxonomy presented in [58] and compare our model with the methods relying on language-image pretraining, also called open-vocabulary." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "text", + "content": "M.Wysoczanska et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "We split the compared baselines into four categories: (1) dataset specific which employ pseudo-labeling and supervised training of a segmentation model on target dataset: NamedMask [50], MaskCLIP + [67]); (2) construct prototypes: ReCO [49], OVDiff [25]; (3) train with text supervision including GroupViT [61], ZeroSeg [47], SegCLIP [33], TCL [6], CLIPpy [44], OVSegmentor [62], which all require access to additional datasets of millions of image/caption pairs (we note in the table the exact datasets used for the training); and finally use frozen CLIP i.e. CLIP-DIY [59] and MaskCLIP [67], which use pre-trained CLIP. Our method falls into the last category as we do not modify CLIP, and do not need access to additional caption annotations as we use only 1k unannotated images." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 134, + 260, + 479, + 495 + ], + "blocks": [ + { + "bbox": [ + 134, + 260, + 479, + 495 + ], + "lines": [ + { + "bbox": [ + 134, + 260, + 479, + 495 + ], + "spans": [ + { + "bbox": [ + 134, + 260, + 479, + 495 + ], + "type": "table", + "html": "
MethodsConcept spec.Extra Inference data backboneNo background promptW/ bkg prompt Cont. Object VOC
VOC20C59StuffCityADE
Dataset specific
MaskCLIP+ [67]✓XIDLv2-31.118.0----
NamedMask [50]✓XIDLv3+-----27.759.2
Build prototypes per visual concept
ReCo [49]✓✓ICLIP57.822.314.821.111.219.915.7
OVDiff [25]✓✓XCLIP+ DINO+SD81.733.7--14.930.134.8
Text/image alignment training with captions
GroupViT [61] [6]XXITCLIP79.723.415.311.19.218.727.5
ZeroSeg [7]XXITCLIP-----21.822.1
SegCLIP [33]XXITCLIP---11.08.724.726.5
TCL [6]XXITCLIP77.530.319.623.114.924.330.4
CLIPpy [44]XXITCLIP----13.5-32.0
OVSegmentor [62]XXITCLIP----5.620.425.1
Frozen CLIP
CLIP-DIY [59]*X✓XCLIP+ DINO79.719.813.311.69.919.731.0
MaskCLIP [67] [6]X✓XCLIP53.723.314.721.610.821.115.5
MaskCLIP*X✓XCLIP61.825.617.625.014.322.916.4
MaskCLIP* †X✓XCLIP71.927.418.623.014.924.021.6
CLIP-DINOiserX✓I(1k)CLIP80.935.924.631.720.032.434.8
", + "image_path": "9afa16e34166ebaead84cea2ede83a357d4435d11cb091eea252e9d6299a4282.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 504, + 482, + 647 + ], + "lines": [ + { + "bbox": [ + 130, + 504, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 482, + 647 + ], + "type": "text", + "content": "Table 1: Open-vocabulary semantic segmentation quantitative comparison using the mIoU metric. We separate in two columns the evaluation datasets: those without a 'background' prompt and those with (noted 'W/ bkg prompt'), as discussed in Sec. 4.1. We report all methods without post-processing. We note with * methods for which we computed scores; we obtained MaskCLIP* scores with OpenCLIP [22] and mark with † the use of MaskCLIP refinement. The first and second best methods are respectively bold and underlined. We specify if a method assumes prior access to names of concepts ('Concept spec.)' and if it employs a frozen backbone (''). We specify what additional data is used at training ('Extra data') ('I' stands for images and 'IT' for image/text aligned data). Our CLIP-DINOiser only needs 1k images from ImageNet to be trained. 'SD' stands for Stable Diffusion [48]. We refer to Sec. 4.1 for more details on baselines and we detail the datasets used for training by each method in the supplementary material." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 384, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 384, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLIP-DINOiser" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 363, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 363, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 363, + 128 + ], + "type": "text", + "content": "4.2 Open-vocabulary semantic segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 163 + ], + "type": "text", + "content": "We discuss in this section state-of-the-art results on the task of open-vocabulary semantic segmentation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 179, + 481, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 179, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 132, + 179, + 481, + 323 + ], + "type": "text", + "content": "Evaluation with no 'background' class. We first compare in Tab. 1 ('No background prompt' column) the results on datasets which aim at the segmentation of most of the pixels in an image and do not consider a 'background' class. We observe that our method CLIP-DINOiser achieves the best results on four datasets yielding " + }, + { + "bbox": [ + 132, + 179, + 481, + 323 + ], + "type": "inline_equation", + "content": "+2.2" + }, + { + "bbox": [ + 132, + 179, + 481, + 323 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 179, + 481, + 323 + ], + "type": "inline_equation", + "content": "+5.0" + }, + { + "bbox": [ + 132, + 179, + 481, + 323 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 179, + 481, + 323 + ], + "type": "inline_equation", + "content": "+6.7" + }, + { + "bbox": [ + 132, + 179, + 481, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 179, + 481, + 323 + ], + "type": "inline_equation", + "content": "+5.1" + }, + { + "bbox": [ + 132, + 179, + 481, + 323 + ], + "type": "text", + "content": " mIoU over the second best performing method. Interestingly, we outperform methods which build expensive prototypes per visual concept on fine-grained datasets, showing the benefit of our lightweight and generalizable method. The only drop " + }, + { + "bbox": [ + 132, + 179, + 481, + 323 + ], + "type": "inline_equation", + "content": "(-0.8" + }, + { + "bbox": [ + 132, + 179, + 481, + 323 + ], + "type": "text", + "content": " mIoU) is seen on VOC20 with respect to OVDiff; we believe it is due to the benefit of generating per-concept negative prototypes which likely benefits this object-centric dataset. An adaptive granularity of feature correlation could help mitigate this drop, which we leave for future work." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 339, + 481, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 339, + 481, + 494 + ], + "spans": [ + { + "bbox": [ + 132, + 339, + 481, + 494 + ], + "type": "text", + "content": "Evaluation with 'background' class. We now compare our method on datasets which include a 'background' query in Tab. 1 ('W/ bkg prompt' column). In this setup, we also apply our background detection mechanism (detailed in Sec. 3.5) on VOC and Object in order to improve the stuff-like background detection. We observe that CLIP-DINOiser significantly outperforms all methods which do not construct prototypes. Moreover, we surpass OVDiff (which uses an ensemble of three models) on Context dataset by " + }, + { + "bbox": [ + 132, + 339, + 481, + 494 + ], + "type": "inline_equation", + "content": "+2.3\\mathrm{mIoU}" + }, + { + "bbox": [ + 132, + 339, + 481, + 494 + ], + "type": "text", + "content": " and are on par on Object. It is to be noted that with a single feature extractor, the performance of OVDiff drops by " + }, + { + "bbox": [ + 132, + 339, + 481, + 494 + ], + "type": "inline_equation", + "content": "-10\\mathrm{mIoU}" + }, + { + "bbox": [ + 132, + 339, + 481, + 494 + ], + "type": "text", + "content": " and the method requires the construction of a 'background' prototype per concept, otherwise losing another " + }, + { + "bbox": [ + 132, + 339, + 481, + 494 + ], + "type": "inline_equation", + "content": "-10\\mathrm{mIoU}" + }, + { + "bbox": [ + 132, + 339, + 481, + 494 + ], + "type": "text", + "content": " on VOC. On the other hand, CLIP-DINOiser produces segmentation masks in a single pass of CLIP with the light addition of two convolutional layers while remaining fully open-vocabulary as it does not require any concept-specific constructs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 510, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 481, + 665 + ], + "type": "text", + "content": "Qualitative results. We qualitatively compare in Fig. 6 CLIP-DINOiser with high-performing TCL [6], CLIP-DIY [59] (two recent methods which provide code) and our baseline method MaskCLIP [67] on images taken from the datasets considered in the evaluation. We observe that our method generates predictions accurate both in terms of localization and assignment. Indeed we obtain fined-grained results on the challenging datasets, e.g. in the Cityscapes example the text query 'car' and in the ADE20k example 'fountain' are accurately located when CLIP-DIY and TCL produce coarser results. Versus MaskCLIP, we can see the denoising capabilities of CLIP-DINOiser as MaskCLIP hallucinations grow with the number of text queries prompted at evaluation. Finally, in Fig. 1 we present 'in the wild' examples, beyond the evaluation benchmarks, and show that CLIP-DINOiser produces accurate segmentation masks for arbitrary and very specific prompts, such as 'wooden table' or 'leather bag'." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 255, + 102 + ], + "type": "text", + "content": "M.Wysoczanska et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 485, + 404 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 485, + 404 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 485, + 404 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 485, + 404 + ], + "type": "image", + "image_path": "1ce59c04146c54e151e6a345fa5d585a1022ce5194e561eec165314b08037cbf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 410, + 482, + 464 + ], + "lines": [ + { + "bbox": [ + 130, + 410, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 482, + 464 + ], + "type": "text", + "content": "Fig.6: Qualitative open-vocabulary segmentation results. We compare ours against CLIP-DIY [59], TCL [6] and MaskCLIP [67]. For a fair comparison, we do not apply post-processing. All pixels annotated in black are from the background class. We observe that our method achieves more accurate results both in terms of localization and class assignment." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 491, + 237, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 237, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 237, + 503 + ], + "type": "text", + "content": "4.3 Ablation study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 516, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 482, + 552 + ], + "type": "text", + "content": "We now conduct an ablation study of the different components of CLIP-DINOiser and investigate the impact of both our feature pooling strategy and background detection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "The impact of the pooling mechanism. We propose with CLIP-DINOiser to combine MaskCLIP features with a well-defined linear combination and compare different solutions in Tab. 2a. In [67], the authors proposed to refine the predictions with a combination weighted by CLIP key embeddings (noted 'CLIP keys (preds.)' in the table) and boost MaskCLIP results by more than " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "+8" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " mIoU on VOC and VOC20, " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "+1.8" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "+1.0" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "+0.6" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " mIoU on the other datasets. However, we show that working directly at the feature level allows us to achieve better results; we obtain consistent improvements ranging from " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "+6" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "+19" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " mIoU on" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 384, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 384, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLIP-DINOiser" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 482, + 140 + ], + "type": "text", + "content": "all datasets when using DINO-based weight " + }, + { + "bbox": [ + 132, + 115, + 482, + 140 + ], + "type": "inline_equation", + "content": "A^{\\xi}" + }, + { + "bbox": [ + 132, + 115, + 482, + 140 + ], + "type": "text", + "content": " and further improve when using trained CLIP-based weights " + }, + { + "bbox": [ + 132, + 115, + 482, + 140 + ], + "type": "inline_equation", + "content": "A^{\\phi}" + }, + { + "bbox": [ + 132, + 115, + 482, + 140 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 134, + 163, + 320, + 215 + ], + "blocks": [ + { + "bbox": [ + 134, + 163, + 320, + 215 + ], + "lines": [ + { + "bbox": [ + 134, + 163, + 320, + 215 + ], + "spans": [ + { + "bbox": [ + 134, + 163, + 320, + 215 + ], + "type": "table", + "html": "
Pooling strategyVOCVOC20C59StuffADE
MaskCLIP [67] - baseline32.961.825.617.614.3
CLIP keys (preds.) [67]41.371.927.418.614.9
ours w. CLIP keys39.273.223.012.67.7
ours w. DINO Aξ53.779.135.524.720.4
ours w. trained Aφ54.080.935.924.620.0
", + "image_path": "ed5c28b969d3133db29603925b8e41c3627b72e40a0f765f0b45961e6333668f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 325, + 161, + 480, + 219 + ], + "blocks": [ + { + "bbox": [ + 187, + 216, + 268, + 225 + ], + "lines": [ + { + "bbox": [ + 187, + 216, + 268, + 225 + ], + "spans": [ + { + "bbox": [ + 187, + 216, + 268, + 225 + ], + "type": "text", + "content": "(a) Pooling strategy" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 161, + 480, + 219 + ], + "lines": [ + { + "bbox": [ + 325, + 161, + 480, + 219 + ], + "spans": [ + { + "bbox": [ + 325, + 161, + 480, + 219 + ], + "type": "table", + "html": "
PoolingBkg det.ObjectVOC
MaskCLIP [67] - baseline16.432.9
ours w. DINO Aξ29.953.7
ours w. DINO AξFOUND32.160.1
ours w. DINO Aξours w. M34.162.1
ours w. DINO Aξours w. Mφ34.261.9
ours w. trained Aφours w. Mφ34.862.1
", + "image_path": "dc7051ededb486a1ab8910ff4db35a1ced00037aa8bb333e9e030e5a35f6a254.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 351, + 220, + 454, + 229 + ], + "lines": [ + { + "bbox": [ + 351, + 220, + 454, + 229 + ], + "spans": [ + { + "bbox": [ + 351, + 220, + 454, + 229 + ], + "type": "text", + "content": "(b) Background detection" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 238, + 481, + 262 + ], + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 262 + ], + "type": "text", + "content": "Table 2: Impact of the pooling strategy (a) and background detection (b) on diverse datasets reported with the mIoU metric." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 283, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 283, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 283, + 482, + 439 + ], + "type": "text", + "content": "The impact of the background detection. We now discuss the improvement provided by our background refinement strategy, which is applied when stuff-like background patches need to be detected. We report such results in Tab. 2b when employing our pooling strategy (either using DINO features, noted 'w. DINO " + }, + { + "bbox": [ + 132, + 283, + 482, + 439 + ], + "type": "inline_equation", + "content": "A^{\\xi}" + }, + { + "bbox": [ + 132, + 283, + 482, + 439 + ], + "type": "text", + "content": " or those extracted from CLIP, noted 'w. trained " + }, + { + "bbox": [ + 132, + 283, + 482, + 439 + ], + "type": "inline_equation", + "content": "A^{\\phi}" + }, + { + "bbox": [ + 132, + 283, + 482, + 439 + ], + "type": "text", + "content": " ). When using solely 'FOUND' for background detection, as in [59], we improve by " + }, + { + "bbox": [ + 132, + 283, + 482, + 439 + ], + "type": "inline_equation", + "content": "+6.4" + }, + { + "bbox": [ + 132, + 283, + 482, + 439 + ], + "type": "text", + "content": " mIoU on VOC (achieving " + }, + { + "bbox": [ + 132, + 283, + 482, + 439 + ], + "type": "inline_equation", + "content": "60.1\\mathrm{mIoU}" + }, + { + "bbox": [ + 132, + 283, + 482, + 439 + ], + "type": "text", + "content": "), but when relaxing FOUND (see Sec. 3.5) with an uncertainty condition, we boost scores up to 62.1 on VOC, showing the limitation of using FOUND alone. We also achieve similar results when using CLIP-based predictions " + }, + { + "bbox": [ + 132, + 283, + 482, + 439 + ], + "type": "inline_equation", + "content": "M^{\\phi}" + }, + { + "bbox": [ + 132, + 283, + 482, + 439 + ], + "type": "text", + "content": " both with DINO-based " + }, + { + "bbox": [ + 132, + 283, + 482, + 439 + ], + "type": "inline_equation", + "content": "A^{\\xi}" + }, + { + "bbox": [ + 132, + 283, + 482, + 439 + ], + "type": "text", + "content": " and trained CLIP-based " + }, + { + "bbox": [ + 132, + 283, + 482, + 439 + ], + "type": "inline_equation", + "content": "A^{\\phi}" + }, + { + "bbox": [ + 132, + 283, + 482, + 439 + ], + "type": "text", + "content": " correlations, although we observe that best results are overall obtained with trained " + }, + { + "bbox": [ + 132, + 283, + 482, + 439 + ], + "type": "inline_equation", + "content": "A^{\\phi}" + }, + { + "bbox": [ + 132, + 283, + 482, + 439 + ], + "type": "text", + "content": ". We visualize CLIP-based mask " + }, + { + "bbox": [ + 132, + 283, + 482, + 439 + ], + "type": "inline_equation", + "content": "M^{\\phi}" + }, + { + "bbox": [ + 132, + 283, + 482, + 439 + ], + "type": "text", + "content": " in Fig. 5a and see high similarity to DINO-based predictions, therefore showing the localization quality of CLIP." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 462, + 226, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 462, + 226, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 462, + 226, + 475 + ], + "type": "text", + "content": "5 Conclusions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 492, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 482, + 589 + ], + "type": "text", + "content": "In this work, we propose to make the most out of CLIP features and show that the features already contain useful localization information. Indeed with light convolutional layers, we are able to learn both good patch-correlation and objectness information by using DINO self-supervised model as a guide. With such information, our method CLIP-DINOiser performs zero-shot open-vocabulary semantic segmentation in a single pass of CLIP model and with two light extra convolutional layers. CLIP-DINOiser reaches state-of-the-art results on complex semantic segmentation datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Limitations. Despite yielding strong results on open-vocabulary semantic segmentation, CLIP-DINOiser is still bounded by the capability of the CLIP model to separate classes, as it inherits its granularity. We believe that better prompt engineering paired with better image-text models could further boost the performance of CLIP-DINOiser." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 255, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 255, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 255, + 101 + ], + "type": "text", + "content": "M.Wysoczanska et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 175 + ], + "type": "text", + "content": "This work was supported by the National Centre of Science (Poland) Grant No. 2022/45/B/ST6/02817 and by the grant from NVIDIA providing one RTX A5000 24GB used for this project." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 217, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 138, + 217, + 480, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 480, + 240 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 480, + 240 + ], + "type": "text", + "content": "1. Abdelreheem, A., Skorokhodov, I., Ovsjanikov, M., Wonka, P.: Satr: Zero-shot semantic segmentation of 3d shapes. In: ICCV (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 240, + 480, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 240, + 480, + 261 + ], + "spans": [ + { + "bbox": [ + 138, + 240, + 480, + 261 + ], + "type": "text", + "content": "2. Bucher, M., Vu, T.H., Cord, M., Pérez, P.: Zero-shot semantic segmentation. In: NeurIPS (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 261, + 480, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 480, + 283 + ], + "type": "text", + "content": "3. Caesar, H., Uijlings, J., Ferrari, V.: Coco-stuff: Thing and stuff classes in context. In: CVPR (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 284, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 284, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 284, + 480, + 304 + ], + "type": "text", + "content": "4. Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., Joulin, A.: Unsupervised learning of visual features by contrasting cluster assignments. In: NeurIPS (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 305, + 480, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 305, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 305, + 480, + 327 + ], + "type": "text", + "content": "5. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 327, + 480, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 327, + 480, + 349 + ], + "spans": [ + { + "bbox": [ + 138, + 327, + 480, + 349 + ], + "type": "text", + "content": "6. Cha, J., Mun, J., Roh, B.: Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs. In: CVPR (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 349, + 480, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 349, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 138, + 349, + 480, + 381 + ], + "type": "text", + "content": "7. Chen, J., Zhu, D., Qian, G., Ghanem, B., Yan, Z., Zhu, C., Xiao, F., Culatana, S.C., Elhoseiny, M.: Exploring open-vocabulary semantic segmentation from clip vision encoder distillation only. In: ICCV (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 381, + 480, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 480, + 414 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 480, + 414 + ], + "type": "text", + "content": "8. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: Clip2scene: Towards label-efficient 3d scene understanding by clip. In: CVPR (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 415, + 480, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 480, + 436 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 480, + 436 + ], + "type": "text", + "content": "9. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 436, + 480, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 436, + 480, + 457 + ], + "spans": [ + { + "bbox": [ + 138, + 436, + 480, + 457 + ], + "type": "text", + "content": "10. Contributors, M.: MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 458, + 480, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 458, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 458, + 480, + 491 + ], + "type": "text", + "content": "1. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: CVPR (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 491, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 512 + ], + "type": "text", + "content": "2. Darcet, T., Oquab, M., Mairal, J., Bojanowski, P.: Vision transformers need registers. arXiv preprint arXiv:2309.16588 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 512, + 480, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 534 + ], + "type": "text", + "content": "3. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "type": "text", + "content": "4. Ding, J., Xue, N., Xia, G.S., Dai, D.: Decoupling zero-shot semantic segmentation. In: CVPR (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 555, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 589 + ], + "type": "text", + "content": "5. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "type": "text", + "content": "6. Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "text", + "content": "7. Fang, A., Ilharco, G., Wortsman, M., Wan, Y., Shankar, V., Dave, A., Schmidt, L.: Data determines distributional robustness in contrastive language image pretraining (CLIP). In: ICML (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "8. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: ECCV (2022)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 384, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 384, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLIP-DINOiser" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "19. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: CVPR (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "20. Gu, Z., Zhou, S., Niu, L., Zhao, Z., Zhang, L.: Context-aware feature generation for zero-shot semantic segmentation. In: ACM MM (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "21. Hu, P., Sclaroff, S., Saenko, K.: Uncertainty-aware learning for zero-shot semantic segmentation. In: NeurIPS (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 225 + ], + "type": "text", + "content": "22. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021). https://doi.org/10.5281/zenodo.5143773, https://doi.org/10.5281/zenodo.5143773" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "23. Jatavallabhula, K.M., Kuwajerwala, A., Gu, Q., Omama, M., Chen, T., Li, S., Iyer, G., Saryazdi, S., Keetha, N., Tewari, A., et al.: Conceptfusion: Open-set multimodal 3d mapping. In: RSS (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "24. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "25. Karazija, L., Laina, I., Vedaldi, A., Rupprecht, C.: Diffusion models for zero-shot open-vocabulary segmentation. arXiv preprint arXiv:2306.09316 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "26. Kato, N., Yamasaki, T., Aizawa, K.: Zero-shot semantic segmentation via variational mapping. In: ICCVW (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "27. Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: Lerf: Language embedded radiance fields. In: ICCV (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "text", + "content": "28. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "29. Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: ICLR (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "text", + "content": "30. Li, P., Wei, Y., Yang, Y.: Consistent structural relation learning for zero-shot segmentation. NeurIPS (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "type": "text", + "content": "31. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: CVPR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "text", + "content": "32. Liu, Q., Wen, Y., Han, J., Xu, C., Xu, H., Liang, X.: Open-world semantic segmentation via contrasting and clustering vision-language embedding. In: ECCV (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 490, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 511 + ], + "type": "text", + "content": "33. Luo, H., Bao, J., Wu, Y., He, X., Li, T.: SegCLIP: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. In: ICML (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "34. Mayilvahanan, P., Wiedemer, T., Rusak, E., Bethge, M., Brendel, W.: Does CLIP's generalization performance mainly stem from high train-test similarity? arXiv preprint arXiv:2310.09562 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 545, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 566 + ], + "type": "text", + "content": "35. Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word representations in vector space. In: ICLR (2013)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "36. Mottaghi, R., Chen, X., Liu, X., Cho, N.G., Lee, S.W., Fidler, S., Urtasun, R., Yuille, A.: The role of context for object detection and semantic segmentation in the wild. In: CVPR (2014)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "37. Mukhoti, J., Lin, T.Y., Poursaeed, O., Wang, R., Shah, A., Torr, P.H., Lim, S.N.: Open vocabulary semantic segmentation with patch aligned contrastive learning. In: CVPR (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "38. Najibi, M., Ji, J., Zhou, Y., Qi, C.R., Yan, X., Ettinger, S., Anguelov, D.: Unsupervised 3d perception with 2d vision-language distillation for autonomous driving. In: ICCV (2023)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "type": "text", + "content": "M.Wysoczanska et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 172 + ], + "type": "text", + "content": "39. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., HAZIZA, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: DINOv2: Learning robust visual features without supervision. TMLR (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "text", + "content": "40. Pastore, G., Cermelli, F., Xian, Y., Mancini, M., Akata, Z., Caputo, B.: A closer look at self-training for zero-label semantic segmentation. In: CVPR (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "41. Peng, S., Genova, K., Jiang, C., Tagliasacchi, A., Pollefeys, M., Funkhouser, T., et al.: Openscene: 3d scene understanding with open vocabularies. In: CVPR (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "42. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: EMNLP (2014)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "text", + "content": "43. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "type": "text", + "content": "44. Ranasinghe, K., McKinzie, B., Ravi, S., Yang, Y., Toshev, A., Shlens, J.: Perceptual grouping in contrastive vision-language models. In: ICCV (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "text", + "content": "45. Rao, Y., Zhao, W., Chen, G., Tang, Y., Zhu, Z., Huang, G., Zhou, J., Lu, J.: Denseclip: Language-guided dense prediction with context-aware prompting. In: CVPR (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "type": "text", + "content": "46. Ren, P., Li, C., Xu, H., Zhu, Y., Wang, G., Liu, J., Chang, X., Liang, X.: Viewco: Discovering text-supervised segmentation masks via multi-view semantic consistency. In: ICLR (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "type": "text", + "content": "47. Rewatbowornwong, P., Chatthee, N., Chuangsuwanich, E., Suwajanakorn, S.: Zero-guidance segmentation using zero segment labels. In: ICCV (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 403 + ], + "type": "text", + "content": "48. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "text", + "content": "49. Shin, G., Xie, W., Albanie, S.: Reco: Retrieve and co-segment for zero-shot transfer. In: NeurIPS (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 424, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 445 + ], + "type": "text", + "content": "50. Shin, G., Xie, W., Albanie, S.: Namedmask: Distilling segmenters from complementary foundation models. In: CVPRW (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "text", + "content": "51. Simeoni, O., Puy, G., Vo, H.V., Roburin, S., Gidaris, S., Bursuc, A., Pérez, P., Marlet, R., Ponce, J.: Localizing objects with self-supervised transformers and no labels. In: BMVC (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "text", + "content": "52. Simeoni, O., Sekkat, C., Puy, G., Vobecky, A., Zablocki, É., Pérez, P.: Unsupervised object localization: Observing the background to discover objects. In: CVPR (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "text", + "content": "53. Simeoni, O., Zablocki, E., Gidaris, S., Puy, G., Pérez, P.: Unsupervised object localization in the era of self-supervised vits: A survey. arXiv preprint arXiv:2310.12904 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 533, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 481, + 566 + ], + "type": "text", + "content": "54. Vobecký, A., Simeoni, O., Hurych, D., Gidaris, S., Bursuc, A., Perez, P., Sivic, J.: Pop-3d: Open-vocabulary 3d occupancy prediction from images. In: NeurIPS (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 566, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 566, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 481, + 589 + ], + "type": "text", + "content": "55. Walmer, M., Suri, S., Gupta, K., Shrivastava, A.: Teaching matters: Investigating the role of supervision in vision transformers. In: CVPR (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "type": "text", + "content": "56. Wang, X., Girdhar, R., Yu, S.X., Misra, I.: Cut and learn for unsupervised object detection and instance segmentation. In: CVPR (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "text", + "content": "57. Wang, Y., Shen, X., Hu, S.X., Yuan, Y., Crowley, J.L., Vaufreydaz, D.: Self-supervised transformers for unsupervised object discovery using normalized cut. In: CVPR (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "58. Wu, J., Li, X., Xu, S., Yuan, H., Ding, H., Yang, Y., Li, X., Zhang, J., Tong, Y., Jiang, X., et al.: Towards open vocabulary learning: A survey. T-PAMI (2024)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 384, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 384, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLIP-DINOiser" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 369 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "text", + "content": "59. Wysoczanska, M., Ramamonjisoa, M., Trzcinski, T., Simeoni, O.: Clip-diy: Clip dense inference yields open-vocabulary semantic segmentation for-free. In: WACV (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "text", + "content": "60. Xian, Y., Choudhury, S., He, Y., Schiele, B., Akata, Z.: Semantic projection network for zero-and few-label semantic segmentation. In: CVPR (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 204 + ], + "type": "text", + "content": "61. Xu, J., De Mello, S., Liu, S., Byeon, W., Breuel, T., Kautz, J., Wang, X.: Groupvit: Semantic segmentation emerges from text supervision. arXiv preprint arXiv:2202.11094 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 236 + ], + "type": "text", + "content": "62. Xu, J., Hou, J., Zhang, Y., Feng, R., Wang, Y., Qiao, Y., Xie, W.: Learning open-vocabulary semantic segmentation models from natural language supervision. In: CVPR (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 258 + ], + "type": "text", + "content": "63. Zhai, X., Mustafa, B., Kolesnikov, A., Beyer, L.: Sigmoid loss for language image pre-training. In: ICCV (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "64. Zhao, H., Puig, X., Zhou, B., Fidler, S., Torralba, A.: Open vocabulary scene parsing. In: ICCV (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 480, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 313 + ], + "type": "text", + "content": "65. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., et al.: Regionclip: Region-based language-image pretraining. In: CVPR (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 314, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 314, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 480, + 335 + ], + "type": "text", + "content": "66. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. IJCV (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 336, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 346 + ], + "type": "text", + "content": "67. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: ECCV (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "text", + "content": "68. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: iBOT: Image bert pre-training with online tokenizer. In: ICLR (2022)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 254, + 101 + ], + "type": "text", + "content": "M.Wysoczanska et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CLIP-DPO_ Vision-Language Models as a Source of Preference for Fixing Hallucinations in LVLMs/174c66bb-6859-4a7e-835b-0b1709492a92_content_list.json b/2024/CLIP-DPO_ Vision-Language Models as a Source of Preference for Fixing Hallucinations in LVLMs/174c66bb-6859-4a7e-835b-0b1709492a92_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fc6df766a36973d708a5716c1ad3fe3c63819155 --- /dev/null +++ b/2024/CLIP-DPO_ Vision-Language Models as a Source of Preference for Fixing Hallucinations in LVLMs/174c66bb-6859-4a7e-835b-0b1709492a92_content_list.json @@ -0,0 +1,2066 @@ +[ + { + "type": "text", + "text": "CLIP-DPO: Vision-Language Models as a Source of Preference for Fixing Hallucinations in LVLMs", + "text_level": 1, + "bbox": [ + 218, + 140, + 782, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yassine Ouali1 Adrian Bulat1,2 Brais Martinez1 Georgios Tzimiropoulos1,3", + "bbox": [ + 228, + 210, + 767, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Samsung AI Center Cambridge, UK $^{2}$ Technical University of Iasi, Romania $^{3}$ Queen Mary University of London, UK", + "bbox": [ + 217, + 241, + 785, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Despite recent successes, LVLMs or Large Vision Language Models are prone to hallucinating details like objects and their properties or relations, limiting their real-world deployment. To address this and improve their robustness, we present CLIP-DPO, a preference optimization method that leverages contrastively pre-trained Vision-Language (VL) embedding models, such as CLIP, for DPO-based optimization of LVLMs. Unlike prior works tackling LVLM hallucinations, our method does not rely on paid-for APIs, and does not require additional training data or the deployment of other external LVLMs. Instead, starting from the initial pool of supervised fine-tuning data, we generate a diverse set of predictions, which are ranked based on their CLIP image-text similarities, and then filtered using a robust rule-based approach to obtain a set of positive and negative pairs for DPO-based training. We applied CLIP-DPO fine-tuning to the MobileVLM-v2 family of models and to LlaVA-1.5, in all cases observing significant improvements in terms of hallucination reduction over baseline models. We also observe better performance for zero-shot classification, suggesting improved grounding capabilities, and verify that the original performance on standard LVLM benchmarks is overall preserved.", + "bbox": [ + 261, + 308, + 743, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 597, + 375, + 614 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A Large Vision Language Model (LVLM) is a Large Language Model (LLM) combined with an independently-trained vision encoder, typically taken from VL embedding models like CLIP [56] and often kept frozen. LVLMs have recently become the state-of-the-art for vision language understanding [5, 16, 42, 65, 81]. Unlike the prior generation of models [34, 70], which were typically tuned for one or a small number of tasks, LVLMs allow free-form dialog in natural language with an image in the loop. This direction largely draws inspiration from the recent success of ChatGPT [50] and respectively, ChatGPT-4V [51], which adapts LLMs to follow human instruction and preferences via various forms of reinforcement learning from human feedback (RFLH) and/or supervised fine-tuning (SFT) on high-quality multi-turn instruction (i.e. chat) data. In a similar fashion, the current generation of LVLMs is trained as part of a two-step process [42]. First, the CLIP vision encoder is aligned with the LLM by training a few adaptation layers only and, then, by fine-tuning the model using SFT on a multi-modal", + "bbox": [ + 212, + 628, + 787, + 843 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "instructional dataset. Despite their remarkable success, a major limitation hindering the wider adoption of LVLMs is the high rate of hallucinated details (e.g. non-existing objects) that these models exhibit in their generated output text.", + "bbox": [ + 215, + 146, + 784, + 191 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivation. Following the footsteps of LLM training, a natural solution to combat hallucinations is given by RFLH. As reinforcement learning-based training is generally expensive, often requiring the training of a policy model, Rafailov et al. [57] proposed a simplified framework, coined Direct Preference optimization (DPO), that allows direct training of the model using a binary cross-entropy objective. The data for DPO is typically obtained either from human preferences or using LLMs to construct preference data automatically. Thanks to its efficiency, DPO has been quickly adopted for LVLMs too [35, 36, 83]. However, these approaches require multiple LVLMs for generation, collecting additional data, and using GPT-4/GPT-4V API for ranking. Moreover, the GPT-4 provided scores are discrete (hence, of reduced granularity) and are themselves prone to hallucinations, which can perpetuate and exacerbate the already high rate of factual errors and poor image grounding [37, 71]. Finally, GPT-4/V is behind a paywall, and using it for generation/ranking is neither scalable nor cost-efficient.", + "bbox": [ + 215, + 196, + 785, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Main idea. To fix hallucinations and address the high data construction and annotation cost exhibited by current methods, we propose a new DPO variant, called CLIP-DPO, that uses a pre-trained CLIP model [56] to rank the LVLM's self-generated captions to construct positive-negative pairs for DPO. Since the CLIP model was trained in a contrastive manner to measure the alignment between image-text pairs, it is naturally suitable for determining the quality of a given output from an LVLM, grounding it to the correct object or attribute. The dataset over which we operate is constructed by running the original pre-trained target LVLM on its own output obtained using prompting, removing both the need for (i) additional external data and (ii) ensembles of external LVLMs. The final data is filtered before training using robust rule-based filtering.", + "bbox": [ + 215, + 412, + 785, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Main results. We applied CLIP-DPO fine-tuning on top of two state-of-the-art models: MobileVLM-v2 (3 models in total) and LlaVA-1.5 7B. We find that, in both cases, our approach is effective in reducing hallucinations, outperforming all baseline models (i.e. the models without CLIP-DPO fine-tuning) by a significant margin. Importantly, CLIP-DPO significantly outperforms our direct competitor HA-DPO [83], outperforms Qwen-VL [5] trained on significantly larger datasets (i.e. 1.4B samples for pre-training and 77M for multitask training; as opposed to CLIP-DPO training on just 0.7M samples). Finally, our model's enhanced object grounding capabilities are also illustrated for zero-shot image classification, all without degrading the original performance of the base LVLM model.", + "bbox": [ + 215, + 584, + 785, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 760, + 380, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large Visual Language Models (LVLMs). Following the unprecedented success of Large Language Models (LLMs) [7, 50, 68, 69, 80], several works have recently proposed to build multi-modal capabilities on top of them [10, 14, 16, 42].", + "bbox": [ + 215, + 794, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Ouali et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LLaVA [42] and FROMAGe [31] directly pass the visual tokens produced by a pre-trained CLIP [56] vision encoder to an LLM, either fine-tuning the LLM or adapting it using LoRA [24]. Notably, LLaVA training includes a pre-training stage that aligns the CLIP features with the LLM input tokens using a simple projection layer (i.e. keeping the rest of the model frozen) using image captioning data. InstructBLIP [16] uses QFormer [33] to reduce the number of vision tokens before passing them to the LLM. As the quality and distribution of the training data plays a crucial role [84], a series of methods [9,10,55,76] introduced improved data construction pipelines. For example, ShareGPT4v [10] uses the API of GPT-4V [51] to first label, then train a model, and finally use it to re-announce a new training set. Another line of work is improving efficiency by reducing the models' size (original LLaVA models have 7B and 13B parameters). LLaVA-Phi [87], MobileVLM [13] and its follow-up, MobileVLM-v2 [14], replace the LLaMA [68] and Vicuna [12] LLMs with the smaller 1.4B and 2.7B variants of MobileLLaMA and Phi [38]. Our work is orthogonal to the aforementioned methods and does not seek to improve the model's architecture. Instead, we propose CLIP-DPO, an improved training approach for LVLMs based on DPO.", + "bbox": [ + 217, + 145, + 785, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preference optimization. Instruction tuning can significantly improve LLMs' perceived output quality and usefulness by aligning their responses to a given task domain or human preferences. This is achieved either by direct fine-tuning on expert data [15, 47, 67] or via reinforcement learning [52, 64, 88]. The latter significantly simplifies the data collection process, but still requires complicated training algorithms based on REINFORCE [73] or PPO [61]. Recently, a much-simplified approach was proposed, Direct Preference optimization (DPO) [57], which bypasses the need to train a reward model and allows direct training using a cross-entropy loss. Multiple improved versions for LLMs have been proposed in the meantime [4, 75, 82]. Following this, a recent wave of works on combining DPO with LVLMs have been proposed [35, 36, 83]. Silkie [35] constructs a multimodal instructional dataset automatically labeled by GPT-4V. Similarly, HA-DPO [83], aiming to reduce the rate of hallucinations, uses a GPT4 model to label and construct positive-negative pairs with and without hallucinations. The work of [36] follows a similar path by using a suite of LLMs and LVLMs (i.e. Gemini-Vision [66]) to generate and label the data. These methods are then primarily evaluated on the LLaVA benchmark and PoPE [72]. In contrast to the aforementioned works, the proposed CLIP-DPO (i) simplifies the pipeline, (ii) removes the need for paid APIs, (iii) removes the need for additional data, and (iv) removes the need for additional external LVLMs. Instead, we make use of a pre-trained CLIP model to rank the generated outputs from a small pool of efficient LVLMs, and show that the proposed CLIP-DPO training provides significant improvements for fixing LVLM hallucinations and, in general, for enhancing the discriminability and robustness of the model as demonstrated by image classification experiments.", + "bbox": [ + 217, + 410, + 785, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LVLM hallucinations. Broadly speaking, in the context of LVLMs, we consider hallucinations to be incorrect or misleading generated text, contradicting the visual evidence provided by the input image. This is an undesirable characteristic", + "bbox": [ + 217, + 794, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "CLIP-DPO", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6a9ae7b1a8a89479afbf2abdd702599b518d1594d17601816cb4fba608e505e0.jpg", + "image_caption": [ + "Fig. 1: An example of the injected hallucinations. Given an image and its caption, we prompt GPT-4 to generate 3 types of hallucination: existence, attributes, and relation." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 364, + 231 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Original Caption: In this image, we see a black dog wearing an orange vest standing on the grass. The dog's head is down and it appears to be in a backyard.", + "bbox": [ + 370, + 143, + 764, + 162 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Existence Hallucination: In this image, we see a black dog wearing an orange vest standing on the grass, with a small orange ball near his feet. The dog's head is down and it appears to be in a backyard.", + "- Attributes Hallucination: In this image, we see a dark brown dog wearing an orange vest standing on the grass. The dog's head is down and it appears to be in a backyard.", + "- Relation Hallucination: In this image, we see a black dog wearing an orange vest sitting on the grass. The dog's head is down and it appears to be in a backyard." + ], + "bbox": [ + 367, + 165, + 782, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "inherited from the pre-trained LLM used, and further exacerbated by the visual-language alignment process [5, 25, 29, 40]. Multiple solutions have been recently proposed by the community with varying degrees of success. The works of [25, 77] attempt to address the data bias by constructing better-grounded annotated image-text pairs. The works of [5, 11] scale the resolution of the image encoder, as this was observed to reduce the amount of hallucination, but at the cost of a high increase in the computational cost. The works of [27, 78] improve the vision encoder by adding extra informational paths. The closest work related to ours is that of [83], which, with the help of GPT-4V, constructs negative-positive pairs for DPO fine-tuning. Different from [83], we use a pre-trained CLIP model to perform the ranking, and no additional data or external LVLMs are required. For evaluation purposes, we use AMBER [71], the most comprehensive benchmark for hallucinations to date, encompassing both a generative and discriminative evaluation component. Importantly, unlike all prior benchmarks [22, 30, 37, 40, 65], AMBER is a high-quality dataset fully annotated by humans for both the generative and discriminative tasks.", + "bbox": [ + 212, + 300, + 787, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 566, + 382, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 The Effectiveness of CLIP in Reducing Hallucinations", + "text_level": 1, + "bbox": [ + 212, + 601, + 705, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, we probe the effectiveness of CLIP in accurately ranking correct captions over those with hallucinated content. This assessment will indicate the efficacy of CLIP-style models in reducing hallucinations with DPO-based training. To this end, initially, we select 1K captions from the Detailed Caption [1] dataset, chosen for its high-quality labels. Then, we instruct GPT-4 [2] to create three hallucinated captions for each image, corresponding to the following types of hallucination: (i) existence: new elements or objects are added to the caption that were not mentioned originally, (ii) attribute: the attributes, characteristics, or features of the original caption's elements are altered, and (iii) relationship: the spatial, contextual, or interactive relationships of the original caption's elements are altered. See Fig. 1 for an example of the injected hallucinations.", + "bbox": [ + 212, + 628, + 787, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we use LLaVA-1.5 7B [41] to compute the likelihood of all captions, including the original and the hallucinated captions. We then retain only the hallucinated captions for which the model predicts a higher likelihood than the", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Ouali et al.", + "bbox": [ + 271, + 114, + 369, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "original, yielding 61 for existence, 76 for attribute, and 119 for relationship hallucinations out of the initial 1K captions. Finally, we compute the CLIP", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "image-text scores for these likely yet incorrect captions. As illustrated in Fig.2, CLIP manages to accurately rank the original caption over the hallucinated ones in at least $59\\%$ of cases, reaching up to $83\\%$ for the existence type of hallucinations. These findings underscore the potential of using VL embedding models to provide a reliable training signal for hallucination reduction in LVLMs.", + "bbox": [ + 212, + 176, + 500, + 340 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fddc432d5c60939e88adc6911e453efe23869fd149a59cfe100c41e56feb3b42.jpg", + "image_caption": [ + "Fig. 2: In dark blue, we show the number of hallucinated captions per type that LLaVA-1.5 7B assigns a higher likelihood than the original out of 1K captions. Light blue shows the portion of samples corrected by CLIP." + ], + "image_footnote": [], + "bbox": [ + 535, + 178, + 761, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Direct Preference Optimization", + "text_level": 1, + "bbox": [ + 214, + 366, + 410, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The standard pipeline for LLM and LVLM alignment consists of three steps: generation, annotation, and op", + "bbox": [ + 212, + 407, + 500, + 453 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "timization. First, a set of $N$ prompts $\\{x_{i}\\}_{i = 1}^{N}$ , each one consisting of an image and a text component in the case of LVLMs, are used to generate a set of response pairs $y_{i}^{1}$ and $y_{i}^{2}$ obtained from a pool of pre-trained LVLMs for each prompt $x_{i}$ . Then, either human annotators or another set of LLMs (i.e. AI annotators) are used to rank the responses, resulting in a preferred $y_{i}^{+}$ and a less preferred response $y_{i}^{-}$ for each prompt $x_{i}$ , and thus a final preference dataset $\\mathcal{D} = \\{(x_i,y_i^+,y_i^-)\\}_{i = 1}^N$ . Finally, Direct Preference Optimization [57] (DPO) can be applied to update the target policy $\\pi_{\\theta}$ parameterized by $\\theta$ directly using the preference dataset $\\mathcal{D}$ . Specifically, the DPO optimization objective is defined as follows:", + "bbox": [ + 212, + 453, + 787, + 603 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\pi_ {\\theta}} \\mathbb {E} _ {(x _ {i}, y _ {i} ^ {+}, y _ {i} ^ {-}) \\sim \\mathcal {D}} \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta} (y _ {i} ^ {+} \\mid x _ {i})}{\\pi_ {\\mathrm {r e f}} (y _ {i} ^ {+} \\mid x _ {i})} - \\beta \\log \\frac {\\pi_ {\\theta} (y _ {i} ^ {-} \\mid x _ {i})}{\\pi_ {\\mathrm {r e f}} (y _ {i} ^ {-} \\mid x _ {i})}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 259, + 612, + 785, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $\\pi_{\\theta}$ as the policy to be learned, $\\pi_{\\mathrm{ref}}$ as the reference SFT policy, and $\\beta$ as a hyperparameter to control the Kullback-Leibler (KL) divergence between the learned $\\pi_{\\theta}$ and reference $\\pi_{\\mathrm{ref}}$ policies [57]. The main benefit of DPO-based optimization is the direct alignment of the LVLM towards the preferences implicit in the preference data $\\mathcal{D}$ .", + "bbox": [ + 212, + 657, + 787, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In contrast to previous DPO-based LVLM preference optimization works [35, 36, 83], next, we will introduce the proposed CLIP-DPO, which (i) simplifies the generation process by limiting the pool of LVLMs used for generation to small and efficient models, (ii) removes the need for external LLMs and LVLMs annotators accessed via paid APIs (e.g. GPT-3.5, GPT-4 or GPT-4V) by using CLIP as the ranking model, and (iii) removes the need for additional data by reusing the same data used during the SFT step.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "CLIP-DPO", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f79e40a20a29c0337b7e8f8f3ec82dc49962d47d9d910f6a5b6742c549fd9540.jpg", + "image_caption": [ + "Fig. 3: CLIP-DPO. Starting from the initial SFT data pool and a set of prompts, an LVLM generates captions. These captions are first ranked using a CLIP model, then filtered to identify the most suitable positive and negative pairs for DPO-based optimization." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 787, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 215, + 354, + 330, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we introduce our DPO variant for LVLMs preference optimization, named CLIP-DPO. As illustrated in Fig. 3, CLIP-DPO follows a three-step process: generation, annotation, and optimization, similar to the standard DPO-based training pipeline. While the final DPO-based optimization step remains unchanged, CLIP-DPO modifies the first two steps.", + "bbox": [ + 212, + 386, + 782, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Firstly, we streamline the generation step by utilizing the same data pool (i.e. images only) as the SFT stage and generate data using either the model itself or a small pool of efficient LVLMs. Secondly, we introduce an annotation step tailored for CLIP-style embedding models, where we rank captions based on their image-text similarities. Then, given that the resulting captions and pairs may vary in quality, we implement a rule-based filtering strategy to remove unsuitable candidates. Finally, with the constructed preference data, the model undergoes DPO-based optimization using Eq. 1, as shown in Fig. 3.", + "bbox": [ + 212, + 462, + 784, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As such, by leveraging VL embedding models for data annotation, we tap into the extensive information captured by these contrastively-pretrained models, which have been exposed to hundreds of millions of unique image-text pairs during training, unlike existing open-sourced LVLMs. Additionally, unlike the discrete GPT-4V ranking used in prior work, CLIP scores are continuous, enabling finer-grained comparisons. This allows us to assess the difficulty of a given pair based on the margin between the score values.", + "bbox": [ + 212, + 583, + 785, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, we will elaborate on each module of our pipeline. Sec. 4.1 details the data generation module, and Sec. 4.2 details the data annotation process.", + "bbox": [ + 212, + 688, + 785, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Data Generation", + "text_level": 1, + "bbox": [ + 215, + 739, + 401, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The first step of our pipeline consists of generating a set of per-image captions that will be later ranked by a pre-trained CLIP model, along with a subsequent filtering approach to select a set of positive and negative pairs, which are then used for DPO-based training. To this end, we start by selecting the pool of LVLM annotators and the data to be annotated. For the annotators and to reduce the", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Ouali et al.", + "bbox": [ + 271, + 114, + 369, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cost of the generation step, we select MobileVLM-v2 [14] family of models, given their efficiency and performance. As for the data, and to avoid introducing any new sources that might bias the results of CLIP-DPO, and to further reduce the method's cost, we opt for the initial pool of data used during the SFT stage of MobileVLM-v2 [14] models, see Table 1 for details. Then, we conduct our data generation pipeline that consists of two steps, the generation of generic captions and per-image questions and answers.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/80e7e79683e7a8556f04ad02adead17f0385bce5093d4a7184e992e8f741389d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetDataset Size
COCO118K
GQA72K
LLaVA-1.0 Pretraining595K
LLaVA-1.5 Pretraining558K
OCR VQA80K
SAM570K
SBU845K
TextVQA22K
VG86K
Web-Celebrity495
Web-Landmark500
Wikiart500
Total2.9M
", + "bbox": [ + 233, + 265, + 472, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1: Data pool used for data generation.", + "bbox": [ + 214, + 417, + 491, + 445 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/05ffd1475c796d2f88ab645289a33b8e1a495813251ed3b521f3a2cd761937b5.jpg", + "image_caption": [ + "Fig. 4: The distribution of CLIP image-text scores per category." + ], + "image_footnote": [], + "bbox": [ + 535, + 257, + 750, + 406 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2d257b53e0d3d9670fa31a085752a1c82d358b45ef8218a48a098507e4cc0800.jpg", + "image_caption": [ + "Fig. 5: Examples of generated generic captions produced by MobileVLM-v2 1.7B and MobileVLM-v2 3B. For MobileVLM-v2 3B's generated captions, we also show the produced question and positive and negative answers generated using Mistral 7B." + ], + "image_footnote": [], + "bbox": [ + 217, + 474, + 338, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MobileVLMv-2 1.7b:", + "text_level": 1, + "bbox": [ + 346, + 477, + 421, + 484 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The image showcases a charming two-story building, characterized by its steep roof and a balcony adorned with red flowers. The building's facade is painted white, accented with a green roof that adds a touch of color to the scene.", + "The image captures a charming scene of a German half-timbered building, bathed in the soft glow of daylight." + ], + "bbox": [ + 352, + 486, + 779, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MobileVLMv-2 3b:", + "text_level": 1, + "bbox": [ + 348, + 522, + 411, + 529 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The image showcases a traditional German half-timbered house, a common architectural feature in the region. The house is adorned with a green roof, a characteristic of half-timbered houses, and is embellished with red flowers that add a vibrant touch to the overall design.", + "- A building with a statue of a knight in the middle of the building." + ], + "bbox": [ + 352, + 531, + 745, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mistral 7b:", + "text_level": 1, + "bbox": [ + 348, + 571, + 395, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- What color is the roof of the building in the image?\n- Positive: Green Negative: Belge", + "bbox": [ + 352, + 579, + 560, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generation of Generic Captions. We start by generating a set of 5 descriptive captions per image. For each of the MobileVLM-v2 models, we prompt with 5 different prompts (e.g. \"identify the setting and note any characters or objects, focusing on visible details.\") to increase the diversity of the generated captions. While these captions can be used for CLIP ranking and DPO-based training, they are still produced by generic prompts that are not image-specific. Next, we perform the second step of our data generation pipeline to obtain a set of per-image questions and answers.", + "bbox": [ + 212, + 674, + 787, + 796 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generation of Per-Image Questions and Answers: To obtain a set of questions per image, we leverage an LLM, i.e. a variant of Mistral 7B Instruct", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "CLIP-DPO", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5e879f6d006e00e68966d810102b196309f7bf5bc56fd93a4cd95bd2db17d628.jpg", + "image_caption": [ + "Fig. 6: Question Filtering. The sample is rejected if the CLIP image-question score is low for a given sample. If the question is relevant and not generic, it is kept. Next, we evaluate the quality of the answer by first parsing the question into a caption, appending the positive answer to it, and computing its CLIP score. If it is a high score, the sample is kept for DPO-based training." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 372, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Question 1: What is the central focus of the image?", + "bbox": [ + 379, + 143, + 558, + 162 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Question CLIP Score: 15.0", + "- Question Filter: Rejected" + ], + "bbox": [ + 383, + 164, + 513, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Question 2: What type of market is depicted in the image?", + "bbox": [ + 580, + 143, + 746, + 162 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Question CLIP Score: 26.7", + "- Question Filter: Kept" + ], + "bbox": [ + 584, + 165, + 712, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Positive: A fruit and vegetable market.", + "bbox": [ + 379, + 190, + 566, + 199 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Caption: type of market is depicted in the image a fruit and vegetable market.", + "Positive CLIP Score: 29.1", + "- Question Filter: Kept" + ], + "bbox": [ + 382, + 200, + 761, + 227 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/921def778516fde821386dde71ff3277789985a0bd5f78379e894c697c348963.jpg", + "image_caption": [ + "Fig. 7: Examples of the final positive and negative pairs obtained post-filtering, to be used for DPO-based training." + ], + "image_footnote": [], + "bbox": [ + 217, + 324, + 364, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt: Provide a brief description of the given image.", + "bbox": [ + 217, + 410, + 400, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Positive: An Apple computer with a laptop computer sitting on a desk.", + "bbox": [ + 217, + 426, + 395, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Negative: A laptop and a desktop computer are sitting on a desk.", + "bbox": [ + 217, + 441, + 395, + 458 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/48e92c389da54ece0c3757c592c986106ec6927b53055777c6054d1d5c245744.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 415, + 324, + 566, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt: Provide a depiction of the picture. Positive: A man in orange shirt kicking a", + "bbox": [ + 413, + 410, + 598, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "soccer ball", + "bbox": [ + 413, + 428, + 462, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Negative: A group of people kicking a", + "bbox": [ + 413, + 435, + 573, + 443 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "soccer ball", + "bbox": [ + 413, + 443, + 465, + 450 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5aa0d5905cac1a7c42a519100dad450e5d7d57fbbfec85b4c81a9b797eea8476.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 324, + 771, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt: What is the name of the paint brand featured on the bucket?", + "bbox": [ + 609, + 410, + 771, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Positive: Homer's All Purpose Bucket", + "bbox": [ + 611, + 426, + 767, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Negative: Home Depot All-Purpose Bucket.", + "bbox": [ + 611, + 435, + 782, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "v0.2 [28] $^1$ , and feed it with the generated captions and prompt it to generate 2 questions for each image, together with positive and negative answers. The LLM is asked to generate the positive ones based on the fed captions and to generate plausible but incorrect negatives given the image description/caption.", + "bbox": [ + 212, + 527, + 785, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For an example of the generated captions, questions, and answers, see Fig. 5.", + "bbox": [ + 214, + 599, + 769, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Data Annotation", + "text_level": 1, + "bbox": [ + 214, + 638, + 401, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Starting with the generated captions, the subsequent data annotation step involves ranking them based on their CLIP image-text similarities and filtering them to obtain a final set of high-quality positive-negative pairs for DPO-based training. Our filtering pipeline includes two stages: global filtering, where we eliminate low-quality images and captions, and pair-filtering, where we select the best positive and negative per-image pairs for DPO-based training.", + "bbox": [ + 212, + 667, + 787, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CLIP Ranking. The initial step of data annotation is CLIP ranking. In this step, we use a pre-trained CLIP model to compute the cosine similarities of the captions with their associated images and then sort them from highest to lowest.", + "bbox": [ + 212, + 768, + 785, + 814 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Ouali et al.", + "bbox": [ + 271, + 114, + 369, + 127 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "1 https://huggingface.co/teknium/OpenHermes-2.5-Mistral7B", + "bbox": [ + 217, + 823, + 661, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Global Filtering. Next, we aim to filter our data and only retain a balanced and higher-quality subset. To achieve this, we start by analyzing the types of images in our data pool (see Tab. 1). We define a set of four generic categories: images of text, people, objects, and scenes. Using CLIP and a set of 10 descriptions per category, we create four class prototypes. Each image is then assigned to one of these prototypes based on the highest cosine similarity. As illustrated in Fig. 4, we observe that CLIP scores for the text category are the highest, while scores for people, objects, and scenes are relatively similar. Consequently, before applying CLIP-based filtering, we down-sample the portion of text images. We then filter out all generic captions below a given CLIP score ( $e.g. < 28.0$ ). Additionally, since CLIP is primarily trained on short text, we remove long captions to ensure more precise CLIP scores. For questions, we compute their CLIP scores and remove all questions with low CLIP scores ( $e.g. < 25.0$ ), thereby eliminating generic questions ( $e.g. \"what is the main object in the image?\")$ that are already covered by the generated captions.", + "bbox": [ + 217, + 145, + 787, + 372 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Pair Filtering. Finally, starting from the remaining high-quality image captions and image question-answers, the final step will be selecting a set of positive and negative pairs for DPO-based training. For the questions, since we already have a set of positive and negative responses generated by our LLM, we only need to filter out the low-quality pairs. To achieve this, we use simple regex matching rules to extract an image description from the question, append the positive answer to it to create a synthetic caption, and compute its CLIP image-text scores. We then reject examples where the scores are low. See Fig. 6 for an example of the questions filtering process. As for the captions, we consider all possible pairs where the CLIP score difference between two captions is larger than a given threshold ( $e.g. > 2.0$ ) and where the length of the two captions is similar to avoid introducing false preferences. We then order them based on the CLIP score difference between the positive and negative captions and select the top-ranking pair per image.", + "bbox": [ + 217, + 381, + 787, + 593 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "After the data annotation step, the resulting DPO training data consists of 750K pairs, of which 50K are question-answer pairs and the rest are caption pairs. For examples of the final pairs, see Fig. 7.", + "bbox": [ + 217, + 602, + 785, + 647 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Implementation details", + "text_level": 1, + "bbox": [ + 217, + 671, + 473, + 688 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Network architecture: In this work, we consider two LVLMs architectures: MobileVLM-v2 [14] and LLaVA-1.5 [42]. Both models follow the same overall structure: a pre-trained CLIP vision encoder and a pre-trained LLM. The visual tokens produced by the frozen vision encoder are projected using either a linear layer or a small projection module and passed as input to the LLM. LLaVA opts for a pre-trained Vicuna LLM [12] while MobileVLM-v2 uses MobileLLaMA [13], except for their 7B variant, which also uses a Vicuna model. Both use the same ViT-L-14 @ 336px CLIP visual encoder [56]. For efficiency purposes, MobileVLM-v2 uses a projection module that halves the number of", + "bbox": [ + 217, + 703, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "CLIP-DPO", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "visual tokens. We consider the following model variants in our comparisons: MobileVLM-v2 (1.7B, 3B and 7B), and LLaVA-1.5 7B. For data annotation, we use ViT-H/14 DFN [19] as our pre-trained CLIP model.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training details: For all of our experiments, unless specified otherwise, we start from the pre-trained MobileVLM-v2 and LLaVA-1.5 models. The models are then fine-tuned for 1 epoch using DPO-based optimization on the constructed CLIP-DPO dataset. We use the following hyperparameters for fine-tuning the models: AdamW optimizer [45] with a batch size of 256, a learning rate of $5e - 7$ , decreased to 0 using a cosine scheduler, a warm-up of 0.01, and a weight decay set to 0. Both during training and testing, the input images are cropped and resized to $336 \\times 336$ px. The training was performed on 8 A100 GPUs using Pytorch [54]. For the larger models, e.g. LLaVA-1.5 7B, to fit them in memory, we use the Zero-3 strategy [58, 59] and decrease the batch size to 64.", + "bbox": [ + 212, + 200, + 782, + 351 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6 Results", + "text_level": 1, + "bbox": [ + 214, + 369, + 326, + 386 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We first evaluate the effectiveness of CLIP-DPO in reducing hallucinations on the recently introduced AMBER [71] benchmark in Sec. 6.1. Moreover, we demonstrate its enhanced grounding capabilities by reporting zero-shot image classification in Sec. 6.2. Finally, in Sec. 6.4, we show that the proposed CLIP-DPO training does not compromise LVLM performance by reporting results on the LLaVA benchmark [87].", + "bbox": [ + 212, + 397, + 782, + 488 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6.1 Evaluation on hallucinations", + "text_level": 1, + "bbox": [ + 214, + 494, + 496, + 508 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We start by evaluating CLIP-DPO in terms of LVLM hallucination reduction, which is our main objective. We use AMBER [71], a comprehensive, high-quality, and LLM-free multidimensional benchmark for LVLM hallucination evaluation, which can be used to evaluate both generative and discriminative tasks. As the results from Tab. 2 show, MobileVLM-v2 trained with CLIP-DPO improves upon MobileVLM-v2 baselines across all model sizes, and sometimes quite significantly, especially for the 1.7B and 7B models. It can also be seen that CLIP-DPO significantly improves when applied on top of LLaVA-1.5 7B. Importantly, CLIP-DPO significantly outperforms HA-DPO [83], the main competing approach, improving the AMBER score 3.2 vs 7.8 when using LLaVA-1.5. Finally, our LlaVA-1.5 7B+CLIP-DPO even outperforms Qwen-VL [5], which is trained on significantly larger datasets (1.4B image-text pairs for pre-training and 77M for multitask training while CLIP-DPO is fine-tuned on just 0.7M samples), and even matches the performance of GPT-4V without using any GPT-4V model outputs during training. Overall, these results on a high-quality state-of-the-art benchmark such as AMBER clearly demonstrate the effectiveness of our CLIP-DPO approach for reducing hallucinations.", + "bbox": [ + 212, + 513, + 784, + 771 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6.2 Zero-shot image classification", + "text_level": 1, + "bbox": [ + 214, + 789, + 504, + 804 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A primary reason for the hallucinatory behavior of LVLMs is the weak alignment between their visual features and the input LLM tokens. A direct way to", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Ouali et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/ec447d08746b98cf77c922183dbeff054dc61052774bf3533f2f78433609fe61.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelGENERATIVE TASKDISCRIMINATIVE TASKAMBER
CHAIR↓Cover↑Hal↓Cog↓Acc.P.R.F1
mPLUG-Owl21.650.176.111.540.192.810.518.948.7
LLaVA11.551.048.85.542.774.121.032.760.6
MiniGPT-413.663.065.311.363.690.550.464.775.6
CogVLM5.657.223.61.369.088.960.972.383.4
mPLUG-Owl210.652.039.94.575.695.066.978.584.0
InstructBLIP8.852.238.24.476.584.579.081.786.5
Qwen-VL5.549.423.61.981.290.879.784.989.7
GPT-4V4.667.130.72.683.484.990.187.491.4
LLaVA-1.5 7B7.851.036.44.272.093.262.474.783.5
+ HA-DPO7.233.619.72.668.368.198.480.586.7
+ CLIP-DPO3.747.816.61.377.884.481.582.989.6
MobileVLM-v2 1.7B3.839.68.90.565.492.651.966.581.4
+ CLIP-DPO4.238.910.80.571.288.764.874.985.3
MobileVLM-v2 3B4.838.711.10.773.592.165.776.786.0
+ CLIP-DPO4.741.513.10.576.791.071.880.387.8
MobileVLM-v2 7B4.438.910.40.671.995.060.874.184.9
+ CLIP-DPO4.038.010.10.577.393.470.880.588.3
", + "bbox": [ + 248, + 143, + 756, + 356 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/dface6c72bb5620056fa4a5798fbac4221056d6c2bcc91a021d2ddc30e73e216.jpg", + "table_caption": [ + "Table 2: Hallucination evaluation results on AMBER for both generative (leftmost set of columns) and discriminative tasks (rightmost set of columns). Our approach offers consistent and large performance improvements across 4 different LVLM models, and also beats HA-DPO by a large margin." + ], + "table_footnote": [], + "table_body": "
MethodStanfordCarsOxfordPetsOxfordFlowersImagenetFood-101EurosatCaltech-101UCF-101SUN397Avg
LLaVA-1.523.234.07.337.945.349.582.750.543.140.0
+ HA-DPO23.233.47.136.945.149.482.750.342.739.7
+ CLIP-DPO30.144.816.342.253.952.184.853.148.847.4
MobileVLM-v2 1.7B17.115.214.532.631.249.377.945.539.333.9
+ CLIP-DPO19.232.623.841.347.448.980.850.249.343.7
MobileVLM-v2 3B14.823.59.135.738.853.281.848.642.036.7
+ CLIP-DPO28.640.919.544.352.056.485.552.350.147.7
MobileVLM-v2 7B27.132.511.537.745.043.181.949.046.641.6
+ CLIP-DPO32.551.335.850.162.659.388.356.254.854.5
", + "bbox": [ + 215, + 439, + 782, + 556 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 3: Zero-shot image recognition results in terms of Top-1 accuracy (\\%). While HA-DPO shows very similar performance to the base model, our CLIP-DPO improves the base model by a very large margin in all cases.", + "bbox": [ + 215, + 568, + 784, + 609 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "evaluate this is through simple zero-shot image classification, which is the goto benchmark for contrastively trained VL models like CLIP [56]. The typical setup follows a closed-set classification problem, where the names of all possible classes are known a priori and are encoded into class prototypes using CLIP. A given image is then assigned to the class with the highest CLIP image-text scores. Herein, we follow the same protocol with the notable difference that, as our goal is LVLM evaluation, we first prompt the model to generate a free-form image caption describing the main object in the image, then encode the caption using the CLIP text encoder and assign the image to the class with the highest text-text (i.e. caption-class) CLIP score. Here, we opted for a different family of VL embedding model, i.e. SigLIP [79], to avoid evaluating using CLIP models similar to those used during the data annotation step. Following [56,85], we evaluate our approach on a suite of 9 diverse datasets: UCF-101 [63], SUN397 [74],", + "bbox": [ + 215, + 643, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "CLIP-DPO", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/050437536d2d8c1dd5c52fc333629b4356b6d39f1bd11558b55e5ad356be51ea.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLLMRes.GQA\\( SQAI \\)\\( VQAT \\)POPE\\( MME^P \\)\\( MMB^{dev} \\)Avg.
BLIP-2 [33]Vicuna-13B22441.061.042.585.31293.8--
InstructBLIP [16]Vicuna-13B22449.563.150.778.91212.8--
Shikra [9]Vicuna-13B224-----58.8-
Openflamingo [3]MPT 7B336--33.6--4.6-
Qwen-VL [5]Qwen 7B44859.367.163.8-1487.638.2-
mPLUG-Owl [76]LLaMA 7B224----967.349.4-
MiniGPT-v2 [8]LLaMA 7B44860.3----12.2-
MiniGPT-4 [86]Vicuna 7B22432.2---581.723.0-
InstructBLIP [16]Vicuna 7B22449.260.550.1--36.0-
ShareGPT4V [10]Vicuna 7B33663.368.460.485.71567.468.870.8
MoE-LLaVA-1.6B×4 [39]StableLM-1.6B33660.462.647.884.31300.8*59.463.3
MoE-LLaVA-2.7B×4 [39]Phi-2.7B33661.168.750.285.01396.4*65.566.7
LLaVA-1.5 [41]Vicuna 7B33662.066.858.285.91510.764.368.8
+ HA-DPOVicuna 7B33661.969.258.384.31505.664.969.0
+ CLIP-DPOVicuna 7B33659.367.656.485.81468.764.967.9
MobileVLM 1.7B [13]MobileLLaMA 1.4B33656.157.341.584.51196.253.258.7
MobileVLM-v2 1.7B [14]MobileLLaMA 1.4B33658.366.752.184.31302.857.764.2
+ CLIP-DPOMobileLLaMA 1.4B33658.668.447.884.31331.157.663.8
MobileVLM- 3B [13]MobileLLaMA 2.7B33659.061.247.584.91288.959.662.8
MobileVLM-v2 3B [14]MobileLLaMA 2.7B33661.170.057.584.71440.563.268.1
+ CLIP-DPOMobileLLaMA 2.7B33660.972.357.385.11425.763.868.4
MobileVLM-v2 [14] 7BVicuna 7B33662.674.862.385.31560.769.272.1
+ CLIP-DPO 7BVicuna 7B33662.577.362.284.71539.170.472.3
", + "bbox": [ + 233, + 143, + 769, + 421 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 4: Results on several vision language tasks from the LlaVA benchmark. Our approach is combined with LVLMs of different sizes, showing consistent performance in all cases.", + "bbox": [ + 215, + 431, + 787, + 473 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Stanford Cars [32], Oxford Pets [53], Oxford flowers [49], ImageNet [17], Food 101 [6], Eurosat [23] and Caltech-101 [20]. As the results from Tab. 3 show, all LVLMs fine-tuned with CLIP-DPO significantly outperform their corresponding baselines, showcasing the increased discriminative properties. Note again that these improvements are obtained without affecting the model's performance on other tasks and datasets.", + "bbox": [ + 212, + 500, + 787, + 590 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6.3 Additional evaluations", + "text_level": 1, + "bbox": [ + 215, + 609, + 446, + 625 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Herein, we evaluate the impact of CLIP-DPO training on other vision language tasks and, more specifically, on the popular LLaVA-Bench (GQA [26], ScienceQA [46], TextVQA [62], MME [21], MMBench [44]). As Tab. 4 shows, overall, CLIP-DPO training does not compromise performance. We further note that accuracy improvements on LLaVA-Bench are heavily tied to the addition of extra training data [10] or architectural changes [13,14,39,41]. As we are not using any additional data & models nor making architectural changes, it is not surprising that the performance after CLIP-DPO training remains largely in line with that of the original baseline model.", + "bbox": [ + 212, + 632, + 787, + 768 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6.4 Qualitative Results", + "text_level": 1, + "bbox": [ + 215, + 787, + 419, + 801 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In Figure 8, we show some qualitative examples comparing the predictions of LLaVA-1.5 7B, LLaVA-1.5 7B+HA-DPO, and LLaVA-1.5 7B+CLIP-DPO. Over", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Ouali et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2a6af99c9dfdbc1d2d4c156f2e4f9ea6ff9fc41ff05667cc675777cddd9fef41.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 245, + 143, + 359, + 203 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Provide a short description of the image.", + "bbox": [ + 372, + 148, + 558, + 159 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "LLava-1.5-7b: A large silver airplane is parked inside a hangar. The plane is positioned on a runway, and there is a ladder leading up to the cockpit. The hangar appears to be a large building, possibly a factory or a maintenance facility.", + "bbox": [ + 370, + 162, + 746, + 188 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "+ HA-DPO: A large silver airplane is parked inside a hangar.", + "+ CLIP-DPO: A fighter jet is parked inside a hangar, with a ladder attached to it. The jet is numbered 157." + ], + "bbox": [ + 377, + 190, + 732, + 220 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b05df5ad4610a422e25687c4927231823f6d240d4be8c0a3e0c1692675623e15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 245, + 233, + 359, + 292 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Provide a short description of the image.", + "bbox": [ + 372, + 238, + 558, + 248 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "LLava-1.5-7b: A green bus is parked in a parking lot, waiting for passengers.", + "+ HA-DPO: A green bus is parked in a parking lot, waiting for passengers.", + "+CLIP-DPO: A green Go3 bus is parked at a bus stop." + ], + "bbox": [ + 372, + 250, + 720, + 285 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/d9f04fbb7f31602eef73f5f1a591f99f4bb00af84571cf1c5504e9b9fd033e28.jpg", + "table_caption": [ + "Fig. 8: Qualitative examples comparing our method with LLaVA and HA-DPO." + ], + "table_footnote": [], + "table_body": "
Loss variantAMBER ScoreCls. Avg.LLaVA bench.
DPO (β = 0.1) [57]85.343.763.8
DPO (β = 0.3) [57]83.342.164.0
IPO [4]83.640.564.5
KTO [18]83.539.564.7
SLIC [43]83.242.764.0
cDPO (β = 0.1, LS = 0.1 [43])83.143.263.9
cDPO (β = 0.3, LS = 0.1 [43])83.140.764.5
", + "bbox": [ + 300, + 330, + 702, + 434 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 5: Effect of DPO loss variant on CLIP-DPO training. Results are reported in terms of AMBER Score, Average Top-1 (\\%) accuracy across all 9 image classification datasets, and the average on LLaVA benchmark. MobileVLM-v2 1.7B was used.", + "bbox": [ + 214, + 444, + 787, + 486 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "all, with CLIP-DPO, the model is more grounded in the visual content, less prone to hallucinations, and more precise and fine-grained in its descriptions.", + "bbox": [ + 214, + 516, + 784, + 547 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "7 Ablation studies", + "text_level": 1, + "bbox": [ + 215, + 571, + 410, + 585 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Herein, we ablate the effect of the proposed components. In all experiments, we use a MobileVLM-v2 1.7B model.", + "bbox": [ + 214, + 604, + 784, + 633 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "7.1 Effect of DPO loss variant", + "text_level": 1, + "bbox": [ + 215, + 657, + 478, + 672 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Our work primarily follows the original DPO formulation proposed in [57]. Here we also consider the following recently proposed variations: KTO [18], ITO [4], SLIC [43] and cDPO [48]. As Tab. 5 shows, aggregated, all losses tend to perform similarly, with DPO marginally outperforming the others.", + "bbox": [ + 212, + 684, + 784, + 744 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "7.2 Effect of CLIP scorer", + "text_level": 1, + "bbox": [ + 215, + 768, + 437, + 782 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Herein, we seek to explore alternatives to ViT-H/14 DFN [19] CLIP model used in previous experiments, analyzing the impact of the scorer used on the overall performance of the model. We consider a diverse set of alternatives, covering", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "CLIP-DPO", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/45ddaa57bd5c0534c3ae4c9a8d3bffb2bea91dd966637d8bff7525d054fd4a06.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ScorerAMBER ScoreCls. Avg.LLaVA bench.
ViT-L/14 [56]79.842.363.0
SigLIP-L/16 [79]85.043.963.2
ViT-H/14 [60]84.344.263.7
ViT-H/14 DFN [19]85.343.763.8
", + "bbox": [ + 330, + 143, + 671, + 213 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 6: Effect of CLIP scorer on CLIP-DPO training. Results are reported in terms of AMBER Score, Average Top-1 (\\%) accuracy across all 9 image classification datasets, and the average on LLaVA benchmark. MobileVLM-v2 1.7B was used.", + "bbox": [ + 212, + 223, + 785, + 265 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "multiple exploratory paths: equally-sized models trained on different data (ViT-H/14 [19] trained on LAION-2B instead of DFN-5B); smaller models (ViT-L/14 [56]) and models trained using different pre-training losses (SigLIP-ViT-L/16 [79]). As the results from Tab. 6 show, our approach is generally robust to the exact scorer used, with the notable exception of the ViT-L/14 [56] model. Notice that this difference is primarily manifesting on the AMBER benchmark. Intuitively, this showcases the importance of a powerful scorer for reducing the amount of hallucinations.", + "bbox": [ + 212, + 294, + 785, + 414 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations and broader impact", + "text_level": 1, + "bbox": [ + 215, + 438, + 526, + 455 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "As an LVLM-based approach, our method is subject to the same general consideration (i.e. potential data bias, susceptibility to hallucinations, etc.). Moreover, as the LVLMs are trained on relatively small datasets compared to LLMs or CLIP, gaps within their knowledge domains are possible. This is especially important as neural networks tend to be overconfident outside their seen input distribution. As with all models from this category, we strongly recommend checking the models and the data carefully before deploying them. Despite these general aspects, our approach is shown to significantly reduce the amount of hallucinated content and improve the model's discriminability, hence resulting in more robust and reliable models.", + "bbox": [ + 212, + 468, + 785, + 619 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "8 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 642, + 369, + 657 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we proposed CLIP-DPO, a simple method that reduces hallucinations in LVLMs based on using a pre-trained CLIP model [56] to rank the LVLM's self-generated captions in order to construct positive-negative pairs for DPO fine-tuning. In contrast to previous works, our proposed CLIP-DPO removes the need of (i) paid-for APIs, (ii) additional data and (iii) additional external LVLMs. Before training, the data is filtered using a newly proposed robust rule-based approach. When applied on top of established LVLMs, it is shown that CLIP-DPO fine-tuning significantly reduces hallucinations, outperforming the baseline models by significant margins and even matching the performance of GPT-4V on the AMBER benchmark. We also observe superior zero-shot object recognition, suggesting improved object grounding capabilities.", + "bbox": [ + 212, + 672, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Ouali et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Detailed caption dataset. https://huggingface.co/datasets/echo840/Detailed_Caption (2024)", + "2. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)", + "3. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., et al.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)", + "4. Azar, M.G., Rowland, M., Piot, B., Guo, D., Calandriello, D., Valko, M., Munos, R.: A general theoretical paradigm to understand learning from human preferences. arXiv preprint arXiv:2310.12036 (2023)", + "5. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-VL: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966 (2023)", + "6. Bossard, L., Guillaumin, M., Van Gool, L.: Food-101-mining discriminative components with random forests. In: European Conference on Computer Vision (2014)", + "7. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances on Neural Information Processing Systems (2020)", + "8. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023)", + "9. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023)", + "10. Chen, L., Li, J., Dong, X., Zhang, P., He, C., Wang, J., Zhao, F., Lin, D.: Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793 (2023)", + "1. Chen, Z., Wu, J., Wang, W., Su, W., Chen, G., Xing, S., Muyan, Z., Zhang, Q., Zhu, X., Lu, L., et al.: Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238 (2023)", + "2. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023)", + "3. Chu, X., Qiao, L., Lin, X., Xu, S., Yang, Y., Hu, Y., Wei, F., Zhang, X., Zhang, B., Wei, X., et al.: MobileVLM: A fast, reproducible and strong vision language assistant for mobile devices. arXiv preprint arXiv:2312.16886 (2023)", + "4. Chu, X., Qiao, L., Zhang, X., Xu, S., Wei, F., Yang, Y., Sun, X., Hu, Y., Lin, X., Zhang, B., et al.: MobileVLM V2: Faster and stronger baseline for vision language model. arXiv preprint arXiv:2402.03766 (2024)", + "5. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022)", + "6. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: InstructBLIP: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023)" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "CLIP-DPO", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: IEEE Conference on Computer Vision and Pattern Recognition (2009)", + "18. Ethayarajh, K., Xu, W., Muennighoff, N., Jurafsky, D., Kiela, D.: Kto: Model alignment as prospect theoretic optimization. arXiv preprint arXiv:2402.01306 (2024)", + "19. Fang, A., Jose, A.M., Jain, A., Schmidt, L., Toshev, A., Shankar, V.: Data filtering networks. arXiv preprint arXiv:2309.17425 (2023)", + "20. Fei-Fei, L., Fergus, R., Perona, P.: Learning generative visual models from few training examples: An incremental Bayesian approach tested on 101 object categories. In: IEEE Conference on Computer Vision and Pattern Recognition - Workshops (2004)", + "21. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023)", + "22. Gunjal, A., Yin, J., Bas, E.: Detecting and preventing hallucinations in large vision language models. arXiv preprint arXiv:2308.06394 (2023)", + "23. Helber, P., Bischke, B., Dengel, A., Borth, D.: EuroSAT: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 12(7), 2217-2226 (2019)", + "24. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)", + "25. Hu, H., Zhang, J., Zhao, M., Sun, Z.: Ciem: Contrastive instruction evaluation method for better instruction tuning. arXiv preprint arXiv:2309.02301 (2023)", + "26. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: IEEE Conference on Computer Vision and Pattern Recognition (2019)", + "27. Jain, J., Yang, J., Shi, H.: Vcoder: Versatile vision encoders for multimodal large language models. arXiv preprint arXiv:2312.14233 (2023)", + "28. Jiang, A.Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D.S., Casas, D.d.l., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., et al.: Mistral 7b. arXiv preprint arXiv:2310.06825 (2023)", + "29. Jiang, C., Xu, H., Dong, M., Chen, J., Ye, W., Yan, M., Ye, Q., Zhang, J., Huang, F., Zhang, S.: Hallucination augmented contrastive learning for multimodal large language model. arXiv preprint arXiv:2312.06968 (2023)", + "30. Jing, L., Li, R., Chen, Y., Jia, M., Du, X.: Faithscore: Evaluating hallucinations in large vision-language models. arXiv preprint arXiv:2311.01477 (2023)", + "31. Koh, J.Y., Salakhutdinov, R., Fried, D.: Grounding language models to images for multimodal inputs and outputs. International Conference on Machine Learning (2023)", + "32. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3D object representations for fine-grained categorization. In: IEEE International Conference on Computer Vision - Workshops (2013)", + "33. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)", + "34. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Learning Representations (2022)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Ouali et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Li, L., Xie, Z., Li, M., Chen, S., Wang, P., Chen, L., Yang, Y., Wang, B., Kong, L.: Silkie: Preference distillation for large visual language models. arXiv preprint arXiv:2312.10665 (2023)", + "36. Li, S., Lin, R., Pei, S.: Multi-modal preference alignment remedies regression of visual instruction tuning on language model. arXiv preprint arXiv:2402.10884 (2024)", + "37. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023)", + "38. Li, Y., Bubeck, S., Eldan, R., Del Giorno, A., Gunasekar, S., Lee, Y.T.: Textbooks are all you need ii: phi-1.5 technical report. arXiv preprint arXiv:2309.05463 (2023)", + "39. Lin, B., Tang, Z., Ye, Y., Cui, J., Zhu, B., Jin, P., Zhang, J., Ning, M., Yuan, L.: Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947 (2024)", + "40. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Mitigating hallucination in large multi-modal models via robust instruction tuning. In: International Conference on Learning Representations (2023)", + "41. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)", + "42. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances on Neural Information Processing Systems (2024)", + "43. Liu, T., Zhao, Y., Joshi, R., Khalman, M., Saleh, M., Liu, P.J., Liu, J.: Statistical rejection sampling improves preference optimization. arXiv preprint arXiv:2309.06657 (2023)", + "44. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: MMBench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281 (2023)", + "45. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)", + "46. Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems (2022)", + "47. Mishra, S., Khashabi, D., Baral, C., Hajishirzi, H.: Cross-task generalization via natural language crowdsourcing instructions. arXiv preprint arXiv:2104.08773 (2021)", + "48. Mitchell, E.: A note on dpo with noisy preferences & relationship to ipo (2024)", + "49. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: Indian conference on computer vision, graphics & image processing. pp. 722-729 (2008)", + "50. OpenAI: Introducing chatgpt (2022)", + "51. OpenAI: Gpt-4v(ision) system card (2023)", + "52. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems (2022)", + "53. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: IEEE Conference on Computer Vision and Pattern Recognition (2012)", + "54. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch (2017)", + "55. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "CLIP-DPO", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "56. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021)", + "57. Rafailov, R., Sharma, A., Mitchell, E., Manning, C.D., Ermon, S., Finn, C.: Direct preference optimization: Your language model is secretly a reward model. Advances on Neural Information Processing Systems (2024)", + "58. Rajbhandari, S., Ruwase, O., Rasley, J., Smith, S., He, Y.: Zero-infinity: Breaking thegpu memory wall for extreme scale deep learning. In: Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis (2021)", + "59. Rasley, J., Rajbhandari, S., Ruwase, O., He, Y.: Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In: Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (2020)", + "60. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems (2022)", + "61. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)", + "62. Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards VQA models that can read. In: IEEE Conference on Computer Vision and Pattern Recognition (2019)", + "63. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)", + "64. Stiennon, N., Ouyang, L., Wu, J., Ziegler, D., Lowe, R., Voss, C., Radford, A., Amodei, D., Christiano, P.F.: Learning to summarize with human feedback. Advances in Neural Information Processing Systems (2020)", + "65. Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L.Y., Wang, Y.X., Yang, Y., et al.: Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525 (2023)", + "66. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)", + "67. Thoppilan, R., De Freitas, D., Hall, J., Shazeer, N., Kulshreshtha, A., Cheng, H.T., Jin, A., Bos, T., Baker, L., Du, Y., et al.: Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239 (2022)", + "68. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)", + "69. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)", + "70. Wang, J., Yang, Z., Hu, X., Li, L., Lin, K., Gan, Z., Liu, Z., Liu, C., Wang, L.: Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100 (2022)", + "71. Wang, J., Wang, Y., Xu, G., Zhang, J., Gu, Y., Jia, H., Yan, M., Zhang, J., Sang, J.: An LLM-free multi-dimensional benchmark for MLLMs hallucination evaluation. arXiv preprint arXiv:2311.07397 (2023)" + ], + "bbox": [ + 215, + 147, + 784, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Ouali et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "72. Wang, J., Zhou, Y., Xu, G., Shi, P., Zhao, C., Xu, H., Ye, Q., Yan, M., Zhang, J., Zhu, J., et al.: Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126 (2023)", + "73. Williams, R.J.: Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning 8, 229-256 (1992)", + "74. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: Sun database: Large-scale scene recognition from abbey to zoo. In: IEEE Conference on Computer Vision and Pattern Recognition (2010)", + "75. Xu, J., Lee, A., Sukhbaatar, S., Weston, J.: Some things are more cringe than others: Preference optimization with the pairwise cringe loss. arXiv preprint arXiv:2312.16682 (2023)", + "76. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)", + "77. You, H., Zhang, H., Gan, Z., Du, X., Zhang, B., Wang, Z., Cao, L., Chang, S.F., Yang, Y.: Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704 (2023)", + "78. Zhai, B., Yang, S., Zhao, X., Xu, C., Shen, S., Zhao, D., Keutzer, K., Li, M., Yan, T., Fan, X.: Halle-switch: Rethinking and controlling object existence hallucinations in large vision language models for detailed caption. arXiv preprint arXiv:2310.01779 (2023)", + "79. Zhai, X., Mustafa, B., Kolesnikov, A., Beyer, L.: Sigmoid loss for language image pre-training. arXiv preprint arXiv:2303.15343 (2023)", + "80. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022)", + "81. Zhao, H., Cai, Z., Si, S., Ma, X., An, K., Chen, L., Liu, Z., Wang, S., Han, W., Chang, B.: Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915 (2023)", + "82. Zhao, Y., Joshi, R., Liu, T., Khalman, M., Saleh, M., Liu, P.J.: Slic-hf: Sequence likelihood calibration with human feedback. arXiv preprint arXiv:2305.10425 (2023)", + "83. Zhao, Z., Wang, B., Ouyang, L., Dong, X., Wang, J., He, C.: Beyond hallucinations: Enhancing LVLMs through hallucination-aware direct preference optimization. arXiv preprint arXiv:2311.16839 (2023)", + "84. Zhou, C., Liu, P., Xu, P., Iyer, S., Sun, J., Mao, Y., Ma, X., Efrat, A., Yu, P., Yu, L., et al.: Lima: Less is more for alignment. Advances in Neural Information Processing Systems (2024)", + "85. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: IEEE Conference on Computer Vision and Pattern Recognition (2022)", + "86. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: MiniGPT-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)", + "87. Zhu, Y., Zhu, M., Liu, N., Ou, Z., Mou, X., Tang, J.: LLaVA-phi: Efficient multi-modal assistant with small language model. arXiv preprint arXiv:2401.02330 (2024)", + "88. Ziegler, D.M., Stiannon, N., Wu, J., Brown, T.B., Radford, A., Amodei, D., Christiano, P., Irving, G.: Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593 (2019)" + ], + "bbox": [ + 215, + 146, + 784, + 825 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "CLIP-DPO", + "bbox": [ + 653, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/CLIP-DPO_ Vision-Language Models as a Source of Preference for Fixing Hallucinations in LVLMs/174c66bb-6859-4a7e-835b-0b1709492a92_model.json b/2024/CLIP-DPO_ Vision-Language Models as a Source of Preference for Fixing Hallucinations in LVLMs/174c66bb-6859-4a7e-835b-0b1709492a92_model.json new file mode 100644 index 0000000000000000000000000000000000000000..eb72f581fb798c0e16550cb113216f1072d8c361 --- /dev/null +++ b/2024/CLIP-DPO_ Vision-Language Models as a Source of Preference for Fixing Hallucinations in LVLMs/174c66bb-6859-4a7e-835b-0b1709492a92_model.json @@ -0,0 +1,3164 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.22, + 0.141, + 0.783, + 0.186 + ], + "angle": 0, + "content": "CLIP-DPO: Vision-Language Models as a Source of Preference for Fixing Hallucinations in LVLMs" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.212, + 0.768, + 0.229 + ], + "angle": 0, + "content": "Yassine Ouali1 Adrian Bulat1,2 Brais Martinez1 Georgios Tzimiropoulos1,3" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.242, + 0.786, + 0.274 + ], + "angle": 0, + "content": "\\(^{1}\\)Samsung AI Center Cambridge, UK \\(^{2}\\) Technical University of Iasi, Romania \\(^{3}\\) Queen Mary University of London, UK" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.309, + 0.744, + 0.573 + ], + "angle": 0, + "content": "Abstract. Despite recent successes, LVLMs or Large Vision Language Models are prone to hallucinating details like objects and their properties or relations, limiting their real-world deployment. To address this and improve their robustness, we present CLIP-DPO, a preference optimization method that leverages contrastively pre-trained Vision-Language (VL) embedding models, such as CLIP, for DPO-based optimization of LVLMs. Unlike prior works tackling LVLM hallucinations, our method does not rely on paid-for APIs, and does not require additional training data or the deployment of other external LVLMs. Instead, starting from the initial pool of supervised fine-tuning data, we generate a diverse set of predictions, which are ranked based on their CLIP image-text similarities, and then filtered using a robust rule-based approach to obtain a set of positive and negative pairs for DPO-based training. We applied CLIP-DPO fine-tuning to the MobileVLM-v2 family of models and to LlaVA-1.5, in all cases observing significant improvements in terms of hallucination reduction over baseline models. We also observe better performance for zero-shot classification, suggesting improved grounding capabilities, and verify that the original performance on standard LVLM benchmarks is overall preserved." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.598, + 0.377, + 0.615 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.844 + ], + "angle": 0, + "content": "A Large Vision Language Model (LVLM) is a Large Language Model (LLM) combined with an independently-trained vision encoder, typically taken from VL embedding models like CLIP [56] and often kept frozen. LVLMs have recently become the state-of-the-art for vision language understanding [5, 16, 42, 65, 81]. Unlike the prior generation of models [34, 70], which were typically tuned for one or a small number of tasks, LVLMs allow free-form dialog in natural language with an image in the loop. This direction largely draws inspiration from the recent success of ChatGPT [50] and respectively, ChatGPT-4V [51], which adapts LLMs to follow human instruction and preferences via various forms of reinforcement learning from human feedback (RFLH) and/or supervised fine-tuning (SFT) on high-quality multi-turn instruction (i.e. chat) data. In a similar fashion, the current generation of LVLMs is trained as part of a two-step process [42]. First, the CLIP vision encoder is aligned with the LLM by training a few adaptation layers only and, then, by fine-tuning the model using SFT on a multi-modal" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "Y. Ouali et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.193 + ], + "angle": 0, + "content": "instructional dataset. Despite their remarkable success, a major limitation hindering the wider adoption of LVLMs is the high rate of hallucinated details (e.g. non-existing objects) that these models exhibit in their generated output text." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.197, + 0.787, + 0.409 + ], + "angle": 0, + "content": "Motivation. Following the footsteps of LLM training, a natural solution to combat hallucinations is given by RFLH. As reinforcement learning-based training is generally expensive, often requiring the training of a policy model, Rafailov et al. [57] proposed a simplified framework, coined Direct Preference optimization (DPO), that allows direct training of the model using a binary cross-entropy objective. The data for DPO is typically obtained either from human preferences or using LLMs to construct preference data automatically. Thanks to its efficiency, DPO has been quickly adopted for LVLMs too [35, 36, 83]. However, these approaches require multiple LVLMs for generation, collecting additional data, and using GPT-4/GPT-4V API for ranking. Moreover, the GPT-4 provided scores are discrete (hence, of reduced granularity) and are themselves prone to hallucinations, which can perpetuate and exacerbate the already high rate of factual errors and poor image grounding [37, 71]. Finally, GPT-4/V is behind a paywall, and using it for generation/ranking is neither scalable nor cost-efficient." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.413, + 0.787, + 0.581 + ], + "angle": 0, + "content": "Main idea. To fix hallucinations and address the high data construction and annotation cost exhibited by current methods, we propose a new DPO variant, called CLIP-DPO, that uses a pre-trained CLIP model [56] to rank the LVLM's self-generated captions to construct positive-negative pairs for DPO. Since the CLIP model was trained in a contrastive manner to measure the alignment between image-text pairs, it is naturally suitable for determining the quality of a given output from an LVLM, grounding it to the correct object or attribute. The dataset over which we operate is constructed by running the original pre-trained target LVLM on its own output obtained using prompting, removing both the need for (i) additional external data and (ii) ensembles of external LVLMs. The final data is filtered before training using robust rule-based filtering." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.585, + 0.787, + 0.736 + ], + "angle": 0, + "content": "Main results. We applied CLIP-DPO fine-tuning on top of two state-of-the-art models: MobileVLM-v2 (3 models in total) and LlaVA-1.5 7B. We find that, in both cases, our approach is effective in reducing hallucinations, outperforming all baseline models (i.e. the models without CLIP-DPO fine-tuning) by a significant margin. Importantly, CLIP-DPO significantly outperforms our direct competitor HA-DPO [83], outperforms Qwen-VL [5] trained on significantly larger datasets (i.e. 1.4B samples for pre-training and 77M for multitask training; as opposed to CLIP-DPO training on just 0.7M samples). Finally, our model's enhanced object grounding capabilities are also illustrated for zero-shot image classification, all without degrading the original performance of the base LVLM model." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.761, + 0.382, + 0.777 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.795, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Large Visual Language Models (LVLMs). Following the unprecedented success of Large Language Models (LLMs) [7, 50, 68, 69, 80], several works have recently proposed to build multi-modal capabilities on top of them [10, 14, 16, 42]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.654, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLIP-DPO" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.146, + 0.786, + 0.403 + ], + "angle": 0, + "content": "LLaVA [42] and FROMAGe [31] directly pass the visual tokens produced by a pre-trained CLIP [56] vision encoder to an LLM, either fine-tuning the LLM or adapting it using LoRA [24]. Notably, LLaVA training includes a pre-training stage that aligns the CLIP features with the LLM input tokens using a simple projection layer (i.e. keeping the rest of the model frozen) using image captioning data. InstructBLIP [16] uses QFormer [33] to reduce the number of vision tokens before passing them to the LLM. As the quality and distribution of the training data plays a crucial role [84], a series of methods [9,10,55,76] introduced improved data construction pipelines. For example, ShareGPT4v [10] uses the API of GPT-4V [51] to first label, then train a model, and finally use it to re-announce a new training set. Another line of work is improving efficiency by reducing the models' size (original LLaVA models have 7B and 13B parameters). LLaVA-Phi [87], MobileVLM [13] and its follow-up, MobileVLM-v2 [14], replace the LLaMA [68] and Vicuna [12] LLMs with the smaller 1.4B and 2.7B variants of MobileLLaMA and Phi [38]. Our work is orthogonal to the aforementioned methods and does not seek to improve the model's architecture. Instead, we propose CLIP-DPO, an improved training approach for LVLMs based on DPO." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.411, + 0.786, + 0.787 + ], + "angle": 0, + "content": "Preference optimization. Instruction tuning can significantly improve LLMs' perceived output quality and usefulness by aligning their responses to a given task domain or human preferences. This is achieved either by direct fine-tuning on expert data [15, 47, 67] or via reinforcement learning [52, 64, 88]. The latter significantly simplifies the data collection process, but still requires complicated training algorithms based on REINFORCE [73] or PPO [61]. Recently, a much-simplified approach was proposed, Direct Preference optimization (DPO) [57], which bypasses the need to train a reward model and allows direct training using a cross-entropy loss. Multiple improved versions for LLMs have been proposed in the meantime [4, 75, 82]. Following this, a recent wave of works on combining DPO with LVLMs have been proposed [35, 36, 83]. Silkie [35] constructs a multimodal instructional dataset automatically labeled by GPT-4V. Similarly, HA-DPO [83], aiming to reduce the rate of hallucinations, uses a GPT4 model to label and construct positive-negative pairs with and without hallucinations. The work of [36] follows a similar path by using a suite of LLMs and LVLMs (i.e. Gemini-Vision [66]) to generate and label the data. These methods are then primarily evaluated on the LLaVA benchmark and PoPE [72]. In contrast to the aforementioned works, the proposed CLIP-DPO (i) simplifies the pipeline, (ii) removes the need for paid APIs, (iii) removes the need for additional data, and (iv) removes the need for additional external LVLMs. Instead, we make use of a pre-trained CLIP model to rank the generated outputs from a small pool of efficient LVLMs, and show that the proposed CLIP-DPO training provides significant improvements for fixing LVLM hallucinations and, in general, for enhancing the discriminability and robustness of the model as demonstrated by image classification experiments." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.795, + 0.786, + 0.84 + ], + "angle": 0, + "content": "LVLM hallucinations. Broadly speaking, in the context of LVLMs, we consider hallucinations to be incorrect or misleading generated text, contradicting the visual evidence provided by the input image. This is an undesirable characteristic" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.128 + ], + "angle": 0, + "content": "Y. Ouali et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.365, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.372, + 0.145, + 0.766, + 0.164 + ], + "angle": 0, + "content": "Original Caption: In this image, we see a black dog wearing an orange vest standing on the grass. The dog's head is down and it appears to be in a backyard." + }, + { + "type": "text", + "bbox": [ + 0.369, + 0.166, + 0.784, + 0.193 + ], + "angle": 0, + "content": "- Existence Hallucination: In this image, we see a black dog wearing an orange vest standing on the grass, with a small orange ball near his feet. The dog's head is down and it appears to be in a backyard." + }, + { + "type": "text", + "bbox": [ + 0.369, + 0.194, + 0.772, + 0.212 + ], + "angle": 0, + "content": "- Attributes Hallucination: In this image, we see a dark brown dog wearing an orange vest standing on the grass. The dog's head is down and it appears to be in a backyard." + }, + { + "type": "text", + "bbox": [ + 0.369, + 0.213, + 0.775, + 0.231 + ], + "angle": 0, + "content": "- Relation Hallucination: In this image, we see a black dog wearing an orange vest sitting on the grass. The dog's head is down and it appears to be in a backyard." + }, + { + "type": "list", + "bbox": [ + 0.369, + 0.166, + 0.784, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.242, + 0.788, + 0.271 + ], + "angle": 0, + "content": "Fig. 1: An example of the injected hallucinations. Given an image and its caption, we prompt GPT-4 to generate 3 types of hallucination: existence, attributes, and relation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.301, + 0.789, + 0.543 + ], + "angle": 0, + "content": "inherited from the pre-trained LLM used, and further exacerbated by the visual-language alignment process [5, 25, 29, 40]. Multiple solutions have been recently proposed by the community with varying degrees of success. The works of [25, 77] attempt to address the data bias by constructing better-grounded annotated image-text pairs. The works of [5, 11] scale the resolution of the image encoder, as this was observed to reduce the amount of hallucination, but at the cost of a high increase in the computational cost. The works of [27, 78] improve the vision encoder by adding extra informational paths. The closest work related to ours is that of [83], which, with the help of GPT-4V, constructs negative-positive pairs for DPO fine-tuning. Different from [83], we use a pre-trained CLIP model to perform the ranking, and no additional data or external LVLMs are required. For evaluation purposes, we use AMBER [71], the most comprehensive benchmark for hallucinations to date, encompassing both a generative and discriminative evaluation component. Importantly, unlike all prior benchmarks [22, 30, 37, 40, 65], AMBER is a high-quality dataset fully annotated by humans for both the generative and discriminative tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.568, + 0.383, + 0.584 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.602, + 0.707, + 0.618 + ], + "angle": 0, + "content": "3.1 The Effectiveness of CLIP in Reducing Hallucinations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.794 + ], + "angle": 0, + "content": "First, we probe the effectiveness of CLIP in accurately ranking correct captions over those with hallucinated content. This assessment will indicate the efficacy of CLIP-style models in reducing hallucinations with DPO-based training. To this end, initially, we select 1K captions from the Detailed Caption [1] dataset, chosen for its high-quality labels. Then, we instruct GPT-4 [2] to create three hallucinated captions for each image, corresponding to the following types of hallucination: (i) existence: new elements or objects are added to the caption that were not mentioned originally, (ii) attribute: the attributes, characteristics, or features of the original caption's elements are altered, and (iii) relationship: the spatial, contextual, or interactive relationships of the original caption's elements are altered. See Fig. 1 for an example of the injected hallucinations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Next, we use LLaVA-1.5 7B [41] to compute the likelihood of all captions, including the original and the hallucinated captions. We then retain only the hallucinated captions for which the model predicts a higher likelihood than the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.654, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLIP-DPO" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "original, yielding 61 for existence, 76 for attribute, and 119 for relationship hallucinations out of the initial 1K captions. Finally, we compute the CLIP" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.5, + 0.342 + ], + "angle": 0, + "content": "image-text scores for these likely yet incorrect captions. As illustrated in Fig.2, CLIP manages to accurately rank the original caption over the hallucinated ones in at least \\(59\\%\\) of cases, reaching up to \\(83\\%\\) for the existence type of hallucinations. These findings underscore the potential of using VL embedding models to provide a reliable training signal for hallucination reduction in LVLMs." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.179, + 0.763, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.367, + 0.411, + 0.397 + ], + "angle": 0, + "content": "3.2 Direct Preference Optimization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.5, + 0.454 + ], + "angle": 0, + "content": "The standard pipeline for LLM and LVLM alignment consists of three steps: generation, annotation, and op" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.454, + 0.788, + 0.604 + ], + "angle": 0, + "content": "timization. First, a set of \\(N\\) prompts \\(\\{x_{i}\\}_{i = 1}^{N}\\), each one consisting of an image and a text component in the case of LVLMs, are used to generate a set of response pairs \\(y_{i}^{1}\\) and \\(y_{i}^{2}\\) obtained from a pool of pre-trained LVLMs for each prompt \\(x_{i}\\). Then, either human annotators or another set of LLMs (i.e. AI annotators) are used to rank the responses, resulting in a preferred \\(y_{i}^{+}\\) and a less preferred response \\(y_{i}^{-}\\) for each prompt \\(x_{i}\\), and thus a final preference dataset \\(\\mathcal{D} = \\{(x_i,y_i^+,y_i^-)\\}_{i = 1}^N\\). Finally, Direct Preference Optimization [57] (DPO) can be applied to update the target policy \\(\\pi_{\\theta}\\) parameterized by \\(\\theta\\) directly using the preference dataset \\(\\mathcal{D}\\). Specifically, the DPO optimization objective is defined as follows:" + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.36, + 0.788, + 0.442 + ], + "angle": 0, + "content": "Fig. 2: In dark blue, we show the number of hallucinated captions per type that LLaVA-1.5 7B assigns a higher likelihood than the original out of 1K captions. Light blue shows the portion of samples corrected by CLIP." + }, + { + "type": "equation", + "bbox": [ + 0.261, + 0.613, + 0.787, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\pi_ {\\theta}} \\mathbb {E} _ {(x _ {i}, y _ {i} ^ {+}, y _ {i} ^ {-}) \\sim \\mathcal {D}} \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta} (y _ {i} ^ {+} \\mid x _ {i})}{\\pi_ {\\mathrm {r e f}} (y _ {i} ^ {+} \\mid x _ {i})} - \\beta \\log \\frac {\\pi_ {\\theta} (y _ {i} ^ {-} \\mid x _ {i})}{\\pi_ {\\mathrm {r e f}} (y _ {i} ^ {-} \\mid x _ {i})}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.734 + ], + "angle": 0, + "content": "with \\(\\pi_{\\theta}\\) as the policy to be learned, \\(\\pi_{\\mathrm{ref}}\\) as the reference SFT policy, and \\(\\beta\\) as a hyperparameter to control the Kullback-Leibler (KL) divergence between the learned \\(\\pi_{\\theta}\\) and reference \\(\\pi_{\\mathrm{ref}}\\) policies [57]. The main benefit of DPO-based optimization is the direct alignment of the LVLM towards the preferences implicit in the preference data \\(\\mathcal{D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In contrast to previous DPO-based LVLM preference optimization works [35, 36, 83], next, we will introduce the proposed CLIP-DPO, which (i) simplifies the generation process by limiting the pool of LVLMs used for generation to small and efficient models, (ii) removes the need for external LLMs and LVLMs annotators accessed via paid APIs (e.g. GPT-3.5, GPT-4 or GPT-4V) by using CLIP as the ranking model, and (iii) removes the need for additional data by reusing the same data used during the SFT step." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.128 + ], + "angle": 0, + "content": "Y. Ouali et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.788, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.329 + ], + "angle": 0, + "content": "Fig. 3: CLIP-DPO. Starting from the initial SFT data pool and a set of prompts, an LVLM generates captions. These captions are first ranked using a CLIP model, then filtered to identify the most suitable positive and negative pairs for DPO-based optimization." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.356, + 0.331, + 0.373 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.784, + 0.463 + ], + "angle": 0, + "content": "In this section, we introduce our DPO variant for LVLMs preference optimization, named CLIP-DPO. As illustrated in Fig. 3, CLIP-DPO follows a three-step process: generation, annotation, and optimization, similar to the standard DPO-based training pipeline. While the final DPO-based optimization step remains unchanged, CLIP-DPO modifies the first two steps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.785, + 0.584 + ], + "angle": 0, + "content": "Firstly, we streamline the generation step by utilizing the same data pool (i.e. images only) as the SFT stage and generate data using either the model itself or a small pool of efficient LVLMs. Secondly, we introduce an annotation step tailored for CLIP-style embedding models, where we rank captions based on their image-text similarities. Then, given that the resulting captions and pairs may vary in quality, we implement a rule-based filtering strategy to remove unsuitable candidates. Finally, with the constructed preference data, the model undergoes DPO-based optimization using Eq. 1, as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.786, + 0.688 + ], + "angle": 0, + "content": "As such, by leveraging VL embedding models for data annotation, we tap into the extensive information captured by these contrastively-pretrained models, which have been exposed to hundreds of millions of unique image-text pairs during training, unlike existing open-sourced LVLMs. Additionally, unlike the discrete GPT-4V ranking used in prior work, CLIP scores are continuous, enabling finer-grained comparisons. This allows us to assess the difficulty of a given pair based on the margin between the score values." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.721 + ], + "angle": 0, + "content": "Next, we will elaborate on each module of our pipeline. Sec. 4.1 details the data generation module, and Sec. 4.2 details the data annotation process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.741, + 0.403, + 0.755 + ], + "angle": 0, + "content": "4.1 Data Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The first step of our pipeline consists of generating a set of per-image captions that will be later ranked by a pre-trained CLIP model, along with a subsequent filtering approach to select a set of positive and negative pairs, which are then used for DPO-based training. To this end, we start by selecting the pool of LVLM annotators and the data to be annotated. For the annotators and to reduce the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.654, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLIP-DPO" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "cost of the generation step, we select MobileVLM-v2 [14] family of models, given their efficiency and performance. As for the data, and to avoid introducing any new sources that might bias the results of CLIP-DPO, and to further reduce the method's cost, we opt for the initial pool of data used during the SFT stage of MobileVLM-v2 [14] models, see Table 1 for details. Then, we conduct our data generation pipeline that consists of two steps, the generation of generic captions and per-image questions and answers." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.266, + 0.473, + 0.409 + ], + "angle": 0, + "content": "
DatasetDataset Size
COCO118K
GQA72K
LLaVA-1.0 Pretraining595K
LLaVA-1.5 Pretraining558K
OCR VQA80K
SAM570K
SBU845K
TextVQA22K
VG86K
Web-Celebrity495
Web-Landmark500
Wikiart500
Total2.9M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.419, + 0.492, + 0.446 + ], + "angle": 0, + "content": "Table 1: Data pool used for data generation." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.258, + 0.75, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.419, + 0.78, + 0.448 + ], + "angle": 0, + "content": "Fig. 4: The distribution of CLIP image-text scores per category." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.475, + 0.339, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.348, + 0.478, + 0.423, + 0.486 + ], + "angle": 0, + "content": "MobileVLMv-2 1.7b:" + }, + { + "type": "text", + "bbox": [ + 0.354, + 0.487, + 0.738, + 0.509 + ], + "angle": 0, + "content": "- The image showcases a charming two-story building, characterized by its steep roof and a balcony adorned with red flowers. The building's facade is painted white, accented with a green roof that adds a touch of color to the scene." + }, + { + "type": "text", + "bbox": [ + 0.354, + 0.51, + 0.78, + 0.52 + ], + "angle": 0, + "content": "The image captures a charming scene of a German half-timbered building, bathed in the soft glow of daylight." + }, + { + "type": "list", + "bbox": [ + 0.354, + 0.487, + 0.78, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.349, + 0.523, + 0.413, + 0.53 + ], + "angle": 0, + "content": "MobileVLMv-2 3b:" + }, + { + "type": "text", + "bbox": [ + 0.354, + 0.532, + 0.746, + 0.555 + ], + "angle": 0, + "content": "- The image showcases a traditional German half-timbered house, a common architectural feature in the region. The house is adorned with a green roof, a characteristic of half-timbered houses, and is embellished with red flowers that add a vibrant touch to the overall design." + }, + { + "type": "text", + "bbox": [ + 0.354, + 0.558, + 0.621, + 0.566 + ], + "angle": 0, + "content": "- A building with a statue of a knight in the middle of the building." + }, + { + "type": "list", + "bbox": [ + 0.354, + 0.532, + 0.746, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.349, + 0.572, + 0.396, + 0.579 + ], + "angle": 0, + "content": "Mistral 7b:" + }, + { + "type": "text", + "bbox": [ + 0.354, + 0.58, + 0.562, + 0.597 + ], + "angle": 0, + "content": "- What color is the roof of the building in the image?\n- Positive: Green Negative: Belge" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.61, + 0.788, + 0.653 + ], + "angle": 0, + "content": "Fig. 5: Examples of generated generic captions produced by MobileVLM-v2 1.7B and MobileVLM-v2 3B. For MobileVLM-v2 3B's generated captions, we also show the produced question and positive and negative answers generated using Mistral 7B." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.797 + ], + "angle": 0, + "content": "Generation of Generic Captions. We start by generating a set of 5 descriptive captions per image. For each of the MobileVLM-v2 models, we prompt with 5 different prompts (e.g. \"identify the setting and note any characters or objects, focusing on visible details.\") to increase the diversity of the generated captions. While these captions can be used for CLIP ranking and DPO-based training, they are still produced by generic prompts that are not image-specific. Next, we perform the second step of our data generation pipeline to obtain a set of per-image questions and answers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Generation of Per-Image Questions and Answers: To obtain a set of questions per image, we leverage an LLM, i.e. a variant of Mistral 7B Instruct" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.128 + ], + "angle": 0, + "content": "Y. Ouali et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.374, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.38, + 0.145, + 0.56, + 0.164 + ], + "angle": 0, + "content": "Question 1: What is the central focus of the image?" + }, + { + "type": "text", + "bbox": [ + 0.384, + 0.165, + 0.514, + 0.174 + ], + "angle": 0, + "content": "- Question CLIP Score: 15.0" + }, + { + "type": "text", + "bbox": [ + 0.384, + 0.175, + 0.514, + 0.183 + ], + "angle": 0, + "content": "- Question Filter: Rejected" + }, + { + "type": "list", + "bbox": [ + 0.384, + 0.165, + 0.514, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.581, + 0.145, + 0.747, + 0.164 + ], + "angle": 0, + "content": "Question 2: What type of market is depicted in the image?" + }, + { + "type": "text", + "bbox": [ + 0.585, + 0.166, + 0.714, + 0.174 + ], + "angle": 0, + "content": "- Question CLIP Score: 26.7" + }, + { + "type": "text", + "bbox": [ + 0.585, + 0.175, + 0.695, + 0.183 + ], + "angle": 0, + "content": "- Question Filter: Kept" + }, + { + "type": "list", + "bbox": [ + 0.585, + 0.166, + 0.714, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.38, + 0.191, + 0.568, + 0.2 + ], + "angle": 0, + "content": "Positive: A fruit and vegetable market." + }, + { + "type": "text", + "bbox": [ + 0.383, + 0.202, + 0.762, + 0.21 + ], + "angle": 0, + "content": "- Caption: type of market is depicted in the image a fruit and vegetable market." + }, + { + "type": "text", + "bbox": [ + 0.384, + 0.211, + 0.512, + 0.219 + ], + "angle": 0, + "content": "Positive CLIP Score: 29.1" + }, + { + "type": "text", + "bbox": [ + 0.384, + 0.22, + 0.494, + 0.228 + ], + "angle": 0, + "content": "- Question Filter: Kept" + }, + { + "type": "list", + "bbox": [ + 0.383, + 0.202, + 0.762, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.24, + 0.787, + 0.31 + ], + "angle": 0, + "content": "Fig. 6: Question Filtering. The sample is rejected if the CLIP image-question score is low for a given sample. If the question is relevant and not generic, it is kept. Next, we evaluate the quality of the answer by first parsing the question into a caption, appending the positive answer to it, and computing its CLIP score. If it is a high score, the sample is kept for DPO-based training." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.325, + 0.365, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.411, + 0.401, + 0.427 + ], + "angle": 0, + "content": "Prompt: Provide a brief description of the given image." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.428, + 0.396, + 0.443 + ], + "angle": 0, + "content": "Positive: An Apple computer with a laptop computer sitting on a desk." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.443, + 0.396, + 0.459 + ], + "angle": 0, + "content": "Negative: A laptop and a desktop computer are sitting on a desk." + }, + { + "type": "image", + "bbox": [ + 0.416, + 0.325, + 0.567, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.415, + 0.411, + 0.599, + 0.428 + ], + "angle": 0, + "content": "Prompt: Provide a depiction of the picture. Positive: A man in orange shirt kicking a" + }, + { + "type": "text", + "bbox": [ + 0.415, + 0.429, + 0.464, + 0.436 + ], + "angle": 0, + "content": "soccer ball" + }, + { + "type": "text", + "bbox": [ + 0.415, + 0.436, + 0.575, + 0.444 + ], + "angle": 0, + "content": "Negative: A group of people kicking a" + }, + { + "type": "text", + "bbox": [ + 0.415, + 0.444, + 0.466, + 0.451 + ], + "angle": 0, + "content": "soccer ball" + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.325, + 0.772, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.611, + 0.411, + 0.772, + 0.427 + ], + "angle": 0, + "content": "Prompt: What is the name of the paint brand featured on the bucket?" + }, + { + "type": "text", + "bbox": [ + 0.612, + 0.428, + 0.768, + 0.436 + ], + "angle": 0, + "content": "Positive: Homer's All Purpose Bucket" + }, + { + "type": "text", + "bbox": [ + 0.612, + 0.436, + 0.783, + 0.443 + ], + "angle": 0, + "content": "Negative: Home Depot All-Purpose Bucket." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.47, + 0.787, + 0.499 + ], + "angle": 0, + "content": "Fig. 7: Examples of the final positive and negative pairs obtained post-filtering, to be used for DPO-based training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.529, + 0.787, + 0.591 + ], + "angle": 0, + "content": "v0.2 [28]\\(^1\\), and feed it with the generated captions and prompt it to generate 2 questions for each image, together with positive and negative answers. The LLM is asked to generate the positive ones based on the fed captions and to generate plausible but incorrect negatives given the image description/caption." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.6, + 0.77, + 0.616 + ], + "angle": 0, + "content": "For an example of the generated captions, questions, and answers, see Fig. 5." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.64, + 0.403, + 0.654 + ], + "angle": 0, + "content": "4.2 Data Annotation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.669, + 0.788, + 0.76 + ], + "angle": 0, + "content": "Starting with the generated captions, the subsequent data annotation step involves ranking them based on their CLIP image-text similarities and filtering them to obtain a final set of high-quality positive-negative pairs for DPO-based training. Our filtering pipeline includes two stages: global filtering, where we eliminate low-quality images and captions, and pair-filtering, where we select the best positive and negative per-image pairs for DPO-based training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.769, + 0.787, + 0.815 + ], + "angle": 0, + "content": "CLIP Ranking. The initial step of data annotation is CLIP ranking. In this step, we use a pre-trained CLIP model to compute the cosine similarities of the captions with their associated images and then sort them from highest to lowest." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.824, + 0.663, + 0.84 + ], + "angle": 0, + "content": "1 https://huggingface.co/teknium/OpenHermes-2.5-Mistral7B" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.654, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLIP-DPO" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.146, + 0.788, + 0.373 + ], + "angle": 0, + "content": "Global Filtering. Next, we aim to filter our data and only retain a balanced and higher-quality subset. To achieve this, we start by analyzing the types of images in our data pool (see Tab. 1). We define a set of four generic categories: images of text, people, objects, and scenes. Using CLIP and a set of 10 descriptions per category, we create four class prototypes. Each image is then assigned to one of these prototypes based on the highest cosine similarity. As illustrated in Fig. 4, we observe that CLIP scores for the text category are the highest, while scores for people, objects, and scenes are relatively similar. Consequently, before applying CLIP-based filtering, we down-sample the portion of text images. We then filter out all generic captions below a given CLIP score (\\(e.g. < 28.0\\)). Additionally, since CLIP is primarily trained on short text, we remove long captions to ensure more precise CLIP scores. For questions, we compute their CLIP scores and remove all questions with low CLIP scores (\\(e.g. < 25.0\\)), thereby eliminating generic questions (\\(e.g. \"what is the main object in the image?\")\\) that are already covered by the generated captions." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.382, + 0.788, + 0.594 + ], + "angle": 0, + "content": "Pair Filtering. Finally, starting from the remaining high-quality image captions and image question-answers, the final step will be selecting a set of positive and negative pairs for DPO-based training. For the questions, since we already have a set of positive and negative responses generated by our LLM, we only need to filter out the low-quality pairs. To achieve this, we use simple regex matching rules to extract an image description from the question, append the positive answer to it to create a synthetic caption, and compute its CLIP image-text scores. We then reject examples where the scores are low. See Fig. 6 for an example of the questions filtering process. As for the captions, we consider all possible pairs where the CLIP score difference between two captions is larger than a given threshold (\\(e.g. > 2.0\\)) and where the length of the two captions is similar to avoid introducing false preferences. We then order them based on the CLIP score difference between the positive and negative captions and select the top-ranking pair per image." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.603, + 0.787, + 0.648 + ], + "angle": 0, + "content": "After the data annotation step, the resulting DPO training data consists of 750K pairs, of which 50K are question-answer pairs and the rest are caption pairs. For examples of the final pairs, see Fig. 7." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.672, + 0.474, + 0.689 + ], + "angle": 0, + "content": "5 Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.704, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Network architecture: In this work, we consider two LVLMs architectures: MobileVLM-v2 [14] and LLaVA-1.5 [42]. Both models follow the same overall structure: a pre-trained CLIP vision encoder and a pre-trained LLM. The visual tokens produced by the frozen vision encoder are projected using either a linear layer or a small projection module and passed as input to the LLM. LLaVA opts for a pre-trained Vicuna LLM [12] while MobileVLM-v2 uses MobileLLaMA [13], except for their 7B variant, which also uses a Vicuna model. Both use the same ViT-L-14 @ 336px CLIP visual encoder [56]. For efficiency purposes, MobileVLM-v2 uses a projection module that halves the number of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "Y. Ouali et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.193 + ], + "angle": 0, + "content": "visual tokens. We consider the following model variants in our comparisons: MobileVLM-v2 (1.7B, 3B and 7B), and LLaVA-1.5 7B. For data annotation, we use ViT-H/14 DFN [19] as our pre-trained CLIP model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.201, + 0.784, + 0.352 + ], + "angle": 0, + "content": "Training details: For all of our experiments, unless specified otherwise, we start from the pre-trained MobileVLM-v2 and LLaVA-1.5 models. The models are then fine-tuned for 1 epoch using DPO-based optimization on the constructed CLIP-DPO dataset. We use the following hyperparameters for fine-tuning the models: AdamW optimizer [45] with a batch size of 256, a learning rate of \\(5e - 7\\), decreased to 0 using a cosine scheduler, a warm-up of 0.01, and a weight decay set to 0. Both during training and testing, the input images are cropped and resized to \\(336 \\times 336\\)px. The training was performed on 8 A100 GPUs using Pytorch [54]. For the larger models, e.g. LLaVA-1.5 7B, to fit them in memory, we use the Zero-3 strategy [58, 59] and decrease the batch size to 64." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.371, + 0.327, + 0.387 + ], + "angle": 0, + "content": "6 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.784, + 0.489 + ], + "angle": 0, + "content": "We first evaluate the effectiveness of CLIP-DPO in reducing hallucinations on the recently introduced AMBER [71] benchmark in Sec. 6.1. Moreover, we demonstrate its enhanced grounding capabilities by reporting zero-shot image classification in Sec. 6.2. Finally, in Sec. 6.4, we show that the proposed CLIP-DPO training does not compromise LVLM performance by reporting results on the LLaVA benchmark [87]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.496, + 0.498, + 0.51 + ], + "angle": 0, + "content": "6.1 Evaluation on hallucinations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.785, + 0.772 + ], + "angle": 0, + "content": "We start by evaluating CLIP-DPO in terms of LVLM hallucination reduction, which is our main objective. We use AMBER [71], a comprehensive, high-quality, and LLM-free multidimensional benchmark for LVLM hallucination evaluation, which can be used to evaluate both generative and discriminative tasks. As the results from Tab. 2 show, MobileVLM-v2 trained with CLIP-DPO improves upon MobileVLM-v2 baselines across all model sizes, and sometimes quite significantly, especially for the 1.7B and 7B models. It can also be seen that CLIP-DPO significantly improves when applied on top of LLaVA-1.5 7B. Importantly, CLIP-DPO significantly outperforms HA-DPO [83], the main competing approach, improving the AMBER score 3.2 vs 7.8 when using LLaVA-1.5. Finally, our LlaVA-1.5 7B+CLIP-DPO even outperforms Qwen-VL [5], which is trained on significantly larger datasets (1.4B image-text pairs for pre-training and 77M for multitask training while CLIP-DPO is fine-tuned on just 0.7M samples), and even matches the performance of GPT-4V without using any GPT-4V model outputs during training. Overall, these results on a high-quality state-of-the-art benchmark such as AMBER clearly demonstrate the effectiveness of our CLIP-DPO approach for reducing hallucinations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.79, + 0.506, + 0.805 + ], + "angle": 0, + "content": "6.2 Zero-shot image classification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "A primary reason for the hallucinatory behavior of LVLMs is the weak alignment between their visual features and the input LLM tokens. A direct way to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.654, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLIP-DPO" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.249, + 0.144, + 0.758, + 0.357 + ], + "angle": 0, + "content": "
ModelGENERATIVE TASKDISCRIMINATIVE TASKAMBER
CHAIR↓Cover↑Hal↓Cog↓Acc.P.R.F1
mPLUG-Owl21.650.176.111.540.192.810.518.948.7
LLaVA11.551.048.85.542.774.121.032.760.6
MiniGPT-413.663.065.311.363.690.550.464.775.6
CogVLM5.657.223.61.369.088.960.972.383.4
mPLUG-Owl210.652.039.94.575.695.066.978.584.0
InstructBLIP8.852.238.24.476.584.579.081.786.5
Qwen-VL5.549.423.61.981.290.879.784.989.7
GPT-4V4.667.130.72.683.484.990.187.491.4
LLaVA-1.5 7B7.851.036.44.272.093.262.474.783.5
+ HA-DPO7.233.619.72.668.368.198.480.586.7
+ CLIP-DPO3.747.816.61.377.884.481.582.989.6
MobileVLM-v2 1.7B3.839.68.90.565.492.651.966.581.4
+ CLIP-DPO4.238.910.80.571.288.764.874.985.3
MobileVLM-v2 3B4.838.711.10.773.592.165.776.786.0
+ CLIP-DPO4.741.513.10.576.791.071.880.387.8
MobileVLM-v2 7B4.438.910.40.671.995.060.874.184.9
+ CLIP-DPO4.038.010.10.577.393.470.880.588.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.368, + 0.785, + 0.424 + ], + "angle": 0, + "content": "Table 2: Hallucination evaluation results on AMBER for both generative (leftmost set of columns) and discriminative tasks (rightmost set of columns). Our approach offers consistent and large performance improvements across 4 different LVLM models, and also beats HA-DPO by a large margin." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.44, + 0.784, + 0.558 + ], + "angle": 0, + "content": "
MethodStanfordCarsOxfordPetsOxfordFlowersImagenetFood-101EurosatCaltech-101UCF-101SUN397Avg
LLaVA-1.523.234.07.337.945.349.582.750.543.140.0
+ HA-DPO23.233.47.136.945.149.482.750.342.739.7
+ CLIP-DPO30.144.816.342.253.952.184.853.148.847.4
MobileVLM-v2 1.7B17.115.214.532.631.249.377.945.539.333.9
+ CLIP-DPO19.232.623.841.347.448.980.850.249.343.7
MobileVLM-v2 3B14.823.59.135.738.853.281.848.642.036.7
+ CLIP-DPO28.640.919.544.352.056.485.552.350.147.7
MobileVLM-v2 7B27.132.511.537.745.043.181.949.046.641.6
+ CLIP-DPO32.551.335.850.162.659.388.356.254.854.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.61 + ], + "angle": 0, + "content": "Table 3: Zero-shot image recognition results in terms of Top-1 accuracy (\\%). While HA-DPO shows very similar performance to the base model, our CLIP-DPO improves the base model by a very large margin in all cases." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.644, + 0.787, + 0.841 + ], + "angle": 0, + "content": "evaluate this is through simple zero-shot image classification, which is the goto benchmark for contrastively trained VL models like CLIP [56]. The typical setup follows a closed-set classification problem, where the names of all possible classes are known a priori and are encoded into class prototypes using CLIP. A given image is then assigned to the class with the highest CLIP image-text scores. Herein, we follow the same protocol with the notable difference that, as our goal is LVLM evaluation, we first prompt the model to generate a free-form image caption describing the main object in the image, then encode the caption using the CLIP text encoder and assign the image to the class with the highest text-text (i.e. caption-class) CLIP score. Here, we opted for a different family of VL embedding model, i.e. SigLIP [79], to avoid evaluating using CLIP models similar to those used during the data annotation step. Following [56,85], we evaluate our approach on a suite of 9 diverse datasets: UCF-101 [63], SUN397 [74]," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "Y. Ouali et al." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.144, + 0.77, + 0.422 + ], + "angle": 0, + "content": "
MethodLLMRes.GQA\\( SQAI \\)\\( VQAT \\)POPE\\( MME^P \\)\\( MMB^{dev} \\)Avg.
BLIP-2 [33]Vicuna-13B22441.061.042.585.31293.8--
InstructBLIP [16]Vicuna-13B22449.563.150.778.91212.8--
Shikra [9]Vicuna-13B224-----58.8-
Openflamingo [3]MPT 7B336--33.6--4.6-
Qwen-VL [5]Qwen 7B44859.367.163.8-1487.638.2-
mPLUG-Owl [76]LLaMA 7B224----967.349.4-
MiniGPT-v2 [8]LLaMA 7B44860.3----12.2-
MiniGPT-4 [86]Vicuna 7B22432.2---581.723.0-
InstructBLIP [16]Vicuna 7B22449.260.550.1--36.0-
ShareGPT4V [10]Vicuna 7B33663.368.460.485.71567.468.870.8
MoE-LLaVA-1.6B×4 [39]StableLM-1.6B33660.462.647.884.31300.8*59.463.3
MoE-LLaVA-2.7B×4 [39]Phi-2.7B33661.168.750.285.01396.4*65.566.7
LLaVA-1.5 [41]Vicuna 7B33662.066.858.285.91510.764.368.8
+ HA-DPOVicuna 7B33661.969.258.384.31505.664.969.0
+ CLIP-DPOVicuna 7B33659.367.656.485.81468.764.967.9
MobileVLM 1.7B [13]MobileLLaMA 1.4B33656.157.341.584.51196.253.258.7
MobileVLM-v2 1.7B [14]MobileLLaMA 1.4B33658.366.752.184.31302.857.764.2
+ CLIP-DPOMobileLLaMA 1.4B33658.668.447.884.31331.157.663.8
MobileVLM- 3B [13]MobileLLaMA 2.7B33659.061.247.584.91288.959.662.8
MobileVLM-v2 3B [14]MobileLLaMA 2.7B33661.170.057.584.71440.563.268.1
+ CLIP-DPOMobileLLaMA 2.7B33660.972.357.385.11425.763.868.4
MobileVLM-v2 [14] 7BVicuna 7B33662.674.862.385.31560.769.272.1
+ CLIP-DPO 7BVicuna 7B33662.577.362.284.71539.170.472.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.433, + 0.788, + 0.474 + ], + "angle": 0, + "content": "Table 4: Results on several vision language tasks from the LlaVA benchmark. Our approach is combined with LVLMs of different sizes, showing consistent performance in all cases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.788, + 0.592 + ], + "angle": 0, + "content": "Stanford Cars [32], Oxford Pets [53], Oxford flowers [49], ImageNet [17], Food 101 [6], Eurosat [23] and Caltech-101 [20]. As the results from Tab. 3 show, all LVLMs fine-tuned with CLIP-DPO significantly outperform their corresponding baselines, showcasing the increased discriminative properties. Note again that these improvements are obtained without affecting the model's performance on other tasks and datasets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.611, + 0.447, + 0.625 + ], + "angle": 0, + "content": "6.3 Additional evaluations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.77 + ], + "angle": 0, + "content": "Herein, we evaluate the impact of CLIP-DPO training on other vision language tasks and, more specifically, on the popular LLaVA-Bench (GQA [26], ScienceQA [46], TextVQA [62], MME [21], MMBench [44]). As Tab. 4 shows, overall, CLIP-DPO training does not compromise performance. We further note that accuracy improvements on LLaVA-Bench are heavily tied to the addition of extra training data [10] or architectural changes [13,14,39,41]. As we are not using any additional data & models nor making architectural changes, it is not surprising that the performance after CLIP-DPO training remains largely in line with that of the original baseline model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.421, + 0.803 + ], + "angle": 0, + "content": "6.4 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In Figure 8, we show some qualitative examples comparing the predictions of LLaVA-1.5 7B, LLaVA-1.5 7B+HA-DPO, and LLaVA-1.5 7B+CLIP-DPO. Over" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.654, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLIP-DPO" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.144, + 0.361, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.15, + 0.56, + 0.16 + ], + "angle": 0, + "content": "Provide a short description of the image." + }, + { + "type": "text", + "bbox": [ + 0.372, + 0.163, + 0.747, + 0.189 + ], + "angle": 0, + "content": "LLava-1.5-7b: A large silver airplane is parked inside a hangar. The plane is positioned on a runway, and there is a ladder leading up to the cockpit. The hangar appears to be a large building, possibly a factory or a maintenance facility." + }, + { + "type": "text", + "bbox": [ + 0.38, + 0.191, + 0.652, + 0.201 + ], + "angle": 0, + "content": "+ HA-DPO: A large silver airplane is parked inside a hangar." + }, + { + "type": "text", + "bbox": [ + 0.379, + 0.204, + 0.733, + 0.222 + ], + "angle": 0, + "content": "+ CLIP-DPO: A fighter jet is parked inside a hangar, with a ladder attached to it. The jet is numbered 157." + }, + { + "type": "list", + "bbox": [ + 0.379, + 0.191, + 0.733, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.234, + 0.361, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.239, + 0.559, + 0.249 + ], + "angle": 0, + "content": "Provide a short description of the image." + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.251, + 0.721, + 0.261 + ], + "angle": 0, + "content": "LLava-1.5-7b: A green bus is parked in a parking lot, waiting for passengers." + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.264, + 0.702, + 0.274 + ], + "angle": 0, + "content": "+ HA-DPO: A green bus is parked in a parking lot, waiting for passengers." + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.276, + 0.607, + 0.286 + ], + "angle": 0, + "content": "+CLIP-DPO: A green Go3 bus is parked at a bus stop." + }, + { + "type": "list", + "bbox": [ + 0.373, + 0.251, + 0.721, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.233, + 0.303, + 0.768, + 0.318 + ], + "angle": 0, + "content": "Fig. 8: Qualitative examples comparing our method with LLaVA and HA-DPO." + }, + { + "type": "table", + "bbox": [ + 0.301, + 0.332, + 0.703, + 0.435 + ], + "angle": 0, + "content": "
Loss variantAMBER ScoreCls. Avg.LLaVA bench.
DPO (β = 0.1) [57]85.343.763.8
DPO (β = 0.3) [57]83.342.164.0
IPO [4]83.640.564.5
KTO [18]83.539.564.7
SLIC [43]83.242.764.0
cDPO (β = 0.1, LS = 0.1 [43])83.143.263.9
cDPO (β = 0.3, LS = 0.1 [43])83.140.764.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.445, + 0.788, + 0.487 + ], + "angle": 0, + "content": "Table 5: Effect of DPO loss variant on CLIP-DPO training. Results are reported in terms of AMBER Score, Average Top-1 (\\%) accuracy across all 9 image classification datasets, and the average on LLaVA benchmark. MobileVLM-v2 1.7B was used." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.517, + 0.785, + 0.548 + ], + "angle": 0, + "content": "all, with CLIP-DPO, the model is more grounded in the visual content, less prone to hallucinations, and more precise and fine-grained in its descriptions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.572, + 0.411, + 0.587 + ], + "angle": 0, + "content": "7 Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.605, + 0.785, + 0.634 + ], + "angle": 0, + "content": "Herein, we ablate the effect of the proposed components. In all experiments, we use a MobileVLM-v2 1.7B model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.659, + 0.479, + 0.673 + ], + "angle": 0, + "content": "7.1 Effect of DPO loss variant" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.685, + 0.785, + 0.746 + ], + "angle": 0, + "content": "Our work primarily follows the original DPO formulation proposed in [57]. Here we also consider the following recently proposed variations: KTO [18], ITO [4], SLIC [43] and cDPO [48]. As Tab. 5 shows, aggregated, all losses tend to perform similarly, with DPO marginally outperforming the others." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.769, + 0.439, + 0.784 + ], + "angle": 0, + "content": "7.2 Effect of CLIP scorer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Herein, we seek to explore alternatives to ViT-H/14 DFN [19] CLIP model used in previous experiments, analyzing the impact of the scorer used on the overall performance of the model. We consider a diverse set of alternatives, covering" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "Y. Ouali et al." + }, + { + "type": "table", + "bbox": [ + 0.331, + 0.144, + 0.673, + 0.214 + ], + "angle": 0, + "content": "
ScorerAMBER ScoreCls. Avg.LLaVA bench.
ViT-L/14 [56]79.842.363.0
SigLIP-L/16 [79]85.043.963.2
ViT-H/14 [60]84.344.263.7
ViT-H/14 DFN [19]85.343.763.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.224, + 0.787, + 0.266 + ], + "angle": 0, + "content": "Table 6: Effect of CLIP scorer on CLIP-DPO training. Results are reported in terms of AMBER Score, Average Top-1 (\\%) accuracy across all 9 image classification datasets, and the average on LLaVA benchmark. MobileVLM-v2 1.7B was used." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.295, + 0.787, + 0.415 + ], + "angle": 0, + "content": "multiple exploratory paths: equally-sized models trained on different data (ViT-H/14 [19] trained on LAION-2B instead of DFN-5B); smaller models (ViT-L/14 [56]) and models trained using different pre-training losses (SigLIP-ViT-L/16 [79]). As the results from Tab. 6 show, our approach is generally robust to the exact scorer used, with the notable exception of the ViT-L/14 [56] model. Notice that this difference is primarily manifesting on the AMBER benchmark. Intuitively, this showcases the importance of a powerful scorer for reducing the amount of hallucinations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.439, + 0.527, + 0.456 + ], + "angle": 0, + "content": "Limitations and broader impact" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.787, + 0.62 + ], + "angle": 0, + "content": "As an LVLM-based approach, our method is subject to the same general consideration (i.e. potential data bias, susceptibility to hallucinations, etc.). Moreover, as the LVLMs are trained on relatively small datasets compared to LLMs or CLIP, gaps within their knowledge domains are possible. This is especially important as neural networks tend to be overconfident outside their seen input distribution. As with all models from this category, we strongly recommend checking the models and the data carefully before deploying them. Despite these general aspects, our approach is shown to significantly reduce the amount of hallucinated content and improve the model's discriminability, hence resulting in more robust and reliable models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.643, + 0.37, + 0.659 + ], + "angle": 0, + "content": "8 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In this work, we proposed CLIP-DPO, a simple method that reduces hallucinations in LVLMs based on using a pre-trained CLIP model [56] to rank the LVLM's self-generated captions in order to construct positive-negative pairs for DPO fine-tuning. In contrast to previous works, our proposed CLIP-DPO removes the need of (i) paid-for APIs, (ii) additional data and (iii) additional external LVLMs. Before training, the data is filtered using a newly proposed robust rule-based approach. When applied on top of established LVLMs, it is shown that CLIP-DPO fine-tuning significantly reduces hallucinations, outperforming the baseline models by significant margins and even matching the performance of GPT-4V on the AMBER benchmark. We also observe superior zero-shot object recognition, suggesting improved object grounding capabilities." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.654, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLIP-DPO" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.2 + ], + "angle": 0, + "content": "1. Detailed caption dataset. https://huggingface.co/datasets/echo840/Detailed_Caption (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.201, + 0.785, + 0.241 + ], + "angle": 0, + "content": "2. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.242, + 0.785, + 0.295 + ], + "angle": 0, + "content": "3. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., et al.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.785, + 0.337 + ], + "angle": 0, + "content": "4. Azar, M.G., Rowland, M., Piot, B., Guo, D., Calandriello, D., Valko, M., Munos, R.: A general theoretical paradigm to understand learning from human preferences. arXiv preprint arXiv:2310.12036 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.337, + 0.785, + 0.377 + ], + "angle": 0, + "content": "5. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-VL: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.404 + ], + "angle": 0, + "content": "6. Bossard, L., Guillaumin, M., Van Gool, L.: Food-101-mining discriminative components with random forests. In: European Conference on Computer Vision (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.785, + 0.445 + ], + "angle": 0, + "content": "7. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances on Neural Information Processing Systems (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.445, + 0.785, + 0.499 + ], + "angle": 0, + "content": "8. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.5, + 0.785, + 0.54 + ], + "angle": 0, + "content": "9. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.541, + 0.785, + 0.581 + ], + "angle": 0, + "content": "10. Chen, L., Li, J., Dong, X., Zhang, P., He, C., Wang, J., Zhao, F., Lin, D.: Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.785, + 0.622 + ], + "angle": 0, + "content": "1. Chen, Z., Wu, J., Wang, W., Su, W., Chen, G., Xing, S., Muyan, Z., Zhang, Q., Zhu, X., Lu, L., et al.: Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.622, + 0.785, + 0.677 + ], + "angle": 0, + "content": "2. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.677, + 0.785, + 0.717 + ], + "angle": 0, + "content": "3. Chu, X., Qiao, L., Lin, X., Xu, S., Yang, Y., Hu, Y., Wei, F., Zhang, X., Zhang, B., Wei, X., et al.: MobileVLM: A fast, reproducible and strong vision language assistant for mobile devices. arXiv preprint arXiv:2312.16886 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.758 + ], + "angle": 0, + "content": "4. Chu, X., Qiao, L., Zhang, X., Xu, S., Wei, F., Yang, Y., Sun, X., Hu, Y., Lin, X., Zhang, B., et al.: MobileVLM V2: Faster and stronger baseline for vision language model. arXiv preprint arXiv:2402.03766 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "5. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: InstructBLIP: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "Y. Ouali et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "17. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: IEEE Conference on Computer Vision and Pattern Recognition (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "18. Ethayarajh, K., Xu, W., Muennighoff, N., Jurafsky, D., Kiela, D.: Kto: Model alignment as prospect theoretic optimization. arXiv preprint arXiv:2402.01306 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.245 + ], + "angle": 0, + "content": "19. Fang, A., Jose, A.M., Jain, A., Schmidt, L., Toshev, A., Shankar, V.: Data filtering networks. arXiv preprint arXiv:2309.17425 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.3 + ], + "angle": 0, + "content": "20. Fei-Fei, L., Fergus, R., Perona, P.: Learning generative visual models from few training examples: An incremental Bayesian approach tested on 101 object categories. In: IEEE Conference on Computer Vision and Pattern Recognition - Workshops (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "21. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.37 + ], + "angle": 0, + "content": "22. Gunjal, A., Yin, J., Bas, E.: Detecting and preventing hallucinations in large vision language models. arXiv preprint arXiv:2308.06394 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.424 + ], + "angle": 0, + "content": "23. Helber, P., Bischke, B., Dengel, A., Borth, D.: EuroSAT: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 12(7), 2217-2226 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.467 + ], + "angle": 0, + "content": "24. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.495 + ], + "angle": 0, + "content": "25. Hu, H., Zhang, J., Zhao, M., Sun, Z.: Ciem: Contrastive instruction evaluation method for better instruction tuning. arXiv preprint arXiv:2309.02301 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.536 + ], + "angle": 0, + "content": "26. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: IEEE Conference on Computer Vision and Pattern Recognition (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.563 + ], + "angle": 0, + "content": "27. Jain, J., Yang, J., Shi, H.: Vcoder: Versatile vision encoders for multimodal large language models. arXiv preprint arXiv:2312.14233 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.605 + ], + "angle": 0, + "content": "28. Jiang, A.Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D.S., Casas, D.d.l., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., et al.: Mistral 7b. arXiv preprint arXiv:2310.06825 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.647 + ], + "angle": 0, + "content": "29. Jiang, C., Xu, H., Dong, M., Chen, J., Ye, W., Yan, M., Ye, Q., Zhang, J., Huang, F., Zhang, S.: Hallucination augmented contrastive learning for multimodal large language model. arXiv preprint arXiv:2312.06968 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "30. Jing, L., Li, R., Chen, Y., Jia, M., Du, X.: Faithscore: Evaluating hallucinations in large vision-language models. arXiv preprint arXiv:2311.01477 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "31. Koh, J.Y., Salakhutdinov, R., Fried, D.: Grounding language models to images for multimodal inputs and outputs. International Conference on Machine Learning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "32. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3D object representations for fine-grained categorization. In: IEEE International Conference on Computer Vision - Workshops (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "33. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "34. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Learning Representations (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.654, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLIP-DPO" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "35. Li, L., Xie, Z., Li, M., Chen, S., Wang, P., Chen, L., Yang, Y., Wang, B., Kong, L.: Silkie: Preference distillation for large visual language models. arXiv preprint arXiv:2312.10665 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "36. Li, S., Lin, R., Pei, S.: Multi-modal preference alignment remedies regression of visual instruction tuning on language model. arXiv preprint arXiv:2402.10884 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.786, + 0.245 + ], + "angle": 0, + "content": "37. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.245, + 0.787, + 0.273 + ], + "angle": 0, + "content": "38. Li, Y., Bubeck, S., Eldan, R., Del Giorno, A., Gunasekar, S., Lee, Y.T.: Textbooks are all you need ii: phi-1.5 technical report. arXiv preprint arXiv:2309.05463 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.273, + 0.787, + 0.313 + ], + "angle": 0, + "content": "39. Lin, B., Tang, Z., Ye, Y., Cui, J., Zhu, B., Jin, P., Zhang, J., Ning, M., Yuan, L.: Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.314, + 0.787, + 0.357 + ], + "angle": 0, + "content": "40. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Mitigating hallucination in large multi-modal models via robust instruction tuning. In: International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.357, + 0.787, + 0.384 + ], + "angle": 0, + "content": "41. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.384, + 0.787, + 0.412 + ], + "angle": 0, + "content": "42. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances on Neural Information Processing Systems (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.412, + 0.787, + 0.453 + ], + "angle": 0, + "content": "43. Liu, T., Zhao, Y., Joshi, R., Khalman, M., Saleh, M., Liu, P.J., Liu, J.: Statistical rejection sampling improves preference optimization. arXiv preprint arXiv:2309.06657 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.453, + 0.787, + 0.495 + ], + "angle": 0, + "content": "44. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: MMBench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.495, + 0.787, + 0.522 + ], + "angle": 0, + "content": "45. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.522, + 0.787, + 0.564 + ], + "angle": 0, + "content": "46. Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.564, + 0.787, + 0.605 + ], + "angle": 0, + "content": "47. Mishra, S., Khashabi, D., Baral, C., Hajishirzi, H.: Cross-task generalization via natural language crowdsourcing instructions. arXiv preprint arXiv:2104.08773 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.606, + 0.787, + 0.619 + ], + "angle": 0, + "content": "48. Mitchell, E.: A note on dpo with noisy preferences & relationship to ipo (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.62, + 0.787, + 0.66 + ], + "angle": 0, + "content": "49. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: Indian conference on computer vision, graphics & image processing. pp. 722-729 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.661, + 0.487, + 0.674 + ], + "angle": 0, + "content": "50. OpenAI: Introducing chatgpt (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.674, + 0.528, + 0.688 + ], + "angle": 0, + "content": "51. OpenAI: Gpt-4v(ision) system card (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.688, + 0.787, + 0.743 + ], + "angle": 0, + "content": "52. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.744, + 0.787, + 0.771 + ], + "angle": 0, + "content": "53. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: IEEE Conference on Computer Vision and Pattern Recognition (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.771, + 0.787, + 0.799 + ], + "angle": 0, + "content": "54. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "55. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "Y. Ouali et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.203 + ], + "angle": 0, + "content": "56. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.785, + 0.245 + ], + "angle": 0, + "content": "57. Rafailov, R., Sharma, A., Mitchell, E., Manning, C.D., Ermon, S., Finn, C.: Direct preference optimization: Your language model is secretly a reward model. Advances on Neural Information Processing Systems (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.3 + ], + "angle": 0, + "content": "58. Rajbhandari, S., Ruwase, O., Rasley, J., Smith, S., He, Y.: Zero-infinity: Breaking thegpu memory wall for extreme scale deep learning. In: Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.357 + ], + "angle": 0, + "content": "59. Rasley, J., Rajbhandari, S., Ruwase, O., He, Y.: Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In: Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.412 + ], + "angle": 0, + "content": "60. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.439 + ], + "angle": 0, + "content": "61. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "62. Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards VQA models that can read. In: IEEE Conference on Computer Vision and Pattern Recognition (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.509 + ], + "angle": 0, + "content": "63. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "64. Stiennon, N., Ouyang, L., Wu, J., Ziegler, D., Lowe, R., Voss, C., Radford, A., Amodei, D., Christiano, P.F.: Learning to summarize with human feedback. Advances in Neural Information Processing Systems (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.592 + ], + "angle": 0, + "content": "65. Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L.Y., Wang, Y.X., Yang, Y., et al.: Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.633 + ], + "angle": 0, + "content": "66. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.675 + ], + "angle": 0, + "content": "67. Thoppilan, R., De Freitas, D., Hall, J., Shazeer, N., Kulshreshtha, A., Cheng, H.T., Jin, A., Bos, T., Baker, L., Du, Y., et al.: Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.716 + ], + "angle": 0, + "content": "68. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.758 + ], + "angle": 0, + "content": "69. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "70. Wang, J., Yang, Z., Hu, X., Li, L., Lin, K., Gan, Z., Liu, Z., Liu, C., Wang, L.: Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "71. Wang, J., Wang, Y., Xu, G., Zhang, J., Gu, Y., Jia, H., Yan, M., Zhang, J., Sang, J.: An LLM-free multi-dimensional benchmark for MLLMs hallucination evaluation. arXiv preprint arXiv:2311.07397 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.654, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLIP-DPO" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "72. Wang, J., Zhou, Y., Xu, G., Shi, P., Zhao, C., Xu, H., Ye, Q., Yan, M., Zhang, J., Zhu, J., et al.: Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "73. Williams, R.J.: Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning 8, 229-256 (1992)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "74. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: Sun database: Large-scale scene recognition from abbey to zoo. In: IEEE Conference on Computer Vision and Pattern Recognition (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.3 + ], + "angle": 0, + "content": "75. Xu, J., Lee, A., Sukhbaatar, S., Weston, J.: Some things are more cringe than others: Preference optimization with the pairwise cringe loss. arXiv preprint arXiv:2312.16682 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "76. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.384 + ], + "angle": 0, + "content": "77. You, H., Zhang, H., Gan, Z., Du, X., Zhang, B., Wang, Z., Cao, L., Chang, S.F., Yang, Y.: Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.438 + ], + "angle": 0, + "content": "78. Zhai, B., Yang, S., Zhao, X., Xu, C., Shen, S., Zhao, D., Keutzer, K., Li, M., Yan, T., Fan, X.: Halle-switch: Rethinking and controlling object existence hallucinations in large vision language models for detailed caption. arXiv preprint arXiv:2310.01779 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.467 + ], + "angle": 0, + "content": "79. Zhai, X., Mustafa, B., Kolesnikov, A., Beyer, L.: Sigmoid loss for language image pre-training. arXiv preprint arXiv:2303.15343 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "80. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.785, + 0.55 + ], + "angle": 0, + "content": "81. Zhao, H., Cai, Z., Si, S., Ma, X., An, K., Chen, L., Liu, Z., Wang, S., Han, W., Chang, B.: Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "82. Zhao, Y., Joshi, R., Liu, T., Khalman, M., Saleh, M., Liu, P.J.: Slic-hf: Sequence likelihood calibration with human feedback. arXiv preprint arXiv:2305.10425 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.632 + ], + "angle": 0, + "content": "83. Zhao, Z., Wang, B., Ouyang, L., Dong, X., Wang, J., He, C.: Beyond hallucinations: Enhancing LVLMs through hallucination-aware direct preference optimization. arXiv preprint arXiv:2311.16839 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.632, + 0.785, + 0.674 + ], + "angle": 0, + "content": "84. Zhou, C., Liu, P., Xu, P., Iyer, S., Sun, J., Mao, Y., Ma, X., Efrat, A., Yu, P., Yu, L., et al.: Lima: Less is more for alignment. Advances in Neural Information Processing Systems (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.785, + 0.715 + ], + "angle": 0, + "content": "85. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: IEEE Conference on Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.715, + 0.785, + 0.757 + ], + "angle": 0, + "content": "86. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: MiniGPT-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.785, + 0.785 + ], + "angle": 0, + "content": "87. Zhu, Y., Zhu, M., Liu, N., Ou, Z., Mou, X., Tang, J.: LLaVA-phi: Efficient multi-modal assistant with small language model. arXiv preprint arXiv:2401.02330 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.827 + ], + "angle": 0, + "content": "88. Ziegler, D.M., Stiannon, N., Wu, J., Brown, T.B., Radford, A., Amodei, D., Christiano, P., Irving, G.: Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.827 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CLIP-DPO_ Vision-Language Models as a Source of Preference for Fixing Hallucinations in LVLMs/174c66bb-6859-4a7e-835b-0b1709492a92_origin.pdf b/2024/CLIP-DPO_ Vision-Language Models as a Source of Preference for Fixing Hallucinations in LVLMs/174c66bb-6859-4a7e-835b-0b1709492a92_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5e3760ccfc3bf112b85e0a75086aab753b1b74f7 --- /dev/null +++ b/2024/CLIP-DPO_ Vision-Language Models as a Source of Preference for Fixing Hallucinations in LVLMs/174c66bb-6859-4a7e-835b-0b1709492a92_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9cccb014719d8195fa1d27e2919ea29a38b359cb668e6d8692454918389eeacd +size 3174938 diff --git a/2024/CLIP-DPO_ Vision-Language Models as a Source of Preference for Fixing Hallucinations in LVLMs/full.md b/2024/CLIP-DPO_ Vision-Language Models as a Source of Preference for Fixing Hallucinations in LVLMs/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2f5e84a9f97cb7328610acb39f904cc2c45fda38 --- /dev/null +++ b/2024/CLIP-DPO_ Vision-Language Models as a Source of Preference for Fixing Hallucinations in LVLMs/full.md @@ -0,0 +1,369 @@ +# CLIP-DPO: Vision-Language Models as a Source of Preference for Fixing Hallucinations in LVLMs + +Yassine Ouali1 Adrian Bulat1,2 Brais Martinez1 Georgios Tzimiropoulos1,3 + +$^{1}$ Samsung AI Center Cambridge, UK $^{2}$ Technical University of Iasi, Romania $^{3}$ Queen Mary University of London, UK + +Abstract. Despite recent successes, LVLMs or Large Vision Language Models are prone to hallucinating details like objects and their properties or relations, limiting their real-world deployment. To address this and improve their robustness, we present CLIP-DPO, a preference optimization method that leverages contrastively pre-trained Vision-Language (VL) embedding models, such as CLIP, for DPO-based optimization of LVLMs. Unlike prior works tackling LVLM hallucinations, our method does not rely on paid-for APIs, and does not require additional training data or the deployment of other external LVLMs. Instead, starting from the initial pool of supervised fine-tuning data, we generate a diverse set of predictions, which are ranked based on their CLIP image-text similarities, and then filtered using a robust rule-based approach to obtain a set of positive and negative pairs for DPO-based training. We applied CLIP-DPO fine-tuning to the MobileVLM-v2 family of models and to LlaVA-1.5, in all cases observing significant improvements in terms of hallucination reduction over baseline models. We also observe better performance for zero-shot classification, suggesting improved grounding capabilities, and verify that the original performance on standard LVLM benchmarks is overall preserved. + +# 1 Introduction + +A Large Vision Language Model (LVLM) is a Large Language Model (LLM) combined with an independently-trained vision encoder, typically taken from VL embedding models like CLIP [56] and often kept frozen. LVLMs have recently become the state-of-the-art for vision language understanding [5, 16, 42, 65, 81]. Unlike the prior generation of models [34, 70], which were typically tuned for one or a small number of tasks, LVLMs allow free-form dialog in natural language with an image in the loop. This direction largely draws inspiration from the recent success of ChatGPT [50] and respectively, ChatGPT-4V [51], which adapts LLMs to follow human instruction and preferences via various forms of reinforcement learning from human feedback (RFLH) and/or supervised fine-tuning (SFT) on high-quality multi-turn instruction (i.e. chat) data. In a similar fashion, the current generation of LVLMs is trained as part of a two-step process [42]. First, the CLIP vision encoder is aligned with the LLM by training a few adaptation layers only and, then, by fine-tuning the model using SFT on a multi-modal + +instructional dataset. Despite their remarkable success, a major limitation hindering the wider adoption of LVLMs is the high rate of hallucinated details (e.g. non-existing objects) that these models exhibit in their generated output text. + +Motivation. Following the footsteps of LLM training, a natural solution to combat hallucinations is given by RFLH. As reinforcement learning-based training is generally expensive, often requiring the training of a policy model, Rafailov et al. [57] proposed a simplified framework, coined Direct Preference optimization (DPO), that allows direct training of the model using a binary cross-entropy objective. The data for DPO is typically obtained either from human preferences or using LLMs to construct preference data automatically. Thanks to its efficiency, DPO has been quickly adopted for LVLMs too [35, 36, 83]. However, these approaches require multiple LVLMs for generation, collecting additional data, and using GPT-4/GPT-4V API for ranking. Moreover, the GPT-4 provided scores are discrete (hence, of reduced granularity) and are themselves prone to hallucinations, which can perpetuate and exacerbate the already high rate of factual errors and poor image grounding [37, 71]. Finally, GPT-4/V is behind a paywall, and using it for generation/ranking is neither scalable nor cost-efficient. + +Main idea. To fix hallucinations and address the high data construction and annotation cost exhibited by current methods, we propose a new DPO variant, called CLIP-DPO, that uses a pre-trained CLIP model [56] to rank the LVLM's self-generated captions to construct positive-negative pairs for DPO. Since the CLIP model was trained in a contrastive manner to measure the alignment between image-text pairs, it is naturally suitable for determining the quality of a given output from an LVLM, grounding it to the correct object or attribute. The dataset over which we operate is constructed by running the original pre-trained target LVLM on its own output obtained using prompting, removing both the need for (i) additional external data and (ii) ensembles of external LVLMs. The final data is filtered before training using robust rule-based filtering. + +Main results. We applied CLIP-DPO fine-tuning on top of two state-of-the-art models: MobileVLM-v2 (3 models in total) and LlaVA-1.5 7B. We find that, in both cases, our approach is effective in reducing hallucinations, outperforming all baseline models (i.e. the models without CLIP-DPO fine-tuning) by a significant margin. Importantly, CLIP-DPO significantly outperforms our direct competitor HA-DPO [83], outperforms Qwen-VL [5] trained on significantly larger datasets (i.e. 1.4B samples for pre-training and 77M for multitask training; as opposed to CLIP-DPO training on just 0.7M samples). Finally, our model's enhanced object grounding capabilities are also illustrated for zero-shot image classification, all without degrading the original performance of the base LVLM model. + +# 2 Related work + +Large Visual Language Models (LVLMs). Following the unprecedented success of Large Language Models (LLMs) [7, 50, 68, 69, 80], several works have recently proposed to build multi-modal capabilities on top of them [10, 14, 16, 42]. + +LLaVA [42] and FROMAGe [31] directly pass the visual tokens produced by a pre-trained CLIP [56] vision encoder to an LLM, either fine-tuning the LLM or adapting it using LoRA [24]. Notably, LLaVA training includes a pre-training stage that aligns the CLIP features with the LLM input tokens using a simple projection layer (i.e. keeping the rest of the model frozen) using image captioning data. InstructBLIP [16] uses QFormer [33] to reduce the number of vision tokens before passing them to the LLM. As the quality and distribution of the training data plays a crucial role [84], a series of methods [9,10,55,76] introduced improved data construction pipelines. For example, ShareGPT4v [10] uses the API of GPT-4V [51] to first label, then train a model, and finally use it to re-announce a new training set. Another line of work is improving efficiency by reducing the models' size (original LLaVA models have 7B and 13B parameters). LLaVA-Phi [87], MobileVLM [13] and its follow-up, MobileVLM-v2 [14], replace the LLaMA [68] and Vicuna [12] LLMs with the smaller 1.4B and 2.7B variants of MobileLLaMA and Phi [38]. Our work is orthogonal to the aforementioned methods and does not seek to improve the model's architecture. Instead, we propose CLIP-DPO, an improved training approach for LVLMs based on DPO. + +Preference optimization. Instruction tuning can significantly improve LLMs' perceived output quality and usefulness by aligning their responses to a given task domain or human preferences. This is achieved either by direct fine-tuning on expert data [15, 47, 67] or via reinforcement learning [52, 64, 88]. The latter significantly simplifies the data collection process, but still requires complicated training algorithms based on REINFORCE [73] or PPO [61]. Recently, a much-simplified approach was proposed, Direct Preference optimization (DPO) [57], which bypasses the need to train a reward model and allows direct training using a cross-entropy loss. Multiple improved versions for LLMs have been proposed in the meantime [4, 75, 82]. Following this, a recent wave of works on combining DPO with LVLMs have been proposed [35, 36, 83]. Silkie [35] constructs a multimodal instructional dataset automatically labeled by GPT-4V. Similarly, HA-DPO [83], aiming to reduce the rate of hallucinations, uses a GPT4 model to label and construct positive-negative pairs with and without hallucinations. The work of [36] follows a similar path by using a suite of LLMs and LVLMs (i.e. Gemini-Vision [66]) to generate and label the data. These methods are then primarily evaluated on the LLaVA benchmark and PoPE [72]. In contrast to the aforementioned works, the proposed CLIP-DPO (i) simplifies the pipeline, (ii) removes the need for paid APIs, (iii) removes the need for additional data, and (iv) removes the need for additional external LVLMs. Instead, we make use of a pre-trained CLIP model to rank the generated outputs from a small pool of efficient LVLMs, and show that the proposed CLIP-DPO training provides significant improvements for fixing LVLM hallucinations and, in general, for enhancing the discriminability and robustness of the model as demonstrated by image classification experiments. + +LVLM hallucinations. Broadly speaking, in the context of LVLMs, we consider hallucinations to be incorrect or misleading generated text, contradicting the visual evidence provided by the input image. This is an undesirable characteristic + +![](images/6a9ae7b1a8a89479afbf2abdd702599b518d1594d17601816cb4fba608e505e0.jpg) +Fig. 1: An example of the injected hallucinations. Given an image and its caption, we prompt GPT-4 to generate 3 types of hallucination: existence, attributes, and relation. + +Original Caption: In this image, we see a black dog wearing an orange vest standing on the grass. The dog's head is down and it appears to be in a backyard. + +- Existence Hallucination: In this image, we see a black dog wearing an orange vest standing on the grass, with a small orange ball near his feet. The dog's head is down and it appears to be in a backyard. +- Attributes Hallucination: In this image, we see a dark brown dog wearing an orange vest standing on the grass. The dog's head is down and it appears to be in a backyard. +- Relation Hallucination: In this image, we see a black dog wearing an orange vest sitting on the grass. The dog's head is down and it appears to be in a backyard. + +inherited from the pre-trained LLM used, and further exacerbated by the visual-language alignment process [5, 25, 29, 40]. Multiple solutions have been recently proposed by the community with varying degrees of success. The works of [25, 77] attempt to address the data bias by constructing better-grounded annotated image-text pairs. The works of [5, 11] scale the resolution of the image encoder, as this was observed to reduce the amount of hallucination, but at the cost of a high increase in the computational cost. The works of [27, 78] improve the vision encoder by adding extra informational paths. The closest work related to ours is that of [83], which, with the help of GPT-4V, constructs negative-positive pairs for DPO fine-tuning. Different from [83], we use a pre-trained CLIP model to perform the ranking, and no additional data or external LVLMs are required. For evaluation purposes, we use AMBER [71], the most comprehensive benchmark for hallucinations to date, encompassing both a generative and discriminative evaluation component. Importantly, unlike all prior benchmarks [22, 30, 37, 40, 65], AMBER is a high-quality dataset fully annotated by humans for both the generative and discriminative tasks. + +# 3 Preliminaries + +# 3.1 The Effectiveness of CLIP in Reducing Hallucinations + +First, we probe the effectiveness of CLIP in accurately ranking correct captions over those with hallucinated content. This assessment will indicate the efficacy of CLIP-style models in reducing hallucinations with DPO-based training. To this end, initially, we select 1K captions from the Detailed Caption [1] dataset, chosen for its high-quality labels. Then, we instruct GPT-4 [2] to create three hallucinated captions for each image, corresponding to the following types of hallucination: (i) existence: new elements or objects are added to the caption that were not mentioned originally, (ii) attribute: the attributes, characteristics, or features of the original caption's elements are altered, and (iii) relationship: the spatial, contextual, or interactive relationships of the original caption's elements are altered. See Fig. 1 for an example of the injected hallucinations. + +Next, we use LLaVA-1.5 7B [41] to compute the likelihood of all captions, including the original and the hallucinated captions. We then retain only the hallucinated captions for which the model predicts a higher likelihood than the + +original, yielding 61 for existence, 76 for attribute, and 119 for relationship hallucinations out of the initial 1K captions. Finally, we compute the CLIP + +image-text scores for these likely yet incorrect captions. As illustrated in Fig.2, CLIP manages to accurately rank the original caption over the hallucinated ones in at least $59\%$ of cases, reaching up to $83\%$ for the existence type of hallucinations. These findings underscore the potential of using VL embedding models to provide a reliable training signal for hallucination reduction in LVLMs. + +![](images/fddc432d5c60939e88adc6911e453efe23869fd149a59cfe100c41e56feb3b42.jpg) +Fig. 2: In dark blue, we show the number of hallucinated captions per type that LLaVA-1.5 7B assigns a higher likelihood than the original out of 1K captions. Light blue shows the portion of samples corrected by CLIP. + +# 3.2 Direct Preference Optimization + +The standard pipeline for LLM and LVLM alignment consists of three steps: generation, annotation, and op + +timization. First, a set of $N$ prompts $\{x_{i}\}_{i = 1}^{N}$ , each one consisting of an image and a text component in the case of LVLMs, are used to generate a set of response pairs $y_{i}^{1}$ and $y_{i}^{2}$ obtained from a pool of pre-trained LVLMs for each prompt $x_{i}$ . Then, either human annotators or another set of LLMs (i.e. AI annotators) are used to rank the responses, resulting in a preferred $y_{i}^{+}$ and a less preferred response $y_{i}^{-}$ for each prompt $x_{i}$ , and thus a final preference dataset $\mathcal{D} = \{(x_i,y_i^+,y_i^-)\}_{i = 1}^N$ . Finally, Direct Preference Optimization [57] (DPO) can be applied to update the target policy $\pi_{\theta}$ parameterized by $\theta$ directly using the preference dataset $\mathcal{D}$ . Specifically, the DPO optimization objective is defined as follows: + +$$ +\max _ {\pi_ {\theta}} \mathbb {E} _ {(x _ {i}, y _ {i} ^ {+}, y _ {i} ^ {-}) \sim \mathcal {D}} \log \sigma \left(\beta \log \frac {\pi_ {\theta} (y _ {i} ^ {+} \mid x _ {i})}{\pi_ {\mathrm {r e f}} (y _ {i} ^ {+} \mid x _ {i})} - \beta \log \frac {\pi_ {\theta} (y _ {i} ^ {-} \mid x _ {i})}{\pi_ {\mathrm {r e f}} (y _ {i} ^ {-} \mid x _ {i})}\right), \tag {1} +$$ + +with $\pi_{\theta}$ as the policy to be learned, $\pi_{\mathrm{ref}}$ as the reference SFT policy, and $\beta$ as a hyperparameter to control the Kullback-Leibler (KL) divergence between the learned $\pi_{\theta}$ and reference $\pi_{\mathrm{ref}}$ policies [57]. The main benefit of DPO-based optimization is the direct alignment of the LVLM towards the preferences implicit in the preference data $\mathcal{D}$ . + +In contrast to previous DPO-based LVLM preference optimization works [35, 36, 83], next, we will introduce the proposed CLIP-DPO, which (i) simplifies the generation process by limiting the pool of LVLMs used for generation to small and efficient models, (ii) removes the need for external LLMs and LVLMs annotators accessed via paid APIs (e.g. GPT-3.5, GPT-4 or GPT-4V) by using CLIP as the ranking model, and (iii) removes the need for additional data by reusing the same data used during the SFT step. + +![](images/f79e40a20a29c0337b7e8f8f3ec82dc49962d47d9d910f6a5b6742c549fd9540.jpg) +Fig. 3: CLIP-DPO. Starting from the initial SFT data pool and a set of prompts, an LVLM generates captions. These captions are first ranked using a CLIP model, then filtered to identify the most suitable positive and negative pairs for DPO-based optimization. + +# 4 Method + +In this section, we introduce our DPO variant for LVLMs preference optimization, named CLIP-DPO. As illustrated in Fig. 3, CLIP-DPO follows a three-step process: generation, annotation, and optimization, similar to the standard DPO-based training pipeline. While the final DPO-based optimization step remains unchanged, CLIP-DPO modifies the first two steps. + +Firstly, we streamline the generation step by utilizing the same data pool (i.e. images only) as the SFT stage and generate data using either the model itself or a small pool of efficient LVLMs. Secondly, we introduce an annotation step tailored for CLIP-style embedding models, where we rank captions based on their image-text similarities. Then, given that the resulting captions and pairs may vary in quality, we implement a rule-based filtering strategy to remove unsuitable candidates. Finally, with the constructed preference data, the model undergoes DPO-based optimization using Eq. 1, as shown in Fig. 3. + +As such, by leveraging VL embedding models for data annotation, we tap into the extensive information captured by these contrastively-pretrained models, which have been exposed to hundreds of millions of unique image-text pairs during training, unlike existing open-sourced LVLMs. Additionally, unlike the discrete GPT-4V ranking used in prior work, CLIP scores are continuous, enabling finer-grained comparisons. This allows us to assess the difficulty of a given pair based on the margin between the score values. + +Next, we will elaborate on each module of our pipeline. Sec. 4.1 details the data generation module, and Sec. 4.2 details the data annotation process. + +# 4.1 Data Generation + +The first step of our pipeline consists of generating a set of per-image captions that will be later ranked by a pre-trained CLIP model, along with a subsequent filtering approach to select a set of positive and negative pairs, which are then used for DPO-based training. To this end, we start by selecting the pool of LVLM annotators and the data to be annotated. For the annotators and to reduce the + +cost of the generation step, we select MobileVLM-v2 [14] family of models, given their efficiency and performance. As for the data, and to avoid introducing any new sources that might bias the results of CLIP-DPO, and to further reduce the method's cost, we opt for the initial pool of data used during the SFT stage of MobileVLM-v2 [14] models, see Table 1 for details. Then, we conduct our data generation pipeline that consists of two steps, the generation of generic captions and per-image questions and answers. + +
DatasetDataset Size
COCO118K
GQA72K
LLaVA-1.0 Pretraining595K
LLaVA-1.5 Pretraining558K
OCR VQA80K
SAM570K
SBU845K
TextVQA22K
VG86K
Web-Celebrity495
Web-Landmark500
Wikiart500
Total2.9M
+ +Table 1: Data pool used for data generation. + +![](images/05ffd1475c796d2f88ab645289a33b8e1a495813251ed3b521f3a2cd761937b5.jpg) +Fig. 4: The distribution of CLIP image-text scores per category. + +![](images/2d257b53e0d3d9670fa31a085752a1c82d358b45ef8218a48a098507e4cc0800.jpg) +Fig. 5: Examples of generated generic captions produced by MobileVLM-v2 1.7B and MobileVLM-v2 3B. For MobileVLM-v2 3B's generated captions, we also show the produced question and positive and negative answers generated using Mistral 7B. + +# MobileVLMv-2 1.7b: + +- The image showcases a charming two-story building, characterized by its steep roof and a balcony adorned with red flowers. The building's facade is painted white, accented with a green roof that adds a touch of color to the scene. +The image captures a charming scene of a German half-timbered building, bathed in the soft glow of daylight. + +# MobileVLMv-2 3b: + +- The image showcases a traditional German half-timbered house, a common architectural feature in the region. The house is adorned with a green roof, a characteristic of half-timbered houses, and is embellished with red flowers that add a vibrant touch to the overall design. +- A building with a statue of a knight in the middle of the building. + +# Mistral 7b: + +- What color is the roof of the building in the image? +- Positive: Green Negative: Belge + +Generation of Generic Captions. We start by generating a set of 5 descriptive captions per image. For each of the MobileVLM-v2 models, we prompt with 5 different prompts (e.g. "identify the setting and note any characters or objects, focusing on visible details.") to increase the diversity of the generated captions. While these captions can be used for CLIP ranking and DPO-based training, they are still produced by generic prompts that are not image-specific. Next, we perform the second step of our data generation pipeline to obtain a set of per-image questions and answers. + +Generation of Per-Image Questions and Answers: To obtain a set of questions per image, we leverage an LLM, i.e. a variant of Mistral 7B Instruct + +![](images/5e879f6d006e00e68966d810102b196309f7bf5bc56fd93a4cd95bd2db17d628.jpg) +Fig. 6: Question Filtering. The sample is rejected if the CLIP image-question score is low for a given sample. If the question is relevant and not generic, it is kept. Next, we evaluate the quality of the answer by first parsing the question into a caption, appending the positive answer to it, and computing its CLIP score. If it is a high score, the sample is kept for DPO-based training. + +Question 1: What is the central focus of the image? + +- Question CLIP Score: 15.0 +- Question Filter: Rejected + +Question 2: What type of market is depicted in the image? + +- Question CLIP Score: 26.7 +- Question Filter: Kept + +Positive: A fruit and vegetable market. + +- Caption: type of market is depicted in the image a fruit and vegetable market. +Positive CLIP Score: 29.1 +- Question Filter: Kept + +![](images/921def778516fde821386dde71ff3277789985a0bd5f78379e894c697c348963.jpg) +Fig. 7: Examples of the final positive and negative pairs obtained post-filtering, to be used for DPO-based training. + +Prompt: Provide a brief description of the given image. + +Positive: An Apple computer with a laptop computer sitting on a desk. + +Negative: A laptop and a desktop computer are sitting on a desk. + +![](images/48e92c389da54ece0c3757c592c986106ec6927b53055777c6054d1d5c245744.jpg) + +Prompt: Provide a depiction of the picture. Positive: A man in orange shirt kicking a + +soccer ball + +Negative: A group of people kicking a + +soccer ball + +![](images/5aa0d5905cac1a7c42a519100dad450e5d7d57fbbfec85b4c81a9b797eea8476.jpg) + +Prompt: What is the name of the paint brand featured on the bucket? + +Positive: Homer's All Purpose Bucket + +Negative: Home Depot All-Purpose Bucket. + +v0.2 [28] $^1$ , and feed it with the generated captions and prompt it to generate 2 questions for each image, together with positive and negative answers. The LLM is asked to generate the positive ones based on the fed captions and to generate plausible but incorrect negatives given the image description/caption. + +For an example of the generated captions, questions, and answers, see Fig. 5. + +# 4.2 Data Annotation + +Starting with the generated captions, the subsequent data annotation step involves ranking them based on their CLIP image-text similarities and filtering them to obtain a final set of high-quality positive-negative pairs for DPO-based training. Our filtering pipeline includes two stages: global filtering, where we eliminate low-quality images and captions, and pair-filtering, where we select the best positive and negative per-image pairs for DPO-based training. + +CLIP Ranking. The initial step of data annotation is CLIP ranking. In this step, we use a pre-trained CLIP model to compute the cosine similarities of the captions with their associated images and then sort them from highest to lowest. + +Global Filtering. Next, we aim to filter our data and only retain a balanced and higher-quality subset. To achieve this, we start by analyzing the types of images in our data pool (see Tab. 1). We define a set of four generic categories: images of text, people, objects, and scenes. Using CLIP and a set of 10 descriptions per category, we create four class prototypes. Each image is then assigned to one of these prototypes based on the highest cosine similarity. As illustrated in Fig. 4, we observe that CLIP scores for the text category are the highest, while scores for people, objects, and scenes are relatively similar. Consequently, before applying CLIP-based filtering, we down-sample the portion of text images. We then filter out all generic captions below a given CLIP score ( $e.g. < 28.0$ ). Additionally, since CLIP is primarily trained on short text, we remove long captions to ensure more precise CLIP scores. For questions, we compute their CLIP scores and remove all questions with low CLIP scores ( $e.g. < 25.0$ ), thereby eliminating generic questions ( $e.g. "what is the main object in the image?")$ that are already covered by the generated captions. + +Pair Filtering. Finally, starting from the remaining high-quality image captions and image question-answers, the final step will be selecting a set of positive and negative pairs for DPO-based training. For the questions, since we already have a set of positive and negative responses generated by our LLM, we only need to filter out the low-quality pairs. To achieve this, we use simple regex matching rules to extract an image description from the question, append the positive answer to it to create a synthetic caption, and compute its CLIP image-text scores. We then reject examples where the scores are low. See Fig. 6 for an example of the questions filtering process. As for the captions, we consider all possible pairs where the CLIP score difference between two captions is larger than a given threshold ( $e.g. > 2.0$ ) and where the length of the two captions is similar to avoid introducing false preferences. We then order them based on the CLIP score difference between the positive and negative captions and select the top-ranking pair per image. + +After the data annotation step, the resulting DPO training data consists of 750K pairs, of which 50K are question-answer pairs and the rest are caption pairs. For examples of the final pairs, see Fig. 7. + +# 5 Implementation details + +Network architecture: In this work, we consider two LVLMs architectures: MobileVLM-v2 [14] and LLaVA-1.5 [42]. Both models follow the same overall structure: a pre-trained CLIP vision encoder and a pre-trained LLM. The visual tokens produced by the frozen vision encoder are projected using either a linear layer or a small projection module and passed as input to the LLM. LLaVA opts for a pre-trained Vicuna LLM [12] while MobileVLM-v2 uses MobileLLaMA [13], except for their 7B variant, which also uses a Vicuna model. Both use the same ViT-L-14 @ 336px CLIP visual encoder [56]. For efficiency purposes, MobileVLM-v2 uses a projection module that halves the number of + +visual tokens. We consider the following model variants in our comparisons: MobileVLM-v2 (1.7B, 3B and 7B), and LLaVA-1.5 7B. For data annotation, we use ViT-H/14 DFN [19] as our pre-trained CLIP model. + +Training details: For all of our experiments, unless specified otherwise, we start from the pre-trained MobileVLM-v2 and LLaVA-1.5 models. The models are then fine-tuned for 1 epoch using DPO-based optimization on the constructed CLIP-DPO dataset. We use the following hyperparameters for fine-tuning the models: AdamW optimizer [45] with a batch size of 256, a learning rate of $5e - 7$ , decreased to 0 using a cosine scheduler, a warm-up of 0.01, and a weight decay set to 0. Both during training and testing, the input images are cropped and resized to $336 \times 336$ px. The training was performed on 8 A100 GPUs using Pytorch [54]. For the larger models, e.g. LLaVA-1.5 7B, to fit them in memory, we use the Zero-3 strategy [58, 59] and decrease the batch size to 64. + +# 6 Results + +We first evaluate the effectiveness of CLIP-DPO in reducing hallucinations on the recently introduced AMBER [71] benchmark in Sec. 6.1. Moreover, we demonstrate its enhanced grounding capabilities by reporting zero-shot image classification in Sec. 6.2. Finally, in Sec. 6.4, we show that the proposed CLIP-DPO training does not compromise LVLM performance by reporting results on the LLaVA benchmark [87]. + +# 6.1 Evaluation on hallucinations + +We start by evaluating CLIP-DPO in terms of LVLM hallucination reduction, which is our main objective. We use AMBER [71], a comprehensive, high-quality, and LLM-free multidimensional benchmark for LVLM hallucination evaluation, which can be used to evaluate both generative and discriminative tasks. As the results from Tab. 2 show, MobileVLM-v2 trained with CLIP-DPO improves upon MobileVLM-v2 baselines across all model sizes, and sometimes quite significantly, especially for the 1.7B and 7B models. It can also be seen that CLIP-DPO significantly improves when applied on top of LLaVA-1.5 7B. Importantly, CLIP-DPO significantly outperforms HA-DPO [83], the main competing approach, improving the AMBER score 3.2 vs 7.8 when using LLaVA-1.5. Finally, our LlaVA-1.5 7B+CLIP-DPO even outperforms Qwen-VL [5], which is trained on significantly larger datasets (1.4B image-text pairs for pre-training and 77M for multitask training while CLIP-DPO is fine-tuned on just 0.7M samples), and even matches the performance of GPT-4V without using any GPT-4V model outputs during training. Overall, these results on a high-quality state-of-the-art benchmark such as AMBER clearly demonstrate the effectiveness of our CLIP-DPO approach for reducing hallucinations. + +# 6.2 Zero-shot image classification + +A primary reason for the hallucinatory behavior of LVLMs is the weak alignment between their visual features and the input LLM tokens. A direct way to + +
ModelGENERATIVE TASKDISCRIMINATIVE TASKAMBER
CHAIR↓Cover↑Hal↓Cog↓Acc.P.R.F1
mPLUG-Owl21.650.176.111.540.192.810.518.948.7
LLaVA11.551.048.85.542.774.121.032.760.6
MiniGPT-413.663.065.311.363.690.550.464.775.6
CogVLM5.657.223.61.369.088.960.972.383.4
mPLUG-Owl210.652.039.94.575.695.066.978.584.0
InstructBLIP8.852.238.24.476.584.579.081.786.5
Qwen-VL5.549.423.61.981.290.879.784.989.7
GPT-4V4.667.130.72.683.484.990.187.491.4
LLaVA-1.5 7B7.851.036.44.272.093.262.474.783.5
+ HA-DPO7.233.619.72.668.368.198.480.586.7
+ CLIP-DPO3.747.816.61.377.884.481.582.989.6
MobileVLM-v2 1.7B3.839.68.90.565.492.651.966.581.4
+ CLIP-DPO4.238.910.80.571.288.764.874.985.3
MobileVLM-v2 3B4.838.711.10.773.592.165.776.786.0
+ CLIP-DPO4.741.513.10.576.791.071.880.387.8
MobileVLM-v2 7B4.438.910.40.671.995.060.874.184.9
+ CLIP-DPO4.038.010.10.577.393.470.880.588.3
+ +Table 2: Hallucination evaluation results on AMBER for both generative (leftmost set of columns) and discriminative tasks (rightmost set of columns). Our approach offers consistent and large performance improvements across 4 different LVLM models, and also beats HA-DPO by a large margin. + +
MethodStanfordCarsOxfordPetsOxfordFlowersImagenetFood-101EurosatCaltech-101UCF-101SUN397Avg
LLaVA-1.523.234.07.337.945.349.582.750.543.140.0
+ HA-DPO23.233.47.136.945.149.482.750.342.739.7
+ CLIP-DPO30.144.816.342.253.952.184.853.148.847.4
MobileVLM-v2 1.7B17.115.214.532.631.249.377.945.539.333.9
+ CLIP-DPO19.232.623.841.347.448.980.850.249.343.7
MobileVLM-v2 3B14.823.59.135.738.853.281.848.642.036.7
+ CLIP-DPO28.640.919.544.352.056.485.552.350.147.7
MobileVLM-v2 7B27.132.511.537.745.043.181.949.046.641.6
+ CLIP-DPO32.551.335.850.162.659.388.356.254.854.5
+ +Table 3: Zero-shot image recognition results in terms of Top-1 accuracy (\%). While HA-DPO shows very similar performance to the base model, our CLIP-DPO improves the base model by a very large margin in all cases. + +evaluate this is through simple zero-shot image classification, which is the goto benchmark for contrastively trained VL models like CLIP [56]. The typical setup follows a closed-set classification problem, where the names of all possible classes are known a priori and are encoded into class prototypes using CLIP. A given image is then assigned to the class with the highest CLIP image-text scores. Herein, we follow the same protocol with the notable difference that, as our goal is LVLM evaluation, we first prompt the model to generate a free-form image caption describing the main object in the image, then encode the caption using the CLIP text encoder and assign the image to the class with the highest text-text (i.e. caption-class) CLIP score. Here, we opted for a different family of VL embedding model, i.e. SigLIP [79], to avoid evaluating using CLIP models similar to those used during the data annotation step. Following [56,85], we evaluate our approach on a suite of 9 diverse datasets: UCF-101 [63], SUN397 [74], + +
MethodLLMRes.GQA\( SQAI \)\( VQAT \)POPE\( MME^P \)\( MMB^{dev} \)Avg.
BLIP-2 [33]Vicuna-13B22441.061.042.585.31293.8--
InstructBLIP [16]Vicuna-13B22449.563.150.778.91212.8--
Shikra [9]Vicuna-13B224-----58.8-
Openflamingo [3]MPT 7B336--33.6--4.6-
Qwen-VL [5]Qwen 7B44859.367.163.8-1487.638.2-
mPLUG-Owl [76]LLaMA 7B224----967.349.4-
MiniGPT-v2 [8]LLaMA 7B44860.3----12.2-
MiniGPT-4 [86]Vicuna 7B22432.2---581.723.0-
InstructBLIP [16]Vicuna 7B22449.260.550.1--36.0-
ShareGPT4V [10]Vicuna 7B33663.368.460.485.71567.468.870.8
MoE-LLaVA-1.6B×4 [39]StableLM-1.6B33660.462.647.884.31300.8*59.463.3
MoE-LLaVA-2.7B×4 [39]Phi-2.7B33661.168.750.285.01396.4*65.566.7
LLaVA-1.5 [41]Vicuna 7B33662.066.858.285.91510.764.368.8
+ HA-DPOVicuna 7B33661.969.258.384.31505.664.969.0
+ CLIP-DPOVicuna 7B33659.367.656.485.81468.764.967.9
MobileVLM 1.7B [13]MobileLLaMA 1.4B33656.157.341.584.51196.253.258.7
MobileVLM-v2 1.7B [14]MobileLLaMA 1.4B33658.366.752.184.31302.857.764.2
+ CLIP-DPOMobileLLaMA 1.4B33658.668.447.884.31331.157.663.8
MobileVLM- 3B [13]MobileLLaMA 2.7B33659.061.247.584.91288.959.662.8
MobileVLM-v2 3B [14]MobileLLaMA 2.7B33661.170.057.584.71440.563.268.1
+ CLIP-DPOMobileLLaMA 2.7B33660.972.357.385.11425.763.868.4
MobileVLM-v2 [14] 7BVicuna 7B33662.674.862.385.31560.769.272.1
+ CLIP-DPO 7BVicuna 7B33662.577.362.284.71539.170.472.3
+ +Table 4: Results on several vision language tasks from the LlaVA benchmark. Our approach is combined with LVLMs of different sizes, showing consistent performance in all cases. + +Stanford Cars [32], Oxford Pets [53], Oxford flowers [49], ImageNet [17], Food 101 [6], Eurosat [23] and Caltech-101 [20]. As the results from Tab. 3 show, all LVLMs fine-tuned with CLIP-DPO significantly outperform their corresponding baselines, showcasing the increased discriminative properties. Note again that these improvements are obtained without affecting the model's performance on other tasks and datasets. + +# 6.3 Additional evaluations + +Herein, we evaluate the impact of CLIP-DPO training on other vision language tasks and, more specifically, on the popular LLaVA-Bench (GQA [26], ScienceQA [46], TextVQA [62], MME [21], MMBench [44]). As Tab. 4 shows, overall, CLIP-DPO training does not compromise performance. We further note that accuracy improvements on LLaVA-Bench are heavily tied to the addition of extra training data [10] or architectural changes [13,14,39,41]. As we are not using any additional data & models nor making architectural changes, it is not surprising that the performance after CLIP-DPO training remains largely in line with that of the original baseline model. + +# 6.4 Qualitative Results + +In Figure 8, we show some qualitative examples comparing the predictions of LLaVA-1.5 7B, LLaVA-1.5 7B+HA-DPO, and LLaVA-1.5 7B+CLIP-DPO. Over + +![](images/2a6af99c9dfdbc1d2d4c156f2e4f9ea6ff9fc41ff05667cc675777cddd9fef41.jpg) + +Provide a short description of the image. + +LLava-1.5-7b: A large silver airplane is parked inside a hangar. The plane is positioned on a runway, and there is a ladder leading up to the cockpit. The hangar appears to be a large building, possibly a factory or a maintenance facility. + ++ HA-DPO: A large silver airplane is parked inside a hangar. ++ CLIP-DPO: A fighter jet is parked inside a hangar, with a ladder attached to it. The jet is numbered 157. + +![](images/b05df5ad4610a422e25687c4927231823f6d240d4be8c0a3e0c1692675623e15.jpg) + +Provide a short description of the image. + +LLava-1.5-7b: A green bus is parked in a parking lot, waiting for passengers. ++ HA-DPO: A green bus is parked in a parking lot, waiting for passengers. ++CLIP-DPO: A green Go3 bus is parked at a bus stop. + +Fig. 8: Qualitative examples comparing our method with LLaVA and HA-DPO. + +
Loss variantAMBER ScoreCls. Avg.LLaVA bench.
DPO (β = 0.1) [57]85.343.763.8
DPO (β = 0.3) [57]83.342.164.0
IPO [4]83.640.564.5
KTO [18]83.539.564.7
SLIC [43]83.242.764.0
cDPO (β = 0.1, LS = 0.1 [43])83.143.263.9
cDPO (β = 0.3, LS = 0.1 [43])83.140.764.5
+ +Table 5: Effect of DPO loss variant on CLIP-DPO training. Results are reported in terms of AMBER Score, Average Top-1 (\%) accuracy across all 9 image classification datasets, and the average on LLaVA benchmark. MobileVLM-v2 1.7B was used. + +all, with CLIP-DPO, the model is more grounded in the visual content, less prone to hallucinations, and more precise and fine-grained in its descriptions. + +# 7 Ablation studies + +Herein, we ablate the effect of the proposed components. In all experiments, we use a MobileVLM-v2 1.7B model. + +# 7.1 Effect of DPO loss variant + +Our work primarily follows the original DPO formulation proposed in [57]. Here we also consider the following recently proposed variations: KTO [18], ITO [4], SLIC [43] and cDPO [48]. As Tab. 5 shows, aggregated, all losses tend to perform similarly, with DPO marginally outperforming the others. + +# 7.2 Effect of CLIP scorer + +Herein, we seek to explore alternatives to ViT-H/14 DFN [19] CLIP model used in previous experiments, analyzing the impact of the scorer used on the overall performance of the model. We consider a diverse set of alternatives, covering + +
ScorerAMBER ScoreCls. Avg.LLaVA bench.
ViT-L/14 [56]79.842.363.0
SigLIP-L/16 [79]85.043.963.2
ViT-H/14 [60]84.344.263.7
ViT-H/14 DFN [19]85.343.763.8
+ +Table 6: Effect of CLIP scorer on CLIP-DPO training. Results are reported in terms of AMBER Score, Average Top-1 (\%) accuracy across all 9 image classification datasets, and the average on LLaVA benchmark. MobileVLM-v2 1.7B was used. + +multiple exploratory paths: equally-sized models trained on different data (ViT-H/14 [19] trained on LAION-2B instead of DFN-5B); smaller models (ViT-L/14 [56]) and models trained using different pre-training losses (SigLIP-ViT-L/16 [79]). As the results from Tab. 6 show, our approach is generally robust to the exact scorer used, with the notable exception of the ViT-L/14 [56] model. Notice that this difference is primarily manifesting on the AMBER benchmark. Intuitively, this showcases the importance of a powerful scorer for reducing the amount of hallucinations. + +# Limitations and broader impact + +As an LVLM-based approach, our method is subject to the same general consideration (i.e. potential data bias, susceptibility to hallucinations, etc.). Moreover, as the LVLMs are trained on relatively small datasets compared to LLMs or CLIP, gaps within their knowledge domains are possible. This is especially important as neural networks tend to be overconfident outside their seen input distribution. As with all models from this category, we strongly recommend checking the models and the data carefully before deploying them. Despite these general aspects, our approach is shown to significantly reduce the amount of hallucinated content and improve the model's discriminability, hence resulting in more robust and reliable models. + +# 8 Conclusions + +In this work, we proposed CLIP-DPO, a simple method that reduces hallucinations in LVLMs based on using a pre-trained CLIP model [56] to rank the LVLM's self-generated captions in order to construct positive-negative pairs for DPO fine-tuning. In contrast to previous works, our proposed CLIP-DPO removes the need of (i) paid-for APIs, (ii) additional data and (iii) additional external LVLMs. Before training, the data is filtered using a newly proposed robust rule-based approach. When applied on top of established LVLMs, it is shown that CLIP-DPO fine-tuning significantly reduces hallucinations, outperforming the baseline models by significant margins and even matching the performance of GPT-4V on the AMBER benchmark. We also observe superior zero-shot object recognition, suggesting improved object grounding capabilities. + +# References + +1. Detailed caption dataset. https://huggingface.co/datasets/echo840/Detailed_Caption (2024) +2. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) +3. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., et al.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023) +4. Azar, M.G., Rowland, M., Piot, B., Guo, D., Calandriello, D., Valko, M., Munos, R.: A general theoretical paradigm to understand learning from human preferences. arXiv preprint arXiv:2310.12036 (2023) +5. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-VL: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966 (2023) +6. Bossard, L., Guillaumin, M., Van Gool, L.: Food-101-mining discriminative components with random forests. In: European Conference on Computer Vision (2014) +7. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances on Neural Information Processing Systems (2020) +8. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023) +9. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023) +10. Chen, L., Li, J., Dong, X., Zhang, P., He, C., Wang, J., Zhao, F., Lin, D.: Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793 (2023) +1. Chen, Z., Wu, J., Wang, W., Su, W., Chen, G., Xing, S., Muyan, Z., Zhang, Q., Zhu, X., Lu, L., et al.: Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238 (2023) +2. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023) +3. Chu, X., Qiao, L., Lin, X., Xu, S., Yang, Y., Hu, Y., Wei, F., Zhang, X., Zhang, B., Wei, X., et al.: MobileVLM: A fast, reproducible and strong vision language assistant for mobile devices. arXiv preprint arXiv:2312.16886 (2023) +4. Chu, X., Qiao, L., Zhang, X., Xu, S., Wei, F., Yang, Y., Sun, X., Hu, Y., Lin, X., Zhang, B., et al.: MobileVLM V2: Faster and stronger baseline for vision language model. arXiv preprint arXiv:2402.03766 (2024) +5. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022) +6. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: InstructBLIP: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023) + +17. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: IEEE Conference on Computer Vision and Pattern Recognition (2009) +18. Ethayarajh, K., Xu, W., Muennighoff, N., Jurafsky, D., Kiela, D.: Kto: Model alignment as prospect theoretic optimization. arXiv preprint arXiv:2402.01306 (2024) +19. Fang, A., Jose, A.M., Jain, A., Schmidt, L., Toshev, A., Shankar, V.: Data filtering networks. arXiv preprint arXiv:2309.17425 (2023) +20. Fei-Fei, L., Fergus, R., Perona, P.: Learning generative visual models from few training examples: An incremental Bayesian approach tested on 101 object categories. In: IEEE Conference on Computer Vision and Pattern Recognition - Workshops (2004) +21. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023) +22. Gunjal, A., Yin, J., Bas, E.: Detecting and preventing hallucinations in large vision language models. arXiv preprint arXiv:2308.06394 (2023) +23. Helber, P., Bischke, B., Dengel, A., Borth, D.: EuroSAT: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 12(7), 2217-2226 (2019) +24. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) +25. Hu, H., Zhang, J., Zhao, M., Sun, Z.: Ciem: Contrastive instruction evaluation method for better instruction tuning. arXiv preprint arXiv:2309.02301 (2023) +26. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: IEEE Conference on Computer Vision and Pattern Recognition (2019) +27. Jain, J., Yang, J., Shi, H.: Vcoder: Versatile vision encoders for multimodal large language models. arXiv preprint arXiv:2312.14233 (2023) +28. Jiang, A.Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D.S., Casas, D.d.l., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., et al.: Mistral 7b. arXiv preprint arXiv:2310.06825 (2023) +29. Jiang, C., Xu, H., Dong, M., Chen, J., Ye, W., Yan, M., Ye, Q., Zhang, J., Huang, F., Zhang, S.: Hallucination augmented contrastive learning for multimodal large language model. arXiv preprint arXiv:2312.06968 (2023) +30. Jing, L., Li, R., Chen, Y., Jia, M., Du, X.: Faithscore: Evaluating hallucinations in large vision-language models. arXiv preprint arXiv:2311.01477 (2023) +31. Koh, J.Y., Salakhutdinov, R., Fried, D.: Grounding language models to images for multimodal inputs and outputs. International Conference on Machine Learning (2023) +32. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3D object representations for fine-grained categorization. In: IEEE International Conference on Computer Vision - Workshops (2013) +33. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) +34. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Learning Representations (2022) + +35. Li, L., Xie, Z., Li, M., Chen, S., Wang, P., Chen, L., Yang, Y., Wang, B., Kong, L.: Silkie: Preference distillation for large visual language models. arXiv preprint arXiv:2312.10665 (2023) +36. Li, S., Lin, R., Pei, S.: Multi-modal preference alignment remedies regression of visual instruction tuning on language model. arXiv preprint arXiv:2402.10884 (2024) +37. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023) +38. Li, Y., Bubeck, S., Eldan, R., Del Giorno, A., Gunasekar, S., Lee, Y.T.: Textbooks are all you need ii: phi-1.5 technical report. arXiv preprint arXiv:2309.05463 (2023) +39. Lin, B., Tang, Z., Ye, Y., Cui, J., Zhu, B., Jin, P., Zhang, J., Ning, M., Yuan, L.: Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947 (2024) +40. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Mitigating hallucination in large multi-modal models via robust instruction tuning. In: International Conference on Learning Representations (2023) +41. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023) +42. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances on Neural Information Processing Systems (2024) +43. Liu, T., Zhao, Y., Joshi, R., Khalman, M., Saleh, M., Liu, P.J., Liu, J.: Statistical rejection sampling improves preference optimization. arXiv preprint arXiv:2309.06657 (2023) +44. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: MMBench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281 (2023) +45. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) +46. Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems (2022) +47. Mishra, S., Khashabi, D., Baral, C., Hajishirzi, H.: Cross-task generalization via natural language crowdsourcing instructions. arXiv preprint arXiv:2104.08773 (2021) +48. Mitchell, E.: A note on dpo with noisy preferences & relationship to ipo (2024) +49. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: Indian conference on computer vision, graphics & image processing. pp. 722-729 (2008) +50. OpenAI: Introducing chatgpt (2022) +51. OpenAI: Gpt-4v(ision) system card (2023) +52. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems (2022) +53. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: IEEE Conference on Computer Vision and Pattern Recognition (2012) +54. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch (2017) +55. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023) + +56. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021) +57. Rafailov, R., Sharma, A., Mitchell, E., Manning, C.D., Ermon, S., Finn, C.: Direct preference optimization: Your language model is secretly a reward model. Advances on Neural Information Processing Systems (2024) +58. Rajbhandari, S., Ruwase, O., Rasley, J., Smith, S., He, Y.: Zero-infinity: Breaking thegpu memory wall for extreme scale deep learning. In: Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis (2021) +59. Rasley, J., Rajbhandari, S., Ruwase, O., He, Y.: Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In: Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (2020) +60. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems (2022) +61. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017) +62. Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards VQA models that can read. In: IEEE Conference on Computer Vision and Pattern Recognition (2019) +63. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012) +64. Stiennon, N., Ouyang, L., Wu, J., Ziegler, D., Lowe, R., Voss, C., Radford, A., Amodei, D., Christiano, P.F.: Learning to summarize with human feedback. Advances in Neural Information Processing Systems (2020) +65. Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L.Y., Wang, Y.X., Yang, Y., et al.: Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525 (2023) +66. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023) +67. Thoppilan, R., De Freitas, D., Hall, J., Shazeer, N., Kulshreshtha, A., Cheng, H.T., Jin, A., Bos, T., Baker, L., Du, Y., et al.: Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239 (2022) +68. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) +69. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) +70. Wang, J., Yang, Z., Hu, X., Li, L., Lin, K., Gan, Z., Liu, Z., Liu, C., Wang, L.: Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100 (2022) +71. Wang, J., Wang, Y., Xu, G., Zhang, J., Gu, Y., Jia, H., Yan, M., Zhang, J., Sang, J.: An LLM-free multi-dimensional benchmark for MLLMs hallucination evaluation. arXiv preprint arXiv:2311.07397 (2023) + +72. Wang, J., Zhou, Y., Xu, G., Shi, P., Zhao, C., Xu, H., Ye, Q., Yan, M., Zhang, J., Zhu, J., et al.: Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126 (2023) +73. Williams, R.J.: Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning 8, 229-256 (1992) +74. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: Sun database: Large-scale scene recognition from abbey to zoo. In: IEEE Conference on Computer Vision and Pattern Recognition (2010) +75. Xu, J., Lee, A., Sukhbaatar, S., Weston, J.: Some things are more cringe than others: Preference optimization with the pairwise cringe loss. arXiv preprint arXiv:2312.16682 (2023) +76. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023) +77. You, H., Zhang, H., Gan, Z., Du, X., Zhang, B., Wang, Z., Cao, L., Chang, S.F., Yang, Y.: Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704 (2023) +78. Zhai, B., Yang, S., Zhao, X., Xu, C., Shen, S., Zhao, D., Keutzer, K., Li, M., Yan, T., Fan, X.: Halle-switch: Rethinking and controlling object existence hallucinations in large vision language models for detailed caption. arXiv preprint arXiv:2310.01779 (2023) +79. Zhai, X., Mustafa, B., Kolesnikov, A., Beyer, L.: Sigmoid loss for language image pre-training. arXiv preprint arXiv:2303.15343 (2023) +80. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022) +81. Zhao, H., Cai, Z., Si, S., Ma, X., An, K., Chen, L., Liu, Z., Wang, S., Han, W., Chang, B.: Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915 (2023) +82. Zhao, Y., Joshi, R., Liu, T., Khalman, M., Saleh, M., Liu, P.J.: Slic-hf: Sequence likelihood calibration with human feedback. arXiv preprint arXiv:2305.10425 (2023) +83. Zhao, Z., Wang, B., Ouyang, L., Dong, X., Wang, J., He, C.: Beyond hallucinations: Enhancing LVLMs through hallucination-aware direct preference optimization. arXiv preprint arXiv:2311.16839 (2023) +84. Zhou, C., Liu, P., Xu, P., Iyer, S., Sun, J., Mao, Y., Ma, X., Efrat, A., Yu, P., Yu, L., et al.: Lima: Less is more for alignment. Advances in Neural Information Processing Systems (2024) +85. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: IEEE Conference on Computer Vision and Pattern Recognition (2022) +86. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: MiniGPT-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023) +87. Zhu, Y., Zhu, M., Liu, N., Ou, Z., Mou, X., Tang, J.: LLaVA-phi: Efficient multi-modal assistant with small language model. arXiv preprint arXiv:2401.02330 (2024) +88. Ziegler, D.M., Stiannon, N., Wu, J., Brown, T.B., Radford, A., Amodei, D., Christiano, P., Irving, G.: Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593 (2019) \ No newline at end of file diff --git a/2024/CLIP-DPO_ Vision-Language Models as a Source of Preference for Fixing Hallucinations in LVLMs/images.zip b/2024/CLIP-DPO_ Vision-Language Models as a Source of Preference for Fixing Hallucinations in LVLMs/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0ea92b548d2090ab7d82dc10084c3fc77a5d4ac3 --- /dev/null +++ b/2024/CLIP-DPO_ Vision-Language Models as a Source of Preference for Fixing Hallucinations in LVLMs/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6136d07171b39323e7c6c54dd24508a9152793005d6fb03b80dff65592b22bb0 +size 449927 diff --git a/2024/CLIP-DPO_ Vision-Language Models as a Source of Preference for Fixing Hallucinations in LVLMs/layout.json b/2024/CLIP-DPO_ Vision-Language Models as a Source of Preference for Fixing Hallucinations in LVLMs/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..81f7a64defc8505848ed8bb13487a282a74863a5 --- /dev/null +++ b/2024/CLIP-DPO_ Vision-Language Models as a Source of Preference for Fixing Hallucinations in LVLMs/layout.json @@ -0,0 +1,10160 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 479, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 479, + 147 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 479, + 147 + ], + "type": "text", + "content": "CLIP-DPO: Vision-Language Models as a Source of Preference for Fixing Hallucinations in LVLMs" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 167, + 470, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 167, + 470, + 181 + ], + "spans": [ + { + "bbox": [ + 140, + 167, + 470, + 181 + ], + "type": "text", + "content": "Yassine Ouali1 Adrian Bulat1,2 Brais Martinez1 Georgios Tzimiropoulos1,3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 191, + 481, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 191, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 191, + 481, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 191, + 481, + 217 + ], + "type": "text", + "content": "Samsung AI Center Cambridge, UK " + }, + { + "bbox": [ + 133, + 191, + 481, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 191, + 481, + 217 + ], + "type": "text", + "content": " Technical University of Iasi, Romania " + }, + { + "bbox": [ + 133, + 191, + 481, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 133, + 191, + 481, + 217 + ], + "type": "text", + "content": " Queen Mary University of London, UK" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 244, + 455, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 244, + 455, + 453 + ], + "spans": [ + { + "bbox": [ + 160, + 244, + 455, + 453 + ], + "type": "text", + "content": "Abstract. Despite recent successes, LVLMs or Large Vision Language Models are prone to hallucinating details like objects and their properties or relations, limiting their real-world deployment. To address this and improve their robustness, we present CLIP-DPO, a preference optimization method that leverages contrastively pre-trained Vision-Language (VL) embedding models, such as CLIP, for DPO-based optimization of LVLMs. Unlike prior works tackling LVLM hallucinations, our method does not rely on paid-for APIs, and does not require additional training data or the deployment of other external LVLMs. Instead, starting from the initial pool of supervised fine-tuning data, we generate a diverse set of predictions, which are ranked based on their CLIP image-text similarities, and then filtered using a robust rule-based approach to obtain a set of positive and negative pairs for DPO-based training. We applied CLIP-DPO fine-tuning to the MobileVLM-v2 family of models and to LlaVA-1.5, in all cases observing significant improvements in terms of hallucination reduction over baseline models. We also observe better performance for zero-shot classification, suggesting improved grounding capabilities, and verify that the original performance on standard LVLM benchmarks is overall preserved." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 473, + 230, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 473, + 230, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 473, + 230, + 487 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": "A Large Vision Language Model (LVLM) is a Large Language Model (LLM) combined with an independently-trained vision encoder, typically taken from VL embedding models like CLIP [56] and often kept frozen. LVLMs have recently become the state-of-the-art for vision language understanding [5, 16, 42, 65, 81]. Unlike the prior generation of models [34, 70], which were typically tuned for one or a small number of tasks, LVLMs allow free-form dialog in natural language with an image in the loop. This direction largely draws inspiration from the recent success of ChatGPT [50] and respectively, ChatGPT-4V [51], which adapts LLMs to follow human instruction and preferences via various forms of reinforcement learning from human feedback (RFLH) and/or supervised fine-tuning (SFT) on high-quality multi-turn instruction (i.e. chat) data. In a similar fashion, the current generation of LVLMs is trained as part of a two-step process [42]. First, the CLIP vision encoder is aligned with the LLM by training a few adaptation layers only and, then, by fine-tuning the model using SFT on a multi-modal" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "content": "instructional dataset. Despite their remarkable success, a major limitation hindering the wider adoption of LVLMs is the high rate of hallucinated details (e.g. non-existing objects) that these models exhibit in their generated output text." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 156, + 481, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 156, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 132, + 156, + 481, + 323 + ], + "type": "text", + "content": "Motivation. Following the footsteps of LLM training, a natural solution to combat hallucinations is given by RFLH. As reinforcement learning-based training is generally expensive, often requiring the training of a policy model, Rafailov et al. [57] proposed a simplified framework, coined Direct Preference optimization (DPO), that allows direct training of the model using a binary cross-entropy objective. The data for DPO is typically obtained either from human preferences or using LLMs to construct preference data automatically. Thanks to its efficiency, DPO has been quickly adopted for LVLMs too [35, 36, 83]. However, these approaches require multiple LVLMs for generation, collecting additional data, and using GPT-4/GPT-4V API for ranking. Moreover, the GPT-4 provided scores are discrete (hence, of reduced granularity) and are themselves prone to hallucinations, which can perpetuate and exacerbate the already high rate of factual errors and poor image grounding [37, 71]. Finally, GPT-4/V is behind a paywall, and using it for generation/ranking is neither scalable nor cost-efficient." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 327, + 481, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 481, + 460 + ], + "type": "text", + "content": "Main idea. To fix hallucinations and address the high data construction and annotation cost exhibited by current methods, we propose a new DPO variant, called CLIP-DPO, that uses a pre-trained CLIP model [56] to rank the LVLM's self-generated captions to construct positive-negative pairs for DPO. Since the CLIP model was trained in a contrastive manner to measure the alignment between image-text pairs, it is naturally suitable for determining the quality of a given output from an LVLM, grounding it to the correct object or attribute. The dataset over which we operate is constructed by running the original pre-trained target LVLM on its own output obtained using prompting, removing both the need for (i) additional external data and (ii) ensembles of external LVLMs. The final data is filtered before training using robust rule-based filtering." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 463, + 481, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 463, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 463, + 481, + 582 + ], + "type": "text", + "content": "Main results. We applied CLIP-DPO fine-tuning on top of two state-of-the-art models: MobileVLM-v2 (3 models in total) and LlaVA-1.5 7B. We find that, in both cases, our approach is effective in reducing hallucinations, outperforming all baseline models (i.e. the models without CLIP-DPO fine-tuning) by a significant margin. Importantly, CLIP-DPO significantly outperforms our direct competitor HA-DPO [83], outperforms Qwen-VL [5] trained on significantly larger datasets (i.e. 1.4B samples for pre-training and 77M for multitask training; as opposed to CLIP-DPO training on just 0.7M samples). Finally, our model's enhanced object grounding capabilities are also illustrated for zero-shot image classification, all without degrading the original performance of the base LVLM model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 602, + 233, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 602, + 233, + 615 + ], + "spans": [ + { + "bbox": [ + 132, + 602, + 233, + 615 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 481, + 666 + ], + "type": "text", + "content": "Large Visual Language Models (LVLMs). Following the unprecedented success of Large Language Models (LLMs) [7, 50, 68, 69, 80], several works have recently proposed to build multi-modal capabilities on top of them [10, 14, 16, 42]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "Y. Ouali et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 481, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 481, + 319 + ], + "type": "text", + "content": "LLaVA [42] and FROMAGe [31] directly pass the visual tokens produced by a pre-trained CLIP [56] vision encoder to an LLM, either fine-tuning the LLM or adapting it using LoRA [24]. Notably, LLaVA training includes a pre-training stage that aligns the CLIP features with the LLM input tokens using a simple projection layer (i.e. keeping the rest of the model frozen) using image captioning data. InstructBLIP [16] uses QFormer [33] to reduce the number of vision tokens before passing them to the LLM. As the quality and distribution of the training data plays a crucial role [84], a series of methods [9,10,55,76] introduced improved data construction pipelines. For example, ShareGPT4v [10] uses the API of GPT-4V [51] to first label, then train a model, and finally use it to re-announce a new training set. Another line of work is improving efficiency by reducing the models' size (original LLaVA models have 7B and 13B parameters). LLaVA-Phi [87], MobileVLM [13] and its follow-up, MobileVLM-v2 [14], replace the LLaMA [68] and Vicuna [12] LLMs with the smaller 1.4B and 2.7B variants of MobileLLaMA and Phi [38]. Our work is orthogonal to the aforementioned methods and does not seek to improve the model's architecture. Instead, we propose CLIP-DPO, an improved training approach for LVLMs based on DPO." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 325, + 481, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 325, + 481, + 623 + ], + "spans": [ + { + "bbox": [ + 133, + 325, + 481, + 623 + ], + "type": "text", + "content": "Preference optimization. Instruction tuning can significantly improve LLMs' perceived output quality and usefulness by aligning their responses to a given task domain or human preferences. This is achieved either by direct fine-tuning on expert data [15, 47, 67] or via reinforcement learning [52, 64, 88]. The latter significantly simplifies the data collection process, but still requires complicated training algorithms based on REINFORCE [73] or PPO [61]. Recently, a much-simplified approach was proposed, Direct Preference optimization (DPO) [57], which bypasses the need to train a reward model and allows direct training using a cross-entropy loss. Multiple improved versions for LLMs have been proposed in the meantime [4, 75, 82]. Following this, a recent wave of works on combining DPO with LVLMs have been proposed [35, 36, 83]. Silkie [35] constructs a multimodal instructional dataset automatically labeled by GPT-4V. Similarly, HA-DPO [83], aiming to reduce the rate of hallucinations, uses a GPT4 model to label and construct positive-negative pairs with and without hallucinations. The work of [36] follows a similar path by using a suite of LLMs and LVLMs (i.e. Gemini-Vision [66]) to generate and label the data. These methods are then primarily evaluated on the LLaVA benchmark and PoPE [72]. In contrast to the aforementioned works, the proposed CLIP-DPO (i) simplifies the pipeline, (ii) removes the need for paid APIs, (iii) removes the need for additional data, and (iv) removes the need for additional external LVLMs. Instead, we make use of a pre-trained CLIP model to rank the generated outputs from a small pool of efficient LVLMs, and show that the proposed CLIP-DPO training provides significant improvements for fixing LVLM hallucinations and, in general, for enhancing the discriminability and robustness of the model as demonstrated by image classification experiments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 629, + 481, + 665 + ], + "type": "text", + "content": "LVLM hallucinations. Broadly speaking, in the context of LVLMs, we consider hallucinations to be incorrect or misleading generated text, contradicting the visual evidence provided by the input image. This is an undesirable characteristic" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLIP-DPO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 223, + 183 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 223, + 183 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 223, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 223, + 183 + ], + "type": "image", + "image_path": "6a9ae7b1a8a89479afbf2abdd702599b518d1594d17601816cb4fba608e505e0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 191, + 482, + 214 + ], + "lines": [ + { + "bbox": [ + 132, + 191, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 191, + 482, + 214 + ], + "type": "text", + "content": "Fig. 1: An example of the injected hallucinations. Given an image and its caption, we prompt GPT-4 to generate 3 types of hallucination: existence, attributes, and relation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 227, + 114, + 468, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 114, + 468, + 129 + ], + "spans": [ + { + "bbox": [ + 227, + 114, + 468, + 129 + ], + "type": "text", + "content": "Original Caption: In this image, we see a black dog wearing an orange vest standing on the grass. The dog's head is down and it appears to be in a backyard." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 225, + 131, + 479, + 182 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 225, + 131, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 131, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 225, + 131, + 479, + 152 + ], + "type": "text", + "content": "- Existence Hallucination: In this image, we see a black dog wearing an orange vest standing on the grass, with a small orange ball near his feet. The dog's head is down and it appears to be in a backyard." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 225, + 153, + 472, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 153, + 472, + 167 + ], + "spans": [ + { + "bbox": [ + 225, + 153, + 472, + 167 + ], + "type": "text", + "content": "- Attributes Hallucination: In this image, we see a dark brown dog wearing an orange vest standing on the grass. The dog's head is down and it appears to be in a backyard." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 225, + 168, + 474, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 168, + 474, + 182 + ], + "spans": [ + { + "bbox": [ + 225, + 168, + 474, + 182 + ], + "type": "text", + "content": "- Relation Hallucination: In this image, we see a black dog wearing an orange vest sitting on the grass. The dog's head is down and it appears to be in a backyard." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 238, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 238, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 482, + 430 + ], + "type": "text", + "content": "inherited from the pre-trained LLM used, and further exacerbated by the visual-language alignment process [5, 25, 29, 40]. Multiple solutions have been recently proposed by the community with varying degrees of success. The works of [25, 77] attempt to address the data bias by constructing better-grounded annotated image-text pairs. The works of [5, 11] scale the resolution of the image encoder, as this was observed to reduce the amount of hallucination, but at the cost of a high increase in the computational cost. The works of [27, 78] improve the vision encoder by adding extra informational paths. The closest work related to ours is that of [83], which, with the help of GPT-4V, constructs negative-positive pairs for DPO fine-tuning. Different from [83], we use a pre-trained CLIP model to perform the ranking, and no additional data or external LVLMs are required. For evaluation purposes, we use AMBER [71], the most comprehensive benchmark for hallucinations to date, encompassing both a generative and discriminative evaluation component. Importantly, unlike all prior benchmarks [22, 30, 37, 40, 65], AMBER is a high-quality dataset fully annotated by humans for both the generative and discriminative tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 449, + 234, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 234, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 234, + 462 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 476, + 432, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 476, + 432, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 476, + 432, + 489 + ], + "type": "text", + "content": "3.1 The Effectiveness of CLIP in Reducing Hallucinations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 498, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 628 + ], + "type": "text", + "content": "First, we probe the effectiveness of CLIP in accurately ranking correct captions over those with hallucinated content. This assessment will indicate the efficacy of CLIP-style models in reducing hallucinations with DPO-based training. To this end, initially, we select 1K captions from the Detailed Caption [1] dataset, chosen for its high-quality labels. Then, we instruct GPT-4 [2] to create three hallucinated captions for each image, corresponding to the following types of hallucination: (i) existence: new elements or objects are added to the caption that were not mentioned originally, (ii) attribute: the attributes, characteristics, or features of the original caption's elements are altered, and (iii) relationship: the spatial, contextual, or interactive relationships of the original caption's elements are altered. See Fig. 1 for an example of the injected hallucinations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "Next, we use LLaVA-1.5 7B [41] to compute the likelihood of all captions, including the original and the hallucinated captions. We then retain only the hallucinated captions for which the model predicts a higher likelihood than the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "text", + "content": "Y. Ouali et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "original, yielding 61 for existence, 76 for attribute, and 119 for relationship hallucinations out of the initial 1K captions. Finally, we compute the CLIP" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 306, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 306, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 306, + 270 + ], + "type": "text", + "content": "image-text scores for these likely yet incorrect captions. As illustrated in Fig.2, CLIP manages to accurately rank the original caption over the hallucinated ones in at least " + }, + { + "bbox": [ + 130, + 140, + 306, + 270 + ], + "type": "inline_equation", + "content": "59\\%" + }, + { + "bbox": [ + 130, + 140, + 306, + 270 + ], + "type": "text", + "content": " of cases, reaching up to " + }, + { + "bbox": [ + 130, + 140, + 306, + 270 + ], + "type": "inline_equation", + "content": "83\\%" + }, + { + "bbox": [ + 130, + 140, + 306, + 270 + ], + "type": "text", + "content": " for the existence type of hallucinations. These findings underscore the potential of using VL embedding models to provide a reliable training signal for hallucination reduction in LVLMs." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 328, + 141, + 466, + 277 + ], + "blocks": [ + { + "bbox": [ + 328, + 141, + 466, + 277 + ], + "lines": [ + { + "bbox": [ + 328, + 141, + 466, + 277 + ], + "spans": [ + { + "bbox": [ + 328, + 141, + 466, + 277 + ], + "type": "image", + "image_path": "fddc432d5c60939e88adc6911e453efe23869fd149a59cfe100c41e56feb3b42.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 285, + 482, + 350 + ], + "lines": [ + { + "bbox": [ + 311, + 285, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 311, + 285, + 482, + 350 + ], + "type": "text", + "content": "Fig. 2: In dark blue, we show the number of hallucinated captions per type that LLaVA-1.5 7B assigns a higher likelihood than the original out of 1K captions. Light blue shows the portion of samples corrected by CLIP." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 290, + 251, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 290, + 251, + 314 + ], + "spans": [ + { + "bbox": [ + 131, + 290, + 251, + 314 + ], + "type": "text", + "content": "3.2 Direct Preference Optimization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 323, + 306, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 306, + 359 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 306, + 359 + ], + "type": "text", + "content": "The standard pipeline for LLM and LVLM alignment consists of three steps: generation, annotation, and op" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "text", + "content": "timization. First, a set of " + }, + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "text", + "content": " prompts " + }, + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "inline_equation", + "content": "\\{x_{i}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "text", + "content": ", each one consisting of an image and a text component in the case of LVLMs, are used to generate a set of response pairs " + }, + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "inline_equation", + "content": "y_{i}^{1}" + }, + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "inline_equation", + "content": "y_{i}^{2}" + }, + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "text", + "content": " obtained from a pool of pre-trained LVLMs for each prompt " + }, + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "text", + "content": ". Then, either human annotators or another set of LLMs (i.e. AI annotators) are used to rank the responses, resulting in a preferred " + }, + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "inline_equation", + "content": "y_{i}^{+}" + }, + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "text", + "content": " and a less preferred response " + }, + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "inline_equation", + "content": "y_{i}^{-}" + }, + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "text", + "content": " for each prompt " + }, + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "text", + "content": ", and thus a final preference dataset " + }, + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(x_i,y_i^+,y_i^-)\\}_{i = 1}^N" + }, + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "text", + "content": ". Finally, Direct Preference Optimization [57] (DPO) can be applied to update the target policy " + }, + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "text", + "content": " directly using the preference dataset " + }, + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 359, + 482, + 478 + ], + "type": "text", + "content": ". Specifically, the DPO optimization objective is defined as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 485, + 481, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 485, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 159, + 485, + 481, + 514 + ], + "type": "interline_equation", + "content": "\\max _ {\\pi_ {\\theta}} \\mathbb {E} _ {(x _ {i}, y _ {i} ^ {+}, y _ {i} ^ {-}) \\sim \\mathcal {D}} \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta} (y _ {i} ^ {+} \\mid x _ {i})}{\\pi_ {\\mathrm {r e f}} (y _ {i} ^ {+} \\mid x _ {i})} - \\beta \\log \\frac {\\pi_ {\\theta} (y _ {i} ^ {-} \\mid x _ {i})}{\\pi_ {\\mathrm {r e f}} (y _ {i} ^ {-} \\mid x _ {i})}\\right), \\tag {1}", + "image_path": "9eba2443cde8449e71dab5221fe54819f71e3d74bf038ab3e3641dab37c45232.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "text", + "content": " as the policy to be learned, " + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ref}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "text", + "content": " as the reference SFT policy, and " + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "text", + "content": " as a hyperparameter to control the Kullback-Leibler (KL) divergence between the learned " + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "text", + "content": " and reference " + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ref}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "text", + "content": " policies [57]. The main benefit of DPO-based optimization is the direct alignment of the LVLM towards the preferences implicit in the preference data " + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "In contrast to previous DPO-based LVLM preference optimization works [35, 36, 83], next, we will introduce the proposed CLIP-DPO, which (i) simplifies the generation process by limiting the pool of LVLMs used for generation to small and efficient models, (ii) removes the need for external LLMs and LVLMs annotators accessed via paid APIs (e.g. GPT-3.5, GPT-4 or GPT-4V) by using CLIP as the ranking model, and (iii) removes the need for additional data by reusing the same data used during the SFT step." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLIP-DPO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 482, + 209 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 482, + 209 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 482, + 209 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 482, + 209 + ], + "type": "image", + "image_path": "f79e40a20a29c0337b7e8f8f3ec82dc49962d47d9d910f6a5b6742c549fd9540.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 216, + 482, + 260 + ], + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 260 + ], + "type": "text", + "content": "Fig. 3: CLIP-DPO. Starting from the initial SFT data pool and a set of prompts, an LVLM generates captions. These captions are first ranked using a CLIP model, then filtered to identify the most suitable positive and negative pairs for DPO-based optimization." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 281, + 202, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 202, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 202, + 295 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 306, + 479, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 479, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 479, + 366 + ], + "type": "text", + "content": "In this section, we introduce our DPO variant for LVLMs preference optimization, named CLIP-DPO. As illustrated in Fig. 3, CLIP-DPO follows a three-step process: generation, annotation, and optimization, similar to the standard DPO-based training pipeline. While the final DPO-based optimization step remains unchanged, CLIP-DPO modifies the first two steps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 366, + 480, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 480, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 480, + 462 + ], + "type": "text", + "content": "Firstly, we streamline the generation step by utilizing the same data pool (i.e. images only) as the SFT stage and generate data using either the model itself or a small pool of efficient LVLMs. Secondly, we introduce an annotation step tailored for CLIP-style embedding models, where we rank captions based on their image-text similarities. Then, given that the resulting captions and pairs may vary in quality, we implement a rule-based filtering strategy to remove unsuitable candidates. Finally, with the constructed preference data, the model undergoes DPO-based optimization using Eq. 1, as shown in Fig. 3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 481, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 481, + 544 + ], + "type": "text", + "content": "As such, by leveraging VL embedding models for data annotation, we tap into the extensive information captured by these contrastively-pretrained models, which have been exposed to hundreds of millions of unique image-text pairs during training, unlike existing open-sourced LVLMs. Additionally, unlike the discrete GPT-4V ranking used in prior work, CLIP scores are continuous, enabling finer-grained comparisons. This allows us to assess the difficulty of a given pair based on the margin between the score values." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 481, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 571 + ], + "type": "text", + "content": "Next, we will elaborate on each module of our pipeline. Sec. 4.1 details the data generation module, and Sec. 4.2 details the data annotation process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 586, + 246, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 246, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 246, + 597 + ], + "type": "text", + "content": "4.1 Data Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "The first step of our pipeline consists of generating a set of per-image captions that will be later ranked by a pre-trained CLIP model, along with a subsequent filtering approach to select a set of positive and negative pairs, which are then used for DPO-based training. To this end, we start by selecting the pool of LVLM annotators and the data to be annotated. For the annotators and to reduce the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "text", + "content": "Y. Ouali et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "cost of the generation step, we select MobileVLM-v2 [14] family of models, given their efficiency and performance. As for the data, and to avoid introducing any new sources that might bias the results of CLIP-DPO, and to further reduce the method's cost, we opt for the initial pool of data used during the SFT stage of MobileVLM-v2 [14] models, see Table 1 for details. Then, we conduct our data generation pipeline that consists of two steps, the generation of generic captions and per-image questions and answers." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 143, + 210, + 289, + 323 + ], + "blocks": [ + { + "bbox": [ + 143, + 210, + 289, + 323 + ], + "lines": [ + { + "bbox": [ + 143, + 210, + 289, + 323 + ], + "spans": [ + { + "bbox": [ + 143, + 210, + 289, + 323 + ], + "type": "table", + "html": "
DatasetDataset Size
COCO118K
GQA72K
LLaVA-1.0 Pretraining595K
LLaVA-1.5 Pretraining558K
OCR VQA80K
SAM570K
SBU845K
TextVQA22K
VG86K
Web-Celebrity495
Web-Landmark500
Wikiart500
Total2.9M
", + "image_path": "80e7e79683e7a8556f04ad02adead17f0385bce5093d4a7184e992e8f741389d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 331, + 301, + 353 + ], + "lines": [ + { + "bbox": [ + 131, + 331, + 301, + 353 + ], + "spans": [ + { + "bbox": [ + 131, + 331, + 301, + 353 + ], + "type": "text", + "content": "Table 1: Data pool used for data generation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 328, + 204, + 459, + 322 + ], + "blocks": [ + { + "bbox": [ + 328, + 204, + 459, + 322 + ], + "lines": [ + { + "bbox": [ + 328, + 204, + 459, + 322 + ], + "spans": [ + { + "bbox": [ + 328, + 204, + 459, + 322 + ], + "type": "image", + "image_path": "05ffd1475c796d2f88ab645289a33b8e1a495813251ed3b521f3a2cd761937b5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 331, + 477, + 354 + ], + "lines": [ + { + "bbox": [ + 307, + 331, + 477, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 477, + 354 + ], + "type": "text", + "content": "Fig. 4: The distribution of CLIP image-text scores per category." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 133, + 376, + 207, + 449 + ], + "blocks": [ + { + "bbox": [ + 133, + 376, + 207, + 449 + ], + "lines": [ + { + "bbox": [ + 133, + 376, + 207, + 449 + ], + "spans": [ + { + "bbox": [ + 133, + 376, + 207, + 449 + ], + "type": "image", + "image_path": "2d257b53e0d3d9670fa31a085752a1c82d358b45ef8218a48a098507e4cc0800.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 483, + 482, + 517 + ], + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 517 + ], + "type": "text", + "content": "Fig. 5: Examples of generated generic captions produced by MobileVLM-v2 1.7B and MobileVLM-v2 3B. For MobileVLM-v2 3B's generated captions, we also show the produced question and positive and negative answers generated using Mistral 7B." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 212, + 378, + 258, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 378, + 258, + 384 + ], + "spans": [ + { + "bbox": [ + 212, + 378, + 258, + 384 + ], + "type": "text", + "content": "MobileVLMv-2 1.7b:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 216, + 385, + 477, + 411 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 216, + 385, + 451, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 385, + 451, + 403 + ], + "spans": [ + { + "bbox": [ + 216, + 385, + 451, + 403 + ], + "type": "text", + "content": "- The image showcases a charming two-story building, characterized by its steep roof and a balcony adorned with red flowers. The building's facade is painted white, accented with a green roof that adds a touch of color to the scene." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 216, + 403, + 477, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 403, + 477, + 411 + ], + "spans": [ + { + "bbox": [ + 216, + 403, + 477, + 411 + ], + "type": "text", + "content": "The image captures a charming scene of a German half-timbered building, bathed in the soft glow of daylight." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 213, + 414, + 252, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 414, + 252, + 419 + ], + "spans": [ + { + "bbox": [ + 213, + 414, + 252, + 419 + ], + "type": "text", + "content": "MobileVLMv-2 3b:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 216, + 421, + 456, + 448 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 216, + 421, + 456, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 421, + 456, + 439 + ], + "spans": [ + { + "bbox": [ + 216, + 421, + 456, + 439 + ], + "type": "text", + "content": "- The image showcases a traditional German half-timbered house, a common architectural feature in the region. The house is adorned with a green roof, a characteristic of half-timbered houses, and is embellished with red flowers that add a vibrant touch to the overall design." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 216, + 441, + 380, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 441, + 380, + 448 + ], + "spans": [ + { + "bbox": [ + 216, + 441, + 380, + 448 + ], + "type": "text", + "content": "- A building with a statue of a knight in the middle of the building." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 213, + 453, + 242, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 453, + 242, + 458 + ], + "spans": [ + { + "bbox": [ + 213, + 453, + 242, + 458 + ], + "type": "text", + "content": "Mistral 7b:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 216, + 459, + 343, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 459, + 343, + 472 + ], + "spans": [ + { + "bbox": [ + 216, + 459, + 343, + 472 + ], + "type": "text", + "content": "- What color is the roof of the building in the image?\n- Positive: Green Negative: Belge" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 534, + 482, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 631 + ], + "type": "text", + "content": "Generation of Generic Captions. We start by generating a set of 5 descriptive captions per image. For each of the MobileVLM-v2 models, we prompt with 5 different prompts (e.g. \"identify the setting and note any characters or objects, focusing on visible details.\") to increase the diversity of the generated captions. While these captions can be used for CLIP ranking and DPO-based training, they are still produced by generic prompts that are not image-specific. Next, we perform the second step of our data generation pipeline to obtain a set of per-image questions and answers." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Generation of Per-Image Questions and Answers: To obtain a set of questions per image, we leverage an LLM, i.e. a variant of Mistral 7B Instruct" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLIP-DPO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 228, + 182 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 228, + 182 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 228, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 228, + 182 + ], + "type": "image", + "image_path": "5e879f6d006e00e68966d810102b196309f7bf5bc56fd93a4cd95bd2db17d628.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 190, + 481, + 245 + ], + "lines": [ + { + "bbox": [ + 130, + 190, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 481, + 245 + ], + "type": "text", + "content": "Fig. 6: Question Filtering. The sample is rejected if the CLIP image-question score is low for a given sample. If the question is relevant and not generic, it is kept. Next, we evaluate the quality of the answer by first parsing the question into a caption, appending the positive answer to it, and computing its CLIP score. If it is a high score, the sample is kept for DPO-based training." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 232, + 114, + 342, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 114, + 342, + 129 + ], + "spans": [ + { + "bbox": [ + 232, + 114, + 342, + 129 + ], + "type": "text", + "content": "Question 1: What is the central focus of the image?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 235, + 130, + 314, + 144 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 235, + 130, + 314, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 130, + 314, + 137 + ], + "spans": [ + { + "bbox": [ + 235, + 130, + 314, + 137 + ], + "type": "text", + "content": "- Question CLIP Score: 15.0" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 235, + 138, + 314, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 138, + 314, + 144 + ], + "spans": [ + { + "bbox": [ + 235, + 138, + 314, + 144 + ], + "type": "text", + "content": "- Question Filter: Rejected" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 355, + 114, + 457, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 114, + 457, + 129 + ], + "spans": [ + { + "bbox": [ + 355, + 114, + 457, + 129 + ], + "type": "text", + "content": "Question 2: What type of market is depicted in the image?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 358, + 131, + 436, + 144 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 358, + 131, + 436, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 131, + 436, + 137 + ], + "spans": [ + { + "bbox": [ + 358, + 131, + 436, + 137 + ], + "type": "text", + "content": "- Question CLIP Score: 26.7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 358, + 138, + 425, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 138, + 425, + 144 + ], + "spans": [ + { + "bbox": [ + 358, + 138, + 425, + 144 + ], + "type": "text", + "content": "- Question Filter: Kept" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 232, + 151, + 347, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 151, + 347, + 158 + ], + "spans": [ + { + "bbox": [ + 232, + 151, + 347, + 158 + ], + "type": "text", + "content": "Positive: A fruit and vegetable market." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 234, + 159, + 466, + 180 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 234, + 159, + 466, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 159, + 466, + 166 + ], + "spans": [ + { + "bbox": [ + 234, + 159, + 466, + 166 + ], + "type": "text", + "content": "- Caption: type of market is depicted in the image a fruit and vegetable market." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 235, + 167, + 313, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 167, + 313, + 173 + ], + "spans": [ + { + "bbox": [ + 235, + 167, + 313, + 173 + ], + "type": "text", + "content": "Positive CLIP Score: 29.1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 235, + 174, + 302, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 174, + 302, + 180 + ], + "spans": [ + { + "bbox": [ + 235, + 174, + 302, + 180 + ], + "type": "text", + "content": "- Question Filter: Kept" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 133, + 257, + 223, + 322 + ], + "blocks": [ + { + "bbox": [ + 133, + 257, + 223, + 322 + ], + "lines": [ + { + "bbox": [ + 133, + 257, + 223, + 322 + ], + "spans": [ + { + "bbox": [ + 133, + 257, + 223, + 322 + ], + "type": "image", + "image_path": "921def778516fde821386dde71ff3277789985a0bd5f78379e894c697c348963.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 372, + 481, + 395 + ], + "lines": [ + { + "bbox": [ + 131, + 372, + 481, + 395 + ], + "spans": [ + { + "bbox": [ + 131, + 372, + 481, + 395 + ], + "type": "text", + "content": "Fig. 7: Examples of the final positive and negative pairs obtained post-filtering, to be used for DPO-based training." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 325, + 245, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 325, + 245, + 338 + ], + "spans": [ + { + "bbox": [ + 133, + 325, + 245, + 338 + ], + "type": "text", + "content": "Prompt: Provide a brief description of the given image." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 338, + 242, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 338, + 242, + 350 + ], + "spans": [ + { + "bbox": [ + 133, + 338, + 242, + 350 + ], + "type": "text", + "content": "Positive: An Apple computer with a laptop computer sitting on a desk." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 350, + 242, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 350, + 242, + 363 + ], + "spans": [ + { + "bbox": [ + 133, + 350, + 242, + 363 + ], + "type": "text", + "content": "Negative: A laptop and a desktop computer are sitting on a desk." + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 254, + 257, + 347, + 322 + ], + "blocks": [ + { + "bbox": [ + 254, + 257, + 347, + 322 + ], + "lines": [ + { + "bbox": [ + 254, + 257, + 347, + 322 + ], + "spans": [ + { + "bbox": [ + 254, + 257, + 347, + 322 + ], + "type": "image", + "image_path": "48e92c389da54ece0c3757c592c986106ec6927b53055777c6054d1d5c245744.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 253, + 325, + 366, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 325, + 366, + 338 + ], + "spans": [ + { + "bbox": [ + 253, + 325, + 366, + 338 + ], + "type": "text", + "content": "Prompt: Provide a depiction of the picture. Positive: A man in orange shirt kicking a" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 253, + 339, + 283, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 339, + 283, + 345 + ], + "spans": [ + { + "bbox": [ + 253, + 339, + 283, + 345 + ], + "type": "text", + "content": "soccer ball" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 253, + 345, + 351, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 345, + 351, + 351 + ], + "spans": [ + { + "bbox": [ + 253, + 345, + 351, + 351 + ], + "type": "text", + "content": "Negative: A group of people kicking a" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 253, + 351, + 285, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 351, + 285, + 357 + ], + "spans": [ + { + "bbox": [ + 253, + 351, + 285, + 357 + ], + "type": "text", + "content": "soccer ball" + } + ] + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 374, + 257, + 472, + 322 + ], + "blocks": [ + { + "bbox": [ + 374, + 257, + 472, + 322 + ], + "lines": [ + { + "bbox": [ + 374, + 257, + 472, + 322 + ], + "spans": [ + { + "bbox": [ + 374, + 257, + 472, + 322 + ], + "type": "image", + "image_path": "5aa0d5905cac1a7c42a519100dad450e5d7d57fbbfec85b4c81a9b797eea8476.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 373, + 325, + 472, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 325, + 472, + 338 + ], + "spans": [ + { + "bbox": [ + 373, + 325, + 472, + 338 + ], + "type": "text", + "content": "Prompt: What is the name of the paint brand featured on the bucket?" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 374, + 338, + 470, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 338, + 470, + 345 + ], + "spans": [ + { + "bbox": [ + 374, + 338, + 470, + 345 + ], + "type": "text", + "content": "Positive: Homer's All Purpose Bucket" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 374, + 345, + 479, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 345, + 479, + 350 + ], + "spans": [ + { + "bbox": [ + 374, + 345, + 479, + 350 + ], + "type": "text", + "content": "Negative: Home Depot All-Purpose Bucket." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 130, + 418, + 481, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 481, + 468 + ], + "type": "text", + "content": "v0.2 [28]" + }, + { + "bbox": [ + 130, + 418, + 481, + 468 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 130, + 418, + 481, + 468 + ], + "type": "text", + "content": ", and feed it with the generated captions and prompt it to generate 2 questions for each image, together with positive and negative answers. The LLM is asked to generate the positive ones based on the fed captions and to generate plausible but incorrect negatives given the image description/caption." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 131, + 475, + 471, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 475, + 471, + 487 + ], + "spans": [ + { + "bbox": [ + 131, + 475, + 471, + 487 + ], + "type": "text", + "content": "For an example of the generated captions, questions, and answers, see Fig. 5." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 131, + 506, + 246, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 506, + 246, + 517 + ], + "spans": [ + { + "bbox": [ + 131, + 506, + 246, + 517 + ], + "type": "text", + "content": "4.2 Data Annotation" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 130, + 529, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 482, + 601 + ], + "type": "text", + "content": "Starting with the generated captions, the subsequent data annotation step involves ranking them based on their CLIP image-text similarities and filtering them to obtain a final set of high-quality positive-negative pairs for DPO-based training. Our filtering pipeline includes two stages: global filtering, where we eliminate low-quality images and captions, and pair-filtering, where we select the best positive and negative per-image pairs for DPO-based training." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 130, + 609, + 481, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 609, + 481, + 645 + ], + "spans": [ + { + "bbox": [ + 130, + 609, + 481, + 645 + ], + "type": "text", + "content": "CLIP Ranking. The initial step of data annotation is CLIP ranking. In this step, we use a pre-trained CLIP model to compute the cosine similarities of the captions with their associated images and then sort them from highest to lowest." + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 101 + ], + "type": "text", + "content": "Y. Ouali et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 652, + 405, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 652, + 405, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 652, + 405, + 665 + ], + "type": "text", + "content": "1 https://huggingface.co/teknium/OpenHermes-2.5-Mistral7B" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 482, + 295 + ], + "type": "text", + "content": "Global Filtering. Next, we aim to filter our data and only retain a balanced and higher-quality subset. To achieve this, we start by analyzing the types of images in our data pool (see Tab. 1). We define a set of four generic categories: images of text, people, objects, and scenes. Using CLIP and a set of 10 descriptions per category, we create four class prototypes. Each image is then assigned to one of these prototypes based on the highest cosine similarity. As illustrated in Fig. 4, we observe that CLIP scores for the text category are the highest, while scores for people, objects, and scenes are relatively similar. Consequently, before applying CLIP-based filtering, we down-sample the portion of text images. We then filter out all generic captions below a given CLIP score (" + }, + { + "bbox": [ + 133, + 115, + 482, + 295 + ], + "type": "inline_equation", + "content": "e.g. < 28.0" + }, + { + "bbox": [ + 133, + 115, + 482, + 295 + ], + "type": "text", + "content": "). Additionally, since CLIP is primarily trained on short text, we remove long captions to ensure more precise CLIP scores. For questions, we compute their CLIP scores and remove all questions with low CLIP scores (" + }, + { + "bbox": [ + 133, + 115, + 482, + 295 + ], + "type": "inline_equation", + "content": "e.g. < 25.0" + }, + { + "bbox": [ + 133, + 115, + 482, + 295 + ], + "type": "text", + "content": "), thereby eliminating generic questions (" + }, + { + "bbox": [ + 133, + 115, + 482, + 295 + ], + "type": "inline_equation", + "content": "e.g. \"what is the main object in the image?\")" + }, + { + "bbox": [ + 133, + 115, + 482, + 295 + ], + "type": "text", + "content": " that are already covered by the generated captions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 302, + 482, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 302, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 133, + 302, + 482, + 470 + ], + "type": "text", + "content": "Pair Filtering. Finally, starting from the remaining high-quality image captions and image question-answers, the final step will be selecting a set of positive and negative pairs for DPO-based training. For the questions, since we already have a set of positive and negative responses generated by our LLM, we only need to filter out the low-quality pairs. To achieve this, we use simple regex matching rules to extract an image description from the question, append the positive answer to it to create a synthetic caption, and compute its CLIP image-text scores. We then reject examples where the scores are low. See Fig. 6 for an example of the questions filtering process. As for the captions, we consider all possible pairs where the CLIP score difference between two captions is larger than a given threshold (" + }, + { + "bbox": [ + 133, + 302, + 482, + 470 + ], + "type": "inline_equation", + "content": "e.g. > 2.0" + }, + { + "bbox": [ + 133, + 302, + 482, + 470 + ], + "type": "text", + "content": ") and where the length of the two captions is similar to avoid introducing false preferences. We then order them based on the CLIP score difference between the positive and negative captions and select the top-ranking pair per image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 477, + 481, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 477, + 481, + 513 + ], + "spans": [ + { + "bbox": [ + 133, + 477, + 481, + 513 + ], + "type": "text", + "content": "After the data annotation step, the resulting DPO training data consists of 750K pairs, of which 50K are question-answer pairs and the rest are caption pairs. For examples of the final pairs, see Fig. 7." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 532, + 290, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 532, + 290, + 545 + ], + "spans": [ + { + "bbox": [ + 133, + 532, + 290, + 545 + ], + "type": "text", + "content": "5 Implementation details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 557, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 557, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 557, + 482, + 665 + ], + "type": "text", + "content": "Network architecture: In this work, we consider two LVLMs architectures: MobileVLM-v2 [14] and LLaVA-1.5 [42]. Both models follow the same overall structure: a pre-trained CLIP vision encoder and a pre-trained LLM. The visual tokens produced by the frozen vision encoder are projected using either a linear layer or a small projection module and passed as input to the LLM. LLaVA opts for a pre-trained Vicuna LLM [12] while MobileVLM-v2 uses MobileLLaMA [13], except for their 7B variant, which also uses a Vicuna model. Both use the same ViT-L-14 @ 336px CLIP visual encoder [56]. For efficiency purposes, MobileVLM-v2 uses a projection module that halves the number of" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLIP-DPO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "visual tokens. We consider the following model variants in our comparisons: MobileVLM-v2 (1.7B, 3B and 7B), and LLaVA-1.5 7B. For data annotation, we use ViT-H/14 DFN [19] as our pre-trained CLIP model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 159, + 479, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 159, + 479, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 159, + 479, + 278 + ], + "type": "text", + "content": "Training details: For all of our experiments, unless specified otherwise, we start from the pre-trained MobileVLM-v2 and LLaVA-1.5 models. The models are then fine-tuned for 1 epoch using DPO-based optimization on the constructed CLIP-DPO dataset. We use the following hyperparameters for fine-tuning the models: AdamW optimizer [45] with a batch size of 256, a learning rate of " + }, + { + "bbox": [ + 130, + 159, + 479, + 278 + ], + "type": "inline_equation", + "content": "5e - 7" + }, + { + "bbox": [ + 130, + 159, + 479, + 278 + ], + "type": "text", + "content": ", decreased to 0 using a cosine scheduler, a warm-up of 0.01, and a weight decay set to 0. Both during training and testing, the input images are cropped and resized to " + }, + { + "bbox": [ + 130, + 159, + 479, + 278 + ], + "type": "inline_equation", + "content": "336 \\times 336" + }, + { + "bbox": [ + 130, + 159, + 479, + 278 + ], + "type": "text", + "content": "px. The training was performed on 8 A100 GPUs using Pytorch [54]. For the larger models, e.g. LLaVA-1.5 7B, to fit them in memory, we use the Zero-3 strategy [58, 59] and decrease the batch size to 64." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 293, + 200, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 293, + 200, + 306 + ], + "spans": [ + { + "bbox": [ + 131, + 293, + 200, + 306 + ], + "type": "text", + "content": "6 Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 315, + 479, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 479, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 479, + 387 + ], + "type": "text", + "content": "We first evaluate the effectiveness of CLIP-DPO in reducing hallucinations on the recently introduced AMBER [71] benchmark in Sec. 6.1. Moreover, we demonstrate its enhanced grounding capabilities by reporting zero-shot image classification in Sec. 6.2. Finally, in Sec. 6.4, we show that the proposed CLIP-DPO training does not compromise LVLM performance by reporting results on the LLaVA benchmark [87]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 392, + 304, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 392, + 304, + 403 + ], + "spans": [ + { + "bbox": [ + 131, + 392, + 304, + 403 + ], + "type": "text", + "content": "6.1 Evaluation on hallucinations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 407, + 480, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 480, + 611 + ], + "type": "text", + "content": "We start by evaluating CLIP-DPO in terms of LVLM hallucination reduction, which is our main objective. We use AMBER [71], a comprehensive, high-quality, and LLM-free multidimensional benchmark for LVLM hallucination evaluation, which can be used to evaluate both generative and discriminative tasks. As the results from Tab. 2 show, MobileVLM-v2 trained with CLIP-DPO improves upon MobileVLM-v2 baselines across all model sizes, and sometimes quite significantly, especially for the 1.7B and 7B models. It can also be seen that CLIP-DPO significantly improves when applied on top of LLaVA-1.5 7B. Importantly, CLIP-DPO significantly outperforms HA-DPO [83], the main competing approach, improving the AMBER score 3.2 vs 7.8 when using LLaVA-1.5. Finally, our LlaVA-1.5 7B+CLIP-DPO even outperforms Qwen-VL [5], which is trained on significantly larger datasets (1.4B image-text pairs for pre-training and 77M for multitask training while CLIP-DPO is fine-tuned on just 0.7M samples), and even matches the performance of GPT-4V without using any GPT-4V model outputs during training. Overall, these results on a high-quality state-of-the-art benchmark such as AMBER clearly demonstrate the effectiveness of our CLIP-DPO approach for reducing hallucinations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 625, + 309, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 625, + 309, + 637 + ], + "spans": [ + { + "bbox": [ + 131, + 625, + 309, + 637 + ], + "type": "text", + "content": "6.2 Zero-shot image classification" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "A primary reason for the hallucinatory behavior of LVLMs is the weak alignment between their visual features and the input LLM tokens. A direct way to" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "Y. Ouali et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 152, + 114, + 463, + 282 + ], + "blocks": [ + { + "bbox": [ + 152, + 114, + 463, + 282 + ], + "lines": [ + { + "bbox": [ + 152, + 114, + 463, + 282 + ], + "spans": [ + { + "bbox": [ + 152, + 114, + 463, + 282 + ], + "type": "table", + "html": "
ModelGENERATIVE TASKDISCRIMINATIVE TASKAMBER
CHAIR↓Cover↑Hal↓Cog↓Acc.P.R.F1
mPLUG-Owl21.650.176.111.540.192.810.518.948.7
LLaVA11.551.048.85.542.774.121.032.760.6
MiniGPT-413.663.065.311.363.690.550.464.775.6
CogVLM5.657.223.61.369.088.960.972.383.4
mPLUG-Owl210.652.039.94.575.695.066.978.584.0
InstructBLIP8.852.238.24.476.584.579.081.786.5
Qwen-VL5.549.423.61.981.290.879.784.989.7
GPT-4V4.667.130.72.683.484.990.187.491.4
LLaVA-1.5 7B7.851.036.44.272.093.262.474.783.5
+ HA-DPO7.233.619.72.668.368.198.480.586.7
+ CLIP-DPO3.747.816.61.377.884.481.582.989.6
MobileVLM-v2 1.7B3.839.68.90.565.492.651.966.581.4
+ CLIP-DPO4.238.910.80.571.288.764.874.985.3
MobileVLM-v2 3B4.838.711.10.773.592.165.776.786.0
+ CLIP-DPO4.741.513.10.576.791.071.880.387.8
MobileVLM-v2 7B4.438.910.40.671.995.060.874.184.9
+ CLIP-DPO4.038.010.10.577.393.470.880.588.3
", + "image_path": "ec447d08746b98cf77c922183dbeff054dc61052774bf3533f2f78433609fe61.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 132, + 348, + 479, + 441 + ], + "blocks": [ + { + "bbox": [ + 132, + 291, + 480, + 335 + ], + "lines": [ + { + "bbox": [ + 132, + 291, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 291, + 480, + 335 + ], + "type": "text", + "content": "Table 2: Hallucination evaluation results on AMBER for both generative (leftmost set of columns) and discriminative tasks (rightmost set of columns). Our approach offers consistent and large performance improvements across 4 different LVLM models, and also beats HA-DPO by a large margin." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 348, + 479, + 441 + ], + "lines": [ + { + "bbox": [ + 132, + 348, + 479, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 479, + 441 + ], + "type": "table", + "html": "
MethodStanfordCarsOxfordPetsOxfordFlowersImagenetFood-101EurosatCaltech-101UCF-101SUN397Avg
LLaVA-1.523.234.07.337.945.349.582.750.543.140.0
+ HA-DPO23.233.47.136.945.149.482.750.342.739.7
+ CLIP-DPO30.144.816.342.253.952.184.853.148.847.4
MobileVLM-v2 1.7B17.115.214.532.631.249.377.945.539.333.9
+ CLIP-DPO19.232.623.841.347.448.980.850.249.343.7
MobileVLM-v2 3B14.823.59.135.738.853.281.848.642.036.7
+ CLIP-DPO28.640.919.544.352.056.485.552.350.147.7
MobileVLM-v2 7B27.132.511.537.745.043.181.949.046.641.6
+ CLIP-DPO32.551.335.850.162.659.388.356.254.854.5
", + "image_path": "dface6c72bb5620056fa4a5798fbac4221056d6c2bcc91a021d2ddc30e73e216.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "type": "text", + "content": "Table 3: Zero-shot image recognition results in terms of Top-1 accuracy (\\%). While HA-DPO shows very similar performance to the base model, our CLIP-DPO improves the base model by a very large margin in all cases." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 510, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 481, + 666 + ], + "type": "text", + "content": "evaluate this is through simple zero-shot image classification, which is the goto benchmark for contrastively trained VL models like CLIP [56]. The typical setup follows a closed-set classification problem, where the names of all possible classes are known a priori and are encoded into class prototypes using CLIP. A given image is then assigned to the class with the highest CLIP image-text scores. Herein, we follow the same protocol with the notable difference that, as our goal is LVLM evaluation, we first prompt the model to generate a free-form image caption describing the main object in the image, then encode the caption using the CLIP text encoder and assign the image to the class with the highest text-text (i.e. caption-class) CLIP score. Here, we opted for a different family of VL embedding model, i.e. SigLIP [79], to avoid evaluating using CLIP models similar to those used during the data annotation step. Following [56,85], we evaluate our approach on a suite of 9 diverse datasets: UCF-101 [63], SUN397 [74]," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLIP-DPO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 114, + 471, + 334 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 471, + 334 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 471, + 334 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 471, + 334 + ], + "type": "table", + "html": "
MethodLLMRes.GQA\\( SQAI \\)\\( VQAT \\)POPE\\( MME^P \\)\\( MMB^{dev} \\)Avg.
BLIP-2 [33]Vicuna-13B22441.061.042.585.31293.8--
InstructBLIP [16]Vicuna-13B22449.563.150.778.91212.8--
Shikra [9]Vicuna-13B224-----58.8-
Openflamingo [3]MPT 7B336--33.6--4.6-
Qwen-VL [5]Qwen 7B44859.367.163.8-1487.638.2-
mPLUG-Owl [76]LLaMA 7B224----967.349.4-
MiniGPT-v2 [8]LLaMA 7B44860.3----12.2-
MiniGPT-4 [86]Vicuna 7B22432.2---581.723.0-
InstructBLIP [16]Vicuna 7B22449.260.550.1--36.0-
ShareGPT4V [10]Vicuna 7B33663.368.460.485.71567.468.870.8
MoE-LLaVA-1.6B×4 [39]StableLM-1.6B33660.462.647.884.31300.8*59.463.3
MoE-LLaVA-2.7B×4 [39]Phi-2.7B33661.168.750.285.01396.4*65.566.7
LLaVA-1.5 [41]Vicuna 7B33662.066.858.285.91510.764.368.8
+ HA-DPOVicuna 7B33661.969.258.384.31505.664.969.0
+ CLIP-DPOVicuna 7B33659.367.656.485.81468.764.967.9
MobileVLM 1.7B [13]MobileLLaMA 1.4B33656.157.341.584.51196.253.258.7
MobileVLM-v2 1.7B [14]MobileLLaMA 1.4B33658.366.752.184.31302.857.764.2
+ CLIP-DPOMobileLLaMA 1.4B33658.668.447.884.31331.157.663.8
MobileVLM- 3B [13]MobileLLaMA 2.7B33659.061.247.584.91288.959.662.8
MobileVLM-v2 3B [14]MobileLLaMA 2.7B33661.170.057.584.71440.563.268.1
+ CLIP-DPOMobileLLaMA 2.7B33660.972.357.385.11425.763.868.4
MobileVLM-v2 [14] 7BVicuna 7B33662.674.862.385.31560.769.272.1
+ CLIP-DPO 7BVicuna 7B33662.577.362.284.71539.170.472.3
", + "image_path": "050437536d2d8c1dd5c52fc333629b4356b6d39f1bd11558b55e5ad356be51ea.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 342, + 482, + 375 + ], + "lines": [ + { + "bbox": [ + 132, + 342, + 482, + 375 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 482, + 375 + ], + "type": "text", + "content": "Table 4: Results on several vision language tasks from the LlaVA benchmark. Our approach is combined with LVLMs of different sizes, showing consistent performance in all cases." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "content": "Stanford Cars [32], Oxford Pets [53], Oxford flowers [49], ImageNet [17], Food 101 [6], Eurosat [23] and Caltech-101 [20]. As the results from Tab. 3 show, all LVLMs fine-tuned with CLIP-DPO significantly outperform their corresponding baselines, showcasing the increased discriminative properties. Note again that these improvements are obtained without affecting the model's performance on other tasks and datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 483, + 273, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 273, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 273, + 495 + ], + "type": "text", + "content": "6.3 Additional evaluations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 501, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 609 + ], + "type": "text", + "content": "Herein, we evaluate the impact of CLIP-DPO training on other vision language tasks and, more specifically, on the popular LLaVA-Bench (GQA [26], ScienceQA [46], TextVQA [62], MME [21], MMBench [44]). As Tab. 4 shows, overall, CLIP-DPO training does not compromise performance. We further note that accuracy improvements on LLaVA-Bench are heavily tied to the addition of extra training data [10] or architectural changes [13,14,39,41]. As we are not using any additional data & models nor making architectural changes, it is not surprising that the performance after CLIP-DPO training remains largely in line with that of the original baseline model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 624, + 257, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 257, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 257, + 635 + ], + "type": "text", + "content": "6.4 Qualitative Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "In Figure 8, we show some qualitative examples comparing the predictions of LLaVA-1.5 7B, LLaVA-1.5 7B+HA-DPO, and LLaVA-1.5 7B+CLIP-DPO. Over" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "Y. Ouali et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 150, + 114, + 220, + 161 + ], + "blocks": [ + { + "bbox": [ + 150, + 114, + 220, + 161 + ], + "lines": [ + { + "bbox": [ + 150, + 114, + 220, + 161 + ], + "spans": [ + { + "bbox": [ + 150, + 114, + 220, + 161 + ], + "type": "image", + "image_path": "2a6af99c9dfdbc1d2d4c156f2e4f9ea6ff9fc41ff05667cc675777cddd9fef41.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 228, + 118, + 342, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 118, + 342, + 126 + ], + "spans": [ + { + "bbox": [ + 228, + 118, + 342, + 126 + ], + "type": "text", + "content": "Provide a short description of the image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 227, + 129, + 457, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 129, + 457, + 149 + ], + "spans": [ + { + "bbox": [ + 227, + 129, + 457, + 149 + ], + "type": "text", + "content": "LLava-1.5-7b: A large silver airplane is parked inside a hangar. The plane is positioned on a runway, and there is a ladder leading up to the cockpit. The hangar appears to be a large building, possibly a factory or a maintenance facility." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 231, + 151, + 448, + 175 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 232, + 151, + 399, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 151, + 399, + 159 + ], + "spans": [ + { + "bbox": [ + 232, + 151, + 399, + 159 + ], + "type": "text", + "content": "+ HA-DPO: A large silver airplane is parked inside a hangar." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 231, + 161, + 448, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 161, + 448, + 175 + ], + "spans": [ + { + "bbox": [ + 231, + 161, + 448, + 175 + ], + "type": "text", + "content": "+ CLIP-DPO: A fighter jet is parked inside a hangar, with a ladder attached to it. The jet is numbered 157." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 150, + 185, + 220, + 232 + ], + "blocks": [ + { + "bbox": [ + 150, + 185, + 220, + 232 + ], + "lines": [ + { + "bbox": [ + 150, + 185, + 220, + 232 + ], + "spans": [ + { + "bbox": [ + 150, + 185, + 220, + 232 + ], + "type": "image", + "image_path": "b05df5ad4610a422e25687c4927231823f6d240d4be8c0a3e0c1692675623e15.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 228, + 189, + 342, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 189, + 342, + 197 + ], + "spans": [ + { + "bbox": [ + 228, + 189, + 342, + 197 + ], + "type": "text", + "content": "Provide a short description of the image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 228, + 198, + 441, + 226 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 228, + 198, + 441, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 198, + 441, + 206 + ], + "spans": [ + { + "bbox": [ + 228, + 198, + 441, + 206 + ], + "type": "text", + "content": "LLava-1.5-7b: A green bus is parked in a parking lot, waiting for passengers." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 228, + 209, + 429, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 209, + 429, + 217 + ], + "spans": [ + { + "bbox": [ + 228, + 209, + 429, + 217 + ], + "type": "text", + "content": "+ HA-DPO: A green bus is parked in a parking lot, waiting for passengers." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 228, + 218, + 371, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 218, + 371, + 226 + ], + "spans": [ + { + "bbox": [ + 228, + 218, + 371, + 226 + ], + "type": "text", + "content": "+CLIP-DPO: A green Go3 bus is parked at a bus stop." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "type": "table", + "bbox": [ + 184, + 262, + 430, + 344 + ], + "blocks": [ + { + "bbox": [ + 142, + 239, + 470, + 251 + ], + "lines": [ + { + "bbox": [ + 142, + 239, + 470, + 251 + ], + "spans": [ + { + "bbox": [ + 142, + 239, + 470, + 251 + ], + "type": "text", + "content": "Fig. 8: Qualitative examples comparing our method with LLaVA and HA-DPO." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 184, + 262, + 430, + 344 + ], + "lines": [ + { + "bbox": [ + 184, + 262, + 430, + 344 + ], + "spans": [ + { + "bbox": [ + 184, + 262, + 430, + 344 + ], + "type": "table", + "html": "
Loss variantAMBER ScoreCls. Avg.LLaVA bench.
DPO (β = 0.1) [57]85.343.763.8
DPO (β = 0.3) [57]83.342.164.0
IPO [4]83.640.564.5
KTO [18]83.539.564.7
SLIC [43]83.242.764.0
cDPO (β = 0.1, LS = 0.1 [43])83.143.263.9
cDPO (β = 0.3, LS = 0.1 [43])83.140.764.5
", + "image_path": "d9f04fbb7f31602eef73f5f1a591f99f4bb00af84571cf1c5504e9b9fd033e28.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 352, + 482, + 385 + ], + "lines": [ + { + "bbox": [ + 131, + 352, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 131, + 352, + 482, + 385 + ], + "type": "text", + "content": "Table 5: Effect of DPO loss variant on CLIP-DPO training. Results are reported in terms of AMBER Score, Average Top-1 (\\%) accuracy across all 9 image classification datasets, and the average on LLaVA benchmark. MobileVLM-v2 1.7B was used." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 131, + 409, + 480, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 409, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 131, + 409, + 480, + 434 + ], + "type": "text", + "content": "all, with CLIP-DPO, the model is more grounded in the visual content, less prone to hallucinations, and more precise and fine-grained in its descriptions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 453, + 251, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 251, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 251, + 464 + ], + "type": "text", + "content": "7 Ablation studies" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 131, + 479, + 480, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 479, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 131, + 479, + 480, + 502 + ], + "type": "text", + "content": "Herein, we ablate the effect of the proposed components. In all experiments, we use a MobileVLM-v2 1.7B model." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 521, + 293, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 293, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 293, + 533 + ], + "type": "text", + "content": "7.1 Effect of DPO loss variant" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 542, + 480, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 542, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 542, + 480, + 590 + ], + "type": "text", + "content": "Our work primarily follows the original DPO formulation proposed in [57]. Here we also consider the following recently proposed variations: KTO [18], ITO [4], SLIC [43] and cDPO [48]. As Tab. 5 shows, aggregated, all losses tend to perform similarly, with DPO marginally outperforming the others." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 609, + 268, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 268, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 268, + 620 + ], + "type": "text", + "content": "7.2 Effect of CLIP scorer" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Herein, we seek to explore alternatives to ViT-H/14 DFN [19] CLIP model used in previous experiments, analyzing the impact of the scorer used on the overall performance of the model. We consider a diverse set of alternatives, covering" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLIP-DPO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 202, + 114, + 411, + 169 + ], + "blocks": [ + { + "bbox": [ + 202, + 114, + 411, + 169 + ], + "lines": [ + { + "bbox": [ + 202, + 114, + 411, + 169 + ], + "spans": [ + { + "bbox": [ + 202, + 114, + 411, + 169 + ], + "type": "table", + "html": "
ScorerAMBER ScoreCls. Avg.LLaVA bench.
ViT-L/14 [56]79.842.363.0
SigLIP-L/16 [79]85.043.963.2
ViT-H/14 [60]84.344.263.7
ViT-H/14 DFN [19]85.343.763.8
", + "image_path": "45ddaa57bd5c0534c3ae4c9a8d3bffb2bea91dd966637d8bff7525d054fd4a06.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 177, + 481, + 210 + ], + "lines": [ + { + "bbox": [ + 130, + 177, + 481, + 210 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 481, + 210 + ], + "type": "text", + "content": "Table 6: Effect of CLIP scorer on CLIP-DPO training. Results are reported in terms of AMBER Score, Average Top-1 (\\%) accuracy across all 9 image classification datasets, and the average on LLaVA benchmark. MobileVLM-v2 1.7B was used." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 233, + 481, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 233, + 481, + 328 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 481, + 328 + ], + "type": "text", + "content": "multiple exploratory paths: equally-sized models trained on different data (ViT-H/14 [19] trained on LAION-2B instead of DFN-5B); smaller models (ViT-L/14 [56]) and models trained using different pre-training losses (SigLIP-ViT-L/16 [79]). As the results from Tab. 6 show, our approach is generally robust to the exact scorer used, with the notable exception of the ViT-L/14 [56] model. Notice that this difference is primarily manifesting on the AMBER benchmark. Intuitively, this showcases the importance of a powerful scorer for reducing the amount of hallucinations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 347, + 322, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 322, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 322, + 361 + ], + "type": "text", + "content": "Limitations and broader impact" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 371, + 481, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 481, + 491 + ], + "type": "text", + "content": "As an LVLM-based approach, our method is subject to the same general consideration (i.e. potential data bias, susceptibility to hallucinations, etc.). Moreover, as the LVLMs are trained on relatively small datasets compared to LLMs or CLIP, gaps within their knowledge domains are possible. This is especially important as neural networks tend to be overconfident outside their seen input distribution. As with all models from this category, we strongly recommend checking the models and the data carefully before deploying them. Despite these general aspects, our approach is shown to significantly reduce the amount of hallucinated content and improve the model's discriminability, hence resulting in more robust and reliable models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 509, + 226, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 226, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 226, + 521 + ], + "type": "text", + "content": "8 Conclusions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": "In this work, we proposed CLIP-DPO, a simple method that reduces hallucinations in LVLMs based on using a pre-trained CLIP model [56] to rank the LVLM's self-generated captions in order to construct positive-negative pairs for DPO fine-tuning. In contrast to previous works, our proposed CLIP-DPO removes the need of (i) paid-for APIs, (ii) additional data and (iii) additional external LVLMs. Before training, the data is filtered using a newly proposed robust rule-based approach. When applied on top of established LVLMs, it is shown that CLIP-DPO fine-tuning significantly reduces hallucinations, outperforming the baseline models by significant margins and even matching the performance of GPT-4V on the AMBER benchmark. We also observe superior zero-shot object recognition, suggesting improved object grounding capabilities." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "Y. Ouali et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "text", + "content": "1. Detailed caption dataset. https://huggingface.co/datasets/echo840/Detailed_Caption (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 159, + 480, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 159, + 480, + 190 + ], + "spans": [ + { + "bbox": [ + 138, + 159, + 480, + 190 + ], + "type": "text", + "content": "2. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 191, + 480, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 191, + 480, + 233 + ], + "spans": [ + { + "bbox": [ + 138, + 191, + 480, + 233 + ], + "type": "text", + "content": "3. Awadalla, A., Gao, I., Gardner, J., Hessel, J., Hanafy, Y., Zhu, W., Marathe, K., Bitton, Y., Gadre, S., Sagawa, S., et al.: Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 234, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 480, + 266 + ], + "type": "text", + "content": "4. Azar, M.G., Rowland, M., Piot, B., Guo, D., Calandriello, D., Valko, M., Munos, R.: A general theoretical paradigm to understand learning from human preferences. arXiv preprint arXiv:2310.12036 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 266, + 480, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 480, + 298 + ], + "type": "text", + "content": "5. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-VL: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 299, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 319 + ], + "type": "text", + "content": "6. Bossard, L., Guillaumin, M., Van Gool, L.: Food-101-mining discriminative components with random forests. In: European Conference on Computer Vision (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 320, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 480, + 352 + ], + "type": "text", + "content": "7. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances on Neural Information Processing Systems (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 352, + 480, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 352, + 480, + 395 + ], + "spans": [ + { + "bbox": [ + 138, + 352, + 480, + 395 + ], + "type": "text", + "content": "8. Chen, J., Zhu, D., Shen, X., Li, X., Liu, Z., Zhang, P., Krishnamoorthi, R., Chandra, V., Xiong, Y., Elhoseiny, M.: Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 396, + 480, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 480, + 427 + ], + "type": "text", + "content": "9. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 428, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 428, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 428, + 480, + 460 + ], + "type": "text", + "content": "10. Chen, L., Li, J., Dong, X., Zhang, P., He, C., Wang, J., Zhao, F., Lin, D.: Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 460, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 480, + 492 + ], + "type": "text", + "content": "1. Chen, Z., Wu, J., Wang, W., Su, W., Chen, G., Xing, S., Muyan, Z., Zhang, Q., Zhu, X., Lu, L., et al.: Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 492, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 492, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 492, + 480, + 536 + ], + "type": "text", + "content": "2. Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., et al.: Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 138, + 492, + 480, + 536 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 138, + 492, + 480, + 536 + ], + "type": "text", + "content": " chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023) (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "type": "text", + "content": "3. Chu, X., Qiao, L., Lin, X., Xu, S., Yang, Y., Hu, Y., Wei, F., Zhang, X., Zhang, B., Wei, X., et al.: MobileVLM: A fast, reproducible and strong vision language assistant for mobile devices. arXiv preprint arXiv:2312.16886 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "type": "text", + "content": "4. Chu, X., Qiao, L., Zhang, X., Xu, S., Wei, F., Yang, Y., Sun, X., Hu, Y., Lin, X., Zhang, B., et al.: MobileVLM V2: Faster and stronger baseline for vision language model. arXiv preprint arXiv:2402.03766 (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "text", + "content": "5. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "6. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: InstructBLIP: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLIP-DPO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "17. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: IEEE Conference on Computer Vision and Pattern Recognition (2009)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "text", + "content": "18. Ethayarajh, K., Xu, W., Muennighoff, N., Jurafsky, D., Kiela, D.: Kto: Model alignment as prospect theoretic optimization. arXiv preprint arXiv:2402.01306 (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 194 + ], + "type": "text", + "content": "19. Fang, A., Jose, A.M., Jain, A., Schmidt, L., Toshev, A., Shankar, V.: Data filtering networks. arXiv preprint arXiv:2309.17425 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 194, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 237 + ], + "type": "text", + "content": "20. Fei-Fei, L., Fergus, R., Perona, P.: Learning generative visual models from few training examples: An incremental Bayesian approach tested on 101 object categories. In: IEEE Conference on Computer Vision and Pattern Recognition - Workshops (2004)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "21. Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., et al.: Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "text", + "content": "22. Gunjal, A., Yin, J., Bas, E.: Detecting and preventing hallucinations in large vision language models. arXiv preprint arXiv:2308.06394 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "type": "text", + "content": "23. Helber, P., Bischke, B., Dengel, A., Borth, D.: EuroSAT: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 12(7), 2217-2226 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "24. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 392 + ], + "type": "text", + "content": "25. Hu, H., Zhang, J., Zhao, M., Sun, Z.: Ciem: Contrastive instruction evaluation method for better instruction tuning. arXiv preprint arXiv:2309.02301 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "text", + "content": "26. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: IEEE Conference on Computer Vision and Pattern Recognition (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "text", + "content": "27. Jain, J., Yang, J., Shi, H.: Vcoder: Versatile vision encoders for multimodal large language models. arXiv preprint arXiv:2312.14233 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "text", + "content": "28. Jiang, A.Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D.S., Casas, D.d.l., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., et al.: Mistral 7b. arXiv preprint arXiv:2310.06825 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 512 + ], + "type": "text", + "content": "29. Jiang, C., Xu, H., Dong, M., Chen, J., Ye, W., Yan, M., Ye, Q., Zhang, J., Huang, F., Zhang, S.: Hallucination augmented contrastive learning for multimodal large language model. arXiv preprint arXiv:2312.06968 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "text", + "content": "30. Jing, L., Li, R., Chen, Y., Jia, M., Du, X.: Faithscore: Evaluating hallucinations in large vision-language models. arXiv preprint arXiv:2311.01477 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "31. Koh, J.Y., Salakhutdinov, R., Fried, D.: Grounding language models to images for multimodal inputs and outputs. International Conference on Machine Learning (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "32. Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3D object representations for fine-grained categorization. In: IEEE International Conference on Computer Vision - Workshops (2013)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "33. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "34. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "Y. Ouali et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "35. Li, L., Xie, Z., Li, M., Chen, S., Wang, P., Chen, L., Yang, Y., Wang, B., Kong, L.: Silkie: Preference distillation for large visual language models. arXiv preprint arXiv:2312.10665 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "text", + "content": "36. Li, S., Lin, R., Pei, S.: Multi-modal preference alignment remedies regression of visual instruction tuning on language model. arXiv preprint arXiv:2402.10884 (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "text", + "content": "37. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "38. Li, Y., Bubeck, S., Eldan, R., Del Giorno, A., Gunasekar, S., Lee, Y.T.: Textbooks are all you need ii: phi-1.5 technical report. arXiv preprint arXiv:2309.05463 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "type": "text", + "content": "39. Lin, B., Tang, Z., Ye, Y., Cui, J., Zhu, B., Jin, P., Zhang, J., Ning, M., Yuan, L.: Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 248, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 481, + 282 + ], + "type": "text", + "content": "40. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Mitigating hallucination in large multi-modal models via robust instruction tuning. In: International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "text", + "content": "41. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 326 + ], + "type": "text", + "content": "42. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances on Neural Information Processing Systems (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 358 + ], + "type": "text", + "content": "43. Liu, T., Zhao, Y., Joshi, R., Khalman, M., Saleh, M., Liu, P.J., Liu, J.: Statistical rejection sampling improves preference optimization. arXiv preprint arXiv:2309.06657 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "text", + "content": "44. Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: MMBench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "text", + "content": "45. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 413, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 446 + ], + "type": "text", + "content": "46. Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "text", + "content": "47. Mishra, S., Khashabi, D., Baral, C., Hajishirzi, H.: Cross-task generalization via natural language crowdsourcing instructions. arXiv preprint arXiv:2104.08773 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 479, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 490 + ], + "type": "text", + "content": "48. Mitchell, E.: A note on dpo with noisy preferences & relationship to ipo (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "text", + "content": "49. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large number of classes. In: Indian conference on computer vision, graphics & image processing. pp. 722-729 (2008)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 523, + 298, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 298, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 298, + 533 + ], + "type": "text", + "content": "50. OpenAI: Introducing chatgpt (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 533, + 323, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 323, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 323, + 544 + ], + "type": "text", + "content": "51. OpenAI: Gpt-4v(ision) system card (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 544, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 588 + ], + "type": "text", + "content": "52. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "type": "text", + "content": "53. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: IEEE Conference on Computer Vision and Pattern Recognition (2012)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "text", + "content": "54. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch (2017)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "55. Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLIP-DPO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 160 + ], + "type": "text", + "content": "56. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 480, + 194 + ], + "type": "text", + "content": "57. Rafailov, R., Sharma, A., Mitchell, E., Manning, C.D., Ermon, S., Finn, C.: Direct preference optimization: Your language model is secretly a reward model. Advances on Neural Information Processing Systems (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 237 + ], + "type": "text", + "content": "58. Rajbhandari, S., Ruwase, O., Rasley, J., Smith, S., He, Y.: Zero-infinity: Breaking thegpu memory wall for extreme scale deep learning. In: Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 237, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 282 + ], + "type": "text", + "content": "59. Rasley, J., Rajbhandari, S., Ruwase, O., He, Y.: Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In: Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 282, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 326 + ], + "type": "text", + "content": "60. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 326, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 347 + ], + "type": "text", + "content": "61. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "62. Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards VQA models that can read. In: IEEE Conference on Computer Vision and Pattern Recognition (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 403 + ], + "type": "text", + "content": "63. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "text", + "content": "64. Stiennon, N., Ouyang, L., Wu, J., Ziegler, D., Lowe, R., Voss, C., Radford, A., Amodei, D., Christiano, P.F.: Learning to summarize with human feedback. Advances in Neural Information Processing Systems (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "65. Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L.Y., Wang, Y.X., Yang, Y., et al.: Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "text", + "content": "66. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "type": "text", + "content": "67. Thoppilan, R., De Freitas, D., Hall, J., Shazeer, N., Kulshreshtha, A., Cheng, H.T., Jin, A., Bos, T., Baker, L., Du, Y., et al.: Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 534, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 567 + ], + "type": "text", + "content": "68. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "text", + "content": "69. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "70. Wang, J., Yang, Z., Hu, X., Li, L., Lin, K., Gan, Z., Liu, Z., Liu, C., Wang, L.: Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "71. Wang, J., Wang, Y., Xu, G., Zhang, J., Gu, Y., Jia, H., Yan, M., Zhang, J., Sang, J.: An LLM-free multi-dimensional benchmark for MLLMs hallucination evaluation. arXiv preprint arXiv:2311.07397 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "Y. Ouali et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 654 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "72. Wang, J., Zhou, Y., Xu, G., Shi, P., Zhao, C., Xu, H., Ye, Q., Yan, M., Zhang, J., Zhu, J., et al.: Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "text", + "content": "73. Williams, R.J.: Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning 8, 229-256 (1992)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "text", + "content": "74. Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: Sun database: Large-scale scene recognition from abbey to zoo. In: IEEE Conference on Computer Vision and Pattern Recognition (2010)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "75. Xu, J., Lee, A., Sukhbaatar, S., Weston, J.: Some things are more cringe than others: Preference optimization with the pairwise cringe loss. arXiv preprint arXiv:2312.16682 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "76. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "type": "text", + "content": "77. You, H., Zhang, H., Gan, Z., Du, X., Zhang, B., Wang, Z., Cao, L., Chang, S.F., Yang, Y.: Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 346 + ], + "type": "text", + "content": "78. Zhai, B., Yang, S., Zhao, X., Xu, C., Shen, S., Zhao, D., Keutzer, K., Li, M., Yan, T., Fan, X.: Halle-switch: Rethinking and controlling object existence hallucinations in large vision language models for detailed caption. arXiv preprint arXiv:2310.01779 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "text", + "content": "79. Zhai, X., Mustafa, B., Kolesnikov, A., Beyer, L.: Sigmoid loss for language image pre-training. arXiv preprint arXiv:2303.15343 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "80. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 402, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 480, + 435 + ], + "type": "text", + "content": "81. Zhao, H., Cai, Z., Si, S., Ma, X., An, K., Chen, L., Liu, Z., Wang, S., Han, W., Chang, B.: Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "82. Zhao, Y., Joshi, R., Liu, T., Khalman, M., Saleh, M., Liu, P.J.: Slic-hf: Sequence likelihood calibration with human feedback. arXiv preprint arXiv:2305.10425 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "text", + "content": "83. Zhao, Z., Wang, B., Ouyang, L., Dong, X., Wang, J., He, C.: Beyond hallucinations: Enhancing LVLMs through hallucination-aware direct preference optimization. arXiv preprint arXiv:2311.16839 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 500, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 480, + 533 + ], + "type": "text", + "content": "84. Zhou, C., Liu, P., Xu, P., Iyer, S., Sun, J., Mao, Y., Ma, X., Efrat, A., Yu, P., Yu, L., et al.: Lima: Less is more for alignment. Advances in Neural Information Processing Systems (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "type": "text", + "content": "85. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: IEEE Conference on Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 566, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 566, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 480, + 599 + ], + "type": "text", + "content": "86. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: MiniGPT-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 599, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 480, + 621 + ], + "type": "text", + "content": "87. Zhu, Y., Zhu, M., Liu, N., Ou, Z., Mou, X., Tang, J.: LLaVA-phi: Efficient multi-modal assistant with small language model. arXiv preprint arXiv:2401.02330 (2024)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 621, + 480, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 654 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 654 + ], + "type": "text", + "content": "88. Ziegler, D.M., Stiannon, N., Wu, J., Brown, T.B., Radford, A., Amodei, D., Christiano, P., Irving, G.: Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593 (2019)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 400, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLIP-DPO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CLIP-Guided Generative Networks for Transferable Targeted Adversarial Attacks/ddf47f95-f1c1-405a-8323-f8a66b1071ea_content_list.json b/2024/CLIP-Guided Generative Networks for Transferable Targeted Adversarial Attacks/ddf47f95-f1c1-405a-8323-f8a66b1071ea_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..df18ce9d6c97b1c85eeaa7147b07bdd6092d02ba --- /dev/null +++ b/2024/CLIP-Guided Generative Networks for Transferable Targeted Adversarial Attacks/ddf47f95-f1c1-405a-8323-f8a66b1071ea_content_list.json @@ -0,0 +1,1702 @@ +[ + { + "type": "text", + "text": "CLIP-Guided Generative Networks for Transferable Targeted Adversarial Attacks", + "text_level": 1, + "bbox": [ + 259, + 140, + 743, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao Fang $^{1\\dagger}$ , Jiawei Kong $^{2\\dagger}$ , Bin Chen $^{2,3,4\\#}$ , Tao Dai $^{5}$ , Hao Wu $^{6}$ , and Shu-Tao Xia $^{1,3}$", + "bbox": [ + 250, + 210, + 754, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tsinghua Shenzhen International Graduate School, Tsinghua University $^{2}$ Harbin Institute of Technology, Shenzhen $^{3}$ Pengcheng Laboratory", + "bbox": [ + 250, + 253, + 753, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Guangdong Provincial Key Laboratory of Novel Security Intelligence Technologies \n5 Shenzhen University 6 Shenzhen Digital Certificate Authority CO., Ltd \nfang-h23@mails.tsinghua.edu.cn, kongjiawei@stu.hit.edu.cn \nchenbin2021@hit.edu.cn, daitao.edu@gmail.com, whpc79@163.com, \nxiast@sz.tsinghua.edu.cn", + "bbox": [ + 223, + 282, + 779, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Transferable targeted adversarial attacks aim to mislead models into outputting adversary-specified predictions in black-box scenarios. Recent studies have introduced single-target attacks that train a generator for each target class to generate highly transferable perturbations, resulting in substantial computational overhead when handling multiple classes. Multi-target attacks address this by training only one class-conditional generator for multiple classes. However, the generator simply uses class labels as conditions, failing to leverage the rich semantic information of the target class. To this end, we design a CLIP-guided Generative Network with Cross-attention modules (CGNC) to enhance multi-target attacks by incorporating textual knowledge of CLIP into the generator. Extensive experiments demonstrate that CGNC yields significant improvements over previous multi-target attacks, e.g., a $21.46\\%$ improvement in success rate from Res-152 to DenseNet-121. Moreover, we propose the masked fine-tuning to further strengthen our method in attacking a single class, which surpasses existing single-target methods.", + "bbox": [ + 259, + 388, + 743, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 635, + 377, + 651 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapid progress of deep learning, deep neural networks (DNNs) have been widely applied in many security-critical fields, such as autonomous driving [15, 37], financial systems [53], and point cloud modeling [69, 70]. However, DNNs are corroborated to be vulnerable to adversarial attacks [20, 26, 58], which attempt to fool models with adversarial examples crafted by adding imperceptible perturbations to the original inputs. Based on the attack goal, adversarial attacks can be categorized into untargeted and targeted attacks. Untargeted attacks attempt to fool DNNs into predicting incorrect labels while targeted attacks aim at triggering the model to output the attacker-desired predictions.", + "bbox": [ + 212, + 666, + 787, + 803 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Equal contribution.", + "bbox": [ + 232, + 810, + 370, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author.", + "bbox": [ + 232, + 824, + 395, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e597ef4133afdf3eb4daa87c93fd652ba2630da50cbeec01b0458fd1c5157721.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 246, + 146, + 553, + 271 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e07c63107175448b06783303ae2d73a0281fb1165b7aef9c1dfcac5908d528e7.jpg", + "image_caption": [ + "(b)", + "Fig. 1: (a) Targeted attacks from 'Panda' to 'Dog'. Fig. (a) illustrates that previous multi-target methods [27,67] generate perturbations simply conditioned by class indices or one-hot vectors and only learn the classification boundary specific to the surrogate model. In contrast, CGNC exploits CLIP's meaningful guidance to effectively capture the feature distribution inherent to the target data. (b) We directly feed the scaled perturbations generated by our CGNC and C-GSP [67] into three black-box classifiers and reveal that our perturbations achieve significantly higher mean confidence of the target class, demonstrating its superiority in modeling the target feature distribution." + ], + "image_footnote": [], + "bbox": [ + 563, + 162, + 754, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent investigations into the adversarial transferability [2, 3, 44, 58] have demonstrated that adversarial examples crafted for a white-box surrogate model can also mislead other unseen black-box models. Since this attack does not require access to the target model, it exposes a serious security threat to real-world applications of DNNs and motivates a wide range of studies [1,5,42,46]. Despite the remarkable performance on untargeted transferable attacks, these approaches produce unsatisfactory results for targeted attacks due to their over-reliance on the white-box surrogate. Existing studies on transferable targeted attacks can be categorized into instance-specific [11, 15, 23, 39, 43, 60, 66] and instance-agnostic attacks [19, 37, 44, 48, 49, 64]. Specifically, instance-specific attacks [11, 12, 65] iteratively perform gradient updating to craft adversarial perturbation tailored to a specific natural sample, yet producing low-targeted transferability due to overfitting [67]. Conversely, instance-agnostic attacks learn a universal perturbation [45, 71] or a perturbation generator [48, 51] based on data distribution rather than the specific instance, alleviating the data-specific over-fitting issues and achieving a higher adversarial transferability. Recent studies further explore generative attacks to produce highly transferable perturbations, which can be divided into single-target and multi-target attacks as follows.", + "bbox": [ + 212, + 443, + 787, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single-target attacks [19, 48, 49, 62] exhibit impressive performance by training perturbation generators for target categories. However, they require training a generator for each target, which can lead to a heavy computation burden when attacking numerous classes, and thus are not applicable in real-world classification systems that usually contain hundreds/thousands of target classes [27]. To address this, [27, 67] propose multi-target attacks that train a single conditional generator for multi-target classes. By specifying the desired label as conditioning input, the trained generator can efficiently generate the corresponding adversar", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H. Fang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ial perturbation. Nevertheless, these methods simply adopt class indices [27] or one-hot vectors [67] of the target labels as conditions, and thus solely rely on the classification information from the surrogate model as the guidance of the target category, hence resulting in only modest black-box fooling rates.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we build upon the research line of multi-target attacks and propose a novel CLIP-guided Generative Network with Cross-attention modules (CGNC) to boost the attack effect. Concretely, we revisit the architecture of the conditional generators used in [27,67] and argue that the simple class conditions, e.g., class indices or handcrafted one-hot vectors, limit the transferability of the generated perturbations (see Fig. 1(a) for comparison). Motivated by the impressive effects of CLIP's encoded text embeddings in multi-modal learning, we introduce concise text descriptions of the target classes to encode them as class-specific representations encapsulated with abundant information, which assist the generator in learning target class distribution and ultimately lead to a fundamental transferability improvement. To better incorporate the text information, we improve the condition-inputting mechanism by adding the cross-attention layers that have been proven effective in models with various input modalities. Results in Fig. 1(b) confirm the capability of our CGNC in capturing the target distribution. Besides, we propose a masked fine-tuning (MFT) technique, which fine-tunes the trained conditional generator with a fixed text condition of the desired class for further improvement in attacking a single class.", + "bbox": [ + 212, + 207, + 784, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With all the above efforts, our method achieves great efficiency and scalability. When there are hundreds of target classes and single-target methods become impractical, the proposed conditional network can be utilized to achieve significant performance over previous multi-target attacks. Conversely, for scenarios involving only a few target classes, the proposed MFT mechanism enhances CGNC to outperform existing single-target methods while substantially reducing computational costs. In summary, our main contributions are as follows:", + "bbox": [ + 212, + 463, + 784, + 569 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose CGNC, a novel CLIP-guided generative network with cross-attention layers that fully exploits the textual knowledge provided by the advanced CLIP model for enhanced multi-target attacks.", + "- We introduce a masked fine-tuning mechanism to improve the single-target performance by adapting the CGNC to an individual target class.", + "- Extensive experiments show that CGNC remarkably enhances the targeted transferability compared to previous attack methods in a range of settings." + ], + "bbox": [ + 225, + 579, + 782, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 707, + 387, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Vision-Language Models", + "text_level": 1, + "bbox": [ + 215, + 739, + 464, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-language models (VLM) drawn great attention [21,22] due to their promising potential in learning general visual and textual representations through contrastive pre-training on large-scale image-text pairs. Given the powerful capacity of text descriptions in modeling multi-modal tasks, these models can be effectively adapted to diverse downstream tasks through appropriately formulated", + "bbox": [ + 212, + 763, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "CLIP-Guided Networks for Transferable Targeted Attacks", + "bbox": [ + 344, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "textual prompts, including video understanding [36], image manipulation [7], and text-to-image synthesis [59]. Inspired by these works, we propose to harness CLIP-based multi-modal learning to empower our perturbation generator, facilitating more effective learning of the target feature distribution.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Adversarial Attacks", + "text_level": 1, + "bbox": [ + 215, + 233, + 426, + 248 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Among various security threats to DNNs [8, 16, 18, 68], the adversarial attack is one of the most formidable and well-known one, which can be classified into instance-specific attacks and instance-agnostic attacks [67].", + "bbox": [ + 212, + 263, + 785, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instance-specific Attacks. Since the pioneering work [58] highlighted the vulnerability of neural networks, numerous gradient-based optimization methods [3, 13, 26, 38, 66] have been proposed to craft image-dependent perturbations. MIM [11] integrates a momentum item into the gradient update for utilizing the previous gradient information to avoid plunging into poor local optimum. DIM [65] enhances the transferability by randomly diversifying the input pattern, and TIM [12] implements the attack by convolving the gradient with a predefined kernel. In addition, intermediate feature space [35, 63] and classifier information [34] are exploited to enhance attack effects, while [72] leverages logit-based loss to achieve competitive results. More advanced works [6,28,30] consider an ensemble of multiple surrogate models to reduce over-fitting.", + "bbox": [ + 212, + 309, + 787, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instance-agnostic Attacks. In contrast to instance-specific attacks, instance-agnostic attacks learn a universal perturbation [17, 45] or a generative model [44, 48, 49, 51, 61, 67] for crafting adversarial examples. By modeling the global data distribution, these methods have shown better transferability and attracted more attention in recent years. Specifically, UAP [45] and AAA [47] learn a universal perturbation to fool the model based on concrete data and compressed impression, respectively. Many subsequent works, such as GAP [51], focus on boosting the transferability of adversarial attacks by training generative models. These generative attacks can be categorized into the following two types.", + "bbox": [ + 212, + 476, + 797, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Single-target Attacks. This type of attack requires training a generator for each target class. [64] first introduce the generative adversarial networks (GAN) [25] to generate adversarial perturbations. Then, CD-AP [48] concentrates on the domain-invariant adversaries and launches highly transferable cross-domain attacks using a relativistic supervisory. TTP [49] modifies the loss function and proposes to match the target distribution to mitigate over-fitting to the surrogate model. Subsequent methods achieve better transferability based on a dynamic network with pattern injection [19] or a feature discriminator [62].", + "bbox": [ + 212, + 612, + 787, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-target Attacks. MAN [27] notes that when dealing with numerous classes, single-target attacks inevitably suffer from extreme computational burdens as they need to train the same number of models as multiple target classes. Therefore, single-target attacks become impractical in attacking real classification systems that often have hundreds of categories. To pursue the extreme speed and storage, MAN trains only one model for 1000 target categories from ImageNet [9]. However, the excess of the target class severely degrades the targeted", + "bbox": [ + 212, + 734, + 790, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H. Fang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "transferability. C-GSP [67] improves the performance by designing a hierarchical partition mechanism to divide all classes into a feasible number of subsets and train generators for each subset. Nonetheless, C-GSP simply conditions the generator with one-hot vectors and only utilizes the classification information to train the generator, failing to use the semantic knowledge of the target class. To overcome the limitation, we propose to incorporate the text information provided by CLIP as significant guidance for the target class.", + "bbox": [ + 212, + 146, + 782, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 273, + 330, + 289 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first introduce the preliminaries of targeted transferable attacks and present the basic paradigm of generative attack methods. Then, we elaborate on the proposed CGNC, which remarkably enhances multi-target black-box attacks. Finally, we detail the proposed masked fine-tuning technique that strengthens our method in single-class attacks.", + "bbox": [ + 212, + 303, + 782, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminary", + "text_level": 1, + "bbox": [ + 215, + 398, + 361, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We denote the white-box image classifier parameterized with $\\theta$ as $f_{\\theta}:\\mathcal{X}\\rightarrow \\mathcal{Y}$ where $\\mathcal{X}\\subset \\mathbb{R}^{N\\times H\\times W}$ represents the image domain and $\\mathcal{V}\\subset \\mathbb{R}^L$ is the output confidence score of different classes $(H,W,N,L$ being height, width, number of channels, and class number). Given a natural image $\\pmb {x}\\in \\mathcal{X}$ and the attacker's desired label $c_{t}\\in \\mathcal{C}$ , the transferable targeted adversarial attacks attempt to craft the imperceptible perturbation $\\delta$ based on the accessible surrogate model $f_{\\theta}$ to mislead another unseen victim model $F_{\\phi}$ into predicting $c_{t}$ , i.e., $\\arg \\max_{i\\in \\mathcal{C}}F_{\\phi}(\\pmb {x} + \\pmb {\\delta})_i = c_t$ . Concurrently, the $l_{\\infty}$ norm is employed to ensure the adversarial samples are indistinguishable from the original images by constraining the perturbation within the range $\\epsilon$ by $\\| \\pmb {\\delta}\\|_{\\infty}\\leq \\epsilon$ .", + "bbox": [ + 212, + 422, + 784, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To boost the fooling rate of targeted black-box attacks, single-target attacks [48, 49, 51] utilize powerful generative models to learn the target data distribution and achieve higher transferability. However, these methods consume great computation time and resources for multi-target scenarios, making them impractical for real-world attacks. C-GSP [67] effectively solves this issue by formulating the multi-target attacks as learning a class-conditional generator $G_{w}$ with weights $w$ , which is capable of generating perturbations for any target class. Given an unlabeled training dataset $\\mathcal{X}_s$ , the optimization objective is as follows:", + "bbox": [ + 212, + 574, + 782, + 695 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {w} \\mathbb {E} _ {\\boldsymbol {x} _ {s} \\sim \\chi_ {s}, c \\sim \\mathcal {C}} [ \\mathcal {L} (f _ {\\theta} (\\boldsymbol {x} _ {s} + G _ {w} (\\boldsymbol {x} _ {s}, c)), c) ], \\text {s . t .} \\| G _ {w} (\\boldsymbol {x} _ {s}, c) \\| _ {\\infty} \\leq \\epsilon , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 269, + 720, + 784, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}(\\cdot, \\cdot)$ denotes the cross-entropy (CE) loss. By minimizing the loss of specified classes using various unlabeled images from $\\mathcal{X}_s$ , we optimize the parameters $w$ of the generative model and finally obtain a targeted conditional generator that can generate perturbation for any given clean image $\\pmb{x}_t$ from the test dataset $\\mathcal{X}_t$ . Specifically, the adversary can simply specify a target label $c$ and craft an corresponding adversarial example via $\\pmb{x}_{adv} = \\pmb{x}_t + G_w(\\pmb{x}_t, c)$ .", + "bbox": [ + 212, + 750, + 784, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "CLIP-Guided Networks for Transferable Targeted Attacks", + "bbox": [ + 344, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6ad7695640e03bc51b244d528ec508da07dde1217990c7f38aaa3b5759aa9e41.jpg", + "image_caption": [ + "Fig. 2: An overview of our proposed architecture of CGNC. Equipped with the three exquisite modules VL-Purifier, F-Encoder, and CA-Decoder, the generator fully leverages the textual representations encoded by CLIP as auxiliary information about the target classes to better probe their data distribution and achieve superior attack effects." + ], + "image_footnote": [], + "bbox": [ + 218, + 148, + 782, + 262 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, current multi-target methods [27, 67] simply condition the generator with class labels and learn the target distribution only relying on the classification information of the surrogate model, thus not fully exploiting the specific characteristics of the target category. Inspired by the efficiency of vision-language (VL) learning [52, 59], we propose a novel generative network that leverages sufficient prior knowledge from the powerful CLIP model by incorporating textual-modality information to promote the target class modeling, which greatly enhances the multi-target transferable attacks.", + "bbox": [ + 212, + 376, + 787, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 CLIP-Guided Generative Network", + "text_level": 1, + "bbox": [ + 214, + 523, + 545, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proposed generative model architecture is presented in Fig. 2. Specifically, CGNC is composed of a Vision-Language feature Purifier (VL-Purifier), a feature Fusion Encoder (F-Encoder), and a Cross-Attention based Decoder (CA-Decoder). We also provide the pseudocode of the training procedure in Algorithm 1. Next, we illustrate the design of each module as follows.", + "bbox": [ + 212, + 551, + 787, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Vision-language feature purifier (VL-Purifier). To utilize CLIP to produce semantic embeddings, we first feed in CLIP's text encoder $\\varPhi$ with queries $t_c$ that follow the handcrafted template \"a photo of a {class}\" which has shown effectiveness in many tasks [33,52]. Since the obtained embeddings $\\pmb{e}_t\\in \\mathbb{R}^{B\\times 512}$ ( $B$ being the batch size) in CLIP's vision-language space are generic representations of the target classes and are not yet tailored to our learning task, we refine them using the VL-Purifier, which is composed of several blocks consisting of a fully-connected layer and a spectral normalization layer. Through this module, we translate the encoded embeddings into more meaningful representations $\\pmb{e}_t^*\\in \\mathbb{R}^{B\\times 16}$ , thereby facilitating the subsequent step of feature fusion.", + "bbox": [ + 212, + 627, + 787, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Feature fusion encoder (F-Encoder). This module aims to fuse the purified features $\\pmb{e}_t^*$ with the learned visual representations. Firstly, a batch of input images $\\pmb{x}_s$ is encoded to capture the visual concepts $\\pmb{h}_s \\in \\mathbb{R}^{B \\times N' \\times H' \\times W'}$ . Then, we expand the text embedding $\\pmb{e}_t^* \\in \\mathbb{R}^{B \\times 16}$ into $\\pmb{e}_t^{*'} \\in \\mathbb{R}^{B \\times 16 \\times H' \\times W'}$ , which are", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H. Fang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Pseudocode of Training the CLIP-guided Generative Network" + ], + "code_body": "Require: $\\mathcal{X}_s$ : the training data; $\\mathcal{C}$ : the target label space; $\\mathcal{T}$ : the text prompts set; $f_{\\theta}$ : the surrogate model; $\\varPhi$ : the CLIP's text encoder; $N$ : the max iteration; \nEnsure: the perturbation generator $G_w$ ; \n1: for $i \\gets 0$ to $N$ do \n2: Sample a batch of images $\\boldsymbol{x}_s \\sim \\mathcal{X}_s$ ; \n3: Obtain $\\boldsymbol{x}_s'$ by processing $\\boldsymbol{x}_s$ with data augmentation; \n4: Sample a batch of target labels $c \\sim \\mathcal{C}$ ; \n5: Obtain the corresponding text prompts $t_c$ from $\\mathcal{T}$ ; \n6: Compute the text embedding $\\boldsymbol{e}_t$ by feeding $t_c$ into $\\varPhi$ ; \n7: Obtain perturbed images by $\\boldsymbol{x}_{adv} = \\boldsymbol{x}_s + G_w(\\boldsymbol{x}_s, \\boldsymbol{e}_t)$ , $\\boldsymbol{x}_{adv}' = \\boldsymbol{x}_s' + G_w(\\boldsymbol{x}_s', \\boldsymbol{e}_t)$ ; \n8: Forward pass $\\boldsymbol{x}_{adv}$ , $\\boldsymbol{x}_{adv}'$ to $f_{\\theta}$ and compute the loss in Eq. (3); \n9: Perform gradient backpropagation and update the generator $G_w$ ; \n10: end for \n11: return the trained generator $G_w$", + "bbox": [ + 215, + 164, + 785, + 362 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "then integrated with the extracted visual concepts $\\pmb{h}_s$ through channel-wise concatenation to obtain the fused representations, i.e., $\\pmb{m} \\in \\mathbb{R}^{B \\times (N' + 16) \\times H' \\times W'}$ . Subsequently, $\\pmb{m}$ undergoes further downsampling, and the resultant features are again concatenated with the expanded embedding $e_t^{*'}$ . By repeating this operation several times, we effectively fuse the visual concepts of input images and the purified CLIP's embedding of the target classes. This mechanism fully exploits both the instance-level and class-level information from visual and textual modalities, thus encouraging the generation of perturbations with better semantic patterns and higher transferability.", + "bbox": [ + 212, + 401, + 787, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-Attention based Decoder(CA-Decoder). Given the multi-modal fused features from the previous module, the decoder attempts to translate them into perturbations of the target class. The network backbone is realized based on the decoder used in previous works [49,67]. To fully explore the semantic priors brought by the CLIP model, we enhance the underlying backbone by introducing the cross-attention mechanism, which is proven to be effective for many multimodal learning tasks. Specifically, we incorporate the textual embedding $\\boldsymbol{e}_t$ from CLIP's latent space into our decoder via the cross-attention layer:", + "bbox": [ + 212, + 541, + 789, + 662 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} Q = z _ {t} W _ {q}, K = e _ {t} W _ {k}, V = e _ {t} W _ {v}, \\\\ \\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) \\cdot V, \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 700, + 784, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{z}_t\\in \\mathbb{R}^{B\\times d_\\alpha}$ denotes the flattened intermediate features of the decoder, $W_{q}\\in \\mathbb{R}^{d_{\\alpha}\\times d}$ , $W_{k}\\in \\mathbb{R}^{512\\times d}$ , $W_{v}\\in \\mathbb{R}^{512\\times d}$ are learnable parameters. Similar to C-GSP, we post-process the output $\\pmb{o}$ via the $\\tanh (\\cdot)$ smooth projection to obtain the $\\ell_{\\infty}$ constrained perturbation with budge $\\epsilon$ , i.e., $\\delta = P(o) = \\epsilon \\cdot \\tanh (o)$ .", + "bbox": [ + 212, + 777, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "CLIP-Guided Networks for Transferable Targeted Attacks", + "bbox": [ + 344, + 113, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/535006a5d511924f267f7362672a9da371d50584ba9b950433af67444248dd16.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 228, + 156, + 519, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7b805cc33c4da89426f50342466f93d72f015ea4339a44f34f6568682834403e.jpg", + "image_caption": [ + "(b)", + "Fig. 3: Illustration of the proposed masked fine-tuning mechanism. (a) We fix the condition input with different text prompts to fine-tune the trained conditional generator $G_{w}$ into multiple generators for single-target attacks. (b) The fooling rate of several target classes with Inc-v3 and Res-152 as substitute models respectively. The results indicate that direct fine-tuning yields inadequate results for certain classes due to overfitting. We efficiently resolve this issue via a patch-wise random mask operation." + ], + "image_footnote": [], + "bbox": [ + 532, + 145, + 787, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Optimization objective. Based on the proposed CLIP-guided architecture, the optimization objective can be formulated as:", + "bbox": [ + 212, + 383, + 782, + 412 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nw ^ {*} \\leftarrow \\arg \\min _ {w} \\mathcal {L} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {s} + G _ {w} (\\boldsymbol {x} _ {s}, \\Phi (\\boldsymbol {t} _ {c}))\\right), c\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 420, + 784, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $t_c$ represents the corresponding text prompts of the target classes $c$ , $\\mathcal{L}(\\cdot, \\cdot)$ denotes the cross-entropy loss. By encouraging the network to craft adversarial samples that can misguide the surrogate model's output toward the desired labels, the generator learns the data distribution of the target class and thus exhibits great generalizability in producing perturbation for any input data.", + "bbox": [ + 212, + 452, + 784, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Masked Fine-Tuning Mechanism", + "text_level": 1, + "bbox": [ + 214, + 546, + 529, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In addition to the proposed CLIP-guided generator for multi-target attack scenarios, we also design a single-target variant for further improved performance. As depicted in Fig. 3a, we fix the conditioning input with the text description of a specific target class, and fine-tune the trained multi-target generator by resolving the objective function in Eq. (3) on the unlabeled training data $\\mathcal{X}_s$ . This strategy further enhances the attack effects by fine-tuning the conditional generator to specialize in a specific target class.", + "bbox": [ + 212, + 568, + 784, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "However, we encounter the overfitting problem that leads to limited improvement in success rates or even performance degradation for certain target classes. This problem is partly attributed to the fact that the generated adversarial perturbation sometimes heavily focuses on specific regions of the input image [62]. To alleviate this issue, we adopt a patch-wise random mask operation to postprocess the adversarial perturbation produced by the generator, which brings a notable increase in the targeted fooling rate as shown in Fig. 3b. This technique improves the capability and flexibility of our method, as an adversary can perform fine-tuning to further augment the generator for single-target attacks, achieving a better trade-off between efficiency and performance compared to previous single-target methods (See Section 4.5 for detailed analysis).", + "bbox": [ + 212, + 674, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H. Fang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To validate the effectiveness of CGNC in boosting transferability, we conduct extensive experiments on various black-box models in a range of scenarios. Our code is available at https://github.com/ffhibnese/CGNC_Targeted_Adversarial_Attacks. Please see the Appendix for more experimental results.", + "bbox": [ + 212, + 175, + 782, + 236 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 215, + 257, + 444, + 272 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. Following [19,67], we train the generator on the ImageNet training set [9] and evaluate the attack performance using ImageNet-NeurIPS (1k) dataset proposed by [50]. We also conduct experiments in a more realistic cross-domain scenario where we train the generator and generate perturbations on MS-COCO [41] or Comics [4] while evaluating on target classifier trained on ImageNet.", + "bbox": [ + 212, + 281, + 782, + 354 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Victim Models. We use various victim models with different architectures. Specifically, the naturally trained models include Inception-v3 (Inc-v3) [57], Inception-v4 (Inc-v4) [55], Inception-ResNet-v2 (Inc-Res-v2) [55], ResNet-152 (Res-152) [29], DenseNet-121 (DN-121) [32], GoogleNet [56], and VGG-16 [54].", + "bbox": [ + 212, + 357, + 782, + 415 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For further evaluation, we also analyze on the robust-trained models, including adv-Inception-v3 (Inc-v3adv) [26], ens-adv-Inception-ResNet-v2 (IR-v2ens) [28], and robust-trained ResNet-50 [24,31], dubbed as Res50SIN (trained on stylized ImageNet), Res50IN (trained on the mixture of stylized and Nature ImageNet), Res50fine (Res50IN plus further finetuning with an auxiliary dataset [24]), and Res50Aug (trained with the advanced data augmentation Augmix [31]).", + "bbox": [ + 212, + 417, + 782, + 503 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baseline Attacks. We reveal the superiority of the proposed CGNC in enhancing multi-target attacks by comparing our method with multiple competitive baselines, including MIM [11], DIM [65], SIM [40], DIM [12], Logit [72], SU [63], and the state-of-the-art (SOTA) multi-target generative attacks CGSP [67] in Section 4.2. For SU attack [63], we choose to compare with its best version DTMI-Logit-SU. Besides, we also provide a comparison of the proposed single-target variant of CGNC with the existing single-target attack methods, i.e., GAP [51], CD-AP [48], TTP [49], and DGTA-PI [19] in Section 4.5.", + "bbox": [ + 212, + 508, + 782, + 625 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. Following [19,67], we employ Inc-v3 and Res-152 as surrogate models and adopt the perturbation budget $\\epsilon$ 16/255. We conduct 10 epochs of generator training using a learning rate of 2e-4. During the masked fine-tuning, we maintain the learning rate and apply a mask ratio of 0.2 to fine-tune the text-conditional generator for an additional 5 epochs.", + "bbox": [ + 212, + 628, + 782, + 700 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Multi-Target Transferability Evaluation", + "text_level": 1, + "bbox": [ + 215, + 724, + 588, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To align the experimental setup to former works [19,67], we first target 8 different classes in [71] and provide the average attack success rates (ASR) of the 8 target categories as evaluation metrics. Attack performance under larger numbers of target categories is presented in the next section.", + "bbox": [ + 212, + 750, + 782, + 806 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Attacks against regularly trained models. We first perform attacks on normal models to evaluate the multi-target attack performance. By observing", + "bbox": [ + 212, + 810, + 782, + 838 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "CLIP-Guided Networks for Transferable Targeted Attacks", + "bbox": [ + 344, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/1a76f83a01e4d0b78b347411fdfedb2577e6f926a5b751400d6f45ca9aeacbe6.jpg", + "table_caption": [ + "Table 1: Attack success rates (\\%) for multi-target attacks against regularly trained models on ImageNet NeurIPS validation set. * represents white-box attacks." + ], + "table_footnote": [], + "table_body": "
SourceMethodInc-v3Inc-v4Inc-Res-v2Res-152DN-121GoogleNetVGG-16
Inc-v3MIM99.90*0.801.000.400.200.200.30
TI-MIM98.50*0.500.500.300.200.400.40
SI-MIM99.80*1.502.000.800.700.700.50
DIM95.60*2.700.500.801.100.400.80
TI-DIM96.00*1.101.200.500.500.500.80
SI-DIM90.20*3.804.402.002.201.701.40
Logit99.60*5.606.501.703.000.801.50
SU99.59*5.807.003.353.502.003.94
C-GSP93.40*46.5836.7441.6046.4040.0045.00
CGNC96.03*59.4348.0642.4862.9851.3352.54
Res-152MIM0.500.400.6099.70*0.300.300.20
TI-MIM0.300.300.3096.50*0.300.400.30
SI-MIM1.301.201.6099.50*1.001.400.70
DIM2.302.203.0092.30*0.200.800.70
TI-DIM0.800.701.0090.60*0.600.800.50
SI-DIM4.204.805.4090.50*4.203.602.00
Logit10.1010.7012.8095.70*12.703.709.20
SU12.3611.3116.1695.08*16.136.5514.28
C-GSP37.7033.3320.2893.20*64.2041.7045.90
CGNC53.3951.5334.2495.85*85.6662.2363.36
", + "bbox": [ + 222, + 185, + 777, + 412 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "the results in Table 1, we demonstrate that the proposed CGNC consistently achieves significant improvement compared with previous methods. Specifically, our method achieves an average improvement of $17.88\\%$ and $10.08\\%$ in ASR over the C-GSP [67] attack regarding Res-152 and Inc-v3 as surrogate models respectively, demonstrating the superiority of leveraging the rich prior knowledge from CLIP's text embedding. We also note that the iterative methods obtain nearly $100\\%$ while-box ASR while receiving poor performance on black-box models due to overfitting the classification boundaries of surrogate models.", + "bbox": [ + 212, + 443, + 784, + 564 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Attacks under defense strategies. For a more thorough analysis and comparison, we then compare these attacks under several widely used defenses. Firstly, we consider attacking six robustly trained networks in Table 2. Generally, our method is still able to deceive the black-box classifiers into predicting the specified classes, significantly outperforming former multi-target attacks, e.g., a $20.87\\%$ increase of fooling rate from Res-152 to $\\mathrm{Res}50_{\\mathrm{Aug}}$ .", + "bbox": [ + 212, + 566, + 785, + 656 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Next, we evaluate the performance of different approaches on models with input preprocessing-based defenses, including a set of image smoothing mechanisms [10] and JPEG compression [14] algorithms. As shown in Table 3, although these defenses eliminate certain valid information in the adversarial samples, the CLIP-empowered CGNC still demonstrates excellent capability in boosting targeted transferability. Particularly when the substitute model is Res-152, CGNC achieves an average fooling rate of $71.18\\%$ and $80.34\\%$ on smoothing methods and JPEG compression, while C-GSP only reaches $49.05\\%$ and $56.16\\%$ respectively, verifying the stability and robustness of the proposed method.", + "bbox": [ + 212, + 657, + 785, + 792 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Perturbation Visualization. We present visualization results in Fig. 4 to unveil the principle of our method. Upon careful inspection, we can see that the generated perturbations mainly focus on the semantic regions of the input", + "bbox": [ + 212, + 795, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H. Fang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/1a4679cee41f507ea71016cc62cea60e88523894bedddc3770a00ac3e8fc6d1e.jpg", + "table_caption": [ + "Table 2: Comparison of the proposed CGNC with the SOTA multi-target attacks against models with robust training mechanism on ImageNet NeurIPS validation set." + ], + "table_footnote": [], + "table_body": "
SourceMethodInc-v3advIR-v2ensRes50SINRes50INRes50fineRes50Aug
Inc-v3MIM0.160.100.200.270.440.19
TI-MIM0.210.190.330.490.680.31
SI-MIM0.130.190.260.430.630.29
DIM0.110.090.160.330.390.19
TI-DIM0.150.130.160.210.330.14
SI-DIM0.190.210.430.710.840.46
Logit0.300.300.701.233.140.86
SU0.490.410.841.753.551.04
C-GSP20.4118.046.9633.7644.5621.95
CGNC24.3622.548.8540.8352.1822.85
Res-152MIM0.190.150.281.582.750.78
TI-MIM0.610.730.502.514.751.76
SI-MIM0.240.240.390.660.840.36
DIM0.630.370.948.5014.223.77
TI-DIM0.230.300.280.761.490.49
SI-DIM0.710.710.752.733.891.37
Logit1.151.181.656.7015.465.93
SU2.121.201.957.5321.146.95
C-GSP14.6016.0116.8460.3065.5142.88
CGNC22.2126.7129.8379.8084.0563.75
", + "bbox": [ + 253, + 184, + 743, + 406 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/8c74f5fd0cd428b79376a4555392ab97ee0a2be21d4db40ec4042d61241d3597.jpg", + "table_caption": [ + "Table 3: Comparison of our method with C-GSP under different defense strategies. Q indicates the quality factor in JPEG compression. Here the target model is DN-121 and results for more victim models are in the Appendix." + ], + "table_footnote": [], + "table_body": "
SourceMethodSmoothingJPEG compression
GaussianMediumAverageQ=70Q=75Q=80Q=85Q=90
Inc-v3C-GSP36.5946.9139.8649.4150.5551.9552.8453.39
CGNC43.0355.3545.9458.3559.2560.2861.2961.94
Res-152C-GSP43.2156.3847.5553.0354.5456.1357.7559.36
CGNC64.4479.8469.2577.3878.8180.5482.0582.94
", + "bbox": [ + 240, + 474, + 761, + 559 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "images and contain sufficient semantic patterns specific to the target category. For instance, when the condition is a photo of a sea lion, the resulting perturbations indeed carry rich patterns closely resembling this marine animal. We also observe that the pattern changes in accordance with the text prompts, which validates our conditioning mechanism of the CLIP-encoded textual embedding.", + "bbox": [ + 215, + 589, + 784, + 662 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Evaluation on More Scenarios", + "text_level": 1, + "bbox": [ + 217, + 690, + 506, + 704 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Cross-Domain scenarios. Next, we explore the more realistic cross-domain scenarios [48, 49] where attackers do not know anything about the data distribution of the training set used by the black-box classifier. Based on this setting, attackers train the generator and generate adversarial perturbations using an auxiliary dataset that follows a different probability distribution from that of the target model. Specifically, we satisfy the cross-domain experimental setting using MS-COCO [41] and Comics [4] datasets respectively. MS-COCO is a large-scale image dataset widely used for object detection and semantic segmentation,", + "bbox": [ + 215, + 719, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "CLIP-Guided Networks for Transferable Targeted Attacks", + "bbox": [ + 344, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/cf93e2fe4dc4048988469f40089162bb690ec6d1950e0b9c3b4e8c9a9315f521.jpg", + "image_caption": [ + "Fig. 4: Visualization results of different input images for different targets. For each text prompt of the target class, the left column shows the perturbation generated by our CGNC while the right column displays the corresponding adversarial examples." + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 772, + 324 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b7330cdcc02e3b51f99848fca2ac8c5c2d3a8d0e7ef9979bd59b5067be897873.jpg", + "image_caption": [ + "Fig.5: Fooling rates (from Res-152 to VGG-16) in attacking 8 target classes on cross-domain scenarios. We also provide the results on ImageNet as a comparison." + ], + "image_footnote": [], + "bbox": [ + 238, + 391, + 472, + 527 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/587f7b2c655aa79a529e370bb3b4f35a4ca3652f7575a9eb6ebfa8e5b62aa964.jpg", + "image_caption": [ + "Fig.6: Fooling rates of our CGNC with C-GSP on larger numbers of attacked classes regarding Res-152 and Inc-v3 as surrogate models." + ], + "image_footnote": [], + "bbox": [ + 531, + 391, + 764, + 527 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "while Comic is composed of a large number of comic images that can be regarded as a stylized version of ImageNet.", + "bbox": [ + 212, + 628, + 785, + 657 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To analyze the cross-domain transferability, we randomly select 1000 images from the source dataset to craft adversarial samples using the trained generator. Fig. 5 shows the results of the cross-domain attacks transferring from three different source datasets to ImageNet. In general, the more challenging cross-domain datasets lead to varying degrees of reduction in the targeted fooling rate due to the domain gap, especially on the Comics dataset which differs greatly from the ImageNet distribution. Benefiting from the text guidance on the target category, CGNC still attains decent attack performance and remarkably outperforms C-GSP. This demonstrates that our method achieves better cross-domain transferability and is partly independent of the training dataset.", + "bbox": [ + 212, + 657, + 787, + 809 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In addition, since the SOTA single-target attacks [19,49,62] require samples of the target class for loss computation, they are not applicable in the cross-", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H. Fang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/b7e086230954a2d9486b04ffa751717093b81f96bb610c451a0e511f4e5ab735.jpg", + "table_caption": [ + "Table 4: Ablation study of CGNC and its two variants on ImageNet NeurIPS validation set for 8 target classes. The substitute model is Res-152." + ], + "table_footnote": [], + "table_body": "
MethodVGG-16GoogleNetInc-v3Res-152DN-201
CGNA-CA-t56.5551.0947.4492.8174.65
CGNA-CA56.6454.2949.7393.3475.99
CGNC63.3662.2353.3995.8582.69
", + "bbox": [ + 272, + 184, + 730, + 250 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "domain scenarios where the source dataset lacks images of the target category. This represents an additional advantage of our method over single-target attacks.", + "bbox": [ + 212, + 273, + 785, + 301 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Larger Numbers of Target Classes. We then increase the number of target categories to verify the effectiveness of our method when handling plenty of classes. Specifically, we condition CGNC with more text inputs corresponding to the increased number of target categories. The performance is evaluated across the aforementioned six black-box models.", + "bbox": [ + 212, + 304, + 785, + 378 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As mentioned before, single-target attacks become impractical for real-world classification systems with hundreds or thousands of target categories [27]. In comparison, our method effectively solves this issue and achieves great improvements in large numbers of target classes over C-GSP as in Fig. 6, e.g., $14.71\\%$ increase for 200 target classes proxy on the Res-152. This again reveals the significance of the fully exploited textual guidance from the CLIP model. Additionally, our method exhibits greater robustness to the varying numbers of target classes, e.g., the performance of our network on Inc-v3 demonstrates a smoother decrease compared to that of C-GSP. This further highlights the superiority of our method in reducing computation costs. For instance, when launching 1000-class targeted attacks using Inc-v3 as the surrogate model, our method requires training only 5 CGNCs with 200 conditions, yet it can achieve comparable performance to 20 generators with 50 conditions trained using C-GSP.", + "bbox": [ + 212, + 380, + 787, + 575 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 595, + 388, + 609 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We conduct ablation experiments on the ImageNet NeurIPS dataset to study the effect of the proposed techniques. Specifically, we introduce two variants of CGNC. CGNC-CA removes the cross-attention module from the original network, and CGNC-CA-t further modifies the conditioning mechanism by replacing the CLIP's text embedding with one-hot labels.", + "bbox": [ + 212, + 617, + 785, + 691 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "From the results in Table 4, we can find that each aforementioned technique can further improve the attack success rates. Moreover, the remarkable improvement from CGNC-CA to CGNC also confirms that these cross-attention modules in the CA-Decoder can make better use of the text guidance provided by the CLIP model to enhance the targeted transferability of crafted perturbations.", + "bbox": [ + 212, + 691, + 785, + 768 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Comparison with Single-Target Attacks", + "text_level": 1, + "bbox": [ + 215, + 787, + 586, + 801 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Next, we compare the single-target variants of the proposed CGNC with multiple state-of-the-art single-target attacks. To obtain our single-target generators, we", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "CLIP-Guided Networks for Transferable Targeted Attacks", + "bbox": [ + 344, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/649d56f21cb00169fdedfb3e1535d8f784f44fc5c86eed0d31884ebaa5671871.jpg", + "table_caption": [ + "Table 5: Comparison of our method with the SOTA single-target attacks. $^\\dagger$ denotes the single-target variant of our CGNC obtained through the masked fine-tuning technique. * represents the white-box attacks." + ], + "table_footnote": [], + "table_body": "
SourceMethodInc-v3Inc-v4Inc-Res-v2Res-152DN-121GoogleNetVGG-16
Inc-v3GAP86.90*45.0634.4834.4841.7426.8934.34
CD-AP94.20*57.6060.1037.1041.6032.3041.70
TTP91.37*46.0439.3716.4033.4725.8025.73
DGTA-PI94.63*67.9555.0350.5047.3847.6748.11
CGNC†98.84*74.7664.4862.0078.9469.0670.74
Res-152GAP30.9931.4320.4884.86*58.3529.8939.70
CD-AP33.3043.7042.7096.60*53.8036.6034.10
TTP62.0349.2038.7095.12*82.9665.0962.82
DGTA-PI66.8353.6247.6196.48*86.6168.2969.58
CGNC†68.8669.4545.7198.61*91.1469.8368.05
", + "bbox": [ + 228, + 199, + 771, + 320 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "conduct the proposed masked fine-tuning to the trained conditional generator eight times using eight different text prompts of the target classes. Quantitative results in Table 5 indicate that our single-target enhanced generators outperform the competing single-target attacks in most cases. Particularly on the surrogate model of Inc-v3, $\\mathrm{CGNC}^{\\dagger}$ achieves a notable increase of $15.36\\%$ in the average black-box fooling rate compared to previous methods, demonstrating the significant effectiveness of the proposed masked fine-tuning. The MFT technique further enhances the single-target performance, thereby improving the adaptability and flexibility of our approach in scenarios with fewer target classes.", + "bbox": [ + 212, + 347, + 785, + 483 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Note that these single-target methods require training 8 generators from scratch for 8 different target classes. In contrast, our method only needs to train a single multi-target generator and perform fine-tuning 8 times, each time with just a few epochs. Based on the typical experimental setup used in the SOTA single-target methods [19,49], our strategy can diminish over 100 training epochs when targeting 8 classes, thus substantially mitigating the computational burden. More experiments of single-target attacks against adversarially robust models and input preprocessing defenses are shown in the Appendix.", + "bbox": [ + 212, + 484, + 787, + 604 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 627, + 359, + 643 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we design a novel generative network CGNC, which improves multi-target transferable adversarial attacks by fully utilizing the rich prior within the CLIP as auxiliary semantic knowledge about the target category. To better incorporate the prior information, we introduce the cross-attention modules and efficiently condition the generator with CLIP's text embeddings. Through extensive experiments, we demonstrate the effectiveness of the proposed CGNC on multiple black-box target models in a variety of scenarios. Moreover, we propose a masked fine-tuning technique to further enhance the CGNC in attacking a single class, which outperforms previous single-target methods in both efficiency and effectiveness. We hope that the proposed method can serve as a reliable tool for evaluating the model robustness under black-box setups, promoting further research on the vulnerability and robustness of DNNs.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H. Fang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported in part by the National Natural Science Foundation of China under grant 62171248, 62301189, Guangdong Basic and Applied Basic Research Foundation under grant 2021A1515110066, the PCNL KEY project (PCL2021A07), and Shenzhen Science and Technology Program under Grant JCYJ20220818101012025, RCBS20221008093124061, GXWD20220811172936001.", + "bbox": [ + 215, + 172, + 792, + 248 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 268, + 321, + 285 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Akhtar, N., Mian, A.: Threat of adversarial attacks on deep learning in computer vision: A survey. IEEE Access 6, 14410-14430 (2018)", + "2. Andriushchenko, M., Croce, F., Flammarion, N., Hein, M.: Square attack: a query-efficient black-box adversarial attack via random search. In: European conference on computer vision. pp. 484-501. Springer (2020)", + "3. Athalye, A., Engstrom, L., Ilyas, A., Kwok, K.: Synthesizing robust adversarial examples. In: International conference on machine learning. pp. 284-293. PMLR (2018)", + "4. BircanoAYlu, C.: https://www.kaggle.com/datasets/cenkbircanoglu/comic-books-classification. Kaggle, 2017", + "5. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: 2017 ieee symposium on security and privacy (sp). pp. 39-57. IEEE (2017)", + "6. Che, Z., Borji, A., Zhai, G., Ling, S., Li, J., Min, X., Guo, G., Le Callet, P.: Smgea: A new ensemble adversarial attack powered by long-term gradient memories. IEEE Transactions on Neural Networks and Learning Systems 33(3), 1051-1065 (2020)", + "7. Chefer, H., Benaim, S., Paiss, R., Wolf, L.: Image-based clip-guided essence transfer. In: European Conference on Computer Vision. pp. 695-711. Springer (2022)", + "8. Chen, B., Feng, Y., Dai, T., Bai, J., Jiang, Y., Xia, S.T., Wang, X.: Adversarial examples generation for deep product quantization networks on image retrieval. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(2), 1388-1404 (2022)", + "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)", + "0. Ding, G.W., Wang, L., Jin, X.: Advertorch v0.1: An adversarial robustness toolbox based on pytorch. arXiv preprint arXiv:1902.07623 (2019)", + "1. Dong, Y., Liao, F., Pang, T., Su, H., Zhu, J., Hu, X., Li, J.: Boosting adversarial attacks with momentum. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9185-9193 (2018)", + "2. Dong, Y., Pang, T., Su, H., Zhu, J.: Evading defenses to transferable adversarial examples by translation-invariant attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4312-4321 (2019)", + "3. Du, J., Zhang, H., Zhou, J.T., Yang, Y., Feng, J.: Query-efficient meta attack to deep neural networks. arXiv preprint arXiv:1906.02398 (2019)", + "4. Dziugaite, G.K., Ghahramani, Z., Roy, D.M.: A study of the effect of jpg compression on adversarial images. arXiv preprint arXiv:1608.00853 (2016)", + "5. Eykholt, K., Evtimov, I., Fernandes, E., Li, B., Rahmati, A., Xiao, C., Prakash, A., Kohno, T., Song, D.: Robust physical-world attacks on deep learning visual classification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1625-1634 (2018)" + ], + "bbox": [ + 225, + 296, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "CLIP-Guided Networks for Transferable Targeted Attacks", + "bbox": [ + 344, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Fang, H., Chen, B., Wang, X., Wang, Z., Xia, S.T.: Gidf: A generative gradient inversion method with feature domain optimization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4967-4976 (2023)", + "17. Fang, H., Kong, J., Yu, W., Chen, B., Li, J., Xia, S., Xu, K.: One perturbation is enough: On generating universal adversarial perturbations against vision-language pre-training models. arXiv preprint arXiv:2406.05491 (2024)", + "18. Fang, H., Qiu, Y., Yu, H., Yu, W., Kong, J., Chong, B., Chen, B., Wang, X., Xia, S.T.: Privacy leakage on dnns: A survey of model inversion attacks and defenses. arXiv preprint arXiv:2402.04013 (2024)", + "19. Feng, W., Xu, N., Zhang, T., Zhang, Y.: Dynamic generative targeted attacks with pattern injection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16404-16414 (2023)", + "20. Gao, K., Bai, Y., Bai, J., Yang, Y., Xia, S.T.: Adversarial robustness for visual grounding of multimodal large language models. In: ICLR Workshop (2024)", + "21. Gao, K., Bai, Y., Gu, J., Xia, S.T., Torr, P., Li, Z., Liu, W.: Inducing high energy-latency of large vision-language models with verbose images. In: ICLR (2024)", + "22. Gao, K., Gu, J., Bai, Y., Xia, S.T., Torr, P., Liu, W., Li, Z.: Energy-latency manipulation of multi-modal large language models via verbose samples. arXiv preprint arXiv:2404.16557 (2024)", + "23. Gao, L., Cheng, Y., Zhang, Q., Xu, X., Song, J.: Feature space targeted attacks by statistic alignment. arXiv preprint arXiv:2105.11645 (2021)", + "24. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231 (2018)", + "25. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014)", + "26. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014)", + "27. Han, J., Dong, X., Zhang, R., Chen, D., Zhang, W., Yu, N., Luo, P., Wang, X.: Once a man: Towards multi-target attack via learning multi-target adversarial network once. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5158-5167 (2019)", + "28. Hang, J., Han, K., Chen, H., Li, Y.: Ensemble adversarial black-box attacks against deep learning systems. Pattern Recognition 101, 107184 (2020)", + "29. He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14. pp. 630-645. Springer (2016)", + "30. He, Z., Wang, W., Dong, J., Tan, T.: Revisiting ensemble adversarial attack. Signal Processing: Image Communication 107, 116747 (2022)", + "31. Hendrycks, D., Mu, N., Cubuk, E.D., Zoph, B., Gilmer, J., Lakshminarayanan, B.: Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781 (2019)", + "32. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017)", + "33. Huang, Z., Zhou, A., Ling, Z., Cai, M., Wang, H., Lee, Y.J.: A sentence speaks a thousand images: Domain generalization through distilling clip with language guidance. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11685-11695 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H. Fang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Inkawich, N., Liang, K., Wang, B., Inkawich, M., Carin, L., Chen, Y.: Perturbing across the feature hierarchy to improve standard and strict blackbox attack transferability. Advances in Neural Information Processing Systems 33, 20791-20801 (2020)", + "35. Inkawich, N., Wen, W., Li, H.H., Chen, Y.: Feature space perturbations yield more transferable adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7066-7074 (2019)", + "36. Ju, C., Han, T., Zheng, K., Zhang, Y., Xie, W.: Prompting visual-language models for efficient video understanding. In: European Conference on Computer Vision. pp. 105-124. Springer (2022)", + "37. Kong, Z., Guo, J., Li, A., Liu, C.: Physgan: Generating physical-world-resilient adversarial examples for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14254-14263 (2020)", + "38. Kurakin, A., Goodfellow, I., Bengio, S.: Adversarial machine learning at scale. arXiv preprint arXiv:1611.01236 (2016)", + "39. Li, Q., Guo, Y., Chen, H.: Yet another intermediate-level attack. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16. pp. 241-257. Springer (2020)", + "40. Lin, J., Song, C., He, K., Wang, L., Hopcroft, J.E.: Nesterov accelerated gradient and scale invariance for adversarial attacks. arXiv preprint arXiv:1908.06281 (2019)", + "41. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "42. Liu, A., Wang, J., Liu, X., Cao, B., Zhang, C., Yu, H.: Bias-based universal adversarial patch attack for automatic check-out. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16. pp. 395-410. Springer (2020)", + "43. Lu, Y., Jia, Y., Wang, J., Li, B., Chai, W., Carin, L., Velipasalar, S.: Enhancing cross-task black-box transferability of adversarial examples with dispersion reduction. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 940-949 (2020)", + "44. Luo, J., Bai, T., Zhao, J.: Generating adversarial yet inconspicuous patches with a single image (student abstract). In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 15837-15838 (2021)", + "45. Moosavi-Dezfooli, S.M., Fawzi, A., Fawzi, O., Frossard, P.: Universal adversarial perturbations. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1765-1773 (2017)", + "46. Moosavi-Dezfooli, S.M., Fawzi, A., Frossard, P.: Deepfool: a simple and accurate method to fool deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2574-2582 (2016)", + "47. Mopuri, K.R., Uppala, P.K., Babu, R.V.: Ask, acquire, and attack: Data-free uap generation using class impressions. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 19-34 (2018)", + "48. Naseer, M.M., Khan, S.H., Khan, M.H., Shahbaz Khan, F., Porikli, F.: Cross-domain transferability of adversarial perturbations. Advances in Neural Information Processing Systems 32 (2019)", + "49. Naseer, M., Khan, S., Hayat, M., Khan, F.S., Porikli, F.: On generating transferable targeted perturbations. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7708-7717 (2021)" + ], + "bbox": [ + 215, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "CLIP-Guided Networks for Transferable Targeted Attacks", + "bbox": [ + 344, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "50. NeurIPS: https://www.kaggle.com/c/nips-2017-defense-against-adversarial-attack/data.Kaggle,2017", + "51. Poursaeed, O., Katsman, I., Gao, B., Belongie, S.: Generative adversarial perturbations. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4422-4431 (2018)", + "52. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "53. Sarkar, S.K., Oshiba, K., Giebisch, D., Singer, Y.: Robust classification of financial risk. arXiv preprint arXiv:1811.11079 (2018)", + "54. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)", + "55. Szegedy, C., Ioffe, S., Vanhoucke, V., Alemi, A.: Inception-v4, inception-resnet and the impact of residual connections on learning. In: Proceedings of the AAAI conference on artificial intelligence. vol. 31 (2017)", + "56. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A.: Going deeper with convolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1-9 (2015)", + "57. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2818-2826 (2016)", + "58. Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I., Fergus, R.: Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199 (2013)", + "59. Tao, M., Bao, B.K., Tang, H., Xu, C.: Galip: Generative adversarial clips for text-to-image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14214-14223 (2023)", + "60. Wang, X., He, K.: Enhancing the transferability of adversarial attacks through variance tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1924-1933 (2021)", + "61. Wang, X., He, K., Hopcroft, J.E.: At-gan: A generative attack model for adversarial transferring on generative adversarial nets. arXiv preprint arXiv:1904.07793 3(4), 3 (2019)", + "62. Wang, Z., Yang, H., Feng, Y., Sun, P., Guo, H., Zhang, Z., Ren, K.: Towards transferable targeted adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20534-20543 (2023)", + "63. Wei, Z., Chen, J., Wu, Z., Jiang, Y.G.: Enhancing the self-universality for transferable targeted attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12281-12290 (2023)", + "64. Xiao, C., Li, B., Zhu, J.Y., He, W., Liu, M., Song, D.: Generating adversarial examples with adversarial networks. arXiv preprint arXiv:1801.02610 (2018)", + "65. Xie, C., Zhang, Z., Zhou, Y., Bai, S., Wang, J., Ren, Z., Yuille, A.L.: Improving transferability of adversarial examples with input diversity. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2730-2739 (2019)", + "66. Xiong, Y., Lin, J., Zhang, M., Hopcroft, J.E., He, K.: Stochastic variance reduced ensemble adversarial attack for boosting the adversarial transferability. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14983-14992 (2022)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "H. Fang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "67. Yang, X., Dong, Y., Pang, T., Su, H., Zhu, J.: Boosting transferability of targeted adversarial examples via hierarchical generative networks. In: European Conference on Computer Vision. pp. 725-742. Springer (2022)", + "68. Yu, W., Fang, H., Chen, B., Sui, X., Chen, C., Wu, H., Xia, S.T., Xu, K.: Gi-nas: Boosting gradient inversion attacks through adaptive neural architecture search. arXiv preprint arXiv:2405.20725 (2024)", + "69. Zha, Y., Ji, H., Li, J., Li, R., Dai, T., Chen, B., Wang, Z., Xia, S.T.: Towards compact 3d representations via point feature enhancement masked autoencoders. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 6962-6970 (2024)", + "70. Zha, Y., Wang, J., Dai, T., Chen, B., Wang, Z., Xia, S.T.: Instance-aware dynamic prompt tuning for pre-trained point cloud models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14161-14170 (2023)", + "71. Zhang, C., Benz, P., Imtiaz, T., Kweon, I.S.: Understanding adversarial examples from the mutual influence of images and perturbations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14521-14530 (2020)", + "72. Zhao, Z., Liu, Z., Larson, M.: On success and simplicity: A second look at transferable targeted attacks. Advances in Neural Information Processing Systems 34, 6115-6128 (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 424 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "CLIP-Guided Networks for Transferable Targeted Attacks", + "bbox": [ + 344, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/CLIP-Guided Generative Networks for Transferable Targeted Adversarial Attacks/ddf47f95-f1c1-405a-8323-f8a66b1071ea_model.json b/2024/CLIP-Guided Generative Networks for Transferable Targeted Adversarial Attacks/ddf47f95-f1c1-405a-8323-f8a66b1071ea_model.json new file mode 100644 index 0000000000000000000000000000000000000000..599618c2c4447e79099476b4e6aae504087b98f4 --- /dev/null +++ b/2024/CLIP-Guided Generative Networks for Transferable Targeted Adversarial Attacks/ddf47f95-f1c1-405a-8323-f8a66b1071ea_model.json @@ -0,0 +1,2581 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.26, + 0.141, + 0.745, + 0.187 + ], + "angle": 0, + "content": "CLIP-Guided Generative Networks for Transferable Targeted Adversarial Attacks" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.211, + 0.756, + 0.245 + ], + "angle": 0, + "content": "Hao Fang\\(^{1\\dagger}\\), Jiawei Kong\\(^{2\\dagger}\\), Bin Chen\\(^{2,3,4\\#}\\), Tao Dai\\(^{5}\\), Hao Wu\\(^{6}\\), and Shu-Tao Xia\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.255, + 0.754, + 0.283 + ], + "angle": 0, + "content": "\\(^{1}\\) Tsinghua Shenzhen International Graduate School, Tsinghua University \\(^{2}\\) Harbin Institute of Technology, Shenzhen \\(^{3}\\) Pengcheng Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.283, + 0.78, + 0.354 + ], + "angle": 0, + "content": "4 Guangdong Provincial Key Laboratory of Novel Security Intelligence Technologies \n5 Shenzhen University 6 Shenzhen Digital Certificate Authority CO., Ltd \nfang-h23@mails.tsinghua.edu.cn, kongjiawei@stu.hit.edu.cn \nchenbin2021@hit.edu.cn, daitao.edu@gmail.com, whpc79@163.com, \nxiast@sz.tsinghua.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.389, + 0.745, + 0.613 + ], + "angle": 0, + "content": "Abstract. Transferable targeted adversarial attacks aim to mislead models into outputting adversary-specified predictions in black-box scenarios. Recent studies have introduced single-target attacks that train a generator for each target class to generate highly transferable perturbations, resulting in substantial computational overhead when handling multiple classes. Multi-target attacks address this by training only one class-conditional generator for multiple classes. However, the generator simply uses class labels as conditions, failing to leverage the rich semantic information of the target class. To this end, we design a CLIP-guided Generative Network with Cross-attention modules (CGNC) to enhance multi-target attacks by incorporating textual knowledge of CLIP into the generator. Extensive experiments demonstrate that CGNC yields significant improvements over previous multi-target attacks, e.g., a \\(21.46\\%\\) improvement in success rate from Res-152 to DenseNet-121. Moreover, we propose the masked fine-tuning to further strengthen our method in attacking a single class, which surpasses existing single-target methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.636, + 0.378, + 0.652 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.788, + 0.804 + ], + "angle": 0, + "content": "With the rapid progress of deep learning, deep neural networks (DNNs) have been widely applied in many security-critical fields, such as autonomous driving [15, 37], financial systems [53], and point cloud modeling [69, 70]. However, DNNs are corroborated to be vulnerable to adversarial attacks [20, 26, 58], which attempt to fool models with adversarial examples crafted by adding imperceptible perturbations to the original inputs. Based on the attack goal, adversarial attacks can be categorized into untargeted and targeted attacks. Untargeted attacks attempt to fool DNNs into predicting incorrect labels while targeted attacks aim at triggering the model to output the attacker-desired predictions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.811, + 0.371, + 0.825 + ], + "angle": 0, + "content": "\\(\\dagger\\)Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.825, + 0.396, + 0.84 + ], + "angle": 0, + "content": "Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.233, + 0.811, + 0.396, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "H. Fang et al." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.147, + 0.554, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.273, + 0.413, + 0.284 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.564, + 0.164, + 0.756, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.272, + 0.69, + 0.284 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.296, + 0.789, + 0.409 + ], + "angle": 0, + "content": "Fig. 1: (a) Targeted attacks from 'Panda' to 'Dog'. Fig. (a) illustrates that previous multi-target methods [27,67] generate perturbations simply conditioned by class indices or one-hot vectors and only learn the classification boundary specific to the surrogate model. In contrast, CGNC exploits CLIP's meaningful guidance to effectively capture the feature distribution inherent to the target data. (b) We directly feed the scaled perturbations generated by our CGNC and C-GSP [67] into three black-box classifiers and reveal that our perturbations achieve significantly higher mean confidence of the target class, demonstrating its superiority in modeling the target feature distribution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.789, + 0.717 + ], + "angle": 0, + "content": "Recent investigations into the adversarial transferability [2, 3, 44, 58] have demonstrated that adversarial examples crafted for a white-box surrogate model can also mislead other unseen black-box models. Since this attack does not require access to the target model, it exposes a serious security threat to real-world applications of DNNs and motivates a wide range of studies [1,5,42,46]. Despite the remarkable performance on untargeted transferable attacks, these approaches produce unsatisfactory results for targeted attacks due to their over-reliance on the white-box surrogate. Existing studies on transferable targeted attacks can be categorized into instance-specific [11, 15, 23, 39, 43, 60, 66] and instance-agnostic attacks [19, 37, 44, 48, 49, 64]. Specifically, instance-specific attacks [11, 12, 65] iteratively perform gradient updating to craft adversarial perturbation tailored to a specific natural sample, yet producing low-targeted transferability due to overfitting [67]. Conversely, instance-agnostic attacks learn a universal perturbation [45, 71] or a perturbation generator [48, 51] based on data distribution rather than the specific instance, alleviating the data-specific over-fitting issues and achieving a higher adversarial transferability. Recent studies further explore generative attacks to produce highly transferable perturbations, which can be divided into single-target and multi-target attacks as follows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Single-target attacks [19, 48, 49, 62] exhibit impressive performance by training perturbation generators for target categories. However, they require training a generator for each target, which can lead to a heavy computation burden when attacking numerous classes, and thus are not applicable in real-world classification systems that usually contain hundreds/thousands of target classes [27]. To address this, [27, 67] propose multi-target attacks that train a single conditional generator for multi-target classes. By specifying the desired label as conditioning input, the trained generator can efficiently generate the corresponding adversar" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.345, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "CLIP-Guided Networks for Transferable Targeted Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "ial perturbation. Nevertheless, these methods simply adopt class indices [27] or one-hot vectors [67] of the target labels as conditions, and thus solely rely on the classification information from the surrogate model as the guidance of the target category, hence resulting in only modest black-box fooling rates." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.463 + ], + "angle": 0, + "content": "In this paper, we build upon the research line of multi-target attacks and propose a novel CLIP-guided Generative Network with Cross-attention modules (CGNC) to boost the attack effect. Concretely, we revisit the architecture of the conditional generators used in [27,67] and argue that the simple class conditions, e.g., class indices or handcrafted one-hot vectors, limit the transferability of the generated perturbations (see Fig. 1(a) for comparison). Motivated by the impressive effects of CLIP's encoded text embeddings in multi-modal learning, we introduce concise text descriptions of the target classes to encode them as class-specific representations encapsulated with abundant information, which assist the generator in learning target class distribution and ultimately lead to a fundamental transferability improvement. To better incorporate the text information, we improve the condition-inputting mechanism by adding the cross-attention layers that have been proven effective in models with various input modalities. Results in Fig. 1(b) confirm the capability of our CGNC in capturing the target distribution. Besides, we propose a masked fine-tuning (MFT) technique, which fine-tunes the trained conditional generator with a fixed text condition of the desired class for further improvement in attacking a single class." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.785, + 0.57 + ], + "angle": 0, + "content": "With all the above efforts, our method achieves great efficiency and scalability. When there are hundreds of target classes and single-target methods become impractical, the proposed conditional network can be utilized to achieve significant performance over previous multi-target attacks. Conversely, for scenarios involving only a few target classes, the proposed MFT mechanism enhances CGNC to outperform existing single-target methods while substantially reducing computational costs. In summary, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.58, + 0.784, + 0.623 + ], + "angle": 0, + "content": "- We propose CGNC, a novel CLIP-guided generative network with cross-attention layers that fully exploits the textual knowledge provided by the advanced CLIP model for enhanced multi-target attacks." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.625, + 0.784, + 0.653 + ], + "angle": 0, + "content": "- We introduce a masked fine-tuning mechanism to improve the single-target performance by adapting the CGNC to an individual target class." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.655, + 0.784, + 0.684 + ], + "angle": 0, + "content": "- Extensive experiments show that CGNC remarkably enhances the targeted transferability compared to previous attack methods in a range of settings." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.58, + 0.784, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.708, + 0.388, + 0.724 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.74, + 0.465, + 0.755 + ], + "angle": 0, + "content": "2.1 Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Vision-language models (VLM) drawn great attention [21,22] due to their promising potential in learning general visual and textual representations through contrastive pre-training on large-scale image-text pairs. Given the powerful capacity of text descriptions in modeling multi-modal tasks, these models can be effectively adapted to diverse downstream tasks through appropriately formulated" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "H. Fang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "textual prompts, including video understanding [36], image manipulation [7], and text-to-image synthesis [59]. Inspired by these works, we propose to harness CLIP-based multi-modal learning to empower our perturbation generator, facilitating more effective learning of the target feature distribution." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.234, + 0.427, + 0.249 + ], + "angle": 0, + "content": "2.2 Adversarial Attacks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.264, + 0.786, + 0.308 + ], + "angle": 0, + "content": "Among various security threats to DNNs [8, 16, 18, 68], the adversarial attack is one of the most formidable and well-known one, which can be classified into instance-specific attacks and instance-agnostic attacks [67]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.31, + 0.788, + 0.476 + ], + "angle": 0, + "content": "Instance-specific Attacks. Since the pioneering work [58] highlighted the vulnerability of neural networks, numerous gradient-based optimization methods [3, 13, 26, 38, 66] have been proposed to craft image-dependent perturbations. MIM [11] integrates a momentum item into the gradient update for utilizing the previous gradient information to avoid plunging into poor local optimum. DIM [65] enhances the transferability by randomly diversifying the input pattern, and TIM [12] implements the attack by convolving the gradient with a predefined kernel. In addition, intermediate feature space [35, 63] and classifier information [34] are exploited to enhance attack effects, while [72] leverages logit-based loss to achieve competitive results. More advanced works [6,28,30] consider an ensemble of multiple surrogate models to reduce over-fitting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.798, + 0.612 + ], + "angle": 0, + "content": "Instance-agnostic Attacks. In contrast to instance-specific attacks, instance-agnostic attacks learn a universal perturbation [17, 45] or a generative model [44, 48, 49, 51, 61, 67] for crafting adversarial examples. By modeling the global data distribution, these methods have shown better transferability and attracted more attention in recent years. Specifically, UAP [45] and AAA [47] learn a universal perturbation to fool the model based on concrete data and compressed impression, respectively. Many subsequent works, such as GAP [51], focus on boosting the transferability of adversarial attacks by training generative models. These generative attacks can be categorized into the following two types." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Single-target Attacks. This type of attack requires training a generator for each target class. [64] first introduce the generative adversarial networks (GAN) [25] to generate adversarial perturbations. Then, CD-AP [48] concentrates on the domain-invariant adversaries and launches highly transferable cross-domain attacks using a relativistic supervisory. TTP [49] modifies the loss function and proposes to match the target distribution to mitigate over-fitting to the surrogate model. Subsequent methods achieve better transferability based on a dynamic network with pattern injection [19] or a feature discriminator [62]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.792, + 0.841 + ], + "angle": 0, + "content": "Multi-target Attacks. MAN [27] notes that when dealing with numerous classes, single-target attacks inevitably suffer from extreme computational burdens as they need to train the same number of models as multiple target classes. Therefore, single-target attacks become impractical in attacking real classification systems that often have hundreds of categories. To pursue the extreme speed and storage, MAN trains only one model for 1000 target categories from ImageNet [9]. However, the excess of the target class severely degrades the targeted" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.345, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "CLIP-Guided Networks for Transferable Targeted Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.253 + ], + "angle": 0, + "content": "transferability. C-GSP [67] improves the performance by designing a hierarchical partition mechanism to divide all classes into a feasible number of subsets and train generators for each subset. Nonetheless, C-GSP simply conditions the generator with one-hot vectors and only utilizes the classification information to train the generator, failing to use the semantic knowledge of the target class. To overcome the limitation, we propose to incorporate the text information provided by CLIP as significant guidance for the target class." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.274, + 0.331, + 0.29 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.304, + 0.784, + 0.38 + ], + "angle": 0, + "content": "In this section, we first introduce the preliminaries of targeted transferable attacks and present the basic paradigm of generative attack methods. Then, we elaborate on the proposed CGNC, which remarkably enhances multi-target black-box attacks. Finally, we detail the proposed masked fine-tuning technique that strengthens our method in single-class attacks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.4, + 0.362, + 0.416 + ], + "angle": 0, + "content": "3.1 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.785, + 0.575 + ], + "angle": 0, + "content": "We denote the white-box image classifier parameterized with \\(\\theta\\) as \\(f_{\\theta}:\\mathcal{X}\\rightarrow \\mathcal{Y}\\) where \\(\\mathcal{X}\\subset \\mathbb{R}^{N\\times H\\times W}\\) represents the image domain and \\(\\mathcal{V}\\subset \\mathbb{R}^L\\) is the output confidence score of different classes \\((H,W,N,L\\) being height, width, number of channels, and class number). Given a natural image \\(\\pmb {x}\\in \\mathcal{X}\\) and the attacker's desired label \\(c_{t}\\in \\mathcal{C}\\), the transferable targeted adversarial attacks attempt to craft the imperceptible perturbation \\(\\delta\\) based on the accessible surrogate model \\(f_{\\theta}\\) to mislead another unseen victim model \\(F_{\\phi}\\) into predicting \\(c_{t}\\), i.e., \\(\\arg \\max_{i\\in \\mathcal{C}}F_{\\phi}(\\pmb {x} + \\pmb {\\delta})_i = c_t\\). Concurrently, the \\(l_{\\infty}\\) norm is employed to ensure the adversarial samples are indistinguishable from the original images by constraining the perturbation within the range \\(\\epsilon\\) by \\(\\| \\pmb {\\delta}\\|_{\\infty}\\leq \\epsilon\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.784, + 0.696 + ], + "angle": 0, + "content": "To boost the fooling rate of targeted black-box attacks, single-target attacks [48, 49, 51] utilize powerful generative models to learn the target data distribution and achieve higher transferability. However, these methods consume great computation time and resources for multi-target scenarios, making them impractical for real-world attacks. C-GSP [67] effectively solves this issue by formulating the multi-target attacks as learning a class-conditional generator \\( G_{w} \\) with weights \\( w \\), which is capable of generating perturbations for any target class. Given an unlabeled training dataset \\( \\mathcal{X}_s \\), the optimization objective is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.271, + 0.721, + 0.785, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\min _ {w} \\mathbb {E} _ {\\boldsymbol {x} _ {s} \\sim \\chi_ {s}, c \\sim \\mathcal {C}} [ \\mathcal {L} (f _ {\\theta} (\\boldsymbol {x} _ {s} + G _ {w} (\\boldsymbol {x} _ {s}, c)), c) ], \\text {s . t .} \\| G _ {w} (\\boldsymbol {x} _ {s}, c) \\| _ {\\infty} \\leq \\epsilon , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.841 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}(\\cdot, \\cdot)\\) denotes the cross-entropy (CE) loss. By minimizing the loss of specified classes using various unlabeled images from \\(\\mathcal{X}_s\\), we optimize the parameters \\(w\\) of the generative model and finally obtain a targeted conditional generator that can generate perturbation for any given clean image \\(\\pmb{x}_t\\) from the test dataset \\(\\mathcal{X}_t\\). Specifically, the adversary can simply specify a target label \\(c\\) and craft an corresponding adversarial example via \\(\\pmb{x}_{adv} = \\pmb{x}_t + G_w(\\pmb{x}_t, c)\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "H. Fang et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.149, + 0.784, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.289, + 0.788, + 0.348 + ], + "angle": 0, + "content": "Fig. 2: An overview of our proposed architecture of CGNC. Equipped with the three exquisite modules VL-Purifier, F-Encoder, and CA-Decoder, the generator fully leverages the textual representations encoded by CLIP as auxiliary information about the target classes to better probe their data distribution and achieve superior attack effects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.788, + 0.499 + ], + "angle": 0, + "content": "However, current multi-target methods [27, 67] simply condition the generator with class labels and learn the target distribution only relying on the classification information of the surrogate model, thus not fully exploiting the specific characteristics of the target category. Inspired by the efficiency of vision-language (VL) learning [52, 59], we propose a novel generative network that leverages sufficient prior knowledge from the powerful CLIP model by incorporating textual-modality information to promote the target class modeling, which greatly enhances the multi-target transferable attacks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.524, + 0.547, + 0.538 + ], + "angle": 0, + "content": "3.2 CLIP-Guided Generative Network" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.627 + ], + "angle": 0, + "content": "The proposed generative model architecture is presented in Fig. 2. Specifically, CGNC is composed of a Vision-Language feature Purifier (VL-Purifier), a feature Fusion Encoder (F-Encoder), and a Cross-Attention based Decoder (CA-Decoder). We also provide the pseudocode of the training procedure in Algorithm 1. Next, we illustrate the design of each module as follows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Vision-language feature purifier (VL-Purifier). To utilize CLIP to produce semantic embeddings, we first feed in CLIP's text encoder \\(\\varPhi\\) with queries \\(t_c\\) that follow the handcrafted template \"a photo of a {class}\" which has shown effectiveness in many tasks [33,52]. Since the obtained embeddings \\(\\pmb{e}_t\\in \\mathbb{R}^{B\\times 512}\\) (\\(B\\) being the batch size) in CLIP's vision-language space are generic representations of the target classes and are not yet tailored to our learning task, we refine them using the VL-Purifier, which is composed of several blocks consisting of a fully-connected layer and a spectral normalization layer. Through this module, we translate the encoded embeddings into more meaningful representations \\(\\pmb{e}_t^*\\in \\mathbb{R}^{B\\times 16}\\), thereby facilitating the subsequent step of feature fusion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Feature fusion encoder (F-Encoder). This module aims to fuse the purified features \\( \\pmb{e}_t^* \\) with the learned visual representations. Firstly, a batch of input images \\( \\pmb{x}_s \\) is encoded to capture the visual concepts \\( \\pmb{h}_s \\in \\mathbb{R}^{B \\times N' \\times H' \\times W'} \\). Then, we expand the text embedding \\( \\pmb{e}_t^* \\in \\mathbb{R}^{B \\times 16} \\) into \\( \\pmb{e}_t^{*'} \\in \\mathbb{R}^{B \\times 16 \\times H' \\times W'} \\), which are" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.345, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "CLIP-Guided Networks for Transferable Targeted Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "code_caption", + "bbox": [ + 0.217, + 0.146, + 0.768, + 0.163 + ], + "angle": 0, + "content": "Algorithm 1 Pseudocode of Training the CLIP-guided Generative Network" + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.165, + 0.787, + 0.363 + ], + "angle": 0, + "content": "Require: \\(\\mathcal{X}_s\\): the training data; \\(\\mathcal{C}\\): the target label space; \\(\\mathcal{T}\\): the text prompts set; \\(f_{\\theta}\\): the surrogate model; \\(\\varPhi\\): the CLIP's text encoder; \\(N\\): the max iteration; \nEnsure: the perturbation generator \\(G_w\\); \n1: for \\(i \\gets 0\\) to \\(N\\) do \n2: Sample a batch of images \\(\\boldsymbol{x}_s \\sim \\mathcal{X}_s\\); \n3: Obtain \\(\\boldsymbol{x}_s'\\) by processing \\(\\boldsymbol{x}_s\\) with data augmentation; \n4: Sample a batch of target labels \\(c \\sim \\mathcal{C}\\); \n5: Obtain the corresponding text prompts \\(t_c\\) from \\(\\mathcal{T}\\); \n6: Compute the text embedding \\(\\boldsymbol{e}_t\\) by feeding \\(t_c\\) into \\(\\varPhi\\); \n7: Obtain perturbed images by \\(\\boldsymbol{x}_{adv} = \\boldsymbol{x}_s + G_w(\\boldsymbol{x}_s, \\boldsymbol{e}_t)\\), \\(\\boldsymbol{x}_{adv}' = \\boldsymbol{x}_s' + G_w(\\boldsymbol{x}_s', \\boldsymbol{e}_t)\\); \n8: Forward pass \\(\\boldsymbol{x}_{adv}\\), \\(\\boldsymbol{x}_{adv}'\\) to \\(f_{\\theta}\\) and compute the loss in Eq. (3); \n9: Perform gradient backpropagation and update the generator \\(G_w\\); \n10: end for \n11: return the trained generator \\(G_w\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.789, + 0.538 + ], + "angle": 0, + "content": "then integrated with the extracted visual concepts \\( \\pmb{h}_s \\) through channel-wise concatenation to obtain the fused representations, i.e., \\( \\pmb{m} \\in \\mathbb{R}^{B \\times (N' + 16) \\times H' \\times W'} \\). Subsequently, \\( \\pmb{m} \\) undergoes further downsampling, and the resultant features are again concatenated with the expanded embedding \\( e_t^{*'} \\). By repeating this operation several times, we effectively fuse the visual concepts of input images and the purified CLIP's embedding of the target classes. This mechanism fully exploits both the instance-level and class-level information from visual and textual modalities, thus encouraging the generation of perturbations with better semantic patterns and higher transferability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.79, + 0.664 + ], + "angle": 0, + "content": "Cross-Attention based Decoder(CA-Decoder). Given the multi-modal fused features from the previous module, the decoder attempts to translate them into perturbations of the target class. The network backbone is realized based on the decoder used in previous works [49,67]. To fully explore the semantic priors brought by the CLIP model, we enhance the underlying backbone by introducing the cross-attention mechanism, which is proven to be effective for many multimodal learning tasks. Specifically, we incorporate the textual embedding \\( \\boldsymbol{e}_t \\) from CLIP's latent space into our decoder via the cross-attention layer:" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.701, + 0.785, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} Q = z _ {t} W _ {q}, K = e _ {t} W _ {k}, V = e _ {t} W _ {v}, \\\\ \\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) \\cdot V, \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.779, + 0.788, + 0.842 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}_t\\in \\mathbb{R}^{B\\times d_\\alpha}\\) denotes the flattened intermediate features of the decoder, \\(W_{q}\\in \\mathbb{R}^{d_{\\alpha}\\times d}\\), \\(W_{k}\\in \\mathbb{R}^{512\\times d}\\), \\(W_{v}\\in \\mathbb{R}^{512\\times d}\\) are learnable parameters. Similar to C-GSP, we post-process the output \\(\\pmb{o}\\) via the \\(\\tanh (\\cdot)\\) smooth projection to obtain the \\(\\ell_{\\infty}\\) constrained perturbation with budge \\(\\epsilon\\), i.e., \\(\\delta = P(o) = \\epsilon \\cdot \\tanh (o)\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "H. Fang et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.157, + 0.52, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.25, + 0.385, + 0.263 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.146, + 0.788, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.649, + 0.251, + 0.671, + 0.263 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.274, + 0.785, + 0.358 + ], + "angle": 0, + "content": "Fig. 3: Illustration of the proposed masked fine-tuning mechanism. (a) We fix the condition input with different text prompts to fine-tune the trained conditional generator \\( G_{w} \\) into multiple generators for single-target attacks. (b) The fooling rate of several target classes with Inc-v3 and Res-152 as substitute models respectively. The results indicate that direct fine-tuning yields inadequate results for certain classes due to overfitting. We efficiently resolve this issue via a patch-wise random mask operation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.784, + 0.414 + ], + "angle": 0, + "content": "Optimization objective. Based on the proposed CLIP-guided architecture, the optimization objective can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.34, + 0.421, + 0.785, + 0.446 + ], + "angle": 0, + "content": "\\[\nw ^ {*} \\leftarrow \\arg \\min _ {w} \\mathcal {L} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {s} + G _ {w} (\\boldsymbol {x} _ {s}, \\Phi (\\boldsymbol {t} _ {c}))\\right), c\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.785, + 0.529 + ], + "angle": 0, + "content": "where \\( t_c \\) represents the corresponding text prompts of the target classes \\( c \\), \\( \\mathcal{L}(\\cdot, \\cdot) \\) denotes the cross-entropy loss. By encouraging the network to craft adversarial samples that can misguide the surrogate model's output toward the desired labels, the generator learns the data distribution of the target class and thus exhibits great generalizability in producing perturbation for any input data." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.547, + 0.53, + 0.563 + ], + "angle": 0, + "content": "3.3 Masked Fine-Tuning Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.785, + 0.674 + ], + "angle": 0, + "content": "In addition to the proposed CLIP-guided generator for multi-target attack scenarios, we also design a single-target variant for further improved performance. As depicted in Fig. 3a, we fix the conditioning input with the text description of a specific target class, and fine-tune the trained multi-target generator by resolving the objective function in Eq. (3) on the unlabeled training data \\(\\mathcal{X}_s\\). This strategy further enhances the attack effects by fine-tuning the conditional generator to specialize in a specific target class." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.787, + 0.841 + ], + "angle": 0, + "content": "However, we encounter the overfitting problem that leads to limited improvement in success rates or even performance degradation for certain target classes. This problem is partly attributed to the fact that the generated adversarial perturbation sometimes heavily focuses on specific regions of the input image [62]. To alleviate this issue, we adopt a patch-wise random mask operation to postprocess the adversarial perturbation produced by the generator, which brings a notable increase in the targeted fooling rate as shown in Fig. 3b. This technique improves the capability and flexibility of our method, as an adversary can perform fine-tuning to further augment the generator for single-target attacks, achieving a better trade-off between efficiency and performance compared to previous single-target methods (See Section 4.5 for detailed analysis)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.345, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "CLIP-Guided Networks for Transferable Targeted Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.377, + 0.163 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.783, + 0.237 + ], + "angle": 0, + "content": "To validate the effectiveness of CGNC in boosting transferability, we conduct extensive experiments on various black-box models in a range of scenarios. Our code is available at https://github.com/ffhibnese/CGNC_Targeted_Adversarial_Attacks. Please see the Appendix for more experimental results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.258, + 0.446, + 0.273 + ], + "angle": 0, + "content": "4.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.282, + 0.783, + 0.356 + ], + "angle": 0, + "content": "Datasets. Following [19,67], we train the generator on the ImageNet training set [9] and evaluate the attack performance using ImageNet-NeurIPS (1k) dataset proposed by [50]. We also conduct experiments in a more realistic cross-domain scenario where we train the generator and generate perturbations on MS-COCO [41] or Comics [4] while evaluating on target classifier trained on ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.783, + 0.416 + ], + "angle": 0, + "content": "Victim Models. We use various victim models with different architectures. Specifically, the naturally trained models include Inception-v3 (Inc-v3) [57], Inception-v4 (Inc-v4) [55], Inception-ResNet-v2 (Inc-Res-v2) [55], ResNet-152 (Res-152) [29], DenseNet-121 (DN-121) [32], GoogleNet [56], and VGG-16 [54]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.783, + 0.505 + ], + "angle": 0, + "content": "For further evaluation, we also analyze on the robust-trained models, including adv-Inception-v3 (Inc-v3adv) [26], ens-adv-Inception-ResNet-v2 (IR-v2ens) [28], and robust-trained ResNet-50 [24,31], dubbed as Res50SIN (trained on stylized ImageNet), Res50IN (trained on the mixture of stylized and Nature ImageNet), Res50fine (Res50IN plus further finetuning with an auxiliary dataset [24]), and Res50Aug (trained with the advanced data augmentation Augmix [31])." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.783, + 0.626 + ], + "angle": 0, + "content": "Baseline Attacks. We reveal the superiority of the proposed CGNC in enhancing multi-target attacks by comparing our method with multiple competitive baselines, including MIM [11], DIM [65], SIM [40], DIM [12], Logit [72], SU [63], and the state-of-the-art (SOTA) multi-target generative attacks CGSP [67] in Section 4.2. For SU attack [63], we choose to compare with its best version DTMI-Logit-SU. Besides, we also provide a comparison of the proposed single-target variant of CGNC with the existing single-target attack methods, i.e., GAP [51], CD-AP [48], TTP [49], and DGTA-PI [19] in Section 4.5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.783, + 0.701 + ], + "angle": 0, + "content": "Implementation Details. Following [19,67], we employ Inc-v3 and Res-152 as surrogate models and adopt the perturbation budget \\(\\epsilon\\) 16/255. We conduct 10 epochs of generator training using a learning rate of 2e-4. During the masked fine-tuning, we maintain the learning rate and apply a mask ratio of 0.2 to fine-tune the text-conditional generator for an additional 5 epochs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.726, + 0.589, + 0.74 + ], + "angle": 0, + "content": "4.2 Multi-Target Transferability Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.783, + 0.807 + ], + "angle": 0, + "content": "To align the experimental setup to former works [19,67], we first target 8 different classes in [71] and provide the average attack success rates (ASR) of the 8 target categories as evaluation metrics. Attack performance under larger numbers of target categories is presented in the next section." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.783, + 0.839 + ], + "angle": 0, + "content": "Attacks against regularly trained models. We first perform attacks on normal models to evaluate the multi-target attack performance. By observing" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "H. Fang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 1: Attack success rates (\\%) for multi-target attacks against regularly trained models on ImageNet NeurIPS validation set. * represents white-box attacks." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.186, + 0.778, + 0.414 + ], + "angle": 0, + "content": "
SourceMethodInc-v3Inc-v4Inc-Res-v2Res-152DN-121GoogleNetVGG-16
Inc-v3MIM99.90*0.801.000.400.200.200.30
TI-MIM98.50*0.500.500.300.200.400.40
SI-MIM99.80*1.502.000.800.700.700.50
DIM95.60*2.700.500.801.100.400.80
TI-DIM96.00*1.101.200.500.500.500.80
SI-DIM90.20*3.804.402.002.201.701.40
Logit99.60*5.606.501.703.000.801.50
SU99.59*5.807.003.353.502.003.94
C-GSP93.40*46.5836.7441.6046.4040.0045.00
CGNC96.03*59.4348.0642.4862.9851.3352.54
Res-152MIM0.500.400.6099.70*0.300.300.20
TI-MIM0.300.300.3096.50*0.300.400.30
SI-MIM1.301.201.6099.50*1.001.400.70
DIM2.302.203.0092.30*0.200.800.70
TI-DIM0.800.701.0090.60*0.600.800.50
SI-DIM4.204.805.4090.50*4.203.602.00
Logit10.1010.7012.8095.70*12.703.709.20
SU12.3611.3116.1695.08*16.136.5514.28
C-GSP37.7033.3320.2893.20*64.2041.7045.90
CGNC53.3951.5334.2495.85*85.6662.2363.36
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.785, + 0.565 + ], + "angle": 0, + "content": "the results in Table 1, we demonstrate that the proposed CGNC consistently achieves significant improvement compared with previous methods. Specifically, our method achieves an average improvement of \\(17.88\\%\\) and \\(10.08\\%\\) in ASR over the C-GSP [67] attack regarding Res-152 and Inc-v3 as surrogate models respectively, demonstrating the superiority of leveraging the rich prior knowledge from CLIP's text embedding. We also note that the iterative methods obtain nearly \\(100\\%\\) while-box ASR while receiving poor performance on black-box models due to overfitting the classification boundaries of surrogate models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.787, + 0.657 + ], + "angle": 0, + "content": "Attacks under defense strategies. For a more thorough analysis and comparison, we then compare these attacks under several widely used defenses. Firstly, we consider attacking six robustly trained networks in Table 2. Generally, our method is still able to deceive the black-box classifiers into predicting the specified classes, significantly outperforming former multi-target attacks, e.g., a \\(20.87\\%\\) increase of fooling rate from Res-152 to \\(\\mathrm{Res}50_{\\mathrm{Aug}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.794 + ], + "angle": 0, + "content": "Next, we evaluate the performance of different approaches on models with input preprocessing-based defenses, including a set of image smoothing mechanisms [10] and JPEG compression [14] algorithms. As shown in Table 3, although these defenses eliminate certain valid information in the adversarial samples, the CLIP-empowered CGNC still demonstrates excellent capability in boosting targeted transferability. Particularly when the substitute model is Res-152, CGNC achieves an average fooling rate of \\(71.18\\%\\) and \\(80.34\\%\\) on smoothing methods and JPEG compression, while C-GSP only reaches \\(49.05\\%\\) and \\(56.16\\%\\) respectively, verifying the stability and robustness of the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Perturbation Visualization. We present visualization results in Fig. 4 to unveil the principle of our method. Upon careful inspection, we can see that the generated perturbations mainly focus on the semantic regions of the input" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.345, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "CLIP-Guided Networks for Transferable Targeted Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 2: Comparison of the proposed CGNC with the SOTA multi-target attacks against models with robust training mechanism on ImageNet NeurIPS validation set." + }, + { + "type": "table", + "bbox": [ + 0.254, + 0.185, + 0.744, + 0.407 + ], + "angle": 0, + "content": "
SourceMethodInc-v3advIR-v2ensRes50SINRes50INRes50fineRes50Aug
Inc-v3MIM0.160.100.200.270.440.19
TI-MIM0.210.190.330.490.680.31
SI-MIM0.130.190.260.430.630.29
DIM0.110.090.160.330.390.19
TI-DIM0.150.130.160.210.330.14
SI-DIM0.190.210.430.710.840.46
Logit0.300.300.701.233.140.86
SU0.490.410.841.753.551.04
C-GSP20.4118.046.9633.7644.5621.95
CGNC24.3622.548.8540.8352.1822.85
Res-152MIM0.190.150.281.582.750.78
TI-MIM0.610.730.502.514.751.76
SI-MIM0.240.240.390.660.840.36
DIM0.630.370.948.5014.223.77
TI-DIM0.230.300.280.761.490.49
SI-DIM0.710.710.752.733.891.37
Logit1.151.181.656.7015.465.93
SU2.121.201.957.5321.146.95
C-GSP14.6016.0116.8460.3065.5142.88
CGNC22.2126.7129.8379.8084.0563.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.422, + 0.784, + 0.464 + ], + "angle": 0, + "content": "Table 3: Comparison of our method with C-GSP under different defense strategies. Q indicates the quality factor in JPEG compression. Here the target model is DN-121 and results for more victim models are in the Appendix." + }, + { + "type": "table", + "bbox": [ + 0.241, + 0.476, + 0.763, + 0.56 + ], + "angle": 0, + "content": "
SourceMethodSmoothingJPEG compression
GaussianMediumAverageQ=70Q=75Q=80Q=85Q=90
Inc-v3C-GSP36.5946.9139.8649.4150.5551.9552.8453.39
CGNC43.0355.3545.9458.3559.2560.2861.2961.94
Res-152C-GSP43.2156.3847.5553.0354.5456.1357.7559.36
CGNC64.4479.8469.2577.3878.8180.5482.0582.94
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.59, + 0.785, + 0.664 + ], + "angle": 0, + "content": "images and contain sufficient semantic patterns specific to the target category. For instance, when the condition is a photo of a sea lion, the resulting perturbations indeed carry rich patterns closely resembling this marine animal. We also observe that the pattern changes in accordance with the text prompts, which validates our conditioning mechanism of the CLIP-encoded textual embedding." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.691, + 0.507, + 0.705 + ], + "angle": 0, + "content": "4.3 Evaluation on More Scenarios" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.72, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Cross-Domain scenarios. Next, we explore the more realistic cross-domain scenarios [48, 49] where attackers do not know anything about the data distribution of the training set used by the black-box classifier. Based on this setting, attackers train the generator and generate adversarial perturbations using an auxiliary dataset that follows a different probability distribution from that of the target model. Specifically, we satisfy the cross-domain experimental setting using MS-COCO [41] and Comics [4] datasets respectively. MS-COCO is a large-scale image dataset widely used for object detection and semantic segmentation," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "H. Fang et al." + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.145, + 0.773, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.334, + 0.788, + 0.378 + ], + "angle": 0, + "content": "Fig. 4: Visualization results of different input images for different targets. For each text prompt of the target class, the left column shows the perturbation generated by our CGNC while the right column displays the corresponding adversarial examples." + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.392, + 0.473, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.542, + 0.496, + 0.6 + ], + "angle": 0, + "content": "Fig.5: Fooling rates (from Res-152 to VGG-16) in attacking 8 target classes on cross-domain scenarios. We also provide the results on ImageNet as a comparison." + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.392, + 0.765, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.542, + 0.788, + 0.6 + ], + "angle": 0, + "content": "Fig.6: Fooling rates of our CGNC with C-GSP on larger numbers of attacked classes regarding Res-152 and Inc-v3 as surrogate models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.787, + 0.658 + ], + "angle": 0, + "content": "while Comic is composed of a large number of comic images that can be regarded as a stylized version of ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.81 + ], + "angle": 0, + "content": "To analyze the cross-domain transferability, we randomly select 1000 images from the source dataset to craft adversarial samples using the trained generator. Fig. 5 shows the results of the cross-domain attacks transferring from three different source datasets to ImageNet. In general, the more challenging cross-domain datasets lead to varying degrees of reduction in the targeted fooling rate due to the domain gap, especially on the Comics dataset which differs greatly from the ImageNet distribution. Benefiting from the text guidance on the target category, CGNC still attains decent attack performance and remarkably outperforms C-GSP. This demonstrates that our method achieves better cross-domain transferability and is partly independent of the training dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In addition, since the SOTA single-target attacks [19,49,62] require samples of the target class for loss computation, they are not applicable in the cross-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.345, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "CLIP-Guided Networks for Transferable Targeted Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 4: Ablation study of CGNC and its two variants on ImageNet NeurIPS validation set for 8 target classes. The substitute model is Res-152." + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.185, + 0.731, + 0.25 + ], + "angle": 0, + "content": "
MethodVGG-16GoogleNetInc-v3Res-152DN-201
CGNA-CA-t56.5551.0947.4492.8174.65
CGNA-CA56.6454.2949.7393.3475.99
CGNC63.3662.2353.3995.8582.69
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.275, + 0.786, + 0.303 + ], + "angle": 0, + "content": "domain scenarios where the source dataset lacks images of the target category. This represents an additional advantage of our method over single-target attacks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.305, + 0.787, + 0.379 + ], + "angle": 0, + "content": "Larger Numbers of Target Classes. We then increase the number of target categories to verify the effectiveness of our method when handling plenty of classes. Specifically, we condition CGNC with more text inputs corresponding to the increased number of target categories. The performance is evaluated across the aforementioned six black-box models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.788, + 0.577 + ], + "angle": 0, + "content": "As mentioned before, single-target attacks become impractical for real-world classification systems with hundreds or thousands of target categories [27]. In comparison, our method effectively solves this issue and achieves great improvements in large numbers of target classes over C-GSP as in Fig. 6, e.g., \\(14.71\\%\\) increase for 200 target classes proxy on the Res-152. This again reveals the significance of the fully exploited textual guidance from the CLIP model. Additionally, our method exhibits greater robustness to the varying numbers of target classes, e.g., the performance of our network on Inc-v3 demonstrates a smoother decrease compared to that of C-GSP. This further highlights the superiority of our method in reducing computation costs. For instance, when launching 1000-class targeted attacks using Inc-v3 as the surrogate model, our method requires training only 5 CGNCs with 200 conditions, yet it can achieve comparable performance to 20 generators with 50 conditions trained using C-GSP." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.596, + 0.389, + 0.611 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.618, + 0.787, + 0.692 + ], + "angle": 0, + "content": "We conduct ablation experiments on the ImageNet NeurIPS dataset to study the effect of the proposed techniques. Specifically, we introduce two variants of CGNC. CGNC-CA removes the cross-attention module from the original network, and CGNC-CA-t further modifies the conditioning mechanism by replacing the CLIP's text embedding with one-hot labels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.693, + 0.787, + 0.769 + ], + "angle": 0, + "content": "From the results in Table 4, we can find that each aforementioned technique can further improve the attack success rates. Moreover, the remarkable improvement from CGNC-CA to CGNC also confirms that these cross-attention modules in the CA-Decoder can make better use of the text guidance provided by the CLIP model to enhance the targeted transferability of crafted perturbations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.588, + 0.803 + ], + "angle": 0, + "content": "4.5 Comparison with Single-Target Attacks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Next, we compare the single-target variants of the proposed CGNC with multiple state-of-the-art single-target attacks. To obtain our single-target generators, we" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "H. Fang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 5: Comparison of our method with the SOTA single-target attacks. \\(^\\dagger\\) denotes the single-target variant of our CGNC obtained through the masked fine-tuning technique. * represents the white-box attacks." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.2, + 0.772, + 0.321 + ], + "angle": 0, + "content": "
SourceMethodInc-v3Inc-v4Inc-Res-v2Res-152DN-121GoogleNetVGG-16
Inc-v3GAP86.90*45.0634.4834.4841.7426.8934.34
CD-AP94.20*57.6060.1037.1041.6032.3041.70
TTP91.37*46.0439.3716.4033.4725.8025.73
DGTA-PI94.63*67.9555.0350.5047.3847.6748.11
CGNC†98.84*74.7664.4862.0078.9469.0670.74
Res-152GAP30.9931.4320.4884.86*58.3529.8939.70
CD-AP33.3043.7042.7096.60*53.8036.6034.10
TTP62.0349.2038.7095.12*82.9665.0962.82
DGTA-PI66.8353.6247.6196.48*86.6168.2969.58
CGNC†68.8669.4545.7198.61*91.1469.8368.05
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.787, + 0.484 + ], + "angle": 0, + "content": "conduct the proposed masked fine-tuning to the trained conditional generator eight times using eight different text prompts of the target classes. Quantitative results in Table 5 indicate that our single-target enhanced generators outperform the competing single-target attacks in most cases. Particularly on the surrogate model of Inc-v3, \\(\\mathrm{CGNC}^{\\dagger}\\) achieves a notable increase of \\(15.36\\%\\) in the average black-box fooling rate compared to previous methods, demonstrating the significant effectiveness of the proposed masked fine-tuning. The MFT technique further enhances the single-target performance, thereby improving the adaptability and flexibility of our approach in scenarios with fewer target classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.788, + 0.606 + ], + "angle": 0, + "content": "Note that these single-target methods require training 8 generators from scratch for 8 different target classes. In contrast, our method only needs to train a single multi-target generator and perform fine-tuning 8 times, each time with just a few epochs. Based on the typical experimental setup used in the SOTA single-target methods [19,49], our strategy can diminish over 100 training epochs when targeting 8 classes, thus substantially mitigating the computational burden. More experiments of single-target attacks against adversarially robust models and input preprocessing defenses are shown in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.628, + 0.36, + 0.644 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we design a novel generative network CGNC, which improves multi-target transferable adversarial attacks by fully utilizing the rich prior within the CLIP as auxiliary semantic knowledge about the target category. To better incorporate the prior information, we introduce the cross-attention modules and efficiently condition the generator with CLIP's text embeddings. Through extensive experiments, we demonstrate the effectiveness of the proposed CGNC on multiple black-box target models in a variety of scenarios. Moreover, we propose a masked fine-tuning technique to further enhance the CGNC in attacking a single class, which outperforms previous single-target methods in both efficiency and effectiveness. We hope that the proposed method can serve as a reliable tool for evaluating the model robustness under black-box setups, promoting further research on the vulnerability and robustness of DNNs." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.345, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "CLIP-Guided Networks for Transferable Targeted Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.174, + 0.794, + 0.249 + ], + "angle": 0, + "content": "This work is supported in part by the National Natural Science Foundation of China under grant 62171248, 62301189, Guangdong Basic and Applied Basic Research Foundation under grant 2021A1515110066, the PCNL KEY project (PCL2021A07), and Shenzhen Science and Technology Program under Grant JCYJ20220818101012025, RCBS20221008093124061, GXWD20220811172936001." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.269, + 0.323, + 0.286 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.297, + 0.785, + 0.325 + ], + "angle": 0, + "content": "1. Akhtar, N., Mian, A.: Threat of adversarial attacks on deep learning in computer vision: A survey. IEEE Access 6, 14410-14430 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.326, + 0.785, + 0.365 + ], + "angle": 0, + "content": "2. Andriushchenko, M., Croce, F., Flammarion, N., Hein, M.: Square attack: a query-efficient black-box adversarial attack via random search. In: European conference on computer vision. pp. 484-501. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.366, + 0.785, + 0.406 + ], + "angle": 0, + "content": "3. Athalye, A., Engstrom, L., Ilyas, A., Kwok, K.: Synthesizing robust adversarial examples. In: International conference on machine learning. pp. 284-293. PMLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.407, + 0.785, + 0.433 + ], + "angle": 0, + "content": "4. BircanoAYlu, C.: https://www.kaggle.com/datasets/cenkbircanoglu/comic-books-classification. Kaggle, 2017" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.46 + ], + "angle": 0, + "content": "5. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: 2017 ieee symposium on security and privacy (sp). pp. 39-57. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.785, + 0.501 + ], + "angle": 0, + "content": "6. Che, Z., Borji, A., Zhai, G., Ling, S., Li, J., Min, X., Guo, G., Le Callet, P.: Smgea: A new ensemble adversarial attack powered by long-term gradient memories. IEEE Transactions on Neural Networks and Learning Systems 33(3), 1051-1065 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.502, + 0.785, + 0.528 + ], + "angle": 0, + "content": "7. Chefer, H., Benaim, S., Paiss, R., Wolf, L.: Image-based clip-guided essence transfer. In: European Conference on Computer Vision. pp. 695-711. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.529, + 0.785, + 0.582 + ], + "angle": 0, + "content": "8. Chen, B., Feng, Y., Dai, T., Bai, J., Jiang, Y., Xia, S.T., Wang, X.: Adversarial examples generation for deep product quantization networks on image retrieval. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(2), 1388-1404 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.583, + 0.785, + 0.624 + ], + "angle": 0, + "content": "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.624, + 0.785, + 0.651 + ], + "angle": 0, + "content": "0. Ding, G.W., Wang, L., Jin, X.: Advertorch v0.1: An adversarial robustness toolbox based on pytorch. arXiv preprint arXiv:1902.07623 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.651, + 0.785, + 0.691 + ], + "angle": 0, + "content": "1. Dong, Y., Liao, F., Pang, T., Su, H., Zhu, J., Hu, X., Li, J.: Boosting adversarial attacks with momentum. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9185-9193 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.785, + 0.732 + ], + "angle": 0, + "content": "2. Dong, Y., Pang, T., Su, H., Zhu, J.: Evading defenses to transferable adversarial examples by translation-invariant attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4312-4321 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.733, + 0.785, + 0.759 + ], + "angle": 0, + "content": "3. Du, J., Zhang, H., Zhou, J.T., Yang, Y., Feng, J.: Query-efficient meta attack to deep neural networks. arXiv preprint arXiv:1906.02398 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.785 + ], + "angle": 0, + "content": "4. Dziugaite, G.K., Ghahramani, Z., Roy, D.M.: A study of the effect of jpg compression on adversarial images. arXiv preprint arXiv:1608.00853 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Eykholt, K., Evtimov, I., Fernandes, E., Li, B., Rahmati, A., Xiao, C., Prakash, A., Kohno, T., Song, D.: Robust physical-world attacks on deep learning visual classification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1625-1634 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.297, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "H. Fang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "16. Fang, H., Chen, B., Wang, X., Wang, Z., Xia, S.T.: Gidf: A generative gradient inversion method with feature domain optimization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4967-4976 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.23 + ], + "angle": 0, + "content": "17. Fang, H., Kong, J., Yu, W., Chen, B., Li, J., Xia, S., Xu, K.: One perturbation is enough: On generating universal adversarial perturbations against vision-language pre-training models. arXiv preprint arXiv:2406.05491 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.787, + 0.271 + ], + "angle": 0, + "content": "18. Fang, H., Qiu, Y., Yu, H., Yu, W., Kong, J., Chong, B., Chen, B., Wang, X., Xia, S.T.: Privacy leakage on dnns: A survey of model inversion attacks and defenses. arXiv preprint arXiv:2402.04013 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.787, + 0.312 + ], + "angle": 0, + "content": "19. Feng, W., Xu, N., Zhang, T., Zhang, Y.: Dynamic generative targeted attacks with pattern injection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16404-16414 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.312, + 0.787, + 0.339 + ], + "angle": 0, + "content": "20. Gao, K., Bai, Y., Bai, J., Yang, Y., Xia, S.T.: Adversarial robustness for visual grounding of multimodal large language models. In: ICLR Workshop (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.787, + 0.366 + ], + "angle": 0, + "content": "21. Gao, K., Bai, Y., Gu, J., Xia, S.T., Torr, P., Li, Z., Liu, W.: Inducing high energy-latency of large vision-language models with verbose images. In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.787, + 0.406 + ], + "angle": 0, + "content": "22. Gao, K., Gu, J., Bai, Y., Xia, S.T., Torr, P., Liu, W., Li, Z.: Energy-latency manipulation of multi-modal large language models via verbose samples. arXiv preprint arXiv:2404.16557 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.787, + 0.433 + ], + "angle": 0, + "content": "23. Gao, L., Cheng, Y., Zhang, Q., Xu, X., Song, J.: Feature space targeted attacks by statistic alignment. arXiv preprint arXiv:2105.11645 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.787, + 0.473 + ], + "angle": 0, + "content": "24. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.514 + ], + "angle": 0, + "content": "25. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.787, + 0.542 + ], + "angle": 0, + "content": "26. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.787, + 0.595 + ], + "angle": 0, + "content": "27. Han, J., Dong, X., Zhang, R., Chen, D., Zhang, W., Yu, N., Luo, P., Wang, X.: Once a man: Towards multi-target attack via learning multi-target adversarial network once. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5158-5167 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.595, + 0.787, + 0.623 + ], + "angle": 0, + "content": "28. Hang, J., Han, K., Chen, H., Li, Y.: Ensemble adversarial black-box attacks against deep learning systems. Pattern Recognition 101, 107184 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.787, + 0.678 + ], + "angle": 0, + "content": "29. He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14. pp. 630-645. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.787, + 0.704 + ], + "angle": 0, + "content": "30. He, Z., Wang, W., Dong, J., Tan, T.: Revisiting ensemble adversarial attack. Signal Processing: Image Communication 107, 116747 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.787, + 0.745 + ], + "angle": 0, + "content": "31. Hendrycks, D., Mu, N., Cubuk, E.D., Zoph, B., Gilmer, J., Lakshminarayanan, B.: Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.787, + 0.786 + ], + "angle": 0, + "content": "32. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "33. Huang, Z., Zhou, A., Ling, Z., Cai, M., Wang, H., Lee, Y.J.: A sentence speaks a thousand images: Domain generalization through distilling clip with language guidance. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11685-11695 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.345, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "CLIP-Guided Networks for Transferable Targeted Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "34. Inkawich, N., Liang, K., Wang, B., Inkawich, M., Carin, L., Chen, Y.: Perturbing across the feature hierarchy to improve standard and strict blackbox attack transferability. Advances in Neural Information Processing Systems 33, 20791-20801 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "35. Inkawich, N., Wen, W., Li, H.H., Chen, Y.: Feature space perturbations yield more transferable adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7066-7074 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.286 + ], + "angle": 0, + "content": "36. Ju, C., Han, T., Zheng, K., Zhang, Y., Xie, W.: Prompting visual-language models for efficient video understanding. In: European Conference on Computer Vision. pp. 105-124. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.328 + ], + "angle": 0, + "content": "37. Kong, Z., Guo, J., Li, A., Liu, C.: Physgan: Generating physical-world-resilient adversarial examples for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14254-14263 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.787, + 0.356 + ], + "angle": 0, + "content": "38. Kurakin, A., Goodfellow, I., Bengio, S.: Adversarial machine learning at scale. arXiv preprint arXiv:1611.01236 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.398 + ], + "angle": 0, + "content": "39. Li, Q., Guo, Y., Chen, H.: Yet another intermediate-level attack. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16. pp. 241-257. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.424 + ], + "angle": 0, + "content": "40. Lin, J., Song, C., He, K., Wang, L., Hopcroft, J.E.: Nesterov accelerated gradient and scale invariance for adversarial attacks. arXiv preprint arXiv:1908.06281 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.48 + ], + "angle": 0, + "content": "41. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.536 + ], + "angle": 0, + "content": "42. Liu, A., Wang, J., Liu, X., Cao, B., Zhang, C., Yu, H.: Bias-based universal adversarial patch attack for automatic check-out. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16. pp. 395-410. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.787, + 0.591 + ], + "angle": 0, + "content": "43. Lu, Y., Jia, Y., Wang, J., Li, B., Chai, W., Carin, L., Velipasalar, S.: Enhancing cross-task black-box transferability of adversarial examples with dispersion reduction. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 940-949 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.787, + 0.632 + ], + "angle": 0, + "content": "44. Luo, J., Bai, T., Zhao, J.: Generating adversarial yet inconspicuous patches with a single image (student abstract). In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 15837-15838 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.674 + ], + "angle": 0, + "content": "45. Moosavi-Dezfooli, S.M., Fawzi, A., Fawzi, O., Frossard, P.: Universal adversarial perturbations. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1765-1773 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.715 + ], + "angle": 0, + "content": "46. Moosavi-Dezfooli, S.M., Fawzi, A., Frossard, P.: Deepfool: a simple and accurate method to fool deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2574-2582 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.757 + ], + "angle": 0, + "content": "47. Mopuri, K.R., Uppala, P.K., Babu, R.V.: Ask, acquire, and attack: Data-free uap generation using class impressions. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 19-34 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "48. Naseer, M.M., Khan, S.H., Khan, M.H., Shahbaz Khan, F., Porikli, F.: Cross-domain transferability of adversarial perturbations. Advances in Neural Information Processing Systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "49. Naseer, M., Khan, S., Hayat, M., Khan, F.S., Porikli, F.: On generating transferable targeted perturbations. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7708-7717 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "H. Fang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "50. NeurIPS: https://www.kaggle.com/c/nips-2017-defense-against-adversarial-attack/data.Kaggle,2017" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.217 + ], + "angle": 0, + "content": "51. Poursaeed, O., Katsman, I., Gao, B., Belongie, S.: Generative adversarial perturbations. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4422-4431 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.272 + ], + "angle": 0, + "content": "52. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "53. Sarkar, S.K., Oshiba, K., Giebisch, D., Singer, Y.: Robust classification of financial risk. arXiv preprint arXiv:1811.11079 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.328 + ], + "angle": 0, + "content": "54. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "55. Szegedy, C., Ioffe, S., Vanhoucke, V., Alemi, A.: Inception-v4, inception-resnet and the impact of residual connections on learning. In: Proceedings of the AAAI conference on artificial intelligence. vol. 31 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.412 + ], + "angle": 0, + "content": "56. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A.: Going deeper with convolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1-9 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.453 + ], + "angle": 0, + "content": "57. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2818-2826 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.494 + ], + "angle": 0, + "content": "58. Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I., Fergus, R.: Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.536 + ], + "angle": 0, + "content": "59. Tao, M., Bao, B.K., Tang, H., Xu, C.: Galip: Generative adversarial clips for text-to-image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14214-14223 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "60. Wang, X., He, K.: Enhancing the transferability of adversarial attacks through variance tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1924-1933 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.618 + ], + "angle": 0, + "content": "61. Wang, X., He, K., Hopcroft, J.E.: At-gan: A generative attack model for adversarial transferring on generative adversarial nets. arXiv preprint arXiv:1904.07793 3(4), 3 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.66 + ], + "angle": 0, + "content": "62. Wang, Z., Yang, H., Feng, Y., Sun, P., Guo, H., Zhang, Z., Ren, K.: Towards transferable targeted adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20534-20543 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.702 + ], + "angle": 0, + "content": "63. Wei, Z., Chen, J., Wu, Z., Jiang, Y.G.: Enhancing the self-universality for transferable targeted attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12281-12290 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.73 + ], + "angle": 0, + "content": "64. Xiao, C., Li, B., Zhu, J.Y., He, W., Liu, M., Song, D.: Generating adversarial examples with adversarial networks. arXiv preprint arXiv:1801.02610 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.785 + ], + "angle": 0, + "content": "65. Xie, C., Zhang, Z., Zhou, Y., Bai, S., Wang, J., Ren, Z., Yuille, A.L.: Improving transferability of adversarial examples with input diversity. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2730-2739 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "66. Xiong, Y., Lin, J., Zhang, M., Hopcroft, J.E., He, K.: Stochastic variance reduced ensemble adversarial attack for boosting the adversarial transferability. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14983-14992 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.345, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "CLIP-Guided Networks for Transferable Targeted Attacks" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "67. Yang, X., Dong, Y., Pang, T., Su, H., Zhu, J.: Boosting transferability of targeted adversarial examples via hierarchical generative networks. In: European Conference on Computer Vision. pp. 725-742. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "68. Yu, W., Fang, H., Chen, B., Sui, X., Chen, C., Wu, H., Xia, S.T., Xu, K.: Gi-nas: Boosting gradient inversion attacks through adaptive neural architecture search. arXiv preprint arXiv:2405.20725 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.286 + ], + "angle": 0, + "content": "69. Zha, Y., Ji, H., Li, J., Li, R., Dai, T., Chen, B., Wang, Z., Xia, S.T.: Towards compact 3d representations via point feature enhancement masked autoencoders. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 6962-6970 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.328 + ], + "angle": 0, + "content": "70. Zha, Y., Wang, J., Dai, T., Chen, B., Wang, Z., Xia, S.T.: Instance-aware dynamic prompt tuning for pre-trained point cloud models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14161-14170 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.383 + ], + "angle": 0, + "content": "71. Zhang, C., Benz, P., Imtiaz, T., Kweon, I.S.: Understanding adversarial examples from the mutual influence of images and perturbations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14521-14530 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.425 + ], + "angle": 0, + "content": "72. Zhao, Z., Liu, Z., Larson, M.: On success and simplicity: A second look at transferable targeted attacks. Advances in Neural Information Processing Systems 34, 6115-6128 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.425 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CLIP-Guided Generative Networks for Transferable Targeted Adversarial Attacks/ddf47f95-f1c1-405a-8323-f8a66b1071ea_origin.pdf b/2024/CLIP-Guided Generative Networks for Transferable Targeted Adversarial Attacks/ddf47f95-f1c1-405a-8323-f8a66b1071ea_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5f29f76fcb5f4fe74d8424f9c0b6181cb67116ef --- /dev/null +++ b/2024/CLIP-Guided Generative Networks for Transferable Targeted Adversarial Attacks/ddf47f95-f1c1-405a-8323-f8a66b1071ea_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d43e3b42146d618ef59046974ccc6987bed4dbacc377a665868331e8bc05cc62 +size 1714645 diff --git a/2024/CLIP-Guided Generative Networks for Transferable Targeted Adversarial Attacks/full.md b/2024/CLIP-Guided Generative Networks for Transferable Targeted Adversarial Attacks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9c7b29adedd5f4551d256587239f19e515a38b76 --- /dev/null +++ b/2024/CLIP-Guided Generative Networks for Transferable Targeted Adversarial Attacks/full.md @@ -0,0 +1,313 @@ +# CLIP-Guided Generative Networks for Transferable Targeted Adversarial Attacks + +Hao Fang $^{1\dagger}$ , Jiawei Kong $^{2\dagger}$ , Bin Chen $^{2,3,4\#}$ , Tao Dai $^{5}$ , Hao Wu $^{6}$ , and Shu-Tao Xia $^{1,3}$ + +$^{1}$ Tsinghua Shenzhen International Graduate School, Tsinghua University $^{2}$ Harbin Institute of Technology, Shenzhen $^{3}$ Pengcheng Laboratory + +4 Guangdong Provincial Key Laboratory of Novel Security Intelligence Technologies +5 Shenzhen University 6 Shenzhen Digital Certificate Authority CO., Ltd +fang-h23@mails.tsinghua.edu.cn, kongjiawei@stu.hit.edu.cn +chenbin2021@hit.edu.cn, daitao.edu@gmail.com, whpc79@163.com, +xiast@sz.tsinghua.edu.cn + +Abstract. Transferable targeted adversarial attacks aim to mislead models into outputting adversary-specified predictions in black-box scenarios. Recent studies have introduced single-target attacks that train a generator for each target class to generate highly transferable perturbations, resulting in substantial computational overhead when handling multiple classes. Multi-target attacks address this by training only one class-conditional generator for multiple classes. However, the generator simply uses class labels as conditions, failing to leverage the rich semantic information of the target class. To this end, we design a CLIP-guided Generative Network with Cross-attention modules (CGNC) to enhance multi-target attacks by incorporating textual knowledge of CLIP into the generator. Extensive experiments demonstrate that CGNC yields significant improvements over previous multi-target attacks, e.g., a $21.46\%$ improvement in success rate from Res-152 to DenseNet-121. Moreover, we propose the masked fine-tuning to further strengthen our method in attacking a single class, which surpasses existing single-target methods. + +# 1 Introduction + +With the rapid progress of deep learning, deep neural networks (DNNs) have been widely applied in many security-critical fields, such as autonomous driving [15, 37], financial systems [53], and point cloud modeling [69, 70]. However, DNNs are corroborated to be vulnerable to adversarial attacks [20, 26, 58], which attempt to fool models with adversarial examples crafted by adding imperceptible perturbations to the original inputs. Based on the attack goal, adversarial attacks can be categorized into untargeted and targeted attacks. Untargeted attacks attempt to fool DNNs into predicting incorrect labels while targeted attacks aim at triggering the model to output the attacker-desired predictions. + +![](images/e597ef4133afdf3eb4daa87c93fd652ba2630da50cbeec01b0458fd1c5157721.jpg) +(a) + +![](images/e07c63107175448b06783303ae2d73a0281fb1165b7aef9c1dfcac5908d528e7.jpg) +(b) +Fig. 1: (a) Targeted attacks from 'Panda' to 'Dog'. Fig. (a) illustrates that previous multi-target methods [27,67] generate perturbations simply conditioned by class indices or one-hot vectors and only learn the classification boundary specific to the surrogate model. In contrast, CGNC exploits CLIP's meaningful guidance to effectively capture the feature distribution inherent to the target data. (b) We directly feed the scaled perturbations generated by our CGNC and C-GSP [67] into three black-box classifiers and reveal that our perturbations achieve significantly higher mean confidence of the target class, demonstrating its superiority in modeling the target feature distribution. + +Recent investigations into the adversarial transferability [2, 3, 44, 58] have demonstrated that adversarial examples crafted for a white-box surrogate model can also mislead other unseen black-box models. Since this attack does not require access to the target model, it exposes a serious security threat to real-world applications of DNNs and motivates a wide range of studies [1,5,42,46]. Despite the remarkable performance on untargeted transferable attacks, these approaches produce unsatisfactory results for targeted attacks due to their over-reliance on the white-box surrogate. Existing studies on transferable targeted attacks can be categorized into instance-specific [11, 15, 23, 39, 43, 60, 66] and instance-agnostic attacks [19, 37, 44, 48, 49, 64]. Specifically, instance-specific attacks [11, 12, 65] iteratively perform gradient updating to craft adversarial perturbation tailored to a specific natural sample, yet producing low-targeted transferability due to overfitting [67]. Conversely, instance-agnostic attacks learn a universal perturbation [45, 71] or a perturbation generator [48, 51] based on data distribution rather than the specific instance, alleviating the data-specific over-fitting issues and achieving a higher adversarial transferability. Recent studies further explore generative attacks to produce highly transferable perturbations, which can be divided into single-target and multi-target attacks as follows. + +Single-target attacks [19, 48, 49, 62] exhibit impressive performance by training perturbation generators for target categories. However, they require training a generator for each target, which can lead to a heavy computation burden when attacking numerous classes, and thus are not applicable in real-world classification systems that usually contain hundreds/thousands of target classes [27]. To address this, [27, 67] propose multi-target attacks that train a single conditional generator for multi-target classes. By specifying the desired label as conditioning input, the trained generator can efficiently generate the corresponding adversar + +ial perturbation. Nevertheless, these methods simply adopt class indices [27] or one-hot vectors [67] of the target labels as conditions, and thus solely rely on the classification information from the surrogate model as the guidance of the target category, hence resulting in only modest black-box fooling rates. + +In this paper, we build upon the research line of multi-target attacks and propose a novel CLIP-guided Generative Network with Cross-attention modules (CGNC) to boost the attack effect. Concretely, we revisit the architecture of the conditional generators used in [27,67] and argue that the simple class conditions, e.g., class indices or handcrafted one-hot vectors, limit the transferability of the generated perturbations (see Fig. 1(a) for comparison). Motivated by the impressive effects of CLIP's encoded text embeddings in multi-modal learning, we introduce concise text descriptions of the target classes to encode them as class-specific representations encapsulated with abundant information, which assist the generator in learning target class distribution and ultimately lead to a fundamental transferability improvement. To better incorporate the text information, we improve the condition-inputting mechanism by adding the cross-attention layers that have been proven effective in models with various input modalities. Results in Fig. 1(b) confirm the capability of our CGNC in capturing the target distribution. Besides, we propose a masked fine-tuning (MFT) technique, which fine-tunes the trained conditional generator with a fixed text condition of the desired class for further improvement in attacking a single class. + +With all the above efforts, our method achieves great efficiency and scalability. When there are hundreds of target classes and single-target methods become impractical, the proposed conditional network can be utilized to achieve significant performance over previous multi-target attacks. Conversely, for scenarios involving only a few target classes, the proposed MFT mechanism enhances CGNC to outperform existing single-target methods while substantially reducing computational costs. In summary, our main contributions are as follows: + +- We propose CGNC, a novel CLIP-guided generative network with cross-attention layers that fully exploits the textual knowledge provided by the advanced CLIP model for enhanced multi-target attacks. +- We introduce a masked fine-tuning mechanism to improve the single-target performance by adapting the CGNC to an individual target class. +- Extensive experiments show that CGNC remarkably enhances the targeted transferability compared to previous attack methods in a range of settings. + +# 2 Related Work + +# 2.1 Vision-Language Models + +Vision-language models (VLM) drawn great attention [21,22] due to their promising potential in learning general visual and textual representations through contrastive pre-training on large-scale image-text pairs. Given the powerful capacity of text descriptions in modeling multi-modal tasks, these models can be effectively adapted to diverse downstream tasks through appropriately formulated + +textual prompts, including video understanding [36], image manipulation [7], and text-to-image synthesis [59]. Inspired by these works, we propose to harness CLIP-based multi-modal learning to empower our perturbation generator, facilitating more effective learning of the target feature distribution. + +# 2.2 Adversarial Attacks + +Among various security threats to DNNs [8, 16, 18, 68], the adversarial attack is one of the most formidable and well-known one, which can be classified into instance-specific attacks and instance-agnostic attacks [67]. + +Instance-specific Attacks. Since the pioneering work [58] highlighted the vulnerability of neural networks, numerous gradient-based optimization methods [3, 13, 26, 38, 66] have been proposed to craft image-dependent perturbations. MIM [11] integrates a momentum item into the gradient update for utilizing the previous gradient information to avoid plunging into poor local optimum. DIM [65] enhances the transferability by randomly diversifying the input pattern, and TIM [12] implements the attack by convolving the gradient with a predefined kernel. In addition, intermediate feature space [35, 63] and classifier information [34] are exploited to enhance attack effects, while [72] leverages logit-based loss to achieve competitive results. More advanced works [6,28,30] consider an ensemble of multiple surrogate models to reduce over-fitting. + +Instance-agnostic Attacks. In contrast to instance-specific attacks, instance-agnostic attacks learn a universal perturbation [17, 45] or a generative model [44, 48, 49, 51, 61, 67] for crafting adversarial examples. By modeling the global data distribution, these methods have shown better transferability and attracted more attention in recent years. Specifically, UAP [45] and AAA [47] learn a universal perturbation to fool the model based on concrete data and compressed impression, respectively. Many subsequent works, such as GAP [51], focus on boosting the transferability of adversarial attacks by training generative models. These generative attacks can be categorized into the following two types. + +Single-target Attacks. This type of attack requires training a generator for each target class. [64] first introduce the generative adversarial networks (GAN) [25] to generate adversarial perturbations. Then, CD-AP [48] concentrates on the domain-invariant adversaries and launches highly transferable cross-domain attacks using a relativistic supervisory. TTP [49] modifies the loss function and proposes to match the target distribution to mitigate over-fitting to the surrogate model. Subsequent methods achieve better transferability based on a dynamic network with pattern injection [19] or a feature discriminator [62]. + +Multi-target Attacks. MAN [27] notes that when dealing with numerous classes, single-target attacks inevitably suffer from extreme computational burdens as they need to train the same number of models as multiple target classes. Therefore, single-target attacks become impractical in attacking real classification systems that often have hundreds of categories. To pursue the extreme speed and storage, MAN trains only one model for 1000 target categories from ImageNet [9]. However, the excess of the target class severely degrades the targeted + +transferability. C-GSP [67] improves the performance by designing a hierarchical partition mechanism to divide all classes into a feasible number of subsets and train generators for each subset. Nonetheless, C-GSP simply conditions the generator with one-hot vectors and only utilizes the classification information to train the generator, failing to use the semantic knowledge of the target class. To overcome the limitation, we propose to incorporate the text information provided by CLIP as significant guidance for the target class. + +# 3 Method + +In this section, we first introduce the preliminaries of targeted transferable attacks and present the basic paradigm of generative attack methods. Then, we elaborate on the proposed CGNC, which remarkably enhances multi-target black-box attacks. Finally, we detail the proposed masked fine-tuning technique that strengthens our method in single-class attacks. + +# 3.1 Preliminary + +We denote the white-box image classifier parameterized with $\theta$ as $f_{\theta}:\mathcal{X}\rightarrow \mathcal{Y}$ where $\mathcal{X}\subset \mathbb{R}^{N\times H\times W}$ represents the image domain and $\mathcal{V}\subset \mathbb{R}^L$ is the output confidence score of different classes $(H,W,N,L$ being height, width, number of channels, and class number). Given a natural image $\pmb {x}\in \mathcal{X}$ and the attacker's desired label $c_{t}\in \mathcal{C}$ , the transferable targeted adversarial attacks attempt to craft the imperceptible perturbation $\delta$ based on the accessible surrogate model $f_{\theta}$ to mislead another unseen victim model $F_{\phi}$ into predicting $c_{t}$ , i.e., $\arg \max_{i\in \mathcal{C}}F_{\phi}(\pmb {x} + \pmb {\delta})_i = c_t$ . Concurrently, the $l_{\infty}$ norm is employed to ensure the adversarial samples are indistinguishable from the original images by constraining the perturbation within the range $\epsilon$ by $\| \pmb {\delta}\|_{\infty}\leq \epsilon$ . + +To boost the fooling rate of targeted black-box attacks, single-target attacks [48, 49, 51] utilize powerful generative models to learn the target data distribution and achieve higher transferability. However, these methods consume great computation time and resources for multi-target scenarios, making them impractical for real-world attacks. C-GSP [67] effectively solves this issue by formulating the multi-target attacks as learning a class-conditional generator $G_{w}$ with weights $w$ , which is capable of generating perturbations for any target class. Given an unlabeled training dataset $\mathcal{X}_s$ , the optimization objective is as follows: + +$$ +\min _ {w} \mathbb {E} _ {\boldsymbol {x} _ {s} \sim \chi_ {s}, c \sim \mathcal {C}} [ \mathcal {L} (f _ {\theta} (\boldsymbol {x} _ {s} + G _ {w} (\boldsymbol {x} _ {s}, c)), c) ], \text {s . t .} \| G _ {w} (\boldsymbol {x} _ {s}, c) \| _ {\infty} \leq \epsilon , \tag {1} +$$ + +where $\mathcal{L}(\cdot, \cdot)$ denotes the cross-entropy (CE) loss. By minimizing the loss of specified classes using various unlabeled images from $\mathcal{X}_s$ , we optimize the parameters $w$ of the generative model and finally obtain a targeted conditional generator that can generate perturbation for any given clean image $\pmb{x}_t$ from the test dataset $\mathcal{X}_t$ . Specifically, the adversary can simply specify a target label $c$ and craft an corresponding adversarial example via $\pmb{x}_{adv} = \pmb{x}_t + G_w(\pmb{x}_t, c)$ . + +![](images/6ad7695640e03bc51b244d528ec508da07dde1217990c7f38aaa3b5759aa9e41.jpg) +Fig. 2: An overview of our proposed architecture of CGNC. Equipped with the three exquisite modules VL-Purifier, F-Encoder, and CA-Decoder, the generator fully leverages the textual representations encoded by CLIP as auxiliary information about the target classes to better probe their data distribution and achieve superior attack effects. + +However, current multi-target methods [27, 67] simply condition the generator with class labels and learn the target distribution only relying on the classification information of the surrogate model, thus not fully exploiting the specific characteristics of the target category. Inspired by the efficiency of vision-language (VL) learning [52, 59], we propose a novel generative network that leverages sufficient prior knowledge from the powerful CLIP model by incorporating textual-modality information to promote the target class modeling, which greatly enhances the multi-target transferable attacks. + +# 3.2 CLIP-Guided Generative Network + +The proposed generative model architecture is presented in Fig. 2. Specifically, CGNC is composed of a Vision-Language feature Purifier (VL-Purifier), a feature Fusion Encoder (F-Encoder), and a Cross-Attention based Decoder (CA-Decoder). We also provide the pseudocode of the training procedure in Algorithm 1. Next, we illustrate the design of each module as follows. + +Vision-language feature purifier (VL-Purifier). To utilize CLIP to produce semantic embeddings, we first feed in CLIP's text encoder $\varPhi$ with queries $t_c$ that follow the handcrafted template "a photo of a {class}" which has shown effectiveness in many tasks [33,52]. Since the obtained embeddings $\pmb{e}_t\in \mathbb{R}^{B\times 512}$ ( $B$ being the batch size) in CLIP's vision-language space are generic representations of the target classes and are not yet tailored to our learning task, we refine them using the VL-Purifier, which is composed of several blocks consisting of a fully-connected layer and a spectral normalization layer. Through this module, we translate the encoded embeddings into more meaningful representations $\pmb{e}_t^*\in \mathbb{R}^{B\times 16}$ , thereby facilitating the subsequent step of feature fusion. + +Feature fusion encoder (F-Encoder). This module aims to fuse the purified features $\pmb{e}_t^*$ with the learned visual representations. Firstly, a batch of input images $\pmb{x}_s$ is encoded to capture the visual concepts $\pmb{h}_s \in \mathbb{R}^{B \times N' \times H' \times W'}$ . Then, we expand the text embedding $\pmb{e}_t^* \in \mathbb{R}^{B \times 16}$ into $\pmb{e}_t^{*'} \in \mathbb{R}^{B \times 16 \times H' \times W'}$ , which are + +Algorithm 1 Pseudocode of Training the CLIP-guided Generative Network +Require: $\mathcal{X}_s$ : the training data; $\mathcal{C}$ : the target label space; $\mathcal{T}$ : the text prompts set; $f_{\theta}$ : the surrogate model; $\varPhi$ : the CLIP's text encoder; $N$ : the max iteration; +Ensure: the perturbation generator $G_w$ ; +1: for $i \gets 0$ to $N$ do +2: Sample a batch of images $\boldsymbol{x}_s \sim \mathcal{X}_s$ ; +3: Obtain $\boldsymbol{x}_s'$ by processing $\boldsymbol{x}_s$ with data augmentation; +4: Sample a batch of target labels $c \sim \mathcal{C}$ ; +5: Obtain the corresponding text prompts $t_c$ from $\mathcal{T}$ ; +6: Compute the text embedding $\boldsymbol{e}_t$ by feeding $t_c$ into $\varPhi$ ; +7: Obtain perturbed images by $\boldsymbol{x}_{adv} = \boldsymbol{x}_s + G_w(\boldsymbol{x}_s, \boldsymbol{e}_t)$ , $\boldsymbol{x}_{adv}' = \boldsymbol{x}_s' + G_w(\boldsymbol{x}_s', \boldsymbol{e}_t)$ ; +8: Forward pass $\boldsymbol{x}_{adv}$ , $\boldsymbol{x}_{adv}'$ to $f_{\theta}$ and compute the loss in Eq. (3); +9: Perform gradient backpropagation and update the generator $G_w$ ; +10: end for +11: return the trained generator $G_w$ + +then integrated with the extracted visual concepts $\pmb{h}_s$ through channel-wise concatenation to obtain the fused representations, i.e., $\pmb{m} \in \mathbb{R}^{B \times (N' + 16) \times H' \times W'}$ . Subsequently, $\pmb{m}$ undergoes further downsampling, and the resultant features are again concatenated with the expanded embedding $e_t^{*'}$ . By repeating this operation several times, we effectively fuse the visual concepts of input images and the purified CLIP's embedding of the target classes. This mechanism fully exploits both the instance-level and class-level information from visual and textual modalities, thus encouraging the generation of perturbations with better semantic patterns and higher transferability. + +Cross-Attention based Decoder(CA-Decoder). Given the multi-modal fused features from the previous module, the decoder attempts to translate them into perturbations of the target class. The network backbone is realized based on the decoder used in previous works [49,67]. To fully explore the semantic priors brought by the CLIP model, we enhance the underlying backbone by introducing the cross-attention mechanism, which is proven to be effective for many multimodal learning tasks. Specifically, we incorporate the textual embedding $\boldsymbol{e}_t$ from CLIP's latent space into our decoder via the cross-attention layer: + +$$ +\begin{array}{l} Q = z _ {t} W _ {q}, K = e _ {t} W _ {k}, V = e _ {t} W _ {v}, \\ \operatorname {A t t e n t i o n} (Q, K, V) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d}}\right) \cdot V, \tag {2} \\ \end{array} +$$ + +where $\mathbf{z}_t\in \mathbb{R}^{B\times d_\alpha}$ denotes the flattened intermediate features of the decoder, $W_{q}\in \mathbb{R}^{d_{\alpha}\times d}$ , $W_{k}\in \mathbb{R}^{512\times d}$ , $W_{v}\in \mathbb{R}^{512\times d}$ are learnable parameters. Similar to C-GSP, we post-process the output $\pmb{o}$ via the $\tanh (\cdot)$ smooth projection to obtain the $\ell_{\infty}$ constrained perturbation with budge $\epsilon$ , i.e., $\delta = P(o) = \epsilon \cdot \tanh (o)$ . + +![](images/535006a5d511924f267f7362672a9da371d50584ba9b950433af67444248dd16.jpg) +(a) + +![](images/7b805cc33c4da89426f50342466f93d72f015ea4339a44f34f6568682834403e.jpg) +(b) +Fig. 3: Illustration of the proposed masked fine-tuning mechanism. (a) We fix the condition input with different text prompts to fine-tune the trained conditional generator $G_{w}$ into multiple generators for single-target attacks. (b) The fooling rate of several target classes with Inc-v3 and Res-152 as substitute models respectively. The results indicate that direct fine-tuning yields inadequate results for certain classes due to overfitting. We efficiently resolve this issue via a patch-wise random mask operation. + +Optimization objective. Based on the proposed CLIP-guided architecture, the optimization objective can be formulated as: + +$$ +w ^ {*} \leftarrow \arg \min _ {w} \mathcal {L} \left(f _ {\theta} \left(\boldsymbol {x} _ {s} + G _ {w} (\boldsymbol {x} _ {s}, \Phi (\boldsymbol {t} _ {c}))\right), c\right), \tag {3} +$$ + +where $t_c$ represents the corresponding text prompts of the target classes $c$ , $\mathcal{L}(\cdot, \cdot)$ denotes the cross-entropy loss. By encouraging the network to craft adversarial samples that can misguide the surrogate model's output toward the desired labels, the generator learns the data distribution of the target class and thus exhibits great generalizability in producing perturbation for any input data. + +# 3.3 Masked Fine-Tuning Mechanism + +In addition to the proposed CLIP-guided generator for multi-target attack scenarios, we also design a single-target variant for further improved performance. As depicted in Fig. 3a, we fix the conditioning input with the text description of a specific target class, and fine-tune the trained multi-target generator by resolving the objective function in Eq. (3) on the unlabeled training data $\mathcal{X}_s$ . This strategy further enhances the attack effects by fine-tuning the conditional generator to specialize in a specific target class. + +However, we encounter the overfitting problem that leads to limited improvement in success rates or even performance degradation for certain target classes. This problem is partly attributed to the fact that the generated adversarial perturbation sometimes heavily focuses on specific regions of the input image [62]. To alleviate this issue, we adopt a patch-wise random mask operation to postprocess the adversarial perturbation produced by the generator, which brings a notable increase in the targeted fooling rate as shown in Fig. 3b. This technique improves the capability and flexibility of our method, as an adversary can perform fine-tuning to further augment the generator for single-target attacks, achieving a better trade-off between efficiency and performance compared to previous single-target methods (See Section 4.5 for detailed analysis). + +# 4 Experiments + +To validate the effectiveness of CGNC in boosting transferability, we conduct extensive experiments on various black-box models in a range of scenarios. Our code is available at https://github.com/ffhibnese/CGNC_Targeted_Adversarial_Attacks. Please see the Appendix for more experimental results. + +# 4.1 Experimental Settings + +Datasets. Following [19,67], we train the generator on the ImageNet training set [9] and evaluate the attack performance using ImageNet-NeurIPS (1k) dataset proposed by [50]. We also conduct experiments in a more realistic cross-domain scenario where we train the generator and generate perturbations on MS-COCO [41] or Comics [4] while evaluating on target classifier trained on ImageNet. + +Victim Models. We use various victim models with different architectures. Specifically, the naturally trained models include Inception-v3 (Inc-v3) [57], Inception-v4 (Inc-v4) [55], Inception-ResNet-v2 (Inc-Res-v2) [55], ResNet-152 (Res-152) [29], DenseNet-121 (DN-121) [32], GoogleNet [56], and VGG-16 [54]. + +For further evaluation, we also analyze on the robust-trained models, including adv-Inception-v3 (Inc-v3adv) [26], ens-adv-Inception-ResNet-v2 (IR-v2ens) [28], and robust-trained ResNet-50 [24,31], dubbed as Res50SIN (trained on stylized ImageNet), Res50IN (trained on the mixture of stylized and Nature ImageNet), Res50fine (Res50IN plus further finetuning with an auxiliary dataset [24]), and Res50Aug (trained with the advanced data augmentation Augmix [31]). + +Baseline Attacks. We reveal the superiority of the proposed CGNC in enhancing multi-target attacks by comparing our method with multiple competitive baselines, including MIM [11], DIM [65], SIM [40], DIM [12], Logit [72], SU [63], and the state-of-the-art (SOTA) multi-target generative attacks CGSP [67] in Section 4.2. For SU attack [63], we choose to compare with its best version DTMI-Logit-SU. Besides, we also provide a comparison of the proposed single-target variant of CGNC with the existing single-target attack methods, i.e., GAP [51], CD-AP [48], TTP [49], and DGTA-PI [19] in Section 4.5. + +Implementation Details. Following [19,67], we employ Inc-v3 and Res-152 as surrogate models and adopt the perturbation budget $\epsilon$ 16/255. We conduct 10 epochs of generator training using a learning rate of 2e-4. During the masked fine-tuning, we maintain the learning rate and apply a mask ratio of 0.2 to fine-tune the text-conditional generator for an additional 5 epochs. + +# 4.2 Multi-Target Transferability Evaluation + +To align the experimental setup to former works [19,67], we first target 8 different classes in [71] and provide the average attack success rates (ASR) of the 8 target categories as evaluation metrics. Attack performance under larger numbers of target categories is presented in the next section. + +Attacks against regularly trained models. We first perform attacks on normal models to evaluate the multi-target attack performance. By observing + +Table 1: Attack success rates (\%) for multi-target attacks against regularly trained models on ImageNet NeurIPS validation set. * represents white-box attacks. + +
SourceMethodInc-v3Inc-v4Inc-Res-v2Res-152DN-121GoogleNetVGG-16
Inc-v3MIM99.90*0.801.000.400.200.200.30
TI-MIM98.50*0.500.500.300.200.400.40
SI-MIM99.80*1.502.000.800.700.700.50
DIM95.60*2.700.500.801.100.400.80
TI-DIM96.00*1.101.200.500.500.500.80
SI-DIM90.20*3.804.402.002.201.701.40
Logit99.60*5.606.501.703.000.801.50
SU99.59*5.807.003.353.502.003.94
C-GSP93.40*46.5836.7441.6046.4040.0045.00
CGNC96.03*59.4348.0642.4862.9851.3352.54
Res-152MIM0.500.400.6099.70*0.300.300.20
TI-MIM0.300.300.3096.50*0.300.400.30
SI-MIM1.301.201.6099.50*1.001.400.70
DIM2.302.203.0092.30*0.200.800.70
TI-DIM0.800.701.0090.60*0.600.800.50
SI-DIM4.204.805.4090.50*4.203.602.00
Logit10.1010.7012.8095.70*12.703.709.20
SU12.3611.3116.1695.08*16.136.5514.28
C-GSP37.7033.3320.2893.20*64.2041.7045.90
CGNC53.3951.5334.2495.85*85.6662.2363.36
+ +the results in Table 1, we demonstrate that the proposed CGNC consistently achieves significant improvement compared with previous methods. Specifically, our method achieves an average improvement of $17.88\%$ and $10.08\%$ in ASR over the C-GSP [67] attack regarding Res-152 and Inc-v3 as surrogate models respectively, demonstrating the superiority of leveraging the rich prior knowledge from CLIP's text embedding. We also note that the iterative methods obtain nearly $100\%$ while-box ASR while receiving poor performance on black-box models due to overfitting the classification boundaries of surrogate models. + +Attacks under defense strategies. For a more thorough analysis and comparison, we then compare these attacks under several widely used defenses. Firstly, we consider attacking six robustly trained networks in Table 2. Generally, our method is still able to deceive the black-box classifiers into predicting the specified classes, significantly outperforming former multi-target attacks, e.g., a $20.87\%$ increase of fooling rate from Res-152 to $\mathrm{Res}50_{\mathrm{Aug}}$ . + +Next, we evaluate the performance of different approaches on models with input preprocessing-based defenses, including a set of image smoothing mechanisms [10] and JPEG compression [14] algorithms. As shown in Table 3, although these defenses eliminate certain valid information in the adversarial samples, the CLIP-empowered CGNC still demonstrates excellent capability in boosting targeted transferability. Particularly when the substitute model is Res-152, CGNC achieves an average fooling rate of $71.18\%$ and $80.34\%$ on smoothing methods and JPEG compression, while C-GSP only reaches $49.05\%$ and $56.16\%$ respectively, verifying the stability and robustness of the proposed method. + +Perturbation Visualization. We present visualization results in Fig. 4 to unveil the principle of our method. Upon careful inspection, we can see that the generated perturbations mainly focus on the semantic regions of the input + +Table 2: Comparison of the proposed CGNC with the SOTA multi-target attacks against models with robust training mechanism on ImageNet NeurIPS validation set. + +
SourceMethodInc-v3advIR-v2ensRes50SINRes50INRes50fineRes50Aug
Inc-v3MIM0.160.100.200.270.440.19
TI-MIM0.210.190.330.490.680.31
SI-MIM0.130.190.260.430.630.29
DIM0.110.090.160.330.390.19
TI-DIM0.150.130.160.210.330.14
SI-DIM0.190.210.430.710.840.46
Logit0.300.300.701.233.140.86
SU0.490.410.841.753.551.04
C-GSP20.4118.046.9633.7644.5621.95
CGNC24.3622.548.8540.8352.1822.85
Res-152MIM0.190.150.281.582.750.78
TI-MIM0.610.730.502.514.751.76
SI-MIM0.240.240.390.660.840.36
DIM0.630.370.948.5014.223.77
TI-DIM0.230.300.280.761.490.49
SI-DIM0.710.710.752.733.891.37
Logit1.151.181.656.7015.465.93
SU2.121.201.957.5321.146.95
C-GSP14.6016.0116.8460.3065.5142.88
CGNC22.2126.7129.8379.8084.0563.75
+ +Table 3: Comparison of our method with C-GSP under different defense strategies. Q indicates the quality factor in JPEG compression. Here the target model is DN-121 and results for more victim models are in the Appendix. + +
SourceMethodSmoothingJPEG compression
GaussianMediumAverageQ=70Q=75Q=80Q=85Q=90
Inc-v3C-GSP36.5946.9139.8649.4150.5551.9552.8453.39
CGNC43.0355.3545.9458.3559.2560.2861.2961.94
Res-152C-GSP43.2156.3847.5553.0354.5456.1357.7559.36
CGNC64.4479.8469.2577.3878.8180.5482.0582.94
+ +images and contain sufficient semantic patterns specific to the target category. For instance, when the condition is a photo of a sea lion, the resulting perturbations indeed carry rich patterns closely resembling this marine animal. We also observe that the pattern changes in accordance with the text prompts, which validates our conditioning mechanism of the CLIP-encoded textual embedding. + +# 4.3 Evaluation on More Scenarios + +Cross-Domain scenarios. Next, we explore the more realistic cross-domain scenarios [48, 49] where attackers do not know anything about the data distribution of the training set used by the black-box classifier. Based on this setting, attackers train the generator and generate adversarial perturbations using an auxiliary dataset that follows a different probability distribution from that of the target model. Specifically, we satisfy the cross-domain experimental setting using MS-COCO [41] and Comics [4] datasets respectively. MS-COCO is a large-scale image dataset widely used for object detection and semantic segmentation, + +![](images/cf93e2fe4dc4048988469f40089162bb690ec6d1950e0b9c3b4e8c9a9315f521.jpg) +Fig. 4: Visualization results of different input images for different targets. For each text prompt of the target class, the left column shows the perturbation generated by our CGNC while the right column displays the corresponding adversarial examples. + +![](images/b7330cdcc02e3b51f99848fca2ac8c5c2d3a8d0e7ef9979bd59b5067be897873.jpg) +Fig.5: Fooling rates (from Res-152 to VGG-16) in attacking 8 target classes on cross-domain scenarios. We also provide the results on ImageNet as a comparison. + +![](images/587f7b2c655aa79a529e370bb3b4f35a4ca3652f7575a9eb6ebfa8e5b62aa964.jpg) +Fig.6: Fooling rates of our CGNC with C-GSP on larger numbers of attacked classes regarding Res-152 and Inc-v3 as surrogate models. + +while Comic is composed of a large number of comic images that can be regarded as a stylized version of ImageNet. + +To analyze the cross-domain transferability, we randomly select 1000 images from the source dataset to craft adversarial samples using the trained generator. Fig. 5 shows the results of the cross-domain attacks transferring from three different source datasets to ImageNet. In general, the more challenging cross-domain datasets lead to varying degrees of reduction in the targeted fooling rate due to the domain gap, especially on the Comics dataset which differs greatly from the ImageNet distribution. Benefiting from the text guidance on the target category, CGNC still attains decent attack performance and remarkably outperforms C-GSP. This demonstrates that our method achieves better cross-domain transferability and is partly independent of the training dataset. + +In addition, since the SOTA single-target attacks [19,49,62] require samples of the target class for loss computation, they are not applicable in the cross- + +Table 4: Ablation study of CGNC and its two variants on ImageNet NeurIPS validation set for 8 target classes. The substitute model is Res-152. + +
MethodVGG-16GoogleNetInc-v3Res-152DN-201
CGNA-CA-t56.5551.0947.4492.8174.65
CGNA-CA56.6454.2949.7393.3475.99
CGNC63.3662.2353.3995.8582.69
+ +domain scenarios where the source dataset lacks images of the target category. This represents an additional advantage of our method over single-target attacks. + +Larger Numbers of Target Classes. We then increase the number of target categories to verify the effectiveness of our method when handling plenty of classes. Specifically, we condition CGNC with more text inputs corresponding to the increased number of target categories. The performance is evaluated across the aforementioned six black-box models. + +As mentioned before, single-target attacks become impractical for real-world classification systems with hundreds or thousands of target categories [27]. In comparison, our method effectively solves this issue and achieves great improvements in large numbers of target classes over C-GSP as in Fig. 6, e.g., $14.71\%$ increase for 200 target classes proxy on the Res-152. This again reveals the significance of the fully exploited textual guidance from the CLIP model. Additionally, our method exhibits greater robustness to the varying numbers of target classes, e.g., the performance of our network on Inc-v3 demonstrates a smoother decrease compared to that of C-GSP. This further highlights the superiority of our method in reducing computation costs. For instance, when launching 1000-class targeted attacks using Inc-v3 as the surrogate model, our method requires training only 5 CGNCs with 200 conditions, yet it can achieve comparable performance to 20 generators with 50 conditions trained using C-GSP. + +# 4.4 Ablation Study + +We conduct ablation experiments on the ImageNet NeurIPS dataset to study the effect of the proposed techniques. Specifically, we introduce two variants of CGNC. CGNC-CA removes the cross-attention module from the original network, and CGNC-CA-t further modifies the conditioning mechanism by replacing the CLIP's text embedding with one-hot labels. + +From the results in Table 4, we can find that each aforementioned technique can further improve the attack success rates. Moreover, the remarkable improvement from CGNC-CA to CGNC also confirms that these cross-attention modules in the CA-Decoder can make better use of the text guidance provided by the CLIP model to enhance the targeted transferability of crafted perturbations. + +# 4.5 Comparison with Single-Target Attacks + +Next, we compare the single-target variants of the proposed CGNC with multiple state-of-the-art single-target attacks. To obtain our single-target generators, we + +Table 5: Comparison of our method with the SOTA single-target attacks. $^\dagger$ denotes the single-target variant of our CGNC obtained through the masked fine-tuning technique. * represents the white-box attacks. + +
SourceMethodInc-v3Inc-v4Inc-Res-v2Res-152DN-121GoogleNetVGG-16
Inc-v3GAP86.90*45.0634.4834.4841.7426.8934.34
CD-AP94.20*57.6060.1037.1041.6032.3041.70
TTP91.37*46.0439.3716.4033.4725.8025.73
DGTA-PI94.63*67.9555.0350.5047.3847.6748.11
CGNC†98.84*74.7664.4862.0078.9469.0670.74
Res-152GAP30.9931.4320.4884.86*58.3529.8939.70
CD-AP33.3043.7042.7096.60*53.8036.6034.10
TTP62.0349.2038.7095.12*82.9665.0962.82
DGTA-PI66.8353.6247.6196.48*86.6168.2969.58
CGNC†68.8669.4545.7198.61*91.1469.8368.05
+ +conduct the proposed masked fine-tuning to the trained conditional generator eight times using eight different text prompts of the target classes. Quantitative results in Table 5 indicate that our single-target enhanced generators outperform the competing single-target attacks in most cases. Particularly on the surrogate model of Inc-v3, $\mathrm{CGNC}^{\dagger}$ achieves a notable increase of $15.36\%$ in the average black-box fooling rate compared to previous methods, demonstrating the significant effectiveness of the proposed masked fine-tuning. The MFT technique further enhances the single-target performance, thereby improving the adaptability and flexibility of our approach in scenarios with fewer target classes. + +Note that these single-target methods require training 8 generators from scratch for 8 different target classes. In contrast, our method only needs to train a single multi-target generator and perform fine-tuning 8 times, each time with just a few epochs. Based on the typical experimental setup used in the SOTA single-target methods [19,49], our strategy can diminish over 100 training epochs when targeting 8 classes, thus substantially mitigating the computational burden. More experiments of single-target attacks against adversarially robust models and input preprocessing defenses are shown in the Appendix. + +# 5 Conclusion + +In this paper, we design a novel generative network CGNC, which improves multi-target transferable adversarial attacks by fully utilizing the rich prior within the CLIP as auxiliary semantic knowledge about the target category. To better incorporate the prior information, we introduce the cross-attention modules and efficiently condition the generator with CLIP's text embeddings. Through extensive experiments, we demonstrate the effectiveness of the proposed CGNC on multiple black-box target models in a variety of scenarios. Moreover, we propose a masked fine-tuning technique to further enhance the CGNC in attacking a single class, which outperforms previous single-target methods in both efficiency and effectiveness. We hope that the proposed method can serve as a reliable tool for evaluating the model robustness under black-box setups, promoting further research on the vulnerability and robustness of DNNs. + +# Acknowledgments + +This work is supported in part by the National Natural Science Foundation of China under grant 62171248, 62301189, Guangdong Basic and Applied Basic Research Foundation under grant 2021A1515110066, the PCNL KEY project (PCL2021A07), and Shenzhen Science and Technology Program under Grant JCYJ20220818101012025, RCBS20221008093124061, GXWD20220811172936001. + +# References + +1. Akhtar, N., Mian, A.: Threat of adversarial attacks on deep learning in computer vision: A survey. IEEE Access 6, 14410-14430 (2018) +2. Andriushchenko, M., Croce, F., Flammarion, N., Hein, M.: Square attack: a query-efficient black-box adversarial attack via random search. In: European conference on computer vision. pp. 484-501. Springer (2020) +3. Athalye, A., Engstrom, L., Ilyas, A., Kwok, K.: Synthesizing robust adversarial examples. In: International conference on machine learning. pp. 284-293. PMLR (2018) +4. BircanoAYlu, C.: https://www.kaggle.com/datasets/cenkbircanoglu/comic-books-classification. Kaggle, 2017 +5. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: 2017 ieee symposium on security and privacy (sp). pp. 39-57. IEEE (2017) +6. Che, Z., Borji, A., Zhai, G., Ling, S., Li, J., Min, X., Guo, G., Le Callet, P.: Smgea: A new ensemble adversarial attack powered by long-term gradient memories. IEEE Transactions on Neural Networks and Learning Systems 33(3), 1051-1065 (2020) +7. Chefer, H., Benaim, S., Paiss, R., Wolf, L.: Image-based clip-guided essence transfer. In: European Conference on Computer Vision. pp. 695-711. Springer (2022) +8. Chen, B., Feng, Y., Dai, T., Bai, J., Jiang, Y., Xia, S.T., Wang, X.: Adversarial examples generation for deep product quantization networks on image retrieval. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(2), 1388-1404 (2022) +9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) +0. Ding, G.W., Wang, L., Jin, X.: Advertorch v0.1: An adversarial robustness toolbox based on pytorch. arXiv preprint arXiv:1902.07623 (2019) +1. Dong, Y., Liao, F., Pang, T., Su, H., Zhu, J., Hu, X., Li, J.: Boosting adversarial attacks with momentum. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9185-9193 (2018) +2. Dong, Y., Pang, T., Su, H., Zhu, J.: Evading defenses to transferable adversarial examples by translation-invariant attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4312-4321 (2019) +3. Du, J., Zhang, H., Zhou, J.T., Yang, Y., Feng, J.: Query-efficient meta attack to deep neural networks. arXiv preprint arXiv:1906.02398 (2019) +4. Dziugaite, G.K., Ghahramani, Z., Roy, D.M.: A study of the effect of jpg compression on adversarial images. arXiv preprint arXiv:1608.00853 (2016) +5. Eykholt, K., Evtimov, I., Fernandes, E., Li, B., Rahmati, A., Xiao, C., Prakash, A., Kohno, T., Song, D.: Robust physical-world attacks on deep learning visual classification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1625-1634 (2018) + +16. Fang, H., Chen, B., Wang, X., Wang, Z., Xia, S.T.: Gidf: A generative gradient inversion method with feature domain optimization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4967-4976 (2023) +17. Fang, H., Kong, J., Yu, W., Chen, B., Li, J., Xia, S., Xu, K.: One perturbation is enough: On generating universal adversarial perturbations against vision-language pre-training models. arXiv preprint arXiv:2406.05491 (2024) +18. Fang, H., Qiu, Y., Yu, H., Yu, W., Kong, J., Chong, B., Chen, B., Wang, X., Xia, S.T.: Privacy leakage on dnns: A survey of model inversion attacks and defenses. arXiv preprint arXiv:2402.04013 (2024) +19. Feng, W., Xu, N., Zhang, T., Zhang, Y.: Dynamic generative targeted attacks with pattern injection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16404-16414 (2023) +20. Gao, K., Bai, Y., Bai, J., Yang, Y., Xia, S.T.: Adversarial robustness for visual grounding of multimodal large language models. In: ICLR Workshop (2024) +21. Gao, K., Bai, Y., Gu, J., Xia, S.T., Torr, P., Li, Z., Liu, W.: Inducing high energy-latency of large vision-language models with verbose images. In: ICLR (2024) +22. Gao, K., Gu, J., Bai, Y., Xia, S.T., Torr, P., Liu, W., Li, Z.: Energy-latency manipulation of multi-modal large language models via verbose samples. arXiv preprint arXiv:2404.16557 (2024) +23. Gao, L., Cheng, Y., Zhang, Q., Xu, X., Song, J.: Feature space targeted attacks by statistic alignment. arXiv preprint arXiv:2105.11645 (2021) +24. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231 (2018) +25. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014) +26. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014) +27. Han, J., Dong, X., Zhang, R., Chen, D., Zhang, W., Yu, N., Luo, P., Wang, X.: Once a man: Towards multi-target attack via learning multi-target adversarial network once. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5158-5167 (2019) +28. Hang, J., Han, K., Chen, H., Li, Y.: Ensemble adversarial black-box attacks against deep learning systems. Pattern Recognition 101, 107184 (2020) +29. He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14. pp. 630-645. Springer (2016) +30. He, Z., Wang, W., Dong, J., Tan, T.: Revisiting ensemble adversarial attack. Signal Processing: Image Communication 107, 116747 (2022) +31. Hendrycks, D., Mu, N., Cubuk, E.D., Zoph, B., Gilmer, J., Lakshminarayanan, B.: Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781 (2019) +32. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017) +33. Huang, Z., Zhou, A., Ling, Z., Cai, M., Wang, H., Lee, Y.J.: A sentence speaks a thousand images: Domain generalization through distilling clip with language guidance. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11685-11695 (2023) + +34. Inkawich, N., Liang, K., Wang, B., Inkawich, M., Carin, L., Chen, Y.: Perturbing across the feature hierarchy to improve standard and strict blackbox attack transferability. Advances in Neural Information Processing Systems 33, 20791-20801 (2020) +35. Inkawich, N., Wen, W., Li, H.H., Chen, Y.: Feature space perturbations yield more transferable adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7066-7074 (2019) +36. Ju, C., Han, T., Zheng, K., Zhang, Y., Xie, W.: Prompting visual-language models for efficient video understanding. In: European Conference on Computer Vision. pp. 105-124. Springer (2022) +37. Kong, Z., Guo, J., Li, A., Liu, C.: Physgan: Generating physical-world-resilient adversarial examples for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14254-14263 (2020) +38. Kurakin, A., Goodfellow, I., Bengio, S.: Adversarial machine learning at scale. arXiv preprint arXiv:1611.01236 (2016) +39. Li, Q., Guo, Y., Chen, H.: Yet another intermediate-level attack. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16. pp. 241-257. Springer (2020) +40. Lin, J., Song, C., He, K., Wang, L., Hopcroft, J.E.: Nesterov accelerated gradient and scale invariance for adversarial attacks. arXiv preprint arXiv:1908.06281 (2019) +41. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +42. Liu, A., Wang, J., Liu, X., Cao, B., Zhang, C., Yu, H.: Bias-based universal adversarial patch attack for automatic check-out. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16. pp. 395-410. Springer (2020) +43. Lu, Y., Jia, Y., Wang, J., Li, B., Chai, W., Carin, L., Velipasalar, S.: Enhancing cross-task black-box transferability of adversarial examples with dispersion reduction. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 940-949 (2020) +44. Luo, J., Bai, T., Zhao, J.: Generating adversarial yet inconspicuous patches with a single image (student abstract). In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 15837-15838 (2021) +45. Moosavi-Dezfooli, S.M., Fawzi, A., Fawzi, O., Frossard, P.: Universal adversarial perturbations. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1765-1773 (2017) +46. Moosavi-Dezfooli, S.M., Fawzi, A., Frossard, P.: Deepfool: a simple and accurate method to fool deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2574-2582 (2016) +47. Mopuri, K.R., Uppala, P.K., Babu, R.V.: Ask, acquire, and attack: Data-free uap generation using class impressions. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 19-34 (2018) +48. Naseer, M.M., Khan, S.H., Khan, M.H., Shahbaz Khan, F., Porikli, F.: Cross-domain transferability of adversarial perturbations. Advances in Neural Information Processing Systems 32 (2019) +49. Naseer, M., Khan, S., Hayat, M., Khan, F.S., Porikli, F.: On generating transferable targeted perturbations. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7708-7717 (2021) + +50. NeurIPS: https://www.kaggle.com/c/nips-2017-defense-against-adversarial-attack/data.Kaggle,2017 +51. Poursaeed, O., Katsman, I., Gao, B., Belongie, S.: Generative adversarial perturbations. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4422-4431 (2018) +52. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +53. Sarkar, S.K., Oshiba, K., Giebisch, D., Singer, Y.: Robust classification of financial risk. arXiv preprint arXiv:1811.11079 (2018) +54. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) +55. Szegedy, C., Ioffe, S., Vanhoucke, V., Alemi, A.: Inception-v4, inception-resnet and the impact of residual connections on learning. In: Proceedings of the AAAI conference on artificial intelligence. vol. 31 (2017) +56. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A.: Going deeper with convolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1-9 (2015) +57. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2818-2826 (2016) +58. Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I., Fergus, R.: Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199 (2013) +59. Tao, M., Bao, B.K., Tang, H., Xu, C.: Galip: Generative adversarial clips for text-to-image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14214-14223 (2023) +60. Wang, X., He, K.: Enhancing the transferability of adversarial attacks through variance tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1924-1933 (2021) +61. Wang, X., He, K., Hopcroft, J.E.: At-gan: A generative attack model for adversarial transferring on generative adversarial nets. arXiv preprint arXiv:1904.07793 3(4), 3 (2019) +62. Wang, Z., Yang, H., Feng, Y., Sun, P., Guo, H., Zhang, Z., Ren, K.: Towards transferable targeted adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20534-20543 (2023) +63. Wei, Z., Chen, J., Wu, Z., Jiang, Y.G.: Enhancing the self-universality for transferable targeted attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12281-12290 (2023) +64. Xiao, C., Li, B., Zhu, J.Y., He, W., Liu, M., Song, D.: Generating adversarial examples with adversarial networks. arXiv preprint arXiv:1801.02610 (2018) +65. Xie, C., Zhang, Z., Zhou, Y., Bai, S., Wang, J., Ren, Z., Yuille, A.L.: Improving transferability of adversarial examples with input diversity. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2730-2739 (2019) +66. Xiong, Y., Lin, J., Zhang, M., Hopcroft, J.E., He, K.: Stochastic variance reduced ensemble adversarial attack for boosting the adversarial transferability. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14983-14992 (2022) + +67. Yang, X., Dong, Y., Pang, T., Su, H., Zhu, J.: Boosting transferability of targeted adversarial examples via hierarchical generative networks. In: European Conference on Computer Vision. pp. 725-742. Springer (2022) +68. Yu, W., Fang, H., Chen, B., Sui, X., Chen, C., Wu, H., Xia, S.T., Xu, K.: Gi-nas: Boosting gradient inversion attacks through adaptive neural architecture search. arXiv preprint arXiv:2405.20725 (2024) +69. Zha, Y., Ji, H., Li, J., Li, R., Dai, T., Chen, B., Wang, Z., Xia, S.T.: Towards compact 3d representations via point feature enhancement masked autoencoders. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 6962-6970 (2024) +70. Zha, Y., Wang, J., Dai, T., Chen, B., Wang, Z., Xia, S.T.: Instance-aware dynamic prompt tuning for pre-trained point cloud models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14161-14170 (2023) +71. Zhang, C., Benz, P., Imtiaz, T., Kweon, I.S.: Understanding adversarial examples from the mutual influence of images and perturbations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14521-14530 (2020) +72. Zhao, Z., Liu, Z., Larson, M.: On success and simplicity: A second look at transferable targeted attacks. Advances in Neural Information Processing Systems 34, 6115-6128 (2021) \ No newline at end of file diff --git a/2024/CLIP-Guided Generative Networks for Transferable Targeted Adversarial Attacks/images.zip b/2024/CLIP-Guided Generative Networks for Transferable Targeted Adversarial Attacks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ef9e9cb9a23d47eda35866ecf93074cac3316413 --- /dev/null +++ b/2024/CLIP-Guided Generative Networks for Transferable Targeted Adversarial Attacks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f582d83ee7440e3d608c8a124f3134232ae8eb7a4100b9319701a7e61e1db767 +size 517194 diff --git a/2024/CLIP-Guided Generative Networks for Transferable Targeted Adversarial Attacks/layout.json b/2024/CLIP-Guided Generative Networks for Transferable Targeted Adversarial Attacks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1a2c827a6149e434bce6053d867c24e1bb9fccd9 --- /dev/null +++ b/2024/CLIP-Guided Generative Networks for Transferable Targeted Adversarial Attacks/layout.json @@ -0,0 +1,9908 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "spans": [ + { + "bbox": [ + 159, + 111, + 455, + 148 + ], + "type": "text", + "content": "CLIP-Guided Generative Networks for Transferable Targeted Adversarial Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "spans": [ + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "text", + "content": "Hao Fang" + }, + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "text", + "content": ", Jiawei Kong" + }, + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "text", + "content": ", Bin Chen" + }, + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "inline_equation", + "content": "^{2,3,4\\#}" + }, + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "text", + "content": ", Tao Dai" + }, + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "text", + "content": ", Hao Wu" + }, + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "text", + "content": ", and Shu-Tao Xia" + }, + { + "bbox": [ + 153, + 167, + 462, + 194 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 153, + 201, + 461, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 201, + 461, + 224 + ], + "spans": [ + { + "bbox": [ + 153, + 201, + 461, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 153, + 201, + 461, + 224 + ], + "type": "text", + "content": " Tsinghua Shenzhen International Graduate School, Tsinghua University " + }, + { + "bbox": [ + 153, + 201, + 461, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 153, + 201, + 461, + 224 + ], + "type": "text", + "content": " Harbin Institute of Technology, Shenzhen " + }, + { + "bbox": [ + 153, + 201, + 461, + 224 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 153, + 201, + 461, + 224 + ], + "type": "text", + "content": " Pengcheng Laboratory" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 224, + 477, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 224, + 477, + 280 + ], + "spans": [ + { + "bbox": [ + 137, + 224, + 477, + 280 + ], + "type": "text", + "content": "4 Guangdong Provincial Key Laboratory of Novel Security Intelligence Technologies \n5 Shenzhen University 6 Shenzhen Digital Certificate Authority CO., Ltd \nfang-h23@mails.tsinghua.edu.cn, kongjiawei@stu.hit.edu.cn \nchenbin2021@hit.edu.cn, daitao.edu@gmail.com, whpc79@163.com, \nxiast@sz.tsinghua.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 308, + 455, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 308, + 455, + 485 + ], + "spans": [ + { + "bbox": [ + 159, + 308, + 455, + 485 + ], + "type": "text", + "content": "Abstract. Transferable targeted adversarial attacks aim to mislead models into outputting adversary-specified predictions in black-box scenarios. Recent studies have introduced single-target attacks that train a generator for each target class to generate highly transferable perturbations, resulting in substantial computational overhead when handling multiple classes. Multi-target attacks address this by training only one class-conditional generator for multiple classes. However, the generator simply uses class labels as conditions, failing to leverage the rich semantic information of the target class. To this end, we design a CLIP-guided Generative Network with Cross-attention modules (CGNC) to enhance multi-target attacks by incorporating textual knowledge of CLIP into the generator. Extensive experiments demonstrate that CGNC yields significant improvements over previous multi-target attacks, e.g., a " + }, + { + "bbox": [ + 159, + 308, + 455, + 485 + ], + "type": "inline_equation", + "content": "21.46\\%" + }, + { + "bbox": [ + 159, + 308, + 455, + 485 + ], + "type": "text", + "content": " improvement in success rate from Res-152 to DenseNet-121. Moreover, we propose the masked fine-tuning to further strengthen our method in attacking a single class, which surpasses existing single-target methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 503, + 231, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 231, + 516 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 231, + 516 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 528, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 482, + 636 + ], + "type": "text", + "content": "With the rapid progress of deep learning, deep neural networks (DNNs) have been widely applied in many security-critical fields, such as autonomous driving [15, 37], financial systems [53], and point cloud modeling [69, 70]. However, DNNs are corroborated to be vulnerable to adversarial attacks [20, 26, 58], which attempt to fool models with adversarial examples crafted by adding imperceptible perturbations to the original inputs. Based on the attack goal, adversarial attacks can be categorized into untargeted and targeted attacks. Untargeted attacks attempt to fool DNNs into predicting incorrect labels while targeted attacks aim at triggering the model to output the attacker-desired predictions." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 642, + 227, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 642, + 227, + 653 + ], + "spans": [ + { + "bbox": [ + 142, + 642, + 227, + 653 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 142, + 642, + 227, + 653 + ], + "type": "text", + "content": "Equal contribution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 142, + 653, + 242, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 653, + 242, + 665 + ], + "spans": [ + { + "bbox": [ + 142, + 653, + 242, + 665 + ], + "type": "text", + "content": "Corresponding author." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 116, + 339, + 215 + ], + "blocks": [ + { + "bbox": [ + 151, + 116, + 339, + 215 + ], + "lines": [ + { + "bbox": [ + 151, + 116, + 339, + 215 + ], + "spans": [ + { + "bbox": [ + 151, + 116, + 339, + 215 + ], + "type": "image", + "image_path": "e597ef4133afdf3eb4daa87c93fd652ba2630da50cbeec01b0458fd1c5157721.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 216, + 252, + 224 + ], + "lines": [ + { + "bbox": [ + 241, + 216, + 252, + 224 + ], + "spans": [ + { + "bbox": [ + 241, + 216, + 252, + 224 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 345, + 129, + 462, + 213 + ], + "blocks": [ + { + "bbox": [ + 345, + 129, + 462, + 213 + ], + "lines": [ + { + "bbox": [ + 345, + 129, + 462, + 213 + ], + "spans": [ + { + "bbox": [ + 345, + 129, + 462, + 213 + ], + "type": "image", + "image_path": "e07c63107175448b06783303ae2d73a0281fb1165b7aef9c1dfcac5908d528e7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 215, + 422, + 224 + ], + "lines": [ + { + "bbox": [ + 410, + 215, + 422, + 224 + ], + "spans": [ + { + "bbox": [ + 410, + 215, + 422, + 224 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 234, + 482, + 323 + ], + "lines": [ + { + "bbox": [ + 130, + 234, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 482, + 323 + ], + "type": "text", + "content": "Fig. 1: (a) Targeted attacks from 'Panda' to 'Dog'. Fig. (a) illustrates that previous multi-target methods [27,67] generate perturbations simply conditioned by class indices or one-hot vectors and only learn the classification boundary specific to the surrogate model. In contrast, CGNC exploits CLIP's meaningful guidance to effectively capture the feature distribution inherent to the target data. (b) We directly feed the scaled perturbations generated by our CGNC and C-GSP [67] into three black-box classifiers and reveal that our perturbations achieve significantly higher mean confidence of the target class, demonstrating its superiority in modeling the target feature distribution." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 351, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 567 + ], + "type": "text", + "content": "Recent investigations into the adversarial transferability [2, 3, 44, 58] have demonstrated that adversarial examples crafted for a white-box surrogate model can also mislead other unseen black-box models. Since this attack does not require access to the target model, it exposes a serious security threat to real-world applications of DNNs and motivates a wide range of studies [1,5,42,46]. Despite the remarkable performance on untargeted transferable attacks, these approaches produce unsatisfactory results for targeted attacks due to their over-reliance on the white-box surrogate. Existing studies on transferable targeted attacks can be categorized into instance-specific [11, 15, 23, 39, 43, 60, 66] and instance-agnostic attacks [19, 37, 44, 48, 49, 64]. Specifically, instance-specific attacks [11, 12, 65] iteratively perform gradient updating to craft adversarial perturbation tailored to a specific natural sample, yet producing low-targeted transferability due to overfitting [67]. Conversely, instance-agnostic attacks learn a universal perturbation [45, 71] or a perturbation generator [48, 51] based on data distribution rather than the specific instance, alleviating the data-specific over-fitting issues and achieving a higher adversarial transferability. Recent studies further explore generative attacks to produce highly transferable perturbations, which can be divided into single-target and multi-target attacks as follows." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Single-target attacks [19, 48, 49, 62] exhibit impressive performance by training perturbation generators for target categories. However, they require training a generator for each target, which can lead to a heavy computation burden when attacking numerous classes, and thus are not applicable in real-world classification systems that usually contain hundreds/thousands of target classes [27]. To address this, [27, 67] propose multi-target attacks that train a single conditional generator for multi-target classes. By specifying the desired label as conditioning input, the trained generator can efficiently generate the corresponding adversar" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "H. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "ial perturbation. Nevertheless, these methods simply adopt class indices [27] or one-hot vectors [67] of the target labels as conditions, and thus solely rely on the classification information from the surrogate model as the guidance of the target category, hence resulting in only modest black-box fooling rates." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 366 + ], + "type": "text", + "content": "In this paper, we build upon the research line of multi-target attacks and propose a novel CLIP-guided Generative Network with Cross-attention modules (CGNC) to boost the attack effect. Concretely, we revisit the architecture of the conditional generators used in [27,67] and argue that the simple class conditions, e.g., class indices or handcrafted one-hot vectors, limit the transferability of the generated perturbations (see Fig. 1(a) for comparison). Motivated by the impressive effects of CLIP's encoded text embeddings in multi-modal learning, we introduce concise text descriptions of the target classes to encode them as class-specific representations encapsulated with abundant information, which assist the generator in learning target class distribution and ultimately lead to a fundamental transferability improvement. To better incorporate the text information, we improve the condition-inputting mechanism by adding the cross-attention layers that have been proven effective in models with various input modalities. Results in Fig. 1(b) confirm the capability of our CGNC in capturing the target distribution. Besides, we propose a masked fine-tuning (MFT) technique, which fine-tunes the trained conditional generator with a fixed text condition of the desired class for further improvement in attacking a single class." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 367, + 480, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 480, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 480, + 451 + ], + "type": "text", + "content": "With all the above efforts, our method achieves great efficiency and scalability. When there are hundreds of target classes and single-target methods become impractical, the proposed conditional network can be utilized to achieve significant performance over previous multi-target attacks. Conversely, for scenarios involving only a few target classes, the proposed MFT mechanism enhances CGNC to outperform existing single-target methods while substantially reducing computational costs. In summary, our main contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 459, + 479, + 541 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 459, + 479, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 479, + 493 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 479, + 493 + ], + "type": "text", + "content": "- We propose CGNC, a novel CLIP-guided generative network with cross-attention layers that fully exploits the textual knowledge provided by the advanced CLIP model for enhanced multi-target attacks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 495, + 479, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 495, + 479, + 517 + ], + "spans": [ + { + "bbox": [ + 138, + 495, + 479, + 517 + ], + "type": "text", + "content": "- We introduce a masked fine-tuning mechanism to improve the single-target performance by adapting the CGNC to an individual target class." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 518, + 479, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 518, + 479, + 541 + ], + "spans": [ + { + "bbox": [ + 138, + 518, + 479, + 541 + ], + "type": "text", + "content": "- Extensive experiments show that CGNC remarkably enhances the targeted transferability compared to previous attack methods in a range of settings." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 560, + 237, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 560, + 237, + 573 + ], + "spans": [ + { + "bbox": [ + 132, + 560, + 237, + 573 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 586, + 284, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 284, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 284, + 597 + ], + "type": "text", + "content": "2.1 Vision-Language Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "content": "Vision-language models (VLM) drawn great attention [21,22] due to their promising potential in learning general visual and textual representations through contrastive pre-training on large-scale image-text pairs. Given the powerful capacity of text descriptions in modeling multi-modal tasks, these models can be effectively adapted to diverse downstream tasks through appropriately formulated" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "type": "text", + "content": "CLIP-Guided Networks for Transferable Targeted Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "textual prompts, including video understanding [36], image manipulation [7], and text-to-image synthesis [59]. Inspired by these works, we propose to harness CLIP-based multi-modal learning to empower our perturbation generator, facilitating more effective learning of the target feature distribution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 185, + 261, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 261, + 197 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 261, + 197 + ], + "type": "text", + "content": "2.2 Adversarial Attacks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 209, + 481, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 209, + 481, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 481, + 243 + ], + "type": "text", + "content": "Among various security threats to DNNs [8, 16, 18, 68], the adversarial attack is one of the most formidable and well-known one, which can be classified into instance-specific attacks and instance-agnostic attacks [67]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 245, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 245, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 482, + 376 + ], + "type": "text", + "content": "Instance-specific Attacks. Since the pioneering work [58] highlighted the vulnerability of neural networks, numerous gradient-based optimization methods [3, 13, 26, 38, 66] have been proposed to craft image-dependent perturbations. MIM [11] integrates a momentum item into the gradient update for utilizing the previous gradient information to avoid plunging into poor local optimum. DIM [65] enhances the transferability by randomly diversifying the input pattern, and TIM [12] implements the attack by convolving the gradient with a predefined kernel. In addition, intermediate feature space [35, 63] and classifier information [34] are exploited to enhance attack effects, while [72] leverages logit-based loss to achieve competitive results. More advanced works [6,28,30] consider an ensemble of multiple surrogate models to reduce over-fitting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 377, + 488, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 488, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 488, + 484 + ], + "type": "text", + "content": "Instance-agnostic Attacks. In contrast to instance-specific attacks, instance-agnostic attacks learn a universal perturbation [17, 45] or a generative model [44, 48, 49, 51, 61, 67] for crafting adversarial examples. By modeling the global data distribution, these methods have shown better transferability and attracted more attention in recent years. Specifically, UAP [45] and AAA [47] learn a universal perturbation to fool the model based on concrete data and compressed impression, respectively. Many subsequent works, such as GAP [51], focus on boosting the transferability of adversarial attacks by training generative models. These generative attacks can be categorized into the following two types." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 485, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 581 + ], + "type": "text", + "content": "Single-target Attacks. This type of attack requires training a generator for each target class. [64] first introduce the generative adversarial networks (GAN) [25] to generate adversarial perturbations. Then, CD-AP [48] concentrates on the domain-invariant adversaries and launches highly transferable cross-domain attacks using a relativistic supervisory. TTP [49] modifies the loss function and proposes to match the target distribution to mitigate over-fitting to the surrogate model. Subsequent methods achieve better transferability based on a dynamic network with pattern injection [19] or a feature discriminator [62]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 484, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 484, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 484, + 666 + ], + "type": "text", + "content": "Multi-target Attacks. MAN [27] notes that when dealing with numerous classes, single-target attacks inevitably suffer from extreme computational burdens as they need to train the same number of models as multiple target classes. Therefore, single-target attacks become impractical in attacking real classification systems that often have hundreds of categories. To pursue the extreme speed and storage, MAN trains only one model for 1000 target categories from ImageNet [9]. However, the excess of the target class severely degrades the targeted" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "H. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": "transferability. C-GSP [67] improves the performance by designing a hierarchical partition mechanism to divide all classes into a feasible number of subsets and train generators for each subset. Nonetheless, C-GSP simply conditions the generator with one-hot vectors and only utilizes the classification information to train the generator, failing to use the semantic knowledge of the target class. To overcome the limitation, we propose to incorporate the text information provided by CLIP as significant guidance for the target class." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 217, + 202, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 217, + 202, + 229 + ], + "spans": [ + { + "bbox": [ + 132, + 217, + 202, + 229 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 240, + 479, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 240, + 479, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 479, + 300 + ], + "type": "text", + "content": "In this section, we first introduce the preliminaries of targeted transferable attacks and present the basic paradigm of generative attack methods. Then, we elaborate on the proposed CGNC, which remarkably enhances multi-target black-box attacks. Finally, we detail the proposed masked fine-tuning technique that strengthens our method in single-class attacks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 316, + 221, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 316, + 221, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 221, + 329 + ], + "type": "text", + "content": "3.1 Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "text", + "content": "We denote the white-box image classifier parameterized with " + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "inline_equation", + "content": "f_{\\theta}:\\mathcal{X}\\rightarrow \\mathcal{Y}" + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{X}\\subset \\mathbb{R}^{N\\times H\\times W}" + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "text", + "content": " represents the image domain and " + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{V}\\subset \\mathbb{R}^L" + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "text", + "content": " is the output confidence score of different classes " + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "inline_equation", + "content": "(H,W,N,L" + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "text", + "content": " being height, width, number of channels, and class number). Given a natural image " + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "inline_equation", + "content": "\\pmb {x}\\in \\mathcal{X}" + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "text", + "content": " and the attacker's desired label " + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "inline_equation", + "content": "c_{t}\\in \\mathcal{C}" + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "text", + "content": ", the transferable targeted adversarial attacks attempt to craft the imperceptible perturbation " + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "text", + "content": " based on the accessible surrogate model " + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "text", + "content": " to mislead another unseen victim model " + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "inline_equation", + "content": "F_{\\phi}" + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "text", + "content": " into predicting " + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "inline_equation", + "content": "\\arg \\max_{i\\in \\mathcal{C}}F_{\\phi}(\\pmb {x} + \\pmb {\\delta})_i = c_t" + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "text", + "content": ". Concurrently, the " + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "text", + "content": " norm is employed to ensure the adversarial samples are indistinguishable from the original images by constraining the perturbation within the range " + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "inline_equation", + "content": "\\| \\pmb {\\delta}\\|_{\\infty}\\leq \\epsilon" + }, + { + "bbox": [ + 130, + 335, + 480, + 455 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 455, + 479, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 479, + 551 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 479, + 551 + ], + "type": "text", + "content": "To boost the fooling rate of targeted black-box attacks, single-target attacks [48, 49, 51] utilize powerful generative models to learn the target data distribution and achieve higher transferability. However, these methods consume great computation time and resources for multi-target scenarios, making them impractical for real-world attacks. C-GSP [67] effectively solves this issue by formulating the multi-target attacks as learning a class-conditional generator " + }, + { + "bbox": [ + 130, + 455, + 479, + 551 + ], + "type": "inline_equation", + "content": "G_{w}" + }, + { + "bbox": [ + 130, + 455, + 479, + 551 + ], + "type": "text", + "content": " with weights " + }, + { + "bbox": [ + 130, + 455, + 479, + 551 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 455, + 479, + 551 + ], + "type": "text", + "content": ", which is capable of generating perturbations for any target class. Given an unlabeled training dataset " + }, + { + "bbox": [ + 130, + 455, + 479, + 551 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_s" + }, + { + "bbox": [ + 130, + 455, + 479, + 551 + ], + "type": "text", + "content": ", the optimization objective is as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 165, + 571, + 480, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 571, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 165, + 571, + 480, + 587 + ], + "type": "interline_equation", + "content": "\\min _ {w} \\mathbb {E} _ {\\boldsymbol {x} _ {s} \\sim \\chi_ {s}, c \\sim \\mathcal {C}} [ \\mathcal {L} (f _ {\\theta} (\\boldsymbol {x} _ {s} + G _ {w} (\\boldsymbol {x} _ {s}, c)), c) ], \\text {s . t .} \\| G _ {w} (\\boldsymbol {x} _ {s}, c) \\| _ {\\infty} \\leq \\epsilon , \\tag {1}", + "image_path": "ddd610ce7d673f48205887f7233f04be7b4777b401ea8c32efd9615dc6b3c2bf.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "text", + "content": " denotes the cross-entropy (CE) loss. By minimizing the loss of specified classes using various unlabeled images from " + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_s" + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "text", + "content": ", we optimize the parameters " + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "text", + "content": " of the generative model and finally obtain a targeted conditional generator that can generate perturbation for any given clean image " + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "text", + "content": " from the test dataset " + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_t" + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "text", + "content": ". Specifically, the adversary can simply specify a target label " + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "text", + "content": " and craft an corresponding adversarial example via " + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{adv} = \\pmb{x}_t + G_w(\\pmb{x}_t, c)" + }, + { + "bbox": [ + 130, + 594, + 480, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "type": "text", + "content": "CLIP-Guided Networks for Transferable Targeted Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 118, + 479, + 208 + ], + "blocks": [ + { + "bbox": [ + 134, + 118, + 479, + 208 + ], + "lines": [ + { + "bbox": [ + 134, + 118, + 479, + 208 + ], + "spans": [ + { + "bbox": [ + 134, + 118, + 479, + 208 + ], + "type": "image", + "image_path": "6ad7695640e03bc51b244d528ec508da07dde1217990c7f38aaa3b5759aa9e41.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 228, + 482, + 275 + ], + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 275 + ], + "type": "text", + "content": "Fig. 2: An overview of our proposed architecture of CGNC. Equipped with the three exquisite modules VL-Purifier, F-Encoder, and CA-Decoder, the generator fully leverages the textual representations encoded by CLIP as auxiliary information about the target classes to better probe their data distribution and achieve superior attack effects." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 298, + 482, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 482, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 482, + 395 + ], + "type": "text", + "content": "However, current multi-target methods [27, 67] simply condition the generator with class labels and learn the target distribution only relying on the classification information of the surrogate model, thus not fully exploiting the specific characteristics of the target category. Inspired by the efficiency of vision-language (VL) learning [52, 59], we propose a novel generative network that leverages sufficient prior knowledge from the powerful CLIP model by incorporating textual-modality information to promote the target class modeling, which greatly enhances the multi-target transferable attacks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 415, + 334, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 415, + 334, + 426 + ], + "spans": [ + { + "bbox": [ + 131, + 415, + 334, + 426 + ], + "type": "text", + "content": "3.2 CLIP-Guided Generative Network" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 496 + ], + "type": "text", + "content": "The proposed generative model architecture is presented in Fig. 2. Specifically, CGNC is composed of a Vision-Language feature Purifier (VL-Purifier), a feature Fusion Encoder (F-Encoder), and a Cross-Attention based Decoder (CA-Decoder). We also provide the pseudocode of the training procedure in Algorithm 1. Next, we illustrate the design of each module as follows." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": "Vision-language feature purifier (VL-Purifier). To utilize CLIP to produce semantic embeddings, we first feed in CLIP's text encoder " + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": " with queries " + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "inline_equation", + "content": "t_c" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": " that follow the handcrafted template \"a photo of a {class}\" which has shown effectiveness in many tasks [33,52]. Since the obtained embeddings " + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{e}_t\\in \\mathbb{R}^{B\\times 512}" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": " being the batch size) in CLIP's vision-language space are generic representations of the target classes and are not yet tailored to our learning task, we refine them using the VL-Purifier, which is composed of several blocks consisting of a fully-connected layer and a spectral normalization layer. Through this module, we translate the encoded embeddings into more meaningful representations " + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{e}_t^*\\in \\mathbb{R}^{B\\times 16}" + }, + { + "bbox": [ + 130, + 497, + 482, + 617 + ], + "type": "text", + "content": ", thereby facilitating the subsequent step of feature fusion." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Feature fusion encoder (F-Encoder). This module aims to fuse the purified features " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{e}_t^*" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " with the learned visual representations. Firstly, a batch of input images " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{x}_s" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " is encoded to capture the visual concepts " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{h}_s \\in \\mathbb{R}^{B \\times N' \\times H' \\times W'}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ". Then, we expand the text embedding " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{e}_t^* \\in \\mathbb{R}^{B \\times 16}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{e}_t^{*'} \\in \\mathbb{R}^{B \\times 16 \\times H' \\times W'}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", which are" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "H. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 130, + 481, + 287 + ], + "blocks": [ + { + "bbox": [ + 132, + 115, + 470, + 129 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 470, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 470, + 129 + ], + "type": "text", + "content": "Algorithm 1 Pseudocode of Training the CLIP-guided Generative Network" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "lines": [ + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "spans": [ + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "text", + "content": "Require: " + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_s" + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "text", + "content": ": the training data; " + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "text", + "content": ": the target label space; " + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "text", + "content": ": the text prompts set; " + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "text", + "content": ": the surrogate model; " + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "text", + "content": ": the CLIP's text encoder; " + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "text", + "content": ": the max iteration; \nEnsure: the perturbation generator " + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "inline_equation", + "content": "G_w" + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "text", + "content": "; \n1: for " + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "inline_equation", + "content": "i \\gets 0" + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "text", + "content": " do \n2: Sample a batch of images " + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_s \\sim \\mathcal{X}_s" + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "text", + "content": "; \n3: Obtain " + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_s'" + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "text", + "content": " by processing " + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_s" + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "text", + "content": " with data augmentation; \n4: Sample a batch of target labels " + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "inline_equation", + "content": "c \\sim \\mathcal{C}" + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "text", + "content": "; \n5: Obtain the corresponding text prompts " + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "inline_equation", + "content": "t_c" + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "text", + "content": "; \n6: Compute the text embedding " + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "inline_equation", + "content": "\\boldsymbol{e}_t" + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "text", + "content": " by feeding " + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "inline_equation", + "content": "t_c" + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "text", + "content": "; \n7: Obtain perturbed images by " + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{adv} = \\boldsymbol{x}_s + G_w(\\boldsymbol{x}_s, \\boldsymbol{e}_t)" + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{adv}' = \\boldsymbol{x}_s' + G_w(\\boldsymbol{x}_s', \\boldsymbol{e}_t)" + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "text", + "content": "; \n8: Forward pass " + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{adv}" + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{adv}'" + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "text", + "content": " and compute the loss in Eq. (3); \n9: Perform gradient backpropagation and update the generator " + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "inline_equation", + "content": "G_w" + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "text", + "content": "; \n10: end for \n11: return the trained generator " + }, + { + "bbox": [ + 132, + 130, + 481, + 287 + ], + "type": "inline_equation", + "content": "G_w" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 318, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 426 + ], + "type": "text", + "content": "then integrated with the extracted visual concepts " + }, + { + "bbox": [ + 130, + 318, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\pmb{h}_s" + }, + { + "bbox": [ + 130, + 318, + 482, + 426 + ], + "type": "text", + "content": " through channel-wise concatenation to obtain the fused representations, i.e., " + }, + { + "bbox": [ + 130, + 318, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\pmb{m} \\in \\mathbb{R}^{B \\times (N' + 16) \\times H' \\times W'}" + }, + { + "bbox": [ + 130, + 318, + 482, + 426 + ], + "type": "text", + "content": ". Subsequently, " + }, + { + "bbox": [ + 130, + 318, + 482, + 426 + ], + "type": "inline_equation", + "content": "\\pmb{m}" + }, + { + "bbox": [ + 130, + 318, + 482, + 426 + ], + "type": "text", + "content": " undergoes further downsampling, and the resultant features are again concatenated with the expanded embedding " + }, + { + "bbox": [ + 130, + 318, + 482, + 426 + ], + "type": "inline_equation", + "content": "e_t^{*'}" + }, + { + "bbox": [ + 130, + 318, + 482, + 426 + ], + "type": "text", + "content": ". By repeating this operation several times, we effectively fuse the visual concepts of input images and the purified CLIP's embedding of the target classes. This mechanism fully exploits both the instance-level and class-level information from visual and textual modalities, thus encouraging the generation of perturbations with better semantic patterns and higher transferability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 429, + 483, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 483, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 483, + 525 + ], + "type": "text", + "content": "Cross-Attention based Decoder(CA-Decoder). Given the multi-modal fused features from the previous module, the decoder attempts to translate them into perturbations of the target class. The network backbone is realized based on the decoder used in previous works [49,67]. To fully explore the semantic priors brought by the CLIP model, we enhance the underlying backbone by introducing the cross-attention mechanism, which is proven to be effective for many multimodal learning tasks. Specifically, we incorporate the textual embedding " + }, + { + "bbox": [ + 130, + 429, + 483, + 525 + ], + "type": "inline_equation", + "content": "\\boldsymbol{e}_t" + }, + { + "bbox": [ + 130, + 429, + 483, + 525 + ], + "type": "text", + "content": " from CLIP's latent space into our decoder via the cross-attention layer:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 213, + 555, + 480, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 555, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 213, + 555, + 480, + 597 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} Q = z _ {t} W _ {q}, K = e _ {t} W _ {k}, V = e _ {t} W _ {v}, \\\\ \\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) \\cdot V, \\tag {2} \\\\ \\end{array}", + "image_path": "adc1452718af86d127fca4a6732da10c04876cceb347c8580821f94d8bc70e17.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t\\in \\mathbb{R}^{B\\times d_\\alpha}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " denotes the flattened intermediate features of the decoder, " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "W_{q}\\in \\mathbb{R}^{d_{\\alpha}\\times d}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "W_{k}\\in \\mathbb{R}^{512\\times d}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "W_{v}\\in \\mathbb{R}^{512\\times d}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " are learnable parameters. Similar to C-GSP, we post-process the output " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{o}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " via the " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tanh (\\cdot)" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " smooth projection to obtain the " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " constrained perturbation with budge " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\delta = P(o) = \\epsilon \\cdot \\tanh (o)" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 211, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 211, + 90, + 448, + 102 + ], + "type": "text", + "content": "CLIP-Guided Networks for Transferable Targeted Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 124, + 318, + 190 + ], + "blocks": [ + { + "bbox": [ + 140, + 124, + 318, + 190 + ], + "lines": [ + { + "bbox": [ + 140, + 124, + 318, + 190 + ], + "spans": [ + { + "bbox": [ + 140, + 124, + 318, + 190 + ], + "type": "image", + "image_path": "535006a5d511924f267f7362672a9da371d50584ba9b950433af67444248dd16.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 198, + 235, + 208 + ], + "lines": [ + { + "bbox": [ + 223, + 198, + 235, + 208 + ], + "spans": [ + { + "bbox": [ + 223, + 198, + 235, + 208 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 326, + 115, + 482, + 196 + ], + "blocks": [ + { + "bbox": [ + 326, + 115, + 482, + 196 + ], + "lines": [ + { + "bbox": [ + 326, + 115, + 482, + 196 + ], + "spans": [ + { + "bbox": [ + 326, + 115, + 482, + 196 + ], + "type": "image", + "image_path": "7b805cc33c4da89426f50342466f93d72f015ea4339a44f34f6568682834403e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 198, + 410, + 208 + ], + "lines": [ + { + "bbox": [ + 397, + 198, + 410, + 208 + ], + "spans": [ + { + "bbox": [ + 397, + 198, + 410, + 208 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 217, + 480, + 283 + ], + "lines": [ + { + "bbox": [ + 130, + 217, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 480, + 283 + ], + "type": "text", + "content": "Fig. 3: Illustration of the proposed masked fine-tuning mechanism. (a) We fix the condition input with different text prompts to fine-tune the trained conditional generator " + }, + { + "bbox": [ + 130, + 217, + 480, + 283 + ], + "type": "inline_equation", + "content": "G_{w}" + }, + { + "bbox": [ + 130, + 217, + 480, + 283 + ], + "type": "text", + "content": " into multiple generators for single-target attacks. (b) The fooling rate of several target classes with Inc-v3 and Res-152 as substitute models respectively. The results indicate that direct fine-tuning yields inadequate results for certain classes due to overfitting. We efficiently resolve this issue via a patch-wise random mask operation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 304, + 479, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 479, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 479, + 327 + ], + "type": "text", + "content": "Optimization objective. Based on the proposed CLIP-guided architecture, the optimization objective can be formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 208, + 333, + 480, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 333, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 208, + 333, + 480, + 353 + ], + "type": "interline_equation", + "content": "w ^ {*} \\leftarrow \\arg \\min _ {w} \\mathcal {L} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {s} + G _ {w} (\\boldsymbol {x} _ {s}, \\Phi (\\boldsymbol {t} _ {c}))\\right), c\\right), \\tag {3}", + "image_path": "557199013c15a457202441a2665eeb55daf626ee368ac05ec729012af56b2cd0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 358, + 480, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 480, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 358, + 480, + 418 + ], + "type": "inline_equation", + "content": "t_c" + }, + { + "bbox": [ + 130, + 358, + 480, + 418 + ], + "type": "text", + "content": " represents the corresponding text prompts of the target classes " + }, + { + "bbox": [ + 130, + 358, + 480, + 418 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 358, + 480, + 418 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 358, + 480, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 358, + 480, + 418 + ], + "type": "text", + "content": " denotes the cross-entropy loss. By encouraging the network to craft adversarial samples that can misguide the surrogate model's output toward the desired labels, the generator learns the data distribution of the target class and thus exhibits great generalizability in producing perturbation for any input data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 433, + 324, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 433, + 324, + 445 + ], + "spans": [ + { + "bbox": [ + 131, + 433, + 324, + 445 + ], + "type": "text", + "content": "3.3 Masked Fine-Tuning Mechanism" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 450, + 480, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 480, + 533 + ], + "type": "text", + "content": "In addition to the proposed CLIP-guided generator for multi-target attack scenarios, we also design a single-target variant for further improved performance. As depicted in Fig. 3a, we fix the conditioning input with the text description of a specific target class, and fine-tune the trained multi-target generator by resolving the objective function in Eq. (3) on the unlabeled training data " + }, + { + "bbox": [ + 130, + 450, + 480, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_s" + }, + { + "bbox": [ + 130, + 450, + 480, + 533 + ], + "type": "text", + "content": ". This strategy further enhances the attack effects by fine-tuning the conditional generator to specialize in a specific target class." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 481, + 666 + ], + "type": "text", + "content": "However, we encounter the overfitting problem that leads to limited improvement in success rates or even performance degradation for certain target classes. This problem is partly attributed to the fact that the generated adversarial perturbation sometimes heavily focuses on specific regions of the input image [62]. To alleviate this issue, we adopt a patch-wise random mask operation to postprocess the adversarial perturbation produced by the generator, which brings a notable increase in the targeted fooling rate as shown in Fig. 3b. This technique improves the capability and flexibility of our method, as an adversary can perform fine-tuning to further augment the generator for single-target attacks, achieving a better trade-off between efficiency and performance compared to previous single-target methods (See Section 4.5 for detailed analysis)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "H. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 479, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 479, + 187 + ], + "type": "text", + "content": "To validate the effectiveness of CGNC in boosting transferability, we conduct extensive experiments on various black-box models in a range of scenarios. Our code is available at https://github.com/ffhibnese/CGNC_Targeted_Adversarial_Attacks. Please see the Appendix for more experimental results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 204, + 272, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 204, + 272, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 204, + 272, + 216 + ], + "type": "text", + "content": "4.1 Experimental Settings" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 223, + 479, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 223, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 479, + 281 + ], + "type": "text", + "content": "Datasets. Following [19,67], we train the generator on the ImageNet training set [9] and evaluate the attack performance using ImageNet-NeurIPS (1k) dataset proposed by [50]. We also conduct experiments in a more realistic cross-domain scenario where we train the generator and generate perturbations on MS-COCO [41] or Comics [4] while evaluating on target classifier trained on ImageNet." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 283, + 479, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 479, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 479, + 329 + ], + "type": "text", + "content": "Victim Models. We use various victim models with different architectures. Specifically, the naturally trained models include Inception-v3 (Inc-v3) [57], Inception-v4 (Inc-v4) [55], Inception-ResNet-v2 (Inc-Res-v2) [55], ResNet-152 (Res-152) [29], DenseNet-121 (DN-121) [32], GoogleNet [56], and VGG-16 [54]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 331, + 479, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 479, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 479, + 399 + ], + "type": "text", + "content": "For further evaluation, we also analyze on the robust-trained models, including adv-Inception-v3 (Inc-v3adv) [26], ens-adv-Inception-ResNet-v2 (IR-v2ens) [28], and robust-trained ResNet-50 [24,31], dubbed as Res50SIN (trained on stylized ImageNet), Res50IN (trained on the mixture of stylized and Nature ImageNet), Res50fine (Res50IN plus further finetuning with an auxiliary dataset [24]), and Res50Aug (trained with the advanced data augmentation Augmix [31])." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 403, + 479, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 479, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 479, + 495 + ], + "type": "text", + "content": "Baseline Attacks. We reveal the superiority of the proposed CGNC in enhancing multi-target attacks by comparing our method with multiple competitive baselines, including MIM [11], DIM [65], SIM [40], DIM [12], Logit [72], SU [63], and the state-of-the-art (SOTA) multi-target generative attacks CGSP [67] in Section 4.2. For SU attack [63], we choose to compare with its best version DTMI-Logit-SU. Besides, we also provide a comparison of the proposed single-target variant of CGNC with the existing single-target attack methods, i.e., GAP [51], CD-AP [48], TTP [49], and DGTA-PI [19] in Section 4.5." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 479, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 479, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 479, + 555 + ], + "type": "text", + "content": "Implementation Details. Following [19,67], we employ Inc-v3 and Res-152 as surrogate models and adopt the perturbation budget " + }, + { + "bbox": [ + 130, + 498, + 479, + 555 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 498, + 479, + 555 + ], + "type": "text", + "content": " 16/255. We conduct 10 epochs of generator training using a learning rate of 2e-4. During the masked fine-tuning, we maintain the learning rate and apply a mask ratio of 0.2 to fine-tune the text-conditional generator for an additional 5 epochs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 574, + 360, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 360, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 360, + 586 + ], + "type": "text", + "content": "4.2 Multi-Target Transferability Evaluation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 479, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 479, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 479, + 639 + ], + "type": "text", + "content": "To align the experimental setup to former works [19,67], we first target 8 different classes in [71] and provide the average attack success rates (ASR) of the 8 target categories as evaluation metrics. Attack performance under larger numbers of target categories is presented in the next section." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 642, + 479, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 479, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 479, + 664 + ], + "type": "text", + "content": "Attacks against regularly trained models. We first perform attacks on normal models to evaluate the multi-target attack performance. By observing" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "type": "text", + "content": "CLIP-Guided Networks for Transferable Targeted Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 147, + 476, + 327 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 1: Attack success rates (\\%) for multi-target attacks against regularly trained models on ImageNet NeurIPS validation set. * represents white-box attacks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 147, + 476, + 327 + ], + "lines": [ + { + "bbox": [ + 136, + 147, + 476, + 327 + ], + "spans": [ + { + "bbox": [ + 136, + 147, + 476, + 327 + ], + "type": "table", + "html": "
SourceMethodInc-v3Inc-v4Inc-Res-v2Res-152DN-121GoogleNetVGG-16
Inc-v3MIM99.90*0.801.000.400.200.200.30
TI-MIM98.50*0.500.500.300.200.400.40
SI-MIM99.80*1.502.000.800.700.700.50
DIM95.60*2.700.500.801.100.400.80
TI-DIM96.00*1.101.200.500.500.500.80
SI-DIM90.20*3.804.402.002.201.701.40
Logit99.60*5.606.501.703.000.801.50
SU99.59*5.807.003.353.502.003.94
C-GSP93.40*46.5836.7441.6046.4040.0045.00
CGNC96.03*59.4348.0642.4862.9851.3352.54
Res-152MIM0.500.400.6099.70*0.300.300.20
TI-MIM0.300.300.3096.50*0.300.400.30
SI-MIM1.301.201.6099.50*1.001.400.70
DIM2.302.203.0092.30*0.200.800.70
TI-DIM0.800.701.0090.60*0.600.800.50
SI-DIM4.204.805.4090.50*4.203.602.00
Logit10.1010.7012.8095.70*12.703.709.20
SU12.3611.3116.1695.08*16.136.5514.28
C-GSP37.7033.3320.2893.20*64.2041.7045.90
CGNC53.3951.5334.2495.85*85.6662.2363.36
", + "image_path": "1a76f83a01e4d0b78b347411fdfedb2577e6f926a5b751400d6f45ca9aeacbe6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 351, + 480, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 480, + 447 + ], + "type": "text", + "content": "the results in Table 1, we demonstrate that the proposed CGNC consistently achieves significant improvement compared with previous methods. Specifically, our method achieves an average improvement of " + }, + { + "bbox": [ + 130, + 351, + 480, + 447 + ], + "type": "inline_equation", + "content": "17.88\\%" + }, + { + "bbox": [ + 130, + 351, + 480, + 447 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 351, + 480, + 447 + ], + "type": "inline_equation", + "content": "10.08\\%" + }, + { + "bbox": [ + 130, + 351, + 480, + 447 + ], + "type": "text", + "content": " in ASR over the C-GSP [67] attack regarding Res-152 and Inc-v3 as surrogate models respectively, demonstrating the superiority of leveraging the rich prior knowledge from CLIP's text embedding. We also note that the iterative methods obtain nearly " + }, + { + "bbox": [ + 130, + 351, + 480, + 447 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 351, + 480, + 447 + ], + "type": "text", + "content": " while-box ASR while receiving poor performance on black-box models due to overfitting the classification boundaries of surrogate models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 449, + 481, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 481, + 520 + ], + "type": "text", + "content": "Attacks under defense strategies. For a more thorough analysis and comparison, we then compare these attacks under several widely used defenses. Firstly, we consider attacking six robustly trained networks in Table 2. Generally, our method is still able to deceive the black-box classifiers into predicting the specified classes, significantly outperforming former multi-target attacks, e.g., a " + }, + { + "bbox": [ + 130, + 449, + 481, + 520 + ], + "type": "inline_equation", + "content": "20.87\\%" + }, + { + "bbox": [ + 130, + 449, + 481, + 520 + ], + "type": "text", + "content": " increase of fooling rate from Res-152 to " + }, + { + "bbox": [ + 130, + 449, + 481, + 520 + ], + "type": "inline_equation", + "content": "\\mathrm{Res}50_{\\mathrm{Aug}}" + }, + { + "bbox": [ + 130, + 449, + 481, + 520 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 628 + ], + "type": "text", + "content": "Next, we evaluate the performance of different approaches on models with input preprocessing-based defenses, including a set of image smoothing mechanisms [10] and JPEG compression [14] algorithms. As shown in Table 3, although these defenses eliminate certain valid information in the adversarial samples, the CLIP-empowered CGNC still demonstrates excellent capability in boosting targeted transferability. Particularly when the substitute model is Res-152, CGNC achieves an average fooling rate of " + }, + { + "bbox": [ + 130, + 521, + 481, + 628 + ], + "type": "inline_equation", + "content": "71.18\\%" + }, + { + "bbox": [ + 130, + 521, + 481, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 481, + 628 + ], + "type": "inline_equation", + "content": "80.34\\%" + }, + { + "bbox": [ + 130, + 521, + 481, + 628 + ], + "type": "text", + "content": " on smoothing methods and JPEG compression, while C-GSP only reaches " + }, + { + "bbox": [ + 130, + 521, + 481, + 628 + ], + "type": "inline_equation", + "content": "49.05\\%" + }, + { + "bbox": [ + 130, + 521, + 481, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 481, + 628 + ], + "type": "inline_equation", + "content": "56.16\\%" + }, + { + "bbox": [ + 130, + 521, + 481, + 628 + ], + "type": "text", + "content": " respectively, verifying the stability and robustness of the proposed method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "content": "Perturbation Visualization. We present visualization results in Fig. 4 to unveil the principle of our method. Upon careful inspection, we can see that the generated perturbations mainly focus on the semantic regions of the input" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "H. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 155, + 146, + 455, + 322 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 2: Comparison of the proposed CGNC with the SOTA multi-target attacks against models with robust training mechanism on ImageNet NeurIPS validation set." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 155, + 146, + 455, + 322 + ], + "lines": [ + { + "bbox": [ + 155, + 146, + 455, + 322 + ], + "spans": [ + { + "bbox": [ + 155, + 146, + 455, + 322 + ], + "type": "table", + "html": "
SourceMethodInc-v3advIR-v2ensRes50SINRes50INRes50fineRes50Aug
Inc-v3MIM0.160.100.200.270.440.19
TI-MIM0.210.190.330.490.680.31
SI-MIM0.130.190.260.430.630.29
DIM0.110.090.160.330.390.19
TI-DIM0.150.130.160.210.330.14
SI-DIM0.190.210.430.710.840.46
Logit0.300.300.701.233.140.86
SU0.490.410.841.753.551.04
C-GSP20.4118.046.9633.7644.5621.95
CGNC24.3622.548.8540.8352.1822.85
Res-152MIM0.190.150.281.582.750.78
TI-MIM0.610.730.502.514.751.76
SI-MIM0.240.240.390.660.840.36
DIM0.630.370.948.5014.223.77
TI-DIM0.230.300.280.761.490.49
SI-DIM0.710.710.752.733.891.37
Logit1.151.181.656.7015.465.93
SU2.121.201.957.5321.146.95
C-GSP14.6016.0116.8460.3065.5142.88
CGNC22.2126.7129.8379.8084.0563.75
", + "image_path": "1a4679cee41f507ea71016cc62cea60e88523894bedddc3770a00ac3e8fc6d1e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 147, + 376, + 466, + 443 + ], + "blocks": [ + { + "bbox": [ + 132, + 334, + 479, + 367 + ], + "lines": [ + { + "bbox": [ + 132, + 334, + 479, + 367 + ], + "spans": [ + { + "bbox": [ + 132, + 334, + 479, + 367 + ], + "type": "text", + "content": "Table 3: Comparison of our method with C-GSP under different defense strategies. Q indicates the quality factor in JPEG compression. Here the target model is DN-121 and results for more victim models are in the Appendix." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 147, + 376, + 466, + 443 + ], + "lines": [ + { + "bbox": [ + 147, + 376, + 466, + 443 + ], + "spans": [ + { + "bbox": [ + 147, + 376, + 466, + 443 + ], + "type": "table", + "html": "
SourceMethodSmoothingJPEG compression
GaussianMediumAverageQ=70Q=75Q=80Q=85Q=90
Inc-v3C-GSP36.5946.9139.8649.4150.5551.9552.8453.39
CGNC43.0355.3545.9458.3559.2560.2861.2961.94
Res-152C-GSP43.2156.3847.5553.0354.5456.1357.7559.36
CGNC64.4479.8469.2577.3878.8180.5482.0582.94
", + "image_path": "8c74f5fd0cd428b79376a4555392ab97ee0a2be21d4db40ec4042d61241d3597.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 467, + 480, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 467, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 467, + 480, + 525 + ], + "type": "text", + "content": "images and contain sufficient semantic patterns specific to the target category. For instance, when the condition is a photo of a sea lion, the resulting perturbations indeed carry rich patterns closely resembling this marine animal. We also observe that the pattern changes in accordance with the text prompts, which validates our conditioning mechanism of the CLIP-encoded textual embedding." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 547, + 310, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 547, + 310, + 558 + ], + "spans": [ + { + "bbox": [ + 133, + 547, + 310, + 558 + ], + "type": "text", + "content": "4.3 Evaluation on More Scenarios" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 570, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 570, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 570, + 480, + 665 + ], + "type": "text", + "content": "Cross-Domain scenarios. Next, we explore the more realistic cross-domain scenarios [48, 49] where attackers do not know anything about the data distribution of the training set used by the black-box classifier. Based on this setting, attackers train the generator and generate adversarial perturbations using an auxiliary dataset that follows a different probability distribution from that of the target model. Specifically, we satisfy the cross-domain experimental setting using MS-COCO [41] and Comics [4] datasets respectively. MS-COCO is a large-scale image dataset widely used for object detection and semantic segmentation," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "type": "text", + "content": "CLIP-Guided Networks for Transferable Targeted Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 473, + 257 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 473, + 257 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 473, + 257 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 473, + 257 + ], + "type": "image", + "image_path": "cf93e2fe4dc4048988469f40089162bb690ec6d1950e0b9c3b4e8c9a9315f521.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 264, + 482, + 299 + ], + "lines": [ + { + "bbox": [ + 130, + 264, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 482, + 299 + ], + "type": "text", + "content": "Fig. 4: Visualization results of different input images for different targets. For each text prompt of the target class, the left column shows the perturbation generated by our CGNC while the right column displays the corresponding adversarial examples." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 146, + 310, + 289, + 418 + ], + "blocks": [ + { + "bbox": [ + 146, + 310, + 289, + 418 + ], + "lines": [ + { + "bbox": [ + 146, + 310, + 289, + 418 + ], + "spans": [ + { + "bbox": [ + 146, + 310, + 289, + 418 + ], + "type": "image", + "image_path": "b7330cdcc02e3b51f99848fca2ac8c5c2d3a8d0e7ef9979bd59b5067be897873.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 429, + 303, + 475 + ], + "lines": [ + { + "bbox": [ + 130, + 429, + 303, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 303, + 475 + ], + "type": "text", + "content": "Fig.5: Fooling rates (from Res-152 to VGG-16) in attacking 8 target classes on cross-domain scenarios. We also provide the results on ImageNet as a comparison." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 325, + 310, + 468, + 418 + ], + "blocks": [ + { + "bbox": [ + 325, + 310, + 468, + 418 + ], + "lines": [ + { + "bbox": [ + 325, + 310, + 468, + 418 + ], + "spans": [ + { + "bbox": [ + 325, + 310, + 468, + 418 + ], + "type": "image", + "image_path": "587f7b2c655aa79a529e370bb3b4f35a4ca3652f7575a9eb6ebfa8e5b62aa964.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 429, + 482, + 475 + ], + "lines": [ + { + "bbox": [ + 310, + 429, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 310, + 429, + 482, + 475 + ], + "type": "text", + "content": "Fig.6: Fooling rates of our CGNC with C-GSP on larger numbers of attacked classes regarding Res-152 and Inc-v3 as surrogate models." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 521 + ], + "type": "text", + "content": "while Comic is composed of a large number of comic images that can be regarded as a stylized version of ImageNet." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "type": "text", + "content": "To analyze the cross-domain transferability, we randomly select 1000 images from the source dataset to craft adversarial samples using the trained generator. Fig. 5 shows the results of the cross-domain attacks transferring from three different source datasets to ImageNet. In general, the more challenging cross-domain datasets lead to varying degrees of reduction in the targeted fooling rate due to the domain gap, especially on the Comics dataset which differs greatly from the ImageNet distribution. Benefiting from the text guidance on the target category, CGNC still attains decent attack performance and remarkably outperforms C-GSP. This demonstrates that our method achieves better cross-domain transferability and is partly independent of the training dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "In addition, since the SOTA single-target attacks [19,49,62] require samples of the target class for loss computation, they are not applicable in the cross-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "H. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 146, + 447, + 198 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 4: Ablation study of CGNC and its two variants on ImageNet NeurIPS validation set for 8 target classes. The substitute model is Res-152." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 146, + 447, + 198 + ], + "lines": [ + { + "bbox": [ + 167, + 146, + 447, + 198 + ], + "spans": [ + { + "bbox": [ + 167, + 146, + 447, + 198 + ], + "type": "table", + "html": "
MethodVGG-16GoogleNetInc-v3Res-152DN-201
CGNA-CA-t56.5551.0947.4492.8174.65
CGNA-CA56.6454.2949.7393.3475.99
CGNC63.3662.2353.3995.8582.69
", + "image_path": "b7e086230954a2d9486b04ffa751717093b81f96bb610c451a0e511f4e5ab735.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 217, + 481, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 481, + 239 + ], + "type": "text", + "content": "domain scenarios where the source dataset lacks images of the target category. This represents an additional advantage of our method over single-target attacks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 241, + 481, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 241, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 481, + 300 + ], + "type": "text", + "content": "Larger Numbers of Target Classes. We then increase the number of target categories to verify the effectiveness of our method when handling plenty of classes. Specifically, we condition CGNC with more text inputs corresponding to the increased number of target categories. The performance is evaluated across the aforementioned six black-box models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 301, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 456 + ], + "type": "text", + "content": "As mentioned before, single-target attacks become impractical for real-world classification systems with hundreds or thousands of target categories [27]. In comparison, our method effectively solves this issue and achieves great improvements in large numbers of target classes over C-GSP as in Fig. 6, e.g., " + }, + { + "bbox": [ + 130, + 301, + 482, + 456 + ], + "type": "inline_equation", + "content": "14.71\\%" + }, + { + "bbox": [ + 130, + 301, + 482, + 456 + ], + "type": "text", + "content": " increase for 200 target classes proxy on the Res-152. This again reveals the significance of the fully exploited textual guidance from the CLIP model. Additionally, our method exhibits greater robustness to the varying numbers of target classes, e.g., the performance of our network on Inc-v3 demonstrates a smoother decrease compared to that of C-GSP. This further highlights the superiority of our method in reducing computation costs. For instance, when launching 1000-class targeted attacks using Inc-v3 as the surrogate model, our method requires training only 5 CGNCs with 200 conditions, yet it can achieve comparable performance to 20 generators with 50 conditions trained using C-GSP." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 472, + 238, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 238, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 238, + 483 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 489, + 481, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 489, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 489, + 481, + 548 + ], + "type": "text", + "content": "We conduct ablation experiments on the ImageNet NeurIPS dataset to study the effect of the proposed techniques. Specifically, we introduce two variants of CGNC. CGNC-CA removes the cross-attention module from the original network, and CGNC-CA-t further modifies the conditioning mechanism by replacing the CLIP's text embedding with one-hot labels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 548, + 481, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 481, + 609 + ], + "type": "text", + "content": "From the results in Table 4, we can find that each aforementioned technique can further improve the attack success rates. Moreover, the remarkable improvement from CGNC-CA to CGNC also confirms that these cross-attention modules in the CA-Decoder can make better use of the text guidance provided by the CLIP model to enhance the targeted transferability of crafted perturbations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 624, + 359, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 359, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 359, + 635 + ], + "type": "text", + "content": "4.5 Comparison with Single-Target Attacks" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Next, we compare the single-target variants of the proposed CGNC with multiple state-of-the-art single-target attacks. To obtain our single-target generators, we" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 211, + 91, + 448, + 102 + ], + "type": "text", + "content": "CLIP-Guided Networks for Transferable Targeted Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 158, + 472, + 254 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 5: Comparison of our method with the SOTA single-target attacks. " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "^\\dagger" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " denotes the single-target variant of our CGNC obtained through the masked fine-tuning technique. * represents the white-box attacks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 158, + 472, + 254 + ], + "lines": [ + { + "bbox": [ + 140, + 158, + 472, + 254 + ], + "spans": [ + { + "bbox": [ + 140, + 158, + 472, + 254 + ], + "type": "table", + "html": "
SourceMethodInc-v3Inc-v4Inc-Res-v2Res-152DN-121GoogleNetVGG-16
Inc-v3GAP86.90*45.0634.4834.4841.7426.8934.34
CD-AP94.20*57.6060.1037.1041.6032.3041.70
TTP91.37*46.0439.3716.4033.4725.8025.73
DGTA-PI94.63*67.9555.0350.5047.3847.6748.11
CGNC†98.84*74.7664.4862.0078.9469.0670.74
Res-152GAP30.9931.4320.4884.86*58.3529.8939.70
CD-AP33.3043.7042.7096.60*53.8036.6034.10
TTP62.0349.2038.7095.12*82.9665.0962.82
DGTA-PI66.8353.6247.6196.48*86.6168.2969.58
CGNC†68.8669.4545.7198.61*91.1469.8368.05
", + "image_path": "649d56f21cb00169fdedfb3e1535d8f784f44fc5c86eed0d31884ebaa5671871.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 275, + 481, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 481, + 383 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 481, + 383 + ], + "type": "text", + "content": "conduct the proposed masked fine-tuning to the trained conditional generator eight times using eight different text prompts of the target classes. Quantitative results in Table 5 indicate that our single-target enhanced generators outperform the competing single-target attacks in most cases. Particularly on the surrogate model of Inc-v3, " + }, + { + "bbox": [ + 130, + 275, + 481, + 383 + ], + "type": "inline_equation", + "content": "\\mathrm{CGNC}^{\\dagger}" + }, + { + "bbox": [ + 130, + 275, + 481, + 383 + ], + "type": "text", + "content": " achieves a notable increase of " + }, + { + "bbox": [ + 130, + 275, + 481, + 383 + ], + "type": "inline_equation", + "content": "15.36\\%" + }, + { + "bbox": [ + 130, + 275, + 481, + 383 + ], + "type": "text", + "content": " in the average black-box fooling rate compared to previous methods, demonstrating the significant effectiveness of the proposed masked fine-tuning. The MFT technique further enhances the single-target performance, thereby improving the adaptability and flexibility of our approach in scenarios with fewer target classes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 384, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 479 + ], + "type": "text", + "content": "Note that these single-target methods require training 8 generators from scratch for 8 different target classes. In contrast, our method only needs to train a single multi-target generator and perform fine-tuning 8 times, each time with just a few epochs. Based on the typical experimental setup used in the SOTA single-target methods [19,49], our strategy can diminish over 100 training epochs when targeting 8 classes, thus substantially mitigating the computational burden. More experiments of single-target attacks against adversarially robust models and input preprocessing defenses are shown in the Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 497, + 220, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 497, + 220, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 497, + 220, + 510 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we design a novel generative network CGNC, which improves multi-target transferable adversarial attacks by fully utilizing the rich prior within the CLIP as auxiliary semantic knowledge about the target category. To better incorporate the prior information, we introduce the cross-attention modules and efficiently condition the generator with CLIP's text embeddings. Through extensive experiments, we demonstrate the effectiveness of the proposed CGNC on multiple black-box target models in a variety of scenarios. Moreover, we propose a masked fine-tuning technique to further enhance the CGNC in attacking a single class, which outperforms previous single-target methods in both efficiency and effectiveness. We hope that the proposed method can serve as a reliable tool for evaluating the model robustness under black-box setups, promoting further research on the vulnerability and robustness of DNNs." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "H. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 485, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 485, + 197 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 485, + 197 + ], + "type": "text", + "content": "This work is supported in part by the National Natural Science Foundation of China under grant 62171248, 62301189, Guangdong Basic and Applied Basic Research Foundation under grant 2021A1515110066, the PCNL KEY project (PCL2021A07), and Shenzhen Science and Technology Program under Grant JCYJ20220818101012025, RCBS20221008093124061, GXWD20220811172936001." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 213, + 197, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 213, + 197, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 213, + 197, + 226 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 235, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 235, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 235, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 138, + 235, + 480, + 257 + ], + "type": "text", + "content": "1. Akhtar, N., Mian, A.: Threat of adversarial attacks on deep learning in computer vision: A survey. IEEE Access 6, 14410-14430 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 258, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 258, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 258, + 480, + 289 + ], + "type": "text", + "content": "2. Andriushchenko, M., Croce, F., Flammarion, N., Hein, M.: Square attack: a query-efficient black-box adversarial attack via random search. In: European conference on computer vision. pp. 484-501. Springer (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 289, + 480, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 480, + 321 + ], + "type": "text", + "content": "3. Athalye, A., Engstrom, L., Ilyas, A., Kwok, K.: Synthesizing robust adversarial examples. In: International conference on machine learning. pp. 284-293. PMLR (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 322, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 322, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 322, + 480, + 342 + ], + "type": "text", + "content": "4. BircanoAYlu, C.: https://www.kaggle.com/datasets/cenkbircanoglu/comic-books-classification. Kaggle, 2017" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 343, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 364 + ], + "type": "text", + "content": "5. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: 2017 ieee symposium on security and privacy (sp). pp. 39-57. IEEE (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 365, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 396 + ], + "type": "text", + "content": "6. Che, Z., Borji, A., Zhai, G., Ling, S., Li, J., Min, X., Guo, G., Le Callet, P.: Smgea: A new ensemble adversarial attack powered by long-term gradient memories. IEEE Transactions on Neural Networks and Learning Systems 33(3), 1051-1065 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 397, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 397, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 397, + 480, + 418 + ], + "type": "text", + "content": "7. Chefer, H., Benaim, S., Paiss, R., Wolf, L.: Image-based clip-guided essence transfer. In: European Conference on Computer Vision. pp. 695-711. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 418, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 480, + 460 + ], + "type": "text", + "content": "8. Chen, B., Feng, Y., Dai, T., Bai, J., Jiang, Y., Xia, S.T., Wang, X.: Adversarial examples generation for deep product quantization networks on image retrieval. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(2), 1388-1404 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 461, + 480, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 461, + 480, + 494 + ], + "spans": [ + { + "bbox": [ + 138, + 461, + 480, + 494 + ], + "type": "text", + "content": "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 494, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 494, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 138, + 494, + 480, + 515 + ], + "type": "text", + "content": "0. Ding, G.W., Wang, L., Jin, X.: Advertorch v0.1: An adversarial robustness toolbox based on pytorch. arXiv preprint arXiv:1902.07623 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 515, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 515, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 138, + 515, + 480, + 547 + ], + "type": "text", + "content": "1. Dong, Y., Liao, F., Pang, T., Su, H., Zhu, J., Hu, X., Li, J.: Boosting adversarial attacks with momentum. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9185-9193 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 547, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 480, + 579 + ], + "type": "text", + "content": "2. Dong, Y., Pang, T., Su, H., Zhu, J.: Evading defenses to transferable adversarial examples by translation-invariant attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4312-4321 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 580, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 580, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 580, + 480, + 601 + ], + "type": "text", + "content": "3. Du, J., Zhang, H., Zhou, J.T., Yang, Y., Feng, J.: Query-efficient meta attack to deep neural networks. arXiv preprint arXiv:1906.02398 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 601, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 621 + ], + "type": "text", + "content": "4. Dziugaite, G.K., Ghahramani, Z., Roy, D.M.: A study of the effect of jpg compression on adversarial images. arXiv preprint arXiv:1608.00853 (2016)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "type": "text", + "content": "5. Eykholt, K., Evtimov, I., Fernandes, E., Li, B., Rahmati, A., Xiao, C., Prakash, A., Kohno, T., Song, D.: Robust physical-world attacks on deep learning visual classification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1625-1634 (2018)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 211, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 211, + 91, + 447, + 102 + ], + "type": "text", + "content": "CLIP-Guided Networks for Transferable Targeted Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "16. Fang, H., Chen, B., Wang, X., Wang, Z., Xia, S.T.: Gidf: A generative gradient inversion method with feature domain optimization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4967-4976 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "text", + "content": "17. Fang, H., Kong, J., Yu, W., Chen, B., Li, J., Xia, S., Xu, K.: One perturbation is enough: On generating universal adversarial perturbations against vision-language pre-training models. arXiv preprint arXiv:2406.05491 (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "type": "text", + "content": "18. Fang, H., Qiu, Y., Yu, H., Yu, W., Kong, J., Chong, B., Chen, B., Wang, X., Xia, S.T.: Privacy leakage on dnns: A survey of model inversion attacks and defenses. arXiv preprint arXiv:2402.04013 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 481, + 247 + ], + "type": "text", + "content": "19. Feng, W., Xu, N., Zhang, T., Zhang, Y.: Dynamic generative targeted attacks with pattern injection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16404-16414 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 247, + 481, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 481, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 481, + 268 + ], + "type": "text", + "content": "20. Gao, K., Bai, Y., Bai, J., Yang, Y., Xia, S.T.: Adversarial robustness for visual grounding of multimodal large language models. In: ICLR Workshop (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 268, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 481, + 289 + ], + "type": "text", + "content": "21. Gao, K., Bai, Y., Gu, J., Xia, S.T., Torr, P., Li, Z., Liu, W.: Inducing high energy-latency of large vision-language models with verbose images. In: ICLR (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 289, + 481, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 481, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 481, + 321 + ], + "type": "text", + "content": "22. Gao, K., Gu, J., Bai, Y., Xia, S.T., Torr, P., Liu, W., Li, Z.: Energy-latency manipulation of multi-modal large language models via verbose samples. arXiv preprint arXiv:2404.16557 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 321, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 481, + 342 + ], + "type": "text", + "content": "23. Gao, L., Cheng, Y., Zhang, Q., Xu, X., Song, J.: Feature space targeted attacks by statistic alignment. arXiv preprint arXiv:2105.11645 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "type": "text", + "content": "24. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "text", + "content": "25. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 407, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 429 + ], + "type": "text", + "content": "26. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 429, + 481, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 481, + 471 + ], + "type": "text", + "content": "27. Han, J., Dong, X., Zhang, R., Chen, D., Zhang, W., Yu, N., Luo, P., Wang, X.: Once a man: Towards multi-target attack via learning multi-target adversarial network once. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5158-5167 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 471, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 481, + 493 + ], + "type": "text", + "content": "28. Hang, J., Han, K., Chen, H., Li, Y.: Ensemble adversarial black-box attacks against deep learning systems. Pattern Recognition 101, 107184 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 493, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 481, + 536 + ], + "type": "text", + "content": "29. He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14. pp. 630-645. Springer (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "type": "text", + "content": "30. He, Z., Wang, W., Dong, J., Tan, T.: Revisiting ensemble adversarial attack. Signal Processing: Image Communication 107, 116747 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 557, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 481, + 590 + ], + "type": "text", + "content": "31. Hendrycks, D., Mu, N., Cubuk, E.D., Zoph, B., Gilmer, J., Lakshminarayanan, B.: Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 590, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 481, + 622 + ], + "type": "text", + "content": "32. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "text", + "content": "33. Huang, Z., Zhou, A., Ling, Z., Cai, M., Wang, H., Lee, Y.J.: A sentence speaks a thousand images: Domain generalization through distilling clip with language guidance. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11685-11695 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "H. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "text", + "content": "34. Inkawich, N., Liang, K., Wang, B., Inkawich, M., Carin, L., Chen, Y.: Perturbing across the feature hierarchy to improve standard and strict blackbox attack transferability. Advances in Neural Information Processing Systems 33, 20791-20801 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "text", + "content": "35. Inkawich, N., Wen, W., Li, H.H., Chen, Y.: Feature space perturbations yield more transferable adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7066-7074 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "36. Ju, C., Han, T., Zheng, K., Zhang, Y., Xie, W.: Prompting visual-language models for efficient video understanding. In: European Conference on Computer Vision. pp. 105-124. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "text", + "content": "37. Kong, Z., Guo, J., Li, A., Liu, C.: Physgan: Generating physical-world-resilient adversarial examples for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14254-14263 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 260, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 481, + 281 + ], + "type": "text", + "content": "38. Kurakin, A., Goodfellow, I., Bengio, S.: Adversarial machine learning at scale. arXiv preprint arXiv:1611.01236 (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "type": "text", + "content": "39. Li, Q., Guo, Y., Chen, H.: Yet another intermediate-level attack. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16. pp. 241-257. Springer (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "type": "text", + "content": "40. Lin, J., Song, C., He, K., Wang, L., Hopcroft, J.E.: Nesterov accelerated gradient and scale invariance for adversarial attacks. arXiv preprint arXiv:1908.06281 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 380 + ], + "type": "text", + "content": "41. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 424 + ], + "type": "text", + "content": "42. Liu, A., Wang, J., Liu, X., Cao, B., Zhang, C., Yu, H.: Bias-based universal adversarial patch attack for automatic check-out. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16. pp. 395-410. Springer (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 468 + ], + "type": "text", + "content": "43. Lu, Y., Jia, Y., Wang, J., Li, B., Chai, W., Carin, L., Velipasalar, S.: Enhancing cross-task black-box transferability of adversarial examples with dispersion reduction. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 940-949 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "text", + "content": "44. Luo, J., Bai, T., Zhao, J.: Generating adversarial yet inconspicuous patches with a single image (student abstract). In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 15837-15838 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "text", + "content": "45. Moosavi-Dezfooli, S.M., Fawzi, A., Fawzi, O., Frossard, P.: Universal adversarial perturbations. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1765-1773 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "text", + "content": "46. Moosavi-Dezfooli, S.M., Fawzi, A., Frossard, P.: Deepfool: a simple and accurate method to fool deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2574-2582 (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "text", + "content": "47. Mopuri, K.R., Uppala, P.K., Babu, R.V.: Ask, acquire, and attack: Data-free uap generation using class impressions. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 19-34 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "48. Naseer, M.M., Khan, S.H., Khan, M.H., Shahbaz Khan, F., Porikli, F.: Cross-domain transferability of adversarial perturbations. Advances in Neural Information Processing Systems 32 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "49. Naseer, M., Khan, S., Hayat, M., Khan, F.S., Porikli, F.: On generating transferable targeted perturbations. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7708-7717 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 211, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 211, + 91, + 447, + 102 + ], + "type": "text", + "content": "CLIP-Guided Networks for Transferable Targeted Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "50. NeurIPS: https://www.kaggle.com/c/nips-2017-defense-against-adversarial-attack/data.Kaggle,2017" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 171 + ], + "type": "text", + "content": "51. Poursaeed, O., Katsman, I., Gao, B., Belongie, S.: Generative adversarial perturbations. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4422-4431 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 215 + ], + "type": "text", + "content": "52. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "53. Sarkar, S.K., Oshiba, K., Giebisch, D., Singer, Y.: Robust classification of financial risk. arXiv preprint arXiv:1811.11079 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "text", + "content": "54. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "55. Szegedy, C., Ioffe, S., Vanhoucke, V., Alemi, A.: Inception-v4, inception-resnet and the impact of residual connections on learning. In: Proceedings of the AAAI conference on artificial intelligence. vol. 31 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "type": "text", + "content": "56. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A.: Going deeper with convolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1-9 (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "type": "text", + "content": "57. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2818-2826 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "type": "text", + "content": "58. Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I., Fergus, R.: Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199 (2013)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "text", + "content": "59. Tao, M., Bao, B.K., Tang, H., Xu, C.: Galip: Generative adversarial clips for text-to-image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14214-14223 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "60. Wang, X., He, K.: Enhancing the transferability of adversarial attacks through variance tuning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1924-1933 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "text", + "content": "61. Wang, X., He, K., Hopcroft, J.E.: At-gan: A generative attack model for adversarial transferring on generative adversarial nets. arXiv preprint arXiv:1904.07793 3(4), 3 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "text", + "content": "62. Wang, Z., Yang, H., Feng, Y., Sun, P., Guo, H., Zhang, Z., Ren, K.: Towards transferable targeted adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20534-20543 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "type": "text", + "content": "63. Wei, Z., Chen, J., Wu, Z., Jiang, Y.G.: Enhancing the self-universality for transferable targeted attacks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12281-12290 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 578 + ], + "type": "text", + "content": "64. Xiao, C., Li, B., Zhu, J.Y., He, W., Liu, M., Song, D.: Generating adversarial examples with adversarial networks. arXiv preprint arXiv:1801.02610 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 578, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 621 + ], + "type": "text", + "content": "65. Xie, C., Zhang, Z., Zhou, Y., Bai, S., Wang, J., Ren, Z., Yuille, A.L.: Improving transferability of adversarial examples with input diversity. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2730-2739 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 665 + ], + "type": "text", + "content": "66. Xiong, Y., Lin, J., Zhang, M., Hopcroft, J.E., He, K.: Stochastic variance reduced ensemble adversarial attack for boosting the adversarial transferability. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14983-14992 (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "H. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 336 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "67. Yang, X., Dong, Y., Pang, T., Su, H., Zhu, J.: Boosting transferability of targeted adversarial examples via hierarchical generative networks. In: European Conference on Computer Vision. pp. 725-742. Springer (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "68. Yu, W., Fang, H., Chen, B., Sui, X., Chen, C., Wu, H., Xia, S.T., Xu, K.: Gi-nas: Boosting gradient inversion attacks through adaptive neural architecture search. arXiv preprint arXiv:2405.20725 (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 226 + ], + "type": "text", + "content": "69. Zha, Y., Ji, H., Li, J., Li, R., Dai, T., Chen, B., Wang, Z., Xia, S.T.: Towards compact 3d representations via point feature enhancement masked autoencoders. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 6962-6970 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "type": "text", + "content": "70. Zha, Y., Wang, J., Dai, T., Chen, B., Wang, Z., Xia, S.T.: Instance-aware dynamic prompt tuning for pre-trained point cloud models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14161-14170 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 259, + 482, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 303 + ], + "type": "text", + "content": "71. Zhang, C., Benz, P., Imtiaz, T., Kweon, I.S.: Understanding adversarial examples from the mutual influence of images and perturbations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14521-14530 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 303, + 482, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 336 + ], + "type": "text", + "content": "72. Zhao, Z., Liu, Z., Larson, M.: On success and simplicity: A second look at transferable targeted attacks. Advances in Neural Information Processing Systems 34, 6115-6128 (2021)" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 211, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 211, + 91, + 447, + 102 + ], + "type": "text", + "content": "CLIP-Guided Networks for Transferable Targeted Attacks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CLOSER_ Towards Better Representation Learning for Few-Shot Class-Incremental Learning/10463618-6ae0-4309-91d1-f1ab71be0dbc_content_list.json b/2024/CLOSER_ Towards Better Representation Learning for Few-Shot Class-Incremental Learning/10463618-6ae0-4309-91d1-f1ab71be0dbc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d4de5ab8ea4513390bda624b4f15bcfa84b77a02 --- /dev/null +++ b/2024/CLOSER_ Towards Better Representation Learning for Few-Shot Class-Incremental Learning/10463618-6ae0-4309-91d1-f1ab71be0dbc_content_list.json @@ -0,0 +1,1756 @@ +[ + { + "type": "text", + "text": "CLOSER: Towards Better Representation Learning for Few-Shot Class-Incremental Learning", + "text_level": 1, + "bbox": [ + 215, + 141, + 787, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junghun Oh $^{*1}$ , Sungyong Baik $^{*3}$ , and Kyoung Mu Lee $^{1,2}$", + "bbox": [ + 276, + 210, + 725, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Dept. of ECE&ASRI, $^{2}$ IPAI, Seoul National University", + "bbox": [ + 312, + 238, + 687, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Dept. of Data Science, Hanyang University", + "bbox": [ + 352, + 253, + 648, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{dh6dh,kyoungmu}@snu.ac.kr, dsybaik@hanyang.ac.kr", + "bbox": [ + 312, + 268, + 684, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Aiming to incrementally learn new classes with only few samples while preserving the knowledge of base (old) classes, few-shot class-incremental learning (FSCIL) faces several challenges, such as overfitting and catastrophic forgetting. Such a challenging problem is often tackled by fixing a feature extractor trained on base classes to reduce the adverse effects of overfitting and forgetting. Under such formulation, our primary focus is representation learning on base classes to tackle the unique challenge of FSCIL: simultaneously achieving the transferability and discriminability of the learned representation. Building upon the recent efforts for enhancing the transferability, such as promoting the spread of features, we find that trying to secure the spread of features within a more confined feature space enables the learned representation to strike a better balance between the transferability and discriminability. Thus, in stark contrast to prior beliefs that the inter-class distance should be maximized, we claim that the CLOSER different classes are, the better for FSCIL. The empirical results and analysis from the perspective of information bottleneck theory justify our simple yet seemingly counterintuitive representation learning method, raising research questions and suggesting alternative research directions. The code is available here.", + "bbox": [ + 259, + 318, + 743, + 582 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Few-shot class incremental learning $\\cdot$ Representation learning $\\cdot$ Transferability", + "bbox": [ + 259, + 595, + 740, + 623 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 648, + 375, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Owing to its strong representation power, deep neural networks (DNNs) boast their outstanding performance across various fields. However, such feats require tremendous human effort and time to collect an immense amount of data with accurate annotation. The data hunger of DNNs poses a challenge, especially in dynamic real-world environments, where DNNs are required to learn new concepts with few examples while retaining previously learned concepts. To tackle the challenge, few-shot class-incremental learning (FSCIL) [41] aims to design artificial intelligence systems that can learn new classes with few examples while maintaining performance on previously seen classes.", + "bbox": [ + 212, + 680, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 238, + 825, + 375, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/169eb00ebbd074501d782d682f0fc5b78e4b5ba58ea7c8213894f5de666ea8d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 263, + 150, + 390, + 222 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5106805d7831abf3b8fa0a8e3d8025779bee3b9805bf901aa0526eba4c6aa5d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 263, + 224, + 387, + 299 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/41a84f1362785251bed844ad8dd6fd608e54f174b74fa575b1cfe0ce748458cd.jpg", + "image_caption": [ + "$A_{B} = 98.61\\%$ $A_{N} = 11.49\\%$", + "(a) Baseline" + ], + "image_footnote": [], + "bbox": [ + 263, + 301, + 392, + 375 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/93456906fdc827458cce25de5e95384bfb3deb29fd3b2f7afe55a5acc42bbd1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 148, + 519, + 222 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c54bdd2f5e3287f43e2a5feee594b7c6cad4949889d4c9f4cba3daa3a98f1b32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 224, + 517, + 299 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/34a0548c6ff7eb19059580bb6282bb2382c4b01dadf1a55240610f3b75f77299.jpg", + "image_caption": [ + "$A_{B} = 87.42\\%$ $A_{N} = 33.34\\%$", + "(b) Baseline $^+$ sensation spreading", + "Fig. 1: Visualization of representation trained on MNIST1. (a) Baseline [20, 48] exhibits great base-class discriminability (large inter-class distance) but weak transferability to the new classes (huge overlap between new and base classes leading to misclassification). (b) Baseline + representation spreading [7, 21, 25] benefits the new classes (less overlap with the base classes), while compromising base-class discriminability in the context of FSCIL (dispersed intra-class features leading to less accurate class representation with class prototypes). (c) CLOSER (Ours): Dispersing features in a narrowed feature space enhances both discriminability on the base classes (less deviation between intra-class features and class prototypes) and transferability to the new classes (even less overlap between the base and new classes). For instance, the 4 and 9 classes are not distinguishable in (b) and even less in (a), but CLOSER can yield representation that successfully discriminates them." + ], + "image_footnote": [], + "bbox": [ + 419, + 301, + 517, + 376 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e65ef29366675c9410b0b44779e95090877361cc8d3032f3259bc0231880d136.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 148, + 653, + 223 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fc592dfb548ce2068b224d5307b7346f9237eb9f7a85d071f08a1c4d16830eb3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 224, + 651, + 299 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c72aa6d7210f35c98de28e513eb0e36269705cef9f8371cffdb7160ded88641c.jpg", + "image_caption": [ + "$A_{R} = 92.47\\%$ $A_{N} = 42.14\\%$", + "(c) CLOSER (Ours)", + "$A_{B}$ : accuracy on base classes $A_N$ : accuracy on new classes" + ], + "image_footnote": [], + "bbox": [ + 557, + 301, + 651, + 376 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/85c9814ec26e4ed9d35e51bde9a503604fa9d4368896730f2c204003733edc86.jpg", + "image_caption": [ + "Base Classes" + ], + "image_footnote": [], + "bbox": [ + 686, + 160, + 730, + 277 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6e4546e4bafde37e6611b0a1a7dc42fe9d1333d3134b1ead2a86b0565ef4ad36.jpg", + "image_caption": [ + "New Classes", + "Class prototypes" + ], + "image_footnote": [], + "bbox": [ + 686, + 295, + 730, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve the goal of FSCIL, we need to address catastrophic forgetting (forgetting of previous knowledge while learning new concepts) [13, 23] and overfitting issues (overfitting to few examples and thus poor generalization) [24, 43]. To bypass this convoluted mixture of issues that hinder flexible adaptation of models, most previous works [20, 32, 47, 48, 51, 55] fix the learned representation after training it on base (old) classes and employ a non-parametric classifier, using the feature-average class prototype representation [39]. However, such formulation leads to heavy reliance on the representation acquired through the optimization of softmax cross-entropy (SCE) loss on base classes, which often leads to collapsed intra-class representation [31] and poor transferability to new classes [21], as shown in Fig. 1a. Therefore, in this paper, we mainly focus on exploring effective representation learning methods, aiming to strike a better balance between discriminability on base classes and transferability to new classes.", + "bbox": [ + 212, + 619, + 787, + 816 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Oh*, S. Baik*, and K.M. Lee", + "bbox": [ + 269, + 114, + 483, + 127 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1 As in [27, 45, 50], we use an angular histogram to visualize 2D features of a DNN.", + "bbox": [ + 238, + 825, + 712, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There have been great advances, particularly self-supervised contrastive (SSC) learning [8,18], in the representation learning field to improve the transferability of the learned representation to downstream tasks. Some works [7,21] have attributed the strong transferability of SSC learning to 'spreading out' of intra-class features. Such representation places more emphasis on low- and mid-level features, which can be effectively transferred to and shared by new tasks. Kornblith et al. [25] have also found the relationship between the temperature of SCE loss and the spread of features, suggesting that lower temperature leads to better transferability. As illustrated in Fig. 1b, we observe that the joint optimization of SCE loss with low temperature and SSC loss encourages the spread of features and better transferability to new classes.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite the foregoing, we observe that the previous methods for improving transferability are not enough to find a good representation for FSCIL; in fact, they harm the performance on base classes. Our experimental analysis suggests that excessive feature spread is very detrimental to base classes in the context of FSCIL because it hinders the feature-average class prototype from effectively representing its corresponding class, as demonstrated in Fig. 1b. Hence, it's crucial to develop a representation learning method tailored specifically for the FCSIL problem, particularly addressing the unique challenge of simultaneously achieving discriminability on seen classes and transferability to unseen classes.", + "bbox": [ + 212, + 311, + 787, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, with the support from our experimental findings and information-bottleneck-theory-based analysis, we argue that the inter-class distance greatly affects the trade-off between the discriminability and transferability of the learned representation in the FSCIL problem. We find that the degraded discriminability due to the spread of features can be greatly regulated by reducing inter-class distance. Moreover, we discover that reducing the inter-class variability leads to enhancing the information bottleneck trade-off. Thus, we claim that attempting to ensure the spread of features within a compressed representation space promotes learning minimal yet intrinsic task-related information.", + "bbox": [ + 212, + 449, + 787, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on our analysis, in contrast to common beliefs and practices of previous FSCIL methods [20, 40, 47, 51] that have attempted to increase the inter-class distance, we propose to decrease it. Incorporating SCE loss with lower temperature, SSC loss, and inter-class distance minimization, our new objective enables the learned representation to strike a better balance between the discriminability and transferability, as illustrated in Fig. 1c. With the simple yet seemingly counter-intuitive idea of bringing classes closer (hence the name CLOSER), the proposed method demonstrates outstanding performance, suggesting a new promising research avenue regarding representation learning for FSCIL.", + "bbox": [ + 212, + 585, + 787, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 744, + 397, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Few-Shot Class-Incremental Learning (FSCIL). Towards the development of real-world artificial intelligence systems, Tao et al. [41] have initially introduced FSCIL problem, subsequently fostering numerous studies in the field [1,6,9-11, 36,53]. Most of the works [20,22,28,32,40,47,48,52,54] bypass both catastrophic", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "CLOSER: Representation Learning for Few-Shot Class-Incremental Learning", + "bbox": [ + 218, + 113, + 730, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "forgetting and overfitting issues by fixing the feature extractor trained on base classes and employing a non-parametric classifier using class prototypes [39]. Hence, recent works have focused on representation learning, where the common approach is to encourage greater separation between base classes to reserve the representation space for future new classes [20, 40, 47, 51]. However, we argue that increasing inter-class distance suppresses the acquisition of shared features among classes, which could be relevant to new classes. Thus, contrary to the prevailing belief, we suggest learning representation with an inter-class distance minimization and theoretically and empirically prove its effectiveness in improving the discriminability-transferability trade-off. Similarly, Zou et al. [55] emphasize the importance of learning shareable features among classes, which they propose to achieve with a negative margin [27]. We note that a negative margin is more related to representation spreading rather than our idea of reducing inter-class distance, as discussed in supplementary material Sec. S4.", + "bbox": [ + 212, + 146, + 787, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transferable Representation Learning. The pursuit of representations that can be effectively transferred to downstream tasks has gained significant attention in recent years. The early works have focused on supervised training on ImageNet dataset [35] and the way to transfer the knowledge to other tasks [5,15,34]. Kornblith et al. [25] find that learned representations with better discriminability on a source dataset tend to exhibit degraded transferability to downstream tasks. To obtain a transferable representation, Liu et al. [27,55] suggest incorporating a negative margin in the softmax cross-entropy loss to promote feature sharing among classes rather than solely focusing on discriminative features. In parallel, several methods have reported the strong transferability of representation yielded by self-supervised contrastive (SSC) learning [8,18]. Islam et al. [21] claim that the enhanced transferability acquired by SSC learning can be attributed to the spread of representation, implying that a network learns more fine-grained and shareable knowledge among tasks [7]. From the perspective of information bottleneck theory, Cui et al. [12] claim that over-compression of mutual information between inputs and latent representations can prevent a network from learning features beneficial for downstream tasks. In this paper, we argue that, along with the representation spreading loss, directly regulating inter-class distance encourages a network to learn discriminative yet shareable features among classes, which can be beneficial for both base and new classes. Through the lens of information bottleneck theory, we theoretically analyze the connection between the joint optimization objective and the information bottleneck trade-off, supporting our claim.", + "bbox": [ + 212, + 358, + 787, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Proposed Method", + "text_level": 1, + "bbox": [ + 215, + 714, + 426, + 732 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Background: Problem Formulation", + "text_level": 1, + "bbox": [ + 215, + 747, + 547, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the standard formulation of few-shot class incremental learning (FS-CIL) [41], we assume a sequence of training sessions with the corresponding datasets $\\{\\mathcal{D}^{(0)},\\mathcal{D}^{(1)},\\dots ,\\mathcal{D}^{(T)}\\}$ . $\\mathcal{D}^{(t)}$ consists of training examples $\\pmb{x}_i^{(t)}$ with its class labels $y_{i}^{(t)}\\in \\mathcal{C}^{(t)}$ (for simplicity, we will exclude the superscript), where $\\mathcal{C}^{(t)}$", + "bbox": [ + 212, + 773, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Oh*, S. Baik*, and K.M. Lee", + "bbox": [ + 271, + 114, + 483, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is the set of classes in its respective dataset $\\mathcal{D}^{(t)}$ and $\\mathcal{C}^{(s)}\\cap \\mathcal{C}^{(t)} = \\emptyset$ for $s\\neq t$ (each dataset has its own distinct classes without overlap). In the first session (a.k.a. base session) with its dataset $\\mathcal{D}^{(0)}$ , there is assumed to be a large number of classes available with an abundant amount of training data for each class. In subsequent sessions (a.k.a. incremental sessions) with its datasets $\\mathcal{D}^{(\\geq 1)}$ , it is assumed that each dataset has few training examples for each class. In particular, FSCIL is said to have a $N$ -way $K$ -shot setting when each incremental session has $N$ classes with $K$ examples for each class. At each $t$ -th training session, only its corresponding dataset $\\mathcal{D}^{(t)}$ is accessible for training. After each $t$ -th session, the evaluation is performed on all previously seen classes $\\mathcal{C}^{(\\leq t)}$ using test datasets $\\mathcal{D}_{test}^{(\\leq t)}$ , which consists of unseen new examples with the class label set $\\mathcal{C}^{(\\leq t)}$ .", + "bbox": [ + 212, + 145, + 787, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Background: Baseline", + "text_level": 1, + "bbox": [ + 215, + 333, + 441, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let a classification network consist of a feature extractor $f_{\\theta}(\\cdot)$ and a classification layer with its weights $\\phi$ . The training objective of the base session is simply the softmax cross-entropy (SCE) loss, with the cosine similarity $\\mathrm{sim}(\\cdot, \\cdot)$ between features $z = f_{\\theta}(x)$ and classification weights $\\phi$ as logits [14]:", + "bbox": [ + 212, + 354, + 787, + 417 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c e}} = \\frac {1}{B} \\sum_ {j = 1} ^ {B} - \\log \\frac {\\exp \\left(\\frac {1}{\\tau} \\operatorname {s i m} \\left(z _ {j} , \\phi_ {y _ {j}}\\right)\\right)}{\\sum_ {k = 1} ^ {| \\mathcal {C} ^ {(0)} |} \\exp \\left(\\frac {1}{\\tau} \\operatorname {s i m} \\left(z _ {j} , \\phi_ {k}\\right)\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 425, + 787, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $B$ is the batch size and $\\tau$ is the temperature parameter. Incrementally updating weights with few examples in incremental sessions can make the network vulnerable to both catastrophic forgetting and overfitting. To bypass the problems, several works [20,48] suggest minimizing weight updates by freezing the feature extractor after the base session and using feature-average class prototype representation [39]. Specifically, after the base session, trained $\\phi$ is replaced with class prototypes, a process we refer to as classifier replacement (CR), and new-class prototypes are obtained in the incremental sessions. The $i$ -th class prototype is acquired by averaging the features of its training samples:", + "bbox": [ + 212, + 473, + 787, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {i} ^ {P} = \\frac {1}{N _ {c _ {i}}} \\sum_ {\\left(\\boldsymbol {x} _ {j}, y _ {j}\\right) \\in \\mathcal {D} ^ {\\left(\\geq 0\\right)}} \\mathbb {1} _ {[ y _ {j} = i ]} f _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {j}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 617, + 787, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N_{c_i}$ is the number of training samples for the $i$ -th class and $\\mathbb{1}_{[\\cdot]}$ indicates 1 if the subscript condition is True and 0 otherwise. For an input $\\pmb{x}$ , the unnormalized log probability for the $i$ -th class is computed by $\\mathrm{sim}(f_{\\pmb{\\theta}}(\\pmb{x}), \\phi_i^P)$ .", + "bbox": [ + 212, + 662, + 787, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although this baseline bypasses the forgetting and overfitting issues, it heavily relies on the quality of the representation trained solely on the base classes. Thus, the main focus of this paper is to investigate the important factors that influence representation learning for FSCIL and strategies to improve them.", + "bbox": [ + 212, + 708, + 787, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Transferability, feature spread, and its adverse effects on FSCIL", + "text_level": 1, + "bbox": [ + 212, + 787, + 787, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 1a, the baseline method exhibits a narrow intra-class distribution, widely perceived as representation collapse [31]. Recent studies [7,21,46] have", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "CLOSER: Representation Learning for Few-Shot Class-Incremental Learning", + "bbox": [ + 218, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/eff9e4754620398410349535a5b9c6c336d8cec809f4950832047dcb3fa4346a.jpg", + "image_caption": [ + "Fig. 2: The impact of the spread of representation. Stronger emphasis on self-supervised contrastive loss (larger $\\lambda_{\\mathrm{ssc}}$ ) and low temperature (skyblue) enhances the new-class performance $A_N$ (left), but at the expense of base-class performance $A_B$ (center). The reduced base-class performance is mainly attributed to the excessive intra-class variation, adversely affecting the class prototype representation (right). The experiments are conducted on CIFAR100 dataset." + ], + "image_footnote": [], + "bbox": [ + 248, + 146, + 754, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "revealed that the collapsed representation shows poor transferability due to the loss of shareable low- and mid-level features that new classes can benefit from. As such, they suggest joint optimization with a self-supervised contrastive (SSC) task [8,18], which promotes the spread of features and thus the sharing of features among classes. SSC learning optimizes infoNCE loss [30], regarding an augmented view from a query image as positive and the other images as negative samples. The SSC loss for a positive pair $(i,j)$ is:", + "bbox": [ + 212, + 375, + 789, + 481 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s s c}} ^ {(i, j)} = - \\log \\frac {\\exp \\left(\\frac {1}{\\tau} \\operatorname {s i m} \\left(\\boldsymbol {z} _ {i} , \\boldsymbol {z} _ {j}\\right)\\right)}{\\sum_ {k = 1} ^ {B} \\mathbb {1} _ {[ k \\neq i ]} \\exp \\left(\\frac {1}{\\tau} \\operatorname {s i m} \\left(\\boldsymbol {z} _ {i} , \\boldsymbol {z} _ {k}\\right)\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 489, + 787, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $B$ is the number of samples, including augmented images, and $z_{j}$ is the feature from an augmented view of $\\pmb{x}_{i}$ . The loss is averaged over all positive pairs.", + "bbox": [ + 212, + 536, + 787, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In parallel, several studies have proposed to reduce the temperature and margin parameters in the softmax cross-entropy loss as another way to encourage feature sharing [25,27,55]. Our empirical analysis in Sec. S4 of the supplementary material demonstrates that the temperature parameter affects the transferability more than the margin. Thus, we employ a low-temperature parameter in the SCE loss to further enhance the transferability of learned representations. Indeed, we observe that the combination of $\\mathcal{L}_{\\mathrm{ssc}}$ and $\\mathcal{L}_{\\mathrm{ce}}$ with low $\\tau$ results in better new-class performance in the FSCIL problem due to a larger spread of features, as demonstrated in Fig. 1b and the left figure of Fig. 2.", + "bbox": [ + 212, + 566, + 787, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, as illustrated in the center figure in Fig. 2, we observe that the pursuit of better transferability results in a degradation in discriminability on the base classes. While exploring the dilemma, we discover that the performance decline on the base classes mainly arises from the base class classifier replacement (CR) strategy, which is introduced in Sec. 3.2. The figure on the right in Fig. 2 shows the accuracy on the base classes subsequent to the base session training $(A_B^0)$ , without considering the new classes, both before and after CR, denoted by $A_B^{0,\\mathrm{bef}}$ and $A_B^{0,\\mathrm{aft}}$ , respectively. Before CR, $A_B^{0,\\mathrm{bef}}$ exhibits a relatively high value and a small variance as $\\tau$ and $\\lambda_{\\mathrm{ssc}}$ vary (indicated by the solid lines). Conversely,", + "bbox": [ + 212, + 702, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Oh*, S. Baik*, and K.M. Lee", + "bbox": [ + 269, + 114, + 483, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3d3e3139b96bcdb2dfe4a3a8e18f54cc7c92b4dcd3437e3089064d87b857defc.jpg", + "image_caption": [ + "Fig. 3: Effect of minimizing inter-class distance. As the weight of $\\mathcal{L}_{\\mathrm{inter}}$ , denoted by $\\lambda_{\\mathrm{inter}}$ , increases, the performance on the new classes increases (skyblue) and the performance loss on the base classes induced by CR is greatly alleviated (purple). The experiments are conducted on CUB200 dataset." + ], + "image_footnote": [], + "bbox": [ + 361, + 146, + 643, + 265 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$A_B^{0,\\mathrm{aft}}$ is considerably lower than $A_B^{0,\\mathrm{bef}}$ with a relatively higher variance (indicated by the dotted lines). Considering that CR is introduced to bypass overfitting and catastrophic forgetting in the FSCIL problem, these findings suggest that the previous methods for improving the transferability of representations are not effective in enhancing the trade-off between transferability and discriminability in the context of the FSCIL problem. Thus, it is essential to develop a representation learning method tailored for the FSCIL problem.", + "bbox": [ + 214, + 361, + 787, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Inter-Class Distance Matters", + "text_level": 1, + "bbox": [ + 215, + 489, + 501, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As discussed in Sec. 3.3, learning shareable features through representation spreading proves advantageous for transferability. Yet, it also harms discriminability in the process of CR. Based on the observation in Fig. 1b, we hypothesize that such a dilemma appears to arise from a large inter-class distance since the representation spreading could push features into the extensive inter-class space, which may dilute the information on the base classes. Furthermore, we argue that the large inter-class distance may impede effective feature sharing among classes, undermining the transferability of learned representations. Consequently, to retain the knowledge on base classes while promoting effective feature sharing, we introduce a novel loss function that minimizes the inter-class distance:", + "bbox": [ + 214, + 512, + 787, + 662 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {i n t e r}} = - \\frac {1}{\\sum_ {i = 1} ^ {B} \\sum_ {j > i} ^ {B} \\mathbb {1} _ {[ y _ {i} \\neq y _ {j} ]}} \\sum_ {i = 1} ^ {B} \\sum_ {j > i} ^ {B} \\mathbb {1} _ {[ y _ {i} \\neq y _ {j} ]} \\operatorname {s i m} \\left(\\boldsymbol {z} _ {i}, \\boldsymbol {z} _ {j}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 669, + 787, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 1c displays that the spread of intra-class features is well regulated by applying $\\mathcal{L}_{\\mathrm{inter}}$ and Fig. 3 demonstrates the performance decline from CR is alleviated by minimizing $\\mathcal{L}_{\\mathrm{inter}}$ (indicated by the purple line). Moreover, the results indicated by the skyblue line show that reducing inter-class distance improves the performance on the new classes, corroborating our hypothesis.", + "bbox": [ + 214, + 734, + 787, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our assertion initially seems counter-intuitive, diverging from the common belief of prior works [20,40,47,51] that maximizing inter-class distance may be", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "CLOSER: Representation Learning for Few-Shot Class-Incremental Learning", + "bbox": [ + 218, + 113, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/86efe6a231a80e37d7d61896f9fa4f656b04acdd899e768562c8e6c44e592204.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 406, + 243 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3c02a39d0008eb38d7b3606114271a6386a2c10a75372f41fe38e7687d706822.jpg", + "image_caption": [ + "(b)", + "Fig. 4: (a) Sanity test for $\\mathcal{T}(f_{\\theta})$ : $\\mathcal{T}(f_{\\theta})$ has a positive correlation with the performance on the new classes. Each data point is obtained by different configurations of $\\tau$ and $\\lambda_{\\mathrm{ssc}}$ (without $\\mathcal{L}_{\\mathrm{inter}}$ ). (b), (c) Relationship between inter-class distance, $\\mathcal{T}(f_{\\theta})$ , and $A_N$ : Integrated with the representation spreading, reducing inter-class distance encourages better transferability (red points). However, the tendency is broken when reducing inter-class distance without representation spreading (blue points). Please refer to Sec. 3.5 for theoretical support for these observations. The dots with greater transparency correspond to smaller $\\lambda_{\\mathrm{inter}}$ , ranging from 0 to 1 with intervals of 0.1. We set $\\lambda_{\\mathrm{ssc}}$ as 0.1 when it is used. The experiments are conducted on CIFAR100 dataset." + ], + "image_footnote": [], + "bbox": [ + 411, + 146, + 591, + 243 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4d42ebafd5b7bf19f328c489499f67d10df6b71943c7ce0ca18c358588847db1.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 598, + 146, + 776, + 243 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "beneficial for reserving representation space for future new classes. To further validate the efficacy of reducing inter-class distance with respect to transferability, we propose a measure to quantify how new class samples are distinct from base-class representations. We define the measure as the averaged relative angular distance between a new-class sample and its nearest base-class prototype with respect to the averaged angular distance among all base-class prototype pairs:", + "bbox": [ + 212, + 429, + 787, + 521 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} (f _ {\\boldsymbol {\\theta}}) = \\frac {\\frac {1}{| \\mathcal {D} _ {t e s t} ^ {(> 0)} |} \\sum_ {(\\boldsymbol {x} _ {j} , y _ {j}) \\in \\mathcal {D} _ {t e s t} ^ {(> 0)} i} \\min _ {i} \\angle (\\boldsymbol {z} _ {j} , \\phi_ {b a s e , i} ^ {P})}{\\sum_ {j = 1} ^ {| \\mathcal {C} ^ {(0)} |} \\sum_ {k > j} ^ {| \\mathcal {C} ^ {(0)} |} \\angle \\left(\\phi_ {j} ^ {P} , \\phi_ {k} ^ {P}\\right) / \\left(^ {| \\mathcal {C} ^ {(0)} |} _ {2}\\right)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 534, + 787, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\phi_{base,i}^{P}$ and $\\angle (\\cdot ,\\cdot)$ indicate the $i$ -th base-class prototype and the angular distance between two input vectors, respectively. We introduce the denominator to normalize the varying sizes of the representation space depending on methods. If the representation produced by $f_{\\theta}$ has a distinguishable representation for new classes, $\\mathcal{T}(f_{\\theta})$ would be large. For the sanity test, we check the relationship between $\\mathcal{T}(f_{\\theta})$ and the accuracy of the new class, which is shown in Fig. 4a.", + "bbox": [ + 212, + 595, + 787, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Using this measure, we analyze the relationship between inter-class distance, the spread of features, and the transferability of learned representation. To do so, we train the feature extractor using $\\mathcal{L}_{\\mathrm{ce}}$ with a low temperature, $\\mathcal{L}_{\\mathrm{ssc}}$ , and $\\mathcal{L}_{\\mathrm{inter}}$ with varying loss weights for $\\mathcal{L}_{\\mathrm{ssc}}$ and $\\mathcal{L}_{\\mathrm{inter}}$ , denoted by $\\lambda_{\\mathrm{ssc}}$ and $\\lambda_{\\mathrm{inter}}$ , respectively. Figs. 4b and 4c show the relationship between the inter-class distance, $\\mathcal{T}(f_{\\theta})$ , and the performance on the new classes. The results demonstrate that the joint optimization of $\\mathcal{L}_{\\mathrm{ssc}}$ and $\\mathcal{L}_{\\mathrm{inter}}$ leads to an increase in both $\\mathcal{T}(f_{\\theta})$ and $A_N$ , indicating that smaller inter-class distances lead to more discriminability between base classes and new classes. The analysis corroborates our seemingly counter-intuitive claim that the closer classes are, the better.", + "bbox": [ + 212, + 689, + 789, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Oh*, S. Baik*, and K.M. Lee", + "bbox": [ + 269, + 114, + 483, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In summary, we have observed that the spread of representation achieved by employing a low temperature in the SCE loss and self-supervised contrastive loss is advantageous for learning transferable representation. However, spread representation itself cannot address the trade-off between transferability to new classes and discriminability on base classes in the context of the FSCIL problem. Our analysis demonstrates that decreasing inter-class distance enhances discriminability by regularizing the intra-class spread and improves the transferability by promoting the effective learning of shareable information among classes when combined with the feature-spread-encouraging loss. Consequently, our final loss function is the combination of cross-entropy loss with a lower temperature parameter, self-supervised contrastive loss, and inter-class distance minimizing loss:", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {c e}} + \\lambda_ {\\mathrm {s s c}} \\mathcal {L} _ {\\mathrm {s s c}} + \\lambda_ {\\mathrm {i n t e r}} \\mathcal {L} _ {\\mathrm {i n t e r}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 328, + 787, + 345 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 Information Bottleneck Theory Perspective", + "text_level": 1, + "bbox": [ + 214, + 362, + 620, + 378 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this subsection, we provide theoretical support for the proposed representation learning objective in Eq. (6) from the perspective of the information bottleneck (IB) theory [3, 37, 42]. The IB theory describes the goal of representation learning as finding a good trade-off between complexity and accuracy through finding minimal information from inputs necessary to preserve maximal information about the targets. As discussed by Cui et al. [12], we consider the objective of IB theory to be closely related to learning transferable representations since while achieving the objective, a network could be guided to learn intrinsic knowledge rather than task-irrelevant shortcuts [17], leading to better transferability.", + "bbox": [ + 212, + 385, + 787, + 521 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For an image classification task, let $X \\in \\mathbb{R}^{h \\times w \\times 3}$ , $Y \\in \\mathbb{R}^C$ , and $Z = \\frac{f_{\\theta}(X)}{\\|f_{\\theta}(X)\\|} \\in \\mathbb{R}^d$ denote the input, label, and normalized latent representation variables, respectively, where $h$ and $w$ are the spatial sizes of the image, $C$ is the number of classes, and $d$ is the dimension of the latent representations. The aforementioned trade-off has been formulated as the following objective:", + "bbox": [ + 212, + 521, + 787, + 602 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\max I (Y; Z) - \\beta I (X; Z), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 612, + 785, + 628 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $I(\\cdot ;\\cdot)$ indicates the mutual information between two random variables and $\\beta > 0$ is a Lagrange multiplier. In this work, we consider an alternative trade-off objective, $\\max \\frac{I(Y;Z)}{\\beta I(X;Z)}$ , which is adopted by several works [29,38]. After omitting $\\beta$ for simplicity and modest assumptions, we derive the following lower bound of the IB trade-off objective:", + "bbox": [ + 214, + 635, + 787, + 715 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {I (Y ; Z)}{I (X ; Z)} \\geq 1 - \\frac {d \\cdot \\log (2 \\pi e) + \\frac {1}{C} \\sum_ {i = 1} ^ {C} \\log | \\Sigma_ {W _ {i}} |}{d \\cdot \\log (2 \\pi e) + \\log | \\Sigma_ {T} |}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 723, + 787, + 773 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\Sigma_{W_i}$ and $\\Sigma_T$ indicate the covariance matrices of $Z$ within the $n$ -th class and overall classes, respectively. As proven by Lemma S1 in the supplementary material, both the numerator and denominator in the fractional term of the lower bound in Eq. (8) are negative, leading to the following theorem:", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "CLOSER: Representation Learning for Few-Shot Class-Incremental Learning", + "bbox": [ + 218, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/7aecea879fcc79723767c60fc778a300051c1e84ac46d210996fd2e48f83abba.jpg", + "table_caption": [ + "Table 1: 10-way 5-shot incremental learning results on CUB200." + ], + "table_footnote": [], + "table_body": "
MethodAcc. in each session (%)PD (%) ↓
012345678910
Baseline79.9276.2373.1869.4567.8365.7464.5463.3361.5661.2760.1019.83
TOPIC [41]68.6862.4954.8149.9945.2541.4038.3535.3632.2228.3126.2642.42
F2M [36]81.0778.1675.5772.8970.8668.1767.0165.2663.3661.7660.2620.81
CEC [48]75.8571.9468.5063.5062.4358.2757.7355.8154.8353.5252.2823.57
IDLVQ-C [6]77.3774.7270.2867.1365.3463.5262.1061.5459.0458.6857.8119.56
ALICE [33]77.4072.7070.6067.2065.9063.4062.9061.9060.5060.6060.1017.30
CLOM [55]79.5776.0772.9469.8267.8065.5663.9462.5960.6260.3459.5819.99
Entropy Reg. [28]75.9072.1468.6463.7662.5859.1157.8255.8954.9253.5852.3923.51
LIMIT [52]75.8973.5571.9968.1467.4263.6162.4061.3559.9158.6657.4118.48
MetaFSCIL [11]75.9072.4168.7864.7862.9659.9958.3056.8554.7853.8252.6423.26
FACT [51]75.9073.2370.8466.1365.5662.1561.7459.8358.4157.8956.9418.96
S3C [22]80.6277.5573.1968.5468.0564.3363.5862.0760.6159.7958.9520.83
SAVC [40]81.8577.9274.9570.2169.9667.0266.1665.3063.8463.1562.5019.35
NC-FSCIL [47]80.4575.9872.3070.2868.1765.1664.4363.2560.6660.0159.4421.01
GKEAL [54]78.8875.6272.3268.6267.2364.2662.9861.8960.2059.2158.6720.21
CABD [49]79.1275.3772.8069.0567.5365.1264.0063.5161.8761.4760.9318.19
CLOSER (Ours)79.4075.9273.5070.4769.2467.2266.7365.6964.0064.0263.5815.82
", + "bbox": [ + 246, + 171, + 751, + 371 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/6de8b99bfbae89fbabfabab40a63976e6aec98df30a48a2b6892e6e0a44886f8.jpg", + "table_caption": [ + "Table 2: 5-way 5-shot incremental learning results on CIFAR100." + ], + "table_footnote": [], + "table_body": "
MethodAcc. in each session (%)PD (%) ↓
012345678
Baseline72.9368.4664.2660.1556.5353.6051.5149.1947.0925.84
ERDIL [16]73.6268.2265.1461.8458.3555.5452.5150.1648.2325.39
CEC [48]73.0768.8865.2661.1958.0955.5753.2251.3449.1423.93
CLOM [55]74.2069.8366.1762.3959.2656.4854.3652.1650.2523.95
Entropy Reg. [28]74.470.266.5462.5159.7156.5854.5252.3950.1424.26
FACT [51]74.6072.0967.5663.5261.3858.3656.2854.2452.1022.50
LIMIT [52]73.8172.0967.8763.8960.7057.7755.6753.5251.2322.58
MetaFSCIL [11]74.5070.1066.8462.7759.4856.5254.3652.5649.9724.53
GKEAL [54]74.0170.4567.0163.0860.0157.3055.5053.3951.4022.61
CLOSER (Ours)75.7271.8368.3264.6261.9159.2557.5355.4353.3222.40
", + "bbox": [ + 274, + 417, + 723, + 554 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Theorem 1. The lower bound of $\\frac{I(Y;Z)}{I(X;Z)}$ in Eq. (8) is a monotonically increasing function of $|\\Sigma_{W_i}|$ and a monotonically decreasing function of $|\\Sigma_T|$ .", + "bbox": [ + 212, + 588, + 787, + 625 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The detailed derivations are presented in Sec. S1 of the supplementary material.", + "bbox": [ + 212, + 640, + 787, + 655 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Generally, the determinant of a covariance matrix is correlated with the spread or variability of variables across all dimensions. Thus, Theorem 1 suggests that the IB trade-off can be enhanced by increasing the intra-class variability $(\\mathcal{L}_{\\mathrm{ce}}$ with a low temperature and $\\mathcal{L}_{\\mathrm{ssc}})$ while suppressing the overall representation space $(\\mathcal{L}_{\\mathrm{inter}})$ , supporting the proposed learning objective in Eq. (6) and our claims for inter-class distance minimization. Intuitively, our objective functions encourage greater overlap in representations among different classes, promoting more shareable and compact representations. Our theoretic analysis also underpins the findings in Fig. 4 that reducing inter-class distance without representation spreading loss is observed to result in less transferable representations. Since compressing the overall feature space diminishes the intra-class variability, the IB trade-off may decrease, leading to degraded transferability.", + "bbox": [ + 212, + 657, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Oh*, S. Baik*, and K.M. Lee", + "bbox": [ + 269, + 114, + 482, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/9acab8849aea997e2eea0fa551a2a4c03f1033f7fcbfe61b7c72e88fa7466fd0.jpg", + "table_caption": [ + "Table 3: 5-way 5-shot incremental learning results on miniImageNet." + ], + "table_footnote": [], + "table_body": "
MethodAcc. in each session (%)PD (%) ↓
012345678
Baseline72.2767.4663.2659.7356.5653.5350.9048.9347.2625.01
TOPIC [41]61.3150.0945.1741.1637.4835.5232.1929.4624.4236.89
F2M [36]67.2863.8060.3857.0654.0851.3948.8246.5844.6522.63
CEC [48]72.0066.8362.9759.4356.7053.7351.1949.2447.6324.37
IDLVQ-C [6]64.7759.8755.9352.6249.8847.5544.8343.1441.8422.93
Subspace Reg. [2]80.3773.7668.3664.0760.3656.2753.1050.4547.5532.83
ALICE [33]80.6070.6067.4064.5062.5060.0057.8056.8055.7024.90
CLOM [55]73.0868.0964.1660.4157.4154.2951.5449.3748.0025.08
Entropy Reg. [28]71.8467.1263.2159.7757.0153.9551.5549.5248.2123.63
LIMIT [52]72.3268.4764.3060.7857.9555.0752.7050.7249.1423.13
MetaFSCIL [11]72.0467.9463.7760.2957.5855.1652.9050.7949.1922.85
FACT [51]72.5669.6366.3862.7760.6057.3354.3452.1650.4922.07
GKEAL [54]73.5968.9065.3362.2959.3956.7054.2052.5951.3122.28
CABD [49]74.6570.4366.2962.7760.7557.2454.7953.6552.2222.43
CLOSER (Ours)76.0271.6167.9964.6961.7058.9456.2354.5253.3322.69
", + "bbox": [ + 274, + 171, + 723, + 354 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 382, + 375, + 398 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Experimental Details", + "text_level": 1, + "bbox": [ + 215, + 412, + 437, + 429 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Dataset. Following the benchmark settings proposed by Tao et al. [41], we evaluate the proposed method on CIFAR100 [26], miniImageNet [43], and CUB200 [44]. For CIFAR100 and miniImageNet, the total number of classes is 100: 60 base classes and 40 new classes. The 40 new classes are split into 8 disjoint sets of 5 classes, each of which is sequentially provided with 5 training examples per class (5-way 5-shot) in each incremental session. As for CUB200, there total number of classes is 200, with 100 base classes and 100 new classes. 100 new classes are split into 10 disjoint sets of 10 classes, each of which is sequentially provided with 5 training examples per class (10-way 5-shot) in each incremental session.", + "bbox": [ + 212, + 438, + 787, + 573 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation. Following Zhang et al. [48], we use ResNet-20 [19] for CIFAR100 experiments and ResNet-18 [19] for both miniImageNet and CUB200 experiments. We follow the conventions to use the ResNet-18 model pre-trained on the ImageNet dataset [35] for CUB200. We set the mini-batch size to 128, 128, and 256 for CIFAR100, miniImageNet, and CUB200 experiments, respectively. The temperature parameter $\\tau$ for the baseline method is 1/16 and 'low temperature' indicates $\\tau = 1 / 32$ . We set $\\lambda_{\\mathrm{ssc}}$ as 0.1,0.1, and 0.01 for CIFAR100, miniImageNet, and CUB200, respectively, and $\\lambda_{\\mathrm{inter}}$ as 1, 0.5, and 1.5 for CIFAR100, miniImageNet, and CUB200, respectively. These hyper-parameters are searched via validation using synthesized validation sets. For mutual information estimation, we adopt MINE [4] method and implement it based on the open-source code2. More details are provided in the supplementary material.", + "bbox": [ + 212, + 574, + 787, + 755 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation. We use the accuracy on base $(A_B)$ , new $(A_N)$ , and the whole classes $(A_W)$ as metrics to assess the discriminability, transferability, and the trade-off between them in learned representations. Additionally, we use the performance drop (PD) between the accuracy at the end of the base session (session 0) and", + "bbox": [ + 212, + 755, + 787, + 816 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "CLOSER: Representation Learning for Few-Shot Class-Incremental Learning", + "bbox": [ + 218, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://github.com/gtegner/mine-pytorch", + "bbox": [ + 236, + 824, + 488, + 839 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/7bbd3b239ce2cb6c62e8a35aabe8f1381f3cfb76c96a67ba1d6550181b61cc31.jpg", + "table_caption": [ + "Table 4: Ablation studies on CIFAR100." + ], + "table_footnote": [], + "table_body": "
lowτL SSCL interAB(%)AN(%)AW(%)PD(%)
XXX70.1320.9550.4624.81
X✓X69.9522.8051.0924.59
XX✓69.9520.0549.9926.78
X✓✓71.4322.8351.9925.26
✓XX68.3324.1050.6423.23
✓✓X66.5825.0849.9823.15
✓X✓70.1722.4051.0625.27
✓✓✓70.7227.2353.3222.40
", + "bbox": [ + 372, + 171, + 630, + 262 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "the accuracy at the last incremental session to evaluate the degree to which old knowledge is forgotten and new knowledge is learned simultaneously. All experimental results of CLOSER are obtained by averaging 3 trials.", + "bbox": [ + 212, + 292, + 785, + 339 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Comparison with the Existing Works", + "text_level": 1, + "bbox": [ + 214, + 368, + 568, + 383 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We compare the proposed method, dubbed CLOSER, with prior arts on CUB200 (Tab. 1), CIFAR100 (Tab. 2), and miniImageNet (Tab. 3). We observe that CLOSER achieves state-of-the-art performance on both CUB200 and CIFAR100 datasets, surpassing the results of previous methods by a large margin with respect to $A_W$ and PD. With miniImageNet, the proposed method exhibits substantially higher $A_W$ than the method with the lowest PD and achieves lower PD than the method with the highest $A_W$ , which means that the proposed method achieves a better balance between the performance on base and the new classes. It is worth noting that CLOSER shows such outstanding performance without any assistance from the storage of previous samples (F2M, ERDIL, IDLVQ-C, and CABD), additional computational modules (CEC, CLOM, NC-FSCIL, SAVC, and MetaFSCIL), and test-time data augmentation (S3C and SAVC), suggesting the critical importance of learning effective representations in FSCIL.", + "bbox": [ + 212, + 398, + 787, + 597 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Studies", + "text_level": 1, + "bbox": [ + 214, + 626, + 401, + 641 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To verify the efficacy of the individual components in the proposed method, we perform ablation studies, which are shown in Tab. 4. The increase in $A_N$ when employing a lower temperature in the softmax cross-entropy loss (low $\\tau$ ) or a self-supervised contrastive loss ( $\\mathcal{L}_{\\mathrm{ssc}}$ ) confirms the advantage of feature spread in enhancing the transferability of representations. While transferability can be greatly improved by using low $\\tau$ and $\\mathcal{L}_{\\mathrm{ssc}}$ , a substantial decline in base-class performance $A_B$ is observed, as discussed in Sec. 3.4. The result from the case where all components are utilized demonstrates that this issue can be effectively resolved by minimizing inter-class distance. Furthermore, as discussed in Sec. 3.4, reducing inter-class distance is observed to improve transferability, especially when used with the representation spreading methods. The comprehensive results of ablation studies confirm and support our claims.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Oh*, S. Baik*, and K.M. Lee", + "bbox": [ + 271, + 114, + 482, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7926c9b824bf0fb31175b52982c9974883d8f2b5324cc38cbc6dfae80a1c2077.jpg", + "image_caption": [ + "(a) Base Classes" + ], + "image_footnote": [], + "bbox": [ + 238, + 145, + 411, + 243 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1f37a6e31454da20ead1a27cc944ddff3b9e03901684c02bf77ff54dab72da1b.jpg", + "image_caption": [ + "(b) New Classes" + ], + "image_footnote": [], + "bbox": [ + 413, + 145, + 581, + 244 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/25abf52628bfec146c877a8eca8aefbc7eb396330ef40be9e114c7154530921d.jpg", + "image_caption": [ + "(c) Whole Classes" + ], + "image_footnote": [], + "bbox": [ + 583, + 146, + 763, + 242 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ad9a42eb4ece99d9a9149c53b04dcf819804a07ef5a6f10d12d7bcd5ba109415.jpg", + "image_caption": [ + "Fig. 5: Information bottleneck trade-off analysis. We compare representations acquired by three different methods by assessing the information bottleneck (IB) tradeoff. 'RS' refers to representation spreading methods. We indicate the final models with black edges. The experiments are conducted on the CIFAR100 dataset.", + "(a) Baseline", + "Fig. 6: T-SNE visualization of learned representations. The blue and red points indicate the base and new class samples, respectively. We measure $\\mathcal{T}(f_{\\theta})$ to quantify the transferability of the learned representation. We conducted the experiments on CIFAR100 with reduced classes (20 base classes and 10 new classes) for better visualization. The classification results for each experiment are as follows: (a): $A_B = 78.85\\%$ , $A_N = 13.50\\%$ , $A_W = 57.07\\%$ . (b): $A_B = 77.85\\%$ , $A_N = 25.50\\%$ , $A_W = 60.22\\%$ . (c): $A_B = 78.88\\%$ , $A_N = 28.80\\%$ , $A_W = 62.18\\%$ ." + ], + "image_footnote": [], + "bbox": [ + 243, + 345, + 410, + 443 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/600483079d620d23719251f778d174dca8165cc3fd04aeddbf8108cff1c7b505.jpg", + "image_caption": [ + "(b) Baseline $+\\mathrm{RS}$" + ], + "image_footnote": [], + "bbox": [ + 418, + 345, + 584, + 441 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/71f15c504e72c9c2765c5970978dd103bbb1c3c0b65a78424d590260aa1855eb.jpg", + "image_caption": [ + "(c) CLOSER (Ours)" + ], + "image_footnote": [], + "bbox": [ + 593, + 345, + 759, + 441 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Information Bottleneck trade-off Analysis", + "text_level": 1, + "bbox": [ + 215, + 603, + 606, + 619 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Sec. 3.5, we show the connection between the proposed objective function and the information bottleneck (IB) trade-off. To validate our analysis, we measure the mutual information between representations and inputs $I(X;Z)$ and the mutual information between representations and targets $I(Y;Z)$ for the baseline, representation spread approach, and our method CLOSER, as shown in Fig. 5. In the figure, the more left (lower $I(X;Z)$ ) and upper (higher $I(Y;Z)$ ) regions imply better IB trade-off. Our proposed method CLOSER demonstrates to have found a better IB trade-off, especially when considering whole classes, including base and new classes. The results are encouraging in that CLOSER is able to find a better IB trade-off for all classes, when the feature extractor is only trained on base classes and fixed afterwards. The results demonstrate that CLOSER is effective in tackling a particularly difficult challenge of FSCIL: learning transferable and discriminative features from base classes.", + "bbox": [ + 212, + 643, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "CLOSER: Representation Learning for Few-Shot Class-Incremental Learning", + "bbox": [ + 218, + 114, + 732, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 T-SNE Analysis", + "text_level": 1, + "bbox": [ + 217, + 145, + 393, + 162 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "For qualitative evaluation, we visualize the learned representation trained by different configurations of the proposed losses, which is illustrated in Fig. 6. The value of $\\mathcal{T}(f_{\\theta})$ in the right below in each figure is for measuring the transferability of representation. The baseline representation (a) shows the characteristics of the base class, represented by the features of the new classes mapped to those of the base classes, also indicated by the low value of $\\mathcal{T}(f_{\\theta})$ . Spreading of features (b) largely resolves the overfitting issues, exhibited by larger distances between new classes and base classes; i.e., the increase in $\\mathcal{T}(f_{\\theta})$ . Finally, $\\mathcal{L}_{\\mathrm{inter}}$ is used to compensate for the decreased performance on the base classes due to the spread representation. Reducing inter-class distance also enhances the separability between the new class samples and the clusters of base classes, as evidenced by the further increase in $\\mathcal{T}(f_{\\theta})$ .", + "bbox": [ + 217, + 174, + 787, + 354 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Limitations and Discussions", + "text_level": 1, + "bbox": [ + 217, + 380, + 519, + 398 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our work focuses on learning representation that is discriminative yet transferable to unseen classes to tackle the challenges of FSCIL. As such, our work does not consider an option of updating representation with new classes. Continual update of representation can improve the performance on new classes, however at the cost of sacrificing the old-class performance. Thus, it is a question of stability-plasticity dilemma, where our method focuses more on stability while improving plasticity through discriminative and transferable representation. Furthermore, similar to other works on FSCIL, our method is limited to classification tasks. But, we believe our work can have an impact on other domains, akin to the impact of representation spread methods, such as contrastive learning.", + "bbox": [ + 217, + 415, + 787, + 566 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 217, + 592, + 356, + 609 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "To tackle convoluted challenges in few-shot class-incremental learning (FSCIL), we focus on representation learning, which plays a crucial role. In contrast to previous FSCIL methods that have focused on maximizing the distance between classes, our experimental and theoretic analysis suggests that the closer classes are, the better for FSCIL, especially when feature sharing is encouraged between classes. Upon the analysis, we propose a simple, yet seemingly counter-intuitive idea: bring classes closer for a better transferability-discriminability Pareto front. Our experimental results and information-bottleneck-theory-based analysis suggest that our work can provide a promising research avenue. As such, we hope that our work will inspire future works and discussions in this research direction.", + "bbox": [ + 217, + 627, + 787, + 777 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported in part by the IITP grants [No. 2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University), No.2021-0-02068, and No.2023-0-00156], the NRF grant [No.2021M3A 9E4080782] funded by the Korean government (MSIT).", + "bbox": [ + 217, + 779, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Oh*, S. Baik*, and K.M. Lee", + "bbox": [ + 271, + 114, + 482, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Achituve, I., Navon, A., Yemini, Y., Chechik, G., Fetaya, E.: Gp-tree: A gaussian process classifier for few-shot incremental learning. In: ICML (2021)", + "2. Akyurek, A.F., Akyurek, E., Wijaya, D.T., Andreas, J.: Subspace regularizers for few-shot class incremental learning. In: ICLR (2022)", + "3. Alemi, A.A., Fischer, I., Dillon, J.V., Murphy, K.: Deep variational information bottleneck. In: ICLR (2017)", + "4. Belghazi, M.I., Baratin, A., Rajeshwar, S., Ozair, S., Bengio, Y., Courville, A., Hjelm, D.: Mutual information neural estimation. In: PMLR (2018)", + "5. Chatfield, K., Simonyan, K., Vedaldi, A., Zisserman, A.: Return of the devil in the details: delving deep into convolutional nets. In: BMVC (2014)", + "6. Chen, K., Lee, C.G.: Incremental few-shot learning via vector quantization in deep embedded space. In: ICLR (2021)", + "7. Chen, M.F., Fu, D.Y., Narayan, A., Zhang, M., Song, Z., Fatahalian, K., Ré, C.: Perfectly balanced: improving transfer and robustness of supervised contrastive learning. In: ICML (2022)", + "8. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML (2020)", + "9. Cheraghian, A., Rahman, S., Fang, P., Roy, S.K., Petersson, L., Harandi, M.: Semantic-aware knowledge distillation for few-shot class-incremental learning. In: CVPR (2021)", + "10. Cheraghian, A., Rahman, S., Ramasinghe, S., Fang, P., Simon, C., Petersson, L., Harandi, M.: Synthesized feature based few-shot class-incremental learning on a mixture of subspaces. In: ICCV (2021)", + "1. Chi, Z., Gu, L., Liu, H., Wang, Y., Yu, Y., Tang, J.: Meta桅: a meta-learning approach for few-shot class incremental learning. In: CVPR (2022)", + "2. Cui, Q., Zhao, B., Chen, Z.M., Zhao, B., Song, R., Liang, J., Zhou, B., Yoshie, O.: Discriminability-transferability trade-off: an information-theoretic perspective. In: ECCV (2022)", + "3. Delange, M., Aljundi, R., Masana, M., Parisot, S., Jia, X., Leonardis, A., Slabaugh, G., Tuytelaars, T.: A continual learning survey: Defying forgetting in classification tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2021)", + "4. Deng, J., Guo, J., Yang, J., Xue, N., Kotsia, I., Zafeiriou, S.: Arcface: additive angular margin loss for deep face recognition. In: CVPR (2019)", + "15. Donahue, J., Jia, Y., Vinyals, O., Hoffman, J., Zhang, N., Tzeng, E., Darrellin, T.: Decaf: a deep convolutional activation feature for generic visual recognition. In: ICML (2014)", + "6. Dong, S., Hong, X., Tao, X., Chang, X., Wei, X., Gong, Y.: Few-shot class-incremental learning via relation knowledge distillation. In: AAAI (2021)", + "7. Geirhos, R., Jacobsen, J.H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., Wichmann, F.A.: Shortcut learning in deep neural networks. Nature Machine Intelligence 2(11), 665-673 (2020)", + "8. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR (2020)", + "9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition In: CVPR (2016)", + "20. Hersche, M., Karunaratne, G., Cherubini, G., Benini, L., Sebastian, A., Rahimi, A.: Constrained few-shot class-incremental learning. In: CVPR (2022)" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "CLOSER: Representation Learning for Few-Shot Class-Incremental Learning", + "bbox": [ + 218, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "21. Islam, A., Chen, C.F., Panda, R., Karlinsky, L., Radke, R., Feris, R.: A broad study on the transferability of visual representations with contrastive learning. In: ICCV (2021)", + "22. Kalla, J., Biswas, S.: S3c: self-supervised stochastic classifiers for few-shot class-incremental learning. In: ECCV (2022)", + "23. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., Hassabis, D., Clopath, C., Kumaran, D., Hadsell, R.: Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences of the United States of America (PNAS) (2017)", + "24. Koch, G., Zemel, R., Salakhutdinov, R.: Siamese neural networks for one-shot image recognition. In: ICMLW (2015)", + "25. Kornblith, S., Chen, T., Lee, H., Norouzi, M.: Why do better loss functions lead to less transferable features? In: NeurIPS (2021)", + "26. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images. University of Toronto (2009)", + "27. Liu, B., Cao, Y., Lin, Y., Li, Q., Zhang, Z., Long, M., Hu, H.: Negative margin matters: Understanding margin in few-shot classification. In: ECCV (2020)", + "28. Liu, H., Gu, L., Chi, Z., Wang, Y., Yu, Y., Chen, J., Tang, J.: Few-shot class-incremental learning via entropy-regularized data-free replay. In: ECCV (2022)", + "29. Ngampruetikorn, V., Schwab, D.J.: Information bottleneck theory of high-dimensional regression: relevancy, efficiency and optimality. In: NeurIPS (2022)", + "30. van den Oord, A., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)", + "31. Papyan, V., Han, X., Donoho, D.L.: Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences of the United States of America (PNAS) (2020)", + "32. Peng, C., Zhao, K., Wang, T., Li, M., Lovell, B.C.: Few-shot class-incremental learning from an open-set perspective. In: ECCV (2022)", + "33. Peng, C., Zhao, K., Wang, T., Li, M., Lovell, B.C.: Few-shot class-incremental learning from an open-set perspective. In: ECCV (2022)", + "34. Razavian, A.S., Azizpour, H., Sullivan, J., Carlsson, S.: Cnn features off-the-shelf: an astounding baseline for recognition. In: CVPRW (2014)", + "35. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV) (2015)", + "36. Shi, G., Chen, J., Zhang, W., Zhan, L.M., Wu, X.M.: Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. In: NeurIPS (2021)", + "37. Shwartz-Ziv, R., Tishby, N.: Opening the black box of deep neural networks via information (2017)", + "38. Slonim, N., Tishby, N.: Agglomerative information bottleneck. In: NIPS (1999)", + "39. Snell, J., Swersky, K., Zemel, R.: Prototypical networks for few-shot learning. In: NIPS (2017)", + "40. Song, Z., Zhao, Y., Shi, Y., Peng, P., Yuan, L., Tian, Y.: Learning with fantasy: semantic-aware virtual contrastive constraint for few-shot class-incremental learning. In: CVPR (2023)", + "41. Tao, X., Hong, X., Chang, X., Dong, S., Wei, X., Gong, Y.: Few-shot class-incremental learning. In: CVPR (2020)" + ], + "bbox": [ + 215, + 146, + 787, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Oh*, S. Baik*, and K.M. Lee", + "bbox": [ + 271, + 114, + 483, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "42. Tishby, N., Zaslavsky, N.: Deep learning and the information bottleneck principle. In: 2015 IEEE Information Theory Workshop (ITW). pp. 1-5 (2015). https://doi.org/10.1109/ITW.2015.7133169", + "43. Vinyals, O., Blundell, C., Lillicrap, T., Kavukcuoglu, K., Wierstra, D.: Matching networks for one shot learning. In: NIPS (2016)", + "44. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The Caltech-UCSD Birds-200-2011 Dataset. Tech. rep., Caltech (2011)", + "45. Wang, H., Wang, Y., Zhou, Z., Ji, X., Gong, D., Zhou, J., Li, Z., Liu, W.: Cosface: large margin cosine loss for deep face recognition. In: CVPR (2018)", + "46. Xue, Y., Joshi, S., Gan, E., Chen, P.Y., Mirzasoleiman, B.: Which features are learnt by contrastive learning? on the role of simplicity bias in class collapse and feature suppression. In: ICML (2023)", + "47. Yang, Y., Yuan, H., Li, X., Lin, Z., Torr, P., Tao, D.: Neural collapse inspired feature-classifier alignment for few-shot class incremental learning. In: ICLR (2023)", + "48. Zhang, C., Song, N., Lin, G., Zheng, Y., Pan, P., Xu, Y.: Few-shot incremental learning with continually evolved classifiers. In: CVPR (2021)", + "49. Zhao, L., Lu, J., Xu, Y., Cheng, Z., Guo, D., Niu, Y., Fang, X.: Few-shot class-incremental learning via class-aware bilateral distillation. In: CVPR (2023)", + "50. Zheng, Y., Pal, D.K., Savvides, M.: Ring loss: convex feature normalization for face recognition. In: CVPR (2018)", + "51. Zhou, D.W., Wang, F.Y., Ye, H.J., Ma, L., Pu, S., Zhan, D.C.: Forward compatible few-shot class-incremental learning. In: CVPR (2022)", + "52. Zhou, D.W., Ye, H.J., Ma, L., Xie, D., Pu, S., Zhan, D.C.: Few-shot class-incremental learning by sampling multi-phase tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2022)", + "53. Zhu, K., Cao, Y., Zhai, W., Cheng, J., Zha, Z.J.: Self-promoted prototype refinement for few-shot class-incremental learning. In: CVPR (2021)", + "54. Zhuang, H., Weng, Z., He, R., Lin, Z., Zeng, Z.: Gkeal: gaussian kernel embedded analytic learning for few-shot class incremental task. In: CVPR (2023)", + "55. Zou, Y., Zhang, S., Li, Y., Li, R.: Margin-based few-shot class-incremental learning with class-level overfitting mitigation. In: NeurIPS (2022)" + ], + "bbox": [ + 215, + 146, + 787, + 575 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "CLOSER: Representation Learning for Few-Shot Class-Incremental Learning", + "bbox": [ + 220, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/CLOSER_ Towards Better Representation Learning for Few-Shot Class-Incremental Learning/10463618-6ae0-4309-91d1-f1ab71be0dbc_model.json b/2024/CLOSER_ Towards Better Representation Learning for Few-Shot Class-Incremental Learning/10463618-6ae0-4309-91d1-f1ab71be0dbc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..263e6096f7805cd7ea29762b3d3d0bf860052a7c --- /dev/null +++ b/2024/CLOSER_ Towards Better Representation Learning for Few-Shot Class-Incremental Learning/10463618-6ae0-4309-91d1-f1ab71be0dbc_model.json @@ -0,0 +1,2522 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.217, + 0.142, + 0.788, + 0.187 + ], + "angle": 0, + "content": "CLOSER: Towards Better Representation Learning for Few-Shot Class-Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.212, + 0.726, + 0.229 + ], + "angle": 0, + "content": "Junghun Oh\\(^{*1}\\), Sungyong Baik\\(^{*3}\\), and Kyoung Mu Lee\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.313, + 0.239, + 0.688, + 0.254 + ], + "angle": 0, + "content": "\\(^{1}\\)Dept. of ECE&ASRI, \\(^{2}\\)IPAI, Seoul National University" + }, + { + "type": "text", + "bbox": [ + 0.354, + 0.255, + 0.649, + 0.269 + ], + "angle": 0, + "content": "\\(^{3}\\)Dept. of Data Science, Hanyang University" + }, + { + "type": "text", + "bbox": [ + 0.313, + 0.27, + 0.686, + 0.282 + ], + "angle": 0, + "content": "{dh6dh,kyoungmu}@snu.ac.kr, dsybaik@hanyang.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.319, + 0.744, + 0.583 + ], + "angle": 0, + "content": "Abstract. Aiming to incrementally learn new classes with only few samples while preserving the knowledge of base (old) classes, few-shot class-incremental learning (FSCIL) faces several challenges, such as overfitting and catastrophic forgetting. Such a challenging problem is often tackled by fixing a feature extractor trained on base classes to reduce the adverse effects of overfitting and forgetting. Under such formulation, our primary focus is representation learning on base classes to tackle the unique challenge of FSCIL: simultaneously achieving the transferability and discriminability of the learned representation. Building upon the recent efforts for enhancing the transferability, such as promoting the spread of features, we find that trying to secure the spread of features within a more confined feature space enables the learned representation to strike a better balance between the transferability and discriminability. Thus, in stark contrast to prior beliefs that the inter-class distance should be maximized, we claim that the CLOSER different classes are, the better for FSCIL. The empirical results and analysis from the perspective of information bottleneck theory justify our simple yet seemingly counterintuitive representation learning method, raising research questions and suggesting alternative research directions. The code is available here." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.596, + 0.741, + 0.624 + ], + "angle": 0, + "content": "Keywords: Few-shot class incremental learning \\(\\cdot\\) Representation learning \\(\\cdot\\) Transferability" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.65, + 0.377, + 0.667 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.681, + 0.788, + 0.818 + ], + "angle": 0, + "content": "Owing to its strong representation power, deep neural networks (DNNs) boast their outstanding performance across various fields. However, such feats require tremendous human effort and time to collect an immense amount of data with accurate annotation. The data hunger of DNNs poses a challenge, especially in dynamic real-world environments, where DNNs are required to learn new concepts with few examples while retaining previously learned concepts. To tackle the challenge, few-shot class-incremental learning (FSCIL) [41] aims to design artificial intelligence systems that can learn new classes with few examples while maintaining performance on previously seen classes." + }, + { + "type": "page_footnote", + "bbox": [ + 0.239, + 0.826, + 0.377, + 0.841 + ], + "angle": 0, + "content": "*Equal contribution" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.484, + 0.128 + ], + "angle": 0, + "content": "J. Oh*, S. Baik*, and K.M. Lee" + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.151, + 0.391, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.225, + 0.388, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.303, + 0.393, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.278, + 0.381, + 0.394, + 0.39 + ], + "angle": 0, + "content": "\\(A_{B} = 98.61\\%\\) \\(A_{N} = 11.49\\%\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.313, + 0.399, + 0.363, + 0.408 + ], + "angle": 0, + "content": "(a) Baseline" + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.149, + 0.521, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.225, + 0.519, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.303, + 0.519, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.381, + 0.532, + 0.39 + ], + "angle": 0, + "content": "\\(A_{B} = 87.42\\%\\) \\(A_{N} = 33.34\\%\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.445, + 0.392, + 0.522, + 0.41 + ], + "angle": 0, + "content": "(b) Baseline \\(^+\\) sensation spreading" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.15, + 0.655, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.226, + 0.653, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.303, + 0.653, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.381, + 0.671, + 0.39 + ], + "angle": 0, + "content": "\\(A_{R} = 92.47\\%\\) \\(A_{N} = 42.14\\%\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.398, + 0.645, + 0.407 + ], + "angle": 0, + "content": "(c) CLOSER (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.149, + 0.736, + 0.157 + ], + "angle": 0, + "content": "Base Classes" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.161, + 0.732, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.684, + 0.284, + 0.733, + 0.292 + ], + "angle": 0, + "content": "New Classes" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.296, + 0.732, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.356, + 0.74, + 0.372 + ], + "angle": 0, + "content": "Class prototypes" + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.376, + 0.739, + 0.408 + ], + "angle": 0, + "content": "\\(A_{B}\\) : accuracy on base classes \\(A_N\\) : accuracy on new classes" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.591 + ], + "angle": 0, + "content": "Fig. 1: Visualization of representation trained on MNIST1. (a) Baseline [20, 48] exhibits great base-class discriminability (large inter-class distance) but weak transferability to the new classes (huge overlap between new and base classes leading to misclassification). (b) Baseline + representation spreading [7, 21, 25] benefits the new classes (less overlap with the base classes), while compromising base-class discriminability in the context of FSCIL (dispersed intra-class features leading to less accurate class representation with class prototypes). (c) CLOSER (Ours): Dispersing features in a narrowed feature space enhances both discriminability on the base classes (less deviation between intra-class features and class prototypes) and transferability to the new classes (even less overlap between the base and new classes). For instance, the 4 and 9 classes are not distinguishable in (b) and even less in (a), but CLOSER can yield representation that successfully discriminates them." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.62, + 0.789, + 0.817 + ], + "angle": 0, + "content": "To achieve the goal of FSCIL, we need to address catastrophic forgetting (forgetting of previous knowledge while learning new concepts) [13, 23] and overfitting issues (overfitting to few examples and thus poor generalization) [24, 43]. To bypass this convoluted mixture of issues that hinder flexible adaptation of models, most previous works [20, 32, 47, 48, 51, 55] fix the learned representation after training it on base (old) classes and employ a non-parametric classifier, using the feature-average class prototype representation [39]. However, such formulation leads to heavy reliance on the representation acquired through the optimization of softmax cross-entropy (SCE) loss on base classes, which often leads to collapsed intra-class representation [31] and poor transferability to new classes [21], as shown in Fig. 1a. Therefore, in this paper, we mainly focus on exploring effective representation learning methods, aiming to strike a better balance between discriminability on base classes and transferability to new classes." + }, + { + "type": "page_footnote", + "bbox": [ + 0.239, + 0.826, + 0.714, + 0.84 + ], + "angle": 0, + "content": "1 As in [27, 45, 50], we use an angular histogram to visualize 2D features of a DNN." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "CLOSER: Representation Learning for Few-Shot Class-Incremental Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.312 + ], + "angle": 0, + "content": "There have been great advances, particularly self-supervised contrastive (SSC) learning [8,18], in the representation learning field to improve the transferability of the learned representation to downstream tasks. Some works [7,21] have attributed the strong transferability of SSC learning to 'spreading out' of intra-class features. Such representation places more emphasis on low- and mid-level features, which can be effectively transferred to and shared by new tasks. Kornblith et al. [25] have also found the relationship between the temperature of SCE loss and the spread of features, suggesting that lower temperature leads to better transferability. As illustrated in Fig. 1b, we observe that the joint optimization of SCE loss with low temperature and SSC loss encourages the spread of features and better transferability to new classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.449 + ], + "angle": 0, + "content": "Despite the foregoing, we observe that the previous methods for improving transferability are not enough to find a good representation for FSCIL; in fact, they harm the performance on base classes. Our experimental analysis suggests that excessive feature spread is very detrimental to base classes in the context of FSCIL because it hinders the feature-average class prototype from effectively representing its corresponding class, as demonstrated in Fig. 1b. Hence, it's crucial to develop a representation learning method tailored specifically for the FCSIL problem, particularly addressing the unique challenge of simultaneously achieving discriminability on seen classes and transferability to unseen classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.585 + ], + "angle": 0, + "content": "In this work, with the support from our experimental findings and information-bottleneck-theory-based analysis, we argue that the inter-class distance greatly affects the trade-off between the discriminability and transferability of the learned representation in the FSCIL problem. We find that the degraded discriminability due to the spread of features can be greatly regulated by reducing inter-class distance. Moreover, we discover that reducing the inter-class variability leads to enhancing the information bottleneck trade-off. Thus, we claim that attempting to ensure the spread of features within a compressed representation space promotes learning minimal yet intrinsic task-related information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.788, + 0.722 + ], + "angle": 0, + "content": "Based on our analysis, in contrast to common beliefs and practices of previous FSCIL methods [20, 40, 47, 51] that have attempted to increase the inter-class distance, we propose to decrease it. Incorporating SCE loss with lower temperature, SSC loss, and inter-class distance minimization, our new objective enables the learned representation to strike a better balance between the discriminability and transferability, as illustrated in Fig. 1c. With the simple yet seemingly counter-intuitive idea of bringing classes closer (hence the name CLOSER), the proposed method demonstrates outstanding performance, suggesting a new promising research avenue regarding representation learning for FSCIL." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.746, + 0.398, + 0.763 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Few-Shot Class-Incremental Learning (FSCIL). Towards the development of real-world artificial intelligence systems, Tao et al. [41] have initially introduced FSCIL problem, subsequently fostering numerous studies in the field [1,6,9-11, 36,53]. Most of the works [20,22,28,32,40,47,48,52,54] bypass both catastrophic" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.484, + 0.129 + ], + "angle": 0, + "content": "J. Oh*, S. Baik*, and K.M. Lee" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.357 + ], + "angle": 0, + "content": "forgetting and overfitting issues by fixing the feature extractor trained on base classes and employing a non-parametric classifier using class prototypes [39]. Hence, recent works have focused on representation learning, where the common approach is to encourage greater separation between base classes to reserve the representation space for future new classes [20, 40, 47, 51]. However, we argue that increasing inter-class distance suppresses the acquisition of shared features among classes, which could be relevant to new classes. Thus, contrary to the prevailing belief, we suggest learning representation with an inter-class distance minimization and theoretically and empirically prove its effectiveness in improving the discriminability-transferability trade-off. Similarly, Zou et al. [55] emphasize the importance of learning shareable features among classes, which they propose to achieve with a negative margin [27]. We note that a negative margin is more related to representation spreading rather than our idea of reducing inter-class distance, as discussed in supplementary material Sec. S4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.789, + 0.691 + ], + "angle": 0, + "content": "Transferable Representation Learning. The pursuit of representations that can be effectively transferred to downstream tasks has gained significant attention in recent years. The early works have focused on supervised training on ImageNet dataset [35] and the way to transfer the knowledge to other tasks [5,15,34]. Kornblith et al. [25] find that learned representations with better discriminability on a source dataset tend to exhibit degraded transferability to downstream tasks. To obtain a transferable representation, Liu et al. [27,55] suggest incorporating a negative margin in the softmax cross-entropy loss to promote feature sharing among classes rather than solely focusing on discriminative features. In parallel, several methods have reported the strong transferability of representation yielded by self-supervised contrastive (SSC) learning [8,18]. Islam et al. [21] claim that the enhanced transferability acquired by SSC learning can be attributed to the spread of representation, implying that a network learns more fine-grained and shareable knowledge among tasks [7]. From the perspective of information bottleneck theory, Cui et al. [12] claim that over-compression of mutual information between inputs and latent representations can prevent a network from learning features beneficial for downstream tasks. In this paper, we argue that, along with the representation spreading loss, directly regulating inter-class distance encourages a network to learn discriminative yet shareable features among classes, which can be beneficial for both base and new classes. Through the lens of information bottleneck theory, we theoretically analyze the connection between the joint optimization objective and the information bottleneck trade-off, supporting our claim." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.715, + 0.427, + 0.733 + ], + "angle": 0, + "content": "3 Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.748, + 0.549, + 0.765 + ], + "angle": 0, + "content": "3.1 Background: Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.774, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Following the standard formulation of few-shot class incremental learning (FS-CIL) [41], we assume a sequence of training sessions with the corresponding datasets \\(\\{\\mathcal{D}^{(0)},\\mathcal{D}^{(1)},\\dots ,\\mathcal{D}^{(T)}\\}\\). \\(\\mathcal{D}^{(t)}\\) consists of training examples \\(\\pmb{x}_i^{(t)}\\) with its class labels \\(y_{i}^{(t)}\\in \\mathcal{C}^{(t)}\\) (for simplicity, we will exclude the superscript), where \\(\\mathcal{C}^{(t)}\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "CLOSER: Representation Learning for Few-Shot Class-Incremental Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.317 + ], + "angle": 0, + "content": "is the set of classes in its respective dataset \\(\\mathcal{D}^{(t)}\\) and \\(\\mathcal{C}^{(s)}\\cap \\mathcal{C}^{(t)} = \\emptyset\\) for \\(s\\neq t\\) (each dataset has its own distinct classes without overlap). In the first session (a.k.a. base session) with its dataset \\(\\mathcal{D}^{(0)}\\), there is assumed to be a large number of classes available with an abundant amount of training data for each class. In subsequent sessions (a.k.a. incremental sessions) with its datasets \\(\\mathcal{D}^{(\\geq 1)}\\), it is assumed that each dataset has few training examples for each class. In particular, FSCIL is said to have a \\(N\\)-way \\(K\\)-shot setting when each incremental session has \\(N\\) classes with \\(K\\) examples for each class. At each \\(t\\)-th training session, only its corresponding dataset \\(\\mathcal{D}^{(t)}\\) is accessible for training. After each \\(t\\)-th session, the evaluation is performed on all previously seen classes \\(\\mathcal{C}^{(\\leq t)}\\) using test datasets \\(\\mathcal{D}_{test}^{(\\leq t)}\\), which consists of unseen new examples with the class label set \\(\\mathcal{C}^{(\\leq t)}\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.334, + 0.442, + 0.349 + ], + "angle": 0, + "content": "3.2 Background: Baseline" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.418 + ], + "angle": 0, + "content": "Let a classification network consist of a feature extractor \\( f_{\\theta}(\\cdot) \\) and a classification layer with its weights \\( \\phi \\). The training objective of the base session is simply the softmax cross-entropy (SCE) loss, with the cosine similarity \\( \\mathrm{sim}(\\cdot, \\cdot) \\) between features \\( z = f_{\\theta}(x) \\) and classification weights \\( \\phi \\) as logits [14]:" + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.426, + 0.788, + 0.468 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c e}} = \\frac {1}{B} \\sum_ {j = 1} ^ {B} - \\log \\frac {\\exp \\left(\\frac {1}{\\tau} \\operatorname {s i m} \\left(z _ {j} , \\phi_ {y _ {j}}\\right)\\right)}{\\sum_ {k = 1} ^ {| \\mathcal {C} ^ {(0)} |} \\exp \\left(\\frac {1}{\\tau} \\operatorname {s i m} \\left(z _ {j} , \\phi_ {k}\\right)\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.788, + 0.611 + ], + "angle": 0, + "content": "where \\( B \\) is the batch size and \\( \\tau \\) is the temperature parameter. Incrementally updating weights with few examples in incremental sessions can make the network vulnerable to both catastrophic forgetting and overfitting. To bypass the problems, several works [20,48] suggest minimizing weight updates by freezing the feature extractor after the base session and using feature-average class prototype representation [39]. Specifically, after the base session, trained \\( \\phi \\) is replaced with class prototypes, a process we refer to as classifier replacement (CR), and new-class prototypes are obtained in the incremental sessions. The \\( i \\)-th class prototype is acquired by averaging the features of its training samples:" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.618, + 0.788, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\phi_ {i} ^ {P} = \\frac {1}{N _ {c _ {i}}} \\sum_ {\\left(\\boldsymbol {x} _ {j}, y _ {j}\\right) \\in \\mathcal {D} ^ {\\left(\\geq 0\\right)}} \\mathbb {1} _ {[ y _ {j} = i ]} f _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {j}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.663, + 0.788, + 0.709 + ], + "angle": 0, + "content": "where \\(N_{c_i}\\) is the number of training samples for the \\(i\\)-th class and \\(\\mathbb{1}_{[\\cdot]}\\) indicates 1 if the subscript condition is True and 0 otherwise. For an input \\(\\pmb{x}\\), the unnormalized log probability for the \\(i\\)-th class is computed by \\(\\mathrm{sim}(f_{\\pmb{\\theta}}(\\pmb{x}), \\phi_i^P)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.709, + 0.788, + 0.77 + ], + "angle": 0, + "content": "Although this baseline bypasses the forgetting and overfitting issues, it heavily relies on the quality of the representation trained solely on the base classes. Thus, the main focus of this paper is to investigate the important factors that influence representation learning for FSCIL and strategies to improve them." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.788, + 0.788, + 0.804 + ], + "angle": 0, + "content": "3.3 Transferability, feature spread, and its adverse effects on FSCIL" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "As shown in Fig. 1a, the baseline method exhibits a narrow intra-class distribution, widely perceived as representation collapse [31]. Recent studies [7,21,46] have" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.484, + 0.129 + ], + "angle": 0, + "content": "J. Oh*, S. Baik*, and K.M. Lee" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.147, + 0.756, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.262, + 0.789, + 0.348 + ], + "angle": 0, + "content": "Fig. 2: The impact of the spread of representation. Stronger emphasis on self-supervised contrastive loss (larger \\(\\lambda_{\\mathrm{ssc}}\\)) and low temperature (skyblue) enhances the new-class performance \\(A_N\\) (left), but at the expense of base-class performance \\(A_B\\) (center). The reduced base-class performance is mainly attributed to the excessive intra-class variation, adversely affecting the class prototype representation (right). The experiments are conducted on CIFAR100 dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.79, + 0.482 + ], + "angle": 0, + "content": "revealed that the collapsed representation shows poor transferability due to the loss of shareable low- and mid-level features that new classes can benefit from. As such, they suggest joint optimization with a self-supervised contrastive (SSC) task [8,18], which promotes the spread of features and thus the sharing of features among classes. SSC learning optimizes infoNCE loss [30], regarding an augmented view from a query image as positive and the other images as negative samples. The SSC loss for a positive pair \\((i,j)\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.338, + 0.49, + 0.788, + 0.529 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s s c}} ^ {(i, j)} = - \\log \\frac {\\exp \\left(\\frac {1}{\\tau} \\operatorname {s i m} \\left(\\boldsymbol {z} _ {i} , \\boldsymbol {z} _ {j}\\right)\\right)}{\\sum_ {k = 1} ^ {B} \\mathbb {1} _ {[ k \\neq i ]} \\exp \\left(\\frac {1}{\\tau} \\operatorname {s i m} \\left(\\boldsymbol {z} _ {i} , \\boldsymbol {z} _ {k}\\right)\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.789, + 0.567 + ], + "angle": 0, + "content": "where \\( B \\) is the number of samples, including augmented images, and \\( z_{j} \\) is the feature from an augmented view of \\( \\pmb{x}_{i} \\). The loss is averaged over all positive pairs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.789, + 0.702 + ], + "angle": 0, + "content": "In parallel, several studies have proposed to reduce the temperature and margin parameters in the softmax cross-entropy loss as another way to encourage feature sharing [25,27,55]. Our empirical analysis in Sec. S4 of the supplementary material demonstrates that the temperature parameter affects the transferability more than the margin. Thus, we employ a low-temperature parameter in the SCE loss to further enhance the transferability of learned representations. Indeed, we observe that the combination of \\(\\mathcal{L}_{\\mathrm{ssc}}\\) and \\(\\mathcal{L}_{\\mathrm{ce}}\\) with low \\(\\tau\\) results in better new-class performance in the FSCIL problem due to a larger spread of features, as demonstrated in Fig. 1b and the left figure of Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.789, + 0.842 + ], + "angle": 0, + "content": "However, as illustrated in the center figure in Fig. 2, we observe that the pursuit of better transferability results in a degradation in discriminability on the base classes. While exploring the dilemma, we discover that the performance decline on the base classes mainly arises from the base class classifier replacement (CR) strategy, which is introduced in Sec. 3.2. The figure on the right in Fig. 2 shows the accuracy on the base classes subsequent to the base session training \\((A_B^0)\\), without considering the new classes, both before and after CR, denoted by \\(A_B^{0,\\mathrm{bef}}\\) and \\(A_B^{0,\\mathrm{aft}}\\), respectively. Before CR, \\(A_B^{0,\\mathrm{bef}}\\) exhibits a relatively high value and a small variance as \\(\\tau\\) and \\(\\lambda_{\\mathrm{ssc}}\\) vary (indicated by the solid lines). Conversely," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "CLOSER: Representation Learning for Few-Shot Class-Incremental Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.147, + 0.644, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.28, + 0.788, + 0.337 + ], + "angle": 0, + "content": "Fig. 3: Effect of minimizing inter-class distance. As the weight of \\(\\mathcal{L}_{\\mathrm{inter}}\\), denoted by \\(\\lambda_{\\mathrm{inter}}\\), increases, the performance on the new classes increases (skyblue) and the performance loss on the base classes induced by CR is greatly alleviated (purple). The experiments are conducted on CUB200 dataset." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.362, + 0.788, + 0.471 + ], + "angle": 0, + "content": "\\(A_B^{0,\\mathrm{aft}}\\) is considerably lower than \\(A_B^{0,\\mathrm{bef}}\\) with a relatively higher variance (indicated by the dotted lines). Considering that CR is introduced to bypass overfitting and catastrophic forgetting in the FSCIL problem, these findings suggest that the previous methods for improving the transferability of representations are not effective in enhancing the trade-off between transferability and discriminability in the context of the FSCIL problem. Thus, it is essential to develop a representation learning method tailored for the FSCIL problem." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.49, + 0.503, + 0.505 + ], + "angle": 0, + "content": "3.4 Inter-Class Distance Matters" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.513, + 0.789, + 0.664 + ], + "angle": 0, + "content": "As discussed in Sec. 3.3, learning shareable features through representation spreading proves advantageous for transferability. Yet, it also harms discriminability in the process of CR. Based on the observation in Fig. 1b, we hypothesize that such a dilemma appears to arise from a large inter-class distance since the representation spreading could push features into the extensive inter-class space, which may dilute the information on the base classes. Furthermore, we argue that the large inter-class distance may impede effective feature sharing among classes, undermining the transferability of learned representations. Consequently, to retain the knowledge on base classes while promoting effective feature sharing, we introduce a novel loss function that minimizes the inter-class distance:" + }, + { + "type": "equation", + "bbox": [ + 0.315, + 0.67, + 0.788, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {i n t e r}} = - \\frac {1}{\\sum_ {i = 1} ^ {B} \\sum_ {j > i} ^ {B} \\mathbb {1} _ {[ y _ {i} \\neq y _ {j} ]}} \\sum_ {i = 1} ^ {B} \\sum_ {j > i} ^ {B} \\mathbb {1} _ {[ y _ {i} \\neq y _ {j} ]} \\operatorname {s i m} \\left(\\boldsymbol {z} _ {i}, \\boldsymbol {z} _ {j}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.735, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Fig. 1c displays that the spread of intra-class features is well regulated by applying \\(\\mathcal{L}_{\\mathrm{inter}}\\) and Fig. 3 demonstrates the performance decline from CR is alleviated by minimizing \\(\\mathcal{L}_{\\mathrm{inter}}\\) (indicated by the purple line). Moreover, the results indicated by the skyblue line show that reducing inter-class distance improves the performance on the new classes, corroborating our hypothesis." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Our assertion initially seems counter-intuitive, diverging from the common belief of prior works [20,40,47,51] that maximizing inter-class distance may be" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.484, + 0.128 + ], + "angle": 0, + "content": "J. Oh*, S. Baik*, and K.M. Lee" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.145, + 0.407, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.305, + 0.248, + 0.324, + 0.26 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.413, + 0.147, + 0.593, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.491, + 0.248, + 0.512, + 0.26 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.147, + 0.777, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.678, + 0.248, + 0.697, + 0.26 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.271, + 0.789, + 0.397 + ], + "angle": 0, + "content": "Fig. 4: (a) Sanity test for \\(\\mathcal{T}(f_{\\theta})\\): \\(\\mathcal{T}(f_{\\theta})\\) has a positive correlation with the performance on the new classes. Each data point is obtained by different configurations of \\(\\tau\\) and \\(\\lambda_{\\mathrm{ssc}}\\) (without \\(\\mathcal{L}_{\\mathrm{inter}}\\)). (b), (c) Relationship between inter-class distance, \\(\\mathcal{T}(f_{\\theta})\\), and \\(A_N\\): Integrated with the representation spreading, reducing inter-class distance encourages better transferability (red points). However, the tendency is broken when reducing inter-class distance without representation spreading (blue points). Please refer to Sec. 3.5 for theoretical support for these observations. The dots with greater transparency correspond to smaller \\(\\lambda_{\\mathrm{inter}}\\), ranging from 0 to 1 with intervals of 0.1. We set \\(\\lambda_{\\mathrm{ssc}}\\) as 0.1 when it is used. The experiments are conducted on CIFAR100 dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.789, + 0.522 + ], + "angle": 0, + "content": "beneficial for reserving representation space for future new classes. To further validate the efficacy of reducing inter-class distance with respect to transferability, we propose a measure to quantify how new class samples are distinct from base-class representations. We define the measure as the averaged relative angular distance between a new-class sample and its nearest base-class prototype with respect to the averaged angular distance among all base-class prototype pairs:" + }, + { + "type": "equation", + "bbox": [ + 0.322, + 0.535, + 0.788, + 0.582 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} (f _ {\\boldsymbol {\\theta}}) = \\frac {\\frac {1}{| \\mathcal {D} _ {t e s t} ^ {(> 0)} |} \\sum_ {(\\boldsymbol {x} _ {j} , y _ {j}) \\in \\mathcal {D} _ {t e s t} ^ {(> 0)} i} \\min _ {i} \\angle (\\boldsymbol {z} _ {j} , \\phi_ {b a s e , i} ^ {P})}{\\sum_ {j = 1} ^ {| \\mathcal {C} ^ {(0)} |} \\sum_ {k > j} ^ {| \\mathcal {C} ^ {(0)} |} \\angle \\left(\\phi_ {j} ^ {P} , \\phi_ {k} ^ {P}\\right) / \\left(^ {| \\mathcal {C} ^ {(0)} |} _ {2}\\right)}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.789, + 0.688 + ], + "angle": 0, + "content": "where \\(\\phi_{base,i}^{P}\\) and \\(\\angle (\\cdot ,\\cdot)\\) indicate the \\(i\\) -th base-class prototype and the angular distance between two input vectors, respectively. We introduce the denominator to normalize the varying sizes of the representation space depending on methods. If the representation produced by \\(f_{\\theta}\\) has a distinguishable representation for new classes, \\(\\mathcal{T}(f_{\\theta})\\) would be large. For the sanity test, we check the relationship between \\(\\mathcal{T}(f_{\\theta})\\) and the accuracy of the new class, which is shown in Fig. 4a." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.79, + 0.84 + ], + "angle": 0, + "content": "Using this measure, we analyze the relationship between inter-class distance, the spread of features, and the transferability of learned representation. To do so, we train the feature extractor using \\(\\mathcal{L}_{\\mathrm{ce}}\\) with a low temperature, \\(\\mathcal{L}_{\\mathrm{ssc}}\\), and \\(\\mathcal{L}_{\\mathrm{inter}}\\) with varying loss weights for \\(\\mathcal{L}_{\\mathrm{ssc}}\\) and \\(\\mathcal{L}_{\\mathrm{inter}}\\), denoted by \\(\\lambda_{\\mathrm{ssc}}\\) and \\(\\lambda_{\\mathrm{inter}}\\), respectively. Figs. 4b and 4c show the relationship between the inter-class distance, \\(\\mathcal{T}(f_{\\theta})\\), and the performance on the new classes. The results demonstrate that the joint optimization of \\(\\mathcal{L}_{\\mathrm{ssc}}\\) and \\(\\mathcal{L}_{\\mathrm{inter}}\\) leads to an increase in both \\(\\mathcal{T}(f_{\\theta})\\) and \\(A_N\\), indicating that smaller inter-class distances lead to more discriminability between base classes and new classes. The analysis corroborates our seemingly counter-intuitive claim that the closer classes are, the better." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "CLOSER: Representation Learning for Few-Shot Class-Incremental Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "In summary, we have observed that the spread of representation achieved by employing a low temperature in the SCE loss and self-supervised contrastive loss is advantageous for learning transferable representation. However, spread representation itself cannot address the trade-off between transferability to new classes and discriminability on base classes in the context of the FSCIL problem. Our analysis demonstrates that decreasing inter-class distance enhances discriminability by regularizing the intra-class spread and improves the transferability by promoting the effective learning of shareable information among classes when combined with the feature-spread-encouraging loss. Consequently, our final loss function is the combination of cross-entropy loss with a lower temperature parameter, self-supervised contrastive loss, and inter-class distance minimizing loss:" + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.329, + 0.788, + 0.346 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {c e}} + \\lambda_ {\\mathrm {s s c}} \\mathcal {L} _ {\\mathrm {s s c}} + \\lambda_ {\\mathrm {i n t e r}} \\mathcal {L} _ {\\mathrm {i n t e r}}. \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.363, + 0.621, + 0.379 + ], + "angle": 0, + "content": "3.5 Information Bottleneck Theory Perspective" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.788, + 0.522 + ], + "angle": 0, + "content": "In this subsection, we provide theoretical support for the proposed representation learning objective in Eq. (6) from the perspective of the information bottleneck (IB) theory [3, 37, 42]. The IB theory describes the goal of representation learning as finding a good trade-off between complexity and accuracy through finding minimal information from inputs necessary to preserve maximal information about the targets. As discussed by Cui et al. [12], we consider the objective of IB theory to be closely related to learning transferable representations since while achieving the objective, a network could be guided to learn intrinsic knowledge rather than task-irrelevant shortcuts [17], leading to better transferability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.603 + ], + "angle": 0, + "content": "For an image classification task, let \\( X \\in \\mathbb{R}^{h \\times w \\times 3} \\), \\( Y \\in \\mathbb{R}^C \\), and \\( Z = \\frac{f_{\\theta}(X)}{\\|f_{\\theta}(X)\\|} \\in \\mathbb{R}^d \\) denote the input, label, and normalized latent representation variables, respectively, where \\( h \\) and \\( w \\) are the spatial sizes of the image, \\( C \\) is the number of classes, and \\( d \\) is the dimension of the latent representations. The aforementioned trade-off has been formulated as the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.414, + 0.613, + 0.787, + 0.629 + ], + "angle": 0, + "content": "\\[\n\\max I (Y; Z) - \\beta I (X; Z), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.636, + 0.788, + 0.716 + ], + "angle": 0, + "content": "where \\( I(\\cdot ;\\cdot) \\) indicates the mutual information between two random variables and \\( \\beta > 0 \\) is a Lagrange multiplier. In this work, we consider an alternative trade-off objective, \\( \\max \\frac{I(Y;Z)}{\\beta I(X;Z)} \\), which is adopted by several works [29,38]. After omitting \\( \\beta \\) for simplicity and modest assumptions, we derive the following lower bound of the IB trade-off objective:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.724, + 0.788, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\frac {I (Y ; Z)}{I (X ; Z)} \\geq 1 - \\frac {d \\cdot \\log (2 \\pi e) + \\frac {1}{C} \\sum_ {i = 1} ^ {C} \\log | \\Sigma_ {W _ {i}} |}{d \\cdot \\log (2 \\pi e) + \\log | \\Sigma_ {T} |}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "where \\(\\Sigma_{W_i}\\) and \\(\\Sigma_T\\) indicate the covariance matrices of \\(Z\\) within the \\(n\\)-th class and overall classes, respectively. As proven by Lemma S1 in the supplementary material, both the numerator and denominator in the fractional term of the lower bound in Eq. (8) are negative, leading to the following theorem:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.483, + 0.128 + ], + "angle": 0, + "content": "J. Oh*, S. Baik*, and K.M. Lee" + }, + { + "type": "table_caption", + "bbox": [ + 0.255, + 0.145, + 0.746, + 0.159 + ], + "angle": 0, + "content": "Table 1: 10-way 5-shot incremental learning results on CUB200." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.172, + 0.753, + 0.372 + ], + "angle": 0, + "content": "
MethodAcc. in each session (%)PD (%) ↓
012345678910
Baseline79.9276.2373.1869.4567.8365.7464.5463.3361.5661.2760.1019.83
TOPIC [41]68.6862.4954.8149.9945.2541.4038.3535.3632.2228.3126.2642.42
F2M [36]81.0778.1675.5772.8970.8668.1767.0165.2663.3661.7660.2620.81
CEC [48]75.8571.9468.5063.5062.4358.2757.7355.8154.8353.5252.2823.57
IDLVQ-C [6]77.3774.7270.2867.1365.3463.5262.1061.5459.0458.6857.8119.56
ALICE [33]77.4072.7070.6067.2065.9063.4062.9061.9060.5060.6060.1017.30
CLOM [55]79.5776.0772.9469.8267.8065.5663.9462.5960.6260.3459.5819.99
Entropy Reg. [28]75.9072.1468.6463.7662.5859.1157.8255.8954.9253.5852.3923.51
LIMIT [52]75.8973.5571.9968.1467.4263.6162.4061.3559.9158.6657.4118.48
MetaFSCIL [11]75.9072.4168.7864.7862.9659.9958.3056.8554.7853.8252.6423.26
FACT [51]75.9073.2370.8466.1365.5662.1561.7459.8358.4157.8956.9418.96
S3C [22]80.6277.5573.1968.5468.0564.3363.5862.0760.6159.7958.9520.83
SAVC [40]81.8577.9274.9570.2169.9667.0266.1665.3063.8463.1562.5019.35
NC-FSCIL [47]80.4575.9872.3070.2868.1765.1664.4363.2560.6660.0159.4421.01
GKEAL [54]78.8875.6272.3268.6267.2364.2662.9861.8960.2059.2158.6720.21
CABD [49]79.1275.3772.8069.0567.5365.1264.0063.5161.8761.4760.9318.19
CLOSER (Ours)79.4075.9273.5070.4769.2467.2266.7365.6964.0064.0263.5815.82
" + }, + { + "type": "table_caption", + "bbox": [ + 0.252, + 0.391, + 0.749, + 0.405 + ], + "angle": 0, + "content": "Table 2: 5-way 5-shot incremental learning results on CIFAR100." + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.418, + 0.724, + 0.555 + ], + "angle": 0, + "content": "
MethodAcc. in each session (%)PD (%) ↓
012345678
Baseline72.9368.4664.2660.1556.5353.6051.5149.1947.0925.84
ERDIL [16]73.6268.2265.1461.8458.3555.5452.5150.1648.2325.39
CEC [48]73.0768.8865.2661.1958.0955.5753.2251.3449.1423.93
CLOM [55]74.2069.8366.1762.3959.2656.4854.3652.1650.2523.95
Entropy Reg. [28]74.470.266.5462.5159.7156.5854.5252.3950.1424.26
FACT [51]74.6072.0967.5663.5261.3858.3656.2854.2452.1022.50
LIMIT [52]73.8172.0967.8763.8960.7057.7755.6753.5251.2322.58
MetaFSCIL [11]74.5070.1066.8462.7759.4856.5254.3652.5649.9724.53
GKEAL [54]74.0170.4567.0163.0860.0157.3055.5053.3951.4022.61
CLOSER (Ours)75.7271.8368.3264.6261.9159.2557.5355.4353.3222.40
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.589, + 0.788, + 0.625 + ], + "angle": 0, + "content": "Theorem 1. The lower bound of \\( \\frac{I(Y;Z)}{I(X;Z)} \\) in Eq. (8) is a monotonically increasing function of \\( |\\Sigma_{W_i}| \\) and a monotonically decreasing function of \\( |\\Sigma_T| \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.788, + 0.656 + ], + "angle": 0, + "content": "The detailed derivations are presented in Sec. S1 of the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Generally, the determinant of a covariance matrix is correlated with the spread or variability of variables across all dimensions. Thus, Theorem 1 suggests that the IB trade-off can be enhanced by increasing the intra-class variability \\((\\mathcal{L}_{\\mathrm{ce}}\\) with a low temperature and \\(\\mathcal{L}_{\\mathrm{ssc}})\\) while suppressing the overall representation space \\((\\mathcal{L}_{\\mathrm{inter}})\\), supporting the proposed learning objective in Eq. (6) and our claims for inter-class distance minimization. Intuitively, our objective functions encourage greater overlap in representations among different classes, promoting more shareable and compact representations. Our theoretic analysis also underpins the findings in Fig. 4 that reducing inter-class distance without representation spreading loss is observed to result in less transferable representations. Since compressing the overall feature space diminishes the intra-class variability, the IB trade-off may decrease, leading to degraded transferability." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "CLOSER: Representation Learning for Few-Shot Class-Incremental Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.236, + 0.145, + 0.765, + 0.159 + ], + "angle": 0, + "content": "Table 3: 5-way 5-shot incremental learning results on miniImageNet." + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.172, + 0.725, + 0.356 + ], + "angle": 0, + "content": "
MethodAcc. in each session (%)PD (%) ↓
012345678
Baseline72.2767.4663.2659.7356.5653.5350.9048.9347.2625.01
TOPIC [41]61.3150.0945.1741.1637.4835.5232.1929.4624.4236.89
F2M [36]67.2863.8060.3857.0654.0851.3948.8246.5844.6522.63
CEC [48]72.0066.8362.9759.4356.7053.7351.1949.2447.6324.37
IDLVQ-C [6]64.7759.8755.9352.6249.8847.5544.8343.1441.8422.93
Subspace Reg. [2]80.3773.7668.3664.0760.3656.2753.1050.4547.5532.83
ALICE [33]80.6070.6067.4064.5062.5060.0057.8056.8055.7024.90
CLOM [55]73.0868.0964.1660.4157.4154.2951.5449.3748.0025.08
Entropy Reg. [28]71.8467.1263.2159.7757.0153.9551.5549.5248.2123.63
LIMIT [52]72.3268.4764.3060.7857.9555.0752.7050.7249.1423.13
MetaFSCIL [11]72.0467.9463.7760.2957.5855.1652.9050.7949.1922.85
FACT [51]72.5669.6366.3862.7760.6057.3354.3452.1650.4922.07
GKEAL [54]73.5968.9065.3362.2959.3956.7054.2052.5951.3122.28
CABD [49]74.6570.4366.2962.7760.7557.2454.7953.6552.2222.43
CLOSER (Ours)76.0271.6167.9964.6961.7058.9456.2354.5253.3322.69
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.383, + 0.377, + 0.4 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.414, + 0.439, + 0.43 + ], + "angle": 0, + "content": "4.1 Experimental Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.574 + ], + "angle": 0, + "content": "Dataset. Following the benchmark settings proposed by Tao et al. [41], we evaluate the proposed method on CIFAR100 [26], miniImageNet [43], and CUB200 [44]. For CIFAR100 and miniImageNet, the total number of classes is 100: 60 base classes and 40 new classes. The 40 new classes are split into 8 disjoint sets of 5 classes, each of which is sequentially provided with 5 training examples per class (5-way 5-shot) in each incremental session. As for CUB200, there total number of classes is 200, with 100 base classes and 100 new classes. 100 new classes are split into 10 disjoint sets of 10 classes, each of which is sequentially provided with 5 training examples per class (10-way 5-shot) in each incremental session." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.789, + 0.756 + ], + "angle": 0, + "content": "Implementation. Following Zhang et al. [48], we use ResNet-20 [19] for CIFAR100 experiments and ResNet-18 [19] for both miniImageNet and CUB200 experiments. We follow the conventions to use the ResNet-18 model pre-trained on the ImageNet dataset [35] for CUB200. We set the mini-batch size to 128, 128, and 256 for CIFAR100, miniImageNet, and CUB200 experiments, respectively. The temperature parameter \\(\\tau\\) for the baseline method is 1/16 and 'low temperature' indicates \\(\\tau = 1 / 32\\). We set \\(\\lambda_{\\mathrm{ssc}}\\) as 0.1,0.1, and 0.01 for CIFAR100, miniImageNet, and CUB200, respectively, and \\(\\lambda_{\\mathrm{inter}}\\) as 1, 0.5, and 1.5 for CIFAR100, miniImageNet, and CUB200, respectively. These hyper-parameters are searched via validation using synthesized validation sets. For mutual information estimation, we adopt MINE [4] method and implement it based on the open-source code2. More details are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.756, + 0.789, + 0.817 + ], + "angle": 0, + "content": "Evaluation. We use the accuracy on base \\((A_B)\\), new \\((A_N)\\), and the whole classes \\((A_W)\\) as metrics to assess the discriminability, transferability, and the trade-off between them in learned representations. Additionally, we use the performance drop (PD) between the accuracy at the end of the base session (session 0) and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.238, + 0.825, + 0.49, + 0.84 + ], + "angle": 0, + "content": "\\(^{2}\\)https://github.com/gtegner/mine-pytorch" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.483, + 0.128 + ], + "angle": 0, + "content": "J. Oh*, S. Baik*, and K.M. Lee" + }, + { + "type": "table_caption", + "bbox": [ + 0.347, + 0.145, + 0.656, + 0.159 + ], + "angle": 0, + "content": "Table 4: Ablation studies on CIFAR100." + }, + { + "type": "table", + "bbox": [ + 0.374, + 0.172, + 0.631, + 0.263 + ], + "angle": 0, + "content": "
lowτL SSCL interAB(%)AN(%)AW(%)PD(%)
XXX70.1320.9550.4624.81
X✓X69.9522.8051.0924.59
XX✓69.9520.0549.9926.78
X✓✓71.4322.8351.9925.26
✓XX68.3324.1050.6423.23
✓✓X66.5825.0849.9823.15
✓X✓70.1722.4051.0625.27
✓✓✓70.7227.2353.3222.40
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.293, + 0.787, + 0.34 + ], + "angle": 0, + "content": "the accuracy at the last incremental session to evaluate the degree to which old knowledge is forgotten and new knowledge is learned simultaneously. All experimental results of CLOSER are obtained by averaging 3 trials." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.369, + 0.57, + 0.384 + ], + "angle": 0, + "content": "4.2 Comparison with the Existing Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.789, + 0.598 + ], + "angle": 0, + "content": "We compare the proposed method, dubbed CLOSER, with prior arts on CUB200 (Tab. 1), CIFAR100 (Tab. 2), and miniImageNet (Tab. 3). We observe that CLOSER achieves state-of-the-art performance on both CUB200 and CIFAR100 datasets, surpassing the results of previous methods by a large margin with respect to \\( A_W \\) and PD. With miniImageNet, the proposed method exhibits substantially higher \\( A_W \\) than the method with the lowest PD and achieves lower PD than the method with the highest \\( A_W \\), which means that the proposed method achieves a better balance between the performance on base and the new classes. It is worth noting that CLOSER shows such outstanding performance without any assistance from the storage of previous samples (F2M, ERDIL, IDLVQ-C, and CABD), additional computational modules (CEC, CLOM, NC-FSCIL, SAVC, and MetaFSCIL), and test-time data augmentation (S3C and SAVC), suggesting the critical importance of learning effective representations in FSCIL." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.627, + 0.403, + 0.642 + ], + "angle": 0, + "content": "4.3 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.841 + ], + "angle": 0, + "content": "To verify the efficacy of the individual components in the proposed method, we perform ablation studies, which are shown in Tab. 4. The increase in \\( A_N \\) when employing a lower temperature in the softmax cross-entropy loss (low \\( \\tau \\)) or a self-supervised contrastive loss (\\( \\mathcal{L}_{\\mathrm{ssc}} \\)) confirms the advantage of feature spread in enhancing the transferability of representations. While transferability can be greatly improved by using low \\( \\tau \\) and \\( \\mathcal{L}_{\\mathrm{ssc}} \\), a substantial decline in base-class performance \\( A_B \\) is observed, as discussed in Sec. 3.4. The result from the case where all components are utilized demonstrates that this issue can be effectively resolved by minimizing inter-class distance. Furthermore, as discussed in Sec. 3.4, reducing inter-class distance is observed to improve transferability, especially when used with the representation spreading methods. The comprehensive results of ablation studies confirm and support our claims." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "CLOSER: Representation Learning for Few-Shot Class-Incremental Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.239, + 0.146, + 0.413, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.246, + 0.375, + 0.258 + ], + "angle": 0, + "content": "(a) Base Classes" + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.146, + 0.583, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.246, + 0.55, + 0.258 + ], + "angle": 0, + "content": "(b) New Classes" + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.147, + 0.764, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.623, + 0.246, + 0.73, + 0.259 + ], + "angle": 0, + "content": "(c) Whole Classes" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.269, + 0.789, + 0.327 + ], + "angle": 0, + "content": "Fig. 5: Information bottleneck trade-off analysis. We compare representations acquired by three different methods by assessing the information bottleneck (IB) tradeoff. 'RS' refers to representation spreading methods. We indicate the final models with black edges. The experiments are conducted on the CIFAR100 dataset." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.347, + 0.411, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.29, + 0.447, + 0.364, + 0.459 + ], + "angle": 0, + "content": "(a) Baseline" + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.347, + 0.586, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.446, + 0.447, + 0.557, + 0.459 + ], + "angle": 0, + "content": "(b) Baseline \\(+\\mathrm{RS}\\)" + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.347, + 0.761, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.447, + 0.736, + 0.459 + ], + "angle": 0, + "content": "(c) CLOSER (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.471, + 0.79, + 0.569 + ], + "angle": 0, + "content": "Fig. 6: T-SNE visualization of learned representations. The blue and red points indicate the base and new class samples, respectively. We measure \\(\\mathcal{T}(f_{\\theta})\\) to quantify the transferability of the learned representation. We conducted the experiments on CIFAR100 with reduced classes (20 base classes and 10 new classes) for better visualization. The classification results for each experiment are as follows: (a): \\(A_B = 78.85\\%\\), \\(A_N = 13.50\\%\\), \\(A_W = 57.07\\%\\). (b): \\(A_B = 77.85\\%\\), \\(A_N = 25.50\\%\\), \\(A_W = 60.22\\%\\). (c): \\(A_B = 78.88\\%\\), \\(A_N = 28.80\\%\\), \\(A_W = 62.18\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.604, + 0.607, + 0.62 + ], + "angle": 0, + "content": "4.4 Information Bottleneck trade-off Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.84 + ], + "angle": 0, + "content": "In Sec. 3.5, we show the connection between the proposed objective function and the information bottleneck (IB) trade-off. To validate our analysis, we measure the mutual information between representations and inputs \\( I(X;Z) \\) and the mutual information between representations and targets \\( I(Y;Z) \\) for the baseline, representation spread approach, and our method CLOSER, as shown in Fig. 5. In the figure, the more left (lower \\( I(X;Z) \\)) and upper (higher \\( I(Y;Z) \\)) regions imply better IB trade-off. Our proposed method CLOSER demonstrates to have found a better IB trade-off, especially when considering whole classes, including base and new classes. The results are encouraging in that CLOSER is able to find a better IB trade-off for all classes, when the feature extractor is only trained on base classes and fixed afterwards. The results demonstrate that CLOSER is effective in tackling a particularly difficult challenge of FSCIL: learning transferable and discriminative features from base classes." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.483, + 0.128 + ], + "angle": 0, + "content": "J. Oh*, S. Baik*, and K.M. Lee" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.146, + 0.395, + 0.163 + ], + "angle": 0, + "content": "4.5 T-SNE Analysis" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.175, + 0.788, + 0.356 + ], + "angle": 0, + "content": "For qualitative evaluation, we visualize the learned representation trained by different configurations of the proposed losses, which is illustrated in Fig. 6. The value of \\(\\mathcal{T}(f_{\\theta})\\) in the right below in each figure is for measuring the transferability of representation. The baseline representation (a) shows the characteristics of the base class, represented by the features of the new classes mapped to those of the base classes, also indicated by the low value of \\(\\mathcal{T}(f_{\\theta})\\). Spreading of features (b) largely resolves the overfitting issues, exhibited by larger distances between new classes and base classes; i.e., the increase in \\(\\mathcal{T}(f_{\\theta})\\). Finally, \\(\\mathcal{L}_{\\mathrm{inter}}\\) is used to compensate for the decreased performance on the base classes due to the spread representation. Reducing inter-class distance also enhances the separability between the new class samples and the clusters of base classes, as evidenced by the further increase in \\(\\mathcal{T}(f_{\\theta})\\)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.381, + 0.521, + 0.399 + ], + "angle": 0, + "content": "5 Limitations and Discussions" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.416, + 0.788, + 0.568 + ], + "angle": 0, + "content": "Our work focuses on learning representation that is discriminative yet transferable to unseen classes to tackle the challenges of FSCIL. As such, our work does not consider an option of updating representation with new classes. Continual update of representation can improve the performance on new classes, however at the cost of sacrificing the old-class performance. Thus, it is a question of stability-plasticity dilemma, where our method focuses more on stability while improving plasticity through discriminative and transferable representation. Furthermore, similar to other works on FSCIL, our method is limited to classification tasks. But, we believe our work can have an impact on other domains, akin to the impact of representation spread methods, such as contrastive learning." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.593, + 0.357, + 0.61 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.628, + 0.788, + 0.778 + ], + "angle": 0, + "content": "To tackle convoluted challenges in few-shot class-incremental learning (FSCIL), we focus on representation learning, which plays a crucial role. In contrast to previous FSCIL methods that have focused on maximizing the distance between classes, our experimental and theoretic analysis suggests that the closer classes are, the better for FSCIL, especially when feature sharing is encouraged between classes. Upon the analysis, we propose a simple, yet seemingly counter-intuitive idea: bring classes closer for a better transferability-discriminability Pareto front. Our experimental results and information-bottleneck-theory-based analysis suggest that our work can provide a promising research avenue. As such, we hope that our work will inspire future works and discussions in this research direction." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported in part by the IITP grants [No. 2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University), No.2021-0-02068, and No.2023-0-00156], the NRF grant [No.2021M3A 9E4080782] funded by the Korean government (MSIT)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "CLOSER: Representation Learning for Few-Shot Class-Incremental Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. Achituve, I., Navon, A., Yemini, Y., Chechik, G., Fetaya, E.: Gp-tree: A gaussian process classifier for few-shot incremental learning. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.232 + ], + "angle": 0, + "content": "2. Akyurek, A.F., Akyurek, E., Wijaya, D.T., Andreas, J.: Subspace regularizers for few-shot class incremental learning. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.233, + 0.785, + 0.259 + ], + "angle": 0, + "content": "3. Alemi, A.A., Fischer, I., Dillon, J.V., Murphy, K.: Deep variational information bottleneck. In: ICLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.286 + ], + "angle": 0, + "content": "4. Belghazi, M.I., Baratin, A., Rajeshwar, S., Ozair, S., Bengio, Y., Courville, A., Hjelm, D.: Mutual information neural estimation. In: PMLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.314 + ], + "angle": 0, + "content": "5. Chatfield, K., Simonyan, K., Vedaldi, A., Zisserman, A.: Return of the devil in the details: delving deep into convolutional nets. In: BMVC (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.316, + 0.785, + 0.342 + ], + "angle": 0, + "content": "6. Chen, K., Lee, C.G.: Incremental few-shot learning via vector quantization in deep embedded space. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "7. Chen, M.F., Fu, D.Y., Narayan, A., Zhang, M., Song, Z., Fatahalian, K., Ré, C.: Perfectly balanced: improving transfer and robustness of supervised contrastive learning. In: ICML (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.785, + 0.411 + ], + "angle": 0, + "content": "8. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "9. Cheraghian, A., Rahman, S., Fang, P., Roy, S.K., Petersson, L., Harandi, M.: Semantic-aware knowledge distillation for few-shot class-incremental learning. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.453, + 0.785, + 0.493 + ], + "angle": 0, + "content": "10. Cheraghian, A., Rahman, S., Ramasinghe, S., Fang, P., Simon, C., Petersson, L., Harandi, M.: Synthesized feature based few-shot class-incremental learning on a mixture of subspaces. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "1. Chi, Z., Gu, L., Liu, H., Wang, Y., Yu, Y., Tang, J.: Meta桅: a meta-learning approach for few-shot class incremental learning. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.785, + 0.563 + ], + "angle": 0, + "content": "2. Cui, Q., Zhao, B., Chen, Z.M., Zhao, B., Song, R., Liang, J., Zhou, B., Yoshie, O.: Discriminability-transferability trade-off: an information-theoretic perspective. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.618 + ], + "angle": 0, + "content": "3. Delange, M., Aljundi, R., Masana, M., Parisot, S., Jia, X., Leonardis, A., Slabaugh, G., Tuytelaars, T.: A continual learning survey: Defying forgetting in classification tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.62, + 0.785, + 0.646 + ], + "angle": 0, + "content": "4. Deng, J., Guo, J., Yang, J., Xue, N., Kotsia, I., Zafeiriou, S.: Arcface: additive angular margin loss for deep face recognition. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.687 + ], + "angle": 0, + "content": "15. Donahue, J., Jia, Y., Vinyals, O., Hoffman, J., Zhang, N., Tzeng, E., Darrellin, T.: Decaf: a deep convolutional activation feature for generic visual recognition. In: ICML (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.715 + ], + "angle": 0, + "content": "6. Dong, S., Hong, X., Tao, X., Chang, X., Wei, X., Gong, Y.: Few-shot class-incremental learning via relation knowledge distillation. In: AAAI (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "7. Geirhos, R., Jacobsen, J.H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., Wichmann, F.A.: Shortcut learning in deep neural networks. Nature Machine Intelligence 2(11), 665-673 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.784 + ], + "angle": 0, + "content": "8. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.812 + ], + "angle": 0, + "content": "9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "20. Hersche, M., Karunaratne, G., Cherubini, G., Benini, L., Sebastian, A., Rahimi, A.: Constrained few-shot class-incremental learning. In: CVPR (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.484, + 0.128 + ], + "angle": 0, + "content": "J. Oh*, S. Baik*, and K.M. Lee" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "21. Islam, A., Chen, C.F., Panda, R., Karlinsky, L., Radke, R., Feris, R.: A broad study on the transferability of visual representations with contrastive learning. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.218 + ], + "angle": 0, + "content": "22. Kalla, J., Biswas, S.: S3c: self-supervised stochastic classifiers for few-shot class-incremental learning. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.788, + 0.286 + ], + "angle": 0, + "content": "23. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., Hassabis, D., Clopath, C., Kumaran, D., Hadsell, R.: Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences of the United States of America (PNAS) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.788, + 0.314 + ], + "angle": 0, + "content": "24. Koch, G., Zemel, R., Salakhutdinov, R.: Siamese neural networks for one-shot image recognition. In: ICMLW (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.788, + 0.342 + ], + "angle": 0, + "content": "25. Kornblith, S., Chen, T., Lee, H., Norouzi, M.: Why do better loss functions lead to less transferable features? In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.788, + 0.37 + ], + "angle": 0, + "content": "26. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images. University of Toronto (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.788, + 0.398 + ], + "angle": 0, + "content": "27. Liu, B., Cao, Y., Lin, Y., Li, Q., Zhang, Z., Long, M., Hu, H.: Negative margin matters: Understanding margin in few-shot classification. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.399, + 0.788, + 0.425 + ], + "angle": 0, + "content": "28. Liu, H., Gu, L., Chi, Z., Wang, Y., Yu, Y., Chen, J., Tang, J.: Few-shot class-incremental learning via entropy-regularized data-free replay. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.426, + 0.788, + 0.452 + ], + "angle": 0, + "content": "29. Ngampruetikorn, V., Schwab, D.J.: Information bottleneck theory of high-dimensional regression: relevancy, efficiency and optimality. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.788, + 0.48 + ], + "angle": 0, + "content": "30. van den Oord, A., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.788, + 0.522 + ], + "angle": 0, + "content": "31. Papyan, V., Han, X., Donoho, D.L.: Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences of the United States of America (PNAS) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.788, + 0.55 + ], + "angle": 0, + "content": "32. Peng, C., Zhao, K., Wang, T., Li, M., Lovell, B.C.: Few-shot class-incremental learning from an open-set perspective. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.788, + 0.577 + ], + "angle": 0, + "content": "33. Peng, C., Zhao, K., Wang, T., Li, M., Lovell, B.C.: Few-shot class-incremental learning from an open-set perspective. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.788, + 0.605 + ], + "angle": 0, + "content": "34. Razavian, A.S., Azizpour, H., Sullivan, J., Carlsson, S.: Cnn features off-the-shelf: an astounding baseline for recognition. In: CVPRW (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.788, + 0.66 + ], + "angle": 0, + "content": "35. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV) (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.788, + 0.702 + ], + "angle": 0, + "content": "36. Shi, G., Chen, J., Zhang, W., Zhan, L.M., Wu, X.M.: Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.788, + 0.73 + ], + "angle": 0, + "content": "37. Shwartz-Ziv, R., Tishby, N.: Opening the black box of deep neural networks via information (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.788, + 0.744 + ], + "angle": 0, + "content": "38. Slonim, N., Tishby, N.: Agglomerative information bottleneck. In: NIPS (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.788, + 0.771 + ], + "angle": 0, + "content": "39. Snell, J., Swersky, K., Zemel, R.: Prototypical networks for few-shot learning. In: NIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.788, + 0.813 + ], + "angle": 0, + "content": "40. Song, Z., Zhao, Y., Shi, Y., Peng, P., Yuan, L., Tian, Y.: Learning with fantasy: semantic-aware virtual contrastive constraint for few-shot class-incremental learning. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.788, + 0.84 + ], + "angle": 0, + "content": "41. Tao, X., Hong, X., Chang, X., Dong, S., Wei, X., Gong, Y.: Few-shot class-incremental learning. In: CVPR (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.221, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "CLOSER: Representation Learning for Few-Shot Class-Incremental Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "42. Tishby, N., Zaslavsky, N.: Deep learning and the information bottleneck principle. In: 2015 IEEE Information Theory Workshop (ITW). pp. 1-5 (2015). https://doi.org/10.1109/ITW.2015.7133169" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "43. Vinyals, O., Blundell, C., Lillicrap, T., Kavukcuoglu, K., Wierstra, D.: Matching networks for one shot learning. In: NIPS (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.245 + ], + "angle": 0, + "content": "44. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The Caltech-UCSD Birds-200-2011 Dataset. Tech. rep., Caltech (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.787, + 0.273 + ], + "angle": 0, + "content": "45. Wang, H., Wang, Y., Zhou, Z., Ji, X., Gong, D., Zhou, J., Li, Z., Liu, W.: Cosface: large margin cosine loss for deep face recognition. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "46. Xue, Y., Joshi, S., Gan, E., Chen, P.Y., Mirzasoleiman, B.: Which features are learnt by contrastive learning? on the role of simplicity bias in class collapse and feature suppression. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.314, + 0.787, + 0.342 + ], + "angle": 0, + "content": "47. Yang, Y., Yuan, H., Li, X., Lin, Z., Torr, P., Tao, D.: Neural collapse inspired feature-classifier alignment for few-shot class incremental learning. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.787, + 0.37 + ], + "angle": 0, + "content": "48. Zhang, C., Song, N., Lin, G., Zheng, Y., Pan, P., Xu, Y.: Few-shot incremental learning with continually evolved classifiers. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.787, + 0.398 + ], + "angle": 0, + "content": "49. Zhao, L., Lu, J., Xu, Y., Cheng, Z., Guo, D., Niu, Y., Fang, X.: Few-shot class-incremental learning via class-aware bilateral distillation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.424 + ], + "angle": 0, + "content": "50. Zheng, Y., Pal, D.K., Savvides, M.: Ring loss: convex feature normalization for face recognition. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.424, + 0.787, + 0.452 + ], + "angle": 0, + "content": "51. Zhou, D.W., Wang, F.Y., Ye, H.J., Ma, L., Pu, S., Zhan, D.C.: Forward compatible few-shot class-incremental learning. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.452, + 0.787, + 0.493 + ], + "angle": 0, + "content": "52. Zhou, D.W., Ye, H.J., Ma, L., Xie, D., Pu, S., Zhan, D.C.: Few-shot class-incremental learning by sampling multi-phase tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.493, + 0.787, + 0.522 + ], + "angle": 0, + "content": "53. Zhu, K., Cao, Y., Zhai, W., Cheng, J., Zha, Z.J.: Self-promoted prototype refinement for few-shot class-incremental learning. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.55 + ], + "angle": 0, + "content": "54. Zhuang, H., Weng, Z., He, R., Lin, Z., Zeng, Z.: Gkeal: gaussian kernel embedded analytic learning for few-shot class incremental task. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.577 + ], + "angle": 0, + "content": "55. Zou, Y., Zhang, S., Li, Y., Li, R.: Margin-based few-shot class-incremental learning with class-level overfitting mitigation. In: NeurIPS (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.577 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CLOSER_ Towards Better Representation Learning for Few-Shot Class-Incremental Learning/10463618-6ae0-4309-91d1-f1ab71be0dbc_origin.pdf b/2024/CLOSER_ Towards Better Representation Learning for Few-Shot Class-Incremental Learning/10463618-6ae0-4309-91d1-f1ab71be0dbc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..19a7ccf5815b074c6e99928174f7a5de8b3f8d97 --- /dev/null +++ b/2024/CLOSER_ Towards Better Representation Learning for Few-Shot Class-Incremental Learning/10463618-6ae0-4309-91d1-f1ab71be0dbc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7cca7bb00f2d84356b161c44e32fa76ed76b520d27de362ec8d9c03c7c52e66c +size 2443412 diff --git a/2024/CLOSER_ Towards Better Representation Learning for Few-Shot Class-Incremental Learning/full.md b/2024/CLOSER_ Towards Better Representation Learning for Few-Shot Class-Incremental Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cfae8f805b4afac275b3883dddbddb657197ff55 --- /dev/null +++ b/2024/CLOSER_ Towards Better Representation Learning for Few-Shot Class-Incremental Learning/full.md @@ -0,0 +1,314 @@ +# CLOSER: Towards Better Representation Learning for Few-Shot Class-Incremental Learning + +Junghun Oh $^{*1}$ , Sungyong Baik $^{*3}$ , and Kyoung Mu Lee $^{1,2}$ + +$^{1}$ Dept. of ECE&ASRI, $^{2}$ IPAI, Seoul National University + +$^{3}$ Dept. of Data Science, Hanyang University + +{dh6dh,kyoungmu}@snu.ac.kr, dsybaik@hanyang.ac.kr + +Abstract. Aiming to incrementally learn new classes with only few samples while preserving the knowledge of base (old) classes, few-shot class-incremental learning (FSCIL) faces several challenges, such as overfitting and catastrophic forgetting. Such a challenging problem is often tackled by fixing a feature extractor trained on base classes to reduce the adverse effects of overfitting and forgetting. Under such formulation, our primary focus is representation learning on base classes to tackle the unique challenge of FSCIL: simultaneously achieving the transferability and discriminability of the learned representation. Building upon the recent efforts for enhancing the transferability, such as promoting the spread of features, we find that trying to secure the spread of features within a more confined feature space enables the learned representation to strike a better balance between the transferability and discriminability. Thus, in stark contrast to prior beliefs that the inter-class distance should be maximized, we claim that the CLOSER different classes are, the better for FSCIL. The empirical results and analysis from the perspective of information bottleneck theory justify our simple yet seemingly counterintuitive representation learning method, raising research questions and suggesting alternative research directions. The code is available here. + +Keywords: Few-shot class incremental learning $\cdot$ Representation learning $\cdot$ Transferability + +# 1 Introduction + +Owing to its strong representation power, deep neural networks (DNNs) boast their outstanding performance across various fields. However, such feats require tremendous human effort and time to collect an immense amount of data with accurate annotation. The data hunger of DNNs poses a challenge, especially in dynamic real-world environments, where DNNs are required to learn new concepts with few examples while retaining previously learned concepts. To tackle the challenge, few-shot class-incremental learning (FSCIL) [41] aims to design artificial intelligence systems that can learn new classes with few examples while maintaining performance on previously seen classes. + +![](images/169eb00ebbd074501d782d682f0fc5b78e4b5ba58ea7c8213894f5de666ea8d7.jpg) + +![](images/5106805d7831abf3b8fa0a8e3d8025779bee3b9805bf901aa0526eba4c6aa5d4.jpg) + +![](images/41a84f1362785251bed844ad8dd6fd608e54f174b74fa575b1cfe0ce748458cd.jpg) +$A_{B} = 98.61\%$ $A_{N} = 11.49\%$ +(a) Baseline + +![](images/93456906fdc827458cce25de5e95384bfb3deb29fd3b2f7afe55a5acc42bbd1d.jpg) + +![](images/c54bdd2f5e3287f43e2a5feee594b7c6cad4949889d4c9f4cba3daa3a98f1b32.jpg) + +![](images/34a0548c6ff7eb19059580bb6282bb2382c4b01dadf1a55240610f3b75f77299.jpg) +$A_{B} = 87.42\%$ $A_{N} = 33.34\%$ +(b) Baseline $^+$ sensation spreading +Fig. 1: Visualization of representation trained on MNIST1. (a) Baseline [20, 48] exhibits great base-class discriminability (large inter-class distance) but weak transferability to the new classes (huge overlap between new and base classes leading to misclassification). (b) Baseline + representation spreading [7, 21, 25] benefits the new classes (less overlap with the base classes), while compromising base-class discriminability in the context of FSCIL (dispersed intra-class features leading to less accurate class representation with class prototypes). (c) CLOSER (Ours): Dispersing features in a narrowed feature space enhances both discriminability on the base classes (less deviation between intra-class features and class prototypes) and transferability to the new classes (even less overlap between the base and new classes). For instance, the 4 and 9 classes are not distinguishable in (b) and even less in (a), but CLOSER can yield representation that successfully discriminates them. + +![](images/e65ef29366675c9410b0b44779e95090877361cc8d3032f3259bc0231880d136.jpg) + +![](images/fc592dfb548ce2068b224d5307b7346f9237eb9f7a85d071f08a1c4d16830eb3.jpg) + +![](images/c72aa6d7210f35c98de28e513eb0e36269705cef9f8371cffdb7160ded88641c.jpg) +$A_{R} = 92.47\%$ $A_{N} = 42.14\%$ +(c) CLOSER (Ours) +$A_{B}$ : accuracy on base classes $A_N$ : accuracy on new classes + +![](images/85c9814ec26e4ed9d35e51bde9a503604fa9d4368896730f2c204003733edc86.jpg) +Base Classes + +![](images/6e4546e4bafde37e6611b0a1a7dc42fe9d1333d3134b1ead2a86b0565ef4ad36.jpg) +New Classes +Class prototypes + +To achieve the goal of FSCIL, we need to address catastrophic forgetting (forgetting of previous knowledge while learning new concepts) [13, 23] and overfitting issues (overfitting to few examples and thus poor generalization) [24, 43]. To bypass this convoluted mixture of issues that hinder flexible adaptation of models, most previous works [20, 32, 47, 48, 51, 55] fix the learned representation after training it on base (old) classes and employ a non-parametric classifier, using the feature-average class prototype representation [39]. However, such formulation leads to heavy reliance on the representation acquired through the optimization of softmax cross-entropy (SCE) loss on base classes, which often leads to collapsed intra-class representation [31] and poor transferability to new classes [21], as shown in Fig. 1a. Therefore, in this paper, we mainly focus on exploring effective representation learning methods, aiming to strike a better balance between discriminability on base classes and transferability to new classes. + +There have been great advances, particularly self-supervised contrastive (SSC) learning [8,18], in the representation learning field to improve the transferability of the learned representation to downstream tasks. Some works [7,21] have attributed the strong transferability of SSC learning to 'spreading out' of intra-class features. Such representation places more emphasis on low- and mid-level features, which can be effectively transferred to and shared by new tasks. Kornblith et al. [25] have also found the relationship between the temperature of SCE loss and the spread of features, suggesting that lower temperature leads to better transferability. As illustrated in Fig. 1b, we observe that the joint optimization of SCE loss with low temperature and SSC loss encourages the spread of features and better transferability to new classes. + +Despite the foregoing, we observe that the previous methods for improving transferability are not enough to find a good representation for FSCIL; in fact, they harm the performance on base classes. Our experimental analysis suggests that excessive feature spread is very detrimental to base classes in the context of FSCIL because it hinders the feature-average class prototype from effectively representing its corresponding class, as demonstrated in Fig. 1b. Hence, it's crucial to develop a representation learning method tailored specifically for the FCSIL problem, particularly addressing the unique challenge of simultaneously achieving discriminability on seen classes and transferability to unseen classes. + +In this work, with the support from our experimental findings and information-bottleneck-theory-based analysis, we argue that the inter-class distance greatly affects the trade-off between the discriminability and transferability of the learned representation in the FSCIL problem. We find that the degraded discriminability due to the spread of features can be greatly regulated by reducing inter-class distance. Moreover, we discover that reducing the inter-class variability leads to enhancing the information bottleneck trade-off. Thus, we claim that attempting to ensure the spread of features within a compressed representation space promotes learning minimal yet intrinsic task-related information. + +Based on our analysis, in contrast to common beliefs and practices of previous FSCIL methods [20, 40, 47, 51] that have attempted to increase the inter-class distance, we propose to decrease it. Incorporating SCE loss with lower temperature, SSC loss, and inter-class distance minimization, our new objective enables the learned representation to strike a better balance between the discriminability and transferability, as illustrated in Fig. 1c. With the simple yet seemingly counter-intuitive idea of bringing classes closer (hence the name CLOSER), the proposed method demonstrates outstanding performance, suggesting a new promising research avenue regarding representation learning for FSCIL. + +# 2 Related Works + +Few-Shot Class-Incremental Learning (FSCIL). Towards the development of real-world artificial intelligence systems, Tao et al. [41] have initially introduced FSCIL problem, subsequently fostering numerous studies in the field [1,6,9-11, 36,53]. Most of the works [20,22,28,32,40,47,48,52,54] bypass both catastrophic + +forgetting and overfitting issues by fixing the feature extractor trained on base classes and employing a non-parametric classifier using class prototypes [39]. Hence, recent works have focused on representation learning, where the common approach is to encourage greater separation between base classes to reserve the representation space for future new classes [20, 40, 47, 51]. However, we argue that increasing inter-class distance suppresses the acquisition of shared features among classes, which could be relevant to new classes. Thus, contrary to the prevailing belief, we suggest learning representation with an inter-class distance minimization and theoretically and empirically prove its effectiveness in improving the discriminability-transferability trade-off. Similarly, Zou et al. [55] emphasize the importance of learning shareable features among classes, which they propose to achieve with a negative margin [27]. We note that a negative margin is more related to representation spreading rather than our idea of reducing inter-class distance, as discussed in supplementary material Sec. S4. + +Transferable Representation Learning. The pursuit of representations that can be effectively transferred to downstream tasks has gained significant attention in recent years. The early works have focused on supervised training on ImageNet dataset [35] and the way to transfer the knowledge to other tasks [5,15,34]. Kornblith et al. [25] find that learned representations with better discriminability on a source dataset tend to exhibit degraded transferability to downstream tasks. To obtain a transferable representation, Liu et al. [27,55] suggest incorporating a negative margin in the softmax cross-entropy loss to promote feature sharing among classes rather than solely focusing on discriminative features. In parallel, several methods have reported the strong transferability of representation yielded by self-supervised contrastive (SSC) learning [8,18]. Islam et al. [21] claim that the enhanced transferability acquired by SSC learning can be attributed to the spread of representation, implying that a network learns more fine-grained and shareable knowledge among tasks [7]. From the perspective of information bottleneck theory, Cui et al. [12] claim that over-compression of mutual information between inputs and latent representations can prevent a network from learning features beneficial for downstream tasks. In this paper, we argue that, along with the representation spreading loss, directly regulating inter-class distance encourages a network to learn discriminative yet shareable features among classes, which can be beneficial for both base and new classes. Through the lens of information bottleneck theory, we theoretically analyze the connection between the joint optimization objective and the information bottleneck trade-off, supporting our claim. + +# 3 Proposed Method + +# 3.1 Background: Problem Formulation + +Following the standard formulation of few-shot class incremental learning (FS-CIL) [41], we assume a sequence of training sessions with the corresponding datasets $\{\mathcal{D}^{(0)},\mathcal{D}^{(1)},\dots ,\mathcal{D}^{(T)}\}$ . $\mathcal{D}^{(t)}$ consists of training examples $\pmb{x}_i^{(t)}$ with its class labels $y_{i}^{(t)}\in \mathcal{C}^{(t)}$ (for simplicity, we will exclude the superscript), where $\mathcal{C}^{(t)}$ + +is the set of classes in its respective dataset $\mathcal{D}^{(t)}$ and $\mathcal{C}^{(s)}\cap \mathcal{C}^{(t)} = \emptyset$ for $s\neq t$ (each dataset has its own distinct classes without overlap). In the first session (a.k.a. base session) with its dataset $\mathcal{D}^{(0)}$ , there is assumed to be a large number of classes available with an abundant amount of training data for each class. In subsequent sessions (a.k.a. incremental sessions) with its datasets $\mathcal{D}^{(\geq 1)}$ , it is assumed that each dataset has few training examples for each class. In particular, FSCIL is said to have a $N$ -way $K$ -shot setting when each incremental session has $N$ classes with $K$ examples for each class. At each $t$ -th training session, only its corresponding dataset $\mathcal{D}^{(t)}$ is accessible for training. After each $t$ -th session, the evaluation is performed on all previously seen classes $\mathcal{C}^{(\leq t)}$ using test datasets $\mathcal{D}_{test}^{(\leq t)}$ , which consists of unseen new examples with the class label set $\mathcal{C}^{(\leq t)}$ . + +# 3.2 Background: Baseline + +Let a classification network consist of a feature extractor $f_{\theta}(\cdot)$ and a classification layer with its weights $\phi$ . The training objective of the base session is simply the softmax cross-entropy (SCE) loss, with the cosine similarity $\mathrm{sim}(\cdot, \cdot)$ between features $z = f_{\theta}(x)$ and classification weights $\phi$ as logits [14]: + +$$ +\mathcal {L} _ {\mathrm {c e}} = \frac {1}{B} \sum_ {j = 1} ^ {B} - \log \frac {\exp \left(\frac {1}{\tau} \operatorname {s i m} \left(z _ {j} , \phi_ {y _ {j}}\right)\right)}{\sum_ {k = 1} ^ {| \mathcal {C} ^ {(0)} |} \exp \left(\frac {1}{\tau} \operatorname {s i m} \left(z _ {j} , \phi_ {k}\right)\right)}, \tag {1} +$$ + +where $B$ is the batch size and $\tau$ is the temperature parameter. Incrementally updating weights with few examples in incremental sessions can make the network vulnerable to both catastrophic forgetting and overfitting. To bypass the problems, several works [20,48] suggest minimizing weight updates by freezing the feature extractor after the base session and using feature-average class prototype representation [39]. Specifically, after the base session, trained $\phi$ is replaced with class prototypes, a process we refer to as classifier replacement (CR), and new-class prototypes are obtained in the incremental sessions. The $i$ -th class prototype is acquired by averaging the features of its training samples: + +$$ +\phi_ {i} ^ {P} = \frac {1}{N _ {c _ {i}}} \sum_ {\left(\boldsymbol {x} _ {j}, y _ {j}\right) \in \mathcal {D} ^ {\left(\geq 0\right)}} \mathbb {1} _ {[ y _ {j} = i ]} f _ {\boldsymbol {\theta}} \left(\boldsymbol {x} _ {j}\right), \tag {2} +$$ + +where $N_{c_i}$ is the number of training samples for the $i$ -th class and $\mathbb{1}_{[\cdot]}$ indicates 1 if the subscript condition is True and 0 otherwise. For an input $\pmb{x}$ , the unnormalized log probability for the $i$ -th class is computed by $\mathrm{sim}(f_{\pmb{\theta}}(\pmb{x}), \phi_i^P)$ . + +Although this baseline bypasses the forgetting and overfitting issues, it heavily relies on the quality of the representation trained solely on the base classes. Thus, the main focus of this paper is to investigate the important factors that influence representation learning for FSCIL and strategies to improve them. + +# 3.3 Transferability, feature spread, and its adverse effects on FSCIL + +As shown in Fig. 1a, the baseline method exhibits a narrow intra-class distribution, widely perceived as representation collapse [31]. Recent studies [7,21,46] have + +![](images/eff9e4754620398410349535a5b9c6c336d8cec809f4950832047dcb3fa4346a.jpg) +Fig. 2: The impact of the spread of representation. Stronger emphasis on self-supervised contrastive loss (larger $\lambda_{\mathrm{ssc}}$ ) and low temperature (skyblue) enhances the new-class performance $A_N$ (left), but at the expense of base-class performance $A_B$ (center). The reduced base-class performance is mainly attributed to the excessive intra-class variation, adversely affecting the class prototype representation (right). The experiments are conducted on CIFAR100 dataset. + +revealed that the collapsed representation shows poor transferability due to the loss of shareable low- and mid-level features that new classes can benefit from. As such, they suggest joint optimization with a self-supervised contrastive (SSC) task [8,18], which promotes the spread of features and thus the sharing of features among classes. SSC learning optimizes infoNCE loss [30], regarding an augmented view from a query image as positive and the other images as negative samples. The SSC loss for a positive pair $(i,j)$ is: + +$$ +\mathcal {L} _ {\mathrm {s s c}} ^ {(i, j)} = - \log \frac {\exp \left(\frac {1}{\tau} \operatorname {s i m} \left(\boldsymbol {z} _ {i} , \boldsymbol {z} _ {j}\right)\right)}{\sum_ {k = 1} ^ {B} \mathbb {1} _ {[ k \neq i ]} \exp \left(\frac {1}{\tau} \operatorname {s i m} \left(\boldsymbol {z} _ {i} , \boldsymbol {z} _ {k}\right)\right)}, \tag {3} +$$ + +where $B$ is the number of samples, including augmented images, and $z_{j}$ is the feature from an augmented view of $\pmb{x}_{i}$ . The loss is averaged over all positive pairs. + +In parallel, several studies have proposed to reduce the temperature and margin parameters in the softmax cross-entropy loss as another way to encourage feature sharing [25,27,55]. Our empirical analysis in Sec. S4 of the supplementary material demonstrates that the temperature parameter affects the transferability more than the margin. Thus, we employ a low-temperature parameter in the SCE loss to further enhance the transferability of learned representations. Indeed, we observe that the combination of $\mathcal{L}_{\mathrm{ssc}}$ and $\mathcal{L}_{\mathrm{ce}}$ with low $\tau$ results in better new-class performance in the FSCIL problem due to a larger spread of features, as demonstrated in Fig. 1b and the left figure of Fig. 2. + +However, as illustrated in the center figure in Fig. 2, we observe that the pursuit of better transferability results in a degradation in discriminability on the base classes. While exploring the dilemma, we discover that the performance decline on the base classes mainly arises from the base class classifier replacement (CR) strategy, which is introduced in Sec. 3.2. The figure on the right in Fig. 2 shows the accuracy on the base classes subsequent to the base session training $(A_B^0)$ , without considering the new classes, both before and after CR, denoted by $A_B^{0,\mathrm{bef}}$ and $A_B^{0,\mathrm{aft}}$ , respectively. Before CR, $A_B^{0,\mathrm{bef}}$ exhibits a relatively high value and a small variance as $\tau$ and $\lambda_{\mathrm{ssc}}$ vary (indicated by the solid lines). Conversely, + +![](images/3d3e3139b96bcdb2dfe4a3a8e18f54cc7c92b4dcd3437e3089064d87b857defc.jpg) +Fig. 3: Effect of minimizing inter-class distance. As the weight of $\mathcal{L}_{\mathrm{inter}}$ , denoted by $\lambda_{\mathrm{inter}}$ , increases, the performance on the new classes increases (skyblue) and the performance loss on the base classes induced by CR is greatly alleviated (purple). The experiments are conducted on CUB200 dataset. + +$A_B^{0,\mathrm{aft}}$ is considerably lower than $A_B^{0,\mathrm{bef}}$ with a relatively higher variance (indicated by the dotted lines). Considering that CR is introduced to bypass overfitting and catastrophic forgetting in the FSCIL problem, these findings suggest that the previous methods for improving the transferability of representations are not effective in enhancing the trade-off between transferability and discriminability in the context of the FSCIL problem. Thus, it is essential to develop a representation learning method tailored for the FSCIL problem. + +# 3.4 Inter-Class Distance Matters + +As discussed in Sec. 3.3, learning shareable features through representation spreading proves advantageous for transferability. Yet, it also harms discriminability in the process of CR. Based on the observation in Fig. 1b, we hypothesize that such a dilemma appears to arise from a large inter-class distance since the representation spreading could push features into the extensive inter-class space, which may dilute the information on the base classes. Furthermore, we argue that the large inter-class distance may impede effective feature sharing among classes, undermining the transferability of learned representations. Consequently, to retain the knowledge on base classes while promoting effective feature sharing, we introduce a novel loss function that minimizes the inter-class distance: + +$$ +\mathcal {L} _ {\text {i n t e r}} = - \frac {1}{\sum_ {i = 1} ^ {B} \sum_ {j > i} ^ {B} \mathbb {1} _ {[ y _ {i} \neq y _ {j} ]}} \sum_ {i = 1} ^ {B} \sum_ {j > i} ^ {B} \mathbb {1} _ {[ y _ {i} \neq y _ {j} ]} \operatorname {s i m} \left(\boldsymbol {z} _ {i}, \boldsymbol {z} _ {j}\right). \tag {4} +$$ + +Fig. 1c displays that the spread of intra-class features is well regulated by applying $\mathcal{L}_{\mathrm{inter}}$ and Fig. 3 demonstrates the performance decline from CR is alleviated by minimizing $\mathcal{L}_{\mathrm{inter}}$ (indicated by the purple line). Moreover, the results indicated by the skyblue line show that reducing inter-class distance improves the performance on the new classes, corroborating our hypothesis. + +Our assertion initially seems counter-intuitive, diverging from the common belief of prior works [20,40,47,51] that maximizing inter-class distance may be + +![](images/86efe6a231a80e37d7d61896f9fa4f656b04acdd899e768562c8e6c44e592204.jpg) +(a) + +![](images/3c02a39d0008eb38d7b3606114271a6386a2c10a75372f41fe38e7687d706822.jpg) +(b) +Fig. 4: (a) Sanity test for $\mathcal{T}(f_{\theta})$ : $\mathcal{T}(f_{\theta})$ has a positive correlation with the performance on the new classes. Each data point is obtained by different configurations of $\tau$ and $\lambda_{\mathrm{ssc}}$ (without $\mathcal{L}_{\mathrm{inter}}$ ). (b), (c) Relationship between inter-class distance, $\mathcal{T}(f_{\theta})$ , and $A_N$ : Integrated with the representation spreading, reducing inter-class distance encourages better transferability (red points). However, the tendency is broken when reducing inter-class distance without representation spreading (blue points). Please refer to Sec. 3.5 for theoretical support for these observations. The dots with greater transparency correspond to smaller $\lambda_{\mathrm{inter}}$ , ranging from 0 to 1 with intervals of 0.1. We set $\lambda_{\mathrm{ssc}}$ as 0.1 when it is used. The experiments are conducted on CIFAR100 dataset. + +![](images/4d42ebafd5b7bf19f328c489499f67d10df6b71943c7ce0ca18c358588847db1.jpg) +(c) + +beneficial for reserving representation space for future new classes. To further validate the efficacy of reducing inter-class distance with respect to transferability, we propose a measure to quantify how new class samples are distinct from base-class representations. We define the measure as the averaged relative angular distance between a new-class sample and its nearest base-class prototype with respect to the averaged angular distance among all base-class prototype pairs: + +$$ +\mathcal {T} (f _ {\boldsymbol {\theta}}) = \frac {\frac {1}{| \mathcal {D} _ {t e s t} ^ {(> 0)} |} \sum_ {(\boldsymbol {x} _ {j} , y _ {j}) \in \mathcal {D} _ {t e s t} ^ {(> 0)} i} \min _ {i} \angle (\boldsymbol {z} _ {j} , \phi_ {b a s e , i} ^ {P})}{\sum_ {j = 1} ^ {| \mathcal {C} ^ {(0)} |} \sum_ {k > j} ^ {| \mathcal {C} ^ {(0)} |} \angle \left(\phi_ {j} ^ {P} , \phi_ {k} ^ {P}\right) / \left(^ {| \mathcal {C} ^ {(0)} |} _ {2}\right)}, \tag {5} +$$ + +where $\phi_{base,i}^{P}$ and $\angle (\cdot ,\cdot)$ indicate the $i$ -th base-class prototype and the angular distance between two input vectors, respectively. We introduce the denominator to normalize the varying sizes of the representation space depending on methods. If the representation produced by $f_{\theta}$ has a distinguishable representation for new classes, $\mathcal{T}(f_{\theta})$ would be large. For the sanity test, we check the relationship between $\mathcal{T}(f_{\theta})$ and the accuracy of the new class, which is shown in Fig. 4a. + +Using this measure, we analyze the relationship between inter-class distance, the spread of features, and the transferability of learned representation. To do so, we train the feature extractor using $\mathcal{L}_{\mathrm{ce}}$ with a low temperature, $\mathcal{L}_{\mathrm{ssc}}$ , and $\mathcal{L}_{\mathrm{inter}}$ with varying loss weights for $\mathcal{L}_{\mathrm{ssc}}$ and $\mathcal{L}_{\mathrm{inter}}$ , denoted by $\lambda_{\mathrm{ssc}}$ and $\lambda_{\mathrm{inter}}$ , respectively. Figs. 4b and 4c show the relationship between the inter-class distance, $\mathcal{T}(f_{\theta})$ , and the performance on the new classes. The results demonstrate that the joint optimization of $\mathcal{L}_{\mathrm{ssc}}$ and $\mathcal{L}_{\mathrm{inter}}$ leads to an increase in both $\mathcal{T}(f_{\theta})$ and $A_N$ , indicating that smaller inter-class distances lead to more discriminability between base classes and new classes. The analysis corroborates our seemingly counter-intuitive claim that the closer classes are, the better. + +In summary, we have observed that the spread of representation achieved by employing a low temperature in the SCE loss and self-supervised contrastive loss is advantageous for learning transferable representation. However, spread representation itself cannot address the trade-off between transferability to new classes and discriminability on base classes in the context of the FSCIL problem. Our analysis demonstrates that decreasing inter-class distance enhances discriminability by regularizing the intra-class spread and improves the transferability by promoting the effective learning of shareable information among classes when combined with the feature-spread-encouraging loss. Consequently, our final loss function is the combination of cross-entropy loss with a lower temperature parameter, self-supervised contrastive loss, and inter-class distance minimizing loss: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {c e}} + \lambda_ {\mathrm {s s c}} \mathcal {L} _ {\mathrm {s s c}} + \lambda_ {\mathrm {i n t e r}} \mathcal {L} _ {\mathrm {i n t e r}}. \tag {6} +$$ + +# 3.5 Information Bottleneck Theory Perspective + +In this subsection, we provide theoretical support for the proposed representation learning objective in Eq. (6) from the perspective of the information bottleneck (IB) theory [3, 37, 42]. The IB theory describes the goal of representation learning as finding a good trade-off between complexity and accuracy through finding minimal information from inputs necessary to preserve maximal information about the targets. As discussed by Cui et al. [12], we consider the objective of IB theory to be closely related to learning transferable representations since while achieving the objective, a network could be guided to learn intrinsic knowledge rather than task-irrelevant shortcuts [17], leading to better transferability. + +For an image classification task, let $X \in \mathbb{R}^{h \times w \times 3}$ , $Y \in \mathbb{R}^C$ , and $Z = \frac{f_{\theta}(X)}{\|f_{\theta}(X)\|} \in \mathbb{R}^d$ denote the input, label, and normalized latent representation variables, respectively, where $h$ and $w$ are the spatial sizes of the image, $C$ is the number of classes, and $d$ is the dimension of the latent representations. The aforementioned trade-off has been formulated as the following objective: + +$$ +\max I (Y; Z) - \beta I (X; Z), \tag {7} +$$ + +where $I(\cdot ;\cdot)$ indicates the mutual information between two random variables and $\beta > 0$ is a Lagrange multiplier. In this work, we consider an alternative trade-off objective, $\max \frac{I(Y;Z)}{\beta I(X;Z)}$ , which is adopted by several works [29,38]. After omitting $\beta$ for simplicity and modest assumptions, we derive the following lower bound of the IB trade-off objective: + +$$ +\frac {I (Y ; Z)}{I (X ; Z)} \geq 1 - \frac {d \cdot \log (2 \pi e) + \frac {1}{C} \sum_ {i = 1} ^ {C} \log | \Sigma_ {W _ {i}} |}{d \cdot \log (2 \pi e) + \log | \Sigma_ {T} |}, \tag {8} +$$ + +where $\Sigma_{W_i}$ and $\Sigma_T$ indicate the covariance matrices of $Z$ within the $n$ -th class and overall classes, respectively. As proven by Lemma S1 in the supplementary material, both the numerator and denominator in the fractional term of the lower bound in Eq. (8) are negative, leading to the following theorem: + +Table 1: 10-way 5-shot incremental learning results on CUB200. + +
MethodAcc. in each session (%)PD (%) ↓
012345678910
Baseline79.9276.2373.1869.4567.8365.7464.5463.3361.5661.2760.1019.83
TOPIC [41]68.6862.4954.8149.9945.2541.4038.3535.3632.2228.3126.2642.42
F2M [36]81.0778.1675.5772.8970.8668.1767.0165.2663.3661.7660.2620.81
CEC [48]75.8571.9468.5063.5062.4358.2757.7355.8154.8353.5252.2823.57
IDLVQ-C [6]77.3774.7270.2867.1365.3463.5262.1061.5459.0458.6857.8119.56
ALICE [33]77.4072.7070.6067.2065.9063.4062.9061.9060.5060.6060.1017.30
CLOM [55]79.5776.0772.9469.8267.8065.5663.9462.5960.6260.3459.5819.99
Entropy Reg. [28]75.9072.1468.6463.7662.5859.1157.8255.8954.9253.5852.3923.51
LIMIT [52]75.8973.5571.9968.1467.4263.6162.4061.3559.9158.6657.4118.48
MetaFSCIL [11]75.9072.4168.7864.7862.9659.9958.3056.8554.7853.8252.6423.26
FACT [51]75.9073.2370.8466.1365.5662.1561.7459.8358.4157.8956.9418.96
S3C [22]80.6277.5573.1968.5468.0564.3363.5862.0760.6159.7958.9520.83
SAVC [40]81.8577.9274.9570.2169.9667.0266.1665.3063.8463.1562.5019.35
NC-FSCIL [47]80.4575.9872.3070.2868.1765.1664.4363.2560.6660.0159.4421.01
GKEAL [54]78.8875.6272.3268.6267.2364.2662.9861.8960.2059.2158.6720.21
CABD [49]79.1275.3772.8069.0567.5365.1264.0063.5161.8761.4760.9318.19
CLOSER (Ours)79.4075.9273.5070.4769.2467.2266.7365.6964.0064.0263.5815.82
+ +Table 2: 5-way 5-shot incremental learning results on CIFAR100. + +
MethodAcc. in each session (%)PD (%) ↓
012345678
Baseline72.9368.4664.2660.1556.5353.6051.5149.1947.0925.84
ERDIL [16]73.6268.2265.1461.8458.3555.5452.5150.1648.2325.39
CEC [48]73.0768.8865.2661.1958.0955.5753.2251.3449.1423.93
CLOM [55]74.2069.8366.1762.3959.2656.4854.3652.1650.2523.95
Entropy Reg. [28]74.470.266.5462.5159.7156.5854.5252.3950.1424.26
FACT [51]74.6072.0967.5663.5261.3858.3656.2854.2452.1022.50
LIMIT [52]73.8172.0967.8763.8960.7057.7755.6753.5251.2322.58
MetaFSCIL [11]74.5070.1066.8462.7759.4856.5254.3652.5649.9724.53
GKEAL [54]74.0170.4567.0163.0860.0157.3055.5053.3951.4022.61
CLOSER (Ours)75.7271.8368.3264.6261.9159.2557.5355.4353.3222.40
+ +Theorem 1. The lower bound of $\frac{I(Y;Z)}{I(X;Z)}$ in Eq. (8) is a monotonically increasing function of $|\Sigma_{W_i}|$ and a monotonically decreasing function of $|\Sigma_T|$ . + +The detailed derivations are presented in Sec. S1 of the supplementary material. + +Generally, the determinant of a covariance matrix is correlated with the spread or variability of variables across all dimensions. Thus, Theorem 1 suggests that the IB trade-off can be enhanced by increasing the intra-class variability $(\mathcal{L}_{\mathrm{ce}}$ with a low temperature and $\mathcal{L}_{\mathrm{ssc}})$ while suppressing the overall representation space $(\mathcal{L}_{\mathrm{inter}})$ , supporting the proposed learning objective in Eq. (6) and our claims for inter-class distance minimization. Intuitively, our objective functions encourage greater overlap in representations among different classes, promoting more shareable and compact representations. Our theoretic analysis also underpins the findings in Fig. 4 that reducing inter-class distance without representation spreading loss is observed to result in less transferable representations. Since compressing the overall feature space diminishes the intra-class variability, the IB trade-off may decrease, leading to degraded transferability. + +Table 3: 5-way 5-shot incremental learning results on miniImageNet. + +
MethodAcc. in each session (%)PD (%) ↓
012345678
Baseline72.2767.4663.2659.7356.5653.5350.9048.9347.2625.01
TOPIC [41]61.3150.0945.1741.1637.4835.5232.1929.4624.4236.89
F2M [36]67.2863.8060.3857.0654.0851.3948.8246.5844.6522.63
CEC [48]72.0066.8362.9759.4356.7053.7351.1949.2447.6324.37
IDLVQ-C [6]64.7759.8755.9352.6249.8847.5544.8343.1441.8422.93
Subspace Reg. [2]80.3773.7668.3664.0760.3656.2753.1050.4547.5532.83
ALICE [33]80.6070.6067.4064.5062.5060.0057.8056.8055.7024.90
CLOM [55]73.0868.0964.1660.4157.4154.2951.5449.3748.0025.08
Entropy Reg. [28]71.8467.1263.2159.7757.0153.9551.5549.5248.2123.63
LIMIT [52]72.3268.4764.3060.7857.9555.0752.7050.7249.1423.13
MetaFSCIL [11]72.0467.9463.7760.2957.5855.1652.9050.7949.1922.85
FACT [51]72.5669.6366.3862.7760.6057.3354.3452.1650.4922.07
GKEAL [54]73.5968.9065.3362.2959.3956.7054.2052.5951.3122.28
CABD [49]74.6570.4366.2962.7760.7557.2454.7953.6552.2222.43
CLOSER (Ours)76.0271.6167.9964.6961.7058.9456.2354.5253.3322.69
+ +# 4 Experiments + +# 4.1 Experimental Details + +Dataset. Following the benchmark settings proposed by Tao et al. [41], we evaluate the proposed method on CIFAR100 [26], miniImageNet [43], and CUB200 [44]. For CIFAR100 and miniImageNet, the total number of classes is 100: 60 base classes and 40 new classes. The 40 new classes are split into 8 disjoint sets of 5 classes, each of which is sequentially provided with 5 training examples per class (5-way 5-shot) in each incremental session. As for CUB200, there total number of classes is 200, with 100 base classes and 100 new classes. 100 new classes are split into 10 disjoint sets of 10 classes, each of which is sequentially provided with 5 training examples per class (10-way 5-shot) in each incremental session. + +Implementation. Following Zhang et al. [48], we use ResNet-20 [19] for CIFAR100 experiments and ResNet-18 [19] for both miniImageNet and CUB200 experiments. We follow the conventions to use the ResNet-18 model pre-trained on the ImageNet dataset [35] for CUB200. We set the mini-batch size to 128, 128, and 256 for CIFAR100, miniImageNet, and CUB200 experiments, respectively. The temperature parameter $\tau$ for the baseline method is 1/16 and 'low temperature' indicates $\tau = 1 / 32$ . We set $\lambda_{\mathrm{ssc}}$ as 0.1,0.1, and 0.01 for CIFAR100, miniImageNet, and CUB200, respectively, and $\lambda_{\mathrm{inter}}$ as 1, 0.5, and 1.5 for CIFAR100, miniImageNet, and CUB200, respectively. These hyper-parameters are searched via validation using synthesized validation sets. For mutual information estimation, we adopt MINE [4] method and implement it based on the open-source code2. More details are provided in the supplementary material. + +Evaluation. We use the accuracy on base $(A_B)$ , new $(A_N)$ , and the whole classes $(A_W)$ as metrics to assess the discriminability, transferability, and the trade-off between them in learned representations. Additionally, we use the performance drop (PD) between the accuracy at the end of the base session (session 0) and + +Table 4: Ablation studies on CIFAR100. + +
lowτL SSCL interAB(%)AN(%)AW(%)PD(%)
XXX70.1320.9550.4624.81
X✓X69.9522.8051.0924.59
XX✓69.9520.0549.9926.78
X✓✓71.4322.8351.9925.26
✓XX68.3324.1050.6423.23
✓✓X66.5825.0849.9823.15
✓X✓70.1722.4051.0625.27
✓✓✓70.7227.2353.3222.40
+ +the accuracy at the last incremental session to evaluate the degree to which old knowledge is forgotten and new knowledge is learned simultaneously. All experimental results of CLOSER are obtained by averaging 3 trials. + +# 4.2 Comparison with the Existing Works + +We compare the proposed method, dubbed CLOSER, with prior arts on CUB200 (Tab. 1), CIFAR100 (Tab. 2), and miniImageNet (Tab. 3). We observe that CLOSER achieves state-of-the-art performance on both CUB200 and CIFAR100 datasets, surpassing the results of previous methods by a large margin with respect to $A_W$ and PD. With miniImageNet, the proposed method exhibits substantially higher $A_W$ than the method with the lowest PD and achieves lower PD than the method with the highest $A_W$ , which means that the proposed method achieves a better balance between the performance on base and the new classes. It is worth noting that CLOSER shows such outstanding performance without any assistance from the storage of previous samples (F2M, ERDIL, IDLVQ-C, and CABD), additional computational modules (CEC, CLOM, NC-FSCIL, SAVC, and MetaFSCIL), and test-time data augmentation (S3C and SAVC), suggesting the critical importance of learning effective representations in FSCIL. + +# 4.3 Ablation Studies + +To verify the efficacy of the individual components in the proposed method, we perform ablation studies, which are shown in Tab. 4. The increase in $A_N$ when employing a lower temperature in the softmax cross-entropy loss (low $\tau$ ) or a self-supervised contrastive loss ( $\mathcal{L}_{\mathrm{ssc}}$ ) confirms the advantage of feature spread in enhancing the transferability of representations. While transferability can be greatly improved by using low $\tau$ and $\mathcal{L}_{\mathrm{ssc}}$ , a substantial decline in base-class performance $A_B$ is observed, as discussed in Sec. 3.4. The result from the case where all components are utilized demonstrates that this issue can be effectively resolved by minimizing inter-class distance. Furthermore, as discussed in Sec. 3.4, reducing inter-class distance is observed to improve transferability, especially when used with the representation spreading methods. The comprehensive results of ablation studies confirm and support our claims. + +![](images/7926c9b824bf0fb31175b52982c9974883d8f2b5324cc38cbc6dfae80a1c2077.jpg) +(a) Base Classes + +![](images/1f37a6e31454da20ead1a27cc944ddff3b9e03901684c02bf77ff54dab72da1b.jpg) +(b) New Classes + +![](images/25abf52628bfec146c877a8eca8aefbc7eb396330ef40be9e114c7154530921d.jpg) +(c) Whole Classes + +![](images/ad9a42eb4ece99d9a9149c53b04dcf819804a07ef5a6f10d12d7bcd5ba109415.jpg) +Fig. 5: Information bottleneck trade-off analysis. We compare representations acquired by three different methods by assessing the information bottleneck (IB) tradeoff. 'RS' refers to representation spreading methods. We indicate the final models with black edges. The experiments are conducted on the CIFAR100 dataset. +(a) Baseline +Fig. 6: T-SNE visualization of learned representations. The blue and red points indicate the base and new class samples, respectively. We measure $\mathcal{T}(f_{\theta})$ to quantify the transferability of the learned representation. We conducted the experiments on CIFAR100 with reduced classes (20 base classes and 10 new classes) for better visualization. The classification results for each experiment are as follows: (a): $A_B = 78.85\%$ , $A_N = 13.50\%$ , $A_W = 57.07\%$ . (b): $A_B = 77.85\%$ , $A_N = 25.50\%$ , $A_W = 60.22\%$ . (c): $A_B = 78.88\%$ , $A_N = 28.80\%$ , $A_W = 62.18\%$ . + +![](images/600483079d620d23719251f778d174dca8165cc3fd04aeddbf8108cff1c7b505.jpg) +(b) Baseline $+\mathrm{RS}$ + +![](images/71f15c504e72c9c2765c5970978dd103bbb1c3c0b65a78424d590260aa1855eb.jpg) +(c) CLOSER (Ours) + +# 4.4 Information Bottleneck trade-off Analysis + +In Sec. 3.5, we show the connection between the proposed objective function and the information bottleneck (IB) trade-off. To validate our analysis, we measure the mutual information between representations and inputs $I(X;Z)$ and the mutual information between representations and targets $I(Y;Z)$ for the baseline, representation spread approach, and our method CLOSER, as shown in Fig. 5. In the figure, the more left (lower $I(X;Z)$ ) and upper (higher $I(Y;Z)$ ) regions imply better IB trade-off. Our proposed method CLOSER demonstrates to have found a better IB trade-off, especially when considering whole classes, including base and new classes. The results are encouraging in that CLOSER is able to find a better IB trade-off for all classes, when the feature extractor is only trained on base classes and fixed afterwards. The results demonstrate that CLOSER is effective in tackling a particularly difficult challenge of FSCIL: learning transferable and discriminative features from base classes. + +# 4.5 T-SNE Analysis + +For qualitative evaluation, we visualize the learned representation trained by different configurations of the proposed losses, which is illustrated in Fig. 6. The value of $\mathcal{T}(f_{\theta})$ in the right below in each figure is for measuring the transferability of representation. The baseline representation (a) shows the characteristics of the base class, represented by the features of the new classes mapped to those of the base classes, also indicated by the low value of $\mathcal{T}(f_{\theta})$ . Spreading of features (b) largely resolves the overfitting issues, exhibited by larger distances between new classes and base classes; i.e., the increase in $\mathcal{T}(f_{\theta})$ . Finally, $\mathcal{L}_{\mathrm{inter}}$ is used to compensate for the decreased performance on the base classes due to the spread representation. Reducing inter-class distance also enhances the separability between the new class samples and the clusters of base classes, as evidenced by the further increase in $\mathcal{T}(f_{\theta})$ . + +# 5 Limitations and Discussions + +Our work focuses on learning representation that is discriminative yet transferable to unseen classes to tackle the challenges of FSCIL. As such, our work does not consider an option of updating representation with new classes. Continual update of representation can improve the performance on new classes, however at the cost of sacrificing the old-class performance. Thus, it is a question of stability-plasticity dilemma, where our method focuses more on stability while improving plasticity through discriminative and transferable representation. Furthermore, similar to other works on FSCIL, our method is limited to classification tasks. But, we believe our work can have an impact on other domains, akin to the impact of representation spread methods, such as contrastive learning. + +# 6 Conclusion + +To tackle convoluted challenges in few-shot class-incremental learning (FSCIL), we focus on representation learning, which plays a crucial role. In contrast to previous FSCIL methods that have focused on maximizing the distance between classes, our experimental and theoretic analysis suggests that the closer classes are, the better for FSCIL, especially when feature sharing is encouraged between classes. Upon the analysis, we propose a simple, yet seemingly counter-intuitive idea: bring classes closer for a better transferability-discriminability Pareto front. Our experimental results and information-bottleneck-theory-based analysis suggest that our work can provide a promising research avenue. As such, we hope that our work will inspire future works and discussions in this research direction. + +Acknowledgments. This work was supported in part by the IITP grants [No. 2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University), No.2021-0-02068, and No.2023-0-00156], the NRF grant [No.2021M3A 9E4080782] funded by the Korean government (MSIT). + +# References + +1. Achituve, I., Navon, A., Yemini, Y., Chechik, G., Fetaya, E.: Gp-tree: A gaussian process classifier for few-shot incremental learning. In: ICML (2021) +2. Akyurek, A.F., Akyurek, E., Wijaya, D.T., Andreas, J.: Subspace regularizers for few-shot class incremental learning. In: ICLR (2022) +3. Alemi, A.A., Fischer, I., Dillon, J.V., Murphy, K.: Deep variational information bottleneck. In: ICLR (2017) +4. Belghazi, M.I., Baratin, A., Rajeshwar, S., Ozair, S., Bengio, Y., Courville, A., Hjelm, D.: Mutual information neural estimation. In: PMLR (2018) +5. Chatfield, K., Simonyan, K., Vedaldi, A., Zisserman, A.: Return of the devil in the details: delving deep into convolutional nets. In: BMVC (2014) +6. Chen, K., Lee, C.G.: Incremental few-shot learning via vector quantization in deep embedded space. In: ICLR (2021) +7. Chen, M.F., Fu, D.Y., Narayan, A., Zhang, M., Song, Z., Fatahalian, K., Ré, C.: Perfectly balanced: improving transfer and robustness of supervised contrastive learning. In: ICML (2022) +8. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML (2020) +9. Cheraghian, A., Rahman, S., Fang, P., Roy, S.K., Petersson, L., Harandi, M.: Semantic-aware knowledge distillation for few-shot class-incremental learning. In: CVPR (2021) +10. Cheraghian, A., Rahman, S., Ramasinghe, S., Fang, P., Simon, C., Petersson, L., Harandi, M.: Synthesized feature based few-shot class-incremental learning on a mixture of subspaces. In: ICCV (2021) +1. Chi, Z., Gu, L., Liu, H., Wang, Y., Yu, Y., Tang, J.: Meta桅: a meta-learning approach for few-shot class incremental learning. In: CVPR (2022) +2. Cui, Q., Zhao, B., Chen, Z.M., Zhao, B., Song, R., Liang, J., Zhou, B., Yoshie, O.: Discriminability-transferability trade-off: an information-theoretic perspective. In: ECCV (2022) +3. Delange, M., Aljundi, R., Masana, M., Parisot, S., Jia, X., Leonardis, A., Slabaugh, G., Tuytelaars, T.: A continual learning survey: Defying forgetting in classification tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2021) +4. Deng, J., Guo, J., Yang, J., Xue, N., Kotsia, I., Zafeiriou, S.: Arcface: additive angular margin loss for deep face recognition. In: CVPR (2019) +15. Donahue, J., Jia, Y., Vinyals, O., Hoffman, J., Zhang, N., Tzeng, E., Darrellin, T.: Decaf: a deep convolutional activation feature for generic visual recognition. In: ICML (2014) +6. Dong, S., Hong, X., Tao, X., Chang, X., Wei, X., Gong, Y.: Few-shot class-incremental learning via relation knowledge distillation. In: AAAI (2021) +7. Geirhos, R., Jacobsen, J.H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., Wichmann, F.A.: Shortcut learning in deep neural networks. Nature Machine Intelligence 2(11), 665-673 (2020) +8. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR (2020) +9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition In: CVPR (2016) +20. Hersche, M., Karunaratne, G., Cherubini, G., Benini, L., Sebastian, A., Rahimi, A.: Constrained few-shot class-incremental learning. In: CVPR (2022) + +21. Islam, A., Chen, C.F., Panda, R., Karlinsky, L., Radke, R., Feris, R.: A broad study on the transferability of visual representations with contrastive learning. In: ICCV (2021) +22. Kalla, J., Biswas, S.: S3c: self-supervised stochastic classifiers for few-shot class-incremental learning. In: ECCV (2022) +23. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., Hassabis, D., Clopath, C., Kumaran, D., Hadsell, R.: Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences of the United States of America (PNAS) (2017) +24. Koch, G., Zemel, R., Salakhutdinov, R.: Siamese neural networks for one-shot image recognition. In: ICMLW (2015) +25. Kornblith, S., Chen, T., Lee, H., Norouzi, M.: Why do better loss functions lead to less transferable features? In: NeurIPS (2021) +26. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images. University of Toronto (2009) +27. Liu, B., Cao, Y., Lin, Y., Li, Q., Zhang, Z., Long, M., Hu, H.: Negative margin matters: Understanding margin in few-shot classification. In: ECCV (2020) +28. Liu, H., Gu, L., Chi, Z., Wang, Y., Yu, Y., Chen, J., Tang, J.: Few-shot class-incremental learning via entropy-regularized data-free replay. In: ECCV (2022) +29. Ngampruetikorn, V., Schwab, D.J.: Information bottleneck theory of high-dimensional regression: relevancy, efficiency and optimality. In: NeurIPS (2022) +30. van den Oord, A., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) +31. Papyan, V., Han, X., Donoho, D.L.: Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences of the United States of America (PNAS) (2020) +32. Peng, C., Zhao, K., Wang, T., Li, M., Lovell, B.C.: Few-shot class-incremental learning from an open-set perspective. In: ECCV (2022) +33. Peng, C., Zhao, K., Wang, T., Li, M., Lovell, B.C.: Few-shot class-incremental learning from an open-set perspective. In: ECCV (2022) +34. Razavian, A.S., Azizpour, H., Sullivan, J., Carlsson, S.: Cnn features off-the-shelf: an astounding baseline for recognition. In: CVPRW (2014) +35. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV) (2015) +36. Shi, G., Chen, J., Zhang, W., Zhan, L.M., Wu, X.M.: Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. In: NeurIPS (2021) +37. Shwartz-Ziv, R., Tishby, N.: Opening the black box of deep neural networks via information (2017) +38. Slonim, N., Tishby, N.: Agglomerative information bottleneck. In: NIPS (1999) +39. Snell, J., Swersky, K., Zemel, R.: Prototypical networks for few-shot learning. In: NIPS (2017) +40. Song, Z., Zhao, Y., Shi, Y., Peng, P., Yuan, L., Tian, Y.: Learning with fantasy: semantic-aware virtual contrastive constraint for few-shot class-incremental learning. In: CVPR (2023) +41. Tao, X., Hong, X., Chang, X., Dong, S., Wei, X., Gong, Y.: Few-shot class-incremental learning. In: CVPR (2020) + +42. Tishby, N., Zaslavsky, N.: Deep learning and the information bottleneck principle. In: 2015 IEEE Information Theory Workshop (ITW). pp. 1-5 (2015). https://doi.org/10.1109/ITW.2015.7133169 +43. Vinyals, O., Blundell, C., Lillicrap, T., Kavukcuoglu, K., Wierstra, D.: Matching networks for one shot learning. In: NIPS (2016) +44. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The Caltech-UCSD Birds-200-2011 Dataset. Tech. rep., Caltech (2011) +45. Wang, H., Wang, Y., Zhou, Z., Ji, X., Gong, D., Zhou, J., Li, Z., Liu, W.: Cosface: large margin cosine loss for deep face recognition. In: CVPR (2018) +46. Xue, Y., Joshi, S., Gan, E., Chen, P.Y., Mirzasoleiman, B.: Which features are learnt by contrastive learning? on the role of simplicity bias in class collapse and feature suppression. In: ICML (2023) +47. Yang, Y., Yuan, H., Li, X., Lin, Z., Torr, P., Tao, D.: Neural collapse inspired feature-classifier alignment for few-shot class incremental learning. In: ICLR (2023) +48. Zhang, C., Song, N., Lin, G., Zheng, Y., Pan, P., Xu, Y.: Few-shot incremental learning with continually evolved classifiers. In: CVPR (2021) +49. Zhao, L., Lu, J., Xu, Y., Cheng, Z., Guo, D., Niu, Y., Fang, X.: Few-shot class-incremental learning via class-aware bilateral distillation. In: CVPR (2023) +50. Zheng, Y., Pal, D.K., Savvides, M.: Ring loss: convex feature normalization for face recognition. In: CVPR (2018) +51. Zhou, D.W., Wang, F.Y., Ye, H.J., Ma, L., Pu, S., Zhan, D.C.: Forward compatible few-shot class-incremental learning. In: CVPR (2022) +52. Zhou, D.W., Ye, H.J., Ma, L., Xie, D., Pu, S., Zhan, D.C.: Few-shot class-incremental learning by sampling multi-phase tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2022) +53. Zhu, K., Cao, Y., Zhai, W., Cheng, J., Zha, Z.J.: Self-promoted prototype refinement for few-shot class-incremental learning. In: CVPR (2021) +54. Zhuang, H., Weng, Z., He, R., Lin, Z., Zeng, Z.: Gkeal: gaussian kernel embedded analytic learning for few-shot class incremental task. In: CVPR (2023) +55. Zou, Y., Zhang, S., Li, Y., Li, R.: Margin-based few-shot class-incremental learning with class-level overfitting mitigation. In: NeurIPS (2022) \ No newline at end of file diff --git a/2024/CLOSER_ Towards Better Representation Learning for Few-Shot Class-Incremental Learning/images.zip b/2024/CLOSER_ Towards Better Representation Learning for Few-Shot Class-Incremental Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4df391eec66a613f1c9cff49520837172d1197ed --- /dev/null +++ b/2024/CLOSER_ Towards Better Representation Learning for Few-Shot Class-Incremental Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83d3ace6ac976622cadd7d22b62879221f2b1eb34326d1158b0445e5855e277f +size 557087 diff --git a/2024/CLOSER_ Towards Better Representation Learning for Few-Shot Class-Incremental Learning/layout.json b/2024/CLOSER_ Towards Better Representation Learning for Few-Shot Class-Incremental Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7bf14fcdf925ea6a83c04552e525eaf63d250a6d --- /dev/null +++ b/2024/CLOSER_ Towards Better Representation Learning for Few-Shot Class-Incremental Learning/layout.json @@ -0,0 +1,10928 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 132, + 112, + 482, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 112, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 112, + 482, + 148 + ], + "type": "text", + "content": "CLOSER: Towards Better Representation Learning for Few-Shot Class-Incremental Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 169, + 167, + 444, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 167, + 444, + 181 + ], + "spans": [ + { + "bbox": [ + 169, + 167, + 444, + 181 + ], + "type": "text", + "content": "Junghun Oh" + }, + { + "bbox": [ + 169, + 167, + 444, + 181 + ], + "type": "inline_equation", + "content": "^{*1}" + }, + { + "bbox": [ + 169, + 167, + 444, + 181 + ], + "type": "text", + "content": ", Sungyong Baik" + }, + { + "bbox": [ + 169, + 167, + 444, + 181 + ], + "type": "inline_equation", + "content": "^{*3}" + }, + { + "bbox": [ + 169, + 167, + 444, + 181 + ], + "type": "text", + "content": ", and Kyoung Mu Lee" + }, + { + "bbox": [ + 169, + 167, + 444, + 181 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 191, + 189, + 421, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 191, + 189, + 421, + 201 + ], + "spans": [ + { + "bbox": [ + 191, + 189, + 421, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 191, + 189, + 421, + 201 + ], + "type": "text", + "content": "Dept. of ECE&ASRI, " + }, + { + "bbox": [ + 191, + 189, + 421, + 201 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 191, + 189, + 421, + 201 + ], + "type": "text", + "content": "IPAI, Seoul National University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 216, + 201, + 397, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 201, + 397, + 213 + ], + "spans": [ + { + "bbox": [ + 216, + 201, + 397, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 216, + 201, + 397, + 213 + ], + "type": "text", + "content": "Dept. of Data Science, Hanyang University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 191, + 213, + 419, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 191, + 213, + 419, + 223 + ], + "spans": [ + { + "bbox": [ + 191, + 213, + 419, + 223 + ], + "type": "text", + "content": "{dh6dh,kyoungmu}@snu.ac.kr, dsybaik@hanyang.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 252, + 455, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 252, + 455, + 461 + ], + "spans": [ + { + "bbox": [ + 159, + 252, + 455, + 461 + ], + "type": "text", + "content": "Abstract. Aiming to incrementally learn new classes with only few samples while preserving the knowledge of base (old) classes, few-shot class-incremental learning (FSCIL) faces several challenges, such as overfitting and catastrophic forgetting. Such a challenging problem is often tackled by fixing a feature extractor trained on base classes to reduce the adverse effects of overfitting and forgetting. Under such formulation, our primary focus is representation learning on base classes to tackle the unique challenge of FSCIL: simultaneously achieving the transferability and discriminability of the learned representation. Building upon the recent efforts for enhancing the transferability, such as promoting the spread of features, we find that trying to secure the spread of features within a more confined feature space enables the learned representation to strike a better balance between the transferability and discriminability. Thus, in stark contrast to prior beliefs that the inter-class distance should be maximized, we claim that the CLOSER different classes are, the better for FSCIL. The empirical results and analysis from the perspective of information bottleneck theory justify our simple yet seemingly counterintuitive representation learning method, raising research questions and suggesting alternative research directions. The code is available here." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 472, + 453, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 472, + 453, + 494 + ], + "spans": [ + { + "bbox": [ + 159, + 472, + 453, + 494 + ], + "type": "text", + "content": "Keywords: Few-shot class incremental learning " + }, + { + "bbox": [ + 159, + 472, + 453, + 494 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 472, + 453, + 494 + ], + "type": "text", + "content": " Representation learning " + }, + { + "bbox": [ + 159, + 472, + 453, + 494 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 472, + 453, + 494 + ], + "type": "text", + "content": " Transferability" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 230, + 528 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 539, + 482, + 647 + ], + "type": "text", + "content": "Owing to its strong representation power, deep neural networks (DNNs) boast their outstanding performance across various fields. However, such feats require tremendous human effort and time to collect an immense amount of data with accurate annotation. The data hunger of DNNs poses a challenge, especially in dynamic real-world environments, where DNNs are required to learn new concepts with few examples while retaining previously learned concepts. To tackle the challenge, few-shot class-incremental learning (FSCIL) [41] aims to design artificial intelligence systems that can learn new classes with few examples while maintaining performance on previously seen classes." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 654, + 230, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 654, + 230, + 666 + ], + "spans": [ + { + "bbox": [ + 146, + 654, + 230, + 666 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 119, + 239, + 176 + ], + "blocks": [ + { + "bbox": [ + 161, + 119, + 239, + 176 + ], + "lines": [ + { + "bbox": [ + 161, + 119, + 239, + 176 + ], + "spans": [ + { + "bbox": [ + 161, + 119, + 239, + 176 + ], + "type": "image", + "image_path": "169eb00ebbd074501d782d682f0fc5b78e4b5ba58ea7c8213894f5de666ea8d7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 161, + 178, + 237, + 237 + ], + "blocks": [ + { + "bbox": [ + 161, + 178, + 237, + 237 + ], + "lines": [ + { + "bbox": [ + 161, + 178, + 237, + 237 + ], + "spans": [ + { + "bbox": [ + 161, + 178, + 237, + 237 + ], + "type": "image", + "image_path": "5106805d7831abf3b8fa0a8e3d8025779bee3b9805bf901aa0526eba4c6aa5d4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 161, + 239, + 240, + 297 + ], + "blocks": [ + { + "bbox": [ + 161, + 239, + 240, + 297 + ], + "lines": [ + { + "bbox": [ + 161, + 239, + 240, + 297 + ], + "spans": [ + { + "bbox": [ + 161, + 239, + 240, + 297 + ], + "type": "image", + "image_path": "41a84f1362785251bed844ad8dd6fd608e54f174b74fa575b1cfe0ce748458cd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 301, + 241, + 308 + ], + "lines": [ + { + "bbox": [ + 170, + 301, + 241, + 308 + ], + "spans": [ + { + "bbox": [ + 170, + 301, + 241, + 308 + ], + "type": "inline_equation", + "content": "A_{B} = 98.61\\%" + }, + { + "bbox": [ + 170, + 301, + 241, + 308 + ], + "type": "inline_equation", + "content": "A_{N} = 11.49\\%" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 191, + 316, + 222, + 323 + ], + "lines": [ + { + "bbox": [ + 191, + 316, + 222, + 323 + ], + "spans": [ + { + "bbox": [ + 191, + 316, + 222, + 323 + ], + "type": "text", + "content": "(a) Baseline" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 257, + 118, + 318, + 176 + ], + "blocks": [ + { + "bbox": [ + 257, + 118, + 318, + 176 + ], + "lines": [ + { + "bbox": [ + 257, + 118, + 318, + 176 + ], + "spans": [ + { + "bbox": [ + 257, + 118, + 318, + 176 + ], + "type": "image", + "image_path": "93456906fdc827458cce25de5e95384bfb3deb29fd3b2f7afe55a5acc42bbd1d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 258, + 178, + 317, + 237 + ], + "blocks": [ + { + "bbox": [ + 258, + 178, + 317, + 237 + ], + "lines": [ + { + "bbox": [ + 258, + 178, + 317, + 237 + ], + "spans": [ + { + "bbox": [ + 258, + 178, + 317, + 237 + ], + "type": "image", + "image_path": "c54bdd2f5e3287f43e2a5feee594b7c6cad4949889d4c9f4cba3daa3a98f1b32.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 257, + 239, + 317, + 298 + ], + "blocks": [ + { + "bbox": [ + 257, + 239, + 317, + 298 + ], + "lines": [ + { + "bbox": [ + 257, + 239, + 317, + 298 + ], + "spans": [ + { + "bbox": [ + 257, + 239, + 317, + 298 + ], + "type": "image", + "image_path": "34a0548c6ff7eb19059580bb6282bb2382c4b01dadf1a55240610f3b75f77299.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 301, + 325, + 308 + ], + "lines": [ + { + "bbox": [ + 255, + 301, + 325, + 308 + ], + "spans": [ + { + "bbox": [ + 255, + 301, + 325, + 308 + ], + "type": "inline_equation", + "content": "A_{B} = 87.42\\%" + }, + { + "bbox": [ + 255, + 301, + 325, + 308 + ], + "type": "inline_equation", + "content": "A_{N} = 33.34\\%" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 272, + 310, + 319, + 324 + ], + "lines": [ + { + "bbox": [ + 272, + 310, + 319, + 324 + ], + "spans": [ + { + "bbox": [ + 272, + 310, + 319, + 324 + ], + "type": "text", + "content": "(b) Baseline " + }, + { + "bbox": [ + 272, + 310, + 319, + 324 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 272, + 310, + 319, + 324 + ], + "type": "text", + "content": " sensation spreading" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 335, + 482, + 468 + ], + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 468 + ], + "type": "text", + "content": "Fig. 1: Visualization of representation trained on MNIST1. (a) Baseline [20, 48] exhibits great base-class discriminability (large inter-class distance) but weak transferability to the new classes (huge overlap between new and base classes leading to misclassification). (b) Baseline + representation spreading [7, 21, 25] benefits the new classes (less overlap with the base classes), while compromising base-class discriminability in the context of FSCIL (dispersed intra-class features leading to less accurate class representation with class prototypes). (c) CLOSER (Ours): Dispersing features in a narrowed feature space enhances both discriminability on the base classes (less deviation between intra-class features and class prototypes) and transferability to the new classes (even less overlap between the base and new classes). For instance, the 4 and 9 classes are not distinguishable in (b) and even less in (a), but CLOSER can yield representation that successfully discriminates them." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 340, + 118, + 400, + 177 + ], + "blocks": [ + { + "bbox": [ + 340, + 118, + 400, + 177 + ], + "lines": [ + { + "bbox": [ + 340, + 118, + 400, + 177 + ], + "spans": [ + { + "bbox": [ + 340, + 118, + 400, + 177 + ], + "type": "image", + "image_path": "e65ef29366675c9410b0b44779e95090877361cc8d3032f3259bc0231880d136.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 341, + 178, + 399, + 237 + ], + "blocks": [ + { + "bbox": [ + 341, + 178, + 399, + 237 + ], + "lines": [ + { + "bbox": [ + 341, + 178, + 399, + 237 + ], + "spans": [ + { + "bbox": [ + 341, + 178, + 399, + 237 + ], + "type": "image", + "image_path": "fc592dfb548ce2068b224d5307b7346f9237eb9f7a85d071f08a1c4d16830eb3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 341, + 239, + 399, + 298 + ], + "blocks": [ + { + "bbox": [ + 341, + 239, + 399, + 298 + ], + "lines": [ + { + "bbox": [ + 341, + 239, + 399, + 298 + ], + "spans": [ + { + "bbox": [ + 341, + 239, + 399, + 298 + ], + "type": "image", + "image_path": "c72aa6d7210f35c98de28e513eb0e36269705cef9f8371cffdb7160ded88641c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 301, + 410, + 308 + ], + "lines": [ + { + "bbox": [ + 340, + 301, + 410, + 308 + ], + "spans": [ + { + "bbox": [ + 340, + 301, + 410, + 308 + ], + "type": "inline_equation", + "content": "A_{R} = 92.47\\%" + }, + { + "bbox": [ + 340, + 301, + 410, + 308 + ], + "type": "inline_equation", + "content": "A_{N} = 42.14\\%" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 346, + 315, + 394, + 322 + ], + "lines": [ + { + "bbox": [ + 346, + 315, + 394, + 322 + ], + "spans": [ + { + "bbox": [ + 346, + 315, + 394, + 322 + ], + "type": "text", + "content": "(c) CLOSER (Ours)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 297, + 452, + 323 + ], + "lines": [ + { + "bbox": [ + 417, + 297, + 452, + 323 + ], + "spans": [ + { + "bbox": [ + 417, + 297, + 452, + 323 + ], + "type": "inline_equation", + "content": "A_{B}" + }, + { + "bbox": [ + 417, + 297, + 452, + 323 + ], + "type": "text", + "content": " : accuracy on base classes " + }, + { + "bbox": [ + 417, + 297, + 452, + 323 + ], + "type": "inline_equation", + "content": "A_N" + }, + { + "bbox": [ + 417, + 297, + 452, + 323 + ], + "type": "text", + "content": " : accuracy on new classes" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 420, + 127, + 447, + 220 + ], + "blocks": [ + { + "bbox": [ + 417, + 118, + 450, + 124 + ], + "lines": [ + { + "bbox": [ + 417, + 118, + 450, + 124 + ], + "spans": [ + { + "bbox": [ + 417, + 118, + 450, + 124 + ], + "type": "text", + "content": "Base Classes" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 420, + 127, + 447, + 220 + ], + "lines": [ + { + "bbox": [ + 420, + 127, + 447, + 220 + ], + "spans": [ + { + "bbox": [ + 420, + 127, + 447, + 220 + ], + "type": "image", + "image_path": "85c9814ec26e4ed9d35e51bde9a503604fa9d4368896730f2c204003733edc86.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 420, + 234, + 447, + 274 + ], + "blocks": [ + { + "bbox": [ + 418, + 224, + 448, + 231 + ], + "lines": [ + { + "bbox": [ + 418, + 224, + 448, + 231 + ], + "spans": [ + { + "bbox": [ + 418, + 224, + 448, + 231 + ], + "type": "text", + "content": "New Classes" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 420, + 234, + 447, + 274 + ], + "lines": [ + { + "bbox": [ + 420, + 234, + 447, + 274 + ], + "spans": [ + { + "bbox": [ + 420, + 234, + 447, + 274 + ], + "type": "image", + "image_path": "6e4546e4bafde37e6611b0a1a7dc42fe9d1333d3134b1ead2a86b0565ef4ad36.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 415, + 281, + 452, + 294 + ], + "lines": [ + { + "bbox": [ + 415, + 281, + 452, + 294 + ], + "spans": [ + { + "bbox": [ + 415, + 281, + 452, + 294 + ], + "type": "text", + "content": "Class prototypes" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 491, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 647 + ], + "type": "text", + "content": "To achieve the goal of FSCIL, we need to address catastrophic forgetting (forgetting of previous knowledge while learning new concepts) [13, 23] and overfitting issues (overfitting to few examples and thus poor generalization) [24, 43]. To bypass this convoluted mixture of issues that hinder flexible adaptation of models, most previous works [20, 32, 47, 48, 51, 55] fix the learned representation after training it on base (old) classes and employ a non-parametric classifier, using the feature-average class prototype representation [39]. However, such formulation leads to heavy reliance on the representation acquired through the optimization of softmax cross-entropy (SCE) loss on base classes, which often leads to collapsed intra-class representation [31] and poor transferability to new classes [21], as shown in Fig. 1a. Therefore, in this paper, we mainly focus on exploring effective representation learning methods, aiming to strike a better balance between discriminability on base classes and transferability to new classes." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 296, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 296, + 101 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 296, + 101 + ], + "type": "text", + "content": "J. Oh*, S. Baik*, and K.M. Lee" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 146, + 654, + 436, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 654, + 436, + 665 + ], + "spans": [ + { + "bbox": [ + 146, + 654, + 436, + 665 + ], + "type": "text", + "content": "1 As in [27, 45, 50], we use an angular histogram to visualize 2D features of a DNN." + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "There have been great advances, particularly self-supervised contrastive (SSC) learning [8,18], in the representation learning field to improve the transferability of the learned representation to downstream tasks. Some works [7,21] have attributed the strong transferability of SSC learning to 'spreading out' of intra-class features. Such representation places more emphasis on low- and mid-level features, which can be effectively transferred to and shared by new tasks. Kornblith et al. [25] have also found the relationship between the temperature of SCE loss and the spread of features, suggesting that lower temperature leads to better transferability. As illustrated in Fig. 1b, we observe that the joint optimization of SCE loss with low temperature and SSC loss encourages the spread of features and better transferability to new classes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 247, + 482, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 355 + ], + "type": "text", + "content": "Despite the foregoing, we observe that the previous methods for improving transferability are not enough to find a good representation for FSCIL; in fact, they harm the performance on base classes. Our experimental analysis suggests that excessive feature spread is very detrimental to base classes in the context of FSCIL because it hinders the feature-average class prototype from effectively representing its corresponding class, as demonstrated in Fig. 1b. Hence, it's crucial to develop a representation learning method tailored specifically for the FCSIL problem, particularly addressing the unique challenge of simultaneously achieving discriminability on seen classes and transferability to unseen classes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 356, + 482, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 463 + ], + "type": "text", + "content": "In this work, with the support from our experimental findings and information-bottleneck-theory-based analysis, we argue that the inter-class distance greatly affects the trade-off between the discriminability and transferability of the learned representation in the FSCIL problem. We find that the degraded discriminability due to the spread of features can be greatly regulated by reducing inter-class distance. Moreover, we discover that reducing the inter-class variability leads to enhancing the information bottleneck trade-off. Thus, we claim that attempting to ensure the spread of features within a compressed representation space promotes learning minimal yet intrinsic task-related information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 464, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 571 + ], + "type": "text", + "content": "Based on our analysis, in contrast to common beliefs and practices of previous FSCIL methods [20, 40, 47, 51] that have attempted to increase the inter-class distance, we propose to decrease it. Incorporating SCE loss with lower temperature, SSC loss, and inter-class distance minimization, our new objective enables the learned representation to strike a better balance between the discriminability and transferability, as illustrated in Fig. 1c. With the simple yet seemingly counter-intuitive idea of bringing classes closer (hence the name CLOSER), the proposed method demonstrates outstanding performance, suggesting a new promising research avenue regarding representation learning for FSCIL." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 590, + 243, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 243, + 604 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 243, + 604 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Few-Shot Class-Incremental Learning (FSCIL). Towards the development of real-world artificial intelligence systems, Tao et al. [41] have initially introduced FSCIL problem, subsequently fostering numerous studies in the field [1,6,9-11, 36,53]. Most of the works [20,22,28,32,40,47,48,52,54] bypass both catastrophic" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 134, + 90, + 447, + 103 + ], + "type": "text", + "content": "CLOSER: Representation Learning for Few-Shot Class-Incremental Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "text", + "content": "forgetting and overfitting issues by fixing the feature extractor trained on base classes and employing a non-parametric classifier using class prototypes [39]. Hence, recent works have focused on representation learning, where the common approach is to encourage greater separation between base classes to reserve the representation space for future new classes [20, 40, 47, 51]. However, we argue that increasing inter-class distance suppresses the acquisition of shared features among classes, which could be relevant to new classes. Thus, contrary to the prevailing belief, we suggest learning representation with an inter-class distance minimization and theoretically and empirically prove its effectiveness in improving the discriminability-transferability trade-off. Similarly, Zou et al. [55] emphasize the importance of learning shareable features among classes, which they propose to achieve with a negative margin [27]. We note that a negative margin is more related to representation spreading rather than our idea of reducing inter-class distance, as discussed in supplementary material Sec. S4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 284, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 547 + ], + "type": "text", + "content": "Transferable Representation Learning. The pursuit of representations that can be effectively transferred to downstream tasks has gained significant attention in recent years. The early works have focused on supervised training on ImageNet dataset [35] and the way to transfer the knowledge to other tasks [5,15,34]. Kornblith et al. [25] find that learned representations with better discriminability on a source dataset tend to exhibit degraded transferability to downstream tasks. To obtain a transferable representation, Liu et al. [27,55] suggest incorporating a negative margin in the softmax cross-entropy loss to promote feature sharing among classes rather than solely focusing on discriminative features. In parallel, several methods have reported the strong transferability of representation yielded by self-supervised contrastive (SSC) learning [8,18]. Islam et al. [21] claim that the enhanced transferability acquired by SSC learning can be attributed to the spread of representation, implying that a network learns more fine-grained and shareable knowledge among tasks [7]. From the perspective of information bottleneck theory, Cui et al. [12] claim that over-compression of mutual information between inputs and latent representations can prevent a network from learning features beneficial for downstream tasks. In this paper, we argue that, along with the representation spreading loss, directly regulating inter-class distance encourages a network to learn discriminative yet shareable features among classes, which can be beneficial for both base and new classes. Through the lens of information bottleneck theory, we theoretically analyze the connection between the joint optimization objective and the information bottleneck trade-off, supporting our claim." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 566, + 261, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 566, + 261, + 580 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 261, + 580 + ], + "type": "text", + "content": "3 Proposed Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 592, + 335, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 592, + 335, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 592, + 335, + 605 + ], + "type": "text", + "content": "3.1 Background: Problem Formulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 613, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 613, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 613, + 482, + 666 + ], + "type": "text", + "content": "Following the standard formulation of few-shot class incremental learning (FS-CIL) [41], we assume a sequence of training sessions with the corresponding datasets " + }, + { + "bbox": [ + 130, + 613, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{D}^{(0)},\\mathcal{D}^{(1)},\\dots ,\\mathcal{D}^{(T)}\\}" + }, + { + "bbox": [ + 130, + 613, + 482, + 666 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 613, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{(t)}" + }, + { + "bbox": [ + 130, + 613, + 482, + 666 + ], + "type": "text", + "content": " consists of training examples " + }, + { + "bbox": [ + 130, + 613, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^{(t)}" + }, + { + "bbox": [ + 130, + 613, + 482, + 666 + ], + "type": "text", + "content": " with its class labels " + }, + { + "bbox": [ + 130, + 613, + 482, + 666 + ], + "type": "inline_equation", + "content": "y_{i}^{(t)}\\in \\mathcal{C}^{(t)}" + }, + { + "bbox": [ + 130, + 613, + 482, + 666 + ], + "type": "text", + "content": " (for simplicity, we will exclude the superscript), where " + }, + { + "bbox": [ + 130, + 613, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{(t)}" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 296, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 296, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 296, + 102 + ], + "type": "text", + "content": "J. Oh*, S. Baik*, and K.M. Lee" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "text", + "content": "is the set of classes in its respective dataset " + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{(t)}" + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{(s)}\\cap \\mathcal{C}^{(t)} = \\emptyset" + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "inline_equation", + "content": "s\\neq t" + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "text", + "content": " (each dataset has its own distinct classes without overlap). In the first session (a.k.a. base session) with its dataset " + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{(0)}" + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "text", + "content": ", there is assumed to be a large number of classes available with an abundant amount of training data for each class. In subsequent sessions (a.k.a. incremental sessions) with its datasets " + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{(\\geq 1)}" + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "text", + "content": ", it is assumed that each dataset has few training examples for each class. In particular, FSCIL is said to have a " + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "text", + "content": "-way " + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "text", + "content": "-shot setting when each incremental session has " + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "text", + "content": " classes with " + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "text", + "content": " examples for each class. At each " + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "text", + "content": "-th training session, only its corresponding dataset " + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{(t)}" + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "text", + "content": " is accessible for training. After each " + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "text", + "content": "-th session, the evaluation is performed on all previously seen classes " + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{(\\leq t)}" + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "text", + "content": " using test datasets " + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}^{(\\leq t)}" + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "text", + "content": ", which consists of unseen new examples with the class label set " + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{(\\leq t)}" + }, + { + "bbox": [ + 130, + 115, + 482, + 251 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 264, + 270, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 264, + 270, + 276 + ], + "spans": [ + { + "bbox": [ + 132, + 264, + 270, + 276 + ], + "type": "text", + "content": "3.2 Background: Baseline" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 281, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 331 + ], + "type": "text", + "content": "Let a classification network consist of a feature extractor " + }, + { + "bbox": [ + 130, + 281, + 482, + 331 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 130, + 281, + 482, + 331 + ], + "type": "text", + "content": " and a classification layer with its weights " + }, + { + "bbox": [ + 130, + 281, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 281, + 482, + 331 + ], + "type": "text", + "content": ". The training objective of the base session is simply the softmax cross-entropy (SCE) loss, with the cosine similarity " + }, + { + "bbox": [ + 130, + 281, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 281, + 482, + 331 + ], + "type": "text", + "content": " between features " + }, + { + "bbox": [ + 130, + 281, + 482, + 331 + ], + "type": "inline_equation", + "content": "z = f_{\\theta}(x)" + }, + { + "bbox": [ + 130, + 281, + 482, + 331 + ], + "type": "text", + "content": " and classification weights " + }, + { + "bbox": [ + 130, + 281, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 281, + 482, + 331 + ], + "type": "text", + "content": " as logits [14]:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 209, + 337, + 482, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 337, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 209, + 337, + 482, + 370 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c e}} = \\frac {1}{B} \\sum_ {j = 1} ^ {B} - \\log \\frac {\\exp \\left(\\frac {1}{\\tau} \\operatorname {s i m} \\left(z _ {j} , \\phi_ {y _ {j}}\\right)\\right)}{\\sum_ {k = 1} ^ {| \\mathcal {C} ^ {(0)} |} \\exp \\left(\\frac {1}{\\tau} \\operatorname {s i m} \\left(z _ {j} , \\phi_ {k}\\right)\\right)}, \\tag {1}", + "image_path": "d2c895c519f075dfbe09fb3490ddcb3a1bbb615aeabc13922da67dc55ea4e7ff.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 375, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 483 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 375, + 482, + 483 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 375, + 482, + 483 + ], + "type": "text", + "content": " is the batch size and " + }, + { + "bbox": [ + 130, + 375, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 375, + 482, + 483 + ], + "type": "text", + "content": " is the temperature parameter. Incrementally updating weights with few examples in incremental sessions can make the network vulnerable to both catastrophic forgetting and overfitting. To bypass the problems, several works [20,48] suggest minimizing weight updates by freezing the feature extractor after the base session and using feature-average class prototype representation [39]. Specifically, after the base session, trained " + }, + { + "bbox": [ + 130, + 375, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 375, + 482, + 483 + ], + "type": "text", + "content": " is replaced with class prototypes, a process we refer to as classifier replacement (CR), and new-class prototypes are obtained in the incremental sessions. The " + }, + { + "bbox": [ + 130, + 375, + 482, + 483 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 375, + 482, + 483 + ], + "type": "text", + "content": "-th class prototype is acquired by averaging the features of its training samples:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 228, + 489, + 482, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 489, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 228, + 489, + 482, + 520 + ], + "type": "interline_equation", + "content": "\\phi_ {i} ^ {P} = \\frac {1}{N _ {c _ {i}}} \\sum_ {\\left(\\boldsymbol {x} _ {j}, y _ {j}\\right) \\in \\mathcal {D} ^ {\\left(\\geq 0\\right)}} \\mathbb {1} _ {[ y _ {j} = i ]} f _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {j}\\right), \\tag {2}", + "image_path": "baba1ee7e49c0710e72101e6d284d8de6305ebc359fe6279f3f6473d9dc99f82.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 525, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 561 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 525, + 482, + 561 + ], + "type": "inline_equation", + "content": "N_{c_i}" + }, + { + "bbox": [ + 130, + 525, + 482, + 561 + ], + "type": "text", + "content": " is the number of training samples for the " + }, + { + "bbox": [ + 130, + 525, + 482, + 561 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 525, + 482, + 561 + ], + "type": "text", + "content": "-th class and " + }, + { + "bbox": [ + 130, + 525, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\mathbb{1}_{[\\cdot]}" + }, + { + "bbox": [ + 130, + 525, + 482, + 561 + ], + "type": "text", + "content": " indicates 1 if the subscript condition is True and 0 otherwise. For an input " + }, + { + "bbox": [ + 130, + 525, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 525, + 482, + 561 + ], + "type": "text", + "content": ", the unnormalized log probability for the " + }, + { + "bbox": [ + 130, + 525, + 482, + 561 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 525, + 482, + 561 + ], + "type": "text", + "content": "-th class is computed by " + }, + { + "bbox": [ + 130, + 525, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(f_{\\pmb{\\theta}}(\\pmb{x}), \\phi_i^P)" + }, + { + "bbox": [ + 130, + 525, + 482, + 561 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 561, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 561, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 561, + 482, + 609 + ], + "type": "text", + "content": "Although this baseline bypasses the forgetting and overfitting issues, it heavily relies on the quality of the representation trained solely on the base classes. Thus, the main focus of this paper is to investigate the important factors that influence representation learning for FSCIL and strategies to improve them." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 624, + 482, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 624, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 624, + 482, + 636 + ], + "type": "text", + "content": "3.3 Transferability, feature spread, and its adverse effects on FSCIL" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "As shown in Fig. 1a, the baseline method exhibits a narrow intra-class distribution, widely perceived as representation collapse [31]. Recent studies [7,21,46] have" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "type": "text", + "content": "CLOSER: Representation Learning for Few-Shot Class-Incremental Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 152, + 116, + 462, + 198 + ], + "blocks": [ + { + "bbox": [ + 152, + 116, + 462, + 198 + ], + "lines": [ + { + "bbox": [ + 152, + 116, + 462, + 198 + ], + "spans": [ + { + "bbox": [ + 152, + 116, + 462, + 198 + ], + "type": "image", + "image_path": "eff9e4754620398410349535a5b9c6c336d8cec809f4950832047dcb3fa4346a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 207, + 482, + 275 + ], + "lines": [ + { + "bbox": [ + 130, + 207, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 207, + 482, + 275 + ], + "type": "text", + "content": "Fig. 2: The impact of the spread of representation. Stronger emphasis on self-supervised contrastive loss (larger " + }, + { + "bbox": [ + 130, + 207, + 482, + 275 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{ssc}}" + }, + { + "bbox": [ + 130, + 207, + 482, + 275 + ], + "type": "text", + "content": ") and low temperature (skyblue) enhances the new-class performance " + }, + { + "bbox": [ + 130, + 207, + 482, + 275 + ], + "type": "inline_equation", + "content": "A_N" + }, + { + "bbox": [ + 130, + 207, + 482, + 275 + ], + "type": "text", + "content": " (left), but at the expense of base-class performance " + }, + { + "bbox": [ + 130, + 207, + 482, + 275 + ], + "type": "inline_equation", + "content": "A_B" + }, + { + "bbox": [ + 130, + 207, + 482, + 275 + ], + "type": "text", + "content": " (center). The reduced base-class performance is mainly attributed to the excessive intra-class variation, adversely affecting the class prototype representation (right). The experiments are conducted on CIFAR100 dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 297, + 483, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 483, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 483, + 381 + ], + "type": "text", + "content": "revealed that the collapsed representation shows poor transferability due to the loss of shareable low- and mid-level features that new classes can benefit from. As such, they suggest joint optimization with a self-supervised contrastive (SSC) task [8,18], which promotes the spread of features and thus the sharing of features among classes. SSC learning optimizes infoNCE loss [30], regarding an augmented view from a query image as positive and the other images as negative samples. The SSC loss for a positive pair " + }, + { + "bbox": [ + 130, + 297, + 483, + 381 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 130, + 297, + 483, + 381 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 206, + 388, + 482, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 388, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 206, + 388, + 482, + 418 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s s c}} ^ {(i, j)} = - \\log \\frac {\\exp \\left(\\frac {1}{\\tau} \\operatorname {s i m} \\left(\\boldsymbol {z} _ {i} , \\boldsymbol {z} _ {j}\\right)\\right)}{\\sum_ {k = 1} ^ {B} \\mathbb {1} _ {[ k \\neq i ]} \\exp \\left(\\frac {1}{\\tau} \\operatorname {s i m} \\left(\\boldsymbol {z} _ {i} , \\boldsymbol {z} _ {k}\\right)\\right)}, \\tag {3}", + "image_path": "9a10d401b6f46a0a17fd1b2f910c6cd44bc6c1bb82e6c08444daf213123060db.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 425, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 425, + 482, + 449 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 425, + 482, + 449 + ], + "type": "text", + "content": " is the number of samples, including augmented images, and " + }, + { + "bbox": [ + 130, + 425, + 482, + 449 + ], + "type": "inline_equation", + "content": "z_{j}" + }, + { + "bbox": [ + 130, + 425, + 482, + 449 + ], + "type": "text", + "content": " is the feature from an augmented view of " + }, + { + "bbox": [ + 130, + 425, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{i}" + }, + { + "bbox": [ + 130, + 425, + 482, + 449 + ], + "type": "text", + "content": ". The loss is averaged over all positive pairs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 449, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 555 + ], + "type": "text", + "content": "In parallel, several studies have proposed to reduce the temperature and margin parameters in the softmax cross-entropy loss as another way to encourage feature sharing [25,27,55]. Our empirical analysis in Sec. S4 of the supplementary material demonstrates that the temperature parameter affects the transferability more than the margin. Thus, we employ a low-temperature parameter in the SCE loss to further enhance the transferability of learned representations. Indeed, we observe that the combination of " + }, + { + "bbox": [ + 130, + 449, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ssc}}" + }, + { + "bbox": [ + 130, + 449, + 482, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 449, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ce}}" + }, + { + "bbox": [ + 130, + 449, + 482, + 555 + ], + "type": "text", + "content": " with low " + }, + { + "bbox": [ + 130, + 449, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 449, + 482, + 555 + ], + "type": "text", + "content": " results in better new-class performance in the FSCIL problem due to a larger spread of features, as demonstrated in Fig. 1b and the left figure of Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "text", + "content": "However, as illustrated in the center figure in Fig. 2, we observe that the pursuit of better transferability results in a degradation in discriminability on the base classes. While exploring the dilemma, we discover that the performance decline on the base classes mainly arises from the base class classifier replacement (CR) strategy, which is introduced in Sec. 3.2. The figure on the right in Fig. 2 shows the accuracy on the base classes subsequent to the base session training " + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "inline_equation", + "content": "(A_B^0)" + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "text", + "content": ", without considering the new classes, both before and after CR, denoted by " + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "inline_equation", + "content": "A_B^{0,\\mathrm{bef}}" + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "inline_equation", + "content": "A_B^{0,\\mathrm{aft}}" + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "text", + "content": ", respectively. Before CR, " + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "inline_equation", + "content": "A_B^{0,\\mathrm{bef}}" + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "text", + "content": " exhibits a relatively high value and a small variance as " + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{ssc}}" + }, + { + "bbox": [ + 130, + 556, + 482, + 666 + ], + "type": "text", + "content": " vary (indicated by the solid lines). Conversely," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 296, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 296, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 296, + 102 + ], + "type": "text", + "content": "J. Oh*, S. Baik*, and K.M. Lee" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 221, + 116, + 394, + 210 + ], + "blocks": [ + { + "bbox": [ + 221, + 116, + 394, + 210 + ], + "lines": [ + { + "bbox": [ + 221, + 116, + 394, + 210 + ], + "spans": [ + { + "bbox": [ + 221, + 116, + 394, + 210 + ], + "type": "image", + "image_path": "3d3e3139b96bcdb2dfe4a3a8e18f54cc7c92b4dcd3437e3089064d87b857defc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 221, + 482, + 266 + ], + "lines": [ + { + "bbox": [ + 131, + 221, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 131, + 221, + 482, + 266 + ], + "type": "text", + "content": "Fig. 3: Effect of minimizing inter-class distance. As the weight of " + }, + { + "bbox": [ + 131, + 221, + 482, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{inter}}" + }, + { + "bbox": [ + 131, + 221, + 482, + 266 + ], + "type": "text", + "content": ", denoted by " + }, + { + "bbox": [ + 131, + 221, + 482, + 266 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{inter}}" + }, + { + "bbox": [ + 131, + 221, + 482, + 266 + ], + "type": "text", + "content": ", increases, the performance on the new classes increases (skyblue) and the performance loss on the base classes induced by CR is greatly alleviated (purple). The experiments are conducted on CUB200 dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 286, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 286, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 131, + 286, + 482, + 373 + ], + "type": "inline_equation", + "content": "A_B^{0,\\mathrm{aft}}" + }, + { + "bbox": [ + 131, + 286, + 482, + 373 + ], + "type": "text", + "content": " is considerably lower than " + }, + { + "bbox": [ + 131, + 286, + 482, + 373 + ], + "type": "inline_equation", + "content": "A_B^{0,\\mathrm{bef}}" + }, + { + "bbox": [ + 131, + 286, + 482, + 373 + ], + "type": "text", + "content": " with a relatively higher variance (indicated by the dotted lines). Considering that CR is introduced to bypass overfitting and catastrophic forgetting in the FSCIL problem, these findings suggest that the previous methods for improving the transferability of representations are not effective in enhancing the trade-off between transferability and discriminability in the context of the FSCIL problem. Thus, it is essential to develop a representation learning method tailored for the FSCIL problem." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 388, + 307, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 388, + 307, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 388, + 307, + 399 + ], + "type": "text", + "content": "3.4 Inter-Class Distance Matters" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 406, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 406, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 131, + 406, + 482, + 525 + ], + "type": "text", + "content": "As discussed in Sec. 3.3, learning shareable features through representation spreading proves advantageous for transferability. Yet, it also harms discriminability in the process of CR. Based on the observation in Fig. 1b, we hypothesize that such a dilemma appears to arise from a large inter-class distance since the representation spreading could push features into the extensive inter-class space, which may dilute the information on the base classes. Furthermore, we argue that the large inter-class distance may impede effective feature sharing among classes, undermining the transferability of learned representations. Consequently, to retain the knowledge on base classes while promoting effective feature sharing, we introduce a novel loss function that minimizes the inter-class distance:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 192, + 530, + 482, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 530, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 192, + 530, + 482, + 574 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {i n t e r}} = - \\frac {1}{\\sum_ {i = 1} ^ {B} \\sum_ {j > i} ^ {B} \\mathbb {1} _ {[ y _ {i} \\neq y _ {j} ]}} \\sum_ {i = 1} ^ {B} \\sum_ {j > i} ^ {B} \\mathbb {1} _ {[ y _ {i} \\neq y _ {j} ]} \\operatorname {s i m} \\left(\\boldsymbol {z} _ {i}, \\boldsymbol {z} _ {j}\\right). \\tag {4}", + "image_path": "4a0bf5788cfeab3bb15596e846478e89456b949ca2cfc58c5dadf89d9528d547.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 582, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 582, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 131, + 582, + 482, + 641 + ], + "type": "text", + "content": "Fig. 1c displays that the spread of intra-class features is well regulated by applying " + }, + { + "bbox": [ + 131, + 582, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{inter}}" + }, + { + "bbox": [ + 131, + 582, + 482, + 641 + ], + "type": "text", + "content": " and Fig. 3 demonstrates the performance decline from CR is alleviated by minimizing " + }, + { + "bbox": [ + 131, + 582, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{inter}}" + }, + { + "bbox": [ + 131, + 582, + 482, + 641 + ], + "type": "text", + "content": " (indicated by the purple line). Moreover, the results indicated by the skyblue line show that reducing inter-class distance improves the performance on the new classes, corroborating our hypothesis." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "Our assertion initially seems counter-intuitive, diverging from the common belief of prior works [20,40,47,51] that maximizing inter-class distance may be" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 134, + 90, + 448, + 103 + ], + "type": "text", + "content": "CLOSER: Representation Learning for Few-Shot Class-Incremental Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 249, + 193 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 249, + 193 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 249, + 193 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 249, + 193 + ], + "type": "image", + "image_path": "86efe6a231a80e37d7d61896f9fa4f656b04acdd899e768562c8e6c44e592204.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 196, + 198, + 205 + ], + "lines": [ + { + "bbox": [ + 186, + 196, + 198, + 205 + ], + "spans": [ + { + "bbox": [ + 186, + 196, + 198, + 205 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 252, + 116, + 362, + 193 + ], + "blocks": [ + { + "bbox": [ + 252, + 116, + 362, + 193 + ], + "lines": [ + { + "bbox": [ + 252, + 116, + 362, + 193 + ], + "spans": [ + { + "bbox": [ + 252, + 116, + 362, + 193 + ], + "type": "image", + "image_path": "3c02a39d0008eb38d7b3606114271a6386a2c10a75372f41fe38e7687d706822.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 300, + 196, + 313, + 205 + ], + "lines": [ + { + "bbox": [ + 300, + 196, + 313, + 205 + ], + "spans": [ + { + "bbox": [ + 300, + 196, + 313, + 205 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 214, + 482, + 314 + ], + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 314 + ], + "type": "text", + "content": "Fig. 4: (a) Sanity test for " + }, + { + "bbox": [ + 130, + 214, + 482, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(f_{\\theta})" + }, + { + "bbox": [ + 130, + 214, + 482, + 314 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 130, + 214, + 482, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(f_{\\theta})" + }, + { + "bbox": [ + 130, + 214, + 482, + 314 + ], + "type": "text", + "content": " has a positive correlation with the performance on the new classes. Each data point is obtained by different configurations of " + }, + { + "bbox": [ + 130, + 214, + 482, + 314 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 214, + 482, + 314 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 214, + 482, + 314 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{ssc}}" + }, + { + "bbox": [ + 130, + 214, + 482, + 314 + ], + "type": "text", + "content": " (without " + }, + { + "bbox": [ + 130, + 214, + 482, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{inter}}" + }, + { + "bbox": [ + 130, + 214, + 482, + 314 + ], + "type": "text", + "content": "). (b), (c) Relationship between inter-class distance, " + }, + { + "bbox": [ + 130, + 214, + 482, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(f_{\\theta})" + }, + { + "bbox": [ + 130, + 214, + 482, + 314 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 214, + 482, + 314 + ], + "type": "inline_equation", + "content": "A_N" + }, + { + "bbox": [ + 130, + 214, + 482, + 314 + ], + "type": "text", + "content": ": Integrated with the representation spreading, reducing inter-class distance encourages better transferability (red points). However, the tendency is broken when reducing inter-class distance without representation spreading (blue points). Please refer to Sec. 3.5 for theoretical support for these observations. The dots with greater transparency correspond to smaller " + }, + { + "bbox": [ + 130, + 214, + 482, + 314 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{inter}}" + }, + { + "bbox": [ + 130, + 214, + 482, + 314 + ], + "type": "text", + "content": ", ranging from 0 to 1 with intervals of 0.1. We set " + }, + { + "bbox": [ + 130, + 214, + 482, + 314 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{ssc}}" + }, + { + "bbox": [ + 130, + 214, + 482, + 314 + ], + "type": "text", + "content": " as 0.1 when it is used. The experiments are conducted on CIFAR100 dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 366, + 116, + 475, + 193 + ], + "blocks": [ + { + "bbox": [ + 366, + 116, + 475, + 193 + ], + "lines": [ + { + "bbox": [ + 366, + 116, + 475, + 193 + ], + "spans": [ + { + "bbox": [ + 366, + 116, + 475, + 193 + ], + "type": "image", + "image_path": "4d42ebafd5b7bf19f328c489499f67d10df6b71943c7ce0ca18c358588847db1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 196, + 426, + 205 + ], + "lines": [ + { + "bbox": [ + 414, + 196, + 426, + 205 + ], + "spans": [ + { + "bbox": [ + 414, + 196, + 426, + 205 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 340, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 413 + ], + "type": "text", + "content": "beneficial for reserving representation space for future new classes. To further validate the efficacy of reducing inter-class distance with respect to transferability, we propose a measure to quantify how new class samples are distinct from base-class representations. We define the measure as the averaged relative angular distance between a new-class sample and its nearest base-class prototype with respect to the averaged angular distance among all base-class prototype pairs:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 197, + 423, + 482, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 423, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 197, + 423, + 482, + 460 + ], + "type": "interline_equation", + "content": "\\mathcal {T} (f _ {\\boldsymbol {\\theta}}) = \\frac {\\frac {1}{| \\mathcal {D} _ {t e s t} ^ {(> 0)} |} \\sum_ {(\\boldsymbol {x} _ {j} , y _ {j}) \\in \\mathcal {D} _ {t e s t} ^ {(> 0)} i} \\min _ {i} \\angle (\\boldsymbol {z} _ {j} , \\phi_ {b a s e , i} ^ {P})}{\\sum_ {j = 1} ^ {| \\mathcal {C} ^ {(0)} |} \\sum_ {k > j} ^ {| \\mathcal {C} ^ {(0)} |} \\angle \\left(\\phi_ {j} ^ {P} , \\phi_ {k} ^ {P}\\right) / \\left(^ {| \\mathcal {C} ^ {(0)} |} _ {2}\\right)}, \\tag {5}", + "image_path": "ff44bfd830f644d5765c2e7021d1811790c5e54898a9091832661c8d421e6d6d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\phi_{base,i}^{P}" + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\angle (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "text", + "content": " indicate the " + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "text", + "content": " -th base-class prototype and the angular distance between two input vectors, respectively. We introduce the denominator to normalize the varying sizes of the representation space depending on methods. If the representation produced by " + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "text", + "content": " has a distinguishable representation for new classes, " + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(f_{\\theta})" + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "text", + "content": " would be large. For the sanity test, we check the relationship between " + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(f_{\\theta})" + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "text", + "content": " and the accuracy of the new class, which is shown in Fig. 4a." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "text", + "content": "Using this measure, we analyze the relationship between inter-class distance, the spread of features, and the transferability of learned representation. To do so, we train the feature extractor using " + }, + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ce}}" + }, + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "text", + "content": " with a low temperature, " + }, + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ssc}}" + }, + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{inter}}" + }, + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "text", + "content": " with varying loss weights for " + }, + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ssc}}" + }, + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{inter}}" + }, + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "text", + "content": ", denoted by " + }, + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{ssc}}" + }, + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{inter}}" + }, + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "text", + "content": ", respectively. Figs. 4b and 4c show the relationship between the inter-class distance, " + }, + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(f_{\\theta})" + }, + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "text", + "content": ", and the performance on the new classes. The results demonstrate that the joint optimization of " + }, + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ssc}}" + }, + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{inter}}" + }, + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "text", + "content": " leads to an increase in both " + }, + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(f_{\\theta})" + }, + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "inline_equation", + "content": "A_N" + }, + { + "bbox": [ + 130, + 546, + 483, + 665 + ], + "type": "text", + "content": ", indicating that smaller inter-class distances lead to more discriminability between base classes and new classes. The analysis corroborates our seemingly counter-intuitive claim that the closer classes are, the better." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 296, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 296, + 101 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 296, + 101 + ], + "type": "text", + "content": "J. Oh*, S. Baik*, and K.M. Lee" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "In summary, we have observed that the spread of representation achieved by employing a low temperature in the SCE loss and self-supervised contrastive loss is advantageous for learning transferable representation. However, spread representation itself cannot address the trade-off between transferability to new classes and discriminability on base classes in the context of the FSCIL problem. Our analysis demonstrates that decreasing inter-class distance enhances discriminability by regularizing the intra-class spread and improves the transferability by promoting the effective learning of shareable information among classes when combined with the feature-spread-encouraging loss. Consequently, our final loss function is the combination of cross-entropy loss with a lower temperature parameter, self-supervised contrastive loss, and inter-class distance minimizing loss:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 236, + 260, + 482, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 260, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 236, + 260, + 482, + 274 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {c e}} + \\lambda_ {\\mathrm {s s c}} \\mathcal {L} _ {\\mathrm {s s c}} + \\lambda_ {\\mathrm {i n t e r}} \\mathcal {L} _ {\\mathrm {i n t e r}}. \\tag {6}", + "image_path": "2a5901d7eee01e0672e497ef7d64ffd2b3873725c0f810456052b60f0d0cbece.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 287, + 380, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 287, + 380, + 300 + ], + "spans": [ + { + "bbox": [ + 131, + 287, + 380, + 300 + ], + "type": "text", + "content": "3.5 Information Bottleneck Theory Perspective" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 482, + 413 + ], + "type": "text", + "content": "In this subsection, we provide theoretical support for the proposed representation learning objective in Eq. (6) from the perspective of the information bottleneck (IB) theory [3, 37, 42]. The IB theory describes the goal of representation learning as finding a good trade-off between complexity and accuracy through finding minimal information from inputs necessary to preserve maximal information about the targets. As discussed by Cui et al. [12], we consider the objective of IB theory to be closely related to learning transferable representations since while achieving the objective, a network could be guided to learn intrinsic knowledge rather than task-irrelevant shortcuts [17], leading to better transferability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 413, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 477 + ], + "type": "text", + "content": "For an image classification task, let " + }, + { + "bbox": [ + 130, + 413, + 482, + 477 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{h \\times w \\times 3}" + }, + { + "bbox": [ + 130, + 413, + 482, + 477 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 413, + 482, + 477 + ], + "type": "inline_equation", + "content": "Y \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 130, + 413, + 482, + 477 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 413, + 482, + 477 + ], + "type": "inline_equation", + "content": "Z = \\frac{f_{\\theta}(X)}{\\|f_{\\theta}(X)\\|} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 413, + 482, + 477 + ], + "type": "text", + "content": " denote the input, label, and normalized latent representation variables, respectively, where " + }, + { + "bbox": [ + 130, + 413, + 482, + 477 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 413, + 482, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 413, + 482, + 477 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 413, + 482, + 477 + ], + "type": "text", + "content": " are the spatial sizes of the image, " + }, + { + "bbox": [ + 130, + 413, + 482, + 477 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 413, + 482, + 477 + ], + "type": "text", + "content": " is the number of classes, and " + }, + { + "bbox": [ + 130, + 413, + 482, + 477 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 413, + 482, + 477 + ], + "type": "text", + "content": " is the dimension of the latent representations. The aforementioned trade-off has been formulated as the following objective:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 253, + 485, + 481, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 485, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 253, + 485, + 481, + 498 + ], + "type": "interline_equation", + "content": "\\max I (Y; Z) - \\beta I (X; Z), \\tag {7}", + "image_path": "cd81351f2aedce22ef2ece8c28a08c9a62f74d7d3f55661e5ab3fdc37a7c804b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 503, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 503, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 131, + 503, + 482, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 503, + 482, + 567 + ], + "type": "inline_equation", + "content": "I(\\cdot ;\\cdot)" + }, + { + "bbox": [ + 131, + 503, + 482, + 567 + ], + "type": "text", + "content": " indicates the mutual information between two random variables and " + }, + { + "bbox": [ + 131, + 503, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\beta > 0" + }, + { + "bbox": [ + 131, + 503, + 482, + 567 + ], + "type": "text", + "content": " is a Lagrange multiplier. In this work, we consider an alternative trade-off objective, " + }, + { + "bbox": [ + 131, + 503, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\max \\frac{I(Y;Z)}{\\beta I(X;Z)}" + }, + { + "bbox": [ + 131, + 503, + 482, + 567 + ], + "type": "text", + "content": ", which is adopted by several works [29,38]. After omitting " + }, + { + "bbox": [ + 131, + 503, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 131, + 503, + 482, + 567 + ], + "type": "text", + "content": " for simplicity and modest assumptions, we derive the following lower bound of the IB trade-off objective:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 211, + 573, + 482, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 573, + 482, + 613 + ], + "spans": [ + { + "bbox": [ + 211, + 573, + 482, + 613 + ], + "type": "interline_equation", + "content": "\\frac {I (Y ; Z)}{I (X ; Z)} \\geq 1 - \\frac {d \\cdot \\log (2 \\pi e) + \\frac {1}{C} \\sum_ {i = 1} ^ {C} \\log | \\Sigma_ {W _ {i}} |}{d \\cdot \\log (2 \\pi e) + \\log | \\Sigma_ {T} |}, \\tag {8}", + "image_path": "ea583001b6b3b5e9f64c5bc998fdbc22a02d6697453d49358c5a96420915b756.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Sigma_{W_i}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\Sigma_T" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " indicate the covariance matrices of " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " within the " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "-th class and overall classes, respectively. As proven by Lemma S1 in the supplementary material, both the numerator and denominator in the fractional term of the lower bound in Eq. (8) are negative, leading to the following theorem:" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "type": "text", + "content": "CLOSER: Representation Learning for Few-Shot Class-Incremental Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 136, + 460, + 294 + ], + "blocks": [ + { + "bbox": [ + 156, + 114, + 456, + 125 + ], + "lines": [ + { + "bbox": [ + 156, + 114, + 456, + 125 + ], + "spans": [ + { + "bbox": [ + 156, + 114, + 456, + 125 + ], + "type": "text", + "content": "Table 1: 10-way 5-shot incremental learning results on CUB200." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 136, + 460, + 294 + ], + "lines": [ + { + "bbox": [ + 151, + 136, + 460, + 294 + ], + "spans": [ + { + "bbox": [ + 151, + 136, + 460, + 294 + ], + "type": "table", + "html": "
MethodAcc. in each session (%)PD (%) ↓
012345678910
Baseline79.9276.2373.1869.4567.8365.7464.5463.3361.5661.2760.1019.83
TOPIC [41]68.6862.4954.8149.9945.2541.4038.3535.3632.2228.3126.2642.42
F2M [36]81.0778.1675.5772.8970.8668.1767.0165.2663.3661.7660.2620.81
CEC [48]75.8571.9468.5063.5062.4358.2757.7355.8154.8353.5252.2823.57
IDLVQ-C [6]77.3774.7270.2867.1365.3463.5262.1061.5459.0458.6857.8119.56
ALICE [33]77.4072.7070.6067.2065.9063.4062.9061.9060.5060.6060.1017.30
CLOM [55]79.5776.0772.9469.8267.8065.5663.9462.5960.6260.3459.5819.99
Entropy Reg. [28]75.9072.1468.6463.7662.5859.1157.8255.8954.9253.5852.3923.51
LIMIT [52]75.8973.5571.9968.1467.4263.6162.4061.3559.9158.6657.4118.48
MetaFSCIL [11]75.9072.4168.7864.7862.9659.9958.3056.8554.7853.8252.6423.26
FACT [51]75.9073.2370.8466.1365.5662.1561.7459.8358.4157.8956.9418.96
S3C [22]80.6277.5573.1968.5468.0564.3363.5862.0760.6159.7958.9520.83
SAVC [40]81.8577.9274.9570.2169.9667.0266.1665.3063.8463.1562.5019.35
NC-FSCIL [47]80.4575.9872.3070.2868.1765.1664.4363.2560.6660.0159.4421.01
GKEAL [54]78.8875.6272.3268.6267.2364.2662.9861.8960.2059.2158.6720.21
CABD [49]79.1275.3772.8069.0567.5365.1264.0063.5161.8761.4760.9318.19
CLOSER (Ours)79.4075.9273.5070.4769.2467.2266.7365.6964.0064.0263.5815.82
", + "image_path": "7aecea879fcc79723767c60fc778a300051c1e84ac46d210996fd2e48f83abba.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 168, + 331, + 443, + 439 + ], + "blocks": [ + { + "bbox": [ + 154, + 309, + 458, + 320 + ], + "lines": [ + { + "bbox": [ + 154, + 309, + 458, + 320 + ], + "spans": [ + { + "bbox": [ + 154, + 309, + 458, + 320 + ], + "type": "text", + "content": "Table 2: 5-way 5-shot incremental learning results on CIFAR100." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 168, + 331, + 443, + 439 + ], + "lines": [ + { + "bbox": [ + 168, + 331, + 443, + 439 + ], + "spans": [ + { + "bbox": [ + 168, + 331, + 443, + 439 + ], + "type": "table", + "html": "
MethodAcc. in each session (%)PD (%) ↓
012345678
Baseline72.9368.4664.2660.1556.5353.6051.5149.1947.0925.84
ERDIL [16]73.6268.2265.1461.8458.3555.5452.5150.1648.2325.39
CEC [48]73.0768.8865.2661.1958.0955.5753.2251.3449.1423.93
CLOM [55]74.2069.8366.1762.3959.2656.4854.3652.1650.2523.95
Entropy Reg. [28]74.470.266.5462.5159.7156.5854.5252.3950.1424.26
FACT [51]74.6072.0967.5663.5261.3858.3656.2854.2452.1022.50
LIMIT [52]73.8172.0967.8763.8960.7057.7755.6753.5251.2322.58
MetaFSCIL [11]74.5070.1066.8462.7759.4856.5254.3652.5649.9724.53
GKEAL [54]74.0170.4567.0163.0860.0157.3055.5053.3951.4022.61
CLOSER (Ours)75.7271.8368.3264.6261.9159.2557.5355.4353.3222.40
", + "image_path": "6de8b99bfbae89fbabfabab40a63976e6aec98df30a48a2b6892e6e0a44886f8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 466, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 466, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 466, + 482, + 495 + ], + "type": "text", + "content": "Theorem 1. The lower bound of " + }, + { + "bbox": [ + 130, + 466, + 482, + 495 + ], + "type": "inline_equation", + "content": "\\frac{I(Y;Z)}{I(X;Z)}" + }, + { + "bbox": [ + 130, + 466, + 482, + 495 + ], + "type": "text", + "content": " in Eq. (8) is a monotonically increasing function of " + }, + { + "bbox": [ + 130, + 466, + 482, + 495 + ], + "type": "inline_equation", + "content": "|\\Sigma_{W_i}|" + }, + { + "bbox": [ + 130, + 466, + 482, + 495 + ], + "type": "text", + "content": " and a monotonically decreasing function of " + }, + { + "bbox": [ + 130, + 466, + 482, + 495 + ], + "type": "inline_equation", + "content": "|\\Sigma_T|" + }, + { + "bbox": [ + 130, + 466, + 482, + 495 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 507, + 482, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 482, + 519 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 482, + 519 + ], + "type": "text", + "content": "The detailed derivations are presented in Sec. S1 of the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": "Generally, the determinant of a covariance matrix is correlated with the spread or variability of variables across all dimensions. Thus, Theorem 1 suggests that the IB trade-off can be enhanced by increasing the intra-class variability " + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{ce}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": " with a low temperature and " + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ssc}})" + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": " while suppressing the overall representation space " + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{inter}})" + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": ", supporting the proposed learning objective in Eq. (6) and our claims for inter-class distance minimization. Intuitively, our objective functions encourage greater overlap in representations among different classes, promoting more shareable and compact representations. Our theoretic analysis also underpins the findings in Fig. 4 that reducing inter-class distance without representation spreading loss is observed to result in less transferable representations. Since compressing the overall feature space diminishes the intra-class variability, the IB trade-off may decrease, leading to degraded transferability." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 295, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 295, + 101 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 295, + 101 + ], + "type": "text", + "content": "J. Oh*, S. Baik*, and K.M. Lee" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 168, + 136, + 443, + 281 + ], + "blocks": [ + { + "bbox": [ + 144, + 114, + 468, + 125 + ], + "lines": [ + { + "bbox": [ + 144, + 114, + 468, + 125 + ], + "spans": [ + { + "bbox": [ + 144, + 114, + 468, + 125 + ], + "type": "text", + "content": "Table 3: 5-way 5-shot incremental learning results on miniImageNet." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 168, + 136, + 443, + 281 + ], + "lines": [ + { + "bbox": [ + 168, + 136, + 443, + 281 + ], + "spans": [ + { + "bbox": [ + 168, + 136, + 443, + 281 + ], + "type": "table", + "html": "
MethodAcc. in each session (%)PD (%) ↓
012345678
Baseline72.2767.4663.2659.7356.5653.5350.9048.9347.2625.01
TOPIC [41]61.3150.0945.1741.1637.4835.5232.1929.4624.4236.89
F2M [36]67.2863.8060.3857.0654.0851.3948.8246.5844.6522.63
CEC [48]72.0066.8362.9759.4356.7053.7351.1949.2447.6324.37
IDLVQ-C [6]64.7759.8755.9352.6249.8847.5544.8343.1441.8422.93
Subspace Reg. [2]80.3773.7668.3664.0760.3656.2753.1050.4547.5532.83
ALICE [33]80.6070.6067.4064.5062.5060.0057.8056.8055.7024.90
CLOM [55]73.0868.0964.1660.4157.4154.2951.5449.3748.0025.08
Entropy Reg. [28]71.8467.1263.2159.7757.0153.9551.5549.5248.2123.63
LIMIT [52]72.3268.4764.3060.7857.9555.0752.7050.7249.1423.13
MetaFSCIL [11]72.0467.9463.7760.2957.5855.1652.9050.7949.1922.85
FACT [51]72.5669.6366.3862.7760.6057.3354.3452.1650.4922.07
GKEAL [54]73.5968.9065.3362.2959.3956.7054.2052.5951.3122.28
CABD [49]74.6570.4366.2962.7760.7557.2454.7953.6552.2222.43
CLOSER (Ours)76.0271.6167.9964.6961.7058.9456.2354.5253.3322.69
", + "image_path": "9acab8849aea997e2eea0fa551a2a4c03f1033f7fcbfe61b7c72e88fa7466fd0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 303, + 230, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 230, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 230, + 316 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 327, + 268, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 268, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 268, + 340 + ], + "type": "text", + "content": "4.1 Experimental Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 347, + 482, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 454 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 454 + ], + "type": "text", + "content": "Dataset. Following the benchmark settings proposed by Tao et al. [41], we evaluate the proposed method on CIFAR100 [26], miniImageNet [43], and CUB200 [44]. For CIFAR100 and miniImageNet, the total number of classes is 100: 60 base classes and 40 new classes. The 40 new classes are split into 8 disjoint sets of 5 classes, each of which is sequentially provided with 5 training examples per class (5-way 5-shot) in each incremental session. As for CUB200, there total number of classes is 200, with 100 base classes and 100 new classes. 100 new classes are split into 10 disjoint sets of 10 classes, each of which is sequentially provided with 5 training examples per class (10-way 5-shot) in each incremental session." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 455, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 482, + 598 + ], + "type": "text", + "content": "Implementation. Following Zhang et al. [48], we use ResNet-20 [19] for CIFAR100 experiments and ResNet-18 [19] for both miniImageNet and CUB200 experiments. We follow the conventions to use the ResNet-18 model pre-trained on the ImageNet dataset [35] for CUB200. We set the mini-batch size to 128, 128, and 256 for CIFAR100, miniImageNet, and CUB200 experiments, respectively. The temperature parameter " + }, + { + "bbox": [ + 130, + 455, + 482, + 598 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 455, + 482, + 598 + ], + "type": "text", + "content": " for the baseline method is 1/16 and 'low temperature' indicates " + }, + { + "bbox": [ + 130, + 455, + 482, + 598 + ], + "type": "inline_equation", + "content": "\\tau = 1 / 32" + }, + { + "bbox": [ + 130, + 455, + 482, + 598 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 130, + 455, + 482, + 598 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{ssc}}" + }, + { + "bbox": [ + 130, + 455, + 482, + 598 + ], + "type": "text", + "content": " as 0.1,0.1, and 0.01 for CIFAR100, miniImageNet, and CUB200, respectively, and " + }, + { + "bbox": [ + 130, + 455, + 482, + 598 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{inter}}" + }, + { + "bbox": [ + 130, + 455, + 482, + 598 + ], + "type": "text", + "content": " as 1, 0.5, and 1.5 for CIFAR100, miniImageNet, and CUB200, respectively. These hyper-parameters are searched via validation using synthesized validation sets. For mutual information estimation, we adopt MINE [4] method and implement it based on the open-source code2. More details are provided in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": "Evaluation. We use the accuracy on base " + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "inline_equation", + "content": "(A_B)" + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": ", new " + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "inline_equation", + "content": "(A_N)" + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": ", and the whole classes " + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "inline_equation", + "content": "(A_W)" + }, + { + "bbox": [ + 130, + 598, + 482, + 647 + ], + "type": "text", + "content": " as metrics to assess the discriminability, transferability, and the trade-off between them in learned representations. Additionally, we use the performance drop (PD) between the accuracy at the end of the base session (session 0) and" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "text", + "content": "CLOSER: Representation Learning for Few-Shot Class-Incremental Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 145, + 653, + 299, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 653, + 299, + 665 + ], + "spans": [ + { + "bbox": [ + 145, + 653, + 299, + 665 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 653, + 299, + 665 + ], + "type": "text", + "content": "https://github.com/gtegner/mine-pytorch" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 228, + 136, + 386, + 208 + ], + "blocks": [ + { + "bbox": [ + 212, + 114, + 401, + 125 + ], + "lines": [ + { + "bbox": [ + 212, + 114, + 401, + 125 + ], + "spans": [ + { + "bbox": [ + 212, + 114, + 401, + 125 + ], + "type": "text", + "content": "Table 4: Ablation studies on CIFAR100." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 228, + 136, + 386, + 208 + ], + "lines": [ + { + "bbox": [ + 228, + 136, + 386, + 208 + ], + "spans": [ + { + "bbox": [ + 228, + 136, + 386, + 208 + ], + "type": "table", + "html": "
lowτL SSCL interAB(%)AN(%)AW(%)PD(%)
XXX70.1320.9550.4624.81
X✓X69.9522.8051.0924.59
XX✓69.9520.0549.9926.78
X✓✓71.4322.8351.9925.26
✓XX68.3324.1050.6423.23
✓✓X66.5825.0849.9823.15
✓X✓70.1722.4051.0625.27
✓✓✓70.7227.2353.3222.40
", + "image_path": "7bbd3b239ce2cb6c62e8a35aabe8f1381f3cfb76c96a67ba1d6550181b61cc31.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 232, + 481, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 232, + 481, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 481, + 269 + ], + "type": "text", + "content": "the accuracy at the last incremental session to evaluate the degree to which old knowledge is forgotten and new knowledge is learned simultaneously. All experimental results of CLOSER are obtained by averaging 3 trials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 292, + 348, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 292, + 348, + 304 + ], + "spans": [ + { + "bbox": [ + 131, + 292, + 348, + 304 + ], + "type": "text", + "content": "4.2 Comparison with the Existing Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 316, + 482, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 473 + ], + "type": "text", + "content": "We compare the proposed method, dubbed CLOSER, with prior arts on CUB200 (Tab. 1), CIFAR100 (Tab. 2), and miniImageNet (Tab. 3). We observe that CLOSER achieves state-of-the-art performance on both CUB200 and CIFAR100 datasets, surpassing the results of previous methods by a large margin with respect to " + }, + { + "bbox": [ + 130, + 316, + 482, + 473 + ], + "type": "inline_equation", + "content": "A_W" + }, + { + "bbox": [ + 130, + 316, + 482, + 473 + ], + "type": "text", + "content": " and PD. With miniImageNet, the proposed method exhibits substantially higher " + }, + { + "bbox": [ + 130, + 316, + 482, + 473 + ], + "type": "inline_equation", + "content": "A_W" + }, + { + "bbox": [ + 130, + 316, + 482, + 473 + ], + "type": "text", + "content": " than the method with the lowest PD and achieves lower PD than the method with the highest " + }, + { + "bbox": [ + 130, + 316, + 482, + 473 + ], + "type": "inline_equation", + "content": "A_W" + }, + { + "bbox": [ + 130, + 316, + 482, + 473 + ], + "type": "text", + "content": ", which means that the proposed method achieves a better balance between the performance on base and the new classes. It is worth noting that CLOSER shows such outstanding performance without any assistance from the storage of previous samples (F2M, ERDIL, IDLVQ-C, and CABD), additional computational modules (CEC, CLOM, NC-FSCIL, SAVC, and MetaFSCIL), and test-time data augmentation (S3C and SAVC), suggesting the critical importance of learning effective representations in FSCIL." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 496, + 246, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 496, + 246, + 508 + ], + "spans": [ + { + "bbox": [ + 131, + 496, + 246, + 508 + ], + "type": "text", + "content": "4.3 Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "To verify the efficacy of the individual components in the proposed method, we perform ablation studies, which are shown in Tab. 4. The increase in " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "A_N" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " when employing a lower temperature in the softmax cross-entropy loss (low " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ") or a self-supervised contrastive loss (" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ssc}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ") confirms the advantage of feature spread in enhancing the transferability of representations. While transferability can be greatly improved by using low " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ssc}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", a substantial decline in base-class performance " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "A_B" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " is observed, as discussed in Sec. 3.4. The result from the case where all components are utilized demonstrates that this issue can be effectively resolved by minimizing inter-class distance. Furthermore, as discussed in Sec. 3.4, reducing inter-class distance is observed to improve transferability, especially when used with the representation spreading methods. The comprehensive results of ablation studies confirm and support our claims." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 295, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 295, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 295, + 101 + ], + "type": "text", + "content": "J. Oh*, S. Baik*, and K.M. Lee" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 146, + 115, + 252, + 193 + ], + "blocks": [ + { + "bbox": [ + 146, + 115, + 252, + 193 + ], + "lines": [ + { + "bbox": [ + 146, + 115, + 252, + 193 + ], + "spans": [ + { + "bbox": [ + 146, + 115, + 252, + 193 + ], + "type": "image", + "image_path": "7926c9b824bf0fb31175b52982c9974883d8f2b5324cc38cbc6dfae80a1c2077.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 194, + 229, + 204 + ], + "lines": [ + { + "bbox": [ + 169, + 194, + 229, + 204 + ], + "spans": [ + { + "bbox": [ + 169, + 194, + 229, + 204 + ], + "type": "text", + "content": "(a) Base Classes" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 253, + 115, + 356, + 194 + ], + "blocks": [ + { + "bbox": [ + 253, + 115, + 356, + 194 + ], + "lines": [ + { + "bbox": [ + 253, + 115, + 356, + 194 + ], + "spans": [ + { + "bbox": [ + 253, + 115, + 356, + 194 + ], + "type": "image", + "image_path": "1f37a6e31454da20ead1a27cc944ddff3b9e03901684c02bf77ff54dab72da1b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 194, + 336, + 204 + ], + "lines": [ + { + "bbox": [ + 277, + 194, + 336, + 204 + ], + "spans": [ + { + "bbox": [ + 277, + 194, + 336, + 204 + ], + "type": "text", + "content": "(b) New Classes" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 357, + 116, + 467, + 192 + ], + "blocks": [ + { + "bbox": [ + 357, + 116, + 467, + 192 + ], + "lines": [ + { + "bbox": [ + 357, + 116, + 467, + 192 + ], + "spans": [ + { + "bbox": [ + 357, + 116, + 467, + 192 + ], + "type": "image", + "image_path": "25abf52628bfec146c877a8eca8aefbc7eb396330ef40be9e114c7154530921d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 194, + 446, + 205 + ], + "lines": [ + { + "bbox": [ + 381, + 194, + 446, + 205 + ], + "spans": [ + { + "bbox": [ + 381, + 194, + 446, + 205 + ], + "type": "text", + "content": "(c) Whole Classes" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 149, + 274, + 251, + 351 + ], + "blocks": [ + { + "bbox": [ + 131, + 213, + 482, + 258 + ], + "lines": [ + { + "bbox": [ + 131, + 213, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 131, + 213, + 482, + 258 + ], + "type": "text", + "content": "Fig. 5: Information bottleneck trade-off analysis. We compare representations acquired by three different methods by assessing the information bottleneck (IB) tradeoff. 'RS' refers to representation spreading methods. We indicate the final models with black edges. The experiments are conducted on the CIFAR100 dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 274, + 251, + 351 + ], + "lines": [ + { + "bbox": [ + 149, + 274, + 251, + 351 + ], + "spans": [ + { + "bbox": [ + 149, + 274, + 251, + 351 + ], + "type": "image", + "image_path": "ad9a42eb4ece99d9a9149c53b04dcf819804a07ef5a6f10d12d7bcd5ba109415.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 177, + 354, + 222, + 363 + ], + "lines": [ + { + "bbox": [ + 177, + 354, + 222, + 363 + ], + "spans": [ + { + "bbox": [ + 177, + 354, + 222, + 363 + ], + "type": "text", + "content": "(a) Baseline" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 373, + 483, + 450 + ], + "lines": [ + { + "bbox": [ + 130, + 373, + 483, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 483, + 450 + ], + "type": "text", + "content": "Fig. 6: T-SNE visualization of learned representations. The blue and red points indicate the base and new class samples, respectively. We measure " + }, + { + "bbox": [ + 130, + 373, + 483, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(f_{\\theta})" + }, + { + "bbox": [ + 130, + 373, + 483, + 450 + ], + "type": "text", + "content": " to quantify the transferability of the learned representation. We conducted the experiments on CIFAR100 with reduced classes (20 base classes and 10 new classes) for better visualization. The classification results for each experiment are as follows: (a): " + }, + { + "bbox": [ + 130, + 373, + 483, + 450 + ], + "type": "inline_equation", + "content": "A_B = 78.85\\%" + }, + { + "bbox": [ + 130, + 373, + 483, + 450 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 373, + 483, + 450 + ], + "type": "inline_equation", + "content": "A_N = 13.50\\%" + }, + { + "bbox": [ + 130, + 373, + 483, + 450 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 373, + 483, + 450 + ], + "type": "inline_equation", + "content": "A_W = 57.07\\%" + }, + { + "bbox": [ + 130, + 373, + 483, + 450 + ], + "type": "text", + "content": ". (b): " + }, + { + "bbox": [ + 130, + 373, + 483, + 450 + ], + "type": "inline_equation", + "content": "A_B = 77.85\\%" + }, + { + "bbox": [ + 130, + 373, + 483, + 450 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 373, + 483, + 450 + ], + "type": "inline_equation", + "content": "A_N = 25.50\\%" + }, + { + "bbox": [ + 130, + 373, + 483, + 450 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 373, + 483, + 450 + ], + "type": "inline_equation", + "content": "A_W = 60.22\\%" + }, + { + "bbox": [ + 130, + 373, + 483, + 450 + ], + "type": "text", + "content": ". (c): " + }, + { + "bbox": [ + 130, + 373, + 483, + 450 + ], + "type": "inline_equation", + "content": "A_B = 78.88\\%" + }, + { + "bbox": [ + 130, + 373, + 483, + 450 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 373, + 483, + 450 + ], + "type": "inline_equation", + "content": "A_N = 28.80\\%" + }, + { + "bbox": [ + 130, + 373, + 483, + 450 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 373, + 483, + 450 + ], + "type": "inline_equation", + "content": "A_W = 62.18\\%" + }, + { + "bbox": [ + 130, + 373, + 483, + 450 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 256, + 274, + 358, + 350 + ], + "blocks": [ + { + "bbox": [ + 256, + 274, + 358, + 350 + ], + "lines": [ + { + "bbox": [ + 256, + 274, + 358, + 350 + ], + "spans": [ + { + "bbox": [ + 256, + 274, + 358, + 350 + ], + "type": "image", + "image_path": "600483079d620d23719251f778d174dca8165cc3fd04aeddbf8108cff1c7b505.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 354, + 340, + 363 + ], + "lines": [ + { + "bbox": [ + 272, + 354, + 340, + 363 + ], + "spans": [ + { + "bbox": [ + 272, + 354, + 340, + 363 + ], + "type": "text", + "content": "(b) Baseline " + }, + { + "bbox": [ + 272, + 354, + 340, + 363 + ], + "type": "inline_equation", + "content": "+\\mathrm{RS}" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 363, + 274, + 465, + 350 + ], + "blocks": [ + { + "bbox": [ + 363, + 274, + 465, + 350 + ], + "lines": [ + { + "bbox": [ + 363, + 274, + 465, + 350 + ], + "spans": [ + { + "bbox": [ + 363, + 274, + 465, + 350 + ], + "type": "image", + "image_path": "71f15c504e72c9c2765c5970978dd103bbb1c3c0b65a78424d590260aa1855eb.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 354, + 450, + 363 + ], + "lines": [ + { + "bbox": [ + 376, + 354, + 450, + 363 + ], + "spans": [ + { + "bbox": [ + 376, + 354, + 450, + 363 + ], + "type": "text", + "content": "(c) CLOSER (Ours)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 478, + 371, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 478, + 371, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 478, + 371, + 491 + ], + "type": "text", + "content": "4.4 Information Bottleneck trade-off Analysis" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": "In Sec. 3.5, we show the connection between the proposed objective function and the information bottleneck (IB) trade-off. To validate our analysis, we measure the mutual information between representations and inputs " + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "I(X;Z)" + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": " and the mutual information between representations and targets " + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "I(Y;Z)" + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": " for the baseline, representation spread approach, and our method CLOSER, as shown in Fig. 5. In the figure, the more left (lower " + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "I(X;Z)" + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": ") and upper (higher " + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "I(Y;Z)" + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": ") regions imply better IB trade-off. Our proposed method CLOSER demonstrates to have found a better IB trade-off, especially when considering whole classes, including base and new classes. The results are encouraging in that CLOSER is able to find a better IB trade-off for all classes, when the feature extractor is only trained on base classes and fixed afterwards. The results demonstrate that CLOSER is effective in tackling a particularly difficult challenge of FSCIL: learning transferable and discriminative features from base classes." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "type": "text", + "content": "CLOSER: Representation Learning for Few-Shot Class-Incremental Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 241, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 241, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 241, + 129 + ], + "type": "text", + "content": "4.5 T-SNE Analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 482, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 482, + 281 + ], + "type": "text", + "content": "For qualitative evaluation, we visualize the learned representation trained by different configurations of the proposed losses, which is illustrated in Fig. 6. The value of " + }, + { + "bbox": [ + 133, + 138, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(f_{\\theta})" + }, + { + "bbox": [ + 133, + 138, + 482, + 281 + ], + "type": "text", + "content": " in the right below in each figure is for measuring the transferability of representation. The baseline representation (a) shows the characteristics of the base class, represented by the features of the new classes mapped to those of the base classes, also indicated by the low value of " + }, + { + "bbox": [ + 133, + 138, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(f_{\\theta})" + }, + { + "bbox": [ + 133, + 138, + 482, + 281 + ], + "type": "text", + "content": ". Spreading of features (b) largely resolves the overfitting issues, exhibited by larger distances between new classes and base classes; i.e., the increase in " + }, + { + "bbox": [ + 133, + 138, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(f_{\\theta})" + }, + { + "bbox": [ + 133, + 138, + 482, + 281 + ], + "type": "text", + "content": ". Finally, " + }, + { + "bbox": [ + 133, + 138, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{inter}}" + }, + { + "bbox": [ + 133, + 138, + 482, + 281 + ], + "type": "text", + "content": " is used to compensate for the decreased performance on the base classes due to the spread representation. Reducing inter-class distance also enhances the separability between the new class samples and the clusters of base classes, as evidenced by the further increase in " + }, + { + "bbox": [ + 133, + 138, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(f_{\\theta})" + }, + { + "bbox": [ + 133, + 138, + 482, + 281 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 301, + 318, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 301, + 318, + 316 + ], + "spans": [ + { + "bbox": [ + 133, + 301, + 318, + 316 + ], + "type": "text", + "content": "5 Limitations and Discussions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 329, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 329, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 133, + 329, + 482, + 449 + ], + "type": "text", + "content": "Our work focuses on learning representation that is discriminative yet transferable to unseen classes to tackle the challenges of FSCIL. As such, our work does not consider an option of updating representation with new classes. Continual update of representation can improve the performance on new classes, however at the cost of sacrificing the old-class performance. Thus, it is a question of stability-plasticity dilemma, where our method focuses more on stability while improving plasticity through discriminative and transferable representation. Furthermore, similar to other works on FSCIL, our method is limited to classification tasks. But, we believe our work can have an impact on other domains, akin to the impact of representation spread methods, such as contrastive learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 469, + 218, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 469, + 218, + 483 + ], + "spans": [ + { + "bbox": [ + 133, + 469, + 218, + 483 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 497, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 497, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 133, + 497, + 482, + 616 + ], + "type": "text", + "content": "To tackle convoluted challenges in few-shot class-incremental learning (FSCIL), we focus on representation learning, which plays a crucial role. In contrast to previous FSCIL methods that have focused on maximizing the distance between classes, our experimental and theoretic analysis suggests that the closer classes are, the better for FSCIL, especially when feature sharing is encouraged between classes. Upon the analysis, we propose a simple, yet seemingly counter-intuitive idea: bring classes closer for a better transferability-discriminability Pareto front. Our experimental results and information-bottleneck-theory-based analysis suggest that our work can provide a promising research avenue. As such, we hope that our work will inspire future works and discussions in this research direction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 617, + 482, + 665 + ], + "type": "text", + "content": "Acknowledgments. This work was supported in part by the IITP grants [No. 2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University), No.2021-0-02068, and No.2023-0-00156], the NRF grant [No.2021M3A 9E4080782] funded by the Korean government (MSIT)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 295, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 295, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 295, + 101 + ], + "type": "text", + "content": "J. Oh*, S. Baik*, and K.M. Lee" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "text", + "content": "1. Achituve, I., Navon, A., Yemini, Y., Chechik, G., Fetaya, E.: Gp-tree: A gaussian process classifier for few-shot incremental learning. In: ICML (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "type": "text", + "content": "2. Akyurek, A.F., Akyurek, E., Wijaya, D.T., Andreas, J.: Subspace regularizers for few-shot class incremental learning. In: ICLR (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 184, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 480, + 205 + ], + "type": "text", + "content": "3. Alemi, A.A., Fischer, I., Dillon, J.V., Murphy, K.: Deep variational information bottleneck. In: ICLR (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 206, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 480, + 226 + ], + "type": "text", + "content": "4. Belghazi, M.I., Baratin, A., Rajeshwar, S., Ozair, S., Bengio, Y., Courville, A., Hjelm, D.: Mutual information neural estimation. In: PMLR (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 228, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 248 + ], + "type": "text", + "content": "5. Chatfield, K., Simonyan, K., Vedaldi, A., Zisserman, A.: Return of the devil in the details: delving deep into convolutional nets. In: BMVC (2014)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 250, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 250, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 250, + 480, + 270 + ], + "type": "text", + "content": "6. Chen, K., Lee, C.G.: Incremental few-shot learning via vector quantization in deep embedded space. In: ICLR (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 303 + ], + "type": "text", + "content": "7. Chen, M.F., Fu, D.Y., Narayan, A., Zhang, M., Song, Z., Fatahalian, K., Ré, C.: Perfectly balanced: improving transfer and robustness of supervised contrastive learning. In: ICML (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 325 + ], + "type": "text", + "content": "8. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 480, + 357 + ], + "type": "text", + "content": "9. Cheraghian, A., Rahman, S., Fang, P., Roy, S.K., Petersson, L., Harandi, M.: Semantic-aware knowledge distillation for few-shot class-incremental learning. In: CVPR (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 358, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 358, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 138, + 358, + 480, + 390 + ], + "type": "text", + "content": "10. Cheraghian, A., Rahman, S., Ramasinghe, S., Fang, P., Simon, C., Petersson, L., Harandi, M.: Synthesized feature based few-shot class-incremental learning on a mixture of subspaces. In: ICCV (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 413 + ], + "type": "text", + "content": "1. Chi, Z., Gu, L., Liu, H., Wang, Y., Yu, Y., Tang, J.: Meta桅: a meta-learning approach for few-shot class incremental learning. In: CVPR (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 414, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 480, + 445 + ], + "type": "text", + "content": "2. Cui, Q., Zhao, B., Chen, Z.M., Zhao, B., Song, R., Liang, J., Zhou, B., Yoshie, O.: Discriminability-transferability trade-off: an information-theoretic perspective. In: ECCV (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 447, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 489 + ], + "type": "text", + "content": "3. Delange, M., Aljundi, R., Masana, M., Parisot, S., Jia, X., Leonardis, A., Slabaugh, G., Tuytelaars, T.: A continual learning survey: Defying forgetting in classification tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "type": "text", + "content": "4. Deng, J., Guo, J., Yang, J., Xue, N., Kotsia, I., Zafeiriou, S.: Arcface: additive angular margin loss for deep face recognition. In: CVPR (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "text", + "content": "15. Donahue, J., Jia, Y., Vinyals, O., Hoffman, J., Zhang, N., Tzeng, E., Darrellin, T.: Decaf: a deep convolutional activation feature for generic visual recognition. In: ICML (2014)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 545, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 566 + ], + "type": "text", + "content": "6. Dong, S., Hong, X., Tao, X., Chang, X., Wei, X., Gong, Y.: Few-shot class-incremental learning via relation knowledge distillation. In: AAAI (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "type": "text", + "content": "7. Geirhos, R., Jacobsen, J.H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., Wichmann, F.A.: Shortcut learning in deep neural networks. Nature Machine Intelligence 2(11), 665-673 (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 600, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 620 + ], + "type": "text", + "content": "8. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR (2020)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "text", + "content": "9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition In: CVPR (2016)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "20. Hersche, M., Karunaratne, G., Cherubini, G., Benini, L., Sebastian, A., Rahimi, A.: Constrained few-shot class-incremental learning. In: CVPR (2022)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "text", + "content": "CLOSER: Representation Learning for Few-Shot Class-Incremental Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "21. Islam, A., Chen, C.F., Panda, R., Karlinsky, L., Radke, R., Feris, R.: A broad study on the transferability of visual representations with contrastive learning. In: ICCV (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 172 + ], + "type": "text", + "content": "22. Kalla, J., Biswas, S.: S3c: self-supervised stochastic classifiers for few-shot class-incremental learning. In: ECCV (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 482, + 226 + ], + "type": "text", + "content": "23. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., Hassabis, D., Clopath, C., Kumaran, D., Hadsell, R.: Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences of the United States of America (PNAS) (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 482, + 248 + ], + "type": "text", + "content": "24. Koch, G., Zemel, R., Salakhutdinov, R.: Siamese neural networks for one-shot image recognition. In: ICMLW (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 482, + 270 + ], + "type": "text", + "content": "25. Kornblith, S., Chen, T., Lee, H., Norouzi, M.: Why do better loss functions lead to less transferable features? In: NeurIPS (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 482, + 293 + ], + "type": "text", + "content": "26. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images. University of Toronto (2009)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 482, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 482, + 315 + ], + "type": "text", + "content": "27. Liu, B., Cao, Y., Lin, Y., Li, Q., Zhang, Z., Long, M., Hu, H.: Negative margin matters: Understanding margin in few-shot classification. In: ECCV (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 316, + 482, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 316, + 482, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 482, + 336 + ], + "type": "text", + "content": "28. Liu, H., Gu, L., Chi, Z., Wang, Y., Yu, Y., Chen, J., Tang, J.: Few-shot class-incremental learning via entropy-regularized data-free replay. In: ECCV (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 337, + 482, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 482, + 357 + ], + "type": "text", + "content": "29. Ngampruetikorn, V., Schwab, D.J.: Information bottleneck theory of high-dimensional regression: relevancy, efficiency and optimality. In: NeurIPS (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 482, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 482, + 380 + ], + "type": "text", + "content": "30. van den Oord, A., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 482, + 413 + ], + "type": "text", + "content": "31. Papyan, V., Han, X., Donoho, D.L.: Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences of the United States of America (PNAS) (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 414, + 482, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 482, + 435 + ], + "type": "text", + "content": "32. Peng, C., Zhao, K., Wang, T., Li, M., Lovell, B.C.: Few-shot class-incremental learning from an open-set perspective. In: ECCV (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 435, + 482, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 482, + 456 + ], + "type": "text", + "content": "33. Peng, C., Zhao, K., Wang, T., Li, M., Lovell, B.C.: Few-shot class-incremental learning from an open-set perspective. In: ECCV (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 457, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 482, + 479 + ], + "type": "text", + "content": "34. Razavian, A.S., Azizpour, H., Sullivan, J., Carlsson, S.: Cnn features off-the-shelf: an astounding baseline for recognition. In: CVPRW (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 479, + 482, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 482, + 522 + ], + "type": "text", + "content": "35. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV) (2015)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 523, + 482, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 482, + 555 + ], + "type": "text", + "content": "36. Shi, G., Chen, J., Zhang, W., Zhan, L.M., Wu, X.M.: Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. In: NeurIPS (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 555, + 482, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 482, + 578 + ], + "type": "text", + "content": "37. Shwartz-Ziv, R., Tishby, N.: Opening the black box of deep neural networks via information (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 578, + 482, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 482, + 589 + ], + "type": "text", + "content": "38. Slonim, N., Tishby, N.: Agglomerative information bottleneck. In: NIPS (1999)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 590, + 482, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 482, + 610 + ], + "type": "text", + "content": "39. Snell, J., Swersky, K., Zemel, R.: Prototypical networks for few-shot learning. In: NIPS (2017)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 611, + 482, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 482, + 643 + ], + "type": "text", + "content": "40. Song, Z., Zhao, Y., Shi, Y., Peng, P., Yuan, L., Tian, Y.: Learning with fantasy: semantic-aware virtual contrastive constraint for few-shot class-incremental learning. In: CVPR (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 644, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 482, + 665 + ], + "type": "text", + "content": "41. Tao, X., Hong, X., Chang, X., Dong, S., Wei, X., Gong, Y.: Few-shot class-incremental learning. In: CVPR (2020)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 296, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 296, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 296, + 101 + ], + "type": "text", + "content": "J. Oh*, S. Baik*, and K.M. Lee" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 456 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "42. Tishby, N., Zaslavsky, N.: Deep learning and the information bottleneck principle. In: 2015 IEEE Information Theory Workshop (ITW). pp. 1-5 (2015). https://doi.org/10.1109/ITW.2015.7133169" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 172 + ], + "type": "text", + "content": "43. Vinyals, O., Blundell, C., Lillicrap, T., Kavukcuoglu, K., Wierstra, D.: Matching networks for one shot learning. In: NIPS (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "text", + "content": "44. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The Caltech-UCSD Birds-200-2011 Dataset. Tech. rep., Caltech (2011)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "45. Wang, H., Wang, Y., Zhou, Z., Ji, X., Gong, D., Zhou, J., Li, Z., Liu, W.: Cosface: large margin cosine loss for deep face recognition. In: CVPR (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "46. Xue, Y., Joshi, S., Gan, E., Chen, P.Y., Mirzasoleiman, B.: Which features are learnt by contrastive learning? on the role of simplicity bias in class collapse and feature suppression. In: ICML (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 248, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 481, + 270 + ], + "type": "text", + "content": "47. Yang, Y., Yuan, H., Li, X., Lin, Z., Torr, P., Tao, D.: Neural collapse inspired feature-classifier alignment for few-shot class incremental learning. In: ICLR (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "type": "text", + "content": "48. Zhang, C., Song, N., Lin, G., Zheng, Y., Pan, P., Xu, Y.: Few-shot incremental learning with continually evolved classifiers. In: CVPR (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 315 + ], + "type": "text", + "content": "49. Zhao, L., Lu, J., Xu, Y., Cheng, Z., Guo, D., Niu, Y., Fang, X.: Few-shot class-incremental learning via class-aware bilateral distillation. In: CVPR (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "type": "text", + "content": "50. Zheng, Y., Pal, D.K., Savvides, M.: Ring loss: convex feature normalization for face recognition. In: CVPR (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 335, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 335, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 335, + 481, + 357 + ], + "type": "text", + "content": "51. Zhou, D.W., Wang, F.Y., Ye, H.J., Ma, L., Pu, S., Zhan, D.C.: Forward compatible few-shot class-incremental learning. In: CVPR (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 357, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 481, + 390 + ], + "type": "text", + "content": "52. Zhou, D.W., Ye, H.J., Ma, L., Xie, D., Pu, S., Zhan, D.C.: Few-shot class-incremental learning by sampling multi-phase tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 390, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 390, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 390, + 481, + 413 + ], + "type": "text", + "content": "53. Zhu, K., Cao, Y., Zhai, W., Cheng, J., Zha, Z.J.: Self-promoted prototype refinement for few-shot class-incremental learning. In: CVPR (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "type": "text", + "content": "54. Zhuang, H., Weng, Z., He, R., Lin, Z., Zeng, Z.: Gkeal: gaussian kernel embedded analytic learning for few-shot class incremental task. In: CVPR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "text", + "content": "55. Zou, Y., Zhang, S., Li, Y., Li, R.: Margin-based few-shot class-incremental learning with class-level overfitting mitigation. In: NeurIPS (2022)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "type": "text", + "content": "CLOSER: Representation Learning for Few-Shot Class-Incremental Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CLR-GAN_ Improving GANs Stability and Quality via Consistent Latent Representation and Reconstruction/c724f22f-d6fb-423d-a760-902dfb6a8a1d_content_list.json b/2024/CLR-GAN_ Improving GANs Stability and Quality via Consistent Latent Representation and Reconstruction/c724f22f-d6fb-423d-a760-902dfb6a8a1d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fd815b7dec2147345b1582e9fbfcd335059011e3 --- /dev/null +++ b/2024/CLR-GAN_ Improving GANs Stability and Quality via Consistent Latent Representation and Reconstruction/c724f22f-d6fb-423d-a760-902dfb6a8a1d_content_list.json @@ -0,0 +1,1799 @@ +[ + { + "type": "text", + "text": "CLR-GAN: Improving GANs Stability and Quality via Consistent Latent Representation and Reconstruction", + "text_level": 1, + "bbox": [ + 217, + 140, + 787, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shengke Sun1*, Ziqian Luan2*, Zhanshan Zhao1,3, Shijie Luo1, and Shuzhen Han1", + "bbox": [ + 233, + 234, + 771, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tiangong University, 300387 Tianjin, China {sunshengke,hanshuzhen,luoshijie}@tiangong.edu.cn, zhzhsh127@163.com", + "bbox": [ + 236, + 276, + 764, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Xidian University, 710126 Xi'an, Shaanxi, China", + "bbox": [ + 331, + 306, + 669, + 318 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tyuwe@foxmail.com", + "bbox": [ + 429, + 321, + 573, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Institute for Sustainable Industries & Liveable Cities, Victory University, Melbourne, VIC 8001, Australia \n†Corresponding author", + "bbox": [ + 250, + 333, + 750, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Generative Adversarial Networks(GANs) have received considerable attention due to its outstanding ability to generate images. However, training a GAN is hard since the game between the Generator(G) and the Discriminator(D) is unfair. Towards making the competition fairer, we propose a new perspective of training GANs, named Consistent Latent Representation and Reconstruction(CLR-GAN). In this paradigm, we treat the G and D as an inverse process, the discriminator has an additional task to restore the pre-defined latent code while the generator also needs to reconstruct the real input, thus obtaining a relationship between the latent space of G and the out-features of D. Based on this prior, we can put D and G on an equal position during training using a new criterion. Experimental results on various datasets and architectures prove our paradigm can make GANs more stable and generate better quality images(31.22% gain of FID on CIFAR10 and 39.5% on AFHQ-Cat, respectively). We hope that the proposed perspective can inspire researchers to explore different ways of viewing GANs training, rather than being limited to a two-player game. The code is publicly available at https://github.com/Petecheco/CLR-GAN.", + "bbox": [ + 259, + 415, + 743, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Generative Adversarial Networks $\\cdot$ Image generation $\\cdot$ Latent space consistency", + "bbox": [ + 259, + 679, + 740, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 734, + 375, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative models are an important part of computer vision, they can be widely used in fields such as image generation [12, 16, 20, 24], video editing [3, 5, 48, 50], style transfer [9, 27, 55] and even drug discovery [7, 35]. Among various generative", + "bbox": [ + 212, + 768, + 787, + 815 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. Listing order is random.", + "bbox": [ + 217, + 824, + 529, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/de061eb9c0fdf1bad66007ead92185b8e996efcac31478bf912a30d288a3203d.jpg", + "image_caption": [ + "Fig. 1: Sketch map of the proposed method, we proposed two additional learning objectives to consider D and G as inverse process. The top is latent consistency loss where D tries to rebuild the latent distribution of G. The bottom is reconstruction loss where G tries to restore the original real images." + ], + "image_footnote": [], + "bbox": [ + 236, + 152, + 772, + 372 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "models, Generative Adversarial Networks(GANs) have received considerable attention due to its remarkable performance in generating realistic images [12]. A conventional GAN comprises of two individual networks named generator(G) and discriminator(D). While training a GAN, the generator aims to produce indistinguishable images, and the discriminator tries to figure out the synthesized images from the input images. By the competition between D and G, ideally, we can eventually train a generator that can recover the real distribution and generate high quality samples.", + "bbox": [ + 212, + 476, + 787, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, it is well known that training a GAN is hard and unstable [26,37]. In general, the instability of training GANs can be attributed to the unequal set up of generator(G) and discriminator(D). Specifically, in the traditional paradigm, the training process is formulated as a two-player game. But in fact, the discriminator seems to dominate the game [2,47]. In practice, the discriminator can easily identify the fake samples from a relatively early stage of the training and maintain this advantage during the whole training, making the generator hard to converge [51]. Many attempts have been done to improve the stability of training GANs. Previous studies attempted to find new metric functions [1, 14, 32, 45], enhance the generator's ability to generate [4,21,40], or use data augment methods [19,30]. These solutions have shown some degree of improvement on training stability, but they did not fundamentally change the core relationship between the generator and discriminator, while ignoring the essence of dynamic balance within GANs architecture. The competition between the generator and discriminator is still unfair. In addition, existing methods have not fully utilized the structural information of the latent space, which limits the further improvement", + "bbox": [ + 212, + 598, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S.Sun, Z.Luan et al.", + "bbox": [ + 271, + 114, + 406, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of GANs generation ability. Therefore, improving the training of GANs still remains an unsolved and challenging problem.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, contrast to existing methods, we proposed a novel perspective of training GANs. By rethinking the training process and network architecture of GANs. We found that the generator and the discriminator got optimized in a considerably large latent space, resulting unstable training and inconsistency between the generator and the discriminator. While traditional methods primarily focus on adjusting network architectures or tuning hyperparameters to stabilize training [19, 40]. Our method is to view the generator and the discriminator as two inverse components. As shown in Fig. 1. In particular, the discriminator can be regarded as outputting a realness score and extracting samples to the latent representation space at the same time. And the generator not only generates images but restores source images from the extracted latent representation. By constraining the generator and discriminator with these additional tasks, we make the discriminator's representation latent space closer to the generator's latent space, thus making the generator and the discriminator more consistent, leading a fairer game between the generator and the discriminator. This not only provides a more robust framework for training GANs but also opens up new possibilities for improving GANs training from different aspects.", + "bbox": [ + 212, + 178, + 787, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 459, + 387, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Improving GANs through discriminators", + "text_level": 1, + "bbox": [ + 214, + 494, + 599, + 510 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As we mentioned above, due to the limitations of original training objective [1, 12], the discriminator tends to easily take over the competition during the early stage of training. Prior works have been devoted to improving the inequality using modified discriminators. Dist-GAN [45] decelerate the convergence of the discriminator by constructing a reconstruction constrain using a auxiliary autoencoder. Wang et al. [47] combined Grad-CAM [42] with GANs to improve the spatial awareness of the generator using the discriminator as a regularizer, lessening the information gap between D and G. DynamicGAN [51] controls the strength of the discriminator by gradually increasing D's learning capacity, thereby preventing the discriminator from being overly dominant in the early stages of training. AdaptiveMix [32] shrinks the feature space of the discriminator by training using constructed hard samples, reducing the feature space of the discriminator and making training easier. There are also some data augmentation methods that enhance model training by providing more data to the model [19, 46]. However, the above methods have not changed the inner properties of training. All gradient during the optimization process is provided by D, so D always occupies a more dominant position during the training process, which will prevent further optimization of the model. While our model, based on limiting D, also constrains D to optimize based on the gradient provided by G. This allows both G and D to optimize themselves using the gradients provided by each other, making the game fairer.", + "bbox": [ + 212, + 522, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "CLR-GAN", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Latent space of GANs", + "text_level": 1, + "bbox": [ + 215, + 146, + 444, + 161 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The latent space of GANs has always been a popular research area due to its rich semantic information. Multi-Code GAN [13] combines multiple latent feature maps with adaptive channels to improve the quality of image reconstruction. ClusterGAN [38] combines latent space encoding and one-hot encoding to obtain a better clustering result. StyleMapGAN [23] utilizes latent space as an accurate feature embedding, improving the performance of image editing and interpolation tasks. Although latent space plays an important role in various downstream tasks. But to our best knowledge, there has been no previous work that directly uses the latent space of GANs as a component to optimize the training of GANs. In our proposed method, we have directly incorporated latent space into the training process of GANs as an additional constrain. The experimental results show that applying this method can effectively improve the quality and diversity of generated images.", + "bbox": [ + 212, + 169, + 787, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Differences to Reconstruction-based Methods", + "text_level": 1, + "bbox": [ + 215, + 387, + 635, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prior works have explored the application of reconstruction-based methods within GANs. HoloGAN [39] and PiGAN [6] utilize latent reconstruction to learn the 3D-aware image representations. GLeaD [2] takes advantage of image reconstruction to improve the training dynamics of GANs. However, these methods either lack a specific focus on stabilizing GANs training or solely concentrate on improving the generator without considering the discriminator. Specifically, in the GLeaD paradigm, given an image, we expect D to extract representative features that can be adequately decoded by G to reconstruct the input, however, this paradigm only improves the generator, the discriminator still controls the game. To address these limitations, we use the latent spatial code extracted by the discriminator and reconstruction loss to achieve stable GAN training and generate higher quality and more diverse images. Thus making GANs training more efficient and practical.", + "bbox": [ + 212, + 410, + 787, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 628, + 330, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As we mentioned before, we try to augment the training fairness of GANs via considering the generative process of the generator (G) and the discriminative process of the discriminator (D) as inverse to each other, thereby enhancing the stability and quality of generated samples. To better understand this new perspective, we first review the conventional formulation of GANs in Sec. 3.1. To obtain a fairer training, Sec. 3.2 introduces our new training paradigm named Consistent Latent Representation that considers G and D as an inverse process. Furthermore, in Sec. 3.3, we also propose a Real Image Reconstruction strategy to achieve mutual connection between generator and the discriminator. Then Sec. 3.4 explained our method from a regularization perspective. Finally, Sec. 3.5 gives the full optimization objective and shows the pseudo code of the new training paradigm.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S.Sun, Z.Luan et al.", + "bbox": [ + 271, + 114, + 406, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminary", + "text_level": 1, + "bbox": [ + 215, + 146, + 359, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A GAN usually consists of two components: a generator $G(\\cdot)$ and a discriminator $D(\\cdot)$ . The generator aims to map a latent code $z$ into an image, while the discriminator tries to distinguish the generated image $G(z)$ from the real one $x$ . The conventional GAN trains via a two-player game by optimizing the learning objective as follows:", + "bbox": [ + 212, + 167, + 782, + 243 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {G} \\max _ {D} V (D, G) = \\mathbb {E} _ {x \\sim p _ {d a t a} (x)} [ \\log D (x) ] + \\mathbb {E} _ {z \\sim p _ {z} (z)} [ \\log (1 - D (G (z))) ] \\quad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 252, + 784, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p_{z}(z)$ and $p_{data}(x)$ represent a random latent representation and real data distribution respectively.", + "bbox": [ + 212, + 280, + 782, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Goodfellow et al. [12] have proved that the optimal solution to this objective is that $G$ can eventually recover the distribution of source data and $D$ can not separate the generated images out from real images. However, the game between two players is not fair since in Eq. 1 that the gradient for optimizing all comes from $D$ , which makes the discriminator a natural dominant [2]. Thus the ideal solution is hard to reach in practice. [11,47]", + "bbox": [ + 212, + 310, + 784, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Consistent Latent Space", + "text_level": 1, + "bbox": [ + 215, + 420, + 462, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We aim to improve the stability of GANs by introducing a novel perspective, that is to consider the generate and discriminate as an inverse process. Recall that in a regular GAN paradigm, the generator and discriminator are treated as two individual components. While in our proposed view, we have refined the concepts of generator and discriminator and consider them as an inverse process. In our perspective, the generator maps a latent representation to the real data distribution, and the discriminator not only outputs a realizness score, but it also transform the high-dimensional distribution into a trainable latent space. By measuring the distance between these two spaces, we can improve GANs training in the following aspects: First, we can to some extent make the generator and discriminator behave more consistent in the mapping process. Second, since the discriminator needs to align with the generator, it can be seen as constraining the discriminator using generator, which makes the game not completely dominated by discriminator. The detailed implementation is shown in Fig. 2.", + "bbox": [ + 212, + 441, + 787, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Extracting Latent Representation through D. In the training of GANs, the discriminator plays a role of downsample high-dimensional data. For any discriminator, it can be divided into two parts, the first part is the feature extractor denoted as $f(\\cdot)$ which extracts features from high-dimensional data samples, another serves as a affine from extracted features to realizness score to tell whether the input is true(real samples) or false(generated samples). For any input data $x$ and latent code $z$ , the whole discriminate process can be denoted as follows:", + "bbox": [ + 212, + 672, + 787, + 792 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD (x) = R (f (x)) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 801, + 784, + 816 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD (z) = R (f (G (z))) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 433, + 820, + 784, + 835 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "CLR-GAN", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ce52d80ecda9abf83c85b0e4a10a31fc0295544a528d3f3e87522430bf962476.jpg", + "image_caption": [ + "Fig. 2: Detailed implementation of Consistent Latent Representation. Given a latent code $z$ , the generator maps $z$ to an image. The synthesised image was then got discriminated by the discriminator, outputting a realness score and a reconstructed latent code $z'$ . The Consistent Latent loss is calculated between the real latent $z$ and the reconstructed latent $z'$ . The right part specifically shows the architecture of D." + ], + "image_footnote": [], + "bbox": [ + 228, + 150, + 781, + 253 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $f(\\cdot)$ is the feature extractor module of discriminator and $R(\\cdot)$ is the affine module to transform feature space to the realness score.", + "bbox": [ + 212, + 375, + 787, + 404 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Through this discriminate process, we can transform structural high-dimensional input into feature map that contains abundant semantic information [31], thus judging the realness of the input data. We discovered both latent code in the generator and feature map in the discriminator contain lots semantic features, making it possible to align the generator and discriminator using a mapping function. The mapping function converts discriminator's feature space to latent space. Finally we can get a reconstructed latent code $\\phi (x)$ from discriminator as follows:", + "bbox": [ + 212, + 407, + 808, + 526 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\phi (x) = \\sigma (f (x)) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 444, + 547, + 785, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $x$ is the real input, $f(\\cdot)$ is the feature extractor of D and $\\sigma(\\cdot)$ is a mapping function that transform feature to latent space, we use an identical affine in this paper for simplicity.", + "bbox": [ + 212, + 575, + 787, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Consistent Latent Representation Loss. As we mentioned above, we now have a real latent code $(z)$ from the generator and a reconstructed latent code(Eq.4) from the discriminator. The latent representation loss can be defined by measuring the distance between these two latent spaces:", + "bbox": [ + 212, + 651, + 795, + 712 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C L R} = \\operatorname {D i s t} (\\phi (x), z) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 727, + 785, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\text{Dist}(\\cdot)$ is distance measurement for any two distributions. In this paper, we use L1-Distance similar to [55] for all distance measurement, so the $\\mathcal{L}_{\\text{clr}}$ can be rewritten as:", + "bbox": [ + 212, + 757, + 787, + 801 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C L R} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\| z _ {i} - \\phi (x _ {i}) \\| _ {1} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 804, + 785, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S.Sun, Z.Luan et al.", + "bbox": [ + 271, + 114, + 406, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Real Image Reconstruction", + "text_level": 1, + "bbox": [ + 215, + 146, + 485, + 161 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As mentioned in Sec. 2, in a regular GAN, the generator optimized through the gradient offered by discriminator. Also during the training, the generator has no access to real data distributions, it can only update its parameters implicitly, while the discriminator can get optimized according to real data distribution. This unfair situation helps the discriminator converge faster at an early stage of the training [47], thus the discriminator can not provide useful gradient in the later stage of training, making the generator not perform well.", + "bbox": [ + 212, + 172, + 787, + 277 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To solve this problem, we proposed a new strategy called Real Image Reconstruction, which enables the generator to also get gradient from the real data distribution. Specifically, since we have strengthened the discriminator to output both realizness and latent code in the last section, we can then use latent code as a bridge to indirectly connect the generator with real distribution. Subsequently, we can use the generated images for a image reconstruction task. The generator then can utilize this gradient that comes from real distribution for better synthesis, making the whole process more efficient. In practice, we can use Eq. 4 to get the reconstructed latent code of real distributions, then we use generator for image restoration.", + "bbox": [ + 212, + 279, + 787, + 429 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nI _ {r e c} = G (\\phi (x)) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 449, + 430, + 784, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After reconstructing the image, we can calculate the distance between real images and reconstructed images. And for the reconstruction loss, we use a regular from [55] as follows:", + "bbox": [ + 214, + 453, + 785, + 497 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\| I (x) - I _ {r e c} \\| _ {1} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 500, + 784, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $I(x)$ is the source image data and $I_{rec}$ is the reconstructed image using generator and reconstructed latent code.", + "bbox": [ + 214, + 544, + 787, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Enhancing Training Stability via Regularization Techniques", + "text_level": 1, + "bbox": [ + 214, + 597, + 751, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To better understand the superiority of our method, we turn to investigate the above training strategies from the perspective of regularization in this section.", + "bbox": [ + 214, + 623, + 785, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With our method, the objective of training GANs can be rewritten as:", + "bbox": [ + 238, + 654, + 746, + 669 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {G} \\max _ {D} V (D, G) = \\mathbb {E} _ {x \\sim p _ {d a t a}} \\log D (x) + \\mathbb {E} _ {z \\sim p _ {z}} \\log [ 1 - D (G (z)) ] \\tag {9} \\\\ + \\mathbb {E} _ {x \\sim p _ {d a t a}} \\| G (D (x)) - x \\| - \\mathbb {E} _ {x, z} \\| D (x) - z \\| \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 272, + 681, + 785, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The first two terms of Eq. 9 are the default training objectives for GANs, while the third and fourth terms correspond to Latent Consistency Loss and Real Image Reconstruction Loss, respectively. Note that these two objectives actually appear in the form of L1 norm, and we will consider them as two types of regularization in this section.", + "bbox": [ + 212, + 733, + 787, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "When optimizing networks, we actually calculate the gradients of objective function. For the training objective we propose, the gradients can be expressed", + "bbox": [ + 214, + 809, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "CLR-GAN", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "as:", + "bbox": [ + 215, + 148, + 240, + 159 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\theta_ {d}, \\theta_ {g}} V (G, D) = \\mathbb {E} _ {x \\sim p _ {d a t a}} [ \\nabla_ {\\theta_ {d}} \\log D (x) ] + \\mathbb {E} _ {z \\sim p _ {z}} [ \\nabla_ {\\theta_ {d}, \\theta_ {g}} \\log D (1 - D (G (z))) ] \\\\ + \\mathbb {E} _ {x \\sim p _ {d a t a}} \\operatorname {s i g n} (G (D (x)) - x) - \\mathbb {E} _ {x, z} \\operatorname {s i g n} (D (x) - z) \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 167, + 785, + 217 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\operatorname{sign}(G(D(x)) - x)$ and $\\operatorname{sign}(D(x) - z)$ is simply the sign of both constrains applied element-wise.", + "bbox": [ + 212, + 224, + 782, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Next, we will discuss the role of our regularization in two different scenarios:", + "bbox": [ + 238, + 255, + 784, + 270 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training the Discriminator with a fixed Generator. When the generator is fixed, we can consider the gradients related to the generator to be 0, so we can obtain the gradients as follows:", + "bbox": [ + 214, + 289, + 784, + 334 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta_ {d}} V (D) = \\mathbb {E} _ {x \\sim p _ {d a t a}} \\left[ \\nabla_ {\\theta_ {d}} \\log D (x) \\right] - \\mathbb {E} _ {x, z} \\operatorname {s i g n} (D (x) - z) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 303, + 361, + 784, + 376 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Considering the general training situation of GANs, when the discriminator is too strong during training. The first term in the gradient tends to be large, resulting in unstable gradient descent. However with our regularization term. Due to the negative sign in the front, we can reduce the gradient of D using this term, making the gradient smaller and can stable the training.", + "bbox": [ + 214, + 385, + 784, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training the Generator with a fixed Discriminator. As in the previous case, we first write the representation for the gradient:", + "bbox": [ + 214, + 479, + 784, + 508 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta_ {g}} V (G) = \\mathbb {E} _ {z \\sim p _ {z}} [ \\nabla_ {\\theta_ {d}, \\theta_ {g}} \\log D (1 - D (G (z))) ] + \\mathbb {E} _ {x \\sim p _ {d a t a}} \\operatorname {s i g n} (G (D (x)) - x) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 534, + 784, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When the discriminator is dominating the game, it can make the first term to become very small, making it difficult for the generator to effectively optimize and generate real images. After adding a regularization term, we found an additional training objective was given to the generator, allowing the generator to optimize through this training objective, and continue to compete with the discriminator instead of collapsing.", + "bbox": [ + 212, + 571, + 782, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we can find that by adding two additional training objective, our model can penalize the discriminator when it is too strong. It can optimize the generator by providing more information through additional tasks, making the game fairer and improving the effectiveness of image generation.", + "bbox": [ + 212, + 662, + 782, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Full Objective", + "text_level": 1, + "bbox": [ + 215, + 741, + 379, + 756 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With the enhanced discriminator and the reconstruction guided generator, the overall training objective can be rewritten as:", + "bbox": [ + 214, + 763, + 782, + 792 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D} ^ {\\prime} = \\mathcal {L} _ {D} + \\lambda_ {1} \\mathcal {L} _ {C L R} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 801, + 782, + 816 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {G} ^ {\\prime} = \\mathcal {L} _ {G} + \\lambda_ {2} \\mathcal {L} _ {r e c} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 821, + 782, + 835 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S.Sun, Z.Luan et al.", + "bbox": [ + 271, + 114, + 406, + 127 + ], + "page_idx": 7 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Algorithm 1 Training a GAN with Consistent Latent Representation and Reconstruction" + ], + "code_body": "Require: G and our D that are initialized with random parameters. Training data $\\{x_{i}\\}$ , latent variable $\\{z\\}$ for $t = 1$ until converge do Sample $z\\sim P(Z),x\\sim P(X)$ real_score, real_forent $= D(x)$ fake_score, fake_forent $= D(G(z))$ reconstruct_images $= G$ (real_forent) Calculate Consistent Latent Loss with Eq.(6) Calculate Reconstruction Loss with Eq. (8) Update D and G with Eq. (9), Eq. (10) end for \nOutput: G with best training set FID.", + "guess_lang": "verilog", + "bbox": [ + 215, + 176, + 785, + 344 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_G$ and $\\mathcal{L}_D$ are the original loss function of generator and discriminator, $\\mathcal{L}_{rec}$ and $\\mathcal{L}_{CLR}$ are the proposed consistent latent space distance and real image reconstruction distance calculated through Eq. 6 and Eq. 8, $\\lambda_1$ and $\\lambda_2$ are weighting coefficient of the additional loss.", + "bbox": [ + 212, + 373, + 784, + 433 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "With this updated objective, we can stabilize GANs training in the following two aspects: First, we align the generator and discriminator during training phase by assigning a consistent latent representation task for the discriminator, which makes the game between generator and discriminator fairer. Second, we enable the generator to get access to the real data through a reconstruction task. This helps generator make full use of alternative information, leading to a more realistic output. The pseudo code of our method is shown in Algorithm 1. Extensive experiments result in Sec 4. showed the effectiveness of our method.", + "bbox": [ + 212, + 434, + 785, + 554 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 574, + 375, + 590 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To validate the effectiveness of the proposed method, we conducted extensive experiments on various GAN architectures and datasets. Sec. 4.1 first introduces the detailed settings of our experiments. In Sec. 4.2, we proved that our method achieves better quality and stability through quantitative evaluation on various image generation datasets. Sec. 4.3 includes ablation study to the proposed Consistent Latent loss and Reconstruction loss to better understand the designed components. In Sec. 4.4 we prove that we can get a better feature extractor by unsupervised classification. At last, we visualize the realness curve of D to validate the improved fairness of GANs in Sec. 4.5. The generated samples can be found in the Supplementary Material.", + "bbox": [ + 212, + 603, + 787, + 753 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 772, + 426, + 789 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. For a more comprehensive evaluation of the proposed method, the model was tested on low resolution images and high resolution images separately. We used DCGAN [40] as the basic architecture for low resolution image", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "CLR-GAN", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/774e4e130ccc7b279695047a810647193fd8861c2e3bb3eb51cada7ebe76f3cb.jpg", + "table_caption": [ + "Table 1: FIDs of DCGAN [40] using different learning objectives on CelebA [33] and CIFAR-10 [25] dataset. The bold numbers indicate the best result for each dataset." + ], + "table_footnote": [], + "table_body": "
Learning ObjectiveCelebACIFAR-10
WGAN [1]36.4755.96
HingeGAN [54]25.5742.4
LSGAN [36]30.7642.01
DCGAN [40]27.0238.56
WGAN-GP [14]70.2841.86
Realness GAN-Obj.1 [49]-36.73
Realness GAN-Obj.2 [49]23.5134.59
Realness GAN-Obj.3 [49]-36.21
AdaptiveMix [32]12.4330.85
CLR-GAN(Ours)13.6323.3
", + "bbox": [ + 300, + 184, + 702, + 340 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "generation, with CIFAR-10 [25] dataset and Celeba64 $\\times$ 64 [33] dataset. For high-resolution image generation, we used StyleGAN-V2 [22] as the basic architecture, and test the generation ability under AFHQ Cat [8], LSUN Church [53], and FFHQ [21] dataset corresponding to animal generation, natural scene generation, and face generation respectively.", + "bbox": [ + 212, + 366, + 784, + 444 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation. During evaluation phase, we mainly used the most common Frechet Inception Distance(FID) [15] for quantitatively represent the image generation quality. Besides, Precision & Recall [29] is also adopted to provide more details of the realness and diversity of the proposed model. In practice, we calculate FID between 20K generated images and all real samples on CIFAR-10 and CelebA dataset. While for AFHQ-Cat, LSUN-Church and FFHQ, we calculate the FID and P&R between 50K generated images and all real samples.", + "bbox": [ + 212, + 462, + 787, + 570 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Other settings. For all baselines, our experimental hyperparameters were set according to the original paper. We used pre-trained Inception-V3 [43] as the feature extractor for FID and P&R calculation. For the weights of the additional loss, we set $\\lambda_1 = 5$ and $\\lambda_2 = 0.5$ .", + "bbox": [ + 212, + 587, + 787, + 647 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Performance on Image Generation", + "text_level": 1, + "bbox": [ + 214, + 666, + 547, + 681 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Low-resolution Image Generation. We first conducted experiments on low resolution image datasets to verify the effectiveness of the proposed method. We used DCGAN [40] as the baseline architecture and compared our method with some well-known learning objectives on CIFAR10 and CelebA datasets.", + "bbox": [ + 212, + 688, + 785, + 748 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tab. 1 presents the results. From the perspective of FID, we can directly see a significant margin between the proposed method and previous learning objectives. Furthermore, compared with the recently proposed learning objective [32], our method outperforms $24.5\\%$ on the CIFAR-10 dataset while remains a competitive result on CIFAR-10 dataset $(8\\%)$ behind), which further demonstrates the superiority of our model.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S.Sun, Z.Luan et al.", + "bbox": [ + 271, + 114, + 406, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e5d155e39b3c5e53605018b757aa0c9f23287726310eb33fc7926c09a7a671f3.jpg", + "table_caption": [ + "Table 2: Comparisons on AFHQ-Cat [8], FFHQ [21] and LSUN-Church [53] with different high-resolution image generative models. P and R denote precision and recall. The bold numbers indicate the best result for each dataset. The blue numbers indicate the improvements." + ], + "table_footnote": [], + "table_body": "
MethodAFHQ-CatFFHQLSUN Church
FID ↓P ↑R ↑FID ↓P ↑R ↑FID ↓P ↑R ↑
StyleGAN-V2 [22]7.920.680.273.860.680.254.040.580.40
LC-Reg [46]6.70--3.93--4.07--
StyleGAN-V2-ADA [19]6.050.660.254.010.660.264.010.610.43
StyleGAN-V2-APA [18]4.880.650.303.750.670.293.920.600.43
StyleGAN-V2 + Ours4.79(-3.13)0.760.283.44(-0.42)0.700.413.52(-0.52)0.630.46
StyleGAN-V2-ADA + Ours4.45(-0.43)0.740.333.37(-0.64)0.710.443.43(-0.58)0.610.48
", + "bbox": [ + 218, + 210, + 784, + 338 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "High-resolution Image Generation. To further demonstrate the effectiveness of the proposed method, we also considered the recently proposed high-resolution image synthesis architecture StyleGAN-V2 [22]. We compared our method with other StyleGAN-V2 variants methods on AFHQ-Cat, LSUN Church and FFHQ datasets, respectively. We also combined our method with recent proposed works that aims to improve GANs training called StyleGAN-V2-ADA [19] to show the compatibility of our method.", + "bbox": [ + 212, + 366, + 782, + 470 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As shown in Tab. 2, there is a substantially improvement with the proposed method on StyleGAN-V2 for both datasets, outperforming other methods with a clear margin. We also found that combined with StyleGAN-V2-ADA, our method can get a further improvement, achieving a better result on various datasets. Therefore, our proposed method can improve the synthesis quality of high-resolution images while also integrating well with previous works.", + "bbox": [ + 212, + 473, + 784, + 563 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As for the Precision and Recall. The improvements among multiple datasets are clear, indicating that our model can not only learn to generate more realistic images, but also enhance the diversity of generated images, to some extent solving the problem of insufficient diversity of conventional GANs. This may suggest that CLR-GAN can benefit from the proposed two constrains that leads to learning a wider distribution, thereby improving the diversity of generated samples. More validation metrics and experimental results on the more complex ImageNet [41] dataset are shown in the Supplementary Material.", + "bbox": [ + 212, + 564, + 787, + 684 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 707, + 388, + 722 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In the ablation experiment, we primarily focus on the impact of the two new training constrains we added during training. To ensure the generality, we selected both the low-resolution CIFAR-10 dataset and the high-resolution FFHQ dataset as two baseline datasets. Recall that we can use $\\lambda_{1}$ and $\\lambda_{2}$ to control the strength of the proposed constrains in Eq. 13. Therefore we can modify different $\\lambda$ values in the ablation study to figure out the optimal values of two constrains. We first set $\\lambda_{1} = \\lambda_{2} = 0$ to get the baseline values.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "CLR-GAN", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/47df8aed4b9651d83105f94782aebf89371cb51b556ad2cdf4fb4231f600e24c.jpg", + "table_caption": [ + "Table 3: Ablation studies of different weights $\\lambda_{1}$ and $\\lambda_{2}$ on low-resolution CIFAR-10 dataset and high-resolution FFHQ dataset. The best FID of each dataset are marked in bold." + ], + "table_footnote": [ + "(a) CIFAR-10 @ $32 \\times 32$" + ], + "table_body": "
λ1λ2FID
0038.56(baseline)
1025.15
0126.09
5125.20
10125.17
15126.30
5525.55
50.523.3
", + "bbox": [ + 241, + 196, + 480, + 328 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/dc16f7387d436d2bcbdfcd4263dc74a4d4e672e3e99615dd312232f36eb9a153.jpg", + "table_caption": [], + "table_footnote": [ + "(b) FFHQ @ $256 \\times 256$" + ], + "table_body": "
λ1λ2FID
003.86(baseline)
103.75
013.77
513.62
1013.84
1514.01
553.80
50.53.44
", + "bbox": [ + 527, + 198, + 756, + 328 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The FID scores are shown in Tab. 3. We can observe that when we set $\\lambda_{1} = 0$ or $\\lambda_{2} = 0$ , the performance of the network significantly drops $9.96\\%$ on CIFAR-10 dataset. The performance even falls below the baseline model on high resolution dataset. Only when both the latent consistent loss and the reconstruction loss are combined together does the model gets its best performance. This proves that the two constrains we added are indispensable with each other.", + "bbox": [ + 212, + 387, + 787, + 477 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Next we will find the optimal strength of the proposed constrains. We first fix $\\lambda_{2} = 1$ and change the value of $\\lambda_{1}$ to find the ideal value of $\\lambda_{1}$ . During the experiments, we found that a relatively high $\\lambda_{1}$ makes the model unable to converge, so we chose 15 as the maximum value. As shown in Tab. 3, $\\lambda_{1} = 5$ turns out to be the best. Consequently, we then fix $\\lambda_{1} = 5$ and search for the best $\\lambda_{2}$ . Ultimately, $\\lambda_{1} = 5$ and $\\lambda_{2} = 0.5$ turns out to be the best strategy.", + "bbox": [ + 212, + 479, + 787, + 569 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "During the search for optimal parameters, we observed some interesting phenomena. As shown in Fig. 3. Let $\\eta = \\frac{\\lambda_1}{\\lambda_2}$ . We found a higher value of $\\eta$ leads to a better result in the early stage of training. However, an excessively high $\\eta$ might lead to poor convergence at the final stages of model training. We infer that when $\\eta$ is large, our model that constrained by the latent distance, can quickly converge to a distribution similar to the real distribution, hence leads to an early convergence. But overly tight constraints might prevent the model's ability to generalize effectively, resulting in a lack of diversity in the generated data thus making the final performance poor. The underlying causes of this phenomenon require further investigation in future research.", + "bbox": [ + 212, + 570, + 787, + 720 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Performance on Unsupervised Visual Recognition", + "text_level": 1, + "bbox": [ + 214, + 750, + 671, + 763 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "GANs are often regarded as an unsupervised training method. To further demonstrate the effectiveness of the proposed method, we did experiment to show our proposed method leads to a better feature extractor. We take the pre-trained discriminator as a feature extractor and add a linear classification head to perform", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S.Sun, Z.Luan et al.", + "bbox": [ + 271, + 114, + 406, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ff7bc499da0340d06845c11206046bd0e7ff0ec508cc9fbcf511276636db84f2.jpg", + "image_caption": [ + "Fig. 3: Training curves at different $\\eta$" + ], + "image_footnote": [], + "bbox": [ + 220, + 147, + 782, + 363 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/d917e2c2fd06e8f1470aeff6e6c55dc5214357b87162a6c73c274debb57e37e3.jpg", + "table_caption": [ + "Table 4: Classification accuracy on CIFAR-10 dataset [25] with various methods with linear classifier." + ], + "table_footnote": [], + "table_body": "
Method#ParametersAccuracy
ViT-H [10]632M99.5%
EfficientNetV2-M [44]55M99%
PyramidNet [28]26M98.6%
NAT-M4 [34]6.2M98.4%
HCGNet [52]3.1M97.7%
CLR-GAN(Ours)3.02M98.8%
", + "bbox": [ + 295, + 443, + 709, + 546 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "classification task on the CIFAR-10 dataset. We compared our feature extractor with other classification models. The results are shown in Tab. 4.", + "bbox": [ + 212, + 570, + 784, + 599 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As listed in Tab. 4. Our proposed pre-train feature extractor can reach a relatively high accuracy after simple fine-tuning. Compared with models which have similar parameter sizes ( $\\leq 10M$ ) [34,52], our model is $0.4\\%$ and $1.1\\%$ higher respectively. In comparison with models several orders of magnitude higher than our parameter size [10], our model is also very competitive, only $0.7\\%$ lower than the best model. This indicates that our strategy can not only improve the quality and diversity of generated images in generative adversarial networks, but also serve as an effective unsupervised training method, benefiting downstream tasks.", + "bbox": [ + 212, + 601, + 787, + 736 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Towards a fairer game", + "text_level": 1, + "bbox": [ + 215, + 756, + 447, + 772 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As we mentioned in the introduction part, the fundamental purpose of our method is to produce more realistic and diverse images by making the training of GANs fairer. By the proposed consistent latent representation and reconstruction, we can shrink the distance between the generator and discriminator.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "CLR-GAN", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/343da88656d16b9e3f248d949aee833c05182e74e0cc25813bda8a6775fb23fc.jpg", + "image_caption": [ + "Fig. 4: Realness scores of the real samples and synthesis samples by various discriminators during training. We visualize the realness score of StyleGAN-V2 [22], StyleGAN-V2-ADA [19] and the proposed CLR-GAN." + ], + "image_footnote": [], + "bbox": [ + 251, + 147, + 504, + 339 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d556ee9befe05df20e0b6b8f735f109d684cd2bb94aab8b46d53eb57893e16af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 148, + 751, + 337 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this section, we demonstrate the enhanced fairness of our method through experiments.", + "bbox": [ + 212, + 424, + 782, + 453 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In order to better quantitatively judge the fairness of our model, we followed the previous two works [2, 47], using the score of the discriminator to represent the realness of the image. We visualize the realness score the model trained on LSUN-Church. We used exponentially weighted averages [17] on the original data for better visualization. The left part of Fig. 4 shows the score of the synthesis images, while the right part shows the score of the real images. We can clearly see that using our CLR-GAN, the realness gap between the generated image and the real image is much smaller than the baseline. While for StyleGAN-V2-ADA, the gap between scores is similar to the baseline, so it only improves FID and can not make the game fairer. We can then draw a conclusion that with the aid of the proposed auxiliary objective functions, the realness scores of real images and generated images become closer, which means that CLR-GAN can improve the realness and diversity of generated images by improving training fairness.", + "bbox": [ + 212, + 455, + 787, + 650 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 672, + 359, + 690 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we proposed a simple, effective and plug and play new objective function for GANs training. The new objective function simultaneously constrains the output distribution of the generator and the discriminator to make them more consistent. Thereby making the competition between the generator and the discriminator more fair. Eventually improving the quality and diversity of generated images. In addition to image generator, we also demonstrated that the discriminator with constrain is an efficient unsupervised feature extractor. Experimental results demonstrate that our proposed method improves the performance of baseline models over various datasets.", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S.Sun, Z.Luan et al.", + "bbox": [ + 271, + 114, + 406, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Arjovsky, M., Chintala, S., Bottou, L.: Wasserstein generative adversarial networks. In: International conference on machine learning. pp. 214-223. PMLR (2017)", + "2. Bai, Q., Yang, C., Xu, Y., Liu, X., Yang, Y., Shen, Y.: Glead: Improving gans with a generator-leading task. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12094-12104 (June 2023)", + "3. Bar-Tal, O., Ofri-Amar, D., Fridman, R., Kasten, Y., Dekel, T.: Text2live: Text-driven layered image and video editing. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 707-723. Springer Nature Switzerland, Cham (2022)", + "4. Brock, A., Donahue, J., Simonyan, K.: Large scale gan training for high fidelity natural image synthesis (2019)", + "5. Ceylan, D., Huang, C.H.P., Mitra, N.J.: Pix2video: Video editing using image diffusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 23206-23217 (October 2023)", + "6. Chan, E.R., Monteiro, M., Kellnhofer, P., Wu, J., Wetzstein, G.: pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5799-5809 (2021)", + "7. Chenthamarakshan, V., Das, P., Hoffman, S., Strobelt, H., Padhi, I., Lim, K.W., Hoover, B., Manica, M., Born, J., Laino, T., Mojsilovic, A.: Cogmol: Target-specific and selective drug design for Covid-19 using deep generative models. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 4320-4332. Curran Associates, Inc. (2020)", + "8. Choi, Y., Uh, Y., Yoo, J., Ha, J.W.: Stargan v2: Diverse image synthesis for multiple domains. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8188-8197 (2020)", + "9. Deng, Y., Tang, F., Dong, W., Ma, C., Pan, X., Wang, L., Xu, C.: Stytr2: Image style transfer with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11326-11336 (June 2022)", + "0. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale (2021)", + "1. Farnia, F., Ozdaglar, A.: Do GANs always have Nash equilibria? In: III, H.D., Singh, A. (eds.) Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 3029-3039. PMLR (13-18 Jul 2020)", + "2. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NeurIPS. vol. 27 (2014)", + "3. Gu, J., Shen, Y., Zhou, B.: Image processing using multi-code gan prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)", + "4. Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A.C.: Improved training of wasserstein gans. Advances in neural information processing systems 30 (2017)", + "5. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + ], + "bbox": [ + 225, + 178, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "CLR-GAN", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 6840-6851. Curran Associates, Inc. (2020)", + "17. Hunter, J.S.: The exponentially weighted moving average. Journal of Quality Technology 18(4), 203-210 (1986)", + "18. Jiang, L., Dai, B., Wu, W., Loy, C.C.: Deceive d: Adaptive pseudo augmentation for gan training with limited data. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems (NeurIPS). vol. 34, pp. 21655-21667. Curran Associates, Inc. (2021)", + "19. Karras, T., Aittala, M., Hellsten, J., Laine, S., Lehtinen, J., Aila, T.: Training generative adversarial networks with limited data (2020)", + "20. Karras, T., Aittala, M., Laine, S., Härkönen, E., Hellsten, J., Lehtinen, J., Aila, T.: Alias-free generative adversarial networks. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 852-863. Curran Associates, Inc. (2021)", + "21. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2019)", + "22. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8110-8119 (2020)", + "23. Kim, H., Choi, Y., Kim, J., Yoo, S., Uh, Y.: Exploiting spatial dimensions of latent in gan for real-time image editing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 852-861 (June 2021)", + "24. Kingma, D.P., Welling, M.: Auto-encoding variational bayes (2022)", + "25. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)", + "26. Kurach, K., Lucic, M., Zhai, X., Michalski, M., Gelly, S.: A large-scale study on regularization and normalization in GANs. In: Chaudhuri, K., Salakhutdinov, R. (eds.) Proceedings of the 36th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 97, pp. 3581-3590. PMLR (09-15 Jun 2019)", + "27. Kwon, G., Ye, J.C.: Clipstyler: Image style transfer with a single text condition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18062-18071 (June 2022)", + "28. Kwon, J., Kim, J., Park, H., Choi, I.K.: Asam: Adaptive sharpness-aware minimization for scale-invariant learning of deep neural networks (2021)", + "29. Kynkänniemi, T., Karras, T., Laine, S., Lehtinen, J., Aila, T.: Improved precision and recall metric for assessing generative models. Advances in Neural Information Processing Systems 32 (2019)", + "30. Lee, G., Kim, H., Kim, J., Kim, S., Ha, J.W., Choi, Y.: Generator knows what discriminator should learn in unconditional gans. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 406-422. Springer Nature Switzerland, Cham (2022)", + "31. Liao, W., Hu, K., Yang, M.Y., Rosenhahn, B.: Text to image generation with semantic-spatial aware gan. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18187-18196 (June 2022)", + "32. Liu, H., Zhang, W., Li, B., Wu, H., He, N., Huang, Y., Li, Y., Ghanem, B., Zheng, Y.: Adaptivemix: Improving gan training via feature space shrinkage. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16219-16229 (June 2023)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S.Sun, Z.Luan et al.", + "bbox": [ + 271, + 114, + 405, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: Proceedings of the IEEE international conference on computer vision. pp. 3730-3738 (2015)", + "34. Lu, Z., Sreekumar, G., Goodman, E., Banzhaf, W., Deb, K., Boddeti, V.N.: Neural architecture transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(9), 2971-2989 (Sep 2021)", + "35. Luo, S., Guan, J., Ma, J., Peng, J.: A 3d generative model for structure-based drug design. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 6229-6239. Curran Associates, Inc. (2021)", + "36. Mao, X., Li, Q., Xie, H., Lau, R.Y., Wang, Z., Paul Smolley, S.: Least squares generative adversarial networks. In: Proceedings of the IEEE international conference on computer vision. pp. 2794-2802 (2017)", + "37. Miyato, T., Kataoka, T., Koyama, M., Yoshida, Y.: Spectral normalization for generative adversarial networks (2018)", + "38. Mukherjee, S., Asnani, H., Lin, E., Kannan, S.: Clustergan: Latent space clustering in generative adversarial networks. Proceedings of the AAAI Conference on Artificial Intelligence 33(01), 4610-4617 (Jul 2019)", + "39. Nguyen-Phuoc, T., Li, C., Theis, L., Richardt, C., Yang, Y.L.: Hologan: Unsupervised learning of 3d representations from natural images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7588-7597 (2019)", + "40. Radford, A., Metz, L., Chintala, S.: Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434 (2015)", + "41. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision 115, 211-252 (2015)", + "42. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (Oct 2017)", + "43. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision (2015)", + "44. Tan, M., Le, Q.: Efficientnetv2: Smaller models and faster training. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 10096-10106. PMLR (18-24 Jul 2021)", + "45. Tran, N.T., Bui, T.A., Cheung, N.M.: Dist-gan: An improved gan using distance constraints. In: Proceedings of the European Conference on Computer Vision (ECCV) (September 2018)", + "46. Tseng, H.Y., Jiang, L., Liu, C., Yang, M.H., Yang, W.: Regularizing generative adversarial networks under limited data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7921-7931 (June 2021)", + "47. Wang, J., Yang, C., Xu, Y., Shen, Y., Li, H., Zhou, B.: Improving gan equilibrium by raising spatial awareness. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11285-11293 (June 2022)", + "48. Wu, J.Z., Ge, Y., Wang, X., Lei, S.W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7623-7633 (October 2023)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "CLR-GAN", + "bbox": [ + 656, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. Xiangli, Y., Deng, Y., Dai, B., Loy, C.C., Lin, D.: Real or not real, that is the question. arXiv preprint arXiv:2002.05512 (2020)", + "50. Xu, Y., AlBahar, B., Huang, J.B.: Temporally consistent semantic video editing. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 357-374. Springer Nature Switzerland, Cham (2022)", + "51. Yang, C., Shen, Y., Xu, Y., Zhao, D., Dai, B., Zhou, B.: Improving gans with a dynamic discriminator. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. vol. 35, pp. 15093-15104. Curran Associates, Inc. (2022)", + "52. Yang, C., An, Z., Zhu, H., Hu, X., Zhang, K., Xu, K., Li, C., Xu, Y.: Gated convolutional networks with hybrid connectivity for image classification. Proceedings of the AAAI Conference on Artificial Intelligence 34(07), 12581-12588 (Apr 2020)", + "53. Yu, F., Seff, A., Zhang, Y., Song, S., Funkhouser, T., Xiao, J.: Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365 (2015)", + "54. Zhao, J., Mathieu, M., LeCun, Y.: Energy-based generative adversarial network. arXiv preprint arXiv:1609.03126 (2016)", + "55. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (Oct 2017)" + ], + "bbox": [ + 215, + 146, + 784, + 424 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "S.Sun, Z.Luan et al.", + "bbox": [ + 271, + 114, + 406, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/CLR-GAN_ Improving GANs Stability and Quality via Consistent Latent Representation and Reconstruction/c724f22f-d6fb-423d-a760-902dfb6a8a1d_model.json b/2024/CLR-GAN_ Improving GANs Stability and Quality via Consistent Latent Representation and Reconstruction/c724f22f-d6fb-423d-a760-902dfb6a8a1d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c38f37ce43ffde2c15dad76bbb7ed81b29a59b19 --- /dev/null +++ b/2024/CLR-GAN_ Improving GANs Stability and Quality via Consistent Latent Representation and Reconstruction/c724f22f-d6fb-423d-a760-902dfb6a8a1d_model.json @@ -0,0 +1,2414 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.141, + 0.788, + 0.207 + ], + "angle": 0, + "content": "CLR-GAN: Improving GANs Stability and Quality via Consistent Latent Representation and Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.234, + 0.235, + 0.772, + 0.266 + ], + "angle": 0, + "content": "Shengke Sun1*, Ziqian Luan2*, Zhanshan Zhao1,3, Shijie Luo1, and Shuzhen Han1" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.277, + 0.765, + 0.307 + ], + "angle": 0, + "content": "\\(^{1}\\) Tiangong University, 300387 Tianjin, China {sunshengke,hanshuzhen,luoshijie}@tiangong.edu.cn, zhzhsh127@163.com" + }, + { + "type": "text", + "bbox": [ + 0.333, + 0.307, + 0.671, + 0.319 + ], + "angle": 0, + "content": "\\(^{2}\\) Xidian University, 710126 Xi'an, Shaanxi, China" + }, + { + "type": "text", + "bbox": [ + 0.43, + 0.323, + 0.574, + 0.334 + ], + "angle": 0, + "content": "tyuwe@foxmail.com" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.334, + 0.751, + 0.376 + ], + "angle": 0, + "content": "3 Institute for Sustainable Industries & Liveable Cities, Victory University, Melbourne, VIC 8001, Australia \n†Corresponding author" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.416, + 0.744, + 0.668 + ], + "angle": 0, + "content": "Abstract. Generative Adversarial Networks(GANs) have received considerable attention due to its outstanding ability to generate images. However, training a GAN is hard since the game between the Generator(G) and the Discriminator(D) is unfair. Towards making the competition fairer, we propose a new perspective of training GANs, named Consistent Latent Representation and Reconstruction(CLR-GAN). In this paradigm, we treat the G and D as an inverse process, the discriminator has an additional task to restore the pre-defined latent code while the generator also needs to reconstruct the real input, thus obtaining a relationship between the latent space of G and the out-features of D. Based on this prior, we can put D and G on an equal position during training using a new criterion. Experimental results on various datasets and architectures prove our paradigm can make GANs more stable and generate better quality images(31.22% gain of FID on CIFAR10 and 39.5% on AFHQ-Cat, respectively). We hope that the proposed perspective can inspire researchers to explore different ways of viewing GANs training, rather than being limited to a two-player game. The code is publicly available at https://github.com/Petecheco/CLR-GAN." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.68, + 0.741, + 0.709 + ], + "angle": 0, + "content": "Keywords: Generative Adversarial Networks \\(\\cdot\\) Image generation \\(\\cdot\\) Latent space consistency" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.736, + 0.376, + 0.751 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.769, + 0.788, + 0.816 + ], + "angle": 0, + "content": "Generative models are an important part of computer vision, they can be widely used in fields such as image generation [12, 16, 20, 24], video editing [3, 5, 48, 50], style transfer [9, 27, 55] and even drug discovery [7, 35]. Among various generative" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.531, + 0.84 + ], + "angle": 0, + "content": "* Equal contribution. Listing order is random." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.407, + 0.129 + ], + "angle": 0, + "content": "S.Sun, Z.Luan et al." + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.153, + 0.773, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.39, + 0.788, + 0.447 + ], + "angle": 0, + "content": "Fig. 1: Sketch map of the proposed method, we proposed two additional learning objectives to consider D and G as inverse process. The top is latent consistency loss where D tries to rebuild the latent distribution of G. The bottom is reconstruction loss where G tries to restore the original real images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.788, + 0.599 + ], + "angle": 0, + "content": "models, Generative Adversarial Networks(GANs) have received considerable attention due to its remarkable performance in generating realistic images [12]. A conventional GAN comprises of two individual networks named generator(G) and discriminator(D). While training a GAN, the generator aims to produce indistinguishable images, and the discriminator tries to figure out the synthesized images from the input images. By the competition between D and G, ideally, we can eventually train a generator that can recover the real distribution and generate high quality samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.842 + ], + "angle": 0, + "content": "However, it is well known that training a GAN is hard and unstable [26,37]. In general, the instability of training GANs can be attributed to the unequal set up of generator(G) and discriminator(D). Specifically, in the traditional paradigm, the training process is formulated as a two-player game. But in fact, the discriminator seems to dominate the game [2,47]. In practice, the discriminator can easily identify the fake samples from a relatively early stage of the training and maintain this advantage during the whole training, making the generator hard to converge [51]. Many attempts have been done to improve the stability of training GANs. Previous studies attempted to find new metric functions [1, 14, 32, 45], enhance the generator's ability to generate [4,21,40], or use data augment methods [19,30]. These solutions have shown some degree of improvement on training stability, but they did not fundamentally change the core relationship between the generator and discriminator, while ignoring the essence of dynamic balance within GANs architecture. The competition between the generator and discriminator is still unfair. In addition, existing methods have not fully utilized the structural information of the latent space, which limits the further improvement" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLR-GAN" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "of GANs generation ability. Therefore, improving the training of GANs still remains an unsolved and challenging problem." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.435 + ], + "angle": 0, + "content": "In this paper, contrast to existing methods, we proposed a novel perspective of training GANs. By rethinking the training process and network architecture of GANs. We found that the generator and the discriminator got optimized in a considerably large latent space, resulting unstable training and inconsistency between the generator and the discriminator. While traditional methods primarily focus on adjusting network architectures or tuning hyperparameters to stabilize training [19, 40]. Our method is to view the generator and the discriminator as two inverse components. As shown in Fig. 1. In particular, the discriminator can be regarded as outputting a realness score and extracting samples to the latent representation space at the same time. And the generator not only generates images but restores source images from the extracted latent representation. By constraining the generator and discriminator with these additional tasks, we make the discriminator's representation latent space closer to the generator's latent space, thus making the generator and the discriminator more consistent, leading a fairer game between the generator and the discriminator. This not only provides a more robust framework for training GANs but also opens up new possibilities for improving GANs training from different aspects." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.46, + 0.388, + 0.476 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.495, + 0.6, + 0.511 + ], + "angle": 0, + "content": "2.1 Improving GANs through discriminators" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As we mentioned above, due to the limitations of original training objective [1, 12], the discriminator tends to easily take over the competition during the early stage of training. Prior works have been devoted to improving the inequality using modified discriminators. Dist-GAN [45] decelerate the convergence of the discriminator by constructing a reconstruction constrain using a auxiliary autoencoder. Wang et al. [47] combined Grad-CAM [42] with GANs to improve the spatial awareness of the generator using the discriminator as a regularizer, lessening the information gap between D and G. DynamicGAN [51] controls the strength of the discriminator by gradually increasing D's learning capacity, thereby preventing the discriminator from being overly dominant in the early stages of training. AdaptiveMix [32] shrinks the feature space of the discriminator by training using constructed hard samples, reducing the feature space of the discriminator and making training easier. There are also some data augmentation methods that enhance model training by providing more data to the model [19, 46]. However, the above methods have not changed the inner properties of training. All gradient during the optimization process is provided by D, so D always occupies a more dominant position during the training process, which will prevent further optimization of the model. While our model, based on limiting D, also constrains D to optimize based on the gradient provided by G. This allows both G and D to optimize themselves using the gradients provided by each other, making the game fairer." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.407, + 0.129 + ], + "angle": 0, + "content": "S.Sun, Z.Luan et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.446, + 0.162 + ], + "angle": 0, + "content": "2.2 Latent space of GANs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.367 + ], + "angle": 0, + "content": "The latent space of GANs has always been a popular research area due to its rich semantic information. Multi-Code GAN [13] combines multiple latent feature maps with adaptive channels to improve the quality of image reconstruction. ClusterGAN [38] combines latent space encoding and one-hot encoding to obtain a better clustering result. StyleMapGAN [23] utilizes latent space as an accurate feature embedding, improving the performance of image editing and interpolation tasks. Although latent space plays an important role in various downstream tasks. But to our best knowledge, there has been no previous work that directly uses the latent space of GANs as a component to optimize the training of GANs. In our proposed method, we have directly incorporated latent space into the training process of GANs as an additional constrain. The experimental results show that applying this method can effectively improve the quality and diversity of generated images." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.388, + 0.637, + 0.403 + ], + "angle": 0, + "content": "2.3 Differences to Reconstruction-based Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.789, + 0.608 + ], + "angle": 0, + "content": "Prior works have explored the application of reconstruction-based methods within GANs. HoloGAN [39] and PiGAN [6] utilize latent reconstruction to learn the 3D-aware image representations. GLeaD [2] takes advantage of image reconstruction to improve the training dynamics of GANs. However, these methods either lack a specific focus on stabilizing GANs training or solely concentrate on improving the generator without considering the discriminator. Specifically, in the GLeaD paradigm, given an image, we expect D to extract representative features that can be adequately decoded by G to reconstruct the input, however, this paradigm only improves the generator, the discriminator still controls the game. To address these limitations, we use the latent spatial code extracted by the discriminator and reconstruction loss to achieve stable GAN training and generate higher quality and more diverse images. Thus making GANs training more efficient and practical." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.629, + 0.331, + 0.645 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As we mentioned before, we try to augment the training fairness of GANs via considering the generative process of the generator (G) and the discriminative process of the discriminator (D) as inverse to each other, thereby enhancing the stability and quality of generated samples. To better understand this new perspective, we first review the conventional formulation of GANs in Sec. 3.1. To obtain a fairer training, Sec. 3.2 introduces our new training paradigm named Consistent Latent Representation that considers G and D as an inverse process. Furthermore, in Sec. 3.3, we also propose a Real Image Reconstruction strategy to achieve mutual connection between generator and the discriminator. Then Sec. 3.4 explained our method from a regularization perspective. Finally, Sec. 3.5 gives the full optimization objective and shows the pseudo code of the new training paradigm." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLR-GAN" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.36, + 0.162 + ], + "angle": 0, + "content": "3.1 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.169, + 0.784, + 0.244 + ], + "angle": 0, + "content": "A GAN usually consists of two components: a generator \\(G(\\cdot)\\) and a discriminator \\(D(\\cdot)\\). The generator aims to map a latent code \\(z\\) into an image, while the discriminator tries to distinguish the generated image \\(G(z)\\) from the real one \\(x\\). The conventional GAN trains via a two-player game by optimizing the learning objective as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.24, + 0.253, + 0.785, + 0.274 + ], + "angle": 0, + "content": "\\[\n\\min _ {G} \\max _ {D} V (D, G) = \\mathbb {E} _ {x \\sim p _ {d a t a} (x)} [ \\log D (x) ] + \\mathbb {E} _ {z \\sim p _ {z} (z)} [ \\log (1 - D (G (z))) ] \\quad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.281, + 0.784, + 0.31 + ], + "angle": 0, + "content": "where \\( p_{z}(z) \\) and \\( p_{data}(x) \\) represent a random latent representation and real data distribution respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.785, + 0.402 + ], + "angle": 0, + "content": "Goodfellow et al. [12] have proved that the optimal solution to this objective is that \\( G \\) can eventually recover the distribution of source data and \\( D \\) can not separate the generated images out from real images. However, the game between two players is not fair since in Eq. 1 that the gradient for optimizing all comes from \\( D \\), which makes the discriminator a natural dominant [2]. Thus the ideal solution is hard to reach in practice. [11,47]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.421, + 0.463, + 0.435 + ], + "angle": 0, + "content": "3.2 Consistent Latent Space" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.788, + 0.653 + ], + "angle": 0, + "content": "We aim to improve the stability of GANs by introducing a novel perspective, that is to consider the generate and discriminate as an inverse process. Recall that in a regular GAN paradigm, the generator and discriminator are treated as two individual components. While in our proposed view, we have refined the concepts of generator and discriminator and consider them as an inverse process. In our perspective, the generator maps a latent representation to the real data distribution, and the discriminator not only outputs a realizness score, but it also transform the high-dimensional distribution into a trainable latent space. By measuring the distance between these two spaces, we can improve GANs training in the following aspects: First, we can to some extent make the generator and discriminator behave more consistent in the mapping process. Second, since the discriminator needs to align with the generator, it can be seen as constraining the discriminator using generator, which makes the game not completely dominated by discriminator. The detailed implementation is shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.793 + ], + "angle": 0, + "content": "Extracting Latent Representation through D. In the training of GANs, the discriminator plays a role of downsample high-dimensional data. For any discriminator, it can be divided into two parts, the first part is the feature extractor denoted as \\( f(\\cdot) \\) which extracts features from high-dimensional data samples, another serves as a affine from extracted features to realizness score to tell whether the input is true(real samples) or false(generated samples). For any input data \\( x \\) and latent code \\( z \\), the whole discriminate process can be denoted as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.433, + 0.803, + 0.785, + 0.818 + ], + "angle": 0, + "content": "\\[\nD (x) = R (f (x)) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.434, + 0.821, + 0.785, + 0.837 + ], + "angle": 0, + "content": "\\[\nD (z) = R (f (G (z))) \\tag {3}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.407, + 0.128 + ], + "angle": 0, + "content": "S.Sun, Z.Luan et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.151, + 0.782, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.343 + ], + "angle": 0, + "content": "Fig. 2: Detailed implementation of Consistent Latent Representation. Given a latent code \\( z \\), the generator maps \\( z \\) to an image. The synthesised image was then got discriminated by the discriminator, outputting a realness score and a reconstructed latent code \\( z' \\). The Consistent Latent loss is calculated between the real latent \\( z \\) and the reconstructed latent \\( z' \\). The right part specifically shows the architecture of D." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.788, + 0.405 + ], + "angle": 0, + "content": "where \\( f(\\cdot) \\) is the feature extractor module of discriminator and \\( R(\\cdot) \\) is the affine module to transform feature space to the realness score." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.809, + 0.527 + ], + "angle": 0, + "content": "Through this discriminate process, we can transform structural high-dimensional input into feature map that contains abundant semantic information [31], thus judging the realness of the input data. We discovered both latent code in the generator and feature map in the discriminator contain lots semantic features, making it possible to align the generator and discriminator using a mapping function. The mapping function converts discriminator's feature space to latent space. Finally we can get a reconstructed latent code \\(\\phi (x)\\) from discriminator as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.445, + 0.549, + 0.786, + 0.567 + ], + "angle": 0, + "content": "\\[\n\\phi (x) = \\sigma (f (x)) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.577, + 0.788, + 0.623 + ], + "angle": 0, + "content": "where \\( x \\) is the real input, \\( f(\\cdot) \\) is the feature extractor of D and \\( \\sigma(\\cdot) \\) is a mapping function that transform feature to latent space, we use an identical affine in this paper for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.796, + 0.713 + ], + "angle": 0, + "content": "Consistent Latent Representation Loss. As we mentioned above, we now have a real latent code \\((z)\\) from the generator and a reconstructed latent code(Eq.4) from the discriminator. The latent representation loss can be defined by measuring the distance between these two latent spaces:" + }, + { + "type": "equation", + "bbox": [ + 0.42, + 0.728, + 0.787, + 0.745 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C L R} = \\operatorname {D i s t} (\\phi (x), z) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.758, + 0.788, + 0.802 + ], + "angle": 0, + "content": "where \\( \\text{Dist}(\\cdot) \\) is distance measurement for any two distributions. In this paper, we use L1-Distance similar to [55] for all distance measurement, so the \\( \\mathcal{L}_{\\text{clr}} \\) can be rewritten as:" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.805, + 0.787, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C L R} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\| z _ {i} - \\phi (x _ {i}) \\| _ {1} \\tag {6}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLR-GAN" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.486, + 0.162 + ], + "angle": 0, + "content": "3.3 Real Image Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.279 + ], + "angle": 0, + "content": "As mentioned in Sec. 2, in a regular GAN, the generator optimized through the gradient offered by discriminator. Also during the training, the generator has no access to real data distributions, it can only update its parameters implicitly, while the discriminator can get optimized according to real data distribution. This unfair situation helps the discriminator converge faster at an early stage of the training [47], thus the discriminator can not provide useful gradient in the later stage of training, making the generator not perform well." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.28, + 0.788, + 0.43 + ], + "angle": 0, + "content": "To solve this problem, we proposed a new strategy called Real Image Reconstruction, which enables the generator to also get gradient from the real data distribution. Specifically, since we have strengthened the discriminator to output both realizness and latent code in the last section, we can then use latent code as a bridge to indirectly connect the generator with real distribution. Subsequently, we can use the generated images for a image reconstruction task. The generator then can utilize this gradient that comes from real distribution for better synthesis, making the whole process more efficient. In practice, we can use Eq. 4 to get the reconstructed latent code of real distributions, then we use generator for image restoration." + }, + { + "type": "equation", + "bbox": [ + 0.45, + 0.431, + 0.785, + 0.447 + ], + "angle": 0, + "content": "\\[\nI _ {r e c} = G (\\phi (x)) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.454, + 0.786, + 0.498 + ], + "angle": 0, + "content": "After reconstructing the image, we can calculate the distance between real images and reconstructed images. And for the reconstruction loss, we use a regular from [55] as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.5, + 0.785, + 0.539 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\| I (x) - I _ {r e c} \\| _ {1} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.545, + 0.788, + 0.575 + ], + "angle": 0, + "content": "where \\( I(x) \\) is the source image data and \\( I_{rec} \\) is the reconstructed image using generator and reconstructed latent code." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.598, + 0.753, + 0.615 + ], + "angle": 0, + "content": "3.4 Enhancing Training Stability via Regularization Techniques" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.624, + 0.786, + 0.654 + ], + "angle": 0, + "content": "To better understand the superiority of our method, we turn to investigate the above training strategies from the perspective of regularization in this section." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.655, + 0.747, + 0.67 + ], + "angle": 0, + "content": "With our method, the objective of training GANs can be rewritten as:" + }, + { + "type": "equation", + "bbox": [ + 0.273, + 0.682, + 0.786, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {G} \\max _ {D} V (D, G) = \\mathbb {E} _ {x \\sim p _ {d a t a}} \\log D (x) + \\mathbb {E} _ {z \\sim p _ {z}} \\log [ 1 - D (G (z)) ] \\tag {9} \\\\ + \\mathbb {E} _ {x \\sim p _ {d a t a}} \\| G (D (x)) - x \\| - \\mathbb {E} _ {x, z} \\| D (x) - z \\| \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.809 + ], + "angle": 0, + "content": "The first two terms of Eq. 9 are the default training objectives for GANs, while the third and fourth terms correspond to Latent Consistency Loss and Real Image Reconstruction Loss, respectively. Note that these two objectives actually appear in the form of L1 norm, and we will consider them as two types of regularization in this section." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "When optimizing networks, we actually calculate the gradients of objective function. For the training objective we propose, the gradients can be expressed" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.407, + 0.128 + ], + "angle": 0, + "content": "S.Sun, Z.Luan et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.15, + 0.241, + 0.16 + ], + "angle": 0, + "content": "as:" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.168, + 0.786, + 0.218 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\theta_ {d}, \\theta_ {g}} V (G, D) = \\mathbb {E} _ {x \\sim p _ {d a t a}} [ \\nabla_ {\\theta_ {d}} \\log D (x) ] + \\mathbb {E} _ {z \\sim p _ {z}} [ \\nabla_ {\\theta_ {d}, \\theta_ {g}} \\log D (1 - D (G (z))) ] \\\\ + \\mathbb {E} _ {x \\sim p _ {d a t a}} \\operatorname {s i g n} (G (D (x)) - x) - \\mathbb {E} _ {x, z} \\operatorname {s i g n} (D (x) - z) \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.226, + 0.784, + 0.255 + ], + "angle": 0, + "content": "where \\(\\operatorname{sign}(G(D(x)) - x)\\) and \\(\\operatorname{sign}(D(x) - z)\\) is simply the sign of both constrains applied element-wise." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.256, + 0.785, + 0.271 + ], + "angle": 0, + "content": "Next, we will discuss the role of our regularization in two different scenarios:" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.29, + 0.785, + 0.335 + ], + "angle": 0, + "content": "Training the Discriminator with a fixed Generator. When the generator is fixed, we can consider the gradients related to the generator to be 0, so we can obtain the gradients as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.304, + 0.362, + 0.785, + 0.377 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta_ {d}} V (D) = \\mathbb {E} _ {x \\sim p _ {d a t a}} \\left[ \\nabla_ {\\theta_ {d}} \\log D (x) \\right] - \\mathbb {E} _ {x, z} \\operatorname {s i g n} (D (x) - z) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.386, + 0.785, + 0.462 + ], + "angle": 0, + "content": "Considering the general training situation of GANs, when the discriminator is too strong during training. The first term in the gradient tends to be large, resulting in unstable gradient descent. However with our regularization term. Due to the negative sign in the front, we can reduce the gradient of D using this term, making the gradient smaller and can stable the training." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.481, + 0.785, + 0.51 + ], + "angle": 0, + "content": "Training the Generator with a fixed Discriminator. As in the previous case, we first write the representation for the gradient:" + }, + { + "type": "equation", + "bbox": [ + 0.238, + 0.535, + 0.785, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta_ {g}} V (G) = \\mathbb {E} _ {z \\sim p _ {z}} [ \\nabla_ {\\theta_ {d}, \\theta_ {g}} \\log D (1 - D (G (z))) ] + \\mathbb {E} _ {x \\sim p _ {d a t a}} \\operatorname {s i g n} (G (D (x)) - x) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.784, + 0.662 + ], + "angle": 0, + "content": "When the discriminator is dominating the game, it can make the first term to become very small, making it difficult for the generator to effectively optimize and generate real images. After adding a regularization term, we found an additional training objective was given to the generator, allowing the generator to optimize through this training objective, and continue to compete with the discriminator instead of collapsing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.663, + 0.784, + 0.723 + ], + "angle": 0, + "content": "Finally, we can find that by adding two additional training objective, our model can penalize the discriminator when it is too strong. It can optimize the generator by providing more information through additional tasks, making the game fairer and improving the effectiveness of image generation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.742, + 0.38, + 0.757 + ], + "angle": 0, + "content": "3.5 Full Objective" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.764, + 0.784, + 0.794 + ], + "angle": 0, + "content": "With the enhanced discriminator and the reconstruction guided generator, the overall training objective can be rewritten as:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.803, + 0.784, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D} ^ {\\prime} = \\mathcal {L} _ {D} + \\lambda_ {1} \\mathcal {L} _ {C L R} \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.822, + 0.784, + 0.836 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {G} ^ {\\prime} = \\mathcal {L} _ {G} + \\lambda_ {2} \\mathcal {L} _ {r e c} \\tag {14}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLR-GAN" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "code_caption", + "bbox": [ + 0.217, + 0.146, + 0.786, + 0.174 + ], + "angle": 0, + "content": "Algorithm 1 Training a GAN with Consistent Latent Representation and Reconstruction" + }, + { + "type": "code", + "bbox": [ + 0.217, + 0.178, + 0.786, + 0.345 + ], + "angle": 0, + "content": "Require: G and our D that are initialized with random parameters. Training data \\(\\{x_{i}\\}\\) , latent variable \\(\\{z\\}\\) for \\(t = 1\\) until converge do Sample \\(z\\sim P(Z),x\\sim P(X)\\) real_score, real_forent \\(= D(x)\\) fake_score, fake_forent \\(= D(G(z))\\) reconstruct_images \\(= G\\) (real_forent) Calculate Consistent Latent Loss with Eq.(6) Calculate Reconstruction Loss with Eq. (8) Update D and G with Eq. (9), Eq. (10) end for \nOutput: G with best training set FID." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.785, + 0.434 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_G\\) and \\(\\mathcal{L}_D\\) are the original loss function of generator and discriminator, \\(\\mathcal{L}_{rec}\\) and \\(\\mathcal{L}_{CLR}\\) are the proposed consistent latent space distance and real image reconstruction distance calculated through Eq. 6 and Eq. 8, \\(\\lambda_1\\) and \\(\\lambda_2\\) are weighting coefficient of the additional loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.787, + 0.555 + ], + "angle": 0, + "content": "With this updated objective, we can stabilize GANs training in the following two aspects: First, we align the generator and discriminator during training phase by assigning a consistent latent representation task for the discriminator, which makes the game between generator and discriminator fairer. Second, we enable the generator to get access to the real data through a reconstruction task. This helps generator make full use of alternative information, leading to a more realistic output. The pseudo code of our method is shown in Algorithm 1. Extensive experiments result in Sec 4. showed the effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.575, + 0.376, + 0.592 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.604, + 0.788, + 0.755 + ], + "angle": 0, + "content": "To validate the effectiveness of the proposed method, we conducted extensive experiments on various GAN architectures and datasets. Sec. 4.1 first introduces the detailed settings of our experiments. In Sec. 4.2, we proved that our method achieves better quality and stability through quantitative evaluation on various image generation datasets. Sec. 4.3 includes ablation study to the proposed Consistent Latent loss and Reconstruction loss to better understand the designed components. In Sec. 4.4 we prove that we can get a better feature extractor by unsupervised classification. At last, we visualize the realness curve of D to validate the improved fairness of GANs in Sec. 4.5. The generated samples can be found in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.773, + 0.427, + 0.79 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Datasets. For a more comprehensive evaluation of the proposed method, the model was tested on low resolution images and high resolution images separately. We used DCGAN [40] as the basic architecture for low resolution image" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.407, + 0.129 + ], + "angle": 0, + "content": "S.Sun, Z.Luan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.144, + 0.787, + 0.174 + ], + "angle": 0, + "content": "Table 1: FIDs of DCGAN [40] using different learning objectives on CelebA [33] and CIFAR-10 [25] dataset. The bold numbers indicate the best result for each dataset." + }, + { + "type": "table", + "bbox": [ + 0.301, + 0.185, + 0.703, + 0.342 + ], + "angle": 0, + "content": "
Learning ObjectiveCelebACIFAR-10
WGAN [1]36.4755.96
HingeGAN [54]25.5742.4
LSGAN [36]30.7642.01
DCGAN [40]27.0238.56
WGAN-GP [14]70.2841.86
Realness GAN-Obj.1 [49]-36.73
Realness GAN-Obj.2 [49]23.5134.59
Realness GAN-Obj.3 [49]-36.21
AdaptiveMix [32]12.4330.85
CLR-GAN(Ours)13.6323.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.785, + 0.445 + ], + "angle": 0, + "content": "generation, with CIFAR-10 [25] dataset and Celeba64 \\(\\times\\) 64 [33] dataset. For high-resolution image generation, we used StyleGAN-V2 [22] as the basic architecture, and test the generation ability under AFHQ Cat [8], LSUN Church [53], and FFHQ [21] dataset corresponding to animal generation, natural scene generation, and face generation respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.571 + ], + "angle": 0, + "content": "Evaluation. During evaluation phase, we mainly used the most common Frechet Inception Distance(FID) [15] for quantitatively represent the image generation quality. Besides, Precision & Recall [29] is also adopted to provide more details of the realness and diversity of the proposed model. In practice, we calculate FID between 20K generated images and all real samples on CIFAR-10 and CelebA dataset. While for AFHQ-Cat, LSUN-Church and FFHQ, we calculate the FID and P&R between 50K generated images and all real samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.588, + 0.788, + 0.648 + ], + "angle": 0, + "content": "Other settings. For all baselines, our experimental hyperparameters were set according to the original paper. We used pre-trained Inception-V3 [43] as the feature extractor for FID and P&R calculation. For the weights of the additional loss, we set \\(\\lambda_1 = 5\\) and \\(\\lambda_2 = 0.5\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.667, + 0.548, + 0.683 + ], + "angle": 0, + "content": "4.2 Performance on Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.749 + ], + "angle": 0, + "content": "Low-resolution Image Generation. We first conducted experiments on low resolution image datasets to verify the effectiveness of the proposed method. We used DCGAN [40] as the baseline architecture and compared our method with some well-known learning objectives on CIFAR10 and CelebA datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Tab. 1 presents the results. From the perspective of FID, we can directly see a significant margin between the proposed method and previous learning objectives. Furthermore, compared with the recently proposed learning objective [32], our method outperforms \\(24.5\\%\\) on the CIFAR-10 dataset while remains a competitive result on CIFAR-10 dataset \\((8\\%)\\) behind), which further demonstrates the superiority of our model." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLR-GAN" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.144, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 2: Comparisons on AFHQ-Cat [8], FFHQ [21] and LSUN-Church [53] with different high-resolution image generative models. P and R denote precision and recall. The bold numbers indicate the best result for each dataset. The blue numbers indicate the improvements." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.212, + 0.785, + 0.339 + ], + "angle": 0, + "content": "
MethodAFHQ-CatFFHQLSUN Church
FID ↓P ↑R ↑FID ↓P ↑R ↑FID ↓P ↑R ↑
StyleGAN-V2 [22]7.920.680.273.860.680.254.040.580.40
LC-Reg [46]6.70--3.93--4.07--
StyleGAN-V2-ADA [19]6.050.660.254.010.660.264.010.610.43
StyleGAN-V2-APA [18]4.880.650.303.750.670.293.920.600.43
StyleGAN-V2 + Ours4.79(-3.13)0.760.283.44(-0.42)0.700.413.52(-0.52)0.630.46
StyleGAN-V2-ADA + Ours4.45(-0.43)0.740.333.37(-0.64)0.710.443.43(-0.58)0.610.48
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.784, + 0.472 + ], + "angle": 0, + "content": "High-resolution Image Generation. To further demonstrate the effectiveness of the proposed method, we also considered the recently proposed high-resolution image synthesis architecture StyleGAN-V2 [22]. We compared our method with other StyleGAN-V2 variants methods on AFHQ-Cat, LSUN Church and FFHQ datasets, respectively. We also combined our method with recent proposed works that aims to improve GANs training called StyleGAN-V2-ADA [19] to show the compatibility of our method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.785, + 0.564 + ], + "angle": 0, + "content": "As shown in Tab. 2, there is a substantially improvement with the proposed method on StyleGAN-V2 for both datasets, outperforming other methods with a clear margin. We also found that combined with StyleGAN-V2-ADA, our method can get a further improvement, achieving a better result on various datasets. Therefore, our proposed method can improve the synthesis quality of high-resolution images while also integrating well with previous works." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.788, + 0.685 + ], + "angle": 0, + "content": "As for the Precision and Recall. The improvements among multiple datasets are clear, indicating that our model can not only learn to generate more realistic images, but also enhance the diversity of generated images, to some extent solving the problem of insufficient diversity of conventional GANs. This may suggest that CLR-GAN can benefit from the proposed two constrains that leads to learning a wider distribution, thereby improving the diversity of generated samples. More validation metrics and experimental results on the more complex ImageNet [41] dataset are shown in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.708, + 0.389, + 0.723 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In the ablation experiment, we primarily focus on the impact of the two new training constrains we added during training. To ensure the generality, we selected both the low-resolution CIFAR-10 dataset and the high-resolution FFHQ dataset as two baseline datasets. Recall that we can use \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) to control the strength of the proposed constrains in Eq. 13. Therefore we can modify different \\(\\lambda\\) values in the ablation study to figure out the optimal values of two constrains. We first set \\(\\lambda_{1} = \\lambda_{2} = 0\\) to get the baseline values." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.407, + 0.129 + ], + "angle": 0, + "content": "S.Sun, Z.Luan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 3: Ablation studies of different weights \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) on low-resolution CIFAR-10 dataset and high-resolution FFHQ dataset. The best FID of each dataset are marked in bold." + }, + { + "type": "table", + "bbox": [ + 0.243, + 0.198, + 0.481, + 0.329 + ], + "angle": 0, + "content": "
λ1λ2FID
0038.56(baseline)
1025.15
0126.09
5125.20
10125.17
15126.30
5525.55
50.523.3
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.291, + 0.331, + 0.432, + 0.342 + ], + "angle": 0, + "content": "(a) CIFAR-10 @ \\(32 \\times 32\\)" + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.199, + 0.758, + 0.329 + ], + "angle": 0, + "content": "
λ1λ2FID
003.86(baseline)
103.75
013.77
513.62
1013.84
1514.01
553.80
50.53.44
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.578, + 0.331, + 0.71, + 0.342 + ], + "angle": 0, + "content": "(b) FFHQ @ \\(256 \\times 256\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.788, + 0.478 + ], + "angle": 0, + "content": "The FID scores are shown in Tab. 3. We can observe that when we set \\(\\lambda_{1} = 0\\) or \\(\\lambda_{2} = 0\\), the performance of the network significantly drops \\(9.96\\%\\) on CIFAR-10 dataset. The performance even falls below the baseline model on high resolution dataset. Only when both the latent consistent loss and the reconstruction loss are combined together does the model gets its best performance. This proves that the two constrains we added are indispensable with each other." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.57 + ], + "angle": 0, + "content": "Next we will find the optimal strength of the proposed constrains. We first fix \\(\\lambda_{2} = 1\\) and change the value of \\(\\lambda_{1}\\) to find the ideal value of \\(\\lambda_{1}\\). During the experiments, we found that a relatively high \\(\\lambda_{1}\\) makes the model unable to converge, so we chose 15 as the maximum value. As shown in Tab. 3, \\(\\lambda_{1} = 5\\) turns out to be the best. Consequently, we then fix \\(\\lambda_{1} = 5\\) and search for the best \\(\\lambda_{2}\\). Ultimately, \\(\\lambda_{1} = 5\\) and \\(\\lambda_{2} = 0.5\\) turns out to be the best strategy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.571, + 0.788, + 0.722 + ], + "angle": 0, + "content": "During the search for optimal parameters, we observed some interesting phenomena. As shown in Fig. 3. Let \\(\\eta = \\frac{\\lambda_1}{\\lambda_2}\\). We found a higher value of \\(\\eta\\) leads to a better result in the early stage of training. However, an excessively high \\(\\eta\\) might lead to poor convergence at the final stages of model training. We infer that when \\(\\eta\\) is large, our model that constrained by the latent distance, can quickly converge to a distribution similar to the real distribution, hence leads to an early convergence. But overly tight constraints might prevent the model's ability to generalize effectively, resulting in a lack of diversity in the generated data thus making the final performance poor. The underlying causes of this phenomenon require further investigation in future research." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.75, + 0.673, + 0.765 + ], + "angle": 0, + "content": "4.4 Performance on Unsupervised Visual Recognition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "GANs are often regarded as an unsupervised training method. To further demonstrate the effectiveness of the proposed method, we did experiment to show our proposed method leads to a better feature extractor. We take the pre-trained discriminator as a feature extractor and add a linear classification head to perform" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLR-GAN" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.148, + 0.784, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.376, + 0.376, + 0.629, + 0.391 + ], + "angle": 0, + "content": "Fig. 3: Training curves at different \\(\\eta\\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.403, + 0.789, + 0.431 + ], + "angle": 0, + "content": "Table 4: Classification accuracy on CIFAR-10 dataset [25] with various methods with linear classifier." + }, + { + "type": "table", + "bbox": [ + 0.296, + 0.444, + 0.71, + 0.547 + ], + "angle": 0, + "content": "
Method#ParametersAccuracy
ViT-H [10]632M99.5%
EfficientNetV2-M [44]55M99%
PyramidNet [28]26M98.6%
NAT-M4 [34]6.2M98.4%
HCGNet [52]3.1M97.7%
CLR-GAN(Ours)3.02M98.8%
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.571, + 0.785, + 0.6 + ], + "angle": 0, + "content": "classification task on the CIFAR-10 dataset. We compared our feature extractor with other classification models. The results are shown in Tab. 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.602, + 0.788, + 0.737 + ], + "angle": 0, + "content": "As listed in Tab. 4. Our proposed pre-train feature extractor can reach a relatively high accuracy after simple fine-tuning. Compared with models which have similar parameter sizes (\\(\\leq 10M\\)) [34,52], our model is \\(0.4\\%\\) and \\(1.1\\%\\) higher respectively. In comparison with models several orders of magnitude higher than our parameter size [10], our model is also very competitive, only \\(0.7\\%\\) lower than the best model. This indicates that our strategy can not only improve the quality and diversity of generated images in generative adversarial networks, but also serve as an effective unsupervised training method, benefiting downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.757, + 0.448, + 0.773 + ], + "angle": 0, + "content": "4.5 Towards a fairer game" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As we mentioned in the introduction part, the fundamental purpose of our method is to produce more realistic and diverse images by making the training of GANs fairer. By the proposed consistent latent representation and reconstruction, we can shrink the distance between the generator and discriminator." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.407, + 0.128 + ], + "angle": 0, + "content": "S.Sun, Z.Luan et al." + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.148, + 0.505, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.149, + 0.753, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.354, + 0.788, + 0.397 + ], + "angle": 0, + "content": "Fig. 4: Realness scores of the real samples and synthesis samples by various discriminators during training. We visualize the realness score of StyleGAN-V2 [22], StyleGAN-V2-ADA [19] and the proposed CLR-GAN." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.425, + 0.784, + 0.454 + ], + "angle": 0, + "content": "In this section, we demonstrate the enhanced fairness of our method through experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.788, + 0.651 + ], + "angle": 0, + "content": "In order to better quantitatively judge the fairness of our model, we followed the previous two works [2, 47], using the score of the discriminator to represent the realness of the image. We visualize the realness score the model trained on LSUN-Church. We used exponentially weighted averages [17] on the original data for better visualization. The left part of Fig. 4 shows the score of the synthesis images, while the right part shows the score of the real images. We can clearly see that using our CLR-GAN, the realness gap between the generated image and the real image is much smaller than the baseline. While for StyleGAN-V2-ADA, the gap between scores is similar to the baseline, so it only improves FID and can not make the game fairer. We can then draw a conclusion that with the aid of the proposed auxiliary objective functions, the realness scores of real images and generated images become closer, which means that CLR-GAN can improve the realness and diversity of generated images by improving training fairness." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.674, + 0.36, + 0.691 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this paper, we proposed a simple, effective and plug and play new objective function for GANs training. The new objective function simultaneously constrains the output distribution of the generator and the discriminator to make them more consistent. Thereby making the competition between the generator and the discriminator more fair. Eventually improving the quality and diversity of generated images. In addition to image generator, we also demonstrated that the discriminator with constrain is an efficient unsupervised feature extractor. Experimental results demonstrate that our proposed method improves the performance of baseline models over various datasets." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLR-GAN" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.208 + ], + "angle": 0, + "content": "1. Arjovsky, M., Chintala, S., Bottou, L.: Wasserstein generative adversarial networks. In: International conference on machine learning. pp. 214-223. PMLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.21, + 0.787, + 0.251 + ], + "angle": 0, + "content": "2. Bai, Q., Yang, C., Xu, Y., Liu, X., Yang, Y., Shen, Y.: Glead: Improving gans with a generator-leading task. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12094-12104 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.252, + 0.785, + 0.306 + ], + "angle": 0, + "content": "3. Bar-Tal, O., Ofri-Amar, D., Fridman, R., Kasten, Y., Dekel, T.: Text2live: Text-driven layered image and video editing. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 707-723. Springer Nature Switzerland, Cham (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.308, + 0.785, + 0.335 + ], + "angle": 0, + "content": "4. Brock, A., Donahue, J., Simonyan, K.: Large scale gan training for high fidelity natural image synthesis (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.336, + 0.785, + 0.377 + ], + "angle": 0, + "content": "5. Ceylan, D., Huang, C.H.P., Mitra, N.J.: Pix2video: Video editing using image diffusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 23206-23217 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.433 + ], + "angle": 0, + "content": "6. Chan, E.R., Monteiro, M., Kellnhofer, P., Wu, J., Wetzstein, G.: pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5799-5809 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.503 + ], + "angle": 0, + "content": "7. Chenthamarakshan, V., Das, P., Hoffman, S., Strobelt, H., Padhi, I., Lim, K.W., Hoover, B., Manica, M., Born, J., Laino, T., Mojsilovic, A.: Cogmol: Target-specific and selective drug design for Covid-19 using deep generative models. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 4320-4332. Curran Associates, Inc. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.504, + 0.785, + 0.545 + ], + "angle": 0, + "content": "8. Choi, Y., Uh, Y., Yoo, J., Ha, J.W.: Stargan v2: Diverse image synthesis for multiple domains. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8188-8197 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.547, + 0.785, + 0.587 + ], + "angle": 0, + "content": "9. Deng, Y., Tang, F., Dong, W., Ma, C., Pan, X., Wang, L., Xu, C.: Stytr2: Image style transfer with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11326-11336 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.589, + 0.785, + 0.629 + ], + "angle": 0, + "content": "0. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.63, + 0.785, + 0.685 + ], + "angle": 0, + "content": "1. Farnia, F., Ozdaglar, A.: Do GANs always have Nash equilibria? In: III, H.D., Singh, A. (eds.) Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 3029-3039. PMLR (13-18 Jul 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.687, + 0.785, + 0.714 + ], + "angle": 0, + "content": "2. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NeurIPS. vol. 27 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.715, + 0.785, + 0.756 + ], + "angle": 0, + "content": "3. Gu, J., Shen, Y., Zhou, B.: Image processing using multi-code gan prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "4. Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A.C.: Improved training of wasserstein gans. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.406, + 0.128 + ], + "angle": 0, + "content": "S.Sun, Z.Luan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "16. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 6840-6851. Curran Associates, Inc. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.217 + ], + "angle": 0, + "content": "17. Hunter, J.S.: The exponentially weighted moving average. Journal of Quality Technology 18(4), 203-210 (1986)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.785, + 0.271 + ], + "angle": 0, + "content": "18. Jiang, L., Dai, B., Wu, W., Loy, C.C.: Deceive d: Adaptive pseudo augmentation for gan training with limited data. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems (NeurIPS). vol. 34, pp. 21655-21667. Curran Associates, Inc. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.785, + 0.298 + ], + "angle": 0, + "content": "19. Karras, T., Aittala, M., Hellsten, J., Laine, S., Lehtinen, J., Aila, T.: Training generative adversarial networks with limited data (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.352 + ], + "angle": 0, + "content": "20. Karras, T., Aittala, M., Laine, S., Härkönen, E., Hellsten, J., Lehtinen, J., Aila, T.: Alias-free generative adversarial networks. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 852-863. Curran Associates, Inc. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.393 + ], + "angle": 0, + "content": "21. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.434 + ], + "angle": 0, + "content": "22. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8110-8119 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.474 + ], + "angle": 0, + "content": "23. Kim, H., Choi, Y., Kim, J., Yoo, S., Uh, Y.: Exploiting spatial dimensions of latent in gan for real-time image editing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 852-861 (June 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.69, + 0.488 + ], + "angle": 0, + "content": "24. Kingma, D.P., Welling, M.: Auto-encoding variational bayes (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.514 + ], + "angle": 0, + "content": "25. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.581 + ], + "angle": 0, + "content": "26. Kurach, K., Lucic, M., Zhai, X., Michalski, M., Gelly, S.: A large-scale study on regularization and normalization in GANs. In: Chaudhuri, K., Salakhutdinov, R. (eds.) Proceedings of the 36th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 97, pp. 3581-3590. PMLR (09-15 Jun 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.785, + 0.623 + ], + "angle": 0, + "content": "27. Kwon, G., Ye, J.C.: Clipstyler: Image style transfer with a single text condition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18062-18071 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.785, + 0.65 + ], + "angle": 0, + "content": "28. Kwon, J., Kim, J., Park, H., Choi, I.K.: Asam: Adaptive sharpness-aware minimization for scale-invariant learning of deep neural networks (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.785, + 0.69 + ], + "angle": 0, + "content": "29. Kynkänniemi, T., Karras, T., Laine, S., Lehtinen, J., Aila, T.: Improved precision and recall metric for assessing generative models. Advances in Neural Information Processing Systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.69, + 0.785, + 0.745 + ], + "angle": 0, + "content": "30. Lee, G., Kim, H., Kim, J., Kim, S., Ha, J.W., Choi, Y.: Generator knows what discriminator should learn in unconditional gans. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 406-422. Springer Nature Switzerland, Cham (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.785, + 0.786 + ], + "angle": 0, + "content": "31. Liao, W., Hu, K., Yang, M.Y., Rosenhahn, B.: Text to image generation with semantic-spatial aware gan. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18187-18196 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "32. Liu, H., Zhang, W., Li, B., Wu, H., He, N., Huang, Y., Li, Y., Ghanem, B., Zheng, Y.: Adaptivemix: Improving gan training via feature space shrinkage. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16219-16229 (June 2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.657, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLR-GAN" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "33. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: Proceedings of the IEEE international conference on computer vision. pp. 3730-3738 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.23 + ], + "angle": 0, + "content": "34. Lu, Z., Sreekumar, G., Goodman, E., Banzhaf, W., Deb, K., Boddeti, V.N.: Neural architecture transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(9), 2971-2989 (Sep 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.23, + 0.788, + 0.285 + ], + "angle": 0, + "content": "35. Luo, S., Guan, J., Ma, J., Peng, J.: A 3d generative model for structure-based drug design. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 6229-6239. Curran Associates, Inc. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.285, + 0.788, + 0.325 + ], + "angle": 0, + "content": "36. Mao, X., Li, Q., Xie, H., Lau, R.Y., Wang, Z., Paul Smolley, S.: Least squares generative adversarial networks. In: Proceedings of the IEEE international conference on computer vision. pp. 2794-2802 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.325, + 0.788, + 0.352 + ], + "angle": 0, + "content": "37. Miyato, T., Kataoka, T., Koyama, M., Yoshida, Y.: Spectral normalization for generative adversarial networks (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.352, + 0.788, + 0.393 + ], + "angle": 0, + "content": "38. Mukherjee, S., Asnani, H., Lin, E., Kannan, S.: Clustergan: Latent space clustering in generative adversarial networks. Proceedings of the AAAI Conference on Artificial Intelligence 33(01), 4610-4617 (Jul 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.393, + 0.788, + 0.433 + ], + "angle": 0, + "content": "39. Nguyen-Phuoc, T., Li, C., Theis, L., Richardt, C., Yang, Y.L.: Hologan: Unsupervised learning of 3d representations from natural images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7588-7597 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.433, + 0.788, + 0.473 + ], + "angle": 0, + "content": "40. Radford, A., Metz, L., Chintala, S.: Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.514 + ], + "angle": 0, + "content": "41. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision 115, 211-252 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.514, + 0.788, + 0.568 + ], + "angle": 0, + "content": "42. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (Oct 2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.595 + ], + "angle": 0, + "content": "43. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.595, + 0.788, + 0.65 + ], + "angle": 0, + "content": "44. Tan, M., Le, Q.: Efficientnetv2: Smaller models and faster training. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 10096-10106. PMLR (18-24 Jul 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.65, + 0.788, + 0.691 + ], + "angle": 0, + "content": "45. Tran, N.T., Bui, T.A., Cheung, N.M.: Dist-gan: An improved gan using distance constraints. In: Proceedings of the European Conference on Computer Vision (ECCV) (September 2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.691, + 0.788, + 0.744 + ], + "angle": 0, + "content": "46. Tseng, H.Y., Jiang, L., Liu, C., Yang, M.H., Yang, W.: Regularizing generative adversarial networks under limited data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7921-7931 (June 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.744, + 0.788, + 0.786 + ], + "angle": 0, + "content": "47. Wang, J., Yang, C., Xu, Y., Shen, Y., Li, H., Zhou, B.: Improving gan equilibrium by raising spatial awareness. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11285-11293 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.786, + 0.788, + 0.84 + ], + "angle": 0, + "content": "48. Wu, J.Z., Ge, Y., Wang, X., Lei, S.W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7623-7633 (October 2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.407, + 0.128 + ], + "angle": 0, + "content": "S.Sun, Z.Luan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "49. Xiangli, Y., Deng, Y., Dai, B., Loy, C.C., Lin, D.: Real or not real, that is the question. arXiv preprint arXiv:2002.05512 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.217 + ], + "angle": 0, + "content": "50. Xu, Y., AlBahar, B., Huang, J.B.: Temporally consistent semantic video editing. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 357-374. Springer Nature Switzerland, Cham (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.272 + ], + "angle": 0, + "content": "51. Yang, C., Shen, Y., Xu, Y., Zhao, D., Dai, B., Zhou, B.: Improving gans with a dynamic discriminator. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. vol. 35, pp. 15093-15104. Curran Associates, Inc. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "52. Yang, C., An, Z., Zhu, H., Hu, X., Zhang, K., Xu, K., Li, C., Xu, Y.: Gated convolutional networks with hybrid connectivity for image classification. Proceedings of the AAAI Conference on Artificial Intelligence 34(07), 12581-12588 (Apr 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "53. Yu, F., Seff, A., Zhang, Y., Song, S., Funkhouser, T., Xiao, J.: Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.383 + ], + "angle": 0, + "content": "54. Zhao, J., Mathieu, M., LeCun, Y.: Energy-based generative adversarial network. arXiv preprint arXiv:1609.03126 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.425 + ], + "angle": 0, + "content": "55. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (Oct 2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.425 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CLR-GAN_ Improving GANs Stability and Quality via Consistent Latent Representation and Reconstruction/c724f22f-d6fb-423d-a760-902dfb6a8a1d_origin.pdf b/2024/CLR-GAN_ Improving GANs Stability and Quality via Consistent Latent Representation and Reconstruction/c724f22f-d6fb-423d-a760-902dfb6a8a1d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0376d68d096f4944487534b31f13ab312acbc5bd --- /dev/null +++ b/2024/CLR-GAN_ Improving GANs Stability and Quality via Consistent Latent Representation and Reconstruction/c724f22f-d6fb-423d-a760-902dfb6a8a1d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2ae9ddf9c03ba0a8f08812a1a2ba5e837d81aec98cfffcc8f3e9808a9497919 +size 1080149 diff --git a/2024/CLR-GAN_ Improving GANs Stability and Quality via Consistent Latent Representation and Reconstruction/full.md b/2024/CLR-GAN_ Improving GANs Stability and Quality via Consistent Latent Representation and Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..24146fddfb6db52ae7801d908b0033a58e55eef0 --- /dev/null +++ b/2024/CLR-GAN_ Improving GANs Stability and Quality via Consistent Latent Representation and Reconstruction/full.md @@ -0,0 +1,329 @@ +# CLR-GAN: Improving GANs Stability and Quality via Consistent Latent Representation and Reconstruction + +Shengke Sun1*, Ziqian Luan2*, Zhanshan Zhao1,3, Shijie Luo1, and Shuzhen Han1 + +$^{1}$ Tiangong University, 300387 Tianjin, China {sunshengke,hanshuzhen,luoshijie}@tiangong.edu.cn, zhzhsh127@163.com + +$^{2}$ Xidian University, 710126 Xi'an, Shaanxi, China + +tyuwe@foxmail.com + +3 Institute for Sustainable Industries & Liveable Cities, Victory University, Melbourne, VIC 8001, Australia +†Corresponding author + +Abstract. Generative Adversarial Networks(GANs) have received considerable attention due to its outstanding ability to generate images. However, training a GAN is hard since the game between the Generator(G) and the Discriminator(D) is unfair. Towards making the competition fairer, we propose a new perspective of training GANs, named Consistent Latent Representation and Reconstruction(CLR-GAN). In this paradigm, we treat the G and D as an inverse process, the discriminator has an additional task to restore the pre-defined latent code while the generator also needs to reconstruct the real input, thus obtaining a relationship between the latent space of G and the out-features of D. Based on this prior, we can put D and G on an equal position during training using a new criterion. Experimental results on various datasets and architectures prove our paradigm can make GANs more stable and generate better quality images(31.22% gain of FID on CIFAR10 and 39.5% on AFHQ-Cat, respectively). We hope that the proposed perspective can inspire researchers to explore different ways of viewing GANs training, rather than being limited to a two-player game. The code is publicly available at https://github.com/Petecheco/CLR-GAN. + +Keywords: Generative Adversarial Networks $\cdot$ Image generation $\cdot$ Latent space consistency + +# 1 Introduction + +Generative models are an important part of computer vision, they can be widely used in fields such as image generation [12, 16, 20, 24], video editing [3, 5, 48, 50], style transfer [9, 27, 55] and even drug discovery [7, 35]. Among various generative + +![](images/de061eb9c0fdf1bad66007ead92185b8e996efcac31478bf912a30d288a3203d.jpg) +Fig. 1: Sketch map of the proposed method, we proposed two additional learning objectives to consider D and G as inverse process. The top is latent consistency loss where D tries to rebuild the latent distribution of G. The bottom is reconstruction loss where G tries to restore the original real images. + +models, Generative Adversarial Networks(GANs) have received considerable attention due to its remarkable performance in generating realistic images [12]. A conventional GAN comprises of two individual networks named generator(G) and discriminator(D). While training a GAN, the generator aims to produce indistinguishable images, and the discriminator tries to figure out the synthesized images from the input images. By the competition between D and G, ideally, we can eventually train a generator that can recover the real distribution and generate high quality samples. + +However, it is well known that training a GAN is hard and unstable [26,37]. In general, the instability of training GANs can be attributed to the unequal set up of generator(G) and discriminator(D). Specifically, in the traditional paradigm, the training process is formulated as a two-player game. But in fact, the discriminator seems to dominate the game [2,47]. In practice, the discriminator can easily identify the fake samples from a relatively early stage of the training and maintain this advantage during the whole training, making the generator hard to converge [51]. Many attempts have been done to improve the stability of training GANs. Previous studies attempted to find new metric functions [1, 14, 32, 45], enhance the generator's ability to generate [4,21,40], or use data augment methods [19,30]. These solutions have shown some degree of improvement on training stability, but they did not fundamentally change the core relationship between the generator and discriminator, while ignoring the essence of dynamic balance within GANs architecture. The competition between the generator and discriminator is still unfair. In addition, existing methods have not fully utilized the structural information of the latent space, which limits the further improvement + +of GANs generation ability. Therefore, improving the training of GANs still remains an unsolved and challenging problem. + +In this paper, contrast to existing methods, we proposed a novel perspective of training GANs. By rethinking the training process and network architecture of GANs. We found that the generator and the discriminator got optimized in a considerably large latent space, resulting unstable training and inconsistency between the generator and the discriminator. While traditional methods primarily focus on adjusting network architectures or tuning hyperparameters to stabilize training [19, 40]. Our method is to view the generator and the discriminator as two inverse components. As shown in Fig. 1. In particular, the discriminator can be regarded as outputting a realness score and extracting samples to the latent representation space at the same time. And the generator not only generates images but restores source images from the extracted latent representation. By constraining the generator and discriminator with these additional tasks, we make the discriminator's representation latent space closer to the generator's latent space, thus making the generator and the discriminator more consistent, leading a fairer game between the generator and the discriminator. This not only provides a more robust framework for training GANs but also opens up new possibilities for improving GANs training from different aspects. + +# 2 Related Work + +# 2.1 Improving GANs through discriminators + +As we mentioned above, due to the limitations of original training objective [1, 12], the discriminator tends to easily take over the competition during the early stage of training. Prior works have been devoted to improving the inequality using modified discriminators. Dist-GAN [45] decelerate the convergence of the discriminator by constructing a reconstruction constrain using a auxiliary autoencoder. Wang et al. [47] combined Grad-CAM [42] with GANs to improve the spatial awareness of the generator using the discriminator as a regularizer, lessening the information gap between D and G. DynamicGAN [51] controls the strength of the discriminator by gradually increasing D's learning capacity, thereby preventing the discriminator from being overly dominant in the early stages of training. AdaptiveMix [32] shrinks the feature space of the discriminator by training using constructed hard samples, reducing the feature space of the discriminator and making training easier. There are also some data augmentation methods that enhance model training by providing more data to the model [19, 46]. However, the above methods have not changed the inner properties of training. All gradient during the optimization process is provided by D, so D always occupies a more dominant position during the training process, which will prevent further optimization of the model. While our model, based on limiting D, also constrains D to optimize based on the gradient provided by G. This allows both G and D to optimize themselves using the gradients provided by each other, making the game fairer. + +# 2.2 Latent space of GANs + +The latent space of GANs has always been a popular research area due to its rich semantic information. Multi-Code GAN [13] combines multiple latent feature maps with adaptive channels to improve the quality of image reconstruction. ClusterGAN [38] combines latent space encoding and one-hot encoding to obtain a better clustering result. StyleMapGAN [23] utilizes latent space as an accurate feature embedding, improving the performance of image editing and interpolation tasks. Although latent space plays an important role in various downstream tasks. But to our best knowledge, there has been no previous work that directly uses the latent space of GANs as a component to optimize the training of GANs. In our proposed method, we have directly incorporated latent space into the training process of GANs as an additional constrain. The experimental results show that applying this method can effectively improve the quality and diversity of generated images. + +# 2.3 Differences to Reconstruction-based Methods + +Prior works have explored the application of reconstruction-based methods within GANs. HoloGAN [39] and PiGAN [6] utilize latent reconstruction to learn the 3D-aware image representations. GLeaD [2] takes advantage of image reconstruction to improve the training dynamics of GANs. However, these methods either lack a specific focus on stabilizing GANs training or solely concentrate on improving the generator without considering the discriminator. Specifically, in the GLeaD paradigm, given an image, we expect D to extract representative features that can be adequately decoded by G to reconstruct the input, however, this paradigm only improves the generator, the discriminator still controls the game. To address these limitations, we use the latent spatial code extracted by the discriminator and reconstruction loss to achieve stable GAN training and generate higher quality and more diverse images. Thus making GANs training more efficient and practical. + +# 3 Method + +As we mentioned before, we try to augment the training fairness of GANs via considering the generative process of the generator (G) and the discriminative process of the discriminator (D) as inverse to each other, thereby enhancing the stability and quality of generated samples. To better understand this new perspective, we first review the conventional formulation of GANs in Sec. 3.1. To obtain a fairer training, Sec. 3.2 introduces our new training paradigm named Consistent Latent Representation that considers G and D as an inverse process. Furthermore, in Sec. 3.3, we also propose a Real Image Reconstruction strategy to achieve mutual connection between generator and the discriminator. Then Sec. 3.4 explained our method from a regularization perspective. Finally, Sec. 3.5 gives the full optimization objective and shows the pseudo code of the new training paradigm. + +# 3.1 Preliminary + +A GAN usually consists of two components: a generator $G(\cdot)$ and a discriminator $D(\cdot)$ . The generator aims to map a latent code $z$ into an image, while the discriminator tries to distinguish the generated image $G(z)$ from the real one $x$ . The conventional GAN trains via a two-player game by optimizing the learning objective as follows: + +$$ +\min _ {G} \max _ {D} V (D, G) = \mathbb {E} _ {x \sim p _ {d a t a} (x)} [ \log D (x) ] + \mathbb {E} _ {z \sim p _ {z} (z)} [ \log (1 - D (G (z))) ] \quad (1) +$$ + +where $p_{z}(z)$ and $p_{data}(x)$ represent a random latent representation and real data distribution respectively. + +Goodfellow et al. [12] have proved that the optimal solution to this objective is that $G$ can eventually recover the distribution of source data and $D$ can not separate the generated images out from real images. However, the game between two players is not fair since in Eq. 1 that the gradient for optimizing all comes from $D$ , which makes the discriminator a natural dominant [2]. Thus the ideal solution is hard to reach in practice. [11,47] + +# 3.2 Consistent Latent Space + +We aim to improve the stability of GANs by introducing a novel perspective, that is to consider the generate and discriminate as an inverse process. Recall that in a regular GAN paradigm, the generator and discriminator are treated as two individual components. While in our proposed view, we have refined the concepts of generator and discriminator and consider them as an inverse process. In our perspective, the generator maps a latent representation to the real data distribution, and the discriminator not only outputs a realizness score, but it also transform the high-dimensional distribution into a trainable latent space. By measuring the distance between these two spaces, we can improve GANs training in the following aspects: First, we can to some extent make the generator and discriminator behave more consistent in the mapping process. Second, since the discriminator needs to align with the generator, it can be seen as constraining the discriminator using generator, which makes the game not completely dominated by discriminator. The detailed implementation is shown in Fig. 2. + +Extracting Latent Representation through D. In the training of GANs, the discriminator plays a role of downsample high-dimensional data. For any discriminator, it can be divided into two parts, the first part is the feature extractor denoted as $f(\cdot)$ which extracts features from high-dimensional data samples, another serves as a affine from extracted features to realizness score to tell whether the input is true(real samples) or false(generated samples). For any input data $x$ and latent code $z$ , the whole discriminate process can be denoted as follows: + +$$ +D (x) = R (f (x)) \tag {2} +$$ + +$$ +D (z) = R (f (G (z))) \tag {3} +$$ + +![](images/ce52d80ecda9abf83c85b0e4a10a31fc0295544a528d3f3e87522430bf962476.jpg) +Fig. 2: Detailed implementation of Consistent Latent Representation. Given a latent code $z$ , the generator maps $z$ to an image. The synthesised image was then got discriminated by the discriminator, outputting a realness score and a reconstructed latent code $z'$ . The Consistent Latent loss is calculated between the real latent $z$ and the reconstructed latent $z'$ . The right part specifically shows the architecture of D. + +where $f(\cdot)$ is the feature extractor module of discriminator and $R(\cdot)$ is the affine module to transform feature space to the realness score. + +Through this discriminate process, we can transform structural high-dimensional input into feature map that contains abundant semantic information [31], thus judging the realness of the input data. We discovered both latent code in the generator and feature map in the discriminator contain lots semantic features, making it possible to align the generator and discriminator using a mapping function. The mapping function converts discriminator's feature space to latent space. Finally we can get a reconstructed latent code $\phi (x)$ from discriminator as follows: + +$$ +\phi (x) = \sigma (f (x)) \tag {4} +$$ + +where $x$ is the real input, $f(\cdot)$ is the feature extractor of D and $\sigma(\cdot)$ is a mapping function that transform feature to latent space, we use an identical affine in this paper for simplicity. + +Consistent Latent Representation Loss. As we mentioned above, we now have a real latent code $(z)$ from the generator and a reconstructed latent code(Eq.4) from the discriminator. The latent representation loss can be defined by measuring the distance between these two latent spaces: + +$$ +\mathcal {L} _ {C L R} = \operatorname {D i s t} (\phi (x), z) \tag {5} +$$ + +where $\text{Dist}(\cdot)$ is distance measurement for any two distributions. In this paper, we use L1-Distance similar to [55] for all distance measurement, so the $\mathcal{L}_{\text{clr}}$ can be rewritten as: + +$$ +\mathcal {L} _ {C L R} = \frac {1}{n} \sum_ {i = 1} ^ {n} \| z _ {i} - \phi (x _ {i}) \| _ {1} \tag {6} +$$ + +# 3.3 Real Image Reconstruction + +As mentioned in Sec. 2, in a regular GAN, the generator optimized through the gradient offered by discriminator. Also during the training, the generator has no access to real data distributions, it can only update its parameters implicitly, while the discriminator can get optimized according to real data distribution. This unfair situation helps the discriminator converge faster at an early stage of the training [47], thus the discriminator can not provide useful gradient in the later stage of training, making the generator not perform well. + +To solve this problem, we proposed a new strategy called Real Image Reconstruction, which enables the generator to also get gradient from the real data distribution. Specifically, since we have strengthened the discriminator to output both realizness and latent code in the last section, we can then use latent code as a bridge to indirectly connect the generator with real distribution. Subsequently, we can use the generated images for a image reconstruction task. The generator then can utilize this gradient that comes from real distribution for better synthesis, making the whole process more efficient. In practice, we can use Eq. 4 to get the reconstructed latent code of real distributions, then we use generator for image restoration. + +$$ +I _ {r e c} = G (\phi (x)) \tag {7} +$$ + +After reconstructing the image, we can calculate the distance between real images and reconstructed images. And for the reconstruction loss, we use a regular from [55] as follows: + +$$ +\mathcal {L} _ {r e c} = \frac {1}{n} \sum_ {i = 1} ^ {n} \| I (x) - I _ {r e c} \| _ {1} \tag {8} +$$ + +where $I(x)$ is the source image data and $I_{rec}$ is the reconstructed image using generator and reconstructed latent code. + +# 3.4 Enhancing Training Stability via Regularization Techniques + +To better understand the superiority of our method, we turn to investigate the above training strategies from the perspective of regularization in this section. + +With our method, the objective of training GANs can be rewritten as: + +$$ +\begin{array}{l} \min _ {G} \max _ {D} V (D, G) = \mathbb {E} _ {x \sim p _ {d a t a}} \log D (x) + \mathbb {E} _ {z \sim p _ {z}} \log [ 1 - D (G (z)) ] \tag {9} \\ + \mathbb {E} _ {x \sim p _ {d a t a}} \| G (D (x)) - x \| - \mathbb {E} _ {x, z} \| D (x) - z \| \\ \end{array} +$$ + +The first two terms of Eq. 9 are the default training objectives for GANs, while the third and fourth terms correspond to Latent Consistency Loss and Real Image Reconstruction Loss, respectively. Note that these two objectives actually appear in the form of L1 norm, and we will consider them as two types of regularization in this section. + +When optimizing networks, we actually calculate the gradients of objective function. For the training objective we propose, the gradients can be expressed + +as: + +$$ +\begin{array}{l} \nabla_ {\theta_ {d}, \theta_ {g}} V (G, D) = \mathbb {E} _ {x \sim p _ {d a t a}} [ \nabla_ {\theta_ {d}} \log D (x) ] + \mathbb {E} _ {z \sim p _ {z}} [ \nabla_ {\theta_ {d}, \theta_ {g}} \log D (1 - D (G (z))) ] \\ + \mathbb {E} _ {x \sim p _ {d a t a}} \operatorname {s i g n} (G (D (x)) - x) - \mathbb {E} _ {x, z} \operatorname {s i g n} (D (x) - z) \tag {10} \\ \end{array} +$$ + +where $\operatorname{sign}(G(D(x)) - x)$ and $\operatorname{sign}(D(x) - z)$ is simply the sign of both constrains applied element-wise. + +Next, we will discuss the role of our regularization in two different scenarios: + +Training the Discriminator with a fixed Generator. When the generator is fixed, we can consider the gradients related to the generator to be 0, so we can obtain the gradients as follows: + +$$ +\nabla_ {\theta_ {d}} V (D) = \mathbb {E} _ {x \sim p _ {d a t a}} \left[ \nabla_ {\theta_ {d}} \log D (x) \right] - \mathbb {E} _ {x, z} \operatorname {s i g n} (D (x) - z) \tag {11} +$$ + +Considering the general training situation of GANs, when the discriminator is too strong during training. The first term in the gradient tends to be large, resulting in unstable gradient descent. However with our regularization term. Due to the negative sign in the front, we can reduce the gradient of D using this term, making the gradient smaller and can stable the training. + +Training the Generator with a fixed Discriminator. As in the previous case, we first write the representation for the gradient: + +$$ +\nabla_ {\theta_ {g}} V (G) = \mathbb {E} _ {z \sim p _ {z}} [ \nabla_ {\theta_ {d}, \theta_ {g}} \log D (1 - D (G (z))) ] + \mathbb {E} _ {x \sim p _ {d a t a}} \operatorname {s i g n} (G (D (x)) - x) \tag {12} +$$ + +When the discriminator is dominating the game, it can make the first term to become very small, making it difficult for the generator to effectively optimize and generate real images. After adding a regularization term, we found an additional training objective was given to the generator, allowing the generator to optimize through this training objective, and continue to compete with the discriminator instead of collapsing. + +Finally, we can find that by adding two additional training objective, our model can penalize the discriminator when it is too strong. It can optimize the generator by providing more information through additional tasks, making the game fairer and improving the effectiveness of image generation. + +# 3.5 Full Objective + +With the enhanced discriminator and the reconstruction guided generator, the overall training objective can be rewritten as: + +$$ +\mathcal {L} _ {D} ^ {\prime} = \mathcal {L} _ {D} + \lambda_ {1} \mathcal {L} _ {C L R} \tag {13} +$$ + +$$ +\mathcal {L} _ {G} ^ {\prime} = \mathcal {L} _ {G} + \lambda_ {2} \mathcal {L} _ {r e c} \tag {14} +$$ + +Algorithm 1 Training a GAN with Consistent Latent Representation and Reconstruction +```verilog +Require: G and our D that are initialized with random parameters. Training data $\{x_{i}\}$ , latent variable $\{z\}$ for $t = 1$ until converge do Sample $z\sim P(Z),x\sim P(X)$ real_score, real_forent $= D(x)$ fake_score, fake_forent $= D(G(z))$ reconstruct_images $= G$ (real_forent) Calculate Consistent Latent Loss with Eq.(6) Calculate Reconstruction Loss with Eq. (8) Update D and G with Eq. (9), Eq. (10) end for +Output: G with best training set FID. +``` + +where $\mathcal{L}_G$ and $\mathcal{L}_D$ are the original loss function of generator and discriminator, $\mathcal{L}_{rec}$ and $\mathcal{L}_{CLR}$ are the proposed consistent latent space distance and real image reconstruction distance calculated through Eq. 6 and Eq. 8, $\lambda_1$ and $\lambda_2$ are weighting coefficient of the additional loss. + +With this updated objective, we can stabilize GANs training in the following two aspects: First, we align the generator and discriminator during training phase by assigning a consistent latent representation task for the discriminator, which makes the game between generator and discriminator fairer. Second, we enable the generator to get access to the real data through a reconstruction task. This helps generator make full use of alternative information, leading to a more realistic output. The pseudo code of our method is shown in Algorithm 1. Extensive experiments result in Sec 4. showed the effectiveness of our method. + +# 4 Experiments + +To validate the effectiveness of the proposed method, we conducted extensive experiments on various GAN architectures and datasets. Sec. 4.1 first introduces the detailed settings of our experiments. In Sec. 4.2, we proved that our method achieves better quality and stability through quantitative evaluation on various image generation datasets. Sec. 4.3 includes ablation study to the proposed Consistent Latent loss and Reconstruction loss to better understand the designed components. In Sec. 4.4 we prove that we can get a better feature extractor by unsupervised classification. At last, we visualize the realness curve of D to validate the improved fairness of GANs in Sec. 4.5. The generated samples can be found in the Supplementary Material. + +# 4.1 Experimental Setup + +Datasets. For a more comprehensive evaluation of the proposed method, the model was tested on low resolution images and high resolution images separately. We used DCGAN [40] as the basic architecture for low resolution image + +Table 1: FIDs of DCGAN [40] using different learning objectives on CelebA [33] and CIFAR-10 [25] dataset. The bold numbers indicate the best result for each dataset. + +
Learning ObjectiveCelebACIFAR-10
WGAN [1]36.4755.96
HingeGAN [54]25.5742.4
LSGAN [36]30.7642.01
DCGAN [40]27.0238.56
WGAN-GP [14]70.2841.86
Realness GAN-Obj.1 [49]-36.73
Realness GAN-Obj.2 [49]23.5134.59
Realness GAN-Obj.3 [49]-36.21
AdaptiveMix [32]12.4330.85
CLR-GAN(Ours)13.6323.3
+ +generation, with CIFAR-10 [25] dataset and Celeba64 $\times$ 64 [33] dataset. For high-resolution image generation, we used StyleGAN-V2 [22] as the basic architecture, and test the generation ability under AFHQ Cat [8], LSUN Church [53], and FFHQ [21] dataset corresponding to animal generation, natural scene generation, and face generation respectively. + +Evaluation. During evaluation phase, we mainly used the most common Frechet Inception Distance(FID) [15] for quantitatively represent the image generation quality. Besides, Precision & Recall [29] is also adopted to provide more details of the realness and diversity of the proposed model. In practice, we calculate FID between 20K generated images and all real samples on CIFAR-10 and CelebA dataset. While for AFHQ-Cat, LSUN-Church and FFHQ, we calculate the FID and P&R between 50K generated images and all real samples. + +Other settings. For all baselines, our experimental hyperparameters were set according to the original paper. We used pre-trained Inception-V3 [43] as the feature extractor for FID and P&R calculation. For the weights of the additional loss, we set $\lambda_1 = 5$ and $\lambda_2 = 0.5$ . + +# 4.2 Performance on Image Generation + +Low-resolution Image Generation. We first conducted experiments on low resolution image datasets to verify the effectiveness of the proposed method. We used DCGAN [40] as the baseline architecture and compared our method with some well-known learning objectives on CIFAR10 and CelebA datasets. + +Tab. 1 presents the results. From the perspective of FID, we can directly see a significant margin between the proposed method and previous learning objectives. Furthermore, compared with the recently proposed learning objective [32], our method outperforms $24.5\%$ on the CIFAR-10 dataset while remains a competitive result on CIFAR-10 dataset $(8\%)$ behind), which further demonstrates the superiority of our model. + +Table 2: Comparisons on AFHQ-Cat [8], FFHQ [21] and LSUN-Church [53] with different high-resolution image generative models. P and R denote precision and recall. The bold numbers indicate the best result for each dataset. The blue numbers indicate the improvements. + +
MethodAFHQ-CatFFHQLSUN Church
FID ↓P ↑R ↑FID ↓P ↑R ↑FID ↓P ↑R ↑
StyleGAN-V2 [22]7.920.680.273.860.680.254.040.580.40
LC-Reg [46]6.70--3.93--4.07--
StyleGAN-V2-ADA [19]6.050.660.254.010.660.264.010.610.43
StyleGAN-V2-APA [18]4.880.650.303.750.670.293.920.600.43
StyleGAN-V2 + Ours4.79(-3.13)0.760.283.44(-0.42)0.700.413.52(-0.52)0.630.46
StyleGAN-V2-ADA + Ours4.45(-0.43)0.740.333.37(-0.64)0.710.443.43(-0.58)0.610.48
+ +High-resolution Image Generation. To further demonstrate the effectiveness of the proposed method, we also considered the recently proposed high-resolution image synthesis architecture StyleGAN-V2 [22]. We compared our method with other StyleGAN-V2 variants methods on AFHQ-Cat, LSUN Church and FFHQ datasets, respectively. We also combined our method with recent proposed works that aims to improve GANs training called StyleGAN-V2-ADA [19] to show the compatibility of our method. + +As shown in Tab. 2, there is a substantially improvement with the proposed method on StyleGAN-V2 for both datasets, outperforming other methods with a clear margin. We also found that combined with StyleGAN-V2-ADA, our method can get a further improvement, achieving a better result on various datasets. Therefore, our proposed method can improve the synthesis quality of high-resolution images while also integrating well with previous works. + +As for the Precision and Recall. The improvements among multiple datasets are clear, indicating that our model can not only learn to generate more realistic images, but also enhance the diversity of generated images, to some extent solving the problem of insufficient diversity of conventional GANs. This may suggest that CLR-GAN can benefit from the proposed two constrains that leads to learning a wider distribution, thereby improving the diversity of generated samples. More validation metrics and experimental results on the more complex ImageNet [41] dataset are shown in the Supplementary Material. + +# 4.3 Ablation Study + +In the ablation experiment, we primarily focus on the impact of the two new training constrains we added during training. To ensure the generality, we selected both the low-resolution CIFAR-10 dataset and the high-resolution FFHQ dataset as two baseline datasets. Recall that we can use $\lambda_{1}$ and $\lambda_{2}$ to control the strength of the proposed constrains in Eq. 13. Therefore we can modify different $\lambda$ values in the ablation study to figure out the optimal values of two constrains. We first set $\lambda_{1} = \lambda_{2} = 0$ to get the baseline values. + +Table 3: Ablation studies of different weights $\lambda_{1}$ and $\lambda_{2}$ on low-resolution CIFAR-10 dataset and high-resolution FFHQ dataset. The best FID of each dataset are marked in bold. + +
λ1λ2FID
0038.56(baseline)
1025.15
0126.09
5125.20
10125.17
15126.30
5525.55
50.523.3
+ +(a) CIFAR-10 @ $32 \times 32$ + +
λ1λ2FID
003.86(baseline)
103.75
013.77
513.62
1013.84
1514.01
553.80
50.53.44
+ +(b) FFHQ @ $256 \times 256$ + +The FID scores are shown in Tab. 3. We can observe that when we set $\lambda_{1} = 0$ or $\lambda_{2} = 0$ , the performance of the network significantly drops $9.96\%$ on CIFAR-10 dataset. The performance even falls below the baseline model on high resolution dataset. Only when both the latent consistent loss and the reconstruction loss are combined together does the model gets its best performance. This proves that the two constrains we added are indispensable with each other. + +Next we will find the optimal strength of the proposed constrains. We first fix $\lambda_{2} = 1$ and change the value of $\lambda_{1}$ to find the ideal value of $\lambda_{1}$ . During the experiments, we found that a relatively high $\lambda_{1}$ makes the model unable to converge, so we chose 15 as the maximum value. As shown in Tab. 3, $\lambda_{1} = 5$ turns out to be the best. Consequently, we then fix $\lambda_{1} = 5$ and search for the best $\lambda_{2}$ . Ultimately, $\lambda_{1} = 5$ and $\lambda_{2} = 0.5$ turns out to be the best strategy. + +During the search for optimal parameters, we observed some interesting phenomena. As shown in Fig. 3. Let $\eta = \frac{\lambda_1}{\lambda_2}$ . We found a higher value of $\eta$ leads to a better result in the early stage of training. However, an excessively high $\eta$ might lead to poor convergence at the final stages of model training. We infer that when $\eta$ is large, our model that constrained by the latent distance, can quickly converge to a distribution similar to the real distribution, hence leads to an early convergence. But overly tight constraints might prevent the model's ability to generalize effectively, resulting in a lack of diversity in the generated data thus making the final performance poor. The underlying causes of this phenomenon require further investigation in future research. + +# 4.4 Performance on Unsupervised Visual Recognition + +GANs are often regarded as an unsupervised training method. To further demonstrate the effectiveness of the proposed method, we did experiment to show our proposed method leads to a better feature extractor. We take the pre-trained discriminator as a feature extractor and add a linear classification head to perform + +![](images/ff7bc499da0340d06845c11206046bd0e7ff0ec508cc9fbcf511276636db84f2.jpg) +Fig. 3: Training curves at different $\eta$ + +Table 4: Classification accuracy on CIFAR-10 dataset [25] with various methods with linear classifier. + +
Method#ParametersAccuracy
ViT-H [10]632M99.5%
EfficientNetV2-M [44]55M99%
PyramidNet [28]26M98.6%
NAT-M4 [34]6.2M98.4%
HCGNet [52]3.1M97.7%
CLR-GAN(Ours)3.02M98.8%
+ +classification task on the CIFAR-10 dataset. We compared our feature extractor with other classification models. The results are shown in Tab. 4. + +As listed in Tab. 4. Our proposed pre-train feature extractor can reach a relatively high accuracy after simple fine-tuning. Compared with models which have similar parameter sizes ( $\leq 10M$ ) [34,52], our model is $0.4\%$ and $1.1\%$ higher respectively. In comparison with models several orders of magnitude higher than our parameter size [10], our model is also very competitive, only $0.7\%$ lower than the best model. This indicates that our strategy can not only improve the quality and diversity of generated images in generative adversarial networks, but also serve as an effective unsupervised training method, benefiting downstream tasks. + +# 4.5 Towards a fairer game + +As we mentioned in the introduction part, the fundamental purpose of our method is to produce more realistic and diverse images by making the training of GANs fairer. By the proposed consistent latent representation and reconstruction, we can shrink the distance between the generator and discriminator. + +![](images/343da88656d16b9e3f248d949aee833c05182e74e0cc25813bda8a6775fb23fc.jpg) +Fig. 4: Realness scores of the real samples and synthesis samples by various discriminators during training. We visualize the realness score of StyleGAN-V2 [22], StyleGAN-V2-ADA [19] and the proposed CLR-GAN. + +![](images/d556ee9befe05df20e0b6b8f735f109d684cd2bb94aab8b46d53eb57893e16af.jpg) + +In this section, we demonstrate the enhanced fairness of our method through experiments. + +In order to better quantitatively judge the fairness of our model, we followed the previous two works [2, 47], using the score of the discriminator to represent the realness of the image. We visualize the realness score the model trained on LSUN-Church. We used exponentially weighted averages [17] on the original data for better visualization. The left part of Fig. 4 shows the score of the synthesis images, while the right part shows the score of the real images. We can clearly see that using our CLR-GAN, the realness gap between the generated image and the real image is much smaller than the baseline. While for StyleGAN-V2-ADA, the gap between scores is similar to the baseline, so it only improves FID and can not make the game fairer. We can then draw a conclusion that with the aid of the proposed auxiliary objective functions, the realness scores of real images and generated images become closer, which means that CLR-GAN can improve the realness and diversity of generated images by improving training fairness. + +# 5 Conclusion + +In this paper, we proposed a simple, effective and plug and play new objective function for GANs training. The new objective function simultaneously constrains the output distribution of the generator and the discriminator to make them more consistent. Thereby making the competition between the generator and the discriminator more fair. Eventually improving the quality and diversity of generated images. In addition to image generator, we also demonstrated that the discriminator with constrain is an efficient unsupervised feature extractor. Experimental results demonstrate that our proposed method improves the performance of baseline models over various datasets. + +# References + +1. Arjovsky, M., Chintala, S., Bottou, L.: Wasserstein generative adversarial networks. In: International conference on machine learning. pp. 214-223. PMLR (2017) +2. Bai, Q., Yang, C., Xu, Y., Liu, X., Yang, Y., Shen, Y.: Glead: Improving gans with a generator-leading task. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12094-12104 (June 2023) +3. Bar-Tal, O., Ofri-Amar, D., Fridman, R., Kasten, Y., Dekel, T.: Text2live: Text-driven layered image and video editing. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 707-723. Springer Nature Switzerland, Cham (2022) +4. Brock, A., Donahue, J., Simonyan, K.: Large scale gan training for high fidelity natural image synthesis (2019) +5. Ceylan, D., Huang, C.H.P., Mitra, N.J.: Pix2video: Video editing using image diffusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 23206-23217 (October 2023) +6. Chan, E.R., Monteiro, M., Kellnhofer, P., Wu, J., Wetzstein, G.: pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5799-5809 (2021) +7. Chenthamarakshan, V., Das, P., Hoffman, S., Strobelt, H., Padhi, I., Lim, K.W., Hoover, B., Manica, M., Born, J., Laino, T., Mojsilovic, A.: Cogmol: Target-specific and selective drug design for Covid-19 using deep generative models. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 4320-4332. Curran Associates, Inc. (2020) +8. Choi, Y., Uh, Y., Yoo, J., Ha, J.W.: Stargan v2: Diverse image synthesis for multiple domains. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8188-8197 (2020) +9. Deng, Y., Tang, F., Dong, W., Ma, C., Pan, X., Wang, L., Xu, C.: Stytr2: Image style transfer with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11326-11336 (June 2022) +0. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale (2021) +1. Farnia, F., Ozdaglar, A.: Do GANs always have Nash equilibria? In: III, H.D., Singh, A. (eds.) Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 3029-3039. PMLR (13-18 Jul 2020) +2. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NeurIPS. vol. 27 (2014) +3. Gu, J., Shen, Y., Zhou, B.: Image processing using multi-code gan prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020) +4. Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A.C.: Improved training of wasserstein gans. Advances in neural information processing systems 30 (2017) +5. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017) + +16. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 6840-6851. Curran Associates, Inc. (2020) +17. Hunter, J.S.: The exponentially weighted moving average. Journal of Quality Technology 18(4), 203-210 (1986) +18. Jiang, L., Dai, B., Wu, W., Loy, C.C.: Deceive d: Adaptive pseudo augmentation for gan training with limited data. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems (NeurIPS). vol. 34, pp. 21655-21667. Curran Associates, Inc. (2021) +19. Karras, T., Aittala, M., Hellsten, J., Laine, S., Lehtinen, J., Aila, T.: Training generative adversarial networks with limited data (2020) +20. Karras, T., Aittala, M., Laine, S., Härkönen, E., Hellsten, J., Lehtinen, J., Aila, T.: Alias-free generative adversarial networks. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 852-863. Curran Associates, Inc. (2021) +21. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2019) +22. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8110-8119 (2020) +23. Kim, H., Choi, Y., Kim, J., Yoo, S., Uh, Y.: Exploiting spatial dimensions of latent in gan for real-time image editing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 852-861 (June 2021) +24. Kingma, D.P., Welling, M.: Auto-encoding variational bayes (2022) +25. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) +26. Kurach, K., Lucic, M., Zhai, X., Michalski, M., Gelly, S.: A large-scale study on regularization and normalization in GANs. In: Chaudhuri, K., Salakhutdinov, R. (eds.) Proceedings of the 36th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 97, pp. 3581-3590. PMLR (09-15 Jun 2019) +27. Kwon, G., Ye, J.C.: Clipstyler: Image style transfer with a single text condition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18062-18071 (June 2022) +28. Kwon, J., Kim, J., Park, H., Choi, I.K.: Asam: Adaptive sharpness-aware minimization for scale-invariant learning of deep neural networks (2021) +29. Kynkänniemi, T., Karras, T., Laine, S., Lehtinen, J., Aila, T.: Improved precision and recall metric for assessing generative models. Advances in Neural Information Processing Systems 32 (2019) +30. Lee, G., Kim, H., Kim, J., Kim, S., Ha, J.W., Choi, Y.: Generator knows what discriminator should learn in unconditional gans. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 406-422. Springer Nature Switzerland, Cham (2022) +31. Liao, W., Hu, K., Yang, M.Y., Rosenhahn, B.: Text to image generation with semantic-spatial aware gan. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18187-18196 (June 2022) +32. Liu, H., Zhang, W., Li, B., Wu, H., He, N., Huang, Y., Li, Y., Ghanem, B., Zheng, Y.: Adaptivemix: Improving gan training via feature space shrinkage. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16219-16229 (June 2023) + +33. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: Proceedings of the IEEE international conference on computer vision. pp. 3730-3738 (2015) +34. Lu, Z., Sreekumar, G., Goodman, E., Banzhaf, W., Deb, K., Boddeti, V.N.: Neural architecture transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(9), 2971-2989 (Sep 2021) +35. Luo, S., Guan, J., Ma, J., Peng, J.: A 3d generative model for structure-based drug design. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 6229-6239. Curran Associates, Inc. (2021) +36. Mao, X., Li, Q., Xie, H., Lau, R.Y., Wang, Z., Paul Smolley, S.: Least squares generative adversarial networks. In: Proceedings of the IEEE international conference on computer vision. pp. 2794-2802 (2017) +37. Miyato, T., Kataoka, T., Koyama, M., Yoshida, Y.: Spectral normalization for generative adversarial networks (2018) +38. Mukherjee, S., Asnani, H., Lin, E., Kannan, S.: Clustergan: Latent space clustering in generative adversarial networks. Proceedings of the AAAI Conference on Artificial Intelligence 33(01), 4610-4617 (Jul 2019) +39. Nguyen-Phuoc, T., Li, C., Theis, L., Richardt, C., Yang, Y.L.: Hologan: Unsupervised learning of 3d representations from natural images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7588-7597 (2019) +40. Radford, A., Metz, L., Chintala, S.: Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434 (2015) +41. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision 115, 211-252 (2015) +42. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (Oct 2017) +43. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision (2015) +44. Tan, M., Le, Q.: Efficientnetv2: Smaller models and faster training. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 10096-10106. PMLR (18-24 Jul 2021) +45. Tran, N.T., Bui, T.A., Cheung, N.M.: Dist-gan: An improved gan using distance constraints. In: Proceedings of the European Conference on Computer Vision (ECCV) (September 2018) +46. Tseng, H.Y., Jiang, L., Liu, C., Yang, M.H., Yang, W.: Regularizing generative adversarial networks under limited data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7921-7931 (June 2021) +47. Wang, J., Yang, C., Xu, Y., Shen, Y., Li, H., Zhou, B.: Improving gan equilibrium by raising spatial awareness. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11285-11293 (June 2022) +48. Wu, J.Z., Ge, Y., Wang, X., Lei, S.W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7623-7633 (October 2023) + +49. Xiangli, Y., Deng, Y., Dai, B., Loy, C.C., Lin, D.: Real or not real, that is the question. arXiv preprint arXiv:2002.05512 (2020) +50. Xu, Y., AlBahar, B., Huang, J.B.: Temporally consistent semantic video editing. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 357-374. Springer Nature Switzerland, Cham (2022) +51. Yang, C., Shen, Y., Xu, Y., Zhao, D., Dai, B., Zhou, B.: Improving gans with a dynamic discriminator. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. vol. 35, pp. 15093-15104. Curran Associates, Inc. (2022) +52. Yang, C., An, Z., Zhu, H., Hu, X., Zhang, K., Xu, K., Li, C., Xu, Y.: Gated convolutional networks with hybrid connectivity for image classification. Proceedings of the AAAI Conference on Artificial Intelligence 34(07), 12581-12588 (Apr 2020) +53. Yu, F., Seff, A., Zhang, Y., Song, S., Funkhouser, T., Xiao, J.: Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365 (2015) +54. Zhao, J., Mathieu, M., LeCun, Y.: Energy-based generative adversarial network. arXiv preprint arXiv:1609.03126 (2016) +55. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (Oct 2017) \ No newline at end of file diff --git a/2024/CLR-GAN_ Improving GANs Stability and Quality via Consistent Latent Representation and Reconstruction/images.zip b/2024/CLR-GAN_ Improving GANs Stability and Quality via Consistent Latent Representation and Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c6bd6f86400c9f31b66d292fae96be04e0291884 --- /dev/null +++ b/2024/CLR-GAN_ Improving GANs Stability and Quality via Consistent Latent Representation and Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66c4363b487aa9a1c893469fe9aa5703ff17039db6b51e914cba318e60bc4471 +size 367200 diff --git a/2024/CLR-GAN_ Improving GANs Stability and Quality via Consistent Latent Representation and Reconstruction/layout.json b/2024/CLR-GAN_ Improving GANs Stability and Quality via Consistent Latent Representation and Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..11671776a884ea0f4c194e90c8c9269cdf685f9e --- /dev/null +++ b/2024/CLR-GAN_ Improving GANs Stability and Quality via Consistent Latent Representation and Reconstruction/layout.json @@ -0,0 +1,8999 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 111, + 482, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 111, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 482, + 163 + ], + "type": "text", + "content": "CLR-GAN: Improving GANs Stability and Quality via Consistent Latent Representation and Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 143, + 186, + 472, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 186, + 472, + 210 + ], + "spans": [ + { + "bbox": [ + 143, + 186, + 472, + 210 + ], + "type": "text", + "content": "Shengke Sun1*, Ziqian Luan2*, Zhanshan Zhao1,3, Shijie Luo1, and Shuzhen Han1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 145, + 219, + 468, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 219, + 468, + 243 + ], + "spans": [ + { + "bbox": [ + 145, + 219, + 468, + 243 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 219, + 468, + 243 + ], + "type": "text", + "content": " Tiangong University, 300387 Tianjin, China {sunshengke,hanshuzhen,luoshijie}@tiangong.edu.cn, zhzhsh127@163.com" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 203, + 243, + 410, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 243, + 410, + 252 + ], + "spans": [ + { + "bbox": [ + 203, + 243, + 410, + 252 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 203, + 243, + 410, + 252 + ], + "type": "text", + "content": " Xidian University, 710126 Xi'an, Shaanxi, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 263, + 255, + 351, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 255, + 351, + 264 + ], + "spans": [ + { + "bbox": [ + 263, + 255, + 351, + 264 + ], + "type": "text", + "content": "tyuwe@foxmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 153, + 264, + 459, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 264, + 459, + 297 + ], + "spans": [ + { + "bbox": [ + 153, + 264, + 459, + 297 + ], + "type": "text", + "content": "3 Institute for Sustainable Industries & Liveable Cities, Victory University, Melbourne, VIC 8001, Australia \n†Corresponding author" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 329, + 455, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 329, + 455, + 529 + ], + "spans": [ + { + "bbox": [ + 159, + 329, + 455, + 529 + ], + "type": "text", + "content": "Abstract. Generative Adversarial Networks(GANs) have received considerable attention due to its outstanding ability to generate images. However, training a GAN is hard since the game between the Generator(G) and the Discriminator(D) is unfair. Towards making the competition fairer, we propose a new perspective of training GANs, named Consistent Latent Representation and Reconstruction(CLR-GAN). In this paradigm, we treat the G and D as an inverse process, the discriminator has an additional task to restore the pre-defined latent code while the generator also needs to reconstruct the real input, thus obtaining a relationship between the latent space of G and the out-features of D. Based on this prior, we can put D and G on an equal position during training using a new criterion. Experimental results on various datasets and architectures prove our paradigm can make GANs more stable and generate better quality images(31.22% gain of FID on CIFAR10 and 39.5% on AFHQ-Cat, respectively). We hope that the proposed perspective can inspire researchers to explore different ways of viewing GANs training, rather than being limited to a two-player game. The code is publicly available at https://github.com/Petecheco/CLR-GAN." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 538, + 453, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 538, + 453, + 561 + ], + "spans": [ + { + "bbox": [ + 159, + 538, + 453, + 561 + ], + "type": "text", + "content": "Keywords: Generative Adversarial Networks " + }, + { + "bbox": [ + 159, + 538, + 453, + 561 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 538, + 453, + 561 + ], + "type": "text", + "content": " Image generation " + }, + { + "bbox": [ + 159, + 538, + 453, + 561 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 538, + 453, + 561 + ], + "type": "text", + "content": " Latent space consistency" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 582, + 230, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 582, + 230, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 582, + 230, + 594 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 609, + 482, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 609, + 482, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 609, + 482, + 646 + ], + "type": "text", + "content": "Generative models are an important part of computer vision, they can be widely used in fields such as image generation [12, 16, 20, 24], video editing [3, 5, 48, 50], style transfer [9, 27, 55] and even drug discovery [7, 35]. Among various generative" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 324, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 324, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 324, + 665 + ], + "type": "text", + "content": "* Equal contribution. Listing order is random." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 121, + 473, + 295 + ], + "blocks": [ + { + "bbox": [ + 145, + 121, + 473, + 295 + ], + "lines": [ + { + "bbox": [ + 145, + 121, + 473, + 295 + ], + "spans": [ + { + "bbox": [ + 145, + 121, + 473, + 295 + ], + "type": "image", + "image_path": "de061eb9c0fdf1bad66007ead92185b8e996efcac31478bf912a30d288a3203d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 308, + 482, + 354 + ], + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 354 + ], + "type": "text", + "content": "Fig. 1: Sketch map of the proposed method, we proposed two additional learning objectives to consider D and G as inverse process. The top is latent consistency loss where D tries to rebuild the latent distribution of G. The bottom is reconstruction loss where G tries to restore the original real images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 377, + 482, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 474 + ], + "type": "text", + "content": "models, Generative Adversarial Networks(GANs) have received considerable attention due to its remarkable performance in generating realistic images [12]. A conventional GAN comprises of two individual networks named generator(G) and discriminator(D). While training a GAN, the generator aims to produce indistinguishable images, and the discriminator tries to figure out the synthesized images from the input images. By the competition between D and G, ideally, we can eventually train a generator that can recover the real distribution and generate high quality samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": "However, it is well known that training a GAN is hard and unstable [26,37]. In general, the instability of training GANs can be attributed to the unequal set up of generator(G) and discriminator(D). Specifically, in the traditional paradigm, the training process is formulated as a two-player game. But in fact, the discriminator seems to dominate the game [2,47]. In practice, the discriminator can easily identify the fake samples from a relatively early stage of the training and maintain this advantage during the whole training, making the generator hard to converge [51]. Many attempts have been done to improve the stability of training GANs. Previous studies attempted to find new metric functions [1, 14, 32, 45], enhance the generator's ability to generate [4,21,40], or use data augment methods [19,30]. These solutions have shown some degree of improvement on training stability, but they did not fundamentally change the core relationship between the generator and discriminator, while ignoring the essence of dynamic balance within GANs architecture. The competition between the generator and discriminator is still unfair. In addition, existing methods have not fully utilized the structural information of the latent space, which limits the further improvement" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "type": "text", + "content": "S.Sun, Z.Luan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "of GANs generation ability. Therefore, improving the training of GANs still remains an unsolved and challenging problem." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 344 + ], + "type": "text", + "content": "In this paper, contrast to existing methods, we proposed a novel perspective of training GANs. By rethinking the training process and network architecture of GANs. We found that the generator and the discriminator got optimized in a considerably large latent space, resulting unstable training and inconsistency between the generator and the discriminator. While traditional methods primarily focus on adjusting network architectures or tuning hyperparameters to stabilize training [19, 40]. Our method is to view the generator and the discriminator as two inverse components. As shown in Fig. 1. In particular, the discriminator can be regarded as outputting a realness score and extracting samples to the latent representation space at the same time. And the generator not only generates images but restores source images from the extracted latent representation. By constraining the generator and discriminator with these additional tasks, we make the discriminator's representation latent space closer to the generator's latent space, thus making the generator and the discriminator more consistent, leading a fairer game between the generator and the discriminator. This not only provides a more robust framework for training GANs but also opens up new possibilities for improving GANs training from different aspects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 364, + 237, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 237, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 237, + 376 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 392, + 367, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 392, + 367, + 404 + ], + "spans": [ + { + "bbox": [ + 131, + 392, + 367, + 404 + ], + "type": "text", + "content": "2.1 Improving GANs through discriminators" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 414, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 666 + ], + "type": "text", + "content": "As we mentioned above, due to the limitations of original training objective [1, 12], the discriminator tends to easily take over the competition during the early stage of training. Prior works have been devoted to improving the inequality using modified discriminators. Dist-GAN [45] decelerate the convergence of the discriminator by constructing a reconstruction constrain using a auxiliary autoencoder. Wang et al. [47] combined Grad-CAM [42] with GANs to improve the spatial awareness of the generator using the discriminator as a regularizer, lessening the information gap between D and G. DynamicGAN [51] controls the strength of the discriminator by gradually increasing D's learning capacity, thereby preventing the discriminator from being overly dominant in the early stages of training. AdaptiveMix [32] shrinks the feature space of the discriminator by training using constructed hard samples, reducing the feature space of the discriminator and making training easier. There are also some data augmentation methods that enhance model training by providing more data to the model [19, 46]. However, the above methods have not changed the inner properties of training. All gradient during the optimization process is provided by D, so D always occupies a more dominant position during the training process, which will prevent further optimization of the model. While our model, based on limiting D, also constrains D to optimize based on the gradient provided by G. This allows both G and D to optimize themselves using the gradients provided by each other, making the game fairer." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLR-GAN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 272, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 272, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 272, + 128 + ], + "type": "text", + "content": "2.2 Latent space of GANs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 290 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 290 + ], + "type": "text", + "content": "The latent space of GANs has always been a popular research area due to its rich semantic information. Multi-Code GAN [13] combines multiple latent feature maps with adaptive channels to improve the quality of image reconstruction. ClusterGAN [38] combines latent space encoding and one-hot encoding to obtain a better clustering result. StyleMapGAN [23] utilizes latent space as an accurate feature embedding, improving the performance of image editing and interpolation tasks. Although latent space plays an important role in various downstream tasks. But to our best knowledge, there has been no previous work that directly uses the latent space of GANs as a component to optimize the training of GANs. In our proposed method, we have directly incorporated latent space into the training process of GANs as an additional constrain. The experimental results show that applying this method can effectively improve the quality and diversity of generated images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 307, + 389, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 307, + 389, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 389, + 319 + ], + "type": "text", + "content": "2.3 Differences to Reconstruction-based Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 325, + 482, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 481 + ], + "type": "text", + "content": "Prior works have explored the application of reconstruction-based methods within GANs. HoloGAN [39] and PiGAN [6] utilize latent reconstruction to learn the 3D-aware image representations. GLeaD [2] takes advantage of image reconstruction to improve the training dynamics of GANs. However, these methods either lack a specific focus on stabilizing GANs training or solely concentrate on improving the generator without considering the discriminator. Specifically, in the GLeaD paradigm, given an image, we expect D to extract representative features that can be adequately decoded by G to reconstruct the input, however, this paradigm only improves the generator, the discriminator still controls the game. To address these limitations, we use the latent spatial code extracted by the discriminator and reconstruction loss to achieve stable GAN training and generate higher quality and more diverse images. Thus making GANs training more efficient and practical." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 498, + 202, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 202, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 202, + 510 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "As we mentioned before, we try to augment the training fairness of GANs via considering the generative process of the generator (G) and the discriminative process of the discriminator (D) as inverse to each other, thereby enhancing the stability and quality of generated samples. To better understand this new perspective, we first review the conventional formulation of GANs in Sec. 3.1. To obtain a fairer training, Sec. 3.2 introduces our new training paradigm named Consistent Latent Representation that considers G and D as an inverse process. Furthermore, in Sec. 3.3, we also propose a Real Image Reconstruction strategy to achieve mutual connection between generator and the discriminator. Then Sec. 3.4 explained our method from a regularization perspective. Finally, Sec. 3.5 gives the full optimization objective and shows the pseudo code of the new training paradigm." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "type": "text", + "content": "S.Sun, Z.Luan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 220, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 220, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 220, + 128 + ], + "type": "text", + "content": "3.1 Preliminary" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 479, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 479, + 193 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 479, + 193 + ], + "type": "text", + "content": "A GAN usually consists of two components: a generator " + }, + { + "bbox": [ + 130, + 133, + 479, + 193 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 130, + 133, + 479, + 193 + ], + "type": "text", + "content": " and a discriminator " + }, + { + "bbox": [ + 130, + 133, + 479, + 193 + ], + "type": "inline_equation", + "content": "D(\\cdot)" + }, + { + "bbox": [ + 130, + 133, + 479, + 193 + ], + "type": "text", + "content": ". The generator aims to map a latent code " + }, + { + "bbox": [ + 130, + 133, + 479, + 193 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 133, + 479, + 193 + ], + "type": "text", + "content": " into an image, while the discriminator tries to distinguish the generated image " + }, + { + "bbox": [ + 130, + 133, + 479, + 193 + ], + "type": "inline_equation", + "content": "G(z)" + }, + { + "bbox": [ + 130, + 133, + 479, + 193 + ], + "type": "text", + "content": " from the real one " + }, + { + "bbox": [ + 130, + 133, + 479, + 193 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 133, + 479, + 193 + ], + "type": "text", + "content": ". The conventional GAN trains via a two-player game by optimizing the learning objective as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 200, + 480, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 200, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 146, + 200, + 480, + 217 + ], + "type": "interline_equation", + "content": "\\min _ {G} \\max _ {D} V (D, G) = \\mathbb {E} _ {x \\sim p _ {d a t a} (x)} [ \\log D (x) ] + \\mathbb {E} _ {z \\sim p _ {z} (z)} [ \\log (1 - D (G (z))) ] \\quad (1)", + "image_path": "1c874a07424e8cd34d2afc92d05bf307be0ce9e04ca75f6594a042639a9e6f35.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 222, + 479, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 222, + 479, + 245 + ], + "spans": [ + { + "bbox": [ + 130, + 222, + 479, + 245 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 222, + 479, + 245 + ], + "type": "inline_equation", + "content": "p_{z}(z)" + }, + { + "bbox": [ + 130, + 222, + 479, + 245 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 222, + 479, + 245 + ], + "type": "inline_equation", + "content": "p_{data}(x)" + }, + { + "bbox": [ + 130, + 222, + 479, + 245 + ], + "type": "text", + "content": " represent a random latent representation and real data distribution respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 246, + 480, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 480, + 318 + ], + "type": "text", + "content": "Goodfellow et al. [12] have proved that the optimal solution to this objective is that " + }, + { + "bbox": [ + 130, + 246, + 480, + 318 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 246, + 480, + 318 + ], + "type": "text", + "content": " can eventually recover the distribution of source data and " + }, + { + "bbox": [ + 130, + 246, + 480, + 318 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 246, + 480, + 318 + ], + "type": "text", + "content": " can not separate the generated images out from real images. However, the game between two players is not fair since in Eq. 1 that the gradient for optimizing all comes from " + }, + { + "bbox": [ + 130, + 246, + 480, + 318 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 246, + 480, + 318 + ], + "type": "text", + "content": ", which makes the discriminator a natural dominant [2]. Thus the ideal solution is hard to reach in practice. [11,47]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 333, + 283, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 283, + 344 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 283, + 344 + ], + "type": "text", + "content": "3.2 Consistent Latent Space" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 350, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 517 + ], + "type": "text", + "content": "We aim to improve the stability of GANs by introducing a novel perspective, that is to consider the generate and discriminate as an inverse process. Recall that in a regular GAN paradigm, the generator and discriminator are treated as two individual components. While in our proposed view, we have refined the concepts of generator and discriminator and consider them as an inverse process. In our perspective, the generator maps a latent representation to the real data distribution, and the discriminator not only outputs a realizness score, but it also transform the high-dimensional distribution into a trainable latent space. By measuring the distance between these two spaces, we can improve GANs training in the following aspects: First, we can to some extent make the generator and discriminator behave more consistent in the mapping process. Second, since the discriminator needs to align with the generator, it can be seen as constraining the discriminator using generator, which makes the game not completely dominated by discriminator. The detailed implementation is shown in Fig. 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "content": "Extracting Latent Representation through D. In the training of GANs, the discriminator plays a role of downsample high-dimensional data. For any discriminator, it can be divided into two parts, the first part is the feature extractor denoted as " + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "content": " which extracts features from high-dimensional data samples, another serves as a affine from extracted features to realizness score to tell whether the input is true(real samples) or false(generated samples). For any input data " + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "content": " and latent code " + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "content": ", the whole discriminate process can be denoted as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 264, + 635, + 480, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 635, + 480, + 647 + ], + "spans": [ + { + "bbox": [ + 264, + 635, + 480, + 647 + ], + "type": "interline_equation", + "content": "D (x) = R (f (x)) \\tag {2}", + "image_path": "d518a859533d71fbfe088b3eb870b2200b9de9fd70a6d07c11933062dd807cd3.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 265, + 650, + 480, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 650, + 480, + 662 + ], + "spans": [ + { + "bbox": [ + 265, + 650, + 480, + 662 + ], + "type": "interline_equation", + "content": "D (z) = R (f (G (z))) \\tag {3}", + "image_path": "41f7515bc33bfe675603e96ea0dcd6d156a816f08860da41418134b967a759e5.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLR-GAN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 119, + 478, + 201 + ], + "blocks": [ + { + "bbox": [ + 140, + 119, + 478, + 201 + ], + "lines": [ + { + "bbox": [ + 140, + 119, + 478, + 201 + ], + "spans": [ + { + "bbox": [ + 140, + 119, + 478, + 201 + ], + "type": "image", + "image_path": "ce52d80ecda9abf83c85b0e4a10a31fc0295544a528d3f3e87522430bf962476.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 216, + 482, + 271 + ], + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 271 + ], + "type": "text", + "content": "Fig. 2: Detailed implementation of Consistent Latent Representation. Given a latent code " + }, + { + "bbox": [ + 130, + 216, + 482, + 271 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 216, + 482, + 271 + ], + "type": "text", + "content": ", the generator maps " + }, + { + "bbox": [ + 130, + 216, + 482, + 271 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 216, + 482, + 271 + ], + "type": "text", + "content": " to an image. The synthesised image was then got discriminated by the discriminator, outputting a realness score and a reconstructed latent code " + }, + { + "bbox": [ + 130, + 216, + 482, + 271 + ], + "type": "inline_equation", + "content": "z'" + }, + { + "bbox": [ + 130, + 216, + 482, + 271 + ], + "type": "text", + "content": ". The Consistent Latent loss is calculated between the real latent " + }, + { + "bbox": [ + 130, + 216, + 482, + 271 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 216, + 482, + 271 + ], + "type": "text", + "content": " and the reconstructed latent " + }, + { + "bbox": [ + 130, + 216, + 482, + 271 + ], + "type": "inline_equation", + "content": "z'" + }, + { + "bbox": [ + 130, + 216, + 482, + 271 + ], + "type": "text", + "content": ". The right part specifically shows the architecture of D." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 297, + 482, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 320 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 297, + 482, + 320 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 130, + 297, + 482, + 320 + ], + "type": "text", + "content": " is the feature extractor module of discriminator and " + }, + { + "bbox": [ + 130, + 297, + 482, + 320 + ], + "type": "inline_equation", + "content": "R(\\cdot)" + }, + { + "bbox": [ + 130, + 297, + 482, + 320 + ], + "type": "text", + "content": " is the affine module to transform feature space to the realness score." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 323, + 495, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 495, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 495, + 417 + ], + "type": "text", + "content": "Through this discriminate process, we can transform structural high-dimensional input into feature map that contains abundant semantic information [31], thus judging the realness of the input data. We discovered both latent code in the generator and feature map in the discriminator contain lots semantic features, making it possible to align the generator and discriminator using a mapping function. The mapping function converts discriminator's feature space to latent space. Finally we can get a reconstructed latent code " + }, + { + "bbox": [ + 130, + 323, + 495, + 417 + ], + "type": "inline_equation", + "content": "\\phi (x)" + }, + { + "bbox": [ + 130, + 323, + 495, + 417 + ], + "type": "text", + "content": " from discriminator as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 272, + 434, + 481, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 434, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 272, + 434, + 481, + 449 + ], + "type": "interline_equation", + "content": "\\phi (x) = \\sigma (f (x)) \\tag {4}", + "image_path": "c957b6730f1936a90c2bfbfe1a857fb39618d4e714308cc8431263a5a7ba4c49.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 456, + 482, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 493 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 456, + 482, + 493 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 456, + 482, + 493 + ], + "type": "text", + "content": " is the real input, " + }, + { + "bbox": [ + 130, + 456, + 482, + 493 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 130, + 456, + 482, + 493 + ], + "type": "text", + "content": " is the feature extractor of D and " + }, + { + "bbox": [ + 130, + 456, + 482, + 493 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 130, + 456, + 482, + 493 + ], + "type": "text", + "content": " is a mapping function that transform feature to latent space, we use an identical affine in this paper for simplicity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 516, + 487, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 487, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 487, + 564 + ], + "type": "text", + "content": "Consistent Latent Representation Loss. As we mentioned above, we now have a real latent code " + }, + { + "bbox": [ + 130, + 516, + 487, + 564 + ], + "type": "inline_equation", + "content": "(z)" + }, + { + "bbox": [ + 130, + 516, + 487, + 564 + ], + "type": "text", + "content": " from the generator and a reconstructed latent code(Eq.4) from the discriminator. The latent representation loss can be defined by measuring the distance between these two latent spaces:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 257, + 576, + 481, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 576, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 257, + 576, + 481, + 590 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C L R} = \\operatorname {D i s t} (\\phi (x), z) \\tag {5}", + "image_path": "7955534f6e3b808ca47d95eb922dd2f4cf839a76b52faa89a3c3352b5e24aab7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 600, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 600, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 600, + 482, + 635 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 600, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\text{Dist}(\\cdot)" + }, + { + "bbox": [ + 130, + 600, + 482, + 635 + ], + "type": "text", + "content": " is distance measurement for any two distributions. In this paper, we use L1-Distance similar to [55] for all distance measurement, so the " + }, + { + "bbox": [ + 130, + 600, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{clr}}" + }, + { + "bbox": [ + 130, + 600, + 482, + 635 + ], + "type": "text", + "content": " can be rewritten as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 244, + 637, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 637, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 244, + 637, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C L R} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\| z _ {i} - \\phi (x _ {i}) \\| _ {1} \\tag {6}", + "image_path": "b822edb271b287766fb958a7aa3f819c03bd93321bb5209919ee61b3429cb8ca.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 249, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 249, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 249, + 101 + ], + "type": "text", + "content": "S.Sun, Z.Luan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 297, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 297, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 297, + 128 + ], + "type": "text", + "content": "3.3 Real Image Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 220 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 220 + ], + "type": "text", + "content": "As mentioned in Sec. 2, in a regular GAN, the generator optimized through the gradient offered by discriminator. Also during the training, the generator has no access to real data distributions, it can only update its parameters implicitly, while the discriminator can get optimized according to real data distribution. This unfair situation helps the discriminator converge faster at an early stage of the training [47], thus the discriminator can not provide useful gradient in the later stage of training, making the generator not perform well." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 221, + 482, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 221, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 482, + 340 + ], + "type": "text", + "content": "To solve this problem, we proposed a new strategy called Real Image Reconstruction, which enables the generator to also get gradient from the real data distribution. Specifically, since we have strengthened the discriminator to output both realizness and latent code in the last section, we can then use latent code as a bridge to indirectly connect the generator with real distribution. Subsequently, we can use the generated images for a image reconstruction task. The generator then can utilize this gradient that comes from real distribution for better synthesis, making the whole process more efficient. In practice, we can use Eq. 4 to get the reconstructed latent code of real distributions, then we use generator for image restoration." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 275, + 341, + 480, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 341, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 275, + 341, + 480, + 354 + ], + "type": "interline_equation", + "content": "I _ {r e c} = G (\\phi (x)) \\tag {7}", + "image_path": "166bb53c79fa3144ae0699cd91bcf7b4d509a219b2abb471164a3492436025b5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 359, + 481, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 359, + 481, + 394 + ], + "spans": [ + { + "bbox": [ + 131, + 359, + 481, + 394 + ], + "type": "text", + "content": "After reconstructing the image, we can calculate the distance between real images and reconstructed images. And for the reconstruction loss, we use a regular from [55] as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 247, + 396, + 480, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 396, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 247, + 396, + 480, + 426 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\| I (x) - I _ {r e c} \\| _ {1} \\tag {8}", + "image_path": "f46d5afecb1b9a3f31dc6f57d5eaf73e21abae367d0ef2f4a911d87f0aa03031.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 431, + 482, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 431, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 131, + 431, + 482, + 455 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 431, + 482, + 455 + ], + "type": "inline_equation", + "content": "I(x)" + }, + { + "bbox": [ + 131, + 431, + 482, + 455 + ], + "type": "text", + "content": " is the source image data and " + }, + { + "bbox": [ + 131, + 431, + 482, + 455 + ], + "type": "inline_equation", + "content": "I_{rec}" + }, + { + "bbox": [ + 131, + 431, + 482, + 455 + ], + "type": "text", + "content": " is the reconstructed image using generator and reconstructed latent code." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 473, + 460, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 473, + 460, + 487 + ], + "spans": [ + { + "bbox": [ + 131, + 473, + 460, + 487 + ], + "type": "text", + "content": "3.4 Enhancing Training Stability via Regularization Techniques" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 494, + 481, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 494, + 481, + 517 + ], + "spans": [ + { + "bbox": [ + 131, + 494, + 481, + 517 + ], + "type": "text", + "content": "To better understand the superiority of our method, we turn to investigate the above training strategies from the perspective of regularization in this section." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 146, + 518, + 457, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 518, + 457, + 530 + ], + "spans": [ + { + "bbox": [ + 146, + 518, + 457, + 530 + ], + "type": "text", + "content": "With our method, the objective of training GANs can be rewritten as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 167, + 540, + 481, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 540, + 481, + 572 + ], + "spans": [ + { + "bbox": [ + 167, + 540, + 481, + 572 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {G} \\max _ {D} V (D, G) = \\mathbb {E} _ {x \\sim p _ {d a t a}} \\log D (x) + \\mathbb {E} _ {z \\sim p _ {z}} \\log [ 1 - D (G (z)) ] \\tag {9} \\\\ + \\mathbb {E} _ {x \\sim p _ {d a t a}} \\| G (D (x)) - x \\| - \\mathbb {E} _ {x, z} \\| D (x) - z \\| \\\\ \\end{array}", + "image_path": "12494ac2ea16d2f783cd7ce21532ccb6b9c1ec5acf56cc5e3550bd74ca316448.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "type": "text", + "content": "The first two terms of Eq. 9 are the default training objectives for GANs, while the third and fourth terms correspond to Latent Consistency Loss and Real Image Reconstruction Loss, respectively. Note that these two objectives actually appear in the form of L1 norm, and we will consider them as two types of regularization in this section." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "content": "When optimizing networks, we actually calculate the gradients of objective function. For the training objective we propose, the gradients can be expressed" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLR-GAN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 118, + 147, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 118, + 147, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 118, + 147, + 126 + ], + "type": "text", + "content": "as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 144, + 133, + 481, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 133, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 144, + 133, + 481, + 172 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\theta_ {d}, \\theta_ {g}} V (G, D) = \\mathbb {E} _ {x \\sim p _ {d a t a}} [ \\nabla_ {\\theta_ {d}} \\log D (x) ] + \\mathbb {E} _ {z \\sim p _ {z}} [ \\nabla_ {\\theta_ {d}, \\theta_ {g}} \\log D (1 - D (G (z))) ] \\\\ + \\mathbb {E} _ {x \\sim p _ {d a t a}} \\operatorname {s i g n} (G (D (x)) - x) - \\mathbb {E} _ {x, z} \\operatorname {s i g n} (D (x) - z) \\tag {10} \\\\ \\end{array}", + "image_path": "5160e3d1a7b675d88bff478c679e488da0822b59a6fec7abe0c9dabef0f2ce47.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 178, + 479, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 178, + 479, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 479, + 201 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 178, + 479, + 201 + ], + "type": "inline_equation", + "content": "\\operatorname{sign}(G(D(x)) - x)" + }, + { + "bbox": [ + 130, + 178, + 479, + 201 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 178, + 479, + 201 + ], + "type": "inline_equation", + "content": "\\operatorname{sign}(D(x) - z)" + }, + { + "bbox": [ + 130, + 178, + 479, + 201 + ], + "type": "text", + "content": " is simply the sign of both constrains applied element-wise." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 202, + 480, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 202, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 146, + 202, + 480, + 214 + ], + "type": "text", + "content": "Next, we will discuss the role of our regularization in two different scenarios:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 229, + 480, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 229, + 480, + 265 + ], + "spans": [ + { + "bbox": [ + 131, + 229, + 480, + 265 + ], + "type": "text", + "content": "Training the Discriminator with a fixed Generator. When the generator is fixed, we can consider the gradients related to the generator to be 0, so we can obtain the gradients as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 186, + 286, + 480, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 286, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 186, + 286, + 480, + 298 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta_ {d}} V (D) = \\mathbb {E} _ {x \\sim p _ {d a t a}} \\left[ \\nabla_ {\\theta_ {d}} \\log D (x) \\right] - \\mathbb {E} _ {x, z} \\operatorname {s i g n} (D (x) - z) \\tag {11}", + "image_path": "fde5989d16e08ddc6f033b01bd3b055f593cccc1de114fc5ac272fabe372e077.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 305, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 305, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 131, + 305, + 480, + 365 + ], + "type": "text", + "content": "Considering the general training situation of GANs, when the discriminator is too strong during training. The first term in the gradient tends to be large, resulting in unstable gradient descent. However with our regularization term. Due to the negative sign in the front, we can reduce the gradient of D using this term, making the gradient smaller and can stable the training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 380, + 480, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 380, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 131, + 380, + 480, + 403 + ], + "type": "text", + "content": "Training the Generator with a fixed Discriminator. As in the previous case, we first write the representation for the gradient:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 145, + 423, + 480, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 423, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 145, + 423, + 480, + 447 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta_ {g}} V (G) = \\mathbb {E} _ {z \\sim p _ {z}} [ \\nabla_ {\\theta_ {d}, \\theta_ {g}} \\log D (1 - D (G (z))) ] + \\mathbb {E} _ {x \\sim p _ {d a t a}} \\operatorname {s i g n} (G (D (x)) - x) \\tag {12}", + "image_path": "ce3f4f8a6ab295daeedffb6bed24fb2ca825e8c2438934a595696f4d0a429072.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 453, + 479, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 479, + 524 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 479, + 524 + ], + "type": "text", + "content": "When the discriminator is dominating the game, it can make the first term to become very small, making it difficult for the generator to effectively optimize and generate real images. After adding a regularization term, we found an additional training objective was given to the generator, allowing the generator to optimize through this training objective, and continue to compete with the discriminator instead of collapsing." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 525, + 479, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 479, + 572 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 479, + 572 + ], + "type": "text", + "content": "Finally, we can find that by adding two additional training objective, our model can penalize the discriminator when it is too strong. It can optimize the generator by providing more information through additional tasks, making the game fairer and improving the effectiveness of image generation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 587, + 232, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 232, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 232, + 599 + ], + "type": "text", + "content": "3.5 Full Objective" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 605, + 479, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 479, + 628 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 479, + 628 + ], + "type": "text", + "content": "With the enhanced discriminator and the reconstruction guided generator, the overall training objective can be rewritten as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 261, + 635, + 479, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 635, + 479, + 647 + ], + "spans": [ + { + "bbox": [ + 261, + 635, + 479, + 647 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D} ^ {\\prime} = \\mathcal {L} _ {D} + \\lambda_ {1} \\mathcal {L} _ {C L R} \\tag {13}", + "image_path": "917af5e13f8bedd2da95eefa3b5c7c49f0a9c7a07c10dc2daea7e2503710be78.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 261, + 651, + 479, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 651, + 479, + 662 + ], + "spans": [ + { + "bbox": [ + 261, + 651, + 479, + 662 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {G} ^ {\\prime} = \\mathcal {L} _ {G} + \\lambda_ {2} \\mathcal {L} _ {r e c} \\tag {14}", + "image_path": "911e2d81977d1529d08ad65216ba563005daa98bdc250898b6f365135fdf053d.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 249, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 249, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 249, + 101 + ], + "type": "text", + "content": "S.Sun, Z.Luan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 140, + 481, + 273 + ], + "blocks": [ + { + "bbox": [ + 132, + 115, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 481, + 137 + ], + "type": "text", + "content": "Algorithm 1 Training a GAN with Consistent Latent Representation and Reconstruction" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 140, + 481, + 273 + ], + "lines": [ + { + "bbox": [ + 132, + 140, + 481, + 273 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 481, + 273 + ], + "type": "text", + "content": "Require: G and our D that are initialized with random parameters. Training data " + }, + { + "bbox": [ + 132, + 140, + 481, + 273 + ], + "type": "inline_equation", + "content": "\\{x_{i}\\}" + }, + { + "bbox": [ + 132, + 140, + 481, + 273 + ], + "type": "text", + "content": " , latent variable " + }, + { + "bbox": [ + 132, + 140, + 481, + 273 + ], + "type": "inline_equation", + "content": "\\{z\\}" + }, + { + "bbox": [ + 132, + 140, + 481, + 273 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 132, + 140, + 481, + 273 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 132, + 140, + 481, + 273 + ], + "type": "text", + "content": " until converge do Sample " + }, + { + "bbox": [ + 132, + 140, + 481, + 273 + ], + "type": "inline_equation", + "content": "z\\sim P(Z),x\\sim P(X)" + }, + { + "bbox": [ + 132, + 140, + 481, + 273 + ], + "type": "text", + "content": " real_score, real_forent " + }, + { + "bbox": [ + 132, + 140, + 481, + 273 + ], + "type": "inline_equation", + "content": "= D(x)" + }, + { + "bbox": [ + 132, + 140, + 481, + 273 + ], + "type": "text", + "content": " fake_score, fake_forent " + }, + { + "bbox": [ + 132, + 140, + 481, + 273 + ], + "type": "inline_equation", + "content": "= D(G(z))" + }, + { + "bbox": [ + 132, + 140, + 481, + 273 + ], + "type": "text", + "content": " reconstruct_images " + }, + { + "bbox": [ + 132, + 140, + 481, + 273 + ], + "type": "inline_equation", + "content": "= G" + }, + { + "bbox": [ + 132, + 140, + 481, + 273 + ], + "type": "text", + "content": " (real_forent) Calculate Consistent Latent Loss with Eq.(6) Calculate Reconstruction Loss with Eq. (8) Update D and G with Eq. (9), Eq. (10) end for \nOutput: G with best training set FID." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "code", + "guess_lang": "verilog" + }, + { + "bbox": [ + 130, + 296, + 480, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 480, + 343 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 296, + 480, + 343 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_G" + }, + { + "bbox": [ + 130, + 296, + 480, + 343 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 296, + 480, + 343 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 130, + 296, + 480, + 343 + ], + "type": "text", + "content": " are the original loss function of generator and discriminator, " + }, + { + "bbox": [ + 130, + 296, + 480, + 343 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 130, + 296, + 480, + 343 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 296, + 480, + 343 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CLR}" + }, + { + "bbox": [ + 130, + 296, + 480, + 343 + ], + "type": "text", + "content": " are the proposed consistent latent space distance and real image reconstruction distance calculated through Eq. 6 and Eq. 8, " + }, + { + "bbox": [ + 130, + 296, + 480, + 343 + ], + "type": "inline_equation", + "content": "\\lambda_1" + }, + { + "bbox": [ + 130, + 296, + 480, + 343 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 296, + 480, + 343 + ], + "type": "inline_equation", + "content": "\\lambda_2" + }, + { + "bbox": [ + 130, + 296, + 480, + 343 + ], + "type": "text", + "content": " are weighting coefficient of the additional loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 344, + 481, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 481, + 439 + ], + "type": "text", + "content": "With this updated objective, we can stabilize GANs training in the following two aspects: First, we align the generator and discriminator during training phase by assigning a consistent latent representation task for the discriminator, which makes the game between generator and discriminator fairer. Second, we enable the generator to get access to the real data through a reconstruction task. This helps generator make full use of alternative information, leading to a more realistic output. The pseudo code of our method is shown in Algorithm 1. Extensive experiments result in Sec 4. showed the effectiveness of our method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 455, + 230, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 455, + 230, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 455, + 230, + 468 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 478, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 478, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 478, + 482, + 597 + ], + "type": "text", + "content": "To validate the effectiveness of the proposed method, we conducted extensive experiments on various GAN architectures and datasets. Sec. 4.1 first introduces the detailed settings of our experiments. In Sec. 4.2, we proved that our method achieves better quality and stability through quantitative evaluation on various image generation datasets. Sec. 4.3 includes ablation study to the proposed Consistent Latent loss and Reconstruction loss to better understand the designed components. In Sec. 4.4 we prove that we can get a better feature extractor by unsupervised classification. At last, we visualize the realness curve of D to validate the improved fairness of GANs in Sec. 4.5. The generated samples can be found in the Supplementary Material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 612, + 261, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 261, + 625 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 261, + 625 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Datasets. For a more comprehensive evaluation of the proposed method, the model was tested on low resolution images and high resolution images separately. We used DCGAN [40] as the basic architecture for low resolution image" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLR-GAN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 184, + 146, + 430, + 270 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 1: FIDs of DCGAN [40] using different learning objectives on CelebA [33] and CIFAR-10 [25] dataset. The bold numbers indicate the best result for each dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 184, + 146, + 430, + 270 + ], + "lines": [ + { + "bbox": [ + 184, + 146, + 430, + 270 + ], + "spans": [ + { + "bbox": [ + 184, + 146, + 430, + 270 + ], + "type": "table", + "html": "
Learning ObjectiveCelebACIFAR-10
WGAN [1]36.4755.96
HingeGAN [54]25.5742.4
LSGAN [36]30.7642.01
DCGAN [40]27.0238.56
WGAN-GP [14]70.2841.86
Realness GAN-Obj.1 [49]-36.73
Realness GAN-Obj.2 [49]23.5134.59
Realness GAN-Obj.3 [49]-36.21
AdaptiveMix [32]12.4330.85
CLR-GAN(Ours)13.6323.3
", + "image_path": "774e4e130ccc7b279695047a810647193fd8861c2e3bb3eb51cada7ebe76f3cb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 290, + 480, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 480, + 352 + ], + "type": "text", + "content": "generation, with CIFAR-10 [25] dataset and Celeba64 " + }, + { + "bbox": [ + 130, + 290, + 480, + 352 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 130, + 290, + 480, + 352 + ], + "type": "text", + "content": " 64 [33] dataset. For high-resolution image generation, we used StyleGAN-V2 [22] as the basic architecture, and test the generation ability under AFHQ Cat [8], LSUN Church [53], and FFHQ [21] dataset corresponding to animal generation, natural scene generation, and face generation respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 366, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 452 + ], + "type": "text", + "content": "Evaluation. During evaluation phase, we mainly used the most common Frechet Inception Distance(FID) [15] for quantitatively represent the image generation quality. Besides, Precision & Recall [29] is also adopted to provide more details of the realness and diversity of the proposed model. In practice, we calculate FID between 20K generated images and all real samples on CIFAR-10 and CelebA dataset. While for AFHQ-Cat, LSUN-Church and FFHQ, we calculate the FID and P&R between 50K generated images and all real samples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 465, + 482, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 465, + 482, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 465, + 482, + 513 + ], + "type": "text", + "content": "Other settings. For all baselines, our experimental hyperparameters were set according to the original paper. We used pre-trained Inception-V3 [43] as the feature extractor for FID and P&R calculation. For the weights of the additional loss, we set " + }, + { + "bbox": [ + 130, + 465, + 482, + 513 + ], + "type": "inline_equation", + "content": "\\lambda_1 = 5" + }, + { + "bbox": [ + 130, + 465, + 482, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 465, + 482, + 513 + ], + "type": "inline_equation", + "content": "\\lambda_2 = 0.5" + }, + { + "bbox": [ + 130, + 465, + 482, + 513 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 528, + 335, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 528, + 335, + 540 + ], + "spans": [ + { + "bbox": [ + 131, + 528, + 335, + 540 + ], + "type": "text", + "content": "4.2 Performance on Image Generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 593 + ], + "type": "text", + "content": "Low-resolution Image Generation. We first conducted experiments on low resolution image datasets to verify the effectiveness of the proposed method. We used DCGAN [40] as the baseline architecture and compared our method with some well-known learning objectives on CIFAR10 and CelebA datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Tab. 1 presents the results. From the perspective of FID, we can directly see a significant margin between the proposed method and previous learning objectives. Furthermore, compared with the recently proposed learning objective [32], our method outperforms " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "24.5\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " on the CIFAR-10 dataset while remains a competitive result on CIFAR-10 dataset " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "(8\\%)" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " behind), which further demonstrates the superiority of our model." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "type": "text", + "content": "S.Sun, Z.Luan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 167, + 480, + 268 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 2: Comparisons on AFHQ-Cat [8], FFHQ [21] and LSUN-Church [53] with different high-resolution image generative models. P and R denote precision and recall. The bold numbers indicate the best result for each dataset. The blue numbers indicate the improvements." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 167, + 480, + 268 + ], + "lines": [ + { + "bbox": [ + 134, + 167, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 134, + 167, + 480, + 268 + ], + "type": "table", + "html": "
MethodAFHQ-CatFFHQLSUN Church
FID ↓P ↑R ↑FID ↓P ↑R ↑FID ↓P ↑R ↑
StyleGAN-V2 [22]7.920.680.273.860.680.254.040.580.40
LC-Reg [46]6.70--3.93--4.07--
StyleGAN-V2-ADA [19]6.050.660.254.010.660.264.010.610.43
StyleGAN-V2-APA [18]4.880.650.303.750.670.293.920.600.43
StyleGAN-V2 + Ours4.79(-3.13)0.760.283.44(-0.42)0.700.413.52(-0.52)0.630.46
StyleGAN-V2-ADA + Ours4.45(-0.43)0.740.333.37(-0.64)0.710.443.43(-0.58)0.610.48
", + "image_path": "e5d155e39b3c5e53605018b757aa0c9f23287726310eb33fc7926c09a7a671f3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 290, + 479, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 479, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 479, + 373 + ], + "type": "text", + "content": "High-resolution Image Generation. To further demonstrate the effectiveness of the proposed method, we also considered the recently proposed high-resolution image synthesis architecture StyleGAN-V2 [22]. We compared our method with other StyleGAN-V2 variants methods on AFHQ-Cat, LSUN Church and FFHQ datasets, respectively. We also combined our method with recent proposed works that aims to improve GANs training called StyleGAN-V2-ADA [19] to show the compatibility of our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 375, + 480, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 480, + 446 + ], + "type": "text", + "content": "As shown in Tab. 2, there is a substantially improvement with the proposed method on StyleGAN-V2 for both datasets, outperforming other methods with a clear margin. We also found that combined with StyleGAN-V2-ADA, our method can get a further improvement, achieving a better result on various datasets. Therefore, our proposed method can improve the synthesis quality of high-resolution images while also integrating well with previous works." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 447, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 542 + ], + "type": "text", + "content": "As for the Precision and Recall. The improvements among multiple datasets are clear, indicating that our model can not only learn to generate more realistic images, but also enhance the diversity of generated images, to some extent solving the problem of insufficient diversity of conventional GANs. This may suggest that CLR-GAN can benefit from the proposed two constrains that leads to learning a wider distribution, thereby improving the diversity of generated samples. More validation metrics and experimental results on the more complex ImageNet [41] dataset are shown in the Supplementary Material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 560, + 238, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 560, + 238, + 572 + ], + "spans": [ + { + "bbox": [ + 132, + 560, + 238, + 572 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "In the ablation experiment, we primarily focus on the impact of the two new training constrains we added during training. To ensure the generality, we selected both the low-resolution CIFAR-10 dataset and the high-resolution FFHQ dataset as two baseline datasets. Recall that we can use " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " to control the strength of the proposed constrains in Eq. 13. Therefore we can modify different " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " values in the ablation study to figure out the optimal values of two constrains. We first set " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = \\lambda_{2} = 0" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " to get the baseline values." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLR-GAN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 148, + 156, + 294, + 260 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 3: Ablation studies of different weights " + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": " on low-resolution CIFAR-10 dataset and high-resolution FFHQ dataset. The best FID of each dataset are marked in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 156, + 294, + 260 + ], + "lines": [ + { + "bbox": [ + 148, + 156, + 294, + 260 + ], + "spans": [ + { + "bbox": [ + 148, + 156, + 294, + 260 + ], + "type": "table", + "html": "
λ1λ2FID
0038.56(baseline)
1025.15
0126.09
5125.20
10125.17
15126.30
5525.55
50.523.3
", + "image_path": "47df8aed4b9651d83105f94782aebf89371cb51b556ad2cdf4fb4231f600e24c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 178, + 262, + 264, + 270 + ], + "lines": [ + { + "bbox": [ + 178, + 262, + 264, + 270 + ], + "spans": [ + { + "bbox": [ + 178, + 262, + 264, + 270 + ], + "type": "text", + "content": "(a) CIFAR-10 @ " + }, + { + "bbox": [ + 178, + 262, + 264, + 270 + ], + "type": "inline_equation", + "content": "32 \\times 32" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 323, + 157, + 463, + 260 + ], + "blocks": [ + { + "bbox": [ + 323, + 157, + 463, + 260 + ], + "lines": [ + { + "bbox": [ + 323, + 157, + 463, + 260 + ], + "spans": [ + { + "bbox": [ + 323, + 157, + 463, + 260 + ], + "type": "table", + "html": "
λ1λ2FID
003.86(baseline)
103.75
013.77
513.62
1013.84
1514.01
553.80
50.53.44
", + "image_path": "dc16f7387d436d2bcbdfcd4263dc74a4d4e672e3e99615dd312232f36eb9a153.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 353, + 262, + 434, + 270 + ], + "lines": [ + { + "bbox": [ + 353, + 262, + 434, + 270 + ], + "spans": [ + { + "bbox": [ + 353, + 262, + 434, + 270 + ], + "type": "text", + "content": "(b) FFHQ @ " + }, + { + "bbox": [ + 353, + 262, + 434, + 270 + ], + "type": "inline_equation", + "content": "256 \\times 256" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 307, + 482, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 482, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 482, + 378 + ], + "type": "text", + "content": "The FID scores are shown in Tab. 3. We can observe that when we set " + }, + { + "bbox": [ + 130, + 307, + 482, + 378 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 0" + }, + { + "bbox": [ + 130, + 307, + 482, + 378 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 307, + 482, + 378 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 0" + }, + { + "bbox": [ + 130, + 307, + 482, + 378 + ], + "type": "text", + "content": ", the performance of the network significantly drops " + }, + { + "bbox": [ + 130, + 307, + 482, + 378 + ], + "type": "inline_equation", + "content": "9.96\\%" + }, + { + "bbox": [ + 130, + 307, + 482, + 378 + ], + "type": "text", + "content": " on CIFAR-10 dataset. The performance even falls below the baseline model on high resolution dataset. Only when both the latent consistent loss and the reconstruction loss are combined together does the model gets its best performance. This proves that the two constrains we added are indispensable with each other." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 380, + 482, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 451 + ], + "type": "text", + "content": "Next we will find the optimal strength of the proposed constrains. We first fix " + }, + { + "bbox": [ + 130, + 380, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 1" + }, + { + "bbox": [ + 130, + 380, + 482, + 451 + ], + "type": "text", + "content": " and change the value of " + }, + { + "bbox": [ + 130, + 380, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 380, + 482, + 451 + ], + "type": "text", + "content": " to find the ideal value of " + }, + { + "bbox": [ + 130, + 380, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 380, + 482, + 451 + ], + "type": "text", + "content": ". During the experiments, we found that a relatively high " + }, + { + "bbox": [ + 130, + 380, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 380, + 482, + 451 + ], + "type": "text", + "content": " makes the model unable to converge, so we chose 15 as the maximum value. As shown in Tab. 3, " + }, + { + "bbox": [ + 130, + 380, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 5" + }, + { + "bbox": [ + 130, + 380, + 482, + 451 + ], + "type": "text", + "content": " turns out to be the best. Consequently, we then fix " + }, + { + "bbox": [ + 130, + 380, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 5" + }, + { + "bbox": [ + 130, + 380, + 482, + 451 + ], + "type": "text", + "content": " and search for the best " + }, + { + "bbox": [ + 130, + 380, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 380, + 482, + 451 + ], + "type": "text", + "content": ". Ultimately, " + }, + { + "bbox": [ + 130, + 380, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 5" + }, + { + "bbox": [ + 130, + 380, + 482, + 451 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 380, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 0.5" + }, + { + "bbox": [ + 130, + 380, + 482, + 451 + ], + "type": "text", + "content": " turns out to be the best strategy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 452, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 452, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 452, + 482, + 571 + ], + "type": "text", + "content": "During the search for optimal parameters, we observed some interesting phenomena. As shown in Fig. 3. Let " + }, + { + "bbox": [ + 130, + 452, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\eta = \\frac{\\lambda_1}{\\lambda_2}" + }, + { + "bbox": [ + 130, + 452, + 482, + 571 + ], + "type": "text", + "content": ". We found a higher value of " + }, + { + "bbox": [ + 130, + 452, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 452, + 482, + 571 + ], + "type": "text", + "content": " leads to a better result in the early stage of training. However, an excessively high " + }, + { + "bbox": [ + 130, + 452, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 452, + 482, + 571 + ], + "type": "text", + "content": " might lead to poor convergence at the final stages of model training. We infer that when " + }, + { + "bbox": [ + 130, + 452, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 452, + 482, + 571 + ], + "type": "text", + "content": " is large, our model that constrained by the latent distance, can quickly converge to a distribution similar to the real distribution, hence leads to an early convergence. But overly tight constraints might prevent the model's ability to generalize effectively, resulting in a lack of diversity in the generated data thus making the final performance poor. The underlying causes of this phenomenon require further investigation in future research." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 594, + 411, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 594, + 411, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 594, + 411, + 605 + ], + "type": "text", + "content": "4.4 Performance on Unsupervised Visual Recognition" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "GANs are often regarded as an unsupervised training method. To further demonstrate the effectiveness of the proposed method, we did experiment to show our proposed method leads to a better feature extractor. We take the pre-trained discriminator as a feature extractor and add a linear classification head to perform" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "type": "text", + "content": "S.Sun, Z.Luan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 117, + 479, + 288 + ], + "blocks": [ + { + "bbox": [ + 135, + 117, + 479, + 288 + ], + "lines": [ + { + "bbox": [ + 135, + 117, + 479, + 288 + ], + "spans": [ + { + "bbox": [ + 135, + 117, + 479, + 288 + ], + "type": "image", + "image_path": "ff7bc499da0340d06845c11206046bd0e7ff0ec508cc9fbcf511276636db84f2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 297, + 384, + 309 + ], + "lines": [ + { + "bbox": [ + 230, + 297, + 384, + 309 + ], + "spans": [ + { + "bbox": [ + 230, + 297, + 384, + 309 + ], + "type": "text", + "content": "Fig. 3: Training curves at different " + }, + { + "bbox": [ + 230, + 297, + 384, + 309 + ], + "type": "inline_equation", + "content": "\\eta" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 181, + 351, + 434, + 433 + ], + "blocks": [ + { + "bbox": [ + 132, + 319, + 482, + 341 + ], + "lines": [ + { + "bbox": [ + 132, + 319, + 482, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 482, + 341 + ], + "type": "text", + "content": "Table 4: Classification accuracy on CIFAR-10 dataset [25] with various methods with linear classifier." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 181, + 351, + 434, + 433 + ], + "lines": [ + { + "bbox": [ + 181, + 351, + 434, + 433 + ], + "spans": [ + { + "bbox": [ + 181, + 351, + 434, + 433 + ], + "type": "table", + "html": "
Method#ParametersAccuracy
ViT-H [10]632M99.5%
EfficientNetV2-M [44]55M99%
PyramidNet [28]26M98.6%
NAT-M4 [34]6.2M98.4%
HCGNet [52]3.1M97.7%
CLR-GAN(Ours)3.02M98.8%
", + "image_path": "d917e2c2fd06e8f1470aeff6e6c55dc5214357b87162a6c73c274debb57e37e3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 452, + 480, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 452, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 452, + 480, + 475 + ], + "type": "text", + "content": "classification task on the CIFAR-10 dataset. We compared our feature extractor with other classification models. The results are shown in Tab. 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 476, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 476, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 476, + 482, + 583 + ], + "type": "text", + "content": "As listed in Tab. 4. Our proposed pre-train feature extractor can reach a relatively high accuracy after simple fine-tuning. Compared with models which have similar parameter sizes (" + }, + { + "bbox": [ + 130, + 476, + 482, + 583 + ], + "type": "inline_equation", + "content": "\\leq 10M" + }, + { + "bbox": [ + 130, + 476, + 482, + 583 + ], + "type": "text", + "content": ") [34,52], our model is " + }, + { + "bbox": [ + 130, + 476, + 482, + 583 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 130, + 476, + 482, + 583 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 476, + 482, + 583 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 130, + 476, + 482, + 583 + ], + "type": "text", + "content": " higher respectively. In comparison with models several orders of magnitude higher than our parameter size [10], our model is also very competitive, only " + }, + { + "bbox": [ + 130, + 476, + 482, + 583 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 130, + 476, + 482, + 583 + ], + "type": "text", + "content": " lower than the best model. This indicates that our strategy can not only improve the quality and diversity of generated images in generative adversarial networks, but also serve as an effective unsupervised training method, benefiting downstream tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 599, + 274, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 274, + 612 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 274, + 612 + ], + "type": "text", + "content": "4.5 Towards a fairer game" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "As we mentioned in the introduction part, the fundamental purpose of our method is to produce more realistic and diverse images by making the training of GANs fairer. By the proposed consistent latent representation and reconstruction, we can shrink the distance between the generator and discriminator." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLR-GAN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 154, + 117, + 309, + 269 + ], + "blocks": [ + { + "bbox": [ + 154, + 117, + 309, + 269 + ], + "lines": [ + { + "bbox": [ + 154, + 117, + 309, + 269 + ], + "spans": [ + { + "bbox": [ + 154, + 117, + 309, + 269 + ], + "type": "image", + "image_path": "343da88656d16b9e3f248d949aee833c05182e74e0cc25813bda8a6775fb23fc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 280, + 482, + 314 + ], + "lines": [ + { + "bbox": [ + 130, + 280, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 482, + 314 + ], + "type": "text", + "content": "Fig. 4: Realness scores of the real samples and synthesis samples by various discriminators during training. We visualize the realness score of StyleGAN-V2 [22], StyleGAN-V2-ADA [19] and the proposed CLR-GAN." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 118, + 460, + 267 + ], + "blocks": [ + { + "bbox": [ + 309, + 118, + 460, + 267 + ], + "lines": [ + { + "bbox": [ + 309, + 118, + 460, + 267 + ], + "spans": [ + { + "bbox": [ + 309, + 118, + 460, + 267 + ], + "type": "image", + "image_path": "d556ee9befe05df20e0b6b8f735f109d684cd2bb94aab8b46d53eb57893e16af.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 336, + 479, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 479, + 359 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 479, + 359 + ], + "type": "text", + "content": "In this section, we demonstrate the enhanced fairness of our method through experiments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 361, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 515 + ], + "type": "text", + "content": "In order to better quantitatively judge the fairness of our model, we followed the previous two works [2, 47], using the score of the discriminator to represent the realness of the image. We visualize the realness score the model trained on LSUN-Church. We used exponentially weighted averages [17] on the original data for better visualization. The left part of Fig. 4 shows the score of the synthesis images, while the right part shows the score of the real images. We can clearly see that using our CLR-GAN, the realness gap between the generated image and the real image is much smaller than the baseline. While for StyleGAN-V2-ADA, the gap between scores is similar to the baseline, so it only improves FID and can not make the game fairer. We can then draw a conclusion that with the aid of the proposed auxiliary objective functions, the realness scores of real images and generated images become closer, which means that CLR-GAN can improve the realness and diversity of generated images by improving training fairness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 533, + 220, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 220, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 220, + 547 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "In this paper, we proposed a simple, effective and plug and play new objective function for GANs training. The new objective function simultaneously constrains the output distribution of the generator and the discriminator to make them more consistent. Thereby making the competition between the generator and the discriminator more fair. Eventually improving the quality and diversity of generated images. In addition to image generator, we also demonstrated that the discriminator with constrain is an efficient unsupervised feature extractor. Experimental results demonstrate that our proposed method improves the performance of baseline models over various datasets." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 249, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 249, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 249, + 101 + ], + "type": "text", + "content": "S.Sun, Z.Luan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "type": "text", + "content": "1. Arjovsky, M., Chintala, S., Bottou, L.: Wasserstein generative adversarial networks. In: International conference on machine learning. pp. 214-223. PMLR (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 166, + 481, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 166, + 481, + 198 + ], + "spans": [ + { + "bbox": [ + 138, + 166, + 481, + 198 + ], + "type": "text", + "content": "2. Bai, Q., Yang, C., Xu, Y., Liu, X., Yang, Y., Shen, Y.: Glead: Improving gans with a generator-leading task. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12094-12104 (June 2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 199, + 480, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 199, + 480, + 242 + ], + "spans": [ + { + "bbox": [ + 138, + 199, + 480, + 242 + ], + "type": "text", + "content": "3. Bar-Tal, O., Ofri-Amar, D., Fridman, R., Kasten, Y., Dekel, T.: Text2live: Text-driven layered image and video editing. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 707-723. Springer Nature Switzerland, Cham (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 243, + 480, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 480, + 265 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 480, + 265 + ], + "type": "text", + "content": "4. Brock, A., Donahue, J., Simonyan, K.: Large scale gan training for high fidelity natural image synthesis (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 266, + 480, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 480, + 298 + ], + "type": "text", + "content": "5. Ceylan, D., Huang, C.H.P., Mitra, N.J.: Pix2video: Video editing using image diffusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 23206-23217 (October 2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 299, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 342 + ], + "type": "text", + "content": "6. Chan, E.R., Monteiro, M., Kellnhofer, P., Wu, J., Wetzstein, G.: pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5799-5809 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 343, + 480, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 398 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 398 + ], + "type": "text", + "content": "7. Chenthamarakshan, V., Das, P., Hoffman, S., Strobelt, H., Padhi, I., Lim, K.W., Hoover, B., Manica, M., Born, J., Laino, T., Mojsilovic, A.: Cogmol: Target-specific and selective drug design for Covid-19 using deep generative models. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 4320-4332. Curran Associates, Inc. (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 399, + 480, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 480, + 431 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 480, + 431 + ], + "type": "text", + "content": "8. Choi, Y., Uh, Y., Yoo, J., Ha, J.W.: Stargan v2: Diverse image synthesis for multiple domains. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8188-8197 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 433, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 433, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 433, + 480, + 464 + ], + "type": "text", + "content": "9. Deng, Y., Tang, F., Dong, W., Ma, C., Pan, X., Wang, L., Xu, C.: Stytr2: Image style transfer with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11326-11336 (June 2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 466, + 480, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 466, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 466, + 480, + 498 + ], + "type": "text", + "content": "0. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 498, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 480, + 542 + ], + "type": "text", + "content": "1. Farnia, F., Ozdaglar, A.: Do GANs always have Nash equilibria? In: III, H.D., Singh, A. (eds.) Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 3029-3039. PMLR (13-18 Jul 2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 544, + 480, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 565 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 565 + ], + "type": "text", + "content": "2. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NeurIPS. vol. 27 (2014)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 566, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 566, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 138, + 566, + 480, + 598 + ], + "type": "text", + "content": "3. Gu, J., Shen, Y., Zhou, B.: Image processing using multi-code gan prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "type": "text", + "content": "4. Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A.C.: Improved training of wasserstein gans. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "5. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLR-GAN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "16. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 6840-6851. Curran Associates, Inc. (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "type": "text", + "content": "17. Hunter, J.S.: The exponentially weighted moving average. Journal of Quality Technology 18(4), 203-210 (1986)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 214 + ], + "type": "text", + "content": "18. Jiang, L., Dai, B., Wu, W., Loy, C.C.: Deceive d: Adaptive pseudo augmentation for gan training with limited data. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems (NeurIPS). vol. 34, pp. 21655-21667. Curran Associates, Inc. (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 480, + 236 + ], + "type": "text", + "content": "19. Karras, T., Aittala, M., Hellsten, J., Laine, S., Lehtinen, J., Aila, T.: Training generative adversarial networks with limited data (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 278 + ], + "type": "text", + "content": "20. Karras, T., Aittala, M., Laine, S., Härkönen, E., Hellsten, J., Lehtinen, J., Aila, T.: Alias-free generative adversarial networks. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 852-863. Curran Associates, Inc. (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 278, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 311 + ], + "type": "text", + "content": "21. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 311, + 480, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 343 + ], + "type": "text", + "content": "22. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8110-8119 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 343, + 480, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 375 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 375 + ], + "type": "text", + "content": "23. Kim, H., Choi, Y., Kim, J., Yoo, S., Uh, Y.: Exploiting spatial dimensions of latent in gan for real-time image editing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 852-861 (June 2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 375, + 422, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 422, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 422, + 386 + ], + "type": "text", + "content": "24. Kingma, D.P., Welling, M.: Auto-encoding variational bayes (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "type": "text", + "content": "25. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 407, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 460 + ], + "type": "text", + "content": "26. Kurach, K., Lucic, M., Zhai, X., Michalski, M., Gelly, S.: A large-scale study on regularization and normalization in GANs. In: Chaudhuri, K., Salakhutdinov, R. (eds.) Proceedings of the 36th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 97, pp. 3581-3590. PMLR (09-15 Jun 2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 460, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 480, + 493 + ], + "type": "text", + "content": "27. Kwon, G., Ye, J.C.: Clipstyler: Image style transfer with a single text condition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18062-18071 (June 2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 493, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 480, + 514 + ], + "type": "text", + "content": "28. Kwon, J., Kim, J., Park, H., Choi, I.K.: Asam: Adaptive sharpness-aware minimization for scale-invariant learning of deep neural networks (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "type": "text", + "content": "29. Kynkänniemi, T., Karras, T., Laine, S., Lehtinen, J., Aila, T.: Improved precision and recall metric for assessing generative models. Advances in Neural Information Processing Systems 32 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 546, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 480, + 590 + ], + "type": "text", + "content": "30. Lee, G., Kim, H., Kim, J., Kim, S., Ha, J.W., Choi, Y.: Generator knows what discriminator should learn in unconditional gans. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 406-422. Springer Nature Switzerland, Cham (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "type": "text", + "content": "31. Liao, W., Hu, K., Yang, M.Y., Rosenhahn, B.: Text to image generation with semantic-spatial aware gan. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18187-18196 (June 2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "text", + "content": "32. Liu, H., Zhang, W., Li, B., Wu, H., He, N., Huang, Y., Li, Y., Ghanem, B., Zheng, Y.: Adaptivemix: Improving gan training via feature space shrinkage. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16219-16229 (June 2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 248, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 248, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 248, + 101 + ], + "type": "text", + "content": "S.Sun, Z.Luan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "33. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: Proceedings of the IEEE international conference on computer vision. pp. 3730-3738 (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "34. Lu, Z., Sreekumar, G., Goodman, E., Banzhaf, W., Deb, K., Boddeti, V.N.: Neural architecture transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(9), 2971-2989 (Sep 2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 225 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 225 + ], + "type": "text", + "content": "35. Luo, S., Guan, J., Ma, J., Peng, J.: A 3d generative model for structure-based drug design. In: Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems. vol. 34, pp. 6229-6239. Curran Associates, Inc. (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 225, + 482, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 482, + 257 + ], + "type": "text", + "content": "36. Mao, X., Li, Q., Xie, H., Lau, R.Y., Wang, Z., Paul Smolley, S.: Least squares generative adversarial networks. In: Proceedings of the IEEE international conference on computer vision. pp. 2794-2802 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 257, + 482, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 278 + ], + "type": "text", + "content": "37. Miyato, T., Kataoka, T., Koyama, M., Yoshida, Y.: Spectral normalization for generative adversarial networks (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 278, + 482, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 311 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 311 + ], + "type": "text", + "content": "38. Mukherjee, S., Asnani, H., Lin, E., Kannan, S.: Clustergan: Latent space clustering in generative adversarial networks. Proceedings of the AAAI Conference on Artificial Intelligence 33(01), 4610-4617 (Jul 2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 311, + 482, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 342 + ], + "type": "text", + "content": "39. Nguyen-Phuoc, T., Li, C., Theis, L., Richardt, C., Yang, Y.L.: Hologan: Unsupervised learning of 3d representations from natural images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7588-7597 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 342, + 482, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 374 + ], + "type": "text", + "content": "40. Radford, A., Metz, L., Chintala, S.: Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434 (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 374, + 482, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 407 + ], + "type": "text", + "content": "41. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision 115, 211-252 (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 407, + 482, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 449 + ], + "type": "text", + "content": "42. Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Gradcam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (Oct 2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 449, + 482, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 471 + ], + "type": "text", + "content": "43. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 471, + 482, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 514 + ], + "type": "text", + "content": "44. Tan, M., Le, Q.: Efficientnetv2: Smaller models and faster training. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 10096-10106. PMLR (18-24 Jul 2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 514, + 482, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 547 + ], + "type": "text", + "content": "45. Tran, N.T., Bui, T.A., Cheung, N.M.: Dist-gan: An improved gan using distance constraints. In: Proceedings of the European Conference on Computer Vision (ECCV) (September 2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 547, + 482, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 547, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 547, + 482, + 589 + ], + "type": "text", + "content": "46. Tseng, H.Y., Jiang, L., Liu, C., Yang, M.H., Yang, W.: Regularizing generative adversarial networks under limited data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7921-7931 (June 2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 589, + 482, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 589, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 589, + 482, + 622 + ], + "type": "text", + "content": "47. Wang, J., Yang, C., Xu, Y., Shen, Y., Li, H., Zhou, B.: Improving gan equilibrium by raising spatial awareness. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11285-11293 (June 2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 622, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 622, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 622, + 482, + 665 + ], + "type": "text", + "content": "48. Wu, J.Z., Ge, Y., Wang, X., Lei, S.W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7623-7633 (October 2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 402, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLR-GAN" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 336 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "49. Xiangli, Y., Deng, Y., Dai, B., Loy, C.C., Lin, D.: Real or not real, that is the question. arXiv preprint arXiv:2002.05512 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "type": "text", + "content": "50. Xu, Y., AlBahar, B., Huang, J.B.: Temporally consistent semantic video editing. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 357-374. Springer Nature Switzerland, Cham (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 215 + ], + "type": "text", + "content": "51. Yang, C., Shen, Y., Xu, Y., Zhao, D., Dai, B., Zhou, B.: Improving gans with a dynamic discriminator. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Neural Information Processing Systems. vol. 35, pp. 15093-15104. Curran Associates, Inc. (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "text", + "content": "52. Yang, C., An, Z., Zhu, H., Hu, X., Zhang, K., Xu, K., Li, C., Xu, Y.: Gated convolutional networks with hybrid connectivity for image classification. Proceedings of the AAAI Conference on Artificial Intelligence 34(07), 12581-12588 (Apr 2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "53. Yu, F., Seff, A., Zhang, Y., Song, S., Funkhouser, T., Xiao, J.: Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365 (2015)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "text", + "content": "54. Zhao, J., Mathieu, M., LeCun, Y.: Energy-based generative adversarial network. arXiv preprint arXiv:1609.03126 (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 336 + ], + "type": "text", + "content": "55. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV) (Oct 2017)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 249, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 249, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 249, + 101 + ], + "type": "text", + "content": "S.Sun, Z.Luan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CMD_ A Cross Mechanism Domain Adaptation Dataset for 3D Object Detection/55286f6e-e84b-4ee2-994b-347fc5f78848_content_list.json b/2024/CMD_ A Cross Mechanism Domain Adaptation Dataset for 3D Object Detection/55286f6e-e84b-4ee2-994b-347fc5f78848_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c897cc6c27d2a6c2d5ffd7c39792945d1ce4d896 --- /dev/null +++ b/2024/CMD_ A Cross Mechanism Domain Adaptation Dataset for 3D Object Detection/55286f6e-e84b-4ee2-994b-347fc5f78848_content_list.json @@ -0,0 +1,1779 @@ +[ + { + "type": "text", + "text": "CMD: A Cross Mechanism Domain Adaptation Dataset for 3D Object Detection", + "text_level": 1, + "bbox": [ + 230, + 140, + 774, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinhao Deng $^{1,2,*}$ , Wei Ye $^{1,2,*}$ , Hai Wu $^{1,2}$ , Xun Huang $^{1,2}$", + "bbox": [ + 267, + 210, + 732, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qiming Xia $^{1,2}$ , Xin Li $^{4}$ , Jin Fang $^{3}$ , Wei Li $^{3,2}$ ,", + "bbox": [ + 313, + 232, + 687, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chenglu Wen $^{1,2,\\text{图}}$ and Cheng Wang $^{1,2}$", + "bbox": [ + 349, + 255, + 653, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Fujian Key Laboratory of Sensing and Computing for Smart Cities, Xiamen University, Xiamen, P.R China", + "bbox": [ + 267, + 282, + 732, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, Xiamen, P.R. China.", + "bbox": [ + 238, + 311, + 759, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Inceptio, Shanghai, P.R China", + "bbox": [ + 392, + 339, + 609, + 352 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Section of Visual Computing and Interactive Media, Texas A&M University, Texas, USA", + "bbox": [ + 320, + 352, + 681, + 380 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Point cloud data, representing the precise 3D layout of the scene, quickly drives the research of 3D object detection. However, the challenge arises due to the rapid iteration of 3D sensors, which leads to significantly different distributions in point clouds. This, in turn, results in subpar performance of 3D cross-sensor object detection. This paper introduces a Cross Mechanism Dataset, named CMD, to support research tackling this challenge. CMD is the first domain adaptation dataset, comprehensively encompassing diverse mechanical sensors and various scenes for 3D object detection. In terms of sensors, CMD includes 32-beam LiDAR, 128-beam LiDAR, solid-state LiDAR, 4D millimeter-wave radar, and cameras, all of which are well-synchronized and calibrated. Regarding the scenes, CMD consists of 50 sequences collocated from different scenarios, ranging from campuses to highways. Furthermore, we validated the effectiveness of various domain adaptation methods in mitigating sensor-based domain differences. We also proposed a DIG method to reduce domain disparities from the perspectives of Density, Intensity, and Geometry, which effectively bridges the domain gap between different sensors. The experimental results on the CMD dataset show that our proposed DIG method outperforms the state-of-the-art techniques, demonstrating the effectiveness of our baseline method. The dataset and the corresponding code are available at https://github.com/im-djh/CMD.", + "bbox": [ + 259, + 406, + 743, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Dataset $\\cdot$ 3D Object Detection $\\cdot$ Domain Adaptation", + "bbox": [ + 261, + 709, + 700, + 723 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 743, + 375, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As a crucial component in robotics and autonomous driving systems, 3D object detection has garnered increasing attention from researchers. Due to the inherent", + "bbox": [ + 212, + 771, + 785, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution", + "bbox": [ + 230, + 808, + 372, + 821 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\text{忍}$ Corresponding authors: W. Li (liweimcc@gmail.com); C. Wen (clwen@xmu.edu.cn)", + "bbox": [ + 230, + 824, + 785, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "depth of information, point cloud data enjoys a unique advantage in the domain of 3D object detection. Notably, researchers have developed several exceptional 3D object detection datasets and benchmarks (e.g., KITTI [13], nuScenes [5], Waymo [30], ONCE [21], etc.). Leveraging these high-quality datasets, numerous outstanding 3D object detection approaches [3,9,18,29,39,41,49] emerged, significantly facilitating the 3D object detection research. However, these prevalent datasets typically comprise mechanical spinning LiDAR only w.r.t. 3D point cloud sensors. The widely equipped low-cost automotive-grade sensors, e.g. solid-state LiDAR, in mass-produced vehicles [27] [25] are heavily overlooked even missing in those datasets.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The truth of outdatedness of existing datasets in terms of sensors behind is that point cloud sensors are undergoing really rapid advancements. Commonly used sensors now include: (1) mechanical spinning LiDAR with low-beam (e.g., 32 beams) or high-beam (e.g., 128 beams); (2) solid-state LiDAR; and (3) the recently acclaimed 4D millimeter-wave radar. These sensors differ either in the number of beams they use or in their world modeling patterns. We refer to these sensor differences as various mechanisms. The data acquired by different mechanisms may exhibit domain disparities in terms of density, intensity, and geometry (see Fig. 1(a), (b) and (c)). As a result, 3D detectors trained with data from one sensor often incur substantial accuracy degradation when directly applied to another sensor (see \"Direct\" in Fig. 1(d), (e) and (f)). For example, when migrating from mechanical spinning LiDAR to a more inexpensive and compact solid-state LiDAR, the cost of annotating new sensor data to re-train detectors is prohibitive. Cross-mechanism domain adaptation is a more efficient yet promising solution with rigid demand to transfer not only detectors/models but also all related assets to new sensors.", + "bbox": [ + 212, + 301, + 787, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meanwhile, prevalent domain adaptation research [19, 26, 35, 36, 45, 46] finds it hard to quantitatively analyze the pivotal sensor factor. This challenge arises because they have been directed at cross-dataset settings (e.g. domain adaptation between KITTI, Waymo, and nuScenes), covering differences in sensors, geographic location, climate, and so on. Therefore, they have not shown optimal performance when applied to cross-mechanism domain adaptation problems (see Fig. 1(d)(e)(f)). The reason lies in the fact that there is a lack of datasets that contain comprehensive sensors to fully decouple the domain disparities caused by locations and sensors. Several inspiring datasets, encompassing various common sensors (e.g., PandaSet [44], KRadar [22], VoD [23], LiDAR-CS [11]), often fall short in providing comprehensive modality coverage. This limitation, in turn, restricts the design of methods to bridge the sensor-based domain gap.", + "bbox": [ + 212, + 547, + 787, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this problem, this paper introduces a Cross Mechanism domain adaptation Dataset (CMD) for 3D object detection. To the best of our knowledge, CMD is the first domain-adaptation dataset in 3D object detection that contains a comprehensive suite of sensors, including: (1) 32-beam low-resolution and 128-beam high-resolution mechanical spinning (MS) LiDAR, (2) automotive-grade solid-state LiDAR, (3) $4D$ millimeter-wave radar, and (4) camera. This combination provides the most extensive modality coverage. All sensors are time-", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Deng et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "synchronized with high accuracy under 1 ms, ensuring that different modalities can capture and model the same scene accurately, even with highly dynamic objects. Furthermore, the CMD comprises 50 sequences. Each sequence has a time span of 20 seconds, with each sensor capturing 10 frames per second, totaling 10,000 frames of data per sensor. In addition, we meticulously annotated 3D objects of 13 categories based on a multi-sensor collaborative annotation system subjected to multiple rounds of human inspection.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e86365b3bc77bba7497ae8073fc1771fd7918d9c880027d8034a9b8d71e20e0f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 232, + 276, + 568, + 411 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5c1e820d28cb03e37f134c29a2ac3688029ad1cb9f411692cd8d4b1fa8e65779.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 276, + 772, + 411 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1c36901429a9508c4a44db8570f96aab853dd38886b243a65eff8a07bfdf79e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 240, + 411, + 769, + 473 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/19648560fae7813fe06a6fcfcef2539e8c87848e05f1ab1b6d8157f71c6502ed.jpg", + "image_caption": [ + "Fig. 1: (a) The points number distribution of different sensors. (b) The intensity distribution of different sensors. (c) Point clouds of the same instance for different sensors. The results show that data acquired by different sensors exhibit substantial differences in point density, intensity, and geometry. Subfigure (d), (e), and (f) show the cross-mechanism detection results of different domain adaptation methods on our CMD. Our DIG outperforms the traditional ST3D by a large margin." + ], + "image_footnote": [], + "bbox": [ + 232, + 477, + 403, + 595 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1087e7ea9d5585bddc4069b58cac5c0feeb8839d2db066bde509b5b6d5110690.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 477, + 558, + 595 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/38126f93d0f81ca5d3b6f56dcf5da01c40fdb763952c287894302212856fc139.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 575, + 477, + 764, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Using the CMD, we investigate the cross-mechanism domain adaptation issues and summarize the gaps into three primary components: Density gap, Intensity gap, and Geometry gap. Subsequently, we introduce a simple yet effective baseline method, named DIG, to address each of these gaps systematically. (1) Density gap: We propose the Beam-Distance Down-sampling (BDS) that enforces a similar data pattern in terms of beam numbers and density distribution along different detection distances. (2) Intensity gap: We introduce the Box-Cox log Normalization (BCN) that initially transforms point clouds from", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "CMD", + "bbox": [ + 691, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "different modalities into near-normal distributions and then normalizes them by logarithms. (3) Geometry gap: We design the Geometry-Aware label Mixing (GAM) that mixes the pseudo-labels from the target domain to narrow down the geometry difference in training. By implementing these three components, DIG achieves optimal results in bridging the identified domain gaps. Our main contributions are as follows.", + "bbox": [ + 212, + 146, + 787, + 234 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present the Cross Mechanism Dataset (CMD), the first domain adaptation dataset that contains a comprehensive suite of LiDAR/4D Radar sensors. CMD is a key piece of the puzzle of cross-mechanism detection as existing public datasets can not adequately evaluate detection performance drops caused by cross-domain disparities due to limited sensor variety.", + "- We construct comprehensive 3D cross-sensor detection baselines by benchmarking the results of SOTA domain adaptation algorithms on CMD.", + "- We provide an in-depth analysis of the key factors leading to cross-sensor domain disparities, i.e. Density, Intensity, and Geometry gaps. Based on this, we propose a novel DIG method, which shows plausible performance on cross-sensor detection." + ], + "bbox": [ + 222, + 234, + 784, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 422, + 387, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Datasets for 3D Object Detection", + "text_level": 1, + "bbox": [ + 215, + 453, + 539, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With increasing research on 3D object detection [7, 8, 17, 40, 42, 43, 47], more datasets have emerged. Widely used datasets such as KITTI [13], Waymo [30], nuScenes [5], and ONCE [21] support conventional 3D object detection research and facilitate domain adaptation across different geographic locations. However, focusing solely on geographical domain differences is inadequate for addressing the domain adaptation problem in detection algorithms; sensor variability is also crucial. To address this, datasets like PandaSet [44], K-Radar [22], TJ4DRadSet [52], VoD [23], Lyft L5 [15], and Cirrus [37] offer diverse sensors, scan beams, locations, and weather conditions. Despite their contributions, these datasets are limited in sensor diversity within individual collections. In contrast, our CMD includes extensive image and point cloud data annotated for detection and tracking, encompassing five different sensor types, thus providing a comprehensive evaluation. Statistical comparisons with other 3D object detection datasets are shown in Table 2.", + "bbox": [ + 212, + 477, + 787, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Domain Adaption in 3D Object Detection", + "text_level": 1, + "bbox": [ + 215, + 710, + 607, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To mitigate the poor generalization of 3D object detectors to unknown data, recent methods have focused on adapting domain gaps [10, 14, 32-34, 48, 50, 51]. Among these methods, some attempt to introduce a teacher-student framework. ST3D [45] and $\\mathrm{ST3D + + }$ [46] design label assignment strategy and pseudo-labels denoising to enable adaptation between domains. MLC-Net [19] employs consistency for cross-domain transfer. Another statistics-based approach adapts distribution gaps, with SN [35] correcting car size distribution gaps. GBA [26]", + "bbox": [ + 212, + 734, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Deng et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "adapts label distribution gap via a Gradual Batch Alternation training strategy. And, SSDA3D [36] adapting point cloud distributions through Inter-domain Alignment module. Further methods address sensor gaps, with LiDAR Distillation [38] and DTS [16] proposing progressive and density-insensitive frameworks to mitigate beam-induce gaps from different sensors, and CL3D [24] using spatial geometry alignment to adapt geometric gaps between sensors.", + "bbox": [ + 212, + 146, + 782, + 236 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In general, current domain adaptation research methods still primarily focus on geographical information and sensor beam migration. Due to dataset limitations, domain adaptation methods for multiple types of sensors have yet to be seen. Therefore, it is necessary to propose a novel, well-annotated dataset including various types of sensors to advance this research field.", + "bbox": [ + 212, + 237, + 784, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Cross Mechanism Dataset", + "text_level": 1, + "bbox": [ + 214, + 337, + 501, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Sensor Setup", + "text_level": 1, + "bbox": [ + 214, + 371, + 372, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sensor specifications. Our CMD comprises data from seven sensors of five types: a 128-beams mechanical scanning (MS) LiDAR, a 32-beams MS LiDAR, a solid-state LiDAR, a 4D millimeter-wave radar, and three cameras. Detailed specifications for our sensors are shown in Table 1. The sensors in our dataset, each with unique beam configurations or modeling modes, are collectively referred to as mechanisms. As shown in Table 2, in contrast to other datasets, our CMD encompasses a diverse array of commonly employed sensors, establishing it as the most comprehensive dataset presently available.", + "bbox": [ + 212, + 397, + 784, + 518 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/bf0258756fce89372fd28d81e17bdeacc780d3e2e1b2c0fe3f09b5b2087bfb24.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SensorTypeHFOV(°)VFOV(°)ResolutionFPS
OS128128 beams MS LiDAR360[-22.5, 22.5]128 * 102410
XT3232 beams MS LiDAR360[-16, 15]32 * 204810
M1SS LiDAR[-60, 60][-12.5, 12.5]126 * 62510
Radar4D Radar[-75, 75][-15, 15]>60010
HIK Cam.Camera[-31.2, 31.2][-27.7, 27.7]1080 * 192010
", + "bbox": [ + 243, + 545, + 759, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: Sensor specifications for CMD. MS and SS denote mechanical spinning and solid-state respectively. FPS refers to frames per second.", + "bbox": [ + 214, + 632, + 782, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sensor layout. In our CMD, all sensors are oriented in the same direction as the vehicle's forward direction. Such a configuration is instrumental in ensuring that the most critical information is captured within the area of greatest overlap of FOV across different sensors. To achieve this alignment, a rigid bracket has been designed, aligning all sensors along a vertically aligned axis. We have deliberately placed all sensors as close to each other as possible to optimize FOV overlap. In addition to this, a camera system comprising three evenly spaced cameras has been integrated. This system spans a $150^{\\circ}$ horizontal field of view (HFOV), as depicted in Fig. 2(c). Such a configuration is pivotal", + "bbox": [ + 212, + 704, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "CMD", + "bbox": [ + 691, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for ensuring data consistency across different mechanisms. The overall FOV is shown in Fig. 2(b).", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b9e908acc9ce7a128336dd9367be8276b438db5a1b413d096b9fe2265c97f61b.jpg", + "image_caption": [ + "Fig. 2: The illustration of sensor layout. (a) demonstrates the specific layout of each sensor; (b) showcases the field of view angles for each sensor; (c) demonstrates the positions of each sensor along the z-axis." + ], + "image_footnote": [], + "bbox": [ + 217, + 205, + 336, + 306 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bbabdac9b0b734d763543ec6b3060bea56092ed4a1bb913dad5f0f8ddcb62e28.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 205, + 437, + 306 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dc3843b9a3bd9db89f83b0e7b647f6b5610299d6b5647aa2d7f8816bf10e650b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 205, + 491, + 305 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6122c5a37371835274ea7fe115519b4181b96137aa84d7212366b417a0e500e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 207, + 630, + 305 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/42a273595bb090827ed95f7186f5bcbcc47e244ab382add5dae2697ef8397307.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 205, + 781, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Synchronization and Calibration", + "text_level": 1, + "bbox": [ + 214, + 417, + 529, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Synchronization and trigger. To achieve synchronization among different sensors when modeling the world, we utilized the Precision Time Protocol Version 2 [1] to establish a time synchronization system with an error margin within $1\\mathrm{ms}$ . Specifically, we employed the CoolShark AUTO 66 unit as the PTP master clock, while other sensors and the host served as PTP slave clocks. Once synchronized, M1 is set to generate a frame every $100\\mathrm{ms}$ , and both OS128 and XT32 generate frames simultaneously using phase-locking. Meanwhile, the radar is triggered by a network signal sent from the host. Due to the differing principles of data acquisition between cameras and LiDAR, the cameras are triggered by OS128. Whenever OS128 scans to $0^{\\circ}$ , it sends out a trigger signal. This signal undergoes a delay process via a microcontroller before being fed into the cameras. For each camera, $T_{delay} = \\frac{\\theta_{cam}}{2\\pi} - \\frac{1}{2} T_{exposure}$ where $T_{exposure}$ means the exposure time for camera and $\\theta_{cam}$ means the angle at which the camera is positioned. By this, the time when OS128 scans the central of the camera's HFOV, is the central of its exposure period, which ensures a lower frame time offset between the cameras and the LiDARs. This low-latency synchronization ensures that the same frame of data captured by different sensors effectively reflects their respective mechanism differences.", + "bbox": [ + 212, + 446, + 787, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Calibration. Achieving high-quality data in a multi-sensor setup relies heavily on calibration. (1) For the LiDARs (i.e. OS128, M1, XT32) and radar, we calibrate their extrinsic parameters using Generalized-ICP [28]. Several corner reflectors in the space are used to obtain more precise corner points for radar calibration. (2) For the cameras, we calibrate their intrinsic parameters and extrinsic parameters with MATLAB Toolkit [12] and OpenCV [4]. The results of calibration are shown in Fig. 4. Subsequent to calibration, we established a well-defined coordinate system, as illustrated in Fig. 2(c). We transfer XT32 and", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Deng et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "radar to the coordinate system of OS128 and move this coordinate system vertically down to the ground to get the ego system. The ego system has the x-axis pointing forward, the y-axis to the left, and the z-axis upward. The camera coordinate system, in 2D, aligns the x-axis with image width and y-axis with height, starting from the top-left corner.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Annotation", + "text_level": 1, + "bbox": [ + 214, + 247, + 356, + 260 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Annotation area. With precise time synchronization and calibration, we can jointly annotate data from multiple sensors, improving accuracy, especially for distant objects. We annotate objects within $160m$ and within the HFOV of M1. We believe that the data within this range is highly representative.", + "bbox": [ + 212, + 273, + 782, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Annotation rules. There are 13 annotated categories, namely: Car, Van, Bus, Truck, Semi-Trailer towing vehicle, Special Vehicles, Motorcycle, Bicycle, Tricycle, Adult Pedestrian, Children Pedestrian, Animal, and Barrier. We labeled each object as a 9-DoF 3D bounding box $(x,y,z,l,w,h,\\theta_x,\\theta_y,\\theta_z)$ . Where $x,y,z$ represents the center coordinates, $l,w,h$ denotes length, width height, and $\\theta_x,\\theta_y,\\theta_z$ are the rotation angles around the x,y,z axis. Additionally, we provide a motion state and tracking ID for all objects. Occlusion situations are also provided. More annotation specifications are provided in the appendix.", + "bbox": [ + 212, + 334, + 784, + 455 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e6e404ab41c84208c456b3be7461a64186a7dd167e1b80e593573fc15f6e84e7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMechanismMS LiDAR(beams)SS LiDAR 4DRadar Cam.Anno.Frames TrackingIlluminationReal
LowHigh
KITTI Det. [13]X64XX✓15KX✓✓
Waymo [30]X64XX✓230K✓✓✓
nuScenes [5]Single32XXX✓40K✓✓✓
Argoverse [6]32*2XXX✓44K✓✓✓
Lyft L5 [15]40*264XX✓30K✓X✓
K-Radar [22]X64, 128X✓✓35K✓✓✓
Lidar-CS [11]16, 3264, 128✓XX14KXXX
PandaSet [44]MultiX64✓X✓8.2K✓✓✓
TJ4DRadSet [52]X64X✓✓7.8K✓✓✓
VoD [23]X64X✓✓8.7KXX✓
CMD32128✓✓✓10K✓✓✓
", + "bbox": [ + 218, + 481, + 782, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2: Comparison with existing datasets. \"Mechanism\" means whether the dataset contains multiple mechanisms of 3D sensor. \"Tracking\" signifies the relevance of annotations for tracking. \"Illumination\" indicate the inclusion of various light conditions. \"Real\" refers to whether the dataset is real or synthetic.", + "bbox": [ + 212, + 614, + 782, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Dataset Analytic", + "text_level": 1, + "bbox": [ + 214, + 737, + 401, + 752 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Diversity of scenes. In CMD, scenes are carefully selected for wide coverage, primarily encompassing urban areas, suburbs, campuses, bridges, and tunnels with different illuminations (see Fig. 3(a)). We select 50 high-quality sequences, each spanning 20 seconds, equating to 200 frames per sensor, culminating in a dataset of 40,000 frames of point cloud data and 30,000 frames of image", + "bbox": [ + 212, + 763, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "CMD", + "bbox": [ + 691, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "data. This comprehensive collection offers a rich resource for research on cross-mechanism domain adaptation. These 50 sequences are evenly divided into 30, 10 and 10 for training, validation, and testing according to different scenarios. This arrangement ensures a balanced distribution of data across environments, crucial for training reliable and generalizable 3D object detection models.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6f5eb51ed399a4c404f634a84090c5285db2626f35e942401b3654d923c4b842.jpg", + "image_caption": [ + "Cross Mechanism domain adaptation Dataset (CMD) for 3D Object Detection" + ], + "image_footnote": [], + "bbox": [ + 220, + 276, + 575, + 473 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9b1c6c08041cd325ed272d6799b9d9874b8cfdbc9cadbf18bb017372efc3839e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 267, + 785, + 366 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/21ce536fbb4f5e2d3e9da15e4cad9723fdab2b134f07aee08105e37988666d36.jpg", + "image_caption": [ + "Fig. 3: The scene and annotation details. (a) The upper section showcases the distribution of road types, while the lower section demonstrates the distribution across various illumination types; (b) presents the distribution of objects' numbers for different categories; (c) presents the distribution of objects' numbers along different distances." + ], + "image_footnote": [], + "bbox": [ + 586, + 375, + 782, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Distribution of annotations. We annotated approximately 230,000 3D bounding boxes based on the rules described in Section 3.3. The distribution of objects' numbers for different categories is presented in Fig. 3(b). The distribution of objects' numbers along different distances is shown in Fig. 3(c). Among these categories, the most concerned \"Car\" has the highest annotation count, reaching approximately 92,000 instances.", + "bbox": [ + 212, + 589, + 787, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparisons between mechanisms. The average number of points for these four sensors is quite different. From most to least, they are M1, OS128, XT32, and 4D radar. For a better understanding, we show the visualization results in Fig. 4. Point clouds from sensors with different mechanisms exhibit substantial differences. (1) OS128 and XT32 both adopt a mechanism spinning style, clearly modeling the rigid vehicles with minimal vertical distortion. The main difference between them is density. (2) Due to the use of a zigzag scan pattern, M1 tends to introduce a bit of horizontal displacement. (3) Point clouds from radar sensors are generally sparse and inaccurate, posing challenges in accurately distinguishing the geometric shapes of objects.", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Deng et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1bda9ae537febd463216a986dabd09069ba31d81a3a8735a05988594fd340c5e.jpg", + "image_caption": [ + "Fig. 4: Different mechanisms and their sample data. The left side of the image denotes the various scenes of data collection, while the bottom indicates the types of sensors involved." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 782, + 424 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 DIG Baseline Method", + "text_level": 1, + "bbox": [ + 215, + 503, + 444, + 517 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we propose DIG (Density-Intensity-Geometry), a streamlined and potent baseline for cross-mechanism domain adaptation. It consists of three key components: Box-Cox Intensity Log Normalization (BCN), Beam-Distance Down Sampling (BDS), and Geometry-Aware Label Mixing (GAM).", + "bbox": [ + 212, + 525, + 782, + 585 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Box-Cox intensity log Normalization (BCN). The distribution of intensity for point clouds may significantly differ across various mechanisms. See Fig. 1(b). Recent studies have either opted to discard [45] [38] intensity or straightforwardly normalized it to [0,1] [36]. However, it is observed that intensity plays a crucial role in detecting foreground objects across diverse domains. We contend that effectively incorporating intensity can lead to enhanced performance in domain adaptation. Motivated by this insight, we introduce the BCN to mitigate disparities in intensity. For both intensities (denoted as $i$ ) for source and target domain point clouds, we initially take the logarithm to address biases stemming from different sensor definitions of intensity ranges. Subsequently, we apply the Box-Cox transformation to enhance their normality and homoscedasticity, respectively. This process can be formulated as:", + "bbox": [ + 212, + 585, + 787, + 767 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\ni (\\lambda) = \\frac {\\log (i + 1) ^ {\\lambda} - 1}{\\lambda}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 772, + 785, + 804 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda$ refers to the box-cox transfer parameter [2]. It's different for source and target domains, and remains constant respectively. After being normalized to", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "CMD", + "bbox": [ + 692, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "[0,1], the transformed intensity information exhibits a good level of similarity between the source and target domains, thus narrowing the domain gap.", + "bbox": [ + 215, + 146, + 782, + 176 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Beam-Distance based down Sampling (BDS). As shown in Fig. 1(a), the density distributions of points vary remarkably among different mechanisms. We observe that, within a relatively short distance, domain adaptation for different mechanisms is more significantly impacted by variations in point density. However, when it comes to longer distances, the point cloud is excessively sparse, which has a greater impact on object detection accuracy. To deal with this problem, we introduce the BDS that assigns different sampling probabilities based on both beam and distances. Denote $P_{\\text{keep}}(p)$ as the probability of a point (p) being retained, and it can be formulated as:", + "bbox": [ + 212, + 176, + 787, + 313 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\text {k e e p}} (p) = 1 - \\mathbf {1} \\left\\{\\text {b e a m} (p) \\notin \\mathcal {B} \\right\\} \\exp \\left(- \\frac {1}{K} | | p | |\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 319, + 785, + 348 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $K$ controls the rate of probability increase, and we set it to 35 empirically. $\\mathcal{B}$ means the set of target beam indexes that are closest to the source domain. Function $\\text{beam}(p)$ denotes the index of the beam that point $p$ belongs to. If this beam information is not directly available, an alternative approach involves calculating the angle of each point, denoted as $\\theta(p) = \\arctan \\left( \\frac{z}{\\sqrt{x^2 + y^2}} \\right)$ . This angle can then be used to apply a clustering algorithm, such as K-means [20], to approximate the beam index for each point.", + "bbox": [ + 212, + 353, + 787, + 465 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Geometry-Aware label Mixing (GAM). Point clouds captured from the same object by diverse sensors frequently exhibit notable geometric variances. To tackle this challenge, we introduce the Geometry-Aware Mix (GAM) module. In our approach, we initially utilize a model trained on the source domain to generate pseudo-labels within the target domain. Subsequently, instances from the target domain with high-confidence pseudo-labels are integrated back into the source domain for further training. Inspired by GBA [26], our second training phase progressively reduces source domain labels while maintaining a constant number of target domain pseudo-labels. This technique aims to gradually adapt the model to the geometric characteristics of the target domain. Through this method, the model becomes more adept at learning and adapting to the unique data characteristics of the target domain, thereby improving its ability for cross-mechanism domain adaptation.", + "bbox": [ + 212, + 465, + 787, + 662 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experimental Result", + "text_level": 1, + "bbox": [ + 215, + 683, + 452, + 700 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experiment Setup", + "text_level": 1, + "bbox": [ + 215, + 710, + 413, + 728 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "All of our experiments were conducted using four Nvidia 3090 GPUs and the open-source code repository OpenPCDet [31]. The Voxel-RCNN [9] served as the detector for most experiments. The batch size was set to 32, and the learning rate was 0.003. Similar to KITTI [13], the grid range was defined as $[0\\mathrm{m},70.4\\mathrm{m}]$ , $[-70.4\\mathrm{m},70.4\\mathrm{m}]$ , and $[-2\\mathrm{m},4\\mathrm{m}]$ along the x, y, and z axis respectively. Point cloud data were cropped to HFOV $[-60^{\\circ},60^{\\circ}]$ with horizontal distances smaller than $70.4\\mathrm{m}$ . Voxel size was set to $[0.1,0.1,0.15]$ along the x, y, and z axis respectively.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Deng et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/01118e19813a7984f94f896e43790c8b47e4e5b661ca8a7eaa6bd77d437286f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 146, + 475, + 233 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/07652d9d90f9f1cbf22863e6755a61f842a8a432a61681ded5a8b8e05cce1f32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 146, + 727, + 232 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ea1fd491f024a4bc8e5c89d75d71b4f7bf345ce1262e684bb71b903768ab0c94.jpg", + "image_caption": [ + "Fig. 5: Similarities and differences among baselines. ST3D [45] leverage random object scaling, generating and denoising pseudo-labels. DTS [16] down sample point clouds before training source-detector. Our DIG jointly downs ample point clouds, transforms intensity, and makes use of pseudo-labels." + ], + "image_footnote": [ + "$\\pmb{X}^{S}$ Points in source domain", + "$\\mathcal{F}(\\theta^s)$ Model in source domain", + "$X^t$ Points in target domain", + "$\\mathcal{F}(\\theta^t)$ Model in target domain" + ], + "bbox": [ + 276, + 234, + 473, + 323 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d56d7b09aeb8f7155d7fb0927d10e3d0468f3010a2422009f71b0f78d8b966f5.jpg", + "image_caption": [], + "image_footnote": [ + "$\\varphi^a$ Points down sampling", + "$\\varphi^i$ Intensity normalization" + ], + "bbox": [ + 480, + 234, + 725, + 321 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Evaluation Metric", + "text_level": 1, + "bbox": [ + 215, + 445, + 413, + 459 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Mean Average Precision (mAP). mAP is a commonly used object detection evaluation metric. Here when calculating mAP, we use 3D IoU thresholds of 0.5, 0.5, 0.25, and 0.25 for Car, Truck, Pedestrian, and Cyclist, respectively. Following the ONCE [21] dataset, we calculated an orientation-aware AP on 50 precision steps. The AP among different detection distances (e.g., overall and 0 to $30\\mathrm{m}$ ) are also computed for detailed analysis. Formally, the AP is defined as:", + "bbox": [ + 212, + 465, + 787, + 556 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nA P = 1 0 0 \\int_ {0} ^ {1} \\max \\left\\{p \\left(r ^ {\\prime} \\mid r ^ {\\prime} \\geq r\\right) \\right\\} d r, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 564, + 785, + 599 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $p(r)$ refers to the precision-recall curve that calculated by 50 recall positions. The mAP is the average of AP across all categories.", + "bbox": [ + 212, + 604, + 785, + 635 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Mean Closed Gap (mCG). Directly applying AP or mAP as evaluation metrics for domain adaptation is not intuitive since different sensors do not share the same domain gap. Closed gap [35] is used for measuring the effectiveness of a method on a single domain adaptation task. CG can be formulated as:", + "bbox": [ + 212, + 633, + 787, + 695 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nC G = 1 0 0 \\frac {m A P _ {\\text {m o d e l}} - m A P _ {D T}}{m A P _ {\\text {O r a c l e}} - m A P _ {D T}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 379, + 700, + 785, + 733 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $DT$ means direct transfer using a model trained sorely on source domain data. Oracle implies results obtained by the full training on target domain. To quantify the universality of domain adaptation methods, we define a new metric named mean closed gap (mCG), formulated as:", + "bbox": [ + 212, + 736, + 787, + 797 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nm C G = \\sum_ {s \\in \\mathbf {M}} \\sum_ {t \\in \\mathbf {M}} \\mathbf {1} (s \\neq t) C G _ {s \\rightarrow t}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 379, + 804, + 785, + 837 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "CMD", + "bbox": [ + 692, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $\\mathbf{M}$ refers to the set of sensors, including OS128, XT32, and M1 in CMD.", + "bbox": [ + 215, + 146, + 782, + 161 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/1061a7cec3ce8022d52c2ebb3eef3995bda8ca4c2b76570a3ce19d5727014677.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodmCGCG
M1→OS128XT32→OS128OS128→M1XT32→M1OS128→XT32M1→XT32
ST3D [45]-16.72-9.2734.00-7.95-5.07-43.39-68.61
ST3D++ [46]09.2312.6554.4902.9917.2307.56-39.54
DTS [16]29.2248.2822.9209.0207.0934.5053.51
DIG(Ours)42.8954.4155.0933.6925.2249.8739.06
", + "bbox": [ + 220, + 186, + 785, + 263 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 3: Overall mean closed gap for baseline methods.", + "bbox": [ + 310, + 263, + 687, + 275 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Discussion. Currently, commonly used CG is tailored for a single task, usually only considering the \"Car\" category [16], and cannot reflect the comprehensive performance of methods across various domain adaptation tasks. In contrast, mCG includes all MS LiDARs with varying beam numbers and SS LiDARs (i.e., OS128, XT32, M1), measuring performance from six experimental groups, thus providing the most comprehensive evaluation of cross-mechanism domain adaptation. See the appendix for 4D radar results.", + "bbox": [ + 212, + 319, + 784, + 426 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Results and Benchmark", + "text_level": 1, + "bbox": [ + 215, + 450, + 459, + 464 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Domain adaptation baselines. We selected four baseline methods. Fig. 5 illustrates the similarities and differences between them. (1) Direct transfer (DT) means directly applying the source model to the target dataset. (2) ST3D [45] and $\\mathrm{ST3D + + }$ [46] share a similar teacher-student architecture. The source model is trained with random object scaling, then used to generate pseudo-labels. The difference is that $\\mathrm{ST3D + + }$ comes with a more effective denoising method to obtain more high-quality pseudo-labels. (3) DTS [16] introduced Random Beam Re-Sampling (RBRS) to enhance the robustness to varying beam densities. We reproduced the RBRS on CMD, providing representative baseline results. (4) Our proposed DIG.", + "bbox": [ + 212, + 476, + 784, + 626 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Overall comparison results. Table 3 shows the overall results. Since the ST3D and $\\mathrm{ST3D + + }$ are mostly designed for cross-geographical domain adaptation that highly relies on pseudo-label quality, the two detectors show lower performance in our cross-mechanism dataset. DTS takes density into consideration, thus outperforming $\\mathrm{ST3D + + }$ by 19.99. Our DIG outperformed all previous methods by a large margin. Its mCG reached 42.89, surpassing ST3D, $\\mathrm{ST3D + + }$ , and DTS by 59.61, 33.66, and 13.67 respectively, as our method considers all the typical domain disparities in the cross-mechanism domain adaptation problem.", + "bbox": [ + 212, + 628, + 784, + 748 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Performance discussion. A more detailed composition of mCG is shown in Table 4. We can get several conclusions from these results. (1) DIG exhibits notable superiority in the categories of pedestrians and bicycles, both across all distances and within a detection range of 30 meters. The similarity in these categories is the fewer points within the targets. We believe that sparse points make intensity the key factor for adapting to different domains. DIG's BCN module", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Deng et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/506c1c7d7b0555784d71efc2c3dbe85aa02d8f4ca8b41154aeab979115d59009.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskMethodCGmAPCarTruckPedCyc
OS128Oracle10026.4436.50/69.3417.45/38.5118.79/38.9933.0/69.22
XT32↓OS128DT006.2911.76/20.5904.61/08.2301.54/02.1507.25/08.90
ST3D [45]34.0013.1422.78/47.3910.41/15.1006.83/16.0112.55/33.05
ST3D++ [46]54.4917.2727.87/61.0213.50/22.6108.57/19.7618.95/49.32
DTS [16]22.9210.9114.26/49.4306.72/27.6709.29/20.9413.37/41.21
DIG(Ours)55.0917.3922.73/60.5913.60/30.1611.56/25.6721.70/57.46
M1↓OS128DT006.9213.23/30.2105.08/07.8601.98/03.6107.38/09.98
ST3D [45]-9.2705.1110.24/39.1104.06/11.3200.00/00.0006.14/21.18
ST3D++ [46]12.5609.3913.46/46.9306.61/23.1605.85/13.6811.65/35.06
DTS [16]48.2516.3421.60/58.9912.67/34.1409.57/23.4921.51/58.70
DIG(Ours)54.4117.5423.03/57.0809.71/25.9509.71/25.9525.00/66.13
M1Oracle10030.0942.01/70.6419.96/40.0117.54/36.9940.84/68.49
XT32↓M1DT013.3220.33/42.8209.47/19.8404.81/11.4718.67/36.06
ST3D [45]-5.0712.4730.34/55.4104.20/03.8200.19/00.2215.13/25.95
ST3D++ [46]17.2316.2132.92/61.0705.24/06.0402.18/04.1724.48/49.39
DTS [16]07.0914.5123.26/56.9408.78/28.0206.99/14.7219.01/48.44
DIG(Ours)25.2217.5527.19/64.7009.26/22.7905.72/12.5728.04/57.34
OS128↓M1DT014.3627.07/56.5409.91/24.2501.13/03.8119.32/33.49
ST3D [45]-7.9513.1128.71/60.3708.26/07.2506.37/17.4809.08/12.72
ST3D++ [46]02.9914.8332.98/68.4506.63/08.4507.28/19.4612.43/17.89
DTS [16]09.0215.7822.11/54.0111.84/30.2611.88/28.5817.27/40.40
DIG(Ours)33.6919.6626.85/62.8911.86/28.1113.69/32.2426.19/60.40
XT32Oracle10026.5936.84/72.4719.26/41.0116.23/32.7434.01/66.59
M1↓XT32DT014.0722.10/59.8610.44/21.4806.29/15.2417.44/45.69
ST3D [45]-68.6105.4811.39/38.0502.92/06.7800.00/00.0002.92/06.78
ST3D++ [46]-39.5409.1214.56/47.9407.47/22.0703.15/05.0211.28/32.18
DTS [16]53.5120.7730.62/68.0315.67/33.0109.48/21.6327.33/57.48
DIG(Ours)39.0618.9625.95/68.7814.77/34.6511.45/26.4723.66/60.92
OS128↓XT32DT011.7722.25/58.2008.45/19.1601.66/04.5014.71/41.05
ST3D [45]-43.9905.3411.46/45.4101.72/05.8700.00/00.0008.18/22.82
ST3D++ [46]07.5612.8917.81/56.9905.31/19.9810.69/24.2217.75/48.29
DTS [16]34.5016.0918.87/59.7813.63/27.7011.26/25.4920.59/58.19
DIG(Ours)49.8719.1627.20/63.8910.75/26.7911.76/27.1626.91/64.22
", + "bbox": [ + 220, + 143, + 787, + 566 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 4: Detailed experimental results for all cross-mechanism settings. For each category, we provide the Average Precision (AP) results for all distances and within a range of 30 meters.", + "bbox": [ + 212, + 566, + 787, + 607 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "aligns intensity well, which explains its strong performance in these cases. (2) The results indicate a significant decrease in mAP while using OS128 as the target domain. This decrease is primarily attributed to its unique intensity range of [0, 512], unlike the [0, 256] range of other sensors. The pre-trained model finds it challenging to handle unfamiliar intensities. Unlike the other three methods, which either discarded the intensity information or did not include a dedicated module for it, we addressed this issue with BCN, resulting in a significant performance improvement. (3) All these methods perform unsatisfactorily for domain adaptation from XT32 to M1. Even DIG, the best one among them, gets CGs of 25.22 on it. This observation means that M1 and XT32 may suffer from larger domain gaps, highlighting the need for further investigation in the future. (4) ST3D++ gets a CG of 54.49 for XT32 to OS128, which is mainly due to the", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "CMD", + "bbox": [ + 692, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "great performance on APs for 0-30m. As they are both MS LiDAR, they suffer more from differences in density. Pseudo-labels from XT32 might look like harder cases for OS128. Refining model parameters with these cases can dramatically improve the performance. (5) It indicates that DTS plays a crucial role when there is a substantial difference in the number of beams. CG on tasks that contain XT32 gets nice performance improvement. Specifically, CG can reach 34.51 and 53.51 for OS128 to XT32 and M1 to XT32 respectively.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/134d9a9a96269de707f8e83e1c464dd99c1f72930ddff5d90995b50c0ef35244.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
modulesM1 ⇒ OS128OS128 ⇒ XT32
BDS BCN GAMmAPCGmAPCG
✓09.2211.7814.6919.70
✓17.1452.5316.4731.71
✓17.4754.0418.9548.45
✓17.5454.4119.1649.87
", + "bbox": [ + 338, + 275, + 666, + 362 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 5: Ablation study for DIG baseline method.", + "bbox": [ + 330, + 363, + 671, + 376 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ablation study for DIG. We conduct ablation studies to analyze the effectiveness of each component in the DIG baseline method. Two representative tasks are selected. $M1 \\Rightarrow OS128$ and $OS128 \\Rightarrow XT32$ respectively refer to domain adaptation for sensors with different scanning patterns and different beams/densities. As shown in Table 5, For both tasks, BCN is anticipated to be the most effective module, contributing 52.53 and 31.71 CG on each respective task. This is attributed to the substantial intensity differences between the source and target domains in both tasks. BDS also exhibits commendable performance in both tasks, achieving 11.78 and 19.70 CG, respectively. Although less conspicuous, GAM proves to be effective for both tasks.", + "bbox": [ + 212, + 417, + 787, + 568 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 585, + 359, + 602 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We presented a dataset specifically designed for cross-mechanism domain adaptation, incorporating mechanical LiDARs with both high-res and low-res beams, solid-state LiDAR, and 4D millimeter-wave radar. These sensors undergo precise time synchronization, allowing them to simultaneously generate point clouds of the same external environment. Data from diverse scenes were selected and meticulously annotated. We believe the proposed dataset would hugely facilitate research of cross-mechanism 3D detection, as it currently stands as the most comprehensive point cloud 3D detection dataset in terms of sensor types. We also provide a novel DIG method as well as complete experimental results of recent methods on our dataset. They can be used as reliable baselines to further benefit the research communities.", + "bbox": [ + 212, + 613, + 787, + 777 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In the future, we will continue to explore more usage of CMD, including utilizing data from multiple mechanisms as the source domain to enhance domain adaptation effects, and employing CMD as a tool for validating the performance of multi-modal 3D object detection.", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Deng et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. IEEE standard for a precision clock synchronization protocol for networked measurement and control systems. IEEE Std 1588-2008 (Revision of IEEE Std 1588-2002) pp. 1-269 (2008). https://doi.org/10.1109/IEEESTD.2008.4579760", + "2. Atkinson, A.C., Riani, M., Corbellini, A.: The box-cox transformation: Review and extensions (2021)", + "3. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "4. Bradski, G.: The opencv library. Dr. Dobb's Journal: Software Tools for the Professional Programmer 25(11), 120-123 (2000)", + "5. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (Jun 2020). https://doi.org/10.1109/cvpr42600.2020.01164, http://dx.doi.org/10.1109/cvpr42600.2020.01164", + "6. Chang, M.F., Lambert, J., Sangkloy, P., Singh, J., Bak, S., Hartnett, A., Wang, D., Carr, P., Lucey, S., Ramanan, D., et al.: Argoverse: 3d tracking and forecasting with rich maps. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8748-8757 (2019)", + "7. Chen, Y., Yu, Z., Chen, Y., Lan, S., Anandkumar, A., Jia, J., Alvarez, J.M.: Focalformer3d: Focusing on hard instance for 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8394-8405 (2023)", + "8. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21674-21683 (2023)", + "9. Deng, J., Shi, S., Li, P., gang Zhou, W., Zhang, Y., Li, H.: Voxel r-cnn: Towards high performance voxel-based 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence (2021)", + "10. Ding, G., Zhang, M., Li, E., Hao, Q.: Jst: Joint self-training for unsupervised domain adaptation on 2d&3d object detection. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 477-483. IEEE (2022)", + "1. Fang, J., Zhou, D., Zhao, J., Tang, C., Xu, C.Z., Zhang, L.: Lidar-cs dataset: Lidar point cloud dataset with cross-sensors for 3d object detection (Jan 2023)", + "2. Fetic, A., Juric, D., Osmankovic, D.: The procedure of a camera calibration using camera calibration toolbox for matlab. In: 2012 Proceedings of the 35th International Convention MIPRO. pp. 1752-1757. IEEE (2012)", + "3. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition (Jun 2012). https://doi.org/10.1109/cvpr.2012.6248074, http://dx.doi.org/10.1109/cvpr.2012.6248074", + "4. Hegde, D., Sindagi, V., Kilic, V., Cooper, A.B., Foster, M., Patel, V.: Uncertainty-aware mean teacher for source-free unsupervised domain adaptive 3d object detection. arXiv preprint arXiv:2109.14651 (2021)", + "5. Houston, J., Zuidhof, G., Bergamini, L., Ye, Y., Chen, L., Jain, A., Omari, S., Iglovikov, V., Ondruska, P.: One thousand and one hours: Self-driving motion prediction dataset. In: Conference on Robot Learning. pp. 409-418. PMLR (2021)" + ], + "bbox": [ + 225, + 176, + 784, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "CMD", + "bbox": [ + 691, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Hu, Q., Liu, D., Hu, W.: Density-insensitive unsupervised domain adaption on 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17556-17566 (2023)", + "17. Huang, X., Wu, H., Li, X., Fan, X., Wen, C., Wang, C.: Sunshine to rainstorm: Cross-weather knowledge distillation for robust 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 2409–2416 (2024)", + "18. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. ArXiv (2022)", + "19. Luo, Z., Cai, Z., Zhou, C., Zhang, G., Zhao, H., Yi, S., Lu, S., Li, H., Zhang, S., Liu, Z.: Unsupervised domain adaptive 3d detection with multi-level consistency. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8866-8875 (2021)", + "20. MacQueen, J., et al.: Some methods for classification and analysis of multivariate observations. In: Proceedings of the fifth Berkeley symposium on mathematical statistics and probability. vol. 1, pp. 281-297. Oakland, CA, USA (1967)", + "21. Mao, J., Niu, M., Jiang, C., Liang, H., Liang, X., Li, Y., Ye, C., Zhang, W., Li, Z., Yu, J., Xu, H., Xu, C.: One million scenes for autonomous driving: Once dataset. Cornell University - arXiv, Cornell University - arXiv (Jun 2021)", + "22. Paek, D.H., Kong, S.H., Wijaya, K.: K-radar: 4d radar object detection dataset and benchmark for autonomous driving in various weather conditions (Jun 2022)", + "23. Palffy, A., Pool, E., Baratam, S., Kooij, J., Gavrila, D.: Multi-class road user detection with 3+1d radar in the view-of-delft dataset", + "24. Peng, X., Zhu, X., Ma, Y.: Cl3d: Unsupervised domain adaptation for cross-lidar 3d detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2047-2055 (2023)", + "25. Raj, T., Hanim Hashim, F., Baseri Huddin, A., Ibrahim, M.F., Hussain, A.: A survey on lidar scanning mechanisms. Electronics 9(5), 741 (2020)", + "26. Rochan, M., Chen, X., Grandhi, A., Corral-Soto, E.R., Liu, B.: Domain adaptation in 3d object detection with gradual batch alternation training. arXiv preprint arXiv:2210.10180 (2022)", + "27. Roriz, R., Cabral, J., Gomes, T.: Automotive lidar technology: A survey. IEEE Transactions on Intelligent Transportation Systems 23(7), 6282-6297 (2021)", + "28. Segal, A., Haehnel, D., Thrun, S.: Generalized-icp. In: Robotics: science and systems. vol. 2, p. 435. Seattle, WA (2009)", + "29. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR). pp. 10526 - 10535 (2020)", + "30. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (Jun 2020). https://doi.org/10.1109/cvpr42600.2020.00252, http://dx.doi.org/10.1109/cvpr42600.2020.00252", + "31. Team, O.D.: Openpcdet: An open-source toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet (2020)", + "32. Tsai, D., Berrio, J.S., Shan, M., Nebot, E., Worrall, S.: Ms3d: Leveraging multiple detectors for unsupervised domain adaptation in 3d object detection. arXiv preprint arXiv:2304.02431 (2023)" + ], + "bbox": [ + 217, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Deng et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Tsai, D., Berrio, J.S., Shan, M., Nebot, E., Worrall, S.: Viewer-centred surface completion for unsupervised domain adaptation in 3d object detection. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 9346-9353. IEEE (2023)", + "34. Tsai, D., Berrio, J.S., Shan, M., Worrall, S., Nebot, E.: See eye to eye: A lidar-agnostic 3d detection framework for unsupervised multi-target domain adaptation. IEEE Robotics and Automation Letters 7(3), 7904-7911 (2022)", + "35. Wang, Y., Chen, X., You, Y., Li, L.E., Hariharan, B., Campbell, M., Weinberger, K.Q., Chao, W.L.: Train in germany, test in the usa: Making 3d object detectors generalize. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11713-11723 (2020)", + "36. Wang, Y., Yin, J., Li, W., Frossard, P., Yang, R., Shen, J.: Ssa3d: Semi-supervised domain adaptation for 3d object detection from point cloud. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2707-2715 (2023)", + "37. Wang, Z., Ding, S., Li, Y., Fenn, J., Roychowdhury, S., Wallin, A., Martin, L., Ryvola, S., Sapiro, G., Qiu, Q.: Cirrus: A long-range bi-pattern lidar dataset. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 5744-5750. IEEE (2021)", + "38. Wei, Y., Wei, Z., Rao, Y., Li, J., Zhou, J., Lu, J.: Lidar distillation: Bridging the beam-induced domain gap for 3d object detection. In: European Conference on Computer Vision. pp. 179-195. Springer (2022)", + "39. Wu, H., Deng, J., Wen, C., Li, X., Wang, C.: Casa: A cascade attention network for 3d object detection from lidar point clouds. IEEE Transactions on Geoscience and Remote Sensing (2022)", + "40. Wu, H., Wen, C., Li, W., Yang, R., Wang, C.: Learning transformation-equivariant features for 3d object detection (2022)", + "41. Wu, H., Wen, C., Shi, S., Li, X., Wang, C.: Virtual sparse convolution for multimodal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21653-21662 (2023)", + "42. Xia, Q., Chen, Y., Cai, G., Chen, G., Xie, D., Su, J., Wang, Z.: 3-d hanet: A flexible 3-d heatmap auxiliary network for object detection. IEEE Transactions on Geoscience and Remote Sensing 61, 1-13 (2023)", + "43. Xia, Q., Deng, J., Wen, C., Wu, H., Shi, S., Li, X., Wang, C.: Coin: Contrastive instance feature mining for outdoor 3d object detection with very limited annotations. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6254-6263 (2023)", + "44. Xiao, P., Shao, Z., Hao, S., Zhang, Z., Chai, X., Jiao, J., Li, Z., Wu, J., Sun, K., Jiang, K., Wang, Y., Yang, D.: Pandaset: Advanced sensor suite dataset for autonomous driving. In: 2021 IEEE International Intelligent Transportation Systems Conference (ITSC) (Sep 2021). https://doi.org/10.1109/itsc48978.2021.9565009, http://dx.doi.org/10.1109/itsc48978.2021.9565009", + "45. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d: Self-training for unsupervised domain adaptation on 3d object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10368-10378 (2021)", + "46. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d++: Denoised self-training for unsupervised domain adaptation on 3d object detection. IEEE transactions on pattern analysis and machine intelligence 45(5), 6354-6371 (2022)", + "47. Yang, Z., Sun, Y., Liu, S., Jia, J.: 3dssd: Point-based 3d single stage object detector. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR). pp. 11040-11048 (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "CMD", + "bbox": [ + 691, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Yihan, Z., Wang, C., Wang, Y., Xu, H., Ye, C., Yang, Z., Ma, C.: Learning transferable features for point cloud detection via 3d contrastive co-training. Advances in Neural Information Processing Systems 34, 21493-21504 (2021)", + "49. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR) (2021)", + "50. You, Y., Diaz-Ruiz, C.A., Wang, Y., Chao, W.L., Hariharan, B., Campbell, M., Weinbergert, K.Q.: Exploiting playbacks in unsupervised domain adaptation for 3d object detection in self-driving cars. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 5070-5077. IEEE (2022)", + "51. You, Y., Phoo, C.P., Luo, K., Zhang, T., Chao, W.L., Hariharan, B., Campbell, M., Weinberger, K.Q.: Unsupervised adaptation from repeated traversals for autonomous driving. Advances in Neural Information Processing Systems 35, 27716-27729 (2022)", + "52. Zheng, L., Ma, Z., Zhu, X., Tan, B., Li, S., Long, K., Sun, W., Chen, S., Zhang, L., Wan, M., Huang, L., Bai, J.: Tj4dradset: A 4d radar dataset for autonomous driving (Apr 2022)" + ], + "bbox": [ + 215, + 146, + 785, + 383 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "J. Deng et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/CMD_ A Cross Mechanism Domain Adaptation Dataset for 3D Object Detection/55286f6e-e84b-4ee2-994b-347fc5f78848_model.json b/2024/CMD_ A Cross Mechanism Domain Adaptation Dataset for 3D Object Detection/55286f6e-e84b-4ee2-994b-347fc5f78848_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8a8d6e8fd0d4116337e97f0e1e234c3b7df759e3 --- /dev/null +++ b/2024/CMD_ A Cross Mechanism Domain Adaptation Dataset for 3D Object Detection/55286f6e-e84b-4ee2-994b-347fc5f78848_model.json @@ -0,0 +1,2414 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.231, + 0.141, + 0.776, + 0.187 + ], + "angle": 0, + "content": "CMD: A Cross Mechanism Domain Adaptation Dataset for 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.212, + 0.733, + 0.228 + ], + "angle": 0, + "content": "Jinhao Deng\\(^{1,2,*}\\), Wei Ye\\(^{1,2,*}\\), Hai Wu\\(^{1,2}\\), Xun Huang\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.233, + 0.689, + 0.251 + ], + "angle": 0, + "content": "Qiming Xia\\(^{1,2}\\), Xin Li\\(^{4}\\), Jin Fang\\(^{3}\\), Wei Li\\(^{3,2}\\)," + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.256, + 0.654, + 0.273 + ], + "angle": 0, + "content": "Chenglu Wen\\(^{1,2,\\text{图}}\\) and Cheng Wang\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.284, + 0.733, + 0.311 + ], + "angle": 0, + "content": "\\(^{1}\\) Fujian Key Laboratory of Sensing and Computing for Smart Cities, Xiamen University, Xiamen, P.R China" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.312, + 0.761, + 0.34 + ], + "angle": 0, + "content": "\\(^{2}\\) Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, Xiamen, P.R. China." + }, + { + "type": "text", + "bbox": [ + 0.393, + 0.34, + 0.61, + 0.353 + ], + "angle": 0, + "content": "3 Inceptio, Shanghai, P.R China" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.353, + 0.682, + 0.381 + ], + "angle": 0, + "content": "\\(^{4}\\) Section of Visual Computing and Interactive Media, Texas A&M University, Texas, USA" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.407, + 0.744, + 0.699 + ], + "angle": 0, + "content": "Abstract. Point cloud data, representing the precise 3D layout of the scene, quickly drives the research of 3D object detection. However, the challenge arises due to the rapid iteration of 3D sensors, which leads to significantly different distributions in point clouds. This, in turn, results in subpar performance of 3D cross-sensor object detection. This paper introduces a Cross Mechanism Dataset, named CMD, to support research tackling this challenge. CMD is the first domain adaptation dataset, comprehensively encompassing diverse mechanical sensors and various scenes for 3D object detection. In terms of sensors, CMD includes 32-beam LiDAR, 128-beam LiDAR, solid-state LiDAR, 4D millimeter-wave radar, and cameras, all of which are well-synchronized and calibrated. Regarding the scenes, CMD consists of 50 sequences collocated from different scenarios, ranging from campuses to highways. Furthermore, we validated the effectiveness of various domain adaptation methods in mitigating sensor-based domain differences. We also proposed a DIG method to reduce domain disparities from the perspectives of Density, Intensity, and Geometry, which effectively bridges the domain gap between different sensors. The experimental results on the CMD dataset show that our proposed DIG method outperforms the state-of-the-art techniques, demonstrating the effectiveness of our baseline method. The dataset and the corresponding code are available at https://github.com/im-djh/CMD." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.71, + 0.702, + 0.724 + ], + "angle": 0, + "content": "Keywords: Dataset \\(\\cdot\\) 3D Object Detection \\(\\cdot\\) Domain Adaptation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.744, + 0.376, + 0.76 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.772, + 0.787, + 0.802 + ], + "angle": 0, + "content": "As a crucial component in robotics and autonomous driving systems, 3D object detection has garnered increasing attention from researchers. Due to the inherent" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.809, + 0.374, + 0.823 + ], + "angle": 0, + "content": "* Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.231, + 0.825, + 0.786, + 0.84 + ], + "angle": 0, + "content": "\\(\\text{忍}\\) Corresponding authors: W. Li (liweimcc@gmail.com); C. Wen (clwen@xmu.edu.cn)" + }, + { + "type": "list", + "bbox": [ + 0.231, + 0.809, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Deng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "depth of information, point cloud data enjoys a unique advantage in the domain of 3D object detection. Notably, researchers have developed several exceptional 3D object detection datasets and benchmarks (e.g., KITTI [13], nuScenes [5], Waymo [30], ONCE [21], etc.). Leveraging these high-quality datasets, numerous outstanding 3D object detection approaches [3,9,18,29,39,41,49] emerged, significantly facilitating the 3D object detection research. However, these prevalent datasets typically comprise mechanical spinning LiDAR only w.r.t. 3D point cloud sensors. The widely equipped low-cost automotive-grade sensors, e.g. solid-state LiDAR, in mass-produced vehicles [27] [25] are heavily overlooked even missing in those datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.303, + 0.788, + 0.543 + ], + "angle": 0, + "content": "The truth of outdatedness of existing datasets in terms of sensors behind is that point cloud sensors are undergoing really rapid advancements. Commonly used sensors now include: (1) mechanical spinning LiDAR with low-beam (e.g., 32 beams) or high-beam (e.g., 128 beams); (2) solid-state LiDAR; and (3) the recently acclaimed 4D millimeter-wave radar. These sensors differ either in the number of beams they use or in their world modeling patterns. We refer to these sensor differences as various mechanisms. The data acquired by different mechanisms may exhibit domain disparities in terms of density, intensity, and geometry (see Fig. 1(a), (b) and (c)). As a result, 3D detectors trained with data from one sensor often incur substantial accuracy degradation when directly applied to another sensor (see \"Direct\" in Fig. 1(d), (e) and (f)). For example, when migrating from mechanical spinning LiDAR to a more inexpensive and compact solid-state LiDAR, the cost of annotating new sensor data to re-train detectors is prohibitive. Cross-mechanism domain adaptation is a more efficient yet promising solution with rigid demand to transfer not only detectors/models but also all related assets to new sensors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.788, + 0.732 + ], + "angle": 0, + "content": "Meanwhile, prevalent domain adaptation research [19, 26, 35, 36, 45, 46] finds it hard to quantitatively analyze the pivotal sensor factor. This challenge arises because they have been directed at cross-dataset settings (e.g. domain adaptation between KITTI, Waymo, and nuScenes), covering differences in sensors, geographic location, climate, and so on. Therefore, they have not shown optimal performance when applied to cross-mechanism domain adaptation problems (see Fig. 1(d)(e)(f)). The reason lies in the fact that there is a lack of datasets that contain comprehensive sensors to fully decouple the domain disparities caused by locations and sensors. Several inspiring datasets, encompassing various common sensors (e.g., PandaSet [44], KRadar [22], VoD [23], LiDAR-CS [11]), often fall short in providing comprehensive modality coverage. This limitation, in turn, restricts the design of methods to bridge the sensor-based domain gap." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To address this problem, this paper introduces a Cross Mechanism domain adaptation Dataset (CMD) for 3D object detection. To the best of our knowledge, CMD is the first domain-adaptation dataset in 3D object detection that contains a comprehensive suite of sensors, including: (1) 32-beam low-resolution and 128-beam high-resolution mechanical spinning (MS) LiDAR, (2) automotive-grade solid-state LiDAR, (3) \\(4D\\) millimeter-wave radar, and (4) camera. This combination provides the most extensive modality coverage. All sensors are time-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.692, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CMD" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "synchronized with high accuracy under 1 ms, ensuring that different modalities can capture and model the same scene accurately, even with highly dynamic objects. Furthermore, the CMD comprises 50 sequences. Each sequence has a time span of 20 seconds, with each sensor capturing 10 frames per second, totaling 10,000 frames of data per sensor. In addition, we meticulously annotated 3D objects of 13 categories based on a multi-sensor collaborative annotation system subjected to multiple rounds of human inspection." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.277, + 0.57, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.573, + 0.277, + 0.773, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.412, + 0.77, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.478, + 0.404, + 0.597 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.478, + 0.56, + 0.597 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.478, + 0.765, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.607, + 0.788, + 0.692 + ], + "angle": 0, + "content": "Fig. 1: (a) The points number distribution of different sensors. (b) The intensity distribution of different sensors. (c) Point clouds of the same instance for different sensors. The results show that data acquired by different sensors exhibit substantial differences in point density, intensity, and geometry. Subfigure (d), (e), and (f) show the cross-mechanism detection results of different domain adaptation methods on our CMD. Our DIG outperforms the traditional ST3D by a large margin." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Using the CMD, we investigate the cross-mechanism domain adaptation issues and summarize the gaps into three primary components: Density gap, Intensity gap, and Geometry gap. Subsequently, we introduce a simple yet effective baseline method, named DIG, to address each of these gaps systematically. (1) Density gap: We propose the Beam-Distance Down-sampling (BDS) that enforces a similar data pattern in terms of beam numbers and density distribution along different detection distances. (2) Intensity gap: We introduce the Box-Cox log Normalization (BCN) that initially transforms point clouds from" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Deng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.235 + ], + "angle": 0, + "content": "different modalities into near-normal distributions and then normalizes them by logarithms. (3) Geometry gap: We design the Geometry-Aware label Mixing (GAM) that mixes the pseudo-labels from the target domain to narrow down the geometry difference in training. By implementing these three components, DIG achieves optimal results in bridging the identified domain gaps. Our main contributions are as follows." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.236, + 0.784, + 0.31 + ], + "angle": 0, + "content": "- We present the Cross Mechanism Dataset (CMD), the first domain adaptation dataset that contains a comprehensive suite of LiDAR/4D Radar sensors. CMD is a key piece of the puzzle of cross-mechanism detection as existing public datasets can not adequately evaluate detection performance drops caused by cross-domain disparities due to limited sensor variety." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.312, + 0.785, + 0.34 + ], + "angle": 0, + "content": "- We construct comprehensive 3D cross-sensor detection baselines by benchmarking the results of SOTA domain adaptation algorithms on CMD." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.342, + 0.785, + 0.4 + ], + "angle": 0, + "content": "- We provide an in-depth analysis of the key factors leading to cross-sensor domain disparities, i.e. Density, Intensity, and Geometry gaps. Based on this, we propose a novel DIG method, which shows plausible performance on cross-sensor detection." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.236, + 0.785, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.424, + 0.388, + 0.439 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.454, + 0.54, + 0.469 + ], + "angle": 0, + "content": "2.1 Datasets for 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.788, + 0.689 + ], + "angle": 0, + "content": "With increasing research on 3D object detection [7, 8, 17, 40, 42, 43, 47], more datasets have emerged. Widely used datasets such as KITTI [13], Waymo [30], nuScenes [5], and ONCE [21] support conventional 3D object detection research and facilitate domain adaptation across different geographic locations. However, focusing solely on geographical domain differences is inadequate for addressing the domain adaptation problem in detection algorithms; sensor variability is also crucial. To address this, datasets like PandaSet [44], K-Radar [22], TJ4DRadSet [52], VoD [23], Lyft L5 [15], and Cirrus [37] offer diverse sensors, scan beams, locations, and weather conditions. Despite their contributions, these datasets are limited in sensor diversity within individual collections. In contrast, our CMD includes extensive image and point cloud data annotated for detection and tracking, encompassing five different sensor types, thus providing a comprehensive evaluation. Statistical comparisons with other 3D object detection datasets are shown in Table 2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.711, + 0.609, + 0.726 + ], + "angle": 0, + "content": "2.2 Domain Adaption in 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.843 + ], + "angle": 0, + "content": "To mitigate the poor generalization of 3D object detectors to unknown data, recent methods have focused on adapting domain gaps [10, 14, 32-34, 48, 50, 51]. Among these methods, some attempt to introduce a teacher-student framework. ST3D [45] and \\(\\mathrm{ST3D + + }\\) [46] design label assignment strategy and pseudo-labels denoising to enable adaptation between domains. MLC-Net [19] employs consistency for cross-domain transfer. Another statistics-based approach adapts distribution gaps, with SN [35] correcting car size distribution gaps. GBA [26]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.692, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CMD" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.237 + ], + "angle": 0, + "content": "adapts label distribution gap via a Gradual Batch Alternation training strategy. And, SSDA3D [36] adapting point cloud distributions through Inter-domain Alignment module. Further methods address sensor gaps, with LiDAR Distillation [38] and DTS [16] proposing progressive and density-insensitive frameworks to mitigate beam-induce gaps from different sensors, and CL3D [24] using spatial geometry alignment to adapt geometric gaps between sensors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.785, + 0.313 + ], + "angle": 0, + "content": "In general, current domain adaptation research methods still primarily focus on geographical information and sensor beam migration. Due to dataset limitations, domain adaptation methods for multiple types of sensors have yet to be seen. Therefore, it is necessary to propose a novel, well-annotated dataset including various types of sensors to advance this research field." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.338, + 0.503, + 0.354 + ], + "angle": 0, + "content": "3 Cross Mechanism Dataset" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.372, + 0.373, + 0.387 + ], + "angle": 0, + "content": "3.1 Sensor Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.785, + 0.519 + ], + "angle": 0, + "content": "Sensor specifications. Our CMD comprises data from seven sensors of five types: a 128-beams mechanical scanning (MS) LiDAR, a 32-beams MS LiDAR, a solid-state LiDAR, a 4D millimeter-wave radar, and three cameras. Detailed specifications for our sensors are shown in Table 1. The sensors in our dataset, each with unique beam configurations or modeling modes, are collectively referred to as mechanisms. As shown in Table 2, in contrast to other datasets, our CMD encompasses a diverse array of commonly employed sensors, establishing it as the most comprehensive dataset presently available." + }, + { + "type": "table", + "bbox": [ + 0.245, + 0.546, + 0.761, + 0.633 + ], + "angle": 0, + "content": "
SensorTypeHFOV(°)VFOV(°)ResolutionFPS
OS128128 beams MS LiDAR360[-22.5, 22.5]128 * 102410
XT3232 beams MS LiDAR360[-16, 15]32 * 204810
M1SS LiDAR[-60, 60][-12.5, 12.5]126 * 62510
Radar4D Radar[-75, 75][-15, 15]>60010
HIK Cam.Camera[-31.2, 31.2][-27.7, 27.7]1080 * 192010
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.633, + 0.784, + 0.661 + ], + "angle": 0, + "content": "Table 1: Sensor specifications for CMD. MS and SS denote mechanical spinning and solid-state respectively. FPS refers to frames per second." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Sensor layout. In our CMD, all sensors are oriented in the same direction as the vehicle's forward direction. Such a configuration is instrumental in ensuring that the most critical information is captured within the area of greatest overlap of FOV across different sensors. To achieve this alignment, a rigid bracket has been designed, aligning all sensors along a vertically aligned axis. We have deliberately placed all sensors as close to each other as possible to optimize FOV overlap. In addition to this, a camera system comprising three evenly spaced cameras has been integrated. This system spans a \\(150^{\\circ}\\) horizontal field of view (HFOV), as depicted in Fig. 2(c). Such a configuration is pivotal" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Deng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "for ensuring data consistency across different mechanisms. The overall FOV is shown in Fig. 2(b)." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.206, + 0.338, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.206, + 0.438, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.207, + 0.492, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.208, + 0.631, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.207, + 0.782, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.32, + 0.785, + 0.362 + ], + "angle": 0, + "content": "Fig. 2: The illustration of sensor layout. (a) demonstrates the specific layout of each sensor; (b) showcases the field of view angles for each sensor; (c) demonstrates the positions of each sensor along the z-axis." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.418, + 0.53, + 0.433 + ], + "angle": 0, + "content": "3.2 Synchronization and Calibration" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.719 + ], + "angle": 0, + "content": "Synchronization and trigger. To achieve synchronization among different sensors when modeling the world, we utilized the Precision Time Protocol Version 2 [1] to establish a time synchronization system with an error margin within \\(1\\mathrm{ms}\\). Specifically, we employed the CoolShark AUTO 66 unit as the PTP master clock, while other sensors and the host served as PTP slave clocks. Once synchronized, M1 is set to generate a frame every \\(100\\mathrm{ms}\\), and both OS128 and XT32 generate frames simultaneously using phase-locking. Meanwhile, the radar is triggered by a network signal sent from the host. Due to the differing principles of data acquisition between cameras and LiDAR, the cameras are triggered by OS128. Whenever OS128 scans to \\(0^{\\circ}\\), it sends out a trigger signal. This signal undergoes a delay process via a microcontroller before being fed into the cameras. For each camera, \\(T_{delay} = \\frac{\\theta_{cam}}{2\\pi} - \\frac{1}{2} T_{exposure}\\) where \\(T_{exposure}\\) means the exposure time for camera and \\(\\theta_{cam}\\) means the angle at which the camera is positioned. By this, the time when OS128 scans the central of the camera's HFOV, is the central of its exposure period, which ensures a lower frame time offset between the cameras and the LiDARs. This low-latency synchronization ensures that the same frame of data captured by different sensors effectively reflects their respective mechanism differences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Calibration. Achieving high-quality data in a multi-sensor setup relies heavily on calibration. (1) For the LiDARs (i.e. OS128, M1, XT32) and radar, we calibrate their extrinsic parameters using Generalized-ICP [28]. Several corner reflectors in the space are used to obtain more precise corner points for radar calibration. (2) For the cameras, we calibrate their intrinsic parameters and extrinsic parameters with MATLAB Toolkit [12] and OpenCV [4]. The results of calibration are shown in Fig. 4. Subsequent to calibration, we established a well-defined coordinate system, as illustrated in Fig. 2(c). We transfer XT32 and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.692, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CMD" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "radar to the coordinate system of OS128 and move this coordinate system vertically down to the ground to get the ego system. The ego system has the x-axis pointing forward, the y-axis to the left, and the z-axis upward. The camera coordinate system, in 2D, aligns the x-axis with image width and y-axis with height, starting from the top-left corner." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.248, + 0.357, + 0.261 + ], + "angle": 0, + "content": "3.3 Annotation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.274, + 0.784, + 0.334 + ], + "angle": 0, + "content": "Annotation area. With precise time synchronization and calibration, we can jointly annotate data from multiple sensors, improving accuracy, especially for distant objects. We annotate objects within \\(160m\\) and within the HFOV of M1. We believe that the data within this range is highly representative." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.785, + 0.456 + ], + "angle": 0, + "content": "Annotation rules. There are 13 annotated categories, namely: Car, Van, Bus, Truck, Semi-Trailer towing vehicle, Special Vehicles, Motorcycle, Bicycle, Tricycle, Adult Pedestrian, Children Pedestrian, Animal, and Barrier. We labeled each object as a 9-DoF 3D bounding box \\((x,y,z,l,w,h,\\theta_x,\\theta_y,\\theta_z)\\). Where \\(x,y,z\\) represents the center coordinates, \\(l,w,h\\) denotes length, width height, and \\(\\theta_x,\\theta_y,\\theta_z\\) are the rotation angles around the x,y,z axis. Additionally, we provide a motion state and tracking ID for all objects. Occlusion situations are also provided. More annotation specifications are provided in the appendix." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.482, + 0.784, + 0.615 + ], + "angle": 0, + "content": "
DatasetMechanismMS LiDAR(beams)SS LiDAR 4DRadar Cam.Anno.Frames TrackingIlluminationReal
LowHigh
KITTI Det. [13]X64XX✓15KX✓✓
Waymo [30]X64XX✓230K✓✓✓
nuScenes [5]Single32XXX✓40K✓✓✓
Argoverse [6]32*2XXX✓44K✓✓✓
Lyft L5 [15]40*264XX✓30K✓X✓
K-Radar [22]X64, 128X✓✓35K✓✓✓
Lidar-CS [11]16, 3264, 128✓XX14KXXX
PandaSet [44]MultiX64✓X✓8.2K✓✓✓
TJ4DRadSet [52]X64X✓✓7.8K✓✓✓
VoD [23]X64X✓✓8.7KXX✓
CMD32128✓✓✓10K✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.615, + 0.784, + 0.669 + ], + "angle": 0, + "content": "Table 2: Comparison with existing datasets. \"Mechanism\" means whether the dataset contains multiple mechanisms of 3D sensor. \"Tracking\" signifies the relevance of annotations for tracking. \"Illumination\" indicate the inclusion of various light conditions. \"Real\" refers to whether the dataset is real or synthetic." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.738, + 0.403, + 0.753 + ], + "angle": 0, + "content": "3.4 Dataset Analytic" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Diversity of scenes. In CMD, scenes are carefully selected for wide coverage, primarily encompassing urban areas, suburbs, campuses, bridges, and tunnels with different illuminations (see Fig. 3(a)). We select 50 high-quality sequences, each spanning 20 seconds, equating to 200 frames per sensor, culminating in a dataset of 40,000 frames of point cloud data and 30,000 frames of image" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Deng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "data. This comprehensive collection offers a rich resource for research on cross-mechanism domain adaptation. These 50 sequences are evenly divided into 30, 10 and 10 for training, validation, and testing according to different scenarios. This arrangement ensures a balanced distribution of data across environments, crucial for training reliable and generalizable 3D object detection models." + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.266, + 0.56, + 0.276 + ], + "angle": 0, + "content": "Cross Mechanism domain adaptation Dataset (CMD) for 3D Object Detection" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.277, + 0.576, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.268, + 0.787, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.375, + 0.784, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.487, + 0.788, + 0.545 + ], + "angle": 0, + "content": "Fig. 3: The scene and annotation details. (a) The upper section showcases the distribution of road types, while the lower section demonstrates the distribution across various illumination types; (b) presents the distribution of objects' numbers for different categories; (c) presents the distribution of objects' numbers along different distances." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.59, + 0.788, + 0.682 + ], + "angle": 0, + "content": "Distribution of annotations. We annotated approximately 230,000 3D bounding boxes based on the rules described in Section 3.3. The distribution of objects' numbers for different categories is presented in Fig. 3(b). The distribution of objects' numbers along different distances is shown in Fig. 3(c). Among these categories, the most concerned \"Car\" has the highest annotation count, reaching approximately 92,000 instances." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Comparisons between mechanisms. The average number of points for these four sensors is quite different. From most to least, they are M1, OS128, XT32, and 4D radar. For a better understanding, we show the visualization results in Fig. 4. Point clouds from sensors with different mechanisms exhibit substantial differences. (1) OS128 and XT32 both adopt a mechanism spinning style, clearly modeling the rigid vehicles with minimal vertical distortion. The main difference between them is density. (2) Due to the use of a zigzag scan pattern, M1 tends to introduce a bit of horizontal displacement. (3) Point clouds from radar sensors are generally sparse and inaccurate, posing challenges in accurately distinguishing the geometric shapes of objects." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.693, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CMD" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.784, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.436, + 0.788, + 0.477 + ], + "angle": 0, + "content": "Fig. 4: Different mechanisms and their sample data. The left side of the image denotes the various scenes of data collection, while the bottom indicates the types of sensors involved." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.504, + 0.445, + 0.518 + ], + "angle": 0, + "content": "3.5 DIG Baseline Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.784, + 0.586 + ], + "angle": 0, + "content": "In this section, we propose DIG (Density-Intensity-Geometry), a streamlined and potent baseline for cross-mechanism domain adaptation. It consists of three key components: Box-Cox Intensity Log Normalization (BCN), Beam-Distance Down Sampling (BDS), and Geometry-Aware Label Mixing (GAM)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.788, + 0.768 + ], + "angle": 0, + "content": "Box-Cox intensity log Normalization (BCN). The distribution of intensity for point clouds may significantly differ across various mechanisms. See Fig. 1(b). Recent studies have either opted to discard [45] [38] intensity or straightforwardly normalized it to [0,1] [36]. However, it is observed that intensity plays a crucial role in detecting foreground objects across diverse domains. We contend that effectively incorporating intensity can lead to enhanced performance in domain adaptation. Motivated by this insight, we introduce the BCN to mitigate disparities in intensity. For both intensities (denoted as \\(i\\)) for source and target domain point clouds, we initially take the logarithm to address biases stemming from different sensor definitions of intensity ranges. Subsequently, we apply the Box-Cox transformation to enhance their normality and homoscedasticity, respectively. This process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.773, + 0.786, + 0.805 + ], + "angle": 0, + "content": "\\[\ni (\\lambda) = \\frac {\\log (i + 1) ^ {\\lambda} - 1}{\\lambda}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.786, + 0.84 + ], + "angle": 0, + "content": "where \\(\\lambda\\) refers to the box-cox transfer parameter [2]. It's different for source and target domains, and remains constant respectively. After being normalized to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Deng et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "[0,1], the transformed intensity information exhibits a good level of similarity between the source and target domains, thus narrowing the domain gap." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Beam-Distance based down Sampling (BDS). As shown in Fig. 1(a), the density distributions of points vary remarkably among different mechanisms. We observe that, within a relatively short distance, domain adaptation for different mechanisms is more significantly impacted by variations in point density. However, when it comes to longer distances, the point cloud is excessively sparse, which has a greater impact on object detection accuracy. To deal with this problem, we introduce the BDS that assigns different sampling probabilities based on both beam and distances. Denote \\( P_{\\text{keep}}(p) \\) as the probability of a point (p) being retained, and it can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.341, + 0.32, + 0.786, + 0.349 + ], + "angle": 0, + "content": "\\[\nP _ {\\text {k e e p}} (p) = 1 - \\mathbf {1} \\left\\{\\text {b e a m} (p) \\notin \\mathcal {B} \\right\\} \\exp \\left(- \\frac {1}{K} | | p | |\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.788, + 0.466 + ], + "angle": 0, + "content": "where \\(K\\) controls the rate of probability increase, and we set it to 35 empirically. \\(\\mathcal{B}\\) means the set of target beam indexes that are closest to the source domain. Function \\(\\text{beam}(p)\\) denotes the index of the beam that point \\(p\\) belongs to. If this beam information is not directly available, an alternative approach involves calculating the angle of each point, denoted as \\(\\theta(p) = \\arctan \\left( \\frac{z}{\\sqrt{x^2 + y^2}} \\right)\\). This angle can then be used to apply a clustering algorithm, such as K-means [20], to approximate the beam index for each point." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.664 + ], + "angle": 0, + "content": "Geometry-Aware label Mixing (GAM). Point clouds captured from the same object by diverse sensors frequently exhibit notable geometric variances. To tackle this challenge, we introduce the Geometry-Aware Mix (GAM) module. In our approach, we initially utilize a model trained on the source domain to generate pseudo-labels within the target domain. Subsequently, instances from the target domain with high-confidence pseudo-labels are integrated back into the source domain for further training. Inspired by GBA [26], our second training phase progressively reduces source domain labels while maintaining a constant number of target domain pseudo-labels. This technique aims to gradually adapt the model to the geometric characteristics of the target domain. Through this method, the model becomes more adept at learning and adapting to the unique data characteristics of the target domain, thereby improving its ability for cross-mechanism domain adaptation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.684, + 0.454, + 0.701 + ], + "angle": 0, + "content": "4 Experimental Result" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.712, + 0.414, + 0.729 + ], + "angle": 0, + "content": "4.1 Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.842 + ], + "angle": 0, + "content": "All of our experiments were conducted using four Nvidia 3090 GPUs and the open-source code repository OpenPCDet [31]. The Voxel-RCNN [9] served as the detector for most experiments. The batch size was set to 32, and the learning rate was 0.003. Similar to KITTI [13], the grid range was defined as \\([0\\mathrm{m},70.4\\mathrm{m}]\\), \\([-70.4\\mathrm{m},70.4\\mathrm{m}]\\), and \\([-2\\mathrm{m},4\\mathrm{m}]\\) along the x, y, and z axis respectively. Point cloud data were cropped to HFOV \\([-60^{\\circ},60^{\\circ}]\\) with horizontal distances smaller than \\(70.4\\mathrm{m}\\). Voxel size was set to \\([0.1,0.1,0.15]\\) along the x, y, and z axis respectively." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.693, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CMD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.147, + 0.477, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.147, + 0.728, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.236, + 0.475, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.236, + 0.726, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.275, + 0.324, + 0.418, + 0.337 + ], + "angle": 0, + "content": "\\(\\pmb{X}^{S}\\) Points in source domain" + }, + { + "type": "image_footnote", + "bbox": [ + 0.425, + 0.324, + 0.582, + 0.337 + ], + "angle": 0, + "content": "\\(\\mathcal{F}(\\theta^s)\\) Model in source domain" + }, + { + "type": "image_footnote", + "bbox": [ + 0.589, + 0.323, + 0.72, + 0.337 + ], + "angle": 0, + "content": "\\(\\varphi^a\\) Points down sampling" + }, + { + "type": "list", + "bbox": [ + 0.275, + 0.324, + 0.72, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.275, + 0.339, + 0.411, + 0.351 + ], + "angle": 0, + "content": "\\(X^t\\) Points in target domain" + }, + { + "type": "image_footnote", + "bbox": [ + 0.426, + 0.339, + 0.577, + 0.351 + ], + "angle": 0, + "content": "\\(\\mathcal{F}(\\theta^t)\\) Model in target domain" + }, + { + "type": "image_footnote", + "bbox": [ + 0.589, + 0.339, + 0.722, + 0.351 + ], + "angle": 0, + "content": "\\(\\varphi^i\\) Intensity normalization" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.364, + 0.788, + 0.42 + ], + "angle": 0, + "content": "Fig. 5: Similarities and differences among baselines. ST3D [45] leverage random object scaling, generating and denoising pseudo-labels. DTS [16] down sample point clouds before training source-detector. Our DIG jointly downs ample point clouds, transforms intensity, and makes use of pseudo-labels." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.446, + 0.414, + 0.46 + ], + "angle": 0, + "content": "4.2 Evaluation Metric" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.789, + 0.558 + ], + "angle": 0, + "content": "Mean Average Precision (mAP). mAP is a commonly used object detection evaluation metric. Here when calculating mAP, we use 3D IoU thresholds of 0.5, 0.5, 0.25, and 0.25 for Car, Truck, Pedestrian, and Cyclist, respectively. Following the ONCE [21] dataset, we calculated an orientation-aware AP on 50 precision steps. The AP among different detection distances (e.g., overall and 0 to \\(30\\mathrm{m}\\)) are also computed for detailed analysis. Formally, the AP is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.565, + 0.787, + 0.6 + ], + "angle": 0, + "content": "\\[\nA P = 1 0 0 \\int_ {0} ^ {1} \\max \\left\\{p \\left(r ^ {\\prime} \\mid r ^ {\\prime} \\geq r\\right) \\right\\} d r, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.605, + 0.787, + 0.636 + ], + "angle": 0, + "content": "where \\( p(r) \\) refers to the precision-recall curve that calculated by 50 recall positions. The mAP is the average of AP across all categories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.635, + 0.788, + 0.696 + ], + "angle": 0, + "content": "Mean Closed Gap (mCG). Directly applying AP or mAP as evaluation metrics for domain adaptation is not intuitive since different sensors do not share the same domain gap. Closed gap [35] is used for measuring the effectiveness of a method on a single domain adaptation task. CG can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.38, + 0.701, + 0.787, + 0.734 + ], + "angle": 0, + "content": "\\[\nC G = 1 0 0 \\frac {m A P _ {\\text {m o d e l}} - m A P _ {D T}}{m A P _ {\\text {O r a c l e}} - m A P _ {D T}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.737, + 0.788, + 0.799 + ], + "angle": 0, + "content": "where \\(DT\\) means direct transfer using a model trained sorely on source domain data. Oracle implies results obtained by the full training on target domain. To quantify the universality of domain adaptation methods, we define a new metric named mean closed gap (mCG), formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.38, + 0.805, + 0.787, + 0.838 + ], + "angle": 0, + "content": "\\[\nm C G = \\sum_ {s \\in \\mathbf {M}} \\sum_ {t \\in \\mathbf {M}} \\mathbf {1} (s \\neq t) C G _ {s \\rightarrow t}, \\tag {5}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Deng et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.162 + ], + "angle": 0, + "content": "where \\(\\mathbf{M}\\) refers to the set of sensors, including OS128, XT32, and M1 in CMD." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.188, + 0.787, + 0.264 + ], + "angle": 0, + "content": "
MethodmCGCG
M1→OS128XT32→OS128OS128→M1XT32→M1OS128→XT32M1→XT32
ST3D [45]-16.72-9.2734.00-7.95-5.07-43.39-68.61
ST3D++ [46]09.2312.6554.4902.9917.2307.56-39.54
DTS [16]29.2248.2822.9209.0207.0934.5053.51
DIG(Ours)42.8954.4155.0933.6925.2249.8739.06
" + }, + { + "type": "table_caption", + "bbox": [ + 0.312, + 0.264, + 0.688, + 0.276 + ], + "angle": 0, + "content": "Table 3: Overall mean closed gap for baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.32, + 0.785, + 0.427 + ], + "angle": 0, + "content": "Discussion. Currently, commonly used CG is tailored for a single task, usually only considering the \"Car\" category [16], and cannot reflect the comprehensive performance of methods across various domain adaptation tasks. In contrast, mCG includes all MS LiDARs with varying beam numbers and SS LiDARs (i.e., OS128, XT32, M1), measuring performance from six experimental groups, thus providing the most comprehensive evaluation of cross-mechanism domain adaptation. See the appendix for 4D radar results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.451, + 0.46, + 0.465 + ], + "angle": 0, + "content": "4.3 Results and Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.785, + 0.627 + ], + "angle": 0, + "content": "Domain adaptation baselines. We selected four baseline methods. Fig. 5 illustrates the similarities and differences between them. (1) Direct transfer (DT) means directly applying the source model to the target dataset. (2) ST3D [45] and \\(\\mathrm{ST3D + + }\\) [46] share a similar teacher-student architecture. The source model is trained with random object scaling, then used to generate pseudo-labels. The difference is that \\(\\mathrm{ST3D + + }\\) comes with a more effective denoising method to obtain more high-quality pseudo-labels. (3) DTS [16] introduced Random Beam Re-Sampling (RBRS) to enhance the robustness to varying beam densities. We reproduced the RBRS on CMD, providing representative baseline results. (4) Our proposed DIG." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.785, + 0.749 + ], + "angle": 0, + "content": "Overall comparison results. Table 3 shows the overall results. Since the ST3D and \\(\\mathrm{ST3D + + }\\) are mostly designed for cross-geographical domain adaptation that highly relies on pseudo-label quality, the two detectors show lower performance in our cross-mechanism dataset. DTS takes density into consideration, thus outperforming \\(\\mathrm{ST3D + + }\\) by 19.99. Our DIG outperformed all previous methods by a large margin. Its mCG reached 42.89, surpassing ST3D, \\(\\mathrm{ST3D + + }\\), and DTS by 59.61, 33.66, and 13.67 respectively, as our method considers all the typical domain disparities in the cross-mechanism domain adaptation problem." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Performance discussion. A more detailed composition of mCG is shown in Table 4. We can get several conclusions from these results. (1) DIG exhibits notable superiority in the categories of pedestrians and bicycles, both across all distances and within a detection range of 30 meters. The similarity in these categories is the fewer points within the targets. We believe that sparse points make intensity the key factor for adapting to different domains. DIG's BCN module" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.693, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CMD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.144, + 0.788, + 0.567 + ], + "angle": 0, + "content": "
TaskMethodCGmAPCarTruckPedCyc
OS128Oracle10026.4436.50/69.3417.45/38.5118.79/38.9933.0/69.22
XT32↓OS128DT006.2911.76/20.5904.61/08.2301.54/02.1507.25/08.90
ST3D [45]34.0013.1422.78/47.3910.41/15.1006.83/16.0112.55/33.05
ST3D++ [46]54.4917.2727.87/61.0213.50/22.6108.57/19.7618.95/49.32
DTS [16]22.9210.9114.26/49.4306.72/27.6709.29/20.9413.37/41.21
DIG(Ours)55.0917.3922.73/60.5913.60/30.1611.56/25.6721.70/57.46
M1↓OS128DT006.9213.23/30.2105.08/07.8601.98/03.6107.38/09.98
ST3D [45]-9.2705.1110.24/39.1104.06/11.3200.00/00.0006.14/21.18
ST3D++ [46]12.5609.3913.46/46.9306.61/23.1605.85/13.6811.65/35.06
DTS [16]48.2516.3421.60/58.9912.67/34.1409.57/23.4921.51/58.70
DIG(Ours)54.4117.5423.03/57.0809.71/25.9509.71/25.9525.00/66.13
M1Oracle10030.0942.01/70.6419.96/40.0117.54/36.9940.84/68.49
XT32↓M1DT013.3220.33/42.8209.47/19.8404.81/11.4718.67/36.06
ST3D [45]-5.0712.4730.34/55.4104.20/03.8200.19/00.2215.13/25.95
ST3D++ [46]17.2316.2132.92/61.0705.24/06.0402.18/04.1724.48/49.39
DTS [16]07.0914.5123.26/56.9408.78/28.0206.99/14.7219.01/48.44
DIG(Ours)25.2217.5527.19/64.7009.26/22.7905.72/12.5728.04/57.34
OS128↓M1DT014.3627.07/56.5409.91/24.2501.13/03.8119.32/33.49
ST3D [45]-7.9513.1128.71/60.3708.26/07.2506.37/17.4809.08/12.72
ST3D++ [46]02.9914.8332.98/68.4506.63/08.4507.28/19.4612.43/17.89
DTS [16]09.0215.7822.11/54.0111.84/30.2611.88/28.5817.27/40.40
DIG(Ours)33.6919.6626.85/62.8911.86/28.1113.69/32.2426.19/60.40
XT32Oracle10026.5936.84/72.4719.26/41.0116.23/32.7434.01/66.59
M1↓XT32DT014.0722.10/59.8610.44/21.4806.29/15.2417.44/45.69
ST3D [45]-68.6105.4811.39/38.0502.92/06.7800.00/00.0002.92/06.78
ST3D++ [46]-39.5409.1214.56/47.9407.47/22.0703.15/05.0211.28/32.18
DTS [16]53.5120.7730.62/68.0315.67/33.0109.48/21.6327.33/57.48
DIG(Ours)39.0618.9625.95/68.7814.77/34.6511.45/26.4723.66/60.92
OS128↓XT32DT011.7722.25/58.2008.45/19.1601.66/04.5014.71/41.05
ST3D [45]-43.9905.3411.46/45.4101.72/05.8700.00/00.0008.18/22.82
ST3D++ [46]07.5612.8917.81/56.9905.31/19.9810.69/24.2217.75/48.29
DTS [16]34.5016.0918.87/59.7813.63/27.7011.26/25.4920.59/58.19
DIG(Ours)49.8719.1627.20/63.8910.75/26.7911.76/27.1626.91/64.22
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.608 + ], + "angle": 0, + "content": "Table 4: Detailed experimental results for all cross-mechanism settings. For each category, we provide the Average Precision (AP) results for all distances and within a range of 30 meters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "aligns intensity well, which explains its strong performance in these cases. (2) The results indicate a significant decrease in mAP while using OS128 as the target domain. This decrease is primarily attributed to its unique intensity range of [0, 512], unlike the [0, 256] range of other sensors. The pre-trained model finds it challenging to handle unfamiliar intensities. Unlike the other three methods, which either discarded the intensity information or did not include a dedicated module for it, we addressed this issue with BCN, resulting in a significant performance improvement. (3) All these methods perform unsatisfactorily for domain adaptation from XT32 to M1. Even DIG, the best one among them, gets CGs of 25.22 on it. This observation means that M1 and XT32 may suffer from larger domain gaps, highlighting the need for further investigation in the future. (4) ST3D++ gets a CG of 54.49 for XT32 to OS128, which is mainly due to the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Deng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "great performance on APs for 0-30m. As they are both MS LiDAR, they suffer more from differences in density. Pseudo-labels from XT32 might look like harder cases for OS128. Refining model parameters with these cases can dramatically improve the performance. (5) It indicates that DTS plays a crucial role when there is a substantial difference in the number of beams. CG on tasks that contain XT32 gets nice performance improvement. Specifically, CG can reach 34.51 and 53.51 for OS128 to XT32 and M1 to XT32 respectively." + }, + { + "type": "table", + "bbox": [ + 0.339, + 0.276, + 0.667, + 0.363 + ], + "angle": 0, + "content": "
modulesM1 ⇒ OS128OS128 ⇒ XT32
BDS BCN GAMmAPCGmAPCG
✓09.2211.7814.6919.70
✓17.1452.5316.4731.71
✓17.4754.0418.9548.45
✓17.5454.4119.1649.87
" + }, + { + "type": "table_caption", + "bbox": [ + 0.331, + 0.364, + 0.672, + 0.377 + ], + "angle": 0, + "content": "Table 5: Ablation study for DIG baseline method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.789, + 0.569 + ], + "angle": 0, + "content": "Ablation study for DIG. We conduct ablation studies to analyze the effectiveness of each component in the DIG baseline method. Two representative tasks are selected. \\( M1 \\Rightarrow OS128 \\) and \\( OS128 \\Rightarrow XT32 \\) respectively refer to domain adaptation for sensors with different scanning patterns and different beams/densities. As shown in Table 5, For both tasks, BCN is anticipated to be the most effective module, contributing 52.53 and 31.71 CG on each respective task. This is attributed to the substantial intensity differences between the source and target domains in both tasks. BDS also exhibits commendable performance in both tasks, achieving 11.78 and 19.70 CG, respectively. Although less conspicuous, GAM proves to be effective for both tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.587, + 0.36, + 0.603 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.779 + ], + "angle": 0, + "content": "We presented a dataset specifically designed for cross-mechanism domain adaptation, incorporating mechanical LiDARs with both high-res and low-res beams, solid-state LiDAR, and 4D millimeter-wave radar. These sensors undergo precise time synchronization, allowing them to simultaneously generate point clouds of the same external environment. Data from diverse scenes were selected and meticulously annotated. We believe the proposed dataset would hugely facilitate research of cross-mechanism 3D detection, as it currently stands as the most comprehensive point cloud 3D detection dataset in terms of sensor types. We also provide a novel DIG method as well as complete experimental results of recent methods on our dataset. They can be used as reliable baselines to further benefit the research communities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.84 + ], + "angle": 0, + "content": "In the future, we will continue to explore more usage of CMD, including utilizing data from multiple mechanisms as the source domain to enhance domain adaptation effects, and employing CMD as a tool for validating the performance of multi-modal 3D object detection." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.692, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CMD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "1. IEEE standard for a precision clock synchronization protocol for networked measurement and control systems. IEEE Std 1588-2008 (Revision of IEEE Std 1588-2002) pp. 1-269 (2008). https://doi.org/10.1109/IEEESTD.2008.4579760" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.245 + ], + "angle": 0, + "content": "2. Atkinson, A.C., Riani, M., Corbellini, A.: The box-cox transformation: Review and extensions (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.246, + 0.785, + 0.3 + ], + "angle": 0, + "content": "3. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.301, + 0.785, + 0.328 + ], + "angle": 0, + "content": "4. Bradski, G.: The opencv library. Dr. Dobb's Journal: Software Tools for the Professional Programmer 25(11), 120-123 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.329, + 0.785, + 0.398 + ], + "angle": 0, + "content": "5. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (Jun 2020). https://doi.org/10.1109/cvpr42600.2020.01164, http://dx.doi.org/10.1109/cvpr42600.2020.01164" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.785, + 0.453 + ], + "angle": 0, + "content": "6. Chang, M.F., Lambert, J., Sangkloy, P., Singh, J., Bak, S., Hartnett, A., Wang, D., Carr, P., Lucey, S., Ramanan, D., et al.: Argoverse: 3d tracking and forecasting with rich maps. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8748-8757 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.454, + 0.785, + 0.508 + ], + "angle": 0, + "content": "7. Chen, Y., Yu, Z., Chen, Y., Lan, S., Anandkumar, A., Jia, J., Alvarez, J.M.: Focalformer3d: Focusing on hard instance for 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8394-8405 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "8. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21674-21683 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.551, + 0.785, + 0.592 + ], + "angle": 0, + "content": "9. Deng, J., Shi, S., Li, P., gang Zhou, W., Zhang, Y., Li, H.: Voxel r-cnn: Towards high performance voxel-based 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.785, + 0.633 + ], + "angle": 0, + "content": "10. Ding, G., Zhang, M., Li, E., Hao, Q.: Jst: Joint self-training for unsupervised domain adaptation on 2d&3d object detection. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 477-483. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.785, + 0.661 + ], + "angle": 0, + "content": "1. Fang, J., Zhou, D., Zhao, J., Tang, C., Xu, C.Z., Zhang, L.: Lidar-cs dataset: Lidar point cloud dataset with cross-sensors for 3d object detection (Jan 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "2. Fetic, A., Juric, D., Osmankovic, D.: The procedure of a camera calibration using camera calibration toolbox for matlab. In: 2012 Proceedings of the 35th International Convention MIPRO. pp. 1752-1757. IEEE (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.757 + ], + "angle": 0, + "content": "3. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition (Jun 2012). https://doi.org/10.1109/cvpr.2012.6248074, http://dx.doi.org/10.1109/cvpr.2012.6248074" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "4. Hegde, D., Sindagi, V., Kilic, V., Cooper, A.B., Foster, M., Patel, V.: Uncertainty-aware mean teacher for source-free unsupervised domain adaptive 3d object detection. arXiv preprint arXiv:2109.14651 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "5. Houston, J., Zuidhof, G., Bergamini, L., Ye, Y., Chen, L., Jain, A., Omari, S., Iglovikov, V., Ondruska, P.: One thousand and one hours: Self-driving motion prediction dataset. In: Conference on Robot Learning. pp. 409-418. PMLR (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Deng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "16. Hu, Q., Liu, D., Hu, W.: Density-insensitive unsupervised domain adaption on 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17556-17566 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "17. Huang, X., Wu, H., Li, X., Fan, X., Wen, C., Wang, C.: Sunshine to rainstorm: Cross-weather knowledge distillation for robust 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 2409–2416 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.785, + 0.271 + ], + "angle": 0, + "content": "18. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. ArXiv (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.785, + 0.325 + ], + "angle": 0, + "content": "19. Luo, Z., Cai, Z., Zhou, C., Zhang, G., Zhao, H., Yi, S., Lu, S., Li, H., Zhang, S., Liu, Z.: Unsupervised domain adaptive 3d detection with multi-level consistency. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8866-8875 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.325, + 0.785, + 0.367 + ], + "angle": 0, + "content": "20. MacQueen, J., et al.: Some methods for classification and analysis of multivariate observations. In: Proceedings of the fifth Berkeley symposium on mathematical statistics and probability. vol. 1, pp. 281-297. Oakland, CA, USA (1967)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.366, + 0.785, + 0.407 + ], + "angle": 0, + "content": "21. Mao, J., Niu, M., Jiang, C., Liang, H., Liang, X., Li, Y., Ye, C., Zhang, W., Li, Z., Yu, J., Xu, H., Xu, C.: One million scenes for autonomous driving: Once dataset. Cornell University - arXiv, Cornell University - arXiv (Jun 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.407, + 0.785, + 0.434 + ], + "angle": 0, + "content": "22. Paek, D.H., Kong, S.H., Wijaya, K.: K-radar: 4d radar object detection dataset and benchmark for autonomous driving in various weather conditions (Jun 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.434, + 0.785, + 0.459 + ], + "angle": 0, + "content": "23. Palffy, A., Pool, E., Baratam, S., Kooij, J., Gavrila, D.: Multi-class road user detection with 3+1d radar in the view-of-delft dataset" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.459, + 0.785, + 0.5 + ], + "angle": 0, + "content": "24. Peng, X., Zhu, X., Ma, Y.: Cl3d: Unsupervised domain adaptation for cross-lidar 3d detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2047-2055 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.501, + 0.785, + 0.528 + ], + "angle": 0, + "content": "25. Raj, T., Hanim Hashim, F., Baseri Huddin, A., Ibrahim, M.F., Hussain, A.: A survey on lidar scanning mechanisms. Electronics 9(5), 741 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.528, + 0.785, + 0.568 + ], + "angle": 0, + "content": "26. Rochan, M., Chen, X., Grandhi, A., Corral-Soto, E.R., Liu, B.: Domain adaptation in 3d object detection with gradual batch alternation training. arXiv preprint arXiv:2210.10180 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.568, + 0.785, + 0.595 + ], + "angle": 0, + "content": "27. Roriz, R., Cabral, J., Gomes, T.: Automotive lidar technology: A survey. IEEE Transactions on Intelligent Transportation Systems 23(7), 6282-6297 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.595, + 0.785, + 0.622 + ], + "angle": 0, + "content": "28. Segal, A., Haehnel, D., Thrun, S.: Generalized-icp. In: Robotics: science and systems. vol. 2, p. 435. Seattle, WA (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.622, + 0.785, + 0.676 + ], + "angle": 0, + "content": "29. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR). pp. 10526 - 10535 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.676, + 0.785, + 0.772 + ], + "angle": 0, + "content": "30. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (Jun 2020). https://doi.org/10.1109/cvpr42600.2020.00252, http://dx.doi.org/10.1109/cvpr42600.2020.00252" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.772, + 0.785, + 0.8 + ], + "angle": 0, + "content": "31. Team, O.D.: Openpcdet: An open-source toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "32. Tsai, D., Berrio, J.S., Shan, M., Nebot, E., Worrall, S.: Ms3d: Leveraging multiple detectors for unsupervised domain adaptation in 3d object detection. arXiv preprint arXiv:2304.02431 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.692, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CMD" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "33. Tsai, D., Berrio, J.S., Shan, M., Nebot, E., Worrall, S.: Viewer-centred surface completion for unsupervised domain adaptation in 3d object detection. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 9346-9353. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "34. Tsai, D., Berrio, J.S., Shan, M., Worrall, S., Nebot, E.: See eye to eye: A lidar-agnostic 3d detection framework for unsupervised multi-target domain adaptation. IEEE Robotics and Automation Letters 7(3), 7904-7911 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.787, + 0.3 + ], + "angle": 0, + "content": "35. Wang, Y., Chen, X., You, Y., Li, L.E., Hariharan, B., Campbell, M., Weinberger, K.Q., Chao, W.L.: Train in germany, test in the usa: Making 3d object detectors generalize. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11713-11723 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.301, + 0.787, + 0.342 + ], + "angle": 0, + "content": "36. Wang, Y., Yin, J., Li, W., Frossard, P., Yang, R., Shen, J.: Ssa3d: Semi-supervised domain adaptation for 3d object detection from point cloud. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2707-2715 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.343, + 0.787, + 0.397 + ], + "angle": 0, + "content": "37. Wang, Z., Ding, S., Li, Y., Fenn, J., Roychowdhury, S., Wallin, A., Martin, L., Ryvola, S., Sapiro, G., Qiu, Q.: Cirrus: A long-range bi-pattern lidar dataset. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 5744-5750. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.398, + 0.787, + 0.439 + ], + "angle": 0, + "content": "38. Wei, Y., Wei, Z., Rao, Y., Li, J., Zhou, J., Lu, J.: Lidar distillation: Bridging the beam-induced domain gap for 3d object detection. In: European Conference on Computer Vision. pp. 179-195. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.44, + 0.787, + 0.481 + ], + "angle": 0, + "content": "39. Wu, H., Deng, J., Wen, C., Li, X., Wang, C.: Casa: A cascade attention network for 3d object detection from lidar point clouds. IEEE Transactions on Geoscience and Remote Sensing (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.482, + 0.787, + 0.508 + ], + "angle": 0, + "content": "40. Wu, H., Wen, C., Li, W., Yang, R., Wang, C.: Learning transformation-equivariant features for 3d object detection (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.509, + 0.787, + 0.55 + ], + "angle": 0, + "content": "41. Wu, H., Wen, C., Shi, S., Li, X., Wang, C.: Virtual sparse convolution for multimodal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21653-21662 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.551, + 0.787, + 0.592 + ], + "angle": 0, + "content": "42. Xia, Q., Chen, Y., Cai, G., Chen, G., Xie, D., Su, J., Wang, Z.: 3-d hanet: A flexible 3-d heatmap auxiliary network for object detection. IEEE Transactions on Geoscience and Remote Sensing 61, 1-13 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.592, + 0.787, + 0.647 + ], + "angle": 0, + "content": "43. Xia, Q., Deng, J., Wen, C., Wu, H., Shi, S., Li, X., Wang, C.: Coin: Contrastive instance feature mining for outdoor 3d object detection with very limited annotations. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6254-6263 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.647, + 0.787, + 0.715 + ], + "angle": 0, + "content": "44. Xiao, P., Shao, Z., Hao, S., Zhang, Z., Chai, X., Jiao, J., Li, Z., Wu, J., Sun, K., Jiang, K., Wang, Y., Yang, D.: Pandaset: Advanced sensor suite dataset for autonomous driving. In: 2021 IEEE International Intelligent Transportation Systems Conference (ITSC) (Sep 2021). https://doi.org/10.1109/itsc48978.2021.9565009, http://dx.doi.org/10.1109/itsc48978.2021.9565009" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.716, + 0.787, + 0.758 + ], + "angle": 0, + "content": "45. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d: Self-training for unsupervised domain adaptation on 3d object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10368-10378 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "46. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d++: Denoised self-training for unsupervised domain adaptation on 3d object detection. IEEE transactions on pattern analysis and machine intelligence 45(5), 6354-6371 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.8, + 0.787, + 0.841 + ], + "angle": 0, + "content": "47. Yang, Z., Sun, Y., Liu, S., Jia, J.: 3dssd: Point-based 3d single stage object detector. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR). pp. 11040-11048 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "J. Deng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "48. Yihan, Z., Wang, C., Wang, Y., Xu, H., Ye, C., Yang, Z., Ma, C.: Learning transferable features for point cloud detection via 3d contrastive co-training. Advances in Neural Information Processing Systems 34, 21493-21504 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.231 + ], + "angle": 0, + "content": "49. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.787, + 0.286 + ], + "angle": 0, + "content": "50. You, Y., Diaz-Ruiz, C.A., Wang, Y., Chao, W.L., Hariharan, B., Campbell, M., Weinbergert, K.Q.: Exploiting playbacks in unsupervised domain adaptation for 3d object detection in self-driving cars. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 5070-5077. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.286, + 0.787, + 0.342 + ], + "angle": 0, + "content": "51. You, Y., Phoo, C.P., Luo, K., Zhang, T., Chao, W.L., Hariharan, B., Campbell, M., Weinberger, K.Q.: Unsupervised adaptation from repeated traversals for autonomous driving. Advances in Neural Information Processing Systems 35, 27716-27729 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.787, + 0.384 + ], + "angle": 0, + "content": "52. Zheng, L., Ma, Z., Zhu, X., Tan, B., Li, S., Long, K., Sun, W., Chen, S., Zhang, L., Wan, M., Huang, L., Bai, J.: Tj4dradset: A 4d radar dataset for autonomous driving (Apr 2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.384 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CMD_ A Cross Mechanism Domain Adaptation Dataset for 3D Object Detection/55286f6e-e84b-4ee2-994b-347fc5f78848_origin.pdf b/2024/CMD_ A Cross Mechanism Domain Adaptation Dataset for 3D Object Detection/55286f6e-e84b-4ee2-994b-347fc5f78848_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9fa02ff1f2bd5eae3c8ae5690275a54e0fa7e2c1 --- /dev/null +++ b/2024/CMD_ A Cross Mechanism Domain Adaptation Dataset for 3D Object Detection/55286f6e-e84b-4ee2-994b-347fc5f78848_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6b784ebf3b174a7eb22506b38e0a9ce678f6cbf24c707ac92cb3f838f7bca4cd +size 5864807 diff --git a/2024/CMD_ A Cross Mechanism Domain Adaptation Dataset for 3D Object Detection/full.md b/2024/CMD_ A Cross Mechanism Domain Adaptation Dataset for 3D Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..56615e321f248119f7513c4eede5c275f2f8f0c8 --- /dev/null +++ b/2024/CMD_ A Cross Mechanism Domain Adaptation Dataset for 3D Object Detection/full.md @@ -0,0 +1,293 @@ +# CMD: A Cross Mechanism Domain Adaptation Dataset for 3D Object Detection + +Jinhao Deng $^{1,2,*}$ , Wei Ye $^{1,2,*}$ , Hai Wu $^{1,2}$ , Xun Huang $^{1,2}$ + +Qiming Xia $^{1,2}$ , Xin Li $^{4}$ , Jin Fang $^{3}$ , Wei Li $^{3,2}$ , + +Chenglu Wen $^{1,2,\text{图}}$ and Cheng Wang $^{1,2}$ + +$^{1}$ Fujian Key Laboratory of Sensing and Computing for Smart Cities, Xiamen University, Xiamen, P.R China + +$^{2}$ Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, Xiamen, P.R. China. + +3 Inceptio, Shanghai, P.R China + +$^{4}$ Section of Visual Computing and Interactive Media, Texas A&M University, Texas, USA + +Abstract. Point cloud data, representing the precise 3D layout of the scene, quickly drives the research of 3D object detection. However, the challenge arises due to the rapid iteration of 3D sensors, which leads to significantly different distributions in point clouds. This, in turn, results in subpar performance of 3D cross-sensor object detection. This paper introduces a Cross Mechanism Dataset, named CMD, to support research tackling this challenge. CMD is the first domain adaptation dataset, comprehensively encompassing diverse mechanical sensors and various scenes for 3D object detection. In terms of sensors, CMD includes 32-beam LiDAR, 128-beam LiDAR, solid-state LiDAR, 4D millimeter-wave radar, and cameras, all of which are well-synchronized and calibrated. Regarding the scenes, CMD consists of 50 sequences collocated from different scenarios, ranging from campuses to highways. Furthermore, we validated the effectiveness of various domain adaptation methods in mitigating sensor-based domain differences. We also proposed a DIG method to reduce domain disparities from the perspectives of Density, Intensity, and Geometry, which effectively bridges the domain gap between different sensors. The experimental results on the CMD dataset show that our proposed DIG method outperforms the state-of-the-art techniques, demonstrating the effectiveness of our baseline method. The dataset and the corresponding code are available at https://github.com/im-djh/CMD. + +Keywords: Dataset $\cdot$ 3D Object Detection $\cdot$ Domain Adaptation + +# 1 Introduction + +As a crucial component in robotics and autonomous driving systems, 3D object detection has garnered increasing attention from researchers. Due to the inherent + +depth of information, point cloud data enjoys a unique advantage in the domain of 3D object detection. Notably, researchers have developed several exceptional 3D object detection datasets and benchmarks (e.g., KITTI [13], nuScenes [5], Waymo [30], ONCE [21], etc.). Leveraging these high-quality datasets, numerous outstanding 3D object detection approaches [3,9,18,29,39,41,49] emerged, significantly facilitating the 3D object detection research. However, these prevalent datasets typically comprise mechanical spinning LiDAR only w.r.t. 3D point cloud sensors. The widely equipped low-cost automotive-grade sensors, e.g. solid-state LiDAR, in mass-produced vehicles [27] [25] are heavily overlooked even missing in those datasets. + +The truth of outdatedness of existing datasets in terms of sensors behind is that point cloud sensors are undergoing really rapid advancements. Commonly used sensors now include: (1) mechanical spinning LiDAR with low-beam (e.g., 32 beams) or high-beam (e.g., 128 beams); (2) solid-state LiDAR; and (3) the recently acclaimed 4D millimeter-wave radar. These sensors differ either in the number of beams they use or in their world modeling patterns. We refer to these sensor differences as various mechanisms. The data acquired by different mechanisms may exhibit domain disparities in terms of density, intensity, and geometry (see Fig. 1(a), (b) and (c)). As a result, 3D detectors trained with data from one sensor often incur substantial accuracy degradation when directly applied to another sensor (see "Direct" in Fig. 1(d), (e) and (f)). For example, when migrating from mechanical spinning LiDAR to a more inexpensive and compact solid-state LiDAR, the cost of annotating new sensor data to re-train detectors is prohibitive. Cross-mechanism domain adaptation is a more efficient yet promising solution with rigid demand to transfer not only detectors/models but also all related assets to new sensors. + +Meanwhile, prevalent domain adaptation research [19, 26, 35, 36, 45, 46] finds it hard to quantitatively analyze the pivotal sensor factor. This challenge arises because they have been directed at cross-dataset settings (e.g. domain adaptation between KITTI, Waymo, and nuScenes), covering differences in sensors, geographic location, climate, and so on. Therefore, they have not shown optimal performance when applied to cross-mechanism domain adaptation problems (see Fig. 1(d)(e)(f)). The reason lies in the fact that there is a lack of datasets that contain comprehensive sensors to fully decouple the domain disparities caused by locations and sensors. Several inspiring datasets, encompassing various common sensors (e.g., PandaSet [44], KRadar [22], VoD [23], LiDAR-CS [11]), often fall short in providing comprehensive modality coverage. This limitation, in turn, restricts the design of methods to bridge the sensor-based domain gap. + +To address this problem, this paper introduces a Cross Mechanism domain adaptation Dataset (CMD) for 3D object detection. To the best of our knowledge, CMD is the first domain-adaptation dataset in 3D object detection that contains a comprehensive suite of sensors, including: (1) 32-beam low-resolution and 128-beam high-resolution mechanical spinning (MS) LiDAR, (2) automotive-grade solid-state LiDAR, (3) $4D$ millimeter-wave radar, and (4) camera. This combination provides the most extensive modality coverage. All sensors are time- + +synchronized with high accuracy under 1 ms, ensuring that different modalities can capture and model the same scene accurately, even with highly dynamic objects. Furthermore, the CMD comprises 50 sequences. Each sequence has a time span of 20 seconds, with each sensor capturing 10 frames per second, totaling 10,000 frames of data per sensor. In addition, we meticulously annotated 3D objects of 13 categories based on a multi-sensor collaborative annotation system subjected to multiple rounds of human inspection. + +![](images/e86365b3bc77bba7497ae8073fc1771fd7918d9c880027d8034a9b8d71e20e0f.jpg) + +![](images/5c1e820d28cb03e37f134c29a2ac3688029ad1cb9f411692cd8d4b1fa8e65779.jpg) + +![](images/1c36901429a9508c4a44db8570f96aab853dd38886b243a65eff8a07bfdf79e1.jpg) + +![](images/19648560fae7813fe06a6fcfcef2539e8c87848e05f1ab1b6d8157f71c6502ed.jpg) +Fig. 1: (a) The points number distribution of different sensors. (b) The intensity distribution of different sensors. (c) Point clouds of the same instance for different sensors. The results show that data acquired by different sensors exhibit substantial differences in point density, intensity, and geometry. Subfigure (d), (e), and (f) show the cross-mechanism detection results of different domain adaptation methods on our CMD. Our DIG outperforms the traditional ST3D by a large margin. + +![](images/1087e7ea9d5585bddc4069b58cac5c0feeb8839d2db066bde509b5b6d5110690.jpg) + +![](images/38126f93d0f81ca5d3b6f56dcf5da01c40fdb763952c287894302212856fc139.jpg) + +Using the CMD, we investigate the cross-mechanism domain adaptation issues and summarize the gaps into three primary components: Density gap, Intensity gap, and Geometry gap. Subsequently, we introduce a simple yet effective baseline method, named DIG, to address each of these gaps systematically. (1) Density gap: We propose the Beam-Distance Down-sampling (BDS) that enforces a similar data pattern in terms of beam numbers and density distribution along different detection distances. (2) Intensity gap: We introduce the Box-Cox log Normalization (BCN) that initially transforms point clouds from + +different modalities into near-normal distributions and then normalizes them by logarithms. (3) Geometry gap: We design the Geometry-Aware label Mixing (GAM) that mixes the pseudo-labels from the target domain to narrow down the geometry difference in training. By implementing these three components, DIG achieves optimal results in bridging the identified domain gaps. Our main contributions are as follows. + +- We present the Cross Mechanism Dataset (CMD), the first domain adaptation dataset that contains a comprehensive suite of LiDAR/4D Radar sensors. CMD is a key piece of the puzzle of cross-mechanism detection as existing public datasets can not adequately evaluate detection performance drops caused by cross-domain disparities due to limited sensor variety. +- We construct comprehensive 3D cross-sensor detection baselines by benchmarking the results of SOTA domain adaptation algorithms on CMD. +- We provide an in-depth analysis of the key factors leading to cross-sensor domain disparities, i.e. Density, Intensity, and Geometry gaps. Based on this, we propose a novel DIG method, which shows plausible performance on cross-sensor detection. + +# 2 Related Work + +# 2.1 Datasets for 3D Object Detection + +With increasing research on 3D object detection [7, 8, 17, 40, 42, 43, 47], more datasets have emerged. Widely used datasets such as KITTI [13], Waymo [30], nuScenes [5], and ONCE [21] support conventional 3D object detection research and facilitate domain adaptation across different geographic locations. However, focusing solely on geographical domain differences is inadequate for addressing the domain adaptation problem in detection algorithms; sensor variability is also crucial. To address this, datasets like PandaSet [44], K-Radar [22], TJ4DRadSet [52], VoD [23], Lyft L5 [15], and Cirrus [37] offer diverse sensors, scan beams, locations, and weather conditions. Despite their contributions, these datasets are limited in sensor diversity within individual collections. In contrast, our CMD includes extensive image and point cloud data annotated for detection and tracking, encompassing five different sensor types, thus providing a comprehensive evaluation. Statistical comparisons with other 3D object detection datasets are shown in Table 2. + +# 2.2 Domain Adaption in 3D Object Detection + +To mitigate the poor generalization of 3D object detectors to unknown data, recent methods have focused on adapting domain gaps [10, 14, 32-34, 48, 50, 51]. Among these methods, some attempt to introduce a teacher-student framework. ST3D [45] and $\mathrm{ST3D + + }$ [46] design label assignment strategy and pseudo-labels denoising to enable adaptation between domains. MLC-Net [19] employs consistency for cross-domain transfer. Another statistics-based approach adapts distribution gaps, with SN [35] correcting car size distribution gaps. GBA [26] + +adapts label distribution gap via a Gradual Batch Alternation training strategy. And, SSDA3D [36] adapting point cloud distributions through Inter-domain Alignment module. Further methods address sensor gaps, with LiDAR Distillation [38] and DTS [16] proposing progressive and density-insensitive frameworks to mitigate beam-induce gaps from different sensors, and CL3D [24] using spatial geometry alignment to adapt geometric gaps between sensors. + +In general, current domain adaptation research methods still primarily focus on geographical information and sensor beam migration. Due to dataset limitations, domain adaptation methods for multiple types of sensors have yet to be seen. Therefore, it is necessary to propose a novel, well-annotated dataset including various types of sensors to advance this research field. + +# 3 Cross Mechanism Dataset + +# 3.1 Sensor Setup + +Sensor specifications. Our CMD comprises data from seven sensors of five types: a 128-beams mechanical scanning (MS) LiDAR, a 32-beams MS LiDAR, a solid-state LiDAR, a 4D millimeter-wave radar, and three cameras. Detailed specifications for our sensors are shown in Table 1. The sensors in our dataset, each with unique beam configurations or modeling modes, are collectively referred to as mechanisms. As shown in Table 2, in contrast to other datasets, our CMD encompasses a diverse array of commonly employed sensors, establishing it as the most comprehensive dataset presently available. + +
SensorTypeHFOV(°)VFOV(°)ResolutionFPS
OS128128 beams MS LiDAR360[-22.5, 22.5]128 * 102410
XT3232 beams MS LiDAR360[-16, 15]32 * 204810
M1SS LiDAR[-60, 60][-12.5, 12.5]126 * 62510
Radar4D Radar[-75, 75][-15, 15]>60010
HIK Cam.Camera[-31.2, 31.2][-27.7, 27.7]1080 * 192010
+ +Table 1: Sensor specifications for CMD. MS and SS denote mechanical spinning and solid-state respectively. FPS refers to frames per second. + +Sensor layout. In our CMD, all sensors are oriented in the same direction as the vehicle's forward direction. Such a configuration is instrumental in ensuring that the most critical information is captured within the area of greatest overlap of FOV across different sensors. To achieve this alignment, a rigid bracket has been designed, aligning all sensors along a vertically aligned axis. We have deliberately placed all sensors as close to each other as possible to optimize FOV overlap. In addition to this, a camera system comprising three evenly spaced cameras has been integrated. This system spans a $150^{\circ}$ horizontal field of view (HFOV), as depicted in Fig. 2(c). Such a configuration is pivotal + +for ensuring data consistency across different mechanisms. The overall FOV is shown in Fig. 2(b). + +![](images/b9e908acc9ce7a128336dd9367be8276b438db5a1b413d096b9fe2265c97f61b.jpg) +Fig. 2: The illustration of sensor layout. (a) demonstrates the specific layout of each sensor; (b) showcases the field of view angles for each sensor; (c) demonstrates the positions of each sensor along the z-axis. + +![](images/bbabdac9b0b734d763543ec6b3060bea56092ed4a1bb913dad5f0f8ddcb62e28.jpg) + +![](images/dc3843b9a3bd9db89f83b0e7b647f6b5610299d6b5647aa2d7f8816bf10e650b.jpg) + +![](images/6122c5a37371835274ea7fe115519b4181b96137aa84d7212366b417a0e500e5.jpg) + +![](images/42a273595bb090827ed95f7186f5bcbcc47e244ab382add5dae2697ef8397307.jpg) + +# 3.2 Synchronization and Calibration + +Synchronization and trigger. To achieve synchronization among different sensors when modeling the world, we utilized the Precision Time Protocol Version 2 [1] to establish a time synchronization system with an error margin within $1\mathrm{ms}$ . Specifically, we employed the CoolShark AUTO 66 unit as the PTP master clock, while other sensors and the host served as PTP slave clocks. Once synchronized, M1 is set to generate a frame every $100\mathrm{ms}$ , and both OS128 and XT32 generate frames simultaneously using phase-locking. Meanwhile, the radar is triggered by a network signal sent from the host. Due to the differing principles of data acquisition between cameras and LiDAR, the cameras are triggered by OS128. Whenever OS128 scans to $0^{\circ}$ , it sends out a trigger signal. This signal undergoes a delay process via a microcontroller before being fed into the cameras. For each camera, $T_{delay} = \frac{\theta_{cam}}{2\pi} - \frac{1}{2} T_{exposure}$ where $T_{exposure}$ means the exposure time for camera and $\theta_{cam}$ means the angle at which the camera is positioned. By this, the time when OS128 scans the central of the camera's HFOV, is the central of its exposure period, which ensures a lower frame time offset between the cameras and the LiDARs. This low-latency synchronization ensures that the same frame of data captured by different sensors effectively reflects their respective mechanism differences. + +Calibration. Achieving high-quality data in a multi-sensor setup relies heavily on calibration. (1) For the LiDARs (i.e. OS128, M1, XT32) and radar, we calibrate their extrinsic parameters using Generalized-ICP [28]. Several corner reflectors in the space are used to obtain more precise corner points for radar calibration. (2) For the cameras, we calibrate their intrinsic parameters and extrinsic parameters with MATLAB Toolkit [12] and OpenCV [4]. The results of calibration are shown in Fig. 4. Subsequent to calibration, we established a well-defined coordinate system, as illustrated in Fig. 2(c). We transfer XT32 and + +radar to the coordinate system of OS128 and move this coordinate system vertically down to the ground to get the ego system. The ego system has the x-axis pointing forward, the y-axis to the left, and the z-axis upward. The camera coordinate system, in 2D, aligns the x-axis with image width and y-axis with height, starting from the top-left corner. + +# 3.3 Annotation + +Annotation area. With precise time synchronization and calibration, we can jointly annotate data from multiple sensors, improving accuracy, especially for distant objects. We annotate objects within $160m$ and within the HFOV of M1. We believe that the data within this range is highly representative. + +Annotation rules. There are 13 annotated categories, namely: Car, Van, Bus, Truck, Semi-Trailer towing vehicle, Special Vehicles, Motorcycle, Bicycle, Tricycle, Adult Pedestrian, Children Pedestrian, Animal, and Barrier. We labeled each object as a 9-DoF 3D bounding box $(x,y,z,l,w,h,\theta_x,\theta_y,\theta_z)$ . Where $x,y,z$ represents the center coordinates, $l,w,h$ denotes length, width height, and $\theta_x,\theta_y,\theta_z$ are the rotation angles around the x,y,z axis. Additionally, we provide a motion state and tracking ID for all objects. Occlusion situations are also provided. More annotation specifications are provided in the appendix. + +
DatasetMechanismMS LiDAR(beams)SS LiDAR 4DRadar Cam.Anno.Frames TrackingIlluminationReal
LowHigh
KITTI Det. [13]X64XX✓15KX✓✓
Waymo [30]X64XX✓230K✓✓✓
nuScenes [5]Single32XXX✓40K✓✓✓
Argoverse [6]32*2XXX✓44K✓✓✓
Lyft L5 [15]40*264XX✓30K✓X✓
K-Radar [22]X64, 128X✓✓35K✓✓✓
Lidar-CS [11]16, 3264, 128✓XX14KXXX
PandaSet [44]MultiX64✓X✓8.2K✓✓✓
TJ4DRadSet [52]X64X✓✓7.8K✓✓✓
VoD [23]X64X✓✓8.7KXX✓
CMD32128✓✓✓10K✓✓✓
+ +Table 2: Comparison with existing datasets. "Mechanism" means whether the dataset contains multiple mechanisms of 3D sensor. "Tracking" signifies the relevance of annotations for tracking. "Illumination" indicate the inclusion of various light conditions. "Real" refers to whether the dataset is real or synthetic. + +# 3.4 Dataset Analytic + +Diversity of scenes. In CMD, scenes are carefully selected for wide coverage, primarily encompassing urban areas, suburbs, campuses, bridges, and tunnels with different illuminations (see Fig. 3(a)). We select 50 high-quality sequences, each spanning 20 seconds, equating to 200 frames per sensor, culminating in a dataset of 40,000 frames of point cloud data and 30,000 frames of image + +data. This comprehensive collection offers a rich resource for research on cross-mechanism domain adaptation. These 50 sequences are evenly divided into 30, 10 and 10 for training, validation, and testing according to different scenarios. This arrangement ensures a balanced distribution of data across environments, crucial for training reliable and generalizable 3D object detection models. + +![](images/6f5eb51ed399a4c404f634a84090c5285db2626f35e942401b3654d923c4b842.jpg) +Cross Mechanism domain adaptation Dataset (CMD) for 3D Object Detection + +![](images/9b1c6c08041cd325ed272d6799b9d9874b8cfdbc9cadbf18bb017372efc3839e.jpg) + +![](images/21ce536fbb4f5e2d3e9da15e4cad9723fdab2b134f07aee08105e37988666d36.jpg) +Fig. 3: The scene and annotation details. (a) The upper section showcases the distribution of road types, while the lower section demonstrates the distribution across various illumination types; (b) presents the distribution of objects' numbers for different categories; (c) presents the distribution of objects' numbers along different distances. + +Distribution of annotations. We annotated approximately 230,000 3D bounding boxes based on the rules described in Section 3.3. The distribution of objects' numbers for different categories is presented in Fig. 3(b). The distribution of objects' numbers along different distances is shown in Fig. 3(c). Among these categories, the most concerned "Car" has the highest annotation count, reaching approximately 92,000 instances. + +Comparisons between mechanisms. The average number of points for these four sensors is quite different. From most to least, they are M1, OS128, XT32, and 4D radar. For a better understanding, we show the visualization results in Fig. 4. Point clouds from sensors with different mechanisms exhibit substantial differences. (1) OS128 and XT32 both adopt a mechanism spinning style, clearly modeling the rigid vehicles with minimal vertical distortion. The main difference between them is density. (2) Due to the use of a zigzag scan pattern, M1 tends to introduce a bit of horizontal displacement. (3) Point clouds from radar sensors are generally sparse and inaccurate, posing challenges in accurately distinguishing the geometric shapes of objects. + +![](images/1bda9ae537febd463216a986dabd09069ba31d81a3a8735a05988594fd340c5e.jpg) +Fig. 4: Different mechanisms and their sample data. The left side of the image denotes the various scenes of data collection, while the bottom indicates the types of sensors involved. + +# 3.5 DIG Baseline Method + +In this section, we propose DIG (Density-Intensity-Geometry), a streamlined and potent baseline for cross-mechanism domain adaptation. It consists of three key components: Box-Cox Intensity Log Normalization (BCN), Beam-Distance Down Sampling (BDS), and Geometry-Aware Label Mixing (GAM). + +Box-Cox intensity log Normalization (BCN). The distribution of intensity for point clouds may significantly differ across various mechanisms. See Fig. 1(b). Recent studies have either opted to discard [45] [38] intensity or straightforwardly normalized it to [0,1] [36]. However, it is observed that intensity plays a crucial role in detecting foreground objects across diverse domains. We contend that effectively incorporating intensity can lead to enhanced performance in domain adaptation. Motivated by this insight, we introduce the BCN to mitigate disparities in intensity. For both intensities (denoted as $i$ ) for source and target domain point clouds, we initially take the logarithm to address biases stemming from different sensor definitions of intensity ranges. Subsequently, we apply the Box-Cox transformation to enhance their normality and homoscedasticity, respectively. This process can be formulated as: + +$$ +i (\lambda) = \frac {\log (i + 1) ^ {\lambda} - 1}{\lambda}, \tag {1} +$$ + +where $\lambda$ refers to the box-cox transfer parameter [2]. It's different for source and target domains, and remains constant respectively. After being normalized to + +[0,1], the transformed intensity information exhibits a good level of similarity between the source and target domains, thus narrowing the domain gap. + +Beam-Distance based down Sampling (BDS). As shown in Fig. 1(a), the density distributions of points vary remarkably among different mechanisms. We observe that, within a relatively short distance, domain adaptation for different mechanisms is more significantly impacted by variations in point density. However, when it comes to longer distances, the point cloud is excessively sparse, which has a greater impact on object detection accuracy. To deal with this problem, we introduce the BDS that assigns different sampling probabilities based on both beam and distances. Denote $P_{\text{keep}}(p)$ as the probability of a point (p) being retained, and it can be formulated as: + +$$ +P _ {\text {k e e p}} (p) = 1 - \mathbf {1} \left\{\text {b e a m} (p) \notin \mathcal {B} \right\} \exp \left(- \frac {1}{K} | | p | |\right), \tag {2} +$$ + +where $K$ controls the rate of probability increase, and we set it to 35 empirically. $\mathcal{B}$ means the set of target beam indexes that are closest to the source domain. Function $\text{beam}(p)$ denotes the index of the beam that point $p$ belongs to. If this beam information is not directly available, an alternative approach involves calculating the angle of each point, denoted as $\theta(p) = \arctan \left( \frac{z}{\sqrt{x^2 + y^2}} \right)$ . This angle can then be used to apply a clustering algorithm, such as K-means [20], to approximate the beam index for each point. + +Geometry-Aware label Mixing (GAM). Point clouds captured from the same object by diverse sensors frequently exhibit notable geometric variances. To tackle this challenge, we introduce the Geometry-Aware Mix (GAM) module. In our approach, we initially utilize a model trained on the source domain to generate pseudo-labels within the target domain. Subsequently, instances from the target domain with high-confidence pseudo-labels are integrated back into the source domain for further training. Inspired by GBA [26], our second training phase progressively reduces source domain labels while maintaining a constant number of target domain pseudo-labels. This technique aims to gradually adapt the model to the geometric characteristics of the target domain. Through this method, the model becomes more adept at learning and adapting to the unique data characteristics of the target domain, thereby improving its ability for cross-mechanism domain adaptation. + +# 4 Experimental Result + +# 4.1 Experiment Setup + +All of our experiments were conducted using four Nvidia 3090 GPUs and the open-source code repository OpenPCDet [31]. The Voxel-RCNN [9] served as the detector for most experiments. The batch size was set to 32, and the learning rate was 0.003. Similar to KITTI [13], the grid range was defined as $[0\mathrm{m},70.4\mathrm{m}]$ , $[-70.4\mathrm{m},70.4\mathrm{m}]$ , and $[-2\mathrm{m},4\mathrm{m}]$ along the x, y, and z axis respectively. Point cloud data were cropped to HFOV $[-60^{\circ},60^{\circ}]$ with horizontal distances smaller than $70.4\mathrm{m}$ . Voxel size was set to $[0.1,0.1,0.15]$ along the x, y, and z axis respectively. + +![](images/01118e19813a7984f94f896e43790c8b47e4e5b661ca8a7eaa6bd77d437286f2.jpg) + +![](images/07652d9d90f9f1cbf22863e6755a61f842a8a432a61681ded5a8b8e05cce1f32.jpg) + +Fig. 5: Similarities and differences among baselines. ST3D [45] leverage random object scaling, generating and denoising pseudo-labels. DTS [16] down sample point clouds before training source-detector. Our DIG jointly downs ample point clouds, transforms intensity, and makes use of pseudo-labels. +![](images/ea1fd491f024a4bc8e5c89d75d71b4f7bf345ce1262e684bb71b903768ab0c94.jpg) +$\pmb{X}^{S}$ Points in source domain +$\mathcal{F}(\theta^s)$ Model in source domain +$X^t$ Points in target domain +$\mathcal{F}(\theta^t)$ Model in target domain + +![](images/d56d7b09aeb8f7155d7fb0927d10e3d0468f3010a2422009f71b0f78d8b966f5.jpg) +$\varphi^a$ Points down sampling +$\varphi^i$ Intensity normalization + +# 4.2 Evaluation Metric + +Mean Average Precision (mAP). mAP is a commonly used object detection evaluation metric. Here when calculating mAP, we use 3D IoU thresholds of 0.5, 0.5, 0.25, and 0.25 for Car, Truck, Pedestrian, and Cyclist, respectively. Following the ONCE [21] dataset, we calculated an orientation-aware AP on 50 precision steps. The AP among different detection distances (e.g., overall and 0 to $30\mathrm{m}$ ) are also computed for detailed analysis. Formally, the AP is defined as: + +$$ +A P = 1 0 0 \int_ {0} ^ {1} \max \left\{p \left(r ^ {\prime} \mid r ^ {\prime} \geq r\right) \right\} d r, \tag {3} +$$ + +where $p(r)$ refers to the precision-recall curve that calculated by 50 recall positions. The mAP is the average of AP across all categories. + +Mean Closed Gap (mCG). Directly applying AP or mAP as evaluation metrics for domain adaptation is not intuitive since different sensors do not share the same domain gap. Closed gap [35] is used for measuring the effectiveness of a method on a single domain adaptation task. CG can be formulated as: + +$$ +C G = 1 0 0 \frac {m A P _ {\text {m o d e l}} - m A P _ {D T}}{m A P _ {\text {O r a c l e}} - m A P _ {D T}}, \tag {4} +$$ + +where $DT$ means direct transfer using a model trained sorely on source domain data. Oracle implies results obtained by the full training on target domain. To quantify the universality of domain adaptation methods, we define a new metric named mean closed gap (mCG), formulated as: + +$$ +m C G = \sum_ {s \in \mathbf {M}} \sum_ {t \in \mathbf {M}} \mathbf {1} (s \neq t) C G _ {s \rightarrow t}, \tag {5} +$$ + +where $\mathbf{M}$ refers to the set of sensors, including OS128, XT32, and M1 in CMD. + +
MethodmCGCG
M1→OS128XT32→OS128OS128→M1XT32→M1OS128→XT32M1→XT32
ST3D [45]-16.72-9.2734.00-7.95-5.07-43.39-68.61
ST3D++ [46]09.2312.6554.4902.9917.2307.56-39.54
DTS [16]29.2248.2822.9209.0207.0934.5053.51
DIG(Ours)42.8954.4155.0933.6925.2249.8739.06
+ +Table 3: Overall mean closed gap for baseline methods. + +Discussion. Currently, commonly used CG is tailored for a single task, usually only considering the "Car" category [16], and cannot reflect the comprehensive performance of methods across various domain adaptation tasks. In contrast, mCG includes all MS LiDARs with varying beam numbers and SS LiDARs (i.e., OS128, XT32, M1), measuring performance from six experimental groups, thus providing the most comprehensive evaluation of cross-mechanism domain adaptation. See the appendix for 4D radar results. + +# 4.3 Results and Benchmark + +Domain adaptation baselines. We selected four baseline methods. Fig. 5 illustrates the similarities and differences between them. (1) Direct transfer (DT) means directly applying the source model to the target dataset. (2) ST3D [45] and $\mathrm{ST3D + + }$ [46] share a similar teacher-student architecture. The source model is trained with random object scaling, then used to generate pseudo-labels. The difference is that $\mathrm{ST3D + + }$ comes with a more effective denoising method to obtain more high-quality pseudo-labels. (3) DTS [16] introduced Random Beam Re-Sampling (RBRS) to enhance the robustness to varying beam densities. We reproduced the RBRS on CMD, providing representative baseline results. (4) Our proposed DIG. + +Overall comparison results. Table 3 shows the overall results. Since the ST3D and $\mathrm{ST3D + + }$ are mostly designed for cross-geographical domain adaptation that highly relies on pseudo-label quality, the two detectors show lower performance in our cross-mechanism dataset. DTS takes density into consideration, thus outperforming $\mathrm{ST3D + + }$ by 19.99. Our DIG outperformed all previous methods by a large margin. Its mCG reached 42.89, surpassing ST3D, $\mathrm{ST3D + + }$ , and DTS by 59.61, 33.66, and 13.67 respectively, as our method considers all the typical domain disparities in the cross-mechanism domain adaptation problem. + +Performance discussion. A more detailed composition of mCG is shown in Table 4. We can get several conclusions from these results. (1) DIG exhibits notable superiority in the categories of pedestrians and bicycles, both across all distances and within a detection range of 30 meters. The similarity in these categories is the fewer points within the targets. We believe that sparse points make intensity the key factor for adapting to different domains. DIG's BCN module + +
TaskMethodCGmAPCarTruckPedCyc
OS128Oracle10026.4436.50/69.3417.45/38.5118.79/38.9933.0/69.22
XT32↓OS128DT006.2911.76/20.5904.61/08.2301.54/02.1507.25/08.90
ST3D [45]34.0013.1422.78/47.3910.41/15.1006.83/16.0112.55/33.05
ST3D++ [46]54.4917.2727.87/61.0213.50/22.6108.57/19.7618.95/49.32
DTS [16]22.9210.9114.26/49.4306.72/27.6709.29/20.9413.37/41.21
DIG(Ours)55.0917.3922.73/60.5913.60/30.1611.56/25.6721.70/57.46
M1↓OS128DT006.9213.23/30.2105.08/07.8601.98/03.6107.38/09.98
ST3D [45]-9.2705.1110.24/39.1104.06/11.3200.00/00.0006.14/21.18
ST3D++ [46]12.5609.3913.46/46.9306.61/23.1605.85/13.6811.65/35.06
DTS [16]48.2516.3421.60/58.9912.67/34.1409.57/23.4921.51/58.70
DIG(Ours)54.4117.5423.03/57.0809.71/25.9509.71/25.9525.00/66.13
M1Oracle10030.0942.01/70.6419.96/40.0117.54/36.9940.84/68.49
XT32↓M1DT013.3220.33/42.8209.47/19.8404.81/11.4718.67/36.06
ST3D [45]-5.0712.4730.34/55.4104.20/03.8200.19/00.2215.13/25.95
ST3D++ [46]17.2316.2132.92/61.0705.24/06.0402.18/04.1724.48/49.39
DTS [16]07.0914.5123.26/56.9408.78/28.0206.99/14.7219.01/48.44
DIG(Ours)25.2217.5527.19/64.7009.26/22.7905.72/12.5728.04/57.34
OS128↓M1DT014.3627.07/56.5409.91/24.2501.13/03.8119.32/33.49
ST3D [45]-7.9513.1128.71/60.3708.26/07.2506.37/17.4809.08/12.72
ST3D++ [46]02.9914.8332.98/68.4506.63/08.4507.28/19.4612.43/17.89
DTS [16]09.0215.7822.11/54.0111.84/30.2611.88/28.5817.27/40.40
DIG(Ours)33.6919.6626.85/62.8911.86/28.1113.69/32.2426.19/60.40
XT32Oracle10026.5936.84/72.4719.26/41.0116.23/32.7434.01/66.59
M1↓XT32DT014.0722.10/59.8610.44/21.4806.29/15.2417.44/45.69
ST3D [45]-68.6105.4811.39/38.0502.92/06.7800.00/00.0002.92/06.78
ST3D++ [46]-39.5409.1214.56/47.9407.47/22.0703.15/05.0211.28/32.18
DTS [16]53.5120.7730.62/68.0315.67/33.0109.48/21.6327.33/57.48
DIG(Ours)39.0618.9625.95/68.7814.77/34.6511.45/26.4723.66/60.92
OS128↓XT32DT011.7722.25/58.2008.45/19.1601.66/04.5014.71/41.05
ST3D [45]-43.9905.3411.46/45.4101.72/05.8700.00/00.0008.18/22.82
ST3D++ [46]07.5612.8917.81/56.9905.31/19.9810.69/24.2217.75/48.29
DTS [16]34.5016.0918.87/59.7813.63/27.7011.26/25.4920.59/58.19
DIG(Ours)49.8719.1627.20/63.8910.75/26.7911.76/27.1626.91/64.22
+ +Table 4: Detailed experimental results for all cross-mechanism settings. For each category, we provide the Average Precision (AP) results for all distances and within a range of 30 meters. + +aligns intensity well, which explains its strong performance in these cases. (2) The results indicate a significant decrease in mAP while using OS128 as the target domain. This decrease is primarily attributed to its unique intensity range of [0, 512], unlike the [0, 256] range of other sensors. The pre-trained model finds it challenging to handle unfamiliar intensities. Unlike the other three methods, which either discarded the intensity information or did not include a dedicated module for it, we addressed this issue with BCN, resulting in a significant performance improvement. (3) All these methods perform unsatisfactorily for domain adaptation from XT32 to M1. Even DIG, the best one among them, gets CGs of 25.22 on it. This observation means that M1 and XT32 may suffer from larger domain gaps, highlighting the need for further investigation in the future. (4) ST3D++ gets a CG of 54.49 for XT32 to OS128, which is mainly due to the + +great performance on APs for 0-30m. As they are both MS LiDAR, they suffer more from differences in density. Pseudo-labels from XT32 might look like harder cases for OS128. Refining model parameters with these cases can dramatically improve the performance. (5) It indicates that DTS plays a crucial role when there is a substantial difference in the number of beams. CG on tasks that contain XT32 gets nice performance improvement. Specifically, CG can reach 34.51 and 53.51 for OS128 to XT32 and M1 to XT32 respectively. + +
modulesM1 ⇒ OS128OS128 ⇒ XT32
BDS BCN GAMmAPCGmAPCG
✓09.2211.7814.6919.70
✓17.1452.5316.4731.71
✓17.4754.0418.9548.45
✓17.5454.4119.1649.87
+ +Table 5: Ablation study for DIG baseline method. + +Ablation study for DIG. We conduct ablation studies to analyze the effectiveness of each component in the DIG baseline method. Two representative tasks are selected. $M1 \Rightarrow OS128$ and $OS128 \Rightarrow XT32$ respectively refer to domain adaptation for sensors with different scanning patterns and different beams/densities. As shown in Table 5, For both tasks, BCN is anticipated to be the most effective module, contributing 52.53 and 31.71 CG on each respective task. This is attributed to the substantial intensity differences between the source and target domains in both tasks. BDS also exhibits commendable performance in both tasks, achieving 11.78 and 19.70 CG, respectively. Although less conspicuous, GAM proves to be effective for both tasks. + +# 5 Conclusion + +We presented a dataset specifically designed for cross-mechanism domain adaptation, incorporating mechanical LiDARs with both high-res and low-res beams, solid-state LiDAR, and 4D millimeter-wave radar. These sensors undergo precise time synchronization, allowing them to simultaneously generate point clouds of the same external environment. Data from diverse scenes were selected and meticulously annotated. We believe the proposed dataset would hugely facilitate research of cross-mechanism 3D detection, as it currently stands as the most comprehensive point cloud 3D detection dataset in terms of sensor types. We also provide a novel DIG method as well as complete experimental results of recent methods on our dataset. They can be used as reliable baselines to further benefit the research communities. + +In the future, we will continue to explore more usage of CMD, including utilizing data from multiple mechanisms as the source domain to enhance domain adaptation effects, and employing CMD as a tool for validating the performance of multi-modal 3D object detection. + +# References + +1. IEEE standard for a precision clock synchronization protocol for networked measurement and control systems. IEEE Std 1588-2008 (Revision of IEEE Std 1588-2002) pp. 1-269 (2008). https://doi.org/10.1109/IEEESTD.2008.4579760 +2. Atkinson, A.C., Riani, M., Corbellini, A.: The box-cox transformation: Review and extensions (2021) +3. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR) (2022) +4. Bradski, G.: The opencv library. Dr. Dobb's Journal: Software Tools for the Professional Programmer 25(11), 120-123 (2000) +5. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (Jun 2020). https://doi.org/10.1109/cvpr42600.2020.01164, http://dx.doi.org/10.1109/cvpr42600.2020.01164 +6. Chang, M.F., Lambert, J., Sangkloy, P., Singh, J., Bak, S., Hartnett, A., Wang, D., Carr, P., Lucey, S., Ramanan, D., et al.: Argoverse: 3d tracking and forecasting with rich maps. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8748-8757 (2019) +7. Chen, Y., Yu, Z., Chen, Y., Lan, S., Anandkumar, A., Jia, J., Alvarez, J.M.: Focalformer3d: Focusing on hard instance for 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8394-8405 (2023) +8. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21674-21683 (2023) +9. Deng, J., Shi, S., Li, P., gang Zhou, W., Zhang, Y., Li, H.: Voxel r-cnn: Towards high performance voxel-based 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence (2021) +10. Ding, G., Zhang, M., Li, E., Hao, Q.: Jst: Joint self-training for unsupervised domain adaptation on 2d&3d object detection. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 477-483. IEEE (2022) +1. Fang, J., Zhou, D., Zhao, J., Tang, C., Xu, C.Z., Zhang, L.: Lidar-cs dataset: Lidar point cloud dataset with cross-sensors for 3d object detection (Jan 2023) +2. Fetic, A., Juric, D., Osmankovic, D.: The procedure of a camera calibration using camera calibration toolbox for matlab. In: 2012 Proceedings of the 35th International Convention MIPRO. pp. 1752-1757. IEEE (2012) +3. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition (Jun 2012). https://doi.org/10.1109/cvpr.2012.6248074, http://dx.doi.org/10.1109/cvpr.2012.6248074 +4. Hegde, D., Sindagi, V., Kilic, V., Cooper, A.B., Foster, M., Patel, V.: Uncertainty-aware mean teacher for source-free unsupervised domain adaptive 3d object detection. arXiv preprint arXiv:2109.14651 (2021) +5. Houston, J., Zuidhof, G., Bergamini, L., Ye, Y., Chen, L., Jain, A., Omari, S., Iglovikov, V., Ondruska, P.: One thousand and one hours: Self-driving motion prediction dataset. In: Conference on Robot Learning. pp. 409-418. PMLR (2021) + +16. Hu, Q., Liu, D., Hu, W.: Density-insensitive unsupervised domain adaption on 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17556-17566 (2023) +17. Huang, X., Wu, H., Li, X., Fan, X., Wen, C., Wang, C.: Sunshine to rainstorm: Cross-weather knowledge distillation for robust 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 2409–2416 (2024) +18. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. ArXiv (2022) +19. Luo, Z., Cai, Z., Zhou, C., Zhang, G., Zhao, H., Yi, S., Lu, S., Li, H., Zhang, S., Liu, Z.: Unsupervised domain adaptive 3d detection with multi-level consistency. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8866-8875 (2021) +20. MacQueen, J., et al.: Some methods for classification and analysis of multivariate observations. In: Proceedings of the fifth Berkeley symposium on mathematical statistics and probability. vol. 1, pp. 281-297. Oakland, CA, USA (1967) +21. Mao, J., Niu, M., Jiang, C., Liang, H., Liang, X., Li, Y., Ye, C., Zhang, W., Li, Z., Yu, J., Xu, H., Xu, C.: One million scenes for autonomous driving: Once dataset. Cornell University - arXiv, Cornell University - arXiv (Jun 2021) +22. Paek, D.H., Kong, S.H., Wijaya, K.: K-radar: 4d radar object detection dataset and benchmark for autonomous driving in various weather conditions (Jun 2022) +23. Palffy, A., Pool, E., Baratam, S., Kooij, J., Gavrila, D.: Multi-class road user detection with 3+1d radar in the view-of-delft dataset +24. Peng, X., Zhu, X., Ma, Y.: Cl3d: Unsupervised domain adaptation for cross-lidar 3d detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2047-2055 (2023) +25. Raj, T., Hanim Hashim, F., Baseri Huddin, A., Ibrahim, M.F., Hussain, A.: A survey on lidar scanning mechanisms. Electronics 9(5), 741 (2020) +26. Rochan, M., Chen, X., Grandhi, A., Corral-Soto, E.R., Liu, B.: Domain adaptation in 3d object detection with gradual batch alternation training. arXiv preprint arXiv:2210.10180 (2022) +27. Roriz, R., Cabral, J., Gomes, T.: Automotive lidar technology: A survey. IEEE Transactions on Intelligent Transportation Systems 23(7), 6282-6297 (2021) +28. Segal, A., Haehnel, D., Thrun, S.: Generalized-icp. In: Robotics: science and systems. vol. 2, p. 435. Seattle, WA (2009) +29. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR). pp. 10526 - 10535 (2020) +30. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (Jun 2020). https://doi.org/10.1109/cvpr42600.2020.00252, http://dx.doi.org/10.1109/cvpr42600.2020.00252 +31. Team, O.D.: Openpcdet: An open-source toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet (2020) +32. Tsai, D., Berrio, J.S., Shan, M., Nebot, E., Worrall, S.: Ms3d: Leveraging multiple detectors for unsupervised domain adaptation in 3d object detection. arXiv preprint arXiv:2304.02431 (2023) + +33. Tsai, D., Berrio, J.S., Shan, M., Nebot, E., Worrall, S.: Viewer-centred surface completion for unsupervised domain adaptation in 3d object detection. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 9346-9353. IEEE (2023) +34. Tsai, D., Berrio, J.S., Shan, M., Worrall, S., Nebot, E.: See eye to eye: A lidar-agnostic 3d detection framework for unsupervised multi-target domain adaptation. IEEE Robotics and Automation Letters 7(3), 7904-7911 (2022) +35. Wang, Y., Chen, X., You, Y., Li, L.E., Hariharan, B., Campbell, M., Weinberger, K.Q., Chao, W.L.: Train in germany, test in the usa: Making 3d object detectors generalize. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11713-11723 (2020) +36. Wang, Y., Yin, J., Li, W., Frossard, P., Yang, R., Shen, J.: Ssa3d: Semi-supervised domain adaptation for 3d object detection from point cloud. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2707-2715 (2023) +37. Wang, Z., Ding, S., Li, Y., Fenn, J., Roychowdhury, S., Wallin, A., Martin, L., Ryvola, S., Sapiro, G., Qiu, Q.: Cirrus: A long-range bi-pattern lidar dataset. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 5744-5750. IEEE (2021) +38. Wei, Y., Wei, Z., Rao, Y., Li, J., Zhou, J., Lu, J.: Lidar distillation: Bridging the beam-induced domain gap for 3d object detection. In: European Conference on Computer Vision. pp. 179-195. Springer (2022) +39. Wu, H., Deng, J., Wen, C., Li, X., Wang, C.: Casa: A cascade attention network for 3d object detection from lidar point clouds. IEEE Transactions on Geoscience and Remote Sensing (2022) +40. Wu, H., Wen, C., Li, W., Yang, R., Wang, C.: Learning transformation-equivariant features for 3d object detection (2022) +41. Wu, H., Wen, C., Shi, S., Li, X., Wang, C.: Virtual sparse convolution for multimodal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21653-21662 (2023) +42. Xia, Q., Chen, Y., Cai, G., Chen, G., Xie, D., Su, J., Wang, Z.: 3-d hanet: A flexible 3-d heatmap auxiliary network for object detection. IEEE Transactions on Geoscience and Remote Sensing 61, 1-13 (2023) +43. Xia, Q., Deng, J., Wen, C., Wu, H., Shi, S., Li, X., Wang, C.: Coin: Contrastive instance feature mining for outdoor 3d object detection with very limited annotations. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6254-6263 (2023) +44. Xiao, P., Shao, Z., Hao, S., Zhang, Z., Chai, X., Jiao, J., Li, Z., Wu, J., Sun, K., Jiang, K., Wang, Y., Yang, D.: Pandaset: Advanced sensor suite dataset for autonomous driving. In: 2021 IEEE International Intelligent Transportation Systems Conference (ITSC) (Sep 2021). https://doi.org/10.1109/itsc48978.2021.9565009, http://dx.doi.org/10.1109/itsc48978.2021.9565009 +45. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d: Self-training for unsupervised domain adaptation on 3d object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10368-10378 (2021) +46. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d++: Denoised self-training for unsupervised domain adaptation on 3d object detection. IEEE transactions on pattern analysis and machine intelligence 45(5), 6354-6371 (2022) +47. Yang, Z., Sun, Y., Liu, S., Jia, J.: 3dssd: Point-based 3d single stage object detector. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR). pp. 11040-11048 (2020) + +48. Yihan, Z., Wang, C., Wang, Y., Xu, H., Ye, C., Yang, Z., Ma, C.: Learning transferable features for point cloud detection via 3d contrastive co-training. Advances in Neural Information Processing Systems 34, 21493-21504 (2021) +49. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR) (2021) +50. You, Y., Diaz-Ruiz, C.A., Wang, Y., Chao, W.L., Hariharan, B., Campbell, M., Weinbergert, K.Q.: Exploiting playbacks in unsupervised domain adaptation for 3d object detection in self-driving cars. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 5070-5077. IEEE (2022) +51. You, Y., Phoo, C.P., Luo, K., Zhang, T., Chao, W.L., Hariharan, B., Campbell, M., Weinberger, K.Q.: Unsupervised adaptation from repeated traversals for autonomous driving. Advances in Neural Information Processing Systems 35, 27716-27729 (2022) +52. Zheng, L., Ma, Z., Zhu, X., Tan, B., Li, S., Long, K., Sun, W., Chen, S., Zhang, L., Wan, M., Huang, L., Bai, J.: Tj4dradset: A 4d radar dataset for autonomous driving (Apr 2022) \ No newline at end of file diff --git a/2024/CMD_ A Cross Mechanism Domain Adaptation Dataset for 3D Object Detection/images.zip b/2024/CMD_ A Cross Mechanism Domain Adaptation Dataset for 3D Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..56262415af87595637db5097351afcf11d54ef97 --- /dev/null +++ b/2024/CMD_ A Cross Mechanism Domain Adaptation Dataset for 3D Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac5768737f1985aa964f3d1f1d3aec89e4a0befc86cfc9d6f51a3e690c3237f8 +size 770592 diff --git a/2024/CMD_ A Cross Mechanism Domain Adaptation Dataset for 3D Object Detection/layout.json b/2024/CMD_ A Cross Mechanism Domain Adaptation Dataset for 3D Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7d1d3000d582f836507b7e621946531c5b1c546c --- /dev/null +++ b/2024/CMD_ A Cross Mechanism Domain Adaptation Dataset for 3D Object Detection/layout.json @@ -0,0 +1,8569 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 141, + 111, + 474, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 111, + 474, + 148 + ], + "spans": [ + { + "bbox": [ + 141, + 111, + 474, + 148 + ], + "type": "text", + "content": "CMD: A Cross Mechanism Domain Adaptation Dataset for 3D Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 164, + 167, + 448, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 167, + 448, + 180 + ], + "spans": [ + { + "bbox": [ + 164, + 167, + 448, + 180 + ], + "type": "text", + "content": "Jinhao Deng" + }, + { + "bbox": [ + 164, + 167, + 448, + 180 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 164, + 167, + 448, + 180 + ], + "type": "text", + "content": ", Wei Ye" + }, + { + "bbox": [ + 164, + 167, + 448, + 180 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 164, + 167, + 448, + 180 + ], + "type": "text", + "content": ", Hai Wu" + }, + { + "bbox": [ + 164, + 167, + 448, + 180 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 164, + 167, + 448, + 180 + ], + "type": "text", + "content": ", Xun Huang" + }, + { + "bbox": [ + 164, + 167, + 448, + 180 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 192, + 184, + 421, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 184, + 421, + 198 + ], + "spans": [ + { + "bbox": [ + 192, + 184, + 421, + 198 + ], + "type": "text", + "content": "Qiming Xia" + }, + { + "bbox": [ + 192, + 184, + 421, + 198 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 192, + 184, + 421, + 198 + ], + "type": "text", + "content": ", Xin Li" + }, + { + "bbox": [ + 192, + 184, + 421, + 198 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 192, + 184, + 421, + 198 + ], + "type": "text", + "content": ", Jin Fang" + }, + { + "bbox": [ + 192, + 184, + 421, + 198 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 192, + 184, + 421, + 198 + ], + "type": "text", + "content": ", Wei Li" + }, + { + "bbox": [ + 192, + 184, + 421, + 198 + ], + "type": "inline_equation", + "content": "^{3,2}" + }, + { + "bbox": [ + 192, + 184, + 421, + 198 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 214, + 202, + 400, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 202, + 400, + 216 + ], + "spans": [ + { + "bbox": [ + 214, + 202, + 400, + 216 + ], + "type": "text", + "content": "Chenglu Wen" + }, + { + "bbox": [ + 214, + 202, + 400, + 216 + ], + "type": "inline_equation", + "content": "^{1,2,\\text{图}}" + }, + { + "bbox": [ + 214, + 202, + 400, + 216 + ], + "type": "text", + "content": " and Cheng Wang" + }, + { + "bbox": [ + 214, + 202, + 400, + 216 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 164, + 224, + 448, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 224, + 448, + 246 + ], + "spans": [ + { + "bbox": [ + 164, + 224, + 448, + 246 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 164, + 224, + 448, + 246 + ], + "type": "text", + "content": " Fujian Key Laboratory of Sensing and Computing for Smart Cities, Xiamen University, Xiamen, P.R China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 247, + 465, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 247, + 465, + 269 + ], + "spans": [ + { + "bbox": [ + 146, + 247, + 465, + 269 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 247, + 465, + 269 + ], + "type": "text", + "content": " Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, Xiamen, P.R. China." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 240, + 269, + 373, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 269, + 373, + 279 + ], + "spans": [ + { + "bbox": [ + 240, + 269, + 373, + 279 + ], + "type": "text", + "content": "3 Inceptio, Shanghai, P.R China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 196, + 279, + 417, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 279, + 417, + 301 + ], + "spans": [ + { + "bbox": [ + 196, + 279, + 417, + 301 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 196, + 279, + 417, + 301 + ], + "type": "text", + "content": " Section of Visual Computing and Interactive Media, Texas A&M University, Texas, USA" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 322, + 455, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 322, + 455, + 553 + ], + "spans": [ + { + "bbox": [ + 159, + 322, + 455, + 553 + ], + "type": "text", + "content": "Abstract. Point cloud data, representing the precise 3D layout of the scene, quickly drives the research of 3D object detection. However, the challenge arises due to the rapid iteration of 3D sensors, which leads to significantly different distributions in point clouds. This, in turn, results in subpar performance of 3D cross-sensor object detection. This paper introduces a Cross Mechanism Dataset, named CMD, to support research tackling this challenge. CMD is the first domain adaptation dataset, comprehensively encompassing diverse mechanical sensors and various scenes for 3D object detection. In terms of sensors, CMD includes 32-beam LiDAR, 128-beam LiDAR, solid-state LiDAR, 4D millimeter-wave radar, and cameras, all of which are well-synchronized and calibrated. Regarding the scenes, CMD consists of 50 sequences collocated from different scenarios, ranging from campuses to highways. Furthermore, we validated the effectiveness of various domain adaptation methods in mitigating sensor-based domain differences. We also proposed a DIG method to reduce domain disparities from the perspectives of Density, Intensity, and Geometry, which effectively bridges the domain gap between different sensors. The experimental results on the CMD dataset show that our proposed DIG method outperforms the state-of-the-art techniques, demonstrating the effectiveness of our baseline method. The dataset and the corresponding code are available at https://github.com/im-djh/CMD." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 160, + 562, + 429, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 562, + 429, + 573 + ], + "spans": [ + { + "bbox": [ + 160, + 562, + 429, + 573 + ], + "type": "text", + "content": "Keywords: Dataset " + }, + { + "bbox": [ + 160, + 562, + 429, + 573 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 562, + 429, + 573 + ], + "type": "text", + "content": " 3D Object Detection " + }, + { + "bbox": [ + 160, + 562, + 429, + 573 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 562, + 429, + 573 + ], + "type": "text", + "content": " Domain Adaptation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 589, + 230, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 230, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 230, + 601 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 611, + 481, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 481, + 635 + ], + "type": "text", + "content": "As a crucial component in robotics and autonomous driving systems, 3D object detection has garnered increasing attention from researchers. Due to the inherent" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 640, + 228, + 651 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 640, + 228, + 651 + ], + "spans": [ + { + "bbox": [ + 141, + 640, + 228, + 651 + ], + "type": "text", + "content": "* Equal contribution" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 653, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 653, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 653, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\text{忍}" + }, + { + "bbox": [ + 141, + 653, + 481, + 665 + ], + "type": "text", + "content": " Corresponding authors: W. Li (liweimcc@gmail.com); C. Wen (clwen@xmu.edu.cn)" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "depth of information, point cloud data enjoys a unique advantage in the domain of 3D object detection. Notably, researchers have developed several exceptional 3D object detection datasets and benchmarks (e.g., KITTI [13], nuScenes [5], Waymo [30], ONCE [21], etc.). Leveraging these high-quality datasets, numerous outstanding 3D object detection approaches [3,9,18,29,39,41,49] emerged, significantly facilitating the 3D object detection research. However, these prevalent datasets typically comprise mechanical spinning LiDAR only w.r.t. 3D point cloud sensors. The widely equipped low-cost automotive-grade sensors, e.g. solid-state LiDAR, in mass-produced vehicles [27] [25] are heavily overlooked even missing in those datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 239, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 430 + ], + "type": "text", + "content": "The truth of outdatedness of existing datasets in terms of sensors behind is that point cloud sensors are undergoing really rapid advancements. Commonly used sensors now include: (1) mechanical spinning LiDAR with low-beam (e.g., 32 beams) or high-beam (e.g., 128 beams); (2) solid-state LiDAR; and (3) the recently acclaimed 4D millimeter-wave radar. These sensors differ either in the number of beams they use or in their world modeling patterns. We refer to these sensor differences as various mechanisms. The data acquired by different mechanisms may exhibit domain disparities in terms of density, intensity, and geometry (see Fig. 1(a), (b) and (c)). As a result, 3D detectors trained with data from one sensor often incur substantial accuracy degradation when directly applied to another sensor (see \"Direct\" in Fig. 1(d), (e) and (f)). For example, when migrating from mechanical spinning LiDAR to a more inexpensive and compact solid-state LiDAR, the cost of annotating new sensor data to re-train detectors is prohibitive. Cross-mechanism domain adaptation is a more efficient yet promising solution with rigid demand to transfer not only detectors/models but also all related assets to new sensors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 434, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 579 + ], + "type": "text", + "content": "Meanwhile, prevalent domain adaptation research [19, 26, 35, 36, 45, 46] finds it hard to quantitatively analyze the pivotal sensor factor. This challenge arises because they have been directed at cross-dataset settings (e.g. domain adaptation between KITTI, Waymo, and nuScenes), covering differences in sensors, geographic location, climate, and so on. Therefore, they have not shown optimal performance when applied to cross-mechanism domain adaptation problems (see Fig. 1(d)(e)(f)). The reason lies in the fact that there is a lack of datasets that contain comprehensive sensors to fully decouple the domain disparities caused by locations and sensors. Several inspiring datasets, encompassing various common sensors (e.g., PandaSet [44], KRadar [22], VoD [23], LiDAR-CS [11]), often fall short in providing comprehensive modality coverage. This limitation, in turn, restricts the design of methods to bridge the sensor-based domain gap." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "To address this problem, this paper introduces a Cross Mechanism domain adaptation Dataset (CMD) for 3D object detection. To the best of our knowledge, CMD is the first domain-adaptation dataset in 3D object detection that contains a comprehensive suite of sensors, including: (1) 32-beam low-resolution and 128-beam high-resolution mechanical spinning (MS) LiDAR, (2) automotive-grade solid-state LiDAR, (3) " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "4D" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " millimeter-wave radar, and (4) camera. This combination provides the most extensive modality coverage. All sensors are time-" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Deng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "synchronized with high accuracy under 1 ms, ensuring that different modalities can capture and model the same scene accurately, even with highly dynamic objects. Furthermore, the CMD comprises 50 sequences. Each sequence has a time span of 20 seconds, with each sensor capturing 10 frames per second, totaling 10,000 frames of data per sensor. In addition, we meticulously annotated 3D objects of 13 categories based on a multi-sensor collaborative annotation system subjected to multiple rounds of human inspection." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 142, + 219, + 348, + 326 + ], + "blocks": [ + { + "bbox": [ + 142, + 219, + 348, + 326 + ], + "lines": [ + { + "bbox": [ + 142, + 219, + 348, + 326 + ], + "spans": [ + { + "bbox": [ + 142, + 219, + 348, + 326 + ], + "type": "image", + "image_path": "e86365b3bc77bba7497ae8073fc1771fd7918d9c880027d8034a9b8d71e20e0f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 350, + 219, + 473, + 326 + ], + "blocks": [ + { + "bbox": [ + 350, + 219, + 473, + 326 + ], + "lines": [ + { + "bbox": [ + 350, + 219, + 473, + 326 + ], + "spans": [ + { + "bbox": [ + 350, + 219, + 473, + 326 + ], + "type": "image", + "image_path": "5c1e820d28cb03e37f134c29a2ac3688029ad1cb9f411692cd8d4b1fa8e65779.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 147, + 326, + 471, + 375 + ], + "blocks": [ + { + "bbox": [ + 147, + 326, + 471, + 375 + ], + "lines": [ + { + "bbox": [ + 147, + 326, + 471, + 375 + ], + "spans": [ + { + "bbox": [ + 147, + 326, + 471, + 375 + ], + "type": "image", + "image_path": "1c36901429a9508c4a44db8570f96aab853dd38886b243a65eff8a07bfdf79e1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 142, + 378, + 247, + 472 + ], + "blocks": [ + { + "bbox": [ + 142, + 378, + 247, + 472 + ], + "lines": [ + { + "bbox": [ + 142, + 378, + 247, + 472 + ], + "spans": [ + { + "bbox": [ + 142, + 378, + 247, + 472 + ], + "type": "image", + "image_path": "19648560fae7813fe06a6fcfcef2539e8c87848e05f1ab1b6d8157f71c6502ed.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 480, + 482, + 548 + ], + "lines": [ + { + "bbox": [ + 130, + 480, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 482, + 548 + ], + "type": "text", + "content": "Fig. 1: (a) The points number distribution of different sensors. (b) The intensity distribution of different sensors. (c) Point clouds of the same instance for different sensors. The results show that data acquired by different sensors exhibit substantial differences in point density, intensity, and geometry. Subfigure (d), (e), and (f) show the cross-mechanism detection results of different domain adaptation methods on our CMD. Our DIG outperforms the traditional ST3D by a large margin." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 256, + 378, + 342, + 472 + ], + "blocks": [ + { + "bbox": [ + 256, + 378, + 342, + 472 + ], + "lines": [ + { + "bbox": [ + 256, + 378, + 342, + 472 + ], + "spans": [ + { + "bbox": [ + 256, + 378, + 342, + 472 + ], + "type": "image", + "image_path": "1087e7ea9d5585bddc4069b58cac5c0feeb8839d2db066bde509b5b6d5110690.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 352, + 378, + 468, + 471 + ], + "blocks": [ + { + "bbox": [ + 352, + 378, + 468, + 471 + ], + "lines": [ + { + "bbox": [ + 352, + 378, + 468, + 471 + ], + "spans": [ + { + "bbox": [ + 352, + 378, + 468, + 471 + ], + "type": "image", + "image_path": "38126f93d0f81ca5d3b6f56dcf5da01c40fdb763952c287894302212856fc139.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Using the CMD, we investigate the cross-mechanism domain adaptation issues and summarize the gaps into three primary components: Density gap, Intensity gap, and Geometry gap. Subsequently, we introduce a simple yet effective baseline method, named DIG, to address each of these gaps systematically. (1) Density gap: We propose the Beam-Distance Down-sampling (BDS) that enforces a similar data pattern in terms of beam numbers and density distribution along different detection distances. (2) Intensity gap: We introduce the Box-Cox log Normalization (BCN) that initially transforms point clouds from" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 423, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 423, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 423, + 91, + 447, + 100 + ], + "type": "text", + "content": "CMD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 186 + ], + "type": "text", + "content": "different modalities into near-normal distributions and then normalizes them by logarithms. (3) Geometry gap: We design the Geometry-Aware label Mixing (GAM) that mixes the pseudo-labels from the target domain to narrow down the geometry difference in training. By implementing these three components, DIG achieves optimal results in bridging the identified domain gaps. Our main contributions are as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 186, + 480, + 316 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 136, + 186, + 479, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 186, + 479, + 245 + ], + "spans": [ + { + "bbox": [ + 136, + 186, + 479, + 245 + ], + "type": "text", + "content": "- We present the Cross Mechanism Dataset (CMD), the first domain adaptation dataset that contains a comprehensive suite of LiDAR/4D Radar sensors. CMD is a key piece of the puzzle of cross-mechanism detection as existing public datasets can not adequately evaluate detection performance drops caused by cross-domain disparities due to limited sensor variety." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 247, + 480, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 247, + 480, + 269 + ], + "spans": [ + { + "bbox": [ + 138, + 247, + 480, + 269 + ], + "type": "text", + "content": "- We construct comprehensive 3D cross-sensor detection baselines by benchmarking the results of SOTA domain adaptation algorithms on CMD." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 270, + 480, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 270, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 270, + 480, + 316 + ], + "type": "text", + "content": "- We provide an in-depth analysis of the key factors leading to cross-sensor domain disparities, i.e. Density, Intensity, and Geometry gaps. Based on this, we propose a novel DIG method, which shows plausible performance on cross-sensor detection." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 335, + 237, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 335, + 237, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 335, + 237, + 347 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 359, + 330, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 359, + 330, + 371 + ], + "spans": [ + { + "bbox": [ + 132, + 359, + 330, + 371 + ], + "type": "text", + "content": "2.1 Datasets for 3D Object Detection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 545 + ], + "type": "text", + "content": "With increasing research on 3D object detection [7, 8, 17, 40, 42, 43, 47], more datasets have emerged. Widely used datasets such as KITTI [13], Waymo [30], nuScenes [5], and ONCE [21] support conventional 3D object detection research and facilitate domain adaptation across different geographic locations. However, focusing solely on geographical domain differences is inadequate for addressing the domain adaptation problem in detection algorithms; sensor variability is also crucial. To address this, datasets like PandaSet [44], K-Radar [22], TJ4DRadSet [52], VoD [23], Lyft L5 [15], and Cirrus [37] offer diverse sensors, scan beams, locations, and weather conditions. Despite their contributions, these datasets are limited in sensor diversity within individual collections. In contrast, our CMD includes extensive image and point cloud data annotated for detection and tracking, encompassing five different sensor types, thus providing a comprehensive evaluation. Statistical comparisons with other 3D object detection datasets are shown in Table 2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 563, + 372, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 563, + 372, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 563, + 372, + 574 + ], + "type": "text", + "content": "2.2 Domain Adaption in 3D Object Detection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": "To mitigate the poor generalization of 3D object detectors to unknown data, recent methods have focused on adapting domain gaps [10, 14, 32-34, 48, 50, 51]. Among these methods, some attempt to introduce a teacher-student framework. ST3D [45] and " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{ST3D + + }" + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": " [46] design label assignment strategy and pseudo-labels denoising to enable adaptation between domains. MLC-Net [19] employs consistency for cross-domain transfer. Another statistics-based approach adapts distribution gaps, with SN [35] correcting car size distribution gaps. GBA [26]" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Deng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": "adapts label distribution gap via a Gradual Batch Alternation training strategy. And, SSDA3D [36] adapting point cloud distributions through Inter-domain Alignment module. Further methods address sensor gaps, with LiDAR Distillation [38] and DTS [16] proposing progressive and density-insensitive frameworks to mitigate beam-induce gaps from different sensors, and CL3D [24] using spatial geometry alignment to adapt geometric gaps between sensors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 480, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 480, + 247 + ], + "type": "text", + "content": "In general, current domain adaptation research methods still primarily focus on geographical information and sensor beam migration. Due to dataset limitations, domain adaptation methods for multiple types of sensors have yet to be seen. Therefore, it is necessary to propose a novel, well-annotated dataset including various types of sensors to advance this research field." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 267, + 307, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 267, + 307, + 280 + ], + "spans": [ + { + "bbox": [ + 131, + 267, + 307, + 280 + ], + "type": "text", + "content": "3 Cross Mechanism Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 294, + 228, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 294, + 228, + 306 + ], + "spans": [ + { + "bbox": [ + 131, + 294, + 228, + 306 + ], + "type": "text", + "content": "3.1 Sensor Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 315, + 480, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 480, + 411 + ], + "type": "text", + "content": "Sensor specifications. Our CMD comprises data from seven sensors of five types: a 128-beams mechanical scanning (MS) LiDAR, a 32-beams MS LiDAR, a solid-state LiDAR, a 4D millimeter-wave radar, and three cameras. Detailed specifications for our sensors are shown in Table 1. The sensors in our dataset, each with unique beam configurations or modeling modes, are collectively referred to as mechanisms. As shown in Table 2, in contrast to other datasets, our CMD encompasses a diverse array of commonly employed sensors, establishing it as the most comprehensive dataset presently available." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 149, + 432, + 465, + 501 + ], + "blocks": [ + { + "bbox": [ + 149, + 432, + 465, + 501 + ], + "lines": [ + { + "bbox": [ + 149, + 432, + 465, + 501 + ], + "spans": [ + { + "bbox": [ + 149, + 432, + 465, + 501 + ], + "type": "table", + "html": "
SensorTypeHFOV(°)VFOV(°)ResolutionFPS
OS128128 beams MS LiDAR360[-22.5, 22.5]128 * 102410
XT3232 beams MS LiDAR360[-16, 15]32 * 204810
M1SS LiDAR[-60, 60][-12.5, 12.5]126 * 62510
Radar4D Radar[-75, 75][-15, 15]>60010
HIK Cam.Camera[-31.2, 31.2][-27.7, 27.7]1080 * 192010
", + "image_path": "bf0258756fce89372fd28d81e17bdeacc780d3e2e1b2c0fe3f09b5b2087bfb24.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 501, + 479, + 523 + ], + "lines": [ + { + "bbox": [ + 131, + 501, + 479, + 523 + ], + "spans": [ + { + "bbox": [ + 131, + 501, + 479, + 523 + ], + "type": "text", + "content": "Table 1: Sensor specifications for CMD. MS and SS denote mechanical spinning and solid-state respectively. FPS refers to frames per second." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 558, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 666 + ], + "type": "text", + "content": "Sensor layout. In our CMD, all sensors are oriented in the same direction as the vehicle's forward direction. Such a configuration is instrumental in ensuring that the most critical information is captured within the area of greatest overlap of FOV across different sensors. To achieve this alignment, a rigid bracket has been designed, aligning all sensors along a vertically aligned axis. We have deliberately placed all sensors as close to each other as possible to optimize FOV overlap. In addition to this, a camera system comprising three evenly spaced cameras has been integrated. This system spans a " + }, + { + "bbox": [ + 130, + 558, + 481, + 666 + ], + "type": "inline_equation", + "content": "150^{\\circ}" + }, + { + "bbox": [ + 130, + 558, + 481, + 666 + ], + "type": "text", + "content": " horizontal field of view (HFOV), as depicted in Fig. 2(c). Such a configuration is pivotal" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 423, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 423, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 423, + 91, + 447, + 100 + ], + "type": "text", + "content": "CMD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "for ensuring data consistency across different mechanisms. The overall FOV is shown in Fig. 2(b)." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 163, + 206, + 243 + ], + "blocks": [ + { + "bbox": [ + 133, + 163, + 206, + 243 + ], + "lines": [ + { + "bbox": [ + 133, + 163, + 206, + 243 + ], + "spans": [ + { + "bbox": [ + 133, + 163, + 206, + 243 + ], + "type": "image", + "image_path": "b9e908acc9ce7a128336dd9367be8276b438db5a1b413d096b9fe2265c97f61b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 253, + 480, + 286 + ], + "lines": [ + { + "bbox": [ + 130, + 253, + 480, + 286 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 480, + 286 + ], + "type": "text", + "content": "Fig. 2: The illustration of sensor layout. (a) demonstrates the specific layout of each sensor; (b) showcases the field of view angles for each sensor; (c) demonstrates the positions of each sensor along the z-axis." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 208, + 163, + 268, + 243 + ], + "blocks": [ + { + "bbox": [ + 208, + 163, + 268, + 243 + ], + "lines": [ + { + "bbox": [ + 208, + 163, + 268, + 243 + ], + "spans": [ + { + "bbox": [ + 208, + 163, + 268, + 243 + ], + "type": "image", + "image_path": "bbabdac9b0b734d763543ec6b3060bea56092ed4a1bb913dad5f0f8ddcb62e28.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 269, + 163, + 301, + 242 + ], + "blocks": [ + { + "bbox": [ + 269, + 163, + 301, + 242 + ], + "lines": [ + { + "bbox": [ + 269, + 163, + 301, + 242 + ], + "spans": [ + { + "bbox": [ + 269, + 163, + 301, + 242 + ], + "type": "image", + "image_path": "dc3843b9a3bd9db89f83b0e7b647f6b5610299d6b5647aa2d7f8816bf10e650b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 302, + 164, + 386, + 242 + ], + "blocks": [ + { + "bbox": [ + 302, + 164, + 386, + 242 + ], + "lines": [ + { + "bbox": [ + 302, + 164, + 386, + 242 + ], + "spans": [ + { + "bbox": [ + 302, + 164, + 386, + 242 + ], + "type": "image", + "image_path": "6122c5a37371835274ea7fe115519b4181b96137aa84d7212366b417a0e500e5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 387, + 163, + 478, + 243 + ], + "blocks": [ + { + "bbox": [ + 387, + 163, + 478, + 243 + ], + "lines": [ + { + "bbox": [ + 387, + 163, + 478, + 243 + ], + "spans": [ + { + "bbox": [ + 387, + 163, + 478, + 243 + ], + "type": "image", + "image_path": "42a273595bb090827ed95f7186f5bcbcc47e244ab382add5dae2697ef8397307.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 331, + 324, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 331, + 324, + 342 + ], + "spans": [ + { + "bbox": [ + 131, + 331, + 324, + 342 + ], + "type": "text", + "content": "3.2 Synchronization and Calibration" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 354, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 569 + ], + "type": "text", + "content": "Synchronization and trigger. To achieve synchronization among different sensors when modeling the world, we utilized the Precision Time Protocol Version 2 [1] to establish a time synchronization system with an error margin within " + }, + { + "bbox": [ + 130, + 354, + 482, + 569 + ], + "type": "inline_equation", + "content": "1\\mathrm{ms}" + }, + { + "bbox": [ + 130, + 354, + 482, + 569 + ], + "type": "text", + "content": ". Specifically, we employed the CoolShark AUTO 66 unit as the PTP master clock, while other sensors and the host served as PTP slave clocks. Once synchronized, M1 is set to generate a frame every " + }, + { + "bbox": [ + 130, + 354, + 482, + 569 + ], + "type": "inline_equation", + "content": "100\\mathrm{ms}" + }, + { + "bbox": [ + 130, + 354, + 482, + 569 + ], + "type": "text", + "content": ", and both OS128 and XT32 generate frames simultaneously using phase-locking. Meanwhile, the radar is triggered by a network signal sent from the host. Due to the differing principles of data acquisition between cameras and LiDAR, the cameras are triggered by OS128. Whenever OS128 scans to " + }, + { + "bbox": [ + 130, + 354, + 482, + 569 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 130, + 354, + 482, + 569 + ], + "type": "text", + "content": ", it sends out a trigger signal. This signal undergoes a delay process via a microcontroller before being fed into the cameras. For each camera, " + }, + { + "bbox": [ + 130, + 354, + 482, + 569 + ], + "type": "inline_equation", + "content": "T_{delay} = \\frac{\\theta_{cam}}{2\\pi} - \\frac{1}{2} T_{exposure}" + }, + { + "bbox": [ + 130, + 354, + 482, + 569 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 354, + 482, + 569 + ], + "type": "inline_equation", + "content": "T_{exposure}" + }, + { + "bbox": [ + 130, + 354, + 482, + 569 + ], + "type": "text", + "content": " means the exposure time for camera and " + }, + { + "bbox": [ + 130, + 354, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\theta_{cam}" + }, + { + "bbox": [ + 130, + 354, + 482, + 569 + ], + "type": "text", + "content": " means the angle at which the camera is positioned. By this, the time when OS128 scans the central of the camera's HFOV, is the central of its exposure period, which ensures a lower frame time offset between the cameras and the LiDARs. This low-latency synchronization ensures that the same frame of data captured by different sensors effectively reflects their respective mechanism differences." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Calibration. Achieving high-quality data in a multi-sensor setup relies heavily on calibration. (1) For the LiDARs (i.e. OS128, M1, XT32) and radar, we calibrate their extrinsic parameters using Generalized-ICP [28]. Several corner reflectors in the space are used to obtain more precise corner points for radar calibration. (2) For the cameras, we calibrate their intrinsic parameters and extrinsic parameters with MATLAB Toolkit [12] and OpenCV [4]. The results of calibration are shown in Fig. 4. Subsequent to calibration, we established a well-defined coordinate system, as illustrated in Fig. 2(c). We transfer XT32 and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Deng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "radar to the coordinate system of OS128 and move this coordinate system vertically down to the ground to get the ego system. The ego system has the x-axis pointing forward, the y-axis to the left, and the z-axis upward. The camera coordinate system, in 2D, aligns the x-axis with image width and y-axis with height, starting from the top-left corner." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 196, + 218, + 206 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 196, + 218, + 206 + ], + "spans": [ + { + "bbox": [ + 131, + 196, + 218, + 206 + ], + "type": "text", + "content": "3.3 Annotation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 217, + 479, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 479, + 264 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 479, + 264 + ], + "type": "text", + "content": "Annotation area. With precise time synchronization and calibration, we can jointly annotate data from multiple sensors, improving accuracy, especially for distant objects. We annotate objects within " + }, + { + "bbox": [ + 130, + 217, + 479, + 264 + ], + "type": "inline_equation", + "content": "160m" + }, + { + "bbox": [ + 130, + 217, + 479, + 264 + ], + "type": "text", + "content": " and within the HFOV of M1. We believe that the data within this range is highly representative." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 265, + 480, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 480, + 361 + ], + "type": "text", + "content": "Annotation rules. There are 13 annotated categories, namely: Car, Van, Bus, Truck, Semi-Trailer towing vehicle, Special Vehicles, Motorcycle, Bicycle, Tricycle, Adult Pedestrian, Children Pedestrian, Animal, and Barrier. We labeled each object as a 9-DoF 3D bounding box " + }, + { + "bbox": [ + 130, + 265, + 480, + 361 + ], + "type": "inline_equation", + "content": "(x,y,z,l,w,h,\\theta_x,\\theta_y,\\theta_z)" + }, + { + "bbox": [ + 130, + 265, + 480, + 361 + ], + "type": "text", + "content": ". Where " + }, + { + "bbox": [ + 130, + 265, + 480, + 361 + ], + "type": "inline_equation", + "content": "x,y,z" + }, + { + "bbox": [ + 130, + 265, + 480, + 361 + ], + "type": "text", + "content": " represents the center coordinates, " + }, + { + "bbox": [ + 130, + 265, + 480, + 361 + ], + "type": "inline_equation", + "content": "l,w,h" + }, + { + "bbox": [ + 130, + 265, + 480, + 361 + ], + "type": "text", + "content": " denotes length, width height, and " + }, + { + "bbox": [ + 130, + 265, + 480, + 361 + ], + "type": "inline_equation", + "content": "\\theta_x,\\theta_y,\\theta_z" + }, + { + "bbox": [ + 130, + 265, + 480, + 361 + ], + "type": "text", + "content": " are the rotation angles around the x,y,z axis. Additionally, we provide a motion state and tracking ID for all objects. Occlusion situations are also provided. More annotation specifications are provided in the appendix." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 134, + 381, + 479, + 487 + ], + "blocks": [ + { + "bbox": [ + 134, + 381, + 479, + 487 + ], + "lines": [ + { + "bbox": [ + 134, + 381, + 479, + 487 + ], + "spans": [ + { + "bbox": [ + 134, + 381, + 479, + 487 + ], + "type": "table", + "html": "
DatasetMechanismMS LiDAR(beams)SS LiDAR 4DRadar Cam.Anno.Frames TrackingIlluminationReal
LowHigh
KITTI Det. [13]X64XX✓15KX✓✓
Waymo [30]X64XX✓230K✓✓✓
nuScenes [5]Single32XXX✓40K✓✓✓
Argoverse [6]32*2XXX✓44K✓✓✓
Lyft L5 [15]40*264XX✓30K✓X✓
K-Radar [22]X64, 128X✓✓35K✓✓✓
Lidar-CS [11]16, 3264, 128✓XX14KXXX
PandaSet [44]MultiX64✓X✓8.2K✓✓✓
TJ4DRadSet [52]X64X✓✓7.8K✓✓✓
VoD [23]X64X✓✓8.7KXX✓
CMD32128✓✓✓10K✓✓✓
", + "image_path": "e6e404ab41c84208c456b3be7461a64186a7dd167e1b80e593573fc15f6e84e7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 487, + 479, + 529 + ], + "lines": [ + { + "bbox": [ + 130, + 487, + 479, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 479, + 529 + ], + "type": "text", + "content": "Table 2: Comparison with existing datasets. \"Mechanism\" means whether the dataset contains multiple mechanisms of 3D sensor. \"Tracking\" signifies the relevance of annotations for tracking. \"Illumination\" indicate the inclusion of various light conditions. \"Real\" refers to whether the dataset is real or synthetic." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 131, + 584, + 246, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 584, + 246, + 596 + ], + "spans": [ + { + "bbox": [ + 131, + 584, + 246, + 596 + ], + "type": "text", + "content": "3.4 Dataset Analytic" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "content": "Diversity of scenes. In CMD, scenes are carefully selected for wide coverage, primarily encompassing urban areas, suburbs, campuses, bridges, and tunnels with different illuminations (see Fig. 3(a)). We select 50 high-quality sequences, each spanning 20 seconds, equating to 200 frames per sensor, culminating in a dataset of 40,000 frames of point cloud data and 30,000 frames of image" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 423, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 423, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 423, + 91, + 447, + 100 + ], + "type": "text", + "content": "CMD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "data. This comprehensive collection offers a rich resource for research on cross-mechanism domain adaptation. These 50 sequences are evenly divided into 30, 10 and 10 for training, validation, and testing according to different scenarios. This arrangement ensures a balanced distribution of data across environments, crucial for training reliable and generalizable 3D object detection models." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 135, + 219, + 352, + 375 + ], + "blocks": [ + { + "bbox": [ + 140, + 210, + 342, + 218 + ], + "lines": [ + { + "bbox": [ + 140, + 210, + 342, + 218 + ], + "spans": [ + { + "bbox": [ + 140, + 210, + 342, + 218 + ], + "type": "text", + "content": "Cross Mechanism domain adaptation Dataset (CMD) for 3D Object Detection" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 135, + 219, + 352, + 375 + ], + "lines": [ + { + "bbox": [ + 135, + 219, + 352, + 375 + ], + "spans": [ + { + "bbox": [ + 135, + 219, + 352, + 375 + ], + "type": "image", + "image_path": "6f5eb51ed399a4c404f634a84090c5285db2626f35e942401b3654d923c4b842.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 359, + 212, + 481, + 290 + ], + "blocks": [ + { + "bbox": [ + 359, + 212, + 481, + 290 + ], + "lines": [ + { + "bbox": [ + 359, + 212, + 481, + 290 + ], + "spans": [ + { + "bbox": [ + 359, + 212, + 481, + 290 + ], + "type": "image", + "image_path": "9b1c6c08041cd325ed272d6799b9d9874b8cfdbc9cadbf18bb017372efc3839e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 359, + 297, + 479, + 376 + ], + "blocks": [ + { + "bbox": [ + 359, + 297, + 479, + 376 + ], + "lines": [ + { + "bbox": [ + 359, + 297, + 479, + 376 + ], + "spans": [ + { + "bbox": [ + 359, + 297, + 479, + 376 + ], + "type": "image", + "image_path": "21ce536fbb4f5e2d3e9da15e4cad9723fdab2b134f07aee08105e37988666d36.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 385, + 482, + 431 + ], + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 431 + ], + "type": "text", + "content": "Fig. 3: The scene and annotation details. (a) The upper section showcases the distribution of road types, while the lower section demonstrates the distribution across various illumination types; (b) presents the distribution of objects' numbers for different categories; (c) presents the distribution of objects' numbers along different distances." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 467, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 467, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 467, + 482, + 540 + ], + "type": "text", + "content": "Distribution of annotations. We annotated approximately 230,000 3D bounding boxes based on the rules described in Section 3.3. The distribution of objects' numbers for different categories is presented in Fig. 3(b). The distribution of objects' numbers along different distances is shown in Fig. 3(c). Among these categories, the most concerned \"Car\" has the highest annotation count, reaching approximately 92,000 instances." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "Comparisons between mechanisms. The average number of points for these four sensors is quite different. From most to least, they are M1, OS128, XT32, and 4D radar. For a better understanding, we show the visualization results in Fig. 4. Point clouds from sensors with different mechanisms exhibit substantial differences. (1) OS128 and XT32 both adopt a mechanism spinning style, clearly modeling the rigid vehicles with minimal vertical distortion. The main difference between them is density. (2) Due to the use of a zigzag scan pattern, M1 tends to introduce a bit of horizontal displacement. (3) Point clouds from radar sensors are generally sparse and inaccurate, posing challenges in accurately distinguishing the geometric shapes of objects." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Deng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 479, + 336 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 479, + 336 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 479, + 336 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 479, + 336 + ], + "type": "image", + "image_path": "1bda9ae537febd463216a986dabd09069ba31d81a3a8735a05988594fd340c5e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 345, + 482, + 377 + ], + "lines": [ + { + "bbox": [ + 130, + 345, + 482, + 377 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 482, + 377 + ], + "type": "text", + "content": "Fig. 4: Different mechanisms and their sample data. The left side of the image denotes the various scenes of data collection, while the bottom indicates the types of sensors involved." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 399, + 272, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 399, + 272, + 410 + ], + "spans": [ + { + "bbox": [ + 132, + 399, + 272, + 410 + ], + "type": "text", + "content": "3.5 DIG Baseline Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 416, + 479, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 479, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 479, + 464 + ], + "type": "text", + "content": "In this section, we propose DIG (Density-Intensity-Geometry), a streamlined and potent baseline for cross-mechanism domain adaptation. It consists of three key components: Box-Cox Intensity Log Normalization (BCN), Beam-Distance Down Sampling (BDS), and Geometry-Aware Label Mixing (GAM)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 464, + 482, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 608 + ], + "type": "text", + "content": "Box-Cox intensity log Normalization (BCN). The distribution of intensity for point clouds may significantly differ across various mechanisms. See Fig. 1(b). Recent studies have either opted to discard [45] [38] intensity or straightforwardly normalized it to [0,1] [36]. However, it is observed that intensity plays a crucial role in detecting foreground objects across diverse domains. We contend that effectively incorporating intensity can lead to enhanced performance in domain adaptation. Motivated by this insight, we introduce the BCN to mitigate disparities in intensity. For both intensities (denoted as " + }, + { + "bbox": [ + 130, + 464, + 482, + 608 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 464, + 482, + 608 + ], + "type": "text", + "content": ") for source and target domain point clouds, we initially take the logarithm to address biases stemming from different sensor definitions of intensity ranges. Subsequently, we apply the Box-Cox transformation to enhance their normality and homoscedasticity, respectively. This process can be formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 255, + 612, + 481, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 612, + 481, + 637 + ], + "spans": [ + { + "bbox": [ + 255, + 612, + 481, + 637 + ], + "type": "interline_equation", + "content": "i (\\lambda) = \\frac {\\log (i + 1) ^ {\\lambda} - 1}{\\lambda}, \\tag {1}", + "image_path": "932dc0a0fe847c4f6d95c4a4a0b79638912a43181de707b97813db0d37a98cf4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " refers to the box-cox transfer parameter [2]. It's different for source and target domains, and remains constant respectively. After being normalized to" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "text", + "content": "CMD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "content": "[0,1], the transformed intensity information exhibits a good level of similarity between the source and target domains, thus narrowing the domain gap." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": "Beam-Distance based down Sampling (BDS). As shown in Fig. 1(a), the density distributions of points vary remarkably among different mechanisms. We observe that, within a relatively short distance, domain adaptation for different mechanisms is more significantly impacted by variations in point density. However, when it comes to longer distances, the point cloud is excessively sparse, which has a greater impact on object detection accuracy. To deal with this problem, we introduce the BDS that assigns different sampling probabilities based on both beam and distances. Denote " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "P_{\\text{keep}}(p)" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": " as the probability of a point (p) being retained, and it can be formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 208, + 253, + 481, + 276 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 253, + 481, + 276 + ], + "spans": [ + { + "bbox": [ + 208, + 253, + 481, + 276 + ], + "type": "interline_equation", + "content": "P _ {\\text {k e e p}} (p) = 1 - \\mathbf {1} \\left\\{\\text {b e a m} (p) \\notin \\mathcal {B} \\right\\} \\exp \\left(- \\frac {1}{K} | | p | |\\right), \\tag {2}", + "image_path": "8b6a9717691ec60195f665edf5be65d2efb76f296a8272b943fc105616f11d13.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 280, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 482, + 369 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 280, + 482, + 369 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 280, + 482, + 369 + ], + "type": "text", + "content": " controls the rate of probability increase, and we set it to 35 empirically. " + }, + { + "bbox": [ + 130, + 280, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 130, + 280, + 482, + 369 + ], + "type": "text", + "content": " means the set of target beam indexes that are closest to the source domain. Function " + }, + { + "bbox": [ + 130, + 280, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\text{beam}(p)" + }, + { + "bbox": [ + 130, + 280, + 482, + 369 + ], + "type": "text", + "content": " denotes the index of the beam that point " + }, + { + "bbox": [ + 130, + 280, + 482, + 369 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 280, + 482, + 369 + ], + "type": "text", + "content": " belongs to. If this beam information is not directly available, an alternative approach involves calculating the angle of each point, denoted as " + }, + { + "bbox": [ + 130, + 280, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\theta(p) = \\arctan \\left( \\frac{z}{\\sqrt{x^2 + y^2}} \\right)" + }, + { + "bbox": [ + 130, + 280, + 482, + 369 + ], + "type": "text", + "content": ". This angle can then be used to apply a clustering algorithm, such as K-means [20], to approximate the beam index for each point." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 369, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 525 + ], + "type": "text", + "content": "Geometry-Aware label Mixing (GAM). Point clouds captured from the same object by diverse sensors frequently exhibit notable geometric variances. To tackle this challenge, we introduce the Geometry-Aware Mix (GAM) module. In our approach, we initially utilize a model trained on the source domain to generate pseudo-labels within the target domain. Subsequently, instances from the target domain with high-confidence pseudo-labels are integrated back into the source domain for further training. Inspired by GBA [26], our second training phase progressively reduces source domain labels while maintaining a constant number of target domain pseudo-labels. This technique aims to gradually adapt the model to the geometric characteristics of the target domain. Through this method, the model becomes more adept at learning and adapting to the unique data characteristics of the target domain, thereby improving its ability for cross-mechanism domain adaptation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 541, + 277, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 541, + 277, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 541, + 277, + 555 + ], + "type": "text", + "content": "4 Experimental Result" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 563, + 253, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 563, + 253, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 563, + 253, + 577 + ], + "type": "text", + "content": "4.1 Experiment Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "All of our experiments were conducted using four Nvidia 3090 GPUs and the open-source code repository OpenPCDet [31]. The Voxel-RCNN [9] served as the detector for most experiments. The batch size was set to 32, and the learning rate was 0.003. Similar to KITTI [13], the grid range was defined as " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "[0\\mathrm{m},70.4\\mathrm{m}]" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "[-70.4\\mathrm{m},70.4\\mathrm{m}]" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "[-2\\mathrm{m},4\\mathrm{m}]" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " along the x, y, and z axis respectively. Point cloud data were cropped to HFOV " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "[-60^{\\circ},60^{\\circ}]" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " with horizontal distances smaller than " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "70.4\\mathrm{m}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ". Voxel size was set to " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "[0.1,0.1,0.15]" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " along the x, y, and z axis respectively." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Deng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 168, + 116, + 291, + 185 + ], + "blocks": [ + { + "bbox": [ + 168, + 116, + 291, + 185 + ], + "lines": [ + { + "bbox": [ + 168, + 116, + 291, + 185 + ], + "spans": [ + { + "bbox": [ + 168, + 116, + 291, + 185 + ], + "type": "image", + "image_path": "01118e19813a7984f94f896e43790c8b47e4e5b661ca8a7eaa6bd77d437286f2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 294, + 116, + 445, + 184 + ], + "blocks": [ + { + "bbox": [ + 294, + 116, + 445, + 184 + ], + "lines": [ + { + "bbox": [ + 294, + 116, + 445, + 184 + ], + "spans": [ + { + "bbox": [ + 294, + 116, + 445, + 184 + ], + "type": "image", + "image_path": "07652d9d90f9f1cbf22863e6755a61f842a8a432a61681ded5a8b8e05cce1f32.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 169, + 186, + 290, + 256 + ], + "blocks": [ + { + "bbox": [ + 169, + 186, + 290, + 256 + ], + "lines": [ + { + "bbox": [ + 169, + 186, + 290, + 256 + ], + "spans": [ + { + "bbox": [ + 169, + 186, + 290, + 256 + ], + "type": "image", + "image_path": "ea1fd491f024a4bc8e5c89d75d71b4f7bf345ce1262e684bb71b903768ab0c94.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 256, + 255, + 266 + ], + "lines": [ + { + "bbox": [ + 168, + 256, + 255, + 266 + ], + "spans": [ + { + "bbox": [ + 168, + 256, + 255, + 266 + ], + "type": "inline_equation", + "content": "\\pmb{X}^{S}" + }, + { + "bbox": [ + 168, + 256, + 255, + 266 + ], + "type": "text", + "content": " Points in source domain" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 260, + 256, + 356, + 266 + ], + "lines": [ + { + "bbox": [ + 260, + 256, + 356, + 266 + ], + "spans": [ + { + "bbox": [ + 260, + 256, + 356, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\theta^s)" + }, + { + "bbox": [ + 260, + 256, + 356, + 266 + ], + "type": "text", + "content": " Model in source domain" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 168, + 268, + 251, + 277 + ], + "lines": [ + { + "bbox": [ + 168, + 268, + 251, + 277 + ], + "spans": [ + { + "bbox": [ + 168, + 268, + 251, + 277 + ], + "type": "inline_equation", + "content": "X^t" + }, + { + "bbox": [ + 168, + 268, + 251, + 277 + ], + "type": "text", + "content": " Points in target domain" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 260, + 268, + 353, + 277 + ], + "lines": [ + { + "bbox": [ + 260, + 268, + 353, + 277 + ], + "spans": [ + { + "bbox": [ + 260, + 268, + 353, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\theta^t)" + }, + { + "bbox": [ + 260, + 268, + 353, + 277 + ], + "type": "text", + "content": " Model in target domain" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 130, + 288, + 482, + 332 + ], + "lines": [ + { + "bbox": [ + 130, + 288, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 482, + 332 + ], + "type": "text", + "content": "Fig. 5: Similarities and differences among baselines. ST3D [45] leverage random object scaling, generating and denoising pseudo-labels. DTS [16] down sample point clouds before training source-detector. Our DIG jointly downs ample point clouds, transforms intensity, and makes use of pseudo-labels." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 294, + 186, + 444, + 255 + ], + "blocks": [ + { + "bbox": [ + 294, + 186, + 444, + 255 + ], + "lines": [ + { + "bbox": [ + 294, + 186, + 444, + 255 + ], + "spans": [ + { + "bbox": [ + 294, + 186, + 444, + 255 + ], + "type": "image", + "image_path": "d56d7b09aeb8f7155d7fb0927d10e3d0468f3010a2422009f71b0f78d8b966f5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 255, + 440, + 266 + ], + "lines": [ + { + "bbox": [ + 360, + 255, + 440, + 266 + ], + "spans": [ + { + "bbox": [ + 360, + 255, + 440, + 266 + ], + "type": "inline_equation", + "content": "\\varphi^a" + }, + { + "bbox": [ + 360, + 255, + 440, + 266 + ], + "type": "text", + "content": " Points down sampling" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 360, + 268, + 441, + 277 + ], + "lines": [ + { + "bbox": [ + 360, + 268, + 441, + 277 + ], + "spans": [ + { + "bbox": [ + 360, + 268, + 441, + 277 + ], + "type": "inline_equation", + "content": "\\varphi^i" + }, + { + "bbox": [ + 360, + 268, + 441, + 277 + ], + "type": "text", + "content": " Intensity normalization" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 353, + 253, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 253, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 253, + 364 + ], + "type": "text", + "content": "4.2 Evaluation Metric" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 369, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 441 + ], + "type": "text", + "content": "Mean Average Precision (mAP). mAP is a commonly used object detection evaluation metric. Here when calculating mAP, we use 3D IoU thresholds of 0.5, 0.5, 0.25, and 0.25 for Car, Truck, Pedestrian, and Cyclist, respectively. Following the ONCE [21] dataset, we calculated an orientation-aware AP on 50 precision steps. The AP among different detection distances (e.g., overall and 0 to " + }, + { + "bbox": [ + 130, + 369, + 482, + 441 + ], + "type": "inline_equation", + "content": "30\\mathrm{m}" + }, + { + "bbox": [ + 130, + 369, + 482, + 441 + ], + "type": "text", + "content": ") are also computed for detailed analysis. Formally, the AP is defined as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 228, + 447, + 481, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 447, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 228, + 447, + 481, + 475 + ], + "type": "interline_equation", + "content": "A P = 1 0 0 \\int_ {0} ^ {1} \\max \\left\\{p \\left(r ^ {\\prime} \\mid r ^ {\\prime} \\geq r\\right) \\right\\} d r, \\tag {3}", + "image_path": "26ff09fb10d0fa47e39b561568c379a3dc03e9c0d3659e3b9c68e3af4875b72a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 479, + 481, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 481, + 503 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 479, + 481, + 503 + ], + "type": "inline_equation", + "content": "p(r)" + }, + { + "bbox": [ + 130, + 479, + 481, + 503 + ], + "type": "text", + "content": " refers to the precision-recall curve that calculated by 50 recall positions. The mAP is the average of AP across all categories." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 502, + 482, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 551 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 551 + ], + "type": "text", + "content": "Mean Closed Gap (mCG). Directly applying AP or mAP as evaluation metrics for domain adaptation is not intuitive since different sensors do not share the same domain gap. Closed gap [35] is used for measuring the effectiveness of a method on a single domain adaptation task. CG can be formulated as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 232, + 555, + 481, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 555, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 232, + 555, + 481, + 581 + ], + "type": "interline_equation", + "content": "C G = 1 0 0 \\frac {m A P _ {\\text {m o d e l}} - m A P _ {D T}}{m A P _ {\\text {O r a c l e}} - m A P _ {D T}}, \\tag {4}", + "image_path": "1658a5807e3a8b2f707ee4266256a89c53dadf363dc4ff6886cf436bbeb0ba1c.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 583, + 482, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 583, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 583, + 482, + 632 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 583, + 482, + 632 + ], + "type": "inline_equation", + "content": "DT" + }, + { + "bbox": [ + 130, + 583, + 482, + 632 + ], + "type": "text", + "content": " means direct transfer using a model trained sorely on source domain data. Oracle implies results obtained by the full training on target domain. To quantify the universality of domain adaptation methods, we define a new metric named mean closed gap (mCG), formulated as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 232, + 637, + 481, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 637, + 481, + 663 + ], + "spans": [ + { + "bbox": [ + 232, + 637, + 481, + 663 + ], + "type": "interline_equation", + "content": "m C G = \\sum_ {s \\in \\mathbf {M}} \\sum_ {t \\in \\mathbf {M}} \\mathbf {1} (s \\neq t) C G _ {s \\rightarrow t}, \\tag {5}", + "image_path": "877da452f44811a8250b548f300ac09c0cb390713f4734eec3923424e3b97fa6.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "text", + "content": "CMD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 128 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 116, + 479, + 128 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 132, + 116, + 479, + 128 + ], + "type": "text", + "content": " refers to the set of sensors, including OS128, XT32, and M1 in CMD." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 135, + 148, + 481, + 209 + ], + "blocks": [ + { + "bbox": [ + 135, + 148, + 481, + 209 + ], + "lines": [ + { + "bbox": [ + 135, + 148, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 135, + 148, + 481, + 209 + ], + "type": "table", + "html": "
MethodmCGCG
M1→OS128XT32→OS128OS128→M1XT32→M1OS128→XT32M1→XT32
ST3D [45]-16.72-9.2734.00-7.95-5.07-43.39-68.61
ST3D++ [46]09.2312.6554.4902.9917.2307.56-39.54
DTS [16]29.2248.2822.9209.0207.0934.5053.51
DIG(Ours)42.8954.4155.0933.6925.2249.8739.06
", + "image_path": "1061a7cec3ce8022d52c2ebb3eef3995bda8ca4c2b76570a3ce19d5727014677.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 190, + 209, + 421, + 218 + ], + "lines": [ + { + "bbox": [ + 190, + 209, + 421, + 218 + ], + "spans": [ + { + "bbox": [ + 190, + 209, + 421, + 218 + ], + "type": "text", + "content": "Table 3: Overall mean closed gap for baseline methods." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 253, + 480, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 253, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 480, + 338 + ], + "type": "text", + "content": "Discussion. Currently, commonly used CG is tailored for a single task, usually only considering the \"Car\" category [16], and cannot reflect the comprehensive performance of methods across various domain adaptation tasks. In contrast, mCG includes all MS LiDARs with varying beam numbers and SS LiDARs (i.e., OS128, XT32, M1), measuring performance from six experimental groups, thus providing the most comprehensive evaluation of cross-mechanism domain adaptation. See the appendix for 4D radar results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 357, + 281, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 281, + 368 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 281, + 368 + ], + "type": "text", + "content": "4.3 Results and Benchmark" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 377, + 480, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 480, + 496 + ], + "type": "text", + "content": "Domain adaptation baselines. We selected four baseline methods. Fig. 5 illustrates the similarities and differences between them. (1) Direct transfer (DT) means directly applying the source model to the target dataset. (2) ST3D [45] and " + }, + { + "bbox": [ + 130, + 377, + 480, + 496 + ], + "type": "inline_equation", + "content": "\\mathrm{ST3D + + }" + }, + { + "bbox": [ + 130, + 377, + 480, + 496 + ], + "type": "text", + "content": " [46] share a similar teacher-student architecture. The source model is trained with random object scaling, then used to generate pseudo-labels. The difference is that " + }, + { + "bbox": [ + 130, + 377, + 480, + 496 + ], + "type": "inline_equation", + "content": "\\mathrm{ST3D + + }" + }, + { + "bbox": [ + 130, + 377, + 480, + 496 + ], + "type": "text", + "content": " comes with a more effective denoising method to obtain more high-quality pseudo-labels. (3) DTS [16] introduced Random Beam Re-Sampling (RBRS) to enhance the robustness to varying beam densities. We reproduced the RBRS on CMD, providing representative baseline results. (4) Our proposed DIG." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 480, + 593 + ], + "type": "text", + "content": "Overall comparison results. Table 3 shows the overall results. Since the ST3D and " + }, + { + "bbox": [ + 130, + 498, + 480, + 593 + ], + "type": "inline_equation", + "content": "\\mathrm{ST3D + + }" + }, + { + "bbox": [ + 130, + 498, + 480, + 593 + ], + "type": "text", + "content": " are mostly designed for cross-geographical domain adaptation that highly relies on pseudo-label quality, the two detectors show lower performance in our cross-mechanism dataset. DTS takes density into consideration, thus outperforming " + }, + { + "bbox": [ + 130, + 498, + 480, + 593 + ], + "type": "inline_equation", + "content": "\\mathrm{ST3D + + }" + }, + { + "bbox": [ + 130, + 498, + 480, + 593 + ], + "type": "text", + "content": " by 19.99. Our DIG outperformed all previous methods by a large margin. Its mCG reached 42.89, surpassing ST3D, " + }, + { + "bbox": [ + 130, + 498, + 480, + 593 + ], + "type": "inline_equation", + "content": "\\mathrm{ST3D + + }" + }, + { + "bbox": [ + 130, + 498, + 480, + 593 + ], + "type": "text", + "content": ", and DTS by 59.61, 33.66, and 13.67 respectively, as our method considers all the typical domain disparities in the cross-mechanism domain adaptation problem." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "Performance discussion. A more detailed composition of mCG is shown in Table 4. We can get several conclusions from these results. (1) DIG exhibits notable superiority in the categories of pedestrians and bicycles, both across all distances and within a detection range of 30 meters. The similarity in these categories is the fewer points within the targets. We believe that sparse points make intensity the key factor for adapting to different domains. DIG's BCN module" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Deng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 114, + 482, + 449 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 482, + 449 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 482, + 449 + ], + "type": "table", + "html": "
TaskMethodCGmAPCarTruckPedCyc
OS128Oracle10026.4436.50/69.3417.45/38.5118.79/38.9933.0/69.22
XT32↓OS128DT006.2911.76/20.5904.61/08.2301.54/02.1507.25/08.90
ST3D [45]34.0013.1422.78/47.3910.41/15.1006.83/16.0112.55/33.05
ST3D++ [46]54.4917.2727.87/61.0213.50/22.6108.57/19.7618.95/49.32
DTS [16]22.9210.9114.26/49.4306.72/27.6709.29/20.9413.37/41.21
DIG(Ours)55.0917.3922.73/60.5913.60/30.1611.56/25.6721.70/57.46
M1↓OS128DT006.9213.23/30.2105.08/07.8601.98/03.6107.38/09.98
ST3D [45]-9.2705.1110.24/39.1104.06/11.3200.00/00.0006.14/21.18
ST3D++ [46]12.5609.3913.46/46.9306.61/23.1605.85/13.6811.65/35.06
DTS [16]48.2516.3421.60/58.9912.67/34.1409.57/23.4921.51/58.70
DIG(Ours)54.4117.5423.03/57.0809.71/25.9509.71/25.9525.00/66.13
M1Oracle10030.0942.01/70.6419.96/40.0117.54/36.9940.84/68.49
XT32↓M1DT013.3220.33/42.8209.47/19.8404.81/11.4718.67/36.06
ST3D [45]-5.0712.4730.34/55.4104.20/03.8200.19/00.2215.13/25.95
ST3D++ [46]17.2316.2132.92/61.0705.24/06.0402.18/04.1724.48/49.39
DTS [16]07.0914.5123.26/56.9408.78/28.0206.99/14.7219.01/48.44
DIG(Ours)25.2217.5527.19/64.7009.26/22.7905.72/12.5728.04/57.34
OS128↓M1DT014.3627.07/56.5409.91/24.2501.13/03.8119.32/33.49
ST3D [45]-7.9513.1128.71/60.3708.26/07.2506.37/17.4809.08/12.72
ST3D++ [46]02.9914.8332.98/68.4506.63/08.4507.28/19.4612.43/17.89
DTS [16]09.0215.7822.11/54.0111.84/30.2611.88/28.5817.27/40.40
DIG(Ours)33.6919.6626.85/62.8911.86/28.1113.69/32.2426.19/60.40
XT32Oracle10026.5936.84/72.4719.26/41.0116.23/32.7434.01/66.59
M1↓XT32DT014.0722.10/59.8610.44/21.4806.29/15.2417.44/45.69
ST3D [45]-68.6105.4811.39/38.0502.92/06.7800.00/00.0002.92/06.78
ST3D++ [46]-39.5409.1214.56/47.9407.47/22.0703.15/05.0211.28/32.18
DTS [16]53.5120.7730.62/68.0315.67/33.0109.48/21.6327.33/57.48
DIG(Ours)39.0618.9625.95/68.7814.77/34.6511.45/26.4723.66/60.92
OS128↓XT32DT011.7722.25/58.2008.45/19.1601.66/04.5014.71/41.05
ST3D [45]-43.9905.3411.46/45.4101.72/05.8700.00/00.0008.18/22.82
ST3D++ [46]07.5612.8917.81/56.9905.31/19.9810.69/24.2217.75/48.29
DTS [16]34.5016.0918.87/59.7813.63/27.7011.26/25.4920.59/58.19
DIG(Ours)49.8719.1627.20/63.8910.75/26.7911.76/27.1626.91/64.22
", + "image_path": "506c1c7d7b0555784d71efc2c3dbe85aa02d8f4ca8b41154aeab979115d59009.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 449, + 482, + 481 + ], + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 481 + ], + "type": "text", + "content": "Table 4: Detailed experimental results for all cross-mechanism settings. For each category, we provide the Average Precision (AP) results for all distances and within a range of 30 meters." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "aligns intensity well, which explains its strong performance in these cases. (2) The results indicate a significant decrease in mAP while using OS128 as the target domain. This decrease is primarily attributed to its unique intensity range of [0, 512], unlike the [0, 256] range of other sensors. The pre-trained model finds it challenging to handle unfamiliar intensities. Unlike the other three methods, which either discarded the intensity information or did not include a dedicated module for it, we addressed this issue with BCN, resulting in a significant performance improvement. (3) All these methods perform unsatisfactorily for domain adaptation from XT32 to M1. Even DIG, the best one among them, gets CGs of 25.22 on it. This observation means that M1 and XT32 may suffer from larger domain gaps, highlighting the need for further investigation in the future. (4) ST3D++ gets a CG of 54.49 for XT32 to OS128, which is mainly due to the" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 424, + 91, + 447, + 100 + ], + "type": "text", + "content": "CMD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "great performance on APs for 0-30m. As they are both MS LiDAR, they suffer more from differences in density. Pseudo-labels from XT32 might look like harder cases for OS128. Refining model parameters with these cases can dramatically improve the performance. (5) It indicates that DTS plays a crucial role when there is a substantial difference in the number of beams. CG on tasks that contain XT32 gets nice performance improvement. Specifically, CG can reach 34.51 and 53.51 for OS128 to XT32 and M1 to XT32 respectively." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 207, + 218, + 408, + 287 + ], + "blocks": [ + { + "bbox": [ + 207, + 218, + 408, + 287 + ], + "lines": [ + { + "bbox": [ + 207, + 218, + 408, + 287 + ], + "spans": [ + { + "bbox": [ + 207, + 218, + 408, + 287 + ], + "type": "table", + "html": "
modulesM1 ⇒ OS128OS128 ⇒ XT32
BDS BCN GAMmAPCGmAPCG
✓09.2211.7814.6919.70
✓17.1452.5316.4731.71
✓17.4754.0418.9548.45
✓17.5454.4119.1649.87
", + "image_path": "134d9a9a96269de707f8e83e1c464dd99c1f72930ddff5d90995b50c0ef35244.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 202, + 288, + 411, + 298 + ], + "lines": [ + { + "bbox": [ + 202, + 288, + 411, + 298 + ], + "spans": [ + { + "bbox": [ + 202, + 288, + 411, + 298 + ], + "type": "text", + "content": "Table 5: Ablation study for DIG baseline method." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 331, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 450 + ], + "type": "text", + "content": "Ablation study for DIG. We conduct ablation studies to analyze the effectiveness of each component in the DIG baseline method. Two representative tasks are selected. " + }, + { + "bbox": [ + 130, + 331, + 482, + 450 + ], + "type": "inline_equation", + "content": "M1 \\Rightarrow OS128" + }, + { + "bbox": [ + 130, + 331, + 482, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 331, + 482, + 450 + ], + "type": "inline_equation", + "content": "OS128 \\Rightarrow XT32" + }, + { + "bbox": [ + 130, + 331, + 482, + 450 + ], + "type": "text", + "content": " respectively refer to domain adaptation for sensors with different scanning patterns and different beams/densities. As shown in Table 5, For both tasks, BCN is anticipated to be the most effective module, contributing 52.53 and 31.71 CG on each respective task. This is attributed to the substantial intensity differences between the source and target domains in both tasks. BDS also exhibits commendable performance in both tasks, achieving 11.78 and 19.70 CG, respectively. Although less conspicuous, GAM proves to be effective for both tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 464, + 220, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 220, + 477 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 220, + 477 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 486, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 616 + ], + "type": "text", + "content": "We presented a dataset specifically designed for cross-mechanism domain adaptation, incorporating mechanical LiDARs with both high-res and low-res beams, solid-state LiDAR, and 4D millimeter-wave radar. These sensors undergo precise time synchronization, allowing them to simultaneously generate point clouds of the same external environment. Data from diverse scenes were selected and meticulously annotated. We believe the proposed dataset would hugely facilitate research of cross-mechanism 3D detection, as it currently stands as the most comprehensive point cloud 3D detection dataset in terms of sensor types. We also provide a novel DIG method as well as complete experimental results of recent methods on our dataset. They can be used as reliable baselines to further benefit the research communities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "In the future, we will continue to explore more usage of CMD, including utilizing data from multiple mechanisms as the source domain to enhance domain adaptation effects, and employing CMD as a tool for validating the performance of multi-modal 3D object detection." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Deng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "text", + "content": "1. IEEE standard for a precision clock synchronization protocol for networked measurement and control systems. IEEE Std 1588-2008 (Revision of IEEE Std 1588-2002) pp. 1-269 (2008). https://doi.org/10.1109/IEEESTD.2008.4579760" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 194 + ], + "type": "text", + "content": "2. Atkinson, A.C., Riani, M., Corbellini, A.: The box-cox transformation: Review and extensions (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 194, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 480, + 237 + ], + "type": "text", + "content": "3. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 238, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 238, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 238, + 480, + 259 + ], + "type": "text", + "content": "4. Bradski, G.: The opencv library. Dr. Dobb's Journal: Software Tools for the Professional Programmer 25(11), 120-123 (2000)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 260, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 260, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 260, + 480, + 315 + ], + "type": "text", + "content": "5. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (Jun 2020). https://doi.org/10.1109/cvpr42600.2020.01164, http://dx.doi.org/10.1109/cvpr42600.2020.01164" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 316, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 480, + 358 + ], + "type": "text", + "content": "6. Chang, M.F., Lambert, J., Sangkloy, P., Singh, J., Bak, S., Hartnett, A., Wang, D., Carr, P., Lucey, S., Ramanan, D., et al.: Argoverse: 3d tracking and forecasting with rich maps. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8748-8757 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 359, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 359, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 138, + 359, + 480, + 402 + ], + "type": "text", + "content": "7. Chen, Y., Yu, Z., Chen, Y., Lan, S., Anandkumar, A., Jia, J., Alvarez, J.M.: Focalformer3d: Focusing on hard instance for 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8394-8405 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "text", + "content": "8. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21674-21683 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 436, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 436, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 436, + 480, + 468 + ], + "type": "text", + "content": "9. Deng, J., Shi, S., Li, P., gang Zhou, W., Zhang, Y., Li, H.: Voxel r-cnn: Towards high performance voxel-based 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "type": "text", + "content": "10. Ding, G., Zhang, M., Li, E., Hao, Q.: Jst: Joint self-training for unsupervised domain adaptation on 2d&3d object detection. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 477-483. IEEE (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 502, + 480, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 523 + ], + "type": "text", + "content": "1. Fang, J., Zhou, D., Zhao, J., Tang, C., Xu, C.Z., Zhang, L.: Lidar-cs dataset: Lidar point cloud dataset with cross-sensors for 3d object detection (Jan 2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "text", + "content": "2. Fetic, A., Juric, D., Osmankovic, D.: The procedure of a camera calibration using camera calibration toolbox for matlab. In: 2012 Proceedings of the 35th International Convention MIPRO. pp. 1752-1757. IEEE (2012)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 555, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 599 + ], + "type": "text", + "content": "3. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition (Jun 2012). https://doi.org/10.1109/cvpr.2012.6248074, http://dx.doi.org/10.1109/cvpr.2012.6248074" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "text", + "content": "4. Hegde, D., Sindagi, V., Kilic, V., Cooper, A.B., Foster, M., Patel, V.: Uncertainty-aware mean teacher for source-free unsupervised domain adaptive 3d object detection. arXiv preprint arXiv:2109.14651 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 666 + ], + "type": "text", + "content": "5. Houston, J., Zuidhof, G., Bergamini, L., Ye, Y., Chen, L., Jain, A., Omari, S., Iglovikov, V., Ondruska, P.: One thousand and one hours: Self-driving motion prediction dataset. In: Conference on Robot Learning. pp. 409-418. PMLR (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 423, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 423, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 423, + 91, + 447, + 100 + ], + "type": "text", + "content": "CMD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "16. Hu, Q., Liu, D., Hu, W.: Density-insensitive unsupervised domain adaption on 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17556-17566 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "text", + "content": "17. Huang, X., Wu, H., Li, X., Fan, X., Wen, C., Wang, C.: Sunshine to rainstorm: Cross-weather knowledge distillation for robust 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 2409–2416 (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "type": "text", + "content": "18. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D., Han, S.: Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. ArXiv (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 480, + 257 + ], + "type": "text", + "content": "19. Luo, Z., Cai, Z., Zhou, C., Zhang, G., Zhao, H., Yi, S., Lu, S., Li, H., Zhang, S., Liu, Z.: Unsupervised domain adaptive 3d detection with multi-level consistency. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8866-8875 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 257, + 480, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 257, + 480, + 290 + ], + "spans": [ + { + "bbox": [ + 133, + 257, + 480, + 290 + ], + "type": "text", + "content": "20. MacQueen, J., et al.: Some methods for classification and analysis of multivariate observations. In: Proceedings of the fifth Berkeley symposium on mathematical statistics and probability. vol. 1, pp. 281-297. Oakland, CA, USA (1967)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 289, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 289, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 133, + 289, + 480, + 322 + ], + "type": "text", + "content": "21. Mao, J., Niu, M., Jiang, C., Liang, H., Liang, X., Li, Y., Ye, C., Zhang, W., Li, Z., Yu, J., Xu, H., Xu, C.: One million scenes for autonomous driving: Once dataset. Cornell University - arXiv, Cornell University - arXiv (Jun 2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 322, + 480, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 322, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 133, + 322, + 480, + 343 + ], + "type": "text", + "content": "22. Paek, D.H., Kong, S.H., Wijaya, K.: K-radar: 4d radar object detection dataset and benchmark for autonomous driving in various weather conditions (Jun 2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 343, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 343, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 133, + 343, + 480, + 363 + ], + "type": "text", + "content": "23. Palffy, A., Pool, E., Baratam, S., Kooij, J., Gavrila, D.: Multi-class road user detection with 3+1d radar in the view-of-delft dataset" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 363, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 363, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 133, + 363, + 480, + 396 + ], + "type": "text", + "content": "24. Peng, X., Zhu, X., Ma, Y.: Cl3d: Unsupervised domain adaptation for cross-lidar 3d detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2047-2055 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 396, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 396, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 133, + 396, + 480, + 418 + ], + "type": "text", + "content": "25. Raj, T., Hanim Hashim, F., Baseri Huddin, A., Ibrahim, M.F., Hussain, A.: A survey on lidar scanning mechanisms. Electronics 9(5), 741 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 418, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 418, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 133, + 418, + 480, + 449 + ], + "type": "text", + "content": "26. Rochan, M., Chen, X., Grandhi, A., Corral-Soto, E.R., Liu, B.: Domain adaptation in 3d object detection with gradual batch alternation training. arXiv preprint arXiv:2210.10180 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 449, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 449, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 133, + 449, + 480, + 471 + ], + "type": "text", + "content": "27. Roriz, R., Cabral, J., Gomes, T.: Automotive lidar technology: A survey. IEEE Transactions on Intelligent Transportation Systems 23(7), 6282-6297 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 471, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 471, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 133, + 471, + 480, + 492 + ], + "type": "text", + "content": "28. Segal, A., Haehnel, D., Thrun, S.: Generalized-icp. In: Robotics: science and systems. vol. 2, p. 435. Seattle, WA (2009)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 492, + 480, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 492, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 133, + 492, + 480, + 535 + ], + "type": "text", + "content": "29. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR). pp. 10526 - 10535 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 535, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 535, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 133, + 535, + 480, + 611 + ], + "type": "text", + "content": "30. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (Jun 2020). https://doi.org/10.1109/cvpr42600.2020.00252, http://dx.doi.org/10.1109/cvpr42600.2020.00252" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 611, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 611, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 133, + 611, + 480, + 633 + ], + "type": "text", + "content": "31. Team, O.D.: Openpcdet: An open-source toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "text", + "content": "32. Tsai, D., Berrio, J.S., Shan, M., Nebot, E., Worrall, S.: Ms3d: Leveraging multiple detectors for unsupervised domain adaptation in 3d object detection. arXiv preprint arXiv:2304.02431 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Deng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "33. Tsai, D., Berrio, J.S., Shan, M., Nebot, E., Worrall, S.: Viewer-centred surface completion for unsupervised domain adaptation in 3d object detection. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 9346-9353. IEEE (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "type": "text", + "content": "34. Tsai, D., Berrio, J.S., Shan, M., Worrall, S., Nebot, E.: See eye to eye: A lidar-agnostic 3d detection framework for unsupervised multi-target domain adaptation. IEEE Robotics and Automation Letters 7(3), 7904-7911 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "type": "text", + "content": "35. Wang, Y., Chen, X., You, Y., Li, L.E., Hariharan, B., Campbell, M., Weinberger, K.Q., Chao, W.L.: Train in germany, test in the usa: Making 3d object detectors generalize. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11713-11723 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "text", + "content": "36. Wang, Y., Yin, J., Li, W., Frossard, P., Yang, R., Shen, J.: Ssa3d: Semi-supervised domain adaptation for 3d object detection from point cloud. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2707-2715 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 314 + ], + "type": "text", + "content": "37. Wang, Z., Ding, S., Li, Y., Fenn, J., Roychowdhury, S., Wallin, A., Martin, L., Ryvola, S., Sapiro, G., Qiu, Q.: Cirrus: A long-range bi-pattern lidar dataset. In: 2021 IEEE International Conference on Robotics and Automation (ICRA). pp. 5744-5750. IEEE (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "text", + "content": "38. Wei, Y., Wei, Z., Rao, Y., Li, J., Zhou, J., Lu, J.: Lidar distillation: Bridging the beam-induced domain gap for 3d object detection. In: European Conference on Computer Vision. pp. 179-195. Springer (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "type": "text", + "content": "39. Wu, H., Deng, J., Wen, C., Li, X., Wang, C.: Casa: A cascade attention network for 3d object detection from lidar point clouds. IEEE Transactions on Geoscience and Remote Sensing (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 381, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 381, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 381, + 481, + 402 + ], + "type": "text", + "content": "40. Wu, H., Wen, C., Li, W., Yang, R., Wang, C.: Learning transformation-equivariant features for 3d object detection (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "text", + "content": "41. Wu, H., Wen, C., Shi, S., Li, X., Wang, C.: Virtual sparse convolution for multimodal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21653-21662 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 436, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 481, + 468 + ], + "type": "text", + "content": "42. Xia, Q., Chen, Y., Cai, G., Chen, G., Xie, D., Su, J., Wang, Z.: 3-d hanet: A flexible 3-d heatmap auxiliary network for object detection. IEEE Transactions on Geoscience and Remote Sensing 61, 1-13 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 512 + ], + "type": "text", + "content": "43. Xia, Q., Deng, J., Wen, C., Wu, H., Shi, S., Li, X., Wang, C.: Coin: Contrastive instance feature mining for outdoor 3d object detection with very limited annotations. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6254-6263 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 512, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 566 + ], + "type": "text", + "content": "44. Xiao, P., Shao, Z., Hao, S., Zhang, Z., Chai, X., Jiao, J., Li, Z., Wu, J., Sun, K., Jiang, K., Wang, Y., Yang, D.: Pandaset: Advanced sensor suite dataset for autonomous driving. In: 2021 IEEE International Intelligent Transportation Systems Conference (ITSC) (Sep 2021). https://doi.org/10.1109/itsc48978.2021.9565009, http://dx.doi.org/10.1109/itsc48978.2021.9565009" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "text", + "content": "45. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d: Self-training for unsupervised domain adaptation on 3d object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10368-10378 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "text", + "content": "46. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d++: Denoised self-training for unsupervised domain adaptation on 3d object detection. IEEE transactions on pattern analysis and machine intelligence 45(5), 6354-6371 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "text", + "content": "47. Yang, Z., Sun, Y., Liu, S., Jia, J.: 3dssd: Point-based 3d single stage object detector. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR). pp. 11040-11048 (2020)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 423, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 423, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 423, + 91, + 447, + 100 + ], + "type": "text", + "content": "CMD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 304 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "48. Yihan, Z., Wang, C., Wang, Y., Xu, H., Ye, C., Yang, Z., Ma, C.: Learning transferable features for point cloud detection via 3d contrastive co-training. Advances in Neural Information Processing Systems 34, 21493-21504 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 182 + ], + "type": "text", + "content": "49. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 226 + ], + "type": "text", + "content": "50. You, Y., Diaz-Ruiz, C.A., Wang, Y., Chao, W.L., Hariharan, B., Campbell, M., Weinbergert, K.Q.: Exploiting playbacks in unsupervised domain adaptation for 3d object detection in self-driving cars. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 5070-5077. IEEE (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 226, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 226, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 226, + 481, + 270 + ], + "type": "text", + "content": "51. You, Y., Phoo, C.P., Luo, K., Zhang, T., Chao, W.L., Hariharan, B., Campbell, M., Weinberger, K.Q.: Unsupervised adaptation from repeated traversals for autonomous driving. Advances in Neural Information Processing Systems 35, 27716-27729 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 270, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 304 + ], + "type": "text", + "content": "52. Zheng, L., Ma, Z., Zhu, X., Tan, B., Li, S., Long, K., Sun, W., Chen, S., Zhang, L., Wan, M., Huang, L., Bai, J.: Tj4dradset: A 4d radar dataset for autonomous driving (Apr 2022)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "J. Deng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CMTA_ Cross-Modal Temporal Alignment for Event-guided Video Deblurring/7cf7582c-c0da-4c34-8883-80e1b8a2e3ef_content_list.json b/2024/CMTA_ Cross-Modal Temporal Alignment for Event-guided Video Deblurring/7cf7582c-c0da-4c34-8883-80e1b8a2e3ef_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dfcbd03907ebedea6cbf9a9571a721a2b24d617f --- /dev/null +++ b/2024/CMTA_ Cross-Modal Temporal Alignment for Event-guided Video Deblurring/7cf7582c-c0da-4c34-8883-80e1b8a2e3ef_content_list.json @@ -0,0 +1,1837 @@ +[ + { + "type": "text", + "text": "CMTA: Cross-Modal Temporal Alignment for Event-guided Video Deblurring", + "text_level": 1, + "bbox": [ + 241, + 140, + 759, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Taewoo Kim $^{1}$ , Hoonhee Cho $^{1}$ , and Kuk-Jin Yoon $^{1}$", + "bbox": [ + 307, + 212, + 696, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Korea Advanced Institute of Science and Technology {intelpro, gngsgsm1, kjyoon}@kaist.ac.kr", + "bbox": [ + 326, + 239, + 676, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Video deblurring aims to enhance the quality of restored results in motion-blurred videos by effectively gathering information from adjacent video frames to compensate for the insufficient data in a single blurred frame. However, when faced with consecutively severe motion blur situations, frame-based video deblurring methods often fail to find accurate temporal correspondence among neighboring video frames, leading to diminished performance. To address this limitation, we aim to solve the video deblurring task by leveraging an event camera with microsecond temporal resolution. To fully exploit the dense temporal resolution of the event camera, we propose two modules: 1) Intra-frame feature enhancement operates within the exposure time of a single blurred frame, iteratively enhancing cross-modality features in a recurrent manner to better utilize the rich temporal information of events, 2) Interframe temporal feature alignment gathers valuable long-range temporal information to target frames, aggregating sharp features leveraging the advantages of the events. In addition, we present a novel dataset composed of real-world blurred RGB videos, corresponding sharp videos, and event data. This dataset serves as a valuable resource for evaluating event-guided deblurring methods. We demonstrate that our proposed methods outperform state-of-the-art frame-based and event-based motion deblurring methods through extensive experiments conducted on both synthetic and real-world deblurring datasets. The code and dataset are available at https://github.com/intelpro/CMTA.", + "bbox": [ + 263, + 305, + 738, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Temporal alignment $\\cdot$ Video Deblurring $\\cdot$ Event cameras", + "bbox": [ + 263, + 637, + 723, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 678, + 374, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Motion blur is a common artifact caused by dynamic movements within a scene or camera motion during exposure. Motion deblurring, which aims to reverse the blurring process, presents significant challenges due to variations in blur intensity influenced by scene structure and depth. To achieve high-quality deblurring, video deblurring has emerged, leveraging information from neighboring frames instead of relying solely on a single blurred image. However, identifying temporal correspondence between blurred video frames becomes challenging with", + "bbox": [ + 217, + 710, + 785, + 815 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 220, + 825, + 364, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "extreme motion blur, hindering the extraction of valuable information from adjacent frames and impeding performance improvement.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Event cameras [9], with their extremely low latency (on the order of microseconds), can offer high-quality guidance for motion deblurring due to their ability to capture high-temporal resolution of brightness change. To effectively utilize the advantages of the events, several event-guided motion deblurring works [15, 36, 50] have been introduced. While these works have typically explored cross-modal feature fusion methods across different modalities, there has been limited work on leveraging the abundant temporal information in videos.", + "bbox": [ + 212, + 179, + 785, + 285 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To obtain high-quality results, we emphasize the event camera's temporal continuity, focusing on its interaction with video frames that exhibit long-term temporal dependencies. Unlike previous event-guided motion deblurring works [36,43,50], which relied on a single blurry image and corresponding events on its exposure time, we further design precise temporal feature alignment methods between neighboring video frames by leveraging the advantages of event data. Specifically, we propose novel modules from two perspectives: intra-frame (interaction between events and frames within the exposure time) and inter-frame (interaction between different frames) perspectives.", + "bbox": [ + 212, + 287, + 787, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "From an intra-frame perspective within exposure time, we propose a Cross-modal Recurrent Intra-frame Feature Enhancement (CRIFE) module to better leverage the rich temporal information of the events by mutually interacting the blur frame and event features within the duration of exposure time. In this module, we perform recurrent attention-based feature enhancement using a transformer [38] that better captures long-range pixel dependencies.", + "bbox": [ + 212, + 426, + 787, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In temporal feature alignment with the second perspective, we propose a novel event-guided temporal feature alignment module, effectively leveraging rich temporal characteristics of the events. Conventional frame-based temporal feature alignment methods rely on optical flow [28] or deformable convolution [5,58]. While optical flow and deformable convolutions aid in achieving temporal feature alignment, the high computational complexity of these operations makes it challenging to execute them at a higher spatial scale. Therefore, video frame alignment, generally conducted at lower spatial scales, leads to sub-optimal deblurring results due to the lack of spatial contexts of features. To overcome these limitations, our temporal feature alignment module avoids relying on optical flow or deformable convolution by effectively leveraging the temporally dense advantages of the events. Therefore, we can effectively perform temporal feature alignment across multiple visual scales as we do not rely on these complex operations but rather efficiently leverage the temporal information from the events. As a result, our temporal feature alignment module demonstrates a significant performance improvement in event-guided video deblurring tasks. Since we have introduced a pioneering method for aligning temporal features using the advantages of the events, it is expected to be effectively applicable to various event-guided video restoration tasks (e.g., event-guided video super-resolution).", + "bbox": [ + 212, + 520, + 787, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, we propose a novel video deblurring dataset, the EVRB dataset, composed of high-quality RGB and event data. It consists of real-world blurry", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Kim et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "videos generated by extreme motion and corresponding sharp videos for generalized applications in real-world scenarios. The network trained on the EVRB dataset can be directly applicable to real-world scenarios, making it a valuable resource for event-guided deblurring research.", + "bbox": [ + 218, + 146, + 784, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 218, + 229, + 393, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Video Deblurring", + "text_level": 1, + "bbox": [ + 218, + 261, + 403, + 276 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Early works [1, 35] employed CNNs that take the concatenation of adjacent frames as input to address video deblurring. Subsequently, to better leverage temporal information, approaches have emerged that employ 3D convolutions [26, 47], temporal alignment modules with deformable convolutions [39] and optical flow [14, 25, 40, 46]. As an alternative video alignment method, some works utilize RNN [11, 42, 53] and transformer [18, 19] structures to propagate information from long-range video frames. However, as the intensity of motion blur in the video increases, frame-based video alignment methods struggle to achieve accurate video alignment, leading to sub-optimal deblurring results.", + "bbox": [ + 218, + 286, + 784, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Event-guided Motion Deblurring", + "text_level": 1, + "bbox": [ + 218, + 445, + 531, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An event camera can effectively be used for motion deblurring as it records motion information corresponding to the brightness differences with high temporal resolution. Efforts to utilize event cameras for motion deblurring [8, 13, 15, 20, 27, 36, 37, 43, 48-50] have been actively ongoing. Recent studies have focused on effectively fusing event and image features of different modalities. To this end, Sun et al. [36] employed a transformer architecture for feature fusion. Zhang et al. [50] effectively combined modalities using a multi-scale architecture. In addition, there are also attempts to address motion deblurring by assuming challenging and general scenarios [8, 15, 33]. However, these studies have primarily focused on single image deblurring and do not exploit the long-range temporal information demonstrated in previous video deblurring tasks. To effectively acquire information that may be missing from sparse events, we introduce a novel method for accurate temporal alignment with events, utilizing information from surrounding adjacent frames in video deblurring.", + "bbox": [ + 218, + 470, + 784, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Event-based Video Deblurring Dataset for Real-world Blur", + "text_level": 1, + "bbox": [ + 218, + 704, + 771, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 Limitation of Synthetic Blur Dataset", + "text_level": 1, + "bbox": [ + 218, + 753, + 563, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Typically, the blurred images synthesis procedure adheres to the methodology outlined in [22-24, 34, 35], involving the averaging of consecutive video frames within a fixed-size window. However, as recent studies [31, 52, 54] have discussed, blur synthesis based on discrete signals may result in shutter artifacts even when", + "bbox": [ + 218, + 779, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Cross-Modal Temporal Alignment for Event-guided Video Deblurring", + "bbox": [ + 269, + 114, + 730, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2aa45a1d09c9f77a410f2e52ea77c3c95b8fe3b80ad3c1d1f948fb53b0bd39b5.jpg", + "image_caption": [ + "Fig. 1: Illustration of a hybrid camera system for real-world event-based video deblurring dataset. $S$ and $B$ denote the cameras for acquiring sharp and blur videos, respectively. (a): The triple-axis camera system to capture real-world blur. (b): A diagram of our hybrid camera system. (c): Samples from our EVRB dataset with natural blur." + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 380, + 291 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c7c93452eb3c45e081aef27acf4743df0783247ecf107f0b49cbb9922989e105.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 143, + 566, + 292 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1876ae1d4afcaad698acabc3cd91eebeb08d9d35ee932ea9a4bab4e377375eda.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 568, + 143, + 771, + 294 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "averaging high-frame-rate videos. Furthermore, simply averaging sharp images disregards essential elements in the blurred images such as pixel saturation [30], limitations due to dynamic range, and physical noise [41, 51] in data acquired during exposure time. Furthermore, in the case of the event-guided motion deblurring research, existing works have enhanced the synthetic aspects by using event simulator [29]. To address these limitations, recent works have attempted to acquire real-world blur datasets using hybrid camera systems [3, 31, 52-54]. Meanwhile, to generalize event-based motion deblurring, Sun et al. [36] introduced the ReBlur dataset acquired in a high-precision optical laboratory using an electronic-controlled slide-rail system. However, despite these successes, the ReBlur dataset has limitations due to its indoor setting, a lack of dynamic objects, and the absence of high-quality RGB data as it is acquired using the DAVIS sensor [2]. Additionally, the sequences are relatively short(minimum six blurred frames in the sequence), making it challenging for use in event-guided video deblurring. We have acquired a new EVRB dataset for evaluating event-guided image and video deblurring methods in real-world blurry videos. The EVRB dataset was captured using a customized hybrid system consisting of two RGB cameras and one event camera. This system encompasses a range of blur magnitude from slight to extreme in various urban environments.", + "bbox": [ + 212, + 402, + 787, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 Triple-axis Hybrid Camera System", + "text_level": 1, + "bbox": [ + 214, + 710, + 549, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 1, we design a hybrid camera system to enable the acquisition of different data sources to be geometrically aligned. Two RGB cameras and one event camera are geometrically aligned using two $50/50$ cube beam splitters, resulting in a minimal baseline. We perform pixel-wise alignment between multiple cameras based on the homography calculated by extrinsic calibration for precise alignment. For the synchronization of the multiple cameras, we use an external trigger system. Each camera receives the falling and rising edges of the trigger", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Kim et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bb7afde7f4dc31ede6443dc3b7f34963307980eabe143e38a3326855983ba012.jpg", + "image_caption": [ + "Fig. 2: Overall framework of CMTA is divided into two main components: Cross-modal Recurrent Intra-frame Feature Enhancement (CRIFE) and Event-guided Cascaded Inter-frame Temporal Feature Alignment (ECITFA). $s$ is the scale factor for multi-scale features. In the figure of the ECITFA module, the description was performed for the case of $P = 2$ for simplicity." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "signal and acquires data synchronized with the period of the signal. This external trigger can precisely control the exposure times of the two RGB cameras, enabling us to capture paired sharp and blurred video frames. Furthermore, for photometric alignment, we physically adjust the amount of incoming light for both cameras to equalize the total irradiance of the multiple cameras using a neutral density filter.", + "bbox": [ + 212, + 460, + 787, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 215, + 573, + 330, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 604, + 341, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overview of the proposed framework is illustrated in Fig. 2. Given consecutive blurred video frames $\\{B_k\\}$ and sets of event streams corresponding to the exposure time of each video frame $\\{\\mathbb{E}_k\\}$ , where $k \\in \\{t - P, \\dots, t, \\dots, t + P\\}$ , our goal is to estimate the latent sharp video frame $S_t$ . To utilize the event stream $\\{\\mathbb{E}_k\\}$ corresponding to the exposure time of $\\{B_k\\}$ as the input for the networks, we first perform embedding using the event voxel grid representation [56] for the event stream $\\{\\mathbb{E}_k\\}$ , resulting in $\\{E_k\\}$ . Our framework consists of two main submodules: (1) Cross-modal Recurrent Intra-frame Feature Enhancement (CRIFE) module and (2) Event-guided Cascaded Inter-frame Temporal Feature Alignment (ECITFA) module. In the first module, we perform cross-modal feature enhancement through recurrent interactions between blurred frame features and event features to leverage the continuous temporal information of events within the exposure time. After obtaining the fused feature $\\{\\mathcal{G}_k\\}$ from the first module, we generate multi-scale features, $\\{\\mathcal{F}_k^s\\}_{s=0}^2$ , through a weight-shared pyra", + "bbox": [ + 212, + 628, + 787, + 843 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Cross-Modal Temporal Alignment for Event-guided Video Deblurring", + "bbox": [ + 269, + 114, + 730, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e417d8010bde286557f1bbfc66ea431af1ce7ad9ff133596eac8865891e6b4b2.jpg", + "image_caption": [ + "Fig. 3: Illustration of Cross-modal Recurrent Intra-frame Feature Enhancement (CRIFE)." + ], + "image_footnote": [], + "bbox": [ + 272, + 145, + 733, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mid encoder. Subsequently, for the temporal feature alignment stage, we encode multi-scale event features, $\\{\\varepsilon_{[k,k + 1]}^s\\}_{s = 0}^2$ , by grouping two consecutive events corresponding to the exposure times of each frame to connect adjacent frames. By utilizing the multi-scale pyramid features $\\{\\mathcal{F}_k^s\\}_{s = 0}^2$ and the encoded event feature $\\{\\varepsilon_{[k,k + 1]}^s\\}_{s = 0}^2$ , we perform temporal feature alignment. Afterward, temporally aligned feature pyramids are fed into the U-Net [32]-based decoder, generating the final sharp video frame $S_{t}$ . Our framework can be extended to cases where $P$ is an arbitrary positive number. However, for the sake of brevity, we will explain it in the main text with $P = 2$ .", + "bbox": [ + 214, + 414, + 787, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Cross-modal Recurrent Intra-frame Feature Enhancement", + "text_level": 1, + "bbox": [ + 214, + 574, + 740, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since event cameras provide rich temporal information on brightness changes, it is crucial to effectively utilize this dense temporal information of the events within the duration of the exposure time. Typically, event data captured during the exposure time is transformed into event embeddings [10,56] as input to the network, followed by feature extraction using 2D CNNs. However, these approaches cannot effectively utilize the continuous temporal information of events, and these limitations can impact the performance of event-guided video deblurring. To address these limitations, we propose a method that leverages the rich temporal nature of events and fuses event and blurred frame features using recurrent-based attention methods.", + "bbox": [ + 212, + 598, + 787, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Recently, researchers have demonstrated the effectiveness of transformer-based architectures [38] by capturing long-range pixel dependencies, which have been proven highly effective in various vision tasks. We leverage transformers' advantages to more effectively utilize the temporal benefits of the events through a recurrent-based approach. For each blur video frame index $k \\in \\{t - 2, \\dots, t + 2\\}$ , we initially partition the exposure time $T_{exp,k}$ into $N$ unit temporal intervals,", + "bbox": [ + 212, + 750, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Kim et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "denoted as $\\Delta t$ , with $\\Delta t = T_{exp,k} / N$ . Based on the time interval $\\Delta t$ , we divide the event voxel grid within the exposure time, denoted as $E_k \\in \\mathbb{R}^{C \\times H \\times W}$ into $N$ temporally divided event voxel grids $\\{E_k^n\\}$ , where $E_k^n \\in \\mathbb{R}^{C / N \\times H \\times W}$ and $n \\in \\{0, \\dots, N - 1\\}$ . After, we extract event features $\\{\\mathcal{F}(E)_k^n\\}$ of the temporally divided event voxel grids $E_k^n$ using weight-sharing event feature extractor. To apply cross-attention, we construct a query encompassing global information of blur and temporally divided event features. That is, we concatenate the blur frame feature $\\mathcal{F}(B)_k$ with the temporally segmented event sets $\\{\\mathcal{F}(E)_k^n\\}$ and then extract the feature for the encoding query information.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Q} _ {k} = \\operatorname {C o n v} _ {\\mathrm {p}} (\\mathcal {F} (B) _ {k} \\| \\left\\{\\mathcal {F} (E) _ {k} ^ {n} \\right\\}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 309, + 785, + 325 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {k} ^ {n = 0} = F _ {R} \\left(\\mathcal {Q} _ {k}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 444, + 328, + 784, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{Conv}_{\\mathrm{p}}$ represents point-wise convolution, $F_{R}$ consists of a sequence of blocks with $3\\times 3$ convolution and a stride of 2, along with ResBlocks. Additionally, $Q_{k}^{n = 0}$ denotes the initial query feature for the recurrent-based attention method.", + "bbox": [ + 212, + 354, + 787, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As illustrated in Fig. 3, we perform cross-attention to update query information iteratively. To find the key and value of cross-attention for query updating, we recursively input temporally separated event features, helping to better utilize the events' rich temporal information. That is, for iteration $n$ , we extract features to be used as keys and values by concatenating the previously updated $Q_{k}^{n}$ with temporally divided $n$ -th event features $\\mathcal{F}(E)_k^n$ as follows:", + "bbox": [ + 212, + 415, + 787, + 507 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {K V} _ {k} ^ {n} = F _ {K V} \\left(Q _ {k} ^ {n} \\| \\mathcal {F} (E) _ {k} ^ {n}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 517, + 784, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\parallel$ channel-wise concatenation and $\\mathrm{KV}_k^n$ denotes the output features for key and value projection, and $F_{KV}$ refers to the ResBlocks layer. We then construct $K_k^n$ , $V_k^n$ as $K_k^n = W^K (\\mathrm{KV}_k^n)$ and $V_{k}^{n} = W^{V}(\\mathrm{KV}_{k}^{n})$ where $W^{(\\cdot)}$ denote $1\\times 1$ convolution layer. Then, attention can be calculated as:", + "bbox": [ + 214, + 542, + 787, + 604 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t n} _ {\\mathrm {k}} ^ {\\mathrm {n}} = \\operatorname {S o f t M a x} \\left(\\frac {Q _ {k} ^ {n} \\left(K _ {k} ^ {n}\\right) ^ {T}}{\\alpha}\\right) V _ {k} ^ {n} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 613, + 785, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\alpha$ is learnable scaling parameter to balance attention weights and $\\mathrm{Att}\\mathfrak{n}_{\\mathrm{n}}$ is outputs of cross-attention at iteration $n$ . To calculate cross-covariance matrix efficiently, we adopted transposed attention [44] for efficient computations. This method enables the efficient computation of attention values at high resolutions by calculating the cross-covariance matrix along the channel axis, leading to a complexity of $O(C^2)$ . Moreover, we reduced the number of channels when encoding step, enabling even more efficient operations. We use the attention value $\\mathrm{Att}\\mathfrak{n}_{\\mathrm{k}}^{\\mathrm{n}}$ to update the query iteratively as follows:", + "bbox": [ + 212, + 651, + 787, + 773 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {k} ^ {n + 1} = Q _ {k} ^ {n} + \\operatorname {A t t n} _ {\\mathrm {k}} ^ {\\mathrm {n}} + \\operatorname {M L P} \\left(\\operatorname {A t t n} _ {\\mathrm {k}} ^ {\\mathrm {n}}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 782, + 785, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where MLP denote multi-layer perceptron. After $N$ iterations, we obtain the updated query feature, $Q_{k}^{N}$ . The final query feature passes through up-sampling", + "bbox": [ + 214, + 809, + 787, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Cross-Modal Temporal Alignment for Event-guided Video Deblurring", + "bbox": [ + 267, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and skip connections. Then, we generate the final fused features $\\{\\mathcal{G}_k\\}$ as $\\mathcal{G}_k = \\mathcal{Q}_k + \\mathrm{Dconv}_{4\\times 4}(Q_k^N)$ where $\\mathrm{Dconv}_{4\\times 4}$ denote deconvolution layer with a kernel size of 4. As illustrated in Fig. 2, the fused features $\\{\\mathcal{G}_k\\}$ pass through a pyramid encoder, leading to the creation of multi-scale pyramid features, $\\{\\mathcal{F}_k^s\\}_{s=0}^2$ . Through the CRIFE module, we can leverage the advantages of abundant temporal information on the events within the exposure time.", + "bbox": [ + 215, + 146, + 787, + 238 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 Event-guided Cascaded Inter-frame Temporal Feature Alignment", + "text_level": 1, + "bbox": [ + 215, + 255, + 712, + 286 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Temporal feature alignment aims to extract valuable information from adjacent video frames. Conventional frame-based video deblurring methods face challenges from inaccurate motion estimation due to motion blur. Conversely, event cameras, thanks to their resistance to motion blur, can offer valuable guidance for aligning video frames. The most straightforward way to align adjacent video frames using the events is to utilize event features to estimate optical flows or deformable offsets between neighboring frames [4,5,58]. While this approach can leverage the advantages of event data for offset and optical flow estimation, it is typically employed after spatial down-sampling due to the high computational costs associated with deformable convolutions [58] and optical flows [28]. Therefore, this approach could restrict performance as it inherently hinders the processing of features at high spatial resolutions in network architectures, limiting access to information across multiple visual scales. To address the aforementioned limitations, we propose a new temporal alignment module that combines the advantages of multiple-visual scale pyramids, leveraging the events' rich temporal contexts.", + "bbox": [ + 212, + 292, + 787, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, our proposed temporal feature alignment modules are structured as multi-level networks that gradually perform coarse-to-fine feature alignment. First, we extract event features, incorporating exposure time information from neighboring video frames to reference frames to facilitate video feature alignment. Specifically, we encode event features encompassing the exposure time interval between $t$ and $t + 1$ , allowing us to connect the frame at time $t$ with the frame at time $t + 1$ . Through this event encoding step, we obtain event feature pyramid set $\\{\\varepsilon_{[m,m + 1]}^s\\}_{s = 0}^2$ where $m \\in \\{t - 2,\\dots,t + 1\\}$ . We use the event feature pyramid containing motion information for adjacent times for the alignment of each blur frame feature pyramid $\\{\\mathcal{F}_k^s\\}_{s = 0}^2$ where $k \\in \\{t - 2,\\dots,t + 2\\}$ .", + "bbox": [ + 212, + 534, + 787, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "After the event encoding step for the feature alignment, we gradually perform temporal feature alignment from the bottom pyramid level (scale factor $s$ of 2) to the top pyramid level ( $s$ of 0). Specifically, as shown in left side of Fig. 4, in each pyramid level $s$ ( $s \\in \\{0,1,2\\}$ ), We first upsample the hidden state of temporally aligned features at the previous scale, $\\hat{\\mathcal{F}}_i^{s+1}$ through a deconvolution layer and receive them as inputs of alignment module, resulting in $h_i^s$ .", + "bbox": [ + 212, + 686, + 787, + 777 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i} ^ {s} = \\operatorname {D c o n v} _ {4 \\times 4} \\left(\\hat {\\mathcal {F}} _ {i} ^ {s + 1}\\right), i \\in \\{t - 1, t, t + 1 \\} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 351, + 784, + 784, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\hat{\\mathcal{F}}_i^{s + 1}$ denote aligned feature at previous scale $s + 1$ , $\\mathrm{Dconv}_{4\\times 4}$ denote $4\\times 4$ deconvolution layer. Note that there is no hidden state at the bottom pyramid", + "bbox": [ + 212, + 808, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Kim et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/359efd6ea5ed3303e309257429825178ab57665602d237ce7dfecf55fa139eb2.jpg", + "image_caption": [ + "Fig. 4: Overview of the Event-guided Cascaded Inter-frame Temporal Feature Alignment (ECITFA). The left figure illustrates temporal alignment for scale $s$ . The key module for each alignment procedure, Cross-modal Temporal Feature Alignment (CTFA) at time $t$ , is illustrated on the right of the figure. The CTFA module operates similarly for reference times $t - 1$ and $t + 1$ as well." + ], + "image_footnote": [], + "bbox": [ + 228, + 145, + 480, + 303 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/20408f853ddba7bd47faf5758dcf8bdd8213f38f17c04faa81a84a5b9e0255b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 143, + 772, + 309 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "level ( $s$ of 2) since there are no features aligned at the previous scale, and for brevity, we focus on the case when $s < 2$ .", + "bbox": [ + 212, + 431, + 784, + 460 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As depicted in the left side of Fig. 4, we progressively group three blurred video frames when performing temporal feature alignment. In other words, we first align with respect to $t - 1$ using $\\{t - 2, t - 1, t\\}$ and simultaneously align with respect to $t + 1$ using $\\{t, t + 1, t + 2\\}$ features. Subsequently, we perform the final temporal feature alignment for the previously aligned $t - 1$ and $t + 1$ with the last target time $t$ . More specifically, when performing frame alignment for the time step $t - 1$ , we first group the three blur video frame features $\\mathcal{F}_{t - 2}^s, \\mathcal{F}_{t - 1}^s, \\mathcal{F}_t^s$ and perform alignment. In this case, we make use of two event features $\\varepsilon_{[t - 1,t]}^s$ and $\\varepsilon_{[t,t + 1]}^s$ , which respectively contain motion information between $t - 1$ and $t$ , and between $t$ and $t + 1$ , respectively.", + "bbox": [ + 212, + 462, + 787, + 613 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {F}} _ {t - 1} ^ {s} = \\mathrm {C T F A} ^ {s} (\\mathcal {F} _ {t} ^ {s}, \\mathcal {F} _ {t - 1} ^ {s}, \\mathcal {F} _ {t - 2} ^ {s}, h _ {t - 1} ^ {s}, \\varepsilon_ {[ t - 2, t - 1 ]} ^ {s}, \\varepsilon_ {[ t - 1, t ]} ^ {s}), \\qquad (7)\n$$\n", + "text_format": "latex", + "bbox": [ + 300, + 623, + 785, + 646 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathrm{CTFA}^s$ denote unit Cross-modal Temporal Feature Alignment (CTFA) module at scale factor $s$ . Across the same scale factor $s$ , CTFA modules are weight-shared. Similarly, we perform alignment for the time step $t + 1$ as:", + "bbox": [ + 212, + 651, + 787, + 696 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {F}} _ {t + 1} ^ {s} = \\mathrm {C T F A} ^ {s} (\\mathcal {F} _ {t} ^ {s}, \\mathcal {F} _ {t + 1} ^ {s}, \\mathcal {F} _ {i + 1} ^ {s}, h _ {t + 1} ^ {s}, \\varepsilon_ {[ t + 1, t + 2 ]} ^ {s}, \\varepsilon_ {[ t, t + 1 ]} ^ {s}). \\qquad (8)\n$$\n", + "text_format": "latex", + "bbox": [ + 302, + 707, + 785, + 729 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Finally, we utilize the temporally aligned results $\\hat{\\mathcal{F}}_{t + 1}^{s}$ and $\\hat{\\mathcal{F}}_{t - 1}^{s}$ to once again group with $\\mathcal{F}_t^s$ and perform alignment:", + "bbox": [ + 212, + 739, + 785, + 771 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {F}} _ {t} ^ {s} = \\mathrm {C T F A} ^ {s} (\\hat {\\mathcal {F}} _ {t - 1} ^ {s}, \\mathcal {F} _ {t} ^ {s}, \\hat {\\mathcal {F}} _ {t + 1} ^ {s}, h _ {t} ^ {s}, \\varepsilon_ {[ t - 1, t ]} ^ {s}, \\varepsilon_ {[ t, t + 1 ]} ^ {s}). \\qquad (9)\n$$\n", + "text_format": "latex", + "bbox": [ + 325, + 780, + 785, + 801 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Through this cascaded feature alignment stage, we effectively propagate non-local video frame information, facilitating temporal feature alignment for both", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Cross-Modal Temporal Alignment for Event-guided Video Deblurring", + "bbox": [ + 269, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "non-local and adjacent video frames. All the results of alignment $\\hat{\\mathcal{F}}_{t - 1}^{s},\\hat{\\mathcal{F}}_{t}^{s},\\hat{\\mathcal{F}}_{t + 1}^{s}$ are passed to the next scale, $s - 1$ . The right side of Fig. 4 illustrates the overall alignment process for the time step $t$ in the CTFA module. Since it is applied similarly for all time steps $i\\in \\{t - 1,t,t + 1\\}$ , we will describe it here specifically when $i = t$ , where previously aligned $\\hat{\\mathcal{F}}_{t - 1}^{s}$ and $\\hat{\\mathcal{F}}_{t + 1}^{s}$ are provided. For leveraging the benefits of coarse-to-fine architecture, we first fuse hidden state aligned feature $h_t^s$ from the previous scale for the time step $t$ at scale factor $s$ , as follows:", + "bbox": [ + 212, + 143, + 787, + 253 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {t} ^ {s} = \\mathcal {N} _ {f} ^ {s} \\left(\\mathcal {F} _ {t} ^ {s}, h _ {t} ^ {s}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 437, + 280, + 785, + 297 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We simplify the sequential operations of Conv, ReLU, and ResBlock as $\\mathcal{N}_f$ . Additionally, to leverage temporal information, we align the features at times $t - 1$ and $t + 1$ to the current feature at time $t$ using event features that encapsulate motion between frames in the following manner:", + "bbox": [ + 214, + 306, + 785, + 368 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} _ {t - 1 \\rightarrow t} ^ {s} = \\mathcal {N} _ {g, f} ^ {s} (\\hat {\\mathcal {F}} _ {t - 1} ^ {s}, \\mathcal {F} _ {t} ^ {s}, \\varepsilon_ {[ t - 1, t ]} ^ {s}),\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 380, + 622, + 398 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} _ {t + 1 \\rightarrow t} ^ {s} = \\mathcal {N} _ {g, b} ^ {s} \\left(\\mathcal {F} _ {t} ^ {s}, \\hat {\\mathcal {F}} _ {t + 1} ^ {s}, \\varepsilon_ {[ t, t + 1 ]} ^ {s}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 402, + 784, + 422 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} _ {t} ^ {s} = \\mathcal {N} _ {h} ^ {s} (\\mathcal {T} _ {t - 1 \\rightarrow t}) \\| \\mathcal {T} _ {t + 1 \\rightarrow t} ^ {s}),\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 425, + 571, + 441 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\parallel$ denotes the channel-wise concatenation, and $\\mathcal{N}_{g,f}^{s},\\mathcal{N}_{g,b}^{s}$ , and $\\mathcal{N}_h^s$ represent convolution blocks, as illustrated on the right side of Fig. 4. While we obtain temporally aligned features $\\mathcal{T}_t^s$ using event features and leveraging the advantages of events, for additional feature refinement, we utilize the spatially variant pixel-wise dynamic filter [21,55] mechanism. Dynamic filter $\\mathcal{D}_t^s$ at scale $s$ can be calculated through filter generation blocks, $\\mathcal{N}_l^s$ , which consists of convolution and resblock, such as $\\mathcal{D}_t^s = \\mathcal{N}_l^s (\\mathcal{T}_t^s)$ where $\\mathcal{D}_t^s\\in \\mathbb{R}^{(s_k\\times s_k)\\times H^s\\times W^s}$ , $s_k$ is kernel size of dynamic convolution filter, $H^{s}$ , $W^{s}$ denote height and width of the feature at scale factor $s$ , respectively. Then, we apply the dynamic convolution operation as follows:", + "bbox": [ + 212, + 452, + 787, + 603 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathcal {T}} _ {t} ^ {s} (h, w) = \\mathcal {D} _ {t} ^ {s} (h, w) \\otimes \\mathcal {T} _ {t} ^ {s} (h, w) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 614, + 785, + 633 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $h \\in \\{1, \\dots, H^s\\}$ , $w \\in \\{1, \\dots, W^s\\}$ , and $\\otimes$ denotes the convolution operation. Then, we employ a cross-attention mechanism [38,44] to effectively combine the information of two features, $S_t^s$ and $\\hat{T}_t^s$ . The cross-attention generally examines the correlation between input features (query) and the key-value features. Therefore, we conduct correlation analysis by applying cross-attention between the current blurred frame features to project query $S_t^s$ and the aligned features $T_t^s$ to project key and values. That is, we generate query, key, and value features, $\\mathbf{Q} = W_Q(S_t^s)$ , $\\mathbf{K} = W_K(\\hat{T}_t^s)$ , $\\mathbf{V} = W_V(\\hat{T}_t^s)$ , where $W_{(\\cdot)}$ is $1 \\times 1$ convolution and $3 \\times 3$ depth-wise convolution. Utilizing these $\\mathbf{Q}$ , $\\mathbf{K}$ , and $\\mathbf{V}$ , we compute the attention matrix similarly to Eq. (4).", + "bbox": [ + 214, + 643, + 787, + 796 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {T}}{\\alpha}\\right) \\cdot \\mathbf {V} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 806, + 785, + 838 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Kim et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/f80c3286b616bd4cc9704d53a055b92234d2d9de9091ead59b6e000557b240ce.jpg", + "table_caption": [ + "Table 1: Quantitative results on the GoPro dataset. The asterisk(*) indicates that the results are not officially on the GoPro so we retrained the official model by us. CMTA-5 and CMTA-7 refer to the results obtained using 5 and 7 input frames, respectively." + ], + "table_footnote": [], + "table_body": "
MethodsMPRNet [45]Restormer [44]NAFNet [6]EDVR [39]ESTRNN [53]RNN-MBP [57]DSTNet [26]VRT [17]
PSNRs32.6632.9233.6926.8331.0233.3234.1634.81
SSIMs0.9590.9610.9670.8430.9110.9630.9680.972
Params(MB)20.126.167.823.62.416.47.518.3
MethodsRVRT [18]Shift-Net [16]UEVD* [15]EFNet [36]REFID [37]SpkNet [7]CMTA-5CMTA-7
PSNRs34.9235.4935.4835.4635.9136.1236.5536.78
SSIMs0.9740.9760.9710.9720.9730.9710.9770.977
Params(MB)10.810.527.98.515.913.59.79.7
", + "bbox": [ + 218, + 198, + 784, + 280 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Finally, temporally aligned feature, $\\hat{\\mathcal{F}}_t^s$ , can be obtained by:", + "bbox": [ + 214, + 306, + 650, + 324 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {F}} _ {t} ^ {s} = \\operatorname {M L P} (\\mathcal {A}) + \\mathcal {A}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 330, + 785, + 349 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where MLP denote multi-layer perceptrons, $\\mathcal{A}$ denote the result of attention operation. Finally, we obtained the aligned feature pyramid $\\{\\hat{\\mathcal{F}}_t^s\\}$ through cascaded temporal feature alignment. As illustrated in Fig. 2, these multi-scale aligned features are fed into the decoder.", + "bbox": [ + 212, + 357, + 787, + 417 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Decoder", + "text_level": 1, + "bbox": [ + 214, + 436, + 333, + 452 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The decoder is designed based on the standard U-Net [32]. It takes multi-scale temporally aligned features $\\{\\hat{\\mathcal{F}}_t^s\\}$ as inputs and produces output feature pyramid $\\{\\mathcal{F}(D)_k^s\\}$ . The final deblurred outputs $S_{t}$ using last scale of output feature $\\mathcal{F}(D)_t^{s = 0}$ is calculated as follows:", + "bbox": [ + 214, + 459, + 787, + 523 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nS _ {t} = B _ {t} + \\operatorname {C o n v} _ {5 \\times 5} (\\mathcal {F} (D) _ {t} ^ {s = 0}) \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 529, + 785, + 547 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $\\mathrm{Conv}_{5\\times 5}$ represents a conv. layer with a filter size of $5\\times 5$ , and $S_{t}$ denotes the final estimated sharp frame.", + "bbox": [ + 212, + 554, + 787, + 585 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 214, + 606, + 375, + 625 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.1 Datasets", + "text_level": 1, + "bbox": [ + 214, + 635, + 334, + 648 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "GoPro Dataset [23]. For a fair comparison with other previous event-guided deblurring methods, we utilize the same raw events provided by the authors of recent work [36]. These events were generated using ESIM [29] with a randomly generated contrast threshold set to a Gaussian normal distribution of parameters as $N(\\mu = 0.2, \\sigma = 0.03)$ . We used the official train and test splits.", + "bbox": [ + 212, + 657, + 787, + 734 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "HighREV Dataset [37]. HighREV consists of high-resolution events and RGB data at $1632 \\times 1224$ resolution, designed for both deblurring and interpolation tasks. To evaluate motion deblurring exclusively, we use the $11 + 1$ split, excluding the interpolation ground truth.", + "bbox": [ + 212, + 734, + 787, + 794 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Real-world Video Deblurring Dataset. EVRB dataset consists of 11 training sequences and 6 test sequences. With each sequence containing 149 frames, it is well-suited for video deblurring tasks.", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Cross-Modal Temporal Alignment for Event-guided Video Deblurring", + "bbox": [ + 267, + 114, + 732, + 130 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d297df95791baffe6f5ee30a8cad71b94ba84face5532a358f688724c884a8c7.jpg", + "image_caption": [ + "Fig. 5: Visual comparison of deblurring results on the GoPro dataset. The qualitative results of other methods were taken from the results provided by the authors." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 268 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/9565fd5b07d4e3590cd4b78be09e690a39017c98c1c40a8b4df3cb8201d075f8.jpg", + "table_caption": [ + "Table 2: Quantitative results on the HighREV dataset." + ], + "table_footnote": [], + "table_body": "
MethodsESTRNN [53]DSTNet [26]UEVD [15]EFNet [36]REFID [37]Ours
PSNRs30.3831.7737.4037.9938.3739.12
SSIMs0.9400.9480.9740.9760.9770.980
", + "bbox": [ + 235, + 349, + 767, + 386 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Comparison on Synthetic Blur Datasets", + "text_level": 1, + "bbox": [ + 215, + 414, + 589, + 429 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We present the quantitative results of our frameworks with other frame-based image and video deblurring methods and event-guided motion deblurring methods using the GoPro dataset as depicted in Table 1. When compared to the existing best performance of video deblurring methods, Shift-Net [16], our approach (CMTA-5 model) shows a significant improvement of $1.06\\mathrm{dB}$ in terms of PSNR, demonstrating that our method effectively leverages the temporal dense characteristics of event modality for accurate temporal feature alignment and cross-modality feature enhancement. Moreover, compared to the best-performing SpkNet [7] among existing event-guided motion deblurring methods, CMTA-5 model exhibits a performance improvement of $0.43\\mathrm{dB}$ with a lower model params of $3.8\\mathrm{MB}$ . Furthermore, by using 7 input video frames (CMTA-7 model) instead of 5 (CMTA-5 model), we achieved an impressive state-of-the-art performance with a PSNR of $36.78\\mathrm{dB}$ in the GoPro, an increase of 0.23. We further demonstrated the superiority of our approach through the qualitative results in Fig. 5. Also, we conduct experiments on HighREV [37], which consists of real events, and report the results in Table 2. Our method still achieves the best performance.", + "bbox": [ + 212, + 441, + 787, + 684 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Comparison on Real-world Blur Datasets", + "text_level": 1, + "bbox": [ + 215, + 707, + 602, + 722 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For comparisons in the EVRB dataset, we trained representative video deblurring methods [12, 16, 18, 26, 39, 53] and event-guided motion deblurring methods [15, 36, 37] on the same training set. Tab 3 presents the quantitative results on the EVRB dataset. The EVRB dataset includes extremely blurred videos captured during exposure times. As a result, frame-based video deblurring methods exhibit subpar performance. For instance, the best-performing network, ShiftNet [16], achieves only $30.56\\mathrm{dB}$ , which is $0.42\\mathrm{dB}$ lower than the best-performing", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Kim et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/b14802cf7bb69e54e4849c004e74abbfcaeb919c3095e18121d2b23762f64499.jpg", + "table_caption": [ + "Table 3: Quantitative results on the EVRB dataset." + ], + "table_footnote": [], + "table_body": "
MethodsEDVR [39]ESTRNN [53]ERDN [12]DSTNet [26]RVRT [18]
PSNRs29.0229.7928.3229.1530.24
SSIMs0.8860.9110.8930.8980.906
MethodsShift-Net [16]UEVD [15]EFNet [36]REFID [37]Ours
PSNRs30.5630.5530.9830.3331.38
SSIMs0.9220.9150.9270.9180.927
", + "bbox": [ + 233, + 170, + 769, + 239 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/34782f7695ccac1073018b3771aeca0fc4244793ee4b28ad2b3fc821b7801a3f.jpg", + "image_caption": [ + "Fig. 6: Visual comparison of deblurring results on the EVRB dataset." + ], + "image_footnote": [], + "bbox": [ + 218, + 252, + 785, + 377 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "event-guided deblurring method, EFNet [36]. In contrast, our approach leverages video and event characteristics, effectively restoring even severe motion blur, which may be difficult to recover. Our method outperforms all approaches, achieving the best performance. We show our qualitative comparison with other methods in Fig. 6.", + "bbox": [ + 212, + 431, + 787, + 506 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 527, + 388, + 544 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We analyzed the performance contribution of the various modules in our frameworks. For a fair ablation study, we trained all the models for 600 epochs with 5 video frame inputs, conducting all experiments on the GoPro dataset.", + "bbox": [ + 212, + 553, + 785, + 598 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "CRIFE module. To demonstrate the effectiveness of the CRIFE module, we replaced it with concatenation and convolution for comparison. When comparing the first column(Ver.1) and the second column(Ver.2) of Tab 4, we observed a performance gain of $+0.35$ dB with a small additional parameter $(+0.08\\mathrm{MB})$ . Similarly, when comparing the performance of the third and fourth rows, we observed performance improvement.", + "bbox": [ + 212, + 598, + 785, + 688 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "ECITFA module is the most crucial component of our model, performing feature alignment by leveraging information of the events. As in the Tab. 4, when comparing the first (Ver.1) with the third column (Ver.3) of the table, we observed a significant performance gain $(+1.69\\mathrm{dB})$ upon the insertion of the ECITFA module. This trend is similarly maintained when the CRIFE module is incorporated. When comparing Ver.2 with Ver.4 of the Tab. 4, we observed a notable performance improvement $(+1.55\\mathrm{dB})$ .", + "bbox": [ + 212, + 688, + 785, + 794 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of components in ECITFA module. In the Tab. 5, we demonstrated an effectiveness analysis for each component of ECITFA. Comparing the 2nd column with the last column labeled 'Ours', we observed a performance", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Cross-Modal Temporal Alignment for Event-guided Video Deblurring", + "bbox": [ + 269, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/98d79ed63db2a5806a5ac275995c92470219ff2fd60b906b7e5fd2307c5e288b.jpg", + "table_caption": [ + "Table 4: Ablation study on the GoPro dataset." + ], + "table_footnote": [], + "table_body": "
MethodsVer.1Ver.2Ver.3Ver.4
CRIFE✓✓
ECITFA✓✓
PSNRs / Params34.65 / 4.57M35.00 / 4.65M36.34 / 9.59M36.55 / 9.68M
", + "bbox": [ + 263, + 170, + 738, + 219 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/7b060511466c96e365dd93c61b47972fabd5fe6a8f7748098ec898868cebfbaa.jpg", + "table_caption": [ + "Table 5: Effect of the component of ECITFA module on the GoPro dataset. DF denotes dynamic filter operation in the Eq.(12)." + ], + "table_footnote": [], + "table_body": "
Methodsw/o Cascadedw/o DFw/o AttentionOurs
PSNRs36.0836.1836.2936.55
", + "bbox": [ + 263, + 272, + 738, + 303 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/e75d7befeecd5a65f7a13d1b640625fafe3e1fd467a74351d0a918dab95a961f.jpg", + "table_caption": [ + "Table 6: Comparison of CRIFE with various module variants. \"Baseline\" refers to Ver.3 of Tab.4. E and F represent event and frame features, respectively." + ], + "table_footnote": [], + "table_body": "
MethodsBaselinew/o concat (E+F)w/concat (E+F)
w/o recurrentw/ recurrent(Ours)
PSNRs36.3436.0435.9936.55
", + "bbox": [ + 259, + 356, + 736, + 393 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "gain of our method $(+0.37\\mathrm{dB})$ when employing spatial pixel-wise dynamic filters, in contrast to not using DF. When utilizing cross-attention (Eq.13) to better leverage long-range pixel-dependencies in the alignment blocks, we observed a performance improvement $(+0.29\\mathrm{dB})$ compared to aggregate feature using ResBlocks. Finally, we confirmed the effectiveness of leveraging non-local video frame information through a cascaded-based temporal feature alignment method for video deblurring. After using the proposed cascaded structure for temporal feature alignment, we can observe a performance gain $(+0.47\\mathrm{dB})$ .", + "bbox": [ + 212, + 422, + 784, + 542 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effectiveness of components in CRIFE module. To evaluate each component's effectiveness of the CRIFE module, we removed the concatenation between RGB and event features, directly matching their features. However, as shown in the 3rd column of Tab. 6, this approach yielded sub-optimal performance. Additionally, combining event features without a recurrent structure degraded performance (see 4th column). These ablations confirm that the recurrent structure effectively utilizes temporal information from events within exposure time.", + "bbox": [ + 212, + 542, + 784, + 648 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 672, + 366, + 688 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper proposes a video deblurring framework, CMTA, that elaborately considers the characteristics of an event and video. Specifically, we achieve significant performance improvement through intra-frame feature enhancement and inter-frame temporal feature alignment. Furthermore, we construct a real-world deblurring dataset, the EVRB dataset, which will be valuable for evaluating event-guided deblurring methods. Finally, CMTA demonstrates state-of-the-art performance across various deblurring datasets.", + "bbox": [ + 212, + 703, + 784, + 809 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by the Technology Innovation Program (1415187329,20024355, Development of autonomous driving connec", + "bbox": [ + 215, + 809, + 782, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Kim et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "tivity technology based on sensor-infrastructure cooperation) funded By the Ministry of Trade, Industry & Energy(MOTIE, Korea) and the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (NRF2022R1A2B5B03002636).", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 232, + 323, + 248 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aittala, M., Durand, F.: Burst image deblurring using permutation invariant convolutional neural networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 731-747 (2018)", + "2. Brandli, C., Berner, R., Yang, M., Liu, S.C., Delbruck, T.: A $240 \\times 180$ 130 db $3\\mu s$ latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits 49(10), 2333-2341 (2014)", + "3. Cao, M., Zhong, Z., Wang, J., Zheng, Y., Yang, Y.: Learning adaptive warping for real-world rolling shutter correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17785-17793 (2022)", + "4. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basic vsr: The search for essential components in video super-resolution and beyond. In: Proceedings of the IEEE conference on computer vision and pattern recognition (2021)", + "5. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment (2021)", + "6. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: European Conference on Computer Vision. pp. 17-33. Springer (2022)", + "7. Chen, S., Zhang, J., Zheng, Y., Huang, T., Yu, Z.: Enhancing motion deblurring in high-speed scenes with spike streams. Advances in Neural Information Processing Systems 36 (2024)", + "8. Cho, H., Jeong, Y., Kim, T., Yoon, K.J.: Non-coaxial event-guided motion deblurring with spatial alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12492-12503 (2023)", + "9. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence 44(1), 154-180 (2020)", + "0. Gehrig, D., Loquercio, A., Derpanis, K.G., Scaramuzza, D.: End-to-end learning of representations for asynchronous event-based data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5633-5643 (2019)", + "1. Hyun Kim, T., Mu Lee, K., Scholkopf, B., Hirsch, M.: Online video deblurring via dynamic temporal blending network. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 4038-4047 (2017)", + "2. Jiang, B., Xie, Z., Xia, Z., Li, S., Liu, S.: Erdn: Equivalent receptive field deformable network for video deblurring. In: European Conference on Computer Vision. pp. 663-678. Springer (2022)", + "3. Jiang, Z., Zhang, Y., Zou, D., Ren, J., Lv, J., Liu, Y.: Learning event-based motion deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3320-3329 (2020)", + "4. Kim, T.H., Sajjadi, M.S., Hirsch, M., Scholkopf, B.: Spatio-temporal transformer network for video restoration. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 106-122 (2018)" + ], + "bbox": [ + 225, + 266, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Cross-Modal Temporal Alignment for Event-guided Video Deblurring", + "bbox": [ + 267, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Kim, T., Lee, J., Wang, L., Yoon, K.J.: Event-guided deblurring of unknown exposure time videos. In: European Conference on Computer Vision. pp. 519-538. Springer (2022)", + "16. Li, D., Shi, X., Zhang, Y., Cheung, K.C., See, S., Wang, X., Qin, H., Li, H.: A simple baseline for video restoration with grouped spatial-temporal shift. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9822-9832 (June 2023)", + "17. Liang, J., Cao, J., Fan, Y., Zhang, K., Ranjan, R., Li, Y., Timofte, R., Van Gool, L.: Vrt: A video restoration transformer. arXiv preprint arXiv:2201.12288 (2022)", + "18. Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems 35, 378-393 (2022)", + "19. Lin, J., Cai, Y., Hu, X., Wang, H., Yan, Y., Zou, X., Ding, H., Zhang, Y., Timofte, R., Van Gool, L.: Flow-guided sparse transformer for video deblurring. arXiv preprint arXiv:2201.01893 (2022)", + "20. Lin, S., Zhang, J., Pan, J., Jiang, Z., Zou, D., Wang, Y., Chen, J., Ren, J.S.: Learning event-driven video deblurring and interpolation. In: ECCV (8). pp. 695-710 (2020)", + "21. Mildenhall, B., Barron, J.T., Chen, J., Sharlet, D., Ng, R., Carroll, R.: Burst denoising with kernel prediction networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2502-2510 (2018)", + "22. Nah, S., Baik, S., Hong, S., Moon, G., Son, S., Timofte, R., Mu Lee, K.: Ntire 2019 challenge on video deblurring and super-resolution: Dataset and study. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)", + "23. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3883-3891 (2017)", + "24. Oh, J., Kim, M.: Demfi: deep joint deblurring and multi-frame interpolation with flow-guided attentive correlation and recursive boosting. In: European Conference on Computer Vision. pp. 198-215. Springer (2022)", + "25. Pan, J., Bai, H., Tang, J.: Cascaded deep video deblurring using temporal sharpness prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3043-3051 (2020)", + "26. Pan, J., Xu, B., Dong, J., Ge, J., Tang, J.: Deep discriminative spatial and temporal network for efficient video deblurring. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (Feb 2023)", + "27. Pan, L., Scheerlinck, C., Yu, X., Hartley, R., Liu, M., Dai, Y.: Bringing a blurry frame alive at high frame-rate with an event camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6820-6829 (2019)", + "28. Ranjan, A., Black, M.J.: Optical flow estimation using a spatial pyramid network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4161-4170 (2017)", + "29. Rebecq, H., Gehrig, D., Scaramuzza, D.: Esim: an open event camera simulator. In: Conference on Robot Learning. pp. 969-982. PMLR (2018)", + "30. Rim, J., Kim, G., Kim, J., Lee, J., Lee, S., Cho, S.: Realistic blur synthesis for learning image deblurring. In: European conference on computer vision. pp. 487-503. Springer (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Kim et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Rim, J., Lee, H., Won, J., Cho, S.: Real-world blur dataset for learning and benchmarking deblurring algorithms. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16. pp. 184-201. Springer (2020)", + "32. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)", + "33. Shang, W., Ren, D., Zou, D., Ren, J.S., Luo, P., Zuo, W.: Bringing events into video deblurring with non-consecutively blurry frames. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4531-4540 (October 2021)", + "34. Shen, W., Bao, W., Zhai, G., Chen, L., Min, X., Gao, Z.: Video frame interpolation and enhancement via pyramid recurrent framework. IEEE Transactions on Image Processing 30, 277-292 (2020)", + "35. Su, S., Delbracio, M., Wang, J., Sapiro, G., Heidrich, W., Wang, O.: Deep video deblurring for hand-held cameras. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1279-1288 (2017)", + "36. Sun, L., Sakaridis, C., Liang, J., Jiang, Q., Yang, K., Sun, P., Ye, Y., Wang, K., Van Gool, L.: Event-based fusion for motion deblurring with cross-modal attention. In: European Conference on Computer Vision (ECCV) (2022)", + "37. Sun, L., Sakaridis, C., Liang, J., Sun, P., Cao, J., Zhang, K., Jiang, Q., Wang, K., Van Gool, L.: Event-based frame interpolation with ad-hoc deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18043-18052 (2023)", + "38. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NIPS. pp. 5998-6008 (2017)", + "39. Wang, X., Chan, K.C., Yu, K., Dong, C., Change Loy, C.: Edvr: Video restoration with enhanced deformable convolutional networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (June 2019)", + "40. Wang, Y., Lu, Y., Gao, Y., Wang, L., Zhong, Z., Zheng, Y., Yamashita, A.: Efficient video deblurring guided by motion magnitude. In: Proceedings of the European Conference on Computer Vision (ECCV) (2022)", + "41. Wei, K., Fu, Y., Yang, J., Huang, H.: A physics-based noise formation model for extreme low-light raw denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2758-2767 (2020)", + "42. Wieschollek, P., Hirsch, M., Scholkopf, B., Lensch, H.: Learning blind motion deblurring. In: ICCV. pp. 231-240 (2017)", + "43. Xu, F., Yu, L., Wang, B., Yang, W., Xia, G.S., Jia, X., Qiao, Z., Liu, J.: Motion deblurring with real events. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2583-2592 (October 2021)", + "44. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5728-5739 (2022)", + "45. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. arXiv preprint arXiv:2102.02808 (2021)", + "46. Zhang, H., Xie, H., Yao, H.: Spatio-temporal deformable attention network for video deblurring. In: ECCV (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Cross-Modal Temporal Alignment for Event-guided Video Deblurring", + "bbox": [ + 269, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "47. Zhang, K., Luo, W., Zhong, Y., Ma, L., Liu, W., Li, H.: Adversarial spatio-temporal learning for video deblurring. IEEE Transactions on Image Processing (2018)", + "48. Zhang, L., Zhang, H., Zhu, C., Guo, S., Chen, J., Wang, L.: Fine-grained video deblurring with event camera. In: ICMM. pp. 352-364. Springer (2021)", + "49. Zhang, X., Yu, L.: Unifying motion deblurring and frame interpolation with events. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17765-17774 (2022)", + "50. Zhang, X., Yu, L., Yang, W., Liu, J., Xia, G.S.: Generalizing event-based motion deblurring in real-world scenarios. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10734-10744 (2023)", + "51. Zhang, Y., Qin, H., Wang, X., Li, H.: Rethinking noise synthesis and modeling in raw denoising. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4593-4601 (2021)", + "52. Zhong, Z., Cao, M., Ji, X., Zheng, Y., Sato, I.: Blur interpolation transformer for real-world motion from blur. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5713-5723 (2023)", + "53. Zhong, Z., Gao, Y., Zheng, Y., Zheng, B.: Efficient spatio-temporal recurrent neural network for video deblurring. In: European Conference on Computer Vision. pp. 191-207. Springer (2020)", + "54. Zhong, Z., Zheng, Y., Sato, I.: Towards rolling shutter correction and deblurring in dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9219-9228 (2021)", + "55. Zhou, J., Jampani, V., Pi, Z., Liu, Q., Yang, M.H.: Decoupled dynamic filter networks. In: IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR) (Jun 2021)", + "56. Zhu, A.Z., Yuan, L., Chaney, K., Daniilidis, K.: Unsupervised event-based learning of optical flow, depth, and egomotion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 989-997 (2019)", + "57. Zhu, C., Dong, H., Pan, J., Liang, B., Huang, Y., Fu, L., Wang, F.: Deep recurrent neural network with multi-scale bi-directional propagation for video deblurring. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 3598-3607 (2022)", + "58. Zhu, X., Hu, H., Lin, S., Dai, J.: Deformable convnets v2: More deformable, better results. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9308-9316 (2019)" + ], + "bbox": [ + 215, + 147, + 784, + 631 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Kim et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/CMTA_ Cross-Modal Temporal Alignment for Event-guided Video Deblurring/7cf7582c-c0da-4c34-8883-80e1b8a2e3ef_model.json b/2024/CMTA_ Cross-Modal Temporal Alignment for Event-guided Video Deblurring/7cf7582c-c0da-4c34-8883-80e1b8a2e3ef_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8b4717472441a2d3a87534a5a2e4e8769a0b4867 --- /dev/null +++ b/2024/CMTA_ Cross-Modal Temporal Alignment for Event-guided Video Deblurring/7cf7582c-c0da-4c34-8883-80e1b8a2e3ef_model.json @@ -0,0 +1,2480 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.242, + 0.141, + 0.761, + 0.186 + ], + "angle": 0, + "content": "CMTA: Cross-Modal Temporal Alignment for Event-guided Video Deblurring" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.213, + 0.697, + 0.228 + ], + "angle": 0, + "content": "Taewoo Kim\\(^{1}\\), Hoonhee Cho\\(^{1}\\), and Kuk-Jin Yoon\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.24, + 0.677, + 0.269 + ], + "angle": 0, + "content": "Korea Advanced Institute of Science and Technology {intelpro, gngsgsm1, kjyoon}@kaist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.306, + 0.74, + 0.625 + ], + "angle": 0, + "content": "Abstract. Video deblurring aims to enhance the quality of restored results in motion-blurred videos by effectively gathering information from adjacent video frames to compensate for the insufficient data in a single blurred frame. However, when faced with consecutively severe motion blur situations, frame-based video deblurring methods often fail to find accurate temporal correspondence among neighboring video frames, leading to diminished performance. To address this limitation, we aim to solve the video deblurring task by leveraging an event camera with microsecond temporal resolution. To fully exploit the dense temporal resolution of the event camera, we propose two modules: 1) Intra-frame feature enhancement operates within the exposure time of a single blurred frame, iteratively enhancing cross-modality features in a recurrent manner to better utilize the rich temporal information of events, 2) Interframe temporal feature alignment gathers valuable long-range temporal information to target frames, aggregating sharp features leveraging the advantages of the events. In addition, we present a novel dataset composed of real-world blurred RGB videos, corresponding sharp videos, and event data. This dataset serves as a valuable resource for evaluating event-guided deblurring methods. We demonstrate that our proposed methods outperform state-of-the-art frame-based and event-based motion deblurring methods through extensive experiments conducted on both synthetic and real-world deblurring datasets. The code and dataset are available at https://github.com/intelpro/CMTA." + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.638, + 0.725, + 0.653 + ], + "angle": 0, + "content": "Keywords: Temporal alignment \\(\\cdot\\) Video Deblurring \\(\\cdot\\) Event cameras" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.679, + 0.375, + 0.695 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.711, + 0.787, + 0.816 + ], + "angle": 0, + "content": "Motion blur is a common artifact caused by dynamic movements within a scene or camera motion during exposure. Motion deblurring, which aims to reverse the blurring process, presents significant challenges due to variations in blur intensity influenced by scene structure and depth. To achieve high-quality deblurring, video deblurring has emerged, leveraging information from neighboring frames instead of relying solely on a single blurred image. However, identifying temporal correspondence between blurred video frames becomes challenging with" + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.826, + 0.365, + 0.84 + ], + "angle": 0, + "content": "* Equal contribution." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.128 + ], + "angle": 0, + "content": "Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "extreme motion blur, hindering the extraction of valuable information from adjacent frames and impeding performance improvement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.18, + 0.787, + 0.286 + ], + "angle": 0, + "content": "Event cameras [9], with their extremely low latency (on the order of microseconds), can offer high-quality guidance for motion deblurring due to their ability to capture high-temporal resolution of brightness change. To effectively utilize the advantages of the events, several event-guided motion deblurring works [15, 36, 50] have been introduced. While these works have typically explored cross-modal feature fusion methods across different modalities, there has been limited work on leveraging the abundant temporal information in videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.289, + 0.788, + 0.424 + ], + "angle": 0, + "content": "To obtain high-quality results, we emphasize the event camera's temporal continuity, focusing on its interaction with video frames that exhibit long-term temporal dependencies. Unlike previous event-guided motion deblurring works [36,43,50], which relied on a single blurry image and corresponding events on its exposure time, we further design precise temporal feature alignment methods between neighboring video frames by leveraging the advantages of event data. Specifically, we propose novel modules from two perspectives: intra-frame (interaction between events and frames within the exposure time) and inter-frame (interaction between different frames) perspectives." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.788, + 0.518 + ], + "angle": 0, + "content": "From an intra-frame perspective within exposure time, we propose a Cross-modal Recurrent Intra-frame Feature Enhancement (CRIFE) module to better leverage the rich temporal information of the events by mutually interacting the blur frame and event features within the duration of exposure time. In this module, we perform recurrent attention-based feature enhancement using a transformer [38] that better captures long-range pixel dependencies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.788, + 0.808 + ], + "angle": 0, + "content": "In temporal feature alignment with the second perspective, we propose a novel event-guided temporal feature alignment module, effectively leveraging rich temporal characteristics of the events. Conventional frame-based temporal feature alignment methods rely on optical flow [28] or deformable convolution [5,58]. While optical flow and deformable convolutions aid in achieving temporal feature alignment, the high computational complexity of these operations makes it challenging to execute them at a higher spatial scale. Therefore, video frame alignment, generally conducted at lower spatial scales, leads to sub-optimal deblurring results due to the lack of spatial contexts of features. To overcome these limitations, our temporal feature alignment module avoids relying on optical flow or deformable convolution by effectively leveraging the temporally dense advantages of the events. Therefore, we can effectively perform temporal feature alignment across multiple visual scales as we do not rely on these complex operations but rather efficiently leverage the temporal information from the events. As a result, our temporal feature alignment module demonstrates a significant performance improvement in event-guided video deblurring tasks. Since we have introduced a pioneering method for aligning temporal features using the advantages of the events, it is expected to be effectively applicable to various event-guided video restoration tasks (e.g., event-guided video super-resolution)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Finally, we propose a novel video deblurring dataset, the EVRB dataset, composed of high-quality RGB and event data. It consists of real-world blurry" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Cross-Modal Temporal Alignment for Event-guided Video Deblurring" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.147, + 0.785, + 0.207 + ], + "angle": 0, + "content": "videos generated by extreme motion and corresponding sharp videos for generalized applications in real-world scenarios. The network trained on the EVRB dataset can be directly applicable to real-world scenarios, making it a valuable resource for event-guided deblurring research." + }, + { + "type": "title", + "bbox": [ + 0.219, + 0.231, + 0.395, + 0.246 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.219, + 0.262, + 0.405, + 0.277 + ], + "angle": 0, + "content": "2.1 Video Deblurring" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.287, + 0.785, + 0.423 + ], + "angle": 0, + "content": "Early works [1, 35] employed CNNs that take the concatenation of adjacent frames as input to address video deblurring. Subsequently, to better leverage temporal information, approaches have emerged that employ 3D convolutions [26, 47], temporal alignment modules with deformable convolutions [39] and optical flow [14, 25, 40, 46]. As an alternative video alignment method, some works utilize RNN [11, 42, 53] and transformer [18, 19] structures to propagate information from long-range video frames. However, as the intensity of motion blur in the video increases, frame-based video alignment methods struggle to achieve accurate video alignment, leading to sub-optimal deblurring results." + }, + { + "type": "title", + "bbox": [ + 0.219, + 0.446, + 0.532, + 0.461 + ], + "angle": 0, + "content": "2.2 Event-guided Motion Deblurring" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.471, + 0.785, + 0.682 + ], + "angle": 0, + "content": "An event camera can effectively be used for motion deblurring as it records motion information corresponding to the brightness differences with high temporal resolution. Efforts to utilize event cameras for motion deblurring [8, 13, 15, 20, 27, 36, 37, 43, 48-50] have been actively ongoing. Recent studies have focused on effectively fusing event and image features of different modalities. To this end, Sun et al. [36] employed a transformer architecture for feature fusion. Zhang et al. [50] effectively combined modalities using a multi-scale architecture. In addition, there are also attempts to address motion deblurring by assuming challenging and general scenarios [8, 15, 33]. However, these studies have primarily focused on single image deblurring and do not exploit the long-range temporal information demonstrated in previous video deblurring tasks. To effectively acquire information that may be missing from sparse events, we introduce a novel method for accurate temporal alignment with events, utilizing information from surrounding adjacent frames in video deblurring." + }, + { + "type": "title", + "bbox": [ + 0.219, + 0.705, + 0.772, + 0.739 + ], + "angle": 0, + "content": "3 Event-based Video Deblurring Dataset for Real-world Blur" + }, + { + "type": "title", + "bbox": [ + 0.219, + 0.755, + 0.565, + 0.77 + ], + "angle": 0, + "content": "3.1 Limitation of Synthetic Blur Dataset" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Typically, the blurred images synthesis procedure adheres to the methodology outlined in [22-24, 34, 35], involving the averaging of consecutive video frames within a fixed-size window. However, as recent studies [31, 52, 54] have discussed, blur synthesis based on discrete signals may result in shutter artifacts even when" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.127 + ], + "angle": 0, + "content": "Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.145, + 0.382, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.144, + 0.568, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.145, + 0.772, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.305, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Fig. 1: Illustration of a hybrid camera system for real-world event-based video deblurring dataset. \\( S \\) and \\( B \\) denote the cameras for acquiring sharp and blur videos, respectively. (a): The triple-axis camera system to capture real-world blur. (b): A diagram of our hybrid camera system. (c): Samples from our EVRB dataset with natural blur." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.789, + 0.691 + ], + "angle": 0, + "content": "averaging high-frame-rate videos. Furthermore, simply averaging sharp images disregards essential elements in the blurred images such as pixel saturation [30], limitations due to dynamic range, and physical noise [41, 51] in data acquired during exposure time. Furthermore, in the case of the event-guided motion deblurring research, existing works have enhanced the synthetic aspects by using event simulator [29]. To address these limitations, recent works have attempted to acquire real-world blur datasets using hybrid camera systems [3, 31, 52-54]. Meanwhile, to generalize event-based motion deblurring, Sun et al. [36] introduced the ReBlur dataset acquired in a high-precision optical laboratory using an electronic-controlled slide-rail system. However, despite these successes, the ReBlur dataset has limitations due to its indoor setting, a lack of dynamic objects, and the absence of high-quality RGB data as it is acquired using the DAVIS sensor [2]. Additionally, the sequences are relatively short(minimum six blurred frames in the sequence), making it challenging for use in event-guided video deblurring. We have acquired a new EVRB dataset for evaluating event-guided image and video deblurring methods in real-world blurry videos. The EVRB dataset was captured using a customized hybrid system consisting of two RGB cameras and one event camera. This system encompasses a range of blur magnitude from slight to extreme in various urban environments." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.711, + 0.55, + 0.727 + ], + "angle": 0, + "content": "3.2 Triple-axis Hybrid Camera System" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "As shown in Fig. 1, we design a hybrid camera system to enable the acquisition of different data sources to be geometrically aligned. Two RGB cameras and one event camera are geometrically aligned using two \\(50/50\\) cube beam splitters, resulting in a minimal baseline. We perform pixel-wise alignment between multiple cameras based on the homography calculated by extrinsic calibration for precise alignment. For the synchronization of the multiple cameras, we use an external trigger system. Each camera receives the falling and rising edges of the trigger" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Cross-Modal Temporal Alignment for Event-guided Video Deblurring" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.35, + 0.788, + 0.421 + ], + "angle": 0, + "content": "Fig. 2: Overall framework of CMTA is divided into two main components: Cross-modal Recurrent Intra-frame Feature Enhancement (CRIFE) and Event-guided Cascaded Inter-frame Temporal Feature Alignment (ECITFA). \\(s\\) is the scale factor for multi-scale features. In the figure of the ECITFA module, the description was performed for the case of \\(P = 2\\) for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.553 + ], + "angle": 0, + "content": "signal and acquires data synchronized with the period of the signal. This external trigger can precisely control the exposure times of the two RGB cameras, enabling us to capture paired sharp and blurred video frames. Furthermore, for photometric alignment, we physically adjust the amount of incoming light for both cameras to equalize the total irradiance of the multiple cameras using a neutral density filter." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.574, + 0.331, + 0.59 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.605, + 0.342, + 0.619 + ], + "angle": 0, + "content": "4.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.844 + ], + "angle": 0, + "content": "The overview of the proposed framework is illustrated in Fig. 2. Given consecutive blurred video frames \\(\\{B_k\\}\\) and sets of event streams corresponding to the exposure time of each video frame \\(\\{\\mathbb{E}_k\\}\\), where \\(k \\in \\{t - P, \\dots, t, \\dots, t + P\\}\\), our goal is to estimate the latent sharp video frame \\(S_t\\). To utilize the event stream \\(\\{\\mathbb{E}_k\\}\\) corresponding to the exposure time of \\(\\{B_k\\}\\) as the input for the networks, we first perform embedding using the event voxel grid representation [56] for the event stream \\(\\{\\mathbb{E}_k\\}\\), resulting in \\(\\{E_k\\}\\). Our framework consists of two main submodules: (1) Cross-modal Recurrent Intra-frame Feature Enhancement (CRIFE) module and (2) Event-guided Cascaded Inter-frame Temporal Feature Alignment (ECITFA) module. In the first module, we perform cross-modal feature enhancement through recurrent interactions between blurred frame features and event features to leverage the continuous temporal information of events within the exposure time. After obtaining the fused feature \\(\\{\\mathcal{G}_k\\}\\) from the first module, we generate multi-scale features, \\(\\{\\mathcal{F}_k^s\\}_{s=0}^2\\), through a weight-shared pyra" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.127 + ], + "angle": 0, + "content": "Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.146, + 0.734, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.345, + 0.788, + 0.374 + ], + "angle": 0, + "content": "Fig. 3: Illustration of Cross-modal Recurrent Intra-frame Feature Enhancement (CRIFE)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.415, + 0.789, + 0.552 + ], + "angle": 0, + "content": "mid encoder. Subsequently, for the temporal feature alignment stage, we encode multi-scale event features, \\(\\{\\varepsilon_{[k,k + 1]}^s\\}_{s = 0}^2\\), by grouping two consecutive events corresponding to the exposure times of each frame to connect adjacent frames. By utilizing the multi-scale pyramid features \\(\\{\\mathcal{F}_k^s\\}_{s = 0}^2\\) and the encoded event feature \\(\\{\\varepsilon_{[k,k + 1]}^s\\}_{s = 0}^2\\), we perform temporal feature alignment. Afterward, temporally aligned feature pyramids are fed into the U-Net [32]-based decoder, generating the final sharp video frame \\(S_{t}\\). Our framework can be extended to cases where \\(P\\) is an arbitrary positive number. However, for the sake of brevity, we will explain it in the main text with \\(P = 2\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.575, + 0.741, + 0.589 + ], + "angle": 0, + "content": "4.2 Cross-modal Recurrent Intra-frame Feature Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.748 + ], + "angle": 0, + "content": "Since event cameras provide rich temporal information on brightness changes, it is crucial to effectively utilize this dense temporal information of the events within the duration of the exposure time. Typically, event data captured during the exposure time is transformed into event embeddings [10,56] as input to the network, followed by feature extraction using 2D CNNs. However, these approaches cannot effectively utilize the continuous temporal information of events, and these limitations can impact the performance of event-guided video deblurring. To address these limitations, we propose a method that leverages the rich temporal nature of events and fuses event and blurred frame features using recurrent-based attention methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Recently, researchers have demonstrated the effectiveness of transformer-based architectures [38] by capturing long-range pixel dependencies, which have been proven highly effective in various vision tasks. We leverage transformers' advantages to more effectively utilize the temporal benefits of the events through a recurrent-based approach. For each blur video frame index \\( k \\in \\{t - 2, \\dots, t + 2\\} \\), we initially partition the exposure time \\( T_{exp,k} \\) into \\( N \\) unit temporal intervals," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.269, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Cross-Modal Temporal Alignment for Event-guided Video Deblurring" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "denoted as \\(\\Delta t\\), with \\(\\Delta t = T_{exp,k} / N\\). Based on the time interval \\(\\Delta t\\), we divide the event voxel grid within the exposure time, denoted as \\(E_k \\in \\mathbb{R}^{C \\times H \\times W}\\) into \\(N\\) temporally divided event voxel grids \\(\\{E_k^n\\}\\), where \\(E_k^n \\in \\mathbb{R}^{C / N \\times H \\times W}\\) and \\(n \\in \\{0, \\dots, N - 1\\}\\). After, we extract event features \\(\\{\\mathcal{F}(E)_k^n\\}\\) of the temporally divided event voxel grids \\(E_k^n\\) using weight-sharing event feature extractor. To apply cross-attention, we construct a query encompassing global information of blur and temporally divided event features. That is, we concatenate the blur frame feature \\(\\mathcal{F}(B)_k\\) with the temporally segmented event sets \\(\\{\\mathcal{F}(E)_k^n\\}\\) and then extract the feature for the encoding query information." + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.31, + 0.786, + 0.326 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Q} _ {k} = \\operatorname {C o n v} _ {\\mathrm {p}} (\\mathcal {F} (B) _ {k} \\| \\left\\{\\mathcal {F} (E) _ {k} ^ {n} \\right\\}) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.445, + 0.329, + 0.785, + 0.346 + ], + "angle": 0, + "content": "\\[\nQ _ {k} ^ {n = 0} = F _ {R} \\left(\\mathcal {Q} _ {k}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.789, + 0.414 + ], + "angle": 0, + "content": "where \\(\\mathrm{Conv}_{\\mathrm{p}}\\) represents point-wise convolution, \\(F_{R}\\) consists of a sequence of blocks with \\(3\\times 3\\) convolution and a stride of 2, along with ResBlocks. Additionally, \\(Q_{k}^{n = 0}\\) denotes the initial query feature for the recurrent-based attention method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.788, + 0.508 + ], + "angle": 0, + "content": "As illustrated in Fig. 3, we perform cross-attention to update query information iteratively. To find the key and value of cross-attention for query updating, we recursively input temporally separated event features, helping to better utilize the events' rich temporal information. That is, for iteration \\( n \\), we extract features to be used as keys and values by concatenating the previously updated \\( Q_{k}^{n} \\) with temporally divided \\( n \\)-th event features \\( \\mathcal{F}(E)_k^n \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.518, + 0.785, + 0.534 + ], + "angle": 0, + "content": "\\[\n\\mathrm {K V} _ {k} ^ {n} = F _ {K V} \\left(Q _ {k} ^ {n} \\| \\mathcal {F} (E) _ {k} ^ {n}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.544, + 0.788, + 0.605 + ], + "angle": 0, + "content": "where \\(\\parallel\\) channel-wise concatenation and \\(\\mathrm{KV}_k^n\\) denotes the output features for key and value projection, and \\(F_{KV}\\) refers to the ResBlocks layer. We then construct \\(K_k^n\\), \\(V_k^n\\) as \\(K_k^n = W^K (\\mathrm{KV}_k^n)\\) and \\(V_{k}^{n} = W^{V}(\\mathrm{KV}_{k}^{n})\\) where \\(W^{(\\cdot)}\\) denote \\(1\\times 1\\) convolution layer. Then, attention can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.614, + 0.786, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t n} _ {\\mathrm {k}} ^ {\\mathrm {n}} = \\operatorname {S o f t M a x} \\left(\\frac {Q _ {k} ^ {n} \\left(K _ {k} ^ {n}\\right) ^ {T}}{\\alpha}\\right) V _ {k} ^ {n} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.788, + 0.775 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is learnable scaling parameter to balance attention weights and \\(\\mathrm{Att}\\mathfrak{n}_{\\mathrm{n}}\\) is outputs of cross-attention at iteration \\(n\\). To calculate cross-covariance matrix efficiently, we adopted transposed attention [44] for efficient computations. This method enables the efficient computation of attention values at high resolutions by calculating the cross-covariance matrix along the channel axis, leading to a complexity of \\(O(C^2)\\). Moreover, we reduced the number of channels when encoding step, enabling even more efficient operations. We use the attention value \\(\\mathrm{Att}\\mathfrak{n}_{\\mathrm{k}}^{\\mathrm{n}}\\) to update the query iteratively as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.376, + 0.784, + 0.786, + 0.801 + ], + "angle": 0, + "content": "\\[\nQ _ {k} ^ {n + 1} = Q _ {k} ^ {n} + \\operatorname {A t t n} _ {\\mathrm {k}} ^ {\\mathrm {n}} + \\operatorname {M L P} \\left(\\operatorname {A t t n} _ {\\mathrm {k}} ^ {\\mathrm {n}}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where MLP denote multi-layer perceptron. After \\(N\\) iterations, we obtain the updated query feature, \\(Q_{k}^{N}\\). The final query feature passes through up-sampling" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.128 + ], + "angle": 0, + "content": "Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "and skip connections. Then, we generate the final fused features \\(\\{\\mathcal{G}_k\\}\\) as \\(\\mathcal{G}_k = \\mathcal{Q}_k + \\mathrm{Dconv}_{4\\times 4}(Q_k^N)\\) where \\(\\mathrm{Dconv}_{4\\times 4}\\) denote deconvolution layer with a kernel size of 4. As illustrated in Fig. 2, the fused features \\(\\{\\mathcal{G}_k\\}\\) pass through a pyramid encoder, leading to the creation of multi-scale pyramid features, \\(\\{\\mathcal{F}_k^s\\}_{s=0}^2\\). Through the CRIFE module, we can leverage the advantages of abundant temporal information on the events within the exposure time." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.256, + 0.714, + 0.287 + ], + "angle": 0, + "content": "4.3 Event-guided Cascaded Inter-frame Temporal Feature Alignment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.293, + 0.789, + 0.534 + ], + "angle": 0, + "content": "Temporal feature alignment aims to extract valuable information from adjacent video frames. Conventional frame-based video deblurring methods face challenges from inaccurate motion estimation due to motion blur. Conversely, event cameras, thanks to their resistance to motion blur, can offer valuable guidance for aligning video frames. The most straightforward way to align adjacent video frames using the events is to utilize event features to estimate optical flows or deformable offsets between neighboring frames [4,5,58]. While this approach can leverage the advantages of event data for offset and optical flow estimation, it is typically employed after spatial down-sampling due to the high computational costs associated with deformable convolutions [58] and optical flows [28]. Therefore, this approach could restrict performance as it inherently hinders the processing of features at high spatial resolutions in network architectures, limiting access to information across multiple visual scales. To address the aforementioned limitations, we propose a new temporal alignment module that combines the advantages of multiple-visual scale pyramids, leveraging the events' rich temporal contexts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.789, + 0.687 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, our proposed temporal feature alignment modules are structured as multi-level networks that gradually perform coarse-to-fine feature alignment. First, we extract event features, incorporating exposure time information from neighboring video frames to reference frames to facilitate video feature alignment. Specifically, we encode event features encompassing the exposure time interval between \\( t \\) and \\( t + 1 \\), allowing us to connect the frame at time \\( t \\) with the frame at time \\( t + 1 \\). Through this event encoding step, we obtain event feature pyramid set \\( \\{\\varepsilon_{[m,m + 1]}^s\\}_{s = 0}^2 \\) where \\( m \\in \\{t - 2,\\dots,t + 1\\} \\). We use the event feature pyramid containing motion information for adjacent times for the alignment of each blur frame feature pyramid \\( \\{\\mathcal{F}_k^s\\}_{s = 0}^2 \\) where \\( k \\in \\{t - 2,\\dots,t + 2\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.789, + 0.779 + ], + "angle": 0, + "content": "After the event encoding step for the feature alignment, we gradually perform temporal feature alignment from the bottom pyramid level (scale factor \\(s\\) of 2) to the top pyramid level (\\(s\\) of 0). Specifically, as shown in left side of Fig. 4, in each pyramid level \\(s\\) (\\(s \\in \\{0,1,2\\}\\)), We first upsample the hidden state of temporally aligned features at the previous scale, \\(\\hat{\\mathcal{F}}_i^{s+1}\\) through a deconvolution layer and receive them as inputs of alignment module, resulting in \\(h_i^s\\)." + }, + { + "type": "equation", + "bbox": [ + 0.352, + 0.785, + 0.785, + 0.803 + ], + "angle": 0, + "content": "\\[\nh _ {i} ^ {s} = \\operatorname {D c o n v} _ {4 \\times 4} \\left(\\hat {\\mathcal {F}} _ {i} ^ {s + 1}\\right), i \\in \\{t - 1, t, t + 1 \\} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.809, + 0.789, + 0.842 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathcal{F}}_i^{s + 1}\\) denote aligned feature at previous scale \\(s + 1\\), \\(\\mathrm{Dconv}_{4\\times 4}\\) denote \\(4\\times 4\\) deconvolution layer. Note that there is no hidden state at the bottom pyramid" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Cross-Modal Temporal Alignment for Event-guided Video Deblurring" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.146, + 0.481, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.144, + 0.774, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.32, + 0.788, + 0.39 + ], + "angle": 0, + "content": "Fig. 4: Overview of the Event-guided Cascaded Inter-frame Temporal Feature Alignment (ECITFA). The left figure illustrates temporal alignment for scale \\( s \\). The key module for each alignment procedure, Cross-modal Temporal Feature Alignment (CTFA) at time \\( t \\), is illustrated on the right of the figure. The CTFA module operates similarly for reference times \\( t - 1 \\) and \\( t + 1 \\) as well." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.785, + 0.461 + ], + "angle": 0, + "content": "level (\\(s\\) of 2) since there are no features aligned at the previous scale, and for brevity, we focus on the case when \\(s < 2\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.614 + ], + "angle": 0, + "content": "As depicted in the left side of Fig. 4, we progressively group three blurred video frames when performing temporal feature alignment. In other words, we first align with respect to \\( t - 1 \\) using \\( \\{t - 2, t - 1, t\\} \\) and simultaneously align with respect to \\( t + 1 \\) using \\( \\{t, t + 1, t + 2\\} \\) features. Subsequently, we perform the final temporal feature alignment for the previously aligned \\( t - 1 \\) and \\( t + 1 \\) with the last target time \\( t \\). More specifically, when performing frame alignment for the time step \\( t - 1 \\), we first group the three blur video frame features \\( \\mathcal{F}_{t - 2}^s, \\mathcal{F}_{t - 1}^s, \\mathcal{F}_t^s \\) and perform alignment. In this case, we make use of two event features \\( \\varepsilon_{[t - 1,t]}^s \\) and \\( \\varepsilon_{[t,t + 1]}^s \\), which respectively contain motion information between \\( t - 1 \\) and \\( t \\), and between \\( t \\) and \\( t + 1 \\), respectively." + }, + { + "type": "equation", + "bbox": [ + 0.302, + 0.624, + 0.787, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {F}} _ {t - 1} ^ {s} = \\mathrm {C T F A} ^ {s} (\\mathcal {F} _ {t} ^ {s}, \\mathcal {F} _ {t - 1} ^ {s}, \\mathcal {F} _ {t - 2} ^ {s}, h _ {t - 1} ^ {s}, \\varepsilon_ {[ t - 2, t - 1 ]} ^ {s}, \\varepsilon_ {[ t - 1, t ]} ^ {s}), \\qquad (7)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.788, + 0.698 + ], + "angle": 0, + "content": "where \\(\\mathrm{CTFA}^s\\) denote unit Cross-modal Temporal Feature Alignment (CTFA) module at scale factor \\(s\\). Across the same scale factor \\(s\\), CTFA modules are weight-shared. Similarly, we perform alignment for the time step \\(t + 1\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.303, + 0.708, + 0.787, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {F}} _ {t + 1} ^ {s} = \\mathrm {C T F A} ^ {s} (\\mathcal {F} _ {t} ^ {s}, \\mathcal {F} _ {t + 1} ^ {s}, \\mathcal {F} _ {i + 1} ^ {s}, h _ {t + 1} ^ {s}, \\varepsilon_ {[ t + 1, t + 2 ]} ^ {s}, \\varepsilon_ {[ t, t + 1 ]} ^ {s}). \\qquad (8)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.74, + 0.787, + 0.772 + ], + "angle": 0, + "content": "Finally, we utilize the temporally aligned results \\(\\hat{\\mathcal{F}}_{t + 1}^{s}\\) and \\(\\hat{\\mathcal{F}}_{t - 1}^{s}\\) to once again group with \\(\\mathcal{F}_t^s\\) and perform alignment:" + }, + { + "type": "equation", + "bbox": [ + 0.326, + 0.781, + 0.787, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {F}} _ {t} ^ {s} = \\mathrm {C T F A} ^ {s} (\\hat {\\mathcal {F}} _ {t - 1} ^ {s}, \\mathcal {F} _ {t} ^ {s}, \\hat {\\mathcal {F}} _ {t + 1} ^ {s}, h _ {t} ^ {s}, \\varepsilon_ {[ t - 1, t ]} ^ {s}, \\varepsilon_ {[ t, t + 1 ]} ^ {s}). \\qquad (9)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Through this cascaded feature alignment stage, we effectively propagate non-local video frame information, facilitating temporal feature alignment for both" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.128 + ], + "angle": 0, + "content": "Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.254 + ], + "angle": 0, + "content": "non-local and adjacent video frames. All the results of alignment \\(\\hat{\\mathcal{F}}_{t - 1}^{s},\\hat{\\mathcal{F}}_{t}^{s},\\hat{\\mathcal{F}}_{t + 1}^{s}\\) are passed to the next scale, \\(s - 1\\). The right side of Fig. 4 illustrates the overall alignment process for the time step \\(t\\) in the CTFA module. Since it is applied similarly for all time steps \\(i\\in \\{t - 1,t,t + 1\\}\\), we will describe it here specifically when \\(i = t\\), where previously aligned \\(\\hat{\\mathcal{F}}_{t - 1}^{s}\\) and \\(\\hat{\\mathcal{F}}_{t + 1}^{s}\\) are provided. For leveraging the benefits of coarse-to-fine architecture, we first fuse hidden state aligned feature \\(h_t^s\\) from the previous scale for the time step \\(t\\) at scale factor \\(s\\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.439, + 0.281, + 0.786, + 0.298 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {t} ^ {s} = \\mathcal {N} _ {f} ^ {s} \\left(\\mathcal {F} _ {t} ^ {s}, h _ {t} ^ {s}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.308, + 0.787, + 0.369 + ], + "angle": 0, + "content": "We simplify the sequential operations of Conv, ReLU, and ResBlock as \\(\\mathcal{N}_f\\). Additionally, to leverage temporal information, we align the features at times \\(t - 1\\) and \\(t + 1\\) to the current feature at time \\(t\\) using event features that encapsulate motion between frames in the following manner:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.381, + 0.623, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} _ {t - 1 \\rightarrow t} ^ {s} = \\mathcal {N} _ {g, f} ^ {s} (\\hat {\\mathcal {F}} _ {t - 1} ^ {s}, \\mathcal {F} _ {t} ^ {s}, \\varepsilon_ {[ t - 1, t ]} ^ {s}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.403, + 0.785, + 0.423 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} _ {t + 1 \\rightarrow t} ^ {s} = \\mathcal {N} _ {g, b} ^ {s} \\left(\\mathcal {F} _ {t} ^ {s}, \\hat {\\mathcal {F}} _ {t + 1} ^ {s}, \\varepsilon_ {[ t, t + 1 ]} ^ {s}\\right), \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.426, + 0.573, + 0.443 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} _ {t} ^ {s} = \\mathcal {N} _ {h} ^ {s} (\\mathcal {T} _ {t - 1 \\rightarrow t}) \\| \\mathcal {T} _ {t + 1 \\rightarrow t} ^ {s}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.604 + ], + "angle": 0, + "content": "where \\( \\parallel \\) denotes the channel-wise concatenation, and \\( \\mathcal{N}_{g,f}^{s},\\mathcal{N}_{g,b}^{s} \\), and \\( \\mathcal{N}_h^s \\) represent convolution blocks, as illustrated on the right side of Fig. 4. While we obtain temporally aligned features \\( \\mathcal{T}_t^s \\) using event features and leveraging the advantages of events, for additional feature refinement, we utilize the spatially variant pixel-wise dynamic filter [21,55] mechanism. Dynamic filter \\( \\mathcal{D}_t^s \\) at scale \\( s \\) can be calculated through filter generation blocks, \\( \\mathcal{N}_l^s \\), which consists of convolution and resblock, such as \\( \\mathcal{D}_t^s = \\mathcal{N}_l^s (\\mathcal{T}_t^s) \\) where \\( \\mathcal{D}_t^s\\in \\mathbb{R}^{(s_k\\times s_k)\\times H^s\\times W^s} \\), \\( s_k \\) is kernel size of dynamic convolution filter, \\( H^{s} \\), \\( W^{s} \\) denote height and width of the feature at scale factor \\( s \\), respectively. Then, we apply the dynamic convolution operation as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.616, + 0.786, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathcal {T}} _ {t} ^ {s} (h, w) = \\mathcal {D} _ {t} ^ {s} (h, w) \\otimes \\mathcal {T} _ {t} ^ {s} (h, w) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.644, + 0.788, + 0.797 + ], + "angle": 0, + "content": "where \\( h \\in \\{1, \\dots, H^s\\} \\), \\( w \\in \\{1, \\dots, W^s\\} \\), and \\( \\otimes \\) denotes the convolution operation. Then, we employ a cross-attention mechanism [38,44] to effectively combine the information of two features, \\( S_t^s \\) and \\( \\hat{T}_t^s \\). The cross-attention generally examines the correlation between input features (query) and the key-value features. Therefore, we conduct correlation analysis by applying cross-attention between the current blurred frame features to project query \\( S_t^s \\) and the aligned features \\( T_t^s \\) to project key and values. That is, we generate query, key, and value features, \\( \\mathbf{Q} = W_Q(S_t^s) \\), \\( \\mathbf{K} = W_K(\\hat{T}_t^s) \\), \\( \\mathbf{V} = W_V(\\hat{T}_t^s) \\), where \\( W_{(\\cdot)} \\) is \\( 1 \\times 1 \\) convolution and \\( 3 \\times 3 \\) depth-wise convolution. Utilizing these \\( \\mathbf{Q} \\), \\( \\mathbf{K} \\), and \\( \\mathbf{V} \\), we compute the attention matrix similarly to Eq. (4)." + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.808, + 0.786, + 0.839 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {T}}{\\alpha}\\right) \\cdot \\mathbf {V} \\tag {13}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.269, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Cross-Modal Temporal Alignment for Event-guided Video Deblurring" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 1: Quantitative results on the GoPro dataset. The asterisk(*) indicates that the results are not officially on the GoPro so we retrained the official model by us. CMTA-5 and CMTA-7 refer to the results obtained using 5 and 7 input frames, respectively." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.199, + 0.785, + 0.281 + ], + "angle": 0, + "content": "
MethodsMPRNet [45]Restormer [44]NAFNet [6]EDVR [39]ESTRNN [53]RNN-MBP [57]DSTNet [26]VRT [17]
PSNRs32.6632.9233.6926.8331.0233.3234.1634.81
SSIMs0.9590.9610.9670.8430.9110.9630.9680.972
Params(MB)20.126.167.823.62.416.47.518.3
MethodsRVRT [18]Shift-Net [16]UEVD* [15]EFNet [36]REFID [37]SpkNet [7]CMTA-5CMTA-7
PSNRs34.9235.4935.4835.4635.9136.1236.5536.78
SSIMs0.9740.9760.9710.9720.9730.9710.9770.977
Params(MB)10.810.527.98.515.913.59.79.7
" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.308, + 0.651, + 0.325 + ], + "angle": 0, + "content": "Finally, temporally aligned feature, \\(\\hat{\\mathcal{F}}_t^s\\), can be obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.331, + 0.786, + 0.35 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {F}} _ {t} ^ {s} = \\operatorname {M L P} (\\mathcal {A}) + \\mathcal {A}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.419 + ], + "angle": 0, + "content": "where MLP denote multi-layer perceptrons, \\(\\mathcal{A}\\) denote the result of attention operation. Finally, we obtained the aligned feature pyramid \\(\\{\\hat{\\mathcal{F}}_t^s\\}\\) through cascaded temporal feature alignment. As illustrated in Fig. 2, these multi-scale aligned features are fed into the decoder." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.438, + 0.334, + 0.453 + ], + "angle": 0, + "content": "4.4 Decoder" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.46, + 0.788, + 0.524 + ], + "angle": 0, + "content": "The decoder is designed based on the standard U-Net [32]. It takes multi-scale temporally aligned features \\(\\{\\hat{\\mathcal{F}}_t^s\\}\\) as inputs and produces output feature pyramid \\(\\{\\mathcal{F}(D)_k^s\\}\\). The final deblurred outputs \\(S_{t}\\) using last scale of output feature \\(\\mathcal{F}(D)_t^{s = 0}\\) is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.394, + 0.53, + 0.786, + 0.548 + ], + "angle": 0, + "content": "\\[\nS _ {t} = B _ {t} + \\operatorname {C o n v} _ {5 \\times 5} (\\mathcal {F} (D) _ {t} ^ {s = 0}) \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.788, + 0.587 + ], + "angle": 0, + "content": "where \\(\\mathrm{Conv}_{5\\times 5}\\) represents a conv. layer with a filter size of \\(5\\times 5\\), and \\(S_{t}\\) denotes the final estimated sharp frame." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.607, + 0.377, + 0.625 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.636, + 0.336, + 0.65 + ], + "angle": 0, + "content": "5.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.735 + ], + "angle": 0, + "content": "GoPro Dataset [23]. For a fair comparison with other previous event-guided deblurring methods, we utilize the same raw events provided by the authors of recent work [36]. These events were generated using ESIM [29] with a randomly generated contrast threshold set to a Gaussian normal distribution of parameters as \\( N(\\mu = 0.2, \\sigma = 0.03) \\). We used the official train and test splits." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.795 + ], + "angle": 0, + "content": "HighREV Dataset [37]. HighREV consists of high-resolution events and RGB data at \\(1632 \\times 1224\\) resolution, designed for both deblurring and interpolation tasks. To evaluate motion deblurring exclusively, we use the \\(11 + 1\\) split, excluding the interpolation ground truth." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Real-world Video Deblurring Dataset. EVRB dataset consists of 11 training sequences and 6 test sequences. With each sequence containing 149 frames, it is well-suited for video deblurring tasks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.127 + ], + "angle": 0, + "content": "Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.281, + 0.785, + 0.309 + ], + "angle": 0, + "content": "Fig. 5: Visual comparison of deblurring results on the GoPro dataset. The qualitative results of other methods were taken from the results provided by the authors." + }, + { + "type": "table_caption", + "bbox": [ + 0.314, + 0.324, + 0.687, + 0.338 + ], + "angle": 0, + "content": "Table 2: Quantitative results on the HighREV dataset." + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.351, + 0.768, + 0.387 + ], + "angle": 0, + "content": "
MethodsESTRNN [53]DSTNet [26]UEVD [15]EFNet [36]REFID [37]Ours
PSNRs30.3831.7737.4037.9938.3739.12
SSIMs0.9400.9480.9740.9760.9770.980
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.415, + 0.59, + 0.43 + ], + "angle": 0, + "content": "5.2 Comparison on Synthetic Blur Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.788, + 0.685 + ], + "angle": 0, + "content": "We present the quantitative results of our frameworks with other frame-based image and video deblurring methods and event-guided motion deblurring methods using the GoPro dataset as depicted in Table 1. When compared to the existing best performance of video deblurring methods, Shift-Net [16], our approach (CMTA-5 model) shows a significant improvement of \\(1.06\\mathrm{dB}\\) in terms of PSNR, demonstrating that our method effectively leverages the temporal dense characteristics of event modality for accurate temporal feature alignment and cross-modality feature enhancement. Moreover, compared to the best-performing SpkNet [7] among existing event-guided motion deblurring methods, CMTA-5 model exhibits a performance improvement of \\(0.43\\mathrm{dB}\\) with a lower model params of \\(3.8\\mathrm{MB}\\). Furthermore, by using 7 input video frames (CMTA-7 model) instead of 5 (CMTA-5 model), we achieved an impressive state-of-the-art performance with a PSNR of \\(36.78\\mathrm{dB}\\) in the GoPro, an increase of 0.23. We further demonstrated the superiority of our approach through the qualitative results in Fig. 5. Also, we conduct experiments on HighREV [37], which consists of real events, and report the results in Table 2. Our method still achieves the best performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.708, + 0.603, + 0.723 + ], + "angle": 0, + "content": "5.3 Comparison on Real-world Blur Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "For comparisons in the EVRB dataset, we trained representative video deblurring methods [12, 16, 18, 26, 39, 53] and event-guided motion deblurring methods [15, 36, 37] on the same training set. Tab 3 presents the quantitative results on the EVRB dataset. The EVRB dataset includes extremely blurred videos captured during exposure times. As a result, frame-based video deblurring methods exhibit subpar performance. For instance, the best-performing network, ShiftNet [16], achieves only \\(30.56\\mathrm{dB}\\), which is \\(0.42\\mathrm{dB}\\) lower than the best-performing" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Cross-Modal Temporal Alignment for Event-guided Video Deblurring" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.324, + 0.145, + 0.678, + 0.159 + ], + "angle": 0, + "content": "Table 3: Quantitative results on the EVRB dataset." + }, + { + "type": "table", + "bbox": [ + 0.234, + 0.171, + 0.77, + 0.24 + ], + "angle": 0, + "content": "
MethodsEDVR [39]ESTRNN [53]ERDN [12]DSTNet [26]RVRT [18]
PSNRs29.0229.7928.3229.1530.24
SSIMs0.8860.9110.8930.8980.906
MethodsShift-Net [16]UEVD [15]EFNet [36]REFID [37]Ours
PSNRs30.5630.5530.9830.3331.38
SSIMs0.9220.9150.9270.9180.927
" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.253, + 0.787, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.388, + 0.735, + 0.404 + ], + "angle": 0, + "content": "Fig. 6: Visual comparison of deblurring results on the EVRB dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.507 + ], + "angle": 0, + "content": "event-guided deblurring method, EFNet [36]. In contrast, our approach leverages video and event characteristics, effectively restoring even severe motion blur, which may be difficult to recover. Our method outperforms all approaches, achieving the best performance. We show our qualitative comparison with other methods in Fig. 6." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.529, + 0.389, + 0.545 + ], + "angle": 0, + "content": "5.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.787, + 0.599 + ], + "angle": 0, + "content": "We analyzed the performance contribution of the various modules in our frameworks. For a fair ablation study, we trained all the models for 600 epochs with 5 video frame inputs, conducting all experiments on the GoPro dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.689 + ], + "angle": 0, + "content": "CRIFE module. To demonstrate the effectiveness of the CRIFE module, we replaced it with concatenation and convolution for comparison. When comparing the first column(Ver.1) and the second column(Ver.2) of Tab 4, we observed a performance gain of \\(+0.35\\) dB with a small additional parameter \\((+0.08\\mathrm{MB})\\). Similarly, when comparing the performance of the third and fourth rows, we observed performance improvement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.795 + ], + "angle": 0, + "content": "ECITFA module is the most crucial component of our model, performing feature alignment by leveraging information of the events. As in the Tab. 4, when comparing the first (Ver.1) with the third column (Ver.3) of the table, we observed a significant performance gain \\((+1.69\\mathrm{dB})\\) upon the insertion of the ECITFA module. This trend is similarly maintained when the CRIFE module is incorporated. When comparing Ver.2 with Ver.4 of the Tab. 4, we observed a notable performance improvement \\((+1.55\\mathrm{dB})\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Effectiveness of components in ECITFA module. In the Tab. 5, we demonstrated an effectiveness analysis for each component of ECITFA. Comparing the 2nd column with the last column labeled 'Ours', we observed a performance" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.127 + ], + "angle": 0, + "content": "Kim et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.34, + 0.145, + 0.66, + 0.159 + ], + "angle": 0, + "content": "Table 4: Ablation study on the GoPro dataset." + }, + { + "type": "table", + "bbox": [ + 0.264, + 0.171, + 0.74, + 0.22 + ], + "angle": 0, + "content": "
MethodsVer.1Ver.2Ver.3Ver.4
CRIFE✓✓
ECITFA✓✓
PSNRs / Params34.65 / 4.57M35.00 / 4.65M36.34 / 9.59M36.55 / 9.68M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.233, + 0.784, + 0.261 + ], + "angle": 0, + "content": "Table 5: Effect of the component of ECITFA module on the GoPro dataset. DF denotes dynamic filter operation in the Eq.(12)." + }, + { + "type": "table", + "bbox": [ + 0.264, + 0.273, + 0.739, + 0.304 + ], + "angle": 0, + "content": "
Methodsw/o Cascadedw/o DFw/o AttentionOurs
PSNRs36.0836.1836.2936.55
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.316, + 0.784, + 0.345 + ], + "angle": 0, + "content": "Table 6: Comparison of CRIFE with various module variants. \"Baseline\" refers to Ver.3 of Tab.4. E and F represent event and frame features, respectively." + }, + { + "type": "table", + "bbox": [ + 0.261, + 0.357, + 0.738, + 0.395 + ], + "angle": 0, + "content": "
MethodsBaselinew/o concat (E+F)w/concat (E+F)
w/o recurrentw/ recurrent(Ours)
PSNRs36.3436.0435.9936.55
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.423, + 0.785, + 0.544 + ], + "angle": 0, + "content": "gain of our method \\((+0.37\\mathrm{dB})\\) when employing spatial pixel-wise dynamic filters, in contrast to not using DF. When utilizing cross-attention (Eq.13) to better leverage long-range pixel-dependencies in the alignment blocks, we observed a performance improvement \\((+0.29\\mathrm{dB})\\) compared to aggregate feature using ResBlocks. Finally, we confirmed the effectiveness of leveraging non-local video frame information through a cascaded-based temporal feature alignment method for video deblurring. After using the proposed cascaded structure for temporal feature alignment, we can observe a performance gain \\((+0.47\\mathrm{dB})\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.544, + 0.785, + 0.65 + ], + "angle": 0, + "content": "Effectiveness of components in CRIFE module. To evaluate each component's effectiveness of the CRIFE module, we removed the concatenation between RGB and event features, directly matching their features. However, as shown in the 3rd column of Tab. 6, this approach yielded sub-optimal performance. Additionally, combining event features without a recurrent structure degraded performance (see 4th column). These ablations confirm that the recurrent structure effectively utilizes temporal information from events within exposure time." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.673, + 0.367, + 0.689 + ], + "angle": 0, + "content": "6 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.785, + 0.81 + ], + "angle": 0, + "content": "This paper proposes a video deblurring framework, CMTA, that elaborately considers the characteristics of an event and video. Specifically, we achieve significant performance improvement through intra-frame feature enhancement and inter-frame temporal feature alignment. Furthermore, we construct a real-world deblurring dataset, the EVRB dataset, which will be valuable for evaluating event-guided deblurring methods. Finally, CMTA demonstrates state-of-the-art performance across various deblurring datasets." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by the Technology Innovation Program (1415187329,20024355, Development of autonomous driving connec" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.269, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Cross-Modal Temporal Alignment for Event-guided Video Deblurring" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "tivity technology based on sensor-infrastructure cooperation) funded By the Ministry of Trade, Industry & Energy(MOTIE, Korea) and the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (NRF2022R1A2B5B03002636)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.233, + 0.324, + 0.249 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.267, + 0.785, + 0.308 + ], + "angle": 0, + "content": "1. Aittala, M., Durand, F.: Burst image deblurring using permutation invariant convolutional neural networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 731-747 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.785, + 0.351 + ], + "angle": 0, + "content": "2. Brandli, C., Berner, R., Yang, M., Liu, S.C., Delbruck, T.: A \\(240 \\times 180\\) 130 db \\(3\\mu s\\) latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits 49(10), 2333-2341 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.352, + 0.785, + 0.393 + ], + "angle": 0, + "content": "3. Cao, M., Zhong, Z., Wang, J., Zheng, Y., Yang, Y.: Learning adaptive warping for real-world rolling shutter correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17785-17793 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.394, + 0.785, + 0.435 + ], + "angle": 0, + "content": "4. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basic vsr: The search for essential components in video super-resolution and beyond. In: Proceedings of the IEEE conference on computer vision and pattern recognition (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.436, + 0.785, + 0.463 + ], + "angle": 0, + "content": "5. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.464, + 0.785, + 0.492 + ], + "angle": 0, + "content": "6. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: European Conference on Computer Vision. pp. 17-33. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.493, + 0.785, + 0.533 + ], + "angle": 0, + "content": "7. Chen, S., Zhang, J., Zheng, Y., Huang, T., Yu, Z.: Enhancing motion deblurring in high-speed scenes with spike streams. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.534, + 0.785, + 0.575 + ], + "angle": 0, + "content": "8. Cho, H., Jeong, Y., Kim, T., Yoon, K.J.: Non-coaxial event-guided motion deblurring with spatial alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12492-12503 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.577, + 0.785, + 0.63 + ], + "angle": 0, + "content": "9. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence 44(1), 154-180 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.632, + 0.785, + 0.673 + ], + "angle": 0, + "content": "0. Gehrig, D., Loquercio, A., Derpanis, K.G., Scaramuzza, D.: End-to-end learning of representations for asynchronous event-based data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5633-5643 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.674, + 0.785, + 0.715 + ], + "angle": 0, + "content": "1. Hyun Kim, T., Mu Lee, K., Scholkopf, B., Hirsch, M.: Online video deblurring via dynamic temporal blending network. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 4038-4047 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "2. Jiang, B., Xie, Z., Xia, Z., Li, S., Liu, S.: Erdn: Equivalent receptive field deformable network for video deblurring. In: European Conference on Computer Vision. pp. 663-678. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "3. Jiang, Z., Zhang, Y., Zou, D., Ren, J., Lv, J., Liu, Y.: Learning event-based motion deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3320-3329 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "4. Kim, T.H., Sajjadi, M.S., Hirsch, M., Scholkopf, B.: Spatio-temporal transformer network for video restoration. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 106-122 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.267, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.127 + ], + "angle": 0, + "content": "Kim et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "15. Kim, T., Lee, J., Wang, L., Yoon, K.J.: Event-guided deblurring of unknown exposure time videos. In: European Conference on Computer Vision. pp. 519-538. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "16. Li, D., Shi, X., Zhang, Y., Cheung, K.C., See, S., Wang, X., Qin, H., Li, H.: A simple baseline for video restoration with grouped spatial-temporal shift. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9822-9832 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.273 + ], + "angle": 0, + "content": "17. Liang, J., Cao, J., Fan, Y., Zhang, K., Ranjan, R., Li, Y., Timofte, R., Van Gool, L.: Vrt: A video restoration transformer. arXiv preprint arXiv:2201.12288 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.785, + 0.327 + ], + "angle": 0, + "content": "18. Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems 35, 378-393 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "19. Lin, J., Cai, Y., Hu, X., Wang, H., Yan, Y., Zou, X., Ding, H., Zhang, Y., Timofte, R., Van Gool, L.: Flow-guided sparse transformer for video deblurring. arXiv preprint arXiv:2201.01893 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.41 + ], + "angle": 0, + "content": "20. Lin, S., Zhang, J., Pan, J., Jiang, Z., Zou, D., Wang, Y., Chen, J., Ren, J.S.: Learning event-driven video deblurring and interpolation. In: ECCV (8). pp. 695-710 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.453 + ], + "angle": 0, + "content": "21. Mildenhall, B., Barron, J.T., Chen, J., Sharlet, D., Ng, R., Carroll, R.: Burst denoising with kernel prediction networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2502-2510 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.508 + ], + "angle": 0, + "content": "22. Nah, S., Baik, S., Hong, S., Moon, G., Son, S., Timofte, R., Mu Lee, K.: Ntire 2019 challenge on video deblurring and super-resolution: Dataset and study. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "23. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3883-3891 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.592 + ], + "angle": 0, + "content": "24. Oh, J., Kim, M.: Demfi: deep joint deblurring and multi-frame interpolation with flow-guided attentive correlation and recursive boosting. In: European Conference on Computer Vision. pp. 198-215. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.633 + ], + "angle": 0, + "content": "25. Pan, J., Bai, H., Tang, J.: Cascaded deep video deblurring using temporal sharpness prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3043-3051 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.675 + ], + "angle": 0, + "content": "26. Pan, J., Xu, B., Dong, J., Ge, J., Tang, J.: Deep discriminative spatial and temporal network for efficient video deblurring. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (Feb 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.73 + ], + "angle": 0, + "content": "27. Pan, L., Scheerlinck, C., Yu, X., Hartley, R., Liu, M., Dai, Y.: Bringing a blurry frame alive at high frame-rate with an event camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6820-6829 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.772 + ], + "angle": 0, + "content": "28. Ranjan, A., Black, M.J.: Optical flow estimation using a spatial pyramid network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4161-4170 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "29. Rebecq, H., Gehrig, D., Scaramuzza, D.: Esim: an open event camera simulator. In: Conference on Robot Learning. pp. 969-982. PMLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "30. Rim, J., Kim, G., Kim, J., Lee, J., Lee, S., Cho, S.: Realistic blur synthesis for learning image deblurring. In: European conference on computer vision. pp. 487-503. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Cross-Modal Temporal Alignment for Event-guided Video Deblurring" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "31. Rim, J., Lee, H., Won, J., Cho, S.: Real-world blur dataset for learning and benchmarking deblurring algorithms. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16. pp. 184-201. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.788, + 0.259 + ], + "angle": 0, + "content": "32. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.313 + ], + "angle": 0, + "content": "33. Shang, W., Ren, D., Zou, D., Ren, J.S., Luo, P., Zuo, W.: Bringing events into video deblurring with non-consecutively blurry frames. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4531-4540 (October 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.357 + ], + "angle": 0, + "content": "34. Shen, W., Bao, W., Zhai, G., Chen, L., Min, X., Gao, Z.: Video frame interpolation and enhancement via pyramid recurrent framework. IEEE Transactions on Image Processing 30, 277-292 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.398 + ], + "angle": 0, + "content": "35. Su, S., Delbracio, M., Wang, J., Sapiro, G., Heidrich, W., Wang, O.: Deep video deblurring for hand-held cameras. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1279-1288 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.439 + ], + "angle": 0, + "content": "36. Sun, L., Sakaridis, C., Liang, J., Jiang, Q., Yang, K., Sun, P., Ye, Y., Wang, K., Van Gool, L.: Event-based fusion for motion deblurring with cross-modal attention. In: European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.495 + ], + "angle": 0, + "content": "37. Sun, L., Sakaridis, C., Liang, J., Sun, P., Cao, J., Zhang, K., Jiang, Q., Wang, K., Van Gool, L.: Event-based frame interpolation with ad-hoc deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18043-18052 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.495, + 0.788, + 0.522 + ], + "angle": 0, + "content": "38. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NIPS. pp. 5998-6008 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.577 + ], + "angle": 0, + "content": "39. Wang, X., Chan, K.C., Yu, K., Dong, C., Change Loy, C.: Edvr: Video restoration with enhanced deformable convolutional networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (June 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.577, + 0.788, + 0.619 + ], + "angle": 0, + "content": "40. Wang, Y., Lu, Y., Gao, Y., Wang, L., Zhong, Z., Zheng, Y., Yamashita, A.: Efficient video deblurring guided by motion magnitude. In: Proceedings of the European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.661 + ], + "angle": 0, + "content": "41. Wei, K., Fu, Y., Yang, J., Huang, H.: A physics-based noise formation model for extreme low-light raw denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2758-2767 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.661, + 0.788, + 0.688 + ], + "angle": 0, + "content": "42. Wieschollek, P., Hirsch, M., Scholkopf, B., Lensch, H.: Learning blind motion deblurring. In: ICCV. pp. 231-240 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.688, + 0.788, + 0.731 + ], + "angle": 0, + "content": "43. Xu, F., Yu, L., Wang, B., Yang, W., Xia, G.S., Jia, X., Qiao, Z., Liu, J.: Motion deblurring with real events. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2583-2592 (October 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.731, + 0.788, + 0.785 + ], + "angle": 0, + "content": "44. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5728-5739 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.785, + 0.788, + 0.813 + ], + "angle": 0, + "content": "45. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. arXiv preprint arXiv:2102.02808 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.813, + 0.788, + 0.84 + ], + "angle": 0, + "content": "46. Zhang, H., Xie, H., Yao, H.: Spatio-temporal deformable attention network for video deblurring. In: ECCV (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.127 + ], + "angle": 0, + "content": "Kim et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "47. Zhang, K., Luo, W., Zhong, Y., Ma, L., Liu, W., Li, H.: Adversarial spatio-temporal learning for video deblurring. IEEE Transactions on Image Processing (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "48. Zhang, L., Zhang, H., Zhu, C., Guo, S., Chen, J., Wang, L.: Fine-grained video deblurring with event camera. In: ICMM. pp. 352-364. Springer (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.785, + 0.244 + ], + "angle": 0, + "content": "49. Zhang, X., Yu, L.: Unifying motion deblurring and frame interpolation with events. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17765-17774 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.785, + 0.285 + ], + "angle": 0, + "content": "50. Zhang, X., Yu, L., Yang, W., Liu, J., Xia, G.S.: Generalizing event-based motion deblurring in real-world scenarios. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10734-10744 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "51. Zhang, Y., Qin, H., Wang, X., Li, H.: Rethinking noise synthesis and modeling in raw denoising. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4593-4601 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.785, + 0.37 + ], + "angle": 0, + "content": "52. Zhong, Z., Cao, M., Ji, X., Zheng, Y., Sato, I.: Blur interpolation transformer for real-world motion from blur. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5713-5723 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "53. Zhong, Z., Gao, Y., Zheng, Y., Zheng, B.: Efficient spatio-temporal recurrent neural network for video deblurring. In: European Conference on Computer Vision. pp. 191-207. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "54. Zhong, Z., Zheng, Y., Sato, I.: Towards rolling shutter correction and deblurring in dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9219-9228 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.493 + ], + "angle": 0, + "content": "55. Zhou, J., Jampani, V., Pi, Z., Liu, Q., Yang, M.H.: Decoupled dynamic filter networks. In: IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR) (Jun 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.535 + ], + "angle": 0, + "content": "56. Zhu, A.Z., Yuan, L., Chaney, K., Daniilidis, K.: Unsupervised event-based learning of optical flow, depth, and egomotion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 989-997 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.589 + ], + "angle": 0, + "content": "57. Zhu, C., Dong, H., Pan, J., Liang, B., Huang, Y., Fu, L., Wang, F.: Deep recurrent neural network with multi-scale bi-directional propagation for video deblurring. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 3598-3607 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.632 + ], + "angle": 0, + "content": "58. Zhu, X., Hu, H., Lin, S., Dai, J.: Deformable convnets v2: More deformable, better results. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9308-9316 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.632 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CMTA_ Cross-Modal Temporal Alignment for Event-guided Video Deblurring/7cf7582c-c0da-4c34-8883-80e1b8a2e3ef_origin.pdf b/2024/CMTA_ Cross-Modal Temporal Alignment for Event-guided Video Deblurring/7cf7582c-c0da-4c34-8883-80e1b8a2e3ef_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..98f223b65f5454ec23d2f61e66a2b144ddf54ec9 --- /dev/null +++ b/2024/CMTA_ Cross-Modal Temporal Alignment for Event-guided Video Deblurring/7cf7582c-c0da-4c34-8883-80e1b8a2e3ef_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29ed4a5e3fd21e16af3ecdcba4a5af682ebe7871448d3c25aa6f0c90c2c04396 +size 1565962 diff --git a/2024/CMTA_ Cross-Modal Temporal Alignment for Event-guided Video Deblurring/full.md b/2024/CMTA_ Cross-Modal Temporal Alignment for Event-guided Video Deblurring/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7d830b49cafde9bec52bd4703f59b6299cd1b27b --- /dev/null +++ b/2024/CMTA_ Cross-Modal Temporal Alignment for Event-guided Video Deblurring/full.md @@ -0,0 +1,339 @@ +# CMTA: Cross-Modal Temporal Alignment for Event-guided Video Deblurring + +Taewoo Kim $^{1}$ , Hoonhee Cho $^{1}$ , and Kuk-Jin Yoon $^{1}$ + +Korea Advanced Institute of Science and Technology {intelpro, gngsgsm1, kjyoon}@kaist.ac.kr + +Abstract. Video deblurring aims to enhance the quality of restored results in motion-blurred videos by effectively gathering information from adjacent video frames to compensate for the insufficient data in a single blurred frame. However, when faced with consecutively severe motion blur situations, frame-based video deblurring methods often fail to find accurate temporal correspondence among neighboring video frames, leading to diminished performance. To address this limitation, we aim to solve the video deblurring task by leveraging an event camera with microsecond temporal resolution. To fully exploit the dense temporal resolution of the event camera, we propose two modules: 1) Intra-frame feature enhancement operates within the exposure time of a single blurred frame, iteratively enhancing cross-modality features in a recurrent manner to better utilize the rich temporal information of events, 2) Interframe temporal feature alignment gathers valuable long-range temporal information to target frames, aggregating sharp features leveraging the advantages of the events. In addition, we present a novel dataset composed of real-world blurred RGB videos, corresponding sharp videos, and event data. This dataset serves as a valuable resource for evaluating event-guided deblurring methods. We demonstrate that our proposed methods outperform state-of-the-art frame-based and event-based motion deblurring methods through extensive experiments conducted on both synthetic and real-world deblurring datasets. The code and dataset are available at https://github.com/intelpro/CMTA. + +Keywords: Temporal alignment $\cdot$ Video Deblurring $\cdot$ Event cameras + +# 1 Introduction + +Motion blur is a common artifact caused by dynamic movements within a scene or camera motion during exposure. Motion deblurring, which aims to reverse the blurring process, presents significant challenges due to variations in blur intensity influenced by scene structure and depth. To achieve high-quality deblurring, video deblurring has emerged, leveraging information from neighboring frames instead of relying solely on a single blurred image. However, identifying temporal correspondence between blurred video frames becomes challenging with + +extreme motion blur, hindering the extraction of valuable information from adjacent frames and impeding performance improvement. + +Event cameras [9], with their extremely low latency (on the order of microseconds), can offer high-quality guidance for motion deblurring due to their ability to capture high-temporal resolution of brightness change. To effectively utilize the advantages of the events, several event-guided motion deblurring works [15, 36, 50] have been introduced. While these works have typically explored cross-modal feature fusion methods across different modalities, there has been limited work on leveraging the abundant temporal information in videos. + +To obtain high-quality results, we emphasize the event camera's temporal continuity, focusing on its interaction with video frames that exhibit long-term temporal dependencies. Unlike previous event-guided motion deblurring works [36,43,50], which relied on a single blurry image and corresponding events on its exposure time, we further design precise temporal feature alignment methods between neighboring video frames by leveraging the advantages of event data. Specifically, we propose novel modules from two perspectives: intra-frame (interaction between events and frames within the exposure time) and inter-frame (interaction between different frames) perspectives. + +From an intra-frame perspective within exposure time, we propose a Cross-modal Recurrent Intra-frame Feature Enhancement (CRIFE) module to better leverage the rich temporal information of the events by mutually interacting the blur frame and event features within the duration of exposure time. In this module, we perform recurrent attention-based feature enhancement using a transformer [38] that better captures long-range pixel dependencies. + +In temporal feature alignment with the second perspective, we propose a novel event-guided temporal feature alignment module, effectively leveraging rich temporal characteristics of the events. Conventional frame-based temporal feature alignment methods rely on optical flow [28] or deformable convolution [5,58]. While optical flow and deformable convolutions aid in achieving temporal feature alignment, the high computational complexity of these operations makes it challenging to execute them at a higher spatial scale. Therefore, video frame alignment, generally conducted at lower spatial scales, leads to sub-optimal deblurring results due to the lack of spatial contexts of features. To overcome these limitations, our temporal feature alignment module avoids relying on optical flow or deformable convolution by effectively leveraging the temporally dense advantages of the events. Therefore, we can effectively perform temporal feature alignment across multiple visual scales as we do not rely on these complex operations but rather efficiently leverage the temporal information from the events. As a result, our temporal feature alignment module demonstrates a significant performance improvement in event-guided video deblurring tasks. Since we have introduced a pioneering method for aligning temporal features using the advantages of the events, it is expected to be effectively applicable to various event-guided video restoration tasks (e.g., event-guided video super-resolution). + +Finally, we propose a novel video deblurring dataset, the EVRB dataset, composed of high-quality RGB and event data. It consists of real-world blurry + +videos generated by extreme motion and corresponding sharp videos for generalized applications in real-world scenarios. The network trained on the EVRB dataset can be directly applicable to real-world scenarios, making it a valuable resource for event-guided deblurring research. + +# 2 Related Works + +# 2.1 Video Deblurring + +Early works [1, 35] employed CNNs that take the concatenation of adjacent frames as input to address video deblurring. Subsequently, to better leverage temporal information, approaches have emerged that employ 3D convolutions [26, 47], temporal alignment modules with deformable convolutions [39] and optical flow [14, 25, 40, 46]. As an alternative video alignment method, some works utilize RNN [11, 42, 53] and transformer [18, 19] structures to propagate information from long-range video frames. However, as the intensity of motion blur in the video increases, frame-based video alignment methods struggle to achieve accurate video alignment, leading to sub-optimal deblurring results. + +# 2.2 Event-guided Motion Deblurring + +An event camera can effectively be used for motion deblurring as it records motion information corresponding to the brightness differences with high temporal resolution. Efforts to utilize event cameras for motion deblurring [8, 13, 15, 20, 27, 36, 37, 43, 48-50] have been actively ongoing. Recent studies have focused on effectively fusing event and image features of different modalities. To this end, Sun et al. [36] employed a transformer architecture for feature fusion. Zhang et al. [50] effectively combined modalities using a multi-scale architecture. In addition, there are also attempts to address motion deblurring by assuming challenging and general scenarios [8, 15, 33]. However, these studies have primarily focused on single image deblurring and do not exploit the long-range temporal information demonstrated in previous video deblurring tasks. To effectively acquire information that may be missing from sparse events, we introduce a novel method for accurate temporal alignment with events, utilizing information from surrounding adjacent frames in video deblurring. + +# 3 Event-based Video Deblurring Dataset for Real-world Blur + +# 3.1 Limitation of Synthetic Blur Dataset + +Typically, the blurred images synthesis procedure adheres to the methodology outlined in [22-24, 34, 35], involving the averaging of consecutive video frames within a fixed-size window. However, as recent studies [31, 52, 54] have discussed, blur synthesis based on discrete signals may result in shutter artifacts even when + +![](images/2aa45a1d09c9f77a410f2e52ea77c3c95b8fe3b80ad3c1d1f948fb53b0bd39b5.jpg) +Fig. 1: Illustration of a hybrid camera system for real-world event-based video deblurring dataset. $S$ and $B$ denote the cameras for acquiring sharp and blur videos, respectively. (a): The triple-axis camera system to capture real-world blur. (b): A diagram of our hybrid camera system. (c): Samples from our EVRB dataset with natural blur. + +![](images/c7c93452eb3c45e081aef27acf4743df0783247ecf107f0b49cbb9922989e105.jpg) + +![](images/1876ae1d4afcaad698acabc3cd91eebeb08d9d35ee932ea9a4bab4e377375eda.jpg) + +averaging high-frame-rate videos. Furthermore, simply averaging sharp images disregards essential elements in the blurred images such as pixel saturation [30], limitations due to dynamic range, and physical noise [41, 51] in data acquired during exposure time. Furthermore, in the case of the event-guided motion deblurring research, existing works have enhanced the synthetic aspects by using event simulator [29]. To address these limitations, recent works have attempted to acquire real-world blur datasets using hybrid camera systems [3, 31, 52-54]. Meanwhile, to generalize event-based motion deblurring, Sun et al. [36] introduced the ReBlur dataset acquired in a high-precision optical laboratory using an electronic-controlled slide-rail system. However, despite these successes, the ReBlur dataset has limitations due to its indoor setting, a lack of dynamic objects, and the absence of high-quality RGB data as it is acquired using the DAVIS sensor [2]. Additionally, the sequences are relatively short(minimum six blurred frames in the sequence), making it challenging for use in event-guided video deblurring. We have acquired a new EVRB dataset for evaluating event-guided image and video deblurring methods in real-world blurry videos. The EVRB dataset was captured using a customized hybrid system consisting of two RGB cameras and one event camera. This system encompasses a range of blur magnitude from slight to extreme in various urban environments. + +# 3.2 Triple-axis Hybrid Camera System + +As shown in Fig. 1, we design a hybrid camera system to enable the acquisition of different data sources to be geometrically aligned. Two RGB cameras and one event camera are geometrically aligned using two $50/50$ cube beam splitters, resulting in a minimal baseline. We perform pixel-wise alignment between multiple cameras based on the homography calculated by extrinsic calibration for precise alignment. For the synchronization of the multiple cameras, we use an external trigger system. Each camera receives the falling and rising edges of the trigger + +![](images/bb7afde7f4dc31ede6443dc3b7f34963307980eabe143e38a3326855983ba012.jpg) +Fig. 2: Overall framework of CMTA is divided into two main components: Cross-modal Recurrent Intra-frame Feature Enhancement (CRIFE) and Event-guided Cascaded Inter-frame Temporal Feature Alignment (ECITFA). $s$ is the scale factor for multi-scale features. In the figure of the ECITFA module, the description was performed for the case of $P = 2$ for simplicity. + +signal and acquires data synchronized with the period of the signal. This external trigger can precisely control the exposure times of the two RGB cameras, enabling us to capture paired sharp and blurred video frames. Furthermore, for photometric alignment, we physically adjust the amount of incoming light for both cameras to equalize the total irradiance of the multiple cameras using a neutral density filter. + +# 4 Method + +# 4.1 Overview + +The overview of the proposed framework is illustrated in Fig. 2. Given consecutive blurred video frames $\{B_k\}$ and sets of event streams corresponding to the exposure time of each video frame $\{\mathbb{E}_k\}$ , where $k \in \{t - P, \dots, t, \dots, t + P\}$ , our goal is to estimate the latent sharp video frame $S_t$ . To utilize the event stream $\{\mathbb{E}_k\}$ corresponding to the exposure time of $\{B_k\}$ as the input for the networks, we first perform embedding using the event voxel grid representation [56] for the event stream $\{\mathbb{E}_k\}$ , resulting in $\{E_k\}$ . Our framework consists of two main submodules: (1) Cross-modal Recurrent Intra-frame Feature Enhancement (CRIFE) module and (2) Event-guided Cascaded Inter-frame Temporal Feature Alignment (ECITFA) module. In the first module, we perform cross-modal feature enhancement through recurrent interactions between blurred frame features and event features to leverage the continuous temporal information of events within the exposure time. After obtaining the fused feature $\{\mathcal{G}_k\}$ from the first module, we generate multi-scale features, $\{\mathcal{F}_k^s\}_{s=0}^2$ , through a weight-shared pyra + +![](images/e417d8010bde286557f1bbfc66ea431af1ce7ad9ff133596eac8865891e6b4b2.jpg) +Fig. 3: Illustration of Cross-modal Recurrent Intra-frame Feature Enhancement (CRIFE). + +mid encoder. Subsequently, for the temporal feature alignment stage, we encode multi-scale event features, $\{\varepsilon_{[k,k + 1]}^s\}_{s = 0}^2$ , by grouping two consecutive events corresponding to the exposure times of each frame to connect adjacent frames. By utilizing the multi-scale pyramid features $\{\mathcal{F}_k^s\}_{s = 0}^2$ and the encoded event feature $\{\varepsilon_{[k,k + 1]}^s\}_{s = 0}^2$ , we perform temporal feature alignment. Afterward, temporally aligned feature pyramids are fed into the U-Net [32]-based decoder, generating the final sharp video frame $S_{t}$ . Our framework can be extended to cases where $P$ is an arbitrary positive number. However, for the sake of brevity, we will explain it in the main text with $P = 2$ . + +# 4.2 Cross-modal Recurrent Intra-frame Feature Enhancement + +Since event cameras provide rich temporal information on brightness changes, it is crucial to effectively utilize this dense temporal information of the events within the duration of the exposure time. Typically, event data captured during the exposure time is transformed into event embeddings [10,56] as input to the network, followed by feature extraction using 2D CNNs. However, these approaches cannot effectively utilize the continuous temporal information of events, and these limitations can impact the performance of event-guided video deblurring. To address these limitations, we propose a method that leverages the rich temporal nature of events and fuses event and blurred frame features using recurrent-based attention methods. + +Recently, researchers have demonstrated the effectiveness of transformer-based architectures [38] by capturing long-range pixel dependencies, which have been proven highly effective in various vision tasks. We leverage transformers' advantages to more effectively utilize the temporal benefits of the events through a recurrent-based approach. For each blur video frame index $k \in \{t - 2, \dots, t + 2\}$ , we initially partition the exposure time $T_{exp,k}$ into $N$ unit temporal intervals, + +denoted as $\Delta t$ , with $\Delta t = T_{exp,k} / N$ . Based on the time interval $\Delta t$ , we divide the event voxel grid within the exposure time, denoted as $E_k \in \mathbb{R}^{C \times H \times W}$ into $N$ temporally divided event voxel grids $\{E_k^n\}$ , where $E_k^n \in \mathbb{R}^{C / N \times H \times W}$ and $n \in \{0, \dots, N - 1\}$ . After, we extract event features $\{\mathcal{F}(E)_k^n\}$ of the temporally divided event voxel grids $E_k^n$ using weight-sharing event feature extractor. To apply cross-attention, we construct a query encompassing global information of blur and temporally divided event features. That is, we concatenate the blur frame feature $\mathcal{F}(B)_k$ with the temporally segmented event sets $\{\mathcal{F}(E)_k^n\}$ and then extract the feature for the encoding query information. + +$$ +\mathcal {Q} _ {k} = \operatorname {C o n v} _ {\mathrm {p}} (\mathcal {F} (B) _ {k} \| \left\{\mathcal {F} (E) _ {k} ^ {n} \right\}) \tag {1} +$$ + +$$ +Q _ {k} ^ {n = 0} = F _ {R} \left(\mathcal {Q} _ {k}\right) \tag {2} +$$ + +where $\mathrm{Conv}_{\mathrm{p}}$ represents point-wise convolution, $F_{R}$ consists of a sequence of blocks with $3\times 3$ convolution and a stride of 2, along with ResBlocks. Additionally, $Q_{k}^{n = 0}$ denotes the initial query feature for the recurrent-based attention method. + +As illustrated in Fig. 3, we perform cross-attention to update query information iteratively. To find the key and value of cross-attention for query updating, we recursively input temporally separated event features, helping to better utilize the events' rich temporal information. That is, for iteration $n$ , we extract features to be used as keys and values by concatenating the previously updated $Q_{k}^{n}$ with temporally divided $n$ -th event features $\mathcal{F}(E)_k^n$ as follows: + +$$ +\mathrm {K V} _ {k} ^ {n} = F _ {K V} \left(Q _ {k} ^ {n} \| \mathcal {F} (E) _ {k} ^ {n}\right) \tag {3} +$$ + +where $\parallel$ channel-wise concatenation and $\mathrm{KV}_k^n$ denotes the output features for key and value projection, and $F_{KV}$ refers to the ResBlocks layer. We then construct $K_k^n$ , $V_k^n$ as $K_k^n = W^K (\mathrm{KV}_k^n)$ and $V_{k}^{n} = W^{V}(\mathrm{KV}_{k}^{n})$ where $W^{(\cdot)}$ denote $1\times 1$ convolution layer. Then, attention can be calculated as: + +$$ +\operatorname {A t t n} _ {\mathrm {k}} ^ {\mathrm {n}} = \operatorname {S o f t M a x} \left(\frac {Q _ {k} ^ {n} \left(K _ {k} ^ {n}\right) ^ {T}}{\alpha}\right) V _ {k} ^ {n} \tag {4} +$$ + +where $\alpha$ is learnable scaling parameter to balance attention weights and $\mathrm{Att}\mathfrak{n}_{\mathrm{n}}$ is outputs of cross-attention at iteration $n$ . To calculate cross-covariance matrix efficiently, we adopted transposed attention [44] for efficient computations. This method enables the efficient computation of attention values at high resolutions by calculating the cross-covariance matrix along the channel axis, leading to a complexity of $O(C^2)$ . Moreover, we reduced the number of channels when encoding step, enabling even more efficient operations. We use the attention value $\mathrm{Att}\mathfrak{n}_{\mathrm{k}}^{\mathrm{n}}$ to update the query iteratively as follows: + +$$ +Q _ {k} ^ {n + 1} = Q _ {k} ^ {n} + \operatorname {A t t n} _ {\mathrm {k}} ^ {\mathrm {n}} + \operatorname {M L P} \left(\operatorname {A t t n} _ {\mathrm {k}} ^ {\mathrm {n}}\right) \tag {5} +$$ + +where MLP denote multi-layer perceptron. After $N$ iterations, we obtain the updated query feature, $Q_{k}^{N}$ . The final query feature passes through up-sampling + +and skip connections. Then, we generate the final fused features $\{\mathcal{G}_k\}$ as $\mathcal{G}_k = \mathcal{Q}_k + \mathrm{Dconv}_{4\times 4}(Q_k^N)$ where $\mathrm{Dconv}_{4\times 4}$ denote deconvolution layer with a kernel size of 4. As illustrated in Fig. 2, the fused features $\{\mathcal{G}_k\}$ pass through a pyramid encoder, leading to the creation of multi-scale pyramid features, $\{\mathcal{F}_k^s\}_{s=0}^2$ . Through the CRIFE module, we can leverage the advantages of abundant temporal information on the events within the exposure time. + +# 4.3 Event-guided Cascaded Inter-frame Temporal Feature Alignment + +Temporal feature alignment aims to extract valuable information from adjacent video frames. Conventional frame-based video deblurring methods face challenges from inaccurate motion estimation due to motion blur. Conversely, event cameras, thanks to their resistance to motion blur, can offer valuable guidance for aligning video frames. The most straightforward way to align adjacent video frames using the events is to utilize event features to estimate optical flows or deformable offsets between neighboring frames [4,5,58]. While this approach can leverage the advantages of event data for offset and optical flow estimation, it is typically employed after spatial down-sampling due to the high computational costs associated with deformable convolutions [58] and optical flows [28]. Therefore, this approach could restrict performance as it inherently hinders the processing of features at high spatial resolutions in network architectures, limiting access to information across multiple visual scales. To address the aforementioned limitations, we propose a new temporal alignment module that combines the advantages of multiple-visual scale pyramids, leveraging the events' rich temporal contexts. + +As illustrated in Fig. 2, our proposed temporal feature alignment modules are structured as multi-level networks that gradually perform coarse-to-fine feature alignment. First, we extract event features, incorporating exposure time information from neighboring video frames to reference frames to facilitate video feature alignment. Specifically, we encode event features encompassing the exposure time interval between $t$ and $t + 1$ , allowing us to connect the frame at time $t$ with the frame at time $t + 1$ . Through this event encoding step, we obtain event feature pyramid set $\{\varepsilon_{[m,m + 1]}^s\}_{s = 0}^2$ where $m \in \{t - 2,\dots,t + 1\}$ . We use the event feature pyramid containing motion information for adjacent times for the alignment of each blur frame feature pyramid $\{\mathcal{F}_k^s\}_{s = 0}^2$ where $k \in \{t - 2,\dots,t + 2\}$ . + +After the event encoding step for the feature alignment, we gradually perform temporal feature alignment from the bottom pyramid level (scale factor $s$ of 2) to the top pyramid level ( $s$ of 0). Specifically, as shown in left side of Fig. 4, in each pyramid level $s$ ( $s \in \{0,1,2\}$ ), We first upsample the hidden state of temporally aligned features at the previous scale, $\hat{\mathcal{F}}_i^{s+1}$ through a deconvolution layer and receive them as inputs of alignment module, resulting in $h_i^s$ . + +$$ +h _ {i} ^ {s} = \operatorname {D c o n v} _ {4 \times 4} \left(\hat {\mathcal {F}} _ {i} ^ {s + 1}\right), i \in \{t - 1, t, t + 1 \} \tag {6} +$$ + +where $\hat{\mathcal{F}}_i^{s + 1}$ denote aligned feature at previous scale $s + 1$ , $\mathrm{Dconv}_{4\times 4}$ denote $4\times 4$ deconvolution layer. Note that there is no hidden state at the bottom pyramid + +![](images/359efd6ea5ed3303e309257429825178ab57665602d237ce7dfecf55fa139eb2.jpg) +Fig. 4: Overview of the Event-guided Cascaded Inter-frame Temporal Feature Alignment (ECITFA). The left figure illustrates temporal alignment for scale $s$ . The key module for each alignment procedure, Cross-modal Temporal Feature Alignment (CTFA) at time $t$ , is illustrated on the right of the figure. The CTFA module operates similarly for reference times $t - 1$ and $t + 1$ as well. + +![](images/20408f853ddba7bd47faf5758dcf8bdd8213f38f17c04faa81a84a5b9e0255b7.jpg) + +level ( $s$ of 2) since there are no features aligned at the previous scale, and for brevity, we focus on the case when $s < 2$ . + +As depicted in the left side of Fig. 4, we progressively group three blurred video frames when performing temporal feature alignment. In other words, we first align with respect to $t - 1$ using $\{t - 2, t - 1, t\}$ and simultaneously align with respect to $t + 1$ using $\{t, t + 1, t + 2\}$ features. Subsequently, we perform the final temporal feature alignment for the previously aligned $t - 1$ and $t + 1$ with the last target time $t$ . More specifically, when performing frame alignment for the time step $t - 1$ , we first group the three blur video frame features $\mathcal{F}_{t - 2}^s, \mathcal{F}_{t - 1}^s, \mathcal{F}_t^s$ and perform alignment. In this case, we make use of two event features $\varepsilon_{[t - 1,t]}^s$ and $\varepsilon_{[t,t + 1]}^s$ , which respectively contain motion information between $t - 1$ and $t$ , and between $t$ and $t + 1$ , respectively. + +$$ +\hat {\mathcal {F}} _ {t - 1} ^ {s} = \mathrm {C T F A} ^ {s} (\mathcal {F} _ {t} ^ {s}, \mathcal {F} _ {t - 1} ^ {s}, \mathcal {F} _ {t - 2} ^ {s}, h _ {t - 1} ^ {s}, \varepsilon_ {[ t - 2, t - 1 ]} ^ {s}, \varepsilon_ {[ t - 1, t ]} ^ {s}), \qquad (7) +$$ + +where $\mathrm{CTFA}^s$ denote unit Cross-modal Temporal Feature Alignment (CTFA) module at scale factor $s$ . Across the same scale factor $s$ , CTFA modules are weight-shared. Similarly, we perform alignment for the time step $t + 1$ as: + +$$ +\hat {\mathcal {F}} _ {t + 1} ^ {s} = \mathrm {C T F A} ^ {s} (\mathcal {F} _ {t} ^ {s}, \mathcal {F} _ {t + 1} ^ {s}, \mathcal {F} _ {i + 1} ^ {s}, h _ {t + 1} ^ {s}, \varepsilon_ {[ t + 1, t + 2 ]} ^ {s}, \varepsilon_ {[ t, t + 1 ]} ^ {s}). \qquad (8) +$$ + +Finally, we utilize the temporally aligned results $\hat{\mathcal{F}}_{t + 1}^{s}$ and $\hat{\mathcal{F}}_{t - 1}^{s}$ to once again group with $\mathcal{F}_t^s$ and perform alignment: + +$$ +\hat {\mathcal {F}} _ {t} ^ {s} = \mathrm {C T F A} ^ {s} (\hat {\mathcal {F}} _ {t - 1} ^ {s}, \mathcal {F} _ {t} ^ {s}, \hat {\mathcal {F}} _ {t + 1} ^ {s}, h _ {t} ^ {s}, \varepsilon_ {[ t - 1, t ]} ^ {s}, \varepsilon_ {[ t, t + 1 ]} ^ {s}). \qquad (9) +$$ + +Through this cascaded feature alignment stage, we effectively propagate non-local video frame information, facilitating temporal feature alignment for both + +non-local and adjacent video frames. All the results of alignment $\hat{\mathcal{F}}_{t - 1}^{s},\hat{\mathcal{F}}_{t}^{s},\hat{\mathcal{F}}_{t + 1}^{s}$ are passed to the next scale, $s - 1$ . The right side of Fig. 4 illustrates the overall alignment process for the time step $t$ in the CTFA module. Since it is applied similarly for all time steps $i\in \{t - 1,t,t + 1\}$ , we will describe it here specifically when $i = t$ , where previously aligned $\hat{\mathcal{F}}_{t - 1}^{s}$ and $\hat{\mathcal{F}}_{t + 1}^{s}$ are provided. For leveraging the benefits of coarse-to-fine architecture, we first fuse hidden state aligned feature $h_t^s$ from the previous scale for the time step $t$ at scale factor $s$ , as follows: + +$$ +\mathcal {S} _ {t} ^ {s} = \mathcal {N} _ {f} ^ {s} \left(\mathcal {F} _ {t} ^ {s}, h _ {t} ^ {s}\right), \tag {10} +$$ + +We simplify the sequential operations of Conv, ReLU, and ResBlock as $\mathcal{N}_f$ . Additionally, to leverage temporal information, we align the features at times $t - 1$ and $t + 1$ to the current feature at time $t$ using event features that encapsulate motion between frames in the following manner: + +$$ +\mathcal {T} _ {t - 1 \rightarrow t} ^ {s} = \mathcal {N} _ {g, f} ^ {s} (\hat {\mathcal {F}} _ {t - 1} ^ {s}, \mathcal {F} _ {t} ^ {s}, \varepsilon_ {[ t - 1, t ]} ^ {s}), +$$ + +$$ +\mathcal {T} _ {t + 1 \rightarrow t} ^ {s} = \mathcal {N} _ {g, b} ^ {s} \left(\mathcal {F} _ {t} ^ {s}, \hat {\mathcal {F}} _ {t + 1} ^ {s}, \varepsilon_ {[ t, t + 1 ]} ^ {s}\right), \tag {11} +$$ + +$$ +\mathcal {T} _ {t} ^ {s} = \mathcal {N} _ {h} ^ {s} (\mathcal {T} _ {t - 1 \rightarrow t}) \| \mathcal {T} _ {t + 1 \rightarrow t} ^ {s}), +$$ + +where $\parallel$ denotes the channel-wise concatenation, and $\mathcal{N}_{g,f}^{s},\mathcal{N}_{g,b}^{s}$ , and $\mathcal{N}_h^s$ represent convolution blocks, as illustrated on the right side of Fig. 4. While we obtain temporally aligned features $\mathcal{T}_t^s$ using event features and leveraging the advantages of events, for additional feature refinement, we utilize the spatially variant pixel-wise dynamic filter [21,55] mechanism. Dynamic filter $\mathcal{D}_t^s$ at scale $s$ can be calculated through filter generation blocks, $\mathcal{N}_l^s$ , which consists of convolution and resblock, such as $\mathcal{D}_t^s = \mathcal{N}_l^s (\mathcal{T}_t^s)$ where $\mathcal{D}_t^s\in \mathbb{R}^{(s_k\times s_k)\times H^s\times W^s}$ , $s_k$ is kernel size of dynamic convolution filter, $H^{s}$ , $W^{s}$ denote height and width of the feature at scale factor $s$ , respectively. Then, we apply the dynamic convolution operation as follows: + +$$ +\tilde {\mathcal {T}} _ {t} ^ {s} (h, w) = \mathcal {D} _ {t} ^ {s} (h, w) \otimes \mathcal {T} _ {t} ^ {s} (h, w) \tag {12} +$$ + +where $h \in \{1, \dots, H^s\}$ , $w \in \{1, \dots, W^s\}$ , and $\otimes$ denotes the convolution operation. Then, we employ a cross-attention mechanism [38,44] to effectively combine the information of two features, $S_t^s$ and $\hat{T}_t^s$ . The cross-attention generally examines the correlation between input features (query) and the key-value features. Therefore, we conduct correlation analysis by applying cross-attention between the current blurred frame features to project query $S_t^s$ and the aligned features $T_t^s$ to project key and values. That is, we generate query, key, and value features, $\mathbf{Q} = W_Q(S_t^s)$ , $\mathbf{K} = W_K(\hat{T}_t^s)$ , $\mathbf{V} = W_V(\hat{T}_t^s)$ , where $W_{(\cdot)}$ is $1 \times 1$ convolution and $3 \times 3$ depth-wise convolution. Utilizing these $\mathbf{Q}$ , $\mathbf{K}$ , and $\mathbf{V}$ , we compute the attention matrix similarly to Eq. (4). + +$$ +\mathcal {A} (\mathbf {Q}, \mathbf {K}, \mathbf {V}) = \operatorname {S o f t m a x} \left(\frac {\mathbf {Q} \mathbf {K} ^ {T}}{\alpha}\right) \cdot \mathbf {V} \tag {13} +$$ + +Table 1: Quantitative results on the GoPro dataset. The asterisk(*) indicates that the results are not officially on the GoPro so we retrained the official model by us. CMTA-5 and CMTA-7 refer to the results obtained using 5 and 7 input frames, respectively. + +
MethodsMPRNet [45]Restormer [44]NAFNet [6]EDVR [39]ESTRNN [53]RNN-MBP [57]DSTNet [26]VRT [17]
PSNRs32.6632.9233.6926.8331.0233.3234.1634.81
SSIMs0.9590.9610.9670.8430.9110.9630.9680.972
Params(MB)20.126.167.823.62.416.47.518.3
MethodsRVRT [18]Shift-Net [16]UEVD* [15]EFNet [36]REFID [37]SpkNet [7]CMTA-5CMTA-7
PSNRs34.9235.4935.4835.4635.9136.1236.5536.78
SSIMs0.9740.9760.9710.9720.9730.9710.9770.977
Params(MB)10.810.527.98.515.913.59.79.7
+ +Finally, temporally aligned feature, $\hat{\mathcal{F}}_t^s$ , can be obtained by: + +$$ +\hat {\mathcal {F}} _ {t} ^ {s} = \operatorname {M L P} (\mathcal {A}) + \mathcal {A}. \tag {14} +$$ + +where MLP denote multi-layer perceptrons, $\mathcal{A}$ denote the result of attention operation. Finally, we obtained the aligned feature pyramid $\{\hat{\mathcal{F}}_t^s\}$ through cascaded temporal feature alignment. As illustrated in Fig. 2, these multi-scale aligned features are fed into the decoder. + +# 4.4 Decoder + +The decoder is designed based on the standard U-Net [32]. It takes multi-scale temporally aligned features $\{\hat{\mathcal{F}}_t^s\}$ as inputs and produces output feature pyramid $\{\mathcal{F}(D)_k^s\}$ . The final deblurred outputs $S_{t}$ using last scale of output feature $\mathcal{F}(D)_t^{s = 0}$ is calculated as follows: + +$$ +S _ {t} = B _ {t} + \operatorname {C o n v} _ {5 \times 5} (\mathcal {F} (D) _ {t} ^ {s = 0}) \tag {15} +$$ + +where $\mathrm{Conv}_{5\times 5}$ represents a conv. layer with a filter size of $5\times 5$ , and $S_{t}$ denotes the final estimated sharp frame. + +# 5 Experiments + +# 5.1 Datasets + +GoPro Dataset [23]. For a fair comparison with other previous event-guided deblurring methods, we utilize the same raw events provided by the authors of recent work [36]. These events were generated using ESIM [29] with a randomly generated contrast threshold set to a Gaussian normal distribution of parameters as $N(\mu = 0.2, \sigma = 0.03)$ . We used the official train and test splits. + +HighREV Dataset [37]. HighREV consists of high-resolution events and RGB data at $1632 \times 1224$ resolution, designed for both deblurring and interpolation tasks. To evaluate motion deblurring exclusively, we use the $11 + 1$ split, excluding the interpolation ground truth. + +Real-world Video Deblurring Dataset. EVRB dataset consists of 11 training sequences and 6 test sequences. With each sequence containing 149 frames, it is well-suited for video deblurring tasks. + +![](images/d297df95791baffe6f5ee30a8cad71b94ba84face5532a358f688724c884a8c7.jpg) +Fig. 5: Visual comparison of deblurring results on the GoPro dataset. The qualitative results of other methods were taken from the results provided by the authors. + +Table 2: Quantitative results on the HighREV dataset. + +
MethodsESTRNN [53]DSTNet [26]UEVD [15]EFNet [36]REFID [37]Ours
PSNRs30.3831.7737.4037.9938.3739.12
SSIMs0.9400.9480.9740.9760.9770.980
+ +# 5.2 Comparison on Synthetic Blur Datasets + +We present the quantitative results of our frameworks with other frame-based image and video deblurring methods and event-guided motion deblurring methods using the GoPro dataset as depicted in Table 1. When compared to the existing best performance of video deblurring methods, Shift-Net [16], our approach (CMTA-5 model) shows a significant improvement of $1.06\mathrm{dB}$ in terms of PSNR, demonstrating that our method effectively leverages the temporal dense characteristics of event modality for accurate temporal feature alignment and cross-modality feature enhancement. Moreover, compared to the best-performing SpkNet [7] among existing event-guided motion deblurring methods, CMTA-5 model exhibits a performance improvement of $0.43\mathrm{dB}$ with a lower model params of $3.8\mathrm{MB}$ . Furthermore, by using 7 input video frames (CMTA-7 model) instead of 5 (CMTA-5 model), we achieved an impressive state-of-the-art performance with a PSNR of $36.78\mathrm{dB}$ in the GoPro, an increase of 0.23. We further demonstrated the superiority of our approach through the qualitative results in Fig. 5. Also, we conduct experiments on HighREV [37], which consists of real events, and report the results in Table 2. Our method still achieves the best performance. + +# 5.3 Comparison on Real-world Blur Datasets + +For comparisons in the EVRB dataset, we trained representative video deblurring methods [12, 16, 18, 26, 39, 53] and event-guided motion deblurring methods [15, 36, 37] on the same training set. Tab 3 presents the quantitative results on the EVRB dataset. The EVRB dataset includes extremely blurred videos captured during exposure times. As a result, frame-based video deblurring methods exhibit subpar performance. For instance, the best-performing network, ShiftNet [16], achieves only $30.56\mathrm{dB}$ , which is $0.42\mathrm{dB}$ lower than the best-performing + +Table 3: Quantitative results on the EVRB dataset. + +
MethodsEDVR [39]ESTRNN [53]ERDN [12]DSTNet [26]RVRT [18]
PSNRs29.0229.7928.3229.1530.24
SSIMs0.8860.9110.8930.8980.906
MethodsShift-Net [16]UEVD [15]EFNet [36]REFID [37]Ours
PSNRs30.5630.5530.9830.3331.38
SSIMs0.9220.9150.9270.9180.927
+ +![](images/34782f7695ccac1073018b3771aeca0fc4244793ee4b28ad2b3fc821b7801a3f.jpg) +Fig. 6: Visual comparison of deblurring results on the EVRB dataset. + +event-guided deblurring method, EFNet [36]. In contrast, our approach leverages video and event characteristics, effectively restoring even severe motion blur, which may be difficult to recover. Our method outperforms all approaches, achieving the best performance. We show our qualitative comparison with other methods in Fig. 6. + +# 5.4 Ablation Study + +We analyzed the performance contribution of the various modules in our frameworks. For a fair ablation study, we trained all the models for 600 epochs with 5 video frame inputs, conducting all experiments on the GoPro dataset. + +CRIFE module. To demonstrate the effectiveness of the CRIFE module, we replaced it with concatenation and convolution for comparison. When comparing the first column(Ver.1) and the second column(Ver.2) of Tab 4, we observed a performance gain of $+0.35$ dB with a small additional parameter $(+0.08\mathrm{MB})$ . Similarly, when comparing the performance of the third and fourth rows, we observed performance improvement. + +ECITFA module is the most crucial component of our model, performing feature alignment by leveraging information of the events. As in the Tab. 4, when comparing the first (Ver.1) with the third column (Ver.3) of the table, we observed a significant performance gain $(+1.69\mathrm{dB})$ upon the insertion of the ECITFA module. This trend is similarly maintained when the CRIFE module is incorporated. When comparing Ver.2 with Ver.4 of the Tab. 4, we observed a notable performance improvement $(+1.55\mathrm{dB})$ . + +Effectiveness of components in ECITFA module. In the Tab. 5, we demonstrated an effectiveness analysis for each component of ECITFA. Comparing the 2nd column with the last column labeled 'Ours', we observed a performance + +Table 4: Ablation study on the GoPro dataset. + +
MethodsVer.1Ver.2Ver.3Ver.4
CRIFE✓✓
ECITFA✓✓
PSNRs / Params34.65 / 4.57M35.00 / 4.65M36.34 / 9.59M36.55 / 9.68M
+ +Table 5: Effect of the component of ECITFA module on the GoPro dataset. DF denotes dynamic filter operation in the Eq.(12). + +
Methodsw/o Cascadedw/o DFw/o AttentionOurs
PSNRs36.0836.1836.2936.55
+ +Table 6: Comparison of CRIFE with various module variants. "Baseline" refers to Ver.3 of Tab.4. E and F represent event and frame features, respectively. + +
MethodsBaselinew/o concat (E+F)w/concat (E+F)
w/o recurrentw/ recurrent(Ours)
PSNRs36.3436.0435.9936.55
+ +gain of our method $(+0.37\mathrm{dB})$ when employing spatial pixel-wise dynamic filters, in contrast to not using DF. When utilizing cross-attention (Eq.13) to better leverage long-range pixel-dependencies in the alignment blocks, we observed a performance improvement $(+0.29\mathrm{dB})$ compared to aggregate feature using ResBlocks. Finally, we confirmed the effectiveness of leveraging non-local video frame information through a cascaded-based temporal feature alignment method for video deblurring. After using the proposed cascaded structure for temporal feature alignment, we can observe a performance gain $(+0.47\mathrm{dB})$ . + +Effectiveness of components in CRIFE module. To evaluate each component's effectiveness of the CRIFE module, we removed the concatenation between RGB and event features, directly matching their features. However, as shown in the 3rd column of Tab. 6, this approach yielded sub-optimal performance. Additionally, combining event features without a recurrent structure degraded performance (see 4th column). These ablations confirm that the recurrent structure effectively utilizes temporal information from events within exposure time. + +# 6 Conclusions + +This paper proposes a video deblurring framework, CMTA, that elaborately considers the characteristics of an event and video. Specifically, we achieve significant performance improvement through intra-frame feature enhancement and inter-frame temporal feature alignment. Furthermore, we construct a real-world deblurring dataset, the EVRB dataset, which will be valuable for evaluating event-guided deblurring methods. Finally, CMTA demonstrates state-of-the-art performance across various deblurring datasets. + +Acknowledgements. This work was supported by the Technology Innovation Program (1415187329,20024355, Development of autonomous driving connec + +tivity technology based on sensor-infrastructure cooperation) funded By the Ministry of Trade, Industry & Energy(MOTIE, Korea) and the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (NRF2022R1A2B5B03002636). + +# References + +1. Aittala, M., Durand, F.: Burst image deblurring using permutation invariant convolutional neural networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 731-747 (2018) +2. Brandli, C., Berner, R., Yang, M., Liu, S.C., Delbruck, T.: A $240 \times 180$ 130 db $3\mu s$ latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits 49(10), 2333-2341 (2014) +3. Cao, M., Zhong, Z., Wang, J., Zheng, Y., Yang, Y.: Learning adaptive warping for real-world rolling shutter correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17785-17793 (2022) +4. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basic vsr: The search for essential components in video super-resolution and beyond. In: Proceedings of the IEEE conference on computer vision and pattern recognition (2021) +5. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment (2021) +6. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: European Conference on Computer Vision. pp. 17-33. Springer (2022) +7. Chen, S., Zhang, J., Zheng, Y., Huang, T., Yu, Z.: Enhancing motion deblurring in high-speed scenes with spike streams. Advances in Neural Information Processing Systems 36 (2024) +8. Cho, H., Jeong, Y., Kim, T., Yoon, K.J.: Non-coaxial event-guided motion deblurring with spatial alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12492-12503 (2023) +9. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence 44(1), 154-180 (2020) +0. Gehrig, D., Loquercio, A., Derpanis, K.G., Scaramuzza, D.: End-to-end learning of representations for asynchronous event-based data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5633-5643 (2019) +1. Hyun Kim, T., Mu Lee, K., Scholkopf, B., Hirsch, M.: Online video deblurring via dynamic temporal blending network. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 4038-4047 (2017) +2. Jiang, B., Xie, Z., Xia, Z., Li, S., Liu, S.: Erdn: Equivalent receptive field deformable network for video deblurring. In: European Conference on Computer Vision. pp. 663-678. Springer (2022) +3. Jiang, Z., Zhang, Y., Zou, D., Ren, J., Lv, J., Liu, Y.: Learning event-based motion deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3320-3329 (2020) +4. Kim, T.H., Sajjadi, M.S., Hirsch, M., Scholkopf, B.: Spatio-temporal transformer network for video restoration. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 106-122 (2018) + +15. Kim, T., Lee, J., Wang, L., Yoon, K.J.: Event-guided deblurring of unknown exposure time videos. In: European Conference on Computer Vision. pp. 519-538. Springer (2022) +16. Li, D., Shi, X., Zhang, Y., Cheung, K.C., See, S., Wang, X., Qin, H., Li, H.: A simple baseline for video restoration with grouped spatial-temporal shift. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9822-9832 (June 2023) +17. Liang, J., Cao, J., Fan, Y., Zhang, K., Ranjan, R., Li, Y., Timofte, R., Van Gool, L.: Vrt: A video restoration transformer. arXiv preprint arXiv:2201.12288 (2022) +18. Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems 35, 378-393 (2022) +19. Lin, J., Cai, Y., Hu, X., Wang, H., Yan, Y., Zou, X., Ding, H., Zhang, Y., Timofte, R., Van Gool, L.: Flow-guided sparse transformer for video deblurring. arXiv preprint arXiv:2201.01893 (2022) +20. Lin, S., Zhang, J., Pan, J., Jiang, Z., Zou, D., Wang, Y., Chen, J., Ren, J.S.: Learning event-driven video deblurring and interpolation. In: ECCV (8). pp. 695-710 (2020) +21. Mildenhall, B., Barron, J.T., Chen, J., Sharlet, D., Ng, R., Carroll, R.: Burst denoising with kernel prediction networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2502-2510 (2018) +22. Nah, S., Baik, S., Hong, S., Moon, G., Son, S., Timofte, R., Mu Lee, K.: Ntire 2019 challenge on video deblurring and super-resolution: Dataset and study. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019) +23. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3883-3891 (2017) +24. Oh, J., Kim, M.: Demfi: deep joint deblurring and multi-frame interpolation with flow-guided attentive correlation and recursive boosting. In: European Conference on Computer Vision. pp. 198-215. Springer (2022) +25. Pan, J., Bai, H., Tang, J.: Cascaded deep video deblurring using temporal sharpness prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3043-3051 (2020) +26. Pan, J., Xu, B., Dong, J., Ge, J., Tang, J.: Deep discriminative spatial and temporal network for efficient video deblurring. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (Feb 2023) +27. Pan, L., Scheerlinck, C., Yu, X., Hartley, R., Liu, M., Dai, Y.: Bringing a blurry frame alive at high frame-rate with an event camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6820-6829 (2019) +28. Ranjan, A., Black, M.J.: Optical flow estimation using a spatial pyramid network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4161-4170 (2017) +29. Rebecq, H., Gehrig, D., Scaramuzza, D.: Esim: an open event camera simulator. In: Conference on Robot Learning. pp. 969-982. PMLR (2018) +30. Rim, J., Kim, G., Kim, J., Lee, J., Lee, S., Cho, S.: Realistic blur synthesis for learning image deblurring. In: European conference on computer vision. pp. 487-503. Springer (2022) + +31. Rim, J., Lee, H., Won, J., Cho, S.: Real-world blur dataset for learning and benchmarking deblurring algorithms. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16. pp. 184-201. Springer (2020) +32. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) +33. Shang, W., Ren, D., Zou, D., Ren, J.S., Luo, P., Zuo, W.: Bringing events into video deblurring with non-consecutively blurry frames. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4531-4540 (October 2021) +34. Shen, W., Bao, W., Zhai, G., Chen, L., Min, X., Gao, Z.: Video frame interpolation and enhancement via pyramid recurrent framework. IEEE Transactions on Image Processing 30, 277-292 (2020) +35. Su, S., Delbracio, M., Wang, J., Sapiro, G., Heidrich, W., Wang, O.: Deep video deblurring for hand-held cameras. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1279-1288 (2017) +36. Sun, L., Sakaridis, C., Liang, J., Jiang, Q., Yang, K., Sun, P., Ye, Y., Wang, K., Van Gool, L.: Event-based fusion for motion deblurring with cross-modal attention. In: European Conference on Computer Vision (ECCV) (2022) +37. Sun, L., Sakaridis, C., Liang, J., Sun, P., Cao, J., Zhang, K., Jiang, Q., Wang, K., Van Gool, L.: Event-based frame interpolation with ad-hoc deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18043-18052 (2023) +38. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NIPS. pp. 5998-6008 (2017) +39. Wang, X., Chan, K.C., Yu, K., Dong, C., Change Loy, C.: Edvr: Video restoration with enhanced deformable convolutional networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (June 2019) +40. Wang, Y., Lu, Y., Gao, Y., Wang, L., Zhong, Z., Zheng, Y., Yamashita, A.: Efficient video deblurring guided by motion magnitude. In: Proceedings of the European Conference on Computer Vision (ECCV) (2022) +41. Wei, K., Fu, Y., Yang, J., Huang, H.: A physics-based noise formation model for extreme low-light raw denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2758-2767 (2020) +42. Wieschollek, P., Hirsch, M., Scholkopf, B., Lensch, H.: Learning blind motion deblurring. In: ICCV. pp. 231-240 (2017) +43. Xu, F., Yu, L., Wang, B., Yang, W., Xia, G.S., Jia, X., Qiao, Z., Liu, J.: Motion deblurring with real events. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2583-2592 (October 2021) +44. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5728-5739 (2022) +45. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. arXiv preprint arXiv:2102.02808 (2021) +46. Zhang, H., Xie, H., Yao, H.: Spatio-temporal deformable attention network for video deblurring. In: ECCV (2022) + +47. Zhang, K., Luo, W., Zhong, Y., Ma, L., Liu, W., Li, H.: Adversarial spatio-temporal learning for video deblurring. IEEE Transactions on Image Processing (2018) +48. Zhang, L., Zhang, H., Zhu, C., Guo, S., Chen, J., Wang, L.: Fine-grained video deblurring with event camera. In: ICMM. pp. 352-364. Springer (2021) +49. Zhang, X., Yu, L.: Unifying motion deblurring and frame interpolation with events. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17765-17774 (2022) +50. Zhang, X., Yu, L., Yang, W., Liu, J., Xia, G.S.: Generalizing event-based motion deblurring in real-world scenarios. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10734-10744 (2023) +51. Zhang, Y., Qin, H., Wang, X., Li, H.: Rethinking noise synthesis and modeling in raw denoising. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4593-4601 (2021) +52. Zhong, Z., Cao, M., Ji, X., Zheng, Y., Sato, I.: Blur interpolation transformer for real-world motion from blur. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5713-5723 (2023) +53. Zhong, Z., Gao, Y., Zheng, Y., Zheng, B.: Efficient spatio-temporal recurrent neural network for video deblurring. In: European Conference on Computer Vision. pp. 191-207. Springer (2020) +54. Zhong, Z., Zheng, Y., Sato, I.: Towards rolling shutter correction and deblurring in dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9219-9228 (2021) +55. Zhou, J., Jampani, V., Pi, Z., Liu, Q., Yang, M.H.: Decoupled dynamic filter networks. In: IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR) (Jun 2021) +56. Zhu, A.Z., Yuan, L., Chaney, K., Daniilidis, K.: Unsupervised event-based learning of optical flow, depth, and egomotion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 989-997 (2019) +57. Zhu, C., Dong, H., Pan, J., Liang, B., Huang, Y., Fu, L., Wang, F.: Deep recurrent neural network with multi-scale bi-directional propagation for video deblurring. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 3598-3607 (2022) +58. Zhu, X., Hu, H., Lin, S., Dai, J.: Deformable convnets v2: More deformable, better results. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9308-9316 (2019) \ No newline at end of file diff --git a/2024/CMTA_ Cross-Modal Temporal Alignment for Event-guided Video Deblurring/images.zip b/2024/CMTA_ Cross-Modal Temporal Alignment for Event-guided Video Deblurring/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..56c661abf5b86443c455a67100cf1036cc5e33cd --- /dev/null +++ b/2024/CMTA_ Cross-Modal Temporal Alignment for Event-guided Video Deblurring/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7668fcd7366c52a73a4ffb797235d95792208fa8445756d4669f22184f055288 +size 481020 diff --git a/2024/CMTA_ Cross-Modal Temporal Alignment for Event-guided Video Deblurring/layout.json b/2024/CMTA_ Cross-Modal Temporal Alignment for Event-guided Video Deblurring/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ff11752252cf01f5f224293b18d0b530cfe185cf --- /dev/null +++ b/2024/CMTA_ Cross-Modal Temporal Alignment for Event-guided Video Deblurring/layout.json @@ -0,0 +1,11387 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 111, + 465, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 111, + 465, + 147 + ], + "spans": [ + { + "bbox": [ + 148, + 111, + 465, + 147 + ], + "type": "text", + "content": "CMTA: Cross-Modal Temporal Alignment for Event-guided Video Deblurring" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 188, + 168, + 426, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 168, + 426, + 180 + ], + "spans": [ + { + "bbox": [ + 188, + 168, + 426, + 180 + ], + "type": "text", + "content": "Taewoo Kim" + }, + { + "bbox": [ + 188, + 168, + 426, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 188, + 168, + 426, + 180 + ], + "type": "text", + "content": ", Hoonhee Cho" + }, + { + "bbox": [ + 188, + 168, + 426, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 188, + 168, + 426, + 180 + ], + "type": "text", + "content": ", and Kuk-Jin Yoon" + }, + { + "bbox": [ + 188, + 168, + 426, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 200, + 190, + 414, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 190, + 414, + 213 + ], + "spans": [ + { + "bbox": [ + 200, + 190, + 414, + 213 + ], + "type": "text", + "content": "Korea Advanced Institute of Science and Technology {intelpro, gngsgsm1, kjyoon}@kaist.ac.kr" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 161, + 242, + 452, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 242, + 452, + 495 + ], + "spans": [ + { + "bbox": [ + 161, + 242, + 452, + 495 + ], + "type": "text", + "content": "Abstract. Video deblurring aims to enhance the quality of restored results in motion-blurred videos by effectively gathering information from adjacent video frames to compensate for the insufficient data in a single blurred frame. However, when faced with consecutively severe motion blur situations, frame-based video deblurring methods often fail to find accurate temporal correspondence among neighboring video frames, leading to diminished performance. To address this limitation, we aim to solve the video deblurring task by leveraging an event camera with microsecond temporal resolution. To fully exploit the dense temporal resolution of the event camera, we propose two modules: 1) Intra-frame feature enhancement operates within the exposure time of a single blurred frame, iteratively enhancing cross-modality features in a recurrent manner to better utilize the rich temporal information of events, 2) Interframe temporal feature alignment gathers valuable long-range temporal information to target frames, aggregating sharp features leveraging the advantages of the events. In addition, we present a novel dataset composed of real-world blurred RGB videos, corresponding sharp videos, and event data. This dataset serves as a valuable resource for evaluating event-guided deblurring methods. We demonstrate that our proposed methods outperform state-of-the-art frame-based and event-based motion deblurring methods through extensive experiments conducted on both synthetic and real-world deblurring datasets. The code and dataset are available at https://github.com/intelpro/CMTA." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 161, + 505, + 443, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 505, + 443, + 517 + ], + "spans": [ + { + "bbox": [ + 161, + 505, + 443, + 517 + ], + "type": "text", + "content": "Keywords: Temporal alignment " + }, + { + "bbox": [ + 161, + 505, + 443, + 517 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 161, + 505, + 443, + 517 + ], + "type": "text", + "content": " Video Deblurring " + }, + { + "bbox": [ + 161, + 505, + 443, + 517 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 161, + 505, + 443, + 517 + ], + "type": "text", + "content": " Event cameras" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 537, + 229, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 537, + 229, + 550 + ], + "spans": [ + { + "bbox": [ + 133, + 537, + 229, + 550 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 563, + 481, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 563, + 481, + 646 + ], + "spans": [ + { + "bbox": [ + 133, + 563, + 481, + 646 + ], + "type": "text", + "content": "Motion blur is a common artifact caused by dynamic movements within a scene or camera motion during exposure. Motion deblurring, which aims to reverse the blurring process, presents significant challenges due to variations in blur intensity influenced by scene structure and depth. To achieve high-quality deblurring, video deblurring has emerged, leveraging information from neighboring frames instead of relying solely on a single blurred image. However, identifying temporal correspondence between blurred video frames becomes challenging with" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 654, + 223, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 654, + 223, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 654, + 223, + 665 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "extreme motion blur, hindering the extraction of valuable information from adjacent frames and impeding performance improvement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 481, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 481, + 226 + ], + "type": "text", + "content": "Event cameras [9], with their extremely low latency (on the order of microseconds), can offer high-quality guidance for motion deblurring due to their ability to capture high-temporal resolution of brightness change. To effectively utilize the advantages of the events, several event-guided motion deblurring works [15, 36, 50] have been introduced. While these works have typically explored cross-modal feature fusion methods across different modalities, there has been limited work on leveraging the abundant temporal information in videos." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 228, + 482, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 335 + ], + "type": "text", + "content": "To obtain high-quality results, we emphasize the event camera's temporal continuity, focusing on its interaction with video frames that exhibit long-term temporal dependencies. Unlike previous event-guided motion deblurring works [36,43,50], which relied on a single blurry image and corresponding events on its exposure time, we further design precise temporal feature alignment methods between neighboring video frames by leveraging the advantages of event data. Specifically, we propose novel modules from two perspectives: intra-frame (interaction between events and frames within the exposure time) and inter-frame (interaction between different frames) perspectives." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 338, + 482, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 410 + ], + "type": "text", + "content": "From an intra-frame perspective within exposure time, we propose a Cross-modal Recurrent Intra-frame Feature Enhancement (CRIFE) module to better leverage the rich temporal information of the events by mutually interacting the blur frame and event features within the duration of exposure time. In this module, we perform recurrent attention-based feature enhancement using a transformer [38] that better captures long-range pixel dependencies." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 412, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 639 + ], + "type": "text", + "content": "In temporal feature alignment with the second perspective, we propose a novel event-guided temporal feature alignment module, effectively leveraging rich temporal characteristics of the events. Conventional frame-based temporal feature alignment methods rely on optical flow [28] or deformable convolution [5,58]. While optical flow and deformable convolutions aid in achieving temporal feature alignment, the high computational complexity of these operations makes it challenging to execute them at a higher spatial scale. Therefore, video frame alignment, generally conducted at lower spatial scales, leads to sub-optimal deblurring results due to the lack of spatial contexts of features. To overcome these limitations, our temporal feature alignment module avoids relying on optical flow or deformable convolution by effectively leveraging the temporally dense advantages of the events. Therefore, we can effectively perform temporal feature alignment across multiple visual scales as we do not rely on these complex operations but rather efficiently leverage the temporal information from the events. As a result, our temporal feature alignment module demonstrates a significant performance improvement in event-guided video deblurring tasks. Since we have introduced a pioneering method for aligning temporal features using the advantages of the events, it is expected to be effectively applicable to various event-guided video restoration tasks (e.g., event-guided video super-resolution)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Finally, we propose a novel video deblurring dataset, the EVRB dataset, composed of high-quality RGB and event data. It consists of real-world blurry" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 116, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 116, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 480, + 163 + ], + "type": "text", + "content": "videos generated by extreme motion and corresponding sharp videos for generalized applications in real-world scenarios. The network trained on the EVRB dataset can be directly applicable to real-world scenarios, making it a valuable resource for event-guided deblurring research." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 182, + 241, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 182, + 241, + 194 + ], + "spans": [ + { + "bbox": [ + 134, + 182, + 241, + 194 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 207, + 247, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 207, + 247, + 219 + ], + "spans": [ + { + "bbox": [ + 134, + 207, + 247, + 219 + ], + "type": "text", + "content": "2.1 Video Deblurring" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 227, + 480, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 227, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 134, + 227, + 480, + 335 + ], + "type": "text", + "content": "Early works [1, 35] employed CNNs that take the concatenation of adjacent frames as input to address video deblurring. Subsequently, to better leverage temporal information, approaches have emerged that employ 3D convolutions [26, 47], temporal alignment modules with deformable convolutions [39] and optical flow [14, 25, 40, 46]. As an alternative video alignment method, some works utilize RNN [11, 42, 53] and transformer [18, 19] structures to propagate information from long-range video frames. However, as the intensity of motion blur in the video increases, frame-based video alignment methods struggle to achieve accurate video alignment, leading to sub-optimal deblurring results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 353, + 325, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 353, + 325, + 365 + ], + "spans": [ + { + "bbox": [ + 134, + 353, + 325, + 365 + ], + "type": "text", + "content": "2.2 Event-guided Motion Deblurring" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 373, + 480, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 373, + 480, + 540 + ], + "spans": [ + { + "bbox": [ + 134, + 373, + 480, + 540 + ], + "type": "text", + "content": "An event camera can effectively be used for motion deblurring as it records motion information corresponding to the brightness differences with high temporal resolution. Efforts to utilize event cameras for motion deblurring [8, 13, 15, 20, 27, 36, 37, 43, 48-50] have been actively ongoing. Recent studies have focused on effectively fusing event and image features of different modalities. To this end, Sun et al. [36] employed a transformer architecture for feature fusion. Zhang et al. [50] effectively combined modalities using a multi-scale architecture. In addition, there are also attempts to address motion deblurring by assuming challenging and general scenarios [8, 15, 33]. However, these studies have primarily focused on single image deblurring and do not exploit the long-range temporal information demonstrated in previous video deblurring tasks. To effectively acquire information that may be missing from sparse events, we introduce a novel method for accurate temporal alignment with events, utilizing information from surrounding adjacent frames in video deblurring." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 558, + 472, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 558, + 472, + 585 + ], + "spans": [ + { + "bbox": [ + 134, + 558, + 472, + 585 + ], + "type": "text", + "content": "3 Event-based Video Deblurring Dataset for Real-world Blur" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 597, + 345, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 597, + 345, + 609 + ], + "spans": [ + { + "bbox": [ + 134, + 597, + 345, + 609 + ], + "type": "text", + "content": "3.1 Limitation of Synthetic Blur Dataset" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 617, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 617, + 480, + 665 + ], + "type": "text", + "content": "Typically, the blurred images synthesis procedure adheres to the methodology outlined in [22-24, 34, 35], involving the averaging of consecutive video frames within a fixed-size window. However, as recent studies [31, 52, 54] have discussed, blur synthesis based on discrete signals may result in shutter artifacts even when" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 103 + ], + "type": "text", + "content": "Cross-Modal Temporal Alignment for Event-guided Video Deblurring" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 233, + 231 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 233, + 231 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 233, + 231 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 233, + 231 + ], + "type": "image", + "image_path": "2aa45a1d09c9f77a410f2e52ea77c3c95b8fe3b80ad3c1d1f948fb53b0bd39b5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 241, + 482, + 297 + ], + "lines": [ + { + "bbox": [ + 130, + 241, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 482, + 297 + ], + "type": "text", + "content": "Fig. 1: Illustration of a hybrid camera system for real-world event-based video deblurring dataset. " + }, + { + "bbox": [ + 130, + 241, + 482, + 297 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 241, + 482, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 241, + 482, + 297 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 241, + 482, + 297 + ], + "type": "text", + "content": " denote the cameras for acquiring sharp and blur videos, respectively. (a): The triple-axis camera system to capture real-world blur. (b): A diagram of our hybrid camera system. (c): Samples from our EVRB dataset with natural blur." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 238, + 114, + 347, + 232 + ], + "blocks": [ + { + "bbox": [ + 238, + 114, + 347, + 232 + ], + "lines": [ + { + "bbox": [ + 238, + 114, + 347, + 232 + ], + "spans": [ + { + "bbox": [ + 238, + 114, + 347, + 232 + ], + "type": "image", + "image_path": "c7c93452eb3c45e081aef27acf4743df0783247ecf107f0b49cbb9922989e105.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 348, + 114, + 472, + 233 + ], + "blocks": [ + { + "bbox": [ + 348, + 114, + 472, + 233 + ], + "lines": [ + { + "bbox": [ + 348, + 114, + 472, + 233 + ], + "spans": [ + { + "bbox": [ + 348, + 114, + 472, + 233 + ], + "type": "image", + "image_path": "1876ae1d4afcaad698acabc3cd91eebeb08d9d35ee932ea9a4bab4e377375eda.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 547 + ], + "type": "text", + "content": "averaging high-frame-rate videos. Furthermore, simply averaging sharp images disregards essential elements in the blurred images such as pixel saturation [30], limitations due to dynamic range, and physical noise [41, 51] in data acquired during exposure time. Furthermore, in the case of the event-guided motion deblurring research, existing works have enhanced the synthetic aspects by using event simulator [29]. To address these limitations, recent works have attempted to acquire real-world blur datasets using hybrid camera systems [3, 31, 52-54]. Meanwhile, to generalize event-based motion deblurring, Sun et al. [36] introduced the ReBlur dataset acquired in a high-precision optical laboratory using an electronic-controlled slide-rail system. However, despite these successes, the ReBlur dataset has limitations due to its indoor setting, a lack of dynamic objects, and the absence of high-quality RGB data as it is acquired using the DAVIS sensor [2]. Additionally, the sequences are relatively short(minimum six blurred frames in the sequence), making it challenging for use in event-guided video deblurring. We have acquired a new EVRB dataset for evaluating event-guided image and video deblurring methods in real-world blurry videos. The EVRB dataset was captured using a customized hybrid system consisting of two RGB cameras and one event camera. This system encompasses a range of blur magnitude from slight to extreme in various urban environments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 563, + 336, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 563, + 336, + 575 + ], + "spans": [ + { + "bbox": [ + 131, + 563, + 336, + 575 + ], + "type": "text", + "content": "3.2 Triple-axis Hybrid Camera System" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "As shown in Fig. 1, we design a hybrid camera system to enable the acquisition of different data sources to be geometrically aligned. Two RGB cameras and one event camera are geometrically aligned using two " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "50/50" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " cube beam splitters, resulting in a minimal baseline. We perform pixel-wise alignment between multiple cameras based on the homography calculated by extrinsic calibration for precise alignment. For the synchronization of the multiple cameras, we use an external trigger system. Each camera receives the falling and rising edges of the trigger" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 270 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 270 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 270 + ], + "type": "image", + "image_path": "bb7afde7f4dc31ede6443dc3b7f34963307980eabe143e38a3326855983ba012.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 277, + 482, + 333 + ], + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 333 + ], + "type": "text", + "content": "Fig. 2: Overall framework of CMTA is divided into two main components: Cross-modal Recurrent Intra-frame Feature Enhancement (CRIFE) and Event-guided Cascaded Inter-frame Temporal Feature Alignment (ECITFA). " + }, + { + "bbox": [ + 130, + 277, + 482, + 333 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 277, + 482, + 333 + ], + "type": "text", + "content": " is the scale factor for multi-scale features. In the figure of the ECITFA module, the description was performed for the case of " + }, + { + "bbox": [ + 130, + 277, + 482, + 333 + ], + "type": "inline_equation", + "content": "P = 2" + }, + { + "bbox": [ + 130, + 277, + 482, + 333 + ], + "type": "text", + "content": " for simplicity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 437 + ], + "type": "text", + "content": "signal and acquires data synchronized with the period of the signal. This external trigger can precisely control the exposure times of the two RGB cameras, enabling us to capture paired sharp and blurred video frames. Furthermore, for photometric alignment, we physically adjust the amount of incoming light for both cameras to equalize the total irradiance of the multiple cameras using a neutral density filter." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 454, + 202, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 454, + 202, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 454, + 202, + 467 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 479, + 209, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 209, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 209, + 490 + ], + "type": "text", + "content": "4.1 Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": "The overview of the proposed framework is illustrated in Fig. 2. Given consecutive blurred video frames " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{B_k\\}" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": " and sets of event streams corresponding to the exposure time of each video frame " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{\\mathbb{E}_k\\}" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "k \\in \\{t - P, \\dots, t, \\dots, t + P\\}" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": ", our goal is to estimate the latent sharp video frame " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "S_t" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": ". To utilize the event stream " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{\\mathbb{E}_k\\}" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": " corresponding to the exposure time of " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{B_k\\}" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": " as the input for the networks, we first perform embedding using the event voxel grid representation [56] for the event stream " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{\\mathbb{E}_k\\}" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{E_k\\}" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": ". Our framework consists of two main submodules: (1) Cross-modal Recurrent Intra-frame Feature Enhancement (CRIFE) module and (2) Event-guided Cascaded Inter-frame Temporal Feature Alignment (ECITFA) module. In the first module, we perform cross-modal feature enhancement through recurrent interactions between blurred frame features and event features to leverage the continuous temporal information of events within the exposure time. After obtaining the fused feature " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{G}_k\\}" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": " from the first module, we generate multi-scale features, " + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}_k^s\\}_{s=0}^2" + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": ", through a weight-shared pyra" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 103 + ], + "type": "text", + "content": "Cross-Modal Temporal Alignment for Event-guided Video Deblurring" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 115, + 449, + 262 + ], + "blocks": [ + { + "bbox": [ + 167, + 115, + 449, + 262 + ], + "lines": [ + { + "bbox": [ + 167, + 115, + 449, + 262 + ], + "spans": [ + { + "bbox": [ + 167, + 115, + 449, + 262 + ], + "type": "image", + "image_path": "e417d8010bde286557f1bbfc66ea431af1ce7ad9ff133596eac8865891e6b4b2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 273, + 482, + 296 + ], + "lines": [ + { + "bbox": [ + 131, + 273, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 131, + 273, + 482, + 296 + ], + "type": "text", + "content": "Fig. 3: Illustration of Cross-modal Recurrent Intra-frame Feature Enhancement (CRIFE)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 328, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 328, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 131, + 328, + 482, + 437 + ], + "type": "text", + "content": "mid encoder. Subsequently, for the temporal feature alignment stage, we encode multi-scale event features, " + }, + { + "bbox": [ + 131, + 328, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\{\\varepsilon_{[k,k + 1]}^s\\}_{s = 0}^2" + }, + { + "bbox": [ + 131, + 328, + 482, + 437 + ], + "type": "text", + "content": ", by grouping two consecutive events corresponding to the exposure times of each frame to connect adjacent frames. By utilizing the multi-scale pyramid features " + }, + { + "bbox": [ + 131, + 328, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}_k^s\\}_{s = 0}^2" + }, + { + "bbox": [ + 131, + 328, + 482, + 437 + ], + "type": "text", + "content": " and the encoded event feature " + }, + { + "bbox": [ + 131, + 328, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\{\\varepsilon_{[k,k + 1]}^s\\}_{s = 0}^2" + }, + { + "bbox": [ + 131, + 328, + 482, + 437 + ], + "type": "text", + "content": ", we perform temporal feature alignment. Afterward, temporally aligned feature pyramids are fed into the U-Net [32]-based decoder, generating the final sharp video frame " + }, + { + "bbox": [ + 131, + 328, + 482, + 437 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 131, + 328, + 482, + 437 + ], + "type": "text", + "content": ". Our framework can be extended to cases where " + }, + { + "bbox": [ + 131, + 328, + 482, + 437 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 131, + 328, + 482, + 437 + ], + "type": "text", + "content": " is an arbitrary positive number. However, for the sake of brevity, we will explain it in the main text with " + }, + { + "bbox": [ + 131, + 328, + 482, + 437 + ], + "type": "inline_equation", + "content": "P = 2" + }, + { + "bbox": [ + 131, + 328, + 482, + 437 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 455, + 453, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 455, + 453, + 466 + ], + "spans": [ + { + "bbox": [ + 131, + 455, + 453, + 466 + ], + "type": "text", + "content": "4.2 Cross-modal Recurrent Intra-frame Feature Enhancement" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 592 + ], + "type": "text", + "content": "Since event cameras provide rich temporal information on brightness changes, it is crucial to effectively utilize this dense temporal information of the events within the duration of the exposure time. Typically, event data captured during the exposure time is transformed into event embeddings [10,56] as input to the network, followed by feature extraction using 2D CNNs. However, these approaches cannot effectively utilize the continuous temporal information of events, and these limitations can impact the performance of event-guided video deblurring. To address these limitations, we propose a method that leverages the rich temporal nature of events and fuses event and blurred frame features using recurrent-based attention methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": "Recently, researchers have demonstrated the effectiveness of transformer-based architectures [38] by capturing long-range pixel dependencies, which have been proven highly effective in various vision tasks. We leverage transformers' advantages to more effectively utilize the temporal benefits of the events through a recurrent-based approach. For each blur video frame index " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "k \\in \\{t - 2, \\dots, t + 2\\}" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": ", we initially partition the exposure time " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "T_{exp,k}" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": " unit temporal intervals," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "denoted as " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\Delta t = T_{exp,k} / N" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ". Based on the time interval " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ", we divide the event voxel grid within the exposure time, denoted as " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "E_k \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " temporally divided event voxel grids " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\{E_k^n\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "E_k^n \\in \\mathbb{R}^{C / N \\times H \\times W}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "n \\in \\{0, \\dots, N - 1\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ". After, we extract event features " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}(E)_k^n\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " of the temporally divided event voxel grids " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "E_k^n" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " using weight-sharing event feature extractor. To apply cross-attention, we construct a query encompassing global information of blur and temporally divided event features. That is, we concatenate the blur frame feature " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(B)_k" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " with the temporally segmented event sets " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}(E)_k^n\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " and then extract the feature for the encoding query information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 240, + 245, + 481, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 245, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 240, + 245, + 481, + 258 + ], + "type": "interline_equation", + "content": "\\mathcal {Q} _ {k} = \\operatorname {C o n v} _ {\\mathrm {p}} (\\mathcal {F} (B) _ {k} \\| \\left\\{\\mathcal {F} (E) _ {k} ^ {n} \\right\\}) \\tag {1}", + "image_path": "42a607af423edcb06dc3d04b628269afa3c9a60ce68ab1cd5afede89281a9057.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 272, + 260, + 480, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 260, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 272, + 260, + 480, + 274 + ], + "type": "interline_equation", + "content": "Q _ {k} ^ {n = 0} = F _ {R} \\left(\\mathcal {Q} _ {k}\\right) \\tag {2}", + "image_path": "d62ab70fbf73545004502f2229bd8f3f9057efa862f9e2884756e6124a373dab.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 281, + 482, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 327 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 281, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}_{\\mathrm{p}}" + }, + { + "bbox": [ + 130, + 281, + 482, + 327 + ], + "type": "text", + "content": " represents point-wise convolution, " + }, + { + "bbox": [ + 130, + 281, + 482, + 327 + ], + "type": "inline_equation", + "content": "F_{R}" + }, + { + "bbox": [ + 130, + 281, + 482, + 327 + ], + "type": "text", + "content": " consists of a sequence of blocks with " + }, + { + "bbox": [ + 130, + 281, + 482, + 327 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 130, + 281, + 482, + 327 + ], + "type": "text", + "content": " convolution and a stride of 2, along with ResBlocks. Additionally, " + }, + { + "bbox": [ + 130, + 281, + 482, + 327 + ], + "type": "inline_equation", + "content": "Q_{k}^{n = 0}" + }, + { + "bbox": [ + 130, + 281, + 482, + 327 + ], + "type": "text", + "content": " denotes the initial query feature for the recurrent-based attention method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 329, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 402 + ], + "type": "text", + "content": "As illustrated in Fig. 3, we perform cross-attention to update query information iteratively. To find the key and value of cross-attention for query updating, we recursively input temporally separated event features, helping to better utilize the events' rich temporal information. That is, for iteration " + }, + { + "bbox": [ + 130, + 329, + 482, + 402 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 329, + 482, + 402 + ], + "type": "text", + "content": ", we extract features to be used as keys and values by concatenating the previously updated " + }, + { + "bbox": [ + 130, + 329, + 482, + 402 + ], + "type": "inline_equation", + "content": "Q_{k}^{n}" + }, + { + "bbox": [ + 130, + 329, + 482, + 402 + ], + "type": "text", + "content": " with temporally divided " + }, + { + "bbox": [ + 130, + 329, + 482, + 402 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 329, + 482, + 402 + ], + "type": "text", + "content": "-th event features " + }, + { + "bbox": [ + 130, + 329, + 482, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(E)_k^n" + }, + { + "bbox": [ + 130, + 329, + 482, + 402 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 252, + 410, + 480, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 410, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 252, + 410, + 480, + 422 + ], + "type": "interline_equation", + "content": "\\mathrm {K V} _ {k} ^ {n} = F _ {K V} \\left(Q _ {k} ^ {n} \\| \\mathcal {F} (E) _ {k} ^ {n}\\right) \\tag {3}", + "image_path": "387a614b1b607e039d96dc7b2a4e1aa9b2d9bef5827a7f33fbc4d9f61a969062.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 430, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 430, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 131, + 430, + 482, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 430, + 482, + 479 + ], + "type": "inline_equation", + "content": "\\parallel" + }, + { + "bbox": [ + 131, + 430, + 482, + 479 + ], + "type": "text", + "content": " channel-wise concatenation and " + }, + { + "bbox": [ + 131, + 430, + 482, + 479 + ], + "type": "inline_equation", + "content": "\\mathrm{KV}_k^n" + }, + { + "bbox": [ + 131, + 430, + 482, + 479 + ], + "type": "text", + "content": " denotes the output features for key and value projection, and " + }, + { + "bbox": [ + 131, + 430, + 482, + 479 + ], + "type": "inline_equation", + "content": "F_{KV}" + }, + { + "bbox": [ + 131, + 430, + 482, + 479 + ], + "type": "text", + "content": " refers to the ResBlocks layer. We then construct " + }, + { + "bbox": [ + 131, + 430, + 482, + 479 + ], + "type": "inline_equation", + "content": "K_k^n" + }, + { + "bbox": [ + 131, + 430, + 482, + 479 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 430, + 482, + 479 + ], + "type": "inline_equation", + "content": "V_k^n" + }, + { + "bbox": [ + 131, + 430, + 482, + 479 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 131, + 430, + 482, + 479 + ], + "type": "inline_equation", + "content": "K_k^n = W^K (\\mathrm{KV}_k^n)" + }, + { + "bbox": [ + 131, + 430, + 482, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 430, + 482, + 479 + ], + "type": "inline_equation", + "content": "V_{k}^{n} = W^{V}(\\mathrm{KV}_{k}^{n})" + }, + { + "bbox": [ + 131, + 430, + 482, + 479 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 131, + 430, + 482, + 479 + ], + "type": "inline_equation", + "content": "W^{(\\cdot)}" + }, + { + "bbox": [ + 131, + 430, + 482, + 479 + ], + "type": "text", + "content": " denote " + }, + { + "bbox": [ + 131, + 430, + 482, + 479 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 131, + 430, + 482, + 479 + ], + "type": "text", + "content": " convolution layer. Then, attention can be calculated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 237, + 486, + 481, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 486, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 237, + 486, + 481, + 510 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t n} _ {\\mathrm {k}} ^ {\\mathrm {n}} = \\operatorname {S o f t M a x} \\left(\\frac {Q _ {k} ^ {n} \\left(K _ {k} ^ {n}\\right) ^ {T}}{\\alpha}\\right) V _ {k} ^ {n} \\tag {4}", + "image_path": "1282be186210f6b2493d865700da5ba5ae226ecb81e5fa0316921393b2c16d0a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 516, + 482, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 482, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 482, + 613 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 516, + 482, + 613 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 516, + 482, + 613 + ], + "type": "text", + "content": " is learnable scaling parameter to balance attention weights and " + }, + { + "bbox": [ + 130, + 516, + 482, + 613 + ], + "type": "inline_equation", + "content": "\\mathrm{Att}\\mathfrak{n}_{\\mathrm{n}}" + }, + { + "bbox": [ + 130, + 516, + 482, + 613 + ], + "type": "text", + "content": " is outputs of cross-attention at iteration " + }, + { + "bbox": [ + 130, + 516, + 482, + 613 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 516, + 482, + 613 + ], + "type": "text", + "content": ". To calculate cross-covariance matrix efficiently, we adopted transposed attention [44] for efficient computations. This method enables the efficient computation of attention values at high resolutions by calculating the cross-covariance matrix along the channel axis, leading to a complexity of " + }, + { + "bbox": [ + 130, + 516, + 482, + 613 + ], + "type": "inline_equation", + "content": "O(C^2)" + }, + { + "bbox": [ + 130, + 516, + 482, + 613 + ], + "type": "text", + "content": ". Moreover, we reduced the number of channels when encoding step, enabling even more efficient operations. We use the attention value " + }, + { + "bbox": [ + 130, + 516, + 482, + 613 + ], + "type": "inline_equation", + "content": "\\mathrm{Att}\\mathfrak{n}_{\\mathrm{k}}^{\\mathrm{n}}" + }, + { + "bbox": [ + 130, + 516, + 482, + 613 + ], + "type": "text", + "content": " to update the query iteratively as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 230, + 620, + 481, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 620, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 230, + 620, + 481, + 634 + ], + "type": "interline_equation", + "content": "Q _ {k} ^ {n + 1} = Q _ {k} ^ {n} + \\operatorname {A t t n} _ {\\mathrm {k}} ^ {\\mathrm {n}} + \\operatorname {M L P} \\left(\\operatorname {A t t n} _ {\\mathrm {k}} ^ {\\mathrm {n}}\\right) \\tag {5}", + "image_path": "3ae03ee0459d7f3e1171262a1ec523b28ce49ce35af965681928a00b45b04f28.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 641, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 667 + ], + "type": "text", + "content": "where MLP denote multi-layer perceptron. After " + }, + { + "bbox": [ + 131, + 641, + 482, + 667 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 641, + 482, + 667 + ], + "type": "text", + "content": " iterations, we obtain the updated query feature, " + }, + { + "bbox": [ + 131, + 641, + 482, + 667 + ], + "type": "inline_equation", + "content": "Q_{k}^{N}" + }, + { + "bbox": [ + 131, + 641, + 482, + 667 + ], + "type": "text", + "content": ". The final query feature passes through up-sampling" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 164, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 164, + 91, + 448, + 103 + ], + "type": "text", + "content": "Cross-Modal Temporal Alignment for Event-guided Video Deblurring" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "text", + "content": "and skip connections. Then, we generate the final fused features " + }, + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{G}_k\\}" + }, + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_k = \\mathcal{Q}_k + \\mathrm{Dconv}_{4\\times 4}(Q_k^N)" + }, + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathrm{Dconv}_{4\\times 4}" + }, + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "text", + "content": " denote deconvolution layer with a kernel size of 4. As illustrated in Fig. 2, the fused features " + }, + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{G}_k\\}" + }, + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "text", + "content": " pass through a pyramid encoder, leading to the creation of multi-scale pyramid features, " + }, + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}_k^s\\}_{s=0}^2" + }, + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "text", + "content": ". Through the CRIFE module, we can leverage the advantages of abundant temporal information on the events within the exposure time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 202, + 436, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 202, + 436, + 227 + ], + "spans": [ + { + "bbox": [ + 132, + 202, + 436, + 227 + ], + "type": "text", + "content": "4.3 Event-guided Cascaded Inter-frame Temporal Feature Alignment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 232, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 232, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 482, + 422 + ], + "type": "text", + "content": "Temporal feature alignment aims to extract valuable information from adjacent video frames. Conventional frame-based video deblurring methods face challenges from inaccurate motion estimation due to motion blur. Conversely, event cameras, thanks to their resistance to motion blur, can offer valuable guidance for aligning video frames. The most straightforward way to align adjacent video frames using the events is to utilize event features to estimate optical flows or deformable offsets between neighboring frames [4,5,58]. While this approach can leverage the advantages of event data for offset and optical flow estimation, it is typically employed after spatial down-sampling due to the high computational costs associated with deformable convolutions [58] and optical flows [28]. Therefore, this approach could restrict performance as it inherently hinders the processing of features at high spatial resolutions in network architectures, limiting access to information across multiple visual scales. To address the aforementioned limitations, we propose a new temporal alignment module that combines the advantages of multiple-visual scale pyramids, leveraging the events' rich temporal contexts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": "As illustrated in Fig. 2, our proposed temporal feature alignment modules are structured as multi-level networks that gradually perform coarse-to-fine feature alignment. First, we extract event features, incorporating exposure time information from neighboring video frames to reference frames to facilitate video feature alignment. Specifically, we encode event features encompassing the exposure time interval between " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": ", allowing us to connect the frame at time " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": " with the frame at time " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": ". Through this event encoding step, we obtain event feature pyramid set " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\{\\varepsilon_{[m,m + 1]}^s\\}_{s = 0}^2" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "m \\in \\{t - 2,\\dots,t + 1\\}" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": ". We use the event feature pyramid containing motion information for adjacent times for the alignment of each blur frame feature pyramid " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}_k^s\\}_{s = 0}^2" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "inline_equation", + "content": "k \\in \\{t - 2,\\dots,t + 2\\}" + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "type": "text", + "content": "After the event encoding step for the feature alignment, we gradually perform temporal feature alignment from the bottom pyramid level (scale factor " + }, + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "type": "text", + "content": " of 2) to the top pyramid level (" + }, + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "type": "text", + "content": " of 0). Specifically, as shown in left side of Fig. 4, in each pyramid level " + }, + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "type": "inline_equation", + "content": "s \\in \\{0,1,2\\}" + }, + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "type": "text", + "content": "), We first upsample the hidden state of temporally aligned features at the previous scale, " + }, + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_i^{s+1}" + }, + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "type": "text", + "content": " through a deconvolution layer and receive them as inputs of alignment module, resulting in " + }, + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "type": "inline_equation", + "content": "h_i^s" + }, + { + "bbox": [ + 130, + 544, + 482, + 616 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 215, + 621, + 480, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 621, + 480, + 635 + ], + "spans": [ + { + "bbox": [ + 215, + 621, + 480, + 635 + ], + "type": "interline_equation", + "content": "h _ {i} ^ {s} = \\operatorname {D c o n v} _ {4 \\times 4} \\left(\\hat {\\mathcal {F}} _ {i} ^ {s + 1}\\right), i \\in \\{t - 1, t, t + 1 \\} \\tag {6}", + "image_path": "f9df798398c344adcffeb8a18381067a10605840f6223ad81c67845a379aaa37.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_i^{s + 1}" + }, + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "text", + "content": " denote aligned feature at previous scale " + }, + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "inline_equation", + "content": "s + 1" + }, + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{Dconv}_{4\\times 4}" + }, + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "text", + "content": " denote " + }, + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "inline_equation", + "content": "4\\times 4" + }, + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "text", + "content": " deconvolution layer. Note that there is no hidden state at the bottom pyramid" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 115, + 294, + 240 + ], + "blocks": [ + { + "bbox": [ + 140, + 115, + 294, + 240 + ], + "lines": [ + { + "bbox": [ + 140, + 115, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 140, + 115, + 294, + 240 + ], + "type": "image", + "image_path": "359efd6ea5ed3303e309257429825178ab57665602d237ce7dfecf55fa139eb2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 253, + 482, + 308 + ], + "lines": [ + { + "bbox": [ + 130, + 253, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 482, + 308 + ], + "type": "text", + "content": "Fig. 4: Overview of the Event-guided Cascaded Inter-frame Temporal Feature Alignment (ECITFA). The left figure illustrates temporal alignment for scale " + }, + { + "bbox": [ + 130, + 253, + 482, + 308 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 253, + 482, + 308 + ], + "type": "text", + "content": ". The key module for each alignment procedure, Cross-modal Temporal Feature Alignment (CTFA) at time " + }, + { + "bbox": [ + 130, + 253, + 482, + 308 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 253, + 482, + 308 + ], + "type": "text", + "content": ", is illustrated on the right of the figure. The CTFA module operates similarly for reference times " + }, + { + "bbox": [ + 130, + 253, + 482, + 308 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 130, + 253, + 482, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 253, + 482, + 308 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 130, + 253, + 482, + 308 + ], + "type": "text", + "content": " as well." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 296, + 114, + 473, + 245 + ], + "blocks": [ + { + "bbox": [ + 296, + 114, + 473, + 245 + ], + "lines": [ + { + "bbox": [ + 296, + 114, + 473, + 245 + ], + "spans": [ + { + "bbox": [ + 296, + 114, + 473, + 245 + ], + "type": "image", + "image_path": "20408f853ddba7bd47faf5758dcf8bdd8213f38f17c04faa81a84a5b9e0255b7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 342, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 480, + 365 + ], + "type": "text", + "content": "level (" + }, + { + "bbox": [ + 130, + 342, + 480, + 365 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 342, + 480, + 365 + ], + "type": "text", + "content": " of 2) since there are no features aligned at the previous scale, and for brevity, we focus on the case when " + }, + { + "bbox": [ + 130, + 342, + 480, + 365 + ], + "type": "inline_equation", + "content": "s < 2" + }, + { + "bbox": [ + 130, + 342, + 480, + 365 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "text", + "content": "As depicted in the left side of Fig. 4, we progressively group three blurred video frames when performing temporal feature alignment. In other words, we first align with respect to " + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "inline_equation", + "content": "\\{t - 2, t - 1, t\\}" + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "text", + "content": " and simultaneously align with respect to " + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "inline_equation", + "content": "\\{t, t + 1, t + 2\\}" + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "text", + "content": " features. Subsequently, we perform the final temporal feature alignment for the previously aligned " + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "text", + "content": " with the last target time " + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "text", + "content": ". More specifically, when performing frame alignment for the time step " + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "text", + "content": ", we first group the three blur video frame features " + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{t - 2}^s, \\mathcal{F}_{t - 1}^s, \\mathcal{F}_t^s" + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "text", + "content": " and perform alignment. In this case, we make use of two event features " + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "inline_equation", + "content": "\\varepsilon_{[t - 1,t]}^s" + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "inline_equation", + "content": "\\varepsilon_{[t,t + 1]}^s" + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "text", + "content": ", which respectively contain motion information between " + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "text", + "content": ", and between " + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 130, + 366, + 482, + 486 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 184, + 494, + 481, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 494, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 184, + 494, + 481, + 512 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {F}} _ {t - 1} ^ {s} = \\mathrm {C T F A} ^ {s} (\\mathcal {F} _ {t} ^ {s}, \\mathcal {F} _ {t - 1} ^ {s}, \\mathcal {F} _ {t - 2} ^ {s}, h _ {t - 1} ^ {s}, \\varepsilon_ {[ t - 2, t - 1 ]} ^ {s}, \\varepsilon_ {[ t - 1, t ]} ^ {s}), \\qquad (7)", + "image_path": "6541b655899d4767de8ad0d6d71521ae883f7317ad82e7407682494b9f16002d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 516, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 482, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 516, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\mathrm{CTFA}^s" + }, + { + "bbox": [ + 130, + 516, + 482, + 552 + ], + "type": "text", + "content": " denote unit Cross-modal Temporal Feature Alignment (CTFA) module at scale factor " + }, + { + "bbox": [ + 130, + 516, + 482, + 552 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 516, + 482, + 552 + ], + "type": "text", + "content": ". Across the same scale factor " + }, + { + "bbox": [ + 130, + 516, + 482, + 552 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 516, + 482, + 552 + ], + "type": "text", + "content": ", CTFA modules are weight-shared. Similarly, we perform alignment for the time step " + }, + { + "bbox": [ + 130, + 516, + 482, + 552 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 130, + 516, + 482, + 552 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 185, + 560, + 481, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 560, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 185, + 560, + 481, + 578 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {F}} _ {t + 1} ^ {s} = \\mathrm {C T F A} ^ {s} (\\mathcal {F} _ {t} ^ {s}, \\mathcal {F} _ {t + 1} ^ {s}, \\mathcal {F} _ {i + 1} ^ {s}, h _ {t + 1} ^ {s}, \\varepsilon_ {[ t + 1, t + 2 ]} ^ {s}, \\varepsilon_ {[ t, t + 1 ]} ^ {s}). \\qquad (8)", + "image_path": "3cc3e40ac82b5abfe6c8207568408c40b98202d81dcf615d40df127ea8859eac.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 586, + 481, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 481, + 611 + ], + "type": "text", + "content": "Finally, we utilize the temporally aligned results " + }, + { + "bbox": [ + 130, + 586, + 481, + 611 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_{t + 1}^{s}" + }, + { + "bbox": [ + 130, + 586, + 481, + 611 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 586, + 481, + 611 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_{t - 1}^{s}" + }, + { + "bbox": [ + 130, + 586, + 481, + 611 + ], + "type": "text", + "content": " to once again group with " + }, + { + "bbox": [ + 130, + 586, + 481, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t^s" + }, + { + "bbox": [ + 130, + 586, + 481, + 611 + ], + "type": "text", + "content": " and perform alignment:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 199, + 618, + 481, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 618, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 199, + 618, + 481, + 635 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {F}} _ {t} ^ {s} = \\mathrm {C T F A} ^ {s} (\\hat {\\mathcal {F}} _ {t - 1} ^ {s}, \\mathcal {F} _ {t} ^ {s}, \\hat {\\mathcal {F}} _ {t + 1} ^ {s}, h _ {t} ^ {s}, \\varepsilon_ {[ t - 1, t ]} ^ {s}, \\varepsilon_ {[ t, t + 1 ]} ^ {s}). \\qquad (9)", + "image_path": "ac2b36b93611795773bbe4d583ffb2435d58b1713dcaff34f97e27e377c041a7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Through this cascaded feature alignment stage, we effectively propagate non-local video frame information, facilitating temporal feature alignment for both" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 448, + 103 + ], + "type": "text", + "content": "Cross-Modal Temporal Alignment for Event-guided Video Deblurring" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": "non-local and adjacent video frames. All the results of alignment " + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_{t - 1}^{s},\\hat{\\mathcal{F}}_{t}^{s},\\hat{\\mathcal{F}}_{t + 1}^{s}" + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": " are passed to the next scale, " + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "s - 1" + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": ". The right side of Fig. 4 illustrates the overall alignment process for the time step " + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": " in the CTFA module. Since it is applied similarly for all time steps " + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "i\\in \\{t - 1,t,t + 1\\}" + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": ", we will describe it here specifically when " + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "i = t" + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": ", where previously aligned " + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_{t - 1}^{s}" + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_{t + 1}^{s}" + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": " are provided. For leveraging the benefits of coarse-to-fine architecture, we first fuse hidden state aligned feature " + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "h_t^s" + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": " from the previous scale for the time step " + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": " at scale factor " + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 114, + 482, + 201 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 268, + 222, + 481, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 222, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 268, + 222, + 481, + 236 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {t} ^ {s} = \\mathcal {N} _ {f} ^ {s} \\left(\\mathcal {F} _ {t} ^ {s}, h _ {t} ^ {s}\\right), \\tag {10}", + "image_path": "33fb7ca8a59318d0608d26de81f90080bc466fd5b9a1e7ac9d1c5d17c63b86bb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 243, + 481, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 243, + 481, + 292 + ], + "spans": [ + { + "bbox": [ + 131, + 243, + 481, + 292 + ], + "type": "text", + "content": "We simplify the sequential operations of Conv, ReLU, and ResBlock as " + }, + { + "bbox": [ + 131, + 243, + 481, + 292 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_f" + }, + { + "bbox": [ + 131, + 243, + 481, + 292 + ], + "type": "text", + "content": ". Additionally, to leverage temporal information, we align the features at times " + }, + { + "bbox": [ + 131, + 243, + 481, + 292 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 131, + 243, + 481, + 292 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 243, + 481, + 292 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 131, + 243, + 481, + 292 + ], + "type": "text", + "content": " to the current feature at time " + }, + { + "bbox": [ + 131, + 243, + 481, + 292 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 243, + 481, + 292 + ], + "type": "text", + "content": " using event features that encapsulate motion between frames in the following manner:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 235, + 301, + 381, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 301, + 381, + 316 + ], + "spans": [ + { + "bbox": [ + 235, + 301, + 381, + 316 + ], + "type": "interline_equation", + "content": "\\mathcal {T} _ {t - 1 \\rightarrow t} ^ {s} = \\mathcal {N} _ {g, f} ^ {s} (\\hat {\\mathcal {F}} _ {t - 1} ^ {s}, \\mathcal {F} _ {t} ^ {s}, \\varepsilon_ {[ t - 1, t ]} ^ {s}),", + "image_path": "6a39f31e229d557e7dbfcf81c8d7421b0d346dbeb2179fa53a39d27ce0f03d15.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 235, + 319, + 480, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 319, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 235, + 319, + 480, + 335 + ], + "type": "interline_equation", + "content": "\\mathcal {T} _ {t + 1 \\rightarrow t} ^ {s} = \\mathcal {N} _ {g, b} ^ {s} \\left(\\mathcal {F} _ {t} ^ {s}, \\hat {\\mathcal {F}} _ {t + 1} ^ {s}, \\varepsilon_ {[ t, t + 1 ]} ^ {s}\\right), \\tag {11}", + "image_path": "d95eb57bdac6ee6e3ef91c7bb42265c42fd3daa920ca5af4bc79223ab3951a4b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 235, + 337, + 350, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 337, + 350, + 350 + ], + "spans": [ + { + "bbox": [ + 235, + 337, + 350, + 350 + ], + "type": "interline_equation", + "content": "\\mathcal {T} _ {t} ^ {s} = \\mathcal {N} _ {h} ^ {s} (\\mathcal {T} _ {t - 1 \\rightarrow t}) \\| \\mathcal {T} _ {t + 1 \\rightarrow t} ^ {s}),", + "image_path": "2d7e60f6043a82183903807ece1991ef2f8731d714b5b309abe2a4dff518bc95.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "inline_equation", + "content": "\\parallel" + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "text", + "content": " denotes the channel-wise concatenation, and " + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{g,f}^{s},\\mathcal{N}_{g,b}^{s}" + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_h^s" + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "text", + "content": " represent convolution blocks, as illustrated on the right side of Fig. 4. While we obtain temporally aligned features " + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t^s" + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "text", + "content": " using event features and leveraging the advantages of events, for additional feature refinement, we utilize the spatially variant pixel-wise dynamic filter [21,55] mechanism. Dynamic filter " + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t^s" + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "text", + "content": " at scale " + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "text", + "content": " can be calculated through filter generation blocks, " + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_l^s" + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "text", + "content": ", which consists of convolution and resblock, such as " + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t^s = \\mathcal{N}_l^s (\\mathcal{T}_t^s)" + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t^s\\in \\mathbb{R}^{(s_k\\times s_k)\\times H^s\\times W^s}" + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "inline_equation", + "content": "s_k" + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "text", + "content": " is kernel size of dynamic convolution filter, " + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "inline_equation", + "content": "H^{s}" + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "inline_equation", + "content": "W^{s}" + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "text", + "content": " denote height and width of the feature at scale factor " + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 358, + 482, + 478 + ], + "type": "text", + "content": ", respectively. Then, we apply the dynamic convolution operation as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 238, + 487, + 481, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 487, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 238, + 487, + 481, + 502 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathcal {T}} _ {t} ^ {s} (h, w) = \\mathcal {D} _ {t} ^ {s} (h, w) \\otimes \\mathcal {T} _ {t} ^ {s} (h, w) \\tag {12}", + "image_path": "e9b0ba05c6217a5c6f2715f3507b1ab3c0096003efc35f0a28fa78c5bbc2b4cb.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "spans": [ + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "inline_equation", + "content": "h \\in \\{1, \\dots, H^s\\}" + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "inline_equation", + "content": "w \\in \\{1, \\dots, W^s\\}" + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "text", + "content": " denotes the convolution operation. Then, we employ a cross-attention mechanism [38,44] to effectively combine the information of two features, " + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "inline_equation", + "content": "S_t^s" + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\hat{T}_t^s" + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "text", + "content": ". The cross-attention generally examines the correlation between input features (query) and the key-value features. Therefore, we conduct correlation analysis by applying cross-attention between the current blurred frame features to project query " + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "inline_equation", + "content": "S_t^s" + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "text", + "content": " and the aligned features " + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "inline_equation", + "content": "T_t^s" + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "text", + "content": " to project key and values. That is, we generate query, key, and value features, " + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} = W_Q(S_t^s)" + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{K} = W_K(\\hat{T}_t^s)" + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = W_V(\\hat{T}_t^s)" + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "inline_equation", + "content": "W_{(\\cdot)}" + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "text", + "content": " convolution and " + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "text", + "content": " depth-wise convolution. Utilizing these " + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 131, + 510, + 482, + 631 + ], + "type": "text", + "content": ", we compute the attention matrix similarly to Eq. (4)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 231, + 639, + 481, + 664 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 639, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 231, + 639, + 481, + 664 + ], + "type": "interline_equation", + "content": "\\mathcal {A} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {T}}{\\alpha}\\right) \\cdot \\mathbf {V} \\tag {13}", + "image_path": "b562d34d7b9a4471eb58ab86b2264ea53e3441ac866d745b59f2690e25ba80d4.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 157, + 480, + 222 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 1: Quantitative results on the GoPro dataset. The asterisk(*) indicates that the results are not officially on the GoPro so we retrained the official model by us. CMTA-5 and CMTA-7 refer to the results obtained using 5 and 7 input frames, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 157, + 480, + 222 + ], + "lines": [ + { + "bbox": [ + 134, + 157, + 480, + 222 + ], + "spans": [ + { + "bbox": [ + 134, + 157, + 480, + 222 + ], + "type": "table", + "html": "
MethodsMPRNet [45]Restormer [44]NAFNet [6]EDVR [39]ESTRNN [53]RNN-MBP [57]DSTNet [26]VRT [17]
PSNRs32.6632.9233.6926.8331.0233.3234.1634.81
SSIMs0.9590.9610.9670.8430.9110.9630.9680.972
Params(MB)20.126.167.823.62.416.47.518.3
MethodsRVRT [18]Shift-Net [16]UEVD* [15]EFNet [36]REFID [37]SpkNet [7]CMTA-5CMTA-7
PSNRs34.9235.4935.4835.4635.9136.1236.5536.78
SSIMs0.9740.9760.9710.9720.9730.9710.9770.977
Params(MB)10.810.527.98.515.913.59.79.7
", + "image_path": "f80c3286b616bd4cc9704d53a055b92234d2d9de9091ead59b6e000557b240ce.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 243, + 398, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 243, + 398, + 257 + ], + "spans": [ + { + "bbox": [ + 131, + 243, + 398, + 257 + ], + "type": "text", + "content": "Finally, temporally aligned feature, " + }, + { + "bbox": [ + 131, + 243, + 398, + 257 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_t^s" + }, + { + "bbox": [ + 131, + 243, + 398, + 257 + ], + "type": "text", + "content": ", can be obtained by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 261, + 262, + 481, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 262, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 261, + 262, + 481, + 277 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {F}} _ {t} ^ {s} = \\operatorname {M L P} (\\mathcal {A}) + \\mathcal {A}. \\tag {14}", + "image_path": "6f8815aa3f3311b60d61237119475fe05b3a2cf6bfdd21dce68f509c3271da84.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 283, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 331 + ], + "type": "text", + "content": "where MLP denote multi-layer perceptrons, " + }, + { + "bbox": [ + 130, + 283, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 283, + 482, + 331 + ], + "type": "text", + "content": " denote the result of attention operation. Finally, we obtained the aligned feature pyramid " + }, + { + "bbox": [ + 130, + 283, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathcal{F}}_t^s\\}" + }, + { + "bbox": [ + 130, + 283, + 482, + 331 + ], + "type": "text", + "content": " through cascaded temporal feature alignment. As illustrated in Fig. 2, these multi-scale aligned features are fed into the decoder." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 346, + 204, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 346, + 204, + 358 + ], + "spans": [ + { + "bbox": [ + 131, + 346, + 204, + 358 + ], + "type": "text", + "content": "4.4 Decoder" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 364, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 364, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 131, + 364, + 482, + 415 + ], + "type": "text", + "content": "The decoder is designed based on the standard U-Net [32]. It takes multi-scale temporally aligned features " + }, + { + "bbox": [ + 131, + 364, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathcal{F}}_t^s\\}" + }, + { + "bbox": [ + 131, + 364, + 482, + 415 + ], + "type": "text", + "content": " as inputs and produces output feature pyramid " + }, + { + "bbox": [ + 131, + 364, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}(D)_k^s\\}" + }, + { + "bbox": [ + 131, + 364, + 482, + 415 + ], + "type": "text", + "content": ". The final deblurred outputs " + }, + { + "bbox": [ + 131, + 364, + 482, + 415 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 131, + 364, + 482, + 415 + ], + "type": "text", + "content": " using last scale of output feature " + }, + { + "bbox": [ + 131, + 364, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(D)_t^{s = 0}" + }, + { + "bbox": [ + 131, + 364, + 482, + 415 + ], + "type": "text", + "content": " is calculated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 241, + 419, + 481, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 419, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 241, + 419, + 481, + 434 + ], + "type": "interline_equation", + "content": "S _ {t} = B _ {t} + \\operatorname {C o n v} _ {5 \\times 5} (\\mathcal {F} (D) _ {t} ^ {s = 0}) \\tag {15}", + "image_path": "14eeb4f665a571d9b3e1e9ff12f82b71ff85d7ce64415f047804d82684b0fe58.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 439, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 482, + 464 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 439, + 482, + 464 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}_{5\\times 5}" + }, + { + "bbox": [ + 130, + 439, + 482, + 464 + ], + "type": "text", + "content": " represents a conv. layer with a filter size of " + }, + { + "bbox": [ + 130, + 439, + 482, + 464 + ], + "type": "inline_equation", + "content": "5\\times 5" + }, + { + "bbox": [ + 130, + 439, + 482, + 464 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 439, + 482, + 464 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 130, + 439, + 482, + 464 + ], + "type": "text", + "content": " denotes the final estimated sharp frame." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 480, + 230, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 480, + 230, + 495 + ], + "spans": [ + { + "bbox": [ + 131, + 480, + 230, + 495 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 503, + 205, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 503, + 205, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 503, + 205, + 514 + ], + "type": "text", + "content": "5.1 Datasets" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": "GoPro Dataset [23]. For a fair comparison with other previous event-guided deblurring methods, we utilize the same raw events provided by the authors of recent work [36]. These events were generated using ESIM [29] with a randomly generated contrast threshold set to a Gaussian normal distribution of parameters as " + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "inline_equation", + "content": "N(\\mu = 0.2, \\sigma = 0.03)" + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": ". We used the official train and test splits." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "text", + "content": "HighREV Dataset [37]. HighREV consists of high-resolution events and RGB data at " + }, + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "inline_equation", + "content": "1632 \\times 1224" + }, + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "text", + "content": " resolution, designed for both deblurring and interpolation tasks. To evaluate motion deblurring exclusively, we use the " + }, + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "inline_equation", + "content": "11 + 1" + }, + { + "bbox": [ + 130, + 582, + 482, + 629 + ], + "type": "text", + "content": " split, excluding the interpolation ground truth." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Real-world Video Deblurring Dataset. EVRB dataset consists of 11 training sequences and 6 test sequences. With each sequence containing 149 frames, it is well-suited for video deblurring tasks." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 164, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 164, + 91, + 448, + 103 + ], + "type": "text", + "content": "Cross-Modal Temporal Alignment for Event-guided Video Deblurring" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 213 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 213 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 213 + ], + "type": "image", + "image_path": "d297df95791baffe6f5ee30a8cad71b94ba84face5532a358f688724c884a8c7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 222, + 480, + 244 + ], + "lines": [ + { + "bbox": [ + 132, + 222, + 480, + 244 + ], + "spans": [ + { + "bbox": [ + 132, + 222, + 480, + 244 + ], + "type": "text", + "content": "Fig. 5: Visual comparison of deblurring results on the GoPro dataset. The qualitative results of other methods were taken from the results provided by the authors." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 144, + 277, + 470, + 306 + ], + "blocks": [ + { + "bbox": [ + 192, + 256, + 420, + 267 + ], + "lines": [ + { + "bbox": [ + 192, + 256, + 420, + 267 + ], + "spans": [ + { + "bbox": [ + 192, + 256, + 420, + 267 + ], + "type": "text", + "content": "Table 2: Quantitative results on the HighREV dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 144, + 277, + 470, + 306 + ], + "lines": [ + { + "bbox": [ + 144, + 277, + 470, + 306 + ], + "spans": [ + { + "bbox": [ + 144, + 277, + 470, + 306 + ], + "type": "table", + "html": "
MethodsESTRNN [53]DSTNet [26]UEVD [15]EFNet [36]REFID [37]Ours
PSNRs30.3831.7737.4037.9938.3739.12
SSIMs0.9400.9480.9740.9760.9770.980
", + "image_path": "9565fd5b07d4e3590cd4b78be09e690a39017c98c1c40a8b4df3cb8201d075f8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 328, + 361, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 328, + 361, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 328, + 361, + 340 + ], + "type": "text", + "content": "5.2 Comparison on Synthetic Blur Datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "text", + "content": "We present the quantitative results of our frameworks with other frame-based image and video deblurring methods and event-guided motion deblurring methods using the GoPro dataset as depicted in Table 1. When compared to the existing best performance of video deblurring methods, Shift-Net [16], our approach (CMTA-5 model) shows a significant improvement of " + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "inline_equation", + "content": "1.06\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "text", + "content": " in terms of PSNR, demonstrating that our method effectively leverages the temporal dense characteristics of event modality for accurate temporal feature alignment and cross-modality feature enhancement. Moreover, compared to the best-performing SpkNet [7] among existing event-guided motion deblurring methods, CMTA-5 model exhibits a performance improvement of " + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "inline_equation", + "content": "0.43\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "text", + "content": " with a lower model params of " + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "inline_equation", + "content": "3.8\\mathrm{MB}" + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "text", + "content": ". Furthermore, by using 7 input video frames (CMTA-7 model) instead of 5 (CMTA-5 model), we achieved an impressive state-of-the-art performance with a PSNR of " + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "inline_equation", + "content": "36.78\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 350, + 482, + 542 + ], + "type": "text", + "content": " in the GoPro, an increase of 0.23. We further demonstrated the superiority of our approach through the qualitative results in Fig. 5. Also, we conduct experiments on HighREV [37], which consists of real events, and report the results in Table 2. Our method still achieves the best performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 560, + 369, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 560, + 369, + 572 + ], + "spans": [ + { + "bbox": [ + 132, + 560, + 369, + 572 + ], + "type": "text", + "content": "5.3 Comparison on Real-world Blur Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "For comparisons in the EVRB dataset, we trained representative video deblurring methods [12, 16, 18, 26, 39, 53] and event-guided motion deblurring methods [15, 36, 37] on the same training set. Tab 3 presents the quantitative results on the EVRB dataset. The EVRB dataset includes extremely blurred videos captured during exposure times. As a result, frame-based video deblurring methods exhibit subpar performance. For instance, the best-performing network, ShiftNet [16], achieves only " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "30.56\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ", which is " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.42\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " lower than the best-performing" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 135, + 471, + 190 + ], + "blocks": [ + { + "bbox": [ + 198, + 114, + 414, + 125 + ], + "lines": [ + { + "bbox": [ + 198, + 114, + 414, + 125 + ], + "spans": [ + { + "bbox": [ + 198, + 114, + 414, + 125 + ], + "type": "text", + "content": "Table 3: Quantitative results on the EVRB dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 135, + 471, + 190 + ], + "lines": [ + { + "bbox": [ + 143, + 135, + 471, + 190 + ], + "spans": [ + { + "bbox": [ + 143, + 135, + 471, + 190 + ], + "type": "table", + "html": "
MethodsEDVR [39]ESTRNN [53]ERDN [12]DSTNet [26]RVRT [18]
PSNRs29.0229.7928.3229.1530.24
SSIMs0.8860.9110.8930.8980.906
MethodsShift-Net [16]UEVD [15]EFNet [36]REFID [37]Ours
PSNRs30.5630.5530.9830.3331.38
SSIMs0.9220.9150.9270.9180.927
", + "image_path": "b14802cf7bb69e54e4849c004e74abbfcaeb919c3095e18121d2b23762f64499.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 134, + 200, + 481, + 299 + ], + "blocks": [ + { + "bbox": [ + 134, + 200, + 481, + 299 + ], + "lines": [ + { + "bbox": [ + 134, + 200, + 481, + 299 + ], + "spans": [ + { + "bbox": [ + 134, + 200, + 481, + 299 + ], + "type": "image", + "image_path": "34782f7695ccac1073018b3771aeca0fc4244793ee4b28ad2b3fc821b7801a3f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 307, + 449, + 319 + ], + "lines": [ + { + "bbox": [ + 163, + 307, + 449, + 319 + ], + "spans": [ + { + "bbox": [ + 163, + 307, + 449, + 319 + ], + "type": "text", + "content": "Fig. 6: Visual comparison of deblurring results on the EVRB dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 342, + 482, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 401 + ], + "type": "text", + "content": "event-guided deblurring method, EFNet [36]. In contrast, our approach leverages video and event characteristics, effectively restoring even severe motion blur, which may be difficult to recover. Our method outperforms all approaches, achieving the best performance. We show our qualitative comparison with other methods in Fig. 6." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 418, + 238, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 418, + 238, + 431 + ], + "spans": [ + { + "bbox": [ + 131, + 418, + 238, + 431 + ], + "type": "text", + "content": "5.4 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 438, + 481, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 481, + 474 + ], + "type": "text", + "content": "We analyzed the performance contribution of the various modules in our frameworks. For a fair ablation study, we trained all the models for 600 epochs with 5 video frame inputs, conducting all experiments on the GoPro dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 474, + 481, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 545 + ], + "type": "text", + "content": "CRIFE module. To demonstrate the effectiveness of the CRIFE module, we replaced it with concatenation and convolution for comparison. When comparing the first column(Ver.1) and the second column(Ver.2) of Tab 4, we observed a performance gain of " + }, + { + "bbox": [ + 130, + 474, + 481, + 545 + ], + "type": "inline_equation", + "content": "+0.35" + }, + { + "bbox": [ + 130, + 474, + 481, + 545 + ], + "type": "text", + "content": " dB with a small additional parameter " + }, + { + "bbox": [ + 130, + 474, + 481, + 545 + ], + "type": "inline_equation", + "content": "(+0.08\\mathrm{MB})" + }, + { + "bbox": [ + 130, + 474, + 481, + 545 + ], + "type": "text", + "content": ". Similarly, when comparing the performance of the third and fourth rows, we observed performance improvement." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 629 + ], + "type": "text", + "content": "ECITFA module is the most crucial component of our model, performing feature alignment by leveraging information of the events. As in the Tab. 4, when comparing the first (Ver.1) with the third column (Ver.3) of the table, we observed a significant performance gain " + }, + { + "bbox": [ + 130, + 545, + 481, + 629 + ], + "type": "inline_equation", + "content": "(+1.69\\mathrm{dB})" + }, + { + "bbox": [ + 130, + 545, + 481, + 629 + ], + "type": "text", + "content": " upon the insertion of the ECITFA module. This trend is similarly maintained when the CRIFE module is incorporated. When comparing Ver.2 with Ver.4 of the Tab. 4, we observed a notable performance improvement " + }, + { + "bbox": [ + 130, + 545, + 481, + 629 + ], + "type": "inline_equation", + "content": "(+1.55\\mathrm{dB})" + }, + { + "bbox": [ + 130, + 545, + 481, + 629 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Effectiveness of components in ECITFA module. In the Tab. 5, we demonstrated an effectiveness analysis for each component of ECITFA. Comparing the 2nd column with the last column labeled 'Ours', we observed a performance" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "text", + "content": "Cross-Modal Temporal Alignment for Event-guided Video Deblurring" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 161, + 135, + 452, + 174 + ], + "blocks": [ + { + "bbox": [ + 208, + 114, + 403, + 125 + ], + "lines": [ + { + "bbox": [ + 208, + 114, + 403, + 125 + ], + "spans": [ + { + "bbox": [ + 208, + 114, + 403, + 125 + ], + "type": "text", + "content": "Table 4: Ablation study on the GoPro dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 161, + 135, + 452, + 174 + ], + "lines": [ + { + "bbox": [ + 161, + 135, + 452, + 174 + ], + "spans": [ + { + "bbox": [ + 161, + 135, + 452, + 174 + ], + "type": "table", + "html": "
MethodsVer.1Ver.2Ver.3Ver.4
CRIFE✓✓
ECITFA✓✓
PSNRs / Params34.65 / 4.57M35.00 / 4.65M36.34 / 9.59M36.55 / 9.68M
", + "image_path": "98d79ed63db2a5806a5ac275995c92470219ff2fd60b906b7e5fd2307c5e288b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 161, + 216, + 452, + 240 + ], + "blocks": [ + { + "bbox": [ + 132, + 184, + 479, + 206 + ], + "lines": [ + { + "bbox": [ + 132, + 184, + 479, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 479, + 206 + ], + "type": "text", + "content": "Table 5: Effect of the component of ECITFA module on the GoPro dataset. DF denotes dynamic filter operation in the Eq.(12)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 161, + 216, + 452, + 240 + ], + "lines": [ + { + "bbox": [ + 161, + 216, + 452, + 240 + ], + "spans": [ + { + "bbox": [ + 161, + 216, + 452, + 240 + ], + "type": "table", + "html": "
Methodsw/o Cascadedw/o DFw/o AttentionOurs
PSNRs36.0836.1836.2936.55
", + "image_path": "7b060511466c96e365dd93c61b47972fabd5fe6a8f7748098ec898868cebfbaa.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 159, + 282, + 451, + 312 + ], + "blocks": [ + { + "bbox": [ + 132, + 250, + 479, + 273 + ], + "lines": [ + { + "bbox": [ + 132, + 250, + 479, + 273 + ], + "spans": [ + { + "bbox": [ + 132, + 250, + 479, + 273 + ], + "type": "text", + "content": "Table 6: Comparison of CRIFE with various module variants. \"Baseline\" refers to Ver.3 of Tab.4. E and F represent event and frame features, respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 282, + 451, + 312 + ], + "lines": [ + { + "bbox": [ + 159, + 282, + 451, + 312 + ], + "spans": [ + { + "bbox": [ + 159, + 282, + 451, + 312 + ], + "type": "table", + "html": "
MethodsBaselinew/o concat (E+F)w/concat (E+F)
w/o recurrentw/ recurrent(Ours)
PSNRs36.3436.0435.9936.55
", + "image_path": "e75d7befeecd5a65f7a13d1b640625fafe3e1fd467a74351d0a918dab95a961f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 335, + 480, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 480, + 430 + ], + "type": "text", + "content": "gain of our method " + }, + { + "bbox": [ + 130, + 335, + 480, + 430 + ], + "type": "inline_equation", + "content": "(+0.37\\mathrm{dB})" + }, + { + "bbox": [ + 130, + 335, + 480, + 430 + ], + "type": "text", + "content": " when employing spatial pixel-wise dynamic filters, in contrast to not using DF. When utilizing cross-attention (Eq.13) to better leverage long-range pixel-dependencies in the alignment blocks, we observed a performance improvement " + }, + { + "bbox": [ + 130, + 335, + 480, + 430 + ], + "type": "inline_equation", + "content": "(+0.29\\mathrm{dB})" + }, + { + "bbox": [ + 130, + 335, + 480, + 430 + ], + "type": "text", + "content": " compared to aggregate feature using ResBlocks. Finally, we confirmed the effectiveness of leveraging non-local video frame information through a cascaded-based temporal feature alignment method for video deblurring. After using the proposed cascaded structure for temporal feature alignment, we can observe a performance gain " + }, + { + "bbox": [ + 130, + 335, + 480, + 430 + ], + "type": "inline_equation", + "content": "(+0.47\\mathrm{dB})" + }, + { + "bbox": [ + 130, + 335, + 480, + 430 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 430, + 480, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 480, + 514 + ], + "type": "text", + "content": "Effectiveness of components in CRIFE module. To evaluate each component's effectiveness of the CRIFE module, we removed the concatenation between RGB and event features, directly matching their features. However, as shown in the 3rd column of Tab. 6, this approach yielded sub-optimal performance. Additionally, combining event features without a recurrent structure degraded performance (see 4th column). These ablations confirm that the recurrent structure effectively utilizes temporal information from events within exposure time." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 533, + 224, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 224, + 545 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 224, + 545 + ], + "type": "text", + "content": "6 Conclusions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 557, + 480, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 480, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 480, + 641 + ], + "type": "text", + "content": "This paper proposes a video deblurring framework, CMTA, that elaborately considers the characteristics of an event and video. Specifically, we achieve significant performance improvement through intra-frame feature enhancement and inter-frame temporal feature alignment. Furthermore, we construct a real-world deblurring dataset, the EVRB dataset, which will be valuable for evaluating event-guided deblurring methods. Finally, CMTA demonstrates state-of-the-art performance across various deblurring datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 479, + 665 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by the Technology Innovation Program (1415187329,20024355, Development of autonomous driving connec" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "tivity technology based on sensor-infrastructure cooperation) funded By the Ministry of Trade, Industry & Energy(MOTIE, Korea) and the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (NRF2022R1A2B5B03002636)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 184, + 198, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 198, + 197 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 198, + 197 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 211, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 211, + 480, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 211, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 138, + 211, + 480, + 243 + ], + "type": "text", + "content": "1. Aittala, M., Durand, F.: Burst image deblurring using permutation invariant convolutional neural networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 731-747 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 245, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 480, + 277 + ], + "type": "text", + "content": "2. Brandli, C., Berner, R., Yang, M., Liu, S.C., Delbruck, T.: A " + }, + { + "bbox": [ + 138, + 245, + 480, + 277 + ], + "type": "inline_equation", + "content": "240 \\times 180" + }, + { + "bbox": [ + 138, + 245, + 480, + 277 + ], + "type": "text", + "content": " 130 db " + }, + { + "bbox": [ + 138, + 245, + 480, + 277 + ], + "type": "inline_equation", + "content": "3\\mu s" + }, + { + "bbox": [ + 138, + 245, + 480, + 277 + ], + "type": "text", + "content": " latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits 49(10), 2333-2341 (2014)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 278, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 278, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 278, + 480, + 311 + ], + "type": "text", + "content": "3. Cao, M., Zhong, Z., Wang, J., Zheng, Y., Yang, Y.: Learning adaptive warping for real-world rolling shutter correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17785-17793 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 312, + 480, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 480, + 344 + ], + "type": "text", + "content": "4. Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Basic vsr: The search for essential components in video super-resolution and beyond. In: Proceedings of the IEEE conference on computer vision and pattern recognition (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 345, + 480, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 345, + 480, + 366 + ], + "spans": [ + { + "bbox": [ + 138, + 345, + 480, + 366 + ], + "type": "text", + "content": "5. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: Basicvsr++: Improving video superresolution with enhanced propagation and alignment (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 367, + 480, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 367, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 138, + 367, + 480, + 389 + ], + "type": "text", + "content": "6. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: European Conference on Computer Vision. pp. 17-33. Springer (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 390, + 480, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 390, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 390, + 480, + 422 + ], + "type": "text", + "content": "7. Chen, S., Zhang, J., Zheng, Y., Huang, T., Yu, Z.: Enhancing motion deblurring in high-speed scenes with spike streams. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 422, + 480, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 480, + 455 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 480, + 455 + ], + "type": "text", + "content": "8. Cho, H., Jeong, Y., Kim, T., Yoon, K.J.: Non-coaxial event-guided motion deblurring with spatial alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12492-12503 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 456, + 480, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 480, + 498 + ], + "type": "text", + "content": "9. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence 44(1), 154-180 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 500, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 500, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 500, + 480, + 533 + ], + "type": "text", + "content": "0. Gehrig, D., Loquercio, A., Derpanis, K.G., Scaramuzza, D.: End-to-end learning of representations for asynchronous event-based data. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5633-5643 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 533, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 480, + 566 + ], + "type": "text", + "content": "1. Hyun Kim, T., Mu Lee, K., Scholkopf, B., Hirsch, M.: Online video deblurring via dynamic temporal blending network. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 4038-4047 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "type": "text", + "content": "2. Jiang, B., Xie, Z., Xia, Z., Li, S., Liu, S.: Erdn: Equivalent receptive field deformable network for video deblurring. In: European Conference on Computer Vision. pp. 663-678. Springer (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "text", + "content": "3. Jiang, Z., Zhang, Y., Zou, D., Ren, J., Lv, J., Liu, Y.: Learning event-based motion deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3320-3329 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "4. Kim, T.H., Sajjadi, M.S., Hirsch, M., Scholkopf, B.: Spatio-temporal transformer network for video restoration. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 106-122 (2018)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 164, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 164, + 91, + 447, + 102 + ], + "type": "text", + "content": "Cross-Modal Temporal Alignment for Event-guided Video Deblurring" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "15. Kim, T., Lee, J., Wang, L., Yoon, K.J.: Event-guided deblurring of unknown exposure time videos. In: European Conference on Computer Vision. pp. 519-538. Springer (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "text", + "content": "16. Li, D., Shi, X., Zhang, Y., Cheung, K.C., See, S., Wang, X., Qin, H., Li, H.: A simple baseline for video restoration with grouped spatial-temporal shift. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9822-9832 (June 2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 216 + ], + "type": "text", + "content": "17. Liang, J., Cao, J., Fan, Y., Zhang, K., Ranjan, R., Li, Y., Timofte, R., Van Gool, L.: Vrt: A video restoration transformer. arXiv preprint arXiv:2201.12288 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 216, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 480, + 258 + ], + "type": "text", + "content": "18. Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems 35, 378-393 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "19. Lin, J., Cai, Y., Hu, X., Wang, H., Yan, Y., Zou, X., Ding, H., Zhang, Y., Timofte, R., Van Gool, L.: Flow-guided sparse transformer for video deblurring. arXiv preprint arXiv:2201.01893 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 480, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 324 + ], + "type": "text", + "content": "20. Lin, S., Zhang, J., Pan, J., Jiang, Z., Zou, D., Wang, Y., Chen, J., Ren, J.S.: Learning event-driven video deblurring and interpolation. In: ECCV (8). pp. 695-710 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 325, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 358 + ], + "type": "text", + "content": "21. Mildenhall, B., Barron, J.T., Chen, J., Sharlet, D., Ng, R., Carroll, R.: Burst denoising with kernel prediction networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2502-2510 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "type": "text", + "content": "22. Nah, S., Baik, S., Hong, S., Moon, G., Son, S., Timofte, R., Mu Lee, K.: Ntire 2019 challenge on video deblurring and super-resolution: Dataset and study. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "text", + "content": "23. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3883-3891 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "24. Oh, J., Kim, M.: Demfi: deep joint deblurring and multi-frame interpolation with flow-guided attentive correlation and recursive boosting. In: European Conference on Computer Vision. pp. 198-215. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "text", + "content": "25. Pan, J., Bai, H., Tang, J.: Cascaded deep video deblurring using temporal sharpness prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3043-3051 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "type": "text", + "content": "26. Pan, J., Xu, B., Dong, J., Ge, J., Tang, J.: Deep discriminative spatial and temporal network for efficient video deblurring. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (Feb 2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 534, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 578 + ], + "type": "text", + "content": "27. Pan, L., Scheerlinck, C., Yu, X., Hartley, R., Liu, M., Dai, Y.: Bringing a blurry frame alive at high frame-rate with an event camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6820-6829 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 578, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 611 + ], + "type": "text", + "content": "28. Ranjan, A., Black, M.J.: Optical flow estimation using a spatial pyramid network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4161-4170 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "29. Rebecq, H., Gehrig, D., Scaramuzza, D.: Esim: an open event camera simulator. In: Conference on Robot Learning. pp. 969-982. PMLR (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "30. Rim, J., Kim, G., Kim, J., Lee, J., Lee, S., Cho, S.: Realistic blur synthesis for learning image deblurring. In: European conference on computer vision. pp. 487-503. Springer (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "31. Rim, J., Lee, H., Won, J., Cho, S.: Real-world blur dataset for learning and benchmarking deblurring algorithms. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16. pp. 184-201. Springer (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 482, + 205 + ], + "type": "text", + "content": "32. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 205, + 482, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 247 + ], + "type": "text", + "content": "33. Shang, W., Ren, D., Zou, D., Ren, J.S., Luo, P., Zuo, W.: Bringing events into video deblurring with non-consecutively blurry frames. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4531-4540 (October 2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 248, + 482, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 282 + ], + "type": "text", + "content": "34. Shen, W., Bao, W., Zhai, G., Chen, L., Min, X., Gao, Z.: Video frame interpolation and enhancement via pyramid recurrent framework. IEEE Transactions on Image Processing 30, 277-292 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 282, + 482, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 315 + ], + "type": "text", + "content": "35. Su, S., Delbracio, M., Wang, J., Sapiro, G., Heidrich, W., Wang, O.: Deep video deblurring for hand-held cameras. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1279-1288 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "type": "text", + "content": "36. Sun, L., Sakaridis, C., Liang, J., Jiang, Q., Yang, K., Sun, P., Ye, Y., Wang, K., Van Gool, L.: Event-based fusion for motion deblurring with cross-modal attention. In: European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 347, + 482, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 392 + ], + "type": "text", + "content": "37. Sun, L., Sakaridis, C., Liang, J., Sun, P., Cao, J., Zhang, K., Jiang, Q., Wang, K., Van Gool, L.: Event-based frame interpolation with ad-hoc deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18043-18052 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 392, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 413 + ], + "type": "text", + "content": "38. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: NIPS. pp. 5998-6008 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 413, + 482, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 456 + ], + "type": "text", + "content": "39. Wang, X., Chan, K.C., Yu, K., Dong, C., Change Loy, C.: Edvr: Video restoration with enhanced deformable convolutional networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (June 2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 456, + 482, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 490 + ], + "type": "text", + "content": "40. Wang, Y., Lu, Y., Gao, Y., Wang, L., Zhong, Z., Zheng, Y., Yamashita, A.: Efficient video deblurring guided by motion magnitude. In: Proceedings of the European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 490, + 482, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 523 + ], + "type": "text", + "content": "41. Wei, K., Fu, Y., Yang, J., Huang, H.: A physics-based noise formation model for extreme low-light raw denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2758-2767 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 523, + 482, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 523, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 523, + 482, + 544 + ], + "type": "text", + "content": "42. Wieschollek, P., Hirsch, M., Scholkopf, B., Lensch, H.: Learning blind motion deblurring. In: ICCV. pp. 231-240 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 544, + 482, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 578 + ], + "type": "text", + "content": "43. Xu, F., Yu, L., Wang, B., Yang, W., Xia, G.S., Jia, X., Qiao, Z., Liu, J.: Motion deblurring with real events. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2583-2592 (October 2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 578, + 482, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 482, + 621 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 482, + 621 + ], + "type": "text", + "content": "44. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5728-5739 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 621, + 482, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 621, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 621, + 482, + 643 + ], + "type": "text", + "content": "45. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. arXiv preprint arXiv:2102.02808 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 643, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 482, + 665 + ], + "type": "text", + "content": "46. Zhang, H., Xie, H., Yao, H.: Spatio-temporal deformable attention network for video deblurring. In: ECCV (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "text", + "content": "Cross-Modal Temporal Alignment for Event-guided Video Deblurring" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 500 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "47. Zhang, K., Luo, W., Zhong, Y., Ma, L., Liu, W., Li, H.: Adversarial spatio-temporal learning for video deblurring. IEEE Transactions on Image Processing (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "text", + "content": "48. Zhang, L., Zhang, H., Zhu, C., Guo, S., Chen, J., Wang, L.: Fine-grained video deblurring with event camera. In: ICMM. pp. 352-364. Springer (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 162, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 480, + 193 + ], + "type": "text", + "content": "49. Zhang, X., Yu, L.: Unifying motion deblurring and frame interpolation with events. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17765-17774 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 194, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 480, + 225 + ], + "type": "text", + "content": "50. Zhang, X., Yu, L., Yang, W., Liu, J., Xia, G.S.: Generalizing event-based motion deblurring in real-world scenarios. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10734-10744 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "51. Zhang, Y., Qin, H., Wang, X., Li, H.: Rethinking noise synthesis and modeling in raw denoising. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4593-4601 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 260, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 480, + 293 + ], + "type": "text", + "content": "52. Zhong, Z., Cao, M., Ji, X., Zheng, Y., Sato, I.: Blur interpolation transformer for real-world motion from blur. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5713-5723 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "53. Zhong, Z., Gao, Y., Zheng, Y., Zheng, B.: Efficient spatio-temporal recurrent neural network for video deblurring. In: European Conference on Computer Vision. pp. 191-207. Springer (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "text", + "content": "54. Zhong, Z., Zheng, Y., Sato, I.: Towards rolling shutter correction and deblurring in dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9219-9228 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "text", + "content": "55. Zhou, J., Jampani, V., Pi, Z., Liu, Q., Yang, M.H.: Decoupled dynamic filter networks. In: IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR) (Jun 2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 391, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 423 + ], + "type": "text", + "content": "56. Zhu, A.Z., Yuan, L., Chaney, K., Daniilidis, K.: Unsupervised event-based learning of optical flow, depth, and egomotion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 989-997 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 425, + 480, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 466 + ], + "type": "text", + "content": "57. Zhu, C., Dong, H., Pan, J., Liang, B., Huang, Y., Fu, L., Wang, F.: Deep recurrent neural network with multi-scale bi-directional propagation for video deblurring. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 3598-3607 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "text", + "content": "58. Zhu, X., Hu, H., Lin, S., Dai, J.: Deformable convnets v2: More deformable, better results. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9308-9316 (2019)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/COD_ Learning Conditional Invariant Representation for Domain Adaptation Regression/5f47c561-6103-4c2b-b6da-ac9368ec9383_content_list.json b/2024/COD_ Learning Conditional Invariant Representation for Domain Adaptation Regression/5f47c561-6103-4c2b-b6da-ac9368ec9383_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c429df2d964e7ed08a618e5ed6f7a71f77c350f2 --- /dev/null +++ b/2024/COD_ Learning Conditional Invariant Representation for Domain Adaptation Regression/5f47c561-6103-4c2b-b6da-ac9368ec9383_content_list.json @@ -0,0 +1,1854 @@ +[ + { + "type": "text", + "text": "COD: Learning Conditional Invariant Representation for Domain Adaptation Regression", + "text_level": 1, + "bbox": [ + 277, + 140, + 725, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao-Ran Yang1, Chuan-Xian Ren1*, and You-Wei Luo1,2", + "bbox": [ + 282, + 234, + 720, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Mathematics, Sun Yat-Sen University, China \n $^{2}$ School of Mathematics, Jiaying University, China \n{yanghr26@mail2, rchuanx@mail, luoyw28@mail2}.sysu.edu.cn", + "bbox": [ + 279, + 262, + 720, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Aiming to generalize the label knowledge from a source domain with continuous outputs to an unlabeled target domain, Domain Adaptation Regression (DAR) is developed for complex practical learning problems. However, due to the continuity problem in regression, existing conditional distribution alignment theory and methods with discrete prior, which are proven to be effective in classification settings, are no longer applicable. In this work, focusing on the feasibility problems in DAR, we establish the sufficiency theory for the regression model, which shows the generalization error can be sufficiently dominated by the cross-domain conditional discrepancy. Further, to characterize conditional discrepancy with continuous conditioning variable, a novel Conditional Operator Discrepancy (COD) is proposed, which admits the metric property on conditional distributions via the kernel embedding theory. Finally, to minimize the discrepancy, a COD-based conditional invariant representation learning model is proposed, and the reformulation is derived to show that reasonable modifications on moment statistics can further improve the discriminability of the adaptation model. Extensive experiments on standard DAR datasets verify the validity of theoretical results and the superiority over SOTA DAR methods.", + "bbox": [ + 261, + 345, + 738, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Domain Adaptation Regression, Optimal Transport, Invariant Representation Learning, Kernel Embedding Theory.", + "bbox": [ + 261, + 622, + 740, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 678, + 375, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning has shown a promising ability for fitting and achieved great success in various application tasks. However, large-scale data with sufficient annotations plays a vital role in training high-quality deep models while the label annotations for unlabeled data are usually expensive. Therefore, aiming to reduce training costs by transferring knowledge, Unsupervised Domain Adaptation (UDA) [2,14,17,28] has received increasing attention in recent years. Basically, UDA aims to transfer task-related information from a labeled source domain", + "bbox": [ + 212, + 710, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 217, + 824, + 385, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to a target domain without annotations, where the domain gap is characterized as the distribution shift across domains [43]. Following this line, considerable methods have been proposed to alleviate the UDA problem by shift correction and distribution alignment, such as adversarial training for cross-domain samples [14, 52] and explicitly distribution matching by statistical discrepancy minimization [28, 49]. From a theoretical perspective, recent literature [6, 60] show that the marginal distribution $P_{X}$ with covariate shift assumption is insufficient to characterize the cross-domain generalization error, while the conditional distribution alignment indeed serves as a sufficient condition to ensure small error and successful transfer. Mathematically, when the conditional distributions $P_{X|y}$ for every $y$ are aligned, a well-trained source predictor can perform well on the target domain, which is described in Fig. 1. Inspired by theoretical advances, recent works [23, 41, 45] consider the statistical methods for conditional alignment, and show promising performance in mitigating negative transfer and improving recognition accuracy. However, most of the existing results are designed for classification settings with discrete prior on label variables, which induces the infeasibility for regression problems. Consequently, it restricts the application of UDA methods for wider real-world problems, e.g., human pose estimation [36], product condition monitoring [27], depth estimation [42].", + "bbox": [ + 212, + 146, + 787, + 434 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Considering the dilemma of UDA, pioneer works have explored the Domain Adaptation Regression (DAR) framework for solving the distribution shift in continuous label scenarios, where theoretical results [7,35] and correction methods [15,47,51] are provided. To further relax the requirement of labeled training target data in pioneer DAR works, advanced works [4,38] propose unsupervised DAR methods to learn transferable regression models without the help of labeled target data. Nonetheless, existing DAR theory and methods are directly facilitated by the covariate shift framework, where the marginal distribution alignment is considered. Therefore, it inevitably induces the negative transfer and misaligned local structure, which are shown in the results in classification settings [6,60]. In fact, there is a significant obstacle for modeling conditional distribution alignment in regression scenarios, i.e., the conditional shift correction induces infinite matching problems for $P_{X|Y=y}$ on all points $y$ . The continuous label variable with infinite possible outcomes makes it infeasible to directly apply the existing conditional alignment methods for DAR. Thus, it is vital to understand the continuity problem in regression, and develop the generalization error theory and statistical methodology for DAR scenarios.", + "bbox": [ + 212, + 439, + 787, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we focus on the theory guarantee and conditional alignment methodology for successful DAR. Theoretically, we explore the generalization upper bound for continuous output scenarios and show the cross-domain joint error can be sufficiently upper-bounded by the conditional discrepancy. It ensures the sufficiency of conditional alignment still holds for regression scenarios and provides reliable guarantees for theory-driven modeling. Methodologically, to deal with the continuous problem in conditional alignment, we treat the conditional distribution as finite statistical moments in Hilbert space. By exploring the distribution embedding theory of Reproducing Kernel Hilbert Space (RKHS), a novel", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H. Yang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/16737da332bfa6f9c5df6a89905a35a2c74bdd50860d8c1d07db50c0d953f06a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 143, + 683, + 170 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4c86c9bb133f0858a85908027199ac9dc5e091a7dc4256d376478eecd27b60ca.jpg", + "image_caption": [ + "Fig. 1: Illustration of the conditional shift in regression setting, where label value $y \\in [0,1]$ and color gradients imply the continuity. (a) Before adaptation, distribution shift exists between source and target representations so the predictor trained on source domain cannot generalize to target domain. (b) After marginal distribution alignment, distributions of source and target domain are globally aligned. However, representations with different labels may be falsely aligned across domains, leading to significant inconsistency between cross-domain labeling rules. Thus, level sets provided by the source predictor are not suitable for target representations. (c) After conditional distribution alignment, label-wise matching is achieved, where the level sets of cross-domain labeling rules are consistent and the source predictor will be reliable." + ], + "image_footnote": [], + "bbox": [ + 277, + 171, + 415, + 275 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/204bbfbb086977bb66063a36f96f52e1e60519e6ce4dd6d841d8c2085d014f69.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 437, + 171, + 581, + 273 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/18e436871d05a6c11880c8428d53d8e409dc9d0bebc9080b485a7915da558ec2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 172, + 727, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "conditional metric called Conditional Operator Discrepancy (COD) is proposed, which can be interpreted as the conditional Wasserstein distance in RKHS. With the sufficiency theory and metric-based learning principle for regression, a COD-based conditional invariant representation learning method is proposed, which aims to correct the conditional shift on the pushforward measures in latent representation space. By reformulating the empirical COD estimation, a modified principle is developed to further improve the discriminability during knowledge transfer. Generally, our contributions can be summarized as follows:", + "bbox": [ + 212, + 455, + 787, + 578 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The sufficiency theory for the conditional shift framework is established for DAR settings, which shows the cross-domain conditional discrepancy on continuous label variable serves as the major term in generalization error.", + "- A novel COD metric for characterizing conditional distribution discrepancy with continuous and infinite conditioning variable is explored, providing a feasible solution for conditional discrepancy optimization in DAR.", + "- A COD-based representation learning model with reformulation on moment statistics is proposed, which admits discriminability on task-related knowledge and shows superior performance in extensive experiments." + ], + "bbox": [ + 223, + 588, + 784, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 747, + 387, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Domain Adaptation Classification. Based on the distribution shift assumption, most of the UDA methods can be roughly categorized into three types: methods that align marginal distribution, conditional distribution and joint distribution. Most UDA methods consider only the marginal distribution align-", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "COD: Learning Conditional Invariant Representation for DAR", + "bbox": [ + 313, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ment. Discrepancy-based methods attempt to directly minimize the distribution discrepancy and different types of discrepancy measurement greatly enrich this type of methodology. For example, these methods adopt moment matching [28, 49] or extension under manifold framework [17, 33]. In particular, Optimal transport (OT) and Wasserstein distance are introduced to UDA as a sound measurement for domain shift, based on which a large number of methods have been proposed [8-10, 24, 46]. Zhang et al. [59] further develop both theory and methodology to RKHS and Luo et al. [31] extends Kernel Bures metric to conditional distribution alignment. Adversarial-based methods [14, 20, 21] learn domain-invariant features via two-player games. Specifically, they alternatively optimize the feature generator and the domain discriminator to achieve domain confusion. In order to obtain better performance, advanced methods have explored how to utilize the label information to achieve conditional distribution alignment [3, 32, 41, 44, 45, 50, 53, 61] or joint distribution alignment [29, 30]. However, UDA methods for shift correction are proposed for discrete and finite conditioning variables, which are inapplicable for regression.", + "bbox": [ + 212, + 146, + 787, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Domain Adaptation Regression. DA for regression has received relatively little attention compared with classification. Mansour et al. [35] and Cortes et al. [7] introduce theoretical analysis for DA in regression scenarios. Zhao et al. [62] discuss the tradeoffs between accuracy and invariance in learning invariant representations in both classification and regression settings. A range of methods in the shallow regime are proposed to tackle the DAR problem. These methods can be roughly categorized into two types: importance weighting [15,37] and learning domain-invariant representation [8,39,40]. Besides, some attempts have been made in deep regime to tackle regression problem [25,47,51,56]. Unfortunately, most of them belong to semi-supervised DA, which means they need to use a small number of target domain labels. Recently, deep UDA methods for regression that get rid of target label are proposed [4,11,38,57]. RSD [4] states that deep neural networks are less robust to feature scaling in regression problems and aligning the deep features directly will impede task performance for altering feature scale. Thus, a strategy of aligning orthogonal bases of both domains is adopted in RSD. Based on RSD, DARE-GRAM [38] aligns the inverse Gram Matrix of deep features of both domains, and ensures promising improvements in accuracy and stability. Meanwhile, several DAR methods for specific tasks have been explored [1,18,22,26,27,55]. However, the theory and method for conditional shift correction are still unexplored for DAR scenarios, which hinders the in-depth reasoning and development for practical modeling.", + "bbox": [ + 212, + 388, + 787, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 COD: Theory and Methodology", + "text_level": 1, + "bbox": [ + 215, + 726, + 563, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminary", + "text_level": 1, + "bbox": [ + 215, + 756, + 361, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Setup. In this paper, we discuss UDA method dealing with regression problems. Let $\\mathrm{X}$ and $\\mathrm{Y}$ be covariate and label variable taken values on $\\mathcal{X}$ and $\\mathcal{Y}$ . $\\mathbb{P}$ denotes the set of distributions and subscripts represent the corresponding random variable while the superscript represents the domain, e.g. $P_{XY}^{s} \\in \\mathbb{P}_{\\mathcal{X}\\mathcal{Y}}$", + "bbox": [ + 212, + 779, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H. Yang et al.", + "bbox": [ + 271, + 114, + 366, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "implies the joint distribution on source domain. In UDA settings, we are provided with labeled source samples $(x^{s},y^{s})\\sim P_{XY}^{s}$ and unlabeled target samples $(x^{t},y^{t})\\sim P_{XY}^{t}$ , which means the target labels are not available during training. Samples of source and target domain follow different distributions, i.e. $P_{XY}^{s}\\neq P_{XY}^{t}$ . Denote $g:X\\to Z$ as the representation extractor, $h:Z\\to Y$ as the predictor, $\\varepsilon_{s}$ and $\\varepsilon_{t}$ as the source and target prediction error, respectively. UDA methods aim to generalize the learner $h\\circ g$ across domains and minimize target risk $\\varepsilon_{t}$ by mitigating the distribution shift problem. In contrast to the discrete label space in classification settings, unsupervised DAR discusses UDA problems with continuous labels, which means the number of classes is uncountable, leading to great difficulty in aligning conditional distribution $P_{X|Y}$ . The proofs are provided in supplementary materials.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "OT in RKHS. Let $(\\mathcal{X},\\mathcal{B})$ be a measure space with Borel $\\sigma$ -field $\\mathcal{B}$ . Denote $(\\mathcal{H}_{\\mathcal{X}},k_{\\mathcal{X}})$ as the RKHS of $\\mathcal{X}$ , which is generated by the positive definite kernel $k_{\\mathcal{X}}$ . Elements from $\\mathcal{X}$ are mapped to $\\mathcal{H}_{\\mathcal{X}}$ via feature map $\\phi (x) = k_{\\mathcal{X}}(x,\\cdot)$ , which is assumed to satisfy the reproducing properties $\\langle \\phi (x),\\phi (x^{\\prime})\\rangle_{\\mathcal{H}_{\\mathcal{X}}} = k_{\\mathcal{X}}(x,x^{\\prime})$ and $\\langle \\phi (x),f\\rangle_{\\mathcal{H}_x} = f(x),\\forall f\\in \\mathcal{H}_x$ . The mean element $\\mu_X$ in $\\mathcal{H}_{\\mathcal{X}}$ with law $P_{X}$ is given by $\\mu_{X} = \\mathbb{E}_{X}[\\phi (X)]$ , and the covariance operator $\\mathcal{C}_{XX}$ is defined as $\\mathcal{C}_{XX} = E_X(\\phi (X) - \\mu_X)\\otimes (\\phi (X) - \\mu_X)$ .", + "bbox": [ + 212, + 327, + 787, + 433 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For any two distributions $P_X^s, P_X^t \\in \\mathbb{P}_X$ , the Kantorovitch optimal transport in RKHS $\\mathcal{H}_{\\mathcal{K}}$ is considered on pushforward measures $\\phi_{\\#}P_X^s$ and $\\phi_{\\#}P_X^t$ :", + "bbox": [ + 212, + 434, + 785, + 465 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd _ {\\mathrm {K W}} ^ {2} \\left(X ^ {s}, X ^ {t}\\right) = \\inf _ {\\pi_ {\\mathcal {K}} \\in \\Pi \\left(\\phi_ {\\#} P _ {X} ^ {s}, \\phi_ {\\#} P _ {X} ^ {t}\\right)} \\int_ {\\mathcal {H} _ {\\mathcal {K}} \\times \\mathcal {H} _ {\\mathcal {K}}} \\left\\| X ^ {s} - X ^ {t} \\right\\| _ {\\mathcal {H} _ {\\mathcal {K}}} ^ {2} d \\pi_ {\\mathcal {K}} \\left(X ^ {s}, X ^ {t}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 477, + 785, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Zhang et al. [59] show that if the pushforward measures are Gaussian, the Kernel Gaussian Wasserstein (KGW) distance between them can be written as", + "bbox": [ + 214, + 518, + 785, + 550 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd _ {\\mathrm {K G W}} ^ {2} \\left(X ^ {s}, X ^ {t}\\right) = \\| u _ {X} ^ {s} - u _ {X} ^ {t} \\| _ {\\mathcal {H} _ {\\kappa}} ^ {2} + t r \\left(\\mathcal {C} _ {X X} ^ {s s} + \\mathcal {C} _ {X X} ^ {t t} - 2 \\mathcal {C} _ {X X} ^ {s t}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 290, + 561, + 785, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{C}_{XX}^{st} = \\sqrt{\\sqrt{\\mathcal{C}_{XX}^{ss}}\\mathcal{C}_{XX}^{tt}\\sqrt{\\mathcal{C}_{XX}^{ss}}}$ and $\\mathcal{C}_{XX}^{ss},\\mathcal{C}_{XX}^{tt}$ are the covariance operators of $P_{X}^{s}$ and $P_{X}^{t}$ on $\\mathcal{H}_{\\mathcal{K}}$ , respectively.", + "bbox": [ + 212, + 593, + 785, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark 1. Note that KGW consists of two terms, i.e., the Maximum Mean Discrepancy (MMD) on first-order statistics and kernel Bures metric on second-order statistics. It provides a basic innovation for exploring conditional discrepancy metric by considering the conditional statistics in RKHS, i.e., first-order statistics in literature [48] and second-order statistics in literature [13].", + "bbox": [ + 212, + 638, + 787, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Generalization Error Analysis for DAR", + "text_level": 1, + "bbox": [ + 214, + 738, + 586, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Most of DAR methods focus on aligning the marginal distribution $P_{X}$ . However, Zhao et al. [60] have shown that in the presence of label shift, only aligning the marginal distribution does not guarantee satisfying generalization performance on target domain. In fact, [6, 60] have proven the following theorem (1) for classification problems:", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "COD: Learning Conditional Invariant Representation for DAR", + "bbox": [ + 313, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 1 ([60]). For representation variable $Z = g(X)$ , suppose inequality $d_{JS}(P_Y^s, P_Y^t) \\geq d_{JS}(P_Z^s, P_Z^t)$ holds, then for predictor $h: Z \\to Y$ , we have", + "bbox": [ + 215, + 146, + 787, + 178 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\varepsilon_ {s} (h \\circ g) + \\varepsilon_ {t} (h \\circ g) \\geq \\frac {1}{2} \\left(\\sqrt {d _ {J S} (P _ {Y} ^ {s} , P _ {Y} ^ {t})} - \\sqrt {d _ {J S} (P _ {Z} ^ {s} , P _ {Z} ^ {t})}\\right) ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 294, + 189, + 785, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $d_{JS}$ denotes the Jensen-Shannon divergence.", + "bbox": [ + 215, + 233, + 584, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Assuming the label distributions differ between source and target domain, i.e., $d_{JS}(P_Y^s,P_Y^t) > 0$ , the lower bound shows that good alignment of representation distributions leads to bad joint error. In other words, when label shift exists, which is rather common in realistic scenarios, only aligning the marginal distribution is not enough for low target error.", + "bbox": [ + 215, + 257, + 782, + 332 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To deal with the insufficiency problem shown in the lower bound analysis, it is intuitive to introduce label information into generalization analysis, e.g., the balanced error rate for classification scenarios [6]. Analogously, for the continuous regression scenarios, we first define the continuous balanced error rate for quantifying maximum conditional risk in DAR.", + "bbox": [ + 215, + 333, + 784, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Definition 1. The continuous balanced error rate of predictor $h$ on distribution $P$ is defined as:", + "bbox": [ + 215, + 417, + 785, + 446 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {c B E R} _ {\\mathrm {P}} (h, Y) := \\sup _ {(x, y) \\in \\mathcal {X} \\oplus \\mathcal {Y}} P (h (x) \\neq Y \\mid Y = y). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 459, + 785, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Intuitively, the continuous balanced error rate can be regarded as the maximum potential risk for predicting sample $x$ with label $y$ . Based on cBER and conditional shift correction with representation transformation $Z = g(X)$ , a sufficient condition for DAR can be induced as follows.", + "bbox": [ + 215, + 496, + 782, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Theorem 2. For representation variable $Z^s = g(X^s)$ and $Z^t = g(X^t)$ , if $g(\\cdot)$ satisfies conditional invariant property, i.e.,", + "bbox": [ + 215, + 564, + 782, + 595 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP _ {Z \\mid Y = y} ^ {s} = P _ {Z \\mid Y = y} ^ {t}, \\quad \\forall y \\in \\mathcal {Y}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 607, + 784, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "then for any predictor $h: Z \\to Y$ , we have", + "bbox": [ + 240, + 633, + 547, + 648 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\varepsilon_ {s} (h) + \\varepsilon_ {t} (h) \\leq 2 \\mathrm {c B E R} _ {\\mathrm {P s}} (h, Y). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 662, + 784, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Therefore, the key for DA, no matter for classification or regression problems, is to achieve conditional distribution alignment. In classification settings, conditional alignment can be achieved by conducting class-wise marginal distribution alignment with the help of target pseudo label. However, in regression scenarios, the label variable is continuous, which means previous methods correcting conditional shift in classification settings are no longer applicable. Obviously, it is impossible to align the $P_{X|y}$ for every single $y$ when $y$ is a continuous variable and the possible number of $y$ is infinite. Thus, it is important and meaningful to explore a new framework that is capable of conditional distribution alignment for continuous label variable.", + "bbox": [ + 215, + 689, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H. Yang et al.", + "bbox": [ + 271, + 114, + 366, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ca6e49b2c95c085cdfb1e4fc57ed3ce0542789eee74f050f932ee74df187057c.jpg", + "image_caption": [ + "Fig. 2: Illustration of COD metric. (a) Classification DA methods usually correct conditional shift by class-wise computations on discrete clusters $P_{X|y}$ , which is infeasible for regression due to the infinite slices $P_{X|y}$ . (b) In COD, the continuous conditional distributions are embedded into the RKHS and characterized by finite statistical moments in kernel spaces, which do not rely on specific conditions $y$ . Under the guarantees of distribution embedding property, the conditional alignment is equivalent to the matching on first-order and second-order statistics, i.e., mean embedding operator $\\mathcal{U}_{X|Y}$ and covariance operator $\\mathcal{C}_{XX|Y}$ . When COD is minimized, conditional alignment is achieved, and the level sets of cross-domain labeling rules are then aligned." + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 769, + 270 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 COD metric for DAR", + "text_level": 1, + "bbox": [ + 215, + 431, + 441, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Theory for COD. Due to the continuity of label variable, it is impractical to correct conditional shift by aligning marginal distribution of $P_{X|y}$ for every single $y$ . For this inevitable obstacle, an intuitive idea is can we align all $P_{X|y}$ as a whole? i.e., $P_{X|Y}$ , rather than align $P_{X|y}$ one by one? It is more than attractive that if we can characterize $P_{X|Y}$ with a well-defined discrepancy, then all $P_{X|y}$ can be aligned by aligning the whole $P_{X|Y}$ .", + "bbox": [ + 212, + 453, + 787, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In kernel embedding theory, a distribution can be characterized by a point in RKHS. Then $P_{X|Y}$ can be seen as a point set in RKHS in which every element corresponds to a distribution $P_{X|y}$ for a fixed $y$ . From another perspective, we can consider $P_{X|Y}$ as the distribution in RKHS and measure the cross-domain conditional discrepancy by defining the COD distance on their statistical moments in RKHS. Following the definitions of conditional statistics in pioneer works, i.e., the conditional mean operator as $\\mathcal{U}_{X|Y} \\coloneqq \\mathcal{C}_{XY}\\mathcal{C}_{YY}^{-1}$ [48] and conditional covariance operator as $\\mathcal{C}_{XX|Y} = \\mathcal{C}_{XX} - \\mathcal{C}_{XY}\\mathcal{C}_{YY}^{-1}\\mathcal{C}_{YX}$ [13], we present the definition of COD as follows.", + "bbox": [ + 212, + 545, + 787, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Definition 2 (COD). Given conditional distributions $P_{X|Y}^{s}$ and $P_{X|Y}^{t}$ , denote $\\mathcal{C}_{XX|Y}^{ss}$ and $\\mathcal{C}_{XX|Y}^{tt}$ as their corresponding covariance operators. The COD is defined as", + "bbox": [ + 212, + 686, + 787, + 734 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nd _ {\\mathrm {C O D}} ^ {2} \\left(P _ {X | Y} ^ {s}, P _ {X | Y} ^ {t}\\right) = \\left\\| \\mathcal {U} _ {X | Y} ^ {s} - \\mathcal {U} _ {X | Y} ^ {t} \\right\\| _ {\\mathcal {H} _ {\\mathcal {K}}} ^ {2} + \\operatorname {t r} \\left(\\mathcal {C} _ {X X | Y} ^ {s s} + \\mathcal {C} _ {X X | Y} ^ {t t} - 2 \\mathcal {C} _ {X X | Y} ^ {s t}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 741, + 785, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{C}_{XX|Y}^{st} = \\sqrt{\\sqrt{\\mathcal{C}_{XX|Y}^{ss}}\\mathcal{C}_{XX|Y}^{tt}\\sqrt{\\mathcal{C}_{XX|Y}^{ss}}}$ .", + "bbox": [ + 215, + 773, + 529, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Theoretically, it can be proved that COD indeed admits the metric property on conditional distributions, i.e., two conditional distributions are the same if", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "COD: Learning Conditional Invariant Representation for DAR", + "bbox": [ + 313, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and only if the COD metric between them is zero. An intuitive illustration of COD is shown in Fig. 2. Formally, COD admits two key properties.", + "bbox": [ + 212, + 146, + 787, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Proposition 1. (a) COD defines a metric between Gaussian measures.", + "bbox": [ + 214, + 183, + 732, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(b) Let $(\\mathcal{X},\\mathcal{B}_{\\mathcal{X}})$ be the locally compact and Hausdorff measurable space and $k$ be $c_{0}$ -universal kernel. Assuming that $(\\phi (X),\\psi (Y))$ is a Gaussian random variable in $\\mathcal{H}_{\\mathcal{X}}\\oplus \\mathcal{H}_{\\mathcal{Y}}$ . For any $P_{X|Y}^{s},P_{X|Y}^{t}\\in \\mathbb{P}_{\\mathcal{X}|\\mathcal{Y}}$ , we have", + "bbox": [ + 214, + 198, + 785, + 246 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nd _ {\\mathrm {C O D}} ^ {2} \\left(P _ {X | Y} ^ {s}, P _ {X | Y} ^ {t}\\right) = 0 \\Rightarrow P _ {X | Y} ^ {s} = P _ {X | Y} ^ {t}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 256, + 785, + 275 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Empirical Estimation of COD. Based on the above properties, we can perform conditional distribution alignment by minimizing the COD metric between source and target domain. To do this, we shall derive the empirical estimation of COD, which can be obtained with the help of the kernel trick.", + "bbox": [ + 212, + 282, + 782, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Let $(\\mathbf{x}_i^s,\\mathbf{y}_i^s)_{i = 1}^{n_s}$ and $(\\mathbf{x}_i^t,\\mathbf{y}_i^t)_{i = 1}^{n_t}$ be two set of samples drawn i.i.d. from source and target domain. For simplicity, $n_s$ and $n_t$ are both set to $n$ . In kernel method, samples are mapped to RKHS $\\mathcal{H}_{\\mathcal{X}}\\oplus \\mathcal{H}_{\\mathcal{Y}}$ by the implicit feature map $(\\phi ,\\psi)$ . Denote the feature map matrices by $(\\Phi^s,\\Psi^s)$ and $(\\Phi^t,\\Psi^t)$ , the explicit kernel matrices $\\mathbf{K}_X^{ss} = \\Phi^s\\Phi^{sT}$ , $\\mathbf{K}_Y^{ss} = \\Psi^s\\Psi^{sT}$ are computed as $(\\mathbf{K}_X^{ss})_{ij} = k_{\\mathcal{X}}(\\mathbf{x}_i^s,\\mathbf{x}_j^s)$ , $(\\mathbf{K}_Y^{ss})_{ij} = k_{\\mathcal{Y}}(\\mathbf{y}_i^s,\\mathbf{y}_j^s)$ , respectively. And so as $\\mathbf{K}_X^{tt}$ , $\\mathbf{K}_X^{st}$ and $\\mathbf{K}_Y^{tt}$ . Let $\\mathbf{I}_n$ be the n-dimensional identity matrix, $\\mathbf{1}_n$ be the n-dimensional vector with all elements equal to 1. The $n$ -dimensional centering matrix can be defined as $\\mathbf{H}_n = \\mathbf{I}_n - \\frac{1}{n}\\mathbf{1}_n\\mathbf{1}_n^T$ . Then the centralized kernel matrix is defined as $\\mathbf{G} = \\mathbf{H}_n\\mathbf{K}\\mathbf{H}_n$ .", + "bbox": [ + 214, + 342, + 787, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With above notations, the covariance operator $\\mathcal{C}_{XY}$ can be empirically estimated as $\\hat{\\mathcal{C}}_{XY} = \\frac{1}{n} (\\pmb {\\Phi} - \\hat{\\mu}_X\\pmb {1}^T)(\\pmb {\\Psi} - \\hat{\\mu}_Y\\pmb {1}^T)^T = \\frac{1}{n}\\pmb {\\Phi}\\pmb {H}\\pmb {\\Psi}^T$ , then the conditional moment statistics can be estimated as", + "bbox": [ + 212, + 479, + 785, + 523 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {U}} _ {X \\mid Y} = (\\frac {1}{n} \\boldsymbol {\\Phi} \\mathbf {H} \\boldsymbol {\\Psi} ^ {T}) (\\frac {1}{n} \\boldsymbol {\\Psi} \\mathbf {H} \\boldsymbol {\\Psi} ^ {T}) ^ {- 1}, \\quad \\hat {\\mathcal {C}} _ {X X \\mid Y} = \\frac {1}{n} \\boldsymbol {\\Phi} \\mathbf {H} _ {n} \\mathbf {A} (\\boldsymbol {\\Phi} \\mathbf {H} _ {n} \\mathbf {A}) ^ {T}, \\quad (9)\n$$\n", + "text_format": "latex", + "bbox": [ + 259, + 530, + 785, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{A}$ satisfies that $\\mathbf{B} = \\mathbf{A}\\mathbf{A}^T$ and $B\\triangleq \\mathbf{I}_n - \\frac{1}{n\\varepsilon}\\left[\\mathbf{G}_Y - \\mathbf{G}_Y\\left(\\mathbf{G}_Y + \\varepsilon n\\mathbf{I}_n\\right)^{-1}\\mathbf{G}_Y\\right]$ . Based on these estimations, the closed-form empirical estimation of the COD metric can be derived as follows.", + "bbox": [ + 212, + 566, + 797, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Proposition 2. The empirical Conditional Operator Discrepancy distance is", + "bbox": [ + 214, + 626, + 772, + 642 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {d} _ {\\mathrm {C O D}} ^ {2} = \\operatorname {t r} \\left(\\mathbf {K} _ {Y} ^ {s s} \\left(\\mathbf {K} _ {Y} ^ {s s}\\right) ^ {- 1} \\mathbf {K} _ {X} ^ {s s} \\left(\\mathbf {K} _ {Y} ^ {s s}\\right) ^ {- 1}\\right) + \\operatorname {t r} \\left(\\mathbf {K} _ {Y} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1} \\mathbf {K} _ {X} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1}\\right) \\\\ - 2 \\operatorname {t r} \\left(\\mathbf {K} _ {Y} ^ {t s} \\left(\\mathbf {K} _ {Y} ^ {s s}\\right) ^ {- 1} \\mathbf {K} _ {X ^ {s} X ^ {t}} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1}\\right) \\\\ + \\varepsilon \\operatorname {t r} \\left[ \\mathbf {G} _ {X} ^ {s} \\left(\\varepsilon n \\mathbf {I} _ {n} + \\mathbf {G} _ {Y} ^ {s}\\right) ^ {- 1} \\right] + \\varepsilon \\operatorname {t r} \\left[ \\mathbf {G} _ {X} ^ {t} \\left(\\varepsilon n \\mathbf {I} _ {n} + \\mathbf {G} _ {Y} ^ {t}\\right) ^ {- 1} \\right] \\tag {10} \\\\ \\left. - \\frac {2}{n} \\left\\| \\left(\\mathbf {H} _ {n} \\mathbf {A} ^ {t}\\right) ^ {T} \\mathbf {K} _ {X} ^ {t s} \\left(\\mathbf {H} _ {n} \\mathbf {A} ^ {s}\\right) \\right\\| _ {*} \\right., \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 251, + 650, + 784, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\|_{*}$ is the nuclear norm. Empirically, $\\mathbf{K}_Y^{-1}$ is approximated by the regularized formulation $(\\mathbf{K}_Y + \\lambda \\mathbf{I})^{-1}$ .", + "bbox": [ + 214, + 757, + 782, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Discriminability of empirical COD. As discussed above, the first term of COD measures the distance of the first-order statistic between conditional distributions in RKHS. However, recent study [5,54] shows that the mean matching", + "bbox": [ + 214, + 794, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H. Yang et al.", + "bbox": [ + 271, + 114, + 366, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "may reduce the discriminability when bringing the two distributions close, which undoubtedly degrades the performance of the algorithm. Therefore, it is worth making an in-depth analysis of the discriminability of COD as well.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Without loss of generality, we first consider Kronechker Delta kernel on conditioning variable $Y$ , i.e., $k(\\mathbf{y}_1,\\mathbf{y}_2) = \\delta (\\mathbf{y}_1,\\mathbf{y}_2)$ . More explicitly, $k(\\mathbf{y}_1,\\mathbf{y}_2)$ equals to 1 if $\\mathbf{y}_1 = \\mathbf{y}_2$ or 0 otherwise. For empirical scenarios with finite samples, the numbers of samples and observed labels are finite, which implies the number of distinct label $c$ will be obviously no more than sample size $n$ . For convenience, we call two samples are in the same class when their label values are the same. Consequently, the first term in COD can be reformulated as follows.", + "bbox": [ + 212, + 191, + 784, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Proposition 3. Under the conditions mentioned above, the empirical estimation of $d_{CMMD}$ can be written as", + "bbox": [ + 214, + 305, + 782, + 335 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\widehat {d} _ {\\mathrm {C M M D}} ^ {2} = \\sum_ {p = 1} ^ {c} \\frac {1}{(\\lambda + n _ {p}) ^ {2}} \\left(\\sum_ {i, j} k (\\mathbf {x} _ {(p), i} ^ {s}, \\mathbf {x} _ {(p), j} ^ {s})\\right) \\\\ + \\sum_ {p = 1} ^ {c} \\frac {1}{(\\lambda + n _ {p}) ^ {2}} \\left(\\sum_ {i, j} k \\left(\\mathbf {x} _ {(p), i} ^ {t}, \\mathbf {x} _ {(p), j} ^ {t}\\right)\\right) \\tag {11} \\\\ - 2 \\sum_ {p = 1} ^ {c} \\frac {1}{(\\lambda + n _ {p}) ^ {2}} \\left(\\sum_ {i, j} k (\\mathbf {x} _ {(p), i} ^ {s}, \\mathbf {x} _ {(p), j} ^ {t})\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 345, + 784, + 494 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $n_i$ denotes the sample-size of $i$ -th label such that $n = \\sum_{i=1}^{c} n_i$ , $\\mathbf{x}_{(p),i}$ denotes the $i$ -th sample of $p$ -th label.", + "bbox": [ + 214, + 503, + 782, + 534 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Eq. (11) contains three terms. The third term characterizes the cross-domain intra-class similarity. When Eq. (11) is minimized, the third term is maximized so that the cross-domain sample with the same label will be aligned, leading to conditional alignment as expected. However, the first two terms of the equation represent the intra-class similarity of both domains. Minimizing them will decrease the intra-class similarity, which means intra-class samples will be pushed away from each other and the discriminability of representations is degraded.", + "bbox": [ + 212, + 541, + 784, + 647 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Remark 2. Note the Kronechker Delta kernel is defined as", + "bbox": [ + 215, + 655, + 635, + 670 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nk \\left(\\mathbf {y} _ {1}, \\mathbf {y} _ {2}\\right) = \\left\\{ \\begin{array}{l} 1 \\text {i f} \\mathbf {y} _ {1} = \\mathbf {y} _ {2} \\\\ 0 \\text {o t h e r w i s e} \\end{array} \\right.. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 676, + 784, + 718 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "When it comes to the Gaussian kernel for empirical modeling, the case of $\\mathbf{y}_1 = \\mathbf{y}_2$ is the same; for other cases, the kernel values can be arbitrarily close to zero, i.e., for any $\\epsilon > 0$ , there always width $\\sigma_0$ , such that", + "bbox": [ + 214, + 726, + 782, + 772 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nk \\left(\\mathbf {y} _ {1}, \\mathbf {y} _ {2}\\right) = \\exp \\left(- \\| \\mathbf {y} _ {1} - \\mathbf {y} _ {2} \\| _ {2} ^ {2} / \\sigma_ {0} ^ {2}\\right) < \\epsilon . \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 782, + 784, + 800 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Thus, it is generally reasonable to use Gaussian kernel, which admits a similar property to Kronechker Delta kernel, for empirical modeling.", + "bbox": [ + 214, + 809, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "COD: Learning Conditional Invariant Representation for DAR", + "bbox": [ + 313, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Remark 3. Note that although we treat the number of distinct labels as finite numbers in mini-batch training, it does not mean class-wise matching can be conducted. In fact, even though the pseudo target label has finite range in minibatch, it still admits the continuous property, which means we can not cluster samples as usually done in classification methods. When two samples have predicted labels with little numerical difference, they can be regarded as samples of the same class in classification settings. Nonetheless, this kind of clustering is infeasible in regression settings since the labels are continuous. Therefore, class-wise alignment is not practical even in mini-batch training.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "With these observation, it is reasonable to reformulate Eq. (11) to promote the discriminability of representations:", + "bbox": [ + 212, + 287, + 787, + 318 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\widehat {d} _ {\\mathrm {C M M D} _ {\\mathrm {m o d}}} = - \\operatorname {t r} \\left(\\mathbf {K} _ {Y} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1} \\mathbf {K} _ {X} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1}\\right) - \\operatorname {t r} \\left(\\mathbf {K} _ {Y} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1} \\mathbf {K} _ {X} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1}\\right) \\\\ - 2 \\operatorname {t r} \\left(\\mathbf {K} _ {Y ^ {t} Y ^ {s}} \\left(\\mathbf {K} _ {Y} ^ {s s}\\right) ^ {- 1} \\mathbf {K} _ {X ^ {s} X ^ {t}} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1}\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 323, + 761, + 363 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Now minimizing Eq. (14) will maximize intra-class similarity regardless of domain, yielding conditional matching with great discriminability. Then the empirical COD is modified as", + "bbox": [ + 212, + 368, + 787, + 412 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\widehat {d} _ {\\mathrm {C O D} _ {\\mathrm {m o d}}} = - \\operatorname {t r} \\left(\\mathbf {K} _ {Y} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1} \\mathbf {K} _ {X} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1}\\right) - \\operatorname {t r} \\left(\\mathbf {K} _ {Y} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1} \\mathbf {K} _ {X} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1}\\right) \\\\ - 2 \\operatorname {t r} \\left(\\mathbf {K} _ {Y ^ {t} Y ^ {s}} \\left(\\mathbf {K} _ {Y} ^ {s s}\\right) ^ {- 1} \\mathbf {K} _ {X ^ {s} X ^ {t}} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1}\\right) \\\\ + \\varepsilon \\operatorname {t r} \\left[ \\mathbf {G} _ {X} ^ {s} \\left(\\varepsilon n \\mathbf {I} _ {n} + \\mathbf {G} _ {Y} ^ {s}\\right) ^ {- 1} \\right] + \\varepsilon \\operatorname {t r} \\left[ \\mathbf {G} _ {X} ^ {t} \\left(\\varepsilon n \\mathbf {I} _ {n} + \\mathbf {G} _ {Y} ^ {t}\\right) ^ {- 1} \\right] \\tag {15} \\\\ - \\frac {2}{n} \\left\\| \\left(\\mathbf {H} _ {n} \\mathbf {A} ^ {t}\\right) ^ {T} \\mathbf {K} _ {X} ^ {t s} \\left(H _ {n} \\mathbf {A} ^ {s}\\right) \\right\\| _ {*} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 420, + 784, + 517 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3.4 COD-based DAR Modeling", + "text_level": 1, + "bbox": [ + 214, + 532, + 488, + 547 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "According to theorem (2), when the conditional distributions $P_{X^s|Y}$ and $P_{X^t|Y}$ are aligned, a well-trained source predictor can gain promising performance on target domain. However, when the shift is significant, it is difficult for predictor to produce high-quality pseudo labels at the beginning. For this reason, the model is likely to miss the optimal optimization direction in the early training stage and thus fall into sub-optimal solutions. Therefore, the marginal alignment can improve the reliability of initial pseudo labels, and further benefit the conditional alignment. Overall, the learning objective of COD can be formulated as", + "bbox": [ + 212, + 553, + 787, + 675 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {g, h} \\mathcal {L} _ {\\mathrm {s r c}} + \\lambda_ {1} \\widehat {d} _ {\\mathrm {C O D} _ {\\mathrm {m o d}}} + \\lambda_ {2} \\widehat {d} _ {\\mathrm {K G W}}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 679, + 784, + 703 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\lambda_1, \\lambda_2$ are the trade-off parameters. For regression task, the source risk $\\mathcal{L}_{\\mathrm{src}}$ is commonly set as Mean Square Error (MSE) over continuous outputs:", + "bbox": [ + 214, + 709, + 785, + 739 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s r c}} = \\frac {1}{n ^ {s}} \\sum_ {i = 1} ^ {n ^ {s}} \\left\\| \\hat {\\mathbf {y}} _ {i} ^ {s} - \\mathbf {y} _ {i} ^ {s} \\right\\| _ {2} ^ {2} \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 747, + 784, + 785 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{y}}_i^s$ is the predicted value of the $i$ -th source sample. Note that the marginal alignment term can be replaced by any other discrepancy metric, while we adopt KGW here to preserve the consistency with OT framework.", + "bbox": [ + 214, + 792, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H. Yang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 162 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We evaluate our method with several state-of-the-art domain adaptation methods on three regression benchmark datasets, implementation details and additional illustrations are provided in supplementary material.", + "bbox": [ + 212, + 179, + 782, + 224 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "dSprites [19] is a standard 2D synthetic dataset for deep representation learning, which contains three domains each with 737,280 images: Color (C), Noisy (N) and Scream (S). Following [4], we consider three factors for regression tasks: scale and two plane coordinates, while the orientation factor is excluded for its complexity. We evaluate all methods on six transfer tasks: $\\mathbf{C} \\rightarrow \\mathbf{N}$ , $\\mathbf{C} \\rightarrow \\mathbf{S}$ , $\\mathbf{N} \\rightarrow \\mathbf{C}$ , $\\mathbf{N} \\rightarrow \\mathbf{S}$ , $\\mathbf{S} \\rightarrow \\mathbf{C}$ and $\\mathbf{S} \\rightarrow \\mathbf{N}$ . Following previous works [4,38], the sum of Mean Absolute Error (MAE) on all three sub-regression tasks is reported as the evaluation metric.", + "bbox": [ + 212, + 224, + 784, + 345 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Biwi kinect [12] is a real-world dataset containing two domains according to gender: Female $(\\mathbf{F})$ with 5874 images and Male $(\\mathbf{M})$ with 9804 images. The three factors that can be employed for regression tasks are pitch, yaw and roll. We evaluate all methods on two transfer tasks: $\\mathbf{F} \\rightarrow \\mathbf{M}$ and $\\mathbf{M} \\rightarrow \\mathbf{F}$ . The sum of MAE on three sub-regression tasks is reported.", + "bbox": [ + 212, + 347, + 784, + 421 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "MPI3D [16] is a simulation-to-real dataset of 3D objects which consist of three different domain: Toy (T), RealistiC (RC) and ReaL (RL). Each domain contains 1,036,800 images. There are two factors in MPI3D that can be employed for regression tasks: a rotation about a vertical axis and a second rotation about a horizontal axis. We evaluate all methods on six transfer tasks: RL $\\rightarrow$ RC, RL $\\rightarrow$ T, RC $\\rightarrow$ T, RC $\\rightarrow$ RL, T $\\rightarrow$ RL and T $\\rightarrow$ RC. The sum of MAE on two sub-regression tasks is reported.", + "bbox": [ + 212, + 422, + 784, + 527 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison methods. SOTA methods for UDA and DAR are selected: TCA [40], MCD [45], DAN [28], DANN [14], JDOT [8], AFN [58], RSD [4], DARE-GRAM [38]. Following previous works [4,38], Mean Absolute Error (MAE) is used as our evaluation metric across all regression tasks. In comparison experiment, the proposed DAR model in Eq. (16) is denoted as COD for convenience.", + "bbox": [ + 212, + 529, + 785, + 604 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Results and Analysis", + "text_level": 1, + "bbox": [ + 215, + 628, + 434, + 645 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison on dSprites. As shown in Tab. 1, COD achieves the best performance among all competing methods in nearly all tasks of dSprites. In $\\mathbf{C} \\rightarrow \\mathbf{N}$ and $\\mathbf{C} \\rightarrow \\mathbf{S}$ , in which the Resnet-18 and UDA methods exhibit terrible performance because of the significant domain gap. Although SOTA DAR methods have greatly boosted the performance compared with classification methods, COD still achieves a significant improvement, reducing the error to only $40\\%$ of that produced by SOTA methods. These results demonstrate that the conditional invariant representations learned by COD indeed ensure better discriminability. Comparison on Biwi Kinect. Since Biwi Kinect is the closest to real-world scenarios among the three benchmark datasets, the performance improvement of model is more challenging. In Tab. 2, the performance gap between UDA methods and DAR methods is relatively smaller, which implies the SOTA DAR", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "COD: Learning Conditional Invariant Representation for DAR", + "bbox": [ + 313, + 113, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/dd667930f755c59575616a898b8275497cd80fbf94fce8f1ccd67d1f48e9d966.jpg", + "table_caption": [ + "Table 1: Domain adaptation regression results (MAE) on dSprites (ResNet-18)." + ], + "table_footnote": [], + "table_body": "
MethodC→NC→SN→CN→SS→CS→NAvg.
Resnet-180.940.900.160.650.080.260.498
TCA0.940.870.190.660.100.230.498
MCD0.810.810.170.650.070.190.450
JDOT0.860.790.190.640.100.230.468
AFN1.000.960.160.620.080.320.523
DAN0.700.770.120.500.060.110.377
DANN0.470.460.160.650.050.100.315
RSD0.310.310.120.530.070.080.237
DARE-GRAM0.300.200.110.250.050.070.164
COD0.120.160.100.230.080.120.134
", + "bbox": [ + 222, + 170, + 781, + 333 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/19b8a74dd7a5312347cdcc535d0e3d9e68391376c35efdbe1903e0c01893d185.jpg", + "table_caption": [ + "Table 2: Domain adaptation regression results (MAE) on Biwi Kinect (ResNet-18)." + ], + "table_footnote": [], + "table_body": "
MethodM→FF→MAvg.
Resnet-180.290.380.335
TCA0.310.390.350
MCD0.310.370.340
JDOT0.290.390.340
AFN0.320.410.365
DAN0.280.370.325
DANN0.300.370.335
RSD0.260.300.280
DARE-GRAM0.230.290.260
COD0.200.210.205
", + "bbox": [ + 348, + 375, + 656, + 536 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "methods with marginal alignment indeed show limitations in the difficult transfer tasks. In such a scenario, the sufficiency of condition alignment is directly validated, where the MAE of COD is significantly lower than other SOTA methods. Thus, for real-world regression applications with complex shift, the COD model indeed shows superiority in both theoretical and methodological aspects. Comparison on MPI3D. In MPI3D, the resolution remains the same with dSprites, while the main challenge is the large sample size. In Tab. 3, the MAE of COD is competitive with the SOTA method DARE-GRAM in the first four tasks, while higher in the last two tasks. However, note that COD still achieves the second-best result in averaged MAE, and obtains the lowest MAE in task RL $\\rightarrow$ T. In conclusion, the overall MAE of COD on three benchmarks is still lower, which demonstrates COD is generally superior to SOTA methods.", + "bbox": [ + 212, + 566, + 784, + 747 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation study. The ablation experiment was conducted on the most challenging tasks from all three datasets: $\\mathbf{C} \\rightarrow \\mathbf{S}$ and $\\mathbf{N} \\rightarrow \\mathbf{S}$ from dSprites, $\\mathbf{T} \\rightarrow \\mathbf{R}\\mathbf{L}$ and $\\mathbf{T} \\rightarrow \\mathbf{RC}$ from MPI3D, and both $\\mathbf{M} \\rightarrow \\mathbf{F}$ and $\\mathbf{F} \\rightarrow \\mathbf{M}$ from Biwi Kinect. The major components of proposed method, i.e., MSE objective $\\mathcal{L}_{src}$ , KGW distance $d_{\\mathrm{KGW}}$ , COD distance $d_{\\mathrm{COD}}$ and modified COD $d_{\\mathrm{COD}_{mod}}$ , are evaluated, where the results are shown in Tab. 4.", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H. Yang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/09bd9b74dfbfd2d5c97a9adb6169e74e23fa80570bacc224b88f5e4bd09c8669.jpg", + "table_caption": [ + "Table 3: Domain adaptation regression results (MAE) on MPI3D (ResNet-18)." + ], + "table_footnote": [], + "table_body": "
MethodRL→RCRL→TRC→RLRC→TT→RLT→RCAvg.
Resnet-180.170.440.190.450.510.500.377
TCA0.170.420.190.420.500.500.373
MCD0.130.400.150.450.520.500.358
JDOT0.160.410.160.410.470.470.353
AFN0.180.450.200.460.530.530.390
DAN0.120.350.120.270.400.410.278
DANN0.090.240.110.410.480.370.283
RSD0.090.190.080.150.360.360.205
DARE-GRAM0.090.150.100.140.240.240.160
COD0.100.110.130.150.370.290.192
", + "bbox": [ + 212, + 171, + 789, + 333 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/4b61c75db04ce08a8763f1e0494ec7093bdf4978230509a588a1b5cd60e90a01.jpg", + "table_caption": [ + "Table 4: Ablation study on different metrics for DAR." + ], + "table_footnote": [], + "table_body": "
MSEObjectivesdSpritesBiwi KinectMPI3D
dKGWdCODdCODmodC→SN→ST→RLT→RCM→FF→M
✓0.900.650.290.380.510.50
✓✓0.180.260.210.250.410.34
✓✓0.260.450.210.220.400.31
✓✓✓0.170.250.210.220.410.29
✓✓✓0.160.230.200.210.370.29
", + "bbox": [ + 218, + 377, + 784, + 487 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "From the $2^{\\mathrm{nd}}$ and $3^{\\mathrm{rd}}$ rows, it can be observed that both OT-based metrics show promising results on DAR datasets, where the conditional alignment with COD shows better performance than marginal alignment with KGW on MPI3D and Biwi Kinect datasets. On the other hand, since dSprites induce more difficult transfer problems, the baseline performance is generally lower, and the pseudo labels for conditional alignment are less reliable. Therefore, the combination of KGW and COD is necessary to ensure higher performance on challenging tasks, where the results in the $4^{\\mathrm{th}}$ row validate the mutually beneficial relation between these two objectives. Finally, the $5^{\\mathrm{th}}$ row shows the effectiveness of rethinking the discriminability of COD, which implies the modifications on the moment statistics indeed ensure lower MAE values on all datasets. Overall, these results demonstrate that conditional alignment is indeed crucial for successful DAR, and the theoretical results and proposed method are generally valid.", + "bbox": [ + 212, + 520, + 787, + 717 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Feature Visualization. t-SNE [34] is employed to visualize the 2-D features of different alignments on dSprites $\\mathrm{C}\\rightarrow \\mathrm{N}$ task. (a) Before adaptation, distributions of source and target domain differ severely, as shown in Fig. 3(a), which implies predictor is infeasible for the target samples with a significant domain gap. (b) After marginal distribution alignment via the SOTA DARE-GRAM method, source and target representations are aligned globally, and the cross-domain distributions are overlapped on the whole. However, there are still misaligned local patterns where source and target samples of different colors are matched.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "COD: Learning Conditional Invariant Representation for DAR", + "bbox": [ + 313, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/afb46b230699af1112d79d9ece81ccaee0e0e8721d8ce341d159b9a7ff6c1c47.jpg", + "image_caption": [ + "Fig. 3: t-SNE visualization of learned representations. Label values are denoted by color gradients. Eight label values are selected from the range of the variable for visualization. $+$ : source samples; $o$ : target samples." + ], + "image_footnote": [], + "bbox": [ + 218, + 150, + 781, + 260 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Therefore, it is hard to fit clear level sets in Fig. 3 (b). (c) After conditional distribution alignment via COD, source and target distributions are aligned not only globally but also locally. As shown in Fig. 3 (c), representations are matched according to label values so the change of color shows a clear gradient and level", + "bbox": [ + 212, + 344, + 787, + 406 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "sets suitable for both domains can be fitted.", + "bbox": [ + 215, + 406, + 535, + 419 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Hyper-parameter. We investigate the sensitivity of hyper-parameters on dSprites $\\mathrm{C}\\rightarrow \\mathrm{N}$ MAE values by varying $\\lambda_{1}$ and $\\lambda_{2}$ ,i.e., parameters of modified COD and KGW, are shown in Fig. 4. The results show that the performance of the COD-based DAR model is generally robust to the different choice of parameters. In fact, the MAE values of the COD-based model are stable in random experiments, which implies the proposed method is both robust from the view of randomness and parameter setting. In conclusion, the sta", + "bbox": [ + 212, + 420, + 542, + 601 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/fde948489ae7c2d8f5bd51ef3ce4e356503a10d0742673fe58a23006019ab258.jpg", + "image_caption": [ + "Fig.4: MAE values under different settings of hyper-parameters." + ], + "image_footnote": [], + "bbox": [ + 555, + 420, + 782, + 551 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "ble performance validates that the proposed model is also empirically reliable.", + "bbox": [ + 212, + 601, + 777, + 618 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 641, + 359, + 657 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we studied the limitations of existing DAR research, i.e., the misaligned intrinsic discriminant structure and negative transfer of marginal distribution matching framework. To deal with the challenges of conditional shift with continuous output space, we provide feasible solutions from both theoretical and methodological views, where the main results show that: 1) the cross-domain generalization regression error can be sufficiently bounded by the conditional discrepancy on continuous labels; 2) the proposed COD serves as a well-defined metric for conditional discrepancy measures and empirical modeling. With these guarantees, a COD-based DAR method is proposed with a discriminability enhancement mechanism. Numerical validations on standard DAR datasets validate the superiority of COD.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H. Yang et al.", + "bbox": [ + 271, + 114, + 366, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported in part by National Natural Science Foundation of China (Grant No. 62376291), in part by Guangdong Basic and Applied Basic Research Foundation (2023B1515020004), in part by Science and Technology Program of Guangzhou (2024A04J6413), in part by the Fundamental Research Funds for the Central Universities, Sun Yat-sen University (24xkjc013), in part by Guangdong Province Key Laboratory of Computational Science at Sun Yat-sen University (2020B1212060032), and in part by Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education.", + "bbox": [ + 212, + 176, + 787, + 297 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 321, + 321, + 338 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bao, Y., Liu, Y., Wang, H., Lu, F.: Generalizing gaze estimation with rotation consistency. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4207-4216 (2022)", + "2. Ben-David, S., Blitzer, J., Crammer, K., Kulesza, A., Pereira, F., Vaughan, J.W.: A theory of learning from different domains. Mach. Learn. 79, 151-175 (2010)", + "3. Chen, L., Chen, H., Wei, Z., Jin, X., Tan, X., Jin, Y., Chen, E.: Reusing the task-specific classifier as a discriminator: Discriminator-free adversarial domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7181-7190 (2022)", + "4. Chen, X., Wang, S., Wang, J., Long, M.: Representation subspace distance for domain adaptation regression. In: Int. Conf. Mach. Learn. pp. 1749-1759 (2021)", + "5. Chen, Y., Song, S., Li, S., Wu, C.: A graph embedding framework for maximum mean discrepancy-based domain adaptation algorithms. IEEE Trans. Image Process. 29, 199-213 (2019)", + "6. Tachet des Combes, R., Zhao, H., Wang, Y.X., Gordon, G.J.: Domain adaptation with conditional distribution matching and generalized label shift. In: Adv. Neural Inform. Process. Syst. vol. 33, pp. 19276-19289 (2020)", + "7. Cortes, C., Mohri, M.: Domain adaptation in regression. In: ICALT. pp. 308-323 (2011)", + "8. Courty, N., Flamary, R., Habrard, A., Rakotomamonjy, A.: Joint distribution optimal transportation for domain adaptation. Adv. Neural Inform. Process. Syst. 30 (2017)", + "9. Courty, N., Flamary, R., Tuia, D.: Domain adaptation with regularized optimal transport. In: ECML PKDD. pp. 274-289 (2014)", + "10. Damodaran, B.B., Kellenberger, B., Flamary, R., Tuia, D., Courty, N.: Deepjdot: Deep joint distribution optimal transport for unsupervised domain adaptation. In: Eur. Conf. Comput. Vis. pp. 447-463 (2018)", + "11. Dhaini, M., Berar, M., Honeine, P., Van Exem, A.: Unsupervised domain adaptation for regression using dictionary learning. Knowl.-Based Syst. 267, 110439 (2023)", + "12. Fanelli, G., Dantone, M., Gall, J., Fossati, A., Van Gool, L.: Random forests for real time 3d face analysis. Int. J. Comput. Vis. **101**, 437–458 (2013)", + "13. Fukumizu, K., Bach, F.R., Jordan, M.I.: Kernel dimension reduction in regression. Ann. Stat. 37(4), 1871-1905 (2009)", + "14. Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., March, M., Lempitsky, V.: Domain-adversarial training of neural networks. J. Mach. Learn. Res. 17(59), 1-35 (2016)" + ], + "bbox": [ + 218, + 353, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "COD: Learning Conditional Invariant Representation for DAR", + "bbox": [ + 313, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Geng, X., Zhou, Z.H., Smith-Miles, K.: Automatic age estimation based on facial aging patterns. IEEE Trans. Pattern Anal. Mach. Intell. 29(12), 2234-2240 (2007)", + "16. Gondal, M.W., Wuthrich, M., Miladinovic, D., Locatello, F., Breidt, M., Volchkov, V., Akpo, J., Bachem, O., Scholkopf, B., Bauer, S.: On the transfer of inductive bias from simulation to the real world: a new disentanglement dataset. Adv. Neural Inform. Process. Syst. 32 (2019)", + "17. Gong, B., Shi, Y., Sha, F., Grauman, K.: Geodesic flow kernel for unsupervised domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2066-2073 (2012)", + "18. Goswami, S., Kontolati, K., Shields, M.D., Karniadakis, G.E.: Deep transfer operator learning for partial differential equations under conditional shift. Nat. Mach. Intell. 4(12), 1155-1164 (2022)", + "19. Higgins, I., Matthew, L., Pal, A., Burgess, C., Glorot, X., Botvinick, M., Mohamed, S., Lerchner, A.: beta-vae: Learning basic visual concepts with a constrained variational framework. In: Int. Conf. Learn. Represent. (2016)", + "20. Hoffman, J., Tzeng, E., Park, T., Zhu, J.Y., Isola, P., Saenko, K., Efros, A., Darrell, T.: Cycada: Cycle-consistent adversarial domain adaptation. In: Int. Conf. Mach. Learn. pp. 1989-1998 (2018)", + "21. Hu, J., Zhong, H., Yang, F., Gong, S., Wu, G., Yan, J.: Learning unbiased transferability for domain adaptation by uncertainty modeling. In: Eur. Conf. Comput. Vis. pp. 223-241 (2022)", + "22. Jiang, J., Ji, Y., Wang, X., Liu, Y., Wang, J., Long, M.: Regressive domain adaptation for unsupervised keypoint detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6780-6789 (2021)", + "23. Kirchmeyer, M., Rakotomamonjy, A., de Bezenac, E., Gallinari, P.: Mapping conditional distributions for domain adaptation under generalized target shift. In: Int. Conf. Learn. Represent. (2022)", + "24. Lee, C.Y., Batra, T., Baig, M.H., Ulbricht, D.: Sliced wasserstein discrepancy for unsupervised domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10285-10295 (2019)", + "25. Li, B., Wang, Y., Zhang, S., Li, D., Keutzer, K., Darrell, T., Zhao, H.: Learning invariant representations and risks for semi-supervised domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1104-1113 (2021)", + "26. Li, X., Li, J., Zuo, L., Zhu, L., Shen, H.T.: Domain adaptive remaining useful life prediction with transformer. IEEE Trans. Instrum. 71, 1-13 (2022)", + "27. Li, Z., Zhang, K., Liu, Y., Zou, Y., Ding, G.: A novel remaining useful life transfer prediction method of rolling bearings based on working conditions common benchmark. IEEE Trans. Instrum. 71, 1-9 (2022)", + "28. Long, M., Cao, Y., Wang, J., Jordan, M.: Learning transferable features with deep adaptation networks. In: Int. Conf. Mach. Learn. pp. 97-105 (2015)", + "29. Long, M., Cao, Z., Wang, J., Jordan, M.I.: Conditional adversarial domain adaptation. Adv. Neural Inform. Process. Syst. 31 (2018)", + "30. Long, M., Zhu, H., Wang, J., Jordan, M.I.: Deep transfer learning with joint adaptation networks. In: Int. Conf. Mach. Learn. pp. 2208-2217 (2017)", + "31. Luo, Y.W., Ren, C.X.: Conditional bures metric for domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 13989-13998 (2021)", + "32. Luo, Y.W., Ren, C.X.: Mot: Masked optimal transport for partial domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3531-3540 (2023)", + "33. Luo, Y.W., Ren, C.X., Dai, D.Q., Yan, H.: Unsupervised domain adaptation via discriminative manifold propagation. IEEE Trans. Pattern Anal. Mach. Intell. 44(3), 1653-1669 (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H. Yang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. J. Mach. Learn. Res. 9(11) (2008)", + "35. Mansour, Y., Mohri, M., Rostamizadeh, A.: Domain adaptation: Learning bounds and algorithms. arXiv preprint arXiv:0902.3430 (2009)", + "36. Mao, W., Ge, Y., Shen, C., Tian, Z., Wang, X., Wang, Z., den Hengel, A.v.: Poseur: Direct human pose regression with transformers. In: Eur. Conf. Comput. Vis. pp. 72-88 (2022)", + "37. de Mathelin, A., Richard, G., Deheeger, F., Mougeot, M., Vayatis, N.: Adversarial weighting for domain adaptation in regression. In: ICTAI. pp. 49-56 (2021)", + "38. Nejjar, I., Wang, Q., Fink, O.: Dare-gram: Unsupervised domain adaptation regression by aligning inverse gram matrices. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 11744–11754 (2023)", + "39. Nikzad-Langerodi, R., Zellinger, W., Saminger-Platz, S., Moser, B.A.: Domain adaptation for regression under beer-lambert's law. Knowl.-Based Syst. 210, 106447 (2020)", + "40. Pan, S.J., Tsang, I.W., Kwok, J.T., Yang, Q.: Domain adaptation via transfer component analysis. IEEE Trans. Neural Netw. 22(2), 199-210 (2010)", + "41. Pei, Z., Cao, Z., Long, M., Wang, J.: Multi-adversarial domain adaptation. In: AAAI. vol. 32 (2018)", + "42. Peng, R., Wang, R., Wang, Z., Lai, Y., Wang, R.: Rethinking depth estimation for multi-view stereo: A unified representation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8645-8654 (2022)", + "43. Quinonero-Candela, J., Sugiyama, M., Schwaighofer, A., Lawrence, N.D.: Dataset shift in machine learning. Mit Press (2008)", + "44. Ren, C.X., Luo, Y.W., Dai, D.Q.: Buresnet: Conditional bures metric for transferable representation learning. IEEE Trans. Pattern Anal. Mach. Intell. 45(4), 4198-4213 (2023)", + "45. Saito, K., Watanabe, K., Ushiku, Y., Harada, T.: Maximum classifier discrepancy for unsupervised domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3723-3732 (2018)", + "46. Shen, J., Qu, Y., Zhang, W., Yu, Y.: Wasserstein distance guided representation learning for domain adaptation. In: AAAI. vol. 32 (2018)", + "47. Singh, A., Chakraborty, S.: Deep domain adaptation for regression. Development and Analysis of Deep Learning Architectures pp. 91-115 (2020)", + "48. Song, L., Huang, J., Smola, A., Fukumizu, K.: Hilbert space embeddings of conditional distributions with applications to dynamical systems. In: Int. Conf. Mach. Learn. pp. 961-968 (2009)", + "49. Sun, B., Feng, J., Saenko, K.: Return of frustratingly easy domain adaptation. In: AAAI. vol. 30 (2016)", + "50. Tang, H., Jia, K.: Discriminative adversarial domain adaptation. In: AAAI. vol. 34, pp. 5940-5947 (2020)", + "51. Teshima, T., Sato, I., Sugiyama, M.: Few-shot domain adaptation by causal mechanism transfer. In: Int. Conf. Mach. Learn. pp. 9458-9469 (2020)", + "52. Tzeng, E., Hoffman, J., Saenko, K., Darrell, T.: Adversarial discriminative domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7167-7176 (2017)", + "53. Wang, H., Shen, T., Zhang, W., Duan, L.Y., Mei, T.: Classes matter: A fine-grained adversarial approach to cross-domain semantic segmentation. In: Eur. Conf. Comput. Vis. pp. 642-659 (2020)", + "54. Wang, W., Li, H., Ding, Z., Nie, F., Chen, J., Dong, X., Wang, Z.: Rethinking maximum mean discrepancy for visual domain adaptation. IEEE Trans. Neural Netw. Learn. Syst. (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "COD: Learning Conditional Invariant Representation for DAR", + "bbox": [ + 315, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "55. Wang, Y., Jiang, Y., Li, J., Ni, B., Dai, W., Li, C., Xiong, H., Li, T.: Contrastive regression for domain adaptation on gaze estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19376-19385 (2022)", + "56. Wei, P., Ke, Y., Ong, Y.S., Ma, Z.: Adaptive transfer kernel learning for transfer gaussian process regression. IEEE Trans. Pattern Anal. Mach. Intell. (2022)", + "57. Wu, J., He, J., Wang, S., Guan, K., Ainsworth, E.: Distribution-informed neural networks for domain adaptation regression. Adv. Neural Inform. Process. Syst. 35, 10040-10054 (2022)", + "58. Xu, R., Li, G., Yang, J., Lin, L.: Larger norm more transferable: An adaptive feature norm approach for unsupervised domain adaptation. In: Int. Conf. Comput. Vis. pp. 1426-1435 (2019)", + "59. Zhang, Z., Wang, M., Nehorai, A.: Optimal transport in reproducing kernel hilbert spaces: Theory and applications. IEEE Trans. Pattern Anal. Mach. Intell. 42(7), 1741-1754 (2019)", + "60. Zhao, H., Combes, R.T.D., Zhang, K., Gordon, G.: On learning invariant representations for domain adaptation. In: Int. Conf. Mach. Learn. pp. 7523-7532 (2019)", + "61. Zhao, H., Coston, A., Adel, T., Gordon, G.J.: Conditional learning of fair representations. In: Int. Conf. Learn. Represent. (2020)", + "62. Zhao, H., Dan, C., Aragam, B., Jaakkola, T.S., Gordon, G.J., Ravikumar, P.: Fundamental limits and tradeoffs in invariant representation learning. J. Mach. Learn. Res. 23(340), 1-49 (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 438 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "H. Yang et al.", + "bbox": [ + 271, + 114, + 366, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/COD_ Learning Conditional Invariant Representation for Domain Adaptation Regression/5f47c561-6103-4c2b-b6da-ac9368ec9383_model.json b/2024/COD_ Learning Conditional Invariant Representation for Domain Adaptation Regression/5f47c561-6103-4c2b-b6da-ac9368ec9383_model.json new file mode 100644 index 0000000000000000000000000000000000000000..39690c677e9a2c4a728f3a512f6aa926370e02ee --- /dev/null +++ b/2024/COD_ Learning Conditional Invariant Representation for Domain Adaptation Regression/5f47c561-6103-4c2b-b6da-ac9368ec9383_model.json @@ -0,0 +1,2546 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.279, + 0.141, + 0.726, + 0.21 + ], + "angle": 0, + "content": "COD: Learning Conditional Invariant Representation for Domain Adaptation Regression" + }, + { + "type": "text", + "bbox": [ + 0.283, + 0.235, + 0.721, + 0.252 + ], + "angle": 0, + "content": "Hao-Ran Yang1, Chuan-Xian Ren1*, and You-Wei Luo1,2" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.263, + 0.722, + 0.306 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Mathematics, Sun Yat-Sen University, China \n\\(^{2}\\) School of Mathematics, Jiaying University, China \n{yanghr26@mail2, rchuanx@mail, luoyw28@mail2}.sysu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.346, + 0.74, + 0.61 + ], + "angle": 0, + "content": "Abstract. Aiming to generalize the label knowledge from a source domain with continuous outputs to an unlabeled target domain, Domain Adaptation Regression (DAR) is developed for complex practical learning problems. However, due to the continuity problem in regression, existing conditional distribution alignment theory and methods with discrete prior, which are proven to be effective in classification settings, are no longer applicable. In this work, focusing on the feasibility problems in DAR, we establish the sufficiency theory for the regression model, which shows the generalization error can be sufficiently dominated by the cross-domain conditional discrepancy. Further, to characterize conditional discrepancy with continuous conditioning variable, a novel Conditional Operator Discrepancy (COD) is proposed, which admits the metric property on conditional distributions via the kernel embedding theory. Finally, to minimize the discrepancy, a COD-based conditional invariant representation learning model is proposed, and the reformulation is derived to show that reasonable modifications on moment statistics can further improve the discriminability of the adaptation model. Extensive experiments on standard DAR datasets verify the validity of theoretical results and the superiority over SOTA DAR methods." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.623, + 0.741, + 0.651 + ], + "angle": 0, + "content": "Keywords: Domain Adaptation Regression, Optimal Transport, Invariant Representation Learning, Kernel Embedding Theory." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.679, + 0.377, + 0.695 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.712, + 0.788, + 0.818 + ], + "angle": 0, + "content": "Deep learning has shown a promising ability for fitting and achieved great success in various application tasks. However, large-scale data with sufficient annotations plays a vital role in training high-quality deep models while the label annotations for unlabeled data are usually expensive. Therefore, aiming to reduce training costs by transferring knowledge, Unsupervised Domain Adaptation (UDA) [2,14,17,28] has received increasing attention in recent years. Basically, UDA aims to transfer task-related information from a labeled source domain" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.386, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "H. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.435 + ], + "angle": 0, + "content": "to a target domain without annotations, where the domain gap is characterized as the distribution shift across domains [43]. Following this line, considerable methods have been proposed to alleviate the UDA problem by shift correction and distribution alignment, such as adversarial training for cross-domain samples [14, 52] and explicitly distribution matching by statistical discrepancy minimization [28, 49]. From a theoretical perspective, recent literature [6, 60] show that the marginal distribution \\( P_{X} \\) with covariate shift assumption is insufficient to characterize the cross-domain generalization error, while the conditional distribution alignment indeed serves as a sufficient condition to ensure small error and successful transfer. Mathematically, when the conditional distributions \\( P_{X|y} \\) for every \\( y \\) are aligned, a well-trained source predictor can perform well on the target domain, which is described in Fig. 1. Inspired by theoretical advances, recent works [23, 41, 45] consider the statistical methods for conditional alignment, and show promising performance in mitigating negative transfer and improving recognition accuracy. However, most of the existing results are designed for classification settings with discrete prior on label variables, which induces the infeasibility for regression problems. Consequently, it restricts the application of UDA methods for wider real-world problems, e.g., human pose estimation [36], product condition monitoring [27], depth estimation [42]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.788, + 0.698 + ], + "angle": 0, + "content": "Considering the dilemma of UDA, pioneer works have explored the Domain Adaptation Regression (DAR) framework for solving the distribution shift in continuous label scenarios, where theoretical results [7,35] and correction methods [15,47,51] are provided. To further relax the requirement of labeled training target data in pioneer DAR works, advanced works [4,38] propose unsupervised DAR methods to learn transferable regression models without the help of labeled target data. Nonetheless, existing DAR theory and methods are directly facilitated by the covariate shift framework, where the marginal distribution alignment is considered. Therefore, it inevitably induces the negative transfer and misaligned local structure, which are shown in the results in classification settings [6,60]. In fact, there is a significant obstacle for modeling conditional distribution alignment in regression scenarios, i.e., the conditional shift correction induces infinite matching problems for \\( P_{X|Y=y} \\) on all points \\( y \\). The continuous label variable with infinite possible outcomes makes it infeasible to directly apply the existing conditional alignment methods for DAR. Thus, it is vital to understand the continuity problem in regression, and develop the generalization error theory and statistical methodology for DAR scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this work, we focus on the theory guarantee and conditional alignment methodology for successful DAR. Theoretically, we explore the generalization upper bound for continuous output scenarios and show the cross-domain joint error can be sufficiently upper-bounded by the conditional discrepancy. It ensures the sufficiency of conditional alignment still holds for regression scenarios and provides reliable guarantees for theory-driven modeling. Methodologically, to deal with the continuous problem in conditional alignment, we treat the conditional distribution as finite statistical moments in Hilbert space. By exploring the distribution embedding theory of Reproducing Kernel Hilbert Space (RKHS), a novel" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.315, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "COD: Learning Conditional Invariant Representation for DAR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.324, + 0.145, + 0.684, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.172, + 0.416, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.439, + 0.172, + 0.582, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.174, + 0.728, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.289, + 0.788, + 0.428 + ], + "angle": 0, + "content": "Fig. 1: Illustration of the conditional shift in regression setting, where label value \\( y \\in [0,1] \\) and color gradients imply the continuity. (a) Before adaptation, distribution shift exists between source and target representations so the predictor trained on source domain cannot generalize to target domain. (b) After marginal distribution alignment, distributions of source and target domain are globally aligned. However, representations with different labels may be falsely aligned across domains, leading to significant inconsistency between cross-domain labeling rules. Thus, level sets provided by the source predictor are not suitable for target representations. (c) After conditional distribution alignment, label-wise matching is achieved, where the level sets of cross-domain labeling rules are consistent and the source predictor will be reliable." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.579 + ], + "angle": 0, + "content": "conditional metric called Conditional Operator Discrepancy (COD) is proposed, which can be interpreted as the conditional Wasserstein distance in RKHS. With the sufficiency theory and metric-based learning principle for regression, a COD-based conditional invariant representation learning method is proposed, which aims to correct the conditional shift on the pushforward measures in latent representation space. By reformulating the empirical COD estimation, a modified principle is developed to further improve the discriminability during knowledge transfer. Generally, our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.589, + 0.785, + 0.633 + ], + "angle": 0, + "content": "- The sufficiency theory for the conditional shift framework is established for DAR settings, which shows the cross-domain conditional discrepancy on continuous label variable serves as the major term in generalization error." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.634, + 0.785, + 0.679 + ], + "angle": 0, + "content": "- A novel COD metric for characterizing conditional distribution discrepancy with continuous and infinite conditioning variable is explored, providing a feasible solution for conditional discrepancy optimization in DAR." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.68, + 0.785, + 0.724 + ], + "angle": 0, + "content": "- A COD-based representation learning model with reformulation on moment statistics is proposed, which admits discriminability on task-related knowledge and shows superior performance in extensive experiments." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.589, + 0.785, + 0.724 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.748, + 0.388, + 0.763 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Domain Adaptation Classification. Based on the distribution shift assumption, most of the UDA methods can be roughly categorized into three types: methods that align marginal distribution, conditional distribution and joint distribution. Most UDA methods consider only the marginal distribution align-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.367, + 0.129 + ], + "angle": 0, + "content": "H. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.388 + ], + "angle": 0, + "content": "ment. Discrepancy-based methods attempt to directly minimize the distribution discrepancy and different types of discrepancy measurement greatly enrich this type of methodology. For example, these methods adopt moment matching [28, 49] or extension under manifold framework [17, 33]. In particular, Optimal transport (OT) and Wasserstein distance are introduced to UDA as a sound measurement for domain shift, based on which a large number of methods have been proposed [8-10, 24, 46]. Zhang et al. [59] further develop both theory and methodology to RKHS and Luo et al. [31] extends Kernel Bures metric to conditional distribution alignment. Adversarial-based methods [14, 20, 21] learn domain-invariant features via two-player games. Specifically, they alternatively optimize the feature generator and the domain discriminator to achieve domain confusion. In order to obtain better performance, advanced methods have explored how to utilize the label information to achieve conditional distribution alignment [3, 32, 41, 44, 45, 50, 53, 61] or joint distribution alignment [29, 30]. However, UDA methods for shift correction are proposed for discrete and finite conditioning variables, which are inapplicable for regression." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.789, + 0.707 + ], + "angle": 0, + "content": "Domain Adaptation Regression. DA for regression has received relatively little attention compared with classification. Mansour et al. [35] and Cortes et al. [7] introduce theoretical analysis for DA in regression scenarios. Zhao et al. [62] discuss the tradeoffs between accuracy and invariance in learning invariant representations in both classification and regression settings. A range of methods in the shallow regime are proposed to tackle the DAR problem. These methods can be roughly categorized into two types: importance weighting [15,37] and learning domain-invariant representation [8,39,40]. Besides, some attempts have been made in deep regime to tackle regression problem [25,47,51,56]. Unfortunately, most of them belong to semi-supervised DA, which means they need to use a small number of target domain labels. Recently, deep UDA methods for regression that get rid of target label are proposed [4,11,38,57]. RSD [4] states that deep neural networks are less robust to feature scaling in regression problems and aligning the deep features directly will impede task performance for altering feature scale. Thus, a strategy of aligning orthogonal bases of both domains is adopted in RSD. Based on RSD, DARE-GRAM [38] aligns the inverse Gram Matrix of deep features of both domains, and ensures promising improvements in accuracy and stability. Meanwhile, several DAR methods for specific tasks have been explored [1,18,22,26,27,55]. However, the theory and method for conditional shift correction are still unexplored for DAR scenarios, which hinders the in-depth reasoning and development for practical modeling." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.727, + 0.564, + 0.746 + ], + "angle": 0, + "content": "3 COD: Theory and Methodology" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.757, + 0.362, + 0.773 + ], + "angle": 0, + "content": "3.1 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Setup. In this paper, we discuss UDA method dealing with regression problems. Let \\( \\mathrm{X} \\) and \\( \\mathrm{Y} \\) be covariate and label variable taken values on \\( \\mathcal{X} \\) and \\( \\mathcal{Y} \\). \\( \\mathbb{P} \\) denotes the set of distributions and subscripts represent the corresponding random variable while the superscript represents the domain, e.g. \\( P_{XY}^{s} \\in \\mathbb{P}_{\\mathcal{X}\\mathcal{Y}} \\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.315, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "COD: Learning Conditional Invariant Representation for DAR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "implies the joint distribution on source domain. In UDA settings, we are provided with labeled source samples \\((x^{s},y^{s})\\sim P_{XY}^{s}\\) and unlabeled target samples \\((x^{t},y^{t})\\sim P_{XY}^{t}\\), which means the target labels are not available during training. Samples of source and target domain follow different distributions, i.e. \\(P_{XY}^{s}\\neq P_{XY}^{t}\\). Denote \\(g:X\\to Z\\) as the representation extractor, \\(h:Z\\to Y\\) as the predictor, \\(\\varepsilon_{s}\\) and \\(\\varepsilon_{t}\\) as the source and target prediction error, respectively. UDA methods aim to generalize the learner \\(h\\circ g\\) across domains and minimize target risk \\(\\varepsilon_{t}\\) by mitigating the distribution shift problem. In contrast to the discrete label space in classification settings, unsupervised DAR discusses UDA problems with continuous labels, which means the number of classes is uncountable, leading to great difficulty in aligning conditional distribution \\(P_{X|Y}\\). The proofs are provided in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.434 + ], + "angle": 0, + "content": "OT in RKHS. Let \\((\\mathcal{X},\\mathcal{B})\\) be a measure space with Borel \\(\\sigma\\)-field \\(\\mathcal{B}\\). Denote \\((\\mathcal{H}_{\\mathcal{X}},k_{\\mathcal{X}})\\) as the RKHS of \\(\\mathcal{X}\\), which is generated by the positive definite kernel \\(k_{\\mathcal{X}}\\). Elements from \\(\\mathcal{X}\\) are mapped to \\(\\mathcal{H}_{\\mathcal{X}}\\) via feature map \\(\\phi (x) = k_{\\mathcal{X}}(x,\\cdot)\\), which is assumed to satisfy the reproducing properties \\(\\langle \\phi (x),\\phi (x^{\\prime})\\rangle_{\\mathcal{H}_{\\mathcal{X}}} = k_{\\mathcal{X}}(x,x^{\\prime})\\) and \\(\\langle \\phi (x),f\\rangle_{\\mathcal{H}_x} = f(x),\\forall f\\in \\mathcal{H}_x\\). The mean element \\(\\mu_X\\) in \\(\\mathcal{H}_{\\mathcal{X}}\\) with law \\(P_{X}\\) is given by \\(\\mu_{X} = \\mathbb{E}_{X}[\\phi (X)]\\), and the covariance operator \\(\\mathcal{C}_{XX}\\) is defined as \\(\\mathcal{C}_{XX} = E_X(\\phi (X) - \\mu_X)\\otimes (\\phi (X) - \\mu_X)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.787, + 0.466 + ], + "angle": 0, + "content": "For any two distributions \\(P_X^s, P_X^t \\in \\mathbb{P}_X\\), the Kantorovitch optimal transport in RKHS \\(\\mathcal{H}_{\\mathcal{K}}\\) is considered on pushforward measures \\(\\phi_{\\#}P_X^s\\) and \\(\\phi_{\\#}P_X^t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.24, + 0.478, + 0.786, + 0.51 + ], + "angle": 0, + "content": "\\[\nd _ {\\mathrm {K W}} ^ {2} \\left(X ^ {s}, X ^ {t}\\right) = \\inf _ {\\pi_ {\\mathcal {K}} \\in \\Pi \\left(\\phi_ {\\#} P _ {X} ^ {s}, \\phi_ {\\#} P _ {X} ^ {t}\\right)} \\int_ {\\mathcal {H} _ {\\mathcal {K}} \\times \\mathcal {H} _ {\\mathcal {K}}} \\left\\| X ^ {s} - X ^ {t} \\right\\| _ {\\mathcal {H} _ {\\mathcal {K}}} ^ {2} d \\pi_ {\\mathcal {K}} \\left(X ^ {s}, X ^ {t}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.52, + 0.787, + 0.551 + ], + "angle": 0, + "content": "Zhang et al. [59] show that if the pushforward measures are Gaussian, the Kernel Gaussian Wasserstein (KGW) distance between them can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.292, + 0.563, + 0.786, + 0.581 + ], + "angle": 0, + "content": "\\[\nd _ {\\mathrm {K G W}} ^ {2} \\left(X ^ {s}, X ^ {t}\\right) = \\| u _ {X} ^ {s} - u _ {X} ^ {t} \\| _ {\\mathcal {H} _ {\\kappa}} ^ {2} + t r \\left(\\mathcal {C} _ {X X} ^ {s s} + \\mathcal {C} _ {X X} ^ {t t} - 2 \\mathcal {C} _ {X X} ^ {s t}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.787, + 0.634 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}_{XX}^{st} = \\sqrt{\\sqrt{\\mathcal{C}_{XX}^{ss}}\\mathcal{C}_{XX}^{tt}\\sqrt{\\mathcal{C}_{XX}^{ss}}}\\) and \\(\\mathcal{C}_{XX}^{ss},\\mathcal{C}_{XX}^{tt}\\) are the covariance operators of \\(P_{X}^{s}\\) and \\(P_{X}^{t}\\) on \\(\\mathcal{H}_{\\mathcal{K}}\\) , respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.64, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Remark 1. Note that KGW consists of two terms, i.e., the Maximum Mean Discrepancy (MMD) on first-order statistics and kernel Bures metric on second-order statistics. It provides a basic innovation for exploring conditional discrepancy metric by considering the conditional statistics in RKHS, i.e., first-order statistics in literature [48] and second-order statistics in literature [13]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.739, + 0.587, + 0.755 + ], + "angle": 0, + "content": "3.2 Generalization Error Analysis for DAR" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Most of DAR methods focus on aligning the marginal distribution \\( P_{X} \\). However, Zhao et al. [60] have shown that in the presence of label shift, only aligning the marginal distribution does not guarantee satisfying generalization performance on target domain. In fact, [6, 60] have proven the following theorem (1) for classification problems:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.367, + 0.129 + ], + "angle": 0, + "content": "H. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.179 + ], + "angle": 0, + "content": "Theorem 1 ([60]). For representation variable \\( Z = g(X) \\), suppose inequality \\( d_{JS}(P_Y^s, P_Y^t) \\geq d_{JS}(P_Z^s, P_Z^t) \\) holds, then for predictor \\( h: Z \\to Y \\), we have" + }, + { + "type": "equation", + "bbox": [ + 0.295, + 0.19, + 0.786, + 0.226 + ], + "angle": 0, + "content": "\\[\n\\varepsilon_ {s} (h \\circ g) + \\varepsilon_ {t} (h \\circ g) \\geq \\frac {1}{2} \\left(\\sqrt {d _ {J S} (P _ {Y} ^ {s} , P _ {Y} ^ {t})} - \\sqrt {d _ {J S} (P _ {Z} ^ {s} , P _ {Z} ^ {t})}\\right) ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.234, + 0.585, + 0.25 + ], + "angle": 0, + "content": "where \\(d_{JS}\\) denotes the Jensen-Shannon divergence." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.258, + 0.784, + 0.333 + ], + "angle": 0, + "content": "Assuming the label distributions differ between source and target domain, i.e., \\( d_{JS}(P_Y^s,P_Y^t) > 0 \\), the lower bound shows that good alignment of representation distributions leads to bad joint error. In other words, when label shift exists, which is rather common in realistic scenarios, only aligning the marginal distribution is not enough for low target error." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.334, + 0.785, + 0.409 + ], + "angle": 0, + "content": "To deal with the insufficiency problem shown in the lower bound analysis, it is intuitive to introduce label information into generalization analysis, e.g., the balanced error rate for classification scenarios [6]. Analogously, for the continuous regression scenarios, we first define the continuous balanced error rate for quantifying maximum conditional risk in DAR." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.418, + 0.786, + 0.448 + ], + "angle": 0, + "content": "Definition 1. The continuous balanced error rate of predictor \\( h \\) on distribution \\( P \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.33, + 0.46, + 0.786, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\operatorname {c B E R} _ {\\mathrm {P}} (h, Y) := \\sup _ {(x, y) \\in \\mathcal {X} \\oplus \\mathcal {Y}} P (h (x) \\neq Y \\mid Y = y). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.497, + 0.784, + 0.556 + ], + "angle": 0, + "content": "Intuitively, the continuous balanced error rate can be regarded as the maximum potential risk for predicting sample \\( x \\) with label \\( y \\). Based on cBER and conditional shift correction with representation transformation \\( Z = g(X) \\), a sufficient condition for DAR can be induced as follows." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.565, + 0.784, + 0.596 + ], + "angle": 0, + "content": "Theorem 2. For representation variable \\( Z^s = g(X^s) \\) and \\( Z^t = g(X^t) \\), if \\( g(\\cdot) \\) satisfies conditional invariant property, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.404, + 0.608, + 0.785, + 0.625 + ], + "angle": 0, + "content": "\\[\nP _ {Z \\mid Y = y} ^ {s} = P _ {Z \\mid Y = y} ^ {t}, \\quad \\forall y \\in \\mathcal {Y}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.635, + 0.549, + 0.65 + ], + "angle": 0, + "content": "then for any predictor \\( h: Z \\to Y \\), we have" + }, + { + "type": "equation", + "bbox": [ + 0.39, + 0.664, + 0.785, + 0.679 + ], + "angle": 0, + "content": "\\[\n\\varepsilon_ {s} (h) + \\varepsilon_ {t} (h) \\leq 2 \\mathrm {c B E R} _ {\\mathrm {P s}} (h, Y). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.69, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Therefore, the key for DA, no matter for classification or regression problems, is to achieve conditional distribution alignment. In classification settings, conditional alignment can be achieved by conducting class-wise marginal distribution alignment with the help of target pseudo label. However, in regression scenarios, the label variable is continuous, which means previous methods correcting conditional shift in classification settings are no longer applicable. Obviously, it is impossible to align the \\( P_{X|y} \\) for every single \\( y \\) when \\( y \\) is a continuous variable and the possible number of \\( y \\) is infinite. Thus, it is important and meaningful to explore a new framework that is capable of conditional distribution alignment for continuous label variable." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.315, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "COD: Learning Conditional Invariant Representation for DAR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.145, + 0.77, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.28, + 0.788, + 0.406 + ], + "angle": 0, + "content": "Fig. 2: Illustration of COD metric. (a) Classification DA methods usually correct conditional shift by class-wise computations on discrete clusters \\( P_{X|y} \\), which is infeasible for regression due to the infinite slices \\( P_{X|y} \\). (b) In COD, the continuous conditional distributions are embedded into the RKHS and characterized by finite statistical moments in kernel spaces, which do not rely on specific conditions \\( y \\). Under the guarantees of distribution embedding property, the conditional alignment is equivalent to the matching on first-order and second-order statistics, i.e., mean embedding operator \\( \\mathcal{U}_{X|Y} \\) and covariance operator \\( \\mathcal{C}_{XX|Y} \\). When COD is minimized, conditional alignment is achieved, and the level sets of cross-domain labeling rules are then aligned." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.432, + 0.442, + 0.446 + ], + "angle": 0, + "content": "3.3 COD metric for DAR" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.454, + 0.788, + 0.546 + ], + "angle": 0, + "content": "Theory for COD. Due to the continuity of label variable, it is impractical to correct conditional shift by aligning marginal distribution of \\( P_{X|y} \\) for every single \\( y \\). For this inevitable obstacle, an intuitive idea is can we align all \\( P_{X|y} \\) as a whole? i.e., \\( P_{X|Y} \\), rather than align \\( P_{X|y} \\) one by one? It is more than attractive that if we can characterize \\( P_{X|Y} \\) with a well-defined discrepancy, then all \\( P_{X|y} \\) can be aligned by aligning the whole \\( P_{X|Y} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.788, + 0.682 + ], + "angle": 0, + "content": "In kernel embedding theory, a distribution can be characterized by a point in RKHS. Then \\( P_{X|Y} \\) can be seen as a point set in RKHS in which every element corresponds to a distribution \\( P_{X|y} \\) for a fixed \\( y \\). From another perspective, we can consider \\( P_{X|Y} \\) as the distribution in RKHS and measure the cross-domain conditional discrepancy by defining the COD distance on their statistical moments in RKHS. Following the definitions of conditional statistics in pioneer works, i.e., the conditional mean operator as \\( \\mathcal{U}_{X|Y} \\coloneqq \\mathcal{C}_{XY}\\mathcal{C}_{YY}^{-1} \\) [48] and conditional covariance operator as \\( \\mathcal{C}_{XX|Y} = \\mathcal{C}_{XX} - \\mathcal{C}_{XY}\\mathcal{C}_{YY}^{-1}\\mathcal{C}_{YX} \\) [13], we present the definition of COD as follows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.788, + 0.736 + ], + "angle": 0, + "content": "Definition 2 (COD). Given conditional distributions \\( P_{X|Y}^{s} \\) and \\( P_{X|Y}^{t} \\), denote \\( \\mathcal{C}_{XX|Y}^{ss} \\) and \\( \\mathcal{C}_{XX|Y}^{tt} \\) as their corresponding covariance operators. The COD is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.742, + 0.786, + 0.775 + ], + "angle": 0, + "content": "\\[\nd _ {\\mathrm {C O D}} ^ {2} \\left(P _ {X | Y} ^ {s}, P _ {X | Y} ^ {t}\\right) = \\left\\| \\mathcal {U} _ {X | Y} ^ {s} - \\mathcal {U} _ {X | Y} ^ {t} \\right\\| _ {\\mathcal {H} _ {\\mathcal {K}}} ^ {2} + \\operatorname {t r} \\left(\\mathcal {C} _ {X X | Y} ^ {s s} + \\mathcal {C} _ {X X | Y} ^ {t t} - 2 \\mathcal {C} _ {X X | Y} ^ {s t}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.775, + 0.531, + 0.807 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}_{XX|Y}^{st} = \\sqrt{\\sqrt{\\mathcal{C}_{XX|Y}^{ss}}\\mathcal{C}_{XX|Y}^{tt}\\sqrt{\\mathcal{C}_{XX|Y}^{ss}}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Theoretically, it can be proved that COD indeed admits the metric property on conditional distributions, i.e., two conditional distributions are the same if" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.367, + 0.129 + ], + "angle": 0, + "content": "H. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.178 + ], + "angle": 0, + "content": "and only if the COD metric between them is zero. An intuitive illustration of COD is shown in Fig. 2. Formally, COD admits two key properties." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.184, + 0.733, + 0.198 + ], + "angle": 0, + "content": "Proposition 1. (a) COD defines a metric between Gaussian measures." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.199, + 0.787, + 0.247 + ], + "angle": 0, + "content": "(b) Let \\((\\mathcal{X},\\mathcal{B}_{\\mathcal{X}})\\) be the locally compact and Hausdorff measurable space and \\(k\\) be \\(c_{0}\\)-universal kernel. Assuming that \\((\\phi (X),\\psi (Y))\\) is a Gaussian random variable in \\(\\mathcal{H}_{\\mathcal{X}}\\oplus \\mathcal{H}_{\\mathcal{Y}}\\). For any \\(P_{X|Y}^{s},P_{X|Y}^{t}\\in \\mathbb{P}_{\\mathcal{X}|\\mathcal{Y}}\\), we have" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.257, + 0.786, + 0.276 + ], + "angle": 0, + "content": "\\[\nd _ {\\mathrm {C O D}} ^ {2} \\left(P _ {X | Y} ^ {s}, P _ {X | Y} ^ {t}\\right) = 0 \\Rightarrow P _ {X | Y} ^ {s} = P _ {X | Y} ^ {t}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.784, + 0.342 + ], + "angle": 0, + "content": "Empirical Estimation of COD. Based on the above properties, we can perform conditional distribution alignment by minimizing the COD metric between source and target domain. To do this, we shall derive the empirical estimation of COD, which can be obtained with the help of the kernel trick." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.343, + 0.788, + 0.48 + ], + "angle": 0, + "content": "Let \\((\\mathbf{x}_i^s,\\mathbf{y}_i^s)_{i = 1}^{n_s}\\) and \\((\\mathbf{x}_i^t,\\mathbf{y}_i^t)_{i = 1}^{n_t}\\) be two set of samples drawn i.i.d. from source and target domain. For simplicity, \\(n_s\\) and \\(n_t\\) are both set to \\(n\\). In kernel method, samples are mapped to RKHS \\(\\mathcal{H}_{\\mathcal{X}}\\oplus \\mathcal{H}_{\\mathcal{Y}}\\) by the implicit feature map \\((\\phi ,\\psi)\\). Denote the feature map matrices by \\((\\Phi^s,\\Psi^s)\\) and \\((\\Phi^t,\\Psi^t)\\), the explicit kernel matrices \\(\\mathbf{K}_X^{ss} = \\Phi^s\\Phi^{sT}\\), \\(\\mathbf{K}_Y^{ss} = \\Psi^s\\Psi^{sT}\\) are computed as \\((\\mathbf{K}_X^{ss})_{ij} = k_{\\mathcal{X}}(\\mathbf{x}_i^s,\\mathbf{x}_j^s)\\), \\((\\mathbf{K}_Y^{ss})_{ij} = k_{\\mathcal{Y}}(\\mathbf{y}_i^s,\\mathbf{y}_j^s)\\), respectively. And so as \\(\\mathbf{K}_X^{tt}\\), \\(\\mathbf{K}_X^{st}\\) and \\(\\mathbf{K}_Y^{tt}\\). Let \\(\\mathbf{I}_n\\) be the n-dimensional identity matrix, \\(\\mathbf{1}_n\\) be the n-dimensional vector with all elements equal to 1. The \\(n\\)-dimensional centering matrix can be defined as \\(\\mathbf{H}_n = \\mathbf{I}_n - \\frac{1}{n}\\mathbf{1}_n\\mathbf{1}_n^T\\). Then the centralized kernel matrix is defined as \\(\\mathbf{G} = \\mathbf{H}_n\\mathbf{K}\\mathbf{H}_n\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.787, + 0.524 + ], + "angle": 0, + "content": "With above notations, the covariance operator \\(\\mathcal{C}_{XY}\\) can be empirically estimated as \\(\\hat{\\mathcal{C}}_{XY} = \\frac{1}{n} (\\pmb {\\Phi} - \\hat{\\mu}_X\\pmb {1}^T)(\\pmb {\\Psi} - \\hat{\\mu}_Y\\pmb {1}^T)^T = \\frac{1}{n}\\pmb {\\Phi}\\pmb {H}\\pmb {\\Psi}^T\\), then the conditional moment statistics can be estimated as" + }, + { + "type": "equation", + "bbox": [ + 0.261, + 0.531, + 0.786, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {U}} _ {X \\mid Y} = (\\frac {1}{n} \\boldsymbol {\\Phi} \\mathbf {H} \\boldsymbol {\\Psi} ^ {T}) (\\frac {1}{n} \\boldsymbol {\\Psi} \\mathbf {H} \\boldsymbol {\\Psi} ^ {T}) ^ {- 1}, \\quad \\hat {\\mathcal {C}} _ {X X \\mid Y} = \\frac {1}{n} \\boldsymbol {\\Phi} \\mathbf {H} _ {n} \\mathbf {A} (\\boldsymbol {\\Phi} \\mathbf {H} _ {n} \\mathbf {A}) ^ {T}, \\quad (9)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.798, + 0.62 + ], + "angle": 0, + "content": "where \\(\\mathbf{A}\\) satisfies that \\(\\mathbf{B} = \\mathbf{A}\\mathbf{A}^T\\) and \\(B\\triangleq \\mathbf{I}_n - \\frac{1}{n\\varepsilon}\\left[\\mathbf{G}_Y - \\mathbf{G}_Y\\left(\\mathbf{G}_Y + \\varepsilon n\\mathbf{I}_n\\right)^{-1}\\mathbf{G}_Y\\right]\\). Based on these estimations, the closed-form empirical estimation of the COD metric can be derived as follows." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.627, + 0.773, + 0.643 + ], + "angle": 0, + "content": "Proposition 2. The empirical Conditional Operator Discrepancy distance is" + }, + { + "type": "equation", + "bbox": [ + 0.252, + 0.651, + 0.785, + 0.749 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {d} _ {\\mathrm {C O D}} ^ {2} = \\operatorname {t r} \\left(\\mathbf {K} _ {Y} ^ {s s} \\left(\\mathbf {K} _ {Y} ^ {s s}\\right) ^ {- 1} \\mathbf {K} _ {X} ^ {s s} \\left(\\mathbf {K} _ {Y} ^ {s s}\\right) ^ {- 1}\\right) + \\operatorname {t r} \\left(\\mathbf {K} _ {Y} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1} \\mathbf {K} _ {X} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1}\\right) \\\\ - 2 \\operatorname {t r} \\left(\\mathbf {K} _ {Y} ^ {t s} \\left(\\mathbf {K} _ {Y} ^ {s s}\\right) ^ {- 1} \\mathbf {K} _ {X ^ {s} X ^ {t}} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1}\\right) \\\\ + \\varepsilon \\operatorname {t r} \\left[ \\mathbf {G} _ {X} ^ {s} \\left(\\varepsilon n \\mathbf {I} _ {n} + \\mathbf {G} _ {Y} ^ {s}\\right) ^ {- 1} \\right] + \\varepsilon \\operatorname {t r} \\left[ \\mathbf {G} _ {X} ^ {t} \\left(\\varepsilon n \\mathbf {I} _ {n} + \\mathbf {G} _ {Y} ^ {t}\\right) ^ {- 1} \\right] \\tag {10} \\\\ \\left. - \\frac {2}{n} \\left\\| \\left(\\mathbf {H} _ {n} \\mathbf {A} ^ {t}\\right) ^ {T} \\mathbf {K} _ {X} ^ {t s} \\left(\\mathbf {H} _ {n} \\mathbf {A} ^ {s}\\right) \\right\\| _ {*} \\right., \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.758, + 0.784, + 0.79 + ], + "angle": 0, + "content": "where \\(\\| \\cdot \\|_{*}\\) is the nuclear norm. Empirically, \\(\\mathbf{K}_Y^{-1}\\) is approximated by the regularized formulation \\((\\mathbf{K}_Y + \\lambda \\mathbf{I})^{-1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Discriminability of empirical COD. As discussed above, the first term of COD measures the distance of the first-order statistic between conditional distributions in RKHS. However, recent study [5,54] shows that the mean matching" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.315, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "COD: Learning Conditional Invariant Representation for DAR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "may reduce the discriminability when bringing the two distributions close, which undoubtedly degrades the performance of the algorithm. Therefore, it is worth making an in-depth analysis of the discriminability of COD as well." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.298 + ], + "angle": 0, + "content": "Without loss of generality, we first consider Kronechker Delta kernel on conditioning variable \\( Y \\), i.e., \\( k(\\mathbf{y}_1,\\mathbf{y}_2) = \\delta (\\mathbf{y}_1,\\mathbf{y}_2) \\). More explicitly, \\( k(\\mathbf{y}_1,\\mathbf{y}_2) \\) equals to 1 if \\( \\mathbf{y}_1 = \\mathbf{y}_2 \\) or 0 otherwise. For empirical scenarios with finite samples, the numbers of samples and observed labels are finite, which implies the number of distinct label \\( c \\) will be obviously no more than sample size \\( n \\). For convenience, we call two samples are in the same class when their label values are the same. Consequently, the first term in COD can be reformulated as follows." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.306, + 0.784, + 0.336 + ], + "angle": 0, + "content": "Proposition 3. Under the conditions mentioned above, the empirical estimation of \\( d_{CMMD} \\) can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.347, + 0.785, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\widehat {d} _ {\\mathrm {C M M D}} ^ {2} = \\sum_ {p = 1} ^ {c} \\frac {1}{(\\lambda + n _ {p}) ^ {2}} \\left(\\sum_ {i, j} k (\\mathbf {x} _ {(p), i} ^ {s}, \\mathbf {x} _ {(p), j} ^ {s})\\right) \\\\ + \\sum_ {p = 1} ^ {c} \\frac {1}{(\\lambda + n _ {p}) ^ {2}} \\left(\\sum_ {i, j} k \\left(\\mathbf {x} _ {(p), i} ^ {t}, \\mathbf {x} _ {(p), j} ^ {t}\\right)\\right) \\tag {11} \\\\ - 2 \\sum_ {p = 1} ^ {c} \\frac {1}{(\\lambda + n _ {p}) ^ {2}} \\left(\\sum_ {i, j} k (\\mathbf {x} _ {(p), i} ^ {s}, \\mathbf {x} _ {(p), j} ^ {t})\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.504, + 0.784, + 0.535 + ], + "angle": 0, + "content": "where \\( n_i \\) denotes the sample-size of \\( i \\)-th label such that \\( n = \\sum_{i=1}^{c} n_i \\), \\( \\mathbf{x}_{(p),i} \\) denotes the \\( i \\)-th sample of \\( p \\)-th label." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.785, + 0.648 + ], + "angle": 0, + "content": "Eq. (11) contains three terms. The third term characterizes the cross-domain intra-class similarity. When Eq. (11) is minimized, the third term is maximized so that the cross-domain sample with the same label will be aligned, leading to conditional alignment as expected. However, the first two terms of the equation represent the intra-class similarity of both domains. Minimizing them will decrease the intra-class similarity, which means intra-class samples will be pushed away from each other and the discriminability of representations is degraded." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.656, + 0.637, + 0.671 + ], + "angle": 0, + "content": "Remark 2. Note the Kronechker Delta kernel is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.678, + 0.785, + 0.719 + ], + "angle": 0, + "content": "\\[\nk \\left(\\mathbf {y} _ {1}, \\mathbf {y} _ {2}\\right) = \\left\\{ \\begin{array}{l} 1 \\text {i f} \\mathbf {y} _ {1} = \\mathbf {y} _ {2} \\\\ 0 \\text {o t h e r w i s e} \\end{array} \\right.. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.727, + 0.784, + 0.773 + ], + "angle": 0, + "content": "When it comes to the Gaussian kernel for empirical modeling, the case of \\(\\mathbf{y}_1 = \\mathbf{y}_2\\) is the same; for other cases, the kernel values can be arbitrarily close to zero, i.e., for any \\(\\epsilon > 0\\), there always width \\(\\sigma_0\\), such that" + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.783, + 0.785, + 0.801 + ], + "angle": 0, + "content": "\\[\nk \\left(\\mathbf {y} _ {1}, \\mathbf {y} _ {2}\\right) = \\exp \\left(- \\| \\mathbf {y} _ {1} - \\mathbf {y} _ {2} \\| _ {2} ^ {2} / \\sigma_ {0} ^ {2}\\right) < \\epsilon . \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Thus, it is generally reasonable to use Gaussian kernel, which admits a similar property to Kronechker Delta kernel, for empirical modeling." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "H. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "Remark 3. Note that although we treat the number of distinct labels as finite numbers in mini-batch training, it does not mean class-wise matching can be conducted. In fact, even though the pseudo target label has finite range in minibatch, it still admits the continuous property, which means we can not cluster samples as usually done in classification methods. When two samples have predicted labels with little numerical difference, they can be regarded as samples of the same class in classification settings. Nonetheless, this kind of clustering is infeasible in regression settings since the labels are continuous. Therefore, class-wise alignment is not practical even in mini-batch training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.289, + 0.788, + 0.319 + ], + "angle": 0, + "content": "With these observation, it is reasonable to reformulate Eq. (11) to promote the discriminability of representations:" + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.324, + 0.763, + 0.364 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\widehat {d} _ {\\mathrm {C M M D} _ {\\mathrm {m o d}}} = - \\operatorname {t r} \\left(\\mathbf {K} _ {Y} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1} \\mathbf {K} _ {X} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1}\\right) - \\operatorname {t r} \\left(\\mathbf {K} _ {Y} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1} \\mathbf {K} _ {X} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1}\\right) \\\\ - 2 \\operatorname {t r} \\left(\\mathbf {K} _ {Y ^ {t} Y ^ {s}} \\left(\\mathbf {K} _ {Y} ^ {s s}\\right) ^ {- 1} \\mathbf {K} _ {X ^ {s} X ^ {t}} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1}\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.788, + 0.414 + ], + "angle": 0, + "content": "Now minimizing Eq. (14) will maximize intra-class similarity regardless of domain, yielding conditional matching with great discriminability. Then the empirical COD is modified as" + }, + { + "type": "equation", + "bbox": [ + 0.238, + 0.421, + 0.785, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\widehat {d} _ {\\mathrm {C O D} _ {\\mathrm {m o d}}} = - \\operatorname {t r} \\left(\\mathbf {K} _ {Y} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1} \\mathbf {K} _ {X} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1}\\right) - \\operatorname {t r} \\left(\\mathbf {K} _ {Y} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1} \\mathbf {K} _ {X} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1}\\right) \\\\ - 2 \\operatorname {t r} \\left(\\mathbf {K} _ {Y ^ {t} Y ^ {s}} \\left(\\mathbf {K} _ {Y} ^ {s s}\\right) ^ {- 1} \\mathbf {K} _ {X ^ {s} X ^ {t}} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1}\\right) \\\\ + \\varepsilon \\operatorname {t r} \\left[ \\mathbf {G} _ {X} ^ {s} \\left(\\varepsilon n \\mathbf {I} _ {n} + \\mathbf {G} _ {Y} ^ {s}\\right) ^ {- 1} \\right] + \\varepsilon \\operatorname {t r} \\left[ \\mathbf {G} _ {X} ^ {t} \\left(\\varepsilon n \\mathbf {I} _ {n} + \\mathbf {G} _ {Y} ^ {t}\\right) ^ {- 1} \\right] \\tag {15} \\\\ - \\frac {2}{n} \\left\\| \\left(\\mathbf {H} _ {n} \\mathbf {A} ^ {t}\\right) ^ {T} \\mathbf {K} _ {X} ^ {t s} \\left(H _ {n} \\mathbf {A} ^ {s}\\right) \\right\\| _ {*} \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.533, + 0.49, + 0.549 + ], + "angle": 0, + "content": "3.4 COD-based DAR Modeling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.676 + ], + "angle": 0, + "content": "According to theorem (2), when the conditional distributions \\( P_{X^s|Y} \\) and \\( P_{X^t|Y} \\) are aligned, a well-trained source predictor can gain promising performance on target domain. However, when the shift is significant, it is difficult for predictor to produce high-quality pseudo labels at the beginning. For this reason, the model is likely to miss the optimal optimization direction in the early training stage and thus fall into sub-optimal solutions. Therefore, the marginal alignment can improve the reliability of initial pseudo labels, and further benefit the conditional alignment. Overall, the learning objective of COD can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.68, + 0.785, + 0.704 + ], + "angle": 0, + "content": "\\[\n\\min _ {g, h} \\mathcal {L} _ {\\mathrm {s r c}} + \\lambda_ {1} \\widehat {d} _ {\\mathrm {C O D} _ {\\mathrm {m o d}}} + \\lambda_ {2} \\widehat {d} _ {\\mathrm {K G W}}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.71, + 0.787, + 0.74 + ], + "angle": 0, + "content": "where \\(\\lambda_1, \\lambda_2\\) are the trade-off parameters. For regression task, the source risk \\(\\mathcal{L}_{\\mathrm{src}}\\) is commonly set as Mean Square Error (MSE) over continuous outputs:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.748, + 0.785, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s r c}} = \\frac {1}{n ^ {s}} \\sum_ {i = 1} ^ {n ^ {s}} \\left\\| \\hat {\\mathbf {y}} _ {i} ^ {s} - \\mathbf {y} _ {i} ^ {s} \\right\\| _ {2} ^ {2} \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.794, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{y}}_i^s\\) is the predicted value of the \\(i\\)-th source sample. Note that the marginal alignment term can be replaced by any other discrepancy metric, while we adopt KGW here to preserve the consistency with OT framework." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.315, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "COD: Learning Conditional Invariant Representation for DAR" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.376, + 0.163 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.18, + 0.784, + 0.225 + ], + "angle": 0, + "content": "We evaluate our method with several state-of-the-art domain adaptation methods on three regression benchmark datasets, implementation details and additional illustrations are provided in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.226, + 0.785, + 0.346 + ], + "angle": 0, + "content": "dSprites [19] is a standard 2D synthetic dataset for deep representation learning, which contains three domains each with 737,280 images: Color (C), Noisy (N) and Scream (S). Following [4], we consider three factors for regression tasks: scale and two plane coordinates, while the orientation factor is excluded for its complexity. We evaluate all methods on six transfer tasks: \\(\\mathbf{C} \\rightarrow \\mathbf{N}\\), \\(\\mathbf{C} \\rightarrow \\mathbf{S}\\), \\(\\mathbf{N} \\rightarrow \\mathbf{C}\\), \\(\\mathbf{N} \\rightarrow \\mathbf{S}\\), \\(\\mathbf{S} \\rightarrow \\mathbf{C}\\) and \\(\\mathbf{S} \\rightarrow \\mathbf{N}\\). Following previous works [4,38], the sum of Mean Absolute Error (MAE) on all three sub-regression tasks is reported as the evaluation metric." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.785, + 0.422 + ], + "angle": 0, + "content": "Biwi kinect [12] is a real-world dataset containing two domains according to gender: Female \\((\\mathbf{F})\\) with 5874 images and Male \\((\\mathbf{M})\\) with 9804 images. The three factors that can be employed for regression tasks are pitch, yaw and roll. We evaluate all methods on two transfer tasks: \\(\\mathbf{F} \\rightarrow \\mathbf{M}\\) and \\(\\mathbf{M} \\rightarrow \\mathbf{F}\\). The sum of MAE on three sub-regression tasks is reported." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.785, + 0.528 + ], + "angle": 0, + "content": "MPI3D [16] is a simulation-to-real dataset of 3D objects which consist of three different domain: Toy (T), RealistiC (RC) and ReaL (RL). Each domain contains 1,036,800 images. There are two factors in MPI3D that can be employed for regression tasks: a rotation about a vertical axis and a second rotation about a horizontal axis. We evaluate all methods on six transfer tasks: RL \\(\\rightarrow\\) RC, RL \\(\\rightarrow\\) T, RC \\(\\rightarrow\\) T, RC \\(\\rightarrow\\) RL, T \\(\\rightarrow\\) RL and T \\(\\rightarrow\\) RC. The sum of MAE on two sub-regression tasks is reported." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.53, + 0.787, + 0.605 + ], + "angle": 0, + "content": "Comparison methods. SOTA methods for UDA and DAR are selected: TCA [40], MCD [45], DAN [28], DANN [14], JDOT [8], AFN [58], RSD [4], DARE-GRAM [38]. Following previous works [4,38], Mean Absolute Error (MAE) is used as our evaluation metric across all regression tasks. In comparison experiment, the proposed DAR model in Eq. (16) is denoted as COD for convenience." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.63, + 0.436, + 0.646 + ], + "angle": 0, + "content": "4.1 Results and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Comparison on dSprites. As shown in Tab. 1, COD achieves the best performance among all competing methods in nearly all tasks of dSprites. In \\(\\mathbf{C} \\rightarrow \\mathbf{N}\\) and \\(\\mathbf{C} \\rightarrow \\mathbf{S}\\), in which the Resnet-18 and UDA methods exhibit terrible performance because of the significant domain gap. Although SOTA DAR methods have greatly boosted the performance compared with classification methods, COD still achieves a significant improvement, reducing the error to only \\(40\\%\\) of that produced by SOTA methods. These results demonstrate that the conditional invariant representations learned by COD indeed ensure better discriminability. Comparison on Biwi Kinect. Since Biwi Kinect is the closest to real-world scenarios among the three benchmark datasets, the performance improvement of model is more challenging. In Tab. 2, the performance gap between UDA methods and DAR methods is relatively smaller, which implies the SOTA DAR" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "H. Yang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.229, + 0.145, + 0.773, + 0.16 + ], + "angle": 0, + "content": "Table 1: Domain adaptation regression results (MAE) on dSprites (ResNet-18)." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.171, + 0.782, + 0.334 + ], + "angle": 0, + "content": "
MethodC→NC→SN→CN→SS→CS→NAvg.
Resnet-180.940.900.160.650.080.260.498
TCA0.940.870.190.660.100.230.498
MCD0.810.810.170.650.070.190.450
JDOT0.860.790.190.640.100.230.468
AFN1.000.960.160.620.080.320.523
DAN0.700.770.120.500.060.110.377
DANN0.470.460.160.650.050.100.315
RSD0.310.310.120.530.070.080.237
DARE-GRAM0.300.200.110.250.050.070.164
COD0.120.160.100.230.080.120.134
" + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.348, + 0.785, + 0.362 + ], + "angle": 0, + "content": "Table 2: Domain adaptation regression results (MAE) on Biwi Kinect (ResNet-18)." + }, + { + "type": "table", + "bbox": [ + 0.349, + 0.375, + 0.657, + 0.537 + ], + "angle": 0, + "content": "
MethodM→FF→MAvg.
Resnet-180.290.380.335
TCA0.310.390.350
MCD0.310.370.340
JDOT0.290.390.340
AFN0.320.410.365
DAN0.280.370.325
DANN0.300.370.335
RSD0.260.300.280
DARE-GRAM0.230.290.260
COD0.200.210.205
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.785, + 0.748 + ], + "angle": 0, + "content": "methods with marginal alignment indeed show limitations in the difficult transfer tasks. In such a scenario, the sufficiency of condition alignment is directly validated, where the MAE of COD is significantly lower than other SOTA methods. Thus, for real-world regression applications with complex shift, the COD model indeed shows superiority in both theoretical and methodological aspects. Comparison on MPI3D. In MPI3D, the resolution remains the same with dSprites, while the main challenge is the large sample size. In Tab. 3, the MAE of COD is competitive with the SOTA method DARE-GRAM in the first four tasks, while higher in the last two tasks. However, note that COD still achieves the second-best result in averaged MAE, and obtains the lowest MAE in task RL \\(\\rightarrow\\) T. In conclusion, the overall MAE of COD on three benchmarks is still lower, which demonstrates COD is generally superior to SOTA methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Ablation study. The ablation experiment was conducted on the most challenging tasks from all three datasets: \\(\\mathbf{C} \\rightarrow \\mathbf{S}\\) and \\(\\mathbf{N} \\rightarrow \\mathbf{S}\\) from dSprites, \\(\\mathbf{T} \\rightarrow \\mathbf{R}\\mathbf{L}\\) and \\(\\mathbf{T} \\rightarrow \\mathbf{RC}\\) from MPI3D, and both \\(\\mathbf{M} \\rightarrow \\mathbf{F}\\) and \\(\\mathbf{F} \\rightarrow \\mathbf{M}\\) from Biwi Kinect. The major components of proposed method, i.e., MSE objective \\(\\mathcal{L}_{src}\\), KGW distance \\(d_{\\mathrm{KGW}}\\), COD distance \\(d_{\\mathrm{COD}}\\) and modified COD \\(d_{\\mathrm{COD}_{mod}}\\), are evaluated, where the results are shown in Tab. 4." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.315, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "COD: Learning Conditional Invariant Representation for DAR" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.231, + 0.145, + 0.772, + 0.16 + ], + "angle": 0, + "content": "Table 3: Domain adaptation regression results (MAE) on MPI3D (ResNet-18)." + }, + { + "type": "table", + "bbox": [ + 0.214, + 0.172, + 0.79, + 0.334 + ], + "angle": 0, + "content": "
MethodRL→RCRL→TRC→RLRC→TT→RLT→RCAvg.
Resnet-180.170.440.190.450.510.500.377
TCA0.170.420.190.420.500.500.373
MCD0.130.400.150.450.520.500.358
JDOT0.160.410.160.410.470.470.353
AFN0.180.450.200.460.530.530.390
DAN0.120.350.120.270.400.410.278
DANN0.090.240.110.410.480.370.283
RSD0.090.190.080.150.360.360.205
DARE-GRAM0.090.150.100.140.240.240.160
COD0.100.110.130.150.370.290.192
" + }, + { + "type": "table_caption", + "bbox": [ + 0.316, + 0.352, + 0.684, + 0.366 + ], + "angle": 0, + "content": "Table 4: Ablation study on different metrics for DAR." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.378, + 0.785, + 0.488 + ], + "angle": 0, + "content": "
MSEObjectivesdSpritesBiwi KinectMPI3D
dKGWdCODdCODmodC→SN→ST→RLT→RCM→FF→M
✓0.900.650.290.380.510.50
✓✓0.180.260.210.250.410.34
✓✓0.260.450.210.220.400.31
✓✓✓0.170.250.210.220.410.29
✓✓✓0.160.230.200.210.370.29
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.788, + 0.718 + ], + "angle": 0, + "content": "From the \\(2^{\\mathrm{nd}}\\) and \\(3^{\\mathrm{rd}}\\) rows, it can be observed that both OT-based metrics show promising results on DAR datasets, where the conditional alignment with COD shows better performance than marginal alignment with KGW on MPI3D and Biwi Kinect datasets. On the other hand, since dSprites induce more difficult transfer problems, the baseline performance is generally lower, and the pseudo labels for conditional alignment are less reliable. Therefore, the combination of KGW and COD is necessary to ensure higher performance on challenging tasks, where the results in the \\(4^{\\mathrm{th}}\\) row validate the mutually beneficial relation between these two objectives. Finally, the \\(5^{\\mathrm{th}}\\) row shows the effectiveness of rethinking the discriminability of COD, which implies the modifications on the moment statistics indeed ensure lower MAE values on all datasets. Overall, these results demonstrate that conditional alignment is indeed crucial for successful DAR, and the theoretical results and proposed method are generally valid." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Feature Visualization. t-SNE [34] is employed to visualize the 2-D features of different alignments on dSprites \\(\\mathrm{C}\\rightarrow \\mathrm{N}\\) task. (a) Before adaptation, distributions of source and target domain differ severely, as shown in Fig. 3(a), which implies predictor is infeasible for the target samples with a significant domain gap. (b) After marginal distribution alignment via the SOTA DARE-GRAM method, source and target representations are aligned globally, and the cross-domain distributions are overlapped on the whole. However, there are still misaligned local patterns where source and target samples of different colors are matched." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.367, + 0.129 + ], + "angle": 0, + "content": "H. Yang et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.151, + 0.782, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.318 + ], + "angle": 0, + "content": "Fig. 3: t-SNE visualization of learned representations. Label values are denoted by color gradients. Eight label values are selected from the range of the variable for visualization. \\( + \\): source samples; \\( o \\): target samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.345, + 0.788, + 0.407 + ], + "angle": 0, + "content": "Therefore, it is hard to fit clear level sets in Fig. 3 (b). (c) After conditional distribution alignment via COD, source and target distributions are aligned not only globally but also locally. As shown in Fig. 3 (c), representations are matched according to label values so the change of color shows a clear gradient and level" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.407, + 0.536, + 0.42 + ], + "angle": 0, + "content": "sets suitable for both domains can be fitted." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.544, + 0.602 + ], + "angle": 0, + "content": "Hyper-parameter. We investigate the sensitivity of hyper-parameters on dSprites \\(\\mathrm{C}\\rightarrow \\mathrm{N}\\) MAE values by varying \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) ,i.e., parameters of modified COD and KGW, are shown in Fig. 4. The results show that the performance of the COD-based DAR model is generally robust to the different choice of parameters. In fact, the MAE values of the COD-based model are stable in random experiments, which implies the proposed method is both robust from the view of randomness and parameter setting. In conclusion, the sta" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.421, + 0.784, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.565, + 0.786, + 0.594 + ], + "angle": 0, + "content": "Fig.4: MAE values under different settings of hyper-parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.602, + 0.778, + 0.619 + ], + "angle": 0, + "content": "ble performance validates that the proposed model is also empirically reliable." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.642, + 0.36, + 0.658 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this work, we studied the limitations of existing DAR research, i.e., the misaligned intrinsic discriminant structure and negative transfer of marginal distribution matching framework. To deal with the challenges of conditional shift with continuous output space, we provide feasible solutions from both theoretical and methodological views, where the main results show that: 1) the cross-domain generalization regression error can be sufficiently bounded by the conditional discrepancy on continuous labels; 2) the proposed COD serves as a well-defined metric for conditional discrepancy measures and empirical modeling. With these guarantees, a COD-based DAR method is proposed with a discriminability enhancement mechanism. Numerical validations on standard DAR datasets validate the superiority of COD." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.315, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "COD: Learning Conditional Invariant Representation for DAR" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.299 + ], + "angle": 0, + "content": "This work is supported in part by National Natural Science Foundation of China (Grant No. 62376291), in part by Guangdong Basic and Applied Basic Research Foundation (2023B1515020004), in part by Science and Technology Program of Guangzhou (2024A04J6413), in part by the Fundamental Research Funds for the Central Universities, Sun Yat-sen University (24xkjc013), in part by Guangdong Province Key Laboratory of Computational Science at Sun Yat-sen University (2020B1212060032), and in part by Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.322, + 0.323, + 0.339 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.354, + 0.785, + 0.381 + ], + "angle": 0, + "content": "1. Bao, Y., Liu, Y., Wang, H., Lu, F.: Generalizing gaze estimation with rotation consistency. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4207-4216 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.383, + 0.785, + 0.41 + ], + "angle": 0, + "content": "2. Ben-David, S., Blitzer, J., Crammer, K., Kulesza, A., Pereira, F., Vaughan, J.W.: A theory of learning from different domains. Mach. Learn. 79, 151-175 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.411, + 0.785, + 0.451 + ], + "angle": 0, + "content": "3. Chen, L., Chen, H., Wei, Z., Jin, X., Tan, X., Jin, Y., Chen, E.: Reusing the task-specific classifier as a discriminator: Discriminator-free adversarial domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7181-7190 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.452, + 0.785, + 0.479 + ], + "angle": 0, + "content": "4. Chen, X., Wang, S., Wang, J., Long, M.: Representation subspace distance for domain adaptation regression. In: Int. Conf. Mach. Learn. pp. 1749-1759 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.48, + 0.785, + 0.52 + ], + "angle": 0, + "content": "5. Chen, Y., Song, S., Li, S., Wu, C.: A graph embedding framework for maximum mean discrepancy-based domain adaptation algorithms. IEEE Trans. Image Process. 29, 199-213 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.522, + 0.785, + 0.562 + ], + "angle": 0, + "content": "6. Tachet des Combes, R., Zhao, H., Wang, Y.X., Gordon, G.J.: Domain adaptation with conditional distribution matching and generalized label shift. In: Adv. Neural Inform. Process. Syst. vol. 33, pp. 19276-19289 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.564, + 0.785, + 0.59 + ], + "angle": 0, + "content": "7. Cortes, C., Mohri, M.: Domain adaptation in regression. In: ICALT. pp. 308-323 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.591, + 0.785, + 0.632 + ], + "angle": 0, + "content": "8. Courty, N., Flamary, R., Habrard, A., Rakotomamonjy, A.: Joint distribution optimal transportation for domain adaptation. Adv. Neural Inform. Process. Syst. 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.785, + 0.66 + ], + "angle": 0, + "content": "9. Courty, N., Flamary, R., Tuia, D.: Domain adaptation with regularized optimal transport. In: ECML PKDD. pp. 274-289 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.66, + 0.785, + 0.701 + ], + "angle": 0, + "content": "10. Damodaran, B.B., Kellenberger, B., Flamary, R., Tuia, D., Courty, N.: Deepjdot: Deep joint distribution optimal transport for unsupervised domain adaptation. In: Eur. Conf. Comput. Vis. pp. 447-463 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.702, + 0.785, + 0.742 + ], + "angle": 0, + "content": "11. Dhaini, M., Berar, M., Honeine, P., Van Exem, A.: Unsupervised domain adaptation for regression using dictionary learning. Knowl.-Based Syst. 267, 110439 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.743, + 0.785, + 0.771 + ], + "angle": 0, + "content": "12. Fanelli, G., Dantone, M., Gall, J., Fossati, A., Van Gool, L.: Random forests for real time 3d face analysis. Int. J. Comput. Vis. **101**, 437–458 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.772, + 0.785, + 0.798 + ], + "angle": 0, + "content": "13. Fukumizu, K., Bach, F.R., Jordan, M.I.: Kernel dimension reduction in regression. Ann. Stat. 37(4), 1871-1905 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "14. Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., March, M., Lempitsky, V.: Domain-adversarial training of neural networks. J. Mach. Learn. Res. 17(59), 1-35 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.354, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.129 + ], + "angle": 0, + "content": "H. Yang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "15. Geng, X., Zhou, Z.H., Smith-Miles, K.: Automatic age estimation based on facial aging patterns. IEEE Trans. Pattern Anal. Mach. Intell. 29(12), 2234-2240 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.787, + 0.229 + ], + "angle": 0, + "content": "16. Gondal, M.W., Wuthrich, M., Miladinovic, D., Locatello, F., Breidt, M., Volchkov, V., Akpo, J., Bachem, O., Scholkopf, B., Bauer, S.: On the transfer of inductive bias from simulation to the real world: a new disentanglement dataset. Adv. Neural Inform. Process. Syst. 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.786, + 0.271 + ], + "angle": 0, + "content": "17. Gong, B., Shi, Y., Sha, F., Grauman, K.: Geodesic flow kernel for unsupervised domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2066-2073 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.786, + 0.311 + ], + "angle": 0, + "content": "18. Goswami, S., Kontolati, K., Shields, M.D., Karniadakis, G.E.: Deep transfer operator learning for partial differential equations under conditional shift. Nat. Mach. Intell. 4(12), 1155-1164 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.312, + 0.786, + 0.352 + ], + "angle": 0, + "content": "19. Higgins, I., Matthew, L., Pal, A., Burgess, C., Glorot, X., Botvinick, M., Mohamed, S., Lerchner, A.: beta-vae: Learning basic visual concepts with a constrained variational framework. In: Int. Conf. Learn. Represent. (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.353, + 0.786, + 0.393 + ], + "angle": 0, + "content": "20. Hoffman, J., Tzeng, E., Park, T., Zhu, J.Y., Isola, P., Saenko, K., Efros, A., Darrell, T.: Cycada: Cycle-consistent adversarial domain adaptation. In: Int. Conf. Mach. Learn. pp. 1989-1998 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.394, + 0.786, + 0.434 + ], + "angle": 0, + "content": "21. Hu, J., Zhong, H., Yang, F., Gong, S., Wu, G., Yan, J.: Learning unbiased transferability for domain adaptation by uncertainty modeling. In: Eur. Conf. Comput. Vis. pp. 223-241 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.435, + 0.786, + 0.474 + ], + "angle": 0, + "content": "22. Jiang, J., Ji, Y., Wang, X., Liu, Y., Wang, J., Long, M.: Regressive domain adaptation for unsupervised keypoint detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6780-6789 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.476, + 0.786, + 0.515 + ], + "angle": 0, + "content": "23. Kirchmeyer, M., Rakotomamonjy, A., de Bezenac, E., Gallinari, P.: Mapping conditional distributions for domain adaptation under generalized target shift. In: Int. Conf. Learn. Represent. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.786, + 0.555 + ], + "angle": 0, + "content": "24. Lee, C.Y., Batra, T., Baig, M.H., Ulbricht, D.: Sliced wasserstein discrepancy for unsupervised domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10285-10295 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.557, + 0.786, + 0.597 + ], + "angle": 0, + "content": "25. Li, B., Wang, Y., Zhang, S., Li, D., Keutzer, K., Darrell, T., Zhao, H.: Learning invariant representations and risks for semi-supervised domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1104-1113 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.598, + 0.786, + 0.623 + ], + "angle": 0, + "content": "26. Li, X., Li, J., Zuo, L., Zhu, L., Shen, H.T.: Domain adaptive remaining useful life prediction with transformer. IEEE Trans. Instrum. 71, 1-13 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.625, + 0.786, + 0.664 + ], + "angle": 0, + "content": "27. Li, Z., Zhang, K., Liu, Y., Zou, Y., Ding, G.: A novel remaining useful life transfer prediction method of rolling bearings based on working conditions common benchmark. IEEE Trans. Instrum. 71, 1-9 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.665, + 0.786, + 0.692 + ], + "angle": 0, + "content": "28. Long, M., Cao, Y., Wang, J., Jordan, M.: Learning transferable features with deep adaptation networks. In: Int. Conf. Mach. Learn. pp. 97-105 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.693, + 0.786, + 0.718 + ], + "angle": 0, + "content": "29. Long, M., Cao, Z., Wang, J., Jordan, M.I.: Conditional adversarial domain adaptation. Adv. Neural Inform. Process. Syst. 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.719, + 0.786, + 0.745 + ], + "angle": 0, + "content": "30. Long, M., Zhu, H., Wang, J., Jordan, M.I.: Deep transfer learning with joint adaptation networks. In: Int. Conf. Mach. Learn. pp. 2208-2217 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.786, + 0.772 + ], + "angle": 0, + "content": "31. Luo, Y.W., Ren, C.X.: Conditional bures metric for domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 13989-13998 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.773, + 0.786, + 0.799 + ], + "angle": 0, + "content": "32. Luo, Y.W., Ren, C.X.: Mot: Masked optimal transport for partial domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3531-3540 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "33. Luo, Y.W., Ren, C.X., Dai, D.Q., Yan, H.: Unsupervised domain adaptation via discriminative manifold propagation. IEEE Trans. Pattern Anal. Mach. Intell. 44(3), 1653-1669 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.316, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "COD: Learning Conditional Invariant Representation for DAR" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "34. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. J. Mach. Learn. Res. 9(11) (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.202 + ], + "angle": 0, + "content": "35. Mansour, Y., Mohri, M., Rostamizadeh, A.: Domain adaptation: Learning bounds and algorithms. arXiv preprint arXiv:0902.3430 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.242 + ], + "angle": 0, + "content": "36. Mao, W., Ge, Y., Shen, C., Tian, Z., Wang, X., Wang, Z., den Hengel, A.v.: Poseur: Direct human pose regression with transformers. In: Eur. Conf. Comput. Vis. pp. 72-88 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.271 + ], + "angle": 0, + "content": "37. de Mathelin, A., Richard, G., Deheeger, F., Mougeot, M., Vayatis, N.: Adversarial weighting for domain adaptation in regression. In: ICTAI. pp. 49-56 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.311 + ], + "angle": 0, + "content": "38. Nejjar, I., Wang, Q., Fink, O.: Dare-gram: Unsupervised domain adaptation regression by aligning inverse gram matrices. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 11744–11754 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.352 + ], + "angle": 0, + "content": "39. Nikzad-Langerodi, R., Zellinger, W., Saminger-Platz, S., Moser, B.A.: Domain adaptation for regression under beer-lambert's law. Knowl.-Based Syst. 210, 106447 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.379 + ], + "angle": 0, + "content": "40. Pan, S.J., Tsang, I.W., Kwok, J.T., Yang, Q.: Domain adaptation via transfer component analysis. IEEE Trans. Neural Netw. 22(2), 199-210 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.406 + ], + "angle": 0, + "content": "41. Pei, Z., Cao, Z., Long, M., Wang, J.: Multi-adversarial domain adaptation. In: AAAI. vol. 32 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.785, + 0.447 + ], + "angle": 0, + "content": "42. Peng, R., Wang, R., Wang, Z., Lai, Y., Wang, R.: Rethinking depth estimation for multi-view stereo: A unified representation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8645-8654 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.473 + ], + "angle": 0, + "content": "43. Quinonero-Candela, J., Sugiyama, M., Schwaighofer, A., Lawrence, N.D.: Dataset shift in machine learning. Mit Press (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "44. Ren, C.X., Luo, Y.W., Dai, D.Q.: Buresnet: Conditional bures metric for transferable representation learning. IEEE Trans. Pattern Anal. Mach. Intell. 45(4), 4198-4213 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.556 + ], + "angle": 0, + "content": "45. Saito, K., Watanabe, K., Ushiku, Y., Harada, T.: Maximum classifier discrepancy for unsupervised domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3723-3732 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.785, + 0.582 + ], + "angle": 0, + "content": "46. Shen, J., Qu, Y., Zhang, W., Yu, Y.: Wasserstein distance guided representation learning for domain adaptation. In: AAAI. vol. 32 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.609 + ], + "angle": 0, + "content": "47. Singh, A., Chakraborty, S.: Deep domain adaptation for regression. Development and Analysis of Deep Learning Architectures pp. 91-115 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.785, + 0.65 + ], + "angle": 0, + "content": "48. Song, L., Huang, J., Smola, A., Fukumizu, K.: Hilbert space embeddings of conditional distributions with applications to dynamical systems. In: Int. Conf. Mach. Learn. pp. 961-968 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.785, + 0.677 + ], + "angle": 0, + "content": "49. Sun, B., Feng, J., Saenko, K.: Return of frustratingly easy domain adaptation. In: AAAI. vol. 30 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.785, + 0.704 + ], + "angle": 0, + "content": "50. Tang, H., Jia, K.: Discriminative adversarial domain adaptation. In: AAAI. vol. 34, pp. 5940-5947 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.785, + 0.731 + ], + "angle": 0, + "content": "51. Teshima, T., Sato, I., Sugiyama, M.: Few-shot domain adaptation by causal mechanism transfer. In: Int. Conf. Mach. Learn. pp. 9458-9469 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.785, + 0.758 + ], + "angle": 0, + "content": "52. Tzeng, E., Hoffman, J., Saenko, K., Darrell, T.: Adversarial discriminative domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7167-7176 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "53. Wang, H., Shen, T., Zhang, W., Duan, L.Y., Mei, T.: Classes matter: A fine-grained adversarial approach to cross-domain semantic segmentation. In: Eur. Conf. Comput. Vis. pp. 642-659 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "54. Wang, W., Li, H., Ding, Z., Nie, F., Chen, J., Dong, X., Wang, Z.: Rethinking maximum mean discrepancy for visual domain adaptation. IEEE Trans. Neural Netw. Learn. Syst. (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.367, + 0.129 + ], + "angle": 0, + "content": "H. Yang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "55. Wang, Y., Jiang, Y., Li, J., Ni, B., Dai, W., Li, C., Xiong, H., Li, T.: Contrastive regression for domain adaptation on gaze estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19376-19385 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "56. Wei, P., Ke, Y., Ong, Y.S., Ma, Z.: Adaptive transfer kernel learning for transfer gaussian process regression. IEEE Trans. Pattern Anal. Mach. Intell. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "57. Wu, J., He, J., Wang, S., Guan, K., Ainsworth, E.: Distribution-informed neural networks for domain adaptation regression. Adv. Neural Inform. Process. Syst. 35, 10040-10054 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.3 + ], + "angle": 0, + "content": "58. Xu, R., Li, G., Yang, J., Lin, L.: Larger norm more transferable: An adaptive feature norm approach for unsupervised domain adaptation. In: Int. Conf. Comput. Vis. pp. 1426-1435 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "59. Zhang, Z., Wang, M., Nehorai, A.: Optimal transport in reproducing kernel hilbert spaces: Theory and applications. IEEE Trans. Pattern Anal. Mach. Intell. 42(7), 1741-1754 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.37 + ], + "angle": 0, + "content": "60. Zhao, H., Combes, R.T.D., Zhang, K., Gordon, G.: On learning invariant representations for domain adaptation. In: Int. Conf. Mach. Learn. pp. 7523-7532 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.398 + ], + "angle": 0, + "content": "61. Zhao, H., Coston, A., Adel, T., Gordon, G.J.: Conditional learning of fair representations. In: Int. Conf. Learn. Represent. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.439 + ], + "angle": 0, + "content": "62. Zhao, H., Dan, C., Aragam, B., Jaakkola, T.S., Gordon, G.J., Ravikumar, P.: Fundamental limits and tradeoffs in invariant representation learning. J. Mach. Learn. Res. 23(340), 1-49 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.439 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/COD_ Learning Conditional Invariant Representation for Domain Adaptation Regression/5f47c561-6103-4c2b-b6da-ac9368ec9383_origin.pdf b/2024/COD_ Learning Conditional Invariant Representation for Domain Adaptation Regression/5f47c561-6103-4c2b-b6da-ac9368ec9383_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..53cc22d8e924cf48236caa2c8c8ac89c3d633328 --- /dev/null +++ b/2024/COD_ Learning Conditional Invariant Representation for Domain Adaptation Regression/5f47c561-6103-4c2b-b6da-ac9368ec9383_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:829b1548b60589f1a1175d17d480dd50e0a7d8470193dbdec98047db49ce9252 +size 1377730 diff --git a/2024/COD_ Learning Conditional Invariant Representation for Domain Adaptation Regression/full.md b/2024/COD_ Learning Conditional Invariant Representation for Domain Adaptation Regression/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b718b1f668a8272e037db377f4719f0762100af3 --- /dev/null +++ b/2024/COD_ Learning Conditional Invariant Representation for Domain Adaptation Regression/full.md @@ -0,0 +1,347 @@ +# COD: Learning Conditional Invariant Representation for Domain Adaptation Regression + +Hao-Ran Yang1, Chuan-Xian Ren1*, and You-Wei Luo1,2 + +$^{1}$ School of Mathematics, Sun Yat-Sen University, China + $^{2}$ School of Mathematics, Jiaying University, China +{yanghr26@mail2, rchuanx@mail, luoyw28@mail2}.sysu.edu.cn + +Abstract. Aiming to generalize the label knowledge from a source domain with continuous outputs to an unlabeled target domain, Domain Adaptation Regression (DAR) is developed for complex practical learning problems. However, due to the continuity problem in regression, existing conditional distribution alignment theory and methods with discrete prior, which are proven to be effective in classification settings, are no longer applicable. In this work, focusing on the feasibility problems in DAR, we establish the sufficiency theory for the regression model, which shows the generalization error can be sufficiently dominated by the cross-domain conditional discrepancy. Further, to characterize conditional discrepancy with continuous conditioning variable, a novel Conditional Operator Discrepancy (COD) is proposed, which admits the metric property on conditional distributions via the kernel embedding theory. Finally, to minimize the discrepancy, a COD-based conditional invariant representation learning model is proposed, and the reformulation is derived to show that reasonable modifications on moment statistics can further improve the discriminability of the adaptation model. Extensive experiments on standard DAR datasets verify the validity of theoretical results and the superiority over SOTA DAR methods. + +Keywords: Domain Adaptation Regression, Optimal Transport, Invariant Representation Learning, Kernel Embedding Theory. + +# 1 Introduction + +Deep learning has shown a promising ability for fitting and achieved great success in various application tasks. However, large-scale data with sufficient annotations plays a vital role in training high-quality deep models while the label annotations for unlabeled data are usually expensive. Therefore, aiming to reduce training costs by transferring knowledge, Unsupervised Domain Adaptation (UDA) [2,14,17,28] has received increasing attention in recent years. Basically, UDA aims to transfer task-related information from a labeled source domain + +to a target domain without annotations, where the domain gap is characterized as the distribution shift across domains [43]. Following this line, considerable methods have been proposed to alleviate the UDA problem by shift correction and distribution alignment, such as adversarial training for cross-domain samples [14, 52] and explicitly distribution matching by statistical discrepancy minimization [28, 49]. From a theoretical perspective, recent literature [6, 60] show that the marginal distribution $P_{X}$ with covariate shift assumption is insufficient to characterize the cross-domain generalization error, while the conditional distribution alignment indeed serves as a sufficient condition to ensure small error and successful transfer. Mathematically, when the conditional distributions $P_{X|y}$ for every $y$ are aligned, a well-trained source predictor can perform well on the target domain, which is described in Fig. 1. Inspired by theoretical advances, recent works [23, 41, 45] consider the statistical methods for conditional alignment, and show promising performance in mitigating negative transfer and improving recognition accuracy. However, most of the existing results are designed for classification settings with discrete prior on label variables, which induces the infeasibility for regression problems. Consequently, it restricts the application of UDA methods for wider real-world problems, e.g., human pose estimation [36], product condition monitoring [27], depth estimation [42]. + +Considering the dilemma of UDA, pioneer works have explored the Domain Adaptation Regression (DAR) framework for solving the distribution shift in continuous label scenarios, where theoretical results [7,35] and correction methods [15,47,51] are provided. To further relax the requirement of labeled training target data in pioneer DAR works, advanced works [4,38] propose unsupervised DAR methods to learn transferable regression models without the help of labeled target data. Nonetheless, existing DAR theory and methods are directly facilitated by the covariate shift framework, where the marginal distribution alignment is considered. Therefore, it inevitably induces the negative transfer and misaligned local structure, which are shown in the results in classification settings [6,60]. In fact, there is a significant obstacle for modeling conditional distribution alignment in regression scenarios, i.e., the conditional shift correction induces infinite matching problems for $P_{X|Y=y}$ on all points $y$ . The continuous label variable with infinite possible outcomes makes it infeasible to directly apply the existing conditional alignment methods for DAR. Thus, it is vital to understand the continuity problem in regression, and develop the generalization error theory and statistical methodology for DAR scenarios. + +In this work, we focus on the theory guarantee and conditional alignment methodology for successful DAR. Theoretically, we explore the generalization upper bound for continuous output scenarios and show the cross-domain joint error can be sufficiently upper-bounded by the conditional discrepancy. It ensures the sufficiency of conditional alignment still holds for regression scenarios and provides reliable guarantees for theory-driven modeling. Methodologically, to deal with the continuous problem in conditional alignment, we treat the conditional distribution as finite statistical moments in Hilbert space. By exploring the distribution embedding theory of Reproducing Kernel Hilbert Space (RKHS), a novel + +![](images/16737da332bfa6f9c5df6a89905a35a2c74bdd50860d8c1d07db50c0d953f06a.jpg) + +![](images/4c86c9bb133f0858a85908027199ac9dc5e091a7dc4256d376478eecd27b60ca.jpg) +Fig. 1: Illustration of the conditional shift in regression setting, where label value $y \in [0,1]$ and color gradients imply the continuity. (a) Before adaptation, distribution shift exists between source and target representations so the predictor trained on source domain cannot generalize to target domain. (b) After marginal distribution alignment, distributions of source and target domain are globally aligned. However, representations with different labels may be falsely aligned across domains, leading to significant inconsistency between cross-domain labeling rules. Thus, level sets provided by the source predictor are not suitable for target representations. (c) After conditional distribution alignment, label-wise matching is achieved, where the level sets of cross-domain labeling rules are consistent and the source predictor will be reliable. + +![](images/204bbfbb086977bb66063a36f96f52e1e60519e6ce4dd6d841d8c2085d014f69.jpg) + +![](images/18e436871d05a6c11880c8428d53d8e409dc9d0bebc9080b485a7915da558ec2.jpg) + +conditional metric called Conditional Operator Discrepancy (COD) is proposed, which can be interpreted as the conditional Wasserstein distance in RKHS. With the sufficiency theory and metric-based learning principle for regression, a COD-based conditional invariant representation learning method is proposed, which aims to correct the conditional shift on the pushforward measures in latent representation space. By reformulating the empirical COD estimation, a modified principle is developed to further improve the discriminability during knowledge transfer. Generally, our contributions can be summarized as follows: + +- The sufficiency theory for the conditional shift framework is established for DAR settings, which shows the cross-domain conditional discrepancy on continuous label variable serves as the major term in generalization error. +- A novel COD metric for characterizing conditional distribution discrepancy with continuous and infinite conditioning variable is explored, providing a feasible solution for conditional discrepancy optimization in DAR. +- A COD-based representation learning model with reformulation on moment statistics is proposed, which admits discriminability on task-related knowledge and shows superior performance in extensive experiments. + +# 2 Related Work + +Domain Adaptation Classification. Based on the distribution shift assumption, most of the UDA methods can be roughly categorized into three types: methods that align marginal distribution, conditional distribution and joint distribution. Most UDA methods consider only the marginal distribution align- + +ment. Discrepancy-based methods attempt to directly minimize the distribution discrepancy and different types of discrepancy measurement greatly enrich this type of methodology. For example, these methods adopt moment matching [28, 49] or extension under manifold framework [17, 33]. In particular, Optimal transport (OT) and Wasserstein distance are introduced to UDA as a sound measurement for domain shift, based on which a large number of methods have been proposed [8-10, 24, 46]. Zhang et al. [59] further develop both theory and methodology to RKHS and Luo et al. [31] extends Kernel Bures metric to conditional distribution alignment. Adversarial-based methods [14, 20, 21] learn domain-invariant features via two-player games. Specifically, they alternatively optimize the feature generator and the domain discriminator to achieve domain confusion. In order to obtain better performance, advanced methods have explored how to utilize the label information to achieve conditional distribution alignment [3, 32, 41, 44, 45, 50, 53, 61] or joint distribution alignment [29, 30]. However, UDA methods for shift correction are proposed for discrete and finite conditioning variables, which are inapplicable for regression. + +Domain Adaptation Regression. DA for regression has received relatively little attention compared with classification. Mansour et al. [35] and Cortes et al. [7] introduce theoretical analysis for DA in regression scenarios. Zhao et al. [62] discuss the tradeoffs between accuracy and invariance in learning invariant representations in both classification and regression settings. A range of methods in the shallow regime are proposed to tackle the DAR problem. These methods can be roughly categorized into two types: importance weighting [15,37] and learning domain-invariant representation [8,39,40]. Besides, some attempts have been made in deep regime to tackle regression problem [25,47,51,56]. Unfortunately, most of them belong to semi-supervised DA, which means they need to use a small number of target domain labels. Recently, deep UDA methods for regression that get rid of target label are proposed [4,11,38,57]. RSD [4] states that deep neural networks are less robust to feature scaling in regression problems and aligning the deep features directly will impede task performance for altering feature scale. Thus, a strategy of aligning orthogonal bases of both domains is adopted in RSD. Based on RSD, DARE-GRAM [38] aligns the inverse Gram Matrix of deep features of both domains, and ensures promising improvements in accuracy and stability. Meanwhile, several DAR methods for specific tasks have been explored [1,18,22,26,27,55]. However, the theory and method for conditional shift correction are still unexplored for DAR scenarios, which hinders the in-depth reasoning and development for practical modeling. + +# 3 COD: Theory and Methodology + +# 3.1 Preliminary + +Setup. In this paper, we discuss UDA method dealing with regression problems. Let $\mathrm{X}$ and $\mathrm{Y}$ be covariate and label variable taken values on $\mathcal{X}$ and $\mathcal{Y}$ . $\mathbb{P}$ denotes the set of distributions and subscripts represent the corresponding random variable while the superscript represents the domain, e.g. $P_{XY}^{s} \in \mathbb{P}_{\mathcal{X}\mathcal{Y}}$ + +implies the joint distribution on source domain. In UDA settings, we are provided with labeled source samples $(x^{s},y^{s})\sim P_{XY}^{s}$ and unlabeled target samples $(x^{t},y^{t})\sim P_{XY}^{t}$ , which means the target labels are not available during training. Samples of source and target domain follow different distributions, i.e. $P_{XY}^{s}\neq P_{XY}^{t}$ . Denote $g:X\to Z$ as the representation extractor, $h:Z\to Y$ as the predictor, $\varepsilon_{s}$ and $\varepsilon_{t}$ as the source and target prediction error, respectively. UDA methods aim to generalize the learner $h\circ g$ across domains and minimize target risk $\varepsilon_{t}$ by mitigating the distribution shift problem. In contrast to the discrete label space in classification settings, unsupervised DAR discusses UDA problems with continuous labels, which means the number of classes is uncountable, leading to great difficulty in aligning conditional distribution $P_{X|Y}$ . The proofs are provided in supplementary materials. + +OT in RKHS. Let $(\mathcal{X},\mathcal{B})$ be a measure space with Borel $\sigma$ -field $\mathcal{B}$ . Denote $(\mathcal{H}_{\mathcal{X}},k_{\mathcal{X}})$ as the RKHS of $\mathcal{X}$ , which is generated by the positive definite kernel $k_{\mathcal{X}}$ . Elements from $\mathcal{X}$ are mapped to $\mathcal{H}_{\mathcal{X}}$ via feature map $\phi (x) = k_{\mathcal{X}}(x,\cdot)$ , which is assumed to satisfy the reproducing properties $\langle \phi (x),\phi (x^{\prime})\rangle_{\mathcal{H}_{\mathcal{X}}} = k_{\mathcal{X}}(x,x^{\prime})$ and $\langle \phi (x),f\rangle_{\mathcal{H}_x} = f(x),\forall f\in \mathcal{H}_x$ . The mean element $\mu_X$ in $\mathcal{H}_{\mathcal{X}}$ with law $P_{X}$ is given by $\mu_{X} = \mathbb{E}_{X}[\phi (X)]$ , and the covariance operator $\mathcal{C}_{XX}$ is defined as $\mathcal{C}_{XX} = E_X(\phi (X) - \mu_X)\otimes (\phi (X) - \mu_X)$ . + +For any two distributions $P_X^s, P_X^t \in \mathbb{P}_X$ , the Kantorovitch optimal transport in RKHS $\mathcal{H}_{\mathcal{K}}$ is considered on pushforward measures $\phi_{\#}P_X^s$ and $\phi_{\#}P_X^t$ : + +$$ +d _ {\mathrm {K W}} ^ {2} \left(X ^ {s}, X ^ {t}\right) = \inf _ {\pi_ {\mathcal {K}} \in \Pi \left(\phi_ {\#} P _ {X} ^ {s}, \phi_ {\#} P _ {X} ^ {t}\right)} \int_ {\mathcal {H} _ {\mathcal {K}} \times \mathcal {H} _ {\mathcal {K}}} \left\| X ^ {s} - X ^ {t} \right\| _ {\mathcal {H} _ {\mathcal {K}}} ^ {2} d \pi_ {\mathcal {K}} \left(X ^ {s}, X ^ {t}\right) \tag {1} +$$ + +Zhang et al. [59] show that if the pushforward measures are Gaussian, the Kernel Gaussian Wasserstein (KGW) distance between them can be written as + +$$ +d _ {\mathrm {K G W}} ^ {2} \left(X ^ {s}, X ^ {t}\right) = \| u _ {X} ^ {s} - u _ {X} ^ {t} \| _ {\mathcal {H} _ {\kappa}} ^ {2} + t r \left(\mathcal {C} _ {X X} ^ {s s} + \mathcal {C} _ {X X} ^ {t t} - 2 \mathcal {C} _ {X X} ^ {s t}\right) \tag {2} +$$ + +where $\mathcal{C}_{XX}^{st} = \sqrt{\sqrt{\mathcal{C}_{XX}^{ss}}\mathcal{C}_{XX}^{tt}\sqrt{\mathcal{C}_{XX}^{ss}}}$ and $\mathcal{C}_{XX}^{ss},\mathcal{C}_{XX}^{tt}$ are the covariance operators of $P_{X}^{s}$ and $P_{X}^{t}$ on $\mathcal{H}_{\mathcal{K}}$ , respectively. + +Remark 1. Note that KGW consists of two terms, i.e., the Maximum Mean Discrepancy (MMD) on first-order statistics and kernel Bures metric on second-order statistics. It provides a basic innovation for exploring conditional discrepancy metric by considering the conditional statistics in RKHS, i.e., first-order statistics in literature [48] and second-order statistics in literature [13]. + +# 3.2 Generalization Error Analysis for DAR + +Most of DAR methods focus on aligning the marginal distribution $P_{X}$ . However, Zhao et al. [60] have shown that in the presence of label shift, only aligning the marginal distribution does not guarantee satisfying generalization performance on target domain. In fact, [6, 60] have proven the following theorem (1) for classification problems: + +Theorem 1 ([60]). For representation variable $Z = g(X)$ , suppose inequality $d_{JS}(P_Y^s, P_Y^t) \geq d_{JS}(P_Z^s, P_Z^t)$ holds, then for predictor $h: Z \to Y$ , we have + +$$ +\varepsilon_ {s} (h \circ g) + \varepsilon_ {t} (h \circ g) \geq \frac {1}{2} \left(\sqrt {d _ {J S} (P _ {Y} ^ {s} , P _ {Y} ^ {t})} - \sqrt {d _ {J S} (P _ {Z} ^ {s} , P _ {Z} ^ {t})}\right) ^ {2}, \tag {3} +$$ + +where $d_{JS}$ denotes the Jensen-Shannon divergence. + +Assuming the label distributions differ between source and target domain, i.e., $d_{JS}(P_Y^s,P_Y^t) > 0$ , the lower bound shows that good alignment of representation distributions leads to bad joint error. In other words, when label shift exists, which is rather common in realistic scenarios, only aligning the marginal distribution is not enough for low target error. + +To deal with the insufficiency problem shown in the lower bound analysis, it is intuitive to introduce label information into generalization analysis, e.g., the balanced error rate for classification scenarios [6]. Analogously, for the continuous regression scenarios, we first define the continuous balanced error rate for quantifying maximum conditional risk in DAR. + +Definition 1. The continuous balanced error rate of predictor $h$ on distribution $P$ is defined as: + +$$ +\operatorname {c B E R} _ {\mathrm {P}} (h, Y) := \sup _ {(x, y) \in \mathcal {X} \oplus \mathcal {Y}} P (h (x) \neq Y \mid Y = y). \tag {4} +$$ + +Intuitively, the continuous balanced error rate can be regarded as the maximum potential risk for predicting sample $x$ with label $y$ . Based on cBER and conditional shift correction with representation transformation $Z = g(X)$ , a sufficient condition for DAR can be induced as follows. + +Theorem 2. For representation variable $Z^s = g(X^s)$ and $Z^t = g(X^t)$ , if $g(\cdot)$ satisfies conditional invariant property, i.e., + +$$ +P _ {Z \mid Y = y} ^ {s} = P _ {Z \mid Y = y} ^ {t}, \quad \forall y \in \mathcal {Y}, \tag {5} +$$ + +then for any predictor $h: Z \to Y$ , we have + +$$ +\varepsilon_ {s} (h) + \varepsilon_ {t} (h) \leq 2 \mathrm {c B E R} _ {\mathrm {P s}} (h, Y). \tag {6} +$$ + +Therefore, the key for DA, no matter for classification or regression problems, is to achieve conditional distribution alignment. In classification settings, conditional alignment can be achieved by conducting class-wise marginal distribution alignment with the help of target pseudo label. However, in regression scenarios, the label variable is continuous, which means previous methods correcting conditional shift in classification settings are no longer applicable. Obviously, it is impossible to align the $P_{X|y}$ for every single $y$ when $y$ is a continuous variable and the possible number of $y$ is infinite. Thus, it is important and meaningful to explore a new framework that is capable of conditional distribution alignment for continuous label variable. + +![](images/ca6e49b2c95c085cdfb1e4fc57ed3ce0542789eee74f050f932ee74df187057c.jpg) +Fig. 2: Illustration of COD metric. (a) Classification DA methods usually correct conditional shift by class-wise computations on discrete clusters $P_{X|y}$ , which is infeasible for regression due to the infinite slices $P_{X|y}$ . (b) In COD, the continuous conditional distributions are embedded into the RKHS and characterized by finite statistical moments in kernel spaces, which do not rely on specific conditions $y$ . Under the guarantees of distribution embedding property, the conditional alignment is equivalent to the matching on first-order and second-order statistics, i.e., mean embedding operator $\mathcal{U}_{X|Y}$ and covariance operator $\mathcal{C}_{XX|Y}$ . When COD is minimized, conditional alignment is achieved, and the level sets of cross-domain labeling rules are then aligned. + +# 3.3 COD metric for DAR + +Theory for COD. Due to the continuity of label variable, it is impractical to correct conditional shift by aligning marginal distribution of $P_{X|y}$ for every single $y$ . For this inevitable obstacle, an intuitive idea is can we align all $P_{X|y}$ as a whole? i.e., $P_{X|Y}$ , rather than align $P_{X|y}$ one by one? It is more than attractive that if we can characterize $P_{X|Y}$ with a well-defined discrepancy, then all $P_{X|y}$ can be aligned by aligning the whole $P_{X|Y}$ . + +In kernel embedding theory, a distribution can be characterized by a point in RKHS. Then $P_{X|Y}$ can be seen as a point set in RKHS in which every element corresponds to a distribution $P_{X|y}$ for a fixed $y$ . From another perspective, we can consider $P_{X|Y}$ as the distribution in RKHS and measure the cross-domain conditional discrepancy by defining the COD distance on their statistical moments in RKHS. Following the definitions of conditional statistics in pioneer works, i.e., the conditional mean operator as $\mathcal{U}_{X|Y} \coloneqq \mathcal{C}_{XY}\mathcal{C}_{YY}^{-1}$ [48] and conditional covariance operator as $\mathcal{C}_{XX|Y} = \mathcal{C}_{XX} - \mathcal{C}_{XY}\mathcal{C}_{YY}^{-1}\mathcal{C}_{YX}$ [13], we present the definition of COD as follows. + +Definition 2 (COD). Given conditional distributions $P_{X|Y}^{s}$ and $P_{X|Y}^{t}$ , denote $\mathcal{C}_{XX|Y}^{ss}$ and $\mathcal{C}_{XX|Y}^{tt}$ as their corresponding covariance operators. The COD is defined as + +$$ +d _ {\mathrm {C O D}} ^ {2} \left(P _ {X | Y} ^ {s}, P _ {X | Y} ^ {t}\right) = \left\| \mathcal {U} _ {X | Y} ^ {s} - \mathcal {U} _ {X | Y} ^ {t} \right\| _ {\mathcal {H} _ {\mathcal {K}}} ^ {2} + \operatorname {t r} \left(\mathcal {C} _ {X X | Y} ^ {s s} + \mathcal {C} _ {X X | Y} ^ {t t} - 2 \mathcal {C} _ {X X | Y} ^ {s t}\right), \tag {7} +$$ + +where $\mathcal{C}_{XX|Y}^{st} = \sqrt{\sqrt{\mathcal{C}_{XX|Y}^{ss}}\mathcal{C}_{XX|Y}^{tt}\sqrt{\mathcal{C}_{XX|Y}^{ss}}}$ . + +Theoretically, it can be proved that COD indeed admits the metric property on conditional distributions, i.e., two conditional distributions are the same if + +and only if the COD metric between them is zero. An intuitive illustration of COD is shown in Fig. 2. Formally, COD admits two key properties. + +Proposition 1. (a) COD defines a metric between Gaussian measures. + +(b) Let $(\mathcal{X},\mathcal{B}_{\mathcal{X}})$ be the locally compact and Hausdorff measurable space and $k$ be $c_{0}$ -universal kernel. Assuming that $(\phi (X),\psi (Y))$ is a Gaussian random variable in $\mathcal{H}_{\mathcal{X}}\oplus \mathcal{H}_{\mathcal{Y}}$ . For any $P_{X|Y}^{s},P_{X|Y}^{t}\in \mathbb{P}_{\mathcal{X}|\mathcal{Y}}$ , we have + +$$ +d _ {\mathrm {C O D}} ^ {2} \left(P _ {X | Y} ^ {s}, P _ {X | Y} ^ {t}\right) = 0 \Rightarrow P _ {X | Y} ^ {s} = P _ {X | Y} ^ {t}. \tag {8} +$$ + +Empirical Estimation of COD. Based on the above properties, we can perform conditional distribution alignment by minimizing the COD metric between source and target domain. To do this, we shall derive the empirical estimation of COD, which can be obtained with the help of the kernel trick. + +Let $(\mathbf{x}_i^s,\mathbf{y}_i^s)_{i = 1}^{n_s}$ and $(\mathbf{x}_i^t,\mathbf{y}_i^t)_{i = 1}^{n_t}$ be two set of samples drawn i.i.d. from source and target domain. For simplicity, $n_s$ and $n_t$ are both set to $n$ . In kernel method, samples are mapped to RKHS $\mathcal{H}_{\mathcal{X}}\oplus \mathcal{H}_{\mathcal{Y}}$ by the implicit feature map $(\phi ,\psi)$ . Denote the feature map matrices by $(\Phi^s,\Psi^s)$ and $(\Phi^t,\Psi^t)$ , the explicit kernel matrices $\mathbf{K}_X^{ss} = \Phi^s\Phi^{sT}$ , $\mathbf{K}_Y^{ss} = \Psi^s\Psi^{sT}$ are computed as $(\mathbf{K}_X^{ss})_{ij} = k_{\mathcal{X}}(\mathbf{x}_i^s,\mathbf{x}_j^s)$ , $(\mathbf{K}_Y^{ss})_{ij} = k_{\mathcal{Y}}(\mathbf{y}_i^s,\mathbf{y}_j^s)$ , respectively. And so as $\mathbf{K}_X^{tt}$ , $\mathbf{K}_X^{st}$ and $\mathbf{K}_Y^{tt}$ . Let $\mathbf{I}_n$ be the n-dimensional identity matrix, $\mathbf{1}_n$ be the n-dimensional vector with all elements equal to 1. The $n$ -dimensional centering matrix can be defined as $\mathbf{H}_n = \mathbf{I}_n - \frac{1}{n}\mathbf{1}_n\mathbf{1}_n^T$ . Then the centralized kernel matrix is defined as $\mathbf{G} = \mathbf{H}_n\mathbf{K}\mathbf{H}_n$ . + +With above notations, the covariance operator $\mathcal{C}_{XY}$ can be empirically estimated as $\hat{\mathcal{C}}_{XY} = \frac{1}{n} (\pmb {\Phi} - \hat{\mu}_X\pmb {1}^T)(\pmb {\Psi} - \hat{\mu}_Y\pmb {1}^T)^T = \frac{1}{n}\pmb {\Phi}\pmb {H}\pmb {\Psi}^T$ , then the conditional moment statistics can be estimated as + +$$ +\hat {\mathcal {U}} _ {X \mid Y} = (\frac {1}{n} \boldsymbol {\Phi} \mathbf {H} \boldsymbol {\Psi} ^ {T}) (\frac {1}{n} \boldsymbol {\Psi} \mathbf {H} \boldsymbol {\Psi} ^ {T}) ^ {- 1}, \quad \hat {\mathcal {C}} _ {X X \mid Y} = \frac {1}{n} \boldsymbol {\Phi} \mathbf {H} _ {n} \mathbf {A} (\boldsymbol {\Phi} \mathbf {H} _ {n} \mathbf {A}) ^ {T}, \quad (9) +$$ + +where $\mathbf{A}$ satisfies that $\mathbf{B} = \mathbf{A}\mathbf{A}^T$ and $B\triangleq \mathbf{I}_n - \frac{1}{n\varepsilon}\left[\mathbf{G}_Y - \mathbf{G}_Y\left(\mathbf{G}_Y + \varepsilon n\mathbf{I}_n\right)^{-1}\mathbf{G}_Y\right]$ . Based on these estimations, the closed-form empirical estimation of the COD metric can be derived as follows. + +Proposition 2. The empirical Conditional Operator Discrepancy distance is + +$$ +\begin{array}{l} \hat {d} _ {\mathrm {C O D}} ^ {2} = \operatorname {t r} \left(\mathbf {K} _ {Y} ^ {s s} \left(\mathbf {K} _ {Y} ^ {s s}\right) ^ {- 1} \mathbf {K} _ {X} ^ {s s} \left(\mathbf {K} _ {Y} ^ {s s}\right) ^ {- 1}\right) + \operatorname {t r} \left(\mathbf {K} _ {Y} ^ {t t} \left(\mathbf {K} _ {Y} ^ {t t}\right) ^ {- 1} \mathbf {K} _ {X} ^ {t t} \left(\mathbf {K} _ {Y} ^ {t t}\right) ^ {- 1}\right) \\ - 2 \operatorname {t r} \left(\mathbf {K} _ {Y} ^ {t s} \left(\mathbf {K} _ {Y} ^ {s s}\right) ^ {- 1} \mathbf {K} _ {X ^ {s} X ^ {t}} \left(\mathbf {K} _ {Y} ^ {t t}\right) ^ {- 1}\right) \\ + \varepsilon \operatorname {t r} \left[ \mathbf {G} _ {X} ^ {s} \left(\varepsilon n \mathbf {I} _ {n} + \mathbf {G} _ {Y} ^ {s}\right) ^ {- 1} \right] + \varepsilon \operatorname {t r} \left[ \mathbf {G} _ {X} ^ {t} \left(\varepsilon n \mathbf {I} _ {n} + \mathbf {G} _ {Y} ^ {t}\right) ^ {- 1} \right] \tag {10} \\ \left. - \frac {2}{n} \left\| \left(\mathbf {H} _ {n} \mathbf {A} ^ {t}\right) ^ {T} \mathbf {K} _ {X} ^ {t s} \left(\mathbf {H} _ {n} \mathbf {A} ^ {s}\right) \right\| _ {*} \right., \\ \end{array} +$$ + +where $\| \cdot \|_{*}$ is the nuclear norm. Empirically, $\mathbf{K}_Y^{-1}$ is approximated by the regularized formulation $(\mathbf{K}_Y + \lambda \mathbf{I})^{-1}$ . + +Discriminability of empirical COD. As discussed above, the first term of COD measures the distance of the first-order statistic between conditional distributions in RKHS. However, recent study [5,54] shows that the mean matching + +may reduce the discriminability when bringing the two distributions close, which undoubtedly degrades the performance of the algorithm. Therefore, it is worth making an in-depth analysis of the discriminability of COD as well. + +Without loss of generality, we first consider Kronechker Delta kernel on conditioning variable $Y$ , i.e., $k(\mathbf{y}_1,\mathbf{y}_2) = \delta (\mathbf{y}_1,\mathbf{y}_2)$ . More explicitly, $k(\mathbf{y}_1,\mathbf{y}_2)$ equals to 1 if $\mathbf{y}_1 = \mathbf{y}_2$ or 0 otherwise. For empirical scenarios with finite samples, the numbers of samples and observed labels are finite, which implies the number of distinct label $c$ will be obviously no more than sample size $n$ . For convenience, we call two samples are in the same class when their label values are the same. Consequently, the first term in COD can be reformulated as follows. + +Proposition 3. Under the conditions mentioned above, the empirical estimation of $d_{CMMD}$ can be written as + +$$ +\begin{array}{l} \widehat {d} _ {\mathrm {C M M D}} ^ {2} = \sum_ {p = 1} ^ {c} \frac {1}{(\lambda + n _ {p}) ^ {2}} \left(\sum_ {i, j} k (\mathbf {x} _ {(p), i} ^ {s}, \mathbf {x} _ {(p), j} ^ {s})\right) \\ + \sum_ {p = 1} ^ {c} \frac {1}{(\lambda + n _ {p}) ^ {2}} \left(\sum_ {i, j} k \left(\mathbf {x} _ {(p), i} ^ {t}, \mathbf {x} _ {(p), j} ^ {t}\right)\right) \tag {11} \\ - 2 \sum_ {p = 1} ^ {c} \frac {1}{(\lambda + n _ {p}) ^ {2}} \left(\sum_ {i, j} k (\mathbf {x} _ {(p), i} ^ {s}, \mathbf {x} _ {(p), j} ^ {t})\right), \\ \end{array} +$$ + +where $n_i$ denotes the sample-size of $i$ -th label such that $n = \sum_{i=1}^{c} n_i$ , $\mathbf{x}_{(p),i}$ denotes the $i$ -th sample of $p$ -th label. + +Eq. (11) contains three terms. The third term characterizes the cross-domain intra-class similarity. When Eq. (11) is minimized, the third term is maximized so that the cross-domain sample with the same label will be aligned, leading to conditional alignment as expected. However, the first two terms of the equation represent the intra-class similarity of both domains. Minimizing them will decrease the intra-class similarity, which means intra-class samples will be pushed away from each other and the discriminability of representations is degraded. + +Remark 2. Note the Kronechker Delta kernel is defined as + +$$ +k \left(\mathbf {y} _ {1}, \mathbf {y} _ {2}\right) = \left\{ \begin{array}{l} 1 \text {i f} \mathbf {y} _ {1} = \mathbf {y} _ {2} \\ 0 \text {o t h e r w i s e} \end{array} \right.. \tag {12} +$$ + +When it comes to the Gaussian kernel for empirical modeling, the case of $\mathbf{y}_1 = \mathbf{y}_2$ is the same; for other cases, the kernel values can be arbitrarily close to zero, i.e., for any $\epsilon > 0$ , there always width $\sigma_0$ , such that + +$$ +k \left(\mathbf {y} _ {1}, \mathbf {y} _ {2}\right) = \exp \left(- \| \mathbf {y} _ {1} - \mathbf {y} _ {2} \| _ {2} ^ {2} / \sigma_ {0} ^ {2}\right) < \epsilon . \tag {13} +$$ + +Thus, it is generally reasonable to use Gaussian kernel, which admits a similar property to Kronechker Delta kernel, for empirical modeling. + +Remark 3. Note that although we treat the number of distinct labels as finite numbers in mini-batch training, it does not mean class-wise matching can be conducted. In fact, even though the pseudo target label has finite range in minibatch, it still admits the continuous property, which means we can not cluster samples as usually done in classification methods. When two samples have predicted labels with little numerical difference, they can be regarded as samples of the same class in classification settings. Nonetheless, this kind of clustering is infeasible in regression settings since the labels are continuous. Therefore, class-wise alignment is not practical even in mini-batch training. + +With these observation, it is reasonable to reformulate Eq. (11) to promote the discriminability of representations: + +$$ +\begin{array}{l} \widehat {d} _ {\mathrm {C M M D} _ {\mathrm {m o d}}} = - \operatorname {t r} \left(\mathbf {K} _ {Y} ^ {t t} \left(\mathbf {K} _ {Y} ^ {t t}\right) ^ {- 1} \mathbf {K} _ {X} ^ {t t} \left(\mathbf {K} _ {Y} ^ {t t}\right) ^ {- 1}\right) - \operatorname {t r} \left(\mathbf {K} _ {Y} ^ {t t} \left(\mathbf {K} _ {Y} ^ {t t}\right) ^ {- 1} \mathbf {K} _ {X} ^ {t t} \left(\mathbf {K} _ {Y} ^ {t t}\right) ^ {- 1}\right) \\ - 2 \operatorname {t r} \left(\mathbf {K} _ {Y ^ {t} Y ^ {s}} \left(\mathbf {K} _ {Y} ^ {s s}\right) ^ {- 1} \mathbf {K} _ {X ^ {s} X ^ {t}} \left(\mathbf {K} _ {Y} ^ {t t}\right) ^ {- 1}\right) \\ \end{array} +$$ + +Now minimizing Eq. (14) will maximize intra-class similarity regardless of domain, yielding conditional matching with great discriminability. Then the empirical COD is modified as + +$$ +\begin{array}{l} \widehat {d} _ {\mathrm {C O D} _ {\mathrm {m o d}}} = - \operatorname {t r} \left(\mathbf {K} _ {Y} ^ {t t} \left(\mathbf {K} _ {Y} ^ {t t}\right) ^ {- 1} \mathbf {K} _ {X} ^ {t t} \left(\mathbf {K} _ {Y} ^ {t t}\right) ^ {- 1}\right) - \operatorname {t r} \left(\mathbf {K} _ {Y} ^ {t t} \left(\mathbf {K} _ {Y} ^ {t t}\right) ^ {- 1} \mathbf {K} _ {X} ^ {t t} \left(\mathbf {K} _ {Y} ^ {t t}\right) ^ {- 1}\right) \\ - 2 \operatorname {t r} \left(\mathbf {K} _ {Y ^ {t} Y ^ {s}} \left(\mathbf {K} _ {Y} ^ {s s}\right) ^ {- 1} \mathbf {K} _ {X ^ {s} X ^ {t}} \left(\mathbf {K} _ {Y} ^ {t t}\right) ^ {- 1}\right) \\ + \varepsilon \operatorname {t r} \left[ \mathbf {G} _ {X} ^ {s} \left(\varepsilon n \mathbf {I} _ {n} + \mathbf {G} _ {Y} ^ {s}\right) ^ {- 1} \right] + \varepsilon \operatorname {t r} \left[ \mathbf {G} _ {X} ^ {t} \left(\varepsilon n \mathbf {I} _ {n} + \mathbf {G} _ {Y} ^ {t}\right) ^ {- 1} \right] \tag {15} \\ - \frac {2}{n} \left\| \left(\mathbf {H} _ {n} \mathbf {A} ^ {t}\right) ^ {T} \mathbf {K} _ {X} ^ {t s} \left(H _ {n} \mathbf {A} ^ {s}\right) \right\| _ {*} \\ \end{array} +$$ + +# 3.4 COD-based DAR Modeling + +According to theorem (2), when the conditional distributions $P_{X^s|Y}$ and $P_{X^t|Y}$ are aligned, a well-trained source predictor can gain promising performance on target domain. However, when the shift is significant, it is difficult for predictor to produce high-quality pseudo labels at the beginning. For this reason, the model is likely to miss the optimal optimization direction in the early training stage and thus fall into sub-optimal solutions. Therefore, the marginal alignment can improve the reliability of initial pseudo labels, and further benefit the conditional alignment. Overall, the learning objective of COD can be formulated as + +$$ +\min _ {g, h} \mathcal {L} _ {\mathrm {s r c}} + \lambda_ {1} \widehat {d} _ {\mathrm {C O D} _ {\mathrm {m o d}}} + \lambda_ {2} \widehat {d} _ {\mathrm {K G W}}, \tag {16} +$$ + +where $\lambda_1, \lambda_2$ are the trade-off parameters. For regression task, the source risk $\mathcal{L}_{\mathrm{src}}$ is commonly set as Mean Square Error (MSE) over continuous outputs: + +$$ +\mathcal {L} _ {\mathrm {s r c}} = \frac {1}{n ^ {s}} \sum_ {i = 1} ^ {n ^ {s}} \left\| \hat {\mathbf {y}} _ {i} ^ {s} - \mathbf {y} _ {i} ^ {s} \right\| _ {2} ^ {2} \tag {17} +$$ + +where $\hat{\mathbf{y}}_i^s$ is the predicted value of the $i$ -th source sample. Note that the marginal alignment term can be replaced by any other discrepancy metric, while we adopt KGW here to preserve the consistency with OT framework. + +# 4 Experiments + +We evaluate our method with several state-of-the-art domain adaptation methods on three regression benchmark datasets, implementation details and additional illustrations are provided in supplementary material. + +dSprites [19] is a standard 2D synthetic dataset for deep representation learning, which contains three domains each with 737,280 images: Color (C), Noisy (N) and Scream (S). Following [4], we consider three factors for regression tasks: scale and two plane coordinates, while the orientation factor is excluded for its complexity. We evaluate all methods on six transfer tasks: $\mathbf{C} \rightarrow \mathbf{N}$ , $\mathbf{C} \rightarrow \mathbf{S}$ , $\mathbf{N} \rightarrow \mathbf{C}$ , $\mathbf{N} \rightarrow \mathbf{S}$ , $\mathbf{S} \rightarrow \mathbf{C}$ and $\mathbf{S} \rightarrow \mathbf{N}$ . Following previous works [4,38], the sum of Mean Absolute Error (MAE) on all three sub-regression tasks is reported as the evaluation metric. + +Biwi kinect [12] is a real-world dataset containing two domains according to gender: Female $(\mathbf{F})$ with 5874 images and Male $(\mathbf{M})$ with 9804 images. The three factors that can be employed for regression tasks are pitch, yaw and roll. We evaluate all methods on two transfer tasks: $\mathbf{F} \rightarrow \mathbf{M}$ and $\mathbf{M} \rightarrow \mathbf{F}$ . The sum of MAE on three sub-regression tasks is reported. + +MPI3D [16] is a simulation-to-real dataset of 3D objects which consist of three different domain: Toy (T), RealistiC (RC) and ReaL (RL). Each domain contains 1,036,800 images. There are two factors in MPI3D that can be employed for regression tasks: a rotation about a vertical axis and a second rotation about a horizontal axis. We evaluate all methods on six transfer tasks: RL $\rightarrow$ RC, RL $\rightarrow$ T, RC $\rightarrow$ T, RC $\rightarrow$ RL, T $\rightarrow$ RL and T $\rightarrow$ RC. The sum of MAE on two sub-regression tasks is reported. + +Comparison methods. SOTA methods for UDA and DAR are selected: TCA [40], MCD [45], DAN [28], DANN [14], JDOT [8], AFN [58], RSD [4], DARE-GRAM [38]. Following previous works [4,38], Mean Absolute Error (MAE) is used as our evaluation metric across all regression tasks. In comparison experiment, the proposed DAR model in Eq. (16) is denoted as COD for convenience. + +# 4.1 Results and Analysis + +Comparison on dSprites. As shown in Tab. 1, COD achieves the best performance among all competing methods in nearly all tasks of dSprites. In $\mathbf{C} \rightarrow \mathbf{N}$ and $\mathbf{C} \rightarrow \mathbf{S}$ , in which the Resnet-18 and UDA methods exhibit terrible performance because of the significant domain gap. Although SOTA DAR methods have greatly boosted the performance compared with classification methods, COD still achieves a significant improvement, reducing the error to only $40\%$ of that produced by SOTA methods. These results demonstrate that the conditional invariant representations learned by COD indeed ensure better discriminability. Comparison on Biwi Kinect. Since Biwi Kinect is the closest to real-world scenarios among the three benchmark datasets, the performance improvement of model is more challenging. In Tab. 2, the performance gap between UDA methods and DAR methods is relatively smaller, which implies the SOTA DAR + +Table 1: Domain adaptation regression results (MAE) on dSprites (ResNet-18). + +
MethodC→NC→SN→CN→SS→CS→NAvg.
Resnet-180.940.900.160.650.080.260.498
TCA0.940.870.190.660.100.230.498
MCD0.810.810.170.650.070.190.450
JDOT0.860.790.190.640.100.230.468
AFN1.000.960.160.620.080.320.523
DAN0.700.770.120.500.060.110.377
DANN0.470.460.160.650.050.100.315
RSD0.310.310.120.530.070.080.237
DARE-GRAM0.300.200.110.250.050.070.164
COD0.120.160.100.230.080.120.134
+ +Table 2: Domain adaptation regression results (MAE) on Biwi Kinect (ResNet-18). + +
MethodM→FF→MAvg.
Resnet-180.290.380.335
TCA0.310.390.350
MCD0.310.370.340
JDOT0.290.390.340
AFN0.320.410.365
DAN0.280.370.325
DANN0.300.370.335
RSD0.260.300.280
DARE-GRAM0.230.290.260
COD0.200.210.205
+ +methods with marginal alignment indeed show limitations in the difficult transfer tasks. In such a scenario, the sufficiency of condition alignment is directly validated, where the MAE of COD is significantly lower than other SOTA methods. Thus, for real-world regression applications with complex shift, the COD model indeed shows superiority in both theoretical and methodological aspects. Comparison on MPI3D. In MPI3D, the resolution remains the same with dSprites, while the main challenge is the large sample size. In Tab. 3, the MAE of COD is competitive with the SOTA method DARE-GRAM in the first four tasks, while higher in the last two tasks. However, note that COD still achieves the second-best result in averaged MAE, and obtains the lowest MAE in task RL $\rightarrow$ T. In conclusion, the overall MAE of COD on three benchmarks is still lower, which demonstrates COD is generally superior to SOTA methods. + +Ablation study. The ablation experiment was conducted on the most challenging tasks from all three datasets: $\mathbf{C} \rightarrow \mathbf{S}$ and $\mathbf{N} \rightarrow \mathbf{S}$ from dSprites, $\mathbf{T} \rightarrow \mathbf{R}\mathbf{L}$ and $\mathbf{T} \rightarrow \mathbf{RC}$ from MPI3D, and both $\mathbf{M} \rightarrow \mathbf{F}$ and $\mathbf{F} \rightarrow \mathbf{M}$ from Biwi Kinect. The major components of proposed method, i.e., MSE objective $\mathcal{L}_{src}$ , KGW distance $d_{\mathrm{KGW}}$ , COD distance $d_{\mathrm{COD}}$ and modified COD $d_{\mathrm{COD}_{mod}}$ , are evaluated, where the results are shown in Tab. 4. + +Table 3: Domain adaptation regression results (MAE) on MPI3D (ResNet-18). + +
MethodRL→RCRL→TRC→RLRC→TT→RLT→RCAvg.
Resnet-180.170.440.190.450.510.500.377
TCA0.170.420.190.420.500.500.373
MCD0.130.400.150.450.520.500.358
JDOT0.160.410.160.410.470.470.353
AFN0.180.450.200.460.530.530.390
DAN0.120.350.120.270.400.410.278
DANN0.090.240.110.410.480.370.283
RSD0.090.190.080.150.360.360.205
DARE-GRAM0.090.150.100.140.240.240.160
COD0.100.110.130.150.370.290.192
+ +Table 4: Ablation study on different metrics for DAR. + +
MSEObjectivesdSpritesBiwi KinectMPI3D
dKGWdCODdCODmodC→SN→ST→RLT→RCM→FF→M
✓0.900.650.290.380.510.50
✓✓0.180.260.210.250.410.34
✓✓0.260.450.210.220.400.31
✓✓✓0.170.250.210.220.410.29
✓✓✓0.160.230.200.210.370.29
+ +From the $2^{\mathrm{nd}}$ and $3^{\mathrm{rd}}$ rows, it can be observed that both OT-based metrics show promising results on DAR datasets, where the conditional alignment with COD shows better performance than marginal alignment with KGW on MPI3D and Biwi Kinect datasets. On the other hand, since dSprites induce more difficult transfer problems, the baseline performance is generally lower, and the pseudo labels for conditional alignment are less reliable. Therefore, the combination of KGW and COD is necessary to ensure higher performance on challenging tasks, where the results in the $4^{\mathrm{th}}$ row validate the mutually beneficial relation between these two objectives. Finally, the $5^{\mathrm{th}}$ row shows the effectiveness of rethinking the discriminability of COD, which implies the modifications on the moment statistics indeed ensure lower MAE values on all datasets. Overall, these results demonstrate that conditional alignment is indeed crucial for successful DAR, and the theoretical results and proposed method are generally valid. + +Feature Visualization. t-SNE [34] is employed to visualize the 2-D features of different alignments on dSprites $\mathrm{C}\rightarrow \mathrm{N}$ task. (a) Before adaptation, distributions of source and target domain differ severely, as shown in Fig. 3(a), which implies predictor is infeasible for the target samples with a significant domain gap. (b) After marginal distribution alignment via the SOTA DARE-GRAM method, source and target representations are aligned globally, and the cross-domain distributions are overlapped on the whole. However, there are still misaligned local patterns where source and target samples of different colors are matched. + +![](images/afb46b230699af1112d79d9ece81ccaee0e0e8721d8ce341d159b9a7ff6c1c47.jpg) +Fig. 3: t-SNE visualization of learned representations. Label values are denoted by color gradients. Eight label values are selected from the range of the variable for visualization. $+$ : source samples; $o$ : target samples. + +Therefore, it is hard to fit clear level sets in Fig. 3 (b). (c) After conditional distribution alignment via COD, source and target distributions are aligned not only globally but also locally. As shown in Fig. 3 (c), representations are matched according to label values so the change of color shows a clear gradient and level + +sets suitable for both domains can be fitted. + +Hyper-parameter. We investigate the sensitivity of hyper-parameters on dSprites $\mathrm{C}\rightarrow \mathrm{N}$ MAE values by varying $\lambda_{1}$ and $\lambda_{2}$ ,i.e., parameters of modified COD and KGW, are shown in Fig. 4. The results show that the performance of the COD-based DAR model is generally robust to the different choice of parameters. In fact, the MAE values of the COD-based model are stable in random experiments, which implies the proposed method is both robust from the view of randomness and parameter setting. In conclusion, the sta + +![](images/fde948489ae7c2d8f5bd51ef3ce4e356503a10d0742673fe58a23006019ab258.jpg) +Fig.4: MAE values under different settings of hyper-parameters. + +ble performance validates that the proposed model is also empirically reliable. + +# 5 Conclusion + +In this work, we studied the limitations of existing DAR research, i.e., the misaligned intrinsic discriminant structure and negative transfer of marginal distribution matching framework. To deal with the challenges of conditional shift with continuous output space, we provide feasible solutions from both theoretical and methodological views, where the main results show that: 1) the cross-domain generalization regression error can be sufficiently bounded by the conditional discrepancy on continuous labels; 2) the proposed COD serves as a well-defined metric for conditional discrepancy measures and empirical modeling. With these guarantees, a COD-based DAR method is proposed with a discriminability enhancement mechanism. Numerical validations on standard DAR datasets validate the superiority of COD. + +# Acknowledgements + +This work is supported in part by National Natural Science Foundation of China (Grant No. 62376291), in part by Guangdong Basic and Applied Basic Research Foundation (2023B1515020004), in part by Science and Technology Program of Guangzhou (2024A04J6413), in part by the Fundamental Research Funds for the Central Universities, Sun Yat-sen University (24xkjc013), in part by Guangdong Province Key Laboratory of Computational Science at Sun Yat-sen University (2020B1212060032), and in part by Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education. + +# References + +1. Bao, Y., Liu, Y., Wang, H., Lu, F.: Generalizing gaze estimation with rotation consistency. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4207-4216 (2022) +2. Ben-David, S., Blitzer, J., Crammer, K., Kulesza, A., Pereira, F., Vaughan, J.W.: A theory of learning from different domains. Mach. Learn. 79, 151-175 (2010) +3. Chen, L., Chen, H., Wei, Z., Jin, X., Tan, X., Jin, Y., Chen, E.: Reusing the task-specific classifier as a discriminator: Discriminator-free adversarial domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7181-7190 (2022) +4. Chen, X., Wang, S., Wang, J., Long, M.: Representation subspace distance for domain adaptation regression. In: Int. Conf. Mach. Learn. pp. 1749-1759 (2021) +5. Chen, Y., Song, S., Li, S., Wu, C.: A graph embedding framework for maximum mean discrepancy-based domain adaptation algorithms. IEEE Trans. Image Process. 29, 199-213 (2019) +6. Tachet des Combes, R., Zhao, H., Wang, Y.X., Gordon, G.J.: Domain adaptation with conditional distribution matching and generalized label shift. In: Adv. Neural Inform. Process. Syst. vol. 33, pp. 19276-19289 (2020) +7. Cortes, C., Mohri, M.: Domain adaptation in regression. In: ICALT. pp. 308-323 (2011) +8. Courty, N., Flamary, R., Habrard, A., Rakotomamonjy, A.: Joint distribution optimal transportation for domain adaptation. Adv. Neural Inform. Process. Syst. 30 (2017) +9. Courty, N., Flamary, R., Tuia, D.: Domain adaptation with regularized optimal transport. In: ECML PKDD. pp. 274-289 (2014) +10. Damodaran, B.B., Kellenberger, B., Flamary, R., Tuia, D., Courty, N.: Deepjdot: Deep joint distribution optimal transport for unsupervised domain adaptation. In: Eur. Conf. Comput. Vis. pp. 447-463 (2018) +11. Dhaini, M., Berar, M., Honeine, P., Van Exem, A.: Unsupervised domain adaptation for regression using dictionary learning. Knowl.-Based Syst. 267, 110439 (2023) +12. Fanelli, G., Dantone, M., Gall, J., Fossati, A., Van Gool, L.: Random forests for real time 3d face analysis. Int. J. Comput. Vis. **101**, 437–458 (2013) +13. Fukumizu, K., Bach, F.R., Jordan, M.I.: Kernel dimension reduction in regression. Ann. Stat. 37(4), 1871-1905 (2009) +14. Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., March, M., Lempitsky, V.: Domain-adversarial training of neural networks. J. Mach. Learn. Res. 17(59), 1-35 (2016) + +15. Geng, X., Zhou, Z.H., Smith-Miles, K.: Automatic age estimation based on facial aging patterns. IEEE Trans. Pattern Anal. Mach. Intell. 29(12), 2234-2240 (2007) +16. Gondal, M.W., Wuthrich, M., Miladinovic, D., Locatello, F., Breidt, M., Volchkov, V., Akpo, J., Bachem, O., Scholkopf, B., Bauer, S.: On the transfer of inductive bias from simulation to the real world: a new disentanglement dataset. Adv. Neural Inform. Process. Syst. 32 (2019) +17. Gong, B., Shi, Y., Sha, F., Grauman, K.: Geodesic flow kernel for unsupervised domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2066-2073 (2012) +18. Goswami, S., Kontolati, K., Shields, M.D., Karniadakis, G.E.: Deep transfer operator learning for partial differential equations under conditional shift. Nat. Mach. Intell. 4(12), 1155-1164 (2022) +19. Higgins, I., Matthew, L., Pal, A., Burgess, C., Glorot, X., Botvinick, M., Mohamed, S., Lerchner, A.: beta-vae: Learning basic visual concepts with a constrained variational framework. In: Int. Conf. Learn. Represent. (2016) +20. Hoffman, J., Tzeng, E., Park, T., Zhu, J.Y., Isola, P., Saenko, K., Efros, A., Darrell, T.: Cycada: Cycle-consistent adversarial domain adaptation. In: Int. Conf. Mach. Learn. pp. 1989-1998 (2018) +21. Hu, J., Zhong, H., Yang, F., Gong, S., Wu, G., Yan, J.: Learning unbiased transferability for domain adaptation by uncertainty modeling. In: Eur. Conf. Comput. Vis. pp. 223-241 (2022) +22. Jiang, J., Ji, Y., Wang, X., Liu, Y., Wang, J., Long, M.: Regressive domain adaptation for unsupervised keypoint detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6780-6789 (2021) +23. Kirchmeyer, M., Rakotomamonjy, A., de Bezenac, E., Gallinari, P.: Mapping conditional distributions for domain adaptation under generalized target shift. In: Int. Conf. Learn. Represent. (2022) +24. Lee, C.Y., Batra, T., Baig, M.H., Ulbricht, D.: Sliced wasserstein discrepancy for unsupervised domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10285-10295 (2019) +25. Li, B., Wang, Y., Zhang, S., Li, D., Keutzer, K., Darrell, T., Zhao, H.: Learning invariant representations and risks for semi-supervised domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1104-1113 (2021) +26. Li, X., Li, J., Zuo, L., Zhu, L., Shen, H.T.: Domain adaptive remaining useful life prediction with transformer. IEEE Trans. Instrum. 71, 1-13 (2022) +27. Li, Z., Zhang, K., Liu, Y., Zou, Y., Ding, G.: A novel remaining useful life transfer prediction method of rolling bearings based on working conditions common benchmark. IEEE Trans. Instrum. 71, 1-9 (2022) +28. Long, M., Cao, Y., Wang, J., Jordan, M.: Learning transferable features with deep adaptation networks. In: Int. Conf. Mach. Learn. pp. 97-105 (2015) +29. Long, M., Cao, Z., Wang, J., Jordan, M.I.: Conditional adversarial domain adaptation. Adv. Neural Inform. Process. Syst. 31 (2018) +30. Long, M., Zhu, H., Wang, J., Jordan, M.I.: Deep transfer learning with joint adaptation networks. In: Int. Conf. Mach. Learn. pp. 2208-2217 (2017) +31. Luo, Y.W., Ren, C.X.: Conditional bures metric for domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 13989-13998 (2021) +32. Luo, Y.W., Ren, C.X.: Mot: Masked optimal transport for partial domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3531-3540 (2023) +33. Luo, Y.W., Ren, C.X., Dai, D.Q., Yan, H.: Unsupervised domain adaptation via discriminative manifold propagation. IEEE Trans. Pattern Anal. Mach. Intell. 44(3), 1653-1669 (2022) + +34. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. J. Mach. Learn. Res. 9(11) (2008) +35. Mansour, Y., Mohri, M., Rostamizadeh, A.: Domain adaptation: Learning bounds and algorithms. arXiv preprint arXiv:0902.3430 (2009) +36. Mao, W., Ge, Y., Shen, C., Tian, Z., Wang, X., Wang, Z., den Hengel, A.v.: Poseur: Direct human pose regression with transformers. In: Eur. Conf. Comput. Vis. pp. 72-88 (2022) +37. de Mathelin, A., Richard, G., Deheeger, F., Mougeot, M., Vayatis, N.: Adversarial weighting for domain adaptation in regression. In: ICTAI. pp. 49-56 (2021) +38. Nejjar, I., Wang, Q., Fink, O.: Dare-gram: Unsupervised domain adaptation regression by aligning inverse gram matrices. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 11744–11754 (2023) +39. Nikzad-Langerodi, R., Zellinger, W., Saminger-Platz, S., Moser, B.A.: Domain adaptation for regression under beer-lambert's law. Knowl.-Based Syst. 210, 106447 (2020) +40. Pan, S.J., Tsang, I.W., Kwok, J.T., Yang, Q.: Domain adaptation via transfer component analysis. IEEE Trans. Neural Netw. 22(2), 199-210 (2010) +41. Pei, Z., Cao, Z., Long, M., Wang, J.: Multi-adversarial domain adaptation. In: AAAI. vol. 32 (2018) +42. Peng, R., Wang, R., Wang, Z., Lai, Y., Wang, R.: Rethinking depth estimation for multi-view stereo: A unified representation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8645-8654 (2022) +43. Quinonero-Candela, J., Sugiyama, M., Schwaighofer, A., Lawrence, N.D.: Dataset shift in machine learning. Mit Press (2008) +44. Ren, C.X., Luo, Y.W., Dai, D.Q.: Buresnet: Conditional bures metric for transferable representation learning. IEEE Trans. Pattern Anal. Mach. Intell. 45(4), 4198-4213 (2023) +45. Saito, K., Watanabe, K., Ushiku, Y., Harada, T.: Maximum classifier discrepancy for unsupervised domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3723-3732 (2018) +46. Shen, J., Qu, Y., Zhang, W., Yu, Y.: Wasserstein distance guided representation learning for domain adaptation. In: AAAI. vol. 32 (2018) +47. Singh, A., Chakraborty, S.: Deep domain adaptation for regression. Development and Analysis of Deep Learning Architectures pp. 91-115 (2020) +48. Song, L., Huang, J., Smola, A., Fukumizu, K.: Hilbert space embeddings of conditional distributions with applications to dynamical systems. In: Int. Conf. Mach. Learn. pp. 961-968 (2009) +49. Sun, B., Feng, J., Saenko, K.: Return of frustratingly easy domain adaptation. In: AAAI. vol. 30 (2016) +50. Tang, H., Jia, K.: Discriminative adversarial domain adaptation. In: AAAI. vol. 34, pp. 5940-5947 (2020) +51. Teshima, T., Sato, I., Sugiyama, M.: Few-shot domain adaptation by causal mechanism transfer. In: Int. Conf. Mach. Learn. pp. 9458-9469 (2020) +52. Tzeng, E., Hoffman, J., Saenko, K., Darrell, T.: Adversarial discriminative domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7167-7176 (2017) +53. Wang, H., Shen, T., Zhang, W., Duan, L.Y., Mei, T.: Classes matter: A fine-grained adversarial approach to cross-domain semantic segmentation. In: Eur. Conf. Comput. Vis. pp. 642-659 (2020) +54. Wang, W., Li, H., Ding, Z., Nie, F., Chen, J., Dong, X., Wang, Z.: Rethinking maximum mean discrepancy for visual domain adaptation. IEEE Trans. Neural Netw. Learn. Syst. (2021) + +55. Wang, Y., Jiang, Y., Li, J., Ni, B., Dai, W., Li, C., Xiong, H., Li, T.: Contrastive regression for domain adaptation on gaze estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19376-19385 (2022) +56. Wei, P., Ke, Y., Ong, Y.S., Ma, Z.: Adaptive transfer kernel learning for transfer gaussian process regression. IEEE Trans. Pattern Anal. Mach. Intell. (2022) +57. Wu, J., He, J., Wang, S., Guan, K., Ainsworth, E.: Distribution-informed neural networks for domain adaptation regression. Adv. Neural Inform. Process. Syst. 35, 10040-10054 (2022) +58. Xu, R., Li, G., Yang, J., Lin, L.: Larger norm more transferable: An adaptive feature norm approach for unsupervised domain adaptation. In: Int. Conf. Comput. Vis. pp. 1426-1435 (2019) +59. Zhang, Z., Wang, M., Nehorai, A.: Optimal transport in reproducing kernel hilbert spaces: Theory and applications. IEEE Trans. Pattern Anal. Mach. Intell. 42(7), 1741-1754 (2019) +60. Zhao, H., Combes, R.T.D., Zhang, K., Gordon, G.: On learning invariant representations for domain adaptation. In: Int. Conf. Mach. Learn. pp. 7523-7532 (2019) +61. Zhao, H., Coston, A., Adel, T., Gordon, G.J.: Conditional learning of fair representations. In: Int. Conf. Learn. Represent. (2020) +62. Zhao, H., Dan, C., Aragam, B., Jaakkola, T.S., Gordon, G.J., Ravikumar, P.: Fundamental limits and tradeoffs in invariant representation learning. J. Mach. Learn. Res. 23(340), 1-49 (2022) \ No newline at end of file diff --git a/2024/COD_ Learning Conditional Invariant Representation for Domain Adaptation Regression/images.zip b/2024/COD_ Learning Conditional Invariant Representation for Domain Adaptation Regression/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fd08a783d5fe2998c275c1685d7905a7015e5de3 --- /dev/null +++ b/2024/COD_ Learning Conditional Invariant Representation for Domain Adaptation Regression/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d09d11c2e1fa46f9507d885cc9465fe717bf6c9e22ad0f21a39bff6a65f1606b +size 476825 diff --git a/2024/COD_ Learning Conditional Invariant Representation for Domain Adaptation Regression/layout.json b/2024/COD_ Learning Conditional Invariant Representation for Domain Adaptation Regression/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cb9a7c2249939f42b526407eb6eea9a934f09571 --- /dev/null +++ b/2024/COD_ Learning Conditional Invariant Representation for Domain Adaptation Regression/layout.json @@ -0,0 +1,11566 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 170, + 111, + 444, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 111, + 444, + 166 + ], + "spans": [ + { + "bbox": [ + 170, + 111, + 444, + 166 + ], + "type": "text", + "content": "COD: Learning Conditional Invariant Representation for Domain Adaptation Regression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 173, + 186, + 441, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 186, + 441, + 199 + ], + "spans": [ + { + "bbox": [ + 173, + 186, + 441, + 199 + ], + "type": "text", + "content": "Hao-Ran Yang1, Chuan-Xian Ren1*, and You-Wei Luo1,2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 171, + 208, + 441, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 208, + 441, + 242 + ], + "spans": [ + { + "bbox": [ + 171, + 208, + 441, + 242 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 171, + 208, + 441, + 242 + ], + "type": "text", + "content": " School of Mathematics, Sun Yat-Sen University, China \n" + }, + { + "bbox": [ + 171, + 208, + 441, + 242 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 171, + 208, + 441, + 242 + ], + "type": "text", + "content": " School of Mathematics, Jiaying University, China \n{yanghr26@mail2, rchuanx@mail, luoyw28@mail2}.sysu.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 274, + 452, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 274, + 452, + 483 + ], + "spans": [ + { + "bbox": [ + 160, + 274, + 452, + 483 + ], + "type": "text", + "content": "Abstract. Aiming to generalize the label knowledge from a source domain with continuous outputs to an unlabeled target domain, Domain Adaptation Regression (DAR) is developed for complex practical learning problems. However, due to the continuity problem in regression, existing conditional distribution alignment theory and methods with discrete prior, which are proven to be effective in classification settings, are no longer applicable. In this work, focusing on the feasibility problems in DAR, we establish the sufficiency theory for the regression model, which shows the generalization error can be sufficiently dominated by the cross-domain conditional discrepancy. Further, to characterize conditional discrepancy with continuous conditioning variable, a novel Conditional Operator Discrepancy (COD) is proposed, which admits the metric property on conditional distributions via the kernel embedding theory. Finally, to minimize the discrepancy, a COD-based conditional invariant representation learning model is proposed, and the reformulation is derived to show that reasonable modifications on moment statistics can further improve the discriminability of the adaptation model. Extensive experiments on standard DAR datasets verify the validity of theoretical results and the superiority over SOTA DAR methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 493, + 453, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 493, + 453, + 515 + ], + "spans": [ + { + "bbox": [ + 160, + 493, + 453, + 515 + ], + "type": "text", + "content": "Keywords: Domain Adaptation Regression, Optimal Transport, Invariant Representation Learning, Kernel Embedding Theory." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 537, + 230, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 537, + 230, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 537, + 230, + 550 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "content": "Deep learning has shown a promising ability for fitting and achieved great success in various application tasks. However, large-scale data with sufficient annotations plays a vital role in training high-quality deep models while the label annotations for unlabeled data are usually expensive. Therefore, aiming to reduce training costs by transferring knowledge, Unsupervised Domain Adaptation (UDA) [2,14,17,28] has received increasing attention in recent years. Basically, UDA aims to transfer task-related information from a labeled source domain" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 236, + 665 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "content": "to a target domain without annotations, where the domain gap is characterized as the distribution shift across domains [43]. Following this line, considerable methods have been proposed to alleviate the UDA problem by shift correction and distribution alignment, such as adversarial training for cross-domain samples [14, 52] and explicitly distribution matching by statistical discrepancy minimization [28, 49]. From a theoretical perspective, recent literature [6, 60] show that the marginal distribution " + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "inline_equation", + "content": "P_{X}" + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "content": " with covariate shift assumption is insufficient to characterize the cross-domain generalization error, while the conditional distribution alignment indeed serves as a sufficient condition to ensure small error and successful transfer. Mathematically, when the conditional distributions " + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "inline_equation", + "content": "P_{X|y}" + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "content": " for every " + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 116, + 482, + 344 + ], + "type": "text", + "content": " are aligned, a well-trained source predictor can perform well on the target domain, which is described in Fig. 1. Inspired by theoretical advances, recent works [23, 41, 45] consider the statistical methods for conditional alignment, and show promising performance in mitigating negative transfer and improving recognition accuracy. However, most of the existing results are designed for classification settings with discrete prior on label variables, which induces the infeasibility for regression problems. Consequently, it restricts the application of UDA methods for wider real-world problems, e.g., human pose estimation [36], product condition monitoring [27], depth estimation [42]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 348, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 482, + 552 + ], + "type": "text", + "content": "Considering the dilemma of UDA, pioneer works have explored the Domain Adaptation Regression (DAR) framework for solving the distribution shift in continuous label scenarios, where theoretical results [7,35] and correction methods [15,47,51] are provided. To further relax the requirement of labeled training target data in pioneer DAR works, advanced works [4,38] propose unsupervised DAR methods to learn transferable regression models without the help of labeled target data. Nonetheless, existing DAR theory and methods are directly facilitated by the covariate shift framework, where the marginal distribution alignment is considered. Therefore, it inevitably induces the negative transfer and misaligned local structure, which are shown in the results in classification settings [6,60]. In fact, there is a significant obstacle for modeling conditional distribution alignment in regression scenarios, i.e., the conditional shift correction induces infinite matching problems for " + }, + { + "bbox": [ + 130, + 348, + 482, + 552 + ], + "type": "inline_equation", + "content": "P_{X|Y=y}" + }, + { + "bbox": [ + 130, + 348, + 482, + 552 + ], + "type": "text", + "content": " on all points " + }, + { + "bbox": [ + 130, + 348, + 482, + 552 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 348, + 482, + 552 + ], + "type": "text", + "content": ". The continuous label variable with infinite possible outcomes makes it infeasible to directly apply the existing conditional alignment methods for DAR. Thus, it is vital to understand the continuity problem in regression, and develop the generalization error theory and statistical methodology for DAR scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "In this work, we focus on the theory guarantee and conditional alignment methodology for successful DAR. Theoretically, we explore the generalization upper bound for continuous output scenarios and show the cross-domain joint error can be sufficiently upper-bounded by the conditional discrepancy. It ensures the sufficiency of conditional alignment still holds for regression scenarios and provides reliable guarantees for theory-driven modeling. Methodologically, to deal with the continuous problem in conditional alignment, we treat the conditional distribution as finite statistical moments in Hilbert space. By exploring the distribution embedding theory of Reproducing Kernel Hilbert Space (RKHS), a novel" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "H. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 198, + 114, + 418, + 135 + ], + "blocks": [ + { + "bbox": [ + 198, + 114, + 418, + 135 + ], + "lines": [ + { + "bbox": [ + 198, + 114, + 418, + 135 + ], + "spans": [ + { + "bbox": [ + 198, + 114, + 418, + 135 + ], + "type": "image", + "image_path": "16737da332bfa6f9c5df6a89905a35a2c74bdd50860d8c1d07db50c0d953f06a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 170, + 136, + 254, + 218 + ], + "blocks": [ + { + "bbox": [ + 170, + 136, + 254, + 218 + ], + "lines": [ + { + "bbox": [ + 170, + 136, + 254, + 218 + ], + "spans": [ + { + "bbox": [ + 170, + 136, + 254, + 218 + ], + "type": "image", + "image_path": "4c86c9bb133f0858a85908027199ac9dc5e091a7dc4256d376478eecd27b60ca.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 228, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 338 + ], + "type": "text", + "content": "Fig. 1: Illustration of the conditional shift in regression setting, where label value " + }, + { + "bbox": [ + 130, + 228, + 482, + 338 + ], + "type": "inline_equation", + "content": "y \\in [0,1]" + }, + { + "bbox": [ + 130, + 228, + 482, + 338 + ], + "type": "text", + "content": " and color gradients imply the continuity. (a) Before adaptation, distribution shift exists between source and target representations so the predictor trained on source domain cannot generalize to target domain. (b) After marginal distribution alignment, distributions of source and target domain are globally aligned. However, representations with different labels may be falsely aligned across domains, leading to significant inconsistency between cross-domain labeling rules. Thus, level sets provided by the source predictor are not suitable for target representations. (c) After conditional distribution alignment, label-wise matching is achieved, where the level sets of cross-domain labeling rules are consistent and the source predictor will be reliable." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 268, + 136, + 356, + 217 + ], + "blocks": [ + { + "bbox": [ + 268, + 136, + 356, + 217 + ], + "lines": [ + { + "bbox": [ + 268, + 136, + 356, + 217 + ], + "spans": [ + { + "bbox": [ + 268, + 136, + 356, + 217 + ], + "type": "image", + "image_path": "204bbfbb086977bb66063a36f96f52e1e60519e6ce4dd6d841d8c2085d014f69.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 362, + 137, + 445, + 217 + ], + "blocks": [ + { + "bbox": [ + 362, + 137, + 445, + 217 + ], + "lines": [ + { + "bbox": [ + 362, + 137, + 445, + 217 + ], + "spans": [ + { + "bbox": [ + 362, + 137, + 445, + 217 + ], + "type": "image", + "image_path": "18e436871d05a6c11880c8428d53d8e409dc9d0bebc9080b485a7915da558ec2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 361, + 482, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 458 + ], + "type": "text", + "content": "conditional metric called Conditional Operator Discrepancy (COD) is proposed, which can be interpreted as the conditional Wasserstein distance in RKHS. With the sufficiency theory and metric-based learning principle for regression, a COD-based conditional invariant representation learning method is proposed, which aims to correct the conditional shift on the pushforward measures in latent representation space. By reformulating the empirical COD estimation, a modified principle is developed to further improve the discriminability during knowledge transfer. Generally, our contributions can be summarized as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 466, + 480, + 573 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 138, + 466, + 480, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 466, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 466, + 480, + 501 + ], + "type": "text", + "content": "- The sufficiency theory for the conditional shift framework is established for DAR settings, which shows the cross-domain conditional discrepancy on continuous label variable serves as the major term in generalization error." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 502, + 480, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 502, + 480, + 537 + ], + "spans": [ + { + "bbox": [ + 137, + 502, + 480, + 537 + ], + "type": "text", + "content": "- A novel COD metric for characterizing conditional distribution discrepancy with continuous and infinite conditioning variable is explored, providing a feasible solution for conditional discrepancy optimization in DAR." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 538, + 480, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 538, + 480, + 573 + ], + "spans": [ + { + "bbox": [ + 137, + 538, + 480, + 573 + ], + "type": "text", + "content": "- A COD-based representation learning model with reformulation on moment statistics is proposed, which admits discriminability on task-related knowledge and shows superior performance in extensive experiments." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 592, + 237, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 592, + 237, + 604 + ], + "spans": [ + { + "bbox": [ + 132, + 592, + 237, + 604 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Domain Adaptation Classification. Based on the distribution shift assumption, most of the UDA methods can be roughly categorized into three types: methods that align marginal distribution, conditional distribution and joint distribution. Most UDA methods consider only the marginal distribution align-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 192, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 192, + 91, + 448, + 102 + ], + "type": "text", + "content": "COD: Learning Conditional Invariant Representation for DAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 307 + ], + "type": "text", + "content": "ment. Discrepancy-based methods attempt to directly minimize the distribution discrepancy and different types of discrepancy measurement greatly enrich this type of methodology. For example, these methods adopt moment matching [28, 49] or extension under manifold framework [17, 33]. In particular, Optimal transport (OT) and Wasserstein distance are introduced to UDA as a sound measurement for domain shift, based on which a large number of methods have been proposed [8-10, 24, 46]. Zhang et al. [59] further develop both theory and methodology to RKHS and Luo et al. [31] extends Kernel Bures metric to conditional distribution alignment. Adversarial-based methods [14, 20, 21] learn domain-invariant features via two-player games. Specifically, they alternatively optimize the feature generator and the domain discriminator to achieve domain confusion. In order to obtain better performance, advanced methods have explored how to utilize the label information to achieve conditional distribution alignment [3, 32, 41, 44, 45, 50, 53, 61] or joint distribution alignment [29, 30]. However, UDA methods for shift correction are proposed for discrete and finite conditioning variables, which are inapplicable for regression." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 308, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 559 + ], + "type": "text", + "content": "Domain Adaptation Regression. DA for regression has received relatively little attention compared with classification. Mansour et al. [35] and Cortes et al. [7] introduce theoretical analysis for DA in regression scenarios. Zhao et al. [62] discuss the tradeoffs between accuracy and invariance in learning invariant representations in both classification and regression settings. A range of methods in the shallow regime are proposed to tackle the DAR problem. These methods can be roughly categorized into two types: importance weighting [15,37] and learning domain-invariant representation [8,39,40]. Besides, some attempts have been made in deep regime to tackle regression problem [25,47,51,56]. Unfortunately, most of them belong to semi-supervised DA, which means they need to use a small number of target domain labels. Recently, deep UDA methods for regression that get rid of target label are proposed [4,11,38,57]. RSD [4] states that deep neural networks are less robust to feature scaling in regression problems and aligning the deep features directly will impede task performance for altering feature scale. Thus, a strategy of aligning orthogonal bases of both domains is adopted in RSD. Based on RSD, DARE-GRAM [38] aligns the inverse Gram Matrix of deep features of both domains, and ensures promising improvements in accuracy and stability. Meanwhile, several DAR methods for specific tasks have been explored [1,18,22,26,27,55]. However, the theory and method for conditional shift correction are still unexplored for DAR scenarios, which hinders the in-depth reasoning and development for practical modeling." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 575, + 345, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 575, + 345, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 575, + 345, + 590 + ], + "type": "text", + "content": "3 COD: Theory and Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 599, + 221, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 221, + 612 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 221, + 612 + ], + "type": "text", + "content": "3.1 Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": "Setup. In this paper, we discuss UDA method dealing with regression problems. Let " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{X}" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{Y}" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " be covariate and label variable taken values on " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbb{P}" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " denotes the set of distributions and subscripts represent the corresponding random variable while the superscript represents the domain, e.g. " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "P_{XY}^{s} \\in \\mathbb{P}_{\\mathcal{X}\\mathcal{Y}}" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "type": "text", + "content": "H. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "implies the joint distribution on source domain. In UDA settings, we are provided with labeled source samples " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "(x^{s},y^{s})\\sim P_{XY}^{s}" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": " and unlabeled target samples " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "(x^{t},y^{t})\\sim P_{XY}^{t}" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": ", which means the target labels are not available during training. Samples of source and target domain follow different distributions, i.e. " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "P_{XY}^{s}\\neq P_{XY}^{t}" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": ". Denote " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "g:X\\to Z" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": " as the representation extractor, " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "h:Z\\to Y" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": " as the predictor, " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\varepsilon_{s}" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\varepsilon_{t}" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": " as the source and target prediction error, respectively. UDA methods aim to generalize the learner " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "h\\circ g" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": " across domains and minimize target risk " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\varepsilon_{t}" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": " by mitigating the distribution shift problem. In contrast to the discrete label space in classification settings, unsupervised DAR discusses UDA problems with continuous labels, which means the number of classes is uncountable, leading to great difficulty in aligning conditional distribution " + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "inline_equation", + "content": "P_{X|Y}" + }, + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": ". The proofs are provided in supplementary materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "text", + "content": "OT in RKHS. Let " + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "inline_equation", + "content": "(\\mathcal{X},\\mathcal{B})" + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "text", + "content": " be a measure space with Borel " + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "text", + "content": "-field " + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "text", + "content": ". Denote " + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "inline_equation", + "content": "(\\mathcal{H}_{\\mathcal{X}},k_{\\mathcal{X}})" + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "text", + "content": " as the RKHS of " + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "text", + "content": ", which is generated by the positive definite kernel " + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "inline_equation", + "content": "k_{\\mathcal{X}}" + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "text", + "content": ". Elements from " + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "text", + "content": " are mapped to " + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathcal{X}}" + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "text", + "content": " via feature map " + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "inline_equation", + "content": "\\phi (x) = k_{\\mathcal{X}}(x,\\cdot)" + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "text", + "content": ", which is assumed to satisfy the reproducing properties " + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "inline_equation", + "content": "\\langle \\phi (x),\\phi (x^{\\prime})\\rangle_{\\mathcal{H}_{\\mathcal{X}}} = k_{\\mathcal{X}}(x,x^{\\prime})" + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "inline_equation", + "content": "\\langle \\phi (x),f\\rangle_{\\mathcal{H}_x} = f(x),\\forall f\\in \\mathcal{H}_x" + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "text", + "content": ". The mean element " + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "inline_equation", + "content": "\\mu_X" + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathcal{X}}" + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "text", + "content": " with law " + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "inline_equation", + "content": "P_{X}" + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "text", + "content": " is given by " + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "inline_equation", + "content": "\\mu_{X} = \\mathbb{E}_{X}[\\phi (X)]" + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "text", + "content": ", and the covariance operator " + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{XX}" + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{XX} = E_X(\\phi (X) - \\mu_X)\\otimes (\\phi (X) - \\mu_X)" + }, + { + "bbox": [ + 130, + 259, + 482, + 343 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 344, + 481, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 481, + 369 + ], + "type": "text", + "content": "For any two distributions " + }, + { + "bbox": [ + 130, + 344, + 481, + 369 + ], + "type": "inline_equation", + "content": "P_X^s, P_X^t \\in \\mathbb{P}_X" + }, + { + "bbox": [ + 130, + 344, + 481, + 369 + ], + "type": "text", + "content": ", the Kantorovitch optimal transport in RKHS " + }, + { + "bbox": [ + 130, + 344, + 481, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathcal{K}}" + }, + { + "bbox": [ + 130, + 344, + 481, + 369 + ], + "type": "text", + "content": " is considered on pushforward measures " + }, + { + "bbox": [ + 130, + 344, + 481, + 369 + ], + "type": "inline_equation", + "content": "\\phi_{\\#}P_X^s" + }, + { + "bbox": [ + 130, + 344, + 481, + 369 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 344, + 481, + 369 + ], + "type": "inline_equation", + "content": "\\phi_{\\#}P_X^t" + }, + { + "bbox": [ + 130, + 344, + 481, + 369 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 378, + 481, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 378, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 146, + 378, + 481, + 403 + ], + "type": "interline_equation", + "content": "d _ {\\mathrm {K W}} ^ {2} \\left(X ^ {s}, X ^ {t}\\right) = \\inf _ {\\pi_ {\\mathcal {K}} \\in \\Pi \\left(\\phi_ {\\#} P _ {X} ^ {s}, \\phi_ {\\#} P _ {X} ^ {t}\\right)} \\int_ {\\mathcal {H} _ {\\mathcal {K}} \\times \\mathcal {H} _ {\\mathcal {K}}} \\left\\| X ^ {s} - X ^ {t} \\right\\| _ {\\mathcal {H} _ {\\mathcal {K}}} ^ {2} d \\pi_ {\\mathcal {K}} \\left(X ^ {s}, X ^ {t}\\right) \\tag {1}", + "image_path": "5e6a4b2c8f495ad0dc3d349ed70de57337e52b4ccea570b0599bd5d5a95a2350.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 411, + 481, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 411, + 481, + 436 + ], + "spans": [ + { + "bbox": [ + 131, + 411, + 481, + 436 + ], + "type": "text", + "content": "Zhang et al. [59] show that if the pushforward measures are Gaussian, the Kernel Gaussian Wasserstein (KGW) distance between them can be written as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 178, + 445, + 481, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 445, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 178, + 445, + 481, + 460 + ], + "type": "interline_equation", + "content": "d _ {\\mathrm {K G W}} ^ {2} \\left(X ^ {s}, X ^ {t}\\right) = \\| u _ {X} ^ {s} - u _ {X} ^ {t} \\| _ {\\mathcal {H} _ {\\kappa}} ^ {2} + t r \\left(\\mathcal {C} _ {X X} ^ {s s} + \\mathcal {C} _ {X X} ^ {t t} - 2 \\mathcal {C} _ {X X} ^ {s t}\\right) \\tag {2}", + "image_path": "49b7edaffca11323f408f0854696c98bf4faffc28f147425faf75935c9c7d949.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 470, + 481, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 481, + 502 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 470, + 481, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{XX}^{st} = \\sqrt{\\sqrt{\\mathcal{C}_{XX}^{ss}}\\mathcal{C}_{XX}^{tt}\\sqrt{\\mathcal{C}_{XX}^{ss}}}" + }, + { + "bbox": [ + 130, + 470, + 481, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 470, + 481, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{XX}^{ss},\\mathcal{C}_{XX}^{tt}" + }, + { + "bbox": [ + 130, + 470, + 481, + 502 + ], + "type": "text", + "content": " are the covariance operators of " + }, + { + "bbox": [ + 130, + 470, + 481, + 502 + ], + "type": "inline_equation", + "content": "P_{X}^{s}" + }, + { + "bbox": [ + 130, + 470, + 481, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 470, + 481, + 502 + ], + "type": "inline_equation", + "content": "P_{X}^{t}" + }, + { + "bbox": [ + 130, + 470, + 481, + 502 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 130, + 470, + 481, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathcal{K}}" + }, + { + "bbox": [ + 130, + 470, + 481, + 502 + ], + "type": "text", + "content": " , respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 506, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 482, + 568 + ], + "type": "text", + "content": "Remark 1. Note that KGW consists of two terms, i.e., the Maximum Mean Discrepancy (MMD) on first-order statistics and kernel Bures metric on second-order statistics. It provides a basic innovation for exploring conditional discrepancy metric by considering the conditional statistics in RKHS, i.e., first-order statistics in literature [48] and second-order statistics in literature [13]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 585, + 359, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 585, + 359, + 597 + ], + "spans": [ + { + "bbox": [ + 131, + 585, + 359, + 597 + ], + "type": "text", + "content": "3.2 Generalization Error Analysis for DAR" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Most of DAR methods focus on aligning the marginal distribution " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "P_{X}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ". However, Zhao et al. [60] have shown that in the presence of label shift, only aligning the marginal distribution does not guarantee satisfying generalization performance on target domain. In fact, [6, 60] have proven the following theorem (1) for classification problems:" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 192, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 192, + 91, + 448, + 102 + ], + "type": "text", + "content": "COD: Learning Conditional Invariant Representation for DAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 141 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 141 + ], + "type": "text", + "content": "Theorem 1 ([60]). For representation variable " + }, + { + "bbox": [ + 132, + 116, + 482, + 141 + ], + "type": "inline_equation", + "content": "Z = g(X)" + }, + { + "bbox": [ + 132, + 116, + 482, + 141 + ], + "type": "text", + "content": ", suppose inequality " + }, + { + "bbox": [ + 132, + 116, + 482, + 141 + ], + "type": "inline_equation", + "content": "d_{JS}(P_Y^s, P_Y^t) \\geq d_{JS}(P_Z^s, P_Z^t)" + }, + { + "bbox": [ + 132, + 116, + 482, + 141 + ], + "type": "text", + "content": " holds, then for predictor " + }, + { + "bbox": [ + 132, + 116, + 482, + 141 + ], + "type": "inline_equation", + "content": "h: Z \\to Y" + }, + { + "bbox": [ + 132, + 116, + 482, + 141 + ], + "type": "text", + "content": ", we have" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 180, + 150, + 481, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 150, + 481, + 178 + ], + "spans": [ + { + "bbox": [ + 180, + 150, + 481, + 178 + ], + "type": "interline_equation", + "content": "\\varepsilon_ {s} (h \\circ g) + \\varepsilon_ {t} (h \\circ g) \\geq \\frac {1}{2} \\left(\\sqrt {d _ {J S} (P _ {Y} ^ {s} , P _ {Y} ^ {t})} - \\sqrt {d _ {J S} (P _ {Z} ^ {s} , P _ {Z} ^ {t})}\\right) ^ {2}, \\tag {3}", + "image_path": "c8d0c4b3bc38a813332689b8082635fabf7fdf379a84b7b4d7cce141ff8d5a45.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 358, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 358, + 198 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 358, + 198 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 185, + 358, + 198 + ], + "type": "inline_equation", + "content": "d_{JS}" + }, + { + "bbox": [ + 132, + 185, + 358, + 198 + ], + "type": "text", + "content": " denotes the Jensen-Shannon divergence." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 204, + 479, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 204, + 479, + 263 + ], + "spans": [ + { + "bbox": [ + 132, + 204, + 479, + 263 + ], + "type": "text", + "content": "Assuming the label distributions differ between source and target domain, i.e., " + }, + { + "bbox": [ + 132, + 204, + 479, + 263 + ], + "type": "inline_equation", + "content": "d_{JS}(P_Y^s,P_Y^t) > 0" + }, + { + "bbox": [ + 132, + 204, + 479, + 263 + ], + "type": "text", + "content": ", the lower bound shows that good alignment of representation distributions leads to bad joint error. In other words, when label shift exists, which is rather common in realistic scenarios, only aligning the marginal distribution is not enough for low target error." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 264, + 480, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 264, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 132, + 264, + 480, + 323 + ], + "type": "text", + "content": "To deal with the insufficiency problem shown in the lower bound analysis, it is intuitive to introduce label information into generalization analysis, e.g., the balanced error rate for classification scenarios [6]. Analogously, for the continuous regression scenarios, we first define the continuous balanced error rate for quantifying maximum conditional risk in DAR." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 331, + 481, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 481, + 354 + ], + "type": "text", + "content": "Definition 1. The continuous balanced error rate of predictor " + }, + { + "bbox": [ + 132, + 331, + 481, + 354 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 132, + 331, + 481, + 354 + ], + "type": "text", + "content": " on distribution " + }, + { + "bbox": [ + 132, + 331, + 481, + 354 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 132, + 331, + 481, + 354 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 201, + 364, + 481, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 364, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 201, + 364, + 481, + 384 + ], + "type": "interline_equation", + "content": "\\operatorname {c B E R} _ {\\mathrm {P}} (h, Y) := \\sup _ {(x, y) \\in \\mathcal {X} \\oplus \\mathcal {Y}} P (h (x) \\neq Y \\mid Y = y). \\tag {4}", + "image_path": "0281b218d7f201bce3a1026b1a7d7d6efd7db5db8889d19451264f33a3799c4a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 393, + 479, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 393, + 479, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 393, + 479, + 440 + ], + "type": "text", + "content": "Intuitively, the continuous balanced error rate can be regarded as the maximum potential risk for predicting sample " + }, + { + "bbox": [ + 132, + 393, + 479, + 440 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 132, + 393, + 479, + 440 + ], + "type": "text", + "content": " with label " + }, + { + "bbox": [ + 132, + 393, + 479, + 440 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 132, + 393, + 479, + 440 + ], + "type": "text", + "content": ". Based on cBER and conditional shift correction with representation transformation " + }, + { + "bbox": [ + 132, + 393, + 479, + 440 + ], + "type": "inline_equation", + "content": "Z = g(X)" + }, + { + "bbox": [ + 132, + 393, + 479, + 440 + ], + "type": "text", + "content": ", a sufficient condition for DAR can be induced as follows." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 447, + 479, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 479, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 479, + 472 + ], + "type": "text", + "content": "Theorem 2. For representation variable " + }, + { + "bbox": [ + 132, + 447, + 479, + 472 + ], + "type": "inline_equation", + "content": "Z^s = g(X^s)" + }, + { + "bbox": [ + 132, + 447, + 479, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 447, + 479, + 472 + ], + "type": "inline_equation", + "content": "Z^t = g(X^t)" + }, + { + "bbox": [ + 132, + 447, + 479, + 472 + ], + "type": "text", + "content": ", if " + }, + { + "bbox": [ + 132, + 447, + 479, + 472 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 132, + 447, + 479, + 472 + ], + "type": "text", + "content": " satisfies conditional invariant property, i.e.," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 247, + 481, + 480, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 481, + 480, + 495 + ], + "spans": [ + { + "bbox": [ + 247, + 481, + 480, + 495 + ], + "type": "interline_equation", + "content": "P _ {Z \\mid Y = y} ^ {s} = P _ {Z \\mid Y = y} ^ {t}, \\quad \\forall y \\in \\mathcal {Y}, \\tag {5}", + "image_path": "dfaa24d1ef084b26f26b83715a03d2c05883d685a184e4b116a5f86fae395c49.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 147, + 502, + 335, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 502, + 335, + 514 + ], + "spans": [ + { + "bbox": [ + 147, + 502, + 335, + 514 + ], + "type": "text", + "content": "then for any predictor " + }, + { + "bbox": [ + 147, + 502, + 335, + 514 + ], + "type": "inline_equation", + "content": "h: Z \\to Y" + }, + { + "bbox": [ + 147, + 502, + 335, + 514 + ], + "type": "text", + "content": ", we have" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 238, + 525, + 480, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 525, + 480, + 537 + ], + "spans": [ + { + "bbox": [ + 238, + 525, + 480, + 537 + ], + "type": "interline_equation", + "content": "\\varepsilon_ {s} (h) + \\varepsilon_ {t} (h) \\leq 2 \\mathrm {c B E R} _ {\\mathrm {P s}} (h, Y). \\tag {6}", + "image_path": "156df16ffa7fd6d29ad3dd03cfe4e991604aeebab8ecce2a1c9f33564fad7c9c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "type": "text", + "content": "Therefore, the key for DA, no matter for classification or regression problems, is to achieve conditional distribution alignment. In classification settings, conditional alignment can be achieved by conducting class-wise marginal distribution alignment with the help of target pseudo label. However, in regression scenarios, the label variable is continuous, which means previous methods correcting conditional shift in classification settings are no longer applicable. Obviously, it is impossible to align the " + }, + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "type": "inline_equation", + "content": "P_{X|y}" + }, + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "type": "text", + "content": " for every single " + }, + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "type": "text", + "content": " is a continuous variable and the possible number of " + }, + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 132, + 546, + 481, + 665 + ], + "type": "text", + "content": " is infinite. Thus, it is important and meaningful to explore a new framework that is capable of conditional distribution alignment for continuous label variable." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "type": "text", + "content": "H. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 471, + 214 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 471, + 214 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 471, + 214 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 471, + 214 + ], + "type": "image", + "image_path": "ca6e49b2c95c085cdfb1e4fc57ed3ce0542789eee74f050f932ee74df187057c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 221, + 482, + 321 + ], + "lines": [ + { + "bbox": [ + 130, + 221, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 482, + 321 + ], + "type": "text", + "content": "Fig. 2: Illustration of COD metric. (a) Classification DA methods usually correct conditional shift by class-wise computations on discrete clusters " + }, + { + "bbox": [ + 130, + 221, + 482, + 321 + ], + "type": "inline_equation", + "content": "P_{X|y}" + }, + { + "bbox": [ + 130, + 221, + 482, + 321 + ], + "type": "text", + "content": ", which is infeasible for regression due to the infinite slices " + }, + { + "bbox": [ + 130, + 221, + 482, + 321 + ], + "type": "inline_equation", + "content": "P_{X|y}" + }, + { + "bbox": [ + 130, + 221, + 482, + 321 + ], + "type": "text", + "content": ". (b) In COD, the continuous conditional distributions are embedded into the RKHS and characterized by finite statistical moments in kernel spaces, which do not rely on specific conditions " + }, + { + "bbox": [ + 130, + 221, + 482, + 321 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 221, + 482, + 321 + ], + "type": "text", + "content": ". Under the guarantees of distribution embedding property, the conditional alignment is equivalent to the matching on first-order and second-order statistics, i.e., mean embedding operator " + }, + { + "bbox": [ + 130, + 221, + 482, + 321 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_{X|Y}" + }, + { + "bbox": [ + 130, + 221, + 482, + 321 + ], + "type": "text", + "content": " and covariance operator " + }, + { + "bbox": [ + 130, + 221, + 482, + 321 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{XX|Y}" + }, + { + "bbox": [ + 130, + 221, + 482, + 321 + ], + "type": "text", + "content": ". When COD is minimized, conditional alignment is achieved, and the level sets of cross-domain labeling rules are then aligned." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 342, + 270, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 270, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 270, + 353 + ], + "type": "text", + "content": "3.3 COD metric for DAR" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 359, + 482, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 359, + 482, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 482, + 432 + ], + "type": "text", + "content": "Theory for COD. Due to the continuity of label variable, it is impractical to correct conditional shift by aligning marginal distribution of " + }, + { + "bbox": [ + 130, + 359, + 482, + 432 + ], + "type": "inline_equation", + "content": "P_{X|y}" + }, + { + "bbox": [ + 130, + 359, + 482, + 432 + ], + "type": "text", + "content": " for every single " + }, + { + "bbox": [ + 130, + 359, + 482, + 432 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 359, + 482, + 432 + ], + "type": "text", + "content": ". For this inevitable obstacle, an intuitive idea is can we align all " + }, + { + "bbox": [ + 130, + 359, + 482, + 432 + ], + "type": "inline_equation", + "content": "P_{X|y}" + }, + { + "bbox": [ + 130, + 359, + 482, + 432 + ], + "type": "text", + "content": " as a whole? i.e., " + }, + { + "bbox": [ + 130, + 359, + 482, + 432 + ], + "type": "inline_equation", + "content": "P_{X|Y}" + }, + { + "bbox": [ + 130, + 359, + 482, + 432 + ], + "type": "text", + "content": ", rather than align " + }, + { + "bbox": [ + 130, + 359, + 482, + 432 + ], + "type": "inline_equation", + "content": "P_{X|y}" + }, + { + "bbox": [ + 130, + 359, + 482, + 432 + ], + "type": "text", + "content": " one by one? It is more than attractive that if we can characterize " + }, + { + "bbox": [ + 130, + 359, + 482, + 432 + ], + "type": "inline_equation", + "content": "P_{X|Y}" + }, + { + "bbox": [ + 130, + 359, + 482, + 432 + ], + "type": "text", + "content": " with a well-defined discrepancy, then all " + }, + { + "bbox": [ + 130, + 359, + 482, + 432 + ], + "type": "inline_equation", + "content": "P_{X|y}" + }, + { + "bbox": [ + 130, + 359, + 482, + 432 + ], + "type": "text", + "content": " can be aligned by aligning the whole " + }, + { + "bbox": [ + 130, + 359, + 482, + 432 + ], + "type": "inline_equation", + "content": "P_{X|Y}" + }, + { + "bbox": [ + 130, + 359, + 482, + 432 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "text", + "content": "In kernel embedding theory, a distribution can be characterized by a point in RKHS. Then " + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "inline_equation", + "content": "P_{X|Y}" + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "text", + "content": " can be seen as a point set in RKHS in which every element corresponds to a distribution " + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "inline_equation", + "content": "P_{X|y}" + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "text", + "content": " for a fixed " + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "text", + "content": ". From another perspective, we can consider " + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "inline_equation", + "content": "P_{X|Y}" + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "text", + "content": " as the distribution in RKHS and measure the cross-domain conditional discrepancy by defining the COD distance on their statistical moments in RKHS. Following the definitions of conditional statistics in pioneer works, i.e., the conditional mean operator as " + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_{X|Y} \\coloneqq \\mathcal{C}_{XY}\\mathcal{C}_{YY}^{-1}" + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "text", + "content": " [48] and conditional covariance operator as " + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{XX|Y} = \\mathcal{C}_{XX} - \\mathcal{C}_{XY}\\mathcal{C}_{YY}^{-1}\\mathcal{C}_{YX}" + }, + { + "bbox": [ + 130, + 432, + 482, + 540 + ], + "type": "text", + "content": " [13], we present the definition of COD as follows." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 544, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 582 + ], + "type": "text", + "content": "Definition 2 (COD). Given conditional distributions " + }, + { + "bbox": [ + 130, + 544, + 482, + 582 + ], + "type": "inline_equation", + "content": "P_{X|Y}^{s}" + }, + { + "bbox": [ + 130, + 544, + 482, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 544, + 482, + 582 + ], + "type": "inline_equation", + "content": "P_{X|Y}^{t}" + }, + { + "bbox": [ + 130, + 544, + 482, + 582 + ], + "type": "text", + "content": ", denote " + }, + { + "bbox": [ + 130, + 544, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{XX|Y}^{ss}" + }, + { + "bbox": [ + 130, + 544, + 482, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 544, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{XX|Y}^{tt}" + }, + { + "bbox": [ + 130, + 544, + 482, + 582 + ], + "type": "text", + "content": " as their corresponding covariance operators. The COD is defined as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 140, + 587, + 481, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 587, + 481, + 613 + ], + "spans": [ + { + "bbox": [ + 140, + 587, + 481, + 613 + ], + "type": "interline_equation", + "content": "d _ {\\mathrm {C O D}} ^ {2} \\left(P _ {X | Y} ^ {s}, P _ {X | Y} ^ {t}\\right) = \\left\\| \\mathcal {U} _ {X | Y} ^ {s} - \\mathcal {U} _ {X | Y} ^ {t} \\right\\| _ {\\mathcal {H} _ {\\mathcal {K}}} ^ {2} + \\operatorname {t r} \\left(\\mathcal {C} _ {X X | Y} ^ {s s} + \\mathcal {C} _ {X X | Y} ^ {t t} - 2 \\mathcal {C} _ {X X | Y} ^ {s t}\\right), \\tag {7}", + "image_path": "cdd58a7ddb829cf18d070839ab8b17fbc4d15e43c8d215ca15a38ce302e44784.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 613, + 324, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 613, + 324, + 639 + ], + "spans": [ + { + "bbox": [ + 132, + 613, + 324, + 639 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 613, + 324, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{XX|Y}^{st} = \\sqrt{\\sqrt{\\mathcal{C}_{XX|Y}^{ss}}\\mathcal{C}_{XX|Y}^{tt}\\sqrt{\\mathcal{C}_{XX|Y}^{ss}}}" + }, + { + "bbox": [ + 132, + 613, + 324, + 639 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "Theoretically, it can be proved that COD indeed admits the metric property on conditional distributions, i.e., two conditional distributions are the same if" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 192, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 192, + 91, + 448, + 102 + ], + "type": "text", + "content": "COD: Learning Conditional Invariant Representation for DAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 140 + ], + "type": "text", + "content": "and only if the COD metric between them is zero. An intuitive illustration of COD is shown in Fig. 2. Formally, COD admits two key properties." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 145, + 448, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 145, + 448, + 156 + ], + "spans": [ + { + "bbox": [ + 131, + 145, + 448, + 156 + ], + "type": "text", + "content": "Proposition 1. (a) COD defines a metric between Gaussian measures." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 157, + 481, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 157, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 131, + 157, + 481, + 195 + ], + "type": "text", + "content": "(b) Let " + }, + { + "bbox": [ + 131, + 157, + 481, + 195 + ], + "type": "inline_equation", + "content": "(\\mathcal{X},\\mathcal{B}_{\\mathcal{X}})" + }, + { + "bbox": [ + 131, + 157, + 481, + 195 + ], + "type": "text", + "content": " be the locally compact and Hausdorff measurable space and " + }, + { + "bbox": [ + 131, + 157, + 481, + 195 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 157, + 481, + 195 + ], + "type": "text", + "content": " be " + }, + { + "bbox": [ + 131, + 157, + 481, + 195 + ], + "type": "inline_equation", + "content": "c_{0}" + }, + { + "bbox": [ + 131, + 157, + 481, + 195 + ], + "type": "text", + "content": "-universal kernel. Assuming that " + }, + { + "bbox": [ + 131, + 157, + 481, + 195 + ], + "type": "inline_equation", + "content": "(\\phi (X),\\psi (Y))" + }, + { + "bbox": [ + 131, + 157, + 481, + 195 + ], + "type": "text", + "content": " is a Gaussian random variable in " + }, + { + "bbox": [ + 131, + 157, + 481, + 195 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathcal{X}}\\oplus \\mathcal{H}_{\\mathcal{Y}}" + }, + { + "bbox": [ + 131, + 157, + 481, + 195 + ], + "type": "text", + "content": ". For any " + }, + { + "bbox": [ + 131, + 157, + 481, + 195 + ], + "type": "inline_equation", + "content": "P_{X|Y}^{s},P_{X|Y}^{t}\\in \\mathbb{P}_{\\mathcal{X}|\\mathcal{Y}}" + }, + { + "bbox": [ + 131, + 157, + 481, + 195 + ], + "type": "text", + "content": ", we have" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 212, + 203, + 481, + 218 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 203, + 481, + 218 + ], + "spans": [ + { + "bbox": [ + 212, + 203, + 481, + 218 + ], + "type": "interline_equation", + "content": "d _ {\\mathrm {C O D}} ^ {2} \\left(P _ {X | Y} ^ {s}, P _ {X | Y} ^ {t}\\right) = 0 \\Rightarrow P _ {X | Y} ^ {s} = P _ {X | Y} ^ {t}. \\tag {8}", + "image_path": "be5d1e3974510c516e84194714570f36067e8ff3843c547c71ce83822e106f80.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 224, + 479, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 479, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 479, + 270 + ], + "type": "text", + "content": "Empirical Estimation of COD. Based on the above properties, we can perform conditional distribution alignment by minimizing the COD metric between source and target domain. To do this, we shall derive the empirical estimation of COD, which can be obtained with the help of the kernel trick." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_i^s,\\mathbf{y}_i^s)_{i = 1}^{n_s}" + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_i^t,\\mathbf{y}_i^t)_{i = 1}^{n_t}" + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "text", + "content": " be two set of samples drawn i.i.d. from source and target domain. For simplicity, " + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "inline_equation", + "content": "n_s" + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "inline_equation", + "content": "n_t" + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "text", + "content": " are both set to " + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "text", + "content": ". In kernel method, samples are mapped to RKHS " + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathcal{X}}\\oplus \\mathcal{H}_{\\mathcal{Y}}" + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "text", + "content": " by the implicit feature map " + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "inline_equation", + "content": "(\\phi ,\\psi)" + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "text", + "content": ". Denote the feature map matrices by " + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "inline_equation", + "content": "(\\Phi^s,\\Psi^s)" + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "inline_equation", + "content": "(\\Phi^t,\\Psi^t)" + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "text", + "content": ", the explicit kernel matrices " + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_X^{ss} = \\Phi^s\\Phi^{sT}" + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_Y^{ss} = \\Psi^s\\Psi^{sT}" + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "text", + "content": " are computed as " + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "inline_equation", + "content": "(\\mathbf{K}_X^{ss})_{ij} = k_{\\mathcal{X}}(\\mathbf{x}_i^s,\\mathbf{x}_j^s)" + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "inline_equation", + "content": "(\\mathbf{K}_Y^{ss})_{ij} = k_{\\mathcal{Y}}(\\mathbf{y}_i^s,\\mathbf{y}_j^s)" + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "text", + "content": ", respectively. And so as " + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_X^{tt}" + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_X^{st}" + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_Y^{tt}" + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_n" + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "text", + "content": " be the n-dimensional identity matrix, " + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{1}_n" + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "text", + "content": " be the n-dimensional vector with all elements equal to 1. The " + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "text", + "content": "-dimensional centering matrix can be defined as " + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_n = \\mathbf{I}_n - \\frac{1}{n}\\mathbf{1}_n\\mathbf{1}_n^T" + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "text", + "content": ". Then the centralized kernel matrix is defined as " + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{G} = \\mathbf{H}_n\\mathbf{K}\\mathbf{H}_n" + }, + { + "bbox": [ + 131, + 271, + 482, + 380 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 380, + 481, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 481, + 415 + ], + "type": "text", + "content": "With above notations, the covariance operator " + }, + { + "bbox": [ + 130, + 380, + 481, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{XY}" + }, + { + "bbox": [ + 130, + 380, + 481, + 415 + ], + "type": "text", + "content": " can be empirically estimated as " + }, + { + "bbox": [ + 130, + 380, + 481, + 415 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{C}}_{XY} = \\frac{1}{n} (\\pmb {\\Phi} - \\hat{\\mu}_X\\pmb {1}^T)(\\pmb {\\Psi} - \\hat{\\mu}_Y\\pmb {1}^T)^T = \\frac{1}{n}\\pmb {\\Phi}\\pmb {H}\\pmb {\\Psi}^T" + }, + { + "bbox": [ + 130, + 380, + 481, + 415 + ], + "type": "text", + "content": ", then the conditional moment statistics can be estimated as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 159, + 420, + 481, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 420, + 481, + 444 + ], + "spans": [ + { + "bbox": [ + 159, + 420, + 481, + 444 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {U}} _ {X \\mid Y} = (\\frac {1}{n} \\boldsymbol {\\Phi} \\mathbf {H} \\boldsymbol {\\Psi} ^ {T}) (\\frac {1}{n} \\boldsymbol {\\Psi} \\mathbf {H} \\boldsymbol {\\Psi} ^ {T}) ^ {- 1}, \\quad \\hat {\\mathcal {C}} _ {X X \\mid Y} = \\frac {1}{n} \\boldsymbol {\\Phi} \\mathbf {H} _ {n} \\mathbf {A} (\\boldsymbol {\\Phi} \\mathbf {H} _ {n} \\mathbf {A}) ^ {T}, \\quad (9)", + "image_path": "252bd325022c7b817cc0fbb84bcf4de41eb8fbdab712586a8cd27e9212ca72d1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 449, + 488, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 488, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 488, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 449, + 488, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 130, + 449, + 488, + 491 + ], + "type": "text", + "content": " satisfies that " + }, + { + "bbox": [ + 130, + 449, + 488, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{B} = \\mathbf{A}\\mathbf{A}^T" + }, + { + "bbox": [ + 130, + 449, + 488, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 449, + 488, + 491 + ], + "type": "inline_equation", + "content": "B\\triangleq \\mathbf{I}_n - \\frac{1}{n\\varepsilon}\\left[\\mathbf{G}_Y - \\mathbf{G}_Y\\left(\\mathbf{G}_Y + \\varepsilon n\\mathbf{I}_n\\right)^{-1}\\mathbf{G}_Y\\right]" + }, + { + "bbox": [ + 130, + 449, + 488, + 491 + ], + "type": "text", + "content": ". Based on these estimations, the closed-form empirical estimation of the COD metric can be derived as follows." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 496, + 473, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 496, + 473, + 509 + ], + "spans": [ + { + "bbox": [ + 131, + 496, + 473, + 509 + ], + "type": "text", + "content": "Proposition 2. The empirical Conditional Operator Discrepancy distance is" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 154, + 515, + 480, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 515, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 154, + 515, + 480, + 593 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {d} _ {\\mathrm {C O D}} ^ {2} = \\operatorname {t r} \\left(\\mathbf {K} _ {Y} ^ {s s} \\left(\\mathbf {K} _ {Y} ^ {s s}\\right) ^ {- 1} \\mathbf {K} _ {X} ^ {s s} \\left(\\mathbf {K} _ {Y} ^ {s s}\\right) ^ {- 1}\\right) + \\operatorname {t r} \\left(\\mathbf {K} _ {Y} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1} \\mathbf {K} _ {X} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1}\\right) \\\\ - 2 \\operatorname {t r} \\left(\\mathbf {K} _ {Y} ^ {t s} \\left(\\mathbf {K} _ {Y} ^ {s s}\\right) ^ {- 1} \\mathbf {K} _ {X ^ {s} X ^ {t}} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1}\\right) \\\\ + \\varepsilon \\operatorname {t r} \\left[ \\mathbf {G} _ {X} ^ {s} \\left(\\varepsilon n \\mathbf {I} _ {n} + \\mathbf {G} _ {Y} ^ {s}\\right) ^ {- 1} \\right] + \\varepsilon \\operatorname {t r} \\left[ \\mathbf {G} _ {X} ^ {t} \\left(\\varepsilon n \\mathbf {I} _ {n} + \\mathbf {G} _ {Y} ^ {t}\\right) ^ {- 1} \\right] \\tag {10} \\\\ \\left. - \\frac {2}{n} \\left\\| \\left(\\mathbf {H} _ {n} \\mathbf {A} ^ {t}\\right) ^ {T} \\mathbf {K} _ {X} ^ {t s} \\left(\\mathbf {H} _ {n} \\mathbf {A} ^ {s}\\right) \\right\\| _ {*} \\right., \\\\ \\end{array}", + "image_path": "531dc4cf7095e20095b2c9ceaa569ae09a0424d1ee5c0b2f6f5e321b61d91308.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 600, + 479, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 600, + 479, + 625 + ], + "spans": [ + { + "bbox": [ + 131, + 600, + 479, + 625 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 600, + 479, + 625 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_{*}" + }, + { + "bbox": [ + 131, + 600, + 479, + 625 + ], + "type": "text", + "content": " is the nuclear norm. Empirically, " + }, + { + "bbox": [ + 131, + 600, + 479, + 625 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_Y^{-1}" + }, + { + "bbox": [ + 131, + 600, + 479, + 625 + ], + "type": "text", + "content": " is approximated by the regularized formulation " + }, + { + "bbox": [ + 131, + 600, + 479, + 625 + ], + "type": "inline_equation", + "content": "(\\mathbf{K}_Y + \\lambda \\mathbf{I})^{-1}" + }, + { + "bbox": [ + 131, + 600, + 479, + 625 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 482, + 667 + ], + "type": "text", + "content": "Discriminability of empirical COD. As discussed above, the first term of COD measures the distance of the first-order statistic between conditional distributions in RKHS. However, recent study [5,54] shows that the mean matching" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "type": "text", + "content": "H. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "may reduce the discriminability when bringing the two distributions close, which undoubtedly degrades the performance of the algorithm. Therefore, it is worth making an in-depth analysis of the discriminability of COD as well." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "text", + "content": "Without loss of generality, we first consider Kronechker Delta kernel on conditioning variable " + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "inline_equation", + "content": "k(\\mathbf{y}_1,\\mathbf{y}_2) = \\delta (\\mathbf{y}_1,\\mathbf{y}_2)" + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "text", + "content": ". More explicitly, " + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "inline_equation", + "content": "k(\\mathbf{y}_1,\\mathbf{y}_2)" + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "text", + "content": " equals to 1 if " + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_1 = \\mathbf{y}_2" + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "text", + "content": " or 0 otherwise. For empirical scenarios with finite samples, the numbers of samples and observed labels are finite, which implies the number of distinct label " + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "text", + "content": " will be obviously no more than sample size " + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "text", + "content": ". For convenience, we call two samples are in the same class when their label values are the same. Consequently, the first term in COD can be reformulated as follows." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 242, + 479, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 242, + 479, + 266 + ], + "spans": [ + { + "bbox": [ + 131, + 242, + 479, + 266 + ], + "type": "text", + "content": "Proposition 3. Under the conditions mentioned above, the empirical estimation of " + }, + { + "bbox": [ + 131, + 242, + 479, + 266 + ], + "type": "inline_equation", + "content": "d_{CMMD}" + }, + { + "bbox": [ + 131, + 242, + 479, + 266 + ], + "type": "text", + "content": " can be written as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 195, + 274, + 480, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 274, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 195, + 274, + 480, + 392 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\widehat {d} _ {\\mathrm {C M M D}} ^ {2} = \\sum_ {p = 1} ^ {c} \\frac {1}{(\\lambda + n _ {p}) ^ {2}} \\left(\\sum_ {i, j} k (\\mathbf {x} _ {(p), i} ^ {s}, \\mathbf {x} _ {(p), j} ^ {s})\\right) \\\\ + \\sum_ {p = 1} ^ {c} \\frac {1}{(\\lambda + n _ {p}) ^ {2}} \\left(\\sum_ {i, j} k \\left(\\mathbf {x} _ {(p), i} ^ {t}, \\mathbf {x} _ {(p), j} ^ {t}\\right)\\right) \\tag {11} \\\\ - 2 \\sum_ {p = 1} ^ {c} \\frac {1}{(\\lambda + n _ {p}) ^ {2}} \\left(\\sum_ {i, j} k (\\mathbf {x} _ {(p), i} ^ {s}, \\mathbf {x} _ {(p), j} ^ {t})\\right), \\\\ \\end{array}", + "image_path": "d1fa97c93ab9ea85a95fce5b67bc5ab62f38834e30b291314c80868503166ff3.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 399, + 479, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 399, + 479, + 423 + ], + "spans": [ + { + "bbox": [ + 131, + 399, + 479, + 423 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 399, + 479, + 423 + ], + "type": "inline_equation", + "content": "n_i" + }, + { + "bbox": [ + 131, + 399, + 479, + 423 + ], + "type": "text", + "content": " denotes the sample-size of " + }, + { + "bbox": [ + 131, + 399, + 479, + 423 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 399, + 479, + 423 + ], + "type": "text", + "content": "-th label such that " + }, + { + "bbox": [ + 131, + 399, + 479, + 423 + ], + "type": "inline_equation", + "content": "n = \\sum_{i=1}^{c} n_i" + }, + { + "bbox": [ + 131, + 399, + 479, + 423 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 399, + 479, + 423 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{(p),i}" + }, + { + "bbox": [ + 131, + 399, + 479, + 423 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 131, + 399, + 479, + 423 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 399, + 479, + 423 + ], + "type": "text", + "content": "-th sample of " + }, + { + "bbox": [ + 131, + 399, + 479, + 423 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 131, + 399, + 479, + 423 + ], + "type": "text", + "content": "-th label." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 429, + 480, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 480, + 513 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 480, + 513 + ], + "type": "text", + "content": "Eq. (11) contains three terms. The third term characterizes the cross-domain intra-class similarity. When Eq. (11) is minimized, the third term is maximized so that the cross-domain sample with the same label will be aligned, leading to conditional alignment as expected. However, the first two terms of the equation represent the intra-class similarity of both domains. Minimizing them will decrease the intra-class similarity, which means intra-class samples will be pushed away from each other and the discriminability of representations is degraded." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 519, + 389, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 519, + 389, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 519, + 389, + 531 + ], + "type": "text", + "content": "Remark 2. Note the Kronechker Delta kernel is defined as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 242, + 536, + 480, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 536, + 480, + 569 + ], + "spans": [ + { + "bbox": [ + 242, + 536, + 480, + 569 + ], + "type": "interline_equation", + "content": "k \\left(\\mathbf {y} _ {1}, \\mathbf {y} _ {2}\\right) = \\left\\{ \\begin{array}{l} 1 \\text {i f} \\mathbf {y} _ {1} = \\mathbf {y} _ {2} \\\\ 0 \\text {o t h e r w i s e} \\end{array} \\right.. \\tag {12}", + "image_path": "169b92329285365fd31297ee5dc99c8d4f1859f757ee3ac2ad13c92f2b38ac7d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 575, + 479, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 575, + 479, + 612 + ], + "spans": [ + { + "bbox": [ + 131, + 575, + 479, + 612 + ], + "type": "text", + "content": "When it comes to the Gaussian kernel for empirical modeling, the case of " + }, + { + "bbox": [ + 131, + 575, + 479, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_1 = \\mathbf{y}_2" + }, + { + "bbox": [ + 131, + 575, + 479, + 612 + ], + "type": "text", + "content": " is the same; for other cases, the kernel values can be arbitrarily close to zero, i.e., for any " + }, + { + "bbox": [ + 131, + 575, + 479, + 612 + ], + "type": "inline_equation", + "content": "\\epsilon > 0" + }, + { + "bbox": [ + 131, + 575, + 479, + 612 + ], + "type": "text", + "content": ", there always width " + }, + { + "bbox": [ + 131, + 575, + 479, + 612 + ], + "type": "inline_equation", + "content": "\\sigma_0" + }, + { + "bbox": [ + 131, + 575, + 479, + 612 + ], + "type": "text", + "content": ", such that" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 222, + 620, + 480, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 620, + 480, + 634 + ], + "spans": [ + { + "bbox": [ + 222, + 620, + 480, + 634 + ], + "type": "interline_equation", + "content": "k \\left(\\mathbf {y} _ {1}, \\mathbf {y} _ {2}\\right) = \\exp \\left(- \\| \\mathbf {y} _ {1} - \\mathbf {y} _ {2} \\| _ {2} ^ {2} / \\sigma_ {0} ^ {2}\\right) < \\epsilon . \\tag {13}", + "image_path": "f619e35618d287ffc45fc613e40ca36da975e5b231312957291b8588817bd19f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "Thus, it is generally reasonable to use Gaussian kernel, which admits a similar property to Kronechker Delta kernel, for empirical modeling." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 192, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 192, + 91, + 448, + 102 + ], + "type": "text", + "content": "COD: Learning Conditional Invariant Representation for DAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "Remark 3. Note that although we treat the number of distinct labels as finite numbers in mini-batch training, it does not mean class-wise matching can be conducted. In fact, even though the pseudo target label has finite range in minibatch, it still admits the continuous property, which means we can not cluster samples as usually done in classification methods. When two samples have predicted labels with little numerical difference, they can be regarded as samples of the same class in classification settings. Nonetheless, this kind of clustering is infeasible in regression settings since the labels are continuous. Therefore, class-wise alignment is not practical even in mini-batch training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 228, + 482, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 252 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 252 + ], + "type": "text", + "content": "With these observation, it is reasonable to reformulate Eq. (11) to promote the discriminability of representations:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 141, + 256, + 466, + 288 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 256, + 466, + 288 + ], + "spans": [ + { + "bbox": [ + 141, + 256, + 466, + 288 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\widehat {d} _ {\\mathrm {C M M D} _ {\\mathrm {m o d}}} = - \\operatorname {t r} \\left(\\mathbf {K} _ {Y} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1} \\mathbf {K} _ {X} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1}\\right) - \\operatorname {t r} \\left(\\mathbf {K} _ {Y} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1} \\mathbf {K} _ {X} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1}\\right) \\\\ - 2 \\operatorname {t r} \\left(\\mathbf {K} _ {Y ^ {t} Y ^ {s}} \\left(\\mathbf {K} _ {Y} ^ {s s}\\right) ^ {- 1} \\mathbf {K} _ {X ^ {s} X ^ {t}} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1}\\right) \\\\ \\end{array}", + "image_path": "b1f993ebb773616bc6b31adbc5d9a719c320b5ff99cc610b56f83c1c7f1e9cc1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 292, + 482, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 482, + 327 + ], + "type": "text", + "content": "Now minimizing Eq. (14) will maximize intra-class similarity regardless of domain, yielding conditional matching with great discriminability. Then the empirical COD is modified as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 145, + 333, + 480, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 333, + 480, + 410 + ], + "spans": [ + { + "bbox": [ + 145, + 333, + 480, + 410 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\widehat {d} _ {\\mathrm {C O D} _ {\\mathrm {m o d}}} = - \\operatorname {t r} \\left(\\mathbf {K} _ {Y} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1} \\mathbf {K} _ {X} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1}\\right) - \\operatorname {t r} \\left(\\mathbf {K} _ {Y} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1} \\mathbf {K} _ {X} ^ {t t} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1}\\right) \\\\ - 2 \\operatorname {t r} \\left(\\mathbf {K} _ {Y ^ {t} Y ^ {s}} \\left(\\mathbf {K} _ {Y} ^ {s s}\\right) ^ {- 1} \\mathbf {K} _ {X ^ {s} X ^ {t}} \\left(\\mathbf {K} _ {Y} ^ {t t}\\right) ^ {- 1}\\right) \\\\ + \\varepsilon \\operatorname {t r} \\left[ \\mathbf {G} _ {X} ^ {s} \\left(\\varepsilon n \\mathbf {I} _ {n} + \\mathbf {G} _ {Y} ^ {s}\\right) ^ {- 1} \\right] + \\varepsilon \\operatorname {t r} \\left[ \\mathbf {G} _ {X} ^ {t} \\left(\\varepsilon n \\mathbf {I} _ {n} + \\mathbf {G} _ {Y} ^ {t}\\right) ^ {- 1} \\right] \\tag {15} \\\\ - \\frac {2}{n} \\left\\| \\left(\\mathbf {H} _ {n} \\mathbf {A} ^ {t}\\right) ^ {T} \\mathbf {K} _ {X} ^ {t s} \\left(H _ {n} \\mathbf {A} ^ {s}\\right) \\right\\| _ {*} \\\\ \\end{array}", + "image_path": "3ab694de8897a112948d180d280b5c68935f8d44931e65e5fee9fbfde2d8abd9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 422, + 299, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 422, + 299, + 434 + ], + "spans": [ + { + "bbox": [ + 131, + 422, + 299, + 434 + ], + "type": "text", + "content": "3.4 COD-based DAR Modeling" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 438, + 482, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 535 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 535 + ], + "type": "text", + "content": "According to theorem (2), when the conditional distributions " + }, + { + "bbox": [ + 130, + 438, + 482, + 535 + ], + "type": "inline_equation", + "content": "P_{X^s|Y}" + }, + { + "bbox": [ + 130, + 438, + 482, + 535 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 438, + 482, + 535 + ], + "type": "inline_equation", + "content": "P_{X^t|Y}" + }, + { + "bbox": [ + 130, + 438, + 482, + 535 + ], + "type": "text", + "content": " are aligned, a well-trained source predictor can gain promising performance on target domain. However, when the shift is significant, it is difficult for predictor to produce high-quality pseudo labels at the beginning. For this reason, the model is likely to miss the optimal optimization direction in the early training stage and thus fall into sub-optimal solutions. Therefore, the marginal alignment can improve the reliability of initial pseudo labels, and further benefit the conditional alignment. Overall, the learning objective of COD can be formulated as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 231, + 538, + 480, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 538, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 231, + 538, + 480, + 557 + ], + "type": "interline_equation", + "content": "\\min _ {g, h} \\mathcal {L} _ {\\mathrm {s r c}} + \\lambda_ {1} \\widehat {d} _ {\\mathrm {C O D} _ {\\mathrm {m o d}}} + \\lambda_ {2} \\widehat {d} _ {\\mathrm {K G W}}, \\tag {16}", + "image_path": "aa7195f3c6e74dc1300c08442cd47ff5eff9e9761ed0790d939054dd3bfca87d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 562, + 481, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 562, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 562, + 481, + 586 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 562, + 481, + 586 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2" + }, + { + "bbox": [ + 131, + 562, + 481, + 586 + ], + "type": "text", + "content": " are the trade-off parameters. For regression task, the source risk " + }, + { + "bbox": [ + 131, + 562, + 481, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{src}}" + }, + { + "bbox": [ + 131, + 562, + 481, + 586 + ], + "type": "text", + "content": " is commonly set as Mean Square Error (MSE) over continuous outputs:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 252, + 592, + 480, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 592, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 252, + 592, + 480, + 622 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s r c}} = \\frac {1}{n ^ {s}} \\sum_ {i = 1} ^ {n ^ {s}} \\left\\| \\hat {\\mathbf {y}} _ {i} ^ {s} - \\mathbf {y} _ {i} ^ {s} \\right\\| _ {2} ^ {2} \\tag {17}", + "image_path": "e56847e79937e09c945a037c1c2e80f0f5dc80cc237dd741f034942ec54a5c72.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 628, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 628, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 628, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_i^s" + }, + { + "bbox": [ + 131, + 628, + 482, + 666 + ], + "type": "text", + "content": " is the predicted value of the " + }, + { + "bbox": [ + 131, + 628, + 482, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 628, + 482, + 666 + ], + "type": "text", + "content": "-th source sample. Note that the marginal alignment term can be replaced by any other discrepancy metric, while we adopt KGW here to preserve the consistency with OT framework." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "H. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 479, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 479, + 178 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 479, + 178 + ], + "type": "text", + "content": "We evaluate our method with several state-of-the-art domain adaptation methods on three regression benchmark datasets, implementation details and additional illustrations are provided in supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 178, + 480, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 178, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 480, + 274 + ], + "type": "text", + "content": "dSprites [19] is a standard 2D synthetic dataset for deep representation learning, which contains three domains each with 737,280 images: Color (C), Noisy (N) and Scream (S). Following [4], we consider three factors for regression tasks: scale and two plane coordinates, while the orientation factor is excluded for its complexity. We evaluate all methods on six transfer tasks: " + }, + { + "bbox": [ + 130, + 178, + 480, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{C} \\rightarrow \\mathbf{N}" + }, + { + "bbox": [ + 130, + 178, + 480, + 274 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 178, + 480, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{C} \\rightarrow \\mathbf{S}" + }, + { + "bbox": [ + 130, + 178, + 480, + 274 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 178, + 480, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{N} \\rightarrow \\mathbf{C}" + }, + { + "bbox": [ + 130, + 178, + 480, + 274 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 178, + 480, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{N} \\rightarrow \\mathbf{S}" + }, + { + "bbox": [ + 130, + 178, + 480, + 274 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 178, + 480, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\rightarrow \\mathbf{C}" + }, + { + "bbox": [ + 130, + 178, + 480, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 178, + 480, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\rightarrow \\mathbf{N}" + }, + { + "bbox": [ + 130, + 178, + 480, + 274 + ], + "type": "text", + "content": ". Following previous works [4,38], the sum of Mean Absolute Error (MAE) on all three sub-regression tasks is reported as the evaluation metric." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 275, + 480, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 480, + 334 + ], + "type": "text", + "content": "Biwi kinect [12] is a real-world dataset containing two domains according to gender: Female " + }, + { + "bbox": [ + 130, + 275, + 480, + 334 + ], + "type": "inline_equation", + "content": "(\\mathbf{F})" + }, + { + "bbox": [ + 130, + 275, + 480, + 334 + ], + "type": "text", + "content": " with 5874 images and Male " + }, + { + "bbox": [ + 130, + 275, + 480, + 334 + ], + "type": "inline_equation", + "content": "(\\mathbf{M})" + }, + { + "bbox": [ + 130, + 275, + 480, + 334 + ], + "type": "text", + "content": " with 9804 images. The three factors that can be employed for regression tasks are pitch, yaw and roll. We evaluate all methods on two transfer tasks: " + }, + { + "bbox": [ + 130, + 275, + 480, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\rightarrow \\mathbf{M}" + }, + { + "bbox": [ + 130, + 275, + 480, + 334 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 275, + 480, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\rightarrow \\mathbf{F}" + }, + { + "bbox": [ + 130, + 275, + 480, + 334 + ], + "type": "text", + "content": ". The sum of MAE on three sub-regression tasks is reported." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 335, + 480, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 480, + 418 + ], + "type": "text", + "content": "MPI3D [16] is a simulation-to-real dataset of 3D objects which consist of three different domain: Toy (T), RealistiC (RC) and ReaL (RL). Each domain contains 1,036,800 images. There are two factors in MPI3D that can be employed for regression tasks: a rotation about a vertical axis and a second rotation about a horizontal axis. We evaluate all methods on six transfer tasks: RL " + }, + { + "bbox": [ + 130, + 335, + 480, + 418 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 335, + 480, + 418 + ], + "type": "text", + "content": " RC, RL " + }, + { + "bbox": [ + 130, + 335, + 480, + 418 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 335, + 480, + 418 + ], + "type": "text", + "content": " T, RC " + }, + { + "bbox": [ + 130, + 335, + 480, + 418 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 335, + 480, + 418 + ], + "type": "text", + "content": " T, RC " + }, + { + "bbox": [ + 130, + 335, + 480, + 418 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 335, + 480, + 418 + ], + "type": "text", + "content": " RL, T " + }, + { + "bbox": [ + 130, + 335, + 480, + 418 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 335, + 480, + 418 + ], + "type": "text", + "content": " RL and T " + }, + { + "bbox": [ + 130, + 335, + 480, + 418 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 335, + 480, + 418 + ], + "type": "text", + "content": " RC. The sum of MAE on two sub-regression tasks is reported." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 419, + 481, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 481, + 479 + ], + "type": "text", + "content": "Comparison methods. SOTA methods for UDA and DAR are selected: TCA [40], MCD [45], DAN [28], DANN [14], JDOT [8], AFN [58], RSD [4], DARE-GRAM [38]. Following previous works [4,38], Mean Absolute Error (MAE) is used as our evaluation metric across all regression tasks. In comparison experiment, the proposed DAR model in Eq. (16) is denoted as COD for convenience." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 498, + 266, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 266, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 266, + 511 + ], + "type": "text", + "content": "4.1 Results and Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Comparison on dSprites. As shown in Tab. 1, COD achieves the best performance among all competing methods in nearly all tasks of dSprites. In " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{C} \\rightarrow \\mathbf{N}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{C} \\rightarrow \\mathbf{S}" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", in which the Resnet-18 and UDA methods exhibit terrible performance because of the significant domain gap. Although SOTA DAR methods have greatly boosted the performance compared with classification methods, COD still achieves a significant improvement, reducing the error to only " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " of that produced by SOTA methods. These results demonstrate that the conditional invariant representations learned by COD indeed ensure better discriminability. Comparison on Biwi Kinect. Since Biwi Kinect is the closest to real-world scenarios among the three benchmark datasets, the performance improvement of model is more challenging. In Tab. 2, the performance gap between UDA methods and DAR methods is relatively smaller, which implies the SOTA DAR" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 192, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 192, + 90, + 447, + 102 + ], + "type": "text", + "content": "COD: Learning Conditional Invariant Representation for DAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 135, + 478, + 264 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 473, + 126 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 473, + 126 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 473, + 126 + ], + "type": "text", + "content": "Table 1: Domain adaptation regression results (MAE) on dSprites (ResNet-18)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 135, + 478, + 264 + ], + "lines": [ + { + "bbox": [ + 136, + 135, + 478, + 264 + ], + "spans": [ + { + "bbox": [ + 136, + 135, + 478, + 264 + ], + "type": "table", + "html": "
MethodC→NC→SN→CN→SS→CS→NAvg.
Resnet-180.940.900.160.650.080.260.498
TCA0.940.870.190.660.100.230.498
MCD0.810.810.170.650.070.190.450
JDOT0.860.790.190.640.100.230.468
AFN1.000.960.160.620.080.320.523
DAN0.700.770.120.500.060.110.377
DANN0.470.460.160.650.050.100.315
RSD0.310.310.120.530.070.080.237
DARE-GRAM0.300.200.110.250.050.070.164
COD0.120.160.100.230.080.120.134
", + "image_path": "dd667930f755c59575616a898b8275497cd80fbf94fce8f1ccd67d1f48e9d966.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 213, + 297, + 402, + 425 + ], + "blocks": [ + { + "bbox": [ + 133, + 275, + 480, + 286 + ], + "lines": [ + { + "bbox": [ + 133, + 275, + 480, + 286 + ], + "spans": [ + { + "bbox": [ + 133, + 275, + 480, + 286 + ], + "type": "text", + "content": "Table 2: Domain adaptation regression results (MAE) on Biwi Kinect (ResNet-18)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 213, + 297, + 402, + 425 + ], + "lines": [ + { + "bbox": [ + 213, + 297, + 402, + 425 + ], + "spans": [ + { + "bbox": [ + 213, + 297, + 402, + 425 + ], + "type": "table", + "html": "
MethodM→FF→MAvg.
Resnet-180.290.380.335
TCA0.310.390.350
MCD0.310.370.340
JDOT0.290.390.340
AFN0.320.410.365
DAN0.280.370.325
DANN0.300.370.335
RSD0.260.300.280
DARE-GRAM0.230.290.260
COD0.200.210.205
", + "image_path": "19b8a74dd7a5312347cdcc535d0e3d9e68391376c35efdbe1903e0c01893d185.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 480, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 480, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 480, + 592 + ], + "type": "text", + "content": "methods with marginal alignment indeed show limitations in the difficult transfer tasks. In such a scenario, the sufficiency of condition alignment is directly validated, where the MAE of COD is significantly lower than other SOTA methods. Thus, for real-world regression applications with complex shift, the COD model indeed shows superiority in both theoretical and methodological aspects. Comparison on MPI3D. In MPI3D, the resolution remains the same with dSprites, while the main challenge is the large sample size. In Tab. 3, the MAE of COD is competitive with the SOTA method DARE-GRAM in the first four tasks, while higher in the last two tasks. However, note that COD still achieves the second-best result in averaged MAE, and obtains the lowest MAE in task RL " + }, + { + "bbox": [ + 130, + 449, + 480, + 592 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 449, + 480, + 592 + ], + "type": "text", + "content": " T. In conclusion, the overall MAE of COD on three benchmarks is still lower, which demonstrates COD is generally superior to SOTA methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "Ablation study. The ablation experiment was conducted on the most challenging tasks from all three datasets: " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{C} \\rightarrow \\mathbf{S}" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{N} \\rightarrow \\mathbf{S}" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " from dSprites, " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{T} \\rightarrow \\mathbf{R}\\mathbf{L}" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{T} \\rightarrow \\mathbf{RC}" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " from MPI3D, and both " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\rightarrow \\mathbf{F}" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\rightarrow \\mathbf{M}" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " from Biwi Kinect. The major components of proposed method, i.e., MSE objective " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{src}" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": ", KGW distance " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{KGW}}" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": ", COD distance " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{COD}}" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " and modified COD " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{COD}_{mod}}" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": ", are evaluated, where the results are shown in Tab. 4." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "H. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 130, + 136, + 483, + 264 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 472, + 126 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 472, + 126 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 472, + 126 + ], + "type": "text", + "content": "Table 3: Domain adaptation regression results (MAE) on MPI3D (ResNet-18)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 130, + 136, + 483, + 264 + ], + "lines": [ + { + "bbox": [ + 130, + 136, + 483, + 264 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 483, + 264 + ], + "type": "table", + "html": "
MethodRL→RCRL→TRC→RLRC→TT→RLT→RCAvg.
Resnet-180.170.440.190.450.510.500.377
TCA0.170.420.190.420.500.500.373
MCD0.130.400.150.450.520.500.358
JDOT0.160.410.160.410.470.470.353
AFN0.180.450.200.460.530.530.390
DAN0.120.350.120.270.400.410.278
DANN0.090.240.110.410.480.370.283
RSD0.090.190.080.150.360.360.205
DARE-GRAM0.090.150.100.140.240.240.160
COD0.100.110.130.150.370.290.192
", + "image_path": "09bd9b74dfbfd2d5c97a9adb6169e74e23fa80570bacc224b88f5e4bd09c8669.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 134, + 299, + 480, + 386 + ], + "blocks": [ + { + "bbox": [ + 193, + 278, + 418, + 289 + ], + "lines": [ + { + "bbox": [ + 193, + 278, + 418, + 289 + ], + "spans": [ + { + "bbox": [ + 193, + 278, + 418, + 289 + ], + "type": "text", + "content": "Table 4: Ablation study on different metrics for DAR." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 299, + 480, + 386 + ], + "lines": [ + { + "bbox": [ + 134, + 299, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 134, + 299, + 480, + 386 + ], + "type": "table", + "html": "
MSEObjectivesdSpritesBiwi KinectMPI3D
dKGWdCODdCODmodC→SN→ST→RLT→RCM→FF→M
✓0.900.650.290.380.510.50
✓✓0.180.260.210.250.410.34
✓✓0.260.450.210.220.400.31
✓✓✓0.170.250.210.220.410.29
✓✓✓0.160.230.200.210.370.29
", + "image_path": "4b61c75db04ce08a8763f1e0494ec7093bdf4978230509a588a1b5cd60e90a01.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 412, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 568 + ], + "type": "text", + "content": "From the " + }, + { + "bbox": [ + 130, + 412, + 482, + 568 + ], + "type": "inline_equation", + "content": "2^{\\mathrm{nd}}" + }, + { + "bbox": [ + 130, + 412, + 482, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 412, + 482, + 568 + ], + "type": "inline_equation", + "content": "3^{\\mathrm{rd}}" + }, + { + "bbox": [ + 130, + 412, + 482, + 568 + ], + "type": "text", + "content": " rows, it can be observed that both OT-based metrics show promising results on DAR datasets, where the conditional alignment with COD shows better performance than marginal alignment with KGW on MPI3D and Biwi Kinect datasets. On the other hand, since dSprites induce more difficult transfer problems, the baseline performance is generally lower, and the pseudo labels for conditional alignment are less reliable. Therefore, the combination of KGW and COD is necessary to ensure higher performance on challenging tasks, where the results in the " + }, + { + "bbox": [ + 130, + 412, + 482, + 568 + ], + "type": "inline_equation", + "content": "4^{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 412, + 482, + 568 + ], + "type": "text", + "content": " row validate the mutually beneficial relation between these two objectives. Finally, the " + }, + { + "bbox": [ + 130, + 412, + 482, + 568 + ], + "type": "inline_equation", + "content": "5^{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 412, + 482, + 568 + ], + "type": "text", + "content": " row shows the effectiveness of rethinking the discriminability of COD, which implies the modifications on the moment statistics indeed ensure lower MAE values on all datasets. Overall, these results demonstrate that conditional alignment is indeed crucial for successful DAR, and the theoretical results and proposed method are generally valid." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Feature Visualization. t-SNE [34] is employed to visualize the 2-D features of different alignments on dSprites " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{C}\\rightarrow \\mathrm{N}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " task. (a) Before adaptation, distributions of source and target domain differ severely, as shown in Fig. 3(a), which implies predictor is infeasible for the target samples with a significant domain gap. (b) After marginal distribution alignment via the SOTA DARE-GRAM method, source and target representations are aligned globally, and the cross-domain distributions are overlapped on the whole. However, there are still misaligned local patterns where source and target samples of different colors are matched." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 192, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 192, + 91, + 447, + 102 + ], + "type": "text", + "content": "COD: Learning Conditional Invariant Representation for DAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 119, + 478, + 206 + ], + "blocks": [ + { + "bbox": [ + 134, + 119, + 478, + 206 + ], + "lines": [ + { + "bbox": [ + 134, + 119, + 478, + 206 + ], + "spans": [ + { + "bbox": [ + 134, + 119, + 478, + 206 + ], + "type": "image", + "image_path": "afb46b230699af1112d79d9ece81ccaee0e0e8721d8ce341d159b9a7ff6c1c47.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 215, + 482, + 251 + ], + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 251 + ], + "type": "text", + "content": "Fig. 3: t-SNE visualization of learned representations. Label values are denoted by color gradients. Eight label values are selected from the range of the variable for visualization. " + }, + { + "bbox": [ + 130, + 215, + 482, + 251 + ], + "type": "inline_equation", + "content": "+" + }, + { + "bbox": [ + 130, + 215, + 482, + 251 + ], + "type": "text", + "content": ": source samples; " + }, + { + "bbox": [ + 130, + 215, + 482, + 251 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 130, + 215, + 482, + 251 + ], + "type": "text", + "content": ": target samples." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 273, + 482, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 273, + 482, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 273, + 482, + 322 + ], + "type": "text", + "content": "Therefore, it is hard to fit clear level sets in Fig. 3 (b). (c) After conditional distribution alignment via COD, source and target distributions are aligned not only globally but also locally. As shown in Fig. 3 (c), representations are matched according to label values so the change of color shows a clear gradient and level" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 322, + 328, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 328, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 328, + 332 + ], + "type": "text", + "content": "sets suitable for both domains can be fitted." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 333, + 332, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 332, + 476 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 332, + 476 + ], + "type": "text", + "content": "Hyper-parameter. We investigate the sensitivity of hyper-parameters on dSprites " + }, + { + "bbox": [ + 130, + 333, + 332, + 476 + ], + "type": "inline_equation", + "content": "\\mathrm{C}\\rightarrow \\mathrm{N}" + }, + { + "bbox": [ + 130, + 333, + 332, + 476 + ], + "type": "text", + "content": " MAE values by varying " + }, + { + "bbox": [ + 130, + 333, + 332, + 476 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 333, + 332, + 476 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 333, + 332, + 476 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 333, + 332, + 476 + ], + "type": "text", + "content": " ,i.e., parameters of modified COD and KGW, are shown in Fig. 4. The results show that the performance of the COD-based DAR model is generally robust to the different choice of parameters. In fact, the MAE values of the COD-based model are stable in random experiments, which implies the proposed method is both robust from the view of randomness and parameter setting. In conclusion, the sta" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 340, + 333, + 479, + 437 + ], + "blocks": [ + { + "bbox": [ + 340, + 333, + 479, + 437 + ], + "lines": [ + { + "bbox": [ + 340, + 333, + 479, + 437 + ], + "spans": [ + { + "bbox": [ + 340, + 333, + 479, + 437 + ], + "type": "image", + "image_path": "fde948489ae7c2d8f5bd51ef3ce4e356503a10d0742673fe58a23006019ab258.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 447, + 481, + 470 + ], + "lines": [ + { + "bbox": [ + 339, + 447, + 481, + 470 + ], + "spans": [ + { + "bbox": [ + 339, + 447, + 481, + 470 + ], + "type": "text", + "content": "Fig.4: MAE values under different settings of hyper-parameters." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 476, + 476, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 476, + 476, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 476, + 476, + 490 + ], + "type": "text", + "content": "ble performance validates that the proposed model is also empirically reliable." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 508, + 220, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 508, + 220, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 508, + 220, + 521 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "In this work, we studied the limitations of existing DAR research, i.e., the misaligned intrinsic discriminant structure and negative transfer of marginal distribution matching framework. To deal with the challenges of conditional shift with continuous output space, we provide feasible solutions from both theoretical and methodological views, where the main results show that: 1) the cross-domain generalization regression error can be sufficiently bounded by the conditional discrepancy on continuous labels; 2) the proposed COD serves as a well-defined metric for conditional discrepancy measures and empirical modeling. With these guarantees, a COD-based DAR method is proposed with a discriminability enhancement mechanism. Numerical validations on standard DAR datasets validate the superiority of COD." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "type": "text", + "content": "H. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 236 + ], + "type": "text", + "content": "This work is supported in part by National Natural Science Foundation of China (Grant No. 62376291), in part by Guangdong Basic and Applied Basic Research Foundation (2023B1515020004), in part by Science and Technology Program of Guangzhou (2024A04J6413), in part by the Fundamental Research Funds for the Central Universities, Sun Yat-sen University (24xkjc013), in part by Guangdong Province Key Laboratory of Computational Science at Sun Yat-sen University (2020B1212060032), and in part by Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 255, + 197, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 255, + 197, + 268 + ], + "spans": [ + { + "bbox": [ + 133, + 255, + 197, + 268 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 280, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 280, + 480, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 280, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 138, + 280, + 480, + 301 + ], + "type": "text", + "content": "1. Bao, Y., Liu, Y., Wang, H., Lu, F.: Generalizing gaze estimation with rotation consistency. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4207-4216 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 303, + 480, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 303, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 138, + 303, + 480, + 324 + ], + "type": "text", + "content": "2. Ben-David, S., Blitzer, J., Crammer, K., Kulesza, A., Pereira, F., Vaughan, J.W.: A theory of learning from different domains. Mach. Learn. 79, 151-175 (2010)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 325, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 325, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 325, + 480, + 357 + ], + "type": "text", + "content": "3. Chen, L., Chen, H., Wei, Z., Jin, X., Tan, X., Jin, Y., Chen, E.: Reusing the task-specific classifier as a discriminator: Discriminator-free adversarial domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7181-7190 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 357, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 480, + 379 + ], + "type": "text", + "content": "4. Chen, X., Wang, S., Wang, J., Long, M.: Representation subspace distance for domain adaptation regression. In: Int. Conf. Mach. Learn. pp. 1749-1759 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 380, + 480, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 480, + 411 + ], + "type": "text", + "content": "5. Chen, Y., Song, S., Li, S., Wu, C.: A graph embedding framework for maximum mean discrepancy-based domain adaptation algorithms. IEEE Trans. Image Process. 29, 199-213 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 413, + 480, + 445 + ], + "type": "text", + "content": "6. Tachet des Combes, R., Zhao, H., Wang, Y.X., Gordon, G.J.: Domain adaptation with conditional distribution matching and generalized label shift. In: Adv. Neural Inform. Process. Syst. vol. 33, pp. 19276-19289 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 446, + 480, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 446, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 138, + 446, + 480, + 467 + ], + "type": "text", + "content": "7. Cortes, C., Mohri, M.: Domain adaptation in regression. In: ICALT. pp. 308-323 (2011)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 500 + ], + "type": "text", + "content": "8. Courty, N., Flamary, R., Habrard, A., Rakotomamonjy, A.: Joint distribution optimal transportation for domain adaptation. Adv. Neural Inform. Process. Syst. 30 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 501, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 480, + 522 + ], + "type": "text", + "content": "9. Courty, N., Flamary, R., Tuia, D.: Domain adaptation with regularized optimal transport. In: ECML PKDD. pp. 274-289 (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 522, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 522, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 134, + 522, + 480, + 555 + ], + "type": "text", + "content": "10. Damodaran, B.B., Kellenberger, B., Flamary, R., Tuia, D., Courty, N.: Deepjdot: Deep joint distribution optimal transport for unsupervised domain adaptation. In: Eur. Conf. Comput. Vis. pp. 447-463 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 555, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 555, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 134, + 555, + 480, + 587 + ], + "type": "text", + "content": "11. Dhaini, M., Berar, M., Honeine, P., Van Exem, A.: Unsupervised domain adaptation for regression using dictionary learning. Knowl.-Based Syst. 267, 110439 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 588, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 588, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 134, + 588, + 480, + 610 + ], + "type": "text", + "content": "12. Fanelli, G., Dantone, M., Gall, J., Fossati, A., Van Gool, L.: Random forests for real time 3d face analysis. Int. J. Comput. Vis. **101**, 437–458 (2013)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 611, + 480, + 632 + ], + "type": "text", + "content": "13. Fukumizu, K., Bach, F.R., Jordan, M.I.: Kernel dimension reduction in regression. Ann. Stat. 37(4), 1871-1905 (2009)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "text", + "content": "14. Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., March, M., Lempitsky, V.: Domain-adversarial training of neural networks. J. Mach. Learn. Res. 17(59), 1-35 (2016)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 192, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 192, + 91, + 447, + 102 + ], + "type": "text", + "content": "COD: Learning Conditional Invariant Representation for DAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "15. Geng, X., Zhou, Z.H., Smith-Miles, K.: Automatic age estimation based on facial aging patterns. IEEE Trans. Pattern Anal. Mach. Intell. 29(12), 2234-2240 (2007)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 181 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 181 + ], + "type": "text", + "content": "16. Gondal, M.W., Wuthrich, M., Miladinovic, D., Locatello, F., Breidt, M., Volchkov, V., Akpo, J., Bachem, O., Scholkopf, B., Bauer, S.: On the transfer of inductive bias from simulation to the real world: a new disentanglement dataset. Adv. Neural Inform. Process. Syst. 32 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 214 + ], + "type": "text", + "content": "17. Gong, B., Shi, Y., Sha, F., Grauman, K.: Geodesic flow kernel for unsupervised domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2066-2073 (2012)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 481, + 246 + ], + "type": "text", + "content": "18. Goswami, S., Kontolati, K., Shields, M.D., Karniadakis, G.E.: Deep transfer operator learning for partial differential equations under conditional shift. Nat. Mach. Intell. 4(12), 1155-1164 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 247, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 481, + 278 + ], + "type": "text", + "content": "19. Higgins, I., Matthew, L., Pal, A., Burgess, C., Glorot, X., Botvinick, M., Mohamed, S., Lerchner, A.: beta-vae: Learning basic visual concepts with a constrained variational framework. In: Int. Conf. Learn. Represent. (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 279, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 481, + 311 + ], + "type": "text", + "content": "20. Hoffman, J., Tzeng, E., Park, T., Zhu, J.Y., Isola, P., Saenko, K., Efros, A., Darrell, T.: Cycada: Cycle-consistent adversarial domain adaptation. In: Int. Conf. Mach. Learn. pp. 1989-1998 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 312, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 481, + 343 + ], + "type": "text", + "content": "21. Hu, J., Zhong, H., Yang, F., Gong, S., Wu, G., Yan, J.: Learning unbiased transferability for domain adaptation by uncertainty modeling. In: Eur. Conf. Comput. Vis. pp. 223-241 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 344, + 481, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 344, + 481, + 375 + ], + "spans": [ + { + "bbox": [ + 132, + 344, + 481, + 375 + ], + "type": "text", + "content": "22. Jiang, J., Ji, Y., Wang, X., Liu, Y., Wang, J., Long, M.: Regressive domain adaptation for unsupervised keypoint detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6780-6789 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 376, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 481, + 407 + ], + "type": "text", + "content": "23. Kirchmeyer, M., Rakotomamonjy, A., de Bezenac, E., Gallinari, P.: Mapping conditional distributions for domain adaptation under generalized target shift. In: Int. Conf. Learn. Represent. (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 408, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 481, + 439 + ], + "type": "text", + "content": "24. Lee, C.Y., Batra, T., Baig, M.H., Ulbricht, D.: Sliced wasserstein discrepancy for unsupervised domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10285-10295 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 441, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 481, + 472 + ], + "type": "text", + "content": "25. Li, B., Wang, Y., Zhang, S., Li, D., Keutzer, K., Darrell, T., Zhao, H.: Learning invariant representations and risks for semi-supervised domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1104-1113 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 473, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 473, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 473, + 481, + 493 + ], + "type": "text", + "content": "26. Li, X., Li, J., Zuo, L., Zhu, L., Shen, H.T.: Domain adaptive remaining useful life prediction with transformer. IEEE Trans. Instrum. 71, 1-13 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 495, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 481, + 525 + ], + "type": "text", + "content": "27. Li, Z., Zhang, K., Liu, Y., Zou, Y., Ding, G.: A novel remaining useful life transfer prediction method of rolling bearings based on working conditions common benchmark. IEEE Trans. Instrum. 71, 1-9 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 526, + 481, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 526, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 481, + 548 + ], + "type": "text", + "content": "28. Long, M., Cao, Y., Wang, J., Jordan, M.: Learning transferable features with deep adaptation networks. In: Int. Conf. Mach. Learn. pp. 97-105 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 548, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 481, + 568 + ], + "type": "text", + "content": "29. Long, M., Cao, Z., Wang, J., Jordan, M.I.: Conditional adversarial domain adaptation. Adv. Neural Inform. Process. Syst. 31 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 569, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 481, + 590 + ], + "type": "text", + "content": "30. Long, M., Zhu, H., Wang, J., Jordan, M.I.: Deep transfer learning with joint adaptation networks. In: Int. Conf. Mach. Learn. pp. 2208-2217 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 590, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 481, + 611 + ], + "type": "text", + "content": "31. Luo, Y.W., Ren, C.X.: Conditional bures metric for domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 13989-13998 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 612, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 481, + 632 + ], + "type": "text", + "content": "32. Luo, Y.W., Ren, C.X.: Mot: Masked optimal transport for partial domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3531-3540 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "33. Luo, Y.W., Ren, C.X., Dai, D.Q., Yan, H.: Unsupervised domain adaptation via discriminative manifold propagation. IEEE Trans. Pattern Anal. Mach. Intell. 44(3), 1653-1669 (2022)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "H. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "34. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. J. Mach. Learn. Res. 9(11) (2008)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 159 + ], + "type": "text", + "content": "35. Mansour, Y., Mohri, M., Rostamizadeh, A.: Domain adaptation: Learning bounds and algorithms. arXiv preprint arXiv:0902.3430 (2009)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 191 + ], + "type": "text", + "content": "36. Mao, W., Ge, Y., Shen, C., Tian, Z., Wang, X., Wang, Z., den Hengel, A.v.: Poseur: Direct human pose regression with transformers. In: Eur. Conf. Comput. Vis. pp. 72-88 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 214 + ], + "type": "text", + "content": "37. de Mathelin, A., Richard, G., Deheeger, F., Mougeot, M., Vayatis, N.: Adversarial weighting for domain adaptation in regression. In: ICTAI. pp. 49-56 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 214, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 246 + ], + "type": "text", + "content": "38. Nejjar, I., Wang, Q., Fink, O.: Dare-gram: Unsupervised domain adaptation regression by aligning inverse gram matrices. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 11744–11754 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 246, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 278 + ], + "type": "text", + "content": "39. Nikzad-Langerodi, R., Zellinger, W., Saminger-Platz, S., Moser, B.A.: Domain adaptation for regression under beer-lambert's law. Knowl.-Based Syst. 210, 106447 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 300 + ], + "type": "text", + "content": "40. Pan, S.J., Tsang, I.W., Kwok, J.T., Yang, Q.: Domain adaptation via transfer component analysis. IEEE Trans. Neural Netw. 22(2), 199-210 (2010)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 300, + 480, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 321 + ], + "type": "text", + "content": "41. Pei, Z., Cao, Z., Long, M., Wang, J.: Multi-adversarial domain adaptation. In: AAAI. vol. 32 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 321, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 480, + 354 + ], + "type": "text", + "content": "42. Peng, R., Wang, R., Wang, Z., Lai, Y., Wang, R.: Rethinking depth estimation for multi-view stereo: A unified representation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8645-8654 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "text", + "content": "43. Quinonero-Candela, J., Sugiyama, M., Schwaighofer, A., Lawrence, N.D.: Dataset shift in machine learning. Mit Press (2008)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "text", + "content": "44. Ren, C.X., Luo, Y.W., Dai, D.Q.: Buresnet: Conditional bures metric for transferable representation learning. IEEE Trans. Pattern Anal. Mach. Intell. 45(4), 4198-4213 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 407, + 480, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 440 + ], + "type": "text", + "content": "45. Saito, K., Watanabe, K., Ushiku, Y., Harada, T.: Maximum classifier discrepancy for unsupervised domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3723-3732 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 440, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 480, + 460 + ], + "type": "text", + "content": "46. Shen, J., Qu, Y., Zhang, W., Yu, Y.: Wasserstein distance guided representation learning for domain adaptation. In: AAAI. vol. 32 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 482 + ], + "type": "text", + "content": "47. Singh, A., Chakraborty, S.: Deep domain adaptation for regression. Development and Analysis of Deep Learning Architectures pp. 91-115 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "type": "text", + "content": "48. Song, L., Huang, J., Smola, A., Fukumizu, K.: Hilbert space embeddings of conditional distributions with applications to dynamical systems. In: Int. Conf. Mach. Learn. pp. 961-968 (2009)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 514, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 536 + ], + "type": "text", + "content": "49. Sun, B., Feng, J., Saenko, K.: Return of frustratingly easy domain adaptation. In: AAAI. vol. 30 (2016)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "type": "text", + "content": "50. Tang, H., Jia, K.: Discriminative adversarial domain adaptation. In: AAAI. vol. 34, pp. 5940-5947 (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 557, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 480, + 578 + ], + "type": "text", + "content": "51. Teshima, T., Sato, I., Sugiyama, M.: Few-shot domain adaptation by causal mechanism transfer. In: Int. Conf. Mach. Learn. pp. 9458-9469 (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 578, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 600 + ], + "type": "text", + "content": "52. Tzeng, E., Hoffman, J., Saenko, K., Darrell, T.: Adversarial discriminative domain adaptation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7167-7176 (2017)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "text", + "content": "53. Wang, H., Shen, T., Zhang, W., Duan, L.Y., Mei, T.: Classes matter: A fine-grained adversarial approach to cross-domain semantic segmentation. In: Eur. Conf. Comput. Vis. pp. 642-659 (2020)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "54. Wang, W., Li, H., Ding, Z., Nie, F., Chen, J., Dong, X., Wang, Z.: Rethinking maximum mean discrepancy for visual domain adaptation. IEEE Trans. Neural Netw. Learn. Syst. (2021)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 193, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 193, + 91, + 447, + 102 + ], + "type": "text", + "content": "COD: Learning Conditional Invariant Representation for DAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 347 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "55. Wang, Y., Jiang, Y., Li, J., Ni, B., Dai, W., Li, C., Xiong, H., Li, T.: Contrastive regression for domain adaptation on gaze estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19376-19385 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "text", + "content": "56. Wei, P., Ke, Y., Ong, Y.S., Ma, Z.: Adaptive transfer kernel learning for transfer gaussian process regression. IEEE Trans. Pattern Anal. Mach. Intell. (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "text", + "content": "57. Wu, J., He, J., Wang, S., Guan, K., Ainsworth, E.: Distribution-informed neural networks for domain adaptation regression. Adv. Neural Inform. Process. Syst. 35, 10040-10054 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "58. Xu, R., Li, G., Yang, J., Lin, L.: Larger norm more transferable: An adaptive feature norm approach for unsupervised domain adaptation. In: Int. Conf. Comput. Vis. pp. 1426-1435 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "59. Zhang, Z., Wang, M., Nehorai, A.: Optimal transport in reproducing kernel hilbert spaces: Theory and applications. IEEE Trans. Pattern Anal. Mach. Intell. 42(7), 1741-1754 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "text", + "content": "60. Zhao, H., Combes, R.T.D., Zhang, K., Gordon, G.: On learning invariant representations for domain adaptation. In: Int. Conf. Mach. Learn. pp. 7523-7532 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "type": "text", + "content": "61. Zhao, H., Coston, A., Adel, T., Gordon, G.J.: Conditional learning of fair representations. In: Int. Conf. Learn. Represent. (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 347 + ], + "type": "text", + "content": "62. Zhao, H., Dan, C., Aragam, B., Jaakkola, T.S., Gordon, G.J., Ravikumar, P.: Fundamental limits and tradeoffs in invariant representation learning. J. Mach. Learn. Res. 23(340), 1-49 (2022)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 224, + 102 + ], + "type": "text", + "content": "H. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/COHO_ Context-Sensitive City-Scale Hierarchical Urban Layout Generation/ad29cc50-ea60-4c58-8c17-9b4b4ddc225f_content_list.json b/2024/COHO_ Context-Sensitive City-Scale Hierarchical Urban Layout Generation/ad29cc50-ea60-4c58-8c17-9b4b4ddc225f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b5ffc16ace9eed1b63198ccb4d8a57172f414616 --- /dev/null +++ b/2024/COHO_ Context-Sensitive City-Scale Hierarchical Urban Layout Generation/ad29cc50-ea60-4c58-8c17-9b4b4ddc225f_content_list.json @@ -0,0 +1,1692 @@ +[ + { + "type": "text", + "text": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation", + "text_level": 1, + "bbox": [ + 218, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Liu He, Daniel Aliaga", + "bbox": [ + 408, + 212, + 594, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Purdue University, USA", + "bbox": [ + 418, + 239, + 583, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{he425,aliaga}@purdue.edu", + "bbox": [ + 403, + 253, + 599, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eeb1e4304f82043ce7508d9d636bb723fd83c5cac48bbe79bb742ab17b5baafa.jpg", + "image_caption": [ + "LayoutDM $(CTS = -2.20)$" + ], + "image_footnote": [], + "bbox": [ + 259, + 301, + 421, + 416 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/36b085177dada633a268f3c5ddb070f50057286d0b5db2990542f7f062b97496.jpg", + "image_caption": [ + "Ours $(CTS = 0.21)$" + ], + "image_footnote": [], + "bbox": [ + 433, + 301, + 581, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ce3753b00c29ac292e17d8c267271a185b6fa36ae11a58b86b5899746e1ebdbf.jpg", + "image_caption": [ + "GlobalMapper $(CTS = 0.62)$" + ], + "image_footnote": [], + "bbox": [ + 593, + 303, + 741, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ca21aa3b26064b52762e7dc614eaae7900fb1e159837833012992d80dc7ee4f1.jpg", + "image_caption": [ + "Random $(CTS\\ll 0)$" + ], + "image_footnote": [], + "bbox": [ + 259, + 430, + 421, + 544 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/edde6205d9b52d281d179b0fe6c2f5c40ad8b41e31b4c6187bf9c2acd950add2.jpg", + "image_caption": [ + "Real Data $(CTS\\rightarrow 0)$" + ], + "image_footnote": [], + "bbox": [ + 433, + 430, + 581, + 544 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1b121afc5408cfebba0a8399974f27efb816b208138162c225ec21c17054eef9.jpg", + "image_caption": [ + "Identical $(CTS\\gg 0)$", + "Fig. 1: Context-Sensitive Generation. Our method (Green) pursues realistic context harmonization among neighboring city blocks as real data (bottom-middle). Other methods (e.g. LayoutDM [29], GlobalMapper [20]) show over-diversity/-similarity in city-scale layout generation (evaluated by Context Score $CTS$ as Eq. 4.). Fully random and identical layouts are synthetically generated to illustrate extreme cases." + ], + "image_footnote": [], + "bbox": [ + 593, + 431, + 741, + 545 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The generation of large-scale urban layouts has garnered substantial interest across various disciplines. Prior methods have utilized procedural generation requiring manual rule coding or deep learning needing abundant data. However, prior approaches have not considered the context-sensitive nature of urban layout generation. Our approach addresses this gap by leveraging a canonical graph representation for the entire city, which facilitates scalability and captures the multilayer semantics inherent in urban layouts. We introduce a novel graph-based masked autoencoder (GMAE) for city-scale urban layout generation. The method encodes attributed buildings, city blocks, communities and cities into a unified graph structure, enabling self-supervised", + "bbox": [ + 259, + 686, + 743, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "masked training for graph autoencoder. Additionally, we employ scheduled iterative sampling for 2.5D layout generation, prioritizing the generation of important city blocks and buildings. Our approach achieves good realism, semantic consistency, and correctness across the heterogeneous urban styles in 330 US cities. Codes and datasets are released at https://github.com/Arking1995/COHO.", + "bbox": [ + 259, + 147, + 743, + 231 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keywords: Layout Generation $\\cdot$ Urban Modeling $\\cdot$ Masked Graph Autoencoder $\\cdot$ Context Sensitivity", + "bbox": [ + 259, + 244, + 740, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 297, + 375, + 314 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large-scale urban layout generation has received significant interest by computer vision, urban planning and related disciplines. In particular, cities easily have 1,000 to 50,000 city blocks spanning a mostly horizontal terrain. The configuration of each city block varies significantly across the city, yet there is some configuration similarity amongst subsets within and among cities. Reconstructing existing layouts and generating future planned layouts are crucial tasks for urban simulation, digital twins, and game/content design.", + "bbox": [ + 212, + 330, + 785, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prior works have generated urban layouts from a variety of sources. Procedural layout generation [39, 57] used hand-coded rules to generate cities, and inverse procedural modeling [5] extracted the rules from data and then enabled generation. More recently, deep learning has promoted data-driven approaches that produce either pixel-based layout generation (e.g., InfiniCity [38], CityDreamer [61], CityGEN [13]) or graph-based approaches for general polygonal shape layouts (e.g., LayoutTransformer [17], VTN [2]) and for urban scenarios (e.g., BlockPlanner [62], GlobalMapper [20], BuildingGAN [11]). However, none of these techniques explicitly recover and consider the context-sensitive nature of urban layout generation (Fig. 1). In other words, the layout of buildings, blocks, and communities cannot be treated in isolation. The multi-layer semantics are critical to generating realistic and plausible urban layouts. While such multilevel structure (e.g., from single building to a community or subset of a city) has been specified via hand-coded rules, it has not been discovered, organized and used for generation in data-driven deep learning approaches.", + "bbox": [ + 212, + 436, + 787, + 662 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach builds on three key observations:", + "bbox": [ + 238, + 662, + 583, + 678 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Representation: Cities, communities, blocks, buildings, and roads are arbitrarily shaped and with complex topology. Hence, a graph-based representation can better capture the regular and non-regular configurations, as opposed to the limiting regular structure of an image. Further, graphs can be more compact which in turn supports better scalability. Altogether, having buildings or city blocks as atomic units, instead of pixels, will contribute to improving correctness, realism, and consistency with priors.", + "2. Context-Sensitivity: Buildings, city blocks, and communities are built considering their neighboring structures and not in isolation. Hence, a generator should consider this multi-layer structure and inter-dependency. In" + ], + "bbox": [ + 220, + 688, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "He, L. & Aliaga, D.", + "bbox": [ + 271, + 114, + 403, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "order words, the style features at one level (e.g., size, shape, location) depend on the surrounding context.", + "bbox": [ + 240, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Prioritization: The stylistic and semantic importance, or priority, of city blocks and buildings varies. This implies that a generation scheme should generate the most important city blocks and buildings first and then fill in the rest, as opposed to a region growing approach for example. Overall, prioritization will improve correctness and realism as well as enable large-scale urban layout generation.", + "bbox": [ + 222, + 176, + 784, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a novel Graph-based Masked AutoEncoder (GMAE) for large-scale multi-layer 2.5D urban layout generation. First, we define a canonical graph-based representation capturing the multi-layer context sensitivity of any city's urban layouts (Fig. 2). The entire graph $G$ corresponds to a city. Subsets of the graph map to communities and each node $b$ amounts to a city block. Further, each node encodes its building layouts, shapes, and heights as a quantized feature by well-trained quantizor. Second, we utilize our GMAE to learn the multi-layer stylistic behavior of many urban layouts (i.e., 330 cities).", + "bbox": [ + 212, + 273, + 782, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Using subsets of the graph (or communities), node features are randomly masked and self-supervised GMAE training is performed to learn node feature reconstruction (Fig. 3). Third, utilizing the well-trained GMAE as a generator, we perform an iterative priority-based scheduling to generate a large layout or to complete a city fragment; i.e., each iteration predicts node features for the entire graph but only the most confident ones are preserved for the next step (Fig. 4). This methodology is able to leverage any percentage of given prior (e.g., $[0,100\\%]$ ) for urban layout completion and generation.", + "bbox": [ + 212, + 396, + 784, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To the best of our knowledge, our approach is the first data-driven deep-learning method to enable large-scale 2.5D layout generation with good realism, semantic consistency, and correctness. We show the capability to generalize to the heterogeneous urban layout styles of 330 cities in the US (e.g., all cities with a population $>100\\mathrm{K}$ , totaling 833,473 city blocks and 17,663,607 buildings). In addition, we provide comparisons to several prior works including SDXL [48]), VTN [2], LayoutDM [29], and GlobalMapper [20] in Sec. 4 and exceed their performance in multiple well-known geometric and perceptual metrics.", + "bbox": [ + 212, + 516, + 784, + 635 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions include the following:", + "bbox": [ + 238, + 637, + 529, + 652 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A canonical graph representation for large-scale 2.5D urban layouts. It encodes global and local node features and neighboring relations for buildings, city blocks, and communities within a city.", + "- A self-supervised Graph-based Masked AutoEncoder (GMAE) enabling arbitrarily shaped city layout generation with consistency, correctness, and realism", + "- Priority-based scheduled iterative sampling for controllable urban layout synthesis including completion, refinement, and generation with any percentage of priors ([0, 100%]).", + "- A open dataset of road network and urban layouts of 330 US cities for benchmarking current and future urban layout generation methods (https://huggingface.co/datasets/Arking95/COHO)." + ], + "bbox": [ + 225, + 659, + 782, + 837 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation", + "bbox": [ + 225, + 113, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 387, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Layout Generation. Numerous data-driven deep learning methods have been proposed for layout generation; for example, generalized layout synthesis [2,17, 31,34,64], document layouts [21,30,46,55,67], urban land lots [22,62,66], indoor scenes [42,44,60], and 3D buildings [3,11,45]. Recently, diffusion models have also been used for layout generation tasks (e.g., [9,21,28,29]). However, most of these works assume 2D layouts generated on a rectangular empty canvas. The position and size of each bounding box is often limited to a finite set of categories and to rectilinear alignments. This results in poor diversity and low realism for urban layout generation tasks. BlockPlanner [62] and GlobalMapper [20] adapt layout generation to either rectilinear or arbitrarily-shaped city block generation. However, all previous works assume each layout, or city block, is independent. This is not the case in urban environments where adjacent and nearby city blocks may share a common style or be positioned in a purposeful multi-block arrangement. Without a contextual guidance, city block generation may result in either unwanted harmonization or unexpected diversity between neighboring blocks (see Fig. 1). Thus, we pursue city-scale urban layout generation with the help of context-aware neighboring information.", + "bbox": [ + 212, + 178, + 787, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Infinite Visual Synthesis. With bombing of visual deep learning works [27, 40, 50, 51, 53, 54], researchers have explored synthesizing large, potentially unbounded images. For example, omnidirectional image synthesis from multiple smaller field-of-view images [36, 37]. Other works focus on unbounded 3D scene generation [8, 12, 49]. Particularly relevant are several recent infinite urban scene generators [13, 38, 61] that directly adapt image synthesis methods (e.g. InfinityGAN [37], MaskGIT [10]). The synthesis performs an autoregressive outpainting process using layout image patches. However, patch-based training and inference doesn't learn global features nor their interdependencies. Rather, it learns a very local behavior. Prior pixel-based methods acknowledge limited realism and semantics [13, 38, 61]. Often, post-processing [13, 61], and human-in-the-loop editing [38] is required to refine the generated images to vector-based urban layouts. Even the latest stable diffusion model SDXL [48] seeks high-resolution image synthesis but may not produce realistic urban layouts (Fig. 5).", + "bbox": [ + 212, + 463, + 787, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learning based on Quantized Representations. Vector quantization has been a prevailing stepping-stone technique for text (e.g., Word2Vec [41]), image (e.g., VQVAE/GAN [16, 56], [15]), and video synthesis (e.g., SORA [6, 47], [63]). Pretrained quantizers provide a scalable token representation benefiting high-level and large-scale synthesis. A well-trained quantizer combined with self-supervised masked training often performs well in reconstructing the original signal during generative tasks. BERT [14] provides plausible text-based representation learning. MAE [19] defines a masked autoencoder for image space which excels at multiple downstream tasks. MAE also provides representation", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "He, L. & Aliaga, D.", + "bbox": [ + 271, + 114, + 403, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "learning for a broad range of node/graph feature synthesis [25, 26]. Specifically, MaskGIT [10] and MAGE [35] utilize MAE frameworks for impressive image generation tasks. The idea of a generative masked autoencoder based on pretrained quantization was inspirational to our city-scale urban layout approach.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 229, + 330, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First we describe our canonical graph representation of 2.5D urban layouts (Fig. 2). Second, we explain the training process for our GMAE based on the graphs for many cities (Fig. 3). Third, we show our priority-based scheduling for iteratively generating city-scale urban layouts given any amount of prior (Fig. 4).", + "bbox": [ + 212, + 263, + 784, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Canonical Graph Representation", + "text_level": 1, + "bbox": [ + 214, + 347, + 532, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A city is represented by a graph $G = \\{B, E\\}$ , where $B$ is a set of nodes $B = \\{b_i | i \\in [1, N]\\}$ such that $b_i$ represents each city block (Fig. 2), and any connected subgraph of $G$ corresponds to a community. $N = |B|$ is the total number of blocks for a given city. $E$ is a set of edges $E = \\{e_{ij} | i, j \\in [1, N]\\}$ representing mutually adjacent nodes (or city blocks). $K = |E|$ is the total number of edges.", + "bbox": [ + 212, + 373, + 784, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Node (or City Block). Each node, or city block, $b_{i}$ is a region defined by an enclosing road network and contains a set of block-level features and building-level features. The per-node features include:", + "bbox": [ + 212, + 472, + 784, + 517 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- $s_i$ is the shape and location feature vector of a city block contour. It includes four features: (1) aspect ratio, (2) total block area, (3) ratio of total block area over the block's convex-hull area, and (4) relative distance from current block to the centroid of the city.", + "- $q_{i}$ is a 512-dimensional vector from a codebook $C$ . It hierarchically captures the buildings layouts and their configuration within $b_{i}$ ." + ], + "bbox": [ + 225, + 531, + 782, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We concatenate these node features to form a 516-dimensional vector $\\left([s_i, q_i]\\right)$ , which is able to represent our large and heterogeneous set of blocks and buildings.", + "bbox": [ + 212, + 635, + 782, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Building Layout Quantization. The 512-dimensional vector $q_{i}$ describing the building layouts within a city block is obtained from a quantized codebook $C = \\{1,2,\\dots,L\\}$ with quantization level $L = |C|$ . To create the codebook, we train a block-level variational autoencoder (BVAE) for building layout within a city block, using a canonical graph structure for representing the multiple building shapes, positions, and heights inside a single city block. Using this pretrained encoder, the latent vector of all possible building layouts will form a numerical distribution for each latent dimension. We further quantize each of these distributions into $L$ bins with equal percentiles. Then, each latent value is replaced by the rank of the bin containing its value. Therefore, any building", + "bbox": [ + 212, + 688, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation", + "bbox": [ + 225, + 113, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/beb3bd749036747d4c0ba7eb9860694e92bf71f2956b85b6de9b26f6a9d12de5.jpg", + "image_caption": [ + "Fig. 2: Canonical Graph Representation. Our method represents a city as a canonical graph $G$ . Each node $b_{i}$ represent a single city block, and each edge $e_{ij}$ connects spatially adjacent blocks. Each block/node corresponds to a set of node features $s_i$ and a quantized vector $q_{i}$ hierarchically capturing enclosed building layouts. Edge feature $d_{ij}$ encodes distances between block centroids. Graph $G$ is used for GMAE training." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 759, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "layout within a city block is quantized into a 512-dimensional categorical index vector $q_{i} = [c_{1},\\dots,c_{512}]$ , where each $c\\in C$ . Note that our codebook is defined after BVAE training and not dynamically trainable as in VQGAN/VQVAE [16,56]. After performing a set of detailed evaluations using different training schemes, encoding models, and quantization levels $L$ (see Sec. 4.3), we selected the Graph Attention Network (GAT) [59] as the backbone for the BVAE. Further, we select $L = 20$ to balance compactness and representation ability after quantization. BVAE and quantization details are found in Supplementary Materials.", + "bbox": [ + 212, + 422, + 787, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Edge. Each edge $e_{ij}$ contains a feature $d_{ij}$ storing the relative distance between the centroids of the adjacent city blocks $b_i$ and $b_j$ . This distance-weighted adjacency helps encode the relationship between neighboring blocks.", + "bbox": [ + 212, + 566, + 785, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Altogether, the shape, style, and multi-layer urban layout of a city with $N$ nodes and $K$ edges is represented by a total of $(516N + K)$ variables. This canonical graph is used for the self-supervised GMAE training described below.", + "bbox": [ + 212, + 611, + 785, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Graph-based Masked AutoEncoder", + "text_level": 1, + "bbox": [ + 214, + 679, + 552, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train our GMAE using a self-supervised process. A road-only city (or community) is a graph (or subgraph) $G$ where building layout features $q_{i}$ of every node (block) are missing, but $s_i$ and $e_{ij}$ are preserved. A context-sensitive node feature generation process corresponds to recreating the purposefully removed building layout features.", + "bbox": [ + 212, + 703, + 785, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During training (Fig. 3), we use dynamic masking ratios of $q_{i}$ in order to learn how to regenerate the original features. In each training iteration, the mask ratio $m$ is sampled from a truncated Gaussian distribution as recommended by Li et al. [35]. The ratio is obtained from the range $m \\in [0.5,1.0]$ and is centered at", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "He, L. & Aliaga, D.", + "bbox": [ + 271, + 114, + 403, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b4bd815f63a9b166abdc9f26433bee9e908635be57d3c82ee2b98cdc8964af06.jpg", + "image_caption": [ + "Fig. 3: Graph-based Masked Autoencoder. Given a canonical city graph, quantized building layout features $Q$ are masked with dynamic masking ratios $m \\in [0.5,1.0]$ , while block shape and location features $S$ are kept. The GNN encoder uses message passing between neighboring nodes to obtain the context-aware node features $F$ . The decoder uses $F$ to reconstruct $Q'$ . The predicted $Q'$ are decoded to 2.5D urban layouts." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 251 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "0.55. We denote building layout features as $Q = \\{q_{i} | i \\in [1, N]\\}$ and those after masking as $Q^{mask}$ . Block features are denoted as $S = \\{s_{i} | i \\in [1, N]\\}$ . The graph neural network (GNN) encoder is denoted as $f_{E}$ , and node feature decoder as $f_{D}$ . Thus the process of our GMAE can be written as:", + "bbox": [ + 212, + 359, + 787, + 421 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nF = f _ {E} ([ Q ^ {m a s k}, S ], E), \\quad Q ^ {\\prime} = f _ {D} (F) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 436, + 785, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $F = \\{f_i | i \\in [1, N]\\}$ indicates the context-sensitive node features obtained by message passing between neighboring nodes by $f_E$ . Extracted $F$ is processed by $f_D$ to reconstruct building layout features $Q'$ .", + "bbox": [ + 212, + 460, + 787, + 507 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The backbones of $f_{E}$ and $f_{D}$ can be any GNN, including GAT [58], GCN [33], and GraphSAGE [18]. After experimentation, we select GAT as $f_{E}$ because of its best context-sensitive feature extraction ability, and we choose a straightforward MLP as $f_{D}$ for efficient processing. We observed that the re-masking trick and complex GNN decoder utilized by [25,26] requires more memory and calculation but has no net benefit in our application. Since $Q$ contains quantized index vectors, we use cross entropy loss for self-supervised reconstruction training on masked nodes. Detailed ablations of masking strategy and model selections are provided in Sec. 4.3.", + "bbox": [ + 212, + 507, + 787, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our objective function is as follows, where $L$ indicates the length of the quantized codebook $C$ :", + "bbox": [ + 212, + 643, + 787, + 672 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {r e c o n} = - \\sum_ {i = 1} ^ {L} \\left[ Q _ {i} \\log \\left(Q _ {i} ^ {\\prime}\\right) \\right] ^ {m a s k} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 686, + 785, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To capture community behavior (and also reduce GPU memory usage), we randomly sample a batch of subgraphs from $G$ for each training iteration. The nodes in each subgraph are sampled using a chosen radius from a random center node. While community sizes vary significantly, we found that a practical compromise is to use 500 meters as the radius value. This radius ensures to cover a reasonable number of the city blocks with a typical census tract from governmental CJEST dataset [1].", + "bbox": [ + 212, + 734, + 787, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation", + "bbox": [ + 225, + 113, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d91eb5da47464a7c083d15174c7970e1f4a263d9ab769e3268343e3253a648ab.jpg", + "image_caption": [ + "Fig. 4: Priority-based Scheduled Generation. We iteratively utilize pretrained GMAE to reconstruct masked node features. In each iteration, we accept a certain ratio of predicted nodes decided by the scheduling function $\\beta(t) = 1 - \\cos(t / T)$ . We obtain a full graph after $T$ iterations." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 260 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Priority-based Scheduled Generation", + "text_level": 1, + "bbox": [ + 215, + 357, + 565, + 372 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While in theory the trained GMAE can predict all nodes in a single pass, the generation quality is much better when performed iteratively. Recent published papers (e.g., [10,35]) have found that gradually sampling with a reasonable number of iterations improves generation quality. In urban scenarios, small number of representative city blocks are important to indicate context behavior. We intend to generate those blocks at early iterations and then span neighboring blocks.", + "bbox": [ + 212, + 385, + 785, + 476 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 4, we propose a $T$ -iteration sampling strategy similar to MaskGIT [10] and MAGE [35]. At each iteration, GMAE predicts per-node building layout features for all remaining nodes. Then we rank generated nodes by their prediction confidence. The generated nodes with the highest prediction confidence are accepted. The accepted nodes after the current iteration $t \\in [1,T]$ are scheduled by a cosine function $\\beta(t) = 1 - \\cos(t / T)$ . When $t$ is small, the ratio is near to 0 causing few nodes to be accepted in initial iterations. As iterations continue, $t$ gets bigger and hence the number of accepted nodes during each iteration will also increase. This behavior follows the prioritization observation in Sec. 1 which causes distinguished, or unique, city blocks to be generated first and essentially guide the layout style for nearby city blocks. Typically we set $T = 12$ which results in dozens of nodes being generated as is the typical size of a community. Detailed ablations of alternative sampling schedules are provided in Sec. 4.3.", + "bbox": [ + 212, + 477, + 787, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 700, + 375, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 734, + 455, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Open Dataset We have created a vector-based urban layout dataset of all US cities with a population of 100K or more. We believe this dataset will be of significant use to the community to further develop urban layout generation methods and thus we will release the dataset together with the paper. For each city, we define a rectangular bounding box containing most of the metropolitan area. Then,", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "He, L. & Aliaga, D.", + "bbox": [ + 271, + 114, + 403, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "we use this bounding box to extract data from OpenStreetMap (OSM) [43], Microsoft Building Footprints (MSF) [23], and Topologically Integrated Geographic Encoding and Referencing (TIGER) dataset [7]. The data we collected includes: (1) road network vectors and attributes, (2) building footprints and heights, and (3) block-level social economic metrics. Specifically, we collect road networks and block contours from TIGER. Then we look for building footprints and heights from OSM and MSF within each city block contour. We choose the one with more and accurate details (e.g., containing more buildings, having features that match those in TIGER). Over all 330 cities, this dataset has a total of 833,473 city blocks, and 17,663,607 buildings. We set training, validation, testing split as $70\\%$ , $20\\%$ , and $10\\%$ respectively.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training Details The training has two stages. We first use all city blocks for self-supervised training of a graph-based Block-level Variational AutoEncoder (BVAE). Given this trained block-level encoder, we perform the latent quantization for all building layouts in each city block. Then our GMAE is trained based on canonical city graph representation. Training time for BVAE and GMAE typically takes 12 hours and 15 hour respectively by a single A5000 GPU. We will release our codes upon acceptance.", + "bbox": [ + 212, + 337, + 787, + 443 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Inferencing using our graph-based approach is seemingly instantaneous for all the results shown in this paper. This is contrast to the VTN [2] and SDXL [48] approaches which take significantly more time to infer and to train.", + "bbox": [ + 212, + 443, + 785, + 489 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Comparisons", + "text_level": 1, + "bbox": [ + 215, + 515, + 370, + 530 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We compare our method to several alternate city-scale layout generators (Tab. 1). Specifically, we randomly select 100 communities from among our test set, total 3700 city blocks. All methods generate the layout of the 100 communities in one pass, and without any post-processing or human-in-the-loop refinement. Implementations are downloaded from the author's repositories and trained with default parameter settings. For fairness, we convert our dataset to token sequences, graphs, or image patches $(1024\\times 1024)$ as needed. VTN [2], LayoutDM [29], GlobalMapper [20] are retrained for single block generation. SDXL [48] is fine-tuned for outpainting by using the upper half to predict the lower half.", + "bbox": [ + 212, + 542, + 787, + 679 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Metrics We define a Context Score $(CTS)$ to evaluate the context relationships among neighboring city blocks (i.e., a community). This metric is a numerical extension of LayoutSim [20] and DocSim index [46] which evaluate the similarity between two city blocks regarding location, size, and category. Specifically, for each generated city block $b_{i}$ , we calculate the average LayoutSim with its neighbors $N(i)$ . Then, the local average of LayoutSim represents the context similarity among the community $(CT)$ . High $CT$ indicates identical blocks in a community and low $CT$ corresponds to high diversity in a community. $CTS$ is the subtraction of the real community's $CT$ from the generated $CT$ :", + "bbox": [ + 212, + 703, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation", + "bbox": [ + 225, + 113, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/d0fdcdfcd0b9e6fb1c0ccc86b21a788ad0ef9b20effa5675ae16eb2ebd8bfd5e.jpg", + "image_caption": [ + "Fig. 5: Qualitative Comparisons. Given the same road network (except SDXL [48]), all above methods generate urban layouts in only one pass without post-processing or human-in-the-loop refinement. The \"even rows\" are a zoom-in of the highlighted areas in the \"odd rows\". Our method generates a realistic distribution of urban layouts with plausible context-dependent behaviors (as indicated by $CTS$ score). VTN [2] and LayoutDM [29] show abnormal style shift among neighboring blocks, and no awareness of road networks. GlobalMapper [20] generates over-similar communities. SDXL [48] provides poor outpainting quality containing semantic errors and undesired overlaps." + ], + "image_footnote": [], + "bbox": [ + 220, + 157, + 782, + 696 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "He, L. & Aliaga, D.", + "bbox": [ + 271, + 114, + 403, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/d08205646244d3b2de1405d1fb8dd29dc90d3c3fe8a7b9191a1a58795f90fe8a.jpg", + "table_caption": [ + "Table 1: Quantitative Comparisons. All methods are compared to the same real urban layouts (except SDXL [48] which cannot take-in a road network). Best values are in bold, second best values are underlined. Our method outperforms other existing methods in all but the overlap metric. See text for an explanation of the metrics." + ], + "table_footnote": [], + "table_body": "
Method\\(CTS_{x\\rightarrow 0}\\)WD-5D↓WD-CO↓Overlap↓O-Blk↓FID↓KID↓LPIPS↓
SDXL [48]-----120.240.0790.48
VTN [2]-1.143.185.811.247.3569.140.0470.32
LayoutDM [29]-2.202.9212.504.561.7266.770.0400.39
GlobalMP [20]0.624.774.142.520.6849.550.0240.34
Ours0.212.281.911.270.4223.630.0050.20
", + "bbox": [ + 217, + 212, + 782, + 320 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nC T = \\frac {1}{\\| N (i) \\|} \\sum_ {j} ^ {N (i)} \\operatorname {L a y o u t S i m} \\left(b _ {i}, b _ {j}\\right), \\quad N (i) = \\{j \\mid (i, j) \\in E \\} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 276, + 361, + 785, + 405 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nC T S = C T _ {g e n} - C T _ {r e a l} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 419, + 785, + 436 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Given a real community, the metric indicates whether a generated community is over-diverse ( $CTS < 0$ ), or over-similar ( $CTS > 0$ ). A low value means more realism.", + "bbox": [ + 212, + 443, + 787, + 487 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In our comparisons, we compute several metrics. One metric is the Wasserstein Distance between the distributions of generated communities and real communities in terms of 2D position (x, y) and 2.5D geometry (length, width, height), and building count (CO) - we name these WD-5D and WD-CO. We define two quality metrics as well: Overlap indicates the percentage by which two building overlaps, and O-Blk represents the percentage of building area that sticks outside of its city block. We also implement several perceptual metrics: FID [24], KID [4], and LPIPS [65] to evaluate the visual quality and realism of generated layouts. For all the above metrics, lower value indicates better quality.", + "bbox": [ + 212, + 489, + 789, + 625 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative Results As Tab. 1 shows, our approach performs better than existing approaches in all but one metric. Note that the poor quality of SDXL results makes it hard to decipher layouts. We only compare to it by pixel-based perceptual metrics.", + "bbox": [ + 212, + 648, + 787, + 710 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Qualitative Results We show in Fig. 5 qualitative results for several exemplary areas (more examples are in Supplemental Material). In general, our method shows good contextual harmonization and has awareness of arbitrary road networks. It generates a realistic distribution of urban layouts with plausible context dependent behaviors. Other methods show improper patterns (e.g., overly diverse or similar) and potentially containing semantic errors or undesired overlaps.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation", + "bbox": [ + 225, + 113, + 732, + 130 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/3ab87a3cd1f94a9bfaf5fbc6e39c2ab166a7fbf2e8ac7e035afc0c890a3de6fb.jpg", + "table_caption": [ + "Table 2: Block Quantization Ablations. We report metrics (all in %) among all alternative ablations of our Block-level graph-based Variational AutoEncoder (BVAE). Using GAT-backbone with post-trained dimensional quantization with $L = 20$ shows the best overall compromise. The best values are in boldface." + ], + "table_footnote": [], + "table_body": "
Encode ModelOverlap↓Out-Blk↓Pos-E.↓Geom-E.↓Ct-E.↓Cov-E.↓
LayoutVAE [32]13.8411.1515.0937.35-24.34
VTN [2]2.704.706.0122.8618.581.97
BlockPlanner [62]2.931.304.7812.063.635.66
GlobalMapper [20]1.041.203.252.830.080.46
BVAE (SAGE [18])1.331.705.013.250.100.35
BVAE (GCN [33])0.451.214.252.740.040.28
BVAE* (GAT [58])0.170.281.000.710.0060.04
QuantizationOverlap↓Out-Blk↓Pos-E.↓Geom-E.↓Ct-E.↓Cov-E.↓
Trainable-VQ3.020.553.755.780.070.60
KMeans-Q2.900.0220.9817.8550.0116.23
DIM-Q (L=5)0.280.627.7116.2025.014.68
DIM-Q (L=10)0.560.285.734.488.334.58
DIM-Q (L=30)0.170.012.680.780.0050.15
DIM-Q* (L=20)0.110.013.500.890.0060.12
", + "bbox": [ + 246, + 212, + 754, + 470 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 501, + 401, + 515 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Block Quantization We ablate block quantization in order to understand and maximize the impact of changing its parameters and design strategy. First, we report in Tab. 2 (Encode Model Table) the reconstruction loss using different block-level variational autoencoders. We evaluate using various geometric metrics: Overlap (percent of building-to-building overlap), Out-Blk (percent of buildings that stick outside of its city block), Pos-E (positional error as a percentage of block diagonal), Geom-E (2.5D building geometry error as percentage of block scale), Ct-E (percentage of building count error), and Cov-E (total building coverage error). They clearly show our approach is the best overall. Thus we choose our BVAE based on Graph Attention Networks (GAT) [58].", + "bbox": [ + 212, + 535, + 787, + 686 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Then, in Tab. 2 (Quantization Table), we ablate various codebook approaches and quantitization levels for block-level encoding. \"Trainable-VQ\" leverages trainable vector quantization codebooks introduced by VQVAE [56]. We also provide an alternative non-training based quantization method using KMeans (\"KMeans-Q\") to cluster BVAE latent space values to a reasonable number of categories in order to handle diverse urban layouts (e.g., 100 classes). The \"DIM-Q\" rows imply quantizing each of the 512 dimensions of the block-level code into a codebook $C = \\{1,2,\\dots,L\\}$ as described in Sec. 3.1. This per-dimension quantization performs clearly the best. Specifically, setting $L$ to infinity produces a lossless representation of BVAE latent vectors, but also causes the curse of dimension-", + "bbox": [ + 212, + 688, + 789, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "He, L. & Aliaga, D.", + "bbox": [ + 271, + 113, + 403, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/7c55e30166b8702645ac7e59d7a92831436b89b3d80c1915cc66bff3582d2cc2.jpg", + "table_caption": [ + "Table 3: City-Scale GMAE Ablations. We ablate backbone message passing layers, maximum messaging passing hops ( $D$ : depth of encoder), masking strategy during training, and scheduling functions of iterative generation ( $T$ indicates total iterations). The overall best setting is marked as *." + ], + "table_footnote": [], + "table_body": "
AblationsContext↓WD-3D↓WD-N↓Overlap↓Out-Blk↓FID↓KID↓LPIPS↓
GMAE (GCN [33])-0.302.742.910.260.3652.950.0250.32
GMAE (GTrans [52])0.743.023.271.961.0743.600.0180.29
GMAE (SAGE [18])0.972.982.361.752.6043.150.0170.29
GMAE (GAT, Dec)0.942.452.962.131.1048.240.0190.30
GMAE (GAT, D = 1)1.253.032.641.730.4443.750.0180.29
GMAE (GAT, D = 2)0.332.662.011.310.5734.820.0140.27
GMAE (GAT, D = 4)0.241.922.310.630.2824.590.0090.22
GMAE (Mask ∈ [0.05, 1])1.302.703.621.702.4543.140.0180.29
GMAE (Mask ∈ [0.05, 0.5])1.103.073.681.681.3046.850.0200.30
GMAE (Mask = 0.15)1.594.913.411.831.2049.500.0220.31
GMAE (Linear, T = 12)0.632.833.791.270.9337.420.0150.28
GMAE (Log, T = 12)-1.312.777.277.413.1463.250.0380.33
GMAE (Cosine, T = 1)-1.677.445.643.564.0176.120.0510.35
GMAE (Cosine, T = 20)0.322.923.981.680.3132.060.0130.24
*(GAT-D3, [0.5,1], Cos-T12)0.212.281.911.270.4223.630.0050.20
", + "bbox": [ + 215, + 212, + 784, + 416 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "ality. We discover that $L = 20$ per dimension is the best overall compromise balancing compactness and representation ability.", + "bbox": [ + 212, + 446, + 785, + 477 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "City-Scale Graph Masked Autoencoder. In Tab. 3, We ablate various model parameters, together with significant training and inference strategies for our GMAE. Each alternatives is evaluated by the same set of metrics in Sec. 4.2.", + "bbox": [ + 212, + 503, + 785, + 549 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "By varying the message passing layers, we find GAT [58] provides the best performance. GCN [33] produces diverse and unrealistic urban layouts, while transformer-based [52], and GraphSAGE [18] backbones generate blocks that are too similar. The depth of the stacked layers in the encoder (D) dictates the maximum message passing hops from each node and thus the span of context behaviors captured for realistic generation. We find that 3 hops is promising and larger contexts cause over-similarity and unnecessary computing burden. We also observed that using stacked GAT layers for decoding with remasking tricks, as described by [25,26], didn't produce a net benefit towards city generation. So we choose MLP as the node feature decoder for both its performance and efficiency.", + "bbox": [ + 212, + 551, + 785, + 700 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The masking strategy during training is crucial for GMAE performance. We find that dynamic and large masking ratios are beneficial, while a fixed small ratio (e.g., 0.15 as in [14]) leads to over similarity. This behavior is also observed by Li et al. [35]. We adopt a similar masking strategy by sampling a Gaussian distribution $N(0.55, 0.25)$ , truncated by [0.5, 1.0], for each training iteration.", + "bbox": [ + 212, + 703, + 785, + 777 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Scheduling for iterative generation is also critical to capture representative city blocks (Sec. 3.3). We evaluate several functions $\\beta(t)$ (e.g., linear, logarithmic, or cosine-based) that provide the scheduled acceptance ratio, and evaluate alternative total iteration counts $T$ . Rapid iteration leads to over-diversity and", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation", + "bbox": [ + 225, + 113, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "long iteration results in over-similarity. Logarithmic functions with reverse pattern produces the worst realism. Metrics show that cosine-based function, which is conservative initially and generous in later iterations, outperforms others.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In summary, this study determines the best configuration for our GMAE as: GAT encoder with $D = 3$ and MLP decoder, training mask ratio $m \\in [0.5,1.0]$ , and priority-based scheduling function $\\beta (t) = 1 - \\cos (t / T)$ with $T = 12$ .", + "bbox": [ + 212, + 191, + 784, + 238 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.4 Limitations", + "text_level": 1, + "bbox": [ + 214, + 260, + 356, + 273 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our method faces limitations where shapes of city blocks or buildings are not simple polygons (e.g. the building with a garden inside). It also struggles to handle very irregular and concave shaped city blocks (e.g., cul-de-sac's). Also, road-network to contextual structure relations are not explicitly considered.", + "bbox": [ + 212, + 285, + 782, + 345 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.5 Additional Results", + "text_level": 1, + "bbox": [ + 214, + 368, + 415, + 382 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We place several other interesting results in Supplementary Materials.", + "bbox": [ + 214, + 393, + 718, + 409 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Controllable Generation of Entire City. Our method enables realistic and efficient city-scale controllable generation given any percentage of priors. We show several large examples spanning a subset of the many cities.", + "- Socio-Economic Prediction. Trained GMAE may act as a powerful encoder for downstream tasks. For example, the encoded GMAE latent vectors of urban layouts can indicate social-economic metrics for each city block.", + "- Semantic Manipulations. Our method enables semantic manipulations over various building layout styles across different cities." + ], + "bbox": [ + 225, + 421, + 782, + 539 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusions and Future Works", + "text_level": 1, + "bbox": [ + 214, + 566, + 549, + 580 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Large-scale urban layout generation has seen significant advancements, spurred by the intersection of computer vision, urban planning, and related disciplines. Previous methodologies often neglect the context-sensitive nature of urban layouts, failing to capture the multi-layer semantics crucial for realism and plausibility. Our proposed Graph-based Masked AutoEncoder (GMAE) addresses these limitations by leveraging a graph-based representation, acknowledging context-sensitivity, and prioritizing the generation process. By training on a diverse dataset of 330 cities, we demonstrate the efficacy of our approach in generating large-scale 2.5D layouts with realism and semantic consistency. Moreover, our method outperforms existing techniques in various metrics, marking a significant advancement in the realm of urban layout generation.", + "bbox": [ + 212, + 598, + 784, + 763 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Moving forward, our approach paves the way for more accurate and efficient urban simulation, digital twin creation, and game/content design. As future work we intend to incorporate our method for large-scale photo-realistic multi-view scene synthesis, and city-scale 3D modeling. Those potentials will contribute to synthetic data generation in autonomous driving and world model training.", + "bbox": [ + 212, + 765, + 784, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "He, L. & Aliaga, D.", + "bbox": [ + 271, + 114, + 403, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This project was funded in part by NSF Grant #2107096 and NSF Grant #1835739.", + "bbox": [ + 215, + 178, + 784, + 208 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 234, + 321, + 250 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Climate and economic justice screening tool. https://screeningtool.geoplatform.gov/", + "2. Arroyo, D.M., Postels, J., Tombari, F.: Variational transformer networks for layout generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13642-13652 (2021)", + "3. Bhatt, M., Kalyanam, R., Nishida, G., He, L., May, C., Niyogi, D., Aliaga, D.: Design and deployment of photo2building: A cloud-based procedural modeling tool as a service. In: Practice and Experience in Advanced Research Computing, pp. 132-138 (2020)", + "4. Binkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying mmd gans. arXiv preprint arXiv:1801.01401 (2018)", + "5. Bokeloh, M., Wand, M., Seidel, H.P.: A connection between partial symmetry and inverse procedural modeling. In: ACM SIGGRAPH 2010 papers, pp. 1-10 (2010)", + "6. Brooks, T., Peebles, B., Holmes, C., DePue, W., Guo, Y., Jing, L., Schnurr, D., Taylor, J., Luhman, T., Luhman, E., Ng, C., Wang, R., Ramesh, A.: Video generation models as world simulators (2024), https://openai.com/research/video-generation-models-as-world-simulators", + "7. Bureau, U.S.C.: Topologically integrated geographic encoding and referencing. https://www.census.gov/geographies/mapping-files/time-series/geo/tiger-line-file.html", + "8. Chai, L., Tucker, R., Li, Z., Isola, P., Snavely, N.: Persistent nature: A generative model of unbounded 3d worlds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20863-20874 (2023)", + "9. Chai, S., Zhuang, L., Yan, F.: Layoutdm: Transformer-based diffusion model for layout generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18349-18358 (2023)", + "0. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked generative image transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11315-11325 (2022)", + "1. Chang, K.H., Cheng, C.Y., Luo, J., Murata, S., Nourbakhsh, M., Tsuji, Y.: Building-gan: Graph-conditioned architectural volumetric design generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11956-11965 (2021)", + "2. Chen, Z., Wang, G., Liu, Z.: Scenedreamer: Unbounded 3d scene generation from 2d image collections. arXiv preprint arXiv:2302.01330 (2023)", + "3. Deng, J., Chai, W., Guo, J., Huang, Q., Hu, W., Hwang, J.N., Wang, G.: Citygen: Infinite and controllable 3d city layout generation. arXiv preprint arXiv:2312.01508 (2023)", + "4. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + ], + "bbox": [ + 225, + 267, + 784, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation", + "bbox": [ + 225, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "16. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873-12883 (2021)", + "17. Gupta, K., Lazarow, J., Achille, A., Davis, L.S., Mahadevan, V., Shrivastava, A.: Layouttransformer: Layout generation and completion with self-attention. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1004-1014 (2021)", + "18. Hamilton, W., Ying, Z., Leskovec, J.: Inductive representation learning on large graphs. Advances in neural information processing systems 30 (2017)", + "19. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)", + "20. He, L., Aliaga, D.: Globalmapper: Arbitrary-shaped urban layout generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 454-464 (2023)", + "21. He, L., Lu, Y., Corring, J., Florencio, D., Zhang, C.: Diffusion-based document layout generation. In: Fink, G.A., Jain, R., Kise, K., Zanibbi, R. (eds.) Document Analysis and Recognition - ICDAR 2023. pp. 361-378. Springer Nature Switzerland, Cham (2023)", + "22. He, L., Shan, J., Aliaga, D.: Generative building feature estimation from satellite images. IEEE Transactions on Geoscience and Remote Sensing 61, 1-13 (2023)", + "23. Heris, M.P., Foks, N.L., Bagstad, K.J., Troy, A., Ancona, Z.H.: A rasterized building footprint dataset for the united states. Scientific data 7(1), 207 (2020)", + "24. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)", + "25. Hou, Z., He, Y., Cen, Y., Liu, X., Dong, Y., Kharlamov, E., Tang, J.: Graphmae2: A decoding-enhanced masked self-supervised graph learner. In: Proceedings of the ACM Web Conference 2023. pp. 737-746 (2023)", + "26. Hou, Z., Liu, X., Cen, Y., Dong, Y., Yang, H., Wang, C., Tang, J.: Graphmae: Self-supervised masked graph autoencoders. In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 594-604 (2022)", + "27. Hua, H., Shi, J., Kafle, K., Jenni, S., Zhang, D., Collomosse, J., Cohen, S., Luo, J.: Finematch: Aspect-based fine-grained image and text mismatch detection and correction. arXiv preprint arXiv:2404.14715 (2024)", + "28. Hui, M., Zhang, Z., Zhang, X., Xie, W., Wang, Y., Lu, Y.: Unifying layout generation with a decoupled diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1942-1951 (2023)", + "29. Inoue, N., Kikuchi, K., Simo-Serra, E., Otani, M., Yamaguchi, K.: Layoutdm: Discrete diffusion model for controllable layout generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10167-10176 (2023)", + "30. Jiang, Z., Guo, J., Sun, S., Deng, H., Wu, Z., Mijovic, V., Yang, Z.J., Lou, J.G., Zhang, D.: Layoutformer++: Conditional graphic layout generation via constraint serialization and decoding space restriction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18403-18412 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "He, L. & Aliaga, D.", + "bbox": [ + 271, + 114, + 403, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Jyothi, A.A., Durand, T., He, J., Sigal, L., Mori, G.: Layoutvae: Stochastic scene layout generation from a label set. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9895-9904 (2019)", + "32. Jyothi, A.A., Durand, T., He, J., Sigal, L., Mori, G.: Layoutvae: Stochastic scene layout generation from a label set. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9894-9903 (2019). https://doi.org/10.1109/ICCV.2019.00999", + "33. Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907 (2016)", + "34. Li, J., Yang, J., Hertzmann, A., Zhang, J., Xu, T.: Layoutgan: Synthesizing graphic layouts with vector-wireframe adversarial networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(7), 2388-2399 (2020)", + "35. Li, T., Chang, H., Mishra, S., Zhang, H., Katabi, D., Krishnan, D.: Mage: Masked generative encoder to unify representation learning and image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2142-2152 (2023)", + "36. Li, Z., Wang, Q., Snavely, N., Kanazawa, A.: Infinitenature-zero: Learning perpetual view generation of natural scenes from single images. In: European Conference on Computer Vision. pp. 515-534. Springer (2022)", + "37. Lin, C.H., Lee, H.Y., Cheng, Y.C., Tulyakov, S., Yang, M.H.: Infinitygan: Towards infinite-pixel image synthesis. arXiv preprint arXiv:2104.03963 (2021)", + "38. Lin, C.H., Lee, H.Y., Menapace, W., Chai, M., Siarohin, A., Yang, M.H., Tulyakov, S.: Infinity: Infinite-scale city synthesis. arXiv preprint arXiv:2301.09637 (2023)", + "39. Lipp, M., Scherzer, D., Wonka, P., Wimmer, M.: Interactive modeling of city layouts using layers of procedural content. In: Computer Graphics Forum. vol. 30, pp. 345-354. Wiley Online Library (2011)", + "40. Ma, H., Zeng, D., Liu, Y.: Learning individualized treatment rules with many treatments: A supervised clustering approach using adaptive fusion. Advances in Neural Information Processing Systems 35, 15956-15969 (2022)", + "41. Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781 (2013)", + "42. Nauata, N., Chang, K.H., Cheng, C.Y., Mori, G., Furukawa, Y.: House-gan: Relational generative adversarial networks for graph-constrained house layout generation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 162-177. Springer (2020)", + "43. OpenStreetMap contributors: Planet dump retrieved from https://planet.osm.org. https://www.openstreetmap.org (2017)", + "44. Para, W., Guerrero, P., Kelly, T., Guibas, L.J., Wonka, P.: Generative layout modeling using constraint graphs. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6690-6700 (2021)", + "45. Patel, P., Kalyanam, R., He, L., Aliaga, D., Niyogi, D.: Deep learning-based urban morphology for city-scale environmental modeling. PNAS nexus 2(3), pgad027 (2023)", + "46. Patil, A.G., Ben-Eliezer, O., Perel, O., Averbuch-Elor, H.: Read: Recursive autoencoders for document layout generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 544-545 (2020)", + "47. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4195-4205 (2023)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation", + "bbox": [ + 225, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)", + "49. Shen, Y., Ma, W.C., Wang, S.: Sgam: Building a virtual 3d world through simultaneous generation and mapping. Advances in Neural Information Processing Systems 35, 22090-22102 (2022)", + "50. Sheng, Y., Liu, Y., Zhang, J., Yin, W., Oztireli, A.C., Zhang, H., Lin, Z., Shechtman, E., Benes, B.: Controllable shadow generation using pixel height maps. In: European Conference on Computer Vision. pp. 240-256. Springer (2022)", + "51. Sheng, Y., Yu, Z., Ling, L., Cao, Z., Zhang, X., Lu, X., Xian, K., Lin, H., Benes, B.: Dr. bokeh: Differentiable occlusion-aware bokeh rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4515-4525 (2024)", + "52. Shi, Y., Huang, Z., Feng, S., Zhong, H., Wang, W., Sun, Y.: Masked label prediction: Unified message passing model for semi-supervised classification. arXiv preprint arXiv:2009.03509 (2020)", + "53. Song, Y., Zhang, Z., Lin, Z., Cohen, S., Price, B., Zhang, J., Kim, S.Y., Aliaga, D.: Objectstitch: Object compositing with diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18310-18319 (2023)", + "54. Song, Y., Zhang, Z., Lin, Z., Cohen, S., Price, B., Zhang, J., Kim, S.Y., Zhang, H., Xiong, W., Aliaga, D.: Imprint: Generative object compositing by learning identity-preserving representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8048-8058 (2024)", + "55. Tabata, S., Yoshihara, H., Maeda, H., Yokoyama, K.: Automatic layout generation for graphical design magazines. In: ACM SIGGRAPH 2019 Posters, pp. 1-2 (2019)", + "56. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017)", + "57. Vanegas, C.A., Kelly, T., Weber, B., Halatsch, J., Aliaga, D.G., Müller, P.: Procedural generation of parcels in urban modeling. In: Computer graphics forum. vol. 31, pp. 681-690. Wiley Online Library (2012)", + "58. Velicković, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y.: Graph attention networks. arXiv preprint arXiv:1710.10903 (2017)", + "59. Velickovic, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y., et al.: Graph attention networks. stat 1050(20), 10-48550 (2017)", + "60. Wu, W., Fu, X.M., Tang, R., Wang, Y., Qi, Y.H., Liu, L.: Data-driven interior plan generation for residential buildings. ACM Transactions on Graphics (TOG) 38(6), 1-12 (2019)", + "61. Xie, H., Chen, Z., Hong, F., Liu, Z.: Citydreamer: Compositional generative model of unbounded 3d cities. arXiv preprint arXiv:2309.00610 (2023)", + "62. Xu, L., Xiangli, Y., Rao, A., Zhao, N., Dai, B., Liu, Z., Lin, D.: Blockplanner: City block generation with vectorized graph representation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5077-5086 (2021)", + "63. Yan, W., Zhang, Y., Abbeel, P., Srinivas, A.: Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157 (2021)", + "64. Yang, C.F., Fan, W.C., Yang, F.E., Wang, Y.C.F.: Layouttransformer: Scene layout generation with conceptual and spatial diversity. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3732-3741 (2021)", + "65. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "He, L. & Aliaga, D.", + "bbox": [ + 271, + 114, + 403, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "66. Zhang, X., Ma, W., Varinlioglu, G., Rauh, N., He, L., Aliaga, D.: Guided pluralistic building contour completion. The Visual Computer 38(9), 3205-3216 (2022)", + "67. Zheng, X., Qiao, X., Cao, Y., Lau, R.W.: Content-aware generative modeling of graphic design layouts. ACM Transactions on Graphics (TOG) 38(4), 1-15 (2019)" + ], + "bbox": [ + 215, + 146, + 784, + 204 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation", + "bbox": [ + 225, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/COHO_ Context-Sensitive City-Scale Hierarchical Urban Layout Generation/ad29cc50-ea60-4c58-8c17-9b4b4ddc225f_model.json b/2024/COHO_ Context-Sensitive City-Scale Hierarchical Urban Layout Generation/ad29cc50-ea60-4c58-8c17-9b4b4ddc225f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1d5f52fdbf9ded98ea4bcd2e8cc03fba4d340c1e --- /dev/null +++ b/2024/COHO_ Context-Sensitive City-Scale Hierarchical Urban Layout Generation/ad29cc50-ea60-4c58-8c17-9b4b4ddc225f_model.json @@ -0,0 +1,2570 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.22, + 0.141, + 0.784, + 0.187 + ], + "angle": 0, + "content": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation" + }, + { + "type": "text", + "bbox": [ + 0.41, + 0.213, + 0.596, + 0.229 + ], + "angle": 0, + "content": "Liu He, Daniel Aliaga" + }, + { + "type": "text", + "bbox": [ + 0.419, + 0.24, + 0.584, + 0.255 + ], + "angle": 0, + "content": "Purdue University, USA" + }, + { + "type": "text", + "bbox": [ + 0.405, + 0.255, + 0.601, + 0.269 + ], + "angle": 0, + "content": "{he425,aliaga}@purdue.edu" + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.302, + 0.422, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.282, + 0.418, + 0.41, + 0.429 + ], + "angle": 0, + "content": "LayoutDM \\((CTS = -2.20)\\)" + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.303, + 0.583, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.461, + 0.419, + 0.551, + 0.429 + ], + "angle": 0, + "content": "Ours \\((CTS = 0.21)\\)" + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.304, + 0.743, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.6, + 0.419, + 0.737, + 0.43 + ], + "angle": 0, + "content": "GlobalMapper \\((CTS = 0.62)\\)" + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.431, + 0.422, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.547, + 0.394, + 0.557 + ], + "angle": 0, + "content": "Random \\((CTS\\ll 0)\\)" + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.431, + 0.583, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.461, + 0.547, + 0.562, + 0.558 + ], + "angle": 0, + "content": "Real Data \\((CTS\\rightarrow 0)\\)" + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.432, + 0.743, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.547, + 0.716, + 0.558 + ], + "angle": 0, + "content": "Identical \\((CTS\\gg 0)\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.639 + ], + "angle": 0, + "content": "Fig. 1: Context-Sensitive Generation. Our method (Green) pursues realistic context harmonization among neighboring city blocks as real data (bottom-middle). Other methods (e.g. LayoutDM [29], GlobalMapper [20]) show over-diversity/-similarity in city-scale layout generation (evaluated by Context Score \\(CTS\\) as Eq. 4.). Fully random and identical layouts are synthetically generated to illustrate extreme cases." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.688, + 0.744, + 0.841 + ], + "angle": 0, + "content": "Abstract. The generation of large-scale urban layouts has garnered substantial interest across various disciplines. Prior methods have utilized procedural generation requiring manual rule coding or deep learning needing abundant data. However, prior approaches have not considered the context-sensitive nature of urban layout generation. Our approach addresses this gap by leveraging a canonical graph representation for the entire city, which facilitates scalability and captures the multilayer semantics inherent in urban layouts. We introduce a novel graph-based masked autoencoder (GMAE) for city-scale urban layout generation. The method encodes attributed buildings, city blocks, communities and cities into a unified graph structure, enabling self-supervised" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.129 + ], + "angle": 0, + "content": "He, L. & Aliaga, D." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.148, + 0.744, + 0.232 + ], + "angle": 0, + "content": "masked training for graph autoencoder. Additionally, we employ scheduled iterative sampling for 2.5D layout generation, prioritizing the generation of important city blocks and buildings. Our approach achieves good realism, semantic consistency, and correctness across the heterogeneous urban styles in 330 US cities. Codes and datasets are released at https://github.com/Arking1995/COHO." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.245, + 0.741, + 0.273 + ], + "angle": 0, + "content": "Keywords: Layout Generation \\(\\cdot\\) Urban Modeling \\(\\cdot\\) Masked Graph Autoencoder \\(\\cdot\\) Context Sensitivity" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.299, + 0.376, + 0.315 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.331, + 0.787, + 0.437 + ], + "angle": 0, + "content": "Large-scale urban layout generation has received significant interest by computer vision, urban planning and related disciplines. In particular, cities easily have 1,000 to 50,000 city blocks spanning a mostly horizontal terrain. The configuration of each city block varies significantly across the city, yet there is some configuration similarity amongst subsets within and among cities. Reconstructing existing layouts and generating future planned layouts are crucial tasks for urban simulation, digital twins, and game/content design." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.788, + 0.664 + ], + "angle": 0, + "content": "Prior works have generated urban layouts from a variety of sources. Procedural layout generation [39, 57] used hand-coded rules to generate cities, and inverse procedural modeling [5] extracted the rules from data and then enabled generation. More recently, deep learning has promoted data-driven approaches that produce either pixel-based layout generation (e.g., InfiniCity [38], CityDreamer [61], CityGEN [13]) or graph-based approaches for general polygonal shape layouts (e.g., LayoutTransformer [17], VTN [2]) and for urban scenarios (e.g., BlockPlanner [62], GlobalMapper [20], BuildingGAN [11]). However, none of these techniques explicitly recover and consider the context-sensitive nature of urban layout generation (Fig. 1). In other words, the layout of buildings, blocks, and communities cannot be treated in isolation. The multi-layer semantics are critical to generating realistic and plausible urban layouts. While such multilevel structure (e.g., from single building to a community or subset of a city) has been specified via hand-coded rules, it has not been discovered, organized and used for generation in data-driven deep learning approaches." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.664, + 0.584, + 0.679 + ], + "angle": 0, + "content": "Our approach builds on three key observations:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.689, + 0.787, + 0.795 + ], + "angle": 0, + "content": "1. Representation: Cities, communities, blocks, buildings, and roads are arbitrarily shaped and with complex topology. Hence, a graph-based representation can better capture the regular and non-regular configurations, as opposed to the limiting regular structure of an image. Further, graphs can be more compact which in turn supports better scalability. Altogether, having buildings or city blocks as atomic units, instead of pixels, will contribute to improving correctness, realism, and consistency with priors." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "2. Context-Sensitivity: Buildings, city blocks, and communities are built considering their neighboring structures and not in isolation. Hence, a generator should consider this multi-layer structure and inter-dependency. In" + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.689, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.226, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "order words, the style features at one level (e.g., size, shape, location) depend on the surrounding context." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.178, + 0.785, + 0.266 + ], + "angle": 0, + "content": "3. Prioritization: The stylistic and semantic importance, or priority, of city blocks and buildings varies. This implies that a generation scheme should generate the most important city blocks and buildings first and then fill in the rest, as opposed to a region growing approach for example. Overall, prioritization will improve correctness and realism as well as enable large-scale urban layout generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.275, + 0.784, + 0.395 + ], + "angle": 0, + "content": "We propose a novel Graph-based Masked AutoEncoder (GMAE) for large-scale multi-layer 2.5D urban layout generation. First, we define a canonical graph-based representation capturing the multi-layer context sensitivity of any city's urban layouts (Fig. 2). The entire graph \\( G \\) corresponds to a city. Subsets of the graph map to communities and each node \\( b \\) amounts to a city block. Further, each node encodes its building layouts, shapes, and heights as a quantized feature by well-trained quantizor. Second, we utilize our GMAE to learn the multi-layer stylistic behavior of many urban layouts (i.e., 330 cities)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.785, + 0.516 + ], + "angle": 0, + "content": "Using subsets of the graph (or communities), node features are randomly masked and self-supervised GMAE training is performed to learn node feature reconstruction (Fig. 3). Third, utilizing the well-trained GMAE as a generator, we perform an iterative priority-based scheduling to generate a large layout or to complete a city fragment; i.e., each iteration predicts node features for the entire graph but only the most confident ones are preserved for the next step (Fig. 4). This methodology is able to leverage any percentage of given prior (e.g., \\([0,100\\%]\\)) for urban layout completion and generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.517, + 0.785, + 0.636 + ], + "angle": 0, + "content": "To the best of our knowledge, our approach is the first data-driven deep-learning method to enable large-scale 2.5D layout generation with good realism, semantic consistency, and correctness. We show the capability to generalize to the heterogeneous urban layout styles of 330 cities in the US (e.g., all cities with a population \\(>100\\mathrm{K}\\), totaling 833,473 city blocks and 17,663,607 buildings). In addition, we provide comparisons to several prior works including SDXL [48]), VTN [2], LayoutDM [29], and GlobalMapper [20] in Sec. 4 and exceed their performance in multiple well-known geometric and perceptual metrics." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.638, + 0.531, + 0.653 + ], + "angle": 0, + "content": "Our contributions include the following:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.66, + 0.783, + 0.702 + ], + "angle": 0, + "content": "- A canonical graph representation for large-scale 2.5D urban layouts. It encodes global and local node features and neighboring relations for buildings, city blocks, and communities within a city." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.706, + 0.783, + 0.745 + ], + "angle": 0, + "content": "- A self-supervised Graph-based Masked AutoEncoder (GMAE) enabling arbitrarily shaped city layout generation with consistency, correctness, and realism" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.75, + 0.783, + 0.792 + ], + "angle": 0, + "content": "- Priority-based scheduled iterative sampling for controllable urban layout synthesis including completion, refinement, and generation with any percentage of priors ([0, 100%])." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.796, + 0.783, + 0.838 + ], + "angle": 0, + "content": "- A open dataset of road network and urban layouts of 330 US cities for benchmarking current and future urban layout generation methods (https://huggingface.co/datasets/Arking95/COHO)." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.66, + 0.783, + 0.838 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.129 + ], + "angle": 0, + "content": "He, L. & Aliaga, D." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.388, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.436 + ], + "angle": 0, + "content": "Layout Generation. Numerous data-driven deep learning methods have been proposed for layout generation; for example, generalized layout synthesis [2,17, 31,34,64], document layouts [21,30,46,55,67], urban land lots [22,62,66], indoor scenes [42,44,60], and 3D buildings [3,11,45]. Recently, diffusion models have also been used for layout generation tasks (e.g., [9,21,28,29]). However, most of these works assume 2D layouts generated on a rectangular empty canvas. The position and size of each bounding box is often limited to a finite set of categories and to rectilinear alignments. This results in poor diversity and low realism for urban layout generation tasks. BlockPlanner [62] and GlobalMapper [20] adapt layout generation to either rectilinear or arbitrarily-shaped city block generation. However, all previous works assume each layout, or city block, is independent. This is not the case in urban environments where adjacent and nearby city blocks may share a common style or be positioned in a purposeful multi-block arrangement. Without a contextual guidance, city block generation may result in either unwanted harmonization or unexpected diversity between neighboring blocks (see Fig. 1). Thus, we pursue city-scale urban layout generation with the help of context-aware neighboring information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.788, + 0.678 + ], + "angle": 0, + "content": "Infinite Visual Synthesis. With bombing of visual deep learning works [27, 40, 50, 51, 53, 54], researchers have explored synthesizing large, potentially unbounded images. For example, omnidirectional image synthesis from multiple smaller field-of-view images [36, 37]. Other works focus on unbounded 3D scene generation [8, 12, 49]. Particularly relevant are several recent infinite urban scene generators [13, 38, 61] that directly adapt image synthesis methods (e.g. InfinityGAN [37], MaskGIT [10]). The synthesis performs an autoregressive outpainting process using layout image patches. However, patch-based training and inference doesn't learn global features nor their interdependencies. Rather, it learns a very local behavior. Prior pixel-based methods acknowledge limited realism and semantics [13, 38, 61]. Often, post-processing [13, 61], and human-in-the-loop editing [38] is required to refine the generated images to vector-based urban layouts. Even the latest stable diffusion model SDXL [48] seeks high-resolution image synthesis but may not produce realistic urban layouts (Fig. 5)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Learning based on Quantized Representations. Vector quantization has been a prevailing stepping-stone technique for text (e.g., Word2Vec [41]), image (e.g., VQVAE/GAN [16, 56], [15]), and video synthesis (e.g., SORA [6, 47], [63]). Pretrained quantizers provide a scalable token representation benefiting high-level and large-scale synthesis. A well-trained quantizer combined with self-supervised masked training often performs well in reconstructing the original signal during generative tasks. BERT [14] provides plausible text-based representation learning. MAE [19] defines a masked autoencoder for image space which excels at multiple downstream tasks. MAE also provides representation" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.226, + 0.114, + 0.731, + 0.129 + ], + "angle": 0, + "content": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.207 + ], + "angle": 0, + "content": "learning for a broad range of node/graph feature synthesis [25, 26]. Specifically, MaskGIT [10] and MAGE [35] utilize MAE frameworks for impressive image generation tasks. The idea of a generative masked autoencoder based on pretrained quantization was inspirational to our city-scale urban layout approach." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.231, + 0.331, + 0.248 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.265, + 0.785, + 0.326 + ], + "angle": 0, + "content": "First we describe our canonical graph representation of 2.5D urban layouts (Fig. 2). Second, we explain the training process for our GMAE based on the graphs for many cities (Fig. 3). Third, we show our priority-based scheduling for iteratively generating city-scale urban layouts given any amount of prior (Fig. 4)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.348, + 0.533, + 0.364 + ], + "angle": 0, + "content": "3.1 Canonical Graph Representation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.785, + 0.45 + ], + "angle": 0, + "content": "A city is represented by a graph \\( G = \\{B, E\\} \\), where \\( B \\) is a set of nodes \\( B = \\{b_i | i \\in [1, N]\\} \\) such that \\( b_i \\) represents each city block (Fig. 2), and any connected subgraph of \\( G \\) corresponds to a community. \\( N = |B| \\) is the total number of blocks for a given city. \\( E \\) is a set of edges \\( E = \\{e_{ij} | i, j \\in [1, N]\\} \\) representing mutually adjacent nodes (or city blocks). \\( K = |E| \\) is the total number of edges." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.785, + 0.518 + ], + "angle": 0, + "content": "Node (or City Block). Each node, or city block, \\( b_{i} \\) is a region defined by an enclosing road network and contains a set of block-level features and building-level features. The per-node features include:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.532, + 0.784, + 0.59 + ], + "angle": 0, + "content": "- \\( s_i \\) is the shape and location feature vector of a city block contour. It includes four features: (1) aspect ratio, (2) total block area, (3) ratio of total block area over the block's convex-hull area, and (4) relative distance from current block to the centroid of the city." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.592, + 0.784, + 0.62 + ], + "angle": 0, + "content": "- \\( q_{i} \\) is a 512-dimensional vector from a codebook \\( C \\). It hierarchically captures the buildings layouts and their configuration within \\( b_{i} \\)." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.532, + 0.784, + 0.62 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.784, + 0.667 + ], + "angle": 0, + "content": "We concatenate these node features to form a 516-dimensional vector \\(\\left([s_i, q_i]\\right)\\), which is able to represent our large and heterogeneous set of blocks and buildings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Building Layout Quantization. The 512-dimensional vector \\( q_{i} \\) describing the building layouts within a city block is obtained from a quantized codebook \\( C = \\{1,2,\\dots,L\\} \\) with quantization level \\( L = |C| \\). To create the codebook, we train a block-level variational autoencoder (BVAE) for building layout within a city block, using a canonical graph structure for representing the multiple building shapes, positions, and heights inside a single city block. Using this pretrained encoder, the latent vector of all possible building layouts will form a numerical distribution for each latent dimension. We further quantize each of these distributions into \\( L \\) bins with equal percentiles. Then, each latent value is replaced by the rank of the bin containing its value. Therefore, any building" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.129 + ], + "angle": 0, + "content": "He, L. & Aliaga, D." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.144, + 0.761, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.324, + 0.788, + 0.395 + ], + "angle": 0, + "content": "Fig. 2: Canonical Graph Representation. Our method represents a city as a canonical graph \\( G \\). Each node \\( b_{i} \\) represent a single city block, and each edge \\( e_{ij} \\) connects spatially adjacent blocks. Each block/node corresponds to a set of node features \\( s_i \\) and a quantized vector \\( q_{i} \\) hierarchically capturing enclosed building layouts. Edge feature \\( d_{ij} \\) encodes distances between block centroids. Graph \\( G \\) is used for GMAE training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.545 + ], + "angle": 0, + "content": "layout within a city block is quantized into a 512-dimensional categorical index vector \\( q_{i} = [c_{1},\\dots,c_{512}] \\), where each \\( c\\in C \\). Note that our codebook is defined after BVAE training and not dynamically trainable as in VQGAN/VQVAE [16,56]. After performing a set of detailed evaluations using different training schemes, encoding models, and quantization levels \\( L \\) (see Sec. 4.3), we selected the Graph Attention Network (GAT) [59] as the backbone for the BVAE. Further, we select \\( L = 20 \\) to balance compactness and representation ability after quantization. BVAE and quantization details are found in Supplementary Materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.787, + 0.612 + ], + "angle": 0, + "content": "Edge. Each edge \\( e_{ij} \\) contains a feature \\( d_{ij} \\) storing the relative distance between the centroids of the adjacent city blocks \\( b_i \\) and \\( b_j \\). This distance-weighted adjacency helps encode the relationship between neighboring blocks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.787, + 0.658 + ], + "angle": 0, + "content": "Altogether, the shape, style, and multi-layer urban layout of a city with \\(N\\) nodes and \\(K\\) edges is represented by a total of \\((516N + K)\\) variables. This canonical graph is used for the self-supervised GMAE training described below." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.68, + 0.553, + 0.695 + ], + "angle": 0, + "content": "3.2 Graph-based Masked AutoEncoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.779 + ], + "angle": 0, + "content": "We train our GMAE using a self-supervised process. A road-only city (or community) is a graph (or subgraph) \\(G\\) where building layout features \\(q_{i}\\) of every node (block) are missing, but \\(s_i\\) and \\(e_{ij}\\) are preserved. A context-sensitive node feature generation process corresponds to recreating the purposefully removed building layout features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "During training (Fig. 3), we use dynamic masking ratios of \\( q_{i} \\) in order to learn how to regenerate the original features. In each training iteration, the mask ratio \\( m \\) is sampled from a truncated Gaussian distribution as recommended by Li et al. [35]. The ratio is obtained from the range \\( m \\in [0.5,1.0] \\) and is centered at" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.226, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.261, + 0.788, + 0.332 + ], + "angle": 0, + "content": "Fig. 3: Graph-based Masked Autoencoder. Given a canonical city graph, quantized building layout features \\( Q \\) are masked with dynamic masking ratios \\( m \\in [0.5,1.0] \\), while block shape and location features \\( S \\) are kept. The GNN encoder uses message passing between neighboring nodes to obtain the context-aware node features \\( F \\). The decoder uses \\( F \\) to reconstruct \\( Q' \\). The predicted \\( Q' \\) are decoded to 2.5D urban layouts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.361, + 0.788, + 0.422 + ], + "angle": 0, + "content": "0.55. We denote building layout features as \\( Q = \\{q_{i} | i \\in [1, N]\\} \\) and those after masking as \\( Q^{mask} \\). Block features are denoted as \\( S = \\{s_{i} | i \\in [1, N]\\} \\). The graph neural network (GNN) encoder is denoted as \\( f_{E} \\), and node feature decoder as \\( f_{D} \\). Thus the process of our GMAE can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.366, + 0.437, + 0.786, + 0.457 + ], + "angle": 0, + "content": "\\[\nF = f _ {E} ([ Q ^ {m a s k}, S ], E), \\quad Q ^ {\\prime} = f _ {D} (F) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.508 + ], + "angle": 0, + "content": "where \\( F = \\{f_i | i \\in [1, N]\\} \\) indicates the context-sensitive node features obtained by message passing between neighboring nodes by \\( f_E \\). Extracted \\( F \\) is processed by \\( f_D \\) to reconstruct building layout features \\( Q' \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.643 + ], + "angle": 0, + "content": "The backbones of \\( f_{E} \\) and \\( f_{D} \\) can be any GNN, including GAT [58], GCN [33], and GraphSAGE [18]. After experimentation, we select GAT as \\( f_{E} \\) because of its best context-sensitive feature extraction ability, and we choose a straightforward MLP as \\( f_{D} \\) for efficient processing. We observed that the re-masking trick and complex GNN decoder utilized by [25,26] requires more memory and calculation but has no net benefit in our application. Since \\( Q \\) contains quantized index vectors, we use cross entropy loss for self-supervised reconstruction training on masked nodes. Detailed ablations of masking strategy and model selections are provided in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Our objective function is as follows, where \\( L \\) indicates the length of the quantized codebook \\( C \\):" + }, + { + "type": "equation", + "bbox": [ + 0.39, + 0.687, + 0.787, + 0.728 + ], + "angle": 0, + "content": "\\[\nL _ {r e c o n} = - \\sum_ {i = 1} ^ {L} \\left[ Q _ {i} \\log \\left(Q _ {i} ^ {\\prime}\\right) \\right] ^ {m a s k} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.843 + ], + "angle": 0, + "content": "To capture community behavior (and also reduce GPU memory usage), we randomly sample a batch of subgraphs from \\( G \\) for each training iteration. The nodes in each subgraph are sampled using a chosen radius from a random center node. While community sizes vary significantly, we found that a practical compromise is to use 500 meters as the radius value. This radius ensures to cover a reasonable number of the city blocks with a typical census tract from governmental CJEST dataset [1]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.129 + ], + "angle": 0, + "content": "He, L. & Aliaga, D." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.788, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.27, + 0.788, + 0.327 + ], + "angle": 0, + "content": "Fig. 4: Priority-based Scheduled Generation. We iteratively utilize pretrained GMAE to reconstruct masked node features. In each iteration, we accept a certain ratio of predicted nodes decided by the scheduling function \\(\\beta(t) = 1 - \\cos(t / T)\\). We obtain a full graph after \\(T\\) iterations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.358, + 0.566, + 0.373 + ], + "angle": 0, + "content": "3.3 Priority-based Scheduled Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.787, + 0.477 + ], + "angle": 0, + "content": "While in theory the trained GMAE can predict all nodes in a single pass, the generation quality is much better when performed iteratively. Recent published papers (e.g., [10,35]) have found that gradually sampling with a reasonable number of iterations improves generation quality. In urban scenarios, small number of representative city blocks are important to indicate context behavior. We intend to generate those blocks at early iterations and then span neighboring blocks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.788, + 0.675 + ], + "angle": 0, + "content": "In Fig. 4, we propose a \\(T\\)-iteration sampling strategy similar to MaskGIT [10] and MAGE [35]. At each iteration, GMAE predicts per-node building layout features for all remaining nodes. Then we rank generated nodes by their prediction confidence. The generated nodes with the highest prediction confidence are accepted. The accepted nodes after the current iteration \\(t \\in [1,T]\\) are scheduled by a cosine function \\(\\beta(t) = 1 - \\cos(t / T)\\). When \\(t\\) is small, the ratio is near to 0 causing few nodes to be accepted in initial iterations. As iterations continue, \\(t\\) gets bigger and hence the number of accepted nodes during each iteration will also increase. This behavior follows the prioritization observation in Sec. 1 which causes distinguished, or unique, city blocks to be generated first and essentially guide the layout style for nearby city blocks. Typically we set \\(T = 12\\) which results in dozens of nodes being generated as is the typical size of a community. Detailed ablations of alternative sampling schedules are provided in Sec. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.701, + 0.376, + 0.718 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.736, + 0.457, + 0.751 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Open Dataset We have created a vector-based urban layout dataset of all US cities with a population of 100K or more. We believe this dataset will be of significant use to the community to further develop urban layout generation methods and thus we will release the dataset together with the paper. For each city, we define a rectangular bounding box containing most of the metropolitan area. Then," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.226, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "we use this bounding box to extract data from OpenStreetMap (OSM) [43], Microsoft Building Footprints (MSF) [23], and Topologically Integrated Geographic Encoding and Referencing (TIGER) dataset [7]. The data we collected includes: (1) road network vectors and attributes, (2) building footprints and heights, and (3) block-level social economic metrics. Specifically, we collect road networks and block contours from TIGER. Then we look for building footprints and heights from OSM and MSF within each city block contour. We choose the one with more and accurate details (e.g., containing more buildings, having features that match those in TIGER). Over all 330 cities, this dataset has a total of 833,473 city blocks, and 17,663,607 buildings. We set training, validation, testing split as \\(70\\%\\), \\(20\\%\\), and \\(10\\%\\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.338, + 0.788, + 0.444 + ], + "angle": 0, + "content": "Training Details The training has two stages. We first use all city blocks for self-supervised training of a graph-based Block-level Variational AutoEncoder (BVAE). Given this trained block-level encoder, we perform the latent quantization for all building layouts in each city block. Then our GMAE is trained based on canonical city graph representation. Training time for BVAE and GMAE typically takes 12 hours and 15 hour respectively by a single A5000 GPU. We will release our codes upon acceptance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.787, + 0.491 + ], + "angle": 0, + "content": "Inferencing using our graph-based approach is seemingly instantaneous for all the results shown in this paper. This is contrast to the VTN [2] and SDXL [48] approaches which take significantly more time to infer and to train." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.516, + 0.371, + 0.531 + ], + "angle": 0, + "content": "4.2 Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.788, + 0.68 + ], + "angle": 0, + "content": "We compare our method to several alternate city-scale layout generators (Tab. 1). Specifically, we randomly select 100 communities from among our test set, total 3700 city blocks. All methods generate the layout of the 100 communities in one pass, and without any post-processing or human-in-the-loop refinement. Implementations are downloaded from the author's repositories and trained with default parameter settings. For fairness, we convert our dataset to token sequences, graphs, or image patches \\((1024\\times 1024)\\) as needed. VTN [2], LayoutDM [29], GlobalMapper [20] are retrained for single block generation. SDXL [48] is fine-tuned for outpainting by using the upper half to predict the lower half." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Evaluation Metrics We define a Context Score \\((CTS)\\) to evaluate the context relationships among neighboring city blocks (i.e., a community). This metric is a numerical extension of LayoutSim [20] and DocSim index [46] which evaluate the similarity between two city blocks regarding location, size, and category. Specifically, for each generated city block \\(b_{i}\\), we calculate the average LayoutSim with its neighbors \\(N(i)\\). Then, the local average of LayoutSim represents the context similarity among the community \\((CT)\\). High \\(CT\\) indicates identical blocks in a community and low \\(CT\\) corresponds to high diversity in a community. \\(CTS\\) is the subtraction of the real community's \\(CT\\) from the generated \\(CT\\):" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.129 + ], + "angle": 0, + "content": "He, L. & Aliaga, D." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.159, + 0.784, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.709, + 0.788, + 0.822 + ], + "angle": 0, + "content": "Fig. 5: Qualitative Comparisons. Given the same road network (except SDXL [48]), all above methods generate urban layouts in only one pass without post-processing or human-in-the-loop refinement. The \"even rows\" are a zoom-in of the highlighted areas in the \"odd rows\". Our method generates a realistic distribution of urban layouts with plausible context-dependent behaviors (as indicated by \\(CTS\\) score). VTN [2] and LayoutDM [29] show abnormal style shift among neighboring blocks, and no awareness of road networks. GlobalMapper [20] generates over-similar communities. SDXL [48] provides poor outpainting quality containing semantic errors and undesired overlaps." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.226, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 1: Quantitative Comparisons. All methods are compared to the same real urban layouts (except SDXL [48] which cannot take-in a road network). Best values are in bold, second best values are underlined. Our method outperforms other existing methods in all but the overlap metric. See text for an explanation of the metrics." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.213, + 0.784, + 0.321 + ], + "angle": 0, + "content": "
Method\\(CTS_{x\\rightarrow 0}\\)WD-5D↓WD-CO↓Overlap↓O-Blk↓FID↓KID↓LPIPS↓
SDXL [48]-----120.240.0790.48
VTN [2]-1.143.185.811.247.3569.140.0470.32
LayoutDM [29]-2.202.9212.504.561.7266.770.0400.39
GlobalMP [20]0.624.774.142.520.6849.550.0240.34
Ours0.212.281.911.270.4223.630.0050.20
" + }, + { + "type": "equation", + "bbox": [ + 0.277, + 0.362, + 0.786, + 0.406 + ], + "angle": 0, + "content": "\\[\nC T = \\frac {1}{\\| N (i) \\|} \\sum_ {j} ^ {N (i)} \\operatorname {L a y o u t S i m} \\left(b _ {i}, b _ {j}\\right), \\quad N (i) = \\{j \\mid (i, j) \\in E \\} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.42, + 0.786, + 0.438 + ], + "angle": 0, + "content": "\\[\nC T S = C T _ {g e n} - C T _ {r e a l} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.789, + 0.488 + ], + "angle": 0, + "content": "Given a real community, the metric indicates whether a generated community is over-diverse (\\(CTS < 0\\)), or over-similar (\\(CTS > 0\\)). A low value means more realism." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.79, + 0.626 + ], + "angle": 0, + "content": "In our comparisons, we compute several metrics. One metric is the Wasserstein Distance between the distributions of generated communities and real communities in terms of 2D position (x, y) and 2.5D geometry (length, width, height), and building count (CO) - we name these WD-5D and WD-CO. We define two quality metrics as well: Overlap indicates the percentage by which two building overlaps, and O-Blk represents the percentage of building area that sticks outside of its city block. We also implement several perceptual metrics: FID [24], KID [4], and LPIPS [65] to evaluate the visual quality and realism of generated layouts. For all the above metrics, lower value indicates better quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.65, + 0.788, + 0.711 + ], + "angle": 0, + "content": "Quantitative Results As Tab. 1 shows, our approach performs better than existing approaches in all but one metric. Note that the poor quality of SDXL results makes it hard to decipher layouts. We only compare to it by pixel-based perceptual metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Qualitative Results We show in Fig. 5 qualitative results for several exemplary areas (more examples are in Supplemental Material). In general, our method shows good contextual harmonization and has awareness of arbitrary road networks. It generates a realistic distribution of urban layouts with plausible context dependent behaviors. Other methods show improper patterns (e.g., overly diverse or similar) and potentially containing semantic errors or undesired overlaps." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.404, + 0.129 + ], + "angle": 0, + "content": "He, L. & Aliaga, D." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.144, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 2: Block Quantization Ablations. We report metrics (all in %) among all alternative ablations of our Block-level graph-based Variational AutoEncoder (BVAE). Using GAT-backbone with post-trained dimensional quantization with \\( L = 20 \\) shows the best overall compromise. The best values are in boldface." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.213, + 0.756, + 0.471 + ], + "angle": 0, + "content": "
Encode ModelOverlap↓Out-Blk↓Pos-E.↓Geom-E.↓Ct-E.↓Cov-E.↓
LayoutVAE [32]13.8411.1515.0937.35-24.34
VTN [2]2.704.706.0122.8618.581.97
BlockPlanner [62]2.931.304.7812.063.635.66
GlobalMapper [20]1.041.203.252.830.080.46
BVAE (SAGE [18])1.331.705.013.250.100.35
BVAE (GCN [33])0.451.214.252.740.040.28
BVAE* (GAT [58])0.170.281.000.710.0060.04
QuantizationOverlap↓Out-Blk↓Pos-E.↓Geom-E.↓Ct-E.↓Cov-E.↓
Trainable-VQ3.020.553.755.780.070.60
KMeans-Q2.900.0220.9817.8550.0116.23
DIM-Q (L=5)0.280.627.7116.2025.014.68
DIM-Q (L=10)0.560.285.734.488.334.58
DIM-Q (L=30)0.170.012.680.780.0050.15
DIM-Q* (L=20)0.110.013.500.890.0060.12
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.502, + 0.402, + 0.516 + ], + "angle": 0, + "content": "4.3 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.789, + 0.688 + ], + "angle": 0, + "content": "Block Quantization We ablate block quantization in order to understand and maximize the impact of changing its parameters and design strategy. First, we report in Tab. 2 (Encode Model Table) the reconstruction loss using different block-level variational autoencoders. We evaluate using various geometric metrics: Overlap (percent of building-to-building overlap), Out-Blk (percent of buildings that stick outside of its city block), Pos-E (positional error as a percentage of block diagonal), Geom-E (2.5D building geometry error as percentage of block scale), Ct-E (percentage of building count error), and Cov-E (total building coverage error). They clearly show our approach is the best overall. Thus we choose our BVAE based on Graph Attention Networks (GAT) [58]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.79, + 0.841 + ], + "angle": 0, + "content": "Then, in Tab. 2 (Quantization Table), we ablate various codebook approaches and quantitization levels for block-level encoding. \"Trainable-VQ\" leverages trainable vector quantization codebooks introduced by VQVAE [56]. We also provide an alternative non-training based quantization method using KMeans (\"KMeans-Q\") to cluster BVAE latent space values to a reasonable number of categories in order to handle diverse urban layouts (e.g., 100 classes). The \"DIM-Q\" rows imply quantizing each of the 512 dimensions of the block-level code into a codebook \\( C = \\{1,2,\\dots,L\\} \\) as described in Sec. 3.1. This per-dimension quantization performs clearly the best. Specifically, setting \\( L \\) to infinity produces a lossless representation of BVAE latent vectors, but also causes the curse of dimension-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.226, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.144, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 3: City-Scale GMAE Ablations. We ablate backbone message passing layers, maximum messaging passing hops (\\(D\\): depth of encoder), masking strategy during training, and scheduling functions of iterative generation (\\(T\\) indicates total iterations). The overall best setting is marked as *." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.213, + 0.785, + 0.417 + ], + "angle": 0, + "content": "
AblationsContext↓WD-3D↓WD-N↓Overlap↓Out-Blk↓FID↓KID↓LPIPS↓
GMAE (GCN [33])-0.302.742.910.260.3652.950.0250.32
GMAE (GTrans [52])0.743.023.271.961.0743.600.0180.29
GMAE (SAGE [18])0.972.982.361.752.6043.150.0170.29
GMAE (GAT, Dec)0.942.452.962.131.1048.240.0190.30
GMAE (GAT, D = 1)1.253.032.641.730.4443.750.0180.29
GMAE (GAT, D = 2)0.332.662.011.310.5734.820.0140.27
GMAE (GAT, D = 4)0.241.922.310.630.2824.590.0090.22
GMAE (Mask ∈ [0.05, 1])1.302.703.621.702.4543.140.0180.29
GMAE (Mask ∈ [0.05, 0.5])1.103.073.681.681.3046.850.0200.30
GMAE (Mask = 0.15)1.594.913.411.831.2049.500.0220.31
GMAE (Linear, T = 12)0.632.833.791.270.9337.420.0150.28
GMAE (Log, T = 12)-1.312.777.277.413.1463.250.0380.33
GMAE (Cosine, T = 1)-1.677.445.643.564.0176.120.0510.35
GMAE (Cosine, T = 20)0.322.923.981.680.3132.060.0130.24
*(GAT-D3, [0.5,1], Cos-T12)0.212.281.911.270.4223.630.0050.20
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.786, + 0.478 + ], + "angle": 0, + "content": "ality. We discover that \\( L = 20 \\) per dimension is the best overall compromise balancing compactness and representation ability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.786, + 0.55 + ], + "angle": 0, + "content": "City-Scale Graph Masked Autoencoder. In Tab. 3, We ablate various model parameters, together with significant training and inference strategies for our GMAE. Each alternatives is evaluated by the same set of metrics in Sec. 4.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.787, + 0.702 + ], + "angle": 0, + "content": "By varying the message passing layers, we find GAT [58] provides the best performance. GCN [33] produces diverse and unrealistic urban layouts, while transformer-based [52], and GraphSAGE [18] backbones generate blocks that are too similar. The depth of the stacked layers in the encoder (D) dictates the maximum message passing hops from each node and thus the span of context behaviors captured for realistic generation. We find that 3 hops is promising and larger contexts cause over-similarity and unnecessary computing burden. We also observed that using stacked GAT layers for decoding with remasking tricks, as described by [25,26], didn't produce a net benefit towards city generation. So we choose MLP as the node feature decoder for both its performance and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.779 + ], + "angle": 0, + "content": "The masking strategy during training is crucial for GMAE performance. We find that dynamic and large masking ratios are beneficial, while a fixed small ratio (e.g., 0.15 as in [14]) leads to over similarity. This behavior is also observed by Li et al. [35]. We adopt a similar masking strategy by sampling a Gaussian distribution \\( N(0.55, 0.25) \\), truncated by [0.5, 1.0], for each training iteration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Scheduling for iterative generation is also critical to capture representative city blocks (Sec. 3.3). We evaluate several functions \\(\\beta(t)\\) (e.g., linear, logarithmic, or cosine-based) that provide the scheduled acceptance ratio, and evaluate alternative total iteration counts \\(T\\). Rapid iteration leads to over-diversity and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.129 + ], + "angle": 0, + "content": "He, L. & Aliaga, D." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "long iteration results in over-similarity. Logarithmic functions with reverse pattern produces the worst realism. Metrics show that cosine-based function, which is conservative initially and generous in later iterations, outperforms others." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.239 + ], + "angle": 0, + "content": "In summary, this study determines the best configuration for our GMAE as: GAT encoder with \\( D = 3 \\) and MLP decoder, training mask ratio \\( m \\in [0.5,1.0] \\), and priority-based scheduling function \\( \\beta (t) = 1 - \\cos (t / T) \\) with \\( T = 12 \\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.261, + 0.357, + 0.274 + ], + "angle": 0, + "content": "4.4 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.286, + 0.784, + 0.346 + ], + "angle": 0, + "content": "Our method faces limitations where shapes of city blocks or buildings are not simple polygons (e.g. the building with a garden inside). It also struggles to handle very irregular and concave shaped city blocks (e.g., cul-de-sac's). Also, road-network to contextual structure relations are not explicitly considered." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.369, + 0.416, + 0.383 + ], + "angle": 0, + "content": "4.5 Additional Results" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.395, + 0.719, + 0.41 + ], + "angle": 0, + "content": "We place several other interesting results in Supplementary Materials." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.422, + 0.784, + 0.467 + ], + "angle": 0, + "content": "- Controllable Generation of Entire City. Our method enables realistic and efficient city-scale controllable generation given any percentage of priors. We show several large examples spanning a subset of the many cities." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.468, + 0.784, + 0.511 + ], + "angle": 0, + "content": "- Socio-Economic Prediction. Trained GMAE may act as a powerful encoder for downstream tasks. For example, the encoded GMAE latent vectors of urban layouts can indicate social-economic metrics for each city block." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.513, + 0.784, + 0.54 + ], + "angle": 0, + "content": "- Semantic Manipulations. Our method enables semantic manipulations over various building layout styles across different cities." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.422, + 0.784, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.567, + 0.55, + 0.582 + ], + "angle": 0, + "content": "5 Conclusions and Future Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.785, + 0.764 + ], + "angle": 0, + "content": "Large-scale urban layout generation has seen significant advancements, spurred by the intersection of computer vision, urban planning, and related disciplines. Previous methodologies often neglect the context-sensitive nature of urban layouts, failing to capture the multi-layer semantics crucial for realism and plausibility. Our proposed Graph-based Masked AutoEncoder (GMAE) addresses these limitations by leveraging a graph-based representation, acknowledging context-sensitivity, and prioritizing the generation process. By training on a diverse dataset of 330 cities, we demonstrate the efficacy of our approach in generating large-scale 2.5D layouts with realism and semantic consistency. Moreover, our method outperforms existing techniques in various metrics, marking a significant advancement in the realm of urban layout generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Moving forward, our approach paves the way for more accurate and efficient urban simulation, digital twin creation, and game/content design. As future work we intend to incorporate our method for large-scale photo-realistic multi-view scene synthesis, and city-scale 3D modeling. Those potentials will contribute to synthetic data generation in autonomous driving and world model training." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.226, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.179, + 0.785, + 0.209 + ], + "angle": 0, + "content": "This project was funded in part by NSF Grant #2107096 and NSF Grant #1835739." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.235, + 0.323, + 0.25 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.268, + 0.784, + 0.296 + ], + "angle": 0, + "content": "1. Climate and economic justice screening tool. https://screeningtool.geoplatform.gov/" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.297, + 0.785, + 0.338 + ], + "angle": 0, + "content": "2. Arroyo, D.M., Postels, J., Tombari, F.: Variational transformer networks for layout generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13642-13652 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.785, + 0.393 + ], + "angle": 0, + "content": "3. Bhatt, M., Kalyanam, R., Nishida, G., He, L., May, C., Niyogi, D., Aliaga, D.: Design and deployment of photo2building: A cloud-based procedural modeling tool as a service. In: Practice and Experience in Advanced Research Computing, pp. 132-138 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.394, + 0.785, + 0.422 + ], + "angle": 0, + "content": "4. Binkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying mmd gans. arXiv preprint arXiv:1801.01401 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.422, + 0.785, + 0.45 + ], + "angle": 0, + "content": "5. Bokeloh, M., Wand, M., Seidel, H.P.: A connection between partial symmetry and inverse procedural modeling. In: ACM SIGGRAPH 2010 papers, pp. 1-10 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.45, + 0.785, + 0.505 + ], + "angle": 0, + "content": "6. Brooks, T., Peebles, B., Holmes, C., DePue, W., Guo, Y., Jing, L., Schnurr, D., Taylor, J., Luhman, T., Luhman, E., Ng, C., Wang, R., Ramesh, A.: Video generation models as world simulators (2024), https://openai.com/research/video-generation-models-as-world-simulators" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.506, + 0.785, + 0.547 + ], + "angle": 0, + "content": "7. Bureau, U.S.C.: Topologically integrated geographic encoding and referencing. https://www.census.gov/geographies/mapping-files/time-series/geo/tiger-line-file.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.548, + 0.785, + 0.589 + ], + "angle": 0, + "content": "8. Chai, L., Tucker, R., Li, Z., Isola, P., Snavely, N.: Persistent nature: A generative model of unbounded 3d worlds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20863-20874 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.59, + 0.785, + 0.631 + ], + "angle": 0, + "content": "9. Chai, S., Zhuang, L., Yan, F.: Layoutdm: Transformer-based diffusion model for layout generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18349-18358 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.632, + 0.785, + 0.673 + ], + "angle": 0, + "content": "0. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked generative image transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11315-11325 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.785, + 0.729 + ], + "angle": 0, + "content": "1. Chang, K.H., Cheng, C.Y., Luo, J., Murata, S., Nourbakhsh, M., Tsuji, Y.: Building-gan: Graph-conditioned architectural volumetric design generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11956-11965 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.785, + 0.757 + ], + "angle": 0, + "content": "2. Chen, Z., Wang, G., Liu, Z.: Scenedreamer: Unbounded 3d scene generation from 2d image collections. arXiv preprint arXiv:2302.01330 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "3. Deng, J., Chai, W., Guo, J., Huang, Q., Hu, W., Hwang, J.N., Wang, G.: Citygen: Infinite and controllable 3d city layout generation. arXiv preprint arXiv:2312.01508 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "4. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.268, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.128 + ], + "angle": 0, + "content": "He, L. & Aliaga, D." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "16. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873-12883 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.245, + 0.787, + 0.3 + ], + "angle": 0, + "content": "17. Gupta, K., Lazarow, J., Achille, A., Davis, L.S., Mahadevan, V., Shrivastava, A.: Layouttransformer: Layout generation and completion with self-attention. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1004-1014 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.3, + 0.787, + 0.328 + ], + "angle": 0, + "content": "18. Hamilton, W., Ying, Z., Leskovec, J.: Inductive representation learning on large graphs. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "19. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.412 + ], + "angle": 0, + "content": "20. He, L., Aliaga, D.: Globalmapper: Arbitrary-shaped urban layout generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 454-464 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.467 + ], + "angle": 0, + "content": "21. He, L., Lu, Y., Corring, J., Florencio, D., Zhang, C.: Diffusion-based document layout generation. In: Fink, G.A., Jain, R., Kise, K., Zanibbi, R. (eds.) Document Analysis and Recognition - ICDAR 2023. pp. 361-378. Springer Nature Switzerland, Cham (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.495 + ], + "angle": 0, + "content": "22. He, L., Shan, J., Aliaga, D.: Generative building feature estimation from satellite images. IEEE Transactions on Geoscience and Remote Sensing 61, 1-13 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.522 + ], + "angle": 0, + "content": "23. Heris, M.P., Foks, N.L., Bagstad, K.J., Troy, A., Ancona, Z.H.: A rasterized building footprint dataset for the united states. Scientific data 7(1), 207 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.564 + ], + "angle": 0, + "content": "24. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.787, + 0.605 + ], + "angle": 0, + "content": "25. Hou, Z., He, Y., Cen, Y., Liu, X., Dong, Y., Kharlamov, E., Tang, J.: Graphmae2: A decoding-enhanced masked self-supervised graph learner. In: Proceedings of the ACM Web Conference 2023. pp. 737-746 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.647 + ], + "angle": 0, + "content": "26. Hou, Z., Liu, X., Cen, Y., Dong, Y., Yang, H., Wang, C., Tang, J.: Graphmae: Self-supervised masked graph autoencoders. In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 594-604 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.688 + ], + "angle": 0, + "content": "27. Hua, H., Shi, J., Kafle, K., Jenni, S., Zhang, D., Collomosse, J., Cohen, S., Luo, J.: Finematch: Aspect-based fine-grained image and text mismatch detection and correction. arXiv preprint arXiv:2404.14715 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.73 + ], + "angle": 0, + "content": "28. Hui, M., Zhang, Z., Zhang, X., Xie, W., Wang, Y., Lu, Y.: Unifying layout generation with a decoupled diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1942-1951 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.787, + 0.785 + ], + "angle": 0, + "content": "29. Inoue, N., Kikuchi, K., Simo-Serra, E., Otani, M., Yamaguchi, K.: Layoutdm: Discrete diffusion model for controllable layout generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10167-10176 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "30. Jiang, Z., Guo, J., Sun, S., Deng, H., Wu, Z., Mijovic, V., Yang, Z.J., Lou, J.G., Zhang, D.: Layoutformer++: Conditional graphic layout generation via constraint serialization and decoding space restriction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18403-18412 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "31. Jyothi, A.A., Durand, T., He, J., Sigal, L., Mori, G.: Layoutvae: Stochastic scene layout generation from a label set. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9895-9904 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.243 + ], + "angle": 0, + "content": "32. Jyothi, A.A., Durand, T., He, J., Sigal, L., Mori, G.: Layoutvae: Stochastic scene layout generation from a label set. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9894-9903 (2019). https://doi.org/10.1109/ICCV.2019.00999" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.272 + ], + "angle": 0, + "content": "33. Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "34. Li, J., Yang, J., Hertzmann, A., Zhang, J., Xu, T.: Layoutgan: Synthesizing graphic layouts with vector-wireframe adversarial networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(7), 2388-2399 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.369 + ], + "angle": 0, + "content": "35. Li, T., Chang, H., Mishra, S., Zhang, H., Katabi, D., Krishnan, D.: Mage: Masked generative encoder to unify representation learning and image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2142-2152 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.412 + ], + "angle": 0, + "content": "36. Li, Z., Wang, Q., Snavely, N., Kanazawa, A.: Infinitenature-zero: Learning perpetual view generation of natural scenes from single images. In: European Conference on Computer Vision. pp. 515-534. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.438 + ], + "angle": 0, + "content": "37. Lin, C.H., Lee, H.Y., Cheng, Y.C., Tulyakov, S., Yang, M.H.: Infinitygan: Towards infinite-pixel image synthesis. arXiv preprint arXiv:2104.03963 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "38. Lin, C.H., Lee, H.Y., Menapace, W., Chai, M., Siarohin, A., Yang, M.H., Tulyakov, S.: Infinity: Infinite-scale city synthesis. arXiv preprint arXiv:2301.09637 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "39. Lipp, M., Scherzer, D., Wonka, P., Wimmer, M.: Interactive modeling of city layouts using layers of procedural content. In: Computer Graphics Forum. vol. 30, pp. 345-354. Wiley Online Library (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.549 + ], + "angle": 0, + "content": "40. Ma, H., Zeng, D., Liu, Y.: Learning individualized treatment rules with many treatments: A supervised clustering approach using adaptive fusion. Advances in Neural Information Processing Systems 35, 15956-15969 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "41. Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.632 + ], + "angle": 0, + "content": "42. Nauata, N., Chang, K.H., Cheng, C.Y., Mori, G., Furukawa, Y.: House-gan: Relational generative adversarial networks for graph-constrained house layout generation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 162-177. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.66 + ], + "angle": 0, + "content": "43. OpenStreetMap contributors: Planet dump retrieved from https://planet.osm.org. https://www.openstreetmap.org (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "44. Para, W., Guerrero, P., Kelly, T., Guibas, L.J., Wonka, P.: Generative layout modeling using constraint graphs. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6690-6700 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.742 + ], + "angle": 0, + "content": "45. Patel, P., Kalyanam, R., He, L., Aliaga, D., Niyogi, D.: Deep learning-based urban morphology for city-scale environmental modeling. PNAS nexus 2(3), pgad027 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.798 + ], + "angle": 0, + "content": "46. Patil, A.G., Ben-Eliezer, O., Perel, O., Averbuch-Elor, H.: Read: Recursive autoencoders for document layout generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 544-545 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "47. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4195-4205 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.128 + ], + "angle": 0, + "content": "He, L. & Aliaga, D." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "48. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "49. Shen, Y., Ma, W.C., Wang, S.: Sgam: Building a virtual 3d world through simultaneous generation and mapping. Advances in Neural Information Processing Systems 35, 22090-22102 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.271 + ], + "angle": 0, + "content": "50. Sheng, Y., Liu, Y., Zhang, J., Yin, W., Oztireli, A.C., Zhang, H., Lin, Z., Shechtman, E., Benes, B.: Controllable shadow generation using pixel height maps. In: European Conference on Computer Vision. pp. 240-256. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.325 + ], + "angle": 0, + "content": "51. Sheng, Y., Yu, Z., Ling, L., Cao, Z., Zhang, X., Lu, X., Xian, K., Lin, H., Benes, B.: Dr. bokeh: Differentiable occlusion-aware bokeh rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4515-4525 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.367 + ], + "angle": 0, + "content": "52. Shi, Y., Huang, Z., Feng, S., Zhong, H., Wang, W., Sun, Y.: Masked label prediction: Unified message passing model for semi-supervised classification. arXiv preprint arXiv:2009.03509 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.785, + 0.42 + ], + "angle": 0, + "content": "53. Song, Y., Zhang, Z., Lin, Z., Cohen, S., Price, B., Zhang, J., Kim, S.Y., Aliaga, D.: Objectstitch: Object compositing with diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18310-18319 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.421, + 0.785, + 0.476 + ], + "angle": 0, + "content": "54. Song, Y., Zhang, Z., Lin, Z., Cohen, S., Price, B., Zhang, J., Kim, S.Y., Zhang, H., Xiong, W., Aliaga, D.: Imprint: Generative object compositing by learning identity-preserving representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8048-8058 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.476, + 0.785, + 0.502 + ], + "angle": 0, + "content": "55. Tabata, S., Yoshihara, H., Maeda, H., Yokoyama, K.: Automatic layout generation for graphical design magazines. In: ACM SIGGRAPH 2019 Posters, pp. 1-2 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.529 + ], + "angle": 0, + "content": "56. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.785, + 0.57 + ], + "angle": 0, + "content": "57. Vanegas, C.A., Kelly, T., Weber, B., Halatsch, J., Aliaga, D.G., Müller, P.: Procedural generation of parcels in urban modeling. In: Computer graphics forum. vol. 31, pp. 681-690. Wiley Online Library (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.597 + ], + "angle": 0, + "content": "58. Velicković, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y.: Graph attention networks. arXiv preprint arXiv:1710.10903 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.785, + 0.623 + ], + "angle": 0, + "content": "59. Velickovic, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y., et al.: Graph attention networks. stat 1050(20), 10-48550 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.785, + 0.664 + ], + "angle": 0, + "content": "60. Wu, W., Fu, X.M., Tang, R., Wang, Y., Qi, Y.H., Liu, L.: Data-driven interior plan generation for residential buildings. ACM Transactions on Graphics (TOG) 38(6), 1-12 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.785, + 0.691 + ], + "angle": 0, + "content": "61. Xie, H., Chen, Z., Hong, F., Liu, Z.: Citydreamer: Compositional generative model of unbounded 3d cities. arXiv preprint arXiv:2309.00610 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.732 + ], + "angle": 0, + "content": "62. Xu, L., Xiangli, Y., Rao, A., Zhao, N., Dai, B., Liu, Z., Lin, D.: Blockplanner: City block generation with vectorized graph representation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5077-5086 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.785, + 0.759 + ], + "angle": 0, + "content": "63. Yan, W., Zhang, Y., Abbeel, P., Srinivas, A.: Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "64. Yang, C.F., Fan, W.C., Yang, F.E., Wang, Y.C.F.: Layouttransformer: Scene layout generation with conceptual and spatial diversity. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3732-3741 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "65. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "66. Zhang, X., Ma, W., Varinlioglu, G., Rauh, N., He, L., Aliaga, D.: Guided pluralistic building contour completion. The Visual Computer 38(9), 3205-3216 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.205 + ], + "angle": 0, + "content": "67. Zheng, X., Qiao, X., Cao, Y., Lau, R.W.: Content-aware generative modeling of graphic design layouts. ACM Transactions on Graphics (TOG) 38(4), 1-15 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.205 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/COHO_ Context-Sensitive City-Scale Hierarchical Urban Layout Generation/ad29cc50-ea60-4c58-8c17-9b4b4ddc225f_origin.pdf b/2024/COHO_ Context-Sensitive City-Scale Hierarchical Urban Layout Generation/ad29cc50-ea60-4c58-8c17-9b4b4ddc225f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3a828cc5d322c3f373602b94ae3b27248c210ae7 --- /dev/null +++ b/2024/COHO_ Context-Sensitive City-Scale Hierarchical Urban Layout Generation/ad29cc50-ea60-4c58-8c17-9b4b4ddc225f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64a9640c99feeb2e3d21b4058f199db654dc3463bcdb69c2329e5678289fcecc +size 7163318 diff --git a/2024/COHO_ Context-Sensitive City-Scale Hierarchical Urban Layout Generation/full.md b/2024/COHO_ Context-Sensitive City-Scale Hierarchical Urban Layout Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9fdf439494ed8bd438cbd756bcc0001a9e62c407 --- /dev/null +++ b/2024/COHO_ Context-Sensitive City-Scale Hierarchical Urban Layout Generation/full.md @@ -0,0 +1,297 @@ +# COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation + +Liu He, Daniel Aliaga + +Purdue University, USA + +{he425,aliaga}@purdue.edu + +![](images/eeb1e4304f82043ce7508d9d636bb723fd83c5cac48bbe79bb742ab17b5baafa.jpg) +LayoutDM $(CTS = -2.20)$ + +![](images/36b085177dada633a268f3c5ddb070f50057286d0b5db2990542f7f062b97496.jpg) +Ours $(CTS = 0.21)$ + +![](images/ce3753b00c29ac292e17d8c267271a185b6fa36ae11a58b86b5899746e1ebdbf.jpg) +GlobalMapper $(CTS = 0.62)$ + +![](images/ca21aa3b26064b52762e7dc614eaae7900fb1e159837833012992d80dc7ee4f1.jpg) +Random $(CTS\ll 0)$ + +![](images/edde6205d9b52d281d179b0fe6c2f5c40ad8b41e31b4c6187bf9c2acd950add2.jpg) +Real Data $(CTS\rightarrow 0)$ + +![](images/1b121afc5408cfebba0a8399974f27efb816b208138162c225ec21c17054eef9.jpg) +Identical $(CTS\gg 0)$ +Fig. 1: Context-Sensitive Generation. Our method (Green) pursues realistic context harmonization among neighboring city blocks as real data (bottom-middle). Other methods (e.g. LayoutDM [29], GlobalMapper [20]) show over-diversity/-similarity in city-scale layout generation (evaluated by Context Score $CTS$ as Eq. 4.). Fully random and identical layouts are synthetically generated to illustrate extreme cases. + +Abstract. The generation of large-scale urban layouts has garnered substantial interest across various disciplines. Prior methods have utilized procedural generation requiring manual rule coding or deep learning needing abundant data. However, prior approaches have not considered the context-sensitive nature of urban layout generation. Our approach addresses this gap by leveraging a canonical graph representation for the entire city, which facilitates scalability and captures the multilayer semantics inherent in urban layouts. We introduce a novel graph-based masked autoencoder (GMAE) for city-scale urban layout generation. The method encodes attributed buildings, city blocks, communities and cities into a unified graph structure, enabling self-supervised + +masked training for graph autoencoder. Additionally, we employ scheduled iterative sampling for 2.5D layout generation, prioritizing the generation of important city blocks and buildings. Our approach achieves good realism, semantic consistency, and correctness across the heterogeneous urban styles in 330 US cities. Codes and datasets are released at https://github.com/Arking1995/COHO. + +Keywords: Layout Generation $\cdot$ Urban Modeling $\cdot$ Masked Graph Autoencoder $\cdot$ Context Sensitivity + +# 1 Introduction + +Large-scale urban layout generation has received significant interest by computer vision, urban planning and related disciplines. In particular, cities easily have 1,000 to 50,000 city blocks spanning a mostly horizontal terrain. The configuration of each city block varies significantly across the city, yet there is some configuration similarity amongst subsets within and among cities. Reconstructing existing layouts and generating future planned layouts are crucial tasks for urban simulation, digital twins, and game/content design. + +Prior works have generated urban layouts from a variety of sources. Procedural layout generation [39, 57] used hand-coded rules to generate cities, and inverse procedural modeling [5] extracted the rules from data and then enabled generation. More recently, deep learning has promoted data-driven approaches that produce either pixel-based layout generation (e.g., InfiniCity [38], CityDreamer [61], CityGEN [13]) or graph-based approaches for general polygonal shape layouts (e.g., LayoutTransformer [17], VTN [2]) and for urban scenarios (e.g., BlockPlanner [62], GlobalMapper [20], BuildingGAN [11]). However, none of these techniques explicitly recover and consider the context-sensitive nature of urban layout generation (Fig. 1). In other words, the layout of buildings, blocks, and communities cannot be treated in isolation. The multi-layer semantics are critical to generating realistic and plausible urban layouts. While such multilevel structure (e.g., from single building to a community or subset of a city) has been specified via hand-coded rules, it has not been discovered, organized and used for generation in data-driven deep learning approaches. + +Our approach builds on three key observations: + +1. Representation: Cities, communities, blocks, buildings, and roads are arbitrarily shaped and with complex topology. Hence, a graph-based representation can better capture the regular and non-regular configurations, as opposed to the limiting regular structure of an image. Further, graphs can be more compact which in turn supports better scalability. Altogether, having buildings or city blocks as atomic units, instead of pixels, will contribute to improving correctness, realism, and consistency with priors. +2. Context-Sensitivity: Buildings, city blocks, and communities are built considering their neighboring structures and not in isolation. Hence, a generator should consider this multi-layer structure and inter-dependency. In + +order words, the style features at one level (e.g., size, shape, location) depend on the surrounding context. + +3. Prioritization: The stylistic and semantic importance, or priority, of city blocks and buildings varies. This implies that a generation scheme should generate the most important city blocks and buildings first and then fill in the rest, as opposed to a region growing approach for example. Overall, prioritization will improve correctness and realism as well as enable large-scale urban layout generation. + +We propose a novel Graph-based Masked AutoEncoder (GMAE) for large-scale multi-layer 2.5D urban layout generation. First, we define a canonical graph-based representation capturing the multi-layer context sensitivity of any city's urban layouts (Fig. 2). The entire graph $G$ corresponds to a city. Subsets of the graph map to communities and each node $b$ amounts to a city block. Further, each node encodes its building layouts, shapes, and heights as a quantized feature by well-trained quantizor. Second, we utilize our GMAE to learn the multi-layer stylistic behavior of many urban layouts (i.e., 330 cities). + +Using subsets of the graph (or communities), node features are randomly masked and self-supervised GMAE training is performed to learn node feature reconstruction (Fig. 3). Third, utilizing the well-trained GMAE as a generator, we perform an iterative priority-based scheduling to generate a large layout or to complete a city fragment; i.e., each iteration predicts node features for the entire graph but only the most confident ones are preserved for the next step (Fig. 4). This methodology is able to leverage any percentage of given prior (e.g., $[0,100\%]$ ) for urban layout completion and generation. + +To the best of our knowledge, our approach is the first data-driven deep-learning method to enable large-scale 2.5D layout generation with good realism, semantic consistency, and correctness. We show the capability to generalize to the heterogeneous urban layout styles of 330 cities in the US (e.g., all cities with a population $>100\mathrm{K}$ , totaling 833,473 city blocks and 17,663,607 buildings). In addition, we provide comparisons to several prior works including SDXL [48]), VTN [2], LayoutDM [29], and GlobalMapper [20] in Sec. 4 and exceed their performance in multiple well-known geometric and perceptual metrics. + +Our contributions include the following: + +- A canonical graph representation for large-scale 2.5D urban layouts. It encodes global and local node features and neighboring relations for buildings, city blocks, and communities within a city. +- A self-supervised Graph-based Masked AutoEncoder (GMAE) enabling arbitrarily shaped city layout generation with consistency, correctness, and realism +- Priority-based scheduled iterative sampling for controllable urban layout synthesis including completion, refinement, and generation with any percentage of priors ([0, 100%]). +- A open dataset of road network and urban layouts of 330 US cities for benchmarking current and future urban layout generation methods (https://huggingface.co/datasets/Arking95/COHO). + +# 2 Related Work + +Layout Generation. Numerous data-driven deep learning methods have been proposed for layout generation; for example, generalized layout synthesis [2,17, 31,34,64], document layouts [21,30,46,55,67], urban land lots [22,62,66], indoor scenes [42,44,60], and 3D buildings [3,11,45]. Recently, diffusion models have also been used for layout generation tasks (e.g., [9,21,28,29]). However, most of these works assume 2D layouts generated on a rectangular empty canvas. The position and size of each bounding box is often limited to a finite set of categories and to rectilinear alignments. This results in poor diversity and low realism for urban layout generation tasks. BlockPlanner [62] and GlobalMapper [20] adapt layout generation to either rectilinear or arbitrarily-shaped city block generation. However, all previous works assume each layout, or city block, is independent. This is not the case in urban environments where adjacent and nearby city blocks may share a common style or be positioned in a purposeful multi-block arrangement. Without a contextual guidance, city block generation may result in either unwanted harmonization or unexpected diversity between neighboring blocks (see Fig. 1). Thus, we pursue city-scale urban layout generation with the help of context-aware neighboring information. + +Infinite Visual Synthesis. With bombing of visual deep learning works [27, 40, 50, 51, 53, 54], researchers have explored synthesizing large, potentially unbounded images. For example, omnidirectional image synthesis from multiple smaller field-of-view images [36, 37]. Other works focus on unbounded 3D scene generation [8, 12, 49]. Particularly relevant are several recent infinite urban scene generators [13, 38, 61] that directly adapt image synthesis methods (e.g. InfinityGAN [37], MaskGIT [10]). The synthesis performs an autoregressive outpainting process using layout image patches. However, patch-based training and inference doesn't learn global features nor their interdependencies. Rather, it learns a very local behavior. Prior pixel-based methods acknowledge limited realism and semantics [13, 38, 61]. Often, post-processing [13, 61], and human-in-the-loop editing [38] is required to refine the generated images to vector-based urban layouts. Even the latest stable diffusion model SDXL [48] seeks high-resolution image synthesis but may not produce realistic urban layouts (Fig. 5). + +Learning based on Quantized Representations. Vector quantization has been a prevailing stepping-stone technique for text (e.g., Word2Vec [41]), image (e.g., VQVAE/GAN [16, 56], [15]), and video synthesis (e.g., SORA [6, 47], [63]). Pretrained quantizers provide a scalable token representation benefiting high-level and large-scale synthesis. A well-trained quantizer combined with self-supervised masked training often performs well in reconstructing the original signal during generative tasks. BERT [14] provides plausible text-based representation learning. MAE [19] defines a masked autoencoder for image space which excels at multiple downstream tasks. MAE also provides representation + +learning for a broad range of node/graph feature synthesis [25, 26]. Specifically, MaskGIT [10] and MAGE [35] utilize MAE frameworks for impressive image generation tasks. The idea of a generative masked autoencoder based on pretrained quantization was inspirational to our city-scale urban layout approach. + +# 3 Method + +First we describe our canonical graph representation of 2.5D urban layouts (Fig. 2). Second, we explain the training process for our GMAE based on the graphs for many cities (Fig. 3). Third, we show our priority-based scheduling for iteratively generating city-scale urban layouts given any amount of prior (Fig. 4). + +# 3.1 Canonical Graph Representation + +A city is represented by a graph $G = \{B, E\}$ , where $B$ is a set of nodes $B = \{b_i | i \in [1, N]\}$ such that $b_i$ represents each city block (Fig. 2), and any connected subgraph of $G$ corresponds to a community. $N = |B|$ is the total number of blocks for a given city. $E$ is a set of edges $E = \{e_{ij} | i, j \in [1, N]\}$ representing mutually adjacent nodes (or city blocks). $K = |E|$ is the total number of edges. + +Node (or City Block). Each node, or city block, $b_{i}$ is a region defined by an enclosing road network and contains a set of block-level features and building-level features. The per-node features include: + +- $s_i$ is the shape and location feature vector of a city block contour. It includes four features: (1) aspect ratio, (2) total block area, (3) ratio of total block area over the block's convex-hull area, and (4) relative distance from current block to the centroid of the city. +- $q_{i}$ is a 512-dimensional vector from a codebook $C$ . It hierarchically captures the buildings layouts and their configuration within $b_{i}$ . + +We concatenate these node features to form a 516-dimensional vector $\left([s_i, q_i]\right)$ , which is able to represent our large and heterogeneous set of blocks and buildings. + +Building Layout Quantization. The 512-dimensional vector $q_{i}$ describing the building layouts within a city block is obtained from a quantized codebook $C = \{1,2,\dots,L\}$ with quantization level $L = |C|$ . To create the codebook, we train a block-level variational autoencoder (BVAE) for building layout within a city block, using a canonical graph structure for representing the multiple building shapes, positions, and heights inside a single city block. Using this pretrained encoder, the latent vector of all possible building layouts will form a numerical distribution for each latent dimension. We further quantize each of these distributions into $L$ bins with equal percentiles. Then, each latent value is replaced by the rank of the bin containing its value. Therefore, any building + +![](images/beb3bd749036747d4c0ba7eb9860694e92bf71f2956b85b6de9b26f6a9d12de5.jpg) +Fig. 2: Canonical Graph Representation. Our method represents a city as a canonical graph $G$ . Each node $b_{i}$ represent a single city block, and each edge $e_{ij}$ connects spatially adjacent blocks. Each block/node corresponds to a set of node features $s_i$ and a quantized vector $q_{i}$ hierarchically capturing enclosed building layouts. Edge feature $d_{ij}$ encodes distances between block centroids. Graph $G$ is used for GMAE training. + +layout within a city block is quantized into a 512-dimensional categorical index vector $q_{i} = [c_{1},\dots,c_{512}]$ , where each $c\in C$ . Note that our codebook is defined after BVAE training and not dynamically trainable as in VQGAN/VQVAE [16,56]. After performing a set of detailed evaluations using different training schemes, encoding models, and quantization levels $L$ (see Sec. 4.3), we selected the Graph Attention Network (GAT) [59] as the backbone for the BVAE. Further, we select $L = 20$ to balance compactness and representation ability after quantization. BVAE and quantization details are found in Supplementary Materials. + +Edge. Each edge $e_{ij}$ contains a feature $d_{ij}$ storing the relative distance between the centroids of the adjacent city blocks $b_i$ and $b_j$ . This distance-weighted adjacency helps encode the relationship between neighboring blocks. + +Altogether, the shape, style, and multi-layer urban layout of a city with $N$ nodes and $K$ edges is represented by a total of $(516N + K)$ variables. This canonical graph is used for the self-supervised GMAE training described below. + +# 3.2 Graph-based Masked AutoEncoder + +We train our GMAE using a self-supervised process. A road-only city (or community) is a graph (or subgraph) $G$ where building layout features $q_{i}$ of every node (block) are missing, but $s_i$ and $e_{ij}$ are preserved. A context-sensitive node feature generation process corresponds to recreating the purposefully removed building layout features. + +During training (Fig. 3), we use dynamic masking ratios of $q_{i}$ in order to learn how to regenerate the original features. In each training iteration, the mask ratio $m$ is sampled from a truncated Gaussian distribution as recommended by Li et al. [35]. The ratio is obtained from the range $m \in [0.5,1.0]$ and is centered at + +![](images/b4bd815f63a9b166abdc9f26433bee9e908635be57d3c82ee2b98cdc8964af06.jpg) +Fig. 3: Graph-based Masked Autoencoder. Given a canonical city graph, quantized building layout features $Q$ are masked with dynamic masking ratios $m \in [0.5,1.0]$ , while block shape and location features $S$ are kept. The GNN encoder uses message passing between neighboring nodes to obtain the context-aware node features $F$ . The decoder uses $F$ to reconstruct $Q'$ . The predicted $Q'$ are decoded to 2.5D urban layouts. + +0.55. We denote building layout features as $Q = \{q_{i} | i \in [1, N]\}$ and those after masking as $Q^{mask}$ . Block features are denoted as $S = \{s_{i} | i \in [1, N]\}$ . The graph neural network (GNN) encoder is denoted as $f_{E}$ , and node feature decoder as $f_{D}$ . Thus the process of our GMAE can be written as: + +$$ +F = f _ {E} ([ Q ^ {m a s k}, S ], E), \quad Q ^ {\prime} = f _ {D} (F) \tag {1} +$$ + +where $F = \{f_i | i \in [1, N]\}$ indicates the context-sensitive node features obtained by message passing between neighboring nodes by $f_E$ . Extracted $F$ is processed by $f_D$ to reconstruct building layout features $Q'$ . + +The backbones of $f_{E}$ and $f_{D}$ can be any GNN, including GAT [58], GCN [33], and GraphSAGE [18]. After experimentation, we select GAT as $f_{E}$ because of its best context-sensitive feature extraction ability, and we choose a straightforward MLP as $f_{D}$ for efficient processing. We observed that the re-masking trick and complex GNN decoder utilized by [25,26] requires more memory and calculation but has no net benefit in our application. Since $Q$ contains quantized index vectors, we use cross entropy loss for self-supervised reconstruction training on masked nodes. Detailed ablations of masking strategy and model selections are provided in Sec. 4.3. + +Our objective function is as follows, where $L$ indicates the length of the quantized codebook $C$ : + +$$ +L _ {r e c o n} = - \sum_ {i = 1} ^ {L} \left[ Q _ {i} \log \left(Q _ {i} ^ {\prime}\right) \right] ^ {m a s k} \tag {2} +$$ + +To capture community behavior (and also reduce GPU memory usage), we randomly sample a batch of subgraphs from $G$ for each training iteration. The nodes in each subgraph are sampled using a chosen radius from a random center node. While community sizes vary significantly, we found that a practical compromise is to use 500 meters as the radius value. This radius ensures to cover a reasonable number of the city blocks with a typical census tract from governmental CJEST dataset [1]. + +![](images/d91eb5da47464a7c083d15174c7970e1f4a263d9ab769e3268343e3253a648ab.jpg) +Fig. 4: Priority-based Scheduled Generation. We iteratively utilize pretrained GMAE to reconstruct masked node features. In each iteration, we accept a certain ratio of predicted nodes decided by the scheduling function $\beta(t) = 1 - \cos(t / T)$ . We obtain a full graph after $T$ iterations. + +# 3.3 Priority-based Scheduled Generation + +While in theory the trained GMAE can predict all nodes in a single pass, the generation quality is much better when performed iteratively. Recent published papers (e.g., [10,35]) have found that gradually sampling with a reasonable number of iterations improves generation quality. In urban scenarios, small number of representative city blocks are important to indicate context behavior. We intend to generate those blocks at early iterations and then span neighboring blocks. + +In Fig. 4, we propose a $T$ -iteration sampling strategy similar to MaskGIT [10] and MAGE [35]. At each iteration, GMAE predicts per-node building layout features for all remaining nodes. Then we rank generated nodes by their prediction confidence. The generated nodes with the highest prediction confidence are accepted. The accepted nodes after the current iteration $t \in [1,T]$ are scheduled by a cosine function $\beta(t) = 1 - \cos(t / T)$ . When $t$ is small, the ratio is near to 0 causing few nodes to be accepted in initial iterations. As iterations continue, $t$ gets bigger and hence the number of accepted nodes during each iteration will also increase. This behavior follows the prioritization observation in Sec. 1 which causes distinguished, or unique, city blocks to be generated first and essentially guide the layout style for nearby city blocks. Typically we set $T = 12$ which results in dozens of nodes being generated as is the typical size of a community. Detailed ablations of alternative sampling schedules are provided in Sec. 4.3. + +# 4 Experiments + +# 4.1 Implementation Details + +Open Dataset We have created a vector-based urban layout dataset of all US cities with a population of 100K or more. We believe this dataset will be of significant use to the community to further develop urban layout generation methods and thus we will release the dataset together with the paper. For each city, we define a rectangular bounding box containing most of the metropolitan area. Then, + +we use this bounding box to extract data from OpenStreetMap (OSM) [43], Microsoft Building Footprints (MSF) [23], and Topologically Integrated Geographic Encoding and Referencing (TIGER) dataset [7]. The data we collected includes: (1) road network vectors and attributes, (2) building footprints and heights, and (3) block-level social economic metrics. Specifically, we collect road networks and block contours from TIGER. Then we look for building footprints and heights from OSM and MSF within each city block contour. We choose the one with more and accurate details (e.g., containing more buildings, having features that match those in TIGER). Over all 330 cities, this dataset has a total of 833,473 city blocks, and 17,663,607 buildings. We set training, validation, testing split as $70\%$ , $20\%$ , and $10\%$ respectively. + +Training Details The training has two stages. We first use all city blocks for self-supervised training of a graph-based Block-level Variational AutoEncoder (BVAE). Given this trained block-level encoder, we perform the latent quantization for all building layouts in each city block. Then our GMAE is trained based on canonical city graph representation. Training time for BVAE and GMAE typically takes 12 hours and 15 hour respectively by a single A5000 GPU. We will release our codes upon acceptance. + +Inferencing using our graph-based approach is seemingly instantaneous for all the results shown in this paper. This is contrast to the VTN [2] and SDXL [48] approaches which take significantly more time to infer and to train. + +# 4.2 Comparisons + +We compare our method to several alternate city-scale layout generators (Tab. 1). Specifically, we randomly select 100 communities from among our test set, total 3700 city blocks. All methods generate the layout of the 100 communities in one pass, and without any post-processing or human-in-the-loop refinement. Implementations are downloaded from the author's repositories and trained with default parameter settings. For fairness, we convert our dataset to token sequences, graphs, or image patches $(1024\times 1024)$ as needed. VTN [2], LayoutDM [29], GlobalMapper [20] are retrained for single block generation. SDXL [48] is fine-tuned for outpainting by using the upper half to predict the lower half. + +Evaluation Metrics We define a Context Score $(CTS)$ to evaluate the context relationships among neighboring city blocks (i.e., a community). This metric is a numerical extension of LayoutSim [20] and DocSim index [46] which evaluate the similarity between two city blocks regarding location, size, and category. Specifically, for each generated city block $b_{i}$ , we calculate the average LayoutSim with its neighbors $N(i)$ . Then, the local average of LayoutSim represents the context similarity among the community $(CT)$ . High $CT$ indicates identical blocks in a community and low $CT$ corresponds to high diversity in a community. $CTS$ is the subtraction of the real community's $CT$ from the generated $CT$ : + +![](images/d0fdcdfcd0b9e6fb1c0ccc86b21a788ad0ef9b20effa5675ae16eb2ebd8bfd5e.jpg) +Fig. 5: Qualitative Comparisons. Given the same road network (except SDXL [48]), all above methods generate urban layouts in only one pass without post-processing or human-in-the-loop refinement. The "even rows" are a zoom-in of the highlighted areas in the "odd rows". Our method generates a realistic distribution of urban layouts with plausible context-dependent behaviors (as indicated by $CTS$ score). VTN [2] and LayoutDM [29] show abnormal style shift among neighboring blocks, and no awareness of road networks. GlobalMapper [20] generates over-similar communities. SDXL [48] provides poor outpainting quality containing semantic errors and undesired overlaps. + +Table 1: Quantitative Comparisons. All methods are compared to the same real urban layouts (except SDXL [48] which cannot take-in a road network). Best values are in bold, second best values are underlined. Our method outperforms other existing methods in all but the overlap metric. See text for an explanation of the metrics. + +
Method\(CTS_{x\rightarrow 0}\)WD-5D↓WD-CO↓Overlap↓O-Blk↓FID↓KID↓LPIPS↓
SDXL [48]-----120.240.0790.48
VTN [2]-1.143.185.811.247.3569.140.0470.32
LayoutDM [29]-2.202.9212.504.561.7266.770.0400.39
GlobalMP [20]0.624.774.142.520.6849.550.0240.34
Ours0.212.281.911.270.4223.630.0050.20
+ +$$ +C T = \frac {1}{\| N (i) \|} \sum_ {j} ^ {N (i)} \operatorname {L a y o u t S i m} \left(b _ {i}, b _ {j}\right), \quad N (i) = \{j \mid (i, j) \in E \} \tag {3} +$$ + +$$ +C T S = C T _ {g e n} - C T _ {r e a l} \tag {4} +$$ + +Given a real community, the metric indicates whether a generated community is over-diverse ( $CTS < 0$ ), or over-similar ( $CTS > 0$ ). A low value means more realism. + +In our comparisons, we compute several metrics. One metric is the Wasserstein Distance between the distributions of generated communities and real communities in terms of 2D position (x, y) and 2.5D geometry (length, width, height), and building count (CO) - we name these WD-5D and WD-CO. We define two quality metrics as well: Overlap indicates the percentage by which two building overlaps, and O-Blk represents the percentage of building area that sticks outside of its city block. We also implement several perceptual metrics: FID [24], KID [4], and LPIPS [65] to evaluate the visual quality and realism of generated layouts. For all the above metrics, lower value indicates better quality. + +Quantitative Results As Tab. 1 shows, our approach performs better than existing approaches in all but one metric. Note that the poor quality of SDXL results makes it hard to decipher layouts. We only compare to it by pixel-based perceptual metrics. + +Qualitative Results We show in Fig. 5 qualitative results for several exemplary areas (more examples are in Supplemental Material). In general, our method shows good contextual harmonization and has awareness of arbitrary road networks. It generates a realistic distribution of urban layouts with plausible context dependent behaviors. Other methods show improper patterns (e.g., overly diverse or similar) and potentially containing semantic errors or undesired overlaps. + +Table 2: Block Quantization Ablations. We report metrics (all in %) among all alternative ablations of our Block-level graph-based Variational AutoEncoder (BVAE). Using GAT-backbone with post-trained dimensional quantization with $L = 20$ shows the best overall compromise. The best values are in boldface. + +
Encode ModelOverlap↓Out-Blk↓Pos-E.↓Geom-E.↓Ct-E.↓Cov-E.↓
LayoutVAE [32]13.8411.1515.0937.35-24.34
VTN [2]2.704.706.0122.8618.581.97
BlockPlanner [62]2.931.304.7812.063.635.66
GlobalMapper [20]1.041.203.252.830.080.46
BVAE (SAGE [18])1.331.705.013.250.100.35
BVAE (GCN [33])0.451.214.252.740.040.28
BVAE* (GAT [58])0.170.281.000.710.0060.04
QuantizationOverlap↓Out-Blk↓Pos-E.↓Geom-E.↓Ct-E.↓Cov-E.↓
Trainable-VQ3.020.553.755.780.070.60
KMeans-Q2.900.0220.9817.8550.0116.23
DIM-Q (L=5)0.280.627.7116.2025.014.68
DIM-Q (L=10)0.560.285.734.488.334.58
DIM-Q (L=30)0.170.012.680.780.0050.15
DIM-Q* (L=20)0.110.013.500.890.0060.12
+ +# 4.3 Ablation Studies + +Block Quantization We ablate block quantization in order to understand and maximize the impact of changing its parameters and design strategy. First, we report in Tab. 2 (Encode Model Table) the reconstruction loss using different block-level variational autoencoders. We evaluate using various geometric metrics: Overlap (percent of building-to-building overlap), Out-Blk (percent of buildings that stick outside of its city block), Pos-E (positional error as a percentage of block diagonal), Geom-E (2.5D building geometry error as percentage of block scale), Ct-E (percentage of building count error), and Cov-E (total building coverage error). They clearly show our approach is the best overall. Thus we choose our BVAE based on Graph Attention Networks (GAT) [58]. + +Then, in Tab. 2 (Quantization Table), we ablate various codebook approaches and quantitization levels for block-level encoding. "Trainable-VQ" leverages trainable vector quantization codebooks introduced by VQVAE [56]. We also provide an alternative non-training based quantization method using KMeans ("KMeans-Q") to cluster BVAE latent space values to a reasonable number of categories in order to handle diverse urban layouts (e.g., 100 classes). The "DIM-Q" rows imply quantizing each of the 512 dimensions of the block-level code into a codebook $C = \{1,2,\dots,L\}$ as described in Sec. 3.1. This per-dimension quantization performs clearly the best. Specifically, setting $L$ to infinity produces a lossless representation of BVAE latent vectors, but also causes the curse of dimension- + +Table 3: City-Scale GMAE Ablations. We ablate backbone message passing layers, maximum messaging passing hops ( $D$ : depth of encoder), masking strategy during training, and scheduling functions of iterative generation ( $T$ indicates total iterations). The overall best setting is marked as *. + +
AblationsContext↓WD-3D↓WD-N↓Overlap↓Out-Blk↓FID↓KID↓LPIPS↓
GMAE (GCN [33])-0.302.742.910.260.3652.950.0250.32
GMAE (GTrans [52])0.743.023.271.961.0743.600.0180.29
GMAE (SAGE [18])0.972.982.361.752.6043.150.0170.29
GMAE (GAT, Dec)0.942.452.962.131.1048.240.0190.30
GMAE (GAT, D = 1)1.253.032.641.730.4443.750.0180.29
GMAE (GAT, D = 2)0.332.662.011.310.5734.820.0140.27
GMAE (GAT, D = 4)0.241.922.310.630.2824.590.0090.22
GMAE (Mask ∈ [0.05, 1])1.302.703.621.702.4543.140.0180.29
GMAE (Mask ∈ [0.05, 0.5])1.103.073.681.681.3046.850.0200.30
GMAE (Mask = 0.15)1.594.913.411.831.2049.500.0220.31
GMAE (Linear, T = 12)0.632.833.791.270.9337.420.0150.28
GMAE (Log, T = 12)-1.312.777.277.413.1463.250.0380.33
GMAE (Cosine, T = 1)-1.677.445.643.564.0176.120.0510.35
GMAE (Cosine, T = 20)0.322.923.981.680.3132.060.0130.24
*(GAT-D3, [0.5,1], Cos-T12)0.212.281.911.270.4223.630.0050.20
+ +ality. We discover that $L = 20$ per dimension is the best overall compromise balancing compactness and representation ability. + +City-Scale Graph Masked Autoencoder. In Tab. 3, We ablate various model parameters, together with significant training and inference strategies for our GMAE. Each alternatives is evaluated by the same set of metrics in Sec. 4.2. + +By varying the message passing layers, we find GAT [58] provides the best performance. GCN [33] produces diverse and unrealistic urban layouts, while transformer-based [52], and GraphSAGE [18] backbones generate blocks that are too similar. The depth of the stacked layers in the encoder (D) dictates the maximum message passing hops from each node and thus the span of context behaviors captured for realistic generation. We find that 3 hops is promising and larger contexts cause over-similarity and unnecessary computing burden. We also observed that using stacked GAT layers for decoding with remasking tricks, as described by [25,26], didn't produce a net benefit towards city generation. So we choose MLP as the node feature decoder for both its performance and efficiency. + +The masking strategy during training is crucial for GMAE performance. We find that dynamic and large masking ratios are beneficial, while a fixed small ratio (e.g., 0.15 as in [14]) leads to over similarity. This behavior is also observed by Li et al. [35]. We adopt a similar masking strategy by sampling a Gaussian distribution $N(0.55, 0.25)$ , truncated by [0.5, 1.0], for each training iteration. + +Scheduling for iterative generation is also critical to capture representative city blocks (Sec. 3.3). We evaluate several functions $\beta(t)$ (e.g., linear, logarithmic, or cosine-based) that provide the scheduled acceptance ratio, and evaluate alternative total iteration counts $T$ . Rapid iteration leads to over-diversity and + +long iteration results in over-similarity. Logarithmic functions with reverse pattern produces the worst realism. Metrics show that cosine-based function, which is conservative initially and generous in later iterations, outperforms others. + +In summary, this study determines the best configuration for our GMAE as: GAT encoder with $D = 3$ and MLP decoder, training mask ratio $m \in [0.5,1.0]$ , and priority-based scheduling function $\beta (t) = 1 - \cos (t / T)$ with $T = 12$ . + +# 4.4 Limitations + +Our method faces limitations where shapes of city blocks or buildings are not simple polygons (e.g. the building with a garden inside). It also struggles to handle very irregular and concave shaped city blocks (e.g., cul-de-sac's). Also, road-network to contextual structure relations are not explicitly considered. + +# 4.5 Additional Results + +We place several other interesting results in Supplementary Materials. + +- Controllable Generation of Entire City. Our method enables realistic and efficient city-scale controllable generation given any percentage of priors. We show several large examples spanning a subset of the many cities. +- Socio-Economic Prediction. Trained GMAE may act as a powerful encoder for downstream tasks. For example, the encoded GMAE latent vectors of urban layouts can indicate social-economic metrics for each city block. +- Semantic Manipulations. Our method enables semantic manipulations over various building layout styles across different cities. + +# 5 Conclusions and Future Works + +Large-scale urban layout generation has seen significant advancements, spurred by the intersection of computer vision, urban planning, and related disciplines. Previous methodologies often neglect the context-sensitive nature of urban layouts, failing to capture the multi-layer semantics crucial for realism and plausibility. Our proposed Graph-based Masked AutoEncoder (GMAE) addresses these limitations by leveraging a graph-based representation, acknowledging context-sensitivity, and prioritizing the generation process. By training on a diverse dataset of 330 cities, we demonstrate the efficacy of our approach in generating large-scale 2.5D layouts with realism and semantic consistency. Moreover, our method outperforms existing techniques in various metrics, marking a significant advancement in the realm of urban layout generation. + +Moving forward, our approach paves the way for more accurate and efficient urban simulation, digital twin creation, and game/content design. As future work we intend to incorporate our method for large-scale photo-realistic multi-view scene synthesis, and city-scale 3D modeling. Those potentials will contribute to synthetic data generation in autonomous driving and world model training. + +# Acknowledgements + +This project was funded in part by NSF Grant #2107096 and NSF Grant #1835739. + +# References + +1. Climate and economic justice screening tool. https://screeningtool.geoplatform.gov/ +2. Arroyo, D.M., Postels, J., Tombari, F.: Variational transformer networks for layout generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13642-13652 (2021) +3. Bhatt, M., Kalyanam, R., Nishida, G., He, L., May, C., Niyogi, D., Aliaga, D.: Design and deployment of photo2building: A cloud-based procedural modeling tool as a service. In: Practice and Experience in Advanced Research Computing, pp. 132-138 (2020) +4. Binkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying mmd gans. arXiv preprint arXiv:1801.01401 (2018) +5. Bokeloh, M., Wand, M., Seidel, H.P.: A connection between partial symmetry and inverse procedural modeling. In: ACM SIGGRAPH 2010 papers, pp. 1-10 (2010) +6. Brooks, T., Peebles, B., Holmes, C., DePue, W., Guo, Y., Jing, L., Schnurr, D., Taylor, J., Luhman, T., Luhman, E., Ng, C., Wang, R., Ramesh, A.: Video generation models as world simulators (2024), https://openai.com/research/video-generation-models-as-world-simulators +7. Bureau, U.S.C.: Topologically integrated geographic encoding and referencing. https://www.census.gov/geographies/mapping-files/time-series/geo/tiger-line-file.html +8. Chai, L., Tucker, R., Li, Z., Isola, P., Snavely, N.: Persistent nature: A generative model of unbounded 3d worlds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20863-20874 (2023) +9. Chai, S., Zhuang, L., Yan, F.: Layoutdm: Transformer-based diffusion model for layout generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18349-18358 (2023) +0. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked generative image transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11315-11325 (2022) +1. Chang, K.H., Cheng, C.Y., Luo, J., Murata, S., Nourbakhsh, M., Tsuji, Y.: Building-gan: Graph-conditioned architectural volumetric design generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11956-11965 (2021) +2. Chen, Z., Wang, G., Liu, Z.: Scenedreamer: Unbounded 3d scene generation from 2d image collections. arXiv preprint arXiv:2302.01330 (2023) +3. Deng, J., Chai, W., Guo, J., Huang, Q., Hu, W., Hwang, J.N., Wang, G.: Citygen: Infinite and controllable 3d city layout generation. arXiv preprint arXiv:2312.01508 (2023) +4. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) + +15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +16. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873-12883 (2021) +17. Gupta, K., Lazarow, J., Achille, A., Davis, L.S., Mahadevan, V., Shrivastava, A.: Layouttransformer: Layout generation and completion with self-attention. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1004-1014 (2021) +18. Hamilton, W., Ying, Z., Leskovec, J.: Inductive representation learning on large graphs. Advances in neural information processing systems 30 (2017) +19. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022) +20. He, L., Aliaga, D.: Globalmapper: Arbitrary-shaped urban layout generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 454-464 (2023) +21. He, L., Lu, Y., Corring, J., Florencio, D., Zhang, C.: Diffusion-based document layout generation. In: Fink, G.A., Jain, R., Kise, K., Zanibbi, R. (eds.) Document Analysis and Recognition - ICDAR 2023. pp. 361-378. Springer Nature Switzerland, Cham (2023) +22. He, L., Shan, J., Aliaga, D.: Generative building feature estimation from satellite images. IEEE Transactions on Geoscience and Remote Sensing 61, 1-13 (2023) +23. Heris, M.P., Foks, N.L., Bagstad, K.J., Troy, A., Ancona, Z.H.: A rasterized building footprint dataset for the united states. Scientific data 7(1), 207 (2020) +24. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017) +25. Hou, Z., He, Y., Cen, Y., Liu, X., Dong, Y., Kharlamov, E., Tang, J.: Graphmae2: A decoding-enhanced masked self-supervised graph learner. In: Proceedings of the ACM Web Conference 2023. pp. 737-746 (2023) +26. Hou, Z., Liu, X., Cen, Y., Dong, Y., Yang, H., Wang, C., Tang, J.: Graphmae: Self-supervised masked graph autoencoders. In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 594-604 (2022) +27. Hua, H., Shi, J., Kafle, K., Jenni, S., Zhang, D., Collomosse, J., Cohen, S., Luo, J.: Finematch: Aspect-based fine-grained image and text mismatch detection and correction. arXiv preprint arXiv:2404.14715 (2024) +28. Hui, M., Zhang, Z., Zhang, X., Xie, W., Wang, Y., Lu, Y.: Unifying layout generation with a decoupled diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1942-1951 (2023) +29. Inoue, N., Kikuchi, K., Simo-Serra, E., Otani, M., Yamaguchi, K.: Layoutdm: Discrete diffusion model for controllable layout generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10167-10176 (2023) +30. Jiang, Z., Guo, J., Sun, S., Deng, H., Wu, Z., Mijovic, V., Yang, Z.J., Lou, J.G., Zhang, D.: Layoutformer++: Conditional graphic layout generation via constraint serialization and decoding space restriction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18403-18412 (2023) + +31. Jyothi, A.A., Durand, T., He, J., Sigal, L., Mori, G.: Layoutvae: Stochastic scene layout generation from a label set. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9895-9904 (2019) +32. Jyothi, A.A., Durand, T., He, J., Sigal, L., Mori, G.: Layoutvae: Stochastic scene layout generation from a label set. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9894-9903 (2019). https://doi.org/10.1109/ICCV.2019.00999 +33. Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907 (2016) +34. Li, J., Yang, J., Hertzmann, A., Zhang, J., Xu, T.: Layoutgan: Synthesizing graphic layouts with vector-wireframe adversarial networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(7), 2388-2399 (2020) +35. Li, T., Chang, H., Mishra, S., Zhang, H., Katabi, D., Krishnan, D.: Mage: Masked generative encoder to unify representation learning and image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2142-2152 (2023) +36. Li, Z., Wang, Q., Snavely, N., Kanazawa, A.: Infinitenature-zero: Learning perpetual view generation of natural scenes from single images. In: European Conference on Computer Vision. pp. 515-534. Springer (2022) +37. Lin, C.H., Lee, H.Y., Cheng, Y.C., Tulyakov, S., Yang, M.H.: Infinitygan: Towards infinite-pixel image synthesis. arXiv preprint arXiv:2104.03963 (2021) +38. Lin, C.H., Lee, H.Y., Menapace, W., Chai, M., Siarohin, A., Yang, M.H., Tulyakov, S.: Infinity: Infinite-scale city synthesis. arXiv preprint arXiv:2301.09637 (2023) +39. Lipp, M., Scherzer, D., Wonka, P., Wimmer, M.: Interactive modeling of city layouts using layers of procedural content. In: Computer Graphics Forum. vol. 30, pp. 345-354. Wiley Online Library (2011) +40. Ma, H., Zeng, D., Liu, Y.: Learning individualized treatment rules with many treatments: A supervised clustering approach using adaptive fusion. Advances in Neural Information Processing Systems 35, 15956-15969 (2022) +41. Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781 (2013) +42. Nauata, N., Chang, K.H., Cheng, C.Y., Mori, G., Furukawa, Y.: House-gan: Relational generative adversarial networks for graph-constrained house layout generation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 162-177. Springer (2020) +43. OpenStreetMap contributors: Planet dump retrieved from https://planet.osm.org. https://www.openstreetmap.org (2017) +44. Para, W., Guerrero, P., Kelly, T., Guibas, L.J., Wonka, P.: Generative layout modeling using constraint graphs. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6690-6700 (2021) +45. Patel, P., Kalyanam, R., He, L., Aliaga, D., Niyogi, D.: Deep learning-based urban morphology for city-scale environmental modeling. PNAS nexus 2(3), pgad027 (2023) +46. Patil, A.G., Ben-Eliezer, O., Perel, O., Averbuch-Elor, H.: Read: Recursive autoencoders for document layout generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 544-545 (2020) +47. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4195-4205 (2023) + +48. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023) +49. Shen, Y., Ma, W.C., Wang, S.: Sgam: Building a virtual 3d world through simultaneous generation and mapping. Advances in Neural Information Processing Systems 35, 22090-22102 (2022) +50. Sheng, Y., Liu, Y., Zhang, J., Yin, W., Oztireli, A.C., Zhang, H., Lin, Z., Shechtman, E., Benes, B.: Controllable shadow generation using pixel height maps. In: European Conference on Computer Vision. pp. 240-256. Springer (2022) +51. Sheng, Y., Yu, Z., Ling, L., Cao, Z., Zhang, X., Lu, X., Xian, K., Lin, H., Benes, B.: Dr. bokeh: Differentiable occlusion-aware bokeh rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4515-4525 (2024) +52. Shi, Y., Huang, Z., Feng, S., Zhong, H., Wang, W., Sun, Y.: Masked label prediction: Unified message passing model for semi-supervised classification. arXiv preprint arXiv:2009.03509 (2020) +53. Song, Y., Zhang, Z., Lin, Z., Cohen, S., Price, B., Zhang, J., Kim, S.Y., Aliaga, D.: Objectstitch: Object compositing with diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18310-18319 (2023) +54. Song, Y., Zhang, Z., Lin, Z., Cohen, S., Price, B., Zhang, J., Kim, S.Y., Zhang, H., Xiong, W., Aliaga, D.: Imprint: Generative object compositing by learning identity-preserving representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8048-8058 (2024) +55. Tabata, S., Yoshihara, H., Maeda, H., Yokoyama, K.: Automatic layout generation for graphical design magazines. In: ACM SIGGRAPH 2019 Posters, pp. 1-2 (2019) +56. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017) +57. Vanegas, C.A., Kelly, T., Weber, B., Halatsch, J., Aliaga, D.G., Müller, P.: Procedural generation of parcels in urban modeling. In: Computer graphics forum. vol. 31, pp. 681-690. Wiley Online Library (2012) +58. Velicković, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y.: Graph attention networks. arXiv preprint arXiv:1710.10903 (2017) +59. Velickovic, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y., et al.: Graph attention networks. stat 1050(20), 10-48550 (2017) +60. Wu, W., Fu, X.M., Tang, R., Wang, Y., Qi, Y.H., Liu, L.: Data-driven interior plan generation for residential buildings. ACM Transactions on Graphics (TOG) 38(6), 1-12 (2019) +61. Xie, H., Chen, Z., Hong, F., Liu, Z.: Citydreamer: Compositional generative model of unbounded 3d cities. arXiv preprint arXiv:2309.00610 (2023) +62. Xu, L., Xiangli, Y., Rao, A., Zhao, N., Dai, B., Liu, Z., Lin, D.: Blockplanner: City block generation with vectorized graph representation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5077-5086 (2021) +63. Yan, W., Zhang, Y., Abbeel, P., Srinivas, A.: Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157 (2021) +64. Yang, C.F., Fan, W.C., Yang, F.E., Wang, Y.C.F.: Layouttransformer: Scene layout generation with conceptual and spatial diversity. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3732-3741 (2021) +65. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018) + +66. Zhang, X., Ma, W., Varinlioglu, G., Rauh, N., He, L., Aliaga, D.: Guided pluralistic building contour completion. The Visual Computer 38(9), 3205-3216 (2022) +67. Zheng, X., Qiao, X., Cao, Y., Lau, R.W.: Content-aware generative modeling of graphic design layouts. ACM Transactions on Graphics (TOG) 38(4), 1-15 (2019) \ No newline at end of file diff --git a/2024/COHO_ Context-Sensitive City-Scale Hierarchical Urban Layout Generation/images.zip b/2024/COHO_ Context-Sensitive City-Scale Hierarchical Urban Layout Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..66d0262f69be4e621710f2ed6bd5b42e33da5abb --- /dev/null +++ b/2024/COHO_ Context-Sensitive City-Scale Hierarchical Urban Layout Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e92fb3cc9b2bd0bd47e6685798c834396ce7bb029099e644cc0a83db7a2e00b +size 807318 diff --git a/2024/COHO_ Context-Sensitive City-Scale Hierarchical Urban Layout Generation/layout.json b/2024/COHO_ Context-Sensitive City-Scale Hierarchical Urban Layout Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0f62f034a47e6bba6ca51f62e9799d44c01b7f68 --- /dev/null +++ b/2024/COHO_ Context-Sensitive City-Scale Hierarchical Urban Layout Generation/layout.json @@ -0,0 +1,10136 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "text", + "content": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 250, + 168, + 364, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 168, + 364, + 181 + ], + "spans": [ + { + "bbox": [ + 250, + 168, + 364, + 181 + ], + "type": "text", + "content": "Liu He, Daniel Aliaga" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 256, + 190, + 357, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 190, + 357, + 201 + ], + "spans": [ + { + "bbox": [ + 256, + 190, + 357, + 201 + ], + "type": "text", + "content": "Purdue University, USA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 247, + 201, + 367, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 201, + 367, + 213 + ], + "spans": [ + { + "bbox": [ + 247, + 201, + 367, + 213 + ], + "type": "text", + "content": "{he425,aliaga}@purdue.edu" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 159, + 239, + 258, + 330 + ], + "blocks": [ + { + "bbox": [ + 159, + 239, + 258, + 330 + ], + "lines": [ + { + "bbox": [ + 159, + 239, + 258, + 330 + ], + "spans": [ + { + "bbox": [ + 159, + 239, + 258, + 330 + ], + "type": "image", + "image_path": "eeb1e4304f82043ce7508d9d636bb723fd83c5cac48bbe79bb742ab17b5baafa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 172, + 331, + 250, + 339 + ], + "lines": [ + { + "bbox": [ + 172, + 331, + 250, + 339 + ], + "spans": [ + { + "bbox": [ + 172, + 331, + 250, + 339 + ], + "type": "text", + "content": "LayoutDM " + }, + { + "bbox": [ + 172, + 331, + 250, + 339 + ], + "type": "inline_equation", + "content": "(CTS = -2.20)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 265, + 239, + 356, + 331 + ], + "blocks": [ + { + "bbox": [ + 265, + 239, + 356, + 331 + ], + "lines": [ + { + "bbox": [ + 265, + 239, + 356, + 331 + ], + "spans": [ + { + "bbox": [ + 265, + 239, + 356, + 331 + ], + "type": "image", + "image_path": "36b085177dada633a268f3c5ddb070f50057286d0b5db2990542f7f062b97496.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 331, + 337, + 339 + ], + "lines": [ + { + "bbox": [ + 282, + 331, + 337, + 339 + ], + "spans": [ + { + "bbox": [ + 282, + 331, + 337, + 339 + ], + "type": "text", + "content": "Ours " + }, + { + "bbox": [ + 282, + 331, + 337, + 339 + ], + "type": "inline_equation", + "content": "(CTS = 0.21)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 363, + 240, + 454, + 331 + ], + "blocks": [ + { + "bbox": [ + 363, + 240, + 454, + 331 + ], + "lines": [ + { + "bbox": [ + 363, + 240, + 454, + 331 + ], + "spans": [ + { + "bbox": [ + 363, + 240, + 454, + 331 + ], + "type": "image", + "image_path": "ce3753b00c29ac292e17d8c267271a185b6fa36ae11a58b86b5899746e1ebdbf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 331, + 451, + 340 + ], + "lines": [ + { + "bbox": [ + 367, + 331, + 451, + 340 + ], + "spans": [ + { + "bbox": [ + 367, + 331, + 451, + 340 + ], + "type": "text", + "content": "GlobalMapper " + }, + { + "bbox": [ + 367, + 331, + 451, + 340 + ], + "type": "inline_equation", + "content": "(CTS = 0.62)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 159, + 341, + 258, + 431 + ], + "blocks": [ + { + "bbox": [ + 159, + 341, + 258, + 431 + ], + "lines": [ + { + "bbox": [ + 159, + 341, + 258, + 431 + ], + "spans": [ + { + "bbox": [ + 159, + 341, + 258, + 431 + ], + "type": "image", + "image_path": "ca21aa3b26064b52762e7dc614eaae7900fb1e159837833012992d80dc7ee4f1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 433, + 241, + 441 + ], + "lines": [ + { + "bbox": [ + 182, + 433, + 241, + 441 + ], + "spans": [ + { + "bbox": [ + 182, + 433, + 241, + 441 + ], + "type": "text", + "content": "Random " + }, + { + "bbox": [ + 182, + 433, + 241, + 441 + ], + "type": "inline_equation", + "content": "(CTS\\ll 0)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 265, + 341, + 356, + 431 + ], + "blocks": [ + { + "bbox": [ + 265, + 341, + 356, + 431 + ], + "lines": [ + { + "bbox": [ + 265, + 341, + 356, + 431 + ], + "spans": [ + { + "bbox": [ + 265, + 341, + 356, + 431 + ], + "type": "image", + "image_path": "edde6205d9b52d281d179b0fe6c2f5c40ad8b41e31b4c6187bf9c2acd950add2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 433, + 343, + 441 + ], + "lines": [ + { + "bbox": [ + 282, + 433, + 343, + 441 + ], + "spans": [ + { + "bbox": [ + 282, + 433, + 343, + 441 + ], + "type": "text", + "content": "Real Data " + }, + { + "bbox": [ + 282, + 433, + 343, + 441 + ], + "type": "inline_equation", + "content": "(CTS\\rightarrow 0)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 363, + 342, + 454, + 432 + ], + "blocks": [ + { + "bbox": [ + 363, + 342, + 454, + 432 + ], + "lines": [ + { + "bbox": [ + 363, + 342, + 454, + 432 + ], + "spans": [ + { + "bbox": [ + 363, + 342, + 454, + 432 + ], + "type": "image", + "image_path": "1b121afc5408cfebba0a8399974f27efb816b208138162c225ec21c17054eef9.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 379, + 433, + 438, + 441 + ], + "lines": [ + { + "bbox": [ + 379, + 433, + 438, + 441 + ], + "spans": [ + { + "bbox": [ + 379, + 433, + 438, + 441 + ], + "type": "text", + "content": "Identical " + }, + { + "bbox": [ + 379, + 433, + 438, + 441 + ], + "type": "inline_equation", + "content": "(CTS\\gg 0)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 450, + 482, + 506 + ], + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 506 + ], + "type": "text", + "content": "Fig. 1: Context-Sensitive Generation. Our method (Green) pursues realistic context harmonization among neighboring city blocks as real data (bottom-middle). Other methods (e.g. LayoutDM [29], GlobalMapper [20]) show over-diversity/-similarity in city-scale layout generation (evaluated by Context Score " + }, + { + "bbox": [ + 130, + 450, + 482, + 506 + ], + "type": "inline_equation", + "content": "CTS" + }, + { + "bbox": [ + 130, + 450, + 482, + 506 + ], + "type": "text", + "content": " as Eq. 4.). Fully random and identical layouts are synthetically generated to illustrate extreme cases." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 159, + 544, + 455, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 544, + 455, + 666 + ], + "spans": [ + { + "bbox": [ + 159, + 544, + 455, + 666 + ], + "type": "text", + "content": "Abstract. The generation of large-scale urban layouts has garnered substantial interest across various disciplines. Prior methods have utilized procedural generation requiring manual rule coding or deep learning needing abundant data. However, prior approaches have not considered the context-sensitive nature of urban layout generation. Our approach addresses this gap by leveraging a canonical graph representation for the entire city, which facilitates scalability and captures the multilayer semantics inherent in urban layouts. We introduce a novel graph-based masked autoencoder (GMAE) for city-scale urban layout generation. The method encodes attributed buildings, city blocks, communities and cities into a unified graph structure, enabling self-supervised" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 117, + 455, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 117, + 455, + 183 + ], + "spans": [ + { + "bbox": [ + 159, + 117, + 455, + 183 + ], + "type": "text", + "content": "masked training for graph autoencoder. Additionally, we employ scheduled iterative sampling for 2.5D layout generation, prioritizing the generation of important city blocks and buildings. Our approach achieves good realism, semantic consistency, and correctness across the heterogeneous urban styles in 330 US cities. Codes and datasets are released at https://github.com/Arking1995/COHO." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 194, + 453, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 194, + 453, + 216 + ], + "spans": [ + { + "bbox": [ + 159, + 194, + 453, + 216 + ], + "type": "text", + "content": "Keywords: Layout Generation " + }, + { + "bbox": [ + 159, + 194, + 453, + 216 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 194, + 453, + 216 + ], + "type": "text", + "content": " Urban Modeling " + }, + { + "bbox": [ + 159, + 194, + 453, + 216 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 194, + 453, + 216 + ], + "type": "text", + "content": " Masked Graph Autoencoder " + }, + { + "bbox": [ + 159, + 194, + 453, + 216 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 194, + 453, + 216 + ], + "type": "text", + "content": " Context Sensitivity" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 236, + 230, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 230, + 249 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 230, + 249 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 262, + 481, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 481, + 346 + ], + "type": "text", + "content": "Large-scale urban layout generation has received significant interest by computer vision, urban planning and related disciplines. In particular, cities easily have 1,000 to 50,000 city blocks spanning a mostly horizontal terrain. The configuration of each city block varies significantly across the city, yet there is some configuration similarity amongst subsets within and among cities. Reconstructing existing layouts and generating future planned layouts are crucial tasks for urban simulation, digital twins, and game/content design." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 346, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 525 + ], + "type": "text", + "content": "Prior works have generated urban layouts from a variety of sources. Procedural layout generation [39, 57] used hand-coded rules to generate cities, and inverse procedural modeling [5] extracted the rules from data and then enabled generation. More recently, deep learning has promoted data-driven approaches that produce either pixel-based layout generation (e.g., InfiniCity [38], CityDreamer [61], CityGEN [13]) or graph-based approaches for general polygonal shape layouts (e.g., LayoutTransformer [17], VTN [2]) and for urban scenarios (e.g., BlockPlanner [62], GlobalMapper [20], BuildingGAN [11]). However, none of these techniques explicitly recover and consider the context-sensitive nature of urban layout generation (Fig. 1). In other words, the layout of buildings, blocks, and communities cannot be treated in isolation. The multi-layer semantics are critical to generating realistic and plausible urban layouts. While such multilevel structure (e.g., from single building to a community or subset of a city) has been specified via hand-coded rules, it has not been discovered, organized and used for generation in data-driven deep learning approaches." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 146, + 525, + 357, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 525, + 357, + 537 + ], + "spans": [ + { + "bbox": [ + 146, + 525, + 357, + 537 + ], + "type": "text", + "content": "Our approach builds on three key observations:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 135, + 545, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 136, + 545, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 545, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 136, + 545, + 481, + 629 + ], + "type": "text", + "content": "1. Representation: Cities, communities, blocks, buildings, and roads are arbitrarily shaped and with complex topology. Hence, a graph-based representation can better capture the regular and non-regular configurations, as opposed to the limiting regular structure of an image. Further, graphs can be more compact which in turn supports better scalability. Altogether, having buildings or city blocks as atomic units, instead of pixels, will contribute to improving correctness, realism, and consistency with priors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 135, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 630, + 481, + 665 + ], + "type": "text", + "content": "2. Context-Sensitivity: Buildings, city blocks, and communities are built considering their neighboring structures and not in isolation. Hence, a generator should consider this multi-layer structure and inter-dependency. In" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "type": "text", + "content": "He, L. & Aliaga, D." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 147, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 479, + 140 + ], + "type": "text", + "content": "order words, the style features at one level (e.g., size, shape, location) depend on the surrounding context." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 140, + 480, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 140, + 480, + 210 + ], + "spans": [ + { + "bbox": [ + 136, + 140, + 480, + 210 + ], + "type": "text", + "content": "3. Prioritization: The stylistic and semantic importance, or priority, of city blocks and buildings varies. This implies that a generation scheme should generate the most important city blocks and buildings first and then fill in the rest, as opposed to a region growing approach for example. Overall, prioritization will improve correctness and realism as well as enable large-scale urban layout generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 217, + 479, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 479, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 479, + 312 + ], + "type": "text", + "content": "We propose a novel Graph-based Masked AutoEncoder (GMAE) for large-scale multi-layer 2.5D urban layout generation. First, we define a canonical graph-based representation capturing the multi-layer context sensitivity of any city's urban layouts (Fig. 2). The entire graph " + }, + { + "bbox": [ + 130, + 217, + 479, + 312 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 217, + 479, + 312 + ], + "type": "text", + "content": " corresponds to a city. Subsets of the graph map to communities and each node " + }, + { + "bbox": [ + 130, + 217, + 479, + 312 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 217, + 479, + 312 + ], + "type": "text", + "content": " amounts to a city block. Further, each node encodes its building layouts, shapes, and heights as a quantized feature by well-trained quantizor. Second, we utilize our GMAE to learn the multi-layer stylistic behavior of many urban layouts (i.e., 330 cities)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 314, + 480, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 480, + 408 + ], + "type": "text", + "content": "Using subsets of the graph (or communities), node features are randomly masked and self-supervised GMAE training is performed to learn node feature reconstruction (Fig. 3). Third, utilizing the well-trained GMAE as a generator, we perform an iterative priority-based scheduling to generate a large layout or to complete a city fragment; i.e., each iteration predicts node features for the entire graph but only the most confident ones are preserved for the next step (Fig. 4). This methodology is able to leverage any percentage of given prior (e.g., " + }, + { + "bbox": [ + 130, + 314, + 480, + 408 + ], + "type": "inline_equation", + "content": "[0,100\\%]" + }, + { + "bbox": [ + 130, + 314, + 480, + 408 + ], + "type": "text", + "content": ") for urban layout completion and generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 409, + 480, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 409, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 480, + 503 + ], + "type": "text", + "content": "To the best of our knowledge, our approach is the first data-driven deep-learning method to enable large-scale 2.5D layout generation with good realism, semantic consistency, and correctness. We show the capability to generalize to the heterogeneous urban layout styles of 330 cities in the US (e.g., all cities with a population " + }, + { + "bbox": [ + 130, + 409, + 480, + 503 + ], + "type": "inline_equation", + "content": ">100\\mathrm{K}" + }, + { + "bbox": [ + 130, + 409, + 480, + 503 + ], + "type": "text", + "content": ", totaling 833,473 city blocks and 17,663,607 buildings). In addition, we provide comparisons to several prior works including SDXL [48]), VTN [2], LayoutDM [29], and GlobalMapper [20] in Sec. 4 and exceed their performance in multiple well-known geometric and perceptual metrics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 146, + 505, + 324, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 505, + 324, + 517 + ], + "spans": [ + { + "bbox": [ + 146, + 505, + 324, + 517 + ], + "type": "text", + "content": "Our contributions include the following:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 522, + 479, + 663 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 138, + 522, + 479, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 479, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 479, + 555 + ], + "type": "text", + "content": "- A canonical graph representation for large-scale 2.5D urban layouts. It encodes global and local node features and neighboring relations for buildings, city blocks, and communities within a city." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 559, + 479, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 559, + 479, + 590 + ], + "spans": [ + { + "bbox": [ + 138, + 559, + 479, + 590 + ], + "type": "text", + "content": "- A self-supervised Graph-based Masked AutoEncoder (GMAE) enabling arbitrarily shaped city layout generation with consistency, correctness, and realism" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 594, + 479, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 594, + 479, + 627 + ], + "spans": [ + { + "bbox": [ + 138, + 594, + 479, + 627 + ], + "type": "text", + "content": "- Priority-based scheduled iterative sampling for controllable urban layout synthesis including completion, refinement, and generation with any percentage of priors ([0, 100%])." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 630, + 479, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 630, + 479, + 663 + ], + "spans": [ + { + "bbox": [ + 138, + 630, + 479, + 663 + ], + "type": "text", + "content": "- A open dataset of road network and urban layouts of 330 US cities for benchmarking current and future urban layout generation methods (https://huggingface.co/datasets/Arking95/COHO)." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 90, + 447, + 102 + ], + "type": "text", + "content": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 345 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 345 + ], + "type": "text", + "content": "Layout Generation. Numerous data-driven deep learning methods have been proposed for layout generation; for example, generalized layout synthesis [2,17, 31,34,64], document layouts [21,30,46,55,67], urban land lots [22,62,66], indoor scenes [42,44,60], and 3D buildings [3,11,45]. Recently, diffusion models have also been used for layout generation tasks (e.g., [9,21,28,29]). However, most of these works assume 2D layouts generated on a rectangular empty canvas. The position and size of each bounding box is often limited to a finite set of categories and to rectilinear alignments. This results in poor diversity and low realism for urban layout generation tasks. BlockPlanner [62] and GlobalMapper [20] adapt layout generation to either rectilinear or arbitrarily-shaped city block generation. However, all previous works assume each layout, or city block, is independent. This is not the case in urban environments where adjacent and nearby city blocks may share a common style or be positioned in a purposeful multi-block arrangement. Without a contextual guidance, city block generation may result in either unwanted harmonization or unexpected diversity between neighboring blocks (see Fig. 1). Thus, we pursue city-scale urban layout generation with the help of context-aware neighboring information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 367, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 482, + 536 + ], + "type": "text", + "content": "Infinite Visual Synthesis. With bombing of visual deep learning works [27, 40, 50, 51, 53, 54], researchers have explored synthesizing large, potentially unbounded images. For example, omnidirectional image synthesis from multiple smaller field-of-view images [36, 37]. Other works focus on unbounded 3D scene generation [8, 12, 49]. Particularly relevant are several recent infinite urban scene generators [13, 38, 61] that directly adapt image synthesis methods (e.g. InfinityGAN [37], MaskGIT [10]). The synthesis performs an autoregressive outpainting process using layout image patches. However, patch-based training and inference doesn't learn global features nor their interdependencies. Rather, it learns a very local behavior. Prior pixel-based methods acknowledge limited realism and semantics [13, 38, 61]. Often, post-processing [13, 61], and human-in-the-loop editing [38] is required to refine the generated images to vector-based urban layouts. Even the latest stable diffusion model SDXL [48] seeks high-resolution image synthesis but may not produce realistic urban layouts (Fig. 5)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Learning based on Quantized Representations. Vector quantization has been a prevailing stepping-stone technique for text (e.g., Word2Vec [41]), image (e.g., VQVAE/GAN [16, 56], [15]), and video synthesis (e.g., SORA [6, 47], [63]). Pretrained quantizers provide a scalable token representation benefiting high-level and large-scale synthesis. A well-trained quantizer combined with self-supervised masked training often performs well in reconstructing the original signal during generative tasks. BERT [14] provides plausible text-based representation learning. MAE [19] defines a masked autoencoder for image space which excels at multiple downstream tasks. MAE also provides representation" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "type": "text", + "content": "He, L. & Aliaga, D." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "learning for a broad range of node/graph feature synthesis [25, 26]. Specifically, MaskGIT [10] and MAGE [35] utilize MAE frameworks for impressive image generation tasks. The idea of a generative masked autoencoder based on pretrained quantization was inspirational to our city-scale urban layout approach." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 182, + 202, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 182, + 202, + 196 + ], + "spans": [ + { + "bbox": [ + 131, + 182, + 202, + 196 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 209, + 480, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 209, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 480, + 258 + ], + "type": "text", + "content": "First we describe our canonical graph representation of 2.5D urban layouts (Fig. 2). Second, we explain the training process for our GMAE based on the graphs for many cities (Fig. 3). Third, we show our priority-based scheduling for iteratively generating city-scale urban layouts given any amount of prior (Fig. 4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 275, + 326, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 275, + 326, + 288 + ], + "spans": [ + { + "bbox": [ + 131, + 275, + 326, + 288 + ], + "type": "text", + "content": "3.1 Canonical Graph Representation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 296, + 480, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 480, + 356 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 480, + 356 + ], + "type": "text", + "content": "A city is represented by a graph " + }, + { + "bbox": [ + 130, + 296, + 480, + 356 + ], + "type": "inline_equation", + "content": "G = \\{B, E\\}" + }, + { + "bbox": [ + 130, + 296, + 480, + 356 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 296, + 480, + 356 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 296, + 480, + 356 + ], + "type": "text", + "content": " is a set of nodes " + }, + { + "bbox": [ + 130, + 296, + 480, + 356 + ], + "type": "inline_equation", + "content": "B = \\{b_i | i \\in [1, N]\\}" + }, + { + "bbox": [ + 130, + 296, + 480, + 356 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 130, + 296, + 480, + 356 + ], + "type": "inline_equation", + "content": "b_i" + }, + { + "bbox": [ + 130, + 296, + 480, + 356 + ], + "type": "text", + "content": " represents each city block (Fig. 2), and any connected subgraph of " + }, + { + "bbox": [ + 130, + 296, + 480, + 356 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 296, + 480, + 356 + ], + "type": "text", + "content": " corresponds to a community. " + }, + { + "bbox": [ + 130, + 296, + 480, + 356 + ], + "type": "inline_equation", + "content": "N = |B|" + }, + { + "bbox": [ + 130, + 296, + 480, + 356 + ], + "type": "text", + "content": " is the total number of blocks for a given city. " + }, + { + "bbox": [ + 130, + 296, + 480, + 356 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 296, + 480, + 356 + ], + "type": "text", + "content": " is a set of edges " + }, + { + "bbox": [ + 130, + 296, + 480, + 356 + ], + "type": "inline_equation", + "content": "E = \\{e_{ij} | i, j \\in [1, N]\\}" + }, + { + "bbox": [ + 130, + 296, + 480, + 356 + ], + "type": "text", + "content": " representing mutually adjacent nodes (or city blocks). " + }, + { + "bbox": [ + 130, + 296, + 480, + 356 + ], + "type": "inline_equation", + "content": "K = |E|" + }, + { + "bbox": [ + 130, + 296, + 480, + 356 + ], + "type": "text", + "content": " is the total number of edges." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 374, + 480, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 480, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 480, + 410 + ], + "type": "text", + "content": "Node (or City Block). Each node, or city block, " + }, + { + "bbox": [ + 130, + 374, + 480, + 410 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 130, + 374, + 480, + 410 + ], + "type": "text", + "content": " is a region defined by an enclosing road network and contains a set of block-level features and building-level features. The per-node features include:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 421, + 479, + 491 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 138, + 421, + 479, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 421, + 479, + 467 + ], + "spans": [ + { + "bbox": [ + 138, + 421, + 479, + 467 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 138, + 421, + 479, + 467 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 138, + 421, + 479, + 467 + ], + "type": "text", + "content": " is the shape and location feature vector of a city block contour. It includes four features: (1) aspect ratio, (2) total block area, (3) ratio of total block area over the block's convex-hull area, and (4) relative distance from current block to the centroid of the city." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 139, + 468, + 479, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 468, + 479, + 491 + ], + "spans": [ + { + "bbox": [ + 139, + 468, + 479, + 491 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 139, + 468, + 479, + 491 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 139, + 468, + 479, + 491 + ], + "type": "text", + "content": " is a 512-dimensional vector from a codebook " + }, + { + "bbox": [ + 139, + 468, + 479, + 491 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 139, + 468, + 479, + 491 + ], + "type": "text", + "content": ". It hierarchically captures the buildings layouts and their configuration within " + }, + { + "bbox": [ + 139, + 468, + 479, + 491 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 139, + 468, + 479, + 491 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 503, + 479, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 479, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 479, + 528 + ], + "type": "text", + "content": "We concatenate these node features to form a 516-dimensional vector " + }, + { + "bbox": [ + 130, + 503, + 479, + 528 + ], + "type": "inline_equation", + "content": "\\left([s_i, q_i]\\right)" + }, + { + "bbox": [ + 130, + 503, + 479, + 528 + ], + "type": "text", + "content": ", which is able to represent our large and heterogeneous set of blocks and buildings." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": "Building Layout Quantization. The 512-dimensional vector " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " describing the building layouts within a city block is obtained from a quantized codebook " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "C = \\{1,2,\\dots,L\\}" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " with quantization level " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "L = |C|" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": ". To create the codebook, we train a block-level variational autoencoder (BVAE) for building layout within a city block, using a canonical graph structure for representing the multiple building shapes, positions, and heights inside a single city block. Using this pretrained encoder, the latent vector of all possible building layouts will form a numerical distribution for each latent dimension. We further quantize each of these distributions into " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " bins with equal percentiles. Then, each latent value is replaced by the rank of the bin containing its value. Therefore, any building" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 90, + 447, + 102 + ], + "type": "text", + "content": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 465, + 248 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 465, + 248 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 465, + 248 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 465, + 248 + ], + "type": "image", + "image_path": "beb3bd749036747d4c0ba7eb9860694e92bf71f2956b85b6de9b26f6a9d12de5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 256, + 482, + 312 + ], + "lines": [ + { + "bbox": [ + 130, + 256, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 482, + 312 + ], + "type": "text", + "content": "Fig. 2: Canonical Graph Representation. Our method represents a city as a canonical graph " + }, + { + "bbox": [ + 130, + 256, + 482, + 312 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 256, + 482, + 312 + ], + "type": "text", + "content": ". Each node " + }, + { + "bbox": [ + 130, + 256, + 482, + 312 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 130, + 256, + 482, + 312 + ], + "type": "text", + "content": " represent a single city block, and each edge " + }, + { + "bbox": [ + 130, + 256, + 482, + 312 + ], + "type": "inline_equation", + "content": "e_{ij}" + }, + { + "bbox": [ + 130, + 256, + 482, + 312 + ], + "type": "text", + "content": " connects spatially adjacent blocks. Each block/node corresponds to a set of node features " + }, + { + "bbox": [ + 130, + 256, + 482, + 312 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 130, + 256, + 482, + 312 + ], + "type": "text", + "content": " and a quantized vector " + }, + { + "bbox": [ + 130, + 256, + 482, + 312 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 130, + 256, + 482, + 312 + ], + "type": "text", + "content": " hierarchically capturing enclosed building layouts. Edge feature " + }, + { + "bbox": [ + 130, + 256, + 482, + 312 + ], + "type": "inline_equation", + "content": "d_{ij}" + }, + { + "bbox": [ + 130, + 256, + 482, + 312 + ], + "type": "text", + "content": " encodes distances between block centroids. Graph " + }, + { + "bbox": [ + 130, + 256, + 482, + 312 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 256, + 482, + 312 + ], + "type": "text", + "content": " is used for GMAE training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 335, + 482, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 431 + ], + "type": "text", + "content": "layout within a city block is quantized into a 512-dimensional categorical index vector " + }, + { + "bbox": [ + 130, + 335, + 482, + 431 + ], + "type": "inline_equation", + "content": "q_{i} = [c_{1},\\dots,c_{512}]" + }, + { + "bbox": [ + 130, + 335, + 482, + 431 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 130, + 335, + 482, + 431 + ], + "type": "inline_equation", + "content": "c\\in C" + }, + { + "bbox": [ + 130, + 335, + 482, + 431 + ], + "type": "text", + "content": ". Note that our codebook is defined after BVAE training and not dynamically trainable as in VQGAN/VQVAE [16,56]. After performing a set of detailed evaluations using different training schemes, encoding models, and quantization levels " + }, + { + "bbox": [ + 130, + 335, + 482, + 431 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 335, + 482, + 431 + ], + "type": "text", + "content": " (see Sec. 4.3), we selected the Graph Attention Network (GAT) [59] as the backbone for the BVAE. Further, we select " + }, + { + "bbox": [ + 130, + 335, + 482, + 431 + ], + "type": "inline_equation", + "content": "L = 20" + }, + { + "bbox": [ + 130, + 335, + 482, + 431 + ], + "type": "text", + "content": " to balance compactness and representation ability after quantization. BVAE and quantization details are found in Supplementary Materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 449, + 481, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 481, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 481, + 484 + ], + "type": "text", + "content": "Edge. Each edge " + }, + { + "bbox": [ + 130, + 449, + 481, + 484 + ], + "type": "inline_equation", + "content": "e_{ij}" + }, + { + "bbox": [ + 130, + 449, + 481, + 484 + ], + "type": "text", + "content": " contains a feature " + }, + { + "bbox": [ + 130, + 449, + 481, + 484 + ], + "type": "inline_equation", + "content": "d_{ij}" + }, + { + "bbox": [ + 130, + 449, + 481, + 484 + ], + "type": "text", + "content": " storing the relative distance between the centroids of the adjacent city blocks " + }, + { + "bbox": [ + 130, + 449, + 481, + 484 + ], + "type": "inline_equation", + "content": "b_i" + }, + { + "bbox": [ + 130, + 449, + 481, + 484 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 449, + 481, + 484 + ], + "type": "inline_equation", + "content": "b_j" + }, + { + "bbox": [ + 130, + 449, + 481, + 484 + ], + "type": "text", + "content": ". This distance-weighted adjacency helps encode the relationship between neighboring blocks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 484, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 481, + 521 + ], + "type": "text", + "content": "Altogether, the shape, style, and multi-layer urban layout of a city with " + }, + { + "bbox": [ + 130, + 484, + 481, + 521 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 484, + 481, + 521 + ], + "type": "text", + "content": " nodes and " + }, + { + "bbox": [ + 130, + 484, + 481, + 521 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 484, + 481, + 521 + ], + "type": "text", + "content": " edges is represented by a total of " + }, + { + "bbox": [ + 130, + 484, + 481, + 521 + ], + "type": "inline_equation", + "content": "(516N + K)" + }, + { + "bbox": [ + 130, + 484, + 481, + 521 + ], + "type": "text", + "content": " variables. This canonical graph is used for the self-supervised GMAE training described below." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 538, + 338, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 538, + 338, + 550 + ], + "spans": [ + { + "bbox": [ + 131, + 538, + 338, + 550 + ], + "type": "text", + "content": "3.2 Graph-based Masked AutoEncoder" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 616 + ], + "type": "text", + "content": "We train our GMAE using a self-supervised process. A road-only city (or community) is a graph (or subgraph) " + }, + { + "bbox": [ + 130, + 557, + 481, + 616 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 557, + 481, + 616 + ], + "type": "text", + "content": " where building layout features " + }, + { + "bbox": [ + 130, + 557, + 481, + 616 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 130, + 557, + 481, + 616 + ], + "type": "text", + "content": " of every node (block) are missing, but " + }, + { + "bbox": [ + 130, + 557, + 481, + 616 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 130, + 557, + 481, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 557, + 481, + 616 + ], + "type": "inline_equation", + "content": "e_{ij}" + }, + { + "bbox": [ + 130, + 557, + 481, + 616 + ], + "type": "text", + "content": " are preserved. A context-sensitive node feature generation process corresponds to recreating the purposefully removed building layout features." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "During training (Fig. 3), we use dynamic masking ratios of " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " in order to learn how to regenerate the original features. In each training iteration, the mask ratio " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " is sampled from a truncated Gaussian distribution as recommended by Li et al. [35]. The ratio is obtained from the range " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "m \\in [0.5,1.0]" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " and is centered at" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "type": "text", + "content": "He, L. & Aliaga, D." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 199 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 199 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 199 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 199 + ], + "type": "image", + "image_path": "b4bd815f63a9b166abdc9f26433bee9e908635be57d3c82ee2b98cdc8964af06.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "lines": [ + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "type": "text", + "content": "Fig. 3: Graph-based Masked Autoencoder. Given a canonical city graph, quantized building layout features " + }, + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "type": "text", + "content": " are masked with dynamic masking ratios " + }, + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "type": "inline_equation", + "content": "m \\in [0.5,1.0]" + }, + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "type": "text", + "content": ", while block shape and location features " + }, + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "type": "text", + "content": " are kept. The GNN encoder uses message passing between neighboring nodes to obtain the context-aware node features " + }, + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "type": "text", + "content": ". The decoder uses " + }, + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "type": "text", + "content": " to reconstruct " + }, + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "type": "inline_equation", + "content": "Q'" + }, + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "type": "text", + "content": ". The predicted " + }, + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "type": "inline_equation", + "content": "Q'" + }, + { + "bbox": [ + 130, + 206, + 482, + 262 + ], + "type": "text", + "content": " are decoded to 2.5D urban layouts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 285, + 482, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 334 + ], + "type": "text", + "content": "0.55. We denote building layout features as " + }, + { + "bbox": [ + 130, + 285, + 482, + 334 + ], + "type": "inline_equation", + "content": "Q = \\{q_{i} | i \\in [1, N]\\}" + }, + { + "bbox": [ + 130, + 285, + 482, + 334 + ], + "type": "text", + "content": " and those after masking as " + }, + { + "bbox": [ + 130, + 285, + 482, + 334 + ], + "type": "inline_equation", + "content": "Q^{mask}" + }, + { + "bbox": [ + 130, + 285, + 482, + 334 + ], + "type": "text", + "content": ". Block features are denoted as " + }, + { + "bbox": [ + 130, + 285, + 482, + 334 + ], + "type": "inline_equation", + "content": "S = \\{s_{i} | i \\in [1, N]\\}" + }, + { + "bbox": [ + 130, + 285, + 482, + 334 + ], + "type": "text", + "content": ". The graph neural network (GNN) encoder is denoted as " + }, + { + "bbox": [ + 130, + 285, + 482, + 334 + ], + "type": "inline_equation", + "content": "f_{E}" + }, + { + "bbox": [ + 130, + 285, + 482, + 334 + ], + "type": "text", + "content": ", and node feature decoder as " + }, + { + "bbox": [ + 130, + 285, + 482, + 334 + ], + "type": "inline_equation", + "content": "f_{D}" + }, + { + "bbox": [ + 130, + 285, + 482, + 334 + ], + "type": "text", + "content": ". Thus the process of our GMAE can be written as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 223, + 346, + 481, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 346, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 223, + 346, + 481, + 361 + ], + "type": "interline_equation", + "content": "F = f _ {E} ([ Q ^ {m a s k}, S ], E), \\quad Q ^ {\\prime} = f _ {D} (F) \\tag {1}", + "image_path": "1a210381045621d13b4dde47760e98f24a2ec372a3aad93aa2543ea3f47ce792.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 402 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 365, + 482, + 402 + ], + "type": "inline_equation", + "content": "F = \\{f_i | i \\in [1, N]\\}" + }, + { + "bbox": [ + 130, + 365, + 482, + 402 + ], + "type": "text", + "content": " indicates the context-sensitive node features obtained by message passing between neighboring nodes by " + }, + { + "bbox": [ + 130, + 365, + 482, + 402 + ], + "type": "inline_equation", + "content": "f_E" + }, + { + "bbox": [ + 130, + 365, + 482, + 402 + ], + "type": "text", + "content": ". Extracted " + }, + { + "bbox": [ + 130, + 365, + 482, + 402 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 365, + 482, + 402 + ], + "type": "text", + "content": " is processed by " + }, + { + "bbox": [ + 130, + 365, + 482, + 402 + ], + "type": "inline_equation", + "content": "f_D" + }, + { + "bbox": [ + 130, + 365, + 482, + 402 + ], + "type": "text", + "content": " to reconstruct building layout features " + }, + { + "bbox": [ + 130, + 365, + 482, + 402 + ], + "type": "inline_equation", + "content": "Q'" + }, + { + "bbox": [ + 130, + 365, + 482, + 402 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 402, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 509 + ], + "type": "text", + "content": "The backbones of " + }, + { + "bbox": [ + 130, + 402, + 482, + 509 + ], + "type": "inline_equation", + "content": "f_{E}" + }, + { + "bbox": [ + 130, + 402, + 482, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 402, + 482, + 509 + ], + "type": "inline_equation", + "content": "f_{D}" + }, + { + "bbox": [ + 130, + 402, + 482, + 509 + ], + "type": "text", + "content": " can be any GNN, including GAT [58], GCN [33], and GraphSAGE [18]. After experimentation, we select GAT as " + }, + { + "bbox": [ + 130, + 402, + 482, + 509 + ], + "type": "inline_equation", + "content": "f_{E}" + }, + { + "bbox": [ + 130, + 402, + 482, + 509 + ], + "type": "text", + "content": " because of its best context-sensitive feature extraction ability, and we choose a straightforward MLP as " + }, + { + "bbox": [ + 130, + 402, + 482, + 509 + ], + "type": "inline_equation", + "content": "f_{D}" + }, + { + "bbox": [ + 130, + 402, + 482, + 509 + ], + "type": "text", + "content": " for efficient processing. We observed that the re-masking trick and complex GNN decoder utilized by [25,26] requires more memory and calculation but has no net benefit in our application. Since " + }, + { + "bbox": [ + 130, + 402, + 482, + 509 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 402, + 482, + 509 + ], + "type": "text", + "content": " contains quantized index vectors, we use cross entropy loss for self-supervised reconstruction training on masked nodes. Detailed ablations of masking strategy and model selections are provided in Sec. 4.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 533 + ], + "type": "text", + "content": "Our objective function is as follows, where " + }, + { + "bbox": [ + 130, + 510, + 482, + 533 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 510, + 482, + 533 + ], + "type": "text", + "content": " indicates the length of the quantized codebook " + }, + { + "bbox": [ + 130, + 510, + 482, + 533 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 510, + 482, + 533 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 238, + 544, + 481, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 544, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 238, + 544, + 481, + 576 + ], + "type": "interline_equation", + "content": "L _ {r e c o n} = - \\sum_ {i = 1} ^ {L} \\left[ Q _ {i} \\log \\left(Q _ {i} ^ {\\prime}\\right) \\right] ^ {m a s k} \\tag {2}", + "image_path": "7fa3d92516bbea12ebf273eefaba7cd399de9c07eea80ae7b749eddb2f8ecdb4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": "To capture community behavior (and also reduce GPU memory usage), we randomly sample a batch of subgraphs from " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": " for each training iteration. The nodes in each subgraph are sampled using a chosen radius from a random center node. While community sizes vary significantly, we found that a practical compromise is to use 500 meters as the radius value. This radius ensures to cover a reasonable number of the city blocks with a typical census tract from governmental CJEST dataset [1]." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 90, + 448, + 102 + ], + "type": "text", + "content": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 206 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 206 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 206 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 206 + ], + "type": "image", + "image_path": "d91eb5da47464a7c083d15174c7970e1f4a263d9ab769e3268343e3253a648ab.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 213, + 482, + 258 + ], + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 258 + ], + "type": "text", + "content": "Fig. 4: Priority-based Scheduled Generation. We iteratively utilize pretrained GMAE to reconstruct masked node features. In each iteration, we accept a certain ratio of predicted nodes decided by the scheduling function " + }, + { + "bbox": [ + 130, + 213, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\beta(t) = 1 - \\cos(t / T)" + }, + { + "bbox": [ + 130, + 213, + 482, + 258 + ], + "type": "text", + "content": ". We obtain a full graph after " + }, + { + "bbox": [ + 130, + 213, + 482, + 258 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 213, + 482, + 258 + ], + "type": "text", + "content": " iterations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 283, + 346, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 283, + 346, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 283, + 346, + 295 + ], + "type": "text", + "content": "3.3 Priority-based Scheduled Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 305, + 481, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 481, + 377 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 481, + 377 + ], + "type": "text", + "content": "While in theory the trained GMAE can predict all nodes in a single pass, the generation quality is much better when performed iteratively. Recent published papers (e.g., [10,35]) have found that gradually sampling with a reasonable number of iterations improves generation quality. In urban scenarios, small number of representative city blocks are important to indicate context behavior. We intend to generate those blocks at early iterations and then span neighboring blocks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 378, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 534 + ], + "type": "text", + "content": "In Fig. 4, we propose a " + }, + { + "bbox": [ + 130, + 378, + 482, + 534 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 378, + 482, + 534 + ], + "type": "text", + "content": "-iteration sampling strategy similar to MaskGIT [10] and MAGE [35]. At each iteration, GMAE predicts per-node building layout features for all remaining nodes. Then we rank generated nodes by their prediction confidence. The generated nodes with the highest prediction confidence are accepted. The accepted nodes after the current iteration " + }, + { + "bbox": [ + 130, + 378, + 482, + 534 + ], + "type": "inline_equation", + "content": "t \\in [1,T]" + }, + { + "bbox": [ + 130, + 378, + 482, + 534 + ], + "type": "text", + "content": " are scheduled by a cosine function " + }, + { + "bbox": [ + 130, + 378, + 482, + 534 + ], + "type": "inline_equation", + "content": "\\beta(t) = 1 - \\cos(t / T)" + }, + { + "bbox": [ + 130, + 378, + 482, + 534 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 130, + 378, + 482, + 534 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 378, + 482, + 534 + ], + "type": "text", + "content": " is small, the ratio is near to 0 causing few nodes to be accepted in initial iterations. As iterations continue, " + }, + { + "bbox": [ + 130, + 378, + 482, + 534 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 378, + 482, + 534 + ], + "type": "text", + "content": " gets bigger and hence the number of accepted nodes during each iteration will also increase. This behavior follows the prioritization observation in Sec. 1 which causes distinguished, or unique, city blocks to be generated first and essentially guide the layout style for nearby city blocks. Typically we set " + }, + { + "bbox": [ + 130, + 378, + 482, + 534 + ], + "type": "inline_equation", + "content": "T = 12" + }, + { + "bbox": [ + 130, + 378, + 482, + 534 + ], + "type": "text", + "content": " which results in dozens of nodes being generated as is the typical size of a community. Detailed ablations of alternative sampling schedules are provided in Sec. 4.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 555, + 230, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 230, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 230, + 568 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 582, + 279, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 582, + 279, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 582, + 279, + 594 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "Open Dataset We have created a vector-based urban layout dataset of all US cities with a population of 100K or more. We believe this dataset will be of significant use to the community to further develop urban layout generation methods and thus we will release the dataset together with the paper. For each city, we define a rectangular bounding box containing most of the metropolitan area. Then," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "type": "text", + "content": "He, L. & Aliaga, D." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "we use this bounding box to extract data from OpenStreetMap (OSM) [43], Microsoft Building Footprints (MSF) [23], and Topologically Integrated Geographic Encoding and Referencing (TIGER) dataset [7]. The data we collected includes: (1) road network vectors and attributes, (2) building footprints and heights, and (3) block-level social economic metrics. Specifically, we collect road networks and block contours from TIGER. Then we look for building footprints and heights from OSM and MSF within each city block contour. We choose the one with more and accurate details (e.g., containing more buildings, having features that match those in TIGER). Over all 330 cities, this dataset has a total of 833,473 city blocks, and 17,663,607 buildings. We set training, validation, testing split as " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 267, + 482, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 482, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 482, + 351 + ], + "type": "text", + "content": "Training Details The training has two stages. We first use all city blocks for self-supervised training of a graph-based Block-level Variational AutoEncoder (BVAE). Given this trained block-level encoder, we perform the latent quantization for all building layouts in each city block. Then our GMAE is trained based on canonical city graph representation. Training time for BVAE and GMAE typically takes 12 hours and 15 hour respectively by a single A5000 GPU. We will release our codes upon acceptance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 351, + 481, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 481, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 481, + 388 + ], + "type": "text", + "content": "Inferencing using our graph-based approach is seemingly instantaneous for all the results shown in this paper. This is contrast to the VTN [2] and SDXL [48] approaches which take significantly more time to infer and to train." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 408, + 227, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 227, + 420 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 227, + 420 + ], + "type": "text", + "content": "4.2 Comparisons" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 430, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 538 + ], + "type": "text", + "content": "We compare our method to several alternate city-scale layout generators (Tab. 1). Specifically, we randomly select 100 communities from among our test set, total 3700 city blocks. All methods generate the layout of the 100 communities in one pass, and without any post-processing or human-in-the-loop refinement. Implementations are downloaded from the author's repositories and trained with default parameter settings. For fairness, we convert our dataset to token sequences, graphs, or image patches " + }, + { + "bbox": [ + 130, + 430, + 482, + 538 + ], + "type": "inline_equation", + "content": "(1024\\times 1024)" + }, + { + "bbox": [ + 130, + 430, + 482, + 538 + ], + "type": "text", + "content": " as needed. VTN [2], LayoutDM [29], GlobalMapper [20] are retrained for single block generation. SDXL [48] is fine-tuned for outpainting by using the upper half to predict the lower half." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": "Evaluation Metrics We define a Context Score " + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "inline_equation", + "content": "(CTS)" + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": " to evaluate the context relationships among neighboring city blocks (i.e., a community). This metric is a numerical extension of LayoutSim [20] and DocSim index [46] which evaluate the similarity between two city blocks regarding location, size, and category. Specifically, for each generated city block " + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": ", we calculate the average LayoutSim with its neighbors " + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "inline_equation", + "content": "N(i)" + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": ". Then, the local average of LayoutSim represents the context similarity among the community " + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "inline_equation", + "content": "(CT)" + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": ". High " + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "inline_equation", + "content": "CT" + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": " indicates identical blocks in a community and low " + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "inline_equation", + "content": "CT" + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": " corresponds to high diversity in a community. " + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "inline_equation", + "content": "CTS" + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": " is the subtraction of the real community's " + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "inline_equation", + "content": "CT" + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": " from the generated " + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "inline_equation", + "content": "CT" + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 90, + 447, + 102 + ], + "type": "text", + "content": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 125, + 479, + 552 + ], + "blocks": [ + { + "bbox": [ + 135, + 125, + 479, + 552 + ], + "lines": [ + { + "bbox": [ + 135, + 125, + 479, + 552 + ], + "spans": [ + { + "bbox": [ + 135, + 125, + 479, + 552 + ], + "type": "image", + "image_path": "d0fdcdfcd0b9e6fb1c0ccc86b21a788ad0ef9b20effa5675ae16eb2ebd8bfd5e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 561, + 482, + 651 + ], + "lines": [ + { + "bbox": [ + 130, + 561, + 482, + 651 + ], + "spans": [ + { + "bbox": [ + 130, + 561, + 482, + 651 + ], + "type": "text", + "content": "Fig. 5: Qualitative Comparisons. Given the same road network (except SDXL [48]), all above methods generate urban layouts in only one pass without post-processing or human-in-the-loop refinement. The \"even rows\" are a zoom-in of the highlighted areas in the \"odd rows\". Our method generates a realistic distribution of urban layouts with plausible context-dependent behaviors (as indicated by " + }, + { + "bbox": [ + 130, + 561, + 482, + 651 + ], + "type": "inline_equation", + "content": "CTS" + }, + { + "bbox": [ + 130, + 561, + 482, + 651 + ], + "type": "text", + "content": " score). VTN [2] and LayoutDM [29] show abnormal style shift among neighboring blocks, and no awareness of road networks. GlobalMapper [20] generates over-similar communities. SDXL [48] provides poor outpainting quality containing semantic errors and undesired overlaps." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "type": "text", + "content": "He, L. & Aliaga, D." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 168, + 479, + 254 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 1: Quantitative Comparisons. All methods are compared to the same real urban layouts (except SDXL [48] which cannot take-in a road network). Best values are in bold, second best values are underlined. Our method outperforms other existing methods in all but the overlap metric. See text for an explanation of the metrics." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 168, + 479, + 254 + ], + "lines": [ + { + "bbox": [ + 133, + 168, + 479, + 254 + ], + "spans": [ + { + "bbox": [ + 133, + 168, + 479, + 254 + ], + "type": "table", + "html": "
Method\\(CTS_{x\\rightarrow 0}\\)WD-5D↓WD-CO↓Overlap↓O-Blk↓FID↓KID↓LPIPS↓
SDXL [48]-----120.240.0790.48
VTN [2]-1.143.185.811.247.3569.140.0470.32
LayoutDM [29]-2.202.9212.504.561.7266.770.0400.39
GlobalMP [20]0.624.774.142.520.6849.550.0240.34
Ours0.212.281.911.270.4223.630.0050.20
", + "image_path": "d08205646244d3b2de1405d1fb8dd29dc90d3c3fe8a7b9191a1a58795f90fe8a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 169, + 286, + 481, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 286, + 481, + 321 + ], + "spans": [ + { + "bbox": [ + 169, + 286, + 481, + 321 + ], + "type": "interline_equation", + "content": "C T = \\frac {1}{\\| N (i) \\|} \\sum_ {j} ^ {N (i)} \\operatorname {L a y o u t S i m} \\left(b _ {i}, b _ {j}\\right), \\quad N (i) = \\{j \\mid (i, j) \\in E \\} \\tag {3}", + "image_path": "bf5d20f1a01f762e9bde60a24bc7c2b826fcdcc2b10a84d90632e6ac663d3e0a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 254, + 332, + 481, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 332, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 254, + 332, + 481, + 346 + ], + "type": "interline_equation", + "content": "C T S = C T _ {g e n} - C T _ {r e a l} \\tag {4}", + "image_path": "fce9fb8b154169310e824c7fc9b4e7f610ed64490d1d6bad7a5dea1fde0dd9f5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 351, + 482, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 386 + ], + "type": "text", + "content": "Given a real community, the metric indicates whether a generated community is over-diverse (" + }, + { + "bbox": [ + 130, + 351, + 482, + 386 + ], + "type": "inline_equation", + "content": "CTS < 0" + }, + { + "bbox": [ + 130, + 351, + 482, + 386 + ], + "type": "text", + "content": "), or over-similar (" + }, + { + "bbox": [ + 130, + 351, + 482, + 386 + ], + "type": "inline_equation", + "content": "CTS > 0" + }, + { + "bbox": [ + 130, + 351, + 482, + 386 + ], + "type": "text", + "content": "). A low value means more realism." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 388, + 483, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 483, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 483, + 495 + ], + "type": "text", + "content": "In our comparisons, we compute several metrics. One metric is the Wasserstein Distance between the distributions of generated communities and real communities in terms of 2D position (x, y) and 2.5D geometry (length, width, height), and building count (CO) - we name these WD-5D and WD-CO. We define two quality metrics as well: Overlap indicates the percentage by which two building overlaps, and O-Blk represents the percentage of building area that sticks outside of its city block. We also implement several perceptual metrics: FID [24], KID [4], and LPIPS [65] to evaluate the visual quality and realism of generated layouts. For all the above metrics, lower value indicates better quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 514, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 563 + ], + "type": "text", + "content": "Quantitative Results As Tab. 1 shows, our approach performs better than existing approaches in all but one metric. Note that the poor quality of SDXL results makes it hard to decipher layouts. We only compare to it by pixel-based perceptual metrics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Qualitative Results We show in Fig. 5 qualitative results for several exemplary areas (more examples are in Supplemental Material). In general, our method shows good contextual harmonization and has awareness of arbitrary road networks. It generates a realistic distribution of urban layouts with plausible context dependent behaviors. Other methods show improper patterns (e.g., overly diverse or similar) and potentially containing semantic errors or undesired overlaps." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 138, + 90, + 448, + 103 + ], + "type": "text", + "content": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 168, + 462, + 373 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 2: Block Quantization Ablations. We report metrics (all in %) among all alternative ablations of our Block-level graph-based Variational AutoEncoder (BVAE). Using GAT-backbone with post-trained dimensional quantization with " + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "L = 20" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": " shows the best overall compromise. The best values are in boldface." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 168, + 462, + 373 + ], + "lines": [ + { + "bbox": [ + 151, + 168, + 462, + 373 + ], + "spans": [ + { + "bbox": [ + 151, + 168, + 462, + 373 + ], + "type": "table", + "html": "
Encode ModelOverlap↓Out-Blk↓Pos-E.↓Geom-E.↓Ct-E.↓Cov-E.↓
LayoutVAE [32]13.8411.1515.0937.35-24.34
VTN [2]2.704.706.0122.8618.581.97
BlockPlanner [62]2.931.304.7812.063.635.66
GlobalMapper [20]1.041.203.252.830.080.46
BVAE (SAGE [18])1.331.705.013.250.100.35
BVAE (GCN [33])0.451.214.252.740.040.28
BVAE* (GAT [58])0.170.281.000.710.0060.04
QuantizationOverlap↓Out-Blk↓Pos-E.↓Geom-E.↓Ct-E.↓Cov-E.↓
Trainable-VQ3.020.553.755.780.070.60
KMeans-Q2.900.0220.9817.8550.0116.23
DIM-Q (L=5)0.280.627.7116.2025.014.68
DIM-Q (L=10)0.560.285.734.488.334.58
DIM-Q (L=30)0.170.012.680.780.0050.15
DIM-Q* (L=20)0.110.013.500.890.0060.12
", + "image_path": "3ab87a3cd1f94a9bfaf5fbc6e39c2ab166a7fbf2e8ac7e035afc0c890a3de6fb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 397, + 246, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 246, + 408 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 246, + 408 + ], + "type": "text", + "content": "4.3 Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 424, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 544 + ], + "type": "text", + "content": "Block Quantization We ablate block quantization in order to understand and maximize the impact of changing its parameters and design strategy. First, we report in Tab. 2 (Encode Model Table) the reconstruction loss using different block-level variational autoencoders. We evaluate using various geometric metrics: Overlap (percent of building-to-building overlap), Out-Blk (percent of buildings that stick outside of its city block), Pos-E (positional error as a percentage of block diagonal), Geom-E (2.5D building geometry error as percentage of block scale), Ct-E (percentage of building count error), and Cov-E (total building coverage error). They clearly show our approach is the best overall. Thus we choose our BVAE based on Graph Attention Networks (GAT) [58]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "type": "text", + "content": "Then, in Tab. 2 (Quantization Table), we ablate various codebook approaches and quantitization levels for block-level encoding. \"Trainable-VQ\" leverages trainable vector quantization codebooks introduced by VQVAE [56]. We also provide an alternative non-training based quantization method using KMeans (\"KMeans-Q\") to cluster BVAE latent space values to a reasonable number of categories in order to handle diverse urban layouts (e.g., 100 classes). The \"DIM-Q\" rows imply quantizing each of the 512 dimensions of the block-level code into a codebook " + }, + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "type": "inline_equation", + "content": "C = \\{1,2,\\dots,L\\}" + }, + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "type": "text", + "content": " as described in Sec. 3.1. This per-dimension quantization performs clearly the best. Specifically, setting " + }, + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 545, + 483, + 666 + ], + "type": "text", + "content": " to infinity produces a lossless representation of BVAE latent vectors, but also causes the curse of dimension-" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 247, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 247, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 247, + 102 + ], + "type": "text", + "content": "He, L. & Aliaga, D." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 168, + 480, + 330 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 3: City-Scale GMAE Ablations. We ablate backbone message passing layers, maximum messaging passing hops (" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": ": depth of encoder), masking strategy during training, and scheduling functions of iterative generation (" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": " indicates total iterations). The overall best setting is marked as *." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 168, + 480, + 330 + ], + "lines": [ + { + "bbox": [ + 132, + 168, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 168, + 480, + 330 + ], + "type": "table", + "html": "
AblationsContext↓WD-3D↓WD-N↓Overlap↓Out-Blk↓FID↓KID↓LPIPS↓
GMAE (GCN [33])-0.302.742.910.260.3652.950.0250.32
GMAE (GTrans [52])0.743.023.271.961.0743.600.0180.29
GMAE (SAGE [18])0.972.982.361.752.6043.150.0170.29
GMAE (GAT, Dec)0.942.452.962.131.1048.240.0190.30
GMAE (GAT, D = 1)1.253.032.641.730.4443.750.0180.29
GMAE (GAT, D = 2)0.332.662.011.310.5734.820.0140.27
GMAE (GAT, D = 4)0.241.922.310.630.2824.590.0090.22
GMAE (Mask ∈ [0.05, 1])1.302.703.621.702.4543.140.0180.29
GMAE (Mask ∈ [0.05, 0.5])1.103.073.681.681.3046.850.0200.30
GMAE (Mask = 0.15)1.594.913.411.831.2049.500.0220.31
GMAE (Linear, T = 12)0.632.833.791.270.9337.420.0150.28
GMAE (Log, T = 12)-1.312.777.277.413.1463.250.0380.33
GMAE (Cosine, T = 1)-1.677.445.643.564.0176.120.0510.35
GMAE (Cosine, T = 20)0.322.923.981.680.3132.060.0130.24
*(GAT-D3, [0.5,1], Cos-T12)0.212.281.911.270.4223.630.0050.20
", + "image_path": "7c55e30166b8702645ac7e59d7a92831436b89b3d80c1915cc66bff3582d2cc2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 354, + 481, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 481, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 481, + 378 + ], + "type": "text", + "content": "ality. We discover that " + }, + { + "bbox": [ + 130, + 354, + 481, + 378 + ], + "type": "inline_equation", + "content": "L = 20" + }, + { + "bbox": [ + 130, + 354, + 481, + 378 + ], + "type": "text", + "content": " per dimension is the best overall compromise balancing compactness and representation ability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 399, + 481, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 481, + 435 + ], + "type": "text", + "content": "City-Scale Graph Masked Autoencoder. In Tab. 3, We ablate various model parameters, together with significant training and inference strategies for our GMAE. Each alternatives is evaluated by the same set of metrics in Sec. 4.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 481, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 481, + 555 + ], + "type": "text", + "content": "By varying the message passing layers, we find GAT [58] provides the best performance. GCN [33] produces diverse and unrealistic urban layouts, while transformer-based [52], and GraphSAGE [18] backbones generate blocks that are too similar. The depth of the stacked layers in the encoder (D) dictates the maximum message passing hops from each node and thus the span of context behaviors captured for realistic generation. We find that 3 hops is promising and larger contexts cause over-similarity and unnecessary computing burden. We also observed that using stacked GAT layers for decoding with remasking tricks, as described by [25,26], didn't produce a net benefit towards city generation. So we choose MLP as the node feature decoder for both its performance and efficiency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 616 + ], + "type": "text", + "content": "The masking strategy during training is crucial for GMAE performance. We find that dynamic and large masking ratios are beneficial, while a fixed small ratio (e.g., 0.15 as in [14]) leads to over similarity. This behavior is also observed by Li et al. [35]. We adopt a similar masking strategy by sampling a Gaussian distribution " + }, + { + "bbox": [ + 130, + 557, + 481, + 616 + ], + "type": "inline_equation", + "content": "N(0.55, 0.25)" + }, + { + "bbox": [ + 130, + 557, + 481, + 616 + ], + "type": "text", + "content": ", truncated by [0.5, 1.0], for each training iteration." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Scheduling for iterative generation is also critical to capture representative city blocks (Sec. 3.3). We evaluate several functions " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\beta(t)" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " (e.g., linear, logarithmic, or cosine-based) that provide the scheduled acceptance ratio, and evaluate alternative total iteration counts " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": ". Rapid iteration leads to over-diversity and" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 90, + 447, + 102 + ], + "type": "text", + "content": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "long iteration results in over-similarity. Logarithmic functions with reverse pattern produces the worst realism. Metrics show that cosine-based function, which is conservative initially and generous in later iterations, outperforms others." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 189 + ], + "type": "text", + "content": "In summary, this study determines the best configuration for our GMAE as: GAT encoder with " + }, + { + "bbox": [ + 130, + 152, + 480, + 189 + ], + "type": "inline_equation", + "content": "D = 3" + }, + { + "bbox": [ + 130, + 152, + 480, + 189 + ], + "type": "text", + "content": " and MLP decoder, training mask ratio " + }, + { + "bbox": [ + 130, + 152, + 480, + 189 + ], + "type": "inline_equation", + "content": "m \\in [0.5,1.0]" + }, + { + "bbox": [ + 130, + 152, + 480, + 189 + ], + "type": "text", + "content": ", and priority-based scheduling function " + }, + { + "bbox": [ + 130, + 152, + 480, + 189 + ], + "type": "inline_equation", + "content": "\\beta (t) = 1 - \\cos (t / T)" + }, + { + "bbox": [ + 130, + 152, + 480, + 189 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 152, + 480, + 189 + ], + "type": "inline_equation", + "content": "T = 12" + }, + { + "bbox": [ + 130, + 152, + 480, + 189 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 206, + 218, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 206, + 218, + 217 + ], + "spans": [ + { + "bbox": [ + 131, + 206, + 218, + 217 + ], + "type": "text", + "content": "4.4 Limitations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 226, + 479, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 479, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 479, + 274 + ], + "type": "text", + "content": "Our method faces limitations where shapes of city blocks or buildings are not simple polygons (e.g. the building with a garden inside). It also struggles to handle very irregular and concave shaped city blocks (e.g., cul-de-sac's). Also, road-network to contextual structure relations are not explicitly considered." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 292, + 254, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 292, + 254, + 303 + ], + "spans": [ + { + "bbox": [ + 131, + 292, + 254, + 303 + ], + "type": "text", + "content": "4.5 Additional Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 312, + 440, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 312, + 440, + 324 + ], + "spans": [ + { + "bbox": [ + 131, + 312, + 440, + 324 + ], + "type": "text", + "content": "We place several other interesting results in Supplementary Materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 334, + 479, + 427 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 138, + 334, + 479, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 334, + 479, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 334, + 479, + 369 + ], + "type": "text", + "content": "- Controllable Generation of Entire City. Our method enables realistic and efficient city-scale controllable generation given any percentage of priors. We show several large examples spanning a subset of the many cities." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 370, + 479, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 479, + 404 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 479, + 404 + ], + "type": "text", + "content": "- Socio-Economic Prediction. Trained GMAE may act as a powerful encoder for downstream tasks. For example, the encoded GMAE latent vectors of urban layouts can indicate social-economic metrics for each city block." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 139, + 406, + 479, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 406, + 479, + 427 + ], + "spans": [ + { + "bbox": [ + 139, + 406, + 479, + 427 + ], + "type": "text", + "content": "- Semantic Manipulations. Our method enables semantic manipulations over various building layout styles across different cities." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 449, + 336, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 449, + 336, + 460 + ], + "spans": [ + { + "bbox": [ + 131, + 449, + 336, + 460 + ], + "type": "text", + "content": "5 Conclusions and Future Works" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 474, + 480, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 480, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 480, + 605 + ], + "type": "text", + "content": "Large-scale urban layout generation has seen significant advancements, spurred by the intersection of computer vision, urban planning, and related disciplines. Previous methodologies often neglect the context-sensitive nature of urban layouts, failing to capture the multi-layer semantics crucial for realism and plausibility. Our proposed Graph-based Masked AutoEncoder (GMAE) addresses these limitations by leveraging a graph-based representation, acknowledging context-sensitivity, and prioritizing the generation process. By training on a diverse dataset of 330 cities, we demonstrate the efficacy of our approach in generating large-scale 2.5D layouts with realism and semantic consistency. Moreover, our method outperforms existing techniques in various metrics, marking a significant advancement in the realm of urban layout generation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "text", + "content": "Moving forward, our approach paves the way for more accurate and efficient urban simulation, digital twin creation, and game/content design. As future work we intend to incorporate our method for large-scale photo-realistic multi-view scene synthesis, and city-scale 3D modeling. Those potentials will contribute to synthetic data generation in autonomous driving and world model training." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 102 + ], + "type": "text", + "content": "He, L. & Aliaga, D." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 480, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 480, + 165 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 480, + 165 + ], + "type": "text", + "content": "This project was funded in part by NSF Grant #2107096 and NSF Grant #1835739." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 186, + 197, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 186, + 197, + 198 + ], + "spans": [ + { + "bbox": [ + 133, + 186, + 197, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 212, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 212, + 479, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 212, + 479, + 234 + ], + "spans": [ + { + "bbox": [ + 138, + 212, + 479, + 234 + ], + "type": "text", + "content": "1. Climate and economic justice screening tool. https://screeningtool.geoplatform.gov/" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 235, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 235, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 235, + 480, + 267 + ], + "type": "text", + "content": "2. Arroyo, D.M., Postels, J., Tombari, F.: Variational transformer networks for layout generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13642-13652 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 267, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 480, + 311 + ], + "type": "text", + "content": "3. Bhatt, M., Kalyanam, R., Nishida, G., He, L., May, C., Niyogi, D., Aliaga, D.: Design and deployment of photo2building: A cloud-based procedural modeling tool as a service. In: Practice and Experience in Advanced Research Computing, pp. 132-138 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 312, + 480, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 480, + 334 + ], + "type": "text", + "content": "4. Binkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying mmd gans. arXiv preprint arXiv:1801.01401 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 334, + 480, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 334, + 480, + 356 + ], + "spans": [ + { + "bbox": [ + 138, + 334, + 480, + 356 + ], + "type": "text", + "content": "5. Bokeloh, M., Wand, M., Seidel, H.P.: A connection between partial symmetry and inverse procedural modeling. In: ACM SIGGRAPH 2010 papers, pp. 1-10 (2010)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 356, + 480, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 356, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 356, + 480, + 399 + ], + "type": "text", + "content": "6. Brooks, T., Peebles, B., Holmes, C., DePue, W., Guo, Y., Jing, L., Schnurr, D., Taylor, J., Luhman, T., Luhman, E., Ng, C., Wang, R., Ramesh, A.: Video generation models as world simulators (2024), https://openai.com/research/video-generation-models-as-world-simulators" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 400, + 480, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 400, + 480, + 433 + ], + "spans": [ + { + "bbox": [ + 138, + 400, + 480, + 433 + ], + "type": "text", + "content": "7. Bureau, U.S.C.: Topologically integrated geographic encoding and referencing. https://www.census.gov/geographies/mapping-files/time-series/geo/tiger-line-file.html" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 434, + 480, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 480, + 466 + ], + "type": "text", + "content": "8. Chai, L., Tucker, R., Li, Z., Isola, P., Snavely, N.: Persistent nature: A generative model of unbounded 3d worlds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20863-20874 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 467, + 480, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 467, + 480, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 467, + 480, + 499 + ], + "type": "text", + "content": "9. Chai, S., Zhuang, L., Yan, F.: Layoutdm: Transformer-based diffusion model for layout generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18349-18358 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 500, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 500, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 500, + 480, + 533 + ], + "type": "text", + "content": "0. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked generative image transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11315-11325 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 533, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 480, + 577 + ], + "type": "text", + "content": "1. Chang, K.H., Cheng, C.Y., Luo, J., Murata, S., Nourbakhsh, M., Tsuji, Y.: Building-gan: Graph-conditioned architectural volumetric design generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11956-11965 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 577, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 480, + 599 + ], + "type": "text", + "content": "2. Chen, Z., Wang, G., Liu, Z.: Scenedreamer: Unbounded 3d scene generation from 2d image collections. arXiv preprint arXiv:2302.01330 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "text", + "content": "3. Deng, J., Chai, W., Guo, J., Huang, Q., Hu, W., Hwang, J.N., Wang, G.: Citygen: Infinite and controllable 3d city layout generation. arXiv preprint arXiv:2312.01508 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 666 + ], + "type": "text", + "content": "4. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "text", + "content": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 481, + 194 + ], + "type": "text", + "content": "16. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873-12883 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 481, + 237 + ], + "type": "text", + "content": "17. Gupta, K., Lazarow, J., Achille, A., Davis, L.S., Mahadevan, V., Shrivastava, A.: Layouttransformer: Layout generation and completion with self-attention. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1004-1014 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 237, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 237, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 133, + 237, + 481, + 259 + ], + "type": "text", + "content": "18. Hamilton, W., Ying, Z., Leskovec, J.: Inductive representation learning on large graphs. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 259, + 481, + 293 + ], + "type": "text", + "content": "19. He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 326 + ], + "type": "text", + "content": "20. He, L., Aliaga, D.: Globalmapper: Arbitrary-shaped urban layout generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 454-464 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 369 + ], + "type": "text", + "content": "21. He, L., Lu, Y., Corring, J., Florencio, D., Zhang, C.: Diffusion-based document layout generation. In: Fink, G.A., Jain, R., Kise, K., Zanibbi, R. (eds.) Document Analysis and Recognition - ICDAR 2023. pp. 361-378. Springer Nature Switzerland, Cham (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 369, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 392 + ], + "type": "text", + "content": "22. He, L., Shan, J., Aliaga, D.: Generative building feature estimation from satellite images. IEEE Transactions on Geoscience and Remote Sensing 61, 1-13 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "text", + "content": "23. Heris, M.P., Foks, N.L., Bagstad, K.J., Troy, A., Ancona, Z.H.: A rasterized building footprint dataset for the united states. Scientific data 7(1), 207 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 446 + ], + "type": "text", + "content": "24. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "text", + "content": "25. Hou, Z., He, Y., Cen, Y., Liu, X., Dong, Y., Kharlamov, E., Tang, J.: Graphmae2: A decoding-enhanced masked self-supervised graph learner. In: Proceedings of the ACM Web Conference 2023. pp. 737-746 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 512 + ], + "type": "text", + "content": "26. Hou, Z., Liu, X., Cen, Y., Dong, Y., Yang, H., Wang, C., Tang, J.: Graphmae: Self-supervised masked graph autoencoders. In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 594-604 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 544 + ], + "type": "text", + "content": "27. Hua, H., Shi, J., Kafle, K., Jenni, S., Zhang, D., Collomosse, J., Cohen, S., Luo, J.: Finematch: Aspect-based fine-grained image and text mismatch detection and correction. arXiv preprint arXiv:2404.14715 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "text", + "content": "28. Hui, M., Zhang, Z., Zhang, X., Xie, W., Wang, Y., Lu, Y.: Unifying layout generation with a decoupled diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1942-1951 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 578, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 621 + ], + "type": "text", + "content": "29. Inoue, N., Kikuchi, K., Simo-Serra, E., Otani, M., Yamaguchi, K.: Layoutdm: Discrete diffusion model for controllable layout generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10167-10176 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "30. Jiang, Z., Guo, J., Sun, S., Deng, H., Wu, Z., Mijovic, V., Yang, Z.J., Lou, J.G., Zhang, D.: Layoutformer++: Conditional graphic layout generation via constraint serialization and decoding space restriction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18403-18412 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "text", + "content": "He, L. & Aliaga, D." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 149 + ], + "type": "text", + "content": "31. Jyothi, A.A., Durand, T., He, J., Sigal, L., Mori, G.: Layoutvae: Stochastic scene layout generation from a label set. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9895-9904 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 192 + ], + "type": "text", + "content": "32. Jyothi, A.A., Durand, T., He, J., Sigal, L., Mori, G.: Layoutvae: Stochastic scene layout generation from a label set. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9894-9903 (2019). https://doi.org/10.1109/ICCV.2019.00999" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "text", + "content": "33. Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907 (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "text", + "content": "34. Li, J., Yang, J., Hertzmann, A., Zhang, J., Xu, T.: Layoutgan: Synthesizing graphic layouts with vector-wireframe adversarial networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(7), 2388-2399 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 292 + ], + "type": "text", + "content": "35. Li, T., Chang, H., Mishra, S., Zhang, H., Katabi, D., Krishnan, D.: Mage: Masked generative encoder to unify representation learning and image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2142-2152 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "type": "text", + "content": "36. Li, Z., Wang, Q., Snavely, N., Kanazawa, A.: Infinitenature-zero: Learning perpetual view generation of natural scenes from single images. In: European Conference on Computer Vision. pp. 515-534. Springer (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "text", + "content": "37. Lin, C.H., Lee, H.Y., Cheng, Y.C., Tulyakov, S., Yang, M.H.: Infinitygan: Towards infinite-pixel image synthesis. arXiv preprint arXiv:2104.03963 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "text", + "content": "38. Lin, C.H., Lee, H.Y., Menapace, W., Chai, M., Siarohin, A., Yang, M.H., Tulyakov, S.: Infinity: Infinite-scale city synthesis. arXiv preprint arXiv:2301.09637 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "39. Lipp, M., Scherzer, D., Wonka, P., Wimmer, M.: Interactive modeling of city layouts using layers of procedural content. In: Computer Graphics Forum. vol. 30, pp. 345-354. Wiley Online Library (2011)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "type": "text", + "content": "40. Ma, H., Zeng, D., Liu, Y.: Learning individualized treatment rules with many treatments: A supervised clustering approach using adaptive fusion. Advances in Neural Information Processing Systems 35, 15956-15969 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "text", + "content": "41. Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781 (2013)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 500 + ], + "type": "text", + "content": "42. Nauata, N., Chang, K.H., Cheng, C.Y., Mori, G., Furukawa, Y.: House-gan: Relational generative adversarial networks for graph-constrained house layout generation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16. pp. 162-177. Springer (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "text", + "content": "43. OpenStreetMap contributors: Planet dump retrieved from https://planet.osm.org. https://www.openstreetmap.org (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "text", + "content": "44. Para, W., Guerrero, P., Kelly, T., Guibas, L.J., Wonka, P.: Generative layout modeling using constraint graphs. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6690-6700 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "type": "text", + "content": "45. Patel, P., Kalyanam, R., He, L., Aliaga, D., Niyogi, D.: Deep learning-based urban morphology for city-scale environmental modeling. PNAS nexus 2(3), pgad027 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 588, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 632 + ], + "type": "text", + "content": "46. Patil, A.G., Ben-Eliezer, O., Perel, O., Averbuch-Elor, H.: Read: Recursive autoencoders for document layout generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 544-545 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "47. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4195-4205 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "text", + "content": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "48. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "text", + "content": "49. Shen, Y., Ma, W.C., Wang, S.: Sgam: Building a virtual 3d world through simultaneous generation and mapping. Advances in Neural Information Processing Systems 35, 22090-22102 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 214 + ], + "type": "text", + "content": "50. Sheng, Y., Liu, Y., Zhang, J., Yin, W., Oztireli, A.C., Zhang, H., Lin, Z., Shechtman, E., Benes, B.: Controllable shadow generation using pixel height maps. In: European Conference on Computer Vision. pp. 240-256. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 257 + ], + "type": "text", + "content": "51. Sheng, Y., Yu, Z., Ling, L., Cao, Z., Zhang, X., Lu, X., Xian, K., Lin, H., Benes, B.: Dr. bokeh: Differentiable occlusion-aware bokeh rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4515-4525 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 257, + 480, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 290 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 290 + ], + "type": "text", + "content": "52. Shi, Y., Huang, Z., Feng, S., Zhong, H., Wang, W., Sun, Y.: Masked label prediction: Unified message passing model for semi-supervised classification. arXiv preprint arXiv:2009.03509 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 289, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 332 + ], + "type": "text", + "content": "53. Song, Y., Zhang, Z., Lin, Z., Cohen, S., Price, B., Zhang, J., Kim, S.Y., Aliaga, D.: Objectstitch: Object compositing with diffusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18310-18319 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 333, + 480, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 480, + 376 + ], + "type": "text", + "content": "54. Song, Y., Zhang, Z., Lin, Z., Cohen, S., Price, B., Zhang, J., Kim, S.Y., Zhang, H., Xiong, W., Aliaga, D.: Imprint: Generative object compositing by learning identity-preserving representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8048-8058 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 376, + 480, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 480, + 397 + ], + "type": "text", + "content": "55. Tabata, S., Yoshihara, H., Maeda, H., Yokoyama, K.: Automatic layout generation for graphical design magazines. In: ACM SIGGRAPH 2019 Posters, pp. 1-2 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "type": "text", + "content": "56. Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 418, + 480, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 451 + ], + "type": "text", + "content": "57. Vanegas, C.A., Kelly, T., Weber, B., Halatsch, J., Aliaga, D.G., Müller, P.: Procedural generation of parcels in urban modeling. In: Computer graphics forum. vol. 31, pp. 681-690. Wiley Online Library (2012)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 450, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 472 + ], + "type": "text", + "content": "58. Velicković, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y.: Graph attention networks. arXiv preprint arXiv:1710.10903 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "text", + "content": "59. Velickovic, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y., et al.: Graph attention networks. stat 1050(20), 10-48550 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 480, + 525 + ], + "type": "text", + "content": "60. Wu, W., Fu, X.M., Tang, R., Wang, Y., Qi, Y.H., Liu, L.: Data-driven interior plan generation for residential buildings. ACM Transactions on Graphics (TOG) 38(6), 1-12 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 525, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 547 + ], + "type": "text", + "content": "61. Xie, H., Chen, Z., Hong, F., Liu, Z.: Citydreamer: Compositional generative model of unbounded 3d cities. arXiv preprint arXiv:2309.00610 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 547, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 579 + ], + "type": "text", + "content": "62. Xu, L., Xiangli, Y., Rao, A., Zhao, N., Dai, B., Liu, Z., Lin, D.: Blockplanner: City block generation with vectorized graph representation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5077-5086 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 579, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 480, + 601 + ], + "type": "text", + "content": "63. Yan, W., Zhang, Y., Abbeel, P., Srinivas, A.: Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "text", + "content": "64. Yang, C.F., Fan, W.C., Yang, F.E., Wang, Y.C.F.: Layouttransformer: Scene layout generation with conceptual and spatial diversity. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3732-3741 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "65. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "text", + "content": "He, L. & Aliaga, D." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 162 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "66. Zhang, X., Ma, W., Varinlioglu, G., Rauh, N., He, L., Aliaga, D.: Guided pluralistic building contour completion. The Visual Computer 38(9), 3205-3216 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 162 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 162 + ], + "type": "text", + "content": "67. Zheng, X., Qiao, X., Cao, Y., Lau, R.W.: Content-aware generative modeling of graphic design layouts. ACM Transactions on Graphics (TOG) 38(4), 1-15 (2019)" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "text", + "content": "COHO: Context-Sensitive City-Scale Hierarchical Urban Layout Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/COIN-Matting_ Confounder Intervention for Image Matting/7e767e14-ba96-40b5-847c-1a51f71239aa_content_list.json b/2024/COIN-Matting_ Confounder Intervention for Image Matting/7e767e14-ba96-40b5-847c-1a51f71239aa_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9fbd8d526abb12a6e278c0998af1b63749254180 --- /dev/null +++ b/2024/COIN-Matting_ Confounder Intervention for Image Matting/7e767e14-ba96-40b5-847c-1a51f71239aa_content_list.json @@ -0,0 +1,1881 @@ +[ + { + "type": "text", + "text": "COIN-Matting: Confounder Intervention for Image Matting", + "text_level": 1, + "bbox": [ + 248, + 140, + 754, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhaohe Liao $^{1}$ , Jiangtong Li $^{2}$ , Jun Lan $^{3}$ , Huijia Zhu $^{3}$ , Weiqiang Wang $^{3}$ , Li Niu $^{*1,4}$ , and Liqing Zhang $^{*1}$", + "bbox": [ + 235, + 210, + 767, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Shanghai Jiao Tong University \n{zhaoheliao, ustcnewly, zhang-lq}@sjtu.edu.cn \n2 Tongji University \njiangtongli@tongji.edu.cn \n3 Ant Group \n{yelan.lj, huijia.zhj, weiqiang.wwq}@antgroup.com \n4 miguo.ai", + "bbox": [ + 310, + 253, + 694, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Deep learning methods have significantly advanced the performance of image matting. However, dataset biases can mislead the matting models to biased behavior. In this paper, we identify the two typical biases in existing matting models, specifically contrast bias and transparency bias, and discuss their origins in matting datasets. To address these biases, we model the image matting task from the perspective of causal inference and identify the root causes of these biases: the confounders. To mitigate the effects of these confounders, we employ causal intervention through backdoor adjustment and introduce a novel model-agnostic cofounder intervened (COIN) matting framework. Extensive experiments across various matting methods and datasets have demonstrated that our COIN framework can significantly diminish such biases, thereby enhancing the performance of existing matting models.", + "bbox": [ + 259, + 388, + 743, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Image matting $\\cdot$ Intervention $\\cdot$ Dataset bias", + "bbox": [ + 261, + 583, + 643, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 623, + 375, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image matting aims at separating accurate foreground objects from natural images and precisely estimating the opacity on the boundary of the object. It serves as the basis for image editing, live-streaming, virtual meetings, and movie production [2,16,22,30,34]. Given an image $\\mathbf{I}$ , it is formulated as:", + "bbox": [ + 212, + 655, + 787, + 717 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {I} = \\boldsymbol {F} \\circ \\boldsymbol {\\alpha} + \\boldsymbol {B} \\circ (1 - \\boldsymbol {\\alpha}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 728, + 785, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "where $\\pmb{F}$ represents the foreground object, $\\pmb{B}$ denotes the background, $\\alpha \\in [0,1]$ is the opacity map, and $\\circ$ denotes element-wise multiplication. As shown in Eq. (1), solving the 7 unknown values in $\\{\\pmb{F},\\pmb{B},\\pmb{\\alpha}\\}$ from the given 3 values in $\\pmb{I}$ is a highly ill-posed problem. Therefore, some matting methods require", + "bbox": [ + 212, + 755, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*The corresponding authors.", + "bbox": [ + 236, + 825, + 431, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/14de91af92126b961678bc5b518355048c06a5977c6b42e37a6e943820fcfffc.jpg", + "image_caption": [ + "Image" + ], + "image_footnote": [], + "bbox": [ + 238, + 143, + 325, + 196 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/93b59149e5d997bb10c7a46c5ee6c0d9642edb8dca81ad0bd926c811feb11b36.jpg", + "image_caption": [ + "$\\alpha$" + ], + "image_footnote": [], + "bbox": [ + 328, + 143, + 411, + 196 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6fdae009ebc94decdb41c076f87db2416bd95ba101fd64fd99f5dcdfcdb3903d.jpg", + "image_caption": [ + "a" + ], + "image_footnote": [], + "bbox": [ + 413, + 143, + 500, + 198 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/245126f2418f0d5cde82433173da08c8bf60ffdbad94e5273e1d99acd8aa5ed1.jpg", + "image_caption": [ + "Image" + ], + "image_footnote": [], + "bbox": [ + 500, + 143, + 589, + 198 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3f25e11c133b67eabd6dc750d8faccf304318fc3be604b10933116464c5125f1.jpg", + "image_caption": [ + "$\\alpha$" + ], + "image_footnote": [], + "bbox": [ + 591, + 143, + 676, + 198 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5639fd6a889a84c138afdc7b02b1f7857f8e7ffd9e562e40f10c6100eb63cd18.jpg", + "image_caption": [ + "$\\hat{\\alpha}$" + ], + "image_footnote": [], + "bbox": [ + 678, + 143, + 764, + 198 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/896556478f5b782d812faf379a4f5fcafbaad205799e9372396f05f4e7cd0b1e.jpg", + "image_caption": [ + "Image" + ], + "image_footnote": [], + "bbox": [ + 238, + 224, + 326, + 291 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e429041996ebb57e5f0459a82a60ee49e6eaaa1773e78aa3c03c2a2d88dadc84.jpg", + "image_caption": [ + "$\\alpha$" + ], + "image_footnote": [], + "bbox": [ + 328, + 224, + 411, + 291 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/388d52e96579fad8af456bf4f2dfdacad8019d3bb8ea8364a38cb13ab48e9396.jpg", + "image_caption": [ + "(a) The contrast bias.", + "a", + "(b) The transparency bias." + ], + "image_footnote": [], + "bbox": [ + 413, + 224, + 500, + 291 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/999460dfdc2888e0ccbc7f6b71320b7443a490e69a38c13ea22ee5e0e7eef765.jpg", + "image_caption": [ + "Image", + "Fig. 1: The observed contrast bias and transparency bias for MatteFormer [25] on the Compositional-1K dataset [40]. $\\alpha$ and $\\hat{\\alpha}$ denote the ground truth and predicted alpha mattes, respectively. Images are best viewed when zoomed in." + ], + "image_footnote": [], + "bbox": [ + 500, + 224, + 586, + 291 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7bdb3f12bab41d7960f98ba9f74084120324d9e25c35d9241bfa572b3471854a.jpg", + "image_caption": [ + "$\\alpha$" + ], + "image_footnote": [], + "bbox": [ + 589, + 224, + 674, + 291 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8a8e32032fce451a9455edadca0fcd66b83cde3e4cc4e2966cf615d6fb477275.jpg", + "image_caption": [ + "$\\hat{\\alpha}$" + ], + "image_footnote": [], + "bbox": [ + 676, + 224, + 763, + 291 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "additional user input as guidance, leading to a classification of matting methods based on whether and what type of user input is used. Trimap-based methods [40] are the most widely used, but obtaining a trimap can be challenging in some scenarios. Therefore, more flexible guidance, such as background [16,30] and user interaction [6,37,38], have been explored. Moreover, some methods [12,38] can also work without user guidance, which is concluded as guidance-free methods.", + "bbox": [ + 212, + 401, + 784, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The rapid advances in deep learning have significantly boosted the performance of image matting. However, the matting dataset may contain biases, which may further lead to the biased behavior of image matting models. We identify two primary types of bias in existing models: contrast bias and transparency bias. To illustrate these biases, we present some results of a state-of-the-art (SOTA) model, namely MatteFormer [25], in Fig. 1. The contrast bias refers to the correlation between the contrast and the predicted alpha value. Specifically, the model tends to predict lower alpha values in areas with low contrast between foreground and background, as highlighted by the red boxes in Fig. 1a. Regarding the transparency bias, it is observed that the background texture tends to emerge in the predicted alpha matte for mid-transparent foregrounds. As highlighted by the red boxes in Fig. 1b, the predicted alpha for mid-transparent area contains the background texture (i.e., the texture of the rider and building for the left and right examples, respectively).", + "bbox": [ + 212, + 492, + 785, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Such biased behavior of models can be explained by the biases present in the dataset. For the contrast bias, we calculate the alpha value at different contrast levels within the training dataset as shown in Fig. 2a. We observe that the mean alpha level in low contrast areas (i.e., contrast $< 20$ in the CIEDE2000 color difference formula [31]) is significantly lower than that in other areas. Such dataset bias directly contributes to the observed contrast bias as the model may falsely learn to predict alpha from contrast. Additionally, the number of samples in the low contrast area is significantly fewer than in other areas, making the model more easily led by such dataset bias in these areas. Regarding the", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Liao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e6322f0919a4f7911d908e23e2c46ae4ef906c65778ef225648025d3c8370208.jpg", + "image_caption": [ + "(a) The contrast bias in the dataset." + ], + "image_footnote": [], + "bbox": [ + 256, + 145, + 470, + 241 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/58238cbadfb1502d2decc14f05da253c660a2cdd89bb59daabc292f0d8c51c33.jpg", + "image_caption": [ + "(b) The transparency bias in the dataset.", + "Fig. 2: The observe biases in matting datasets." + ], + "image_footnote": [], + "bbox": [ + 511, + 146, + 728, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "transparency bias, we analyze the distribution of transparency levels within the training dataset in Fig. 2b. We observe that most foreground objects are either highly salient or very opaque. Therefore, the alpha matte is expected to be either close to zero or exhibit textures in the image. For the mid-transparency area, whose alpha matte clearly differs from zero, the model is misled to focus on the texture of the image. Thus, the texture of the background in such areas may appear in the alpha matte, explaining the appearance of transparency bias.", + "bbox": [ + 212, + 320, + 784, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these biases, we model the image matting task from the perspective of causal inference, and utilize the power of intervention to reduce the influence of the biases. The causal graph for image matting is depicted in Fig. 3a, where $\\mathcal{I},\\mathcal{F},\\mathcal{B}$ , and $\\mathcal{A}$ are variables representing the image, foreground, background, and alpha matte respectively, while $\\mathcal{C}$ and $\\mathcal{T}$ are confounder variables denoting the contrast and the transparency of foreground. The notation $\\mathcal{B} \\gets \\mathcal{I} \\to \\mathcal{F}$ suggests that an image can be decomposed into foreground and background, underpinning the basic assumption of image matting. $\\mathcal{F} \\to \\mathcal{A}$ implies that the alpha matte is directly influenced by the foreground object, indicating that changes in the background should not affect the foreground's representation. The confounder variables, representing the common causes affecting multiple variables in the foreground, background, and image, serve as the source of biases. Typically, these can include properties of visual objects. Given that both the foreground and background consist of objects, these properties typically cause the presentation of the foreground, background, and images are presented.", + "bbox": [ + 212, + 429, + 785, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, directly measuring all possible confounders may be impractical, as we can not enumerate all properties of visual objects. Nevertheless, based on observed dataset biases, we can identify the most typical confounders. As highlighted in Fig. 3a, we focus on two typical confounders: contrast $\\mathcal{C}$ and transparency $\\mathcal{T}$ , which underlie the observed contrast bias and transparency bias, respectively. Contrast measures the color difference between foreground and background objects, while transparency refers to the opacity of the foreground. Therefore, we conclude $\\mathcal{F} \\gets \\mathcal{T} \\to \\mathcal{I}$ and $\\mathcal{C} \\to \\mathcal{X}$ for all $\\mathcal{X} \\in \\{\\mathcal{F}, \\mathcal{I}, \\mathcal{B}\\}$ by their definition. The dotted line indicates a potential correlation between $\\mathcal{C}$ and $\\mathcal{T}$ , suggesting that we cannot definitively assert that $\\mathcal{C} \\perp \\mathcal{T}$ . While existing matting methods generally focus on maximizing the likelihood $P(\\mathcal{A}|\\mathcal{I})$ , the presence of backdoor paths $\\mathcal{B} \\gets \\mathcal{C} \\to \\mathcal{F} \\to \\mathcal{A}$ , $\\mathcal{B} \\gets \\mathcal{I} \\gets \\mathcal{C} \\to \\mathcal{F} \\to \\mathcal{A}$ , and $\\mathcal{B} \\gets \\mathcal{I} \\gets \\mathcal{T} \\to$", + "bbox": [ + 212, + 657, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "COIN-Matting: Confounder Intervention for Image Matting", + "bbox": [ + 333, + 114, + 732, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/319a89db7530423775c1f64f267bd014cdcb0f50c2627b014e1cd3ed8a42c83a.jpg", + "image_caption": [ + "(a) The causal graph." + ], + "image_footnote": [], + "bbox": [ + 312, + 146, + 446, + 215 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5aa8e90be0b0a511691bd86a7b1475e95476982e5fe4be345a976d66d1ebdf43.jpg", + "image_caption": [ + "(b) The backdoor adjustment.", + "Fig. 3: The causal graph and causal intervention for image matting." + ], + "image_footnote": [], + "bbox": [ + 540, + 146, + 676, + 214 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{F} \\to \\mathcal{A}$ results in $\\mathcal{B} \\not\\sqcup \\mathcal{A}|\\mathcal{I}$ . This lack of conditional d-separation between $\\mathcal{B}$ and $\\mathcal{A}$ elucidates the co-occurrence of background textures in specific transparency regions and the contrast-induced alpha bias.", + "bbox": [ + 212, + 282, + 784, + 328 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To remove the effect of these confounders, we propose our model-agnostic cofounder intervened (COIN) matting framework. Specifically, we explore the do-calculus to intervene in the influence of $\\mathcal{T}$ and $\\mathcal{C}$ via backdoor adjustment, as shown in Fig. 3b. The do-calculus actively assigns values to variables, rather than passively observing them from the data. Therefore, the do-calculus is an intervention on the influence of their direct causes, which is used for decofounding. While the ideal way to calculate the do-operator involves collecting all images with varying contrast and transparency into the training data, this approach is evidently impractical. Furthermore, the high cost of labeling the alpha matte makes it exceedingly difficult to even approximate the aforementioned intervention \"physically\". To address this issue, we apply backdoor adjustment to calculate $P(\\mathcal{A}|do(\\mathcal{F},\\mathcal{B},\\mathcal{I}))$ , as described in Secs. 3.2 to 3.4. By actively assigning values to $\\mathcal{F},\\mathcal{B}$ , and $\\mathcal{I}$ , we block backdoor paths and magnify the influence of cofounders, thus enhancing the robustness of the matting results against the previously mentioned dataset biases. By applying our framework to state-of-the-art (SOTA) image matting methods with different types of user guidance (i.e., MatteFormer [25], DIIM [6], and dugMatting [38]) and two datasets (i.e., Comp1K and Distinct-646), we verified the effectiveness of our framework through extensive experiments. Our contributions are summarized as follows:", + "bbox": [ + 212, + 329, + 787, + 616 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We examine the biases in matting methods, analyzing how dataset biases cause them and affect alpha matte prediction.", + "2. We introduce a novel and model-agnostic confounder-intervened matting framework that employs causal intervention to sever the influence of confounders, thereby mitigating the aforementioned biases.", + "3. Through comprehensive experiments on various matting methods and multiple datasets, we demonstrate that our framework significantly enhances the performance of SOTA image matting methods." + ], + "bbox": [ + 220, + 621, + 784, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 760, + 387, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Image Matting", + "text_level": 1, + "bbox": [ + 215, + 787, + 385, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As described in Sec. 1, image matting presents a highly ill-posed problem. The solutions can be classified based on the user input prior they rely upon.", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Liao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Among the prior-based image matting methods, the trimap is the most widely utilized prior [1,3,4,9,14,18,21,23,23,25,33,40-42]. Since [40] introduced the first well-known deep learning-based image matting method, featuring an encoder-decoder architecture, this framework has been extensively adopted and improved by subsequent methods. Researchers have enhanced this architecture from various perspectives, including generative adversarial training [23], matting with indices [21], affinity-based methods [3, 14], context-aware matting [9], the introduction of vision transformers [1, 25, 41], prior alignment [18], expansion to ultra-high resolution via sparsity [33], and data augmentation [4, 42].", + "bbox": [ + 212, + 146, + 782, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, trimap is not always available for real-world applications. Thus, recent researchers seek for more flexible and accessible user inputs [5, 6, 8, 13, 16, 30, 37, 38, 43]. Background matting [30] uses the background image as the prior, which can be more accessible for scenes like virtual meeting and livestreaming. It is further enhanced by recursive excitation [5] and extended into real-time version [16]. The coarse mask is used instead of trimap in [43] which shows better penitential of cooperating with the segmentation models [11]. The user feedbacks, such as clicks [6], scribbles [8], their combination [37], proposal selection [38] and even natural language [13], are also informative for matting.", + "bbox": [ + 212, + 282, + 784, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.2 Causal Inference", + "text_level": 1, + "bbox": [ + 215, + 441, + 401, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Causal inference [27, 29] plays a pivotal role in analyzing dataset bias and mitigating counter-causal correlations. It can be categorized into two main types: deconfounding and counterfactual inference. Serving as a fundamental statistical tool, it finds widespread application across various fields, such as image segmentation [44], image classification [35], image and video question answering [10, 15, 24], dialog systems [45], and recommendation systems [36]. Typically, causal-related methods conceptualize the task within a causal graph, where nodes symbolize variables, and edges denote their causal interactions. By examining the causal relations among variables, the causal graph facilitates identifying confounders and backdoor paths, thereby enabling the control of biases.", + "bbox": [ + 212, + 465, + 787, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 641, + 380, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section outlines our intervention approach to mitigating biases introduced by confounders, as discussed in Sec. 1. In Sec. 3.1, we examine the role of confounder variables within the causal graph (i.e., Fig. 3), illustrating their impact on the performance of mating models through mathematical analysis. In Sec. 3.2, the concept of backdoor adjustment is introduced to counteract biases. In Sec. 3.3, we discuss the construction of space representations for each variable in the causal graph to facilitate backdoor adjustment implementation. In Sec. 3.4, the calculation of expectations, as derived in Sec. 3.2, is explained to execute the causal intervention. Finally, in Sec. 3.5, we integrate the above components into our model-agnostic COIN mating framework. Throughout this section, we denote variables by calligraphic letters (e.g., $\\mathcal{A}$ , $\\mathcal{F}$ , $\\mathcal{I}$ ), feature spaces", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "COIN-Matting: Confounder Intervention for Image Matting", + "bbox": [ + 333, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "by boldface capital letters (e.g., $\\mathbf{A}$ , $\\mathbf{F}$ , $\\mathbf{I}$ ), and feature vectors by boldface lowercase letters (e.g., $\\mathbf{a}$ , $\\mathbf{f}$ , $\\mathbf{i}$ ). Additional introduction of background and more detailed derivations are provided in the supplementary material.", + "bbox": [ + 212, + 146, + 787, + 194 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Confounders and the Backdoor Paths", + "text_level": 1, + "bbox": [ + 214, + 217, + 571, + 233 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the task of image matting, confounders are variables that are the direct cause of at least two of the following: the foreground, background, and the image. As illustrated in Fig. 3a, the observational probability can be expressed as", + "bbox": [ + 212, + 246, + 787, + 292 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP (\\mathcal {A} | \\mathcal {I}) = \\sum_ {\\boldsymbol {t}} \\sum_ {\\boldsymbol {c}} P (\\mathcal {A} | \\mathcal {I}, \\boldsymbol {t}, \\boldsymbol {c}) P (\\boldsymbol {c}, \\boldsymbol {t} | \\mathcal {I}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 305, + 785, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\pmb{t}$ and $\\pmb{c}$ represent the split of confounder $\\mathcal{C}$ and $\\mathcal{T}$ , respectively. Given that the observed distribution $P(\\pmb{c},\\pmb{t}|\\mathcal{I})$ significantly deviates from a uniform distribution (as depicted in Fig. 2), a specific partition of $\\langle \\mathcal{C},\\mathcal{T}\\rangle$ predominantly influences the observed probability $P(\\mathcal{A}|\\mathcal{I})$ through a large $P(\\pmb{c},\\pmb{t}|\\mathcal{I})$ . For instance, in the training dataset, regions with high contrast and low transparency are prevalent, leading to their dominance in the observed probability and observed biases. Specifically, in low-contrast regions, the alpha values are typically to be lower, causing the observed contrast bias. Furthermore, the model is trained to discern detailed textures in regions with opaque foregrounds and to assign nearly zero alpha values in highly transparent areas. Since these conditions consist most of the dataset, $P(\\pmb{c},\\pmb{t}|\\mathcal{I})$ with extreme values of $\\pmb{t}$ significantly contributes to $P(\\mathcal{A}|\\mathcal{I})$ . Consequently, the model tends to extract highly textured patterns in medium transparency areas as in opaque foreground areas, where alpha values are apparently not as minimal as in transparent regions. Such background textures in the predicted alpha matte are concluded as transparency bias.", + "bbox": [ + 212, + 345, + 787, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By adopting a causal perspective on Fig. 3a, we identify the presence of backdoor paths introduced by the confounder variables $\\mathcal{C}$ and $\\mathcal{T}$ . In particular, the backdoor paths $\\mathcal{A} \\gets \\mathcal{F} \\gets \\mathcal{C} \\to \\mathcal{B}$ , $\\mathcal{B} \\gets \\mathcal{I} \\gets \\mathcal{C} \\to \\mathcal{F} \\to \\mathcal{A}$ , and $\\mathcal{A} \\gets \\mathcal{F} \\gets \\mathcal{T} \\to \\mathcal{I} \\to \\mathcal{B}$ result in $\\mathcal{A} \\not\\perp \\mathcal{B}|\\mathcal{I}$ , elucidating the influence of the background on the predicted alpha matte. More specifically, the former two paths and the latter are directly accountable for the contrast bias and transparency bias, respectively. To intervene the effects of the confounders and sever the backdoor paths they establish, we employ the backdoor adjustment [27,29] as detailed in Sec. 3.2.", + "bbox": [ + 212, + 574, + 789, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Causal Intervention via Backdoor Adjustment", + "text_level": 1, + "bbox": [ + 214, + 720, + 640, + 737 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The do-calculus [27, 28] signifies active assignment of values to variables, distinguishing it from passive observation within the data, which constitutes intervention on the causal factors of variables. For brevity, we denote $P(do(\\mathcal{X} = x))$ as $P(do(\\mathcal{X}))$ , implying active assignment of a value to variable $\\mathcal{X}$ . As illustrated in Fig. 3b, to sever all incoming causal links of $\\mathcal{I}, \\mathcal{F}, \\mathcal{B}$ (i.e., nullifying the influence of $\\mathcal{C}$ and $\\mathcal{T}$ ), we focus on computing $P(\\mathcal{A}|do(\\mathcal{I},\\mathcal{F},\\mathcal{B}))$ . Given that $\\mathcal{T}$ and", + "bbox": [ + 212, + 750, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Liao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathcal{C}$ block all specified backdoor paths, we can derive:", + "bbox": [ + 214, + 146, + 584, + 161 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} P (\\mathcal {A} | d o (\\mathcal {I}, \\mathcal {F}, \\mathcal {B})) = \\sum_ {\\boldsymbol {t}} \\sum_ {\\boldsymbol {c}} P (\\mathcal {A} | d o (\\mathcal {I}, \\mathcal {F}, \\mathcal {B}), \\boldsymbol {t}, \\boldsymbol {c}) P (\\boldsymbol {c}, \\boldsymbol {t} | d o (\\mathcal {I}, \\mathcal {F}, \\mathcal {B})) \\\\ = \\sum_ {t} ^ {t} \\sum_ {c} ^ {c} P (\\mathcal {A} | i, \\boldsymbol {f}, \\boldsymbol {b}, t, c) P (\\boldsymbol {c}, t). \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 263, + 171, + 784, + 234 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Moreover, based on Fig. 3b, we observe $\\mathcal{A} \\perp \\mathbf{b}|(\\mathbf{f},\\mathbf{i},\\mathbf{c},\\mathbf{t})$ and $\\mathcal{A} \\perp \\mathbf{i}|(\\mathbf{f},\\mathbf{c},\\mathbf{t})$ . Thus, the equation simplifies to:", + "bbox": [ + 214, + 243, + 782, + 273 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nP (\\mathcal {A} | d o (\\mathcal {I}, \\mathcal {F}, \\mathcal {B})) = \\sum_ {\\boldsymbol {t}} \\sum_ {\\boldsymbol {c}} P (\\mathcal {A} | \\boldsymbol {f}, \\boldsymbol {t}, \\boldsymbol {c}) P (\\boldsymbol {t}, \\boldsymbol {c}) = \\mathbb {E} _ {\\boldsymbol {t}, \\boldsymbol {c}} [ P (\\mathcal {A} | \\boldsymbol {f}, \\boldsymbol {t}, \\boldsymbol {c}) ]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 264, + 284, + 784, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Therefore, to predict the alpha matte with confounders intervened, we need to model the relevant spaces and calculate the expectations in Eq. (4), which are further discussed in Sec. 3.3 and Sec. 3.4, respectively.", + "bbox": [ + 214, + 321, + 782, + 366 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "However, to calculate the expectation in Eq. (4), obtaining the foreground representation is also necessary. Although the only input is the image (and the possible user guidance), directly inferring the foreground representation from the image representation is biased, as the confounders similarly facilitate backdoor paths to $\\mathcal{F}$ as to $\\mathcal{A}$ . Specifically, the backdoor paths $\\mathcal{F} \\gets \\mathcal{C} \\to \\mathcal{B}$ , $\\mathcal{B} \\gets \\mathcal{I} \\gets \\mathcal{C} \\to \\mathcal{F}$ , and $\\mathcal{F} \\gets \\mathcal{T} \\to \\mathcal{I} \\to \\mathcal{B}$ lead to $\\mathcal{F} \\not\\perp \\mathcal{B}|\\mathcal{I}$ , indicating that the background's transparency and contrast similarly affect the foreground as influencing the prediction of the alpha matte. To address this issue, we also conduct an intervention for predicting the foreground, yielding:", + "bbox": [ + 214, + 367, + 784, + 503 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} P (\\mathcal {F} | d o (\\mathcal {I}, \\mathcal {B})) = \\sum_ {\\boldsymbol {t}} \\sum_ {\\boldsymbol {c}} P (\\mathcal {F} | \\boldsymbol {i}, \\boldsymbol {b}, \\boldsymbol {t}, \\boldsymbol {c}) P (\\boldsymbol {c}, \\boldsymbol {t}) \\\\ = \\sum_ {t} ^ {t} \\sum_ {c} ^ {c} P (\\mathcal {F} | \\boldsymbol {i}, \\boldsymbol {t}, \\boldsymbol {c}) P (\\boldsymbol {c}, \\boldsymbol {t}) = \\mathbb {E} _ {\\boldsymbol {c}} \\mathbb {E} _ {\\boldsymbol {t}} [ P (\\mathcal {F} | \\boldsymbol {i}, \\boldsymbol {t}, \\boldsymbol {c}) ], \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 274, + 512, + 784, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the multi-scale image representation $\\pmb{i} = \\langle \\pmb{i}_1, \\pmb{i}_2, \\dots, \\pmb{i}_n \\rangle$ is parameterized by the matting backbone, aligned with the scales of $\\pmb{C}$ and $\\pmb{T}$ as described in Sec. 3.3. The expectations are calculated as detailed in Sec. 3.4.", + "bbox": [ + 214, + 585, + 782, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Space Representations", + "text_level": 1, + "bbox": [ + 214, + 650, + 446, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Transparency Space $\\pmb{T}$ . Given the importance of multi-scale information in image matting, we divide the transparency space into several sub-spaces across different scales. Specifically, the transparency space $\\pmb{T} = \\langle T_1, T_2, \\dots, T_n \\rangle$ , with $\\pmb{T}_i$ representing the transparency at the $i$ -th scale, and $n$ indicating the total number of sub-spaces. The $i$ -th scale denotes the transparency of the image patches sized $2^i \\times 2^i$ . Considering the average alpha value as an approximation for the \"transparency level\" of an area, we define $\\pmb{T}_i$ as the average alpha value of pixels within the $i$ -th scale patches. This formulation allows us to discretize the transparency sub-spaces further by categorizing the transparency level into $K_t$ distinct classes, given its bounded nature within [0, 1]. The construction of the feature space entails the following steps:", + "bbox": [ + 214, + 672, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "COIN-Matting: Confounder Intervention for Image Matting", + "bbox": [ + 333, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Utilize a pretrained image feature extractor to derive the patch features at the $i$ -th scale, for each $i \\in \\{1, \\dots, n\\}$ .", + "2. Determine the transparency level category $k \\in \\{1, \\dots, K_t\\}$ for every patch.", + "3. Compute the average patch representation for all patches within the $k$ -th category as the representation of this transparency category." + ], + "bbox": [ + 220, + 146, + 782, + 219 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hence, we formulate the feature space $\\pmb{T}_i\\in \\mathbb{R}^{d_i\\times K_t}$ , where $d_{i}$ signifies the dimensionality of the feature extracted at the $i$ -th scale.", + "bbox": [ + 212, + 223, + 782, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Contrast Space $C$ . Similarly to the transparency space, the contrast space is partitioned into multiple sub-spaces across various scales. We define $C = \\langle C_1, C_2, \\ldots, C_n \\rangle$ , with each $C_i$ representing contrast at the $i$ -th scale, and $n$ marking the total number of sub-spaces. Each $i$ -th scale denotes the contrast of the foreground and background within $2^i \\times 2^i$ image patches.", + "bbox": [ + 212, + 255, + 784, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantifying contrast, however, presents more complexity than measuring transparency. Initially, we estimate the color of a patch using the average color vector of all pixels within it in the color space. Subsequently, the contrast between foreground and background patches is determined by the color distance between their average color vectors. Given the multifaceted nature of color perception, we employ the CIEDE2000 color difference formula [31] for calculating color distance, as it closely aligns with human visual perception of color differences. This formula, incorporating lightness, chroma, and hue, yields a distance within the range of [0, 100]. Following this approach, we categorize the contrast level into $K_{c}$ discrete classes and construct the feature space $C_i \\in \\mathbb{R}^{d_i \\times K_c}$ , mirroring the methodology used for the transparency space. The primary adjustment involves transitioning from transparency to contrast categories for the classification.", + "bbox": [ + 212, + 330, + 782, + 510 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Joint Confounder Space $\\mathbf{Z}$ . Given that $\\mathcal{T}$ and $\\mathcal{C}$ may not be independent, we cannot decompose the joint expectation in Eq. (4) into separate expectations for each. Therefore, it becomes necessary to model the joint space of $\\mathbf{T}$ and $\\mathbf{C}$ . An effective approach is to define the joint space as the Cartesian product of the sub-spaces. Consequently, the joint confounder space $\\mathbf{Z} = \\langle \\mathbf{Z}_1, \\mathbf{Z}_2, \\dots, \\mathbf{Z}_n \\rangle = \\langle \\mathbf{T}_1 \\times \\mathbf{C}_1, \\mathbf{T}_2 \\times \\mathbf{C}_2, \\dots, \\mathbf{T}_n \\times \\mathbf{C}_n \\rangle$ , where each $\\mathbf{T}_i \\times \\mathbf{C}_i \\in \\mathbb{R}^{d_i \\times (K_t \\cdot K_c)}$ represents a multidimensional space formed by the Cartesian product of transparency and contrast spaces at the $i$ -th scale. Thus, Eqs. (4) and (5) are simplified as following:", + "bbox": [ + 212, + 511, + 782, + 632 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} P(\\mathcal{A}|do(\\mathcal{I},\\mathcal{F},\\mathcal{B})) = \\sum_{\\boldsymbol {z}}P(\\mathcal{A}|\\boldsymbol {f},\\boldsymbol {z})P(\\boldsymbol {z}) = \\mathbb{E}_{\\boldsymbol{z}}[P(\\mathcal{A}|\\boldsymbol {f},\\boldsymbol {z})], \\\\ P (\\mathcal {F} | d o (\\mathcal {I}, \\mathcal {B})) = \\sum_ {\\boldsymbol {z}} P (\\mathcal {F} | \\boldsymbol {i}, \\boldsymbol {z}) P (\\boldsymbol {z}) = \\mathbb {E} _ {\\boldsymbol {z}} [ P (\\mathcal {F} | \\boldsymbol {i}, \\boldsymbol {z}) ], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 302, + 651, + 700, + 713 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $z = \\langle \\pmb {c},\\pmb {t}\\rangle$ represents the split of joint observed confounders $\\mathcal{Z} = \\langle \\mathcal{T},\\mathcal{C}\\rangle$", + "bbox": [ + 212, + 719, + 772, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Expectations", + "text_level": 1, + "bbox": [ + 214, + 752, + 370, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The expectations are calculated over the joint confounder space $\\mathbf{Z}$ . Utilizing the Normalized Weighted Geometric Mean (NWGM) approach [32,39], we reformulate the expectation as follows:", + "bbox": [ + 212, + 773, + 782, + 818 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nP (\\mathcal {A} | d o (\\mathcal {I}, \\mathcal {F}, \\mathcal {B})) = \\mathbb {E} _ {\\boldsymbol {z}} P (\\mathcal {A} | \\boldsymbol {f}, \\boldsymbol {z}) = \\mathbb {E} _ {\\boldsymbol {z}} \\sigma (g (\\boldsymbol {f}, \\boldsymbol {z})) \\approx \\sigma (g (\\boldsymbol {f}, \\mathbb {E} _ {\\boldsymbol {z}} [ \\boldsymbol {z} ])), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 263, + 825, + 782, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Liao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0d5db88a6f0a52bf1cca642701e284c995386ef7b531195b5404e588304da10d.jpg", + "image_caption": [ + "Fig. 4: The COIN framework. $\\pmb{i}$ and $\\pmb{f}$ denote the image feature and foreground feature. $T, C$ and $Z$ are transparency space, contrast space, and their joint confounder space respectively, which are constructed in Sec. 3.3. Spaces corresponding to the confounder variable are marked gray. EXP represents the expectations described in Sec. 3.4. $\\sigma(\\cdot)$ and $g(\\cdot)$ are ignored for simplicity. $\\odot$ represents the dot production." + ], + "image_footnote": [], + "bbox": [ + 212, + 145, + 781, + 270 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\sigma(\\cdot)$ denotes a sigmoid function and $g(\\cdot)$ represents a linear transformation for each sub-space. Similarly, for the foreground parameterization:", + "bbox": [ + 212, + 375, + 787, + 406 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nP (\\mathcal {F} | d o (\\mathcal {I}, \\mathcal {B})) = \\mathbb {E} _ {z} P (\\mathcal {F} | \\boldsymbol {i}, z) = \\mathbb {E} _ {z} g (\\boldsymbol {i}, z) = g (\\boldsymbol {i}, \\mathbb {E} _ {z} [ z ]). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 297, + 412, + 785, + 429 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This approach shifts the expectations to the feature level. To prevent model collapse, we use $\\mathbb{E}_{\\boldsymbol{z}|\\mathcal{F}}$ and $\\mathbb{E}_{\\boldsymbol{z}|\\mathcal{I}}$ as approximations for $\\mathbb{E}_z$ respectively. To emulate these feature-level expectations, we leverage a sub-space-wise attention mechanism. Take $\\mathbb{E}_{\\boldsymbol{z}|\\mathcal{I}}[\\boldsymbol{z}]$ as an example. As spaces $Z_{i}$ are heterogeneous with the feature spaces of $\\pmb{i}_j$ when $i\\neq j$ , we apply the attention mechanism for each scale respectively. Therefore, the expectation is", + "bbox": [ + 212, + 435, + 787, + 526 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\boldsymbol {z} \\mid \\mathcal {I}} [ \\boldsymbol {z} ] = \\left\\langle \\mathbb {E} _ {\\boldsymbol {z} _ {1} \\mid \\mathcal {I} _ {1}} [ \\boldsymbol {z} _ {1} ], \\mathbb {E} _ {\\boldsymbol {z} _ {2} \\mid \\mathcal {I} _ {2}} [ \\boldsymbol {z} _ {2} ], \\dots , \\mathbb {E} _ {\\boldsymbol {z} _ {n} \\mid \\mathcal {I} _ {n}} [ \\boldsymbol {z} _ {n} ] \\right\\rangle . \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 532, + 785, + 550 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $n$ is the number of sub-spaces. For attention on image patches at the $i$ -th scale, the input is the contrast sub-space $Z_{i} \\in \\mathbb{R}^{d_{i} \\times (K_{t} \\cdot K_{c})}$ and the image patch representation $\\hat{\\pmb{i}}_i \\in \\mathbb{R}^{d_i}$ parameterized by the matting backbone, where $d_{i}$ is the dimension of the latent representations. During such attention, the attention score emulates the distribution of $P(\\pmb{z}|\\mathcal{I})$ , thus the weighted sum on values (i.e., $z$ ) emulates the expectation $\\mathbb{E}_{\\pmb{z}|\\mathcal{I}}[\\pmb{z}]$ . The mathematical formulation is given by:", + "bbox": [ + 212, + 556, + 787, + 648 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\pmb {h e a d} _ {i} ^ {k} = \\mathrm {s o f t m a x} (\\hat {\\pmb {i}} _ {i} ^ {k} \\pmb {W} _ {i, 1} ^ {k} (\\pmb {Z} _ {i} \\pmb {W} _ {i, 2} ^ {k}) ^ {T} / \\sqrt {d}) (\\pmb {Z} _ {i} \\pmb {W} _ {i, 3} ^ {k})\n$$\n", + "text_format": "latex", + "bbox": [ + 272, + 657, + 635, + 676 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {z} _ {\\mathcal {I} _ {i}} = \\mathbb {E} _ {\\boldsymbol {z} _ {i} \\mid \\mathcal {I} _ {i}} [ \\boldsymbol {z} _ {i} ] = \\operatorname {C o n c a t} \\left(\\boldsymbol {h e a d} _ {i} ^ {1}, \\boldsymbol {h e a d} _ {i} ^ {2}, \\dots , \\boldsymbol {h e a d} _ {i} ^ {N _ {h}}\\right) \\boldsymbol {W} _ {i, o} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 299, + 680, + 784, + 699 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {z} _ {\\mathcal {I}} = \\operatorname {E X P} (\\hat {\\boldsymbol {i}}, \\boldsymbol {Z}) = \\mathbb {E} _ {\\boldsymbol {z} | \\mathcal {I}} [ \\boldsymbol {z} ] = \\left\\langle \\bar {z} _ {\\mathcal {I} _ {1}}, \\bar {z} _ {\\mathcal {I} _ {2}}, \\dots , \\bar {z} _ {\\mathcal {I} _ {n}} \\right\\rangle ,\n$$\n", + "text_format": "latex", + "bbox": [ + 305, + 702, + 647, + 720 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\pmb{W}_{i,1}^{k}, \\pmb{W}_{i,2}^{k}, \\pmb{W}_{i,3}^{k}, \\pmb{W}_{i,o}$ are trainable parameters, $N_{h}$ represents the number of heads, and $\\mathrm{Concat}(\\cdot)$ is the concatenation operation. This process effectively captures the expectations across different confounder sub-spaces.", + "bbox": [ + 212, + 723, + 787, + 771 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 COIN Matting Framework", + "text_level": 1, + "bbox": [ + 214, + 787, + 483, + 803 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The overall structure of our COIN framework is shown in Fig. 4. The confounder space $\\pmb{T}$ is similarly constructed as $\\mathbf{C}$ , as described in Sec. 3.3. The mat", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "COIN-Matting: Confounder Intervention for Image Matting", + "bbox": [ + 333, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ting backbone is employed to parameterize the image representation $\\hat{\\pmb{i}}$ . Rather than directly predicting $P(\\mathcal{A}|\\mathcal{I})$ as the matting backbone does, we engage in confounder intervention through backdoor adjustment to address the issues of contrast bias and transparency bias. The backdoor adjustment is implemented by the feature-level expectations. Specifically, we first process the joint space $\\pmb{Z}$ alongside the image representation $\\pmb{i}$ to compute $\\mathbb{E}_{z|\\mathcal{I}}[z]$ via attention. This expectation is subsequently concatenated with the image representations to yield the foreground representation. Similarly, the foreground representation is used to calculate $\\mathbb{E}_{z|\\mathcal{F}}[z]$ with joint confounder space $\\pmb{Z}$ , and the expectations are further concatenated with the foreground representation to produce the alpha matte at various scales. Thus, we can express the overall framework as follows:", + "bbox": [ + 212, + 145, + 787, + 313 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {i}} = \\text {M a t t i n g B a c k b o n e} (\\boldsymbol {I}); \\quad \\bar {z} _ {\\mathcal {I}} = \\operatorname {E X P} (\\hat {\\boldsymbol {i}}, \\boldsymbol {Z}); \\quad \\hat {\\boldsymbol {f}} = g _ {f} (\\hat {\\boldsymbol {i}}, \\bar {z} _ {\\mathcal {I}}); \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 274, + 323, + 785, + 349 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {z} _ {\\mathcal {F}} = \\operatorname {E X P} (\\hat {\\boldsymbol {f}}, \\boldsymbol {Z}); \\quad \\hat {\\alpha} = P (\\mathcal {A} | d o (\\mathcal {I}, \\mathcal {F}, \\mathcal {B})) = \\sigma (g _ {\\alpha} (\\hat {\\boldsymbol {f}}, \\bar {z} _ {\\mathcal {F}})),\n$$\n", + "text_format": "latex", + "bbox": [ + 261, + 347, + 681, + 363 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $g_{f}(\\cdot)$ and $g_{\\alpha}(\\cdot)$ represent per-sub-space linear transformations, and $\\mathrm{EXP}(\\cdot, \\cdot)$ denotes the feature-level expectations described in Sec. 3.4.", + "bbox": [ + 212, + 375, + 782, + 402 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "During the training of this framework, we supervise the alpha matte at each scale, similar to the approach taken by the backbone. Furthermore, we supervise the attention score, which approximates the probability distribution of the confounders when calculating expectations, with the ground truth categories of contrast and transparency. It is important to note that a closer category index of contrast and transparency signifies a more similar meaning, as these are discretized from the continuous values of color distance and transparency. Therefore, we employ Gaussian label smoothing prior to the cross-entropy for supervising the attention score. Additional details regarding the framework can be found in the supplementary material.", + "bbox": [ + 212, + 405, + 787, + 556 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 579, + 375, + 595 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we evaluate our COIN framework on extensive image matting baselines across various datasets. Specifically, we integrate our framework with SOTA image matting methods across all three types of matting methods: trimap-based methods (i.e., MatteFormer [25]), flexible guidance-based methods (i.e., DIIM [6] and dugMatting [38]) and guidance-free methods (i.e., dugMatting without human interaction [38]). We evaluate the improvement of our framework on Composition-1k [40], and Distinctions-646 [43] datasets. Additionally, we elaborate on the benefits of intervening on each confounder and the computational cost of implementing our framework in Sec. 4.3. The qualitative results are in Sec. 4.4.", + "bbox": [ + 212, + 611, + 787, + 748 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experiment Settings", + "text_level": 1, + "bbox": [ + 215, + 768, + 431, + 786 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. We conduct the experiments on Composition-1k [40] and Distinctions-646 [43] datasets following [25]. Composition-1k dataset consists of 431 and 50 foreground objects in the training set and test set respectively. These foregrounds", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Liao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "are composited with MS COCO [17] images for training and with VOC 2012 [7] images for testing. For testing, each foreground object is combined with 20 different images, generating a total of 1000 test images. The Distinctions-646 dataset contains 646 unique foreground objects, which are divided into 596 for training and 50 for testing. The composition process with background images follows the same procedure as in the Composition-1k dataset.", + "bbox": [ + 212, + 146, + 782, + 236 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Metrics. For measuring the overall performance of our framework, we adopt the widely used image matting metrics [12,25,26,43]: sum of absolute difference (SAD), mean square error (MSE), and the gradient (Grad) and connectivity (Conn). Additionally, to assess the improvement in reducing transparency and contrast biases, we measure SAD conditioned on each contrast and transparency category. The number of parameters and FLOPs are also reported to evaluate the computational complexity of implementing our framework.", + "bbox": [ + 212, + 238, + 784, + 343 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation Details. We set the number of sub-spaces as $n = 5$ for $I, F, A, T$ , and $C$ . For transparency and contrast levels, we set both $K_{t}$ and $K_{c}$ to 25. The contrast and transparency space is initialized by the average multiscale feature extracted by pretrained Swin-B [19, 20] as described in Sec. 3.3. We employ a $\\mathcal{N}(0, 1.2)$ Gaussian distribution to smooth the label of attention score. More details can be found in the supplementary material.", + "bbox": [ + 212, + 345, + 785, + 434 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Main Results", + "text_level": 1, + "bbox": [ + 215, + 462, + 372, + 476 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The improvement of our COIN framework is demonstrated in Tab. 1. It is evident that by intervening on the confounder variables, our COIN framework significantly enhances the performance across all matting baselines with varying guidance mechanisms on both the Composition-1k and Distinctions-646 datasets. Our framework substantially improves results for both guidance-free methods (e.g., dugMatting (0-sel)) and flexible-guidance based methods (e.g., dugMatting (1-sel) and DIIM). Additionally, even for the SOTA trimap-based baseline (e.g., MatteFormer), our COIN framework achieves significant performance gains, decreasing SAD by 1.9 and 2.6, and MSE by 0.7 and 1.5 on the two datasets respectively. Notably, our framework tends to exhibit greater improvements on the Distinctions-646 dataset compared to the Composition-1k dataset. This discrepancy is likely attributable to the more pronounced biases present within the Distinctions-646 dataset. Furthermore, enhancements are generally more substantial for flexible guidance-based or guidance-free matting methods than for trimap-based methods, since the former methods are typically more susceptible to biases in the absence of the detailed information provided by a trimap.", + "bbox": [ + 212, + 491, + 787, + 733 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To demonstrate that the enhancements are truly attributable to the reduction of contrast bias and transparency bias, we further evaluate the SAD improvements across each contrast and transparency category, as shown in Fig. 5. We selected MatteFormer as the baseline because it presents the most substantial challenge and exhibits the highest performance among the baselines. Regarding transparency, it is observable that our framework enhances the performance of baseline matting methods across all transparency levels. Furthermore, we note", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "COIN-Matting: Confounder Intervention for Image Matting", + "bbox": [ + 333, + 114, + 732, + 130 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/224a3ba2856cdff259e0ed0333a2af014edaef2a9d38ebc713aa01784f139b7c.jpg", + "table_caption": [ + "Table 1: The comparison of applying our COIN framework on Composition-1k and Distinctions-646 dataset. The improvements are highlighted as superscripts." + ], + "table_footnote": [ + "* $k$ -sel denotes $k$ times human interaction during matting." + ], + "table_body": "
MethodComposition-1KDistinctions-646
SAD ↓MSE ↓Grad ↓Conn ↓SAD ↓MSE ↓Grad ↓Conn ↓
DIIM37.510.318.534.536.211.819.635.6
dugMatting (0-sel)*34.15.715.631.233.28.818.233.2
dugMatting (1-sel)*25.84.39.722.324.17.112.424.0
MatteFormer23.84.08.718.921.96.611.220.5
COIN(DIIM)33.8-3.78.1-2.232.0-2.59.6-2.2
COIN(dugMatting (0-sel)*)30.8-3.34.5-1.228.0-3.27.4-1.4
COIN(dugMatting (1-sel)*)23.7-2.13.5-0.819.1-3.26.3-0.8
COIN(MatteFormer)21.9-1.93.3-0.716.8-2.15.1-1.5
", + "bbox": [ + 218, + 184, + 759, + 303 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1d89b42cc1878961ec7078efaaed3d0f0cc1a7fa41d65aeb5fcf7a7c1464aede.jpg", + "image_caption": [ + "(a) The SAD conditioned on transparency." + ], + "image_footnote": [], + "bbox": [ + 238, + 353, + 486, + 474 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a66a2a6a20115ec1fd03fa9f05831862e1142af7efb7ce26871be4b32e8b06f4.jpg", + "image_caption": [ + "(b) The SAD conditioned on contrast.", + "Fig. 5: The improvement of applying our framework to MatteFormer on Composition-1K dataset conditioned on each contrast and transparency category." + ], + "image_footnote": [], + "bbox": [ + 514, + 353, + 764, + 474 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "that the improvement is particularly pronounced in regions where the transparency level is within the range of [50, 85]. The training data within this range, as discussed in Sec. 1, is notably scarce but exhibits a unique alpha pattern which differs from the pattern in highly opaque or transparent foregrounds. The observed probability distribution, dominated by data from regions with high opacity or transparency, causes the transparency bias we identified. The significant enhancements in these regions indicate that our framework is reducing the transparency bias through confounder intervention. In terms of contrast, our framework not only significantly reduces the SAD error across all contrast levels compared to the baseline method but also shows that the most notable improvement occurs in areas of relatively low contrast, particularly where the contrast level ranges from [10, 25]. In these regions, the detail patterns increase rapidly as average contrast value and mean alpha value raises (as shown in Fig. 2a). However, owing to the insufficient data relative to other regions, these areas are most susceptible to contrast bias. Thus, the most notable improvements in low-contrast areas suggest that our framework enhances performance by alleviating the contrast bias through intervention on the contrast confounder.", + "bbox": [ + 212, + 583, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Liao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/4b0f68b65c47a12db5e5e2537054ffa062fea9faa2c67df13373f8e3a8c05014.jpg", + "table_caption": [ + "Table 2: The comparison of applying intervention on different confounder variables on the Distinctions-646 dataset using MatteFormer as the baseline. The improvements are highlighted as superscripts. The GFLOPs is calculated on $1024 \\times 1024$ image." + ], + "table_footnote": [], + "table_body": "
MethodErrorsComplexity
SAD ↓MSE ↓Grad ↓Conn ↓ParametersFLOPs
MatteFormer21.9-1.56.6-0.811.2-0.920.544.8M459G
MatteFormer + T intervene20.45.810.319.1-1.447.7M
MatteFormer + C intervene20.2-1.75.7-0.910.0-1.218.8-1.747.7M
COIN(MatteFormer)19.3-2.65.1-1.59.4-1.817.8-2.749.5M
", + "bbox": [ + 233, + 198, + 769, + 282 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 309, + 388, + 325 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Confounder Spaces. We discuss the improvement from intervening in each confounder space individually rather than their joint confounder in Tab. 2. By comparing line 3 and line 2 with line 1, it is observed that intervening on either $\\mathcal{C}$ or $\\mathcal{T}$ can improve the performance of the baseline methods. However, because such interventions do not block the backdoor path opened by $\\mathcal{T}$ and $\\mathcal{C}$ respectively, the condition $\\mathcal{A} \\nsubseteq \\mathcal{B}|\\mathcal{T}$ still exists, which leads to transparency bias and contrast bias respectively, resulting in a limited performance of the intervention framework. Moreover, it was noted that the performance of intervening on $\\mathcal{C}$ is slightly better than that of intervening on $\\mathcal{T}$ . This may be due to the fact that the contrast bias is more pronounced than the transparency bias in the dataset. By jointly intervening both $\\mathcal{C}$ and $\\mathcal{T}$ , we may block all aforementioned backdoor paths, thus further performance improvement in the COIN framework.", + "bbox": [ + 212, + 335, + 787, + 517 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Complexity Analysis. As we introduce a framework atop existing matting methods, we further analyze its space and time complexity in terms of both parameters and FLOPs to delineate the cost of integrating our framework. As indicated in Tab. 2, the additional cost imposed by our framework is minimal when compared to the baseline methods. The incremental parameters (about $10.5\\%$ ) stem exclusively from the feature space and its corresponding attention mechanism, which are negligible relative to the baseline model. The only extra FLOPs (about $10.6\\%$ ) originate from the computation of expectations within feature spaces, which are characterized as $n$ attentions across $d_{i} \\times (K_{t} \\cdot K_{c})$ matrices and $d_{i} \\times L$ image features, where $L$ denotes the number of image patches at the $i$ -th scale. Such attention is also lightweight compared to the matting baseline. Overall, our framework improves the performance of the SOTA image matting baseline between $11.6\\%$ and $22.7\\%$ in different metrics with about $10\\%$ computation cost, which is remarkably efficient for integrating with the baselines.", + "bbox": [ + 212, + 518, + 787, + 729 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Qualitative Results", + "text_level": 1, + "bbox": [ + 215, + 753, + 421, + 768 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we present qualitative comparisons of our COIN framework. In Fig. 6, we compare the alpha matting results from baseline methods with those obtained using our COIN framework. Rows 1 showcase examples illustrating the reduction of contrast bias. It is evident that the low-contrast areas", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "COIN-Matting: Confounder Intervention for Image Matting", + "bbox": [ + 333, + 114, + 732, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a7e5c4ac4cb85e594eff51389db6f21fbe9b57776fe97819aeb86dca8984bca8.jpg", + "image_caption": [ + "Image" + ], + "image_footnote": [], + "bbox": [ + 233, + 142, + 339, + 338 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3289c8ca65ad8c01aa4a9fed4969cb5b1289153a7c276b5dd66c756b3a28f755.jpg", + "image_caption": [ + "Trimap" + ], + "image_footnote": [], + "bbox": [ + 341, + 143, + 446, + 338 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/df03e1bbec89ea81c0ab77ac5d28a4abd6c0ae2e73d231b1cdf61d20b426231c.jpg", + "image_caption": [ + "Alpha matte" + ], + "image_footnote": [], + "bbox": [ + 449, + 143, + 553, + 338 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5f15eef4afc378abdfe2720863b31a6e8ba706c6314a3e984e3624c906cbe1ba.jpg", + "image_caption": [ + "MatteFormer", + "Fig. 6: The visualized comparison of applying our COIN framework on Matte-Former [25]. Images are best viewed when zoomed in." + ], + "image_footnote": [], + "bbox": [ + 555, + 143, + 661, + 338 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5f5e8270e9aa68e41c5d5ac858d2b2466c4eeb52ae62db1e31cf73303c1c4545.jpg", + "image_caption": [ + "COIN" + ], + "image_footnote": [], + "bbox": [ + 663, + 143, + 769, + 338 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "highlighted in the red boxes yield alpha mattes with lower values when processed by MatteFormer. However, the integration of our COIN framework significantly diminishes this contrast bias, resulting in more precise alpha mattes. Rows 2 and 3 demonstrate the reduction of transparency bias. The highlighted mid-transparent regions in the red boxes contain background textures (i.e., the texture of text and the car cage respectively) in the predicted alpha mattes. Nevertheless, with the application of our COIN framework, the alpha mattes become markedly more accurate, effectively eliminating such textures from the mid-transparent areas. Such qualitative results prove that our COIN framework can effectively reduce both contrast bias and transparency bias, therefore improving the performance of the existing image matting methods.", + "bbox": [ + 212, + 420, + 787, + 587 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 607, + 359, + 625 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we have concentrated on the contrast bias and transparency bias in existing matting methods and demonstrated how dataset biases lead to biased behaviors of models. We analyzed these biases from a causal perspective and designed the confounder intervention framework to alleviate them. Extensive experiments have demonstrated that our framework markedly alleviates these biases and enhances SOTA matting methods with minor computational cost.", + "bbox": [ + 212, + 637, + 787, + 729 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 750, + 403, + 767 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The work was supported by the National Natural Science Foundation of China (Grant No. 62076162), the Shanghai Municipal Science and Technology Major Project, China (Grant No. 2021SHZDZX0102) and the Postdoctoral Fellowship Program of CPSF (Grant No. GZC20241225).", + "bbox": [ + 214, + 779, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Liao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Cai, H., Xue, F., Xu, L., Guo, L.: Transmatting: Enhancing transparent objects matting with transformers. In: ECCV (2022)", + "2. Chen, J., Li, X., Luo, L., Ma, J.: Multi-focus image fusion based on multi-scale gradients and image matting. IEEE Trans. Multim. 24, 655-667 (2022)", + "3. Dai, Y., Lu, H., Shen, C.: Learning affinity-aware upsampling for deep image matting. In: CVPR (2021)", + "4. Dai, Y., Price, B., Zhang, H., Shen, C.: Boosting robustness of image matting with context assembling and strong data augmentation. In: CVPR (2022)", + "5. Deng, J., Xu, Y., Zhou, Z., He, S.: Background matting via recursive excitation. In: ICME (2022)", + "6. Ding, H., Zhang, H., Liu, C., Jiang, X.: Deep interactive image matting with feature propagation. IEEE Trans. Image Process. 31, 2421-2432 (2022)", + "7. Everingham, M., Gool, L.V., Williams, C.K.I., Winn, J.M., Zisserman, A.: The pascal visual object classes (VOC) challenge. Int. J. Comput. Vis. (2010)", + "8. Fang, X., Zhang, S., Chen, T., Wu, X., Shamir, A., Hu, S.: User-guided deep human image matting using arbitrary trimaps. IEEE Trans. Image Process. 31, 2040-2052 (2022)", + "9. Hou, Q., Liu, F.: Context-aware image matting for simultaneous foreground and alpha estimation. In: ICCV (2019)", + "0. Li, J., Niu, L., Zhang, L.: Knowledge proxy intervention for deconfounded video question answering. In: ICCV (2023)", + "1. Li, J., Wang, W., Chen, J., Niu, L., Si, J., Qian, C., Zhang, L.: Video semantic segmentation via sparse temporal transformer. In: ACM MM (2021)", + "2. Li, J., Zhang, J., Tao, D.: Deep automatic natural image matting. In: IJCAI (2021)", + "3. Li, J., Zhang, J., Tao, D.: Referring image matting. In: Proceedings of the IEEE Computer Vision and Pattern Recognition (2023)", + "4. Li, Y., Lu, H.: Natural image matting via guided contextual attention. In: AAAI (2020)", + "5. Li, Y., Wang, X., Xiao, J., Chua, T.: Equivariant and invariant grounding for video question answering. In: ACM MM (2022)", + "6. Lin, S., Ryabtsev, A., Sengupta, S., Curless, B.L., Seitz, S.M., Kemelmacher-Shlizerman, I.: Real-time high-resolution background matting. In: CVPR (2021)", + "7. Lin, T., Maire, M., Belongie, S.J., Hays, J., Perona, P., Ramanan, D., Dóllar, P., Zitnick, C.L.: Microsoft COCO: common objects in context. In: ECCV (2014)", + "8. Liu, Y., Xie, J., Qiao, Y., Tang, Y., Yang, X.: Prior-induced information alignment for image matting. IEEE Trans. Multim. 24, 2727-2738 (2022)", + "9. Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., Wei, F., Guo, B.: Swin transformer v2: Scaling up capacity and resolution. In: CVPR (2022)", + "20. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021)", + "21. Lu, H., Dai, Y., Shen, C., Xu, S.: Indices matter: Learning to index for deep image matting. In: ICCV (2019)", + "22. Lu, L., Li, J., Cao, J., Niu, L., Zhang, L.: Painterly image harmonization using diffusion model. In: ACM MM (2023)", + "23. Lutz, S., Amplanitis, K., Smolic, A.: Alphagan: Generative adversarial networks for natural image matting. In: BMVC (2018)" + ], + "bbox": [ + 225, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "COIN-Matting: Confounder Intervention for Image Matting", + "bbox": [ + 333, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "24. Niu, Y., Tang, K., Zhang, H., Lu, Z., Hua, X., Wen, J.: Counterfactual VQA: A cause-effect look at language bias. In: CVPR (2021)", + "25. Park, G., Son, S., Yoo, J., Kim, S., Kwak, N.: Matteformer: Transformer-based image matting via prior-tokens. In: CVPR (2022)", + "26. Park, K., Woo, S., Oh, S.W., Kweon, I.S., Lee, J.Y.: Mask-guided matting in the wild. In: CVPR (2023)", + "27. Pearl, J., Glymour, M., Jewell, N.P.: Causal inference in statistics: A primer (2016)", + "28. Pearl, J., Mackenzie, D.: The Book of Why: The New Science of Cause and Effect. Basic Books, Inc. (2018)", + "29. Rubin, D.B.: Essential concepts of causal inference: a remarkable history and an intriguing future. Biostatistics & Epidemiology (2019)", + "30. Sengupta, S., Jayaram, V., Curless, B., Seitz, S.M., Kemelmacher-Shlizerman, I.: Background matting: The world is your green screen. In: CVPR (2020)", + "31. Sharma, G., Wu, W., Dalal, E.N.: The ciede2000 color-difference formula: Implementation notes, supplementary test data, and mathematical observations. Color Research and Application 30, 21-30 (2005), https://api-semanticscholar.org/ CorpusID:29119937", + "32. Srivastava, N., Hinton, G.E., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. J. Mach. Learn. Res. (2014)", + "33. Sun, Y., Tang, C., Tai, Y.: Ultrahigh resolution image/video matting with spatiotemporal sparsity. In: CVPR (2023)", + "34. Tan, L., Li, J., Niu, L., Zhang, L.: Deep image harmonization in dual color spaces. In: ACM MM (2023)", + "35. Wang, T., Zhou, C., Sun, Q., Zhang, H.: Causal attention for unbiased visual recognition. In: ICCV (2021)", + "36. Wang, W., Feng, F., He, X., Zhang, H., seng Chua, T.: Clicks can be cheating: Counterfactual recommendation for mitigating clickbait issue (2020)", + "37. Wei, T., Chen, D., Zhou, W., Liao, J., Zhao, H., Zhang, W., Hua, G., Yu, N.: Deep image matting with sparse user interactions. TPAMI (2024)", + "38. Wu, J., Zhang, C., Li, Z., Fu, H., Peng, X., Zhou, J.T.: dugMatting: Decomposed-uncertainty-guided matting. In: ICML (2023)", + "39. Xu, K., Ba, J., Kiros, R., Cho, K., Courville, A.C., Salakhutdinov, R., Zemel, R.S., Bengio, Y.: Show, attend and tell: Neural image caption generation with visual attention. In: ICML (2015)", + "40. Xu, N., Price, B.L., Cohen, S., Huang, T.S.: Deep image matting. In: CVPR (2017)", + "41. Yao, J., Wang, X., Yang, S., Wang, B.: Vitmatte: Boosting image matting with pre-trained plain vision transformers. Inf. Fusion 103, 102091 (2024)", + "42. Ye, Z., Dai, Y., Hong, C., Cao, Z., Lu, H.: Infusing definiteness into randomness: Rethinking composition styles for deep image matting. In: AAAI (2023)", + "43. Yu, Q., Zhang, J., Zhang, H., Wang, Y., Lin, Z., Xu, N., Bai, Y., Yuille, A.L.: Mask guided matting via progressive refinement network. In: CVPR (2021)", + "44. Zhang, D., Zhang, H., Tang, J., Hua, X., Sun, Q.: Causal intervention for weakly-supervised semantic segmentation. In: NeurIPS (2020)", + "45. Zhu, Q., Zhang, W.N., Liu, T., Wang, W.Y.: Counterfactual off-policy training for neural dialogue generation. In: EMNLP (2020)" + ], + "bbox": [ + 215, + 146, + 784, + 784 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Liao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/2024/COIN-Matting_ Confounder Intervention for Image Matting/7e767e14-ba96-40b5-847c-1a51f71239aa_model.json b/2024/COIN-Matting_ Confounder Intervention for Image Matting/7e767e14-ba96-40b5-847c-1a51f71239aa_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3cf8344b921761f35900db9a2d07be340b1fe2bf --- /dev/null +++ b/2024/COIN-Matting_ Confounder Intervention for Image Matting/7e767e14-ba96-40b5-847c-1a51f71239aa_model.json @@ -0,0 +1,2641 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.249, + 0.141, + 0.756, + 0.187 + ], + "angle": 0, + "content": "COIN-Matting: Confounder Intervention for Image Matting" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.212, + 0.768, + 0.246 + ], + "angle": 0, + "content": "Zhaohe Liao\\(^{1}\\), Jiangtong Li\\(^{2}\\), Jun Lan\\(^{3}\\), Huijia Zhu\\(^{3}\\), Weiqiang Wang\\(^{3}\\), Li Niu\\(^{*1,4}\\), and Liqing Zhang\\(^{*1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.255, + 0.696, + 0.354 + ], + "angle": 0, + "content": "1 Shanghai Jiao Tong University \n{zhaoheliao, ustcnewly, zhang-lq}@sjtu.edu.cn \n2 Tongji University \njiangtongli@tongji.edu.cn \n3 Ant Group \n{yelan.lj, huijia.zhj, weiqiang.wwq}@antgroup.com \n4 miguo.ai" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.39, + 0.744, + 0.571 + ], + "angle": 0, + "content": "Abstract. Deep learning methods have significantly advanced the performance of image matting. However, dataset biases can mislead the matting models to biased behavior. In this paper, we identify the two typical biases in existing matting models, specifically contrast bias and transparency bias, and discuss their origins in matting datasets. To address these biases, we model the image matting task from the perspective of causal inference and identify the root causes of these biases: the confounders. To mitigate the effects of these confounders, we employ causal intervention through backdoor adjustment and introduce a novel model-agnostic cofounder intervened (COIN) matting framework. Extensive experiments across various matting methods and datasets have demonstrated that our COIN framework can significantly diminish such biases, thereby enhancing the performance of existing matting models." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.584, + 0.644, + 0.599 + ], + "angle": 0, + "content": "Keywords: Image matting \\(\\cdot\\) Intervention \\(\\cdot\\) Dataset bias" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.624, + 0.377, + 0.64 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.656, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Image matting aims at separating accurate foreground objects from natural images and precisely estimating the opacity on the boundary of the object. It serves as the basis for image editing, live-streaming, virtual meetings, and movie production [2,16,22,30,34]. Given an image \\( \\mathbf{I} \\), it is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.729, + 0.786, + 0.745 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {I} = \\boldsymbol {F} \\circ \\boldsymbol {\\alpha} + \\boldsymbol {B} \\circ (1 - \\boldsymbol {\\alpha}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.756, + 0.788, + 0.819 + ], + "angle": 0, + "content": "where \\(\\pmb{F}\\) represents the foreground object, \\(\\pmb{B}\\) denotes the background, \\(\\alpha \\in [0,1]\\) is the opacity map, and \\(\\circ\\) denotes element-wise multiplication. As shown in Eq. (1), solving the 7 unknown values in \\(\\{\\pmb{F},\\pmb{B},\\pmb{\\alpha}\\}\\) from the given 3 values in \\(\\pmb{I}\\) is a highly ill-posed problem. Therefore, some matting methods require" + }, + { + "type": "page_footnote", + "bbox": [ + 0.238, + 0.826, + 0.432, + 0.841 + ], + "angle": 0, + "content": "*The corresponding authors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Z. Liao et al." + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.144, + 0.326, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.201, + 0.302, + 0.211 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.144, + 0.413, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.201, + 0.377, + 0.21 + ], + "angle": 0, + "content": "\\(\\alpha\\)" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.144, + 0.501, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.452, + 0.201, + 0.462, + 0.21 + ], + "angle": 0, + "content": "a" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.144, + 0.59, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.201, + 0.565, + 0.211 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.592, + 0.144, + 0.678, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.629, + 0.201, + 0.639, + 0.21 + ], + "angle": 0, + "content": "\\(\\alpha\\)" + }, + { + "type": "image", + "bbox": [ + 0.679, + 0.144, + 0.765, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.716, + 0.201, + 0.726, + 0.21 + ], + "angle": 0, + "content": "\\(\\hat{\\alpha}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.439, + 0.214, + 0.566, + 0.224 + ], + "angle": 0, + "content": "(a) The contrast bias." + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.225, + 0.327, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.294, + 0.301, + 0.304 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.225, + 0.413, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.366, + 0.295, + 0.377, + 0.303 + ], + "angle": 0, + "content": "\\(\\alpha\\)" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.225, + 0.5, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.294, + 0.462, + 0.303 + ], + "angle": 0, + "content": "a" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.225, + 0.588, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.294, + 0.565, + 0.304 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.225, + 0.676, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.294, + 0.639, + 0.303 + ], + "angle": 0, + "content": "\\(\\alpha\\)" + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.225, + 0.764, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.716, + 0.294, + 0.726, + 0.303 + ], + "angle": 0, + "content": "\\(\\hat{\\alpha}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.423, + 0.306, + 0.581, + 0.319 + ], + "angle": 0, + "content": "(b) The transparency bias." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.33, + 0.788, + 0.372 + ], + "angle": 0, + "content": "Fig. 1: The observed contrast bias and transparency bias for MatteFormer [25] on the Compositional-1K dataset [40]. \\(\\alpha\\) and \\(\\hat{\\alpha}\\) denote the ground truth and predicted alpha mattes, respectively. Images are best viewed when zoomed in." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.785, + 0.492 + ], + "angle": 0, + "content": "additional user input as guidance, leading to a classification of matting methods based on whether and what type of user input is used. Trimap-based methods [40] are the most widely used, but obtaining a trimap can be challenging in some scenarios. Therefore, more flexible guidance, such as background [16,30] and user interaction [6,37,38], have been explored. Moreover, some methods [12,38] can also work without user guidance, which is concluded as guidance-free methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.787, + 0.704 + ], + "angle": 0, + "content": "The rapid advances in deep learning have significantly boosted the performance of image matting. However, the matting dataset may contain biases, which may further lead to the biased behavior of image matting models. We identify two primary types of bias in existing models: contrast bias and transparency bias. To illustrate these biases, we present some results of a state-of-the-art (SOTA) model, namely MatteFormer [25], in Fig. 1. The contrast bias refers to the correlation between the contrast and the predicted alpha value. Specifically, the model tends to predict lower alpha values in areas with low contrast between foreground and background, as highlighted by the red boxes in Fig. 1a. Regarding the transparency bias, it is observed that the background texture tends to emerge in the predicted alpha matte for mid-transparent foregrounds. As highlighted by the red boxes in Fig. 1b, the predicted alpha for mid-transparent area contains the background texture (i.e., the texture of the rider and building for the left and right examples, respectively)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Such biased behavior of models can be explained by the biases present in the dataset. For the contrast bias, we calculate the alpha value at different contrast levels within the training dataset as shown in Fig. 2a. We observe that the mean alpha level in low contrast areas (i.e., contrast \\(< 20\\) in the CIEDE2000 color difference formula [31]) is significantly lower than that in other areas. Such dataset bias directly contributes to the observed contrast bias as the model may falsely learn to predict alpha from contrast. Additionally, the number of samples in the low contrast area is significantly fewer than in other areas, making the model more easily led by such dataset bias in these areas. Regarding the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.334, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "COIN-Matting: Confounder Intervention for Image Matting" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.146, + 0.472, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.246, + 0.478, + 0.259 + ], + "angle": 0, + "content": "(a) The contrast bias in the dataset." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.147, + 0.73, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.246, + 0.751, + 0.259 + ], + "angle": 0, + "content": "(b) The transparency bias in the dataset." + }, + { + "type": "image_caption", + "bbox": [ + 0.342, + 0.27, + 0.658, + 0.284 + ], + "angle": 0, + "content": "Fig. 2: The observe biases in matting datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.321, + 0.785, + 0.427 + ], + "angle": 0, + "content": "transparency bias, we analyze the distribution of transparency levels within the training dataset in Fig. 2b. We observe that most foreground objects are either highly salient or very opaque. Therefore, the alpha matte is expected to be either close to zero or exhibit textures in the image. For the mid-transparency area, whose alpha matte clearly differs from zero, the model is misled to focus on the texture of the image. Thus, the texture of the background in such areas may appear in the alpha matte, explaining the appearance of transparency bias." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.786, + 0.656 + ], + "angle": 0, + "content": "To address these biases, we model the image matting task from the perspective of causal inference, and utilize the power of intervention to reduce the influence of the biases. The causal graph for image matting is depicted in Fig. 3a, where \\(\\mathcal{I},\\mathcal{F},\\mathcal{B}\\), and \\(\\mathcal{A}\\) are variables representing the image, foreground, background, and alpha matte respectively, while \\(\\mathcal{C}\\) and \\(\\mathcal{T}\\) are confounder variables denoting the contrast and the transparency of foreground. The notation \\(\\mathcal{B} \\gets \\mathcal{I} \\to \\mathcal{F}\\) suggests that an image can be decomposed into foreground and background, underpinning the basic assumption of image matting. \\(\\mathcal{F} \\to \\mathcal{A}\\) implies that the alpha matte is directly influenced by the foreground object, indicating that changes in the background should not affect the foreground's representation. The confounder variables, representing the common causes affecting multiple variables in the foreground, background, and image, serve as the source of biases. Typically, these can include properties of visual objects. Given that both the foreground and background consist of objects, these properties typically cause the presentation of the foreground, background, and images are presented." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.84 + ], + "angle": 0, + "content": "However, directly measuring all possible confounders may be impractical, as we can not enumerate all properties of visual objects. Nevertheless, based on observed dataset biases, we can identify the most typical confounders. As highlighted in Fig. 3a, we focus on two typical confounders: contrast \\(\\mathcal{C}\\) and transparency \\(\\mathcal{T}\\), which underlie the observed contrast bias and transparency bias, respectively. Contrast measures the color difference between foreground and background objects, while transparency refers to the opacity of the foreground. Therefore, we conclude \\(\\mathcal{F} \\gets \\mathcal{T} \\to \\mathcal{I}\\) and \\(\\mathcal{C} \\to \\mathcal{X}\\) for all \\(\\mathcal{X} \\in \\{\\mathcal{F}, \\mathcal{I}, \\mathcal{B}\\}\\) by their definition. The dotted line indicates a potential correlation between \\(\\mathcal{C}\\) and \\(\\mathcal{T}\\), suggesting that we cannot definitively assert that \\(\\mathcal{C} \\perp \\mathcal{T}\\). While existing matting methods generally focus on maximizing the likelihood \\(P(\\mathcal{A}|\\mathcal{I})\\), the presence of backdoor paths \\(\\mathcal{B} \\gets \\mathcal{C} \\to \\mathcal{F} \\to \\mathcal{A}\\), \\(\\mathcal{B} \\gets \\mathcal{I} \\gets \\mathcal{C} \\to \\mathcal{F} \\to \\mathcal{A}\\), and \\(\\mathcal{B} \\gets \\mathcal{I} \\gets \\mathcal{T} \\to\\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Z. Liao et al." + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.147, + 0.447, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.323, + 0.22, + 0.45, + 0.233 + ], + "angle": 0, + "content": "(a) The causal graph." + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.147, + 0.677, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.22, + 0.706, + 0.233 + ], + "angle": 0, + "content": "(b) The backdoor adjustment." + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.243, + 0.728, + 0.258 + ], + "angle": 0, + "content": "Fig. 3: The causal graph and causal intervention for image matting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.284, + 0.785, + 0.329 + ], + "angle": 0, + "content": "\\(\\mathcal{F} \\to \\mathcal{A}\\) results in \\(\\mathcal{B} \\not\\sqcup \\mathcal{A}|\\mathcal{I}\\). This lack of conditional d-separation between \\(\\mathcal{B}\\) and \\(\\mathcal{A}\\) elucidates the co-occurrence of background textures in specific transparency regions and the contrast-induced alpha bias." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.617 + ], + "angle": 0, + "content": "To remove the effect of these confounders, we propose our model-agnostic cofounder intervened (COIN) matting framework. Specifically, we explore the do-calculus to intervene in the influence of \\(\\mathcal{T}\\) and \\(\\mathcal{C}\\) via backdoor adjustment, as shown in Fig. 3b. The do-calculus actively assigns values to variables, rather than passively observing them from the data. Therefore, the do-calculus is an intervention on the influence of their direct causes, which is used for decofounding. While the ideal way to calculate the do-operator involves collecting all images with varying contrast and transparency into the training data, this approach is evidently impractical. Furthermore, the high cost of labeling the alpha matte makes it exceedingly difficult to even approximate the aforementioned intervention \"physically\". To address this issue, we apply backdoor adjustment to calculate \\(P(\\mathcal{A}|do(\\mathcal{F},\\mathcal{B},\\mathcal{I}))\\), as described in Secs. 3.2 to 3.4. By actively assigning values to \\(\\mathcal{F},\\mathcal{B}\\), and \\(\\mathcal{I}\\), we block backdoor paths and magnify the influence of cofounders, thus enhancing the robustness of the matting results against the previously mentioned dataset biases. By applying our framework to state-of-the-art (SOTA) image matting methods with different types of user guidance (i.e., MatteFormer [25], DIIM [6], and dugMatting [38]) and two datasets (i.e., Comp1K and Distinct-646), we verified the effectiveness of our framework through extensive experiments. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.622, + 0.785, + 0.651 + ], + "angle": 0, + "content": "1. We examine the biases in matting methods, analyzing how dataset biases cause them and affect alpha matte prediction." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.652, + 0.785, + 0.695 + ], + "angle": 0, + "content": "2. We introduce a novel and model-agnostic confounder-intervened matting framework that employs causal intervention to sever the influence of confounders, thereby mitigating the aforementioned biases." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.696, + 0.785, + 0.74 + ], + "angle": 0, + "content": "3. Through comprehensive experiments on various matting methods and multiple datasets, we demonstrate that our framework significantly enhances the performance of SOTA image matting methods." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.622, + 0.785, + 0.74 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.761, + 0.388, + 0.777 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.789, + 0.387, + 0.804 + ], + "angle": 0, + "content": "2.1 Image Matting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "As described in Sec. 1, image matting presents a highly ill-posed problem. The solutions can be classified based on the user input prior they rely upon." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.334, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "COIN-Matting: Confounder Intervention for Image Matting" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.283 + ], + "angle": 0, + "content": "Among the prior-based image matting methods, the trimap is the most widely utilized prior [1,3,4,9,14,18,21,23,23,25,33,40-42]. Since [40] introduced the first well-known deep learning-based image matting method, featuring an encoder-decoder architecture, this framework has been extensively adopted and improved by subsequent methods. Researchers have enhanced this architecture from various perspectives, including generative adversarial training [23], matting with indices [21], affinity-based methods [3, 14], context-aware matting [9], the introduction of vision transformers [1, 25, 41], prior alignment [18], expansion to ultra-high resolution via sparsity [33], and data augmentation [4, 42]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.785, + 0.419 + ], + "angle": 0, + "content": "However, trimap is not always available for real-world applications. Thus, recent researchers seek for more flexible and accessible user inputs [5, 6, 8, 13, 16, 30, 37, 38, 43]. Background matting [30] uses the background image as the prior, which can be more accessible for scenes like virtual meeting and livestreaming. It is further enhanced by recursive excitation [5] and extended into real-time version [16]. The coarse mask is used instead of trimap in [43] which shows better penitential of cooperating with the segmentation models [11]. The user feedbacks, such as clicks [6], scribbles [8], their combination [37], proposal selection [38] and even natural language [13], are also informative for matting." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.442, + 0.403, + 0.456 + ], + "angle": 0, + "content": "2.2 Causal Inference" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.618 + ], + "angle": 0, + "content": "Causal inference [27, 29] plays a pivotal role in analyzing dataset bias and mitigating counter-causal correlations. It can be categorized into two main types: deconfounding and counterfactual inference. Serving as a fundamental statistical tool, it finds widespread application across various fields, such as image segmentation [44], image classification [35], image and video question answering [10, 15, 24], dialog systems [45], and recommendation systems [36]. Typically, causal-related methods conceptualize the task within a causal graph, where nodes symbolize variables, and edges denote their causal interactions. By examining the causal relations among variables, the causal graph facilitates identifying confounders and backdoor paths, thereby enabling the control of biases." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.642, + 0.381, + 0.66 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This section outlines our intervention approach to mitigating biases introduced by confounders, as discussed in Sec. 1. In Sec. 3.1, we examine the role of confounder variables within the causal graph (i.e., Fig. 3), illustrating their impact on the performance of mating models through mathematical analysis. In Sec. 3.2, the concept of backdoor adjustment is introduced to counteract biases. In Sec. 3.3, we discuss the construction of space representations for each variable in the causal graph to facilitate backdoor adjustment implementation. In Sec. 3.4, the calculation of expectations, as derived in Sec. 3.2, is explained to execute the causal intervention. Finally, in Sec. 3.5, we integrate the above components into our model-agnostic COIN mating framework. Throughout this section, we denote variables by calligraphic letters (e.g., \\(\\mathcal{A}\\), \\(\\mathcal{F}\\), \\(\\mathcal{I}\\)), feature spaces" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "Z. Liao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.195 + ], + "angle": 0, + "content": "by boldface capital letters (e.g., \\( \\mathbf{A} \\), \\( \\mathbf{F} \\), \\( \\mathbf{I} \\)), and feature vectors by boldface lowercase letters (e.g., \\( \\mathbf{a} \\), \\( \\mathbf{f} \\), \\( \\mathbf{i} \\)). Additional introduction of background and more detailed derivations are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.218, + 0.573, + 0.234 + ], + "angle": 0, + "content": "3.1 Confounders and the Backdoor Paths" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.247, + 0.789, + 0.294 + ], + "angle": 0, + "content": "In the task of image matting, confounders are variables that are the direct cause of at least two of the following: the foreground, background, and the image. As illustrated in Fig. 3a, the observational probability can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.306, + 0.786, + 0.336 + ], + "angle": 0, + "content": "\\[\nP (\\mathcal {A} | \\mathcal {I}) = \\sum_ {\\boldsymbol {t}} \\sum_ {\\boldsymbol {c}} P (\\mathcal {A} | \\mathcal {I}, \\boldsymbol {t}, \\boldsymbol {c}) P (\\boldsymbol {c}, \\boldsymbol {t} | \\mathcal {I}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.347, + 0.789, + 0.574 + ], + "angle": 0, + "content": "where \\( \\pmb{t} \\) and \\( \\pmb{c} \\) represent the split of confounder \\( \\mathcal{C} \\) and \\( \\mathcal{T} \\), respectively. Given that the observed distribution \\( P(\\pmb{c},\\pmb{t}|\\mathcal{I}) \\) significantly deviates from a uniform distribution (as depicted in Fig. 2), a specific partition of \\( \\langle \\mathcal{C},\\mathcal{T}\\rangle \\) predominantly influences the observed probability \\( P(\\mathcal{A}|\\mathcal{I}) \\) through a large \\( P(\\pmb{c},\\pmb{t}|\\mathcal{I}) \\). For instance, in the training dataset, regions with high contrast and low transparency are prevalent, leading to their dominance in the observed probability and observed biases. Specifically, in low-contrast regions, the alpha values are typically to be lower, causing the observed contrast bias. Furthermore, the model is trained to discern detailed textures in regions with opaque foregrounds and to assign nearly zero alpha values in highly transparent areas. Since these conditions consist most of the dataset, \\( P(\\pmb{c},\\pmb{t}|\\mathcal{I}) \\) with extreme values of \\( \\pmb{t} \\) significantly contributes to \\( P(\\mathcal{A}|\\mathcal{I}) \\). Consequently, the model tends to extract highly textured patterns in medium transparency areas as in opaque foreground areas, where alpha values are apparently not as minimal as in transparent regions. Such background textures in the predicted alpha matte are concluded as transparency bias." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.79, + 0.697 + ], + "angle": 0, + "content": "By adopting a causal perspective on Fig. 3a, we identify the presence of backdoor paths introduced by the confounder variables \\(\\mathcal{C}\\) and \\(\\mathcal{T}\\). In particular, the backdoor paths \\(\\mathcal{A} \\gets \\mathcal{F} \\gets \\mathcal{C} \\to \\mathcal{B}\\), \\(\\mathcal{B} \\gets \\mathcal{I} \\gets \\mathcal{C} \\to \\mathcal{F} \\to \\mathcal{A}\\), and \\(\\mathcal{A} \\gets \\mathcal{F} \\gets \\mathcal{T} \\to \\mathcal{I} \\to \\mathcal{B}\\) result in \\(\\mathcal{A} \\not\\perp \\mathcal{B}|\\mathcal{I}\\), elucidating the influence of the background on the predicted alpha matte. More specifically, the former two paths and the latter are directly accountable for the contrast bias and transparency bias, respectively. To intervene the effects of the confounders and sever the backdoor paths they establish, we employ the backdoor adjustment [27,29] as detailed in Sec. 3.2." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.721, + 0.642, + 0.738 + ], + "angle": 0, + "content": "3.2 Causal Intervention via Backdoor Adjustment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.843 + ], + "angle": 0, + "content": "The do-calculus [27, 28] signifies active assignment of values to variables, distinguishing it from passive observation within the data, which constitutes intervention on the causal factors of variables. For brevity, we denote \\( P(do(\\mathcal{X} = x)) \\) as \\( P(do(\\mathcal{X})) \\), implying active assignment of a value to variable \\( \\mathcal{X} \\). As illustrated in Fig. 3b, to sever all incoming causal links of \\( \\mathcal{I}, \\mathcal{F}, \\mathcal{B} \\) (i.e., nullifying the influence of \\( \\mathcal{C} \\) and \\( \\mathcal{T} \\)), we focus on computing \\( P(\\mathcal{A}|do(\\mathcal{I},\\mathcal{F},\\mathcal{B})) \\). Given that \\( \\mathcal{T} \\) and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.334, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "COIN-Matting: Confounder Intervention for Image Matting" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.586, + 0.162 + ], + "angle": 0, + "content": "\\(\\mathcal{C}\\) block all specified backdoor paths, we can derive:" + }, + { + "type": "equation", + "bbox": [ + 0.264, + 0.172, + 0.785, + 0.235 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} P (\\mathcal {A} | d o (\\mathcal {I}, \\mathcal {F}, \\mathcal {B})) = \\sum_ {\\boldsymbol {t}} \\sum_ {\\boldsymbol {c}} P (\\mathcal {A} | d o (\\mathcal {I}, \\mathcal {F}, \\mathcal {B}), \\boldsymbol {t}, \\boldsymbol {c}) P (\\boldsymbol {c}, \\boldsymbol {t} | d o (\\mathcal {I}, \\mathcal {F}, \\mathcal {B})) \\\\ = \\sum_ {t} ^ {t} \\sum_ {c} ^ {c} P (\\mathcal {A} | i, \\boldsymbol {f}, \\boldsymbol {b}, t, c) P (\\boldsymbol {c}, t). \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.244, + 0.784, + 0.275 + ], + "angle": 0, + "content": "Moreover, based on Fig. 3b, we observe \\(\\mathcal{A} \\perp \\mathbf{b}|(\\mathbf{f},\\mathbf{i},\\mathbf{c},\\mathbf{t})\\) and \\(\\mathcal{A} \\perp \\mathbf{i}|(\\mathbf{f},\\mathbf{c},\\mathbf{t})\\). Thus, the equation simplifies to:" + }, + { + "type": "equation", + "bbox": [ + 0.266, + 0.285, + 0.785, + 0.314 + ], + "angle": 0, + "content": "\\[\nP (\\mathcal {A} | d o (\\mathcal {I}, \\mathcal {F}, \\mathcal {B})) = \\sum_ {\\boldsymbol {t}} \\sum_ {\\boldsymbol {c}} P (\\mathcal {A} | \\boldsymbol {f}, \\boldsymbol {t}, \\boldsymbol {c}) P (\\boldsymbol {t}, \\boldsymbol {c}) = \\mathbb {E} _ {\\boldsymbol {t}, \\boldsymbol {c}} [ P (\\mathcal {A} | \\boldsymbol {f}, \\boldsymbol {t}, \\boldsymbol {c}) ]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.322, + 0.784, + 0.367 + ], + "angle": 0, + "content": "Therefore, to predict the alpha matte with confounders intervened, we need to model the relevant spaces and calculate the expectations in Eq. (4), which are further discussed in Sec. 3.3 and Sec. 3.4, respectively." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.368, + 0.785, + 0.504 + ], + "angle": 0, + "content": "However, to calculate the expectation in Eq. (4), obtaining the foreground representation is also necessary. Although the only input is the image (and the possible user guidance), directly inferring the foreground representation from the image representation is biased, as the confounders similarly facilitate backdoor paths to \\(\\mathcal{F}\\) as to \\(\\mathcal{A}\\). Specifically, the backdoor paths \\(\\mathcal{F} \\gets \\mathcal{C} \\to \\mathcal{B}\\), \\(\\mathcal{B} \\gets \\mathcal{I} \\gets \\mathcal{C} \\to \\mathcal{F}\\), and \\(\\mathcal{F} \\gets \\mathcal{T} \\to \\mathcal{I} \\to \\mathcal{B}\\) lead to \\(\\mathcal{F} \\not\\perp \\mathcal{B}|\\mathcal{I}\\), indicating that the background's transparency and contrast similarly affect the foreground as influencing the prediction of the alpha matte. To address this issue, we also conduct an intervention for predicting the foreground, yielding:" + }, + { + "type": "equation", + "bbox": [ + 0.276, + 0.513, + 0.785, + 0.576 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} P (\\mathcal {F} | d o (\\mathcal {I}, \\mathcal {B})) = \\sum_ {\\boldsymbol {t}} \\sum_ {\\boldsymbol {c}} P (\\mathcal {F} | \\boldsymbol {i}, \\boldsymbol {b}, \\boldsymbol {t}, \\boldsymbol {c}) P (\\boldsymbol {c}, \\boldsymbol {t}) \\\\ = \\sum_ {t} ^ {t} \\sum_ {c} ^ {c} P (\\mathcal {F} | \\boldsymbol {i}, \\boldsymbol {t}, \\boldsymbol {c}) P (\\boldsymbol {c}, \\boldsymbol {t}) = \\mathbb {E} _ {\\boldsymbol {c}} \\mathbb {E} _ {\\boldsymbol {t}} [ P (\\mathcal {F} | \\boldsymbol {i}, \\boldsymbol {t}, \\boldsymbol {c}) ], \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.586, + 0.784, + 0.631 + ], + "angle": 0, + "content": "where the multi-scale image representation \\(\\pmb{i} = \\langle \\pmb{i}_1, \\pmb{i}_2, \\dots, \\pmb{i}_n \\rangle\\) is parameterized by the matting backbone, aligned with the scales of \\(\\pmb{C}\\) and \\(\\pmb{T}\\) as described in Sec. 3.3. The expectations are calculated as detailed in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.651, + 0.447, + 0.667 + ], + "angle": 0, + "content": "3.3 Space Representations" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.674, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Transparency Space \\( \\pmb{T} \\). Given the importance of multi-scale information in image matting, we divide the transparency space into several sub-spaces across different scales. Specifically, the transparency space \\( \\pmb{T} = \\langle T_1, T_2, \\dots, T_n \\rangle \\), with \\( \\pmb{T}_i \\) representing the transparency at the \\( i \\)-th scale, and \\( n \\) indicating the total number of sub-spaces. The \\( i \\)-th scale denotes the transparency of the image patches sized \\( 2^i \\times 2^i \\). Considering the average alpha value as an approximation for the \"transparency level\" of an area, we define \\( \\pmb{T}_i \\) as the average alpha value of pixels within the \\( i \\)-th scale patches. This formulation allows us to discretize the transparency sub-spaces further by categorizing the transparency level into \\( K_t \\) distinct classes, given its bounded nature within [0, 1]. The construction of the feature space entails the following steps:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "Z. Liao et al." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "1. Utilize a pretrained image feature extractor to derive the patch features at the \\(i\\)-th scale, for each \\(i \\in \\{1, \\dots, n\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.177, + 0.78, + 0.19 + ], + "angle": 0, + "content": "2. Determine the transparency level category \\( k \\in \\{1, \\dots, K_t\\} \\) for every patch." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.19, + 0.784, + 0.22 + ], + "angle": 0, + "content": "3. Compute the average patch representation for all patches within the \\(k\\)-th category as the representation of this transparency category." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.147, + 0.784, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.224, + 0.784, + 0.255 + ], + "angle": 0, + "content": "Hence, we formulate the feature space \\(\\pmb{T}_i\\in \\mathbb{R}^{d_i\\times K_t}\\), where \\(d_{i}\\) signifies the dimensionality of the feature extracted at the \\(i\\)-th scale." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.256, + 0.785, + 0.331 + ], + "angle": 0, + "content": "Contrast Space \\( C \\). Similarly to the transparency space, the contrast space is partitioned into multiple sub-spaces across various scales. We define \\( C = \\langle C_1, C_2, \\ldots, C_n \\rangle \\), with each \\( C_i \\) representing contrast at the \\( i \\)-th scale, and \\( n \\) marking the total number of sub-spaces. Each \\( i \\)-th scale denotes the contrast of the foreground and background within \\( 2^i \\times 2^i \\) image patches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.332, + 0.784, + 0.511 + ], + "angle": 0, + "content": "Quantifying contrast, however, presents more complexity than measuring transparency. Initially, we estimate the color of a patch using the average color vector of all pixels within it in the color space. Subsequently, the contrast between foreground and background patches is determined by the color distance between their average color vectors. Given the multifaceted nature of color perception, we employ the CIEDE2000 color difference formula [31] for calculating color distance, as it closely aligns with human visual perception of color differences. This formula, incorporating lightness, chroma, and hue, yields a distance within the range of [0, 100]. Following this approach, we categorize the contrast level into \\( K_{c} \\) discrete classes and construct the feature space \\( C_i \\in \\mathbb{R}^{d_i \\times K_c} \\), mirroring the methodology used for the transparency space. The primary adjustment involves transitioning from transparency to contrast categories for the classification." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.784, + 0.633 + ], + "angle": 0, + "content": "Joint Confounder Space \\( \\mathbf{Z} \\). Given that \\( \\mathcal{T} \\) and \\( \\mathcal{C} \\) may not be independent, we cannot decompose the joint expectation in Eq. (4) into separate expectations for each. Therefore, it becomes necessary to model the joint space of \\( \\mathbf{T} \\) and \\( \\mathbf{C} \\). An effective approach is to define the joint space as the Cartesian product of the sub-spaces. Consequently, the joint confounder space \\( \\mathbf{Z} = \\langle \\mathbf{Z}_1, \\mathbf{Z}_2, \\dots, \\mathbf{Z}_n \\rangle = \\langle \\mathbf{T}_1 \\times \\mathbf{C}_1, \\mathbf{T}_2 \\times \\mathbf{C}_2, \\dots, \\mathbf{T}_n \\times \\mathbf{C}_n \\rangle \\), where each \\( \\mathbf{T}_i \\times \\mathbf{C}_i \\in \\mathbb{R}^{d_i \\times (K_t \\cdot K_c)} \\) represents a multidimensional space formed by the Cartesian product of transparency and contrast spaces at the \\( i \\)-th scale. Thus, Eqs. (4) and (5) are simplified as following:" + }, + { + "type": "equation", + "bbox": [ + 0.303, + 0.652, + 0.701, + 0.714 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} P(\\mathcal{A}|do(\\mathcal{I},\\mathcal{F},\\mathcal{B})) = \\sum_{\\boldsymbol {z}}P(\\mathcal{A}|\\boldsymbol {f},\\boldsymbol {z})P(\\boldsymbol {z}) = \\mathbb{E}_{\\boldsymbol{z}}[P(\\mathcal{A}|\\boldsymbol {f},\\boldsymbol {z})], \\\\ P (\\mathcal {F} | d o (\\mathcal {I}, \\mathcal {B})) = \\sum_ {\\boldsymbol {z}} P (\\mathcal {F} | \\boldsymbol {i}, \\boldsymbol {z}) P (\\boldsymbol {z}) = \\mathbb {E} _ {\\boldsymbol {z}} [ P (\\mathcal {F} | \\boldsymbol {i}, \\boldsymbol {z}) ], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.774, + 0.737 + ], + "angle": 0, + "content": "where \\(z = \\langle \\pmb {c},\\pmb {t}\\rangle\\) represents the split of joint observed confounders \\(\\mathcal{Z} = \\langle \\mathcal{T},\\mathcal{C}\\rangle\\)" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.753, + 0.371, + 0.768 + ], + "angle": 0, + "content": "3.4 Expectations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.774, + 0.784, + 0.819 + ], + "angle": 0, + "content": "The expectations are calculated over the joint confounder space \\( \\mathbf{Z} \\). Utilizing the Normalized Weighted Geometric Mean (NWGM) approach [32,39], we reformulate the expectation as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.264, + 0.826, + 0.784, + 0.842 + ], + "angle": 0, + "content": "\\[\nP (\\mathcal {A} | d o (\\mathcal {I}, \\mathcal {F}, \\mathcal {B})) = \\mathbb {E} _ {\\boldsymbol {z}} P (\\mathcal {A} | \\boldsymbol {f}, \\boldsymbol {z}) = \\mathbb {E} _ {\\boldsymbol {z}} \\sigma (g (\\boldsymbol {f}, \\boldsymbol {z})) \\approx \\sigma (g (\\boldsymbol {f}, \\mathbb {E} _ {\\boldsymbol {z}} [ \\boldsymbol {z} ])), \\tag {7}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.334, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "COIN-Matting: Confounder Intervention for Image Matting" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.146, + 0.782, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.28, + 0.788, + 0.351 + ], + "angle": 0, + "content": "Fig. 4: The COIN framework. \\( \\pmb{i} \\) and \\( \\pmb{f} \\) denote the image feature and foreground feature. \\( T, C \\) and \\( Z \\) are transparency space, contrast space, and their joint confounder space respectively, which are constructed in Sec. 3.3. Spaces corresponding to the confounder variable are marked gray. EXP represents the expectations described in Sec. 3.4. \\( \\sigma(\\cdot) \\) and \\( g(\\cdot) \\) are ignored for simplicity. \\( \\odot \\) represents the dot production." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.788, + 0.407 + ], + "angle": 0, + "content": "where \\(\\sigma(\\cdot)\\) denotes a sigmoid function and \\(g(\\cdot)\\) represents a linear transformation for each sub-space. Similarly, for the foreground parameterization:" + }, + { + "type": "equation", + "bbox": [ + 0.299, + 0.413, + 0.786, + 0.43 + ], + "angle": 0, + "content": "\\[\nP (\\mathcal {F} | d o (\\mathcal {I}, \\mathcal {B})) = \\mathbb {E} _ {z} P (\\mathcal {F} | \\boldsymbol {i}, z) = \\mathbb {E} _ {z} g (\\boldsymbol {i}, z) = g (\\boldsymbol {i}, \\mathbb {E} _ {z} [ z ]). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.788, + 0.527 + ], + "angle": 0, + "content": "This approach shifts the expectations to the feature level. To prevent model collapse, we use \\(\\mathbb{E}_{\\boldsymbol{z}|\\mathcal{F}}\\) and \\(\\mathbb{E}_{\\boldsymbol{z}|\\mathcal{I}}\\) as approximations for \\(\\mathbb{E}_z\\) respectively. To emulate these feature-level expectations, we leverage a sub-space-wise attention mechanism. Take \\(\\mathbb{E}_{\\boldsymbol{z}|\\mathcal{I}}[\\boldsymbol{z}]\\) as an example. As spaces \\(Z_{i}\\) are heterogeneous with the feature spaces of \\(\\pmb{i}_j\\) when \\(i\\neq j\\), we apply the attention mechanism for each scale respectively. Therefore, the expectation is" + }, + { + "type": "equation", + "bbox": [ + 0.324, + 0.533, + 0.786, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\boldsymbol {z} \\mid \\mathcal {I}} [ \\boldsymbol {z} ] = \\left\\langle \\mathbb {E} _ {\\boldsymbol {z} _ {1} \\mid \\mathcal {I} _ {1}} [ \\boldsymbol {z} _ {1} ], \\mathbb {E} _ {\\boldsymbol {z} _ {2} \\mid \\mathcal {I} _ {2}} [ \\boldsymbol {z} _ {2} ], \\dots , \\mathbb {E} _ {\\boldsymbol {z} _ {n} \\mid \\mathcal {I} _ {n}} [ \\boldsymbol {z} _ {n} ] \\right\\rangle . \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.557, + 0.788, + 0.65 + ], + "angle": 0, + "content": "where \\( n \\) is the number of sub-spaces. For attention on image patches at the \\( i \\)-th scale, the input is the contrast sub-space \\( Z_{i} \\in \\mathbb{R}^{d_{i} \\times (K_{t} \\cdot K_{c})} \\) and the image patch representation \\( \\hat{\\pmb{i}}_i \\in \\mathbb{R}^{d_i} \\) parameterized by the matting backbone, where \\( d_{i} \\) is the dimension of the latent representations. During such attention, the attention score emulates the distribution of \\( P(\\pmb{z}|\\mathcal{I}) \\), thus the weighted sum on values (i.e., \\( z \\)) emulates the expectation \\( \\mathbb{E}_{\\pmb{z}|\\mathcal{I}}[\\pmb{z}] \\). The mathematical formulation is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.273, + 0.658, + 0.637, + 0.678 + ], + "angle": 0, + "content": "\\[\n\\pmb {h e a d} _ {i} ^ {k} = \\mathrm {s o f t m a x} (\\hat {\\pmb {i}} _ {i} ^ {k} \\pmb {W} _ {i, 1} ^ {k} (\\pmb {Z} _ {i} \\pmb {W} _ {i, 2} ^ {k}) ^ {T} / \\sqrt {d}) (\\pmb {Z} _ {i} \\pmb {W} _ {i, 3} ^ {k})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.3, + 0.681, + 0.785, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\bar {z} _ {\\mathcal {I} _ {i}} = \\mathbb {E} _ {\\boldsymbol {z} _ {i} \\mid \\mathcal {I} _ {i}} [ \\boldsymbol {z} _ {i} ] = \\operatorname {C o n c a t} \\left(\\boldsymbol {h e a d} _ {i} ^ {1}, \\boldsymbol {h e a d} _ {i} ^ {2}, \\dots , \\boldsymbol {h e a d} _ {i} ^ {N _ {h}}\\right) \\boldsymbol {W} _ {i, o} \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.306, + 0.703, + 0.648, + 0.721 + ], + "angle": 0, + "content": "\\[\n\\bar {z} _ {\\mathcal {I}} = \\operatorname {E X P} (\\hat {\\boldsymbol {i}}, \\boldsymbol {Z}) = \\mathbb {E} _ {\\boldsymbol {z} | \\mathcal {I}} [ \\boldsymbol {z} ] = \\left\\langle \\bar {z} _ {\\mathcal {I} _ {1}}, \\bar {z} _ {\\mathcal {I} _ {2}}, \\dots , \\bar {z} _ {\\mathcal {I} _ {n}} \\right\\rangle ,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.724, + 0.788, + 0.772 + ], + "angle": 0, + "content": "where \\( \\pmb{W}_{i,1}^{k}, \\pmb{W}_{i,2}^{k}, \\pmb{W}_{i,3}^{k}, \\pmb{W}_{i,o} \\) are trainable parameters, \\( N_{h} \\) represents the number of heads, and \\( \\mathrm{Concat}(\\cdot) \\) is the concatenation operation. This process effectively captures the expectations across different confounder sub-spaces." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.789, + 0.485, + 0.804 + ], + "angle": 0, + "content": "3.5 COIN Matting Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The overall structure of our COIN framework is shown in Fig. 4. The confounder space \\(\\pmb{T}\\) is similarly constructed as \\(\\mathbf{C}\\), as described in Sec. 3.3. The mat" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "Z. Liao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.314 + ], + "angle": 0, + "content": "ting backbone is employed to parameterize the image representation \\(\\hat{\\pmb{i}}\\). Rather than directly predicting \\(P(\\mathcal{A}|\\mathcal{I})\\) as the matting backbone does, we engage in confounder intervention through backdoor adjustment to address the issues of contrast bias and transparency bias. The backdoor adjustment is implemented by the feature-level expectations. Specifically, we first process the joint space \\(\\pmb{Z}\\) alongside the image representation \\(\\pmb{i}\\) to compute \\(\\mathbb{E}_{z|\\mathcal{I}}[z]\\) via attention. This expectation is subsequently concatenated with the image representations to yield the foreground representation. Similarly, the foreground representation is used to calculate \\(\\mathbb{E}_{z|\\mathcal{F}}[z]\\) with joint confounder space \\(\\pmb{Z}\\), and the expectations are further concatenated with the foreground representation to produce the alpha matte at various scales. Thus, we can express the overall framework as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.276, + 0.324, + 0.786, + 0.351 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {i}} = \\text {M a t t i n g B a c k b o n e} (\\boldsymbol {I}); \\quad \\bar {z} _ {\\mathcal {I}} = \\operatorname {E X P} (\\hat {\\boldsymbol {i}}, \\boldsymbol {Z}); \\quad \\hat {\\boldsymbol {f}} = g _ {f} (\\hat {\\boldsymbol {i}}, \\bar {z} _ {\\mathcal {I}}); \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.262, + 0.348, + 0.682, + 0.364 + ], + "angle": 0, + "content": "\\[\n\\bar {z} _ {\\mathcal {F}} = \\operatorname {E X P} (\\hat {\\boldsymbol {f}}, \\boldsymbol {Z}); \\quad \\hat {\\alpha} = P (\\mathcal {A} | d o (\\mathcal {I}, \\mathcal {F}, \\mathcal {B})) = \\sigma (g _ {\\alpha} (\\hat {\\boldsymbol {f}}, \\bar {z} _ {\\mathcal {F}})),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.784, + 0.404 + ], + "angle": 0, + "content": "where \\( g_{f}(\\cdot) \\) and \\( g_{\\alpha}(\\cdot) \\) represent per-sub-space linear transformations, and \\( \\mathrm{EXP}(\\cdot, \\cdot) \\) denotes the feature-level expectations described in Sec. 3.4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.788, + 0.557 + ], + "angle": 0, + "content": "During the training of this framework, we supervise the alpha matte at each scale, similar to the approach taken by the backbone. Furthermore, we supervise the attention score, which approximates the probability distribution of the confounders when calculating expectations, with the ground truth categories of contrast and transparency. It is important to note that a closer category index of contrast and transparency signifies a more similar meaning, as these are discretized from the continuous values of color distance and transparency. Therefore, we employ Gaussian label smoothing prior to the cross-entropy for supervising the attention score. Additional details regarding the framework can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.58, + 0.376, + 0.597 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.788, + 0.749 + ], + "angle": 0, + "content": "In this section, we evaluate our COIN framework on extensive image matting baselines across various datasets. Specifically, we integrate our framework with SOTA image matting methods across all three types of matting methods: trimap-based methods (i.e., MatteFormer [25]), flexible guidance-based methods (i.e., DIIM [6] and dugMatting [38]) and guidance-free methods (i.e., dugMatting without human interaction [38]). We evaluate the improvement of our framework on Composition-1k [40], and Distinctions-646 [43] datasets. Additionally, we elaborate on the benefits of intervening on each confounder and the computational cost of implementing our framework in Sec. 4.3. The qualitative results are in Sec. 4.4." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.77, + 0.433, + 0.787 + ], + "angle": 0, + "content": "4.1 Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Datasets. We conduct the experiments on Composition-1k [40] and Distinctions-646 [43] datasets following [25]. Composition-1k dataset consists of 431 and 50 foreground objects in the training set and test set respectively. These foregrounds" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.334, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "COIN-Matting: Confounder Intervention for Image Matting" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.237 + ], + "angle": 0, + "content": "are composited with MS COCO [17] images for training and with VOC 2012 [7] images for testing. For testing, each foreground object is combined with 20 different images, generating a total of 1000 test images. The Distinctions-646 dataset contains 646 unique foreground objects, which are divided into 596 for training and 50 for testing. The composition process with background images follows the same procedure as in the Composition-1k dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.785, + 0.344 + ], + "angle": 0, + "content": "Metrics. For measuring the overall performance of our framework, we adopt the widely used image matting metrics [12,25,26,43]: sum of absolute difference (SAD), mean square error (MSE), and the gradient (Grad) and connectivity (Conn). Additionally, to assess the improvement in reducing transparency and contrast biases, we measure SAD conditioned on each contrast and transparency category. The number of parameters and FLOPs are also reported to evaluate the computational complexity of implementing our framework." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.787, + 0.435 + ], + "angle": 0, + "content": "Implementation Details. We set the number of sub-spaces as \\( n = 5 \\) for \\( I, F, A, T \\), and \\( C \\). For transparency and contrast levels, we set both \\( K_{t} \\) and \\( K_{c} \\) to 25. The contrast and transparency space is initialized by the average multiscale feature extracted by pretrained Swin-B [19, 20] as described in Sec. 3.3. We employ a \\( \\mathcal{N}(0, 1.2) \\) Gaussian distribution to smooth the label of attention score. More details can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.463, + 0.373, + 0.477 + ], + "angle": 0, + "content": "4.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.734 + ], + "angle": 0, + "content": "The improvement of our COIN framework is demonstrated in Tab. 1. It is evident that by intervening on the confounder variables, our COIN framework significantly enhances the performance across all matting baselines with varying guidance mechanisms on both the Composition-1k and Distinctions-646 datasets. Our framework substantially improves results for both guidance-free methods (e.g., dugMatting (0-sel)) and flexible-guidance based methods (e.g., dugMatting (1-sel) and DIIM). Additionally, even for the SOTA trimap-based baseline (e.g., MatteFormer), our COIN framework achieves significant performance gains, decreasing SAD by 1.9 and 2.6, and MSE by 0.7 and 1.5 on the two datasets respectively. Notably, our framework tends to exhibit greater improvements on the Distinctions-646 dataset compared to the Composition-1k dataset. This discrepancy is likely attributable to the more pronounced biases present within the Distinctions-646 dataset. Furthermore, enhancements are generally more substantial for flexible guidance-based or guidance-free matting methods than for trimap-based methods, since the former methods are typically more susceptible to biases in the absence of the detailed information provided by a trimap." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To demonstrate that the enhancements are truly attributable to the reduction of contrast bias and transparency bias, we further evaluate the SAD improvements across each contrast and transparency category, as shown in Fig. 5. We selected MatteFormer as the baseline because it presents the most substantial challenge and exhibits the highest performance among the baselines. Regarding transparency, it is observable that our framework enhances the performance of baseline matting methods across all transparency levels. Furthermore, we note" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "Z. Liao et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 1: The comparison of applying our COIN framework on Composition-1k and Distinctions-646 dataset. The improvements are highlighted as superscripts." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.185, + 0.76, + 0.304 + ], + "angle": 0, + "content": "
MethodComposition-1KDistinctions-646
SAD ↓MSE ↓Grad ↓Conn ↓SAD ↓MSE ↓Grad ↓Conn ↓
DIIM37.510.318.534.536.211.819.635.6
dugMatting (0-sel)*34.15.715.631.233.28.818.233.2
dugMatting (1-sel)*25.84.39.722.324.17.112.424.0
MatteFormer23.84.08.718.921.96.611.220.5
COIN(DIIM)33.8-3.78.1-2.232.0-2.59.6-2.2
COIN(dugMatting (0-sel)*)30.8-3.34.5-1.228.0-3.27.4-1.4
COIN(dugMatting (1-sel)*)23.7-2.13.5-0.819.1-3.26.3-0.8
COIN(MatteFormer)21.9-1.93.3-0.716.8-2.15.1-1.5
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.22, + 0.305, + 0.61, + 0.319 + ], + "angle": 0, + "content": "* \\( k \\)-sel denotes \\( k \\) times human interaction during matting." + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.354, + 0.488, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.481, + 0.487, + 0.494 + ], + "angle": 0, + "content": "(a) The SAD conditioned on transparency." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.354, + 0.765, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.481, + 0.751, + 0.493 + ], + "angle": 0, + "content": "(b) The SAD conditioned on contrast." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.505, + 0.787, + 0.534 + ], + "angle": 0, + "content": "Fig. 5: The improvement of applying our framework to MatteFormer on Composition-1K dataset conditioned on each contrast and transparency category." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.841 + ], + "angle": 0, + "content": "that the improvement is particularly pronounced in regions where the transparency level is within the range of [50, 85]. The training data within this range, as discussed in Sec. 1, is notably scarce but exhibits a unique alpha pattern which differs from the pattern in highly opaque or transparent foregrounds. The observed probability distribution, dominated by data from regions with high opacity or transparency, causes the transparency bias we identified. The significant enhancements in these regions indicate that our framework is reducing the transparency bias through confounder intervention. In terms of contrast, our framework not only significantly reduces the SAD error across all contrast levels compared to the baseline method but also shows that the most notable improvement occurs in areas of relatively low contrast, particularly where the contrast level ranges from [10, 25]. In these regions, the detail patterns increase rapidly as average contrast value and mean alpha value raises (as shown in Fig. 2a). However, owing to the insufficient data relative to other regions, these areas are most susceptible to contrast bias. Thus, the most notable improvements in low-contrast areas suggest that our framework enhances performance by alleviating the contrast bias through intervention on the contrast confounder." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.334, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "COIN-Matting: Confounder Intervention for Image Matting" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 2: The comparison of applying intervention on different confounder variables on the Distinctions-646 dataset using MatteFormer as the baseline. The improvements are highlighted as superscripts. The GFLOPs is calculated on \\(1024 \\times 1024\\) image." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.199, + 0.77, + 0.283 + ], + "angle": 0, + "content": "
MethodErrorsComplexity
SAD ↓MSE ↓Grad ↓Conn ↓ParametersFLOPs
MatteFormer21.9-1.56.6-0.811.2-0.920.544.8M459G
MatteFormer + T intervene20.45.810.319.1-1.447.7M
MatteFormer + C intervene20.2-1.75.7-0.910.0-1.218.8-1.747.7M
COIN(MatteFormer)19.3-2.65.1-1.59.4-1.817.8-2.749.5M
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.31, + 0.39, + 0.327 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.788, + 0.518 + ], + "angle": 0, + "content": "Confounder Spaces. We discuss the improvement from intervening in each confounder space individually rather than their joint confounder in Tab. 2. By comparing line 3 and line 2 with line 1, it is observed that intervening on either \\(\\mathcal{C}\\) or \\(\\mathcal{T}\\) can improve the performance of the baseline methods. However, because such interventions do not block the backdoor path opened by \\(\\mathcal{T}\\) and \\(\\mathcal{C}\\) respectively, the condition \\(\\mathcal{A} \\nsubseteq \\mathcal{B}|\\mathcal{T}\\) still exists, which leads to transparency bias and contrast bias respectively, resulting in a limited performance of the intervention framework. Moreover, it was noted that the performance of intervening on \\(\\mathcal{C}\\) is slightly better than that of intervening on \\(\\mathcal{T}\\). This may be due to the fact that the contrast bias is more pronounced than the transparency bias in the dataset. By jointly intervening both \\(\\mathcal{C}\\) and \\(\\mathcal{T}\\), we may block all aforementioned backdoor paths, thus further performance improvement in the COIN framework." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.788, + 0.73 + ], + "angle": 0, + "content": "Complexity Analysis. As we introduce a framework atop existing matting methods, we further analyze its space and time complexity in terms of both parameters and FLOPs to delineate the cost of integrating our framework. As indicated in Tab. 2, the additional cost imposed by our framework is minimal when compared to the baseline methods. The incremental parameters (about \\(10.5\\%\\)) stem exclusively from the feature space and its corresponding attention mechanism, which are negligible relative to the baseline model. The only extra FLOPs (about \\(10.6\\%\\)) originate from the computation of expectations within feature spaces, which are characterized as \\(n\\) attentions across \\(d_{i} \\times (K_{t} \\cdot K_{c})\\) matrices and \\(d_{i} \\times L\\) image features, where \\(L\\) denotes the number of image patches at the \\(i\\)-th scale. Such attention is also lightweight compared to the matting baseline. Overall, our framework improves the performance of the SOTA image matting baseline between \\(11.6\\%\\) and \\(22.7\\%\\) in different metrics with about \\(10\\%\\) computation cost, which is remarkably efficient for integrating with the baselines." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.754, + 0.422, + 0.769 + ], + "angle": 0, + "content": "4.4 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this section, we present qualitative comparisons of our COIN framework. In Fig. 6, we compare the alpha matting results from baseline methods with those obtained using our COIN framework. Rows 1 showcase examples illustrating the reduction of contrast bias. It is evident that the low-contrast areas" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Z. Liao et al." + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.143, + 0.34, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.269, + 0.342, + 0.306, + 0.352 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.144, + 0.447, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.374, + 0.342, + 0.418, + 0.352 + ], + "angle": 0, + "content": "Trimap" + }, + { + "type": "image", + "bbox": [ + 0.45, + 0.144, + 0.555, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.341, + 0.54, + 0.352 + ], + "angle": 0, + "content": "Alpha matte" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.144, + 0.662, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.341, + 0.648, + 0.352 + ], + "angle": 0, + "content": "MatteFormer" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.144, + 0.77, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.341, + 0.734, + 0.351 + ], + "angle": 0, + "content": "COIN" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.364, + 0.788, + 0.394 + ], + "angle": 0, + "content": "Fig. 6: The visualized comparison of applying our COIN framework on Matte-Former [25]. Images are best viewed when zoomed in." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.789, + 0.588 + ], + "angle": 0, + "content": "highlighted in the red boxes yield alpha mattes with lower values when processed by MatteFormer. However, the integration of our COIN framework significantly diminishes this contrast bias, resulting in more precise alpha mattes. Rows 2 and 3 demonstrate the reduction of transparency bias. The highlighted mid-transparent regions in the red boxes contain background textures (i.e., the texture of text and the car cage respectively) in the predicted alpha mattes. Nevertheless, with the application of our COIN framework, the alpha mattes become markedly more accurate, effectively eliminating such textures from the mid-transparent areas. Such qualitative results prove that our COIN framework can effectively reduce both contrast bias and transparency bias, therefore improving the performance of the existing image matting methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.608, + 0.36, + 0.625 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.638, + 0.788, + 0.73 + ], + "angle": 0, + "content": "In this work, we have concentrated on the contrast bias and transparency bias in existing matting methods and demonstrated how dataset biases lead to biased behaviors of models. We analyzed these biases from a causal perspective and designed the confounder intervention framework to alleviate them. Extensive experiments have demonstrated that our framework markedly alleviates these biases and enhances SOTA matting methods with minor computational cost." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.751, + 0.405, + 0.768 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The work was supported by the National Natural Science Foundation of China (Grant No. 62076162), the Shanghai Municipal Science and Technology Major Project, China (Grant No. 2021SHZDZX0102) and the Postdoctoral Fellowship Program of CPSF (Grant No. GZC20241225)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.334, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "COIN-Matting: Confounder Intervention for Image Matting" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.207 + ], + "angle": 0, + "content": "1. Cai, H., Xue, F., Xu, L., Guo, L.: Transmatting: Enhancing transparent objects matting with transformers. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.208, + 0.785, + 0.235 + ], + "angle": 0, + "content": "2. Chen, J., Li, X., Luo, L., Ma, J.: Multi-focus image fusion based on multi-scale gradients and image matting. IEEE Trans. Multim. 24, 655-667 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.236, + 0.785, + 0.263 + ], + "angle": 0, + "content": "3. Dai, Y., Lu, H., Shen, C.: Learning affinity-aware upsampling for deep image matting. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.264, + 0.785, + 0.292 + ], + "angle": 0, + "content": "4. Dai, Y., Price, B., Zhang, H., Shen, C.: Boosting robustness of image matting with context assembling and strong data augmentation. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.293, + 0.785, + 0.319 + ], + "angle": 0, + "content": "5. Deng, J., Xu, Y., Zhou, Z., He, S.: Background matting via recursive excitation. In: ICME (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.321, + 0.785, + 0.348 + ], + "angle": 0, + "content": "6. Ding, H., Zhang, H., Liu, C., Jiang, X.: Deep interactive image matting with feature propagation. IEEE Trans. Image Process. 31, 2421-2432 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.349, + 0.785, + 0.376 + ], + "angle": 0, + "content": "7. Everingham, M., Gool, L.V., Williams, C.K.I., Winn, J.M., Zisserman, A.: The pascal visual object classes (VOC) challenge. Int. J. Comput. Vis. (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.377, + 0.785, + 0.417 + ], + "angle": 0, + "content": "8. Fang, X., Zhang, S., Chen, T., Wu, X., Shamir, A., Hu, S.: User-guided deep human image matting using arbitrary trimaps. IEEE Trans. Image Process. 31, 2040-2052 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.419, + 0.785, + 0.446 + ], + "angle": 0, + "content": "9. Hou, Q., Liu, F.: Context-aware image matting for simultaneous foreground and alpha estimation. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.447, + 0.785, + 0.473 + ], + "angle": 0, + "content": "0. Li, J., Niu, L., Zhang, L.: Knowledge proxy intervention for deconfounded video question answering. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.475, + 0.785, + 0.503 + ], + "angle": 0, + "content": "1. Li, J., Wang, W., Chen, J., Niu, L., Si, J., Qian, C., Zhang, L.: Video semantic segmentation via sparse temporal transformer. In: ACM MM (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.504, + 0.785, + 0.517 + ], + "angle": 0, + "content": "2. Li, J., Zhang, J., Tao, D.: Deep automatic natural image matting. In: IJCAI (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.518, + 0.785, + 0.544 + ], + "angle": 0, + "content": "3. Li, J., Zhang, J., Tao, D.: Referring image matting. In: Proceedings of the IEEE Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.546, + 0.785, + 0.572 + ], + "angle": 0, + "content": "4. Li, Y., Lu, H.: Natural image matting via guided contextual attention. In: AAAI (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.574, + 0.785, + 0.601 + ], + "angle": 0, + "content": "5. Li, Y., Wang, X., Xiao, J., Chua, T.: Equivariant and invariant grounding for video question answering. In: ACM MM (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.602, + 0.785, + 0.629 + ], + "angle": 0, + "content": "6. Lin, S., Ryabtsev, A., Sengupta, S., Curless, B.L., Seitz, S.M., Kemelmacher-Shlizerman, I.: Real-time high-resolution background matting. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.63, + 0.785, + 0.657 + ], + "angle": 0, + "content": "7. Lin, T., Maire, M., Belongie, S.J., Hays, J., Perona, P., Ramanan, D., Dóllar, P., Zitnick, C.L.: Microsoft COCO: common objects in context. In: ECCV (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.785, + 0.685 + ], + "angle": 0, + "content": "8. Liu, Y., Xie, J., Qiao, Y., Tang, Y., Yang, X.: Prior-induced information alignment for image matting. IEEE Trans. Multim. 24, 2727-2738 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.687, + 0.785, + 0.727 + ], + "angle": 0, + "content": "9. Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., Wei, F., Guo, B.: Swin transformer v2: Scaling up capacity and resolution. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.728, + 0.785, + 0.755 + ], + "angle": 0, + "content": "20. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.784 + ], + "angle": 0, + "content": "21. Lu, H., Dai, Y., Shen, C., Xu, S.: Indices matter: Learning to index for deep image matting. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.812 + ], + "angle": 0, + "content": "22. Lu, L., Li, J., Cao, J., Niu, L., Zhang, L.: Painterly image harmonization using diffusion model. In: ACM MM (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "23. Lutz, S., Amplanitis, K., Smolic, A.: Alphagan: Generative adversarial networks for natural image matting. In: BMVC (2018)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "Z. Liao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "24. Niu, Y., Tang, K., Zhang, H., Lu, Z., Hua, X., Wen, J.: Counterfactual VQA: A cause-effect look at language bias. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.203 + ], + "angle": 0, + "content": "25. Park, G., Son, S., Yoo, J., Kim, S., Kwak, N.: Matteformer: Transformer-based image matting via prior-tokens. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "26. Park, K., Woo, S., Oh, S.W., Kweon, I.S., Lee, J.Y.: Mask-guided matting in the wild. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.245 + ], + "angle": 0, + "content": "27. Pearl, J., Glymour, M., Jewell, N.P.: Causal inference in statistics: A primer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.272 + ], + "angle": 0, + "content": "28. Pearl, J., Mackenzie, D.: The Book of Why: The New Science of Cause and Effect. Basic Books, Inc. (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "29. Rubin, D.B.: Essential concepts of causal inference: a remarkable history and an intriguing future. Biostatistics & Epidemiology (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.328 + ], + "angle": 0, + "content": "30. Sengupta, S., Jayaram, V., Curless, B., Seitz, S.M., Kemelmacher-Shlizerman, I.: Background matting: The world is your green screen. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.785, + 0.383 + ], + "angle": 0, + "content": "31. Sharma, G., Wu, W., Dalal, E.N.: The ciede2000 color-difference formula: Implementation notes, supplementary test data, and mathematical observations. Color Research and Application 30, 21-30 (2005), https://api-semanticscholar.org/ CorpusID:29119937" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "32. Srivastava, N., Hinton, G.E., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. J. Mach. Learn. Res. (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "33. Sun, Y., Tang, C., Tai, Y.: Ultrahigh resolution image/video matting with spatiotemporal sparsity. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "34. Tan, L., Li, J., Niu, L., Zhang, L.: Deep image harmonization in dual color spaces. In: ACM MM (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.785, + 0.508 + ], + "angle": 0, + "content": "35. Wang, T., Zhou, C., Sun, Q., Zhang, H.: Causal attention for unbiased visual recognition. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "36. Wang, W., Feng, F., He, X., Zhang, H., seng Chua, T.: Clicks can be cheating: Counterfactual recommendation for mitigating clickbait issue (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.563 + ], + "angle": 0, + "content": "37. Wei, T., Chen, D., Zhou, W., Liao, J., Zhao, H., Zhang, W., Hua, G., Yu, N.: Deep image matting with sparse user interactions. TPAMI (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.591 + ], + "angle": 0, + "content": "38. Wu, J., Zhang, C., Li, Z., Fu, H., Peng, X., Zhou, J.T.: dugMatting: Decomposed-uncertainty-guided matting. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.632 + ], + "angle": 0, + "content": "39. Xu, K., Ba, J., Kiros, R., Cho, K., Courville, A.C., Salakhutdinov, R., Zemel, R.S., Bengio, Y.: Show, attend and tell: Neural image caption generation with visual attention. In: ICML (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.647 + ], + "angle": 0, + "content": "40. Xu, N., Price, B.L., Cohen, S., Huang, T.S.: Deep image matting. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.648, + 0.785, + 0.674 + ], + "angle": 0, + "content": "41. Yao, J., Wang, X., Yang, S., Wang, B.: Vitmatte: Boosting image matting with pre-trained plain vision transformers. Inf. Fusion 103, 102091 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "42. Ye, Z., Dai, Y., Hong, C., Cao, Z., Lu, H.: Infusing definiteness into randomness: Rethinking composition styles for deep image matting. In: AAAI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.785, + 0.73 + ], + "angle": 0, + "content": "43. Yu, Q., Zhang, J., Zhang, H., Wang, Y., Lin, Z., Xu, N., Bai, Y., Yuille, A.L.: Mask guided matting via progressive refinement network. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.756 + ], + "angle": 0, + "content": "44. Zhang, D., Zhang, H., Tang, J., Hua, X., Sun, Q.: Causal intervention for weakly-supervised semantic segmentation. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.785, + 0.785 + ], + "angle": 0, + "content": "45. Zhu, Q., Zhang, W.N., Liu, T., Wang, W.Y.: Counterfactual off-policy training for neural dialogue generation. In: EMNLP (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.785 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/COIN-Matting_ Confounder Intervention for Image Matting/7e767e14-ba96-40b5-847c-1a51f71239aa_origin.pdf b/2024/COIN-Matting_ Confounder Intervention for Image Matting/7e767e14-ba96-40b5-847c-1a51f71239aa_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..011ee546eedce74781c016523a2a5881638fcd16 --- /dev/null +++ b/2024/COIN-Matting_ Confounder Intervention for Image Matting/7e767e14-ba96-40b5-847c-1a51f71239aa_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:979205df5d41ee78c31d3ab83e2564a06fa2cce2635f2b160cf80670ee40cea4 +size 45276674 diff --git a/2024/COIN-Matting_ Confounder Intervention for Image Matting/full.md b/2024/COIN-Matting_ Confounder Intervention for Image Matting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dfa853ea2540ad75a8d4fc631b3fc96a3f02b721 --- /dev/null +++ b/2024/COIN-Matting_ Confounder Intervention for Image Matting/full.md @@ -0,0 +1,359 @@ +# COIN-Matting: Confounder Intervention for Image Matting + +Zhaohe Liao $^{1}$ , Jiangtong Li $^{2}$ , Jun Lan $^{3}$ , Huijia Zhu $^{3}$ , Weiqiang Wang $^{3}$ , Li Niu $^{*1,4}$ , and Liqing Zhang $^{*1}$ + +1 Shanghai Jiao Tong University +{zhaoheliao, ustcnewly, zhang-lq}@sjtu.edu.cn +2 Tongji University +jiangtongli@tongji.edu.cn +3 Ant Group +{yelan.lj, huijia.zhj, weiqiang.wwq}@antgroup.com +4 miguo.ai + +Abstract. Deep learning methods have significantly advanced the performance of image matting. However, dataset biases can mislead the matting models to biased behavior. In this paper, we identify the two typical biases in existing matting models, specifically contrast bias and transparency bias, and discuss their origins in matting datasets. To address these biases, we model the image matting task from the perspective of causal inference and identify the root causes of these biases: the confounders. To mitigate the effects of these confounders, we employ causal intervention through backdoor adjustment and introduce a novel model-agnostic cofounder intervened (COIN) matting framework. Extensive experiments across various matting methods and datasets have demonstrated that our COIN framework can significantly diminish such biases, thereby enhancing the performance of existing matting models. + +Keywords: Image matting $\cdot$ Intervention $\cdot$ Dataset bias + +# 1 Introduction + +Image matting aims at separating accurate foreground objects from natural images and precisely estimating the opacity on the boundary of the object. It serves as the basis for image editing, live-streaming, virtual meetings, and movie production [2,16,22,30,34]. Given an image $\mathbf{I}$ , it is formulated as: + +$$ +\boldsymbol {I} = \boldsymbol {F} \circ \boldsymbol {\alpha} + \boldsymbol {B} \circ (1 - \boldsymbol {\alpha}), \tag {1} +$$ + +where $\pmb{F}$ represents the foreground object, $\pmb{B}$ denotes the background, $\alpha \in [0,1]$ is the opacity map, and $\circ$ denotes element-wise multiplication. As shown in Eq. (1), solving the 7 unknown values in $\{\pmb{F},\pmb{B},\pmb{\alpha}\}$ from the given 3 values in $\pmb{I}$ is a highly ill-posed problem. Therefore, some matting methods require + +![](images/14de91af92126b961678bc5b518355048c06a5977c6b42e37a6e943820fcfffc.jpg) +Image + +![](images/93b59149e5d997bb10c7a46c5ee6c0d9642edb8dca81ad0bd926c811feb11b36.jpg) +$\alpha$ + +![](images/6fdae009ebc94decdb41c076f87db2416bd95ba101fd64fd99f5dcdfcdb3903d.jpg) +a + +![](images/245126f2418f0d5cde82433173da08c8bf60ffdbad94e5273e1d99acd8aa5ed1.jpg) +Image + +![](images/3f25e11c133b67eabd6dc750d8faccf304318fc3be604b10933116464c5125f1.jpg) +$\alpha$ + +![](images/5639fd6a889a84c138afdc7b02b1f7857f8e7ffd9e562e40f10c6100eb63cd18.jpg) +$\hat{\alpha}$ + +![](images/896556478f5b782d812faf379a4f5fcafbaad205799e9372396f05f4e7cd0b1e.jpg) +Image + +![](images/e429041996ebb57e5f0459a82a60ee49e6eaaa1773e78aa3c03c2a2d88dadc84.jpg) +$\alpha$ + +![](images/388d52e96579fad8af456bf4f2dfdacad8019d3bb8ea8364a38cb13ab48e9396.jpg) +(a) The contrast bias. +a +(b) The transparency bias. + +![](images/999460dfdc2888e0ccbc7f6b71320b7443a490e69a38c13ea22ee5e0e7eef765.jpg) +Image +Fig. 1: The observed contrast bias and transparency bias for MatteFormer [25] on the Compositional-1K dataset [40]. $\alpha$ and $\hat{\alpha}$ denote the ground truth and predicted alpha mattes, respectively. Images are best viewed when zoomed in. + +![](images/7bdb3f12bab41d7960f98ba9f74084120324d9e25c35d9241bfa572b3471854a.jpg) +$\alpha$ + +![](images/8a8e32032fce451a9455edadca0fcd66b83cde3e4cc4e2966cf615d6fb477275.jpg) +$\hat{\alpha}$ + +additional user input as guidance, leading to a classification of matting methods based on whether and what type of user input is used. Trimap-based methods [40] are the most widely used, but obtaining a trimap can be challenging in some scenarios. Therefore, more flexible guidance, such as background [16,30] and user interaction [6,37,38], have been explored. Moreover, some methods [12,38] can also work without user guidance, which is concluded as guidance-free methods. + +The rapid advances in deep learning have significantly boosted the performance of image matting. However, the matting dataset may contain biases, which may further lead to the biased behavior of image matting models. We identify two primary types of bias in existing models: contrast bias and transparency bias. To illustrate these biases, we present some results of a state-of-the-art (SOTA) model, namely MatteFormer [25], in Fig. 1. The contrast bias refers to the correlation between the contrast and the predicted alpha value. Specifically, the model tends to predict lower alpha values in areas with low contrast between foreground and background, as highlighted by the red boxes in Fig. 1a. Regarding the transparency bias, it is observed that the background texture tends to emerge in the predicted alpha matte for mid-transparent foregrounds. As highlighted by the red boxes in Fig. 1b, the predicted alpha for mid-transparent area contains the background texture (i.e., the texture of the rider and building for the left and right examples, respectively). + +Such biased behavior of models can be explained by the biases present in the dataset. For the contrast bias, we calculate the alpha value at different contrast levels within the training dataset as shown in Fig. 2a. We observe that the mean alpha level in low contrast areas (i.e., contrast $< 20$ in the CIEDE2000 color difference formula [31]) is significantly lower than that in other areas. Such dataset bias directly contributes to the observed contrast bias as the model may falsely learn to predict alpha from contrast. Additionally, the number of samples in the low contrast area is significantly fewer than in other areas, making the model more easily led by such dataset bias in these areas. Regarding the + +![](images/e6322f0919a4f7911d908e23e2c46ae4ef906c65778ef225648025d3c8370208.jpg) +(a) The contrast bias in the dataset. + +![](images/58238cbadfb1502d2decc14f05da253c660a2cdd89bb59daabc292f0d8c51c33.jpg) +(b) The transparency bias in the dataset. +Fig. 2: The observe biases in matting datasets. + +transparency bias, we analyze the distribution of transparency levels within the training dataset in Fig. 2b. We observe that most foreground objects are either highly salient or very opaque. Therefore, the alpha matte is expected to be either close to zero or exhibit textures in the image. For the mid-transparency area, whose alpha matte clearly differs from zero, the model is misled to focus on the texture of the image. Thus, the texture of the background in such areas may appear in the alpha matte, explaining the appearance of transparency bias. + +To address these biases, we model the image matting task from the perspective of causal inference, and utilize the power of intervention to reduce the influence of the biases. The causal graph for image matting is depicted in Fig. 3a, where $\mathcal{I},\mathcal{F},\mathcal{B}$ , and $\mathcal{A}$ are variables representing the image, foreground, background, and alpha matte respectively, while $\mathcal{C}$ and $\mathcal{T}$ are confounder variables denoting the contrast and the transparency of foreground. The notation $\mathcal{B} \gets \mathcal{I} \to \mathcal{F}$ suggests that an image can be decomposed into foreground and background, underpinning the basic assumption of image matting. $\mathcal{F} \to \mathcal{A}$ implies that the alpha matte is directly influenced by the foreground object, indicating that changes in the background should not affect the foreground's representation. The confounder variables, representing the common causes affecting multiple variables in the foreground, background, and image, serve as the source of biases. Typically, these can include properties of visual objects. Given that both the foreground and background consist of objects, these properties typically cause the presentation of the foreground, background, and images are presented. + +However, directly measuring all possible confounders may be impractical, as we can not enumerate all properties of visual objects. Nevertheless, based on observed dataset biases, we can identify the most typical confounders. As highlighted in Fig. 3a, we focus on two typical confounders: contrast $\mathcal{C}$ and transparency $\mathcal{T}$ , which underlie the observed contrast bias and transparency bias, respectively. Contrast measures the color difference between foreground and background objects, while transparency refers to the opacity of the foreground. Therefore, we conclude $\mathcal{F} \gets \mathcal{T} \to \mathcal{I}$ and $\mathcal{C} \to \mathcal{X}$ for all $\mathcal{X} \in \{\mathcal{F}, \mathcal{I}, \mathcal{B}\}$ by their definition. The dotted line indicates a potential correlation between $\mathcal{C}$ and $\mathcal{T}$ , suggesting that we cannot definitively assert that $\mathcal{C} \perp \mathcal{T}$ . While existing matting methods generally focus on maximizing the likelihood $P(\mathcal{A}|\mathcal{I})$ , the presence of backdoor paths $\mathcal{B} \gets \mathcal{C} \to \mathcal{F} \to \mathcal{A}$ , $\mathcal{B} \gets \mathcal{I} \gets \mathcal{C} \to \mathcal{F} \to \mathcal{A}$ , and $\mathcal{B} \gets \mathcal{I} \gets \mathcal{T} \to$ + +![](images/319a89db7530423775c1f64f267bd014cdcb0f50c2627b014e1cd3ed8a42c83a.jpg) +(a) The causal graph. + +![](images/5aa8e90be0b0a511691bd86a7b1475e95476982e5fe4be345a976d66d1ebdf43.jpg) +(b) The backdoor adjustment. +Fig. 3: The causal graph and causal intervention for image matting. + +$\mathcal{F} \to \mathcal{A}$ results in $\mathcal{B} \not\sqcup \mathcal{A}|\mathcal{I}$ . This lack of conditional d-separation between $\mathcal{B}$ and $\mathcal{A}$ elucidates the co-occurrence of background textures in specific transparency regions and the contrast-induced alpha bias. + +To remove the effect of these confounders, we propose our model-agnostic cofounder intervened (COIN) matting framework. Specifically, we explore the do-calculus to intervene in the influence of $\mathcal{T}$ and $\mathcal{C}$ via backdoor adjustment, as shown in Fig. 3b. The do-calculus actively assigns values to variables, rather than passively observing them from the data. Therefore, the do-calculus is an intervention on the influence of their direct causes, which is used for decofounding. While the ideal way to calculate the do-operator involves collecting all images with varying contrast and transparency into the training data, this approach is evidently impractical. Furthermore, the high cost of labeling the alpha matte makes it exceedingly difficult to even approximate the aforementioned intervention "physically". To address this issue, we apply backdoor adjustment to calculate $P(\mathcal{A}|do(\mathcal{F},\mathcal{B},\mathcal{I}))$ , as described in Secs. 3.2 to 3.4. By actively assigning values to $\mathcal{F},\mathcal{B}$ , and $\mathcal{I}$ , we block backdoor paths and magnify the influence of cofounders, thus enhancing the robustness of the matting results against the previously mentioned dataset biases. By applying our framework to state-of-the-art (SOTA) image matting methods with different types of user guidance (i.e., MatteFormer [25], DIIM [6], and dugMatting [38]) and two datasets (i.e., Comp1K and Distinct-646), we verified the effectiveness of our framework through extensive experiments. Our contributions are summarized as follows: + +1. We examine the biases in matting methods, analyzing how dataset biases cause them and affect alpha matte prediction. +2. We introduce a novel and model-agnostic confounder-intervened matting framework that employs causal intervention to sever the influence of confounders, thereby mitigating the aforementioned biases. +3. Through comprehensive experiments on various matting methods and multiple datasets, we demonstrate that our framework significantly enhances the performance of SOTA image matting methods. + +# 2 Related Work + +# 2.1 Image Matting + +As described in Sec. 1, image matting presents a highly ill-posed problem. The solutions can be classified based on the user input prior they rely upon. + +Among the prior-based image matting methods, the trimap is the most widely utilized prior [1,3,4,9,14,18,21,23,23,25,33,40-42]. Since [40] introduced the first well-known deep learning-based image matting method, featuring an encoder-decoder architecture, this framework has been extensively adopted and improved by subsequent methods. Researchers have enhanced this architecture from various perspectives, including generative adversarial training [23], matting with indices [21], affinity-based methods [3, 14], context-aware matting [9], the introduction of vision transformers [1, 25, 41], prior alignment [18], expansion to ultra-high resolution via sparsity [33], and data augmentation [4, 42]. + +However, trimap is not always available for real-world applications. Thus, recent researchers seek for more flexible and accessible user inputs [5, 6, 8, 13, 16, 30, 37, 38, 43]. Background matting [30] uses the background image as the prior, which can be more accessible for scenes like virtual meeting and livestreaming. It is further enhanced by recursive excitation [5] and extended into real-time version [16]. The coarse mask is used instead of trimap in [43] which shows better penitential of cooperating with the segmentation models [11]. The user feedbacks, such as clicks [6], scribbles [8], their combination [37], proposal selection [38] and even natural language [13], are also informative for matting. + +# 2.2 Causal Inference + +Causal inference [27, 29] plays a pivotal role in analyzing dataset bias and mitigating counter-causal correlations. It can be categorized into two main types: deconfounding and counterfactual inference. Serving as a fundamental statistical tool, it finds widespread application across various fields, such as image segmentation [44], image classification [35], image and video question answering [10, 15, 24], dialog systems [45], and recommendation systems [36]. Typically, causal-related methods conceptualize the task within a causal graph, where nodes symbolize variables, and edges denote their causal interactions. By examining the causal relations among variables, the causal graph facilitates identifying confounders and backdoor paths, thereby enabling the control of biases. + +# 3 Methodology + +This section outlines our intervention approach to mitigating biases introduced by confounders, as discussed in Sec. 1. In Sec. 3.1, we examine the role of confounder variables within the causal graph (i.e., Fig. 3), illustrating their impact on the performance of mating models through mathematical analysis. In Sec. 3.2, the concept of backdoor adjustment is introduced to counteract biases. In Sec. 3.3, we discuss the construction of space representations for each variable in the causal graph to facilitate backdoor adjustment implementation. In Sec. 3.4, the calculation of expectations, as derived in Sec. 3.2, is explained to execute the causal intervention. Finally, in Sec. 3.5, we integrate the above components into our model-agnostic COIN mating framework. Throughout this section, we denote variables by calligraphic letters (e.g., $\mathcal{A}$ , $\mathcal{F}$ , $\mathcal{I}$ ), feature spaces + +by boldface capital letters (e.g., $\mathbf{A}$ , $\mathbf{F}$ , $\mathbf{I}$ ), and feature vectors by boldface lowercase letters (e.g., $\mathbf{a}$ , $\mathbf{f}$ , $\mathbf{i}$ ). Additional introduction of background and more detailed derivations are provided in the supplementary material. + +# 3.1 Confounders and the Backdoor Paths + +In the task of image matting, confounders are variables that are the direct cause of at least two of the following: the foreground, background, and the image. As illustrated in Fig. 3a, the observational probability can be expressed as + +$$ +P (\mathcal {A} | \mathcal {I}) = \sum_ {\boldsymbol {t}} \sum_ {\boldsymbol {c}} P (\mathcal {A} | \mathcal {I}, \boldsymbol {t}, \boldsymbol {c}) P (\boldsymbol {c}, \boldsymbol {t} | \mathcal {I}), \tag {2} +$$ + +where $\pmb{t}$ and $\pmb{c}$ represent the split of confounder $\mathcal{C}$ and $\mathcal{T}$ , respectively. Given that the observed distribution $P(\pmb{c},\pmb{t}|\mathcal{I})$ significantly deviates from a uniform distribution (as depicted in Fig. 2), a specific partition of $\langle \mathcal{C},\mathcal{T}\rangle$ predominantly influences the observed probability $P(\mathcal{A}|\mathcal{I})$ through a large $P(\pmb{c},\pmb{t}|\mathcal{I})$ . For instance, in the training dataset, regions with high contrast and low transparency are prevalent, leading to their dominance in the observed probability and observed biases. Specifically, in low-contrast regions, the alpha values are typically to be lower, causing the observed contrast bias. Furthermore, the model is trained to discern detailed textures in regions with opaque foregrounds and to assign nearly zero alpha values in highly transparent areas. Since these conditions consist most of the dataset, $P(\pmb{c},\pmb{t}|\mathcal{I})$ with extreme values of $\pmb{t}$ significantly contributes to $P(\mathcal{A}|\mathcal{I})$ . Consequently, the model tends to extract highly textured patterns in medium transparency areas as in opaque foreground areas, where alpha values are apparently not as minimal as in transparent regions. Such background textures in the predicted alpha matte are concluded as transparency bias. + +By adopting a causal perspective on Fig. 3a, we identify the presence of backdoor paths introduced by the confounder variables $\mathcal{C}$ and $\mathcal{T}$ . In particular, the backdoor paths $\mathcal{A} \gets \mathcal{F} \gets \mathcal{C} \to \mathcal{B}$ , $\mathcal{B} \gets \mathcal{I} \gets \mathcal{C} \to \mathcal{F} \to \mathcal{A}$ , and $\mathcal{A} \gets \mathcal{F} \gets \mathcal{T} \to \mathcal{I} \to \mathcal{B}$ result in $\mathcal{A} \not\perp \mathcal{B}|\mathcal{I}$ , elucidating the influence of the background on the predicted alpha matte. More specifically, the former two paths and the latter are directly accountable for the contrast bias and transparency bias, respectively. To intervene the effects of the confounders and sever the backdoor paths they establish, we employ the backdoor adjustment [27,29] as detailed in Sec. 3.2. + +# 3.2 Causal Intervention via Backdoor Adjustment + +The do-calculus [27, 28] signifies active assignment of values to variables, distinguishing it from passive observation within the data, which constitutes intervention on the causal factors of variables. For brevity, we denote $P(do(\mathcal{X} = x))$ as $P(do(\mathcal{X}))$ , implying active assignment of a value to variable $\mathcal{X}$ . As illustrated in Fig. 3b, to sever all incoming causal links of $\mathcal{I}, \mathcal{F}, \mathcal{B}$ (i.e., nullifying the influence of $\mathcal{C}$ and $\mathcal{T}$ ), we focus on computing $P(\mathcal{A}|do(\mathcal{I},\mathcal{F},\mathcal{B}))$ . Given that $\mathcal{T}$ and + +$\mathcal{C}$ block all specified backdoor paths, we can derive: + +$$ +\begin{array}{l} P (\mathcal {A} | d o (\mathcal {I}, \mathcal {F}, \mathcal {B})) = \sum_ {\boldsymbol {t}} \sum_ {\boldsymbol {c}} P (\mathcal {A} | d o (\mathcal {I}, \mathcal {F}, \mathcal {B}), \boldsymbol {t}, \boldsymbol {c}) P (\boldsymbol {c}, \boldsymbol {t} | d o (\mathcal {I}, \mathcal {F}, \mathcal {B})) \\ = \sum_ {t} ^ {t} \sum_ {c} ^ {c} P (\mathcal {A} | i, \boldsymbol {f}, \boldsymbol {b}, t, c) P (\boldsymbol {c}, t). \tag {3} \\ \end{array} +$$ + +Moreover, based on Fig. 3b, we observe $\mathcal{A} \perp \mathbf{b}|(\mathbf{f},\mathbf{i},\mathbf{c},\mathbf{t})$ and $\mathcal{A} \perp \mathbf{i}|(\mathbf{f},\mathbf{c},\mathbf{t})$ . Thus, the equation simplifies to: + +$$ +P (\mathcal {A} | d o (\mathcal {I}, \mathcal {F}, \mathcal {B})) = \sum_ {\boldsymbol {t}} \sum_ {\boldsymbol {c}} P (\mathcal {A} | \boldsymbol {f}, \boldsymbol {t}, \boldsymbol {c}) P (\boldsymbol {t}, \boldsymbol {c}) = \mathbb {E} _ {\boldsymbol {t}, \boldsymbol {c}} [ P (\mathcal {A} | \boldsymbol {f}, \boldsymbol {t}, \boldsymbol {c}) ]. \tag {4} +$$ + +Therefore, to predict the alpha matte with confounders intervened, we need to model the relevant spaces and calculate the expectations in Eq. (4), which are further discussed in Sec. 3.3 and Sec. 3.4, respectively. + +However, to calculate the expectation in Eq. (4), obtaining the foreground representation is also necessary. Although the only input is the image (and the possible user guidance), directly inferring the foreground representation from the image representation is biased, as the confounders similarly facilitate backdoor paths to $\mathcal{F}$ as to $\mathcal{A}$ . Specifically, the backdoor paths $\mathcal{F} \gets \mathcal{C} \to \mathcal{B}$ , $\mathcal{B} \gets \mathcal{I} \gets \mathcal{C} \to \mathcal{F}$ , and $\mathcal{F} \gets \mathcal{T} \to \mathcal{I} \to \mathcal{B}$ lead to $\mathcal{F} \not\perp \mathcal{B}|\mathcal{I}$ , indicating that the background's transparency and contrast similarly affect the foreground as influencing the prediction of the alpha matte. To address this issue, we also conduct an intervention for predicting the foreground, yielding: + +$$ +\begin{array}{l} P (\mathcal {F} | d o (\mathcal {I}, \mathcal {B})) = \sum_ {\boldsymbol {t}} \sum_ {\boldsymbol {c}} P (\mathcal {F} | \boldsymbol {i}, \boldsymbol {b}, \boldsymbol {t}, \boldsymbol {c}) P (\boldsymbol {c}, \boldsymbol {t}) \\ = \sum_ {t} ^ {t} \sum_ {c} ^ {c} P (\mathcal {F} | \boldsymbol {i}, \boldsymbol {t}, \boldsymbol {c}) P (\boldsymbol {c}, \boldsymbol {t}) = \mathbb {E} _ {\boldsymbol {c}} \mathbb {E} _ {\boldsymbol {t}} [ P (\mathcal {F} | \boldsymbol {i}, \boldsymbol {t}, \boldsymbol {c}) ], \tag {5} \\ \end{array} +$$ + +where the multi-scale image representation $\pmb{i} = \langle \pmb{i}_1, \pmb{i}_2, \dots, \pmb{i}_n \rangle$ is parameterized by the matting backbone, aligned with the scales of $\pmb{C}$ and $\pmb{T}$ as described in Sec. 3.3. The expectations are calculated as detailed in Sec. 3.4. + +# 3.3 Space Representations + +Transparency Space $\pmb{T}$ . Given the importance of multi-scale information in image matting, we divide the transparency space into several sub-spaces across different scales. Specifically, the transparency space $\pmb{T} = \langle T_1, T_2, \dots, T_n \rangle$ , with $\pmb{T}_i$ representing the transparency at the $i$ -th scale, and $n$ indicating the total number of sub-spaces. The $i$ -th scale denotes the transparency of the image patches sized $2^i \times 2^i$ . Considering the average alpha value as an approximation for the "transparency level" of an area, we define $\pmb{T}_i$ as the average alpha value of pixels within the $i$ -th scale patches. This formulation allows us to discretize the transparency sub-spaces further by categorizing the transparency level into $K_t$ distinct classes, given its bounded nature within [0, 1]. The construction of the feature space entails the following steps: + +1. Utilize a pretrained image feature extractor to derive the patch features at the $i$ -th scale, for each $i \in \{1, \dots, n\}$ . +2. Determine the transparency level category $k \in \{1, \dots, K_t\}$ for every patch. +3. Compute the average patch representation for all patches within the $k$ -th category as the representation of this transparency category. + +Hence, we formulate the feature space $\pmb{T}_i\in \mathbb{R}^{d_i\times K_t}$ , where $d_{i}$ signifies the dimensionality of the feature extracted at the $i$ -th scale. + +Contrast Space $C$ . Similarly to the transparency space, the contrast space is partitioned into multiple sub-spaces across various scales. We define $C = \langle C_1, C_2, \ldots, C_n \rangle$ , with each $C_i$ representing contrast at the $i$ -th scale, and $n$ marking the total number of sub-spaces. Each $i$ -th scale denotes the contrast of the foreground and background within $2^i \times 2^i$ image patches. + +Quantifying contrast, however, presents more complexity than measuring transparency. Initially, we estimate the color of a patch using the average color vector of all pixels within it in the color space. Subsequently, the contrast between foreground and background patches is determined by the color distance between their average color vectors. Given the multifaceted nature of color perception, we employ the CIEDE2000 color difference formula [31] for calculating color distance, as it closely aligns with human visual perception of color differences. This formula, incorporating lightness, chroma, and hue, yields a distance within the range of [0, 100]. Following this approach, we categorize the contrast level into $K_{c}$ discrete classes and construct the feature space $C_i \in \mathbb{R}^{d_i \times K_c}$ , mirroring the methodology used for the transparency space. The primary adjustment involves transitioning from transparency to contrast categories for the classification. + +Joint Confounder Space $\mathbf{Z}$ . Given that $\mathcal{T}$ and $\mathcal{C}$ may not be independent, we cannot decompose the joint expectation in Eq. (4) into separate expectations for each. Therefore, it becomes necessary to model the joint space of $\mathbf{T}$ and $\mathbf{C}$ . An effective approach is to define the joint space as the Cartesian product of the sub-spaces. Consequently, the joint confounder space $\mathbf{Z} = \langle \mathbf{Z}_1, \mathbf{Z}_2, \dots, \mathbf{Z}_n \rangle = \langle \mathbf{T}_1 \times \mathbf{C}_1, \mathbf{T}_2 \times \mathbf{C}_2, \dots, \mathbf{T}_n \times \mathbf{C}_n \rangle$ , where each $\mathbf{T}_i \times \mathbf{C}_i \in \mathbb{R}^{d_i \times (K_t \cdot K_c)}$ represents a multidimensional space formed by the Cartesian product of transparency and contrast spaces at the $i$ -th scale. Thus, Eqs. (4) and (5) are simplified as following: + +$$ +\begin{array}{l} P(\mathcal{A}|do(\mathcal{I},\mathcal{F},\mathcal{B})) = \sum_{\boldsymbol {z}}P(\mathcal{A}|\boldsymbol {f},\boldsymbol {z})P(\boldsymbol {z}) = \mathbb{E}_{\boldsymbol{z}}[P(\mathcal{A}|\boldsymbol {f},\boldsymbol {z})], \\ P (\mathcal {F} | d o (\mathcal {I}, \mathcal {B})) = \sum_ {\boldsymbol {z}} P (\mathcal {F} | \boldsymbol {i}, \boldsymbol {z}) P (\boldsymbol {z}) = \mathbb {E} _ {\boldsymbol {z}} [ P (\mathcal {F} | \boldsymbol {i}, \boldsymbol {z}) ], \\ \end{array} +$$ + +where $z = \langle \pmb {c},\pmb {t}\rangle$ represents the split of joint observed confounders $\mathcal{Z} = \langle \mathcal{T},\mathcal{C}\rangle$ + +# 3.4 Expectations + +The expectations are calculated over the joint confounder space $\mathbf{Z}$ . Utilizing the Normalized Weighted Geometric Mean (NWGM) approach [32,39], we reformulate the expectation as follows: + +$$ +P (\mathcal {A} | d o (\mathcal {I}, \mathcal {F}, \mathcal {B})) = \mathbb {E} _ {\boldsymbol {z}} P (\mathcal {A} | \boldsymbol {f}, \boldsymbol {z}) = \mathbb {E} _ {\boldsymbol {z}} \sigma (g (\boldsymbol {f}, \boldsymbol {z})) \approx \sigma (g (\boldsymbol {f}, \mathbb {E} _ {\boldsymbol {z}} [ \boldsymbol {z} ])), \tag {7} +$$ + +![](images/0d5db88a6f0a52bf1cca642701e284c995386ef7b531195b5404e588304da10d.jpg) +Fig. 4: The COIN framework. $\pmb{i}$ and $\pmb{f}$ denote the image feature and foreground feature. $T, C$ and $Z$ are transparency space, contrast space, and their joint confounder space respectively, which are constructed in Sec. 3.3. Spaces corresponding to the confounder variable are marked gray. EXP represents the expectations described in Sec. 3.4. $\sigma(\cdot)$ and $g(\cdot)$ are ignored for simplicity. $\odot$ represents the dot production. + +where $\sigma(\cdot)$ denotes a sigmoid function and $g(\cdot)$ represents a linear transformation for each sub-space. Similarly, for the foreground parameterization: + +$$ +P (\mathcal {F} | d o (\mathcal {I}, \mathcal {B})) = \mathbb {E} _ {z} P (\mathcal {F} | \boldsymbol {i}, z) = \mathbb {E} _ {z} g (\boldsymbol {i}, z) = g (\boldsymbol {i}, \mathbb {E} _ {z} [ z ]). \tag {8} +$$ + +This approach shifts the expectations to the feature level. To prevent model collapse, we use $\mathbb{E}_{\boldsymbol{z}|\mathcal{F}}$ and $\mathbb{E}_{\boldsymbol{z}|\mathcal{I}}$ as approximations for $\mathbb{E}_z$ respectively. To emulate these feature-level expectations, we leverage a sub-space-wise attention mechanism. Take $\mathbb{E}_{\boldsymbol{z}|\mathcal{I}}[\boldsymbol{z}]$ as an example. As spaces $Z_{i}$ are heterogeneous with the feature spaces of $\pmb{i}_j$ when $i\neq j$ , we apply the attention mechanism for each scale respectively. Therefore, the expectation is + +$$ +\mathbb {E} _ {\boldsymbol {z} \mid \mathcal {I}} [ \boldsymbol {z} ] = \left\langle \mathbb {E} _ {\boldsymbol {z} _ {1} \mid \mathcal {I} _ {1}} [ \boldsymbol {z} _ {1} ], \mathbb {E} _ {\boldsymbol {z} _ {2} \mid \mathcal {I} _ {2}} [ \boldsymbol {z} _ {2} ], \dots , \mathbb {E} _ {\boldsymbol {z} _ {n} \mid \mathcal {I} _ {n}} [ \boldsymbol {z} _ {n} ] \right\rangle . \tag {9} +$$ + +where $n$ is the number of sub-spaces. For attention on image patches at the $i$ -th scale, the input is the contrast sub-space $Z_{i} \in \mathbb{R}^{d_{i} \times (K_{t} \cdot K_{c})}$ and the image patch representation $\hat{\pmb{i}}_i \in \mathbb{R}^{d_i}$ parameterized by the matting backbone, where $d_{i}$ is the dimension of the latent representations. During such attention, the attention score emulates the distribution of $P(\pmb{z}|\mathcal{I})$ , thus the weighted sum on values (i.e., $z$ ) emulates the expectation $\mathbb{E}_{\pmb{z}|\mathcal{I}}[\pmb{z}]$ . The mathematical formulation is given by: + +$$ +\pmb {h e a d} _ {i} ^ {k} = \mathrm {s o f t m a x} (\hat {\pmb {i}} _ {i} ^ {k} \pmb {W} _ {i, 1} ^ {k} (\pmb {Z} _ {i} \pmb {W} _ {i, 2} ^ {k}) ^ {T} / \sqrt {d}) (\pmb {Z} _ {i} \pmb {W} _ {i, 3} ^ {k}) +$$ + +$$ +\bar {z} _ {\mathcal {I} _ {i}} = \mathbb {E} _ {\boldsymbol {z} _ {i} \mid \mathcal {I} _ {i}} [ \boldsymbol {z} _ {i} ] = \operatorname {C o n c a t} \left(\boldsymbol {h e a d} _ {i} ^ {1}, \boldsymbol {h e a d} _ {i} ^ {2}, \dots , \boldsymbol {h e a d} _ {i} ^ {N _ {h}}\right) \boldsymbol {W} _ {i, o} \tag {10} +$$ + +$$ +\bar {z} _ {\mathcal {I}} = \operatorname {E X P} (\hat {\boldsymbol {i}}, \boldsymbol {Z}) = \mathbb {E} _ {\boldsymbol {z} | \mathcal {I}} [ \boldsymbol {z} ] = \left\langle \bar {z} _ {\mathcal {I} _ {1}}, \bar {z} _ {\mathcal {I} _ {2}}, \dots , \bar {z} _ {\mathcal {I} _ {n}} \right\rangle , +$$ + +where $\pmb{W}_{i,1}^{k}, \pmb{W}_{i,2}^{k}, \pmb{W}_{i,3}^{k}, \pmb{W}_{i,o}$ are trainable parameters, $N_{h}$ represents the number of heads, and $\mathrm{Concat}(\cdot)$ is the concatenation operation. This process effectively captures the expectations across different confounder sub-spaces. + +# 3.5 COIN Matting Framework + +The overall structure of our COIN framework is shown in Fig. 4. The confounder space $\pmb{T}$ is similarly constructed as $\mathbf{C}$ , as described in Sec. 3.3. The mat + +ting backbone is employed to parameterize the image representation $\hat{\pmb{i}}$ . Rather than directly predicting $P(\mathcal{A}|\mathcal{I})$ as the matting backbone does, we engage in confounder intervention through backdoor adjustment to address the issues of contrast bias and transparency bias. The backdoor adjustment is implemented by the feature-level expectations. Specifically, we first process the joint space $\pmb{Z}$ alongside the image representation $\pmb{i}$ to compute $\mathbb{E}_{z|\mathcal{I}}[z]$ via attention. This expectation is subsequently concatenated with the image representations to yield the foreground representation. Similarly, the foreground representation is used to calculate $\mathbb{E}_{z|\mathcal{F}}[z]$ with joint confounder space $\pmb{Z}$ , and the expectations are further concatenated with the foreground representation to produce the alpha matte at various scales. Thus, we can express the overall framework as follows: + +$$ +\hat {\boldsymbol {i}} = \text {M a t t i n g B a c k b o n e} (\boldsymbol {I}); \quad \bar {z} _ {\mathcal {I}} = \operatorname {E X P} (\hat {\boldsymbol {i}}, \boldsymbol {Z}); \quad \hat {\boldsymbol {f}} = g _ {f} (\hat {\boldsymbol {i}}, \bar {z} _ {\mathcal {I}}); \tag {11} +$$ + +$$ +\bar {z} _ {\mathcal {F}} = \operatorname {E X P} (\hat {\boldsymbol {f}}, \boldsymbol {Z}); \quad \hat {\alpha} = P (\mathcal {A} | d o (\mathcal {I}, \mathcal {F}, \mathcal {B})) = \sigma (g _ {\alpha} (\hat {\boldsymbol {f}}, \bar {z} _ {\mathcal {F}})), +$$ + +where $g_{f}(\cdot)$ and $g_{\alpha}(\cdot)$ represent per-sub-space linear transformations, and $\mathrm{EXP}(\cdot, \cdot)$ denotes the feature-level expectations described in Sec. 3.4. + +During the training of this framework, we supervise the alpha matte at each scale, similar to the approach taken by the backbone. Furthermore, we supervise the attention score, which approximates the probability distribution of the confounders when calculating expectations, with the ground truth categories of contrast and transparency. It is important to note that a closer category index of contrast and transparency signifies a more similar meaning, as these are discretized from the continuous values of color distance and transparency. Therefore, we employ Gaussian label smoothing prior to the cross-entropy for supervising the attention score. Additional details regarding the framework can be found in the supplementary material. + +# 4 Experiments + +In this section, we evaluate our COIN framework on extensive image matting baselines across various datasets. Specifically, we integrate our framework with SOTA image matting methods across all three types of matting methods: trimap-based methods (i.e., MatteFormer [25]), flexible guidance-based methods (i.e., DIIM [6] and dugMatting [38]) and guidance-free methods (i.e., dugMatting without human interaction [38]). We evaluate the improvement of our framework on Composition-1k [40], and Distinctions-646 [43] datasets. Additionally, we elaborate on the benefits of intervening on each confounder and the computational cost of implementing our framework in Sec. 4.3. The qualitative results are in Sec. 4.4. + +# 4.1 Experiment Settings + +Datasets. We conduct the experiments on Composition-1k [40] and Distinctions-646 [43] datasets following [25]. Composition-1k dataset consists of 431 and 50 foreground objects in the training set and test set respectively. These foregrounds + +are composited with MS COCO [17] images for training and with VOC 2012 [7] images for testing. For testing, each foreground object is combined with 20 different images, generating a total of 1000 test images. The Distinctions-646 dataset contains 646 unique foreground objects, which are divided into 596 for training and 50 for testing. The composition process with background images follows the same procedure as in the Composition-1k dataset. + +Metrics. For measuring the overall performance of our framework, we adopt the widely used image matting metrics [12,25,26,43]: sum of absolute difference (SAD), mean square error (MSE), and the gradient (Grad) and connectivity (Conn). Additionally, to assess the improvement in reducing transparency and contrast biases, we measure SAD conditioned on each contrast and transparency category. The number of parameters and FLOPs are also reported to evaluate the computational complexity of implementing our framework. + +Implementation Details. We set the number of sub-spaces as $n = 5$ for $I, F, A, T$ , and $C$ . For transparency and contrast levels, we set both $K_{t}$ and $K_{c}$ to 25. The contrast and transparency space is initialized by the average multiscale feature extracted by pretrained Swin-B [19, 20] as described in Sec. 3.3. We employ a $\mathcal{N}(0, 1.2)$ Gaussian distribution to smooth the label of attention score. More details can be found in the supplementary material. + +# 4.2 Main Results + +The improvement of our COIN framework is demonstrated in Tab. 1. It is evident that by intervening on the confounder variables, our COIN framework significantly enhances the performance across all matting baselines with varying guidance mechanisms on both the Composition-1k and Distinctions-646 datasets. Our framework substantially improves results for both guidance-free methods (e.g., dugMatting (0-sel)) and flexible-guidance based methods (e.g., dugMatting (1-sel) and DIIM). Additionally, even for the SOTA trimap-based baseline (e.g., MatteFormer), our COIN framework achieves significant performance gains, decreasing SAD by 1.9 and 2.6, and MSE by 0.7 and 1.5 on the two datasets respectively. Notably, our framework tends to exhibit greater improvements on the Distinctions-646 dataset compared to the Composition-1k dataset. This discrepancy is likely attributable to the more pronounced biases present within the Distinctions-646 dataset. Furthermore, enhancements are generally more substantial for flexible guidance-based or guidance-free matting methods than for trimap-based methods, since the former methods are typically more susceptible to biases in the absence of the detailed information provided by a trimap. + +To demonstrate that the enhancements are truly attributable to the reduction of contrast bias and transparency bias, we further evaluate the SAD improvements across each contrast and transparency category, as shown in Fig. 5. We selected MatteFormer as the baseline because it presents the most substantial challenge and exhibits the highest performance among the baselines. Regarding transparency, it is observable that our framework enhances the performance of baseline matting methods across all transparency levels. Furthermore, we note + +Table 1: The comparison of applying our COIN framework on Composition-1k and Distinctions-646 dataset. The improvements are highlighted as superscripts. + +
MethodComposition-1KDistinctions-646
SAD ↓MSE ↓Grad ↓Conn ↓SAD ↓MSE ↓Grad ↓Conn ↓
DIIM37.510.318.534.536.211.819.635.6
dugMatting (0-sel)*34.15.715.631.233.28.818.233.2
dugMatting (1-sel)*25.84.39.722.324.17.112.424.0
MatteFormer23.84.08.718.921.96.611.220.5
COIN(DIIM)33.8-3.78.1-2.232.0-2.59.6-2.2
COIN(dugMatting (0-sel)*)30.8-3.34.5-1.228.0-3.27.4-1.4
COIN(dugMatting (1-sel)*)23.7-2.13.5-0.819.1-3.26.3-0.8
COIN(MatteFormer)21.9-1.93.3-0.716.8-2.15.1-1.5
+ +* $k$ -sel denotes $k$ times human interaction during matting. + +![](images/1d89b42cc1878961ec7078efaaed3d0f0cc1a7fa41d65aeb5fcf7a7c1464aede.jpg) +(a) The SAD conditioned on transparency. + +![](images/a66a2a6a20115ec1fd03fa9f05831862e1142af7efb7ce26871be4b32e8b06f4.jpg) +(b) The SAD conditioned on contrast. +Fig. 5: The improvement of applying our framework to MatteFormer on Composition-1K dataset conditioned on each contrast and transparency category. + +that the improvement is particularly pronounced in regions where the transparency level is within the range of [50, 85]. The training data within this range, as discussed in Sec. 1, is notably scarce but exhibits a unique alpha pattern which differs from the pattern in highly opaque or transparent foregrounds. The observed probability distribution, dominated by data from regions with high opacity or transparency, causes the transparency bias we identified. The significant enhancements in these regions indicate that our framework is reducing the transparency bias through confounder intervention. In terms of contrast, our framework not only significantly reduces the SAD error across all contrast levels compared to the baseline method but also shows that the most notable improvement occurs in areas of relatively low contrast, particularly where the contrast level ranges from [10, 25]. In these regions, the detail patterns increase rapidly as average contrast value and mean alpha value raises (as shown in Fig. 2a). However, owing to the insufficient data relative to other regions, these areas are most susceptible to contrast bias. Thus, the most notable improvements in low-contrast areas suggest that our framework enhances performance by alleviating the contrast bias through intervention on the contrast confounder. + +Table 2: The comparison of applying intervention on different confounder variables on the Distinctions-646 dataset using MatteFormer as the baseline. The improvements are highlighted as superscripts. The GFLOPs is calculated on $1024 \times 1024$ image. + +
MethodErrorsComplexity
SAD ↓MSE ↓Grad ↓Conn ↓ParametersFLOPs
MatteFormer21.9-1.56.6-0.811.2-0.920.544.8M459G
MatteFormer + T intervene20.45.810.319.1-1.447.7M
MatteFormer + C intervene20.2-1.75.7-0.910.0-1.218.8-1.747.7M
COIN(MatteFormer)19.3-2.65.1-1.59.4-1.817.8-2.749.5M
+ +# 4.3 Ablation Study + +Confounder Spaces. We discuss the improvement from intervening in each confounder space individually rather than their joint confounder in Tab. 2. By comparing line 3 and line 2 with line 1, it is observed that intervening on either $\mathcal{C}$ or $\mathcal{T}$ can improve the performance of the baseline methods. However, because such interventions do not block the backdoor path opened by $\mathcal{T}$ and $\mathcal{C}$ respectively, the condition $\mathcal{A} \nsubseteq \mathcal{B}|\mathcal{T}$ still exists, which leads to transparency bias and contrast bias respectively, resulting in a limited performance of the intervention framework. Moreover, it was noted that the performance of intervening on $\mathcal{C}$ is slightly better than that of intervening on $\mathcal{T}$ . This may be due to the fact that the contrast bias is more pronounced than the transparency bias in the dataset. By jointly intervening both $\mathcal{C}$ and $\mathcal{T}$ , we may block all aforementioned backdoor paths, thus further performance improvement in the COIN framework. + +Complexity Analysis. As we introduce a framework atop existing matting methods, we further analyze its space and time complexity in terms of both parameters and FLOPs to delineate the cost of integrating our framework. As indicated in Tab. 2, the additional cost imposed by our framework is minimal when compared to the baseline methods. The incremental parameters (about $10.5\%$ ) stem exclusively from the feature space and its corresponding attention mechanism, which are negligible relative to the baseline model. The only extra FLOPs (about $10.6\%$ ) originate from the computation of expectations within feature spaces, which are characterized as $n$ attentions across $d_{i} \times (K_{t} \cdot K_{c})$ matrices and $d_{i} \times L$ image features, where $L$ denotes the number of image patches at the $i$ -th scale. Such attention is also lightweight compared to the matting baseline. Overall, our framework improves the performance of the SOTA image matting baseline between $11.6\%$ and $22.7\%$ in different metrics with about $10\%$ computation cost, which is remarkably efficient for integrating with the baselines. + +# 4.4 Qualitative Results + +In this section, we present qualitative comparisons of our COIN framework. In Fig. 6, we compare the alpha matting results from baseline methods with those obtained using our COIN framework. Rows 1 showcase examples illustrating the reduction of contrast bias. It is evident that the low-contrast areas + +![](images/a7e5c4ac4cb85e594eff51389db6f21fbe9b57776fe97819aeb86dca8984bca8.jpg) +Image + +![](images/3289c8ca65ad8c01aa4a9fed4969cb5b1289153a7c276b5dd66c756b3a28f755.jpg) +Trimap + +![](images/df03e1bbec89ea81c0ab77ac5d28a4abd6c0ae2e73d231b1cdf61d20b426231c.jpg) +Alpha matte + +![](images/5f15eef4afc378abdfe2720863b31a6e8ba706c6314a3e984e3624c906cbe1ba.jpg) +MatteFormer +Fig. 6: The visualized comparison of applying our COIN framework on Matte-Former [25]. Images are best viewed when zoomed in. + +![](images/5f5e8270e9aa68e41c5d5ac858d2b2466c4eeb52ae62db1e31cf73303c1c4545.jpg) +COIN + +highlighted in the red boxes yield alpha mattes with lower values when processed by MatteFormer. However, the integration of our COIN framework significantly diminishes this contrast bias, resulting in more precise alpha mattes. Rows 2 and 3 demonstrate the reduction of transparency bias. The highlighted mid-transparent regions in the red boxes contain background textures (i.e., the texture of text and the car cage respectively) in the predicted alpha mattes. Nevertheless, with the application of our COIN framework, the alpha mattes become markedly more accurate, effectively eliminating such textures from the mid-transparent areas. Such qualitative results prove that our COIN framework can effectively reduce both contrast bias and transparency bias, therefore improving the performance of the existing image matting methods. + +# 5 Conclusion + +In this work, we have concentrated on the contrast bias and transparency bias in existing matting methods and demonstrated how dataset biases lead to biased behaviors of models. We analyzed these biases from a causal perspective and designed the confounder intervention framework to alleviate them. Extensive experiments have demonstrated that our framework markedly alleviates these biases and enhances SOTA matting methods with minor computational cost. + +# Acknowledgements + +The work was supported by the National Natural Science Foundation of China (Grant No. 62076162), the Shanghai Municipal Science and Technology Major Project, China (Grant No. 2021SHZDZX0102) and the Postdoctoral Fellowship Program of CPSF (Grant No. GZC20241225). + +# References + +1. Cai, H., Xue, F., Xu, L., Guo, L.: Transmatting: Enhancing transparent objects matting with transformers. In: ECCV (2022) +2. Chen, J., Li, X., Luo, L., Ma, J.: Multi-focus image fusion based on multi-scale gradients and image matting. IEEE Trans. Multim. 24, 655-667 (2022) +3. Dai, Y., Lu, H., Shen, C.: Learning affinity-aware upsampling for deep image matting. In: CVPR (2021) +4. Dai, Y., Price, B., Zhang, H., Shen, C.: Boosting robustness of image matting with context assembling and strong data augmentation. In: CVPR (2022) +5. Deng, J., Xu, Y., Zhou, Z., He, S.: Background matting via recursive excitation. In: ICME (2022) +6. Ding, H., Zhang, H., Liu, C., Jiang, X.: Deep interactive image matting with feature propagation. IEEE Trans. Image Process. 31, 2421-2432 (2022) +7. Everingham, M., Gool, L.V., Williams, C.K.I., Winn, J.M., Zisserman, A.: The pascal visual object classes (VOC) challenge. Int. J. Comput. Vis. (2010) +8. Fang, X., Zhang, S., Chen, T., Wu, X., Shamir, A., Hu, S.: User-guided deep human image matting using arbitrary trimaps. IEEE Trans. Image Process. 31, 2040-2052 (2022) +9. Hou, Q., Liu, F.: Context-aware image matting for simultaneous foreground and alpha estimation. In: ICCV (2019) +0. Li, J., Niu, L., Zhang, L.: Knowledge proxy intervention for deconfounded video question answering. In: ICCV (2023) +1. Li, J., Wang, W., Chen, J., Niu, L., Si, J., Qian, C., Zhang, L.: Video semantic segmentation via sparse temporal transformer. In: ACM MM (2021) +2. Li, J., Zhang, J., Tao, D.: Deep automatic natural image matting. In: IJCAI (2021) +3. Li, J., Zhang, J., Tao, D.: Referring image matting. In: Proceedings of the IEEE Computer Vision and Pattern Recognition (2023) +4. Li, Y., Lu, H.: Natural image matting via guided contextual attention. In: AAAI (2020) +5. Li, Y., Wang, X., Xiao, J., Chua, T.: Equivariant and invariant grounding for video question answering. In: ACM MM (2022) +6. Lin, S., Ryabtsev, A., Sengupta, S., Curless, B.L., Seitz, S.M., Kemelmacher-Shlizerman, I.: Real-time high-resolution background matting. In: CVPR (2021) +7. Lin, T., Maire, M., Belongie, S.J., Hays, J., Perona, P., Ramanan, D., Dóllar, P., Zitnick, C.L.: Microsoft COCO: common objects in context. In: ECCV (2014) +8. Liu, Y., Xie, J., Qiao, Y., Tang, Y., Yang, X.: Prior-induced information alignment for image matting. IEEE Trans. Multim. 24, 2727-2738 (2022) +9. Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., Wei, F., Guo, B.: Swin transformer v2: Scaling up capacity and resolution. In: CVPR (2022) +20. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021) +21. Lu, H., Dai, Y., Shen, C., Xu, S.: Indices matter: Learning to index for deep image matting. In: ICCV (2019) +22. Lu, L., Li, J., Cao, J., Niu, L., Zhang, L.: Painterly image harmonization using diffusion model. In: ACM MM (2023) +23. Lutz, S., Amplanitis, K., Smolic, A.: Alphagan: Generative adversarial networks for natural image matting. In: BMVC (2018) + +24. Niu, Y., Tang, K., Zhang, H., Lu, Z., Hua, X., Wen, J.: Counterfactual VQA: A cause-effect look at language bias. In: CVPR (2021) +25. Park, G., Son, S., Yoo, J., Kim, S., Kwak, N.: Matteformer: Transformer-based image matting via prior-tokens. In: CVPR (2022) +26. Park, K., Woo, S., Oh, S.W., Kweon, I.S., Lee, J.Y.: Mask-guided matting in the wild. In: CVPR (2023) +27. Pearl, J., Glymour, M., Jewell, N.P.: Causal inference in statistics: A primer (2016) +28. Pearl, J., Mackenzie, D.: The Book of Why: The New Science of Cause and Effect. Basic Books, Inc. (2018) +29. Rubin, D.B.: Essential concepts of causal inference: a remarkable history and an intriguing future. Biostatistics & Epidemiology (2019) +30. Sengupta, S., Jayaram, V., Curless, B., Seitz, S.M., Kemelmacher-Shlizerman, I.: Background matting: The world is your green screen. In: CVPR (2020) +31. Sharma, G., Wu, W., Dalal, E.N.: The ciede2000 color-difference formula: Implementation notes, supplementary test data, and mathematical observations. Color Research and Application 30, 21-30 (2005), https://api-semanticscholar.org/ CorpusID:29119937 +32. Srivastava, N., Hinton, G.E., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. J. Mach. Learn. Res. (2014) +33. Sun, Y., Tang, C., Tai, Y.: Ultrahigh resolution image/video matting with spatiotemporal sparsity. In: CVPR (2023) +34. Tan, L., Li, J., Niu, L., Zhang, L.: Deep image harmonization in dual color spaces. In: ACM MM (2023) +35. Wang, T., Zhou, C., Sun, Q., Zhang, H.: Causal attention for unbiased visual recognition. In: ICCV (2021) +36. Wang, W., Feng, F., He, X., Zhang, H., seng Chua, T.: Clicks can be cheating: Counterfactual recommendation for mitigating clickbait issue (2020) +37. Wei, T., Chen, D., Zhou, W., Liao, J., Zhao, H., Zhang, W., Hua, G., Yu, N.: Deep image matting with sparse user interactions. TPAMI (2024) +38. Wu, J., Zhang, C., Li, Z., Fu, H., Peng, X., Zhou, J.T.: dugMatting: Decomposed-uncertainty-guided matting. In: ICML (2023) +39. Xu, K., Ba, J., Kiros, R., Cho, K., Courville, A.C., Salakhutdinov, R., Zemel, R.S., Bengio, Y.: Show, attend and tell: Neural image caption generation with visual attention. In: ICML (2015) +40. Xu, N., Price, B.L., Cohen, S., Huang, T.S.: Deep image matting. In: CVPR (2017) +41. Yao, J., Wang, X., Yang, S., Wang, B.: Vitmatte: Boosting image matting with pre-trained plain vision transformers. Inf. Fusion 103, 102091 (2024) +42. Ye, Z., Dai, Y., Hong, C., Cao, Z., Lu, H.: Infusing definiteness into randomness: Rethinking composition styles for deep image matting. In: AAAI (2023) +43. Yu, Q., Zhang, J., Zhang, H., Wang, Y., Lin, Z., Xu, N., Bai, Y., Yuille, A.L.: Mask guided matting via progressive refinement network. In: CVPR (2021) +44. Zhang, D., Zhang, H., Tang, J., Hua, X., Sun, Q.: Causal intervention for weakly-supervised semantic segmentation. In: NeurIPS (2020) +45. Zhu, Q., Zhang, W.N., Liu, T., Wang, W.Y.: Counterfactual off-policy training for neural dialogue generation. In: EMNLP (2020) \ No newline at end of file diff --git a/2024/COIN-Matting_ Confounder Intervention for Image Matting/images.zip b/2024/COIN-Matting_ Confounder Intervention for Image Matting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5b218119e86d982ad2ff451df35dacb11f9aea57 --- /dev/null +++ b/2024/COIN-Matting_ Confounder Intervention for Image Matting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ea6e58c8bf6069182bc276559bd9a14f411424852c0cb96a796a9107f6ba417 +size 426267 diff --git a/2024/COIN-Matting_ Confounder Intervention for Image Matting/layout.json b/2024/COIN-Matting_ Confounder Intervention for Image Matting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..95cc316f8be49c3cb387bef169e603bd0e21b9e6 --- /dev/null +++ b/2024/COIN-Matting_ Confounder Intervention for Image Matting/layout.json @@ -0,0 +1,12125 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 152, + 111, + 462, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 111, + 462, + 148 + ], + "spans": [ + { + "bbox": [ + 152, + 111, + 462, + 148 + ], + "type": "text", + "content": "COIN-Matting: Confounder Intervention for Image Matting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 144, + 167, + 470, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 167, + 470, + 194 + ], + "spans": [ + { + "bbox": [ + 144, + 167, + 470, + 194 + ], + "type": "text", + "content": "Zhaohe Liao" + }, + { + "bbox": [ + 144, + 167, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 167, + 470, + 194 + ], + "type": "text", + "content": ", Jiangtong Li" + }, + { + "bbox": [ + 144, + 167, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 144, + 167, + 470, + 194 + ], + "type": "text", + "content": ", Jun Lan" + }, + { + "bbox": [ + 144, + 167, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 144, + 167, + 470, + 194 + ], + "type": "text", + "content": ", Huijia Zhu" + }, + { + "bbox": [ + 144, + 167, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 144, + 167, + 470, + 194 + ], + "type": "text", + "content": ", Weiqiang Wang" + }, + { + "bbox": [ + 144, + 167, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 144, + 167, + 470, + 194 + ], + "type": "text", + "content": ", Li Niu" + }, + { + "bbox": [ + 144, + 167, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{*1,4}" + }, + { + "bbox": [ + 144, + 167, + 470, + 194 + ], + "type": "text", + "content": ", and Liqing Zhang" + }, + { + "bbox": [ + 144, + 167, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{*1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 190, + 201, + 425, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 201, + 425, + 280 + ], + "spans": [ + { + "bbox": [ + 190, + 201, + 425, + 280 + ], + "type": "text", + "content": "1 Shanghai Jiao Tong University \n{zhaoheliao, ustcnewly, zhang-lq}@sjtu.edu.cn \n2 Tongji University \njiangtongli@tongji.edu.cn \n3 Ant Group \n{yelan.lj, huijia.zhj, weiqiang.wwq}@antgroup.com \n4 miguo.ai" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 308, + 455, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 308, + 455, + 452 + ], + "spans": [ + { + "bbox": [ + 159, + 308, + 455, + 452 + ], + "type": "text", + "content": "Abstract. Deep learning methods have significantly advanced the performance of image matting. However, dataset biases can mislead the matting models to biased behavior. In this paper, we identify the two typical biases in existing matting models, specifically contrast bias and transparency bias, and discuss their origins in matting datasets. To address these biases, we model the image matting task from the perspective of causal inference and identify the root causes of these biases: the confounders. To mitigate the effects of these confounders, we employ causal intervention through backdoor adjustment and introduce a novel model-agnostic cofounder intervened (COIN) matting framework. Extensive experiments across various matting methods and datasets have demonstrated that our COIN framework can significantly diminish such biases, thereby enhancing the performance of existing matting models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 462, + 394, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 462, + 394, + 474 + ], + "spans": [ + { + "bbox": [ + 160, + 462, + 394, + 474 + ], + "type": "text", + "content": "Keywords: Image matting " + }, + { + "bbox": [ + 160, + 462, + 394, + 474 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 462, + 394, + 474 + ], + "type": "text", + "content": " Intervention " + }, + { + "bbox": [ + 160, + 462, + 394, + 474 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 462, + 394, + 474 + ], + "type": "text", + "content": " Dataset bias" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 494, + 230, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 230, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 230, + 506 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 519, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 519, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 482, + 568 + ], + "type": "text", + "content": "Image matting aims at separating accurate foreground objects from natural images and precisely estimating the opacity on the boundary of the object. It serves as the basis for image editing, live-streaming, virtual meetings, and movie production [2,16,22,30,34]. Given an image " + }, + { + "bbox": [ + 130, + 519, + 482, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 519, + 482, + 568 + ], + "type": "text", + "content": ", it is formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 249, + 577, + 481, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 577, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 249, + 577, + 481, + 590 + ], + "type": "interline_equation", + "content": "\\boldsymbol {I} = \\boldsymbol {F} \\circ \\boldsymbol {\\alpha} + \\boldsymbol {B} \\circ (1 - \\boldsymbol {\\alpha}), \\tag {1}", + "image_path": "e39b8e29a7884cc92ad716735cc66560f24b773eed6d99dfe7fdfb3b0a92f9c8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 598, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 598, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 598, + 482, + 648 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 598, + 482, + 648 + ], + "type": "inline_equation", + "content": "\\pmb{F}" + }, + { + "bbox": [ + 130, + 598, + 482, + 648 + ], + "type": "text", + "content": " represents the foreground object, " + }, + { + "bbox": [ + 130, + 598, + 482, + 648 + ], + "type": "inline_equation", + "content": "\\pmb{B}" + }, + { + "bbox": [ + 130, + 598, + 482, + 648 + ], + "type": "text", + "content": " denotes the background, " + }, + { + "bbox": [ + 130, + 598, + 482, + 648 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 130, + 598, + 482, + 648 + ], + "type": "text", + "content": " is the opacity map, and " + }, + { + "bbox": [ + 130, + 598, + 482, + 648 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 130, + 598, + 482, + 648 + ], + "type": "text", + "content": " denotes element-wise multiplication. As shown in Eq. (1), solving the 7 unknown values in " + }, + { + "bbox": [ + 130, + 598, + 482, + 648 + ], + "type": "inline_equation", + "content": "\\{\\pmb{F},\\pmb{B},\\pmb{\\alpha}\\}" + }, + { + "bbox": [ + 130, + 598, + 482, + 648 + ], + "type": "text", + "content": " from the given 3 values in " + }, + { + "bbox": [ + 130, + 598, + 482, + 648 + ], + "type": "inline_equation", + "content": "\\pmb{I}" + }, + { + "bbox": [ + 130, + 598, + 482, + 648 + ], + "type": "text", + "content": " is a highly ill-posed problem. Therefore, some matting methods require" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 654, + 264, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 654, + 264, + 666 + ], + "spans": [ + { + "bbox": [ + 145, + 654, + 264, + 666 + ], + "type": "text", + "content": "*The corresponding authors." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 146, + 114, + 199, + 156 + ], + "blocks": [ + { + "bbox": [ + 146, + 114, + 199, + 156 + ], + "lines": [ + { + "bbox": [ + 146, + 114, + 199, + 156 + ], + "spans": [ + { + "bbox": [ + 146, + 114, + 199, + 156 + ], + "type": "image", + "image_path": "14de91af92126b961678bc5b518355048c06a5977c6b42e37a6e943820fcfffc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 159, + 184, + 167 + ], + "lines": [ + { + "bbox": [ + 162, + 159, + 184, + 167 + ], + "spans": [ + { + "bbox": [ + 162, + 159, + 184, + 167 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 201, + 114, + 252, + 156 + ], + "blocks": [ + { + "bbox": [ + 201, + 114, + 252, + 156 + ], + "lines": [ + { + "bbox": [ + 201, + 114, + 252, + 156 + ], + "spans": [ + { + "bbox": [ + 201, + 114, + 252, + 156 + ], + "type": "image", + "image_path": "93b59149e5d997bb10c7a46c5ee6c0d9642edb8dca81ad0bd926c811feb11b36.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 159, + 230, + 166 + ], + "lines": [ + { + "bbox": [ + 223, + 159, + 230, + 166 + ], + "spans": [ + { + "bbox": [ + 223, + 159, + 230, + 166 + ], + "type": "inline_equation", + "content": "\\alpha" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 253, + 114, + 306, + 157 + ], + "blocks": [ + { + "bbox": [ + 253, + 114, + 306, + 157 + ], + "lines": [ + { + "bbox": [ + 253, + 114, + 306, + 157 + ], + "spans": [ + { + "bbox": [ + 253, + 114, + 306, + 157 + ], + "type": "image", + "image_path": "6fdae009ebc94decdb41c076f87db2416bd95ba101fd64fd99f5dcdfcdb3903d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 159, + 282, + 166 + ], + "lines": [ + { + "bbox": [ + 276, + 159, + 282, + 166 + ], + "spans": [ + { + "bbox": [ + 276, + 159, + 282, + 166 + ], + "type": "text", + "content": "a" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 114, + 361, + 157 + ], + "blocks": [ + { + "bbox": [ + 306, + 114, + 361, + 157 + ], + "lines": [ + { + "bbox": [ + 306, + 114, + 361, + 157 + ], + "spans": [ + { + "bbox": [ + 306, + 114, + 361, + 157 + ], + "type": "image", + "image_path": "245126f2418f0d5cde82433173da08c8bf60ffdbad94e5273e1d99acd8aa5ed1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 159, + 345, + 167 + ], + "lines": [ + { + "bbox": [ + 323, + 159, + 345, + 167 + ], + "spans": [ + { + "bbox": [ + 323, + 159, + 345, + 167 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 362, + 114, + 414, + 157 + ], + "blocks": [ + { + "bbox": [ + 362, + 114, + 414, + 157 + ], + "lines": [ + { + "bbox": [ + 362, + 114, + 414, + 157 + ], + "spans": [ + { + "bbox": [ + 362, + 114, + 414, + 157 + ], + "type": "image", + "image_path": "3f25e11c133b67eabd6dc750d8faccf304318fc3be604b10933116464c5125f1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 159, + 391, + 166 + ], + "lines": [ + { + "bbox": [ + 384, + 159, + 391, + 166 + ], + "spans": [ + { + "bbox": [ + 384, + 159, + 391, + 166 + ], + "type": "inline_equation", + "content": "\\alpha" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 415, + 114, + 468, + 157 + ], + "blocks": [ + { + "bbox": [ + 415, + 114, + 468, + 157 + ], + "lines": [ + { + "bbox": [ + 415, + 114, + 468, + 157 + ], + "spans": [ + { + "bbox": [ + 415, + 114, + 468, + 157 + ], + "type": "image", + "image_path": "5639fd6a889a84c138afdc7b02b1f7857f8e7ffd9e562e40f10c6100eb63cd18.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 159, + 444, + 166 + ], + "lines": [ + { + "bbox": [ + 438, + 159, + 444, + 166 + ], + "spans": [ + { + "bbox": [ + 438, + 159, + 444, + 166 + ], + "type": "inline_equation", + "content": "\\hat{\\alpha}" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 146, + 178, + 200, + 231 + ], + "blocks": [ + { + "bbox": [ + 146, + 178, + 200, + 231 + ], + "lines": [ + { + "bbox": [ + 146, + 178, + 200, + 231 + ], + "spans": [ + { + "bbox": [ + 146, + 178, + 200, + 231 + ], + "type": "image", + "image_path": "896556478f5b782d812faf379a4f5fcafbaad205799e9372396f05f4e7cd0b1e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 232, + 184, + 240 + ], + "lines": [ + { + "bbox": [ + 162, + 232, + 184, + 240 + ], + "spans": [ + { + "bbox": [ + 162, + 232, + 184, + 240 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 201, + 178, + 252, + 231 + ], + "blocks": [ + { + "bbox": [ + 201, + 178, + 252, + 231 + ], + "lines": [ + { + "bbox": [ + 201, + 178, + 252, + 231 + ], + "spans": [ + { + "bbox": [ + 201, + 178, + 252, + 231 + ], + "type": "image", + "image_path": "e429041996ebb57e5f0459a82a60ee49e6eaaa1773e78aa3c03c2a2d88dadc84.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 233, + 230, + 239 + ], + "lines": [ + { + "bbox": [ + 223, + 233, + 230, + 239 + ], + "spans": [ + { + "bbox": [ + 223, + 233, + 230, + 239 + ], + "type": "inline_equation", + "content": "\\alpha" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 253, + 178, + 306, + 231 + ], + "blocks": [ + { + "bbox": [ + 268, + 169, + 346, + 177 + ], + "lines": [ + { + "bbox": [ + 268, + 169, + 346, + 177 + ], + "spans": [ + { + "bbox": [ + 268, + 169, + 346, + 177 + ], + "type": "text", + "content": "(a) The contrast bias." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 253, + 178, + 306, + 231 + ], + "lines": [ + { + "bbox": [ + 253, + 178, + 306, + 231 + ], + "spans": [ + { + "bbox": [ + 253, + 178, + 306, + 231 + ], + "type": "image", + "image_path": "388d52e96579fad8af456bf4f2dfdacad8019d3bb8ea8364a38cb13ab48e9396.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 232, + 282, + 239 + ], + "lines": [ + { + "bbox": [ + 277, + 232, + 282, + 239 + ], + "spans": [ + { + "bbox": [ + 277, + 232, + 282, + 239 + ], + "type": "text", + "content": "a" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 258, + 242, + 355, + 252 + ], + "lines": [ + { + "bbox": [ + 258, + 242, + 355, + 252 + ], + "spans": [ + { + "bbox": [ + 258, + 242, + 355, + 252 + ], + "type": "text", + "content": "(b) The transparency bias." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 306, + 178, + 359, + 231 + ], + "blocks": [ + { + "bbox": [ + 306, + 178, + 359, + 231 + ], + "lines": [ + { + "bbox": [ + 306, + 178, + 359, + 231 + ], + "spans": [ + { + "bbox": [ + 306, + 178, + 359, + 231 + ], + "type": "image", + "image_path": "999460dfdc2888e0ccbc7f6b71320b7443a490e69a38c13ea22ee5e0e7eef765.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 232, + 345, + 240 + ], + "lines": [ + { + "bbox": [ + 323, + 232, + 345, + 240 + ], + "spans": [ + { + "bbox": [ + 323, + 232, + 345, + 240 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 261, + 482, + 294 + ], + "lines": [ + { + "bbox": [ + 131, + 261, + 482, + 294 + ], + "spans": [ + { + "bbox": [ + 131, + 261, + 482, + 294 + ], + "type": "text", + "content": "Fig. 1: The observed contrast bias and transparency bias for MatteFormer [25] on the Compositional-1K dataset [40]. " + }, + { + "bbox": [ + 131, + 261, + 482, + 294 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 261, + 482, + 294 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 261, + 482, + 294 + ], + "type": "inline_equation", + "content": "\\hat{\\alpha}" + }, + { + "bbox": [ + 131, + 261, + 482, + 294 + ], + "type": "text", + "content": " denote the ground truth and predicted alpha mattes, respectively. Images are best viewed when zoomed in." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 361, + 178, + 413, + 231 + ], + "blocks": [ + { + "bbox": [ + 361, + 178, + 413, + 231 + ], + "lines": [ + { + "bbox": [ + 361, + 178, + 413, + 231 + ], + "spans": [ + { + "bbox": [ + 361, + 178, + 413, + 231 + ], + "type": "image", + "image_path": "7bdb3f12bab41d7960f98ba9f74084120324d9e25c35d9241bfa572b3471854a.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 232, + 391, + 239 + ], + "lines": [ + { + "bbox": [ + 386, + 232, + 391, + 239 + ], + "spans": [ + { + "bbox": [ + 386, + 232, + 391, + 239 + ], + "type": "inline_equation", + "content": "\\alpha" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 414, + 178, + 467, + 231 + ], + "blocks": [ + { + "bbox": [ + 414, + 178, + 467, + 231 + ], + "lines": [ + { + "bbox": [ + 414, + 178, + 467, + 231 + ], + "spans": [ + { + "bbox": [ + 414, + 178, + 467, + 231 + ], + "type": "image", + "image_path": "8a8e32032fce451a9455edadca0fcd66b83cde3e4cc4e2966cf615d6fb477275.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 232, + 444, + 239 + ], + "lines": [ + { + "bbox": [ + 438, + 232, + 444, + 239 + ], + "spans": [ + { + "bbox": [ + 438, + 232, + 444, + 239 + ], + "type": "inline_equation", + "content": "\\hat{\\alpha}" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 130, + 318, + 480, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 480, + 389 + ], + "type": "text", + "content": "additional user input as guidance, leading to a classification of matting methods based on whether and what type of user input is used. Trimap-based methods [40] are the most widely used, but obtaining a trimap can be challenging in some scenarios. Therefore, more flexible guidance, such as background [16,30] and user interaction [6,37,38], have been explored. Moreover, some methods [12,38] can also work without user guidance, which is concluded as guidance-free methods." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 130, + 390, + 481, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 481, + 557 + ], + "type": "text", + "content": "The rapid advances in deep learning have significantly boosted the performance of image matting. However, the matting dataset may contain biases, which may further lead to the biased behavior of image matting models. We identify two primary types of bias in existing models: contrast bias and transparency bias. To illustrate these biases, we present some results of a state-of-the-art (SOTA) model, namely MatteFormer [25], in Fig. 1. The contrast bias refers to the correlation between the contrast and the predicted alpha value. Specifically, the model tends to predict lower alpha values in areas with low contrast between foreground and background, as highlighted by the red boxes in Fig. 1a. Regarding the transparency bias, it is observed that the background texture tends to emerge in the predicted alpha matte for mid-transparent foregrounds. As highlighted by the red boxes in Fig. 1b, the predicted alpha for mid-transparent area contains the background texture (i.e., the texture of the rider and building for the left and right examples, respectively)." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Such biased behavior of models can be explained by the biases present in the dataset. For the contrast bias, we calculate the alpha value at different contrast levels within the training dataset as shown in Fig. 2a. We observe that the mean alpha level in low contrast areas (i.e., contrast " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "< 20" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " in the CIEDE2000 color difference formula [31]) is significantly lower than that in other areas. Such dataset bias directly contributes to the observed contrast bias as the model may falsely learn to predict alpha from contrast. Additionally, the number of samples in the low contrast area is significantly fewer than in other areas, making the model more easily led by such dataset bias in these areas. Regarding the" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Z. Liao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 115, + 288, + 191 + ], + "blocks": [ + { + "bbox": [ + 157, + 115, + 288, + 191 + ], + "lines": [ + { + "bbox": [ + 157, + 115, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 157, + 115, + 288, + 191 + ], + "type": "image", + "image_path": "e6322f0919a4f7911d908e23e2c46ae4ef906c65778ef225648025d3c8370208.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 194, + 292, + 205 + ], + "lines": [ + { + "bbox": [ + 162, + 194, + 292, + 205 + ], + "spans": [ + { + "bbox": [ + 162, + 194, + 292, + 205 + ], + "type": "text", + "content": "(a) The contrast bias in the dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 313, + 116, + 446, + 191 + ], + "blocks": [ + { + "bbox": [ + 313, + 116, + 446, + 191 + ], + "lines": [ + { + "bbox": [ + 313, + 116, + 446, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 116, + 446, + 191 + ], + "type": "image", + "image_path": "58238cbadfb1502d2decc14f05da253c660a2cdd89bb59daabc292f0d8c51c33.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 194, + 459, + 205 + ], + "lines": [ + { + "bbox": [ + 313, + 194, + 459, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 459, + 205 + ], + "type": "text", + "content": "(b) The transparency bias in the dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 209, + 213, + 402, + 224 + ], + "lines": [ + { + "bbox": [ + 209, + 213, + 402, + 224 + ], + "spans": [ + { + "bbox": [ + 209, + 213, + 402, + 224 + ], + "type": "text", + "content": "Fig. 2: The observe biases in matting datasets." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 254, + 480, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 254, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 480, + 338 + ], + "type": "text", + "content": "transparency bias, we analyze the distribution of transparency levels within the training dataset in Fig. 2b. We observe that most foreground objects are either highly salient or very opaque. Therefore, the alpha matte is expected to be either close to zero or exhibit textures in the image. For the mid-transparency area, whose alpha matte clearly differs from zero, the model is misled to focus on the texture of the image. Thus, the texture of the background in such areas may appear in the alpha matte, explaining the appearance of transparency bias." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 340, + 481, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 481, + 519 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 481, + 519 + ], + "type": "text", + "content": "To address these biases, we model the image matting task from the perspective of causal inference, and utilize the power of intervention to reduce the influence of the biases. The causal graph for image matting is depicted in Fig. 3a, where " + }, + { + "bbox": [ + 130, + 340, + 481, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{I},\\mathcal{F},\\mathcal{B}" + }, + { + "bbox": [ + 130, + 340, + 481, + 519 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 340, + 481, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 340, + 481, + 519 + ], + "type": "text", + "content": " are variables representing the image, foreground, background, and alpha matte respectively, while " + }, + { + "bbox": [ + 130, + 340, + 481, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 340, + 481, + 519 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 340, + 481, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 340, + 481, + 519 + ], + "type": "text", + "content": " are confounder variables denoting the contrast and the transparency of foreground. The notation " + }, + { + "bbox": [ + 130, + 340, + 481, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{B} \\gets \\mathcal{I} \\to \\mathcal{F}" + }, + { + "bbox": [ + 130, + 340, + 481, + 519 + ], + "type": "text", + "content": " suggests that an image can be decomposed into foreground and background, underpinning the basic assumption of image matting. " + }, + { + "bbox": [ + 130, + 340, + 481, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{F} \\to \\mathcal{A}" + }, + { + "bbox": [ + 130, + 340, + 481, + 519 + ], + "type": "text", + "content": " implies that the alpha matte is directly influenced by the foreground object, indicating that changes in the background should not affect the foreground's representation. The confounder variables, representing the common causes affecting multiple variables in the foreground, background, and image, serve as the source of biases. Typically, these can include properties of visual objects. Given that both the foreground and background consist of objects, these properties typically cause the presentation of the foreground, background, and images are presented." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": "However, directly measuring all possible confounders may be impractical, as we can not enumerate all properties of visual objects. Nevertheless, based on observed dataset biases, we can identify the most typical confounders. As highlighted in Fig. 3a, we focus on two typical confounders: contrast " + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": " and transparency " + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": ", which underlie the observed contrast bias and transparency bias, respectively. Contrast measures the color difference between foreground and background objects, while transparency refers to the opacity of the foreground. Therefore, we conclude " + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{F} \\gets \\mathcal{T} \\to \\mathcal{I}" + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{C} \\to \\mathcal{X}" + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\in \\{\\mathcal{F}, \\mathcal{I}, \\mathcal{B}\\}" + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": " by their definition. The dotted line indicates a potential correlation between " + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": ", suggesting that we cannot definitively assert that " + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{C} \\perp \\mathcal{T}" + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": ". While existing matting methods generally focus on maximizing the likelihood " + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "inline_equation", + "content": "P(\\mathcal{A}|\\mathcal{I})" + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": ", the presence of backdoor paths " + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{B} \\gets \\mathcal{C} \\to \\mathcal{F} \\to \\mathcal{A}" + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{B} \\gets \\mathcal{I} \\gets \\mathcal{C} \\to \\mathcal{F} \\to \\mathcal{A}" + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 521, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{B} \\gets \\mathcal{I} \\gets \\mathcal{T} \\to" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 204, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 204, + 91, + 448, + 103 + ], + "type": "text", + "content": "COIN-Matting: Confounder Intervention for Image Matting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 191, + 116, + 273, + 171 + ], + "blocks": [ + { + "bbox": [ + 191, + 116, + 273, + 171 + ], + "lines": [ + { + "bbox": [ + 191, + 116, + 273, + 171 + ], + "spans": [ + { + "bbox": [ + 191, + 116, + 273, + 171 + ], + "type": "image", + "image_path": "319a89db7530423775c1f64f267bd014cdcb0f50c2627b014e1cd3ed8a42c83a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 197, + 174, + 275, + 184 + ], + "lines": [ + { + "bbox": [ + 197, + 174, + 275, + 184 + ], + "spans": [ + { + "bbox": [ + 197, + 174, + 275, + 184 + ], + "type": "text", + "content": "(a) The causal graph." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 331, + 116, + 414, + 170 + ], + "blocks": [ + { + "bbox": [ + 331, + 116, + 414, + 170 + ], + "lines": [ + { + "bbox": [ + 331, + 116, + 414, + 170 + ], + "spans": [ + { + "bbox": [ + 331, + 116, + 414, + 170 + ], + "type": "image", + "image_path": "5aa8e90be0b0a511691bd86a7b1475e95476982e5fe4be345a976d66d1ebdf43.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 174, + 432, + 184 + ], + "lines": [ + { + "bbox": [ + 323, + 174, + 432, + 184 + ], + "spans": [ + { + "bbox": [ + 323, + 174, + 432, + 184 + ], + "type": "text", + "content": "(b) The backdoor adjustment." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 166, + 192, + 445, + 204 + ], + "lines": [ + { + "bbox": [ + 166, + 192, + 445, + 204 + ], + "spans": [ + { + "bbox": [ + 166, + 192, + 445, + 204 + ], + "type": "text", + "content": "Fig. 3: The causal graph and causal intervention for image matting." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 224, + 480, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 480, + 260 + ], + "type": "inline_equation", + "content": "\\mathcal{F} \\to \\mathcal{A}" + }, + { + "bbox": [ + 130, + 224, + 480, + 260 + ], + "type": "text", + "content": " results in " + }, + { + "bbox": [ + 130, + 224, + 480, + 260 + ], + "type": "inline_equation", + "content": "\\mathcal{B} \\not\\sqcup \\mathcal{A}|\\mathcal{I}" + }, + { + "bbox": [ + 130, + 224, + 480, + 260 + ], + "type": "text", + "content": ". This lack of conditional d-separation between " + }, + { + "bbox": [ + 130, + 224, + 480, + 260 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 130, + 224, + 480, + 260 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 224, + 480, + 260 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 224, + 480, + 260 + ], + "type": "text", + "content": " elucidates the co-occurrence of background textures in specific transparency regions and the contrast-induced alpha bias." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 261, + 482, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 488 + ], + "type": "text", + "content": "To remove the effect of these confounders, we propose our model-agnostic cofounder intervened (COIN) matting framework. Specifically, we explore the do-calculus to intervene in the influence of " + }, + { + "bbox": [ + 130, + 261, + 482, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 261, + 482, + 488 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 261, + 482, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 261, + 482, + 488 + ], + "type": "text", + "content": " via backdoor adjustment, as shown in Fig. 3b. The do-calculus actively assigns values to variables, rather than passively observing them from the data. Therefore, the do-calculus is an intervention on the influence of their direct causes, which is used for decofounding. While the ideal way to calculate the do-operator involves collecting all images with varying contrast and transparency into the training data, this approach is evidently impractical. Furthermore, the high cost of labeling the alpha matte makes it exceedingly difficult to even approximate the aforementioned intervention \"physically\". To address this issue, we apply backdoor adjustment to calculate " + }, + { + "bbox": [ + 130, + 261, + 482, + 488 + ], + "type": "inline_equation", + "content": "P(\\mathcal{A}|do(\\mathcal{F},\\mathcal{B},\\mathcal{I}))" + }, + { + "bbox": [ + 130, + 261, + 482, + 488 + ], + "type": "text", + "content": ", as described in Secs. 3.2 to 3.4. By actively assigning values to " + }, + { + "bbox": [ + 130, + 261, + 482, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{F},\\mathcal{B}" + }, + { + "bbox": [ + 130, + 261, + 482, + 488 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 261, + 482, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 261, + 482, + 488 + ], + "type": "text", + "content": ", we block backdoor paths and magnify the influence of cofounders, thus enhancing the robustness of the matting results against the previously mentioned dataset biases. By applying our framework to state-of-the-art (SOTA) image matting methods with different types of user guidance (i.e., MatteFormer [25], DIIM [6], and dugMatting [38]) and two datasets (i.e., Comp1K and Distinct-646), we verified the effectiveness of our framework through extensive experiments. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 135, + 492, + 480, + 586 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 136, + 492, + 480, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 492, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 136, + 492, + 480, + 515 + ], + "type": "text", + "content": "1. We examine the biases in matting methods, analyzing how dataset biases cause them and affect alpha matte prediction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 135, + 516, + 480, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 516, + 480, + 550 + ], + "spans": [ + { + "bbox": [ + 135, + 516, + 480, + 550 + ], + "type": "text", + "content": "2. We introduce a novel and model-agnostic confounder-intervened matting framework that employs causal intervention to sever the influence of confounders, thereby mitigating the aforementioned biases." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 135, + 551, + 480, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 551, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 135, + 551, + 480, + 586 + ], + "type": "text", + "content": "3. Through comprehensive experiments on various matting methods and multiple datasets, we demonstrate that our framework significantly enhances the performance of SOTA image matting methods." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 602, + 237, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 602, + 237, + 615 + ], + "spans": [ + { + "bbox": [ + 132, + 602, + 237, + 615 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 624, + 236, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 236, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 236, + 636 + ], + "type": "text", + "content": "2.1 Image Matting" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "As described in Sec. 1, image matting presents a highly ill-posed problem. The solutions can be classified based on the user input prior they rely upon." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Z. Liao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": "Among the prior-based image matting methods, the trimap is the most widely utilized prior [1,3,4,9,14,18,21,23,23,25,33,40-42]. Since [40] introduced the first well-known deep learning-based image matting method, featuring an encoder-decoder architecture, this framework has been extensively adopted and improved by subsequent methods. Researchers have enhanced this architecture from various perspectives, including generative adversarial training [23], matting with indices [21], affinity-based methods [3, 14], context-aware matting [9], the introduction of vision transformers [1, 25, 41], prior alignment [18], expansion to ultra-high resolution via sparsity [33], and data augmentation [4, 42]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 480, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 480, + 331 + ], + "type": "text", + "content": "However, trimap is not always available for real-world applications. Thus, recent researchers seek for more flexible and accessible user inputs [5, 6, 8, 13, 16, 30, 37, 38, 43]. Background matting [30] uses the background image as the prior, which can be more accessible for scenes like virtual meeting and livestreaming. It is further enhanced by recursive excitation [5] and extended into real-time version [16]. The coarse mask is used instead of trimap in [43] which shows better penitential of cooperating with the segmentation models [11]. The user feedbacks, such as clicks [6], scribbles [8], their combination [37], proposal selection [38] and even natural language [13], are also informative for matting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 350, + 246, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 350, + 246, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 350, + 246, + 361 + ], + "type": "text", + "content": "2.2 Causal Inference" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 369, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 489 + ], + "type": "text", + "content": "Causal inference [27, 29] plays a pivotal role in analyzing dataset bias and mitigating counter-causal correlations. It can be categorized into two main types: deconfounding and counterfactual inference. Serving as a fundamental statistical tool, it finds widespread application across various fields, such as image segmentation [44], image classification [35], image and video question answering [10, 15, 24], dialog systems [45], and recommendation systems [36]. Typically, causal-related methods conceptualize the task within a causal graph, where nodes symbolize variables, and edges denote their causal interactions. By examining the causal relations among variables, the causal graph facilitates identifying confounders and backdoor paths, thereby enabling the control of biases." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 508, + 233, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 508, + 233, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 508, + 233, + 522 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "This section outlines our intervention approach to mitigating biases introduced by confounders, as discussed in Sec. 1. In Sec. 3.1, we examine the role of confounder variables within the causal graph (i.e., Fig. 3), illustrating their impact on the performance of mating models through mathematical analysis. In Sec. 3.2, the concept of backdoor adjustment is introduced to counteract biases. In Sec. 3.3, we discuss the construction of space representations for each variable in the causal graph to facilitate backdoor adjustment implementation. In Sec. 3.4, the calculation of expectations, as derived in Sec. 3.2, is explained to execute the causal intervention. Finally, in Sec. 3.5, we integrate the above components into our model-agnostic COIN mating framework. Throughout this section, we denote variables by calligraphic letters (e.g., " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "), feature spaces" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 204, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 204, + 91, + 448, + 103 + ], + "type": "text", + "content": "COIN-Matting: Confounder Intervention for Image Matting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "content": "by boldface capital letters (e.g., " + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "content": "), and feature vectors by boldface lowercase letters (e.g., " + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "inline_equation", + "content": "\\mathbf{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "content": "). Additional introduction of background and more detailed derivations are provided in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 172, + 350, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 172, + 350, + 185 + ], + "spans": [ + { + "bbox": [ + 131, + 172, + 350, + 185 + ], + "type": "text", + "content": "3.1 Confounders and the Backdoor Paths" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 195, + 482, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 195, + 482, + 232 + ], + "spans": [ + { + "bbox": [ + 130, + 195, + 482, + 232 + ], + "type": "text", + "content": "In the task of image matting, confounders are variables that are the direct cause of at least two of the following: the foreground, background, and the image. As illustrated in Fig. 3a, the observational probability can be expressed as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 222, + 242, + 481, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 242, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 222, + 242, + 481, + 266 + ], + "type": "interline_equation", + "content": "P (\\mathcal {A} | \\mathcal {I}) = \\sum_ {\\boldsymbol {t}} \\sum_ {\\boldsymbol {c}} P (\\mathcal {A} | \\mathcal {I}, \\boldsymbol {t}, \\boldsymbol {c}) P (\\boldsymbol {c}, \\boldsymbol {t} | \\mathcal {I}), \\tag {2}", + "image_path": "af8b091b7ab085f8b3a5851969c98e6f595e83ddc1cf842c26bd4dacc9a19538.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 274, + 482, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 454 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 454 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 274, + 482, + 454 + ], + "type": "inline_equation", + "content": "\\pmb{t}" + }, + { + "bbox": [ + 130, + 274, + 482, + 454 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 274, + 482, + 454 + ], + "type": "inline_equation", + "content": "\\pmb{c}" + }, + { + "bbox": [ + 130, + 274, + 482, + 454 + ], + "type": "text", + "content": " represent the split of confounder " + }, + { + "bbox": [ + 130, + 274, + 482, + 454 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 274, + 482, + 454 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 274, + 482, + 454 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 274, + 482, + 454 + ], + "type": "text", + "content": ", respectively. Given that the observed distribution " + }, + { + "bbox": [ + 130, + 274, + 482, + 454 + ], + "type": "inline_equation", + "content": "P(\\pmb{c},\\pmb{t}|\\mathcal{I})" + }, + { + "bbox": [ + 130, + 274, + 482, + 454 + ], + "type": "text", + "content": " significantly deviates from a uniform distribution (as depicted in Fig. 2), a specific partition of " + }, + { + "bbox": [ + 130, + 274, + 482, + 454 + ], + "type": "inline_equation", + "content": "\\langle \\mathcal{C},\\mathcal{T}\\rangle" + }, + { + "bbox": [ + 130, + 274, + 482, + 454 + ], + "type": "text", + "content": " predominantly influences the observed probability " + }, + { + "bbox": [ + 130, + 274, + 482, + 454 + ], + "type": "inline_equation", + "content": "P(\\mathcal{A}|\\mathcal{I})" + }, + { + "bbox": [ + 130, + 274, + 482, + 454 + ], + "type": "text", + "content": " through a large " + }, + { + "bbox": [ + 130, + 274, + 482, + 454 + ], + "type": "inline_equation", + "content": "P(\\pmb{c},\\pmb{t}|\\mathcal{I})" + }, + { + "bbox": [ + 130, + 274, + 482, + 454 + ], + "type": "text", + "content": ". For instance, in the training dataset, regions with high contrast and low transparency are prevalent, leading to their dominance in the observed probability and observed biases. Specifically, in low-contrast regions, the alpha values are typically to be lower, causing the observed contrast bias. Furthermore, the model is trained to discern detailed textures in regions with opaque foregrounds and to assign nearly zero alpha values in highly transparent areas. Since these conditions consist most of the dataset, " + }, + { + "bbox": [ + 130, + 274, + 482, + 454 + ], + "type": "inline_equation", + "content": "P(\\pmb{c},\\pmb{t}|\\mathcal{I})" + }, + { + "bbox": [ + 130, + 274, + 482, + 454 + ], + "type": "text", + "content": " with extreme values of " + }, + { + "bbox": [ + 130, + 274, + 482, + 454 + ], + "type": "inline_equation", + "content": "\\pmb{t}" + }, + { + "bbox": [ + 130, + 274, + 482, + 454 + ], + "type": "text", + "content": " significantly contributes to " + }, + { + "bbox": [ + 130, + 274, + 482, + 454 + ], + "type": "inline_equation", + "content": "P(\\mathcal{A}|\\mathcal{I})" + }, + { + "bbox": [ + 130, + 274, + 482, + 454 + ], + "type": "text", + "content": ". Consequently, the model tends to extract highly textured patterns in medium transparency areas as in opaque foreground areas, where alpha values are apparently not as minimal as in transparent regions. Such background textures in the predicted alpha matte are concluded as transparency bias." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 455, + 483, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 483, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 483, + 552 + ], + "type": "text", + "content": "By adopting a causal perspective on Fig. 3a, we identify the presence of backdoor paths introduced by the confounder variables " + }, + { + "bbox": [ + 130, + 455, + 483, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 455, + 483, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 455, + 483, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 455, + 483, + 552 + ], + "type": "text", + "content": ". In particular, the backdoor paths " + }, + { + "bbox": [ + 130, + 455, + 483, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{A} \\gets \\mathcal{F} \\gets \\mathcal{C} \\to \\mathcal{B}" + }, + { + "bbox": [ + 130, + 455, + 483, + 552 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 455, + 483, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{B} \\gets \\mathcal{I} \\gets \\mathcal{C} \\to \\mathcal{F} \\to \\mathcal{A}" + }, + { + "bbox": [ + 130, + 455, + 483, + 552 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 455, + 483, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{A} \\gets \\mathcal{F} \\gets \\mathcal{T} \\to \\mathcal{I} \\to \\mathcal{B}" + }, + { + "bbox": [ + 130, + 455, + 483, + 552 + ], + "type": "text", + "content": " result in " + }, + { + "bbox": [ + 130, + 455, + 483, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{A} \\not\\perp \\mathcal{B}|\\mathcal{I}" + }, + { + "bbox": [ + 130, + 455, + 483, + 552 + ], + "type": "text", + "content": ", elucidating the influence of the background on the predicted alpha matte. More specifically, the former two paths and the latter are directly accountable for the contrast bias and transparency bias, respectively. To intervene the effects of the confounders and sever the backdoor paths they establish, we employ the backdoor adjustment [27,29] as detailed in Sec. 3.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 571, + 392, + 584 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 571, + 392, + 584 + ], + "spans": [ + { + "bbox": [ + 131, + 571, + 392, + 584 + ], + "type": "text", + "content": "3.2 Causal Intervention via Backdoor Adjustment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": "The do-calculus [27, 28] signifies active assignment of values to variables, distinguishing it from passive observation within the data, which constitutes intervention on the causal factors of variables. For brevity, we denote " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "P(do(\\mathcal{X} = x))" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "P(do(\\mathcal{X}))" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": ", implying active assignment of a value to variable " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": ". As illustrated in Fig. 3b, to sever all incoming causal links of " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{I}, \\mathcal{F}, \\mathcal{B}" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": " (i.e., nullifying the influence of " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": "), we focus on computing " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "P(\\mathcal{A}|do(\\mathcal{I},\\mathcal{F},\\mathcal{B}))" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": ". Given that " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "Z. Liao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 358, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 358, + 128 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 358, + 128 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 131, + 116, + 358, + 128 + ], + "type": "text", + "content": " block all specified backdoor paths, we can derive:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 161, + 136, + 480, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 136, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 161, + 136, + 480, + 186 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} P (\\mathcal {A} | d o (\\mathcal {I}, \\mathcal {F}, \\mathcal {B})) = \\sum_ {\\boldsymbol {t}} \\sum_ {\\boldsymbol {c}} P (\\mathcal {A} | d o (\\mathcal {I}, \\mathcal {F}, \\mathcal {B}), \\boldsymbol {t}, \\boldsymbol {c}) P (\\boldsymbol {c}, \\boldsymbol {t} | d o (\\mathcal {I}, \\mathcal {F}, \\mathcal {B})) \\\\ = \\sum_ {t} ^ {t} \\sum_ {c} ^ {c} P (\\mathcal {A} | i, \\boldsymbol {f}, \\boldsymbol {b}, t, c) P (\\boldsymbol {c}, t). \\tag {3} \\\\ \\end{array}", + "image_path": "b47d7d5f753832a08d0aa47e275484f7198e54c9178a9e834af7909e58e816a1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 193, + 479, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 193, + 479, + 217 + ], + "spans": [ + { + "bbox": [ + 131, + 193, + 479, + 217 + ], + "type": "text", + "content": "Moreover, based on Fig. 3b, we observe " + }, + { + "bbox": [ + 131, + 193, + 479, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{A} \\perp \\mathbf{b}|(\\mathbf{f},\\mathbf{i},\\mathbf{c},\\mathbf{t})" + }, + { + "bbox": [ + 131, + 193, + 479, + 217 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 193, + 479, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{A} \\perp \\mathbf{i}|(\\mathbf{f},\\mathbf{c},\\mathbf{t})" + }, + { + "bbox": [ + 131, + 193, + 479, + 217 + ], + "type": "text", + "content": ". Thus, the equation simplifies to:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 162, + 225, + 480, + 248 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 225, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 162, + 225, + 480, + 248 + ], + "type": "interline_equation", + "content": "P (\\mathcal {A} | d o (\\mathcal {I}, \\mathcal {F}, \\mathcal {B})) = \\sum_ {\\boldsymbol {t}} \\sum_ {\\boldsymbol {c}} P (\\mathcal {A} | \\boldsymbol {f}, \\boldsymbol {t}, \\boldsymbol {c}) P (\\boldsymbol {t}, \\boldsymbol {c}) = \\mathbb {E} _ {\\boldsymbol {t}, \\boldsymbol {c}} [ P (\\mathcal {A} | \\boldsymbol {f}, \\boldsymbol {t}, \\boldsymbol {c}) ]. \\tag {4}", + "image_path": "768da45f7264a99a808f4cf2b07d86b489e733306e9660555092f2a097180421.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 255, + 479, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 255, + 479, + 290 + ], + "spans": [ + { + "bbox": [ + 131, + 255, + 479, + 290 + ], + "type": "text", + "content": "Therefore, to predict the alpha matte with confounders intervened, we need to model the relevant spaces and calculate the expectations in Eq. (4), which are further discussed in Sec. 3.3 and Sec. 3.4, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 291, + 480, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 291, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 131, + 291, + 480, + 399 + ], + "type": "text", + "content": "However, to calculate the expectation in Eq. (4), obtaining the foreground representation is also necessary. Although the only input is the image (and the possible user guidance), directly inferring the foreground representation from the image representation is biased, as the confounders similarly facilitate backdoor paths to " + }, + { + "bbox": [ + 131, + 291, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 131, + 291, + 480, + 399 + ], + "type": "text", + "content": " as to " + }, + { + "bbox": [ + 131, + 291, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 131, + 291, + 480, + 399 + ], + "type": "text", + "content": ". Specifically, the backdoor paths " + }, + { + "bbox": [ + 131, + 291, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{F} \\gets \\mathcal{C} \\to \\mathcal{B}" + }, + { + "bbox": [ + 131, + 291, + 480, + 399 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 291, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{B} \\gets \\mathcal{I} \\gets \\mathcal{C} \\to \\mathcal{F}" + }, + { + "bbox": [ + 131, + 291, + 480, + 399 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 291, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{F} \\gets \\mathcal{T} \\to \\mathcal{I} \\to \\mathcal{B}" + }, + { + "bbox": [ + 131, + 291, + 480, + 399 + ], + "type": "text", + "content": " lead to " + }, + { + "bbox": [ + 131, + 291, + 480, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{F} \\not\\perp \\mathcal{B}|\\mathcal{I}" + }, + { + "bbox": [ + 131, + 291, + 480, + 399 + ], + "type": "text", + "content": ", indicating that the background's transparency and contrast similarly affect the foreground as influencing the prediction of the alpha matte. To address this issue, we also conduct an intervention for predicting the foreground, yielding:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 168, + 406, + 480, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 406, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 168, + 406, + 480, + 456 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} P (\\mathcal {F} | d o (\\mathcal {I}, \\mathcal {B})) = \\sum_ {\\boldsymbol {t}} \\sum_ {\\boldsymbol {c}} P (\\mathcal {F} | \\boldsymbol {i}, \\boldsymbol {b}, \\boldsymbol {t}, \\boldsymbol {c}) P (\\boldsymbol {c}, \\boldsymbol {t}) \\\\ = \\sum_ {t} ^ {t} \\sum_ {c} ^ {c} P (\\mathcal {F} | \\boldsymbol {i}, \\boldsymbol {t}, \\boldsymbol {c}) P (\\boldsymbol {c}, \\boldsymbol {t}) = \\mathbb {E} _ {\\boldsymbol {c}} \\mathbb {E} _ {\\boldsymbol {t}} [ P (\\mathcal {F} | \\boldsymbol {i}, \\boldsymbol {t}, \\boldsymbol {c}) ], \\tag {5} \\\\ \\end{array}", + "image_path": "a452024dffb12c7cd97da5dacb17fae7ebc396990653d29147459fc10d2a6133.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 464, + 479, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 464, + 479, + 499 + ], + "spans": [ + { + "bbox": [ + 131, + 464, + 479, + 499 + ], + "type": "text", + "content": "where the multi-scale image representation " + }, + { + "bbox": [ + 131, + 464, + 479, + 499 + ], + "type": "inline_equation", + "content": "\\pmb{i} = \\langle \\pmb{i}_1, \\pmb{i}_2, \\dots, \\pmb{i}_n \\rangle" + }, + { + "bbox": [ + 131, + 464, + 479, + 499 + ], + "type": "text", + "content": " is parameterized by the matting backbone, aligned with the scales of " + }, + { + "bbox": [ + 131, + 464, + 479, + 499 + ], + "type": "inline_equation", + "content": "\\pmb{C}" + }, + { + "bbox": [ + 131, + 464, + 479, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 464, + 479, + 499 + ], + "type": "inline_equation", + "content": "\\pmb{T}" + }, + { + "bbox": [ + 131, + 464, + 479, + 499 + ], + "type": "text", + "content": " as described in Sec. 3.3. The expectations are calculated as detailed in Sec. 3.4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 515, + 273, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 515, + 273, + 528 + ], + "spans": [ + { + "bbox": [ + 131, + 515, + 273, + 528 + ], + "type": "text", + "content": "3.3 Space Representations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 533, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 533, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 533, + 481, + 666 + ], + "type": "text", + "content": "Transparency Space " + }, + { + "bbox": [ + 131, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{T}" + }, + { + "bbox": [ + 131, + 533, + 481, + 666 + ], + "type": "text", + "content": ". Given the importance of multi-scale information in image matting, we divide the transparency space into several sub-spaces across different scales. Specifically, the transparency space " + }, + { + "bbox": [ + 131, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{T} = \\langle T_1, T_2, \\dots, T_n \\rangle" + }, + { + "bbox": [ + 131, + 533, + 481, + 666 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 131, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{T}_i" + }, + { + "bbox": [ + 131, + 533, + 481, + 666 + ], + "type": "text", + "content": " representing the transparency at the " + }, + { + "bbox": [ + 131, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 533, + 481, + 666 + ], + "type": "text", + "content": "-th scale, and " + }, + { + "bbox": [ + 131, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 131, + 533, + 481, + 666 + ], + "type": "text", + "content": " indicating the total number of sub-spaces. The " + }, + { + "bbox": [ + 131, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 533, + 481, + 666 + ], + "type": "text", + "content": "-th scale denotes the transparency of the image patches sized " + }, + { + "bbox": [ + 131, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "2^i \\times 2^i" + }, + { + "bbox": [ + 131, + 533, + 481, + 666 + ], + "type": "text", + "content": ". Considering the average alpha value as an approximation for the \"transparency level\" of an area, we define " + }, + { + "bbox": [ + 131, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{T}_i" + }, + { + "bbox": [ + 131, + 533, + 481, + 666 + ], + "type": "text", + "content": " as the average alpha value of pixels within the " + }, + { + "bbox": [ + 131, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 533, + 481, + 666 + ], + "type": "text", + "content": "-th scale patches. This formulation allows us to discretize the transparency sub-spaces further by categorizing the transparency level into " + }, + { + "bbox": [ + 131, + 533, + 481, + 666 + ], + "type": "inline_equation", + "content": "K_t" + }, + { + "bbox": [ + 131, + 533, + 481, + 666 + ], + "type": "text", + "content": " distinct classes, given its bounded nature within [0, 1]. The construction of the feature space entails the following steps:" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 204, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 204, + 91, + 448, + 103 + ], + "type": "text", + "content": "COIN-Matting: Confounder Intervention for Image Matting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 135, + 116, + 479, + 174 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 136, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 479, + 140 + ], + "type": "text", + "content": "1. Utilize a pretrained image feature extractor to derive the patch features at the " + }, + { + "bbox": [ + 136, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 136, + 116, + 479, + 140 + ], + "type": "text", + "content": "-th scale, for each " + }, + { + "bbox": [ + 136, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "i \\in \\{1, \\dots, n\\}" + }, + { + "bbox": [ + 136, + 116, + 479, + 140 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 140, + 477, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 140, + 477, + 150 + ], + "spans": [ + { + "bbox": [ + 135, + 140, + 477, + 150 + ], + "type": "text", + "content": "2. Determine the transparency level category " + }, + { + "bbox": [ + 135, + 140, + 477, + 150 + ], + "type": "inline_equation", + "content": "k \\in \\{1, \\dots, K_t\\}" + }, + { + "bbox": [ + 135, + 140, + 477, + 150 + ], + "type": "text", + "content": " for every patch." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 150, + 479, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 150, + 479, + 174 + ], + "spans": [ + { + "bbox": [ + 135, + 150, + 479, + 174 + ], + "type": "text", + "content": "3. Compute the average patch representation for all patches within the " + }, + { + "bbox": [ + 135, + 150, + 479, + 174 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 135, + 150, + 479, + 174 + ], + "type": "text", + "content": "-th category as the representation of this transparency category." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 177, + 479, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 177, + 479, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 479, + 201 + ], + "type": "text", + "content": "Hence, we formulate the feature space " + }, + { + "bbox": [ + 130, + 177, + 479, + 201 + ], + "type": "inline_equation", + "content": "\\pmb{T}_i\\in \\mathbb{R}^{d_i\\times K_t}" + }, + { + "bbox": [ + 130, + 177, + 479, + 201 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 177, + 479, + 201 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 130, + 177, + 479, + 201 + ], + "type": "text", + "content": " signifies the dimensionality of the feature extracted at the " + }, + { + "bbox": [ + 130, + 177, + 479, + 201 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 177, + 479, + 201 + ], + "type": "text", + "content": "-th scale." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 202, + 480, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 202, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 480, + 262 + ], + "type": "text", + "content": "Contrast Space " + }, + { + "bbox": [ + 130, + 202, + 480, + 262 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 202, + 480, + 262 + ], + "type": "text", + "content": ". Similarly to the transparency space, the contrast space is partitioned into multiple sub-spaces across various scales. We define " + }, + { + "bbox": [ + 130, + 202, + 480, + 262 + ], + "type": "inline_equation", + "content": "C = \\langle C_1, C_2, \\ldots, C_n \\rangle" + }, + { + "bbox": [ + 130, + 202, + 480, + 262 + ], + "type": "text", + "content": ", with each " + }, + { + "bbox": [ + 130, + 202, + 480, + 262 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 130, + 202, + 480, + 262 + ], + "type": "text", + "content": " representing contrast at the " + }, + { + "bbox": [ + 130, + 202, + 480, + 262 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 202, + 480, + 262 + ], + "type": "text", + "content": "-th scale, and " + }, + { + "bbox": [ + 130, + 202, + 480, + 262 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 202, + 480, + 262 + ], + "type": "text", + "content": " marking the total number of sub-spaces. Each " + }, + { + "bbox": [ + 130, + 202, + 480, + 262 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 202, + 480, + 262 + ], + "type": "text", + "content": "-th scale denotes the contrast of the foreground and background within " + }, + { + "bbox": [ + 130, + 202, + 480, + 262 + ], + "type": "inline_equation", + "content": "2^i \\times 2^i" + }, + { + "bbox": [ + 130, + 202, + 480, + 262 + ], + "type": "text", + "content": " image patches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 262, + 479, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 479, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 479, + 404 + ], + "type": "text", + "content": "Quantifying contrast, however, presents more complexity than measuring transparency. Initially, we estimate the color of a patch using the average color vector of all pixels within it in the color space. Subsequently, the contrast between foreground and background patches is determined by the color distance between their average color vectors. Given the multifaceted nature of color perception, we employ the CIEDE2000 color difference formula [31] for calculating color distance, as it closely aligns with human visual perception of color differences. This formula, incorporating lightness, chroma, and hue, yields a distance within the range of [0, 100]. Following this approach, we categorize the contrast level into " + }, + { + "bbox": [ + 130, + 262, + 479, + 404 + ], + "type": "inline_equation", + "content": "K_{c}" + }, + { + "bbox": [ + 130, + 262, + 479, + 404 + ], + "type": "text", + "content": " discrete classes and construct the feature space " + }, + { + "bbox": [ + 130, + 262, + 479, + 404 + ], + "type": "inline_equation", + "content": "C_i \\in \\mathbb{R}^{d_i \\times K_c}" + }, + { + "bbox": [ + 130, + 262, + 479, + 404 + ], + "type": "text", + "content": ", mirroring the methodology used for the transparency space. The primary adjustment involves transitioning from transparency to contrast categories for the classification." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 405, + 479, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 479, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 479, + 501 + ], + "type": "text", + "content": "Joint Confounder Space " + }, + { + "bbox": [ + 130, + 405, + 479, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 130, + 405, + 479, + 501 + ], + "type": "text", + "content": ". Given that " + }, + { + "bbox": [ + 130, + 405, + 479, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 405, + 479, + 501 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 405, + 479, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 405, + 479, + 501 + ], + "type": "text", + "content": " may not be independent, we cannot decompose the joint expectation in Eq. (4) into separate expectations for each. Therefore, it becomes necessary to model the joint space of " + }, + { + "bbox": [ + 130, + 405, + 479, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 130, + 405, + 479, + 501 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 405, + 479, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 130, + 405, + 479, + 501 + ], + "type": "text", + "content": ". An effective approach is to define the joint space as the Cartesian product of the sub-spaces. Consequently, the joint confounder space " + }, + { + "bbox": [ + 130, + 405, + 479, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} = \\langle \\mathbf{Z}_1, \\mathbf{Z}_2, \\dots, \\mathbf{Z}_n \\rangle = \\langle \\mathbf{T}_1 \\times \\mathbf{C}_1, \\mathbf{T}_2 \\times \\mathbf{C}_2, \\dots, \\mathbf{T}_n \\times \\mathbf{C}_n \\rangle" + }, + { + "bbox": [ + 130, + 405, + 479, + 501 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 130, + 405, + 479, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_i \\times \\mathbf{C}_i \\in \\mathbb{R}^{d_i \\times (K_t \\cdot K_c)}" + }, + { + "bbox": [ + 130, + 405, + 479, + 501 + ], + "type": "text", + "content": " represents a multidimensional space formed by the Cartesian product of transparency and contrast spaces at the " + }, + { + "bbox": [ + 130, + 405, + 479, + 501 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 405, + 479, + 501 + ], + "type": "text", + "content": "-th scale. Thus, Eqs. (4) and (5) are simplified as following:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 185, + 516, + 429, + 565 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 516, + 429, + 565 + ], + "spans": [ + { + "bbox": [ + 185, + 516, + 429, + 565 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} P(\\mathcal{A}|do(\\mathcal{I},\\mathcal{F},\\mathcal{B})) = \\sum_{\\boldsymbol {z}}P(\\mathcal{A}|\\boldsymbol {f},\\boldsymbol {z})P(\\boldsymbol {z}) = \\mathbb{E}_{\\boldsymbol{z}}[P(\\mathcal{A}|\\boldsymbol {f},\\boldsymbol {z})], \\\\ P (\\mathcal {F} | d o (\\mathcal {I}, \\mathcal {B})) = \\sum_ {\\boldsymbol {z}} P (\\mathcal {F} | \\boldsymbol {i}, \\boldsymbol {z}) P (\\boldsymbol {z}) = \\mathbb {E} _ {\\boldsymbol {z}} [ P (\\mathcal {F} | \\boldsymbol {i}, \\boldsymbol {z}) ], \\\\ \\end{array}", + "image_path": "4f32305574f50b1a6cc29b62af79ee38bd76cf9f3ed512b2ba376a0f828f1f01.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 473, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 473, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 473, + 583 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 570, + 473, + 583 + ], + "type": "inline_equation", + "content": "z = \\langle \\pmb {c},\\pmb {t}\\rangle" + }, + { + "bbox": [ + 130, + 570, + 473, + 583 + ], + "type": "text", + "content": " represents the split of joint observed confounders " + }, + { + "bbox": [ + 130, + 570, + 473, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} = \\langle \\mathcal{T},\\mathcal{C}\\rangle" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 596, + 227, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 596, + 227, + 608 + ], + "spans": [ + { + "bbox": [ + 131, + 596, + 227, + 608 + ], + "type": "text", + "content": "3.4 Expectations" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 613, + 479, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 613, + 479, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 613, + 479, + 648 + ], + "type": "text", + "content": "The expectations are calculated over the joint confounder space " + }, + { + "bbox": [ + 130, + 613, + 479, + 648 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 130, + 613, + 479, + 648 + ], + "type": "text", + "content": ". Utilizing the Normalized Weighted Geometric Mean (NWGM) approach [32,39], we reformulate the expectation as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 161, + 654, + 479, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 654, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 161, + 654, + 479, + 666 + ], + "type": "interline_equation", + "content": "P (\\mathcal {A} | d o (\\mathcal {I}, \\mathcal {F}, \\mathcal {B})) = \\mathbb {E} _ {\\boldsymbol {z}} P (\\mathcal {A} | \\boldsymbol {f}, \\boldsymbol {z}) = \\mathbb {E} _ {\\boldsymbol {z}} \\sigma (g (\\boldsymbol {f}, \\boldsymbol {z})) \\approx \\sigma (g (\\boldsymbol {f}, \\mathbb {E} _ {\\boldsymbol {z}} [ \\boldsymbol {z} ])), \\tag {7}", + "image_path": "cdbf1b74f793e7ff544be59f56e5bf1f22efb3da9910bdbdad2d8619111fbcf9.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "Z. Liao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 130, + 115, + 478, + 214 + ], + "blocks": [ + { + "bbox": [ + 130, + 115, + 478, + 214 + ], + "lines": [ + { + "bbox": [ + 130, + 115, + 478, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 478, + 214 + ], + "type": "image", + "image_path": "0d5db88a6f0a52bf1cca642701e284c995386ef7b531195b5404e588304da10d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 221, + 482, + 277 + ], + "lines": [ + { + "bbox": [ + 130, + 221, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 482, + 277 + ], + "type": "text", + "content": "Fig. 4: The COIN framework. " + }, + { + "bbox": [ + 130, + 221, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\pmb{i}" + }, + { + "bbox": [ + 130, + 221, + 482, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 221, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\pmb{f}" + }, + { + "bbox": [ + 130, + 221, + 482, + 277 + ], + "type": "text", + "content": " denote the image feature and foreground feature. " + }, + { + "bbox": [ + 130, + 221, + 482, + 277 + ], + "type": "inline_equation", + "content": "T, C" + }, + { + "bbox": [ + 130, + 221, + 482, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 221, + 482, + 277 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 221, + 482, + 277 + ], + "type": "text", + "content": " are transparency space, contrast space, and their joint confounder space respectively, which are constructed in Sec. 3.3. Spaces corresponding to the confounder variable are marked gray. EXP represents the expectations described in Sec. 3.4. " + }, + { + "bbox": [ + 130, + 221, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 130, + 221, + 482, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 221, + 482, + 277 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 130, + 221, + 482, + 277 + ], + "type": "text", + "content": " are ignored for simplicity. " + }, + { + "bbox": [ + 130, + 221, + 482, + 277 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 221, + 482, + 277 + ], + "type": "text", + "content": " represents the dot production." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 297, + 482, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 322 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 297, + 482, + 322 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 130, + 297, + 482, + 322 + ], + "type": "text", + "content": " denotes a sigmoid function and " + }, + { + "bbox": [ + 130, + 297, + 482, + 322 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 130, + 297, + 482, + 322 + ], + "type": "text", + "content": " represents a linear transformation for each sub-space. Similarly, for the foreground parameterization:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 182, + 327, + 481, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 327, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 182, + 327, + 481, + 340 + ], + "type": "interline_equation", + "content": "P (\\mathcal {F} | d o (\\mathcal {I}, \\mathcal {B})) = \\mathbb {E} _ {z} P (\\mathcal {F} | \\boldsymbol {i}, z) = \\mathbb {E} _ {z} g (\\boldsymbol {i}, z) = g (\\boldsymbol {i}, \\mathbb {E} _ {z} [ z ]). \\tag {8}", + "image_path": "7621c9949deeeee385965ef2e5d133a09bb8c80b09b46969718f54f1980303c4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 345, + 482, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 482, + 417 + ], + "type": "text", + "content": "This approach shifts the expectations to the feature level. To prevent model collapse, we use " + }, + { + "bbox": [ + 130, + 345, + 482, + 417 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\boldsymbol{z}|\\mathcal{F}}" + }, + { + "bbox": [ + 130, + 345, + 482, + 417 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 345, + 482, + 417 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\boldsymbol{z}|\\mathcal{I}}" + }, + { + "bbox": [ + 130, + 345, + 482, + 417 + ], + "type": "text", + "content": " as approximations for " + }, + { + "bbox": [ + 130, + 345, + 482, + 417 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_z" + }, + { + "bbox": [ + 130, + 345, + 482, + 417 + ], + "type": "text", + "content": " respectively. To emulate these feature-level expectations, we leverage a sub-space-wise attention mechanism. Take " + }, + { + "bbox": [ + 130, + 345, + 482, + 417 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\boldsymbol{z}|\\mathcal{I}}[\\boldsymbol{z}]" + }, + { + "bbox": [ + 130, + 345, + 482, + 417 + ], + "type": "text", + "content": " as an example. As spaces " + }, + { + "bbox": [ + 130, + 345, + 482, + 417 + ], + "type": "inline_equation", + "content": "Z_{i}" + }, + { + "bbox": [ + 130, + 345, + 482, + 417 + ], + "type": "text", + "content": " are heterogeneous with the feature spaces of " + }, + { + "bbox": [ + 130, + 345, + 482, + 417 + ], + "type": "inline_equation", + "content": "\\pmb{i}_j" + }, + { + "bbox": [ + 130, + 345, + 482, + 417 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 130, + 345, + 482, + 417 + ], + "type": "inline_equation", + "content": "i\\neq j" + }, + { + "bbox": [ + 130, + 345, + 482, + 417 + ], + "type": "text", + "content": ", we apply the attention mechanism for each scale respectively. Therefore, the expectation is" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 198, + 422, + 481, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 422, + 481, + 436 + ], + "spans": [ + { + "bbox": [ + 198, + 422, + 481, + 436 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\boldsymbol {z} \\mid \\mathcal {I}} [ \\boldsymbol {z} ] = \\left\\langle \\mathbb {E} _ {\\boldsymbol {z} _ {1} \\mid \\mathcal {I} _ {1}} [ \\boldsymbol {z} _ {1} ], \\mathbb {E} _ {\\boldsymbol {z} _ {2} \\mid \\mathcal {I} _ {2}} [ \\boldsymbol {z} _ {2} ], \\dots , \\mathbb {E} _ {\\boldsymbol {z} _ {n} \\mid \\mathcal {I} _ {n}} [ \\boldsymbol {z} _ {n} ] \\right\\rangle . \\tag {9}", + "image_path": "77a5bc373527d65678db3f60d4c3f7cc9406c761cad37ac917e606dc8c7985f5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "text", + "content": " is the number of sub-spaces. For attention on image patches at the " + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "text", + "content": "-th scale, the input is the contrast sub-space " + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "inline_equation", + "content": "Z_{i} \\in \\mathbb{R}^{d_{i} \\times (K_{t} \\cdot K_{c})}" + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "text", + "content": " and the image patch representation " + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{i}}_i \\in \\mathbb{R}^{d_i}" + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "text", + "content": " parameterized by the matting backbone, where " + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "text", + "content": " is the dimension of the latent representations. During such attention, the attention score emulates the distribution of " + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "inline_equation", + "content": "P(\\pmb{z}|\\mathcal{I})" + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "text", + "content": ", thus the weighted sum on values (i.e., " + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "text", + "content": ") emulates the expectation " + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\pmb{z}|\\mathcal{I}}[\\pmb{z}]" + }, + { + "bbox": [ + 130, + 441, + 482, + 514 + ], + "type": "text", + "content": ". The mathematical formulation is given by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 167, + 521, + 389, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 521, + 389, + 536 + ], + "spans": [ + { + "bbox": [ + 167, + 521, + 389, + 536 + ], + "type": "interline_equation", + "content": "\\pmb {h e a d} _ {i} ^ {k} = \\mathrm {s o f t m a x} (\\hat {\\pmb {i}} _ {i} ^ {k} \\pmb {W} _ {i, 1} ^ {k} (\\pmb {Z} _ {i} \\pmb {W} _ {i, 2} ^ {k}) ^ {T} / \\sqrt {d}) (\\pmb {Z} _ {i} \\pmb {W} _ {i, 3} ^ {k})", + "image_path": "71f065f8d516dfaa6a4b9b4e4e55139bc00a80d618c786afefafcda9aabf9925.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 183, + 539, + 480, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 539, + 480, + 554 + ], + "spans": [ + { + "bbox": [ + 183, + 539, + 480, + 554 + ], + "type": "interline_equation", + "content": "\\bar {z} _ {\\mathcal {I} _ {i}} = \\mathbb {E} _ {\\boldsymbol {z} _ {i} \\mid \\mathcal {I} _ {i}} [ \\boldsymbol {z} _ {i} ] = \\operatorname {C o n c a t} \\left(\\boldsymbol {h e a d} _ {i} ^ {1}, \\boldsymbol {h e a d} _ {i} ^ {2}, \\dots , \\boldsymbol {h e a d} _ {i} ^ {N _ {h}}\\right) \\boldsymbol {W} _ {i, o} \\tag {10}", + "image_path": "65fa79e5aabfbdd1edb19d0c87fc9ede7a4f14589fdf61b4d0f3469024b42595.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 187, + 556, + 396, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 556, + 396, + 571 + ], + "spans": [ + { + "bbox": [ + 187, + 556, + 396, + 571 + ], + "type": "interline_equation", + "content": "\\bar {z} _ {\\mathcal {I}} = \\operatorname {E X P} (\\hat {\\boldsymbol {i}}, \\boldsymbol {Z}) = \\mathbb {E} _ {\\boldsymbol {z} | \\mathcal {I}} [ \\boldsymbol {z} ] = \\left\\langle \\bar {z} _ {\\mathcal {I} _ {1}}, \\bar {z} _ {\\mathcal {I} _ {2}}, \\dots , \\bar {z} _ {\\mathcal {I} _ {n}} \\right\\rangle ,", + "image_path": "c1eed71c928654e95f76a52e2c263cf739056dd98d50f849d77b74b0b0b12625.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 573, + 482, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 573, + 482, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 573, + 482, + 611 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 573, + 482, + 611 + ], + "type": "inline_equation", + "content": "\\pmb{W}_{i,1}^{k}, \\pmb{W}_{i,2}^{k}, \\pmb{W}_{i,3}^{k}, \\pmb{W}_{i,o}" + }, + { + "bbox": [ + 130, + 573, + 482, + 611 + ], + "type": "text", + "content": " are trainable parameters, " + }, + { + "bbox": [ + 130, + 573, + 482, + 611 + ], + "type": "inline_equation", + "content": "N_{h}" + }, + { + "bbox": [ + 130, + 573, + 482, + 611 + ], + "type": "text", + "content": " represents the number of heads, and " + }, + { + "bbox": [ + 130, + 573, + 482, + 611 + ], + "type": "inline_equation", + "content": "\\mathrm{Concat}(\\cdot)" + }, + { + "bbox": [ + 130, + 573, + 482, + 611 + ], + "type": "text", + "content": " is the concatenation operation. This process effectively captures the expectations across different confounder sub-spaces." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 624, + 296, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 296, + 636 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 296, + 636 + ], + "type": "text", + "content": "3.5 COIN Matting Framework" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "The overall structure of our COIN framework is shown in Fig. 4. The confounder space " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{T}" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": " is similarly constructed as " + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": ", as described in Sec. 3.3. The mat" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 204, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 204, + 91, + 448, + 103 + ], + "type": "text", + "content": "COIN-Matting: Confounder Intervention for Image Matting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "content": "ting backbone is employed to parameterize the image representation " + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{i}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "content": ". Rather than directly predicting " + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "inline_equation", + "content": "P(\\mathcal{A}|\\mathcal{I})" + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "content": " as the matting backbone does, we engage in confounder intervention through backdoor adjustment to address the issues of contrast bias and transparency bias. The backdoor adjustment is implemented by the feature-level expectations. Specifically, we first process the joint space " + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\pmb{Z}" + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "content": " alongside the image representation " + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\pmb{i}" + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "content": " to compute " + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{z|\\mathcal{I}}[z]" + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "content": " via attention. This expectation is subsequently concatenated with the image representations to yield the foreground representation. Similarly, the foreground representation is used to calculate " + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{z|\\mathcal{F}}[z]" + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "content": " with joint confounder space " + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\pmb{Z}" + }, + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "content": ", and the expectations are further concatenated with the foreground representation to produce the alpha matte at various scales. Thus, we can express the overall framework as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 168, + 256, + 481, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 256, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 168, + 256, + 481, + 277 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {i}} = \\text {M a t t i n g B a c k b o n e} (\\boldsymbol {I}); \\quad \\bar {z} _ {\\mathcal {I}} = \\operatorname {E X P} (\\hat {\\boldsymbol {i}}, \\boldsymbol {Z}); \\quad \\hat {\\boldsymbol {f}} = g _ {f} (\\hat {\\boldsymbol {i}}, \\bar {z} _ {\\mathcal {I}}); \\tag {11}", + "image_path": "6b9229f5fc9042a4cfd7544069fa42c78edb91127a256f3e7802ab2180190ed0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 275, + 417, + 288 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 275, + 417, + 288 + ], + "spans": [ + { + "bbox": [ + 160, + 275, + 417, + 288 + ], + "type": "interline_equation", + "content": "\\bar {z} _ {\\mathcal {F}} = \\operatorname {E X P} (\\hat {\\boldsymbol {f}}, \\boldsymbol {Z}); \\quad \\hat {\\alpha} = P (\\mathcal {A} | d o (\\mathcal {I}, \\mathcal {F}, \\mathcal {B})) = \\sigma (g _ {\\alpha} (\\hat {\\boldsymbol {f}}, \\bar {z} _ {\\mathcal {F}})),", + "image_path": "90ffe1caa443acfb24585a867b9b2fd6b3619716bb7453590296346a7ded27bc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 297, + 479, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 479, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 479, + 319 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 297, + 479, + 319 + ], + "type": "inline_equation", + "content": "g_{f}(\\cdot)" + }, + { + "bbox": [ + 130, + 297, + 479, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 297, + 479, + 319 + ], + "type": "inline_equation", + "content": "g_{\\alpha}(\\cdot)" + }, + { + "bbox": [ + 130, + 297, + 479, + 319 + ], + "type": "text", + "content": " represent per-sub-space linear transformations, and " + }, + { + "bbox": [ + 130, + 297, + 479, + 319 + ], + "type": "inline_equation", + "content": "\\mathrm{EXP}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 297, + 479, + 319 + ], + "type": "text", + "content": " denotes the feature-level expectations described in Sec. 3.4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 321, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 441 + ], + "type": "text", + "content": "During the training of this framework, we supervise the alpha matte at each scale, similar to the approach taken by the backbone. Furthermore, we supervise the attention score, which approximates the probability distribution of the confounders when calculating expectations, with the ground truth categories of contrast and transparency. It is important to note that a closer category index of contrast and transparency signifies a more similar meaning, as these are discretized from the continuous values of color distance and transparency. Therefore, we employ Gaussian label smoothing prior to the cross-entropy for supervising the attention score. Additional details regarding the framework can be found in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 459, + 230, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 459, + 230, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 459, + 230, + 472 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 484, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 593 + ], + "type": "text", + "content": "In this section, we evaluate our COIN framework on extensive image matting baselines across various datasets. Specifically, we integrate our framework with SOTA image matting methods across all three types of matting methods: trimap-based methods (i.e., MatteFormer [25]), flexible guidance-based methods (i.e., DIIM [6] and dugMatting [38]) and guidance-free methods (i.e., dugMatting without human interaction [38]). We evaluate the improvement of our framework on Composition-1k [40], and Distinctions-646 [43] datasets. Additionally, we elaborate on the benefits of intervening on each confounder and the computational cost of implementing our framework in Sec. 4.3. The qualitative results are in Sec. 4.4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 609, + 264, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 264, + 623 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 264, + 623 + ], + "type": "text", + "content": "4.1 Experiment Settings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Datasets. We conduct the experiments on Composition-1k [40] and Distinctions-646 [43] datasets following [25]. Composition-1k dataset consists of 431 and 50 foreground objects in the training set and test set respectively. These foregrounds" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "Z. Liao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": "are composited with MS COCO [17] images for training and with VOC 2012 [7] images for testing. For testing, each foreground object is combined with 20 different images, generating a total of 1000 test images. The Distinctions-646 dataset contains 646 unique foreground objects, which are divided into 596 for training and 50 for testing. The composition process with background images follows the same procedure as in the Composition-1k dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 189, + 480, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 480, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 480, + 272 + ], + "type": "text", + "content": "Metrics. For measuring the overall performance of our framework, we adopt the widely used image matting metrics [12,25,26,43]: sum of absolute difference (SAD), mean square error (MSE), and the gradient (Grad) and connectivity (Conn). Additionally, to assess the improvement in reducing transparency and contrast biases, we measure SAD conditioned on each contrast and transparency category. The number of parameters and FLOPs are also reported to evaluate the computational complexity of implementing our framework." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 274, + 481, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 481, + 344 + ], + "type": "text", + "content": "Implementation Details. We set the number of sub-spaces as " + }, + { + "bbox": [ + 130, + 274, + 481, + 344 + ], + "type": "inline_equation", + "content": "n = 5" + }, + { + "bbox": [ + 130, + 274, + 481, + 344 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 274, + 481, + 344 + ], + "type": "inline_equation", + "content": "I, F, A, T" + }, + { + "bbox": [ + 130, + 274, + 481, + 344 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 274, + 481, + 344 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 274, + 481, + 344 + ], + "type": "text", + "content": ". For transparency and contrast levels, we set both " + }, + { + "bbox": [ + 130, + 274, + 481, + 344 + ], + "type": "inline_equation", + "content": "K_{t}" + }, + { + "bbox": [ + 130, + 274, + 481, + 344 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 274, + 481, + 344 + ], + "type": "inline_equation", + "content": "K_{c}" + }, + { + "bbox": [ + 130, + 274, + 481, + 344 + ], + "type": "text", + "content": " to 25. The contrast and transparency space is initialized by the average multiscale feature extracted by pretrained Swin-B [19, 20] as described in Sec. 3.3. We employ a " + }, + { + "bbox": [ + 130, + 274, + 481, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0, 1.2)" + }, + { + "bbox": [ + 130, + 274, + 481, + 344 + ], + "type": "text", + "content": " Gaussian distribution to smooth the label of attention score. More details can be found in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 366, + 228, + 377 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 366, + 228, + 377 + ], + "spans": [ + { + "bbox": [ + 132, + 366, + 228, + 377 + ], + "type": "text", + "content": "4.2 Main Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "text", + "content": "The improvement of our COIN framework is demonstrated in Tab. 1. It is evident that by intervening on the confounder variables, our COIN framework significantly enhances the performance across all matting baselines with varying guidance mechanisms on both the Composition-1k and Distinctions-646 datasets. Our framework substantially improves results for both guidance-free methods (e.g., dugMatting (0-sel)) and flexible-guidance based methods (e.g., dugMatting (1-sel) and DIIM). Additionally, even for the SOTA trimap-based baseline (e.g., MatteFormer), our COIN framework achieves significant performance gains, decreasing SAD by 1.9 and 2.6, and MSE by 0.7 and 1.5 on the two datasets respectively. Notably, our framework tends to exhibit greater improvements on the Distinctions-646 dataset compared to the Composition-1k dataset. This discrepancy is likely attributable to the more pronounced biases present within the Distinctions-646 dataset. Furthermore, enhancements are generally more substantial for flexible guidance-based or guidance-free matting methods than for trimap-based methods, since the former methods are typically more susceptible to biases in the absence of the detailed information provided by a trimap." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "To demonstrate that the enhancements are truly attributable to the reduction of contrast bias and transparency bias, we further evaluate the SAD improvements across each contrast and transparency category, as shown in Fig. 5. We selected MatteFormer as the baseline because it presents the most substantial challenge and exhibits the highest performance among the baselines. Regarding transparency, it is observable that our framework enhances the performance of baseline matting methods across all transparency levels. Furthermore, we note" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 204, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 204, + 91, + 448, + 103 + ], + "type": "text", + "content": "COIN-Matting: Confounder Intervention for Image Matting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 146, + 465, + 240 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 1: The comparison of applying our COIN framework on Composition-1k and Distinctions-646 dataset. The improvements are highlighted as superscripts." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 146, + 465, + 240 + ], + "lines": [ + { + "bbox": [ + 134, + 146, + 465, + 240 + ], + "spans": [ + { + "bbox": [ + 134, + 146, + 465, + 240 + ], + "type": "table", + "html": "
MethodComposition-1KDistinctions-646
SAD ↓MSE ↓Grad ↓Conn ↓SAD ↓MSE ↓Grad ↓Conn ↓
DIIM37.510.318.534.536.211.819.635.6
dugMatting (0-sel)*34.15.715.631.233.28.818.233.2
dugMatting (1-sel)*25.84.39.722.324.17.112.424.0
MatteFormer23.84.08.718.921.96.611.220.5
COIN(DIIM)33.8-3.78.1-2.232.0-2.59.6-2.2
COIN(dugMatting (0-sel)*)30.8-3.34.5-1.228.0-3.27.4-1.4
COIN(dugMatting (1-sel)*)23.7-2.13.5-0.819.1-3.26.3-0.8
COIN(MatteFormer)21.9-1.93.3-0.716.8-2.15.1-1.5
", + "image_path": "224a3ba2856cdff259e0ed0333a2af014edaef2a9d38ebc713aa01784f139b7c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 134, + 241, + 373, + 252 + ], + "lines": [ + { + "bbox": [ + 134, + 241, + 373, + 252 + ], + "spans": [ + { + "bbox": [ + 134, + 241, + 373, + 252 + ], + "type": "text", + "content": "* " + }, + { + "bbox": [ + 134, + 241, + 373, + 252 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 134, + 241, + 373, + 252 + ], + "type": "text", + "content": "-sel denotes " + }, + { + "bbox": [ + 134, + 241, + 373, + 252 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 134, + 241, + 373, + 252 + ], + "type": "text", + "content": " times human interaction during matting." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 146, + 280, + 298, + 376 + ], + "blocks": [ + { + "bbox": [ + 146, + 280, + 298, + 376 + ], + "lines": [ + { + "bbox": [ + 146, + 280, + 298, + 376 + ], + "spans": [ + { + "bbox": [ + 146, + 280, + 298, + 376 + ], + "type": "image", + "image_path": "1d89b42cc1878961ec7078efaaed3d0f0cc1a7fa41d65aeb5fcf7a7c1464aede.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 380, + 298, + 391 + ], + "lines": [ + { + "bbox": [ + 146, + 380, + 298, + 391 + ], + "spans": [ + { + "bbox": [ + 146, + 380, + 298, + 391 + ], + "type": "text", + "content": "(a) The SAD conditioned on transparency." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 315, + 280, + 468, + 376 + ], + "blocks": [ + { + "bbox": [ + 315, + 280, + 468, + 376 + ], + "lines": [ + { + "bbox": [ + 315, + 280, + 468, + 376 + ], + "spans": [ + { + "bbox": [ + 315, + 280, + 468, + 376 + ], + "type": "image", + "image_path": "a66a2a6a20115ec1fd03fa9f05831862e1142af7efb7ce26871be4b32e8b06f4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 380, + 459, + 390 + ], + "lines": [ + { + "bbox": [ + 323, + 380, + 459, + 390 + ], + "spans": [ + { + "bbox": [ + 323, + 380, + 459, + 390 + ], + "type": "text", + "content": "(b) The SAD conditioned on contrast." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 399, + 481, + 422 + ], + "lines": [ + { + "bbox": [ + 131, + 399, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 131, + 399, + 481, + 422 + ], + "type": "text", + "content": "Fig. 5: The improvement of applying our framework to MatteFormer on Composition-1K dataset conditioned on each contrast and transparency category." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": "that the improvement is particularly pronounced in regions where the transparency level is within the range of [50, 85]. The training data within this range, as discussed in Sec. 1, is notably scarce but exhibits a unique alpha pattern which differs from the pattern in highly opaque or transparent foregrounds. The observed probability distribution, dominated by data from regions with high opacity or transparency, causes the transparency bias we identified. The significant enhancements in these regions indicate that our framework is reducing the transparency bias through confounder intervention. In terms of contrast, our framework not only significantly reduces the SAD error across all contrast levels compared to the baseline method but also shows that the most notable improvement occurs in areas of relatively low contrast, particularly where the contrast level ranges from [10, 25]. In these regions, the detail patterns increase rapidly as average contrast value and mean alpha value raises (as shown in Fig. 2a). However, owing to the insufficient data relative to other regions, these areas are most susceptible to contrast bias. Thus, the most notable improvements in low-contrast areas suggest that our framework enhances performance by alleviating the contrast bias through intervention on the contrast confounder." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "Z. Liao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 157, + 471, + 224 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 2: The comparison of applying intervention on different confounder variables on the Distinctions-646 dataset using MatteFormer as the baseline. The improvements are highlighted as superscripts. The GFLOPs is calculated on " + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": " image." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 157, + 471, + 224 + ], + "lines": [ + { + "bbox": [ + 143, + 157, + 471, + 224 + ], + "spans": [ + { + "bbox": [ + 143, + 157, + 471, + 224 + ], + "type": "table", + "html": "
MethodErrorsComplexity
SAD ↓MSE ↓Grad ↓Conn ↓ParametersFLOPs
MatteFormer21.9-1.56.6-0.811.2-0.920.544.8M459G
MatteFormer + T intervene20.45.810.319.1-1.447.7M
MatteFormer + C intervene20.2-1.75.7-0.910.0-1.218.8-1.747.7M
COIN(MatteFormer)19.3-2.65.1-1.59.4-1.817.8-2.749.5M
", + "image_path": "4b0f68b65c47a12db5e5e2537054ffa062fea9faa2c67df13373f8e3a8c05014.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 245, + 238, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 245, + 238, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 245, + 238, + 258 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "content": "Confounder Spaces. We discuss the improvement from intervening in each confounder space individually rather than their joint confounder in Tab. 2. By comparing line 3 and line 2 with line 1, it is observed that intervening on either " + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "content": " can improve the performance of the baseline methods. However, because such interventions do not block the backdoor path opened by " + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "content": " respectively, the condition " + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{A} \\nsubseteq \\mathcal{B}|\\mathcal{T}" + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "content": " still exists, which leads to transparency bias and contrast bias respectively, resulting in a limited performance of the intervention framework. Moreover, it was noted that the performance of intervening on " + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "content": " is slightly better than that of intervening on " + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "content": ". This may be due to the fact that the contrast bias is more pronounced than the transparency bias in the dataset. By jointly intervening both " + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 266, + 482, + 410 + ], + "type": "text", + "content": ", we may block all aforementioned backdoor paths, thus further performance improvement in the COIN framework." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 411, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 578 + ], + "type": "text", + "content": "Complexity Analysis. As we introduce a framework atop existing matting methods, we further analyze its space and time complexity in terms of both parameters and FLOPs to delineate the cost of integrating our framework. As indicated in Tab. 2, the additional cost imposed by our framework is minimal when compared to the baseline methods. The incremental parameters (about " + }, + { + "bbox": [ + 130, + 411, + 482, + 578 + ], + "type": "inline_equation", + "content": "10.5\\%" + }, + { + "bbox": [ + 130, + 411, + 482, + 578 + ], + "type": "text", + "content": ") stem exclusively from the feature space and its corresponding attention mechanism, which are negligible relative to the baseline model. The only extra FLOPs (about " + }, + { + "bbox": [ + 130, + 411, + 482, + 578 + ], + "type": "inline_equation", + "content": "10.6\\%" + }, + { + "bbox": [ + 130, + 411, + 482, + 578 + ], + "type": "text", + "content": ") originate from the computation of expectations within feature spaces, which are characterized as " + }, + { + "bbox": [ + 130, + 411, + 482, + 578 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 411, + 482, + 578 + ], + "type": "text", + "content": " attentions across " + }, + { + "bbox": [ + 130, + 411, + 482, + 578 + ], + "type": "inline_equation", + "content": "d_{i} \\times (K_{t} \\cdot K_{c})" + }, + { + "bbox": [ + 130, + 411, + 482, + 578 + ], + "type": "text", + "content": " matrices and " + }, + { + "bbox": [ + 130, + 411, + 482, + 578 + ], + "type": "inline_equation", + "content": "d_{i} \\times L" + }, + { + "bbox": [ + 130, + 411, + 482, + 578 + ], + "type": "text", + "content": " image features, where " + }, + { + "bbox": [ + 130, + 411, + 482, + 578 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 411, + 482, + 578 + ], + "type": "text", + "content": " denotes the number of image patches at the " + }, + { + "bbox": [ + 130, + 411, + 482, + 578 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 411, + 482, + 578 + ], + "type": "text", + "content": "-th scale. Such attention is also lightweight compared to the matting baseline. Overall, our framework improves the performance of the SOTA image matting baseline between " + }, + { + "bbox": [ + 130, + 411, + 482, + 578 + ], + "type": "inline_equation", + "content": "11.6\\%" + }, + { + "bbox": [ + 130, + 411, + 482, + 578 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 411, + 482, + 578 + ], + "type": "inline_equation", + "content": "22.7\\%" + }, + { + "bbox": [ + 130, + 411, + 482, + 578 + ], + "type": "text", + "content": " in different metrics with about " + }, + { + "bbox": [ + 130, + 411, + 482, + 578 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 411, + 482, + 578 + ], + "type": "text", + "content": " computation cost, which is remarkably efficient for integrating with the baselines." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 597, + 258, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 258, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 258, + 609 + ], + "type": "text", + "content": "4.4 Qualitative Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "In this section, we present qualitative comparisons of our COIN framework. In Fig. 6, we compare the alpha matting results from baseline methods with those obtained using our COIN framework. Rows 1 showcase examples illustrating the reduction of contrast bias. It is evident that the low-contrast areas" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 204, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 204, + 91, + 448, + 103 + ], + "type": "text", + "content": "COIN-Matting: Confounder Intervention for Image Matting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 113, + 208, + 268 + ], + "blocks": [ + { + "bbox": [ + 143, + 113, + 208, + 268 + ], + "lines": [ + { + "bbox": [ + 143, + 113, + 208, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 113, + 208, + 268 + ], + "type": "image", + "image_path": "a7e5c4ac4cb85e594eff51389db6f21fbe9b57776fe97819aeb86dca8984bca8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 270, + 187, + 278 + ], + "lines": [ + { + "bbox": [ + 164, + 270, + 187, + 278 + ], + "spans": [ + { + "bbox": [ + 164, + 270, + 187, + 278 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 209, + 114, + 273, + 268 + ], + "blocks": [ + { + "bbox": [ + 209, + 114, + 273, + 268 + ], + "lines": [ + { + "bbox": [ + 209, + 114, + 273, + 268 + ], + "spans": [ + { + "bbox": [ + 209, + 114, + 273, + 268 + ], + "type": "image", + "image_path": "3289c8ca65ad8c01aa4a9fed4969cb5b1289153a7c276b5dd66c756b3a28f755.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 270, + 255, + 278 + ], + "lines": [ + { + "bbox": [ + 228, + 270, + 255, + 278 + ], + "spans": [ + { + "bbox": [ + 228, + 270, + 255, + 278 + ], + "type": "text", + "content": "Trimap" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 275, + 114, + 339, + 268 + ], + "blocks": [ + { + "bbox": [ + 275, + 114, + 339, + 268 + ], + "lines": [ + { + "bbox": [ + 275, + 114, + 339, + 268 + ], + "spans": [ + { + "bbox": [ + 275, + 114, + 339, + 268 + ], + "type": "image", + "image_path": "df03e1bbec89ea81c0ab77ac5d28a4abd6c0ae2e73d231b1cdf61d20b426231c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 284, + 270, + 330, + 278 + ], + "lines": [ + { + "bbox": [ + 284, + 270, + 330, + 278 + ], + "spans": [ + { + "bbox": [ + 284, + 270, + 330, + 278 + ], + "type": "text", + "content": "Alpha matte" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 340, + 114, + 405, + 268 + ], + "blocks": [ + { + "bbox": [ + 340, + 114, + 405, + 268 + ], + "lines": [ + { + "bbox": [ + 340, + 114, + 405, + 268 + ], + "spans": [ + { + "bbox": [ + 340, + 114, + 405, + 268 + ], + "type": "image", + "image_path": "5f15eef4afc378abdfe2720863b31a6e8ba706c6314a3e984e3624c906cbe1ba.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 270, + 396, + 278 + ], + "lines": [ + { + "bbox": [ + 348, + 270, + 396, + 278 + ], + "spans": [ + { + "bbox": [ + 348, + 270, + 396, + 278 + ], + "type": "text", + "content": "MatteFormer" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 288, + 482, + 312 + ], + "lines": [ + { + "bbox": [ + 131, + 288, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 131, + 288, + 482, + 312 + ], + "type": "text", + "content": "Fig. 6: The visualized comparison of applying our COIN framework on Matte-Former [25]. Images are best viewed when zoomed in." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 406, + 114, + 471, + 268 + ], + "blocks": [ + { + "bbox": [ + 406, + 114, + 471, + 268 + ], + "lines": [ + { + "bbox": [ + 406, + 114, + 471, + 268 + ], + "spans": [ + { + "bbox": [ + 406, + 114, + 471, + 268 + ], + "type": "image", + "image_path": "5f5e8270e9aa68e41c5d5ac858d2b2466c4eeb52ae62db1e31cf73303c1c4545.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 270, + 449, + 277 + ], + "lines": [ + { + "bbox": [ + 427, + 270, + 449, + 277 + ], + "spans": [ + { + "bbox": [ + 427, + 270, + 449, + 277 + ], + "type": "text", + "content": "COIN" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 333, + 482, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 482, + 465 + ], + "type": "text", + "content": "highlighted in the red boxes yield alpha mattes with lower values when processed by MatteFormer. However, the integration of our COIN framework significantly diminishes this contrast bias, resulting in more precise alpha mattes. Rows 2 and 3 demonstrate the reduction of transparency bias. The highlighted mid-transparent regions in the red boxes contain background textures (i.e., the texture of text and the car cage respectively) in the predicted alpha mattes. Nevertheless, with the application of our COIN framework, the alpha mattes become markedly more accurate, effectively eliminating such textures from the mid-transparent areas. Such qualitative results prove that our COIN framework can effectively reduce both contrast bias and transparency bias, therefore improving the performance of the existing image matting methods." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 481, + 220, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 481, + 220, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 481, + 220, + 495 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 505, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 505, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 505, + 482, + 578 + ], + "type": "text", + "content": "In this work, we have concentrated on the contrast bias and transparency bias in existing matting methods and demonstrated how dataset biases lead to biased behaviors of models. We analyzed these biases from a causal perspective and designed the confounder intervention framework to alleviate them. Extensive experiments have demonstrated that our framework markedly alleviates these biases and enhances SOTA matting methods with minor computational cost." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 594, + 247, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 247, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 247, + 608 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "content": "The work was supported by the National Natural Science Foundation of China (Grant No. 62076162), the Shanghai Municipal Science and Technology Major Project, China (Grant No. 2021SHZDZX0102) and the Postdoctoral Fellowship Program of CPSF (Grant No. GZC20241225)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Z. Liao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 163 + ], + "type": "text", + "content": "1. Cai, H., Xue, F., Xu, L., Guo, L.: Transmatting: Enhancing transparent objects matting with transformers. In: ECCV (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 164, + 480, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 164, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 164, + 480, + 186 + ], + "type": "text", + "content": "2. Chen, J., Li, X., Luo, L., Ma, J.: Multi-focus image fusion based on multi-scale gradients and image matting. IEEE Trans. Multim. 24, 655-667 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 186, + 480, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 186, + 480, + 208 + ], + "spans": [ + { + "bbox": [ + 138, + 186, + 480, + 208 + ], + "type": "text", + "content": "3. Dai, Y., Lu, H., Shen, C.: Learning affinity-aware upsampling for deep image matting. In: CVPR (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 209, + 480, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 480, + 231 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 480, + 231 + ], + "type": "text", + "content": "4. Dai, Y., Price, B., Zhang, H., Shen, C.: Boosting robustness of image matting with context assembling and strong data augmentation. In: CVPR (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 232, + 480, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 480, + 252 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 480, + 252 + ], + "type": "text", + "content": "5. Deng, J., Xu, Y., Zhou, Z., He, S.: Background matting via recursive excitation. In: ICME (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 254, + 480, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 254, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 138, + 254, + 480, + 275 + ], + "type": "text", + "content": "6. Ding, H., Zhang, H., Liu, C., Jiang, X.: Deep interactive image matting with feature propagation. IEEE Trans. Image Process. 31, 2421-2432 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 276, + 480, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 276, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 276, + 480, + 297 + ], + "type": "text", + "content": "7. Everingham, M., Gool, L.V., Williams, C.K.I., Winn, J.M., Zisserman, A.: The pascal visual object classes (VOC) challenge. Int. J. Comput. Vis. (2010)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 298, + 480, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 298, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 138, + 298, + 480, + 330 + ], + "type": "text", + "content": "8. Fang, X., Zhang, S., Chen, T., Wu, X., Shamir, A., Hu, S.: User-guided deep human image matting using arbitrary trimaps. IEEE Trans. Image Process. 31, 2040-2052 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 331, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 480, + 353 + ], + "type": "text", + "content": "9. Hou, Q., Liu, F.: Context-aware image matting for simultaneous foreground and alpha estimation. In: ICCV (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 354, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 480, + 374 + ], + "type": "text", + "content": "0. Li, J., Niu, L., Zhang, L.: Knowledge proxy intervention for deconfounded video question answering. In: ICCV (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 376, + 480, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 480, + 398 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 480, + 398 + ], + "type": "text", + "content": "1. Li, J., Wang, W., Chen, J., Niu, L., Si, J., Qian, C., Zhang, L.: Video semantic segmentation via sparse temporal transformer. In: ACM MM (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 399, + 480, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 480, + 409 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 480, + 409 + ], + "type": "text", + "content": "2. Li, J., Zhang, J., Tao, D.: Deep automatic natural image matting. In: IJCAI (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 410, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 410, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 138, + 410, + 480, + 430 + ], + "type": "text", + "content": "3. Li, J., Zhang, J., Tao, D.: Referring image matting. In: Proceedings of the IEEE Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 432, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 432, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 432, + 480, + 453 + ], + "type": "text", + "content": "4. Li, Y., Lu, H.: Natural image matting via guided contextual attention. In: AAAI (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 454, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 454, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 454, + 480, + 475 + ], + "type": "text", + "content": "5. Li, Y., Wang, X., Xiao, J., Chua, T.: Equivariant and invariant grounding for video question answering. In: ACM MM (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 476, + 480, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 476, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 476, + 480, + 498 + ], + "type": "text", + "content": "6. Lin, S., Ryabtsev, A., Sengupta, S., Curless, B.L., Seitz, S.M., Kemelmacher-Shlizerman, I.: Real-time high-resolution background matting. In: CVPR (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 498, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 480, + 520 + ], + "type": "text", + "content": "7. Lin, T., Maire, M., Belongie, S.J., Hays, J., Perona, P., Ramanan, D., Dóllar, P., Zitnick, C.L.: Microsoft COCO: common objects in context. In: ECCV (2014)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 521, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 480, + 542 + ], + "type": "text", + "content": "8. Liu, Y., Xie, J., Qiao, Y., Tang, Y., Yang, X.: Prior-induced information alignment for image matting. IEEE Trans. Multim. 24, 2727-2738 (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 544, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 575 + ], + "type": "text", + "content": "9. Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., Wei, F., Guo, B.: Swin transformer v2: Scaling up capacity and resolution. In: CVPR (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 576, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 576, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 576, + 480, + 597 + ], + "type": "text", + "content": "20. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "type": "text", + "content": "21. Lu, H., Dai, Y., Shen, C., Xu, S.: Indices matter: Learning to index for deep image matting. In: ICCV (2019)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "text", + "content": "22. Lu, L., Li, J., Cao, J., Niu, L., Zhang, L.: Painterly image harmonization using diffusion model. In: ACM MM (2023)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "23. Lutz, S., Amplanitis, K., Smolic, A.: Alphagan: Generative adversarial networks for natural image matting. In: BMVC (2018)" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 204, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 204, + 91, + 447, + 102 + ], + "type": "text", + "content": "COIN-Matting: Confounder Intervention for Image Matting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 621 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "24. Niu, Y., Tang, K., Zhang, H., Lu, Z., Hua, X., Wen, J.: Counterfactual VQA: A cause-effect look at language bias. In: CVPR (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "text", + "content": "25. Park, G., Son, S., Yoo, J., Kim, S., Kwak, N.: Matteformer: Transformer-based image matting via prior-tokens. In: CVPR (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "text", + "content": "26. Park, K., Woo, S., Oh, S.W., Kweon, I.S., Lee, J.Y.: Mask-guided matting in the wild. In: CVPR (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 194 + ], + "type": "text", + "content": "27. Pearl, J., Glymour, M., Jewell, N.P.: Causal inference in statistics: A primer (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "text", + "content": "28. Pearl, J., Mackenzie, D.: The Book of Why: The New Science of Cause and Effect. Basic Books, Inc. (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "29. Rubin, D.B.: Essential concepts of causal inference: a remarkable history and an intriguing future. Biostatistics & Epidemiology (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 238, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 259 + ], + "type": "text", + "content": "30. Sengupta, S., Jayaram, V., Curless, B., Seitz, S.M., Kemelmacher-Shlizerman, I.: Background matting: The world is your green screen. In: CVPR (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 260, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 480, + 303 + ], + "type": "text", + "content": "31. Sharma, G., Wu, W., Dalal, E.N.: The ciede2000 color-difference formula: Implementation notes, supplementary test data, and mathematical observations. Color Research and Application 30, 21-30 (2005), https://api-semanticscholar.org/ CorpusID:29119937" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "32. Srivastava, N., Hinton, G.E., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. J. Mach. Learn. Res. (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "33. Sun, Y., Tang, C., Tai, Y.: Ultrahigh resolution image/video matting with spatiotemporal sparsity. In: CVPR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "text", + "content": "34. Tan, L., Li, J., Niu, L., Zhang, L.: Deep image harmonization in dual color spaces. In: ACM MM (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "35. Wang, T., Zhou, C., Sun, Q., Zhang, H.: Causal attention for unbiased visual recognition. In: ICCV (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "text", + "content": "36. Wang, W., Feng, F., He, X., Zhang, H., seng Chua, T.: Clicks can be cheating: Counterfactual recommendation for mitigating clickbait issue (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "text", + "content": "37. Wei, T., Chen, D., Zhou, W., Liao, J., Zhao, H., Zhang, W., Hua, G., Yu, N.: Deep image matting with sparse user interactions. TPAMI (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "text", + "content": "38. Wu, J., Zhang, C., Li, Z., Fu, H., Peng, X., Zhou, J.T.: dugMatting: Decomposed-uncertainty-guided matting. In: ICML (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "text", + "content": "39. Xu, K., Ba, J., Kiros, R., Cho, K., Courville, A.C., Salakhutdinov, R., Zemel, R.S., Bengio, Y.: Show, attend and tell: Neural image caption generation with visual attention. In: ICML (2015)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 501, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 512 + ], + "type": "text", + "content": "40. Xu, N., Price, B.L., Cohen, S., Huang, T.S.: Deep image matting. In: CVPR (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 513, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 513, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 513, + 480, + 533 + ], + "type": "text", + "content": "41. Yao, J., Wang, X., Yang, S., Wang, B.: Vitmatte: Boosting image matting with pre-trained plain vision transformers. Inf. Fusion 103, 102091 (2024)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "text", + "content": "42. Ye, Z., Dai, Y., Hong, C., Cao, Z., Lu, H.: Infusing definiteness into randomness: Rethinking composition styles for deep image matting. In: AAAI (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 556, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 480, + 578 + ], + "type": "text", + "content": "43. Yu, Q., Zhang, J., Zhang, H., Wang, Y., Lin, Z., Xu, N., Bai, Y., Yuille, A.L.: Mask guided matting via progressive refinement network. In: CVPR (2021)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 578, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 598 + ], + "type": "text", + "content": "44. Zhang, D., Zhang, H., Tang, J., Hua, X., Sun, Q.: Causal intervention for weakly-supervised semantic segmentation. In: NeurIPS (2020)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 599, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 480, + 621 + ], + "type": "text", + "content": "45. Zhu, Q., Zhang, W.N., Liu, T., Wang, W.Y.: Counterfactual off-policy training for neural dialogue generation. In: EMNLP (2020)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "Z. Liao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/COIN_ Control-Inpainting Diffusion Prior for Human and Camera Motion Estimation/213c3e91-2355-4081-9e3c-28b7d1b0600c_content_list.json b/2024/COIN_ Control-Inpainting Diffusion Prior for Human and Camera Motion Estimation/213c3e91-2355-4081-9e3c-28b7d1b0600c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..05df3fb40481118fd98a276fbb308bbe7c36d95b --- /dev/null +++ b/2024/COIN_ Control-Inpainting Diffusion Prior for Human and Camera Motion Estimation/213c3e91-2355-4081-9e3c-28b7d1b0600c_content_list.json @@ -0,0 +1,1768 @@ +[ + { + "type": "text", + "text": "COIN: Control-Inpainting Diffusion Prior for Human and Camera Motion Estimation", + "text_level": 1, + "bbox": [ + 243, + 140, + 759, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiefeng Li $^{1,2}$ , Ye Yuan $^{1}$ , Davis Rempe $^{1}$ , Haotian Zhang $^{1}$ , Pavlo Molchanov $^{1}$ , Cewu Lu $^{2}$ , Jan Kautz $^{1}$ , and Umar Iqbal $^{1}$", + "bbox": [ + 225, + 193, + 776, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ NVIDIA $^{2}$ Shanghai Jiao Tong University https://nvlabs.github.io/COIN/", + "bbox": [ + 351, + 234, + 650, + 267 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9fb51b864a9ead7976a41d75c7841abdf6d41033d1beac89fc244f7e0bf84e22.jpg", + "image_caption": [ + "Fig. 1: Capturing global human and camera motion from a dynamic camera presents unique challenges. In the input video, a person is riding a skateboard - while the local body motion may remain relatively constant, the global position of the individual changes significantly. Current state-of-the-art methods such as PACE [41] and WHAM [81] fail catastrophically on such out-of-distribution motions. Our approach, COIN, gracefully handles such challenging cases, owing to our control-inpainting motion diffusion prior and novel human-scene relation loss." + ], + "image_footnote": [], + "bbox": [ + 217, + 273, + 787, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Estimating global human motion from moving cameras is challenging due to the entanglement of human and camera motions. To mitigate the ambiguity, existing methods leverage learned human motion priors, which however often result in oversmoothed motions with misaligned 2D projections. To tackle this problem, we propose COIN, a control-inpainting motion diffusion prior that enables fine-grained control to disentangle human and camera motions. Although pre-trained motion diffusion models encode rich motion priors, we find it non-trivial to leverage such knowledge to guide global motion estimation from RGB videos. COIN introduces a novel control-inpainting score distillation sampling method to ensure well-aligned, consistent, and high-quality motion from the diffusion prior within a joint optimization framework. Furthermore, we introduce a new human-scene relation loss to alleviate the scale ambiguity by enforcing consistency among the humans, camera, and scene. Experiments on three challenging benchmarks demonstrate the effectiveness of COIN, which outperforms the state-of-the-art methods in terms of global human motion estimation and camera motion estimation. As an illustrative example, COIN outperforms the state-of-the-art method by $33\\%$ in world joint position error (W-MPJPE) on the RICH dataset.", + "bbox": [ + 259, + 577, + 740, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 202, + 375, + 218 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recovering global human and camera motion from dynamic RGB videos is an important problem with many applications, such as animation, human-computer interaction, mixed reality, and robotics. However, it is a very challenging problem due to the entanglement of human and camera motion.", + "bbox": [ + 212, + 234, + 782, + 294 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are only a few works [41, 81, 99, 100] that try to address this problem. Earlier methods [47, 100] only focus on human motion and ignore the camera motion. Their core insight is that the global body motion is highly correlated with the local motion. Thus, they can use the local body movements to estimate the global orientation and trajectory with a regression model [100] or by combining them with physics constraints [47]. However, these regression models ignore the camera movements, so they fail to maintain consistency with the input video, whereas physics-based methods fail to model complex in-the-wild environments so are limited to controlled scenarios. Recent works [41, 99] try to jointly estimate the human and camera motion by exploiting learned motion priors [21, 76] and SLAM [66, 89, 90]. They try to constrain the human body motion in a low-dimensional latent space of a motion prior model, which results in reconstructed motions that are overly smooth and do not align well with video observations. Moreover, the optimization of the camera motion is only based on the global human motion from the motion prior. Hence, they fail catastrophically if the initial human motion predictions are significantly incorrect (as shown in Fig. 1).", + "bbox": [ + 212, + 295, + 785, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More recently, Denoising Diffusion Models [24,91] have emerged as a powerful family of generative models that can model high-quality data priors. Nonetheless, effectively leveraging the learned priors remains an ongoing challenge. Score Distillation Sampling (SDS) is commonly employed for this purpose [74], but we find that naive application of SDS also results in inconsistencies with the available observations (see Sec. 4.2). The root cause of this problem lies in the inconsistency of randomly sampled motions during SDS optimization. Without constraints, these motions may not align with observed evidence, leading to overly smoothed results that lack detail due to the mode-averaging effect.", + "bbox": [ + 212, + 537, + 785, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose COIN, a hybrid COntrol-INpainting score distillation sampling method to address the aforementioned limitations of vanilla SDS. First, we use the partially observed human motion from the video as control signals to guide motion sampling. To tackle noisy observations which may be out of distribution for the prior, we propose a dynamic controlled sampling technique that iteratively refines the observed motions and updates the control signals to ensure effective distillation from the motion prior. Second, to further improve the consistency of the sampled motions, we also develop a novel soft inpainting strategy. We automatically identify the high-confidence regions of the initial predicted global motion from the video and use them as soft constraints during optimization. Concretely, we sample less confident regions from scratch using the", + "bbox": [ + 212, + 672, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "L. Jiefeng et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Keywords: Global Human Motion Estimation $\\cdot$ Human Motion Prior $\\cdot$ Score Distillation Sampling", + "bbox": [ + 261, + 147, + 738, + 174 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "motion model, while the confident regions are only slightly refined. This ensures that the reconstructed motions do not deviate from the available observations. Our new SDS formulation is used to jointly optimize the human and camera motion by finding the most plausible solution that explains the observed evidence. Finally, to prevent catastrophic failure in cases where the initial body motion fails significantly, we propose a human-scene relation loss to consider the human-scene depth relations. This novel loss provides complementary information to the human motion prior by using local motion and scene features. It regularizes the camera scale by enforcing consistency among the human motion, camera motion, and scene features.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We benchmark our approach on the synthetic HCM [41] dataset and the real-world RICH [25] and EMDB [37] datasets. We demonstrate that our approach significantly outperforms the state-of-the-art methods in terms of human motion estimation and camera motion estimation. Overall, the contributions of this paper can be summarized as follows:", + "bbox": [ + 212, + 297, + 787, + 375 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel control-inpainting motion prior specifically designed for global human motion estimation, which enhances score distillation sampling with dynamic control and soft inpainting to reconstruct well-aligned, consistent, and high-quality motions from video observations.", + "- We propose a new human-scene relation loss to resolve the scale ambiguity of the camera motion by enforcing consistency among the human motion, camera motion, and scene features.", + "- Our approach significantly outperforms the state-of-the-art methods in terms of human motion estimation and camera motion estimation on both synthetic and real-world datasets. In terms of global human motion estimation in world space, we outperform the state-of-the-art method PACE [41] by $44\\%$ and $33\\%$ on the HCM [41] and RICH [25] datasets, respectively. We also compare with the contemporary work WHAM [81] and outperform it by $49\\%$ and $7\\%$ on the RICH and EMDB datasets, respectively." + ], + "bbox": [ + 225, + 385, + 784, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 625, + 387, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Camera-Space Human Pose Estimation. Most existing works focus on root-relative local human pose estimation to bypass the difficulty in monocular depth estimation [1, 4, 6-8, 16, 32-34, 39, 42-46, 51, 54, 63, 64, 67, 68, 70, 76, 78, 84, 85, 88, 94, 97, 106, 110, 116, 118]. These methods ignore the position of the person in the camera coordinates. To overcome this limitation, recent methods estimate camera-space human poses by regression [28, 31, 48, 52, 62, 72, 75, 80, 93, 95, 107, 107, 109, 111] or optimization [59-61, 77, 108]. Physics-based constraints are widely used to ensure the plausibility of the estimated poses [10, 13, 27, 29, 80, 95, 105]. In addition to direct regression, heatmap-based representations have also been used to predict the absolute depths of multiple people [11, 87, 117]. A few methods improve absolution depth estimation by using predicted camera parameters [40, 49, 109] instead of the predefined focal length. Despite the promising results", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "COIN: Control-Inpainting Diffusion", + "bbox": [ + 491, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for camera-space pose estimation, how to decouple the camera movement and estimate global human motions is still an open problem.", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Monocular Global Human Pose Estimation. Recovering the global human motion from a monocular moving camera is challenging due to the entanglement of human and camera motions. To disentangle the camera movement, several methods use IMU sensors or pre-scanned environments to recover global human motions [17,20,57,69], which is impractical for large-scale adoption. Recent works use human motion priors [100] or physics-based constraints [47,55] to recover human motions from monocular videos, but do not consider background scene features, which limits performance on in-the-wild videos. Sun et al. [86] use optical flow as a motion cue to estimate the global motions. A contemporary work, WHAM [81], uses a lifting network to estimate global human motions from 2D keypoints and camera angular velocities. While these works can estimate accurate global human motions, they do not recover camera motions. To explicitly recover the camera motion, Liu et al. [52] use SLAM and convert the local pose from the camera to global coordinates. BodySLAM [22] jointly optimizes the human and camera motion using features of both humans and scenes. Along this line, SLAHMR [99] and PACE [41] use SLAM to initialize camera motions and optimize the camera using human motion priors [21,76]. However, these methods rely on the human motion priors to regularize the camera motion, which may lead to inaccurate camera motion when the human motion is not well initialized (as shown in Fig. 1). Such wrong camera trajectories will further affect the optimization of human motions. In contrast, our approach relies on the consistency among the local human motion, scene features, and the camera for optimization, which provides information that complements the human motion priors and enables accurate estimation of both human and camera motions.", + "bbox": [ + 215, + 205, + 785, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human Motion Priors. There are a significant amount of approaches proposed to study 3D human dynamics for motion prediction and synthesis [2, 3, 5, 12, 15, 18, 19, 23, 30, 36, 38, 50, 58, 71, 73, 92, 98, 101-103]. These learned human motion priors are used to help resolve pose ambiguity [21, 39, 76, 115] in human pose estimation. Recently, diffusion models [82] have also been used as priors for motion synthesis and infilling [26, 35, 91, 104, 113]. RoHM [114], adopts motion diffusion model to recover human motions from noisy and occluded input data. Xie et al. [96] use spatial control signals to guide motion generation. They focus on generating realistic human motion given clean spatial constraints. Müller et al. [65] build a diffusion model to learn the joint distribution over the poses of two people. They use the SDS loss to guide the generation of static poses. In contrast, we focus on dynamic human motions. We find adopting SDS directly for temporal human motions encounters the inconsistency issue. Therefore, to distill knowledge from the motion diffusion model, we propose a novel control inpainting SDS to generate high-quality and consistent motion that aligns with observed evidence.", + "bbox": [ + 215, + 598, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "L. Jiefeng et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/de1e8367b8e3c658791af30b4878cabc66b63e7163ab60536295821c1707354f.jpg", + "image_caption": [ + "Fig. 2: Overview. Given a video with a moving camera, we recover the global human motion $\\mathbf{H}$ and camera motion $\\mathcal{C}$ using an iterative optimization framework. We propose a novel Control-Inpainting SDS loss ( $\\mathcal{L}_{\\mathrm{COIN - SDS}}$ ) to leverage motion diffusion models as a prior. COIN-SDS is designed such that the sampled motions from the motion prior are consistent with video observations. We achieve this by controlling and constraining the sampling process of the motion diffusion model through novel control and soft-inpainting branches. We also propose a novel human-scene relation loss ( $\\mathcal{L}_{\\mathrm{HSR}}$ ) to encourage consistency among the human motion, camera motion, and scene features." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 782, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 498, + 328, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall framework of COIN is illustrated in Fig. 2. Given an in-the-wild RGB video with $T$ frames captured by a dynamic camera, our goal is to estimate both the global human motion $\\mathbf{H} = \\{\\mathbf{h}^{(1)},\\mathbf{h}^{(2)},\\dots,\\mathbf{h}^{(T)}\\}$ and the camera motion $\\mathcal{C} = \\{\\mathbf{c}^{(1)},\\mathbf{c}^{(2)},\\dots,\\mathbf{c}^{(T)}\\}$ in a global world coordinate system. We use off-the-shelf 3D human pose and shape estimation method HybrIK [48] to obtain per-frame initial SMPL parameters in the camera space and DROID-SLAM [89] to obtain the initial per-frame camera-to-world transforms. We convert the local human motion to the world coordinates with the estimated camera. However, because the camera trajectories from SLAM are up to an unknown scale, the initial global human motion will abnormally drift and float in the world space. To resolve the scale ambiguity and place the person in the correct global position, we jointly optimize the human and camera motion to minimize the discrepancy between the observed evidence and the estimated motion, while maintaining the plausibility of the human motion with a diffusion prior through the proposed control-inpainting SDS.", + "bbox": [ + 212, + 529, + 787, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motion Representation. The camera motion is represented by the trajectory $\\mathcal{C} = \\{(\\mathbf{R}^{(i)},\\mathbf{t}^{(i)})\\}_{i = 1}^{T}$ , where $[\\mathbf{R}^{(i)},\\mathbf{t}^{(i)}]$ is the camera pose at the $i$ -th frame, consisting of the rotation matrix $\\mathbf{R}^{(i)}\\in \\mathbb{R}^{3\\times 3}$ and the translation vector $\\mathbf{t}^{(i)}\\in \\mathbb{R}^3$ . The human motion is represented by the human trajectory $\\mathbf{H} = \\{\\mathbf{h}^{(i)}\\}_{i = 1}^{T}$ where $\\mathbf{h}^{(i)} = [\\tau^{(i)},\\varPhi^{(i)},\\theta^{(i)},f^{(i)},\\beta ]$ is the human pose at the $i$ -th frame, consisting", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "COIN: Control-Inpainting Diffusion", + "bbox": [ + 491, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of the global translation $\\tau^{(i)}\\in \\mathbb{R}^3$ , global orientation $\\varPhi^{(i)}\\in \\mathbb{R}^3$ , body pose parameters $\\theta^{(i)}\\in \\mathbb{R}^{23\\times 3}$ , foot contact labels $f\\in \\{0,1\\} ^4$ , and the body shape parameters $\\beta \\in \\mathbb{R}^{10}$ . We use the SMPL model [53] to represent the human pose and shape. Before introducing our approach, we first revisit the formulation and drawbacks of SDS in Sec. 3.1. Then we introduce the proposed control-inpainting SDS in Sec. 3.2. Finally, we present the global optimization pipeline with the proposed human-scene interaction loss in Sec. 3.3.", + "bbox": [ + 212, + 145, + 787, + 252 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Revisiting SDS", + "text_level": 1, + "bbox": [ + 215, + 267, + 388, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Score Distillation Sampling (SDS) was first introduced to distill 3D assets from pre-trained 2D text-to-image diffusion models [74]. It exploits the knowledge from the diffusion models by seeking modes for the conditional distribution in the DDPM latent space to optimize the 3D scene representation. Similarly, we can optimize global human motion by distilling knowledge from a pre-trained motion diffusion model.", + "bbox": [ + 212, + 291, + 784, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given an global human motion $\\mathbf{H}$ , the marginal distribution of noisy latent $\\mathbf{H}_t$ at timestep $t \\in \\mathcal{U}(0,1)$ is defined as:", + "bbox": [ + 214, + 382, + 785, + 412 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nq (\\mathbf {H} _ {t} | \\mathbf {H}) = \\mathcal {N} (\\mathbf {H} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {H}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 369, + 422, + 785, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_t\\in (0,1)$ is a hyperparameter controlled by the variance schedule of the diffusion model. SDS adopts the pre-trained diffusion model $\\mathcal{D}_{\\phi}(\\mathbf{H}_t,t,y)$ , which takes in $\\mathbf{H}_t$ and is used to model the conditional density of the human motion, where $\\phi$ are the parameters of the diffusion model and $y$ is the condition. Then, SDS aims to distill global human motion $\\mathbf{H}$ via seeking modes of the learned conditional density, which can be achieved by a weighted denoising score matching objective:", + "bbox": [ + 214, + 450, + 787, + 556 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\mathbf {H}} {\\min } \\mathcal {L} _ {\\mathrm {S D S}} := \\mathbb {E} _ {t, \\epsilon} \\left[ \\omega (t) \\| \\epsilon_ {\\phi} ^ {t} - \\epsilon \\| _ {2} ^ {2} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 566, + 785, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\phi}^{t}$ is the predicted denoising direction from the diffusion model, $\\mathbf{H}_t \\sim q(\\mathbf{H}_t|\\mathbf{H})$ is sampled using the reparameterization trick, $\\epsilon$ is the corresponding sampled noise, and $\\omega(t)$ is a weighting function that depends on the timestep $t$ .", + "bbox": [ + 214, + 599, + 785, + 646 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To clearly review the effect of SDS, we can reparameterize Eq. 2 as:", + "bbox": [ + 238, + 646, + 728, + 660 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {H}} \\mathcal {L} _ {\\mathrm {S D S}} := \\mathbb {E} _ {t} \\left[ \\frac {\\omega (t) \\sqrt {\\bar {\\alpha} _ {t}}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\left\\| \\mathbf {H} - \\hat {\\mathbf {H}} _ {0} ^ {t} \\right\\| _ {2} ^ {2} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 670, + 785, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 215, + 714, + 263, + 726 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {H}} _ {0} ^ {t} = \\frac {\\mathbf {H} _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon} _ {\\phi} ^ {t}}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 724, + 785, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on this reparameterization, we can see that the SDS objective is to minimize the discrepancy between global human motion $\\mathbf{H}$ and the denoised global human motion $\\hat{\\mathbf{H}}_0^t$ from the motion diffusion model in a single step. The denoised motion $\\hat{\\mathbf{H}}_0^t$ serves as the pseudo ground truth. However, at each optimization step, we randomly sample $t$ and $\\epsilon$ to generate the noisy latent $\\mathbf{H}_t$ , and we found", + "bbox": [ + 214, + 763, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "L. Jiefeng et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the pre-trained diffusion model is sensitive to the input. Minor fluctuations in the input latent would substantially change the denoised motion, which leads to inconsistency in $\\hat{\\mathbf{H}}_0^t$ across different time steps.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Although randomness can help generate diverse plausible motions to infer occluded regions and unknown information, we do not need it for well-observed regions, such as simple body poses in a clean background. Such randomness in the denoising steps makes the generated $\\hat{\\mathbf{H}}_0^t$ difficult to align with the local 2D observations and results in wrong global human motion. Moreover, this pseudo ground truth $\\hat{\\mathbf{H}}_0^t$ is generated from only a single denoising step, where the diffusion models may not produce high-quality motions, resulting in foot sliding and floating. Although sampling with a smaller timestep $t$ can alleviate these issues, the initial motion is usually inaccurate and the denoiser is not able to remove artifacts with a small $t$ . To exploit the knowledge of the motion diffusion model and denoise the initial motion, we must allow the SDS to sample with a larger timestep $t$ while maintaining high quality, consistency, and alignment with the local 2D observations.", + "bbox": [ + 212, + 191, + 787, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Control-Inpainting SDS", + "text_level": 1, + "bbox": [ + 215, + 404, + 455, + 419 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The limitations of SDS originate from the randomness and inconsistency of the denoised motion $\\hat{\\mathbf{H}}_0^t$ , which serves as the pseudo ground truth in the objective function. To overcome this issue, we propose a novel COntrol-INpainting SDS (COIN-SDS) to generate high-quality and consistent pseudo-ground-truth motions. Our solution has three key ingredients (shown in Alg. 1). First, to achieve high-quality motions, we seek to produce the pseudo ground truth with multiple DDIM denoising steps. Second, to encourage consistent motions, we propose to use partially observed evidence from the video as a control signal to dynamically guide the diffusion model and align the generated motions with the observations. Third, to further align the motion with observed regions, we propose a soft inpainting strategy within the denoising process.", + "bbox": [ + 212, + 430, + 787, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multiple Denoising Steps. Intuitively, to obtain high-quality pseudo ground truth for SDS, we can replace the single-step denoised motion $\\hat{\\mathbf{H}}_0^t$ with a multi-step one $\\tilde{\\mathbf{H}}_0^t \\coloneqq \\tilde{\\mathbf{H}}_0$ , following the multi-step DDIM denoising process [83]:", + "bbox": [ + 212, + 606, + 785, + 652 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {H}} _ {t - \\Delta t} = \\sqrt {\\bar {\\alpha} _ {t - \\Delta t}} \\cdot \\hat {\\mathbf {H}} _ {0} ^ {t} + \\sqrt {1 - \\bar {\\alpha} _ {t - \\Delta t}} \\cdot \\boldsymbol {\\epsilon} _ {\\phi} ^ {t}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 662, + 784, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "until $\\tilde{\\mathbf{H}}_0 = \\tilde{\\mathbf{H}}_{t - \\Delta t}$ is obtained. By replacing $\\tilde{\\mathbf{H}}_0^t$ in Eq. 3 with $\\tilde{\\mathbf{H}}_0$ , we can obtain a new objective for SDS:", + "bbox": [ + 212, + 693, + 785, + 724 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {H}} \\mathcal {L} _ {\\mathrm {S D S}} := \\mathbb {E} _ {t} \\left[ \\frac {\\omega (t) \\sqrt {\\bar {\\alpha} _ {t}}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\left\\| \\mathbf {H} - \\tilde {\\mathbf {H}} _ {0} \\right\\| _ {2} ^ {2} \\right]. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 737, + 784, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Although the multi-step denoising process can produce high-quality pseudo ground truth, it is computationally expensive to perform multiple denoising steps during optimization, which limits the practicality of increasing the number of denoising steps. In our experiments, we find that using 10 denoising steps is sufficient to produce high-quality pseudo ground truth.", + "bbox": [ + 212, + 779, + 795, + 854 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "COIN: Control-Inpainting Diffusion", + "bbox": [ + 491, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: COIN-SDS" + ], + "code_body": "Input: Latest human motion H, confidence score S, visible mask M \nOutput: $\\mathcal{L}_{\\mathrm{COIN - SDS}}$ \n1 Sample: $\\tilde{\\mathbf{H}}_t \\sim \\mathcal{N}(\\mathbf{H}_t; \\sqrt{\\bar{\\alpha}_t}\\mathbf{H}, (1 - \\bar{\\alpha}_t)\\mathbf{I}), t \\sim \\mathcal{U}(0,1)$ ; \n2 for $\\bar{t} = [t, t - \\Delta t, \\dots, \\Delta t]$ do // multi-step DDIM denoising \n3 $\\tilde{\\mathbf{H}}_0^{\\bar{t},\\mathrm{known}} \\gets \\mathbf{H}$ ; \n4 $\\tilde{\\mathbf{H}}_0^{\\bar{t},\\mathrm{un-known}} \\gets \\mathcal{D}_{\\phi,\\phi_c}(\\tilde{\\mathbf{H}}_{\\bar{t}}, \\bar{t}, \\mathbf{H} \\odot \\mathbf{M})$ ; // controlled denoising \n5 $\\tilde{\\mathbf{M}} \\gets w(\\bar{t}) * \\mathbf{S} \\odot \\mathbf{M}$ ; \n6 $\\tilde{\\mathbf{H}}_0^{\\bar{t}} \\gets \\tilde{\\mathbf{M}} \\odot \\tilde{\\mathbf{H}}_0^{\\bar{t},\\mathrm{known}} + (1 - \\tilde{\\mathbf{M}}) \\odot \\tilde{\\mathbf{H}}_0^{\\bar{t},\\mathrm{unknown}}$ ; // soft inpainting \n7 $\\epsilon_{\\phi}^{\\bar{t}} \\gets \\frac{\\tilde{\\mathbf{H}}_{\\bar{t}} - \\sqrt{\\bar{\\alpha}_{\\bar{t}}} \\tilde{\\mathbf{H}}_0^{\\bar{t}}}{\\sqrt{1 - \\bar{\\alpha}_{\\bar{t}}}}$ ; \n8 $\\tilde{\\mathbf{H}}_{\\bar{t} -\\Delta t} \\gets \\sqrt{\\bar{\\alpha}_{\\bar{t} -\\Delta t}} \\cdot \\tilde{\\mathbf{H}}_0^{\\bar{t}} + \\sqrt{1 - \\bar{\\alpha}_{\\bar{t} -\\Delta t}} \\cdot \\epsilon_{\\phi}^{\\bar{t}}$ ; // update latent motion \n9 $\\mathcal{L}_{\\mathrm{COIN - SDS}} = \\frac{\\omega(t) \\sqrt{\\bar{\\alpha}_t}}{\\sqrt{1 - \\bar{\\alpha}_t}}\\left\\| \\mathbf{H} - \\tilde{\\mathbf{H}}_0 \\right\\|_2^2$", + "bbox": [ + 225, + 166, + 764, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dynamic Controlled Sampling. To generate consistent motions that are aligned with the observed evidence, we propose to attach a control branch $\\phi_c$ to the pre-trained diffusion model $\\mathcal{D}_{\\phi}$ to guide the motion generation. Given a latent motion $\\tilde{\\mathbf{H}}_t$ , control signal $\\mathbf{c}$ , and the visible mask $\\mathbf{M}$ , we train a controlled denoiser $\\mathcal{D}_{\\phi,\\phi_c}$ to generate intermediate denoised motion $\\tilde{\\mathbf{H}}_0^t$ for DDIM denoising:", + "bbox": [ + 212, + 364, + 787, + 444 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {H}} _ {0} ^ {t} = \\mathcal {D} _ {\\phi , \\phi_ {c}} \\left(\\tilde {\\mathbf {H}} _ {t}, t, \\mathbf {c} \\odot \\mathbf {M}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 459, + 785, + 478 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{c}$ and $\\mathbf{M}$ are the same size as the motion, $\\mathbf{M}$ is a binary mask with ones in observed pose dimensions, and $\\odot$ denotes the element-wise multiplication. During training, we synthesize noise and occlusions by randomly adding Gaussian noise to the control signal $\\mathbf{c}$ and randomly masking the pose and trajectory dimensions in $\\mathbf{M}$ . Details are provided in the supplementary document.", + "bbox": [ + 212, + 488, + 787, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When using the denoiser in the optimization stage, instead of always using the initial noisy estimation from HybrIK [48] as the fixed control signal, we propose to use a dynamic control strategy. Specifically, we use the optimized human motion from the previous iteration as the control signal, i.e., $\\mathbf{c} = \\mathbf{H}$ . This strategy prevents performance degradation due to inaccurate initializations. A better control signal can guide the model to generate a more plausible pseudo motion, which in turn helps to optimize the global human motion and provides a control signal that better aligns with the input videos. Such self-evolving control signals can help to generate well-aligned global human motions.", + "bbox": [ + 212, + 566, + 787, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The pre-trained motion diffusion model adopts a transformer encoder structure like [91]. We follow ControlNet [112] to encode the control signals and guide the denoiser output. We create a trainable copy of 4 encoding blocks of the pretrained motion diffusion model followed with zero convolutions. The input to the control branch is the concatenation of the latent and the control signals. The AMASS [56] dataset is used to train the motion diffusion model. The finetuning of the controlled denoiser is computationally efficient since the pre-trained branch is frozen and only the control branch is trained. See the supplementary document for more details on the architecture and training settings.", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "L. Jiefeng et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Soft Inpainting. While guiding motion generation encourages outputs to align with the conditions, it is often not strong enough. We further seek to improve the consistency by masking the known regions and inpainting the unknown regions. Given a binary mask $\\mathbf{M}$ that indicates the observed and unobserved regions, we can use the diffusion model to generate the inpainted motion $\\tilde{\\mathbf{H}}_0^t$ following the DDIM denoising process:", + "bbox": [ + 214, + 146, + 782, + 237 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {H}} _ {0} ^ {t, \\text {k n o w n}} = \\mathbf {H}, \\tag {8a}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 244, + 782, + 261 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {H}} _ {0} ^ {t, \\text {u n k n o w n}} = \\mathcal {D} _ {\\phi , \\phi_ {c}} \\left(\\tilde {\\mathbf {H}} _ {t}, t, \\mathbf {H} \\odot \\mathbf {M}\\right), \\tag {8b}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 266, + 782, + 282 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {H}} _ {0} ^ {t} = \\mathbf {M} \\odot \\tilde {\\mathbf {H}} _ {0} ^ {t, \\text {k n o w n}} + (1 - \\mathbf {M}) \\odot \\tilde {\\mathbf {H}} _ {0} ^ {t, \\text {u n k n o w n}}. \\tag {8c}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 287, + 782, + 305 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Thus, the known regions are overwritten with the observations, while the unknown regions are sampled from the diffusion model. However, the above formulation keeps the observed parts unchanged during the denoising process. In practice, the observed parts can be noisy and not perfect. We still want the diffusion model to refine the observed parts but not change them significantly.", + "bbox": [ + 214, + 311, + 782, + 387 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Here, we present a soft inpainting strategy to infill the unobserved regions while refining the observed regions by dynamically reweighting the denoised direction from the diffusion model. Specifically, instead of using a binary mask, we adopt a continuous mask $\\tilde{\\mathbf{M}}$ depending on both the confidence score of the observations $\\mathbf{S}$ and the denoising time step $t$ :", + "bbox": [ + 214, + 388, + 782, + 463 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} = w (t) * \\mathbf {S} \\odot \\mathbf {M}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 472, + 782, + 487 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where we set $w(t) = \\max(0, \\frac{t - 0.5}{0.5})$ to linearly decrease the weight of the observations as the denoising time step decreases. As the time step decreases, the denoising process will be more deterministic and the model will be more certain of the generated motions.", + "bbox": [ + 214, + 494, + 782, + 554 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Combining the three components of our solution, the final objective for COIN-SDS is formulated as:", + "bbox": [ + 214, + 556, + 782, + 585 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {H}} \\mathcal {L} _ {\\mathrm {C O I N - S D S}} := \\mathbb {E} _ {t} \\left[ \\frac {\\omega (t) \\sqrt {\\bar {\\alpha} _ {t}}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\left\\| \\mathbf {H} - \\tilde {\\mathbf {H}} _ {0} (\\mathbf {H}, \\mathbf {M}, \\mathbf {S}, t) \\right\\| _ {2} ^ {2} \\right]. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 299, + 590, + 782, + 625 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We summarize COIN in Alg. 1.", + "bbox": [ + 215, + 630, + 442, + 643 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.3 Global Optimization", + "text_level": 1, + "bbox": [ + 215, + 652, + 429, + 667 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Here we present the overall optimization pipeline for the joint estimation of global human and camera motion with the proposed COIN-SDS loss. Note that we use SLAM to initialize the camera poses, which is scale-ambiguous. Therefore, we need to jointly optimize the camera scale $s$ with the human and camera motions. Furthermore, the SLAM method assumes the camera in the first frame to be at the origin. To put the human motion in the correct positions, we follow PACE [41] and also optimize the camera height $h_0$ and the orientation $R_0$ for the first frame. The global human motions are initialized by the estimated local motions and the camera poses. The overall optimization objective is:", + "bbox": [ + 214, + 675, + 782, + 811 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {H}, \\mathcal {C}, s, h _ {0}, R _ {0}, \\beta} \\mathcal {L} _ {\\text {b o d y}} + \\mathcal {L} _ {\\text {C O I N - S D S}} + \\mathcal {L} _ {\\text {H S R}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 821, + 782, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "COIN: Control-Inpainting Diffusion", + "bbox": [ + 491, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 215, + 146, + 264, + 159 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {b o d y}} = \\mathcal {L} _ {2 \\mathrm {D}} + \\mathcal {L} _ {3 \\mathrm {D}} + \\mathcal {L} _ {\\beta} + \\mathcal {L} _ {\\text {s m o o t h}} + \\mathcal {L} _ {\\text {c o n t a c t}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 161, + 785, + 176 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{2\\mathrm{D}}$ measures the 2D reprojection error between the projected 2D body joints of the estimated human motion and the detected 2D keypoints from an off-the-shelf 2D joint detector [9]. $\\mathcal{L}_{3\\mathrm{D}}$ measures the distance between the estimated local 3D joints and the detected 3D joints from an off-the-shelf 3D joint detector [79]. $\\mathcal{L}_{\\beta}$ is the shape regularization loss. $\\mathcal{L}_{\\mathrm{smooth}}$ is the temporal smoothness loss. $\\mathcal{L}_{\\mathrm{contact}}$ is the foot contact loss to encourage zero velocities for contact joints. The contact labels are obtained from the pseudo ground truth motion from COIN-SDS. Please refer to the supplementary document for more details.", + "bbox": [ + 212, + 183, + 787, + 305 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Human-scene Relation Loss. The camera trajectories recovered from SLAM are scale-ambiguous. Previous works [41,99] optimize the camera scale by projecting the global human motion to the camera space using the camera poses and minimizing the reprojection error. However, such a method entirely relies on the global human motions, which is in turn affected by the camera scale. If the human motion is not initialized well, the camera scale will also be inaccurate. To solve this problem, instead of the global human motions, we propose a new human-scene relation loss that uses the depth relation between the human and scene in the camera space, which disentangles the effect of the camera itself.", + "bbox": [ + 212, + 309, + 787, + 444 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Specifically, we use the point cloud of the scene recovered by SLAM as a constraint. First of all, the scale of the scene point cloud is the same as the camera scale, so optimizing the scene scale is equivalent to optimizing the camera scale. Second, the scene points that are projected onto the visible vertices of the body mesh should be occluded by the person. Otherwise, the corresponding body parts are invisible. Therefore, we can constrain the depth of the occluded scene points to be larger than the depth of the human body vertices. While finding the corresponding body vertices for each scene point is time-consuming, we propose to use the depth of its nearest body joint as a proxy. Given the scene point cloud $\\mathcal{P}$ and the camera scale $s$ , the human-scene relation loss is formulated as:", + "bbox": [ + 212, + 445, + 789, + 595 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {H S R}} = - \\frac {1}{| \\mathcal {P} |} \\sum_ {i = 1} ^ {T} \\sum_ {p \\in \\mathcal {P} ^ {*}} \\min (0, \\mathcal {T} ^ {(i)} (p) _ {z} - j ^ {(i)} (p) _ {z}) \\cdot \\mathbb {1} (\\mathcal {T} ^ {(i)} (p) \\text {i s i n v i s i b l e}), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 606, + 785, + 650 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\mathcal{P}^* = \\mathcal{P}*s$ is the scaled point cloud of the scene, $\\mathcal{T}^{(i)}(p) = \\mathbf{R}^{(i)}p + \\mathbf{t}^{(i)}$ is the transformed point in the $i$ -th frame, $j^{(i)}(p)$ is body joint that has the nearest 2D projection to the scene point $p$ in the $i$ -th frame, and $z$ denotes the depth of a given point. If the depth order is correct, i.e., $\\mathcal{T}^{(i)}(p)_z - j^{(i)}(p)_z > 0$ , the loss is zero. The proposed human-scene relation loss uses the relation between the local motions and the scene to alleviate the scale ambiguity. The depth relation regularizes consistency among humans, cameras, and scenes.", + "bbox": [ + 212, + 661, + 787, + 768 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 777, + 375, + 792 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. We perform experiments on three human motion datasets. First is the real-world dataset RICH [25]. We follow previous works [41,81,100] to assess the performance of global human motion estimation using this dataset. The second", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "L. Jiefeng et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/bb3c11c7ab28c1a80f2f087f56c70678ef27ee13a680353a92ec7222474283d0.jpg", + "table_caption": [ + "Table 1: Global human motion estimation on the RICH dataset." + ], + "table_footnote": [], + "table_body": "
MethodPA-MPJPE ↓W-MPJPE ↓WA-MPJPE ↓W-RJE ↓ACCEL ↓
HybrIK [48] + SLAM [89]46.71073.1404.41166.220.2
GLAMR [100]79.9653.7365.1646.6107.7
SLAHMR [99]52.5571.6323.7400.59.4
WHAM [81]46.2497.6272.7478.26.7
PACE [41]49.3380.0197.2370.88.8
Guided Sampling132.81384.6502.21440.924.2
Noise Optimization66.7414.8195.3429.28.4
Vanilla SDS78.81453.5497.21458.012.7
COIN w/o Controlled Sampling44.0825.0291.8848.510.8
COIN w/o Dynamic Control49.5293.8180.6299.18.6
COIN w/o Soft Inpainting47.6325.8196.0324.99.6
COIN w/o LHSR43.6273.0176.3281.18.2
COIN42.9254.5169.5249.97.5
", + "bbox": [ + 218, + 159, + 782, + 345 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "one is EMDB [37]. We follow previous works [81] to evaluate on a subset of EMDB for which they provide ground truth global motion with dynamic cameras. The third dataset is HCM [41], a synthetic dataset. Compared to real-world datasets, HCM contains more challenging camera motions. We follow previous work [41] to evaluate the global human motion and the camera motion using this dataset.", + "bbox": [ + 212, + 357, + 784, + 434 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Metrics. We report various metrics for both human and camera motion. For human motion, standard metrics W-MPJPE and WA-MPJPE are used to evaluate global motion, while PA-MPJPE evaluates local motion. We also include an ACCEL metric to measure the joint acceleration error. For evaluation on EMDB, we follow previous work [81] to split sequences into smaller chunks of 100 frames and align each output segment with the ground-truth data using the first two frames W-MPJPE100 or the entire segment WA-MPJPE100. Root Orientation Error (ROE in deg) and Root Translation Error (RTE in $m$ ) evaluate the error over the entire trajectory after aligning with the initial camera pose.", + "bbox": [ + 212, + 439, + 784, + 575 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For camera motion, we report the average translation error after scale alignment (ATE), without scale alignment (ATE-S), and the camera acceleration error (CAM ACCEL). ATE-S more accurately reflects inaccuracies in the captured scale of the camera.", + "bbox": [ + 212, + 575, + 784, + 636 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Baselines. As discussed in Sec. 2, there are different ways to use the pre-trained motion diffusion model as a motion prior. Here, we summarize the three main solutions and compare them with COIN in Sec. 4.2. (1) Guided Sampling, which embeds analytical guidance within the denoising procedure using objective functions, such as 2D projection and foot contact consistency. This does not suit our task because the camera trajectories are also unknown and guided gradients from the wrong camera will lead to unrealistic human motions. We need to optimize the human motion and camera motion simultaneously. (2) Noise Optimization, which represents the motion as latent noise and directly optimizes it. This is similar to other motion priors such as VAE [21]. At each optimization step, we need to calculate the gradients of the latent w.r.t. the generated motion, which is computationally expensive and we find the performance is not good enough. (3) Vanilla SDS, which removes our design and directly uses SDS for optimization.", + "bbox": [ + 212, + 643, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "COIN: Control-Inpainting Diffusion", + "bbox": [ + 491, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/726e55e2aae0e72f866febae22229525b4a2e2f154dabadd27badac029090694.jpg", + "table_caption": [ + "Table 2: Global human motion estimation on the EMDB dataset." + ], + "table_footnote": [], + "table_body": "
MethodPA-MPJPE ↓W-MPJPE100 ↓WA-MPJPE100 ↓RTE ↓ROE ↓
HMR2.0 [14] + DPVO [90]49.62320.9662.917.544.4
GLAMR [100]56.0756.1286.216.774.9
TRACE [86]58.02244.9544.118.972.7
SLAHMR [99]61.5807.4336.913.867.9
WHAM [81]41.9439.2166.18.436.3
COIN32.7407.3152.83.534.1
", + "bbox": [ + 218, + 159, + 782, + 262 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/f760c1e05ff5c04a01aeee5aeaba45c7d699d9cf9bc0b130a9d7bc64a0f0ae4d.jpg", + "table_caption": [ + "Table 3: Global human motion estimation on the HCM dataset." + ], + "table_footnote": [], + "table_body": "
MethodPA-MPJE ↓W-MPJE ↓WA-MPJE ↓W-RJE ↓ACCEL ↓
HybrIK [48] + SLAM [89]67.61137.3780.31100.951.3
GLAMR [100]86.01977.6653.81958.033.4
SLAHMR [99]69.9888.9483.5862.214.9
PACE [41]65.3861.2478.3839.516.7
WHAM [81]47.9588.9279.3579.213.1
COIN45.5479.9212.1470.710.1
", + "bbox": [ + 220, + 306, + 782, + 411 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.1 Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 215, + 416, + 620, + 430 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Human Motion Estimation. We compare COIN against state-of-the-art methods on the RICH, EMDB, and HCM datasets. Quantitative results are shown in Tabs. 1, 2, and 3. We observe that COIN significantly outperforms the state-of-the-art methods on all datasets. On the RICH dataset, COIN outperforms the state-of-the-art method, PACE [41], by $125.5\\mathrm{mm}$ in terms of W-MPJPE, showing $33.0\\%$ relative improvement. On the EMDB and HCM datasets, COIN shows $29.1\\mathrm{mm}$ and $109.0\\mathrm{mm}$ improvement in terms of W-MPJPE, respectively. Qualitative comparisons with state-of-the-art methods, PACE [41] and WHAM [81], are shown in Figs. 1 and 3. More qualitative results are shown in the supplementary video.", + "bbox": [ + 212, + 444, + 787, + 595 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "COIN not only improves global motion, but also improves local body motion. In terms of PA-MPJPE, COIN shows $3.3\\mathrm{mm}$ , $9.2\\mathrm{mm}$ , and $19.8\\mathrm{mm}$ improvement on the RICH, EMDB, and HCM datasets, respectively. This demonstrates that COIN is able to distill high-quality motion priors from the diffusion model and help both local and global motion estimation. Regarding the joint acceleration error, COIN is $0.8\\mathrm{mm/s}^2$ higher than WHAM [81]. Note that the joint acceleration error reflects the smoothness. Human motion can be over-smoothed but not accurate, hence it is important to look at acceleration error in conjunction with other metrics.", + "bbox": [ + 212, + 598, + 787, + 733 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Camera Motion Estimation. We further evaluate the performance of COIN on camera motion estimation on the HCM dataset. Quantitative results are shown in Tab. 4. COIN substantially surpasses the state-of-the-art camera motion estimation methods. Specifically, COIN reduces the absolute camera translation error, ATE-S by $73.8\\mathrm{mm}$ . This demonstrates that COIN is able to disentangle human and camera motions and accurately estimate the camera motion.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "L. Jiefeng et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/7b4406d857db9a8537e5a8da0f2d23435b968e5fe2637b321c66001bf9c6fbe9.jpg", + "table_caption": [ + "Table 4: Camera motion estimation on the HCM dataset." + ], + "table_footnote": [], + "table_body": "
MethodATE ↓ATE-S ↓CAMACCEL ↓
HybrIK [48] + SLAM [89]155.81670.717.1
GLAMR [100]1295.21714.6282.9
SLAHMR [99]155.8506.517.6
PACE [41]137.5459.716.2
Guided Sampling335.6992.415.6
Noise Optimization206.0500.912.3
Vanilla SDS306.4656.413.7
COIN w/o Controlled Sampling299.6553.814.0
COIN w/o Dynamic Control149.8397.711.3
COIN w/o Soft Inpainting167.7423.811.4
COIN w/o LHSR147.8402.112.0
COIN135.3385.911.3
", + "bbox": [ + 263, + 160, + 738, + 377 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.2 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 388, + 388, + 404 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we conduct ablation studies on the RICH and HCM datasets to evaluate the impact of each component on human and camera motions, respectively. More comparisons on other datasets are provided in the supplementary document.", + "bbox": [ + 212, + 412, + 784, + 470 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Baselines with Motion Diffusion Model. We first compare the aforementioned baselines with COIN. To use guided sampling in our tasks, we jointly update the camera poses using the gradients from the objective function during denoising. Quantitative results of human and camera motion estimation are shown in Tabs. 1 and 4, respectively. We observe that COIN outperforms all the baselines in terms of all metrics. As expected, guided sampling shows a terrible performance because it is not able to accurately estimate the camera trajectories. Noise optimization is better than vanilla SDS but worse than COIN, which generates consistent motion priors. Vanilla SDS is replacing $\\mathcal{L}_{\\mathrm{COIN}}$ with the SDS loss and keeping the rest of the settings the same. These results demonstrate the effectiveness of COIN over other motion diffusion baselines.", + "bbox": [ + 212, + 479, + 787, + 645 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Impact of Dynamic Controlled Sampling. To study the effectiveness of controlled sampling, we compare COIN with and without using the controlled denoiser. Quantitative results are shown in Tabs. 1 and 4. We observe that controlled sampling significantly improves the performance of COIN. Specifically, COIN with controlled sampling reduces the W-MPJPE and ATE-S by $570.5\\mathrm{mm}$ and $167.9\\mathrm{mm}$ , showing $69.3\\%$ and $30.3\\%$ relative improvement, respectively. This demonstrates that controlled sampling is able to generate high-quality motion priors that align with the observed evidence and help improve human and camera motion estimation.", + "bbox": [ + 212, + 651, + 787, + 787 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Impact of Soft Inpainting. We further study the effectiveness of soft inpainting. Quantitative comparisons are shown in Tabs. 1 and 4. We observe that while soft inpainting also improves W-MPJPE, it is much more effective than", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "COIN: Control-Inpainting Diffusion", + "bbox": [ + 491, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3d507f94e555b157a84a398fb6034eb37a4e580bf6d730f47ec2d1f84b3455d2.jpg", + "image_caption": [ + "PACE (Human Motion) PACE (Camera Motion)" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 500, + 340 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/b2d1ac4a7282d5b7072db75d882ccf786ccc55ba582646a308f5bcb8bdb1e758.jpg", + "image_caption": [ + "WHAM", + "COIN (Human Motion) COIN (Camera Motion)", + "Fig. 3: Qualitative comparisons with state-of-the-art methods. PACE [41] fails to recover a correct trajectory (left). WHAM [81] estimates the wrong walking direction of the person (right). Our approach, COIN, recovers the human and camera motion accurately in both scenarios." + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 785, + 339 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "controlled sampling in terms of local body motions. Specifically, COIN with soft inpainting reduces the PA-MPJPE by $4.7 \\mathrm{~mm}$ .", + "bbox": [ + 212, + 416, + 785, + 446 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 465, + 359, + 479 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose COIN, a diffusion-based optimization framework for global human and camera motion estimation from dynamic cameras. We identify the inconsistency problem of distilling knowledge from the diffusion model with conventional SDS loss. To address this issue, COIN uses a controlled denoiser combined with soft inpainting to distill a high-quality, well-aligned, and consistent motion prior. To further address the scale ambiguity of the camera trajectory, we develop a novel human-scene relation loss that imposes consistency among the human motion, camera motion, and scene features. Comprehensive experiments on challenging synthetic and real-world datasets demonstrate the effectiveness of COIN, which outperforms the SOTA by a large margin in recovering accurate global human motion and camera motion.", + "bbox": [ + 212, + 492, + 787, + 657 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations and Future Works: While COIN is able to jointly optimize camera trajectories and global human motions, it requires initialization from SLAM. If the SLAM method fails catastrophically, COIN may not be effective. Additionally, we found COIN fails under severe occlusions where there are several unseen frames and the diffusion model cannot provide consistent guidance. Another limitation is that COIN is an optimization framework, so it is unsuitable for real-time applications. Since the denoising diffusion models have shown their power to model data distributions in many different domains, looking forward we can learn the joint distribution of humans and cameras with the diffusion model to additionally denoise the artifacts in camera motions. Such a joint distribution model also holds the potential of real-time human and camera motion estimation via guided sampling with few DDIM denoising steps.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "L. Jiefeng et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Akhter, I., Black, M.J.: Pose-conditioned joint angle limits for 3D human pose reconstruction. In: CVPR (2015) 3", + "2. Aksan, E., Kaufmann, M., Hilliges, O.: Structured prediction helps 3d human motion modelling. In: ICCV (2019) 4", + "3. Barsoum, E., Kender, J., Liu, Z.: Hp-gan: Probabilistic 3d human motion prediction via gan. In: CVPR Workshops (2018) 4", + "4. Bogo, F., Kanazawa, A., Lassner, C., Gehler, P., Romero, J., Black, M.J.: Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In: ECCV (2016) 3", + "5. Cao, Z., Gao, H., Mangalam, K., Cai, Q.Z., Vo, M., Malik, J.: Long-term human motion prediction with scene context. In: ECCV. pp. 387-404. Springer (2020) 4", + "6. Choi, H., Moon, G., Lee, K.M.: Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In: ECCV (2020) 3", + "7. Choi, H., Moon, G., Lee, K.M.: Beyond static features for temporally consistent 3d human pose and shape from a video. In: CVPR (2021) 3", + "8. Choutas, V., Pavlakos, G., Bolkart, T., Tzionas, D., Black, M.J.: Monocular expressive body regression through body-driven attention. In: ECCV (2020) 3", + "9. Contributors, M.: Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose (2020) 10", + "10. Dabral, R., Shimada, S., Jain, A., Theobalt, C., Golyanik, V.: Gravity-aware monocular 3d human-object reconstruction. In: ICCV (2021) 3", + "11. Fabbri, M., Lanzi, F., Calderara, S., Alletto, S., Cucchiara, R.: Compressed volumetric heatmaps for multi-person 3d pose estimation. In: CVPR (June 2020) 3", + "12. Fragkiadaki, K., Levine, S., Felsen, P., Malik, J.: Recurrent network models for human dynamics. In: ICCV (2015) 4", + "13. Gartner, E., Andriluka, M., Xu, H., Sminchisescu, C.: Trajectory optimization for physics-based reconstruction of 3d human pose from monocular video. In: CVPR (2022) 3", + "14. Goel, S., Pavlakos, G., Rajasegaran, J., Kanazawa, A., Malik, J.: Humans in 4d: Reconstructing and tracking humans with transformers. In: ICCV (2023) 12", + "15. Gopalakrishnan, A., Mali, A., Kifer, D., Giles, L., Ororbia, A.G.: A neural temporal model for human motion prediction. In: CVPR (2019) 4", + "16. Guler, R.A., Kokkinos, I.: HoloPose: Holistic 3d human reconstruction in-the-wild. In: CVPR (2019) 3", + "17. Guzov, V., Mir, A., Sattler, T., Pons-Moll, G.: Human POSEitioning system (HPS): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In: CVPR (2021) 4", + "18. Harvey, F.G., Yurick, M., Nowrouzezahrai, D., Pal, C.: Robust motion in betweening. ACM Transactions on Graphics (TOG) 39(4), 60-1 (2020) 4", + "19. Hassan, M., Ceylan, D., Villegas, R., Saito, J., Yang, J., Zhou, Y., Black, M.J.: Stochastic scene-aware motion prediction. In: ICCV (2021) 4", + "20. Hassan, M., Choutas, V., Tzionas, D., Black, M.J.: Resolving 3d human pose ambiguities with 3d scene constraints. In: ICCV. pp. 2282-2292 (2019) 4", + "21. He, C., Saito, J., Zachary, J., Rushmeier, H., Zhou, Y.: Nemf: Neural motion fields for kinematic animation. In: NeurIPS (2022) 2, 4, 11", + "22. Henning, D.F., Laidlow, T., Leutenegger, S.: Bodyslam: Joint camera localisation, mapping, and human motion tracking. In: ECCV (2022) 4" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "COIN: Control-Inpainting Diffusion", + "bbox": [ + 491, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "23. Hernandez, A., Gall, J., Moreno-Noguer, F.: Human motion prediction via spatiotemporal inpainting. In: CVPR (2019) 4", + "24. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPs (2020) 2", + "25. Huang, C.H.P., Yi, H., Höschle, M., Safroshkin, M., Alexiadis, T., Polikovsky, S., Scharstein, D., Black, M.J.: Capturing and inferring dense full-body human-scene contact. In: CVPR (2022) 3, 10", + "26. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: CVPR (2023) 4", + "27. Iqbal, U., Molchanov, P., Kautz, J.: Weakly-supervised 3d human pose learning via multi-view images in the wild. In: CVPR (2020) 3", + "28. Iqbal, U., Xie, K., Guo, Y., Kautz, J., Molchanov, P.: KAMA: 3D keypoint aware body mesh articulation. In: 3DV (2021) 3", + "29. Isogawa, M., Yuan, Y., O'Toole, M., Kitani, K.M.: Optical non-line-of-sight physics-based 3d human pose estimation. In: CVPR (2020) 3", + "30. Jain, A., Zamir, A.R., Savarese, S., Saxena, A.: Structural-rnn: Deep learning on spatio-temporal graphs. In: CVPR (2016) 4", + "31. Jiang, W., Kolotouros, N., Pavlakos, G., Zhou, X., Daniilidis, K.: Coherent reconstruction of multiple humans from a single image. In: CVPR (2020) 3", + "32. Joo, H., Neverova, N., Vedaldi, A.: Exemplar fine-tuning for 3d human pose fitting towards in-the-wild 3d human pose estimation. In: 3DV (2021) 3", + "33. Kanazawa, A., Black, M.J., Jacobs, D.W., Malik, J.: End-to-end recovery of human shape and pose. In: CVPR (2018) 3", + "34. Kanazawa, A., Zhang, J.Y., Felsen, P., Malik, J.: Learning 3d human dynamics from video. In: CVPR (2019) 3", + "35. Karunratanakul, K., Preechakul, K., Aksan, E., Beeler, T., Suwajanakorn, S., Tang, S.: Optimizing diffusion noise can serve as universal motion priors. arXiv preprint arXiv:2312.11994 (2023) 4", + "36. Kaufmann, M., Aksan, E., Song, J., Pece, F., Ziegler, R., Hilliges, O.: Convolutional autoencoders for human motion infilling. In: 3DV (2020) 4", + "37. Kaufmann, M., Song, J., Guo, C., Shen, K., Jiang, T., Tang, C., Zárate, J.J., Hilliges, O.: EMDB: The Electromagnetic Database of Global 3D Human Pose and Shape in the Wild. In: ICCV (2023) 3, 11", + "38. Khurana, T., Dave, A., Ramanan, D.: Detecting invisible people. In: ICCV. pp. 3174-3184 (2021) 4", + "39. Kocabas, M., Athanasiou, N., Black, M.J.: VIBE: Video inference for human body pose and shape estimation. In: CVPR (2020) 3, 4", + "40. Kocabas, M., Huang, C.H.P., Tesch, J., Müller, L., Hilliges, O., Black, M.J.: SPEC: Seeing people in the wild with an estimated camera. In: ICCV (2021) 3", + "41. Kocabas, M., Yuan, Y., Molchanov, P., Guo, Y., Black, M.J., Hilliges, O., Kautz, J., Iqbal, U.: PACE: Human and motion estimation from in-the-wild videos. In: 3DV (2024) 1, 2, 3, 4, 9, 10, 11, 12, 13, 14", + "42. Kolotouros, N., Pavlakos, G., Black, M.J., Daniilidis, K.: Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In: ICCV (2019) 3", + "43. Kolotouros, N., Pavlakos, G., Daniilidis, K.: Convolutional mesh regression for single-image human shape reconstruction. In: CVPR (2019) 3", + "44. Kolotouros, N., Pavlakos, G., Jayaraman, D., Daniilidis, K.: Probabilistic modeling for human mesh recovery. In: ICCV (2021) 3" + ], + "bbox": [ + 223, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "L. Jiefeng et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "45. Kundu, J.N., Rakesh, M., Jampani, V., Venkatesh, R.M., Babu1, R.V.: Appearance consensus driven self-supervised human mesh recovery. In: ECCV (2020) 3", + "46. Lassner, C., Romero, J., Kiefel, M., Bogo, F., Black, M.J., Gehler, P.V.: Unite the people: Closing the loop between 3D and 2D human representations. In: CVPR (2017) 3", + "47. Li, J., Bian, S., Xu, C., Liu, G., Yu, G., Lu, C.: D &d: Learning human dynamics from dynamic camera. In: ECCV (2022) 2, 4", + "48. Li, J., Xu, C., Chen, Z., Bian, S., Yang, L., Lu, C.: Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In: CVPR (2021) 3, 5, 8, 11, 12, 13", + "49. Li, Z., Liu, J., Zhang, Z., Xu, S., Yan, Y.: Cliff: Carrying location information in full frames into human pose and shape estimation. In: ECCV (2022) 3", + "50. Li, Z., Zhou, Y., Xiao, S., He, C., Huang, Z., Li, H.: Auto-conditioned recurrent networks for extended complex human motion synthesis. arXiv preprint arXiv:1707.05363 (2017) 4", + "51. Lin, K., Wang, L., Liu, Z.: End-to-end human pose and mesh reconstruction with transformers. In: CVPR (2021) 3", + "52. Liu, M., Yang, D., Zhang, Y., Cui, Z., Rehg, J.M., Tang, S.: 4d human body capture from egocentric video via 3d scene grounding. In: 3DV (2021) 3, 4", + "53. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. SIGGRAPH Asia 34(6), 248:1-248:16 (2015) 6", + "54. Luo, Z., Golestaneh, S.A., Kitani, K.M.: 3d human motion estimation via motion compression and refinement. In: ACCV (2020) 3", + "55. Luo, Z., Hachiuma, R., Yuan, Y., Kitani, K.: Dynamics-regulated kinematic policy for egocentric pose estimation. NeurIPS 34 (2021) 4", + "56. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: AMASS: Archive of motion capture as surface shapes. In: ICCV (2019) 8", + "57. von Marcard, T., Henschel, R., Black, M., Rosenhahn, B., Pons-Moll, G.: Recovering accurate 3d human pose in the wild using imus and a moving camera. In: ECCV (2018) 4", + "58. Martinez, J., Black, M.J., Romero, J.: On human motion prediction using recurrent neural networks. In: CVPR (2017) 4", + "59. Mehta, D., Rhodin, H., Casas, D., Fua, P., Sotnychenko, O., Xu, W., Theobalt, C.: Monocular 3d human pose estimation in the wild using improved cnn supervision. In: 3DV (2017) 3", + "60. Mehta, D., Sotnychenko, O., Mueller, F., Xu, W., Elgharib, M., Fua, P., Seidel, H.P., Rhodin, H., Pons-Moll, G., Theobalt, C.: XNect: Real-time multi-person 3D motion capture with a single RGB camera. In: SIGGRAPH (2020) 3", + "61. Mehta, D., Sridhar, S., Sotnychenko, O., Rhodin, H., Shafiei, M., Seidel, H., Xu, W., Casas, D., Theobalt, C.: VNect: Real-time 3D human pose estimation with a single RGB camera. In: SIGGRAPH (2017) 3", + "62. Moon, G., Chang, J.Y., Lee, K.M.: Camera distance-aware top-down approach for 3d multi-person pose estimation from a single rgb image. In: ICCV (2019) 3", + "63. Moon, G., Lee, K.M.: I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In: ECCV (2020) 3", + "64. Müller, L., Osman, A.A.A., Tang, S., Huang, C.H.P., Black, M.J.: On self contact and human pose. In: CVPR (2021) 3" + ], + "bbox": [ + 225, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "COIN: Control-Inpainting Diffusion", + "bbox": [ + 491, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "65. Müller, L., Ye, V., Pavlakos, G., Black, M., Kanazawa, A.: Generative proxemics: A prior for 3d social interaction from images. arXiv preprint arXiv:2306.09337 (2023) 4", + "66. Opensfm - a structure from motion library. https://github.com/mapillary/OpenSfM (2021), https://github.com/mapillary/OpenSfM 2", + "67. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: CVPR (2019) 3", + "68. Pavlakos, G., Kolotouros, N., Daniilidis, K.: Texturepose: Supervising human mesh estimation with texture consistency. In: ICCV (2019) 3", + "69. Pavlakos, G., Weber, E., Tancik, M., Kanazawa, A.: The one where they reconstructed 3d humans and environments in tv shows. In: ECCV (2022) 4", + "70. Pavlakos, G., Zhu, L., Zhou, X., Daniilidis, K.: Learning to estimate 3D human pose and shape from a single color image. In: CVPR (2018) 3", + "71. Pavlo, D., Grangier, D., Auli, M.: Quaternet: A quaternion-based recurrent model for human motion. In: BMVC (2018) 4", + "72. Payer, C., Neff, T., Bischof, H., Urschler, M., Stern, D.: Simultaneous multi-person detection and single-person pose estimation with a single heatmap regression network. In: ICCV PoseTrack Workshop (2017) 3", + "73. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: ICCV (2021) 4", + "74. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: ICLR (2023) 2, 6", + "75. Reddy, N.D., Guigues, L., Pischulini, L., Eledath, J., Narasimhan, S.: Tessetrack: End-to-end learnable multi-person articulated 3d pose tracking. In: CVPR (2021) 3", + "76. Rempe, D., Birdal, T., Hertzmann, A., Yang, J., Sridhar, S., Guibas, L.J.: Humor: 3d human motion model for robust pose estimation. In: ICCV (2021) 2, 3, 4", + "77. Rogez, G., Weinzaepfel, P., Schmid, C.: LCR-Net: Localization-classification-regression for human pose. In: CVPR (2017) 3", + "78. Rong, Y., Liu, Z., Li, C., Cao, K., Change Loy, C.: Delving deep into hybrid annotations for 3d human recovery in the wild. In: ICCV (2019) 3", + "79. Sárándi, I., Hermans, A., Leibe, B.: Learning 3D human pose estimation from dozens of datasets using a geometry-aware autoencoder to bridge between skeleton formats. In: WACV (2023) 10", + "80. Shimada, S., Golyanik, V., Xu, W., Theobalt, C.: Physcap: Physically plausible monocular 3d motion capture in real time. In: SIGGRAPH (2020) 3", + "81. Shin, S., Kim, J., Halilaj, E., Black, M.J.: Wham: Reconstructing world-grounded humans with accurate 3d motion. In: CVPR (2024) 1, 2, 3, 4, 10, 11, 12, 14", + "82. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015) 4", + "83. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2021) 7", + "84. Song, J., Chen, X., Hilliges, O.: Human body model fitting by learned gradient descent. In: ECCV (2020) 3", + "85. Sun, Y., Bao, Q., Liu, W., Fu, Y., Black, M.J., Mei, T.: Monocular, one-stage, regression of multiple 3d people. In: ICCV (2021) 3", + "86. Sun, Y., Bao, Q., Liu, W., Mei, T., Black, M.J.: Trace: 5d temporal regression of avatars with dynamic cameras in 3d environments. In: CVPR (2023) 4, 12", + "87. Sun, Y., Liu, W., Bao, Q., Fu, Y., Mei, T., Black, M.J.: Putting people in their place: Monocular regression of 3D people in depth. In: CVPR (2022) 3" + ], + "bbox": [ + 223, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "L. Jiefeng et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "88. Sun, Y., Ye, Y., Liu, W., Gao, W., Fu, Y., Mei, T.: Human mesh recovery from monocular images via a skeleton-disentangled representation. In: ICCV (2019) 3", + "89. Teed, Z., Deng, J.: DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. NeurIPs (2021) 2, 5, 11, 12, 13", + "90. Teed, Z., Lipson, L., Deng, J.: Deep patch visual odometry. NeurIPS (2023) 2, 12", + "91. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. In: ICLR 2023 (2022) 2, 4, 8", + "92. Villegas, R., Yang, J., Zou, Y., Sohn, S., Lin, X., Lee, H.: Learning to generate long-term future via hierarchical prediction. In: ICML (2017) 4", + "93. Weng, Z., Yeung, S.: Holistic 3d human and scene mesh estimation from single view images. In: CVPR (2021) 3", + "94. Xiang, D., Joo, H., Sheikh, Y.: Monocular total capture: Posing face, body and hands in the wild. In: CVPR (2019) 3", + "95. Xie, K., Wang, T., Iqbal, U., Guo, Y., Fidler, S., Shkurti, F.: Physics-based human motion estimation and synthesis from videos. In: ICCV (2021) 3", + "96. Xie, Y., Jampani, V., Zhong, L., Sun, D., Jiang, H.: Omnicontrol: Control any joint at any time for human motion generation. In: ICLR (2024) 4", + "97. Xu, Y., Zhu, S.C., Tung, T.: Denserac: Joint 3d pose and shape estimation by dense render-and-compare. In: ICCV (2019) 3", + "98. Yan, X., Rastogi, A., Villegas, R., Sunkavalli, K., Shechtman, E., Hadap, S., Yumer, E., Lee, H.: Mt-vae: Learning motion transformations to generate multimodal human dynamics. In: ECCV (2018) 4", + "99. Ye, V., Pavlakos, G., Malik, J., Kanazawa, A.: Decoupling human and camera motion from videos in the wild. In: CVPR (2023) 2, 4, 10, 11, 12, 13", + "00. Yuan, Y., Iqbal, U., Molchanov, P., Kitani, K., Kautz, J.: Glamr: Global occlusion-aware human mesh recovery with dynamic cameras. In: CVPR (2022) 2, 4, 10, 11, 12, 13", + "01. Yuan, Y., Kitani, K.: Diverse trajectory forecasting with determinantal point processes. In: ICLR 2020 (2019) 4", + "02. Yuan, Y., Kitani, K.: Dlow: Diversifying latent flows for diverse human motion prediction. In: ECCV (2020) 4", + "03. Yuan, Y., Kitani, K.: Residual force control for agile human behavior imitation and extended motion synthesis. In: NeurIPS (2020) 4", + "04. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: ICCV (2023) 4", + "05. Yuan, Y., Wei, S.E., Simon, T., Kitani, K., Saragih, J.: Simpoe: Simulated character control for 3d human pose estimation. In: CVPR (2021) 3", + "06. Zanfir, A., Bazavan, E.G., Xu, H., Freeman, W.T., Sukthankar, R., Sminchisescu, C.: Weakly supervised 3d human pose and shape reconstruction with normalizing flows. In: ECCV (2020) 3", + "07. Zanfir, A., Marinoiu, E., Sminchisescu, C.: Monocular 3d pose and shape estimation of multiple people in natural scenes the importance of multiple scene constraints. In: CVPR (2018) 3", + "08. Zanfir, A., Marinoiu, E., Zanfir, M., Popa, A.I., Sminchisescu, C.: Deep network for the integrated 3d sensing of multiple people in natural images. In: NeurIPS (2018) 3", + "09. Zanfir, M., Zanfir, A., Bazavan, E.G., Freeman, W.T., Sukthankar, R., Sminchisscu, C.: Thundr: Transformer-based 3d human reconstruction with markers. In: ICCV (2021) 3" + ], + "bbox": [ + 225, + 146, + 784, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "COIN: Control-Inpainting Diffusion", + "bbox": [ + 491, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "110. Zhang, H., Tian, Y., Zhou, X., Ouyang, W., Liu, Y., Wang, L., Sun, Z.: Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In: ICCV (2021) 3", + "111. Zhang, J., Yu, D., Liew, J.H., Nie, X., Feng, J.: Body meshes as points. In: CVPR (2021) 3", + "112. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models (2023) 8", + "113. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022) 4", + "114. Zhang, S., Bhatnagar, B.L., Xu, Y., Winkler, A., Kadlecek, P., Tang, S., Bogo, F.: Rohm: Robust human motion reconstruction via diffusion. In: CVPR (2024) 4", + "115. Zhang, S., Zhang, Y., Bogo, F., Pollefeys, M., Tang, S.: Learning motion priors for 4d human body capture in 3d scenes. In: ICCV (2021) 4", + "116. Zhang, T., Huang, B., Wang, Y.: Object-occluded human shape and pose estimation from a single color image. In: CVPR (2020) 3", + "117. Zhen, J., Fang, Q., Sun, J., Liu, W., Jiang, W., Bao, H., Zhou, X.: SMAP: Single-shot multi-person absolute 3d pose estimation. In: ECCV (2020) 3", + "118. Zhou, Y., Habermann, M., Habibie, I., Tewari, A., Theobalt, C., Xu, F.: Monocular real-time full body capture with inter-part correlations. In: CVPR (2021) 3" + ], + "bbox": [ + 215, + 146, + 785, + 448 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "L. Jiefeng et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/COIN_ Control-Inpainting Diffusion Prior for Human and Camera Motion Estimation/213c3e91-2355-4081-9e3c-28b7d1b0600c_model.json b/2024/COIN_ Control-Inpainting Diffusion Prior for Human and Camera Motion Estimation/213c3e91-2355-4081-9e3c-28b7d1b0600c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9ce1f6880b00e354e9a2e850afb9f7673fd65150 --- /dev/null +++ b/2024/COIN_ Control-Inpainting Diffusion Prior for Human and Camera Motion Estimation/213c3e91-2355-4081-9e3c-28b7d1b0600c_model.json @@ -0,0 +1,3056 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.245, + 0.141, + 0.761, + 0.185 + ], + "angle": 0, + "content": "COIN: Control-Inpainting Diffusion Prior for Human and Camera Motion Estimation" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.194, + 0.777, + 0.226 + ], + "angle": 0, + "content": "Jiefeng Li\\(^{1,2}\\), Ye Yuan\\(^{1}\\), Davis Rempe\\(^{1}\\), Haotian Zhang\\(^{1}\\), Pavlo Molchanov\\(^{1}\\), Cewu Lu\\(^{2}\\), Jan Kautz\\(^{1}\\), and Umar Iqbal\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.352, + 0.236, + 0.651, + 0.268 + ], + "angle": 0, + "content": "\\(^{1}\\)NVIDIA \\(^{2}\\)Shanghai Jiao Tong University https://nvlabs.github.io/COIN/" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.275, + 0.789, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.461, + 0.79, + 0.558 + ], + "angle": 0, + "content": "Fig. 1: Capturing global human and camera motion from a dynamic camera presents unique challenges. In the input video, a person is riding a skateboard - while the local body motion may remain relatively constant, the global position of the individual changes significantly. Current state-of-the-art methods such as PACE [41] and WHAM [81] fail catastrophically on such out-of-distribution motions. Our approach, COIN, gracefully handles such challenging cases, owing to our control-inpainting motion diffusion prior and novel human-scene relation loss." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.578, + 0.741, + 0.841 + ], + "angle": 0, + "content": "Abstract. Estimating global human motion from moving cameras is challenging due to the entanglement of human and camera motions. To mitigate the ambiguity, existing methods leverage learned human motion priors, which however often result in oversmoothed motions with misaligned 2D projections. To tackle this problem, we propose COIN, a control-inpainting motion diffusion prior that enables fine-grained control to disentangle human and camera motions. Although pre-trained motion diffusion models encode rich motion priors, we find it non-trivial to leverage such knowledge to guide global motion estimation from RGB videos. COIN introduces a novel control-inpainting score distillation sampling method to ensure well-aligned, consistent, and high-quality motion from the diffusion prior within a joint optimization framework. Furthermore, we introduce a new human-scene relation loss to alleviate the scale ambiguity by enforcing consistency among the humans, camera, and scene. Experiments on three challenging benchmarks demonstrate the effectiveness of COIN, which outperforms the state-of-the-art methods in terms of global human motion estimation and camera motion estimation. As an illustrative example, COIN outperforms the state-of-the-art method by \\(33\\%\\) in world joint position error (W-MPJPE) on the RICH dataset." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "L. Jiefeng et al." + }, + { + "type": "header", + "bbox": [ + 0.262, + 0.148, + 0.74, + 0.175 + ], + "angle": 0, + "content": "Keywords: Global Human Motion Estimation \\(\\cdot\\) Human Motion Prior \\(\\cdot\\) Score Distillation Sampling" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.203, + 0.377, + 0.219 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.236, + 0.784, + 0.295 + ], + "angle": 0, + "content": "Recovering global human and camera motion from dynamic RGB videos is an important problem with many applications, such as animation, human-computer interaction, mixed reality, and robotics. However, it is a very challenging problem due to the entanglement of human and camera motion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.296, + 0.787, + 0.538 + ], + "angle": 0, + "content": "There are only a few works [41, 81, 99, 100] that try to address this problem. Earlier methods [47, 100] only focus on human motion and ignore the camera motion. Their core insight is that the global body motion is highly correlated with the local motion. Thus, they can use the local body movements to estimate the global orientation and trajectory with a regression model [100] or by combining them with physics constraints [47]. However, these regression models ignore the camera movements, so they fail to maintain consistency with the input video, whereas physics-based methods fail to model complex in-the-wild environments so are limited to controlled scenarios. Recent works [41, 99] try to jointly estimate the human and camera motion by exploiting learned motion priors [21, 76] and SLAM [66, 89, 90]. They try to constrain the human body motion in a low-dimensional latent space of a motion prior model, which results in reconstructed motions that are overly smooth and do not align well with video observations. Moreover, the optimization of the camera motion is only based on the global human motion from the motion prior. Hence, they fail catastrophically if the initial human motion predictions are significantly incorrect (as shown in Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.787, + 0.674 + ], + "angle": 0, + "content": "More recently, Denoising Diffusion Models [24,91] have emerged as a powerful family of generative models that can model high-quality data priors. Nonetheless, effectively leveraging the learned priors remains an ongoing challenge. Score Distillation Sampling (SDS) is commonly employed for this purpose [74], but we find that naive application of SDS also results in inconsistencies with the available observations (see Sec. 4.2). The root cause of this problem lies in the inconsistency of randomly sampled motions during SDS optimization. Without constraints, these motions may not align with observed evidence, leading to overly smoothed results that lack detail due to the mode-averaging effect." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In this work, we propose COIN, a hybrid COntrol-INpainting score distillation sampling method to address the aforementioned limitations of vanilla SDS. First, we use the partially observed human motion from the video as control signals to guide motion sampling. To tackle noisy observations which may be out of distribution for the prior, we propose a dynamic controlled sampling technique that iteratively refines the observed motions and updates the control signals to ensure effective distillation from the motion prior. Second, to further improve the consistency of the sampled motions, we also develop a novel soft inpainting strategy. We automatically identify the high-confidence regions of the initial predicted global motion from the video and use them as soft constraints during optimization. Concretely, we sample less confident regions from scratch using the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.493, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "COIN: Control-Inpainting Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "motion model, while the confident regions are only slightly refined. This ensures that the reconstructed motions do not deviate from the available observations. Our new SDS formulation is used to jointly optimize the human and camera motion by finding the most plausible solution that explains the observed evidence. Finally, to prevent catastrophic failure in cases where the initial body motion fails significantly, we propose a human-scene relation loss to consider the human-scene depth relations. This novel loss provides complementary information to the human motion prior by using local motion and scene features. It regularizes the camera scale by enforcing consistency among the human motion, camera motion, and scene features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.788, + 0.375 + ], + "angle": 0, + "content": "We benchmark our approach on the synthetic HCM [41] dataset and the real-world RICH [25] and EMDB [37] datasets. We demonstrate that our approach significantly outperforms the state-of-the-art methods in terms of human motion estimation and camera motion estimation. Overall, the contributions of this paper can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.386, + 0.784, + 0.445 + ], + "angle": 0, + "content": "- We propose a novel control-inpainting motion prior specifically designed for global human motion estimation, which enhances score distillation sampling with dynamic control and soft inpainting to reconstruct well-aligned, consistent, and high-quality motions from video observations." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.448, + 0.784, + 0.49 + ], + "angle": 0, + "content": "- We propose a new human-scene relation loss to resolve the scale ambiguity of the camera motion by enforcing consistency among the human motion, camera motion, and scene features." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.493, + 0.785, + 0.597 + ], + "angle": 0, + "content": "- Our approach significantly outperforms the state-of-the-art methods in terms of human motion estimation and camera motion estimation on both synthetic and real-world datasets. In terms of global human motion estimation in world space, we outperform the state-of-the-art method PACE [41] by \\(44\\%\\) and \\(33\\%\\) on the HCM [41] and RICH [25] datasets, respectively. We also compare with the contemporary work WHAM [81] and outperform it by \\(49\\%\\) and \\(7\\%\\) on the RICH and EMDB datasets, respectively." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.386, + 0.785, + 0.597 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.625, + 0.388, + 0.64 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Camera-Space Human Pose Estimation. Most existing works focus on root-relative local human pose estimation to bypass the difficulty in monocular depth estimation [1, 4, 6-8, 16, 32-34, 39, 42-46, 51, 54, 63, 64, 67, 68, 70, 76, 78, 84, 85, 88, 94, 97, 106, 110, 116, 118]. These methods ignore the position of the person in the camera coordinates. To overcome this limitation, recent methods estimate camera-space human poses by regression [28, 31, 48, 52, 62, 72, 75, 80, 93, 95, 107, 107, 109, 111] or optimization [59-61, 77, 108]. Physics-based constraints are widely used to ensure the plausibility of the estimated poses [10, 13, 27, 29, 80, 95, 105]. In addition to direct regression, heatmap-based representations have also been used to predict the absolute depths of multiple people [11, 87, 117]. A few methods improve absolution depth estimation by using predicted camera parameters [40, 49, 109] instead of the predefined focal length. Despite the promising results" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "L. Jiefeng et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "for camera-space pose estimation, how to decouple the camera movement and estimate global human motions is still an open problem." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.207, + 0.787, + 0.569 + ], + "angle": 0, + "content": "Monocular Global Human Pose Estimation. Recovering the global human motion from a monocular moving camera is challenging due to the entanglement of human and camera motions. To disentangle the camera movement, several methods use IMU sensors or pre-scanned environments to recover global human motions [17,20,57,69], which is impractical for large-scale adoption. Recent works use human motion priors [100] or physics-based constraints [47,55] to recover human motions from monocular videos, but do not consider background scene features, which limits performance on in-the-wild videos. Sun et al. [86] use optical flow as a motion cue to estimate the global motions. A contemporary work, WHAM [81], uses a lifting network to estimate global human motions from 2D keypoints and camera angular velocities. While these works can estimate accurate global human motions, they do not recover camera motions. To explicitly recover the camera motion, Liu et al. [52] use SLAM and convert the local pose from the camera to global coordinates. BodySLAM [22] jointly optimizes the human and camera motion using features of both humans and scenes. Along this line, SLAHMR [99] and PACE [41] use SLAM to initialize camera motions and optimize the camera using human motion priors [21,76]. However, these methods rely on the human motion priors to regularize the camera motion, which may lead to inaccurate camera motion when the human motion is not well initialized (as shown in Fig. 1). Such wrong camera trajectories will further affect the optimization of human motions. In contrast, our approach relies on the consistency among the local human motion, scene features, and the camera for optimization, which provides information that complements the human motion priors and enables accurate estimation of both human and camera motions." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.599, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Human Motion Priors. There are a significant amount of approaches proposed to study 3D human dynamics for motion prediction and synthesis [2, 3, 5, 12, 15, 18, 19, 23, 30, 36, 38, 50, 58, 71, 73, 92, 98, 101-103]. These learned human motion priors are used to help resolve pose ambiguity [21, 39, 76, 115] in human pose estimation. Recently, diffusion models [82] have also been used as priors for motion synthesis and infilling [26, 35, 91, 104, 113]. RoHM [114], adopts motion diffusion model to recover human motions from noisy and occluded input data. Xie et al. [96] use spatial control signals to guide motion generation. They focus on generating realistic human motion given clean spatial constraints. Müller et al. [65] build a diffusion model to learn the joint distribution over the poses of two people. They use the SDS loss to guide the generation of static poses. In contrast, we focus on dynamic human motions. We find adopting SDS directly for temporal human motions encounters the inconsistency issue. Therefore, to distill knowledge from the motion diffusion model, we propose a novel control inpainting SDS to generate high-quality and consistent motion that aligns with observed evidence." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.493, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "COIN: Control-Inpainting Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.784, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.381, + 0.788, + 0.491 + ], + "angle": 0, + "content": "Fig. 2: Overview. Given a video with a moving camera, we recover the global human motion \\(\\mathbf{H}\\) and camera motion \\(\\mathcal{C}\\) using an iterative optimization framework. We propose a novel Control-Inpainting SDS loss (\\(\\mathcal{L}_{\\mathrm{COIN - SDS}}\\)) to leverage motion diffusion models as a prior. COIN-SDS is designed such that the sampled motions from the motion prior are consistent with video observations. We achieve this by controlling and constraining the sampling process of the motion diffusion model through novel control and soft-inpainting branches. We also propose a novel human-scene relation loss (\\(\\mathcal{L}_{\\mathrm{HSR}}\\)) to encourage consistency among the human motion, camera motion, and scene features." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.499, + 0.33, + 0.515 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.53, + 0.788, + 0.756 + ], + "angle": 0, + "content": "The overall framework of COIN is illustrated in Fig. 2. Given an in-the-wild RGB video with \\(T\\) frames captured by a dynamic camera, our goal is to estimate both the global human motion \\(\\mathbf{H} = \\{\\mathbf{h}^{(1)},\\mathbf{h}^{(2)},\\dots,\\mathbf{h}^{(T)}\\}\\) and the camera motion \\(\\mathcal{C} = \\{\\mathbf{c}^{(1)},\\mathbf{c}^{(2)},\\dots,\\mathbf{c}^{(T)}\\}\\) in a global world coordinate system. We use off-the-shelf 3D human pose and shape estimation method HybrIK [48] to obtain per-frame initial SMPL parameters in the camera space and DROID-SLAM [89] to obtain the initial per-frame camera-to-world transforms. We convert the local human motion to the world coordinates with the estimated camera. However, because the camera trajectories from SLAM are up to an unknown scale, the initial global human motion will abnormally drift and float in the world space. To resolve the scale ambiguity and place the person in the correct global position, we jointly optimize the human and camera motion to minimize the discrepancy between the observed evidence and the estimated motion, while maintaining the plausibility of the human motion with a diffusion prior through the proposed control-inpainting SDS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Motion Representation. The camera motion is represented by the trajectory \\(\\mathcal{C} = \\{(\\mathbf{R}^{(i)},\\mathbf{t}^{(i)})\\}_{i = 1}^{T}\\), where \\([\\mathbf{R}^{(i)},\\mathbf{t}^{(i)}]\\) is the camera pose at the \\(i\\)-th frame, consisting of the rotation matrix \\(\\mathbf{R}^{(i)}\\in \\mathbb{R}^{3\\times 3}\\) and the translation vector \\(\\mathbf{t}^{(i)}\\in \\mathbb{R}^3\\). The human motion is represented by the human trajectory \\(\\mathbf{H} = \\{\\mathbf{h}^{(i)}\\}_{i = 1}^{T}\\) where \\(\\mathbf{h}^{(i)} = [\\tau^{(i)},\\varPhi^{(i)},\\theta^{(i)},f^{(i)},\\beta ]\\) is the human pose at the \\(i\\)-th frame, consisting" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "L. Jiefeng et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.253 + ], + "angle": 0, + "content": "of the global translation \\(\\tau^{(i)}\\in \\mathbb{R}^3\\), global orientation \\(\\varPhi^{(i)}\\in \\mathbb{R}^3\\), body pose parameters \\(\\theta^{(i)}\\in \\mathbb{R}^{23\\times 3}\\), foot contact labels \\(f\\in \\{0,1\\} ^4\\), and the body shape parameters \\(\\beta \\in \\mathbb{R}^{10}\\). We use the SMPL model [53] to represent the human pose and shape. Before introducing our approach, we first revisit the formulation and drawbacks of SDS in Sec. 3.1. Then we introduce the proposed control-inpainting SDS in Sec. 3.2. Finally, we present the global optimization pipeline with the proposed human-scene interaction loss in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.268, + 0.389, + 0.284 + ], + "angle": 0, + "content": "3.1 Revisiting SDS" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.292, + 0.785, + 0.381 + ], + "angle": 0, + "content": "Score Distillation Sampling (SDS) was first introduced to distill 3D assets from pre-trained 2D text-to-image diffusion models [74]. It exploits the knowledge from the diffusion models by seeking modes for the conditional distribution in the DDPM latent space to optimize the 3D scene representation. Similarly, we can optimize global human motion by distilling knowledge from a pre-trained motion diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.383, + 0.786, + 0.414 + ], + "angle": 0, + "content": "Given an global human motion \\(\\mathbf{H}\\), the marginal distribution of noisy latent \\(\\mathbf{H}_t\\) at timestep \\(t \\in \\mathcal{U}(0,1)\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.37, + 0.424, + 0.786, + 0.442 + ], + "angle": 0, + "content": "\\[\nq (\\mathbf {H} _ {t} | \\mathbf {H}) = \\mathcal {N} (\\mathbf {H} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {H}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.451, + 0.788, + 0.557 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_t\\in (0,1)\\) is a hyperparameter controlled by the variance schedule of the diffusion model. SDS adopts the pre-trained diffusion model \\(\\mathcal{D}_{\\phi}(\\mathbf{H}_t,t,y)\\), which takes in \\(\\mathbf{H}_t\\) and is used to model the conditional density of the human motion, where \\(\\phi\\) are the parameters of the diffusion model and \\(y\\) is the condition. Then, SDS aims to distill global human motion \\(\\mathbf{H}\\) via seeking modes of the learned conditional density, which can be achieved by a weighted denoising score matching objective:" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.568, + 0.786, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\underset {\\mathbf {H}} {\\min } \\mathcal {L} _ {\\mathrm {S D S}} := \\mathbb {E} _ {t, \\epsilon} \\left[ \\omega (t) \\| \\epsilon_ {\\phi} ^ {t} - \\epsilon \\| _ {2} ^ {2} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.601, + 0.787, + 0.647 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\phi}^{t}\\) is the predicted denoising direction from the diffusion model, \\(\\mathbf{H}_t \\sim q(\\mathbf{H}_t|\\mathbf{H})\\) is sampled using the reparameterization trick, \\(\\epsilon\\) is the corresponding sampled noise, and \\(\\omega(t)\\) is a weighting function that depends on the timestep \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.647, + 0.729, + 0.661 + ], + "angle": 0, + "content": "To clearly review the effect of SDS, we can reparameterize Eq. 2 as:" + }, + { + "type": "equation", + "bbox": [ + 0.356, + 0.671, + 0.786, + 0.706 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {H}} \\mathcal {L} _ {\\mathrm {S D S}} := \\mathbb {E} _ {t} \\left[ \\frac {\\omega (t) \\sqrt {\\bar {\\alpha} _ {t}}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\left\\| \\mathbf {H} - \\hat {\\mathbf {H}} _ {0} ^ {t} \\right\\| _ {2} ^ {2} \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.715, + 0.264, + 0.727 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.415, + 0.725, + 0.786, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {H}} _ {0} ^ {t} = \\frac {\\mathbf {H} _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon} _ {\\phi} ^ {t}}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Based on this reparameterization, we can see that the SDS objective is to minimize the discrepancy between global human motion \\(\\mathbf{H}\\) and the denoised global human motion \\(\\hat{\\mathbf{H}}_0^t\\) from the motion diffusion model in a single step. The denoised motion \\(\\hat{\\mathbf{H}}_0^t\\) serves as the pseudo ground truth. However, at each optimization step, we randomly sample \\(t\\) and \\(\\epsilon\\) to generate the noisy latent \\(\\mathbf{H}_t\\), and we found" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.493, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "COIN: Control-Inpainting Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "the pre-trained diffusion model is sensitive to the input. Minor fluctuations in the input latent would substantially change the denoised motion, which leads to inconsistency in \\(\\hat{\\mathbf{H}}_0^t\\) across different time steps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.788, + 0.389 + ], + "angle": 0, + "content": "Although randomness can help generate diverse plausible motions to infer occluded regions and unknown information, we do not need it for well-observed regions, such as simple body poses in a clean background. Such randomness in the denoising steps makes the generated \\(\\hat{\\mathbf{H}}_0^t\\) difficult to align with the local 2D observations and results in wrong global human motion. Moreover, this pseudo ground truth \\(\\hat{\\mathbf{H}}_0^t\\) is generated from only a single denoising step, where the diffusion models may not produce high-quality motions, resulting in foot sliding and floating. Although sampling with a smaller timestep \\(t\\) can alleviate these issues, the initial motion is usually inaccurate and the denoiser is not able to remove artifacts with a small \\(t\\). To exploit the knowledge of the motion diffusion model and denoise the initial motion, we must allow the SDS to sample with a larger timestep \\(t\\) while maintaining high quality, consistency, and alignment with the local 2D observations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.405, + 0.457, + 0.42 + ], + "angle": 0, + "content": "3.2 Control-Inpainting SDS" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.598 + ], + "angle": 0, + "content": "The limitations of SDS originate from the randomness and inconsistency of the denoised motion \\(\\hat{\\mathbf{H}}_0^t\\), which serves as the pseudo ground truth in the objective function. To overcome this issue, we propose a novel COntrol-INpainting SDS (COIN-SDS) to generate high-quality and consistent pseudo-ground-truth motions. Our solution has three key ingredients (shown in Alg. 1). First, to achieve high-quality motions, we seek to produce the pseudo ground truth with multiple DDIM denoising steps. Second, to encourage consistent motions, we propose to use partially observed evidence from the video as a control signal to dynamically guide the diffusion model and align the generated motions with the observations. Third, to further align the motion with observed regions, we propose a soft inpainting strategy within the denoising process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.787, + 0.653 + ], + "angle": 0, + "content": "Multiple Denoising Steps. Intuitively, to obtain high-quality pseudo ground truth for SDS, we can replace the single-step denoised motion \\(\\hat{\\mathbf{H}}_0^t\\) with a multi-step one \\(\\tilde{\\mathbf{H}}_0^t \\coloneqq \\tilde{\\mathbf{H}}_0\\), following the multi-step DDIM denoising process [83]:" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.664, + 0.785, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {H}} _ {t - \\Delta t} = \\sqrt {\\bar {\\alpha} _ {t - \\Delta t}} \\cdot \\hat {\\mathbf {H}} _ {0} ^ {t} + \\sqrt {1 - \\bar {\\alpha} _ {t - \\Delta t}} \\cdot \\boldsymbol {\\epsilon} _ {\\phi} ^ {t}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.694, + 0.787, + 0.726 + ], + "angle": 0, + "content": "until \\(\\tilde{\\mathbf{H}}_0 = \\tilde{\\mathbf{H}}_{t - \\Delta t}\\) is obtained. By replacing \\(\\tilde{\\mathbf{H}}_0^t\\) in Eq. 3 with \\(\\tilde{\\mathbf{H}}_0\\), we can obtain a new objective for SDS:" + }, + { + "type": "equation", + "bbox": [ + 0.36, + 0.738, + 0.785, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {H}} \\mathcal {L} _ {\\mathrm {S D S}} := \\mathbb {E} _ {t} \\left[ \\frac {\\omega (t) \\sqrt {\\bar {\\alpha} _ {t}}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\left\\| \\mathbf {H} - \\tilde {\\mathbf {H}} _ {0} \\right\\| _ {2} ^ {2} \\right]. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.797, + 0.856 + ], + "angle": 0, + "content": "Although the multi-step denoising process can produce high-quality pseudo ground truth, it is computationally expensive to perform multiple denoising steps during optimization, which limits the practicality of increasing the number of denoising steps. In our experiments, we find that using 10 denoising steps is sufficient to produce high-quality pseudo ground truth." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "L. Jiefeng et al." + }, + { + "type": "code_caption", + "bbox": [ + 0.23, + 0.15, + 0.422, + 0.164 + ], + "angle": 0, + "content": "Algorithm 1: COIN-SDS" + }, + { + "type": "algorithm", + "bbox": [ + 0.227, + 0.167, + 0.765, + 0.36 + ], + "angle": 0, + "content": "Input: Latest human motion H, confidence score S, visible mask M \nOutput: \\(\\mathcal{L}_{\\mathrm{COIN - SDS}}\\) \n1 Sample: \\(\\tilde{\\mathbf{H}}_t \\sim \\mathcal{N}(\\mathbf{H}_t; \\sqrt{\\bar{\\alpha}_t}\\mathbf{H}, (1 - \\bar{\\alpha}_t)\\mathbf{I}), t \\sim \\mathcal{U}(0,1)\\); \n2 for \\(\\bar{t} = [t, t - \\Delta t, \\dots, \\Delta t]\\) do // multi-step DDIM denoising \n3 \\(\\tilde{\\mathbf{H}}_0^{\\bar{t},\\mathrm{known}} \\gets \\mathbf{H}\\); \n4 \\(\\tilde{\\mathbf{H}}_0^{\\bar{t},\\mathrm{un-known}} \\gets \\mathcal{D}_{\\phi,\\phi_c}(\\tilde{\\mathbf{H}}_{\\bar{t}}, \\bar{t}, \\mathbf{H} \\odot \\mathbf{M})\\); // controlled denoising \n5 \\(\\tilde{\\mathbf{M}} \\gets w(\\bar{t}) * \\mathbf{S} \\odot \\mathbf{M}\\); \n6 \\(\\tilde{\\mathbf{H}}_0^{\\bar{t}} \\gets \\tilde{\\mathbf{M}} \\odot \\tilde{\\mathbf{H}}_0^{\\bar{t},\\mathrm{known}} + (1 - \\tilde{\\mathbf{M}}) \\odot \\tilde{\\mathbf{H}}_0^{\\bar{t},\\mathrm{unknown}}\\); // soft inpainting \n7 \\(\\epsilon_{\\phi}^{\\bar{t}} \\gets \\frac{\\tilde{\\mathbf{H}}_{\\bar{t}} - \\sqrt{\\bar{\\alpha}_{\\bar{t}}} \\tilde{\\mathbf{H}}_0^{\\bar{t}}}{\\sqrt{1 - \\bar{\\alpha}_{\\bar{t}}}}\\); \n8 \\(\\tilde{\\mathbf{H}}_{\\bar{t} -\\Delta t} \\gets \\sqrt{\\bar{\\alpha}_{\\bar{t} -\\Delta t}} \\cdot \\tilde{\\mathbf{H}}_0^{\\bar{t}} + \\sqrt{1 - \\bar{\\alpha}_{\\bar{t} -\\Delta t}} \\cdot \\epsilon_{\\phi}^{\\bar{t}}\\); // update latent motion \n9 \\(\\mathcal{L}_{\\mathrm{COIN - SDS}} = \\frac{\\omega(t) \\sqrt{\\bar{\\alpha}_t}}{\\sqrt{1 - \\bar{\\alpha}_t}}\\left\\| \\mathbf{H} - \\tilde{\\mathbf{H}}_0 \\right\\|_2^2\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.445 + ], + "angle": 0, + "content": "Dynamic Controlled Sampling. To generate consistent motions that are aligned with the observed evidence, we propose to attach a control branch \\(\\phi_c\\) to the pre-trained diffusion model \\(\\mathcal{D}_{\\phi}\\) to guide the motion generation. Given a latent motion \\(\\tilde{\\mathbf{H}}_t\\), control signal \\(\\mathbf{c}\\), and the visible mask \\(\\mathbf{M}\\), we train a controlled denoiser \\(\\mathcal{D}_{\\phi,\\phi_c}\\) to generate intermediate denoised motion \\(\\tilde{\\mathbf{H}}_0^t\\) for DDIM denoising:" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.46, + 0.786, + 0.479 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {H}} _ {0} ^ {t} = \\mathcal {D} _ {\\phi , \\phi_ {c}} \\left(\\tilde {\\mathbf {H}} _ {t}, t, \\mathbf {c} \\odot \\mathbf {M}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.788, + 0.567 + ], + "angle": 0, + "content": "where \\(\\mathbf{c}\\) and \\(\\mathbf{M}\\) are the same size as the motion, \\(\\mathbf{M}\\) is a binary mask with ones in observed pose dimensions, and \\(\\odot\\) denotes the element-wise multiplication. During training, we synthesize noise and occlusions by randomly adding Gaussian noise to the control signal \\(\\mathbf{c}\\) and randomly masking the pose and trajectory dimensions in \\(\\mathbf{M}\\). Details are provided in the supplementary document." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.704 + ], + "angle": 0, + "content": "When using the denoiser in the optimization stage, instead of always using the initial noisy estimation from HybrIK [48] as the fixed control signal, we propose to use a dynamic control strategy. Specifically, we use the optimized human motion from the previous iteration as the control signal, i.e., \\(\\mathbf{c} = \\mathbf{H}\\). This strategy prevents performance degradation due to inaccurate initializations. A better control signal can guide the model to generate a more plausible pseudo motion, which in turn helps to optimize the global human motion and provides a control signal that better aligns with the input videos. Such self-evolving control signals can help to generate well-aligned global human motions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The pre-trained motion diffusion model adopts a transformer encoder structure like [91]. We follow ControlNet [112] to encode the control signals and guide the denoiser output. We create a trainable copy of 4 encoding blocks of the pretrained motion diffusion model followed with zero convolutions. The input to the control branch is the concatenation of the latent and the control signals. The AMASS [56] dataset is used to train the motion diffusion model. The finetuning of the controlled denoiser is computationally efficient since the pre-trained branch is frozen and only the control branch is trained. See the supplementary document for more details on the architecture and training settings." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.493, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "COIN: Control-Inpainting Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.238 + ], + "angle": 0, + "content": "Soft Inpainting. While guiding motion generation encourages outputs to align with the conditions, it is often not strong enough. We further seek to improve the consistency by masking the known regions and inpainting the unknown regions. Given a binary mask \\(\\mathbf{M}\\) that indicates the observed and unobserved regions, we can use the diffusion model to generate the inpainted motion \\(\\tilde{\\mathbf{H}}_0^t\\) following the DDIM denoising process:" + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.245, + 0.784, + 0.262 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {H}} _ {0} ^ {t, \\text {k n o w n}} = \\mathbf {H}, \\tag {8a}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.267, + 0.784, + 0.284 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {H}} _ {0} ^ {t, \\text {u n k n o w n}} = \\mathcal {D} _ {\\phi , \\phi_ {c}} \\left(\\tilde {\\mathbf {H}} _ {t}, t, \\mathbf {H} \\odot \\mathbf {M}\\right), \\tag {8b}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.288, + 0.784, + 0.306 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {H}} _ {0} ^ {t} = \\mathbf {M} \\odot \\tilde {\\mathbf {H}} _ {0} ^ {t, \\text {k n o w n}} + (1 - \\mathbf {M}) \\odot \\tilde {\\mathbf {H}} _ {0} ^ {t, \\text {u n k n o w n}}. \\tag {8c}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.313, + 0.784, + 0.388 + ], + "angle": 0, + "content": "Thus, the known regions are overwritten with the observations, while the unknown regions are sampled from the diffusion model. However, the above formulation keeps the observed parts unchanged during the denoising process. In practice, the observed parts can be noisy and not perfect. We still want the diffusion model to refine the observed parts but not change them significantly." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.39, + 0.784, + 0.464 + ], + "angle": 0, + "content": "Here, we present a soft inpainting strategy to infill the unobserved regions while refining the observed regions by dynamically reweighting the denoised direction from the diffusion model. Specifically, instead of using a binary mask, we adopt a continuous mask \\(\\tilde{\\mathbf{M}}\\) depending on both the confidence score of the observations \\(\\mathbf{S}\\) and the denoising time step \\(t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.433, + 0.473, + 0.784, + 0.488 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} = w (t) * \\mathbf {S} \\odot \\mathbf {M}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.496, + 0.784, + 0.555 + ], + "angle": 0, + "content": "where we set \\( w(t) = \\max(0, \\frac{t - 0.5}{0.5}) \\) to linearly decrease the weight of the observations as the denoising time step decreases. As the time step decreases, the denoising process will be more deterministic and the model will be more certain of the generated motions." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.557, + 0.784, + 0.586 + ], + "angle": 0, + "content": "Combining the three components of our solution, the final objective for COIN-SDS is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.3, + 0.592, + 0.784, + 0.625 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {H}} \\mathcal {L} _ {\\mathrm {C O I N - S D S}} := \\mathbb {E} _ {t} \\left[ \\frac {\\omega (t) \\sqrt {\\bar {\\alpha} _ {t}}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\left\\| \\mathbf {H} - \\tilde {\\mathbf {H}} _ {0} (\\mathbf {H}, \\mathbf {M}, \\mathbf {S}, t) \\right\\| _ {2} ^ {2} \\right]. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.631, + 0.444, + 0.645 + ], + "angle": 0, + "content": "We summarize COIN in Alg. 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.654, + 0.431, + 0.669 + ], + "angle": 0, + "content": "3.3 Global Optimization" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.676, + 0.784, + 0.812 + ], + "angle": 0, + "content": "Here we present the overall optimization pipeline for the joint estimation of global human and camera motion with the proposed COIN-SDS loss. Note that we use SLAM to initialize the camera poses, which is scale-ambiguous. Therefore, we need to jointly optimize the camera scale \\( s \\) with the human and camera motions. Furthermore, the SLAM method assumes the camera in the first frame to be at the origin. To put the human motion in the correct positions, we follow PACE [41] and also optimize the camera height \\( h_0 \\) and the orientation \\( R_0 \\) for the first frame. The global human motions are initialized by the estimated local motions and the camera poses. The overall optimization objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.822, + 0.784, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {H}, \\mathcal {C}, s, h _ {0}, R _ {0}, \\beta} \\mathcal {L} _ {\\text {b o d y}} + \\mathcal {L} _ {\\text {C O I N - S D S}} + \\mathcal {L} _ {\\text {H S R}}, \\tag {11}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "L. Jiefeng et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.265, + 0.16 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.162, + 0.786, + 0.178 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {b o d y}} = \\mathcal {L} _ {2 \\mathrm {D}} + \\mathcal {L} _ {3 \\mathrm {D}} + \\mathcal {L} _ {\\beta} + \\mathcal {L} _ {\\text {s m o o t h}} + \\mathcal {L} _ {\\text {c o n t a c t}}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.184, + 0.789, + 0.306 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{2\\mathrm{D}}\\) measures the 2D reprojection error between the projected 2D body joints of the estimated human motion and the detected 2D keypoints from an off-the-shelf 2D joint detector [9]. \\(\\mathcal{L}_{3\\mathrm{D}}\\) measures the distance between the estimated local 3D joints and the detected 3D joints from an off-the-shelf 3D joint detector [79]. \\(\\mathcal{L}_{\\beta}\\) is the shape regularization loss. \\(\\mathcal{L}_{\\mathrm{smooth}}\\) is the temporal smoothness loss. \\(\\mathcal{L}_{\\mathrm{contact}}\\) is the foot contact loss to encourage zero velocities for contact joints. The contact labels are obtained from the pseudo ground truth motion from COIN-SDS. Please refer to the supplementary document for more details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.31, + 0.789, + 0.445 + ], + "angle": 0, + "content": "Human-scene Relation Loss. The camera trajectories recovered from SLAM are scale-ambiguous. Previous works [41,99] optimize the camera scale by projecting the global human motion to the camera space using the camera poses and minimizing the reprojection error. However, such a method entirely relies on the global human motions, which is in turn affected by the camera scale. If the human motion is not initialized well, the camera scale will also be inaccurate. To solve this problem, instead of the global human motions, we propose a new human-scene relation loss that uses the depth relation between the human and scene in the camera space, which disentangles the effect of the camera itself." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.79, + 0.596 + ], + "angle": 0, + "content": "Specifically, we use the point cloud of the scene recovered by SLAM as a constraint. First of all, the scale of the scene point cloud is the same as the camera scale, so optimizing the scene scale is equivalent to optimizing the camera scale. Second, the scene points that are projected onto the visible vertices of the body mesh should be occluded by the person. Otherwise, the corresponding body parts are invisible. Therefore, we can constrain the depth of the occluded scene points to be larger than the depth of the human body vertices. While finding the corresponding body vertices for each scene point is time-consuming, we propose to use the depth of its nearest body joint as a proxy. Given the scene point cloud \\(\\mathcal{P}\\) and the camera scale \\(s\\), the human-scene relation loss is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.227, + 0.607, + 0.787, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {H S R}} = - \\frac {1}{| \\mathcal {P} |} \\sum_ {i = 1} ^ {T} \\sum_ {p \\in \\mathcal {P} ^ {*}} \\min (0, \\mathcal {T} ^ {(i)} (p) _ {z} - j ^ {(i)} (p) _ {z}) \\cdot \\mathbb {1} (\\mathcal {T} ^ {(i)} (p) \\text {i s i n v i s i b l e}), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.662, + 0.789, + 0.77 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}^* = \\mathcal{P}*s\\) is the scaled point cloud of the scene, \\(\\mathcal{T}^{(i)}(p) = \\mathbf{R}^{(i)}p + \\mathbf{t}^{(i)}\\) is the transformed point in the \\(i\\)-th frame, \\(j^{(i)}(p)\\) is body joint that has the nearest 2D projection to the scene point \\(p\\) in the \\(i\\)-th frame, and \\(z\\) denotes the depth of a given point. If the depth order is correct, i.e., \\(\\mathcal{T}^{(i)}(p)_z - j^{(i)}(p)_z > 0\\), the loss is zero. The proposed human-scene relation loss uses the relation between the local motions and the scene to alleviate the scale ambiguity. The depth relation regularizes consistency among humans, cameras, and scenes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.778, + 0.377, + 0.794 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Datasets. We perform experiments on three human motion datasets. First is the real-world dataset RICH [25]. We follow previous works [41,81,100] to assess the performance of global human motion estimation using this dataset. The second" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.493, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "COIN: Control-Inpainting Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.255, + 0.145, + 0.747, + 0.158 + ], + "angle": 0, + "content": "Table 1: Global human motion estimation on the RICH dataset." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.16, + 0.783, + 0.346 + ], + "angle": 0, + "content": "
MethodPA-MPJPE ↓W-MPJPE ↓WA-MPJPE ↓W-RJE ↓ACCEL ↓
HybrIK [48] + SLAM [89]46.71073.1404.41166.220.2
GLAMR [100]79.9653.7365.1646.6107.7
SLAHMR [99]52.5571.6323.7400.59.4
WHAM [81]46.2497.6272.7478.26.7
PACE [41]49.3380.0197.2370.88.8
Guided Sampling132.81384.6502.21440.924.2
Noise Optimization66.7414.8195.3429.28.4
Vanilla SDS78.81453.5497.21458.012.7
COIN w/o Controlled Sampling44.0825.0291.8848.510.8
COIN w/o Dynamic Control49.5293.8180.6299.18.6
COIN w/o Soft Inpainting47.6325.8196.0324.99.6
COIN w/o LHSR43.6273.0176.3281.18.2
COIN42.9254.5169.5249.97.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.785, + 0.435 + ], + "angle": 0, + "content": "one is EMDB [37]. We follow previous works [81] to evaluate on a subset of EMDB for which they provide ground truth global motion with dynamic cameras. The third dataset is HCM [41], a synthetic dataset. Compared to real-world datasets, HCM contains more challenging camera motions. We follow previous work [41] to evaluate the global human motion and the camera motion using this dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.785, + 0.576 + ], + "angle": 0, + "content": "Metrics. We report various metrics for both human and camera motion. For human motion, standard metrics W-MPJPE and WA-MPJPE are used to evaluate global motion, while PA-MPJPE evaluates local motion. We also include an ACCEL metric to measure the joint acceleration error. For evaluation on EMDB, we follow previous work [81] to split sequences into smaller chunks of 100 frames and align each output segment with the ground-truth data using the first two frames W-MPJPE100 or the entire segment WA-MPJPE100. Root Orientation Error (ROE in deg) and Root Translation Error (RTE in \\( m \\)) evaluate the error over the entire trajectory after aligning with the initial camera pose." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.577, + 0.785, + 0.637 + ], + "angle": 0, + "content": "For camera motion, we report the average translation error after scale alignment (ATE), without scale alignment (ATE-S), and the camera acceleration error (CAM ACCEL). ATE-S more accurately reflects inaccuracies in the captured scale of the camera." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Baselines. As discussed in Sec. 2, there are different ways to use the pre-trained motion diffusion model as a motion prior. Here, we summarize the three main solutions and compare them with COIN in Sec. 4.2. (1) Guided Sampling, which embeds analytical guidance within the denoising procedure using objective functions, such as 2D projection and foot contact consistency. This does not suit our task because the camera trajectories are also unknown and guided gradients from the wrong camera will lead to unrealistic human motions. We need to optimize the human motion and camera motion simultaneously. (2) Noise Optimization, which represents the motion as latent noise and directly optimizes it. This is similar to other motion priors such as VAE [21]. At each optimization step, we need to calculate the gradients of the latent w.r.t. the generated motion, which is computationally expensive and we find the performance is not good enough. (3) Vanilla SDS, which removes our design and directly uses SDS for optimization." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "L. Jiefeng et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.251, + 0.145, + 0.751, + 0.158 + ], + "angle": 0, + "content": "Table 2: Global human motion estimation on the EMDB dataset." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.16, + 0.783, + 0.263 + ], + "angle": 0, + "content": "
MethodPA-MPJPE ↓W-MPJPE100 ↓WA-MPJPE100 ↓RTE ↓ROE ↓
HMR2.0 [14] + DPVO [90]49.62320.9662.917.544.4
GLAMR [100]56.0756.1286.216.774.9
TRACE [86]58.02244.9544.118.972.7
SLAHMR [99]61.5807.4336.913.867.9
WHAM [81]41.9439.2166.18.436.3
COIN32.7407.3152.83.534.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.256, + 0.293, + 0.744, + 0.306 + ], + "angle": 0, + "content": "Table 3: Global human motion estimation on the HCM dataset." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.308, + 0.783, + 0.412 + ], + "angle": 0, + "content": "
MethodPA-MPJE ↓W-MPJE ↓WA-MPJE ↓W-RJE ↓ACCEL ↓
HybrIK [48] + SLAM [89]67.61137.3780.31100.951.3
GLAMR [100]86.01977.6653.81958.033.4
SLAHMR [99]69.9888.9483.5862.214.9
PACE [41]65.3861.2478.3839.516.7
WHAM [81]47.9588.9279.3579.213.1
COIN45.5479.9212.1470.710.1
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.417, + 0.621, + 0.431 + ], + "angle": 0, + "content": "4.1 Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.596 + ], + "angle": 0, + "content": "Human Motion Estimation. We compare COIN against state-of-the-art methods on the RICH, EMDB, and HCM datasets. Quantitative results are shown in Tabs. 1, 2, and 3. We observe that COIN significantly outperforms the state-of-the-art methods on all datasets. On the RICH dataset, COIN outperforms the state-of-the-art method, PACE [41], by \\(125.5\\mathrm{mm}\\) in terms of W-MPJPE, showing \\(33.0\\%\\) relative improvement. On the EMDB and HCM datasets, COIN shows \\(29.1\\mathrm{mm}\\) and \\(109.0\\mathrm{mm}\\) improvement in terms of W-MPJPE, respectively. Qualitative comparisons with state-of-the-art methods, PACE [41] and WHAM [81], are shown in Figs. 1 and 3. More qualitative results are shown in the supplementary video." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.734 + ], + "angle": 0, + "content": "COIN not only improves global motion, but also improves local body motion. In terms of PA-MPJPE, COIN shows \\(3.3\\mathrm{mm}\\), \\(9.2\\mathrm{mm}\\), and \\(19.8\\mathrm{mm}\\) improvement on the RICH, EMDB, and HCM datasets, respectively. This demonstrates that COIN is able to distill high-quality motion priors from the diffusion model and help both local and global motion estimation. Regarding the joint acceleration error, COIN is \\(0.8\\mathrm{mm/s}^2\\) higher than WHAM [81]. Note that the joint acceleration error reflects the smoothness. Human motion can be over-smoothed but not accurate, hence it is important to look at acceleration error in conjunction with other metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Camera Motion Estimation. We further evaluate the performance of COIN on camera motion estimation on the HCM dataset. Quantitative results are shown in Tab. 4. COIN substantially surpasses the state-of-the-art camera motion estimation methods. Specifically, COIN reduces the absolute camera translation error, ATE-S by \\(73.8\\mathrm{mm}\\). This demonstrates that COIN is able to disentangle human and camera motions and accurately estimate the camera motion." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.493, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "COIN: Control-Inpainting Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.28, + 0.145, + 0.722, + 0.159 + ], + "angle": 0, + "content": "Table 4: Camera motion estimation on the HCM dataset." + }, + { + "type": "table", + "bbox": [ + 0.264, + 0.161, + 0.74, + 0.378 + ], + "angle": 0, + "content": "
MethodATE ↓ATE-S ↓CAMACCEL ↓
HybrIK [48] + SLAM [89]155.81670.717.1
GLAMR [100]1295.21714.6282.9
SLAHMR [99]155.8506.517.6
PACE [41]137.5459.716.2
Guided Sampling335.6992.415.6
Noise Optimization206.0500.912.3
Vanilla SDS306.4656.413.7
COIN w/o Controlled Sampling299.6553.814.0
COIN w/o Dynamic Control149.8397.711.3
COIN w/o Soft Inpainting167.7423.811.4
COIN w/o LHSR147.8402.112.0
COIN135.3385.911.3
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.389, + 0.39, + 0.405 + ], + "angle": 0, + "content": "4.2 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.785, + 0.472 + ], + "angle": 0, + "content": "In this section, we conduct ablation studies on the RICH and HCM datasets to evaluate the impact of each component on human and camera motions, respectively. More comparisons on other datasets are provided in the supplementary document." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.646 + ], + "angle": 0, + "content": "Baselines with Motion Diffusion Model. We first compare the aforementioned baselines with COIN. To use guided sampling in our tasks, we jointly update the camera poses using the gradients from the objective function during denoising. Quantitative results of human and camera motion estimation are shown in Tabs. 1 and 4, respectively. We observe that COIN outperforms all the baselines in terms of all metrics. As expected, guided sampling shows a terrible performance because it is not able to accurately estimate the camera trajectories. Noise optimization is better than vanilla SDS but worse than COIN, which generates consistent motion priors. Vanilla SDS is replacing \\(\\mathcal{L}_{\\mathrm{COIN}}\\) with the SDS loss and keeping the rest of the settings the same. These results demonstrate the effectiveness of COIN over other motion diffusion baselines." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.788, + 0.789 + ], + "angle": 0, + "content": "Impact of Dynamic Controlled Sampling. To study the effectiveness of controlled sampling, we compare COIN with and without using the controlled denoiser. Quantitative results are shown in Tabs. 1 and 4. We observe that controlled sampling significantly improves the performance of COIN. Specifically, COIN with controlled sampling reduces the W-MPJPE and ATE-S by \\(570.5\\mathrm{mm}\\) and \\(167.9\\mathrm{mm}\\), showing \\(69.3\\%\\) and \\(30.3\\%\\) relative improvement, respectively. This demonstrates that controlled sampling is able to generate high-quality motion priors that align with the observed evidence and help improve human and camera motion estimation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Impact of Soft Inpainting. We further study the effectiveness of soft inpainting. Quantitative comparisons are shown in Tabs. 1 and 4. We observe that while soft inpainting also improves W-MPJPE, it is much more effective than" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "L. Jiefeng et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.501, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.303, + 0.342, + 0.5, + 0.349 + ], + "angle": 0, + "content": "PACE (Human Motion) PACE (Camera Motion)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.144, + 0.787, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.342, + 0.563, + 0.349 + ], + "angle": 0, + "content": "WHAM" + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.342, + 0.786, + 0.349 + ], + "angle": 0, + "content": "COIN (Human Motion) COIN (Camera Motion)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.353, + 0.788, + 0.409 + ], + "angle": 0, + "content": "Fig. 3: Qualitative comparisons with state-of-the-art methods. PACE [41] fails to recover a correct trajectory (left). WHAM [81] estimates the wrong walking direction of the person (right). Our approach, COIN, recovers the human and camera motion accurately in both scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.786, + 0.448 + ], + "angle": 0, + "content": "controlled sampling in terms of local body motions. Specifically, COIN with soft inpainting reduces the PA-MPJPE by \\(4.7 \\mathrm{~mm}\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.466, + 0.36, + 0.481 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.658 + ], + "angle": 0, + "content": "In this paper, we propose COIN, a diffusion-based optimization framework for global human and camera motion estimation from dynamic cameras. We identify the inconsistency problem of distilling knowledge from the diffusion model with conventional SDS loss. To address this issue, COIN uses a controlled denoiser combined with soft inpainting to distill a high-quality, well-aligned, and consistent motion prior. To further address the scale ambiguity of the camera trajectory, we develop a novel human-scene relation loss that imposes consistency among the human motion, camera motion, and scene features. Comprehensive experiments on challenging synthetic and real-world datasets demonstrate the effectiveness of COIN, which outperforms the SOTA by a large margin in recovering accurate global human motion and camera motion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Limitations and Future Works: While COIN is able to jointly optimize camera trajectories and global human motions, it requires initialization from SLAM. If the SLAM method fails catastrophically, COIN may not be effective. Additionally, we found COIN fails under severe occlusions where there are several unseen frames and the diffusion model cannot provide consistent guidance. Another limitation is that COIN is an optimization framework, so it is unsuitable for real-time applications. Since the denoising diffusion models have shown their power to model data distributions in many different domains, looking forward we can learn the joint distribution of humans and cameras with the diffusion model to additionally denoise the artifacts in camera motions. Such a joint distribution model also holds the potential of real-time human and camera motion estimation via guided sampling with few DDIM denoising steps." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.493, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "COIN: Control-Inpainting Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. Akhter, I., Black, M.J.: Pose-conditioned joint angle limits for 3D human pose reconstruction. In: CVPR (2015) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.205, + 0.785, + 0.232 + ], + "angle": 0, + "content": "2. Aksan, E., Kaufmann, M., Hilliges, O.: Structured prediction helps 3d human motion modelling. In: ICCV (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.233, + 0.785, + 0.259 + ], + "angle": 0, + "content": "3. Barsoum, E., Kender, J., Liu, Z.: Hp-gan: Probabilistic 3d human motion prediction via gan. In: CVPR Workshops (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.261, + 0.785, + 0.3 + ], + "angle": 0, + "content": "4. Bogo, F., Kanazawa, A., Lassner, C., Gehler, P., Romero, J., Black, M.J.: Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In: ECCV (2016) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.302, + 0.785, + 0.329 + ], + "angle": 0, + "content": "5. Cao, Z., Gao, H., Mangalam, K., Cai, Q.Z., Vo, M., Malik, J.: Long-term human motion prediction with scene context. In: ECCV. pp. 387-404. Springer (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.33, + 0.785, + 0.356 + ], + "angle": 0, + "content": "6. Choi, H., Moon, G., Lee, K.M.: Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In: ECCV (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.358, + 0.785, + 0.384 + ], + "angle": 0, + "content": "7. Choi, H., Moon, G., Lee, K.M.: Beyond static features for temporally consistent 3d human pose and shape from a video. In: CVPR (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.385, + 0.785, + 0.412 + ], + "angle": 0, + "content": "8. Choutas, V., Pavlakos, G., Bolkart, T., Tzionas, D., Black, M.J.: Monocular expressive body regression through body-driven attention. In: ECCV (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.413, + 0.785, + 0.439 + ], + "angle": 0, + "content": "9. Contributors, M.: Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose (2020) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.44, + 0.785, + 0.467 + ], + "angle": 0, + "content": "10. Dabral, R., Shimada, S., Jain, A., Theobalt, C., Golyanik, V.: Gravity-aware monocular 3d human-object reconstruction. In: ICCV (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.468, + 0.785, + 0.506 + ], + "angle": 0, + "content": "11. Fabbri, M., Lanzi, F., Calderara, S., Alletto, S., Cucchiara, R.: Compressed volumetric heatmaps for multi-person 3d pose estimation. In: CVPR (June 2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "12. Fragkiadaki, K., Levine, S., Felsen, P., Malik, J.: Recurrent network models for human dynamics. In: ICCV (2015) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.537, + 0.785, + 0.577 + ], + "angle": 0, + "content": "13. Gartner, E., Andriluka, M., Xu, H., Sminchisescu, C.: Trajectory optimization for physics-based reconstruction of 3d human pose from monocular video. In: CVPR (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.579, + 0.785, + 0.605 + ], + "angle": 0, + "content": "14. Goel, S., Pavlakos, G., Rajasegaran, J., Kanazawa, A., Malik, J.: Humans in 4d: Reconstructing and tracking humans with transformers. In: ICCV (2023) 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.607, + 0.785, + 0.633 + ], + "angle": 0, + "content": "15. Gopalakrishnan, A., Mali, A., Kifer, D., Giles, L., Ororbia, A.G.: A neural temporal model for human motion prediction. In: CVPR (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.634, + 0.785, + 0.66 + ], + "angle": 0, + "content": "16. Guler, R.A., Kokkinos, I.: HoloPose: Holistic 3d human reconstruction in-the-wild. In: CVPR (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "17. Guzov, V., Mir, A., Sattler, T., Pons-Moll, G.: Human POSEitioning system (HPS): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In: CVPR (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.703, + 0.785, + 0.73 + ], + "angle": 0, + "content": "18. Harvey, F.G., Yurick, M., Nowrouzezahrai, D., Pal, C.: Robust motion in betweening. ACM Transactions on Graphics (TOG) 39(4), 60-1 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.731, + 0.785, + 0.757 + ], + "angle": 0, + "content": "19. Hassan, M., Ceylan, D., Villegas, R., Saito, J., Yang, J., Zhou, Y., Black, M.J.: Stochastic scene-aware motion prediction. In: ICCV (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "20. Hassan, M., Choutas, V., Tzionas, D., Black, M.J.: Resolving 3d human pose ambiguities with 3d scene constraints. In: ICCV. pp. 2282-2292 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.812 + ], + "angle": 0, + "content": "21. He, C., Saito, J., Zachary, J., Rushmeier, H., Zhou, Y.: Nemf: Neural motion fields for kinematic animation. In: NeurIPS (2022) 2, 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "22. Henning, D.F., Laidlow, T., Leutenegger, S.: Bodyslam: Joint camera localisation, mapping, and human motion tracking. In: ECCV (2022) 4" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "L. Jiefeng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.147, + 0.784, + 0.175 + ], + "angle": 0, + "content": "23. Hernandez, A., Gall, J., Moreno-Noguer, F.: Human motion prediction via spatiotemporal inpainting. In: CVPR (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "24. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPs (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.203, + 0.785, + 0.245 + ], + "angle": 0, + "content": "25. Huang, C.H.P., Yi, H., Höschle, M., Safroshkin, M., Alexiadis, T., Polikovsky, S., Scharstein, D., Black, M.J.: Capturing and inferring dense full-body human-scene contact. In: CVPR (2022) 3, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.246, + 0.785, + 0.285 + ], + "angle": 0, + "content": "26. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: CVPR (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "27. Iqbal, U., Molchanov, P., Kautz, J.: Weakly-supervised 3d human pose learning via multi-view images in the wild. In: CVPR (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "28. Iqbal, U., Xie, K., Guo, Y., Kautz, J., Molchanov, P.: KAMA: 3D keypoint aware body mesh articulation. In: 3DV (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "29. Isogawa, M., Yuan, Y., O'Toole, M., Kitani, K.M.: Optical non-line-of-sight physics-based 3d human pose estimation. In: CVPR (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "30. Jain, A., Zamir, A.R., Savarese, S., Saxena, A.: Structural-rnn: Deep learning on spatio-temporal graphs. In: CVPR (2016) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "31. Jiang, W., Kolotouros, N., Pavlakos, G., Zhou, X., Daniilidis, K.: Coherent reconstruction of multiple humans from a single image. In: CVPR (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "32. Joo, H., Neverova, N., Vedaldi, A.: Exemplar fine-tuning for 3d human pose fitting towards in-the-wild 3d human pose estimation. In: 3DV (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "33. Kanazawa, A., Black, M.J., Jacobs, D.W., Malik, J.: End-to-end recovery of human shape and pose. In: CVPR (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.48, + 0.785, + 0.508 + ], + "angle": 0, + "content": "34. Kanazawa, A., Zhang, J.Y., Felsen, P., Malik, J.: Learning 3d human dynamics from video. In: CVPR (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "35. Karunratanakul, K., Preechakul, K., Aksan, E., Beeler, T., Suwajanakorn, S., Tang, S.: Optimizing diffusion noise can serve as universal motion priors. arXiv preprint arXiv:2312.11994 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "36. Kaufmann, M., Aksan, E., Song, J., Pece, F., Ziegler, R., Hilliges, O.: Convolutional autoencoders for human motion infilling. In: 3DV (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.578, + 0.785, + 0.618 + ], + "angle": 0, + "content": "37. Kaufmann, M., Song, J., Guo, C., Shen, K., Jiang, T., Tang, C., Zárate, J.J., Hilliges, O.: EMDB: The Electromagnetic Database of Global 3D Human Pose and Shape in the Wild. In: ICCV (2023) 3, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.619, + 0.785, + 0.646 + ], + "angle": 0, + "content": "38. Khurana, T., Dave, A., Ramanan, D.: Detecting invisible people. In: ICCV. pp. 3174-3184 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "39. Kocabas, M., Athanasiou, N., Black, M.J.: VIBE: Video inference for human body pose and shape estimation. In: CVPR (2020) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.714 + ], + "angle": 0, + "content": "40. Kocabas, M., Huang, C.H.P., Tesch, J., Müller, L., Hilliges, O., Black, M.J.: SPEC: Seeing people in the wild with an estimated camera. In: ICCV (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.715, + 0.785, + 0.757 + ], + "angle": 0, + "content": "41. Kocabas, M., Yuan, Y., Molchanov, P., Guo, Y., Black, M.J., Hilliges, O., Kautz, J., Iqbal, U.: PACE: Human and motion estimation from in-the-wild videos. In: 3DV (2024) 1, 2, 3, 4, 9, 10, 11, 12, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "42. Kolotouros, N., Pavlakos, G., Black, M.J., Daniilidis, K.: Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In: ICCV (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "43. Kolotouros, N., Pavlakos, G., Daniilidis, K.: Convolutional mesh regression for single-image human shape reconstruction. In: CVPR (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "44. Kolotouros, N., Pavlakos, G., Jayaraman, D., Daniilidis, K.: Probabilistic modeling for human mesh recovery. In: ICCV (2021) 3" + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.493, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "COIN: Control-Inpainting Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.188 + ], + "angle": 0, + "content": "45. Kundu, J.N., Rakesh, M., Jampani, V., Venkatesh, R.M., Babu1, R.V.: Appearance consensus driven self-supervised human mesh recovery. In: ECCV (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "46. Lassner, C., Romero, J., Kiefel, M., Bogo, F., Black, M.J., Gehler, P.V.: Unite the people: Closing the loop between 3D and 2D human representations. In: CVPR (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.233, + 0.785, + 0.26 + ], + "angle": 0, + "content": "47. Li, J., Bian, S., Xu, C., Liu, G., Yu, G., Lu, C.: D &d: Learning human dynamics from dynamic camera. In: ECCV (2022) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.302 + ], + "angle": 0, + "content": "48. Li, J., Xu, C., Chen, Z., Bian, S., Yang, L., Lu, C.: Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In: CVPR (2021) 3, 5, 8, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.304, + 0.785, + 0.331 + ], + "angle": 0, + "content": "49. Li, Z., Liu, J., Zhang, Z., Xu, S., Yan, Y.: Cliff: Carrying location information in full frames into human pose and shape estimation. In: ECCV (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.785, + 0.373 + ], + "angle": 0, + "content": "50. Li, Z., Zhou, Y., Xiao, S., He, C., Huang, Z., Li, H.: Auto-conditioned recurrent networks for extended complex human motion synthesis. arXiv preprint arXiv:1707.05363 (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.374, + 0.785, + 0.401 + ], + "angle": 0, + "content": "51. Lin, K., Wang, L., Liu, Z.: End-to-end human pose and mesh reconstruction with transformers. In: CVPR (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.403, + 0.785, + 0.43 + ], + "angle": 0, + "content": "52. Liu, M., Yang, D., Zhang, Y., Cui, Z., Rehg, J.M., Tang, S.: 4d human body capture from egocentric video via 3d scene grounding. In: 3DV (2021) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.431, + 0.785, + 0.47 + ], + "angle": 0, + "content": "53. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. SIGGRAPH Asia 34(6), 248:1-248:16 (2015) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.785, + 0.501 + ], + "angle": 0, + "content": "54. Luo, Z., Golestaneh, S.A., Kitani, K.M.: 3d human motion estimation via motion compression and refinement. In: ACCV (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.502, + 0.785, + 0.529 + ], + "angle": 0, + "content": "55. Luo, Z., Hachiuma, R., Yuan, Y., Kitani, K.: Dynamics-regulated kinematic policy for egocentric pose estimation. NeurIPS 34 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.53, + 0.785, + 0.558 + ], + "angle": 0, + "content": "56. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: AMASS: Archive of motion capture as surface shapes. In: ICCV (2019) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.559, + 0.785, + 0.599 + ], + "angle": 0, + "content": "57. von Marcard, T., Henschel, R., Black, M., Rosenhahn, B., Pons-Moll, G.: Recovering accurate 3d human pose in the wild using imus and a moving camera. In: ECCV (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.601, + 0.785, + 0.628 + ], + "angle": 0, + "content": "58. Martinez, J., Black, M.J., Romero, J.: On human motion prediction using recurrent neural networks. In: CVPR (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.629, + 0.785, + 0.67 + ], + "angle": 0, + "content": "59. Mehta, D., Rhodin, H., Casas, D., Fua, P., Sotnychenko, O., Xu, W., Theobalt, C.: Monocular 3d human pose estimation in the wild using improved cnn supervision. In: 3DV (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.671, + 0.785, + 0.713 + ], + "angle": 0, + "content": "60. Mehta, D., Sotnychenko, O., Mueller, F., Xu, W., Elgharib, M., Fua, P., Seidel, H.P., Rhodin, H., Pons-Moll, G., Theobalt, C.: XNect: Real-time multi-person 3D motion capture with a single RGB camera. In: SIGGRAPH (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.714, + 0.785, + 0.755 + ], + "angle": 0, + "content": "61. Mehta, D., Sridhar, S., Sotnychenko, O., Rhodin, H., Shafiei, M., Seidel, H., Xu, W., Casas, D., Theobalt, C.: VNect: Real-time 3D human pose estimation with a single RGB camera. In: SIGGRAPH (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.783 + ], + "angle": 0, + "content": "62. Moon, G., Chang, J.Y., Lee, K.M.: Camera distance-aware top-down approach for 3d multi-person pose estimation from a single rgb image. In: ICCV (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.812 + ], + "angle": 0, + "content": "63. Moon, G., Lee, K.M.: I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In: ECCV (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "64. Müller, L., Osman, A.A.A., Tang, S., Huang, C.H.P., Black, M.J.: On self contact and human pose. In: CVPR (2021) 3" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "L. Jiefeng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "65. Müller, L., Ye, V., Pavlakos, G., Black, M., Kanazawa, A.: Generative proxemics: A prior for 3d social interaction from images. arXiv preprint arXiv:2306.09337 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.189, + 0.787, + 0.217 + ], + "angle": 0, + "content": "66. Opensfm - a structure from motion library. https://github.com/mapillary/OpenSfM (2021), https://github.com/mapillary/OpenSfM 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.217, + 0.787, + 0.257 + ], + "angle": 0, + "content": "67. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: CVPR (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.257, + 0.787, + 0.284 + ], + "angle": 0, + "content": "68. Pavlakos, G., Kolotouros, N., Daniilidis, K.: Texturepose: Supervising human mesh estimation with texture consistency. In: ICCV (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.284, + 0.787, + 0.311 + ], + "angle": 0, + "content": "69. Pavlakos, G., Weber, E., Tancik, M., Kanazawa, A.: The one where they reconstructed 3d humans and environments in tv shows. In: ECCV (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.311, + 0.787, + 0.338 + ], + "angle": 0, + "content": "70. Pavlakos, G., Zhu, L., Zhou, X., Daniilidis, K.: Learning to estimate 3D human pose and shape from a single color image. In: CVPR (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.338, + 0.787, + 0.365 + ], + "angle": 0, + "content": "71. Pavlo, D., Grangier, D., Auli, M.: Quaternet: A quaternion-based recurrent model for human motion. In: BMVC (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.365, + 0.787, + 0.407 + ], + "angle": 0, + "content": "72. Payer, C., Neff, T., Bischof, H., Urschler, M., Stern, D.: Simultaneous multi-person detection and single-person pose estimation with a single heatmap regression network. In: ICCV PoseTrack Workshop (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.407, + 0.787, + 0.433 + ], + "angle": 0, + "content": "73. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: ICCV (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.433, + 0.787, + 0.46 + ], + "angle": 0, + "content": "74. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: ICLR (2023) 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.46, + 0.787, + 0.5 + ], + "angle": 0, + "content": "75. Reddy, N.D., Guigues, L., Pischulini, L., Eledath, J., Narasimhan, S.: Tessetrack: End-to-end learnable multi-person articulated 3d pose tracking. In: CVPR (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.501, + 0.787, + 0.528 + ], + "angle": 0, + "content": "76. Rempe, D., Birdal, T., Hertzmann, A., Yang, J., Sridhar, S., Guibas, L.J.: Humor: 3d human motion model for robust pose estimation. In: ICCV (2021) 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.528, + 0.787, + 0.556 + ], + "angle": 0, + "content": "77. Rogez, G., Weinzaepfel, P., Schmid, C.: LCR-Net: Localization-classification-regression for human pose. In: CVPR (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.556, + 0.787, + 0.582 + ], + "angle": 0, + "content": "78. Rong, Y., Liu, Z., Li, C., Cao, K., Change Loy, C.: Delving deep into hybrid annotations for 3d human recovery in the wild. In: ICCV (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.582, + 0.787, + 0.623 + ], + "angle": 0, + "content": "79. Sárándi, I., Hermans, A., Leibe, B.: Learning 3D human pose estimation from dozens of datasets using a geometry-aware autoencoder to bridge between skeleton formats. In: WACV (2023) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.623, + 0.787, + 0.651 + ], + "angle": 0, + "content": "80. Shimada, S., Golyanik, V., Xu, W., Theobalt, C.: Physcap: Physically plausible monocular 3d motion capture in real time. In: SIGGRAPH (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.651, + 0.787, + 0.677 + ], + "angle": 0, + "content": "81. Shin, S., Kim, J., Halilaj, E., Black, M.J.: Wham: Reconstructing world-grounded humans with accurate 3d motion. In: CVPR (2024) 1, 2, 3, 4, 10, 11, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.677, + 0.787, + 0.704 + ], + "angle": 0, + "content": "82. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.705, + 0.787, + 0.731 + ], + "angle": 0, + "content": "83. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2021) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.731, + 0.787, + 0.759 + ], + "angle": 0, + "content": "84. Song, J., Chen, X., Hilliges, O.: Human body model fitting by learned gradient descent. In: ECCV (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.759, + 0.787, + 0.786 + ], + "angle": 0, + "content": "85. Sun, Y., Bao, Q., Liu, W., Fu, Y., Black, M.J., Mei, T.: Monocular, one-stage, regression of multiple 3d people. In: ICCV (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.786, + 0.787, + 0.813 + ], + "angle": 0, + "content": "86. Sun, Y., Bao, Q., Liu, W., Mei, T., Black, M.J.: Trace: 5d temporal regression of avatars with dynamic cameras in 3d environments. In: CVPR (2023) 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "87. Sun, Y., Liu, W., Bao, Q., Fu, Y., Mei, T., Black, M.J.: Putting people in their place: Monocular regression of 3D people in depth. In: CVPR (2022) 3" + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.493, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "COIN: Control-Inpainting Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "88. Sun, Y., Ye, Y., Liu, W., Gao, W., Fu, Y., Mei, T.: Human mesh recovery from monocular images via a skeleton-disentangled representation. In: ICCV (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "89. Teed, Z., Deng, J.: DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. NeurIPs (2021) 2, 5, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.219 + ], + "angle": 0, + "content": "90. Teed, Z., Lipson, L., Deng, J.: Deep patch visual odometry. NeurIPS (2023) 2, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.22, + 0.785, + 0.246 + ], + "angle": 0, + "content": "91. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. In: ICLR 2023 (2022) 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.248, + 0.785, + 0.275 + ], + "angle": 0, + "content": "92. Villegas, R., Yang, J., Zou, Y., Sohn, S., Lin, X., Lee, H.: Learning to generate long-term future via hierarchical prediction. In: ICML (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.277, + 0.785, + 0.303 + ], + "angle": 0, + "content": "93. Weng, Z., Yeung, S.: Holistic 3d human and scene mesh estimation from single view images. In: CVPR (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.305, + 0.785, + 0.332 + ], + "angle": 0, + "content": "94. Xiang, D., Joo, H., Sheikh, Y.: Monocular total capture: Posing face, body and hands in the wild. In: CVPR (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.334, + 0.785, + 0.36 + ], + "angle": 0, + "content": "95. Xie, K., Wang, T., Iqbal, U., Guo, Y., Fidler, S., Shkurti, F.: Physics-based human motion estimation and synthesis from videos. In: ICCV (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.362, + 0.785, + 0.388 + ], + "angle": 0, + "content": "96. Xie, Y., Jampani, V., Zhong, L., Sun, D., Jiang, H.: Omnicontrol: Control any joint at any time for human motion generation. In: ICLR (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.39, + 0.785, + 0.417 + ], + "angle": 0, + "content": "97. Xu, Y., Zhu, S.C., Tung, T.: Denserac: Joint 3d pose and shape estimation by dense render-and-compare. In: ICCV (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.419, + 0.785, + 0.459 + ], + "angle": 0, + "content": "98. Yan, X., Rastogi, A., Villegas, R., Sunkavalli, K., Shechtman, E., Hadap, S., Yumer, E., Lee, H.: Mt-vae: Learning motion transformations to generate multimodal human dynamics. In: ECCV (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.785, + 0.487 + ], + "angle": 0, + "content": "99. Ye, V., Pavlakos, G., Malik, J., Kanazawa, A.: Decoupling human and camera motion from videos in the wild. In: CVPR (2023) 2, 4, 10, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.489, + 0.785, + 0.529 + ], + "angle": 0, + "content": "00. Yuan, Y., Iqbal, U., Molchanov, P., Kitani, K., Kautz, J.: Glamr: Global occlusion-aware human mesh recovery with dynamic cameras. In: CVPR (2022) 2, 4, 10, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.531, + 0.785, + 0.558 + ], + "angle": 0, + "content": "01. Yuan, Y., Kitani, K.: Diverse trajectory forecasting with determinantal point processes. In: ICLR 2020 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.56, + 0.785, + 0.586 + ], + "angle": 0, + "content": "02. Yuan, Y., Kitani, K.: Dlow: Diversifying latent flows for diverse human motion prediction. In: ECCV (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.588, + 0.785, + 0.614 + ], + "angle": 0, + "content": "03. Yuan, Y., Kitani, K.: Residual force control for agile human behavior imitation and extended motion synthesis. In: NeurIPS (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.616, + 0.785, + 0.643 + ], + "angle": 0, + "content": "04. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: ICCV (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.645, + 0.785, + 0.671 + ], + "angle": 0, + "content": "05. Yuan, Y., Wei, S.E., Simon, T., Kitani, K., Saragih, J.: Simpoe: Simulated character control for 3d human pose estimation. In: CVPR (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.785, + 0.713 + ], + "angle": 0, + "content": "06. Zanfir, A., Bazavan, E.G., Xu, H., Freeman, W.T., Sukthankar, R., Sminchisescu, C.: Weakly supervised 3d human pose and shape reconstruction with normalizing flows. In: ECCV (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.715, + 0.785, + 0.755 + ], + "angle": 0, + "content": "07. Zanfir, A., Marinoiu, E., Sminchisescu, C.: Monocular 3d pose and shape estimation of multiple people in natural scenes the importance of multiple scene constraints. In: CVPR (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "08. Zanfir, A., Marinoiu, E., Zanfir, M., Popa, A.I., Sminchisescu, C.: Deep network for the integrated 3d sensing of multiple people in natural images. In: NeurIPS (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "09. Zanfir, M., Zanfir, A., Bazavan, E.G., Freeman, W.T., Sukthankar, R., Sminchisscu, C.: Thundr: Transformer-based 3d human reconstruction with markers. In: ICCV (2021) 3" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "L. Jiefeng et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "110. Zhang, H., Tian, Y., Zhou, X., Ouyang, W., Liu, Y., Wang, L., Sun, Z.: Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In: ICCV (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.217 + ], + "angle": 0, + "content": "111. Zhang, J., Yu, D., Liew, J.H., Nie, X., Feng, J.: Body meshes as points. In: CVPR (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.787, + 0.244 + ], + "angle": 0, + "content": "112. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models (2023) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.245, + 0.787, + 0.286 + ], + "angle": 0, + "content": "113. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.327 + ], + "angle": 0, + "content": "114. Zhang, S., Bhatnagar, B.L., Xu, Y., Winkler, A., Kadlecek, P., Tang, S., Bogo, F.: Rohm: Robust human motion reconstruction via diffusion. In: CVPR (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.356 + ], + "angle": 0, + "content": "115. Zhang, S., Zhang, Y., Bogo, F., Pollefeys, M., Tang, S.: Learning motion priors for 4d human body capture in 3d scenes. In: ICCV (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.383 + ], + "angle": 0, + "content": "116. Zhang, T., Huang, B., Wang, Y.: Object-occluded human shape and pose estimation from a single color image. In: CVPR (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.411 + ], + "angle": 0, + "content": "117. Zhen, J., Fang, Q., Sun, J., Liu, W., Jiang, W., Bao, H., Zhou, X.: SMAP: Single-shot multi-person absolute 3d pose estimation. In: ECCV (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.412, + 0.787, + 0.449 + ], + "angle": 0, + "content": "118. Zhou, Y., Habermann, M., Habibie, I., Tewari, A., Theobalt, C., Xu, F.: Monocular real-time full body capture with inter-part correlations. In: CVPR (2021) 3" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.449 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/COIN_ Control-Inpainting Diffusion Prior for Human and Camera Motion Estimation/213c3e91-2355-4081-9e3c-28b7d1b0600c_origin.pdf b/2024/COIN_ Control-Inpainting Diffusion Prior for Human and Camera Motion Estimation/213c3e91-2355-4081-9e3c-28b7d1b0600c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..37beacccce4fc2f5c33dd70c8d398ed6dc355444 --- /dev/null +++ b/2024/COIN_ Control-Inpainting Diffusion Prior for Human and Camera Motion Estimation/213c3e91-2355-4081-9e3c-28b7d1b0600c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae7e09fd3682c4a0a35874fcc0154b3828c1ae021e4460cced47089dc6853e48 +size 6581292 diff --git a/2024/COIN_ Control-Inpainting Diffusion Prior for Human and Camera Motion Estimation/full.md b/2024/COIN_ Control-Inpainting Diffusion Prior for Human and Camera Motion Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6965e51a6dc4cc4d48056930bc7e4affe070aeb0 --- /dev/null +++ b/2024/COIN_ Control-Inpainting Diffusion Prior for Human and Camera Motion Estimation/full.md @@ -0,0 +1,374 @@ +# COIN: Control-Inpainting Diffusion Prior for Human and Camera Motion Estimation + +Jiefeng Li $^{1,2}$ , Ye Yuan $^{1}$ , Davis Rempe $^{1}$ , Haotian Zhang $^{1}$ , Pavlo Molchanov $^{1}$ , Cewu Lu $^{2}$ , Jan Kautz $^{1}$ , and Umar Iqbal $^{1}$ + +$^{1}$ NVIDIA $^{2}$ Shanghai Jiao Tong University https://nvlabs.github.io/COIN/ + +![](images/9fb51b864a9ead7976a41d75c7841abdf6d41033d1beac89fc244f7e0bf84e22.jpg) +Fig. 1: Capturing global human and camera motion from a dynamic camera presents unique challenges. In the input video, a person is riding a skateboard - while the local body motion may remain relatively constant, the global position of the individual changes significantly. Current state-of-the-art methods such as PACE [41] and WHAM [81] fail catastrophically on such out-of-distribution motions. Our approach, COIN, gracefully handles such challenging cases, owing to our control-inpainting motion diffusion prior and novel human-scene relation loss. + +Abstract. Estimating global human motion from moving cameras is challenging due to the entanglement of human and camera motions. To mitigate the ambiguity, existing methods leverage learned human motion priors, which however often result in oversmoothed motions with misaligned 2D projections. To tackle this problem, we propose COIN, a control-inpainting motion diffusion prior that enables fine-grained control to disentangle human and camera motions. Although pre-trained motion diffusion models encode rich motion priors, we find it non-trivial to leverage such knowledge to guide global motion estimation from RGB videos. COIN introduces a novel control-inpainting score distillation sampling method to ensure well-aligned, consistent, and high-quality motion from the diffusion prior within a joint optimization framework. Furthermore, we introduce a new human-scene relation loss to alleviate the scale ambiguity by enforcing consistency among the humans, camera, and scene. Experiments on three challenging benchmarks demonstrate the effectiveness of COIN, which outperforms the state-of-the-art methods in terms of global human motion estimation and camera motion estimation. As an illustrative example, COIN outperforms the state-of-the-art method by $33\%$ in world joint position error (W-MPJPE) on the RICH dataset. + +# 1 Introduction + +Recovering global human and camera motion from dynamic RGB videos is an important problem with many applications, such as animation, human-computer interaction, mixed reality, and robotics. However, it is a very challenging problem due to the entanglement of human and camera motion. + +There are only a few works [41, 81, 99, 100] that try to address this problem. Earlier methods [47, 100] only focus on human motion and ignore the camera motion. Their core insight is that the global body motion is highly correlated with the local motion. Thus, they can use the local body movements to estimate the global orientation and trajectory with a regression model [100] or by combining them with physics constraints [47]. However, these regression models ignore the camera movements, so they fail to maintain consistency with the input video, whereas physics-based methods fail to model complex in-the-wild environments so are limited to controlled scenarios. Recent works [41, 99] try to jointly estimate the human and camera motion by exploiting learned motion priors [21, 76] and SLAM [66, 89, 90]. They try to constrain the human body motion in a low-dimensional latent space of a motion prior model, which results in reconstructed motions that are overly smooth and do not align well with video observations. Moreover, the optimization of the camera motion is only based on the global human motion from the motion prior. Hence, they fail catastrophically if the initial human motion predictions are significantly incorrect (as shown in Fig. 1). + +More recently, Denoising Diffusion Models [24,91] have emerged as a powerful family of generative models that can model high-quality data priors. Nonetheless, effectively leveraging the learned priors remains an ongoing challenge. Score Distillation Sampling (SDS) is commonly employed for this purpose [74], but we find that naive application of SDS also results in inconsistencies with the available observations (see Sec. 4.2). The root cause of this problem lies in the inconsistency of randomly sampled motions during SDS optimization. Without constraints, these motions may not align with observed evidence, leading to overly smoothed results that lack detail due to the mode-averaging effect. + +In this work, we propose COIN, a hybrid COntrol-INpainting score distillation sampling method to address the aforementioned limitations of vanilla SDS. First, we use the partially observed human motion from the video as control signals to guide motion sampling. To tackle noisy observations which may be out of distribution for the prior, we propose a dynamic controlled sampling technique that iteratively refines the observed motions and updates the control signals to ensure effective distillation from the motion prior. Second, to further improve the consistency of the sampled motions, we also develop a novel soft inpainting strategy. We automatically identify the high-confidence regions of the initial predicted global motion from the video and use them as soft constraints during optimization. Concretely, we sample less confident regions from scratch using the + +motion model, while the confident regions are only slightly refined. This ensures that the reconstructed motions do not deviate from the available observations. Our new SDS formulation is used to jointly optimize the human and camera motion by finding the most plausible solution that explains the observed evidence. Finally, to prevent catastrophic failure in cases where the initial body motion fails significantly, we propose a human-scene relation loss to consider the human-scene depth relations. This novel loss provides complementary information to the human motion prior by using local motion and scene features. It regularizes the camera scale by enforcing consistency among the human motion, camera motion, and scene features. + +We benchmark our approach on the synthetic HCM [41] dataset and the real-world RICH [25] and EMDB [37] datasets. We demonstrate that our approach significantly outperforms the state-of-the-art methods in terms of human motion estimation and camera motion estimation. Overall, the contributions of this paper can be summarized as follows: + +- We propose a novel control-inpainting motion prior specifically designed for global human motion estimation, which enhances score distillation sampling with dynamic control and soft inpainting to reconstruct well-aligned, consistent, and high-quality motions from video observations. +- We propose a new human-scene relation loss to resolve the scale ambiguity of the camera motion by enforcing consistency among the human motion, camera motion, and scene features. +- Our approach significantly outperforms the state-of-the-art methods in terms of human motion estimation and camera motion estimation on both synthetic and real-world datasets. In terms of global human motion estimation in world space, we outperform the state-of-the-art method PACE [41] by $44\%$ and $33\%$ on the HCM [41] and RICH [25] datasets, respectively. We also compare with the contemporary work WHAM [81] and outperform it by $49\%$ and $7\%$ on the RICH and EMDB datasets, respectively. + +# 2 Related Work + +Camera-Space Human Pose Estimation. Most existing works focus on root-relative local human pose estimation to bypass the difficulty in monocular depth estimation [1, 4, 6-8, 16, 32-34, 39, 42-46, 51, 54, 63, 64, 67, 68, 70, 76, 78, 84, 85, 88, 94, 97, 106, 110, 116, 118]. These methods ignore the position of the person in the camera coordinates. To overcome this limitation, recent methods estimate camera-space human poses by regression [28, 31, 48, 52, 62, 72, 75, 80, 93, 95, 107, 107, 109, 111] or optimization [59-61, 77, 108]. Physics-based constraints are widely used to ensure the plausibility of the estimated poses [10, 13, 27, 29, 80, 95, 105]. In addition to direct regression, heatmap-based representations have also been used to predict the absolute depths of multiple people [11, 87, 117]. A few methods improve absolution depth estimation by using predicted camera parameters [40, 49, 109] instead of the predefined focal length. Despite the promising results + +for camera-space pose estimation, how to decouple the camera movement and estimate global human motions is still an open problem. + +Monocular Global Human Pose Estimation. Recovering the global human motion from a monocular moving camera is challenging due to the entanglement of human and camera motions. To disentangle the camera movement, several methods use IMU sensors or pre-scanned environments to recover global human motions [17,20,57,69], which is impractical for large-scale adoption. Recent works use human motion priors [100] or physics-based constraints [47,55] to recover human motions from monocular videos, but do not consider background scene features, which limits performance on in-the-wild videos. Sun et al. [86] use optical flow as a motion cue to estimate the global motions. A contemporary work, WHAM [81], uses a lifting network to estimate global human motions from 2D keypoints and camera angular velocities. While these works can estimate accurate global human motions, they do not recover camera motions. To explicitly recover the camera motion, Liu et al. [52] use SLAM and convert the local pose from the camera to global coordinates. BodySLAM [22] jointly optimizes the human and camera motion using features of both humans and scenes. Along this line, SLAHMR [99] and PACE [41] use SLAM to initialize camera motions and optimize the camera using human motion priors [21,76]. However, these methods rely on the human motion priors to regularize the camera motion, which may lead to inaccurate camera motion when the human motion is not well initialized (as shown in Fig. 1). Such wrong camera trajectories will further affect the optimization of human motions. In contrast, our approach relies on the consistency among the local human motion, scene features, and the camera for optimization, which provides information that complements the human motion priors and enables accurate estimation of both human and camera motions. + +Human Motion Priors. There are a significant amount of approaches proposed to study 3D human dynamics for motion prediction and synthesis [2, 3, 5, 12, 15, 18, 19, 23, 30, 36, 38, 50, 58, 71, 73, 92, 98, 101-103]. These learned human motion priors are used to help resolve pose ambiguity [21, 39, 76, 115] in human pose estimation. Recently, diffusion models [82] have also been used as priors for motion synthesis and infilling [26, 35, 91, 104, 113]. RoHM [114], adopts motion diffusion model to recover human motions from noisy and occluded input data. Xie et al. [96] use spatial control signals to guide motion generation. They focus on generating realistic human motion given clean spatial constraints. Müller et al. [65] build a diffusion model to learn the joint distribution over the poses of two people. They use the SDS loss to guide the generation of static poses. In contrast, we focus on dynamic human motions. We find adopting SDS directly for temporal human motions encounters the inconsistency issue. Therefore, to distill knowledge from the motion diffusion model, we propose a novel control inpainting SDS to generate high-quality and consistent motion that aligns with observed evidence. + +![](images/de1e8367b8e3c658791af30b4878cabc66b63e7163ab60536295821c1707354f.jpg) +Fig. 2: Overview. Given a video with a moving camera, we recover the global human motion $\mathbf{H}$ and camera motion $\mathcal{C}$ using an iterative optimization framework. We propose a novel Control-Inpainting SDS loss ( $\mathcal{L}_{\mathrm{COIN - SDS}}$ ) to leverage motion diffusion models as a prior. COIN-SDS is designed such that the sampled motions from the motion prior are consistent with video observations. We achieve this by controlling and constraining the sampling process of the motion diffusion model through novel control and soft-inpainting branches. We also propose a novel human-scene relation loss ( $\mathcal{L}_{\mathrm{HSR}}$ ) to encourage consistency among the human motion, camera motion, and scene features. + +# 3 Method + +The overall framework of COIN is illustrated in Fig. 2. Given an in-the-wild RGB video with $T$ frames captured by a dynamic camera, our goal is to estimate both the global human motion $\mathbf{H} = \{\mathbf{h}^{(1)},\mathbf{h}^{(2)},\dots,\mathbf{h}^{(T)}\}$ and the camera motion $\mathcal{C} = \{\mathbf{c}^{(1)},\mathbf{c}^{(2)},\dots,\mathbf{c}^{(T)}\}$ in a global world coordinate system. We use off-the-shelf 3D human pose and shape estimation method HybrIK [48] to obtain per-frame initial SMPL parameters in the camera space and DROID-SLAM [89] to obtain the initial per-frame camera-to-world transforms. We convert the local human motion to the world coordinates with the estimated camera. However, because the camera trajectories from SLAM are up to an unknown scale, the initial global human motion will abnormally drift and float in the world space. To resolve the scale ambiguity and place the person in the correct global position, we jointly optimize the human and camera motion to minimize the discrepancy between the observed evidence and the estimated motion, while maintaining the plausibility of the human motion with a diffusion prior through the proposed control-inpainting SDS. + +Motion Representation. The camera motion is represented by the trajectory $\mathcal{C} = \{(\mathbf{R}^{(i)},\mathbf{t}^{(i)})\}_{i = 1}^{T}$ , where $[\mathbf{R}^{(i)},\mathbf{t}^{(i)}]$ is the camera pose at the $i$ -th frame, consisting of the rotation matrix $\mathbf{R}^{(i)}\in \mathbb{R}^{3\times 3}$ and the translation vector $\mathbf{t}^{(i)}\in \mathbb{R}^3$ . The human motion is represented by the human trajectory $\mathbf{H} = \{\mathbf{h}^{(i)}\}_{i = 1}^{T}$ where $\mathbf{h}^{(i)} = [\tau^{(i)},\varPhi^{(i)},\theta^{(i)},f^{(i)},\beta ]$ is the human pose at the $i$ -th frame, consisting + +of the global translation $\tau^{(i)}\in \mathbb{R}^3$ , global orientation $\varPhi^{(i)}\in \mathbb{R}^3$ , body pose parameters $\theta^{(i)}\in \mathbb{R}^{23\times 3}$ , foot contact labels $f\in \{0,1\} ^4$ , and the body shape parameters $\beta \in \mathbb{R}^{10}$ . We use the SMPL model [53] to represent the human pose and shape. Before introducing our approach, we first revisit the formulation and drawbacks of SDS in Sec. 3.1. Then we introduce the proposed control-inpainting SDS in Sec. 3.2. Finally, we present the global optimization pipeline with the proposed human-scene interaction loss in Sec. 3.3. + +# 3.1 Revisiting SDS + +Score Distillation Sampling (SDS) was first introduced to distill 3D assets from pre-trained 2D text-to-image diffusion models [74]. It exploits the knowledge from the diffusion models by seeking modes for the conditional distribution in the DDPM latent space to optimize the 3D scene representation. Similarly, we can optimize global human motion by distilling knowledge from a pre-trained motion diffusion model. + +Given an global human motion $\mathbf{H}$ , the marginal distribution of noisy latent $\mathbf{H}_t$ at timestep $t \in \mathcal{U}(0,1)$ is defined as: + +$$ +q (\mathbf {H} _ {t} | \mathbf {H}) = \mathcal {N} (\mathbf {H} _ {t}; \sqrt {\bar {\alpha} _ {t}} \mathbf {H}, (1 - \bar {\alpha} _ {t}) \mathbf {I}), \tag {1} +$$ + +where $\bar{\alpha}_t\in (0,1)$ is a hyperparameter controlled by the variance schedule of the diffusion model. SDS adopts the pre-trained diffusion model $\mathcal{D}_{\phi}(\mathbf{H}_t,t,y)$ , which takes in $\mathbf{H}_t$ and is used to model the conditional density of the human motion, where $\phi$ are the parameters of the diffusion model and $y$ is the condition. Then, SDS aims to distill global human motion $\mathbf{H}$ via seeking modes of the learned conditional density, which can be achieved by a weighted denoising score matching objective: + +$$ +\underset {\mathbf {H}} {\min } \mathcal {L} _ {\mathrm {S D S}} := \mathbb {E} _ {t, \epsilon} \left[ \omega (t) \| \epsilon_ {\phi} ^ {t} - \epsilon \| _ {2} ^ {2} \right], \tag {2} +$$ + +where $\epsilon_{\phi}^{t}$ is the predicted denoising direction from the diffusion model, $\mathbf{H}_t \sim q(\mathbf{H}_t|\mathbf{H})$ is sampled using the reparameterization trick, $\epsilon$ is the corresponding sampled noise, and $\omega(t)$ is a weighting function that depends on the timestep $t$ . + +To clearly review the effect of SDS, we can reparameterize Eq. 2 as: + +$$ +\min _ {\mathbf {H}} \mathcal {L} _ {\mathrm {S D S}} := \mathbb {E} _ {t} \left[ \frac {\omega (t) \sqrt {\bar {\alpha} _ {t}}}{\sqrt {1 - \bar {\alpha} _ {t}}} \left\| \mathbf {H} - \hat {\mathbf {H}} _ {0} ^ {t} \right\| _ {2} ^ {2} \right], \tag {3} +$$ + +where + +$$ +\hat {\mathbf {H}} _ {0} ^ {t} = \frac {\mathbf {H} _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \boldsymbol {\epsilon} _ {\phi} ^ {t}}{\sqrt {\bar {\alpha} _ {t}}}. \tag {4} +$$ + +Based on this reparameterization, we can see that the SDS objective is to minimize the discrepancy between global human motion $\mathbf{H}$ and the denoised global human motion $\hat{\mathbf{H}}_0^t$ from the motion diffusion model in a single step. The denoised motion $\hat{\mathbf{H}}_0^t$ serves as the pseudo ground truth. However, at each optimization step, we randomly sample $t$ and $\epsilon$ to generate the noisy latent $\mathbf{H}_t$ , and we found + +the pre-trained diffusion model is sensitive to the input. Minor fluctuations in the input latent would substantially change the denoised motion, which leads to inconsistency in $\hat{\mathbf{H}}_0^t$ across different time steps. + +Although randomness can help generate diverse plausible motions to infer occluded regions and unknown information, we do not need it for well-observed regions, such as simple body poses in a clean background. Such randomness in the denoising steps makes the generated $\hat{\mathbf{H}}_0^t$ difficult to align with the local 2D observations and results in wrong global human motion. Moreover, this pseudo ground truth $\hat{\mathbf{H}}_0^t$ is generated from only a single denoising step, where the diffusion models may not produce high-quality motions, resulting in foot sliding and floating. Although sampling with a smaller timestep $t$ can alleviate these issues, the initial motion is usually inaccurate and the denoiser is not able to remove artifacts with a small $t$ . To exploit the knowledge of the motion diffusion model and denoise the initial motion, we must allow the SDS to sample with a larger timestep $t$ while maintaining high quality, consistency, and alignment with the local 2D observations. + +# 3.2 Control-Inpainting SDS + +The limitations of SDS originate from the randomness and inconsistency of the denoised motion $\hat{\mathbf{H}}_0^t$ , which serves as the pseudo ground truth in the objective function. To overcome this issue, we propose a novel COntrol-INpainting SDS (COIN-SDS) to generate high-quality and consistent pseudo-ground-truth motions. Our solution has three key ingredients (shown in Alg. 1). First, to achieve high-quality motions, we seek to produce the pseudo ground truth with multiple DDIM denoising steps. Second, to encourage consistent motions, we propose to use partially observed evidence from the video as a control signal to dynamically guide the diffusion model and align the generated motions with the observations. Third, to further align the motion with observed regions, we propose a soft inpainting strategy within the denoising process. + +Multiple Denoising Steps. Intuitively, to obtain high-quality pseudo ground truth for SDS, we can replace the single-step denoised motion $\hat{\mathbf{H}}_0^t$ with a multi-step one $\tilde{\mathbf{H}}_0^t \coloneqq \tilde{\mathbf{H}}_0$ , following the multi-step DDIM denoising process [83]: + +$$ +\tilde {\mathbf {H}} _ {t - \Delta t} = \sqrt {\bar {\alpha} _ {t - \Delta t}} \cdot \hat {\mathbf {H}} _ {0} ^ {t} + \sqrt {1 - \bar {\alpha} _ {t - \Delta t}} \cdot \boldsymbol {\epsilon} _ {\phi} ^ {t}, \tag {5} +$$ + +until $\tilde{\mathbf{H}}_0 = \tilde{\mathbf{H}}_{t - \Delta t}$ is obtained. By replacing $\tilde{\mathbf{H}}_0^t$ in Eq. 3 with $\tilde{\mathbf{H}}_0$ , we can obtain a new objective for SDS: + +$$ +\min _ {\mathbf {H}} \mathcal {L} _ {\mathrm {S D S}} := \mathbb {E} _ {t} \left[ \frac {\omega (t) \sqrt {\bar {\alpha} _ {t}}}{\sqrt {1 - \bar {\alpha} _ {t}}} \left\| \mathbf {H} - \tilde {\mathbf {H}} _ {0} \right\| _ {2} ^ {2} \right]. \tag {6} +$$ + +Although the multi-step denoising process can produce high-quality pseudo ground truth, it is computationally expensive to perform multiple denoising steps during optimization, which limits the practicality of increasing the number of denoising steps. In our experiments, we find that using 10 denoising steps is sufficient to produce high-quality pseudo ground truth. + +Algorithm 1: COIN-SDS +Input: Latest human motion H, confidence score S, visible mask M +Output: $\mathcal{L}_{\mathrm{COIN - SDS}}$ +1 Sample: $\tilde{\mathbf{H}}_t \sim \mathcal{N}(\mathbf{H}_t; \sqrt{\bar{\alpha}_t}\mathbf{H}, (1 - \bar{\alpha}_t)\mathbf{I}), t \sim \mathcal{U}(0,1)$ ; +2 for $\bar{t} = [t, t - \Delta t, \dots, \Delta t]$ do // multi-step DDIM denoising +3 $\tilde{\mathbf{H}}_0^{\bar{t},\mathrm{known}} \gets \mathbf{H}$ ; +4 $\tilde{\mathbf{H}}_0^{\bar{t},\mathrm{un-known}} \gets \mathcal{D}_{\phi,\phi_c}(\tilde{\mathbf{H}}_{\bar{t}}, \bar{t}, \mathbf{H} \odot \mathbf{M})$ ; // controlled denoising +5 $\tilde{\mathbf{M}} \gets w(\bar{t}) * \mathbf{S} \odot \mathbf{M}$ ; +6 $\tilde{\mathbf{H}}_0^{\bar{t}} \gets \tilde{\mathbf{M}} \odot \tilde{\mathbf{H}}_0^{\bar{t},\mathrm{known}} + (1 - \tilde{\mathbf{M}}) \odot \tilde{\mathbf{H}}_0^{\bar{t},\mathrm{unknown}}$ ; // soft inpainting +7 $\epsilon_{\phi}^{\bar{t}} \gets \frac{\tilde{\mathbf{H}}_{\bar{t}} - \sqrt{\bar{\alpha}_{\bar{t}}} \tilde{\mathbf{H}}_0^{\bar{t}}}{\sqrt{1 - \bar{\alpha}_{\bar{t}}}}$ ; +8 $\tilde{\mathbf{H}}_{\bar{t} -\Delta t} \gets \sqrt{\bar{\alpha}_{\bar{t} -\Delta t}} \cdot \tilde{\mathbf{H}}_0^{\bar{t}} + \sqrt{1 - \bar{\alpha}_{\bar{t} -\Delta t}} \cdot \epsilon_{\phi}^{\bar{t}}$ ; // update latent motion +9 $\mathcal{L}_{\mathrm{COIN - SDS}} = \frac{\omega(t) \sqrt{\bar{\alpha}_t}}{\sqrt{1 - \bar{\alpha}_t}}\left\| \mathbf{H} - \tilde{\mathbf{H}}_0 \right\|_2^2$ + +Dynamic Controlled Sampling. To generate consistent motions that are aligned with the observed evidence, we propose to attach a control branch $\phi_c$ to the pre-trained diffusion model $\mathcal{D}_{\phi}$ to guide the motion generation. Given a latent motion $\tilde{\mathbf{H}}_t$ , control signal $\mathbf{c}$ , and the visible mask $\mathbf{M}$ , we train a controlled denoiser $\mathcal{D}_{\phi,\phi_c}$ to generate intermediate denoised motion $\tilde{\mathbf{H}}_0^t$ for DDIM denoising: + +$$ +\tilde {\mathbf {H}} _ {0} ^ {t} = \mathcal {D} _ {\phi , \phi_ {c}} \left(\tilde {\mathbf {H}} _ {t}, t, \mathbf {c} \odot \mathbf {M}\right), \tag {7} +$$ + +where $\mathbf{c}$ and $\mathbf{M}$ are the same size as the motion, $\mathbf{M}$ is a binary mask with ones in observed pose dimensions, and $\odot$ denotes the element-wise multiplication. During training, we synthesize noise and occlusions by randomly adding Gaussian noise to the control signal $\mathbf{c}$ and randomly masking the pose and trajectory dimensions in $\mathbf{M}$ . Details are provided in the supplementary document. + +When using the denoiser in the optimization stage, instead of always using the initial noisy estimation from HybrIK [48] as the fixed control signal, we propose to use a dynamic control strategy. Specifically, we use the optimized human motion from the previous iteration as the control signal, i.e., $\mathbf{c} = \mathbf{H}$ . This strategy prevents performance degradation due to inaccurate initializations. A better control signal can guide the model to generate a more plausible pseudo motion, which in turn helps to optimize the global human motion and provides a control signal that better aligns with the input videos. Such self-evolving control signals can help to generate well-aligned global human motions. + +The pre-trained motion diffusion model adopts a transformer encoder structure like [91]. We follow ControlNet [112] to encode the control signals and guide the denoiser output. We create a trainable copy of 4 encoding blocks of the pretrained motion diffusion model followed with zero convolutions. The input to the control branch is the concatenation of the latent and the control signals. The AMASS [56] dataset is used to train the motion diffusion model. The finetuning of the controlled denoiser is computationally efficient since the pre-trained branch is frozen and only the control branch is trained. See the supplementary document for more details on the architecture and training settings. + +Soft Inpainting. While guiding motion generation encourages outputs to align with the conditions, it is often not strong enough. We further seek to improve the consistency by masking the known regions and inpainting the unknown regions. Given a binary mask $\mathbf{M}$ that indicates the observed and unobserved regions, we can use the diffusion model to generate the inpainted motion $\tilde{\mathbf{H}}_0^t$ following the DDIM denoising process: + +$$ +\tilde {\mathbf {H}} _ {0} ^ {t, \text {k n o w n}} = \mathbf {H}, \tag {8a} +$$ + +$$ +\tilde {\mathbf {H}} _ {0} ^ {t, \text {u n k n o w n}} = \mathcal {D} _ {\phi , \phi_ {c}} \left(\tilde {\mathbf {H}} _ {t}, t, \mathbf {H} \odot \mathbf {M}\right), \tag {8b} +$$ + +$$ +\tilde {\mathbf {H}} _ {0} ^ {t} = \mathbf {M} \odot \tilde {\mathbf {H}} _ {0} ^ {t, \text {k n o w n}} + (1 - \mathbf {M}) \odot \tilde {\mathbf {H}} _ {0} ^ {t, \text {u n k n o w n}}. \tag {8c} +$$ + +Thus, the known regions are overwritten with the observations, while the unknown regions are sampled from the diffusion model. However, the above formulation keeps the observed parts unchanged during the denoising process. In practice, the observed parts can be noisy and not perfect. We still want the diffusion model to refine the observed parts but not change them significantly. + +Here, we present a soft inpainting strategy to infill the unobserved regions while refining the observed regions by dynamically reweighting the denoised direction from the diffusion model. Specifically, instead of using a binary mask, we adopt a continuous mask $\tilde{\mathbf{M}}$ depending on both the confidence score of the observations $\mathbf{S}$ and the denoising time step $t$ : + +$$ +\mathbf {M} = w (t) * \mathbf {S} \odot \mathbf {M}, \tag {9} +$$ + +where we set $w(t) = \max(0, \frac{t - 0.5}{0.5})$ to linearly decrease the weight of the observations as the denoising time step decreases. As the time step decreases, the denoising process will be more deterministic and the model will be more certain of the generated motions. + +Combining the three components of our solution, the final objective for COIN-SDS is formulated as: + +$$ +\min _ {\mathbf {H}} \mathcal {L} _ {\mathrm {C O I N - S D S}} := \mathbb {E} _ {t} \left[ \frac {\omega (t) \sqrt {\bar {\alpha} _ {t}}}{\sqrt {1 - \bar {\alpha} _ {t}}} \left\| \mathbf {H} - \tilde {\mathbf {H}} _ {0} (\mathbf {H}, \mathbf {M}, \mathbf {S}, t) \right\| _ {2} ^ {2} \right]. \tag {10} +$$ + +We summarize COIN in Alg. 1. + +# 3.3 Global Optimization + +Here we present the overall optimization pipeline for the joint estimation of global human and camera motion with the proposed COIN-SDS loss. Note that we use SLAM to initialize the camera poses, which is scale-ambiguous. Therefore, we need to jointly optimize the camera scale $s$ with the human and camera motions. Furthermore, the SLAM method assumes the camera in the first frame to be at the origin. To put the human motion in the correct positions, we follow PACE [41] and also optimize the camera height $h_0$ and the orientation $R_0$ for the first frame. The global human motions are initialized by the estimated local motions and the camera poses. The overall optimization objective is: + +$$ +\min _ {\mathbf {H}, \mathcal {C}, s, h _ {0}, R _ {0}, \beta} \mathcal {L} _ {\text {b o d y}} + \mathcal {L} _ {\text {C O I N - S D S}} + \mathcal {L} _ {\text {H S R}}, \tag {11} +$$ + +where + +$$ +\mathcal {L} _ {\text {b o d y}} = \mathcal {L} _ {2 \mathrm {D}} + \mathcal {L} _ {3 \mathrm {D}} + \mathcal {L} _ {\beta} + \mathcal {L} _ {\text {s m o o t h}} + \mathcal {L} _ {\text {c o n t a c t}}. \tag {12} +$$ + +$\mathcal{L}_{2\mathrm{D}}$ measures the 2D reprojection error between the projected 2D body joints of the estimated human motion and the detected 2D keypoints from an off-the-shelf 2D joint detector [9]. $\mathcal{L}_{3\mathrm{D}}$ measures the distance between the estimated local 3D joints and the detected 3D joints from an off-the-shelf 3D joint detector [79]. $\mathcal{L}_{\beta}$ is the shape regularization loss. $\mathcal{L}_{\mathrm{smooth}}$ is the temporal smoothness loss. $\mathcal{L}_{\mathrm{contact}}$ is the foot contact loss to encourage zero velocities for contact joints. The contact labels are obtained from the pseudo ground truth motion from COIN-SDS. Please refer to the supplementary document for more details. + +Human-scene Relation Loss. The camera trajectories recovered from SLAM are scale-ambiguous. Previous works [41,99] optimize the camera scale by projecting the global human motion to the camera space using the camera poses and minimizing the reprojection error. However, such a method entirely relies on the global human motions, which is in turn affected by the camera scale. If the human motion is not initialized well, the camera scale will also be inaccurate. To solve this problem, instead of the global human motions, we propose a new human-scene relation loss that uses the depth relation between the human and scene in the camera space, which disentangles the effect of the camera itself. + +Specifically, we use the point cloud of the scene recovered by SLAM as a constraint. First of all, the scale of the scene point cloud is the same as the camera scale, so optimizing the scene scale is equivalent to optimizing the camera scale. Second, the scene points that are projected onto the visible vertices of the body mesh should be occluded by the person. Otherwise, the corresponding body parts are invisible. Therefore, we can constrain the depth of the occluded scene points to be larger than the depth of the human body vertices. While finding the corresponding body vertices for each scene point is time-consuming, we propose to use the depth of its nearest body joint as a proxy. Given the scene point cloud $\mathcal{P}$ and the camera scale $s$ , the human-scene relation loss is formulated as: + +$$ +\mathcal {L} _ {\mathrm {H S R}} = - \frac {1}{| \mathcal {P} |} \sum_ {i = 1} ^ {T} \sum_ {p \in \mathcal {P} ^ {*}} \min (0, \mathcal {T} ^ {(i)} (p) _ {z} - j ^ {(i)} (p) _ {z}) \cdot \mathbb {1} (\mathcal {T} ^ {(i)} (p) \text {i s i n v i s i b l e}), \tag {13} +$$ + +where $\mathcal{P}^* = \mathcal{P}*s$ is the scaled point cloud of the scene, $\mathcal{T}^{(i)}(p) = \mathbf{R}^{(i)}p + \mathbf{t}^{(i)}$ is the transformed point in the $i$ -th frame, $j^{(i)}(p)$ is body joint that has the nearest 2D projection to the scene point $p$ in the $i$ -th frame, and $z$ denotes the depth of a given point. If the depth order is correct, i.e., $\mathcal{T}^{(i)}(p)_z - j^{(i)}(p)_z > 0$ , the loss is zero. The proposed human-scene relation loss uses the relation between the local motions and the scene to alleviate the scale ambiguity. The depth relation regularizes consistency among humans, cameras, and scenes. + +# 4 Experiments + +Datasets. We perform experiments on three human motion datasets. First is the real-world dataset RICH [25]. We follow previous works [41,81,100] to assess the performance of global human motion estimation using this dataset. The second + +Table 1: Global human motion estimation on the RICH dataset. + +
MethodPA-MPJPE ↓W-MPJPE ↓WA-MPJPE ↓W-RJE ↓ACCEL ↓
HybrIK [48] + SLAM [89]46.71073.1404.41166.220.2
GLAMR [100]79.9653.7365.1646.6107.7
SLAHMR [99]52.5571.6323.7400.59.4
WHAM [81]46.2497.6272.7478.26.7
PACE [41]49.3380.0197.2370.88.8
Guided Sampling132.81384.6502.21440.924.2
Noise Optimization66.7414.8195.3429.28.4
Vanilla SDS78.81453.5497.21458.012.7
COIN w/o Controlled Sampling44.0825.0291.8848.510.8
COIN w/o Dynamic Control49.5293.8180.6299.18.6
COIN w/o Soft Inpainting47.6325.8196.0324.99.6
COIN w/o LHSR43.6273.0176.3281.18.2
COIN42.9254.5169.5249.97.5
+ +one is EMDB [37]. We follow previous works [81] to evaluate on a subset of EMDB for which they provide ground truth global motion with dynamic cameras. The third dataset is HCM [41], a synthetic dataset. Compared to real-world datasets, HCM contains more challenging camera motions. We follow previous work [41] to evaluate the global human motion and the camera motion using this dataset. + +Metrics. We report various metrics for both human and camera motion. For human motion, standard metrics W-MPJPE and WA-MPJPE are used to evaluate global motion, while PA-MPJPE evaluates local motion. We also include an ACCEL metric to measure the joint acceleration error. For evaluation on EMDB, we follow previous work [81] to split sequences into smaller chunks of 100 frames and align each output segment with the ground-truth data using the first two frames W-MPJPE100 or the entire segment WA-MPJPE100. Root Orientation Error (ROE in deg) and Root Translation Error (RTE in $m$ ) evaluate the error over the entire trajectory after aligning with the initial camera pose. + +For camera motion, we report the average translation error after scale alignment (ATE), without scale alignment (ATE-S), and the camera acceleration error (CAM ACCEL). ATE-S more accurately reflects inaccuracies in the captured scale of the camera. + +Baselines. As discussed in Sec. 2, there are different ways to use the pre-trained motion diffusion model as a motion prior. Here, we summarize the three main solutions and compare them with COIN in Sec. 4.2. (1) Guided Sampling, which embeds analytical guidance within the denoising procedure using objective functions, such as 2D projection and foot contact consistency. This does not suit our task because the camera trajectories are also unknown and guided gradients from the wrong camera will lead to unrealistic human motions. We need to optimize the human motion and camera motion simultaneously. (2) Noise Optimization, which represents the motion as latent noise and directly optimizes it. This is similar to other motion priors such as VAE [21]. At each optimization step, we need to calculate the gradients of the latent w.r.t. the generated motion, which is computationally expensive and we find the performance is not good enough. (3) Vanilla SDS, which removes our design and directly uses SDS for optimization. + +Table 2: Global human motion estimation on the EMDB dataset. + +
MethodPA-MPJPE ↓W-MPJPE100 ↓WA-MPJPE100 ↓RTE ↓ROE ↓
HMR2.0 [14] + DPVO [90]49.62320.9662.917.544.4
GLAMR [100]56.0756.1286.216.774.9
TRACE [86]58.02244.9544.118.972.7
SLAHMR [99]61.5807.4336.913.867.9
WHAM [81]41.9439.2166.18.436.3
COIN32.7407.3152.83.534.1
+ +Table 3: Global human motion estimation on the HCM dataset. + +
MethodPA-MPJE ↓W-MPJE ↓WA-MPJE ↓W-RJE ↓ACCEL ↓
HybrIK [48] + SLAM [89]67.61137.3780.31100.951.3
GLAMR [100]86.01977.6653.81958.033.4
SLAHMR [99]69.9888.9483.5862.214.9
PACE [41]65.3861.2478.3839.516.7
WHAM [81]47.9588.9279.3579.213.1
COIN45.5479.9212.1470.710.1
+ +# 4.1 Comparison with State-of-the-Art Methods + +Human Motion Estimation. We compare COIN against state-of-the-art methods on the RICH, EMDB, and HCM datasets. Quantitative results are shown in Tabs. 1, 2, and 3. We observe that COIN significantly outperforms the state-of-the-art methods on all datasets. On the RICH dataset, COIN outperforms the state-of-the-art method, PACE [41], by $125.5\mathrm{mm}$ in terms of W-MPJPE, showing $33.0\%$ relative improvement. On the EMDB and HCM datasets, COIN shows $29.1\mathrm{mm}$ and $109.0\mathrm{mm}$ improvement in terms of W-MPJPE, respectively. Qualitative comparisons with state-of-the-art methods, PACE [41] and WHAM [81], are shown in Figs. 1 and 3. More qualitative results are shown in the supplementary video. + +COIN not only improves global motion, but also improves local body motion. In terms of PA-MPJPE, COIN shows $3.3\mathrm{mm}$ , $9.2\mathrm{mm}$ , and $19.8\mathrm{mm}$ improvement on the RICH, EMDB, and HCM datasets, respectively. This demonstrates that COIN is able to distill high-quality motion priors from the diffusion model and help both local and global motion estimation. Regarding the joint acceleration error, COIN is $0.8\mathrm{mm/s}^2$ higher than WHAM [81]. Note that the joint acceleration error reflects the smoothness. Human motion can be over-smoothed but not accurate, hence it is important to look at acceleration error in conjunction with other metrics. + +Camera Motion Estimation. We further evaluate the performance of COIN on camera motion estimation on the HCM dataset. Quantitative results are shown in Tab. 4. COIN substantially surpasses the state-of-the-art camera motion estimation methods. Specifically, COIN reduces the absolute camera translation error, ATE-S by $73.8\mathrm{mm}$ . This demonstrates that COIN is able to disentangle human and camera motions and accurately estimate the camera motion. + +Table 4: Camera motion estimation on the HCM dataset. + +
MethodATE ↓ATE-S ↓CAMACCEL ↓
HybrIK [48] + SLAM [89]155.81670.717.1
GLAMR [100]1295.21714.6282.9
SLAHMR [99]155.8506.517.6
PACE [41]137.5459.716.2
Guided Sampling335.6992.415.6
Noise Optimization206.0500.912.3
Vanilla SDS306.4656.413.7
COIN w/o Controlled Sampling299.6553.814.0
COIN w/o Dynamic Control149.8397.711.3
COIN w/o Soft Inpainting167.7423.811.4
COIN w/o LHSR147.8402.112.0
COIN135.3385.911.3
+ +# 4.2 Ablation Study + +In this section, we conduct ablation studies on the RICH and HCM datasets to evaluate the impact of each component on human and camera motions, respectively. More comparisons on other datasets are provided in the supplementary document. + +Baselines with Motion Diffusion Model. We first compare the aforementioned baselines with COIN. To use guided sampling in our tasks, we jointly update the camera poses using the gradients from the objective function during denoising. Quantitative results of human and camera motion estimation are shown in Tabs. 1 and 4, respectively. We observe that COIN outperforms all the baselines in terms of all metrics. As expected, guided sampling shows a terrible performance because it is not able to accurately estimate the camera trajectories. Noise optimization is better than vanilla SDS but worse than COIN, which generates consistent motion priors. Vanilla SDS is replacing $\mathcal{L}_{\mathrm{COIN}}$ with the SDS loss and keeping the rest of the settings the same. These results demonstrate the effectiveness of COIN over other motion diffusion baselines. + +Impact of Dynamic Controlled Sampling. To study the effectiveness of controlled sampling, we compare COIN with and without using the controlled denoiser. Quantitative results are shown in Tabs. 1 and 4. We observe that controlled sampling significantly improves the performance of COIN. Specifically, COIN with controlled sampling reduces the W-MPJPE and ATE-S by $570.5\mathrm{mm}$ and $167.9\mathrm{mm}$ , showing $69.3\%$ and $30.3\%$ relative improvement, respectively. This demonstrates that controlled sampling is able to generate high-quality motion priors that align with the observed evidence and help improve human and camera motion estimation. + +Impact of Soft Inpainting. We further study the effectiveness of soft inpainting. Quantitative comparisons are shown in Tabs. 1 and 4. We observe that while soft inpainting also improves W-MPJPE, it is much more effective than + +![](images/3d507f94e555b157a84a398fb6034eb37a4e580bf6d730f47ec2d1f84b3455d2.jpg) +PACE (Human Motion) PACE (Camera Motion) + +![](images/b2d1ac4a7282d5b7072db75d882ccf786ccc55ba582646a308f5bcb8bdb1e758.jpg) +WHAM +COIN (Human Motion) COIN (Camera Motion) +Fig. 3: Qualitative comparisons with state-of-the-art methods. PACE [41] fails to recover a correct trajectory (left). WHAM [81] estimates the wrong walking direction of the person (right). Our approach, COIN, recovers the human and camera motion accurately in both scenarios. + +controlled sampling in terms of local body motions. Specifically, COIN with soft inpainting reduces the PA-MPJPE by $4.7 \mathrm{~mm}$ . + +# 5 Conclusion + +In this paper, we propose COIN, a diffusion-based optimization framework for global human and camera motion estimation from dynamic cameras. We identify the inconsistency problem of distilling knowledge from the diffusion model with conventional SDS loss. To address this issue, COIN uses a controlled denoiser combined with soft inpainting to distill a high-quality, well-aligned, and consistent motion prior. To further address the scale ambiguity of the camera trajectory, we develop a novel human-scene relation loss that imposes consistency among the human motion, camera motion, and scene features. Comprehensive experiments on challenging synthetic and real-world datasets demonstrate the effectiveness of COIN, which outperforms the SOTA by a large margin in recovering accurate global human motion and camera motion. + +Limitations and Future Works: While COIN is able to jointly optimize camera trajectories and global human motions, it requires initialization from SLAM. If the SLAM method fails catastrophically, COIN may not be effective. Additionally, we found COIN fails under severe occlusions where there are several unseen frames and the diffusion model cannot provide consistent guidance. Another limitation is that COIN is an optimization framework, so it is unsuitable for real-time applications. Since the denoising diffusion models have shown their power to model data distributions in many different domains, looking forward we can learn the joint distribution of humans and cameras with the diffusion model to additionally denoise the artifacts in camera motions. Such a joint distribution model also holds the potential of real-time human and camera motion estimation via guided sampling with few DDIM denoising steps. + +# References + +1. Akhter, I., Black, M.J.: Pose-conditioned joint angle limits for 3D human pose reconstruction. In: CVPR (2015) 3 +2. Aksan, E., Kaufmann, M., Hilliges, O.: Structured prediction helps 3d human motion modelling. In: ICCV (2019) 4 +3. Barsoum, E., Kender, J., Liu, Z.: Hp-gan: Probabilistic 3d human motion prediction via gan. In: CVPR Workshops (2018) 4 +4. Bogo, F., Kanazawa, A., Lassner, C., Gehler, P., Romero, J., Black, M.J.: Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In: ECCV (2016) 3 +5. Cao, Z., Gao, H., Mangalam, K., Cai, Q.Z., Vo, M., Malik, J.: Long-term human motion prediction with scene context. In: ECCV. pp. 387-404. Springer (2020) 4 +6. Choi, H., Moon, G., Lee, K.M.: Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In: ECCV (2020) 3 +7. Choi, H., Moon, G., Lee, K.M.: Beyond static features for temporally consistent 3d human pose and shape from a video. In: CVPR (2021) 3 +8. Choutas, V., Pavlakos, G., Bolkart, T., Tzionas, D., Black, M.J.: Monocular expressive body regression through body-driven attention. In: ECCV (2020) 3 +9. Contributors, M.: Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose (2020) 10 +10. Dabral, R., Shimada, S., Jain, A., Theobalt, C., Golyanik, V.: Gravity-aware monocular 3d human-object reconstruction. In: ICCV (2021) 3 +11. Fabbri, M., Lanzi, F., Calderara, S., Alletto, S., Cucchiara, R.: Compressed volumetric heatmaps for multi-person 3d pose estimation. In: CVPR (June 2020) 3 +12. Fragkiadaki, K., Levine, S., Felsen, P., Malik, J.: Recurrent network models for human dynamics. In: ICCV (2015) 4 +13. Gartner, E., Andriluka, M., Xu, H., Sminchisescu, C.: Trajectory optimization for physics-based reconstruction of 3d human pose from monocular video. In: CVPR (2022) 3 +14. Goel, S., Pavlakos, G., Rajasegaran, J., Kanazawa, A., Malik, J.: Humans in 4d: Reconstructing and tracking humans with transformers. In: ICCV (2023) 12 +15. Gopalakrishnan, A., Mali, A., Kifer, D., Giles, L., Ororbia, A.G.: A neural temporal model for human motion prediction. In: CVPR (2019) 4 +16. Guler, R.A., Kokkinos, I.: HoloPose: Holistic 3d human reconstruction in-the-wild. In: CVPR (2019) 3 +17. Guzov, V., Mir, A., Sattler, T., Pons-Moll, G.: Human POSEitioning system (HPS): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In: CVPR (2021) 4 +18. Harvey, F.G., Yurick, M., Nowrouzezahrai, D., Pal, C.: Robust motion in betweening. ACM Transactions on Graphics (TOG) 39(4), 60-1 (2020) 4 +19. Hassan, M., Ceylan, D., Villegas, R., Saito, J., Yang, J., Zhou, Y., Black, M.J.: Stochastic scene-aware motion prediction. In: ICCV (2021) 4 +20. Hassan, M., Choutas, V., Tzionas, D., Black, M.J.: Resolving 3d human pose ambiguities with 3d scene constraints. In: ICCV. pp. 2282-2292 (2019) 4 +21. He, C., Saito, J., Zachary, J., Rushmeier, H., Zhou, Y.: Nemf: Neural motion fields for kinematic animation. In: NeurIPS (2022) 2, 4, 11 +22. Henning, D.F., Laidlow, T., Leutenegger, S.: Bodyslam: Joint camera localisation, mapping, and human motion tracking. In: ECCV (2022) 4 + +23. Hernandez, A., Gall, J., Moreno-Noguer, F.: Human motion prediction via spatiotemporal inpainting. In: CVPR (2019) 4 +24. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPs (2020) 2 +25. Huang, C.H.P., Yi, H., Höschle, M., Safroshkin, M., Alexiadis, T., Polikovsky, S., Scharstein, D., Black, M.J.: Capturing and inferring dense full-body human-scene contact. In: CVPR (2022) 3, 10 +26. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: CVPR (2023) 4 +27. Iqbal, U., Molchanov, P., Kautz, J.: Weakly-supervised 3d human pose learning via multi-view images in the wild. In: CVPR (2020) 3 +28. Iqbal, U., Xie, K., Guo, Y., Kautz, J., Molchanov, P.: KAMA: 3D keypoint aware body mesh articulation. In: 3DV (2021) 3 +29. Isogawa, M., Yuan, Y., O'Toole, M., Kitani, K.M.: Optical non-line-of-sight physics-based 3d human pose estimation. In: CVPR (2020) 3 +30. Jain, A., Zamir, A.R., Savarese, S., Saxena, A.: Structural-rnn: Deep learning on spatio-temporal graphs. In: CVPR (2016) 4 +31. Jiang, W., Kolotouros, N., Pavlakos, G., Zhou, X., Daniilidis, K.: Coherent reconstruction of multiple humans from a single image. In: CVPR (2020) 3 +32. Joo, H., Neverova, N., Vedaldi, A.: Exemplar fine-tuning for 3d human pose fitting towards in-the-wild 3d human pose estimation. In: 3DV (2021) 3 +33. Kanazawa, A., Black, M.J., Jacobs, D.W., Malik, J.: End-to-end recovery of human shape and pose. In: CVPR (2018) 3 +34. Kanazawa, A., Zhang, J.Y., Felsen, P., Malik, J.: Learning 3d human dynamics from video. In: CVPR (2019) 3 +35. Karunratanakul, K., Preechakul, K., Aksan, E., Beeler, T., Suwajanakorn, S., Tang, S.: Optimizing diffusion noise can serve as universal motion priors. arXiv preprint arXiv:2312.11994 (2023) 4 +36. Kaufmann, M., Aksan, E., Song, J., Pece, F., Ziegler, R., Hilliges, O.: Convolutional autoencoders for human motion infilling. In: 3DV (2020) 4 +37. Kaufmann, M., Song, J., Guo, C., Shen, K., Jiang, T., Tang, C., Zárate, J.J., Hilliges, O.: EMDB: The Electromagnetic Database of Global 3D Human Pose and Shape in the Wild. In: ICCV (2023) 3, 11 +38. Khurana, T., Dave, A., Ramanan, D.: Detecting invisible people. In: ICCV. pp. 3174-3184 (2021) 4 +39. Kocabas, M., Athanasiou, N., Black, M.J.: VIBE: Video inference for human body pose and shape estimation. In: CVPR (2020) 3, 4 +40. Kocabas, M., Huang, C.H.P., Tesch, J., Müller, L., Hilliges, O., Black, M.J.: SPEC: Seeing people in the wild with an estimated camera. In: ICCV (2021) 3 +41. Kocabas, M., Yuan, Y., Molchanov, P., Guo, Y., Black, M.J., Hilliges, O., Kautz, J., Iqbal, U.: PACE: Human and motion estimation from in-the-wild videos. In: 3DV (2024) 1, 2, 3, 4, 9, 10, 11, 12, 13, 14 +42. Kolotouros, N., Pavlakos, G., Black, M.J., Daniilidis, K.: Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In: ICCV (2019) 3 +43. Kolotouros, N., Pavlakos, G., Daniilidis, K.: Convolutional mesh regression for single-image human shape reconstruction. In: CVPR (2019) 3 +44. Kolotouros, N., Pavlakos, G., Jayaraman, D., Daniilidis, K.: Probabilistic modeling for human mesh recovery. In: ICCV (2021) 3 + +45. Kundu, J.N., Rakesh, M., Jampani, V., Venkatesh, R.M., Babu1, R.V.: Appearance consensus driven self-supervised human mesh recovery. In: ECCV (2020) 3 +46. Lassner, C., Romero, J., Kiefel, M., Bogo, F., Black, M.J., Gehler, P.V.: Unite the people: Closing the loop between 3D and 2D human representations. In: CVPR (2017) 3 +47. Li, J., Bian, S., Xu, C., Liu, G., Yu, G., Lu, C.: D &d: Learning human dynamics from dynamic camera. In: ECCV (2022) 2, 4 +48. Li, J., Xu, C., Chen, Z., Bian, S., Yang, L., Lu, C.: Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In: CVPR (2021) 3, 5, 8, 11, 12, 13 +49. Li, Z., Liu, J., Zhang, Z., Xu, S., Yan, Y.: Cliff: Carrying location information in full frames into human pose and shape estimation. In: ECCV (2022) 3 +50. Li, Z., Zhou, Y., Xiao, S., He, C., Huang, Z., Li, H.: Auto-conditioned recurrent networks for extended complex human motion synthesis. arXiv preprint arXiv:1707.05363 (2017) 4 +51. Lin, K., Wang, L., Liu, Z.: End-to-end human pose and mesh reconstruction with transformers. In: CVPR (2021) 3 +52. Liu, M., Yang, D., Zhang, Y., Cui, Z., Rehg, J.M., Tang, S.: 4d human body capture from egocentric video via 3d scene grounding. In: 3DV (2021) 3, 4 +53. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. SIGGRAPH Asia 34(6), 248:1-248:16 (2015) 6 +54. Luo, Z., Golestaneh, S.A., Kitani, K.M.: 3d human motion estimation via motion compression and refinement. In: ACCV (2020) 3 +55. Luo, Z., Hachiuma, R., Yuan, Y., Kitani, K.: Dynamics-regulated kinematic policy for egocentric pose estimation. NeurIPS 34 (2021) 4 +56. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: AMASS: Archive of motion capture as surface shapes. In: ICCV (2019) 8 +57. von Marcard, T., Henschel, R., Black, M., Rosenhahn, B., Pons-Moll, G.: Recovering accurate 3d human pose in the wild using imus and a moving camera. In: ECCV (2018) 4 +58. Martinez, J., Black, M.J., Romero, J.: On human motion prediction using recurrent neural networks. In: CVPR (2017) 4 +59. Mehta, D., Rhodin, H., Casas, D., Fua, P., Sotnychenko, O., Xu, W., Theobalt, C.: Monocular 3d human pose estimation in the wild using improved cnn supervision. In: 3DV (2017) 3 +60. Mehta, D., Sotnychenko, O., Mueller, F., Xu, W., Elgharib, M., Fua, P., Seidel, H.P., Rhodin, H., Pons-Moll, G., Theobalt, C.: XNect: Real-time multi-person 3D motion capture with a single RGB camera. In: SIGGRAPH (2020) 3 +61. Mehta, D., Sridhar, S., Sotnychenko, O., Rhodin, H., Shafiei, M., Seidel, H., Xu, W., Casas, D., Theobalt, C.: VNect: Real-time 3D human pose estimation with a single RGB camera. In: SIGGRAPH (2017) 3 +62. Moon, G., Chang, J.Y., Lee, K.M.: Camera distance-aware top-down approach for 3d multi-person pose estimation from a single rgb image. In: ICCV (2019) 3 +63. Moon, G., Lee, K.M.: I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In: ECCV (2020) 3 +64. Müller, L., Osman, A.A.A., Tang, S., Huang, C.H.P., Black, M.J.: On self contact and human pose. In: CVPR (2021) 3 + +65. Müller, L., Ye, V., Pavlakos, G., Black, M., Kanazawa, A.: Generative proxemics: A prior for 3d social interaction from images. arXiv preprint arXiv:2306.09337 (2023) 4 +66. Opensfm - a structure from motion library. https://github.com/mapillary/OpenSfM (2021), https://github.com/mapillary/OpenSfM 2 +67. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: CVPR (2019) 3 +68. Pavlakos, G., Kolotouros, N., Daniilidis, K.: Texturepose: Supervising human mesh estimation with texture consistency. In: ICCV (2019) 3 +69. Pavlakos, G., Weber, E., Tancik, M., Kanazawa, A.: The one where they reconstructed 3d humans and environments in tv shows. In: ECCV (2022) 4 +70. Pavlakos, G., Zhu, L., Zhou, X., Daniilidis, K.: Learning to estimate 3D human pose and shape from a single color image. In: CVPR (2018) 3 +71. Pavlo, D., Grangier, D., Auli, M.: Quaternet: A quaternion-based recurrent model for human motion. In: BMVC (2018) 4 +72. Payer, C., Neff, T., Bischof, H., Urschler, M., Stern, D.: Simultaneous multi-person detection and single-person pose estimation with a single heatmap regression network. In: ICCV PoseTrack Workshop (2017) 3 +73. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: ICCV (2021) 4 +74. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: ICLR (2023) 2, 6 +75. Reddy, N.D., Guigues, L., Pischulini, L., Eledath, J., Narasimhan, S.: Tessetrack: End-to-end learnable multi-person articulated 3d pose tracking. In: CVPR (2021) 3 +76. Rempe, D., Birdal, T., Hertzmann, A., Yang, J., Sridhar, S., Guibas, L.J.: Humor: 3d human motion model for robust pose estimation. In: ICCV (2021) 2, 3, 4 +77. Rogez, G., Weinzaepfel, P., Schmid, C.: LCR-Net: Localization-classification-regression for human pose. In: CVPR (2017) 3 +78. Rong, Y., Liu, Z., Li, C., Cao, K., Change Loy, C.: Delving deep into hybrid annotations for 3d human recovery in the wild. In: ICCV (2019) 3 +79. Sárándi, I., Hermans, A., Leibe, B.: Learning 3D human pose estimation from dozens of datasets using a geometry-aware autoencoder to bridge between skeleton formats. In: WACV (2023) 10 +80. Shimada, S., Golyanik, V., Xu, W., Theobalt, C.: Physcap: Physically plausible monocular 3d motion capture in real time. In: SIGGRAPH (2020) 3 +81. Shin, S., Kim, J., Halilaj, E., Black, M.J.: Wham: Reconstructing world-grounded humans with accurate 3d motion. In: CVPR (2024) 1, 2, 3, 4, 10, 11, 12, 14 +82. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015) 4 +83. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2021) 7 +84. Song, J., Chen, X., Hilliges, O.: Human body model fitting by learned gradient descent. In: ECCV (2020) 3 +85. Sun, Y., Bao, Q., Liu, W., Fu, Y., Black, M.J., Mei, T.: Monocular, one-stage, regression of multiple 3d people. In: ICCV (2021) 3 +86. Sun, Y., Bao, Q., Liu, W., Mei, T., Black, M.J.: Trace: 5d temporal regression of avatars with dynamic cameras in 3d environments. In: CVPR (2023) 4, 12 +87. Sun, Y., Liu, W., Bao, Q., Fu, Y., Mei, T., Black, M.J.: Putting people in their place: Monocular regression of 3D people in depth. In: CVPR (2022) 3 + +88. Sun, Y., Ye, Y., Liu, W., Gao, W., Fu, Y., Mei, T.: Human mesh recovery from monocular images via a skeleton-disentangled representation. In: ICCV (2019) 3 +89. Teed, Z., Deng, J.: DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. NeurIPs (2021) 2, 5, 11, 12, 13 +90. Teed, Z., Lipson, L., Deng, J.: Deep patch visual odometry. NeurIPS (2023) 2, 12 +91. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. In: ICLR 2023 (2022) 2, 4, 8 +92. Villegas, R., Yang, J., Zou, Y., Sohn, S., Lin, X., Lee, H.: Learning to generate long-term future via hierarchical prediction. In: ICML (2017) 4 +93. Weng, Z., Yeung, S.: Holistic 3d human and scene mesh estimation from single view images. In: CVPR (2021) 3 +94. Xiang, D., Joo, H., Sheikh, Y.: Monocular total capture: Posing face, body and hands in the wild. In: CVPR (2019) 3 +95. Xie, K., Wang, T., Iqbal, U., Guo, Y., Fidler, S., Shkurti, F.: Physics-based human motion estimation and synthesis from videos. In: ICCV (2021) 3 +96. Xie, Y., Jampani, V., Zhong, L., Sun, D., Jiang, H.: Omnicontrol: Control any joint at any time for human motion generation. In: ICLR (2024) 4 +97. Xu, Y., Zhu, S.C., Tung, T.: Denserac: Joint 3d pose and shape estimation by dense render-and-compare. In: ICCV (2019) 3 +98. Yan, X., Rastogi, A., Villegas, R., Sunkavalli, K., Shechtman, E., Hadap, S., Yumer, E., Lee, H.: Mt-vae: Learning motion transformations to generate multimodal human dynamics. In: ECCV (2018) 4 +99. Ye, V., Pavlakos, G., Malik, J., Kanazawa, A.: Decoupling human and camera motion from videos in the wild. In: CVPR (2023) 2, 4, 10, 11, 12, 13 +00. Yuan, Y., Iqbal, U., Molchanov, P., Kitani, K., Kautz, J.: Glamr: Global occlusion-aware human mesh recovery with dynamic cameras. In: CVPR (2022) 2, 4, 10, 11, 12, 13 +01. Yuan, Y., Kitani, K.: Diverse trajectory forecasting with determinantal point processes. In: ICLR 2020 (2019) 4 +02. Yuan, Y., Kitani, K.: Dlow: Diversifying latent flows for diverse human motion prediction. In: ECCV (2020) 4 +03. Yuan, Y., Kitani, K.: Residual force control for agile human behavior imitation and extended motion synthesis. In: NeurIPS (2020) 4 +04. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: ICCV (2023) 4 +05. Yuan, Y., Wei, S.E., Simon, T., Kitani, K., Saragih, J.: Simpoe: Simulated character control for 3d human pose estimation. In: CVPR (2021) 3 +06. Zanfir, A., Bazavan, E.G., Xu, H., Freeman, W.T., Sukthankar, R., Sminchisescu, C.: Weakly supervised 3d human pose and shape reconstruction with normalizing flows. In: ECCV (2020) 3 +07. Zanfir, A., Marinoiu, E., Sminchisescu, C.: Monocular 3d pose and shape estimation of multiple people in natural scenes the importance of multiple scene constraints. In: CVPR (2018) 3 +08. Zanfir, A., Marinoiu, E., Zanfir, M., Popa, A.I., Sminchisescu, C.: Deep network for the integrated 3d sensing of multiple people in natural images. In: NeurIPS (2018) 3 +09. Zanfir, M., Zanfir, A., Bazavan, E.G., Freeman, W.T., Sukthankar, R., Sminchisscu, C.: Thundr: Transformer-based 3d human reconstruction with markers. In: ICCV (2021) 3 + +110. Zhang, H., Tian, Y., Zhou, X., Ouyang, W., Liu, Y., Wang, L., Sun, Z.: Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In: ICCV (2021) 3 +111. Zhang, J., Yu, D., Liew, J.H., Nie, X., Feng, J.: Body meshes as points. In: CVPR (2021) 3 +112. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models (2023) 8 +113. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022) 4 +114. Zhang, S., Bhatnagar, B.L., Xu, Y., Winkler, A., Kadlecek, P., Tang, S., Bogo, F.: Rohm: Robust human motion reconstruction via diffusion. In: CVPR (2024) 4 +115. Zhang, S., Zhang, Y., Bogo, F., Pollefeys, M., Tang, S.: Learning motion priors for 4d human body capture in 3d scenes. In: ICCV (2021) 4 +116. Zhang, T., Huang, B., Wang, Y.: Object-occluded human shape and pose estimation from a single color image. In: CVPR (2020) 3 +117. Zhen, J., Fang, Q., Sun, J., Liu, W., Jiang, W., Bao, H., Zhou, X.: SMAP: Single-shot multi-person absolute 3d pose estimation. In: ECCV (2020) 3 +118. Zhou, Y., Habermann, M., Habibie, I., Tewari, A., Theobalt, C., Xu, F.: Monocular real-time full body capture with inter-part correlations. In: CVPR (2021) 3 \ No newline at end of file diff --git a/2024/COIN_ Control-Inpainting Diffusion Prior for Human and Camera Motion Estimation/images.zip b/2024/COIN_ Control-Inpainting Diffusion Prior for Human and Camera Motion Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e61c7e9197df8736086ff6e1d25d1c0efa586484 --- /dev/null +++ b/2024/COIN_ Control-Inpainting Diffusion Prior for Human and Camera Motion Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e553720128342a72a7b718e67a5bfe966450699703129aaeeef7341da43abafb +size 476542 diff --git a/2024/COIN_ Control-Inpainting Diffusion Prior for Human and Camera Motion Estimation/layout.json b/2024/COIN_ Control-Inpainting Diffusion Prior for Human and Camera Motion Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0027bab8cb8553830de1d25f378dc5075f127830 --- /dev/null +++ b/2024/COIN_ Control-Inpainting Diffusion Prior for Human and Camera Motion Estimation/layout.json @@ -0,0 +1,11882 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 111, + 465, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 111, + 465, + 146 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 465, + 146 + ], + "type": "text", + "content": "COIN: Control-Inpainting Diffusion Prior for Human and Camera Motion Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 153, + 475, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 153, + 475, + 178 + ], + "spans": [ + { + "bbox": [ + 138, + 153, + 475, + 178 + ], + "type": "text", + "content": "Jiefeng Li" + }, + { + "bbox": [ + 138, + 153, + 475, + 178 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 138, + 153, + 475, + 178 + ], + "type": "text", + "content": ", Ye Yuan" + }, + { + "bbox": [ + 138, + 153, + 475, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 153, + 475, + 178 + ], + "type": "text", + "content": ", Davis Rempe" + }, + { + "bbox": [ + 138, + 153, + 475, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 153, + 475, + 178 + ], + "type": "text", + "content": ", Haotian Zhang" + }, + { + "bbox": [ + 138, + 153, + 475, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 153, + 475, + 178 + ], + "type": "text", + "content": ", Pavlo Molchanov" + }, + { + "bbox": [ + 138, + 153, + 475, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 153, + 475, + 178 + ], + "type": "text", + "content": ", Cewu Lu" + }, + { + "bbox": [ + 138, + 153, + 475, + 178 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 153, + 475, + 178 + ], + "type": "text", + "content": ", Jan Kautz" + }, + { + "bbox": [ + 138, + 153, + 475, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 153, + 475, + 178 + ], + "type": "text", + "content": ", and Umar Iqbal" + }, + { + "bbox": [ + 138, + 153, + 475, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 215, + 186, + 398, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 186, + 398, + 212 + ], + "spans": [ + { + "bbox": [ + 215, + 186, + 398, + 212 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 215, + 186, + 398, + 212 + ], + "type": "text", + "content": "NVIDIA " + }, + { + "bbox": [ + 215, + 186, + 398, + 212 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 215, + 186, + 398, + 212 + ], + "type": "text", + "content": "Shanghai Jiao Tong University https://nvlabs.github.io/COIN/" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 217, + 482, + 361 + ], + "blocks": [ + { + "bbox": [ + 133, + 217, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 133, + 217, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 133, + 217, + 482, + 361 + ], + "type": "image", + "image_path": "9fb51b864a9ead7976a41d75c7841abdf6d41033d1beac89fc244f7e0bf84e22.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 365, + 483, + 441 + ], + "lines": [ + { + "bbox": [ + 130, + 365, + 483, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 483, + 441 + ], + "type": "text", + "content": "Fig. 1: Capturing global human and camera motion from a dynamic camera presents unique challenges. In the input video, a person is riding a skateboard - while the local body motion may remain relatively constant, the global position of the individual changes significantly. Current state-of-the-art methods such as PACE [41] and WHAM [81] fail catastrophically on such out-of-distribution motions. Our approach, COIN, gracefully handles such challenging cases, owing to our control-inpainting motion diffusion prior and novel human-scene relation loss." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 457, + 453, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 457, + 453, + 666 + ], + "spans": [ + { + "bbox": [ + 159, + 457, + 453, + 666 + ], + "type": "text", + "content": "Abstract. Estimating global human motion from moving cameras is challenging due to the entanglement of human and camera motions. To mitigate the ambiguity, existing methods leverage learned human motion priors, which however often result in oversmoothed motions with misaligned 2D projections. To tackle this problem, we propose COIN, a control-inpainting motion diffusion prior that enables fine-grained control to disentangle human and camera motions. Although pre-trained motion diffusion models encode rich motion priors, we find it non-trivial to leverage such knowledge to guide global motion estimation from RGB videos. COIN introduces a novel control-inpainting score distillation sampling method to ensure well-aligned, consistent, and high-quality motion from the diffusion prior within a joint optimization framework. Furthermore, we introduce a new human-scene relation loss to alleviate the scale ambiguity by enforcing consistency among the humans, camera, and scene. Experiments on three challenging benchmarks demonstrate the effectiveness of COIN, which outperforms the state-of-the-art methods in terms of global human motion estimation and camera motion estimation. As an illustrative example, COIN outperforms the state-of-the-art method by " + }, + { + "bbox": [ + 159, + 457, + 453, + 666 + ], + "type": "inline_equation", + "content": "33\\%" + }, + { + "bbox": [ + 159, + 457, + 453, + 666 + ], + "type": "text", + "content": " in world joint position error (W-MPJPE) on the RICH dataset." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 160, + 230, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 230, + 173 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 230, + 173 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 186, + 479, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 186, + 479, + 233 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 479, + 233 + ], + "type": "text", + "content": "Recovering global human and camera motion from dynamic RGB videos is an important problem with many applications, such as animation, human-computer interaction, mixed reality, and robotics. However, it is a very challenging problem due to the entanglement of human and camera motion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 234, + 481, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 234, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 481, + 426 + ], + "type": "text", + "content": "There are only a few works [41, 81, 99, 100] that try to address this problem. Earlier methods [47, 100] only focus on human motion and ignore the camera motion. Their core insight is that the global body motion is highly correlated with the local motion. Thus, they can use the local body movements to estimate the global orientation and trajectory with a regression model [100] or by combining them with physics constraints [47]. However, these regression models ignore the camera movements, so they fail to maintain consistency with the input video, whereas physics-based methods fail to model complex in-the-wild environments so are limited to controlled scenarios. Recent works [41, 99] try to jointly estimate the human and camera motion by exploiting learned motion priors [21, 76] and SLAM [66, 89, 90]. They try to constrain the human body motion in a low-dimensional latent space of a motion prior model, which results in reconstructed motions that are overly smooth and do not align well with video observations. Moreover, the optimization of the camera motion is only based on the global human motion from the motion prior. Hence, they fail catastrophically if the initial human motion predictions are significantly incorrect (as shown in Fig. 1)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 481, + 533 + ], + "type": "text", + "content": "More recently, Denoising Diffusion Models [24,91] have emerged as a powerful family of generative models that can model high-quality data priors. Nonetheless, effectively leveraging the learned priors remains an ongoing challenge. Score Distillation Sampling (SDS) is commonly employed for this purpose [74], but we find that naive application of SDS also results in inconsistencies with the available observations (see Sec. 4.2). The root cause of this problem lies in the inconsistency of randomly sampled motions during SDS optimization. Without constraints, these motions may not align with observed evidence, leading to overly smoothed results that lack detail due to the mode-averaging effect." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": "In this work, we propose COIN, a hybrid COntrol-INpainting score distillation sampling method to address the aforementioned limitations of vanilla SDS. First, we use the partially observed human motion from the video as control signals to guide motion sampling. To tackle noisy observations which may be out of distribution for the prior, we propose a dynamic controlled sampling technique that iteratively refines the observed motions and updates the control signals to ensure effective distillation from the motion prior. Second, to further improve the consistency of the sampled motions, we also develop a novel soft inpainting strategy. We automatically identify the high-confidence regions of the initial predicted global motion from the video and use them as soft constraints during optimization. Concretely, we sample less confident regions from scratch using the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "L. Jiefeng et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 160, + 117, + 452, + 138 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 117, + 452, + 138 + ], + "spans": [ + { + "bbox": [ + 160, + 117, + 452, + 138 + ], + "type": "text", + "content": "Keywords: Global Human Motion Estimation " + }, + { + "bbox": [ + 160, + 117, + 452, + 138 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 117, + 452, + 138 + ], + "type": "text", + "content": " Human Motion Prior " + }, + { + "bbox": [ + 160, + 117, + 452, + 138 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 117, + 452, + 138 + ], + "type": "text", + "content": " Score Distillation Sampling" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "motion model, while the confident regions are only slightly refined. This ensures that the reconstructed motions do not deviate from the available observations. Our new SDS formulation is used to jointly optimize the human and camera motion by finding the most plausible solution that explains the observed evidence. Finally, to prevent catastrophic failure in cases where the initial body motion fails significantly, we propose a human-scene relation loss to consider the human-scene depth relations. This novel loss provides complementary information to the human motion prior by using local motion and scene features. It regularizes the camera scale by enforcing consistency among the human motion, camera motion, and scene features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 297 + ], + "type": "text", + "content": "We benchmark our approach on the synthetic HCM [41] dataset and the real-world RICH [25] and EMDB [37] datasets. We demonstrate that our approach significantly outperforms the state-of-the-art methods in terms of human motion estimation and camera motion estimation. Overall, the contributions of this paper can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 305, + 480, + 472 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 305, + 479, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 305, + 479, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 305, + 479, + 352 + ], + "type": "text", + "content": "- We propose a novel control-inpainting motion prior specifically designed for global human motion estimation, which enhances score distillation sampling with dynamic control and soft inpainting to reconstruct well-aligned, consistent, and high-quality motions from video observations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 354, + 479, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 479, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 479, + 388 + ], + "type": "text", + "content": "- We propose a new human-scene relation loss to resolve the scale ambiguity of the camera motion by enforcing consistency among the human motion, camera motion, and scene features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 390, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 390, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 138, + 390, + 480, + 472 + ], + "type": "text", + "content": "- Our approach significantly outperforms the state-of-the-art methods in terms of human motion estimation and camera motion estimation on both synthetic and real-world datasets. In terms of global human motion estimation in world space, we outperform the state-of-the-art method PACE [41] by " + }, + { + "bbox": [ + 138, + 390, + 480, + 472 + ], + "type": "inline_equation", + "content": "44\\%" + }, + { + "bbox": [ + 138, + 390, + 480, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 138, + 390, + 480, + 472 + ], + "type": "inline_equation", + "content": "33\\%" + }, + { + "bbox": [ + 138, + 390, + 480, + 472 + ], + "type": "text", + "content": " on the HCM [41] and RICH [25] datasets, respectively. We also compare with the contemporary work WHAM [81] and outperform it by " + }, + { + "bbox": [ + 138, + 390, + 480, + 472 + ], + "type": "inline_equation", + "content": "49\\%" + }, + { + "bbox": [ + 138, + 390, + 480, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 138, + 390, + 480, + 472 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 138, + 390, + 480, + 472 + ], + "type": "text", + "content": " on the RICH and EMDB datasets, respectively." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 495, + 237, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 237, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 237, + 506 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Camera-Space Human Pose Estimation. Most existing works focus on root-relative local human pose estimation to bypass the difficulty in monocular depth estimation [1, 4, 6-8, 16, 32-34, 39, 42-46, 51, 54, 63, 64, 67, 68, 70, 76, 78, 84, 85, 88, 94, 97, 106, 110, 116, 118]. These methods ignore the position of the person in the camera coordinates. To overcome this limitation, recent methods estimate camera-space human poses by regression [28, 31, 48, 52, 62, 72, 75, 80, 93, 95, 107, 107, 109, 111] or optimization [59-61, 77, 108]. Physics-based constraints are widely used to ensure the plausibility of the estimated poses [10, 13, 27, 29, 80, 95, 105]. In addition to direct regression, heatmap-based representations have also been used to predict the absolute depths of multiple people [11, 87, 117]. A few methods improve absolution depth estimation by using predicted camera parameters [40, 49, 109] instead of the predefined focal length. Despite the promising results" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "text", + "content": "COIN: Control-Inpainting Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "for camera-space pose estimation, how to decouple the camera movement and estimate global human motions is still an open problem." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 163, + 481, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 163, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 163, + 481, + 450 + ], + "type": "text", + "content": "Monocular Global Human Pose Estimation. Recovering the global human motion from a monocular moving camera is challenging due to the entanglement of human and camera motions. To disentangle the camera movement, several methods use IMU sensors or pre-scanned environments to recover global human motions [17,20,57,69], which is impractical for large-scale adoption. Recent works use human motion priors [100] or physics-based constraints [47,55] to recover human motions from monocular videos, but do not consider background scene features, which limits performance on in-the-wild videos. Sun et al. [86] use optical flow as a motion cue to estimate the global motions. A contemporary work, WHAM [81], uses a lifting network to estimate global human motions from 2D keypoints and camera angular velocities. While these works can estimate accurate global human motions, they do not recover camera motions. To explicitly recover the camera motion, Liu et al. [52] use SLAM and convert the local pose from the camera to global coordinates. BodySLAM [22] jointly optimizes the human and camera motion using features of both humans and scenes. Along this line, SLAHMR [99] and PACE [41] use SLAM to initialize camera motions and optimize the camera using human motion priors [21,76]. However, these methods rely on the human motion priors to regularize the camera motion, which may lead to inaccurate camera motion when the human motion is not well initialized (as shown in Fig. 1). Such wrong camera trajectories will further affect the optimization of human motions. In contrast, our approach relies on the consistency among the local human motion, scene features, and the camera for optimization, which provides information that complements the human motion priors and enables accurate estimation of both human and camera motions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 474, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 474, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 474, + 482, + 665 + ], + "type": "text", + "content": "Human Motion Priors. There are a significant amount of approaches proposed to study 3D human dynamics for motion prediction and synthesis [2, 3, 5, 12, 15, 18, 19, 23, 30, 36, 38, 50, 58, 71, 73, 92, 98, 101-103]. These learned human motion priors are used to help resolve pose ambiguity [21, 39, 76, 115] in human pose estimation. Recently, diffusion models [82] have also been used as priors for motion synthesis and infilling [26, 35, 91, 104, 113]. RoHM [114], adopts motion diffusion model to recover human motions from noisy and occluded input data. Xie et al. [96] use spatial control signals to guide motion generation. They focus on generating realistic human motion given clean spatial constraints. Müller et al. [65] build a diffusion model to learn the joint distribution over the poses of two people. They use the SDS loss to guide the generation of static poses. In contrast, we focus on dynamic human motions. We find adopting SDS directly for temporal human motions encounters the inconsistency issue. Therefore, to distill knowledge from the motion diffusion model, we propose a novel control inpainting SDS to generate high-quality and consistent motion that aligns with observed evidence." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "L. Jiefeng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 479, + 299 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 479, + 299 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 479, + 299 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 479, + 299 + ], + "type": "image", + "image_path": "de1e8367b8e3c658791af30b4878cabc66b63e7163ab60536295821c1707354f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 301, + 482, + 388 + ], + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 388 + ], + "type": "text", + "content": "Fig. 2: Overview. Given a video with a moving camera, we recover the global human motion " + }, + { + "bbox": [ + 130, + 301, + 482, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 130, + 301, + 482, + 388 + ], + "type": "text", + "content": " and camera motion " + }, + { + "bbox": [ + 130, + 301, + 482, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 301, + 482, + 388 + ], + "type": "text", + "content": " using an iterative optimization framework. We propose a novel Control-Inpainting SDS loss (" + }, + { + "bbox": [ + 130, + 301, + 482, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{COIN - SDS}}" + }, + { + "bbox": [ + 130, + 301, + 482, + 388 + ], + "type": "text", + "content": ") to leverage motion diffusion models as a prior. COIN-SDS is designed such that the sampled motions from the motion prior are consistent with video observations. We achieve this by controlling and constraining the sampling process of the motion diffusion model through novel control and soft-inpainting branches. We also propose a novel human-scene relation loss (" + }, + { + "bbox": [ + 130, + 301, + 482, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{HSR}}" + }, + { + "bbox": [ + 130, + 301, + 482, + 388 + ], + "type": "text", + "content": ") to encourage consistency among the human motion, camera motion, and scene features." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 395, + 201, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 395, + 201, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 395, + 201, + 407 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 419, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 482, + 598 + ], + "type": "text", + "content": "The overall framework of COIN is illustrated in Fig. 2. Given an in-the-wild RGB video with " + }, + { + "bbox": [ + 130, + 419, + 482, + 598 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 419, + 482, + 598 + ], + "type": "text", + "content": " frames captured by a dynamic camera, our goal is to estimate both the global human motion " + }, + { + "bbox": [ + 130, + 419, + 482, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{H} = \\{\\mathbf{h}^{(1)},\\mathbf{h}^{(2)},\\dots,\\mathbf{h}^{(T)}\\}" + }, + { + "bbox": [ + 130, + 419, + 482, + 598 + ], + "type": "text", + "content": " and the camera motion " + }, + { + "bbox": [ + 130, + 419, + 482, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{\\mathbf{c}^{(1)},\\mathbf{c}^{(2)},\\dots,\\mathbf{c}^{(T)}\\}" + }, + { + "bbox": [ + 130, + 419, + 482, + 598 + ], + "type": "text", + "content": " in a global world coordinate system. We use off-the-shelf 3D human pose and shape estimation method HybrIK [48] to obtain per-frame initial SMPL parameters in the camera space and DROID-SLAM [89] to obtain the initial per-frame camera-to-world transforms. We convert the local human motion to the world coordinates with the estimated camera. However, because the camera trajectories from SLAM are up to an unknown scale, the initial global human motion will abnormally drift and float in the world space. To resolve the scale ambiguity and place the person in the correct global position, we jointly optimize the human and camera motion to minimize the discrepancy between the observed evidence and the estimated motion, while maintaining the plausibility of the human motion with a diffusion prior through the proposed control-inpainting SDS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Motion Representation. The camera motion is represented by the trajectory " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{(\\mathbf{R}^{(i)},\\mathbf{t}^{(i)})\\}_{i = 1}^{T}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "[\\mathbf{R}^{(i)},\\mathbf{t}^{(i)}]" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " is the camera pose at the " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "-th frame, consisting of the rotation matrix " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^{(i)}\\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " and the translation vector " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{t}^{(i)}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ". The human motion is represented by the human trajectory " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{H} = \\{\\mathbf{h}^{(i)}\\}_{i = 1}^{T}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{h}^{(i)} = [\\tau^{(i)},\\varPhi^{(i)},\\theta^{(i)},f^{(i)},\\beta ]" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " is the human pose at the " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "-th frame, consisting" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "text", + "content": "COIN: Control-Inpainting Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": "of the global translation " + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\tau^{(i)}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": ", global orientation " + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\varPhi^{(i)}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": ", body pose parameters " + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\theta^{(i)}\\in \\mathbb{R}^{23\\times 3}" + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": ", foot contact labels " + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "f\\in \\{0,1\\} ^4" + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": ", and the body shape parameters " + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\beta \\in \\mathbb{R}^{10}" + }, + { + "bbox": [ + 130, + 115, + 482, + 200 + ], + "type": "text", + "content": ". We use the SMPL model [53] to represent the human pose and shape. Before introducing our approach, we first revisit the formulation and drawbacks of SDS in Sec. 3.1. Then we introduce the proposed control-inpainting SDS in Sec. 3.2. Finally, we present the global optimization pipeline with the proposed human-scene interaction loss in Sec. 3.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 212, + 238, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 212, + 238, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 212, + 238, + 224 + ], + "type": "text", + "content": "3.1 Revisiting SDS" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 231, + 480, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 231, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 480, + 301 + ], + "type": "text", + "content": "Score Distillation Sampling (SDS) was first introduced to distill 3D assets from pre-trained 2D text-to-image diffusion models [74]. It exploits the knowledge from the diffusion models by seeking modes for the conditional distribution in the DDPM latent space to optimize the 3D scene representation. Similarly, we can optimize global human motion by distilling knowledge from a pre-trained motion diffusion model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 303, + 481, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 303, + 481, + 327 + ], + "spans": [ + { + "bbox": [ + 131, + 303, + 481, + 327 + ], + "type": "text", + "content": "Given an global human motion " + }, + { + "bbox": [ + 131, + 303, + 481, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 131, + 303, + 481, + 327 + ], + "type": "text", + "content": ", the marginal distribution of noisy latent " + }, + { + "bbox": [ + 131, + 303, + 481, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_t" + }, + { + "bbox": [ + 131, + 303, + 481, + 327 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 131, + 303, + 481, + 327 + ], + "type": "inline_equation", + "content": "t \\in \\mathcal{U}(0,1)" + }, + { + "bbox": [ + 131, + 303, + 481, + 327 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 226, + 335, + 481, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 335, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 226, + 335, + 481, + 350 + ], + "type": "interline_equation", + "content": "q (\\mathbf {H} _ {t} | \\mathbf {H}) = \\mathcal {N} (\\mathbf {H} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {H}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}), \\tag {1}", + "image_path": "b755c9ab906b7fa4d561cd4e1c82f90649448a8d69a9e62ef322ccc275528225.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 357, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 357, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 131, + 357, + 482, + 441 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 357, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t\\in (0,1)" + }, + { + "bbox": [ + 131, + 357, + 482, + 441 + ], + "type": "text", + "content": " is a hyperparameter controlled by the variance schedule of the diffusion model. SDS adopts the pre-trained diffusion model " + }, + { + "bbox": [ + 131, + 357, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\phi}(\\mathbf{H}_t,t,y)" + }, + { + "bbox": [ + 131, + 357, + 482, + 441 + ], + "type": "text", + "content": ", which takes in " + }, + { + "bbox": [ + 131, + 357, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_t" + }, + { + "bbox": [ + 131, + 357, + 482, + 441 + ], + "type": "text", + "content": " and is used to model the conditional density of the human motion, where " + }, + { + "bbox": [ + 131, + 357, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 131, + 357, + 482, + 441 + ], + "type": "text", + "content": " are the parameters of the diffusion model and " + }, + { + "bbox": [ + 131, + 357, + 482, + 441 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 131, + 357, + 482, + 441 + ], + "type": "text", + "content": " is the condition. Then, SDS aims to distill global human motion " + }, + { + "bbox": [ + 131, + 357, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 131, + 357, + 482, + 441 + ], + "type": "text", + "content": " via seeking modes of the learned conditional density, which can be achieved by a weighted denoising score matching objective:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 231, + 449, + 481, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 449, + 481, + 467 + ], + "spans": [ + { + "bbox": [ + 231, + 449, + 481, + 467 + ], + "type": "interline_equation", + "content": "\\underset {\\mathbf {H}} {\\min } \\mathcal {L} _ {\\mathrm {S D S}} := \\mathbb {E} _ {t, \\epsilon} \\left[ \\omega (t) \\| \\epsilon_ {\\phi} ^ {t} - \\epsilon \\| _ {2} ^ {2} \\right], \\tag {2}", + "image_path": "05d669baf97716d8dde097f299b11bf454e3796b41c07fbecbe563d70dd3684d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 475, + 481, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 475, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 131, + 475, + 481, + 512 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 475, + 481, + 512 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{t}" + }, + { + "bbox": [ + 131, + 475, + 481, + 512 + ], + "type": "text", + "content": " is the predicted denoising direction from the diffusion model, " + }, + { + "bbox": [ + 131, + 475, + 481, + 512 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_t \\sim q(\\mathbf{H}_t|\\mathbf{H})" + }, + { + "bbox": [ + 131, + 475, + 481, + 512 + ], + "type": "text", + "content": " is sampled using the reparameterization trick, " + }, + { + "bbox": [ + 131, + 475, + 481, + 512 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 131, + 475, + 481, + 512 + ], + "type": "text", + "content": " is the corresponding sampled noise, and " + }, + { + "bbox": [ + 131, + 475, + 481, + 512 + ], + "type": "inline_equation", + "content": "\\omega(t)" + }, + { + "bbox": [ + 131, + 475, + 481, + 512 + ], + "type": "text", + "content": " is a weighting function that depends on the timestep " + }, + { + "bbox": [ + 131, + 475, + 481, + 512 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 475, + 481, + 512 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 146, + 512, + 446, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 512, + 446, + 523 + ], + "spans": [ + { + "bbox": [ + 146, + 512, + 446, + 523 + ], + "type": "text", + "content": "To clearly review the effect of SDS, we can reparameterize Eq. 2 as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 217, + 531, + 481, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 531, + 481, + 559 + ], + "spans": [ + { + "bbox": [ + 217, + 531, + 481, + 559 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {H}} \\mathcal {L} _ {\\mathrm {S D S}} := \\mathbb {E} _ {t} \\left[ \\frac {\\omega (t) \\sqrt {\\bar {\\alpha} _ {t}}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\left\\| \\mathbf {H} - \\hat {\\mathbf {H}} _ {0} ^ {t} \\right\\| _ {2} ^ {2} \\right], \\tag {3}", + "image_path": "c40dd253fcf43a8917ecc0acc83f3882ec366af859287192e97b1486085c12d2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 566, + 161, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 566, + 161, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 161, + 575 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 253, + 574, + 481, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 574, + 481, + 602 + ], + "spans": [ + { + "bbox": [ + 253, + 574, + 481, + 602 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {H}} _ {0} ^ {t} = \\frac {\\mathbf {H} _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon} _ {\\phi} ^ {t}}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {4}", + "image_path": "6b0e463bc7fa0af7bfa79e7860b746540a5e3c8ec3e4e55ca721c8a554f71fbb.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 482, + 666 + ], + "type": "text", + "content": "Based on this reparameterization, we can see that the SDS objective is to minimize the discrepancy between global human motion " + }, + { + "bbox": [ + 131, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 131, + 605, + 482, + 666 + ], + "type": "text", + "content": " and the denoised global human motion " + }, + { + "bbox": [ + 131, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{H}}_0^t" + }, + { + "bbox": [ + 131, + 605, + 482, + 666 + ], + "type": "text", + "content": " from the motion diffusion model in a single step. The denoised motion " + }, + { + "bbox": [ + 131, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{H}}_0^t" + }, + { + "bbox": [ + 131, + 605, + 482, + 666 + ], + "type": "text", + "content": " serves as the pseudo ground truth. However, at each optimization step, we randomly sample " + }, + { + "bbox": [ + 131, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 605, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 131, + 605, + 482, + 666 + ], + "type": "text", + "content": " to generate the noisy latent " + }, + { + "bbox": [ + 131, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_t" + }, + { + "bbox": [ + 131, + 605, + 482, + 666 + ], + "type": "text", + "content": ", and we found" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "L. Jiefeng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "the pre-trained diffusion model is sensitive to the input. Minor fluctuations in the input latent would substantially change the denoised motion, which leads to inconsistency in " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{H}}_0^t" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " across different time steps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 308 + ], + "type": "text", + "content": "Although randomness can help generate diverse plausible motions to infer occluded regions and unknown information, we do not need it for well-observed regions, such as simple body poses in a clean background. Such randomness in the denoising steps makes the generated " + }, + { + "bbox": [ + 130, + 152, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{H}}_0^t" + }, + { + "bbox": [ + 130, + 152, + 482, + 308 + ], + "type": "text", + "content": " difficult to align with the local 2D observations and results in wrong global human motion. Moreover, this pseudo ground truth " + }, + { + "bbox": [ + 130, + 152, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{H}}_0^t" + }, + { + "bbox": [ + 130, + 152, + 482, + 308 + ], + "type": "text", + "content": " is generated from only a single denoising step, where the diffusion models may not produce high-quality motions, resulting in foot sliding and floating. Although sampling with a smaller timestep " + }, + { + "bbox": [ + 130, + 152, + 482, + 308 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 152, + 482, + 308 + ], + "type": "text", + "content": " can alleviate these issues, the initial motion is usually inaccurate and the denoiser is not able to remove artifacts with a small " + }, + { + "bbox": [ + 130, + 152, + 482, + 308 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 152, + 482, + 308 + ], + "type": "text", + "content": ". To exploit the knowledge of the motion diffusion model and denoise the initial motion, we must allow the SDS to sample with a larger timestep " + }, + { + "bbox": [ + 130, + 152, + 482, + 308 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 152, + 482, + 308 + ], + "type": "text", + "content": " while maintaining high quality, consistency, and alignment with the local 2D observations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 320, + 279, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 320, + 279, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 279, + 332 + ], + "type": "text", + "content": "3.2 Control-Inpainting SDS" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 341, + 482, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 473 + ], + "type": "text", + "content": "The limitations of SDS originate from the randomness and inconsistency of the denoised motion " + }, + { + "bbox": [ + 130, + 341, + 482, + 473 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{H}}_0^t" + }, + { + "bbox": [ + 130, + 341, + 482, + 473 + ], + "type": "text", + "content": ", which serves as the pseudo ground truth in the objective function. To overcome this issue, we propose a novel COntrol-INpainting SDS (COIN-SDS) to generate high-quality and consistent pseudo-ground-truth motions. Our solution has three key ingredients (shown in Alg. 1). First, to achieve high-quality motions, we seek to produce the pseudo ground truth with multiple DDIM denoising steps. Second, to encourage consistent motions, we propose to use partially observed evidence from the video as a control signal to dynamically guide the diffusion model and align the generated motions with the observations. Third, to further align the motion with observed regions, we propose a soft inpainting strategy within the denoising process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 480, + 481, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 481, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 481, + 517 + ], + "type": "text", + "content": "Multiple Denoising Steps. Intuitively, to obtain high-quality pseudo ground truth for SDS, we can replace the single-step denoised motion " + }, + { + "bbox": [ + 130, + 480, + 481, + 517 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{H}}_0^t" + }, + { + "bbox": [ + 130, + 480, + 481, + 517 + ], + "type": "text", + "content": " with a multi-step one " + }, + { + "bbox": [ + 130, + 480, + 481, + 517 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{H}}_0^t \\coloneqq \\tilde{\\mathbf{H}}_0" + }, + { + "bbox": [ + 130, + 480, + 481, + 517 + ], + "type": "text", + "content": ", following the multi-step DDIM denoising process [83]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 216, + 525, + 480, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 525, + 480, + 539 + ], + "spans": [ + { + "bbox": [ + 216, + 525, + 480, + 539 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {H}} _ {t - \\Delta t} = \\sqrt {\\bar {\\alpha} _ {t - \\Delta t}} \\cdot \\hat {\\mathbf {H}} _ {0} ^ {t} + \\sqrt {1 - \\bar {\\alpha} _ {t - \\Delta t}} \\cdot \\boldsymbol {\\epsilon} _ {\\phi} ^ {t}, \\tag {5}", + "image_path": "4de9eaff6597f8886904e1c6ef372dea24b472896c3af4fd6718c7433421a618.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 549, + 481, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 549, + 481, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 549, + 481, + 574 + ], + "type": "text", + "content": "until " + }, + { + "bbox": [ + 130, + 549, + 481, + 574 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{H}}_0 = \\tilde{\\mathbf{H}}_{t - \\Delta t}" + }, + { + "bbox": [ + 130, + 549, + 481, + 574 + ], + "type": "text", + "content": " is obtained. By replacing " + }, + { + "bbox": [ + 130, + 549, + 481, + 574 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{H}}_0^t" + }, + { + "bbox": [ + 130, + 549, + 481, + 574 + ], + "type": "text", + "content": " in Eq. 3 with " + }, + { + "bbox": [ + 130, + 549, + 481, + 574 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{H}}_0" + }, + { + "bbox": [ + 130, + 549, + 481, + 574 + ], + "type": "text", + "content": ", we can obtain a new objective for SDS:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 220, + 584, + 480, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 584, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 220, + 584, + 480, + 609 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {H}} \\mathcal {L} _ {\\mathrm {S D S}} := \\mathbb {E} _ {t} \\left[ \\frac {\\omega (t) \\sqrt {\\bar {\\alpha} _ {t}}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\left\\| \\mathbf {H} - \\tilde {\\mathbf {H}} _ {0} \\right\\| _ {2} ^ {2} \\right]. \\tag {6}", + "image_path": "b6a3a60307ca6c703a843756df3dcde8187412655c2fd64b75b2676eaf0f9f0a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 487, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 487, + 677 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 487, + 677 + ], + "type": "text", + "content": "Although the multi-step denoising process can produce high-quality pseudo ground truth, it is computationally expensive to perform multiple denoising steps during optimization, which limits the practicality of increasing the number of denoising steps. In our experiments, we find that using 10 denoising steps is sufficient to produce high-quality pseudo ground truth." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "text", + "content": "COIN: Control-Inpainting Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 138, + 132, + 468, + 285 + ], + "blocks": [ + { + "bbox": [ + 140, + 118, + 258, + 129 + ], + "lines": [ + { + "bbox": [ + 140, + 118, + 258, + 129 + ], + "spans": [ + { + "bbox": [ + 140, + 118, + 258, + 129 + ], + "type": "text", + "content": "Algorithm 1: COIN-SDS" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 138, + 132, + 468, + 285 + ], + "lines": [ + { + "bbox": [ + 138, + 132, + 468, + 285 + ], + "spans": [ + { + "bbox": [ + 138, + 132, + 468, + 285 + ], + "type": "text", + "content": "Input: Latest human motion H, confidence score S, visible mask M \nOutput: " + }, + { + "bbox": [ + 138, + 132, + 468, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{COIN - SDS}}" + }, + { + "bbox": [ + 138, + 132, + 468, + 285 + ], + "type": "text", + "content": " \n1 Sample: " + }, + { + "bbox": [ + 138, + 132, + 468, + 285 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{H}}_t \\sim \\mathcal{N}(\\mathbf{H}_t; \\sqrt{\\bar{\\alpha}_t}\\mathbf{H}, (1 - \\bar{\\alpha}_t)\\mathbf{I}), t \\sim \\mathcal{U}(0,1)" + }, + { + "bbox": [ + 138, + 132, + 468, + 285 + ], + "type": "text", + "content": "; \n2 for " + }, + { + "bbox": [ + 138, + 132, + 468, + 285 + ], + "type": "inline_equation", + "content": "\\bar{t} = [t, t - \\Delta t, \\dots, \\Delta t]" + }, + { + "bbox": [ + 138, + 132, + 468, + 285 + ], + "type": "text", + "content": " do // multi-step DDIM denoising \n3 " + }, + { + "bbox": [ + 138, + 132, + 468, + 285 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{H}}_0^{\\bar{t},\\mathrm{known}} \\gets \\mathbf{H}" + }, + { + "bbox": [ + 138, + 132, + 468, + 285 + ], + "type": "text", + "content": "; \n4 " + }, + { + "bbox": [ + 138, + 132, + 468, + 285 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{H}}_0^{\\bar{t},\\mathrm{un-known}} \\gets \\mathcal{D}_{\\phi,\\phi_c}(\\tilde{\\mathbf{H}}_{\\bar{t}}, \\bar{t}, \\mathbf{H} \\odot \\mathbf{M})" + }, + { + "bbox": [ + 138, + 132, + 468, + 285 + ], + "type": "text", + "content": "; // controlled denoising \n5 " + }, + { + "bbox": [ + 138, + 132, + 468, + 285 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{M}} \\gets w(\\bar{t}) * \\mathbf{S} \\odot \\mathbf{M}" + }, + { + "bbox": [ + 138, + 132, + 468, + 285 + ], + "type": "text", + "content": "; \n6 " + }, + { + "bbox": [ + 138, + 132, + 468, + 285 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{H}}_0^{\\bar{t}} \\gets \\tilde{\\mathbf{M}} \\odot \\tilde{\\mathbf{H}}_0^{\\bar{t},\\mathrm{known}} + (1 - \\tilde{\\mathbf{M}}) \\odot \\tilde{\\mathbf{H}}_0^{\\bar{t},\\mathrm{unknown}}" + }, + { + "bbox": [ + 138, + 132, + 468, + 285 + ], + "type": "text", + "content": "; // soft inpainting \n7 " + }, + { + "bbox": [ + 138, + 132, + 468, + 285 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{\\bar{t}} \\gets \\frac{\\tilde{\\mathbf{H}}_{\\bar{t}} - \\sqrt{\\bar{\\alpha}_{\\bar{t}}} \\tilde{\\mathbf{H}}_0^{\\bar{t}}}{\\sqrt{1 - \\bar{\\alpha}_{\\bar{t}}}}" + }, + { + "bbox": [ + 138, + 132, + 468, + 285 + ], + "type": "text", + "content": "; \n8 " + }, + { + "bbox": [ + 138, + 132, + 468, + 285 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{H}}_{\\bar{t} -\\Delta t} \\gets \\sqrt{\\bar{\\alpha}_{\\bar{t} -\\Delta t}} \\cdot \\tilde{\\mathbf{H}}_0^{\\bar{t}} + \\sqrt{1 - \\bar{\\alpha}_{\\bar{t} -\\Delta t}} \\cdot \\epsilon_{\\phi}^{\\bar{t}}" + }, + { + "bbox": [ + 138, + 132, + 468, + 285 + ], + "type": "text", + "content": "; // update latent motion \n9 " + }, + { + "bbox": [ + 138, + 132, + 468, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{COIN - SDS}} = \\frac{\\omega(t) \\sqrt{\\bar{\\alpha}_t}}{\\sqrt{1 - \\bar{\\alpha}_t}}\\left\\| \\mathbf{H} - \\tilde{\\mathbf{H}}_0 \\right\\|_2^2" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 289, + 482, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 352 + ], + "type": "text", + "content": "Dynamic Controlled Sampling. To generate consistent motions that are aligned with the observed evidence, we propose to attach a control branch " + }, + { + "bbox": [ + 130, + 289, + 482, + 352 + ], + "type": "inline_equation", + "content": "\\phi_c" + }, + { + "bbox": [ + 130, + 289, + 482, + 352 + ], + "type": "text", + "content": " to the pre-trained diffusion model " + }, + { + "bbox": [ + 130, + 289, + 482, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\phi}" + }, + { + "bbox": [ + 130, + 289, + 482, + 352 + ], + "type": "text", + "content": " to guide the motion generation. Given a latent motion " + }, + { + "bbox": [ + 130, + 289, + 482, + 352 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{H}}_t" + }, + { + "bbox": [ + 130, + 289, + 482, + 352 + ], + "type": "text", + "content": ", control signal " + }, + { + "bbox": [ + 130, + 289, + 482, + 352 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 289, + 482, + 352 + ], + "type": "text", + "content": ", and the visible mask " + }, + { + "bbox": [ + 130, + 289, + 482, + 352 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 289, + 482, + 352 + ], + "type": "text", + "content": ", we train a controlled denoiser " + }, + { + "bbox": [ + 130, + 289, + 482, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\phi,\\phi_c}" + }, + { + "bbox": [ + 130, + 289, + 482, + 352 + ], + "type": "text", + "content": " to generate intermediate denoised motion " + }, + { + "bbox": [ + 130, + 289, + 482, + 352 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{H}}_0^t" + }, + { + "bbox": [ + 130, + 289, + 482, + 352 + ], + "type": "text", + "content": " for DDIM denoising:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 247, + 364, + 481, + 379 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 364, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 247, + 364, + 481, + 379 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {H}} _ {0} ^ {t} = \\mathcal {D} _ {\\phi , \\phi_ {c}} \\left(\\tilde {\\mathbf {H}} _ {t}, t, \\mathbf {c} \\odot \\mathbf {M}\\right), \\tag {7}", + "image_path": "8d2e9bda16dfad10c0ec94e144345202ed9e23c54c5d3354e100cd4759c6653c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 387, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 387, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 387, + 482, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 387, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 387, + 482, + 449 + ], + "type": "text", + "content": " are the same size as the motion, " + }, + { + "bbox": [ + 130, + 387, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 387, + 482, + 449 + ], + "type": "text", + "content": " is a binary mask with ones in observed pose dimensions, and " + }, + { + "bbox": [ + 130, + 387, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 387, + 482, + 449 + ], + "type": "text", + "content": " denotes the element-wise multiplication. During training, we synthesize noise and occlusions by randomly adding Gaussian noise to the control signal " + }, + { + "bbox": [ + 130, + 387, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 387, + 482, + 449 + ], + "type": "text", + "content": " and randomly masking the pose and trajectory dimensions in " + }, + { + "bbox": [ + 130, + 387, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 387, + 482, + 449 + ], + "type": "text", + "content": ". Details are provided in the supplementary document." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": "When using the denoiser in the optimization stage, instead of always using the initial noisy estimation from HybrIK [48] as the fixed control signal, we propose to use a dynamic control strategy. Specifically, we use the optimized human motion from the previous iteration as the control signal, i.e., " + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{c} = \\mathbf{H}" + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": ". This strategy prevents performance degradation due to inaccurate initializations. A better control signal can guide the model to generate a more plausible pseudo motion, which in turn helps to optimize the global human motion and provides a control signal that better aligns with the input videos. Such self-evolving control signals can help to generate well-aligned global human motions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "The pre-trained motion diffusion model adopts a transformer encoder structure like [91]. We follow ControlNet [112] to encode the control signals and guide the denoiser output. We create a trainable copy of 4 encoding blocks of the pretrained motion diffusion model followed with zero convolutions. The input to the control branch is the concatenation of the latent and the control signals. The AMASS [56] dataset is used to train the motion diffusion model. The finetuning of the controlled denoiser is computationally efficient since the pre-trained branch is frozen and only the control branch is trained. See the supplementary document for more details on the architecture and training settings." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "L. Jiefeng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "text", + "content": "Soft Inpainting. While guiding motion generation encourages outputs to align with the conditions, it is often not strong enough. We further seek to improve the consistency by masking the known regions and inpainting the unknown regions. Given a binary mask " + }, + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "text", + "content": " that indicates the observed and unobserved regions, we can use the diffusion model to generate the inpainted motion " + }, + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{H}}_0^t" + }, + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "text", + "content": " following the DDIM denoising process:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 209, + 194, + 479, + 207 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 194, + 479, + 207 + ], + "spans": [ + { + "bbox": [ + 209, + 194, + 479, + 207 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {H}} _ {0} ^ {t, \\text {k n o w n}} = \\mathbf {H}, \\tag {8a}", + "image_path": "38cc5d6ad94895006a3b1a668427b13850587e3d9964643628cca381a1a77585.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 209, + 211, + 479, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 211, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 209, + 211, + 479, + 224 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {H}} _ {0} ^ {t, \\text {u n k n o w n}} = \\mathcal {D} _ {\\phi , \\phi_ {c}} \\left(\\tilde {\\mathbf {H}} _ {t}, t, \\mathbf {H} \\odot \\mathbf {M}\\right), \\tag {8b}", + "image_path": "85862fc7763574cae540bc32697b335ea386fe676dd0c5e441b1220530b14ffa.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 209, + 228, + 479, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 228, + 479, + 242 + ], + "spans": [ + { + "bbox": [ + 209, + 228, + 479, + 242 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {H}} _ {0} ^ {t} = \\mathbf {M} \\odot \\tilde {\\mathbf {H}} _ {0} ^ {t, \\text {k n o w n}} + (1 - \\mathbf {M}) \\odot \\tilde {\\mathbf {H}} _ {0} ^ {t, \\text {u n k n o w n}}. \\tag {8c}", + "image_path": "5aeaf8ece4ff4d273aedd934bc328ff5fff1a537ca3efe05c57d1afcc524626f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 247, + 479, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 247, + 479, + 307 + ], + "spans": [ + { + "bbox": [ + 131, + 247, + 479, + 307 + ], + "type": "text", + "content": "Thus, the known regions are overwritten with the observations, while the unknown regions are sampled from the diffusion model. However, the above formulation keeps the observed parts unchanged during the denoising process. In practice, the observed parts can be noisy and not perfect. We still want the diffusion model to refine the observed parts but not change them significantly." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 308, + 479, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 308, + 479, + 367 + ], + "spans": [ + { + "bbox": [ + 131, + 308, + 479, + 367 + ], + "type": "text", + "content": "Here, we present a soft inpainting strategy to infill the unobserved regions while refining the observed regions by dynamically reweighting the denoised direction from the diffusion model. Specifically, instead of using a binary mask, we adopt a continuous mask " + }, + { + "bbox": [ + 131, + 308, + 479, + 367 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{M}}" + }, + { + "bbox": [ + 131, + 308, + 479, + 367 + ], + "type": "text", + "content": " depending on both the confidence score of the observations " + }, + { + "bbox": [ + 131, + 308, + 479, + 367 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 131, + 308, + 479, + 367 + ], + "type": "text", + "content": " and the denoising time step " + }, + { + "bbox": [ + 131, + 308, + 479, + 367 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 308, + 479, + 367 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 264, + 374, + 479, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 374, + 479, + 386 + ], + "spans": [ + { + "bbox": [ + 264, + 374, + 479, + 386 + ], + "type": "interline_equation", + "content": "\\mathbf {M} = w (t) * \\mathbf {S} \\odot \\mathbf {M}, \\tag {9}", + "image_path": "444f76586bcd0de9a611f09556138d45ac13a8c4d1b7bdb8303f1bdd4ebc708f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 392, + 479, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 392, + 479, + 439 + ], + "spans": [ + { + "bbox": [ + 131, + 392, + 479, + 439 + ], + "type": "text", + "content": "where we set " + }, + { + "bbox": [ + 131, + 392, + 479, + 439 + ], + "type": "inline_equation", + "content": "w(t) = \\max(0, \\frac{t - 0.5}{0.5})" + }, + { + "bbox": [ + 131, + 392, + 479, + 439 + ], + "type": "text", + "content": " to linearly decrease the weight of the observations as the denoising time step decreases. As the time step decreases, the denoising process will be more deterministic and the model will be more certain of the generated motions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 441, + 479, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 441, + 479, + 464 + ], + "spans": [ + { + "bbox": [ + 131, + 441, + 479, + 464 + ], + "type": "text", + "content": "Combining the three components of our solution, the final objective for COIN-SDS is formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 183, + 468, + 479, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 468, + 479, + 495 + ], + "spans": [ + { + "bbox": [ + 183, + 468, + 479, + 495 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {H}} \\mathcal {L} _ {\\mathrm {C O I N - S D S}} := \\mathbb {E} _ {t} \\left[ \\frac {\\omega (t) \\sqrt {\\bar {\\alpha} _ {t}}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\left\\| \\mathbf {H} - \\tilde {\\mathbf {H}} _ {0} (\\mathbf {H}, \\mathbf {M}, \\mathbf {S}, t) \\right\\| _ {2} ^ {2} \\right]. \\tag {10}", + "image_path": "54f95d8e25471a5d3cfaa6ac367599cff6285be7662bd1af159ce89046f64d08.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 499, + 271, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 499, + 271, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 499, + 271, + 510 + ], + "type": "text", + "content": "We summarize COIN in Alg. 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 517, + 263, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 517, + 263, + 529 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 263, + 529 + ], + "type": "text", + "content": "3.3 Global Optimization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 535, + 479, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 535, + 479, + 643 + ], + "spans": [ + { + "bbox": [ + 131, + 535, + 479, + 643 + ], + "type": "text", + "content": "Here we present the overall optimization pipeline for the joint estimation of global human and camera motion with the proposed COIN-SDS loss. Note that we use SLAM to initialize the camera poses, which is scale-ambiguous. Therefore, we need to jointly optimize the camera scale " + }, + { + "bbox": [ + 131, + 535, + 479, + 643 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 131, + 535, + 479, + 643 + ], + "type": "text", + "content": " with the human and camera motions. Furthermore, the SLAM method assumes the camera in the first frame to be at the origin. To put the human motion in the correct positions, we follow PACE [41] and also optimize the camera height " + }, + { + "bbox": [ + 131, + 535, + 479, + 643 + ], + "type": "inline_equation", + "content": "h_0" + }, + { + "bbox": [ + 131, + 535, + 479, + 643 + ], + "type": "text", + "content": " and the orientation " + }, + { + "bbox": [ + 131, + 535, + 479, + 643 + ], + "type": "inline_equation", + "content": "R_0" + }, + { + "bbox": [ + 131, + 535, + 479, + 643 + ], + "type": "text", + "content": " for the first frame. The global human motions are initialized by the estimated local motions and the camera poses. The overall optimization objective is:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 220, + 651, + 479, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 651, + 479, + 667 + ], + "spans": [ + { + "bbox": [ + 220, + 651, + 479, + 667 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {H}, \\mathcal {C}, s, h _ {0}, R _ {0}, \\beta} \\mathcal {L} _ {\\text {b o d y}} + \\mathcal {L} _ {\\text {C O I N - S D S}} + \\mathcal {L} _ {\\text {H S R}}, \\tag {11}", + "image_path": "a9e8777bb7006f523e81efad42b30c660db7a9347432befe17cfdebf2d2cb8de.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "text", + "content": "COIN: Control-Inpainting Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 162, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 162, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 162, + 126 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 203, + 128, + 481, + 140 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 128, + 481, + 140 + ], + "spans": [ + { + "bbox": [ + 203, + 128, + 481, + 140 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {b o d y}} = \\mathcal {L} _ {2 \\mathrm {D}} + \\mathcal {L} _ {3 \\mathrm {D}} + \\mathcal {L} _ {\\beta} + \\mathcal {L} _ {\\text {s m o o t h}} + \\mathcal {L} _ {\\text {c o n t a c t}}. \\tag {12}", + "image_path": "b3fbf29342184a25ea4cf394df23bca2f85450ef1e571a0d1488cfceb138160f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 145, + 482, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 145, + 482, + 242 + ], + "spans": [ + { + "bbox": [ + 130, + 145, + 482, + 242 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{2\\mathrm{D}}" + }, + { + "bbox": [ + 130, + 145, + 482, + 242 + ], + "type": "text", + "content": " measures the 2D reprojection error between the projected 2D body joints of the estimated human motion and the detected 2D keypoints from an off-the-shelf 2D joint detector [9]. " + }, + { + "bbox": [ + 130, + 145, + 482, + 242 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{3\\mathrm{D}}" + }, + { + "bbox": [ + 130, + 145, + 482, + 242 + ], + "type": "text", + "content": " measures the distance between the estimated local 3D joints and the detected 3D joints from an off-the-shelf 3D joint detector [79]. " + }, + { + "bbox": [ + 130, + 145, + 482, + 242 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\beta}" + }, + { + "bbox": [ + 130, + 145, + 482, + 242 + ], + "type": "text", + "content": " is the shape regularization loss. " + }, + { + "bbox": [ + 130, + 145, + 482, + 242 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{smooth}}" + }, + { + "bbox": [ + 130, + 145, + 482, + 242 + ], + "type": "text", + "content": " is the temporal smoothness loss. " + }, + { + "bbox": [ + 130, + 145, + 482, + 242 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{contact}}" + }, + { + "bbox": [ + 130, + 145, + 482, + 242 + ], + "type": "text", + "content": " is the foot contact loss to encourage zero velocities for contact joints. The contact labels are obtained from the pseudo ground truth motion from COIN-SDS. Please refer to the supplementary document for more details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 245, + 482, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 245, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 482, + 352 + ], + "type": "text", + "content": "Human-scene Relation Loss. The camera trajectories recovered from SLAM are scale-ambiguous. Previous works [41,99] optimize the camera scale by projecting the global human motion to the camera space using the camera poses and minimizing the reprojection error. However, such a method entirely relies on the global human motions, which is in turn affected by the camera scale. If the human motion is not initialized well, the camera scale will also be inaccurate. To solve this problem, instead of the global human motions, we propose a new human-scene relation loss that uses the depth relation between the human and scene in the camera space, which disentangles the effect of the camera itself." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 353, + 483, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 483, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 483, + 472 + ], + "type": "text", + "content": "Specifically, we use the point cloud of the scene recovered by SLAM as a constraint. First of all, the scale of the scene point cloud is the same as the camera scale, so optimizing the scene scale is equivalent to optimizing the camera scale. Second, the scene points that are projected onto the visible vertices of the body mesh should be occluded by the person. Otherwise, the corresponding body parts are invisible. Therefore, we can constrain the depth of the occluded scene points to be larger than the depth of the human body vertices. While finding the corresponding body vertices for each scene point is time-consuming, we propose to use the depth of its nearest body joint as a proxy. Given the scene point cloud " + }, + { + "bbox": [ + 130, + 353, + 483, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 353, + 483, + 472 + ], + "type": "text", + "content": " and the camera scale " + }, + { + "bbox": [ + 130, + 353, + 483, + 472 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 353, + 483, + 472 + ], + "type": "text", + "content": ", the human-scene relation loss is formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 480, + 481, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 480, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 138, + 480, + 481, + 515 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {H S R}} = - \\frac {1}{| \\mathcal {P} |} \\sum_ {i = 1} ^ {T} \\sum_ {p \\in \\mathcal {P} ^ {*}} \\min (0, \\mathcal {T} ^ {(i)} (p) _ {z} - j ^ {(i)} (p) _ {z}) \\cdot \\mathbb {1} (\\mathcal {T} ^ {(i)} (p) \\text {i s i n v i s i b l e}), \\tag {13}", + "image_path": "0af2e0af99aaabd8f5459f8311af0c6dfb18ab230d94d105e598a8f1ad9cc979.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^* = \\mathcal{P}*s" + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "text", + "content": " is the scaled point cloud of the scene, " + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{(i)}(p) = \\mathbf{R}^{(i)}p + \\mathbf{t}^{(i)}" + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "text", + "content": " is the transformed point in the " + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "text", + "content": "-th frame, " + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "inline_equation", + "content": "j^{(i)}(p)" + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "text", + "content": " is body joint that has the nearest 2D projection to the scene point " + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "text", + "content": "-th frame, and " + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "text", + "content": " denotes the depth of a given point. If the depth order is correct, i.e., " + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{(i)}(p)_z - j^{(i)}(p)_z > 0" + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "text", + "content": ", the loss is zero. The proposed human-scene relation loss uses the relation between the local motions and the scene to alleviate the scale ambiguity. The depth relation regularizes consistency among humans, cameras, and scenes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 616, + 230, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 616, + 230, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 616, + 230, + 628 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Datasets. We perform experiments on three human motion datasets. First is the real-world dataset RICH [25]. We follow previous works [41,81,100] to assess the performance of global human motion estimation using this dataset. The second" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "L. Jiefeng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 126, + 479, + 274 + ], + "blocks": [ + { + "bbox": [ + 156, + 114, + 457, + 125 + ], + "lines": [ + { + "bbox": [ + 156, + 114, + 457, + 125 + ], + "spans": [ + { + "bbox": [ + 156, + 114, + 457, + 125 + ], + "type": "text", + "content": "Table 1: Global human motion estimation on the RICH dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 126, + 479, + 274 + ], + "lines": [ + { + "bbox": [ + 134, + 126, + 479, + 274 + ], + "spans": [ + { + "bbox": [ + 134, + 126, + 479, + 274 + ], + "type": "table", + "html": "
MethodPA-MPJPE ↓W-MPJPE ↓WA-MPJPE ↓W-RJE ↓ACCEL ↓
HybrIK [48] + SLAM [89]46.71073.1404.41166.220.2
GLAMR [100]79.9653.7365.1646.6107.7
SLAHMR [99]52.5571.6323.7400.59.4
WHAM [81]46.2497.6272.7478.26.7
PACE [41]49.3380.0197.2370.88.8
Guided Sampling132.81384.6502.21440.924.2
Noise Optimization66.7414.8195.3429.28.4
Vanilla SDS78.81453.5497.21458.012.7
COIN w/o Controlled Sampling44.0825.0291.8848.510.8
COIN w/o Dynamic Control49.5293.8180.6299.18.6
COIN w/o Soft Inpainting47.6325.8196.0324.99.6
COIN w/o LHSR43.6273.0176.3281.18.2
COIN42.9254.5169.5249.97.5
", + "image_path": "bb3c11c7ab28c1a80f2f087f56c70678ef27ee13a680353a92ec7222474283d0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 283, + 480, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 480, + 344 + ], + "type": "text", + "content": "one is EMDB [37]. We follow previous works [81] to evaluate on a subset of EMDB for which they provide ground truth global motion with dynamic cameras. The third dataset is HCM [41], a synthetic dataset. Compared to real-world datasets, HCM contains more challenging camera motions. We follow previous work [41] to evaluate the global human motion and the camera motion using this dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 348, + 480, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 480, + 456 + ], + "type": "text", + "content": "Metrics. We report various metrics for both human and camera motion. For human motion, standard metrics W-MPJPE and WA-MPJPE are used to evaluate global motion, while PA-MPJPE evaluates local motion. We also include an ACCEL metric to measure the joint acceleration error. For evaluation on EMDB, we follow previous work [81] to split sequences into smaller chunks of 100 frames and align each output segment with the ground-truth data using the first two frames W-MPJPE100 or the entire segment WA-MPJPE100. Root Orientation Error (ROE in deg) and Root Translation Error (RTE in " + }, + { + "bbox": [ + 130, + 348, + 480, + 456 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 348, + 480, + 456 + ], + "type": "text", + "content": ") evaluate the error over the entire trajectory after aligning with the initial camera pose." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 456, + 480, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 480, + 504 + ], + "type": "text", + "content": "For camera motion, we report the average translation error after scale alignment (ATE), without scale alignment (ATE-S), and the camera acceleration error (CAM ACCEL). ATE-S more accurately reflects inaccuracies in the captured scale of the camera." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 665 + ], + "type": "text", + "content": "Baselines. As discussed in Sec. 2, there are different ways to use the pre-trained motion diffusion model as a motion prior. Here, we summarize the three main solutions and compare them with COIN in Sec. 4.2. (1) Guided Sampling, which embeds analytical guidance within the denoising procedure using objective functions, such as 2D projection and foot contact consistency. This does not suit our task because the camera trajectories are also unknown and guided gradients from the wrong camera will lead to unrealistic human motions. We need to optimize the human motion and camera motion simultaneously. (2) Noise Optimization, which represents the motion as latent noise and directly optimizes it. This is similar to other motion priors such as VAE [21]. At each optimization step, we need to calculate the gradients of the latent w.r.t. the generated motion, which is computationally expensive and we find the performance is not good enough. (3) Vanilla SDS, which removes our design and directly uses SDS for optimization." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "text", + "content": "COIN: Control-Inpainting Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 126, + 479, + 208 + ], + "blocks": [ + { + "bbox": [ + 153, + 114, + 459, + 125 + ], + "lines": [ + { + "bbox": [ + 153, + 114, + 459, + 125 + ], + "spans": [ + { + "bbox": [ + 153, + 114, + 459, + 125 + ], + "type": "text", + "content": "Table 2: Global human motion estimation on the EMDB dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 126, + 479, + 208 + ], + "lines": [ + { + "bbox": [ + 134, + 126, + 479, + 208 + ], + "spans": [ + { + "bbox": [ + 134, + 126, + 479, + 208 + ], + "type": "table", + "html": "
MethodPA-MPJPE ↓W-MPJPE100 ↓WA-MPJPE100 ↓RTE ↓ROE ↓
HMR2.0 [14] + DPVO [90]49.62320.9662.917.544.4
GLAMR [100]56.0756.1286.216.774.9
TRACE [86]58.02244.9544.118.972.7
SLAHMR [99]61.5807.4336.913.867.9
WHAM [81]41.9439.2166.18.436.3
COIN32.7407.3152.83.534.1
", + "image_path": "726e55e2aae0e72f866febae22229525b4a2e2f154dabadd27badac029090694.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 243, + 479, + 326 + ], + "blocks": [ + { + "bbox": [ + 156, + 232, + 455, + 242 + ], + "lines": [ + { + "bbox": [ + 156, + 232, + 455, + 242 + ], + "spans": [ + { + "bbox": [ + 156, + 232, + 455, + 242 + ], + "type": "text", + "content": "Table 3: Global human motion estimation on the HCM dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 243, + 479, + 326 + ], + "lines": [ + { + "bbox": [ + 135, + 243, + 479, + 326 + ], + "spans": [ + { + "bbox": [ + 135, + 243, + 479, + 326 + ], + "type": "table", + "html": "
MethodPA-MPJE ↓W-MPJE ↓WA-MPJE ↓W-RJE ↓ACCEL ↓
HybrIK [48] + SLAM [89]67.61137.3780.31100.951.3
GLAMR [100]86.01977.6653.81958.033.4
SLAHMR [99]69.9888.9483.5862.214.9
PACE [41]65.3861.2478.3839.516.7
WHAM [81]47.9588.9279.3579.213.1
COIN45.5479.9212.1470.710.1
", + "image_path": "f760c1e05ff5c04a01aeee5aeaba45c7d699d9cf9bc0b130a9d7bc64a0f0ae4d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 330, + 380, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 380, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 380, + 341 + ], + "type": "text", + "content": "4.1 Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 352, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 472 + ], + "type": "text", + "content": "Human Motion Estimation. We compare COIN against state-of-the-art methods on the RICH, EMDB, and HCM datasets. Quantitative results are shown in Tabs. 1, 2, and 3. We observe that COIN significantly outperforms the state-of-the-art methods on all datasets. On the RICH dataset, COIN outperforms the state-of-the-art method, PACE [41], by " + }, + { + "bbox": [ + 130, + 352, + 482, + 472 + ], + "type": "inline_equation", + "content": "125.5\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 352, + 482, + 472 + ], + "type": "text", + "content": " in terms of W-MPJPE, showing " + }, + { + "bbox": [ + 130, + 352, + 482, + 472 + ], + "type": "inline_equation", + "content": "33.0\\%" + }, + { + "bbox": [ + 130, + 352, + 482, + 472 + ], + "type": "text", + "content": " relative improvement. On the EMDB and HCM datasets, COIN shows " + }, + { + "bbox": [ + 130, + 352, + 482, + 472 + ], + "type": "inline_equation", + "content": "29.1\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 352, + 482, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 352, + 482, + 472 + ], + "type": "inline_equation", + "content": "109.0\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 352, + 482, + 472 + ], + "type": "text", + "content": " improvement in terms of W-MPJPE, respectively. Qualitative comparisons with state-of-the-art methods, PACE [41] and WHAM [81], are shown in Figs. 1 and 3. More qualitative results are shown in the supplementary video." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "text", + "content": "COIN not only improves global motion, but also improves local body motion. In terms of PA-MPJPE, COIN shows " + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "inline_equation", + "content": "3.3\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "inline_equation", + "content": "9.2\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "inline_equation", + "content": "19.8\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "text", + "content": " improvement on the RICH, EMDB, and HCM datasets, respectively. This demonstrates that COIN is able to distill high-quality motion priors from the diffusion model and help both local and global motion estimation. Regarding the joint acceleration error, COIN is " + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "inline_equation", + "content": "0.8\\mathrm{mm/s}^2" + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "text", + "content": " higher than WHAM [81]. Note that the joint acceleration error reflects the smoothness. Human motion can be over-smoothed but not accurate, hence it is important to look at acceleration error in conjunction with other metrics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Camera Motion Estimation. We further evaluate the performance of COIN on camera motion estimation on the HCM dataset. Quantitative results are shown in Tab. 4. COIN substantially surpasses the state-of-the-art camera motion estimation methods. Specifically, COIN reduces the absolute camera translation error, ATE-S by " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "73.8\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": ". This demonstrates that COIN is able to disentangle human and camera motions and accurately estimate the camera motion." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "L. Jiefeng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 161, + 127, + 452, + 299 + ], + "blocks": [ + { + "bbox": [ + 171, + 114, + 441, + 125 + ], + "lines": [ + { + "bbox": [ + 171, + 114, + 441, + 125 + ], + "spans": [ + { + "bbox": [ + 171, + 114, + 441, + 125 + ], + "type": "text", + "content": "Table 4: Camera motion estimation on the HCM dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 161, + 127, + 452, + 299 + ], + "lines": [ + { + "bbox": [ + 161, + 127, + 452, + 299 + ], + "spans": [ + { + "bbox": [ + 161, + 127, + 452, + 299 + ], + "type": "table", + "html": "
MethodATE ↓ATE-S ↓CAMACCEL ↓
HybrIK [48] + SLAM [89]155.81670.717.1
GLAMR [100]1295.21714.6282.9
SLAHMR [99]155.8506.517.6
PACE [41]137.5459.716.2
Guided Sampling335.6992.415.6
Noise Optimization206.0500.912.3
Vanilla SDS306.4656.413.7
COIN w/o Controlled Sampling299.6553.814.0
COIN w/o Dynamic Control149.8397.711.3
COIN w/o Soft Inpainting167.7423.811.4
COIN w/o LHSR147.8402.112.0
COIN135.3385.911.3
", + "image_path": "7b4406d857db9a8537e5a8da0f2d23435b968e5fe2637b321c66001bf9c6fbe9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 308, + 238, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 238, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 238, + 320 + ], + "type": "text", + "content": "4.2 Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 327, + 480, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 480, + 373 + ], + "type": "text", + "content": "In this section, we conduct ablation studies on the RICH and HCM datasets to evaluate the impact of each component on human and camera motions, respectively. More comparisons on other datasets are provided in the supplementary document." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 380, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 511 + ], + "type": "text", + "content": "Baselines with Motion Diffusion Model. We first compare the aforementioned baselines with COIN. To use guided sampling in our tasks, we jointly update the camera poses using the gradients from the objective function during denoising. Quantitative results of human and camera motion estimation are shown in Tabs. 1 and 4, respectively. We observe that COIN outperforms all the baselines in terms of all metrics. As expected, guided sampling shows a terrible performance because it is not able to accurately estimate the camera trajectories. Noise optimization is better than vanilla SDS but worse than COIN, which generates consistent motion priors. Vanilla SDS is replacing " + }, + { + "bbox": [ + 130, + 380, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{COIN}}" + }, + { + "bbox": [ + 130, + 380, + 482, + 511 + ], + "type": "text", + "content": " with the SDS loss and keeping the rest of the settings the same. These results demonstrate the effectiveness of COIN over other motion diffusion baselines." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "text", + "content": "Impact of Dynamic Controlled Sampling. To study the effectiveness of controlled sampling, we compare COIN with and without using the controlled denoiser. Quantitative results are shown in Tabs. 1 and 4. We observe that controlled sampling significantly improves the performance of COIN. Specifically, COIN with controlled sampling reduces the W-MPJPE and ATE-S by " + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "inline_equation", + "content": "570.5\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "inline_equation", + "content": "167.9\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "text", + "content": ", showing " + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "inline_equation", + "content": "69.3\\%" + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "inline_equation", + "content": "30.3\\%" + }, + { + "bbox": [ + 130, + 516, + 482, + 624 + ], + "type": "text", + "content": " relative improvement, respectively. This demonstrates that controlled sampling is able to generate high-quality motion priors that align with the observed evidence and help improve human and camera motion estimation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Impact of Soft Inpainting. We further study the effectiveness of soft inpainting. Quantitative comparisons are shown in Tabs. 1 and 4. We observe that while soft inpainting also improves W-MPJPE, it is much more effective than" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "text", + "content": "COIN: Control-Inpainting Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 306, + 270 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 306, + 270 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 306, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 306, + 270 + ], + "type": "image", + "image_path": "3d507f94e555b157a84a398fb6034eb37a4e580bf6d730f47ec2d1f84b3455d2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 185, + 270, + 306, + 276 + ], + "lines": [ + { + "bbox": [ + 185, + 270, + 306, + 276 + ], + "spans": [ + { + "bbox": [ + 185, + 270, + 306, + 276 + ], + "type": "text", + "content": "PACE (Human Motion) PACE (Camera Motion)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 481, + 269 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 481, + 269 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 481, + 269 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 481, + 269 + ], + "type": "image", + "image_path": "b2d1ac4a7282d5b7072db75d882ccf786ccc55ba582646a308f5bcb8bdb1e758.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 270, + 344, + 276 + ], + "lines": [ + { + "bbox": [ + 316, + 270, + 344, + 276 + ], + "spans": [ + { + "bbox": [ + 316, + 270, + 344, + 276 + ], + "type": "text", + "content": "WHAM" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 360, + 270, + 481, + 276 + ], + "lines": [ + { + "bbox": [ + 360, + 270, + 481, + 276 + ], + "spans": [ + { + "bbox": [ + 360, + 270, + 481, + 276 + ], + "type": "text", + "content": "COIN (Human Motion) COIN (Camera Motion)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 279, + 482, + 323 + ], + "lines": [ + { + "bbox": [ + 130, + 279, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 482, + 323 + ], + "type": "text", + "content": "Fig. 3: Qualitative comparisons with state-of-the-art methods. PACE [41] fails to recover a correct trajectory (left). WHAM [81] estimates the wrong walking direction of the person (right). Our approach, COIN, recovers the human and camera motion accurately in both scenarios." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 330, + 481, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 481, + 354 + ], + "type": "text", + "content": "controlled sampling in terms of local body motions. Specifically, COIN with soft inpainting reduces the PA-MPJPE by " + }, + { + "bbox": [ + 130, + 330, + 481, + 354 + ], + "type": "inline_equation", + "content": "4.7 \\mathrm{~mm}" + }, + { + "bbox": [ + 130, + 330, + 481, + 354 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 369, + 220, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 220, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 220, + 380 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "text", + "content": "In this paper, we propose COIN, a diffusion-based optimization framework for global human and camera motion estimation from dynamic cameras. We identify the inconsistency problem of distilling knowledge from the diffusion model with conventional SDS loss. To address this issue, COIN uses a controlled denoiser combined with soft inpainting to distill a high-quality, well-aligned, and consistent motion prior. To further address the scale ambiguity of the camera trajectory, we develop a novel human-scene relation loss that imposes consistency among the human motion, camera motion, and scene features. Comprehensive experiments on challenging synthetic and real-world datasets demonstrate the effectiveness of COIN, which outperforms the SOTA by a large margin in recovering accurate global human motion and camera motion." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Limitations and Future Works: While COIN is able to jointly optimize camera trajectories and global human motions, it requires initialization from SLAM. If the SLAM method fails catastrophically, COIN may not be effective. Additionally, we found COIN fails under severe occlusions where there are several unseen frames and the diffusion model cannot provide consistent guidance. Another limitation is that COIN is an optimization framework, so it is unsuitable for real-time applications. Since the denoising diffusion models have shown their power to model data distributions in many different domains, looking forward we can learn the joint distribution of humans and cameras with the diffusion model to additionally denoise the artifacts in camera motions. Such a joint distribution model also holds the potential of real-time human and camera motion estimation via guided sampling with few DDIM denoising steps." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "L. Jiefeng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 142, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 142, + 140, + 480, + 161 + ], + "type": "text", + "content": "1. Akhter, I., Black, M.J.: Pose-conditioned joint angle limits for 3D human pose reconstruction. In: CVPR (2015) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 142, + 162, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 162, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 142, + 162, + 480, + 183 + ], + "type": "text", + "content": "2. Aksan, E., Kaufmann, M., Hilliges, O.: Structured prediction helps 3d human motion modelling. In: ICCV (2019) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 142, + 184, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 184, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 142, + 184, + 480, + 205 + ], + "type": "text", + "content": "3. Barsoum, E., Kender, J., Liu, Z.: Hp-gan: Probabilistic 3d human motion prediction via gan. In: CVPR Workshops (2018) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 142, + 206, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 206, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 142, + 206, + 480, + 237 + ], + "type": "text", + "content": "4. Bogo, F., Kanazawa, A., Lassner, C., Gehler, P., Romero, J., Black, M.J.: Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In: ECCV (2016) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 239, + 480, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 239, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 142, + 239, + 480, + 260 + ], + "type": "text", + "content": "5. Cao, Z., Gao, H., Mangalam, K., Cai, Q.Z., Vo, M., Malik, J.: Long-term human motion prediction with scene context. In: ECCV. pp. 387-404. Springer (2020) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 142, + 261, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 261, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 142, + 261, + 480, + 281 + ], + "type": "text", + "content": "6. Choi, H., Moon, G., Lee, K.M.: Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In: ECCV (2020) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 142, + 283, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 283, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 142, + 283, + 480, + 304 + ], + "type": "text", + "content": "7. Choi, H., Moon, G., Lee, K.M.: Beyond static features for temporally consistent 3d human pose and shape from a video. In: CVPR (2021) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 304, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 304, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 142, + 304, + 480, + 326 + ], + "type": "text", + "content": "8. Choutas, V., Pavlakos, G., Bolkart, T., Tzionas, D., Black, M.J.: Monocular expressive body regression through body-driven attention. In: ECCV (2020) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 327, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 327, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 142, + 327, + 480, + 347 + ], + "type": "text", + "content": "9. Contributors, M.: Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose (2020) 10" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 348, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 480, + 369 + ], + "type": "text", + "content": "10. Dabral, R., Shimada, S., Jain, A., Theobalt, C., Golyanik, V.: Gravity-aware monocular 3d human-object reconstruction. In: ICCV (2021) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 370, + 480, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 480, + 400 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 480, + 400 + ], + "type": "text", + "content": "11. Fabbri, M., Lanzi, F., Calderara, S., Alletto, S., Cucchiara, R.: Compressed volumetric heatmaps for multi-person 3d pose estimation. In: CVPR (June 2020) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "type": "text", + "content": "12. Fragkiadaki, K., Levine, S., Felsen, P., Malik, J.: Recurrent network models for human dynamics. In: ICCV (2015) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 425, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 480, + 456 + ], + "type": "text", + "content": "13. Gartner, E., Andriluka, M., Xu, H., Sminchisescu, C.: Trajectory optimization for physics-based reconstruction of 3d human pose from monocular video. In: CVPR (2022) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 458, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 458, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 458, + 480, + 479 + ], + "type": "text", + "content": "14. Goel, S., Pavlakos, G., Rajasegaran, J., Kanazawa, A., Malik, J.: Humans in 4d: Reconstructing and tracking humans with transformers. In: ICCV (2023) 12" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 480, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 480, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 480, + 480, + 501 + ], + "type": "text", + "content": "15. Gopalakrishnan, A., Mali, A., Kifer, D., Giles, L., Ororbia, A.G.: A neural temporal model for human motion prediction. In: CVPR (2019) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 502, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 522 + ], + "type": "text", + "content": "16. Guler, R.A., Kokkinos, I.: HoloPose: Holistic 3d human reconstruction in-the-wild. In: CVPR (2019) 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 555 + ], + "type": "text", + "content": "17. Guzov, V., Mir, A., Sattler, T., Pons-Moll, G.: Human POSEitioning system (HPS): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In: CVPR (2021) 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 556, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 480, + 578 + ], + "type": "text", + "content": "18. Harvey, F.G., Yurick, M., Nowrouzezahrai, D., Pal, C.: Robust motion in betweening. ACM Transactions on Graphics (TOG) 39(4), 60-1 (2020) 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "type": "text", + "content": "19. Hassan, M., Ceylan, D., Villegas, R., Saito, J., Yang, J., Zhou, Y., Black, M.J.: Stochastic scene-aware motion prediction. In: ICCV (2021) 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "type": "text", + "content": "20. Hassan, M., Choutas, V., Tzionas, D., Black, M.J.: Resolving 3d human pose ambiguities with 3d scene constraints. In: ICCV. pp. 2282-2292 (2019) 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "text", + "content": "21. He, C., Saito, J., Zachary, J., Rushmeier, H., Zhou, Y.: Nemf: Neural motion fields for kinematic animation. In: NeurIPS (2022) 2, 4, 11" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "22. Henning, D.F., Laidlow, T., Leutenegger, S.: Bodyslam: Joint camera localisation, mapping, and human motion tracking. In: ECCV (2022) 4" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "text", + "content": "COIN: Control-Inpainting Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 137, + 116, + 479, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 479, + 138 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 479, + 138 + ], + "type": "text", + "content": "23. Hernandez, A., Gall, J., Moreno-Noguer, F.: Human motion prediction via spatiotemporal inpainting. In: CVPR (2019) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 137, + 139, + 480, + 159 + ], + "type": "text", + "content": "24. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPs (2020) 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 160, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 160, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 160, + 480, + 194 + ], + "type": "text", + "content": "25. Huang, C.H.P., Yi, H., Höschle, M., Safroshkin, M., Alexiadis, T., Polikovsky, S., Scharstein, D., Black, M.J.: Capturing and inferring dense full-body human-scene contact. In: CVPR (2022) 3, 10" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 194, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 480, + 225 + ], + "type": "text", + "content": "26. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: CVPR (2023) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 138, + 227, + 480, + 248 + ], + "type": "text", + "content": "27. Iqbal, U., Molchanov, P., Kautz, J.: Weakly-supervised 3d human pose learning via multi-view images in the wild. In: CVPR (2020) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 249, + 480, + 270 + ], + "type": "text", + "content": "28. Iqbal, U., Xie, K., Guo, Y., Kautz, J., Molchanov, P.: KAMA: 3D keypoint aware body mesh articulation. In: 3DV (2021) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "text", + "content": "29. Isogawa, M., Yuan, Y., O'Toole, M., Kitani, K.M.: Optical non-line-of-sight physics-based 3d human pose estimation. In: CVPR (2020) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 314 + ], + "type": "text", + "content": "30. Jain, A., Zamir, A.R., Savarese, S., Saxena, A.: Structural-rnn: Deep learning on spatio-temporal graphs. In: CVPR (2016) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 138, + 315, + 480, + 335 + ], + "type": "text", + "content": "31. Jiang, W., Kolotouros, N., Pavlakos, G., Zhou, X., Daniilidis, K.: Coherent reconstruction of multiple humans from a single image. In: CVPR (2020) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 336, + 480, + 357 + ], + "type": "text", + "content": "32. Joo, H., Neverova, N., Vedaldi, A.: Exemplar fine-tuning for 3d human pose fitting towards in-the-wild 3d human pose estimation. In: 3DV (2021) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 138, + 358, + 480, + 379 + ], + "type": "text", + "content": "33. Kanazawa, A., Black, M.J., Jacobs, D.W., Malik, J.: End-to-end recovery of human shape and pose. In: CVPR (2018) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 480, + 402 + ], + "type": "text", + "content": "34. Kanazawa, A., Zhang, J.Y., Felsen, P., Malik, J.: Learning 3d human dynamics from video. In: CVPR (2019) 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "text", + "content": "35. Karunratanakul, K., Preechakul, K., Aksan, E., Beeler, T., Suwajanakorn, S., Tang, S.: Optimizing diffusion noise can serve as universal motion priors. arXiv preprint arXiv:2312.11994 (2023) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 435, + 480, + 456 + ], + "type": "text", + "content": "36. Kaufmann, M., Aksan, E., Song, J., Pece, F., Ziegler, R., Hilliges, O.: Convolutional autoencoders for human motion infilling. In: 3DV (2020) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 457, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 457, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 480, + 489 + ], + "type": "text", + "content": "37. Kaufmann, M., Song, J., Guo, C., Shen, K., Jiang, T., Tang, C., Zárate, J.J., Hilliges, O.: EMDB: The Electromagnetic Database of Global 3D Human Pose and Shape in the Wild. In: ICCV (2023) 3, 11" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 490, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 490, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 490, + 480, + 511 + ], + "type": "text", + "content": "38. Khurana, T., Dave, A., Ramanan, D.: Detecting invisible people. In: ICCV. pp. 3174-3184 (2021) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "type": "text", + "content": "39. Kocabas, M., Athanasiou, N., Black, M.J.: VIBE: Video inference for human body pose and shape estimation. In: CVPR (2020) 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 534, + 480, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 565 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 565 + ], + "type": "text", + "content": "40. Kocabas, M., Huang, C.H.P., Tesch, J., Müller, L., Hilliges, O., Black, M.J.: SPEC: Seeing people in the wild with an estimated camera. In: ICCV (2021) 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 566, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 566, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 566, + 480, + 599 + ], + "type": "text", + "content": "41. Kocabas, M., Yuan, Y., Molchanov, P., Guo, Y., Black, M.J., Hilliges, O., Kautz, J., Iqbal, U.: PACE: Human and motion estimation from in-the-wild videos. In: 3DV (2024) 1, 2, 3, 4, 9, 10, 11, 12, 13, 14" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "type": "text", + "content": "42. Kolotouros, N., Pavlakos, G., Black, M.J., Daniilidis, K.: Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In: ICCV (2019) 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "text", + "content": "43. Kolotouros, N., Pavlakos, G., Daniilidis, K.: Convolutional mesh regression for single-image human shape reconstruction. In: CVPR (2019) 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "text", + "content": "44. Kolotouros, N., Pavlakos, G., Jayaraman, D., Daniilidis, K.: Probabilistic modeling for human mesh recovery. In: ICCV (2021) 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "L. Jiefeng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 148 + ], + "type": "text", + "content": "45. Kundu, J.N., Rakesh, M., Jampani, V., Venkatesh, R.M., Babu1, R.V.: Appearance consensus driven self-supervised human mesh recovery. In: ECCV (2020) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 150, + 480, + 183 + ], + "type": "text", + "content": "46. Lassner, C., Romero, J., Kiefel, M., Bogo, F., Black, M.J., Gehler, P.V.: Unite the people: Closing the loop between 3D and 2D human representations. In: CVPR (2017) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 184, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 480, + 205 + ], + "type": "text", + "content": "47. Li, J., Bian, S., Xu, C., Liu, G., Yu, G., Lu, C.: D &d: Learning human dynamics from dynamic camera. In: ECCV (2022) 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 206, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 480, + 239 + ], + "type": "text", + "content": "48. Li, J., Xu, C., Chen, Z., Bian, S., Yang, L., Lu, C.: Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In: CVPR (2021) 3, 5, 8, 11, 12, 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 240, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 240, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 240, + 480, + 262 + ], + "type": "text", + "content": "49. Li, Z., Liu, J., Zhang, Z., Xu, S., Yan, Y.: Cliff: Carrying location information in full frames into human pose and shape estimation. In: ECCV (2022) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 262, + 480, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 480, + 295 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 480, + 295 + ], + "type": "text", + "content": "50. Li, Z., Zhou, Y., Xiao, S., He, C., Huang, Z., Li, H.: Auto-conditioned recurrent networks for extended complex human motion synthesis. arXiv preprint arXiv:1707.05363 (2017) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 296, + 480, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 296, + 480, + 317 + ], + "spans": [ + { + "bbox": [ + 138, + 296, + 480, + 317 + ], + "type": "text", + "content": "51. Lin, K., Wang, L., Liu, Z.: End-to-end human pose and mesh reconstruction with transformers. In: CVPR (2021) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 319, + 480, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 319, + 480, + 340 + ], + "spans": [ + { + "bbox": [ + 138, + 319, + 480, + 340 + ], + "type": "text", + "content": "52. Liu, M., Yang, D., Zhang, Y., Cui, Z., Rehg, J.M., Tang, S.: 4d human body capture from egocentric video via 3d scene grounding. In: 3DV (2021) 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 341, + 480, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 341, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 138, + 341, + 480, + 372 + ], + "type": "text", + "content": "53. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. SIGGRAPH Asia 34(6), 248:1-248:16 (2015) 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 374, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 480, + 396 + ], + "type": "text", + "content": "54. Luo, Z., Golestaneh, S.A., Kitani, K.M.: 3d human motion estimation via motion compression and refinement. In: ACCV (2020) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 397, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 397, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 397, + 480, + 418 + ], + "type": "text", + "content": "55. Luo, Z., Hachiuma, R., Yuan, Y., Kitani, K.: Dynamics-regulated kinematic policy for egocentric pose estimation. NeurIPS 34 (2021) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 419, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 419, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 138, + 419, + 480, + 441 + ], + "type": "text", + "content": "56. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: AMASS: Archive of motion capture as surface shapes. In: ICCV (2019) 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 442, + 480, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 442, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 138, + 442, + 480, + 474 + ], + "type": "text", + "content": "57. von Marcard, T., Henschel, R., Black, M., Rosenhahn, B., Pons-Moll, G.: Recovering accurate 3d human pose in the wild using imus and a moving camera. In: ECCV (2018) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 475, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 475, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 138, + 475, + 480, + 497 + ], + "type": "text", + "content": "58. Martinez, J., Black, M.J., Romero, J.: On human motion prediction using recurrent neural networks. In: CVPR (2017) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 498, + 480, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 480, + 530 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 480, + 530 + ], + "type": "text", + "content": "59. Mehta, D., Rhodin, H., Casas, D., Fua, P., Sotnychenko, O., Xu, W., Theobalt, C.: Monocular 3d human pose estimation in the wild using improved cnn supervision. In: 3DV (2017) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 531, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 531, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 531, + 480, + 564 + ], + "type": "text", + "content": "60. Mehta, D., Sotnychenko, O., Mueller, F., Xu, W., Elgharib, M., Fua, P., Seidel, H.P., Rhodin, H., Pons-Moll, G., Theobalt, C.: XNect: Real-time multi-person 3D motion capture with a single RGB camera. In: SIGGRAPH (2020) 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 565, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 565, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 565, + 480, + 597 + ], + "type": "text", + "content": "61. Mehta, D., Sridhar, S., Sotnychenko, O., Rhodin, H., Shafiei, M., Seidel, H., Xu, W., Casas, D., Theobalt, C.: VNect: Real-time 3D human pose estimation with a single RGB camera. In: SIGGRAPH (2017) 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 620 + ], + "type": "text", + "content": "62. Moon, G., Chang, J.Y., Lee, K.M.: Camera distance-aware top-down approach for 3d multi-person pose estimation from a single rgb image. In: ICCV (2019) 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "text", + "content": "63. Moon, G., Lee, K.M.: I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In: ECCV (2020) 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "64. Müller, L., Osman, A.A.A., Tang, S., Huang, C.H.P., Black, M.J.: On self contact and human pose. In: CVPR (2021) 3" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "text", + "content": "COIN: Control-Inpainting Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 137, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 481, + 149 + ], + "type": "text", + "content": "65. Müller, L., Ye, V., Pavlakos, G., Black, M., Kanazawa, A.: Generative proxemics: A prior for 3d social interaction from images. arXiv preprint arXiv:2306.09337 (2023) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 149, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 149, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 137, + 149, + 481, + 171 + ], + "type": "text", + "content": "66. Opensfm - a structure from motion library. https://github.com/mapillary/OpenSfM (2021), https://github.com/mapillary/OpenSfM 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 171, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 171, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 137, + 171, + 481, + 203 + ], + "type": "text", + "content": "67. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: CVPR (2019) 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 203, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 203, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 137, + 203, + 481, + 224 + ], + "type": "text", + "content": "68. Pavlakos, G., Kolotouros, N., Daniilidis, K.: Texturepose: Supervising human mesh estimation with texture consistency. In: ICCV (2019) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 224, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 224, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 137, + 224, + 481, + 246 + ], + "type": "text", + "content": "69. Pavlakos, G., Weber, E., Tancik, M., Kanazawa, A.: The one where they reconstructed 3d humans and environments in tv shows. In: ECCV (2022) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 246, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 246, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 137, + 246, + 481, + 267 + ], + "type": "text", + "content": "70. Pavlakos, G., Zhu, L., Zhou, X., Daniilidis, K.: Learning to estimate 3D human pose and shape from a single color image. In: CVPR (2018) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 267, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 267, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 137, + 267, + 481, + 289 + ], + "type": "text", + "content": "71. Pavlo, D., Grangier, D., Auli, M.: Quaternet: A quaternion-based recurrent model for human motion. In: BMVC (2018) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 289, + 481, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 289, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 137, + 289, + 481, + 322 + ], + "type": "text", + "content": "72. Payer, C., Neff, T., Bischof, H., Urschler, M., Stern, D.: Simultaneous multi-person detection and single-person pose estimation with a single heatmap regression network. In: ICCV PoseTrack Workshop (2017) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 322, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 322, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 137, + 322, + 481, + 342 + ], + "type": "text", + "content": "73. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: ICCV (2021) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 137, + 342, + 481, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 342, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 137, + 342, + 481, + 364 + ], + "type": "text", + "content": "74. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. In: ICLR (2023) 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 137, + 364, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 364, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 137, + 364, + 481, + 396 + ], + "type": "text", + "content": "75. Reddy, N.D., Guigues, L., Pischulini, L., Eledath, J., Narasimhan, S.: Tessetrack: End-to-end learnable multi-person articulated 3d pose tracking. In: CVPR (2021) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 137, + 396, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 396, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 137, + 396, + 481, + 418 + ], + "type": "text", + "content": "76. Rempe, D., Birdal, T., Hertzmann, A., Yang, J., Sridhar, S., Guibas, L.J.: Humor: 3d human motion model for robust pose estimation. In: ICCV (2021) 2, 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 137, + 418, + 481, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 418, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 137, + 418, + 481, + 440 + ], + "type": "text", + "content": "77. Rogez, G., Weinzaepfel, P., Schmid, C.: LCR-Net: Localization-classification-regression for human pose. In: CVPR (2017) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 137, + 440, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 440, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 137, + 440, + 481, + 460 + ], + "type": "text", + "content": "78. Rong, Y., Liu, Z., Li, C., Cao, K., Change Loy, C.: Delving deep into hybrid annotations for 3d human recovery in the wild. In: ICCV (2019) 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 137, + 460, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 460, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 137, + 460, + 481, + 493 + ], + "type": "text", + "content": "79. Sárándi, I., Hermans, A., Leibe, B.: Learning 3D human pose estimation from dozens of datasets using a geometry-aware autoencoder to bridge between skeleton formats. In: WACV (2023) 10" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 137, + 493, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 493, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 137, + 493, + 481, + 515 + ], + "type": "text", + "content": "80. Shimada, S., Golyanik, V., Xu, W., Theobalt, C.: Physcap: Physically plausible monocular 3d motion capture in real time. In: SIGGRAPH (2020) 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 137, + 515, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 515, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 137, + 515, + 481, + 536 + ], + "type": "text", + "content": "81. Shin, S., Kim, J., Halilaj, E., Black, M.J.: Wham: Reconstructing world-grounded humans with accurate 3d motion. In: CVPR (2024) 1, 2, 3, 4, 10, 11, 12, 14" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 137, + 536, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 536, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 137, + 536, + 481, + 557 + ], + "type": "text", + "content": "82. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: ICML (2015) 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 137, + 558, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 558, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 137, + 558, + 481, + 578 + ], + "type": "text", + "content": "83. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2021) 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 137, + 578, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 578, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 137, + 578, + 481, + 601 + ], + "type": "text", + "content": "84. Song, J., Chen, X., Hilliges, O.: Human body model fitting by learned gradient descent. In: ECCV (2020) 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 137, + 601, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 601, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 137, + 601, + 481, + 622 + ], + "type": "text", + "content": "85. Sun, Y., Bao, Q., Liu, W., Fu, Y., Black, M.J., Mei, T.: Monocular, one-stage, regression of multiple 3d people. In: ICCV (2021) 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 137, + 622, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 622, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 137, + 622, + 481, + 643 + ], + "type": "text", + "content": "86. Sun, Y., Bao, Q., Liu, W., Mei, T., Black, M.J.: Trace: 5d temporal regression of avatars with dynamic cameras in 3d environments. In: CVPR (2023) 4, 12" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 137, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 643, + 481, + 665 + ], + "type": "text", + "content": "87. Sun, Y., Liu, W., Bao, Q., Fu, Y., Mei, T., Black, M.J.: Putting people in their place: Monocular regression of 3D people in depth. In: CVPR (2022) 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "L. Jiefeng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 138 + ], + "type": "text", + "content": "88. Sun, Y., Ye, Y., Liu, W., Gao, W., Fu, Y., Mei, T.: Human mesh recovery from monocular images via a skeleton-disentangled representation. In: ICCV (2019) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "text", + "content": "89. Teed, Z., Deng, J.: DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. NeurIPs (2021) 2, 5, 11, 12, 13" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 173 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 173 + ], + "type": "text", + "content": "90. Teed, Z., Lipson, L., Deng, J.: Deep patch visual odometry. NeurIPS (2023) 2, 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 174, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 174, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 174, + 480, + 194 + ], + "type": "text", + "content": "91. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. In: ICLR 2023 (2022) 2, 4, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 196, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 196, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 138, + 196, + 480, + 217 + ], + "type": "text", + "content": "92. Villegas, R., Yang, J., Zou, Y., Sohn, S., Lin, X., Lee, H.: Learning to generate long-term future via hierarchical prediction. In: ICML (2017) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 219, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 219, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 138, + 219, + 480, + 239 + ], + "type": "text", + "content": "93. Weng, Z., Yeung, S.: Holistic 3d human and scene mesh estimation from single view images. In: CVPR (2021) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 241, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 241, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 241, + 480, + 262 + ], + "type": "text", + "content": "94. Xiang, D., Joo, H., Sheikh, Y.: Monocular total capture: Posing face, body and hands in the wild. In: CVPR (2019) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 264, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 264, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 138, + 264, + 480, + 285 + ], + "type": "text", + "content": "95. Xie, K., Wang, T., Iqbal, U., Guo, Y., Fidler, S., Shkurti, F.: Physics-based human motion estimation and synthesis from videos. In: ICCV (2021) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 286, + 480, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 286, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 138, + 286, + 480, + 307 + ], + "type": "text", + "content": "96. Xie, Y., Jampani, V., Zhong, L., Sun, D., Jiang, H.: Omnicontrol: Control any joint at any time for human motion generation. In: ICLR (2024) 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 308, + 480, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 308, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 138, + 308, + 480, + 330 + ], + "type": "text", + "content": "97. Xu, Y., Zhu, S.C., Tung, T.: Denserac: Joint 3d pose and shape estimation by dense render-and-compare. In: ICCV (2019) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 331, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 480, + 363 + ], + "type": "text", + "content": "98. Yan, X., Rastogi, A., Villegas, R., Sunkavalli, K., Shechtman, E., Hadap, S., Yumer, E., Lee, H.: Mt-vae: Learning motion transformations to generate multimodal human dynamics. In: ECCV (2018) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 365, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 385 + ], + "type": "text", + "content": "99. Ye, V., Pavlakos, G., Malik, J., Kanazawa, A.: Decoupling human and camera motion from videos in the wild. In: CVPR (2023) 2, 4, 10, 11, 12, 13" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 387, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 387, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 387, + 480, + 418 + ], + "type": "text", + "content": "00. Yuan, Y., Iqbal, U., Molchanov, P., Kitani, K., Kautz, J.: Glamr: Global occlusion-aware human mesh recovery with dynamic cameras. In: CVPR (2022) 2, 4, 10, 11, 12, 13" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 420, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 420, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 138, + 420, + 480, + 441 + ], + "type": "text", + "content": "01. Yuan, Y., Kitani, K.: Diverse trajectory forecasting with determinantal point processes. In: ICLR 2020 (2019) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 443, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 443, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 443, + 480, + 464 + ], + "type": "text", + "content": "02. Yuan, Y., Kitani, K.: Dlow: Diversifying latent flows for diverse human motion prediction. In: ECCV (2020) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 465, + 480, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 465, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 138, + 465, + 480, + 486 + ], + "type": "text", + "content": "03. Yuan, Y., Kitani, K.: Residual force control for agile human behavior imitation and extended motion synthesis. In: NeurIPS (2020) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 487, + 480, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 480, + 509 + ], + "type": "text", + "content": "04. Yuan, Y., Song, J., Iqbal, U., Vahdat, A., Kautz, J.: Physdiff: Physics-guided human motion diffusion model. In: ICCV (2023) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 510, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 480, + 531 + ], + "type": "text", + "content": "05. Yuan, Y., Wei, S.E., Simon, T., Kitani, K., Saragih, J.: Simpoe: Simulated character control for 3d human pose estimation. In: CVPR (2021) 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 533, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 480, + 564 + ], + "type": "text", + "content": "06. Zanfir, A., Bazavan, E.G., Xu, H., Freeman, W.T., Sukthankar, R., Sminchisescu, C.: Weakly supervised 3d human pose and shape reconstruction with normalizing flows. In: ECCV (2020) 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 566, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 566, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 566, + 480, + 597 + ], + "type": "text", + "content": "07. Zanfir, A., Marinoiu, E., Sminchisescu, C.: Monocular 3d pose and shape estimation of multiple people in natural scenes the importance of multiple scene constraints. In: CVPR (2018) 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "type": "text", + "content": "08. Zanfir, A., Marinoiu, E., Zanfir, M., Popa, A.I., Sminchisescu, C.: Deep network for the integrated 3d sensing of multiple people in natural images. In: NeurIPS (2018) 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "09. Zanfir, M., Zanfir, A., Bazavan, E.G., Freeman, W.T., Sukthankar, R., Sminchisscu, C.: Thundr: Transformer-based 3d human reconstruction with markers. In: ICCV (2021) 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "text", + "content": "COIN: Control-Inpainting Diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 355 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "110. Zhang, H., Tian, Y., Zhou, X., Ouyang, W., Liu, Y., Wang, L., Sun, Z.: Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In: ICCV (2021) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 171 + ], + "type": "text", + "content": "111. Zhang, J., Yu, D., Liew, J.H., Nie, X., Feng, J.: Body meshes as points. In: CVPR (2021) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 193 + ], + "type": "text", + "content": "112. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models (2023) 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 481, + 226 + ], + "type": "text", + "content": "113. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "114. Zhang, S., Bhatnagar, B.L., Xu, Y., Winkler, A., Kadlecek, P., Tang, S., Bogo, F.: Rohm: Robust human motion reconstruction via diffusion. In: CVPR (2024) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "text", + "content": "115. Zhang, S., Zhang, Y., Bogo, F., Pollefeys, M., Tang, S.: Learning motion priors for 4d human body capture in 3d scenes. In: ICCV (2021) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "type": "text", + "content": "116. Zhang, T., Huang, B., Wang, Y.: Object-occluded human shape and pose estimation from a single color image. In: CVPR (2020) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "text", + "content": "117. Zhen, J., Fang, Q., Sun, J., Liu, W., Jiang, W., Bao, H., Zhou, X.: SMAP: Single-shot multi-person absolute 3d pose estimation. In: ECCV (2020) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 326, + 481, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 326, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 133, + 326, + 481, + 355 + ], + "type": "text", + "content": "118. Zhou, Y., Habermann, M., Habibie, I., Tewari, A., Theobalt, C., Xu, F.: Monocular real-time full body capture with inter-part correlations. In: CVPR (2021) 3" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "L. Jiefeng et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/COM Kitchens_ An Unedited Overhead-view Procedural Videos Dataset a Vision-Language Benchmark/82c10f54-7bc3-400b-b41c-eee014f604fb_content_list.json b/2024/COM Kitchens_ An Unedited Overhead-view Procedural Videos Dataset a Vision-Language Benchmark/82c10f54-7bc3-400b-b41c-eee014f604fb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..860b677cb03e13278c04108ac42096a21f618ba4 --- /dev/null +++ b/2024/COM Kitchens_ An Unedited Overhead-view Procedural Videos Dataset a Vision-Language Benchmark/82c10f54-7bc3-400b-b41c-eee014f604fb_content_list.json @@ -0,0 +1,1827 @@ +[ + { + "type": "text", + "text": "COM Kitchens: An Unedited Overhead-view Video Dataset as a Vision-Language Benchmark", + "text_level": 1, + "bbox": [ + 225, + 141, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Koki Maeda\\*1,2, Tosho Hirasawa\\*1,3, Atsushi Hashimoto1, Jun Harashima4, Leszek Rybicki4, Yusuke Fukasawa4, and Yoshitaka Ushiku1", + "bbox": [ + 215, + 210, + 787, + 243 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1 OMRON SINIC X Corp., Tokyo, Japan (*equally contributed)", + "2 Tokyo Institute of Technology, Tokyo, Japan", + "3 Tokyo Metropolitan University, Tokyo, Japan,", + "4 Cookpad Inc., Yokohama, Japan" + ], + "bbox": [ + 282, + 253, + 715, + 310 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/406e9df3490a685c380c25a907cb86b5f105393aacd0c2b06230c983494745c7.jpg", + "image_caption": [ + "Fig. 1: Sample from COM Kitchens: the dataset includes unedited overhead-view cooking videos, each manually annotated with a visual action graph that links instructional texts to visual elements through edges from before (□) to after (□) bounding boxes (BBs). dist. BBs (□) represent mixing. Details are provided in Sec. 3.2." + ], + "image_footnote": [], + "bbox": [ + 222, + 330, + 781, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Procedural video understanding is gaining attention in the vision and language community. Deep learning-based video analysis requires extensive data. Consequently, existing works often use web videos as training resources, making it challenging to query instructional contents from raw video observations. To address this issue, we propose a new dataset, COM Kitchens. The dataset consists of unedited overhead-view videos captured by smartphones, in which participants performed food preparation based on given recipes. Fixed-viewpoint video datasets often lack environmental diversity due to high camera setup costs. We used modern wide-angle smartphone lenses to cover cooking counters from sink to cooktop in an overhead view, capturing activity without in-person assistance. With this setup, we collected a diverse dataset by distributing smartphones to participants. With this dataset, we propose the novel video-to-text retrieval task Online Recipe Retrieval (OnRR) and new video captioning domain Dense Video Captioning on unedited Overhead-View videos (DVC-OV). Our experiments verified the capabilities and limitations of current web-video-based SOTA methods in handling these tasks. The dataset and code are available at https://doi.org/10.32130/rdata.6.1 and https://github.com/omron-sinicx/com_kitchens, respectively.", + "bbox": [ + 259, + 561, + 743, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Creating higher-value products from raw materials is fundamental to material wealth in society. Understanding these processes through video observation has gained increased attention from the computer vision community [3, 6, 17, 22, 37, 40]. Recent advancements in procedural video analysis rely on large-scale datasets collected from the Web [30, 57, 58] or through ego-vision [6, 11, 14, 37]. These datasets play a crucial role in developing robust techniques and pre-trained models [1, 9, 24-26, 29]. Fixed-viewpoint observation was a major video format before the deep learning era [22, 43, 46]. However, it has received little attention in the last decade because it is hard to find videos with such a format on the Web. A cold-start problem exists: sufficient data is prerequisite to implement a raw-video-based retrieval system. Once available, the system encourages users to upload unedited videos for querying, reinforcing the dataset size.", + "bbox": [ + 212, + 178, + 787, + 358 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome this cold-start problem, we revisit unedited video datasets with fixed viewpoints (FV). To efficiently enhance environmental diversity, we leveraged recent hardware advancements. Modern smartphones are equipped with well-calibrated wide-angle cameras, allowing us to capture large work areas from an overhead view. This way, we can observe processes with minimal occlusion regardless of kitchen layout, making it preferable for practical applications. Additionally, people are familiar with smartphone UIs, allowing us to collect videos without in-person assistance. To prove this concept of new-style dataset development, we scaled the FV procedural video datasets in this way, resulting in 145 videos, which are 40 hours in total footage, as the first collection.", + "bbox": [ + 212, + 359, + 787, + 510 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Revisiting FV videos in this deep learning era offers challenges in understanding long procedural contexts rather than frame-wise image processing. To tackle this, we provide a new dataset, COM Kitchens, with a manually annotated visual action graph [44], linking visual events and text instructions with a workflow graph (Fig. 1). Using this structured annotation, we introduce a novel video2text retrieval task, online cross-modal recipe retrieval (OnRR), and a new domain for video captioning, dense video captioning on unedited overhead-view videos (DVC-OV). OnRR is an online cross-modal task for retrieving corresponding recipes during cooking, designed to develop practical smartphone applications. DVC-OV is an offline cross-modal task that generates instructional text from demonstrations, intended to analyze the domain gap between web- and overhead-view videos through a traditional format of the DVC task.", + "bbox": [ + 212, + 511, + 787, + 691 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contribution of this paper is four-fold.", + "bbox": [ + 240, + 691, + 547, + 707 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce a novel approach to construct datasets of unedited fixed-view videos in diverse environments, leveraging modern smartphones.", + "2. We provide visual action graph annotations for the first time on unedited videos.", + "3. We propose a novel task of online recipe retrieval, including its target recipe pool and baselines.", + "4. We analyzed the SOTA dense video captioning method on unedited overhead-view videos to reveal future challenges." + ], + "bbox": [ + 220, + 717, + 784, + 838 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "K. Maeda et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 143, + 382, + 160 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This paper proposes a new vision-language video dataset with one novel tasks and one novel domain for DVC. We present comparisons to clarify the dataset novelty in Sec. 2.1, and the task/setting novelties in Sec. 2.2.", + "bbox": [ + 212, + 180, + 782, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Datasets for Procedural Video Understanding", + "text_level": 1, + "bbox": [ + 214, + 250, + 640, + 266 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/14b6a8a64bd0f4ead186e3f2921f2b05a99d1510cf62956f2feb73cb8f571f7d.jpg", + "table_caption": [ + "Table 1: Comparison between instructional video datasets with fixed-viewpoint (FV) cameras. We categorized temporal segment type into action (e.g., \"put a bowl,\" \"crack the egg,\" \"beat the egg\"), and step, a higher-level action (e.g., \"whisk eggs in a bowl\")." + ], + "table_footnote": [], + "table_body": "
datasetyeartopictasks#env.#videos total (h)avg. (m)seg. typeseg. description
MMAC [45]2008Cooking1132815.0action130 actions cls.
MPII [39]2012Cooking14144813.4action65 actions cls.
ACE [43]2012Cooking513523.6action8 actions cls.
50 salads [46]2013Cooking215055.4action51 actions cls.
Breakfast [22]2014Cooking-181,712772.7action10 actions cls.
IKEA ASM [3]2021Furniture45371355.7actionnoun+verb (n+v)
Assembly101 [40]2022Assembly1514,3215137.1act./step1,380 act. cls./n+v
COM KitchensOursCooking139701454016.6act./stepinstructional text
", + "bbox": [ + 220, + 361, + 781, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Tab. 1 summarizes datasets with FV procedural videos. They all target manufacturing tasks. Among these datasets, COM Kitchens has a significantly large diversity in tasks and environments and is the only one with linguistic annotations. We omitted the EgoExo4D [15] dataset from the table because the paper does not provide organized statistics.", + "bbox": [ + 212, + 506, + 784, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Breakfast [22], EgoExo4D [15], and our dataset have environmental diversity, while others struggle with the cost of in-person technical support for their setup. The Breakfast dataset collected data from 18 environments (overhead or side view), but its tasks are limited to two salad recipes with only 10 action classes. EgoExo4D addressed the high cost by collaborating with 12 institutes. Our dataset is competitive with EgoExo4D in terms of the number of environments and participants for the cooking scenario but offers more variety in tasks (i.e., recipes). Another difference is camera views. EgoExo4D used multiple cameras from the front, side, and back view angles, aiming to capture activities beyond the kitchen counter to bridge the gap between egocentric and egocentric views. In contrast, our setup captures detailed food manipulations at the counter with minimal occlusions from the overhead view.", + "bbox": [ + 212, + 582, + 784, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Tab. 2 compares our dataset to other procedural video datasets with linguistic annotations. We omit video datasets without manual annotations, such as HowTo100M [30] and YT-Temporal-1B [55], as they are for pre-training, not for downstream tasks. The pioneering work of YouCookII [57] provides linguistic annotations of coarse instructions (e.g., \"whisk egg, flour\"), and many other works", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "COM Kitchens", + "bbox": [ + 629, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/b0286c6dbe201fca0aee94f548ed629b485ea691239911818b7266112f0615f8.jpg", + "table_caption": [ + "Table 2: Comparison between instructional vision-language video datasets. Only our dataset uses a fixed viewpoint. Coarse instruction minimally describes steps with verb(s) and noun(s) in a YouCookII style [57] (e.g., \"tier dried squid\"), whereas fine instruction comes from real instructional texts (e.g., \"tear dried squid by hand into small pieces\"). 'manual*' annotates only the start of intervals." + ], + "table_footnote": [], + "table_body": "
datasetyeartypetopictasks#videos total (h) avg. (m)seg. descriptioninterval# seg.
YouCookII [57]2018WebCook.892,0001765.3 coarse instruction manual4,325
ProceL [10]2019WebMulti.12720473.9 coarse instruction manual498
COIN [47]2019WebMulti.18011,8274762.4 coarse instruction manual46,354
CrossTask [58]2019WebMulti.834,7003764.8 coarse instruction manual19,278
MMAC-Captions [33]2021EgoCook.51461613.4 coarse instruction manual5,002
Epic Kitchens [7]2022EgoCook.707001008.6 narrationutterance39,596
Ego4D [14]2022EgoOpen--3,670- narrationmanual*-
BioVL2 [38]2022EgoBio.53235.3 fine instruction manual408
VRF [44]2022WebCook.20020020.7 fine instruction manual3,705
FineBio [51]2024EgoBio.722614.53.9 fine instruction manual3,541
COM KitchensOursFVCook.1391454016.6 fine instruction manual2,852
", + "bbox": [ + 220, + 226, + 782, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "follow this manner. Epic Kitchens and Ego4D have narrations as their linguistic annotation, but they tend to describe the details of each action (e.g., action name and target objects), which is still similar to coarse instruction. While captioning for procedural image sequences [4,34,36] uses commercial recipe sites as the dataset resource for generating fine instructions (e.g., \"Beat the egg whites with a mixer, starting on low speed\"), there is a gap with current video datasets' linguistic resource.", + "bbox": [ + 212, + 398, + 782, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "BioVL2 [37], FineBio [51], and VRF [44] are datasets with fine instructions, as this work. BioVL2 is a rare dataset capturing biochemical experiments, but it is limited in size. FineBio is a five times larger than BioVL2 in total footage, but its task variation is still limited and the average footage is shorter than general cooking tasks. VRF collected one-minute videos focused on food state changes, excluding human actions. COM Kitchens has $70\\%$ of the diversity against the web-based VRF dataset, consisting of unedited videos with 16.6 minutes on average.", + "bbox": [ + 212, + 506, + 784, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our dataset provides its annotation as visual action graphs. This structured representation of manufacturing instruction was first proposed in 1980 by Momouchi [31] as an estimation target for a natural language processing task. Later, Flow graph [32,52] was proposed with a fine-grained graph as a comprehensive representation of understanding. Merging tree was the simplest structure of the manufacturing process [18], which selects only actions that merge materials as nodes [18]. Action graph is an intermediate representation between flow graph and merging tree, represents both actions on single materials and merging actions [21].", + "bbox": [ + 212, + 628, + 784, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These structures have been independently extended to vision-language setups. Visual action graphs were defined as a prediction target in an unsupervised task [17] using web videos. Unfortunately, their visual action graph dataset is only for testing and is not publicly available. A small dataset of visual merging trees has been provided in [36] for semi-supervised learning, where visual", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "K. Maeda et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "data consists of image sequences instead of videos. VRF [44] provides visual flow graphs with videos, but the videos are only one minute long. COM Kitchens is the only dataset that annotates graph structures for action segments. Note that the graph represents dynamic changes across frames, including merging and splitting processes. This point is an essential difference from video scene graph datasets [19, 54].", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.2 Video-Text Retrieval and Video Captioning", + "text_level": 1, + "bbox": [ + 214, + 258, + 622, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Retrieval is one of the fundamental tasks for cross-modal problems. For videos, video-text retrieval [5,25-27], a task to find the video whose entire contents fit a text query, is the principal retrieval task. The literature often evaluates the video-to-text scenario in addition to the text-to-video condition; web videos usually have text metadata, making video-to-text evaluation less practical. Instead, the OnRR task assumes retrieving web content by its text from raw video observation. Since our dataset is sourced from smartphones, developed techniques should be directly applicable to smartphone videos.", + "bbox": [ + 212, + 285, + 787, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video paragraph captioning is a video captioning task specially designed for procedural videos; it assumes that event segments are given [23,35,42,50]. However, for unedited videos, it is not practical to assume such given event segments. Hence, we focus on dense video captioning (DVC) [20, 56], a joint task of event detection and event description generation, as a fundamental task of video comprehension. The recent main challenge of DVC is to suppress redundant detection [12]. To overcome this problem, \"detect-then-describe\" [8,41] and \"describe-then-detect\" [49] approaches have been studied. The current SOTA method of Vid2Seq [53] describes captions and detects events in one stage by outputting segments as a time token with its caption. Its performance is supported by the largest video dataset of YT-Temporal-1B [55].", + "bbox": [ + 212, + 406, + 787, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We test our dataset with Vid2Seq to investigate the domain gap between Web and FV videos. The most significant gap is in video length and repetitive actions. Tab. 2 shows that videos in COM Kitchens are about three times longer than usual web videos on average (and 20 times longer than VRF, which are the TikTok style). This difference is mainly due to repetitive actions, often eliminated in web videos. Another gap is the location of event-related objects within each frame. In web videos (or ego vision), the object of interest tends to be in the center of the frame, whereas FV videos do not dynamically focus on objects. These repetitions and lack of focus on important objects provide additional challenges for aligning linguistic instructions to video events.", + "bbox": [ + 212, + 571, + 787, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 The COM Kitchens Dataset", + "text_level": 1, + "bbox": [ + 214, + 747, + 524, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data collection and annotation are pivotal in determining the dataset's utility. This section presents our data collection approach (Sec. 3.1), a detailed definition of the visual action graph (Sec. 3.2), annotation procedure (Sec. 3.3), and dataset statistics (Sec. 3.4).", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "COM Kitchens", + "bbox": [ + 629, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Collecting Videos with their Corresponding Procedural Text", + "text_level": 1, + "bbox": [ + 215, + 146, + 759, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We selected candidate recipes for filming the cooking process from the Cookpad Recipe Dataset (CRD) [16]. The CRD is a comprehensive Japanese home cuisine database, including over 1.7 million recipes with ingredient lists. In the selection process, we prioritized recipes estimated to take under 30 minutes to prepare and to have a moderate complexity level. We excluded recipes using off-counter facilities (e.g., microwaves) because such actions would be recognizable with sound-based event identification, which is not our focus. The selected recipes are typically but not limited to European, Chinese, and Japanese cuisines (with a certain level of localization).", + "bbox": [ + 212, + 172, + 787, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To collect videos, we hired 110 participants from August to October 2021. Each participant recorded up to four recipes at home. Instead of in-person guidance, we provided an instructional document and video that directed the way of recording (e.g., cooking alone, where to capture, camera mode, and turning off TVs) and how to exclude privacy-related information (e.g., do not have a conversation with family, set the camera view not to capture the actor's face, and remove any private documents from camera view). Participants signed a consent form acknowledging that the data would be used publicly for academic purpose. Participants received appropriate financial compensation.", + "bbox": [ + 212, + 308, + 787, + 443 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We utilized the iPhone 11 Pro fixed to a tripod and recorded videos using the rear camera. We instructed participants to set the camera to 30 fps, Full HD resolution of $1920 \\times 1080$ , and ultra-wide field of view mode (equivalent to a focal length of $13~\\mathrm{mm}$ ). Despite the ultra-wide angle setting, there was almost no distortion; therefore, we applied no extra intrinsic calibration. Appendix B provides filmed video examples.", + "bbox": [ + 212, + 444, + 787, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Out of the 410 videos recorded, we excluded any that ignored the instructions, curating a total of 210 videos. See Appendix C for detailed statistics. The most common reason for rejection was wrong observation areas. The second was unintentionally observed faces of actors. We made our best effort to avoid involving private information, considering that participants are not professionals and the environment is personal. This emphasis on privacy made data collection more challenging.", + "bbox": [ + 212, + 535, + 787, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Among the 210 available videos, we have annotated 145 videos, which amount to a total of 40 hours. We will publish the remaining 85 videos as an unsupervised resource. Two videos were recorded at 60fps; however, we retained them in the dataset as they exhibited no quality issues. In the experiment, we used videos scaled to $640 \\times 480$ resolution, but we will release the original resolution together publicly.", + "bbox": [ + 212, + 641, + 787, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Grounding Text to Video via Visual Action Graph", + "text_level": 1, + "bbox": [ + 215, + 753, + 679, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A visual action graph delineates intentional actions termed action-by-person (AP) into a temporal graph format. Here, AP (a.k.a., action-by-chef [32]) refers to human actions on ingredients, like cutting or mixing, as opposed to natural processes, such as the browning of ingredients, which occur without human", + "bbox": [ + 215, + 779, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "K. Maeda et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5586dc17f48bea841bb51260c2f09a58cfd67cd75d7b2d1b4f05be2311d12d00.jpg", + "image_caption": [ + "Fig. 2: A partial view of our visual action graph: AP7 consists of two sub-APs (7-1 and 7-2). All bounding boxes (BBs) mark foods (e.g., the destination BB in AP7-1 is oil heated in AP6-1). The duration is defined by the first and last BBs of the sub-APs." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 782, + 244 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "intervention. We first present the formal definition of the visual action graph, followed by an illustrative example to contextualize its concept.", + "bbox": [ + 212, + 330, + 782, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Define $\\mathcal{V}$ as the set of videos in our dataset, and each video $v\\in \\mathcal{V}$ is a sequence of $|v|$ images, denoted $v = \\{v^{1},v^{2},\\ldots ,v^{|v|}\\}$ . Each video is paired with a corresponding sequence of procedural text, denoted as $t = \\{t^1,t^2,\\dots ,t^{|t|}\\}$ . We define a visual action graph for a pair of video and text as $G(v,t) = (A,E)$ , where $A$ symbolizes the set of APs and $E$ the set of edges, respectively.", + "bbox": [ + 212, + 361, + 784, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Let $a_{k}$ be the $k$ -th AP in the procedural text. We tie $a_{k}$ to visual content for encompassing the time segment in which the AP is performed and bounding boxes showing movements of objects, which can be denoted as", + "bbox": [ + 212, + 436, + 784, + 481 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\na _ {k} = \\left(w _ {k}, A _ {k}, s _ {k}, f _ {k}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 421, + 494, + 578, + 511 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$w_{k}$ is a word sequence of the AP (e.g., $w_{7} =$ \"add the onion and the celery stalk, and\" in Fig. 2). Due to the nature of the videos, $a_{k}$ may involve multiple object instances and multiple action instances. $A_{k}$ represents them as a set of actions in the video (e.g., AP7-1 and AP7-2 for $a_7$ ). We denote the $r$ -th element of $a_{k}$ as $a_{k}^{(r)} = (b_{k}^{(r),\\mathrm{bef}}, b_{k}^{(r),\\mathrm{aft}}, b_{k}^{(r),\\mathrm{dest}})$ (e.g., $a_7^{(1)} = \\mathrm{AP7 - 1}$ ), representing bounding boxes of the target ingredient before/after the action ( $\\square / \\square$ ), and the destination ( $\\square$ ) where the ingredient in $b_{k}^{(r),\\mathrm{bef}}$ is mixed by the action. Note that $b_{k}^{(r),\\mathrm{dest}}$ is not mandatory for $a_{k}^{(r)}$ . $s_k$ and $f_{k}$ denote the start and finish frame of the AP, defined by the first and last bounding box in $A_{k}$ (e.g., $s_7 = 05:39$ and $f_7 = 05:59$ ).", + "bbox": [ + 212, + 522, + 785, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Nodes in the visual action graph are bounding boxes in APs defined above. Edges $E$ track ingredients throughout the cooking activity. Here, $E$ consists of intra-action edges and inter-action edges. Intra-action edges connect $b_{k}^{(r),\\mathrm{bef}}$ to $b_{k}^{(r),\\mathrm{aft}}$ $(\\square \\rightarrow \\square)$ and $b_{k}^{(r),\\mathrm{dest}}$ to $b_{k}^{(r),\\mathrm{aft}}$ $(\\square \\rightarrow \\square)$ . They are always within the same action, and we can automatically identify these edges based on the shared action index, $(k,r)$ , where these indices are labeled instead of object names for this annotation. In contrast, we manually identified inter-action edges, which is from $b_{k}^{(r),\\mathrm{aft}}$ to $b_{k'}^{(r'),\\mathrm{bef}/\\mathrm{dest}}$ $(\\square \\rightarrow \\square$ or $\\square$ ).", + "bbox": [ + 212, + 681, + 785, + 810 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To automate the intra-action edge annotation, we labeled each bounding box with an action index instead of ingredient names. Thus, the graph has", + "bbox": [ + 212, + 810, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "COM Kitchens", + "bbox": [ + 629, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "no ingredient information in this form. To fix this problem, we manually gave ingredient names as leaf nodes and connected them to the AP that first processes the ingredient (ingredient-action edges), following the VRF dataset [44]. This method naturally represents the composition of ingredients at each node by tracing back the edges to the leaf nodes.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/370d6549dd4a83101dbf094861a482b7fda6e1e95d359c5da479dddacdff7911.jpg", + "image_caption": [ + "Fig.3: Distributions of duration; the averages are 16.6 min and 46.7 sec. for videos and APs, respectively." + ], + "image_footnote": [], + "bbox": [ + 217, + 244, + 496, + 325 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/54dc1907671729e52dfc4d6a6bcb40d58a16a58aaaa0fe28a23c6d2b78b95977.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 244, + 785, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Annotation Procedure", + "text_level": 1, + "bbox": [ + 214, + 410, + 447, + 424 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The visual action graph defined above was annotated for each video from February to September 2023. A domain specialist annotated the graph to ensure consistency, with one of the authors reviewing it. The annotator completed the task in about 430 hours.", + "bbox": [ + 212, + 430, + 785, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Initially, the annotator reviewed the videos and revised the procedural instruction texts based on the performance in each video. Despite instructions to follow the recipes faithfully, the participants often deviated due to the complexity of the cooking task. Instead, we used the revised recipes to simulate participants following them. Simultaneously, the annotator identified and tagged the APs within the instructions. All instructions were initially written in Japanese and translated into English by experts. Each AP corresponded directly between the Japanese and English versions.", + "bbox": [ + 212, + 489, + 785, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Subsequently, the annotator delineated the start and finish times with bounding boxes $(b_{k}^{(r),\\mathrm{bef}}, b_{k}^{(r),\\mathrm{aft}}, b_{k}^{(r),\\mathrm{dest}})$ for each action. We used the Computer Vision Annotation Tool (CVAT)5 for this annotation. The video was examined every five frames to identify clear images of the target ingredients.", + "bbox": [ + 212, + 611, + 785, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, the annotator has assigned inter-action edges to the bounding boxes. Following the automatic generation of edges in intra-action edges, they construct visual action graphs. Ingredients and ingredient-action edges have also been annotated together with inter-action edges. A total of 6,826 bounding boxes and 8,061 relationships were annotated to create the visual action graphs.", + "bbox": [ + 212, + 674, + 785, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Statistical Analysis", + "text_level": 1, + "bbox": [ + 214, + 767, + 419, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Recording Statistics. We aim to collect a large-scale dataset of fixed-viewpoint videos featuring structured, fine-grained annotations in diverse environments.", + "bbox": [ + 212, + 787, + 785, + 820 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "K. Maeda et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "5 https://github.com/opencv/cvat", + "bbox": [ + 217, + 823, + 467, + 840 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/650c4993bd5e608152636b2b2e120689009989e80bf38fe4511445930f38c593.jpg", + "image_caption": [ + "Fig. 4: Distribution of the length of word sequences; the averages are 87.2 and 13.3 words for recipes and sentences." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 496, + 223 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/7d97bc20989f403cc6e3e45e94e113e26cdf81ac8021a8027feef619dfea3084.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 147, + 784, + 223 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9f7a3c8852b8f681475fdeee953034b826f676a736bbef12968fc5d9818f501c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 276, + 478, + 415 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4bd878f0e2bd585b6c8b26e755e26c5894b3d0edace3f45fe1ead4e13dcede3b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 276, + 771, + 414 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/caa1e12cee3435d7f03d12f60d7c2b54e63466e6f09ceecd21f29dfc037a16ea.jpg", + "image_caption": [ + "AP" + ], + "image_footnote": [], + "bbox": [ + 233, + 415, + 478, + 536 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/8741711702f95de7b3e213625a4bfcd4cd3af2df2dc1f040c16484400fe2f48a.jpg", + "image_caption": [ + "Ingredients", + "Fig. 5: Word cloud and frequency plots. The word clouds visualize our target of fine-grained procedure comprehension well. There are 131 types of APs and 1,187 types of ingredients, both in a clear long-tail trend." + ], + "image_footnote": [], + "bbox": [ + 526, + 415, + 769, + 537 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The COM Kitchens dataset includes 145 videos from 70 kitchens, totaling 40 hours. Kitchens and actors are in one-to-one correspondence; providing 70 unique actors. The average video duration is 16.6 minutes, as illustrated in Fig. 3 (left). $80.0\\%$ of the video frames are labeled with APs, with a $15.0\\%$ overlap, which is comparable to other datasets6. On average, each AP lasted 46.7 seconds (Fig. 3 right). Tab. 1 and Tab. 2 show comparisons with similar recorded datasets. Our dataset stands out for its diverse environments among fixed-viewpoint video datasets. The average video length is longest among the datasets with linguistic annotations.", + "bbox": [ + 212, + 636, + 787, + 772 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Recipe Data. Turning to the linguistic side, we assigned fine instructions comprising 949 sentences, averaging 6.5 sentences per video. The average word lengths", + "bbox": [ + 214, + 786, + 785, + 816 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "COM Kitchens", + "bbox": [ + 629, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "6 Referred to Tab. 1 of the Assembly101 [40].", + "bbox": [ + 217, + 824, + 524, + 840 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Query: Visual Observation", + "bbox": [ + 248, + 148, + 468, + 162 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/133b21742564d64607bcf05d24d19ff8a39898b1fec1cb5aee0949415c62d88d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 248, + 164, + 748, + 229 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Document: Instructional Text", + "bbox": [ + 251, + 244, + 496, + 258 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/e3c5e815aa7382dbd2b984b2f922d7668ea339030408aff0c233c3fab62d66df.jpg", + "image_caption": [ + "Fig. 6: Data related to OnRR: The query of OnRR sub-tasks is the first $Z\\%$ of a video. For feasible recipe retrieval, we added an extra recipe resource to enhance the dataset of retrieval targets alongside our test set." + ], + "image_footnote": [], + "bbox": [ + 248, + 261, + 488, + 412 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/36eda48772eb29a5f0bde41ae52f2870aaca3bebc08e0fbbae9be73ff2c30185.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 261, + 754, + 412 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "are 87.2 and 13.3 for recipes and sentences (Fig. 4). The instructions included 131 types and 2,286 distinct APs, averaging 15.8 per recipe and 2.4 per sentence. An average of 1.24 repetitions was found in $14\\%$ of the APs. The top of Fig. 5 shows a wordcloud of AP, featuring terms beyond coarse instructions of other datasets in Tab. 2, including fine-grained cooking-specific words like \"Saute.\" The annotated instructions contained 305 types and 1,187 ingredients, averaging 8.2 per recipe. The bottom graphs illustrate a distinct long-tail trend in AP and ingredients.", + "bbox": [ + 212, + 497, + 787, + 619 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 643, + 375, + 662 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Online Recipe Retrieval", + "text_level": 1, + "bbox": [ + 215, + 676, + 460, + 693 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Online Recipe Retrieval (OnRR) involves analyzing cooking videos up to the midpoint to determine the recipe type and the cooking stage achieved. Fig. 6 shows data related to this task; retrieving recipe texts that align with the video content up to a time point $Z$ , using the video segment up to the last started AP $Y$ as the search query. This task is more challenging than traditional video-to-text retrieval as it requires both classification of video type and comprehension of the cooking process. We can decompose OnRR into two subtasks: feasible recipe retrieval and recipe stage identification. We assess these subtasks separately because solving them together is challenging for current SOTA methods.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "K. Maeda et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Feasible recipe retrieval evaluates OnRR from a task categorization perspective in an online setting. We consider a recipe feasible for a task in the middle of the workflow when an actor can shift to the recipe without disposing of foods. This setup assumes an application of online recipe recommendation. Task success is measured by whether the retrieved recipe is in the feasible set.", + "bbox": [ + 218, + 146, + 785, + 220 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Recipe stage identification is evaluated with a given recipe. The task is to spot the exact cooking stage in the recipe from the video. It involves aligning the video content with the corresponding part of the recipe text, up to the verb $w_{y}$ that matches AP $Y$ , within a given recipe ID $X$ . This subtask focuses on accurately matching the video content with the specific steps in the recipe.", + "bbox": [ + 218, + 223, + 785, + 297 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Settings. In our experiment, $Z$ is set at 25%, 50%, 75%, and 100% from the initial point by frame. We labeled them early-, middle-, late-, and full-stage $^{7}$ , respectively. The video set was split into 90/26/29 videos for the train/val/test sets, and the test set was further divided into $29 \\times 4 = 116$ stages. We compared the performance of state-of-the-art models (UniVL [25], CLIP4Clip [26], and X-CLIP [27]) against random selection. We set the maximum token length, maximum frame length, batch size, and number of training epochs to 48, 48, 32, 5 for UniVL and 32, 12, 32, 3 for CLIP4Clip and X-CLIP. These methods often cannot handle all the words in a recipe at once, as it often exceeds 100 words, as shown in Fig. 4. Hence, we shortened the recipe texts by POS tagging with spaCy, extracting only nouns and verbs to cover all steps with the 100 words.", + "bbox": [ + 218, + 313, + 785, + 479 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Extra Recipe Resource. The number of retrieval targets should be sufficiently large; however, the COM Kitchens test set size is limited. Thus, we have increased retrieval targets with an extra recipe resource from CRD. First, we identified the ingredients appearing in the early stage ( $Z = 25\\%$ ) by referring to the action graph for each test set video. Based on the early-stage ingredients, we then extracted up to 100 candidate recipes from CRD for each test recipe. This operation collected 1,828 recipes as this extra resource. Among them, we manually discovered $991/243/19/5$ feasible recipes for each stage $Z$ , tied with their corresponding test set recipe. Nonfeasible recipes remained as hard negatives, often overlapping in ingredients with feasible recipes. Note that extracted recipes only include recipe texts and ingredients, not videos. We machine-translated them from Japanese to English8.", + "bbox": [ + 218, + 496, + 785, + 676 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Metrics. Following the conventional video-text retrieval task, we employed Recall at rank K (R@K, higher is better) and median rank (MdR, lower is better) to evaluate the retrieval performances at the two subtasks.", + "bbox": [ + 218, + 694, + 785, + 739 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results. Tab. 3 showcases the result of the OnRR benchmark in the early- and middle-stage settings. Results for the other two stages and no-fine-tune settings are provided in Appendix D. None of the baseline models outperformed", + "bbox": [ + 218, + 756, + 785, + 801 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "COM Kitchens", + "bbox": [ + 629, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_footnote", + "text": "The full-stage task is equal to conventional video-to-text retrieval.", + "bbox": [ + 220, + 811, + 673, + 824 + ], + "page_idx": 10 + }, + { + "type": "page_footnote", + "text": "8 We used the DeepL API to translate from Japanese to English (US).", + "bbox": [ + 220, + 824, + 689, + 839 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/5558d537a1d770011ff90230689004d0f6dff3833f7486815104f46ac63be02a.jpg", + "table_caption": [ + "Table 3: Online recipe retrieval (OnRR) performances of baseline models. R@K and MdR represent recall at rank K $(\\uparrow)$ and median rank $(\\downarrow)$ , respectively. This table provides only the early- and middel-stage settings (using the first $25\\%$ and $50\\%$ of the video as input); results in other stages and settings are detailed in Appendix D." + ], + "table_footnote": [], + "table_body": "
TaskMethodEarly (25%)Middle (50%)
R@1R@5R@10MdRR@1R@5R@10MdR
Feasible Recipe RetrievalRandom1.88.615.8-0.41.83.1-
UniVL [25]3.45.79.2227.03.45.79.2231.0
CLIP4Clip [26]0.00.010.379.00.00.06.885.0
X-CLIP [27]0.06.810.389.00.03.43.4320.0
Recipe Stage IdentificationRandom6.331.663.38.06.331.663.38.0
UniVL [25]17.248.268.95.09.263.389.23.0
CLIP4Clip [26]6.848.268.95.010.355.186.24.0
X-CLIP [27]10.351.768.94.017.237.993.16.0
", + "bbox": [ + 246, + 212, + 754, + 375 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "random selection in feasible recipe retrieval. Conversely, all models outperformed random selection in recipe stage identification and improved with fine-tuning. These results suggest that simple contrastive learning cannot solve these two tasks simultaneously. In other words, the OnRR task serves as a benchmark for video-text retrieval based on a procedural comprehension.", + "bbox": [ + 212, + 402, + 787, + 479 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Dense Video Captioning on Unedited Fixed-viewpoint Videos", + "text_level": 1, + "bbox": [ + 212, + 500, + 769, + 515 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Dense video captioning is another task for procedural videos, where a system generates multiple detailed captions for different segments within a video. This approach involves detecting distinct events in the video timeline and then generating descriptive and accurate captions for each identified event. The objective is to provide a more comprehensive and segmented understanding of the video content, which is beneficial for offline applications, such as accessibility, content analysis, and enhanced video search capabilities.", + "bbox": [ + 212, + 523, + 787, + 628 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For the COM Kitchens dataset, we use APs as the segments of a video; the content is the recipe text of each AP and the entire duration covering all repetitions under each AP. The main challenge of this task is the domain gap between our unedited overhead-view videos and traditional DVC targets such as web videos.", + "bbox": [ + 212, + 628, + 787, + 704 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Settings. In this experiment, we used the same train-valid/test split with the OnRR task. We selected two DVC systems as our baselines: PDVC [49] and Vid2Seq [53]. We first tested their zero-shot performance on the COM Kitchens dataset. Then, we fine-tuned the Vid2Seq model, the SOTA model on the YouCookII dataset, on our dataset and evaluated its performance.", + "bbox": [ + 212, + 719, + 807, + 794 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Besides, we examined two supervision approaches to leverage the action graph in the DVC task: (i) action graph as relation labels (RL) and (ii) action graph as attention supervision (AS). When we took the action graph as relation labels,", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "K. Maeda et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/83afe14133031ae8d0326448f4929c397f03e9331ba599b8d2a2614424ae7c03.jpg", + "table_caption": [ + "Table 4: Comparison between DVC performances of baseline models. The rows with 'FT' of ' $\\checkmark$ ' show the results of models fine-tuned on COM Kitchens. \"AG\" shows the choice of methods to leverage action graph information during the fine-tuning." + ], + "table_footnote": [], + "table_body": "
ModelFTAGSODA_c(↑)CIDEr(↑)METEOR(↑)
PDVC [49]--0.0220.0000.000
Vid2Seq [53]--0.0170.0660.010
Vid2Seq✓-0.3692.8320.642
Vid2Seq✓RL0.2111.3810.285
Vid2Seq✓AS0.2662.5130.423
Vid2Seq✓RL+AS0.5816.1951.142
", + "bbox": [ + 292, + 198, + 709, + 319 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "we employed the decoder of TablERT-CNN [28] as our module to predict the relation (a.k.a. the type of edges) between APs and trained the model in a multitask learning manner. For attention supervision, following Garg et al. [13], we minimize the Kullback-Leibler divergence between the self-attention matrix at the last encoder layer and the alignment matrix indicating the existence of edges.", + "bbox": [ + 212, + 349, + 782, + 424 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Metrics. We employed SODA_c [12], CIDEr [48], and METEOR [2] scores to evaluate model performance, as these are the commonly used for DVC tasks.", + "bbox": [ + 214, + 441, + 782, + 472 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results. Tab. 4 shows the result of the DVC benchmark. The zero-shot performances of both the PDVC and Vid2Seq models are extremely low compared to those on the YouCookII dataset (4.9 and 7.9 SODA_c scores, respectively). This deterioration indicates that the COM Kitchens dataset is challenging for the DVC task. While speech information is proved to be a crucial modality to generate better captions [53], videos in the COM Kitchens dataset have no speech information than web-based procedural video datasets.", + "bbox": [ + 212, + 489, + 784, + 595 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We also found that the supervision obtained from the action graph benefits the model performance. While RL or AS supervision individually does not bring improvement, their combination enhances model performance across all metrics. This performance gain is attributed to improved attention. As shown in the Fig. 7, RL or AS supervision aligns most frames to a small subset, but the $\\mathrm{RL + AS}$ supervision aligns frames to their relative frames (defined by the action graph) and their surrounding frames. This smoother alignment aids in generating recipe texts and determining segment boundaries.", + "bbox": [ + 212, + 597, + 785, + 717 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Discussion", + "text_level": 1, + "bbox": [ + 215, + 744, + 354, + 758 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We provided baseline results for two specific tasks: OnRR and DVC-OV, extensions of conventional cross-modal retrieval and captioning tasks. OnRR aims to create practical video comprehension for online applications. In contrast, DVC-OV focuses on understanding overhead-view procedural videos as offline tasks.", + "bbox": [ + 212, + 779, + 782, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "COM Kitchens", + "bbox": [ + 629, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1783b55b711f38e236e61d9ed97b44880c2a71591ea5c8a4fd8fef1c1beb07a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 143, + 346, + 255 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/2f67ee35a58b681b9a2f8c867aa30938ecb4160bff10626c30c1875cddb11d31.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 143, + 493, + 255 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5c5550d15dd3f120cec54a20bc20232ac7c40a030f134df8c4a0bc91a4a48a09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 145, + 638, + 255 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/619a0f1d89c22d3503c120ac56200755286a596de31ab0e0556c4d3fc0689d98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 145, + 784, + 255 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/4180d0f75e26967e3b855933c7e5ac3b913e8d4244de67e5b6bf648e4a6ec92a.jpg", + "image_caption": [ + "Fig. 7: Visualization of the attention from the first head in the last encoder layer, based on randomly selected samples. Additional examples can be found in Appendix E." + ], + "image_footnote": [], + "bbox": [ + 218, + 258, + 346, + 357 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/b1b06dd9ff2bb7632b2a09419d34e518bf78212e8c9fa9c94c5f0cc10bd3a84f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 258, + 493, + 357 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/084df8601ec153e213edc4ebd180f5f7984185b5bf3992e644640cfae4bb414f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 258, + 638, + 357 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d8831163f55968550eb695f6f2ed8681ea83255b4a0298f1d16088f3bb6f4910.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 258, + 782, + 357 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "A potential application involves constructing the visual action graph using video and recipe text inputs. It aims to structurally understand cooking procedures by focusing on temporal dependencies and local human-object interactions. Another possible task is episodic memory in the procedural domain, a linguistic query-driven reasoning task tailored originally for long-duration ego-centric videos [14]. The COM Kitchens dataset is particularly suited for this task, as it comprises long videos (avg. 16.6 min., Tab. 1) and the visual action graph can generate various queries with spatio-temporal ground truth.", + "bbox": [ + 212, + 429, + 782, + 549 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our dataset, similar to many others, is limited by its size. Furthermore, benchmark results indicate that conventional global alignment methods are ineffective for these tasks, prompting us to explore new pre-training and fine-tuning methods. Fortunately, our data collection costs are significantly lower compared to datasets like EgoExo4D. We aim to further explore the potential of visual action graphs and to expand the dataset both with and without supervision.", + "bbox": [ + 212, + 550, + 784, + 641 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 667, + 359, + 684 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have introduced COM Kitchens, a dataset that facilitates vision-language understanding with overhead-view recordings, procedural recipe texts, and visual action graphs. This dataset reflects real kitchen conditions, providing rich insights into the sequence of actions and states of ingredients. Our experiments on the OnRR and DVC-OV benchmarks revealed the limitations of existing cross-modal retrieval models in handling long sequences and temporal dependencies. We plan to expand COM Kitchens for further challenges, believing that this dataset will contribute to the advancement of complex video content interpretation.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "K. Maeda et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work would not have been possible without the meticulous and precise annotations by Yasuko Sonoda. We are deeply grateful for her significant contributions to the dataset construction. We also express our gratitude to the IDR office at the National Institute of Informatics for their continuous support in our dataset releases. This work was supported by JSPS KAKENHI Grant Number 21H04910 and JST Moonshot R&D Grant Number JPMJMS2236.", + "bbox": [ + 212, + 180, + 787, + 270 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 297, + 321, + 313 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ashutosh, K., Girdhar, R., Torresani, L., Grauman, K.: HierVL: Learning hierarchical video-language embeddings. In: CVPR. pp. 23066-23078 (June 2023)", + "2. Banerjee, S., Lavie, A.: METEOR: an automatic metric for MT evaluation with improved correlation with human judgments. In: Proc. ACL Workshop IEEMMTS. pp. 65-72 (2005)", + "3. Ben-Shabat, Y., Yu, X., Saleh, F., Campbell, D., Rodriguez-Opazo, C., Li, H., Gould, S.: The IKEA ASM dataset: Understanding people assembling furniture through actions, objects and pose. In: WACV. pp. 847-859 (2021)", + "4. Chandu, K., Nyberg, E., Black, A.W.: Storyboarding of recipes: Grounded contextual generation. In: Annual Meeting of the Association for Computational Linguistics. pp. 6040-6046 (Jul 2019)", + "5. Chen, S., Zhao, Y., Jin, Q., Wu, Q.: Fine-grained video-text retrieval with hierarchical graph reasoning. In: CVPR (June 2020)", + "6. Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., Wray, M.: Scaling egocentric vision: The EPIC-KITCHENS dataset. In: ECCV (2018)", + "7. Damen, D., Doughty, H., Farinella, G.M., Furnari, A., Ma, J., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., Wray, M.: Rescaling egocentric vision: Collection, pipeline and challenges for EPIC-KITCHENS-100. IJCV 130, 33-55 (2022)", + "8. Deng, C., Chen, S., Chen, D., He, Y., Wu, Q.: Sketch, ground, and refine: Top-down dense video captioning. In: CVPR. pp. 234-243 (2021)", + "9. Dvornik, N., Hadji, I., Zhang, R., Derpanis, K.G., Wildes, R.P., Jepson, A.D.: StepFormer: Self-supervised step discovery and localization in instructional videos. In: CVPR. pp. 18952-18961 (June 2023)", + "0. Elhamifar, E., Naing, Z.: Unsupervised procedure learning via joint dynamic summarization. In: ICCV (2019)", + "1. Fathi, A., Li, Y., Rehg, J.M.: Learning to recognize daily actions using gaze. In: ECCV. pp. 314-327 (2012)", + "2. Fujita, S., Hirao, T., Kamigaito, H., Okumura, M., Nagata, M.: SODA: Story oriented dense video captioning evaluation framework. In: ECCV. pp. 517-531 (2020)", + "3. Garg, S., Peitz, S., Nallasamy, U., Paulik, M.: Jointly learning to align and translate with transformer models. In: Conference on Empirical Methods in Natural Language Processing and International Joint Conference on Natural Language Processing. pp. 4453-4462. Hong Kong, China (2019)" + ], + "bbox": [ + 225, + 332, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "COM Kitchens", + "bbox": [ + 629, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., Martin, M., Nagarajan, T., Radosavovic, I., Ramakrishnan, S.K., Ryan, F., Sharma, J., Wray, M., Xu, M., Xu, E.Z., Zhao, C., Bansal, S., Batra, D., Cartillier, V., Crane, S., Do, T., Doulaty, M., Erapalli, A., Feichtenhofer, C., Fragomeni, A., Fu, Q., Gebreselasie, A., Gonzalez, C., Hillis, J., Huang, X., Huang, Y., Jia, W., Khoo, W., Kolář, J., Kottur, S., Kumar, A., Landini, F., Li, C., Li, Y., Li, Z., Mangalam, K., Modhugu, R., Munro, J., Murrell, T., Nishiyasu, T., Price, W., Ruiz, P., Ramazanova, M., Sari, L., Somasundaram, K., Southerland, A., Sugano, Y., Tao, R., Vo, M., Wang, Y., Wu, X., Yagi, T., Zhao, Z., Zhu, Y., Arbeláez, P., Crandall, D., Damen, D., Farinella, G.M., Fuegen, C., Ghanem, B., Ithapu, V.K., Jawahar, C.V., Joo, H., Kitani, K., Li, H., Newcombe, R., Oliva, A., Park, H.S., Rehg, J.M., Sato, Y., Shi, J., Shou, M.Z., Torralba, A., Torresani, L., Yan, M., Malik, J.: Ego4D: Around the world in 3,000 hours of egocentric video. In: CVPR. pp. 18995-19012 (June 2022)", + "15. Grauman, K., Westbury, A., Torresani, L., Kitani, K., Malik, J., Afouras, T., Ashutosh, K., Baiyya, V., Bansal, S., Boote, B., Byrne, E., Chavis, Z., Chen, J., Cheng, F., Chu, F.J., Crane, S., Dasgupta, A., Dong, J., Escobar, M., Forigua, C., Gebreselasie, A., Haresh, S., Huang, J., Islam, M.M., Jain, S., Khirodkar, R., Kukreja, D., Liang, K.J., Liu, J.W., Majumder, S., Mao, Y., Martin, M., Mavroudi, E., Nagarajan, T., Ragusa, F., Ramakrishnan, S.K., Seminara, L., Somayazulu, A., Song, Y., Su, S., Xue, Z., Zhang, E., Zhang, J., Castillo, A., Chen, C., Fu, X., Furuta, R., Gonzalez, C., Gupta, P., Hu, J., Huang, Y., Huang, Y., Khoo, W., Kumar, A., Kuo, R., Lakhavani, S., Liu, M., Luo, M., Luo, Z., Meredith, B., Miller, A., Oguntola, O., Pan, X., Peng, P., Pramanick, S., Ramazanova, M., Ryan, F., Shan, W., Somasundaram, K., Song, C., Southerland, A., Tateno, M., Wang, H., Wang, Y., Yagi, T., Yan, M., Yang, X., Yu, Z., Zha, S.C., Zhao, C., Zhao, Z., Zhu, Z., Zhuo, J., Arbelaez, P., Bertasius, G., Crandall, D., Damen, D., Engel, J., Farinella, G.M., Furnari, A., Ghanem, B., Hoffman, J., Jawahar, C.V., Newcombe, R., Park, H.S., Rehg, J.M., Sato, Y., Savva, M., Shi, J., Shou, M.Z., Wray, M.: Ego-exo4d: Understanding skilled human activity from first- and third-person perspectives (2023)", + "16. Harashima, J., Ariga, M., Murata, K., Ioki, M.: A large-scale recipe and meal data collection as infrastructure for food research. In: International Conference on Language Resources and Evaluation. pp. 2455-2459 (2016)", + "17. Huang, D.A., Lim, J.J., Fei-Fei, L., Niebles, J.C.: Unsupervised visual-linguistic reference resolution in instructional videos. In: CVPR. pp. 2183-2192 (2017)", + "18. Jermsurawong, J., Habash, N.: Predicting the structure of cooking recipes. In: Conference on Empirical Methods in Natural Language Processing. pp. 781-786 (2015)", + "19. Ji, W., Li, Y., Wei, M., Shang, X., Xiao, J., Ren, T., Chua, T.S.: VidVRD 2021: The third grand challenge on video relation detection. In: ACM MM. p. 4779-4783 (2021)", + "20. Johnson, J., Karpathy, A., Fei-Fei, L.: DenseCap: Fully convolutional localization networks for dense captioning. In: CVPR. pp. 4565-4574 (2016)", + "21. Kiddon, C., Ponnuraj, G.T., Zettlemoyer, L., Choi, Y.: Mise en place: Unsupervised interpretation of instructional recipes. In: Conference on Empirical Methods in Natural Language Processing. pp. 982-992 (2015)", + "22. Kuehne, H., Arslan, A., Serre, T.: The language of actions: Recovering the syntax and semantics of goal-directed human activities. In: CVPR. pp. 780-787 (2014)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "K. Maeda et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "23. Lei, J., Wang, L., Shen, Y., Yu, D., Berg, T., Bansal, M.: MART: memory-augmented recurrent transformer for coherent video paragraph captioning. In: Annual Meeting of the Association for Computational Linguistics. pp. 2603-2614 (2020)", + "24. Lin, K.Q., Wang, J., Soldan, M., Wray, M., Yan, R., Xu, Z., Gao, D., Tu, R.C., Zhao, W., Kong, W., Cai, C., HongFa, W., Damen, D., Ghanem, B., Liu, W., Shou, M.Z.: Egocentric video-language pretraining. In: NeurIPS (2022)", + "25. Luo, H., Ji, L., Shi, B., Huang, H., Duan, N., Li, T., Li, J., Bharti, T., Zhou, M.: UniVL: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353 (2020)", + "26. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: CLIP4Clip: An empirical study of CLIP for end to end video clip retrieval. Neurocomputing 508, 293-304 (2022)", + "27. Ma, Y., Xu, G., Sun, X., Yan, M., Zhang, J., Ji, R.: X-CLIP: End-to-end multi-grained contrastive learning for video-text retrieval. In: ACM MM. pp. 638-647 (2022)", + "28. Ma, Y., Hiraoka, T., Okazaki, N.: Joint entity and relation extraction based on table labeling using convolutional neural networks. In: Workshop on Structured Prediction for NLP. pp. 11-21 (2022)", + "29. Miech, A., Alayrac, J.B., Smaira, L., Laptev, I., Sivic, J., Zisserman, A.: End-to-end learning of visual Representations from uncurated instructional videos. In: CVPR (2020)", + "30. Miech, A., Zhukov, D., Alayrac, J.B., Tapaswi, M., Laptev, I., Sivic, J.: HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In: ICCV (2019)", + "31. Momouchi, Y.: Control structures for actions in procedural texts and PT-chart. In: International Conference on Computational Linguistics. pp. 108-114 (1980)", + "32. Mori, S., Maeta, H., Yamakata, Y., Sasada, T.: Flow graph corpus from recipe texts. In: International Conference on Language Resources and Evaluation. pp. 2370-2377 (2014)", + "33. Nakamura, K., Ohashi, H., Okada, M.: Sensor-augmented egocentric-video captioning with dynamic modal attention. In: ACM MM. p. 4220-4229 (2021)", + "34. Nishimura, T., Hashimoto, A., Mori, S.: Procedural text generation from a photo sequence. In: International Natural Language Generation Conference. pp. 409-414 (2019)", + "35. Nishimura, T., Hashimoto, A., Ushiku, Y., Kameko, H., Mori, S.: State-aware video procedural captioning. In: ACM MM. pp. 1766-1774 (2021)", + "36. Nishimura, T., Hashimoto, A., Ushiku, Y., Kameko, H., Yamakata, Y., Mori, S.: Structure-aware procedural text generation from an image sequence. IEEE Access 9, 2125-2141 (2020)", + "37. Nishimura, T., Sakoda, K., Hashimoto, A., Ushiku, Y., Tanaka, N., Ono, F., Kameko, H., Mori, S.: Egocentric biochemical video-and-language dataset. In: ICCV Workshop. pp. 3129-3133 (October 2021)", + "38. Nishimura, T., Sakoda, K., Ushiku, A., Hashimoto, A., Okuda, N., Ono, F., Kameko, H., Mori, S.: BioVL2 dataset: Egocentric biochemical video-and-language dataset. Journal of Natural Language Processing (2022)", + "39. Schiele, B.: A database for fine grained activity detection of cooking activities. In: CVPR. p. 1194-1201 (2012)", + "40. Sener, F., Chatterjee, D., Shelepov, D., He, K., Singhania, D., Wang, R., Yao, A.: Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In: CVPR. pp. 21064-21074 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "COM Kitchens", + "bbox": [ + 629, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "41. Shi, B., Ji, L., Liang, Y., Duan, N., Chen, P., Niu, Z., Zhou, M.: Dense procedure captioning in narrated instructional videos. In: Annual Meeting of the Association for Computational Linguistics. pp. 6382-6391 (2019)", + "42. Shi, B., Ji, L., Niu, Z., Duan, N., Zhou, M., Chen, X.: Learning semantic concepts and temporal alignment for narrated video procedural captioning. In: ACM MM. pp. 4355-4363 (2020)", + "43. Shimada, A., Kondo, K., Deguchi, D., Morin, G., Stern, H.: Kitchen scene context based gesture recognition: A contest in ICPR2012. In: International Workshop on Advances in Depth Image Analysis and Applications. pp. 168-185 (2013). https://doi.org/10.1007/978-3-642-40303-3_18", + "44. Shirai, K., Hashimoto, A., Nishimura, T., Kameko, H., Kurita, S., Ushiku, Y., Mori, S.: Visual Recipe Flow: A dataset for learning visual state changes of objects with recipe flows. In: International Conference on Computational Linguistics. pp. 3570-3577 (2022)", + "45. Spriggs, E.H., De La Torre, F., Hebert, M.: Temporal segmentation and activity classification from first-person sensing. In: CVPRW. pp. 17-24 (2009)", + "46. Stein, S., McKenna, S.J.: Combining embedded accelerometers with computer vision for recognizing food preparation activities. In: ACM international joint conference on Pervasive and ubiquitous computing. p. 729-738 (2013)", + "47. Tang, Y., Ding, D., Rao, Y., Zheng, Y., Zhang, D., Zhao, L., Lu, J., Zhou, J.: COIN: A large-scale dataset for comprehensive instructional video analysis. In: CVPR (2019)", + "48. Vedantam, R., Zitnick, C.L., Parikh, D.: CIDEr: consensus-based image description evaluation. In: CVPR. pp. 4566-4575 (2015)", + "49. Wang, T., Zhang, R., Lu, Z., Zheng, F., Cheng, R., Luo, P.: End-to-end dense video captioning with parallel decoding. In: ICCV. pp. 6847-6857 (2021)", + "50. Wu, J., Pan, L., Chen, J., Jiang, Y.G.: Ingredient-enriched recipe generation from cooking videos. In: ACM International Conference on Multimedia Retrieval. pp. 249-257 (2022)", + "51. Yagi, T., Ohashi, M., Huang, Y., Furuta, R., Adachi, S., Mitsuyama, T., Sato, Y.: FineBio: A fine-grained video dataset of biological experiments with hierarchical annotation. arXiv preprint arXiv:2402.00293 (2024)", + "52. Yamakata, Y., Mori, S., Carroll, J.: English recipe flow graph corpus. In: International Conference on Language Resources and Evaluation. pp. 5187-5194 (2020)", + "53. Yang, A., Nagrani, A., Seo, P.H., Miech, A., Pont-Tuset, J., Laptev, I., Sivic, J., Schmid, C.: Vid2Seq: Large-scale pretraining of a visual language model for dense video captioning. In: CVPR (2023)", + "54. Yang, J., Peng, W., Li, X., Guo, Z., Chen, L., Li, B., Ma, Z., Zhou, K., Zhang, W., Loy, C.C., Liu, Z.: Panoptic video scene graph generation. In: CVPR. pp. 18675-18685 (2023)", + "55. Zellers, R., Lu, J., Lu, X., Yu, Y., Zhao, Y., Salehi, M., Kusupati, A., Hessel, J., Farhadi, A., Choi, Y.: MERLOT Reserve: Neural script knowledge through vision and language and sound. In: CVPR. pp. 16375-16387 (June 2022)", + "56. Zhou, L., Kalantidis, Y., Chen, X., Corso, J.J., Rohrbach, M.: Grounded video description. In: CVPR. pp. 6578-6587 (2019)", + "57. Zhou, L., Xu, C., Corso, J.J.: Towards automatic learning of procedures from web instructional videos. In: AAAI (2018)", + "58. Zhukov, D., Alayrac, J.B., Cinbis, R.G., Fouhey, D., Laptev, I., Sivic, J.: Cross-task weakly supervised learning from instructional videos. In: CVPR. pp. 3537-3545 (2019)" + ], + "bbox": [ + 215, + 146, + 785, + 830 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "K. Maeda et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/COM Kitchens_ An Unedited Overhead-view Procedural Videos Dataset a Vision-Language Benchmark/82c10f54-7bc3-400b-b41c-eee014f604fb_model.json b/2024/COM Kitchens_ An Unedited Overhead-view Procedural Videos Dataset a Vision-Language Benchmark/82c10f54-7bc3-400b-b41c-eee014f604fb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7c6f06871aae30e7005d7fe70a46a09edb80b19e --- /dev/null +++ b/2024/COM Kitchens_ An Unedited Overhead-view Procedural Videos Dataset a Vision-Language Benchmark/82c10f54-7bc3-400b-b41c-eee014f604fb_model.json @@ -0,0 +1,2568 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.226, + 0.142, + 0.78, + 0.187 + ], + "angle": 0, + "content": "COM Kitchens: An Unedited Overhead-view Video Dataset as a Vision-Language Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.212, + 0.788, + 0.244 + ], + "angle": 0, + "content": "Koki Maeda\\*1,2, Tosho Hirasawa\\*1,3, Atsushi Hashimoto1, Jun Harashima4, Leszek Rybicki4, Yusuke Fukasawa4, and Yoshitaka Ushiku1" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.255, + 0.717, + 0.27 + ], + "angle": 0, + "content": "1 OMRON SINIC X Corp., Tokyo, Japan (*equally contributed)" + }, + { + "type": "text", + "bbox": [ + 0.344, + 0.27, + 0.658, + 0.283 + ], + "angle": 0, + "content": "2 Tokyo Institute of Technology, Tokyo, Japan" + }, + { + "type": "text", + "bbox": [ + 0.341, + 0.283, + 0.662, + 0.298 + ], + "angle": 0, + "content": "3 Tokyo Metropolitan University, Tokyo, Japan," + }, + { + "type": "text", + "bbox": [ + 0.385, + 0.298, + 0.617, + 0.311 + ], + "angle": 0, + "content": "4 Cookpad Inc., Yokohama, Japan" + }, + { + "type": "list", + "bbox": [ + 0.284, + 0.255, + 0.717, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.332, + 0.782, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.485, + 0.789, + 0.542 + ], + "angle": 0, + "content": "Fig. 1: Sample from COM Kitchens: the dataset includes unedited overhead-view cooking videos, each manually annotated with a visual action graph that links instructional texts to visual elements through edges from before (□) to after (□) bounding boxes (BBs). dist. BBs (□) represent mixing. Details are provided in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.563, + 0.744, + 0.842 + ], + "angle": 0, + "content": "Abstract. Procedural video understanding is gaining attention in the vision and language community. Deep learning-based video analysis requires extensive data. Consequently, existing works often use web videos as training resources, making it challenging to query instructional contents from raw video observations. To address this issue, we propose a new dataset, COM Kitchens. The dataset consists of unedited overhead-view videos captured by smartphones, in which participants performed food preparation based on given recipes. Fixed-viewpoint video datasets often lack environmental diversity due to high camera setup costs. We used modern wide-angle smartphone lenses to cover cooking counters from sink to cooktop in an overhead view, capturing activity without in-person assistance. With this setup, we collected a diverse dataset by distributing smartphones to participants. With this dataset, we propose the novel video-to-text retrieval task Online Recipe Retrieval (OnRR) and new video captioning domain Dense Video Captioning on unedited Overhead-View videos (DVC-OV). Our experiments verified the capabilities and limitations of current web-video-based SOTA methods in handling these tasks. The dataset and code are available at https://doi.org/10.32130/rdata.6.1 and https://github.com/omron-sinicx/com_kitchens, respectively." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "K. Maeda et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.376, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.359 + ], + "angle": 0, + "content": "Creating higher-value products from raw materials is fundamental to material wealth in society. Understanding these processes through video observation has gained increased attention from the computer vision community [3, 6, 17, 22, 37, 40]. Recent advancements in procedural video analysis rely on large-scale datasets collected from the Web [30, 57, 58] or through ego-vision [6, 11, 14, 37]. These datasets play a crucial role in developing robust techniques and pre-trained models [1, 9, 24-26, 29]. Fixed-viewpoint observation was a major video format before the deep learning era [22, 43, 46]. However, it has received little attention in the last decade because it is hard to find videos with such a format on the Web. A cold-start problem exists: sufficient data is prerequisite to implement a raw-video-based retrieval system. Once available, the system encourages users to upload unedited videos for querying, reinforcing the dataset size." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.361, + 0.788, + 0.511 + ], + "angle": 0, + "content": "To overcome this cold-start problem, we revisit unedited video datasets with fixed viewpoints (FV). To efficiently enhance environmental diversity, we leveraged recent hardware advancements. Modern smartphones are equipped with well-calibrated wide-angle cameras, allowing us to capture large work areas from an overhead view. This way, we can observe processes with minimal occlusion regardless of kitchen layout, making it preferable for practical applications. Additionally, people are familiar with smartphone UIs, allowing us to collect videos without in-person assistance. To prove this concept of new-style dataset development, we scaled the FV procedural video datasets in this way, resulting in 145 videos, which are 40 hours in total footage, as the first collection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.788, + 0.692 + ], + "angle": 0, + "content": "Revisiting FV videos in this deep learning era offers challenges in understanding long procedural contexts rather than frame-wise image processing. To tackle this, we provide a new dataset, COM Kitchens, with a manually annotated visual action graph [44], linking visual events and text instructions with a workflow graph (Fig. 1). Using this structured annotation, we introduce a novel video2text retrieval task, online cross-modal recipe retrieval (OnRR), and a new domain for video captioning, dense video captioning on unedited overhead-view videos (DVC-OV). OnRR is an online cross-modal task for retrieving corresponding recipes during cooking, designed to develop practical smartphone applications. DVC-OV is an offline cross-modal task that generates instructional text from demonstrations, intended to analyze the domain gap between web- and overhead-view videos through a traditional format of the DVC task." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.693, + 0.549, + 0.708 + ], + "angle": 0, + "content": "The contribution of this paper is four-fold." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.718, + 0.784, + 0.747 + ], + "angle": 0, + "content": "1. We introduce a novel approach to construct datasets of unedited fixed-view videos in diverse environments, leveraging modern smartphones." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.749, + 0.784, + 0.777 + ], + "angle": 0, + "content": "2. We provide visual action graph annotations for the first time on unedited videos." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.779, + 0.784, + 0.808 + ], + "angle": 0, + "content": "3. We propose a novel task of online recipe retrieval, including its target recipe pool and baselines." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.81, + 0.785, + 0.839 + ], + "angle": 0, + "content": "4. We analyzed the SOTA dense video captioning method on unedited overhead-view videos to reveal future challenges." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.718, + 0.785, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.63, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "COM Kitchens" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.383, + 0.161 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.784, + 0.226 + ], + "angle": 0, + "content": "This paper proposes a new vision-language video dataset with one novel tasks and one novel domain for DVC. We present comparisons to clarify the dataset novelty in Sec. 2.1, and the task/setting novelties in Sec. 2.2." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.251, + 0.642, + 0.267 + ], + "angle": 0, + "content": "2.1 Datasets for Procedural Video Understanding" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.307, + 0.785, + 0.35 + ], + "angle": 0, + "content": "Table 1: Comparison between instructional video datasets with fixed-viewpoint (FV) cameras. We categorized temporal segment type into action (e.g., \"put a bowl,\" \"crack the egg,\" \"beat the egg\"), and step, a higher-level action (e.g., \"whisk eggs in a bowl\")." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.362, + 0.782, + 0.476 + ], + "angle": 0, + "content": "
datasetyeartopictasks#env.#videos total (h)avg. (m)seg. typeseg. description
MMAC [45]2008Cooking1132815.0action130 actions cls.
MPII [39]2012Cooking14144813.4action65 actions cls.
ACE [43]2012Cooking513523.6action8 actions cls.
50 salads [46]2013Cooking215055.4action51 actions cls.
Breakfast [22]2014Cooking-181,712772.7action10 actions cls.
IKEA ASM [3]2021Furniture45371355.7actionnoun+verb (n+v)
Assembly101 [40]2022Assembly1514,3215137.1act./step1,380 act. cls./n+v
COM KitchensOursCooking139701454016.6act./stepinstructional text
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.785, + 0.581 + ], + "angle": 0, + "content": "Tab. 1 summarizes datasets with FV procedural videos. They all target manufacturing tasks. Among these datasets, COM Kitchens has a significantly large diversity in tasks and environments and is the only one with linguistic annotations. We omitted the EgoExo4D [15] dataset from the table because the paper does not provide organized statistics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.785, + 0.763 + ], + "angle": 0, + "content": "Breakfast [22], EgoExo4D [15], and our dataset have environmental diversity, while others struggle with the cost of in-person technical support for their setup. The Breakfast dataset collected data from 18 environments (overhead or side view), but its tasks are limited to two salad recipes with only 10 action classes. EgoExo4D addressed the high cost by collaborating with 12 institutes. Our dataset is competitive with EgoExo4D in terms of the number of environments and participants for the cooking scenario but offers more variety in tasks (i.e., recipes). Another difference is camera views. EgoExo4D used multiple cameras from the front, side, and back view angles, aiming to capture activities beyond the kitchen counter to bridge the gap between egocentric and egocentric views. In contrast, our setup captures detailed food manipulations at the counter with minimal occlusions from the overhead view." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Tab. 2 compares our dataset to other procedural video datasets with linguistic annotations. We omit video datasets without manual annotations, such as HowTo100M [30] and YT-Temporal-1B [55], as they are for pre-training, not for downstream tasks. The pioneering work of YouCookII [57] provides linguistic annotations of coarse instructions (e.g., \"whisk egg, flour\"), and many other works" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "K. Maeda et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 2: Comparison between instructional vision-language video datasets. Only our dataset uses a fixed viewpoint. Coarse instruction minimally describes steps with verb(s) and noun(s) in a YouCookII style [57] (e.g., \"tier dried squid\"), whereas fine instruction comes from real instructional texts (e.g., \"tear dried squid by hand into small pieces\"). 'manual*' annotates only the start of intervals." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.227, + 0.784, + 0.37 + ], + "angle": 0, + "content": "
datasetyeartypetopictasks#videos total (h) avg. (m)seg. descriptioninterval# seg.
YouCookII [57]2018WebCook.892,0001765.3 coarse instruction manual4,325
ProceL [10]2019WebMulti.12720473.9 coarse instruction manual498
COIN [47]2019WebMulti.18011,8274762.4 coarse instruction manual46,354
CrossTask [58]2019WebMulti.834,7003764.8 coarse instruction manual19,278
MMAC-Captions [33]2021EgoCook.51461613.4 coarse instruction manual5,002
Epic Kitchens [7]2022EgoCook.707001008.6 narrationutterance39,596
Ego4D [14]2022EgoOpen--3,670- narrationmanual*-
BioVL2 [38]2022EgoBio.53235.3 fine instruction manual408
VRF [44]2022WebCook.20020020.7 fine instruction manual3,705
FineBio [51]2024EgoBio.722614.53.9 fine instruction manual3,541
COM KitchensOursFVCook.1391454016.6 fine instruction manual2,852
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.784, + 0.505 + ], + "angle": 0, + "content": "follow this manner. Epic Kitchens and Ego4D have narrations as their linguistic annotation, but they tend to describe the details of each action (e.g., action name and target objects), which is still similar to coarse instruction. While captioning for procedural image sequences [4,34,36] uses commercial recipe sites as the dataset resource for generating fine instructions (e.g., \"Beat the egg whites with a mixer, starting on low speed\"), there is a gap with current video datasets' linguistic resource." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.785, + 0.627 + ], + "angle": 0, + "content": "BioVL2 [37], FineBio [51], and VRF [44] are datasets with fine instructions, as this work. BioVL2 is a rare dataset capturing biochemical experiments, but it is limited in size. FineBio is a five times larger than BioVL2 in total footage, but its task variation is still limited and the average footage is shorter than general cooking tasks. VRF collected one-minute videos focused on food state changes, excluding human actions. COM Kitchens has \\(70\\%\\) of the diversity against the web-based VRF dataset, consisting of unedited videos with 16.6 minutes on average." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.785, + 0.764 + ], + "angle": 0, + "content": "Our dataset provides its annotation as visual action graphs. This structured representation of manufacturing instruction was first proposed in 1980 by Momouchi [31] as an estimation target for a natural language processing task. Later, Flow graph [32,52] was proposed with a fine-grained graph as a comprehensive representation of understanding. Merging tree was the simplest structure of the manufacturing process [18], which selects only actions that merge materials as nodes [18]. Action graph is an intermediate representation between flow graph and merging tree, represents both actions on single materials and merging actions [21]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "These structures have been independently extended to vision-language setups. Visual action graphs were defined as a prediction target in an unsupervised task [17] using web videos. Unfortunately, their visual action graph dataset is only for testing and is not publicly available. A small dataset of visual merging trees has been provided in [36] for semi-supervised learning, where visual" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.63, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "COM Kitchens" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "data consists of image sequences instead of videos. VRF [44] provides visual flow graphs with videos, but the videos are only one minute long. COM Kitchens is the only dataset that annotates graph structures for action segments. Note that the graph represents dynamic changes across frames, including merging and splitting processes. This point is an essential difference from video scene graph datasets [19, 54]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.26, + 0.624, + 0.277 + ], + "angle": 0, + "content": "2.2 Video-Text Retrieval and Video Captioning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.407 + ], + "angle": 0, + "content": "Retrieval is one of the fundamental tasks for cross-modal problems. For videos, video-text retrieval [5,25-27], a task to find the video whose entire contents fit a text query, is the principal retrieval task. The literature often evaluates the video-to-text scenario in addition to the text-to-video condition; web videos usually have text metadata, making video-to-text evaluation less practical. Instead, the OnRR task assumes retrieving web content by its text from raw video observation. Since our dataset is sourced from smartphones, developed techniques should be directly applicable to smartphone videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.788, + 0.573 + ], + "angle": 0, + "content": "Video paragraph captioning is a video captioning task specially designed for procedural videos; it assumes that event segments are given [23,35,42,50]. However, for unedited videos, it is not practical to assume such given event segments. Hence, we focus on dense video captioning (DVC) [20, 56], a joint task of event detection and event description generation, as a fundamental task of video comprehension. The recent main challenge of DVC is to suppress redundant detection [12]. To overcome this problem, \"detect-then-describe\" [8,41] and \"describe-then-detect\" [49] approaches have been studied. The current SOTA method of Vid2Seq [53] describes captions and detects events in one stage by outputting segments as a time token with its caption. Its performance is supported by the largest video dataset of YT-Temporal-1B [55]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.573, + 0.788, + 0.725 + ], + "angle": 0, + "content": "We test our dataset with Vid2Seq to investigate the domain gap between Web and FV videos. The most significant gap is in video length and repetitive actions. Tab. 2 shows that videos in COM Kitchens are about three times longer than usual web videos on average (and 20 times longer than VRF, which are the TikTok style). This difference is mainly due to repetitive actions, often eliminated in web videos. Another gap is the location of event-related objects within each frame. In web videos (or ego vision), the object of interest tends to be in the center of the frame, whereas FV videos do not dynamically focus on objects. These repetitions and lack of focus on important objects provide additional challenges for aligning linguistic instructions to video events." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.748, + 0.525, + 0.765 + ], + "angle": 0, + "content": "3 The COM Kitchens Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Data collection and annotation are pivotal in determining the dataset's utility. This section presents our data collection approach (Sec. 3.1), a detailed definition of the visual action graph (Sec. 3.2), annotation procedure (Sec. 3.3), and dataset statistics (Sec. 3.4)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "K. Maeda et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.761, + 0.162 + ], + "angle": 0, + "content": "3.1 Collecting Videos with their Corresponding Procedural Text" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.308 + ], + "angle": 0, + "content": "We selected candidate recipes for filming the cooking process from the Cookpad Recipe Dataset (CRD) [16]. The CRD is a comprehensive Japanese home cuisine database, including over 1.7 million recipes with ingredient lists. In the selection process, we prioritized recipes estimated to take under 30 minutes to prepare and to have a moderate complexity level. We excluded recipes using off-counter facilities (e.g., microwaves) because such actions would be recognizable with sound-based event identification, which is not our focus. The selected recipes are typically but not limited to European, Chinese, and Japanese cuisines (with a certain level of localization)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.309, + 0.788, + 0.444 + ], + "angle": 0, + "content": "To collect videos, we hired 110 participants from August to October 2021. Each participant recorded up to four recipes at home. Instead of in-person guidance, we provided an instructional document and video that directed the way of recording (e.g., cooking alone, where to capture, camera mode, and turning off TVs) and how to exclude privacy-related information (e.g., do not have a conversation with family, set the camera view not to capture the actor's face, and remove any private documents from camera view). Participants signed a consent form acknowledging that the data would be used publicly for academic purpose. Participants received appropriate financial compensation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.535 + ], + "angle": 0, + "content": "We utilized the iPhone 11 Pro fixed to a tripod and recorded videos using the rear camera. We instructed participants to set the camera to 30 fps, Full HD resolution of \\(1920 \\times 1080\\), and ultra-wide field of view mode (equivalent to a focal length of \\(13~\\mathrm{mm}\\)). Despite the ultra-wide angle setting, there was almost no distortion; therefore, we applied no extra intrinsic calibration. Appendix B provides filmed video examples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.641 + ], + "angle": 0, + "content": "Out of the 410 videos recorded, we excluded any that ignored the instructions, curating a total of 210 videos. See Appendix C for detailed statistics. The most common reason for rejection was wrong observation areas. The second was unintentionally observed faces of actors. We made our best effort to avoid involving private information, considering that participants are not professionals and the environment is personal. This emphasis on privacy made data collection more challenging." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.788, + 0.732 + ], + "angle": 0, + "content": "Among the 210 available videos, we have annotated 145 videos, which amount to a total of 40 hours. We will publish the remaining 85 videos as an unsupervised resource. Two videos were recorded at 60fps; however, we retained them in the dataset as they exhibited no quality issues. In the experiment, we used videos scaled to \\(640 \\times 480\\) resolution, but we will release the original resolution together publicly." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.755, + 0.681, + 0.77 + ], + "angle": 0, + "content": "3.2 Grounding Text to Video via Visual Action Graph" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "A visual action graph delineates intentional actions termed action-by-person (AP) into a temporal graph format. Here, AP (a.k.a., action-by-chef [32]) refers to human actions on ingredients, like cutting or mixing, as opposed to natural processes, such as the browning of ingredients, which occur without human" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.63, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "COM Kitchens" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.784, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.258, + 0.784, + 0.3 + ], + "angle": 0, + "content": "Fig. 2: A partial view of our visual action graph: AP7 consists of two sub-APs (7-1 and 7-2). All bounding boxes (BBs) mark foods (e.g., the destination BB in AP7-1 is oil heated in AP6-1). The duration is defined by the first and last BBs of the sub-APs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.331, + 0.784, + 0.361 + ], + "angle": 0, + "content": "intervention. We first present the formal definition of the visual action graph, followed by an illustrative example to contextualize its concept." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.362, + 0.785, + 0.437 + ], + "angle": 0, + "content": "Define \\(\\mathcal{V}\\) as the set of videos in our dataset, and each video \\(v\\in \\mathcal{V}\\) is a sequence of \\(|v|\\) images, denoted \\(v = \\{v^{1},v^{2},\\ldots ,v^{|v|}\\}\\). Each video is paired with a corresponding sequence of procedural text, denoted as \\(t = \\{t^1,t^2,\\dots ,t^{|t|}\\}\\). We define a visual action graph for a pair of video and text as \\(G(v,t) = (A,E)\\), where \\(A\\) symbolizes the set of APs and \\(E\\) the set of edges, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.785, + 0.482 + ], + "angle": 0, + "content": "Let \\( a_{k} \\) be the \\( k \\)-th AP in the procedural text. We tie \\( a_{k} \\) to visual content for encompassing the time segment in which the AP is performed and bounding boxes showing movements of objects, which can be denoted as" + }, + { + "type": "equation", + "bbox": [ + 0.423, + 0.495, + 0.579, + 0.512 + ], + "angle": 0, + "content": "\\[\na _ {k} = \\left(w _ {k}, A _ {k}, s _ {k}, f _ {k}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.787, + 0.682 + ], + "angle": 0, + "content": "\\(w_{k}\\) is a word sequence of the AP (e.g., \\(w_{7} =\\) \"add the onion and the celery stalk, and\" in Fig. 2). Due to the nature of the videos, \\(a_{k}\\) may involve multiple object instances and multiple action instances. \\(A_{k}\\) represents them as a set of actions in the video (e.g., AP7-1 and AP7-2 for \\(a_7\\)). We denote the \\(r\\)-th element of \\(a_{k}\\) as \\(a_{k}^{(r)} = (b_{k}^{(r),\\mathrm{bef}}, b_{k}^{(r),\\mathrm{aft}}, b_{k}^{(r),\\mathrm{dest}})\\) (e.g., \\(a_7^{(1)} = \\mathrm{AP7 - 1}\\)), representing bounding boxes of the target ingredient before/after the action (\\(\\square / \\square\\)), and the destination (\\(\\square\\)) where the ingredient in \\(b_{k}^{(r),\\mathrm{bef}}\\) is mixed by the action. Note that \\(b_{k}^{(r),\\mathrm{dest}}\\) is not mandatory for \\(a_{k}^{(r)}\\). \\(s_k\\) and \\(f_{k}\\) denote the start and finish frame of the AP, defined by the first and last bounding box in \\(A_{k}\\) (e.g., \\(s_7 = 05:39\\) and \\(f_7 = 05:59\\))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.682, + 0.787, + 0.811 + ], + "angle": 0, + "content": "Nodes in the visual action graph are bounding boxes in APs defined above. Edges \\(E\\) track ingredients throughout the cooking activity. Here, \\(E\\) consists of intra-action edges and inter-action edges. Intra-action edges connect \\(b_{k}^{(r),\\mathrm{bef}}\\) to \\(b_{k}^{(r),\\mathrm{aft}}\\) \\((\\square \\rightarrow \\square)\\) and \\(b_{k}^{(r),\\mathrm{dest}}\\) to \\(b_{k}^{(r),\\mathrm{aft}}\\) \\((\\square \\rightarrow \\square)\\). They are always within the same action, and we can automatically identify these edges based on the shared action index, \\((k,r)\\), where these indices are labeled instead of object names for this annotation. In contrast, we manually identified inter-action edges, which is from \\(b_{k}^{(r),\\mathrm{aft}}\\) to \\(b_{k'}^{(r'),\\mathrm{bef}/\\mathrm{dest}}\\) \\((\\square \\rightarrow \\square\\) or \\(\\square\\))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.785, + 0.84 + ], + "angle": 0, + "content": "To automate the intra-action edge annotation, we labeled each bounding box with an action index instead of ingredient names. Thus, the graph has" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "K. Maeda et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "no ingredient information in this form. To fix this problem, we manually gave ingredient names as leaf nodes and connected them to the AP that first processes the ingredient (ingredient-action edges), following the VRF dataset [44]. This method naturally represents the composition of ingredients at each node by tracing back the edges to the leaf nodes." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.245, + 0.498, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.245, + 0.787, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.336, + 0.788, + 0.365 + ], + "angle": 0, + "content": "Fig.3: Distributions of duration; the averages are 16.6 min and 46.7 sec. for videos and APs, respectively." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.411, + 0.449, + 0.425 + ], + "angle": 0, + "content": "3.3 Annotation Procedure" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.786, + 0.49 + ], + "angle": 0, + "content": "The visual action graph defined above was annotated for each video from February to September 2023. A domain specialist annotated the graph to ensure consistency, with one of the authors reviewing it. The annotator completed the task in about 430 hours." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.787, + 0.612 + ], + "angle": 0, + "content": "Initially, the annotator reviewed the videos and revised the procedural instruction texts based on the performance in each video. Despite instructions to follow the recipes faithfully, the participants often deviated due to the complexity of the cooking task. Instead, we used the revised recipes to simulate participants following them. Simultaneously, the annotator identified and tagged the APs within the instructions. All instructions were initially written in Japanese and translated into English by experts. Each AP corresponded directly between the Japanese and English versions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.787, + 0.675 + ], + "angle": 0, + "content": "Subsequently, the annotator delineated the start and finish times with bounding boxes \\((b_{k}^{(r),\\mathrm{bef}}, b_{k}^{(r),\\mathrm{aft}}, b_{k}^{(r),\\mathrm{dest}})\\) for each action. We used the Computer Vision Annotation Tool (CVAT)5 for this annotation. The video was examined every five frames to identify clear images of the target ingredients." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.787, + 0.751 + ], + "angle": 0, + "content": "Finally, the annotator has assigned inter-action edges to the bounding boxes. Following the automatic generation of edges in intra-action edges, they construct visual action graphs. Ingredients and ingredient-action edges have also been annotated together with inter-action edges. A total of 6,826 bounding boxes and 8,061 relationships were annotated to create the visual action graphs." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.768, + 0.421, + 0.784 + ], + "angle": 0, + "content": "3.4 Statistical Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.789, + 0.787, + 0.821 + ], + "angle": 0, + "content": "Recording Statistics. We aim to collect a large-scale dataset of fixed-viewpoint videos featuring structured, fine-grained annotations in diverse environments." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.824, + 0.468, + 0.841 + ], + "angle": 0, + "content": "5 https://github.com/opencv/cvat" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.63, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "COM Kitchens" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.144, + 0.497, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.148, + 0.785, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.236, + 0.785, + 0.264 + ], + "angle": 0, + "content": "Fig. 4: Distribution of the length of word sequences; the averages are 87.2 and 13.3 words for recipes and sentences." + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.277, + 0.48, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.277, + 0.772, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.416, + 0.48, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.542, + 0.365, + 0.552 + ], + "angle": 0, + "content": "AP" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.416, + 0.77, + 0.538 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.544, + 0.682, + 0.555 + ], + "angle": 0, + "content": "Ingredients" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.566, + 0.788, + 0.608 + ], + "angle": 0, + "content": "Fig. 5: Word cloud and frequency plots. The word clouds visualize our target of fine-grained procedure comprehension well. There are 131 types of APs and 1,187 types of ingredients, both in a clear long-tail trend." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.637, + 0.788, + 0.773 + ], + "angle": 0, + "content": "The COM Kitchens dataset includes 145 videos from 70 kitchens, totaling 40 hours. Kitchens and actors are in one-to-one correspondence; providing 70 unique actors. The average video duration is 16.6 minutes, as illustrated in Fig. 3 (left). \\(80.0\\%\\) of the video frames are labeled with APs, with a \\(15.0\\%\\) overlap, which is comparable to other datasets6. On average, each AP lasted 46.7 seconds (Fig. 3 right). Tab. 1 and Tab. 2 show comparisons with similar recorded datasets. Our dataset stands out for its diverse environments among fixed-viewpoint video datasets. The average video length is longest among the datasets with linguistic annotations." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.787, + 0.787, + 0.818 + ], + "angle": 0, + "content": "Recipe Data. Turning to the linguistic side, we assigned fine instructions comprising 949 sentences, averaging 6.5 sentences per video. The average word lengths" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.525, + 0.841 + ], + "angle": 0, + "content": "6 Referred to Tab. 1 of the Assembly101 [40]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "K. Maeda et al." + }, + { + "type": "text", + "bbox": [ + 0.249, + 0.149, + 0.47, + 0.164 + ], + "angle": 0, + "content": "Query: Visual Observation" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.165, + 0.749, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.246, + 0.497, + 0.26 + ], + "angle": 0, + "content": "Document: Instructional Text" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.262, + 0.489, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.262, + 0.756, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.426, + 0.788, + 0.468 + ], + "angle": 0, + "content": "Fig. 6: Data related to OnRR: The query of OnRR sub-tasks is the first \\( Z\\% \\) of a video. For feasible recipe retrieval, we added an extra recipe resource to enhance the dataset of retrieval targets alongside our test set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.789, + 0.621 + ], + "angle": 0, + "content": "are 87.2 and 13.3 for recipes and sentences (Fig. 4). The instructions included 131 types and 2,286 distinct APs, averaging 15.8 per recipe and 2.4 per sentence. An average of 1.24 repetitions was found in \\(14\\%\\) of the APs. The top of Fig. 5 shows a wordcloud of AP, featuring terms beyond coarse instructions of other datasets in Tab. 2, including fine-grained cooking-specific words like \"Saute.\" The annotated instructions contained 305 types and 1,187 ingredients, averaging 8.2 per recipe. The bottom graphs illustrate a distinct long-tail trend in AP and ingredients." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.645, + 0.376, + 0.663 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.678, + 0.461, + 0.694 + ], + "angle": 0, + "content": "4.1 Online Recipe Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Online Recipe Retrieval (OnRR) involves analyzing cooking videos up to the midpoint to determine the recipe type and the cooking stage achieved. Fig. 6 shows data related to this task; retrieving recipe texts that align with the video content up to a time point \\( Z \\), using the video segment up to the last started AP \\( Y \\) as the search query. This task is more challenging than traditional video-to-text retrieval as it requires both classification of video type and comprehension of the cooking process. We can decompose OnRR into two subtasks: feasible recipe retrieval and recipe stage identification. We assess these subtasks separately because solving them together is challenging for current SOTA methods." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.63, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "COM Kitchens" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.147, + 0.786, + 0.221 + ], + "angle": 0, + "content": "Feasible recipe retrieval evaluates OnRR from a task categorization perspective in an online setting. We consider a recipe feasible for a task in the middle of the workflow when an actor can shift to the recipe without disposing of foods. This setup assumes an application of online recipe recommendation. Task success is measured by whether the retrieved recipe is in the feasible set." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.224, + 0.786, + 0.298 + ], + "angle": 0, + "content": "Recipe stage identification is evaluated with a given recipe. The task is to spot the exact cooking stage in the recipe from the video. It involves aligning the video content with the corresponding part of the recipe text, up to the verb \\( w_{y} \\) that matches AP \\( Y \\), within a given recipe ID \\( X \\). This subtask focuses on accurately matching the video content with the specific steps in the recipe." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.314, + 0.786, + 0.48 + ], + "angle": 0, + "content": "Settings. In our experiment, \\( Z \\) is set at 25%, 50%, 75%, and 100% from the initial point by frame. We labeled them early-, middle-, late-, and full-stage\\(^{7}\\), respectively. The video set was split into 90/26/29 videos for the train/val/test sets, and the test set was further divided into \\( 29 \\times 4 = 116 \\) stages. We compared the performance of state-of-the-art models (UniVL [25], CLIP4Clip [26], and X-CLIP [27]) against random selection. We set the maximum token length, maximum frame length, batch size, and number of training epochs to 48, 48, 32, 5 for UniVL and 32, 12, 32, 3 for CLIP4Clip and X-CLIP. These methods often cannot handle all the words in a recipe at once, as it often exceeds 100 words, as shown in Fig. 4. Hence, we shortened the recipe texts by POS tagging with spaCy, extracting only nouns and verbs to cover all steps with the 100 words." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.497, + 0.786, + 0.678 + ], + "angle": 0, + "content": "Extra Recipe Resource. The number of retrieval targets should be sufficiently large; however, the COM Kitchens test set size is limited. Thus, we have increased retrieval targets with an extra recipe resource from CRD. First, we identified the ingredients appearing in the early stage (\\(Z = 25\\%\\)) by referring to the action graph for each test set video. Based on the early-stage ingredients, we then extracted up to 100 candidate recipes from CRD for each test recipe. This operation collected 1,828 recipes as this extra resource. Among them, we manually discovered \\(991/243/19/5\\) feasible recipes for each stage \\(Z\\), tied with their corresponding test set recipe. Nonfeasible recipes remained as hard negatives, often overlapping in ingredients with feasible recipes. Note that extracted recipes only include recipe texts and ingredients, not videos. We machine-translated them from Japanese to English8." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.695, + 0.786, + 0.74 + ], + "angle": 0, + "content": "Metrics. Following the conventional video-text retrieval task, we employed Recall at rank K (R@K, higher is better) and median rank (MdR, lower is better) to evaluate the retrieval performances at the two subtasks." + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.757, + 0.786, + 0.802 + ], + "angle": 0, + "content": "Results. Tab. 3 showcases the result of the OnRR benchmark in the early- and middle-stage settings. Results for the other two stages and no-fine-tune settings are provided in Appendix D. None of the baseline models outperformed" + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.812, + 0.674, + 0.825 + ], + "angle": 0, + "content": "The full-stage task is equal to conventional video-to-text retrieval." + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.825, + 0.69, + 0.84 + ], + "angle": 0, + "content": "8 We used the DeepL API to translate from Japanese to English (US)." + }, + { + "type": "list", + "bbox": [ + 0.221, + 0.812, + 0.69, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "K. Maeda et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 3: Online recipe retrieval (OnRR) performances of baseline models. R@K and MdR represent recall at rank K \\((\\uparrow)\\) and median rank \\((\\downarrow)\\), respectively. This table provides only the early- and middel-stage settings (using the first \\(25\\%\\) and \\(50\\%\\) of the video as input); results in other stages and settings are detailed in Appendix D." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.213, + 0.756, + 0.376 + ], + "angle": 0, + "content": "
TaskMethodEarly (25%)Middle (50%)
R@1R@5R@10MdRR@1R@5R@10MdR
Feasible Recipe RetrievalRandom1.88.615.8-0.41.83.1-
UniVL [25]3.45.79.2227.03.45.79.2231.0
CLIP4Clip [26]0.00.010.379.00.00.06.885.0
X-CLIP [27]0.06.810.389.00.03.43.4320.0
Recipe Stage IdentificationRandom6.331.663.38.06.331.663.38.0
UniVL [25]17.248.268.95.09.263.389.23.0
CLIP4Clip [26]6.848.268.95.010.355.186.24.0
X-CLIP [27]10.351.768.94.017.237.993.16.0
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.788, + 0.48 + ], + "angle": 0, + "content": "random selection in feasible recipe retrieval. Conversely, all models outperformed random selection in recipe stage identification and improved with fine-tuning. These results suggest that simple contrastive learning cannot solve these two tasks simultaneously. In other words, the OnRR task serves as a benchmark for video-text retrieval based on a procedural comprehension." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.5, + 0.771, + 0.516 + ], + "angle": 0, + "content": "4.2 Dense Video Captioning on Unedited Fixed-viewpoint Videos" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.629 + ], + "angle": 0, + "content": "Dense video captioning is another task for procedural videos, where a system generates multiple detailed captions for different segments within a video. This approach involves detecting distinct events in the video timeline and then generating descriptive and accurate captions for each identified event. The objective is to provide a more comprehensive and segmented understanding of the video content, which is beneficial for offline applications, such as accessibility, content analysis, and enhanced video search capabilities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.705 + ], + "angle": 0, + "content": "For the COM Kitchens dataset, we use APs as the segments of a video; the content is the recipe text of each AP and the entire duration covering all repetitions under each AP. The main challenge of this task is the domain gap between our unedited overhead-view videos and traditional DVC targets such as web videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.808, + 0.795 + ], + "angle": 0, + "content": "Settings. In this experiment, we used the same train-valid/test split with the OnRR task. We selected two DVC systems as our baselines: PDVC [49] and Vid2Seq [53]. We first tested their zero-shot performance on the COM Kitchens dataset. Then, we fine-tuned the Vid2Seq model, the SOTA model on the YouCookII dataset, on our dataset and evaluated its performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Besides, we examined two supervision approaches to leverage the action graph in the DVC task: (i) action graph as relation labels (RL) and (ii) action graph as attention supervision (AS). When we took the action graph as relation labels," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.63, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "COM Kitchens" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 4: Comparison between DVC performances of baseline models. The rows with 'FT' of '\\(\\checkmark\\)' show the results of models fine-tuned on COM Kitchens. \"AG\" shows the choice of methods to leverage action graph information during the fine-tuning." + }, + { + "type": "table", + "bbox": [ + 0.294, + 0.199, + 0.71, + 0.32 + ], + "angle": 0, + "content": "
ModelFTAGSODA_c(↑)CIDEr(↑)METEOR(↑)
PDVC [49]--0.0220.0000.000
Vid2Seq [53]--0.0170.0660.010
Vid2Seq✓-0.3692.8320.642
Vid2Seq✓RL0.2111.3810.285
Vid2Seq✓AS0.2662.5130.423
Vid2Seq✓RL+AS0.5816.1951.142
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.784, + 0.425 + ], + "angle": 0, + "content": "we employed the decoder of TablERT-CNN [28] as our module to predict the relation (a.k.a. the type of edges) between APs and trained the model in a multitask learning manner. For attention supervision, following Garg et al. [13], we minimize the Kullback-Leibler divergence between the self-attention matrix at the last encoder layer and the alignment matrix indicating the existence of edges." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.443, + 0.784, + 0.473 + ], + "angle": 0, + "content": "Metrics. We employed SODA_c [12], CIDEr [48], and METEOR [2] scores to evaluate model performance, as these are the commonly used for DVC tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.785, + 0.596 + ], + "angle": 0, + "content": "Results. Tab. 4 shows the result of the DVC benchmark. The zero-shot performances of both the PDVC and Vid2Seq models are extremely low compared to those on the YouCookII dataset (4.9 and 7.9 SODA_c scores, respectively). This deterioration indicates that the COM Kitchens dataset is challenging for the DVC task. While speech information is proved to be a crucial modality to generate better captions [53], videos in the COM Kitchens dataset have no speech information than web-based procedural video datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.787, + 0.718 + ], + "angle": 0, + "content": "We also found that the supervision obtained from the action graph benefits the model performance. While RL or AS supervision individually does not bring improvement, their combination enhances model performance across all metrics. This performance gain is attributed to improved attention. As shown in the Fig. 7, RL or AS supervision aligns most frames to a small subset, but the \\(\\mathrm{RL + AS}\\) supervision aligns frames to their relative frames (defined by the action graph) and their surrounding frames. This smoother alignment aids in generating recipe texts and determining segment boundaries." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.745, + 0.355, + 0.76 + ], + "angle": 0, + "content": "5 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.84 + ], + "angle": 0, + "content": "We provided baseline results for two specific tasks: OnRR and DVC-OV, extensions of conventional cross-modal retrieval and captioning tasks. OnRR aims to create practical video comprehension for online applications. In contrast, DVC-OV focuses on understanding overhead-view procedural videos as offline tasks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "K. Maeda et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.348, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.145, + 0.495, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.146, + 0.64, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.146, + 0.785, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.26, + 0.347, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.26, + 0.494, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.26, + 0.639, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.659, + 0.26, + 0.784, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.369, + 0.788, + 0.399 + ], + "angle": 0, + "content": "Fig. 7: Visualization of the attention from the first head in the last encoder layer, based on randomly selected samples. Additional examples can be found in Appendix E." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.784, + 0.55 + ], + "angle": 0, + "content": "A potential application involves constructing the visual action graph using video and recipe text inputs. It aims to structurally understand cooking procedures by focusing on temporal dependencies and local human-object interactions. Another possible task is episodic memory in the procedural domain, a linguistic query-driven reasoning task tailored originally for long-duration ego-centric videos [14]. The COM Kitchens dataset is particularly suited for this task, as it comprises long videos (avg. 16.6 min., Tab. 1) and the visual action graph can generate various queries with spatio-temporal ground truth." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.785, + 0.642 + ], + "angle": 0, + "content": "Our dataset, similar to many others, is limited by its size. Furthermore, benchmark results indicate that conventional global alignment methods are ineffective for these tasks, prompting us to explore new pre-training and fine-tuning methods. Fortunately, our data collection costs are significantly lower compared to datasets like EgoExo4D. We aim to further explore the potential of visual action graphs and to expand the dataset both with and without supervision." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.669, + 0.36, + 0.685 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We have introduced COM Kitchens, a dataset that facilitates vision-language understanding with overhead-view recordings, procedural recipe texts, and visual action graphs. This dataset reflects real kitchen conditions, providing rich insights into the sequence of actions and states of ingredients. Our experiments on the OnRR and DVC-OV benchmarks revealed the limitations of existing cross-modal retrieval models in handling long sequences and temporal dependencies. We plan to expand COM Kitchens for further challenges, believing that this dataset will contribute to the advancement of complex video content interpretation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.63, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "COM Kitchens" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.788, + 0.271 + ], + "angle": 0, + "content": "This work would not have been possible without the meticulous and precise annotations by Yasuko Sonoda. We are deeply grateful for her significant contributions to the dataset construction. We also express our gratitude to the IDR office at the National Institute of Informatics for their continuous support in our dataset releases. This work was supported by JSPS KAKENHI Grant Number 21H04910 and JST Moonshot R&D Grant Number JPMJMS2236." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.298, + 0.323, + 0.314 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.333, + 0.785, + 0.361 + ], + "angle": 0, + "content": "1. Ashutosh, K., Girdhar, R., Torresani, L., Grauman, K.: HierVL: Learning hierarchical video-language embeddings. In: CVPR. pp. 23066-23078 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.363, + 0.785, + 0.403 + ], + "angle": 0, + "content": "2. Banerjee, S., Lavie, A.: METEOR: an automatic metric for MT evaluation with improved correlation with human judgments. In: Proc. ACL Workshop IEEMMTS. pp. 65-72 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.785, + 0.445 + ], + "angle": 0, + "content": "3. Ben-Shabat, Y., Yu, X., Saleh, F., Campbell, D., Rodriguez-Opazo, C., Li, H., Gould, S.: The IKEA ASM dataset: Understanding people assembling furniture through actions, objects and pose. In: WACV. pp. 847-859 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.447, + 0.785, + 0.487 + ], + "angle": 0, + "content": "4. Chandu, K., Nyberg, E., Black, A.W.: Storyboarding of recipes: Grounded contextual generation. In: Annual Meeting of the Association for Computational Linguistics. pp. 6040-6046 (Jul 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.489, + 0.785, + 0.516 + ], + "angle": 0, + "content": "5. Chen, S., Zhao, Y., Jin, Q., Wu, Q.: Fine-grained video-text retrieval with hierarchical graph reasoning. In: CVPR (June 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.518, + 0.785, + 0.558 + ], + "angle": 0, + "content": "6. Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., Wray, M.: Scaling egocentric vision: The EPIC-KITCHENS dataset. In: ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.56, + 0.785, + 0.613 + ], + "angle": 0, + "content": "7. Damen, D., Doughty, H., Farinella, G.M., Furnari, A., Ma, J., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., Wray, M.: Rescaling egocentric vision: Collection, pipeline and challenges for EPIC-KITCHENS-100. IJCV 130, 33-55 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.616, + 0.785, + 0.643 + ], + "angle": 0, + "content": "8. Deng, C., Chen, S., Chen, D., He, Y., Wu, Q.: Sketch, ground, and refine: Top-down dense video captioning. In: CVPR. pp. 234-243 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.645, + 0.785, + 0.685 + ], + "angle": 0, + "content": "9. Dvornik, N., Hadji, I., Zhang, R., Derpanis, K.G., Wildes, R.P., Jepson, A.D.: StepFormer: Self-supervised step discovery and localization in instructional videos. In: CVPR. pp. 18952-18961 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.687, + 0.785, + 0.713 + ], + "angle": 0, + "content": "0. Elhamifar, E., Naing, Z.: Unsupervised procedure learning via joint dynamic summarization. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.715, + 0.785, + 0.742 + ], + "angle": 0, + "content": "1. Fathi, A., Li, Y., Rehg, J.M.: Learning to recognize daily actions using gaze. In: ECCV. pp. 314-327 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.743, + 0.785, + 0.784 + ], + "angle": 0, + "content": "2. Fujita, S., Hirao, T., Kamigaito, H., Okumura, M., Nagata, M.: SODA: Story oriented dense video captioning evaluation framework. In: ECCV. pp. 517-531 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "3. Garg, S., Peitz, S., Nallasamy, U., Paulik, M.: Jointly learning to align and translate with transformer models. In: Conference on Empirical Methods in Natural Language Processing and International Joint Conference on Natural Language Processing. pp. 4453-4462. Hong Kong, China (2019)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.333, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "K. Maeda et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.342 + ], + "angle": 0, + "content": "14. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., Martin, M., Nagarajan, T., Radosavovic, I., Ramakrishnan, S.K., Ryan, F., Sharma, J., Wray, M., Xu, M., Xu, E.Z., Zhao, C., Bansal, S., Batra, D., Cartillier, V., Crane, S., Do, T., Doulaty, M., Erapalli, A., Feichtenhofer, C., Fragomeni, A., Fu, Q., Gebreselasie, A., Gonzalez, C., Hillis, J., Huang, X., Huang, Y., Jia, W., Khoo, W., Kolář, J., Kottur, S., Kumar, A., Landini, F., Li, C., Li, Y., Li, Z., Mangalam, K., Modhugu, R., Munro, J., Murrell, T., Nishiyasu, T., Price, W., Ruiz, P., Ramazanova, M., Sari, L., Somasundaram, K., Southerland, A., Sugano, Y., Tao, R., Vo, M., Wang, Y., Wu, X., Yagi, T., Zhao, Z., Zhu, Y., Arbeláez, P., Crandall, D., Damen, D., Farinella, G.M., Fuegen, C., Ghanem, B., Ithapu, V.K., Jawahar, C.V., Joo, H., Kitani, K., Li, H., Newcombe, R., Oliva, A., Park, H.S., Rehg, J.M., Sato, Y., Shi, J., Shou, M.Z., Torralba, A., Torresani, L., Yan, M., Malik, J.: Ego4D: Around the world in 3,000 hours of egocentric video. In: CVPR. pp. 18995-19012 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.344, + 0.787, + 0.579 + ], + "angle": 0, + "content": "15. Grauman, K., Westbury, A., Torresani, L., Kitani, K., Malik, J., Afouras, T., Ashutosh, K., Baiyya, V., Bansal, S., Boote, B., Byrne, E., Chavis, Z., Chen, J., Cheng, F., Chu, F.J., Crane, S., Dasgupta, A., Dong, J., Escobar, M., Forigua, C., Gebreselasie, A., Haresh, S., Huang, J., Islam, M.M., Jain, S., Khirodkar, R., Kukreja, D., Liang, K.J., Liu, J.W., Majumder, S., Mao, Y., Martin, M., Mavroudi, E., Nagarajan, T., Ragusa, F., Ramakrishnan, S.K., Seminara, L., Somayazulu, A., Song, Y., Su, S., Xue, Z., Zhang, E., Zhang, J., Castillo, A., Chen, C., Fu, X., Furuta, R., Gonzalez, C., Gupta, P., Hu, J., Huang, Y., Huang, Y., Khoo, W., Kumar, A., Kuo, R., Lakhavani, S., Liu, M., Luo, M., Luo, Z., Meredith, B., Miller, A., Oguntola, O., Pan, X., Peng, P., Pramanick, S., Ramazanova, M., Ryan, F., Shan, W., Somasundaram, K., Song, C., Southerland, A., Tateno, M., Wang, H., Wang, Y., Yagi, T., Yan, M., Yang, X., Yu, Z., Zha, S.C., Zhao, C., Zhao, Z., Zhu, Z., Zhuo, J., Arbelaez, P., Bertasius, G., Crandall, D., Damen, D., Engel, J., Farinella, G.M., Furnari, A., Ghanem, B., Hoffman, J., Jawahar, C.V., Newcombe, R., Park, H.S., Rehg, J.M., Sato, Y., Savva, M., Shi, J., Shou, M.Z., Wray, M.: Ego-exo4d: Understanding skilled human activity from first- and third-person perspectives (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.581, + 0.787, + 0.623 + ], + "angle": 0, + "content": "16. Harashima, J., Ariga, M., Murata, K., Ioki, M.: A large-scale recipe and meal data collection as infrastructure for food research. In: International Conference on Language Resources and Evaluation. pp. 2455-2459 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.624, + 0.787, + 0.653 + ], + "angle": 0, + "content": "17. Huang, D.A., Lim, J.J., Fei-Fei, L., Niebles, J.C.: Unsupervised visual-linguistic reference resolution in instructional videos. In: CVPR. pp. 2183-2192 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.654, + 0.787, + 0.695 + ], + "angle": 0, + "content": "18. Jermsurawong, J., Habash, N.: Predicting the structure of cooking recipes. In: Conference on Empirical Methods in Natural Language Processing. pp. 781-786 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.697, + 0.787, + 0.738 + ], + "angle": 0, + "content": "19. Ji, W., Li, Y., Wei, M., Shang, X., Xiao, J., Ren, T., Chua, T.S.: VidVRD 2021: The third grand challenge on video relation detection. In: ACM MM. p. 4779-4783 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.74, + 0.786, + 0.768 + ], + "angle": 0, + "content": "20. Johnson, J., Karpathy, A., Fei-Fei, L.: DenseCap: Fully convolutional localization networks for dense captioning. In: CVPR. pp. 4565-4574 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.769, + 0.787, + 0.811 + ], + "angle": 0, + "content": "21. Kiddon, C., Ponnuraj, G.T., Zettlemoyer, L., Choi, Y.: Mise en place: Unsupervised interpretation of instructional recipes. In: Conference on Empirical Methods in Natural Language Processing. pp. 982-992 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.787, + 0.841 + ], + "angle": 0, + "content": "22. Kuehne, H., Arslan, A., Serre, T.: The language of actions: Recovering the syntax and semantics of goal-directed human activities. In: CVPR. pp. 780-787 (2014)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.63, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "COM Kitchens" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.202 + ], + "angle": 0, + "content": "23. Lei, J., Wang, L., Shen, Y., Yu, D., Berg, T., Bansal, M.: MART: memory-augmented recurrent transformer for coherent video paragraph captioning. In: Annual Meeting of the Association for Computational Linguistics. pp. 2603-2614 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.787, + 0.243 + ], + "angle": 0, + "content": "24. Lin, K.Q., Wang, J., Soldan, M., Wray, M., Yan, R., Xu, Z., Gao, D., Tu, R.C., Zhao, W., Kong, W., Cai, C., HongFa, W., Damen, D., Ghanem, B., Liu, W., Shou, M.Z.: Egocentric video-language pretraining. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.244, + 0.786, + 0.284 + ], + "angle": 0, + "content": "25. Luo, H., Ji, L., Shi, B., Huang, H., Duan, N., Li, T., Li, J., Bharti, T., Zhou, M.: UniVL: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.285, + 0.787, + 0.324 + ], + "angle": 0, + "content": "26. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: CLIP4Clip: An empirical study of CLIP for end to end video clip retrieval. Neurocomputing 508, 293-304 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.326, + 0.787, + 0.365 + ], + "angle": 0, + "content": "27. Ma, Y., Xu, G., Sun, X., Yan, M., Zhang, J., Ji, R.: X-CLIP: End-to-end multi-grained contrastive learning for video-text retrieval. In: ACM MM. pp. 638-647 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.367, + 0.787, + 0.407 + ], + "angle": 0, + "content": "28. Ma, Y., Hiraoka, T., Okazaki, N.: Joint entity and relation extraction based on table labeling using convolutional neural networks. In: Workshop on Structured Prediction for NLP. pp. 11-21 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.408, + 0.787, + 0.446 + ], + "angle": 0, + "content": "29. Miech, A., Alayrac, J.B., Smaira, L., Laptev, I., Sivic, J., Zisserman, A.: End-to-end learning of visual Representations from uncurated instructional videos. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.448, + 0.787, + 0.488 + ], + "angle": 0, + "content": "30. Miech, A., Zhukov, D., Alayrac, J.B., Tapaswi, M., Laptev, I., Sivic, J.: HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.489, + 0.787, + 0.515 + ], + "angle": 0, + "content": "31. Momouchi, Y.: Control structures for actions in procedural texts and PT-chart. In: International Conference on Computational Linguistics. pp. 108-114 (1980)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.516, + 0.787, + 0.555 + ], + "angle": 0, + "content": "32. Mori, S., Maeta, H., Yamakata, Y., Sasada, T.: Flow graph corpus from recipe texts. In: International Conference on Language Resources and Evaluation. pp. 2370-2377 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.556, + 0.787, + 0.582 + ], + "angle": 0, + "content": "33. Nakamura, K., Ohashi, H., Okada, M.: Sensor-augmented egocentric-video captioning with dynamic modal attention. In: ACM MM. p. 4220-4229 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.583, + 0.787, + 0.623 + ], + "angle": 0, + "content": "34. Nishimura, T., Hashimoto, A., Mori, S.: Procedural text generation from a photo sequence. In: International Natural Language Generation Conference. pp. 409-414 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.624, + 0.787, + 0.651 + ], + "angle": 0, + "content": "35. Nishimura, T., Hashimoto, A., Ushiku, Y., Kameko, H., Mori, S.: State-aware video procedural captioning. In: ACM MM. pp. 1766-1774 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.651, + 0.787, + 0.69 + ], + "angle": 0, + "content": "36. Nishimura, T., Hashimoto, A., Ushiku, Y., Kameko, H., Yamakata, Y., Mori, S.: Structure-aware procedural text generation from an image sequence. IEEE Access 9, 2125-2141 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.691, + 0.787, + 0.732 + ], + "angle": 0, + "content": "37. Nishimura, T., Sakoda, K., Hashimoto, A., Ushiku, Y., Tanaka, N., Ono, F., Kameko, H., Mori, S.: Egocentric biochemical video-and-language dataset. In: ICCV Workshop. pp. 3129-3133 (October 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.733, + 0.787, + 0.772 + ], + "angle": 0, + "content": "38. Nishimura, T., Sakoda, K., Ushiku, A., Hashimoto, A., Okuda, N., Ono, F., Kameko, H., Mori, S.: BioVL2 dataset: Egocentric biochemical video-and-language dataset. Journal of Natural Language Processing (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.773, + 0.787, + 0.799 + ], + "angle": 0, + "content": "39. Schiele, B.: A database for fine grained activity detection of cooking activities. In: CVPR. p. 1194-1201 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "40. Sener, F., Chatterjee, D., Shelepov, D., He, K., Singhania, D., Wang, R., Yao, A.: Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In: CVPR. pp. 21064-21074 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "K. Maeda et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "41. Shi, B., Ji, L., Liang, Y., Duan, N., Chen, P., Niu, Z., Zhou, M.: Dense procedure captioning in narrated instructional videos. In: Annual Meeting of the Association for Computational Linguistics. pp. 6382-6391 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.786, + 0.23 + ], + "angle": 0, + "content": "42. Shi, B., Ji, L., Niu, Z., Duan, N., Zhou, M., Chen, X.: Learning semantic concepts and temporal alignment for narrated video procedural captioning. In: ACM MM. pp. 4355-4363 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.787, + 0.285 + ], + "angle": 0, + "content": "43. Shimada, A., Kondo, K., Deguchi, D., Morin, G., Stern, H.: Kitchen scene context based gesture recognition: A contest in ICPR2012. In: International Workshop on Advances in Depth Image Analysis and Applications. pp. 168-185 (2013). https://doi.org/10.1007/978-3-642-40303-3_18" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.787, + 0.34 + ], + "angle": 0, + "content": "44. Shirai, K., Hashimoto, A., Nishimura, T., Kameko, H., Kurita, S., Ushiku, Y., Mori, S.: Visual Recipe Flow: A dataset for learning visual state changes of objects with recipe flows. In: International Conference on Computational Linguistics. pp. 3570-3577 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.34, + 0.787, + 0.368 + ], + "angle": 0, + "content": "45. Spriggs, E.H., De La Torre, F., Hebert, M.: Temporal segmentation and activity classification from first-person sensing. In: CVPRW. pp. 17-24 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.368, + 0.787, + 0.408 + ], + "angle": 0, + "content": "46. Stein, S., McKenna, S.J.: Combining embedded accelerometers with computer vision for recognizing food preparation activities. In: ACM international joint conference on Pervasive and ubiquitous computing. p. 729-738 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.408, + 0.787, + 0.449 + ], + "angle": 0, + "content": "47. Tang, Y., Ding, D., Rao, Y., Zheng, Y., Zhang, D., Zhao, L., Lu, J., Zhou, J.: COIN: A large-scale dataset for comprehensive instructional video analysis. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.449, + 0.787, + 0.476 + ], + "angle": 0, + "content": "48. Vedantam, R., Zitnick, C.L., Parikh, D.: CIDEr: consensus-based image description evaluation. In: CVPR. pp. 4566-4575 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.476, + 0.787, + 0.504 + ], + "angle": 0, + "content": "49. Wang, T., Zhang, R., Lu, Z., Zheng, F., Cheng, R., Luo, P.: End-to-end dense video captioning with parallel decoding. In: ICCV. pp. 6847-6857 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.504, + 0.787, + 0.544 + ], + "angle": 0, + "content": "50. Wu, J., Pan, L., Chen, J., Jiang, Y.G.: Ingredient-enriched recipe generation from cooking videos. In: ACM International Conference on Multimedia Retrieval. pp. 249-257 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.787, + 0.585 + ], + "angle": 0, + "content": "51. Yagi, T., Ohashi, M., Huang, Y., Furuta, R., Adachi, S., Mitsuyama, T., Sato, Y.: FineBio: A fine-grained video dataset of biological experiments with hierarchical annotation. arXiv preprint arXiv:2402.00293 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.585, + 0.787, + 0.612 + ], + "angle": 0, + "content": "52. Yamakata, Y., Mori, S., Carroll, J.: English recipe flow graph corpus. In: International Conference on Language Resources and Evaluation. pp. 5187-5194 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.612, + 0.787, + 0.653 + ], + "angle": 0, + "content": "53. Yang, A., Nagrani, A., Seo, P.H., Miech, A., Pont-Tuset, J., Laptev, I., Sivic, J., Schmid, C.: Vid2Seq: Large-scale pretraining of a visual language model for dense video captioning. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.653, + 0.787, + 0.695 + ], + "angle": 0, + "content": "54. Yang, J., Peng, W., Li, X., Guo, Z., Chen, L., Li, B., Ma, Z., Zhou, K., Zhang, W., Loy, C.C., Liu, Z.: Panoptic video scene graph generation. In: CVPR. pp. 18675-18685 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.695, + 0.787, + 0.736 + ], + "angle": 0, + "content": "55. Zellers, R., Lu, J., Lu, X., Yu, Y., Zhao, Y., Salehi, M., Kusupati, A., Hessel, J., Farhadi, A., Choi, Y.: MERLOT Reserve: Neural script knowledge through vision and language and sound. In: CVPR. pp. 16375-16387 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.736, + 0.787, + 0.763 + ], + "angle": 0, + "content": "56. Zhou, L., Kalantidis, Y., Chen, X., Corso, J.J., Rohrbach, M.: Grounded video description. In: CVPR. pp. 6578-6587 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.763, + 0.787, + 0.789 + ], + "angle": 0, + "content": "57. Zhou, L., Xu, C., Corso, J.J.: Towards automatic learning of procedures from web instructional videos. In: AAAI (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.789, + 0.787, + 0.831 + ], + "angle": 0, + "content": "58. Zhukov, D., Alayrac, J.B., Cinbis, R.G., Fouhey, D., Laptev, I., Sivic, J.: Cross-task weakly supervised learning from instructional videos. In: CVPR. pp. 3537-3545 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.831 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/COM Kitchens_ An Unedited Overhead-view Procedural Videos Dataset a Vision-Language Benchmark/82c10f54-7bc3-400b-b41c-eee014f604fb_origin.pdf b/2024/COM Kitchens_ An Unedited Overhead-view Procedural Videos Dataset a Vision-Language Benchmark/82c10f54-7bc3-400b-b41c-eee014f604fb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9857093e03dfa07bbea9cb57711ad5c6d1bb22cd --- /dev/null +++ b/2024/COM Kitchens_ An Unedited Overhead-view Procedural Videos Dataset a Vision-Language Benchmark/82c10f54-7bc3-400b-b41c-eee014f604fb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7545a3d4a271547f5741f31a3bceace8a3c46a99fecc0df28f72b8ecab08872d +size 2463322 diff --git a/2024/COM Kitchens_ An Unedited Overhead-view Procedural Videos Dataset a Vision-Language Benchmark/full.md b/2024/COM Kitchens_ An Unedited Overhead-view Procedural Videos Dataset a Vision-Language Benchmark/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a837ceaed28dcdaf3adf4f0026c0bf557f2e4fc0 --- /dev/null +++ b/2024/COM Kitchens_ An Unedited Overhead-view Procedural Videos Dataset a Vision-Language Benchmark/full.md @@ -0,0 +1,302 @@ +# COM Kitchens: An Unedited Overhead-view Video Dataset as a Vision-Language Benchmark + +Koki Maeda\*1,2, Tosho Hirasawa\*1,3, Atsushi Hashimoto1, Jun Harashima4, Leszek Rybicki4, Yusuke Fukasawa4, and Yoshitaka Ushiku1 + +1 OMRON SINIC X Corp., Tokyo, Japan (*equally contributed) +2 Tokyo Institute of Technology, Tokyo, Japan +3 Tokyo Metropolitan University, Tokyo, Japan, +4 Cookpad Inc., Yokohama, Japan + +![](images/406e9df3490a685c380c25a907cb86b5f105393aacd0c2b06230c983494745c7.jpg) +Fig. 1: Sample from COM Kitchens: the dataset includes unedited overhead-view cooking videos, each manually annotated with a visual action graph that links instructional texts to visual elements through edges from before (□) to after (□) bounding boxes (BBs). dist. BBs (□) represent mixing. Details are provided in Sec. 3.2. + +Abstract. Procedural video understanding is gaining attention in the vision and language community. Deep learning-based video analysis requires extensive data. Consequently, existing works often use web videos as training resources, making it challenging to query instructional contents from raw video observations. To address this issue, we propose a new dataset, COM Kitchens. The dataset consists of unedited overhead-view videos captured by smartphones, in which participants performed food preparation based on given recipes. Fixed-viewpoint video datasets often lack environmental diversity due to high camera setup costs. We used modern wide-angle smartphone lenses to cover cooking counters from sink to cooktop in an overhead view, capturing activity without in-person assistance. With this setup, we collected a diverse dataset by distributing smartphones to participants. With this dataset, we propose the novel video-to-text retrieval task Online Recipe Retrieval (OnRR) and new video captioning domain Dense Video Captioning on unedited Overhead-View videos (DVC-OV). Our experiments verified the capabilities and limitations of current web-video-based SOTA methods in handling these tasks. The dataset and code are available at https://doi.org/10.32130/rdata.6.1 and https://github.com/omron-sinicx/com_kitchens, respectively. + +# 1 Introduction + +Creating higher-value products from raw materials is fundamental to material wealth in society. Understanding these processes through video observation has gained increased attention from the computer vision community [3, 6, 17, 22, 37, 40]. Recent advancements in procedural video analysis rely on large-scale datasets collected from the Web [30, 57, 58] or through ego-vision [6, 11, 14, 37]. These datasets play a crucial role in developing robust techniques and pre-trained models [1, 9, 24-26, 29]. Fixed-viewpoint observation was a major video format before the deep learning era [22, 43, 46]. However, it has received little attention in the last decade because it is hard to find videos with such a format on the Web. A cold-start problem exists: sufficient data is prerequisite to implement a raw-video-based retrieval system. Once available, the system encourages users to upload unedited videos for querying, reinforcing the dataset size. + +To overcome this cold-start problem, we revisit unedited video datasets with fixed viewpoints (FV). To efficiently enhance environmental diversity, we leveraged recent hardware advancements. Modern smartphones are equipped with well-calibrated wide-angle cameras, allowing us to capture large work areas from an overhead view. This way, we can observe processes with minimal occlusion regardless of kitchen layout, making it preferable for practical applications. Additionally, people are familiar with smartphone UIs, allowing us to collect videos without in-person assistance. To prove this concept of new-style dataset development, we scaled the FV procedural video datasets in this way, resulting in 145 videos, which are 40 hours in total footage, as the first collection. + +Revisiting FV videos in this deep learning era offers challenges in understanding long procedural contexts rather than frame-wise image processing. To tackle this, we provide a new dataset, COM Kitchens, with a manually annotated visual action graph [44], linking visual events and text instructions with a workflow graph (Fig. 1). Using this structured annotation, we introduce a novel video2text retrieval task, online cross-modal recipe retrieval (OnRR), and a new domain for video captioning, dense video captioning on unedited overhead-view videos (DVC-OV). OnRR is an online cross-modal task for retrieving corresponding recipes during cooking, designed to develop practical smartphone applications. DVC-OV is an offline cross-modal task that generates instructional text from demonstrations, intended to analyze the domain gap between web- and overhead-view videos through a traditional format of the DVC task. + +The contribution of this paper is four-fold. + +1. We introduce a novel approach to construct datasets of unedited fixed-view videos in diverse environments, leveraging modern smartphones. +2. We provide visual action graph annotations for the first time on unedited videos. +3. We propose a novel task of online recipe retrieval, including its target recipe pool and baselines. +4. We analyzed the SOTA dense video captioning method on unedited overhead-view videos to reveal future challenges. + +# 2 Related work + +This paper proposes a new vision-language video dataset with one novel tasks and one novel domain for DVC. We present comparisons to clarify the dataset novelty in Sec. 2.1, and the task/setting novelties in Sec. 2.2. + +# 2.1 Datasets for Procedural Video Understanding + +Table 1: Comparison between instructional video datasets with fixed-viewpoint (FV) cameras. We categorized temporal segment type into action (e.g., "put a bowl," "crack the egg," "beat the egg"), and step, a higher-level action (e.g., "whisk eggs in a bowl"). + +
datasetyeartopictasks#env.#videos total (h)avg. (m)seg. typeseg. description
MMAC [45]2008Cooking1132815.0action130 actions cls.
MPII [39]2012Cooking14144813.4action65 actions cls.
ACE [43]2012Cooking513523.6action8 actions cls.
50 salads [46]2013Cooking215055.4action51 actions cls.
Breakfast [22]2014Cooking-181,712772.7action10 actions cls.
IKEA ASM [3]2021Furniture45371355.7actionnoun+verb (n+v)
Assembly101 [40]2022Assembly1514,3215137.1act./step1,380 act. cls./n+v
COM KitchensOursCooking139701454016.6act./stepinstructional text
+ +Tab. 1 summarizes datasets with FV procedural videos. They all target manufacturing tasks. Among these datasets, COM Kitchens has a significantly large diversity in tasks and environments and is the only one with linguistic annotations. We omitted the EgoExo4D [15] dataset from the table because the paper does not provide organized statistics. + +Breakfast [22], EgoExo4D [15], and our dataset have environmental diversity, while others struggle with the cost of in-person technical support for their setup. The Breakfast dataset collected data from 18 environments (overhead or side view), but its tasks are limited to two salad recipes with only 10 action classes. EgoExo4D addressed the high cost by collaborating with 12 institutes. Our dataset is competitive with EgoExo4D in terms of the number of environments and participants for the cooking scenario but offers more variety in tasks (i.e., recipes). Another difference is camera views. EgoExo4D used multiple cameras from the front, side, and back view angles, aiming to capture activities beyond the kitchen counter to bridge the gap between egocentric and egocentric views. In contrast, our setup captures detailed food manipulations at the counter with minimal occlusions from the overhead view. + +Tab. 2 compares our dataset to other procedural video datasets with linguistic annotations. We omit video datasets without manual annotations, such as HowTo100M [30] and YT-Temporal-1B [55], as they are for pre-training, not for downstream tasks. The pioneering work of YouCookII [57] provides linguistic annotations of coarse instructions (e.g., "whisk egg, flour"), and many other works + +Table 2: Comparison between instructional vision-language video datasets. Only our dataset uses a fixed viewpoint. Coarse instruction minimally describes steps with verb(s) and noun(s) in a YouCookII style [57] (e.g., "tier dried squid"), whereas fine instruction comes from real instructional texts (e.g., "tear dried squid by hand into small pieces"). 'manual*' annotates only the start of intervals. + +
datasetyeartypetopictasks#videos total (h) avg. (m)seg. descriptioninterval# seg.
YouCookII [57]2018WebCook.892,0001765.3 coarse instruction manual4,325
ProceL [10]2019WebMulti.12720473.9 coarse instruction manual498
COIN [47]2019WebMulti.18011,8274762.4 coarse instruction manual46,354
CrossTask [58]2019WebMulti.834,7003764.8 coarse instruction manual19,278
MMAC-Captions [33]2021EgoCook.51461613.4 coarse instruction manual5,002
Epic Kitchens [7]2022EgoCook.707001008.6 narrationutterance39,596
Ego4D [14]2022EgoOpen--3,670- narrationmanual*-
BioVL2 [38]2022EgoBio.53235.3 fine instruction manual408
VRF [44]2022WebCook.20020020.7 fine instruction manual3,705
FineBio [51]2024EgoBio.722614.53.9 fine instruction manual3,541
COM KitchensOursFVCook.1391454016.6 fine instruction manual2,852
+ +follow this manner. Epic Kitchens and Ego4D have narrations as their linguistic annotation, but they tend to describe the details of each action (e.g., action name and target objects), which is still similar to coarse instruction. While captioning for procedural image sequences [4,34,36] uses commercial recipe sites as the dataset resource for generating fine instructions (e.g., "Beat the egg whites with a mixer, starting on low speed"), there is a gap with current video datasets' linguistic resource. + +BioVL2 [37], FineBio [51], and VRF [44] are datasets with fine instructions, as this work. BioVL2 is a rare dataset capturing biochemical experiments, but it is limited in size. FineBio is a five times larger than BioVL2 in total footage, but its task variation is still limited and the average footage is shorter than general cooking tasks. VRF collected one-minute videos focused on food state changes, excluding human actions. COM Kitchens has $70\%$ of the diversity against the web-based VRF dataset, consisting of unedited videos with 16.6 minutes on average. + +Our dataset provides its annotation as visual action graphs. This structured representation of manufacturing instruction was first proposed in 1980 by Momouchi [31] as an estimation target for a natural language processing task. Later, Flow graph [32,52] was proposed with a fine-grained graph as a comprehensive representation of understanding. Merging tree was the simplest structure of the manufacturing process [18], which selects only actions that merge materials as nodes [18]. Action graph is an intermediate representation between flow graph and merging tree, represents both actions on single materials and merging actions [21]. + +These structures have been independently extended to vision-language setups. Visual action graphs were defined as a prediction target in an unsupervised task [17] using web videos. Unfortunately, their visual action graph dataset is only for testing and is not publicly available. A small dataset of visual merging trees has been provided in [36] for semi-supervised learning, where visual + +data consists of image sequences instead of videos. VRF [44] provides visual flow graphs with videos, but the videos are only one minute long. COM Kitchens is the only dataset that annotates graph structures for action segments. Note that the graph represents dynamic changes across frames, including merging and splitting processes. This point is an essential difference from video scene graph datasets [19, 54]. + +# 2.2 Video-Text Retrieval and Video Captioning + +Retrieval is one of the fundamental tasks for cross-modal problems. For videos, video-text retrieval [5,25-27], a task to find the video whose entire contents fit a text query, is the principal retrieval task. The literature often evaluates the video-to-text scenario in addition to the text-to-video condition; web videos usually have text metadata, making video-to-text evaluation less practical. Instead, the OnRR task assumes retrieving web content by its text from raw video observation. Since our dataset is sourced from smartphones, developed techniques should be directly applicable to smartphone videos. + +Video paragraph captioning is a video captioning task specially designed for procedural videos; it assumes that event segments are given [23,35,42,50]. However, for unedited videos, it is not practical to assume such given event segments. Hence, we focus on dense video captioning (DVC) [20, 56], a joint task of event detection and event description generation, as a fundamental task of video comprehension. The recent main challenge of DVC is to suppress redundant detection [12]. To overcome this problem, "detect-then-describe" [8,41] and "describe-then-detect" [49] approaches have been studied. The current SOTA method of Vid2Seq [53] describes captions and detects events in one stage by outputting segments as a time token with its caption. Its performance is supported by the largest video dataset of YT-Temporal-1B [55]. + +We test our dataset with Vid2Seq to investigate the domain gap between Web and FV videos. The most significant gap is in video length and repetitive actions. Tab. 2 shows that videos in COM Kitchens are about three times longer than usual web videos on average (and 20 times longer than VRF, which are the TikTok style). This difference is mainly due to repetitive actions, often eliminated in web videos. Another gap is the location of event-related objects within each frame. In web videos (or ego vision), the object of interest tends to be in the center of the frame, whereas FV videos do not dynamically focus on objects. These repetitions and lack of focus on important objects provide additional challenges for aligning linguistic instructions to video events. + +# 3 The COM Kitchens Dataset + +Data collection and annotation are pivotal in determining the dataset's utility. This section presents our data collection approach (Sec. 3.1), a detailed definition of the visual action graph (Sec. 3.2), annotation procedure (Sec. 3.3), and dataset statistics (Sec. 3.4). + +# 3.1 Collecting Videos with their Corresponding Procedural Text + +We selected candidate recipes for filming the cooking process from the Cookpad Recipe Dataset (CRD) [16]. The CRD is a comprehensive Japanese home cuisine database, including over 1.7 million recipes with ingredient lists. In the selection process, we prioritized recipes estimated to take under 30 minutes to prepare and to have a moderate complexity level. We excluded recipes using off-counter facilities (e.g., microwaves) because such actions would be recognizable with sound-based event identification, which is not our focus. The selected recipes are typically but not limited to European, Chinese, and Japanese cuisines (with a certain level of localization). + +To collect videos, we hired 110 participants from August to October 2021. Each participant recorded up to four recipes at home. Instead of in-person guidance, we provided an instructional document and video that directed the way of recording (e.g., cooking alone, where to capture, camera mode, and turning off TVs) and how to exclude privacy-related information (e.g., do not have a conversation with family, set the camera view not to capture the actor's face, and remove any private documents from camera view). Participants signed a consent form acknowledging that the data would be used publicly for academic purpose. Participants received appropriate financial compensation. + +We utilized the iPhone 11 Pro fixed to a tripod and recorded videos using the rear camera. We instructed participants to set the camera to 30 fps, Full HD resolution of $1920 \times 1080$ , and ultra-wide field of view mode (equivalent to a focal length of $13~\mathrm{mm}$ ). Despite the ultra-wide angle setting, there was almost no distortion; therefore, we applied no extra intrinsic calibration. Appendix B provides filmed video examples. + +Out of the 410 videos recorded, we excluded any that ignored the instructions, curating a total of 210 videos. See Appendix C for detailed statistics. The most common reason for rejection was wrong observation areas. The second was unintentionally observed faces of actors. We made our best effort to avoid involving private information, considering that participants are not professionals and the environment is personal. This emphasis on privacy made data collection more challenging. + +Among the 210 available videos, we have annotated 145 videos, which amount to a total of 40 hours. We will publish the remaining 85 videos as an unsupervised resource. Two videos were recorded at 60fps; however, we retained them in the dataset as they exhibited no quality issues. In the experiment, we used videos scaled to $640 \times 480$ resolution, but we will release the original resolution together publicly. + +# 3.2 Grounding Text to Video via Visual Action Graph + +A visual action graph delineates intentional actions termed action-by-person (AP) into a temporal graph format. Here, AP (a.k.a., action-by-chef [32]) refers to human actions on ingredients, like cutting or mixing, as opposed to natural processes, such as the browning of ingredients, which occur without human + +![](images/5586dc17f48bea841bb51260c2f09a58cfd67cd75d7b2d1b4f05be2311d12d00.jpg) +Fig. 2: A partial view of our visual action graph: AP7 consists of two sub-APs (7-1 and 7-2). All bounding boxes (BBs) mark foods (e.g., the destination BB in AP7-1 is oil heated in AP6-1). The duration is defined by the first and last BBs of the sub-APs. + +intervention. We first present the formal definition of the visual action graph, followed by an illustrative example to contextualize its concept. + +Define $\mathcal{V}$ as the set of videos in our dataset, and each video $v\in \mathcal{V}$ is a sequence of $|v|$ images, denoted $v = \{v^{1},v^{2},\ldots ,v^{|v|}\}$ . Each video is paired with a corresponding sequence of procedural text, denoted as $t = \{t^1,t^2,\dots ,t^{|t|}\}$ . We define a visual action graph for a pair of video and text as $G(v,t) = (A,E)$ , where $A$ symbolizes the set of APs and $E$ the set of edges, respectively. + +Let $a_{k}$ be the $k$ -th AP in the procedural text. We tie $a_{k}$ to visual content for encompassing the time segment in which the AP is performed and bounding boxes showing movements of objects, which can be denoted as + +$$ +a _ {k} = \left(w _ {k}, A _ {k}, s _ {k}, f _ {k}\right). +$$ + +$w_{k}$ is a word sequence of the AP (e.g., $w_{7} =$ "add the onion and the celery stalk, and" in Fig. 2). Due to the nature of the videos, $a_{k}$ may involve multiple object instances and multiple action instances. $A_{k}$ represents them as a set of actions in the video (e.g., AP7-1 and AP7-2 for $a_7$ ). We denote the $r$ -th element of $a_{k}$ as $a_{k}^{(r)} = (b_{k}^{(r),\mathrm{bef}}, b_{k}^{(r),\mathrm{aft}}, b_{k}^{(r),\mathrm{dest}})$ (e.g., $a_7^{(1)} = \mathrm{AP7 - 1}$ ), representing bounding boxes of the target ingredient before/after the action ( $\square / \square$ ), and the destination ( $\square$ ) where the ingredient in $b_{k}^{(r),\mathrm{bef}}$ is mixed by the action. Note that $b_{k}^{(r),\mathrm{dest}}$ is not mandatory for $a_{k}^{(r)}$ . $s_k$ and $f_{k}$ denote the start and finish frame of the AP, defined by the first and last bounding box in $A_{k}$ (e.g., $s_7 = 05:39$ and $f_7 = 05:59$ ). + +Nodes in the visual action graph are bounding boxes in APs defined above. Edges $E$ track ingredients throughout the cooking activity. Here, $E$ consists of intra-action edges and inter-action edges. Intra-action edges connect $b_{k}^{(r),\mathrm{bef}}$ to $b_{k}^{(r),\mathrm{aft}}$ $(\square \rightarrow \square)$ and $b_{k}^{(r),\mathrm{dest}}$ to $b_{k}^{(r),\mathrm{aft}}$ $(\square \rightarrow \square)$ . They are always within the same action, and we can automatically identify these edges based on the shared action index, $(k,r)$ , where these indices are labeled instead of object names for this annotation. In contrast, we manually identified inter-action edges, which is from $b_{k}^{(r),\mathrm{aft}}$ to $b_{k'}^{(r'),\mathrm{bef}/\mathrm{dest}}$ $(\square \rightarrow \square$ or $\square$ ). + +To automate the intra-action edge annotation, we labeled each bounding box with an action index instead of ingredient names. Thus, the graph has + +no ingredient information in this form. To fix this problem, we manually gave ingredient names as leaf nodes and connected them to the AP that first processes the ingredient (ingredient-action edges), following the VRF dataset [44]. This method naturally represents the composition of ingredients at each node by tracing back the edges to the leaf nodes. + +![](images/370d6549dd4a83101dbf094861a482b7fda6e1e95d359c5da479dddacdff7911.jpg) +Fig.3: Distributions of duration; the averages are 16.6 min and 46.7 sec. for videos and APs, respectively. + +![](images/54dc1907671729e52dfc4d6a6bcb40d58a16a58aaaa0fe28a23c6d2b78b95977.jpg) + +# 3.3 Annotation Procedure + +The visual action graph defined above was annotated for each video from February to September 2023. A domain specialist annotated the graph to ensure consistency, with one of the authors reviewing it. The annotator completed the task in about 430 hours. + +Initially, the annotator reviewed the videos and revised the procedural instruction texts based on the performance in each video. Despite instructions to follow the recipes faithfully, the participants often deviated due to the complexity of the cooking task. Instead, we used the revised recipes to simulate participants following them. Simultaneously, the annotator identified and tagged the APs within the instructions. All instructions were initially written in Japanese and translated into English by experts. Each AP corresponded directly between the Japanese and English versions. + +Subsequently, the annotator delineated the start and finish times with bounding boxes $(b_{k}^{(r),\mathrm{bef}}, b_{k}^{(r),\mathrm{aft}}, b_{k}^{(r),\mathrm{dest}})$ for each action. We used the Computer Vision Annotation Tool (CVAT)5 for this annotation. The video was examined every five frames to identify clear images of the target ingredients. + +Finally, the annotator has assigned inter-action edges to the bounding boxes. Following the automatic generation of edges in intra-action edges, they construct visual action graphs. Ingredients and ingredient-action edges have also been annotated together with inter-action edges. A total of 6,826 bounding boxes and 8,061 relationships were annotated to create the visual action graphs. + +# 3.4 Statistical Analysis + +Recording Statistics. We aim to collect a large-scale dataset of fixed-viewpoint videos featuring structured, fine-grained annotations in diverse environments. + +![](images/650c4993bd5e608152636b2b2e120689009989e80bf38fe4511445930f38c593.jpg) +Fig. 4: Distribution of the length of word sequences; the averages are 87.2 and 13.3 words for recipes and sentences. + +![](images/7d97bc20989f403cc6e3e45e94e113e26cdf81ac8021a8027feef619dfea3084.jpg) + +![](images/9f7a3c8852b8f681475fdeee953034b826f676a736bbef12968fc5d9818f501c.jpg) + +![](images/4bd878f0e2bd585b6c8b26e755e26c5894b3d0edace3f45fe1ead4e13dcede3b.jpg) + +![](images/caa1e12cee3435d7f03d12f60d7c2b54e63466e6f09ceecd21f29dfc037a16ea.jpg) +AP + +![](images/8741711702f95de7b3e213625a4bfcd4cd3af2df2dc1f040c16484400fe2f48a.jpg) +Ingredients +Fig. 5: Word cloud and frequency plots. The word clouds visualize our target of fine-grained procedure comprehension well. There are 131 types of APs and 1,187 types of ingredients, both in a clear long-tail trend. + +The COM Kitchens dataset includes 145 videos from 70 kitchens, totaling 40 hours. Kitchens and actors are in one-to-one correspondence; providing 70 unique actors. The average video duration is 16.6 minutes, as illustrated in Fig. 3 (left). $80.0\%$ of the video frames are labeled with APs, with a $15.0\%$ overlap, which is comparable to other datasets6. On average, each AP lasted 46.7 seconds (Fig. 3 right). Tab. 1 and Tab. 2 show comparisons with similar recorded datasets. Our dataset stands out for its diverse environments among fixed-viewpoint video datasets. The average video length is longest among the datasets with linguistic annotations. + +Recipe Data. Turning to the linguistic side, we assigned fine instructions comprising 949 sentences, averaging 6.5 sentences per video. The average word lengths + +Query: Visual Observation + +![](images/133b21742564d64607bcf05d24d19ff8a39898b1fec1cb5aee0949415c62d88d.jpg) + +Document: Instructional Text + +![](images/e3c5e815aa7382dbd2b984b2f922d7668ea339030408aff0c233c3fab62d66df.jpg) +Fig. 6: Data related to OnRR: The query of OnRR sub-tasks is the first $Z\%$ of a video. For feasible recipe retrieval, we added an extra recipe resource to enhance the dataset of retrieval targets alongside our test set. + +![](images/36eda48772eb29a5f0bde41ae52f2870aaca3bebc08e0fbbae9be73ff2c30185.jpg) + +are 87.2 and 13.3 for recipes and sentences (Fig. 4). The instructions included 131 types and 2,286 distinct APs, averaging 15.8 per recipe and 2.4 per sentence. An average of 1.24 repetitions was found in $14\%$ of the APs. The top of Fig. 5 shows a wordcloud of AP, featuring terms beyond coarse instructions of other datasets in Tab. 2, including fine-grained cooking-specific words like "Saute." The annotated instructions contained 305 types and 1,187 ingredients, averaging 8.2 per recipe. The bottom graphs illustrate a distinct long-tail trend in AP and ingredients. + +# 4 Experiments + +# 4.1 Online Recipe Retrieval + +Online Recipe Retrieval (OnRR) involves analyzing cooking videos up to the midpoint to determine the recipe type and the cooking stage achieved. Fig. 6 shows data related to this task; retrieving recipe texts that align with the video content up to a time point $Z$ , using the video segment up to the last started AP $Y$ as the search query. This task is more challenging than traditional video-to-text retrieval as it requires both classification of video type and comprehension of the cooking process. We can decompose OnRR into two subtasks: feasible recipe retrieval and recipe stage identification. We assess these subtasks separately because solving them together is challenging for current SOTA methods. + +Feasible recipe retrieval evaluates OnRR from a task categorization perspective in an online setting. We consider a recipe feasible for a task in the middle of the workflow when an actor can shift to the recipe without disposing of foods. This setup assumes an application of online recipe recommendation. Task success is measured by whether the retrieved recipe is in the feasible set. + +Recipe stage identification is evaluated with a given recipe. The task is to spot the exact cooking stage in the recipe from the video. It involves aligning the video content with the corresponding part of the recipe text, up to the verb $w_{y}$ that matches AP $Y$ , within a given recipe ID $X$ . This subtask focuses on accurately matching the video content with the specific steps in the recipe. + +Settings. In our experiment, $Z$ is set at 25%, 50%, 75%, and 100% from the initial point by frame. We labeled them early-, middle-, late-, and full-stage $^{7}$ , respectively. The video set was split into 90/26/29 videos for the train/val/test sets, and the test set was further divided into $29 \times 4 = 116$ stages. We compared the performance of state-of-the-art models (UniVL [25], CLIP4Clip [26], and X-CLIP [27]) against random selection. We set the maximum token length, maximum frame length, batch size, and number of training epochs to 48, 48, 32, 5 for UniVL and 32, 12, 32, 3 for CLIP4Clip and X-CLIP. These methods often cannot handle all the words in a recipe at once, as it often exceeds 100 words, as shown in Fig. 4. Hence, we shortened the recipe texts by POS tagging with spaCy, extracting only nouns and verbs to cover all steps with the 100 words. + +Extra Recipe Resource. The number of retrieval targets should be sufficiently large; however, the COM Kitchens test set size is limited. Thus, we have increased retrieval targets with an extra recipe resource from CRD. First, we identified the ingredients appearing in the early stage ( $Z = 25\%$ ) by referring to the action graph for each test set video. Based on the early-stage ingredients, we then extracted up to 100 candidate recipes from CRD for each test recipe. This operation collected 1,828 recipes as this extra resource. Among them, we manually discovered $991/243/19/5$ feasible recipes for each stage $Z$ , tied with their corresponding test set recipe. Nonfeasible recipes remained as hard negatives, often overlapping in ingredients with feasible recipes. Note that extracted recipes only include recipe texts and ingredients, not videos. We machine-translated them from Japanese to English8. + +Metrics. Following the conventional video-text retrieval task, we employed Recall at rank K (R@K, higher is better) and median rank (MdR, lower is better) to evaluate the retrieval performances at the two subtasks. + +Results. Tab. 3 showcases the result of the OnRR benchmark in the early- and middle-stage settings. Results for the other two stages and no-fine-tune settings are provided in Appendix D. None of the baseline models outperformed + +Table 3: Online recipe retrieval (OnRR) performances of baseline models. R@K and MdR represent recall at rank K $(\uparrow)$ and median rank $(\downarrow)$ , respectively. This table provides only the early- and middel-stage settings (using the first $25\%$ and $50\%$ of the video as input); results in other stages and settings are detailed in Appendix D. + +
TaskMethodEarly (25%)Middle (50%)
R@1R@5R@10MdRR@1R@5R@10MdR
Feasible Recipe RetrievalRandom1.88.615.8-0.41.83.1-
UniVL [25]3.45.79.2227.03.45.79.2231.0
CLIP4Clip [26]0.00.010.379.00.00.06.885.0
X-CLIP [27]0.06.810.389.00.03.43.4320.0
Recipe Stage IdentificationRandom6.331.663.38.06.331.663.38.0
UniVL [25]17.248.268.95.09.263.389.23.0
CLIP4Clip [26]6.848.268.95.010.355.186.24.0
X-CLIP [27]10.351.768.94.017.237.993.16.0
+ +random selection in feasible recipe retrieval. Conversely, all models outperformed random selection in recipe stage identification and improved with fine-tuning. These results suggest that simple contrastive learning cannot solve these two tasks simultaneously. In other words, the OnRR task serves as a benchmark for video-text retrieval based on a procedural comprehension. + +# 4.2 Dense Video Captioning on Unedited Fixed-viewpoint Videos + +Dense video captioning is another task for procedural videos, where a system generates multiple detailed captions for different segments within a video. This approach involves detecting distinct events in the video timeline and then generating descriptive and accurate captions for each identified event. The objective is to provide a more comprehensive and segmented understanding of the video content, which is beneficial for offline applications, such as accessibility, content analysis, and enhanced video search capabilities. + +For the COM Kitchens dataset, we use APs as the segments of a video; the content is the recipe text of each AP and the entire duration covering all repetitions under each AP. The main challenge of this task is the domain gap between our unedited overhead-view videos and traditional DVC targets such as web videos. + +Settings. In this experiment, we used the same train-valid/test split with the OnRR task. We selected two DVC systems as our baselines: PDVC [49] and Vid2Seq [53]. We first tested their zero-shot performance on the COM Kitchens dataset. Then, we fine-tuned the Vid2Seq model, the SOTA model on the YouCookII dataset, on our dataset and evaluated its performance. + +Besides, we examined two supervision approaches to leverage the action graph in the DVC task: (i) action graph as relation labels (RL) and (ii) action graph as attention supervision (AS). When we took the action graph as relation labels, + +Table 4: Comparison between DVC performances of baseline models. The rows with 'FT' of ' $\checkmark$ ' show the results of models fine-tuned on COM Kitchens. "AG" shows the choice of methods to leverage action graph information during the fine-tuning. + +
ModelFTAGSODA_c(↑)CIDEr(↑)METEOR(↑)
PDVC [49]--0.0220.0000.000
Vid2Seq [53]--0.0170.0660.010
Vid2Seq✓-0.3692.8320.642
Vid2Seq✓RL0.2111.3810.285
Vid2Seq✓AS0.2662.5130.423
Vid2Seq✓RL+AS0.5816.1951.142
+ +we employed the decoder of TablERT-CNN [28] as our module to predict the relation (a.k.a. the type of edges) between APs and trained the model in a multitask learning manner. For attention supervision, following Garg et al. [13], we minimize the Kullback-Leibler divergence between the self-attention matrix at the last encoder layer and the alignment matrix indicating the existence of edges. + +Metrics. We employed SODA_c [12], CIDEr [48], and METEOR [2] scores to evaluate model performance, as these are the commonly used for DVC tasks. + +Results. Tab. 4 shows the result of the DVC benchmark. The zero-shot performances of both the PDVC and Vid2Seq models are extremely low compared to those on the YouCookII dataset (4.9 and 7.9 SODA_c scores, respectively). This deterioration indicates that the COM Kitchens dataset is challenging for the DVC task. While speech information is proved to be a crucial modality to generate better captions [53], videos in the COM Kitchens dataset have no speech information than web-based procedural video datasets. + +We also found that the supervision obtained from the action graph benefits the model performance. While RL or AS supervision individually does not bring improvement, their combination enhances model performance across all metrics. This performance gain is attributed to improved attention. As shown in the Fig. 7, RL or AS supervision aligns most frames to a small subset, but the $\mathrm{RL + AS}$ supervision aligns frames to their relative frames (defined by the action graph) and their surrounding frames. This smoother alignment aids in generating recipe texts and determining segment boundaries. + +# 5 Discussion + +We provided baseline results for two specific tasks: OnRR and DVC-OV, extensions of conventional cross-modal retrieval and captioning tasks. OnRR aims to create practical video comprehension for online applications. In contrast, DVC-OV focuses on understanding overhead-view procedural videos as offline tasks. + +![](images/1783b55b711f38e236e61d9ed97b44880c2a71591ea5c8a4fd8fef1c1beb07a1.jpg) + +![](images/2f67ee35a58b681b9a2f8c867aa30938ecb4160bff10626c30c1875cddb11d31.jpg) + +![](images/5c5550d15dd3f120cec54a20bc20232ac7c40a030f134df8c4a0bc91a4a48a09.jpg) + +![](images/619a0f1d89c22d3503c120ac56200755286a596de31ab0e0556c4d3fc0689d98.jpg) + +![](images/4180d0f75e26967e3b855933c7e5ac3b913e8d4244de67e5b6bf648e4a6ec92a.jpg) +Fig. 7: Visualization of the attention from the first head in the last encoder layer, based on randomly selected samples. Additional examples can be found in Appendix E. + +![](images/b1b06dd9ff2bb7632b2a09419d34e518bf78212e8c9fa9c94c5f0cc10bd3a84f.jpg) + +![](images/084df8601ec153e213edc4ebd180f5f7984185b5bf3992e644640cfae4bb414f.jpg) + +![](images/d8831163f55968550eb695f6f2ed8681ea83255b4a0298f1d16088f3bb6f4910.jpg) + +A potential application involves constructing the visual action graph using video and recipe text inputs. It aims to structurally understand cooking procedures by focusing on temporal dependencies and local human-object interactions. Another possible task is episodic memory in the procedural domain, a linguistic query-driven reasoning task tailored originally for long-duration ego-centric videos [14]. The COM Kitchens dataset is particularly suited for this task, as it comprises long videos (avg. 16.6 min., Tab. 1) and the visual action graph can generate various queries with spatio-temporal ground truth. + +Our dataset, similar to many others, is limited by its size. Furthermore, benchmark results indicate that conventional global alignment methods are ineffective for these tasks, prompting us to explore new pre-training and fine-tuning methods. Fortunately, our data collection costs are significantly lower compared to datasets like EgoExo4D. We aim to further explore the potential of visual action graphs and to expand the dataset both with and without supervision. + +# 6 Conclusion + +We have introduced COM Kitchens, a dataset that facilitates vision-language understanding with overhead-view recordings, procedural recipe texts, and visual action graphs. This dataset reflects real kitchen conditions, providing rich insights into the sequence of actions and states of ingredients. Our experiments on the OnRR and DVC-OV benchmarks revealed the limitations of existing cross-modal retrieval models in handling long sequences and temporal dependencies. We plan to expand COM Kitchens for further challenges, believing that this dataset will contribute to the advancement of complex video content interpretation. + +# Acknowledgement + +This work would not have been possible without the meticulous and precise annotations by Yasuko Sonoda. We are deeply grateful for her significant contributions to the dataset construction. We also express our gratitude to the IDR office at the National Institute of Informatics for their continuous support in our dataset releases. This work was supported by JSPS KAKENHI Grant Number 21H04910 and JST Moonshot R&D Grant Number JPMJMS2236. + +# References + +1. Ashutosh, K., Girdhar, R., Torresani, L., Grauman, K.: HierVL: Learning hierarchical video-language embeddings. In: CVPR. pp. 23066-23078 (June 2023) +2. Banerjee, S., Lavie, A.: METEOR: an automatic metric for MT evaluation with improved correlation with human judgments. In: Proc. ACL Workshop IEEMMTS. pp. 65-72 (2005) +3. Ben-Shabat, Y., Yu, X., Saleh, F., Campbell, D., Rodriguez-Opazo, C., Li, H., Gould, S.: The IKEA ASM dataset: Understanding people assembling furniture through actions, objects and pose. In: WACV. pp. 847-859 (2021) +4. Chandu, K., Nyberg, E., Black, A.W.: Storyboarding of recipes: Grounded contextual generation. In: Annual Meeting of the Association for Computational Linguistics. pp. 6040-6046 (Jul 2019) +5. Chen, S., Zhao, Y., Jin, Q., Wu, Q.: Fine-grained video-text retrieval with hierarchical graph reasoning. In: CVPR (June 2020) +6. Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., Wray, M.: Scaling egocentric vision: The EPIC-KITCHENS dataset. In: ECCV (2018) +7. Damen, D., Doughty, H., Farinella, G.M., Furnari, A., Ma, J., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., Wray, M.: Rescaling egocentric vision: Collection, pipeline and challenges for EPIC-KITCHENS-100. IJCV 130, 33-55 (2022) +8. Deng, C., Chen, S., Chen, D., He, Y., Wu, Q.: Sketch, ground, and refine: Top-down dense video captioning. In: CVPR. pp. 234-243 (2021) +9. Dvornik, N., Hadji, I., Zhang, R., Derpanis, K.G., Wildes, R.P., Jepson, A.D.: StepFormer: Self-supervised step discovery and localization in instructional videos. In: CVPR. pp. 18952-18961 (June 2023) +0. Elhamifar, E., Naing, Z.: Unsupervised procedure learning via joint dynamic summarization. In: ICCV (2019) +1. Fathi, A., Li, Y., Rehg, J.M.: Learning to recognize daily actions using gaze. In: ECCV. pp. 314-327 (2012) +2. Fujita, S., Hirao, T., Kamigaito, H., Okumura, M., Nagata, M.: SODA: Story oriented dense video captioning evaluation framework. In: ECCV. pp. 517-531 (2020) +3. Garg, S., Peitz, S., Nallasamy, U., Paulik, M.: Jointly learning to align and translate with transformer models. In: Conference on Empirical Methods in Natural Language Processing and International Joint Conference on Natural Language Processing. pp. 4453-4462. Hong Kong, China (2019) + +14. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., Martin, M., Nagarajan, T., Radosavovic, I., Ramakrishnan, S.K., Ryan, F., Sharma, J., Wray, M., Xu, M., Xu, E.Z., Zhao, C., Bansal, S., Batra, D., Cartillier, V., Crane, S., Do, T., Doulaty, M., Erapalli, A., Feichtenhofer, C., Fragomeni, A., Fu, Q., Gebreselasie, A., Gonzalez, C., Hillis, J., Huang, X., Huang, Y., Jia, W., Khoo, W., Kolář, J., Kottur, S., Kumar, A., Landini, F., Li, C., Li, Y., Li, Z., Mangalam, K., Modhugu, R., Munro, J., Murrell, T., Nishiyasu, T., Price, W., Ruiz, P., Ramazanova, M., Sari, L., Somasundaram, K., Southerland, A., Sugano, Y., Tao, R., Vo, M., Wang, Y., Wu, X., Yagi, T., Zhao, Z., Zhu, Y., Arbeláez, P., Crandall, D., Damen, D., Farinella, G.M., Fuegen, C., Ghanem, B., Ithapu, V.K., Jawahar, C.V., Joo, H., Kitani, K., Li, H., Newcombe, R., Oliva, A., Park, H.S., Rehg, J.M., Sato, Y., Shi, J., Shou, M.Z., Torralba, A., Torresani, L., Yan, M., Malik, J.: Ego4D: Around the world in 3,000 hours of egocentric video. In: CVPR. pp. 18995-19012 (June 2022) +15. Grauman, K., Westbury, A., Torresani, L., Kitani, K., Malik, J., Afouras, T., Ashutosh, K., Baiyya, V., Bansal, S., Boote, B., Byrne, E., Chavis, Z., Chen, J., Cheng, F., Chu, F.J., Crane, S., Dasgupta, A., Dong, J., Escobar, M., Forigua, C., Gebreselasie, A., Haresh, S., Huang, J., Islam, M.M., Jain, S., Khirodkar, R., Kukreja, D., Liang, K.J., Liu, J.W., Majumder, S., Mao, Y., Martin, M., Mavroudi, E., Nagarajan, T., Ragusa, F., Ramakrishnan, S.K., Seminara, L., Somayazulu, A., Song, Y., Su, S., Xue, Z., Zhang, E., Zhang, J., Castillo, A., Chen, C., Fu, X., Furuta, R., Gonzalez, C., Gupta, P., Hu, J., Huang, Y., Huang, Y., Khoo, W., Kumar, A., Kuo, R., Lakhavani, S., Liu, M., Luo, M., Luo, Z., Meredith, B., Miller, A., Oguntola, O., Pan, X., Peng, P., Pramanick, S., Ramazanova, M., Ryan, F., Shan, W., Somasundaram, K., Song, C., Southerland, A., Tateno, M., Wang, H., Wang, Y., Yagi, T., Yan, M., Yang, X., Yu, Z., Zha, S.C., Zhao, C., Zhao, Z., Zhu, Z., Zhuo, J., Arbelaez, P., Bertasius, G., Crandall, D., Damen, D., Engel, J., Farinella, G.M., Furnari, A., Ghanem, B., Hoffman, J., Jawahar, C.V., Newcombe, R., Park, H.S., Rehg, J.M., Sato, Y., Savva, M., Shi, J., Shou, M.Z., Wray, M.: Ego-exo4d: Understanding skilled human activity from first- and third-person perspectives (2023) +16. Harashima, J., Ariga, M., Murata, K., Ioki, M.: A large-scale recipe and meal data collection as infrastructure for food research. In: International Conference on Language Resources and Evaluation. pp. 2455-2459 (2016) +17. Huang, D.A., Lim, J.J., Fei-Fei, L., Niebles, J.C.: Unsupervised visual-linguistic reference resolution in instructional videos. In: CVPR. pp. 2183-2192 (2017) +18. Jermsurawong, J., Habash, N.: Predicting the structure of cooking recipes. In: Conference on Empirical Methods in Natural Language Processing. pp. 781-786 (2015) +19. Ji, W., Li, Y., Wei, M., Shang, X., Xiao, J., Ren, T., Chua, T.S.: VidVRD 2021: The third grand challenge on video relation detection. In: ACM MM. p. 4779-4783 (2021) +20. Johnson, J., Karpathy, A., Fei-Fei, L.: DenseCap: Fully convolutional localization networks for dense captioning. In: CVPR. pp. 4565-4574 (2016) +21. Kiddon, C., Ponnuraj, G.T., Zettlemoyer, L., Choi, Y.: Mise en place: Unsupervised interpretation of instructional recipes. In: Conference on Empirical Methods in Natural Language Processing. pp. 982-992 (2015) +22. Kuehne, H., Arslan, A., Serre, T.: The language of actions: Recovering the syntax and semantics of goal-directed human activities. In: CVPR. pp. 780-787 (2014) + +23. Lei, J., Wang, L., Shen, Y., Yu, D., Berg, T., Bansal, M.: MART: memory-augmented recurrent transformer for coherent video paragraph captioning. In: Annual Meeting of the Association for Computational Linguistics. pp. 2603-2614 (2020) +24. Lin, K.Q., Wang, J., Soldan, M., Wray, M., Yan, R., Xu, Z., Gao, D., Tu, R.C., Zhao, W., Kong, W., Cai, C., HongFa, W., Damen, D., Ghanem, B., Liu, W., Shou, M.Z.: Egocentric video-language pretraining. In: NeurIPS (2022) +25. Luo, H., Ji, L., Shi, B., Huang, H., Duan, N., Li, T., Li, J., Bharti, T., Zhou, M.: UniVL: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353 (2020) +26. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: CLIP4Clip: An empirical study of CLIP for end to end video clip retrieval. Neurocomputing 508, 293-304 (2022) +27. Ma, Y., Xu, G., Sun, X., Yan, M., Zhang, J., Ji, R.: X-CLIP: End-to-end multi-grained contrastive learning for video-text retrieval. In: ACM MM. pp. 638-647 (2022) +28. Ma, Y., Hiraoka, T., Okazaki, N.: Joint entity and relation extraction based on table labeling using convolutional neural networks. In: Workshop on Structured Prediction for NLP. pp. 11-21 (2022) +29. Miech, A., Alayrac, J.B., Smaira, L., Laptev, I., Sivic, J., Zisserman, A.: End-to-end learning of visual Representations from uncurated instructional videos. In: CVPR (2020) +30. Miech, A., Zhukov, D., Alayrac, J.B., Tapaswi, M., Laptev, I., Sivic, J.: HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In: ICCV (2019) +31. Momouchi, Y.: Control structures for actions in procedural texts and PT-chart. In: International Conference on Computational Linguistics. pp. 108-114 (1980) +32. Mori, S., Maeta, H., Yamakata, Y., Sasada, T.: Flow graph corpus from recipe texts. In: International Conference on Language Resources and Evaluation. pp. 2370-2377 (2014) +33. Nakamura, K., Ohashi, H., Okada, M.: Sensor-augmented egocentric-video captioning with dynamic modal attention. In: ACM MM. p. 4220-4229 (2021) +34. Nishimura, T., Hashimoto, A., Mori, S.: Procedural text generation from a photo sequence. In: International Natural Language Generation Conference. pp. 409-414 (2019) +35. Nishimura, T., Hashimoto, A., Ushiku, Y., Kameko, H., Mori, S.: State-aware video procedural captioning. In: ACM MM. pp. 1766-1774 (2021) +36. Nishimura, T., Hashimoto, A., Ushiku, Y., Kameko, H., Yamakata, Y., Mori, S.: Structure-aware procedural text generation from an image sequence. IEEE Access 9, 2125-2141 (2020) +37. Nishimura, T., Sakoda, K., Hashimoto, A., Ushiku, Y., Tanaka, N., Ono, F., Kameko, H., Mori, S.: Egocentric biochemical video-and-language dataset. In: ICCV Workshop. pp. 3129-3133 (October 2021) +38. Nishimura, T., Sakoda, K., Ushiku, A., Hashimoto, A., Okuda, N., Ono, F., Kameko, H., Mori, S.: BioVL2 dataset: Egocentric biochemical video-and-language dataset. Journal of Natural Language Processing (2022) +39. Schiele, B.: A database for fine grained activity detection of cooking activities. In: CVPR. p. 1194-1201 (2012) +40. Sener, F., Chatterjee, D., Shelepov, D., He, K., Singhania, D., Wang, R., Yao, A.: Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In: CVPR. pp. 21064-21074 (2022) + +41. Shi, B., Ji, L., Liang, Y., Duan, N., Chen, P., Niu, Z., Zhou, M.: Dense procedure captioning in narrated instructional videos. In: Annual Meeting of the Association for Computational Linguistics. pp. 6382-6391 (2019) +42. Shi, B., Ji, L., Niu, Z., Duan, N., Zhou, M., Chen, X.: Learning semantic concepts and temporal alignment for narrated video procedural captioning. In: ACM MM. pp. 4355-4363 (2020) +43. Shimada, A., Kondo, K., Deguchi, D., Morin, G., Stern, H.: Kitchen scene context based gesture recognition: A contest in ICPR2012. In: International Workshop on Advances in Depth Image Analysis and Applications. pp. 168-185 (2013). https://doi.org/10.1007/978-3-642-40303-3_18 +44. Shirai, K., Hashimoto, A., Nishimura, T., Kameko, H., Kurita, S., Ushiku, Y., Mori, S.: Visual Recipe Flow: A dataset for learning visual state changes of objects with recipe flows. In: International Conference on Computational Linguistics. pp. 3570-3577 (2022) +45. Spriggs, E.H., De La Torre, F., Hebert, M.: Temporal segmentation and activity classification from first-person sensing. In: CVPRW. pp. 17-24 (2009) +46. Stein, S., McKenna, S.J.: Combining embedded accelerometers with computer vision for recognizing food preparation activities. In: ACM international joint conference on Pervasive and ubiquitous computing. p. 729-738 (2013) +47. Tang, Y., Ding, D., Rao, Y., Zheng, Y., Zhang, D., Zhao, L., Lu, J., Zhou, J.: COIN: A large-scale dataset for comprehensive instructional video analysis. In: CVPR (2019) +48. Vedantam, R., Zitnick, C.L., Parikh, D.: CIDEr: consensus-based image description evaluation. In: CVPR. pp. 4566-4575 (2015) +49. Wang, T., Zhang, R., Lu, Z., Zheng, F., Cheng, R., Luo, P.: End-to-end dense video captioning with parallel decoding. In: ICCV. pp. 6847-6857 (2021) +50. Wu, J., Pan, L., Chen, J., Jiang, Y.G.: Ingredient-enriched recipe generation from cooking videos. In: ACM International Conference on Multimedia Retrieval. pp. 249-257 (2022) +51. Yagi, T., Ohashi, M., Huang, Y., Furuta, R., Adachi, S., Mitsuyama, T., Sato, Y.: FineBio: A fine-grained video dataset of biological experiments with hierarchical annotation. arXiv preprint arXiv:2402.00293 (2024) +52. Yamakata, Y., Mori, S., Carroll, J.: English recipe flow graph corpus. In: International Conference on Language Resources and Evaluation. pp. 5187-5194 (2020) +53. Yang, A., Nagrani, A., Seo, P.H., Miech, A., Pont-Tuset, J., Laptev, I., Sivic, J., Schmid, C.: Vid2Seq: Large-scale pretraining of a visual language model for dense video captioning. In: CVPR (2023) +54. Yang, J., Peng, W., Li, X., Guo, Z., Chen, L., Li, B., Ma, Z., Zhou, K., Zhang, W., Loy, C.C., Liu, Z.: Panoptic video scene graph generation. In: CVPR. pp. 18675-18685 (2023) +55. Zellers, R., Lu, J., Lu, X., Yu, Y., Zhao, Y., Salehi, M., Kusupati, A., Hessel, J., Farhadi, A., Choi, Y.: MERLOT Reserve: Neural script knowledge through vision and language and sound. In: CVPR. pp. 16375-16387 (June 2022) +56. Zhou, L., Kalantidis, Y., Chen, X., Corso, J.J., Rohrbach, M.: Grounded video description. In: CVPR. pp. 6578-6587 (2019) +57. Zhou, L., Xu, C., Corso, J.J.: Towards automatic learning of procedures from web instructional videos. In: AAAI (2018) +58. Zhukov, D., Alayrac, J.B., Cinbis, R.G., Fouhey, D., Laptev, I., Sivic, J.: Cross-task weakly supervised learning from instructional videos. In: CVPR. pp. 3537-3545 (2019) \ No newline at end of file diff --git a/2024/COM Kitchens_ An Unedited Overhead-view Procedural Videos Dataset a Vision-Language Benchmark/images.zip b/2024/COM Kitchens_ An Unedited Overhead-view Procedural Videos Dataset a Vision-Language Benchmark/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8e7e7bb6759ddda0932474c168cf9c47eb5da37e --- /dev/null +++ b/2024/COM Kitchens_ An Unedited Overhead-view Procedural Videos Dataset a Vision-Language Benchmark/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fedd910caa23ddc5913147b01ea83e52a58fccf212af257475d0604f2a99b059 +size 575169 diff --git a/2024/COM Kitchens_ An Unedited Overhead-view Procedural Videos Dataset a Vision-Language Benchmark/layout.json b/2024/COM Kitchens_ An Unedited Overhead-view Procedural Videos Dataset a Vision-Language Benchmark/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..eff5894bff0069af7b9113a03154b0f34940f936 --- /dev/null +++ b/2024/COM Kitchens_ An Unedited Overhead-view Procedural Videos Dataset a Vision-Language Benchmark/layout.json @@ -0,0 +1,9291 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 112, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 112, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 138, + 112, + 477, + 148 + ], + "type": "text", + "content": "COM Kitchens: An Unedited Overhead-view Video Dataset as a Vision-Language Benchmark" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 167, + 482, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 482, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 482, + 193 + ], + "type": "text", + "content": "Koki Maeda\\*1,2, Tosho Hirasawa\\*1,3, Atsushi Hashimoto1, Jun Harashima4, Leszek Rybicki4, Yusuke Fukasawa4, and Yoshitaka Ushiku1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 173, + 201, + 438, + 246 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 173, + 201, + 438, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 201, + 438, + 213 + ], + "spans": [ + { + "bbox": [ + 173, + 201, + 438, + 213 + ], + "type": "text", + "content": "1 OMRON SINIC X Corp., Tokyo, Japan (*equally contributed)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 210, + 213, + 402, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 213, + 402, + 224 + ], + "spans": [ + { + "bbox": [ + 210, + 213, + 402, + 224 + ], + "type": "text", + "content": "2 Tokyo Institute of Technology, Tokyo, Japan" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 208, + 224, + 405, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 224, + 405, + 236 + ], + "spans": [ + { + "bbox": [ + 208, + 224, + 405, + 236 + ], + "type": "text", + "content": "3 Tokyo Metropolitan University, Tokyo, Japan," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 235, + 236, + 377, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 236, + 377, + 246 + ], + "spans": [ + { + "bbox": [ + 235, + 236, + 377, + 246 + ], + "type": "text", + "content": "4 Cookpad Inc., Yokohama, Japan" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 136, + 262, + 478, + 373 + ], + "blocks": [ + { + "bbox": [ + 136, + 262, + 478, + 373 + ], + "lines": [ + { + "bbox": [ + 136, + 262, + 478, + 373 + ], + "spans": [ + { + "bbox": [ + 136, + 262, + 478, + 373 + ], + "type": "image", + "image_path": "406e9df3490a685c380c25a907cb86b5f105393aacd0c2b06230c983494745c7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 384, + 482, + 429 + ], + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 429 + ], + "type": "text", + "content": "Fig. 1: Sample from COM Kitchens: the dataset includes unedited overhead-view cooking videos, each manually annotated with a visual action graph that links instructional texts to visual elements through edges from before (□) to after (□) bounding boxes (BBs). dist. BBs (□) represent mixing. Details are provided in Sec. 3.2." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 445, + 455, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 445, + 455, + 666 + ], + "spans": [ + { + "bbox": [ + 159, + 445, + 455, + 666 + ], + "type": "text", + "content": "Abstract. Procedural video understanding is gaining attention in the vision and language community. Deep learning-based video analysis requires extensive data. Consequently, existing works often use web videos as training resources, making it challenging to query instructional contents from raw video observations. To address this issue, we propose a new dataset, COM Kitchens. The dataset consists of unedited overhead-view videos captured by smartphones, in which participants performed food preparation based on given recipes. Fixed-viewpoint video datasets often lack environmental diversity due to high camera setup costs. We used modern wide-angle smartphone lenses to cover cooking counters from sink to cooktop in an overhead view, capturing activity without in-person assistance. With this setup, we collected a diverse dataset by distributing smartphones to participants. With this dataset, we propose the novel video-to-text retrieval task Online Recipe Retrieval (OnRR) and new video captioning domain Dense Video Captioning on unedited Overhead-View videos (DVC-OV). Our experiments verified the capabilities and limitations of current web-video-based SOTA methods in handling these tasks. The dataset and code are available at https://doi.org/10.32130/rdata.6.1 and https://github.com/omron-sinicx/com_kitchens, respectively." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 284 + ], + "type": "text", + "content": "Creating higher-value products from raw materials is fundamental to material wealth in society. Understanding these processes through video observation has gained increased attention from the computer vision community [3, 6, 17, 22, 37, 40]. Recent advancements in procedural video analysis rely on large-scale datasets collected from the Web [30, 57, 58] or through ego-vision [6, 11, 14, 37]. These datasets play a crucial role in developing robust techniques and pre-trained models [1, 9, 24-26, 29]. Fixed-viewpoint observation was a major video format before the deep learning era [22, 43, 46]. However, it has received little attention in the last decade because it is hard to find videos with such a format on the Web. A cold-start problem exists: sufficient data is prerequisite to implement a raw-video-based retrieval system. Once available, the system encourages users to upload unedited videos for querying, reinforcing the dataset size." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 285, + 482, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 404 + ], + "type": "text", + "content": "To overcome this cold-start problem, we revisit unedited video datasets with fixed viewpoints (FV). To efficiently enhance environmental diversity, we leveraged recent hardware advancements. Modern smartphones are equipped with well-calibrated wide-angle cameras, allowing us to capture large work areas from an overhead view. This way, we can observe processes with minimal occlusion regardless of kitchen layout, making it preferable for practical applications. Additionally, people are familiar with smartphone UIs, allowing us to collect videos without in-person assistance. To prove this concept of new-style dataset development, we scaled the FV procedural video datasets in this way, resulting in 145 videos, which are 40 hours in total footage, as the first collection." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 405, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 482, + 548 + ], + "type": "text", + "content": "Revisiting FV videos in this deep learning era offers challenges in understanding long procedural contexts rather than frame-wise image processing. To tackle this, we provide a new dataset, COM Kitchens, with a manually annotated visual action graph [44], linking visual events and text instructions with a workflow graph (Fig. 1). Using this structured annotation, we introduce a novel video2text retrieval task, online cross-modal recipe retrieval (OnRR), and a new domain for video captioning, dense video captioning on unedited overhead-view videos (DVC-OV). OnRR is an online cross-modal task for retrieving corresponding recipes during cooking, designed to develop practical smartphone applications. DVC-OV is an offline cross-modal task that generates instructional text from demonstrations, intended to analyze the domain gap between web- and overhead-view videos through a traditional format of the DVC task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 147, + 548, + 335, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 548, + 335, + 560 + ], + "spans": [ + { + "bbox": [ + 147, + 548, + 335, + 560 + ], + "type": "text", + "content": "The contribution of this paper is four-fold." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 568, + 480, + 664 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 136, + 568, + 479, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 568, + 479, + 591 + ], + "spans": [ + { + "bbox": [ + 136, + 568, + 479, + 591 + ], + "type": "text", + "content": "1. We introduce a novel approach to construct datasets of unedited fixed-view videos in diverse environments, leveraging modern smartphones." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 593, + 479, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 593, + 479, + 615 + ], + "spans": [ + { + "bbox": [ + 136, + 593, + 479, + 615 + ], + "type": "text", + "content": "2. We provide visual action graph annotations for the first time on unedited videos." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 135, + 616, + 479, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 616, + 479, + 639 + ], + "spans": [ + { + "bbox": [ + 135, + 616, + 479, + 639 + ], + "type": "text", + "content": "3. We propose a novel task of online recipe retrieval, including its target recipe pool and baselines." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 135, + 641, + 480, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 641, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 135, + 641, + 480, + 664 + ], + "type": "text", + "content": "4. We analyzed the SOTA dense video captioning method on unedited overhead-view videos to reveal future challenges." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "K. Maeda et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 479, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 479, + 178 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 479, + 178 + ], + "type": "text", + "content": "This paper proposes a new vision-language video dataset with one novel tasks and one novel domain for DVC. We present comparisons to clarify the dataset novelty in Sec. 2.1, and the task/setting novelties in Sec. 2.2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 198, + 392, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 198, + 392, + 211 + ], + "spans": [ + { + "bbox": [ + 131, + 198, + 392, + 211 + ], + "type": "text", + "content": "2.1 Datasets for Procedural Video Understanding" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 135, + 286, + 478, + 376 + ], + "blocks": [ + { + "bbox": [ + 130, + 243, + 480, + 277 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 480, + 277 + ], + "type": "text", + "content": "Table 1: Comparison between instructional video datasets with fixed-viewpoint (FV) cameras. We categorized temporal segment type into action (e.g., \"put a bowl,\" \"crack the egg,\" \"beat the egg\"), and step, a higher-level action (e.g., \"whisk eggs in a bowl\")." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 286, + 478, + 376 + ], + "lines": [ + { + "bbox": [ + 135, + 286, + 478, + 376 + ], + "spans": [ + { + "bbox": [ + 135, + 286, + 478, + 376 + ], + "type": "table", + "html": "
datasetyeartopictasks#env.#videos total (h)avg. (m)seg. typeseg. description
MMAC [45]2008Cooking1132815.0action130 actions cls.
MPII [39]2012Cooking14144813.4action65 actions cls.
ACE [43]2012Cooking513523.6action8 actions cls.
50 salads [46]2013Cooking215055.4action51 actions cls.
Breakfast [22]2014Cooking-181,712772.7action10 actions cls.
IKEA ASM [3]2021Furniture45371355.7actionnoun+verb (n+v)
Assembly101 [40]2022Assembly1514,3215137.1act./step1,380 act. cls./n+v
COM KitchensOursCooking139701454016.6act./stepinstructional text
", + "image_path": "14b6a8a64bd0f4ead186e3f2921f2b05a99d1510cf62956f2feb73cb8f571f7d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 401, + 480, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 480, + 460 + ], + "type": "text", + "content": "Tab. 1 summarizes datasets with FV procedural videos. They all target manufacturing tasks. Among these datasets, COM Kitchens has a significantly large diversity in tasks and environments and is the only one with linguistic annotations. We omitted the EgoExo4D [15] dataset from the table because the paper does not provide organized statistics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 461, + 480, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 480, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 480, + 604 + ], + "type": "text", + "content": "Breakfast [22], EgoExo4D [15], and our dataset have environmental diversity, while others struggle with the cost of in-person technical support for their setup. The Breakfast dataset collected data from 18 environments (overhead or side view), but its tasks are limited to two salad recipes with only 10 action classes. EgoExo4D addressed the high cost by collaborating with 12 institutes. Our dataset is competitive with EgoExo4D in terms of the number of environments and participants for the cooking scenario but offers more variety in tasks (i.e., recipes). Another difference is camera views. EgoExo4D used multiple cameras from the front, side, and back view angles, aiming to capture activities beyond the kitchen counter to bridge the gap between egocentric and egocentric views. In contrast, our setup captures detailed food manipulations at the counter with minimal occlusions from the overhead view." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Tab. 2 compares our dataset to other procedural video datasets with linguistic annotations. We omit video datasets without manual annotations, such as HowTo100M [30] and YT-Temporal-1B [55], as they are for pre-training, not for downstream tasks. The pioneering work of YouCookII [57] provides linguistic annotations of coarse instructions (e.g., \"whisk egg, flour\"), and many other works" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "type": "text", + "content": "COM Kitchens" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 179, + 479, + 293 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 2: Comparison between instructional vision-language video datasets. Only our dataset uses a fixed viewpoint. Coarse instruction minimally describes steps with verb(s) and noun(s) in a YouCookII style [57] (e.g., \"tier dried squid\"), whereas fine instruction comes from real instructional texts (e.g., \"tear dried squid by hand into small pieces\"). 'manual*' annotates only the start of intervals." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 179, + 479, + 293 + ], + "lines": [ + { + "bbox": [ + 135, + 179, + 479, + 293 + ], + "spans": [ + { + "bbox": [ + 135, + 179, + 479, + 293 + ], + "type": "table", + "html": "
datasetyeartypetopictasks#videos total (h) avg. (m)seg. descriptioninterval# seg.
YouCookII [57]2018WebCook.892,0001765.3 coarse instruction manual4,325
ProceL [10]2019WebMulti.12720473.9 coarse instruction manual498
COIN [47]2019WebMulti.18011,8274762.4 coarse instruction manual46,354
CrossTask [58]2019WebMulti.834,7003764.8 coarse instruction manual19,278
MMAC-Captions [33]2021EgoCook.51461613.4 coarse instruction manual5,002
Epic Kitchens [7]2022EgoCook.707001008.6 narrationutterance39,596
Ego4D [14]2022EgoOpen--3,670- narrationmanual*-
BioVL2 [38]2022EgoBio.53235.3 fine instruction manual408
VRF [44]2022WebCook.20020020.7 fine instruction manual3,705
FineBio [51]2024EgoBio.722614.53.9 fine instruction manual3,541
COM KitchensOursFVCook.1391454016.6 fine instruction manual2,852
", + "image_path": "b0286c6dbe201fca0aee94f548ed629b485ea691239911818b7266112f0615f8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 316, + 479, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 479, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 479, + 399 + ], + "type": "text", + "content": "follow this manner. Epic Kitchens and Ego4D have narrations as their linguistic annotation, but they tend to describe the details of each action (e.g., action name and target objects), which is still similar to coarse instruction. While captioning for procedural image sequences [4,34,36] uses commercial recipe sites as the dataset resource for generating fine instructions (e.g., \"Beat the egg whites with a mixer, starting on low speed\"), there is a gap with current video datasets' linguistic resource." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 401, + 480, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 480, + 496 + ], + "type": "text", + "content": "BioVL2 [37], FineBio [51], and VRF [44] are datasets with fine instructions, as this work. BioVL2 is a rare dataset capturing biochemical experiments, but it is limited in size. FineBio is a five times larger than BioVL2 in total footage, but its task variation is still limited and the average footage is shorter than general cooking tasks. VRF collected one-minute videos focused on food state changes, excluding human actions. COM Kitchens has " + }, + { + "bbox": [ + 130, + 401, + 480, + 496 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 130, + 401, + 480, + 496 + ], + "type": "text", + "content": " of the diversity against the web-based VRF dataset, consisting of unedited videos with 16.6 minutes on average." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 480, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 480, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 480, + 605 + ], + "type": "text", + "content": "Our dataset provides its annotation as visual action graphs. This structured representation of manufacturing instruction was first proposed in 1980 by Momouchi [31] as an estimation target for a natural language processing task. Later, Flow graph [32,52] was proposed with a fine-grained graph as a comprehensive representation of understanding. Merging tree was the simplest structure of the manufacturing process [18], which selects only actions that merge materials as nodes [18]. Action graph is an intermediate representation between flow graph and merging tree, represents both actions on single materials and merging actions [21]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "These structures have been independently extended to vision-language setups. Visual action graphs were defined as a prediction target in an unsupervised task [17] using web videos. Unfortunately, their visual action graph dataset is only for testing and is not publicly available. A small dataset of visual merging trees has been provided in [36] for semi-supervised learning, where visual" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "K. Maeda et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "data consists of image sequences instead of videos. VRF [44] provides visual flow graphs with videos, but the videos are only one minute long. COM Kitchens is the only dataset that annotates graph structures for action segments. Note that the graph represents dynamic changes across frames, including merging and splitting processes. This point is an essential difference from video scene graph datasets [19, 54]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 205, + 381, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 205, + 381, + 219 + ], + "spans": [ + { + "bbox": [ + 131, + 205, + 381, + 219 + ], + "type": "text", + "content": "2.2 Video-Text Retrieval and Video Captioning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 226, + 482, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 322 + ], + "type": "text", + "content": "Retrieval is one of the fundamental tasks for cross-modal problems. For videos, video-text retrieval [5,25-27], a task to find the video whose entire contents fit a text query, is the principal retrieval task. The literature often evaluates the video-to-text scenario in addition to the text-to-video condition; web videos usually have text metadata, making video-to-text evaluation less practical. Instead, the OnRR task assumes retrieving web content by its text from raw video observation. Since our dataset is sourced from smartphones, developed techniques should be directly applicable to smartphone videos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 322, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 482, + 453 + ], + "type": "text", + "content": "Video paragraph captioning is a video captioning task specially designed for procedural videos; it assumes that event segments are given [23,35,42,50]. However, for unedited videos, it is not practical to assume such given event segments. Hence, we focus on dense video captioning (DVC) [20, 56], a joint task of event detection and event description generation, as a fundamental task of video comprehension. The recent main challenge of DVC is to suppress redundant detection [12]. To overcome this problem, \"detect-then-describe\" [8,41] and \"describe-then-detect\" [49] approaches have been studied. The current SOTA method of Vid2Seq [53] describes captions and detects events in one stage by outputting segments as a time token with its caption. Its performance is supported by the largest video dataset of YT-Temporal-1B [55]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 453, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 574 + ], + "type": "text", + "content": "We test our dataset with Vid2Seq to investigate the domain gap between Web and FV videos. The most significant gap is in video length and repetitive actions. Tab. 2 shows that videos in COM Kitchens are about three times longer than usual web videos on average (and 20 times longer than VRF, which are the TikTok style). This difference is mainly due to repetitive actions, often eliminated in web videos. Another gap is the location of event-related objects within each frame. In web videos (or ego vision), the object of interest tends to be in the center of the frame, whereas FV videos do not dynamically focus on objects. These repetitions and lack of focus on important objects provide additional challenges for aligning linguistic instructions to video events." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 592, + 321, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 592, + 321, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 592, + 321, + 605 + ], + "type": "text", + "content": "3 The COM Kitchens Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Data collection and annotation are pivotal in determining the dataset's utility. This section presents our data collection approach (Sec. 3.1), a detailed definition of the visual action graph (Sec. 3.2), annotation procedure (Sec. 3.3), and dataset statistics (Sec. 3.4)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 385, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 385, + 91, + 447, + 101 + ], + "type": "text", + "content": "COM Kitchens" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 465, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 465, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 465, + 128 + ], + "type": "text", + "content": "3.1 Collecting Videos with their Corresponding Procedural Text" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 243 + ], + "type": "text", + "content": "We selected candidate recipes for filming the cooking process from the Cookpad Recipe Dataset (CRD) [16]. The CRD is a comprehensive Japanese home cuisine database, including over 1.7 million recipes with ingredient lists. In the selection process, we prioritized recipes estimated to take under 30 minutes to prepare and to have a moderate complexity level. We excluded recipes using off-counter facilities (e.g., microwaves) because such actions would be recognizable with sound-based event identification, which is not our focus. The selected recipes are typically but not limited to European, Chinese, and Japanese cuisines (with a certain level of localization)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 244, + 482, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 244, + 482, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 482, + 351 + ], + "type": "text", + "content": "To collect videos, we hired 110 participants from August to October 2021. Each participant recorded up to four recipes at home. Instead of in-person guidance, we provided an instructional document and video that directed the way of recording (e.g., cooking alone, where to capture, camera mode, and turning off TVs) and how to exclude privacy-related information (e.g., do not have a conversation with family, set the camera view not to capture the actor's face, and remove any private documents from camera view). Participants signed a consent form acknowledging that the data would be used publicly for academic purpose. Participants received appropriate financial compensation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 352, + 482, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 423 + ], + "type": "text", + "content": "We utilized the iPhone 11 Pro fixed to a tripod and recorded videos using the rear camera. We instructed participants to set the camera to 30 fps, Full HD resolution of " + }, + { + "bbox": [ + 130, + 352, + 482, + 423 + ], + "type": "inline_equation", + "content": "1920 \\times 1080" + }, + { + "bbox": [ + 130, + 352, + 482, + 423 + ], + "type": "text", + "content": ", and ultra-wide field of view mode (equivalent to a focal length of " + }, + { + "bbox": [ + 130, + 352, + 482, + 423 + ], + "type": "inline_equation", + "content": "13~\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 352, + 482, + 423 + ], + "type": "text", + "content": "). Despite the ultra-wide angle setting, there was almost no distortion; therefore, we applied no extra intrinsic calibration. Appendix B provides filmed video examples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 424, + 482, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 507 + ], + "type": "text", + "content": "Out of the 410 videos recorded, we excluded any that ignored the instructions, curating a total of 210 videos. See Appendix C for detailed statistics. The most common reason for rejection was wrong observation areas. The second was unintentionally observed faces of actors. We made our best effort to avoid involving private information, considering that participants are not professionals and the environment is personal. This emphasis on privacy made data collection more challenging." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 508, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 482, + 579 + ], + "type": "text", + "content": "Among the 210 available videos, we have annotated 145 videos, which amount to a total of 40 hours. We will publish the remaining 85 videos as an unsupervised resource. Two videos were recorded at 60fps; however, we retained them in the dataset as they exhibited no quality issues. In the experiment, we used videos scaled to " + }, + { + "bbox": [ + 130, + 508, + 482, + 579 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 130, + 508, + 482, + 579 + ], + "type": "text", + "content": " resolution, but we will release the original resolution together publicly." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 597, + 416, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 416, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 416, + 609 + ], + "type": "text", + "content": "3.2 Grounding Text to Video via Visual Action Graph" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 482, + 666 + ], + "type": "text", + "content": "A visual action graph delineates intentional actions termed action-by-person (AP) into a temporal graph format. Here, AP (a.k.a., action-by-chef [32]) refers to human actions on ingredients, like cutting or mixing, as opposed to natural processes, such as the browning of ingredients, which occur without human" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "K. Maeda et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 479, + 194 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 479, + 194 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 479, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 479, + 194 + ], + "type": "image", + "image_path": "5586dc17f48bea841bb51260c2f09a58cfd67cd75d7b2d1b4f05be2311d12d00.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 204, + 479, + 237 + ], + "lines": [ + { + "bbox": [ + 130, + 204, + 479, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 479, + 237 + ], + "type": "text", + "content": "Fig. 2: A partial view of our visual action graph: AP7 consists of two sub-APs (7-1 and 7-2). All bounding boxes (BBs) mark foods (e.g., the destination BB in AP7-1 is oil heated in AP6-1). The duration is defined by the first and last BBs of the sub-APs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 262, + 479, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 479, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 479, + 285 + ], + "type": "text", + "content": "intervention. We first present the formal definition of the visual action graph, followed by an illustrative example to contextualize its concept." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "text", + "content": "Define " + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "text", + "content": " as the set of videos in our dataset, and each video " + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "inline_equation", + "content": "v\\in \\mathcal{V}" + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "text", + "content": " is a sequence of " + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "inline_equation", + "content": "|v|" + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "text", + "content": " images, denoted " + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "inline_equation", + "content": "v = \\{v^{1},v^{2},\\ldots ,v^{|v|}\\}" + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "text", + "content": ". Each video is paired with a corresponding sequence of procedural text, denoted as " + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "inline_equation", + "content": "t = \\{t^1,t^2,\\dots ,t^{|t|}\\}" + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "text", + "content": ". We define a visual action graph for a pair of video and text as " + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "inline_equation", + "content": "G(v,t) = (A,E)" + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "text", + "content": " symbolizes the set of APs and " + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 286, + 480, + 346 + ], + "type": "text", + "content": " the set of edges, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 346, + 480, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 480, + 381 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 346, + 480, + 381 + ], + "type": "inline_equation", + "content": "a_{k}" + }, + { + "bbox": [ + 130, + 346, + 480, + 381 + ], + "type": "text", + "content": " be the " + }, + { + "bbox": [ + 130, + 346, + 480, + 381 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 346, + 480, + 381 + ], + "type": "text", + "content": "-th AP in the procedural text. We tie " + }, + { + "bbox": [ + 130, + 346, + 480, + 381 + ], + "type": "inline_equation", + "content": "a_{k}" + }, + { + "bbox": [ + 130, + 346, + 480, + 381 + ], + "type": "text", + "content": " to visual content for encompassing the time segment in which the AP is performed and bounding boxes showing movements of objects, which can be denoted as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 258, + 392, + 354, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 392, + 354, + 405 + ], + "spans": [ + { + "bbox": [ + 258, + 392, + 354, + 405 + ], + "type": "interline_equation", + "content": "a _ {k} = \\left(w _ {k}, A _ {k}, s _ {k}, f _ {k}\\right).", + "image_path": "9fee03f8b65ad5fa5f47708ee67ab4a6504565ad340a78579c30fd86828e3de9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "inline_equation", + "content": "w_{k}" + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "text", + "content": " is a word sequence of the AP (e.g., " + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "inline_equation", + "content": "w_{7} =" + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "text", + "content": " \"add the onion and the celery stalk, and\" in Fig. 2). Due to the nature of the videos, " + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "inline_equation", + "content": "a_{k}" + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "text", + "content": " may involve multiple object instances and multiple action instances. " + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "inline_equation", + "content": "A_{k}" + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "text", + "content": " represents them as a set of actions in the video (e.g., AP7-1 and AP7-2 for " + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "inline_equation", + "content": "a_7" + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "text", + "content": "). We denote the " + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "text", + "content": "-th element of " + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "inline_equation", + "content": "a_{k}" + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "inline_equation", + "content": "a_{k}^{(r)} = (b_{k}^{(r),\\mathrm{bef}}, b_{k}^{(r),\\mathrm{aft}}, b_{k}^{(r),\\mathrm{dest}})" + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "text", + "content": " (e.g., " + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "inline_equation", + "content": "a_7^{(1)} = \\mathrm{AP7 - 1}" + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "text", + "content": "), representing bounding boxes of the target ingredient before/after the action (" + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "inline_equation", + "content": "\\square / \\square" + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "text", + "content": "), and the destination (" + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "inline_equation", + "content": "\\square" + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "text", + "content": ") where the ingredient in " + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "inline_equation", + "content": "b_{k}^{(r),\\mathrm{bef}}" + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "text", + "content": " is mixed by the action. Note that " + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "inline_equation", + "content": "b_{k}^{(r),\\mathrm{dest}}" + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "text", + "content": " is not mandatory for " + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "inline_equation", + "content": "a_{k}^{(r)}" + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "inline_equation", + "content": "s_k" + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "inline_equation", + "content": "f_{k}" + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "text", + "content": " denote the start and finish frame of the AP, defined by the first and last bounding box in " + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "inline_equation", + "content": "A_{k}" + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "text", + "content": " (e.g., " + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "inline_equation", + "content": "s_7 = 05:39" + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "inline_equation", + "content": "f_7 = 05:59" + }, + { + "bbox": [ + 130, + 414, + 481, + 540 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "type": "text", + "content": "Nodes in the visual action graph are bounding boxes in APs defined above. Edges " + }, + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "type": "text", + "content": " track ingredients throughout the cooking activity. Here, " + }, + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "type": "text", + "content": " consists of intra-action edges and inter-action edges. Intra-action edges connect " + }, + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "type": "inline_equation", + "content": "b_{k}^{(r),\\mathrm{bef}}" + }, + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "type": "inline_equation", + "content": "b_{k}^{(r),\\mathrm{aft}}" + }, + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "type": "inline_equation", + "content": "(\\square \\rightarrow \\square)" + }, + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "type": "inline_equation", + "content": "b_{k}^{(r),\\mathrm{dest}}" + }, + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "type": "inline_equation", + "content": "b_{k}^{(r),\\mathrm{aft}}" + }, + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "type": "inline_equation", + "content": "(\\square \\rightarrow \\square)" + }, + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "type": "text", + "content": ". They are always within the same action, and we can automatically identify these edges based on the shared action index, " + }, + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "type": "inline_equation", + "content": "(k,r)" + }, + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "type": "text", + "content": ", where these indices are labeled instead of object names for this annotation. In contrast, we manually identified inter-action edges, which is from " + }, + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "type": "inline_equation", + "content": "b_{k}^{(r),\\mathrm{aft}}" + }, + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "type": "inline_equation", + "content": "b_{k'}^{(r'),\\mathrm{bef}/\\mathrm{dest}}" + }, + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "type": "inline_equation", + "content": "(\\square \\rightarrow \\square" + }, + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "type": "inline_equation", + "content": "\\square" + }, + { + "bbox": [ + 130, + 540, + 481, + 642 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 642, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 480, + 665 + ], + "type": "text", + "content": "To automate the intra-action edge annotation, we labeled each bounding box with an action index instead of ingredient names. Thus, the graph has" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "type": "text", + "content": "COM Kitchens" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "no ingredient information in this form. To fix this problem, we manually gave ingredient names as leaf nodes and connected them to the AP that first processes the ingredient (ingredient-action edges), following the VRF dataset [44]. This method naturally represents the composition of ingredients at each node by tracing back the edges to the leaf nodes." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 194, + 304, + 258 + ], + "blocks": [ + { + "bbox": [ + 133, + 194, + 304, + 258 + ], + "lines": [ + { + "bbox": [ + 133, + 194, + 304, + 258 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 304, + 258 + ], + "type": "image", + "image_path": "370d6549dd4a83101dbf094861a482b7fda6e1e95d359c5da479dddacdff7911.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 266, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 289 + ], + "type": "text", + "content": "Fig.3: Distributions of duration; the averages are 16.6 min and 46.7 sec. for videos and APs, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 308, + 194, + 481, + 258 + ], + "blocks": [ + { + "bbox": [ + 308, + 194, + 481, + 258 + ], + "lines": [ + { + "bbox": [ + 308, + 194, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 194, + 481, + 258 + ], + "type": "image", + "image_path": "54dc1907671729e52dfc4d6a6bcb40d58a16a58aaaa0fe28a23c6d2b78b95977.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 325, + 274, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 325, + 274, + 336 + ], + "spans": [ + { + "bbox": [ + 131, + 325, + 274, + 336 + ], + "type": "text", + "content": "3.3 Annotation Procedure" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 341, + 481, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 481, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 481, + 388 + ], + "type": "text", + "content": "The visual action graph defined above was annotated for each video from February to September 2023. A domain specialist annotated the graph to ensure consistency, with one of the authors reviewing it. The annotator completed the task in about 430 hours." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 388, + 481, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 481, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 481, + 484 + ], + "type": "text", + "content": "Initially, the annotator reviewed the videos and revised the procedural instruction texts based on the performance in each video. Despite instructions to follow the recipes faithfully, the participants often deviated due to the complexity of the cooking task. Instead, we used the revised recipes to simulate participants following them. Simultaneously, the annotator identified and tagged the APs within the instructions. All instructions were initially written in Japanese and translated into English by experts. Each AP corresponded directly between the Japanese and English versions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 484, + 481, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 481, + 534 + ], + "type": "text", + "content": "Subsequently, the annotator delineated the start and finish times with bounding boxes " + }, + { + "bbox": [ + 130, + 484, + 481, + 534 + ], + "type": "inline_equation", + "content": "(b_{k}^{(r),\\mathrm{bef}}, b_{k}^{(r),\\mathrm{aft}}, b_{k}^{(r),\\mathrm{dest}})" + }, + { + "bbox": [ + 130, + 484, + 481, + 534 + ], + "type": "text", + "content": " for each action. We used the Computer Vision Annotation Tool (CVAT)5 for this annotation. The video was examined every five frames to identify clear images of the target ingredients." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 534, + 481, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 481, + 594 + ], + "type": "text", + "content": "Finally, the annotator has assigned inter-action edges to the bounding boxes. Following the automatic generation of edges in intra-action edges, they construct visual action graphs. Ingredients and ingredient-action edges have also been annotated together with inter-action edges. A total of 6,826 bounding boxes and 8,061 relationships were annotated to create the visual action graphs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 608, + 257, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 608, + 257, + 620 + ], + "spans": [ + { + "bbox": [ + 131, + 608, + 257, + 620 + ], + "type": "text", + "content": "3.4 Statistical Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 624, + 481, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 624, + 481, + 650 + ], + "spans": [ + { + "bbox": [ + 130, + 624, + 481, + 650 + ], + "type": "text", + "content": "Recording Statistics. We aim to collect a large-scale dataset of fixed-viewpoint videos featuring structured, fine-grained annotations in diverse environments." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "K. Maeda et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 652, + 286, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 652, + 286, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 652, + 286, + 666 + ], + "type": "text", + "content": "5 https://github.com/opencv/cvat" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 304, + 177 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 304, + 177 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 304, + 177 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 304, + 177 + ], + "type": "image", + "image_path": "650c4993bd5e608152636b2b2e120689009989e80bf38fe4511445930f38c593.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 186, + 480, + 209 + ], + "lines": [ + { + "bbox": [ + 131, + 186, + 480, + 209 + ], + "spans": [ + { + "bbox": [ + 131, + 186, + 480, + 209 + ], + "type": "text", + "content": "Fig. 4: Distribution of the length of word sequences; the averages are 87.2 and 13.3 words for recipes and sentences." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 117, + 480, + 177 + ], + "blocks": [ + { + "bbox": [ + 309, + 117, + 480, + 177 + ], + "lines": [ + { + "bbox": [ + 309, + 117, + 480, + 177 + ], + "spans": [ + { + "bbox": [ + 309, + 117, + 480, + 177 + ], + "type": "image", + "image_path": "7d97bc20989f403cc6e3e45e94e113e26cdf81ac8021a8027feef619dfea3084.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 141, + 219, + 293, + 329 + ], + "blocks": [ + { + "bbox": [ + 141, + 219, + 293, + 329 + ], + "lines": [ + { + "bbox": [ + 141, + 219, + 293, + 329 + ], + "spans": [ + { + "bbox": [ + 141, + 219, + 293, + 329 + ], + "type": "image", + "image_path": "9f7a3c8852b8f681475fdeee953034b826f676a736bbef12968fc5d9818f501c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 321, + 219, + 472, + 328 + ], + "blocks": [ + { + "bbox": [ + 321, + 219, + 472, + 328 + ], + "lines": [ + { + "bbox": [ + 321, + 219, + 472, + 328 + ], + "spans": [ + { + "bbox": [ + 321, + 219, + 472, + 328 + ], + "type": "image", + "image_path": "4bd878f0e2bd585b6c8b26e755e26c5894b3d0edace3f45fe1ead4e13dcede3b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 143, + 329, + 293, + 425 + ], + "blocks": [ + { + "bbox": [ + 143, + 329, + 293, + 425 + ], + "lines": [ + { + "bbox": [ + 143, + 329, + 293, + 425 + ], + "spans": [ + { + "bbox": [ + 143, + 329, + 293, + 425 + ], + "type": "image", + "image_path": "caa1e12cee3435d7f03d12f60d7c2b54e63466e6f09ceecd21f29dfc037a16ea.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 429, + 223, + 437 + ], + "lines": [ + { + "bbox": [ + 211, + 429, + 223, + 437 + ], + "spans": [ + { + "bbox": [ + 211, + 429, + 223, + 437 + ], + "type": "text", + "content": "AP" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 322, + 329, + 471, + 426 + ], + "blocks": [ + { + "bbox": [ + 322, + 329, + 471, + 426 + ], + "lines": [ + { + "bbox": [ + 322, + 329, + 471, + 426 + ], + "spans": [ + { + "bbox": [ + 322, + 329, + 471, + 426 + ], + "type": "image", + "image_path": "8741711702f95de7b3e213625a4bfcd4cd3af2df2dc1f040c16484400fe2f48a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 430, + 417, + 439 + ], + "lines": [ + { + "bbox": [ + 376, + 430, + 417, + 439 + ], + "spans": [ + { + "bbox": [ + 376, + 430, + 417, + 439 + ], + "type": "text", + "content": "Ingredients" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 448, + 482, + 481 + ], + "lines": [ + { + "bbox": [ + 131, + 448, + 482, + 481 + ], + "spans": [ + { + "bbox": [ + 131, + 448, + 482, + 481 + ], + "type": "text", + "content": "Fig. 5: Word cloud and frequency plots. The word clouds visualize our target of fine-grained procedure comprehension well. There are 131 types of APs and 1,187 types of ingredients, both in a clear long-tail trend." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 504, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 482, + 612 + ], + "type": "text", + "content": "The COM Kitchens dataset includes 145 videos from 70 kitchens, totaling 40 hours. Kitchens and actors are in one-to-one correspondence; providing 70 unique actors. The average video duration is 16.6 minutes, as illustrated in Fig. 3 (left). " + }, + { + "bbox": [ + 130, + 504, + 482, + 612 + ], + "type": "inline_equation", + "content": "80.0\\%" + }, + { + "bbox": [ + 130, + 504, + 482, + 612 + ], + "type": "text", + "content": " of the video frames are labeled with APs, with a " + }, + { + "bbox": [ + 130, + 504, + 482, + 612 + ], + "type": "inline_equation", + "content": "15.0\\%" + }, + { + "bbox": [ + 130, + 504, + 482, + 612 + ], + "type": "text", + "content": " overlap, which is comparable to other datasets6. On average, each AP lasted 46.7 seconds (Fig. 3 right). Tab. 1 and Tab. 2 show comparisons with similar recorded datasets. Our dataset stands out for its diverse environments among fixed-viewpoint video datasets. The average video length is longest among the datasets with linguistic annotations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 623, + 481, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 623, + 481, + 647 + ], + "spans": [ + { + "bbox": [ + 131, + 623, + 481, + 647 + ], + "type": "text", + "content": "Recipe Data. Turning to the linguistic side, we assigned fine instructions comprising 949 sentences, averaging 6.5 sentences per video. The average word lengths" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "type": "text", + "content": "COM Kitchens" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 321, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 321, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 321, + 666 + ], + "type": "text", + "content": "6 Referred to Tab. 1 of the Assembly101 [40]." + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 152, + 118, + 287, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 118, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 152, + 118, + 287, + 129 + ], + "type": "text", + "content": "Query: Visual Observation" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 152, + 130, + 458, + 182 + ], + "blocks": [ + { + "bbox": [ + 152, + 130, + 458, + 182 + ], + "lines": [ + { + "bbox": [ + 152, + 130, + 458, + 182 + ], + "spans": [ + { + "bbox": [ + 152, + 130, + 458, + 182 + ], + "type": "image", + "image_path": "133b21742564d64607bcf05d24d19ff8a39898b1fec1cb5aee0949415c62d88d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 154, + 194, + 304, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 194, + 304, + 205 + ], + "spans": [ + { + "bbox": [ + 154, + 194, + 304, + 205 + ], + "type": "text", + "content": "Document: Instructional Text" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 152, + 207, + 299, + 327 + ], + "blocks": [ + { + "bbox": [ + 152, + 207, + 299, + 327 + ], + "lines": [ + { + "bbox": [ + 152, + 207, + 299, + 327 + ], + "spans": [ + { + "bbox": [ + 152, + 207, + 299, + 327 + ], + "type": "image", + "image_path": "e3c5e815aa7382dbd2b984b2f922d7668ea339030408aff0c233c3fab62d66df.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 337, + 482, + 370 + ], + "lines": [ + { + "bbox": [ + 130, + 337, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 482, + 370 + ], + "type": "text", + "content": "Fig. 6: Data related to OnRR: The query of OnRR sub-tasks is the first " + }, + { + "bbox": [ + 130, + 337, + 482, + 370 + ], + "type": "inline_equation", + "content": "Z\\%" + }, + { + "bbox": [ + 130, + 337, + 482, + 370 + ], + "type": "text", + "content": " of a video. For feasible recipe retrieval, we added an extra recipe resource to enhance the dataset of retrieval targets alongside our test set." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 207, + 462, + 327 + ], + "blocks": [ + { + "bbox": [ + 307, + 207, + 462, + 327 + ], + "lines": [ + { + "bbox": [ + 307, + 207, + 462, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 462, + 327 + ], + "type": "image", + "image_path": "36eda48772eb29a5f0bde41ae52f2870aaca3bebc08e0fbbae9be73ff2c30185.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 394, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 482, + 491 + ], + "type": "text", + "content": "are 87.2 and 13.3 for recipes and sentences (Fig. 4). The instructions included 131 types and 2,286 distinct APs, averaging 15.8 per recipe and 2.4 per sentence. An average of 1.24 repetitions was found in " + }, + { + "bbox": [ + 130, + 394, + 482, + 491 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 130, + 394, + 482, + 491 + ], + "type": "text", + "content": " of the APs. The top of Fig. 5 shows a wordcloud of AP, featuring terms beyond coarse instructions of other datasets in Tab. 2, including fine-grained cooking-specific words like \"Saute.\" The annotated instructions contained 305 types and 1,187 ingredients, averaging 8.2 per recipe. The bottom graphs illustrate a distinct long-tail trend in AP and ingredients." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 510, + 230, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 230, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 230, + 525 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 536, + 282, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 282, + 549 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 282, + 549 + ], + "type": "text", + "content": "4.1 Online Recipe Retrieval" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Online Recipe Retrieval (OnRR) involves analyzing cooking videos up to the midpoint to determine the recipe type and the cooking stage achieved. Fig. 6 shows data related to this task; retrieving recipe texts that align with the video content up to a time point " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": ", using the video segment up to the last started AP " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " as the search query. This task is more challenging than traditional video-to-text retrieval as it requires both classification of video type and comprehension of the cooking process. We can decompose OnRR into two subtasks: feasible recipe retrieval and recipe stage identification. We assess these subtasks separately because solving them together is challenging for current SOTA methods." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "K. Maeda et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 116, + 481, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 116, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 481, + 175 + ], + "type": "text", + "content": "Feasible recipe retrieval evaluates OnRR from a task categorization perspective in an online setting. We consider a recipe feasible for a task in the middle of the workflow when an actor can shift to the recipe without disposing of foods. This setup assumes an application of online recipe recommendation. Task success is measured by whether the retrieved recipe is in the feasible set." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 177, + 481, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 177, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 134, + 177, + 481, + 236 + ], + "type": "text", + "content": "Recipe stage identification is evaluated with a given recipe. The task is to spot the exact cooking stage in the recipe from the video. It involves aligning the video content with the corresponding part of the recipe text, up to the verb " + }, + { + "bbox": [ + 134, + 177, + 481, + 236 + ], + "type": "inline_equation", + "content": "w_{y}" + }, + { + "bbox": [ + 134, + 177, + 481, + 236 + ], + "type": "text", + "content": " that matches AP " + }, + { + "bbox": [ + 134, + 177, + 481, + 236 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 134, + 177, + 481, + 236 + ], + "type": "text", + "content": ", within a given recipe ID " + }, + { + "bbox": [ + 134, + 177, + 481, + 236 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 134, + 177, + 481, + 236 + ], + "type": "text", + "content": ". This subtask focuses on accurately matching the video content with the specific steps in the recipe." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 248, + 481, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 248, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 134, + 248, + 481, + 380 + ], + "type": "text", + "content": "Settings. In our experiment, " + }, + { + "bbox": [ + 134, + 248, + 481, + 380 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 134, + 248, + 481, + 380 + ], + "type": "text", + "content": " is set at 25%, 50%, 75%, and 100% from the initial point by frame. We labeled them early-, middle-, late-, and full-stage" + }, + { + "bbox": [ + 134, + 248, + 481, + 380 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 134, + 248, + 481, + 380 + ], + "type": "text", + "content": ", respectively. The video set was split into 90/26/29 videos for the train/val/test sets, and the test set was further divided into " + }, + { + "bbox": [ + 134, + 248, + 481, + 380 + ], + "type": "inline_equation", + "content": "29 \\times 4 = 116" + }, + { + "bbox": [ + 134, + 248, + 481, + 380 + ], + "type": "text", + "content": " stages. We compared the performance of state-of-the-art models (UniVL [25], CLIP4Clip [26], and X-CLIP [27]) against random selection. We set the maximum token length, maximum frame length, batch size, and number of training epochs to 48, 48, 32, 5 for UniVL and 32, 12, 32, 3 for CLIP4Clip and X-CLIP. These methods often cannot handle all the words in a recipe at once, as it often exceeds 100 words, as shown in Fig. 4. Hence, we shortened the recipe texts by POS tagging with spaCy, extracting only nouns and verbs to cover all steps with the 100 words." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 393, + 481, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 393, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 134, + 393, + 481, + 536 + ], + "type": "text", + "content": "Extra Recipe Resource. The number of retrieval targets should be sufficiently large; however, the COM Kitchens test set size is limited. Thus, we have increased retrieval targets with an extra recipe resource from CRD. First, we identified the ingredients appearing in the early stage (" + }, + { + "bbox": [ + 134, + 393, + 481, + 536 + ], + "type": "inline_equation", + "content": "Z = 25\\%" + }, + { + "bbox": [ + 134, + 393, + 481, + 536 + ], + "type": "text", + "content": ") by referring to the action graph for each test set video. Based on the early-stage ingredients, we then extracted up to 100 candidate recipes from CRD for each test recipe. This operation collected 1,828 recipes as this extra resource. Among them, we manually discovered " + }, + { + "bbox": [ + 134, + 393, + 481, + 536 + ], + "type": "inline_equation", + "content": "991/243/19/5" + }, + { + "bbox": [ + 134, + 393, + 481, + 536 + ], + "type": "text", + "content": " feasible recipes for each stage " + }, + { + "bbox": [ + 134, + 393, + 481, + 536 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 134, + 393, + 481, + 536 + ], + "type": "text", + "content": ", tied with their corresponding test set recipe. Nonfeasible recipes remained as hard negatives, often overlapping in ingredients with feasible recipes. Note that extracted recipes only include recipe texts and ingredients, not videos. We machine-translated them from Japanese to English8." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 550, + 481, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 550, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 134, + 550, + 481, + 586 + ], + "type": "text", + "content": "Metrics. Following the conventional video-text retrieval task, we employed Recall at rank K (R@K, higher is better) and median rank (MdR, lower is better) to evaluate the retrieval performances at the two subtasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 599, + 481, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 599, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 134, + 599, + 481, + 635 + ], + "type": "text", + "content": "Results. Tab. 3 showcases the result of the OnRR benchmark in the early- and middle-stage settings. Results for the other two stages and no-fine-tune settings are provided in Appendix D. None of the baseline models outperformed" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "type": "text", + "content": "COM Kitchens" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 135, + 643, + 412, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 643, + 412, + 653 + ], + "spans": [ + { + "bbox": [ + 135, + 643, + 412, + 653 + ], + "type": "text", + "content": "The full-stage task is equal to conventional video-to-text retrieval." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 135, + 653, + 422, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 653, + 422, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 653, + 422, + 665 + ], + "type": "text", + "content": "8 We used the DeepL API to translate from Japanese to English (US)." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 168, + 462, + 297 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 3: Online recipe retrieval (OnRR) performances of baseline models. R@K and MdR represent recall at rank K " + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": " and median rank " + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": ", respectively. This table provides only the early- and middel-stage settings (using the first " + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": " of the video as input); results in other stages and settings are detailed in Appendix D." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 168, + 462, + 297 + ], + "lines": [ + { + "bbox": [ + 151, + 168, + 462, + 297 + ], + "spans": [ + { + "bbox": [ + 151, + 168, + 462, + 297 + ], + "type": "table", + "html": "
TaskMethodEarly (25%)Middle (50%)
R@1R@5R@10MdRR@1R@5R@10MdR
Feasible Recipe RetrievalRandom1.88.615.8-0.41.83.1-
UniVL [25]3.45.79.2227.03.45.79.2231.0
CLIP4Clip [26]0.00.010.379.00.00.06.885.0
X-CLIP [27]0.06.810.389.00.03.43.4320.0
Recipe Stage IdentificationRandom6.331.663.38.06.331.663.38.0
UniVL [25]17.248.268.95.09.263.389.23.0
CLIP4Clip [26]6.848.268.95.010.355.186.24.0
X-CLIP [27]10.351.768.94.017.237.993.16.0
", + "image_path": "5558d537a1d770011ff90230689004d0f6dff3833f7486815104f46ac63be02a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 319, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 380 + ], + "type": "text", + "content": "random selection in feasible recipe retrieval. Conversely, all models outperformed random selection in recipe stage identification and improved with fine-tuning. These results suggest that simple contrastive learning cannot solve these two tasks simultaneously. In other words, the OnRR task serves as a benchmark for video-text retrieval based on a procedural comprehension." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 396, + 471, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 471, + 408 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 471, + 408 + ], + "type": "text", + "content": "4.2 Dense Video Captioning on Unedited Fixed-viewpoint Videos" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 415, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 498 + ], + "type": "text", + "content": "Dense video captioning is another task for procedural videos, where a system generates multiple detailed captions for different segments within a video. This approach involves detecting distinct events in the video timeline and then generating descriptive and accurate captions for each identified event. The objective is to provide a more comprehensive and segmented understanding of the video content, which is beneficial for offline applications, such as accessibility, content analysis, and enhanced video search capabilities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 558 + ], + "type": "text", + "content": "For the COM Kitchens dataset, we use APs as the segments of a video; the content is the recipe text of each AP and the entire duration covering all repetitions under each AP. The main challenge of this task is the domain gap between our unedited overhead-view videos and traditional DVC targets such as web videos." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 494, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 494, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 494, + 629 + ], + "type": "text", + "content": "Settings. In this experiment, we used the same train-valid/test split with the OnRR task. We selected two DVC systems as our baselines: PDVC [49] and Vid2Seq [53]. We first tested their zero-shot performance on the COM Kitchens dataset. Then, we fine-tuned the Vid2Seq model, the SOTA model on the YouCookII dataset, on our dataset and evaluated its performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Besides, we examined two supervision approaches to leverage the action graph in the DVC task: (i) action graph as relation labels (RL) and (ii) action graph as attention supervision (AS). When we took the action graph as relation labels," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "K. Maeda et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 179, + 157, + 434, + 253 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 4: Comparison between DVC performances of baseline models. The rows with 'FT' of '" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "' show the results of models fine-tuned on COM Kitchens. \"AG\" shows the choice of methods to leverage action graph information during the fine-tuning." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 179, + 157, + 434, + 253 + ], + "lines": [ + { + "bbox": [ + 179, + 157, + 434, + 253 + ], + "spans": [ + { + "bbox": [ + 179, + 157, + 434, + 253 + ], + "type": "table", + "html": "
ModelFTAGSODA_c(↑)CIDEr(↑)METEOR(↑)
PDVC [49]--0.0220.0000.000
Vid2Seq [53]--0.0170.0660.010
Vid2Seq✓-0.3692.8320.642
Vid2Seq✓RL0.2111.3810.285
Vid2Seq✓AS0.2662.5130.423
Vid2Seq✓RL+AS0.5816.1951.142
", + "image_path": "83afe14133031ae8d0326448f4929c397f03e9331ba599b8d2a2614424ae7c03.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 277, + 479, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 479, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 479, + 336 + ], + "type": "text", + "content": "we employed the decoder of TablERT-CNN [28] as our module to predict the relation (a.k.a. the type of edges) between APs and trained the model in a multitask learning manner. For attention supervision, following Garg et al. [13], we minimize the Kullback-Leibler divergence between the self-attention matrix at the last encoder layer and the alignment matrix indicating the existence of edges." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 350, + 479, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 350, + 479, + 374 + ], + "spans": [ + { + "bbox": [ + 131, + 350, + 479, + 374 + ], + "type": "text", + "content": "Metrics. We employed SODA_c [12], CIDEr [48], and METEOR [2] scores to evaluate model performance, as these are the commonly used for DVC tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 388, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 480, + 472 + ], + "type": "text", + "content": "Results. Tab. 4 shows the result of the DVC benchmark. The zero-shot performances of both the PDVC and Vid2Seq models are extremely low compared to those on the YouCookII dataset (4.9 and 7.9 SODA_c scores, respectively). This deterioration indicates that the COM Kitchens dataset is challenging for the DVC task. While speech information is proved to be a crucial modality to generate better captions [53], videos in the COM Kitchens dataset have no speech information than web-based procedural video datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 473, + 481, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 481, + 568 + ], + "type": "text", + "content": "We also found that the supervision obtained from the action graph benefits the model performance. While RL or AS supervision individually does not bring improvement, their combination enhances model performance across all metrics. This performance gain is attributed to improved attention. As shown in the Fig. 7, RL or AS supervision aligns most frames to a small subset, but the " + }, + { + "bbox": [ + 130, + 473, + 481, + 568 + ], + "type": "inline_equation", + "content": "\\mathrm{RL + AS}" + }, + { + "bbox": [ + 130, + 473, + 481, + 568 + ], + "type": "text", + "content": " supervision aligns frames to their relative frames (defined by the action graph) and their surrounding frames. This smoother alignment aids in generating recipe texts and determining segment boundaries." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 590, + 217, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 217, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 217, + 601 + ], + "type": "text", + "content": "5 Discussion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": "We provided baseline results for two specific tasks: OnRR and DVC-OV, extensions of conventional cross-modal retrieval and captioning tasks. OnRR aims to create practical video comprehension for online applications. In contrast, DVC-OV focuses on understanding overhead-view procedural videos as offline tasks." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "type": "text", + "content": "COM Kitchens" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 212, + 202 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 212, + 202 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 212, + 202 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 212, + 202 + ], + "type": "image", + "image_path": "1783b55b711f38e236e61d9ed97b44880c2a71591ea5c8a4fd8fef1c1beb07a1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 224, + 114, + 302, + 202 + ], + "blocks": [ + { + "bbox": [ + 224, + 114, + 302, + 202 + ], + "lines": [ + { + "bbox": [ + 224, + 114, + 302, + 202 + ], + "spans": [ + { + "bbox": [ + 224, + 114, + 302, + 202 + ], + "type": "image", + "image_path": "2f67ee35a58b681b9a2f8c867aa30938ecb4160bff10626c30c1875cddb11d31.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 313, + 115, + 391, + 202 + ], + "blocks": [ + { + "bbox": [ + 313, + 115, + 391, + 202 + ], + "lines": [ + { + "bbox": [ + 313, + 115, + 391, + 202 + ], + "spans": [ + { + "bbox": [ + 313, + 115, + 391, + 202 + ], + "type": "image", + "image_path": "5c5550d15dd3f120cec54a20bc20232ac7c40a030f134df8c4a0bc91a4a48a09.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 402, + 115, + 480, + 202 + ], + "blocks": [ + { + "bbox": [ + 402, + 115, + 480, + 202 + ], + "lines": [ + { + "bbox": [ + 402, + 115, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 402, + 115, + 480, + 202 + ], + "type": "image", + "image_path": "619a0f1d89c22d3503c120ac56200755286a596de31ab0e0556c4d3fc0689d98.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 134, + 205, + 212, + 283 + ], + "blocks": [ + { + "bbox": [ + 134, + 205, + 212, + 283 + ], + "lines": [ + { + "bbox": [ + 134, + 205, + 212, + 283 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 212, + 283 + ], + "type": "image", + "image_path": "4180d0f75e26967e3b855933c7e5ac3b913e8d4244de67e5b6bf648e4a6ec92a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 292, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 131, + 292, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 131, + 292, + 482, + 316 + ], + "type": "text", + "content": "Fig. 7: Visualization of the attention from the first head in the last encoder layer, based on randomly selected samples. Additional examples can be found in Appendix E." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 224, + 205, + 302, + 283 + ], + "blocks": [ + { + "bbox": [ + 224, + 205, + 302, + 283 + ], + "lines": [ + { + "bbox": [ + 224, + 205, + 302, + 283 + ], + "spans": [ + { + "bbox": [ + 224, + 205, + 302, + 283 + ], + "type": "image", + "image_path": "b1b06dd9ff2bb7632b2a09419d34e518bf78212e8c9fa9c94c5f0cc10bd3a84f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 313, + 205, + 391, + 283 + ], + "blocks": [ + { + "bbox": [ + 313, + 205, + 391, + 283 + ], + "lines": [ + { + "bbox": [ + 313, + 205, + 391, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 391, + 283 + ], + "type": "image", + "image_path": "084df8601ec153e213edc4ebd180f5f7984185b5bf3992e644640cfae4bb414f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 403, + 205, + 479, + 283 + ], + "blocks": [ + { + "bbox": [ + 403, + 205, + 479, + 283 + ], + "lines": [ + { + "bbox": [ + 403, + 205, + 479, + 283 + ], + "spans": [ + { + "bbox": [ + 403, + 205, + 479, + 283 + ], + "type": "image", + "image_path": "d8831163f55968550eb695f6f2ed8681ea83255b4a0298f1d16088f3bb6f4910.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 340, + 479, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 479, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 479, + 435 + ], + "type": "text", + "content": "A potential application involves constructing the visual action graph using video and recipe text inputs. It aims to structurally understand cooking procedures by focusing on temporal dependencies and local human-object interactions. Another possible task is episodic memory in the procedural domain, a linguistic query-driven reasoning task tailored originally for long-duration ego-centric videos [14]. The COM Kitchens dataset is particularly suited for this task, as it comprises long videos (avg. 16.6 min., Tab. 1) and the visual action graph can generate various queries with spatio-temporal ground truth." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 436, + 480, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 480, + 508 + ], + "type": "text", + "content": "Our dataset, similar to many others, is limited by its size. Furthermore, benchmark results indicate that conventional global alignment methods are ineffective for these tasks, prompting us to explore new pre-training and fine-tuning methods. Fortunately, our data collection costs are significantly lower compared to datasets like EgoExo4D. We aim to further explore the potential of visual action graphs and to expand the dataset both with and without supervision." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 529, + 220, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 529, + 220, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 529, + 220, + 542 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "We have introduced COM Kitchens, a dataset that facilitates vision-language understanding with overhead-view recordings, procedural recipe texts, and visual action graphs. This dataset reflects real kitchen conditions, providing rich insights into the sequence of actions and states of ingredients. Our experiments on the OnRR and DVC-OV benchmarks revealed the limitations of existing cross-modal retrieval models in handling long sequences and temporal dependencies. We plan to expand COM Kitchens for further challenges, believing that this dataset will contribute to the advancement of complex video content interpretation." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "K. Maeda et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 482, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 482, + 214 + ], + "type": "text", + "content": "This work would not have been possible without the meticulous and precise annotations by Yasuko Sonoda. We are deeply grateful for her significant contributions to the dataset construction. We also express our gratitude to the IDR office at the National Institute of Informatics for their continuous support in our dataset releases. This work was supported by JSPS KAKENHI Grant Number 21H04910 and JST Moonshot R&D Grant Number JPMJMS2236." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 236, + 197, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 197, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 197, + 248 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 263, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 263, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 263, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 138, + 263, + 480, + 285 + ], + "type": "text", + "content": "1. Ashutosh, K., Girdhar, R., Torresani, L., Grauman, K.: HierVL: Learning hierarchical video-language embeddings. In: CVPR. pp. 23066-23078 (June 2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 287, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 287, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 287, + 480, + 319 + ], + "type": "text", + "content": "2. Banerjee, S., Lavie, A.: METEOR: an automatic metric for MT evaluation with improved correlation with human judgments. In: Proc. ACL Workshop IEEMMTS. pp. 65-72 (2005)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 320, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 480, + 352 + ], + "type": "text", + "content": "3. Ben-Shabat, Y., Yu, X., Saleh, F., Campbell, D., Rodriguez-Opazo, C., Li, H., Gould, S.: The IKEA ASM dataset: Understanding people assembling furniture through actions, objects and pose. In: WACV. pp. 847-859 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 354, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 480, + 385 + ], + "type": "text", + "content": "4. Chandu, K., Nyberg, E., Black, A.W.: Storyboarding of recipes: Grounded contextual generation. In: Annual Meeting of the Association for Computational Linguistics. pp. 6040-6046 (Jul 2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 387, + 480, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 387, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 138, + 387, + 480, + 408 + ], + "type": "text", + "content": "5. Chen, S., Zhao, Y., Jin, Q., Wu, Q.: Fine-grained video-text retrieval with hierarchical graph reasoning. In: CVPR (June 2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 410, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 410, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 138, + 410, + 480, + 441 + ], + "type": "text", + "content": "6. Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., Wray, M.: Scaling egocentric vision: The EPIC-KITCHENS dataset. In: ECCV (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 443, + 480, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 443, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 138, + 443, + 480, + 485 + ], + "type": "text", + "content": "7. Damen, D., Doughty, H., Farinella, G.M., Furnari, A., Ma, J., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., Wray, M.: Rescaling egocentric vision: Collection, pipeline and challenges for EPIC-KITCHENS-100. IJCV 130, 33-55 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 487, + 480, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 480, + 509 + ], + "type": "text", + "content": "8. Deng, C., Chen, S., Chen, D., He, Y., Wu, Q.: Sketch, ground, and refine: Top-down dense video captioning. In: CVPR. pp. 234-243 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 510, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 480, + 542 + ], + "type": "text", + "content": "9. Dvornik, N., Hadji, I., Zhang, R., Derpanis, K.G., Wildes, R.P., Jepson, A.D.: StepFormer: Self-supervised step discovery and localization in instructional videos. In: CVPR. pp. 18952-18961 (June 2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 544, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 564 + ], + "type": "text", + "content": "0. Elhamifar, E., Naing, Z.: Unsupervised procedure learning via joint dynamic summarization. In: ICCV (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 566, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 566, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 566, + 480, + 587 + ], + "type": "text", + "content": "1. Fathi, A., Li, Y., Rehg, J.M.: Learning to recognize daily actions using gaze. In: ECCV. pp. 314-327 (2012)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 588, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 480, + 620 + ], + "type": "text", + "content": "2. Fujita, S., Hirao, T., Kamigaito, H., Okumura, M., Nagata, M.: SODA: Story oriented dense video captioning evaluation framework. In: ECCV. pp. 517-531 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "type": "text", + "content": "3. Garg, S., Peitz, S., Nallasamy, U., Paulik, M.: Jointly learning to align and translate with transformer models. In: Conference on Empirical Methods in Natural Language Processing and International Joint Conference on Natural Language Processing. pp. 4453-4462. Hong Kong, China (2019)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "type": "text", + "content": "COM Kitchens" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 270 + ], + "type": "text", + "content": "14. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., Martin, M., Nagarajan, T., Radosavovic, I., Ramakrishnan, S.K., Ryan, F., Sharma, J., Wray, M., Xu, M., Xu, E.Z., Zhao, C., Bansal, S., Batra, D., Cartillier, V., Crane, S., Do, T., Doulaty, M., Erapalli, A., Feichtenhofer, C., Fragomeni, A., Fu, Q., Gebreselasie, A., Gonzalez, C., Hillis, J., Huang, X., Huang, Y., Jia, W., Khoo, W., Kolář, J., Kottur, S., Kumar, A., Landini, F., Li, C., Li, Y., Li, Z., Mangalam, K., Modhugu, R., Munro, J., Murrell, T., Nishiyasu, T., Price, W., Ruiz, P., Ramazanova, M., Sari, L., Somasundaram, K., Southerland, A., Sugano, Y., Tao, R., Vo, M., Wang, Y., Wu, X., Yagi, T., Zhao, Z., Zhu, Y., Arbeláez, P., Crandall, D., Damen, D., Farinella, G.M., Fuegen, C., Ghanem, B., Ithapu, V.K., Jawahar, C.V., Joo, H., Kitani, K., Li, H., Newcombe, R., Oliva, A., Park, H.S., Rehg, J.M., Sato, Y., Shi, J., Shou, M.Z., Torralba, A., Torresani, L., Yan, M., Malik, J.: Ego4D: Around the world in 3,000 hours of egocentric video. In: CVPR. pp. 18995-19012 (June 2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 272, + 481, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 272, + 481, + 458 + ], + "spans": [ + { + "bbox": [ + 133, + 272, + 481, + 458 + ], + "type": "text", + "content": "15. Grauman, K., Westbury, A., Torresani, L., Kitani, K., Malik, J., Afouras, T., Ashutosh, K., Baiyya, V., Bansal, S., Boote, B., Byrne, E., Chavis, Z., Chen, J., Cheng, F., Chu, F.J., Crane, S., Dasgupta, A., Dong, J., Escobar, M., Forigua, C., Gebreselasie, A., Haresh, S., Huang, J., Islam, M.M., Jain, S., Khirodkar, R., Kukreja, D., Liang, K.J., Liu, J.W., Majumder, S., Mao, Y., Martin, M., Mavroudi, E., Nagarajan, T., Ragusa, F., Ramakrishnan, S.K., Seminara, L., Somayazulu, A., Song, Y., Su, S., Xue, Z., Zhang, E., Zhang, J., Castillo, A., Chen, C., Fu, X., Furuta, R., Gonzalez, C., Gupta, P., Hu, J., Huang, Y., Huang, Y., Khoo, W., Kumar, A., Kuo, R., Lakhavani, S., Liu, M., Luo, M., Luo, Z., Meredith, B., Miller, A., Oguntola, O., Pan, X., Peng, P., Pramanick, S., Ramazanova, M., Ryan, F., Shan, W., Somasundaram, K., Song, C., Southerland, A., Tateno, M., Wang, H., Wang, Y., Yagi, T., Yan, M., Yang, X., Yu, Z., Zha, S.C., Zhao, C., Zhao, Z., Zhu, Z., Zhuo, J., Arbelaez, P., Bertasius, G., Crandall, D., Damen, D., Engel, J., Farinella, G.M., Furnari, A., Ghanem, B., Hoffman, J., Jawahar, C.V., Newcombe, R., Park, H.S., Rehg, J.M., Sato, Y., Savva, M., Shi, J., Shou, M.Z., Wray, M.: Ego-exo4d: Understanding skilled human activity from first- and third-person perspectives (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 460, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 460, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 133, + 460, + 481, + 493 + ], + "type": "text", + "content": "16. Harashima, J., Ariga, M., Murata, K., Ioki, M.: A large-scale recipe and meal data collection as infrastructure for food research. In: International Conference on Language Resources and Evaluation. pp. 2455-2459 (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 494, + 481, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 494, + 481, + 517 + ], + "spans": [ + { + "bbox": [ + 133, + 494, + 481, + 517 + ], + "type": "text", + "content": "17. Huang, D.A., Lim, J.J., Fei-Fei, L., Niebles, J.C.: Unsupervised visual-linguistic reference resolution in instructional videos. In: CVPR. pp. 2183-2192 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 517, + 481, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 517, + 481, + 550 + ], + "spans": [ + { + "bbox": [ + 133, + 517, + 481, + 550 + ], + "type": "text", + "content": "18. Jermsurawong, J., Habash, N.: Predicting the structure of cooking recipes. In: Conference on Empirical Methods in Natural Language Processing. pp. 781-786 (2015)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 552, + 481, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 552, + 481, + 584 + ], + "spans": [ + { + "bbox": [ + 133, + 552, + 481, + 584 + ], + "type": "text", + "content": "19. Ji, W., Li, Y., Wei, M., Shang, X., Xiao, J., Ren, T., Chua, T.S.: VidVRD 2021: The third grand challenge on video relation detection. In: ACM MM. p. 4779-4783 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 586, + 481, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 481, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 481, + 608 + ], + "type": "text", + "content": "20. Johnson, J., Karpathy, A., Fei-Fei, L.: DenseCap: Fully convolutional localization networks for dense captioning. In: CVPR. pp. 4565-4574 (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 609, + 481, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 481, + 642 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 481, + 642 + ], + "type": "text", + "content": "21. Kiddon, C., Ponnuraj, G.T., Zettlemoyer, L., Choi, Y.: Mise en place: Unsupervised interpretation of instructional recipes. In: Conference on Empirical Methods in Natural Language Processing. pp. 982-992 (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "22. Kuehne, H., Arslan, A., Serre, T.: The language of actions: Recovering the syntax and semantics of goal-directed human activities. In: CVPR. pp. 780-787 (2014)" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "K. Maeda et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 159 + ], + "type": "text", + "content": "23. Lei, J., Wang, L., Shen, Y., Yu, D., Berg, T., Bansal, M.: MART: memory-augmented recurrent transformer for coherent video paragraph captioning. In: Annual Meeting of the Association for Computational Linguistics. pp. 2603-2614 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 160, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 481, + 192 + ], + "type": "text", + "content": "24. Lin, K.Q., Wang, J., Soldan, M., Wray, M., Yan, R., Xu, Z., Gao, D., Tu, R.C., Zhao, W., Kong, W., Cai, C., HongFa, W., Damen, D., Ghanem, B., Liu, W., Shou, M.Z.: Egocentric video-language pretraining. In: NeurIPS (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 193, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 193, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 481, + 224 + ], + "type": "text", + "content": "25. Luo, H., Ji, L., Shi, B., Huang, H., Duan, N., Li, T., Li, J., Bharti, T., Zhou, M.: UniVL: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 225, + 481, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 481, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 481, + 256 + ], + "type": "text", + "content": "26. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: CLIP4Clip: An empirical study of CLIP for end to end video clip retrieval. Neurocomputing 508, 293-304 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 258, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 258, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 258, + 481, + 289 + ], + "type": "text", + "content": "27. Ma, Y., Xu, G., Sun, X., Yan, M., Zhang, J., Ji, R.: X-CLIP: End-to-end multi-grained contrastive learning for video-text retrieval. In: ACM MM. pp. 638-647 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 290, + 481, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 290, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 290, + 481, + 322 + ], + "type": "text", + "content": "28. Ma, Y., Hiraoka, T., Okazaki, N.: Joint entity and relation extraction based on table labeling using convolutional neural networks. In: Workshop on Structured Prediction for NLP. pp. 11-21 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 323, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 323, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 323, + 481, + 353 + ], + "type": "text", + "content": "29. Miech, A., Alayrac, J.B., Smaira, L., Laptev, I., Sivic, J., Zisserman, A.: End-to-end learning of visual Representations from uncurated instructional videos. In: CVPR (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 481, + 386 + ], + "type": "text", + "content": "30. Miech, A., Zhukov, D., Alayrac, J.B., Tapaswi, M., Laptev, I., Sivic, J.: HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In: ICCV (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 387, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 481, + 407 + ], + "type": "text", + "content": "31. Momouchi, Y.: Control structures for actions in procedural texts and PT-chart. In: International Conference on Computational Linguistics. pp. 108-114 (1980)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 408, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 481, + 439 + ], + "type": "text", + "content": "32. Mori, S., Maeta, H., Yamakata, Y., Sasada, T.: Flow graph corpus from recipe texts. In: International Conference on Language Resources and Evaluation. pp. 2370-2377 (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 440, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 481, + 460 + ], + "type": "text", + "content": "33. Nakamura, K., Ohashi, H., Okada, M.: Sensor-augmented egocentric-video captioning with dynamic modal attention. In: ACM MM. p. 4220-4229 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 461, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 481, + 493 + ], + "type": "text", + "content": "34. Nishimura, T., Hashimoto, A., Mori, S.: Procedural text generation from a photo sequence. In: International Natural Language Generation Conference. pp. 409-414 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 494, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 481, + 515 + ], + "type": "text", + "content": "35. Nishimura, T., Hashimoto, A., Ushiku, Y., Kameko, H., Mori, S.: State-aware video procedural captioning. In: ACM MM. pp. 1766-1774 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 515, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 481, + 546 + ], + "type": "text", + "content": "36. Nishimura, T., Hashimoto, A., Ushiku, Y., Kameko, H., Yamakata, Y., Mori, S.: Structure-aware procedural text generation from an image sequence. IEEE Access 9, 2125-2141 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 547, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 481, + 579 + ], + "type": "text", + "content": "37. Nishimura, T., Sakoda, K., Hashimoto, A., Ushiku, Y., Tanaka, N., Ono, F., Kameko, H., Mori, S.: Egocentric biochemical video-and-language dataset. In: ICCV Workshop. pp. 3129-3133 (October 2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 580, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 580, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 481, + 611 + ], + "type": "text", + "content": "38. Nishimura, T., Sakoda, K., Ushiku, A., Hashimoto, A., Okuda, N., Ono, F., Kameko, H., Mori, S.: BioVL2 dataset: Egocentric biochemical video-and-language dataset. Journal of Natural Language Processing (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 612, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 481, + 632 + ], + "type": "text", + "content": "39. Schiele, B.: A database for fine grained activity detection of cooking activities. In: CVPR. p. 1194-1201 (2012)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "40. Sener, F., Chatterjee, D., Shelepov, D., He, K., Singhania, D., Wang, R., Yao, A.: Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In: CVPR. pp. 21064-21074 (2022)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 385, + 91, + 447, + 100 + ], + "type": "text", + "content": "COM Kitchens" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 658 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "41. Shi, B., Ji, L., Liang, Y., Duan, N., Chen, P., Niu, Z., Zhou, M.: Dense procedure captioning in narrated instructional videos. In: Annual Meeting of the Association for Computational Linguistics. pp. 6382-6391 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 182 + ], + "type": "text", + "content": "42. Shi, B., Ji, L., Niu, Z., Duan, N., Zhou, M., Chen, X.: Learning semantic concepts and temporal alignment for narrated video procedural captioning. In: ACM MM. pp. 4355-4363 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 225 + ], + "type": "text", + "content": "43. Shimada, A., Kondo, K., Deguchi, D., Morin, G., Stern, H.: Kitchen scene context based gesture recognition: A contest in ICPR2012. In: International Workshop on Advances in Depth Image Analysis and Applications. pp. 168-185 (2013). https://doi.org/10.1007/978-3-642-40303-3_18" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 225, + 481, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 481, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 481, + 269 + ], + "type": "text", + "content": "44. Shirai, K., Hashimoto, A., Nishimura, T., Kameko, H., Kurita, S., Ushiku, Y., Mori, S.: Visual Recipe Flow: A dataset for learning visual state changes of objects with recipe flows. In: International Conference on Computational Linguistics. pp. 3570-3577 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 269, + 481, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 269, + 481, + 291 + ], + "spans": [ + { + "bbox": [ + 132, + 269, + 481, + 291 + ], + "type": "text", + "content": "45. Spriggs, E.H., De La Torre, F., Hebert, M.: Temporal segmentation and activity classification from first-person sensing. In: CVPRW. pp. 17-24 (2009)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 291, + 481, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 291, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 132, + 291, + 481, + 323 + ], + "type": "text", + "content": "46. Stein, S., McKenna, S.J.: Combining embedded accelerometers with computer vision for recognizing food preparation activities. In: ACM international joint conference on Pervasive and ubiquitous computing. p. 729-738 (2013)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 323, + 481, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 323, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 323, + 481, + 355 + ], + "type": "text", + "content": "47. Tang, Y., Ding, D., Rao, Y., Zheng, Y., Zhang, D., Zhao, L., Lu, J., Zhou, J.: COIN: A large-scale dataset for comprehensive instructional video analysis. In: CVPR (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 355, + 481, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 355, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 355, + 481, + 376 + ], + "type": "text", + "content": "48. Vedantam, R., Zitnick, C.L., Parikh, D.: CIDEr: consensus-based image description evaluation. In: CVPR. pp. 4566-4575 (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 376, + 481, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 481, + 399 + ], + "type": "text", + "content": "49. Wang, T., Zhang, R., Lu, Z., Zheng, F., Cheng, R., Luo, P.: End-to-end dense video captioning with parallel decoding. In: ICCV. pp. 6847-6857 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 399, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 399, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 399, + 481, + 430 + ], + "type": "text", + "content": "50. Wu, J., Pan, L., Chen, J., Jiang, Y.G.: Ingredient-enriched recipe generation from cooking videos. In: ACM International Conference on Multimedia Retrieval. pp. 249-257 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 430, + 481, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 463 + ], + "type": "text", + "content": "51. Yagi, T., Ohashi, M., Huang, Y., Furuta, R., Adachi, S., Mitsuyama, T., Sato, Y.: FineBio: A fine-grained video dataset of biological experiments with hierarchical annotation. arXiv preprint arXiv:2402.00293 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 463, + 481, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 463, + 481, + 484 + ], + "spans": [ + { + "bbox": [ + 132, + 463, + 481, + 484 + ], + "type": "text", + "content": "52. Yamakata, Y., Mori, S., Carroll, J.: English recipe flow graph corpus. In: International Conference on Language Resources and Evaluation. pp. 5187-5194 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 484, + 481, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 484, + 481, + 517 + ], + "spans": [ + { + "bbox": [ + 132, + 484, + 481, + 517 + ], + "type": "text", + "content": "53. Yang, A., Nagrani, A., Seo, P.H., Miech, A., Pont-Tuset, J., Laptev, I., Sivic, J., Schmid, C.: Vid2Seq: Large-scale pretraining of a visual language model for dense video captioning. In: CVPR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 517, + 481, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 517, + 481, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 481, + 550 + ], + "type": "text", + "content": "54. Yang, J., Peng, W., Li, X., Guo, Z., Chen, L., Li, B., Ma, Z., Zhou, K., Zhang, W., Loy, C.C., Liu, Z.: Panoptic video scene graph generation. In: CVPR. pp. 18675-18685 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 550, + 481, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 550, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 550, + 481, + 582 + ], + "type": "text", + "content": "55. Zellers, R., Lu, J., Lu, X., Yu, Y., Zhao, Y., Salehi, M., Kusupati, A., Hessel, J., Farhadi, A., Choi, Y.: MERLOT Reserve: Neural script knowledge through vision and language and sound. In: CVPR. pp. 16375-16387 (June 2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 582, + 481, + 604 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 582, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 132, + 582, + 481, + 604 + ], + "type": "text", + "content": "56. Zhou, L., Kalantidis, Y., Chen, X., Corso, J.J., Rohrbach, M.: Grounded video description. In: CVPR. pp. 6578-6587 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 604, + 481, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 604, + 481, + 624 + ], + "spans": [ + { + "bbox": [ + 132, + 604, + 481, + 624 + ], + "type": "text", + "content": "57. Zhou, L., Xu, C., Corso, J.J.: Towards automatic learning of procedures from web instructional videos. In: AAAI (2018)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 624, + 481, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 481, + 658 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 481, + 658 + ], + "type": "text", + "content": "58. Zhukov, D., Alayrac, J.B., Cinbis, R.G., Fouhey, D., Laptev, I., Sivic, J.: Cross-task weakly supervised learning from instructional videos. In: CVPR. pp. 3537-3545 (2019)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "K. Maeda et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/COMO_ Compact Mapping and Odometry/005f4f87-73cd-41e4-ad7b-b179ff03449c_content_list.json b/2024/COMO_ Compact Mapping and Odometry/005f4f87-73cd-41e4-ad7b-b179ff03449c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..288193ace4441ad7a847f3e0b3314812231d8762 --- /dev/null +++ b/2024/COMO_ Compact Mapping and Odometry/005f4f87-73cd-41e4-ad7b-b179ff03449c_content_list.json @@ -0,0 +1,1942 @@ +[ + { + "type": "text", + "text": "COMO: Compact Mapping and Odometry", + "text_level": 1, + "bbox": [ + 259, + 141, + 741, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Eric Dexheimer1 and Andrew J. Davison1", + "bbox": [ + 339, + 189, + 663, + 204 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dyson Robotics Lab, Imperial College London {e.dexheimer21, a.davison}@imperial.ac.uk", + "bbox": [ + 341, + 217, + 660, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We present COMO, a real-time monocular mapping and odometry system that encodes dense geometry via a compact set of 3D anchor points. Decoding anchor point projections into dense geometry via per-keyframe depth covariance functions guarantees that depth maps are joined together at visible anchor points. The representation enables joint optimization of camera poses and dense geometry, intrinsic 3D consistency, and efficient second-order inference. To maintain a compact yet expressive map, we introduce a frontend that leverages the covariance function for tracking and initializing potentially visually indistinct 3D points across frames. Altogether, we introduce a real-time system capable of estimating accurate poses and consistent geometry.", + "bbox": [ + 261, + 275, + 738, + 428 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: SLAM $\\cdot$ Multi-view geometry $\\cdot$ 3D representations", + "bbox": [ + 261, + 440, + 686, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 478, + 375, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Achieving accurate and consistent poses and dense geometry from monocular images in real-time is a challenging yet essential undertaking to push forward state-of-the-art in robotics and augmented reality. Monocular cameras are the key to achieving low-cost, energy-efficient, and compact intelligent platforms. While images contain rich visual information, reconstructing the world is challenging due to the lack of direct geometric observations.", + "bbox": [ + 212, + 508, + 785, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The ideal representation for real-time monocular visual odometry (VO) and simultaneous localization and mapping (SLAM) remains elusive. While sparse methods jointly optimize camera poses and a set of conditionally independent 3D points given poses, the map lacks consistent dense geometry for downstream tasks and pose estimation cannot benefit from all visual information. Dense SLAM uses and reconstruct all pixels, but the sheer number of variables relative to measurements renders joint optimization infeasible and inference ill-posed. Recently, learned priors over compact depth map representations, such as codes and sparse 2D depths, have enabled joint optimization. However, representations with 2D depth priors have thus far lagged in accuracy, as consistency between depth maps in 3D is not guaranteed. Volumetric representations, such as voxel grids and neural fields, directly enforce dense 3D consistency by construction, but cannot perform real-time joint optimization due to expensive rendering. Furthermore, compact, efficient, and expressive 3D priors for general scenes are not yet suitable. Achieving accurate and consistent 3D geometry in real-time would represent a significant advance for monocular SLAM.", + "bbox": [ + 212, + 598, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ea6a588415295b00661b95c8da0edc65005b10b9ae85a74954e87a208bdd08f9.jpg", + "image_caption": [ + "(a) Point cloud reconstruction from the slidingwindow monocular odometry defined by 240 points and 8 keyframes." + ], + "image_footnote": [], + "bbox": [ + 217, + 148, + 488, + 253 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/141eb0d4228048625b58d4e3c468459e7961b84a3a79bb602a0136ef30099705.jpg", + "image_caption": [ + "(b) Dense point cloud from 40 keyframes across a large office during live operation of a monocular camera.", + "Fig. 1: COMO encodes scene geometry via a compact set of 3D anchor points and decodes dense geometry via per-keyframe depth covariance functions. The 3D points visualized in red anchor depth maps together from multiple views while the covariance function generates dense geometry by conditioning on sparse point projections." + ], + "image_footnote": [], + "bbox": [ + 509, + 143, + 784, + 253 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a representation that achieves the three desired properties: joint optimization of poses and dense geometry, intrinsic 3D consistency, and low-latency, real-time operation. We encode the scene as a compact set of 3D anchor points, which are shared across frames. To decode into dense geometry, 3D points are first projected into keyframes. We then use image-conditioned depth covariance functions [9] per keyframe to generate dense depth maps that intersect each visible 3D point. Depth maps are backprojected into 3D for calculating any-pixel photometric error. All steps are efficient and have analytic Jacobians for second-order optimization. Updates to dense geometry are propagated into our encoded scene representation and ensure suitable regularization for the ill-posed problem of dense reconstruction.", + "bbox": [ + 212, + 381, + 784, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Maintaining a compact map in an incremental VO setting requires a compatible frontend. Keeping a small number of points benefits efficiency, while the distribution of points affects expressiveness. Feature-based systems focus on visually distinct features, but this often leads to many points on edges which are not suitable for modeling dense geometry. Therefore, we leverage depth covariance for determining anchor point visibility in new keyframes, actively initializing new 3D points, and encoding the current dense geometry into new 3D points.", + "bbox": [ + 212, + 547, + 784, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our system demonstrates improved robustness over methods lacking joint pose and dense geometry optimization, as well as more accurate pose and geometry estimation compared to compact 2D representations. Altogether, we propose a real-time VO system that achieves robust and accurate odometry along with consistent dense geometry as shown in Fig. 1.", + "bbox": [ + 212, + 654, + 784, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the contributions of our work are:", + "bbox": [ + 238, + 729, + 580, + 743 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- An efficient representation of dense geometry encoded by a compact set of 3D anchor points and decoded by depth covariance functions.", + "- A frontend for the compact map that leverages depth covariance for visibility, active initialization, and encoding dense geometry.", + "- A real-time monocular visual odometry and mapping system that produces accurate and consistent poses and dense geometry." + ], + "bbox": [ + 223, + 750, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "E. Dexheimer and A.J. Davison", + "bbox": [ + 271, + 114, + 483, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 388, + 160 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sparse vs. Dense Visual SLAM Sparse SLAM systems optimize 3D points that are conditionally independent given poses. Exploiting the sparsity inherent in the information form of structure-from-motion is the key to real-time algorithms with a significant number of points [17]. Modern sparse VO methods [10,12,24] and large-scale bundle adjustment [31] rely on the Schur complement which greatly reduces optimization complexity. Sparse systems lack dense reconstruction, which is useful for many downstream tasks. Thus, it is common to first perform pose estimation followed by dense mapping given poses and potentially sparse landmark estimates [20,23,25,28]. MVS approaches estimate depth maps via a reference cost volume and multiple supporting frames, and have shown progress when integrated into a learning pipeline [18]. However, the sparse followed by dense paradigm has two major issues. Dense information is ignored for pose estimation, which can result in failure in many real-world scenarios, while the mapping step is heavily reliant on accurate pose estimation. Joint optimization of poses and dense geometry avoids the brittle two-step process and has the potential to achieve more accurate odometry and mapping [29].", + "bbox": [ + 212, + 186, + 787, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dense methods [11, 27] brought the promise of estimating poses and complete geometry simultaneously. However, joint optimization proved difficult due to the ill-posed nature of reconstructing all points. Hand-crafted geometry priors regularize the problem, but disrupt the sparsity and thus the tractability of joint optimization as discussed in [10]. In practice, these methods resort to alternating pose and geometry estimation or move away from second-order optimization. More recently, DROID-SLAM [37] reconstructs depths for all pixels, but similar to sparse systems, does so with no geometric correlation. This results in highly parallelized bundle adjustment, but results in many noisy points and no guarantee of consistent depths within a frame and across frames, as shown in Fig. 2a. All real-time state-of-the-art SLAM systems [6, 10, 37, 38] still rely on the Schur complement, and dense methods with geometric correlation have traditionally not demonstrated comparable pose accuracy, which also limits geometry estimation.", + "bbox": [ + 212, + 430, + 787, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learned Depth Priors for Compact Optimization To achieve consistent poses and geometry, monocular depth priors in SLAM focus on test-time optimization to avoid irreparable errors from single-view depth prediction. Therefore, depth priors in monocular SLAM often predict a subspace of potential depths given an image via a compact latent code [3,7,20], or a depth map basis [13,36]. Even with this flexibility, depth maps are not guaranteed to be consistent across multiple views since latent variables live only in frames as shown in Fig. 2b. While per-pixel depth and keypoint reprojection losses may encourage 3D consistency [7], this may introduce bias and does not scale computationally since it requires per-pixel pairwise frame comparisons. Most importantly, it lacks intrinsic 3D consistency, and depth maps often produce a layering effect in 3D as in Fig. 2b. Predicting a view-based mesh topology permits optimizing depth vertices [4], but it is unclear how to move to a 3D topology.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "COMO", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f69143eda7fa70c6e561e08af00d0bbe115f3f33b72c77134efcff92696bfa02.jpg", + "image_caption": [ + "(a) Dense bundle adjustment with 3D landmarks [37]" + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 401, + 212 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6be085474a9c4deffa7ed1e293b2cac8ae196a818611117c1402a0efd86da8ff.jpg", + "image_caption": [ + "(b) Per-frame dense depth priors via compact codes [7]", + "Fig. 2: Reconstructions on Replica and geometry properties of different dense map representations. The dotted line represents the true surface. (a) Densely reconstructing a large number of conditionally independent 3D points given poses enables accurate pose estimation and many accurate points, but there is no guarantee of coherent geometry. (b) Depth priors produce smooth depth maps, but even with inter-frame consistency losses, can produce inconsistent geometry and bias preventing global consistency. (c) A compact set of 3D points and depth covariance functions anchor depth maps together, leading to consistent pose estimation and dense geometry." + ], + "image_footnote": [], + "bbox": [ + 408, + 143, + 594, + 212 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8bb3b6047f89b566c558cf0e3798e197d9d4e17b6ee1d1e70ba738627eafba9c.jpg", + "image_caption": [ + "(c) Ours: 3D points and depth covariance functions" + ], + "image_footnote": [], + "bbox": [ + 602, + 143, + 787, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we leverage the recently proposed learned depth covariance function [9]. While [9] introduces an odometry formulation, it only optimizes per-frame 2D inducing depths, which has the same limitations as per-frame codes, and lacks intrinsic 3D consistency. Our key insight is that depth covariance permits sharing latent 3D points across frames, so that our representation lives in 3D rather than separate 2D image planes. While codes and bases are a fixed subspace that cannot guarantee a depth map passes through any set of 3D points, our covariance formulation guarantees that depth maps go through any desired 3D points as shown in Fig. 2c. Therefore, we can anchor depth maps together by construction for dense 3D consistency. We also exploit the covariance function for our frontend in tracking and initializing potentially non-visually distinct points.", + "bbox": [ + 212, + 393, + 787, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Keyframe-based vs. Volumetric Maps Keyframe-based maps host dense depths in camera frames, which maintain flexible level-of-detail rooted in the camera resolution and focus only on surfaces. However, representing space using per-pixel depths cannot guarantee that inter-frame correspondences refer to the same 3D point. While pairwise depth constraints can mitigate inconsistency [7], these exhaustive per-pixel errors are expensive to compute, lack intrinsic consistency, and may introduce bias when balancing with other losses. Methods that project hosted depths into neighboring frames and predict flow updates [37] are asymmetric with no guarantee of cycle consistency.", + "bbox": [ + 212, + 564, + 787, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Volumetric maps live in 3D space and unknown quantities are shared across frames. 3D consistency is achieved by construction but it is challenging to balance fidelity and efficiency. Voxel grids using backward sensor models [15,26] assume known poses and independence between cells, and thus are unsuitable for efficient and consistent optimization. While forward models in voxel grids [39], neural fields [21,35], and 3D Gaussian splatting [16,19] permit consistent geometry and pose optimization, these representations are limited in low-latency monocular VO and SLAM due to expensive rendering, a large number of correlated variables per ray, and alternating and first-order optimization. Volumetric methods are", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "E. Dexheimer and A.J. Davison", + "bbox": [ + 271, + 114, + 483, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "also subject to a resolution trade-off. Higher resolution permits greater detail at the expense of memory, runtime complexity, and higher levels of noise.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our 3D representation achieves the benefits of both frame-based and volumetric maps for odometry and mapping: the per-pixel resolution and efficiency of depth maps with the intrinsic 3D consistency of volumetric representations. The efficient compact-to-dense model permits joint optimization of poses and geometry via any-pixel constraints in real-time.", + "bbox": [ + 212, + 176, + 784, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Compact Mapping Backend", + "text_level": 1, + "bbox": [ + 215, + 272, + 522, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The mapping backend maintains a compact set of 3D anchor points along with keyframes hosting per-pixel covariance function parameters. A dense world-centric point cloud is decoded from anchor points and poses, which is refined by second-order minimization of photometric error. VO operates in a sliding-window fashion with a fixed number of keyframes. Fig. 3 provides an overview.", + "bbox": [ + 212, + 301, + 782, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantities in the world frame and camera frame, such as points, are denoted by subscripts $\\mathbf{P}_{\\mathrm{W}}$ and $\\mathbf{P}_{\\mathrm{C}}$ , respectively. Anchor points are indexed by superscripts $m$ , while dense points are indexed $n$ , such as pixels $\\mathbf{p}^m$ and $\\mathbf{p}^n$ . Stacked vector forms of sparse and dense variables use capital M and N, as in log-depth vectors $\\mathbf{d}_{\\mathrm{M}}$ and $\\mathbf{d}_{\\mathrm{N}}$ . The notation $\\|\\mathbf{r}\\|_{\\Sigma}^{2}$ denotes Mahalanobis distance of residual vector $\\mathbf{r}$ with measurement covariance matrix $\\boldsymbol{\\Sigma}$ .", + "bbox": [ + 212, + 377, + 784, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminaries: Depth Covariance Function", + "text_level": 1, + "bbox": [ + 215, + 486, + 607, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, we introduce the depth covariance function from [9]. This covariance function models the distribution over all possible log-depth functions for any finite set of pixels via a Gaussian process (GP) [30]. First, a CNN takes in an RGB image and outputs a per-pixel feature map $\\phi$ . Compared to [9], we use a zero-mean GP prior for simplicity, such that the distribution is defined only by the covariance function $k$ that takes in two pixels and respective CNN features:", + "bbox": [ + 212, + 508, + 782, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd (\\mathbf {p}) \\sim \\mathcal {G P} (0, k ([ \\mathbf {p}; \\phi (\\mathbf {p}) ], [ \\mathbf {p} ^ {\\prime}; \\phi (\\mathbf {p} ^ {\\prime}) ])). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 608, + 784, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We found that the mean log-depth variable proposed in [9] is unnecessary, as the largest eigenvector of the covariance function corresponds to scale of the entire depth map. More information on the exact covariance function $k$ is detailed in [9], but the key idea is that larger covariance is an indicator of similarity in log-depth. Since the covariance function takes in both pixel locations and CNN features, it is nonstationary which permits varying levels of smoothness and discontinuities conditioned on the image content. The zero-mean prior defines a planar depth prior in the absence of observations. To define a covariance matrix $\\mathbf{K}$ for a set of pixels, the $(i^{\\mathrm{th}}, j^{\\mathrm{th}})$ entry of $\\mathbf{K}$ is filled in by $k\\left([ \\mathbf{p}^i; \\phi(\\mathbf{p}^i) ], [ \\mathbf{p}^j; \\phi(\\mathbf{p}^j) ] \\right)$ .", + "bbox": [ + 212, + 632, + 784, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Compact-to-Dense Geometry", + "text_level": 1, + "bbox": [ + 215, + 787, + 503, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given anchor point to keyframe visibility, we generate a dense depth map given anchor points $\\mathbf{P}_{\\mathrm{W}}^{m}$ , the pose of keyframe $r$ with rotation $\\mathbf{R}_{\\mathrm{WC}_r}$ and translation", + "bbox": [ + 212, + 809, + 782, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "COMO", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d55bbe92cff36c4db002afbf5f3ab8c93ef98f98a98033e342d55a2fb109e085.jpg", + "image_caption": [ + "(a) Anchor point projection" + ], + "image_footnote": [], + "bbox": [ + 215, + 145, + 352, + 208 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a3fb4a15fada429a43cb0542c2e0917a3cb9bb6571e3c87b2c048c47bb5b88eb.jpg", + "image_caption": [ + "(b) Dense geometry construction", + "Fig. 3: Overview of compact mapping framework. (a) Anchor points are projected to visible keyframes. (b) Keyframes decode into dense depth and backproject geometry to 3D. (c) Target frames enforce dense photo-consistency. (d) Optimizing dense alignment error leads to updated poses and geometry with greater 3D consistency." + ], + "image_footnote": [], + "bbox": [ + 359, + 143, + 493, + 208 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/270e83fd8c2d0e96f009a7a1567b7bd9c069d9fae7a8d536c52e2a421f835f56.jpg", + "image_caption": [ + "(c) Photometric error calculation" + ], + "image_footnote": [], + "bbox": [ + 504, + 143, + 637, + 208 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/764685c9739176877aaeb68fb9c57a6985437e4348fd6824c164eb2c81fc9f00.jpg", + "image_caption": [ + "(d) Pose and geometry update" + ], + "image_footnote": [], + "bbox": [ + 651, + 143, + 785, + 208 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathbf{t}_{\\mathrm{WC}_r}$ , and the keyframe image. Anchor points are transformed and projected to pixel coordinates $\\mathbf{p}$ via camera projection $\\pi$ :", + "bbox": [ + 214, + 330, + 782, + 362 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {\\mathrm {C} _ {r}} ^ {m} = \\mathbf {R} _ {\\mathrm {W C} _ {r}} ^ {T} \\left(\\mathbf {P} _ {\\mathrm {W}} ^ {m} - \\mathbf {t} _ {\\mathrm {W C} _ {r}}\\right), \\quad \\mathbf {p} ^ {m} = \\pi \\left(\\mathbf {P} _ {\\mathrm {C} _ {r}} ^ {m}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 373, + 785, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "which is also illustrated in Fig. 3a. At this point, we create a dense log-depth map from log-depth observations $d^{m} \\triangleq \\log \\left( [\\mathbf{P}_{\\mathrm{Cr}}^{m}]_{z} \\right)$ and projected pixel coordinates $\\mathbf{p}^{m}$ . Given an image and the covariance function $k$ for any two pixels from Sec. 3.1, we can construct the covariance matrix between all visible anchor point pixels in $\\mathbf{K}_{\\mathrm{MM}}$ , and the covariance between all image pixels and anchor point pixels in $\\mathbf{K}_{\\mathrm{NM}}$ . We also stack all training $d^{m}$ and test $d^{n}$ log-depths into $\\mathbf{d}_{\\mathrm{M}}$ and $\\mathbf{d}_{\\mathrm{N}}$ , respectively. The GP formulation allows efficient dense log-depth prediction via the linear Gaussian conditioning equation:", + "bbox": [ + 214, + 404, + 787, + 526 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {d} _ {\\mathrm {N}} = \\mathbf {K} _ {\\mathrm {N M}} \\left(\\mathbf {K} _ {\\mathrm {M M}}\\right) ^ {- 1} \\mathbf {d} _ {\\mathrm {M}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 537, + 785, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The GP guarantees that each dense log-depth map goes through the anchor points, creating a consistent surface across multiple views. Next, we backproject points for each pixel and transform to world coordinates:", + "bbox": [ + 214, + 568, + 782, + 614 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {\\mathrm {W}} ^ {n} = \\mathbf {R} _ {\\mathrm {W C} _ {r}} \\pi^ {- 1} \\left(\\mathbf {p} ^ {n}, e ^ {d ^ {n}}\\right) + \\mathbf {t} _ {\\mathrm {W C} _ {r}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 626, + 785, + 646 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $n$ indexes a query pixel. Note that each keyframe generates a dense depth map via its observed anchor points and covariance parameter feature maps as in Fig. 3b. This compact-to-dense formulation only requires transformations, camera projection, and linear GP conditioning. Therefore, we can efficiently calculate analytical Jacobians for dense per-pixel constraints with respect to poses and anchor points. We show the chain rule for these steps in Section A.1.", + "bbox": [ + 212, + 657, + 787, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In practice, we ignore the Jacobians with respect to pixel coordinates when constructing $\\mathbf{K}_{\\mathrm{MM}}$ and $\\mathbf{K}_{\\mathrm{NM}}$ and calculate the covariance matrices once at initialization. In addition to being more efficient, this ensures that depth maps do not undergo unstable changes if a point moves across a depth discontinuity. If a visible anchor point moves behind a camera during optimization, we reinitialize it to the median depth of the keyframe.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "E. Dexheimer and A.J. Davison", + "bbox": [ + 271, + 114, + 483, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Photometric Residuals", + "text_level": 1, + "bbox": [ + 215, + 146, + 450, + 159 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While only keyframes are used to create dense geometry, additional small baseline support frames are included to improve convergence. We define a set of edges $\\mathcal{E}$ for photometric error that consists of temporally adjacent keyframe-keyframe pairs, and support frames to the two nearest keyframes in time. Given a target image $I_{t}$ , which may be either a keyframe or support frame, with pose $\\mathbf{T}_{\\mathrm{WC}_t}$ , we first transform world points from keyframe view $r$ and project:", + "bbox": [ + 212, + 166, + 782, + 257 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {t} ^ {n} = \\pi \\left(\\mathbf {R} _ {\\mathrm {W C} _ {t}} ^ {T} \\left(\\mathbf {P} _ {\\mathrm {W}} ^ {n} - \\mathbf {t} _ {\\mathrm {W C} _ {t}}\\right)\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 263, + 785, + 281 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given these projective correspondences and accounting for exposure and global illumination changes via affine brightness parameters, $a$ and $b$ , similar to [10], we calculate the photometric error across pixels and frame edges:", + "bbox": [ + 214, + 285, + 782, + 330 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nE = \\sum_ {r, t \\in \\mathcal {E}} \\sum_ {n} \\left\\| \\mathbf {r} _ {r, t} ^ {n} \\right\\| _ {\\sigma_ {\\mathbf {r}} ^ {2} I} ^ {2}, \\quad \\mathbf {r} _ {r, t} ^ {n} = I _ {t} \\left(\\mathbf {p} _ {t} ^ {n}\\right) + b _ {t} - \\left(\\frac {e ^ {- a _ {r}}}{e ^ {- a _ {t}}} I _ {r} \\left(\\mathbf {p} _ {r} ^ {n}\\right) + b _ {r}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 337, + 785, + 372 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "which is illustrated in Fig. 3c. These constraints depend on each frame's pose and affine parameters, and the anchor points viewed by the reference keyframe. For robustness, we perform iteratively reweighted least squares (IRLS) with a Huber cost function and set the photometric standard deviation $\\sigma_{\\mathbf{r}}$ to be proportional to the median absolute residual [1] to handle diverse scenes with different error scales. Compared to RGB residuals, we found that grayscale images were more efficient and similarly robust. The Jacobian chain rule for photometric residuals is in Section A.2.", + "bbox": [ + 212, + 378, + 784, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Additional Constraints", + "text_level": 1, + "bbox": [ + 215, + 517, + 450, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While photometric error is the only data term, we introduce additional priors to remove ambiguities. Priors on the oldest keyframe's pose and affine brightness parameters remove gauge freedom in the sliding-window. After removing a keyframe from the window, we include priors on its still-observed landmarks to ensure global consistency as a simpler alternative to first-estimate Jacobians [10].", + "bbox": [ + 212, + 537, + 782, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since keyframes may have regions unobserved by other frames, which causes anchor points to lack photometric constraints, we include two geometry priors. The first is a weak prior on anchor points to match the log median depth $s$ of the keyframe it was first observed in. The second is a GP prior $||\\mathbf{d}_{\\mathrm{M}} - s||_{\\mathbf{K}_{\\mathrm{MM}}}^2$ on every keyframe to regularize anchor points that have large correlation, which gives improved estimates for underconstrained anchor points lying on surfaces with well-constrained ones, such as on textureless walls.", + "bbox": [ + 212, + 614, + 784, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "During optimization, there is ambiguity in whether an anchor point changes depth or moves laterally along a surface, such as when viewing a plane at an angle. Thus, we include a pixel prior encouraging the projection of the point in its initializing keyframe not to change from its first observation, which defines a ray-surface intersection. As mentioned previously, this is more efficient than calculating Jacobians through Sec. 3.2 and more stable since GP inducing point optimization has little benefit when points start in suitable locations [5] as ensured by our frontend in Sec. 4.", + "bbox": [ + 212, + 719, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "COMO", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bf2b1436a9bd43ef487970386dda0de40fa46524793e44d7149c6d906323cf10.jpg", + "image_caption": [ + "(a) Anchor points visible in KF #1" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 352, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c96b1c6c85ac56ac8790cd63e74819fd78715220f95ce9e92ed5a6ec6e07a935.jpg", + "image_caption": [ + "(b) KF #1 depths projected to KF #2", + "Fig. 4: Overview of visibility checks between two keyframes (KF). Matches are shown in blue, rejected matches in red, and newly initialized points in green. Note that occluded edges are rejected, while non-visually distinct points are often matched. New points are allocated to geometrically complex regions while the table is already well-represented." + ], + "image_footnote": [], + "bbox": [ + 359, + 143, + 496, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dea814d6b88eb52d91e1c2636e193b468772fa1b8a095c1a96722a7219fffc52.jpg", + "image_caption": [ + "(c) Anchor point matches in KF #2" + ], + "image_footnote": [], + "bbox": [ + 503, + 143, + 638, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c52c02513e3d47af233accf282aa43f65847462be8b4d25371c2b54437aebe52.jpg", + "image_caption": [ + "(d) Newly initialized points in KF #2" + ], + "image_footnote": [], + "bbox": [ + 651, + 143, + 787, + 222 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Second-Order Optimization", + "text_level": 1, + "bbox": [ + 215, + 338, + 488, + 353 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We calculate analytical Jacobians for all constraints in the optimization. The compactness of our linear system permits real-time joint optimization of poses and correlated geometry that is infeasible for traditional dense VO. Stacking all Jacobians into $\\mathbf{J}$ , weights from IRLS and noise models into the diagonal matrix $\\mathbf{W}$ , and residuals into $\\mathbf{r}$ , we solve for parameter updates $\\Delta \\mathbf{x}$ shown in Fig. 3d via dense Cholesky factorization of the Gauss-Newton normal equations:", + "bbox": [ + 212, + 359, + 787, + 450 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\mathbf {J} ^ {T} \\mathbf {W} \\mathbf {J}\\right) \\Delta \\mathbf {x} = - \\mathbf {J} ^ {T} \\mathbf {W r}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 455, + 785, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pose updates are minimally parameterized as Lie-algebra elements $\\mathfrak{se}(3)$ as is standard in SLAM [32]. We found that the most time consuming step of backend optimization is accumulating geometry blocks into the Hessian $(\\mathbf{J}^T\\mathbf{W}\\mathbf{J})$ . This can be sped up by factoring out constant components of the Jacobians and reducing the dimensionality of the accumulation as detailed in Section A.4.", + "bbox": [ + 212, + 479, + 787, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Visual Frontend", + "text_level": 1, + "bbox": [ + 215, + 575, + 408, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The visual frontend tracks the current frame with respect to the map, selects keyframes and support frames for the backend, determines 3D anchor point visibility in new keyframes, and initializes new 3D anchor points. Compared to feature-based systems, the ideal points for an expressive yet compact map may be visually indistinct, so we utilize depth covariance for robust visibility checks.", + "bbox": [ + 212, + 603, + 787, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Dense Photometric Tracking", + "text_level": 1, + "bbox": [ + 215, + 696, + 500, + 713 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given the current estimate of the newest keyframe's depth, we optimize the pose and affine brightness parameters using grayscale image alignment in the IRLS framework [2]. For efficiency, we perform inverse compositional tracking and use a multi-scale image pyramid. After tracking, a new keyframe is added to the backend if there is either significant translation relative to the median scene depth or if the number of unique projected pixels in the tracked frame falls below a threshold to handle rotation and occlusion. Support frames are added via modified thresholds according to the desired frequency between keyframes.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "E. Dexheimer and A.J. Davison", + "bbox": [ + 271, + 114, + 483, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Anchor Point Visibility", + "text_level": 1, + "bbox": [ + 215, + 146, + 455, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "When a new keyframe is added, we perform explicit visibility checks to 3D anchor points viewed in the last keyframe to achieve 3D consistency. Note that we only check for visibility as anchor points have no explicit constraint on its location in the image. Compared to feature-based methods [24], anchor points often lack visual distinctiveness, so matching via image information is problematic. Instead, we leverage the interpretability of the depth covariance function to determine visibility. In general, an anchor point should be compatible with all depth covariance functions that use it to generate dense geometry.", + "bbox": [ + 212, + 170, + 787, + 290 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "First, the 3D anchor points and dense depth map from the last keyframe, shown in Fig. 4a, are projected into the current frame, shown in Fig. 4b. Given the projected sparse and dense pixel coordinates, we may query covariance parameters $\\phi (\\mathbf{p})$ and construct $\\mathbf{K}_{\\mathrm{MM}}$ and $\\mathbf{K}_{\\mathrm{MN}}$ for the new keyframe. We leverage the linearity of the GP to compress dense log-depth observations into the sparse projected anchor point coordinates. Assuming conditional independence of log-depth observations $\\mathbf{d}_n$ given the sparse log-depths $\\mathbf{d}_m$ , we can solve an efficient least-squares problem:", + "bbox": [ + 212, + 291, + 787, + 412 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {d} _ {\\mathrm {M}}} \\left| \\left| \\mathbf {K} _ {\\mathrm {N M}} \\left(\\mathbf {K} _ {\\mathrm {M M}}\\right) ^ {- 1} \\mathbf {d} _ {\\mathrm {M}} - \\mathbf {d} _ {\\mathrm {N}} \\right| \\right| _ {\\sigma_ {d} ^ {2}} ^ {2} + \\left| \\left| \\mathbf {d} _ {\\mathrm {M}} \\right| \\right| _ {\\mathbf {K} _ {\\mathrm {M M}}} ^ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 421, + 784, + 445 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where the first conditional term ensures that $\\mathbf{d}_{\\mathrm{M}}$ fits the projected dense observations with standard deviation $\\sigma_{d}$ , while the second prior term regularizes sparse log-depths according to the GP training covariance $\\mathbf{K}_{\\mathrm{MM}}$ . We perform a consistency check on the log-depth difference between the current 3D anchor points and those solved in Eq. (8), and prune points at depth discontinuities. This yields potential matches the are consistent between the two keyframes.", + "bbox": [ + 212, + 454, + 787, + 544 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In practice, we set a maximum number of visible anchor points per-frame to allow efficient batched mapping operations on the GPU. If all points are matched, it could prevent new ones from being initialized and parts of the scene would lack capacity to adequately represent geometry. Thus, we perform pruning of matches so that we have a suitable distribution of points before initializing new anchors. Similar to the GP conditional mean from Eq. (3), we may also obtain conditional variances of all pixels with respect to the first $j$ pixels via:", + "bbox": [ + 212, + 545, + 787, + 650 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\sigma} _ {j} ^ {2} = \\operatorname {d i a g} \\left[ \\mathbf {K} _ {\\mathrm {N N}} \\right] - \\mathbf {K} _ {\\mathrm {N M} _ {1: j}} \\left(\\mathbf {K} _ {\\mathrm {M} _ {1: j} \\mathrm {M} _ {1: j}}\\right) ^ {- 1} \\mathbf {K} _ {\\mathrm {M} _ {1: j} \\mathrm {N}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 325, + 661, + 784, + 680 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We perform active sampling via conditional variance reduction (CVR) [14, 22], where at the $(j + 1)^{\\mathrm{th}}$ iteration, the pixel in domain $P$ with the largest variance is selected $\\mathbf{p}_{j + 1} = \\arg \\max_{\\mathbf{p}\\in P}\\sigma_j^2 (\\mathbf{p})$ . We actively sample from the set of potential visible points until a variance threshold is reached, which gives the final set of visible anchor points for the current keyframe. Compared to [9], we found that a minimum distance threshold between pixel locations and from the image border is required to avoid accumulating pixels near the same discontinuity. This is essential for improving depth estimates and maintaining a good set of candidate points, as otherwise many points near edges may be occluded at the same time. We show the set of matches between two keyframes in Fig. 4c.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "COMO", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/56b0853a238dff6d20aab3b8a4748b88afbf94a7b5982e685223e2357894f749.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 143, + 316, + 202 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/c2f05c3b6d3c4e0a72eb49baedec167687f5851e8815f2b9b0be9f6b495379a4.jpg", + "image_caption": [ + "Fig. 5: Visualization of sliding-window odometry on TUM $fr2/desk$ . Top images show tracked and newly initialized anchor points in blue and red, respectively. Depth maps are below, while the dense point cloud with 248 anchor points is shown on the right. Note the consistent point tracks on textureless surfaces, such as the table and plate, as well as the long-term tracks on the cube, mug, and book." + ], + "image_footnote": [], + "bbox": [ + 217, + 205, + 315, + 260 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/2a44ec5c85e607d5515af8b0f78286a651d50804b41ecd175e5bc8e2fe690a05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 145, + 418, + 202 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d51115a4ffb4b9a4917f662252f8ad3e1bacdff8753d9c6e6d208180936df3f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 205, + 418, + 260 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/0461e0d585cbc0f7a2a144b675a90b51aaf5154a231c556ca6de222063648ec6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 145, + 521, + 202 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/482264346b9ade53b0a1c67a9cdab492dcc1ae8557215146b8e2d74c3a5ed108.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 205, + 521, + 260 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/0f8562bc9a69aa9ef3ce4e8b3b25f30df79fc97687994bac87ab8de8fb828f91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 145, + 622, + 202 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/2aff15b414a1195164d0b78fdedcb40db594370f8c1b24263b73bd4eba40aa4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 205, + 622, + 260 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/165e0447c1fb62b31213ee9ba4ac35b0b083cc68aa4ed3c57a7eaa20a01455ef.jpg", + "image_caption": [ + "#" + ], + "image_footnote": [], + "bbox": [ + 633, + 143, + 782, + 250 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Anchor Point Initialization", + "text_level": 1, + "bbox": [ + 215, + 372, + 483, + 386 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As the camera explores unobserved parts of the scene, new anchor points must be initialized. First, we again perform active sampling, but now across all pixels given the current anchor point locations from the previous section. This ensures new points will be initialized for parts of the scene that are unobserved, lack matches, or require additional geometric capacity, as shown in Fig. 4d.", + "bbox": [ + 212, + 396, + 782, + 470 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Once the pixel locations of new anchor points are known, we initialize the log-depth and backproject into 3D. Defining the set of known log-depths from the anchor points $\\mathbf{d}_{\\mathrm{m}_1}$ , the unknown log-depths of newly initialized points $\\mathbf{d}_{\\mathrm{m}_2}$ and all log-depths as $\\mathbf{d}_{\\mathrm{M}} = (\\mathbf{d}_{\\mathrm{m}_1}^T,\\mathbf{d}_{\\mathrm{m}_2}^T)^T$ , we solve another least-squares problem:", + "bbox": [ + 212, + 472, + 784, + 534 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {d} _ {\\mathrm {m} _ {2}}} \\left\\| \\mathbf {K} _ {\\mathrm {N M}} \\left(\\mathbf {K} _ {\\mathrm {M M}}\\right) ^ {- 1} \\mathbf {d} _ {\\mathrm {M}} - \\mathbf {d} _ {\\mathrm {N}} \\right\\| _ {\\sigma_ {d} ^ {2}} ^ {2} + \\left\\| \\mathbf {d} _ {\\mathrm {m} _ {2}} - s \\right\\| _ {\\sigma_ {s} ^ {2}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 320, + 555, + 784, + 583 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where the first term promotes fitting reprojected dense depth observations given matched anchor log-depths and the new log-depth initializations, while the second term is a prior that new log-depths are close to the log median depth $s$ of the previous keyframe. We found that using an isotropic prior with standard deviation $\\sigma_d$ , rather than $\\mathbf{K}_{\\mathrm{MM}}$ , avoids bias for parts of the image that were not previously observed and would otherwise be underconstrained. We define $\\sigma_d$ as the standard deviation of the residuals from Eq. (8) to better calibrate the model. Given the new pixel locations, log-depth initializations, and the new keyframe's pose estimate, 3D points are initialized for the backend. A qualitative example of tracked anchor points, depth maps, and reconstruction is shown in Fig. 5.", + "bbox": [ + 212, + 590, + 787, + 758 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 779, + 375, + 796 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We evaluate both the pose and geometry estimation of our method on multiple datasets. For sparse methods, we compare against feature-based ORB-", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "E. Dexheimer and A.J. Davison", + "bbox": [ + 271, + 114, + 483, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/f9fe0afff8493118966b86cc780ff1f9dbad9eca31f1012b535425eaef18fe38.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
office0office1office2office3office4room0room1room2mean
SORB-SLAM3 [6]0.40.38.50.44.50.30.30.41.9
DSO [10]0.30.50.30.25.30.20.80.21.0
DDeepFactors [7]42.938.645.641.450.132.426.144.140.2
DROID-VO [37]5.34.16.511.27.19.04.87.06.9
COMO-NC2.95.07.46.18.14.44.53.75.3
COMO2.42.44.33.65.52.52.73.43.4
", + "bbox": [ + 235, + 143, + 767, + 217 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 1: Mean ATE (cm) over 5 runs for monocular VO on the Replica dataset. S and D in the first column refer to sparse and dense methods, respectively.", + "bbox": [ + 215, + 217, + 785, + 246 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "SLAM3 [6], direct DSO [10], and learning-based DPVO [38] in trajectory estimation. We also compare against a representative set of dense methods. TANDEM [18] uses DSO for odometry and fuses learned MVS predictions into a TSDF for tracking. MonoGS [19] is a 3D Gaussian splatting-based monocular SLAM system. DeepFactors [7] is a code-based system with photometric, geometric, and keypoint losses to promote consistency across depth maps. DROIDVO [37] uses learned correspondence and flow updates for bundle adjustment, such that all depths are reconstructed but lack geometric correlation. DepthCov [9] optimizes sparse 2D depths under the depth covariance framework, and in a related manner, COMO-NC (no correspondence) is our full framework but no 3D anchor points are tracked across frames. We disable loop closure for DeepFactors and DROID-SLAM to isolate the odometry and mapping performance.", + "bbox": [ + 212, + 284, + 787, + 468 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 486, + 455, + 502 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We use a fixed configuration across all three datasets to demonstrate the robustness of the system. We use the off-the-shelf depth covariance function [9] that was trained on the ScanNet training set. Our sliding window odometry uses 9 keyframes and 3 support frames between keyframes for a total of 24. Each keyframe views a maximum of 64 anchor points and initializes new points if it fails to track all from the previous keyframe. We use 256x192 images for all operations, and the mapping backend samples pixels with the highest gradient magnitude in 4x4 images patches for photometric error.", + "bbox": [ + 212, + 508, + 787, + 630 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Trajectory Evaluation", + "text_level": 1, + "bbox": [ + 215, + 650, + 442, + 666 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Replica We first test our monocular odometry on the Replica [33] sequences recorded in [35], which provides a photorealistic environment with ground-truth geometry. As shown in Tab. 1, sparse methods without priors perform best, as conditions are favorable for SLAM: few image artifacts, no exposure changes, and sufficient baselines for multi-view geometry. Among methods with learned priors, COMO achieves the lowest ATE. The representation has little bias as compared to the code-based DeepFactors, which drifts despite photometric, geometric, and keypoint constraints. Sharing 3D points between frames in COMO shows significant improvement over COMO-NC which disables correspondence. TUM The TUM RGBD dataset [34] is a challenging dataset for monocular VO due to significant motion blur, exposure changes, rolling shutter artifacts,", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "COMO", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/eab9f23954ccf4d6bf424b815991a8e2204dab666029f6a65d36733403ed0f86.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
fr1fr2fr3 longmean
360deskdesk2plantroomrpyteddyxyzxyzdesk
SORB-SLAM3 [6]X2.0X11.8X5.6X1.00.51.31.7X
DSO [10]X27.266.06.058.6XX3.80.32.29.9X
DPVO [38]13.19.46.53.039.83.56.21.30.53.55.58.4
DTANDEM [18]X4.333.7XX4.943.12.40.32.08.3X
MonoGS [19]14.26.374.09.364.93.435.61.64.5133.13.331.8
DeepFactors [7]17.915.920.231.938.33.856.05.98.426.349.024.9
DepthCov [9]12.85.64.826.125.75.247.55.61.215.968.819.9
DROID-VO [37]15.75.211.16.033.43.219.15.610.77.97.311.4
COMO-NC16.14.210.919.328.65.268.74.10.78.846.819.4
COMO12.94.99.513.827.04.824.54.00.76.310.510.8
", + "bbox": [ + 220, + 143, + 781, + 267 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 2: Mean ATE (cm) over 5 runs for monocular VO on the TUM dataset. S and D in the first column refer to sparse and dense methods, respectively.", + "bbox": [ + 215, + 268, + 782, + 294 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "and heavy rotations. Results comparing our method against state-of-the-art VO methods on 11 sequences is shown in Tab. 2. For sparse methods, ORB-SLAM3 and DSO both lose tracking on multiple sequences, while DPVO, which uses learned context for tracking 64 sparse patches and performing bundle adjustment, performs best among all methods. In this work, we are focused on also reconstructing dense geometry, but it would be interesting to combine the learned features of DPVO with our 3D representation. For dense methods, TANDEM fails due to relying on DSO for odometry, which can still fail despite using a TSDF with MVS fusion for tracking. This demonstrates the need for joint optimization of poses and dense image information. While MonoGS can produce accurate trajectories for some sequences, the lack of priors reduces its robustness, and volumetric rendering is slower than all depth-based systems. Despite promoting consistency between neighboring depth maps, DeepFactors has relatively large error. Both DepthCov and COMO-NC are similar in terms of optimizing sparse depths in each frame, while leveraging a true 3D representation in COMO reduces error by almost $50\\%$ . COMO outperforms DROID-VO, which uses learned features for matching and reconstructs all pixels from downsized images via sparse bundle adjustment. As the assumptions of brightness constancy are violated in many real-world datasets, learned features could further the accuracy of COMO, but it is interesting that our compact representation and simple photometric error can produce the lowest ATE of dense methods.", + "bbox": [ + 217, + 340, + 785, + 656 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "ScanNet Test Lastly, we evaluate the trajectory error on 18 diverse medium-length scenes from the ScanNet [8] test set. Sequences include homes, offices, schools, businesses, and outdoors. Methods that fail are assigned a maximum $100\\mathrm{cm}$ error for averages. Sequences have challenging rotational motion, high image noise, and specular surfaces. We summarize the results in Fig. 6, which shows the number of sequences below varying ATE thresholds, while Tab. 3 contains summary statistics. COMO has the lowest mean ATE and area-underthe-curve (AUC). While DPVO has a higher number of trajectories under low ATE thresholds, it also shows high variance results on some sequences. Despite using photometric error that is often violated due to image noise and specular surfaces, our representation with 3D consistency proves valuable in challenging real-world data. An example reconstruction is shown in Fig. 1a.", + "bbox": [ + 217, + 659, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "E. Dexheimer and A.J. Davison", + "bbox": [ + 271, + 114, + 482, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5b0644cb0379c11ad6848ac680648dd18d2298bfb7ffa3edb0824a115ff97cd5.jpg", + "image_caption": [ + "Fig. 6: Number of successful trajectories below ATE threshold for selected ScanNet test sequences.", + "Table 3: ScanNet Mean ATE (cm) and AUC." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 614, + 263 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/ce2b87cfd9b70424895e2ac21af29a6bb5a31089d4ae818df735a8c23df7a3da.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodATEAUC
ORB-SLAM379.91.63
DSO54.72.46
DPVO15.16.34
TANDEM41.73.79
DeepFactors36.12.97
DROID-VO13.96.51
COMO-NC15.86.16
COMO13.06.67
", + "bbox": [ + 619, + 146, + 774, + 239 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/bdd62a13c12f9d0c5d6d3de026b5f1cbf7c34047d06258bd3f363fa4c57af1d1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodReplicaScanNet
TANDEMN/A0.325
DeepFactors0.2630.232
DROID-VO1.1291.389
COMO-NC0.0690.155
COMO0.0460.128
", + "bbox": [ + 218, + 345, + 375, + 395 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f59ebe50f0c9a3ce41a1a9f1b25f69c0520fe5b4bba68021f54bf122f360d53d.jpg", + "image_caption": [ + "Table 4: Depth absolute relative error.", + "Fig. 7: ScanNet depth accuracy and consistency metrics for fraction of pixels falling below threshold $\\delta$ ." + ], + "image_footnote": [], + "bbox": [ + 385, + 320, + 578, + 417 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7ccca15e2c6c832ec16a50df124b221a83f84cfbcedda4ef82291a7508cc9f66.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 321, + 779, + 417 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Geometry Evaluation", + "text_level": 1, + "bbox": [ + 215, + 492, + 437, + 508 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Since we are interested in both consistent pose estimation and dense geometry, we evaluate dense depth predictions on Replica and ScanNet. To evaluate accuracy, we perform the same global scale alignment that is used to produce ATE for monocular methods and transform estimated poses and depth maps to be aligned to the metric ground-truth. Beyond global accuracy, we also measure consistency, which checks whether neighboring frames viewing the same 3D points have consistent depth estimates. Note that consistency does not care about the correctness of depth. To check valid pixels, we backproject ground-truth depth maps and project these points into neighboring frames, and check if the depths agree within $1\\mathrm{cm}$ . Then, for all valid 3D correspondences, we calculate metrics between pairs of estimated depth images. If $N$ is the number of keyframes for that sequence, then $N$ depth images are used for accuracy metrics, while $2(N - 1)$ depth images are used for consistency since pairs require both directions.", + "bbox": [ + 212, + 521, + 785, + 718 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "For estimated depth $\\hat{D}$ and ground-truth depth $D$ we show the mean absolute relative error $(1 / N)\\sum_{i}|\\hat{D}_{i} - D_{i}| / D_{i}$ across all sequences for Replica and ScanNet in Tab. 4. We do not evaluate TANDEM on Replica since it is used in the training data. COMO outperforms other methods due to its ability to jointly optimize dense, correlated geometry along with poses. Despite accurate pose estimation, DROID-VO has outliers with significant depth errors. To further evaluate the distribution of depth accuracy and remove the effect of large outliers, in Fig. 7, we use the $\\delta$ threshold commonly used in depth estimation, which measures what", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "COMO", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7f478f47dfe1f7c654ec4f147afdc2d47451e77e57ebdbf1468efe96d3bdfbf3.jpg", + "image_caption": [ + "Fig. 8: Mesh and surface normals from TSDF fusion of a live desk sequence." + ], + "image_footnote": [], + "bbox": [ + 225, + 143, + 777, + 204 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/85f665e026a3e998ae5b11fdec72ec79a24e3fd3db86d694802572aaffc31ea7.jpg", + "image_caption": [ + "Table 5: Runtime in seconds of different methods on TUM $fr3/long$ using an RTX 3080. COMO SP and MP denote single and multiprocessing, respectively." + ], + "image_footnote": [], + "bbox": [ + 238, + 239, + 763, + 263 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "fraction of pixels satisfy $\\max \\left(\\hat{D} / D, D / \\hat{D}\\right) < \\delta$ . Interestingly, accurate depth maps do not necessarily imply consistency. For example, DROID-VO has the second best accuracy behind COMO, while DeepFactors has the second most consistent geometry. COMO achieves both the best accuracy and consistency without explicit geometric and keypoint constraints.", + "bbox": [ + 212, + 330, + 784, + 412 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.4 Real-time System and Runtime Details", + "text_level": 1, + "bbox": [ + 215, + 431, + 584, + 448 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "When evaluating our system, we use a sequential, single-threaded mode that runs at around 14 FPS with an RTX 3080. For live demos and reconstructions, we also develop a multiprocessing version that runs at 35-50 FPS as shown in the FPS counter in the top right of the supplementary video, with tracking flexible to be on either CPU or GPU. We show examples of reconstructions while operating a monocular camera in real-time. Fig. 1b displays a large-scale raw point cloud of an office, while Fig. 8 displays consistent TSDF fusion of a desk sequence.", + "bbox": [ + 212, + 455, + 784, + 561 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Average times for tracking a frame, adding a keyframe (KF), and updating the map are 20.1, 50.4, and $39.8~\\mathrm{ms}$ . Keyframing is the bottleneck during fast motion, but 35-50 FPS in multiprocessing mode for normal motion is standard as shown in the video. We measure the wall time of several methods in Tab. 5. COMO achieves real-time processing with its efficient representation, but further optimization is still possible since almost all code is implemented in PyTorch.", + "bbox": [ + 212, + 561, + 785, + 652 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 674, + 359, + 690 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we present a compact representation of 3D anchor points and frame-based depth covariance functions to achieve real-time estimation of consistent poses and geometry. Future work includes training the depth covariance function on more diverse data and replacing photometric constraints with the more robust learned correspondence from DROID-VO. Lastly, we believe the anchor point representation could be used in a full map-centric SLAM system with relocalization, which would allow continually improving geometry upon revisiting. With the advance of learned priors, exploiting geometric correlation is an interesting direction to achieve efficient, robust, and consistent algorithms.", + "bbox": [ + 212, + 704, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "E. Dexheimer and A.J. Davison", + "bbox": [ + 271, + 114, + 483, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 145, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research is funded by EPSRC Prosperity Partnerships (EP/S036636/1) and Dyson Technology Ltd. We would like to thank Riku Murai, Hidenobu Matsuki, Gwangbin Bae, and members of the Dyson Robotics Lab for insightful discussions and feedback.", + "bbox": [ + 212, + 175, + 787, + 236 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 258, + 321, + 273 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alismail, H.: Direct pose estimation and refinement. Ph.D. thesis, Robotics Institute, Carnegie Mellon University. (2016)", + "2. Baker, S., Matthews, I.: Lucas-Kanade 20 years on: A unifying framework. International Journal of Computer Vision (IJCV) (2004)", + "3. Bloesch, M., Czarnowski, J., Clark, R., Leutenegger, S., Davison, A.J.: CodeSLAM - learning a compact, optimisable representation for dense visual SLAM. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)", + "4. Bloesch, M., Laidlow, T., Clark, R., Leutenegger, S., Davison, A.: Learning meshes for dense visual SLAM. In: Proceedings of the International Conference on Computer Vision (ICCV) (2019)", + "5. Burt, D.R., Rasmussen, C.E., van der Wilk, M.: Convergence of sparse variational inference in Gaussian processes regression. The Journal of Machine Learning Research (2020)", + "6. Campos, C., Elvira, R., Rodríguez, J.J.G., M. Montiel, J.M., D. Tardós, J.: ORBSLAM3: An accurate open-source library for visual, visual-inertial, and multimap SLAM. IEEE Transactions on Robotics (T-RO) (2021)", + "7. Czarnowski, J., Laidlow, T., Clark, R., Davison, A.J.: DeepFactors: Real-time probabilistic dense monocular SLAM. IEEE Robotics and Automation Letters (RA-L) (2020)", + "8. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017)", + "9. Dexheimer, E., Davison, A.J.: Learning a depth covariance function. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)", + "10. Engel, J., Koltun, V., Cremers, D.: Direct sparse odometry. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2018)", + "1. Engel, J., Schöps, T., Cremers, D.: LSD-SLAM: Large-scale direct monocular SLAM. In: Proceedings of the European Conference on Computer Vision (ECCV) (2014)", + "2. Forster, C., Zhang, Z., Gassner, M., Werlberger, M., Scaramuzza, D.: SVO: Semidirect visual odometry for monocular and multicamera systems. IEEE Transactions on Robotics (T-RO) (2017)", + "3. Graham, B., Novotny, D.: RidgeSfM: Structure from motion via robust pairwise matching under depth uncertainty. In: Proceedings of the International Conference on 3D Vision (3DV) (2020)", + "4. Guestrin, C., Krause, A., Singh, A.P.: Near-optimal sensor placements in Gaussian processes. In: Proceedings of the International Conference on Machine Learning (ICML) (2005)" + ], + "bbox": [ + 225, + 287, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "COMO", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Hornung, A., Wurm, K.M., Bennewitz, M., Stachniss, C., Burgard, W.: OctoMap: An efficient probabilistic 3D mapping framework based on octrees. Autonomous Robots (2013)", + "16. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3D Gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/", + "17. Klein, G., Murray, D.: Parallel tracking and mapping for small AR workspaces. In: Proceedings of the International Symposium on Mixed and Augmented Reality (ISMAR) (2007)", + "18. Koestler, L., Yang, N., Zeller, N., Cremers, D.: TANDEM: Tracking and dense mapping in real-time using deep multi-view stereo. In: Conference on Robot Learning (CoRL) (2022)", + "19. Matsuki, H., Murai, R., Kelly, P.H.J., Davison, A.J.: Gaussian Splatting SLAM. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2024)", + "20. Matsuki, H., Scona, R., Czarnowski, J., Davison, A.J.: CodeMapping: Real-time dense mapping for sparse SLAM using compact scene representations. IEEE Robotics and Automation Letters (RA-L) (2021)", + "21. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: Proceedings of the European Conference on Computer Vision (ECCV) (2020)", + "22. Moss, H.B., Ober, S.W., Picheny, V.: Inducing point allocation for sparse Gaussian processes in high-throughput bayesian optimisation. In: International Conference on Artificial Intelligence and Statistics (AISTATS) (2023)", + "23. Mur-Artal, R., Tardós, J.D.: Probabilistic semi-dense mapping from highly accurate feature-based monocular SLAM. In: Proceedings of Robotics: Science and Systems (RSS) (2015)", + "24. Mur-Artal, R., Tardós, J.D.: ORB-SLAM2: An open-source SLAM system for monocular, stereo, and RGB-D cameras. IEEE Transactions on Robotics (T-RO) (2017)", + "25. Newcombe, R.A., Davison, A.J.: Live dense reconstruction with a single moving camera. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2010)", + "26. Newcombe, R.A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohi, P., Shotton, J., Hodges, S., Fitzgibbon, A.: KinectFusion: Real-time dense surface mapping and tracking. In: Proceedings of the International Symposium on Mixed and Augmented Reality (ISMAR) (2011)", + "27. Newcombe, R.A., Lovegrove, S.J., Davison, A.J.: DTAM: Dense tracking and mapping in real-time. In: Proceedings of the International Conference on Computer Vision (ICCV) (2011)", + "28. Pizzoli, M., Forster, C., Scaramuzza, D.: REMODE: Probabilistic, monocular dense reconstruction in real time. In: Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) (2014)", + "29. Platinsky, L., Davison, A.J., Leutenegger, S.: Monocular visual odometry: Sparse joint optimisation or dense alternation? In: Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) (2017)", + "30. Rasmussen, C.E., Williams, C.K.I.: Gaussian Processes for Machine Learning. MIT Press (2005)", + "31. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "E. Dexheimer and A.J. Davison", + "bbox": [ + 271, + 114, + 483, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Sola, J., Deray, J., Atchuthan, D.: A micro lie theory for state estimation in robotics. arXiv preprint arXiv:1812.01537 (2018)", + "33. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., Clarkson, A., Yan, M., Budge, B., Yan, Y., Pan, X., Yon, J., Zou, Y., Leon, K., Carter, N., Brales, J., Gillingham, T., Mueggler, E., Pesqueira, L., Savva, M., Batra, D., Strasdat, H.M., Nardi, R.D., Goesele, M., Lovegrove, S., Newcombe, R.: The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)", + "34. Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for the evaluation of RGB-D SLAM systems. In: Proceedings of the IEEE/RSJ Conference on Intelligent Robots and Systems (IROS) (2012)", + "35. Sucar, E., Liu, S., Ortiz, J., Davison, A.J.: iMAP: Implicit mapping and positioning in real-time. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021)", + "36. Tang, C., Tan, P.: BA-Net: Dense bundle adjustment networks. In: Proceedings of the International Conference on Learning Representations (ICLR) (2019)", + "37. Teed, Z., Deng, J.: DROID-SLAM: Deep visual SLAM for monocular, stereo, and RGB-D cameras. In: Neural Information Processing Systems (NeurIPS) (2021)", + "38. Teed, Z., Lipson, L., Deng, J.: Deep patch visual odometry. In: Neural Information Processing Systems (NeurIPS) (2023)", + "39. Thrun, S.: Learning occupancy grids with forward models. In: Proceedings of the IEEE/RSJ Conference on Intelligent Robots and Systems (IROS) (2001)" + ], + "bbox": [ + 212, + 146, + 787, + 452 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "COMO", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/COMO_ Compact Mapping and Odometry/005f4f87-73cd-41e4-ad7b-b179ff03449c_model.json b/2024/COMO_ Compact Mapping and Odometry/005f4f87-73cd-41e4-ad7b-b179ff03449c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bfa0f55775f10155aaf12a5e1fbeffaa43f15213 --- /dev/null +++ b/2024/COMO_ Compact Mapping and Odometry/005f4f87-73cd-41e4-ad7b-b179ff03449c_model.json @@ -0,0 +1,2522 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.26, + 0.142, + 0.743, + 0.163 + ], + "angle": 0, + "content": "COMO: Compact Mapping and Odometry" + }, + { + "type": "text", + "bbox": [ + 0.34, + 0.19, + 0.664, + 0.205 + ], + "angle": 0, + "content": "Eric Dexheimer1 and Andrew J. Davison1" + }, + { + "type": "text", + "bbox": [ + 0.343, + 0.218, + 0.661, + 0.246 + ], + "angle": 0, + "content": "Dyson Robotics Lab, Imperial College London {e.dexheimer21, a.davison}@imperial.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.276, + 0.74, + 0.429 + ], + "angle": 0, + "content": "Abstract. We present COMO, a real-time monocular mapping and odometry system that encodes dense geometry via a compact set of 3D anchor points. Decoding anchor point projections into dense geometry via per-keyframe depth covariance functions guarantees that depth maps are joined together at visible anchor points. The representation enables joint optimization of camera poses and dense geometry, intrinsic 3D consistency, and efficient second-order inference. To maintain a compact yet expressive map, we introduce a frontend that leverages the covariance function for tracking and initializing potentially visually indistinct 3D points across frames. Altogether, we introduce a real-time system capable of estimating accurate poses and consistent geometry." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.441, + 0.687, + 0.456 + ], + "angle": 0, + "content": "Keywords: SLAM \\(\\cdot\\) Multi-view geometry \\(\\cdot\\) 3D representations" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.479, + 0.377, + 0.495 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.787, + 0.598 + ], + "angle": 0, + "content": "Achieving accurate and consistent poses and dense geometry from monocular images in real-time is a challenging yet essential undertaking to push forward state-of-the-art in robotics and augmented reality. Monocular cameras are the key to achieving low-cost, energy-efficient, and compact intelligent platforms. While images contain rich visual information, reconstructing the world is challenging due to the lack of direct geometric observations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The ideal representation for real-time monocular visual odometry (VO) and simultaneous localization and mapping (SLAM) remains elusive. While sparse methods jointly optimize camera poses and a set of conditionally independent 3D points given poses, the map lacks consistent dense geometry for downstream tasks and pose estimation cannot benefit from all visual information. Dense SLAM uses and reconstruct all pixels, but the sheer number of variables relative to measurements renders joint optimization infeasible and inference ill-posed. Recently, learned priors over compact depth map representations, such as codes and sparse 2D depths, have enabled joint optimization. However, representations with 2D depth priors have thus far lagged in accuracy, as consistency between depth maps in 3D is not guaranteed. Volumetric representations, such as voxel grids and neural fields, directly enforce dense 3D consistency by construction, but cannot perform real-time joint optimization due to expensive rendering. Furthermore, compact, efficient, and expressive 3D priors for general scenes are not yet suitable. Achieving accurate and consistent 3D geometry in real-time would represent a significant advance for monocular SLAM." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.484, + 0.128 + ], + "angle": 0, + "content": "E. Dexheimer and A.J. Davison" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.149, + 0.489, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.256, + 0.493, + 0.288 + ], + "angle": 0, + "content": "(a) Point cloud reconstruction from the slidingwindow monocular odometry defined by 240 points and 8 keyframes." + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.145, + 0.785, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.256, + 0.786, + 0.287 + ], + "angle": 0, + "content": "(b) Dense point cloud from 40 keyframes across a large office during live operation of a monocular camera." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.299, + 0.785, + 0.356 + ], + "angle": 0, + "content": "Fig. 1: COMO encodes scene geometry via a compact set of 3D anchor points and decodes dense geometry via per-keyframe depth covariance functions. The 3D points visualized in red anchor depth maps together from multiple views while the covariance function generates dense geometry by conditioning on sparse point projections." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.785, + 0.547 + ], + "angle": 0, + "content": "In this work, we propose a representation that achieves the three desired properties: joint optimization of poses and dense geometry, intrinsic 3D consistency, and low-latency, real-time operation. We encode the scene as a compact set of 3D anchor points, which are shared across frames. To decode into dense geometry, 3D points are first projected into keyframes. We then use image-conditioned depth covariance functions [9] per keyframe to generate dense depth maps that intersect each visible 3D point. Depth maps are backprojected into 3D for calculating any-pixel photometric error. All steps are efficient and have analytic Jacobians for second-order optimization. Updates to dense geometry are propagated into our encoded scene representation and ensure suitable regularization for the ill-posed problem of dense reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.785, + 0.653 + ], + "angle": 0, + "content": "Maintaining a compact map in an incremental VO setting requires a compatible frontend. Keeping a small number of points benefits efficiency, while the distribution of points affects expressiveness. Feature-based systems focus on visually distinct features, but this often leads to many points on edges which are not suitable for modeling dense geometry. Therefore, we leverage depth covariance for determining anchor point visibility in new keyframes, actively initializing new 3D points, and encoding the current dense geometry into new 3D points." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.655, + 0.785, + 0.729 + ], + "angle": 0, + "content": "Our system demonstrates improved robustness over methods lacking joint pose and dense geometry optimization, as well as more accurate pose and geometry estimation compared to compact 2D representations. Altogether, we propose a real-time VO system that achieves robust and accurate odometry along with consistent dense geometry as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.73, + 0.581, + 0.744 + ], + "angle": 0, + "content": "In summary, the contributions of our work are:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.751, + 0.786, + 0.78 + ], + "angle": 0, + "content": "- An efficient representation of dense geometry encoded by a compact set of 3D anchor points and decoded by depth covariance functions." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.781, + 0.784, + 0.81 + ], + "angle": 0, + "content": "- A frontend for the compact map that leverages depth covariance for visibility, active initialization, and encoding dense geometry." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.811, + 0.784, + 0.84 + ], + "angle": 0, + "content": "- A real-time monocular visual odometry and mapping system that produces accurate and consistent poses and dense geometry." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.751, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.68, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "COMO" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.389, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.187, + 0.788, + 0.429 + ], + "angle": 0, + "content": "Sparse vs. Dense Visual SLAM Sparse SLAM systems optimize 3D points that are conditionally independent given poses. Exploiting the sparsity inherent in the information form of structure-from-motion is the key to real-time algorithms with a significant number of points [17]. Modern sparse VO methods [10,12,24] and large-scale bundle adjustment [31] rely on the Schur complement which greatly reduces optimization complexity. Sparse systems lack dense reconstruction, which is useful for many downstream tasks. Thus, it is common to first perform pose estimation followed by dense mapping given poses and potentially sparse landmark estimates [20,23,25,28]. MVS approaches estimate depth maps via a reference cost volume and multiple supporting frames, and have shown progress when integrated into a learning pipeline [18]. However, the sparse followed by dense paradigm has two major issues. Dense information is ignored for pose estimation, which can result in failure in many real-world scenarios, while the mapping step is heavily reliant on accurate pose estimation. Joint optimization of poses and dense geometry avoids the brittle two-step process and has the potential to achieve more accurate odometry and mapping [29]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.642 + ], + "angle": 0, + "content": "Dense methods [11, 27] brought the promise of estimating poses and complete geometry simultaneously. However, joint optimization proved difficult due to the ill-posed nature of reconstructing all points. Hand-crafted geometry priors regularize the problem, but disrupt the sparsity and thus the tractability of joint optimization as discussed in [10]. In practice, these methods resort to alternating pose and geometry estimation or move away from second-order optimization. More recently, DROID-SLAM [37] reconstructs depths for all pixels, but similar to sparse systems, does so with no geometric correlation. This results in highly parallelized bundle adjustment, but results in many noisy points and no guarantee of consistent depths within a frame and across frames, as shown in Fig. 2a. All real-time state-of-the-art SLAM systems [6, 10, 37, 38] still rely on the Schur complement, and dense methods with geometric correlation have traditionally not demonstrated comparable pose accuracy, which also limits geometry estimation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Learned Depth Priors for Compact Optimization To achieve consistent poses and geometry, monocular depth priors in SLAM focus on test-time optimization to avoid irreparable errors from single-view depth prediction. Therefore, depth priors in monocular SLAM often predict a subspace of potential depths given an image via a compact latent code [3,7,20], or a depth map basis [13,36]. Even with this flexibility, depth maps are not guaranteed to be consistent across multiple views since latent variables live only in frames as shown in Fig. 2b. While per-pixel depth and keypoint reprojection losses may encourage 3D consistency [7], this may introduce bias and does not scale computationally since it requires per-pixel pairwise frame comparisons. Most importantly, it lacks intrinsic 3D consistency, and depth maps often produce a layering effect in 3D as in Fig. 2b. Predicting a view-based mesh topology permits optimizing depth vertices [4], but it is unclear how to move to a 3D topology." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.484, + 0.127 + ], + "angle": 0, + "content": "E. Dexheimer and A.J. Davison" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.402, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.214, + 0.395, + 0.236 + ], + "angle": 0, + "content": "(a) Dense bundle adjustment with 3D landmarks [37]" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.144, + 0.596, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.421, + 0.214, + 0.582, + 0.236 + ], + "angle": 0, + "content": "(b) Per-frame dense depth priors via compact codes [7]" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.144, + 0.788, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.605, + 0.214, + 0.783, + 0.235 + ], + "angle": 0, + "content": "(c) Ours: 3D points and depth covariance functions" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.247, + 0.788, + 0.359 + ], + "angle": 0, + "content": "Fig. 2: Reconstructions on Replica and geometry properties of different dense map representations. The dotted line represents the true surface. (a) Densely reconstructing a large number of conditionally independent 3D points given poses enables accurate pose estimation and many accurate points, but there is no guarantee of coherent geometry. (b) Depth priors produce smooth depth maps, but even with inter-frame consistency losses, can produce inconsistent geometry and bias preventing global consistency. (c) A compact set of 3D points and depth covariance functions anchor depth maps together, leading to consistent pose estimation and dense geometry." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.789, + 0.562 + ], + "angle": 0, + "content": "In this work, we leverage the recently proposed learned depth covariance function [9]. While [9] introduces an odometry formulation, it only optimizes per-frame 2D inducing depths, which has the same limitations as per-frame codes, and lacks intrinsic 3D consistency. Our key insight is that depth covariance permits sharing latent 3D points across frames, so that our representation lives in 3D rather than separate 2D image planes. While codes and bases are a fixed subspace that cannot guarantee a depth map passes through any set of 3D points, our covariance formulation guarantees that depth maps go through any desired 3D points as shown in Fig. 2c. Therefore, we can anchor depth maps together by construction for dense 3D consistency. We also exploit the covariance function for our frontend in tracking and initializing potentially non-visually distinct points." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.789, + 0.702 + ], + "angle": 0, + "content": "Keyframe-based vs. Volumetric Maps Keyframe-based maps host dense depths in camera frames, which maintain flexible level-of-detail rooted in the camera resolution and focus only on surfaces. However, representing space using per-pixel depths cannot guarantee that inter-frame correspondences refer to the same 3D point. While pairwise depth constraints can mitigate inconsistency [7], these exhaustive per-pixel errors are expensive to compute, lack intrinsic consistency, and may introduce bias when balancing with other losses. Methods that project hosted depths into neighboring frames and predict flow updates [37] are asymmetric with no guarantee of cycle consistency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Volumetric maps live in 3D space and unknown quantities are shared across frames. 3D consistency is achieved by construction but it is challenging to balance fidelity and efficiency. Voxel grids using backward sensor models [15,26] assume known poses and independence between cells, and thus are unsuitable for efficient and consistent optimization. While forward models in voxel grids [39], neural fields [21,35], and 3D Gaussian splatting [16,19] permit consistent geometry and pose optimization, these representations are limited in low-latency monocular VO and SLAM due to expensive rendering, a large number of correlated variables per ray, and alternating and first-order optimization. Volumetric methods are" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.68, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "COMO" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "also subject to a resolution trade-off. Higher resolution permits greater detail at the expense of memory, runtime complexity, and higher levels of noise." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.253 + ], + "angle": 0, + "content": "Our 3D representation achieves the benefits of both frame-based and volumetric maps for odometry and mapping: the per-pixel resolution and efficiency of depth maps with the intrinsic 3D consistency of volumetric representations. The efficient compact-to-dense model permits joint optimization of poses and geometry via any-pixel constraints in real-time." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.273, + 0.523, + 0.291 + ], + "angle": 0, + "content": "3 Compact Mapping Backend" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.302, + 0.784, + 0.377 + ], + "angle": 0, + "content": "The mapping backend maintains a compact set of 3D anchor points along with keyframes hosting per-pixel covariance function parameters. A dense world-centric point cloud is decoded from anchor points and poses, which is refined by second-order minimization of photometric error. VO operates in a sliding-window fashion with a fixed number of keyframes. Fig. 3 provides an overview." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.378, + 0.785, + 0.468 + ], + "angle": 0, + "content": "Quantities in the world frame and camera frame, such as points, are denoted by subscripts \\(\\mathbf{P}_{\\mathrm{W}}\\) and \\(\\mathbf{P}_{\\mathrm{C}}\\), respectively. Anchor points are indexed by superscripts \\(m\\), while dense points are indexed \\(n\\), such as pixels \\(\\mathbf{p}^m\\) and \\(\\mathbf{p}^n\\). Stacked vector forms of sparse and dense variables use capital M and N, as in log-depth vectors \\(\\mathbf{d}_{\\mathrm{M}}\\) and \\(\\mathbf{d}_{\\mathrm{N}}\\). The notation \\(\\|\\mathbf{r}\\|_{\\Sigma}^{2}\\) denotes Mahalanobis distance of residual vector \\(\\mathbf{r}\\) with measurement covariance matrix \\(\\boldsymbol{\\Sigma}\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.487, + 0.609, + 0.502 + ], + "angle": 0, + "content": "3.1 Preliminaries: Depth Covariance Function" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.784, + 0.6 + ], + "angle": 0, + "content": "First, we introduce the depth covariance function from [9]. This covariance function models the distribution over all possible log-depth functions for any finite set of pixels via a Gaussian process (GP) [30]. First, a CNN takes in an RGB image and outputs a per-pixel feature map \\(\\phi\\). Compared to [9], we use a zero-mean GP prior for simplicity, such that the distribution is defined only by the covariance function \\(k\\) that takes in two pixels and respective CNN features:" + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.609, + 0.785, + 0.626 + ], + "angle": 0, + "content": "\\[\nd (\\mathbf {p}) \\sim \\mathcal {G P} (0, k ([ \\mathbf {p}; \\phi (\\mathbf {p}) ], [ \\mathbf {p} ^ {\\prime}; \\phi (\\mathbf {p} ^ {\\prime}) ])). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.633, + 0.785, + 0.77 + ], + "angle": 0, + "content": "We found that the mean log-depth variable proposed in [9] is unnecessary, as the largest eigenvector of the covariance function corresponds to scale of the entire depth map. More information on the exact covariance function \\( k \\) is detailed in [9], but the key idea is that larger covariance is an indicator of similarity in log-depth. Since the covariance function takes in both pixel locations and CNN features, it is nonstationary which permits varying levels of smoothness and discontinuities conditioned on the image content. The zero-mean prior defines a planar depth prior in the absence of observations. To define a covariance matrix \\( \\mathbf{K} \\) for a set of pixels, the \\( (i^{\\mathrm{th}}, j^{\\mathrm{th}}) \\) entry of \\( \\mathbf{K} \\) is filled in by \\( k\\left([ \\mathbf{p}^i; \\phi(\\mathbf{p}^i) ], [ \\mathbf{p}^j; \\phi(\\mathbf{p}^j) ] \\right) \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.504, + 0.804 + ], + "angle": 0, + "content": "3.2 Compact-to-Dense Geometry" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.841 + ], + "angle": 0, + "content": "Given anchor point to keyframe visibility, we generate a dense depth map given anchor points \\(\\mathbf{P}_{\\mathrm{W}}^{m}\\), the pose of keyframe \\(r\\) with rotation \\(\\mathbf{R}_{\\mathrm{WC}_r}\\) and translation" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.484, + 0.128 + ], + "angle": 0, + "content": "E. Dexheimer and A.J. Davison" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.146, + 0.353, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.211, + 0.334, + 0.233 + ], + "angle": 0, + "content": "(a) Anchor point projection" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.144, + 0.495, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.369, + 0.211, + 0.487, + 0.232 + ], + "angle": 0, + "content": "(b) Dense geometry construction" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.144, + 0.638, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.211, + 0.636, + 0.232 + ], + "angle": 0, + "content": "(c) Photometric error calculation" + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.144, + 0.787, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.211, + 0.769, + 0.233 + ], + "angle": 0, + "content": "(d) Pose and geometry update" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.244, + 0.787, + 0.3 + ], + "angle": 0, + "content": "Fig. 3: Overview of compact mapping framework. (a) Anchor points are projected to visible keyframes. (b) Keyframes decode into dense depth and backproject geometry to 3D. (c) Target frames enforce dense photo-consistency. (d) Optimizing dense alignment error leads to updated poses and geometry with greater 3D consistency." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.332, + 0.784, + 0.363 + ], + "angle": 0, + "content": "\\(\\mathbf{t}_{\\mathrm{WC}_r}\\), and the keyframe image. Anchor points are transformed and projected to pixel coordinates \\(\\mathbf{p}\\) via camera projection \\(\\pi\\):" + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.374, + 0.786, + 0.394 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {\\mathrm {C} _ {r}} ^ {m} = \\mathbf {R} _ {\\mathrm {W C} _ {r}} ^ {T} \\left(\\mathbf {P} _ {\\mathrm {W}} ^ {m} - \\mathbf {t} _ {\\mathrm {W C} _ {r}}\\right), \\quad \\mathbf {p} ^ {m} = \\pi \\left(\\mathbf {P} _ {\\mathrm {C} _ {r}} ^ {m}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.405, + 0.788, + 0.527 + ], + "angle": 0, + "content": "which is also illustrated in Fig. 3a. At this point, we create a dense log-depth map from log-depth observations \\( d^{m} \\triangleq \\log \\left( [\\mathbf{P}_{\\mathrm{Cr}}^{m}]_{z} \\right) \\) and projected pixel coordinates \\( \\mathbf{p}^{m} \\). Given an image and the covariance function \\( k \\) for any two pixels from Sec. 3.1, we can construct the covariance matrix between all visible anchor point pixels in \\( \\mathbf{K}_{\\mathrm{MM}} \\), and the covariance between all image pixels and anchor point pixels in \\( \\mathbf{K}_{\\mathrm{NM}} \\). We also stack all training \\( d^{m} \\) and test \\( d^{n} \\) log-depths into \\( \\mathbf{d}_{\\mathrm{M}} \\) and \\( \\mathbf{d}_{\\mathrm{N}} \\), respectively. The GP formulation allows efficient dense log-depth prediction via the linear Gaussian conditioning equation:" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.538, + 0.786, + 0.558 + ], + "angle": 0, + "content": "\\[\n\\mathbf {d} _ {\\mathrm {N}} = \\mathbf {K} _ {\\mathrm {N M}} \\left(\\mathbf {K} _ {\\mathrm {M M}}\\right) ^ {- 1} \\mathbf {d} _ {\\mathrm {M}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.569, + 0.784, + 0.616 + ], + "angle": 0, + "content": "The GP guarantees that each dense log-depth map goes through the anchor points, creating a consistent surface across multiple views. Next, we backproject points for each pixel and transform to world coordinates:" + }, + { + "type": "equation", + "bbox": [ + 0.376, + 0.627, + 0.787, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {\\mathrm {W}} ^ {n} = \\mathbf {R} _ {\\mathrm {W C} _ {r}} \\pi^ {- 1} \\left(\\mathbf {p} ^ {n}, e ^ {d ^ {n}}\\right) + \\mathbf {t} _ {\\mathrm {W C} _ {r}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.748 + ], + "angle": 0, + "content": "where \\( n \\) indexes a query pixel. Note that each keyframe generates a dense depth map via its observed anchor points and covariance parameter feature maps as in Fig. 3b. This compact-to-dense formulation only requires transformations, camera projection, and linear GP conditioning. Therefore, we can efficiently calculate analytical Jacobians for dense per-pixel constraints with respect to poses and anchor points. We show the chain rule for these steps in Section A.1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In practice, we ignore the Jacobians with respect to pixel coordinates when constructing \\(\\mathbf{K}_{\\mathrm{MM}}\\) and \\(\\mathbf{K}_{\\mathrm{NM}}\\) and calculate the covariance matrices once at initialization. In addition to being more efficient, this ensures that depth maps do not undergo unstable changes if a point moves across a depth discontinuity. If a visible anchor point moves behind a camera during optimization, we reinitialize it to the median depth of the keyframe." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.68, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "COMO" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.451, + 0.16 + ], + "angle": 0, + "content": "3.3 Photometric Residuals" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.167, + 0.784, + 0.258 + ], + "angle": 0, + "content": "While only keyframes are used to create dense geometry, additional small baseline support frames are included to improve convergence. We define a set of edges \\(\\mathcal{E}\\) for photometric error that consists of temporally adjacent keyframe-keyframe pairs, and support frames to the two nearest keyframes in time. Given a target image \\(I_{t}\\), which may be either a keyframe or support frame, with pose \\(\\mathbf{T}_{\\mathrm{WC}_t}\\), we first transform world points from keyframe view \\(r\\) and project:" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.264, + 0.786, + 0.282 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {t} ^ {n} = \\pi \\left(\\mathbf {R} _ {\\mathrm {W C} _ {t}} ^ {T} \\left(\\mathbf {P} _ {\\mathrm {W}} ^ {n} - \\mathbf {t} _ {\\mathrm {W C} _ {t}}\\right)\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.286, + 0.784, + 0.332 + ], + "angle": 0, + "content": "Given these projective correspondences and accounting for exposure and global illumination changes via affine brightness parameters, \\( a \\) and \\( b \\), similar to [10], we calculate the photometric error across pixels and frame edges:" + }, + { + "type": "equation", + "bbox": [ + 0.231, + 0.338, + 0.786, + 0.373 + ], + "angle": 0, + "content": "\\[\nE = \\sum_ {r, t \\in \\mathcal {E}} \\sum_ {n} \\left\\| \\mathbf {r} _ {r, t} ^ {n} \\right\\| _ {\\sigma_ {\\mathbf {r}} ^ {2} I} ^ {2}, \\quad \\mathbf {r} _ {r, t} ^ {n} = I _ {t} \\left(\\mathbf {p} _ {t} ^ {n}\\right) + b _ {t} - \\left(\\frac {e ^ {- a _ {r}}}{e ^ {- a _ {t}}} I _ {r} \\left(\\mathbf {p} _ {r} ^ {n}\\right) + b _ {r}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.785, + 0.499 + ], + "angle": 0, + "content": "which is illustrated in Fig. 3c. These constraints depend on each frame's pose and affine parameters, and the anchor points viewed by the reference keyframe. For robustness, we perform iteratively reweighted least squares (IRLS) with a Huber cost function and set the photometric standard deviation \\(\\sigma_{\\mathbf{r}}\\) to be proportional to the median absolute residual [1] to handle diverse scenes with different error scales. Compared to RGB residuals, we found that grayscale images were more efficient and similarly robust. The Jacobian chain rule for photometric residuals is in Section A.2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.518, + 0.451, + 0.532 + ], + "angle": 0, + "content": "3.4 Additional Constraints" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.784, + 0.613 + ], + "angle": 0, + "content": "While photometric error is the only data term, we introduce additional priors to remove ambiguities. Priors on the oldest keyframe's pose and affine brightness parameters remove gauge freedom in the sliding-window. After removing a keyframe from the window, we include priors on its still-observed landmarks to ensure global consistency as a simpler alternative to first-estimate Jacobians [10]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.785, + 0.718 + ], + "angle": 0, + "content": "Since keyframes may have regions unobserved by other frames, which causes anchor points to lack photometric constraints, we include two geometry priors. The first is a weak prior on anchor points to match the log median depth \\( s \\) of the keyframe it was first observed in. The second is a GP prior \\( ||\\mathbf{d}_{\\mathrm{M}} - s||_{\\mathbf{K}_{\\mathrm{MM}}}^2 \\) on every keyframe to regularize anchor points that have large correlation, which gives improved estimates for underconstrained anchor points lying on surfaces with well-constrained ones, such as on textureless walls." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.784, + 0.84 + ], + "angle": 0, + "content": "During optimization, there is ambiguity in whether an anchor point changes depth or moves laterally along a surface, such as when viewing a plane at an angle. Thus, we include a pixel prior encouraging the projection of the point in its initializing keyframe not to change from its first observation, which defines a ray-surface intersection. As mentioned previously, this is more efficient than calculating Jacobians through Sec. 3.2 and more stable since GP inducing point optimization has little benefit when points start in suitable locations [5] as ensured by our frontend in Sec. 4." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.484, + 0.128 + ], + "angle": 0, + "content": "E. Dexheimer and A.J. Davison" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.353, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.224, + 0.337, + 0.245 + ], + "angle": 0, + "content": "(a) Anchor points visible in KF #1" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.144, + 0.497, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.369, + 0.224, + 0.487, + 0.245 + ], + "angle": 0, + "content": "(b) KF #1 depths projected to KF #2" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.144, + 0.639, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.224, + 0.626, + 0.245 + ], + "angle": 0, + "content": "(c) Anchor point matches in KF #2" + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.144, + 0.788, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.224, + 0.782, + 0.245 + ], + "angle": 0, + "content": "(d) Newly initialized points in KF #2" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.256, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Fig. 4: Overview of visibility checks between two keyframes (KF). Matches are shown in blue, rejected matches in red, and newly initialized points in green. Note that occluded edges are rejected, while non-visually distinct points are often matched. New points are allocated to geometrically complex regions while the table is already well-represented." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.339, + 0.49, + 0.354 + ], + "angle": 0, + "content": "3.5 Second-Order Optimization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.788, + 0.451 + ], + "angle": 0, + "content": "We calculate analytical Jacobians for all constraints in the optimization. The compactness of our linear system permits real-time joint optimization of poses and correlated geometry that is infeasible for traditional dense VO. Stacking all Jacobians into \\(\\mathbf{J}\\), weights from IRLS and noise models into the diagonal matrix \\(\\mathbf{W}\\), and residuals into \\(\\mathbf{r}\\), we solve for parameter updates \\(\\Delta \\mathbf{x}\\) shown in Fig. 3d via dense Cholesky factorization of the Gauss-Newton normal equations:" + }, + { + "type": "equation", + "bbox": [ + 0.415, + 0.456, + 0.786, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\left(\\mathbf {J} ^ {T} \\mathbf {W} \\mathbf {J}\\right) \\Delta \\mathbf {x} = - \\mathbf {J} ^ {T} \\mathbf {W r}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.557 + ], + "angle": 0, + "content": "Pose updates are minimally parameterized as Lie-algebra elements \\(\\mathfrak{se}(3)\\) as is standard in SLAM [32]. We found that the most time consuming step of backend optimization is accumulating geometry blocks into the Hessian \\((\\mathbf{J}^T\\mathbf{W}\\mathbf{J})\\). This can be sped up by factoring out constant components of the Jacobians and reducing the dimensionality of the accumulation as detailed in Section A.4." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.576, + 0.409, + 0.591 + ], + "angle": 0, + "content": "4 Visual Frontend" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.604, + 0.788, + 0.68 + ], + "angle": 0, + "content": "The visual frontend tracks the current frame with respect to the map, selects keyframes and support frames for the backend, determines 3D anchor point visibility in new keyframes, and initializes new 3D anchor points. Compared to feature-based systems, the ideal points for an expressive yet compact map may be visually indistinct, so we utilize depth covariance for robust visibility checks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.698, + 0.5, + 0.714 + ], + "angle": 0, + "content": "4.1 Dense Photometric Tracking" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Given the current estimate of the newest keyframe's depth, we optimize the pose and affine brightness parameters using grayscale image alignment in the IRLS framework [2]. For efficiency, we perform inverse compositional tracking and use a multi-scale image pyramid. After tracking, a new keyframe is added to the backend if there is either significant translation relative to the median scene depth or if the number of unique projected pixels in the tracked frame falls below a threshold to handle rotation and occlusion. Support frames are added via modified thresholds according to the desired frequency between keyframes." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.68, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "COMO" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.456, + 0.163 + ], + "angle": 0, + "content": "4.2 Anchor Point Visibility" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.171, + 0.788, + 0.291 + ], + "angle": 0, + "content": "When a new keyframe is added, we perform explicit visibility checks to 3D anchor points viewed in the last keyframe to achieve 3D consistency. Note that we only check for visibility as anchor points have no explicit constraint on its location in the image. Compared to feature-based methods [24], anchor points often lack visual distinctiveness, so matching via image information is problematic. Instead, we leverage the interpretability of the depth covariance function to determine visibility. In general, an anchor point should be compatible with all depth covariance functions that use it to generate dense geometry." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.292, + 0.789, + 0.413 + ], + "angle": 0, + "content": "First, the 3D anchor points and dense depth map from the last keyframe, shown in Fig. 4a, are projected into the current frame, shown in Fig. 4b. Given the projected sparse and dense pixel coordinates, we may query covariance parameters \\(\\phi (\\mathbf{p})\\) and construct \\(\\mathbf{K}_{\\mathrm{MM}}\\) and \\(\\mathbf{K}_{\\mathrm{MN}}\\) for the new keyframe. We leverage the linearity of the GP to compress dense log-depth observations into the sparse projected anchor point coordinates. Assuming conditional independence of log-depth observations \\(\\mathbf{d}_n\\) given the sparse log-depths \\(\\mathbf{d}_m\\), we can solve an efficient least-squares problem:" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.422, + 0.785, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {d} _ {\\mathrm {M}}} \\left| \\left| \\mathbf {K} _ {\\mathrm {N M}} \\left(\\mathbf {K} _ {\\mathrm {M M}}\\right) ^ {- 1} \\mathbf {d} _ {\\mathrm {M}} - \\mathbf {d} _ {\\mathrm {N}} \\right| \\right| _ {\\sigma_ {d} ^ {2}} ^ {2} + \\left| \\left| \\mathbf {d} _ {\\mathrm {M}} \\right| \\right| _ {\\mathbf {K} _ {\\mathrm {M M}}} ^ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.788, + 0.545 + ], + "angle": 0, + "content": "where the first conditional term ensures that \\(\\mathbf{d}_{\\mathrm{M}}\\) fits the projected dense observations with standard deviation \\(\\sigma_{d}\\), while the second prior term regularizes sparse log-depths according to the GP training covariance \\(\\mathbf{K}_{\\mathrm{MM}}\\). We perform a consistency check on the log-depth difference between the current 3D anchor points and those solved in Eq. (8), and prune points at depth discontinuities. This yields potential matches the are consistent between the two keyframes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.788, + 0.651 + ], + "angle": 0, + "content": "In practice, we set a maximum number of visible anchor points per-frame to allow efficient batched mapping operations on the GPU. If all points are matched, it could prevent new ones from being initialized and parts of the scene would lack capacity to adequately represent geometry. Thus, we perform pruning of matches so that we have a suitable distribution of points before initializing new anchors. Similar to the GP conditional mean from Eq. (3), we may also obtain conditional variances of all pixels with respect to the first \\( j \\) pixels via:" + }, + { + "type": "equation", + "bbox": [ + 0.326, + 0.662, + 0.785, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\sigma} _ {j} ^ {2} = \\operatorname {d i a g} \\left[ \\mathbf {K} _ {\\mathrm {N N}} \\right] - \\mathbf {K} _ {\\mathrm {N M} _ {1: j}} \\left(\\mathbf {K} _ {\\mathrm {M} _ {1: j} \\mathrm {M} _ {1: j}}\\right) ^ {- 1} \\mathbf {K} _ {\\mathrm {M} _ {1: j} \\mathrm {N}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We perform active sampling via conditional variance reduction (CVR) [14, 22], where at the \\((j + 1)^{\\mathrm{th}}\\) iteration, the pixel in domain \\(P\\) with the largest variance is selected \\(\\mathbf{p}_{j + 1} = \\arg \\max_{\\mathbf{p}\\in P}\\sigma_j^2 (\\mathbf{p})\\). We actively sample from the set of potential visible points until a variance threshold is reached, which gives the final set of visible anchor points for the current keyframe. Compared to [9], we found that a minimum distance threshold between pixel locations and from the image border is required to avoid accumulating pixels near the same discontinuity. This is essential for improving depth estimates and maintaining a good set of candidate points, as otherwise many points near edges may be occluded at the same time. We show the set of matches between two keyframes in Fig. 4c." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.484, + 0.128 + ], + "angle": 0, + "content": "E. Dexheimer and A.J. Davison" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.317, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.206, + 0.316, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.322, + 0.146, + 0.419, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.322, + 0.206, + 0.419, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.146, + 0.522, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.206, + 0.522, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.146, + 0.624, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.206, + 0.624, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.145, + 0.784, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.251, + 0.676, + 0.261 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.275, + 0.788, + 0.345 + ], + "angle": 0, + "content": "Fig. 5: Visualization of sliding-window odometry on TUM \\(fr2/desk\\). Top images show tracked and newly initialized anchor points in blue and red, respectively. Depth maps are below, while the dense point cloud with 248 anchor points is shown on the right. Note the consistent point tracks on textureless surfaces, such as the table and plate, as well as the long-term tracks on the cube, mug, and book." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.373, + 0.484, + 0.387 + ], + "angle": 0, + "content": "4.3 Anchor Point Initialization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.784, + 0.472 + ], + "angle": 0, + "content": "As the camera explores unobserved parts of the scene, new anchor points must be initialized. First, we again perform active sampling, but now across all pixels given the current anchor point locations from the previous section. This ensures new points will be initialized for parts of the scene that are unobserved, lack matches, or require additional geometric capacity, as shown in Fig. 4d." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.785, + 0.535 + ], + "angle": 0, + "content": "Once the pixel locations of new anchor points are known, we initialize the log-depth and backproject into 3D. Defining the set of known log-depths from the anchor points \\(\\mathbf{d}_{\\mathrm{m}_1}\\), the unknown log-depths of newly initialized points \\(\\mathbf{d}_{\\mathrm{m}_2}\\) and all log-depths as \\(\\mathbf{d}_{\\mathrm{M}} = (\\mathbf{d}_{\\mathrm{m}_1}^T,\\mathbf{d}_{\\mathrm{m}_2}^T)^T\\), we solve another least-squares problem:" + }, + { + "type": "equation", + "bbox": [ + 0.321, + 0.556, + 0.785, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {d} _ {\\mathrm {m} _ {2}}} \\left\\| \\mathbf {K} _ {\\mathrm {N M}} \\left(\\mathbf {K} _ {\\mathrm {M M}}\\right) ^ {- 1} \\mathbf {d} _ {\\mathrm {M}} - \\mathbf {d} _ {\\mathrm {N}} \\right\\| _ {\\sigma_ {d} ^ {2}} ^ {2} + \\left\\| \\mathbf {d} _ {\\mathrm {m} _ {2}} - s \\right\\| _ {\\sigma_ {s} ^ {2}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.788, + 0.759 + ], + "angle": 0, + "content": "where the first term promotes fitting reprojected dense depth observations given matched anchor log-depths and the new log-depth initializations, while the second term is a prior that new log-depths are close to the log median depth \\( s \\) of the previous keyframe. We found that using an isotropic prior with standard deviation \\( \\sigma_d \\), rather than \\( \\mathbf{K}_{\\mathrm{MM}} \\), avoids bias for parts of the image that were not previously observed and would otherwise be underconstrained. We define \\( \\sigma_d \\) as the standard deviation of the residuals from Eq. (8) to better calibrate the model. Given the new pixel locations, log-depth initializations, and the new keyframe's pose estimate, 3D points are initialized for the backend. A qualitative example of tracked anchor points, depth maps, and reconstruction is shown in Fig. 5." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.78, + 0.376, + 0.797 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "We evaluate both the pose and geometry estimation of our method on multiple datasets. For sparse methods, we compare against feature-based ORB-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.68, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "COMO" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.144, + 0.768, + 0.218 + ], + "angle": 0, + "content": "
office0office1office2office3office4room0room1room2mean
SORB-SLAM3 [6]0.40.38.50.44.50.30.30.41.9
DSO [10]0.30.50.30.25.30.20.80.21.0
DDeepFactors [7]42.938.645.641.450.132.426.144.140.2
DROID-VO [37]5.34.16.511.27.19.04.87.06.9
COMO-NC2.95.07.46.18.14.44.53.75.3
COMO2.42.44.33.65.52.52.73.43.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.218, + 0.786, + 0.247 + ], + "angle": 0, + "content": "Table 1: Mean ATE (cm) over 5 runs for monocular VO on the Replica dataset. S and D in the first column refer to sparse and dense methods, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.285, + 0.789, + 0.469 + ], + "angle": 0, + "content": "SLAM3 [6], direct DSO [10], and learning-based DPVO [38] in trajectory estimation. We also compare against a representative set of dense methods. TANDEM [18] uses DSO for odometry and fuses learned MVS predictions into a TSDF for tracking. MonoGS [19] is a 3D Gaussian splatting-based monocular SLAM system. DeepFactors [7] is a code-based system with photometric, geometric, and keypoint losses to promote consistency across depth maps. DROIDVO [37] uses learned correspondence and flow updates for bundle adjustment, such that all depths are reconstructed but lack geometric correlation. DepthCov [9] optimizes sparse 2D depths under the depth covariance framework, and in a related manner, COMO-NC (no correspondence) is our full framework but no 3D anchor points are tracked across frames. We disable loop closure for DeepFactors and DROID-SLAM to isolate the odometry and mapping performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.487, + 0.457, + 0.503 + ], + "angle": 0, + "content": "5.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.789, + 0.631 + ], + "angle": 0, + "content": "We use a fixed configuration across all three datasets to demonstrate the robustness of the system. We use the off-the-shelf depth covariance function [9] that was trained on the ScanNet training set. Our sliding window odometry uses 9 keyframes and 3 support frames between keyframes for a total of 24. Each keyframe views a maximum of 64 anchor points and initializes new points if it fails to track all from the previous keyframe. We use 256x192 images for all operations, and the mapping backend samples pixels with the highest gradient magnitude in 4x4 images patches for photometric error." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.651, + 0.444, + 0.667 + ], + "angle": 0, + "content": "5.2 Trajectory Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Replica We first test our monocular odometry on the Replica [33] sequences recorded in [35], which provides a photorealistic environment with ground-truth geometry. As shown in Tab. 1, sparse methods without priors perform best, as conditions are favorable for SLAM: few image artifacts, no exposure changes, and sufficient baselines for multi-view geometry. Among methods with learned priors, COMO achieves the lowest ATE. The representation has little bias as compared to the code-based DeepFactors, which drifts despite photometric, geometric, and keypoint constraints. Sharing 3D points between frames in COMO shows significant improvement over COMO-NC which disables correspondence. TUM The TUM RGBD dataset [34] is a challenging dataset for monocular VO due to significant motion blur, exposure changes, rolling shutter artifacts," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.483, + 0.127 + ], + "angle": 0, + "content": "E. Dexheimer and A.J. Davison" + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.144, + 0.782, + 0.268 + ], + "angle": 0, + "content": "
fr1fr2fr3 longmean
360deskdesk2plantroomrpyteddyxyzxyzdesk
SORB-SLAM3 [6]X2.0X11.8X5.6X1.00.51.31.7X
DSO [10]X27.266.06.058.6XX3.80.32.29.9X
DPVO [38]13.19.46.53.039.83.56.21.30.53.55.58.4
DTANDEM [18]X4.333.7XX4.943.12.40.32.08.3X
MonoGS [19]14.26.374.09.364.93.435.61.64.5133.13.331.8
DeepFactors [7]17.915.920.231.938.33.856.05.98.426.349.024.9
DepthCov [9]12.85.64.826.125.75.247.55.61.215.968.819.9
DROID-VO [37]15.75.211.16.033.43.219.15.610.77.97.311.4
COMO-NC16.14.210.919.328.65.268.74.10.78.846.819.4
COMO12.94.99.513.827.04.824.54.00.76.310.510.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.269, + 0.784, + 0.295 + ], + "angle": 0, + "content": "Table 2: Mean ATE (cm) over 5 runs for monocular VO on the TUM dataset. S and D in the first column refer to sparse and dense methods, respectively." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.341, + 0.787, + 0.657 + ], + "angle": 0, + "content": "and heavy rotations. Results comparing our method against state-of-the-art VO methods on 11 sequences is shown in Tab. 2. For sparse methods, ORB-SLAM3 and DSO both lose tracking on multiple sequences, while DPVO, which uses learned context for tracking 64 sparse patches and performing bundle adjustment, performs best among all methods. In this work, we are focused on also reconstructing dense geometry, but it would be interesting to combine the learned features of DPVO with our 3D representation. For dense methods, TANDEM fails due to relying on DSO for odometry, which can still fail despite using a TSDF with MVS fusion for tracking. This demonstrates the need for joint optimization of poses and dense image information. While MonoGS can produce accurate trajectories for some sequences, the lack of priors reduces its robustness, and volumetric rendering is slower than all depth-based systems. Despite promoting consistency between neighboring depth maps, DeepFactors has relatively large error. Both DepthCov and COMO-NC are similar in terms of optimizing sparse depths in each frame, while leveraging a true 3D representation in COMO reduces error by almost \\(50\\%\\). COMO outperforms DROID-VO, which uses learned features for matching and reconstructs all pixels from downsized images via sparse bundle adjustment. As the assumptions of brightness constancy are violated in many real-world datasets, learned features could further the accuracy of COMO, but it is interesting that our compact representation and simple photometric error can produce the lowest ATE of dense methods." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.66, + 0.786, + 0.84 + ], + "angle": 0, + "content": "ScanNet Test Lastly, we evaluate the trajectory error on 18 diverse medium-length scenes from the ScanNet [8] test set. Sequences include homes, offices, schools, businesses, and outdoors. Methods that fail are assigned a maximum \\(100\\mathrm{cm}\\) error for averages. Sequences have challenging rotational motion, high image noise, and specular surfaces. We summarize the results in Fig. 6, which shows the number of sequences below varying ATE thresholds, while Tab. 3 contains summary statistics. COMO has the lowest mean ATE and area-underthe-curve (AUC). While DPVO has a higher number of trajectories under low ATE thresholds, it also shows high variance results on some sequences. Despite using photometric error that is often violated due to image noise and specular surfaces, our representation with 3D consistency proves valuable in challenging real-world data. An example reconstruction is shown in Fig. 1a." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.68, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "COMO" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.616, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.62, + 0.147, + 0.775, + 0.241 + ], + "angle": 0, + "content": "
MethodATEAUC
ORB-SLAM379.91.63
DSO54.72.46
DPVO15.16.34
TANDEM41.73.79
DeepFactors36.12.97
DROID-VO13.96.51
COMO-NC15.86.16
COMO13.06.67
" + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.275, + 0.593, + 0.304 + ], + "angle": 0, + "content": "Fig. 6: Number of successful trajectories below ATE threshold for selected ScanNet test sequences." + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.275, + 0.782, + 0.304 + ], + "angle": 0, + "content": "Table 3: ScanNet Mean ATE (cm) and AUC." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.347, + 0.376, + 0.396 + ], + "angle": 0, + "content": "
MethodReplicaScanNet
TANDEMN/A0.325
DeepFactors0.2630.232
DROID-VO1.1291.389
COMO-NC0.0690.155
COMO0.0460.128
" + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.321, + 0.58, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.322, + 0.781, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.433, + 0.374, + 0.46 + ], + "angle": 0, + "content": "Table 4: Depth absolute relative error." + }, + { + "type": "image_caption", + "bbox": [ + 0.384, + 0.433, + 0.783, + 0.462 + ], + "angle": 0, + "content": "Fig. 7: ScanNet depth accuracy and consistency metrics for fraction of pixels falling below threshold \\(\\delta\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.493, + 0.438, + 0.509 + ], + "angle": 0, + "content": "5.3 Geometry Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.787, + 0.719 + ], + "angle": 0, + "content": "Since we are interested in both consistent pose estimation and dense geometry, we evaluate dense depth predictions on Replica and ScanNet. To evaluate accuracy, we perform the same global scale alignment that is used to produce ATE for monocular methods and transform estimated poses and depth maps to be aligned to the metric ground-truth. Beyond global accuracy, we also measure consistency, which checks whether neighboring frames viewing the same 3D points have consistent depth estimates. Note that consistency does not care about the correctness of depth. To check valid pixels, we backproject ground-truth depth maps and project these points into neighboring frames, and check if the depths agree within \\(1\\mathrm{cm}\\). Then, for all valid 3D correspondences, we calculate metrics between pairs of estimated depth images. If \\(N\\) is the number of keyframes for that sequence, then \\(N\\) depth images are used for accuracy metrics, while \\(2(N - 1)\\) depth images are used for consistency since pairs require both directions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "For estimated depth \\(\\hat{D}\\) and ground-truth depth \\(D\\) we show the mean absolute relative error \\((1 / N)\\sum_{i}|\\hat{D}_{i} - D_{i}| / D_{i}\\) across all sequences for Replica and ScanNet in Tab. 4. We do not evaluate TANDEM on Replica since it is used in the training data. COMO outperforms other methods due to its ability to jointly optimize dense, correlated geometry along with poses. Despite accurate pose estimation, DROID-VO has outliers with significant depth errors. To further evaluate the distribution of depth accuracy and remove the effect of large outliers, in Fig. 7, we use the \\(\\delta\\) threshold commonly used in depth estimation, which measures what" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.484, + 0.128 + ], + "angle": 0, + "content": "E. Dexheimer and A.J. Davison" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.144, + 0.778, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.215, + 0.757, + 0.23 + ], + "angle": 0, + "content": "Fig. 8: Mesh and surface normals from TSDF fusion of a live desk sequence." + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.241, + 0.764, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.265, + 0.786, + 0.293 + ], + "angle": 0, + "content": "Table 5: Runtime in seconds of different methods on TUM \\(fr3/long\\) using an RTX 3080. COMO SP and MP denote single and multiprocessing, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.332, + 0.785, + 0.414 + ], + "angle": 0, + "content": "fraction of pixels satisfy \\(\\max \\left(\\hat{D} / D, D / \\hat{D}\\right) < \\delta\\). Interestingly, accurate depth maps do not necessarily imply consistency. For example, DROID-VO has the second best accuracy behind COMO, while DeepFactors has the second most consistent geometry. COMO achieves both the best accuracy and consistency without explicit geometric and keypoint constraints." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.433, + 0.586, + 0.449 + ], + "angle": 0, + "content": "5.4 Real-time System and Runtime Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.785, + 0.562 + ], + "angle": 0, + "content": "When evaluating our system, we use a sequential, single-threaded mode that runs at around 14 FPS with an RTX 3080. For live demos and reconstructions, we also develop a multiprocessing version that runs at 35-50 FPS as shown in the FPS counter in the top right of the supplementary video, with tracking flexible to be on either CPU or GPU. We show examples of reconstructions while operating a monocular camera in real-time. Fig. 1b displays a large-scale raw point cloud of an office, while Fig. 8 displays consistent TSDF fusion of a desk sequence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.563, + 0.787, + 0.653 + ], + "angle": 0, + "content": "Average times for tracking a frame, adding a keyframe (KF), and updating the map are 20.1, 50.4, and \\(39.8~\\mathrm{ms}\\). Keyframing is the bottleneck during fast motion, but 35-50 FPS in multiprocessing mode for normal motion is standard as shown in the video. We measure the wall time of several methods in Tab. 5. COMO achieves real-time processing with its efficient representation, but further optimization is still possible since almost all code is implemented in PyTorch." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.675, + 0.36, + 0.691 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this work, we present a compact representation of 3D anchor points and frame-based depth covariance functions to achieve real-time estimation of consistent poses and geometry. Future work includes training the depth covariance function on more diverse data and replacing photometric constraints with the more robust learned correspondence from DROID-VO. Lastly, we believe the anchor point representation could be used in a full map-centric SLAM system with relocalization, which would allow continually improving geometry upon revisiting. With the advance of learned priors, exploiting geometric correlation is an interesting direction to achieve efficient, robust, and consistent algorithms." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.681, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "COMO" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.146, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.237 + ], + "angle": 0, + "content": "This research is funded by EPSRC Prosperity Partnerships (EP/S036636/1) and Dyson Technology Ltd. We would like to thank Riku Murai, Hidenobu Matsuki, Gwangbin Bae, and members of the Dyson Robotics Lab for insightful discussions and feedback." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.259, + 0.323, + 0.275 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.317 + ], + "angle": 0, + "content": "1. Alismail, H.: Direct pose estimation and refinement. Ph.D. thesis, Robotics Institute, Carnegie Mellon University. (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.344 + ], + "angle": 0, + "content": "2. Baker, S., Matthews, I.: Lucas-Kanade 20 years on: A unifying framework. International Journal of Computer Vision (IJCV) (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.346, + 0.785, + 0.4 + ], + "angle": 0, + "content": "3. Bloesch, M., Czarnowski, J., Clark, R., Leutenegger, S., Davison, A.J.: CodeSLAM - learning a compact, optimisable representation for dense visual SLAM. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.401, + 0.785, + 0.441 + ], + "angle": 0, + "content": "4. Bloesch, M., Laidlow, T., Clark, R., Leutenegger, S., Davison, A.: Learning meshes for dense visual SLAM. In: Proceedings of the International Conference on Computer Vision (ICCV) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.442, + 0.785, + 0.482 + ], + "angle": 0, + "content": "5. Burt, D.R., Rasmussen, C.E., van der Wilk, M.: Convergence of sparse variational inference in Gaussian processes regression. The Journal of Machine Learning Research (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.484, + 0.785, + 0.523 + ], + "angle": 0, + "content": "6. Campos, C., Elvira, R., Rodríguez, J.J.G., M. Montiel, J.M., D. Tardós, J.: ORBSLAM3: An accurate open-source library for visual, visual-inertial, and multimap SLAM. IEEE Transactions on Robotics (T-RO) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.525, + 0.785, + 0.564 + ], + "angle": 0, + "content": "7. Czarnowski, J., Laidlow, T., Clark, R., Davison, A.J.: DeepFactors: Real-time probabilistic dense monocular SLAM. IEEE Robotics and Automation Letters (RA-L) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.607 + ], + "angle": 0, + "content": "8. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.608, + 0.785, + 0.647 + ], + "angle": 0, + "content": "9. Dexheimer, E., Davison, A.J.: Learning a depth covariance function. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.649, + 0.785, + 0.674 + ], + "angle": 0, + "content": "10. Engel, J., Koltun, V., Cremers, D.: Direct sparse odometry. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.676, + 0.785, + 0.716 + ], + "angle": 0, + "content": "1. Engel, J., Schöps, T., Cremers, D.: LSD-SLAM: Large-scale direct monocular SLAM. In: Proceedings of the European Conference on Computer Vision (ECCV) (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.758 + ], + "angle": 0, + "content": "2. Forster, C., Zhang, Z., Gassner, M., Werlberger, M., Scaramuzza, D.: SVO: Semidirect visual odometry for monocular and multicamera systems. IEEE Transactions on Robotics (T-RO) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "3. Graham, B., Novotny, D.: RidgeSfM: Structure from motion via robust pairwise matching under depth uncertainty. In: Proceedings of the International Conference on 3D Vision (3DV) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "4. Guestrin, C., Krause, A., Singh, A.P.: Near-optimal sensor placements in Gaussian processes. In: Proceedings of the International Conference on Machine Learning (ICML) (2005)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.484, + 0.127 + ], + "angle": 0, + "content": "E. Dexheimer and A.J. Davison" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.188 + ], + "angle": 0, + "content": "15. Hornung, A., Wurm, K.M., Bennewitz, M., Stachniss, C., Burgard, W.: OctoMap: An efficient probabilistic 3D mapping framework based on octrees. Autonomous Robots (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "16. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3D Gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.785, + 0.271 + ], + "angle": 0, + "content": "17. Klein, G., Murray, D.: Parallel tracking and mapping for small AR workspaces. In: Proceedings of the International Symposium on Mixed and Augmented Reality (ISMAR) (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.785, + 0.311 + ], + "angle": 0, + "content": "18. Koestler, L., Yang, N., Zeller, N., Cremers, D.: TANDEM: Tracking and dense mapping in real-time using deep multi-view stereo. In: Conference on Robot Learning (CoRL) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.311, + 0.785, + 0.352 + ], + "angle": 0, + "content": "19. Matsuki, H., Murai, R., Kelly, P.H.J., Davison, A.J.: Gaussian Splatting SLAM. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.393 + ], + "angle": 0, + "content": "20. Matsuki, H., Scona, R., Czarnowski, J., Davison, A.J.: CodeMapping: Real-time dense mapping for sparse SLAM using compact scene representations. IEEE Robotics and Automation Letters (RA-L) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.434 + ], + "angle": 0, + "content": "21. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: Proceedings of the European Conference on Computer Vision (ECCV) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.473 + ], + "angle": 0, + "content": "22. Moss, H.B., Ober, S.W., Picheny, V.: Inducing point allocation for sparse Gaussian processes in high-throughput bayesian optimisation. In: International Conference on Artificial Intelligence and Statistics (AISTATS) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "23. Mur-Artal, R., Tardós, J.D.: Probabilistic semi-dense mapping from highly accurate feature-based monocular SLAM. In: Proceedings of Robotics: Science and Systems (RSS) (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.555 + ], + "angle": 0, + "content": "24. Mur-Artal, R., Tardós, J.D.: ORB-SLAM2: An open-source SLAM system for monocular, stereo, and RGB-D cameras. IEEE Transactions on Robotics (T-RO) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.785, + 0.596 + ], + "angle": 0, + "content": "25. Newcombe, R.A., Davison, A.J.: Live dense reconstruction with a single moving camera. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.785, + 0.651 + ], + "angle": 0, + "content": "26. Newcombe, R.A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohi, P., Shotton, J., Hodges, S., Fitzgibbon, A.: KinectFusion: Real-time dense surface mapping and tracking. In: Proceedings of the International Symposium on Mixed and Augmented Reality (ISMAR) (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.785, + 0.691 + ], + "angle": 0, + "content": "27. Newcombe, R.A., Lovegrove, S.J., Davison, A.J.: DTAM: Dense tracking and mapping in real-time. In: Proceedings of the International Conference on Computer Vision (ICCV) (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.732 + ], + "angle": 0, + "content": "28. Pizzoli, M., Forster, C., Scaramuzza, D.: REMODE: Probabilistic, monocular dense reconstruction in real time. In: Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.785, + 0.772 + ], + "angle": 0, + "content": "29. Platinsky, L., Davison, A.J., Leutenegger, S.: Monocular visual odometry: Sparse joint optimisation or dense alternation? In: Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "30. Rasmussen, C.E., Williams, C.K.I.: Gaussian Processes for Machine Learning. MIT Press (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "31. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.681, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "COMO" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "32. Sola, J., Deray, J., Atchuthan, D.: A micro lie theory for state estimation in robotics. arXiv preprint arXiv:1812.01537 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.259 + ], + "angle": 0, + "content": "33. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., Clarkson, A., Yan, M., Budge, B., Yan, Y., Pan, X., Yon, J., Zou, Y., Leon, K., Carter, N., Brales, J., Gillingham, T., Mueggler, E., Pesqueira, L., Savva, M., Batra, D., Strasdat, H.M., Nardi, R.D., Goesele, M., Lovegrove, S., Newcombe, R.: The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.3 + ], + "angle": 0, + "content": "34. Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for the evaluation of RGB-D SLAM systems. In: Proceedings of the IEEE/RSJ Conference on Intelligent Robots and Systems (IROS) (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.342 + ], + "angle": 0, + "content": "35. Sucar, E., Liu, S., Ortiz, J., Davison, A.J.: iMAP: Implicit mapping and positioning in real-time. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.37 + ], + "angle": 0, + "content": "36. Tang, C., Tan, P.: BA-Net: Dense bundle adjustment networks. In: Proceedings of the International Conference on Learning Representations (ICLR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.398 + ], + "angle": 0, + "content": "37. Teed, Z., Deng, J.: DROID-SLAM: Deep visual SLAM for monocular, stereo, and RGB-D cameras. In: Neural Information Processing Systems (NeurIPS) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.425 + ], + "angle": 0, + "content": "38. Teed, Z., Lipson, L., Deng, J.: Deep patch visual odometry. In: Neural Information Processing Systems (NeurIPS) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.453 + ], + "angle": 0, + "content": "39. Thrun, S.: Learning occupancy grids with forward models. In: Proceedings of the IEEE/RSJ Conference on Intelligent Robots and Systems (IROS) (2001)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.453 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/COMO_ Compact Mapping and Odometry/005f4f87-73cd-41e4-ad7b-b179ff03449c_origin.pdf b/2024/COMO_ Compact Mapping and Odometry/005f4f87-73cd-41e4-ad7b-b179ff03449c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d54f005b1bf072d1d84a5cc6c6caf787dd30b43d --- /dev/null +++ b/2024/COMO_ Compact Mapping and Odometry/005f4f87-73cd-41e4-ad7b-b179ff03449c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51dc65c4aa7275eeb1396e21d083729f77db9bc2140b02f1313b53678f61d5d6 +size 7690439 diff --git a/2024/COMO_ Compact Mapping and Odometry/full.md b/2024/COMO_ Compact Mapping and Odometry/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dcb3ba3f3e47e282bfa4205223178dac137fe0fd --- /dev/null +++ b/2024/COMO_ Compact Mapping and Odometry/full.md @@ -0,0 +1,336 @@ +# COMO: Compact Mapping and Odometry + +Eric Dexheimer1 and Andrew J. Davison1 + +Dyson Robotics Lab, Imperial College London {e.dexheimer21, a.davison}@imperial.ac.uk + +Abstract. We present COMO, a real-time monocular mapping and odometry system that encodes dense geometry via a compact set of 3D anchor points. Decoding anchor point projections into dense geometry via per-keyframe depth covariance functions guarantees that depth maps are joined together at visible anchor points. The representation enables joint optimization of camera poses and dense geometry, intrinsic 3D consistency, and efficient second-order inference. To maintain a compact yet expressive map, we introduce a frontend that leverages the covariance function for tracking and initializing potentially visually indistinct 3D points across frames. Altogether, we introduce a real-time system capable of estimating accurate poses and consistent geometry. + +Keywords: SLAM $\cdot$ Multi-view geometry $\cdot$ 3D representations + +# 1 Introduction + +Achieving accurate and consistent poses and dense geometry from monocular images in real-time is a challenging yet essential undertaking to push forward state-of-the-art in robotics and augmented reality. Monocular cameras are the key to achieving low-cost, energy-efficient, and compact intelligent platforms. While images contain rich visual information, reconstructing the world is challenging due to the lack of direct geometric observations. + +The ideal representation for real-time monocular visual odometry (VO) and simultaneous localization and mapping (SLAM) remains elusive. While sparse methods jointly optimize camera poses and a set of conditionally independent 3D points given poses, the map lacks consistent dense geometry for downstream tasks and pose estimation cannot benefit from all visual information. Dense SLAM uses and reconstruct all pixels, but the sheer number of variables relative to measurements renders joint optimization infeasible and inference ill-posed. Recently, learned priors over compact depth map representations, such as codes and sparse 2D depths, have enabled joint optimization. However, representations with 2D depth priors have thus far lagged in accuracy, as consistency between depth maps in 3D is not guaranteed. Volumetric representations, such as voxel grids and neural fields, directly enforce dense 3D consistency by construction, but cannot perform real-time joint optimization due to expensive rendering. Furthermore, compact, efficient, and expressive 3D priors for general scenes are not yet suitable. Achieving accurate and consistent 3D geometry in real-time would represent a significant advance for monocular SLAM. + +![](images/ea6a588415295b00661b95c8da0edc65005b10b9ae85a74954e87a208bdd08f9.jpg) +(a) Point cloud reconstruction from the slidingwindow monocular odometry defined by 240 points and 8 keyframes. + +![](images/141eb0d4228048625b58d4e3c468459e7961b84a3a79bb602a0136ef30099705.jpg) +(b) Dense point cloud from 40 keyframes across a large office during live operation of a monocular camera. +Fig. 1: COMO encodes scene geometry via a compact set of 3D anchor points and decodes dense geometry via per-keyframe depth covariance functions. The 3D points visualized in red anchor depth maps together from multiple views while the covariance function generates dense geometry by conditioning on sparse point projections. + +In this work, we propose a representation that achieves the three desired properties: joint optimization of poses and dense geometry, intrinsic 3D consistency, and low-latency, real-time operation. We encode the scene as a compact set of 3D anchor points, which are shared across frames. To decode into dense geometry, 3D points are first projected into keyframes. We then use image-conditioned depth covariance functions [9] per keyframe to generate dense depth maps that intersect each visible 3D point. Depth maps are backprojected into 3D for calculating any-pixel photometric error. All steps are efficient and have analytic Jacobians for second-order optimization. Updates to dense geometry are propagated into our encoded scene representation and ensure suitable regularization for the ill-posed problem of dense reconstruction. + +Maintaining a compact map in an incremental VO setting requires a compatible frontend. Keeping a small number of points benefits efficiency, while the distribution of points affects expressiveness. Feature-based systems focus on visually distinct features, but this often leads to many points on edges which are not suitable for modeling dense geometry. Therefore, we leverage depth covariance for determining anchor point visibility in new keyframes, actively initializing new 3D points, and encoding the current dense geometry into new 3D points. + +Our system demonstrates improved robustness over methods lacking joint pose and dense geometry optimization, as well as more accurate pose and geometry estimation compared to compact 2D representations. Altogether, we propose a real-time VO system that achieves robust and accurate odometry along with consistent dense geometry as shown in Fig. 1. + +In summary, the contributions of our work are: + +- An efficient representation of dense geometry encoded by a compact set of 3D anchor points and decoded by depth covariance functions. +- A frontend for the compact map that leverages depth covariance for visibility, active initialization, and encoding dense geometry. +- A real-time monocular visual odometry and mapping system that produces accurate and consistent poses and dense geometry. + +# 2 Related Work + +Sparse vs. Dense Visual SLAM Sparse SLAM systems optimize 3D points that are conditionally independent given poses. Exploiting the sparsity inherent in the information form of structure-from-motion is the key to real-time algorithms with a significant number of points [17]. Modern sparse VO methods [10,12,24] and large-scale bundle adjustment [31] rely on the Schur complement which greatly reduces optimization complexity. Sparse systems lack dense reconstruction, which is useful for many downstream tasks. Thus, it is common to first perform pose estimation followed by dense mapping given poses and potentially sparse landmark estimates [20,23,25,28]. MVS approaches estimate depth maps via a reference cost volume and multiple supporting frames, and have shown progress when integrated into a learning pipeline [18]. However, the sparse followed by dense paradigm has two major issues. Dense information is ignored for pose estimation, which can result in failure in many real-world scenarios, while the mapping step is heavily reliant on accurate pose estimation. Joint optimization of poses and dense geometry avoids the brittle two-step process and has the potential to achieve more accurate odometry and mapping [29]. + +Dense methods [11, 27] brought the promise of estimating poses and complete geometry simultaneously. However, joint optimization proved difficult due to the ill-posed nature of reconstructing all points. Hand-crafted geometry priors regularize the problem, but disrupt the sparsity and thus the tractability of joint optimization as discussed in [10]. In practice, these methods resort to alternating pose and geometry estimation or move away from second-order optimization. More recently, DROID-SLAM [37] reconstructs depths for all pixels, but similar to sparse systems, does so with no geometric correlation. This results in highly parallelized bundle adjustment, but results in many noisy points and no guarantee of consistent depths within a frame and across frames, as shown in Fig. 2a. All real-time state-of-the-art SLAM systems [6, 10, 37, 38] still rely on the Schur complement, and dense methods with geometric correlation have traditionally not demonstrated comparable pose accuracy, which also limits geometry estimation. + +Learned Depth Priors for Compact Optimization To achieve consistent poses and geometry, monocular depth priors in SLAM focus on test-time optimization to avoid irreparable errors from single-view depth prediction. Therefore, depth priors in monocular SLAM often predict a subspace of potential depths given an image via a compact latent code [3,7,20], or a depth map basis [13,36]. Even with this flexibility, depth maps are not guaranteed to be consistent across multiple views since latent variables live only in frames as shown in Fig. 2b. While per-pixel depth and keypoint reprojection losses may encourage 3D consistency [7], this may introduce bias and does not scale computationally since it requires per-pixel pairwise frame comparisons. Most importantly, it lacks intrinsic 3D consistency, and depth maps often produce a layering effect in 3D as in Fig. 2b. Predicting a view-based mesh topology permits optimizing depth vertices [4], but it is unclear how to move to a 3D topology. + +![](images/f69143eda7fa70c6e561e08af00d0bbe115f3f33b72c77134efcff92696bfa02.jpg) +(a) Dense bundle adjustment with 3D landmarks [37] + +![](images/6be085474a9c4deffa7ed1e293b2cac8ae196a818611117c1402a0efd86da8ff.jpg) +(b) Per-frame dense depth priors via compact codes [7] +Fig. 2: Reconstructions on Replica and geometry properties of different dense map representations. The dotted line represents the true surface. (a) Densely reconstructing a large number of conditionally independent 3D points given poses enables accurate pose estimation and many accurate points, but there is no guarantee of coherent geometry. (b) Depth priors produce smooth depth maps, but even with inter-frame consistency losses, can produce inconsistent geometry and bias preventing global consistency. (c) A compact set of 3D points and depth covariance functions anchor depth maps together, leading to consistent pose estimation and dense geometry. + +![](images/8bb3b6047f89b566c558cf0e3798e197d9d4e17b6ee1d1e70ba738627eafba9c.jpg) +(c) Ours: 3D points and depth covariance functions + +In this work, we leverage the recently proposed learned depth covariance function [9]. While [9] introduces an odometry formulation, it only optimizes per-frame 2D inducing depths, which has the same limitations as per-frame codes, and lacks intrinsic 3D consistency. Our key insight is that depth covariance permits sharing latent 3D points across frames, so that our representation lives in 3D rather than separate 2D image planes. While codes and bases are a fixed subspace that cannot guarantee a depth map passes through any set of 3D points, our covariance formulation guarantees that depth maps go through any desired 3D points as shown in Fig. 2c. Therefore, we can anchor depth maps together by construction for dense 3D consistency. We also exploit the covariance function for our frontend in tracking and initializing potentially non-visually distinct points. + +Keyframe-based vs. Volumetric Maps Keyframe-based maps host dense depths in camera frames, which maintain flexible level-of-detail rooted in the camera resolution and focus only on surfaces. However, representing space using per-pixel depths cannot guarantee that inter-frame correspondences refer to the same 3D point. While pairwise depth constraints can mitigate inconsistency [7], these exhaustive per-pixel errors are expensive to compute, lack intrinsic consistency, and may introduce bias when balancing with other losses. Methods that project hosted depths into neighboring frames and predict flow updates [37] are asymmetric with no guarantee of cycle consistency. + +Volumetric maps live in 3D space and unknown quantities are shared across frames. 3D consistency is achieved by construction but it is challenging to balance fidelity and efficiency. Voxel grids using backward sensor models [15,26] assume known poses and independence between cells, and thus are unsuitable for efficient and consistent optimization. While forward models in voxel grids [39], neural fields [21,35], and 3D Gaussian splatting [16,19] permit consistent geometry and pose optimization, these representations are limited in low-latency monocular VO and SLAM due to expensive rendering, a large number of correlated variables per ray, and alternating and first-order optimization. Volumetric methods are + +also subject to a resolution trade-off. Higher resolution permits greater detail at the expense of memory, runtime complexity, and higher levels of noise. + +Our 3D representation achieves the benefits of both frame-based and volumetric maps for odometry and mapping: the per-pixel resolution and efficiency of depth maps with the intrinsic 3D consistency of volumetric representations. The efficient compact-to-dense model permits joint optimization of poses and geometry via any-pixel constraints in real-time. + +# 3 Compact Mapping Backend + +The mapping backend maintains a compact set of 3D anchor points along with keyframes hosting per-pixel covariance function parameters. A dense world-centric point cloud is decoded from anchor points and poses, which is refined by second-order minimization of photometric error. VO operates in a sliding-window fashion with a fixed number of keyframes. Fig. 3 provides an overview. + +Quantities in the world frame and camera frame, such as points, are denoted by subscripts $\mathbf{P}_{\mathrm{W}}$ and $\mathbf{P}_{\mathrm{C}}$ , respectively. Anchor points are indexed by superscripts $m$ , while dense points are indexed $n$ , such as pixels $\mathbf{p}^m$ and $\mathbf{p}^n$ . Stacked vector forms of sparse and dense variables use capital M and N, as in log-depth vectors $\mathbf{d}_{\mathrm{M}}$ and $\mathbf{d}_{\mathrm{N}}$ . The notation $\|\mathbf{r}\|_{\Sigma}^{2}$ denotes Mahalanobis distance of residual vector $\mathbf{r}$ with measurement covariance matrix $\boldsymbol{\Sigma}$ . + +# 3.1 Preliminaries: Depth Covariance Function + +First, we introduce the depth covariance function from [9]. This covariance function models the distribution over all possible log-depth functions for any finite set of pixels via a Gaussian process (GP) [30]. First, a CNN takes in an RGB image and outputs a per-pixel feature map $\phi$ . Compared to [9], we use a zero-mean GP prior for simplicity, such that the distribution is defined only by the covariance function $k$ that takes in two pixels and respective CNN features: + +$$ +d (\mathbf {p}) \sim \mathcal {G P} (0, k ([ \mathbf {p}; \phi (\mathbf {p}) ], [ \mathbf {p} ^ {\prime}; \phi (\mathbf {p} ^ {\prime}) ])). \tag {1} +$$ + +We found that the mean log-depth variable proposed in [9] is unnecessary, as the largest eigenvector of the covariance function corresponds to scale of the entire depth map. More information on the exact covariance function $k$ is detailed in [9], but the key idea is that larger covariance is an indicator of similarity in log-depth. Since the covariance function takes in both pixel locations and CNN features, it is nonstationary which permits varying levels of smoothness and discontinuities conditioned on the image content. The zero-mean prior defines a planar depth prior in the absence of observations. To define a covariance matrix $\mathbf{K}$ for a set of pixels, the $(i^{\mathrm{th}}, j^{\mathrm{th}})$ entry of $\mathbf{K}$ is filled in by $k\left([ \mathbf{p}^i; \phi(\mathbf{p}^i) ], [ \mathbf{p}^j; \phi(\mathbf{p}^j) ] \right)$ . + +# 3.2 Compact-to-Dense Geometry + +Given anchor point to keyframe visibility, we generate a dense depth map given anchor points $\mathbf{P}_{\mathrm{W}}^{m}$ , the pose of keyframe $r$ with rotation $\mathbf{R}_{\mathrm{WC}_r}$ and translation + +![](images/d55bbe92cff36c4db002afbf5f3ab8c93ef98f98a98033e342d55a2fb109e085.jpg) +(a) Anchor point projection + +![](images/a3fb4a15fada429a43cb0542c2e0917a3cb9bb6571e3c87b2c048c47bb5b88eb.jpg) +(b) Dense geometry construction +Fig. 3: Overview of compact mapping framework. (a) Anchor points are projected to visible keyframes. (b) Keyframes decode into dense depth and backproject geometry to 3D. (c) Target frames enforce dense photo-consistency. (d) Optimizing dense alignment error leads to updated poses and geometry with greater 3D consistency. + +![](images/270e83fd8c2d0e96f009a7a1567b7bd9c069d9fae7a8d536c52e2a421f835f56.jpg) +(c) Photometric error calculation + +![](images/764685c9739176877aaeb68fb9c57a6985437e4348fd6824c164eb2c81fc9f00.jpg) +(d) Pose and geometry update + +$\mathbf{t}_{\mathrm{WC}_r}$ , and the keyframe image. Anchor points are transformed and projected to pixel coordinates $\mathbf{p}$ via camera projection $\pi$ : + +$$ +\mathbf {P} _ {\mathrm {C} _ {r}} ^ {m} = \mathbf {R} _ {\mathrm {W C} _ {r}} ^ {T} \left(\mathbf {P} _ {\mathrm {W}} ^ {m} - \mathbf {t} _ {\mathrm {W C} _ {r}}\right), \quad \mathbf {p} ^ {m} = \pi \left(\mathbf {P} _ {\mathrm {C} _ {r}} ^ {m}\right), \tag {2} +$$ + +which is also illustrated in Fig. 3a. At this point, we create a dense log-depth map from log-depth observations $d^{m} \triangleq \log \left( [\mathbf{P}_{\mathrm{Cr}}^{m}]_{z} \right)$ and projected pixel coordinates $\mathbf{p}^{m}$ . Given an image and the covariance function $k$ for any two pixels from Sec. 3.1, we can construct the covariance matrix between all visible anchor point pixels in $\mathbf{K}_{\mathrm{MM}}$ , and the covariance between all image pixels and anchor point pixels in $\mathbf{K}_{\mathrm{NM}}$ . We also stack all training $d^{m}$ and test $d^{n}$ log-depths into $\mathbf{d}_{\mathrm{M}}$ and $\mathbf{d}_{\mathrm{N}}$ , respectively. The GP formulation allows efficient dense log-depth prediction via the linear Gaussian conditioning equation: + +$$ +\mathbf {d} _ {\mathrm {N}} = \mathbf {K} _ {\mathrm {N M}} \left(\mathbf {K} _ {\mathrm {M M}}\right) ^ {- 1} \mathbf {d} _ {\mathrm {M}}. \tag {3} +$$ + +The GP guarantees that each dense log-depth map goes through the anchor points, creating a consistent surface across multiple views. Next, we backproject points for each pixel and transform to world coordinates: + +$$ +\mathbf {P} _ {\mathrm {W}} ^ {n} = \mathbf {R} _ {\mathrm {W C} _ {r}} \pi^ {- 1} \left(\mathbf {p} ^ {n}, e ^ {d ^ {n}}\right) + \mathbf {t} _ {\mathrm {W C} _ {r}}, \tag {4} +$$ + +where $n$ indexes a query pixel. Note that each keyframe generates a dense depth map via its observed anchor points and covariance parameter feature maps as in Fig. 3b. This compact-to-dense formulation only requires transformations, camera projection, and linear GP conditioning. Therefore, we can efficiently calculate analytical Jacobians for dense per-pixel constraints with respect to poses and anchor points. We show the chain rule for these steps in Section A.1. + +In practice, we ignore the Jacobians with respect to pixel coordinates when constructing $\mathbf{K}_{\mathrm{MM}}$ and $\mathbf{K}_{\mathrm{NM}}$ and calculate the covariance matrices once at initialization. In addition to being more efficient, this ensures that depth maps do not undergo unstable changes if a point moves across a depth discontinuity. If a visible anchor point moves behind a camera during optimization, we reinitialize it to the median depth of the keyframe. + +# 3.3 Photometric Residuals + +While only keyframes are used to create dense geometry, additional small baseline support frames are included to improve convergence. We define a set of edges $\mathcal{E}$ for photometric error that consists of temporally adjacent keyframe-keyframe pairs, and support frames to the two nearest keyframes in time. Given a target image $I_{t}$ , which may be either a keyframe or support frame, with pose $\mathbf{T}_{\mathrm{WC}_t}$ , we first transform world points from keyframe view $r$ and project: + +$$ +\mathbf {p} _ {t} ^ {n} = \pi \left(\mathbf {R} _ {\mathrm {W C} _ {t}} ^ {T} \left(\mathbf {P} _ {\mathrm {W}} ^ {n} - \mathbf {t} _ {\mathrm {W C} _ {t}}\right)\right). \tag {5} +$$ + +Given these projective correspondences and accounting for exposure and global illumination changes via affine brightness parameters, $a$ and $b$ , similar to [10], we calculate the photometric error across pixels and frame edges: + +$$ +E = \sum_ {r, t \in \mathcal {E}} \sum_ {n} \left\| \mathbf {r} _ {r, t} ^ {n} \right\| _ {\sigma_ {\mathbf {r}} ^ {2} I} ^ {2}, \quad \mathbf {r} _ {r, t} ^ {n} = I _ {t} \left(\mathbf {p} _ {t} ^ {n}\right) + b _ {t} - \left(\frac {e ^ {- a _ {r}}}{e ^ {- a _ {t}}} I _ {r} \left(\mathbf {p} _ {r} ^ {n}\right) + b _ {r}\right), \tag {6} +$$ + +which is illustrated in Fig. 3c. These constraints depend on each frame's pose and affine parameters, and the anchor points viewed by the reference keyframe. For robustness, we perform iteratively reweighted least squares (IRLS) with a Huber cost function and set the photometric standard deviation $\sigma_{\mathbf{r}}$ to be proportional to the median absolute residual [1] to handle diverse scenes with different error scales. Compared to RGB residuals, we found that grayscale images were more efficient and similarly robust. The Jacobian chain rule for photometric residuals is in Section A.2. + +# 3.4 Additional Constraints + +While photometric error is the only data term, we introduce additional priors to remove ambiguities. Priors on the oldest keyframe's pose and affine brightness parameters remove gauge freedom in the sliding-window. After removing a keyframe from the window, we include priors on its still-observed landmarks to ensure global consistency as a simpler alternative to first-estimate Jacobians [10]. + +Since keyframes may have regions unobserved by other frames, which causes anchor points to lack photometric constraints, we include two geometry priors. The first is a weak prior on anchor points to match the log median depth $s$ of the keyframe it was first observed in. The second is a GP prior $||\mathbf{d}_{\mathrm{M}} - s||_{\mathbf{K}_{\mathrm{MM}}}^2$ on every keyframe to regularize anchor points that have large correlation, which gives improved estimates for underconstrained anchor points lying on surfaces with well-constrained ones, such as on textureless walls. + +During optimization, there is ambiguity in whether an anchor point changes depth or moves laterally along a surface, such as when viewing a plane at an angle. Thus, we include a pixel prior encouraging the projection of the point in its initializing keyframe not to change from its first observation, which defines a ray-surface intersection. As mentioned previously, this is more efficient than calculating Jacobians through Sec. 3.2 and more stable since GP inducing point optimization has little benefit when points start in suitable locations [5] as ensured by our frontend in Sec. 4. + +![](images/bf2b1436a9bd43ef487970386dda0de40fa46524793e44d7149c6d906323cf10.jpg) +(a) Anchor points visible in KF #1 + +![](images/c96b1c6c85ac56ac8790cd63e74819fd78715220f95ce9e92ed5a6ec6e07a935.jpg) +(b) KF #1 depths projected to KF #2 +Fig. 4: Overview of visibility checks between two keyframes (KF). Matches are shown in blue, rejected matches in red, and newly initialized points in green. Note that occluded edges are rejected, while non-visually distinct points are often matched. New points are allocated to geometrically complex regions while the table is already well-represented. + +![](images/dea814d6b88eb52d91e1c2636e193b468772fa1b8a095c1a96722a7219fffc52.jpg) +(c) Anchor point matches in KF #2 + +![](images/c52c02513e3d47af233accf282aa43f65847462be8b4d25371c2b54437aebe52.jpg) +(d) Newly initialized points in KF #2 + +# 3.5 Second-Order Optimization + +We calculate analytical Jacobians for all constraints in the optimization. The compactness of our linear system permits real-time joint optimization of poses and correlated geometry that is infeasible for traditional dense VO. Stacking all Jacobians into $\mathbf{J}$ , weights from IRLS and noise models into the diagonal matrix $\mathbf{W}$ , and residuals into $\mathbf{r}$ , we solve for parameter updates $\Delta \mathbf{x}$ shown in Fig. 3d via dense Cholesky factorization of the Gauss-Newton normal equations: + +$$ +\left(\mathbf {J} ^ {T} \mathbf {W} \mathbf {J}\right) \Delta \mathbf {x} = - \mathbf {J} ^ {T} \mathbf {W r}. \tag {7} +$$ + +Pose updates are minimally parameterized as Lie-algebra elements $\mathfrak{se}(3)$ as is standard in SLAM [32]. We found that the most time consuming step of backend optimization is accumulating geometry blocks into the Hessian $(\mathbf{J}^T\mathbf{W}\mathbf{J})$ . This can be sped up by factoring out constant components of the Jacobians and reducing the dimensionality of the accumulation as detailed in Section A.4. + +# 4 Visual Frontend + +The visual frontend tracks the current frame with respect to the map, selects keyframes and support frames for the backend, determines 3D anchor point visibility in new keyframes, and initializes new 3D anchor points. Compared to feature-based systems, the ideal points for an expressive yet compact map may be visually indistinct, so we utilize depth covariance for robust visibility checks. + +# 4.1 Dense Photometric Tracking + +Given the current estimate of the newest keyframe's depth, we optimize the pose and affine brightness parameters using grayscale image alignment in the IRLS framework [2]. For efficiency, we perform inverse compositional tracking and use a multi-scale image pyramid. After tracking, a new keyframe is added to the backend if there is either significant translation relative to the median scene depth or if the number of unique projected pixels in the tracked frame falls below a threshold to handle rotation and occlusion. Support frames are added via modified thresholds according to the desired frequency between keyframes. + +# 4.2 Anchor Point Visibility + +When a new keyframe is added, we perform explicit visibility checks to 3D anchor points viewed in the last keyframe to achieve 3D consistency. Note that we only check for visibility as anchor points have no explicit constraint on its location in the image. Compared to feature-based methods [24], anchor points often lack visual distinctiveness, so matching via image information is problematic. Instead, we leverage the interpretability of the depth covariance function to determine visibility. In general, an anchor point should be compatible with all depth covariance functions that use it to generate dense geometry. + +First, the 3D anchor points and dense depth map from the last keyframe, shown in Fig. 4a, are projected into the current frame, shown in Fig. 4b. Given the projected sparse and dense pixel coordinates, we may query covariance parameters $\phi (\mathbf{p})$ and construct $\mathbf{K}_{\mathrm{MM}}$ and $\mathbf{K}_{\mathrm{MN}}$ for the new keyframe. We leverage the linearity of the GP to compress dense log-depth observations into the sparse projected anchor point coordinates. Assuming conditional independence of log-depth observations $\mathbf{d}_n$ given the sparse log-depths $\mathbf{d}_m$ , we can solve an efficient least-squares problem: + +$$ +\min _ {\mathbf {d} _ {\mathrm {M}}} \left| \left| \mathbf {K} _ {\mathrm {N M}} \left(\mathbf {K} _ {\mathrm {M M}}\right) ^ {- 1} \mathbf {d} _ {\mathrm {M}} - \mathbf {d} _ {\mathrm {N}} \right| \right| _ {\sigma_ {d} ^ {2}} ^ {2} + \left| \left| \mathbf {d} _ {\mathrm {M}} \right| \right| _ {\mathbf {K} _ {\mathrm {M M}}} ^ {2}, \tag {8} +$$ + +where the first conditional term ensures that $\mathbf{d}_{\mathrm{M}}$ fits the projected dense observations with standard deviation $\sigma_{d}$ , while the second prior term regularizes sparse log-depths according to the GP training covariance $\mathbf{K}_{\mathrm{MM}}$ . We perform a consistency check on the log-depth difference between the current 3D anchor points and those solved in Eq. (8), and prune points at depth discontinuities. This yields potential matches the are consistent between the two keyframes. + +In practice, we set a maximum number of visible anchor points per-frame to allow efficient batched mapping operations on the GPU. If all points are matched, it could prevent new ones from being initialized and parts of the scene would lack capacity to adequately represent geometry. Thus, we perform pruning of matches so that we have a suitable distribution of points before initializing new anchors. Similar to the GP conditional mean from Eq. (3), we may also obtain conditional variances of all pixels with respect to the first $j$ pixels via: + +$$ +\boldsymbol {\sigma} _ {j} ^ {2} = \operatorname {d i a g} \left[ \mathbf {K} _ {\mathrm {N N}} \right] - \mathbf {K} _ {\mathrm {N M} _ {1: j}} \left(\mathbf {K} _ {\mathrm {M} _ {1: j} \mathrm {M} _ {1: j}}\right) ^ {- 1} \mathbf {K} _ {\mathrm {M} _ {1: j} \mathrm {N}}. \tag {9} +$$ + +We perform active sampling via conditional variance reduction (CVR) [14, 22], where at the $(j + 1)^{\mathrm{th}}$ iteration, the pixel in domain $P$ with the largest variance is selected $\mathbf{p}_{j + 1} = \arg \max_{\mathbf{p}\in P}\sigma_j^2 (\mathbf{p})$ . We actively sample from the set of potential visible points until a variance threshold is reached, which gives the final set of visible anchor points for the current keyframe. Compared to [9], we found that a minimum distance threshold between pixel locations and from the image border is required to avoid accumulating pixels near the same discontinuity. This is essential for improving depth estimates and maintaining a good set of candidate points, as otherwise many points near edges may be occluded at the same time. We show the set of matches between two keyframes in Fig. 4c. + +![](images/56b0853a238dff6d20aab3b8a4748b88afbf94a7b5982e685223e2357894f749.jpg) + +![](images/c2f05c3b6d3c4e0a72eb49baedec167687f5851e8815f2b9b0be9f6b495379a4.jpg) +Fig. 5: Visualization of sliding-window odometry on TUM $fr2/desk$ . Top images show tracked and newly initialized anchor points in blue and red, respectively. Depth maps are below, while the dense point cloud with 248 anchor points is shown on the right. Note the consistent point tracks on textureless surfaces, such as the table and plate, as well as the long-term tracks on the cube, mug, and book. + +![](images/2a44ec5c85e607d5515af8b0f78286a651d50804b41ecd175e5bc8e2fe690a05.jpg) + +![](images/d51115a4ffb4b9a4917f662252f8ad3e1bacdff8753d9c6e6d208180936df3f1.jpg) + +![](images/0461e0d585cbc0f7a2a144b675a90b51aaf5154a231c556ca6de222063648ec6.jpg) + +![](images/482264346b9ade53b0a1c67a9cdab492dcc1ae8557215146b8e2d74c3a5ed108.jpg) + +![](images/0f8562bc9a69aa9ef3ce4e8b3b25f30df79fc97687994bac87ab8de8fb828f91.jpg) + +![](images/2aff15b414a1195164d0b78fdedcb40db594370f8c1b24263b73bd4eba40aa4b.jpg) + +![](images/165e0447c1fb62b31213ee9ba4ac35b0b083cc68aa4ed3c57a7eaa20a01455ef.jpg) +# + +# 4.3 Anchor Point Initialization + +As the camera explores unobserved parts of the scene, new anchor points must be initialized. First, we again perform active sampling, but now across all pixels given the current anchor point locations from the previous section. This ensures new points will be initialized for parts of the scene that are unobserved, lack matches, or require additional geometric capacity, as shown in Fig. 4d. + +Once the pixel locations of new anchor points are known, we initialize the log-depth and backproject into 3D. Defining the set of known log-depths from the anchor points $\mathbf{d}_{\mathrm{m}_1}$ , the unknown log-depths of newly initialized points $\mathbf{d}_{\mathrm{m}_2}$ and all log-depths as $\mathbf{d}_{\mathrm{M}} = (\mathbf{d}_{\mathrm{m}_1}^T,\mathbf{d}_{\mathrm{m}_2}^T)^T$ , we solve another least-squares problem: + +$$ +\min _ {\mathbf {d} _ {\mathrm {m} _ {2}}} \left\| \mathbf {K} _ {\mathrm {N M}} \left(\mathbf {K} _ {\mathrm {M M}}\right) ^ {- 1} \mathbf {d} _ {\mathrm {M}} - \mathbf {d} _ {\mathrm {N}} \right\| _ {\sigma_ {d} ^ {2}} ^ {2} + \left\| \mathbf {d} _ {\mathrm {m} _ {2}} - s \right\| _ {\sigma_ {s} ^ {2}}, \tag {10} +$$ + +where the first term promotes fitting reprojected dense depth observations given matched anchor log-depths and the new log-depth initializations, while the second term is a prior that new log-depths are close to the log median depth $s$ of the previous keyframe. We found that using an isotropic prior with standard deviation $\sigma_d$ , rather than $\mathbf{K}_{\mathrm{MM}}$ , avoids bias for parts of the image that were not previously observed and would otherwise be underconstrained. We define $\sigma_d$ as the standard deviation of the residuals from Eq. (8) to better calibrate the model. Given the new pixel locations, log-depth initializations, and the new keyframe's pose estimate, 3D points are initialized for the backend. A qualitative example of tracked anchor points, depth maps, and reconstruction is shown in Fig. 5. + +# 5 Experiments + +We evaluate both the pose and geometry estimation of our method on multiple datasets. For sparse methods, we compare against feature-based ORB- + +
office0office1office2office3office4room0room1room2mean
SORB-SLAM3 [6]0.40.38.50.44.50.30.30.41.9
DSO [10]0.30.50.30.25.30.20.80.21.0
DDeepFactors [7]42.938.645.641.450.132.426.144.140.2
DROID-VO [37]5.34.16.511.27.19.04.87.06.9
COMO-NC2.95.07.46.18.14.44.53.75.3
COMO2.42.44.33.65.52.52.73.43.4
+ +Table 1: Mean ATE (cm) over 5 runs for monocular VO on the Replica dataset. S and D in the first column refer to sparse and dense methods, respectively. + +SLAM3 [6], direct DSO [10], and learning-based DPVO [38] in trajectory estimation. We also compare against a representative set of dense methods. TANDEM [18] uses DSO for odometry and fuses learned MVS predictions into a TSDF for tracking. MonoGS [19] is a 3D Gaussian splatting-based monocular SLAM system. DeepFactors [7] is a code-based system with photometric, geometric, and keypoint losses to promote consistency across depth maps. DROIDVO [37] uses learned correspondence and flow updates for bundle adjustment, such that all depths are reconstructed but lack geometric correlation. DepthCov [9] optimizes sparse 2D depths under the depth covariance framework, and in a related manner, COMO-NC (no correspondence) is our full framework but no 3D anchor points are tracked across frames. We disable loop closure for DeepFactors and DROID-SLAM to isolate the odometry and mapping performance. + +# 5.1 Implementation Details + +We use a fixed configuration across all three datasets to demonstrate the robustness of the system. We use the off-the-shelf depth covariance function [9] that was trained on the ScanNet training set. Our sliding window odometry uses 9 keyframes and 3 support frames between keyframes for a total of 24. Each keyframe views a maximum of 64 anchor points and initializes new points if it fails to track all from the previous keyframe. We use 256x192 images for all operations, and the mapping backend samples pixels with the highest gradient magnitude in 4x4 images patches for photometric error. + +# 5.2 Trajectory Evaluation + +Replica We first test our monocular odometry on the Replica [33] sequences recorded in [35], which provides a photorealistic environment with ground-truth geometry. As shown in Tab. 1, sparse methods without priors perform best, as conditions are favorable for SLAM: few image artifacts, no exposure changes, and sufficient baselines for multi-view geometry. Among methods with learned priors, COMO achieves the lowest ATE. The representation has little bias as compared to the code-based DeepFactors, which drifts despite photometric, geometric, and keypoint constraints. Sharing 3D points between frames in COMO shows significant improvement over COMO-NC which disables correspondence. TUM The TUM RGBD dataset [34] is a challenging dataset for monocular VO due to significant motion blur, exposure changes, rolling shutter artifacts, + +
fr1fr2fr3 longmean
360deskdesk2plantroomrpyteddyxyzxyzdesk
SORB-SLAM3 [6]X2.0X11.8X5.6X1.00.51.31.7X
DSO [10]X27.266.06.058.6XX3.80.32.29.9X
DPVO [38]13.19.46.53.039.83.56.21.30.53.55.58.4
DTANDEM [18]X4.333.7XX4.943.12.40.32.08.3X
MonoGS [19]14.26.374.09.364.93.435.61.64.5133.13.331.8
DeepFactors [7]17.915.920.231.938.33.856.05.98.426.349.024.9
DepthCov [9]12.85.64.826.125.75.247.55.61.215.968.819.9
DROID-VO [37]15.75.211.16.033.43.219.15.610.77.97.311.4
COMO-NC16.14.210.919.328.65.268.74.10.78.846.819.4
COMO12.94.99.513.827.04.824.54.00.76.310.510.8
+ +Table 2: Mean ATE (cm) over 5 runs for monocular VO on the TUM dataset. S and D in the first column refer to sparse and dense methods, respectively. + +and heavy rotations. Results comparing our method against state-of-the-art VO methods on 11 sequences is shown in Tab. 2. For sparse methods, ORB-SLAM3 and DSO both lose tracking on multiple sequences, while DPVO, which uses learned context for tracking 64 sparse patches and performing bundle adjustment, performs best among all methods. In this work, we are focused on also reconstructing dense geometry, but it would be interesting to combine the learned features of DPVO with our 3D representation. For dense methods, TANDEM fails due to relying on DSO for odometry, which can still fail despite using a TSDF with MVS fusion for tracking. This demonstrates the need for joint optimization of poses and dense image information. While MonoGS can produce accurate trajectories for some sequences, the lack of priors reduces its robustness, and volumetric rendering is slower than all depth-based systems. Despite promoting consistency between neighboring depth maps, DeepFactors has relatively large error. Both DepthCov and COMO-NC are similar in terms of optimizing sparse depths in each frame, while leveraging a true 3D representation in COMO reduces error by almost $50\%$ . COMO outperforms DROID-VO, which uses learned features for matching and reconstructs all pixels from downsized images via sparse bundle adjustment. As the assumptions of brightness constancy are violated in many real-world datasets, learned features could further the accuracy of COMO, but it is interesting that our compact representation and simple photometric error can produce the lowest ATE of dense methods. + +ScanNet Test Lastly, we evaluate the trajectory error on 18 diverse medium-length scenes from the ScanNet [8] test set. Sequences include homes, offices, schools, businesses, and outdoors. Methods that fail are assigned a maximum $100\mathrm{cm}$ error for averages. Sequences have challenging rotational motion, high image noise, and specular surfaces. We summarize the results in Fig. 6, which shows the number of sequences below varying ATE thresholds, while Tab. 3 contains summary statistics. COMO has the lowest mean ATE and area-underthe-curve (AUC). While DPVO has a higher number of trajectories under low ATE thresholds, it also shows high variance results on some sequences. Despite using photometric error that is often violated due to image noise and specular surfaces, our representation with 3D consistency proves valuable in challenging real-world data. An example reconstruction is shown in Fig. 1a. + +![](images/5b0644cb0379c11ad6848ac680648dd18d2298bfb7ffa3edb0824a115ff97cd5.jpg) +Fig. 6: Number of successful trajectories below ATE threshold for selected ScanNet test sequences. +Table 3: ScanNet Mean ATE (cm) and AUC. + +
MethodATEAUC
ORB-SLAM379.91.63
DSO54.72.46
DPVO15.16.34
TANDEM41.73.79
DeepFactors36.12.97
DROID-VO13.96.51
COMO-NC15.86.16
COMO13.06.67
+ +
MethodReplicaScanNet
TANDEMN/A0.325
DeepFactors0.2630.232
DROID-VO1.1291.389
COMO-NC0.0690.155
COMO0.0460.128
+ +![](images/f59ebe50f0c9a3ce41a1a9f1b25f69c0520fe5b4bba68021f54bf122f360d53d.jpg) +Table 4: Depth absolute relative error. +Fig. 7: ScanNet depth accuracy and consistency metrics for fraction of pixels falling below threshold $\delta$ . + +![](images/7ccca15e2c6c832ec16a50df124b221a83f84cfbcedda4ef82291a7508cc9f66.jpg) + +# 5.3 Geometry Evaluation + +Since we are interested in both consistent pose estimation and dense geometry, we evaluate dense depth predictions on Replica and ScanNet. To evaluate accuracy, we perform the same global scale alignment that is used to produce ATE for monocular methods and transform estimated poses and depth maps to be aligned to the metric ground-truth. Beyond global accuracy, we also measure consistency, which checks whether neighboring frames viewing the same 3D points have consistent depth estimates. Note that consistency does not care about the correctness of depth. To check valid pixels, we backproject ground-truth depth maps and project these points into neighboring frames, and check if the depths agree within $1\mathrm{cm}$ . Then, for all valid 3D correspondences, we calculate metrics between pairs of estimated depth images. If $N$ is the number of keyframes for that sequence, then $N$ depth images are used for accuracy metrics, while $2(N - 1)$ depth images are used for consistency since pairs require both directions. + +For estimated depth $\hat{D}$ and ground-truth depth $D$ we show the mean absolute relative error $(1 / N)\sum_{i}|\hat{D}_{i} - D_{i}| / D_{i}$ across all sequences for Replica and ScanNet in Tab. 4. We do not evaluate TANDEM on Replica since it is used in the training data. COMO outperforms other methods due to its ability to jointly optimize dense, correlated geometry along with poses. Despite accurate pose estimation, DROID-VO has outliers with significant depth errors. To further evaluate the distribution of depth accuracy and remove the effect of large outliers, in Fig. 7, we use the $\delta$ threshold commonly used in depth estimation, which measures what + +![](images/7f478f47dfe1f7c654ec4f147afdc2d47451e77e57ebdbf1468efe96d3bdfbf3.jpg) +Fig. 8: Mesh and surface normals from TSDF fusion of a live desk sequence. + +![](images/85f665e026a3e998ae5b11fdec72ec79a24e3fd3db86d694802572aaffc31ea7.jpg) +Table 5: Runtime in seconds of different methods on TUM $fr3/long$ using an RTX 3080. COMO SP and MP denote single and multiprocessing, respectively. + +fraction of pixels satisfy $\max \left(\hat{D} / D, D / \hat{D}\right) < \delta$ . Interestingly, accurate depth maps do not necessarily imply consistency. For example, DROID-VO has the second best accuracy behind COMO, while DeepFactors has the second most consistent geometry. COMO achieves both the best accuracy and consistency without explicit geometric and keypoint constraints. + +# 5.4 Real-time System and Runtime Details + +When evaluating our system, we use a sequential, single-threaded mode that runs at around 14 FPS with an RTX 3080. For live demos and reconstructions, we also develop a multiprocessing version that runs at 35-50 FPS as shown in the FPS counter in the top right of the supplementary video, with tracking flexible to be on either CPU or GPU. We show examples of reconstructions while operating a monocular camera in real-time. Fig. 1b displays a large-scale raw point cloud of an office, while Fig. 8 displays consistent TSDF fusion of a desk sequence. + +Average times for tracking a frame, adding a keyframe (KF), and updating the map are 20.1, 50.4, and $39.8~\mathrm{ms}$ . Keyframing is the bottleneck during fast motion, but 35-50 FPS in multiprocessing mode for normal motion is standard as shown in the video. We measure the wall time of several methods in Tab. 5. COMO achieves real-time processing with its efficient representation, but further optimization is still possible since almost all code is implemented in PyTorch. + +# 6 Conclusion + +In this work, we present a compact representation of 3D anchor points and frame-based depth covariance functions to achieve real-time estimation of consistent poses and geometry. Future work includes training the depth covariance function on more diverse data and replacing photometric constraints with the more robust learned correspondence from DROID-VO. Lastly, we believe the anchor point representation could be used in a full map-centric SLAM system with relocalization, which would allow continually improving geometry upon revisiting. With the advance of learned priors, exploiting geometric correlation is an interesting direction to achieve efficient, robust, and consistent algorithms. + +# Acknowledgements + +This research is funded by EPSRC Prosperity Partnerships (EP/S036636/1) and Dyson Technology Ltd. We would like to thank Riku Murai, Hidenobu Matsuki, Gwangbin Bae, and members of the Dyson Robotics Lab for insightful discussions and feedback. + +# References + +1. Alismail, H.: Direct pose estimation and refinement. Ph.D. thesis, Robotics Institute, Carnegie Mellon University. (2016) +2. Baker, S., Matthews, I.: Lucas-Kanade 20 years on: A unifying framework. International Journal of Computer Vision (IJCV) (2004) +3. Bloesch, M., Czarnowski, J., Clark, R., Leutenegger, S., Davison, A.J.: CodeSLAM - learning a compact, optimisable representation for dense visual SLAM. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018) +4. Bloesch, M., Laidlow, T., Clark, R., Leutenegger, S., Davison, A.: Learning meshes for dense visual SLAM. In: Proceedings of the International Conference on Computer Vision (ICCV) (2019) +5. Burt, D.R., Rasmussen, C.E., van der Wilk, M.: Convergence of sparse variational inference in Gaussian processes regression. The Journal of Machine Learning Research (2020) +6. Campos, C., Elvira, R., Rodríguez, J.J.G., M. Montiel, J.M., D. Tardós, J.: ORBSLAM3: An accurate open-source library for visual, visual-inertial, and multimap SLAM. IEEE Transactions on Robotics (T-RO) (2021) +7. Czarnowski, J., Laidlow, T., Clark, R., Davison, A.J.: DeepFactors: Real-time probabilistic dense monocular SLAM. IEEE Robotics and Automation Letters (RA-L) (2020) +8. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017) +9. Dexheimer, E., Davison, A.J.: Learning a depth covariance function. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023) +10. Engel, J., Koltun, V., Cremers, D.: Direct sparse odometry. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2018) +1. Engel, J., Schöps, T., Cremers, D.: LSD-SLAM: Large-scale direct monocular SLAM. In: Proceedings of the European Conference on Computer Vision (ECCV) (2014) +2. Forster, C., Zhang, Z., Gassner, M., Werlberger, M., Scaramuzza, D.: SVO: Semidirect visual odometry for monocular and multicamera systems. IEEE Transactions on Robotics (T-RO) (2017) +3. Graham, B., Novotny, D.: RidgeSfM: Structure from motion via robust pairwise matching under depth uncertainty. In: Proceedings of the International Conference on 3D Vision (3DV) (2020) +4. Guestrin, C., Krause, A., Singh, A.P.: Near-optimal sensor placements in Gaussian processes. In: Proceedings of the International Conference on Machine Learning (ICML) (2005) + +15. Hornung, A., Wurm, K.M., Bennewitz, M., Stachniss, C., Burgard, W.: OctoMap: An efficient probabilistic 3D mapping framework based on octrees. Autonomous Robots (2013) +16. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3D Gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/ +17. Klein, G., Murray, D.: Parallel tracking and mapping for small AR workspaces. In: Proceedings of the International Symposium on Mixed and Augmented Reality (ISMAR) (2007) +18. Koestler, L., Yang, N., Zeller, N., Cremers, D.: TANDEM: Tracking and dense mapping in real-time using deep multi-view stereo. In: Conference on Robot Learning (CoRL) (2022) +19. Matsuki, H., Murai, R., Kelly, P.H.J., Davison, A.J.: Gaussian Splatting SLAM. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2024) +20. Matsuki, H., Scona, R., Czarnowski, J., Davison, A.J.: CodeMapping: Real-time dense mapping for sparse SLAM using compact scene representations. IEEE Robotics and Automation Letters (RA-L) (2021) +21. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: Proceedings of the European Conference on Computer Vision (ECCV) (2020) +22. Moss, H.B., Ober, S.W., Picheny, V.: Inducing point allocation for sparse Gaussian processes in high-throughput bayesian optimisation. In: International Conference on Artificial Intelligence and Statistics (AISTATS) (2023) +23. Mur-Artal, R., Tardós, J.D.: Probabilistic semi-dense mapping from highly accurate feature-based monocular SLAM. In: Proceedings of Robotics: Science and Systems (RSS) (2015) +24. Mur-Artal, R., Tardós, J.D.: ORB-SLAM2: An open-source SLAM system for monocular, stereo, and RGB-D cameras. IEEE Transactions on Robotics (T-RO) (2017) +25. Newcombe, R.A., Davison, A.J.: Live dense reconstruction with a single moving camera. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2010) +26. Newcombe, R.A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohi, P., Shotton, J., Hodges, S., Fitzgibbon, A.: KinectFusion: Real-time dense surface mapping and tracking. In: Proceedings of the International Symposium on Mixed and Augmented Reality (ISMAR) (2011) +27. Newcombe, R.A., Lovegrove, S.J., Davison, A.J.: DTAM: Dense tracking and mapping in real-time. In: Proceedings of the International Conference on Computer Vision (ICCV) (2011) +28. Pizzoli, M., Forster, C., Scaramuzza, D.: REMODE: Probabilistic, monocular dense reconstruction in real time. In: Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) (2014) +29. Platinsky, L., Davison, A.J., Leutenegger, S.: Monocular visual odometry: Sparse joint optimisation or dense alternation? In: Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) (2017) +30. Rasmussen, C.E., Williams, C.K.I.: Gaussian Processes for Machine Learning. MIT Press (2005) +31. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016) + +32. Sola, J., Deray, J., Atchuthan, D.: A micro lie theory for state estimation in robotics. arXiv preprint arXiv:1812.01537 (2018) +33. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., Clarkson, A., Yan, M., Budge, B., Yan, Y., Pan, X., Yon, J., Zou, Y., Leon, K., Carter, N., Brales, J., Gillingham, T., Mueggler, E., Pesqueira, L., Savva, M., Batra, D., Strasdat, H.M., Nardi, R.D., Goesele, M., Lovegrove, S., Newcombe, R.: The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019) +34. Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for the evaluation of RGB-D SLAM systems. In: Proceedings of the IEEE/RSJ Conference on Intelligent Robots and Systems (IROS) (2012) +35. Sucar, E., Liu, S., Ortiz, J., Davison, A.J.: iMAP: Implicit mapping and positioning in real-time. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021) +36. Tang, C., Tan, P.: BA-Net: Dense bundle adjustment networks. In: Proceedings of the International Conference on Learning Representations (ICLR) (2019) +37. Teed, Z., Deng, J.: DROID-SLAM: Deep visual SLAM for monocular, stereo, and RGB-D cameras. In: Neural Information Processing Systems (NeurIPS) (2021) +38. Teed, Z., Lipson, L., Deng, J.: Deep patch visual odometry. In: Neural Information Processing Systems (NeurIPS) (2023) +39. Thrun, S.: Learning occupancy grids with forward models. In: Proceedings of the IEEE/RSJ Conference on Intelligent Robots and Systems (IROS) (2001) \ No newline at end of file diff --git a/2024/COMO_ Compact Mapping and Odometry/images.zip b/2024/COMO_ Compact Mapping and Odometry/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7501ae93440ad42d16039ccd6228f7ea341ccc31 --- /dev/null +++ b/2024/COMO_ Compact Mapping and Odometry/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a695637a6bcd04739c5713f103245ed23ce833329c323464d6344ea07f4aacdc +size 424616 diff --git a/2024/COMO_ Compact Mapping and Odometry/layout.json b/2024/COMO_ Compact Mapping and Odometry/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6d020386d4cfd60a5dbda710074056d2fac22905 --- /dev/null +++ b/2024/COMO_ Compact Mapping and Odometry/layout.json @@ -0,0 +1,9690 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 112, + 454, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 112, + 454, + 129 + ], + "spans": [ + { + "bbox": [ + 159, + 112, + 454, + 129 + ], + "type": "text", + "content": "COMO: Compact Mapping and Odometry" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 208, + 150, + 406, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 150, + 406, + 162 + ], + "spans": [ + { + "bbox": [ + 208, + 150, + 406, + 162 + ], + "type": "text", + "content": "Eric Dexheimer1 and Andrew J. Davison1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 209, + 172, + 404, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 172, + 404, + 194 + ], + "spans": [ + { + "bbox": [ + 209, + 172, + 404, + 194 + ], + "type": "text", + "content": "Dyson Robotics Lab, Imperial College London {e.dexheimer21, a.davison}@imperial.ac.uk" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 218, + 452, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 218, + 452, + 339 + ], + "spans": [ + { + "bbox": [ + 160, + 218, + 452, + 339 + ], + "type": "text", + "content": "Abstract. We present COMO, a real-time monocular mapping and odometry system that encodes dense geometry via a compact set of 3D anchor points. Decoding anchor point projections into dense geometry via per-keyframe depth covariance functions guarantees that depth maps are joined together at visible anchor points. The representation enables joint optimization of camera poses and dense geometry, intrinsic 3D consistency, and efficient second-order inference. To maintain a compact yet expressive map, we introduce a frontend that leverages the covariance function for tracking and initializing potentially visually indistinct 3D points across frames. Altogether, we introduce a real-time system capable of estimating accurate poses and consistent geometry." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 349, + 420, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 349, + 420, + 361 + ], + "spans": [ + { + "bbox": [ + 160, + 349, + 420, + 361 + ], + "type": "text", + "content": "Keywords: SLAM " + }, + { + "bbox": [ + 160, + 349, + 420, + 361 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 349, + 420, + 361 + ], + "type": "text", + "content": " Multi-view geometry " + }, + { + "bbox": [ + 160, + 349, + 420, + 361 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 349, + 420, + 361 + ], + "type": "text", + "content": " 3D representations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 379, + 230, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 379, + 230, + 392 + ], + "spans": [ + { + "bbox": [ + 133, + 379, + 230, + 392 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 403, + 481, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 481, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 481, + 473 + ], + "type": "text", + "content": "Achieving accurate and consistent poses and dense geometry from monocular images in real-time is a challenging yet essential undertaking to push forward state-of-the-art in robotics and augmented reality. Monocular cameras are the key to achieving low-cost, energy-efficient, and compact intelligent platforms. While images contain rich visual information, reconstructing the world is challenging due to the lack of direct geometric observations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 666 + ], + "type": "text", + "content": "The ideal representation for real-time monocular visual odometry (VO) and simultaneous localization and mapping (SLAM) remains elusive. While sparse methods jointly optimize camera poses and a set of conditionally independent 3D points given poses, the map lacks consistent dense geometry for downstream tasks and pose estimation cannot benefit from all visual information. Dense SLAM uses and reconstruct all pixels, but the sheer number of variables relative to measurements renders joint optimization infeasible and inference ill-posed. Recently, learned priors over compact depth map representations, such as codes and sparse 2D depths, have enabled joint optimization. However, representations with 2D depth priors have thus far lagged in accuracy, as consistency between depth maps in 3D is not guaranteed. Volumetric representations, such as voxel grids and neural fields, directly enforce dense 3D consistency by construction, but cannot perform real-time joint optimization due to expensive rendering. Furthermore, compact, efficient, and expressive 3D priors for general scenes are not yet suitable. Achieving accurate and consistent 3D geometry in real-time would represent a significant advance for monocular SLAM." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 118, + 299, + 201 + ], + "blocks": [ + { + "bbox": [ + 133, + 118, + 299, + 201 + ], + "lines": [ + { + "bbox": [ + 133, + 118, + 299, + 201 + ], + "spans": [ + { + "bbox": [ + 133, + 118, + 299, + 201 + ], + "type": "image", + "image_path": "ea6a588415295b00661b95c8da0edc65005b10b9ae85a74954e87a208bdd08f9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 202, + 301, + 228 + ], + "lines": [ + { + "bbox": [ + 132, + 202, + 301, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 202, + 301, + 228 + ], + "type": "text", + "content": "(a) Point cloud reconstruction from the slidingwindow monocular odometry defined by 240 points and 8 keyframes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 312, + 114, + 480, + 201 + ], + "blocks": [ + { + "bbox": [ + 312, + 114, + 480, + 201 + ], + "lines": [ + { + "bbox": [ + 312, + 114, + 480, + 201 + ], + "spans": [ + { + "bbox": [ + 312, + 114, + 480, + 201 + ], + "type": "image", + "image_path": "141eb0d4228048625b58d4e3c468459e7961b84a3a79bb602a0136ef30099705.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 202, + 481, + 227 + ], + "lines": [ + { + "bbox": [ + 312, + 202, + 481, + 227 + ], + "spans": [ + { + "bbox": [ + 312, + 202, + 481, + 227 + ], + "type": "text", + "content": "(b) Dense point cloud from 40 keyframes across a large office during live operation of a monocular camera." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 236, + 480, + 281 + ], + "lines": [ + { + "bbox": [ + 130, + 236, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 480, + 281 + ], + "type": "text", + "content": "Fig. 1: COMO encodes scene geometry via a compact set of 3D anchor points and decodes dense geometry via per-keyframe depth covariance functions. The 3D points visualized in red anchor depth maps together from multiple views while the covariance function generates dense geometry by conditioning on sparse point projections." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 302, + 480, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 480, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 480, + 433 + ], + "type": "text", + "content": "In this work, we propose a representation that achieves the three desired properties: joint optimization of poses and dense geometry, intrinsic 3D consistency, and low-latency, real-time operation. We encode the scene as a compact set of 3D anchor points, which are shared across frames. To decode into dense geometry, 3D points are first projected into keyframes. We then use image-conditioned depth covariance functions [9] per keyframe to generate dense depth maps that intersect each visible 3D point. Depth maps are backprojected into 3D for calculating any-pixel photometric error. All steps are efficient and have analytic Jacobians for second-order optimization. Updates to dense geometry are propagated into our encoded scene representation and ensure suitable regularization for the ill-posed problem of dense reconstruction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 434, + 480, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 480, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 480, + 517 + ], + "type": "text", + "content": "Maintaining a compact map in an incremental VO setting requires a compatible frontend. Keeping a small number of points benefits efficiency, while the distribution of points affects expressiveness. Feature-based systems focus on visually distinct features, but this often leads to many points on edges which are not suitable for modeling dense geometry. Therefore, we leverage depth covariance for determining anchor point visibility in new keyframes, actively initializing new 3D points, and encoding the current dense geometry into new 3D points." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 518, + 480, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 518, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 518, + 480, + 577 + ], + "type": "text", + "content": "Our system demonstrates improved robustness over methods lacking joint pose and dense geometry optimization, as well as more accurate pose and geometry estimation compared to compact 2D representations. Altogether, we propose a real-time VO system that achieves robust and accurate odometry along with consistent dense geometry as shown in Fig. 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 146, + 578, + 355, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 578, + 355, + 589 + ], + "spans": [ + { + "bbox": [ + 146, + 578, + 355, + 589 + ], + "type": "text", + "content": "In summary, the contributions of our work are:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 137, + 594, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 138, + 594, + 481, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 594, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 138, + 594, + 481, + 617 + ], + "type": "text", + "content": "- An efficient representation of dense geometry encoded by a compact set of 3D anchor points and decoded by depth covariance functions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 137, + 618, + 479, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 618, + 479, + 641 + ], + "spans": [ + { + "bbox": [ + 137, + 618, + 479, + 641 + ], + "type": "text", + "content": "- A frontend for the compact map that leverages depth covariance for visibility, active initialization, and encoding dense geometry." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 137, + 642, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 642, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 642, + 479, + 665 + ], + "type": "text", + "content": "- A real-time monocular visual odometry and mapping system that produces accurate and consistent poses and dense geometry." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 296, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 296, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 296, + 101 + ], + "type": "text", + "content": "E. Dexheimer and A.J. Davison" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 148, + 482, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 148, + 482, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 148, + 482, + 339 + ], + "type": "text", + "content": "Sparse vs. Dense Visual SLAM Sparse SLAM systems optimize 3D points that are conditionally independent given poses. Exploiting the sparsity inherent in the information form of structure-from-motion is the key to real-time algorithms with a significant number of points [17]. Modern sparse VO methods [10,12,24] and large-scale bundle adjustment [31] rely on the Schur complement which greatly reduces optimization complexity. Sparse systems lack dense reconstruction, which is useful for many downstream tasks. Thus, it is common to first perform pose estimation followed by dense mapping given poses and potentially sparse landmark estimates [20,23,25,28]. MVS approaches estimate depth maps via a reference cost volume and multiple supporting frames, and have shown progress when integrated into a learning pipeline [18]. However, the sparse followed by dense paradigm has two major issues. Dense information is ignored for pose estimation, which can result in failure in many real-world scenarios, while the mapping step is heavily reliant on accurate pose estimation. Joint optimization of poses and dense geometry avoids the brittle two-step process and has the potential to achieve more accurate odometry and mapping [29]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 341, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 508 + ], + "type": "text", + "content": "Dense methods [11, 27] brought the promise of estimating poses and complete geometry simultaneously. However, joint optimization proved difficult due to the ill-posed nature of reconstructing all points. Hand-crafted geometry priors regularize the problem, but disrupt the sparsity and thus the tractability of joint optimization as discussed in [10]. In practice, these methods resort to alternating pose and geometry estimation or move away from second-order optimization. More recently, DROID-SLAM [37] reconstructs depths for all pixels, but similar to sparse systems, does so with no geometric correlation. This results in highly parallelized bundle adjustment, but results in many noisy points and no guarantee of consistent depths within a frame and across frames, as shown in Fig. 2a. All real-time state-of-the-art SLAM systems [6, 10, 37, 38] still rely on the Schur complement, and dense methods with geometric correlation have traditionally not demonstrated comparable pose accuracy, which also limits geometry estimation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Learned Depth Priors for Compact Optimization To achieve consistent poses and geometry, monocular depth priors in SLAM focus on test-time optimization to avoid irreparable errors from single-view depth prediction. Therefore, depth priors in monocular SLAM often predict a subspace of potential depths given an image via a compact latent code [3,7,20], or a depth map basis [13,36]. Even with this flexibility, depth maps are not guaranteed to be consistent across multiple views since latent variables live only in frames as shown in Fig. 2b. While per-pixel depth and keypoint reprojection losses may encourage 3D consistency [7], this may introduce bias and does not scale computationally since it requires per-pixel pairwise frame comparisons. Most importantly, it lacks intrinsic 3D consistency, and depth maps often produce a layering effect in 3D as in Fig. 2b. Predicting a view-based mesh topology permits optimizing depth vertices [4], but it is unclear how to move to a 3D topology." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "COMO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 246, + 168 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 246, + 168 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 246, + 168 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 246, + 168 + ], + "type": "image", + "image_path": "f69143eda7fa70c6e561e08af00d0bbe115f3f33b72c77134efcff92696bfa02.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 169, + 241, + 186 + ], + "lines": [ + { + "bbox": [ + 137, + 169, + 241, + 186 + ], + "spans": [ + { + "bbox": [ + 137, + 169, + 241, + 186 + ], + "type": "text", + "content": "(a) Dense bundle adjustment with 3D landmarks [37]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 250, + 114, + 364, + 168 + ], + "blocks": [ + { + "bbox": [ + 250, + 114, + 364, + 168 + ], + "lines": [ + { + "bbox": [ + 250, + 114, + 364, + 168 + ], + "spans": [ + { + "bbox": [ + 250, + 114, + 364, + 168 + ], + "type": "image", + "image_path": "6be085474a9c4deffa7ed1e293b2cac8ae196a818611117c1402a0efd86da8ff.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 169, + 356, + 186 + ], + "lines": [ + { + "bbox": [ + 257, + 169, + 356, + 186 + ], + "spans": [ + { + "bbox": [ + 257, + 169, + 356, + 186 + ], + "type": "text", + "content": "(b) Per-frame dense depth priors via compact codes [7]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 195, + 482, + 284 + ], + "lines": [ + { + "bbox": [ + 130, + 195, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 195, + 482, + 284 + ], + "type": "text", + "content": "Fig. 2: Reconstructions on Replica and geometry properties of different dense map representations. The dotted line represents the true surface. (a) Densely reconstructing a large number of conditionally independent 3D points given poses enables accurate pose estimation and many accurate points, but there is no guarantee of coherent geometry. (b) Depth priors produce smooth depth maps, but even with inter-frame consistency losses, can produce inconsistent geometry and bias preventing global consistency. (c) A compact set of 3D points and depth covariance functions anchor depth maps together, leading to consistent pose estimation and dense geometry." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 369, + 114, + 482, + 168 + ], + "blocks": [ + { + "bbox": [ + 369, + 114, + 482, + 168 + ], + "lines": [ + { + "bbox": [ + 369, + 114, + 482, + 168 + ], + "spans": [ + { + "bbox": [ + 369, + 114, + 482, + 168 + ], + "type": "image", + "image_path": "8bb3b6047f89b566c558cf0e3798e197d9d4e17b6ee1d1e70ba738627eafba9c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 169, + 479, + 186 + ], + "lines": [ + { + "bbox": [ + 370, + 169, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 370, + 169, + 479, + 186 + ], + "type": "text", + "content": "(c) Ours: 3D points and depth covariance functions" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 312, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 445 + ], + "type": "text", + "content": "In this work, we leverage the recently proposed learned depth covariance function [9]. While [9] introduces an odometry formulation, it only optimizes per-frame 2D inducing depths, which has the same limitations as per-frame codes, and lacks intrinsic 3D consistency. Our key insight is that depth covariance permits sharing latent 3D points across frames, so that our representation lives in 3D rather than separate 2D image planes. While codes and bases are a fixed subspace that cannot guarantee a depth map passes through any set of 3D points, our covariance formulation guarantees that depth maps go through any desired 3D points as shown in Fig. 2c. Therefore, we can anchor depth maps together by construction for dense 3D consistency. We also exploit the covariance function for our frontend in tracking and initializing potentially non-visually distinct points." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 447, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 555 + ], + "type": "text", + "content": "Keyframe-based vs. Volumetric Maps Keyframe-based maps host dense depths in camera frames, which maintain flexible level-of-detail rooted in the camera resolution and focus only on surfaces. However, representing space using per-pixel depths cannot guarantee that inter-frame correspondences refer to the same 3D point. While pairwise depth constraints can mitigate inconsistency [7], these exhaustive per-pixel errors are expensive to compute, lack intrinsic consistency, and may introduce bias when balancing with other losses. Methods that project hosted depths into neighboring frames and predict flow updates [37] are asymmetric with no guarantee of cycle consistency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Volumetric maps live in 3D space and unknown quantities are shared across frames. 3D consistency is achieved by construction but it is challenging to balance fidelity and efficiency. Voxel grids using backward sensor models [15,26] assume known poses and independence between cells, and thus are unsuitable for efficient and consistent optimization. While forward models in voxel grids [39], neural fields [21,35], and 3D Gaussian splatting [16,19] permit consistent geometry and pose optimization, these representations are limited in low-latency monocular VO and SLAM due to expensive rendering, a large number of correlated variables per ray, and alternating and first-order optimization. Volumetric methods are" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 296, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 296, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 296, + 100 + ], + "type": "text", + "content": "E. Dexheimer and A.J. Davison" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "also subject to a resolution trade-off. Higher resolution permits greater detail at the expense of memory, runtime complexity, and higher levels of noise." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": "Our 3D representation achieves the benefits of both frame-based and volumetric maps for odometry and mapping: the per-pixel resolution and efficiency of depth maps with the intrinsic 3D consistency of volumetric representations. The efficient compact-to-dense model permits joint optimization of poses and geometry via any-pixel constraints in real-time." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 216, + 320, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 320, + 230 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 320, + 230 + ], + "type": "text", + "content": "3 Compact Mapping Backend" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 239, + 479, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 479, + 298 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 479, + 298 + ], + "type": "text", + "content": "The mapping backend maintains a compact set of 3D anchor points along with keyframes hosting per-pixel covariance function parameters. A dense world-centric point cloud is decoded from anchor points and poses, which is refined by second-order minimization of photometric error. VO operates in a sliding-window fashion with a fixed number of keyframes. Fig. 3 provides an overview." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 299, + 480, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 299, + 480, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 480, + 370 + ], + "type": "text", + "content": "Quantities in the world frame and camera frame, such as points, are denoted by subscripts " + }, + { + "bbox": [ + 130, + 299, + 480, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{W}}" + }, + { + "bbox": [ + 130, + 299, + 480, + 370 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 299, + 480, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{C}}" + }, + { + "bbox": [ + 130, + 299, + 480, + 370 + ], + "type": "text", + "content": ", respectively. Anchor points are indexed by superscripts " + }, + { + "bbox": [ + 130, + 299, + 480, + 370 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 299, + 480, + 370 + ], + "type": "text", + "content": ", while dense points are indexed " + }, + { + "bbox": [ + 130, + 299, + 480, + 370 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 299, + 480, + 370 + ], + "type": "text", + "content": ", such as pixels " + }, + { + "bbox": [ + 130, + 299, + 480, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^m" + }, + { + "bbox": [ + 130, + 299, + 480, + 370 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 299, + 480, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^n" + }, + { + "bbox": [ + 130, + 299, + 480, + 370 + ], + "type": "text", + "content": ". Stacked vector forms of sparse and dense variables use capital M and N, as in log-depth vectors " + }, + { + "bbox": [ + 130, + 299, + 480, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{\\mathrm{M}}" + }, + { + "bbox": [ + 130, + 299, + 480, + 370 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 299, + 480, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{\\mathrm{N}}" + }, + { + "bbox": [ + 130, + 299, + 480, + 370 + ], + "type": "text", + "content": ". The notation " + }, + { + "bbox": [ + 130, + 299, + 480, + 370 + ], + "type": "inline_equation", + "content": "\\|\\mathbf{r}\\|_{\\Sigma}^{2}" + }, + { + "bbox": [ + 130, + 299, + 480, + 370 + ], + "type": "text", + "content": " denotes Mahalanobis distance of residual vector " + }, + { + "bbox": [ + 130, + 299, + 480, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 130, + 299, + 480, + 370 + ], + "type": "text", + "content": " with measurement covariance matrix " + }, + { + "bbox": [ + 130, + 299, + 480, + 370 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}" + }, + { + "bbox": [ + 130, + 299, + 480, + 370 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 385, + 372, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 372, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 372, + 397 + ], + "type": "text", + "content": "3.1 Preliminaries: Depth Covariance Function" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 403, + 479, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 479, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 479, + 475 + ], + "type": "text", + "content": "First, we introduce the depth covariance function from [9]. This covariance function models the distribution over all possible log-depth functions for any finite set of pixels via a Gaussian process (GP) [30]. First, a CNN takes in an RGB image and outputs a per-pixel feature map " + }, + { + "bbox": [ + 130, + 403, + 479, + 475 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 403, + 479, + 475 + ], + "type": "text", + "content": ". Compared to [9], we use a zero-mean GP prior for simplicity, such that the distribution is defined only by the covariance function " + }, + { + "bbox": [ + 130, + 403, + 479, + 475 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 403, + 479, + 475 + ], + "type": "text", + "content": " that takes in two pixels and respective CNN features:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 222, + 482, + 480, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 482, + 480, + 495 + ], + "spans": [ + { + "bbox": [ + 222, + 482, + 480, + 495 + ], + "type": "interline_equation", + "content": "d (\\mathbf {p}) \\sim \\mathcal {G P} (0, k ([ \\mathbf {p}; \\phi (\\mathbf {p}) ], [ \\mathbf {p} ^ {\\prime}; \\phi (\\mathbf {p} ^ {\\prime}) ])). \\tag {1}", + "image_path": "02f3939feb1ccfeba074138a5a23aa7ceb66251d9765caafa2eeaaa733a9425c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 501, + 480, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 480, + 609 + ], + "type": "text", + "content": "We found that the mean log-depth variable proposed in [9] is unnecessary, as the largest eigenvector of the covariance function corresponds to scale of the entire depth map. More information on the exact covariance function " + }, + { + "bbox": [ + 130, + 501, + 480, + 609 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 501, + 480, + 609 + ], + "type": "text", + "content": " is detailed in [9], but the key idea is that larger covariance is an indicator of similarity in log-depth. Since the covariance function takes in both pixel locations and CNN features, it is nonstationary which permits varying levels of smoothness and discontinuities conditioned on the image content. The zero-mean prior defines a planar depth prior in the absence of observations. To define a covariance matrix " + }, + { + "bbox": [ + 130, + 501, + 480, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 501, + 480, + 609 + ], + "type": "text", + "content": " for a set of pixels, the " + }, + { + "bbox": [ + 130, + 501, + 480, + 609 + ], + "type": "inline_equation", + "content": "(i^{\\mathrm{th}}, j^{\\mathrm{th}})" + }, + { + "bbox": [ + 130, + 501, + 480, + 609 + ], + "type": "text", + "content": " entry of " + }, + { + "bbox": [ + 130, + 501, + 480, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 501, + 480, + 609 + ], + "type": "text", + "content": " is filled in by " + }, + { + "bbox": [ + 130, + 501, + 480, + 609 + ], + "type": "inline_equation", + "content": "k\\left([ \\mathbf{p}^i; \\phi(\\mathbf{p}^i) ], [ \\mathbf{p}^j; \\phi(\\mathbf{p}^j) ] \\right)" + }, + { + "bbox": [ + 130, + 501, + 480, + 609 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 624, + 308, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 308, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 308, + 636 + ], + "type": "text", + "content": "3.2 Compact-to-Dense Geometry" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "content": "Given anchor point to keyframe visibility, we generate a dense depth map given anchor points " + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{W}}^{m}" + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "content": ", the pose of keyframe " + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "content": " with rotation " + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\mathrm{WC}_r}" + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "content": " and translation" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "COMO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 115, + 216, + 165 + ], + "blocks": [ + { + "bbox": [ + 132, + 115, + 216, + 165 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 216, + 165 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 216, + 165 + ], + "type": "image", + "image_path": "d55bbe92cff36c4db002afbf5f3ab8c93ef98f98a98033e342d55a2fb109e085.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 167, + 204, + 184 + ], + "lines": [ + { + "bbox": [ + 143, + 167, + 204, + 184 + ], + "spans": [ + { + "bbox": [ + 143, + 167, + 204, + 184 + ], + "type": "text", + "content": "(a) Anchor point projection" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 220, + 114, + 302, + 165 + ], + "blocks": [ + { + "bbox": [ + 220, + 114, + 302, + 165 + ], + "lines": [ + { + "bbox": [ + 220, + 114, + 302, + 165 + ], + "spans": [ + { + "bbox": [ + 220, + 114, + 302, + 165 + ], + "type": "image", + "image_path": "a3fb4a15fada429a43cb0542c2e0917a3cb9bb6571e3c87b2c048c47bb5b88eb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 167, + 298, + 183 + ], + "lines": [ + { + "bbox": [ + 225, + 167, + 298, + 183 + ], + "spans": [ + { + "bbox": [ + 225, + 167, + 298, + 183 + ], + "type": "text", + "content": "(b) Dense geometry construction" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 193, + 481, + 237 + ], + "lines": [ + { + "bbox": [ + 131, + 193, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 131, + 193, + 481, + 237 + ], + "type": "text", + "content": "Fig. 3: Overview of compact mapping framework. (a) Anchor points are projected to visible keyframes. (b) Keyframes decode into dense depth and backproject geometry to 3D. (c) Target frames enforce dense photo-consistency. (d) Optimizing dense alignment error leads to updated poses and geometry with greater 3D consistency." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 390, + 165 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 390, + 165 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 390, + 165 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 390, + 165 + ], + "type": "image", + "image_path": "270e83fd8c2d0e96f009a7a1567b7bd9c069d9fae7a8d536c52e2a421f835f56.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 167, + 389, + 183 + ], + "lines": [ + { + "bbox": [ + 310, + 167, + 389, + 183 + ], + "spans": [ + { + "bbox": [ + 310, + 167, + 389, + 183 + ], + "type": "text", + "content": "(c) Photometric error calculation" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 399, + 114, + 481, + 165 + ], + "blocks": [ + { + "bbox": [ + 399, + 114, + 481, + 165 + ], + "lines": [ + { + "bbox": [ + 399, + 114, + 481, + 165 + ], + "spans": [ + { + "bbox": [ + 399, + 114, + 481, + 165 + ], + "type": "image", + "image_path": "764685c9739176877aaeb68fb9c57a6985437e4348fd6824c164eb2c81fc9f00.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 167, + 470, + 184 + ], + "lines": [ + { + "bbox": [ + 409, + 167, + 470, + 184 + ], + "spans": [ + { + "bbox": [ + 409, + 167, + 470, + 184 + ], + "type": "text", + "content": "(d) Pose and geometry update" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 262, + 479, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 262, + 479, + 287 + ], + "spans": [ + { + "bbox": [ + 131, + 262, + 479, + 287 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{\\mathrm{WC}_r}" + }, + { + "bbox": [ + 131, + 262, + 479, + 287 + ], + "type": "text", + "content": ", and the keyframe image. Anchor points are transformed and projected to pixel coordinates " + }, + { + "bbox": [ + 131, + 262, + 479, + 287 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 131, + 262, + 479, + 287 + ], + "type": "text", + "content": " via camera projection " + }, + { + "bbox": [ + 131, + 262, + 479, + 287 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 131, + 262, + 479, + 287 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 202, + 296, + 481, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 296, + 481, + 312 + ], + "spans": [ + { + "bbox": [ + 202, + 296, + 481, + 312 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {\\mathrm {C} _ {r}} ^ {m} = \\mathbf {R} _ {\\mathrm {W C} _ {r}} ^ {T} \\left(\\mathbf {P} _ {\\mathrm {W}} ^ {m} - \\mathbf {t} _ {\\mathrm {W C} _ {r}}\\right), \\quad \\mathbf {p} ^ {m} = \\pi \\left(\\mathbf {P} _ {\\mathrm {C} _ {r}} ^ {m}\\right), \\tag {2}", + "image_path": "bb4e0e3e68006af6551778d1a918677e252928a4ff7f2ebcdaff4adf2558686c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 320, + 482, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 320, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 131, + 320, + 482, + 417 + ], + "type": "text", + "content": "which is also illustrated in Fig. 3a. At this point, we create a dense log-depth map from log-depth observations " + }, + { + "bbox": [ + 131, + 320, + 482, + 417 + ], + "type": "inline_equation", + "content": "d^{m} \\triangleq \\log \\left( [\\mathbf{P}_{\\mathrm{Cr}}^{m}]_{z} \\right)" + }, + { + "bbox": [ + 131, + 320, + 482, + 417 + ], + "type": "text", + "content": " and projected pixel coordinates " + }, + { + "bbox": [ + 131, + 320, + 482, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^{m}" + }, + { + "bbox": [ + 131, + 320, + 482, + 417 + ], + "type": "text", + "content": ". Given an image and the covariance function " + }, + { + "bbox": [ + 131, + 320, + 482, + 417 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 320, + 482, + 417 + ], + "type": "text", + "content": " for any two pixels from Sec. 3.1, we can construct the covariance matrix between all visible anchor point pixels in " + }, + { + "bbox": [ + 131, + 320, + 482, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{\\mathrm{MM}}" + }, + { + "bbox": [ + 131, + 320, + 482, + 417 + ], + "type": "text", + "content": ", and the covariance between all image pixels and anchor point pixels in " + }, + { + "bbox": [ + 131, + 320, + 482, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{\\mathrm{NM}}" + }, + { + "bbox": [ + 131, + 320, + 482, + 417 + ], + "type": "text", + "content": ". We also stack all training " + }, + { + "bbox": [ + 131, + 320, + 482, + 417 + ], + "type": "inline_equation", + "content": "d^{m}" + }, + { + "bbox": [ + 131, + 320, + 482, + 417 + ], + "type": "text", + "content": " and test " + }, + { + "bbox": [ + 131, + 320, + 482, + 417 + ], + "type": "inline_equation", + "content": "d^{n}" + }, + { + "bbox": [ + 131, + 320, + 482, + 417 + ], + "type": "text", + "content": " log-depths into " + }, + { + "bbox": [ + 131, + 320, + 482, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{\\mathrm{M}}" + }, + { + "bbox": [ + 131, + 320, + 482, + 417 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 320, + 482, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{\\mathrm{N}}" + }, + { + "bbox": [ + 131, + 320, + 482, + 417 + ], + "type": "text", + "content": ", respectively. The GP formulation allows efficient dense log-depth prediction via the linear Gaussian conditioning equation:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 249, + 426, + 481, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 426, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 249, + 426, + 481, + 441 + ], + "type": "interline_equation", + "content": "\\mathbf {d} _ {\\mathrm {N}} = \\mathbf {K} _ {\\mathrm {N M}} \\left(\\mathbf {K} _ {\\mathrm {M M}}\\right) ^ {- 1} \\mathbf {d} _ {\\mathrm {M}}. \\tag {3}", + "image_path": "1a4fd65049bf3b745554455f3fd0fb8342996f836c6e05904e626cd6e2facf40.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 450, + 479, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 450, + 479, + 487 + ], + "spans": [ + { + "bbox": [ + 131, + 450, + 479, + 487 + ], + "type": "text", + "content": "The GP guarantees that each dense log-depth map goes through the anchor points, creating a consistent surface across multiple views. Next, we backproject points for each pixel and transform to world coordinates:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 230, + 496, + 481, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 496, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 230, + 496, + 481, + 512 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {\\mathrm {W}} ^ {n} = \\mathbf {R} _ {\\mathrm {W C} _ {r}} \\pi^ {- 1} \\left(\\mathbf {p} ^ {n}, e ^ {d ^ {n}}\\right) + \\mathbf {t} _ {\\mathrm {W C} _ {r}}, \\tag {4}", + "image_path": "dc13673ceb8dfc70c21f6472de24d8b11a384c563bfbaacad217653e56b0aa4b.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 521, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 592 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 521, + 482, + 592 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 521, + 482, + 592 + ], + "type": "text", + "content": " indexes a query pixel. Note that each keyframe generates a dense depth map via its observed anchor points and covariance parameter feature maps as in Fig. 3b. This compact-to-dense formulation only requires transformations, camera projection, and linear GP conditioning. Therefore, we can efficiently calculate analytical Jacobians for dense per-pixel constraints with respect to poses and anchor points. We show the chain rule for these steps in Section A.1." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "In practice, we ignore the Jacobians with respect to pixel coordinates when constructing " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{\\mathrm{MM}}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{\\mathrm{NM}}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " and calculate the covariance matrices once at initialization. In addition to being more efficient, this ensures that depth maps do not undergo unstable changes if a point moves across a depth discontinuity. If a visible anchor point moves behind a camera during optimization, we reinitialize it to the median depth of the keyframe." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 296, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 296, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 296, + 101 + ], + "type": "text", + "content": "E. Dexheimer and A.J. Davison" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 276, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 276, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 276, + 126 + ], + "type": "text", + "content": "3.3 Photometric Residuals" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 132, + 479, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 132, + 479, + 204 + ], + "spans": [ + { + "bbox": [ + 130, + 132, + 479, + 204 + ], + "type": "text", + "content": "While only keyframes are used to create dense geometry, additional small baseline support frames are included to improve convergence. We define a set of edges " + }, + { + "bbox": [ + 130, + 132, + 479, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 132, + 479, + 204 + ], + "type": "text", + "content": " for photometric error that consists of temporally adjacent keyframe-keyframe pairs, and support frames to the two nearest keyframes in time. Given a target image " + }, + { + "bbox": [ + 130, + 132, + 479, + 204 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 130, + 132, + 479, + 204 + ], + "type": "text", + "content": ", which may be either a keyframe or support frame, with pose " + }, + { + "bbox": [ + 130, + 132, + 479, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{WC}_t}" + }, + { + "bbox": [ + 130, + 132, + 479, + 204 + ], + "type": "text", + "content": ", we first transform world points from keyframe view " + }, + { + "bbox": [ + 130, + 132, + 479, + 204 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 132, + 479, + 204 + ], + "type": "text", + "content": " and project:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 244, + 209, + 481, + 223 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 209, + 481, + 223 + ], + "spans": [ + { + "bbox": [ + 244, + 209, + 481, + 223 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {t} ^ {n} = \\pi \\left(\\mathbf {R} _ {\\mathrm {W C} _ {t}} ^ {T} \\left(\\mathbf {P} _ {\\mathrm {W}} ^ {n} - \\mathbf {t} _ {\\mathrm {W C} _ {t}}\\right)\\right). \\tag {5}", + "image_path": "0859e48170734f1abd2044c11a241fdfbb6d834a9e97c6388584f43e64733ee5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 226, + 479, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 226, + 479, + 262 + ], + "spans": [ + { + "bbox": [ + 131, + 226, + 479, + 262 + ], + "type": "text", + "content": "Given these projective correspondences and accounting for exposure and global illumination changes via affine brightness parameters, " + }, + { + "bbox": [ + 131, + 226, + 479, + 262 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 131, + 226, + 479, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 226, + 479, + 262 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 131, + 226, + 479, + 262 + ], + "type": "text", + "content": ", similar to [10], we calculate the photometric error across pixels and frame edges:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 267, + 481, + 295 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 267, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 481, + 295 + ], + "type": "interline_equation", + "content": "E = \\sum_ {r, t \\in \\mathcal {E}} \\sum_ {n} \\left\\| \\mathbf {r} _ {r, t} ^ {n} \\right\\| _ {\\sigma_ {\\mathbf {r}} ^ {2} I} ^ {2}, \\quad \\mathbf {r} _ {r, t} ^ {n} = I _ {t} \\left(\\mathbf {p} _ {t} ^ {n}\\right) + b _ {t} - \\left(\\frac {e ^ {- a _ {r}}}{e ^ {- a _ {t}}} I _ {r} \\left(\\mathbf {p} _ {r} ^ {n}\\right) + b _ {r}\\right), \\tag {6}", + "image_path": "64e6a83beaa2edb2044f1d161f15ffbdc81cd4999847d9cf782dd0b44a76cb36.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 300, + 480, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 480, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 480, + 395 + ], + "type": "text", + "content": "which is illustrated in Fig. 3c. These constraints depend on each frame's pose and affine parameters, and the anchor points viewed by the reference keyframe. For robustness, we perform iteratively reweighted least squares (IRLS) with a Huber cost function and set the photometric standard deviation " + }, + { + "bbox": [ + 130, + 300, + 480, + 395 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathbf{r}}" + }, + { + "bbox": [ + 130, + 300, + 480, + 395 + ], + "type": "text", + "content": " to be proportional to the median absolute residual [1] to handle diverse scenes with different error scales. Compared to RGB residuals, we found that grayscale images were more efficient and similarly robust. The Jacobian chain rule for photometric residuals is in Section A.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 410, + 276, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 410, + 276, + 421 + ], + "spans": [ + { + "bbox": [ + 132, + 410, + 276, + 421 + ], + "type": "text", + "content": "3.4 Additional Constraints" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 426, + 479, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 479, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 479, + 485 + ], + "type": "text", + "content": "While photometric error is the only data term, we introduce additional priors to remove ambiguities. Priors on the oldest keyframe's pose and affine brightness parameters remove gauge freedom in the sliding-window. After removing a keyframe from the window, we include priors on its still-observed landmarks to ensure global consistency as a simpler alternative to first-estimate Jacobians [10]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 487, + 480, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 480, + 568 + ], + "type": "text", + "content": "Since keyframes may have regions unobserved by other frames, which causes anchor points to lack photometric constraints, we include two geometry priors. The first is a weak prior on anchor points to match the log median depth " + }, + { + "bbox": [ + 130, + 487, + 480, + 568 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 487, + 480, + 568 + ], + "type": "text", + "content": " of the keyframe it was first observed in. The second is a GP prior " + }, + { + "bbox": [ + 130, + 487, + 480, + 568 + ], + "type": "inline_equation", + "content": "||\\mathbf{d}_{\\mathrm{M}} - s||_{\\mathbf{K}_{\\mathrm{MM}}}^2" + }, + { + "bbox": [ + 130, + 487, + 480, + 568 + ], + "type": "text", + "content": " on every keyframe to regularize anchor points that have large correlation, which gives improved estimates for underconstrained anchor points lying on surfaces with well-constrained ones, such as on textureless walls." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 479, + 665 + ], + "type": "text", + "content": "During optimization, there is ambiguity in whether an anchor point changes depth or moves laterally along a surface, such as when viewing a plane at an angle. Thus, we include a pixel prior encouraging the projection of the point in its initializing keyframe not to change from its first observation, which defines a ray-surface intersection. As mentioned previously, this is more efficient than calculating Jacobians through Sec. 3.2 and more stable since GP inducing point optimization has little benefit when points start in suitable locations [5] as ensured by our frontend in Sec. 4." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "COMO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 216, + 176 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 216, + 176 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 216, + 176 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 216, + 176 + ], + "type": "image", + "image_path": "bf2b1436a9bd43ef487970386dda0de40fa46524793e44d7149c6d906323cf10.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 177, + 206, + 194 + ], + "lines": [ + { + "bbox": [ + 141, + 177, + 206, + 194 + ], + "spans": [ + { + "bbox": [ + 141, + 177, + 206, + 194 + ], + "type": "text", + "content": "(a) Anchor points visible in KF #1" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 220, + 114, + 304, + 176 + ], + "blocks": [ + { + "bbox": [ + 220, + 114, + 304, + 176 + ], + "lines": [ + { + "bbox": [ + 220, + 114, + 304, + 176 + ], + "spans": [ + { + "bbox": [ + 220, + 114, + 304, + 176 + ], + "type": "image", + "image_path": "c96b1c6c85ac56ac8790cd63e74819fd78715220f95ce9e92ed5a6ec6e07a935.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 177, + 298, + 194 + ], + "lines": [ + { + "bbox": [ + 225, + 177, + 298, + 194 + ], + "spans": [ + { + "bbox": [ + 225, + 177, + 298, + 194 + ], + "type": "text", + "content": "(b) KF #1 depths projected to KF #2" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 202, + 482, + 248 + ], + "lines": [ + { + "bbox": [ + 130, + 202, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 482, + 248 + ], + "type": "text", + "content": "Fig. 4: Overview of visibility checks between two keyframes (KF). Matches are shown in blue, rejected matches in red, and newly initialized points in green. Note that occluded edges are rejected, while non-visually distinct points are often matched. New points are allocated to geometrically complex regions while the table is already well-represented." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 308, + 114, + 391, + 176 + ], + "blocks": [ + { + "bbox": [ + 308, + 114, + 391, + 176 + ], + "lines": [ + { + "bbox": [ + 308, + 114, + 391, + 176 + ], + "spans": [ + { + "bbox": [ + 308, + 114, + 391, + 176 + ], + "type": "image", + "image_path": "dea814d6b88eb52d91e1c2636e193b468772fa1b8a095c1a96722a7219fffc52.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 177, + 383, + 194 + ], + "lines": [ + { + "bbox": [ + 316, + 177, + 383, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 177, + 383, + 194 + ], + "type": "text", + "content": "(c) Anchor point matches in KF #2" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 399, + 114, + 482, + 176 + ], + "blocks": [ + { + "bbox": [ + 399, + 114, + 482, + 176 + ], + "lines": [ + { + "bbox": [ + 399, + 114, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 399, + 114, + 482, + 176 + ], + "type": "image", + "image_path": "c52c02513e3d47af233accf282aa43f65847462be8b4d25371c2b54437aebe52.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 177, + 478, + 194 + ], + "lines": [ + { + "bbox": [ + 402, + 177, + 478, + 194 + ], + "spans": [ + { + "bbox": [ + 402, + 177, + 478, + 194 + ], + "type": "text", + "content": "(d) Newly initialized points in KF #2" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 268, + 299, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 299, + 280 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 299, + 280 + ], + "type": "text", + "content": "3.5 Second-Order Optimization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 285, + 482, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 357 + ], + "type": "text", + "content": "We calculate analytical Jacobians for all constraints in the optimization. The compactness of our linear system permits real-time joint optimization of poses and correlated geometry that is infeasible for traditional dense VO. Stacking all Jacobians into " + }, + { + "bbox": [ + 130, + 285, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{J}" + }, + { + "bbox": [ + 130, + 285, + 482, + 357 + ], + "type": "text", + "content": ", weights from IRLS and noise models into the diagonal matrix " + }, + { + "bbox": [ + 130, + 285, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 130, + 285, + 482, + 357 + ], + "type": "text", + "content": ", and residuals into " + }, + { + "bbox": [ + 130, + 285, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 130, + 285, + 482, + 357 + ], + "type": "text", + "content": ", we solve for parameter updates " + }, + { + "bbox": [ + 130, + 285, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{x}" + }, + { + "bbox": [ + 130, + 285, + 482, + 357 + ], + "type": "text", + "content": " shown in Fig. 3d via dense Cholesky factorization of the Gauss-Newton normal equations:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 253, + 361, + 481, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 361, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 253, + 361, + 481, + 376 + ], + "type": "interline_equation", + "content": "\\left(\\mathbf {J} ^ {T} \\mathbf {W} \\mathbf {J}\\right) \\Delta \\mathbf {x} = - \\mathbf {J} ^ {T} \\mathbf {W r}. \\tag {7}", + "image_path": "7928d6f1da504e07e3e098040da2fc3c618afbff3f5405bb7f08a52db50b4fcb.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 380, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 441 + ], + "type": "text", + "content": "Pose updates are minimally parameterized as Lie-algebra elements " + }, + { + "bbox": [ + 130, + 380, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathfrak{se}(3)" + }, + { + "bbox": [ + 130, + 380, + 482, + 441 + ], + "type": "text", + "content": " as is standard in SLAM [32]. We found that the most time consuming step of backend optimization is accumulating geometry blocks into the Hessian " + }, + { + "bbox": [ + 130, + 380, + 482, + 441 + ], + "type": "inline_equation", + "content": "(\\mathbf{J}^T\\mathbf{W}\\mathbf{J})" + }, + { + "bbox": [ + 130, + 380, + 482, + 441 + ], + "type": "text", + "content": ". This can be sped up by factoring out constant components of the Jacobians and reducing the dimensionality of the accumulation as detailed in Section A.4." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 456, + 250, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 250, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 250, + 468 + ], + "type": "text", + "content": "4 Visual Frontend" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 478, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 478, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 478, + 482, + 538 + ], + "type": "text", + "content": "The visual frontend tracks the current frame with respect to the map, selects keyframes and support frames for the backend, determines 3D anchor point visibility in new keyframes, and initializes new 3D anchor points. Compared to feature-based systems, the ideal points for an expressive yet compact map may be visually indistinct, so we utilize depth covariance for robust visibility checks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 552, + 306, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 552, + 306, + 565 + ], + "spans": [ + { + "bbox": [ + 132, + 552, + 306, + 565 + ], + "type": "text", + "content": "4.1 Dense Photometric Tracking" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Given the current estimate of the newest keyframe's depth, we optimize the pose and affine brightness parameters using grayscale image alignment in the IRLS framework [2]. For efficiency, we perform inverse compositional tracking and use a multi-scale image pyramid. After tracking, a new keyframe is added to the backend if there is either significant translation relative to the median scene depth or if the number of unique projected pixels in the tracked frame falls below a threshold to handle rotation and occlusion. Support frames are added via modified thresholds according to the desired frequency between keyframes." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 296, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 296, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 296, + 101 + ], + "type": "text", + "content": "E. Dexheimer and A.J. Davison" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 279, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 279, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 279, + 129 + ], + "type": "text", + "content": "4.2 Anchor Point Visibility" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 135, + 482, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 135, + 482, + 230 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 482, + 230 + ], + "type": "text", + "content": "When a new keyframe is added, we perform explicit visibility checks to 3D anchor points viewed in the last keyframe to achieve 3D consistency. Note that we only check for visibility as anchor points have no explicit constraint on its location in the image. Compared to feature-based methods [24], anchor points often lack visual distinctiveness, so matching via image information is problematic. Instead, we leverage the interpretability of the depth covariance function to determine visibility. In general, an anchor point should be compatible with all depth covariance functions that use it to generate dense geometry." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 231, + 482, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 231, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 482, + 327 + ], + "type": "text", + "content": "First, the 3D anchor points and dense depth map from the last keyframe, shown in Fig. 4a, are projected into the current frame, shown in Fig. 4b. Given the projected sparse and dense pixel coordinates, we may query covariance parameters " + }, + { + "bbox": [ + 130, + 231, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\phi (\\mathbf{p})" + }, + { + "bbox": [ + 130, + 231, + 482, + 327 + ], + "type": "text", + "content": " and construct " + }, + { + "bbox": [ + 130, + 231, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{\\mathrm{MM}}" + }, + { + "bbox": [ + 130, + 231, + 482, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 231, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{\\mathrm{MN}}" + }, + { + "bbox": [ + 130, + 231, + 482, + 327 + ], + "type": "text", + "content": " for the new keyframe. We leverage the linearity of the GP to compress dense log-depth observations into the sparse projected anchor point coordinates. Assuming conditional independence of log-depth observations " + }, + { + "bbox": [ + 130, + 231, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_n" + }, + { + "bbox": [ + 130, + 231, + 482, + 327 + ], + "type": "text", + "content": " given the sparse log-depths " + }, + { + "bbox": [ + 130, + 231, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_m" + }, + { + "bbox": [ + 130, + 231, + 482, + 327 + ], + "type": "text", + "content": ", we can solve an efficient least-squares problem:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 203, + 334, + 480, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 334, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 203, + 334, + 480, + 353 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {d} _ {\\mathrm {M}}} \\left| \\left| \\mathbf {K} _ {\\mathrm {N M}} \\left(\\mathbf {K} _ {\\mathrm {M M}}\\right) ^ {- 1} \\mathbf {d} _ {\\mathrm {M}} - \\mathbf {d} _ {\\mathrm {N}} \\right| \\right| _ {\\sigma_ {d} ^ {2}} ^ {2} + \\left| \\left| \\mathbf {d} _ {\\mathrm {M}} \\right| \\right| _ {\\mathbf {K} _ {\\mathrm {M M}}} ^ {2}, \\tag {8}", + "image_path": "715aa1f01579384e476e26f4e7c06113ea1944c4a21dd594d771c264f5a8a3f6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 360, + 482, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 482, + 431 + ], + "type": "text", + "content": "where the first conditional term ensures that " + }, + { + "bbox": [ + 130, + 360, + 482, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{\\mathrm{M}}" + }, + { + "bbox": [ + 130, + 360, + 482, + 431 + ], + "type": "text", + "content": " fits the projected dense observations with standard deviation " + }, + { + "bbox": [ + 130, + 360, + 482, + 431 + ], + "type": "inline_equation", + "content": "\\sigma_{d}" + }, + { + "bbox": [ + 130, + 360, + 482, + 431 + ], + "type": "text", + "content": ", while the second prior term regularizes sparse log-depths according to the GP training covariance " + }, + { + "bbox": [ + 130, + 360, + 482, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{\\mathrm{MM}}" + }, + { + "bbox": [ + 130, + 360, + 482, + 431 + ], + "type": "text", + "content": ". We perform a consistency check on the log-depth difference between the current 3D anchor points and those solved in Eq. (8), and prune points at depth discontinuities. This yields potential matches the are consistent between the two keyframes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 432, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 482, + 515 + ], + "type": "text", + "content": "In practice, we set a maximum number of visible anchor points per-frame to allow efficient batched mapping operations on the GPU. If all points are matched, it could prevent new ones from being initialized and parts of the scene would lack capacity to adequately represent geometry. Thus, we perform pruning of matches so that we have a suitable distribution of points before initializing new anchors. Similar to the GP conditional mean from Eq. (3), we may also obtain conditional variances of all pixels with respect to the first " + }, + { + "bbox": [ + 130, + 432, + 482, + 515 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 432, + 482, + 515 + ], + "type": "text", + "content": " pixels via:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 199, + 524, + 480, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 524, + 480, + 539 + ], + "spans": [ + { + "bbox": [ + 199, + 524, + 480, + 539 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\sigma} _ {j} ^ {2} = \\operatorname {d i a g} \\left[ \\mathbf {K} _ {\\mathrm {N N}} \\right] - \\mathbf {K} _ {\\mathrm {N M} _ {1: j}} \\left(\\mathbf {K} _ {\\mathrm {M} _ {1: j} \\mathrm {M} _ {1: j}}\\right) ^ {- 1} \\mathbf {K} _ {\\mathrm {M} _ {1: j} \\mathrm {N}}. \\tag {9}", + "image_path": "9c55f1cd5b489f82209ab6e43e96ade0402392a586f05449cb8134bf3a77f4de.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "We perform active sampling via conditional variance reduction (CVR) [14, 22], where at the " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "(j + 1)^{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " iteration, the pixel in domain " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " with the largest variance is selected " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{j + 1} = \\arg \\max_{\\mathbf{p}\\in P}\\sigma_j^2 (\\mathbf{p})" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ". We actively sample from the set of potential visible points until a variance threshold is reached, which gives the final set of visible anchor points for the current keyframe. Compared to [9], we found that a minimum distance threshold between pixel locations and from the image border is required to avoid accumulating pixels near the same discontinuity. This is essential for improving depth estimates and maintaining a good set of candidate points, as otherwise many points near edges may be occluded at the same time. We show the set of matches between two keyframes in Fig. 4c." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "COMO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 194, + 160 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 194, + 160 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 194, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 194, + 160 + ], + "type": "image", + "image_path": "56b0853a238dff6d20aab3b8a4748b88afbf94a7b5982e685223e2357894f749.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 163, + 193, + 206 + ], + "blocks": [ + { + "bbox": [ + 133, + 163, + 193, + 206 + ], + "lines": [ + { + "bbox": [ + 133, + 163, + 193, + 206 + ], + "spans": [ + { + "bbox": [ + 133, + 163, + 193, + 206 + ], + "type": "image", + "image_path": "c2f05c3b6d3c4e0a72eb49baedec167687f5851e8815f2b9b0be9f6b495379a4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 217, + 482, + 273 + ], + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 273 + ], + "type": "text", + "content": "Fig. 5: Visualization of sliding-window odometry on TUM " + }, + { + "bbox": [ + 130, + 217, + 482, + 273 + ], + "type": "inline_equation", + "content": "fr2/desk" + }, + { + "bbox": [ + 130, + 217, + 482, + 273 + ], + "type": "text", + "content": ". Top images show tracked and newly initialized anchor points in blue and red, respectively. Depth maps are below, while the dense point cloud with 248 anchor points is shown on the right. Note the consistent point tracks on textureless surfaces, such as the table and plate, as well as the long-term tracks on the cube, mug, and book." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 197, + 115, + 256, + 160 + ], + "blocks": [ + { + "bbox": [ + 197, + 115, + 256, + 160 + ], + "lines": [ + { + "bbox": [ + 197, + 115, + 256, + 160 + ], + "spans": [ + { + "bbox": [ + 197, + 115, + 256, + 160 + ], + "type": "image", + "image_path": "2a44ec5c85e607d5515af8b0f78286a651d50804b41ecd175e5bc8e2fe690a05.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 197, + 163, + 256, + 206 + ], + "blocks": [ + { + "bbox": [ + 197, + 163, + 256, + 206 + ], + "lines": [ + { + "bbox": [ + 197, + 163, + 256, + 206 + ], + "spans": [ + { + "bbox": [ + 197, + 163, + 256, + 206 + ], + "type": "image", + "image_path": "d51115a4ffb4b9a4917f662252f8ad3e1bacdff8753d9c6e6d208180936df3f1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 260, + 115, + 319, + 160 + ], + "blocks": [ + { + "bbox": [ + 260, + 115, + 319, + 160 + ], + "lines": [ + { + "bbox": [ + 260, + 115, + 319, + 160 + ], + "spans": [ + { + "bbox": [ + 260, + 115, + 319, + 160 + ], + "type": "image", + "image_path": "0461e0d585cbc0f7a2a144b675a90b51aaf5154a231c556ca6de222063648ec6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 260, + 163, + 319, + 206 + ], + "blocks": [ + { + "bbox": [ + 260, + 163, + 319, + 206 + ], + "lines": [ + { + "bbox": [ + 260, + 163, + 319, + 206 + ], + "spans": [ + { + "bbox": [ + 260, + 163, + 319, + 206 + ], + "type": "image", + "image_path": "482264346b9ade53b0a1c67a9cdab492dcc1ae8557215146b8e2d74c3a5ed108.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 323, + 115, + 381, + 160 + ], + "blocks": [ + { + "bbox": [ + 323, + 115, + 381, + 160 + ], + "lines": [ + { + "bbox": [ + 323, + 115, + 381, + 160 + ], + "spans": [ + { + "bbox": [ + 323, + 115, + 381, + 160 + ], + "type": "image", + "image_path": "0f8562bc9a69aa9ef3ce4e8b3b25f30df79fc97687994bac87ab8de8fb828f91.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 323, + 163, + 381, + 206 + ], + "blocks": [ + { + "bbox": [ + 323, + 163, + 381, + 206 + ], + "lines": [ + { + "bbox": [ + 323, + 163, + 381, + 206 + ], + "spans": [ + { + "bbox": [ + 323, + 163, + 381, + 206 + ], + "type": "image", + "image_path": "2aff15b414a1195164d0b78fdedcb40db594370f8c1b24263b73bd4eba40aa4b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 388, + 114, + 479, + 198 + ], + "blocks": [ + { + "bbox": [ + 388, + 114, + 479, + 198 + ], + "lines": [ + { + "bbox": [ + 388, + 114, + 479, + 198 + ], + "spans": [ + { + "bbox": [ + 388, + 114, + 479, + 198 + ], + "type": "image", + "image_path": "165e0447c1fb62b31213ee9ba4ac35b0b083cc68aa4ed3c57a7eaa20a01455ef.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 198, + 413, + 206 + ], + "lines": [ + { + "bbox": [ + 403, + 198, + 413, + 206 + ], + "spans": [ + { + "bbox": [ + 403, + 198, + 413, + 206 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 295, + 296, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 295, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 295, + 296, + 306 + ], + "type": "text", + "content": "4.3 Anchor Point Initialization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 314, + 479, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 479, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 479, + 373 + ], + "type": "text", + "content": "As the camera explores unobserved parts of the scene, new anchor points must be initialized. First, we again perform active sampling, but now across all pixels given the current anchor point locations from the previous section. This ensures new points will be initialized for parts of the scene that are unobserved, lack matches, or require additional geometric capacity, as shown in Fig. 4d." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 374, + 480, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 480, + 423 + ], + "type": "text", + "content": "Once the pixel locations of new anchor points are known, we initialize the log-depth and backproject into 3D. Defining the set of known log-depths from the anchor points " + }, + { + "bbox": [ + 130, + 374, + 480, + 423 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{\\mathrm{m}_1}" + }, + { + "bbox": [ + 130, + 374, + 480, + 423 + ], + "type": "text", + "content": ", the unknown log-depths of newly initialized points " + }, + { + "bbox": [ + 130, + 374, + 480, + 423 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{\\mathrm{m}_2}" + }, + { + "bbox": [ + 130, + 374, + 480, + 423 + ], + "type": "text", + "content": " and all log-depths as " + }, + { + "bbox": [ + 130, + 374, + 480, + 423 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{\\mathrm{M}} = (\\mathbf{d}_{\\mathrm{m}_1}^T,\\mathbf{d}_{\\mathrm{m}_2}^T)^T" + }, + { + "bbox": [ + 130, + 374, + 480, + 423 + ], + "type": "text", + "content": ", we solve another least-squares problem:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 196, + 440, + 480, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 440, + 480, + 462 + ], + "spans": [ + { + "bbox": [ + 196, + 440, + 480, + 462 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {d} _ {\\mathrm {m} _ {2}}} \\left\\| \\mathbf {K} _ {\\mathrm {N M}} \\left(\\mathbf {K} _ {\\mathrm {M M}}\\right) ^ {- 1} \\mathbf {d} _ {\\mathrm {M}} - \\mathbf {d} _ {\\mathrm {N}} \\right\\| _ {\\sigma_ {d} ^ {2}} ^ {2} + \\left\\| \\mathbf {d} _ {\\mathrm {m} _ {2}} - s \\right\\| _ {\\sigma_ {s} ^ {2}}, \\tag {10}", + "image_path": "ef0c95ddbc48dccfb6dea00662371739bde2269572de1816851758cc9c6e58d6.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 468, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 601 + ], + "type": "text", + "content": "where the first term promotes fitting reprojected dense depth observations given matched anchor log-depths and the new log-depth initializations, while the second term is a prior that new log-depths are close to the log median depth " + }, + { + "bbox": [ + 130, + 468, + 482, + 601 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 468, + 482, + 601 + ], + "type": "text", + "content": " of the previous keyframe. We found that using an isotropic prior with standard deviation " + }, + { + "bbox": [ + 130, + 468, + 482, + 601 + ], + "type": "inline_equation", + "content": "\\sigma_d" + }, + { + "bbox": [ + 130, + 468, + 482, + 601 + ], + "type": "text", + "content": ", rather than " + }, + { + "bbox": [ + 130, + 468, + 482, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{\\mathrm{MM}}" + }, + { + "bbox": [ + 130, + 468, + 482, + 601 + ], + "type": "text", + "content": ", avoids bias for parts of the image that were not previously observed and would otherwise be underconstrained. We define " + }, + { + "bbox": [ + 130, + 468, + 482, + 601 + ], + "type": "inline_equation", + "content": "\\sigma_d" + }, + { + "bbox": [ + 130, + 468, + 482, + 601 + ], + "type": "text", + "content": " as the standard deviation of the residuals from Eq. (8) to better calibrate the model. Given the new pixel locations, log-depth initializations, and the new keyframe's pose estimate, 3D points are initialized for the backend. A qualitative example of tracked anchor points, depth maps, and reconstruction is shown in Fig. 5." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 617, + 230, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 230, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 230, + 631 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "We evaluate both the pose and geometry estimation of our method on multiple datasets. For sparse methods, we compare against feature-based ORB-" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 296, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 296, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 296, + 101 + ], + "type": "text", + "content": "E. Dexheimer and A.J. Davison" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 144, + 114, + 470, + 172 + ], + "blocks": [ + { + "bbox": [ + 144, + 114, + 470, + 172 + ], + "lines": [ + { + "bbox": [ + 144, + 114, + 470, + 172 + ], + "spans": [ + { + "bbox": [ + 144, + 114, + 470, + 172 + ], + "type": "table", + "html": "
office0office1office2office3office4room0room1room2mean
SORB-SLAM3 [6]0.40.38.50.44.50.30.30.41.9
DSO [10]0.30.50.30.25.30.20.80.21.0
DDeepFactors [7]42.938.645.641.450.132.426.144.140.2
DROID-VO [37]5.34.16.511.27.19.04.87.06.9
COMO-NC2.95.07.46.18.14.44.53.75.3
COMO2.42.44.33.65.52.52.73.43.4
", + "image_path": "f9fe0afff8493118966b86cc780ff1f9dbad9eca31f1012b535425eaef18fe38.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 172, + 481, + 195 + ], + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 195 + ], + "type": "text", + "content": "Table 1: Mean ATE (cm) over 5 runs for monocular VO on the Replica dataset. S and D in the first column refer to sparse and dense methods, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 225, + 482, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 482, + 371 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 482, + 371 + ], + "type": "text", + "content": "SLAM3 [6], direct DSO [10], and learning-based DPVO [38] in trajectory estimation. We also compare against a representative set of dense methods. TANDEM [18] uses DSO for odometry and fuses learned MVS predictions into a TSDF for tracking. MonoGS [19] is a 3D Gaussian splatting-based monocular SLAM system. DeepFactors [7] is a code-based system with photometric, geometric, and keypoint losses to promote consistency across depth maps. DROIDVO [37] uses learned correspondence and flow updates for bundle adjustment, such that all depths are reconstructed but lack geometric correlation. DepthCov [9] optimizes sparse 2D depths under the depth covariance framework, and in a related manner, COMO-NC (no correspondence) is our full framework but no 3D anchor points are tracked across frames. We disable loop closure for DeepFactors and DROID-SLAM to isolate the odometry and mapping performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 385, + 279, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 279, + 398 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 279, + 398 + ], + "type": "text", + "content": "5.1 Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "text", + "content": "We use a fixed configuration across all three datasets to demonstrate the robustness of the system. We use the off-the-shelf depth covariance function [9] that was trained on the ScanNet training set. Our sliding window odometry uses 9 keyframes and 3 support frames between keyframes for a total of 24. Each keyframe views a maximum of 64 anchor points and initializes new points if it fails to track all from the previous keyframe. We use 256x192 images for all operations, and the mapping backend samples pixels with the highest gradient magnitude in 4x4 images patches for photometric error." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 515, + 271, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 271, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 271, + 528 + ], + "type": "text", + "content": "5.2 Trajectory Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Replica We first test our monocular odometry on the Replica [33] sequences recorded in [35], which provides a photorealistic environment with ground-truth geometry. As shown in Tab. 1, sparse methods without priors perform best, as conditions are favorable for SLAM: few image artifacts, no exposure changes, and sufficient baselines for multi-view geometry. Among methods with learned priors, COMO achieves the lowest ATE. The representation has little bias as compared to the code-based DeepFactors, which drifts despite photometric, geometric, and keypoint constraints. Sharing 3D points between frames in COMO shows significant improvement over COMO-NC which disables correspondence. TUM The TUM RGBD dataset [34] is a challenging dataset for monocular VO due to significant motion blur, exposure changes, rolling shutter artifacts," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "COMO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 114, + 478, + 212 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 478, + 212 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 478, + 212 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 478, + 212 + ], + "type": "table", + "html": "
fr1fr2fr3 longmean
360deskdesk2plantroomrpyteddyxyzxyzdesk
SORB-SLAM3 [6]X2.0X11.8X5.6X1.00.51.31.7X
DSO [10]X27.266.06.058.6XX3.80.32.29.9X
DPVO [38]13.19.46.53.039.83.56.21.30.53.55.58.4
DTANDEM [18]X4.333.7XX4.943.12.40.32.08.3X
MonoGS [19]14.26.374.09.364.93.435.61.64.5133.13.331.8
DeepFactors [7]17.915.920.231.938.33.856.05.98.426.349.024.9
DepthCov [9]12.85.64.826.125.75.247.55.61.215.968.819.9
DROID-VO [37]15.75.211.16.033.43.219.15.610.77.97.311.4
COMO-NC16.14.210.919.328.65.268.74.10.78.846.819.4
COMO12.94.99.513.827.04.824.54.00.76.310.510.8
", + "image_path": "eab9f23954ccf4d6bf424b815991a8e2204dab666029f6a65d36733403ed0f86.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 213, + 479, + 233 + ], + "lines": [ + { + "bbox": [ + 132, + 213, + 479, + 233 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 479, + 233 + ], + "type": "text", + "content": "Table 2: Mean ATE (cm) over 5 runs for monocular VO on the TUM dataset. S and D in the first column refer to sparse and dense methods, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 133, + 270, + 481, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 270, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 133, + 270, + 481, + 520 + ], + "type": "text", + "content": "and heavy rotations. Results comparing our method against state-of-the-art VO methods on 11 sequences is shown in Tab. 2. For sparse methods, ORB-SLAM3 and DSO both lose tracking on multiple sequences, while DPVO, which uses learned context for tracking 64 sparse patches and performing bundle adjustment, performs best among all methods. In this work, we are focused on also reconstructing dense geometry, but it would be interesting to combine the learned features of DPVO with our 3D representation. For dense methods, TANDEM fails due to relying on DSO for odometry, which can still fail despite using a TSDF with MVS fusion for tracking. This demonstrates the need for joint optimization of poses and dense image information. While MonoGS can produce accurate trajectories for some sequences, the lack of priors reduces its robustness, and volumetric rendering is slower than all depth-based systems. Despite promoting consistency between neighboring depth maps, DeepFactors has relatively large error. Both DepthCov and COMO-NC are similar in terms of optimizing sparse depths in each frame, while leveraging a true 3D representation in COMO reduces error by almost " + }, + { + "bbox": [ + 133, + 270, + 481, + 520 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 133, + 270, + 481, + 520 + ], + "type": "text", + "content": ". COMO outperforms DROID-VO, which uses learned features for matching and reconstructs all pixels from downsized images via sparse bundle adjustment. As the assumptions of brightness constancy are violated in many real-world datasets, learned features could further the accuracy of COMO, but it is interesting that our compact representation and simple photometric error can produce the lowest ATE of dense methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 522, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 522, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 522, + 481, + 665 + ], + "type": "text", + "content": "ScanNet Test Lastly, we evaluate the trajectory error on 18 diverse medium-length scenes from the ScanNet [8] test set. Sequences include homes, offices, schools, businesses, and outdoors. Methods that fail are assigned a maximum " + }, + { + "bbox": [ + 133, + 522, + 481, + 665 + ], + "type": "inline_equation", + "content": "100\\mathrm{cm}" + }, + { + "bbox": [ + 133, + 522, + 481, + 665 + ], + "type": "text", + "content": " error for averages. Sequences have challenging rotational motion, high image noise, and specular surfaces. We summarize the results in Fig. 6, which shows the number of sequences below varying ATE thresholds, while Tab. 3 contains summary statistics. COMO has the lowest mean ATE and area-underthe-curve (AUC). While DPVO has a higher number of trajectories under low ATE thresholds, it also shows high variance results on some sequences. Despite using photometric error that is often violated due to image noise and specular surfaces, our representation with 3D consistency proves valuable in challenging real-world data. An example reconstruction is shown in Fig. 1a." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 295, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 295, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 295, + 100 + ], + "type": "text", + "content": "E. Dexheimer and A.J. Davison" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 376, + 209 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 376, + 209 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 376, + 209 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 376, + 209 + ], + "type": "image", + "image_path": "5b0644cb0379c11ad6848ac680648dd18d2298bfb7ffa3edb0824a115ff97cd5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 217, + 362, + 240 + ], + "lines": [ + { + "bbox": [ + 146, + 217, + 362, + 240 + ], + "spans": [ + { + "bbox": [ + 146, + 217, + 362, + 240 + ], + "type": "text", + "content": "Fig. 6: Number of successful trajectories below ATE threshold for selected ScanNet test sequences." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 376, + 217, + 478, + 240 + ], + "lines": [ + { + "bbox": [ + 376, + 217, + 478, + 240 + ], + "spans": [ + { + "bbox": [ + 376, + 217, + 478, + 240 + ], + "type": "text", + "content": "Table 3: ScanNet Mean ATE (cm) and AUC." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 379, + 116, + 474, + 190 + ], + "blocks": [ + { + "bbox": [ + 379, + 116, + 474, + 190 + ], + "lines": [ + { + "bbox": [ + 379, + 116, + 474, + 190 + ], + "spans": [ + { + "bbox": [ + 379, + 116, + 474, + 190 + ], + "type": "table", + "html": "
MethodATEAUC
ORB-SLAM379.91.63
DSO54.72.46
DPVO15.16.34
TANDEM41.73.79
DeepFactors36.12.97
DROID-VO13.96.51
COMO-NC15.86.16
COMO13.06.67
", + "image_path": "ce2b87cfd9b70424895e2ac21af29a6bb5a31089d4ae818df735a8c23df7a3da.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 134, + 274, + 230, + 313 + ], + "blocks": [ + { + "bbox": [ + 134, + 274, + 230, + 313 + ], + "lines": [ + { + "bbox": [ + 134, + 274, + 230, + 313 + ], + "spans": [ + { + "bbox": [ + 134, + 274, + 230, + 313 + ], + "type": "table", + "html": "
MethodReplicaScanNet
TANDEMN/A0.325
DeepFactors0.2630.232
DROID-VO1.1291.389
COMO-NC0.0690.155
COMO0.0460.128
", + "image_path": "bdd62a13c12f9d0c5d6d3de026b5f1cbf7c34047d06258bd3f363fa4c57af1d1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 236, + 254, + 354, + 331 + ], + "blocks": [ + { + "bbox": [ + 236, + 254, + 354, + 331 + ], + "lines": [ + { + "bbox": [ + 236, + 254, + 354, + 331 + ], + "spans": [ + { + "bbox": [ + 236, + 254, + 354, + 331 + ], + "type": "image", + "image_path": "f59ebe50f0c9a3ce41a1a9f1b25f69c0520fe5b4bba68021f54bf122f360d53d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 342, + 228, + 364 + ], + "lines": [ + { + "bbox": [ + 135, + 342, + 228, + 364 + ], + "spans": [ + { + "bbox": [ + 135, + 342, + 228, + 364 + ], + "type": "text", + "content": "Table 4: Depth absolute relative error." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 235, + 342, + 479, + 365 + ], + "lines": [ + { + "bbox": [ + 235, + 342, + 479, + 365 + ], + "spans": [ + { + "bbox": [ + 235, + 342, + 479, + 365 + ], + "type": "text", + "content": "Fig. 7: ScanNet depth accuracy and consistency metrics for fraction of pixels falling below threshold " + }, + { + "bbox": [ + 235, + 342, + 479, + 365 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 235, + 342, + 479, + 365 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 360, + 255, + 477, + 331 + ], + "blocks": [ + { + "bbox": [ + 360, + 255, + 477, + 331 + ], + "lines": [ + { + "bbox": [ + 360, + 255, + 477, + 331 + ], + "spans": [ + { + "bbox": [ + 360, + 255, + 477, + 331 + ], + "type": "image", + "image_path": "7ccca15e2c6c832ec16a50df124b221a83f84cfbcedda4ef82291a7508cc9f66.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 390, + 268, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 390, + 268, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 390, + 268, + 403 + ], + "type": "text", + "content": "5.3 Geometry Evaluation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 413, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 481, + 569 + ], + "type": "text", + "content": "Since we are interested in both consistent pose estimation and dense geometry, we evaluate dense depth predictions on Replica and ScanNet. To evaluate accuracy, we perform the same global scale alignment that is used to produce ATE for monocular methods and transform estimated poses and depth maps to be aligned to the metric ground-truth. Beyond global accuracy, we also measure consistency, which checks whether neighboring frames viewing the same 3D points have consistent depth estimates. Note that consistency does not care about the correctness of depth. To check valid pixels, we backproject ground-truth depth maps and project these points into neighboring frames, and check if the depths agree within " + }, + { + "bbox": [ + 130, + 413, + 481, + 569 + ], + "type": "inline_equation", + "content": "1\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 413, + 481, + 569 + ], + "type": "text", + "content": ". Then, for all valid 3D correspondences, we calculate metrics between pairs of estimated depth images. If " + }, + { + "bbox": [ + 130, + 413, + 481, + 569 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 413, + 481, + 569 + ], + "type": "text", + "content": " is the number of keyframes for that sequence, then " + }, + { + "bbox": [ + 130, + 413, + 481, + 569 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 413, + 481, + 569 + ], + "type": "text", + "content": " depth images are used for accuracy metrics, while " + }, + { + "bbox": [ + 130, + 413, + 481, + 569 + ], + "type": "inline_equation", + "content": "2(N - 1)" + }, + { + "bbox": [ + 130, + 413, + 481, + 569 + ], + "type": "text", + "content": " depth images are used for consistency since pairs require both directions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "For estimated depth " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{D}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " and ground-truth depth " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " we show the mean absolute relative error " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "(1 / N)\\sum_{i}|\\hat{D}_{i} - D_{i}| / D_{i}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " across all sequences for Replica and ScanNet in Tab. 4. We do not evaluate TANDEM on Replica since it is used in the training data. COMO outperforms other methods due to its ability to jointly optimize dense, correlated geometry along with poses. Despite accurate pose estimation, DROID-VO has outliers with significant depth errors. To further evaluate the distribution of depth accuracy and remove the effect of large outliers, in Fig. 7, we use the " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " threshold commonly used in depth estimation, which measures what" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "COMO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 476, + 162 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 476, + 162 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 476, + 162 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 476, + 162 + ], + "type": "image", + "image_path": "7f478f47dfe1f7c654ec4f147afdc2d47451e77e57ebdbf1468efe96d3bdfbf3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 170, + 463, + 182 + ], + "lines": [ + { + "bbox": [ + 149, + 170, + 463, + 182 + ], + "spans": [ + { + "bbox": [ + 149, + 170, + 463, + 182 + ], + "type": "text", + "content": "Fig. 8: Mesh and surface normals from TSDF fusion of a live desk sequence." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 146, + 190, + 467, + 209 + ], + "blocks": [ + { + "bbox": [ + 146, + 190, + 467, + 209 + ], + "lines": [ + { + "bbox": [ + 146, + 190, + 467, + 209 + ], + "spans": [ + { + "bbox": [ + 146, + 190, + 467, + 209 + ], + "type": "image", + "image_path": "85f665e026a3e998ae5b11fdec72ec79a24e3fd3db86d694802572aaffc31ea7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 209, + 481, + 232 + ], + "lines": [ + { + "bbox": [ + 132, + 209, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 209, + 481, + 232 + ], + "type": "text", + "content": "Table 5: Runtime in seconds of different methods on TUM " + }, + { + "bbox": [ + 132, + 209, + 481, + 232 + ], + "type": "inline_equation", + "content": "fr3/long" + }, + { + "bbox": [ + 132, + 209, + 481, + 232 + ], + "type": "text", + "content": " using an RTX 3080. COMO SP and MP denote single and multiprocessing, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 262, + 480, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 480, + 327 + ], + "type": "text", + "content": "fraction of pixels satisfy " + }, + { + "bbox": [ + 130, + 262, + 480, + 327 + ], + "type": "inline_equation", + "content": "\\max \\left(\\hat{D} / D, D / \\hat{D}\\right) < \\delta" + }, + { + "bbox": [ + 130, + 262, + 480, + 327 + ], + "type": "text", + "content": ". Interestingly, accurate depth maps do not necessarily imply consistency. For example, DROID-VO has the second best accuracy behind COMO, while DeepFactors has the second most consistent geometry. COMO achieves both the best accuracy and consistency without explicit geometric and keypoint constraints." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 342, + 358, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 358, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 358, + 355 + ], + "type": "text", + "content": "5.4 Real-time System and Runtime Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 361, + 480, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 480, + 445 + ], + "type": "text", + "content": "When evaluating our system, we use a sequential, single-threaded mode that runs at around 14 FPS with an RTX 3080. For live demos and reconstructions, we also develop a multiprocessing version that runs at 35-50 FPS as shown in the FPS counter in the top right of the supplementary video, with tracking flexible to be on either CPU or GPU. We show examples of reconstructions while operating a monocular camera in real-time. Fig. 1b displays a large-scale raw point cloud of an office, while Fig. 8 displays consistent TSDF fusion of a desk sequence." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 445, + 481, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 481, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 481, + 517 + ], + "type": "text", + "content": "Average times for tracking a frame, adding a keyframe (KF), and updating the map are 20.1, 50.4, and " + }, + { + "bbox": [ + 130, + 445, + 481, + 517 + ], + "type": "inline_equation", + "content": "39.8~\\mathrm{ms}" + }, + { + "bbox": [ + 130, + 445, + 481, + 517 + ], + "type": "text", + "content": ". Keyframing is the bottleneck during fast motion, but 35-50 FPS in multiprocessing mode for normal motion is standard as shown in the video. We measure the wall time of several methods in Tab. 5. COMO achieves real-time processing with its efficient representation, but further optimization is still possible since almost all code is implemented in PyTorch." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 534, + 220, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 220, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 220, + 547 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": "In this work, we present a compact representation of 3D anchor points and frame-based depth covariance functions to achieve real-time estimation of consistent poses and geometry. Future work includes training the depth covariance function on more diverse data and replacing photometric constraints with the more robust learned correspondence from DROID-VO. Lastly, we believe the anchor point representation could be used in a full map-centric SLAM system with relocalization, which would allow continually improving geometry upon revisiting. With the advance of learned priors, exploiting geometric correlation is an interesting direction to achieve efficient, robust, and consistent algorithms." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 296, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 296, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 296, + 101 + ], + "type": "text", + "content": "E. Dexheimer and A.J. Davison" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 187 + ], + "type": "text", + "content": "This research is funded by EPSRC Prosperity Partnerships (EP/S036636/1) and Dyson Technology Ltd. We would like to thank Riku Murai, Hidenobu Matsuki, Gwangbin Bae, and members of the Dyson Robotics Lab for insightful discussions and feedback." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 228, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "type": "text", + "content": "1. Alismail, H.: Direct pose estimation and refinement. Ph.D. thesis, Robotics Institute, Carnegie Mellon University. (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 480, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 272 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 272 + ], + "type": "text", + "content": "2. Baker, S., Matthews, I.: Lucas-Kanade 20 years on: A unifying framework. International Journal of Computer Vision (IJCV) (2004)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 274, + 480, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 274, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 480, + 316 + ], + "type": "text", + "content": "3. Bloesch, M., Czarnowski, J., Clark, R., Leutenegger, S., Davison, A.J.: CodeSLAM - learning a compact, optimisable representation for dense visual SLAM. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 317, + 480, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 317, + 480, + 349 + ], + "spans": [ + { + "bbox": [ + 138, + 317, + 480, + 349 + ], + "type": "text", + "content": "4. Bloesch, M., Laidlow, T., Clark, R., Leutenegger, S., Davison, A.: Learning meshes for dense visual SLAM. In: Proceedings of the International Conference on Computer Vision (ICCV) (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 350, + 480, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 350, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 138, + 350, + 480, + 381 + ], + "type": "text", + "content": "5. Burt, D.R., Rasmussen, C.E., van der Wilk, M.: Convergence of sparse variational inference in Gaussian processes regression. The Journal of Machine Learning Research (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 383, + 480, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 383, + 480, + 414 + ], + "spans": [ + { + "bbox": [ + 138, + 383, + 480, + 414 + ], + "type": "text", + "content": "6. Campos, C., Elvira, R., Rodríguez, J.J.G., M. Montiel, J.M., D. Tardós, J.: ORBSLAM3: An accurate open-source library for visual, visual-inertial, and multimap SLAM. IEEE Transactions on Robotics (T-RO) (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "type": "text", + "content": "7. Czarnowski, J., Laidlow, T., Clark, R., Davison, A.J.: DeepFactors: Real-time probabilistic dense monocular SLAM. IEEE Robotics and Automation Letters (RA-L) (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 447, + 480, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 480 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 480 + ], + "type": "text", + "content": "8. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 481, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 481, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 481, + 480, + 512 + ], + "type": "text", + "content": "9. Dexheimer, E., Davison, A.J.: Learning a depth covariance function. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 514, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 533 + ], + "type": "text", + "content": "10. Engel, J., Koltun, V., Cremers, D.: Direct sparse odometry. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI) (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 535, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 535, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 535, + 480, + 567 + ], + "type": "text", + "content": "1. Engel, J., Schöps, T., Cremers, D.: LSD-SLAM: Large-scale direct monocular SLAM. In: Proceedings of the European Conference on Computer Vision (ECCV) (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 567, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 600 + ], + "type": "text", + "content": "2. Forster, C., Zhang, Z., Gassner, M., Werlberger, M., Scaramuzza, D.: SVO: Semidirect visual odometry for monocular and multicamera systems. IEEE Transactions on Robotics (T-RO) (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "text", + "content": "3. Graham, B., Novotny, D.: RidgeSfM: Structure from motion via robust pairwise matching under depth uncertainty. In: Proceedings of the International Conference on 3D Vision (3DV) (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "4. Guestrin, C., Krause, A., Singh, A.P.: Near-optimal sensor placements in Gaussian processes. In: Proceedings of the International Conference on Machine Learning (ICML) (2005)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "COMO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 148 + ], + "type": "text", + "content": "15. Hornung, A., Wurm, K.M., Bennewitz, M., Stachniss, C., Burgard, W.: OctoMap: An efficient probabilistic 3D mapping framework based on octrees. Autonomous Robots (2013)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "text", + "content": "16. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3D Gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 480, + 214 + ], + "type": "text", + "content": "17. Klein, G., Murray, D.: Parallel tracking and mapping for small AR workspaces. In: Proceedings of the International Symposium on Mixed and Augmented Reality (ISMAR) (2007)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 480, + 246 + ], + "type": "text", + "content": "18. Koestler, L., Yang, N., Zeller, N., Cremers, D.: TANDEM: Tracking and dense mapping in real-time using deep multi-view stereo. In: Conference on Robot Learning (CoRL) (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 246, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 246, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 133, + 246, + 480, + 278 + ], + "type": "text", + "content": "19. Matsuki, H., Murai, R., Kelly, P.H.J., Davison, A.J.: Gaussian Splatting SLAM. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 278, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 311 + ], + "type": "text", + "content": "20. Matsuki, H., Scona, R., Czarnowski, J., Davison, A.J.: CodeMapping: Real-time dense mapping for sparse SLAM using compact scene representations. IEEE Robotics and Automation Letters (RA-L) (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 311, + 480, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 343 + ], + "type": "text", + "content": "21. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: Proceedings of the European Conference on Computer Vision (ECCV) (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 343, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 374 + ], + "type": "text", + "content": "22. Moss, H.B., Ober, S.W., Picheny, V.: Inducing point allocation for sparse Gaussian processes in high-throughput bayesian optimisation. In: International Conference on Artificial Intelligence and Statistics (AISTATS) (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "text", + "content": "23. Mur-Artal, R., Tardós, J.D.: Probabilistic semi-dense mapping from highly accurate feature-based monocular SLAM. In: Proceedings of Robotics: Science and Systems (RSS) (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 439 + ], + "type": "text", + "content": "24. Mur-Artal, R., Tardós, J.D.: ORB-SLAM2: An open-source SLAM system for monocular, stereo, and RGB-D cameras. IEEE Transactions on Robotics (T-RO) (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 439, + 480, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 480, + 472 + ], + "type": "text", + "content": "25. Newcombe, R.A., Davison, A.J.: Live dense reconstruction with a single moving camera. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2010)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 472, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 515 + ], + "type": "text", + "content": "26. Newcombe, R.A., Izadi, S., Hilliges, O., Molyneaux, D., Kim, D., Davison, A.J., Kohi, P., Shotton, J., Hodges, S., Fitzgibbon, A.: KinectFusion: Real-time dense surface mapping and tracking. In: Proceedings of the International Symposium on Mixed and Augmented Reality (ISMAR) (2011)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 515, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 480, + 547 + ], + "type": "text", + "content": "27. Newcombe, R.A., Lovegrove, S.J., Davison, A.J.: DTAM: Dense tracking and mapping in real-time. In: Proceedings of the International Conference on Computer Vision (ICCV) (2011)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 547, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 579 + ], + "type": "text", + "content": "28. Pizzoli, M., Forster, C., Scaramuzza, D.: REMODE: Probabilistic, monocular dense reconstruction in real time. In: Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "type": "text", + "content": "29. Platinsky, L., Davison, A.J., Leutenegger, S.: Monocular visual odometry: Sparse joint optimisation or dense alternation? In: Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "30. Rasmussen, C.E., Williams, C.K.I.: Gaussian Processes for Machine Learning. MIT Press (2005)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "31. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 296, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 296, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 296, + 100 + ], + "type": "text", + "content": "E. Dexheimer and A.J. Davison" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 358 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "32. Sola, J., Deray, J., Atchuthan, D.: A micro lie theory for state estimation in robotics. arXiv preprint arXiv:1812.01537 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 205 + ], + "type": "text", + "content": "33. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., Clarkson, A., Yan, M., Budge, B., Yan, Y., Pan, X., Yon, J., Zou, Y., Leon, K., Carter, N., Brales, J., Gillingham, T., Mueggler, E., Pesqueira, L., Savva, M., Batra, D., Strasdat, H.M., Nardi, R.D., Goesele, M., Lovegrove, S., Newcombe, R.: The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "text", + "content": "34. Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for the evaluation of RGB-D SLAM systems. In: Proceedings of the IEEE/RSJ Conference on Intelligent Robots and Systems (IROS) (2012)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "text", + "content": "35. Sucar, E., Liu, S., Ortiz, J., Davison, A.J.: iMAP: Implicit mapping and positioning in real-time. In: Proceedings of the International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "type": "text", + "content": "36. Tang, C., Tan, P.: BA-Net: Dense bundle adjustment networks. In: Proceedings of the International Conference on Learning Representations (ICLR) (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 482, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 315 + ], + "type": "text", + "content": "37. Teed, Z., Deng, J.: DROID-SLAM: Deep visual SLAM for monocular, stereo, and RGB-D cameras. In: Neural Information Processing Systems (NeurIPS) (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 315, + 482, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 336 + ], + "type": "text", + "content": "38. Teed, Z., Lipson, L., Deng, J.: Deep patch visual odometry. In: Neural Information Processing Systems (NeurIPS) (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 336, + 482, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 358 + ], + "type": "text", + "content": "39. Thrun, S.: Learning occupancy grids with forward models. In: Proceedings of the IEEE/RSJ Conference on Intelligent Robots and Systems (IROS) (2001)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "COMO" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/COMPOSE_ Comprehensive Portrait Shadow Editing/2be02136-d81b-443f-a244-263becec430b_content_list.json b/2024/COMPOSE_ Comprehensive Portrait Shadow Editing/2be02136-d81b-443f-a244-263becec430b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..026d7fb079f4fa5065cd02076fd971765be14609 --- /dev/null +++ b/2024/COMPOSE_ Comprehensive Portrait Shadow Editing/2be02136-d81b-443f-a244-263becec430b_content_list.json @@ -0,0 +1,1630 @@ +[ + { + "type": "text", + "text": "COMPOSE: Comprehensive Portrait Shadow Editing", + "text_level": 1, + "bbox": [ + 243, + 140, + 759, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Andrew Hou $^{1,2}$ , Zhixin Shu $^{2}$ , Xuaner Zhang $^{2}$ , He Zhang $^{2}$ , Yannick Hold-Geoffroy $^{2}$ , Jae Shin Yoon $^{2}$ , and Xiaoming Liu $^{1}$", + "bbox": [ + 261, + 210, + 743, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Michigan State University", + "bbox": [ + 405, + 253, + 596, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{houandr1, liuxm}@msu.edu", + "bbox": [ + 406, + 270, + 594, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Adobe Research", + "bbox": [ + 439, + 282, + 562, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zshu,cezhang,hezhan,holdgeof,jaeyoon}@adobe.com", + "bbox": [ + 315, + 297, + 687, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Existing portrait relighting methods struggle with precise control over facial shadows, particularly when faced with challenges such as handling hard shadows from directional light sources or adjusting shadows while remaining in harmony with existing lighting conditions. In many situations, completely altering input lighting is undesirable for portrait retouching applications: one may want to preserve some authenticity in the captured environment. Existing shadow editing methods typically restrict their application to just the facial region and often offer limited lighting control options, such as shadow softening or rotation. In this paper, we introduce COMPOSE: a novel shadow editing pipeline for human portraits, offering precise control over shadow attributes such as shape, intensity, and position, all while preserving the original environmental illumination of the portrait. This level of disentanglement and controllability is obtained thanks to a novel decomposition of the environment map representation into ambient light and an editable gaussian dominant light source. COMPOSE is a four-stage pipeline that consists of light estimation and editing, light diffusion, shadow synthesis, and finally shadow editing. We define facial shadows as the result of a dominant light source, encoded using our novel gaussian environment map representation. Utilizing an OLAT dataset, we have trained models to: (1) predict this light source representation from images, and (2) generate realistic shadows using this representation. We also demonstrate comprehensive and intuitive shadow editing with our pipeline. Through extensive quantitative and qualitative evaluations, we have demonstrated the robust capability of our system in shadow editing.", + "bbox": [ + 261, + 351, + 743, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Face Relighting $\\cdot$ Shadow Editing $\\cdot$ Lighting Decomposition", + "bbox": [ + 261, + 710, + 740, + 726 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 776, + 375, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single-image portrait relighting is a well-studied problem with steady interest from the computer vision and graphics communities. Portrait relighting has a", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e92076f726db6df6b454297cea98e599ed66f0e16c2f3ace97c12c87159ea937.jpg", + "image_caption": [ + "a) Source Image" + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 331, + 257 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2dac28d700bc188aca1e09dbd6b263d9df18caee00c321e0b2622af0c39878da.jpg", + "image_caption": [ + "b) Soften \nShadow" + ], + "image_footnote": [], + "bbox": [ + 334, + 143, + 444, + 257 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7f5066a52f3298bd87b9d1c057af746bd2568e06283a4407200818ec200df3ad.jpg", + "image_caption": [ + "c) Intensify Shadow", + "Fig. 1: Overview. COMPOSE is the first single-image portrait shadow editing method that achieves complete shadow editing control, i.e. adjusting shadow intensity, modifying light size, and changing shadow positions, all while preserving the source image's other lighting attributes (e.g. ambient light)." + ], + "image_footnote": [], + "bbox": [ + 446, + 143, + 555, + 257 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0c6b65d77d59b90229177bb8ce0bf63e46af09e54217990d8e9a4a8cc1636dee.jpg", + "image_caption": [ + "d) Modify Light Size" + ], + "image_footnote": [], + "bbox": [ + 558, + 143, + 669, + 257 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ffe6fcd76a686c49eca0658ac41ba5911471a525a22c44fc90d87c8a579075ce.jpg", + "image_caption": [ + "e) Rotate Shadow" + ], + "image_footnote": [], + "bbox": [ + 669, + 143, + 782, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "wide range of applications in computational photography [31], AR/VR [4,5,19], and downstream tasks in the face domain [11,25]. While many methods for single image portrait relighting have been proposed in recent years [17,18,29,31,33,41, 49,54], they often face challenges when dealing with facial shadows.", + "bbox": [ + 212, + 364, + 787, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One popular genre of face relighting methods [31, 41] uses low resolution HDR environment maps to represent lighting. This lighting representation is used in a conditional image translation framework to control the lighting effect of the output face image, thereby achieving facial relighting. The requirement for an environment map restricts users seeking to fine-tune lighting, such as minor modifications to shadow smoothness or direction. Additionally, a low-resolution environment map often lacks the precision needed for accurately depicting complex, high-frequency shadow effects. Conversely, several approaches use low-dimensional spherical harmonics (SH) for their lighting representation [37, 54]. These methods, based on simplified lighting and reflectance models, frequently produce unrealistic relit images. Moreover, they overlook occlusions in the image formation process, disallowing the generation and editing of cast shadows. SunStage [44] generates realistic shadows on faces but lacks generalizability and requires videos as input. Some recent works have focused on shadow editing tasks [11, 26, 53] using state-of-the-art image generators such as diffusion models. However, achieving precise control with these sophisticated generators remains a challenging task. Consequently, these methods often provide limited control in shadow synthesis, typically handling only shadow softening or completely removing the shadow, without precise editing of shadow intensity, shape, or position.", + "bbox": [ + 212, + 428, + 787, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the challenges in modeling and controlling facial shadows in portrait relighting, we introduce COMPOSE, a system that separates shadow modeling from ambient light in the lighting representation. COMPOSE enables flexible manipulation of shadow attributes, including position (lighting direction), smoothness, and intensity (See Fig. 1). The key innovation in COMPOSE is its shadow representation, which is compatible with environment map-based representations, making it easily integrable into existing relighting networks. Building on the concept of image-based lighting representations (e.g. HDR en", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "A. Hou et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "vironment maps), our approach divides lighting effects into two components: shadows caused by dominant light sources and ambient lighting. We simplify the shadow component by attributing it to a single dominant light source on an environment map. This is encoded using a 2D isotropic Gaussian, with variables representing light position, standard deviation (indicating light size or diffusion), and scale (light intensity). The remaining lighting effects are attributed to a diffused environment map, modeling the image's ambient lighting.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For a given face image, we first predict the lighting as a panorama environment map, from which we estimate the dominant light (shadow) parameters. The system then removes shadows to obtain a diffused, albedo-like image. Utilizing this diffuse image and the shadow representation, we train a network to synthesize shadows on the subject in the input image. This shadow synthesis process is adaptable, accepting various shadow-related parameters for controllable shadow synthesis, including position and shape. The synthesized shadow can be linearly blended with the diffuse lighting to create shadows of varying intensities.", + "bbox": [ + 212, + 252, + 787, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To train models for shadow diffusion and synthesis, we utilize an OLAT (one-light-at-a-time) dataset, captured through a light stage system. This dataset, combined with the environment map-based lighting representation, ensures our system's compatibility with previously proposed face relighting techniques. Our approach also uniquely offers the capability to accurately model shadows.", + "bbox": [ + 212, + 373, + 787, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions are thus as follows:", + "bbox": [ + 238, + 449, + 514, + 462 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$\\diamond$ We propose the first single image portrait shadow editing method with precise control over various shadow attributes, including shadow shape, intensity, and position. Our system provides users with the flexibility to control shading while also preserving the other lighting attributes of the input images.", + "$\\diamond$ We propose a novel lighting decomposition into ambient light and an ed-itable dominant gaussian light source, which enables disentangled shadow editing while preserving the remaining lighting attributes.", + "$\\diamond$ We achieve state-of-the-art relighting performance quantitatively and qualitatively, including on a light-stage-captured test set and in-the-wild images." + ], + "bbox": [ + 214, + 463, + 787, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 619, + 388, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Portrait Relighting", + "text_level": 1, + "bbox": [ + 215, + 650, + 421, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Many portrait relighting methods have been proposed in recent years that generally fall into one of two categories: focusing on reducing the data requirement and improving the relighting quality as much as possible with limited data [6, 10, 17, 18, 20, 24, 33, 34, 36-39, 43-45, 47, 49, 54] or using light stage data but modeling more complex lighting effects [4, 12, 28, 29, 31, 41, 42, 48, 50, 52]. Among single image portrait relighting methods, a substantial limitation is that they cannot preserve the original environmental lighting while editing shadows, and instead opt to relight by replacing the original lighting with a completely new environment [11], which is undesirable in many portrait editing applications. Existing methods also struggle to handle more directed lights and fail in the presence of substantial hard or intense shadows [37, 41, 54].", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "COMPOSE: Comprehensive Portrait Shadow Editing", + "bbox": [ + 374, + 113, + 732, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/da4be91a1b0c7392faf38906e446e5662637c181577a9027be877216a23874f3.jpg", + "image_caption": [ + "Fig. 2: Method Overview. COMPOSE is a 4-stage shadow editing method consisting of single-image lighting estimation, light diffusion, shadow synthesis, and image compositing. By first estimating the dominant light position using the environment map regressed from the input image, COMPOSE can control shadow shape and intensity by controlling light spread and light intensity respectively as well as shadow position by changing the location of the dominant light. By estimating a diffuse image $\\mathbf{I}_D$ as well as a shadowed image $\\mathbf{I}_S$ , COMPOSE can generate the final edited image $\\mathbf{I}_E$ through image compositing." + ], + "image_footnote": [], + "bbox": [ + 220, + 148, + 781, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Among portrait relighting methods that take multiple images or videos as input, the method SunStage [44] is able to maintain the lighting attributes of the input video's environment when performing relighting. However, the method must be trained using a video of a subject on a sunny day and involves an expensive optimization procedure. The model is also subject-specific and must be retrained or fine-tuned even for the same subject if they change their hairstyle, accessories, or clothing. COMPOSE has the advantage of being a single-image portrait relighting method, and is both generalizable and practical for in-the-wild photo editing applications.", + "bbox": [ + 212, + 430, + 789, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Shadow Editing", + "text_level": 1, + "bbox": [ + 215, + 588, + 393, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another line of work closely related to COMPOSE are methods that edit only facial shadows, including shadow removal [26,53] and shadow softening [11]. Most similar to our work is the method of [11], which is able to directly control the degree of shadow softening while preserving the original environmental lighting. However, unlike COMPOSE, they are restricted to only shadow softening and cannot handle other shadow editing operations like shadow intensifying, modifying shadow shape, and shadow rotation. Shadow removal methods [14,26,53] are often even more limited since they cannot control the degree of shadow softening and attempt to remove the shadow completely.", + "bbox": [ + 212, + 612, + 787, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Lighting Representations", + "text_level": 1, + "bbox": [ + 215, + 770, + 470, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing portrait relighting methods generally utilize one of three lighting representations: Spherical Harmonics (SH) [18,33,34,37,54], dominant lighting direction [17,29], and environment maps [31,41,49]. SH lighting tends to only model", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "A. Hou et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the lower frequencies of lighting effects, and thus is unsuitable for modeling non-Lambertian hard shadows and specularities and is better suited for diffuse lightings. Representing the light as a dominant lighting direction can often handle the non-Lambertian effects of directional and point light sources, but they often leave other lighting attributes such as light size out of the equation (e.g. Hou et al. [17]). Moreover, they often do not handle diffuse lightings well and are intended for directional lights [17, 29]. The environment map representation is more flexible and is able to model light intensity, light size, and light position and can handle both diffuse and directional lights. However, what is largely missing in this domain is a way to properly decompose the environment map that enables precise controllability over different lighting components. Existing methods only perform relighting by completely changing the scene illumination with a new environment map without disentangling the effects of ambient, diffuse, and directional light [31, 41, 49]. This can be undesirable for many computational photography applications where the user wishes to preserve the ambience of the scene and edit only the facial shadow alone [11]. Moreover, the default environment map representation is unsuitable for some shadow editing applications such as shrinking the light. While enlarging the light can be performed by applying a gaussian blur, there is no equivalent operation that can be performed to shrink the light on an environment map. Our lighting representation decomposes the environment map into the ambient light and an editable gaussian dominant light source, which is adaptable in light intensity, size, and position. COMPOSE is thus able to perform a full suite of shadow editing operations thanks to its editable gaussian component and can edit the shadow alone without disrupting other lighting attributes (e.g. ambient light).", + "bbox": [ + 212, + 146, + 787, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 542, + 380, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Problem Formulation", + "text_level": 1, + "bbox": [ + 215, + 571, + 439, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our framework, shadows are treated as a composable lighting effect, which can be independently predicted and altered based on a specific face image. This shadow representation is integrated into the overall lighting representation, and our framework is designed to (1) accurately predict the shadow from an image, and (2) apply a controllable shadow onto a shadow-free face image. We define the portrait shadow editing problem based on properties of shadows that users can manipulate, namely shadow position, shadow intensity, and shadow shape. These shadow properties are connected to the lighting attribute that is easily controllable. Specifically, given source image $\\mathbf{I}_N$ and desired edited output image $\\mathbf{I}_E$ with modified shadows, our problem can be written as:", + "bbox": [ + 212, + 593, + 787, + 744 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {E} = F _ {\\theta} \\left(\\mathbf {I} _ {N}, x, y, \\sigma , \\gamma\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 755, + 785, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $F_{\\theta}$ is our shadow editing method, $x$ and $y$ are the desired light position on an environment map, $\\sigma$ is the light size, and $\\gamma$ is the light intensity. By editing the parameters, $(x,y)$ , $\\sigma$ , and $\\gamma$ , we can control the shadow position, shape, and intensity respectively. We describe our solution COMPOSE in detail in Sec. 3.2.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "COMPOSE: Comprehensive Portrait Shadow Editing", + "bbox": [ + 374, + 113, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 COMPOSE Framework", + "text_level": 1, + "bbox": [ + 215, + 146, + 457, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "COMPOSE is a 4-stage shadow editing framework, consisting of single image lighting estimation, light diffusion, shadow synthesis, and finally image compositing. See Fig. 2 for an overview of our method.", + "bbox": [ + 212, + 169, + 784, + 214 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Lighting Estimation In the lighting estimation stage, we train a variational autoencoder (VAE) [23] to estimate an environment map (LDR) from a single portrait image as its lighting representation $\\mathbf{I}_N$ . From this environment map, we then estimate the dominant light position. Naturally, the dominant light, such as the sun, represents the brightest intensity on a image-based lighting representation. Therefore we can simply obtain the dominant light position by fitting a 2D isotropic Gaussian on the predicted environment map. The center of the Gaussian will indicate the position of the dominant light on the 2D environment map. This estimated light position corresponds to the shadow information on the input image, and therefore is useful for \"in-place\" shadow editing (e.g., shadow softening, shadow intensifying) in the later stage that does not alter the direction of the dominant light. We argue this is desirable in many portrait editing applications where the user only wants to edit existing shadows and does not want to alter the ambient lighting.", + "bbox": [ + 212, + 215, + 787, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Light Diffusion In the light diffusion stage, we remove all existing hard shadows and specular highlights from the input image and output a \"diffused\" image $\\mathbf{I}_D$ with very smooth shading. This represents the input face under only an ambient illumination condition. Our light diffusion network consists of a hierarchical transformer encoder that generates multi-level features, which are fed to a decoder with transposed convolutional layers. Our network takes as input the original input image, a body parsing mask, and a binary foreground mask.", + "bbox": [ + 212, + 430, + 787, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "**Shadow Synthesis** In the shadow synthesis stage, we aim to produce a shadowed image $\\mathbf{I}_S$ illuminated by an edited light source given the diffuse image $\\mathbf{I}_D$ estimated from the light diffusion stage and an edited environment map. Using the estimated dominant light position from the light estimation stage, we have full control over light attributes such as light size, intensity, and the dominant light position. We can produce a new environment map using a Gaussian to represent the dominant light, where the light size is modeled by the standard deviation of the Gaussian (a larger standard deviation represents a larger, more diffuse light) and the light intensity can be adjusted by multiplying the Gaussian with a scalar. This allows our method to control shadow shape, light intensity, and position. To regress the shadowed image $\\mathbf{I}_S$ , we propose a two-stage architecture: a U-Net followed by a conditional DDPM (See Fig. 3). For the first stage, we adopt the U-Net architecture of [41]. During training, instead of using a reshaped environment map-like representation for the dominant light, we design a feature map-like representation that's parameterized by the shadow attributes. Specifically, the lighting representation for shadows has 4 parameters. The first two parameters $x$ and $y$ encode the coordinates of the light center on the environment map, the third parameter $\\sigma$ encodes the light size, and the fourth parameter $\\gamma$ encodes the light intensity. All four parameters are normalized between 0 and 1 and repeated spatially as $32 \\times 32$ channels to be suitable", + "bbox": [ + 212, + 537, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "A. Hou et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/da5d8b375587fd13bd1aaa8eea2f8d3ada2444b94e726b4b7738256e4d2c6071.jpg", + "image_caption": [ + "Fig. 3: Shadow Synthesis Pipeline. Our shadow synthesis step consists of two stages: a U-Net followed by a conditional DDPM. The U-Net takes ambient image $\\mathbf{I}_D$ and environment map $\\mathbf{E}_T$ as input and outputs the coarse relit image $\\mathbf{I}_U$ . With $\\mathbf{I}_U$ and $\\mathbf{E}_T$ as spatial conditions, the conditional DDPM outputs the final shadowed image $\\mathbf{I}_S$ , refining shadow boundaries and improving image quality." + ], + "image_footnote": [], + "bbox": [ + 232, + 148, + 777, + 251 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "inputs for the U-Net. We find that this lighting representation leads to faster convergence, especially when the desired light mimics a point light and occupies a tiny portion of the environment map. We feed the diffuse image $\\mathbf{I}_D$ and our desired lighting feature map to the U-Net, and the output is a relit image $\\mathbf{I}_U$ . While the U-Net architecture has merits such as complete geometric consistency for shadow synthesis, one demerit is that the synthesized shadow boundaries are often not very sharp and the overall image quality could be improved. We thus employ a DDPM [7,15] as the second stage of our shadow synthesis model. The role of the DDPM here is to take the output image of the U-Net $\\mathbf{I}_U$ as a spatial condition along with the lighting parameters and perform image refinement to generate the final shadowed image $\\mathbf{I}_S$ . The training objective of our shadow synthesis DDPM thus becomes:", + "bbox": [ + 212, + 334, + 787, + 516 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {t, \\mathbf {x} _ {0}, \\boldsymbol {\\epsilon}} \\left[ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\mathbf {x} _ {t}, t, \\mathbf {I} _ {U}, x, y, \\sigma , \\gamma\\right) \\| ^ {2} \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 521, + 785, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our condition $\\mathbf{I}_U$ is spatially concatenated with $\\mathbf{x}_t$ and our lighting parameters are repeated spatially as channels of the same resolution and similarly spatially concatenated, where $\\mathbf{x}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\epsilon$ . We find that the quality of the edited images is noticeably enhanced by the DDPM compared to the U-Net alone and that adding the DDPM sharpens the shadow boundaries.", + "bbox": [ + 212, + 551, + 787, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Image Compositing With $\\mathbf{I}_D$ from our light diffusion stage and $\\mathbf{I}_S$ from our shadow synthesis stage, we can then perform image compositing between $\\mathbf{I}_D$ and $\\mathbf{I}_S$ to achieve our final edited image $\\mathbf{I}_E$ , by assigning weights $\\omega_{D}$ and $\\omega_{S}$ to $\\mathbf{I}_D$ and $\\mathbf{I}_S$ respectively. The final edited image $\\mathbf{I}_E$ is thus generated as:", + "bbox": [ + 212, + 628, + 787, + 690 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {E} = \\omega_ {D} \\mathbf {I} _ {D} + \\omega_ {S} \\mathbf {I} _ {S}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 699, + 785, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\omega_{D} + \\omega_{S} = 1$ . By adjusting $\\omega_{D}$ and $\\omega_{S}$ , we can further tune shadow intensity: a higher $\\omega_{D}$ softens shadows and a higher $\\omega_{S}$ results in darker shadows.", + "bbox": [ + 212, + 724, + 787, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Loss Functions", + "text_level": 1, + "bbox": [ + 214, + 773, + 385, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To train our single image lighting estimation network, we apply two loss functions commonly used in VAE training [23]: $\\mathcal{L}_{\\mathrm{recon}}$ and $\\mathcal{L}_{\\mathrm{KLD}}$ . $\\mathcal{L}_{\\mathrm{recon}}$ is a reconstruction loss between the predicted environment map $\\mathbf{E}_P$ and the groundtruth", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "COMPOSE: Comprehensive Portrait Shadow Editing", + "bbox": [ + 374, + 113, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "environment map $\\mathbf{E}_G$ defined as follows:", + "bbox": [ + 215, + 146, + 509, + 162 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c o n}} = \\frac {1}{3 H W} \\left(\\left\\| \\mathbf {E} _ {P} - \\mathbf {E} _ {G} \\right\\| _ {2} ^ {2}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 172, + 785, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $H$ and $W$ are the size of the environment maps. The Kullback-Leibler divergence (KLD) loss $\\mathcal{L}_{\\mathrm{KLD}}$ is:", + "bbox": [ + 214, + 209, + 784, + 239 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {K L D}} = - \\frac {1}{2 N} \\sum_ {i = 1} ^ {N} \\left(1 + \\log \\left(\\sigma_ {i} ^ {2}\\right) - \\mu_ {i} ^ {2} - \\sigma_ {i} ^ {2}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 252, + 785, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $N$ is the dimensionality of the latent vector, $\\mu$ is the batch mean, and $\\sigma^2$ is the batch variance. The final loss for our lighting estimation network is thus:", + "bbox": [ + 214, + 303, + 784, + 333 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {L E}} = \\lambda_ {1} \\mathcal {L} _ {\\text {r e c o n}} + \\lambda_ {2} \\mathcal {L} _ {\\mathrm {K L D}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 345, + 785, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\lambda_1 = 1$ and $\\lambda_{2} = 2.5\\times 10^{-4}$", + "bbox": [ + 215, + 372, + 470, + 387 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our light diffusion network is also trained with two losses $\\mathcal{L}_{\\mathrm{recon}}$ and $\\mathcal{L}_{\\mathrm{perceptual}}$ .", + "bbox": [ + 238, + 388, + 792, + 404 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c o n}} = \\frac {1}{3 H W} \\left(\\left\\| \\mathbf {I} _ {D} - \\mathbf {I} _ {G} \\right\\| _ {1}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 414, + 785, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $H$ and $W$ are the size of the training images, $\\mathbf{I}_D$ is the predicted diffuse image, and $\\mathbf{I}_G$ is the groundtruth diffuse image. $\\mathcal{L}_{\\mathrm{perceptual}}$ enforces visual similarity [51] and is computed as the distance between VGG [40] features computed for $\\mathbf{I}_D$ and $\\mathbf{I}_G$ . The final loss for the light diffusion network is thus:", + "bbox": [ + 214, + 446, + 785, + 506 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {L D}} = \\mathcal {L} _ {\\text {r e c o n}} + \\mathcal {L} _ {\\text {p e r c e p t u a l}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 520, + 785, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When training the U-Net of our shadow synthesis pipeline, we employ a single reconstruction loss $\\mathcal{L}_{\\mathrm{recon}}$ between the predicted relit image $\\mathbf{I}_S$ and the groundtruth $\\mathbf{I}_G$ , defined as follows:", + "bbox": [ + 214, + 546, + 785, + 590 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c o n}} = \\frac {1}{3 H W} \\left(\\| \\mathbf {I} _ {S} - \\mathbf {I} _ {G} \\| _ {2} ^ {2}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 602, + 785, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Data Collection and Generation", + "text_level": 1, + "bbox": [ + 215, + 648, + 526, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To train out networks, we use a light stage to capture one-light-at-a-time (OLAT) images of 107 subjects with diverse skin tones, body poses, expressions, and accessories using a light stage [9] with 160 lights. Each subject is captured under 10-20 sessions with varied body poses, expressions, and accessories. Each session is captured by 4 cameras with varying camera poses.", + "bbox": [ + 214, + 672, + 785, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To generate input images to train our lighting estimation VAE and our light diffusion network, we render our OLAT images using a diverse set of environment maps collected from the Polyhaven and Laval Outdoor HDR datasets [16]. During rendering, we apply random augmentations to the environment map including rotations and adjusting intensity to improve our model's generalization to different lighting. The groundtruth $\\mathbf{E}_G$ for our lighting estimation VAE is", + "bbox": [ + 214, + 750, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "A. Hou et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "simply the environment map used for rendering, and the groundtruth $\\mathbf{I}_G$ for the light diffusion network is generated by rendering OLAT images with heavily blurred, diffuse environment maps.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To generate images for our shadow synthesis stage, we render groundtruth relit images by using environment maps consisting of Gaussian lights with various positions, intensities, and sizes (See Sec. 3.2) along with our OLAT images. As the input to the shadow synthesis stage is meant to be the diffuse image $\\mathbf{I}_D$ predicted by the light diffusion stage, we generate diffuse images for training our shadow synthesis networks by rendering OLAT images with heavily blurred environment maps in the Polyhaven and Laval Outdoor HDR datasets.", + "bbox": [ + 212, + 191, + 787, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 316, + 455, + 330 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We implement all components of COMPOSE using PyTorch [32]. When training our lighting estimation VAE, we use a batch size of 128 and a learning rate of $10^{-4}$ on 4 24GB A10G GPUs. For our light diffusion network, we train with a batch size of 88 and a learning rate of $10^{-4}$ using 8 A100 GPUs. For our shadow synthesis U-Net, we train with a batch size of 100 and a learning rate of $10^{-4}$ using 4 24GB A10G GPUs. For all networks, we train using the Adam Optimizer [22]. The image resolution used to train the lighting estimation and shadow synthesis networks is $256 \\times 256$ and the resolution used to train the light diffusion network is $768 \\times 768$ . Please see the supplementary materials for architectural details of the lighting estimation network, the lighting diffusion network, and the shadow synthesis network.", + "bbox": [ + 212, + 337, + 787, + 503 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 523, + 375, + 540 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 214, + 551, + 334, + 565 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To train our lighting estimation and shadow synthesis networks, we generate our training sets using light stage images and 521 outdoor environment maps. Our training set consists of 82 subjects with significant variation in factors such as skin tones, pose, expressions, and accessories.", + "bbox": [ + 212, + 573, + 785, + 633 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For evaluation, we use two separate evaluation sets: the first is generated using our light stage images and has corresponding groundtruth, and the second consists of in-the-wild images that represent more realistic application settings and do not have groundtruth. Our light stage evaluation set consists of 48 images from a total of 16 subjects with diverse skin tones, body poses, expressions, and accessories. All evaluation subjects and lightings are held out and unseen during training. Our in-the-wild evaluation set also consists of unseen test subjects and applies unseen lighting conditions in all qualitative results.", + "bbox": [ + 212, + 633, + 787, + 753 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Shadow Editing Evaluation", + "text_level": 1, + "bbox": [ + 214, + 772, + 486, + 787 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluate our shadow editing performance (synthesis and removal) by using randomly sampled, unseen lighting positions as the target lights to generate relit light stage images of our 16 test subjects, which serve as groundtruth.", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "COMPOSE: Comprehensive Portrait Shadow Editing", + "bbox": [ + 374, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4ff79da6fd3733f1cbc65c51d4e3adacfff9cb73a3e0e397cb3c7c011be1bc5e.jpg", + "image_caption": [ + "Fig. 4: Shadow Synthesis. Our model synthesizes more plausible shadows than the baselines, and, unlike [17], is also able to properly remove shadows from the source image before relighting. Environment maps are shown to help visualize each test lighting." + ], + "image_footnote": [], + "bbox": [ + 263, + 143, + 743, + 503 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/4113e79ba13721c6b58c266f87dd7f4f62e3bfdc9dae20a429086c3d8d6b10b3.jpg", + "table_caption": [ + "Table 1: Shadow Editing Performance. We evaluate shadow editing using unseen lightings with varied intensity and size on 16 held out highly diverse light stage test subjects with groundtruth relit images. COMPOSE outperforms all baselines quantitatively in shadow editing performance across all metrics (mean ± standard deviation)." + ], + "table_footnote": [], + "table_body": "
MAEMSESSIMLPIPS
Hou et al. [17]0.1172 ± 0.07540.0438 ± 0.05030.7059 ± 0.10610.2371 ± 0.0871
Total Relighting [31]0.1008 ± 0.07430.0379 ± 0.04470.7754 ± 0.07380.2120 ± 0.0572
COMPOSE (Ours)0.0965 ± 0.06860.0349 ± 0.03680.7780 ± 0.08260.1973 ± 0.0678
", + "bbox": [ + 232, + 622, + 779, + 670 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Each input image is rendered with a randomly selected environment map out of 125 unseen outdoor environment maps and is first passed to our deshading network to generate diffuse image $\\mathbf{I}_D$ . $\\mathbf{I}_D$ and target environment map $\\mathbf{E}_T$ are then fed to our two-stage shadow synthesis pipeline to generate the newly shadowed image $\\mathbf{I}_S$ , which is evaluated against the groundtruth light stage image.", + "bbox": [ + 212, + 686, + 787, + 763 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tab. 1 compares our shadow editing performance against prior relighting methods. For both baselines [17, 31], the test results were provided by the authors. Our method achieves state-of-the-art results on all metrics (MAE, MSE, SSIM [46], and LPIPS [51]). As seen in Fig. 4, our model can synthesize appropriate shadows for various light positions. The method of Hou et al. [17] cannot", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "A. Hou et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/a4aac90790bd26aafb1f059b02033460c6695eb59b2d464b475866ddebbdd7c0.jpg", + "image_caption": [ + "Fig. 5: Shadow Editing. Our method achieves complete shadow editing control, including softening/intensifying shadows, changing light size to alter shape and intensity, and changing light position, all while preserving the source image's ambient light." + ], + "image_footnote": [], + "bbox": [ + 263, + 143, + 738, + 542 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "remove existing shadows in the source image and the shadow traces carry over as artifacts to the relit images. In addition, Hou et al. [17] only models the lighting direction and does not model the light size as a parameter, which leads to inaccurate shadow shape when the light size is varied. Total Relighting [31] is often unable to synthesize physically plausible shadows, and will sometimes overshadow (Fig. 4, rows 1, 3, and 4) or undershadow the image (Fig. 4, row 2). Moreover, their shadows are often blurry and not as sharp as our method. Compared to the baselines, COMPOSE can properly remove existing shadows from the source image and synthesize geometrically plausible and realistic shadows for a wide variety of lighting conditions.", + "bbox": [ + 212, + 604, + 787, + 755 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Portrait Shadow Editing Applications", + "text_level": 1, + "bbox": [ + 214, + 768, + 571, + 782 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We demonstrate the full performance of COMPOSE by performing all forms of shadow editing, including shadow softening, shadow intensifying, modifying light size (which changes both shadow shape and intensity), and shadow rotation", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "COMPOSE: Comprehensive Portrait Shadow Editing", + "bbox": [ + 375, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ac06b94f709f8530bd8dd5e9253be67ce2f592f0927ecc3dff4782577d449afd.jpg", + "image_caption": [ + "Fig.6: Degree of Shadow Editing. We demonstrate COMPOSE's controllable shadow editing by varying shadow intensity (top row), light diffusion for modifying shadow shape and intensity (middle row), and light position (bottom row)." + ], + "image_footnote": [], + "bbox": [ + 263, + 143, + 738, + 428 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "by changing the dominant light position. For each image, we first estimate the diffuse image $\\mathbf{I}_D$ and the shadowed image $\\mathbf{I}_S$ and perform image compositing between these two images to generate the final edited image $\\mathbf{I}_E$ . As seen in Fig. 5, COMPOSE properly edits the shadows of in-the-wild images. To soften or intensify the shadow without changing the shadow shape, we rely on our image compositing coefficients $\\omega_d$ and $\\omega_s$ , which control the contributions of $\\mathbf{I}_D$ and $\\mathbf{I}_S$ respectively. We increase the weight of $\\omega_d$ to soften the shadow and increase the weight of $\\omega_s$ to strengthen the shadow. We further show that we can modify light size by tuning the Gaussian spread $\\sigma$ , which changes both shadow shape and intensity. Enlarging the light makes the shadow more diffuse and less intense, whereas shrinking the light generates a harder, more intense shadow. For the subjects in rows 2, 3, and 4 we enlarged the light, resulting in a more diffuse shadow especially noticeable around the nose. For the subject in row 1 we instead shrunk the light and appropriately produce an edited image with more prominent and intense hard shadows. Finally, we also demonstrate shadow rotation by changing the dominant light position in $\\mathbf{I}_S$ 's environment map.", + "bbox": [ + 212, + 491, + 787, + 733 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To demonstrate COMPOSE can properly control shadow editing, we visualize various degrees of shadow softening/intensifying, light diffusion, and shadow rotation in Fig. 6. The first row shows varying shadow intensity, where the shadows become less intense from left to right using $\\omega_{d}$ and $\\omega_{s}$ . Our results demonstrate that these flexible weights allow for a wide range of shadow intensities. The second row shows increasing degrees of light diffusion from left to right, which corresponds to increasing the light size. The shadows in the generated images", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "A. Hou et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ffaadec496b1350ecde03d9439a9a2705103120165e8fd6b6f3d5d3789ba7e70.jpg", + "image_caption": [ + "Fig. 7: Controllable Shadow Softening. For each subject, we show the shadow softening results of [11] in the first row and COMPOSE in the second. COMPOSE leaves less shadow traces at level 0.0, where the goal is to remove the shadows completely." + ], + "image_footnote": [], + "bbox": [ + 261, + 143, + 738, + 445 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "correctly transition from harder shadows to more diffuse shadows as the light size increases. The third row shows shadow rotation, where the light is rotated horizontally along the environment map and around the subject. COMPOSE is able to produce physically plausible shadows as the light position changes.", + "bbox": [ + 212, + 503, + 784, + 565 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "It is important to note that this level of controllable shadow editing is not achievable by existing portrait shadow editing methods [11, 53] that are only able to handle shadow softening or complete shadow removal as well as existing portrait relighting methods that completely modify the existing environment and cannot preserve the original scene illumination [31], do not properly model all shadow attributes (e.g. intensity, shape, and position) [17, 33] or are not designed to handle lighting conditions with more substantial shadows [37, 41, 54].", + "bbox": [ + 212, + 565, + 787, + 672 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Shadow Softening Comparison", + "text_level": 1, + "bbox": [ + 214, + 680, + 514, + 696 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We compare on in-the-wild shadow softening with the state-of-the-art portrait shadow softening method Futschik et al. [11]. Using results provided by the authors, we visualize different degrees of shadow softening for [11] by varying the light diffusion parameter (0.75, 0.5, 0.2, and 0.0) and compare with equivalent degrees of shadow softening by COMPOSE, as shown in Fig. 7. Higher parameter values correspond to less softening whereas lower values correspond to more softening. At 0.0, the goal is to completely remove the source image's shadows. Across all degrees of light diffusion, both methods gradually soften the shadow, but COMPOSE leaves less shadow traces than [11] when completely removing", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "COMPOSE: Comprehensive Portrait Shadow Editing", + "bbox": [ + 374, + 113, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "the source image's shadows at level 0.0, especially for images with darker shadows. This is thanks to our hierarchical transformer design for our light diffusion network, which better handles removing the effects of shadows at all scales.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 213, + 359, + 229 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have proposed COMPOSE: the first single image portrait shadow editing method that can perform comprehensive variations of shadow editing including editing shadow intensity, light size, and shadow position while preserving the source image's lighting environment. This is largely enabled by our novel lighting decomposition into ambient light and an editable gaussian dominant light, which enables new applications like shrinking the light and intensifying shadows as well as disentangled shadow editing that preserves the scene ambience. We have demonstrated the novel photo editing applications enabled by our method over prior work qualitatively and that our method achieves state-of-the-art shadow editing performance quantitatively over prior relighting methods. We hope that our work can inspire and motivate more research in the exciting new research area of controllable and comprehensive portrait shadow editing.", + "bbox": [ + 212, + 244, + 787, + 426 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. Good directions for future work include handling lighting environments with multiple light sources and indoor environments, as our framework is intended mostly for outdoor settings, where the sun is the single dominant light. In addition, while $\\mathbf{I}_D$ can preserve the ambience, we still notice there are sometimes color shifts with the environmental lighting during shadow editing, largely due to the compositing step with shadowed image $\\mathbf{I}_S$ (which adopts the light stage's environmental lighting). This could be alleviated by modeling the diffuse lighting tint similar to [11] or by training with white balanced data to remove the color bias of the light stage. We also believe that one can continue to improve the generalizability of our model. Similar to other relighting methods that rely on light stage data [11, 31, 41], we train with a limited number of subjects. Future work may incorporate synthetic data to expand the training set or adopt latent diffusion models, which naturally have high generalizability [35]. Finally, one can further improve the identity preservation of our method. We find that high frequency details are sometimes lost at the initial U-Net stage of shadow synthesis, and ways to better preserve the identity include passing $\\mathbf{I}_D$ or the input image $\\mathbf{I}_N$ , which contain high frequency details, as additional conditions to the DDPM or to use stronger backbones than [41].", + "bbox": [ + 212, + 429, + 787, + 700 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Potential Societal Impacts. In image editing, there are always concerns of generating malicious deepfakes. However, we only alter illumination, not subject identity. Even so, there may be concerns that altering the illumination and especially adding dark shadows to faces could be used as a malicious attack to worsen the performance of downstream tasks such as face recognition. Users can detect these attacks using existing deepfake detection methods [1-3,8,13,21,27,30] or by training one using the edited images generated by COMPOSE. State-of-the-art shadow removal methods [26,53] are also an option to counter these potential attacks since they can remove any facial shadows that COMPOSE synthesizes.", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "A. Hou et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aghasanli, A., Kangin, D., Angelov, P.: Interpretable-through-prototypes deepfake detection for diffusion models. In: ICCVW (2023)", + "2. Asnani, V., Yin, X., Hassner, T., Liu, X.: MaLP: Manipulation localization using a proactive scheme. In: CVPR (2023)", + "3. Asnani, V., Yin, X., Hassner, T., Liu, X.: Reverse engineering of generative models: Inferring model hyperparameters from generated images. PAMI (2023)", + "4. Bi, S., Lombardi, S., Saito, S., Simon, T., Wei, S.E., McPhail, K., Ramamoorthi, R., Sheikh, Y., Saragih, J.: Deep relightable appearance models for animatable faces. TOG (2021)", + "5. Caselles, P., Ramon, E., Garcia, J., i Nieto, X.G., Moreno-Noguer, F., Triginer, G.: SIRA: Relightable avatars from a single image. In: WACV (2023)", + "6. Chandran, S., Hold-Geoffroy, Y., Sunkavalli, K., Shu, Z., Jayasuriya, S.: Temporally consistent relighting for portrait videos. In: WACV (2022)", + "7. Choi, J., Kim, S., Jeong, Y., Gwon, Y., Yoon, S.: ILVR: Conditioning method for denoising diffusion probabilistic models. In: ICCV (2021)", + "8. Corvi, R., Cozzolino, D., Zingarini, G., Poggi, G., Nagano, K., Verdoliva, L.: On the detection of synthetic images generated by diffusion models. In: ICASSP (2023)", + "9. Debevec, P., Hawkins, T., Tchou, C., Duiker, H.P., Sarokin, W., Sagar, M.: Acquiring the reflectance field of a human face. In: SIGGRAPH (2000)", + "10. Ding, Z., Zhang, C., Xia, Z., Jebe, L., Tu, Z., Zhang, X.: DiffusionRig: Learning personalized priors for facial appearance editing. In: CVPR (2023)", + "1. Futschik, D., Ritland, K., Vecore, J., Fanello, S., Orts-Escolano, S., Curless, B., Sykora, D., Pandey, R.: Controllable light diffusion for portraits. In: CVPR (2023)", + "12. Guo, K., Lincoln, P., Davidson, P., Busch, J., Yu, X., Whalen, M., Harvey, G., Orts-Escolano, S., Pandey, R., Dourgarian, J., DuVall, M., Tang, D., Tkach, A., Kowdle, A., Cooper, E., Dou, M., Fanello, S., Fyffe, G., Rhemann, C., Taylor, J., Debevecc, P., Izadi, S.: The Relighttables: Volumetric performance capture of humans with realistic relighting. In: SIGGRAPH Asia (2019)", + "13. Guo, X., Liu, X., Ren, Z., Grosz, S., Masi, I., Liu, X.: Hierarchical fine-grained image forgery detection and localization. In: CVPR (2023)", + "14. He, Y., Xing, Y., Zhang, T., Chen, Q.: Unsupervised portrait shadow removal via generative priors. In: MM (2021)", + "15. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)", + "16. Hold-Geoffroy, Y., Athawale, A., Lalonde, J.F.: Deep sky modeling for single image outdoor lighting estimation. In: CVPR (2019)", + "17. Hou, A., Sarkis, M., Bi, N., Tong, Y., Liu, X.: Face relighting with geometrically consistent shadows. In: CVPR (2022)", + "18. Hou, A., Zhang, Z., Sarkis, M., Bi, N., Tong, Y., Liu, X.: Towards high fidelity face relighting with realistic shadows. In: CVPR (2021)", + "19. Iqbal, U., Caliskan, A., Nagano, K., Khamis, S., Molchanov, P., Kautz, J.: RANA: Relightable articulated neural avatars. In: ICCV (2023)", + "20. Ji, C., Yu, T., Guo, K., Liu, J., Liu, Y.: Geometry-aware single-image full-body human relighting. In: ECCV (2022)", + "21. Kamat, S., Agarwal, S., Darrell, T., Rohrbach, A.: Revisiting generalizability in deepfake detection: Improving metrics and stabilizing transfer. In: ICCVW (2023)", + "22. Kingma, D., Adam, J.B.: A method for stochastic optimization. In: ICLR (2014)", + "23. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: ICLR (2014)" + ], + "bbox": [ + 218, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "COMPOSE: Comprehensive Portrait Shadow Editing", + "bbox": [ + 374, + 114, + 732, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "24. Lagunas, M., Sun, X., Yang, J., Villegas, R., Zhang, J., Shu, Z., Masia, B., Gutierrez, D.: Single-image full-body human relighting. In: EGSR (2021)", + "25. Le, H., Kakadiaris, I.: Illumination-invariant face recognition with deep relit face images. In: WACV (2019)", + "26. Liu, Y., Hou, A., Huang, X., Ren, L., Liu, X.: Blind removal of facial foreign shadows. In: BMVC (2022)", + "27. Lorenz, P., Durall, R.L., Keuper, J.: Detecting images generated by deep diffusion models using their local intrinsic dimensionality. In: ICCVW (2023)", + "28. Mei, Y., Zhang, H., Zhang, X., Zhang, J., Shu, Z., Wang, Y., Wei, Z., Yan, S., Jung, H., Patel, V.M.: LightPainter: Interactive portrait relighting with freehand scribble. In: CVPR (2023)", + "29. Nestmeyer, T., Lalonde, J.F., Matthews, I., Lehrmann, A.: Learning physics-guided face relighting under directional light. In: CVPR (2020)", + "30. Ojha, U., Li, Y., Lee, Y.J.: Towards universal fake image detectors that generalize across generative models. In: CVPR (2023)", + "31. Pandey, R., Orts-Escalano, S., LeGendre, C., Haene, C., Bouaziz, S., Rhemann, C., Debevec, P., Fanello, S.: Total relighting: Learning to relight portraits for background replacement. In: SIGGRAPH (2021)", + "32. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in PyTorch. In: NeurIPSW (2017)", + "33. Ponglertnapakorn, P., Tritrong, N., Suwajanakorn, S.: DiFaReli: Diffusion face relighting. In: ICCV (2023)", + "34. Ranjan, A., Yi, K.M., Chang, J.H.R., Tuzel, O.: FaceLit: Neural 3D relightable faces. In: CVPR (June 2023)", + "35. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)", + "36. Sengupta, R., Curless, B., Kemelmacher-Shlizerman, I., Seitz, S.: A light stage on every desk. In: ICCV (2021)", + "37. Sengupta, S., Kanazawa, A., Castillo, C.D., Jacobs, D.W.: SfSNet: Learning shape, reflectance and illuminance of faces in the wild. In: CVPR (2018)", + "38. Shih, Y., Paris, S., Barnes, C., Freeman, W.T., Durand, F.: Style transfer for headshot portraits. TOG (2014)", + "39. Shu, Z., Hadap, S., Shechtman, E., Sunkavalli, K., Paris, S., Samaras, D.: Portrait lighting transfer using a mass transport approach. TOG (2017)", + "40. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: ICLR (2015)", + "41. Sun, T., Barron, J.T., Tsai, Y.T., Xu, Z., Yu, X., Fyffe, G., Rhemann, C., Busch, J., Debevec, P., Ramamoorthi, R.: Single image portrait relighting. In: SIGGRAPH (2019)", + "42. Sun, T., Lin, K.E., Bi, S., Xu, Z., Ramamoorthi, R.: NeLF: Neural light-transport field for portrait view synthesis and relighting. In: EGSR (2021)", + "43. Tan, F., Fanello, S., Meka, A., OrtsEscolano, S., Tang, D., Pandey, R., Taylor, J., Tan, P., Zhang., Y.: VoLux-GAN: A generative model for 3D face synthesis with HDRI relighting. In: SIGGRAPH (2022)", + "44. Wang, Y., Holynski, A., Zhang, X., Zhang, X.: Sunstage: Portrait reconstruction and relighting using the sun as a light stage. In: CVPR (2023)", + "45. Wang, Z., Yu, X., Lu, M., Wang, Q., Qian, C., Xu, F.: Single image portrait relighting via explicit multiple reflectance channel modeling. TOG (2020)", + "46. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. TIP (2004)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "A. Hou et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "47. Weir, J., Zhao, J., Chalmers, A., Rhee, T.: Deep portrait delighting. In: ECCV (2022)", + "48. Yang, X., Taketomi, T.: BareSkinNet: Demakeup and de-lighting via 3D face reconstruction. CGF (2022)", + "49. Yeh, Y.Y., Nagano, K., Khamis, S., Kautz, J., Liu, M.Y., Wang, T.C.: Learning to relight portrait images via a virtual light stage and synthetic-to-real adaptation. TOG (2022)", + "50. Zhang, L., Zhang, Q., Wu, M., Yu, J., Xu, L.: Neural video portrait relighting in real-time via consistency modeling. In: ICCV (2021)", + "51. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)", + "52. Zhang, X., Fanello, S., Tsai, Y.T., Sun, T., Xue, T., Pandey, R., Orts-Escalano, S., Davidson, P., Rhemann, C., Debevec, P., Barron, J.T., Ramamoorthi, R., Freeman, W.T.: Neural light transport for relighting and view synthesis. TOG (2021)", + "53. Zhang, X., Barron, J.T., Tsai, Y.T., Pandey, R., Zhang, X., Ng, R., Jacobs, D.E.: Portrait Shadow Manipulation. TOG (2020)", + "54. Zhou, H., Hadap, S., Sunkavalli, K., Jacobs, D.W.: Deep single portrait image relighting. In: ICCV (2019)" + ], + "bbox": [ + 215, + 146, + 784, + 396 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "COMPOSE: Comprehensive Portrait Shadow Editing", + "bbox": [ + 374, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/COMPOSE_ Comprehensive Portrait Shadow Editing/2be02136-d81b-443f-a244-263becec430b_model.json b/2024/COMPOSE_ Comprehensive Portrait Shadow Editing/2be02136-d81b-443f-a244-263becec430b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..49df8d7af9e6dc49627eb8146132d8e855d72961 --- /dev/null +++ b/2024/COMPOSE_ Comprehensive Portrait Shadow Editing/2be02136-d81b-443f-a244-263becec430b_model.json @@ -0,0 +1,2291 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.245, + 0.141, + 0.761, + 0.187 + ], + "angle": 0, + "content": "COMPOSE: Comprehensive Portrait Shadow Editing" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.212, + 0.744, + 0.245 + ], + "angle": 0, + "content": "Andrew Hou\\(^{1,2}\\), Zhixin Shu\\(^{2}\\), Xuaner Zhang\\(^{2}\\), He Zhang\\(^{2}\\), Yannick Hold-Geoffroy\\(^{2}\\), Jae Shin Yoon\\(^{2}\\), and Xiaoming Liu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.406, + 0.255, + 0.597, + 0.269 + ], + "angle": 0, + "content": "1 Michigan State University" + }, + { + "type": "text", + "bbox": [ + 0.408, + 0.271, + 0.596, + 0.283 + ], + "angle": 0, + "content": "{houandr1, liuxm}@msu.edu" + }, + { + "type": "text", + "bbox": [ + 0.44, + 0.283, + 0.563, + 0.296 + ], + "angle": 0, + "content": "2 Adobe Research" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.298, + 0.688, + 0.312 + ], + "angle": 0, + "content": "{zshu,cezhang,hezhan,holdgeof,jaeyoon}@adobe.com" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.352, + 0.744, + 0.699 + ], + "angle": 0, + "content": "Abstract. Existing portrait relighting methods struggle with precise control over facial shadows, particularly when faced with challenges such as handling hard shadows from directional light sources or adjusting shadows while remaining in harmony with existing lighting conditions. In many situations, completely altering input lighting is undesirable for portrait retouching applications: one may want to preserve some authenticity in the captured environment. Existing shadow editing methods typically restrict their application to just the facial region and often offer limited lighting control options, such as shadow softening or rotation. In this paper, we introduce COMPOSE: a novel shadow editing pipeline for human portraits, offering precise control over shadow attributes such as shape, intensity, and position, all while preserving the original environmental illumination of the portrait. This level of disentanglement and controllability is obtained thanks to a novel decomposition of the environment map representation into ambient light and an editable gaussian dominant light source. COMPOSE is a four-stage pipeline that consists of light estimation and editing, light diffusion, shadow synthesis, and finally shadow editing. We define facial shadows as the result of a dominant light source, encoded using our novel gaussian environment map representation. Utilizing an OLAT dataset, we have trained models to: (1) predict this light source representation from images, and (2) generate realistic shadows using this representation. We also demonstrate comprehensive and intuitive shadow editing with our pipeline. Through extensive quantitative and qualitative evaluations, we have demonstrated the robust capability of our system in shadow editing." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.711, + 0.741, + 0.727 + ], + "angle": 0, + "content": "Keywords: Face Relighting \\(\\cdot\\) Shadow Editing \\(\\cdot\\) Lighting Decomposition" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.777, + 0.377, + 0.793 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Single-image portrait relighting is a well-studied problem with steady interest from the computer vision and graphics communities. Portrait relighting has a" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "A. Hou et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.144, + 0.332, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.26, + 0.332, + 0.274 + ], + "angle": 0, + "content": "a) Source Image" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.144, + 0.445, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.359, + 0.26, + 0.421, + 0.286 + ], + "angle": 0, + "content": "b) Soften \nShadow" + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.144, + 0.556, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.462, + 0.26, + 0.541, + 0.286 + ], + "angle": 0, + "content": "c) Intensify Shadow" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.144, + 0.67, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.26, + 0.669, + 0.286 + ], + "angle": 0, + "content": "d) Modify Light Size" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.144, + 0.783, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.694, + 0.26, + 0.76, + 0.286 + ], + "angle": 0, + "content": "e) Rotate Shadow" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.29, + 0.789, + 0.348 + ], + "angle": 0, + "content": "Fig. 1: Overview. COMPOSE is the first single-image portrait shadow editing method that achieves complete shadow editing control, i.e. adjusting shadow intensity, modifying light size, and changing shadow positions, all while preserving the source image's other lighting attributes (e.g. ambient light)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.365, + 0.789, + 0.426 + ], + "angle": 0, + "content": "wide range of applications in computational photography [31], AR/VR [4,5,19], and downstream tasks in the face domain [11,25]. While many methods for single image portrait relighting have been proposed in recent years [17,18,29,31,33,41, 49,54], they often face challenges when dealing with facial shadows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.789, + 0.717 + ], + "angle": 0, + "content": "One popular genre of face relighting methods [31, 41] uses low resolution HDR environment maps to represent lighting. This lighting representation is used in a conditional image translation framework to control the lighting effect of the output face image, thereby achieving facial relighting. The requirement for an environment map restricts users seeking to fine-tune lighting, such as minor modifications to shadow smoothness or direction. Additionally, a low-resolution environment map often lacks the precision needed for accurately depicting complex, high-frequency shadow effects. Conversely, several approaches use low-dimensional spherical harmonics (SH) for their lighting representation [37, 54]. These methods, based on simplified lighting and reflectance models, frequently produce unrealistic relit images. Moreover, they overlook occlusions in the image formation process, disallowing the generation and editing of cast shadows. SunStage [44] generates realistic shadows on faces but lacks generalizability and requires videos as input. Some recent works have focused on shadow editing tasks [11, 26, 53] using state-of-the-art image generators such as diffusion models. However, achieving precise control with these sophisticated generators remains a challenging task. Consequently, these methods often provide limited control in shadow synthesis, typically handling only shadow softening or completely removing the shadow, without precise editing of shadow intensity, shape, or position." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "To overcome the challenges in modeling and controlling facial shadows in portrait relighting, we introduce COMPOSE, a system that separates shadow modeling from ambient light in the lighting representation. COMPOSE enables flexible manipulation of shadow attributes, including position (lighting direction), smoothness, and intensity (See Fig. 1). The key innovation in COMPOSE is its shadow representation, which is compatible with environment map-based representations, making it easily integrable into existing relighting networks. Building on the concept of image-based lighting representations (e.g. HDR en" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.375, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "COMPOSE: Comprehensive Portrait Shadow Editing" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "vironment maps), our approach divides lighting effects into two components: shadows caused by dominant light sources and ambient lighting. We simplify the shadow component by attributing it to a single dominant light source on an environment map. This is encoded using a 2D isotropic Gaussian, with variables representing light position, standard deviation (indicating light size or diffusion), and scale (light intensity). The remaining lighting effects are attributed to a diffused environment map, modeling the image's ambient lighting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.373 + ], + "angle": 0, + "content": "For a given face image, we first predict the lighting as a panorama environment map, from which we estimate the dominant light (shadow) parameters. The system then removes shadows to obtain a diffused, albedo-like image. Utilizing this diffuse image and the shadow representation, we train a network to synthesize shadows on the subject in the input image. This shadow synthesis process is adaptable, accepting various shadow-related parameters for controllable shadow synthesis, including position and shape. The synthesized shadow can be linearly blended with the diffuse lighting to create shadows of varying intensities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.449 + ], + "angle": 0, + "content": "To train models for shadow diffusion and synthesis, we utilize an OLAT (one-light-at-a-time) dataset, captured through a light stage system. This dataset, combined with the environment map-based lighting representation, ensures our system's compatibility with previously proposed face relighting techniques. Our approach also uniquely offers the capability to accurately model shadows." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.45, + 0.516, + 0.463 + ], + "angle": 0, + "content": "Our contributions are thus as follows:" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.464, + 0.788, + 0.525 + ], + "angle": 0, + "content": "\\(\\diamond\\) We propose the first single image portrait shadow editing method with precise control over various shadow attributes, including shadow shape, intensity, and position. Our system provides users with the flexibility to control shading while also preserving the other lighting attributes of the input images." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.525, + 0.788, + 0.569 + ], + "angle": 0, + "content": "\\(\\diamond\\) We propose a novel lighting decomposition into ambient light and an ed-itable dominant gaussian light source, which enables disentangled shadow editing while preserving the remaining lighting attributes." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.57, + 0.788, + 0.601 + ], + "angle": 0, + "content": "\\(\\diamond\\) We achieve state-of-the-art relighting performance quantitatively and qualitatively, including on a light-stage-captured test set and in-the-wild images." + }, + { + "type": "list", + "bbox": [ + 0.215, + 0.464, + 0.788, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.621, + 0.39, + 0.637 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.651, + 0.422, + 0.667 + ], + "angle": 0, + "content": "2.1 Portrait Relighting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Many portrait relighting methods have been proposed in recent years that generally fall into one of two categories: focusing on reducing the data requirement and improving the relighting quality as much as possible with limited data [6, 10, 17, 18, 20, 24, 33, 34, 36-39, 43-45, 47, 49, 54] or using light stage data but modeling more complex lighting effects [4, 12, 28, 29, 31, 41, 42, 48, 50, 52]. Among single image portrait relighting methods, a substantial limitation is that they cannot preserve the original environmental lighting while editing shadows, and instead opt to relight by replacing the original lighting with a completely new environment [11], which is undesirable in many portrait editing applications. Existing methods also struggle to handle more directed lights and fail in the presence of substantial hard or intense shadows [37, 41, 54]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "A. Hou et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.15, + 0.782, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.307, + 0.789, + 0.419 + ], + "angle": 0, + "content": "Fig. 2: Method Overview. COMPOSE is a 4-stage shadow editing method consisting of single-image lighting estimation, light diffusion, shadow synthesis, and image compositing. By first estimating the dominant light position using the environment map regressed from the input image, COMPOSE can control shadow shape and intensity by controlling light spread and light intensity respectively as well as shadow position by changing the location of the dominant light. By estimating a diffuse image \\(\\mathbf{I}_D\\) as well as a shadowed image \\(\\mathbf{I}_S\\), COMPOSE can generate the final edited image \\(\\mathbf{I}_E\\) through image compositing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.79, + 0.568 + ], + "angle": 0, + "content": "Among portrait relighting methods that take multiple images or videos as input, the method SunStage [44] is able to maintain the lighting attributes of the input video's environment when performing relighting. However, the method must be trained using a video of a subject on a sunny day and involves an expensive optimization procedure. The model is also subject-specific and must be retrained or fine-tuned even for the same subject if they change their hairstyle, accessories, or clothing. COMPOSE has the advantage of being a single-image portrait relighting method, and is both generalizable and practical for in-the-wild photo editing applications." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.589, + 0.395, + 0.604 + ], + "angle": 0, + "content": "2.2 Shadow Editing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.789, + 0.75 + ], + "angle": 0, + "content": "Another line of work closely related to COMPOSE are methods that edit only facial shadows, including shadow removal [26,53] and shadow softening [11]. Most similar to our work is the method of [11], which is able to directly control the degree of shadow softening while preserving the original environmental lighting. However, unlike COMPOSE, they are restricted to only shadow softening and cannot handle other shadow editing operations like shadow intensifying, modifying shadow shape, and shadow rotation. Shadow removal methods [14,26,53] are often even more limited since they cannot control the degree of shadow softening and attempt to remove the shadow completely." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.771, + 0.471, + 0.787 + ], + "angle": 0, + "content": "2.3 Lighting Representations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Existing portrait relighting methods generally utilize one of three lighting representations: Spherical Harmonics (SH) [18,33,34,37,54], dominant lighting direction [17,29], and environment maps [31,41,49]. SH lighting tends to only model" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.375, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "COMPOSE: Comprehensive Portrait Shadow Editing" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.526 + ], + "angle": 0, + "content": "the lower frequencies of lighting effects, and thus is unsuitable for modeling non-Lambertian hard shadows and specularities and is better suited for diffuse lightings. Representing the light as a dominant lighting direction can often handle the non-Lambertian effects of directional and point light sources, but they often leave other lighting attributes such as light size out of the equation (e.g. Hou et al. [17]). Moreover, they often do not handle diffuse lightings well and are intended for directional lights [17, 29]. The environment map representation is more flexible and is able to model light intensity, light size, and light position and can handle both diffuse and directional lights. However, what is largely missing in this domain is a way to properly decompose the environment map that enables precise controllability over different lighting components. Existing methods only perform relighting by completely changing the scene illumination with a new environment map without disentangling the effects of ambient, diffuse, and directional light [31, 41, 49]. This can be undesirable for many computational photography applications where the user wishes to preserve the ambience of the scene and edit only the facial shadow alone [11]. Moreover, the default environment map representation is unsuitable for some shadow editing applications such as shrinking the light. While enlarging the light can be performed by applying a gaussian blur, there is no equivalent operation that can be performed to shrink the light on an environment map. Our lighting representation decomposes the environment map into the ambient light and an editable gaussian dominant light source, which is adaptable in light intensity, size, and position. COMPOSE is thus able to perform a full suite of shadow editing operations thanks to its editable gaussian component and can edit the shadow alone without disrupting other lighting attributes (e.g. ambient light)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.544, + 0.381, + 0.563 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.572, + 0.441, + 0.587 + ], + "angle": 0, + "content": "3.1 Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.788, + 0.746 + ], + "angle": 0, + "content": "In our framework, shadows are treated as a composable lighting effect, which can be independently predicted and altered based on a specific face image. This shadow representation is integrated into the overall lighting representation, and our framework is designed to (1) accurately predict the shadow from an image, and (2) apply a controllable shadow onto a shadow-free face image. We define the portrait shadow editing problem based on properties of shadows that users can manipulate, namely shadow position, shadow intensity, and shadow shape. These shadow properties are connected to the lighting attribute that is easily controllable. Specifically, given source image \\(\\mathbf{I}_N\\) and desired edited output image \\(\\mathbf{I}_E\\) with modified shadows, our problem can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.421, + 0.756, + 0.786, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {E} = F _ {\\theta} \\left(\\mathbf {I} _ {N}, x, y, \\sigma , \\gamma\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "where \\(F_{\\theta}\\) is our shadow editing method, \\(x\\) and \\(y\\) are the desired light position on an environment map, \\(\\sigma\\) is the light size, and \\(\\gamma\\) is the light intensity. By editing the parameters, \\((x,y)\\), \\(\\sigma\\), and \\(\\gamma\\), we can control the shadow position, shape, and intensity respectively. We describe our solution COMPOSE in detail in Sec. 3.2." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "A. Hou et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.458, + 0.162 + ], + "angle": 0, + "content": "3.2 COMPOSE Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.785, + 0.215 + ], + "angle": 0, + "content": "COMPOSE is a 4-stage shadow editing framework, consisting of single image lighting estimation, light diffusion, shadow synthesis, and finally image compositing. See Fig. 2 for an overview of our method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.217, + 0.788, + 0.429 + ], + "angle": 0, + "content": "Lighting Estimation In the lighting estimation stage, we train a variational autoencoder (VAE) [23] to estimate an environment map (LDR) from a single portrait image as its lighting representation \\(\\mathbf{I}_N\\). From this environment map, we then estimate the dominant light position. Naturally, the dominant light, such as the sun, represents the brightest intensity on a image-based lighting representation. Therefore we can simply obtain the dominant light position by fitting a 2D isotropic Gaussian on the predicted environment map. The center of the Gaussian will indicate the position of the dominant light on the 2D environment map. This estimated light position corresponds to the shadow information on the input image, and therefore is useful for \"in-place\" shadow editing (e.g., shadow softening, shadow intensifying) in the later stage that does not alter the direction of the dominant light. We argue this is desirable in many portrait editing applications where the user only wants to edit existing shadows and does not want to alter the ambient lighting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.537 + ], + "angle": 0, + "content": "Light Diffusion In the light diffusion stage, we remove all existing hard shadows and specular highlights from the input image and output a \"diffused\" image \\(\\mathbf{I}_D\\) with very smooth shading. This represents the input face under only an ambient illumination condition. Our light diffusion network consists of a hierarchical transformer encoder that generates multi-level features, which are fed to a decoder with transposed convolutional layers. Our network takes as input the original input image, a body parsing mask, and a binary foreground mask." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.841 + ], + "angle": 0, + "content": "**Shadow Synthesis** In the shadow synthesis stage, we aim to produce a shadowed image \\(\\mathbf{I}_S\\) illuminated by an edited light source given the diffuse image \\(\\mathbf{I}_D\\) estimated from the light diffusion stage and an edited environment map. Using the estimated dominant light position from the light estimation stage, we have full control over light attributes such as light size, intensity, and the dominant light position. We can produce a new environment map using a Gaussian to represent the dominant light, where the light size is modeled by the standard deviation of the Gaussian (a larger standard deviation represents a larger, more diffuse light) and the light intensity can be adjusted by multiplying the Gaussian with a scalar. This allows our method to control shadow shape, light intensity, and position. To regress the shadowed image \\(\\mathbf{I}_S\\), we propose a two-stage architecture: a U-Net followed by a conditional DDPM (See Fig. 3). For the first stage, we adopt the U-Net architecture of [41]. During training, instead of using a reshaped environment map-like representation for the dominant light, we design a feature map-like representation that's parameterized by the shadow attributes. Specifically, the lighting representation for shadows has 4 parameters. The first two parameters \\(x\\) and \\(y\\) encode the coordinates of the light center on the environment map, the third parameter \\(\\sigma\\) encodes the light size, and the fourth parameter \\(\\gamma\\) encodes the light intensity. All four parameters are normalized between 0 and 1 and repeated spatially as \\(32 \\times 32\\) channels to be suitable" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.375, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "COMPOSE: Comprehensive Portrait Shadow Editing" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.15, + 0.778, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.255, + 0.788, + 0.326 + ], + "angle": 0, + "content": "Fig. 3: Shadow Synthesis Pipeline. Our shadow synthesis step consists of two stages: a U-Net followed by a conditional DDPM. The U-Net takes ambient image \\(\\mathbf{I}_D\\) and environment map \\(\\mathbf{E}_T\\) as input and outputs the coarse relit image \\(\\mathbf{I}_U\\). With \\(\\mathbf{I}_U\\) and \\(\\mathbf{E}_T\\) as spatial conditions, the conditional DDPM outputs the final shadowed image \\(\\mathbf{I}_S\\), refining shadow boundaries and improving image quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.789, + 0.517 + ], + "angle": 0, + "content": "inputs for the U-Net. We find that this lighting representation leads to faster convergence, especially when the desired light mimics a point light and occupies a tiny portion of the environment map. We feed the diffuse image \\(\\mathbf{I}_D\\) and our desired lighting feature map to the U-Net, and the output is a relit image \\(\\mathbf{I}_U\\). While the U-Net architecture has merits such as complete geometric consistency for shadow synthesis, one demerit is that the synthesized shadow boundaries are often not very sharp and the overall image quality could be improved. We thus employ a DDPM [7,15] as the second stage of our shadow synthesis model. The role of the DDPM here is to take the output image of the U-Net \\(\\mathbf{I}_U\\) as a spatial condition along with the lighting parameters and perform image refinement to generate the final shadowed image \\(\\mathbf{I}_S\\). The training objective of our shadow synthesis DDPM thus becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.522, + 0.787, + 0.548 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {t, \\mathbf {x} _ {0}, \\boldsymbol {\\epsilon}} \\left[ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\mathbf {x} _ {t}, t, \\mathbf {I} _ {U}, x, y, \\sigma , \\gamma\\right) \\| ^ {2} \\right]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.628 + ], + "angle": 0, + "content": "Our condition \\(\\mathbf{I}_U\\) is spatially concatenated with \\(\\mathbf{x}_t\\) and our lighting parameters are repeated spatially as channels of the same resolution and similarly spatially concatenated, where \\(\\mathbf{x}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\epsilon\\). We find that the quality of the edited images is noticeably enhanced by the DDPM compared to the U-Net alone and that adding the DDPM sharpens the shadow boundaries." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.691 + ], + "angle": 0, + "content": "Image Compositing With \\(\\mathbf{I}_D\\) from our light diffusion stage and \\(\\mathbf{I}_S\\) from our shadow synthesis stage, we can then perform image compositing between \\(\\mathbf{I}_D\\) and \\(\\mathbf{I}_S\\) to achieve our final edited image \\(\\mathbf{I}_E\\), by assigning weights \\(\\omega_{D}\\) and \\(\\omega_{S}\\) to \\(\\mathbf{I}_D\\) and \\(\\mathbf{I}_S\\) respectively. The final edited image \\(\\mathbf{I}_E\\) is thus generated as:" + }, + { + "type": "equation", + "bbox": [ + 0.427, + 0.7, + 0.786, + 0.716 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {E} = \\omega_ {D} \\mathbf {I} _ {D} + \\omega_ {S} \\mathbf {I} _ {S}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.725, + 0.788, + 0.756 + ], + "angle": 0, + "content": "where \\(\\omega_{D} + \\omega_{S} = 1\\). By adjusting \\(\\omega_{D}\\) and \\(\\omega_{S}\\), we can further tune shadow intensity: a higher \\(\\omega_{D}\\) softens shadows and a higher \\(\\omega_{S}\\) results in darker shadows." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.774, + 0.386, + 0.789 + ], + "angle": 0, + "content": "3.3 Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "To train our single image lighting estimation network, we apply two loss functions commonly used in VAE training [23]: \\(\\mathcal{L}_{\\mathrm{recon}}\\) and \\(\\mathcal{L}_{\\mathrm{KLD}}\\). \\(\\mathcal{L}_{\\mathrm{recon}}\\) is a reconstruction loss between the predicted environment map \\(\\mathbf{E}_P\\) and the groundtruth" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "A. Hou et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.51, + 0.163 + ], + "angle": 0, + "content": "environment map \\(\\mathbf{E}_G\\) defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.173, + 0.786, + 0.202 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c o n}} = \\frac {1}{3 H W} \\left(\\left\\| \\mathbf {E} _ {P} - \\mathbf {E} _ {G} \\right\\| _ {2} ^ {2}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.21, + 0.785, + 0.24 + ], + "angle": 0, + "content": "where \\(H\\) and \\(W\\) are the size of the environment maps. The Kullback-Leibler divergence (KLD) loss \\(\\mathcal{L}_{\\mathrm{KLD}}\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.253, + 0.786, + 0.293 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {K L D}} = - \\frac {1}{2 N} \\sum_ {i = 1} ^ {N} \\left(1 + \\log \\left(\\sigma_ {i} ^ {2}\\right) - \\mu_ {i} ^ {2} - \\sigma_ {i} ^ {2}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.304, + 0.785, + 0.334 + ], + "angle": 0, + "content": "where \\(N\\) is the dimensionality of the latent vector, \\(\\mu\\) is the batch mean, and \\(\\sigma^2\\) is the batch variance. The final loss for our lighting estimation network is thus:" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.347, + 0.786, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {L E}} = \\lambda_ {1} \\mathcal {L} _ {\\text {r e c o n}} + \\lambda_ {2} \\mathcal {L} _ {\\mathrm {K L D}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.373, + 0.471, + 0.388 + ], + "angle": 0, + "content": "where \\(\\lambda_1 = 1\\) and \\(\\lambda_{2} = 2.5\\times 10^{-4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.389, + 0.794, + 0.405 + ], + "angle": 0, + "content": "Our light diffusion network is also trained with two losses \\(\\mathcal{L}_{\\mathrm{recon}}\\) and \\(\\mathcal{L}_{\\mathrm{perceptual}}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.415, + 0.786, + 0.443 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c o n}} = \\frac {1}{3 H W} \\left(\\left\\| \\mathbf {I} _ {D} - \\mathbf {I} _ {G} \\right\\| _ {1}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.447, + 0.786, + 0.507 + ], + "angle": 0, + "content": "where \\(H\\) and \\(W\\) are the size of the training images, \\(\\mathbf{I}_D\\) is the predicted diffuse image, and \\(\\mathbf{I}_G\\) is the groundtruth diffuse image. \\(\\mathcal{L}_{\\mathrm{perceptual}}\\) enforces visual similarity [51] and is computed as the distance between VGG [40] features computed for \\(\\mathbf{I}_D\\) and \\(\\mathbf{I}_G\\). The final loss for the light diffusion network is thus:" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.521, + 0.786, + 0.537 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {L D}} = \\mathcal {L} _ {\\text {r e c o n}} + \\mathcal {L} _ {\\text {p e r c e p t u a l}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.547, + 0.786, + 0.592 + ], + "angle": 0, + "content": "When training the U-Net of our shadow synthesis pipeline, we employ a single reconstruction loss \\(\\mathcal{L}_{\\mathrm{recon}}\\) between the predicted relit image \\(\\mathbf{I}_S\\) and the groundtruth \\(\\mathbf{I}_G\\), defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.4, + 0.603, + 0.786, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c o n}} = \\frac {1}{3 H W} \\left(\\| \\mathbf {I} _ {S} - \\mathbf {I} _ {G} \\| _ {2} ^ {2}\\right). \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.65, + 0.527, + 0.663 + ], + "angle": 0, + "content": "3.4 Data Collection and Generation" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.674, + 0.786, + 0.749 + ], + "angle": 0, + "content": "To train out networks, we use a light stage to capture one-light-at-a-time (OLAT) images of 107 subjects with diverse skin tones, body poses, expressions, and accessories using a light stage [9] with 160 lights. Each subject is captured under 10-20 sessions with varied body poses, expressions, and accessories. Each session is captured by 4 cameras with varying camera poses." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.75, + 0.787, + 0.841 + ], + "angle": 0, + "content": "To generate input images to train our lighting estimation VAE and our light diffusion network, we render our OLAT images using a diverse set of environment maps collected from the Polyhaven and Laval Outdoor HDR datasets [16]. During rendering, we apply random augmentations to the environment map including rotations and adjusting intensity to improve our model's generalization to different lighting. The groundtruth \\(\\mathbf{E}_G\\) for our lighting estimation VAE is" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.375, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "COMPOSE: Comprehensive Portrait Shadow Editing" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "simply the environment map used for rendering, and the groundtruth \\(\\mathbf{I}_G\\) for the light diffusion network is generated by rendering OLAT images with heavily blurred, diffuse environment maps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.788, + 0.298 + ], + "angle": 0, + "content": "To generate images for our shadow synthesis stage, we render groundtruth relit images by using environment maps consisting of Gaussian lights with various positions, intensities, and sizes (See Sec. 3.2) along with our OLAT images. As the input to the shadow synthesis stage is meant to be the diffuse image \\(\\mathbf{I}_D\\) predicted by the light diffusion stage, we generate diffuse images for training our shadow synthesis networks by rendering OLAT images with heavily blurred environment maps in the Polyhaven and Laval Outdoor HDR datasets." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.317, + 0.457, + 0.332 + ], + "angle": 0, + "content": "3.5 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.338, + 0.788, + 0.504 + ], + "angle": 0, + "content": "We implement all components of COMPOSE using PyTorch [32]. When training our lighting estimation VAE, we use a batch size of 128 and a learning rate of \\(10^{-4}\\) on 4 24GB A10G GPUs. For our light diffusion network, we train with a batch size of 88 and a learning rate of \\(10^{-4}\\) using 8 A100 GPUs. For our shadow synthesis U-Net, we train with a batch size of 100 and a learning rate of \\(10^{-4}\\) using 4 24GB A10G GPUs. For all networks, we train using the Adam Optimizer [22]. The image resolution used to train the lighting estimation and shadow synthesis networks is \\(256 \\times 256\\) and the resolution used to train the light diffusion network is \\(768 \\times 768\\). Please see the supplementary materials for architectural details of the lighting estimation network, the lighting diffusion network, and the shadow synthesis network." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.524, + 0.377, + 0.541 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.552, + 0.335, + 0.566 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.574, + 0.787, + 0.634 + ], + "angle": 0, + "content": "To train our lighting estimation and shadow synthesis networks, we generate our training sets using light stage images and 521 outdoor environment maps. Our training set consists of 82 subjects with significant variation in factors such as skin tones, pose, expressions, and accessories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.635, + 0.788, + 0.755 + ], + "angle": 0, + "content": "For evaluation, we use two separate evaluation sets: the first is generated using our light stage images and has corresponding groundtruth, and the second consists of in-the-wild images that represent more realistic application settings and do not have groundtruth. Our light stage evaluation set consists of 48 images from a total of 16 subjects with diverse skin tones, body poses, expressions, and accessories. All evaluation subjects and lightings are held out and unseen during training. Our in-the-wild evaluation set also consists of unseen test subjects and applies unseen lighting conditions in all qualitative results." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.773, + 0.487, + 0.789 + ], + "angle": 0, + "content": "4.2 Shadow Editing Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.841 + ], + "angle": 0, + "content": "We evaluate our shadow editing performance (synthesis and removal) by using randomly sampled, unseen lighting positions as the target lights to generate relit light stage images of our 16 test subjects, which serve as groundtruth." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "A. Hou et al." + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.144, + 0.744, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.506, + 0.788, + 0.549 + ], + "angle": 0, + "content": "Fig. 4: Shadow Synthesis. Our model synthesizes more plausible shadows than the baselines, and, unlike [17], is also able to properly remove shadows from the source image before relighting. Environment maps are shown to help visualize each test lighting." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.561, + 0.788, + 0.618 + ], + "angle": 0, + "content": "Table 1: Shadow Editing Performance. We evaluate shadow editing using unseen lightings with varied intensity and size on 16 held out highly diverse light stage test subjects with groundtruth relit images. COMPOSE outperforms all baselines quantitatively in shadow editing performance across all metrics (mean ± standard deviation)." + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.623, + 0.78, + 0.671 + ], + "angle": 0, + "content": "
MAEMSESSIMLPIPS
Hou et al. [17]0.1172 ± 0.07540.0438 ± 0.05030.7059 ± 0.10610.2371 ± 0.0871
Total Relighting [31]0.1008 ± 0.07430.0379 ± 0.04470.7754 ± 0.07380.2120 ± 0.0572
COMPOSE (Ours)0.0965 ± 0.06860.0349 ± 0.03680.7780 ± 0.08260.1973 ± 0.0678
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.788, + 0.764 + ], + "angle": 0, + "content": "Each input image is rendered with a randomly selected environment map out of 125 unseen outdoor environment maps and is first passed to our deshading network to generate diffuse image \\(\\mathbf{I}_D\\). \\(\\mathbf{I}_D\\) and target environment map \\(\\mathbf{E}_T\\) are then fed to our two-stage shadow synthesis pipeline to generate the newly shadowed image \\(\\mathbf{I}_S\\), which is evaluated against the groundtruth light stage image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Tab. 1 compares our shadow editing performance against prior relighting methods. For both baselines [17, 31], the test results were provided by the authors. Our method achieves state-of-the-art results on all metrics (MAE, MSE, SSIM [46], and LPIPS [51]). As seen in Fig. 4, our model can synthesize appropriate shadows for various light positions. The method of Hou et al. [17] cannot" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.376, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "COMPOSE: Comprehensive Portrait Shadow Editing" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.145, + 0.74, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.548, + 0.785, + 0.59 + ], + "angle": 0, + "content": "Fig. 5: Shadow Editing. Our method achieves complete shadow editing control, including softening/intensifying shadows, changing light size to alter shape and intensity, and changing light position, all while preserving the source image's ambient light." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.605, + 0.788, + 0.756 + ], + "angle": 0, + "content": "remove existing shadows in the source image and the shadow traces carry over as artifacts to the relit images. In addition, Hou et al. [17] only models the lighting direction and does not model the light size as a parameter, which leads to inaccurate shadow shape when the light size is varied. Total Relighting [31] is often unable to synthesize physically plausible shadows, and will sometimes overshadow (Fig. 4, rows 1, 3, and 4) or undershadow the image (Fig. 4, row 2). Moreover, their shadows are often blurry and not as sharp as our method. Compared to the baselines, COMPOSE can properly remove existing shadows from the source image and synthesize geometrically plausible and realistic shadows for a wide variety of lighting conditions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.769, + 0.573, + 0.784 + ], + "angle": 0, + "content": "4.3 Portrait Shadow Editing Applications" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We demonstrate the full performance of COMPOSE by performing all forms of shadow editing, including shadow softening, shadow intensifying, modifying light size (which changes both shadow shape and intensity), and shadow rotation" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "A. Hou et al." + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.144, + 0.74, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.476 + ], + "angle": 0, + "content": "Fig.6: Degree of Shadow Editing. We demonstrate COMPOSE's controllable shadow editing by varying shadow intensity (top row), light diffusion for modifying shadow shape and intensity (middle row), and light position (bottom row)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.734 + ], + "angle": 0, + "content": "by changing the dominant light position. For each image, we first estimate the diffuse image \\(\\mathbf{I}_D\\) and the shadowed image \\(\\mathbf{I}_S\\) and perform image compositing between these two images to generate the final edited image \\(\\mathbf{I}_E\\). As seen in Fig. 5, COMPOSE properly edits the shadows of in-the-wild images. To soften or intensify the shadow without changing the shadow shape, we rely on our image compositing coefficients \\(\\omega_d\\) and \\(\\omega_s\\), which control the contributions of \\(\\mathbf{I}_D\\) and \\(\\mathbf{I}_S\\) respectively. We increase the weight of \\(\\omega_d\\) to soften the shadow and increase the weight of \\(\\omega_s\\) to strengthen the shadow. We further show that we can modify light size by tuning the Gaussian spread \\(\\sigma\\), which changes both shadow shape and intensity. Enlarging the light makes the shadow more diffuse and less intense, whereas shrinking the light generates a harder, more intense shadow. For the subjects in rows 2, 3, and 4 we enlarged the light, resulting in a more diffuse shadow especially noticeable around the nose. For the subject in row 1 we instead shrunk the light and appropriately produce an edited image with more prominent and intense hard shadows. Finally, we also demonstrate shadow rotation by changing the dominant light position in \\(\\mathbf{I}_S\\)'s environment map." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To demonstrate COMPOSE can properly control shadow editing, we visualize various degrees of shadow softening/intensifying, light diffusion, and shadow rotation in Fig. 6. The first row shows varying shadow intensity, where the shadows become less intense from left to right using \\(\\omega_{d}\\) and \\(\\omega_{s}\\). Our results demonstrate that these flexible weights allow for a wide range of shadow intensities. The second row shows increasing degrees of light diffusion from left to right, which corresponds to increasing the light size. The shadows in the generated images" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.375, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "COMPOSE: Comprehensive Portrait Shadow Editing" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.144, + 0.74, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.489 + ], + "angle": 0, + "content": "Fig. 7: Controllable Shadow Softening. For each subject, we show the shadow softening results of [11] in the first row and COMPOSE in the second. COMPOSE leaves less shadow traces at level 0.0, where the goal is to remove the shadows completely." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.785, + 0.566 + ], + "angle": 0, + "content": "correctly transition from harder shadows to more diffuse shadows as the light size increases. The third row shows shadow rotation, where the light is rotated horizontally along the environment map and around the subject. COMPOSE is able to produce physically plausible shadows as the light position changes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.673 + ], + "angle": 0, + "content": "It is important to note that this level of controllable shadow editing is not achievable by existing portrait shadow editing methods [11, 53] that are only able to handle shadow softening or complete shadow removal as well as existing portrait relighting methods that completely modify the existing environment and cannot preserve the original scene illumination [31], do not properly model all shadow attributes (e.g. intensity, shape, and position) [17, 33] or are not designed to handle lighting conditions with more substantial shadows [37, 41, 54]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.681, + 0.515, + 0.697 + ], + "angle": 0, + "content": "4.4 Shadow Softening Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.842 + ], + "angle": 0, + "content": "We compare on in-the-wild shadow softening with the state-of-the-art portrait shadow softening method Futschik et al. [11]. Using results provided by the authors, we visualize different degrees of shadow softening for [11] by varying the light diffusion parameter (0.75, 0.5, 0.2, and 0.0) and compare with equivalent degrees of shadow softening by COMPOSE, as shown in Fig. 7. Higher parameter values correspond to less softening whereas lower values correspond to more softening. At 0.0, the goal is to completely remove the source image's shadows. Across all degrees of light diffusion, both methods gradually soften the shadow, but COMPOSE leaves less shadow traces than [11] when completely removing" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "A. Hou et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.193 + ], + "angle": 0, + "content": "the source image's shadows at level 0.0, especially for images with darker shadows. This is thanks to our hierarchical transformer design for our light diffusion network, which better handles removing the effects of shadows at all scales." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.214, + 0.36, + 0.23 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.245, + 0.788, + 0.427 + ], + "angle": 0, + "content": "We have proposed COMPOSE: the first single image portrait shadow editing method that can perform comprehensive variations of shadow editing including editing shadow intensity, light size, and shadow position while preserving the source image's lighting environment. This is largely enabled by our novel lighting decomposition into ambient light and an editable gaussian dominant light, which enables new applications like shrinking the light and intensifying shadows as well as disentangled shadow editing that preserves the scene ambience. We have demonstrated the novel photo editing applications enabled by our method over prior work qualitatively and that our method achieves state-of-the-art shadow editing performance quantitatively over prior relighting methods. We hope that our work can inspire and motivate more research in the exciting new research area of controllable and comprehensive portrait shadow editing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.702 + ], + "angle": 0, + "content": "Limitations. Good directions for future work include handling lighting environments with multiple light sources and indoor environments, as our framework is intended mostly for outdoor settings, where the sun is the single dominant light. In addition, while \\(\\mathbf{I}_D\\) can preserve the ambience, we still notice there are sometimes color shifts with the environmental lighting during shadow editing, largely due to the compositing step with shadowed image \\(\\mathbf{I}_S\\) (which adopts the light stage's environmental lighting). This could be alleviated by modeling the diffuse lighting tint similar to [11] or by training with white balanced data to remove the color bias of the light stage. We also believe that one can continue to improve the generalizability of our model. Similar to other relighting methods that rely on light stage data [11, 31, 41], we train with a limited number of subjects. Future work may incorporate synthetic data to expand the training set or adopt latent diffusion models, which naturally have high generalizability [35]. Finally, one can further improve the identity preservation of our method. We find that high frequency details are sometimes lost at the initial U-Net stage of shadow synthesis, and ways to better preserve the identity include passing \\(\\mathbf{I}_D\\) or the input image \\(\\mathbf{I}_N\\), which contain high frequency details, as additional conditions to the DDPM or to use stronger backbones than [41]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Potential Societal Impacts. In image editing, there are always concerns of generating malicious deepfakes. However, we only alter illumination, not subject identity. Even so, there may be concerns that altering the illumination and especially adding dark shadows to faces could be used as a malicious attack to worsen the performance of downstream tasks such as face recognition. Users can detect these attacks using existing deepfake detection methods [1-3,8,13,21,27,30] or by training one using the edited images generated by COMPOSE. State-of-the-art shadow removal methods [26,53] are also an option to counter these potential attacks since they can remove any facial shadows that COMPOSE synthesizes." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.375, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "COMPOSE: Comprehensive Portrait Shadow Editing" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. Aghasanli, A., Kangin, D., Angelov, P.: Interpretable-through-prototypes deepfake detection for diffusion models. In: ICCVW (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.232 + ], + "angle": 0, + "content": "2. Asnani, V., Yin, X., Hassner, T., Liu, X.: MaLP: Manipulation localization using a proactive scheme. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.233, + 0.785, + 0.26 + ], + "angle": 0, + "content": "3. Asnani, V., Yin, X., Hassner, T., Liu, X.: Reverse engineering of generative models: Inferring model hyperparameters from generated images. PAMI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.3 + ], + "angle": 0, + "content": "4. Bi, S., Lombardi, S., Saito, S., Simon, T., Wei, S.E., McPhail, K., Ramamoorthi, R., Sheikh, Y., Saragih, J.: Deep relightable appearance models for animatable faces. TOG (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.301, + 0.785, + 0.328 + ], + "angle": 0, + "content": "5. Caselles, P., Ramon, E., Garcia, J., i Nieto, X.G., Moreno-Noguer, F., Triginer, G.: SIRA: Relightable avatars from a single image. In: WACV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.785, + 0.357 + ], + "angle": 0, + "content": "6. Chandran, S., Hold-Geoffroy, Y., Sunkavalli, K., Shu, Z., Jayasuriya, S.: Temporally consistent relighting for portrait videos. In: WACV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.358, + 0.785, + 0.384 + ], + "angle": 0, + "content": "7. Choi, J., Kim, S., Jeong, Y., Gwon, Y., Yoon, S.: ILVR: Conditioning method for denoising diffusion probabilistic models. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.785, + 0.412 + ], + "angle": 0, + "content": "8. Corvi, R., Cozzolino, D., Zingarini, G., Poggi, G., Nagano, K., Verdoliva, L.: On the detection of synthetic images generated by diffusion models. In: ICASSP (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.413, + 0.785, + 0.439 + ], + "angle": 0, + "content": "9. Debevec, P., Hawkins, T., Tchou, C., Duiker, H.P., Sarokin, W., Sagar, M.: Acquiring the reflectance field of a human face. In: SIGGRAPH (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.785, + 0.467 + ], + "angle": 0, + "content": "10. Ding, Z., Zhang, C., Xia, Z., Jebe, L., Tu, Z., Zhang, X.: DiffusionRig: Learning personalized priors for facial appearance editing. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.785, + 0.495 + ], + "angle": 0, + "content": "1. Futschik, D., Ritland, K., Vecore, J., Fanello, S., Orts-Escolano, S., Curless, B., Sykora, D., Pandey, R.: Controllable light diffusion for portraits. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.496, + 0.785, + 0.564 + ], + "angle": 0, + "content": "12. Guo, K., Lincoln, P., Davidson, P., Busch, J., Yu, X., Whalen, M., Harvey, G., Orts-Escolano, S., Pandey, R., Dourgarian, J., DuVall, M., Tang, D., Tkach, A., Kowdle, A., Cooper, E., Dou, M., Fanello, S., Fyffe, G., Rhemann, C., Taylor, J., Debevecc, P., Izadi, S.: The Relighttables: Volumetric performance capture of humans with realistic relighting. In: SIGGRAPH Asia (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.565, + 0.785, + 0.592 + ], + "angle": 0, + "content": "13. Guo, X., Liu, X., Ren, Z., Grosz, S., Masi, I., Liu, X.: Hierarchical fine-grained image forgery detection and localization. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.593, + 0.785, + 0.619 + ], + "angle": 0, + "content": "14. He, Y., Xing, Y., Zhang, T., Chen, Q.: Unsupervised portrait shadow removal via generative priors. In: MM (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.62, + 0.785, + 0.646 + ], + "angle": 0, + "content": "15. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "16. Hold-Geoffroy, Y., Athawale, A., Lalonde, J.F.: Deep sky modeling for single image outdoor lighting estimation. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "17. Hou, A., Sarkis, M., Bi, N., Tong, Y., Liu, X.: Face relighting with geometrically consistent shadows. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.703, + 0.785, + 0.73 + ], + "angle": 0, + "content": "18. Hou, A., Zhang, Z., Sarkis, M., Bi, N., Tong, Y., Liu, X.: Towards high fidelity face relighting with realistic shadows. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.731, + 0.785, + 0.757 + ], + "angle": 0, + "content": "19. Iqbal, U., Caliskan, A., Nagano, K., Khamis, S., Molchanov, P., Kautz, J.: RANA: Relightable articulated neural avatars. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "20. Ji, C., Yu, T., Guo, K., Liu, J., Liu, Y.: Geometry-aware single-image full-body human relighting. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "21. Kamat, S., Agarwal, S., Darrell, T., Rohrbach, A.: Revisiting generalizability in deepfake detection: Improving metrics and stabilizing transfer. In: ICCVW (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.814, + 0.785, + 0.826 + ], + "angle": 0, + "content": "22. Kingma, D., Adam, J.B.: A method for stochastic optimization. In: ICLR (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.827, + 0.764, + 0.84 + ], + "angle": 0, + "content": "23. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: ICLR (2014)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "A. Hou et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "24. Lagunas, M., Sun, X., Yang, J., Villegas, R., Zhang, J., Shu, Z., Masia, B., Gutierrez, D.: Single-image full-body human relighting. In: EGSR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "25. Le, H., Kakadiaris, I.: Illumination-invariant face recognition with deep relit face images. In: WACV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.229 + ], + "angle": 0, + "content": "26. Liu, Y., Hou, A., Huang, X., Ren, L., Liu, X.: Blind removal of facial foreign shadows. In: BMVC (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.256 + ], + "angle": 0, + "content": "27. Lorenz, P., Durall, R.L., Keuper, J.: Detecting images generated by deep diffusion models using their local intrinsic dimensionality. In: ICCVW (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.297 + ], + "angle": 0, + "content": "28. Mei, Y., Zhang, H., Zhang, X., Zhang, J., Shu, Z., Wang, Y., Wei, Z., Yan, S., Jung, H., Patel, V.M.: LightPainter: Interactive portrait relighting with freehand scribble. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.325 + ], + "angle": 0, + "content": "29. Nestmeyer, T., Lalonde, J.F., Matthews, I., Lehrmann, A.: Learning physics-guided face relighting under directional light. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.326, + 0.785, + 0.352 + ], + "angle": 0, + "content": "30. Ojha, U., Li, Y., Lee, Y.J.: Towards universal fake image detectors that generalize across generative models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.353, + 0.785, + 0.392 + ], + "angle": 0, + "content": "31. Pandey, R., Orts-Escalano, S., LeGendre, C., Haene, C., Bouaziz, S., Rhemann, C., Debevec, P., Fanello, S.: Total relighting: Learning to relight portraits for background replacement. In: SIGGRAPH (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.433 + ], + "angle": 0, + "content": "32. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in PyTorch. In: NeurIPSW (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.46 + ], + "angle": 0, + "content": "33. Ponglertnapakorn, P., Tritrong, N., Suwajanakorn, S.: DiFaReli: Diffusion face relighting. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.487 + ], + "angle": 0, + "content": "34. Ranjan, A., Yi, K.M., Chang, J.H.R., Tuzel, O.: FaceLit: Neural 3D relightable faces. In: CVPR (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.514 + ], + "angle": 0, + "content": "35. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.785, + 0.541 + ], + "angle": 0, + "content": "36. Sengupta, R., Curless, B., Kemelmacher-Shlizerman, I., Seitz, S.: A light stage on every desk. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.568 + ], + "angle": 0, + "content": "37. Sengupta, S., Kanazawa, A., Castillo, C.D., Jacobs, D.W.: SfSNet: Learning shape, reflectance and illuminance of faces in the wild. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.595 + ], + "angle": 0, + "content": "38. Shih, Y., Paris, S., Barnes, C., Freeman, W.T., Durand, F.: Style transfer for headshot portraits. TOG (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.785, + 0.623 + ], + "angle": 0, + "content": "39. Shu, Z., Hadap, S., Shechtman, E., Sunkavalli, K., Paris, S., Samaras, D.: Portrait lighting transfer using a mass transport approach. TOG (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.624, + 0.785, + 0.65 + ], + "angle": 0, + "content": "40. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: ICLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.785, + 0.69 + ], + "angle": 0, + "content": "41. Sun, T., Barron, J.T., Tsai, Y.T., Xu, Z., Yu, X., Fyffe, G., Rhemann, C., Busch, J., Debevec, P., Ramamoorthi, R.: Single image portrait relighting. In: SIGGRAPH (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.718 + ], + "angle": 0, + "content": "42. Sun, T., Lin, K.E., Bi, S., Xu, Z., Ramamoorthi, R.: NeLF: Neural light-transport field for portrait view synthesis and relighting. In: EGSR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.719, + 0.785, + 0.759 + ], + "angle": 0, + "content": "43. Tan, F., Fanello, S., Meka, A., OrtsEscolano, S., Tang, D., Pandey, R., Taylor, J., Tan, P., Zhang., Y.: VoLux-GAN: A generative model for 3D face synthesis with HDRI relighting. In: SIGGRAPH (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.76, + 0.785, + 0.785 + ], + "angle": 0, + "content": "44. Wang, Y., Holynski, A., Zhang, X., Zhang, X.: Sunstage: Portrait reconstruction and relighting using the sun as a light stage. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "45. Wang, Z., Yu, X., Lu, M., Wang, Q., Qian, C., Xu, F.: Single image portrait relighting via explicit multiple reflectance channel modeling. TOG (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "46. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. TIP (2004)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.375, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "COMPOSE: Comprehensive Portrait Shadow Editing" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "47. Weir, J., Zhao, J., Chalmers, A., Rhee, T.: Deep portrait delighting. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "48. Yang, X., Taketomi, T.: BareSkinNet: Demakeup and de-lighting via 3D face reconstruction. CGF (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.245 + ], + "angle": 0, + "content": "49. Yeh, Y.Y., Nagano, K., Khamis, S., Kautz, J., Liu, M.Y., Wang, T.C.: Learning to relight portrait images via a virtual light stage and synthetic-to-real adaptation. TOG (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.272 + ], + "angle": 0, + "content": "50. Zhang, L., Zhang, Q., Wu, M., Yu, J., Xu, L.: Neural video portrait relighting in real-time via consistency modeling. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "51. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "52. Zhang, X., Fanello, S., Tsai, Y.T., Sun, T., Xue, T., Pandey, R., Orts-Escalano, S., Davidson, P., Rhemann, C., Debevec, P., Barron, J.T., Ramamoorthi, R., Freeman, W.T.: Neural light transport for relighting and view synthesis. TOG (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "53. Zhang, X., Barron, J.T., Tsai, Y.T., Pandey, R., Zhang, X., Ng, R., Jacobs, D.E.: Portrait Shadow Manipulation. TOG (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "54. Zhou, H., Hadap, S., Sunkavalli, K., Jacobs, D.W.: Deep single portrait image relighting. In: ICCV (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.397 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/COMPOSE_ Comprehensive Portrait Shadow Editing/2be02136-d81b-443f-a244-263becec430b_origin.pdf b/2024/COMPOSE_ Comprehensive Portrait Shadow Editing/2be02136-d81b-443f-a244-263becec430b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0f6b570a4905d97c004c63348ca639f49dc5ba52 --- /dev/null +++ b/2024/COMPOSE_ Comprehensive Portrait Shadow Editing/2be02136-d81b-443f-a244-263becec430b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b1dc614a3d06d841624eb00f85d505a80a689c402d2a316d82c40a24433d77ac +size 8419687 diff --git a/2024/COMPOSE_ Comprehensive Portrait Shadow Editing/full.md b/2024/COMPOSE_ Comprehensive Portrait Shadow Editing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f8ad81a92f501ad9ec384d76a72314bebcc1c987 --- /dev/null +++ b/2024/COMPOSE_ Comprehensive Portrait Shadow Editing/full.md @@ -0,0 +1,291 @@ +# COMPOSE: Comprehensive Portrait Shadow Editing + +Andrew Hou $^{1,2}$ , Zhixin Shu $^{2}$ , Xuaner Zhang $^{2}$ , He Zhang $^{2}$ , Yannick Hold-Geoffroy $^{2}$ , Jae Shin Yoon $^{2}$ , and Xiaoming Liu $^{1}$ + +1 Michigan State University + +{houandr1, liuxm}@msu.edu + +2 Adobe Research + +{zshu,cezhang,hezhan,holdgeof,jaeyoon}@adobe.com + +Abstract. Existing portrait relighting methods struggle with precise control over facial shadows, particularly when faced with challenges such as handling hard shadows from directional light sources or adjusting shadows while remaining in harmony with existing lighting conditions. In many situations, completely altering input lighting is undesirable for portrait retouching applications: one may want to preserve some authenticity in the captured environment. Existing shadow editing methods typically restrict their application to just the facial region and often offer limited lighting control options, such as shadow softening or rotation. In this paper, we introduce COMPOSE: a novel shadow editing pipeline for human portraits, offering precise control over shadow attributes such as shape, intensity, and position, all while preserving the original environmental illumination of the portrait. This level of disentanglement and controllability is obtained thanks to a novel decomposition of the environment map representation into ambient light and an editable gaussian dominant light source. COMPOSE is a four-stage pipeline that consists of light estimation and editing, light diffusion, shadow synthesis, and finally shadow editing. We define facial shadows as the result of a dominant light source, encoded using our novel gaussian environment map representation. Utilizing an OLAT dataset, we have trained models to: (1) predict this light source representation from images, and (2) generate realistic shadows using this representation. We also demonstrate comprehensive and intuitive shadow editing with our pipeline. Through extensive quantitative and qualitative evaluations, we have demonstrated the robust capability of our system in shadow editing. + +Keywords: Face Relighting $\cdot$ Shadow Editing $\cdot$ Lighting Decomposition + +# 1 Introduction + +Single-image portrait relighting is a well-studied problem with steady interest from the computer vision and graphics communities. Portrait relighting has a + +![](images/e92076f726db6df6b454297cea98e599ed66f0e16c2f3ace97c12c87159ea937.jpg) +a) Source Image + +![](images/2dac28d700bc188aca1e09dbd6b263d9df18caee00c321e0b2622af0c39878da.jpg) +b) Soften +Shadow + +![](images/7f5066a52f3298bd87b9d1c057af746bd2568e06283a4407200818ec200df3ad.jpg) +c) Intensify Shadow +Fig. 1: Overview. COMPOSE is the first single-image portrait shadow editing method that achieves complete shadow editing control, i.e. adjusting shadow intensity, modifying light size, and changing shadow positions, all while preserving the source image's other lighting attributes (e.g. ambient light). + +![](images/0c6b65d77d59b90229177bb8ce0bf63e46af09e54217990d8e9a4a8cc1636dee.jpg) +d) Modify Light Size + +![](images/ffe6fcd76a686c49eca0658ac41ba5911471a525a22c44fc90d87c8a579075ce.jpg) +e) Rotate Shadow + +wide range of applications in computational photography [31], AR/VR [4,5,19], and downstream tasks in the face domain [11,25]. While many methods for single image portrait relighting have been proposed in recent years [17,18,29,31,33,41, 49,54], they often face challenges when dealing with facial shadows. + +One popular genre of face relighting methods [31, 41] uses low resolution HDR environment maps to represent lighting. This lighting representation is used in a conditional image translation framework to control the lighting effect of the output face image, thereby achieving facial relighting. The requirement for an environment map restricts users seeking to fine-tune lighting, such as minor modifications to shadow smoothness or direction. Additionally, a low-resolution environment map often lacks the precision needed for accurately depicting complex, high-frequency shadow effects. Conversely, several approaches use low-dimensional spherical harmonics (SH) for their lighting representation [37, 54]. These methods, based on simplified lighting and reflectance models, frequently produce unrealistic relit images. Moreover, they overlook occlusions in the image formation process, disallowing the generation and editing of cast shadows. SunStage [44] generates realistic shadows on faces but lacks generalizability and requires videos as input. Some recent works have focused on shadow editing tasks [11, 26, 53] using state-of-the-art image generators such as diffusion models. However, achieving precise control with these sophisticated generators remains a challenging task. Consequently, these methods often provide limited control in shadow synthesis, typically handling only shadow softening or completely removing the shadow, without precise editing of shadow intensity, shape, or position. + +To overcome the challenges in modeling and controlling facial shadows in portrait relighting, we introduce COMPOSE, a system that separates shadow modeling from ambient light in the lighting representation. COMPOSE enables flexible manipulation of shadow attributes, including position (lighting direction), smoothness, and intensity (See Fig. 1). The key innovation in COMPOSE is its shadow representation, which is compatible with environment map-based representations, making it easily integrable into existing relighting networks. Building on the concept of image-based lighting representations (e.g. HDR en + +vironment maps), our approach divides lighting effects into two components: shadows caused by dominant light sources and ambient lighting. We simplify the shadow component by attributing it to a single dominant light source on an environment map. This is encoded using a 2D isotropic Gaussian, with variables representing light position, standard deviation (indicating light size or diffusion), and scale (light intensity). The remaining lighting effects are attributed to a diffused environment map, modeling the image's ambient lighting. + +For a given face image, we first predict the lighting as a panorama environment map, from which we estimate the dominant light (shadow) parameters. The system then removes shadows to obtain a diffused, albedo-like image. Utilizing this diffuse image and the shadow representation, we train a network to synthesize shadows on the subject in the input image. This shadow synthesis process is adaptable, accepting various shadow-related parameters for controllable shadow synthesis, including position and shape. The synthesized shadow can be linearly blended with the diffuse lighting to create shadows of varying intensities. + +To train models for shadow diffusion and synthesis, we utilize an OLAT (one-light-at-a-time) dataset, captured through a light stage system. This dataset, combined with the environment map-based lighting representation, ensures our system's compatibility with previously proposed face relighting techniques. Our approach also uniquely offers the capability to accurately model shadows. + +Our contributions are thus as follows: + +$\diamond$ We propose the first single image portrait shadow editing method with precise control over various shadow attributes, including shadow shape, intensity, and position. Our system provides users with the flexibility to control shading while also preserving the other lighting attributes of the input images. +$\diamond$ We propose a novel lighting decomposition into ambient light and an ed-itable dominant gaussian light source, which enables disentangled shadow editing while preserving the remaining lighting attributes. +$\diamond$ We achieve state-of-the-art relighting performance quantitatively and qualitatively, including on a light-stage-captured test set and in-the-wild images. + +# 2 Related Work + +# 2.1 Portrait Relighting + +Many portrait relighting methods have been proposed in recent years that generally fall into one of two categories: focusing on reducing the data requirement and improving the relighting quality as much as possible with limited data [6, 10, 17, 18, 20, 24, 33, 34, 36-39, 43-45, 47, 49, 54] or using light stage data but modeling more complex lighting effects [4, 12, 28, 29, 31, 41, 42, 48, 50, 52]. Among single image portrait relighting methods, a substantial limitation is that they cannot preserve the original environmental lighting while editing shadows, and instead opt to relight by replacing the original lighting with a completely new environment [11], which is undesirable in many portrait editing applications. Existing methods also struggle to handle more directed lights and fail in the presence of substantial hard or intense shadows [37, 41, 54]. + +![](images/da4be91a1b0c7392faf38906e446e5662637c181577a9027be877216a23874f3.jpg) +Fig. 2: Method Overview. COMPOSE is a 4-stage shadow editing method consisting of single-image lighting estimation, light diffusion, shadow synthesis, and image compositing. By first estimating the dominant light position using the environment map regressed from the input image, COMPOSE can control shadow shape and intensity by controlling light spread and light intensity respectively as well as shadow position by changing the location of the dominant light. By estimating a diffuse image $\mathbf{I}_D$ as well as a shadowed image $\mathbf{I}_S$ , COMPOSE can generate the final edited image $\mathbf{I}_E$ through image compositing. + +Among portrait relighting methods that take multiple images or videos as input, the method SunStage [44] is able to maintain the lighting attributes of the input video's environment when performing relighting. However, the method must be trained using a video of a subject on a sunny day and involves an expensive optimization procedure. The model is also subject-specific and must be retrained or fine-tuned even for the same subject if they change their hairstyle, accessories, or clothing. COMPOSE has the advantage of being a single-image portrait relighting method, and is both generalizable and practical for in-the-wild photo editing applications. + +# 2.2 Shadow Editing + +Another line of work closely related to COMPOSE are methods that edit only facial shadows, including shadow removal [26,53] and shadow softening [11]. Most similar to our work is the method of [11], which is able to directly control the degree of shadow softening while preserving the original environmental lighting. However, unlike COMPOSE, they are restricted to only shadow softening and cannot handle other shadow editing operations like shadow intensifying, modifying shadow shape, and shadow rotation. Shadow removal methods [14,26,53] are often even more limited since they cannot control the degree of shadow softening and attempt to remove the shadow completely. + +# 2.3 Lighting Representations + +Existing portrait relighting methods generally utilize one of three lighting representations: Spherical Harmonics (SH) [18,33,34,37,54], dominant lighting direction [17,29], and environment maps [31,41,49]. SH lighting tends to only model + +the lower frequencies of lighting effects, and thus is unsuitable for modeling non-Lambertian hard shadows and specularities and is better suited for diffuse lightings. Representing the light as a dominant lighting direction can often handle the non-Lambertian effects of directional and point light sources, but they often leave other lighting attributes such as light size out of the equation (e.g. Hou et al. [17]). Moreover, they often do not handle diffuse lightings well and are intended for directional lights [17, 29]. The environment map representation is more flexible and is able to model light intensity, light size, and light position and can handle both diffuse and directional lights. However, what is largely missing in this domain is a way to properly decompose the environment map that enables precise controllability over different lighting components. Existing methods only perform relighting by completely changing the scene illumination with a new environment map without disentangling the effects of ambient, diffuse, and directional light [31, 41, 49]. This can be undesirable for many computational photography applications where the user wishes to preserve the ambience of the scene and edit only the facial shadow alone [11]. Moreover, the default environment map representation is unsuitable for some shadow editing applications such as shrinking the light. While enlarging the light can be performed by applying a gaussian blur, there is no equivalent operation that can be performed to shrink the light on an environment map. Our lighting representation decomposes the environment map into the ambient light and an editable gaussian dominant light source, which is adaptable in light intensity, size, and position. COMPOSE is thus able to perform a full suite of shadow editing operations thanks to its editable gaussian component and can edit the shadow alone without disrupting other lighting attributes (e.g. ambient light). + +# 3 Methodology + +# 3.1 Problem Formulation + +In our framework, shadows are treated as a composable lighting effect, which can be independently predicted and altered based on a specific face image. This shadow representation is integrated into the overall lighting representation, and our framework is designed to (1) accurately predict the shadow from an image, and (2) apply a controllable shadow onto a shadow-free face image. We define the portrait shadow editing problem based on properties of shadows that users can manipulate, namely shadow position, shadow intensity, and shadow shape. These shadow properties are connected to the lighting attribute that is easily controllable. Specifically, given source image $\mathbf{I}_N$ and desired edited output image $\mathbf{I}_E$ with modified shadows, our problem can be written as: + +$$ +\mathbf {I} _ {E} = F _ {\theta} \left(\mathbf {I} _ {N}, x, y, \sigma , \gamma\right), \tag {1} +$$ + +where $F_{\theta}$ is our shadow editing method, $x$ and $y$ are the desired light position on an environment map, $\sigma$ is the light size, and $\gamma$ is the light intensity. By editing the parameters, $(x,y)$ , $\sigma$ , and $\gamma$ , we can control the shadow position, shape, and intensity respectively. We describe our solution COMPOSE in detail in Sec. 3.2. + +# 3.2 COMPOSE Framework + +COMPOSE is a 4-stage shadow editing framework, consisting of single image lighting estimation, light diffusion, shadow synthesis, and finally image compositing. See Fig. 2 for an overview of our method. + +Lighting Estimation In the lighting estimation stage, we train a variational autoencoder (VAE) [23] to estimate an environment map (LDR) from a single portrait image as its lighting representation $\mathbf{I}_N$ . From this environment map, we then estimate the dominant light position. Naturally, the dominant light, such as the sun, represents the brightest intensity on a image-based lighting representation. Therefore we can simply obtain the dominant light position by fitting a 2D isotropic Gaussian on the predicted environment map. The center of the Gaussian will indicate the position of the dominant light on the 2D environment map. This estimated light position corresponds to the shadow information on the input image, and therefore is useful for "in-place" shadow editing (e.g., shadow softening, shadow intensifying) in the later stage that does not alter the direction of the dominant light. We argue this is desirable in many portrait editing applications where the user only wants to edit existing shadows and does not want to alter the ambient lighting. + +Light Diffusion In the light diffusion stage, we remove all existing hard shadows and specular highlights from the input image and output a "diffused" image $\mathbf{I}_D$ with very smooth shading. This represents the input face under only an ambient illumination condition. Our light diffusion network consists of a hierarchical transformer encoder that generates multi-level features, which are fed to a decoder with transposed convolutional layers. Our network takes as input the original input image, a body parsing mask, and a binary foreground mask. + +**Shadow Synthesis** In the shadow synthesis stage, we aim to produce a shadowed image $\mathbf{I}_S$ illuminated by an edited light source given the diffuse image $\mathbf{I}_D$ estimated from the light diffusion stage and an edited environment map. Using the estimated dominant light position from the light estimation stage, we have full control over light attributes such as light size, intensity, and the dominant light position. We can produce a new environment map using a Gaussian to represent the dominant light, where the light size is modeled by the standard deviation of the Gaussian (a larger standard deviation represents a larger, more diffuse light) and the light intensity can be adjusted by multiplying the Gaussian with a scalar. This allows our method to control shadow shape, light intensity, and position. To regress the shadowed image $\mathbf{I}_S$ , we propose a two-stage architecture: a U-Net followed by a conditional DDPM (See Fig. 3). For the first stage, we adopt the U-Net architecture of [41]. During training, instead of using a reshaped environment map-like representation for the dominant light, we design a feature map-like representation that's parameterized by the shadow attributes. Specifically, the lighting representation for shadows has 4 parameters. The first two parameters $x$ and $y$ encode the coordinates of the light center on the environment map, the third parameter $\sigma$ encodes the light size, and the fourth parameter $\gamma$ encodes the light intensity. All four parameters are normalized between 0 and 1 and repeated spatially as $32 \times 32$ channels to be suitable + +![](images/da5d8b375587fd13bd1aaa8eea2f8d3ada2444b94e726b4b7738256e4d2c6071.jpg) +Fig. 3: Shadow Synthesis Pipeline. Our shadow synthesis step consists of two stages: a U-Net followed by a conditional DDPM. The U-Net takes ambient image $\mathbf{I}_D$ and environment map $\mathbf{E}_T$ as input and outputs the coarse relit image $\mathbf{I}_U$ . With $\mathbf{I}_U$ and $\mathbf{E}_T$ as spatial conditions, the conditional DDPM outputs the final shadowed image $\mathbf{I}_S$ , refining shadow boundaries and improving image quality. + +inputs for the U-Net. We find that this lighting representation leads to faster convergence, especially when the desired light mimics a point light and occupies a tiny portion of the environment map. We feed the diffuse image $\mathbf{I}_D$ and our desired lighting feature map to the U-Net, and the output is a relit image $\mathbf{I}_U$ . While the U-Net architecture has merits such as complete geometric consistency for shadow synthesis, one demerit is that the synthesized shadow boundaries are often not very sharp and the overall image quality could be improved. We thus employ a DDPM [7,15] as the second stage of our shadow synthesis model. The role of the DDPM here is to take the output image of the U-Net $\mathbf{I}_U$ as a spatial condition along with the lighting parameters and perform image refinement to generate the final shadowed image $\mathbf{I}_S$ . The training objective of our shadow synthesis DDPM thus becomes: + +$$ +\mathbb {E} _ {t, \mathbf {x} _ {0}, \boldsymbol {\epsilon}} \left[ \| \boldsymbol {\epsilon} - \boldsymbol {\epsilon} _ {\theta} \left(\mathbf {x} _ {t}, t, \mathbf {I} _ {U}, x, y, \sigma , \gamma\right) \| ^ {2} \right]. \tag {2} +$$ + +Our condition $\mathbf{I}_U$ is spatially concatenated with $\mathbf{x}_t$ and our lighting parameters are repeated spatially as channels of the same resolution and similarly spatially concatenated, where $\mathbf{x}_t = \sqrt{\bar{\alpha}_t}\mathbf{x}_0 + \sqrt{1 - \bar{\alpha}_t}\epsilon$ . We find that the quality of the edited images is noticeably enhanced by the DDPM compared to the U-Net alone and that adding the DDPM sharpens the shadow boundaries. + +Image Compositing With $\mathbf{I}_D$ from our light diffusion stage and $\mathbf{I}_S$ from our shadow synthesis stage, we can then perform image compositing between $\mathbf{I}_D$ and $\mathbf{I}_S$ to achieve our final edited image $\mathbf{I}_E$ , by assigning weights $\omega_{D}$ and $\omega_{S}$ to $\mathbf{I}_D$ and $\mathbf{I}_S$ respectively. The final edited image $\mathbf{I}_E$ is thus generated as: + +$$ +\mathbf {I} _ {E} = \omega_ {D} \mathbf {I} _ {D} + \omega_ {S} \mathbf {I} _ {S}, \tag {3} +$$ + +where $\omega_{D} + \omega_{S} = 1$ . By adjusting $\omega_{D}$ and $\omega_{S}$ , we can further tune shadow intensity: a higher $\omega_{D}$ softens shadows and a higher $\omega_{S}$ results in darker shadows. + +# 3.3 Loss Functions + +To train our single image lighting estimation network, we apply two loss functions commonly used in VAE training [23]: $\mathcal{L}_{\mathrm{recon}}$ and $\mathcal{L}_{\mathrm{KLD}}$ . $\mathcal{L}_{\mathrm{recon}}$ is a reconstruction loss between the predicted environment map $\mathbf{E}_P$ and the groundtruth + +environment map $\mathbf{E}_G$ defined as follows: + +$$ +\mathcal {L} _ {\text {r e c o n}} = \frac {1}{3 H W} \left(\left\| \mathbf {E} _ {P} - \mathbf {E} _ {G} \right\| _ {2} ^ {2}\right), \tag {4} +$$ + +where $H$ and $W$ are the size of the environment maps. The Kullback-Leibler divergence (KLD) loss $\mathcal{L}_{\mathrm{KLD}}$ is: + +$$ +\mathcal {L} _ {\mathrm {K L D}} = - \frac {1}{2 N} \sum_ {i = 1} ^ {N} \left(1 + \log \left(\sigma_ {i} ^ {2}\right) - \mu_ {i} ^ {2} - \sigma_ {i} ^ {2}\right), \tag {5} +$$ + +where $N$ is the dimensionality of the latent vector, $\mu$ is the batch mean, and $\sigma^2$ is the batch variance. The final loss for our lighting estimation network is thus: + +$$ +\mathcal {L} _ {\mathrm {L E}} = \lambda_ {1} \mathcal {L} _ {\text {r e c o n}} + \lambda_ {2} \mathcal {L} _ {\mathrm {K L D}}, \tag {6} +$$ + +where $\lambda_1 = 1$ and $\lambda_{2} = 2.5\times 10^{-4}$ + +Our light diffusion network is also trained with two losses $\mathcal{L}_{\mathrm{recon}}$ and $\mathcal{L}_{\mathrm{perceptual}}$ . + +$$ +\mathcal {L} _ {\text {r e c o n}} = \frac {1}{3 H W} \left(\left\| \mathbf {I} _ {D} - \mathbf {I} _ {G} \right\| _ {1}\right), \tag {7} +$$ + +where $H$ and $W$ are the size of the training images, $\mathbf{I}_D$ is the predicted diffuse image, and $\mathbf{I}_G$ is the groundtruth diffuse image. $\mathcal{L}_{\mathrm{perceptual}}$ enforces visual similarity [51] and is computed as the distance between VGG [40] features computed for $\mathbf{I}_D$ and $\mathbf{I}_G$ . The final loss for the light diffusion network is thus: + +$$ +\mathcal {L} _ {\mathrm {L D}} = \mathcal {L} _ {\text {r e c o n}} + \mathcal {L} _ {\text {p e r c e p t u a l}}. \tag {8} +$$ + +When training the U-Net of our shadow synthesis pipeline, we employ a single reconstruction loss $\mathcal{L}_{\mathrm{recon}}$ between the predicted relit image $\mathbf{I}_S$ and the groundtruth $\mathbf{I}_G$ , defined as follows: + +$$ +\mathcal {L} _ {\text {r e c o n}} = \frac {1}{3 H W} \left(\| \mathbf {I} _ {S} - \mathbf {I} _ {G} \| _ {2} ^ {2}\right). \tag {9} +$$ + +# 3.4 Data Collection and Generation + +To train out networks, we use a light stage to capture one-light-at-a-time (OLAT) images of 107 subjects with diverse skin tones, body poses, expressions, and accessories using a light stage [9] with 160 lights. Each subject is captured under 10-20 sessions with varied body poses, expressions, and accessories. Each session is captured by 4 cameras with varying camera poses. + +To generate input images to train our lighting estimation VAE and our light diffusion network, we render our OLAT images using a diverse set of environment maps collected from the Polyhaven and Laval Outdoor HDR datasets [16]. During rendering, we apply random augmentations to the environment map including rotations and adjusting intensity to improve our model's generalization to different lighting. The groundtruth $\mathbf{E}_G$ for our lighting estimation VAE is + +simply the environment map used for rendering, and the groundtruth $\mathbf{I}_G$ for the light diffusion network is generated by rendering OLAT images with heavily blurred, diffuse environment maps. + +To generate images for our shadow synthesis stage, we render groundtruth relit images by using environment maps consisting of Gaussian lights with various positions, intensities, and sizes (See Sec. 3.2) along with our OLAT images. As the input to the shadow synthesis stage is meant to be the diffuse image $\mathbf{I}_D$ predicted by the light diffusion stage, we generate diffuse images for training our shadow synthesis networks by rendering OLAT images with heavily blurred environment maps in the Polyhaven and Laval Outdoor HDR datasets. + +# 3.5 Implementation Details + +We implement all components of COMPOSE using PyTorch [32]. When training our lighting estimation VAE, we use a batch size of 128 and a learning rate of $10^{-4}$ on 4 24GB A10G GPUs. For our light diffusion network, we train with a batch size of 88 and a learning rate of $10^{-4}$ using 8 A100 GPUs. For our shadow synthesis U-Net, we train with a batch size of 100 and a learning rate of $10^{-4}$ using 4 24GB A10G GPUs. For all networks, we train using the Adam Optimizer [22]. The image resolution used to train the lighting estimation and shadow synthesis networks is $256 \times 256$ and the resolution used to train the light diffusion network is $768 \times 768$ . Please see the supplementary materials for architectural details of the lighting estimation network, the lighting diffusion network, and the shadow synthesis network. + +# 4 Experiments + +# 4.1 Datasets + +To train our lighting estimation and shadow synthesis networks, we generate our training sets using light stage images and 521 outdoor environment maps. Our training set consists of 82 subjects with significant variation in factors such as skin tones, pose, expressions, and accessories. + +For evaluation, we use two separate evaluation sets: the first is generated using our light stage images and has corresponding groundtruth, and the second consists of in-the-wild images that represent more realistic application settings and do not have groundtruth. Our light stage evaluation set consists of 48 images from a total of 16 subjects with diverse skin tones, body poses, expressions, and accessories. All evaluation subjects and lightings are held out and unseen during training. Our in-the-wild evaluation set also consists of unseen test subjects and applies unseen lighting conditions in all qualitative results. + +# 4.2 Shadow Editing Evaluation + +We evaluate our shadow editing performance (synthesis and removal) by using randomly sampled, unseen lighting positions as the target lights to generate relit light stage images of our 16 test subjects, which serve as groundtruth. + +![](images/4ff79da6fd3733f1cbc65c51d4e3adacfff9cb73a3e0e397cb3c7c011be1bc5e.jpg) +Fig. 4: Shadow Synthesis. Our model synthesizes more plausible shadows than the baselines, and, unlike [17], is also able to properly remove shadows from the source image before relighting. Environment maps are shown to help visualize each test lighting. + +Table 1: Shadow Editing Performance. We evaluate shadow editing using unseen lightings with varied intensity and size on 16 held out highly diverse light stage test subjects with groundtruth relit images. COMPOSE outperforms all baselines quantitatively in shadow editing performance across all metrics (mean ± standard deviation). + +
MAEMSESSIMLPIPS
Hou et al. [17]0.1172 ± 0.07540.0438 ± 0.05030.7059 ± 0.10610.2371 ± 0.0871
Total Relighting [31]0.1008 ± 0.07430.0379 ± 0.04470.7754 ± 0.07380.2120 ± 0.0572
COMPOSE (Ours)0.0965 ± 0.06860.0349 ± 0.03680.7780 ± 0.08260.1973 ± 0.0678
+ +Each input image is rendered with a randomly selected environment map out of 125 unseen outdoor environment maps and is first passed to our deshading network to generate diffuse image $\mathbf{I}_D$ . $\mathbf{I}_D$ and target environment map $\mathbf{E}_T$ are then fed to our two-stage shadow synthesis pipeline to generate the newly shadowed image $\mathbf{I}_S$ , which is evaluated against the groundtruth light stage image. + +Tab. 1 compares our shadow editing performance against prior relighting methods. For both baselines [17, 31], the test results were provided by the authors. Our method achieves state-of-the-art results on all metrics (MAE, MSE, SSIM [46], and LPIPS [51]). As seen in Fig. 4, our model can synthesize appropriate shadows for various light positions. The method of Hou et al. [17] cannot + +![](images/a4aac90790bd26aafb1f059b02033460c6695eb59b2d464b475866ddebbdd7c0.jpg) +Fig. 5: Shadow Editing. Our method achieves complete shadow editing control, including softening/intensifying shadows, changing light size to alter shape and intensity, and changing light position, all while preserving the source image's ambient light. + +remove existing shadows in the source image and the shadow traces carry over as artifacts to the relit images. In addition, Hou et al. [17] only models the lighting direction and does not model the light size as a parameter, which leads to inaccurate shadow shape when the light size is varied. Total Relighting [31] is often unable to synthesize physically plausible shadows, and will sometimes overshadow (Fig. 4, rows 1, 3, and 4) or undershadow the image (Fig. 4, row 2). Moreover, their shadows are often blurry and not as sharp as our method. Compared to the baselines, COMPOSE can properly remove existing shadows from the source image and synthesize geometrically plausible and realistic shadows for a wide variety of lighting conditions. + +# 4.3 Portrait Shadow Editing Applications + +We demonstrate the full performance of COMPOSE by performing all forms of shadow editing, including shadow softening, shadow intensifying, modifying light size (which changes both shadow shape and intensity), and shadow rotation + +![](images/ac06b94f709f8530bd8dd5e9253be67ce2f592f0927ecc3dff4782577d449afd.jpg) +Fig.6: Degree of Shadow Editing. We demonstrate COMPOSE's controllable shadow editing by varying shadow intensity (top row), light diffusion for modifying shadow shape and intensity (middle row), and light position (bottom row). + +by changing the dominant light position. For each image, we first estimate the diffuse image $\mathbf{I}_D$ and the shadowed image $\mathbf{I}_S$ and perform image compositing between these two images to generate the final edited image $\mathbf{I}_E$ . As seen in Fig. 5, COMPOSE properly edits the shadows of in-the-wild images. To soften or intensify the shadow without changing the shadow shape, we rely on our image compositing coefficients $\omega_d$ and $\omega_s$ , which control the contributions of $\mathbf{I}_D$ and $\mathbf{I}_S$ respectively. We increase the weight of $\omega_d$ to soften the shadow and increase the weight of $\omega_s$ to strengthen the shadow. We further show that we can modify light size by tuning the Gaussian spread $\sigma$ , which changes both shadow shape and intensity. Enlarging the light makes the shadow more diffuse and less intense, whereas shrinking the light generates a harder, more intense shadow. For the subjects in rows 2, 3, and 4 we enlarged the light, resulting in a more diffuse shadow especially noticeable around the nose. For the subject in row 1 we instead shrunk the light and appropriately produce an edited image with more prominent and intense hard shadows. Finally, we also demonstrate shadow rotation by changing the dominant light position in $\mathbf{I}_S$ 's environment map. + +To demonstrate COMPOSE can properly control shadow editing, we visualize various degrees of shadow softening/intensifying, light diffusion, and shadow rotation in Fig. 6. The first row shows varying shadow intensity, where the shadows become less intense from left to right using $\omega_{d}$ and $\omega_{s}$ . Our results demonstrate that these flexible weights allow for a wide range of shadow intensities. The second row shows increasing degrees of light diffusion from left to right, which corresponds to increasing the light size. The shadows in the generated images + +![](images/ffaadec496b1350ecde03d9439a9a2705103120165e8fd6b6f3d5d3789ba7e70.jpg) +Fig. 7: Controllable Shadow Softening. For each subject, we show the shadow softening results of [11] in the first row and COMPOSE in the second. COMPOSE leaves less shadow traces at level 0.0, where the goal is to remove the shadows completely. + +correctly transition from harder shadows to more diffuse shadows as the light size increases. The third row shows shadow rotation, where the light is rotated horizontally along the environment map and around the subject. COMPOSE is able to produce physically plausible shadows as the light position changes. + +It is important to note that this level of controllable shadow editing is not achievable by existing portrait shadow editing methods [11, 53] that are only able to handle shadow softening or complete shadow removal as well as existing portrait relighting methods that completely modify the existing environment and cannot preserve the original scene illumination [31], do not properly model all shadow attributes (e.g. intensity, shape, and position) [17, 33] or are not designed to handle lighting conditions with more substantial shadows [37, 41, 54]. + +# 4.4 Shadow Softening Comparison + +We compare on in-the-wild shadow softening with the state-of-the-art portrait shadow softening method Futschik et al. [11]. Using results provided by the authors, we visualize different degrees of shadow softening for [11] by varying the light diffusion parameter (0.75, 0.5, 0.2, and 0.0) and compare with equivalent degrees of shadow softening by COMPOSE, as shown in Fig. 7. Higher parameter values correspond to less softening whereas lower values correspond to more softening. At 0.0, the goal is to completely remove the source image's shadows. Across all degrees of light diffusion, both methods gradually soften the shadow, but COMPOSE leaves less shadow traces than [11] when completely removing + +the source image's shadows at level 0.0, especially for images with darker shadows. This is thanks to our hierarchical transformer design for our light diffusion network, which better handles removing the effects of shadows at all scales. + +# 5 Conclusion + +We have proposed COMPOSE: the first single image portrait shadow editing method that can perform comprehensive variations of shadow editing including editing shadow intensity, light size, and shadow position while preserving the source image's lighting environment. This is largely enabled by our novel lighting decomposition into ambient light and an editable gaussian dominant light, which enables new applications like shrinking the light and intensifying shadows as well as disentangled shadow editing that preserves the scene ambience. We have demonstrated the novel photo editing applications enabled by our method over prior work qualitatively and that our method achieves state-of-the-art shadow editing performance quantitatively over prior relighting methods. We hope that our work can inspire and motivate more research in the exciting new research area of controllable and comprehensive portrait shadow editing. + +Limitations. Good directions for future work include handling lighting environments with multiple light sources and indoor environments, as our framework is intended mostly for outdoor settings, where the sun is the single dominant light. In addition, while $\mathbf{I}_D$ can preserve the ambience, we still notice there are sometimes color shifts with the environmental lighting during shadow editing, largely due to the compositing step with shadowed image $\mathbf{I}_S$ (which adopts the light stage's environmental lighting). This could be alleviated by modeling the diffuse lighting tint similar to [11] or by training with white balanced data to remove the color bias of the light stage. We also believe that one can continue to improve the generalizability of our model. Similar to other relighting methods that rely on light stage data [11, 31, 41], we train with a limited number of subjects. Future work may incorporate synthetic data to expand the training set or adopt latent diffusion models, which naturally have high generalizability [35]. Finally, one can further improve the identity preservation of our method. We find that high frequency details are sometimes lost at the initial U-Net stage of shadow synthesis, and ways to better preserve the identity include passing $\mathbf{I}_D$ or the input image $\mathbf{I}_N$ , which contain high frequency details, as additional conditions to the DDPM or to use stronger backbones than [41]. + +Potential Societal Impacts. In image editing, there are always concerns of generating malicious deepfakes. However, we only alter illumination, not subject identity. Even so, there may be concerns that altering the illumination and especially adding dark shadows to faces could be used as a malicious attack to worsen the performance of downstream tasks such as face recognition. Users can detect these attacks using existing deepfake detection methods [1-3,8,13,21,27,30] or by training one using the edited images generated by COMPOSE. State-of-the-art shadow removal methods [26,53] are also an option to counter these potential attacks since they can remove any facial shadows that COMPOSE synthesizes. + +# References + +1. Aghasanli, A., Kangin, D., Angelov, P.: Interpretable-through-prototypes deepfake detection for diffusion models. In: ICCVW (2023) +2. Asnani, V., Yin, X., Hassner, T., Liu, X.: MaLP: Manipulation localization using a proactive scheme. In: CVPR (2023) +3. Asnani, V., Yin, X., Hassner, T., Liu, X.: Reverse engineering of generative models: Inferring model hyperparameters from generated images. PAMI (2023) +4. Bi, S., Lombardi, S., Saito, S., Simon, T., Wei, S.E., McPhail, K., Ramamoorthi, R., Sheikh, Y., Saragih, J.: Deep relightable appearance models for animatable faces. TOG (2021) +5. Caselles, P., Ramon, E., Garcia, J., i Nieto, X.G., Moreno-Noguer, F., Triginer, G.: SIRA: Relightable avatars from a single image. In: WACV (2023) +6. Chandran, S., Hold-Geoffroy, Y., Sunkavalli, K., Shu, Z., Jayasuriya, S.: Temporally consistent relighting for portrait videos. In: WACV (2022) +7. Choi, J., Kim, S., Jeong, Y., Gwon, Y., Yoon, S.: ILVR: Conditioning method for denoising diffusion probabilistic models. In: ICCV (2021) +8. Corvi, R., Cozzolino, D., Zingarini, G., Poggi, G., Nagano, K., Verdoliva, L.: On the detection of synthetic images generated by diffusion models. In: ICASSP (2023) +9. Debevec, P., Hawkins, T., Tchou, C., Duiker, H.P., Sarokin, W., Sagar, M.: Acquiring the reflectance field of a human face. In: SIGGRAPH (2000) +10. Ding, Z., Zhang, C., Xia, Z., Jebe, L., Tu, Z., Zhang, X.: DiffusionRig: Learning personalized priors for facial appearance editing. In: CVPR (2023) +1. Futschik, D., Ritland, K., Vecore, J., Fanello, S., Orts-Escolano, S., Curless, B., Sykora, D., Pandey, R.: Controllable light diffusion for portraits. In: CVPR (2023) +12. Guo, K., Lincoln, P., Davidson, P., Busch, J., Yu, X., Whalen, M., Harvey, G., Orts-Escolano, S., Pandey, R., Dourgarian, J., DuVall, M., Tang, D., Tkach, A., Kowdle, A., Cooper, E., Dou, M., Fanello, S., Fyffe, G., Rhemann, C., Taylor, J., Debevecc, P., Izadi, S.: The Relighttables: Volumetric performance capture of humans with realistic relighting. In: SIGGRAPH Asia (2019) +13. Guo, X., Liu, X., Ren, Z., Grosz, S., Masi, I., Liu, X.: Hierarchical fine-grained image forgery detection and localization. In: CVPR (2023) +14. He, Y., Xing, Y., Zhang, T., Chen, Q.: Unsupervised portrait shadow removal via generative priors. In: MM (2021) +15. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020) +16. Hold-Geoffroy, Y., Athawale, A., Lalonde, J.F.: Deep sky modeling for single image outdoor lighting estimation. In: CVPR (2019) +17. Hou, A., Sarkis, M., Bi, N., Tong, Y., Liu, X.: Face relighting with geometrically consistent shadows. In: CVPR (2022) +18. Hou, A., Zhang, Z., Sarkis, M., Bi, N., Tong, Y., Liu, X.: Towards high fidelity face relighting with realistic shadows. In: CVPR (2021) +19. Iqbal, U., Caliskan, A., Nagano, K., Khamis, S., Molchanov, P., Kautz, J.: RANA: Relightable articulated neural avatars. In: ICCV (2023) +20. Ji, C., Yu, T., Guo, K., Liu, J., Liu, Y.: Geometry-aware single-image full-body human relighting. In: ECCV (2022) +21. Kamat, S., Agarwal, S., Darrell, T., Rohrbach, A.: Revisiting generalizability in deepfake detection: Improving metrics and stabilizing transfer. In: ICCVW (2023) +22. Kingma, D., Adam, J.B.: A method for stochastic optimization. In: ICLR (2014) +23. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: ICLR (2014) + +24. Lagunas, M., Sun, X., Yang, J., Villegas, R., Zhang, J., Shu, Z., Masia, B., Gutierrez, D.: Single-image full-body human relighting. In: EGSR (2021) +25. Le, H., Kakadiaris, I.: Illumination-invariant face recognition with deep relit face images. In: WACV (2019) +26. Liu, Y., Hou, A., Huang, X., Ren, L., Liu, X.: Blind removal of facial foreign shadows. In: BMVC (2022) +27. Lorenz, P., Durall, R.L., Keuper, J.: Detecting images generated by deep diffusion models using their local intrinsic dimensionality. In: ICCVW (2023) +28. Mei, Y., Zhang, H., Zhang, X., Zhang, J., Shu, Z., Wang, Y., Wei, Z., Yan, S., Jung, H., Patel, V.M.: LightPainter: Interactive portrait relighting with freehand scribble. In: CVPR (2023) +29. Nestmeyer, T., Lalonde, J.F., Matthews, I., Lehrmann, A.: Learning physics-guided face relighting under directional light. In: CVPR (2020) +30. Ojha, U., Li, Y., Lee, Y.J.: Towards universal fake image detectors that generalize across generative models. In: CVPR (2023) +31. Pandey, R., Orts-Escalano, S., LeGendre, C., Haene, C., Bouaziz, S., Rhemann, C., Debevec, P., Fanello, S.: Total relighting: Learning to relight portraits for background replacement. In: SIGGRAPH (2021) +32. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in PyTorch. In: NeurIPSW (2017) +33. Ponglertnapakorn, P., Tritrong, N., Suwajanakorn, S.: DiFaReli: Diffusion face relighting. In: ICCV (2023) +34. Ranjan, A., Yi, K.M., Chang, J.H.R., Tuzel, O.: FaceLit: Neural 3D relightable faces. In: CVPR (June 2023) +35. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) +36. Sengupta, R., Curless, B., Kemelmacher-Shlizerman, I., Seitz, S.: A light stage on every desk. In: ICCV (2021) +37. Sengupta, S., Kanazawa, A., Castillo, C.D., Jacobs, D.W.: SfSNet: Learning shape, reflectance and illuminance of faces in the wild. In: CVPR (2018) +38. Shih, Y., Paris, S., Barnes, C., Freeman, W.T., Durand, F.: Style transfer for headshot portraits. TOG (2014) +39. Shu, Z., Hadap, S., Shechtman, E., Sunkavalli, K., Paris, S., Samaras, D.: Portrait lighting transfer using a mass transport approach. TOG (2017) +40. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: ICLR (2015) +41. Sun, T., Barron, J.T., Tsai, Y.T., Xu, Z., Yu, X., Fyffe, G., Rhemann, C., Busch, J., Debevec, P., Ramamoorthi, R.: Single image portrait relighting. In: SIGGRAPH (2019) +42. Sun, T., Lin, K.E., Bi, S., Xu, Z., Ramamoorthi, R.: NeLF: Neural light-transport field for portrait view synthesis and relighting. In: EGSR (2021) +43. Tan, F., Fanello, S., Meka, A., OrtsEscolano, S., Tang, D., Pandey, R., Taylor, J., Tan, P., Zhang., Y.: VoLux-GAN: A generative model for 3D face synthesis with HDRI relighting. In: SIGGRAPH (2022) +44. Wang, Y., Holynski, A., Zhang, X., Zhang, X.: Sunstage: Portrait reconstruction and relighting using the sun as a light stage. In: CVPR (2023) +45. Wang, Z., Yu, X., Lu, M., Wang, Q., Qian, C., Xu, F.: Single image portrait relighting via explicit multiple reflectance channel modeling. TOG (2020) +46. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. TIP (2004) + +47. Weir, J., Zhao, J., Chalmers, A., Rhee, T.: Deep portrait delighting. In: ECCV (2022) +48. Yang, X., Taketomi, T.: BareSkinNet: Demakeup and de-lighting via 3D face reconstruction. CGF (2022) +49. Yeh, Y.Y., Nagano, K., Khamis, S., Kautz, J., Liu, M.Y., Wang, T.C.: Learning to relight portrait images via a virtual light stage and synthetic-to-real adaptation. TOG (2022) +50. Zhang, L., Zhang, Q., Wu, M., Yu, J., Xu, L.: Neural video portrait relighting in real-time via consistency modeling. In: ICCV (2021) +51. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) +52. Zhang, X., Fanello, S., Tsai, Y.T., Sun, T., Xue, T., Pandey, R., Orts-Escalano, S., Davidson, P., Rhemann, C., Debevec, P., Barron, J.T., Ramamoorthi, R., Freeman, W.T.: Neural light transport for relighting and view synthesis. TOG (2021) +53. Zhang, X., Barron, J.T., Tsai, Y.T., Pandey, R., Zhang, X., Ng, R., Jacobs, D.E.: Portrait Shadow Manipulation. TOG (2020) +54. Zhou, H., Hadap, S., Sunkavalli, K., Jacobs, D.W.: Deep single portrait image relighting. In: ICCV (2019) \ No newline at end of file diff --git a/2024/COMPOSE_ Comprehensive Portrait Shadow Editing/images.zip b/2024/COMPOSE_ Comprehensive Portrait Shadow Editing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2faeeb9825873d91f479126923c3a60a2e98e277 --- /dev/null +++ b/2024/COMPOSE_ Comprehensive Portrait Shadow Editing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80d70e5d67a8582f71804fdf893578fa7cd4d21e570a7064069456901976dbb3 +size 516553 diff --git a/2024/COMPOSE_ Comprehensive Portrait Shadow Editing/layout.json b/2024/COMPOSE_ Comprehensive Portrait Shadow Editing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e3439e511281bd9f777f322552ab2bb4b136539f --- /dev/null +++ b/2024/COMPOSE_ Comprehensive Portrait Shadow Editing/layout.json @@ -0,0 +1,9325 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 465, + 148 + ], + "type": "text", + "content": "COMPOSE: Comprehensive Portrait Shadow Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 160, + 167, + 455, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 167, + 455, + 194 + ], + "spans": [ + { + "bbox": [ + 160, + 167, + 455, + 194 + ], + "type": "text", + "content": "Andrew Hou" + }, + { + "bbox": [ + 160, + 167, + 455, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 160, + 167, + 455, + 194 + ], + "type": "text", + "content": ", Zhixin Shu" + }, + { + "bbox": [ + 160, + 167, + 455, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 160, + 167, + 455, + 194 + ], + "type": "text", + "content": ", Xuaner Zhang" + }, + { + "bbox": [ + 160, + 167, + 455, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 160, + 167, + 455, + 194 + ], + "type": "text", + "content": ", He Zhang" + }, + { + "bbox": [ + 160, + 167, + 455, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 160, + 167, + 455, + 194 + ], + "type": "text", + "content": ", Yannick Hold-Geoffroy" + }, + { + "bbox": [ + 160, + 167, + 455, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 160, + 167, + 455, + 194 + ], + "type": "text", + "content": ", Jae Shin Yoon" + }, + { + "bbox": [ + 160, + 167, + 455, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 160, + 167, + 455, + 194 + ], + "type": "text", + "content": ", and Xiaoming Liu" + }, + { + "bbox": [ + 160, + 167, + 455, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 248, + 201, + 365, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 201, + 365, + 213 + ], + "spans": [ + { + "bbox": [ + 248, + 201, + 365, + 213 + ], + "type": "text", + "content": "1 Michigan State University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 249, + 214, + 364, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 214, + 364, + 224 + ], + "spans": [ + { + "bbox": [ + 249, + 214, + 364, + 224 + ], + "type": "text", + "content": "{houandr1, liuxm}@msu.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 269, + 224, + 344, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 224, + 344, + 234 + ], + "spans": [ + { + "bbox": [ + 269, + 224, + 344, + 234 + ], + "type": "text", + "content": "2 Adobe Research" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 193, + 236, + 421, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 236, + 421, + 247 + ], + "spans": [ + { + "bbox": [ + 193, + 236, + 421, + 247 + ], + "type": "text", + "content": "{zshu,cezhang,hezhan,holdgeof,jaeyoon}@adobe.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 278, + 455, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 278, + 455, + 553 + ], + "spans": [ + { + "bbox": [ + 160, + 278, + 455, + 553 + ], + "type": "text", + "content": "Abstract. Existing portrait relighting methods struggle with precise control over facial shadows, particularly when faced with challenges such as handling hard shadows from directional light sources or adjusting shadows while remaining in harmony with existing lighting conditions. In many situations, completely altering input lighting is undesirable for portrait retouching applications: one may want to preserve some authenticity in the captured environment. Existing shadow editing methods typically restrict their application to just the facial region and often offer limited lighting control options, such as shadow softening or rotation. In this paper, we introduce COMPOSE: a novel shadow editing pipeline for human portraits, offering precise control over shadow attributes such as shape, intensity, and position, all while preserving the original environmental illumination of the portrait. This level of disentanglement and controllability is obtained thanks to a novel decomposition of the environment map representation into ambient light and an editable gaussian dominant light source. COMPOSE is a four-stage pipeline that consists of light estimation and editing, light diffusion, shadow synthesis, and finally shadow editing. We define facial shadows as the result of a dominant light source, encoded using our novel gaussian environment map representation. Utilizing an OLAT dataset, we have trained models to: (1) predict this light source representation from images, and (2) generate realistic shadows using this representation. We also demonstrate comprehensive and intuitive shadow editing with our pipeline. Through extensive quantitative and qualitative evaluations, we have demonstrated the robust capability of our system in shadow editing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 563, + 453, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 563, + 453, + 575 + ], + "spans": [ + { + "bbox": [ + 160, + 563, + 453, + 575 + ], + "type": "text", + "content": "Keywords: Face Relighting " + }, + { + "bbox": [ + 160, + 563, + 453, + 575 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 563, + 453, + 575 + ], + "type": "text", + "content": " Shadow Editing " + }, + { + "bbox": [ + 160, + 563, + 453, + 575 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 563, + 453, + 575 + ], + "type": "text", + "content": " Lighting Decomposition" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 615, + 230, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 615, + 230, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 615, + 230, + 628 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Single-image portrait relighting is a well-studied problem with steady interest from the computer vision and graphics communities. Portrait relighting has a" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 203, + 204 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 203, + 204 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 203, + 204 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 203, + 204 + ], + "type": "image", + "image_path": "e92076f726db6df6b454297cea98e599ed66f0e16c2f3ace97c12c87159ea937.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 205, + 203, + 217 + ], + "lines": [ + { + "bbox": [ + 136, + 205, + 203, + 217 + ], + "spans": [ + { + "bbox": [ + 136, + 205, + 203, + 217 + ], + "type": "text", + "content": "a) Source Image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 205, + 114, + 272, + 204 + ], + "blocks": [ + { + "bbox": [ + 205, + 114, + 272, + 204 + ], + "lines": [ + { + "bbox": [ + 205, + 114, + 272, + 204 + ], + "spans": [ + { + "bbox": [ + 205, + 114, + 272, + 204 + ], + "type": "image", + "image_path": "2dac28d700bc188aca1e09dbd6b263d9df18caee00c321e0b2622af0c39878da.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 219, + 205, + 257, + 226 + ], + "lines": [ + { + "bbox": [ + 219, + 205, + 257, + 226 + ], + "spans": [ + { + "bbox": [ + 219, + 205, + 257, + 226 + ], + "type": "text", + "content": "b) Soften \nShadow" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 273, + 114, + 340, + 204 + ], + "blocks": [ + { + "bbox": [ + 273, + 114, + 340, + 204 + ], + "lines": [ + { + "bbox": [ + 273, + 114, + 340, + 204 + ], + "spans": [ + { + "bbox": [ + 273, + 114, + 340, + 204 + ], + "type": "image", + "image_path": "7f5066a52f3298bd87b9d1c057af746bd2568e06283a4407200818ec200df3ad.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 205, + 331, + 226 + ], + "lines": [ + { + "bbox": [ + 282, + 205, + 331, + 226 + ], + "spans": [ + { + "bbox": [ + 282, + 205, + 331, + 226 + ], + "type": "text", + "content": "c) Intensify Shadow" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 229, + 482, + 275 + ], + "lines": [ + { + "bbox": [ + 131, + 229, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 131, + 229, + 482, + 275 + ], + "type": "text", + "content": "Fig. 1: Overview. COMPOSE is the first single-image portrait shadow editing method that achieves complete shadow editing control, i.e. adjusting shadow intensity, modifying light size, and changing shadow positions, all while preserving the source image's other lighting attributes (e.g. ambient light)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 342, + 114, + 410, + 204 + ], + "blocks": [ + { + "bbox": [ + 342, + 114, + 410, + 204 + ], + "lines": [ + { + "bbox": [ + 342, + 114, + 410, + 204 + ], + "spans": [ + { + "bbox": [ + 342, + 114, + 410, + 204 + ], + "type": "image", + "image_path": "0c6b65d77d59b90229177bb8ce0bf63e46af09e54217990d8e9a4a8cc1636dee.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 205, + 409, + 226 + ], + "lines": [ + { + "bbox": [ + 343, + 205, + 409, + 226 + ], + "spans": [ + { + "bbox": [ + 343, + 205, + 409, + 226 + ], + "type": "text", + "content": "d) Modify Light Size" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 410, + 114, + 479, + 204 + ], + "blocks": [ + { + "bbox": [ + 410, + 114, + 479, + 204 + ], + "lines": [ + { + "bbox": [ + 410, + 114, + 479, + 204 + ], + "spans": [ + { + "bbox": [ + 410, + 114, + 479, + 204 + ], + "type": "image", + "image_path": "ffe6fcd76a686c49eca0658ac41ba5911471a525a22c44fc90d87c8a579075ce.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 205, + 465, + 226 + ], + "lines": [ + { + "bbox": [ + 424, + 205, + 465, + 226 + ], + "spans": [ + { + "bbox": [ + 424, + 205, + 465, + 226 + ], + "type": "text", + "content": "e) Rotate Shadow" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 289, + 482, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 337 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 337 + ], + "type": "text", + "content": "wide range of applications in computational photography [31], AR/VR [4,5,19], and downstream tasks in the face domain [11,25]. While many methods for single image portrait relighting have been proposed in recent years [17,18,29,31,33,41, 49,54], they often face challenges when dealing with facial shadows." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 339, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 482, + 567 + ], + "type": "text", + "content": "One popular genre of face relighting methods [31, 41] uses low resolution HDR environment maps to represent lighting. This lighting representation is used in a conditional image translation framework to control the lighting effect of the output face image, thereby achieving facial relighting. The requirement for an environment map restricts users seeking to fine-tune lighting, such as minor modifications to shadow smoothness or direction. Additionally, a low-resolution environment map often lacks the precision needed for accurately depicting complex, high-frequency shadow effects. Conversely, several approaches use low-dimensional spherical harmonics (SH) for their lighting representation [37, 54]. These methods, based on simplified lighting and reflectance models, frequently produce unrealistic relit images. Moreover, they overlook occlusions in the image formation process, disallowing the generation and editing of cast shadows. SunStage [44] generates realistic shadows on faces but lacks generalizability and requires videos as input. Some recent works have focused on shadow editing tasks [11, 26, 53] using state-of-the-art image generators such as diffusion models. However, achieving precise control with these sophisticated generators remains a challenging task. Consequently, these methods often provide limited control in shadow synthesis, typically handling only shadow softening or completely removing the shadow, without precise editing of shadow intensity, shape, or position." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "To overcome the challenges in modeling and controlling facial shadows in portrait relighting, we introduce COMPOSE, a system that separates shadow modeling from ambient light in the lighting representation. COMPOSE enables flexible manipulation of shadow attributes, including position (lighting direction), smoothness, and intensity (See Fig. 1). The key innovation in COMPOSE is its shadow representation, which is compatible with environment map-based representations, making it easily integrable into existing relighting networks. Building on the concept of image-based lighting representations (e.g. HDR en" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "A. Hou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "vironment maps), our approach divides lighting effects into two components: shadows caused by dominant light sources and ambient lighting. We simplify the shadow component by attributing it to a single dominant light source on an environment map. This is encoded using a 2D isotropic Gaussian, with variables representing light position, standard deviation (indicating light size or diffusion), and scale (light intensity). The remaining lighting effects are attributed to a diffused environment map, modeling the image's ambient lighting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 295 + ], + "type": "text", + "content": "For a given face image, we first predict the lighting as a panorama environment map, from which we estimate the dominant light (shadow) parameters. The system then removes shadows to obtain a diffused, albedo-like image. Utilizing this diffuse image and the shadow representation, we train a network to synthesize shadows on the subject in the input image. This shadow synthesis process is adaptable, accepting various shadow-related parameters for controllable shadow synthesis, including position and shape. The synthesized shadow can be linearly blended with the diffuse lighting to create shadows of varying intensities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 296, + 482, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 355 + ], + "type": "text", + "content": "To train models for shadow diffusion and synthesis, we utilize an OLAT (one-light-at-a-time) dataset, captured through a light stage system. This dataset, combined with the environment map-based lighting representation, ensures our system's compatibility with previously proposed face relighting techniques. Our approach also uniquely offers the capability to accurately model shadows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 356, + 315, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 356, + 315, + 366 + ], + "spans": [ + { + "bbox": [ + 146, + 356, + 315, + 366 + ], + "type": "text", + "content": "Our contributions are thus as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 367, + 482, + 475 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 131, + 367, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 367, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 131, + 367, + 482, + 415 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 131, + 367, + 482, + 415 + ], + "type": "text", + "content": " We propose the first single image portrait shadow editing method with precise control over various shadow attributes, including shadow shape, intensity, and position. Our system provides users with the flexibility to control shading while also preserving the other lighting attributes of the input images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 415, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 415, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 131, + 415, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 131, + 415, + 482, + 450 + ], + "type": "text", + "content": " We propose a novel lighting decomposition into ambient light and an ed-itable dominant gaussian light source, which enables disentangled shadow editing while preserving the remaining lighting attributes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 451, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 451, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 451, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 131, + 451, + 482, + 475 + ], + "type": "text", + "content": " We achieve state-of-the-art relighting performance quantitatively and qualitatively, including on a light-stage-captured test set and in-the-wild images." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 491, + 238, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 238, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 238, + 504 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 515, + 258, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 258, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 258, + 528 + ], + "type": "text", + "content": "2.1 Portrait Relighting" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Many portrait relighting methods have been proposed in recent years that generally fall into one of two categories: focusing on reducing the data requirement and improving the relighting quality as much as possible with limited data [6, 10, 17, 18, 20, 24, 33, 34, 36-39, 43-45, 47, 49, 54] or using light stage data but modeling more complex lighting effects [4, 12, 28, 29, 31, 41, 42, 48, 50, 52]. Among single image portrait relighting methods, a substantial limitation is that they cannot preserve the original environmental lighting while editing shadows, and instead opt to relight by replacing the original lighting with a completely new environment [11], which is undesirable in many portrait editing applications. Existing methods also struggle to handle more directed lights and fail in the presence of substantial hard or intense shadows [37, 41, 54]." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 229, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 229, + 90, + 448, + 103 + ], + "type": "text", + "content": "COMPOSE: Comprehensive Portrait Shadow Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 118, + 478, + 232 + ], + "blocks": [ + { + "bbox": [ + 135, + 118, + 478, + 232 + ], + "lines": [ + { + "bbox": [ + 135, + 118, + 478, + 232 + ], + "spans": [ + { + "bbox": [ + 135, + 118, + 478, + 232 + ], + "type": "image", + "image_path": "da4be91a1b0c7392faf38906e446e5662637c181577a9027be877216a23874f3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 243, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 331 + ], + "type": "text", + "content": "Fig. 2: Method Overview. COMPOSE is a 4-stage shadow editing method consisting of single-image lighting estimation, light diffusion, shadow synthesis, and image compositing. By first estimating the dominant light position using the environment map regressed from the input image, COMPOSE can control shadow shape and intensity by controlling light spread and light intensity respectively as well as shadow position by changing the location of the dominant light. By estimating a diffuse image " + }, + { + "bbox": [ + 130, + 243, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_D" + }, + { + "bbox": [ + 130, + 243, + 482, + 331 + ], + "type": "text", + "content": " as well as a shadowed image " + }, + { + "bbox": [ + 130, + 243, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_S" + }, + { + "bbox": [ + 130, + 243, + 482, + 331 + ], + "type": "text", + "content": ", COMPOSE can generate the final edited image " + }, + { + "bbox": [ + 130, + 243, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_E" + }, + { + "bbox": [ + 130, + 243, + 482, + 331 + ], + "type": "text", + "content": " through image compositing." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 341, + 483, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 483, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 483, + 449 + ], + "type": "text", + "content": "Among portrait relighting methods that take multiple images or videos as input, the method SunStage [44] is able to maintain the lighting attributes of the input video's environment when performing relighting. However, the method must be trained using a video of a subject on a sunny day and involves an expensive optimization procedure. The model is also subject-specific and must be retrained or fine-tuned even for the same subject if they change their hairstyle, accessories, or clothing. COMPOSE has the advantage of being a single-image portrait relighting method, and is both generalizable and practical for in-the-wild photo editing applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 466, + 241, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 466, + 241, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 466, + 241, + 478 + ], + "type": "text", + "content": "2.2 Shadow Editing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 594 + ], + "type": "text", + "content": "Another line of work closely related to COMPOSE are methods that edit only facial shadows, including shadow removal [26,53] and shadow softening [11]. Most similar to our work is the method of [11], which is able to directly control the degree of shadow softening while preserving the original environmental lighting. However, unlike COMPOSE, they are restricted to only shadow softening and cannot handle other shadow editing operations like shadow intensifying, modifying shadow shape, and shadow rotation. Shadow removal methods [14,26,53] are often even more limited since they cannot control the degree of shadow softening and attempt to remove the shadow completely." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 610, + 288, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 288, + 623 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 288, + 623 + ], + "type": "text", + "content": "2.3 Lighting Representations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Existing portrait relighting methods generally utilize one of three lighting representations: Spherical Harmonics (SH) [18,33,34,37,54], dominant lighting direction [17,29], and environment maps [31,41,49]. SH lighting tends to only model" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "A. Hou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 416 + ], + "type": "text", + "content": "the lower frequencies of lighting effects, and thus is unsuitable for modeling non-Lambertian hard shadows and specularities and is better suited for diffuse lightings. Representing the light as a dominant lighting direction can often handle the non-Lambertian effects of directional and point light sources, but they often leave other lighting attributes such as light size out of the equation (e.g. Hou et al. [17]). Moreover, they often do not handle diffuse lightings well and are intended for directional lights [17, 29]. The environment map representation is more flexible and is able to model light intensity, light size, and light position and can handle both diffuse and directional lights. However, what is largely missing in this domain is a way to properly decompose the environment map that enables precise controllability over different lighting components. Existing methods only perform relighting by completely changing the scene illumination with a new environment map without disentangling the effects of ambient, diffuse, and directional light [31, 41, 49]. This can be undesirable for many computational photography applications where the user wishes to preserve the ambience of the scene and edit only the facial shadow alone [11]. Moreover, the default environment map representation is unsuitable for some shadow editing applications such as shrinking the light. While enlarging the light can be performed by applying a gaussian blur, there is no equivalent operation that can be performed to shrink the light on an environment map. Our lighting representation decomposes the environment map into the ambient light and an editable gaussian dominant light source, which is adaptable in light intensity, size, and position. COMPOSE is thus able to perform a full suite of shadow editing operations thanks to its editable gaussian component and can edit the shadow alone without disrupting other lighting attributes (e.g. ambient light)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 430, + 233, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 233, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 233, + 445 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 453, + 269, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 269, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 269, + 464 + ], + "type": "text", + "content": "3.1 Problem Formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 470, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 482, + 590 + ], + "type": "text", + "content": "In our framework, shadows are treated as a composable lighting effect, which can be independently predicted and altered based on a specific face image. This shadow representation is integrated into the overall lighting representation, and our framework is designed to (1) accurately predict the shadow from an image, and (2) apply a controllable shadow onto a shadow-free face image. We define the portrait shadow editing problem based on properties of shadows that users can manipulate, namely shadow position, shadow intensity, and shadow shape. These shadow properties are connected to the lighting attribute that is easily controllable. Specifically, given source image " + }, + { + "bbox": [ + 130, + 470, + 482, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_N" + }, + { + "bbox": [ + 130, + 470, + 482, + 590 + ], + "type": "text", + "content": " and desired edited output image " + }, + { + "bbox": [ + 130, + 470, + 482, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_E" + }, + { + "bbox": [ + 130, + 470, + 482, + 590 + ], + "type": "text", + "content": " with modified shadows, our problem can be written as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 257, + 598, + 481, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 598, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 257, + 598, + 481, + 611 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {E} = F _ {\\theta} \\left(\\mathbf {I} _ {N}, x, y, \\sigma , \\gamma\\right), \\tag {1}", + "image_path": "cc72a7046e72b3d2d138cfc3666dcc89a6d2d56540bd329e73cd47649aef989a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " is our shadow editing method, " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " are the desired light position on an environment map, " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " is the light size, and " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " is the light intensity. By editing the parameters, " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", we can control the shadow position, shape, and intensity respectively. We describe our solution COMPOSE in detail in Sec. 3.2." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 229, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 229, + 90, + 448, + 103 + ], + "type": "text", + "content": "COMPOSE: Comprehensive Portrait Shadow Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 280, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 280, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 280, + 128 + ], + "type": "text", + "content": "3.2 COMPOSE Framework" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 480, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 480, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 480, + 170 + ], + "type": "text", + "content": "COMPOSE is a 4-stage shadow editing framework, consisting of single image lighting estimation, light diffusion, shadow synthesis, and finally image compositing. See Fig. 2 for an overview of our method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 171, + 482, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 171, + 482, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 171, + 482, + 339 + ], + "type": "text", + "content": "Lighting Estimation In the lighting estimation stage, we train a variational autoencoder (VAE) [23] to estimate an environment map (LDR) from a single portrait image as its lighting representation " + }, + { + "bbox": [ + 130, + 171, + 482, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_N" + }, + { + "bbox": [ + 130, + 171, + 482, + 339 + ], + "type": "text", + "content": ". From this environment map, we then estimate the dominant light position. Naturally, the dominant light, such as the sun, represents the brightest intensity on a image-based lighting representation. Therefore we can simply obtain the dominant light position by fitting a 2D isotropic Gaussian on the predicted environment map. The center of the Gaussian will indicate the position of the dominant light on the 2D environment map. This estimated light position corresponds to the shadow information on the input image, and therefore is useful for \"in-place\" shadow editing (e.g., shadow softening, shadow intensifying) in the later stage that does not alter the direction of the dominant light. We argue this is desirable in many portrait editing applications where the user only wants to edit existing shadows and does not want to alter the ambient lighting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 341, + 482, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 425 + ], + "type": "text", + "content": "Light Diffusion In the light diffusion stage, we remove all existing hard shadows and specular highlights from the input image and output a \"diffused\" image " + }, + { + "bbox": [ + 130, + 341, + 482, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_D" + }, + { + "bbox": [ + 130, + 341, + 482, + 425 + ], + "type": "text", + "content": " with very smooth shading. This represents the input face under only an ambient illumination condition. Our light diffusion network consists of a hierarchical transformer encoder that generates multi-level features, which are fed to a decoder with transposed convolutional layers. Our network takes as input the original input image, a body parsing mask, and a binary foreground mask." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": "**Shadow Synthesis** In the shadow synthesis stage, we aim to produce a shadowed image " + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_S" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": " illuminated by an edited light source given the diffuse image " + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_D" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": " estimated from the light diffusion stage and an edited environment map. Using the estimated dominant light position from the light estimation stage, we have full control over light attributes such as light size, intensity, and the dominant light position. We can produce a new environment map using a Gaussian to represent the dominant light, where the light size is modeled by the standard deviation of the Gaussian (a larger standard deviation represents a larger, more diffuse light) and the light intensity can be adjusted by multiplying the Gaussian with a scalar. This allows our method to control shadow shape, light intensity, and position. To regress the shadowed image " + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_S" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": ", we propose a two-stage architecture: a U-Net followed by a conditional DDPM (See Fig. 3). For the first stage, we adopt the U-Net architecture of [41]. During training, instead of using a reshaped environment map-like representation for the dominant light, we design a feature map-like representation that's parameterized by the shadow attributes. Specifically, the lighting representation for shadows has 4 parameters. The first two parameters " + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": " encode the coordinates of the light center on the environment map, the third parameter " + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": " encodes the light size, and the fourth parameter " + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": " encodes the light intensity. All four parameters are normalized between 0 and 1 and repeated spatially as " + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": " channels to be suitable" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "A. Hou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 118, + 476, + 199 + ], + "blocks": [ + { + "bbox": [ + 142, + 118, + 476, + 199 + ], + "lines": [ + { + "bbox": [ + 142, + 118, + 476, + 199 + ], + "spans": [ + { + "bbox": [ + 142, + 118, + 476, + 199 + ], + "type": "image", + "image_path": "da5d8b375587fd13bd1aaa8eea2f8d3ada2444b94e726b4b7738256e4d2c6071.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 201, + 482, + 258 + ], + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 258 + ], + "type": "text", + "content": "Fig. 3: Shadow Synthesis Pipeline. Our shadow synthesis step consists of two stages: a U-Net followed by a conditional DDPM. The U-Net takes ambient image " + }, + { + "bbox": [ + 130, + 201, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_D" + }, + { + "bbox": [ + 130, + 201, + 482, + 258 + ], + "type": "text", + "content": " and environment map " + }, + { + "bbox": [ + 130, + 201, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_T" + }, + { + "bbox": [ + 130, + 201, + 482, + 258 + ], + "type": "text", + "content": " as input and outputs the coarse relit image " + }, + { + "bbox": [ + 130, + 201, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_U" + }, + { + "bbox": [ + 130, + 201, + 482, + 258 + ], + "type": "text", + "content": ". With " + }, + { + "bbox": [ + 130, + 201, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_U" + }, + { + "bbox": [ + 130, + 201, + 482, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 201, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_T" + }, + { + "bbox": [ + 130, + 201, + 482, + 258 + ], + "type": "text", + "content": " as spatial conditions, the conditional DDPM outputs the final shadowed image " + }, + { + "bbox": [ + 130, + 201, + 482, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_S" + }, + { + "bbox": [ + 130, + 201, + 482, + 258 + ], + "type": "text", + "content": ", refining shadow boundaries and improving image quality." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 265, + 482, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 482, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 482, + 409 + ], + "type": "text", + "content": "inputs for the U-Net. We find that this lighting representation leads to faster convergence, especially when the desired light mimics a point light and occupies a tiny portion of the environment map. We feed the diffuse image " + }, + { + "bbox": [ + 130, + 265, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_D" + }, + { + "bbox": [ + 130, + 265, + 482, + 409 + ], + "type": "text", + "content": " and our desired lighting feature map to the U-Net, and the output is a relit image " + }, + { + "bbox": [ + 130, + 265, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_U" + }, + { + "bbox": [ + 130, + 265, + 482, + 409 + ], + "type": "text", + "content": ". While the U-Net architecture has merits such as complete geometric consistency for shadow synthesis, one demerit is that the synthesized shadow boundaries are often not very sharp and the overall image quality could be improved. We thus employ a DDPM [7,15] as the second stage of our shadow synthesis model. The role of the DDPM here is to take the output image of the U-Net " + }, + { + "bbox": [ + 130, + 265, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_U" + }, + { + "bbox": [ + 130, + 265, + 482, + 409 + ], + "type": "text", + "content": " as a spatial condition along with the lighting parameters and perform image refinement to generate the final shadowed image " + }, + { + "bbox": [ + 130, + 265, + 482, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_S" + }, + { + "bbox": [ + 130, + 265, + 482, + 409 + ], + "type": "text", + "content": ". The training objective of our shadow synthesis DDPM thus becomes:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 222, + 413, + 481, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 413, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 222, + 413, + 481, + 434 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {t, \\mathbf {x} _ {0}, \\boldsymbol {\\epsilon}} \\left[ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\mathbf {x} _ {t}, t, \\mathbf {I} _ {U}, x, y, \\sigma , \\gamma\\right) \\| ^ {2} \\right]. \\tag {2}", + "image_path": "7efaf1b7f948c107e8e835858e6141b73f20574c35490ef03d3c7f0054bec2e5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 482, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 497 + ], + "type": "text", + "content": "Our condition " + }, + { + "bbox": [ + 130, + 437, + 482, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_U" + }, + { + "bbox": [ + 130, + 437, + 482, + 497 + ], + "type": "text", + "content": " is spatially concatenated with " + }, + { + "bbox": [ + 130, + 437, + 482, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 130, + 437, + 482, + 497 + ], + "type": "text", + "content": " and our lighting parameters are repeated spatially as channels of the same resolution and similarly spatially concatenated, where " + }, + { + "bbox": [ + 130, + 437, + 482, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t}\\epsilon" + }, + { + "bbox": [ + 130, + 437, + 482, + 497 + ], + "type": "text", + "content": ". We find that the quality of the edited images is noticeably enhanced by the DDPM compared to the U-Net alone and that adding the DDPM sharpens the shadow boundaries." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 547 + ], + "type": "text", + "content": "Image Compositing With " + }, + { + "bbox": [ + 130, + 498, + 482, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_D" + }, + { + "bbox": [ + 130, + 498, + 482, + 547 + ], + "type": "text", + "content": " from our light diffusion stage and " + }, + { + "bbox": [ + 130, + 498, + 482, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_S" + }, + { + "bbox": [ + 130, + 498, + 482, + 547 + ], + "type": "text", + "content": " from our shadow synthesis stage, we can then perform image compositing between " + }, + { + "bbox": [ + 130, + 498, + 482, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_D" + }, + { + "bbox": [ + 130, + 498, + 482, + 547 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 498, + 482, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_S" + }, + { + "bbox": [ + 130, + 498, + 482, + 547 + ], + "type": "text", + "content": " to achieve our final edited image " + }, + { + "bbox": [ + 130, + 498, + 482, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_E" + }, + { + "bbox": [ + 130, + 498, + 482, + 547 + ], + "type": "text", + "content": ", by assigning weights " + }, + { + "bbox": [ + 130, + 498, + 482, + 547 + ], + "type": "inline_equation", + "content": "\\omega_{D}" + }, + { + "bbox": [ + 130, + 498, + 482, + 547 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 498, + 482, + 547 + ], + "type": "inline_equation", + "content": "\\omega_{S}" + }, + { + "bbox": [ + 130, + 498, + 482, + 547 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 498, + 482, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_D" + }, + { + "bbox": [ + 130, + 498, + 482, + 547 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 498, + 482, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_S" + }, + { + "bbox": [ + 130, + 498, + 482, + 547 + ], + "type": "text", + "content": " respectively. The final edited image " + }, + { + "bbox": [ + 130, + 498, + 482, + 547 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_E" + }, + { + "bbox": [ + 130, + 498, + 482, + 547 + ], + "type": "text", + "content": " is thus generated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 261, + 554, + 481, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 554, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 261, + 554, + 481, + 567 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {E} = \\omega_ {D} \\mathbf {I} _ {D} + \\omega_ {S} \\mathbf {I} _ {S}, \\tag {3}", + "image_path": "19c1c0a2f6e6a048cf1db526ab9df370e9540eef0ca6544bdebcc708d9d29977.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 574, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 482, + 598 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 574, + 482, + 598 + ], + "type": "inline_equation", + "content": "\\omega_{D} + \\omega_{S} = 1" + }, + { + "bbox": [ + 130, + 574, + 482, + 598 + ], + "type": "text", + "content": ". By adjusting " + }, + { + "bbox": [ + 130, + 574, + 482, + 598 + ], + "type": "inline_equation", + "content": "\\omega_{D}" + }, + { + "bbox": [ + 130, + 574, + 482, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 574, + 482, + 598 + ], + "type": "inline_equation", + "content": "\\omega_{S}" + }, + { + "bbox": [ + 130, + 574, + 482, + 598 + ], + "type": "text", + "content": ", we can further tune shadow intensity: a higher " + }, + { + "bbox": [ + 130, + 574, + 482, + 598 + ], + "type": "inline_equation", + "content": "\\omega_{D}" + }, + { + "bbox": [ + 130, + 574, + 482, + 598 + ], + "type": "text", + "content": " softens shadows and a higher " + }, + { + "bbox": [ + 130, + 574, + 482, + 598 + ], + "type": "inline_equation", + "content": "\\omega_{S}" + }, + { + "bbox": [ + 130, + 574, + 482, + 598 + ], + "type": "text", + "content": " results in darker shadows." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 613, + 236, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 613, + 236, + 624 + ], + "spans": [ + { + "bbox": [ + 131, + 613, + 236, + 624 + ], + "type": "text", + "content": "3.3 Loss Functions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "To train our single image lighting estimation network, we apply two loss functions commonly used in VAE training [23]: " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{recon}}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KLD}}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{recon}}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " is a reconstruction loss between the predicted environment map " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_P" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " and the groundtruth" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 229, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 229, + 90, + 448, + 103 + ], + "type": "text", + "content": "COMPOSE: Comprehensive Portrait Shadow Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 312, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 312, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 312, + 129 + ], + "type": "text", + "content": "environment map " + }, + { + "bbox": [ + 132, + 116, + 312, + 129 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_G" + }, + { + "bbox": [ + 132, + 116, + 312, + 129 + ], + "type": "text", + "content": " defined as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 241, + 137, + 481, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 137, + 481, + 159 + ], + "spans": [ + { + "bbox": [ + 241, + 137, + 481, + 159 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c o n}} = \\frac {1}{3 H W} \\left(\\left\\| \\mathbf {E} _ {P} - \\mathbf {E} _ {G} \\right\\| _ {2} ^ {2}\\right), \\tag {4}", + "image_path": "5d00a37ad772e8f6590f3125d017084603591e35d2dd8c4692c7fe5a2ce6e012.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 166, + 480, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 166, + 480, + 190 + ], + "spans": [ + { + "bbox": [ + 131, + 166, + 480, + 190 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 166, + 480, + 190 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 131, + 166, + 480, + 190 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 166, + 480, + 190 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 131, + 166, + 480, + 190 + ], + "type": "text", + "content": " are the size of the environment maps. The Kullback-Leibler divergence (KLD) loss " + }, + { + "bbox": [ + 131, + 166, + 480, + 190 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KLD}}" + }, + { + "bbox": [ + 131, + 166, + 480, + 190 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 216, + 200, + 481, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 200, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 216, + 200, + 481, + 232 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {K L D}} = - \\frac {1}{2 N} \\sum_ {i = 1} ^ {N} \\left(1 + \\log \\left(\\sigma_ {i} ^ {2}\\right) - \\mu_ {i} ^ {2} - \\sigma_ {i} ^ {2}\\right), \\tag {5}", + "image_path": "b57d793c9f1d96a0cd9bc0e334254f0a304f6c147c3d81efcf4dd642f84f1a8e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 240, + 480, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 240, + 480, + 264 + ], + "spans": [ + { + "bbox": [ + 131, + 240, + 480, + 264 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 240, + 480, + 264 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 240, + 480, + 264 + ], + "type": "text", + "content": " is the dimensionality of the latent vector, " + }, + { + "bbox": [ + 131, + 240, + 480, + 264 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 131, + 240, + 480, + 264 + ], + "type": "text", + "content": " is the batch mean, and " + }, + { + "bbox": [ + 131, + 240, + 480, + 264 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 131, + 240, + 480, + 264 + ], + "type": "text", + "content": " is the batch variance. The final loss for our lighting estimation network is thus:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 249, + 274, + 481, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 274, + 481, + 287 + ], + "spans": [ + { + "bbox": [ + 249, + 274, + 481, + 287 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {L E}} = \\lambda_ {1} \\mathcal {L} _ {\\text {r e c o n}} + \\lambda_ {2} \\mathcal {L} _ {\\mathrm {K L D}}, \\tag {6}", + "image_path": "e7003d86544076e019290fe0e11469dbdc46a306044fb21cb5df8861e0c3cd20.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 295, + 288, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 295, + 288, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 295, + 288, + 307 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 295, + 288, + 307 + ], + "type": "inline_equation", + "content": "\\lambda_1 = 1" + }, + { + "bbox": [ + 132, + 295, + 288, + 307 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 295, + 288, + 307 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 2.5\\times 10^{-4}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 146, + 308, + 485, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 308, + 485, + 320 + ], + "spans": [ + { + "bbox": [ + 146, + 308, + 485, + 320 + ], + "type": "text", + "content": "Our light diffusion network is also trained with two losses " + }, + { + "bbox": [ + 146, + 308, + 485, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{recon}}" + }, + { + "bbox": [ + 146, + 308, + 485, + 320 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 146, + 308, + 485, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{perceptual}}" + }, + { + "bbox": [ + 146, + 308, + 485, + 320 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 244, + 328, + 481, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 328, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 244, + 328, + 481, + 350 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c o n}} = \\frac {1}{3 H W} \\left(\\left\\| \\mathbf {I} _ {D} - \\mathbf {I} _ {G} \\right\\| _ {1}\\right), \\tag {7}", + "image_path": "d0e3acfedb8dae69c59c0b90e748dd535d57236b538ae9497ef1d212e6facd31.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 354, + 481, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 354, + 481, + 401 + ], + "spans": [ + { + "bbox": [ + 131, + 354, + 481, + 401 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 354, + 481, + 401 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 131, + 354, + 481, + 401 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 354, + 481, + 401 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 131, + 354, + 481, + 401 + ], + "type": "text", + "content": " are the size of the training images, " + }, + { + "bbox": [ + 131, + 354, + 481, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_D" + }, + { + "bbox": [ + 131, + 354, + 481, + 401 + ], + "type": "text", + "content": " is the predicted diffuse image, and " + }, + { + "bbox": [ + 131, + 354, + 481, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_G" + }, + { + "bbox": [ + 131, + 354, + 481, + 401 + ], + "type": "text", + "content": " is the groundtruth diffuse image. " + }, + { + "bbox": [ + 131, + 354, + 481, + 401 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{perceptual}}" + }, + { + "bbox": [ + 131, + 354, + 481, + 401 + ], + "type": "text", + "content": " enforces visual similarity [51] and is computed as the distance between VGG [40] features computed for " + }, + { + "bbox": [ + 131, + 354, + 481, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_D" + }, + { + "bbox": [ + 131, + 354, + 481, + 401 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 354, + 481, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_G" + }, + { + "bbox": [ + 131, + 354, + 481, + 401 + ], + "type": "text", + "content": ". The final loss for the light diffusion network is thus:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 249, + 412, + 481, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 412, + 481, + 425 + ], + "spans": [ + { + "bbox": [ + 249, + 412, + 481, + 425 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {L D}} = \\mathcal {L} _ {\\text {r e c o n}} + \\mathcal {L} _ {\\text {p e r c e p t u a l}}. \\tag {8}", + "image_path": "a383f492ed72fb66e266097f7927396e3a54ae73b08daba05cfcfd49e25355b0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 433, + 481, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 433, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 131, + 433, + 481, + 468 + ], + "type": "text", + "content": "When training the U-Net of our shadow synthesis pipeline, we employ a single reconstruction loss " + }, + { + "bbox": [ + 131, + 433, + 481, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{recon}}" + }, + { + "bbox": [ + 131, + 433, + 481, + 468 + ], + "type": "text", + "content": " between the predicted relit image " + }, + { + "bbox": [ + 131, + 433, + 481, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_S" + }, + { + "bbox": [ + 131, + 433, + 481, + 468 + ], + "type": "text", + "content": " and the groundtruth " + }, + { + "bbox": [ + 131, + 433, + 481, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_G" + }, + { + "bbox": [ + 131, + 433, + 481, + 468 + ], + "type": "text", + "content": ", defined as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 244, + 477, + 481, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 477, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 244, + 477, + 481, + 500 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c o n}} = \\frac {1}{3 H W} \\left(\\| \\mathbf {I} _ {S} - \\mathbf {I} _ {G} \\| _ {2} ^ {2}\\right). \\tag {9}", + "image_path": "36621bed070c63002badde9337f3b8acdec7da8c8c96289df9a964ac0276b3f1.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 514, + 322, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 322, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 322, + 525 + ], + "type": "text", + "content": "3.4 Data Collection and Generation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 533, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 533, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 131, + 533, + 481, + 593 + ], + "type": "text", + "content": "To train out networks, we use a light stage to capture one-light-at-a-time (OLAT) images of 107 subjects with diverse skin tones, body poses, expressions, and accessories using a light stage [9] with 160 lights. Each subject is captured under 10-20 sessions with varied body poses, expressions, and accessories. Each session is captured by 4 cameras with varying camera poses." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 594, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 594, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 594, + 481, + 666 + ], + "type": "text", + "content": "To generate input images to train our lighting estimation VAE and our light diffusion network, we render our OLAT images using a diverse set of environment maps collected from the Polyhaven and Laval Outdoor HDR datasets [16]. During rendering, we apply random augmentations to the environment map including rotations and adjusting intensity to improve our model's generalization to different lighting. The groundtruth " + }, + { + "bbox": [ + 131, + 594, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_G" + }, + { + "bbox": [ + 131, + 594, + 481, + 666 + ], + "type": "text", + "content": " for our lighting estimation VAE is" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "A. Hou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "simply the environment map used for rendering, and the groundtruth " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_G" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " for the light diffusion network is generated by rendering OLAT images with heavily blurred, diffuse environment maps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 236 + ], + "type": "text", + "content": "To generate images for our shadow synthesis stage, we render groundtruth relit images by using environment maps consisting of Gaussian lights with various positions, intensities, and sizes (See Sec. 3.2) along with our OLAT images. As the input to the shadow synthesis stage is meant to be the diffuse image " + }, + { + "bbox": [ + 130, + 152, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_D" + }, + { + "bbox": [ + 130, + 152, + 482, + 236 + ], + "type": "text", + "content": " predicted by the light diffusion stage, we generate diffuse images for training our shadow synthesis networks by rendering OLAT images with heavily blurred environment maps in the Polyhaven and Laval Outdoor HDR datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 251, + 279, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 251, + 279, + 262 + ], + "spans": [ + { + "bbox": [ + 131, + 251, + 279, + 262 + ], + "type": "text", + "content": "3.5 Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 267, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 482, + 399 + ], + "type": "text", + "content": "We implement all components of COMPOSE using PyTorch [32]. When training our lighting estimation VAE, we use a batch size of 128 and a learning rate of " + }, + { + "bbox": [ + 130, + 267, + 482, + 399 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 130, + 267, + 482, + 399 + ], + "type": "text", + "content": " on 4 24GB A10G GPUs. For our light diffusion network, we train with a batch size of 88 and a learning rate of " + }, + { + "bbox": [ + 130, + 267, + 482, + 399 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 130, + 267, + 482, + 399 + ], + "type": "text", + "content": " using 8 A100 GPUs. For our shadow synthesis U-Net, we train with a batch size of 100 and a learning rate of " + }, + { + "bbox": [ + 130, + 267, + 482, + 399 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 130, + 267, + 482, + 399 + ], + "type": "text", + "content": " using 4 24GB A10G GPUs. For all networks, we train using the Adam Optimizer [22]. The image resolution used to train the lighting estimation and shadow synthesis networks is " + }, + { + "bbox": [ + 130, + 267, + 482, + 399 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 267, + 482, + 399 + ], + "type": "text", + "content": " and the resolution used to train the light diffusion network is " + }, + { + "bbox": [ + 130, + 267, + 482, + 399 + ], + "type": "inline_equation", + "content": "768 \\times 768" + }, + { + "bbox": [ + 130, + 267, + 482, + 399 + ], + "type": "text", + "content": ". Please see the supplementary materials for architectural details of the lighting estimation network, the lighting diffusion network, and the shadow synthesis network." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 415, + 230, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 415, + 230, + 428 + ], + "spans": [ + { + "bbox": [ + 131, + 415, + 230, + 428 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 437, + 205, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 437, + 205, + 448 + ], + "spans": [ + { + "bbox": [ + 131, + 437, + 205, + 448 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 454, + 481, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 454, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 454, + 481, + 502 + ], + "type": "text", + "content": "To train our lighting estimation and shadow synthesis networks, we generate our training sets using light stage images and 521 outdoor environment maps. Our training set consists of 82 subjects with significant variation in factors such as skin tones, pose, expressions, and accessories." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 502, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 597 + ], + "type": "text", + "content": "For evaluation, we use two separate evaluation sets: the first is generated using our light stage images and has corresponding groundtruth, and the second consists of in-the-wild images that represent more realistic application settings and do not have groundtruth. Our light stage evaluation set consists of 48 images from a total of 16 subjects with diverse skin tones, body poses, expressions, and accessories. All evaluation subjects and lightings are held out and unseen during training. Our in-the-wild evaluation set also consists of unseen test subjects and applies unseen lighting conditions in all qualitative results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 612, + 298, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 612, + 298, + 624 + ], + "spans": [ + { + "bbox": [ + 131, + 612, + 298, + 624 + ], + "type": "text", + "content": "4.2 Shadow Editing Evaluation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "We evaluate our shadow editing performance (synthesis and removal) by using randomly sampled, unseen lighting positions as the target lights to generate relit light stage images of our 16 test subjects, which serve as groundtruth." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 229, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 229, + 91, + 448, + 103 + ], + "type": "text", + "content": "COMPOSE: Comprehensive Portrait Shadow Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 114, + 455, + 399 + ], + "blocks": [ + { + "bbox": [ + 161, + 114, + 455, + 399 + ], + "lines": [ + { + "bbox": [ + 161, + 114, + 455, + 399 + ], + "spans": [ + { + "bbox": [ + 161, + 114, + 455, + 399 + ], + "type": "image", + "image_path": "4ff79da6fd3733f1cbc65c51d4e3adacfff9cb73a3e0e397cb3c7c011be1bc5e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 400, + 482, + 434 + ], + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 434 + ], + "type": "text", + "content": "Fig. 4: Shadow Synthesis. Our model synthesizes more plausible shadows than the baselines, and, unlike [17], is also able to properly remove shadows from the source image before relighting. Environment maps are shown to help visualize each test lighting." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 142, + 493, + 477, + 531 + ], + "blocks": [ + { + "bbox": [ + 130, + 444, + 482, + 489 + ], + "lines": [ + { + "bbox": [ + 130, + 444, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 482, + 489 + ], + "type": "text", + "content": "Table 1: Shadow Editing Performance. We evaluate shadow editing using unseen lightings with varied intensity and size on 16 held out highly diverse light stage test subjects with groundtruth relit images. COMPOSE outperforms all baselines quantitatively in shadow editing performance across all metrics (mean ± standard deviation)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 142, + 493, + 477, + 531 + ], + "lines": [ + { + "bbox": [ + 142, + 493, + 477, + 531 + ], + "spans": [ + { + "bbox": [ + 142, + 493, + 477, + 531 + ], + "type": "table", + "html": "
MAEMSESSIMLPIPS
Hou et al. [17]0.1172 ± 0.07540.0438 ± 0.05030.7059 ± 0.10610.2371 ± 0.0871
Total Relighting [31]0.1008 ± 0.07430.0379 ± 0.04470.7754 ± 0.07380.2120 ± 0.0572
COMPOSE (Ours)0.0965 ± 0.06860.0349 ± 0.03680.7780 ± 0.08260.1973 ± 0.0678
", + "image_path": "4113e79ba13721c6b58c266f87dd7f4f62e3bfdc9dae20a429086c3d8d6b10b3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 544, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 605 + ], + "type": "text", + "content": "Each input image is rendered with a randomly selected environment map out of 125 unseen outdoor environment maps and is first passed to our deshading network to generate diffuse image " + }, + { + "bbox": [ + 130, + 544, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_D" + }, + { + "bbox": [ + 130, + 544, + 482, + 605 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 544, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_D" + }, + { + "bbox": [ + 130, + 544, + 482, + 605 + ], + "type": "text", + "content": " and target environment map " + }, + { + "bbox": [ + 130, + 544, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_T" + }, + { + "bbox": [ + 130, + 544, + 482, + 605 + ], + "type": "text", + "content": " are then fed to our two-stage shadow synthesis pipeline to generate the newly shadowed image " + }, + { + "bbox": [ + 130, + 544, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_S" + }, + { + "bbox": [ + 130, + 544, + 482, + 605 + ], + "type": "text", + "content": ", which is evaluated against the groundtruth light stage image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Tab. 1 compares our shadow editing performance against prior relighting methods. For both baselines [17, 31], the test results were provided by the authors. Our method achieves state-of-the-art results on all metrics (MAE, MSE, SSIM [46], and LPIPS [51]). As seen in Fig. 4, our model can synthesize appropriate shadows for various light positions. The method of Hou et al. [17] cannot" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "A. Hou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 114, + 452, + 430 + ], + "blocks": [ + { + "bbox": [ + 161, + 114, + 452, + 430 + ], + "lines": [ + { + "bbox": [ + 161, + 114, + 452, + 430 + ], + "spans": [ + { + "bbox": [ + 161, + 114, + 452, + 430 + ], + "type": "image", + "image_path": "a4aac90790bd26aafb1f059b02033460c6695eb59b2d464b475866ddebbdd7c0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 434, + 480, + 467 + ], + "lines": [ + { + "bbox": [ + 130, + 434, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 480, + 467 + ], + "type": "text", + "content": "Fig. 5: Shadow Editing. Our method achieves complete shadow editing control, including softening/intensifying shadows, changing light size to alter shape and intensity, and changing light position, all while preserving the source image's ambient light." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 479, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 598 + ], + "type": "text", + "content": "remove existing shadows in the source image and the shadow traces carry over as artifacts to the relit images. In addition, Hou et al. [17] only models the lighting direction and does not model the light size as a parameter, which leads to inaccurate shadow shape when the light size is varied. Total Relighting [31] is often unable to synthesize physically plausible shadows, and will sometimes overshadow (Fig. 4, rows 1, 3, and 4) or undershadow the image (Fig. 4, row 2). Moreover, their shadows are often blurry and not as sharp as our method. Compared to the baselines, COMPOSE can properly remove existing shadows from the source image and synthesize geometrically plausible and realistic shadows for a wide variety of lighting conditions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 609, + 350, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 350, + 620 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 350, + 620 + ], + "type": "text", + "content": "4.3 Portrait Shadow Editing Applications" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "We demonstrate the full performance of COMPOSE by performing all forms of shadow editing, including shadow softening, shadow intensifying, modifying light size (which changes both shadow shape and intensity), and shadow rotation" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 230, + 91, + 448, + 102 + ], + "type": "text", + "content": "COMPOSE: Comprehensive Portrait Shadow Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 114, + 452, + 339 + ], + "blocks": [ + { + "bbox": [ + 161, + 114, + 452, + 339 + ], + "lines": [ + { + "bbox": [ + 161, + 114, + 452, + 339 + ], + "spans": [ + { + "bbox": [ + 161, + 114, + 452, + 339 + ], + "type": "image", + "image_path": "ac06b94f709f8530bd8dd5e9253be67ce2f592f0927ecc3dff4782577d449afd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 342, + 482, + 376 + ], + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 376 + ], + "type": "text", + "content": "Fig.6: Degree of Shadow Editing. We demonstrate COMPOSE's controllable shadow editing by varying shadow intensity (top row), light diffusion for modifying shadow shape and intensity (middle row), and light position (bottom row)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "text", + "content": "by changing the dominant light position. For each image, we first estimate the diffuse image " + }, + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_D" + }, + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "text", + "content": " and the shadowed image " + }, + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_S" + }, + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "text", + "content": " and perform image compositing between these two images to generate the final edited image " + }, + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_E" + }, + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "text", + "content": ". As seen in Fig. 5, COMPOSE properly edits the shadows of in-the-wild images. To soften or intensify the shadow without changing the shadow shape, we rely on our image compositing coefficients " + }, + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\omega_d" + }, + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\omega_s" + }, + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "text", + "content": ", which control the contributions of " + }, + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_D" + }, + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_S" + }, + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "text", + "content": " respectively. We increase the weight of " + }, + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\omega_d" + }, + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "text", + "content": " to soften the shadow and increase the weight of " + }, + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\omega_s" + }, + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "text", + "content": " to strengthen the shadow. We further show that we can modify light size by tuning the Gaussian spread " + }, + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "text", + "content": ", which changes both shadow shape and intensity. Enlarging the light makes the shadow more diffuse and less intense, whereas shrinking the light generates a harder, more intense shadow. For the subjects in rows 2, 3, and 4 we enlarged the light, resulting in a more diffuse shadow especially noticeable around the nose. For the subject in row 1 we instead shrunk the light and appropriately produce an edited image with more prominent and intense hard shadows. Finally, we also demonstrate shadow rotation by changing the dominant light position in " + }, + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_S" + }, + { + "bbox": [ + 130, + 389, + 482, + 581 + ], + "type": "text", + "content": "'s environment map." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "To demonstrate COMPOSE can properly control shadow editing, we visualize various degrees of shadow softening/intensifying, light diffusion, and shadow rotation in Fig. 6. The first row shows varying shadow intensity, where the shadows become less intense from left to right using " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\omega_{d}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\omega_{s}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ". Our results demonstrate that these flexible weights allow for a wide range of shadow intensities. The second row shows increasing degrees of light diffusion from left to right, which corresponds to increasing the light size. The shadows in the generated images" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "A. Hou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 160, + 114, + 452, + 353 + ], + "blocks": [ + { + "bbox": [ + 160, + 114, + 452, + 353 + ], + "lines": [ + { + "bbox": [ + 160, + 114, + 452, + 353 + ], + "spans": [ + { + "bbox": [ + 160, + 114, + 452, + 353 + ], + "type": "image", + "image_path": "ffaadec496b1350ecde03d9439a9a2705103120165e8fd6b6f3d5d3789ba7e70.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 354, + 482, + 387 + ], + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 387 + ], + "type": "text", + "content": "Fig. 7: Controllable Shadow Softening. For each subject, we show the shadow softening results of [11] in the first row and COMPOSE in the second. COMPOSE leaves less shadow traces at level 0.0, where the goal is to remove the shadows completely." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 399, + 480, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 480, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 480, + 448 + ], + "type": "text", + "content": "correctly transition from harder shadows to more diffuse shadows as the light size increases. The third row shows shadow rotation, where the light is rotated horizontally along the environment map and around the subject. COMPOSE is able to produce physically plausible shadows as the light position changes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 448, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 533 + ], + "type": "text", + "content": "It is important to note that this level of controllable shadow editing is not achievable by existing portrait shadow editing methods [11, 53] that are only able to handle shadow softening or complete shadow removal as well as existing portrait relighting methods that completely modify the existing environment and cannot preserve the original scene illumination [31], do not properly model all shadow attributes (e.g. intensity, shape, and position) [17, 33] or are not designed to handle lighting conditions with more substantial shadows [37, 41, 54]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 539, + 315, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 539, + 315, + 552 + ], + "spans": [ + { + "bbox": [ + 131, + 539, + 315, + 552 + ], + "type": "text", + "content": "4.4 Shadow Softening Comparison" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "We compare on in-the-wild shadow softening with the state-of-the-art portrait shadow softening method Futschik et al. [11]. Using results provided by the authors, we visualize different degrees of shadow softening for [11] by varying the light diffusion parameter (0.75, 0.5, 0.2, and 0.0) and compare with equivalent degrees of shadow softening by COMPOSE, as shown in Fig. 7. Higher parameter values correspond to less softening whereas lower values correspond to more softening. At 0.0, the goal is to completely remove the source image's shadows. Across all degrees of light diffusion, both methods gradually soften the shadow, but COMPOSE leaves less shadow traces than [11] when completely removing" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 229, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 229, + 90, + 448, + 102 + ], + "type": "text", + "content": "COMPOSE: Comprehensive Portrait Shadow Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "the source image's shadows at level 0.0, especially for images with darker shadows. This is thanks to our hierarchical transformer design for our light diffusion network, which better handles removing the effects of shadows at all scales." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 169, + 220, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 169, + 220, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 169, + 220, + 182 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 482, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 338 + ], + "type": "text", + "content": "We have proposed COMPOSE: the first single image portrait shadow editing method that can perform comprehensive variations of shadow editing including editing shadow intensity, light size, and shadow position while preserving the source image's lighting environment. This is largely enabled by our novel lighting decomposition into ambient light and an editable gaussian dominant light, which enables new applications like shrinking the light and intensifying shadows as well as disentangled shadow editing that preserves the scene ambience. We have demonstrated the novel photo editing applications enabled by our method over prior work qualitatively and that our method achieves state-of-the-art shadow editing performance quantitatively over prior relighting methods. We hope that our work can inspire and motivate more research in the exciting new research area of controllable and comprehensive portrait shadow editing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 340, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 555 + ], + "type": "text", + "content": "Limitations. Good directions for future work include handling lighting environments with multiple light sources and indoor environments, as our framework is intended mostly for outdoor settings, where the sun is the single dominant light. In addition, while " + }, + { + "bbox": [ + 130, + 340, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_D" + }, + { + "bbox": [ + 130, + 340, + 482, + 555 + ], + "type": "text", + "content": " can preserve the ambience, we still notice there are sometimes color shifts with the environmental lighting during shadow editing, largely due to the compositing step with shadowed image " + }, + { + "bbox": [ + 130, + 340, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_S" + }, + { + "bbox": [ + 130, + 340, + 482, + 555 + ], + "type": "text", + "content": " (which adopts the light stage's environmental lighting). This could be alleviated by modeling the diffuse lighting tint similar to [11] or by training with white balanced data to remove the color bias of the light stage. We also believe that one can continue to improve the generalizability of our model. Similar to other relighting methods that rely on light stage data [11, 31, 41], we train with a limited number of subjects. Future work may incorporate synthetic data to expand the training set or adopt latent diffusion models, which naturally have high generalizability [35]. Finally, one can further improve the identity preservation of our method. We find that high frequency details are sometimes lost at the initial U-Net stage of shadow synthesis, and ways to better preserve the identity include passing " + }, + { + "bbox": [ + 130, + 340, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_D" + }, + { + "bbox": [ + 130, + 340, + 482, + 555 + ], + "type": "text", + "content": " or the input image " + }, + { + "bbox": [ + 130, + 340, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_N" + }, + { + "bbox": [ + 130, + 340, + 482, + 555 + ], + "type": "text", + "content": ", which contain high frequency details, as additional conditions to the DDPM or to use stronger backbones than [41]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "Potential Societal Impacts. In image editing, there are always concerns of generating malicious deepfakes. However, we only alter illumination, not subject identity. Even so, there may be concerns that altering the illumination and especially adding dark shadows to faces could be used as a malicious attack to worsen the performance of downstream tasks such as face recognition. Users can detect these attacks using existing deepfake detection methods [1-3,8,13,21,27,30] or by training one using the edited images generated by COMPOSE. State-of-the-art shadow removal methods [26,53] are also an option to counter these potential attacks since they can remove any facial shadows that COMPOSE synthesizes." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "A. Hou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "text", + "content": "1. Aghasanli, A., Kangin, D., Angelov, P.: Interpretable-through-prototypes deepfake detection for diffusion models. In: ICCVW (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "type": "text", + "content": "2. Asnani, V., Yin, X., Hassner, T., Liu, X.: MaLP: Manipulation localization using a proactive scheme. In: CVPR (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 184, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 480, + 205 + ], + "type": "text", + "content": "3. Asnani, V., Yin, X., Hassner, T., Liu, X.: Reverse engineering of generative models: Inferring model hyperparameters from generated images. PAMI (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 206, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 480, + 237 + ], + "type": "text", + "content": "4. Bi, S., Lombardi, S., Saito, S., Simon, T., Wei, S.E., McPhail, K., Ramamoorthi, R., Sheikh, Y., Saragih, J.: Deep relightable appearance models for animatable faces. TOG (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 238, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 238, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 238, + 480, + 259 + ], + "type": "text", + "content": "5. Caselles, P., Ramon, E., Garcia, J., i Nieto, X.G., Moreno-Noguer, F., Triginer, G.: SIRA: Relightable avatars from a single image. In: WACV (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 261, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 480, + 282 + ], + "type": "text", + "content": "6. Chandran, S., Hold-Geoffroy, Y., Sunkavalli, K., Shu, Z., Jayasuriya, S.: Temporally consistent relighting for portrait videos. In: WACV (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 283, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 283, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 283, + 480, + 304 + ], + "type": "text", + "content": "7. Choi, J., Kim, S., Jeong, Y., Gwon, Y., Yoon, S.: ILVR: Conditioning method for denoising diffusion probabilistic models. In: ICCV (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 304, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 326 + ], + "type": "text", + "content": "8. Corvi, R., Cozzolino, D., Zingarini, G., Poggi, G., Nagano, K., Verdoliva, L.: On the detection of synthetic images generated by diffusion models. In: ICASSP (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 327, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 327, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 327, + 480, + 347 + ], + "type": "text", + "content": "9. Debevec, P., Hawkins, T., Tchou, C., Duiker, H.P., Sarokin, W., Sagar, M.: Acquiring the reflectance field of a human face. In: SIGGRAPH (2000)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 348, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 480, + 369 + ], + "type": "text", + "content": "10. Ding, Z., Zhang, C., Xia, Z., Jebe, L., Tu, Z., Zhang, X.: DiffusionRig: Learning personalized priors for facial appearance editing. In: CVPR (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 370, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 480, + 392 + ], + "type": "text", + "content": "1. Futschik, D., Ritland, K., Vecore, J., Fanello, S., Orts-Escolano, S., Curless, B., Sykora, D., Pandey, R.: Controllable light diffusion for portraits. In: CVPR (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 392, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 392, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 134, + 392, + 480, + 446 + ], + "type": "text", + "content": "12. Guo, K., Lincoln, P., Davidson, P., Busch, J., Yu, X., Whalen, M., Harvey, G., Orts-Escolano, S., Pandey, R., Dourgarian, J., DuVall, M., Tang, D., Tkach, A., Kowdle, A., Cooper, E., Dou, M., Fanello, S., Fyffe, G., Rhemann, C., Taylor, J., Debevecc, P., Izadi, S.: The Relighttables: Volumetric performance capture of humans with realistic relighting. In: SIGGRAPH Asia (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 134, + 447, + 480, + 468 + ], + "type": "text", + "content": "13. Guo, X., Liu, X., Ren, Z., Grosz, S., Masi, I., Liu, X.: Hierarchical fine-grained image forgery detection and localization. In: CVPR (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 469, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 469, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 134, + 469, + 480, + 490 + ], + "type": "text", + "content": "14. He, Y., Xing, Y., Zhang, T., Chen, Q.: Unsupervised portrait shadow removal via generative priors. In: MM (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 134, + 491, + 480, + 511 + ], + "type": "text", + "content": "15. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 134, + 512, + 480, + 533 + ], + "type": "text", + "content": "16. Hold-Geoffroy, Y., Athawale, A., Lalonde, J.F.: Deep sky modeling for single image outdoor lighting estimation. In: CVPR (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 134, + 534, + 480, + 555 + ], + "type": "text", + "content": "17. Hou, A., Sarkis, M., Bi, N., Tong, Y., Liu, X.: Face relighting with geometrically consistent shadows. In: CVPR (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 134, + 556, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 556, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 134, + 556, + 480, + 578 + ], + "type": "text", + "content": "18. Hou, A., Zhang, Z., Sarkis, M., Bi, N., Tong, Y., Liu, X.: Towards high fidelity face relighting with realistic shadows. In: CVPR (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 134, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 134, + 578, + 480, + 599 + ], + "type": "text", + "content": "19. Iqbal, U., Caliskan, A., Nagano, K., Khamis, S., Molchanov, P., Kautz, J.: RANA: Relightable articulated neural avatars. In: ICCV (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 134, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 480, + 621 + ], + "type": "text", + "content": "20. Ji, C., Yu, T., Guo, K., Liu, J., Liu, Y.: Geometry-aware single-image full-body human relighting. In: ECCV (2022)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 134, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 480, + 643 + ], + "type": "text", + "content": "21. Kamat, S., Agarwal, S., Darrell, T., Rohrbach, A.: Revisiting generalizability in deepfake detection: Improving metrics and stabilizing transfer. In: ICCVW (2023)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 134, + 644, + 480, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 644, + 480, + 654 + ], + "spans": [ + { + "bbox": [ + 134, + 644, + 480, + 654 + ], + "type": "text", + "content": "22. Kingma, D., Adam, J.B.: A method for stochastic optimization. In: ICLR (2014)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 134, + 654, + 467, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 467, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 467, + 665 + ], + "type": "text", + "content": "23. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: ICLR (2014)" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 229, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 229, + 91, + 448, + 102 + ], + "type": "text", + "content": "COMPOSE: Comprehensive Portrait Shadow Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "24. Lagunas, M., Sun, X., Yang, J., Villegas, R., Zhang, J., Shu, Z., Masia, B., Gutierrez, D.: Single-image full-body human relighting. In: EGSR (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "text", + "content": "25. Le, H., Kakadiaris, I.: Illumination-invariant face recognition with deep relit face images. In: WACV (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "type": "text", + "content": "26. Liu, Y., Hou, A., Huang, X., Ren, L., Liu, X.: Blind removal of facial foreign shadows. In: BMVC (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "text", + "content": "27. Lorenz, P., Durall, R.L., Keuper, J.: Detecting images generated by deep diffusion models using their local intrinsic dimensionality. In: ICCVW (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 203, + 480, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 235 + ], + "type": "text", + "content": "28. Mei, Y., Zhang, H., Zhang, X., Zhang, J., Shu, Z., Wang, Y., Wei, Z., Yan, S., Jung, H., Patel, V.M.: LightPainter: Interactive portrait relighting with freehand scribble. In: CVPR (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 236, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 257 + ], + "type": "text", + "content": "29. Nestmeyer, T., Lalonde, J.F., Matthews, I., Lehrmann, A.: Learning physics-guided face relighting under directional light. In: CVPR (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 258, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 258, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 258, + 480, + 278 + ], + "type": "text", + "content": "30. Ojha, U., Li, Y., Lee, Y.J.: Towards universal fake image detectors that generalize across generative models. In: CVPR (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 279, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 480, + 310 + ], + "type": "text", + "content": "31. Pandey, R., Orts-Escalano, S., LeGendre, C., Haene, C., Bouaziz, S., Rhemann, C., Debevec, P., Fanello, S.: Total relighting: Learning to relight portraits for background replacement. In: SIGGRAPH (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 311, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 342 + ], + "type": "text", + "content": "32. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in PyTorch. In: NeurIPSW (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 343, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 364 + ], + "type": "text", + "content": "33. Ponglertnapakorn, P., Tritrong, N., Suwajanakorn, S.: DiFaReli: Diffusion face relighting. In: ICCV (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "type": "text", + "content": "34. Ranjan, A., Yi, K.M., Chang, J.H.R., Tuzel, O.: FaceLit: Neural 3D relightable faces. In: CVPR (June 2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "type": "text", + "content": "35. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 407, + 480, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 428 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 428 + ], + "type": "text", + "content": "36. Sengupta, R., Curless, B., Kemelmacher-Shlizerman, I., Seitz, S.: A light stage on every desk. In: ICCV (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 429, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 449 + ], + "type": "text", + "content": "37. Sengupta, S., Kanazawa, A., Castillo, C.D., Jacobs, D.W.: SfSNet: Learning shape, reflectance and illuminance of faces in the wild. In: CVPR (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 450, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 471 + ], + "type": "text", + "content": "38. Shih, Y., Paris, S., Barnes, C., Freeman, W.T., Durand, F.: Style transfer for headshot portraits. TOG (2014)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "text", + "content": "39. Shu, Z., Hadap, S., Shechtman, E., Sunkavalli, K., Paris, S., Samaras, D.: Portrait lighting transfer using a mass transport approach. TOG (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 494, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 480, + 514 + ], + "type": "text", + "content": "40. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: ICLR (2015)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 515, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 480, + 546 + ], + "type": "text", + "content": "41. Sun, T., Barron, J.T., Tsai, Y.T., Xu, Z., Yu, X., Fyffe, G., Rhemann, C., Busch, J., Debevec, P., Ramamoorthi, R.: Single image portrait relighting. In: SIGGRAPH (2019)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "type": "text", + "content": "42. Sun, T., Lin, K.E., Bi, S., Xu, Z., Ramamoorthi, R.: NeLF: Neural light-transport field for portrait view synthesis and relighting. In: EGSR (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 569, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 480, + 601 + ], + "type": "text", + "content": "43. Tan, F., Fanello, S., Meka, A., OrtsEscolano, S., Tang, D., Pandey, R., Taylor, J., Tan, P., Zhang., Y.: VoLux-GAN: A generative model for 3D face synthesis with HDRI relighting. In: SIGGRAPH (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 601, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 621 + ], + "type": "text", + "content": "44. Wang, Y., Holynski, A., Zhang, X., Zhang, X.: Sunstage: Portrait reconstruction and relighting using the sun as a light stage. In: CVPR (2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "45. Wang, Z., Yu, X., Lu, M., Wang, Q., Qian, C., Xu, F.: Single image portrait relighting via explicit multiple reflectance channel modeling. TOG (2020)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "46. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. TIP (2004)" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "A. Hou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 314 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "47. Weir, J., Zhao, J., Chalmers, A., Rhee, T.: Deep portrait delighting. In: ECCV (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "text", + "content": "48. Yang, X., Taketomi, T.: BareSkinNet: Demakeup and de-lighting via 3D face reconstruction. CGF (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 194 + ], + "type": "text", + "content": "49. Yeh, Y.Y., Nagano, K., Khamis, S., Kautz, J., Liu, M.Y., Wang, T.C.: Learning to relight portrait images via a virtual light stage and synthetic-to-real adaptation. TOG (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "text", + "content": "50. Zhang, L., Zhang, Q., Wu, M., Yu, J., Xu, L.: Neural video portrait relighting in real-time via consistency modeling. In: ICCV (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "51. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "52. Zhang, X., Fanello, S., Tsai, Y.T., Sun, T., Xue, T., Pandey, R., Orts-Escalano, S., Davidson, P., Rhemann, C., Debevec, P., Barron, J.T., Ramamoorthi, R., Freeman, W.T.: Neural light transport for relighting and view synthesis. TOG (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 293 + ], + "type": "text", + "content": "53. Zhang, X., Barron, J.T., Tsai, Y.T., Pandey, R., Zhang, X., Ng, R., Jacobs, D.E.: Portrait Shadow Manipulation. TOG (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "54. Zhou, H., Hadap, S., Sunkavalli, K., Jacobs, D.W.: Deep single portrait image relighting. In: ICCV (2019)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 229, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 229, + 91, + 447, + 102 + ], + "type": "text", + "content": "COMPOSE: Comprehensive Portrait Shadow Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CONDA_ Condensed Deep Association Learning for Co-Salient Object Detection./4b7d9a37-9ffb-4ebc-9e75-7e4ada571495_content_list.json b/2024/CONDA_ Condensed Deep Association Learning for Co-Salient Object Detection./4b7d9a37-9ffb-4ebc-9e75-7e4ada571495_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..297722bbcea22b60470b3eba5195fcbe91e60489 --- /dev/null +++ b/2024/CONDA_ Condensed Deep Association Learning for Co-Salient Object Detection./4b7d9a37-9ffb-4ebc-9e75-7e4ada571495_content_list.json @@ -0,0 +1,1928 @@ +[ + { + "type": "text", + "text": "CONDA: Condensed Deep Association Learning for Co-Salient Object Detection", + "text_level": 1, + "bbox": [ + 225, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Long Li $^{1}$ , Nian Liu $^{3,*}$ , Dingwen Zhang $^{1}$ , Zhongyu Li $^{4}$ , Salman Khan $^{3,5}$ , Rao Anwer $^{3}$ , Hisham Cholakkal $^{3}$ , Junwei Han $^{1,2,*}$ , and Fahad Shahbaz Khan $^{3,6}$", + "bbox": [ + 220, + 210, + 781, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Northwestern Polytechnical University", + "bbox": [ + 364, + 268, + 638, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Institute of Artificial Intelligence, Hefei Comprehensive National Science Center", + "bbox": [ + 228, + 282, + 772, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Mohamed bin Zayed University of Artificial Intelligence", + "bbox": [ + 308, + 297, + 692, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Xi'an Jiaotong University", + "bbox": [ + 406, + 311, + 594, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5 Australian National University", + "bbox": [ + 390, + 325, + 611, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{6}$ CVL, Linköping University", + "bbox": [ + 401, + 339, + 599, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Inter-image association modeling is crucial for co-salient object detection. Despite satisfactory performance, previous methods still have limitations on sufficient inter-image association modeling. Because most of them focus on image feature optimization under the guidance of heuristically calculated raw inter-image associations. They directly rely on raw associations which are not reliable in complex scenarios, and their image feature optimization approach is not explicit for inter-image association modeling. To alleviate these limitations, this paper proposes a deep association learning strategy that deploys deep networks on raw associations to explicitly transform them into deep association features. Specifically, we first create hyperassociations to collect dense pixel-pair-wise raw associations and then deploys deep aggregation networks on them. We design a progressive association generation module for this purpose with additional enhancement of the hyperassociation calculation. More importantly, we propose a correspondence-induced association condensation module that introduces a pretext task, i.e. semantic correspondence estimation, to condense the hyperassociations for computational burden reduction and noise elimination. We also design an object-aware cycle consistency loss for high-quality correspondence estimations. Experimental results in three benchmark datasets demonstrate the remarkable effectiveness of our proposed method with various training settings. The code is available at: https://github.com/dragonlee258079/CONDA.", + "bbox": [ + 259, + 393, + 743, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Co-salient Object Detection $\\cdot$ Deep Association Learning", + "bbox": [ + 261, + 710, + 730, + 726 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 752, + 375, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Co-Salient Object Detection (CoSOD) aims to segment salient objects that appear commonly across a group of related images. Compared to traditional Salient", + "bbox": [ + 212, + 785, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding authors: {liunian228, junweihan2010}@gmail.com", + "bbox": [ + 230, + 825, + 676, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/404824b20790bd64d1da0a79eb3ae6b9a0ab43b402bf8d6b4670d49ac69f6027.jpg", + "image_caption": [ + "Fig. 1: Difference of raw-association-based image feature optimization strategy (a) and our proposed deep association learning strategy (b). Our deep association learning deploys deep learning networks on raw associations to achieve deep association features. We also present visual samples of our calculated raw associations (Raw Asso.), optimized image feature (Opt. Fea.), and our generated deep association features (Asso. Fea.) in (c)." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 588, + 305 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7f9cc2f91034f063da20762cb03c1295457ec68c8b6828e05a7cdde337347203.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 593, + 146, + 782, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Object Detection (SOD) [19-23, 26, 49, 50, 52], CoSOD is a more challenging task because it requires sufficient inter-image association modeling [9].", + "bbox": [ + 212, + 398, + 785, + 430 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, many advanced works [8,13,16,33,38,41,46,47,51] have emerged and achieved impressive performance. These methods first use related image features to acquire raw inter-image associations (also called consensus representations), and then leverage them as guidance to optimize each image feature, as shown in Figure 1 (a). This approach enables the final image feature to implicitly capture inter-image cues, thereby achieving the purpose of inter-image association modeling. However, we find this raw-association-based feature optimization strategy still has two limitations: i) they directly rely on raw associations, which are acquired in heuristic manners, such as pixel-wise [8,33,41,51], region-wise [16,46], or image-wise [13] similarity measurements. Although high-quality raw associations can be derived from high-level semantic information in image features, their revelation of common saliency regions still relies on similarity measures, which are unreliable when encountering complex scenarios, such as significant differences between co-salient objects or high foreground-background similarity. ii) the primary focus of building their deep models is on optimizing image features. Compared to directly modeling association relationship, image feature optimization is not an explicit approach for inter-image association modeling and will increase the learning difficulty.", + "bbox": [ + 212, + 431, + 787, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To alleviate these limitations, we propose a deep association learning strategy for CoSOD, as shown in Figure 1 (b). Instead of directly using raw associations to optimize image features, we deploy deep networks on raw associations to learn deep association features. This is a more explicit strategy for inter-image association modeling. Moreover, our deep association features can capture high-level inter-image association knowledge, making them more robust in complex scenarios than raw associations, as shown in Figure 1 (c). Technically, we start by collecting all pixel-wise raw associations across the entire image group as hyperassociations. Then, we propose a Progressive Association Generation (PAG)", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "L Li et al.", + "bbox": [ + 271, + 114, + 341, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "module to transform hyperassociations into deep association features. PAG progressively generates association features on varying scales, allowing us to use previous association features to enhance the hyperassociation calculation at the next scale, thereby improving the association quality from the very beginning.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although deep association learning strategy allows more sufficient interimage association modeling, it significantly increases the computational burden and reduces the practicality of this approach. Additionally, this study finds that it is not necessary to utilize all pixel associations to generate deep association features. In fact, there are even some noisy pixels that negatively impact the quality of the deep association features. Therefore, we propose a method based on Correspondence-induced Association Condensation (CAC) to condense the original full-pixel hyperassociations. This not only alleviates the computational burden but also further enhances the quality of deep association features.", + "bbox": [ + 212, + 207, + 784, + 342 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, CAC performs the condensation operation by selectively associating pixels that have semantic correspondence in other images, as well as their surrounding contextual pixels, thereby creating lightweight yet more accurate hyperassociations. Here, we introduce a pretext task, i.e. semantic correspondence estimation, into the CoSOD, not only improving the model performance but also delving deeper into the essence of CoSOD. Co-salient objects inherently possess object-level semantic correspondence. However, in this paper, we aim to further explore the finer pixel-level correspondence. Although highly accurate correspondence estimation remains a challenge, we believe it will pave a new way for CoSOD research. We also provide an object-aware cycle consistency (OCC) loss to aid in learning correspondences within co-salient pixels.", + "bbox": [ + 212, + 343, + 785, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, the contributions of this paper are as follows:", + "bbox": [ + 238, + 508, + 661, + 523 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- For the first time, we introduce a deep association learning approach for CoSOD, applying deep networks to transform raw associations into deep association features for sufficient inter-image association modeling. Specifically, we develop a CONdensed Deep Association (CONDA) learning model.", + "- We propose a PAG module to progressively generate deep association features. It enhances image features with previous association features to improve hyperassociation calculation.", + "- We introduce semantic correspondence estimation into the CoSOD task to condense the original hyperassociation for alleviating the computational burden and further improving the performance. We also propose an OCC loss for effective correspondence estimation.", + "- Experimental results demonstrate that our model achieves significantly improved state-of-the-art performance on three benchmark datasets across different training settings." + ], + "bbox": [ + 223, + 526, + 782, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 756, + 387, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Co-Salient Object Detection", + "text_level": 1, + "bbox": [ + 215, + 786, + 495, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, there has been a surge of excellent methods for CoSOD [8,10,13,16,33, 38,41-43,45-47,51]. These methods initially acquire raw inter-image associations", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "CONDA", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with related image features and then utilize them to optimize each image feature. Most methods generate pixel-wise [8, 33, 41, 51], region-wise [16, 46], and image-wise [13] raw associations through similarity-based manners, e.g. inner product calculations between image features. Even for transformer-based methods [16, 33], they also rely on inner product calculation to produce attention maps [36] as raw associations. The image feature optimization manners include spatial or channel calibration [8, 13, 51], dynamic convolution [41, 46], feature fusion [16, 41], and cross-attention [16, 33], etc. However, they lack the learning of high-level association knowledge and heavily focus on optimized image features. Unlike them, this paper proposes a new research direction that deploys deep networks on associations to achieve deep association features for CoSOD.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Inter-image Relation Modeling", + "text_level": 1, + "bbox": [ + 214, + 334, + 517, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Apart from CoSOD, other tasks necessitating the consideration of inter-image relationships, such as few-shot segmentation [25, 27, 28, 39], stereo matching [2, 40], video semantic segmentation [34], etc. These tasks have made significant advancements recently by effectively modeling inter-image relations. Most of these methods [2, 4, 12, 27, 39, 40] initially create cost volumes to capture dense inter-image pixel-wise similarities and subsequently use various modules to convert these hyper volumes into task-specific features. Our approach distinguishes itself from prior methods in three aspects. Firstly, most of them create 4D cost volumes between two images while we create 6D hyperassociations between all related images. Secondly, they calculate hyper volumes based on original image features, whereas We propose PAG to progressively enhance image features for better hyperassociation calculation. Last and most importantly, they rely on the full-pixel cost volume. We consider the condensation of hyperassociations using semantic correspondences to eliminate noisy pixel associations.", + "bbox": [ + 212, + 359, + 787, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Semantic Correspondence Estimation", + "text_level": 1, + "bbox": [ + 214, + 593, + 571, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic correspondence estimation [24] aims to establish reliable pixel correspondences between different instances of the same object category. Most works performed this task using fully supervised training [14, 24]. Some recent works have utilized unsupervised learning with photometric, forward-backward consistency, and warp-supervision losses [30, 35, 48]. However, they implement these losses on the entire image, where background pixels may affect the performance. In this paper, we introduce this task to condense hyperassociations for Co-SOD and tailor the cycle consistency loss by only applying it to co-salient pixels, hence effectively avoiding the influence of background and extraneous objects.", + "bbox": [ + 212, + 618, + 787, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Proposed Method", + "text_level": 1, + "bbox": [ + 214, + 777, + 428, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 2, CONDA integrates the deep association learning process into an FPN framework. Specifically, given a group of related images $\\{I_i\\}_{i=1}^N$ , we", + "bbox": [ + 212, + 809, + 785, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "L Li et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ac6fd1de06c08cc86d786e4ae99a47c98f05aa45788c8c35362c4a5254b32dd3.jpg", + "image_caption": [ + "Fig. 2: Overall flowchart of our CONDA model. Specifically, CONDA first utilizes the image features to calculate hyperassociations. Then, the full-pixel hyperassociations are condensed by CAC and fed into the aggregation networks to achieve deep association features. These features are then used in the FPN decoder process for the final prediction. To be concise, only three related images are shown." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 782, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "first input them into a VGG-16 [32] backbone to collect its intermediate features for PAG and FPN decoding. In detail, we collect all features of the last three stages for PAG and the last feature of each stage for the FPN decoder as follows:", + "bbox": [ + 212, + 484, + 787, + 531 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbb {F} ^ {P} = \\left\\{\\boldsymbol {F} _ {s, l} \\mid s \\in \\{3, \\dots , 5 \\}, l \\in \\{1, \\dots , L _ {s} \\} \\right\\}, \\\\ \\mathbb {F} ^ {D} = \\left\\{\\boldsymbol {F} _ {s, L _ {s}} \\mid s \\in \\{1, \\dots 5 \\} \\right\\}, \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 542, + 784, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbb{F}^P$ and $\\mathbb{F}^D$ are the feature collections for PAG and FPN decoder, respectively. $F_{s,l}\\in \\mathbb{R}^{N\\times H_s\\times W_s\\times C_s}$ represents the VGG feature of the $l$ -th layer in the $s$ -th stage. There are in total $L_{s}$ layers in the $s$ -th stage. $H_{s}$ , $W_{s}$ and $C_s$ represent the height, width, and channel of the $s$ -th stage, respectively.", + "bbox": [ + 212, + 590, + 787, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we input $\\mathbb{F}^P$ into PAG to calculate hyperassociations and generate deep association features $\\{\\pmb {F}_s^A\\in \\mathbb{R}^{N\\times H_s\\times W_s\\times C_s}\\}_{s = 3}^5$ . Finally, these association features will be fused with $\\mathbb{F}^D$ for the FPN decoder process, formulated as:", + "bbox": [ + 212, + 652, + 787, + 699 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {F} _ {s, L _ {s}} = \\boldsymbol {F} _ {s, L _ {s}} + \\phi (\\boldsymbol {F} _ {s} ^ {A}), \\quad s = 3, \\dots , 5, \\\\ \\boldsymbol {F} = \\operatorname {D e c o d e r} \\left(\\left\\{\\boldsymbol {F} _ {s, L _ {s}} \\right\\} _ {s = 1} ^ {5}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 710, + 642, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\phi$ is a convolution layer. $\\pmb{F} \\in \\mathbb{R}^{N \\times H \\times W \\times C}$ is the final feature for the final co-saliency prediction. We adopt BCE and IoU losses for supervision.", + "bbox": [ + 212, + 762, + 787, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The rest of this section will introduce PAG with full-pixel hyperassociation and the condensation of hyperassociations by plugging the Correspondence-induced Association Condensation (CAC) module into PAG.", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "CONDA", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Progressive Association Generation", + "text_level": 1, + "bbox": [ + 215, + 146, + 553, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our deep association learning involves two steps: 1) acquiring the raw hyper-associations $\\{A_s\\}_{s=3}^5$ from three stages; 2) employing aggregation networks on $\\{A_s\\}_{s=3}^5$ to obtain association features $\\{F_s^A\\}_{s=3}^5$ .", + "bbox": [ + 212, + 172, + 782, + 217 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Earlier methods [4, 11, 31] directly utilize the original backbone features, i.e. $\\mathbb{F}^P$ , to calculate inter-image interactions, such as the so-called cost volume. We argue that hyperassociations derived straight from backbone features might obstruct further improvement of deep association learning, given that the current backbone is pre-trained without any consideration for inter-image associations.", + "bbox": [ + 212, + 218, + 782, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To alleviate this problem, we propose the PAG module to progressively generate pyramid association features so that we can utilize the high-level association feature, e.g. $\\pmb{F}_{s+1}^{A}$ from the $(s+1)$ -th stage, to enhance the neighbouring low-level VGG features in $\\mathbb{F}^P$ , e.g. $\\pmb{F}_{s,l}$ , for attaining association-enhanced features $\\hat{\\pmb{F}}_{s,l}$ , based on which we can calculate high-quality hyperassociations $\\pmb{A}_s$ in the $s$ -th stage. After that, we execute the subsequent aggregation networks on $\\pmb{A}_s$ to achieve association features $\\pmb{F}_s^A$ , which will continue to enhance the VGG features of the next stage and carry out progressive association generation. The whole process of our PAG can be formulated as follows:", + "bbox": [ + 212, + 292, + 784, + 430 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {A} _ {s} = \\operatorname {H A C} \\left(\\left\\{\\tilde {\\boldsymbol {F}} _ {s, l} \\right\\} _ {l = 1} ^ {L _ {s}}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 443, + 584, + 460 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} _ {s} ^ {A} = \\operatorname {A g g} \\left(\\boldsymbol {A} _ {s}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 442, + 464, + 784, + 481 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\{\\hat {\\boldsymbol {F}} _ {s - 1, l} \\} _ {l = 1} ^ {L _ {s - 1}} = \\operatorname {E n h} \\left(\\left\\{\\boldsymbol {F} _ {s - 1, l} \\right\\} _ {l = 1} ^ {L _ {s - 1}}; \\boldsymbol {F} _ {s} ^ {A}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 484, + 645, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $s$ ranges from 5 to 3 and $\\{\\hat{F}_{5,l}\\}_{l=1}^{L_5} = \\{\\pmb{F}_{5,l}\\}_{l=1}^{L_5}$ . The HAC, Agg, and Enh represent the hyperassociation calculation, aggregation network, and association-induced feature enhancement, respectively. Next, we explain them in detail.", + "bbox": [ + 212, + 517, + 782, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hyperassociation Calculation. For each stage, we first compute the raw associations at each layer using the inner product between $l$ -2 normalized association-enhanced features of $N$ related images. After that, we stack the raw associations of all layers to form the final hyperassociation of this stage. The hyperassociation for the $s$ -th stage, i.e. $A_{s} \\in \\mathbb{R}^{N \\times H_{s} \\times W_{s} \\times N \\times H_{s} \\times W_{s} \\times L_{s}}$ , can be calculated via:", + "bbox": [ + 212, + 564, + 784, + 638 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {A} _ {s} = \\operatorname {H A C} \\left(\\left\\{\\hat {\\boldsymbol {F}} _ {s, l} \\right\\} _ {l = 1} ^ {L _ {s}}\\right) \\\\ = \\operatorname {S t a c k} \\left(\\left\\{\\operatorname {R e L U} \\left(\\frac {\\hat {\\boldsymbol {F}} _ {s , l} \\cdot \\hat {\\boldsymbol {F}} _ {s , l} ^ {\\top}}{\\| \\hat {\\boldsymbol {F}} _ {s , l} \\| \\| \\hat {\\boldsymbol {F}} _ {s , l} ^ {\\top} \\|}\\right) \\right\\} _ {l = 1} ^ {L _ {s}}\\right), \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 651, + 784, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{\\pmb{F}}_{s,l}\\cdot \\hat{\\pmb{F}}_{s,l}^{\\top}\\in \\mathbb{R}^{N\\times H_s\\times W_s\\times N\\times H_s\\times W_s}$ . The $\\top$ indicates transposing the last dimension and the first three dimensions. $\\| \\cdot \\|$ represents the $l$ -2 norm. We employ ReLU to suppress noisy association values.", + "bbox": [ + 212, + 729, + 782, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Aggregation Network. The raw hyperassociation $\\mathbf{A}_s \\in \\mathbb{R}^{N \\times H_s \\times W_s \\times N \\times H_s \\times W_s \\times L_s}$ is a hypercube with a nested structure, where each pixel position is characterized by a 4D tensor $(N \\times H_s \\times W_s \\times L_s)$ . Each 4D tensor documents the associations of the respective pixel with all other pixels in $N$ related images. For", + "bbox": [ + 212, + 777, + 782, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "L Li et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "clarity, we designate the first and second $N \\times H_s \\times W_s$ dimensions in $A_s$ as the source and target dimensions, respectively. Although these 4D tensors are crucial for exploring the consensus information for co-saliency detection, they essentially comprise pixel-to-pixel similarity values, seen in (4), which may be suboptimal and unreliable in complex scenarios. Therefore, we propose using deep networks to transform these pixel-wise similarities into deep association features with contextual and high-order association knowledge. This has never been explored in previous CoSOD methods. This is implemented via context aggregations on $A_s$ to squeeze these 4D tensors as $C_s$ -dimensional vectors, formulated as:", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nN \\times H _ {s} \\times W _ {s} \\times N \\times H _ {s} \\times W _ {s} \\times L _ {s} \\rightarrow N \\times H _ {s} \\times W _ {s} \\times C _ {s}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 295, + 785, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In detail, we first deploy several association aggregation layers on $\\mathbf{A}_s$ to progressively aggregate context information, enlarging $L_s$ as $C_s$ , and eliminate the target $H_s \\times W_s$ dimension in 4D tensors. Each aggregation layer consists of 2D convolution layers and a downsampling operation. Specifically, focusing on the first aggregation layer for technical explanation, we first aggregate context information in the target $H_s \\times W_s$ dimension by applying a 2D convolution layer on all 4D tensors. The operations on the 4D tensor at pixel position $(h_i, w_i)$ in image $I_i$ can be formulated as:", + "bbox": [ + 212, + 321, + 787, + 443 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {A} _ {s} ^ {1} (i, h _ {i}, w _ {i}, j, \\therefore , \\therefore): = \\mathcal {C} _ {1} \\left(\\boldsymbol {A} _ {s} (i, h _ {i}, w _ {i}, j, \\therefore , \\therefore)\\right), \\quad j = 1, \\dots , N, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 281, + 454, + 785, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{C}_1$ is a $3\\times 3$ 2D convolution layer. Here, $j$ is the index of other related images in the 4D tensor, and $\\pmb {A}_s(i,h_i,w_i,j,:;:)\\in \\mathbb{R}^{H_s\\times W_s\\times L_s}$ illustrates the associations between the pixel $(h_i,w_i)$ in $I_{i}$ and all pixels in image $I_{j}$ . This interpretation applies to other similar symbols.", + "bbox": [ + 212, + 482, + 787, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Then, a downsampling operation $\\mathcal{D}$ , i.e. bilinear interpolation, is applied on $\\pmb{A}_s^1$ to reduce the spatial dimension of the 4D tensor by a scaling factor:", + "bbox": [ + 215, + 542, + 785, + 574 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {A} _ {s} ^ {2} (i, h _ {i}, w _ {i}, j, \\therefore , \\therefore ,:) = \\mathcal {D} (\\boldsymbol {A} _ {s} ^ {1} (i, h _ {i}, w _ {i}, j, \\therefore , \\therefore)), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 585, + 785, + 602 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $A_{s}^{2}(i,h_{i},w_{i},j,:,:) \\in \\mathbb{R}^{H_{s}^{\\prime} \\times W_{s}^{\\prime} \\times C_{s}^{\\prime}}$ , $H_{s}^{\\prime}$ and $W_{s}^{\\prime}$ are downsampled height and width. $C_{s}^{\\prime}$ is the channel number after convolution $\\mathcal{C}_1$ .", + "bbox": [ + 212, + 613, + 785, + 646 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, we also aggregate context information in the source $H_{s} \\times W_{s}$ dimension. Specifically, we extract the values along the source dimension and channel dimension in $A_{s}^{2}$ to form 4D tensors and apply a 2D convolution layer on them. For instance, $A_{s}^{2}(:, :, :, j, h_{j}, w_{j},:) \\in \\mathbb{R}^{N \\times H_{s} \\times W_{s} \\times C_{s}^{\\prime}}$ is such a 4D tensor, where $(h_{j}, w_{j})$ is a pixel position in the target dimension. This can be formulated as:", + "bbox": [ + 212, + 646, + 787, + 723 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {A} _ {s} ^ {3} (i,:,:) j, h _ {j}, w _ {j,:}) = \\mathcal {C} _ {2} \\left(\\boldsymbol {A} _ {s} ^ {2} (i,:,:) j, h _ {j}, w _ {j,:})\\right), \\quad i = 1, \\dots , N, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 279, + 732, + 785, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{C}_2$ is a $3\\times 3$ 2D convolution layer. $i$ is the related image index.", + "bbox": [ + 212, + 760, + 717, + 775 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After several association aggregation layers, as shown in (6)-(8), we can achieve the aggregated association features with the target $H_{s} \\times W_{s}$ dimension eliminated, denoted as $\\pmb{F}_{s}^{A^{\\prime}} \\in \\mathbb{R}^{N \\times H_{s} \\times W_{s} \\times N \\times C}$ . Subsequently, we average $\\pmb{F}_{s}^{A^{\\prime}}$ on its second $N$ dimension and obtain the final association features", + "bbox": [ + 212, + 776, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "CONDA", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e979f8630c372e8cd48870e7fafda8137440dd82e40736aff359426acda2b2dc.jpg", + "image_caption": [ + "Fig. 3: Difference between full-pixel hyperassociation (a) and condensed hyperassociation (b). We provide an example of collecting the pixel associations from image $I_{j}$ for a pixel $(h_{i},w_{i})$ in image $I_{i}$ . Full-pixel hyperassociation collects all pixel associations in $I_{j}$ , while our condensed hyperassociation only collects the associations of its correspondence pixel $(h_{j},w_{j})$ (red dot) and surrounding pixels (green dots). We first heuristically find an initial pixel $(h_j^0,w_j^0)$ with a fixed surrounding window and then learn coordinate offsets to locate the optimized correspondence and surrounding pixels." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 754, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\pmb{F}_s^A \\in \\mathbb{R}^{N \\times H_s \\times W_s \\times C}$ , formulated as:", + "bbox": [ + 215, + 362, + 478, + 380 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} _ {s} ^ {A} = \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\boldsymbol {F} _ {s} ^ {A ^ {\\prime}} (:,:,:, j,:). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 390, + 785, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Association-induced Feature Enhancement. Once we have obtained association feature $\\mathbf{F}_s^A$ of the $s$ -th stage, we will use it to enhance the VGG feature of the $(s - 1)$ -th stage, i.e. $\\{\\mathbf{F}_{s - 1,l}\\}_{l = 1}^{L_{s - 1}}$ . Specifically, we upsample $\\mathbf{F}_s^A$ to align the spatial size of features in the $(s - 1)$ -th stage, and then add it to $\\{\\mathbf{F}_{s - 1,l}\\}_{l = 1}^{L_{s - 1}}$ followed by a convolution layer, formulated as:", + "bbox": [ + 214, + 429, + 787, + 510 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {F}} _ {s - 1, l} = \\mathcal {C} _ {3} \\left(\\boldsymbol {F} _ {s - 1, l} + \\mathcal {U} \\left(\\boldsymbol {F} _ {s} ^ {A}\\right)\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 520, + 785, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{C}_3$ and $\\mathcal{U}$ represent a 2D convolution layer and the bilinear upsampling operation, respectively.", + "bbox": [ + 212, + 550, + 787, + 582 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.2 Correspondence-induced Association Condensation", + "text_level": 1, + "bbox": [ + 214, + 603, + 683, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although PAG based on full-pixel hyperassociations can deliver satisfactory performance in CoSOD, it also introduces substantial computational overhead. Additionally, we argue that for each pixel in an image, it is unnecessary to gather its associations with all pixels of other related images to form hyperassociations. Some pixel associations may even impair the final performance, such as those between ambiguous regions. To this end, this subsection tries to condense the original full-pixel hyperassociations to retain only informative pixel associations.", + "bbox": [ + 212, + 628, + 787, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This subsection will focus on explaining the condensation of pixel associations of a pixel (e.g. $(h_i, w_i)$ in $I_i$ ) to ones of the other images (e.g. image $I_j$ ), i.e. $\\mathbf{A}_s(i, h_i, w_i, j, :, :, :) \\in \\mathbb{R}^{H_s \\times W_s \\times L_s}$ , as shown in Figure 3. We will simplify the symbol $\\mathbf{A}_s(i, h_i, w_i, j, :, :, :)$ as $\\mathbf{a}_s^j$ in the subsequent text for convenience.", + "bbox": [ + 214, + 734, + 787, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specifically, CAC opts to select $K \\times K$ ( $K < H_{s}, K < W_{s}$ ) informative pixel associations from $\\pmb{a}_{s}^{j}$ to form its condensed representation, i.e. $\\tilde{\\pmb{a}}_{s}^{j} \\in \\mathbb{R}^{K \\times K \\times L_{s}}$ . Thus, the entire condensed hyperassocaton can be symbolized as $\\pmb{A}_{s} \\in \\mathbb{R}^{N \\times H_{s} \\times W_{s} \\times N \\times L_{s}}$ .", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "L Li et al.", + "bbox": [ + 271, + 114, + 341, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$K \\times K \\times L_s$ . To ensure the proper selection of $K \\times K$ pixels, we introduce a pretext task, i.e. semantic correspondence estimation [14,24]. This allows us to first locate the corresponding pixel of $(h_i, w_i)$ in image $I_j$ , i.e. $(h_j, w_j)$ , and then combine $(h_j, w_j)$ with its surrounding pixels as the $K \\times K$ pixel set. We design this approach based on the observation that the co-salient objects across $N$ related images belong to the same semantic category, and the pixels within them should have semantic correspondences to each other, as shown in Figure 4. Therefore, the introduction of semantic correspondence in CAC not only improves the CoSOD performance but also delves deeper into the core nature of the CoSOD task. As far as we know, this is the first work to use semantic correspondence in the CoSOD task.", + "bbox": [ + 212, + 145, + 787, + 311 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Correspondence Estimation. To estimate the correspondence pixel $(h_j, w_j)$ in $I_j$ for $(h_i, w_i)$ , we first identify an initial pixel $(h_j^0, w_j^0)$ via a heuristic approach. Subsequently, we produce a spatial offset to refine $(h_j^0, w_j^0)$ into $(h_j, w_j)$ . To achieve this purpose, all initial correspondence pixels should be utilized to form the initial condensed hyperassociations, with which the initial deep association features can be generated for spatial offset prediction.", + "bbox": [ + 212, + 311, + 787, + 404 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Specifically, we pick out $(h_j^0, w_j^0)$ that has the largest feature similarity value with $(h_i, w_i)$ . Since we have calculated the feature similarities in $\\pmb{a}_s^j$ , the $(h_j^0, w_j^0)$ can be obtained as follows:", + "bbox": [ + 212, + 404, + 787, + 450 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\pmb {a}} _ {s} ^ {j} = \\sum_ {l = 1} ^ {L _ {s}} \\pmb {a} _ {s} ^ {j} (:,: l),\n$$\n", + "text_format": "latex", + "bbox": [ + 437, + 460, + 589, + 487 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\left(h _ {j} ^ {0}, w _ {j} ^ {0}\\right) = \\underset {x, y} {\\operatorname {a r g m a x}} \\left(\\bar {\\boldsymbol {a}} _ {s} ^ {j} (x, y)\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 400, + 479, + 784, + 516 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\bar{\\pmb{a}}_s^j\\in \\mathbb{R}^{H_s\\times W_s}$ is a similarity matrix obtained by eliminating the last dimension of $\\pmb {a}_s^j\\in \\mathbb{R}^{H_s\\times W_s\\times L_s}$ through summation. argmax returns the coordinate of the maximum value.", + "bbox": [ + 212, + 527, + 787, + 571 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Next, we select $K \\times K$ pixels within the square region centered around the initial pixels, e.g. $(h_j^0, w_j^0)$ , to construct the initial condensed hyperassociation $\\tilde{\\pmb{A}}_s'$ . Then, we can feed it into the aggregation network, described in Sec 3.1, and achieve the initial aggregated association features $\\pmb{F}_s^{A'} \\in \\mathbb{R}^{N \\times H_s \\times W_s \\times N \\times C_s}$ . It can be regarded as the association features of each pixel to $N$ other related images. We utilize the feature of $(h_i, w_i)$ to $I_j$ , i.e. $\\pmb{F}_s^{A'}(i, h_i, w_i, j,:) \\in \\mathbb{R}^{C_s}$ , to predict the offsets for $(h_j, w_j)$ via a linear layer, formulated as:", + "bbox": [ + 212, + 574, + 787, + 689 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Delta} _ {s} ^ {j} = \\mathcal {O} \\left(\\boldsymbol {F} _ {s} ^ {A ^ {\\prime}} (i, h _ {i}, w _ {i}, j,:)\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 700, + 784, + 720 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathcal{O}$ is a linear layer for offset generation. $\\pmb{\\Delta}_{s}^{j} \\in \\mathbb{R}^{K \\times K \\times 2}$ consists of $K \\times K$ offsets, besides the center offset $\\pmb{\\Delta}_{s}^{j}(k_{c}, k_{c},:)$ for correspondence estimation, i.e. refining $(h_{j}^{0}, w_{j}^{0})$ as $(h_{j}, w_{j})$ , we also generate other offsets for surrounding pixel selection. Thus, the corresponding pixel $(h_{j}, w_{j})$ can be obtained by adding the offset to the initial pixel $(h_{j}^{0}, w_{j}^{0})$ , formulated as:", + "bbox": [ + 212, + 732, + 787, + 811 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\left(h _ {j}, w _ {j}\\right) = \\left(h _ {j} ^ {0}, w _ {j} ^ {0}\\right) + \\boldsymbol {\\Delta} _ {s} ^ {j} \\left(k _ {c}, k _ {c}, :\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 823, + 784, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "CONDA", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $(k_c, k_c)$ is the center position of $K \\times K$ square.", + "bbox": [ + 212, + 146, + 596, + 161 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Condensation Operation. Given the estimated correspondence pixel $(h_j, w_j)$ and other offsets in $\\pmb{\\Delta}_{s}^{j}$ , we can obtain the surrounding pixels and combine them with $(h_j, w_j)$ to form $K \\times K$ pixel set. Their coordinates are stored in $\\pmb{n}_s^j \\in \\mathbb{R}^{K \\times K \\times 2}$ . This process can be formulated as:", + "bbox": [ + 212, + 161, + 787, + 223 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {n} _ {s} ^ {j} (x, y,:) = \\left(h _ {j}, w _ {j}\\right) + \\boldsymbol {\\Delta} _ {s} ^ {j} (x, y,:) \\quad x, y \\in \\{1, \\dots , K \\}; x, y \\neq k _ {c}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 256, + 234, + 785, + 258 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {n} _ {s} ^ {j} (k _ {c}, k _ {c};) = (h _ {j}, w _ {j}),\n$$\n", + "text_format": "latex", + "bbox": [ + 256, + 256, + 419, + 273 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $(k_c, k_c)$ is the center position of $K \\times K$ square.", + "bbox": [ + 212, + 284, + 596, + 299 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Finally, we can perform the condensation operation via the index selection on $\\pmb{a}_s^j$ , formulated as:", + "bbox": [ + 212, + 299, + 784, + 330 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {a}} _ {s} ^ {j} = \\boldsymbol {a} _ {s} ^ {j} \\left(\\boldsymbol {n} _ {s} ^ {j}\\right), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 455, + 329, + 784, + 347 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\tilde{\\pmb{a}}_s^j$ is the condensed representation of $\\pmb{a}_s^j$ , i.e. pixel associations of $(h_i, w_i)$ to image $I_j$ . Furthermore, we also illustrate the condensation for pixel associations of $(h_i, w_i)$ to all images, i.e. $\\pmb{a}_s \\in \\mathbb{R}^{N \\times H_s \\times W_s \\times L_s}$ , in Figure 2. By applying such a condensation process to all pixel associations, we can obtain the final condensed hyperassociation $\\tilde{\\pmb{A}}_s$ .", + "bbox": [ + 212, + 354, + 787, + 431 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3.3 Object-aware Cycle Consistency Loss", + "text_level": 1, + "bbox": [ + 214, + 454, + 570, + 469 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To achieve accurate correspondence estimations, applying effective supervisions on them is necessary. As explicit semantic correspondence annotations are not available, we can only rely on unsupervised losses by imposing correspondence-related constraints on estimated correspondences. Previous unsupervised approaches apply constraints to all pixels [30, 35, 48], including those on the background and extraneous objects that don't have mutual correspondences, hence harming the model effectiveness. To avoid this problem, we propose an object-aware constraint to only access losses on the co-salient pixels.", + "bbox": [ + 212, + 479, + 787, + 599 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We propose an Object-aware Cycle Consistency (OCC) loss for the supervision of correspondence estimations in CoSOD. The cycle consistency can be explained as: if a co-salient pixel $(h_i, w_i)$ of image $I_i$ corresponds to the pixel $(h_j, w_j)$ in image $I_j$ , then the pixel $(h_j, w_j)$ should also semantically correspond to pixel $(h_i, w_i)$ .", + "bbox": [ + 212, + 599, + 787, + 675 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Based on this cycle consistency constraint, we adopt image warping operations to conduct the OCC loss. Specifically, we first warp the image $I_{i}^{s}$ (a resized $I_{i}$ to align the scales in the $s$ -th stage) as $I_{i\\rightarrow j}^{s}$ using the $I_{i}^{s}\\to I_{j}^{s}$ correspondence estimations. Next, we warp $I_{i\\rightarrow j}^{s}$ backward to $I_{i\\rightarrow j\\rightarrow i}^{s}$ using $I_{j}^{s}\\to I_{i}^{s}$ correspondence estimations. Finally, we can utilize the SSIM loss between $I_{i}^{s}$ and $I_{i\\rightarrow j\\rightarrow i}^{s}$ to measure the cycle consistency for mutual correspondence pixels in $I_{i}^{s}$ and $I_{j}^{s}$ . Moreover, to ensure the constraints are only conducted on co-salient objects, we mask the images with their ground truth masks, formulated as:", + "bbox": [ + 212, + 676, + 787, + 796 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} ^ {C} = \\frac {1}{N ^ {2}} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N} \\mathcal {L} _ {S S I M} \\left(I _ {i} ^ {s} \\cdot G _ {i} ^ {s}, I _ {i \\rightarrow j \\rightarrow i} ^ {s} \\cdot G _ {i} ^ {s}\\right), \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 801, + 785, + 844 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "L Li et al.", + "bbox": [ + 271, + 114, + 339, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/0abe3315e7ce9fdde14ed458ce324683c74f248539502589eb2062e08dc79c1a.jpg", + "table_caption": [ + "Table 1: Ablation Study of our proposed modules. SAG, SAC, and FCC are ablation modules for PAG, CAC, and OCC, respectively." + ], + "table_footnote": [], + "table_body": "
IDModulesCoCA [47]
SAGPAGSACCACFCCOCCSm↑Eξ↑Fβ↑M↓
10.69360.76420.57290.1206
2✓0.72360.80290.63570.1106
3✓0.73080.81220.64590.1075
4✓✓0.73040.81230.65000.1085
5✓✓0.74730.81550.65910.0956
6✓✓0.73980.81380.65060.1013
7✓✓0.75700.82480.67510.0924
", + "bbox": [ + 271, + 174, + 723, + 297 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $G_{i}^{s}$ is the resized ground truth of image $I_{i}$ . The total OCC loss $\\mathcal{L}^C$ is the sum of three stages, formulated as: $\\mathcal{L}^C = \\mathcal{L}_3^C +\\mathcal{L}_4^C +\\mathcal{L}_5^C$ . More details can be found in the supplementary materials.", + "bbox": [ + 212, + 310, + 787, + 357 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 378, + 375, + 395 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Evaluation Datasets and Metrics", + "text_level": 1, + "bbox": [ + 214, + 409, + 532, + 424 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We follow [8] to evaluate our model on three benchmark datasets, i.e. CoCA [47] (1295 images of 80 groups), CoSal2015 [44] (2015 images of 50 groups), and CoSOD3k [7] (3316 images of 160 groups). We adopt four widely-used metrics for the quantitative evaluation, i.e. Structure-measure $(S_{m})$ [5], Maximum enhanced-alignment measure $(E_{\\xi})$ [6], Maximum F-measure $(\\mathrm{F}_{\\beta})$ [1], and Mean Absolute Error $(\\mathcal{M})$ [3].", + "bbox": [ + 212, + 431, + 787, + 525 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 544, + 457, + 559 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To construct the training data, we follow [51] to use different combinations of three commonly used training datasets, i.e. DUTS class [47] (8250 images of 291 groups), COCO-9k [18] (9213 images of 65 groups), and COCO-SEG [37] (200,000 images of 78 groups), for a fair comparison with other state-of-the-art (SOTA) works. We also implement the synthesis strategy for the DUTS class dataset following [46].", + "bbox": [ + 212, + 568, + 787, + 657 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For training specifics, we employ the data augmentation strategy in [21] and use $256 \\times 256$ as the input size for the network. We employ the Adam optimizer [15] with $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.99$ to optimize the network. We train our CONDA model for 300 epochs, starting with an initial learning rate of $1e - 4$ , which is divided by 10 at the $60,000^{th}$ iteration. Our experiments are implemented based on PyTorch [29] on a single Tesla A40 GPU, with the batchsize set to $N = 6$ . The hyperparameter $K$ in CAC is set to 9.", + "bbox": [ + 212, + 659, + 787, + 765 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 785, + 390, + 801 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We conduct ablation studies on the most challenging CoCA [47] dataset. To construct the baseline, we use the FPN [17] (with VGG-16 as encoder) as the", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "CONDA", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/40395a7bb7331fffc5891dbbab1881e79c2baa445aa0a5bdba710a534942373c.jpg", + "image_caption": [ + "Fig. 4: Visual samples for the correspondence estimations. Correspondences I, II, and III visually display estimated correspondences between the main image and three related images. Sparse co-salient pixels were selected and connected to their corresponding pixels using colored lines for clear visualization." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 789, + 279 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "foundational segmentation network and enhance it with the Region-to-Region correlation module (R2R) [16] to simply capture inter-image connections. Then, as shown in Table 1, we incrementally incorporate PAG, CAC, and OCC into the baseline for effectiveness analysis. We trained all ablation models with the DUTS class and COCO9k datasets.", + "bbox": [ + 212, + 353, + 787, + 428 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effectiveness of PAG. PAG first uses intermediate image features from the FPN encoder to calculate full-pixel hyperassociations. Then, aggregation networks are applied to generate deep association features for the decoder process. PAG effectively utilizes deep learning to model pair-wise pixel associations, achieving higher-level association knowledge compared to previous image feature optimization strategies. As shown in the 3rd line of Table 1, PAG shows significant performance boosts compared to the baseline model, with respective gains of $3.72\\%$ , $4.80\\%$ , $7.30\\%$ , and $1.31\\%$ in $S_{m}$ , $E_{\\xi}$ , $F_{\\beta}$ , and $\\mathcal{M}$ .", + "bbox": [ + 212, + 430, + 787, + 551 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Furthermore, to validate our approach of progressively enhancing image features with the previously generated association feature for improving hyperassociation calculation, we conduct an ablation experiment called Separate Association Generation (SAG) where association features are generated for three stages without image feature enhancements. As shown in the 2nd&3rd lines in Table 1, PAG outperforms SAG, indicating that our progressive enhancement design can obtain better hyperassociation.", + "bbox": [ + 212, + 551, + 787, + 657 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effectiveness of CAC. CAC aims to condense full-pixel hyperassociations in PAG by selecting corresponding pixels and their surrounding contexts. Results in the 3rd&5th lines of Table 1 show that introducing the CAC module improves the performance. Moreover, it reduces the multiply-accumulate operations (MACs) of aggregation networks from 91.38G in the full-pixel PAG to $77.19\\mathrm{G}^{1}$ . This indicates that utilizing correspondence estimation to condense the hyperassociations not only effectively reduces the computational burden but also helps obtain more accurate pixel associations.", + "bbox": [ + 212, + 659, + 787, + 779 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We also analyze CAC in detail. As deep association features are necessary for reliable correspondence estimation, CAC first pre-condenses the hyperasso-", + "bbox": [ + 212, + 781, + 787, + 811 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "L Li et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 11 + }, + { + "type": "page_footnote", + "text": "We input a group of 6 related $256 \\times 256$ images to measure MACs.", + "bbox": [ + 230, + 825, + 668, + 839 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/adef6c3da6b359b8d6a99e5a95463d00c32fcd0b9610ff3f7214d3e0c830edc4.jpg", + "table_caption": [ + "Table 2: Quantitative comparison of our model with other SOTA methods. DC, C9, and CS are DUTS class, COCO9k, and COCO-SEG training data, respectively. bold and underline mark the best and second-best excellent results, respectively." + ], + "table_footnote": [], + "table_body": "
MethodsTrainingCoCA [47]CoSal2015 [44]CoSOD3k [7]
SetSm↑Eξ↑Fβ↑M↓Sm↑Eξ↑Fβ↑M↓Sm↑Eξ↑Fβ↑M↓
GICDDC0.6580.7180.5130.1260.8440.8870.8440.0710.7970.8480.7700.079
GCoNetDC0.6730.7600.5440.1100.8450.8880.8470.0680.8020.8600.7780.071
GCoNet+DC0.6910.7860.5740.1130.8750.9180.8760.0540.8280.8810.8070.068
CONDADC0.7170.7740.6000.1020.8900.9260.8940.0490.8320.8730.8070.067
ICNetC90.6540.7040.5130.1470.8570.9010.8580.0580.7940.8450.7620.089
DCFMC90.7100.7830.5980.0850.8380.8930.8560.0670.8090.8740.8050.067
GCoNet+C90.7170.7980.6050.0980.8530.9020.8570.0730.8190.8770.7960.075
CONDAC90.7300.8010.6220.0920.8650.9100.8750.0590.8250.8770.8100.068
CADCDC+C90.6800.7440.5490.1330.8670.9060.8650.0640.8150.8540.7780.088
DMTDC+C90.7250.8000.6190.1080.8970.9360.9050.0450.8510.8950.8350.063
GCoNet+DC+C90.7340.8080.6260.0880.8760.9200.8800.0570.8390.8940.8220.064
CONDADC+C90.7570.8250.6750.0920.9040.9400.9120.0420.8570.8990.8440.060
UGEMDC+CS0.7260.8080.5990.0960.8850.9350.8820.0510.8530.9110.8290.060
GCoNet+DC+CS0.7380.8140.6370.0810.8810.9260.8910.0550.8430.9010.8340.061
CONDADC+CS0.7630.8390.6850.0890.9000.9380.9080.0450.8620.9110.8530.056
", + "bbox": [ + 217, + 188, + 785, + 400 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "ciations using a maximum similarity approach to obtain initial deep association features and then performs further condensation based on the correspondences predicted by these initial deep association features. The hyperassociation condensation process with only the pre-condensation operation is called Similarity-induced Association Condensation (SAC). In Table 1, SAC only brings slight performance improvements due to the heuristic nature of the correspondence estimation. Nevertheless, SAC can provide the initial association feature for CAC to predict reliable correspondence.", + "bbox": [ + 212, + 410, + 784, + 530 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of OCC. OCC provides self-supervision for CAC to enable more precise correspondence estimation. Results in the 5th&7th lines of Table 1 show OCC further improves the performance by leveraging more precise correspondence estimations to condense hyperassociations effectively. In addition, we conducted an ablation experiment to validate our object-aware design in CAC by replacing OCC with full-pixel cycle consistency (FCC) loss. Comparing the 5th&6th lines of Table 1, FCC leads to a notable performance decrease due to background pixels disrupting the correspondence learning.", + "bbox": [ + 212, + 531, + 784, + 652 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Visualization of Correspondence Estimations. We present some visual samples of correspondence estimations for some co-salient pixels in Figure 4. Our semantic correspondence estimations are meaningful and can effectively depict the common attributes of co-salient objects at the pixel level.", + "bbox": [ + 212, + 652, + 784, + 712 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 215, + 737, + 619, + 752 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We compare our model with eight recent SOTA methods, i.e. GICD [47], ICNet [13], GCoNet [8], CADC [46], DCFM [41], DMT [16], UGEM [38], and GCoNet+ [51]. We directly utilize their officially released saliency maps for comparison. To ensure fairness, we trained our model with different combinations of three training datasets, following [51], to align with other compared methods. We", + "bbox": [ + 212, + 763, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "CONDA", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7d0b384ecf46d41b5b3fd790f7ba1310cf13090bae0c24b23cda3f353d2fa847.jpg", + "image_caption": [ + "Fig. 5: Qualitative comparisons of our model with other SOTA methods." + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 782, + 401 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "denote three training datasets, i.e. DUTS class [47], COCO9k [18], and COCOSEG [37], as DC, C9, and CS, respectively, for convenience. Our training sets include DC, C9, DC+C9, DC+CS. As shown in Table 2, we can observe that our model achieves the best performance with each training set in most benchmark datasets. What is even more exciting is that we achieve excellent results in the most challenging CoCA dataset, surpassing the second-best models by large margins, e.g. $2.5\\%$ $S_{m}$ , $2.5\\%$ $E_{\\xi}$ , and $4.8\\%$ $\\mathrm{F}_{\\beta}$ with the DC+CS training set.", + "bbox": [ + 212, + 429, + 787, + 536 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We also present visual comparisons in Figure 5. Our model can accurately detect co-salient objects in complex scenarios, such as irregularly shaped accordions accompanied by extraneous objects (people). However, other models easily fail to accurately segment co-salient objects.", + "bbox": [ + 212, + 537, + 787, + 598 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 623, + 359, + 638 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper proposes a deep association learning strategy for CoSOD that directly embeds hyperassociations into deep association features. Correspondence estimation is also introduced to condense hyperassociations, enabling the exploration of pixel-level correspondences for CoSOD. We also utilize an object-aware cycle consistency loss to further refine correspondence estimations. Extensive experiments have verified the effectiveness of our method.", + "bbox": [ + 212, + 657, + 787, + 747 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported in part by Key-Area Research and Development Program of Guangdong Province under Grant 2021B0101200001, the National Natural Science Foundation of China under Grant 62136007, 62036011, U20B2065, 6202781, 62036005, the Key R&D Program of Shaanxi Province under Grant 2021ZDLGY01-08, and the MBZUAI-WIS Joint Program for AI Research under Grants WIS P008 and P009.", + "bbox": [ + 212, + 750, + 803, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "L Li et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Achanta, R., Hemami, S., Estrada, F., Susstrunk, S.: Frequency-tuned salient region detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1597-1604 (2009)", + "2. Chen, W., Xu, H., Zhou, Z., Liu, Y., Sun, B., Kang, W., Xie, X.: Costformer: Cost transformer for cost aggregation in multi-view stereo. arXiv preprint arXiv:2305.10320 (2023)", + "3. Cheng, M.M., Warrell, J., Lin, W.Y., Zheng, S., Vineet, V., Crook, N.: Efficient salient region detection with soft image abstraction. In: Int. Conf. Comput. Vis. pp. 1529-1536 (2013)", + "4. Cho, S., Hong, S., Jeon, S., Lee, Y., Sohn, K., Kim, S.: Cats: Cost aggregation transformers for visual correspondence 34, 9011-9023 (2021)", + "5. Fan, D.P., Cheng, M.M., Liu, Y., Li, T., Borji, A.: Structure-measure: A new way to evaluate foreground maps. In: Int. Conf. Comput. Vis. pp. 4548-4557 (2017)", + "6. Fan, D.P., Gong, C., Cao, Y., Ren, B., Cheng, M.M., Borji, A.: Enhanced-alignment measure for binary foreground map evaluation. In: IJCAI. pp. 698-704 (2018)", + "7. Fan, D.P., Li, T., Lin, Z., Ji, G.P., Zhang, D., Cheng, M.M., Fu, H., Shen, J.: Re-thinking co-salient object detection. IEEE Trans. Pattern Anal. Mach. Intell. 44(8), 4339-4354 (2021)", + "8. Fan, Q., Fan, D.P., Fu, H., Tang, C.K., Shao, L., Tai, Y.W.: Group collaborative learning for co-salient object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12288-12298 (2021)", + "9. Fu, H., Cao, X., Tu, Z.: Cluster-based co-saliency detection. IEEE Trans. Image Process. 22(10), 3766-3778 (2013)", + "10. Han, J., Cheng, G., Li, Z., Zhang, D.: A unified metric learning-based framework for co-saliency detection. IEEE Trans. Circuit Syst. Video Technol. 28(10), 2473-2483 (2017)", + "1. Hong, S., Cho, S., Nam, J., Lin, S., Kim, S.: Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In: Eur. Conf. Comput. Vis. pp. 108-126 (2022)", + "2. Hong, S., Nam, J., Cho, S., Hong, S., Jeon, S., Min, D., Kim, S.: Neural matching fields: Implicit representation of matching fields for visual correspondence 35, 13512-13526 (2022)", + "3. Jin, W.D., Xu, J., Cheng, M.M., Zhang, Y., Guo, W.: Icnet: Intra-saliency correlation network for co-saliency detection. Adv. Neural Inform. Process. Syst. (2020)", + "4. Kim, S., Min, J., Cho, M.: Transformers: Match-to-match attention for semantic correspondence. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8697-8707 (2022)", + "5. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "6. Li, L., Han, J., Zhang, N., Liu, N., Khan, S., Cholakkal, H., Anwer, R.M., Khan, F.S.: Discriminative co-saliency and background mining transformer for co-salient object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7247-7256 (2023)", + "7. Lin, T.Y., Dólar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2117-2125 (2017)", + "8. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dóllar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Eur. Conf. Comput. Vis. pp. 740-755 (2014)" + ], + "bbox": [ + 225, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "CONDA", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Liu, J.J., Hou, Q., Cheng, M.M., Feng, J., Jiang, J.: A simple pooling-based design for real-time salient object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3917-3926 (2019)", + "20. Liu, N., Han, J.: Dhsnet: Deep hierarchical saliency network for salient object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 678-686 (2016)", + "21. Liu, N., Han, J., Yang, M.H.: Picanet: Learning pixel-wise contextual attention for saliency detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3089-3098 (2018)", + "22. Liu, N., Luo, Z., Zhang, N., Han, J.: Vst++: Efficient and stronger visual saliency transformer. IEEE Trans. Pattern Anal. Mach. Intell. (2024)", + "23. Liu, N., Zhang, N., Wan, K., Shao, L., Han, J.: Visual saliency transformer. In: Int. Conf. Comput. Vis. pp. 4722-4732 (2021)", + "24. Liu, Y., Zhu, L., Yamada, M., Yang, Y.: Semantic correspondence as an optimal transport problem. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4463-4472 (2020)", + "25. Liu, Y., Liu, N., Yao, X., Han, J.: Intermediate prototype mining transformer for few-shot semantic segmentation 35, 38020-38031 (2022)", + "26. Luo, Z., Liu, N., Zhao, W., Yang, X., Zhang, D., Fan, D.P., Khan, F., Han, J.: Vscode: General visual salient and camouflaged object detection with 2d prompt learning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2024)", + "27. Min, J., Kang, D., Cho, M.: Hypercorrelation squeeze for few-shot segmentation. In: Int. Conf. Comput. Vis. pp. 6941-6952 (2021)", + "28. Moon, S., Sohn, S.S., Zhou, H., Yoon, S., Pavlovic, V., Khan, M.H., Kapadia, M.: Hm: Hybrid masking for few-shot segmentation. In: Eur. Conf. Comput. Vis. pp. 506-523. Springer (2022)", + "29. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library 32 (2019)", + "30. Shen, X., Darmon, F., Efros, A.A., Aubry, M.: Ransac-flow: generic two-stage image alignment. In: Eur. Conf. Comput. Vis. pp. 618-637. Springer (2020)", + "31. Shi, X., Huang, Z., Bian, W., Li, D., Zhang, M., Cheung, K.C., See, S., Qin, H., Dai, J., Li, H.: Videoflow: Exploiting temporal cues for multi-frame optical flow estimation. arXiv preprint arXiv:2303.08340 (2023)", + "32. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: Int. Conf. Learn. Represent. (2015)", + "33. Su, Y., Deng, J., Sun, R., Lin, G., Su, H., Wu, Q.: A unified transformer framework for group-based segmentation: Co-segmentation, co-saliency detection and video salient object detection. IEEE Trans. Multimedia (2023)", + "34. Sun, G., Liu, Y., Tang, H., Chhatkuli, A., Zhang, L., Van Gool, L.: Mining relations among cross-frame affinities for video semantic segmentation. In: Eur. Conf. Comput. Vis. pp. 522-539. Springer (2022)", + "35. Truong, P., Danelljan, M., Yu, F., Van Gool, L.: Warp consistency for unsupervised learning of dense correspondences. In: Int. Conf. Comput. Vis. pp. 10346-10356 (2021)", + "36. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Adv. Neural Inform. Process. Syst. 30 (2017)", + "37. Wang, C., Zha, Z.J., Liu, D., Xie, H.: Robust deep co-saliency detection with group semantic. In: AAAI. vol. 33, pp. 8917-8924 (2019)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "L Li et al.", + "bbox": [ + 271, + 114, + 339, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "38. Wu, Y., Song, H., Liu, B., Zhang, K., Liu, D.: Co-salient object detection with uncertainty-aware group exchange-masking. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19639-19648 (2023)", + "39. Xiong, Z., Li, H., Zhu, X.X.: Doubly deformable aggregation of covariance matrices for few-shot segmentation. In: Eur. Conf. Comput. Vis. pp. 133-150. Springer (2022)", + "40. Xu, G., Wang, X., Ding, X., Yang, X.: Iterative geometry encoding volume for stereo matching. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 21919-21928 (2023)", + "41. Yu, S., Xiao, J., Zhang, B., Lim, E.G.: Democracy does matter: Comprehensive feature mining for co-salient object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 979-988 (2022)", + "42. Zhang, D., Fu, H., Han, J., Borji, A., Li, X.: A review of co-saliency detection algorithms: Fundamentals, applications, and challenges. ACM Trans Intell Syst Technol 9(4), 1-31 (2018)", + "43. Zhang, D., Han, J., Han, J., Shao, L.: Cosaliency detection based on intrasaliency prior transfer and deep intersaliency mining. IEEE Trans. Neural Networks Learn. Syst. 27(6), 1163-1176 (2015)", + "44. Zhang, D., Han, J., Li, C., Wang, J., Li, X.: Detection of co-salient objects by looking deep and wide. Int. J. Comput. Vis. 120(2), 215-232 (2016)", + "45. Zhang, D., Meng, D., Han, J.: Co-saliency detection via a self-paced multiple-instance learning framework. IEEE Trans. Pattern Anal. Mach. Intell. 39(5), 865-878 (2016)", + "46. Zhang, N., Han, J., Liu, N., Shao, L.: Summarize and search: Learning consensus-aware dynamic convolution for co-saliency detection. In: Int. Conf. Comput. Vis. pp. 4167-4176 (2021)", + "47. Zhang, Z., Jin, W., Xu, J., Cheng, M.M.: Gradient-induced co-saliency detection. In: Eur. Conf. Comput. Vis. pp. 455-472 (2020)", + "48. Zhao, D., Song, Z., Ji, Z., Zhao, G., Ge, W., Yu, Y.: Multi-scale matching networks for semantic correspondence. In: Int. Conf. Comput. Vis. pp. 3354-3364 (2021)", + "49. Zhao, J.X., Liu, J.J., Fan, D.P., Cao, Y., Yang, J., Cheng, M.M.: Egnet: Edge guidance network for salient object detection. In: Int. Conf. Comput. Vis. pp. 8779-8788 (2019)", + "50. Zhao, W., Zhang, J., Li, L., Barnes, N., Liu, N., Han, J.: Weakly supervised video salient object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 16826-16835 (2021)", + "51. Zheng, P., Fu, H., Fan, D.P., Fan, Q., Qin, J., Tai, Y.W., Tang, C.K., Van Gool, L.: Gconet+: A stronger group collaborative co-salient object detector. IEEE Trans. Pattern Anal. Mach. Intell. (2023)", + "52. Zhuge, M., Fan, D.P., Liu, N., Zhang, D., Xu, D., Shao, L.: Salient object detection via integrity learning. IEEE Trans. Pattern Anal. Mach. Intell. 45(3), 3738-3752 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 729 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "CONDA", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/CONDA_ Condensed Deep Association Learning for Co-Salient Object Detection./4b7d9a37-9ffb-4ebc-9e75-7e4ada571495_model.json b/2024/CONDA_ Condensed Deep Association Learning for Co-Salient Object Detection./4b7d9a37-9ffb-4ebc-9e75-7e4ada571495_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2371ae19c03cd0ceea66b11dca748346a800a15e --- /dev/null +++ b/2024/CONDA_ Condensed Deep Association Learning for Co-Salient Object Detection./4b7d9a37-9ffb-4ebc-9e75-7e4ada571495_model.json @@ -0,0 +1,2522 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.227, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "CONDA: Condensed Deep Association Learning for Co-Salient Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.212, + 0.782, + 0.258 + ], + "angle": 0, + "content": "Long Li\\(^{1}\\), Nian Liu\\(^{3,*}\\), Dingwen Zhang\\(^{1}\\), Zhongyu Li\\(^{4}\\), Salman Khan\\(^{3,5}\\), Rao Anwer\\(^{3}\\), Hisham Cholakkal\\(^{3}\\), Junwei Han\\(^{1,2,*}\\), and Fahad Shahbaz Khan\\(^{3,6}\\)" + }, + { + "type": "text", + "bbox": [ + 0.365, + 0.27, + 0.639, + 0.285 + ], + "angle": 0, + "content": "1 Northwestern Polytechnical University" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.284, + 0.774, + 0.299 + ], + "angle": 0, + "content": "\\(^{2}\\) Institute of Artificial Intelligence, Hefei Comprehensive National Science Center" + }, + { + "type": "text", + "bbox": [ + 0.309, + 0.298, + 0.694, + 0.312 + ], + "angle": 0, + "content": "3 Mohamed bin Zayed University of Artificial Intelligence" + }, + { + "type": "text", + "bbox": [ + 0.408, + 0.312, + 0.596, + 0.326 + ], + "angle": 0, + "content": "\\(^{4}\\) Xi'an Jiaotong University" + }, + { + "type": "text", + "bbox": [ + 0.392, + 0.326, + 0.612, + 0.34 + ], + "angle": 0, + "content": "5 Australian National University" + }, + { + "type": "text", + "bbox": [ + 0.403, + 0.34, + 0.6, + 0.355 + ], + "angle": 0, + "content": "\\(^{6}\\) CVL, Linköping University" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.394, + 0.744, + 0.699 + ], + "angle": 0, + "content": "Abstract. Inter-image association modeling is crucial for co-salient object detection. Despite satisfactory performance, previous methods still have limitations on sufficient inter-image association modeling. Because most of them focus on image feature optimization under the guidance of heuristically calculated raw inter-image associations. They directly rely on raw associations which are not reliable in complex scenarios, and their image feature optimization approach is not explicit for inter-image association modeling. To alleviate these limitations, this paper proposes a deep association learning strategy that deploys deep networks on raw associations to explicitly transform them into deep association features. Specifically, we first create hyperassociations to collect dense pixel-pair-wise raw associations and then deploys deep aggregation networks on them. We design a progressive association generation module for this purpose with additional enhancement of the hyperassociation calculation. More importantly, we propose a correspondence-induced association condensation module that introduces a pretext task, i.e. semantic correspondence estimation, to condense the hyperassociations for computational burden reduction and noise elimination. We also design an object-aware cycle consistency loss for high-quality correspondence estimations. Experimental results in three benchmark datasets demonstrate the remarkable effectiveness of our proposed method with various training settings. The code is available at: https://github.com/dragonlee258079/CONDA." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.712, + 0.731, + 0.727 + ], + "angle": 0, + "content": "Keywords: Co-salient Object Detection \\(\\cdot\\) Deep Association Learning" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.753, + 0.376, + 0.769 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.786, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Co-Salient Object Detection (CoSOD) aims to segment salient objects that appear commonly across a group of related images. Compared to traditional Salient" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.678, + 0.841 + ], + "angle": 0, + "content": "* Corresponding authors: {liunian228, junweihan2010}@gmail.com" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.342, + 0.128 + ], + "angle": 0, + "content": "L Li et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.145, + 0.589, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.147, + 0.784, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.307, + 0.788, + 0.392 + ], + "angle": 0, + "content": "Fig. 1: Difference of raw-association-based image feature optimization strategy (a) and our proposed deep association learning strategy (b). Our deep association learning deploys deep learning networks on raw associations to achieve deep association features. We also present visual samples of our calculated raw associations (Raw Asso.), optimized image feature (Opt. Fea.), and our generated deep association features (Asso. Fea.) in (c)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.786, + 0.431 + ], + "angle": 0, + "content": "Object Detection (SOD) [19-23, 26, 49, 50, 52], CoSOD is a more challenging task because it requires sufficient inter-image association modeling [9]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.704 + ], + "angle": 0, + "content": "Recently, many advanced works [8,13,16,33,38,41,46,47,51] have emerged and achieved impressive performance. These methods first use related image features to acquire raw inter-image associations (also called consensus representations), and then leverage them as guidance to optimize each image feature, as shown in Figure 1 (a). This approach enables the final image feature to implicitly capture inter-image cues, thereby achieving the purpose of inter-image association modeling. However, we find this raw-association-based feature optimization strategy still has two limitations: i) they directly rely on raw associations, which are acquired in heuristic manners, such as pixel-wise [8,33,41,51], region-wise [16,46], or image-wise [13] similarity measurements. Although high-quality raw associations can be derived from high-level semantic information in image features, their revelation of common saliency regions still relies on similarity measures, which are unreliable when encountering complex scenarios, such as significant differences between co-salient objects or high foreground-background similarity. ii) the primary focus of building their deep models is on optimizing image features. Compared to directly modeling association relationship, image feature optimization is not an explicit approach for inter-image association modeling and will increase the learning difficulty." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To alleviate these limitations, we propose a deep association learning strategy for CoSOD, as shown in Figure 1 (b). Instead of directly using raw associations to optimize image features, we deploy deep networks on raw associations to learn deep association features. This is a more explicit strategy for inter-image association modeling. Moreover, our deep association features can capture high-level inter-image association knowledge, making them more robust in complex scenarios than raw associations, as shown in Figure 1 (c). Technically, we start by collecting all pixel-wise raw associations across the entire image group as hyperassociations. Then, we propose a Progressive Association Generation (PAG)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CONDA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "module to transform hyperassociations into deep association features. PAG progressively generates association features on varying scales, allowing us to use previous association features to enhance the hyperassociation calculation at the next scale, thereby improving the association quality from the very beginning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.343 + ], + "angle": 0, + "content": "Although deep association learning strategy allows more sufficient interimage association modeling, it significantly increases the computational burden and reduces the practicality of this approach. Additionally, this study finds that it is not necessary to utilize all pixel associations to generate deep association features. In fact, there are even some noisy pixels that negatively impact the quality of the deep association features. Therefore, we propose a method based on Correspondence-induced Association Condensation (CAC) to condense the original full-pixel hyperassociations. This not only alleviates the computational burden but also further enhances the quality of deep association features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.787, + 0.509 + ], + "angle": 0, + "content": "Specifically, CAC performs the condensation operation by selectively associating pixels that have semantic correspondence in other images, as well as their surrounding contextual pixels, thereby creating lightweight yet more accurate hyperassociations. Here, we introduce a pretext task, i.e. semantic correspondence estimation, into the CoSOD, not only improving the model performance but also delving deeper into the essence of CoSOD. Co-salient objects inherently possess object-level semantic correspondence. However, in this paper, we aim to further explore the finer pixel-level correspondence. Although highly accurate correspondence estimation remains a challenge, we believe it will pave a new way for CoSOD research. We also provide an object-aware cycle consistency (OCC) loss to aid in learning correspondences within co-salient pixels." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.51, + 0.663, + 0.524 + ], + "angle": 0, + "content": "In summary, the contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.527, + 0.784, + 0.586 + ], + "angle": 0, + "content": "- For the first time, we introduce a deep association learning approach for CoSOD, applying deep networks to transform raw associations into deep association features for sufficient inter-image association modeling. Specifically, we develop a CONdensed Deep Association (CONDA) learning model." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.587, + 0.784, + 0.631 + ], + "angle": 0, + "content": "- We propose a PAG module to progressively generate deep association features. It enhances image features with previous association features to improve hyperassociation calculation." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.632, + 0.784, + 0.69 + ], + "angle": 0, + "content": "- We introduce semantic correspondence estimation into the CoSOD task to condense the original hyperassociation for alleviating the computational burden and further improving the performance. We also propose an OCC loss for effective correspondence estimation." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.691, + 0.784, + 0.735 + ], + "angle": 0, + "content": "- Experimental results demonstrate that our model achieves significantly improved state-of-the-art performance on three benchmark datasets across different training settings." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.527, + 0.784, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.757, + 0.388, + 0.773 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.787, + 0.496, + 0.803 + ], + "angle": 0, + "content": "2.1 Co-Salient Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Recently, there has been a surge of excellent methods for CoSOD [8,10,13,16,33, 38,41-43,45-47,51]. These methods initially acquire raw inter-image associations" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.127 + ], + "angle": 0, + "content": "L Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "with related image features and then utilize them to optimize each image feature. Most methods generate pixel-wise [8, 33, 41, 51], region-wise [16, 46], and image-wise [13] raw associations through similarity-based manners, e.g. inner product calculations between image features. Even for transformer-based methods [16, 33], they also rely on inner product calculation to produce attention maps [36] as raw associations. The image feature optimization manners include spatial or channel calibration [8, 13, 51], dynamic convolution [41, 46], feature fusion [16, 41], and cross-attention [16, 33], etc. However, they lack the learning of high-level association knowledge and heavily focus on optimized image features. Unlike them, this paper proposes a new research direction that deploys deep networks on associations to achieve deep association features for CoSOD." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.335, + 0.519, + 0.351 + ], + "angle": 0, + "content": "2.2 Inter-image Relation Modeling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.789, + 0.573 + ], + "angle": 0, + "content": "Apart from CoSOD, other tasks necessitating the consideration of inter-image relationships, such as few-shot segmentation [25, 27, 28, 39], stereo matching [2, 40], video semantic segmentation [34], etc. These tasks have made significant advancements recently by effectively modeling inter-image relations. Most of these methods [2, 4, 12, 27, 39, 40] initially create cost volumes to capture dense inter-image pixel-wise similarities and subsequently use various modules to convert these hyper volumes into task-specific features. Our approach distinguishes itself from prior methods in three aspects. Firstly, most of them create 4D cost volumes between two images while we create 6D hyperassociations between all related images. Secondly, they calculate hyper volumes based on original image features, whereas We propose PAG to progressively enhance image features for better hyperassociation calculation. Last and most importantly, they rely on the full-pixel cost volume. We consider the condensation of hyperassociations using semantic correspondences to eliminate noisy pixel associations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.594, + 0.573, + 0.61 + ], + "angle": 0, + "content": "2.3 Semantic Correspondence Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.756 + ], + "angle": 0, + "content": "Semantic correspondence estimation [24] aims to establish reliable pixel correspondences between different instances of the same object category. Most works performed this task using fully supervised training [14, 24]. Some recent works have utilized unsupervised learning with photometric, forward-backward consistency, and warp-supervision losses [30, 35, 48]. However, they implement these losses on the entire image, where background pixels may affect the performance. In this paper, we introduce this task to condense hyperassociations for Co-SOD and tailor the cycle consistency loss by only applying it to co-salient pixels, hence effectively avoiding the influence of background and extraneous objects." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.778, + 0.429, + 0.796 + ], + "angle": 0, + "content": "3 Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.843 + ], + "angle": 0, + "content": "As shown in Figure 2, CONDA integrates the deep association learning process into an FPN framework. Specifically, given a group of related images \\(\\{I_i\\}_{i=1}^N\\), we" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CONDA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.784, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.406, + 0.788, + 0.477 + ], + "angle": 0, + "content": "Fig. 2: Overall flowchart of our CONDA model. Specifically, CONDA first utilizes the image features to calculate hyperassociations. Then, the full-pixel hyperassociations are condensed by CAC and fed into the aggregation networks to achieve deep association features. These features are then used in the FPN decoder process for the final prediction. To be concise, only three related images are shown." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.788, + 0.532 + ], + "angle": 0, + "content": "first input them into a VGG-16 [32] backbone to collect its intermediate features for PAG and FPN decoding. In detail, we collect all features of the last three stages for PAG and the last feature of each stage for the FPN decoder as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.338, + 0.544, + 0.785, + 0.583 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbb {F} ^ {P} = \\left\\{\\boldsymbol {F} _ {s, l} \\mid s \\in \\{3, \\dots , 5 \\}, l \\in \\{1, \\dots , L _ {s} \\} \\right\\}, \\\\ \\mathbb {F} ^ {D} = \\left\\{\\boldsymbol {F} _ {s, L _ {s}} \\mid s \\in \\{1, \\dots 5 \\} \\right\\}, \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.788, + 0.654 + ], + "angle": 0, + "content": "where \\(\\mathbb{F}^P\\) and \\(\\mathbb{F}^D\\) are the feature collections for PAG and FPN decoder, respectively. \\(F_{s,l}\\in \\mathbb{R}^{N\\times H_s\\times W_s\\times C_s}\\) represents the VGG feature of the \\(l\\)-th layer in the \\(s\\)-th stage. There are in total \\(L_{s}\\) layers in the \\(s\\)-th stage. \\(H_{s}\\), \\(W_{s}\\) and \\(C_s\\) represent the height, width, and channel of the \\(s\\)-th stage, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.654, + 0.788, + 0.7 + ], + "angle": 0, + "content": "Then, we input \\(\\mathbb{F}^P\\) into PAG to calculate hyperassociations and generate deep association features \\(\\{\\pmb {F}_s^A\\in \\mathbb{R}^{N\\times H_s\\times W_s\\times C_s}\\}_{s = 3}^5\\). Finally, these association features will be fused with \\(\\mathbb{F}^D\\) for the FPN decoder process, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.712, + 0.643, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {F} _ {s, L _ {s}} = \\boldsymbol {F} _ {s, L _ {s}} + \\phi (\\boldsymbol {F} _ {s} ^ {A}), \\quad s = 3, \\dots , 5, \\\\ \\boldsymbol {F} = \\operatorname {D e c o d e r} \\left(\\left\\{\\boldsymbol {F} _ {s, L _ {s}} \\right\\} _ {s = 1} ^ {5}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.763, + 0.788, + 0.794 + ], + "angle": 0, + "content": "where \\(\\phi\\) is a convolution layer. \\(\\pmb{F} \\in \\mathbb{R}^{N \\times H \\times W \\times C}\\) is the final feature for the final co-saliency prediction. We adopt BCE and IoU losses for supervision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "The rest of this section will introduce PAG with full-pixel hyperassociation and the condensation of hyperassociations by plugging the Correspondence-induced Association Condensation (CAC) module into PAG." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.128 + ], + "angle": 0, + "content": "L Li et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.555, + 0.162 + ], + "angle": 0, + "content": "3.1 Progressive Association Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.784, + 0.218 + ], + "angle": 0, + "content": "Our deep association learning involves two steps: 1) acquiring the raw hyper-associations \\(\\{A_s\\}_{s=3}^5\\) from three stages; 2) employing aggregation networks on \\(\\{A_s\\}_{s=3}^5\\) to obtain association features \\(\\{F_s^A\\}_{s=3}^5\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.219, + 0.784, + 0.293 + ], + "angle": 0, + "content": "Earlier methods [4, 11, 31] directly utilize the original backbone features, i.e. \\(\\mathbb{F}^P\\), to calculate inter-image interactions, such as the so-called cost volume. We argue that hyperassociations derived straight from backbone features might obstruct further improvement of deep association learning, given that the current backbone is pre-trained without any consideration for inter-image associations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.294, + 0.785, + 0.431 + ], + "angle": 0, + "content": "To alleviate this problem, we propose the PAG module to progressively generate pyramid association features so that we can utilize the high-level association feature, e.g. \\( \\pmb{F}_{s+1}^{A} \\) from the \\( (s+1) \\)-th stage, to enhance the neighbouring low-level VGG features in \\( \\mathbb{F}^P \\), e.g. \\( \\pmb{F}_{s,l} \\), for attaining association-enhanced features \\( \\hat{\\pmb{F}}_{s,l} \\), based on which we can calculate high-quality hyperassociations \\( \\pmb{A}_s \\) in the \\( s \\)-th stage. After that, we execute the subsequent aggregation networks on \\( \\pmb{A}_s \\) to achieve association features \\( \\pmb{F}_s^A \\), which will continue to enhance the VGG features of the next stage and carry out progressive association generation. The whole process of our PAG can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.444, + 0.586, + 0.461 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {A} _ {s} = \\operatorname {H A C} \\left(\\left\\{\\tilde {\\boldsymbol {F}} _ {s, l} \\right\\} _ {l = 1} ^ {L _ {s}}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.444, + 0.465, + 0.785, + 0.482 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} _ {s} ^ {A} = \\operatorname {A g g} \\left(\\boldsymbol {A} _ {s}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.486, + 0.647, + 0.505 + ], + "angle": 0, + "content": "\\[\n\\{\\hat {\\boldsymbol {F}} _ {s - 1, l} \\} _ {l = 1} ^ {L _ {s - 1}} = \\operatorname {E n h} \\left(\\left\\{\\boldsymbol {F} _ {s - 1, l} \\right\\} _ {l = 1} ^ {L _ {s - 1}}; \\boldsymbol {F} _ {s} ^ {A}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.784, + 0.563 + ], + "angle": 0, + "content": "where \\(s\\) ranges from 5 to 3 and \\(\\{\\hat{F}_{5,l}\\}_{l=1}^{L_5} = \\{\\pmb{F}_{5,l}\\}_{l=1}^{L_5}\\). The HAC, Agg, and Enh represent the hyperassociation calculation, aggregation network, and association-induced feature enhancement, respectively. Next, we explain them in detail." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.785, + 0.639 + ], + "angle": 0, + "content": "Hyperassociation Calculation. For each stage, we first compute the raw associations at each layer using the inner product between \\( l \\)-2 normalized association-enhanced features of \\( N \\) related images. After that, we stack the raw associations of all layers to form the final hyperassociation of this stage. The hyperassociation for the \\( s \\)-th stage, i.e. \\( A_{s} \\in \\mathbb{R}^{N \\times H_{s} \\times W_{s} \\times N \\times H_{s} \\times W_{s} \\times L_{s}} \\), can be calculated via:" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.652, + 0.785, + 0.717 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {A} _ {s} = \\operatorname {H A C} \\left(\\left\\{\\hat {\\boldsymbol {F}} _ {s, l} \\right\\} _ {l = 1} ^ {L _ {s}}\\right) \\\\ = \\operatorname {S t a c k} \\left(\\left\\{\\operatorname {R e L U} \\left(\\frac {\\hat {\\boldsymbol {F}} _ {s , l} \\cdot \\hat {\\boldsymbol {F}} _ {s , l} ^ {\\top}}{\\| \\hat {\\boldsymbol {F}} _ {s , l} \\| \\| \\hat {\\boldsymbol {F}} _ {s , l} ^ {\\top} \\|}\\right) \\right\\} _ {l = 1} ^ {L _ {s}}\\right), \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.731, + 0.784, + 0.779 + ], + "angle": 0, + "content": "where \\(\\hat{\\pmb{F}}_{s,l}\\cdot \\hat{\\pmb{F}}_{s,l}^{\\top}\\in \\mathbb{R}^{N\\times H_s\\times W_s\\times N\\times H_s\\times W_s}\\). The \\(\\top\\) indicates transposing the last dimension and the first three dimensions. \\(\\| \\cdot \\|\\) represents the \\(l\\)-2 norm. We employ ReLU to suppress noisy association values." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.779, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Aggregation Network. The raw hyperassociation \\( \\mathbf{A}_s \\in \\mathbb{R}^{N \\times H_s \\times W_s \\times N \\times H_s \\times W_s \\times L_s} \\) is a hypercube with a nested structure, where each pixel position is characterized by a 4D tensor \\( (N \\times H_s \\times W_s \\times L_s) \\). Each 4D tensor documents the associations of the respective pixel with all other pixels in \\( N \\) related images. For" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CONDA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "clarity, we designate the first and second \\( N \\times H_s \\times W_s \\) dimensions in \\( A_s \\) as the source and target dimensions, respectively. Although these 4D tensors are crucial for exploring the consensus information for co-saliency detection, they essentially comprise pixel-to-pixel similarity values, seen in (4), which may be suboptimal and unreliable in complex scenarios. Therefore, we propose using deep networks to transform these pixel-wise similarities into deep association features with contextual and high-order association knowledge. This has never been explored in previous CoSOD methods. This is implemented via context aggregations on \\( A_s \\) to squeeze these 4D tensors as \\( C_s \\)-dimensional vectors, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.296, + 0.786, + 0.312 + ], + "angle": 0, + "content": "\\[\nN \\times H _ {s} \\times W _ {s} \\times N \\times H _ {s} \\times W _ {s} \\times L _ {s} \\rightarrow N \\times H _ {s} \\times W _ {s} \\times C _ {s}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.323, + 0.788, + 0.444 + ], + "angle": 0, + "content": "In detail, we first deploy several association aggregation layers on \\( \\mathbf{A}_s \\) to progressively aggregate context information, enlarging \\( L_s \\) as \\( C_s \\), and eliminate the target \\( H_s \\times W_s \\) dimension in 4D tensors. Each aggregation layer consists of 2D convolution layers and a downsampling operation. Specifically, focusing on the first aggregation layer for technical explanation, we first aggregate context information in the target \\( H_s \\times W_s \\) dimension by applying a 2D convolution layer on all 4D tensors. The operations on the 4D tensor at pixel position \\( (h_i, w_i) \\) in image \\( I_i \\) can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.282, + 0.455, + 0.786, + 0.473 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {A} _ {s} ^ {1} (i, h _ {i}, w _ {i}, j, \\therefore , \\therefore): = \\mathcal {C} _ {1} \\left(\\boldsymbol {A} _ {s} (i, h _ {i}, w _ {i}, j, \\therefore , \\therefore)\\right), \\quad j = 1, \\dots , N, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.788, + 0.543 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}_1\\) is a \\(3\\times 3\\) 2D convolution layer. Here, \\(j\\) is the index of other related images in the 4D tensor, and \\(\\pmb {A}_s(i,h_i,w_i,j,:;:)\\in \\mathbb{R}^{H_s\\times W_s\\times L_s}\\) illustrates the associations between the pixel \\((h_i,w_i)\\) in \\(I_{i}\\) and all pixels in image \\(I_{j}\\). This interpretation applies to other similar symbols." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.544, + 0.787, + 0.575 + ], + "angle": 0, + "content": "Then, a downsampling operation \\(\\mathcal{D}\\), i.e. bilinear interpolation, is applied on \\(\\pmb{A}_s^1\\) to reduce the spatial dimension of the 4D tensor by a scaling factor:" + }, + { + "type": "equation", + "bbox": [ + 0.341, + 0.586, + 0.786, + 0.603 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {A} _ {s} ^ {2} (i, h _ {i}, w _ {i}, j, \\therefore , \\therefore ,:) = \\mathcal {D} (\\boldsymbol {A} _ {s} ^ {1} (i, h _ {i}, w _ {i}, j, \\therefore , \\therefore)), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.787, + 0.647 + ], + "angle": 0, + "content": "where \\(A_{s}^{2}(i,h_{i},w_{i},j,:,:) \\in \\mathbb{R}^{H_{s}^{\\prime} \\times W_{s}^{\\prime} \\times C_{s}^{\\prime}}\\), \\(H_{s}^{\\prime}\\) and \\(W_{s}^{\\prime}\\) are downsampled height and width. \\(C_{s}^{\\prime}\\) is the channel number after convolution \\(\\mathcal{C}_1\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.647, + 0.788, + 0.724 + ], + "angle": 0, + "content": "Finally, we also aggregate context information in the source \\( H_{s} \\times W_{s} \\) dimension. Specifically, we extract the values along the source dimension and channel dimension in \\( A_{s}^{2} \\) to form 4D tensors and apply a 2D convolution layer on them. For instance, \\( A_{s}^{2}(:, :, :, j, h_{j}, w_{j},:) \\in \\mathbb{R}^{N \\times H_{s} \\times W_{s} \\times C_{s}^{\\prime}} \\) is such a 4D tensor, where \\( (h_{j}, w_{j}) \\) is a pixel position in the target dimension. This can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.28, + 0.733, + 0.786, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {A} _ {s} ^ {3} (i,:,:) j, h _ {j}, w _ {j,:}) = \\mathcal {C} _ {2} \\left(\\boldsymbol {A} _ {s} ^ {2} (i,:,:) j, h _ {j}, w _ {j,:})\\right), \\quad i = 1, \\dots , N, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.761, + 0.718, + 0.776 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}_2\\) is a \\(3\\times 3\\) 2D convolution layer. \\(i\\) is the related image index." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.777, + 0.788, + 0.842 + ], + "angle": 0, + "content": "After several association aggregation layers, as shown in (6)-(8), we can achieve the aggregated association features with the target \\( H_{s} \\times W_{s} \\) dimension eliminated, denoted as \\( \\pmb{F}_{s}^{A^{\\prime}} \\in \\mathbb{R}^{N \\times H_{s} \\times W_{s} \\times N \\times C} \\). Subsequently, we average \\( \\pmb{F}_{s}^{A^{\\prime}} \\) on its second \\( N \\) dimension and obtain the final association features" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.342, + 0.127 + ], + "angle": 0, + "content": "L Li et al." + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.144, + 0.756, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.246, + 0.788, + 0.356 + ], + "angle": 0, + "content": "Fig. 3: Difference between full-pixel hyperassociation (a) and condensed hyperassociation (b). We provide an example of collecting the pixel associations from image \\( I_{j} \\) for a pixel \\( (h_{i},w_{i}) \\) in image \\( I_{i} \\). Full-pixel hyperassociation collects all pixel associations in \\( I_{j} \\), while our condensed hyperassociation only collects the associations of its correspondence pixel \\( (h_{j},w_{j}) \\) (red dot) and surrounding pixels (green dots). We first heuristically find an initial pixel \\( (h_j^0,w_j^0) \\) with a fixed surrounding window and then learn coordinate offsets to locate the optimized correspondence and surrounding pixels." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.363, + 0.48, + 0.381 + ], + "angle": 0, + "content": "\\(\\pmb{F}_s^A \\in \\mathbb{R}^{N \\times H_s \\times W_s \\times C}\\), formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.39, + 0.391, + 0.787, + 0.422 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} _ {s} ^ {A} = \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\boldsymbol {F} _ {s} ^ {A ^ {\\prime}} (:,:,:, j,:). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.43, + 0.788, + 0.511 + ], + "angle": 0, + "content": "Association-induced Feature Enhancement. Once we have obtained association feature \\( \\mathbf{F}_s^A \\) of the \\( s \\)-th stage, we will use it to enhance the VGG feature of the \\( (s - 1) \\)-th stage, i.e. \\( \\{\\mathbf{F}_{s - 1,l}\\}_{l = 1}^{L_{s - 1}} \\). Specifically, we upsample \\( \\mathbf{F}_s^A \\) to align the spatial size of features in the \\( (s - 1) \\)-th stage, and then add it to \\( \\{\\mathbf{F}_{s - 1,l}\\}_{l = 1}^{L_{s - 1}} \\) followed by a convolution layer, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.39, + 0.521, + 0.787, + 0.542 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {F}} _ {s - 1, l} = \\mathcal {C} _ {3} \\left(\\boldsymbol {F} _ {s - 1, l} + \\mathcal {U} \\left(\\boldsymbol {F} _ {s} ^ {A}\\right)\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.583 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}_3\\) and \\(\\mathcal{U}\\) represent a 2D convolution layer and the bilinear upsampling operation, respectively." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.604, + 0.684, + 0.619 + ], + "angle": 0, + "content": "3.2 Correspondence-induced Association Condensation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Although PAG based on full-pixel hyperassociations can deliver satisfactory performance in CoSOD, it also introduces substantial computational overhead. Additionally, we argue that for each pixel in an image, it is unnecessary to gather its associations with all pixels of other related images to form hyperassociations. Some pixel associations may even impair the final performance, such as those between ambiguous regions. To this end, this subsection tries to condense the original full-pixel hyperassociations to retain only informative pixel associations." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.735, + 0.788, + 0.796 + ], + "angle": 0, + "content": "This subsection will focus on explaining the condensation of pixel associations of a pixel (e.g. \\((h_i, w_i)\\) in \\(I_i\\)) to ones of the other images (e.g. image \\(I_j\\)), i.e. \\(\\mathbf{A}_s(i, h_i, w_i, j, :, :, :) \\in \\mathbb{R}^{H_s \\times W_s \\times L_s}\\), as shown in Figure 3. We will simplify the symbol \\(\\mathbf{A}_s(i, h_i, w_i, j, :, :, :)\\) as \\(\\mathbf{a}_s^j\\) in the subsequent text for convenience." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Specifically, CAC opts to select \\( K \\times K \\) (\\( K < H_{s}, K < W_{s} \\)) informative pixel associations from \\( \\pmb{a}_{s}^{j} \\) to form its condensed representation, i.e. \\( \\tilde{\\pmb{a}}_{s}^{j} \\in \\mathbb{R}^{K \\times K \\times L_{s}} \\). Thus, the entire condensed hyperassocaton can be symbolized as \\( \\pmb{A}_{s} \\in \\mathbb{R}^{N \\times H_{s} \\times W_{s} \\times N \\times L_{s}} \\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CONDA" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.312 + ], + "angle": 0, + "content": "\\(K \\times K \\times L_s\\). To ensure the proper selection of \\(K \\times K\\) pixels, we introduce a pretext task, i.e. semantic correspondence estimation [14,24]. This allows us to first locate the corresponding pixel of \\((h_i, w_i)\\) in image \\(I_j\\), i.e. \\((h_j, w_j)\\), and then combine \\((h_j, w_j)\\) with its surrounding pixels as the \\(K \\times K\\) pixel set. We design this approach based on the observation that the co-salient objects across \\(N\\) related images belong to the same semantic category, and the pixels within them should have semantic correspondences to each other, as shown in Figure 4. Therefore, the introduction of semantic correspondence in CAC not only improves the CoSOD performance but also delves deeper into the core nature of the CoSOD task. As far as we know, this is the first work to use semantic correspondence in the CoSOD task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.405 + ], + "angle": 0, + "content": "Correspondence Estimation. To estimate the correspondence pixel \\((h_j, w_j)\\) in \\(I_j\\) for \\((h_i, w_i)\\), we first identify an initial pixel \\((h_j^0, w_j^0)\\) via a heuristic approach. Subsequently, we produce a spatial offset to refine \\((h_j^0, w_j^0)\\) into \\((h_j, w_j)\\). To achieve this purpose, all initial correspondence pixels should be utilized to form the initial condensed hyperassociations, with which the initial deep association features can be generated for spatial offset prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.788, + 0.452 + ], + "angle": 0, + "content": "Specifically, we pick out \\((h_j^0, w_j^0)\\) that has the largest feature similarity value with \\((h_i, w_i)\\). Since we have calculated the feature similarities in \\(\\pmb{a}_s^j\\), the \\((h_j^0, w_j^0)\\) can be obtained as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.439, + 0.462, + 0.59, + 0.488 + ], + "angle": 0, + "content": "\\[\n\\bar {\\pmb {a}} _ {s} ^ {j} = \\sum_ {l = 1} ^ {L _ {s}} \\pmb {a} _ {s} ^ {j} (:,: l),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.401, + 0.481, + 0.785, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\left(h _ {j} ^ {0}, w _ {j} ^ {0}\\right) = \\underset {x, y} {\\operatorname {a r g m a x}} \\left(\\bar {\\boldsymbol {a}} _ {s} ^ {j} (x, y)\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.573 + ], + "angle": 0, + "content": "where \\(\\bar{\\pmb{a}}_s^j\\in \\mathbb{R}^{H_s\\times W_s}\\) is a similarity matrix obtained by eliminating the last dimension of \\(\\pmb {a}_s^j\\in \\mathbb{R}^{H_s\\times W_s\\times L_s}\\) through summation. argmax returns the coordinate of the maximum value." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.788, + 0.69 + ], + "angle": 0, + "content": "Next, we select \\(K \\times K\\) pixels within the square region centered around the initial pixels, e.g. \\((h_j^0, w_j^0)\\), to construct the initial condensed hyperassociation \\(\\tilde{\\pmb{A}}_s'\\). Then, we can feed it into the aggregation network, described in Sec 3.1, and achieve the initial aggregated association features \\(\\pmb{F}_s^{A'} \\in \\mathbb{R}^{N \\times H_s \\times W_s \\times N \\times C_s}\\). It can be regarded as the association features of each pixel to \\(N\\) other related images. We utilize the feature of \\((h_i, w_i)\\) to \\(I_j\\), i.e. \\(\\pmb{F}_s^{A'}(i, h_i, w_i, j,:) \\in \\mathbb{R}^{C_s}\\), to predict the offsets for \\((h_j, w_j)\\) via a linear layer, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.701, + 0.785, + 0.721 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Delta} _ {s} ^ {j} = \\mathcal {O} \\left(\\boldsymbol {F} _ {s} ^ {A ^ {\\prime}} (i, h _ {i}, w _ {i}, j,:)\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.733, + 0.788, + 0.812 + ], + "angle": 0, + "content": "where \\(\\mathcal{O}\\) is a linear layer for offset generation. \\(\\pmb{\\Delta}_{s}^{j} \\in \\mathbb{R}^{K \\times K \\times 2}\\) consists of \\(K \\times K\\) offsets, besides the center offset \\(\\pmb{\\Delta}_{s}^{j}(k_{c}, k_{c},:)\\) for correspondence estimation, i.e. refining \\((h_{j}^{0}, w_{j}^{0})\\) as \\((h_{j}, w_{j})\\), we also generate other offsets for surrounding pixel selection. Thus, the corresponding pixel \\((h_{j}, w_{j})\\) can be obtained by adding the offset to the initial pixel \\((h_{j}^{0}, w_{j}^{0})\\), formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.824, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\left(h _ {j}, w _ {j}\\right) = \\left(h _ {j} ^ {0}, w _ {j} ^ {0}\\right) + \\boldsymbol {\\Delta} _ {s} ^ {j} \\left(k _ {c}, k _ {c}, :\\right), \\tag {13}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.128 + ], + "angle": 0, + "content": "L Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.597, + 0.162 + ], + "angle": 0, + "content": "where \\((k_c, k_c)\\) is the center position of \\(K \\times K\\) square." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.162, + 0.788, + 0.224 + ], + "angle": 0, + "content": "Condensation Operation. Given the estimated correspondence pixel \\((h_j, w_j)\\) and other offsets in \\(\\pmb{\\Delta}_{s}^{j}\\), we can obtain the surrounding pixels and combine them with \\((h_j, w_j)\\) to form \\(K \\times K\\) pixel set. Their coordinates are stored in \\(\\pmb{n}_s^j \\in \\mathbb{R}^{K \\times K \\times 2}\\). This process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.258, + 0.236, + 0.786, + 0.26 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {n} _ {s} ^ {j} (x, y,:) = \\left(h _ {j}, w _ {j}\\right) + \\boldsymbol {\\Delta} _ {s} ^ {j} (x, y,:) \\quad x, y \\in \\{1, \\dots , K \\}; x, y \\neq k _ {c}, \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.258, + 0.257, + 0.42, + 0.274 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {n} _ {s} ^ {j} (k _ {c}, k _ {c};) = (h _ {j}, w _ {j}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.285, + 0.597, + 0.3 + ], + "angle": 0, + "content": "where \\((k_c, k_c)\\) is the center position of \\(K \\times K\\) square." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.3, + 0.785, + 0.331 + ], + "angle": 0, + "content": "Finally, we can perform the condensation operation via the index selection on \\( \\pmb{a}_s^j \\), formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.456, + 0.33, + 0.785, + 0.348 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {a}} _ {s} ^ {j} = \\boldsymbol {a} _ {s} ^ {j} \\left(\\boldsymbol {n} _ {s} ^ {j}\\right), \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.788, + 0.432 + ], + "angle": 0, + "content": "where \\(\\tilde{\\pmb{a}}_s^j\\) is the condensed representation of \\(\\pmb{a}_s^j\\), i.e. pixel associations of \\((h_i, w_i)\\) to image \\(I_j\\). Furthermore, we also illustrate the condensation for pixel associations of \\((h_i, w_i)\\) to all images, i.e. \\(\\pmb{a}_s \\in \\mathbb{R}^{N \\times H_s \\times W_s \\times L_s}\\), in Figure 2. By applying such a condensation process to all pixel associations, we can obtain the final condensed hyperassociation \\(\\tilde{\\pmb{A}}_s\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.455, + 0.571, + 0.47 + ], + "angle": 0, + "content": "3.3 Object-aware Cycle Consistency Loss" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.6 + ], + "angle": 0, + "content": "To achieve accurate correspondence estimations, applying effective supervisions on them is necessary. As explicit semantic correspondence annotations are not available, we can only rely on unsupervised losses by imposing correspondence-related constraints on estimated correspondences. Previous unsupervised approaches apply constraints to all pixels [30, 35, 48], including those on the background and extraneous objects that don't have mutual correspondences, hence harming the model effectiveness. To avoid this problem, we propose an object-aware constraint to only access losses on the co-salient pixels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.788, + 0.676 + ], + "angle": 0, + "content": "We propose an Object-aware Cycle Consistency (OCC) loss for the supervision of correspondence estimations in CoSOD. The cycle consistency can be explained as: if a co-salient pixel \\((h_i, w_i)\\) of image \\(I_i\\) corresponds to the pixel \\((h_j, w_j)\\) in image \\(I_j\\), then the pixel \\((h_j, w_j)\\) should also semantically correspond to pixel \\((h_i, w_i)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.677, + 0.788, + 0.797 + ], + "angle": 0, + "content": "Based on this cycle consistency constraint, we adopt image warping operations to conduct the OCC loss. Specifically, we first warp the image \\( I_{i}^{s} \\) (a resized \\( I_{i} \\) to align the scales in the \\( s \\)-th stage) as \\( I_{i\\rightarrow j}^{s} \\) using the \\( I_{i}^{s}\\to I_{j}^{s} \\) correspondence estimations. Next, we warp \\( I_{i\\rightarrow j}^{s} \\) backward to \\( I_{i\\rightarrow j\\rightarrow i}^{s} \\) using \\( I_{j}^{s}\\to I_{i}^{s} \\) correspondence estimations. Finally, we can utilize the SSIM loss between \\( I_{i}^{s} \\) and \\( I_{i\\rightarrow j\\rightarrow i}^{s} \\) to measure the cycle consistency for mutual correspondence pixels in \\( I_{i}^{s} \\) and \\( I_{j}^{s} \\). Moreover, to ensure the constraints are only conducted on co-salient objects, we mask the images with their ground truth masks, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.803, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} ^ {C} = \\frac {1}{N ^ {2}} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N} \\mathcal {L} _ {S S I M} \\left(I _ {i} ^ {s} \\cdot G _ {i} ^ {s}, I _ {i \\rightarrow j \\rightarrow i} ^ {s} \\cdot G _ {i} ^ {s}\\right), \\tag {16}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CONDA" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.173 + ], + "angle": 0, + "content": "Table 1: Ablation Study of our proposed modules. SAG, SAC, and FCC are ablation modules for PAG, CAC, and OCC, respectively." + }, + { + "type": "table", + "bbox": [ + 0.272, + 0.175, + 0.725, + 0.299 + ], + "angle": 0, + "content": "
IDModulesCoCA [47]
SAGPAGSACCACFCCOCCSm↑Eξ↑Fβ↑M↓
10.69360.76420.57290.1206
2✓0.72360.80290.63570.1106
3✓0.73080.81220.64590.1075
4✓✓0.73040.81230.65000.1085
5✓✓0.74730.81550.65910.0956
6✓✓0.73980.81380.65060.1013
7✓✓0.75700.82480.67510.0924
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.788, + 0.358 + ], + "angle": 0, + "content": "where \\( G_{i}^{s} \\) is the resized ground truth of image \\( I_{i} \\). The total OCC loss \\( \\mathcal{L}^C \\) is the sum of three stages, formulated as: \\( \\mathcal{L}^C = \\mathcal{L}_3^C +\\mathcal{L}_4^C +\\mathcal{L}_5^C \\). More details can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.379, + 0.377, + 0.396 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.41, + 0.533, + 0.425 + ], + "angle": 0, + "content": "4.1 Evaluation Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.788, + 0.526 + ], + "angle": 0, + "content": "We follow [8] to evaluate our model on three benchmark datasets, i.e. CoCA [47] (1295 images of 80 groups), CoSal2015 [44] (2015 images of 50 groups), and CoSOD3k [7] (3316 images of 160 groups). We adopt four widely-used metrics for the quantitative evaluation, i.e. Structure-measure \\((S_{m})\\) [5], Maximum enhanced-alignment measure \\((E_{\\xi})\\) [6], Maximum F-measure \\((\\mathrm{F}_{\\beta})\\) [1], and Mean Absolute Error \\((\\mathcal{M})\\) [3]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.545, + 0.458, + 0.56 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.659 + ], + "angle": 0, + "content": "To construct the training data, we follow [51] to use different combinations of three commonly used training datasets, i.e. DUTS class [47] (8250 images of 291 groups), COCO-9k [18] (9213 images of 65 groups), and COCO-SEG [37] (200,000 images of 78 groups), for a fair comparison with other state-of-the-art (SOTA) works. We also implement the synthesis strategy for the DUTS class dataset following [46]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.766 + ], + "angle": 0, + "content": "For training specifics, we employ the data augmentation strategy in [21] and use \\( 256 \\times 256 \\) as the input size for the network. We employ the Adam optimizer [15] with \\( \\beta_{1} = 0.9 \\) and \\( \\beta_{2} = 0.99 \\) to optimize the network. We train our CONDA model for 300 epochs, starting with an initial learning rate of \\( 1e - 4 \\), which is divided by 10 at the \\( 60,000^{th} \\) iteration. Our experiments are implemented based on PyTorch [29] on a single Tesla A40 GPU, with the batchsize set to \\( N = 6 \\). The hyperparameter \\( K \\) in CAC is set to 9." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.786, + 0.391, + 0.802 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We conduct ablation studies on the most challenging CoCA [47] dataset. To construct the baseline, we use the FPN [17] (with VGG-16 as encoder) as the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.127 + ], + "angle": 0, + "content": "L Li et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.79, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.284, + 0.788, + 0.34 + ], + "angle": 0, + "content": "Fig. 4: Visual samples for the correspondence estimations. Correspondences I, II, and III visually display estimated correspondences between the main image and three related images. Sparse co-salient pixels were selected and connected to their corresponding pixels using colored lines for clear visualization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.788, + 0.429 + ], + "angle": 0, + "content": "foundational segmentation network and enhance it with the Region-to-Region correlation module (R2R) [16] to simply capture inter-image connections. Then, as shown in Table 1, we incrementally incorporate PAG, CAC, and OCC into the baseline for effectiveness analysis. We trained all ablation models with the DUTS class and COCO9k datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.552 + ], + "angle": 0, + "content": "Effectiveness of PAG. PAG first uses intermediate image features from the FPN encoder to calculate full-pixel hyperassociations. Then, aggregation networks are applied to generate deep association features for the decoder process. PAG effectively utilizes deep learning to model pair-wise pixel associations, achieving higher-level association knowledge compared to previous image feature optimization strategies. As shown in the 3rd line of Table 1, PAG shows significant performance boosts compared to the baseline model, with respective gains of \\(3.72\\%\\), \\(4.80\\%\\), \\(7.30\\%\\), and \\(1.31\\%\\) in \\(S_{m}\\), \\(E_{\\xi}\\), \\(F_{\\beta}\\), and \\(\\mathcal{M}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.658 + ], + "angle": 0, + "content": "Furthermore, to validate our approach of progressively enhancing image features with the previously generated association feature for improving hyperassociation calculation, we conduct an ablation experiment called Separate Association Generation (SAG) where association features are generated for three stages without image feature enhancements. As shown in the 2nd&3rd lines in Table 1, PAG outperforms SAG, indicating that our progressive enhancement design can obtain better hyperassociation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Effectiveness of CAC. CAC aims to condense full-pixel hyperassociations in PAG by selecting corresponding pixels and their surrounding contexts. Results in the 3rd&5th lines of Table 1 show that introducing the CAC module improves the performance. Moreover, it reduces the multiply-accumulate operations (MACs) of aggregation networks from 91.38G in the full-pixel PAG to \\(77.19\\mathrm{G}^{1}\\). This indicates that utilizing correspondence estimation to condense the hyperassociations not only effectively reduces the computational burden but also helps obtain more accurate pixel associations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.782, + 0.788, + 0.813 + ], + "angle": 0, + "content": "We also analyze CAC in detail. As deep association features are necessary for reliable correspondence estimation, CAC first pre-condenses the hyperasso-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.669, + 0.84 + ], + "angle": 0, + "content": "We input a group of 6 related \\(256 \\times 256\\) images to measure MACs." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CONDA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.187 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison of our model with other SOTA methods. DC, C9, and CS are DUTS class, COCO9k, and COCO-SEG training data, respectively. bold and underline mark the best and second-best excellent results, respectively." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.401 + ], + "angle": 0, + "content": "
MethodsTrainingCoCA [47]CoSal2015 [44]CoSOD3k [7]
SetSm↑Eξ↑Fβ↑M↓Sm↑Eξ↑Fβ↑M↓Sm↑Eξ↑Fβ↑M↓
GICDDC0.6580.7180.5130.1260.8440.8870.8440.0710.7970.8480.7700.079
GCoNetDC0.6730.7600.5440.1100.8450.8880.8470.0680.8020.8600.7780.071
GCoNet+DC0.6910.7860.5740.1130.8750.9180.8760.0540.8280.8810.8070.068
CONDADC0.7170.7740.6000.1020.8900.9260.8940.0490.8320.8730.8070.067
ICNetC90.6540.7040.5130.1470.8570.9010.8580.0580.7940.8450.7620.089
DCFMC90.7100.7830.5980.0850.8380.8930.8560.0670.8090.8740.8050.067
GCoNet+C90.7170.7980.6050.0980.8530.9020.8570.0730.8190.8770.7960.075
CONDAC90.7300.8010.6220.0920.8650.9100.8750.0590.8250.8770.8100.068
CADCDC+C90.6800.7440.5490.1330.8670.9060.8650.0640.8150.8540.7780.088
DMTDC+C90.7250.8000.6190.1080.8970.9360.9050.0450.8510.8950.8350.063
GCoNet+DC+C90.7340.8080.6260.0880.8760.9200.8800.0570.8390.8940.8220.064
CONDADC+C90.7570.8250.6750.0920.9040.9400.9120.0420.8570.8990.8440.060
UGEMDC+CS0.7260.8080.5990.0960.8850.9350.8820.0510.8530.9110.8290.060
GCoNet+DC+CS0.7380.8140.6370.0810.8810.9260.8910.0550.8430.9010.8340.061
CONDADC+CS0.7630.8390.6850.0890.9000.9380.9080.0450.8620.9110.8530.056
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.785, + 0.531 + ], + "angle": 0, + "content": "ciations using a maximum similarity approach to obtain initial deep association features and then performs further condensation based on the correspondences predicted by these initial deep association features. The hyperassociation condensation process with only the pre-condensation operation is called Similarity-induced Association Condensation (SAC). In Table 1, SAC only brings slight performance improvements due to the heuristic nature of the correspondence estimation. Nevertheless, SAC can provide the initial association feature for CAC to predict reliable correspondence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.532, + 0.785, + 0.653 + ], + "angle": 0, + "content": "Effectiveness of OCC. OCC provides self-supervision for CAC to enable more precise correspondence estimation. Results in the 5th&7th lines of Table 1 show OCC further improves the performance by leveraging more precise correspondence estimations to condense hyperassociations effectively. In addition, we conducted an ablation experiment to validate our object-aware design in CAC by replacing OCC with full-pixel cycle consistency (FCC) loss. Comparing the 5th&6th lines of Table 1, FCC leads to a notable performance decrease due to background pixels disrupting the correspondence learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.653, + 0.785, + 0.713 + ], + "angle": 0, + "content": "Visualization of Correspondence Estimations. We present some visual samples of correspondence estimations for some co-salient pixels in Figure 4. Our semantic correspondence estimations are meaningful and can effectively depict the common attributes of co-salient objects at the pixel level." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.738, + 0.62, + 0.753 + ], + "angle": 0, + "content": "4.4 Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.84 + ], + "angle": 0, + "content": "We compare our model with eight recent SOTA methods, i.e. GICD [47], ICNet [13], GCoNet [8], CADC [46], DCFM [41], DMT [16], UGEM [38], and GCoNet+ [51]. We directly utilize their officially released saliency maps for comparison. To ensure fairness, we trained our model with different combinations of three training datasets, following [51], to align with other compared methods. We" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.127 + ], + "angle": 0, + "content": "L Li et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.143, + 0.784, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.407, + 0.778, + 0.421 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparisons of our model with other SOTA methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.537 + ], + "angle": 0, + "content": "denote three training datasets, i.e. DUTS class [47], COCO9k [18], and COCOSEG [37], as DC, C9, and CS, respectively, for convenience. Our training sets include DC, C9, DC+C9, DC+CS. As shown in Table 2, we can observe that our model achieves the best performance with each training set in most benchmark datasets. What is even more exciting is that we achieve excellent results in the most challenging CoCA dataset, surpassing the second-best models by large margins, e.g. \\(2.5\\%\\) \\(S_{m}\\), \\(2.5\\%\\) \\(E_{\\xi}\\), and \\(4.8\\%\\) \\(\\mathrm{F}_{\\beta}\\) with the DC+CS training set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.599 + ], + "angle": 0, + "content": "We also present visual comparisons in Figure 5. Our model can accurately detect co-salient objects in complex scenarios, such as irregularly shaped accordions accompanied by extraneous objects (people). However, other models easily fail to accurately segment co-salient objects." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.624, + 0.36, + 0.64 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.748 + ], + "angle": 0, + "content": "This paper proposes a deep association learning strategy for CoSOD that directly embeds hyperassociations into deep association features. Correspondence estimation is also introduced to condense hyperassociations, enabling the exploration of pixel-level correspondences for CoSOD. We also utilize an object-aware cycle consistency loss to further refine correspondence estimations. Extensive experiments have verified the effectiveness of our method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.804, + 0.84 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported in part by Key-Area Research and Development Program of Guangdong Province under Grant 2021B0101200001, the National Natural Science Foundation of China under Grant 62136007, 62036011, U20B2065, 6202781, 62036005, the Key R&D Program of Shaanxi Province under Grant 2021ZDLGY01-08, and the MBZUAI-WIS Joint Program for AI Research under Grants WIS P008 and P009." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CONDA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.208 + ], + "angle": 0, + "content": "1. Achanta, R., Hemami, S., Estrada, F., Susstrunk, S.: Frequency-tuned salient region detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1597-1604 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.209, + 0.785, + 0.249 + ], + "angle": 0, + "content": "2. Chen, W., Xu, H., Zhou, Z., Liu, Y., Sun, B., Kang, W., Xie, X.: Costformer: Cost transformer for cost aggregation in multi-view stereo. arXiv preprint arXiv:2305.10320 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.25, + 0.785, + 0.292 + ], + "angle": 0, + "content": "3. Cheng, M.M., Warrell, J., Lin, W.Y., Zheng, S., Vineet, V., Crook, N.: Efficient salient region detection with soft image abstraction. In: Int. Conf. Comput. Vis. pp. 1529-1536 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.293, + 0.785, + 0.32 + ], + "angle": 0, + "content": "4. Cho, S., Hong, S., Jeon, S., Lee, Y., Sohn, K., Kim, S.: Cats: Cost aggregation transformers for visual correspondence 34, 9011-9023 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.321, + 0.785, + 0.348 + ], + "angle": 0, + "content": "5. Fan, D.P., Cheng, M.M., Liu, Y., Li, T., Borji, A.: Structure-measure: A new way to evaluate foreground maps. In: Int. Conf. Comput. Vis. pp. 4548-4557 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.349, + 0.785, + 0.377 + ], + "angle": 0, + "content": "6. Fan, D.P., Gong, C., Cao, Y., Ren, B., Cheng, M.M., Borji, A.: Enhanced-alignment measure for binary foreground map evaluation. In: IJCAI. pp. 698-704 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.418 + ], + "angle": 0, + "content": "7. Fan, D.P., Li, T., Lin, Z., Ji, G.P., Zhang, D., Cheng, M.M., Fu, H., Shen, J.: Re-thinking co-salient object detection. IEEE Trans. Pattern Anal. Mach. Intell. 44(8), 4339-4354 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.785, + 0.461 + ], + "angle": 0, + "content": "8. Fan, Q., Fan, D.P., Fu, H., Tang, C.K., Shao, L., Tai, Y.W.: Group collaborative learning for co-salient object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12288-12298 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.462, + 0.785, + 0.489 + ], + "angle": 0, + "content": "9. Fu, H., Cao, X., Tu, Z.: Cluster-based co-saliency detection. IEEE Trans. Image Process. 22(10), 3766-3778 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.49, + 0.785, + 0.531 + ], + "angle": 0, + "content": "10. Han, J., Cheng, G., Li, Z., Zhang, D.: A unified metric learning-based framework for co-saliency detection. IEEE Trans. Circuit Syst. Video Technol. 28(10), 2473-2483 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.532, + 0.785, + 0.573 + ], + "angle": 0, + "content": "1. Hong, S., Cho, S., Nam, J., Lin, S., Kim, S.: Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In: Eur. Conf. Comput. Vis. pp. 108-126 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.574, + 0.785, + 0.615 + ], + "angle": 0, + "content": "2. Hong, S., Nam, J., Cho, S., Hong, S., Jeon, S., Min, D., Kim, S.: Neural matching fields: Implicit representation of matching fields for visual correspondence 35, 13512-13526 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.616, + 0.785, + 0.644 + ], + "angle": 0, + "content": "3. Jin, W.D., Xu, J., Cheng, M.M., Zhang, Y., Guo, W.: Icnet: Intra-saliency correlation network for co-saliency detection. Adv. Neural Inform. Process. Syst. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.645, + 0.785, + 0.672 + ], + "angle": 0, + "content": "4. Kim, S., Min, J., Cho, M.: Transformers: Match-to-match attention for semantic correspondence. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8697-8707 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.785, + 0.7 + ], + "angle": 0, + "content": "5. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.701, + 0.785, + 0.756 + ], + "angle": 0, + "content": "6. Li, L., Han, J., Zhang, N., Liu, N., Khan, S., Cholakkal, H., Anwer, R.M., Khan, F.S.: Discriminative co-saliency and background mining transformer for co-salient object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7247-7256 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.799 + ], + "angle": 0, + "content": "7. Lin, T.Y., Dólar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2117-2125 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "8. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dóllar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Eur. Conf. Comput. Vis. pp. 740-755 (2014)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.341, + 0.127 + ], + "angle": 0, + "content": "L Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "19. Liu, J.J., Hou, Q., Cheng, M.M., Feng, J., Jiang, J.: A simple pooling-based design for real-time salient object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3917-3926 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "20. Liu, N., Han, J.: Dhsnet: Deep hierarchical saliency network for salient object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 678-686 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.219, + 0.785, + 0.26 + ], + "angle": 0, + "content": "21. Liu, N., Han, J., Yang, M.H.: Picanet: Learning pixel-wise contextual attention for saliency detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3089-3098 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.785, + 0.288 + ], + "angle": 0, + "content": "22. Liu, N., Luo, Z., Zhang, N., Han, J.: Vst++: Efficient and stronger visual saliency transformer. IEEE Trans. Pattern Anal. Mach. Intell. (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.785, + 0.317 + ], + "angle": 0, + "content": "23. Liu, N., Zhang, N., Wan, K., Shao, L., Han, J.: Visual saliency transformer. In: Int. Conf. Comput. Vis. pp. 4722-4732 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.785, + 0.359 + ], + "angle": 0, + "content": "24. Liu, Y., Zhu, L., Yamada, M., Yang, Y.: Semantic correspondence as an optimal transport problem. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4463-4472 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.36, + 0.785, + 0.387 + ], + "angle": 0, + "content": "25. Liu, Y., Liu, N., Yao, X., Han, J.: Intermediate prototype mining transformer for few-shot semantic segmentation 35, 38020-38031 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.388, + 0.785, + 0.431 + ], + "angle": 0, + "content": "26. Luo, Z., Liu, N., Zhao, W., Yang, X., Zhang, D., Fan, D.P., Khan, F., Han, J.: Vscode: General visual salient and camouflaged object detection with 2d prompt learning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.432, + 0.785, + 0.459 + ], + "angle": 0, + "content": "27. Min, J., Kang, D., Cho, M.: Hypercorrelation squeeze for few-shot segmentation. In: Int. Conf. Comput. Vis. pp. 6941-6952 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.459, + 0.785, + 0.5 + ], + "angle": 0, + "content": "28. Moon, S., Sohn, S.S., Zhou, H., Yoon, S., Pavlovic, V., Khan, M.H., Kapadia, M.: Hm: Hybrid masking for few-shot segmentation. In: Eur. Conf. Comput. Vis. pp. 506-523. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.543 + ], + "angle": 0, + "content": "29. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.785, + 0.572 + ], + "angle": 0, + "content": "30. Shen, X., Darmon, F., Efros, A.A., Aubry, M.: Ransac-flow: generic two-stage image alignment. In: Eur. Conf. Comput. Vis. pp. 618-637. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.785, + 0.613 + ], + "angle": 0, + "content": "31. Shi, X., Huang, Z., Bian, W., Li, D., Zhang, M., Cheung, K.C., See, S., Qin, H., Dai, J., Li, H.: Videoflow: Exploiting temporal cues for multi-frame optical flow estimation. arXiv preprint arXiv:2303.08340 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.614, + 0.785, + 0.643 + ], + "angle": 0, + "content": "32. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: Int. Conf. Learn. Represent. (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.785, + 0.685 + ], + "angle": 0, + "content": "33. Su, Y., Deng, J., Sun, R., Lin, G., Su, H., Wu, Q.: A unified transformer framework for group-based segmentation: Co-segmentation, co-saliency detection and video salient object detection. IEEE Trans. Multimedia (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.785, + 0.728 + ], + "angle": 0, + "content": "34. Sun, G., Liu, Y., Tang, H., Chhatkuli, A., Zhang, L., Van Gool, L.: Mining relations among cross-frame affinities for video semantic segmentation. In: Eur. Conf. Comput. Vis. pp. 522-539. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.785, + 0.77 + ], + "angle": 0, + "content": "35. Truong, P., Danelljan, M., Yu, F., Van Gool, L.: Warp consistency for unsupervised learning of dense correspondences. In: Int. Conf. Comput. Vis. pp. 10346-10356 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.785, + 0.812 + ], + "angle": 0, + "content": "36. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Adv. Neural Inform. Process. Syst. 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.785, + 0.84 + ], + "angle": 0, + "content": "37. Wang, C., Zha, Z.J., Liu, D., Xie, H.: Robust deep co-saliency detection with group semantic. In: AAAI. vol. 33, pp. 8917-8924 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CONDA" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "38. Wu, Y., Song, H., Liu, B., Zhang, K., Liu, D.: Co-salient object detection with uncertainty-aware group exchange-masking. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19639-19648 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "39. Xiong, Z., Li, H., Zhu, X.X.: Doubly deformable aggregation of covariance matrices for few-shot segmentation. In: Eur. Conf. Comput. Vis. pp. 133-150. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.232, + 0.787, + 0.272 + ], + "angle": 0, + "content": "40. Xu, G., Wang, X., Ding, X., Yang, X.: Iterative geometry encoding volume for stereo matching. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 21919-21928 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "41. Yu, S., Xiao, J., Zhang, B., Lim, E.G.: Democracy does matter: Comprehensive feature mining for co-salient object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 979-988 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.315, + 0.787, + 0.357 + ], + "angle": 0, + "content": "42. Zhang, D., Fu, H., Han, J., Borji, A., Li, X.: A review of co-saliency detection algorithms: Fundamentals, applications, and challenges. ACM Trans Intell Syst Technol 9(4), 1-31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.357, + 0.787, + 0.398 + ], + "angle": 0, + "content": "43. Zhang, D., Han, J., Han, J., Shao, L.: Cosaliency detection based on intrasaliency prior transfer and deep intersaliency mining. IEEE Trans. Neural Networks Learn. Syst. 27(6), 1163-1176 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.398, + 0.787, + 0.424 + ], + "angle": 0, + "content": "44. Zhang, D., Han, J., Li, C., Wang, J., Li, X.: Detection of co-salient objects by looking deep and wide. Int. J. Comput. Vis. 120(2), 215-232 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.425, + 0.787, + 0.467 + ], + "angle": 0, + "content": "45. Zhang, D., Meng, D., Han, J.: Co-saliency detection via a self-paced multiple-instance learning framework. IEEE Trans. Pattern Anal. Mach. Intell. 39(5), 865-878 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.467, + 0.787, + 0.509 + ], + "angle": 0, + "content": "46. Zhang, N., Han, J., Liu, N., Shao, L.: Summarize and search: Learning consensus-aware dynamic convolution for co-saliency detection. In: Int. Conf. Comput. Vis. pp. 4167-4176 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.509, + 0.787, + 0.536 + ], + "angle": 0, + "content": "47. Zhang, Z., Jin, W., Xu, J., Cheng, M.M.: Gradient-induced co-saliency detection. In: Eur. Conf. Comput. Vis. pp. 455-472 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.536, + 0.787, + 0.564 + ], + "angle": 0, + "content": "48. Zhao, D., Song, Z., Ji, Z., Zhao, G., Ge, W., Yu, Y.: Multi-scale matching networks for semantic correspondence. In: Int. Conf. Comput. Vis. pp. 3354-3364 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.564, + 0.787, + 0.605 + ], + "angle": 0, + "content": "49. Zhao, J.X., Liu, J.J., Fan, D.P., Cao, Y., Yang, J., Cheng, M.M.: Egnet: Edge guidance network for salient object detection. In: Int. Conf. Comput. Vis. pp. 8779-8788 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "50. Zhao, W., Zhang, J., Li, L., Barnes, N., Liu, N., Han, J.: Weakly supervised video salient object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 16826-16835 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.646, + 0.787, + 0.688 + ], + "angle": 0, + "content": "51. Zheng, P., Fu, H., Fan, D.P., Fan, Q., Qin, J., Tai, Y.W., Tang, C.K., Van Gool, L.: Gconet+: A stronger group collaborative co-salient object detector. IEEE Trans. Pattern Anal. Mach. Intell. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.688, + 0.787, + 0.73 + ], + "angle": 0, + "content": "52. Zhuge, M., Fan, D.P., Liu, N., Zhang, D., Xu, D., Shao, L.: Salient object detection via integrity learning. IEEE Trans. Pattern Anal. Mach. Intell. 45(3), 3738-3752 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.73 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CONDA_ Condensed Deep Association Learning for Co-Salient Object Detection./4b7d9a37-9ffb-4ebc-9e75-7e4ada571495_origin.pdf b/2024/CONDA_ Condensed Deep Association Learning for Co-Salient Object Detection./4b7d9a37-9ffb-4ebc-9e75-7e4ada571495_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6a41992a2e6e49d95ff34ab212cf019115f98e0f --- /dev/null +++ b/2024/CONDA_ Condensed Deep Association Learning for Co-Salient Object Detection./4b7d9a37-9ffb-4ebc-9e75-7e4ada571495_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34c896ec7190094aced343238773ff1dd97672021601c148b0df5f8c987b8b07 +size 5199943 diff --git a/2024/CONDA_ Condensed Deep Association Learning for Co-Salient Object Detection./full.md b/2024/CONDA_ Condensed Deep Association Learning for Co-Salient Object Detection./full.md new file mode 100644 index 0000000000000000000000000000000000000000..198826fce31f67ced509912b1106a7b805e1f3ef --- /dev/null +++ b/2024/CONDA_ Condensed Deep Association Learning for Co-Salient Object Detection./full.md @@ -0,0 +1,358 @@ +# CONDA: Condensed Deep Association Learning for Co-Salient Object Detection + +Long Li $^{1}$ , Nian Liu $^{3,*}$ , Dingwen Zhang $^{1}$ , Zhongyu Li $^{4}$ , Salman Khan $^{3,5}$ , Rao Anwer $^{3}$ , Hisham Cholakkal $^{3}$ , Junwei Han $^{1,2,*}$ , and Fahad Shahbaz Khan $^{3,6}$ + +1 Northwestern Polytechnical University + +$^{2}$ Institute of Artificial Intelligence, Hefei Comprehensive National Science Center + +3 Mohamed bin Zayed University of Artificial Intelligence + +$^{4}$ Xi'an Jiaotong University + +5 Australian National University + +$^{6}$ CVL, Linköping University + +Abstract. Inter-image association modeling is crucial for co-salient object detection. Despite satisfactory performance, previous methods still have limitations on sufficient inter-image association modeling. Because most of them focus on image feature optimization under the guidance of heuristically calculated raw inter-image associations. They directly rely on raw associations which are not reliable in complex scenarios, and their image feature optimization approach is not explicit for inter-image association modeling. To alleviate these limitations, this paper proposes a deep association learning strategy that deploys deep networks on raw associations to explicitly transform them into deep association features. Specifically, we first create hyperassociations to collect dense pixel-pair-wise raw associations and then deploys deep aggregation networks on them. We design a progressive association generation module for this purpose with additional enhancement of the hyperassociation calculation. More importantly, we propose a correspondence-induced association condensation module that introduces a pretext task, i.e. semantic correspondence estimation, to condense the hyperassociations for computational burden reduction and noise elimination. We also design an object-aware cycle consistency loss for high-quality correspondence estimations. Experimental results in three benchmark datasets demonstrate the remarkable effectiveness of our proposed method with various training settings. The code is available at: https://github.com/dragonlee258079/CONDA. + +Keywords: Co-salient Object Detection $\cdot$ Deep Association Learning + +# 1 Introduction + +Co-Salient Object Detection (CoSOD) aims to segment salient objects that appear commonly across a group of related images. Compared to traditional Salient + +![](images/404824b20790bd64d1da0a79eb3ae6b9a0ab43b402bf8d6b4670d49ac69f6027.jpg) +Fig. 1: Difference of raw-association-based image feature optimization strategy (a) and our proposed deep association learning strategy (b). Our deep association learning deploys deep learning networks on raw associations to achieve deep association features. We also present visual samples of our calculated raw associations (Raw Asso.), optimized image feature (Opt. Fea.), and our generated deep association features (Asso. Fea.) in (c). + +![](images/7f9cc2f91034f063da20762cb03c1295457ec68c8b6828e05a7cdde337347203.jpg) + +Object Detection (SOD) [19-23, 26, 49, 50, 52], CoSOD is a more challenging task because it requires sufficient inter-image association modeling [9]. + +Recently, many advanced works [8,13,16,33,38,41,46,47,51] have emerged and achieved impressive performance. These methods first use related image features to acquire raw inter-image associations (also called consensus representations), and then leverage them as guidance to optimize each image feature, as shown in Figure 1 (a). This approach enables the final image feature to implicitly capture inter-image cues, thereby achieving the purpose of inter-image association modeling. However, we find this raw-association-based feature optimization strategy still has two limitations: i) they directly rely on raw associations, which are acquired in heuristic manners, such as pixel-wise [8,33,41,51], region-wise [16,46], or image-wise [13] similarity measurements. Although high-quality raw associations can be derived from high-level semantic information in image features, their revelation of common saliency regions still relies on similarity measures, which are unreliable when encountering complex scenarios, such as significant differences between co-salient objects or high foreground-background similarity. ii) the primary focus of building their deep models is on optimizing image features. Compared to directly modeling association relationship, image feature optimization is not an explicit approach for inter-image association modeling and will increase the learning difficulty. + +To alleviate these limitations, we propose a deep association learning strategy for CoSOD, as shown in Figure 1 (b). Instead of directly using raw associations to optimize image features, we deploy deep networks on raw associations to learn deep association features. This is a more explicit strategy for inter-image association modeling. Moreover, our deep association features can capture high-level inter-image association knowledge, making them more robust in complex scenarios than raw associations, as shown in Figure 1 (c). Technically, we start by collecting all pixel-wise raw associations across the entire image group as hyperassociations. Then, we propose a Progressive Association Generation (PAG) + +module to transform hyperassociations into deep association features. PAG progressively generates association features on varying scales, allowing us to use previous association features to enhance the hyperassociation calculation at the next scale, thereby improving the association quality from the very beginning. + +Although deep association learning strategy allows more sufficient interimage association modeling, it significantly increases the computational burden and reduces the practicality of this approach. Additionally, this study finds that it is not necessary to utilize all pixel associations to generate deep association features. In fact, there are even some noisy pixels that negatively impact the quality of the deep association features. Therefore, we propose a method based on Correspondence-induced Association Condensation (CAC) to condense the original full-pixel hyperassociations. This not only alleviates the computational burden but also further enhances the quality of deep association features. + +Specifically, CAC performs the condensation operation by selectively associating pixels that have semantic correspondence in other images, as well as their surrounding contextual pixels, thereby creating lightweight yet more accurate hyperassociations. Here, we introduce a pretext task, i.e. semantic correspondence estimation, into the CoSOD, not only improving the model performance but also delving deeper into the essence of CoSOD. Co-salient objects inherently possess object-level semantic correspondence. However, in this paper, we aim to further explore the finer pixel-level correspondence. Although highly accurate correspondence estimation remains a challenge, we believe it will pave a new way for CoSOD research. We also provide an object-aware cycle consistency (OCC) loss to aid in learning correspondences within co-salient pixels. + +In summary, the contributions of this paper are as follows: + +- For the first time, we introduce a deep association learning approach for CoSOD, applying deep networks to transform raw associations into deep association features for sufficient inter-image association modeling. Specifically, we develop a CONdensed Deep Association (CONDA) learning model. +- We propose a PAG module to progressively generate deep association features. It enhances image features with previous association features to improve hyperassociation calculation. +- We introduce semantic correspondence estimation into the CoSOD task to condense the original hyperassociation for alleviating the computational burden and further improving the performance. We also propose an OCC loss for effective correspondence estimation. +- Experimental results demonstrate that our model achieves significantly improved state-of-the-art performance on three benchmark datasets across different training settings. + +# 2 Related Work + +# 2.1 Co-Salient Object Detection + +Recently, there has been a surge of excellent methods for CoSOD [8,10,13,16,33, 38,41-43,45-47,51]. These methods initially acquire raw inter-image associations + +with related image features and then utilize them to optimize each image feature. Most methods generate pixel-wise [8, 33, 41, 51], region-wise [16, 46], and image-wise [13] raw associations through similarity-based manners, e.g. inner product calculations between image features. Even for transformer-based methods [16, 33], they also rely on inner product calculation to produce attention maps [36] as raw associations. The image feature optimization manners include spatial or channel calibration [8, 13, 51], dynamic convolution [41, 46], feature fusion [16, 41], and cross-attention [16, 33], etc. However, they lack the learning of high-level association knowledge and heavily focus on optimized image features. Unlike them, this paper proposes a new research direction that deploys deep networks on associations to achieve deep association features for CoSOD. + +# 2.2 Inter-image Relation Modeling + +Apart from CoSOD, other tasks necessitating the consideration of inter-image relationships, such as few-shot segmentation [25, 27, 28, 39], stereo matching [2, 40], video semantic segmentation [34], etc. These tasks have made significant advancements recently by effectively modeling inter-image relations. Most of these methods [2, 4, 12, 27, 39, 40] initially create cost volumes to capture dense inter-image pixel-wise similarities and subsequently use various modules to convert these hyper volumes into task-specific features. Our approach distinguishes itself from prior methods in three aspects. Firstly, most of them create 4D cost volumes between two images while we create 6D hyperassociations between all related images. Secondly, they calculate hyper volumes based on original image features, whereas We propose PAG to progressively enhance image features for better hyperassociation calculation. Last and most importantly, they rely on the full-pixel cost volume. We consider the condensation of hyperassociations using semantic correspondences to eliminate noisy pixel associations. + +# 2.3 Semantic Correspondence Estimation + +Semantic correspondence estimation [24] aims to establish reliable pixel correspondences between different instances of the same object category. Most works performed this task using fully supervised training [14, 24]. Some recent works have utilized unsupervised learning with photometric, forward-backward consistency, and warp-supervision losses [30, 35, 48]. However, they implement these losses on the entire image, where background pixels may affect the performance. In this paper, we introduce this task to condense hyperassociations for Co-SOD and tailor the cycle consistency loss by only applying it to co-salient pixels, hence effectively avoiding the influence of background and extraneous objects. + +# 3 Proposed Method + +As shown in Figure 2, CONDA integrates the deep association learning process into an FPN framework. Specifically, given a group of related images $\{I_i\}_{i=1}^N$ , we + +![](images/ac6fd1de06c08cc86d786e4ae99a47c98f05aa45788c8c35362c4a5254b32dd3.jpg) +Fig. 2: Overall flowchart of our CONDA model. Specifically, CONDA first utilizes the image features to calculate hyperassociations. Then, the full-pixel hyperassociations are condensed by CAC and fed into the aggregation networks to achieve deep association features. These features are then used in the FPN decoder process for the final prediction. To be concise, only three related images are shown. + +first input them into a VGG-16 [32] backbone to collect its intermediate features for PAG and FPN decoding. In detail, we collect all features of the last three stages for PAG and the last feature of each stage for the FPN decoder as follows: + +$$ +\begin{array}{l} \mathbb {F} ^ {P} = \left\{\boldsymbol {F} _ {s, l} \mid s \in \{3, \dots , 5 \}, l \in \{1, \dots , L _ {s} \} \right\}, \\ \mathbb {F} ^ {D} = \left\{\boldsymbol {F} _ {s, L _ {s}} \mid s \in \{1, \dots 5 \} \right\}, \tag {1} \\ \end{array} +$$ + +where $\mathbb{F}^P$ and $\mathbb{F}^D$ are the feature collections for PAG and FPN decoder, respectively. $F_{s,l}\in \mathbb{R}^{N\times H_s\times W_s\times C_s}$ represents the VGG feature of the $l$ -th layer in the $s$ -th stage. There are in total $L_{s}$ layers in the $s$ -th stage. $H_{s}$ , $W_{s}$ and $C_s$ represent the height, width, and channel of the $s$ -th stage, respectively. + +Then, we input $\mathbb{F}^P$ into PAG to calculate hyperassociations and generate deep association features $\{\pmb {F}_s^A\in \mathbb{R}^{N\times H_s\times W_s\times C_s}\}_{s = 3}^5$ . Finally, these association features will be fused with $\mathbb{F}^D$ for the FPN decoder process, formulated as: + +$$ +\begin{array}{l} \boldsymbol {F} _ {s, L _ {s}} = \boldsymbol {F} _ {s, L _ {s}} + \phi (\boldsymbol {F} _ {s} ^ {A}), \quad s = 3, \dots , 5, \\ \boldsymbol {F} = \operatorname {D e c o d e r} \left(\left\{\boldsymbol {F} _ {s, L _ {s}} \right\} _ {s = 1} ^ {5}\right), \\ \end{array} +$$ + +where $\phi$ is a convolution layer. $\pmb{F} \in \mathbb{R}^{N \times H \times W \times C}$ is the final feature for the final co-saliency prediction. We adopt BCE and IoU losses for supervision. + +The rest of this section will introduce PAG with full-pixel hyperassociation and the condensation of hyperassociations by plugging the Correspondence-induced Association Condensation (CAC) module into PAG. + +# 3.1 Progressive Association Generation + +Our deep association learning involves two steps: 1) acquiring the raw hyper-associations $\{A_s\}_{s=3}^5$ from three stages; 2) employing aggregation networks on $\{A_s\}_{s=3}^5$ to obtain association features $\{F_s^A\}_{s=3}^5$ . + +Earlier methods [4, 11, 31] directly utilize the original backbone features, i.e. $\mathbb{F}^P$ , to calculate inter-image interactions, such as the so-called cost volume. We argue that hyperassociations derived straight from backbone features might obstruct further improvement of deep association learning, given that the current backbone is pre-trained without any consideration for inter-image associations. + +To alleviate this problem, we propose the PAG module to progressively generate pyramid association features so that we can utilize the high-level association feature, e.g. $\pmb{F}_{s+1}^{A}$ from the $(s+1)$ -th stage, to enhance the neighbouring low-level VGG features in $\mathbb{F}^P$ , e.g. $\pmb{F}_{s,l}$ , for attaining association-enhanced features $\hat{\pmb{F}}_{s,l}$ , based on which we can calculate high-quality hyperassociations $\pmb{A}_s$ in the $s$ -th stage. After that, we execute the subsequent aggregation networks on $\pmb{A}_s$ to achieve association features $\pmb{F}_s^A$ , which will continue to enhance the VGG features of the next stage and carry out progressive association generation. The whole process of our PAG can be formulated as follows: + +$$ +\boldsymbol {A} _ {s} = \operatorname {H A C} \left(\left\{\tilde {\boldsymbol {F}} _ {s, l} \right\} _ {l = 1} ^ {L _ {s}}\right), +$$ + +$$ +\boldsymbol {F} _ {s} ^ {A} = \operatorname {A g g} \left(\boldsymbol {A} _ {s}\right), \tag {3} +$$ + +$$ +\{\hat {\boldsymbol {F}} _ {s - 1, l} \} _ {l = 1} ^ {L _ {s - 1}} = \operatorname {E n h} \left(\left\{\boldsymbol {F} _ {s - 1, l} \right\} _ {l = 1} ^ {L _ {s - 1}}; \boldsymbol {F} _ {s} ^ {A}\right), +$$ + +where $s$ ranges from 5 to 3 and $\{\hat{F}_{5,l}\}_{l=1}^{L_5} = \{\pmb{F}_{5,l}\}_{l=1}^{L_5}$ . The HAC, Agg, and Enh represent the hyperassociation calculation, aggregation network, and association-induced feature enhancement, respectively. Next, we explain them in detail. + +Hyperassociation Calculation. For each stage, we first compute the raw associations at each layer using the inner product between $l$ -2 normalized association-enhanced features of $N$ related images. After that, we stack the raw associations of all layers to form the final hyperassociation of this stage. The hyperassociation for the $s$ -th stage, i.e. $A_{s} \in \mathbb{R}^{N \times H_{s} \times W_{s} \times N \times H_{s} \times W_{s} \times L_{s}}$ , can be calculated via: + +$$ +\begin{array}{l} \boldsymbol {A} _ {s} = \operatorname {H A C} \left(\left\{\hat {\boldsymbol {F}} _ {s, l} \right\} _ {l = 1} ^ {L _ {s}}\right) \\ = \operatorname {S t a c k} \left(\left\{\operatorname {R e L U} \left(\frac {\hat {\boldsymbol {F}} _ {s , l} \cdot \hat {\boldsymbol {F}} _ {s , l} ^ {\top}}{\| \hat {\boldsymbol {F}} _ {s , l} \| \| \hat {\boldsymbol {F}} _ {s , l} ^ {\top} \|}\right) \right\} _ {l = 1} ^ {L _ {s}}\right), \tag {4} \\ \end{array} +$$ + +where $\hat{\pmb{F}}_{s,l}\cdot \hat{\pmb{F}}_{s,l}^{\top}\in \mathbb{R}^{N\times H_s\times W_s\times N\times H_s\times W_s}$ . The $\top$ indicates transposing the last dimension and the first three dimensions. $\| \cdot \|$ represents the $l$ -2 norm. We employ ReLU to suppress noisy association values. + +Aggregation Network. The raw hyperassociation $\mathbf{A}_s \in \mathbb{R}^{N \times H_s \times W_s \times N \times H_s \times W_s \times L_s}$ is a hypercube with a nested structure, where each pixel position is characterized by a 4D tensor $(N \times H_s \times W_s \times L_s)$ . Each 4D tensor documents the associations of the respective pixel with all other pixels in $N$ related images. For + +clarity, we designate the first and second $N \times H_s \times W_s$ dimensions in $A_s$ as the source and target dimensions, respectively. Although these 4D tensors are crucial for exploring the consensus information for co-saliency detection, they essentially comprise pixel-to-pixel similarity values, seen in (4), which may be suboptimal and unreliable in complex scenarios. Therefore, we propose using deep networks to transform these pixel-wise similarities into deep association features with contextual and high-order association knowledge. This has never been explored in previous CoSOD methods. This is implemented via context aggregations on $A_s$ to squeeze these 4D tensors as $C_s$ -dimensional vectors, formulated as: + +$$ +N \times H _ {s} \times W _ {s} \times N \times H _ {s} \times W _ {s} \times L _ {s} \rightarrow N \times H _ {s} \times W _ {s} \times C _ {s}. \tag {5} +$$ + +In detail, we first deploy several association aggregation layers on $\mathbf{A}_s$ to progressively aggregate context information, enlarging $L_s$ as $C_s$ , and eliminate the target $H_s \times W_s$ dimension in 4D tensors. Each aggregation layer consists of 2D convolution layers and a downsampling operation. Specifically, focusing on the first aggregation layer for technical explanation, we first aggregate context information in the target $H_s \times W_s$ dimension by applying a 2D convolution layer on all 4D tensors. The operations on the 4D tensor at pixel position $(h_i, w_i)$ in image $I_i$ can be formulated as: + +$$ +\boldsymbol {A} _ {s} ^ {1} (i, h _ {i}, w _ {i}, j, \therefore , \therefore): = \mathcal {C} _ {1} \left(\boldsymbol {A} _ {s} (i, h _ {i}, w _ {i}, j, \therefore , \therefore)\right), \quad j = 1, \dots , N, \tag {6} +$$ + +where $\mathcal{C}_1$ is a $3\times 3$ 2D convolution layer. Here, $j$ is the index of other related images in the 4D tensor, and $\pmb {A}_s(i,h_i,w_i,j,:;:)\in \mathbb{R}^{H_s\times W_s\times L_s}$ illustrates the associations between the pixel $(h_i,w_i)$ in $I_{i}$ and all pixels in image $I_{j}$ . This interpretation applies to other similar symbols. + +Then, a downsampling operation $\mathcal{D}$ , i.e. bilinear interpolation, is applied on $\pmb{A}_s^1$ to reduce the spatial dimension of the 4D tensor by a scaling factor: + +$$ +\boldsymbol {A} _ {s} ^ {2} (i, h _ {i}, w _ {i}, j, \therefore , \therefore ,:) = \mathcal {D} (\boldsymbol {A} _ {s} ^ {1} (i, h _ {i}, w _ {i}, j, \therefore , \therefore)), \tag {7} +$$ + +where $A_{s}^{2}(i,h_{i},w_{i},j,:,:) \in \mathbb{R}^{H_{s}^{\prime} \times W_{s}^{\prime} \times C_{s}^{\prime}}$ , $H_{s}^{\prime}$ and $W_{s}^{\prime}$ are downsampled height and width. $C_{s}^{\prime}$ is the channel number after convolution $\mathcal{C}_1$ . + +Finally, we also aggregate context information in the source $H_{s} \times W_{s}$ dimension. Specifically, we extract the values along the source dimension and channel dimension in $A_{s}^{2}$ to form 4D tensors and apply a 2D convolution layer on them. For instance, $A_{s}^{2}(:, :, :, j, h_{j}, w_{j},:) \in \mathbb{R}^{N \times H_{s} \times W_{s} \times C_{s}^{\prime}}$ is such a 4D tensor, where $(h_{j}, w_{j})$ is a pixel position in the target dimension. This can be formulated as: + +$$ +\boldsymbol {A} _ {s} ^ {3} (i,:,:) j, h _ {j}, w _ {j,:}) = \mathcal {C} _ {2} \left(\boldsymbol {A} _ {s} ^ {2} (i,:,:) j, h _ {j}, w _ {j,:})\right), \quad i = 1, \dots , N, \tag {8} +$$ + +where $\mathcal{C}_2$ is a $3\times 3$ 2D convolution layer. $i$ is the related image index. + +After several association aggregation layers, as shown in (6)-(8), we can achieve the aggregated association features with the target $H_{s} \times W_{s}$ dimension eliminated, denoted as $\pmb{F}_{s}^{A^{\prime}} \in \mathbb{R}^{N \times H_{s} \times W_{s} \times N \times C}$ . Subsequently, we average $\pmb{F}_{s}^{A^{\prime}}$ on its second $N$ dimension and obtain the final association features + +![](images/e979f8630c372e8cd48870e7fafda8137440dd82e40736aff359426acda2b2dc.jpg) +Fig. 3: Difference between full-pixel hyperassociation (a) and condensed hyperassociation (b). We provide an example of collecting the pixel associations from image $I_{j}$ for a pixel $(h_{i},w_{i})$ in image $I_{i}$ . Full-pixel hyperassociation collects all pixel associations in $I_{j}$ , while our condensed hyperassociation only collects the associations of its correspondence pixel $(h_{j},w_{j})$ (red dot) and surrounding pixels (green dots). We first heuristically find an initial pixel $(h_j^0,w_j^0)$ with a fixed surrounding window and then learn coordinate offsets to locate the optimized correspondence and surrounding pixels. + +$\pmb{F}_s^A \in \mathbb{R}^{N \times H_s \times W_s \times C}$ , formulated as: + +$$ +\boldsymbol {F} _ {s} ^ {A} = \frac {1}{N} \sum_ {j = 1} ^ {N} \boldsymbol {F} _ {s} ^ {A ^ {\prime}} (:,:,:, j,:). \tag {9} +$$ + +Association-induced Feature Enhancement. Once we have obtained association feature $\mathbf{F}_s^A$ of the $s$ -th stage, we will use it to enhance the VGG feature of the $(s - 1)$ -th stage, i.e. $\{\mathbf{F}_{s - 1,l}\}_{l = 1}^{L_{s - 1}}$ . Specifically, we upsample $\mathbf{F}_s^A$ to align the spatial size of features in the $(s - 1)$ -th stage, and then add it to $\{\mathbf{F}_{s - 1,l}\}_{l = 1}^{L_{s - 1}}$ followed by a convolution layer, formulated as: + +$$ +\hat {\boldsymbol {F}} _ {s - 1, l} = \mathcal {C} _ {3} \left(\boldsymbol {F} _ {s - 1, l} + \mathcal {U} \left(\boldsymbol {F} _ {s} ^ {A}\right)\right), \tag {10} +$$ + +where $\mathcal{C}_3$ and $\mathcal{U}$ represent a 2D convolution layer and the bilinear upsampling operation, respectively. + +# 3.2 Correspondence-induced Association Condensation + +Although PAG based on full-pixel hyperassociations can deliver satisfactory performance in CoSOD, it also introduces substantial computational overhead. Additionally, we argue that for each pixel in an image, it is unnecessary to gather its associations with all pixels of other related images to form hyperassociations. Some pixel associations may even impair the final performance, such as those between ambiguous regions. To this end, this subsection tries to condense the original full-pixel hyperassociations to retain only informative pixel associations. + +This subsection will focus on explaining the condensation of pixel associations of a pixel (e.g. $(h_i, w_i)$ in $I_i$ ) to ones of the other images (e.g. image $I_j$ ), i.e. $\mathbf{A}_s(i, h_i, w_i, j, :, :, :) \in \mathbb{R}^{H_s \times W_s \times L_s}$ , as shown in Figure 3. We will simplify the symbol $\mathbf{A}_s(i, h_i, w_i, j, :, :, :)$ as $\mathbf{a}_s^j$ in the subsequent text for convenience. + +Specifically, CAC opts to select $K \times K$ ( $K < H_{s}, K < W_{s}$ ) informative pixel associations from $\pmb{a}_{s}^{j}$ to form its condensed representation, i.e. $\tilde{\pmb{a}}_{s}^{j} \in \mathbb{R}^{K \times K \times L_{s}}$ . Thus, the entire condensed hyperassocaton can be symbolized as $\pmb{A}_{s} \in \mathbb{R}^{N \times H_{s} \times W_{s} \times N \times L_{s}}$ . + +$K \times K \times L_s$ . To ensure the proper selection of $K \times K$ pixels, we introduce a pretext task, i.e. semantic correspondence estimation [14,24]. This allows us to first locate the corresponding pixel of $(h_i, w_i)$ in image $I_j$ , i.e. $(h_j, w_j)$ , and then combine $(h_j, w_j)$ with its surrounding pixels as the $K \times K$ pixel set. We design this approach based on the observation that the co-salient objects across $N$ related images belong to the same semantic category, and the pixels within them should have semantic correspondences to each other, as shown in Figure 4. Therefore, the introduction of semantic correspondence in CAC not only improves the CoSOD performance but also delves deeper into the core nature of the CoSOD task. As far as we know, this is the first work to use semantic correspondence in the CoSOD task. + +Correspondence Estimation. To estimate the correspondence pixel $(h_j, w_j)$ in $I_j$ for $(h_i, w_i)$ , we first identify an initial pixel $(h_j^0, w_j^0)$ via a heuristic approach. Subsequently, we produce a spatial offset to refine $(h_j^0, w_j^0)$ into $(h_j, w_j)$ . To achieve this purpose, all initial correspondence pixels should be utilized to form the initial condensed hyperassociations, with which the initial deep association features can be generated for spatial offset prediction. + +Specifically, we pick out $(h_j^0, w_j^0)$ that has the largest feature similarity value with $(h_i, w_i)$ . Since we have calculated the feature similarities in $\pmb{a}_s^j$ , the $(h_j^0, w_j^0)$ can be obtained as follows: + +$$ +\bar {\pmb {a}} _ {s} ^ {j} = \sum_ {l = 1} ^ {L _ {s}} \pmb {a} _ {s} ^ {j} (:,: l), +$$ + +$$ +\left(h _ {j} ^ {0}, w _ {j} ^ {0}\right) = \underset {x, y} {\operatorname {a r g m a x}} \left(\bar {\boldsymbol {a}} _ {s} ^ {j} (x, y)\right), \tag {11} +$$ + +where $\bar{\pmb{a}}_s^j\in \mathbb{R}^{H_s\times W_s}$ is a similarity matrix obtained by eliminating the last dimension of $\pmb {a}_s^j\in \mathbb{R}^{H_s\times W_s\times L_s}$ through summation. argmax returns the coordinate of the maximum value. + +Next, we select $K \times K$ pixels within the square region centered around the initial pixels, e.g. $(h_j^0, w_j^0)$ , to construct the initial condensed hyperassociation $\tilde{\pmb{A}}_s'$ . Then, we can feed it into the aggregation network, described in Sec 3.1, and achieve the initial aggregated association features $\pmb{F}_s^{A'} \in \mathbb{R}^{N \times H_s \times W_s \times N \times C_s}$ . It can be regarded as the association features of each pixel to $N$ other related images. We utilize the feature of $(h_i, w_i)$ to $I_j$ , i.e. $\pmb{F}_s^{A'}(i, h_i, w_i, j,:) \in \mathbb{R}^{C_s}$ , to predict the offsets for $(h_j, w_j)$ via a linear layer, formulated as: + +$$ +\boldsymbol {\Delta} _ {s} ^ {j} = \mathcal {O} \left(\boldsymbol {F} _ {s} ^ {A ^ {\prime}} (i, h _ {i}, w _ {i}, j,:)\right), \tag {12} +$$ + +where $\mathcal{O}$ is a linear layer for offset generation. $\pmb{\Delta}_{s}^{j} \in \mathbb{R}^{K \times K \times 2}$ consists of $K \times K$ offsets, besides the center offset $\pmb{\Delta}_{s}^{j}(k_{c}, k_{c},:)$ for correspondence estimation, i.e. refining $(h_{j}^{0}, w_{j}^{0})$ as $(h_{j}, w_{j})$ , we also generate other offsets for surrounding pixel selection. Thus, the corresponding pixel $(h_{j}, w_{j})$ can be obtained by adding the offset to the initial pixel $(h_{j}^{0}, w_{j}^{0})$ , formulated as: + +$$ +\left(h _ {j}, w _ {j}\right) = \left(h _ {j} ^ {0}, w _ {j} ^ {0}\right) + \boldsymbol {\Delta} _ {s} ^ {j} \left(k _ {c}, k _ {c}, :\right), \tag {13} +$$ + +where $(k_c, k_c)$ is the center position of $K \times K$ square. + +Condensation Operation. Given the estimated correspondence pixel $(h_j, w_j)$ and other offsets in $\pmb{\Delta}_{s}^{j}$ , we can obtain the surrounding pixels and combine them with $(h_j, w_j)$ to form $K \times K$ pixel set. Their coordinates are stored in $\pmb{n}_s^j \in \mathbb{R}^{K \times K \times 2}$ . This process can be formulated as: + +$$ +\boldsymbol {n} _ {s} ^ {j} (x, y,:) = \left(h _ {j}, w _ {j}\right) + \boldsymbol {\Delta} _ {s} ^ {j} (x, y,:) \quad x, y \in \{1, \dots , K \}; x, y \neq k _ {c}, \tag {14} +$$ + +$$ +\boldsymbol {n} _ {s} ^ {j} (k _ {c}, k _ {c};) = (h _ {j}, w _ {j}), +$$ + +where $(k_c, k_c)$ is the center position of $K \times K$ square. + +Finally, we can perform the condensation operation via the index selection on $\pmb{a}_s^j$ , formulated as: + +$$ +\tilde {\boldsymbol {a}} _ {s} ^ {j} = \boldsymbol {a} _ {s} ^ {j} \left(\boldsymbol {n} _ {s} ^ {j}\right), \tag {15} +$$ + +where $\tilde{\pmb{a}}_s^j$ is the condensed representation of $\pmb{a}_s^j$ , i.e. pixel associations of $(h_i, w_i)$ to image $I_j$ . Furthermore, we also illustrate the condensation for pixel associations of $(h_i, w_i)$ to all images, i.e. $\pmb{a}_s \in \mathbb{R}^{N \times H_s \times W_s \times L_s}$ , in Figure 2. By applying such a condensation process to all pixel associations, we can obtain the final condensed hyperassociation $\tilde{\pmb{A}}_s$ . + +# 3.3 Object-aware Cycle Consistency Loss + +To achieve accurate correspondence estimations, applying effective supervisions on them is necessary. As explicit semantic correspondence annotations are not available, we can only rely on unsupervised losses by imposing correspondence-related constraints on estimated correspondences. Previous unsupervised approaches apply constraints to all pixels [30, 35, 48], including those on the background and extraneous objects that don't have mutual correspondences, hence harming the model effectiveness. To avoid this problem, we propose an object-aware constraint to only access losses on the co-salient pixels. + +We propose an Object-aware Cycle Consistency (OCC) loss for the supervision of correspondence estimations in CoSOD. The cycle consistency can be explained as: if a co-salient pixel $(h_i, w_i)$ of image $I_i$ corresponds to the pixel $(h_j, w_j)$ in image $I_j$ , then the pixel $(h_j, w_j)$ should also semantically correspond to pixel $(h_i, w_i)$ . + +Based on this cycle consistency constraint, we adopt image warping operations to conduct the OCC loss. Specifically, we first warp the image $I_{i}^{s}$ (a resized $I_{i}$ to align the scales in the $s$ -th stage) as $I_{i\rightarrow j}^{s}$ using the $I_{i}^{s}\to I_{j}^{s}$ correspondence estimations. Next, we warp $I_{i\rightarrow j}^{s}$ backward to $I_{i\rightarrow j\rightarrow i}^{s}$ using $I_{j}^{s}\to I_{i}^{s}$ correspondence estimations. Finally, we can utilize the SSIM loss between $I_{i}^{s}$ and $I_{i\rightarrow j\rightarrow i}^{s}$ to measure the cycle consistency for mutual correspondence pixels in $I_{i}^{s}$ and $I_{j}^{s}$ . Moreover, to ensure the constraints are only conducted on co-salient objects, we mask the images with their ground truth masks, formulated as: + +$$ +\mathcal {L} _ {s} ^ {C} = \frac {1}{N ^ {2}} \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {N} \mathcal {L} _ {S S I M} \left(I _ {i} ^ {s} \cdot G _ {i} ^ {s}, I _ {i \rightarrow j \rightarrow i} ^ {s} \cdot G _ {i} ^ {s}\right), \tag {16} +$$ + +Table 1: Ablation Study of our proposed modules. SAG, SAC, and FCC are ablation modules for PAG, CAC, and OCC, respectively. + +
IDModulesCoCA [47]
SAGPAGSACCACFCCOCCSm↑Eξ↑Fβ↑M↓
10.69360.76420.57290.1206
2✓0.72360.80290.63570.1106
3✓0.73080.81220.64590.1075
4✓✓0.73040.81230.65000.1085
5✓✓0.74730.81550.65910.0956
6✓✓0.73980.81380.65060.1013
7✓✓0.75700.82480.67510.0924
+ +where $G_{i}^{s}$ is the resized ground truth of image $I_{i}$ . The total OCC loss $\mathcal{L}^C$ is the sum of three stages, formulated as: $\mathcal{L}^C = \mathcal{L}_3^C +\mathcal{L}_4^C +\mathcal{L}_5^C$ . More details can be found in the supplementary materials. + +# 4 Experiments + +# 4.1 Evaluation Datasets and Metrics + +We follow [8] to evaluate our model on three benchmark datasets, i.e. CoCA [47] (1295 images of 80 groups), CoSal2015 [44] (2015 images of 50 groups), and CoSOD3k [7] (3316 images of 160 groups). We adopt four widely-used metrics for the quantitative evaluation, i.e. Structure-measure $(S_{m})$ [5], Maximum enhanced-alignment measure $(E_{\xi})$ [6], Maximum F-measure $(\mathrm{F}_{\beta})$ [1], and Mean Absolute Error $(\mathcal{M})$ [3]. + +# 4.2 Implementation Details + +To construct the training data, we follow [51] to use different combinations of three commonly used training datasets, i.e. DUTS class [47] (8250 images of 291 groups), COCO-9k [18] (9213 images of 65 groups), and COCO-SEG [37] (200,000 images of 78 groups), for a fair comparison with other state-of-the-art (SOTA) works. We also implement the synthesis strategy for the DUTS class dataset following [46]. + +For training specifics, we employ the data augmentation strategy in [21] and use $256 \times 256$ as the input size for the network. We employ the Adam optimizer [15] with $\beta_{1} = 0.9$ and $\beta_{2} = 0.99$ to optimize the network. We train our CONDA model for 300 epochs, starting with an initial learning rate of $1e - 4$ , which is divided by 10 at the $60,000^{th}$ iteration. Our experiments are implemented based on PyTorch [29] on a single Tesla A40 GPU, with the batchsize set to $N = 6$ . The hyperparameter $K$ in CAC is set to 9. + +# 4.3 Ablation Study + +We conduct ablation studies on the most challenging CoCA [47] dataset. To construct the baseline, we use the FPN [17] (with VGG-16 as encoder) as the + +![](images/40395a7bb7331fffc5891dbbab1881e79c2baa445aa0a5bdba710a534942373c.jpg) +Fig. 4: Visual samples for the correspondence estimations. Correspondences I, II, and III visually display estimated correspondences between the main image and three related images. Sparse co-salient pixels were selected and connected to their corresponding pixels using colored lines for clear visualization. + +foundational segmentation network and enhance it with the Region-to-Region correlation module (R2R) [16] to simply capture inter-image connections. Then, as shown in Table 1, we incrementally incorporate PAG, CAC, and OCC into the baseline for effectiveness analysis. We trained all ablation models with the DUTS class and COCO9k datasets. + +Effectiveness of PAG. PAG first uses intermediate image features from the FPN encoder to calculate full-pixel hyperassociations. Then, aggregation networks are applied to generate deep association features for the decoder process. PAG effectively utilizes deep learning to model pair-wise pixel associations, achieving higher-level association knowledge compared to previous image feature optimization strategies. As shown in the 3rd line of Table 1, PAG shows significant performance boosts compared to the baseline model, with respective gains of $3.72\%$ , $4.80\%$ , $7.30\%$ , and $1.31\%$ in $S_{m}$ , $E_{\xi}$ , $F_{\beta}$ , and $\mathcal{M}$ . + +Furthermore, to validate our approach of progressively enhancing image features with the previously generated association feature for improving hyperassociation calculation, we conduct an ablation experiment called Separate Association Generation (SAG) where association features are generated for three stages without image feature enhancements. As shown in the 2nd&3rd lines in Table 1, PAG outperforms SAG, indicating that our progressive enhancement design can obtain better hyperassociation. + +Effectiveness of CAC. CAC aims to condense full-pixel hyperassociations in PAG by selecting corresponding pixels and their surrounding contexts. Results in the 3rd&5th lines of Table 1 show that introducing the CAC module improves the performance. Moreover, it reduces the multiply-accumulate operations (MACs) of aggregation networks from 91.38G in the full-pixel PAG to $77.19\mathrm{G}^{1}$ . This indicates that utilizing correspondence estimation to condense the hyperassociations not only effectively reduces the computational burden but also helps obtain more accurate pixel associations. + +We also analyze CAC in detail. As deep association features are necessary for reliable correspondence estimation, CAC first pre-condenses the hyperasso- + +Table 2: Quantitative comparison of our model with other SOTA methods. DC, C9, and CS are DUTS class, COCO9k, and COCO-SEG training data, respectively. bold and underline mark the best and second-best excellent results, respectively. + +
MethodsTrainingCoCA [47]CoSal2015 [44]CoSOD3k [7]
SetSm↑Eξ↑Fβ↑M↓Sm↑Eξ↑Fβ↑M↓Sm↑Eξ↑Fβ↑M↓
GICDDC0.6580.7180.5130.1260.8440.8870.8440.0710.7970.8480.7700.079
GCoNetDC0.6730.7600.5440.1100.8450.8880.8470.0680.8020.8600.7780.071
GCoNet+DC0.6910.7860.5740.1130.8750.9180.8760.0540.8280.8810.8070.068
CONDADC0.7170.7740.6000.1020.8900.9260.8940.0490.8320.8730.8070.067
ICNetC90.6540.7040.5130.1470.8570.9010.8580.0580.7940.8450.7620.089
DCFMC90.7100.7830.5980.0850.8380.8930.8560.0670.8090.8740.8050.067
GCoNet+C90.7170.7980.6050.0980.8530.9020.8570.0730.8190.8770.7960.075
CONDAC90.7300.8010.6220.0920.8650.9100.8750.0590.8250.8770.8100.068
CADCDC+C90.6800.7440.5490.1330.8670.9060.8650.0640.8150.8540.7780.088
DMTDC+C90.7250.8000.6190.1080.8970.9360.9050.0450.8510.8950.8350.063
GCoNet+DC+C90.7340.8080.6260.0880.8760.9200.8800.0570.8390.8940.8220.064
CONDADC+C90.7570.8250.6750.0920.9040.9400.9120.0420.8570.8990.8440.060
UGEMDC+CS0.7260.8080.5990.0960.8850.9350.8820.0510.8530.9110.8290.060
GCoNet+DC+CS0.7380.8140.6370.0810.8810.9260.8910.0550.8430.9010.8340.061
CONDADC+CS0.7630.8390.6850.0890.9000.9380.9080.0450.8620.9110.8530.056
+ +ciations using a maximum similarity approach to obtain initial deep association features and then performs further condensation based on the correspondences predicted by these initial deep association features. The hyperassociation condensation process with only the pre-condensation operation is called Similarity-induced Association Condensation (SAC). In Table 1, SAC only brings slight performance improvements due to the heuristic nature of the correspondence estimation. Nevertheless, SAC can provide the initial association feature for CAC to predict reliable correspondence. + +Effectiveness of OCC. OCC provides self-supervision for CAC to enable more precise correspondence estimation. Results in the 5th&7th lines of Table 1 show OCC further improves the performance by leveraging more precise correspondence estimations to condense hyperassociations effectively. In addition, we conducted an ablation experiment to validate our object-aware design in CAC by replacing OCC with full-pixel cycle consistency (FCC) loss. Comparing the 5th&6th lines of Table 1, FCC leads to a notable performance decrease due to background pixels disrupting the correspondence learning. + +Visualization of Correspondence Estimations. We present some visual samples of correspondence estimations for some co-salient pixels in Figure 4. Our semantic correspondence estimations are meaningful and can effectively depict the common attributes of co-salient objects at the pixel level. + +# 4.4 Comparison with State-of-the-Art Methods + +We compare our model with eight recent SOTA methods, i.e. GICD [47], ICNet [13], GCoNet [8], CADC [46], DCFM [41], DMT [16], UGEM [38], and GCoNet+ [51]. We directly utilize their officially released saliency maps for comparison. To ensure fairness, we trained our model with different combinations of three training datasets, following [51], to align with other compared methods. We + +![](images/7d0b384ecf46d41b5b3fd790f7ba1310cf13090bae0c24b23cda3f353d2fa847.jpg) +Fig. 5: Qualitative comparisons of our model with other SOTA methods. + +denote three training datasets, i.e. DUTS class [47], COCO9k [18], and COCOSEG [37], as DC, C9, and CS, respectively, for convenience. Our training sets include DC, C9, DC+C9, DC+CS. As shown in Table 2, we can observe that our model achieves the best performance with each training set in most benchmark datasets. What is even more exciting is that we achieve excellent results in the most challenging CoCA dataset, surpassing the second-best models by large margins, e.g. $2.5\%$ $S_{m}$ , $2.5\%$ $E_{\xi}$ , and $4.8\%$ $\mathrm{F}_{\beta}$ with the DC+CS training set. + +We also present visual comparisons in Figure 5. Our model can accurately detect co-salient objects in complex scenarios, such as irregularly shaped accordions accompanied by extraneous objects (people). However, other models easily fail to accurately segment co-salient objects. + +# 5 Conclusion + +This paper proposes a deep association learning strategy for CoSOD that directly embeds hyperassociations into deep association features. Correspondence estimation is also introduced to condense hyperassociations, enabling the exploration of pixel-level correspondences for CoSOD. We also utilize an object-aware cycle consistency loss to further refine correspondence estimations. Extensive experiments have verified the effectiveness of our method. + +Acknowledgments. This work was supported in part by Key-Area Research and Development Program of Guangdong Province under Grant 2021B0101200001, the National Natural Science Foundation of China under Grant 62136007, 62036011, U20B2065, 6202781, 62036005, the Key R&D Program of Shaanxi Province under Grant 2021ZDLGY01-08, and the MBZUAI-WIS Joint Program for AI Research under Grants WIS P008 and P009. + +# References + +1. Achanta, R., Hemami, S., Estrada, F., Susstrunk, S.: Frequency-tuned salient region detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1597-1604 (2009) +2. Chen, W., Xu, H., Zhou, Z., Liu, Y., Sun, B., Kang, W., Xie, X.: Costformer: Cost transformer for cost aggregation in multi-view stereo. arXiv preprint arXiv:2305.10320 (2023) +3. Cheng, M.M., Warrell, J., Lin, W.Y., Zheng, S., Vineet, V., Crook, N.: Efficient salient region detection with soft image abstraction. In: Int. Conf. Comput. Vis. pp. 1529-1536 (2013) +4. Cho, S., Hong, S., Jeon, S., Lee, Y., Sohn, K., Kim, S.: Cats: Cost aggregation transformers for visual correspondence 34, 9011-9023 (2021) +5. Fan, D.P., Cheng, M.M., Liu, Y., Li, T., Borji, A.: Structure-measure: A new way to evaluate foreground maps. In: Int. Conf. Comput. Vis. pp. 4548-4557 (2017) +6. Fan, D.P., Gong, C., Cao, Y., Ren, B., Cheng, M.M., Borji, A.: Enhanced-alignment measure for binary foreground map evaluation. In: IJCAI. pp. 698-704 (2018) +7. Fan, D.P., Li, T., Lin, Z., Ji, G.P., Zhang, D., Cheng, M.M., Fu, H., Shen, J.: Re-thinking co-salient object detection. IEEE Trans. Pattern Anal. Mach. Intell. 44(8), 4339-4354 (2021) +8. Fan, Q., Fan, D.P., Fu, H., Tang, C.K., Shao, L., Tai, Y.W.: Group collaborative learning for co-salient object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12288-12298 (2021) +9. Fu, H., Cao, X., Tu, Z.: Cluster-based co-saliency detection. IEEE Trans. Image Process. 22(10), 3766-3778 (2013) +10. Han, J., Cheng, G., Li, Z., Zhang, D.: A unified metric learning-based framework for co-saliency detection. IEEE Trans. Circuit Syst. Video Technol. 28(10), 2473-2483 (2017) +1. Hong, S., Cho, S., Nam, J., Lin, S., Kim, S.: Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In: Eur. Conf. Comput. Vis. pp. 108-126 (2022) +2. Hong, S., Nam, J., Cho, S., Hong, S., Jeon, S., Min, D., Kim, S.: Neural matching fields: Implicit representation of matching fields for visual correspondence 35, 13512-13526 (2022) +3. Jin, W.D., Xu, J., Cheng, M.M., Zhang, Y., Guo, W.: Icnet: Intra-saliency correlation network for co-saliency detection. Adv. Neural Inform. Process. Syst. (2020) +4. Kim, S., Min, J., Cho, M.: Transformers: Match-to-match attention for semantic correspondence. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8697-8707 (2022) +5. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +6. Li, L., Han, J., Zhang, N., Liu, N., Khan, S., Cholakkal, H., Anwer, R.M., Khan, F.S.: Discriminative co-saliency and background mining transformer for co-salient object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7247-7256 (2023) +7. Lin, T.Y., Dólar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2117-2125 (2017) +8. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dóllar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Eur. Conf. Comput. Vis. pp. 740-755 (2014) + +19. Liu, J.J., Hou, Q., Cheng, M.M., Feng, J., Jiang, J.: A simple pooling-based design for real-time salient object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3917-3926 (2019) +20. Liu, N., Han, J.: Dhsnet: Deep hierarchical saliency network for salient object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 678-686 (2016) +21. Liu, N., Han, J., Yang, M.H.: Picanet: Learning pixel-wise contextual attention for saliency detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3089-3098 (2018) +22. Liu, N., Luo, Z., Zhang, N., Han, J.: Vst++: Efficient and stronger visual saliency transformer. IEEE Trans. Pattern Anal. Mach. Intell. (2024) +23. Liu, N., Zhang, N., Wan, K., Shao, L., Han, J.: Visual saliency transformer. In: Int. Conf. Comput. Vis. pp. 4722-4732 (2021) +24. Liu, Y., Zhu, L., Yamada, M., Yang, Y.: Semantic correspondence as an optimal transport problem. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4463-4472 (2020) +25. Liu, Y., Liu, N., Yao, X., Han, J.: Intermediate prototype mining transformer for few-shot semantic segmentation 35, 38020-38031 (2022) +26. Luo, Z., Liu, N., Zhao, W., Yang, X., Zhang, D., Fan, D.P., Khan, F., Han, J.: Vscode: General visual salient and camouflaged object detection with 2d prompt learning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2024) +27. Min, J., Kang, D., Cho, M.: Hypercorrelation squeeze for few-shot segmentation. In: Int. Conf. Comput. Vis. pp. 6941-6952 (2021) +28. Moon, S., Sohn, S.S., Zhou, H., Yoon, S., Pavlovic, V., Khan, M.H., Kapadia, M.: Hm: Hybrid masking for few-shot segmentation. In: Eur. Conf. Comput. Vis. pp. 506-523. Springer (2022) +29. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library 32 (2019) +30. Shen, X., Darmon, F., Efros, A.A., Aubry, M.: Ransac-flow: generic two-stage image alignment. In: Eur. Conf. Comput. Vis. pp. 618-637. Springer (2020) +31. Shi, X., Huang, Z., Bian, W., Li, D., Zhang, M., Cheung, K.C., See, S., Qin, H., Dai, J., Li, H.: Videoflow: Exploiting temporal cues for multi-frame optical flow estimation. arXiv preprint arXiv:2303.08340 (2023) +32. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: Int. Conf. Learn. Represent. (2015) +33. Su, Y., Deng, J., Sun, R., Lin, G., Su, H., Wu, Q.: A unified transformer framework for group-based segmentation: Co-segmentation, co-saliency detection and video salient object detection. IEEE Trans. Multimedia (2023) +34. Sun, G., Liu, Y., Tang, H., Chhatkuli, A., Zhang, L., Van Gool, L.: Mining relations among cross-frame affinities for video semantic segmentation. In: Eur. Conf. Comput. Vis. pp. 522-539. Springer (2022) +35. Truong, P., Danelljan, M., Yu, F., Van Gool, L.: Warp consistency for unsupervised learning of dense correspondences. In: Int. Conf. Comput. Vis. pp. 10346-10356 (2021) +36. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Adv. Neural Inform. Process. Syst. 30 (2017) +37. Wang, C., Zha, Z.J., Liu, D., Xie, H.: Robust deep co-saliency detection with group semantic. In: AAAI. vol. 33, pp. 8917-8924 (2019) + +38. Wu, Y., Song, H., Liu, B., Zhang, K., Liu, D.: Co-salient object detection with uncertainty-aware group exchange-masking. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19639-19648 (2023) +39. Xiong, Z., Li, H., Zhu, X.X.: Doubly deformable aggregation of covariance matrices for few-shot segmentation. In: Eur. Conf. Comput. Vis. pp. 133-150. Springer (2022) +40. Xu, G., Wang, X., Ding, X., Yang, X.: Iterative geometry encoding volume for stereo matching. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 21919-21928 (2023) +41. Yu, S., Xiao, J., Zhang, B., Lim, E.G.: Democracy does matter: Comprehensive feature mining for co-salient object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 979-988 (2022) +42. Zhang, D., Fu, H., Han, J., Borji, A., Li, X.: A review of co-saliency detection algorithms: Fundamentals, applications, and challenges. ACM Trans Intell Syst Technol 9(4), 1-31 (2018) +43. Zhang, D., Han, J., Han, J., Shao, L.: Cosaliency detection based on intrasaliency prior transfer and deep intersaliency mining. IEEE Trans. Neural Networks Learn. Syst. 27(6), 1163-1176 (2015) +44. Zhang, D., Han, J., Li, C., Wang, J., Li, X.: Detection of co-salient objects by looking deep and wide. Int. J. Comput. Vis. 120(2), 215-232 (2016) +45. Zhang, D., Meng, D., Han, J.: Co-saliency detection via a self-paced multiple-instance learning framework. IEEE Trans. Pattern Anal. Mach. Intell. 39(5), 865-878 (2016) +46. Zhang, N., Han, J., Liu, N., Shao, L.: Summarize and search: Learning consensus-aware dynamic convolution for co-saliency detection. In: Int. Conf. Comput. Vis. pp. 4167-4176 (2021) +47. Zhang, Z., Jin, W., Xu, J., Cheng, M.M.: Gradient-induced co-saliency detection. In: Eur. Conf. Comput. Vis. pp. 455-472 (2020) +48. Zhao, D., Song, Z., Ji, Z., Zhao, G., Ge, W., Yu, Y.: Multi-scale matching networks for semantic correspondence. In: Int. Conf. Comput. Vis. pp. 3354-3364 (2021) +49. Zhao, J.X., Liu, J.J., Fan, D.P., Cao, Y., Yang, J., Cheng, M.M.: Egnet: Edge guidance network for salient object detection. In: Int. Conf. Comput. Vis. pp. 8779-8788 (2019) +50. Zhao, W., Zhang, J., Li, L., Barnes, N., Liu, N., Han, J.: Weakly supervised video salient object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 16826-16835 (2021) +51. Zheng, P., Fu, H., Fan, D.P., Fan, Q., Qin, J., Tai, Y.W., Tang, C.K., Van Gool, L.: Gconet+: A stronger group collaborative co-salient object detector. IEEE Trans. Pattern Anal. Mach. Intell. (2023) +52. Zhuge, M., Fan, D.P., Liu, N., Zhang, D., Xu, D., Shao, L.: Salient object detection via integrity learning. IEEE Trans. Pattern Anal. Mach. Intell. 45(3), 3738-3752 (2022) \ No newline at end of file diff --git a/2024/CONDA_ Condensed Deep Association Learning for Co-Salient Object Detection./images.zip b/2024/CONDA_ Condensed Deep Association Learning for Co-Salient Object Detection./images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2bd9deffbfe6824b2b36e61760b2e3f0058279bb --- /dev/null +++ b/2024/CONDA_ Condensed Deep Association Learning for Co-Salient Object Detection./images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a780560259c4c9b6bd2209a750d633e3ff72a6e979681b1606fc1c1628889a7 +size 584930 diff --git a/2024/CONDA_ Condensed Deep Association Learning for Co-Salient Object Detection./layout.json b/2024/CONDA_ Condensed Deep Association Learning for Co-Salient Object Detection./layout.json new file mode 100644 index 0000000000000000000000000000000000000000..05ef7ef48b543f485906dfee1612f76ffe8d3333 --- /dev/null +++ b/2024/CONDA_ Condensed Deep Association Learning for Co-Salient Object Detection./layout.json @@ -0,0 +1,12043 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "type": "text", + "content": "CONDA: Condensed Deep Association Learning for Co-Salient Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 135, + 167, + 478, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 167, + 478, + 204 + ], + "spans": [ + { + "bbox": [ + 135, + 167, + 478, + 204 + ], + "type": "text", + "content": "Long Li" + }, + { + "bbox": [ + 135, + 167, + 478, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 167, + 478, + 204 + ], + "type": "text", + "content": ", Nian Liu" + }, + { + "bbox": [ + 135, + 167, + 478, + 204 + ], + "type": "inline_equation", + "content": "^{3,*}" + }, + { + "bbox": [ + 135, + 167, + 478, + 204 + ], + "type": "text", + "content": ", Dingwen Zhang" + }, + { + "bbox": [ + 135, + 167, + 478, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 167, + 478, + 204 + ], + "type": "text", + "content": ", Zhongyu Li" + }, + { + "bbox": [ + 135, + 167, + 478, + 204 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 135, + 167, + 478, + 204 + ], + "type": "text", + "content": ", Salman Khan" + }, + { + "bbox": [ + 135, + 167, + 478, + 204 + ], + "type": "inline_equation", + "content": "^{3,5}" + }, + { + "bbox": [ + 135, + 167, + 478, + 204 + ], + "type": "text", + "content": ", Rao Anwer" + }, + { + "bbox": [ + 135, + 167, + 478, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 135, + 167, + 478, + 204 + ], + "type": "text", + "content": ", Hisham Cholakkal" + }, + { + "bbox": [ + 135, + 167, + 478, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 135, + 167, + 478, + 204 + ], + "type": "text", + "content": ", Junwei Han" + }, + { + "bbox": [ + 135, + 167, + 478, + 204 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 135, + 167, + 478, + 204 + ], + "type": "text", + "content": ", and Fahad Shahbaz Khan" + }, + { + "bbox": [ + 135, + 167, + 478, + 204 + ], + "type": "inline_equation", + "content": "^{3,6}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 223, + 213, + 391, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 213, + 391, + 225 + ], + "spans": [ + { + "bbox": [ + 223, + 213, + 391, + 225 + ], + "type": "text", + "content": "1 Northwestern Polytechnical University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 224, + 473, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 224, + 473, + 236 + ], + "spans": [ + { + "bbox": [ + 140, + 224, + 473, + 236 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 224, + 473, + 236 + ], + "type": "text", + "content": " Institute of Artificial Intelligence, Hefei Comprehensive National Science Center" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 189, + 236, + 424, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 236, + 424, + 247 + ], + "spans": [ + { + "bbox": [ + 189, + 236, + 424, + 247 + ], + "type": "text", + "content": "3 Mohamed bin Zayed University of Artificial Intelligence" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 249, + 247, + 364, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 247, + 364, + 258 + ], + "spans": [ + { + "bbox": [ + 249, + 247, + 364, + 258 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 249, + 247, + 364, + 258 + ], + "type": "text", + "content": " Xi'an Jiaotong University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 239, + 258, + 374, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 258, + 374, + 269 + ], + "spans": [ + { + "bbox": [ + 239, + 258, + 374, + 269 + ], + "type": "text", + "content": "5 Australian National University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 246, + 269, + 367, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 269, + 367, + 281 + ], + "spans": [ + { + "bbox": [ + 246, + 269, + 367, + 281 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 246, + 269, + 367, + 281 + ], + "type": "text", + "content": " CVL, Linköping University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 312, + 455, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 312, + 455, + 553 + ], + "spans": [ + { + "bbox": [ + 159, + 312, + 455, + 553 + ], + "type": "text", + "content": "Abstract. Inter-image association modeling is crucial for co-salient object detection. Despite satisfactory performance, previous methods still have limitations on sufficient inter-image association modeling. Because most of them focus on image feature optimization under the guidance of heuristically calculated raw inter-image associations. They directly rely on raw associations which are not reliable in complex scenarios, and their image feature optimization approach is not explicit for inter-image association modeling. To alleviate these limitations, this paper proposes a deep association learning strategy that deploys deep networks on raw associations to explicitly transform them into deep association features. Specifically, we first create hyperassociations to collect dense pixel-pair-wise raw associations and then deploys deep aggregation networks on them. We design a progressive association generation module for this purpose with additional enhancement of the hyperassociation calculation. More importantly, we propose a correspondence-induced association condensation module that introduces a pretext task, i.e. semantic correspondence estimation, to condense the hyperassociations for computational burden reduction and noise elimination. We also design an object-aware cycle consistency loss for high-quality correspondence estimations. Experimental results in three benchmark datasets demonstrate the remarkable effectiveness of our proposed method with various training settings. The code is available at: https://github.com/dragonlee258079/CONDA." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 160, + 563, + 447, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 563, + 447, + 575 + ], + "spans": [ + { + "bbox": [ + 160, + 563, + 447, + 575 + ], + "type": "text", + "content": "Keywords: Co-salient Object Detection " + }, + { + "bbox": [ + 160, + 563, + 447, + 575 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 563, + 447, + 575 + ], + "type": "text", + "content": " Deep Association Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 596, + 230, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 596, + 230, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 596, + 230, + 609 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 622, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 622, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 622, + 482, + 647 + ], + "type": "text", + "content": "Co-Salient Object Detection (CoSOD) aims to segment salient objects that appear commonly across a group of related images. Compared to traditional Salient" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 414, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 414, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 414, + 666 + ], + "type": "text", + "content": "* Corresponding authors: {liunian228, junweihan2010}@gmail.com" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 360, + 242 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 360, + 242 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 360, + 242 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 360, + 242 + ], + "type": "image", + "image_path": "404824b20790bd64d1da0a79eb3ae6b9a0ab43b402bf8d6b4670d49ac69f6027.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 243, + 482, + 310 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 310 + ], + "type": "text", + "content": "Fig. 1: Difference of raw-association-based image feature optimization strategy (a) and our proposed deep association learning strategy (b). Our deep association learning deploys deep learning networks on raw associations to achieve deep association features. We also present visual samples of our calculated raw associations (Raw Asso.), optimized image feature (Opt. Fea.), and our generated deep association features (Asso. Fea.) in (c)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 363, + 116, + 479, + 239 + ], + "blocks": [ + { + "bbox": [ + 363, + 116, + 479, + 239 + ], + "lines": [ + { + "bbox": [ + 363, + 116, + 479, + 239 + ], + "spans": [ + { + "bbox": [ + 363, + 116, + 479, + 239 + ], + "type": "image", + "image_path": "7f9cc2f91034f063da20762cb03c1295457ec68c8b6828e05a7cdde337347203.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 316, + 481, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 481, + 341 + ], + "type": "text", + "content": "Object Detection (SOD) [19-23, 26, 49, 50, 52], CoSOD is a more challenging task because it requires sufficient inter-image association modeling [9]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 342, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 557 + ], + "type": "text", + "content": "Recently, many advanced works [8,13,16,33,38,41,46,47,51] have emerged and achieved impressive performance. These methods first use related image features to acquire raw inter-image associations (also called consensus representations), and then leverage them as guidance to optimize each image feature, as shown in Figure 1 (a). This approach enables the final image feature to implicitly capture inter-image cues, thereby achieving the purpose of inter-image association modeling. However, we find this raw-association-based feature optimization strategy still has two limitations: i) they directly rely on raw associations, which are acquired in heuristic manners, such as pixel-wise [8,33,41,51], region-wise [16,46], or image-wise [13] similarity measurements. Although high-quality raw associations can be derived from high-level semantic information in image features, their revelation of common saliency regions still relies on similarity measures, which are unreliable when encountering complex scenarios, such as significant differences between co-salient objects or high foreground-background similarity. ii) the primary focus of building their deep models is on optimizing image features. Compared to directly modeling association relationship, image feature optimization is not an explicit approach for inter-image association modeling and will increase the learning difficulty." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "To alleviate these limitations, we propose a deep association learning strategy for CoSOD, as shown in Figure 1 (b). Instead of directly using raw associations to optimize image features, we deploy deep networks on raw associations to learn deep association features. This is a more explicit strategy for inter-image association modeling. Moreover, our deep association features can capture high-level inter-image association knowledge, making them more robust in complex scenarios than raw associations, as shown in Figure 1 (c). Technically, we start by collecting all pixel-wise raw associations across the entire image group as hyperassociations. Then, we propose a Progressive Association Generation (PAG)" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 209, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 209, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 209, + 101 + ], + "type": "text", + "content": "L Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "module to transform hyperassociations into deep association features. PAG progressively generates association features on varying scales, allowing us to use previous association features to enhance the hyperassociation calculation at the next scale, thereby improving the association quality from the very beginning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 271 + ], + "type": "text", + "content": "Although deep association learning strategy allows more sufficient interimage association modeling, it significantly increases the computational burden and reduces the practicality of this approach. Additionally, this study finds that it is not necessary to utilize all pixel associations to generate deep association features. In fact, there are even some noisy pixels that negatively impact the quality of the deep association features. Therefore, we propose a method based on Correspondence-induced Association Condensation (CAC) to condense the original full-pixel hyperassociations. This not only alleviates the computational burden but also further enhances the quality of deep association features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 272, + 481, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 481, + 403 + ], + "type": "text", + "content": "Specifically, CAC performs the condensation operation by selectively associating pixels that have semantic correspondence in other images, as well as their surrounding contextual pixels, thereby creating lightweight yet more accurate hyperassociations. Here, we introduce a pretext task, i.e. semantic correspondence estimation, into the CoSOD, not only improving the model performance but also delving deeper into the essence of CoSOD. Co-salient objects inherently possess object-level semantic correspondence. However, in this paper, we aim to further explore the finer pixel-level correspondence. Although highly accurate correspondence estimation remains a challenge, we believe it will pave a new way for CoSOD research. We also provide an object-aware cycle consistency (OCC) loss to aid in learning correspondences within co-salient pixels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 403, + 405, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 403, + 405, + 415 + ], + "spans": [ + { + "bbox": [ + 146, + 403, + 405, + 415 + ], + "type": "text", + "content": "In summary, the contributions of this paper are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 417, + 479, + 582 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 138, + 417, + 479, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 417, + 479, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 417, + 479, + 464 + ], + "type": "text", + "content": "- For the first time, we introduce a deep association learning approach for CoSOD, applying deep networks to transform raw associations into deep association features for sufficient inter-image association modeling. Specifically, we develop a CONdensed Deep Association (CONDA) learning model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 464, + 479, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 464, + 479, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 464, + 479, + 499 + ], + "type": "text", + "content": "- We propose a PAG module to progressively generate deep association features. It enhances image features with previous association features to improve hyperassociation calculation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 500, + 479, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 500, + 479, + 546 + ], + "spans": [ + { + "bbox": [ + 137, + 500, + 479, + 546 + ], + "type": "text", + "content": "- We introduce semantic correspondence estimation into the CoSOD task to condense the original hyperassociation for alleviating the computational burden and further improving the performance. We also propose an OCC loss for effective correspondence estimation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 547, + 479, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 547, + 479, + 582 + ], + "spans": [ + { + "bbox": [ + 137, + 547, + 479, + 582 + ], + "type": "text", + "content": "- Experimental results demonstrate that our model achieves significantly improved state-of-the-art performance on three benchmark datasets across different training settings." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 599, + 237, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 237, + 612 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 237, + 612 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 623, + 303, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 303, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 303, + 635 + ], + "type": "text", + "content": "2.1 Co-Salient Object Detection" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Recently, there has been a surge of excellent methods for CoSOD [8,10,13,16,33, 38,41-43,45-47,51]. These methods initially acquire raw inter-image associations" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "CONDA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "with related image features and then utilize them to optimize each image feature. Most methods generate pixel-wise [8, 33, 41, 51], region-wise [16, 46], and image-wise [13] raw associations through similarity-based manners, e.g. inner product calculations between image features. Even for transformer-based methods [16, 33], they also rely on inner product calculation to produce attention maps [36] as raw associations. The image feature optimization manners include spatial or channel calibration [8, 13, 51], dynamic convolution [41, 46], feature fusion [16, 41], and cross-attention [16, 33], etc. However, they lack the learning of high-level association knowledge and heavily focus on optimized image features. Unlike them, this paper proposes a new research direction that deploys deep networks on associations to achieve deep association features for CoSOD." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 265, + 317, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 265, + 317, + 277 + ], + "spans": [ + { + "bbox": [ + 131, + 265, + 317, + 277 + ], + "type": "text", + "content": "2.2 Inter-image Relation Modeling" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 285, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 453 + ], + "type": "text", + "content": "Apart from CoSOD, other tasks necessitating the consideration of inter-image relationships, such as few-shot segmentation [25, 27, 28, 39], stereo matching [2, 40], video semantic segmentation [34], etc. These tasks have made significant advancements recently by effectively modeling inter-image relations. Most of these methods [2, 4, 12, 27, 39, 40] initially create cost volumes to capture dense inter-image pixel-wise similarities and subsequently use various modules to convert these hyper volumes into task-specific features. Our approach distinguishes itself from prior methods in three aspects. Firstly, most of them create 4D cost volumes between two images while we create 6D hyperassociations between all related images. Secondly, they calculate hyper volumes based on original image features, whereas We propose PAG to progressively enhance image features for better hyperassociation calculation. Last and most importantly, they rely on the full-pixel cost volume. We consider the condensation of hyperassociations using semantic correspondences to eliminate noisy pixel associations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 470, + 350, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 470, + 350, + 483 + ], + "spans": [ + { + "bbox": [ + 131, + 470, + 350, + 483 + ], + "type": "text", + "content": "2.3 Semantic Correspondence Estimation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 490, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 598 + ], + "type": "text", + "content": "Semantic correspondence estimation [24] aims to establish reliable pixel correspondences between different instances of the same object category. Most works performed this task using fully supervised training [14, 24]. Some recent works have utilized unsupervised learning with photometric, forward-backward consistency, and warp-supervision losses [30, 35, 48]. However, they implement these losses on the entire image, where background pixels may affect the performance. In this paper, we introduce this task to condense hyperassociations for Co-SOD and tailor the cycle consistency loss by only applying it to co-salient pixels, hence effectively avoiding the influence of background and extraneous objects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 616, + 262, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 616, + 262, + 630 + ], + "spans": [ + { + "bbox": [ + 131, + 616, + 262, + 630 + ], + "type": "text", + "content": "3 Proposed Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "text", + "content": "As shown in Figure 2, CONDA integrates the deep association learning process into an FPN framework. Specifically, given a group of related images " + }, + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "inline_equation", + "content": "\\{I_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "text", + "content": ", we" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "L Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 479, + 318 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 479, + 318 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 479, + 318 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 479, + 318 + ], + "type": "image", + "image_path": "ac6fd1de06c08cc86d786e4ae99a47c98f05aa45788c8c35362c4a5254b32dd3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 321, + 482, + 377 + ], + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 377 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 377 + ], + "type": "text", + "content": "Fig. 2: Overall flowchart of our CONDA model. Specifically, CONDA first utilizes the image features to calculate hyperassociations. Then, the full-pixel hyperassociations are condensed by CAC and fed into the aggregation networks to achieve deep association features. These features are then used in the FPN decoder process for the final prediction. To be concise, only three related images are shown." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 384, + 482, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 421 + ], + "type": "text", + "content": "first input them into a VGG-16 [32] backbone to collect its intermediate features for PAG and FPN decoding. In detail, we collect all features of the last three stages for PAG and the last feature of each stage for the FPN decoder as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 206, + 430, + 480, + 461 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 430, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 206, + 430, + 480, + 461 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbb {F} ^ {P} = \\left\\{\\boldsymbol {F} _ {s, l} \\mid s \\in \\{3, \\dots , 5 \\}, l \\in \\{1, \\dots , L _ {s} \\} \\right\\}, \\\\ \\mathbb {F} ^ {D} = \\left\\{\\boldsymbol {F} _ {s, L _ {s}} \\mid s \\in \\{1, \\dots 5 \\} \\right\\}, \\tag {1} \\\\ \\end{array}", + "image_path": "f58edb3002b15af5f1a8024648c5d8ba79058d07a994af6694f7b9a7398b3d7b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "inline_equation", + "content": "\\mathbb{F}^P" + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "inline_equation", + "content": "\\mathbb{F}^D" + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "text", + "content": " are the feature collections for PAG and FPN decoder, respectively. " + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "inline_equation", + "content": "F_{s,l}\\in \\mathbb{R}^{N\\times H_s\\times W_s\\times C_s}" + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "text", + "content": " represents the VGG feature of the " + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "text", + "content": "-th layer in the " + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "text", + "content": "-th stage. There are in total " + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "text", + "content": " layers in the " + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "text", + "content": "-th stage. " + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "inline_equation", + "content": "H_{s}" + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "inline_equation", + "content": "W_{s}" + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "inline_equation", + "content": "C_s" + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "text", + "content": " represent the height, width, and channel of the " + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "text", + "content": "-th stage, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 517, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 482, + 554 + ], + "type": "text", + "content": "Then, we input " + }, + { + "bbox": [ + 130, + 517, + 482, + 554 + ], + "type": "inline_equation", + "content": "\\mathbb{F}^P" + }, + { + "bbox": [ + 130, + 517, + 482, + 554 + ], + "type": "text", + "content": " into PAG to calculate hyperassociations and generate deep association features " + }, + { + "bbox": [ + 130, + 517, + 482, + 554 + ], + "type": "inline_equation", + "content": "\\{\\pmb {F}_s^A\\in \\mathbb{R}^{N\\times H_s\\times W_s\\times C_s}\\}_{s = 3}^5" + }, + { + "bbox": [ + 130, + 517, + 482, + 554 + ], + "type": "text", + "content": ". Finally, these association features will be fused with " + }, + { + "bbox": [ + 130, + 517, + 482, + 554 + ], + "type": "inline_equation", + "content": "\\mathbb{F}^D" + }, + { + "bbox": [ + 130, + 517, + 482, + 554 + ], + "type": "text", + "content": " for the FPN decoder process, formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 219, + 563, + 393, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 563, + 393, + 594 + ], + "spans": [ + { + "bbox": [ + 219, + 563, + 393, + 594 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {F} _ {s, L _ {s}} = \\boldsymbol {F} _ {s, L _ {s}} + \\phi (\\boldsymbol {F} _ {s} ^ {A}), \\quad s = 3, \\dots , 5, \\\\ \\boldsymbol {F} = \\operatorname {D e c o d e r} \\left(\\left\\{\\boldsymbol {F} _ {s, L _ {s}} \\right\\} _ {s = 1} ^ {5}\\right), \\\\ \\end{array}", + "image_path": "85d8c453a6781faceb432022fa8073503e91ac5a1d745fba847a589fa042c463.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 604, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 604, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 604, + 482, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 604, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 604, + 482, + 628 + ], + "type": "text", + "content": " is a convolution layer. " + }, + { + "bbox": [ + 130, + 604, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\pmb{F} \\in \\mathbb{R}^{N \\times H \\times W \\times C}" + }, + { + "bbox": [ + 130, + 604, + 482, + 628 + ], + "type": "text", + "content": " is the final feature for the final co-saliency prediction. We adopt BCE and IoU losses for supervision." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "The rest of this section will introduce PAG with full-pixel hyperassociation and the condensation of hyperassociations by plugging the Correspondence-induced Association Condensation (CAC) module into PAG." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "CONDA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 339, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 339, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 339, + 128 + ], + "type": "text", + "content": "3.1 Progressive Association Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 479, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 479, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 479, + 172 + ], + "type": "text", + "content": "Our deep association learning involves two steps: 1) acquiring the raw hyper-associations " + }, + { + "bbox": [ + 130, + 137, + 479, + 172 + ], + "type": "inline_equation", + "content": "\\{A_s\\}_{s=3}^5" + }, + { + "bbox": [ + 130, + 137, + 479, + 172 + ], + "type": "text", + "content": " from three stages; 2) employing aggregation networks on " + }, + { + "bbox": [ + 130, + 137, + 479, + 172 + ], + "type": "inline_equation", + "content": "\\{A_s\\}_{s=3}^5" + }, + { + "bbox": [ + 130, + 137, + 479, + 172 + ], + "type": "text", + "content": " to obtain association features " + }, + { + "bbox": [ + 130, + 137, + 479, + 172 + ], + "type": "inline_equation", + "content": "\\{F_s^A\\}_{s=3}^5" + }, + { + "bbox": [ + 130, + 137, + 479, + 172 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 173, + 479, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 173, + 479, + 232 + ], + "spans": [ + { + "bbox": [ + 130, + 173, + 479, + 232 + ], + "type": "text", + "content": "Earlier methods [4, 11, 31] directly utilize the original backbone features, i.e. " + }, + { + "bbox": [ + 130, + 173, + 479, + 232 + ], + "type": "inline_equation", + "content": "\\mathbb{F}^P" + }, + { + "bbox": [ + 130, + 173, + 479, + 232 + ], + "type": "text", + "content": ", to calculate inter-image interactions, such as the so-called cost volume. We argue that hyperassociations derived straight from backbone features might obstruct further improvement of deep association learning, given that the current backbone is pre-trained without any consideration for inter-image associations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 232, + 480, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 232, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 480, + 341 + ], + "type": "text", + "content": "To alleviate this problem, we propose the PAG module to progressively generate pyramid association features so that we can utilize the high-level association feature, e.g. " + }, + { + "bbox": [ + 130, + 232, + 480, + 341 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{s+1}^{A}" + }, + { + "bbox": [ + 130, + 232, + 480, + 341 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 130, + 232, + 480, + 341 + ], + "type": "inline_equation", + "content": "(s+1)" + }, + { + "bbox": [ + 130, + 232, + 480, + 341 + ], + "type": "text", + "content": "-th stage, to enhance the neighbouring low-level VGG features in " + }, + { + "bbox": [ + 130, + 232, + 480, + 341 + ], + "type": "inline_equation", + "content": "\\mathbb{F}^P" + }, + { + "bbox": [ + 130, + 232, + 480, + 341 + ], + "type": "text", + "content": ", e.g. " + }, + { + "bbox": [ + 130, + 232, + 480, + 341 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{s,l}" + }, + { + "bbox": [ + 130, + 232, + 480, + 341 + ], + "type": "text", + "content": ", for attaining association-enhanced features " + }, + { + "bbox": [ + 130, + 232, + 480, + 341 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{F}}_{s,l}" + }, + { + "bbox": [ + 130, + 232, + 480, + 341 + ], + "type": "text", + "content": ", based on which we can calculate high-quality hyperassociations " + }, + { + "bbox": [ + 130, + 232, + 480, + 341 + ], + "type": "inline_equation", + "content": "\\pmb{A}_s" + }, + { + "bbox": [ + 130, + 232, + 480, + 341 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 130, + 232, + 480, + 341 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 232, + 480, + 341 + ], + "type": "text", + "content": "-th stage. After that, we execute the subsequent aggregation networks on " + }, + { + "bbox": [ + 130, + 232, + 480, + 341 + ], + "type": "inline_equation", + "content": "\\pmb{A}_s" + }, + { + "bbox": [ + 130, + 232, + 480, + 341 + ], + "type": "text", + "content": " to achieve association features " + }, + { + "bbox": [ + 130, + 232, + 480, + 341 + ], + "type": "inline_equation", + "content": "\\pmb{F}_s^A" + }, + { + "bbox": [ + 130, + 232, + 480, + 341 + ], + "type": "text", + "content": ", which will continue to enhance the VGG features of the next stage and carry out progressive association generation. The whole process of our PAG can be formulated as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 256, + 351, + 358, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 351, + 358, + 365 + ], + "spans": [ + { + "bbox": [ + 256, + 351, + 358, + 365 + ], + "type": "interline_equation", + "content": "\\boldsymbol {A} _ {s} = \\operatorname {H A C} \\left(\\left\\{\\tilde {\\boldsymbol {F}} _ {s, l} \\right\\} _ {l = 1} ^ {L _ {s}}\\right),", + "image_path": "ffbb404d7bd2d9562a83e9eefa11ad8a263592bcc4182df0dc6285318f158359.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 271, + 368, + 480, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 368, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 271, + 368, + 480, + 381 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} _ {s} ^ {A} = \\operatorname {A g g} \\left(\\boldsymbol {A} _ {s}\\right), \\tag {3}", + "image_path": "658a000c8e33824e50e99798de10d042ec502da8f944ffb3c228c5f580b041e6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 219, + 384, + 395, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 384, + 395, + 399 + ], + "spans": [ + { + "bbox": [ + 219, + 384, + 395, + 399 + ], + "type": "interline_equation", + "content": "\\{\\hat {\\boldsymbol {F}} _ {s - 1, l} \\} _ {l = 1} ^ {L _ {s - 1}} = \\operatorname {E n h} \\left(\\left\\{\\boldsymbol {F} _ {s - 1, l} \\right\\} _ {l = 1} ^ {L _ {s - 1}}; \\boldsymbol {F} _ {s} ^ {A}\\right),", + "image_path": "e1ce7b5a29ad1a72a0143f12355f23e961712aed11af39cec89c00e660780969.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 410, + 479, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 479, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 479, + 445 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 410, + 479, + 445 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 410, + 479, + 445 + ], + "type": "text", + "content": " ranges from 5 to 3 and " + }, + { + "bbox": [ + 130, + 410, + 479, + 445 + ], + "type": "inline_equation", + "content": "\\{\\hat{F}_{5,l}\\}_{l=1}^{L_5} = \\{\\pmb{F}_{5,l}\\}_{l=1}^{L_5}" + }, + { + "bbox": [ + 130, + 410, + 479, + 445 + ], + "type": "text", + "content": ". The HAC, Agg, and Enh represent the hyperassociation calculation, aggregation network, and association-induced feature enhancement, respectively. Next, we explain them in detail." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 447, + 480, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 480, + 506 + ], + "type": "text", + "content": "Hyperassociation Calculation. For each stage, we first compute the raw associations at each layer using the inner product between " + }, + { + "bbox": [ + 130, + 447, + 480, + 506 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 447, + 480, + 506 + ], + "type": "text", + "content": "-2 normalized association-enhanced features of " + }, + { + "bbox": [ + 130, + 447, + 480, + 506 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 447, + 480, + 506 + ], + "type": "text", + "content": " related images. After that, we stack the raw associations of all layers to form the final hyperassociation of this stage. The hyperassociation for the " + }, + { + "bbox": [ + 130, + 447, + 480, + 506 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 447, + 480, + 506 + ], + "type": "text", + "content": "-th stage, i.e. " + }, + { + "bbox": [ + 130, + 447, + 480, + 506 + ], + "type": "inline_equation", + "content": "A_{s} \\in \\mathbb{R}^{N \\times H_{s} \\times W_{s} \\times N \\times H_{s} \\times W_{s} \\times L_{s}}" + }, + { + "bbox": [ + 130, + 447, + 480, + 506 + ], + "type": "text", + "content": ", can be calculated via:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 213, + 516, + 480, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 516, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 213, + 516, + 480, + 567 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {A} _ {s} = \\operatorname {H A C} \\left(\\left\\{\\hat {\\boldsymbol {F}} _ {s, l} \\right\\} _ {l = 1} ^ {L _ {s}}\\right) \\\\ = \\operatorname {S t a c k} \\left(\\left\\{\\operatorname {R e L U} \\left(\\frac {\\hat {\\boldsymbol {F}} _ {s , l} \\cdot \\hat {\\boldsymbol {F}} _ {s , l} ^ {\\top}}{\\| \\hat {\\boldsymbol {F}} _ {s , l} \\| \\| \\hat {\\boldsymbol {F}} _ {s , l} ^ {\\top} \\|}\\right) \\right\\} _ {l = 1} ^ {L _ {s}}\\right), \\tag {4} \\\\ \\end{array}", + "image_path": "dfb4b0afc19500e431f378c15eb14e37ee1c572f0eddd0090b8cfc0027752546.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 578, + 479, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 479, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 479, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 578, + 479, + 616 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{F}}_{s,l}\\cdot \\hat{\\pmb{F}}_{s,l}^{\\top}\\in \\mathbb{R}^{N\\times H_s\\times W_s\\times N\\times H_s\\times W_s}" + }, + { + "bbox": [ + 130, + 578, + 479, + 616 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 130, + 578, + 479, + 616 + ], + "type": "inline_equation", + "content": "\\top" + }, + { + "bbox": [ + 130, + 578, + 479, + 616 + ], + "type": "text", + "content": " indicates transposing the last dimension and the first three dimensions. " + }, + { + "bbox": [ + 130, + 578, + 479, + 616 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|" + }, + { + "bbox": [ + 130, + 578, + 479, + 616 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 130, + 578, + 479, + 616 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 578, + 479, + 616 + ], + "type": "text", + "content": "-2 norm. We employ ReLU to suppress noisy association values." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 616, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 616, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 616, + 479, + 665 + ], + "type": "text", + "content": "Aggregation Network. The raw hyperassociation " + }, + { + "bbox": [ + 130, + 616, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_s \\in \\mathbb{R}^{N \\times H_s \\times W_s \\times N \\times H_s \\times W_s \\times L_s}" + }, + { + "bbox": [ + 130, + 616, + 479, + 665 + ], + "type": "text", + "content": " is a hypercube with a nested structure, where each pixel position is characterized by a 4D tensor " + }, + { + "bbox": [ + 130, + 616, + 479, + 665 + ], + "type": "inline_equation", + "content": "(N \\times H_s \\times W_s \\times L_s)" + }, + { + "bbox": [ + 130, + 616, + 479, + 665 + ], + "type": "text", + "content": ". Each 4D tensor documents the associations of the respective pixel with all other pixels in " + }, + { + "bbox": [ + 130, + 616, + 479, + 665 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 616, + 479, + 665 + ], + "type": "text", + "content": " related images. For" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "L Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "clarity, we designate the first and second " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "N \\times H_s \\times W_s" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " dimensions in " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "A_s" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " as the source and target dimensions, respectively. Although these 4D tensors are crucial for exploring the consensus information for co-saliency detection, they essentially comprise pixel-to-pixel similarity values, seen in (4), which may be suboptimal and unreliable in complex scenarios. Therefore, we propose using deep networks to transform these pixel-wise similarities into deep association features with contextual and high-order association knowledge. This has never been explored in previous CoSOD methods. This is implemented via context aggregations on " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "A_s" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " to squeeze these 4D tensors as " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "C_s" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "-dimensional vectors, formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 194, + 234, + 481, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 234, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 194, + 234, + 481, + 247 + ], + "type": "interline_equation", + "content": "N \\times H _ {s} \\times W _ {s} \\times N \\times H _ {s} \\times W _ {s} \\times L _ {s} \\rightarrow N \\times H _ {s} \\times W _ {s} \\times C _ {s}. \\tag {5}", + "image_path": "e16b36f7023d1b9537f9dc3875769ad4f581e58ff5664b27edaef5585a6de38d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 255, + 482, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 482, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 482, + 351 + ], + "type": "text", + "content": "In detail, we first deploy several association aggregation layers on " + }, + { + "bbox": [ + 130, + 255, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_s" + }, + { + "bbox": [ + 130, + 255, + 482, + 351 + ], + "type": "text", + "content": " to progressively aggregate context information, enlarging " + }, + { + "bbox": [ + 130, + 255, + 482, + 351 + ], + "type": "inline_equation", + "content": "L_s" + }, + { + "bbox": [ + 130, + 255, + 482, + 351 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 255, + 482, + 351 + ], + "type": "inline_equation", + "content": "C_s" + }, + { + "bbox": [ + 130, + 255, + 482, + 351 + ], + "type": "text", + "content": ", and eliminate the target " + }, + { + "bbox": [ + 130, + 255, + 482, + 351 + ], + "type": "inline_equation", + "content": "H_s \\times W_s" + }, + { + "bbox": [ + 130, + 255, + 482, + 351 + ], + "type": "text", + "content": " dimension in 4D tensors. Each aggregation layer consists of 2D convolution layers and a downsampling operation. Specifically, focusing on the first aggregation layer for technical explanation, we first aggregate context information in the target " + }, + { + "bbox": [ + 130, + 255, + 482, + 351 + ], + "type": "inline_equation", + "content": "H_s \\times W_s" + }, + { + "bbox": [ + 130, + 255, + 482, + 351 + ], + "type": "text", + "content": " dimension by applying a 2D convolution layer on all 4D tensors. The operations on the 4D tensor at pixel position " + }, + { + "bbox": [ + 130, + 255, + 482, + 351 + ], + "type": "inline_equation", + "content": "(h_i, w_i)" + }, + { + "bbox": [ + 130, + 255, + 482, + 351 + ], + "type": "text", + "content": " in image " + }, + { + "bbox": [ + 130, + 255, + 482, + 351 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 130, + 255, + 482, + 351 + ], + "type": "text", + "content": " can be formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 172, + 360, + 481, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 360, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 172, + 360, + 481, + 374 + ], + "type": "interline_equation", + "content": "\\boldsymbol {A} _ {s} ^ {1} (i, h _ {i}, w _ {i}, j, \\therefore , \\therefore): = \\mathcal {C} _ {1} \\left(\\boldsymbol {A} _ {s} (i, h _ {i}, w _ {i}, j, \\therefore , \\therefore)\\right), \\quad j = 1, \\dots , N, \\tag {6}", + "image_path": "1ea6287f0d436d9822e65343489147bcb6f361f5f22950ee160f1ed074d1bdd8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 382, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 482, + 430 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 382, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_1" + }, + { + "bbox": [ + 130, + 382, + 482, + 430 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 130, + 382, + 482, + 430 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 130, + 382, + 482, + 430 + ], + "type": "text", + "content": " 2D convolution layer. Here, " + }, + { + "bbox": [ + 130, + 382, + 482, + 430 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 382, + 482, + 430 + ], + "type": "text", + "content": " is the index of other related images in the 4D tensor, and " + }, + { + "bbox": [ + 130, + 382, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\pmb {A}_s(i,h_i,w_i,j,:;:)\\in \\mathbb{R}^{H_s\\times W_s\\times L_s}" + }, + { + "bbox": [ + 130, + 382, + 482, + 430 + ], + "type": "text", + "content": " illustrates the associations between the pixel " + }, + { + "bbox": [ + 130, + 382, + 482, + 430 + ], + "type": "inline_equation", + "content": "(h_i,w_i)" + }, + { + "bbox": [ + 130, + 382, + 482, + 430 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 382, + 482, + 430 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 130, + 382, + 482, + 430 + ], + "type": "text", + "content": " and all pixels in image " + }, + { + "bbox": [ + 130, + 382, + 482, + 430 + ], + "type": "inline_equation", + "content": "I_{j}" + }, + { + "bbox": [ + 130, + 382, + 482, + 430 + ], + "type": "text", + "content": ". This interpretation applies to other similar symbols." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 430, + 481, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 455 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 455 + ], + "type": "text", + "content": "Then, a downsampling operation " + }, + { + "bbox": [ + 132, + 430, + 481, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 132, + 430, + 481, + 455 + ], + "type": "text", + "content": ", i.e. bilinear interpolation, is applied on " + }, + { + "bbox": [ + 132, + 430, + 481, + 455 + ], + "type": "inline_equation", + "content": "\\pmb{A}_s^1" + }, + { + "bbox": [ + 132, + 430, + 481, + 455 + ], + "type": "text", + "content": " to reduce the spatial dimension of the 4D tensor by a scaling factor:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 208, + 464, + 481, + 477 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 464, + 481, + 477 + ], + "spans": [ + { + "bbox": [ + 208, + 464, + 481, + 477 + ], + "type": "interline_equation", + "content": "\\boldsymbol {A} _ {s} ^ {2} (i, h _ {i}, w _ {i}, j, \\therefore , \\therefore ,:) = \\mathcal {D} (\\boldsymbol {A} _ {s} ^ {1} (i, h _ {i}, w _ {i}, j, \\therefore , \\therefore)), \\tag {7}", + "image_path": "d936f5e8002e3aaff31d191b491b52ccee8e86026104f9f8cdda9634a23fa392.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 486, + 481, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 481, + 512 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 486, + 481, + 512 + ], + "type": "inline_equation", + "content": "A_{s}^{2}(i,h_{i},w_{i},j,:,:) \\in \\mathbb{R}^{H_{s}^{\\prime} \\times W_{s}^{\\prime} \\times C_{s}^{\\prime}}" + }, + { + "bbox": [ + 130, + 486, + 481, + 512 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 486, + 481, + 512 + ], + "type": "inline_equation", + "content": "H_{s}^{\\prime}" + }, + { + "bbox": [ + 130, + 486, + 481, + 512 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 486, + 481, + 512 + ], + "type": "inline_equation", + "content": "W_{s}^{\\prime}" + }, + { + "bbox": [ + 130, + 486, + 481, + 512 + ], + "type": "text", + "content": " are downsampled height and width. " + }, + { + "bbox": [ + 130, + 486, + 481, + 512 + ], + "type": "inline_equation", + "content": "C_{s}^{\\prime}" + }, + { + "bbox": [ + 130, + 486, + 481, + 512 + ], + "type": "text", + "content": " is the channel number after convolution " + }, + { + "bbox": [ + 130, + 486, + 481, + 512 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_1" + }, + { + "bbox": [ + 130, + 486, + 481, + 512 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 512, + 482, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 482, + 573 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 482, + 573 + ], + "type": "text", + "content": "Finally, we also aggregate context information in the source " + }, + { + "bbox": [ + 130, + 512, + 482, + 573 + ], + "type": "inline_equation", + "content": "H_{s} \\times W_{s}" + }, + { + "bbox": [ + 130, + 512, + 482, + 573 + ], + "type": "text", + "content": " dimension. Specifically, we extract the values along the source dimension and channel dimension in " + }, + { + "bbox": [ + 130, + 512, + 482, + 573 + ], + "type": "inline_equation", + "content": "A_{s}^{2}" + }, + { + "bbox": [ + 130, + 512, + 482, + 573 + ], + "type": "text", + "content": " to form 4D tensors and apply a 2D convolution layer on them. For instance, " + }, + { + "bbox": [ + 130, + 512, + 482, + 573 + ], + "type": "inline_equation", + "content": "A_{s}^{2}(:, :, :, j, h_{j}, w_{j},:) \\in \\mathbb{R}^{N \\times H_{s} \\times W_{s} \\times C_{s}^{\\prime}}" + }, + { + "bbox": [ + 130, + 512, + 482, + 573 + ], + "type": "text", + "content": " is such a 4D tensor, where " + }, + { + "bbox": [ + 130, + 512, + 482, + 573 + ], + "type": "inline_equation", + "content": "(h_{j}, w_{j})" + }, + { + "bbox": [ + 130, + 512, + 482, + 573 + ], + "type": "text", + "content": " is a pixel position in the target dimension. This can be formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 171, + 580, + 481, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 580, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 171, + 580, + 481, + 594 + ], + "type": "interline_equation", + "content": "\\boldsymbol {A} _ {s} ^ {3} (i,:,:) j, h _ {j}, w _ {j,:}) = \\mathcal {C} _ {2} \\left(\\boldsymbol {A} _ {s} ^ {2} (i,:,:) j, h _ {j}, w _ {j,:})\\right), \\quad i = 1, \\dots , N, \\tag {8}", + "image_path": "5e333e2665282db1fb6db4a204d54a292115d20a884cef16c0e93084de63ac36.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 602, + 439, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 602, + 439, + 614 + ], + "spans": [ + { + "bbox": [ + 130, + 602, + 439, + 614 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 602, + 439, + 614 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_2" + }, + { + "bbox": [ + 130, + 602, + 439, + 614 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 130, + 602, + 439, + 614 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 130, + 602, + 439, + 614 + ], + "type": "text", + "content": " 2D convolution layer. " + }, + { + "bbox": [ + 130, + 602, + 439, + 614 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 602, + 439, + 614 + ], + "type": "text", + "content": " is the related image index." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 615, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 615, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 615, + 482, + 666 + ], + "type": "text", + "content": "After several association aggregation layers, as shown in (6)-(8), we can achieve the aggregated association features with the target " + }, + { + "bbox": [ + 130, + 615, + 482, + 666 + ], + "type": "inline_equation", + "content": "H_{s} \\times W_{s}" + }, + { + "bbox": [ + 130, + 615, + 482, + 666 + ], + "type": "text", + "content": " dimension eliminated, denoted as " + }, + { + "bbox": [ + 130, + 615, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{s}^{A^{\\prime}} \\in \\mathbb{R}^{N \\times H_{s} \\times W_{s} \\times N \\times C}" + }, + { + "bbox": [ + 130, + 615, + 482, + 666 + ], + "type": "text", + "content": ". Subsequently, we average " + }, + { + "bbox": [ + 130, + 615, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{s}^{A^{\\prime}}" + }, + { + "bbox": [ + 130, + 615, + 482, + 666 + ], + "type": "text", + "content": " on its second " + }, + { + "bbox": [ + 130, + 615, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 615, + 482, + 666 + ], + "type": "text", + "content": " dimension and obtain the final association features" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "CONDA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 462, + 194 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 462, + 194 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 462, + 194 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 462, + 194 + ], + "type": "image", + "image_path": "e979f8630c372e8cd48870e7fafda8137440dd82e40736aff359426acda2b2dc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 194, + 482, + 281 + ], + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 281 + ], + "type": "text", + "content": "Fig. 3: Difference between full-pixel hyperassociation (a) and condensed hyperassociation (b). We provide an example of collecting the pixel associations from image " + }, + { + "bbox": [ + 130, + 194, + 482, + 281 + ], + "type": "inline_equation", + "content": "I_{j}" + }, + { + "bbox": [ + 130, + 194, + 482, + 281 + ], + "type": "text", + "content": " for a pixel " + }, + { + "bbox": [ + 130, + 194, + 482, + 281 + ], + "type": "inline_equation", + "content": "(h_{i},w_{i})" + }, + { + "bbox": [ + 130, + 194, + 482, + 281 + ], + "type": "text", + "content": " in image " + }, + { + "bbox": [ + 130, + 194, + 482, + 281 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 130, + 194, + 482, + 281 + ], + "type": "text", + "content": ". Full-pixel hyperassociation collects all pixel associations in " + }, + { + "bbox": [ + 130, + 194, + 482, + 281 + ], + "type": "inline_equation", + "content": "I_{j}" + }, + { + "bbox": [ + 130, + 194, + 482, + 281 + ], + "type": "text", + "content": ", while our condensed hyperassociation only collects the associations of its correspondence pixel " + }, + { + "bbox": [ + 130, + 194, + 482, + 281 + ], + "type": "inline_equation", + "content": "(h_{j},w_{j})" + }, + { + "bbox": [ + 130, + 194, + 482, + 281 + ], + "type": "text", + "content": " (red dot) and surrounding pixels (green dots). We first heuristically find an initial pixel " + }, + { + "bbox": [ + 130, + 194, + 482, + 281 + ], + "type": "inline_equation", + "content": "(h_j^0,w_j^0)" + }, + { + "bbox": [ + 130, + 194, + 482, + 281 + ], + "type": "text", + "content": " with a fixed surrounding window and then learn coordinate offsets to locate the optimized correspondence and surrounding pixels." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 287, + 293, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 287, + 293, + 301 + ], + "spans": [ + { + "bbox": [ + 132, + 287, + 293, + 301 + ], + "type": "inline_equation", + "content": "\\pmb{F}_s^A \\in \\mathbb{R}^{N \\times H_s \\times W_s \\times C}" + }, + { + "bbox": [ + 132, + 287, + 293, + 301 + ], + "type": "text", + "content": ", formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 238, + 309, + 481, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 309, + 481, + 334 + ], + "spans": [ + { + "bbox": [ + 238, + 309, + 481, + 334 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} _ {s} ^ {A} = \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\boldsymbol {F} _ {s} ^ {A ^ {\\prime}} (:,:,:, j,:). \\tag {9}", + "image_path": "5ec1ca8129bb9e973de545dbecc0807ba5a7d6c4a87b0e45ede885d0b4dfb5f7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 340, + 482, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 340, + 482, + 404 + ], + "spans": [ + { + "bbox": [ + 131, + 340, + 482, + 404 + ], + "type": "text", + "content": "Association-induced Feature Enhancement. Once we have obtained association feature " + }, + { + "bbox": [ + 131, + 340, + 482, + 404 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_s^A" + }, + { + "bbox": [ + 131, + 340, + 482, + 404 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 131, + 340, + 482, + 404 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 131, + 340, + 482, + 404 + ], + "type": "text", + "content": "-th stage, we will use it to enhance the VGG feature of the " + }, + { + "bbox": [ + 131, + 340, + 482, + 404 + ], + "type": "inline_equation", + "content": "(s - 1)" + }, + { + "bbox": [ + 131, + 340, + 482, + 404 + ], + "type": "text", + "content": "-th stage, i.e. " + }, + { + "bbox": [ + 131, + 340, + 482, + 404 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}_{s - 1,l}\\}_{l = 1}^{L_{s - 1}}" + }, + { + "bbox": [ + 131, + 340, + 482, + 404 + ], + "type": "text", + "content": ". Specifically, we upsample " + }, + { + "bbox": [ + 131, + 340, + 482, + 404 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_s^A" + }, + { + "bbox": [ + 131, + 340, + 482, + 404 + ], + "type": "text", + "content": " to align the spatial size of features in the " + }, + { + "bbox": [ + 131, + 340, + 482, + 404 + ], + "type": "inline_equation", + "content": "(s - 1)" + }, + { + "bbox": [ + 131, + 340, + 482, + 404 + ], + "type": "text", + "content": "-th stage, and then add it to " + }, + { + "bbox": [ + 131, + 340, + 482, + 404 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}_{s - 1,l}\\}_{l = 1}^{L_{s - 1}}" + }, + { + "bbox": [ + 131, + 340, + 482, + 404 + ], + "type": "text", + "content": " followed by a convolution layer, formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 238, + 412, + 481, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 412, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 238, + 412, + 481, + 429 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {F}} _ {s - 1, l} = \\mathcal {C} _ {3} \\left(\\boldsymbol {F} _ {s - 1, l} + \\mathcal {U} \\left(\\boldsymbol {F} _ {s} ^ {A}\\right)\\right), \\tag {10}", + "image_path": "6f548da4488209a90ecaf646e59a83ea5ac52417ebd358e90434376f4467edad.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 436, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 461 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 436, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_3" + }, + { + "bbox": [ + 130, + 436, + 482, + 461 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 436, + 482, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 130, + 436, + 482, + 461 + ], + "type": "text", + "content": " represent a 2D convolution layer and the bilinear upsampling operation, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 478, + 418, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 478, + 418, + 490 + ], + "spans": [ + { + "bbox": [ + 131, + 478, + 418, + 490 + ], + "type": "text", + "content": "3.2 Correspondence-induced Association Condensation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 498, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 582 + ], + "type": "text", + "content": "Although PAG based on full-pixel hyperassociations can deliver satisfactory performance in CoSOD, it also introduces substantial computational overhead. Additionally, we argue that for each pixel in an image, it is unnecessary to gather its associations with all pixels of other related images to form hyperassociations. Some pixel associations may even impair the final performance, such as those between ambiguous regions. To this end, this subsection tries to condense the original full-pixel hyperassociations to retain only informative pixel associations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 582, + 482, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 582, + 482, + 630 + ], + "spans": [ + { + "bbox": [ + 131, + 582, + 482, + 630 + ], + "type": "text", + "content": "This subsection will focus on explaining the condensation of pixel associations of a pixel (e.g. " + }, + { + "bbox": [ + 131, + 582, + 482, + 630 + ], + "type": "inline_equation", + "content": "(h_i, w_i)" + }, + { + "bbox": [ + 131, + 582, + 482, + 630 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 131, + 582, + 482, + 630 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 131, + 582, + 482, + 630 + ], + "type": "text", + "content": ") to ones of the other images (e.g. image " + }, + { + "bbox": [ + 131, + 582, + 482, + 630 + ], + "type": "inline_equation", + "content": "I_j" + }, + { + "bbox": [ + 131, + 582, + 482, + 630 + ], + "type": "text", + "content": "), i.e. " + }, + { + "bbox": [ + 131, + 582, + 482, + 630 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_s(i, h_i, w_i, j, :, :, :) \\in \\mathbb{R}^{H_s \\times W_s \\times L_s}" + }, + { + "bbox": [ + 131, + 582, + 482, + 630 + ], + "type": "text", + "content": ", as shown in Figure 3. We will simplify the symbol " + }, + { + "bbox": [ + 131, + 582, + 482, + 630 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_s(i, h_i, w_i, j, :, :, :)" + }, + { + "bbox": [ + 131, + 582, + 482, + 630 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 131, + 582, + 482, + 630 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_s^j" + }, + { + "bbox": [ + 131, + 582, + 482, + 630 + ], + "type": "text", + "content": " in the subsequent text for convenience." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Specifically, CAC opts to select " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "K \\times K" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "K < H_{s}, K < W_{s}" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": ") informative pixel associations from " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{a}_{s}^{j}" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": " to form its condensed representation, i.e. " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{a}}_{s}^{j} \\in \\mathbb{R}^{K \\times K \\times L_{s}}" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": ". Thus, the entire condensed hyperassocaton can be symbolized as " + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{A}_{s} \\in \\mathbb{R}^{N \\times H_{s} \\times W_{s} \\times N \\times L_{s}}" + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 209, + 100 + ], + "type": "text", + "content": "L Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 247 + ], + "type": "inline_equation", + "content": "K \\times K \\times L_s" + }, + { + "bbox": [ + 130, + 115, + 482, + 247 + ], + "type": "text", + "content": ". To ensure the proper selection of " + }, + { + "bbox": [ + 130, + 115, + 482, + 247 + ], + "type": "inline_equation", + "content": "K \\times K" + }, + { + "bbox": [ + 130, + 115, + 482, + 247 + ], + "type": "text", + "content": " pixels, we introduce a pretext task, i.e. semantic correspondence estimation [14,24]. This allows us to first locate the corresponding pixel of " + }, + { + "bbox": [ + 130, + 115, + 482, + 247 + ], + "type": "inline_equation", + "content": "(h_i, w_i)" + }, + { + "bbox": [ + 130, + 115, + 482, + 247 + ], + "type": "text", + "content": " in image " + }, + { + "bbox": [ + 130, + 115, + 482, + 247 + ], + "type": "inline_equation", + "content": "I_j" + }, + { + "bbox": [ + 130, + 115, + 482, + 247 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 130, + 115, + 482, + 247 + ], + "type": "inline_equation", + "content": "(h_j, w_j)" + }, + { + "bbox": [ + 130, + 115, + 482, + 247 + ], + "type": "text", + "content": ", and then combine " + }, + { + "bbox": [ + 130, + 115, + 482, + 247 + ], + "type": "inline_equation", + "content": "(h_j, w_j)" + }, + { + "bbox": [ + 130, + 115, + 482, + 247 + ], + "type": "text", + "content": " with its surrounding pixels as the " + }, + { + "bbox": [ + 130, + 115, + 482, + 247 + ], + "type": "inline_equation", + "content": "K \\times K" + }, + { + "bbox": [ + 130, + 115, + 482, + 247 + ], + "type": "text", + "content": " pixel set. We design this approach based on the observation that the co-salient objects across " + }, + { + "bbox": [ + 130, + 115, + 482, + 247 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 115, + 482, + 247 + ], + "type": "text", + "content": " related images belong to the same semantic category, and the pixels within them should have semantic correspondences to each other, as shown in Figure 4. Therefore, the introduction of semantic correspondence in CAC not only improves the CoSOD performance but also delves deeper into the core nature of the CoSOD task. As far as we know, this is the first work to use semantic correspondence in the CoSOD task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 247, + 482, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 320 + ], + "type": "text", + "content": "Correspondence Estimation. To estimate the correspondence pixel " + }, + { + "bbox": [ + 130, + 247, + 482, + 320 + ], + "type": "inline_equation", + "content": "(h_j, w_j)" + }, + { + "bbox": [ + 130, + 247, + 482, + 320 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 247, + 482, + 320 + ], + "type": "inline_equation", + "content": "I_j" + }, + { + "bbox": [ + 130, + 247, + 482, + 320 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 247, + 482, + 320 + ], + "type": "inline_equation", + "content": "(h_i, w_i)" + }, + { + "bbox": [ + 130, + 247, + 482, + 320 + ], + "type": "text", + "content": ", we first identify an initial pixel " + }, + { + "bbox": [ + 130, + 247, + 482, + 320 + ], + "type": "inline_equation", + "content": "(h_j^0, w_j^0)" + }, + { + "bbox": [ + 130, + 247, + 482, + 320 + ], + "type": "text", + "content": " via a heuristic approach. Subsequently, we produce a spatial offset to refine " + }, + { + "bbox": [ + 130, + 247, + 482, + 320 + ], + "type": "inline_equation", + "content": "(h_j^0, w_j^0)" + }, + { + "bbox": [ + 130, + 247, + 482, + 320 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 247, + 482, + 320 + ], + "type": "inline_equation", + "content": "(h_j, w_j)" + }, + { + "bbox": [ + 130, + 247, + 482, + 320 + ], + "type": "text", + "content": ". To achieve this purpose, all initial correspondence pixels should be utilized to form the initial condensed hyperassociations, with which the initial deep association features can be generated for spatial offset prediction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 320, + 482, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 357 + ], + "type": "text", + "content": "Specifically, we pick out " + }, + { + "bbox": [ + 130, + 320, + 482, + 357 + ], + "type": "inline_equation", + "content": "(h_j^0, w_j^0)" + }, + { + "bbox": [ + 130, + 320, + 482, + 357 + ], + "type": "text", + "content": " that has the largest feature similarity value with " + }, + { + "bbox": [ + 130, + 320, + 482, + 357 + ], + "type": "inline_equation", + "content": "(h_i, w_i)" + }, + { + "bbox": [ + 130, + 320, + 482, + 357 + ], + "type": "text", + "content": ". Since we have calculated the feature similarities in " + }, + { + "bbox": [ + 130, + 320, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\pmb{a}_s^j" + }, + { + "bbox": [ + 130, + 320, + 482, + 357 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 130, + 320, + 482, + 357 + ], + "type": "inline_equation", + "content": "(h_j^0, w_j^0)" + }, + { + "bbox": [ + 130, + 320, + 482, + 357 + ], + "type": "text", + "content": " can be obtained as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 268, + 365, + 361, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 365, + 361, + 386 + ], + "spans": [ + { + "bbox": [ + 268, + 365, + 361, + 386 + ], + "type": "interline_equation", + "content": "\\bar {\\pmb {a}} _ {s} ^ {j} = \\sum_ {l = 1} ^ {L _ {s}} \\pmb {a} _ {s} ^ {j} (:,: l),", + "image_path": "c6614eb15b0bcde60bfc149038abcc3de9d1aa3d8d61be772e7af656bb97ee49.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 245, + 380, + 480, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 380, + 480, + 409 + ], + "spans": [ + { + "bbox": [ + 245, + 380, + 480, + 409 + ], + "type": "interline_equation", + "content": "\\left(h _ {j} ^ {0}, w _ {j} ^ {0}\\right) = \\underset {x, y} {\\operatorname {a r g m a x}} \\left(\\bar {\\boldsymbol {a}} _ {s} ^ {j} (x, y)\\right), \\tag {11}", + "image_path": "4b8fa1ce4087b2df296212d605dbfea03635d1ec1dac37cf96d655e7e8205794.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 418, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 418, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{a}}_s^j\\in \\mathbb{R}^{H_s\\times W_s}" + }, + { + "bbox": [ + 130, + 418, + 482, + 453 + ], + "type": "text", + "content": " is a similarity matrix obtained by eliminating the last dimension of " + }, + { + "bbox": [ + 130, + 418, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\pmb {a}_s^j\\in \\mathbb{R}^{H_s\\times W_s\\times L_s}" + }, + { + "bbox": [ + 130, + 418, + 482, + 453 + ], + "type": "text", + "content": " through summation. argmax returns the coordinate of the maximum value." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 455, + 482, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 482, + 546 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 482, + 546 + ], + "type": "text", + "content": "Next, we select " + }, + { + "bbox": [ + 130, + 455, + 482, + 546 + ], + "type": "inline_equation", + "content": "K \\times K" + }, + { + "bbox": [ + 130, + 455, + 482, + 546 + ], + "type": "text", + "content": " pixels within the square region centered around the initial pixels, e.g. " + }, + { + "bbox": [ + 130, + 455, + 482, + 546 + ], + "type": "inline_equation", + "content": "(h_j^0, w_j^0)" + }, + { + "bbox": [ + 130, + 455, + 482, + 546 + ], + "type": "text", + "content": ", to construct the initial condensed hyperassociation " + }, + { + "bbox": [ + 130, + 455, + 482, + 546 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{A}}_s'" + }, + { + "bbox": [ + 130, + 455, + 482, + 546 + ], + "type": "text", + "content": ". Then, we can feed it into the aggregation network, described in Sec 3.1, and achieve the initial aggregated association features " + }, + { + "bbox": [ + 130, + 455, + 482, + 546 + ], + "type": "inline_equation", + "content": "\\pmb{F}_s^{A'} \\in \\mathbb{R}^{N \\times H_s \\times W_s \\times N \\times C_s}" + }, + { + "bbox": [ + 130, + 455, + 482, + 546 + ], + "type": "text", + "content": ". It can be regarded as the association features of each pixel to " + }, + { + "bbox": [ + 130, + 455, + 482, + 546 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 455, + 482, + 546 + ], + "type": "text", + "content": " other related images. We utilize the feature of " + }, + { + "bbox": [ + 130, + 455, + 482, + 546 + ], + "type": "inline_equation", + "content": "(h_i, w_i)" + }, + { + "bbox": [ + 130, + 455, + 482, + 546 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 455, + 482, + 546 + ], + "type": "inline_equation", + "content": "I_j" + }, + { + "bbox": [ + 130, + 455, + 482, + 546 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 130, + 455, + 482, + 546 + ], + "type": "inline_equation", + "content": "\\pmb{F}_s^{A'}(i, h_i, w_i, j,:) \\in \\mathbb{R}^{C_s}" + }, + { + "bbox": [ + 130, + 455, + 482, + 546 + ], + "type": "text", + "content": ", to predict the offsets for " + }, + { + "bbox": [ + 130, + 455, + 482, + 546 + ], + "type": "inline_equation", + "content": "(h_j, w_j)" + }, + { + "bbox": [ + 130, + 455, + 482, + 546 + ], + "type": "text", + "content": " via a linear layer, formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 247, + 555, + 480, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 555, + 480, + 571 + ], + "spans": [ + { + "bbox": [ + 247, + 555, + 480, + 571 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Delta} _ {s} ^ {j} = \\mathcal {O} \\left(\\boldsymbol {F} _ {s} ^ {A ^ {\\prime}} (i, h _ {i}, w _ {i}, j,:)\\right), \\tag {12}", + "image_path": "77f1321b6e737d08ee061fb54a19e2652862ae0ced7f8fab72cc257e2fd9173b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 580, + 482, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 580, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 580, + 482, + 643 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 580, + 482, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 130, + 580, + 482, + 643 + ], + "type": "text", + "content": " is a linear layer for offset generation. " + }, + { + "bbox": [ + 130, + 580, + 482, + 643 + ], + "type": "inline_equation", + "content": "\\pmb{\\Delta}_{s}^{j} \\in \\mathbb{R}^{K \\times K \\times 2}" + }, + { + "bbox": [ + 130, + 580, + 482, + 643 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 130, + 580, + 482, + 643 + ], + "type": "inline_equation", + "content": "K \\times K" + }, + { + "bbox": [ + 130, + 580, + 482, + 643 + ], + "type": "text", + "content": " offsets, besides the center offset " + }, + { + "bbox": [ + 130, + 580, + 482, + 643 + ], + "type": "inline_equation", + "content": "\\pmb{\\Delta}_{s}^{j}(k_{c}, k_{c},:)" + }, + { + "bbox": [ + 130, + 580, + 482, + 643 + ], + "type": "text", + "content": " for correspondence estimation, i.e. refining " + }, + { + "bbox": [ + 130, + 580, + 482, + 643 + ], + "type": "inline_equation", + "content": "(h_{j}^{0}, w_{j}^{0})" + }, + { + "bbox": [ + 130, + 580, + 482, + 643 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 580, + 482, + 643 + ], + "type": "inline_equation", + "content": "(h_{j}, w_{j})" + }, + { + "bbox": [ + 130, + 580, + 482, + 643 + ], + "type": "text", + "content": ", we also generate other offsets for surrounding pixel selection. Thus, the corresponding pixel " + }, + { + "bbox": [ + 130, + 580, + 482, + 643 + ], + "type": "inline_equation", + "content": "(h_{j}, w_{j})" + }, + { + "bbox": [ + 130, + 580, + 482, + 643 + ], + "type": "text", + "content": " can be obtained by adding the offset to the initial pixel " + }, + { + "bbox": [ + 130, + 580, + 482, + 643 + ], + "type": "inline_equation", + "content": "(h_{j}^{0}, w_{j}^{0})" + }, + { + "bbox": [ + 130, + 580, + 482, + 643 + ], + "type": "text", + "content": ", formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 233, + 652, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 652, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 233, + 652, + 480, + 667 + ], + "type": "interline_equation", + "content": "\\left(h _ {j}, w _ {j}\\right) = \\left(h _ {j} ^ {0}, w _ {j} ^ {0}\\right) + \\boldsymbol {\\Delta} _ {s} ^ {j} \\left(k _ {c}, k _ {c}, :\\right), \\tag {13}", + "image_path": "08e9a35845bd6c31f93e4a4331b37842ff54a1641f3dde12f6d04388cc112e1c.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "CONDA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 365, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 365, + 128 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 365, + 128 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 365, + 128 + ], + "type": "inline_equation", + "content": "(k_c, k_c)" + }, + { + "bbox": [ + 130, + 116, + 365, + 128 + ], + "type": "text", + "content": " is the center position of " + }, + { + "bbox": [ + 130, + 116, + 365, + 128 + ], + "type": "inline_equation", + "content": "K \\times K" + }, + { + "bbox": [ + 130, + 116, + 365, + 128 + ], + "type": "text", + "content": " square." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 128, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 128, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 128, + 482, + 177 + ], + "type": "text", + "content": "Condensation Operation. Given the estimated correspondence pixel " + }, + { + "bbox": [ + 130, + 128, + 482, + 177 + ], + "type": "inline_equation", + "content": "(h_j, w_j)" + }, + { + "bbox": [ + 130, + 128, + 482, + 177 + ], + "type": "text", + "content": " and other offsets in " + }, + { + "bbox": [ + 130, + 128, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\pmb{\\Delta}_{s}^{j}" + }, + { + "bbox": [ + 130, + 128, + 482, + 177 + ], + "type": "text", + "content": ", we can obtain the surrounding pixels and combine them with " + }, + { + "bbox": [ + 130, + 128, + 482, + 177 + ], + "type": "inline_equation", + "content": "(h_j, w_j)" + }, + { + "bbox": [ + 130, + 128, + 482, + 177 + ], + "type": "text", + "content": " to form " + }, + { + "bbox": [ + 130, + 128, + 482, + 177 + ], + "type": "inline_equation", + "content": "K \\times K" + }, + { + "bbox": [ + 130, + 128, + 482, + 177 + ], + "type": "text", + "content": " pixel set. Their coordinates are stored in " + }, + { + "bbox": [ + 130, + 128, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\pmb{n}_s^j \\in \\mathbb{R}^{K \\times K \\times 2}" + }, + { + "bbox": [ + 130, + 128, + 482, + 177 + ], + "type": "text", + "content": ". This process can be formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 157, + 186, + 481, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 186, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 157, + 186, + 481, + 205 + ], + "type": "interline_equation", + "content": "\\boldsymbol {n} _ {s} ^ {j} (x, y,:) = \\left(h _ {j}, w _ {j}\\right) + \\boldsymbol {\\Delta} _ {s} ^ {j} (x, y,:) \\quad x, y \\in \\{1, \\dots , K \\}; x, y \\neq k _ {c}, \\tag {14}", + "image_path": "2f8e648c09dc75128d526d70edec998fd180a088852600006ecea654afe915f4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 157, + 203, + 257, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 203, + 257, + 217 + ], + "spans": [ + { + "bbox": [ + 157, + 203, + 257, + 217 + ], + "type": "interline_equation", + "content": "\\boldsymbol {n} _ {s} ^ {j} (k _ {c}, k _ {c};) = (h _ {j}, w _ {j}),", + "image_path": "e122da2be7f7a2f32fb641cbc7501482fb658257ed820264a919fb98439914f6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 225, + 365, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 365, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 365, + 237 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 225, + 365, + 237 + ], + "type": "inline_equation", + "content": "(k_c, k_c)" + }, + { + "bbox": [ + 130, + 225, + 365, + 237 + ], + "type": "text", + "content": " is the center position of " + }, + { + "bbox": [ + 130, + 225, + 365, + 237 + ], + "type": "inline_equation", + "content": "K \\times K" + }, + { + "bbox": [ + 130, + 225, + 365, + 237 + ], + "type": "text", + "content": " square." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 237, + 480, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 480, + 262 + ], + "type": "text", + "content": "Finally, we can perform the condensation operation via the index selection on " + }, + { + "bbox": [ + 130, + 237, + 480, + 262 + ], + "type": "inline_equation", + "content": "\\pmb{a}_s^j" + }, + { + "bbox": [ + 130, + 237, + 480, + 262 + ], + "type": "text", + "content": ", formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 279, + 261, + 480, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 261, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 279, + 261, + 480, + 275 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {a}} _ {s} ^ {j} = \\boldsymbol {a} _ {s} ^ {j} \\left(\\boldsymbol {n} _ {s} ^ {j}\\right), \\tag {15}", + "image_path": "dbc80f1bdfd9395b3e7c5eec6eb73a0e45eaccece43883e2b82c18b3b9c620af.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 281, + 482, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 342 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 281, + 482, + 342 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{a}}_s^j" + }, + { + "bbox": [ + 130, + 281, + 482, + 342 + ], + "type": "text", + "content": " is the condensed representation of " + }, + { + "bbox": [ + 130, + 281, + 482, + 342 + ], + "type": "inline_equation", + "content": "\\pmb{a}_s^j" + }, + { + "bbox": [ + 130, + 281, + 482, + 342 + ], + "type": "text", + "content": ", i.e. pixel associations of " + }, + { + "bbox": [ + 130, + 281, + 482, + 342 + ], + "type": "inline_equation", + "content": "(h_i, w_i)" + }, + { + "bbox": [ + 130, + 281, + 482, + 342 + ], + "type": "text", + "content": " to image " + }, + { + "bbox": [ + 130, + 281, + 482, + 342 + ], + "type": "inline_equation", + "content": "I_j" + }, + { + "bbox": [ + 130, + 281, + 482, + 342 + ], + "type": "text", + "content": ". Furthermore, we also illustrate the condensation for pixel associations of " + }, + { + "bbox": [ + 130, + 281, + 482, + 342 + ], + "type": "inline_equation", + "content": "(h_i, w_i)" + }, + { + "bbox": [ + 130, + 281, + 482, + 342 + ], + "type": "text", + "content": " to all images, i.e. " + }, + { + "bbox": [ + 130, + 281, + 482, + 342 + ], + "type": "inline_equation", + "content": "\\pmb{a}_s \\in \\mathbb{R}^{N \\times H_s \\times W_s \\times L_s}" + }, + { + "bbox": [ + 130, + 281, + 482, + 342 + ], + "type": "text", + "content": ", in Figure 2. By applying such a condensation process to all pixel associations, we can obtain the final condensed hyperassociation " + }, + { + "bbox": [ + 130, + 281, + 482, + 342 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{A}}_s" + }, + { + "bbox": [ + 130, + 281, + 482, + 342 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 360, + 349, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 360, + 349, + 372 + ], + "spans": [ + { + "bbox": [ + 131, + 360, + 349, + 372 + ], + "type": "text", + "content": "3.3 Object-aware Cycle Consistency Loss" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 380, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 475 + ], + "type": "text", + "content": "To achieve accurate correspondence estimations, applying effective supervisions on them is necessary. As explicit semantic correspondence annotations are not available, we can only rely on unsupervised losses by imposing correspondence-related constraints on estimated correspondences. Previous unsupervised approaches apply constraints to all pixels [30, 35, 48], including those on the background and extraneous objects that don't have mutual correspondences, hence harming the model effectiveness. To avoid this problem, we propose an object-aware constraint to only access losses on the co-salient pixels." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 475, + 482, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 535 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 535 + ], + "type": "text", + "content": "We propose an Object-aware Cycle Consistency (OCC) loss for the supervision of correspondence estimations in CoSOD. The cycle consistency can be explained as: if a co-salient pixel " + }, + { + "bbox": [ + 130, + 475, + 482, + 535 + ], + "type": "inline_equation", + "content": "(h_i, w_i)" + }, + { + "bbox": [ + 130, + 475, + 482, + 535 + ], + "type": "text", + "content": " of image " + }, + { + "bbox": [ + 130, + 475, + 482, + 535 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 130, + 475, + 482, + 535 + ], + "type": "text", + "content": " corresponds to the pixel " + }, + { + "bbox": [ + 130, + 475, + 482, + 535 + ], + "type": "inline_equation", + "content": "(h_j, w_j)" + }, + { + "bbox": [ + 130, + 475, + 482, + 535 + ], + "type": "text", + "content": " in image " + }, + { + "bbox": [ + 130, + 475, + 482, + 535 + ], + "type": "inline_equation", + "content": "I_j" + }, + { + "bbox": [ + 130, + 475, + 482, + 535 + ], + "type": "text", + "content": ", then the pixel " + }, + { + "bbox": [ + 130, + 475, + 482, + 535 + ], + "type": "inline_equation", + "content": "(h_j, w_j)" + }, + { + "bbox": [ + 130, + 475, + 482, + 535 + ], + "type": "text", + "content": " should also semantically correspond to pixel " + }, + { + "bbox": [ + 130, + 475, + 482, + 535 + ], + "type": "inline_equation", + "content": "(h_i, w_i)" + }, + { + "bbox": [ + 130, + 475, + 482, + 535 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "text", + "content": "Based on this cycle consistency constraint, we adopt image warping operations to conduct the OCC loss. Specifically, we first warp the image " + }, + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "inline_equation", + "content": "I_{i}^{s}" + }, + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "text", + "content": " (a resized " + }, + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "text", + "content": " to align the scales in the " + }, + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "text", + "content": "-th stage) as " + }, + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "inline_equation", + "content": "I_{i\\rightarrow j}^{s}" + }, + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "text", + "content": " using the " + }, + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "inline_equation", + "content": "I_{i}^{s}\\to I_{j}^{s}" + }, + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "text", + "content": " correspondence estimations. Next, we warp " + }, + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "inline_equation", + "content": "I_{i\\rightarrow j}^{s}" + }, + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "text", + "content": " backward to " + }, + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "inline_equation", + "content": "I_{i\\rightarrow j\\rightarrow i}^{s}" + }, + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "inline_equation", + "content": "I_{j}^{s}\\to I_{i}^{s}" + }, + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "text", + "content": " correspondence estimations. Finally, we can utilize the SSIM loss between " + }, + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "inline_equation", + "content": "I_{i}^{s}" + }, + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "inline_equation", + "content": "I_{i\\rightarrow j\\rightarrow i}^{s}" + }, + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "text", + "content": " to measure the cycle consistency for mutual correspondence pixels in " + }, + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "inline_equation", + "content": "I_{i}^{s}" + }, + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "inline_equation", + "content": "I_{j}^{s}" + }, + { + "bbox": [ + 130, + 536, + 482, + 631 + ], + "type": "text", + "content": ". Moreover, to ensure the constraints are only conducted on co-salient objects, we mask the images with their ground truth masks, formulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 207, + 635, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 635, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 207, + 635, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} ^ {C} = \\frac {1}{N ^ {2}} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N} \\mathcal {L} _ {S S I M} \\left(I _ {i} ^ {s} \\cdot G _ {i} ^ {s}, I _ {i \\rightarrow j \\rightarrow i} ^ {s} \\cdot G _ {i} ^ {s}\\right), \\tag {16}", + "image_path": "4802e7d00694e1409a1dae660d75485442836d3f4e715d5865e8e77399d08883.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 101 + ], + "type": "text", + "content": "L Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 166, + 138, + 443, + 236 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 1: Ablation Study of our proposed modules. SAG, SAC, and FCC are ablation modules for PAG, CAC, and OCC, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 166, + 138, + 443, + 236 + ], + "lines": [ + { + "bbox": [ + 166, + 138, + 443, + 236 + ], + "spans": [ + { + "bbox": [ + 166, + 138, + 443, + 236 + ], + "type": "table", + "html": "
IDModulesCoCA [47]
SAGPAGSACCACFCCOCCSm↑Eξ↑Fβ↑M↓
10.69360.76420.57290.1206
2✓0.72360.80290.63570.1106
3✓0.73080.81220.64590.1075
4✓✓0.73040.81230.65000.1085
5✓✓0.74730.81550.65910.0956
6✓✓0.73980.81380.65060.1013
7✓✓0.75700.82480.67510.0924
", + "image_path": "0abe3315e7ce9fdde14ed458ce324683c74f248539502589eb2062e08dc79c1a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 246, + 482, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 283 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 246, + 482, + 283 + ], + "type": "inline_equation", + "content": "G_{i}^{s}" + }, + { + "bbox": [ + 130, + 246, + 482, + 283 + ], + "type": "text", + "content": " is the resized ground truth of image " + }, + { + "bbox": [ + 130, + 246, + 482, + 283 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 130, + 246, + 482, + 283 + ], + "type": "text", + "content": ". The total OCC loss " + }, + { + "bbox": [ + 130, + 246, + 482, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^C" + }, + { + "bbox": [ + 130, + 246, + 482, + 283 + ], + "type": "text", + "content": " is the sum of three stages, formulated as: " + }, + { + "bbox": [ + 130, + 246, + 482, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^C = \\mathcal{L}_3^C +\\mathcal{L}_4^C +\\mathcal{L}_5^C" + }, + { + "bbox": [ + 130, + 246, + 482, + 283 + ], + "type": "text", + "content": ". More details can be found in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 300, + 230, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 300, + 230, + 313 + ], + "spans": [ + { + "bbox": [ + 131, + 300, + 230, + 313 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 324, + 326, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 324, + 326, + 336 + ], + "spans": [ + { + "bbox": [ + 131, + 324, + 326, + 336 + ], + "type": "text", + "content": "4.1 Evaluation Datasets and Metrics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 342, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 416 + ], + "type": "text", + "content": "We follow [8] to evaluate our model on three benchmark datasets, i.e. CoCA [47] (1295 images of 80 groups), CoSal2015 [44] (2015 images of 50 groups), and CoSOD3k [7] (3316 images of 160 groups). We adopt four widely-used metrics for the quantitative evaluation, i.e. Structure-measure " + }, + { + "bbox": [ + 130, + 342, + 482, + 416 + ], + "type": "inline_equation", + "content": "(S_{m})" + }, + { + "bbox": [ + 130, + 342, + 482, + 416 + ], + "type": "text", + "content": " [5], Maximum enhanced-alignment measure " + }, + { + "bbox": [ + 130, + 342, + 482, + 416 + ], + "type": "inline_equation", + "content": "(E_{\\xi})" + }, + { + "bbox": [ + 130, + 342, + 482, + 416 + ], + "type": "text", + "content": " [6], Maximum F-measure " + }, + { + "bbox": [ + 130, + 342, + 482, + 416 + ], + "type": "inline_equation", + "content": "(\\mathrm{F}_{\\beta})" + }, + { + "bbox": [ + 130, + 342, + 482, + 416 + ], + "type": "text", + "content": " [1], and Mean Absolute Error " + }, + { + "bbox": [ + 130, + 342, + 482, + 416 + ], + "type": "inline_equation", + "content": "(\\mathcal{M})" + }, + { + "bbox": [ + 130, + 342, + 482, + 416 + ], + "type": "text", + "content": " [3]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 431, + 280, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 431, + 280, + 443 + ], + "spans": [ + { + "bbox": [ + 131, + 431, + 280, + 443 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 450, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 521 + ], + "type": "text", + "content": "To construct the training data, we follow [51] to use different combinations of three commonly used training datasets, i.e. DUTS class [47] (8250 images of 291 groups), COCO-9k [18] (9213 images of 65 groups), and COCO-SEG [37] (200,000 images of 78 groups), for a fair comparison with other state-of-the-art (SOTA) works. We also implement the synthesis strategy for the DUTS class dataset following [46]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 522, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 606 + ], + "type": "text", + "content": "For training specifics, we employ the data augmentation strategy in [21] and use " + }, + { + "bbox": [ + 130, + 522, + 482, + 606 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 522, + 482, + 606 + ], + "type": "text", + "content": " as the input size for the network. We employ the Adam optimizer [15] with " + }, + { + "bbox": [ + 130, + 522, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 130, + 522, + 482, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 522, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.99" + }, + { + "bbox": [ + 130, + 522, + 482, + 606 + ], + "type": "text", + "content": " to optimize the network. We train our CONDA model for 300 epochs, starting with an initial learning rate of " + }, + { + "bbox": [ + 130, + 522, + 482, + 606 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 130, + 522, + 482, + 606 + ], + "type": "text", + "content": ", which is divided by 10 at the " + }, + { + "bbox": [ + 130, + 522, + 482, + 606 + ], + "type": "inline_equation", + "content": "60,000^{th}" + }, + { + "bbox": [ + 130, + 522, + 482, + 606 + ], + "type": "text", + "content": " iteration. Our experiments are implemented based on PyTorch [29] on a single Tesla A40 GPU, with the batchsize set to " + }, + { + "bbox": [ + 130, + 522, + 482, + 606 + ], + "type": "inline_equation", + "content": "N = 6" + }, + { + "bbox": [ + 130, + 522, + 482, + 606 + ], + "type": "text", + "content": ". The hyperparameter " + }, + { + "bbox": [ + 130, + 522, + 482, + 606 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 522, + 482, + 606 + ], + "type": "text", + "content": " in CAC is set to 9." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 622, + 239, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 622, + 239, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 622, + 239, + 635 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "We conduct ablation studies on the most challenging CoCA [47] dataset. To construct the baseline, we use the FPN [17] (with VGG-16 as encoder) as the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "CONDA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 483, + 221 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 483, + 221 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 483, + 221 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 483, + 221 + ], + "type": "image", + "image_path": "40395a7bb7331fffc5891dbbab1881e79c2baa445aa0a5bdba710a534942373c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 224, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 269 + ], + "type": "text", + "content": "Fig. 4: Visual samples for the correspondence estimations. Correspondences I, II, and III visually display estimated correspondences between the main image and three related images. Sparse co-salient pixels were selected and connected to their corresponding pixels using colored lines for clear visualization." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 280, + 482, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 482, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 482, + 339 + ], + "type": "text", + "content": "foundational segmentation network and enhance it with the Region-to-Region correlation module (R2R) [16] to simply capture inter-image connections. Then, as shown in Table 1, we incrementally incorporate PAG, CAC, and OCC into the baseline for effectiveness analysis. We trained all ablation models with the DUTS class and COCO9k datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "text", + "content": "Effectiveness of PAG. PAG first uses intermediate image features from the FPN encoder to calculate full-pixel hyperassociations. Then, aggregation networks are applied to generate deep association features for the decoder process. PAG effectively utilizes deep learning to model pair-wise pixel associations, achieving higher-level association knowledge compared to previous image feature optimization strategies. As shown in the 3rd line of Table 1, PAG shows significant performance boosts compared to the baseline model, with respective gains of " + }, + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "inline_equation", + "content": "3.72\\%" + }, + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "inline_equation", + "content": "4.80\\%" + }, + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "inline_equation", + "content": "7.30\\%" + }, + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "inline_equation", + "content": "1.31\\%" + }, + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "inline_equation", + "content": "S_{m}" + }, + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "inline_equation", + "content": "E_{\\xi}" + }, + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 341, + 482, + 437 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 521 + ], + "type": "text", + "content": "Furthermore, to validate our approach of progressively enhancing image features with the previously generated association feature for improving hyperassociation calculation, we conduct an ablation experiment called Separate Association Generation (SAG) where association features are generated for three stages without image feature enhancements. As shown in the 2nd&3rd lines in Table 1, PAG outperforms SAG, indicating that our progressive enhancement design can obtain better hyperassociation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 522, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 617 + ], + "type": "text", + "content": "Effectiveness of CAC. CAC aims to condense full-pixel hyperassociations in PAG by selecting corresponding pixels and their surrounding contexts. Results in the 3rd&5th lines of Table 1 show that introducing the CAC module improves the performance. Moreover, it reduces the multiply-accumulate operations (MACs) of aggregation networks from 91.38G in the full-pixel PAG to " + }, + { + "bbox": [ + 130, + 522, + 482, + 617 + ], + "type": "inline_equation", + "content": "77.19\\mathrm{G}^{1}" + }, + { + "bbox": [ + 130, + 522, + 482, + 617 + ], + "type": "text", + "content": ". This indicates that utilizing correspondence estimation to condense the hyperassociations not only effectively reduces the computational burden but also helps obtain more accurate pixel associations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 619, + 482, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 619, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 619, + 482, + 643 + ], + "type": "text", + "content": "We also analyze CAC in detail. As deep association features are necessary for reliable correspondence estimation, CAC first pre-condenses the hyperasso-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "L Li et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 141, + 654, + 409, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 409, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 409, + 665 + ], + "type": "text", + "content": "We input a group of 6 related " + }, + { + "bbox": [ + 141, + 654, + 409, + 665 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 141, + 654, + 409, + 665 + ], + "type": "text", + "content": " images to measure MACs." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 149, + 481, + 317 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 148 + ], + "type": "text", + "content": "Table 2: Quantitative comparison of our model with other SOTA methods. DC, C9, and CS are DUTS class, COCO9k, and COCO-SEG training data, respectively. bold and underline mark the best and second-best excellent results, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 149, + 481, + 317 + ], + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 317 + ], + "type": "table", + "html": "
MethodsTrainingCoCA [47]CoSal2015 [44]CoSOD3k [7]
SetSm↑Eξ↑Fβ↑M↓Sm↑Eξ↑Fβ↑M↓Sm↑Eξ↑Fβ↑M↓
GICDDC0.6580.7180.5130.1260.8440.8870.8440.0710.7970.8480.7700.079
GCoNetDC0.6730.7600.5440.1100.8450.8880.8470.0680.8020.8600.7780.071
GCoNet+DC0.6910.7860.5740.1130.8750.9180.8760.0540.8280.8810.8070.068
CONDADC0.7170.7740.6000.1020.8900.9260.8940.0490.8320.8730.8070.067
ICNetC90.6540.7040.5130.1470.8570.9010.8580.0580.7940.8450.7620.089
DCFMC90.7100.7830.5980.0850.8380.8930.8560.0670.8090.8740.8050.067
GCoNet+C90.7170.7980.6050.0980.8530.9020.8570.0730.8190.8770.7960.075
CONDAC90.7300.8010.6220.0920.8650.9100.8750.0590.8250.8770.8100.068
CADCDC+C90.6800.7440.5490.1330.8670.9060.8650.0640.8150.8540.7780.088
DMTDC+C90.7250.8000.6190.1080.8970.9360.9050.0450.8510.8950.8350.063
GCoNet+DC+C90.7340.8080.6260.0880.8760.9200.8800.0570.8390.8940.8220.064
CONDADC+C90.7570.8250.6750.0920.9040.9400.9120.0420.8570.8990.8440.060
UGEMDC+CS0.7260.8080.5990.0960.8850.9350.8820.0510.8530.9110.8290.060
GCoNet+DC+CS0.7380.8140.6370.0810.8810.9260.8910.0550.8430.9010.8340.061
CONDADC+CS0.7630.8390.6850.0890.9000.9380.9080.0450.8620.9110.8530.056
", + "image_path": "adef6c3da6b359b8d6a99e5a95463d00c32fcd0b9610ff3f7214d3e0c830edc4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 325, + 480, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 480, + 420 + ], + "type": "text", + "content": "ciations using a maximum similarity approach to obtain initial deep association features and then performs further condensation based on the correspondences predicted by these initial deep association features. The hyperassociation condensation process with only the pre-condensation operation is called Similarity-induced Association Condensation (SAC). In Table 1, SAC only brings slight performance improvements due to the heuristic nature of the correspondence estimation. Nevertheless, SAC can provide the initial association feature for CAC to predict reliable correspondence." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 421, + 480, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 421, + 480, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 421, + 480, + 517 + ], + "type": "text", + "content": "Effectiveness of OCC. OCC provides self-supervision for CAC to enable more precise correspondence estimation. Results in the 5th&7th lines of Table 1 show OCC further improves the performance by leveraging more precise correspondence estimations to condense hyperassociations effectively. In addition, we conducted an ablation experiment to validate our object-aware design in CAC by replacing OCC with full-pixel cycle consistency (FCC) loss. Comparing the 5th&6th lines of Table 1, FCC leads to a notable performance decrease due to background pixels disrupting the correspondence learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 517, + 480, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 480, + 564 + ], + "type": "text", + "content": "Visualization of Correspondence Estimations. We present some visual samples of correspondence estimations for some co-salient pixels in Figure 4. Our semantic correspondence estimations are meaningful and can effectively depict the common attributes of co-salient objects at the pixel level." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 584, + 379, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 584, + 379, + 596 + ], + "spans": [ + { + "bbox": [ + 132, + 584, + 379, + 596 + ], + "type": "text", + "content": "4.4 Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "content": "We compare our model with eight recent SOTA methods, i.e. GICD [47], ICNet [13], GCoNet [8], CADC [46], DCFM [41], DMT [16], UGEM [38], and GCoNet+ [51]. We directly utilize their officially released saliency maps for comparison. To ensure fairness, we trained our model with different combinations of three training datasets, following [51], to align with other compared methods. We" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "CONDA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 479, + 318 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 479, + 318 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 479, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 479, + 318 + ], + "type": "image", + "image_path": "7d0b384ecf46d41b5b3fd790f7ba1310cf13090bae0c24b23cda3f353d2fa847.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 322, + 476, + 333 + ], + "lines": [ + { + "bbox": [ + 136, + 322, + 476, + 333 + ], + "spans": [ + { + "bbox": [ + 136, + 322, + 476, + 333 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparisons of our model with other SOTA methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 340, + 482, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 425 + ], + "type": "text", + "content": "denote three training datasets, i.e. DUTS class [47], COCO9k [18], and COCOSEG [37], as DC, C9, and CS, respectively, for convenience. Our training sets include DC, C9, DC+C9, DC+CS. As shown in Table 2, we can observe that our model achieves the best performance with each training set in most benchmark datasets. What is even more exciting is that we achieve excellent results in the most challenging CoCA dataset, surpassing the second-best models by large margins, e.g. " + }, + { + "bbox": [ + 130, + 340, + 482, + 425 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 130, + 340, + 482, + 425 + ], + "type": "inline_equation", + "content": "S_{m}" + }, + { + "bbox": [ + 130, + 340, + 482, + 425 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 340, + 482, + 425 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 130, + 340, + 482, + 425 + ], + "type": "inline_equation", + "content": "E_{\\xi}" + }, + { + "bbox": [ + 130, + 340, + 482, + 425 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 340, + 482, + 425 + ], + "type": "inline_equation", + "content": "4.8\\%" + }, + { + "bbox": [ + 130, + 340, + 482, + 425 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_{\\beta}" + }, + { + "bbox": [ + 130, + 340, + 482, + 425 + ], + "type": "text", + "content": " with the DC+CS training set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 426, + 482, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 474 + ], + "type": "text", + "content": "We also present visual comparisons in Figure 5. Our model can accurately detect co-salient objects in complex scenarios, such as irregularly shaped accordions accompanied by extraneous objects (people). However, other models easily fail to accurately segment co-salient objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 494, + 220, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 220, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 220, + 506 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 592 + ], + "type": "text", + "content": "This paper proposes a deep association learning strategy for CoSOD that directly embeds hyperassociations into deep association features. Correspondence estimation is also introduced to condense hyperassociations, enabling the exploration of pixel-level correspondences for CoSOD. We also utilize an object-aware cycle consistency loss to further refine correspondence estimations. Extensive experiments have verified the effectiveness of our method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 492, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 492, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 492, + 665 + ], + "type": "text", + "content": "Acknowledgments. This work was supported in part by Key-Area Research and Development Program of Guangdong Province under Grant 2021B0101200001, the National Natural Science Foundation of China under Grant 62136007, 62036011, U20B2065, 6202781, 62036005, the Key R&D Program of Shaanxi Province under Grant 2021ZDLGY01-08, and the MBZUAI-WIS Joint Program for AI Research under Grants WIS P008 and P009." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "L Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 164 + ], + "type": "text", + "content": "1. Achanta, R., Hemami, S., Estrada, F., Susstrunk, S.: Frequency-tuned salient region detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1597-1604 (2009)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 165, + 480, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 480, + 197 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 480, + 197 + ], + "type": "text", + "content": "2. Chen, W., Xu, H., Zhou, Z., Liu, Y., Sun, B., Kang, W., Xie, X.: Costformer: Cost transformer for cost aggregation in multi-view stereo. arXiv preprint arXiv:2305.10320 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 198, + 480, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 198, + 480, + 231 + ], + "spans": [ + { + "bbox": [ + 138, + 198, + 480, + 231 + ], + "type": "text", + "content": "3. Cheng, M.M., Warrell, J., Lin, W.Y., Zheng, S., Vineet, V., Crook, N.: Efficient salient region detection with soft image abstraction. In: Int. Conf. Comput. Vis. pp. 1529-1536 (2013)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 232, + 480, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 480, + 253 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 480, + 253 + ], + "type": "text", + "content": "4. Cho, S., Hong, S., Jeon, S., Lee, Y., Sohn, K., Kim, S.: Cats: Cost aggregation transformers for visual correspondence 34, 9011-9023 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 254, + 480, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 254, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 138, + 254, + 480, + 275 + ], + "type": "text", + "content": "5. Fan, D.P., Cheng, M.M., Liu, Y., Li, T., Borji, A.: Structure-measure: A new way to evaluate foreground maps. In: Int. Conf. Comput. Vis. pp. 4548-4557 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 276, + 480, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 276, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 276, + 480, + 298 + ], + "type": "text", + "content": "6. Fan, D.P., Gong, C., Cao, Y., Ren, B., Cheng, M.M., Borji, A.: Enhanced-alignment measure for binary foreground map evaluation. In: IJCAI. pp. 698-704 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "type": "text", + "content": "7. Fan, D.P., Li, T., Lin, Z., Ji, G.P., Zhang, D., Cheng, M.M., Fu, H., Shen, J.: Re-thinking co-salient object detection. IEEE Trans. Pattern Anal. Mach. Intell. 44(8), 4339-4354 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 332, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 480, + 365 + ], + "type": "text", + "content": "8. Fan, Q., Fan, D.P., Fu, H., Tang, C.K., Shao, L., Tai, Y.W.: Group collaborative learning for co-salient object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12288-12298 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 365, + 480, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 387 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 387 + ], + "type": "text", + "content": "9. Fu, H., Cao, X., Tu, Z.: Cluster-based co-saliency detection. IEEE Trans. Image Process. 22(10), 3766-3778 (2013)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 388, + 480, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 480, + 420 + ], + "type": "text", + "content": "10. Han, J., Cheng, G., Li, Z., Zhang, D.: A unified metric learning-based framework for co-saliency detection. IEEE Trans. Circuit Syst. Video Technol. 28(10), 2473-2483 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 421, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 421, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 421, + 480, + 453 + ], + "type": "text", + "content": "1. Hong, S., Cho, S., Nam, J., Lin, S., Kim, S.: Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In: Eur. Conf. Comput. Vis. pp. 108-126 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 454, + 480, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 454, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 454, + 480, + 487 + ], + "type": "text", + "content": "2. Hong, S., Nam, J., Cho, S., Hong, S., Jeon, S., Min, D., Kim, S.: Neural matching fields: Implicit representation of matching fields for visual correspondence 35, 13512-13526 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 487, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 480, + 510 + ], + "type": "text", + "content": "3. Jin, W.D., Xu, J., Cheng, M.M., Zhang, Y., Guo, W.: Icnet: Intra-saliency correlation network for co-saliency detection. Adv. Neural Inform. Process. Syst. (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 510, + 480, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 480, + 532 + ], + "type": "text", + "content": "4. Kim, S., Min, J., Cho, M.: Transformers: Match-to-match attention for semantic correspondence. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8697-8707 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 533, + 480, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 480, + 554 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 480, + 554 + ], + "type": "text", + "content": "5. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 555, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 598 + ], + "type": "text", + "content": "6. Li, L., Han, J., Zhang, N., Liu, N., Khan, S., Cholakkal, H., Anwer, R.M., Khan, F.S.: Discriminative co-saliency and background mining transformer for co-salient object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7247-7256 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "type": "text", + "content": "7. Lin, T.Y., Dólar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2117-2125 (2017)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "8. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dóllar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Eur. Conf. Comput. Vis. pp. 740-755 (2014)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "CONDA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "19. Liu, J.J., Hou, Q., Cheng, M.M., Feng, J., Jiang, J.: A simple pooling-based design for real-time salient object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3917-3926 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "text", + "content": "20. Liu, N., Han, J.: Dhsnet: Deep hierarchical saliency network for salient object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 678-686 (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 173, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 173, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 480, + 205 + ], + "type": "text", + "content": "21. Liu, N., Han, J., Yang, M.H.: Picanet: Learning pixel-wise contextual attention for saliency detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3089-3098 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 206, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 480, + 228 + ], + "type": "text", + "content": "22. Liu, N., Luo, Z., Zhang, N., Han, J.: Vst++: Efficient and stronger visual saliency transformer. IEEE Trans. Pattern Anal. Mach. Intell. (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 229, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 480, + 251 + ], + "type": "text", + "content": "23. Liu, N., Zhang, N., Wan, K., Shao, L., Han, J.: Visual saliency transformer. In: Int. Conf. Comput. Vis. pp. 4722-4732 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 251, + 480, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 480, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 480, + 284 + ], + "type": "text", + "content": "24. Liu, Y., Zhu, L., Yamada, M., Yang, Y.: Semantic correspondence as an optimal transport problem. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4463-4472 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 285, + 480, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 306 + ], + "type": "text", + "content": "25. Liu, Y., Liu, N., Yao, X., Han, J.: Intermediate prototype mining transformer for few-shot semantic segmentation 35, 38020-38031 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 307, + 480, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 307, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 480, + 341 + ], + "type": "text", + "content": "26. Luo, Z., Liu, N., Zhao, W., Yang, X., Zhang, D., Fan, D.P., Khan, F., Han, J.: Vscode: General visual salient and camouflaged object detection with 2d prompt learning. In: IEEE Conf. Comput. Vis. Pattern Recog. (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 342, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 363 + ], + "type": "text", + "content": "27. Min, J., Kang, D., Cho, M.: Hypercorrelation squeeze for few-shot segmentation. In: Int. Conf. Comput. Vis. pp. 6941-6952 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 363, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 480, + 396 + ], + "type": "text", + "content": "28. Moon, S., Sohn, S.S., Zhou, H., Yoon, S., Pavlovic, V., Khan, M.H., Kapadia, M.: Hm: Hybrid masking for few-shot segmentation. In: Eur. Conf. Comput. Vis. pp. 506-523. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 396, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 430 + ], + "type": "text", + "content": "29. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library 32 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 430, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 480, + 453 + ], + "type": "text", + "content": "30. Shen, X., Darmon, F., Efros, A.A., Aubry, M.: Ransac-flow: generic two-stage image alignment. In: Eur. Conf. Comput. Vis. pp. 618-637. Springer (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 453, + 480, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 480, + 485 + ], + "type": "text", + "content": "31. Shi, X., Huang, Z., Bian, W., Li, D., Zhang, M., Cheung, K.C., See, S., Qin, H., Dai, J., Li, H.: Videoflow: Exploiting temporal cues for multi-frame optical flow estimation. arXiv preprint arXiv:2303.08340 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 486, + 480, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 480, + 509 + ], + "type": "text", + "content": "32. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: Int. Conf. Learn. Represent. (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 509, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 480, + 542 + ], + "type": "text", + "content": "33. Su, Y., Deng, J., Sun, R., Lin, G., Su, H., Wu, Q.: A unified transformer framework for group-based segmentation: Co-segmentation, co-saliency detection and video salient object detection. IEEE Trans. Multimedia (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 543, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 480, + 576 + ], + "type": "text", + "content": "34. Sun, G., Liu, Y., Tang, H., Chhatkuli, A., Zhang, L., Van Gool, L.: Mining relations among cross-frame affinities for video semantic segmentation. In: Eur. Conf. Comput. Vis. pp. 522-539. Springer (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 577, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 480, + 609 + ], + "type": "text", + "content": "35. Truong, P., Danelljan, M., Yu, F., Van Gool, L.: Warp consistency for unsupervised learning of dense correspondences. In: Int. Conf. Comput. Vis. pp. 10346-10356 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 610, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 480, + 643 + ], + "type": "text", + "content": "36. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Adv. Neural Inform. Process. Syst. 30 (2017)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "37. Wang, C., Zha, Z.J., Liu, D., Xie, H.: Robust deep co-saliency detection with group semantic. In: AAAI. vol. 33, pp. 8917-8924 (2019)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 208, + 100 + ], + "type": "text", + "content": "L Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 578 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "38. Wu, Y., Song, H., Liu, B., Zhang, K., Liu, D.: Co-salient object detection with uncertainty-aware group exchange-masking. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19639-19648 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "text", + "content": "39. Xiong, Z., Li, H., Zhu, X.X.: Doubly deformable aggregation of covariance matrices for few-shot segmentation. In: Eur. Conf. Comput. Vis. pp. 133-150. Springer (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "text", + "content": "40. Xu, G., Wang, X., Ding, X., Yang, X.: Iterative geometry encoding volume for stereo matching. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 21919-21928 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "41. Yu, S., Xiao, J., Zhang, B., Lim, E.G.: Democracy does matter: Comprehensive feature mining for co-salient object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 979-988 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "text", + "content": "42. Zhang, D., Fu, H., Han, J., Borji, A., Li, X.: A review of co-saliency detection algorithms: Fundamentals, applications, and challenges. ACM Trans Intell Syst Technol 9(4), 1-31 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "type": "text", + "content": "43. Zhang, D., Han, J., Han, J., Shao, L.: Cosaliency detection based on intrasaliency prior transfer and deep intersaliency mining. IEEE Trans. Neural Networks Learn. Syst. 27(6), 1163-1176 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "type": "text", + "content": "44. Zhang, D., Han, J., Li, C., Wang, J., Li, X.: Detection of co-salient objects by looking deep and wide. Int. J. Comput. Vis. 120(2), 215-232 (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "45. Zhang, D., Meng, D., Han, J.: Co-saliency detection via a self-paced multiple-instance learning framework. IEEE Trans. Pattern Anal. Mach. Intell. 39(5), 865-878 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 403 + ], + "type": "text", + "content": "46. Zhang, N., Han, J., Liu, N., Shao, L.: Summarize and search: Learning consensus-aware dynamic convolution for co-saliency detection. In: Int. Conf. Comput. Vis. pp. 4167-4176 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "text", + "content": "47. Zhang, Z., Jin, W., Xu, J., Cheng, M.M.: Gradient-induced co-saliency detection. In: Eur. Conf. Comput. Vis. pp. 455-472 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 446 + ], + "type": "text", + "content": "48. Zhao, D., Song, Z., Ji, Z., Zhao, G., Ge, W., Yu, Y.: Multi-scale matching networks for semantic correspondence. In: Int. Conf. Comput. Vis. pp. 3354-3364 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "text", + "content": "49. Zhao, J.X., Liu, J.J., Fan, D.P., Cao, Y., Yang, J., Cheng, M.M.: Egnet: Edge guidance network for salient object detection. In: Int. Conf. Comput. Vis. pp. 8779-8788 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "50. Zhao, W., Zhang, J., Li, L., Barnes, N., Liu, N., Han, J.: Weakly supervised video salient object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 16826-16835 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 511, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 481, + 544 + ], + "type": "text", + "content": "51. Zheng, P., Fu, H., Fan, D.P., Fan, Q., Qin, J., Tai, Y.W., Tang, C.K., Van Gool, L.: Gconet+: A stronger group collaborative co-salient object detector. IEEE Trans. Pattern Anal. Mach. Intell. (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 544, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 578 + ], + "type": "text", + "content": "52. Zhuge, M., Fan, D.P., Liu, N., Zhang, D., Xu, D., Shao, L.: Salient object detection via integrity learning. IEEE Trans. Pattern Anal. Mach. Intell. 45(3), 3738-3752 (2022)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "CONDA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/COSMU_ Complete 3D human shape from monocular unconstrained images/5adea58c-9270-4078-881c-e504ff34f51f_content_list.json b/2024/COSMU_ Complete 3D human shape from monocular unconstrained images/5adea58c-9270-4078-881c-e504ff34f51f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b4f0b5f730fc3106bedde7a971d792adaa87e5cb --- /dev/null +++ b/2024/COSMU_ Complete 3D human shape from monocular unconstrained images/5adea58c-9270-4078-881c-e504ff34f51f_content_list.json @@ -0,0 +1,1505 @@ +[ + { + "type": "text", + "text": "COSMU: Complete 3D human shape from monocular unconstrained images", + "text_level": 1, + "bbox": [ + 142, + 47, + 854, + 107 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Marco Pesavento, Marco Volino, and Adrian Hilton", + "bbox": [ + 200, + 139, + 796, + 159 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Centre for Vision, Speech and Signal Processing (CVSSP), University of Surrey, Guildford, UK {m.pesavento,m.volino,a.hilton} $@$ surrey.ac.uk", + "bbox": [ + 108, + 175, + 891, + 211 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We present a novel framework to reconstruct complete 3D human shapes from a given target image by leveraging monocular unconstrained images. The objective of this work is to reproduce high-quality details in regions of the reconstructed human body that are not visible in the input target. The proposed methodology addresses the limitations of existing approaches for reconstructing 3D human shapes from a single image, which cannot reproduce shape details in occluded body regions. The missing information of the monocular input can be recovered by using multiple views captured from multiple cameras. However, multi-view reconstruction methods necessitate accurately calibrated and registered images, which can be challenging to obtain in real-world scenarios. Given a target RGB image and a collection of multiple uncalibrated and unregistered images of the same individual, acquired using a single camera, we propose a novel framework to generate complete 3D human shapes. We introduce a novel module to generate 2D multi-view normal maps of the person registered with the target input image. The module consists of body part-based reference selection and body part-based registration. The generated 2D normal maps are then processed by a multi-view attention-based neural implicit model that estimates an implicit representation of the 3D shape, ensuring the reproduction of details in both observed and occluded regions. Extensive experiments demonstrate that the proposed approach estimates higher quality details in the non-visible regions of the 3D clothed human shapes compared to related methods, without using parametric models.", + "bbox": [ + 147, + 255, + 849, + 682 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 82, + 713, + 307, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Democratizing the creation of high-quality human avatars is essential to unlock a wide range of immersive technologies including self-representation in the metaverse, virtual conferencing, and virtual try-on. Realistic human avatars are crucial in these applications, making 3D human reconstruction a prominent topic in computer vision and graphics. Research focuses on finding practical solutions to replace existing sophisticated multi-view systems for 3D reconstruction, which are inaccessible to the general community. Single-view 3D human reconstruction methods have been developed to reconstruct 3D human shapes using a simple and cost-effective set-up, such as a single camera.", + "bbox": [ + 79, + 754, + 917, + 926 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent methods [5, 11, 14, 15, 18, 40] use neural implicit models to estimate 3D", + "bbox": [ + 79, + 928, + 917, + 947 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f2389ca12bc43c05236a13c98905f15d86cdd3ea520c80c9206abf9205fb33e2.jpg", + "image_caption": [ + "Fig. 1: COSMU reconstructs complete 3D human shape from a target RGB input by leveraging a set of monocular images collected from various scenarios, including monocular video (a), images with different poses (b) and different clothing (c)." + ], + "image_footnote": [], + "bbox": [ + 101, + 50, + 903, + 295 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "human shapes with high-quality details in visible regions from single RGB images. These works fail to reproduce fine detail in non-visible regions of the input image, resulting in overly smooth reconstruction.", + "bbox": [ + 72, + 404, + 917, + 461 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this problem, solutions extend single-view to multi-view reconstruction, using multiple images of various orientations of the human [7,27,53,59,60]. Multi-view methods represent details across the entire body but require calibrated cameras and registered images, with the subject in the same pose and clothing. Systems to capture such images are inaccessible to the general community in real-world scenarios, significantly limiting reproducibility.", + "bbox": [ + 72, + 463, + 917, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent works [9,16,55,56,58] reconstruct humans from a single target image by leveraging multiple monocular unconstrained images. We define images as unconstrained if camera calibration parameters are not known (uncalibrated) and the images are captured at different time instances resulting in pose and clothing variations without spatial alignment (unregistered). These methods directly fuse features from multiple images, which is a challenging task since diverse views and poses must be fused into a single coherent reconstruction [58]. They also rely on parametric models and cannot use images of the person in different clothing.", + "bbox": [ + 72, + 578, + 917, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose COSMU, a novel framework that reconstructs CComplete 3D human Shape from Monocular Unconstrained images. Our objective is to reconstruct a 3D clothed human shape from a single target image, focusing on reproducing high-quality details in regions of the reconstructed human body that are not visible in the input target. To achieve this, we leverage a collection of monocular unconstrained images of the same individual. COSMU does not rely on parametric models in the reconstruction process as related works.", + "bbox": [ + 72, + 734, + 917, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The multi-view scenario used by the multi-view reconstruction approaches is simulated from a collection of unconstrained images, with the subject in different poses and clothing and without camera calibration or correspondences. We present a novel pipeline that applies an optimization-based algorithm to select,", + "bbox": [ + 72, + 871, + 917, + 947 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 79, + 17, + 94, + 29 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "M. Pesavento et al.", + "bbox": [ + 159, + 16, + 346, + 30 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "from the set of unconstrained images, $N_r$ references for each $N_p$ body part of the target image following an orientation-visibility-similarity criteria. A neural implicit model generates a coarse point cloud $\\zeta_t$ of the target and of each selected reference $\\zeta_{ref}$ . The 3D points of these point clouds are classified into $N_p$ body parts using a novel pixel-aligned neural network. The corresponding reference body part selected for a specific target part is registered to it. To obtain $N_r + 1$ views of the subject, 2D normal maps are generated from the registered point clouds. The multi-view scenario is simulated and the 2D normal maps are processed by an attention-based implicit model that estimates the implicit function $f_f$ of the complete 3D human shape. COSMU generates 2D multi-view normal maps instead of RGB images because the references are not constrained to the same clothing as the target and RGB value estimation of 3D points is avoided. Extensive experiments show COSMU's ability to reconstruct complete 3D human shapes from monocular images taken from various scenarios, including monocular videos (Fig. 1a), images with different poses and same clothing (Fig. 1b) and images with different poses and clothing (Fig. 1c). Thanks to the flexibility in creating the reference set, COSMU represents a significant advance for high-quality 3D avatar reconstruction.", + "bbox": [ + 72, + 55, + 922, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The contributions presented in this paper include:", + "bbox": [ + 75, + 406, + 610, + 424 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A novel framework to reconstruct complete 3D human shapes from a target RGB image by leveraging a set of monocular uncalibrated and unregistered images without the use of parametric models.", + "- A novel pipeline to simulate a multi-view scenario from a collection of unconstrained reference images via reference selection and body part registration.", + "- Improved performance for human shape reconstruction compared to state-of-the-art methods with higher quality details in the regions of the reconstructed human shape that are occluded in the target image." + ], + "bbox": [ + 89, + 435, + 917, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 75, + 617, + 343, + 639 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D human reconstruction from monocular cameras Early works on 3D human reconstruction from monocular cameras represent the body shape without capturing shape details [1,3,4,6,21,22,35,47,51,57]. The introduction of neural implicit models [10,33,34] revolutionized 3D modeling by estimating an implicit representation, which reproduces higher-quality shapes due to its inherent property of defining a surface as a level set of an occupancy probability function. PiFU [40] introduced pixel-aligned implicit function for human reconstruction, aligning 3D points with features of the input image to achieve an unprecedented level of detail in the reconstructed shapes. Since then, neural implicit models have become the standard representation for 3D human shapes [5,11,14,18,28]. A significant limitation of these approaches is their inability to reproduce high-quality details in non-visible regions. SuRS [37] enhances shape quality by learning a map from a low to a high-resolution shape, hallucinating details in occluded regions, which are still overly smoothed as no information is provided to the implicit model. To improve the quality of non-visible regions, PiFU-HD [41]", + "bbox": [ + 72, + 656, + 922, + 949 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "COSMU: Complete 3D human shape from monocular unconstrained images", + "bbox": [ + 92, + 14, + 837, + 34 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 903, + 16, + 920, + 30 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "generates front and back surface normals of the human from the RGB image. This estimation introduces noise, resulting in smoothed surfaces of occluded regions. ICON [50] improves normal estimation by guiding it with SMPL [30] fitted to the image. Similarly, ECON [49] alleviates the SMPL problem of clothing reconstruction by feeding the estimated front and back normals into a d-BiNI optimizer. Conditioning the learning process on SMPL body poses a challenge for the network, which cannot fully exploit 2D image features. These methods estimate only an additional view, which is insufficient to represent the entire body and its accuracy influences the reconstruction quality of the non-visible regions. TeCH [17] proposes to estimate the unseen view from the input through Score Distillation Sampling [39] (SDS) and represents the human with drivable tetrahedral representation (DMTet [43]). Subject-tailored fine-tuning is required, limiting its generalizability. Other works focus only on 3D reconstruction from monocular videos, requiring to optimize their algorithms on a single monocular video, limiting the generalization and their application scenarios [1-3,20]. NERF-based methods are applied to 3D human reconstruction from sparse views, but their subject-specific nature limits their utility [8,13,19,23,36,46,48]. For a fair comparison with the works that require subject-specific optimization, our approach should be fine-tuned with the data of each specific subject, significantly limiting generalizability. Therefore, these works are not evaluated in this paper.", + "bbox": [ + 75, + 56, + 920, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D human shape from unconstrained RGB images", + "text_level": 1, + "bbox": [ + 77, + 445, + 692, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "PaMIR [56] extended single-view reconstruction to multiple unconstrained images by using the SMPL model as a body reference to establish correspondences across different images. Latent embedding vectors from the correspondent frames are aggregated into a single embedding using mean pooling and then processed by a multi-layer perception (MLP) to estimate the implicit representation. Similarly, MVP-Human [58] fits the SMPL model to each frame to establish correspondences between unconstrained images. Skinning weights of a canonical pose are estimated to guide the fusion of the features from multiple images through an attention-based implicit model. UnconFuse [16] builds upon MVP-Human, with a focus on estimating T-pose shapes from unconstrained images, while posed shapes cannot be reconstructed. CrossHuman [9] selects references following PoseFusion [25] and leveraging the SMPL model. Features are then extracted from both the fitted SMPL and RGB input and fused with a transformer to estimate the implicit representation. These works extract features from unconstrained images and fuse them with a neural model, losing details in regions where the body poses differ among frames. In contrast, COSMU generates 2D multi-view normals of the human in a spatially aligned pose, facilitating the estimation of the final implicit function. Moreover, these works rely on parametric models, whose fitting can introduce noise. COSMU distinguishes itself by not relying on parametric models.", + "bbox": [ + 75, + 465, + 920, + 850 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Have-Fun [52] estimates a DMTet [43] representation by integrating SDS [39] and SMPL skinning weights to adapt the unconstrained images to the target one. Similarly to Tech [17], fine-tuning on the target subject is required, limiting generalizability. Other works focus on 3D human reconstruction from other", + "bbox": [ + 75, + 852, + 917, + 933 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 79, + 17, + 94, + 29 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "M. Pesavento et al.", + "bbox": [ + 159, + 16, + 346, + 30 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/69f9791d2a5e6317f14aa40b8878a9a84498d417dc26ec2042289c31f701062a.jpg", + "image_caption": [ + "Fig. 2: COSMU framework: given a single RGB image as input and a collection of monocular unconstrained images, a multi-view scenario simulation module, which consists of part-based reference selection and registration, is first applied to generate registered multi-view 2D normal maps of the subject in the same pose. An implicit function $\\hat{s}_f$ is then estimated by processing these normal maps with a multi-view attention-based neural implicit model. The 3D mesh is reconstructed via marching cube." + ], + "image_footnote": [], + "bbox": [ + 79, + 53, + 915, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "unconstrained data such as depth images [25, 38] or 3D meshes [26], which is beyond the scope of this paper.", + "bbox": [ + 72, + 453, + 922, + 495 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 3D human shape reconstruction from monocular unconstrained images", + "text_level": 1, + "bbox": [ + 75, + 518, + 832, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce a novel framework that reconstructs COcomplete 3D human Shape from Monocular Unconstrained images (COSMU). To be complete, a 3D human shape should contain high-quality details in every body region. A single image lacks the necessary information to reproduce details across the entire body, which can instead be extracted from multi-view images. In contrast to multiview reconstruction, which requires calibrated and registered images captured with inaccessible systems, COSMU, given a single target image, reproduces fine details in the regions of the 3D shape that are occluded in the input image. This is achieved by leveraging a collection of uncalibrated and unregistered monocular images, referred to as unconstrained images. The input is a target image $I_{t}$ of a person together with multiple views of the same person in various poses and either same or different clothing, from which we extract body part detail non-visible in the target view. As shown in Fig. 2, given a set $\\mathbf{U} = \\{I_u^m\\}_{m=1}^{N_u}$ of $N_{u}$ monocular unconstrained images, COSMU simulates the multi-view scenario of traditional multi-view works using a novel module that employs body part-based reference selection and registration to create 2D multi-view normal maps registered with the target input. A multi-view attention-based implicit model regresses the implicit representation $f_{f}$ of the complete 3D shape from the normals.", + "bbox": [ + 72, + 578, + 924, + 947 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "COSMU: Complete 3D human shape from monocular unconstrained images", + "bbox": [ + 92, + 14, + 840, + 34 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 903, + 16, + 917, + 30 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Multi-view scenario simulation module", + "text_level": 1, + "bbox": [ + 77, + 53, + 617, + 74 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Directly processing unconstrained images with a multi-view neural model deteriorates the performance due to their unregistered nature (Sec. 4.1). To address this, we propose a novel pipeline that simulates a multi-view scenario in which the multi-view images represent various orientations of the subject in the same body pose. Given a target image $I_{t}$ , we select $N_{p} \\times N_{r}$ references $\\left\\{\\{I_{ref}^{(p,r)}\\}_{p=1}^{N_{p}}\\right\\}_{r=1}^{N_{r}}$ , $I_{ref} \\in \\mathbf{U}$ , from the collection $\\mathbf{U}$ of unconstrained images applying a body part-based reference selection. 2D registered multi-view normal maps are then created by registering references to the target input via body part-based registration. Set notation is dropped for simplicity.", + "bbox": [ + 77, + 87, + 920, + 276 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Body part-based reference selection: To improve the body part-based registration, we select references observed from complementary orientations to the target one and with body part poses similar to the corresponding target body part ones. $N_{r}$ references are selected for each $N_{p}$ body part of the target $I_{t}$ via energy optimization following an orientation-visibility-similarity criteria. Semantic masks $M$ and 3D poses $P$ are extracted from the input images. To ensure that the entire body is represented in the selected references, we first divide the $N_{u}$ unconstrained images into $N_{r}$ groups based on the full-body orientation angle $\\theta_{u}$ computed between the reference $P_{u}$ and the target $P_{t}$ 3D poses with the Rodrigues rotation formula. Each group contains a variable number $N_{g}$ of unconstrained images $I_{u}^{g}, g \\in [1,\\dots,N_{g}]$ . $N_{p}$ references $I_{ref}$ are selected from each group following 3 criteria:", + "bbox": [ + 77, + 278, + 920, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1. Body orientation angle: for a complete view of the body, the selected images should complement each other in the body orientation. An orientation metric is introduced to enforce the selection of images in which the body orientation is closer to the central angle $\\theta_c^r$ of the $r^{th}$ group:", + "bbox": [ + 77, + 510, + 920, + 588 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\no _ {g} = \\left| \\theta_ {c} ^ {r} - \\theta_ {u} \\right|, \\quad \\theta_ {c} ^ {r} = r \\cdot \\frac {3 6 0}{N _ {r} + 1}, \\quad r \\in [ 1, \\dots , N _ {r} ] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 239, + 604, + 915, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2. Body part visibility: every body part should be visible in the selected references to have a complete view of the body. We use a visibility metric that counts the total number of pixels $N_{M}$ of each body part $p$ in the semantic masks $M$ .", + "bbox": [ + 77, + 666, + 917, + 725 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nv _ {g} ^ {p} = N _ {M} ^ {p}, \\quad p \\in [ 1, \\dots , N _ {p} ] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 363, + 743, + 915, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3. Body part pose similarity: similarity between each body part of $I_{u}^{g}$ and the corresponding body part of $I_{t}$ is evaluated as 3D Euclidean distance between the corresponding 3D bones of $P_{g}$ and $P_{t}$ , ensuring spatial consistency:", + "bbox": [ + 77, + 795, + 917, + 855 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nb _ {g} ^ {p} = \\left\\| P _ {t} ^ {p} - P _ {g} ^ {p} \\right\\|, \\quad p \\in [ 1, \\dots , N _ {p} ] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 873, + 915, + 892 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$o_{g} = o(I_{u}^{g}), v_{g}^{p} = v^{p}(I_{u}^{g}), b_{g}^{p} = b^{p}(I_{u}^{g})$ for the $g^{th}$ image $I_{u}^{g}$ of the $p^{th}$ body part. $N_{p}$ references $I_{ref}$ are selected by minimizing the energy $E_{g}^{p}$ over the $N_{g}$ images", + "bbox": [ + 77, + 907, + 917, + 949 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 77, + 16, + 94, + 30 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "M. Pesavento et al.", + "bbox": [ + 157, + 14, + 348, + 30 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$I_u^g$ for each body part $p$ :", + "bbox": [ + 77, + 55, + 336, + 76 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nI _ {r e f} = \\underset {g} {\\operatorname {a r g m i n}} \\left(E _ {g} ^ {p}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 91, + 917, + 128 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nE _ {g} ^ {p} = i d x _ {m i n} \\left(o _ {g}\\right) + i d x _ {m a x} \\left(v _ {g} ^ {p}\\right) + i d x _ {m i n} \\left(b _ {g} ^ {p}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 128, + 769, + 152 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$idx_{min}$ indicates the position $idx$ of the $g^{th}$ image $I_u^g$ in the arrays created by sorting all the $N_g$ images with respect to their specific metric $K = o_g, v_g^p, b_g^p$ . $idx_{max}$ sorts the array in inverse order:", + "bbox": [ + 75, + 167, + 917, + 226 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\ni d x _ {m i n} (K) = \\mathrm {i d x} _ {I _ {u} ^ {g}} \\left\\{I _ {u} ^ {1}, \\dots , I _ {u} ^ {N _ {g}} \\right\\}, K \\left(I _ {u} ^ {1}\\right) \\leqslant \\dots \\leqslant K \\left(I _ {u} ^ {N _ {g}}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 240, + 912, + 276 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\ni d x _ {m a x} (K) = \\mathrm {i d x} _ {I _ {u} ^ {g}} \\left\\{I _ {u} ^ {1}, \\dots , I _ {u} ^ {N _ {g}} \\right\\}, K \\left(I _ {u} ^ {1}\\right) \\geqslant \\dots \\geqslant K \\left(I _ {u} ^ {N _ {g}}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 271, + 811, + 295 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$N_{p} \\times N_{r}$ references are selected by repeating the body part-based reference selection for every group.", + "bbox": [ + 75, + 308, + 917, + 346 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Body part-based registration: To simulate the multi-view scenario given the selected $N_{p} \\times N_{r}$ references, we register each 3D reference body part to the corresponding 3D body part in $I_{t}$ . This process begins with the reconstruction of a coarse point cloud $\\zeta_{t}$ for $I_{t}$ and for each of the $N_{p}$ reference images $I_{ref}$ previously selected for the $N_{r}$ groups $(\\zeta_{ref})$ . We use a neural implicit model $f_{c}$ comprising of a U-Net that increases the resolution of the input features $\\phi(I)$ as in SuRS [37]. The coarse implicit function is estimated as:", + "bbox": [ + 72, + 347, + 920, + 484 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {s} _ {c} = f _ {c} \\left(\\phi (\\pi (I)), z (x)\\right), \\quad \\hat {s} _ {c} \\in \\mathbb {R} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 499, + 917, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $I$ is either $I_{t}$ or $I_{ref}$ , $\\pi$ represents the orthographic projection of 3D points $x$ into the input images $I$ and $z(x)$ is the depth of $x$ . $f_{c}$ is trained with:", + "bbox": [ + 75, + 543, + 917, + 583 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left| f _ {c} \\left(X _ {c} ^ {i}\\right) - f _ {c} ^ {g t} \\left(X _ {c} ^ {i}\\right) \\right| ^ {2} + | I - I _ {H R} | \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 263, + 598, + 917, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $X_{c}^{i} = (\\phi^{i}(\\pi (I)),z(x))$ , $N$ is the number of points in space and $I_{HR}$ is the up-sampled ground-truth version of $I$ used to train the U-Net feature extractor. We then retrieve 3D body parts by classifying each 3D point $x_{\\zeta}$ of the point clouds $\\zeta_{t / ref}$ with a novel pixel-aligned classification network $f_{\\zeta}$ . An image feature embedding $\\phi (I||M)$ is extracted from the concatenation between the input RGB image and its semantic mask $M$ . Each $x_{\\zeta}$ is projected using orthographic projection $\\pi$ onto the input $I||M$ and indexed to $\\phi (I||M)$ . An MLP processes this embedding to estimate the body part $\\hat{s}_{\\zeta}$ to which $x_{\\zeta}$ belongs:", + "bbox": [ + 72, + 663, + 920, + 821 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {s} _ {\\zeta} = f _ {\\zeta} \\left(\\phi \\left(\\pi (I | | M)\\right), z \\left(x _ {\\zeta}\\right)\\right), \\quad \\hat {s} _ {\\zeta} \\in [ 1, \\dots , N _ {p} ] \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 263, + 837, + 917, + 856 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$N_{p}$ 3D body parts $v_{ref}$ are defined for each reference point cloud as well as for the target $v_{t}$ . Next, the specific $p^{th}$ body part $v_{ref}$ selected for the corresponding $p^{th}$ target body part $v_{t}$ is registered with it as $v_{reg} = R(v_{ref}) + T$ , where $R$ is the rotation and $T$ the translation computed via rigid registration. For each", + "bbox": [ + 72, + 869, + 920, + 949 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "COSMU: Complete 3D human shape from monocular unconstrained images", + "bbox": [ + 92, + 14, + 837, + 32 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 903, + 16, + 917, + 30 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "group, the registered $N_{p}$ body parts are combined to create a point cloud $\\zeta_{f}$ . $N_{r}$ point clouds $\\zeta_{f}$ registered with $\\zeta_{t}$ are generated, rotated by their $\\theta_{u}$ and, together with $\\zeta_{t}$ , used to retrieve $N_{r} + 1$ 2D normal maps $S_{v}$ , one for each view $v \\in [1, \\dots, N_{r} + 1]$ . Normal maps are generated instead of RGB images to avoid the estimation of color values for each 3D point of $\\zeta_{t / ref}$ and to allow different clothing in the references. Directly fusing the registered point clouds to obtain the 3D shape is challenging since correspondences are unknown. We train $f_{\\zeta}$ with a Mean Squared Error $MSE$ loss:", + "bbox": [ + 75, + 56, + 922, + 211 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p a r t}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left| f _ {\\zeta} \\left(X _ {\\zeta} ^ {i}\\right) - f _ {\\zeta} ^ {g t} \\left(X _ {\\zeta} ^ {i}\\right) \\right| ^ {2} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 314, + 229, + 917, + 279 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $X_{\\zeta}^{i} = \\left(\\phi^{i}(\\pi (I||M)),z(x_{\\zeta})\\right)$", + "bbox": [ + 75, + 294, + 435, + 326 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.2 Multi-view implicit representation", + "text_level": 1, + "bbox": [ + 75, + 354, + 561, + 373 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Neural implicit models have proven to be effective in reproducing high-quality details in reconstructed 3D shapes. This is attributed to the characteristic of the implicit function that does not discretize the 3D space like other representations [40]. Leveraging multiple views to learn the implicit representation further enhances the quality of the reconstruction, as it enables the incorporation of information about the entire body. Concatenating features from multiple images is a suboptimal approach because it imposes an order on the views [59]. Fusion approaches based on statistics have been proposed [12,44] but their pooling operation causes a loss of information regarding the individual view contributions. Inspired by Zins et al. [59], we propose to fuse the features extracted from the multi-view 2D normals using an attention-based architecture. An attention score is computed for each view $v$ based on the compatibility of a query with a corresponding key, resulting in the generation of $N_r + 1$ features. Each feature contains the original information from the corresponding view, merged with the information from all the other views. An MLP then models a pixel-aligned implicit function $f_f$ to learn the occupancy value $\\hat{s}_f$ of the 3D points in space:", + "bbox": [ + 72, + 388, + 922, + 704 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {s} _ {f} = f _ {f} \\left(A \\left(\\phi_ {v} \\left(\\pi \\left(S _ {v}\\right)\\right), z (x)\\right)\\right), \\quad \\hat {s} _ {f} \\in \\mathbb {R} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 290, + 721, + 915, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\phi_v$ are the feature extracted from each view $v$ of the surface normal $S_v$ , $\\pi$ indicates the pixel-alignment between the 3D points of the space and each $\\phi_v$ , $A$ is the attention operation and $z(x)$ is the depth value of the 3D points $x$ . We train the multi-view architecture with an MSE loss:", + "bbox": [ + 75, + 749, + 917, + 827 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left| f _ {f} \\left(X _ {f} ^ {i}\\right) - f _ {f} ^ {g t} \\left(X _ {f} ^ {i}\\right) \\right| ^ {2} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 324, + 842, + 917, + 891 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $X_{f}^{i} = A\\left(\\phi_{v}^{i}(\\pi (S_{v})),z(x^{i})\\right)$", + "bbox": [ + 75, + 907, + 435, + 928 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inference: $N_{p} \\times N_{r}$ references are selected from the collection of unconstrained", + "bbox": [ + 75, + 928, + 917, + 949 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 77, + 16, + 94, + 29 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "M. Pesavento et al.", + "bbox": [ + 157, + 14, + 348, + 30 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f13704363cab77fc46968819c620237595c0bb84d3a5147d1bba4f1902bc1eaa.jpg", + "image_caption": [ + "Fig. 3: Qualitative comparisons between 3D shapes from THuman3.0 [45] reconstructed by changing the energy optimization process to select reference images. Front, back and side views of the shapes are illustrated." + ], + "image_footnote": [], + "bbox": [ + 96, + 50, + 895, + 240 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "images. $N_r + 1$ 2D normal maps are generated using body part-based registration to simulate a multi-view scenario with registered views of the subject. The normals are processed by the multi-view attention-based model to estimate the implicit representation. The 3D mesh is reconstructed via marching cube [32]. To achieve pixel-alignment through projection, multi-view works require the parameters of calibrated cameras in specialized capture systems. In COSMU, instead of using the camera projection matrix for pixel-alignment, 3D points of the space are rotated by $\\theta_u$ computed by the multi-view scenario simulation module and projected onto the images using an identity matrix. Existing approaches create the inference 3D grid centered at the origin. If the subject is not centered in the input images, the multi-view pixel-alignment of COSMU is inaccurate since the same rotated 3D point may not be projected in the corresponding pixel in each view. To tackle this, we set the boundary limits of the 3D grid as the boundaries of the coarse point cloud $\\zeta_t$ , ensuring multi-view pixel-alignment.", + "bbox": [ + 72, + 347, + 922, + 621 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 75, + 652, + 309, + 673 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conduct a comprehensive evaluation of COSMU for the task of 3D human shape reconstruction from a target image by leveraging a collection of monocular unconstrained images of the same individual. The neural models of the framework are trained with THuman2.0 dataset [54], which comprises 524 high-resolution surfaces. We render 360-degree view images around each surface, which serves as input data. For the quantitative and qualitative evaluation, we use the THuman3.0 dataset [45], which consists of 20 individuals each captured in various poses (400 models), and the testing set of X-Humans [42] (41 models). For each pose, we render 360-degree views of the subject and we apply the reference selection to select the $N_{p} \\times N_{r}$ references among the images of the subject in different poses, excluding the target one. We evaluate the reconstruction accuracy with three quantitative metrics: the normal reprojection error [40], the average point-to-surface Euclidean distance (P2S) and the Chamfer distance", + "bbox": [ + 72, + 695, + 924, + 949 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "COSMU: Complete 3D human shape from monocular unconstrained images", + "bbox": [ + 92, + 14, + 840, + 34 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 900, + 16, + 917, + 30 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "(CD), expressed in cm. We further test COSMU with real data collected from various scenarios, including monocular videos (People-Snapshot [3], iPER [29] datasets) or images of people in different poses and clothing from the Internet, demonstrating its superiority over related works to reconstruct complete 3D human shapes with higher quality details in regions of the body occluded in the input target image (Sec. 4.2). In contrast to prior works, no parametric models are used. We set $N_{p} = 10$ and $N_{r} = 3$ . See supplementary for additional implementation details.", + "bbox": [ + 72, + 55, + 920, + 211 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Ablation Studies", + "text_level": 1, + "bbox": [ + 75, + 235, + 348, + 253 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Energy function for reference selection: We conduct an ablation study on the criteria used for selecting the $N_{p} \\times N_{r}$ references in the body part-based reference selection. The variants explored are: (i) Random: no criteria are applied. $N_{p}$ references are randomly selected for each group; (ii) Only $\\mathbf{b_p}$ : only the body part pose similarity is considered; (iii) Only $\\mathbf{o}$ : the body orientation angle is the only criterion used; (iv) Only $\\mathbf{v_p}$ : only the body part visibility is considered; (v) $\\mathbf{o} + \\mathbf{v_p}$ : the body part pose similarity is excluded; (vi) $\\mathbf{o} + \\mathbf{b_p}$ : the body part visibility is omitted; (vii) $\\mathbf{v_p} + \\mathbf{b_p}$ : the orientation angle is not used.", + "bbox": [ + 72, + 263, + 920, + 419 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As illustrated in Fig. 3 and supported by the quantitative results of Tab. 1, the 3D shapes obtained with COSMU exhibit fewer artifacts and finer details. The reference selection plays a crucial role in selecting the most suitable references to facilitate the body part-based registration. Omitting any of the criteria in the energy optimization influences the selection of references, leading to suboptimal choices and, consequently, deteriorating the body part-based registration accuracy and introducing artifacts in the reconstruction.", + "bbox": [ + 72, + 419, + 920, + 554 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/5176f2b39c0bdf4b254fb0e95d996e7bfd2fe89edd45b83c7535059b290c6217.jpg", + "table_caption": [ + "Table 1: Quantitative evaluation of the application of different energy optimization processes to select reference images from THuman3.0 [45]." + ], + "table_footnote": [], + "table_body": "
RandomOnly bpOnly oOnly vpo + vpo + bpvp + bpCOSMU
CD1.4851.4411.4211.4101.3461.3521.3261.304
Normal.1683.1598.1641.1654.1523.1529.1503.1475
P2S1.4241.4151.4031.4061.3321.2791.3081.279
", + "bbox": [ + 118, + 637, + 874, + 712 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Body part-based registration effectiveness: Directly using $N_{r}$ selected references for the multi-view reconstruction deteriorates the performance since fusing different poses is challenging in the absence of known correspondences. To avoid this, we introduce a novel body part-based registration method to simulate a multi-view scenario. To prove its effectiveness, we modify the algorithm for registering the $N_{p} \\times N_{r}$ coarse reference point clouds $\\zeta_{ref}$ to the target $\\zeta_{t}$ :", + "bbox": [ + 72, + 744, + 917, + 863 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "- No Reg: To demonstrate the effectiveness of registering the reference point clouds to the target, we omit the body part-based registration step. We adjust the reference selection to select only $N_r$ references by averaging over $p$ the criteria $v_g^p$ and $b_g^p$ in Eq. (4).", + "bbox": [ + 87, + 869, + 920, + 952 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 79, + 16, + 104, + 29 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "M. Pesavento et al.", + "bbox": [ + 157, + 14, + 348, + 30 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/56d50d131e0eee1d1dffdf60e5ef248fd0cae2d45bbadce8c7e10aeba2761b42.jpg", + "table_caption": [ + "Table 2: Quantitative evaluation of the application of different configurations of the body part-based registration algorithm on THuman3.0 [45]." + ], + "table_footnote": [], + "table_body": "
No RegNon rigidAverageNo RGBNo SegmCOSMU
CD3.5466.9321.3142.2492.1891.304
Normal.2280.4832.1569.2351.2114.1475
P2S2.8206.0241.3092.1692.0051.279
", + "bbox": [ + 159, + 105, + 828, + 175 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d902fa8dc4d22485f0c1bc63bfd67d37853198d640d6d6361026aa401cc9753a.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparisons between 3D shapes from THuman3.0 [45] reconstructed from 2D normal maps generated using different configurations of the body part-based registration algorithm. Front, back and side views of the shapes are shown." + ], + "image_footnote": [], + "bbox": [ + 123, + 193, + 876, + 432 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Non Rigid: Instead of applying rigid registration among the correspondent body parts of the target and reference point clouds, we apply the state-of-the-art work by Li and Harada [24] to perform non-rigid registration between the full-body point clouds. We use $N_r$ references as in No reg.", + "- Average: COSMU registers a target body part with the corresponding body part of a reference selected specifically for that body part. In this configuration, we select only $N_r$ references as in No reg and register all the body parts of the single reference for each group.", + "- No RGB: The body part classification model $f_{\\zeta}$ processes a concatenation of 2D RGB images and segmentation masks. To demonstrate the advantages of using these data, we train and test $f_{\\zeta}$ with only segmentation masks.", + "- No Segm: we train and test $f_{\\zeta}$ with only RGB images." + ], + "bbox": [ + 89, + 543, + 917, + 777 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We apply these configurations to generate 2D multi-view normal maps used by the multi-view attention-based implicit model to reconstruct 3D human shapes. Quantitative (Tab. 2) and qualitative results (Fig. 4) show that COSMU configuration is the most effective. Multi-view reconstruction requires registered input since pixel-alignment across different views can be incorrect with unregistered views. Estimating correspondences between point clouds is challenging and the non-rigid registration cannot produce reliable results. Selecting $N_{p}$ references for each body part facilitates the body part registration compared to using a sin", + "bbox": [ + 72, + 793, + 920, + 949 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "COSMU: Complete 3D human shape from monocular unconstrained images", + "bbox": [ + 92, + 14, + 840, + 34 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 893, + 16, + 915, + 30 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/597d8d14e54ca9470031e308290ea8884d3e79abc01f3963785a3df62071aa1e.jpg", + "table_caption": [ + "Table 3: Quantitative comparisons with related works for 3D human reconstruction." + ], + "table_footnote": [], + "table_body": "
MethodsTHuman3.0X-Humans
CDNormalP2SCDNormalP2S
1 viewPiFU [40]1.698.17851.6592.516.25912.492
SuRS [37]1.573.16931.5542.406.24872.385
Baseline1.566.17321.5702.375.24862.351
2 viewsPiFu-HD [41]1.501.16631.5122.139.22762.121
ICON [50]1.474.16951.4632.210.23952.208
ECON [49]1.368.16321.3652.153.23052.114
COSMU-21.355.15931.3732.139.22582.104
4 viewsMulti-PaMIR [56]1.401.16501.3852.145.23192.132
MVP [58]1.348.15581.3652.063.22041.972
COSMU1.304.14751.2792.006.21461.929
", + "bbox": [ + 200, + 87, + 786, + 294 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "gle reference. Leveraging RGB and segmentation masks improves the body part classification since the model learns both semantic and appearance information.", + "bbox": [ + 72, + 326, + 920, + 365 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Comparisons", + "text_level": 1, + "bbox": [ + 75, + 393, + 302, + 413 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conduct a comparative evaluation of COSMU against related works that reconstruct complete 3D human shapes from monocular RGB images. We evaluate MVP-Human [58] (MVP) and Multi-Pamir [56] that reconstruct 3D shapes from unconstrained images leveraging SMPL [31] in the process. The same $N_{r} = 3$ images selected in the ablation study (Sec. 4.1) are used as additional views. We evaluate methods that estimate an unseen view of the subject from the target image (PiFU-HD [41], ICON [50] and ECON [49]). We then compare COSMU with PiFU [40] and SuRS [37], which process a single view without any information on hidden regions. To demonstrate the effect of using a lower number of references, we compare COSMU with COSMU-2, which is trained and evaluated by selecting $N_{r} = 1$ reference ( $\\theta_{c} = 180^{\\circ}$ ). Finally, we evaluate $f_{c}$ (Baseline) that reconstructs the coarse target point clouds $\\zeta_{t}$ without leveraging additional views. We use the released weights of ECON and PiFU-HD due to the absence of the training code while we train the other approaches with THuman2.0 [54]. Quantitative evaluation is conducted with the 3D shapes reconstructed from single images of THuman3.0 [45] and X-Humans [42]. The results presented in Tab. 3 demonstrate the superiority of COSMU across all the considered metrics for both datasets. The best scores overall are highlighted in red while the best scores for approaches that use 2 views are blue. COSMU achieves the highest accuracy while COSMU-2 outperforms approaches that rely on 2 views, demonstrating the benefit of leveraging views from 2D images rather than estimating them from the input. The worst results are obtained with single-view approaches. Qualitative results: Fig. 5 illustrates the 3D shapes reconstructed with the considered methods across several scenarios. The first set of shapes is reconstructed by leveraging a collection of images of the same individual in different poses from THuman3.0 [45]. The target and references of the second set are taken from the same monocular video of a person rotating from People-Snapshot [3].", + "bbox": [ + 72, + 422, + 922, + 949 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "M. Pesavento et al.", + "bbox": [ + 157, + 14, + 353, + 32 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In the third set, random images of a person in different poses and with a combination of both identical and different clothing are selected from the Internet. The fourth set of unconstrained images is created with images of the individual in different poses and clothing. This scenario is the most challenging since there are no images of the person in identical clothing. In all these scenarios, COSMU consistently reproduces higher-quality details in the regions of the 3D shape occluded in the target image, significantly outperforming related works. In addition, the quality of regions visible in the input image is comparable to that of related works. MVP struggles to reproduce details and introduces noise when the subject is wearing different clothing in the references ( $3^{rd}$ and $4^{th}$ sets). ECON and PIFu-HD represent details on unseen parts of the model at a significantly lower quality compared to COSMU, which is the only approach that reconstructs the hood on the jumper ( $1^{st}$ set). The shapes reconstructed with PIFu-HD and ECON exhibit excessive smoothing in the regions of the human body occluded in the target image. The shapes reconstructed by ECON and MVP are influenced by the incorrect estimation and fitting of the SMPL model, resulting in an unnatural bending of the 3D body. SuRS reproduces high-quality details only in visible regions, with excessive smoothed reconstruction in other regions. See supplementary materials for additional results.", + "bbox": [ + 77, + 56, + 922, + 424 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Limitations: COSMU's ability to reconstruct high-quality details is contingent on the availability of reference images. COSMU cannot reproduce high-quality details in a region that is not visible in the reference images. Performance deteriorates when the reference images significantly differ from the target. See supplementary for a study on choosing references with varying degrees of similarity to the target. In some examples, COSMU cannot reproduce the same high-quality level of details in small regions of the body, such as the face and hands. The registration algorithm struggles to register small body parts while the multi-view transformer struggles to fuse multiple views of those regions.", + "bbox": [ + 77, + 426, + 922, + 601 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 79, + 632, + 288, + 653 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We introduce a novel framework to reconstruct complete 3D human shapes from a target RGB image by leveraging a collection of monocular unconstrained images without using parametric models. We first developed a novel multi-view scenario simulation module to generate a set of registered 2D normal maps of the target subject. A multi-view attention-based neural implicit model then estimates an implicit function that represents the subject 3D shape. The evaluation demonstrates that the proposed approach can be applied in various scenarios. Significantly higher quality details are reproduced in non-visible regions of the 3D shape compared to state-of-the-art works. In future works, a solution to enhance the quality of smaller body regions will be investigated and the approach will be extended to reconstruct temporally-coherent dynamic shapes.", + "bbox": [ + 77, + 676, + 922, + 889 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Acknowledgement: Work supported by BBC Prosperity Partnership AI4ME: Future Personalised Object-Based Media Experiences Delivered at Scale Anywhere EP/V038087.", + "bbox": [ + 77, + 891, + 937, + 931 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "COSMU: Complete 3D human shape from monocular unconstrained images", + "bbox": [ + 94, + 14, + 837, + 34 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 893, + 16, + 920, + 30 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/eaccd5ddcb08d29634a2442ef5c5d517236a62d04476943f55f261d8b5d04d48.jpg", + "image_caption": [ + "Fig. 5: Qualitative examples of 3D shapes reconstructed in several scenarios. Static images are from THuman3.0 [45] in the first example and from the Internet in the others; monocular videos are from People-Snapshot [3]. Front, back and side views are shown." + ], + "image_footnote": [], + "bbox": [ + 77, + 61, + 922, + 847 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "M. Pesavento et al.", + "bbox": [ + 159, + 16, + 348, + 30 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 77, + 53, + 237, + 74 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alldieck, T., Magnor, M., Bhatnagar, B.L., Theobalt, C., Pons-Moll, G.: Learning to reconstruct people in clothing from a single rgb camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1175-1186 (2019)", + "2. Alldieck, T., Magnor, M., Xu, W., Theobalt, C., Pons-Moll, G.: Detailed human avatars from monocular video. In: 2018 International Conference on 3D Vision (3DV). pp. 98-109. IEEE (2018)", + "3. Alldieck, T., Magnor, M., Xu, W., Theobalt, C., Pons-Moll, G.: Video based reconstruction of 3d people models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8387-8397 (2018)", + "4. Alldieck, T., Pons-Moll, G., Theobalt, C., Magnor, M.: Tex2shape: Detailed full human body geometry from a single image. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2293-2303 (2019)", + "5. Alldieck, T., Zanfir, M., Sminchisescu, C.: Photorealistic monocular 3D reconstruction of humans wearing clothing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1506-1515 (2022)", + "6. Bogo, F., Kanazawa, A., Lassner, C., Gehler, P., Romero, J., Black, M.J.: Keep it SMPL: Automatic estimation of 3d human pose and shape from a single image. In: ECCV (2016)", + "7. Cao, Y., Han, K., Wong, K.Y.K.: Sesdf: Self-evolved signed distance field for implicit 3d clothed human reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4647-4657 (2023)", + "8. Chen, J., Zhang, Y., Kang, D., Zhe, X., Bao, L., Jia, X., Lu, H.: Animatable neural radiance fields from monocular rgb videos. arXiv preprint arXiv:2106.13629 (2021)", + "9. Chen, L., Li, J., Huang, H., Guo, Y.: Crosshuman: learning cross-guidance from multi-frame images for human reconstruction. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 2483-2494 (2022)", + "10. Chen, Z., Zhang, H.: Learning implicit fields for generative shape modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5939-5948 (2019)", + "11. Corona, E., Zanfir, M., Alldieck, T., Bazavan, E.G., Zanfir, A., Sminchisescu, C.: Structured 3d features for reconstructing controllable avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16954-16964 (2023)", + "12. Gardner, A., Elhami, N., Selmic, R.R.: Classifying unordered feature sets with convolutional deep averaging networks. In: 2019 IEEE International Conference on Systems, Man and Cybernetics (SMC). pp. 3447-3453. IEEE (2019)", + "13. Guo, C., Jiang, T., Chen, X., Song, J., Hilliges, O.: Vid2 avatar: 3d avatar reconstruction from videos in the wild via self-supervised scene decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12858-12868 (2023)", + "14. He, T., Collomosse, J., Jin, H., Soatto, S.: Geo-pifu: Geometry and pixel aligned implicit functions for single-view human reconstruction (2020)", + "15. He, T., Xu, Y., Saito, S., Soatto, S., Tung, T.: Arch++: Animation-ready clothed human reconstruction revisited. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11046-11056 (2021)", + "16. Huang, H., Chen, L., Wang, X.: Unconfuse: Avatar reconstruction from unconstrained images. In: European Conference on Computer Vision. pp. 376-381. Springer (2022)" + ], + "bbox": [ + 79, + 87, + 920, + 949 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "COSMU: Complete 3D human shape from monocular unconstrained images", + "bbox": [ + 92, + 14, + 840, + 34 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 893, + 16, + 917, + 30 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Huang, Y., Yi, H., Xiu, Y., Liao, T., Tang, J., Cai, D., Thies, J.: Tech: Text-guided reconstruction of lifelike clothed humans. arXiv preprint arXiv:2308.08545 (2023)", + "18. Huang, Z., Xu, Y., Lassner, C., Li, H., Tung, T.: Arch: Animatable reconstruction of clothed humans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3093-3102 (2020)", + "19. Jayasundara, V., Agrawal, A., Heron, N., Shrivastava, A., Davis, L.S.: Flexnerf: Photorealistic free-viewpoint rendering of moving humans from sparse views. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21118-21127 (2023)", + "20. Jiang, B., Hong, Y., Bao, H., Zhang, J.: Selfrecon: Self reconstruction your digital avatar from monocular video. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "21. Kanazawa, A., Black, M.J., Jacobs, D.W., Malik, J.: End-to-end recovery of human shape and pose. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7122-7131 (2018)", + "22. Kocabas, M., Athanasiou, N., Black, M.J.: Vibe: Video inference for human body pose and shape estimation. In: CVPR (2020)", + "23. Li, C., Lin, J., Lee, G.H.: Ghunerf: Generalizable human nerf from a monocular video. arXiv preprint arXiv:2308.16576 (2023)", + "24. Li, Y., Harada, T.: Non-rigid point cloud registration with neural deformation pyramid. arXiv preprint arXiv:2205.12796 (2022)", + "25. Li, Z., Yu, T., Zheng, Z., Guo, K., Liu, Y.: Posefusion: Pose-guided selective fusion for single-view human volumetric capture. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14162-14172 (2021)", + "26. Li, Z., Zheng, Z., Zhang, H., Ji, C., Liu, Y.: Avatarcap: Animatable avatar conditioned monocular human volumetric capture. In: European Conference on Computer Vision. pp. 322-341. Springer (2022)", + "27. Li, Z., Oskarsson, M., Heyden, A.: Detailed 3d human body reconstruction from multi-view images combining voxel super-resolution and learned implicit representation. Applied Intelligence 52(6), 6739-6759 (2022)", + "28. Liao, T., Zhang, X., Xiu, Y., Yi, H., Liu, X., Qi, G.J., Zhang, Y., Wang, X., Zhu, X., Lei, Z.: High-fidelity clothed avatar reconstruction from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8662-8672 (2023)", + "29. Liu, W., Piao, Z., Min, J., Luo, W., Ma, L., Gao, S.: Liquid warping gan: A unified framework for human motion imitation, appearance transfer and novel view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5904-5913 (2019)", + "30. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM TOG 34(6), 248 (2015)", + "31. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia) 34(6), 248:1-248:16 (Oct 2015)", + "32. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics 21(4), 163-169 (1987)", + "33. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4460-4470 (2019)" + ], + "bbox": [ + 77, + 56, + 920, + 947 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "M. Pesavento et al.", + "bbox": [ + 157, + 14, + 348, + 30 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 165-174 (2019)", + "35. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10975-10985 (2019)", + "36. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: CVPR (2021)", + "37. Pesavento, M., Volino, M., Hilton, A.: Super-resolution 3d human shape from a single low-resolution image. In: ECCV (2022)", + "38. Pesavento, M., Xu, Y., Sarafianos, N., Maier, R., Wang, Z., Yao, C.H., Volino, M., Boyer, E., Hilton, A., Tung, T.: Anim: Accurate neural implicit model for human reconstruction from a single rgb-d image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5448-5458 (2024)", + "39. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)", + "40. Saito, S., Huang, Z., Natsume, R., Morishima, S., Kanazawa, A., Li, H.: Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2304-2314 (2019)", + "41. Saito, S., Simon, T., Saragih, J., Joo, H.: Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 84-93 (2020)", + "42. Shen, K., Guo, C., Kaufmann, M., Zarate, J.J., Valentin, J., Song, J., Hilliges, O.: X-ayar: Expressive human avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16911-16921 (2023)", + "43. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems 34, 6087-6101 (2021)", + "44. Su, H., Maji, S., Kalogerakis, E., Learned-Miller, E.: Multi-view convolutional neural networks for 3d shape recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 945-953 (2015)", + "45. Su, Z., Yu, T., Wang, Y., Liu, Y.: Deepcloth: Neural garment representation for shape and style editing. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(2), 1581-1593 (2023). https://doi.org/10.1109/TPAMI.2022.3168569", + "46. Te, G., Li, X., Li, X., Wang, J., Hu, W., Lu, Y.: Neural capture of animatable 3d human from monocular video. In: European Conference on Computer Vision. pp. 275-291. Springer (2022)", + "47. Varol, G., Ceylan, D., Russell, B., Yang, J., Yumer, E., Laptev, I., Schmid, C.: Bodynet: Volumetric inference of 3d human body shapes. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 20-36 (2018)", + "48. Wang, S., Schwarz, K., Geiger, A., Tang, S.: Arah: Animatable volume rendering of articulated human sdfs. In: European conference on computer vision. pp. 1-19. Springer (2022)", + "49. Xiu, Y., Yang, J., Cao, X., Tzionas, D., Black, M.J.: Econ: Explicit clothed humans optimized via normal integration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 512-523 (2023)" + ], + "bbox": [ + 75, + 56, + 920, + 949 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "COSMU: Complete 3D human shape from monocular unconstrained images", + "bbox": [ + 92, + 14, + 840, + 34 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 893, + 16, + 920, + 30 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "50. Xiu, Y., Yang, J., Tzionas, D., Black, M.J.: Icon: Implicit clothed humans obtained from normals. In: CVPR (2022)", + "51. Xu, X., Loy, C.C.: 3d human texture estimation from a single image with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13849-13858 (2021)", + "52. Yang, X., Chen, X., Wang, S., Gao, D., Han, X., Wang, B.: Have-fun: Human avatar reconstruction from few-shot unconstrained images. arXiv preprint arXiv:2311.15672 (2023)", + "53. Yu, H., Cheang, C., Fu, Y., Xue, X.: Multi-view shape generation for a 3d human-like body. ACM Transactions on Multimedia Computing, Communications and Applications 19(1), 1-22 (2023)", + "54. Yu, T., Zheng, Z., Guo, K., Liu, P., Dai, Q., Liu, Y.: Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR2021) (June 2021)", + "55. Yu, Z., Zhang, L., Xu, Y., Tang, C., Tran, L., Keskin, C., Park, H.S.: Multiview human body reconstruction from uncalibrated cameras. Advances in Neural Information Processing Systems 35, 7879-7891 (2022)", + "56. Zheng, Z., Yu, T., Liu, Y., Dai, Q.: Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)", + "57. Zheng, Z., Yu, T., Wei, Y., Dai, Q., Liu, Y.: Deephuman: 3d human reconstruction from a single image. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7739-7749 (2019)", + "58. Zhu, X., Liao, T., Lyu, J., Yan, X., Wang, Y., Guo, K., Cao, Q., Li, S.Z., Lei, Z.: Mvp-human dataset for 3d human avatar reconstruction from unconstrained frames. arXiv preprint arXiv:2204.11184 (2022)", + "59. Zins, P., Xu, Y., Boyer, E., Wuhrer, S., Tung, T.: Data-driven 3d reconstruction of dressed humans from sparse views. In: 2021 International Conference on 3D Vision (3DV). pp. 494-504. IEEE (2021)", + "60. Zins, P., Xu, Y., Boyer, E., Wuhrer, S., Tung, T.: Multi-view reconstruction using signed ray distance functions (srdf). In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16696-16706 (2023)" + ], + "bbox": [ + 77, + 56, + 917, + 627 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "M. Pesavento et al.", + "bbox": [ + 157, + 14, + 348, + 30 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/COSMU_ Complete 3D human shape from monocular unconstrained images/5adea58c-9270-4078-881c-e504ff34f51f_model.json b/2024/COSMU_ Complete 3D human shape from monocular unconstrained images/5adea58c-9270-4078-881c-e504ff34f51f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5378a3798d0aa7fcc44c511a1c1c1b68775e3a98 --- /dev/null +++ b/2024/COSMU_ Complete 3D human shape from monocular unconstrained images/5adea58c-9270-4078-881c-e504ff34f51f_model.json @@ -0,0 +1,2227 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.143, + 0.048, + 0.856, + 0.108 + ], + "angle": 0, + "content": "COSMU: Complete 3D human shape from monocular unconstrained images" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.141, + 0.797, + 0.16 + ], + "angle": 0, + "content": "Marco Pesavento, Marco Volino, and Adrian Hilton" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.176, + 0.892, + 0.213 + ], + "angle": 0, + "content": "Centre for Vision, Speech and Signal Processing (CVSSP), University of Surrey, Guildford, UK {m.pesavento,m.volino,a.hilton} \\(@\\) surrey.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.256, + 0.851, + 0.683 + ], + "angle": 0, + "content": "Abstract. We present a novel framework to reconstruct complete 3D human shapes from a given target image by leveraging monocular unconstrained images. The objective of this work is to reproduce high-quality details in regions of the reconstructed human body that are not visible in the input target. The proposed methodology addresses the limitations of existing approaches for reconstructing 3D human shapes from a single image, which cannot reproduce shape details in occluded body regions. The missing information of the monocular input can be recovered by using multiple views captured from multiple cameras. However, multi-view reconstruction methods necessitate accurately calibrated and registered images, which can be challenging to obtain in real-world scenarios. Given a target RGB image and a collection of multiple uncalibrated and unregistered images of the same individual, acquired using a single camera, we propose a novel framework to generate complete 3D human shapes. We introduce a novel module to generate 2D multi-view normal maps of the person registered with the target input image. The module consists of body part-based reference selection and body part-based registration. The generated 2D normal maps are then processed by a multi-view attention-based neural implicit model that estimates an implicit representation of the 3D shape, ensuring the reproduction of details in both observed and occluded regions. Extensive experiments demonstrate that the proposed approach estimates higher quality details in the non-visible regions of the 3D clothed human shapes compared to related methods, without using parametric models." + }, + { + "type": "title", + "bbox": [ + 0.083, + 0.715, + 0.309, + 0.736 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.755, + 0.92, + 0.928 + ], + "angle": 0, + "content": "Democratizing the creation of high-quality human avatars is essential to unlock a wide range of immersive technologies including self-representation in the metaverse, virtual conferencing, and virtual try-on. Realistic human avatars are crucial in these applications, making 3D human reconstruction a prominent topic in computer vision and graphics. Research focuses on finding practical solutions to replace existing sophisticated multi-view systems for 3D reconstruction, which are inaccessible to the general community. Single-view 3D human reconstruction methods have been developed to reconstruct 3D human shapes using a simple and cost-effective set-up, such as a single camera." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.93, + 0.919, + 0.948 + ], + "angle": 0, + "content": "Recent methods [5, 11, 14, 15, 18, 40] use neural implicit models to estimate 3D" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.018, + 0.095, + 0.03 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.16, + 0.017, + 0.348, + 0.031 + ], + "angle": 0, + "content": "M. Pesavento et al." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.051, + 0.904, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.31, + 0.921, + 0.365 + ], + "angle": 0, + "content": "Fig. 1: COSMU reconstructs complete 3D human shape from a target RGB input by leveraging a set of monocular images collected from various scenarios, including monocular video (a), images with different poses (b) and different clothing (c)." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.405, + 0.918, + 0.463 + ], + "angle": 0, + "content": "human shapes with high-quality details in visible regions from single RGB images. These works fail to reproduce fine detail in non-visible regions of the input image, resulting in overly smooth reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.464, + 0.918, + 0.579 + ], + "angle": 0, + "content": "To address this problem, solutions extend single-view to multi-view reconstruction, using multiple images of various orientations of the human [7,27,53,59,60]. Multi-view methods represent details across the entire body but require calibrated cameras and registered images, with the subject in the same pose and clothing. Systems to capture such images are inaccessible to the general community in real-world scenarios, significantly limiting reproducibility." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.58, + 0.918, + 0.735 + ], + "angle": 0, + "content": "Recent works [9,16,55,56,58] reconstruct humans from a single target image by leveraging multiple monocular unconstrained images. We define images as unconstrained if camera calibration parameters are not known (uncalibrated) and the images are captured at different time instances resulting in pose and clothing variations without spatial alignment (unregistered). These methods directly fuse features from multiple images, which is a challenging task since diverse views and poses must be fused into a single coherent reconstruction [58]. They also rely on parametric models and cannot use images of the person in different clothing." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.736, + 0.918, + 0.87 + ], + "angle": 0, + "content": "We propose COSMU, a novel framework that reconstructs CComplete 3D human Shape from Monocular Unconstrained images. Our objective is to reconstruct a 3D clothed human shape from a single target image, focusing on reproducing high-quality details in regions of the reconstructed human body that are not visible in the input target. To achieve this, we leverage a collection of monocular unconstrained images of the same individual. COSMU does not rely on parametric models in the reconstruction process as related works." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.872, + 0.918, + 0.949 + ], + "angle": 0, + "content": "The multi-view scenario used by the multi-view reconstruction approaches is simulated from a collection of unconstrained images, with the subject in different poses and clothing and without camera calibration or correspondences. We present a novel pipeline that applies an optimization-based algorithm to select," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.093, + 0.015, + 0.84, + 0.035 + ], + "angle": 0, + "content": "COSMU: Complete 3D human shape from monocular unconstrained images" + }, + { + "type": "page_number", + "bbox": [ + 0.904, + 0.017, + 0.921, + 0.031 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.056, + 0.923, + 0.406 + ], + "angle": 0, + "content": "from the set of unconstrained images, \\( N_r \\) references for each \\( N_p \\) body part of the target image following an orientation-visibility-similarity criteria. A neural implicit model generates a coarse point cloud \\( \\zeta_t \\) of the target and of each selected reference \\( \\zeta_{ref} \\). The 3D points of these point clouds are classified into \\( N_p \\) body parts using a novel pixel-aligned neural network. The corresponding reference body part selected for a specific target part is registered to it. To obtain \\( N_r + 1 \\) views of the subject, 2D normal maps are generated from the registered point clouds. The multi-view scenario is simulated and the 2D normal maps are processed by an attention-based implicit model that estimates the implicit function \\( f_f \\) of the complete 3D human shape. COSMU generates 2D multi-view normal maps instead of RGB images because the references are not constrained to the same clothing as the target and RGB value estimation of 3D points is avoided. Extensive experiments show COSMU's ability to reconstruct complete 3D human shapes from monocular images taken from various scenarios, including monocular videos (Fig. 1a), images with different poses and same clothing (Fig. 1b) and images with different poses and clothing (Fig. 1c). Thanks to the flexibility in creating the reference set, COSMU represents a significant advance for high-quality 3D avatar reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.407, + 0.611, + 0.425 + ], + "angle": 0, + "content": "The contributions presented in this paper include:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.437, + 0.918, + 0.493 + ], + "angle": 0, + "content": "- A novel framework to reconstruct complete 3D human shapes from a target RGB image by leveraging a set of monocular uncalibrated and unregistered images without the use of parametric models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.495, + 0.918, + 0.533 + ], + "angle": 0, + "content": "- A novel pipeline to simulate a multi-view scenario from a collection of unconstrained reference images via reference selection and body part registration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.918, + 0.591 + ], + "angle": 0, + "content": "- Improved performance for human shape reconstruction compared to state-of-the-art methods with higher quality details in the regions of the reconstructed human shape that are occluded in the target image." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.437, + 0.918, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.619, + 0.345, + 0.64 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.657, + 0.923, + 0.951 + ], + "angle": 0, + "content": "3D human reconstruction from monocular cameras Early works on 3D human reconstruction from monocular cameras represent the body shape without capturing shape details [1,3,4,6,21,22,35,47,51,57]. The introduction of neural implicit models [10,33,34] revolutionized 3D modeling by estimating an implicit representation, which reproduces higher-quality shapes due to its inherent property of defining a surface as a level set of an occupancy probability function. PiFU [40] introduced pixel-aligned implicit function for human reconstruction, aligning 3D points with features of the input image to achieve an unprecedented level of detail in the reconstructed shapes. Since then, neural implicit models have become the standard representation for 3D human shapes [5,11,14,18,28]. A significant limitation of these approaches is their inability to reproduce high-quality details in non-visible regions. SuRS [37] enhances shape quality by learning a map from a low to a high-resolution shape, hallucinating details in occluded regions, which are still overly smoothed as no information is provided to the implicit model. To improve the quality of non-visible regions, PiFU-HD [41]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.018, + 0.095, + 0.03 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.16, + 0.017, + 0.348, + 0.032 + ], + "angle": 0, + "content": "M. Pesavento et al." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.057, + 0.921, + 0.446 + ], + "angle": 0, + "content": "generates front and back surface normals of the human from the RGB image. This estimation introduces noise, resulting in smoothed surfaces of occluded regions. ICON [50] improves normal estimation by guiding it with SMPL [30] fitted to the image. Similarly, ECON [49] alleviates the SMPL problem of clothing reconstruction by feeding the estimated front and back normals into a d-BiNI optimizer. Conditioning the learning process on SMPL body poses a challenge for the network, which cannot fully exploit 2D image features. These methods estimate only an additional view, which is insufficient to represent the entire body and its accuracy influences the reconstruction quality of the non-visible regions. TeCH [17] proposes to estimate the unseen view from the input through Score Distillation Sampling [39] (SDS) and represents the human with drivable tetrahedral representation (DMTet [43]). Subject-tailored fine-tuning is required, limiting its generalizability. Other works focus only on 3D reconstruction from monocular videos, requiring to optimize their algorithms on a single monocular video, limiting the generalization and their application scenarios [1-3,20]. NERF-based methods are applied to 3D human reconstruction from sparse views, but their subject-specific nature limits their utility [8,13,19,23,36,46,48]. For a fair comparison with the works that require subject-specific optimization, our approach should be fine-tuned with the data of each specific subject, significantly limiting generalizability. Therefore, these works are not evaluated in this paper." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.447, + 0.694, + 0.464 + ], + "angle": 0, + "content": "3D human shape from unconstrained RGB images" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.921, + 0.852 + ], + "angle": 0, + "content": "PaMIR [56] extended single-view reconstruction to multiple unconstrained images by using the SMPL model as a body reference to establish correspondences across different images. Latent embedding vectors from the correspondent frames are aggregated into a single embedding using mean pooling and then processed by a multi-layer perception (MLP) to estimate the implicit representation. Similarly, MVP-Human [58] fits the SMPL model to each frame to establish correspondences between unconstrained images. Skinning weights of a canonical pose are estimated to guide the fusion of the features from multiple images through an attention-based implicit model. UnconFuse [16] builds upon MVP-Human, with a focus on estimating T-pose shapes from unconstrained images, while posed shapes cannot be reconstructed. CrossHuman [9] selects references following PoseFusion [25] and leveraging the SMPL model. Features are then extracted from both the fitted SMPL and RGB input and fused with a transformer to estimate the implicit representation. These works extract features from unconstrained images and fuse them with a neural model, losing details in regions where the body poses differ among frames. In contrast, COSMU generates 2D multi-view normals of the human in a spatially aligned pose, facilitating the estimation of the final implicit function. Moreover, these works rely on parametric models, whose fitting can introduce noise. COSMU distinguishes itself by not relying on parametric models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.853, + 0.92, + 0.934 + ], + "angle": 0, + "content": "Have-Fun [52] estimates a DMTet [43] representation by integrating SDS [39] and SMPL skinning weights to adapt the unconstrained images to the target one. Similarly to Tech [17], fine-tuning on the target subject is required, limiting generalizability. Other works focus on 3D human reconstruction from other" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.093, + 0.015, + 0.842, + 0.035 + ], + "angle": 0, + "content": "COSMU: Complete 3D human shape from monocular unconstrained images" + }, + { + "type": "page_number", + "bbox": [ + 0.904, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.054, + 0.916, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.312, + 0.925, + 0.42 + ], + "angle": 0, + "content": "Fig. 2: COSMU framework: given a single RGB image as input and a collection of monocular unconstrained images, a multi-view scenario simulation module, which consists of part-based reference selection and registration, is first applied to generate registered multi-view 2D normal maps of the subject in the same pose. An implicit function \\(\\hat{s}_f\\) is then estimated by processing these normal maps with a multi-view attention-based neural implicit model. The 3D mesh is reconstructed via marching cube." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.454, + 0.923, + 0.496 + ], + "angle": 0, + "content": "unconstrained data such as depth images [25, 38] or 3D meshes [26], which is beyond the scope of this paper." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.52, + 0.833, + 0.566 + ], + "angle": 0, + "content": "3 3D human shape reconstruction from monocular unconstrained images" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.58, + 0.925, + 0.949 + ], + "angle": 0, + "content": "We introduce a novel framework that reconstructs COcomplete 3D human Shape from Monocular Unconstrained images (COSMU). To be complete, a 3D human shape should contain high-quality details in every body region. A single image lacks the necessary information to reproduce details across the entire body, which can instead be extracted from multi-view images. In contrast to multiview reconstruction, which requires calibrated and registered images captured with inaccessible systems, COSMU, given a single target image, reproduces fine details in the regions of the 3D shape that are occluded in the input image. This is achieved by leveraging a collection of uncalibrated and unregistered monocular images, referred to as unconstrained images. The input is a target image \\( I_{t} \\) of a person together with multiple views of the same person in various poses and either same or different clothing, from which we extract body part detail non-visible in the target view. As shown in Fig. 2, given a set \\( \\mathbf{U} = \\{I_u^m\\}_{m=1}^{N_u} \\) of \\( N_{u} \\) monocular unconstrained images, COSMU simulates the multi-view scenario of traditional multi-view works using a novel module that employs body part-based reference selection and registration to create 2D multi-view normal maps registered with the target input. A multi-view attention-based implicit model regresses the implicit representation \\( f_{f} \\) of the complete 3D shape from the normals." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.078, + 0.017, + 0.096, + 0.031 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.35, + 0.032 + ], + "angle": 0, + "content": "M. Pesavento et al." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.055, + 0.619, + 0.075 + ], + "angle": 0, + "content": "3.1 Multi-view scenario simulation module" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.088, + 0.922, + 0.277 + ], + "angle": 0, + "content": "Directly processing unconstrained images with a multi-view neural model deteriorates the performance due to their unregistered nature (Sec. 4.1). To address this, we propose a novel pipeline that simulates a multi-view scenario in which the multi-view images represent various orientations of the subject in the same body pose. Given a target image \\( I_{t} \\), we select \\( N_{p} \\times N_{r} \\) references \\( \\left\\{\\{I_{ref}^{(p,r)}\\}_{p=1}^{N_{p}}\\right\\}_{r=1}^{N_{r}} \\), \\( I_{ref} \\in \\mathbf{U} \\), from the collection \\( \\mathbf{U} \\) of unconstrained images applying a body part-based reference selection. 2D registered multi-view normal maps are then created by registering references to the target input via body part-based registration. Set notation is dropped for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.279, + 0.922, + 0.51 + ], + "angle": 0, + "content": "Body part-based reference selection: To improve the body part-based registration, we select references observed from complementary orientations to the target one and with body part poses similar to the corresponding target body part ones. \\( N_{r} \\) references are selected for each \\( N_{p} \\) body part of the target \\( I_{t} \\) via energy optimization following an orientation-visibility-similarity criteria. Semantic masks \\( M \\) and 3D poses \\( P \\) are extracted from the input images. To ensure that the entire body is represented in the selected references, we first divide the \\( N_{u} \\) unconstrained images into \\( N_{r} \\) groups based on the full-body orientation angle \\( \\theta_{u} \\) computed between the reference \\( P_{u} \\) and the target \\( P_{t} \\) 3D poses with the Rodrigues rotation formula. Each group contains a variable number \\( N_{g} \\) of unconstrained images \\( I_{u}^{g}, g \\in [1,\\dots,N_{g}] \\). \\( N_{p} \\) references \\( I_{ref} \\) are selected from each group following 3 criteria:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.511, + 0.922, + 0.59 + ], + "angle": 0, + "content": "1. Body orientation angle: for a complete view of the body, the selected images should complement each other in the body orientation. An orientation metric is introduced to enforce the selection of images in which the body orientation is closer to the central angle \\(\\theta_c^r\\) of the \\(r^{th}\\) group:" + }, + { + "type": "equation", + "bbox": [ + 0.242, + 0.605, + 0.916, + 0.642 + ], + "angle": 0, + "content": "\\[\no _ {g} = \\left| \\theta_ {c} ^ {r} - \\theta_ {u} \\right|, \\quad \\theta_ {c} ^ {r} = r \\cdot \\frac {3 6 0}{N _ {r} + 1}, \\quad r \\in [ 1, \\dots , N _ {r} ] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.667, + 0.92, + 0.726 + ], + "angle": 0, + "content": "2. Body part visibility: every body part should be visible in the selected references to have a complete view of the body. We use a visibility metric that counts the total number of pixels \\( N_{M} \\) of each body part \\( p \\) in the semantic masks \\( M \\)." + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.744, + 0.916, + 0.764 + ], + "angle": 0, + "content": "\\[\nv _ {g} ^ {p} = N _ {M} ^ {p}, \\quad p \\in [ 1, \\dots , N _ {p} ] \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.796, + 0.92, + 0.856 + ], + "angle": 0, + "content": "3. Body part pose similarity: similarity between each body part of \\( I_{u}^{g} \\) and the corresponding body part of \\( I_{t} \\) is evaluated as 3D Euclidean distance between the corresponding 3D bones of \\( P_{g} \\) and \\( P_{t} \\), ensuring spatial consistency:" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.874, + 0.916, + 0.894 + ], + "angle": 0, + "content": "\\[\nb _ {g} ^ {p} = \\left\\| P _ {t} ^ {p} - P _ {g} ^ {p} \\right\\|, \\quad p \\in [ 1, \\dots , N _ {p} ] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.908, + 0.92, + 0.951 + ], + "angle": 0, + "content": "\\(o_{g} = o(I_{u}^{g}), v_{g}^{p} = v^{p}(I_{u}^{g}), b_{g}^{p} = b^{p}(I_{u}^{g})\\) for the \\(g^{th}\\) image \\(I_{u}^{g}\\) of the \\(p^{th}\\) body part. \\(N_{p}\\) references \\(I_{ref}\\) are selected by minimizing the energy \\(E_{g}^{p}\\) over the \\(N_{g}\\) images" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.093, + 0.015, + 0.84, + 0.033 + ], + "angle": 0, + "content": "COSMU: Complete 3D human shape from monocular unconstrained images" + }, + { + "type": "page_number", + "bbox": [ + 0.904, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.056, + 0.338, + 0.077 + ], + "angle": 0, + "content": "\\(I_u^g\\) for each body part \\(p\\):" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.092, + 0.919, + 0.129 + ], + "angle": 0, + "content": "\\[\nI _ {r e f} = \\underset {g} {\\operatorname {a r g m i n}} \\left(E _ {g} ^ {p}\\right) \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.227, + 0.129, + 0.77, + 0.154 + ], + "angle": 0, + "content": "\\[\nE _ {g} ^ {p} = i d x _ {m i n} \\left(o _ {g}\\right) + i d x _ {m a x} \\left(v _ {g} ^ {p}\\right) + i d x _ {m i n} \\left(b _ {g} ^ {p}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.92, + 0.227 + ], + "angle": 0, + "content": "\\(idx_{min}\\) indicates the position \\(idx\\) of the \\(g^{th}\\) image \\(I_u^g\\) in the arrays created by sorting all the \\(N_g\\) images with respect to their specific metric \\(K = o_g, v_g^p, b_g^p\\). \\(idx_{max}\\) sorts the array in inverse order:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.241, + 0.915, + 0.277 + ], + "angle": 0, + "content": "\\[\ni d x _ {m i n} (K) = \\mathrm {i d x} _ {I _ {u} ^ {g}} \\left\\{I _ {u} ^ {1}, \\dots , I _ {u} ^ {N _ {g}} \\right\\}, K \\left(I _ {u} ^ {1}\\right) \\leqslant \\dots \\leqslant K \\left(I _ {u} ^ {N _ {g}}\\right) \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.272, + 0.812, + 0.296 + ], + "angle": 0, + "content": "\\[\ni d x _ {m a x} (K) = \\mathrm {i d x} _ {I _ {u} ^ {g}} \\left\\{I _ {u} ^ {1}, \\dots , I _ {u} ^ {N _ {g}} \\right\\}, K \\left(I _ {u} ^ {1}\\right) \\geqslant \\dots \\geqslant K \\left(I _ {u} ^ {N _ {g}}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.309, + 0.919, + 0.347 + ], + "angle": 0, + "content": "\\(N_{p} \\times N_{r}\\) references are selected by repeating the body part-based reference selection for every group." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.349, + 0.922, + 0.485 + ], + "angle": 0, + "content": "Body part-based registration: To simulate the multi-view scenario given the selected \\( N_{p} \\times N_{r} \\) references, we register each 3D reference body part to the corresponding 3D body part in \\( I_{t} \\). This process begins with the reconstruction of a coarse point cloud \\( \\zeta_{t} \\) for \\( I_{t} \\) and for each of the \\( N_{p} \\) reference images \\( I_{ref} \\) previously selected for the \\( N_{r} \\) groups \\( (\\zeta_{ref}) \\). We use a neural implicit model \\( f_{c} \\) comprising of a U-Net that increases the resolution of the input features \\( \\phi(I) \\) as in SuRS [37]. The coarse implicit function is estimated as:" + }, + { + "type": "equation", + "bbox": [ + 0.328, + 0.5, + 0.918, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\hat {s} _ {c} = f _ {c} \\left(\\phi (\\pi (I)), z (x)\\right), \\quad \\hat {s} _ {c} \\in \\mathbb {R} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.92, + 0.584 + ], + "angle": 0, + "content": "where \\(I\\) is either \\(I_{t}\\) or \\(I_{ref}\\), \\(\\pi\\) represents the orthographic projection of 3D points \\(x\\) into the input images \\(I\\) and \\(z(x)\\) is the depth of \\(x\\). \\(f_{c}\\) is trained with:" + }, + { + "type": "equation", + "bbox": [ + 0.266, + 0.599, + 0.918, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left| f _ {c} \\left(X _ {c} ^ {i}\\right) - f _ {c} ^ {g t} \\left(X _ {c} ^ {i}\\right) \\right| ^ {2} + | I - I _ {H R} | \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.664, + 0.922, + 0.822 + ], + "angle": 0, + "content": "where \\( X_{c}^{i} = (\\phi^{i}(\\pi (I)),z(x)) \\), \\( N \\) is the number of points in space and \\( I_{HR} \\) is the up-sampled ground-truth version of \\( I \\) used to train the U-Net feature extractor. We then retrieve 3D body parts by classifying each 3D point \\( x_{\\zeta} \\) of the point clouds \\( \\zeta_{t / ref} \\) with a novel pixel-aligned classification network \\( f_{\\zeta} \\). An image feature embedding \\( \\phi (I||M) \\) is extracted from the concatenation between the input RGB image and its semantic mask \\( M \\). Each \\( x_{\\zeta} \\) is projected using orthographic projection \\( \\pi \\) onto the input \\( I||M \\) and indexed to \\( \\phi (I||M) \\). An MLP processes this embedding to estimate the body part \\( \\hat{s}_{\\zeta} \\) to which \\( x_{\\zeta} \\) belongs:" + }, + { + "type": "equation", + "bbox": [ + 0.264, + 0.838, + 0.918, + 0.857 + ], + "angle": 0, + "content": "\\[\n\\hat {s} _ {\\zeta} = f _ {\\zeta} \\left(\\phi \\left(\\pi (I | | M)\\right), z \\left(x _ {\\zeta}\\right)\\right), \\quad \\hat {s} _ {\\zeta} \\in [ 1, \\dots , N _ {p} ] \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.871, + 0.922, + 0.95 + ], + "angle": 0, + "content": "\\(N_{p}\\) 3D body parts \\(v_{ref}\\) are defined for each reference point cloud as well as for the target \\(v_{t}\\). Next, the specific \\(p^{th}\\) body part \\(v_{ref}\\) selected for the corresponding \\(p^{th}\\) target body part \\(v_{t}\\) is registered with it as \\(v_{reg} = R(v_{ref}) + T\\), where \\(R\\) is the rotation and \\(T\\) the translation computed via rigid registration. For each" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.079, + 0.017, + 0.096, + 0.03 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.35, + 0.032 + ], + "angle": 0, + "content": "M. Pesavento et al." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.057, + 0.923, + 0.212 + ], + "angle": 0, + "content": "group, the registered \\( N_{p} \\) body parts are combined to create a point cloud \\( \\zeta_{f} \\). \\( N_{r} \\) point clouds \\( \\zeta_{f} \\) registered with \\( \\zeta_{t} \\) are generated, rotated by their \\( \\theta_{u} \\) and, together with \\( \\zeta_{t} \\), used to retrieve \\( N_{r} + 1 \\) 2D normal maps \\( S_{v} \\), one for each view \\( v \\in [1, \\dots, N_{r} + 1] \\). Normal maps are generated instead of RGB images to avoid the estimation of color values for each 3D point of \\( \\zeta_{t / ref} \\) and to allow different clothing in the references. Directly fusing the registered point clouds to obtain the 3D shape is challenging since correspondences are unknown. We train \\( f_{\\zeta} \\) with a Mean Squared Error \\( MSE \\) loss:" + }, + { + "type": "equation", + "bbox": [ + 0.315, + 0.23, + 0.918, + 0.28 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p a r t}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left| f _ {\\zeta} \\left(X _ {\\zeta} ^ {i}\\right) - f _ {\\zeta} ^ {g t} \\left(X _ {\\zeta} ^ {i}\\right) \\right| ^ {2} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.295, + 0.436, + 0.328 + ], + "angle": 0, + "content": "where \\(X_{\\zeta}^{i} = \\left(\\phi^{i}(\\pi (I||M)),z(x_{\\zeta})\\right)\\)" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.355, + 0.563, + 0.375 + ], + "angle": 0, + "content": "3.2 Multi-view implicit representation" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.389, + 0.924, + 0.705 + ], + "angle": 0, + "content": "Neural implicit models have proven to be effective in reproducing high-quality details in reconstructed 3D shapes. This is attributed to the characteristic of the implicit function that does not discretize the 3D space like other representations [40]. Leveraging multiple views to learn the implicit representation further enhances the quality of the reconstruction, as it enables the incorporation of information about the entire body. Concatenating features from multiple images is a suboptimal approach because it imposes an order on the views [59]. Fusion approaches based on statistics have been proposed [12,44] but their pooling operation causes a loss of information regarding the individual view contributions. Inspired by Zins et al. [59], we propose to fuse the features extracted from the multi-view 2D normals using an attention-based architecture. An attention score is computed for each view \\( v \\) based on the compatibility of a query with a corresponding key, resulting in the generation of \\( N_r + 1 \\) features. Each feature contains the original information from the corresponding view, merged with the information from all the other views. An MLP then models a pixel-aligned implicit function \\( f_f \\) to learn the occupancy value \\( \\hat{s}_f \\) of the 3D points in space:" + }, + { + "type": "equation", + "bbox": [ + 0.292, + 0.722, + 0.917, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\hat {s} _ {f} = f _ {f} \\left(A \\left(\\phi_ {v} \\left(\\pi \\left(S _ {v}\\right)\\right), z (x)\\right)\\right), \\quad \\hat {s} _ {f} \\in \\mathbb {R} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.751, + 0.92, + 0.829 + ], + "angle": 0, + "content": "where \\(\\phi_v\\) are the feature extracted from each view \\(v\\) of the surface normal \\(S_v\\), \\(\\pi\\) indicates the pixel-alignment between the 3D points of the space and each \\(\\phi_v\\), \\(A\\) is the attention operation and \\(z(x)\\) is the depth value of the 3D points \\(x\\). We train the multi-view architecture with an MSE loss:" + }, + { + "type": "equation", + "bbox": [ + 0.325, + 0.843, + 0.918, + 0.892 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left| f _ {f} \\left(X _ {f} ^ {i}\\right) - f _ {f} ^ {g t} \\left(X _ {f} ^ {i}\\right) \\right| ^ {2} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.908, + 0.436, + 0.93 + ], + "angle": 0, + "content": "where \\(X_{f}^{i} = A\\left(\\phi_{v}^{i}(\\pi (S_{v})),z(x^{i})\\right)\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.93, + 0.92, + 0.95 + ], + "angle": 0, + "content": "Inference: \\( N_{p} \\times N_{r} \\) references are selected from the collection of unconstrained" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.093, + 0.015, + 0.842, + 0.035 + ], + "angle": 0, + "content": "COSMU: Complete 3D human shape from monocular unconstrained images" + }, + { + "type": "page_number", + "bbox": [ + 0.903, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.052, + 0.898, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.254, + 0.921, + 0.309 + ], + "angle": 0, + "content": "Fig. 3: Qualitative comparisons between 3D shapes from THuman3.0 [45] reconstructed by changing the energy optimization process to select reference images. Front, back and side views of the shapes are illustrated." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.348, + 0.923, + 0.622 + ], + "angle": 0, + "content": "images. \\( N_r + 1 \\) 2D normal maps are generated using body part-based registration to simulate a multi-view scenario with registered views of the subject. The normals are processed by the multi-view attention-based model to estimate the implicit representation. The 3D mesh is reconstructed via marching cube [32]. To achieve pixel-alignment through projection, multi-view works require the parameters of calibrated cameras in specialized capture systems. In COSMU, instead of using the camera projection matrix for pixel-alignment, 3D points of the space are rotated by \\( \\theta_u \\) computed by the multi-view scenario simulation module and projected onto the images using an identity matrix. Existing approaches create the inference 3D grid centered at the origin. If the subject is not centered in the input images, the multi-view pixel-alignment of COSMU is inaccurate since the same rotated 3D point may not be projected in the corresponding pixel in each view. To tackle this, we set the boundary limits of the 3D grid as the boundaries of the coarse point cloud \\( \\zeta_t \\), ensuring multi-view pixel-alignment." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.653, + 0.312, + 0.674 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.696, + 0.925, + 0.95 + ], + "angle": 0, + "content": "We conduct a comprehensive evaluation of COSMU for the task of 3D human shape reconstruction from a target image by leveraging a collection of monocular unconstrained images of the same individual. The neural models of the framework are trained with THuman2.0 dataset [54], which comprises 524 high-resolution surfaces. We render 360-degree view images around each surface, which serves as input data. For the quantitative and qualitative evaluation, we use the THuman3.0 dataset [45], which consists of 20 individuals each captured in various poses (400 models), and the testing set of X-Humans [42] (41 models). For each pose, we render 360-degree views of the subject and we apply the reference selection to select the \\( N_{p} \\times N_{r} \\) references among the images of the subject in different poses, excluding the target one. We evaluate the reconstruction accuracy with three quantitative metrics: the normal reprojection error [40], the average point-to-surface Euclidean distance (P2S) and the Chamfer distance" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.03 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.349, + 0.032 + ], + "angle": 0, + "content": "M. Pesavento et al." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.056, + 0.922, + 0.212 + ], + "angle": 0, + "content": "(CD), expressed in cm. We further test COSMU with real data collected from various scenarios, including monocular videos (People-Snapshot [3], iPER [29] datasets) or images of people in different poses and clothing from the Internet, demonstrating its superiority over related works to reconstruct complete 3D human shapes with higher quality details in regions of the body occluded in the input target image (Sec. 4.2). In contrast to prior works, no parametric models are used. We set \\( N_{p} = 10 \\) and \\( N_{r} = 3 \\). See supplementary for additional implementation details." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.237, + 0.351, + 0.254 + ], + "angle": 0, + "content": "4.1 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.265, + 0.921, + 0.42 + ], + "angle": 0, + "content": "Energy function for reference selection: We conduct an ablation study on the criteria used for selecting the \\( N_{p} \\times N_{r} \\) references in the body part-based reference selection. The variants explored are: (i) Random: no criteria are applied. \\( N_{p} \\) references are randomly selected for each group; (ii) Only \\( \\mathbf{b_p} \\): only the body part pose similarity is considered; (iii) Only \\( \\mathbf{o} \\): the body orientation angle is the only criterion used; (iv) Only \\( \\mathbf{v_p} \\): only the body part visibility is considered; (v) \\( \\mathbf{o} + \\mathbf{v_p} \\): the body part pose similarity is excluded; (vi) \\( \\mathbf{o} + \\mathbf{b_p} \\): the body part visibility is omitted; (vii) \\( \\mathbf{v_p} + \\mathbf{b_p} \\): the orientation angle is not used." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.421, + 0.922, + 0.556 + ], + "angle": 0, + "content": "As illustrated in Fig. 3 and supported by the quantitative results of Tab. 1, the 3D shapes obtained with COSMU exhibit fewer artifacts and finer details. The reference selection plays a crucial role in selecting the most suitable references to facilitate the body part-based registration. Omitting any of the criteria in the energy optimization influences the selection of references, leading to suboptimal choices and, consequently, deteriorating the body part-based registration accuracy and introducing artifacts in the reconstruction." + }, + { + "type": "table_caption", + "bbox": [ + 0.075, + 0.588, + 0.92, + 0.626 + ], + "angle": 0, + "content": "Table 1: Quantitative evaluation of the application of different energy optimization processes to select reference images from THuman3.0 [45]." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.639, + 0.875, + 0.713 + ], + "angle": 0, + "content": "
RandomOnly bpOnly oOnly vpo + vpo + bpvp + bpCOSMU
CD1.4851.4411.4211.4101.3461.3521.3261.304
Normal.1683.1598.1641.1654.1523.1529.1503.1475
P2S1.4241.4151.4031.4061.3321.2791.3081.279
" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.746, + 0.92, + 0.865 + ], + "angle": 0, + "content": "Body part-based registration effectiveness: Directly using \\( N_{r} \\) selected references for the multi-view reconstruction deteriorates the performance since fusing different poses is challenging in the absence of known correspondences. To avoid this, we introduce a novel body part-based registration method to simulate a multi-view scenario. To prove its effectiveness, we modify the algorithm for registering the \\( N_{p} \\times N_{r} \\) coarse reference point clouds \\( \\zeta_{ref} \\) to the target \\( \\zeta_{t} \\):" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.871, + 0.922, + 0.953 + ], + "angle": 0, + "content": "- No Reg: To demonstrate the effectiveness of registering the reference point clouds to the target, we omit the body part-based registration step. We adjust the reference selection to select only \\( N_r \\) references by averaging over \\( p \\) the criteria \\( v_g^p \\) and \\( b_g^p \\) in Eq. (4)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.093, + 0.015, + 0.842, + 0.035 + ], + "angle": 0, + "content": "COSMU: Complete 3D human shape from monocular unconstrained images" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.917, + 0.031 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.054, + 0.921, + 0.091 + ], + "angle": 0, + "content": "Table 2: Quantitative evaluation of the application of different configurations of the body part-based registration algorithm on THuman3.0 [45]." + }, + { + "type": "table", + "bbox": [ + 0.162, + 0.106, + 0.829, + 0.177 + ], + "angle": 0, + "content": "
No RegNon rigidAverageNo RGBNo SegmCOSMU
CD3.5466.9321.3142.2492.1891.304
Normal.2280.4832.1569.2351.2114.1475
P2S2.8206.0241.3092.1692.0051.279
" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.194, + 0.877, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.449, + 0.924, + 0.507 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparisons between 3D shapes from THuman3.0 [45] reconstructed from 2D normal maps generated using different configurations of the body part-based registration algorithm. Front, back and side views of the shapes are shown." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.544, + 0.92, + 0.621 + ], + "angle": 0, + "content": "- Non Rigid: Instead of applying rigid registration among the correspondent body parts of the target and reference point clouds, we apply the state-of-the-art work by Li and Harada [24] to perform non-rigid registration between the full-body point clouds. We use \\( N_r \\) references as in No reg." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.622, + 0.92, + 0.699 + ], + "angle": 0, + "content": "- Average: COSMU registers a target body part with the corresponding body part of a reference selected specifically for that body part. In this configuration, we select only \\( N_r \\) references as in No reg and register all the body parts of the single reference for each group." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.92, + 0.758 + ], + "angle": 0, + "content": "- No RGB: The body part classification model \\( f_{\\zeta} \\) processes a concatenation of 2D RGB images and segmentation masks. To demonstrate the advantages of using these data, we train and test \\( f_{\\zeta} \\) with only segmentation masks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.759, + 0.714, + 0.778 + ], + "angle": 0, + "content": "- No Segm: we train and test \\( f_{\\zeta} \\) with only RGB images." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.544, + 0.92, + 0.778 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.794, + 0.922, + 0.95 + ], + "angle": 0, + "content": "We apply these configurations to generate 2D multi-view normal maps used by the multi-view attention-based implicit model to reconstruct 3D human shapes. Quantitative (Tab. 2) and qualitative results (Fig. 4) show that COSMU configuration is the most effective. Multi-view reconstruction requires registered input since pixel-alignment across different views can be incorrect with unregistered views. Estimating correspondences between point clouds is challenging and the non-rigid registration cannot produce reliable results. Selecting \\( N_{p} \\) references for each body part facilitates the body part registration compared to using a sin" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.106, + 0.032 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.354, + 0.034 + ], + "angle": 0, + "content": "M. Pesavento et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.054, + 0.917, + 0.072 + ], + "angle": 0, + "content": "Table 3: Quantitative comparisons with related works for 3D human reconstruction." + }, + { + "type": "table", + "bbox": [ + 0.202, + 0.088, + 0.787, + 0.295 + ], + "angle": 0, + "content": "
MethodsTHuman3.0X-Humans
CDNormalP2SCDNormalP2S
1 viewPiFU [40]1.698.17851.6592.516.25912.492
SuRS [37]1.573.16931.5542.406.24872.385
Baseline1.566.17321.5702.375.24862.351
2 viewsPiFu-HD [41]1.501.16631.5122.139.22762.121
ICON [50]1.474.16951.4632.210.23952.208
ECON [49]1.368.16321.3652.153.23052.114
COSMU-21.355.15931.3732.139.22582.104
4 viewsMulti-PaMIR [56]1.401.16501.3852.145.23192.132
MVP [58]1.348.15581.3652.063.22041.972
COSMU1.304.14751.2792.006.21461.929
" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.327, + 0.921, + 0.366 + ], + "angle": 0, + "content": "gle reference. Leveraging RGB and segmentation masks improves the body part classification since the model learns both semantic and appearance information." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.394, + 0.304, + 0.414 + ], + "angle": 0, + "content": "4.2 Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.424, + 0.923, + 0.951 + ], + "angle": 0, + "content": "We conduct a comparative evaluation of COSMU against related works that reconstruct complete 3D human shapes from monocular RGB images. We evaluate MVP-Human [58] (MVP) and Multi-Pamir [56] that reconstruct 3D shapes from unconstrained images leveraging SMPL [31] in the process. The same \\( N_{r} = 3 \\) images selected in the ablation study (Sec. 4.1) are used as additional views. We evaluate methods that estimate an unseen view of the subject from the target image (PiFU-HD [41], ICON [50] and ECON [49]). We then compare COSMU with PiFU [40] and SuRS [37], which process a single view without any information on hidden regions. To demonstrate the effect of using a lower number of references, we compare COSMU with COSMU-2, which is trained and evaluated by selecting \\( N_{r} = 1 \\) reference (\\( \\theta_{c} = 180^{\\circ} \\)). Finally, we evaluate \\( f_{c} \\) (Baseline) that reconstructs the coarse target point clouds \\( \\zeta_{t} \\) without leveraging additional views. We use the released weights of ECON and PiFU-HD due to the absence of the training code while we train the other approaches with THuman2.0 [54]. Quantitative evaluation is conducted with the 3D shapes reconstructed from single images of THuman3.0 [45] and X-Humans [42]. The results presented in Tab. 3 demonstrate the superiority of COSMU across all the considered metrics for both datasets. The best scores overall are highlighted in red while the best scores for approaches that use 2 views are blue. COSMU achieves the highest accuracy while COSMU-2 outperforms approaches that rely on 2 views, demonstrating the benefit of leveraging views from 2D images rather than estimating them from the input. The worst results are obtained with single-view approaches. Qualitative results: Fig. 5 illustrates the 3D shapes reconstructed with the considered methods across several scenarios. The first set of shapes is reconstructed by leveraging a collection of images of the same individual in different poses from THuman3.0 [45]. The target and references of the second set are taken from the same monocular video of a person rotating from People-Snapshot [3]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.096, + 0.015, + 0.84, + 0.035 + ], + "angle": 0, + "content": "COSMU: Complete 3D human shape from monocular unconstrained images" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.921, + 0.032 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.057, + 0.923, + 0.426 + ], + "angle": 0, + "content": "In the third set, random images of a person in different poses and with a combination of both identical and different clothing are selected from the Internet. The fourth set of unconstrained images is created with images of the individual in different poses and clothing. This scenario is the most challenging since there are no images of the person in identical clothing. In all these scenarios, COSMU consistently reproduces higher-quality details in the regions of the 3D shape occluded in the target image, significantly outperforming related works. In addition, the quality of regions visible in the input image is comparable to that of related works. MVP struggles to reproduce details and introduces noise when the subject is wearing different clothing in the references (\\(3^{rd}\\) and \\(4^{th}\\) sets). ECON and PIFu-HD represent details on unseen parts of the model at a significantly lower quality compared to COSMU, which is the only approach that reconstructs the hood on the jumper (\\(1^{st}\\) set). The shapes reconstructed with PIFu-HD and ECON exhibit excessive smoothing in the regions of the human body occluded in the target image. The shapes reconstructed by ECON and MVP are influenced by the incorrect estimation and fitting of the SMPL model, resulting in an unnatural bending of the 3D body. SuRS reproduces high-quality details only in visible regions, with excessive smoothed reconstruction in other regions. See supplementary materials for additional results." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.427, + 0.923, + 0.603 + ], + "angle": 0, + "content": "Limitations: COSMU's ability to reconstruct high-quality details is contingent on the availability of reference images. COSMU cannot reproduce high-quality details in a region that is not visible in the reference images. Performance deteriorates when the reference images significantly differ from the target. See supplementary for a study on choosing references with varying degrees of similarity to the target. In some examples, COSMU cannot reproduce the same high-quality level of details in small regions of the body, such as the face and hands. The registration algorithm struggles to register small body parts while the multi-view transformer struggles to fuse multiple views of those regions." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.633, + 0.289, + 0.655 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.677, + 0.923, + 0.891 + ], + "angle": 0, + "content": "We introduce a novel framework to reconstruct complete 3D human shapes from a target RGB image by leveraging a collection of monocular unconstrained images without using parametric models. We first developed a novel multi-view scenario simulation module to generate a set of registered 2D normal maps of the target subject. A multi-view attention-based neural implicit model then estimates an implicit function that represents the subject 3D shape. The evaluation demonstrates that the proposed approach can be applied in various scenarios. Significantly higher quality details are reproduced in non-visible regions of the 3D shape compared to state-of-the-art works. In future works, a solution to enhance the quality of smaller body regions will be investigated and the approach will be extended to reconstruct temporally-coherent dynamic shapes." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.892, + 0.938, + 0.932 + ], + "angle": 0, + "content": "Acknowledgement: Work supported by BBC Prosperity Partnership AI4ME: Future Personalised Object-Based Media Experiences Delivered at Scale Anywhere EP/V038087." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.031 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.16, + 0.017, + 0.35, + 0.032 + ], + "angle": 0, + "content": "M. Pesavento et al." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.062, + 0.924, + 0.848 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.859, + 0.924, + 0.93 + ], + "angle": 0, + "content": "Fig. 5: Qualitative examples of 3D shapes reconstructed in several scenarios. Static images are from THuman3.0 [45] in the first example and from the Internet in the others; monocular videos are from People-Snapshot [3]. Front, back and side views are shown." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.093, + 0.015, + 0.842, + 0.035 + ], + "angle": 0, + "content": "COSMU: Complete 3D human shape from monocular unconstrained images" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.054, + 0.238, + 0.075 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.089, + 0.92, + 0.161 + ], + "angle": 0, + "content": "1. Alldieck, T., Magnor, M., Bhatnagar, B.L., Theobalt, C., Pons-Moll, G.: Learning to reconstruct people in clothing from a single rgb camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1175-1186 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.161, + 0.92, + 0.214 + ], + "angle": 0, + "content": "2. Alldieck, T., Magnor, M., Xu, W., Theobalt, C., Pons-Moll, G.: Detailed human avatars from monocular video. In: 2018 International Conference on 3D Vision (3DV). pp. 98-109. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.214, + 0.921, + 0.267 + ], + "angle": 0, + "content": "3. Alldieck, T., Magnor, M., Xu, W., Theobalt, C., Pons-Moll, G.: Video based reconstruction of 3d people models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8387-8397 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.267, + 0.921, + 0.319 + ], + "angle": 0, + "content": "4. Alldieck, T., Pons-Moll, G., Theobalt, C., Magnor, M.: Tex2shape: Detailed full human body geometry from a single image. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2293-2303 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.319, + 0.921, + 0.372 + ], + "angle": 0, + "content": "5. Alldieck, T., Zanfir, M., Sminchisescu, C.: Photorealistic monocular 3D reconstruction of humans wearing clothing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1506-1515 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.372, + 0.921, + 0.424 + ], + "angle": 0, + "content": "6. Bogo, F., Kanazawa, A., Lassner, C., Gehler, P., Romero, J., Black, M.J.: Keep it SMPL: Automatic estimation of 3d human pose and shape from a single image. In: ECCV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.424, + 0.921, + 0.477 + ], + "angle": 0, + "content": "7. Cao, Y., Han, K., Wong, K.Y.K.: Sesdf: Self-evolved signed distance field for implicit 3d clothed human reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4647-4657 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.477, + 0.921, + 0.511 + ], + "angle": 0, + "content": "8. Chen, J., Zhang, Y., Kang, D., Zhe, X., Bao, L., Jia, X., Lu, H.: Animatable neural radiance fields from monocular rgb videos. arXiv preprint arXiv:2106.13629 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.511, + 0.921, + 0.564 + ], + "angle": 0, + "content": "9. Chen, L., Li, J., Huang, H., Guo, Y.: Crosshuman: learning cross-guidance from multi-frame images for human reconstruction. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 2483-2494 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.564, + 0.92, + 0.616 + ], + "angle": 0, + "content": "10. Chen, Z., Zhang, H.: Learning implicit fields for generative shape modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5939-5948 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.616, + 0.921, + 0.687 + ], + "angle": 0, + "content": "11. Corona, E., Zanfir, M., Alldieck, T., Bazavan, E.G., Zanfir, A., Sminchisescu, C.: Structured 3d features for reconstructing controllable avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16954-16964 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.687, + 0.92, + 0.739 + ], + "angle": 0, + "content": "12. Gardner, A., Elhami, N., Selmic, R.R.: Classifying unordered feature sets with convolutional deep averaging networks. In: 2019 IEEE International Conference on Systems, Man and Cybernetics (SMC). pp. 3447-3453. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.739, + 0.92, + 0.809 + ], + "angle": 0, + "content": "13. Guo, C., Jiang, T., Chen, X., Song, J., Hilliges, O.: Vid2 avatar: 3d avatar reconstruction from videos in the wild via self-supervised scene decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12858-12868 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.809, + 0.92, + 0.844 + ], + "angle": 0, + "content": "14. He, T., Collomosse, J., Jin, H., Soatto, S.: Geo-pifu: Geometry and pixel aligned implicit functions for single-view human reconstruction (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.844, + 0.92, + 0.897 + ], + "angle": 0, + "content": "15. He, T., Xu, Y., Saito, S., Soatto, S., Tung, T.: Arch++: Animation-ready clothed human reconstruction revisited. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11046-11056 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.897, + 0.92, + 0.95 + ], + "angle": 0, + "content": "16. Huang, H., Chen, L., Wang, X.: Unconfuse: Avatar reconstruction from unconstrained images. In: European Conference on Computer Vision. pp. 376-381. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.089, + 0.921, + 0.95 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.106, + 0.031 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.35, + 0.032 + ], + "angle": 0, + "content": "M. Pesavento et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.057, + 0.918, + 0.093 + ], + "angle": 0, + "content": "17. Huang, Y., Yi, H., Xiu, Y., Liao, T., Tang, J., Cai, D., Thies, J.: Tech: Text-guided reconstruction of lifelike clothed humans. arXiv preprint arXiv:2308.08545 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.096, + 0.92, + 0.147 + ], + "angle": 0, + "content": "18. Huang, Z., Xu, Y., Lassner, C., Li, H., Tung, T.: Arch: Animatable reconstruction of clothed humans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3093-3102 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.921, + 0.22 + ], + "angle": 0, + "content": "19. Jayasundara, V., Agrawal, A., Heron, N., Shrivastava, A., Davis, L.S.: Flexnerf: Photorealistic free-viewpoint rendering of moving humans from sparse views. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21118-21127 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.222, + 0.92, + 0.276 + ], + "angle": 0, + "content": "20. Jiang, B., Hong, Y., Bao, H., Zhang, J.: Selfrecon: Self reconstruction your digital avatar from monocular video. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.92, + 0.33 + ], + "angle": 0, + "content": "21. Kanazawa, A., Black, M.J., Jacobs, D.W., Malik, J.: End-to-end recovery of human shape and pose. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7122-7131 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.332, + 0.918, + 0.367 + ], + "angle": 0, + "content": "22. Kocabas, M., Athanasiou, N., Black, M.J.: Vibe: Video inference for human body pose and shape estimation. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.368, + 0.918, + 0.403 + ], + "angle": 0, + "content": "23. Li, C., Lin, J., Lee, G.H.: Ghunerf: Generalizable human nerf from a monocular video. arXiv preprint arXiv:2308.16576 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.918, + 0.44 + ], + "angle": 0, + "content": "24. Li, Y., Harada, T.: Non-rigid point cloud registration with neural deformation pyramid. arXiv preprint arXiv:2205.12796 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.918, + 0.495 + ], + "angle": 0, + "content": "25. Li, Z., Yu, T., Zheng, Z., Guo, K., Liu, Y.: Posefusion: Pose-guided selective fusion for single-view human volumetric capture. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14162-14172 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.918, + 0.549 + ], + "angle": 0, + "content": "26. Li, Z., Zheng, Z., Zhang, H., Ji, C., Liu, Y.: Avatarcap: Animatable avatar conditioned monocular human volumetric capture. In: European Conference on Computer Vision. pp. 322-341. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.551, + 0.918, + 0.603 + ], + "angle": 0, + "content": "27. Li, Z., Oskarsson, M., Heyden, A.: Detailed 3d human body reconstruction from multi-view images combining voxel super-resolution and learned implicit representation. Applied Intelligence 52(6), 6739-6759 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.918, + 0.676 + ], + "angle": 0, + "content": "28. Liao, T., Zhang, X., Xiu, Y., Yi, H., Liu, X., Qi, G.J., Zhang, Y., Wang, X., Zhu, X., Lei, Z.: High-fidelity clothed avatar reconstruction from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8662-8672 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.918, + 0.748 + ], + "angle": 0, + "content": "29. Liu, W., Piao, Z., Min, J., Luo, W., Ma, L., Gao, S.: Liquid warping gan: A unified framework for human motion imitation, appearance transfer and novel view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5904-5913 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.75, + 0.918, + 0.785 + ], + "angle": 0, + "content": "30. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM TOG 34(6), 248 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.918, + 0.84 + ], + "angle": 0, + "content": "31. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia) 34(6), 248:1-248:16 (Oct 2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.842, + 0.918, + 0.877 + ], + "angle": 0, + "content": "32. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics 21(4), 163-169 (1987)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.879, + 0.918, + 0.949 + ], + "angle": 0, + "content": "33. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4460-4470 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.057, + 0.921, + 0.949 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.093, + 0.015, + 0.842, + 0.035 + ], + "angle": 0, + "content": "COSMU: Complete 3D human shape from monocular unconstrained images" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.921, + 0.031 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.057, + 0.922, + 0.129 + ], + "angle": 0, + "content": "34. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 165-174 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.077, + 0.129, + 0.921, + 0.199 + ], + "angle": 0, + "content": "35. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10975-10985 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.199, + 0.921, + 0.251 + ], + "angle": 0, + "content": "36. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.251, + 0.921, + 0.285 + ], + "angle": 0, + "content": "37. Pesavento, M., Volino, M., Hilton, A.: Super-resolution 3d human shape from a single low-resolution image. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.285, + 0.921, + 0.356 + ], + "angle": 0, + "content": "38. Pesavento, M., Xu, Y., Sarafianos, N., Maier, R., Wang, Z., Yao, C.H., Volino, M., Boyer, E., Hilton, A., Tung, T.: Anim: Accurate neural implicit model for human reconstruction from a single rgb-d image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5448-5458 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.356, + 0.921, + 0.39 + ], + "angle": 0, + "content": "39. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.39, + 0.921, + 0.46 + ], + "angle": 0, + "content": "40. Saito, S., Huang, Z., Natsume, R., Morishima, S., Kanazawa, A., Li, H.: Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2304-2314 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.46, + 0.921, + 0.53 + ], + "angle": 0, + "content": "41. Saito, S., Simon, T., Saragih, J., Joo, H.: Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 84-93 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.53, + 0.921, + 0.583 + ], + "angle": 0, + "content": "42. Shen, K., Guo, C., Kaufmann, M., Zarate, J.J., Valentin, J., Song, J., Hilliges, O.: X-ayar: Expressive human avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16911-16921 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.583, + 0.921, + 0.635 + ], + "angle": 0, + "content": "43. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems 34, 6087-6101 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.634, + 0.921, + 0.687 + ], + "angle": 0, + "content": "44. Su, H., Maji, S., Kalogerakis, E., Learned-Miller, E.: Multi-view convolutional neural networks for 3d shape recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 945-953 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.687, + 0.921, + 0.74 + ], + "angle": 0, + "content": "45. Su, Z., Yu, T., Wang, Y., Liu, Y.: Deepcloth: Neural garment representation for shape and style editing. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(2), 1581-1593 (2023). https://doi.org/10.1109/TPAMI.2022.3168569" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.74, + 0.921, + 0.792 + ], + "angle": 0, + "content": "46. Te, G., Li, X., Li, X., Wang, J., Hu, W., Lu, Y.: Neural capture of animatable 3d human from monocular video. In: European Conference on Computer Vision. pp. 275-291. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.792, + 0.921, + 0.845 + ], + "angle": 0, + "content": "47. Varol, G., Ceylan, D., Russell, B., Yang, J., Yumer, E., Laptev, I., Schmid, C.: Bodynet: Volumetric inference of 3d human body shapes. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 20-36 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.844, + 0.921, + 0.897 + ], + "angle": 0, + "content": "48. Wang, S., Schwarz, K., Geiger, A., Tang, S.: Arah: Animatable volume rendering of articulated human sdfs. In: European conference on computer vision. pp. 1-19. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.897, + 0.921, + 0.95 + ], + "angle": 0, + "content": "49. Xiu, Y., Yang, J., Cao, X., Tzionas, D., Black, M.J.: Econ: Explicit clothed humans optimized via normal integration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 512-523 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.057, + 0.922, + 0.95 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.031 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.35, + 0.032 + ], + "angle": 0, + "content": "M. Pesavento et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.058, + 0.918, + 0.093 + ], + "angle": 0, + "content": "50. Xiu, Y., Yang, J., Tzionas, D., Black, M.J.: Icon: Implicit clothed humans obtained from normals. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.094, + 0.92, + 0.147 + ], + "angle": 0, + "content": "51. Xu, X., Loy, C.C.: 3d human texture estimation from a single image with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13849-13858 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.92, + 0.2 + ], + "angle": 0, + "content": "52. Yang, X., Chen, X., Wang, S., Gao, D., Han, X., Wang, B.: Have-fun: Human avatar reconstruction from few-shot unconstrained images. arXiv preprint arXiv:2311.15672 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.201, + 0.92, + 0.254 + ], + "angle": 0, + "content": "53. Yu, H., Cheang, C., Fu, Y., Xue, X.: Multi-view shape generation for a 3d human-like body. ACM Transactions on Multimedia Computing, Communications and Applications 19(1), 1-22 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.255, + 0.918, + 0.307 + ], + "angle": 0, + "content": "54. Yu, T., Zheng, Z., Guo, K., Liu, P., Dai, Q., Liu, Y.: Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR2021) (June 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.308, + 0.918, + 0.36 + ], + "angle": 0, + "content": "55. Yu, Z., Zhang, L., Xu, Y., Tang, C., Tran, L., Keskin, C., Park, H.S.: Multiview human body reconstruction from uncalibrated cameras. Advances in Neural Information Processing Systems 35, 7879-7891 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.361, + 0.918, + 0.413 + ], + "angle": 0, + "content": "56. Zheng, Z., Yu, T., Liu, Y., Dai, Q.: Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.415, + 0.918, + 0.467 + ], + "angle": 0, + "content": "57. Zheng, Z., Yu, T., Wei, Y., Dai, Q., Liu, Y.: Deephuman: 3d human reconstruction from a single image. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7739-7749 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.468, + 0.918, + 0.521 + ], + "angle": 0, + "content": "58. Zhu, X., Liao, T., Lyu, J., Yan, X., Wang, Y., Guo, K., Cao, Q., Li, S.Z., Lei, Z.: Mvp-human dataset for 3d human avatar reconstruction from unconstrained frames. arXiv preprint arXiv:2204.11184 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.522, + 0.92, + 0.575 + ], + "angle": 0, + "content": "59. Zins, P., Xu, Y., Boyer, E., Wuhrer, S., Tung, T.: Data-driven 3d reconstruction of dressed humans from sparse views. In: 2021 International Conference on 3D Vision (3DV). pp. 494-504. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.918, + 0.628 + ], + "angle": 0, + "content": "60. Zins, P., Xu, Y., Boyer, E., Wuhrer, S., Tung, T.: Multi-view reconstruction using signed ray distance functions (srdf). In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16696-16706 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.058, + 0.92, + 0.628 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/COSMU_ Complete 3D human shape from monocular unconstrained images/5adea58c-9270-4078-881c-e504ff34f51f_origin.pdf b/2024/COSMU_ Complete 3D human shape from monocular unconstrained images/5adea58c-9270-4078-881c-e504ff34f51f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ba982da3a0534acc5c4f05c312c5b2ed7d2f9d7e --- /dev/null +++ b/2024/COSMU_ Complete 3D human shape from monocular unconstrained images/5adea58c-9270-4078-881c-e504ff34f51f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51ade96044928c7c0d11c21fe6948b09a8b28b6cda03d907e5927268da21a03c +size 3585120 diff --git a/2024/COSMU_ Complete 3D human shape from monocular unconstrained images/full.md b/2024/COSMU_ Complete 3D human shape from monocular unconstrained images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..72b1a28e3279987cc10f3eb8a5939fd58432dd8e --- /dev/null +++ b/2024/COSMU_ Complete 3D human shape from monocular unconstrained images/full.md @@ -0,0 +1,277 @@ +# COSMU: Complete 3D human shape from monocular unconstrained images + +Marco Pesavento, Marco Volino, and Adrian Hilton + +Centre for Vision, Speech and Signal Processing (CVSSP), University of Surrey, Guildford, UK {m.pesavento,m.volino,a.hilton} $@$ surrey.ac.uk + +Abstract. We present a novel framework to reconstruct complete 3D human shapes from a given target image by leveraging monocular unconstrained images. The objective of this work is to reproduce high-quality details in regions of the reconstructed human body that are not visible in the input target. The proposed methodology addresses the limitations of existing approaches for reconstructing 3D human shapes from a single image, which cannot reproduce shape details in occluded body regions. The missing information of the monocular input can be recovered by using multiple views captured from multiple cameras. However, multi-view reconstruction methods necessitate accurately calibrated and registered images, which can be challenging to obtain in real-world scenarios. Given a target RGB image and a collection of multiple uncalibrated and unregistered images of the same individual, acquired using a single camera, we propose a novel framework to generate complete 3D human shapes. We introduce a novel module to generate 2D multi-view normal maps of the person registered with the target input image. The module consists of body part-based reference selection and body part-based registration. The generated 2D normal maps are then processed by a multi-view attention-based neural implicit model that estimates an implicit representation of the 3D shape, ensuring the reproduction of details in both observed and occluded regions. Extensive experiments demonstrate that the proposed approach estimates higher quality details in the non-visible regions of the 3D clothed human shapes compared to related methods, without using parametric models. + +# 1 Introduction + +Democratizing the creation of high-quality human avatars is essential to unlock a wide range of immersive technologies including self-representation in the metaverse, virtual conferencing, and virtual try-on. Realistic human avatars are crucial in these applications, making 3D human reconstruction a prominent topic in computer vision and graphics. Research focuses on finding practical solutions to replace existing sophisticated multi-view systems for 3D reconstruction, which are inaccessible to the general community. Single-view 3D human reconstruction methods have been developed to reconstruct 3D human shapes using a simple and cost-effective set-up, such as a single camera. + +Recent methods [5, 11, 14, 15, 18, 40] use neural implicit models to estimate 3D + +![](images/f2389ca12bc43c05236a13c98905f15d86cdd3ea520c80c9206abf9205fb33e2.jpg) +Fig. 1: COSMU reconstructs complete 3D human shape from a target RGB input by leveraging a set of monocular images collected from various scenarios, including monocular video (a), images with different poses (b) and different clothing (c). + +human shapes with high-quality details in visible regions from single RGB images. These works fail to reproduce fine detail in non-visible regions of the input image, resulting in overly smooth reconstruction. + +To address this problem, solutions extend single-view to multi-view reconstruction, using multiple images of various orientations of the human [7,27,53,59,60]. Multi-view methods represent details across the entire body but require calibrated cameras and registered images, with the subject in the same pose and clothing. Systems to capture such images are inaccessible to the general community in real-world scenarios, significantly limiting reproducibility. + +Recent works [9,16,55,56,58] reconstruct humans from a single target image by leveraging multiple monocular unconstrained images. We define images as unconstrained if camera calibration parameters are not known (uncalibrated) and the images are captured at different time instances resulting in pose and clothing variations without spatial alignment (unregistered). These methods directly fuse features from multiple images, which is a challenging task since diverse views and poses must be fused into a single coherent reconstruction [58]. They also rely on parametric models and cannot use images of the person in different clothing. + +We propose COSMU, a novel framework that reconstructs CComplete 3D human Shape from Monocular Unconstrained images. Our objective is to reconstruct a 3D clothed human shape from a single target image, focusing on reproducing high-quality details in regions of the reconstructed human body that are not visible in the input target. To achieve this, we leverage a collection of monocular unconstrained images of the same individual. COSMU does not rely on parametric models in the reconstruction process as related works. + +The multi-view scenario used by the multi-view reconstruction approaches is simulated from a collection of unconstrained images, with the subject in different poses and clothing and without camera calibration or correspondences. We present a novel pipeline that applies an optimization-based algorithm to select, + +from the set of unconstrained images, $N_r$ references for each $N_p$ body part of the target image following an orientation-visibility-similarity criteria. A neural implicit model generates a coarse point cloud $\zeta_t$ of the target and of each selected reference $\zeta_{ref}$ . The 3D points of these point clouds are classified into $N_p$ body parts using a novel pixel-aligned neural network. The corresponding reference body part selected for a specific target part is registered to it. To obtain $N_r + 1$ views of the subject, 2D normal maps are generated from the registered point clouds. The multi-view scenario is simulated and the 2D normal maps are processed by an attention-based implicit model that estimates the implicit function $f_f$ of the complete 3D human shape. COSMU generates 2D multi-view normal maps instead of RGB images because the references are not constrained to the same clothing as the target and RGB value estimation of 3D points is avoided. Extensive experiments show COSMU's ability to reconstruct complete 3D human shapes from monocular images taken from various scenarios, including monocular videos (Fig. 1a), images with different poses and same clothing (Fig. 1b) and images with different poses and clothing (Fig. 1c). Thanks to the flexibility in creating the reference set, COSMU represents a significant advance for high-quality 3D avatar reconstruction. + +The contributions presented in this paper include: + +- A novel framework to reconstruct complete 3D human shapes from a target RGB image by leveraging a set of monocular uncalibrated and unregistered images without the use of parametric models. +- A novel pipeline to simulate a multi-view scenario from a collection of unconstrained reference images via reference selection and body part registration. +- Improved performance for human shape reconstruction compared to state-of-the-art methods with higher quality details in the regions of the reconstructed human shape that are occluded in the target image. + +# 2 Related Works + +3D human reconstruction from monocular cameras Early works on 3D human reconstruction from monocular cameras represent the body shape without capturing shape details [1,3,4,6,21,22,35,47,51,57]. The introduction of neural implicit models [10,33,34] revolutionized 3D modeling by estimating an implicit representation, which reproduces higher-quality shapes due to its inherent property of defining a surface as a level set of an occupancy probability function. PiFU [40] introduced pixel-aligned implicit function for human reconstruction, aligning 3D points with features of the input image to achieve an unprecedented level of detail in the reconstructed shapes. Since then, neural implicit models have become the standard representation for 3D human shapes [5,11,14,18,28]. A significant limitation of these approaches is their inability to reproduce high-quality details in non-visible regions. SuRS [37] enhances shape quality by learning a map from a low to a high-resolution shape, hallucinating details in occluded regions, which are still overly smoothed as no information is provided to the implicit model. To improve the quality of non-visible regions, PiFU-HD [41] + +generates front and back surface normals of the human from the RGB image. This estimation introduces noise, resulting in smoothed surfaces of occluded regions. ICON [50] improves normal estimation by guiding it with SMPL [30] fitted to the image. Similarly, ECON [49] alleviates the SMPL problem of clothing reconstruction by feeding the estimated front and back normals into a d-BiNI optimizer. Conditioning the learning process on SMPL body poses a challenge for the network, which cannot fully exploit 2D image features. These methods estimate only an additional view, which is insufficient to represent the entire body and its accuracy influences the reconstruction quality of the non-visible regions. TeCH [17] proposes to estimate the unseen view from the input through Score Distillation Sampling [39] (SDS) and represents the human with drivable tetrahedral representation (DMTet [43]). Subject-tailored fine-tuning is required, limiting its generalizability. Other works focus only on 3D reconstruction from monocular videos, requiring to optimize their algorithms on a single monocular video, limiting the generalization and their application scenarios [1-3,20]. NERF-based methods are applied to 3D human reconstruction from sparse views, but their subject-specific nature limits their utility [8,13,19,23,36,46,48]. For a fair comparison with the works that require subject-specific optimization, our approach should be fine-tuned with the data of each specific subject, significantly limiting generalizability. Therefore, these works are not evaluated in this paper. + +# 3D human shape from unconstrained RGB images + +PaMIR [56] extended single-view reconstruction to multiple unconstrained images by using the SMPL model as a body reference to establish correspondences across different images. Latent embedding vectors from the correspondent frames are aggregated into a single embedding using mean pooling and then processed by a multi-layer perception (MLP) to estimate the implicit representation. Similarly, MVP-Human [58] fits the SMPL model to each frame to establish correspondences between unconstrained images. Skinning weights of a canonical pose are estimated to guide the fusion of the features from multiple images through an attention-based implicit model. UnconFuse [16] builds upon MVP-Human, with a focus on estimating T-pose shapes from unconstrained images, while posed shapes cannot be reconstructed. CrossHuman [9] selects references following PoseFusion [25] and leveraging the SMPL model. Features are then extracted from both the fitted SMPL and RGB input and fused with a transformer to estimate the implicit representation. These works extract features from unconstrained images and fuse them with a neural model, losing details in regions where the body poses differ among frames. In contrast, COSMU generates 2D multi-view normals of the human in a spatially aligned pose, facilitating the estimation of the final implicit function. Moreover, these works rely on parametric models, whose fitting can introduce noise. COSMU distinguishes itself by not relying on parametric models. + +Have-Fun [52] estimates a DMTet [43] representation by integrating SDS [39] and SMPL skinning weights to adapt the unconstrained images to the target one. Similarly to Tech [17], fine-tuning on the target subject is required, limiting generalizability. Other works focus on 3D human reconstruction from other + +![](images/69f9791d2a5e6317f14aa40b8878a9a84498d417dc26ec2042289c31f701062a.jpg) +Fig. 2: COSMU framework: given a single RGB image as input and a collection of monocular unconstrained images, a multi-view scenario simulation module, which consists of part-based reference selection and registration, is first applied to generate registered multi-view 2D normal maps of the subject in the same pose. An implicit function $\hat{s}_f$ is then estimated by processing these normal maps with a multi-view attention-based neural implicit model. The 3D mesh is reconstructed via marching cube. + +unconstrained data such as depth images [25, 38] or 3D meshes [26], which is beyond the scope of this paper. + +# 3 3D human shape reconstruction from monocular unconstrained images + +We introduce a novel framework that reconstructs COcomplete 3D human Shape from Monocular Unconstrained images (COSMU). To be complete, a 3D human shape should contain high-quality details in every body region. A single image lacks the necessary information to reproduce details across the entire body, which can instead be extracted from multi-view images. In contrast to multiview reconstruction, which requires calibrated and registered images captured with inaccessible systems, COSMU, given a single target image, reproduces fine details in the regions of the 3D shape that are occluded in the input image. This is achieved by leveraging a collection of uncalibrated and unregistered monocular images, referred to as unconstrained images. The input is a target image $I_{t}$ of a person together with multiple views of the same person in various poses and either same or different clothing, from which we extract body part detail non-visible in the target view. As shown in Fig. 2, given a set $\mathbf{U} = \{I_u^m\}_{m=1}^{N_u}$ of $N_{u}$ monocular unconstrained images, COSMU simulates the multi-view scenario of traditional multi-view works using a novel module that employs body part-based reference selection and registration to create 2D multi-view normal maps registered with the target input. A multi-view attention-based implicit model regresses the implicit representation $f_{f}$ of the complete 3D shape from the normals. + +# 3.1 Multi-view scenario simulation module + +Directly processing unconstrained images with a multi-view neural model deteriorates the performance due to their unregistered nature (Sec. 4.1). To address this, we propose a novel pipeline that simulates a multi-view scenario in which the multi-view images represent various orientations of the subject in the same body pose. Given a target image $I_{t}$ , we select $N_{p} \times N_{r}$ references $\left\{\{I_{ref}^{(p,r)}\}_{p=1}^{N_{p}}\right\}_{r=1}^{N_{r}}$ , $I_{ref} \in \mathbf{U}$ , from the collection $\mathbf{U}$ of unconstrained images applying a body part-based reference selection. 2D registered multi-view normal maps are then created by registering references to the target input via body part-based registration. Set notation is dropped for simplicity. + +Body part-based reference selection: To improve the body part-based registration, we select references observed from complementary orientations to the target one and with body part poses similar to the corresponding target body part ones. $N_{r}$ references are selected for each $N_{p}$ body part of the target $I_{t}$ via energy optimization following an orientation-visibility-similarity criteria. Semantic masks $M$ and 3D poses $P$ are extracted from the input images. To ensure that the entire body is represented in the selected references, we first divide the $N_{u}$ unconstrained images into $N_{r}$ groups based on the full-body orientation angle $\theta_{u}$ computed between the reference $P_{u}$ and the target $P_{t}$ 3D poses with the Rodrigues rotation formula. Each group contains a variable number $N_{g}$ of unconstrained images $I_{u}^{g}, g \in [1,\dots,N_{g}]$ . $N_{p}$ references $I_{ref}$ are selected from each group following 3 criteria: + +1. Body orientation angle: for a complete view of the body, the selected images should complement each other in the body orientation. An orientation metric is introduced to enforce the selection of images in which the body orientation is closer to the central angle $\theta_c^r$ of the $r^{th}$ group: + +$$ +o _ {g} = \left| \theta_ {c} ^ {r} - \theta_ {u} \right|, \quad \theta_ {c} ^ {r} = r \cdot \frac {3 6 0}{N _ {r} + 1}, \quad r \in [ 1, \dots , N _ {r} ] \tag {1} +$$ + +2. Body part visibility: every body part should be visible in the selected references to have a complete view of the body. We use a visibility metric that counts the total number of pixels $N_{M}$ of each body part $p$ in the semantic masks $M$ . + +$$ +v _ {g} ^ {p} = N _ {M} ^ {p}, \quad p \in [ 1, \dots , N _ {p} ] \tag {2} +$$ + +3. Body part pose similarity: similarity between each body part of $I_{u}^{g}$ and the corresponding body part of $I_{t}$ is evaluated as 3D Euclidean distance between the corresponding 3D bones of $P_{g}$ and $P_{t}$ , ensuring spatial consistency: + +$$ +b _ {g} ^ {p} = \left\| P _ {t} ^ {p} - P _ {g} ^ {p} \right\|, \quad p \in [ 1, \dots , N _ {p} ] \tag {3} +$$ + +$o_{g} = o(I_{u}^{g}), v_{g}^{p} = v^{p}(I_{u}^{g}), b_{g}^{p} = b^{p}(I_{u}^{g})$ for the $g^{th}$ image $I_{u}^{g}$ of the $p^{th}$ body part. $N_{p}$ references $I_{ref}$ are selected by minimizing the energy $E_{g}^{p}$ over the $N_{g}$ images + +$I_u^g$ for each body part $p$ : + +$$ +I _ {r e f} = \underset {g} {\operatorname {a r g m i n}} \left(E _ {g} ^ {p}\right) \tag {4} +$$ + +$$ +E _ {g} ^ {p} = i d x _ {m i n} \left(o _ {g}\right) + i d x _ {m a x} \left(v _ {g} ^ {p}\right) + i d x _ {m i n} \left(b _ {g} ^ {p}\right) +$$ + +$idx_{min}$ indicates the position $idx$ of the $g^{th}$ image $I_u^g$ in the arrays created by sorting all the $N_g$ images with respect to their specific metric $K = o_g, v_g^p, b_g^p$ . $idx_{max}$ sorts the array in inverse order: + +$$ +i d x _ {m i n} (K) = \mathrm {i d x} _ {I _ {u} ^ {g}} \left\{I _ {u} ^ {1}, \dots , I _ {u} ^ {N _ {g}} \right\}, K \left(I _ {u} ^ {1}\right) \leqslant \dots \leqslant K \left(I _ {u} ^ {N _ {g}}\right) \tag {5} +$$ + +$$ +i d x _ {m a x} (K) = \mathrm {i d x} _ {I _ {u} ^ {g}} \left\{I _ {u} ^ {1}, \dots , I _ {u} ^ {N _ {g}} \right\}, K \left(I _ {u} ^ {1}\right) \geqslant \dots \geqslant K \left(I _ {u} ^ {N _ {g}}\right) +$$ + +$N_{p} \times N_{r}$ references are selected by repeating the body part-based reference selection for every group. + +Body part-based registration: To simulate the multi-view scenario given the selected $N_{p} \times N_{r}$ references, we register each 3D reference body part to the corresponding 3D body part in $I_{t}$ . This process begins with the reconstruction of a coarse point cloud $\zeta_{t}$ for $I_{t}$ and for each of the $N_{p}$ reference images $I_{ref}$ previously selected for the $N_{r}$ groups $(\zeta_{ref})$ . We use a neural implicit model $f_{c}$ comprising of a U-Net that increases the resolution of the input features $\phi(I)$ as in SuRS [37]. The coarse implicit function is estimated as: + +$$ +\hat {s} _ {c} = f _ {c} \left(\phi (\pi (I)), z (x)\right), \quad \hat {s} _ {c} \in \mathbb {R} \tag {6} +$$ + +where $I$ is either $I_{t}$ or $I_{ref}$ , $\pi$ represents the orthographic projection of 3D points $x$ into the input images $I$ and $z(x)$ is the depth of $x$ . $f_{c}$ is trained with: + +$$ +\mathcal {L} _ {c} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left| f _ {c} \left(X _ {c} ^ {i}\right) - f _ {c} ^ {g t} \left(X _ {c} ^ {i}\right) \right| ^ {2} + | I - I _ {H R} | \tag {7} +$$ + +where $X_{c}^{i} = (\phi^{i}(\pi (I)),z(x))$ , $N$ is the number of points in space and $I_{HR}$ is the up-sampled ground-truth version of $I$ used to train the U-Net feature extractor. We then retrieve 3D body parts by classifying each 3D point $x_{\zeta}$ of the point clouds $\zeta_{t / ref}$ with a novel pixel-aligned classification network $f_{\zeta}$ . An image feature embedding $\phi (I||M)$ is extracted from the concatenation between the input RGB image and its semantic mask $M$ . Each $x_{\zeta}$ is projected using orthographic projection $\pi$ onto the input $I||M$ and indexed to $\phi (I||M)$ . An MLP processes this embedding to estimate the body part $\hat{s}_{\zeta}$ to which $x_{\zeta}$ belongs: + +$$ +\hat {s} _ {\zeta} = f _ {\zeta} \left(\phi \left(\pi (I | | M)\right), z \left(x _ {\zeta}\right)\right), \quad \hat {s} _ {\zeta} \in [ 1, \dots , N _ {p} ] \tag {8} +$$ + +$N_{p}$ 3D body parts $v_{ref}$ are defined for each reference point cloud as well as for the target $v_{t}$ . Next, the specific $p^{th}$ body part $v_{ref}$ selected for the corresponding $p^{th}$ target body part $v_{t}$ is registered with it as $v_{reg} = R(v_{ref}) + T$ , where $R$ is the rotation and $T$ the translation computed via rigid registration. For each + +group, the registered $N_{p}$ body parts are combined to create a point cloud $\zeta_{f}$ . $N_{r}$ point clouds $\zeta_{f}$ registered with $\zeta_{t}$ are generated, rotated by their $\theta_{u}$ and, together with $\zeta_{t}$ , used to retrieve $N_{r} + 1$ 2D normal maps $S_{v}$ , one for each view $v \in [1, \dots, N_{r} + 1]$ . Normal maps are generated instead of RGB images to avoid the estimation of color values for each 3D point of $\zeta_{t / ref}$ and to allow different clothing in the references. Directly fusing the registered point clouds to obtain the 3D shape is challenging since correspondences are unknown. We train $f_{\zeta}$ with a Mean Squared Error $MSE$ loss: + +$$ +\mathcal {L} _ {\text {p a r t}} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left| f _ {\zeta} \left(X _ {\zeta} ^ {i}\right) - f _ {\zeta} ^ {g t} \left(X _ {\zeta} ^ {i}\right) \right| ^ {2} \tag {9} +$$ + +where $X_{\zeta}^{i} = \left(\phi^{i}(\pi (I||M)),z(x_{\zeta})\right)$ + +# 3.2 Multi-view implicit representation + +Neural implicit models have proven to be effective in reproducing high-quality details in reconstructed 3D shapes. This is attributed to the characteristic of the implicit function that does not discretize the 3D space like other representations [40]. Leveraging multiple views to learn the implicit representation further enhances the quality of the reconstruction, as it enables the incorporation of information about the entire body. Concatenating features from multiple images is a suboptimal approach because it imposes an order on the views [59]. Fusion approaches based on statistics have been proposed [12,44] but their pooling operation causes a loss of information regarding the individual view contributions. Inspired by Zins et al. [59], we propose to fuse the features extracted from the multi-view 2D normals using an attention-based architecture. An attention score is computed for each view $v$ based on the compatibility of a query with a corresponding key, resulting in the generation of $N_r + 1$ features. Each feature contains the original information from the corresponding view, merged with the information from all the other views. An MLP then models a pixel-aligned implicit function $f_f$ to learn the occupancy value $\hat{s}_f$ of the 3D points in space: + +$$ +\hat {s} _ {f} = f _ {f} \left(A \left(\phi_ {v} \left(\pi \left(S _ {v}\right)\right), z (x)\right)\right), \quad \hat {s} _ {f} \in \mathbb {R} \tag {10} +$$ + +where $\phi_v$ are the feature extracted from each view $v$ of the surface normal $S_v$ , $\pi$ indicates the pixel-alignment between the 3D points of the space and each $\phi_v$ , $A$ is the attention operation and $z(x)$ is the depth value of the 3D points $x$ . We train the multi-view architecture with an MSE loss: + +$$ +\mathcal {L} _ {f} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left| f _ {f} \left(X _ {f} ^ {i}\right) - f _ {f} ^ {g t} \left(X _ {f} ^ {i}\right) \right| ^ {2} \tag {11} +$$ + +where $X_{f}^{i} = A\left(\phi_{v}^{i}(\pi (S_{v})),z(x^{i})\right)$ + +Inference: $N_{p} \times N_{r}$ references are selected from the collection of unconstrained + +![](images/f13704363cab77fc46968819c620237595c0bb84d3a5147d1bba4f1902bc1eaa.jpg) +Fig. 3: Qualitative comparisons between 3D shapes from THuman3.0 [45] reconstructed by changing the energy optimization process to select reference images. Front, back and side views of the shapes are illustrated. + +images. $N_r + 1$ 2D normal maps are generated using body part-based registration to simulate a multi-view scenario with registered views of the subject. The normals are processed by the multi-view attention-based model to estimate the implicit representation. The 3D mesh is reconstructed via marching cube [32]. To achieve pixel-alignment through projection, multi-view works require the parameters of calibrated cameras in specialized capture systems. In COSMU, instead of using the camera projection matrix for pixel-alignment, 3D points of the space are rotated by $\theta_u$ computed by the multi-view scenario simulation module and projected onto the images using an identity matrix. Existing approaches create the inference 3D grid centered at the origin. If the subject is not centered in the input images, the multi-view pixel-alignment of COSMU is inaccurate since the same rotated 3D point may not be projected in the corresponding pixel in each view. To tackle this, we set the boundary limits of the 3D grid as the boundaries of the coarse point cloud $\zeta_t$ , ensuring multi-view pixel-alignment. + +# 4 Experiments + +We conduct a comprehensive evaluation of COSMU for the task of 3D human shape reconstruction from a target image by leveraging a collection of monocular unconstrained images of the same individual. The neural models of the framework are trained with THuman2.0 dataset [54], which comprises 524 high-resolution surfaces. We render 360-degree view images around each surface, which serves as input data. For the quantitative and qualitative evaluation, we use the THuman3.0 dataset [45], which consists of 20 individuals each captured in various poses (400 models), and the testing set of X-Humans [42] (41 models). For each pose, we render 360-degree views of the subject and we apply the reference selection to select the $N_{p} \times N_{r}$ references among the images of the subject in different poses, excluding the target one. We evaluate the reconstruction accuracy with three quantitative metrics: the normal reprojection error [40], the average point-to-surface Euclidean distance (P2S) and the Chamfer distance + +(CD), expressed in cm. We further test COSMU with real data collected from various scenarios, including monocular videos (People-Snapshot [3], iPER [29] datasets) or images of people in different poses and clothing from the Internet, demonstrating its superiority over related works to reconstruct complete 3D human shapes with higher quality details in regions of the body occluded in the input target image (Sec. 4.2). In contrast to prior works, no parametric models are used. We set $N_{p} = 10$ and $N_{r} = 3$ . See supplementary for additional implementation details. + +# 4.1 Ablation Studies + +Energy function for reference selection: We conduct an ablation study on the criteria used for selecting the $N_{p} \times N_{r}$ references in the body part-based reference selection. The variants explored are: (i) Random: no criteria are applied. $N_{p}$ references are randomly selected for each group; (ii) Only $\mathbf{b_p}$ : only the body part pose similarity is considered; (iii) Only $\mathbf{o}$ : the body orientation angle is the only criterion used; (iv) Only $\mathbf{v_p}$ : only the body part visibility is considered; (v) $\mathbf{o} + \mathbf{v_p}$ : the body part pose similarity is excluded; (vi) $\mathbf{o} + \mathbf{b_p}$ : the body part visibility is omitted; (vii) $\mathbf{v_p} + \mathbf{b_p}$ : the orientation angle is not used. + +As illustrated in Fig. 3 and supported by the quantitative results of Tab. 1, the 3D shapes obtained with COSMU exhibit fewer artifacts and finer details. The reference selection plays a crucial role in selecting the most suitable references to facilitate the body part-based registration. Omitting any of the criteria in the energy optimization influences the selection of references, leading to suboptimal choices and, consequently, deteriorating the body part-based registration accuracy and introducing artifacts in the reconstruction. + +Table 1: Quantitative evaluation of the application of different energy optimization processes to select reference images from THuman3.0 [45]. + +
RandomOnly bpOnly oOnly vpo + vpo + bpvp + bpCOSMU
CD1.4851.4411.4211.4101.3461.3521.3261.304
Normal.1683.1598.1641.1654.1523.1529.1503.1475
P2S1.4241.4151.4031.4061.3321.2791.3081.279
+ +Body part-based registration effectiveness: Directly using $N_{r}$ selected references for the multi-view reconstruction deteriorates the performance since fusing different poses is challenging in the absence of known correspondences. To avoid this, we introduce a novel body part-based registration method to simulate a multi-view scenario. To prove its effectiveness, we modify the algorithm for registering the $N_{p} \times N_{r}$ coarse reference point clouds $\zeta_{ref}$ to the target $\zeta_{t}$ : + +- No Reg: To demonstrate the effectiveness of registering the reference point clouds to the target, we omit the body part-based registration step. We adjust the reference selection to select only $N_r$ references by averaging over $p$ the criteria $v_g^p$ and $b_g^p$ in Eq. (4). + +Table 2: Quantitative evaluation of the application of different configurations of the body part-based registration algorithm on THuman3.0 [45]. + +
No RegNon rigidAverageNo RGBNo SegmCOSMU
CD3.5466.9321.3142.2492.1891.304
Normal.2280.4832.1569.2351.2114.1475
P2S2.8206.0241.3092.1692.0051.279
+ +![](images/d902fa8dc4d22485f0c1bc63bfd67d37853198d640d6d6361026aa401cc9753a.jpg) +Fig. 4: Qualitative comparisons between 3D shapes from THuman3.0 [45] reconstructed from 2D normal maps generated using different configurations of the body part-based registration algorithm. Front, back and side views of the shapes are shown. + +- Non Rigid: Instead of applying rigid registration among the correspondent body parts of the target and reference point clouds, we apply the state-of-the-art work by Li and Harada [24] to perform non-rigid registration between the full-body point clouds. We use $N_r$ references as in No reg. +- Average: COSMU registers a target body part with the corresponding body part of a reference selected specifically for that body part. In this configuration, we select only $N_r$ references as in No reg and register all the body parts of the single reference for each group. +- No RGB: The body part classification model $f_{\zeta}$ processes a concatenation of 2D RGB images and segmentation masks. To demonstrate the advantages of using these data, we train and test $f_{\zeta}$ with only segmentation masks. +- No Segm: we train and test $f_{\zeta}$ with only RGB images. + +We apply these configurations to generate 2D multi-view normal maps used by the multi-view attention-based implicit model to reconstruct 3D human shapes. Quantitative (Tab. 2) and qualitative results (Fig. 4) show that COSMU configuration is the most effective. Multi-view reconstruction requires registered input since pixel-alignment across different views can be incorrect with unregistered views. Estimating correspondences between point clouds is challenging and the non-rigid registration cannot produce reliable results. Selecting $N_{p}$ references for each body part facilitates the body part registration compared to using a sin + +Table 3: Quantitative comparisons with related works for 3D human reconstruction. + +
MethodsTHuman3.0X-Humans
CDNormalP2SCDNormalP2S
1 viewPiFU [40]1.698.17851.6592.516.25912.492
SuRS [37]1.573.16931.5542.406.24872.385
Baseline1.566.17321.5702.375.24862.351
2 viewsPiFu-HD [41]1.501.16631.5122.139.22762.121
ICON [50]1.474.16951.4632.210.23952.208
ECON [49]1.368.16321.3652.153.23052.114
COSMU-21.355.15931.3732.139.22582.104
4 viewsMulti-PaMIR [56]1.401.16501.3852.145.23192.132
MVP [58]1.348.15581.3652.063.22041.972
COSMU1.304.14751.2792.006.21461.929
+ +gle reference. Leveraging RGB and segmentation masks improves the body part classification since the model learns both semantic and appearance information. + +# 4.2 Comparisons + +We conduct a comparative evaluation of COSMU against related works that reconstruct complete 3D human shapes from monocular RGB images. We evaluate MVP-Human [58] (MVP) and Multi-Pamir [56] that reconstruct 3D shapes from unconstrained images leveraging SMPL [31] in the process. The same $N_{r} = 3$ images selected in the ablation study (Sec. 4.1) are used as additional views. We evaluate methods that estimate an unseen view of the subject from the target image (PiFU-HD [41], ICON [50] and ECON [49]). We then compare COSMU with PiFU [40] and SuRS [37], which process a single view without any information on hidden regions. To demonstrate the effect of using a lower number of references, we compare COSMU with COSMU-2, which is trained and evaluated by selecting $N_{r} = 1$ reference ( $\theta_{c} = 180^{\circ}$ ). Finally, we evaluate $f_{c}$ (Baseline) that reconstructs the coarse target point clouds $\zeta_{t}$ without leveraging additional views. We use the released weights of ECON and PiFU-HD due to the absence of the training code while we train the other approaches with THuman2.0 [54]. Quantitative evaluation is conducted with the 3D shapes reconstructed from single images of THuman3.0 [45] and X-Humans [42]. The results presented in Tab. 3 demonstrate the superiority of COSMU across all the considered metrics for both datasets. The best scores overall are highlighted in red while the best scores for approaches that use 2 views are blue. COSMU achieves the highest accuracy while COSMU-2 outperforms approaches that rely on 2 views, demonstrating the benefit of leveraging views from 2D images rather than estimating them from the input. The worst results are obtained with single-view approaches. Qualitative results: Fig. 5 illustrates the 3D shapes reconstructed with the considered methods across several scenarios. The first set of shapes is reconstructed by leveraging a collection of images of the same individual in different poses from THuman3.0 [45]. The target and references of the second set are taken from the same monocular video of a person rotating from People-Snapshot [3]. + +In the third set, random images of a person in different poses and with a combination of both identical and different clothing are selected from the Internet. The fourth set of unconstrained images is created with images of the individual in different poses and clothing. This scenario is the most challenging since there are no images of the person in identical clothing. In all these scenarios, COSMU consistently reproduces higher-quality details in the regions of the 3D shape occluded in the target image, significantly outperforming related works. In addition, the quality of regions visible in the input image is comparable to that of related works. MVP struggles to reproduce details and introduces noise when the subject is wearing different clothing in the references ( $3^{rd}$ and $4^{th}$ sets). ECON and PIFu-HD represent details on unseen parts of the model at a significantly lower quality compared to COSMU, which is the only approach that reconstructs the hood on the jumper ( $1^{st}$ set). The shapes reconstructed with PIFu-HD and ECON exhibit excessive smoothing in the regions of the human body occluded in the target image. The shapes reconstructed by ECON and MVP are influenced by the incorrect estimation and fitting of the SMPL model, resulting in an unnatural bending of the 3D body. SuRS reproduces high-quality details only in visible regions, with excessive smoothed reconstruction in other regions. See supplementary materials for additional results. + +Limitations: COSMU's ability to reconstruct high-quality details is contingent on the availability of reference images. COSMU cannot reproduce high-quality details in a region that is not visible in the reference images. Performance deteriorates when the reference images significantly differ from the target. See supplementary for a study on choosing references with varying degrees of similarity to the target. In some examples, COSMU cannot reproduce the same high-quality level of details in small regions of the body, such as the face and hands. The registration algorithm struggles to register small body parts while the multi-view transformer struggles to fuse multiple views of those regions. + +# 5 Conclusion + +We introduce a novel framework to reconstruct complete 3D human shapes from a target RGB image by leveraging a collection of monocular unconstrained images without using parametric models. We first developed a novel multi-view scenario simulation module to generate a set of registered 2D normal maps of the target subject. A multi-view attention-based neural implicit model then estimates an implicit function that represents the subject 3D shape. The evaluation demonstrates that the proposed approach can be applied in various scenarios. Significantly higher quality details are reproduced in non-visible regions of the 3D shape compared to state-of-the-art works. In future works, a solution to enhance the quality of smaller body regions will be investigated and the approach will be extended to reconstruct temporally-coherent dynamic shapes. + +Acknowledgement: Work supported by BBC Prosperity Partnership AI4ME: Future Personalised Object-Based Media Experiences Delivered at Scale Anywhere EP/V038087. + +![](images/eaccd5ddcb08d29634a2442ef5c5d517236a62d04476943f55f261d8b5d04d48.jpg) +Fig. 5: Qualitative examples of 3D shapes reconstructed in several scenarios. Static images are from THuman3.0 [45] in the first example and from the Internet in the others; monocular videos are from People-Snapshot [3]. Front, back and side views are shown. + +# References + +1. Alldieck, T., Magnor, M., Bhatnagar, B.L., Theobalt, C., Pons-Moll, G.: Learning to reconstruct people in clothing from a single rgb camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1175-1186 (2019) +2. Alldieck, T., Magnor, M., Xu, W., Theobalt, C., Pons-Moll, G.: Detailed human avatars from monocular video. In: 2018 International Conference on 3D Vision (3DV). pp. 98-109. IEEE (2018) +3. Alldieck, T., Magnor, M., Xu, W., Theobalt, C., Pons-Moll, G.: Video based reconstruction of 3d people models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8387-8397 (2018) +4. Alldieck, T., Pons-Moll, G., Theobalt, C., Magnor, M.: Tex2shape: Detailed full human body geometry from a single image. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2293-2303 (2019) +5. Alldieck, T., Zanfir, M., Sminchisescu, C.: Photorealistic monocular 3D reconstruction of humans wearing clothing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1506-1515 (2022) +6. Bogo, F., Kanazawa, A., Lassner, C., Gehler, P., Romero, J., Black, M.J.: Keep it SMPL: Automatic estimation of 3d human pose and shape from a single image. In: ECCV (2016) +7. Cao, Y., Han, K., Wong, K.Y.K.: Sesdf: Self-evolved signed distance field for implicit 3d clothed human reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4647-4657 (2023) +8. Chen, J., Zhang, Y., Kang, D., Zhe, X., Bao, L., Jia, X., Lu, H.: Animatable neural radiance fields from monocular rgb videos. arXiv preprint arXiv:2106.13629 (2021) +9. Chen, L., Li, J., Huang, H., Guo, Y.: Crosshuman: learning cross-guidance from multi-frame images for human reconstruction. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 2483-2494 (2022) +10. Chen, Z., Zhang, H.: Learning implicit fields for generative shape modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5939-5948 (2019) +11. Corona, E., Zanfir, M., Alldieck, T., Bazavan, E.G., Zanfir, A., Sminchisescu, C.: Structured 3d features for reconstructing controllable avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16954-16964 (2023) +12. Gardner, A., Elhami, N., Selmic, R.R.: Classifying unordered feature sets with convolutional deep averaging networks. In: 2019 IEEE International Conference on Systems, Man and Cybernetics (SMC). pp. 3447-3453. IEEE (2019) +13. Guo, C., Jiang, T., Chen, X., Song, J., Hilliges, O.: Vid2 avatar: 3d avatar reconstruction from videos in the wild via self-supervised scene decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12858-12868 (2023) +14. He, T., Collomosse, J., Jin, H., Soatto, S.: Geo-pifu: Geometry and pixel aligned implicit functions for single-view human reconstruction (2020) +15. He, T., Xu, Y., Saito, S., Soatto, S., Tung, T.: Arch++: Animation-ready clothed human reconstruction revisited. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11046-11056 (2021) +16. Huang, H., Chen, L., Wang, X.: Unconfuse: Avatar reconstruction from unconstrained images. In: European Conference on Computer Vision. pp. 376-381. Springer (2022) + +17. Huang, Y., Yi, H., Xiu, Y., Liao, T., Tang, J., Cai, D., Thies, J.: Tech: Text-guided reconstruction of lifelike clothed humans. arXiv preprint arXiv:2308.08545 (2023) +18. Huang, Z., Xu, Y., Lassner, C., Li, H., Tung, T.: Arch: Animatable reconstruction of clothed humans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3093-3102 (2020) +19. Jayasundara, V., Agrawal, A., Heron, N., Shrivastava, A., Davis, L.S.: Flexnerf: Photorealistic free-viewpoint rendering of moving humans from sparse views. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21118-21127 (2023) +20. Jiang, B., Hong, Y., Bao, H., Zhang, J.: Selfrecon: Self reconstruction your digital avatar from monocular video. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +21. Kanazawa, A., Black, M.J., Jacobs, D.W., Malik, J.: End-to-end recovery of human shape and pose. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7122-7131 (2018) +22. Kocabas, M., Athanasiou, N., Black, M.J.: Vibe: Video inference for human body pose and shape estimation. In: CVPR (2020) +23. Li, C., Lin, J., Lee, G.H.: Ghunerf: Generalizable human nerf from a monocular video. arXiv preprint arXiv:2308.16576 (2023) +24. Li, Y., Harada, T.: Non-rigid point cloud registration with neural deformation pyramid. arXiv preprint arXiv:2205.12796 (2022) +25. Li, Z., Yu, T., Zheng, Z., Guo, K., Liu, Y.: Posefusion: Pose-guided selective fusion for single-view human volumetric capture. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14162-14172 (2021) +26. Li, Z., Zheng, Z., Zhang, H., Ji, C., Liu, Y.: Avatarcap: Animatable avatar conditioned monocular human volumetric capture. In: European Conference on Computer Vision. pp. 322-341. Springer (2022) +27. Li, Z., Oskarsson, M., Heyden, A.: Detailed 3d human body reconstruction from multi-view images combining voxel super-resolution and learned implicit representation. Applied Intelligence 52(6), 6739-6759 (2022) +28. Liao, T., Zhang, X., Xiu, Y., Yi, H., Liu, X., Qi, G.J., Zhang, Y., Wang, X., Zhu, X., Lei, Z.: High-fidelity clothed avatar reconstruction from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8662-8672 (2023) +29. Liu, W., Piao, Z., Min, J., Luo, W., Ma, L., Gao, S.: Liquid warping gan: A unified framework for human motion imitation, appearance transfer and novel view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5904-5913 (2019) +30. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM TOG 34(6), 248 (2015) +31. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia) 34(6), 248:1-248:16 (Oct 2015) +32. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics 21(4), 163-169 (1987) +33. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4460-4470 (2019) + +34. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 165-174 (2019) +35. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10975-10985 (2019) +36. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: CVPR (2021) +37. Pesavento, M., Volino, M., Hilton, A.: Super-resolution 3d human shape from a single low-resolution image. In: ECCV (2022) +38. Pesavento, M., Xu, Y., Sarafianos, N., Maier, R., Wang, Z., Yao, C.H., Volino, M., Boyer, E., Hilton, A., Tung, T.: Anim: Accurate neural implicit model for human reconstruction from a single rgb-d image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5448-5458 (2024) +39. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022) +40. Saito, S., Huang, Z., Natsume, R., Morishima, S., Kanazawa, A., Li, H.: Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2304-2314 (2019) +41. Saito, S., Simon, T., Saragih, J., Joo, H.: Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 84-93 (2020) +42. Shen, K., Guo, C., Kaufmann, M., Zarate, J.J., Valentin, J., Song, J., Hilliges, O.: X-ayar: Expressive human avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16911-16921 (2023) +43. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems 34, 6087-6101 (2021) +44. Su, H., Maji, S., Kalogerakis, E., Learned-Miller, E.: Multi-view convolutional neural networks for 3d shape recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 945-953 (2015) +45. Su, Z., Yu, T., Wang, Y., Liu, Y.: Deepcloth: Neural garment representation for shape and style editing. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(2), 1581-1593 (2023). https://doi.org/10.1109/TPAMI.2022.3168569 +46. Te, G., Li, X., Li, X., Wang, J., Hu, W., Lu, Y.: Neural capture of animatable 3d human from monocular video. In: European Conference on Computer Vision. pp. 275-291. Springer (2022) +47. Varol, G., Ceylan, D., Russell, B., Yang, J., Yumer, E., Laptev, I., Schmid, C.: Bodynet: Volumetric inference of 3d human body shapes. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 20-36 (2018) +48. Wang, S., Schwarz, K., Geiger, A., Tang, S.: Arah: Animatable volume rendering of articulated human sdfs. In: European conference on computer vision. pp. 1-19. Springer (2022) +49. Xiu, Y., Yang, J., Cao, X., Tzionas, D., Black, M.J.: Econ: Explicit clothed humans optimized via normal integration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 512-523 (2023) + +50. Xiu, Y., Yang, J., Tzionas, D., Black, M.J.: Icon: Implicit clothed humans obtained from normals. In: CVPR (2022) +51. Xu, X., Loy, C.C.: 3d human texture estimation from a single image with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13849-13858 (2021) +52. Yang, X., Chen, X., Wang, S., Gao, D., Han, X., Wang, B.: Have-fun: Human avatar reconstruction from few-shot unconstrained images. arXiv preprint arXiv:2311.15672 (2023) +53. Yu, H., Cheang, C., Fu, Y., Xue, X.: Multi-view shape generation for a 3d human-like body. ACM Transactions on Multimedia Computing, Communications and Applications 19(1), 1-22 (2023) +54. Yu, T., Zheng, Z., Guo, K., Liu, P., Dai, Q., Liu, Y.: Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR2021) (June 2021) +55. Yu, Z., Zhang, L., Xu, Y., Tang, C., Tran, L., Keskin, C., Park, H.S.: Multiview human body reconstruction from uncalibrated cameras. Advances in Neural Information Processing Systems 35, 7879-7891 (2022) +56. Zheng, Z., Yu, T., Liu, Y., Dai, Q.: Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021) +57. Zheng, Z., Yu, T., Wei, Y., Dai, Q., Liu, Y.: Deephuman: 3d human reconstruction from a single image. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7739-7749 (2019) +58. Zhu, X., Liao, T., Lyu, J., Yan, X., Wang, Y., Guo, K., Cao, Q., Li, S.Z., Lei, Z.: Mvp-human dataset for 3d human avatar reconstruction from unconstrained frames. arXiv preprint arXiv:2204.11184 (2022) +59. Zins, P., Xu, Y., Boyer, E., Wuhrer, S., Tung, T.: Data-driven 3d reconstruction of dressed humans from sparse views. In: 2021 International Conference on 3D Vision (3DV). pp. 494-504. IEEE (2021) +60. Zins, P., Xu, Y., Boyer, E., Wuhrer, S., Tung, T.: Multi-view reconstruction using signed ray distance functions (srdf). In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16696-16706 (2023) \ No newline at end of file diff --git a/2024/COSMU_ Complete 3D human shape from monocular unconstrained images/images.zip b/2024/COSMU_ Complete 3D human shape from monocular unconstrained images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..63cfc8661141d864f291a2ad7839bf1d33faaf7f --- /dev/null +++ b/2024/COSMU_ Complete 3D human shape from monocular unconstrained images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f8553738c44e3001e47c435273b53a0d6fa0b9e757b57f44e0ffd319d8b53dd +size 633995 diff --git a/2024/COSMU_ Complete 3D human shape from monocular unconstrained images/layout.json b/2024/COSMU_ Complete 3D human shape from monocular unconstrained images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a81343ddb1f6243d44f16cb9c60be21bb04122de --- /dev/null +++ b/2024/COSMU_ Complete 3D human shape from monocular unconstrained images/layout.json @@ -0,0 +1,10038 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 29, + 353, + 66 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 29, + 353, + 66 + ], + "spans": [ + { + "bbox": [ + 59, + 29, + 353, + 66 + ], + "type": "text", + "content": "COSMU: Complete 3D human shape from monocular unconstrained images" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 83, + 86, + 329, + 98 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 86, + 329, + 98 + ], + "spans": [ + { + "bbox": [ + 83, + 86, + 329, + 98 + ], + "type": "text", + "content": "Marco Pesavento, Marco Volino, and Adrian Hilton" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 108, + 368, + 130 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 108, + 368, + 130 + ], + "spans": [ + { + "bbox": [ + 45, + 108, + 368, + 130 + ], + "type": "text", + "content": "Centre for Vision, Speech and Signal Processing (CVSSP), University of Surrey, Guildford, UK {m.pesavento,m.volino,a.hilton} " + }, + { + "bbox": [ + 45, + 108, + 368, + 130 + ], + "type": "inline_equation", + "content": "@" + }, + { + "bbox": [ + 45, + 108, + 368, + 130 + ], + "type": "text", + "content": " surrey.ac.uk" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 157, + 351, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 157, + 351, + 420 + ], + "spans": [ + { + "bbox": [ + 61, + 157, + 351, + 420 + ], + "type": "text", + "content": "Abstract. We present a novel framework to reconstruct complete 3D human shapes from a given target image by leveraging monocular unconstrained images. The objective of this work is to reproduce high-quality details in regions of the reconstructed human body that are not visible in the input target. The proposed methodology addresses the limitations of existing approaches for reconstructing 3D human shapes from a single image, which cannot reproduce shape details in occluded body regions. The missing information of the monocular input can be recovered by using multiple views captured from multiple cameras. However, multi-view reconstruction methods necessitate accurately calibrated and registered images, which can be challenging to obtain in real-world scenarios. Given a target RGB image and a collection of multiple uncalibrated and unregistered images of the same individual, acquired using a single camera, we propose a novel framework to generate complete 3D human shapes. We introduce a novel module to generate 2D multi-view normal maps of the person registered with the target input image. The module consists of body part-based reference selection and body part-based registration. The generated 2D normal maps are then processed by a multi-view attention-based neural implicit model that estimates an implicit representation of the 3D shape, ensuring the reproduction of details in both observed and occluded regions. Extensive experiments demonstrate that the proposed approach estimates higher quality details in the non-visible regions of the 3D clothed human shapes compared to related methods, without using parametric models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 34, + 439, + 127, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 34, + 439, + 127, + 452 + ], + "spans": [ + { + "bbox": [ + 34, + 439, + 127, + 452 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 464, + 379, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 464, + 379, + 570 + ], + "spans": [ + { + "bbox": [ + 33, + 464, + 379, + 570 + ], + "type": "text", + "content": "Democratizing the creation of high-quality human avatars is essential to unlock a wide range of immersive technologies including self-representation in the metaverse, virtual conferencing, and virtual try-on. Realistic human avatars are crucial in these applications, making 3D human reconstruction a prominent topic in computer vision and graphics. Research focuses on finding practical solutions to replace existing sophisticated multi-view systems for 3D reconstruction, which are inaccessible to the general community. Single-view 3D human reconstruction methods have been developed to reconstruct 3D human shapes using a simple and cost-effective set-up, such as a single camera." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 571, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 571, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 571, + 379, + 583 + ], + "type": "text", + "content": "Recent methods [5, 11, 14, 15, 18, 40] use neural implicit models to estimate 3D" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 413, + 615 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 42, + 31, + 373, + 182 + ], + "blocks": [ + { + "bbox": [ + 42, + 31, + 373, + 182 + ], + "lines": [ + { + "bbox": [ + 42, + 31, + 373, + 182 + ], + "spans": [ + { + "bbox": [ + 42, + 31, + 373, + 182 + ], + "type": "image", + "image_path": "f2389ca12bc43c05236a13c98905f15d86cdd3ea520c80c9206abf9205fb33e2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 31, + 190, + 380, + 224 + ], + "lines": [ + { + "bbox": [ + 31, + 190, + 380, + 224 + ], + "spans": [ + { + "bbox": [ + 31, + 190, + 380, + 224 + ], + "type": "text", + "content": "Fig. 1: COSMU reconstructs complete 3D human shape from a target RGB input by leveraging a set of monocular images collected from various scenarios, including monocular video (a), images with different poses (b) and different clothing (c)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 30, + 249, + 379, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 249, + 379, + 284 + ], + "spans": [ + { + "bbox": [ + 30, + 249, + 379, + 284 + ], + "type": "text", + "content": "human shapes with high-quality details in visible regions from single RGB images. These works fail to reproduce fine detail in non-visible regions of the input image, resulting in overly smooth reconstruction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 30, + 285, + 379, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 285, + 379, + 356 + ], + "spans": [ + { + "bbox": [ + 30, + 285, + 379, + 356 + ], + "type": "text", + "content": "To address this problem, solutions extend single-view to multi-view reconstruction, using multiple images of various orientations of the human [7,27,53,59,60]. Multi-view methods represent details across the entire body but require calibrated cameras and registered images, with the subject in the same pose and clothing. Systems to capture such images are inaccessible to the general community in real-world scenarios, significantly limiting reproducibility." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 356, + 379, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 356, + 379, + 452 + ], + "spans": [ + { + "bbox": [ + 30, + 356, + 379, + 452 + ], + "type": "text", + "content": "Recent works [9,16,55,56,58] reconstruct humans from a single target image by leveraging multiple monocular unconstrained images. We define images as unconstrained if camera calibration parameters are not known (uncalibrated) and the images are captured at different time instances resulting in pose and clothing variations without spatial alignment (unregistered). These methods directly fuse features from multiple images, which is a challenging task since diverse views and poses must be fused into a single coherent reconstruction [58]. They also rely on parametric models and cannot use images of the person in different clothing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 30, + 452, + 379, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 452, + 379, + 535 + ], + "spans": [ + { + "bbox": [ + 30, + 452, + 379, + 535 + ], + "type": "text", + "content": "We propose COSMU, a novel framework that reconstructs CComplete 3D human Shape from Monocular Unconstrained images. Our objective is to reconstruct a 3D clothed human shape from a single target image, focusing on reproducing high-quality details in regions of the reconstructed human body that are not visible in the input target. To achieve this, we leverage a collection of monocular unconstrained images of the same individual. COSMU does not rely on parametric models in the reconstruction process as related works." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 30, + 536, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 536, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 30, + 536, + 379, + 583 + ], + "type": "text", + "content": "The multi-view scenario used by the multi-view reconstruction approaches is simulated from a collection of unconstrained images, with the subject in different poses and clothing and without camera calibration or correspondences. We present a novel pipeline that applies an optimization-based algorithm to select," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "spans": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 10, + 143, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 10, + 143, + 19 + ], + "spans": [ + { + "bbox": [ + 66, + 10, + 143, + 19 + ], + "type": "text", + "content": "M. Pesavento et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 30, + 34, + 381, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 34, + 381, + 249 + ], + "spans": [ + { + "bbox": [ + 30, + 34, + 381, + 249 + ], + "type": "text", + "content": "from the set of unconstrained images, " + }, + { + "bbox": [ + 30, + 34, + 381, + 249 + ], + "type": "inline_equation", + "content": "N_r" + }, + { + "bbox": [ + 30, + 34, + 381, + 249 + ], + "type": "text", + "content": " references for each " + }, + { + "bbox": [ + 30, + 34, + 381, + 249 + ], + "type": "inline_equation", + "content": "N_p" + }, + { + "bbox": [ + 30, + 34, + 381, + 249 + ], + "type": "text", + "content": " body part of the target image following an orientation-visibility-similarity criteria. A neural implicit model generates a coarse point cloud " + }, + { + "bbox": [ + 30, + 34, + 381, + 249 + ], + "type": "inline_equation", + "content": "\\zeta_t" + }, + { + "bbox": [ + 30, + 34, + 381, + 249 + ], + "type": "text", + "content": " of the target and of each selected reference " + }, + { + "bbox": [ + 30, + 34, + 381, + 249 + ], + "type": "inline_equation", + "content": "\\zeta_{ref}" + }, + { + "bbox": [ + 30, + 34, + 381, + 249 + ], + "type": "text", + "content": ". The 3D points of these point clouds are classified into " + }, + { + "bbox": [ + 30, + 34, + 381, + 249 + ], + "type": "inline_equation", + "content": "N_p" + }, + { + "bbox": [ + 30, + 34, + 381, + 249 + ], + "type": "text", + "content": " body parts using a novel pixel-aligned neural network. The corresponding reference body part selected for a specific target part is registered to it. To obtain " + }, + { + "bbox": [ + 30, + 34, + 381, + 249 + ], + "type": "inline_equation", + "content": "N_r + 1" + }, + { + "bbox": [ + 30, + 34, + 381, + 249 + ], + "type": "text", + "content": " views of the subject, 2D normal maps are generated from the registered point clouds. The multi-view scenario is simulated and the 2D normal maps are processed by an attention-based implicit model that estimates the implicit function " + }, + { + "bbox": [ + 30, + 34, + 381, + 249 + ], + "type": "inline_equation", + "content": "f_f" + }, + { + "bbox": [ + 30, + 34, + 381, + 249 + ], + "type": "text", + "content": " of the complete 3D human shape. COSMU generates 2D multi-view normal maps instead of RGB images because the references are not constrained to the same clothing as the target and RGB value estimation of 3D points is avoided. Extensive experiments show COSMU's ability to reconstruct complete 3D human shapes from monocular images taken from various scenarios, including monocular videos (Fig. 1a), images with different poses and same clothing (Fig. 1b) and images with different poses and clothing (Fig. 1c). Thanks to the flexibility in creating the reference set, COSMU represents a significant advance for high-quality 3D avatar reconstruction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 31, + 250, + 252, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 250, + 252, + 261 + ], + "spans": [ + { + "bbox": [ + 31, + 250, + 252, + 261 + ], + "type": "text", + "content": "The contributions presented in this paper include:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 37, + 268, + 379, + 363 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 37, + 268, + 379, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 268, + 379, + 303 + ], + "spans": [ + { + "bbox": [ + 37, + 268, + 379, + 303 + ], + "type": "text", + "content": "- A novel framework to reconstruct complete 3D human shapes from a target RGB image by leveraging a set of monocular uncalibrated and unregistered images without the use of parametric models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 37, + 304, + 379, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 304, + 379, + 327 + ], + "spans": [ + { + "bbox": [ + 37, + 304, + 379, + 327 + ], + "type": "text", + "content": "- A novel pipeline to simulate a multi-view scenario from a collection of unconstrained reference images via reference selection and body part registration." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 37, + 328, + 379, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 328, + 379, + 363 + ], + "spans": [ + { + "bbox": [ + 37, + 328, + 379, + 363 + ], + "type": "text", + "content": "- Improved performance for human shape reconstruction compared to state-of-the-art methods with higher quality details in the regions of the reconstructed human shape that are occluded in the target image." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 31, + 380, + 142, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 380, + 142, + 393 + ], + "spans": [ + { + "bbox": [ + 31, + 380, + 142, + 393 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 30, + 404, + 381, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 404, + 381, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 404, + 381, + 584 + ], + "type": "text", + "content": "3D human reconstruction from monocular cameras Early works on 3D human reconstruction from monocular cameras represent the body shape without capturing shape details [1,3,4,6,21,22,35,47,51,57]. The introduction of neural implicit models [10,33,34] revolutionized 3D modeling by estimating an implicit representation, which reproduces higher-quality shapes due to its inherent property of defining a surface as a level set of an occupancy probability function. PiFU [40] introduced pixel-aligned implicit function for human reconstruction, aligning 3D points with features of the input image to achieve an unprecedented level of detail in the reconstructed shapes. Since then, neural implicit models have become the standard representation for 3D human shapes [5,11,14,18,28]. A significant limitation of these approaches is their inability to reproduce high-quality details in non-visible regions. SuRS [37] enhances shape quality by learning a map from a low to a high-resolution shape, hallucinating details in occluded regions, which are still overly smoothed as no information is provided to the implicit model. To improve the quality of non-visible regions, PiFU-HD [41]" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 38, + 9, + 346, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 9, + 346, + 21 + ], + "spans": [ + { + "bbox": [ + 38, + 9, + 346, + 21 + ], + "type": "text", + "content": "COSMU: Complete 3D human shape from monocular unconstrained images" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 373, + 10, + 380, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 10, + 380, + 19 + ], + "spans": [ + { + "bbox": [ + 373, + 10, + 380, + 19 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 31, + 35, + 380, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 35, + 380, + 274 + ], + "spans": [ + { + "bbox": [ + 31, + 35, + 380, + 274 + ], + "type": "text", + "content": "generates front and back surface normals of the human from the RGB image. This estimation introduces noise, resulting in smoothed surfaces of occluded regions. ICON [50] improves normal estimation by guiding it with SMPL [30] fitted to the image. Similarly, ECON [49] alleviates the SMPL problem of clothing reconstruction by feeding the estimated front and back normals into a d-BiNI optimizer. Conditioning the learning process on SMPL body poses a challenge for the network, which cannot fully exploit 2D image features. These methods estimate only an additional view, which is insufficient to represent the entire body and its accuracy influences the reconstruction quality of the non-visible regions. TeCH [17] proposes to estimate the unseen view from the input through Score Distillation Sampling [39] (SDS) and represents the human with drivable tetrahedral representation (DMTet [43]). Subject-tailored fine-tuning is required, limiting its generalizability. Other works focus only on 3D reconstruction from monocular videos, requiring to optimize their algorithms on a single monocular video, limiting the generalization and their application scenarios [1-3,20]. NERF-based methods are applied to 3D human reconstruction from sparse views, but their subject-specific nature limits their utility [8,13,19,23,36,46,48]. For a fair comparison with the works that require subject-specific optimization, our approach should be fine-tuned with the data of each specific subject, significantly limiting generalizability. Therefore, these works are not evaluated in this paper." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 32, + 274, + 286, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 274, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 32, + 274, + 286, + 285 + ], + "type": "text", + "content": "3D human shape from unconstrained RGB images" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 31, + 286, + 380, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 286, + 380, + 523 + ], + "spans": [ + { + "bbox": [ + 31, + 286, + 380, + 523 + ], + "type": "text", + "content": "PaMIR [56] extended single-view reconstruction to multiple unconstrained images by using the SMPL model as a body reference to establish correspondences across different images. Latent embedding vectors from the correspondent frames are aggregated into a single embedding using mean pooling and then processed by a multi-layer perception (MLP) to estimate the implicit representation. Similarly, MVP-Human [58] fits the SMPL model to each frame to establish correspondences between unconstrained images. Skinning weights of a canonical pose are estimated to guide the fusion of the features from multiple images through an attention-based implicit model. UnconFuse [16] builds upon MVP-Human, with a focus on estimating T-pose shapes from unconstrained images, while posed shapes cannot be reconstructed. CrossHuman [9] selects references following PoseFusion [25] and leveraging the SMPL model. Features are then extracted from both the fitted SMPL and RGB input and fused with a transformer to estimate the implicit representation. These works extract features from unconstrained images and fuse them with a neural model, losing details in regions where the body poses differ among frames. In contrast, COSMU generates 2D multi-view normals of the human in a spatially aligned pose, facilitating the estimation of the final implicit function. Moreover, these works rely on parametric models, whose fitting can introduce noise. COSMU distinguishes itself by not relying on parametric models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 31, + 524, + 379, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 524, + 379, + 574 + ], + "spans": [ + { + "bbox": [ + 31, + 524, + 379, + 574 + ], + "type": "text", + "content": "Have-Fun [52] estimates a DMTet [43] representation by integrating SDS [39] and SMPL skinning weights to adapt the unconstrained images to the target one. Similarly to Tech [17], fine-tuning on the target subject is required, limiting generalizability. Other works focus on 3D human reconstruction from other" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "spans": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 10, + 143, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 10, + 143, + 19 + ], + "spans": [ + { + "bbox": [ + 66, + 10, + 143, + 19 + ], + "type": "text", + "content": "M. Pesavento et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 33, + 33, + 378, + 182 + ], + "blocks": [ + { + "bbox": [ + 33, + 33, + 378, + 182 + ], + "lines": [ + { + "bbox": [ + 33, + 33, + 378, + 182 + ], + "spans": [ + { + "bbox": [ + 33, + 33, + 378, + 182 + ], + "type": "image", + "image_path": "69f9791d2a5e6317f14aa40b8878a9a84498d417dc26ec2042289c31f701062a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 31, + 191, + 382, + 258 + ], + "lines": [ + { + "bbox": [ + 31, + 191, + 382, + 258 + ], + "spans": [ + { + "bbox": [ + 31, + 191, + 382, + 258 + ], + "type": "text", + "content": "Fig. 2: COSMU framework: given a single RGB image as input and a collection of monocular unconstrained images, a multi-view scenario simulation module, which consists of part-based reference selection and registration, is first applied to generate registered multi-view 2D normal maps of the subject in the same pose. An implicit function " + }, + { + "bbox": [ + 31, + 191, + 382, + 258 + ], + "type": "inline_equation", + "content": "\\hat{s}_f" + }, + { + "bbox": [ + 31, + 191, + 382, + 258 + ], + "type": "text", + "content": " is then estimated by processing these normal maps with a multi-view attention-based neural implicit model. The 3D mesh is reconstructed via marching cube." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 30, + 279, + 381, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 279, + 381, + 305 + ], + "spans": [ + { + "bbox": [ + 30, + 279, + 381, + 305 + ], + "type": "text", + "content": "unconstrained data such as depth images [25, 38] or 3D meshes [26], which is beyond the scope of this paper." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 31, + 319, + 344, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 319, + 344, + 348 + ], + "spans": [ + { + "bbox": [ + 31, + 319, + 344, + 348 + ], + "type": "text", + "content": "3 3D human shape reconstruction from monocular unconstrained images" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 356, + 382, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 356, + 382, + 583 + ], + "spans": [ + { + "bbox": [ + 30, + 356, + 382, + 583 + ], + "type": "text", + "content": "We introduce a novel framework that reconstructs COcomplete 3D human Shape from Monocular Unconstrained images (COSMU). To be complete, a 3D human shape should contain high-quality details in every body region. A single image lacks the necessary information to reproduce details across the entire body, which can instead be extracted from multi-view images. In contrast to multiview reconstruction, which requires calibrated and registered images captured with inaccessible systems, COSMU, given a single target image, reproduces fine details in the regions of the 3D shape that are occluded in the input image. This is achieved by leveraging a collection of uncalibrated and unregistered monocular images, referred to as unconstrained images. The input is a target image " + }, + { + "bbox": [ + 30, + 356, + 382, + 583 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 30, + 356, + 382, + 583 + ], + "type": "text", + "content": " of a person together with multiple views of the same person in various poses and either same or different clothing, from which we extract body part detail non-visible in the target view. As shown in Fig. 2, given a set " + }, + { + "bbox": [ + 30, + 356, + 382, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{U} = \\{I_u^m\\}_{m=1}^{N_u}" + }, + { + "bbox": [ + 30, + 356, + 382, + 583 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 30, + 356, + 382, + 583 + ], + "type": "inline_equation", + "content": "N_{u}" + }, + { + "bbox": [ + 30, + 356, + 382, + 583 + ], + "type": "text", + "content": " monocular unconstrained images, COSMU simulates the multi-view scenario of traditional multi-view works using a novel module that employs body part-based reference selection and registration to create 2D multi-view normal maps registered with the target input. A multi-view attention-based implicit model regresses the implicit representation " + }, + { + "bbox": [ + 30, + 356, + 382, + 583 + ], + "type": "inline_equation", + "content": "f_{f}" + }, + { + "bbox": [ + 30, + 356, + 382, + 583 + ], + "type": "text", + "content": " of the complete 3D shape from the normals." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 38, + 9, + 347, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 9, + 347, + 21 + ], + "spans": [ + { + "bbox": [ + 38, + 9, + 347, + 21 + ], + "type": "text", + "content": "COSMU: Complete 3D human shape from monocular unconstrained images" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 33, + 255, + 46 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 33, + 255, + 46 + ], + "spans": [ + { + "bbox": [ + 32, + 33, + 255, + 46 + ], + "type": "text", + "content": "3.1 Multi-view scenario simulation module" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 32, + 54, + 380, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 54, + 380, + 170 + ], + "spans": [ + { + "bbox": [ + 32, + 54, + 380, + 170 + ], + "type": "text", + "content": "Directly processing unconstrained images with a multi-view neural model deteriorates the performance due to their unregistered nature (Sec. 4.1). To address this, we propose a novel pipeline that simulates a multi-view scenario in which the multi-view images represent various orientations of the subject in the same body pose. Given a target image " + }, + { + "bbox": [ + 32, + 54, + 380, + 170 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 32, + 54, + 380, + 170 + ], + "type": "text", + "content": ", we select " + }, + { + "bbox": [ + 32, + 54, + 380, + 170 + ], + "type": "inline_equation", + "content": "N_{p} \\times N_{r}" + }, + { + "bbox": [ + 32, + 54, + 380, + 170 + ], + "type": "text", + "content": " references " + }, + { + "bbox": [ + 32, + 54, + 380, + 170 + ], + "type": "inline_equation", + "content": "\\left\\{\\{I_{ref}^{(p,r)}\\}_{p=1}^{N_{p}}\\right\\}_{r=1}^{N_{r}}" + }, + { + "bbox": [ + 32, + 54, + 380, + 170 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 32, + 54, + 380, + 170 + ], + "type": "inline_equation", + "content": "I_{ref} \\in \\mathbf{U}" + }, + { + "bbox": [ + 32, + 54, + 380, + 170 + ], + "type": "text", + "content": ", from the collection " + }, + { + "bbox": [ + 32, + 54, + 380, + 170 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 32, + 54, + 380, + 170 + ], + "type": "text", + "content": " of unconstrained images applying a body part-based reference selection. 2D registered multi-view normal maps are then created by registering references to the target input via body part-based registration. Set notation is dropped for simplicity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "spans": [ + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "text", + "content": "Body part-based reference selection: To improve the body part-based registration, we select references observed from complementary orientations to the target one and with body part poses similar to the corresponding target body part ones. " + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "inline_equation", + "content": "N_{r}" + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "text", + "content": " references are selected for each " + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "text", + "content": " body part of the target " + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "text", + "content": " via energy optimization following an orientation-visibility-similarity criteria. Semantic masks " + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "text", + "content": " and 3D poses " + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "text", + "content": " are extracted from the input images. To ensure that the entire body is represented in the selected references, we first divide the " + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "inline_equation", + "content": "N_{u}" + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "text", + "content": " unconstrained images into " + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "inline_equation", + "content": "N_{r}" + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "text", + "content": " groups based on the full-body orientation angle " + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "inline_equation", + "content": "\\theta_{u}" + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "text", + "content": " computed between the reference " + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "inline_equation", + "content": "P_{u}" + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "text", + "content": " and the target " + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "inline_equation", + "content": "P_{t}" + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "text", + "content": " 3D poses with the Rodrigues rotation formula. Each group contains a variable number " + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "inline_equation", + "content": "N_{g}" + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "text", + "content": " of unconstrained images " + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "inline_equation", + "content": "I_{u}^{g}, g \\in [1,\\dots,N_{g}]" + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "text", + "content": " references " + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "inline_equation", + "content": "I_{ref}" + }, + { + "bbox": [ + 32, + 171, + 380, + 313 + ], + "type": "text", + "content": " are selected from each group following 3 criteria:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 32, + 314, + 380, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 314, + 380, + 362 + ], + "spans": [ + { + "bbox": [ + 32, + 314, + 380, + 362 + ], + "type": "text", + "content": "1. Body orientation angle: for a complete view of the body, the selected images should complement each other in the body orientation. An orientation metric is introduced to enforce the selection of images in which the body orientation is closer to the central angle " + }, + { + "bbox": [ + 32, + 314, + 380, + 362 + ], + "type": "inline_equation", + "content": "\\theta_c^r" + }, + { + "bbox": [ + 32, + 314, + 380, + 362 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 32, + 314, + 380, + 362 + ], + "type": "inline_equation", + "content": "r^{th}" + }, + { + "bbox": [ + 32, + 314, + 380, + 362 + ], + "type": "text", + "content": " group:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 372, + 378, + 394 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 372, + 378, + 394 + ], + "spans": [ + { + "bbox": [ + 99, + 372, + 378, + 394 + ], + "type": "interline_equation", + "content": "o _ {g} = \\left| \\theta_ {c} ^ {r} - \\theta_ {u} \\right|, \\quad \\theta_ {c} ^ {r} = r \\cdot \\frac {3 6 0}{N _ {r} + 1}, \\quad r \\in [ 1, \\dots , N _ {r} ] \\tag {1}", + "image_path": "a1ccaae9f5a3a6bf2923869b81681afb2da09f70cb3af9d82eb84cac6179e2c5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 32, + 410, + 379, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 410, + 379, + 446 + ], + "spans": [ + { + "bbox": [ + 32, + 410, + 379, + 446 + ], + "type": "text", + "content": "2. Body part visibility: every body part should be visible in the selected references to have a complete view of the body. We use a visibility metric that counts the total number of pixels " + }, + { + "bbox": [ + 32, + 410, + 379, + 446 + ], + "type": "inline_equation", + "content": "N_{M}" + }, + { + "bbox": [ + 32, + 410, + 379, + 446 + ], + "type": "text", + "content": " of each body part " + }, + { + "bbox": [ + 32, + 410, + 379, + 446 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 32, + 410, + 379, + 446 + ], + "type": "text", + "content": " in the semantic masks " + }, + { + "bbox": [ + 32, + 410, + 379, + 446 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 32, + 410, + 379, + 446 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 150, + 457, + 378, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 457, + 378, + 469 + ], + "spans": [ + { + "bbox": [ + 150, + 457, + 378, + 469 + ], + "type": "interline_equation", + "content": "v _ {g} ^ {p} = N _ {M} ^ {p}, \\quad p \\in [ 1, \\dots , N _ {p} ] \\tag {2}", + "image_path": "4d0ee1e19d05b703dc4ab1af494a5098bc529e5b67a997665116294f68cbc451.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 32, + 489, + 379, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 489, + 379, + 526 + ], + "spans": [ + { + "bbox": [ + 32, + 489, + 379, + 526 + ], + "type": "text", + "content": "3. Body part pose similarity: similarity between each body part of " + }, + { + "bbox": [ + 32, + 489, + 379, + 526 + ], + "type": "inline_equation", + "content": "I_{u}^{g}" + }, + { + "bbox": [ + 32, + 489, + 379, + 526 + ], + "type": "text", + "content": " and the corresponding body part of " + }, + { + "bbox": [ + 32, + 489, + 379, + 526 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 32, + 489, + 379, + 526 + ], + "type": "text", + "content": " is evaluated as 3D Euclidean distance between the corresponding 3D bones of " + }, + { + "bbox": [ + 32, + 489, + 379, + 526 + ], + "type": "inline_equation", + "content": "P_{g}" + }, + { + "bbox": [ + 32, + 489, + 379, + 526 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 32, + 489, + 379, + 526 + ], + "type": "inline_equation", + "content": "P_{t}" + }, + { + "bbox": [ + 32, + 489, + 379, + 526 + ], + "type": "text", + "content": ", ensuring spatial consistency:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 135, + 537, + 378, + 549 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 537, + 378, + 549 + ], + "spans": [ + { + "bbox": [ + 135, + 537, + 378, + 549 + ], + "type": "interline_equation", + "content": "b _ {g} ^ {p} = \\left\\| P _ {t} ^ {p} - P _ {g} ^ {p} \\right\\|, \\quad p \\in [ 1, \\dots , N _ {p} ] \\tag {3}", + "image_path": "4934af64e7acd783576d3a4de98c0bf1677a1cd352660ea97f6a54a93204ab3b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 32, + 558, + 379, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 558, + 379, + 584 + ], + "spans": [ + { + "bbox": [ + 32, + 558, + 379, + 584 + ], + "type": "inline_equation", + "content": "o_{g} = o(I_{u}^{g}), v_{g}^{p} = v^{p}(I_{u}^{g}), b_{g}^{p} = b^{p}(I_{u}^{g})" + }, + { + "bbox": [ + 32, + 558, + 379, + 584 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 32, + 558, + 379, + 584 + ], + "type": "inline_equation", + "content": "g^{th}" + }, + { + "bbox": [ + 32, + 558, + 379, + 584 + ], + "type": "text", + "content": " image " + }, + { + "bbox": [ + 32, + 558, + 379, + 584 + ], + "type": "inline_equation", + "content": "I_{u}^{g}" + }, + { + "bbox": [ + 32, + 558, + 379, + 584 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 32, + 558, + 379, + 584 + ], + "type": "inline_equation", + "content": "p^{th}" + }, + { + "bbox": [ + 32, + 558, + 379, + 584 + ], + "type": "text", + "content": " body part. " + }, + { + "bbox": [ + 32, + 558, + 379, + 584 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 32, + 558, + 379, + 584 + ], + "type": "text", + "content": " references " + }, + { + "bbox": [ + 32, + 558, + 379, + 584 + ], + "type": "inline_equation", + "content": "I_{ref}" + }, + { + "bbox": [ + 32, + 558, + 379, + 584 + ], + "type": "text", + "content": " are selected by minimizing the energy " + }, + { + "bbox": [ + 32, + 558, + 379, + 584 + ], + "type": "inline_equation", + "content": "E_{g}^{p}" + }, + { + "bbox": [ + 32, + 558, + 379, + 584 + ], + "type": "text", + "content": " over the " + }, + { + "bbox": [ + 32, + 558, + 379, + 584 + ], + "type": "inline_equation", + "content": "N_{g}" + }, + { + "bbox": [ + 32, + 558, + 379, + 584 + ], + "type": "text", + "content": " images" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 32, + 10, + 39, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 10, + 39, + 19 + ], + "spans": [ + { + "bbox": [ + 32, + 10, + 39, + 19 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 144, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 144, + 19 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 144, + 19 + ], + "type": "text", + "content": "M. Pesavento et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 34, + 139, + 47 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 34, + 139, + 47 + ], + "spans": [ + { + "bbox": [ + 32, + 34, + 139, + 47 + ], + "type": "inline_equation", + "content": "I_u^g" + }, + { + "bbox": [ + 32, + 34, + 139, + 47 + ], + "type": "text", + "content": " for each body part " + }, + { + "bbox": [ + 32, + 34, + 139, + 47 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 32, + 34, + 139, + 47 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 164, + 56, + 379, + 79 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 56, + 379, + 79 + ], + "spans": [ + { + "bbox": [ + 164, + 56, + 379, + 79 + ], + "type": "interline_equation", + "content": "I _ {r e f} = \\underset {g} {\\operatorname {a r g m i n}} \\left(E _ {g} ^ {p}\\right) \\tag {4}", + "image_path": "0a31b247a41182708672137a6c4e9babe53c566b96652a7a266869e508f47ec5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 79, + 318, + 94 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 79, + 318, + 94 + ], + "spans": [ + { + "bbox": [ + 93, + 79, + 318, + 94 + ], + "type": "interline_equation", + "content": "E _ {g} ^ {p} = i d x _ {m i n} \\left(o _ {g}\\right) + i d x _ {m a x} \\left(v _ {g} ^ {p}\\right) + i d x _ {m i n} \\left(b _ {g} ^ {p}\\right)", + "image_path": "ad2243e5f4fe27c25bb04ad7d3142fd9fb118d1819a5dd7bf0fb942686759019.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 31, + 103, + 379, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 103, + 379, + 139 + ], + "spans": [ + { + "bbox": [ + 31, + 103, + 379, + 139 + ], + "type": "inline_equation", + "content": "idx_{min}" + }, + { + "bbox": [ + 31, + 103, + 379, + 139 + ], + "type": "text", + "content": " indicates the position " + }, + { + "bbox": [ + 31, + 103, + 379, + 139 + ], + "type": "inline_equation", + "content": "idx" + }, + { + "bbox": [ + 31, + 103, + 379, + 139 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 31, + 103, + 379, + 139 + ], + "type": "inline_equation", + "content": "g^{th}" + }, + { + "bbox": [ + 31, + 103, + 379, + 139 + ], + "type": "text", + "content": " image " + }, + { + "bbox": [ + 31, + 103, + 379, + 139 + ], + "type": "inline_equation", + "content": "I_u^g" + }, + { + "bbox": [ + 31, + 103, + 379, + 139 + ], + "type": "text", + "content": " in the arrays created by sorting all the " + }, + { + "bbox": [ + 31, + 103, + 379, + 139 + ], + "type": "inline_equation", + "content": "N_g" + }, + { + "bbox": [ + 31, + 103, + 379, + 139 + ], + "type": "text", + "content": " images with respect to their specific metric " + }, + { + "bbox": [ + 31, + 103, + 379, + 139 + ], + "type": "inline_equation", + "content": "K = o_g, v_g^p, b_g^p" + }, + { + "bbox": [ + 31, + 103, + 379, + 139 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 31, + 103, + 379, + 139 + ], + "type": "inline_equation", + "content": "idx_{max}" + }, + { + "bbox": [ + 31, + 103, + 379, + 139 + ], + "type": "text", + "content": " sorts the array in inverse order:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 148, + 377, + 170 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 148, + 377, + 170 + ], + "spans": [ + { + "bbox": [ + 76, + 148, + 377, + 170 + ], + "type": "interline_equation", + "content": "i d x _ {m i n} (K) = \\mathrm {i d x} _ {I _ {u} ^ {g}} \\left\\{I _ {u} ^ {1}, \\dots , I _ {u} ^ {N _ {g}} \\right\\}, K \\left(I _ {u} ^ {1}\\right) \\leqslant \\dots \\leqslant K \\left(I _ {u} ^ {N _ {g}}\\right) \\tag {5}", + "image_path": "c9896543034404b9cad4f4a259fdeb776e91875ba861d011e0d42047452fdca1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 167, + 335, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 167, + 335, + 182 + ], + "spans": [ + { + "bbox": [ + 76, + 167, + 335, + 182 + ], + "type": "interline_equation", + "content": "i d x _ {m a x} (K) = \\mathrm {i d x} _ {I _ {u} ^ {g}} \\left\\{I _ {u} ^ {1}, \\dots , I _ {u} ^ {N _ {g}} \\right\\}, K \\left(I _ {u} ^ {1}\\right) \\geqslant \\dots \\geqslant K \\left(I _ {u} ^ {N _ {g}}\\right)", + "image_path": "d417dae5512281c77b3685a898c895f819f56817b695c04ec9faeb5e488d9a5f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 31, + 190, + 379, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 190, + 379, + 213 + ], + "spans": [ + { + "bbox": [ + 31, + 190, + 379, + 213 + ], + "type": "inline_equation", + "content": "N_{p} \\times N_{r}" + }, + { + "bbox": [ + 31, + 190, + 379, + 213 + ], + "type": "text", + "content": " references are selected by repeating the body part-based reference selection for every group." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 30, + 214, + 380, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 214, + 380, + 298 + ], + "spans": [ + { + "bbox": [ + 30, + 214, + 380, + 298 + ], + "type": "text", + "content": "Body part-based registration: To simulate the multi-view scenario given the selected " + }, + { + "bbox": [ + 30, + 214, + 380, + 298 + ], + "type": "inline_equation", + "content": "N_{p} \\times N_{r}" + }, + { + "bbox": [ + 30, + 214, + 380, + 298 + ], + "type": "text", + "content": " references, we register each 3D reference body part to the corresponding 3D body part in " + }, + { + "bbox": [ + 30, + 214, + 380, + 298 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 30, + 214, + 380, + 298 + ], + "type": "text", + "content": ". This process begins with the reconstruction of a coarse point cloud " + }, + { + "bbox": [ + 30, + 214, + 380, + 298 + ], + "type": "inline_equation", + "content": "\\zeta_{t}" + }, + { + "bbox": [ + 30, + 214, + 380, + 298 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 30, + 214, + 380, + 298 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 30, + 214, + 380, + 298 + ], + "type": "text", + "content": " and for each of the " + }, + { + "bbox": [ + 30, + 214, + 380, + 298 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 30, + 214, + 380, + 298 + ], + "type": "text", + "content": " reference images " + }, + { + "bbox": [ + 30, + 214, + 380, + 298 + ], + "type": "inline_equation", + "content": "I_{ref}" + }, + { + "bbox": [ + 30, + 214, + 380, + 298 + ], + "type": "text", + "content": " previously selected for the " + }, + { + "bbox": [ + 30, + 214, + 380, + 298 + ], + "type": "inline_equation", + "content": "N_{r}" + }, + { + "bbox": [ + 30, + 214, + 380, + 298 + ], + "type": "text", + "content": " groups " + }, + { + "bbox": [ + 30, + 214, + 380, + 298 + ], + "type": "inline_equation", + "content": "(\\zeta_{ref})" + }, + { + "bbox": [ + 30, + 214, + 380, + 298 + ], + "type": "text", + "content": ". We use a neural implicit model " + }, + { + "bbox": [ + 30, + 214, + 380, + 298 + ], + "type": "inline_equation", + "content": "f_{c}" + }, + { + "bbox": [ + 30, + 214, + 380, + 298 + ], + "type": "text", + "content": " comprising of a U-Net that increases the resolution of the input features " + }, + { + "bbox": [ + 30, + 214, + 380, + 298 + ], + "type": "inline_equation", + "content": "\\phi(I)" + }, + { + "bbox": [ + 30, + 214, + 380, + 298 + ], + "type": "text", + "content": " as in SuRS [37]. The coarse implicit function is estimated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 135, + 307, + 379, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 307, + 379, + 326 + ], + "spans": [ + { + "bbox": [ + 135, + 307, + 379, + 326 + ], + "type": "interline_equation", + "content": "\\hat {s} _ {c} = f _ {c} \\left(\\phi (\\pi (I)), z (x)\\right), \\quad \\hat {s} _ {c} \\in \\mathbb {R} \\tag {6}", + "image_path": "53dc4058566923a382aa8f1f42e54a46091f32befecfd58dc76f44b8bab702e9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 31, + 334, + 379, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 334, + 379, + 359 + ], + "spans": [ + { + "bbox": [ + 31, + 334, + 379, + 359 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 31, + 334, + 379, + 359 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 31, + 334, + 379, + 359 + ], + "type": "text", + "content": " is either " + }, + { + "bbox": [ + 31, + 334, + 379, + 359 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 31, + 334, + 379, + 359 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 31, + 334, + 379, + 359 + ], + "type": "inline_equation", + "content": "I_{ref}" + }, + { + "bbox": [ + 31, + 334, + 379, + 359 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 31, + 334, + 379, + 359 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 31, + 334, + 379, + 359 + ], + "type": "text", + "content": " represents the orthographic projection of 3D points " + }, + { + "bbox": [ + 31, + 334, + 379, + 359 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 31, + 334, + 379, + 359 + ], + "type": "text", + "content": " into the input images " + }, + { + "bbox": [ + 31, + 334, + 379, + 359 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 31, + 334, + 379, + 359 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 31, + 334, + 379, + 359 + ], + "type": "inline_equation", + "content": "z(x)" + }, + { + "bbox": [ + 31, + 334, + 379, + 359 + ], + "type": "text", + "content": " is the depth of " + }, + { + "bbox": [ + 31, + 334, + 379, + 359 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 31, + 334, + 379, + 359 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 31, + 334, + 379, + 359 + ], + "type": "inline_equation", + "content": "f_{c}" + }, + { + "bbox": [ + 31, + 334, + 379, + 359 + ], + "type": "text", + "content": " is trained with:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 109, + 368, + 379, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 368, + 379, + 399 + ], + "spans": [ + { + "bbox": [ + 109, + 368, + 379, + 399 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left| f _ {c} \\left(X _ {c} ^ {i}\\right) - f _ {c} ^ {g t} \\left(X _ {c} ^ {i}\\right) \\right| ^ {2} + | I - I _ {H R} | \\tag {7}", + "image_path": "a615700c129b3a78f6fb0d069f5be5621ed250ecb8aa97296190257478ae548e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "spans": [ + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "inline_equation", + "content": "X_{c}^{i} = (\\phi^{i}(\\pi (I)),z(x))" + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "text", + "content": " is the number of points in space and " + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "inline_equation", + "content": "I_{HR}" + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "text", + "content": " is the up-sampled ground-truth version of " + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "text", + "content": " used to train the U-Net feature extractor. We then retrieve 3D body parts by classifying each 3D point " + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "inline_equation", + "content": "x_{\\zeta}" + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "text", + "content": " of the point clouds " + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "inline_equation", + "content": "\\zeta_{t / ref}" + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "text", + "content": " with a novel pixel-aligned classification network " + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "inline_equation", + "content": "f_{\\zeta}" + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "text", + "content": ". An image feature embedding " + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "inline_equation", + "content": "\\phi (I||M)" + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "text", + "content": " is extracted from the concatenation between the input RGB image and its semantic mask " + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "text", + "content": ". Each " + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "inline_equation", + "content": "x_{\\zeta}" + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "text", + "content": " is projected using orthographic projection " + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "text", + "content": " onto the input " + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "inline_equation", + "content": "I||M" + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "text", + "content": " and indexed to " + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "inline_equation", + "content": "\\phi (I||M)" + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "text", + "content": ". An MLP processes this embedding to estimate the body part " + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "inline_equation", + "content": "\\hat{s}_{\\zeta}" + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "text", + "content": " to which " + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "inline_equation", + "content": "x_{\\zeta}" + }, + { + "bbox": [ + 30, + 408, + 380, + 505 + ], + "type": "text", + "content": " belongs:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 109, + 515, + 379, + 527 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 515, + 379, + 527 + ], + "spans": [ + { + "bbox": [ + 109, + 515, + 379, + 527 + ], + "type": "interline_equation", + "content": "\\hat {s} _ {\\zeta} = f _ {\\zeta} \\left(\\phi \\left(\\pi (I | | M)\\right), z \\left(x _ {\\zeta}\\right)\\right), \\quad \\hat {s} _ {\\zeta} \\in [ 1, \\dots , N _ {p} ] \\tag {8}", + "image_path": "3c85f44385199d582ebd4de59e3a9186b03d324c09aaeb4683dafed6a0f4c9e8.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 30, + 535, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 535, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 535, + 380, + 584 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 30, + 535, + 380, + 584 + ], + "type": "text", + "content": " 3D body parts " + }, + { + "bbox": [ + 30, + 535, + 380, + 584 + ], + "type": "inline_equation", + "content": "v_{ref}" + }, + { + "bbox": [ + 30, + 535, + 380, + 584 + ], + "type": "text", + "content": " are defined for each reference point cloud as well as for the target " + }, + { + "bbox": [ + 30, + 535, + 380, + 584 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 30, + 535, + 380, + 584 + ], + "type": "text", + "content": ". Next, the specific " + }, + { + "bbox": [ + 30, + 535, + 380, + 584 + ], + "type": "inline_equation", + "content": "p^{th}" + }, + { + "bbox": [ + 30, + 535, + 380, + 584 + ], + "type": "text", + "content": " body part " + }, + { + "bbox": [ + 30, + 535, + 380, + 584 + ], + "type": "inline_equation", + "content": "v_{ref}" + }, + { + "bbox": [ + 30, + 535, + 380, + 584 + ], + "type": "text", + "content": " selected for the corresponding " + }, + { + "bbox": [ + 30, + 535, + 380, + 584 + ], + "type": "inline_equation", + "content": "p^{th}" + }, + { + "bbox": [ + 30, + 535, + 380, + 584 + ], + "type": "text", + "content": " target body part " + }, + { + "bbox": [ + 30, + 535, + 380, + 584 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 30, + 535, + 380, + 584 + ], + "type": "text", + "content": " is registered with it as " + }, + { + "bbox": [ + 30, + 535, + 380, + 584 + ], + "type": "inline_equation", + "content": "v_{reg} = R(v_{ref}) + T" + }, + { + "bbox": [ + 30, + 535, + 380, + 584 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 30, + 535, + 380, + 584 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 30, + 535, + 380, + 584 + ], + "type": "text", + "content": " is the rotation and " + }, + { + "bbox": [ + 30, + 535, + 380, + 584 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 30, + 535, + 380, + 584 + ], + "type": "text", + "content": " the translation computed via rigid registration. For each" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 38, + 9, + 346, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 9, + 346, + 20 + ], + "spans": [ + { + "bbox": [ + 38, + 9, + 346, + 20 + ], + "type": "text", + "content": "COSMU: Complete 3D human shape from monocular unconstrained images" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "spans": [ + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "text", + "content": "group, the registered " + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "text", + "content": " body parts are combined to create a point cloud " + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "inline_equation", + "content": "\\zeta_{f}" + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "inline_equation", + "content": "N_{r}" + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "text", + "content": " point clouds " + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "inline_equation", + "content": "\\zeta_{f}" + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "text", + "content": " registered with " + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "inline_equation", + "content": "\\zeta_{t}" + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "text", + "content": " are generated, rotated by their " + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "inline_equation", + "content": "\\theta_{u}" + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "text", + "content": " and, together with " + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "inline_equation", + "content": "\\zeta_{t}" + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "text", + "content": ", used to retrieve " + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "inline_equation", + "content": "N_{r} + 1" + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "text", + "content": " 2D normal maps " + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "inline_equation", + "content": "S_{v}" + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "text", + "content": ", one for each view " + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "inline_equation", + "content": "v \\in [1, \\dots, N_{r} + 1]" + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "text", + "content": ". Normal maps are generated instead of RGB images to avoid the estimation of color values for each 3D point of " + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "inline_equation", + "content": "\\zeta_{t / ref}" + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "text", + "content": " and to allow different clothing in the references. Directly fusing the registered point clouds to obtain the 3D shape is challenging since correspondences are unknown. We train " + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "inline_equation", + "content": "f_{\\zeta}" + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "text", + "content": " with a Mean Squared Error " + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "inline_equation", + "content": "MSE" + }, + { + "bbox": [ + 31, + 35, + 381, + 130 + ], + "type": "text", + "content": " loss:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 379, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 379, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 379, + 172 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p a r t}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left| f _ {\\zeta} \\left(X _ {\\zeta} ^ {i}\\right) - f _ {\\zeta} ^ {g t} \\left(X _ {\\zeta} ^ {i}\\right) \\right| ^ {2} \\tag {9}", + "image_path": "cd3e6836ad700a09000b7c8579d04bbddc65e8cba5e6415adcc894bbee554087.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 31, + 181, + 180, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 181, + 180, + 201 + ], + "spans": [ + { + "bbox": [ + 31, + 181, + 180, + 201 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 31, + 181, + 180, + 201 + ], + "type": "inline_equation", + "content": "X_{\\zeta}^{i} = \\left(\\phi^{i}(\\pi (I||M)),z(x_{\\zeta})\\right)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 31, + 218, + 232, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 218, + 232, + 230 + ], + "spans": [ + { + "bbox": [ + 31, + 218, + 232, + 230 + ], + "type": "text", + "content": "3.2 Multi-view implicit representation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 239, + 381, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 239, + 381, + 433 + ], + "spans": [ + { + "bbox": [ + 30, + 239, + 381, + 433 + ], + "type": "text", + "content": "Neural implicit models have proven to be effective in reproducing high-quality details in reconstructed 3D shapes. This is attributed to the characteristic of the implicit function that does not discretize the 3D space like other representations [40]. Leveraging multiple views to learn the implicit representation further enhances the quality of the reconstruction, as it enables the incorporation of information about the entire body. Concatenating features from multiple images is a suboptimal approach because it imposes an order on the views [59]. Fusion approaches based on statistics have been proposed [12,44] but their pooling operation causes a loss of information regarding the individual view contributions. Inspired by Zins et al. [59], we propose to fuse the features extracted from the multi-view 2D normals using an attention-based architecture. An attention score is computed for each view " + }, + { + "bbox": [ + 30, + 239, + 381, + 433 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 30, + 239, + 381, + 433 + ], + "type": "text", + "content": " based on the compatibility of a query with a corresponding key, resulting in the generation of " + }, + { + "bbox": [ + 30, + 239, + 381, + 433 + ], + "type": "inline_equation", + "content": "N_r + 1" + }, + { + "bbox": [ + 30, + 239, + 381, + 433 + ], + "type": "text", + "content": " features. Each feature contains the original information from the corresponding view, merged with the information from all the other views. An MLP then models a pixel-aligned implicit function " + }, + { + "bbox": [ + 30, + 239, + 381, + 433 + ], + "type": "inline_equation", + "content": "f_f" + }, + { + "bbox": [ + 30, + 239, + 381, + 433 + ], + "type": "text", + "content": " to learn the occupancy value " + }, + { + "bbox": [ + 30, + 239, + 381, + 433 + ], + "type": "inline_equation", + "content": "\\hat{s}_f" + }, + { + "bbox": [ + 30, + 239, + 381, + 433 + ], + "type": "text", + "content": " of the 3D points in space:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 120, + 444, + 378, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 444, + 378, + 456 + ], + "spans": [ + { + "bbox": [ + 120, + 444, + 378, + 456 + ], + "type": "interline_equation", + "content": "\\hat {s} _ {f} = f _ {f} \\left(A \\left(\\phi_ {v} \\left(\\pi \\left(S _ {v}\\right)\\right), z (x)\\right)\\right), \\quad \\hat {s} _ {f} \\in \\mathbb {R} \\tag {10}", + "image_path": "4b4ad2f756e9cc11a132fcb7a7652994077d9d903aafbab684edea2e4084b7cf.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 31, + 461, + 379, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 461, + 379, + 509 + ], + "spans": [ + { + "bbox": [ + 31, + 461, + 379, + 509 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 31, + 461, + 379, + 509 + ], + "type": "inline_equation", + "content": "\\phi_v" + }, + { + "bbox": [ + 31, + 461, + 379, + 509 + ], + "type": "text", + "content": " are the feature extracted from each view " + }, + { + "bbox": [ + 31, + 461, + 379, + 509 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 31, + 461, + 379, + 509 + ], + "type": "text", + "content": " of the surface normal " + }, + { + "bbox": [ + 31, + 461, + 379, + 509 + ], + "type": "inline_equation", + "content": "S_v" + }, + { + "bbox": [ + 31, + 461, + 379, + 509 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 31, + 461, + 379, + 509 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 31, + 461, + 379, + 509 + ], + "type": "text", + "content": " indicates the pixel-alignment between the 3D points of the space and each " + }, + { + "bbox": [ + 31, + 461, + 379, + 509 + ], + "type": "inline_equation", + "content": "\\phi_v" + }, + { + "bbox": [ + 31, + 461, + 379, + 509 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 31, + 461, + 379, + 509 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 31, + 461, + 379, + 509 + ], + "type": "text", + "content": " is the attention operation and " + }, + { + "bbox": [ + 31, + 461, + 379, + 509 + ], + "type": "inline_equation", + "content": "z(x)" + }, + { + "bbox": [ + 31, + 461, + 379, + 509 + ], + "type": "text", + "content": " is the depth value of the 3D points " + }, + { + "bbox": [ + 31, + 461, + 379, + 509 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 31, + 461, + 379, + 509 + ], + "type": "text", + "content": ". We train the multi-view architecture with an MSE loss:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 518, + 379, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 518, + 379, + 548 + ], + "spans": [ + { + "bbox": [ + 134, + 518, + 379, + 548 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left| f _ {f} \\left(X _ {f} ^ {i}\\right) - f _ {f} ^ {g t} \\left(X _ {f} ^ {i}\\right) \\right| ^ {2} \\tag {11}", + "image_path": "f732f1a1c6f91b75702e0c37ba735e056ce21586fb6f0c35ad47176d75190072.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 31, + 558, + 180, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 558, + 180, + 571 + ], + "spans": [ + { + "bbox": [ + 31, + 558, + 180, + 571 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 31, + 558, + 180, + 571 + ], + "type": "inline_equation", + "content": "X_{f}^{i} = A\\left(\\phi_{v}^{i}(\\pi (S_{v})),z(x^{i})\\right)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 31, + 571, + 379, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 571, + 379, + 584 + ], + "spans": [ + { + "bbox": [ + 31, + 571, + 379, + 584 + ], + "type": "text", + "content": "Inference: " + }, + { + "bbox": [ + 31, + 571, + 379, + 584 + ], + "type": "inline_equation", + "content": "N_{p} \\times N_{r}" + }, + { + "bbox": [ + 31, + 571, + 379, + 584 + ], + "type": "text", + "content": " references are selected from the collection of unconstrained" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "spans": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 144, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 144, + 19 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 144, + 19 + ], + "type": "text", + "content": "M. Pesavento et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 40, + 31, + 370, + 148 + ], + "blocks": [ + { + "bbox": [ + 40, + 31, + 370, + 148 + ], + "lines": [ + { + "bbox": [ + 40, + 31, + 370, + 148 + ], + "spans": [ + { + "bbox": [ + 40, + 31, + 370, + 148 + ], + "type": "image", + "image_path": "f13704363cab77fc46968819c620237595c0bb84d3a5147d1bba4f1902bc1eaa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 31, + 156, + 380, + 190 + ], + "lines": [ + { + "bbox": [ + 31, + 156, + 380, + 190 + ], + "spans": [ + { + "bbox": [ + 31, + 156, + 380, + 190 + ], + "type": "text", + "content": "Fig. 3: Qualitative comparisons between 3D shapes from THuman3.0 [45] reconstructed by changing the energy optimization process to select reference images. Front, back and side views of the shapes are illustrated." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 30, + 214, + 381, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 214, + 381, + 382 + ], + "spans": [ + { + "bbox": [ + 30, + 214, + 381, + 382 + ], + "type": "text", + "content": "images. " + }, + { + "bbox": [ + 30, + 214, + 381, + 382 + ], + "type": "inline_equation", + "content": "N_r + 1" + }, + { + "bbox": [ + 30, + 214, + 381, + 382 + ], + "type": "text", + "content": " 2D normal maps are generated using body part-based registration to simulate a multi-view scenario with registered views of the subject. The normals are processed by the multi-view attention-based model to estimate the implicit representation. The 3D mesh is reconstructed via marching cube [32]. To achieve pixel-alignment through projection, multi-view works require the parameters of calibrated cameras in specialized capture systems. In COSMU, instead of using the camera projection matrix for pixel-alignment, 3D points of the space are rotated by " + }, + { + "bbox": [ + 30, + 214, + 381, + 382 + ], + "type": "inline_equation", + "content": "\\theta_u" + }, + { + "bbox": [ + 30, + 214, + 381, + 382 + ], + "type": "text", + "content": " computed by the multi-view scenario simulation module and projected onto the images using an identity matrix. Existing approaches create the inference 3D grid centered at the origin. If the subject is not centered in the input images, the multi-view pixel-alignment of COSMU is inaccurate since the same rotated 3D point may not be projected in the corresponding pixel in each view. To tackle this, we set the boundary limits of the 3D grid as the boundaries of the coarse point cloud " + }, + { + "bbox": [ + 30, + 214, + 381, + 382 + ], + "type": "inline_equation", + "content": "\\zeta_t" + }, + { + "bbox": [ + 30, + 214, + 381, + 382 + ], + "type": "text", + "content": ", ensuring multi-view pixel-alignment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 31, + 401, + 128, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 401, + 128, + 414 + ], + "spans": [ + { + "bbox": [ + 31, + 401, + 128, + 414 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 428, + 382, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 428, + 382, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 428, + 382, + 584 + ], + "type": "text", + "content": "We conduct a comprehensive evaluation of COSMU for the task of 3D human shape reconstruction from a target image by leveraging a collection of monocular unconstrained images of the same individual. The neural models of the framework are trained with THuman2.0 dataset [54], which comprises 524 high-resolution surfaces. We render 360-degree view images around each surface, which serves as input data. For the quantitative and qualitative evaluation, we use the THuman3.0 dataset [45], which consists of 20 individuals each captured in various poses (400 models), and the testing set of X-Humans [42] (41 models). For each pose, we render 360-degree views of the subject and we apply the reference selection to select the " + }, + { + "bbox": [ + 30, + 428, + 382, + 584 + ], + "type": "inline_equation", + "content": "N_{p} \\times N_{r}" + }, + { + "bbox": [ + 30, + 428, + 382, + 584 + ], + "type": "text", + "content": " references among the images of the subject in different poses, excluding the target one. We evaluate the reconstruction accuracy with three quantitative metrics: the normal reprojection error [40], the average point-to-surface Euclidean distance (P2S) and the Chamfer distance" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 38, + 9, + 347, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 9, + 347, + 21 + ], + "spans": [ + { + "bbox": [ + 38, + 9, + 347, + 21 + ], + "type": "text", + "content": "COSMU: Complete 3D human shape from monocular unconstrained images" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 30, + 34, + 380, + 130 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 34, + 380, + 130 + ], + "spans": [ + { + "bbox": [ + 30, + 34, + 380, + 130 + ], + "type": "text", + "content": "(CD), expressed in cm. We further test COSMU with real data collected from various scenarios, including monocular videos (People-Snapshot [3], iPER [29] datasets) or images of people in different poses and clothing from the Internet, demonstrating its superiority over related works to reconstruct complete 3D human shapes with higher quality details in regions of the body occluded in the input target image (Sec. 4.2). In contrast to prior works, no parametric models are used. We set " + }, + { + "bbox": [ + 30, + 34, + 380, + 130 + ], + "type": "inline_equation", + "content": "N_{p} = 10" + }, + { + "bbox": [ + 30, + 34, + 380, + 130 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 34, + 380, + 130 + ], + "type": "inline_equation", + "content": "N_{r} = 3" + }, + { + "bbox": [ + 30, + 34, + 380, + 130 + ], + "type": "text", + "content": ". See supplementary for additional implementation details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 31, + 145, + 144, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 145, + 144, + 156 + ], + "spans": [ + { + "bbox": [ + 31, + 145, + 144, + 156 + ], + "type": "text", + "content": "4.1 Ablation Studies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 30, + 162, + 380, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 162, + 380, + 258 + ], + "spans": [ + { + "bbox": [ + 30, + 162, + 380, + 258 + ], + "type": "text", + "content": "Energy function for reference selection: We conduct an ablation study on the criteria used for selecting the " + }, + { + "bbox": [ + 30, + 162, + 380, + 258 + ], + "type": "inline_equation", + "content": "N_{p} \\times N_{r}" + }, + { + "bbox": [ + 30, + 162, + 380, + 258 + ], + "type": "text", + "content": " references in the body part-based reference selection. The variants explored are: (i) Random: no criteria are applied. " + }, + { + "bbox": [ + 30, + 162, + 380, + 258 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 30, + 162, + 380, + 258 + ], + "type": "text", + "content": " references are randomly selected for each group; (ii) Only " + }, + { + "bbox": [ + 30, + 162, + 380, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{b_p}" + }, + { + "bbox": [ + 30, + 162, + 380, + 258 + ], + "type": "text", + "content": ": only the body part pose similarity is considered; (iii) Only " + }, + { + "bbox": [ + 30, + 162, + 380, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 30, + 162, + 380, + 258 + ], + "type": "text", + "content": ": the body orientation angle is the only criterion used; (iv) Only " + }, + { + "bbox": [ + 30, + 162, + 380, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{v_p}" + }, + { + "bbox": [ + 30, + 162, + 380, + 258 + ], + "type": "text", + "content": ": only the body part visibility is considered; (v) " + }, + { + "bbox": [ + 30, + 162, + 380, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{o} + \\mathbf{v_p}" + }, + { + "bbox": [ + 30, + 162, + 380, + 258 + ], + "type": "text", + "content": ": the body part pose similarity is excluded; (vi) " + }, + { + "bbox": [ + 30, + 162, + 380, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{o} + \\mathbf{b_p}" + }, + { + "bbox": [ + 30, + 162, + 380, + 258 + ], + "type": "text", + "content": ": the body part visibility is omitted; (vii) " + }, + { + "bbox": [ + 30, + 162, + 380, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{v_p} + \\mathbf{b_p}" + }, + { + "bbox": [ + 30, + 162, + 380, + 258 + ], + "type": "text", + "content": ": the orientation angle is not used." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 30, + 258, + 380, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 258, + 380, + 341 + ], + "spans": [ + { + "bbox": [ + 30, + 258, + 380, + 341 + ], + "type": "text", + "content": "As illustrated in Fig. 3 and supported by the quantitative results of Tab. 1, the 3D shapes obtained with COSMU exhibit fewer artifacts and finer details. The reference selection plays a crucial role in selecting the most suitable references to facilitate the body part-based registration. Omitting any of the criteria in the energy optimization influences the selection of references, leading to suboptimal choices and, consequently, deteriorating the body part-based registration accuracy and introducing artifacts in the reconstruction." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 49, + 392, + 361, + 438 + ], + "blocks": [ + { + "bbox": [ + 30, + 361, + 379, + 384 + ], + "lines": [ + { + "bbox": [ + 30, + 361, + 379, + 384 + ], + "spans": [ + { + "bbox": [ + 30, + 361, + 379, + 384 + ], + "type": "text", + "content": "Table 1: Quantitative evaluation of the application of different energy optimization processes to select reference images from THuman3.0 [45]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 392, + 361, + 438 + ], + "lines": [ + { + "bbox": [ + 49, + 392, + 361, + 438 + ], + "spans": [ + { + "bbox": [ + 49, + 392, + 361, + 438 + ], + "type": "table", + "html": "
RandomOnly bpOnly oOnly vpo + vpo + bpvp + bpCOSMU
CD1.4851.4411.4211.4101.3461.3521.3261.304
Normal.1683.1598.1641.1654.1523.1529.1503.1475
P2S1.4241.4151.4031.4061.3321.2791.3081.279
", + "image_path": "5176f2b39c0bdf4b254fb0e95d996e7bfd2fe89edd45b83c7535059b290c6217.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 30, + 458, + 379, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 458, + 379, + 531 + ], + "spans": [ + { + "bbox": [ + 30, + 458, + 379, + 531 + ], + "type": "text", + "content": "Body part-based registration effectiveness: Directly using " + }, + { + "bbox": [ + 30, + 458, + 379, + 531 + ], + "type": "inline_equation", + "content": "N_{r}" + }, + { + "bbox": [ + 30, + 458, + 379, + 531 + ], + "type": "text", + "content": " selected references for the multi-view reconstruction deteriorates the performance since fusing different poses is challenging in the absence of known correspondences. To avoid this, we introduce a novel body part-based registration method to simulate a multi-view scenario. To prove its effectiveness, we modify the algorithm for registering the " + }, + { + "bbox": [ + 30, + 458, + 379, + 531 + ], + "type": "inline_equation", + "content": "N_{p} \\times N_{r}" + }, + { + "bbox": [ + 30, + 458, + 379, + 531 + ], + "type": "text", + "content": " coarse reference point clouds " + }, + { + "bbox": [ + 30, + 458, + 379, + 531 + ], + "type": "inline_equation", + "content": "\\zeta_{ref}" + }, + { + "bbox": [ + 30, + 458, + 379, + 531 + ], + "type": "text", + "content": " to the target " + }, + { + "bbox": [ + 30, + 458, + 379, + 531 + ], + "type": "inline_equation", + "content": "\\zeta_{t}" + }, + { + "bbox": [ + 30, + 458, + 379, + 531 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 36, + 535, + 380, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 535, + 380, + 586 + ], + "spans": [ + { + "bbox": [ + 36, + 535, + 380, + 586 + ], + "type": "text", + "content": "- No Reg: To demonstrate the effectiveness of registering the reference point clouds to the target, we omit the body part-based registration step. We adjust the reference selection to select only " + }, + { + "bbox": [ + 36, + 535, + 380, + 586 + ], + "type": "inline_equation", + "content": "N_r" + }, + { + "bbox": [ + 36, + 535, + 380, + 586 + ], + "type": "text", + "content": " references by averaging over " + }, + { + "bbox": [ + 36, + 535, + 380, + 586 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 36, + 535, + 380, + 586 + ], + "type": "text", + "content": " the criteria " + }, + { + "bbox": [ + 36, + 535, + 380, + 586 + ], + "type": "inline_equation", + "content": "v_g^p" + }, + { + "bbox": [ + 36, + 535, + 380, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 36, + 535, + 380, + 586 + ], + "type": "inline_equation", + "content": "b_g^p" + }, + { + "bbox": [ + 36, + 535, + 380, + 586 + ], + "type": "text", + "content": " in Eq. (4)." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 18 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 18 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 144, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 144, + 19 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 144, + 19 + ], + "type": "text", + "content": "M. Pesavento et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 65, + 342, + 108 + ], + "blocks": [ + { + "bbox": [ + 31, + 33, + 380, + 55 + ], + "lines": [ + { + "bbox": [ + 31, + 33, + 380, + 55 + ], + "spans": [ + { + "bbox": [ + 31, + 33, + 380, + 55 + ], + "type": "text", + "content": "Table 2: Quantitative evaluation of the application of different configurations of the body part-based registration algorithm on THuman3.0 [45]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 65, + 342, + 108 + ], + "lines": [ + { + "bbox": [ + 66, + 65, + 342, + 108 + ], + "spans": [ + { + "bbox": [ + 66, + 65, + 342, + 108 + ], + "type": "table", + "html": "
No RegNon rigidAverageNo RGBNo SegmCOSMU
CD3.5466.9321.3142.2492.1891.304
Normal.2280.4832.1569.2351.2114.1475
P2S2.8206.0241.3092.1692.0051.279
", + "image_path": "56d50d131e0eee1d1dffdf60e5ef248fd0cae2d45bbadce8c7e10aeba2761b42.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 51, + 119, + 362, + 266 + ], + "blocks": [ + { + "bbox": [ + 51, + 119, + 362, + 266 + ], + "lines": [ + { + "bbox": [ + 51, + 119, + 362, + 266 + ], + "spans": [ + { + "bbox": [ + 51, + 119, + 362, + 266 + ], + "type": "image", + "image_path": "d902fa8dc4d22485f0c1bc63bfd67d37853198d640d6d6361026aa401cc9753a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 30, + 276, + 381, + 311 + ], + "lines": [ + { + "bbox": [ + 30, + 276, + 381, + 311 + ], + "spans": [ + { + "bbox": [ + 30, + 276, + 381, + 311 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparisons between 3D shapes from THuman3.0 [45] reconstructed from 2D normal maps generated using different configurations of the body part-based registration algorithm. Front, back and side views of the shapes are shown." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 37, + 334, + 379, + 478 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 37, + 334, + 379, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 334, + 379, + 381 + ], + "spans": [ + { + "bbox": [ + 37, + 334, + 379, + 381 + ], + "type": "text", + "content": "- Non Rigid: Instead of applying rigid registration among the correspondent body parts of the target and reference point clouds, we apply the state-of-the-art work by Li and Harada [24] to perform non-rigid registration between the full-body point clouds. We use " + }, + { + "bbox": [ + 37, + 334, + 379, + 381 + ], + "type": "inline_equation", + "content": "N_r" + }, + { + "bbox": [ + 37, + 334, + 379, + 381 + ], + "type": "text", + "content": " references as in No reg." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 37, + 382, + 379, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 382, + 379, + 429 + ], + "spans": [ + { + "bbox": [ + 37, + 382, + 379, + 429 + ], + "type": "text", + "content": "- Average: COSMU registers a target body part with the corresponding body part of a reference selected specifically for that body part. In this configuration, we select only " + }, + { + "bbox": [ + 37, + 382, + 379, + 429 + ], + "type": "inline_equation", + "content": "N_r" + }, + { + "bbox": [ + 37, + 382, + 379, + 429 + ], + "type": "text", + "content": " references as in No reg and register all the body parts of the single reference for each group." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 37, + 430, + 379, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 430, + 379, + 466 + ], + "spans": [ + { + "bbox": [ + 37, + 430, + 379, + 466 + ], + "type": "text", + "content": "- No RGB: The body part classification model " + }, + { + "bbox": [ + 37, + 430, + 379, + 466 + ], + "type": "inline_equation", + "content": "f_{\\zeta}" + }, + { + "bbox": [ + 37, + 430, + 379, + 466 + ], + "type": "text", + "content": " processes a concatenation of 2D RGB images and segmentation masks. To demonstrate the advantages of using these data, we train and test " + }, + { + "bbox": [ + 37, + 430, + 379, + 466 + ], + "type": "inline_equation", + "content": "f_{\\zeta}" + }, + { + "bbox": [ + 37, + 430, + 379, + 466 + ], + "type": "text", + "content": " with only segmentation masks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 37, + 466, + 294, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 466, + 294, + 478 + ], + "spans": [ + { + "bbox": [ + 37, + 466, + 294, + 478 + ], + "type": "text", + "content": "- No Segm: we train and test " + }, + { + "bbox": [ + 37, + 466, + 294, + 478 + ], + "type": "inline_equation", + "content": "f_{\\zeta}" + }, + { + "bbox": [ + 37, + 466, + 294, + 478 + ], + "type": "text", + "content": " with only RGB images." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 30, + 488, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 488, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 488, + 380, + 584 + ], + "type": "text", + "content": "We apply these configurations to generate 2D multi-view normal maps used by the multi-view attention-based implicit model to reconstruct 3D human shapes. Quantitative (Tab. 2) and qualitative results (Fig. 4) show that COSMU configuration is the most effective. Multi-view reconstruction requires registered input since pixel-alignment across different views can be incorrect with unregistered views. Estimating correspondences between point clouds is challenging and the non-rigid registration cannot produce reliable results. Selecting " + }, + { + "bbox": [ + 30, + 488, + 380, + 584 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 30, + 488, + 380, + 584 + ], + "type": "text", + "content": " references for each body part facilitates the body part registration compared to using a sin" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 38, + 9, + 347, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 9, + 347, + 21 + ], + "spans": [ + { + "bbox": [ + 38, + 9, + 347, + 21 + ], + "type": "text", + "content": "COSMU: Complete 3D human shape from monocular unconstrained images" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 378, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 378, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 378, + 19 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 54, + 325, + 181 + ], + "blocks": [ + { + "bbox": [ + 32, + 33, + 378, + 44 + ], + "lines": [ + { + "bbox": [ + 32, + 33, + 378, + 44 + ], + "spans": [ + { + "bbox": [ + 32, + 33, + 378, + 44 + ], + "type": "text", + "content": "Table 3: Quantitative comparisons with related works for 3D human reconstruction." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 54, + 325, + 181 + ], + "lines": [ + { + "bbox": [ + 83, + 54, + 325, + 181 + ], + "spans": [ + { + "bbox": [ + 83, + 54, + 325, + 181 + ], + "type": "table", + "html": "
MethodsTHuman3.0X-Humans
CDNormalP2SCDNormalP2S
1 viewPiFU [40]1.698.17851.6592.516.25912.492
SuRS [37]1.573.16931.5542.406.24872.385
Baseline1.566.17321.5702.375.24862.351
2 viewsPiFu-HD [41]1.501.16631.5122.139.22762.121
ICON [50]1.474.16951.4632.210.23952.208
ECON [49]1.368.16321.3652.153.23052.114
COSMU-21.355.15931.3732.139.22582.104
4 viewsMulti-PaMIR [56]1.401.16501.3852.145.23192.132
MVP [58]1.348.15581.3652.063.22041.972
COSMU1.304.14751.2792.006.21461.929
", + "image_path": "597d8d14e54ca9470031e308290ea8884d3e79abc01f3963785a3df62071aa1e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 30, + 201, + 380, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 201, + 380, + 225 + ], + "spans": [ + { + "bbox": [ + 30, + 201, + 380, + 225 + ], + "type": "text", + "content": "gle reference. Leveraging RGB and segmentation masks improves the body part classification since the model learns both semantic and appearance information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 31, + 242, + 125, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 242, + 125, + 254 + ], + "spans": [ + { + "bbox": [ + 31, + 242, + 125, + 254 + ], + "type": "text", + "content": "4.2 Comparisons" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 260, + 381, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 260, + 381, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 260, + 381, + 584 + ], + "type": "text", + "content": "We conduct a comparative evaluation of COSMU against related works that reconstruct complete 3D human shapes from monocular RGB images. We evaluate MVP-Human [58] (MVP) and Multi-Pamir [56] that reconstruct 3D shapes from unconstrained images leveraging SMPL [31] in the process. The same " + }, + { + "bbox": [ + 30, + 260, + 381, + 584 + ], + "type": "inline_equation", + "content": "N_{r} = 3" + }, + { + "bbox": [ + 30, + 260, + 381, + 584 + ], + "type": "text", + "content": " images selected in the ablation study (Sec. 4.1) are used as additional views. We evaluate methods that estimate an unseen view of the subject from the target image (PiFU-HD [41], ICON [50] and ECON [49]). We then compare COSMU with PiFU [40] and SuRS [37], which process a single view without any information on hidden regions. To demonstrate the effect of using a lower number of references, we compare COSMU with COSMU-2, which is trained and evaluated by selecting " + }, + { + "bbox": [ + 30, + 260, + 381, + 584 + ], + "type": "inline_equation", + "content": "N_{r} = 1" + }, + { + "bbox": [ + 30, + 260, + 381, + 584 + ], + "type": "text", + "content": " reference (" + }, + { + "bbox": [ + 30, + 260, + 381, + 584 + ], + "type": "inline_equation", + "content": "\\theta_{c} = 180^{\\circ}" + }, + { + "bbox": [ + 30, + 260, + 381, + 584 + ], + "type": "text", + "content": "). Finally, we evaluate " + }, + { + "bbox": [ + 30, + 260, + 381, + 584 + ], + "type": "inline_equation", + "content": "f_{c}" + }, + { + "bbox": [ + 30, + 260, + 381, + 584 + ], + "type": "text", + "content": " (Baseline) that reconstructs the coarse target point clouds " + }, + { + "bbox": [ + 30, + 260, + 381, + 584 + ], + "type": "inline_equation", + "content": "\\zeta_{t}" + }, + { + "bbox": [ + 30, + 260, + 381, + 584 + ], + "type": "text", + "content": " without leveraging additional views. We use the released weights of ECON and PiFU-HD due to the absence of the training code while we train the other approaches with THuman2.0 [54]. Quantitative evaluation is conducted with the 3D shapes reconstructed from single images of THuman3.0 [45] and X-Humans [42]. The results presented in Tab. 3 demonstrate the superiority of COSMU across all the considered metrics for both datasets. The best scores overall are highlighted in red while the best scores for approaches that use 2 views are blue. COSMU achieves the highest accuracy while COSMU-2 outperforms approaches that rely on 2 views, demonstrating the benefit of leveraging views from 2D images rather than estimating them from the input. The worst results are obtained with single-view approaches. Qualitative results: Fig. 5 illustrates the 3D shapes reconstructed with the considered methods across several scenarios. The first set of shapes is reconstructed by leveraging a collection of images of the same individual in different poses from THuman3.0 [45]. The target and references of the second set are taken from the same monocular video of a person rotating from People-Snapshot [3]." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 146, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 146, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 146, + 20 + ], + "type": "text", + "content": "M. Pesavento et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 35, + 381, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 35, + 381, + 261 + ], + "spans": [ + { + "bbox": [ + 32, + 35, + 381, + 261 + ], + "type": "text", + "content": "In the third set, random images of a person in different poses and with a combination of both identical and different clothing are selected from the Internet. The fourth set of unconstrained images is created with images of the individual in different poses and clothing. This scenario is the most challenging since there are no images of the person in identical clothing. In all these scenarios, COSMU consistently reproduces higher-quality details in the regions of the 3D shape occluded in the target image, significantly outperforming related works. In addition, the quality of regions visible in the input image is comparable to that of related works. MVP struggles to reproduce details and introduces noise when the subject is wearing different clothing in the references (" + }, + { + "bbox": [ + 32, + 35, + 381, + 261 + ], + "type": "inline_equation", + "content": "3^{rd}" + }, + { + "bbox": [ + 32, + 35, + 381, + 261 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 32, + 35, + 381, + 261 + ], + "type": "inline_equation", + "content": "4^{th}" + }, + { + "bbox": [ + 32, + 35, + 381, + 261 + ], + "type": "text", + "content": " sets). ECON and PIFu-HD represent details on unseen parts of the model at a significantly lower quality compared to COSMU, which is the only approach that reconstructs the hood on the jumper (" + }, + { + "bbox": [ + 32, + 35, + 381, + 261 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 32, + 35, + 381, + 261 + ], + "type": "text", + "content": " set). The shapes reconstructed with PIFu-HD and ECON exhibit excessive smoothing in the regions of the human body occluded in the target image. The shapes reconstructed by ECON and MVP are influenced by the incorrect estimation and fitting of the SMPL model, resulting in an unnatural bending of the 3D body. SuRS reproduces high-quality details only in visible regions, with excessive smoothed reconstruction in other regions. See supplementary materials for additional results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 32, + 262, + 381, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 262, + 381, + 370 + ], + "spans": [ + { + "bbox": [ + 32, + 262, + 381, + 370 + ], + "type": "text", + "content": "Limitations: COSMU's ability to reconstruct high-quality details is contingent on the availability of reference images. COSMU cannot reproduce high-quality details in a region that is not visible in the reference images. Performance deteriorates when the reference images significantly differ from the target. See supplementary for a study on choosing references with varying degrees of similarity to the target. In some examples, COSMU cannot reproduce the same high-quality level of details in small regions of the body, such as the face and hands. The registration algorithm struggles to register small body parts while the multi-view transformer struggles to fuse multiple views of those regions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 389, + 119, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 389, + 119, + 402 + ], + "spans": [ + { + "bbox": [ + 33, + 389, + 119, + 402 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 32, + 416, + 381, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 416, + 381, + 547 + ], + "spans": [ + { + "bbox": [ + 32, + 416, + 381, + 547 + ], + "type": "text", + "content": "We introduce a novel framework to reconstruct complete 3D human shapes from a target RGB image by leveraging a collection of monocular unconstrained images without using parametric models. We first developed a novel multi-view scenario simulation module to generate a set of registered 2D normal maps of the target subject. A multi-view attention-based neural implicit model then estimates an implicit function that represents the subject 3D shape. The evaluation demonstrates that the proposed approach can be applied in various scenarios. Significantly higher quality details are reproduced in non-visible regions of the 3D shape compared to state-of-the-art works. In future works, a solution to enhance the quality of smaller body regions will be investigated and the approach will be extended to reconstruct temporally-coherent dynamic shapes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 32, + 548, + 387, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 548, + 387, + 573 + ], + "spans": [ + { + "bbox": [ + 32, + 548, + 387, + 573 + ], + "type": "text", + "content": "Acknowledgement: Work supported by BBC Prosperity Partnership AI4ME: Future Personalised Object-Based Media Experiences Delivered at Scale Anywhere EP/V038087." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 39, + 9, + 346, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 39, + 9, + 346, + 21 + ], + "spans": [ + { + "bbox": [ + 39, + 9, + 346, + 21 + ], + "type": "text", + "content": "COSMU: Complete 3D human shape from monocular unconstrained images" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 380, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 380, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 380, + 19 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 32, + 38, + 381, + 521 + ], + "blocks": [ + { + "bbox": [ + 32, + 38, + 381, + 521 + ], + "lines": [ + { + "bbox": [ + 32, + 38, + 381, + 521 + ], + "spans": [ + { + "bbox": [ + 32, + 38, + 381, + 521 + ], + "type": "image", + "image_path": "eaccd5ddcb08d29634a2442ef5c5d517236a62d04476943f55f261d8b5d04d48.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 30, + 528, + 381, + 571 + ], + "lines": [ + { + "bbox": [ + 30, + 528, + 381, + 571 + ], + "spans": [ + { + "bbox": [ + 30, + 528, + 381, + 571 + ], + "type": "text", + "content": "Fig. 5: Qualitative examples of 3D shapes reconstructed in several scenarios. Static images are from THuman3.0 [45] in the first example and from the Internet in the others; monocular videos are from People-Snapshot [3]. Front, back and side views are shown." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 10, + 144, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 10, + 144, + 19 + ], + "spans": [ + { + "bbox": [ + 66, + 10, + 144, + 19 + ], + "type": "text", + "content": "M. Pesavento et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 33, + 98, + 46 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 33, + 98, + 46 + ], + "spans": [ + { + "bbox": [ + 32, + 33, + 98, + 46 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 54, + 380, + 584 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 36, + 54, + 379, + 99 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 54, + 379, + 99 + ], + "spans": [ + { + "bbox": [ + 36, + 54, + 379, + 99 + ], + "type": "text", + "content": "1. Alldieck, T., Magnor, M., Bhatnagar, B.L., Theobalt, C., Pons-Moll, G.: Learning to reconstruct people in clothing from a single rgb camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1175-1186 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 36, + 99, + 379, + 131 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 99, + 379, + 131 + ], + "spans": [ + { + "bbox": [ + 36, + 99, + 379, + 131 + ], + "type": "text", + "content": "2. Alldieck, T., Magnor, M., Xu, W., Theobalt, C., Pons-Moll, G.: Detailed human avatars from monocular video. In: 2018 International Conference on 3D Vision (3DV). pp. 98-109. IEEE (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 37, + 131, + 380, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 131, + 380, + 164 + ], + "spans": [ + { + "bbox": [ + 37, + 131, + 380, + 164 + ], + "type": "text", + "content": "3. Alldieck, T., Magnor, M., Xu, W., Theobalt, C., Pons-Moll, G.: Video based reconstruction of 3d people models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8387-8397 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 37, + 164, + 380, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 164, + 380, + 196 + ], + "spans": [ + { + "bbox": [ + 37, + 164, + 380, + 196 + ], + "type": "text", + "content": "4. Alldieck, T., Pons-Moll, G., Theobalt, C., Magnor, M.: Tex2shape: Detailed full human body geometry from a single image. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2293-2303 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 37, + 196, + 380, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 196, + 380, + 228 + ], + "spans": [ + { + "bbox": [ + 37, + 196, + 380, + 228 + ], + "type": "text", + "content": "5. Alldieck, T., Zanfir, M., Sminchisescu, C.: Photorealistic monocular 3D reconstruction of humans wearing clothing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1506-1515 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 37, + 228, + 380, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 228, + 380, + 260 + ], + "spans": [ + { + "bbox": [ + 37, + 228, + 380, + 260 + ], + "type": "text", + "content": "6. Bogo, F., Kanazawa, A., Lassner, C., Gehler, P., Romero, J., Black, M.J.: Keep it SMPL: Automatic estimation of 3d human pose and shape from a single image. In: ECCV (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 37, + 260, + 380, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 260, + 380, + 293 + ], + "spans": [ + { + "bbox": [ + 37, + 260, + 380, + 293 + ], + "type": "text", + "content": "7. Cao, Y., Han, K., Wong, K.Y.K.: Sesdf: Self-evolved signed distance field for implicit 3d clothed human reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4647-4657 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 37, + 293, + 380, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 293, + 380, + 314 + ], + "spans": [ + { + "bbox": [ + 37, + 293, + 380, + 314 + ], + "type": "text", + "content": "8. Chen, J., Zhang, Y., Kang, D., Zhe, X., Bao, L., Jia, X., Lu, H.: Animatable neural radiance fields from monocular rgb videos. arXiv preprint arXiv:2106.13629 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 37, + 314, + 380, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 314, + 380, + 346 + ], + "spans": [ + { + "bbox": [ + 37, + 314, + 380, + 346 + ], + "type": "text", + "content": "9. Chen, L., Li, J., Huang, H., Guo, Y.: Crosshuman: learning cross-guidance from multi-frame images for human reconstruction. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 2483-2494 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 33, + 346, + 379, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 346, + 379, + 378 + ], + "spans": [ + { + "bbox": [ + 33, + 346, + 379, + 378 + ], + "type": "text", + "content": "10. Chen, Z., Zhang, H.: Learning implicit fields for generative shape modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5939-5948 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 33, + 378, + 380, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 378, + 380, + 422 + ], + "spans": [ + { + "bbox": [ + 33, + 378, + 380, + 422 + ], + "type": "text", + "content": "11. Corona, E., Zanfir, M., Alldieck, T., Bazavan, E.G., Zanfir, A., Sminchisescu, C.: Structured 3d features for reconstructing controllable avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16954-16964 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 33, + 422, + 379, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 422, + 379, + 454 + ], + "spans": [ + { + "bbox": [ + 33, + 422, + 379, + 454 + ], + "type": "text", + "content": "12. Gardner, A., Elhami, N., Selmic, R.R.: Classifying unordered feature sets with convolutional deep averaging networks. In: 2019 IEEE International Conference on Systems, Man and Cybernetics (SMC). pp. 3447-3453. IEEE (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 33, + 454, + 379, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 454, + 379, + 497 + ], + "spans": [ + { + "bbox": [ + 33, + 454, + 379, + 497 + ], + "type": "text", + "content": "13. Guo, C., Jiang, T., Chen, X., Song, J., Hilliges, O.: Vid2 avatar: 3d avatar reconstruction from videos in the wild via self-supervised scene decomposition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12858-12868 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 33, + 497, + 379, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 497, + 379, + 519 + ], + "spans": [ + { + "bbox": [ + 33, + 497, + 379, + 519 + ], + "type": "text", + "content": "14. He, T., Collomosse, J., Jin, H., Soatto, S.: Geo-pifu: Geometry and pixel aligned implicit functions for single-view human reconstruction (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 33, + 519, + 379, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 519, + 379, + 551 + ], + "spans": [ + { + "bbox": [ + 33, + 519, + 379, + 551 + ], + "type": "text", + "content": "15. He, T., Xu, Y., Saito, S., Soatto, S., Tung, T.: Arch++: Animation-ready clothed human reconstruction revisited. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11046-11056 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 33, + 551, + 379, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 551, + 379, + 584 + ], + "spans": [ + { + "bbox": [ + 33, + 551, + 379, + 584 + ], + "type": "text", + "content": "16. Huang, H., Chen, L., Wang, X.: Unconfuse: Avatar reconstruction from unconstrained images. In: European Conference on Computer Vision. pp. 376-381. Springer (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 38, + 9, + 347, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 9, + 347, + 21 + ], + "spans": [ + { + "bbox": [ + 38, + 9, + 347, + 21 + ], + "type": "text", + "content": "COSMU: Complete 3D human shape from monocular unconstrained images" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 35, + 380, + 583 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 33, + 35, + 379, + 57 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 35, + 379, + 57 + ], + "spans": [ + { + "bbox": [ + 33, + 35, + 379, + 57 + ], + "type": "text", + "content": "17. Huang, Y., Yi, H., Xiu, Y., Liao, T., Tang, J., Cai, D., Thies, J.: Tech: Text-guided reconstruction of lifelike clothed humans. arXiv preprint arXiv:2308.08545 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 32, + 59, + 379, + 90 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 59, + 379, + 90 + ], + "spans": [ + { + "bbox": [ + 32, + 59, + 379, + 90 + ], + "type": "text", + "content": "18. Huang, Z., Xu, Y., Lassner, C., Li, H., Tung, T.: Arch: Animatable reconstruction of clothed humans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3093-3102 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 91, + 380, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 91, + 380, + 135 + ], + "spans": [ + { + "bbox": [ + 33, + 91, + 380, + 135 + ], + "type": "text", + "content": "19. Jayasundara, V., Agrawal, A., Heron, N., Shrivastava, A., Davis, L.S.: Flexnerf: Photorealistic free-viewpoint rendering of moving humans from sparse views. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21118-21127 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 136, + 379, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 136, + 379, + 169 + ], + "spans": [ + { + "bbox": [ + 33, + 136, + 379, + 169 + ], + "type": "text", + "content": "20. Jiang, B., Hong, Y., Bao, H., Zhang, J.: Selfrecon: Self reconstruction your digital avatar from monocular video. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 170, + 379, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 170, + 379, + 202 + ], + "spans": [ + { + "bbox": [ + 33, + 170, + 379, + 202 + ], + "type": "text", + "content": "21. Kanazawa, A., Black, M.J., Jacobs, D.W., Malik, J.: End-to-end recovery of human shape and pose. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7122-7131 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 204, + 379, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 204, + 379, + 225 + ], + "spans": [ + { + "bbox": [ + 33, + 204, + 379, + 225 + ], + "type": "text", + "content": "22. Kocabas, M., Athanasiou, N., Black, M.J.: Vibe: Video inference for human body pose and shape estimation. In: CVPR (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 33, + 226, + 379, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 226, + 379, + 247 + ], + "spans": [ + { + "bbox": [ + 33, + 226, + 379, + 247 + ], + "type": "text", + "content": "23. Li, C., Lin, J., Lee, G.H.: Ghunerf: Generalizable human nerf from a monocular video. arXiv preprint arXiv:2308.16576 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 33, + 249, + 379, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 249, + 379, + 270 + ], + "spans": [ + { + "bbox": [ + 33, + 249, + 379, + 270 + ], + "type": "text", + "content": "24. Li, Y., Harada, T.: Non-rigid point cloud registration with neural deformation pyramid. arXiv preprint arXiv:2205.12796 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 33, + 271, + 379, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 271, + 379, + 304 + ], + "spans": [ + { + "bbox": [ + 33, + 271, + 379, + 304 + ], + "type": "text", + "content": "25. Li, Z., Yu, T., Zheng, Z., Guo, K., Liu, Y.: Posefusion: Pose-guided selective fusion for single-view human volumetric capture. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14162-14172 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 33, + 305, + 379, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 305, + 379, + 337 + ], + "spans": [ + { + "bbox": [ + 33, + 305, + 379, + 337 + ], + "type": "text", + "content": "26. Li, Z., Zheng, Z., Zhang, H., Ji, C., Liu, Y.: Avatarcap: Animatable avatar conditioned monocular human volumetric capture. In: European Conference on Computer Vision. pp. 322-341. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 33, + 338, + 379, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 338, + 379, + 370 + ], + "spans": [ + { + "bbox": [ + 33, + 338, + 379, + 370 + ], + "type": "text", + "content": "27. Li, Z., Oskarsson, M., Heyden, A.: Detailed 3d human body reconstruction from multi-view images combining voxel super-resolution and learned implicit representation. Applied Intelligence 52(6), 6739-6759 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 33, + 372, + 379, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 372, + 379, + 415 + ], + "spans": [ + { + "bbox": [ + 33, + 372, + 379, + 415 + ], + "type": "text", + "content": "28. Liao, T., Zhang, X., Xiu, Y., Yi, H., Liu, X., Qi, G.J., Zhang, Y., Wang, X., Zhu, X., Lei, Z.: High-fidelity clothed avatar reconstruction from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8662-8672 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 33, + 416, + 379, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 416, + 379, + 460 + ], + "spans": [ + { + "bbox": [ + 33, + 416, + 379, + 460 + ], + "type": "text", + "content": "29. Liu, W., Piao, Z., Min, J., Luo, W., Ma, L., Gao, S.: Liquid warping gan: A unified framework for human motion imitation, appearance transfer and novel view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5904-5913 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 33, + 461, + 379, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 461, + 379, + 482 + ], + "spans": [ + { + "bbox": [ + 33, + 461, + 379, + 482 + ], + "type": "text", + "content": "30. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM TOG 34(6), 248 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 33, + 484, + 379, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 484, + 379, + 516 + ], + "spans": [ + { + "bbox": [ + 33, + 484, + 379, + 516 + ], + "type": "text", + "content": "31. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia) 34(6), 248:1-248:16 (Oct 2015)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 33, + 517, + 379, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 517, + 379, + 539 + ], + "spans": [ + { + "bbox": [ + 33, + 517, + 379, + 539 + ], + "type": "text", + "content": "32. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics 21(4), 163-169 (1987)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 33, + 540, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 540, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 540, + 379, + 583 + ], + "type": "text", + "content": "33. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4460-4470 (2019)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 144, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 144, + 19 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 144, + 19 + ], + "type": "text", + "content": "M. Pesavento et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 31, + 35, + 380, + 584 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 31, + 35, + 380, + 79 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 35, + 380, + 79 + ], + "spans": [ + { + "bbox": [ + 31, + 35, + 380, + 79 + ], + "type": "text", + "content": "34. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 165-174 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 31, + 79, + 380, + 122 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 79, + 380, + 122 + ], + "spans": [ + { + "bbox": [ + 31, + 79, + 380, + 122 + ], + "type": "text", + "content": "35. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10975-10985 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 32, + 122, + 380, + 154 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 122, + 380, + 154 + ], + "spans": [ + { + "bbox": [ + 32, + 122, + 380, + 154 + ], + "type": "text", + "content": "36. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: CVPR (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 32, + 154, + 380, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 154, + 380, + 175 + ], + "spans": [ + { + "bbox": [ + 32, + 154, + 380, + 175 + ], + "type": "text", + "content": "37. Pesavento, M., Volino, M., Hilton, A.: Super-resolution 3d human shape from a single low-resolution image. In: ECCV (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 32, + 175, + 380, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 175, + 380, + 218 + ], + "spans": [ + { + "bbox": [ + 32, + 175, + 380, + 218 + ], + "type": "text", + "content": "38. Pesavento, M., Xu, Y., Sarafianos, N., Maier, R., Wang, Z., Yao, C.H., Volino, M., Boyer, E., Hilton, A., Tung, T.: Anim: Accurate neural implicit model for human reconstruction from a single rgb-d image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5448-5458 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 32, + 218, + 380, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 218, + 380, + 239 + ], + "spans": [ + { + "bbox": [ + 32, + 218, + 380, + 239 + ], + "type": "text", + "content": "39. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 32, + 239, + 380, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 239, + 380, + 282 + ], + "spans": [ + { + "bbox": [ + 32, + 239, + 380, + 282 + ], + "type": "text", + "content": "40. Saito, S., Huang, Z., Natsume, R., Morishima, S., Kanazawa, A., Li, H.: Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2304-2314 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 32, + 282, + 380, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 282, + 380, + 325 + ], + "spans": [ + { + "bbox": [ + 32, + 282, + 380, + 325 + ], + "type": "text", + "content": "41. Saito, S., Simon, T., Saragih, J., Joo, H.: Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 84-93 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 32, + 325, + 380, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 325, + 380, + 358 + ], + "spans": [ + { + "bbox": [ + 32, + 325, + 380, + 358 + ], + "type": "text", + "content": "42. Shen, K., Guo, C., Kaufmann, M., Zarate, J.J., Valentin, J., Song, J., Hilliges, O.: X-ayar: Expressive human avatars. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16911-16921 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 32, + 358, + 380, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 358, + 380, + 390 + ], + "spans": [ + { + "bbox": [ + 32, + 358, + 380, + 390 + ], + "type": "text", + "content": "43. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems 34, 6087-6101 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 32, + 389, + 380, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 389, + 380, + 422 + ], + "spans": [ + { + "bbox": [ + 32, + 389, + 380, + 422 + ], + "type": "text", + "content": "44. Su, H., Maji, S., Kalogerakis, E., Learned-Miller, E.: Multi-view convolutional neural networks for 3d shape recognition. In: Proceedings of the IEEE international conference on computer vision. pp. 945-953 (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 32, + 422, + 380, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 422, + 380, + 455 + ], + "spans": [ + { + "bbox": [ + 32, + 422, + 380, + 455 + ], + "type": "text", + "content": "45. Su, Z., Yu, T., Wang, Y., Liu, Y.: Deepcloth: Neural garment representation for shape and style editing. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(2), 1581-1593 (2023). https://doi.org/10.1109/TPAMI.2022.3168569" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 32, + 455, + 380, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 455, + 380, + 487 + ], + "spans": [ + { + "bbox": [ + 32, + 455, + 380, + 487 + ], + "type": "text", + "content": "46. Te, G., Li, X., Li, X., Wang, J., Hu, W., Lu, Y.: Neural capture of animatable 3d human from monocular video. In: European Conference on Computer Vision. pp. 275-291. Springer (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 32, + 487, + 380, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 487, + 380, + 519 + ], + "spans": [ + { + "bbox": [ + 32, + 487, + 380, + 519 + ], + "type": "text", + "content": "47. Varol, G., Ceylan, D., Russell, B., Yang, J., Yumer, E., Laptev, I., Schmid, C.: Bodynet: Volumetric inference of 3d human body shapes. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 20-36 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 32, + 519, + 380, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 519, + 380, + 551 + ], + "spans": [ + { + "bbox": [ + 32, + 519, + 380, + 551 + ], + "type": "text", + "content": "48. Wang, S., Schwarz, K., Geiger, A., Tang, S.: Arah: Animatable volume rendering of articulated human sdfs. In: European conference on computer vision. pp. 1-19. Springer (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 32, + 551, + 380, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 551, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 32, + 551, + 380, + 584 + ], + "type": "text", + "content": "49. Xiu, Y., Yang, J., Cao, X., Tzionas, D., Black, M.J.: Econ: Explicit clothed humans optimized via normal integration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 512-523 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 38, + 9, + 347, + 21 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 9, + 347, + 21 + ], + "spans": [ + { + "bbox": [ + 38, + 9, + 347, + 21 + ], + "type": "text", + "content": "COSMU: Complete 3D human shape from monocular unconstrained images" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 380, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 380, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 380, + 19 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 35, + 379, + 386 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 33, + 35, + 379, + 57 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 35, + 379, + 57 + ], + "spans": [ + { + "bbox": [ + 33, + 35, + 379, + 57 + ], + "type": "text", + "content": "50. Xiu, Y., Yang, J., Tzionas, D., Black, M.J.: Icon: Implicit clothed humans obtained from normals. In: CVPR (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 32, + 57, + 379, + 90 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 57, + 379, + 90 + ], + "spans": [ + { + "bbox": [ + 32, + 57, + 379, + 90 + ], + "type": "text", + "content": "51. Xu, X., Loy, C.C.: 3d human texture estimation from a single image with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13849-13858 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 91, + 379, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 91, + 379, + 123 + ], + "spans": [ + { + "bbox": [ + 33, + 91, + 379, + 123 + ], + "type": "text", + "content": "52. Yang, X., Chen, X., Wang, S., Gao, D., Han, X., Wang, B.: Have-fun: Human avatar reconstruction from few-shot unconstrained images. arXiv preprint arXiv:2311.15672 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 123, + 379, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 123, + 379, + 156 + ], + "spans": [ + { + "bbox": [ + 33, + 123, + 379, + 156 + ], + "type": "text", + "content": "53. Yu, H., Cheang, C., Fu, Y., Xue, X.: Multi-view shape generation for a 3d human-like body. ACM Transactions on Multimedia Computing, Communications and Applications 19(1), 1-22 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 156, + 379, + 188 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 156, + 379, + 188 + ], + "spans": [ + { + "bbox": [ + 33, + 156, + 379, + 188 + ], + "type": "text", + "content": "54. Yu, T., Zheng, Z., Guo, K., Liu, P., Dai, Q., Liu, Y.: Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR2021) (June 2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 189, + 379, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 189, + 379, + 221 + ], + "spans": [ + { + "bbox": [ + 33, + 189, + 379, + 221 + ], + "type": "text", + "content": "55. Yu, Z., Zhang, L., Xu, Y., Tang, C., Tran, L., Keskin, C., Park, H.S.: Multiview human body reconstruction from uncalibrated cameras. Advances in Neural Information Processing Systems 35, 7879-7891 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 33, + 222, + 379, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 222, + 379, + 253 + ], + "spans": [ + { + "bbox": [ + 33, + 222, + 379, + 253 + ], + "type": "text", + "content": "56. Zheng, Z., Yu, T., Liu, Y., Dai, Q.: Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 33, + 255, + 379, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 255, + 379, + 287 + ], + "spans": [ + { + "bbox": [ + 33, + 255, + 379, + 287 + ], + "type": "text", + "content": "57. Zheng, Z., Yu, T., Wei, Y., Dai, Q., Liu, Y.: Deephuman: 3d human reconstruction from a single image. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7739-7749 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 33, + 287, + 379, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 287, + 379, + 320 + ], + "spans": [ + { + "bbox": [ + 33, + 287, + 379, + 320 + ], + "type": "text", + "content": "58. Zhu, X., Liao, T., Lyu, J., Yan, X., Wang, Y., Guo, K., Cao, Q., Li, S.Z., Lei, Z.: Mvp-human dataset for 3d human avatar reconstruction from unconstrained frames. arXiv preprint arXiv:2204.11184 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 33, + 321, + 379, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 321, + 379, + 353 + ], + "spans": [ + { + "bbox": [ + 33, + 321, + 379, + 353 + ], + "type": "text", + "content": "59. Zins, P., Xu, Y., Boyer, E., Wuhrer, S., Tung, T.: Data-driven 3d reconstruction of dressed humans from sparse views. In: 2021 International Conference on 3D Vision (3DV). pp. 494-504. IEEE (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 33, + 354, + 379, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 354, + 379, + 386 + ], + "spans": [ + { + "bbox": [ + 33, + 354, + 379, + 386 + ], + "type": "text", + "content": "60. Zins, P., Xu, Y., Boyer, E., Wuhrer, S., Tung, T.: Multi-view reconstruction using signed ray distance functions (srdf). In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16696-16706 (2023)" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 144, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 144, + 19 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 144, + 19 + ], + "type": "text", + "content": "M. Pesavento et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CPM_ Class-conditional Prompting Machine for Audio-visual Segmentation/462b4fbe-3dda-4385-88f4-05811a4623a9_content_list.json b/2024/CPM_ Class-conditional Prompting Machine for Audio-visual Segmentation/462b4fbe-3dda-4385-88f4-05811a4623a9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1c267f7ba06b99b30642771dfe53a42531e346fe --- /dev/null +++ b/2024/CPM_ Class-conditional Prompting Machine for Audio-visual Segmentation/462b4fbe-3dda-4385-88f4-05811a4623a9_content_list.json @@ -0,0 +1,1569 @@ +[ + { + "type": "text", + "text": "CPM: Class-conditional Prompting Machine for Audio-visual Segmentation", + "text_level": 1, + "bbox": [ + 228, + 140, + 774, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuanhong Chen1, Chong Wang1, Yuyuan Liu1, Hu Wang2, and Gustavo Carneiro3", + "bbox": [ + 233, + 210, + 767, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Australian Institute for Machine Learning, University of Adelaide, Australia", + "bbox": [ + 228, + 253, + 771, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Mohamed bin Zayed University of Artificial Intelligence, United Arab Emirates", + "bbox": [ + 228, + 268, + 771, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Centre for Vision, Speech and Signal Processing, University of Surrey yuanhong.chen@adelaide.edu.au", + "bbox": [ + 261, + 281, + 738, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Audio-visual segmentation (AVS) is an emerging task that aims to accurately segment sounding objects based on audio-visual cues. The success of AVS learning systems depends on the effectiveness of cross-modal interaction. Such a requirement can be naturally fulfilled by leveraging transformer-based segmentation architecture due to its inherent ability to capture long-range dependencies and flexibility in handling different modalities. However, the inherent training issues of transformer-based methods, such as the low efficacy of cross-attention and unstable bipartite matching, can be amplified in AVS, particularly when the learned audio query does not provide a clear semantic clue. In this paper, we address these two issues with the new Class-conditional Prompting Machine (CPM). CPM improves the bipartite matching with a learning strategy combining class-agnostic queries with class-conditional queries. The efficacy of cross-modal attention is upgraded with new learning objectives for the audio, visual and joint modalities. We conduct experiments on AVS benchmarks, demonstrating that our method achieves state-of-the-art (SOTA) segmentation accuracy4.", + "bbox": [ + 261, + 347, + 740, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Audio-visual Learning $\\cdot$ Segmentation $\\cdot$ Multi-modal Learning", + "bbox": [ + 259, + 595, + 740, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 650, + 375, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recognition and integration of auditory and visual data are fundamental to human cognitive processes, playing a critical role in facilitating meaningful communication [44]. Audio-visual segmentation (AVS) is an emerging cross-modal reasoning task that mimics such cognitive processes, aiming to localize visual objects based on audio-visual cues. AVS has many applications, such as the automatic localisation and identification of sounding objects to improve the accessibility of videos for the blind and visually impaired [34]. A major challenge in AVS is achieving effective cross-modal interaction between sound and visual", + "bbox": [ + 212, + 681, + 787, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This project is supported by the Australian Research Council (ARC) through grant FT190100525.", + "bbox": [ + 217, + 810, + 785, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6fa6c9fc6131d3e7717ec93b0fa998a611d80393b2f55380e0e67d2b8777a19d.jpg", + "image_caption": [ + "Fig. 1: Comparing conventional AVS methods [7, 14] with our CPM approach, CPM inherits the class-agnostic query from transformer-based methods and integrates class-conditional prompts sampled from the learned joint-modal data distribution to achieve three objectives: 1) learn disentangled audio partitioning, 2) facilitate semantic-guided object identification, and 3) promote more explicit audio-visual contrastive learning." + ], + "image_footnote": [], + "bbox": [ + 261, + 145, + 736, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "objects [46]. Many AVS methods [7, 37, 57] generally adopt a traditional per-pixel classification framework [8, 52], utilising early fusion strategies (e.g., cross-attention fusion [49]) together with an FCN decoder [4, 35] to make predictions. Such per-pixel design has achieved good performance, but it tends to underutilise the audio data due to its lower informativeness compared with the visual data [7, 46]. Contrastive learning can mitigate this issue by matching informative audio-visual pairs [7]. However, another critical limitation of the per-pixel design is its failure to capture instance-level visual information, resulting in inconsistent segmentation predictions within or between frames in a video sequence [52].", + "bbox": [ + 212, + 416, + 787, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These two issues have been addressed by transformer-based AVS methods designed to capture instance-level information and to rely on more effective contrastive learning [14, 24, 28, 30, 31, 52, 54]. Nevertheless, these AVS approaches still show slow convergence and relatively poor accuracy [23, 33, 55], primarily attributed to the low efficacy of the cross-attention [48] and the unstable bipartite matching [23]. Solutions for these problems are based on integrated masked-attention [8], which focuses on features around predicted segments, and anchor de-noising [23] to reconstruct the noisy ground-truth bounding box detection. However, these solutions may not work well in AVS because of the weak constraint provided by global audio features that contain a mixture of sound sources [14, 26, 30] resulting in more instability during training. It is worth noting that all methods above have the common issue of relying on class-agnostic prompts that provide little guidance to the bipartite matching process, thereby reducing training efficacy.", + "bbox": [ + 212, + 553, + 787, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce the Class-conditional Prompting Machine (CPM), an audio-visual segmentation training approach that leverages class-conditional prompts to enhance bipartite matching stability and improve cross-modal attention efficacy. To enhance bipartite matching, we introduce a novel learning method, combining class-agnostic queries [8,14,26] with class-conditional queries,", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sampled from our iteratively updated generative model of class-specific embeddings, with the former queries matched to ground-truth labels using the Hungarian Algorithm [1], followed by individual processing of class-conditional queries in two modalities. To improve cross-modal attention efficacy, new learning objectives are proposed for both audio and visual modalities. In the audio modality (audio conditional prompting, denoted as ACP), the original spectrogram is corrupted with auditory off-the-screen noise and class-conditional queries are employed to reconstruct the original spectrogram, while in the visual modality (visual conditional prompting, denoted as VCP), noisy class-conditional queries sampled from our generative model are used to probe semantically similar content in the image space. To further upgrade the performance of cross-modal attention, we introduce a new prompting-based audio-visual contrastive learning (PCL) task, guided by class-specific queries to densely constrain the cross-modal representations. To summarise, our main contributions are:", + "bbox": [ + 212, + 146, + 787, + 357 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A new AVS training approach to enhance bipartite matching stability and enhance the efficacy of cross-modal attention. Our core innovation lies in the development of a Class-conditional Prompting Machine (CPM).", + "- To improve the bipartite matching, we propose a new AVS learning strategy that combines class-agnostic queries with class-conditional queries, sampled from our iteratively updated generative model of class-specific embeddings.", + "- The efficacy of cross-modal attention is upgraded with new learning objectives for the audio, visual and joint modalities. For audio, we present ACP that perturbs the original spectrogram and uses class-conditional queries to reconstruct the spectrogram; for visual, we introduce VCP that explores noisy class-conditional queries sampled from our generative model to probe the corresponding semantic in visual feature map. To enhance the cross-modal attention efficacy further, we propose PCL, consisting of a new prompting-based joint (audio-visual) contrastive learning task." + ], + "bbox": [ + 225, + 368, + 784, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We firstly show the effectiveness of our CPM model through rigorous evaluation on established benchmarks such as AVSBench-Objects [57] and AVSBench-Semantics [56]. Furthermore, we extend our evaluation by including VPO synthetic benchmarks [7], aiming to enhance our comprehension of AVS methods' capacity to capture audio-visual correlations. Our findings across these benchmarks consistently demonstrate that our approach yields better classification accuracy compared to existing methods.", + "bbox": [ + 212, + 590, + 787, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 720, + 397, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Transformer-based architecture", + "text_level": 1, + "bbox": [ + 215, + 753, + 519, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformer-based methods have shown promising performance in detection [1] and segmentation [8,9] benchmarks. The fundamental concept is to leverage the object query to probe image features from the output of transformer encoders and bipartite graph matching to perform set-based box/mask prediction. It is", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Class-conditional Prompting Machine (CPM)", + "bbox": [ + 428, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "also evidenced that such a framework can benefit multi-modal learning due to its attention mechanism and flexibility in data modelling [53]. Despite its successful application in various domains [1, 8], such frameworks have also been reported to exhibit a poorer convergence rate compared to traditional CNN-based methods [8,23,33]. In exploring the reasons behind the poor convergence, previous studies have emphasized enhancing the interpretability of the learnable query [33, 39] (often referring to them as anchor points/boxes). Alternatively, the utilisation of denoising methods [23, 55] has also been adopted to facilitate the learning of bounding box offset [23] and improve the utilisation of the transformer decoder layers [55]. Furthermore, masked attention has proven effective in enhancing both convergence rate and model performance [8]. The successful implementation of denoising methods and masked attention mechanisms has inspired us to devise a mitigation strategy for bipartite matching that promotes improved cross-modal understanding within the cross-attention process.", + "bbox": [ + 212, + 146, + 787, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Audio-visual Segmentation (AVS)", + "text_level": 1, + "bbox": [ + 215, + 369, + 540, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Audio-visual Segmentation is a dense classification task or detection of sounding visual objects in videos, using image sequences and audio cues. Zhou et al. [57] introduced the AVSBench-Object and AVSBench-Semantics benchmarks [56], which enable the evaluations of single-source and multi-source AVS tasks for salient and multi-class object segmentation. To address concerns related to the high annotation cost and dataset diversity, Chen et al. [7] introduced a cost-effective strategy to build a relatively unbiased AVS dataset, named Visual Post-production (VPO). Mainstream AVS methods use audio as the reference query [7, 14, 19, 24, 29–32] or prompt [43, 51]. For example, some methods adopt MaskFormer [9] or segment anything model (SAM) [22] to perform image segmentation using audio queries or encoded audio prompts and cross-attention layers. These methods benefit from the attention mechanism's ability and mask-classification's features to capture long-range dependencies and enhance image segmentation ability [9], spatial-temporal reasoning [24] and task-related features [14, 24, 29, 30, 37, 57]. The adaptation of the Maskformer-based framework for AVS relies on methods to encourage the audio-visual semantic alignment [24, 30, 31, 54]. Such strategy mitigates the poor audio semantic information caused by modality imbalance [5] and learns disentangled multi-modal representations [26, 54] via vector quantization [15] and bidirectional attention mechanism. Even though the query-based transformer architecture [1] has shown great success in semantic segmentation tasks, certain weaknesses such as low efficacy of cross-attention and unstable bipartite matching process [23] are still only partially addressed. These two issues may be exacerbated in AVS due to the poor audio semantic information.", + "bbox": [ + 212, + 395, + 787, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Audio-visual Contrastive Learning", + "text_level": 1, + "bbox": [ + 215, + 768, + 547, + 784 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Contrastive Learning has shown promising results in audio-visual learning (AVL) methods [2, 7, 18, 40, 41]. These methods bring together augmented representations from the same instance as positives while separating representation pairs", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from different instances as negatives within a batch. A reported issue with current AVL contrastive learning is its reliance on self-supervision [6] to connect audio and visual representations of the same class. To overcome this issue, CAVP [7] propose a supervised contrastive learning [21, 25, 50] method that mines informative contrastive pairs from arbitrary audio-visual pairings to constrain the learning of audio-visual embeddings. Nevertheless, these previous approaches predominantly leverage global audio representations, thereby limiting the model's capacity to discern individual audio sources. Consequently, this constraint reduces the model's effectiveness in scenarios involving multiple sound sources. In our work, we employ class-specific queries to retrieve the corresponding audio representations, which facilitates a more explicit form of contrastive learning between audio and visual modalities.", + "bbox": [ + 212, + 146, + 787, + 328 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 349, + 330, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We denote a multi-class audio-visual dataset as $\\mathcal{D} = \\{(\\mathbf{a}_i, \\mathbf{x}_i, \\mathbf{y}_i, \\mathbf{t}_i))\\}_{i=1}^{|\\mathcal{D}|}$ , where $\\mathbf{x}_i \\in \\mathcal{X} \\subset \\mathbb{R}^{H \\times W \\times 3}$ is an RGB image with resolution $H \\times W$ , $\\mathbf{a} \\in \\mathcal{A} \\subset \\mathbb{R}^{T \\times F}$ denotes the magnitude spectrogram of the audio data with $T$ time and $F$ frequency bins, $\\mathbf{y}_i \\in \\mathcal{Y} \\subset \\{0,1\\}^{H \\times W \\times C}$ denotes the pixel-level ground truth for the $C$ classes (the background class is included in these $C$ classes), and $\\mathbf{t}_i \\in \\mathcal{Y} \\subset \\{0,1\\}^C$ is a multi-label ground truth audio annotation. We use Mask2former [8] as the segmentation framework.", + "bbox": [ + 212, + 378, + 787, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminaries using Class-agnostic Queries", + "text_level": 1, + "bbox": [ + 214, + 496, + 612, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Like other Maskformer-based methods [14,30,31,52], we aim to learn the parameters $\\theta \\in \\Theta$ for the model $f_{\\theta}:\\mathcal{X}\\times \\mathcal{A}\\to [0,1]^{H\\times W\\times C}$ , which comprises the image and audio backbones that extract features with $\\mathbf{u}_a = f_\\gamma (\\mathbf{a})$ and $\\mathbf{u}_v = f_\\phi (\\mathbf{x})$ , respectively, where $\\gamma ,\\phi \\in \\theta$ , and $\\mathbf{u}_a,\\mathbf{u}_v\\in \\mathcal{U}$ , with $\\mathcal{U}$ denoting a unified feature space. A set of learnable query features (comprising the object query feature and positional query embeddings) are defined as the joint-modal output embeddings similar to the Perceiver model [20]. We define the class-agnostic query feature as $\\mathbf{q}\\in \\mathcal{Q}\\subset \\mathbb{R}^{N\\times D_q}$ , where $N$ denotes the number of class-agnostic queries, and $D_{q}$ represents the dimensionality of the feature space. As depicted in Fig. 2, given $\\mathbf{q}$ , we aim to group pixels with matched semantic information from $\\mathbf{u}_a$ and $\\mathbf{u}_v$ through consecutive transformer decoder layers [8,49] and generate $N$ mask embeddings $\\tilde{\\mathbf{q}}\\in \\mathcal{Q}$ and pixel embeddings $\\tilde{\\mathbf{u}}_v\\in \\mathbb{R}^{H\\times W\\times D_q}$ . Then, the model independently predicts the embeddings into $N$ set of class predictions (via Softmax(·)) and mask predictions (via Sigmoid(·)) denoted as $\\mathcal{E}_{\\mathrm{pred}} = \\left\\{\\left(\\mathbf{m}_i,\\mathbf{p}_i\\right)\\right\\}_{i = 1}^N$ , where $\\mathbf{m}\\in \\mathcal{M}\\subset \\{0,1\\}^{H\\times W}$ and $\\mathbf{p}\\in \\mathcal{P}\\subset \\{0,1\\}^C$ . We denote the ground-truth set derived from the training set $\\mathcal{D}$ with $\\mathcal{E}_{\\mathrm{gt}} = \\left\\{\\left(\\mathbf{m}_i^{\\mathrm{gt}},\\mathbf{p}_i^{\\mathrm{gt}}\\right)\\right\\}_{i = 1}^{N^{\\mathrm{gt}}}$ . During training, we use the Hungarian algorithm to perform optimal matching between $\\mathcal{E}_{\\mathrm{pred}}$ and $\\mathcal{E}_{\\mathrm{gt}}$ [1,8] to facilitate the label assignment. Since $|\\mathcal{E}_{\\mathrm{pred}}| > |\\mathcal{E}_{\\mathrm{gt}}|$ , we pad $\\mathcal{E}_{\\mathrm{gt}}$ with no-object class $\\varnothing$ to achieve one-to-one matching [8].", + "bbox": [ + 212, + 520, + 787, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The loss $\\ell$ to train the model $f_{\\theta}(.)$ includes a cross-entropy query classification loss $\\ell_{ce}$ , and a binary mask loss $\\ell_{mask} = \\ell_{focal} + \\ell_{dice}$ , which combines a focal loss", + "bbox": [ + 212, + 810, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Class-conditional Prompting Machine (CPM)", + "bbox": [ + 428, + 113, + 730, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/81a0499895e82667d6bdd7d30964ddf0da46833ba0f91263495d8710a773d9dc.jpg", + "image_caption": [ + "Fig. 2: Illustration of our CPM method. Starting with a scene with a mixture of sound sources including Male, Female and Guitar, the training alternates between the use of learnable class-agnostic queries, and queries sampled from a list of class-specific query features from the GMM, denoted as $\\mathbf{z}^{\\mathrm{male}}$ , $\\mathbf{z}^{\\mathrm{female}}$ and $\\mathbf{z}^{\\mathrm{guitar}}$ . The overall training objective is composed of three learning tasks: 1) in audio conditional prompting (ACP), we aim to use $\\mathbf{z}^{\\mathrm{male}}$ , $\\mathbf{z}^{\\mathrm{female}}$ and $\\mathbf{z}^{\\mathrm{guitar}}$ to recover the original magnitude spectrogram $\\mathbf{a}_i$ from the noise spectrogram (i.e., $\\mathbf{a}_i + \\mathbf{a}_j$ ) that is corrupted by another Dog audio signal $\\mathbf{a}_j$ ; 2) a visual conditional prompting (VCP) that aim to probe the corresponding pixels w.r.t to the class-specific query features; and 3) a contrastive learning task that target to densely constrain the audio and visual representations. For training, both the CPM Workflow, indicated by the orange arrow, and the Class-agnostic Workflow, marked by the black arrow, are utilized. However, only the Class-agnostic Workflow is used for inference." + ], + "image_footnote": [], + "bbox": [ + 261, + 145, + 738, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\ell_{focal}$ and a dice loss $\\ell_{dice}$ [8,14,30]. The overall training loss for the class-agnostic query feature is defined as $\\ell_{agn} = \\ell_{ce} + \\ell_{mask}$ . During testing, $\\mathbf{p}$ and $\\mathbf{m}$ are merged via the multiplication $\\arg \\max_{i:c_i\\neq \\emptyset}\\mathbf{p}(c_i)\\cdot \\mathbf{m}_i$ , where $c_{i}$ is the class label with maximum likelihood $c_{i} = \\arg \\max_{c\\in \\{1,\\dots,C,\\emptyset \\}}\\mathbf{p}_{i}(c)$ for each probability-mask pair indexed by $i$ .", + "bbox": [ + 212, + 559, + 787, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Class-conditional Prompting Machine (CPM)", + "text_level": 1, + "bbox": [ + 214, + 647, + 635, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proposed CPM training builds upon the class-agnostic training method from Sec. 3.1 by introducing a new training stage based on class-conditional prompting. This allows the probing of both the magnitude spectrogram and the image feature map, aiming to mitigate unstable training issues and improve cross-attention efficacy. In some cases, the prompts can be manually crafted (i.e., text label, bounding box) based on domain knowledge or specific task requirements [22]. Prompts can also be automatically learned to form a fixed set of prototypical embeddings for each class [42]. However, it is impossible to manually define class conditional prompts in high-dimensional latent spaces, and using a limited-size set of learned prompts may fail to capture the comprehensive distribution of class-specific prompts. Hence, it is desirable that these", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "class-specific prompts can be sampled from a generative model [3,27]. that comprehensively represents the respective class. We adopt the Gaussian Mixture Models (GMMs) as such generative model [45], which improves the intra-class variability and increases robustness to class imbalances when compared to the alternative approaches mentioned before. Before delving into the methodology of the CPM, we first introduce the generation process of the class-conditional query features.", + "bbox": [ + 212, + 146, + 782, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Class Conditional Distribution Modelling (CCDM) Instead of constructing a posterior $p(c|\\tilde{\\mathbf{q}})$ with Softmax classifier for the mask embeddings $\\tilde{\\mathbf{q}}$ , the generative classifier employs the Bayes rule for label prediction, estimating the class-conditional distribution $p(\\tilde{\\mathbf{q}}|c)$ alongside the class prior $p(c)$ [27], with:", + "bbox": [ + 212, + 258, + 784, + 321 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np (c | \\tilde {\\mathbf {q}}) = \\frac {p (\\tilde {\\mathbf {q}} | c) p (c)}{\\sum_ {c ^ {\\prime}} p (\\tilde {\\mathbf {q}} | c ^ {\\prime}) p \\left(c ^ {\\prime}\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 329, + 784, + 362 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the class probability prior $p(c)$ is uniform. To enable GMM modelling in Maskformer architecture [8, 9], we replace the $\\mathbf{p} = \\operatorname{Softmax}(\\tilde{\\mathbf{q}})$ activation function with $f_{\\mathrm{GMM}}(\\tilde{\\mathbf{q}}) = p(c|\\tilde{\\mathbf{q}})$ defined in (1) which maps the input mask embeddings $\\tilde{\\mathbf{q}}$ to probability data density over $C$ classes. Assuming that after the Hungarian matching between the predictions set $\\{(\\mathbf{m}_i,\\mathbf{p}_i)\\}_{i=1}^N$ and $\\mathbf{y}$ , we get a new one-to-one matched ground-truth set $\\tilde{\\mathbf{y}}$ w.r.t each prediction by padding the $\\mathbf{y}$ with no-object class $\\varnothing$ . Subsequently, we can form the GMM training dataset $\\mathcal{F} = \\{(\\tilde{\\mathbf{q}}_n,\\tilde{\\mathbf{y}}_n)\\}_{n=1}^F$ by pairing the mask embeddings with the assigned label. In our method, the goal of GMM is to model the data distribution of the joint-modal mask embedding $\\tilde{\\mathbf{q}}$ for each class $C$ in the $D_q$ -dimensional space by employing a weighted mixture of $M$ multivariate Gaussians, defined as follows:", + "bbox": [ + 212, + 369, + 784, + 535 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np (\\tilde {\\mathbf {q}} | c) = \\sum_ {m = 1} ^ {M} p (m | c; \\boldsymbol {\\pi} _ {c}) p (\\tilde {\\mathbf {q}} | c, m; \\boldsymbol {\\mu} _ {c}, \\boldsymbol {\\Sigma} _ {c}) = \\sum_ {m = 1} ^ {M} \\boldsymbol {\\pi} _ {c m} \\mathcal {N} (\\tilde {\\mathbf {q}}; \\boldsymbol {\\mu} _ {c m}, \\boldsymbol {\\Sigma} _ {c m}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 246, + 545, + 784, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\sum_{m}\\pi_{cm} = 1$ represents the mixing coefficients, $\\pmb{\\mu}_{cm}\\in \\mathbb{R}^{D_q}$ denotes the mean vector, and $\\Sigma_{cm}\\in \\mathbb{R}^{D_q\\times D_q}$ is the covariance matrix. The optimisation of the GMM parameters is performed by the Expectation Maximisation (EM) algorithm [10]. In the E-step, we iterate through $\\mathcal{F}$ , computing the responsibilities using the current parameters of the GMM for each class $c$ :", + "bbox": [ + 212, + 593, + 782, + 667 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma_ {c, n} ^ {(t)} (m) = \\frac {\\pi_ {c m} ^ {(t - 1)} \\mathcal {N} \\left(\\tilde {\\mathbf {q}} _ {n} ; \\boldsymbol {\\mu} _ {c m} ^ {(t - 1)} , \\boldsymbol {\\Sigma} _ {c m} ^ {(t - 1)}\\right)}{\\sum_ {m ^ {\\prime} = 1} ^ {M} \\pi_ {c m ^ {\\prime}} ^ {(t - 1)} \\mathcal {N} \\left(\\tilde {\\mathbf {q}} _ {n} ; \\boldsymbol {\\mu} _ {c m ^ {\\prime}} ^ {(t - 1)} , \\boldsymbol {\\Sigma} _ {c m ^ {\\prime}} ^ {(t - 1)}\\right)}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 676, + 784, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We re-estimate the parameters using the calculated responsibility in the M-step.", + "bbox": [ + 215, + 723, + 782, + 739 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} _ {c m} ^ {(t)} = \\frac {1}{N _ {c m} ^ {(t)}} \\sum_ {(\\tilde {\\mathbf {q}} _ {n}, \\tilde {\\mathbf {y}} _ {n}) \\in \\mathcal {F} _ {c}} \\gamma_ {c, n} ^ {(t)} (m) \\tilde {\\mathbf {q}}, \\quad \\boldsymbol {\\pi} _ {c m} ^ {(t)} = \\frac {N _ {c m} ^ {(t)}}{| \\mathcal {F} |}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 303, + 760, + 784, + 808 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} _ {c m} ^ {(t)} = \\frac {1}{N _ {c m} ^ {(t)}} \\sum_ {(\\tilde {\\mathbf {q}} _ {n}, \\tilde {\\mathbf {y}} _ {n}) \\in \\mathcal {F} _ {c}} \\gamma_ {c, n} ^ {(t)} (m) (\\tilde {\\mathbf {q}} _ {n} - \\boldsymbol {\\mu} _ {c m} ^ {(t)}) (\\tilde {\\mathbf {q}} _ {n} - \\boldsymbol {\\mu} _ {c m} ^ {(t)}) ^ {\\top},\n$$\n", + "text_format": "latex", + "bbox": [ + 303, + 806, + 702, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Class-conditional Prompting Machine (CPM)", + "bbox": [ + 428, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $N_{cm}^{(t)} = \\sum_{(\\tilde{\\mathbf{q}}_n,\\tilde{\\mathbf{y}}_n)\\in \\mathcal{F}_c}\\gamma_{c,n}^{(t)}(m)$ . We maintain an external memory bank containing GMM training samples, defined above with the set $\\mathcal{F}$ , and the GMM parameters are robustly estimated with momentum [27].", + "bbox": [ + 212, + 143, + 782, + 193 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Audio Conditional Prompting (ACP) Motivated by the concept of \"mix-and-separate\" [13, 18], we design the ACP process to improve cross-attention interaction for the dense audio feature representations. We consider an off-the-screen noise dataset $\\mathcal{D}_r = \\{(\\mathbf{a}_j)\\}_{j=1}^{|D_r|}$ . For each training iteration, we sample a clean audio sample $\\mathbf{a}_i$ from the $\\mathcal{D}$ and an off-the-screen noisy audio sample $\\mathbf{a}_j$ from $\\mathcal{D}_r$ to build the mixture of magnitude spectrogram $\\mathbf{a}_p = \\mathbf{a}_i + \\mathbf{a}_j$ . The whole ACP process is illustrated in the first section of Fig. 2. Before delving into the audio recovery process, we first sample a set of class-conditional prompts $\\mathbf{z}^k \\sim \\{\\mathcal{N}(\\boldsymbol{\\mu}_k, \\boldsymbol{\\Sigma}_k)\\}_{k \\in \\mathcal{K}}$ (where $\\mathcal{K} = \\{k | \\mathbf{t}_i(k) = 1\\}$ represent the indices for ground truth labels) via the GMM model from the last iteration according to the target semantic classes $\\mathbf{t}_i$ that we want to recover. The major process of the audio recovery process shares a similar concept with the visual part described in Sec. 3.1; the difference is that we use the joint-modal mask embeddings $\\tilde{\\mathbf{z}}^k$ to search the semantically similar sound source on the decoded dense audio feature map $\\tilde{\\mathbf{u}}_a$ and generate their corresponding mask predictions $\\{\\mathbf{m}_k^a | k \\in \\mathcal{K}\\}$ .", + "bbox": [ + 212, + 204, + 784, + 433 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The learning objective of the ACP process is to encourage the consistency between the predicted masks $\\mathbf{m}_k^a$ with the derived ground-truth spectrogram ratio mask $\\frac{\\mathbf{a}_i}{\\mathbf{a}_p}$ . The AVS datasets generally do not provide the separated sound source data for each semantic class (i.e., in Fig. 2, we do not have the ground-truth audio data of Male, Female and Guitar), hence we adopted a summation operation over the predicted masks to integrate all mixed sound sources. Finally, a Sigmoid activation function $\\sigma(\\cdot)$ is applied to constrain the prediction into a binary mask. We denoted the mean squared error (MSE) loss of the ACP process as follows:", + "bbox": [ + 212, + 434, + 785, + 569 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\mathrm {A C P}} = \\left\\| \\sigma \\left(\\sum_ {k \\in \\mathcal {K}} \\mathbf {m} _ {k} ^ {a}\\right) - \\frac {\\mathbf {a} _ {i}}{\\mathbf {a} _ {p}} \\right\\| _ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 570, + 784, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual Conditional Prompting (VCP) The VCP module is processed simultaneously with ACP. The design of VCP aims to bypass bipartite matching via the generated class-conditional prompts $\\mathbf{z}^k$ to ease the model training as it can provide a more stable learning target for each query feature to mitigate the instability brought by the bipartite matching with the class-agnostic queries. Given a set of class-conditional prompts derived from the ground-truth image labels, our training objective of VCP is to correctly segment the corresponding image regions as well as successfully classify these prompts after the consecutive transformer decoder layers. The loss function is similar to $\\ell_{\\text{agn}}$ in Sec. 3.1, where we replace $\\mathbf{q}$ with $\\mathbf{z}^k$ , denoted as $\\ell_{\\text{VCP}} = \\ell_{ce} + \\ell_{mask}$ (6).", + "bbox": [ + 212, + 630, + 787, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompting-based Contrastive Learning (PCL) The ability to learn discriminative feature representation is critical for the audio-visual system. One limitation of the previous audio-visual contrastive learning methods [7,36] is that", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "they can only leverage the global audio representation due to the lack of class conditional prompts for class-specific feature disentanglement. By taking advantage of class-conditional distribution modelling, we can overcome this limitation by utilising the predicted spectrogram saliency mask $\\mathbf{s}_k^a = \\sigma (\\mathbf{m}_k^a)$ and its associated class label $k$ of each sound source, denoted as $\\{\\mathbf{s}_k^a |\\mathbf{s}_k^a\\in \\mathcal{M}\\subseteq [0,1]^T\\times F,k\\in \\mathcal{K}\\}$ . To disentangle the class-specific representations for the audio feature map, we iteratively apply the masked average pooling (MAP) [47, 58] to $\\mathbf{u}_a$ based on $|\\mathcal{K}|$ saliency masks $\\mathbf{s}_k^a$ for the feature extraction process. We first threshold $\\mathbf{s}_k^a$ and convert it to a binary mask via $\\tilde{\\mathbf{s}}_k^a = \\mathbb{1}(\\mathbf{s}_k^a >\\bar{\\mathbf{s}}_k^a)$ , where $\\bar{\\mathbf{s}}_k^a$ represent the mean value of the saliency mask. Then, for the $k$ -th category, we extract its region-level mean feature $\\mathbf{f}(k)\\in \\mathbb{R}^{D_q}$ by MAP defined as follows:", + "bbox": [ + 212, + 146, + 782, + 311 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} (k) = \\frac {\\sum_ {\\psi \\in \\Psi} \\tilde {\\mathbf {s}} _ {k} ^ {a} \\mathbf {u} _ {a} (\\psi)}{\\sum_ {\\psi} \\tilde {\\mathbf {s}} _ {k} ^ {a}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 324, + 785, + 359 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\varPsi$ is the lattice of size $T\\times F$ . We combine the $\\mathbf{f}(k)$ with the ground-truth class label $k$ to form the anchor set $\\mathcal{E}_{\\mathrm{anch}} = \\{(\\mathbf{f}(k),k)|\\mathbf{f}(k)\\in \\mathcal{U},k\\in \\mathcal{K}\\}$ , with $\\mathcal{U}$ denoting the audio and visual feature spaces defined in Sec. 3.1. Since the pixel-level label $\\mathbf{y}$ is available, we can directly form the visual contrastive set $\\mathcal{E}_{\\mathrm{cont}} = \\{(\\mathbf{u}_v(\\omega),\\mathbf{y}_v(\\omega))|\\mathbf{u}_v(\\omega)\\in \\mathcal{U},\\mathbf{y}(\\omega)\\in \\mathcal{V}\\}$ . Hence, we can derive the positive and negative sets as:", + "bbox": [ + 212, + 371, + 784, + 462 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {P} (\\mathbf {u} _ {v} (\\omega)) = \\left\\{\\mathbf {u} _ {v} (\\omega) \\mid \\mathbf {u} _ {v} (\\omega) \\in \\mathcal {U}, \\mathbf {y} (\\omega , k) = 1 \\right\\}, \\\\ \\mathcal {U} (v, v) = \\left\\{v: (v) \\mid v: (v) \\in \\mathcal {U}, (v - 1) = 0 \\right\\}. \\end{array} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 476, + 784, + 498 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {N} (\\mathbf {u} _ {v} (\\omega)) = \\left\\{\\mathbf {u} _ {v} (\\omega) | \\mathbf {u} _ {v} (\\omega) \\in \\mathcal {U}, \\mathbf {y} (\\omega , k) = 0 \\right\\}.\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 493, + 661, + 508 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Adopting the supervised InfoNCE [21] as the objective function to pull the anchor $\\mathbf{f} \\in \\mathcal{E}_{\\mathrm{anch}}$ and respective positive visual features closer while repelling anchors and their negative visual features, we define the following loss:", + "bbox": [ + 212, + 521, + 782, + 566 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\mathrm {P C L}} (\\mathbf {f}) = \\frac {1}{| \\mathcal {P} \\left(\\mathbf {u} _ {v} (\\omega)\\right) |} \\sum_ {\\mathbf {f} _ {p} \\in \\mathcal {P} \\left(\\mathbf {u} _ {v} (\\omega)\\right)} - \\log \\frac {\\exp \\left(\\mathbf {f} \\cdot \\mathbf {f} _ {p} / \\tau\\right)}{\\exp \\left(\\mathbf {f} \\cdot \\mathbf {f} _ {p} / \\tau\\right) + \\sum_ {\\mathbf {f} _ {n} \\in \\mathcal {N} \\left(\\mathbf {u} _ {v} (\\omega)\\right)} \\exp \\left(\\mathbf {f} \\cdot \\mathbf {f} _ {n} / \\tau\\right)}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 578, + 784, + 611 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbf{f}$ is an anchor feature, and $\\tau$ is the temperature hyperparameter. The combination of all sub-objective $\\ell_{\\mathrm{ACP}}$ , $\\ell_{\\mathrm{VCP}}$ and $\\ell_{\\mathrm{PCL}}$ form the CPM loss $\\ell_{\\mathrm{CPM}} = \\ell_{\\mathrm{ACP}} + \\ell_{\\mathrm{VCP}} + \\ell_{\\mathrm{PCL}}$ .", + "bbox": [ + 212, + 623, + 782, + 667 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "1. Overall training The overall training objective is $\\ell = \\ell_{\\text{agn}} + \\lambda \\ell_{\\text{CPM}}$ , where $\\lambda$ is the weight coefficient.", + "bbox": [ + 212, + 685, + 782, + 715 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 727, + 374, + 744 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 753, + 455, + 768 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Protocols We utilize standard evaluation protocols from AVSBench datasets [56, 57] for single-source (SS) and multi-source (MS) scenarios with binary labels, as well as for AVSBench-Semantics with multi-class labels. Image sizes are standardized to $224 \\times 224$ for fair comparison. Additionally, we", + "bbox": [ + 212, + 779, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Class-conditional Prompting Machine (CPM)", + "bbox": [ + 428, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/77b2707f40699657f856fb1e078d3b5f65358d400e0053ca9681b41592eca0a3.jpg", + "table_caption": [ + "Table 1: Quantitative (mIoU, $F_{\\beta}$ ) audio-visual segmentation results (\\%) for the AVS-Bench test sets [56, 57] (resized to $224 \\times 224$ ) with ResNet50 [17] backbone. Best results in bold, $2^{nd}$ best underlined. Improvements against the $2^{nd}$ best are in the last row." + ], + "table_footnote": [], + "table_body": "
Eval. MethodsMethodAVSBench-Obj.(SS)AVSBench-Obj.(MS)AVSBench-S
mIoU ↑Fβ ↑mIoU ↑Fβ ↑mIoU ↑Fβ ↑
Per-image [56,57]TPAVI [57]72.7984.8047.8857.8020.1825.20
AVSBG [16]74.1385.4044.9556.80--
ECMVAE [37]76.3386.5048.6960.70--
DiffusionAVS [36]75.8086.9049.7762.10--
CATR [24]74.8086.6052.8065.30--
AuTR [31]75.0085.2049.4061.20--
AQFormer [19]77.0086.4055.7066.90--
AVSegFormer [14]76.5484.8049.5362.8024.9329.30
AVSC [29]77.0285.2449.5861.51--
BAVS [30]77.9685.2950.2362.3724.6829.63
AVSAC [5]76.9086.9553.9565.8125.4329.71
QSD [26]77.6086.0059.6063.5046.60-
COMBO [54]81.7090.1054.5066.6033.3037.30
Per-dataset [12]AVSegFormer* [14]79.8091.4153.2864.7641.6455.05
CAVP [7]85.7792.8662.3973.6244.7057.76
CPM87.6493.5365.2275.2248.3964.01
", + "bbox": [ + 259, + 186, + 736, + 383 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "adopt the original image resolution for testing, following [7], to demonstrate optimal model performance. Evaluation is extended to the Visual Post-production (VPO) benchmark [7] for challenging cases. We employ mean Intersection over Union (mIoU) [11] and $F_{\\beta}$ score with $\\beta^2 = 0.3$ [38, 57] to assess segmentation quality, precision, and recall performance at pixel level. Models are stratified into CNN-based pre-pixel classification and transformer-based mask classification to demonstrate architectural capabilities. Official training splits are used for AVSBench datasets [56, 57] and VPO [7], with results reported on the respective testing sets. Training is performed on the entire AVSBench-Semantics dataset, and testing is conducted on subsets as well as the entire testing set to demonstrate partitioned model performance. For further details on training and inference, please refer to the Supplementary Material.", + "bbox": [ + 212, + 396, + 787, + 578 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results We collected the experimental results from the existing benchmark [7] and updated it with recent works [5,26,54]. We modified the AVSegFormer [14] with Mask2former [8] and denoted it as AVSegFormer* in the tables to encourage a fair comparison with our method. Please note that Tab. 1 includes two evaluation protocols. We employ standard semantic segmentation protocols to compute both mIoU and $F_{\\beta}$ same as PascalVOC, initially mentioned in [57].", + "bbox": [ + 212, + 584, + 787, + 676 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Performance on Low-resolution AVSBench Videos", + "text_level": 1, + "bbox": [ + 215, + 681, + 674, + 696 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We adopt established methodologies [7, 30, 31, 37] for conducting performance evaluations on down-sampled image benchmarks, such as AVSBench-Objects [57], which includes single-source (SS) and multi-source (MS) splits with binary annotations, as well as the AVSBench-Semantics dataset [56], which contains multiclass annotations. We compare the performance of state-of-the-art (SOTA) methods with our CPM in Tab. 1 using mIoU and $F_{\\beta}$ . The results demonstrate that our model surpasses the second-best CAVP [7] in terms of mIoU by $1.87\\%$ on AVSBench-Object (SS) [57], $2.83\\%$ on AVSBench-Object (MS) [57] and $1.79\\%$ on AVSBench-Semantics [57] using the ResNet-50 [17] backbone.", + "bbox": [ + 212, + 703, + 787, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/d7c01bd7e5822334db778a916bb4b0eb49ba56821a42f3e046b7c5745bfc97da.jpg", + "table_caption": [ + "Table 2: Quantitative (mIoU, $F_{\\beta}$ ) audio-visual segmentation results (in %) for the AVSBench-Semantic (AVSS) test sets [56] (original resolution) with ResNet50 [17] backbone. Best results in bold, $2^{nd}$ best underlined. Improvements against the $2^{nd}$ best are in the last row." + ], + "table_footnote": [], + "table_body": "
D-ResNet50 [17]MethodAVSS (SS)AVSS (MS)AVSS
mIoU ↑Fβ ↑mIoU ↑Fβ ↑mIoU ↑Fβ ↑
Per-pixel ClassificationTPAVI [57]42.1061.4626.3340.9943.3959.24
CAVP [7]56.9169.1538.6152.9250.7564.57
TransformerAVSegFormer [14]46.2559.7627.2141.3841.4856.21
AVSegFormer* [14]50.5263.7531.4042.8145.8059.16
CPM61.7172.9443.1156.2857.2570.54
ImprovementCPM+4.80+3.79+4.50+3.36+6.50+5.97
", + "bbox": [ + 274, + 200, + 722, + 296 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/5f27222fcf904db2238c12d81bccaffec47182983902d76aaab931e669b2069a.jpg", + "table_caption": [ + "Table 3: Quantitative (mIoU, $F_{\\beta}$ ) audio-visual segmentation results (in %) for the VPO test sets (original resolution) with ResNet50 [17] backbone. Best results in bold, $2^{nd}$ best underlined. Improvements against the $2^{nd}$ best are in the last row." + ], + "table_footnote": [], + "table_body": "
D-ResNet50 [17]MethodVPO (SS)VPO (MS)VPO (MSMI)
mIoU ↑Fβ ↑mIoU ↑Fβ ↑mIoU ↑Fβ ↑
Per-pixel ClassificationTPAVI [57]52.7569.5454.3071.9551.7368.85
CAVP62.3178.4664.3178.9260.3675.60
TransformerAVSegFormer [14]57.5573.0358.3374.2854.2270.39
AVSegFormer* [14]60.5174.8162.9177.3356.2472.67
CPM67.0979.8865.9179.9060.5575.58
ImprovementsCPM+4.78+1.42+1.60+0.98+0.19-0.02
", + "bbox": [ + 274, + 339, + 722, + 436 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Performance on Original AVSBench Videos", + "text_level": 1, + "bbox": [ + 212, + 449, + 620, + 465 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The benchmark mentioned above adopted low-resolution benchmarks using significantly resized input images (from 720p to $224 \\times 224$ ). While this simplifies model training, disregarding the original image aspect ratio can lead to a degradation in model performance, which is not advisable for segmentation tasks. Therefore, we follow the evaluation method outlined in [7] for training and testing on the raw AVSBench videos. This includes using random resized crops for training and performing frame-by-frame evaluation during testing. The initial published version of AVSBench [57] only offers single-source and multisource partitioning. However, this partitioning was not conducted in the later version [56], resulting in some cases being overlooked. We re-organised the previous SS and MS with the newly added video data, resulting in 1278 single-source videos (AVSS-SS) and 276 multi-source videos (AVSS-MS). We compare the results with mIoU and $F_{\\beta}$ to demonstrate the comprehensive model performance in Tab. 2. Our method shows a significant mIoU improvement of $4.80\\%$ on AVSS-SS, $4.50\\%$ on AVSS-MS and $+6.50\\%$ on the entire AVSS dataset. To further demonstrate the effectiveness of our method, we show a visualisation of 6-second video clip in Fig. 5 that displays a qualitative comparison between TPAVI, AVSegFormer, CAVP and our CPM. Our method can successfully approximate the ground truth segmentation of the target sound source within a group of other semantic objects.", + "bbox": [ + 212, + 465, + 787, + 768 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Performance on VPO Dataset", + "text_level": 1, + "bbox": [ + 212, + 787, + 509, + 801 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We also compare the model performance on the VPO benchmark [7], equipped with synthesized stereo audios. We adopted ResNet-50 [17] backbone for all", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Class-conditional Prompting Machine (CPM)", + "bbox": [ + 428, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/f1aecbcb4a60bc22c9f039aeafe13eaa23b9ac8c1bd9d2ff4b23258e507271a8.jpg", + "table_caption": [ + "Table 4: Ablation study of the model components on AVSBench-Semantics [31]." + ], + "table_footnote": [], + "table_body": "
MethodMetrics
BaselineCCDMVCPACPPCLmIoU ↑Fβ ↑
✓53.0465.29
✓✓54.1266.65
✓✓✓55.7968.17
✓✓✓55.0768.03
✓✓✓✓56.4169.07
✓✓✓✓✓57.2570.54
", + "bbox": [ + 243, + 159, + 746, + 255 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/26a18b65d97af0e551be591b6aa13fb28b79e7304bbee5c7e72fc29a7292de34.jpg", + "image_caption": [ + "(a) Visual comparison." + ], + "image_footnote": [], + "bbox": [ + 256, + 256, + 496, + 330 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6846e96eea8630f3724ef00e3f1716a43504c81c0cd83464be8a5e18e380199e.jpg", + "image_caption": [ + "(b) Coverage score comparison.", + "Fig. 3: Qualitative (3a) and quantitative (3b) comparisons between model components in a multi-source scenario (i.e., male singing, female singing and guitar)" + ], + "image_footnote": [], + "bbox": [ + 500, + 258, + 740, + 330 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "three subsets, and the results are shown in Tab. 3. To facilitate stereo audio encoding, we adopt the approach outlined in [7] by adjusting the number of input channels in the first layer to 2. Our approach surpasses the SOTA method CAVP [7] by $4.78\\%$ and $1.60\\%$ in terms of mIoU on VPO (SS) and VPO (MS), respectively. However, we observe a marginal improvement of $0.19\\%$ on the VPO (MSMI) setup compared with the SS and MS subsets. The possible explanation for this phenomenon may stem from the overly simplistic integration of stereo audio within the transformer architecture, as the foundational AVS transformer architectures [14, 24, 26, 30] did not account for a distinct positional encoding scheme tailored for stereo data. We will explore the design of a dedicated stereo AVS transformer architecture as part of our future work.", + "bbox": [ + 212, + 388, + 787, + 554 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 566, + 388, + 580 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation of Key Components We first perform the key components analysis of CPM on AVSBench-Semantics [56] in Tab. 4. The baseline is AVSegformer* [14] in the 1st row. We replace the Softmax classifier in baseline with CCDM defined in (1), and observe an mIoU improvement of $+1.08\\%$ . By integrating the model with VCP (3rd row), using training loss in (6), and ACP (4th row), using loss in (5), separately, we achieve an improvement of $+1.26\\%$ and $+1.39\\%$ , respectively. Subsequently, when we apply both ACP and VCP methods (5th row) the mIoU performance improves $+2.04\\%$ compared to the 2nd row. The final row displays the complete CPM method, incorporating the dense contrastive learning approach we introduced, which leads to an additional mIoU improvement of $+1.09\\%$ .", + "bbox": [ + 212, + 585, + 787, + 752 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation of Bipartite Matching Stability To study the stability of the bipartite matching process, we design an entropy-based stability score STS to quantify such performance on AVSBench-Semantics [56]. Intuitively, the STS measures the average assignment consistency after bipartite matching across all the classes. During the training process, we collect the assigned label $\\tilde{\\mathbf{y}}_i \\in$", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c8f2d72823cf817151f3c48a18aca44174943092ad494ff965295b1b42bcdf80.jpg", + "image_caption": [ + "Fig.4: Matching stability (STS $\\downarrow$ ) comparison on AVSBench-Semantics [31]" + ], + "image_footnote": [], + "bbox": [ + 230, + 181, + 478, + 285 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/a13f8acf23366f7b33e25c0d55285bb020f275b621e931d26ec058ee45c1b788.jpg", + "table_caption": [ + "Table 5: Ablation of GMM modelling on AVSBench-Semantics [31]." + ], + "table_footnote": [], + "table_body": "
MethodmIoU ↑Fβ ↓
Point Rep.55.9668.58
Dist Rep.57.2570.54
MmIoUFβ
155.8368.68
357.2570.54
556.3169.04
756.0568.59
", + "bbox": [ + 521, + 202, + 743, + 304 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "$\\{0, \\dots, C\\}^N$ to the class-agnostic query $\\tilde{\\mathbf{q}}_n$ produced by the Hungarian algorithm for each training image, resulting in $\\mathbf{R} = [\\tilde{\\mathbf{y}}_0 \\oplus \\tilde{\\mathbf{y}}_1 \\oplus, \\dots, \\oplus \\tilde{\\mathbf{y}}_D]$ , where $\\oplus$ represents the concatenation operator and $\\mathbf{R}$ has size $D \\times N$ . We then use the indicator function to extract the assignment information over all semantic classes. We denote the resulting set as $\\{\\mathbf{s}_0, \\mathbf{s}_1, \\dots, \\mathbf{s}_c\\}$ , where $\\mathbf{s}_c(n) = \\sum_{d=1}^{D} \\mathbb{1}(\\mathbf{R}(d,n) = c)$ for $n \\in \\{1, \\dots, N\\}$ , which forms $\\mathbf{s}_c \\in [0,1]^N$ after normalising it to be a probability distribution. Finally, we calculate the stability score $\\mathsf{STS} = \\frac{1}{C} \\sum_{c=1}^{C} H(\\mathbf{s}_c)$ , where $H(\\cdot)$ computes the entropy. We compare our CPM with AVSegformer* over 100 training epochs as illustrated in Fig. 4. Our findings reveal that our CPM consistently improves the stability score (showing lower average entropy) throughout the training process in comparison to AVSegformer*, thereby validating the efficacy of our approach in enhancing bipartite matching stability. Please refer to the Supplementary Material for the Pseudo-code of STS.", + "bbox": [ + 212, + 325, + 787, + 525 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation of Audio Contribution While experimenting with the baseline method mentioned above, we empirically observed that the classification after the audio transformer decoder (i.e., $f^{\\mathrm{TD - A}}$ in Fig. 2) has low accuracy, with noisy predictions. We also found that these noisy predictions are progressively refined by the vision transformer decoder layer based solely on visual clues. Such an observation illustrates that in some hard cases, we may fail to extract useful semantic information from audio, leading to the model being overly reliant on the visual content, which may cause erroneous testing predictions, as shown in Fig. 3a-(i), (ii), (iii). To monitor the amount of valid information retained by the query following interaction with the audio branch, we introduce a classification coverage score. This score is computed by comparing the predictions generated after passing through the audio transformer decoder with the ground-truth class label, using the Jaccard index. We show qualitative results in Fig. 3a for the designated scenario and quantitative results in Fig. 3b for the AVSBench-Semantics testing set. The comparison starts with AVSegFormer as the baseline method with progressive addition of our proposed modules, similar to Tab. 4. Notice that each CPM sub-module contributes to refining key semantic information from the audio modality, as evidenced by the segmentation mask predictions and the improvement in Jaccard indexes. This shows the effectiveness of CPM.", + "bbox": [ + 212, + 551, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Class-conditional Prompting Machine (CPM)", + "bbox": [ + 428, + 113, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/215ea39162a9a98b1b8b6cd2fd660eb079218ae58c13dd8d6b376358c38c42bf.jpg", + "image_caption": [ + "Fig. 5: Qualitative audio-visual segmentation results on AVSBench-Semantics [56] by TPAVI [57], AVSegFormer [14], CAVP [7] and our CPM, which can be compared with the ground truth (GT) Ambulance of the first row." + ], + "image_footnote": [], + "bbox": [ + 272, + 132, + 725, + 297 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ablation of Distribution Representation Table 5 provides a study of the GMM data distribution representation on AVSBench-Semantics [56]. The first row (Point Rep.) replaces the GMM module with a linear SoftMax for the classification output, introducing a learnable $C \\times D_q$ feature map (i.e., $D_q$ -dimensional feature embeddings for $C$ semantic classes for the audio and visual prompting task). The second row (Dist Rep.) shows our approach with the CCDM module, showing that the utilisation of the distribution modelling leads to considerable improvements of $1.29\\%$ and $1.96\\%$ on mIoU and $F_{\\beta}$ respectively. The next rows of Tab. 5 show a hyper-parameter analysis on the number of GMM components (from $M = 1$ to $M = 7$ components). Results reveal that 3 GMM components enable the best fitting for the data distribution, leading to the top performance with a maximum improvement of $1.42\\%$ and $1.95\\%$ on mIoU and $F_{\\beta}$ .", + "bbox": [ + 212, + 357, + 787, + 539 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 215, + 549, + 509, + 565 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduced CPM, a new audio-visual training method designed for the Maskformer based framework to enhance bipartite matching stability and improve the efficacy of cross-modal attention for audio-visual segmentation. We proposed a class-conditional prompting learning strategy that combines class-agnostic queries with class-conditional queries, sampled from our iteratively updated generative model. The generated class-conditional queries are utilised to probe both the magnitude spectrogram and the image feature map, aiming to remove off-the-screen noise and bypass bipartite matching to produce a more stable learning process. Lastly, we extend the class-conditional queries to a new prompting-based audio-visual contrastive learning to explicitly constrain the cross-modal representations. SOTA results on AVS benchmarks suggest that CPM can be a valuable resource for future AVS research.", + "bbox": [ + 212, + 568, + 787, + 747 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations and future work. We acknowledge that the current adaptation of stereo audio into the transformer-based method has limitations, as it encodes positional and semantic information jointly within the transformer block. This contrasts with the typical approach of separately encoding these two types of information in the transformer-based framework. Our future work will concentrate on optimizing this framework through the integration of spatial reasoning.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)", + "2. Chen, H., Xie, W., Afouras, T., Nagrani, A., Vedaldi, A., Zisserman, A.: Localizing visual sounds the hard way. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16867-16876 (2021)", + "3. Chen, J., Lu, J., Zhu, X., Zhang, L.: Generative semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7111-7120 (2023)", + "4. Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 801-818 (2018)", + "5. Chen, T., Tan, Z., Gong, T., Chu, Q., Wu, Y., Liu, B., Lu, L., Ye, J., Yu, N.: Bootstrapping audio-visual segmentation by strengthening audio cues. arXiv preprint arXiv:2402.02327 (2024)", + "6. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)", + "7. Chen, Y., Liu, Y., Wang, H., Liu, F., Wang, C., Carneiro, G.: A closer look at audio-visual semantic segmentation. arXiv e-prints pp. arXiv-2304 (2023)", + "8. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290–1299 (2022)", + "9. Cheng, B., Schwing, A., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems 34, 17864-17875 (2021)", + "0. Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society: series B (methodological) 39(1), 1-22 (1977)", + "1. Everingham, M., Eslami, S.A., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes challenge: A retrospective. International journal of computer vision 111, 98-136 (2015)", + "2. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. International journal of computer vision 88, 303-338 (2010)", + "3. Gao, R., Grauman, K.: Co-separating sounds of visual objects. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3879-3888 (2019)", + "4. Gao, S., Chen, Z., Chen, G., Wang, W., Lu, T.: Avsegformer: Audio-visual segmentation with transformer. arXiv preprint arXiv:2307.01146 (2023)", + "5. Gray, R.: Vector quantization. IEEE Assp Magazine 1(2), 4-29 (1984)", + "6. Hao, D., Mao, Y., He, B., Han, X., Dai, Y., Zhong, Y.: Improving audio-visual segmentation with bidirectional generation. arXiv preprint arXiv:2308.08288 (2023)", + "7. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + ], + "bbox": [ + 225, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Class-conditional Prompting Machine (CPM)", + "bbox": [ + 428, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Hu, X., Chen, Z., Owens, A.: Mix and localize: Localizing sound sources in mixtures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10483-10492 (2022)", + "19. Huang, S., Li, H., Wang, Y., Zhu, H., Dai, J., Han, J., Rong, W., Liu, S.: Discovering sounding objects by audio queries for audio visual segmentation. arXiv preprint arXiv:2309.09501 (2023)", + "20. Jaegle, A., Borgeaud, S., Alayrac, J.B., Doersch, C., Ionescu, C., Ding, D., Koppula, S., Zoran, D., Brock, A., Shelhamer, E., et al.: Perceiver io: A general architecture for structured inputs & outputs. arXiv preprint arXiv:2107.14795 (2021)", + "21. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in neural information processing systems 33, 18661-18673 (2020)", + "22. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)", + "23. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13619-13627 (2022)", + "24. Li, K., Yang, Z., Chen, L., Yang, Y., Xun, J.: Catr: Combinatorial-dependence audio-queried transformer for audio-visual video segmentation. arXiv preprint arXiv:2309.09709 (2023)", + "25. Li, T., Cao, P., Yuan, Y., Fan, L., Yang, Y., Feris, R.S., Indyk, P., Katabi, D.: Targeted supervised contrastive learning for long-tailed recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6918-6928 (2022)", + "26. Li, X., Wang, J., Xu, X., Peng, X., Singh, R., Lu, Y., Raj, B.: Towards robust audiovisual segmentation in complex environments with quantization-based semantic decomposition. arXiv preprint arXiv:2310.00132 (2023)", + "27. Liang, C., Wang, W., Miao, J., Yang, Y.: Gmmseg: Gaussian mixture based generative semantic segmentation models. Advances in Neural Information Processing Systems 35, 31360-31375 (2022)", + "28. Liu, C., Ding, H., Jiang, X.: Gres: Generalized referring expression segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23592-23601 (2023)", + "29. Liu, C., Li, P., Qi, X., Zhang, H., Li, L., Wang, D., Yu, X.: Audio-visual segmentation by exploring cross-modal mutual semantics (2023)", + "30. Liu, C., Li, P., Zhang, H., Li, L., Huang, Z., Wang, D., Yu, X.: Bavs: Bootstrapping audio-visual segmentation by integrating foundation knowledge. arXiv preprint arXiv:2308.10175 (2023)", + "31. Liu, J., Ju, C., Ma, C., Wang, Y., Wang, Y., Zhang, Y.: Audio-aware query-enhanced transformer for audio-visual segmentation. arXiv preprint arXiv:2307.13236 (2023)", + "32. Liu, J., Wang, Y., Ju, C., Zhang, Y., Xie, W.: Annotation-free audio-visual segmentation. arXiv preprint arXiv:2305.11019 (2023)", + "33. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329 (2022)", + "34. Liu, X., Carrington, P., Chen, X., Pavel, A.: What makes videos accessible to blind and visually impaired people? In: Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. pp. 1-14 (2021)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015)", + "36. Mao, Y., Zhang, J., Xiang, M., Lv, Y., Zhong, Y., Dai, Y.: Contrastive conditional latent diffusion for audio-visual segmentation. arXiv preprint arXiv:2307.16579 (2023)", + "37. Mao, Y., Zhang, J., Xiang, M., Zhong, Y., Dai, Y.: Multimodal variational autoencoder based audio-visual segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 954–965 (2023)", + "38. Martin, D.R., Fowlkes, C.C., Malik, J.: Learning to detect natural image boundaries using local brightness, color, and texture cues. IEEE transactions on pattern analysis and machine intelligence 26(5), 530-549 (2004)", + "39. Meng, D., Chen, X., Fan, Z., Zeng, G., Li, H., Yuan, Y., Sun, L., Wang, J.: Conditional detr for fast training convergence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3651-3660 (2021)", + "40. Mo, S., Morgado, P.: A closer look at weakly-supervised audio-visual source localization. arXiv preprint arXiv:2209.09634 (2022)", + "41. Mo, S., Morgado, P.: Localizing visual sounds the easy way. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII. pp. 218-234. Springer (2022)", + "42. Mo, S., Tian, Y.: Audio-visual grouping network for sound localization from mixtures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10565–10574 (2023)", + "43. Mo, S., Tian, Y.: Av-sam: Segment anything model meets audio-visual localization and segmentation. arXiv preprint arXiv:2305.01836 (2023)", + "44. Murray, M.M., Wallace, M.T.: The neural bases of multisensory processes. Frontiers in Neuroscience (2011)", + "45. Reynolds, D.A., et al.: Gaussian mixture models. Encyclopedia of biometrics 741(659-663) (2009)", + "46. Senocak, A., Ryu, H., Kim, J., Oh, T.H., Pfister, H., Chung, J.S.: Sound source localization is all about cross-modal alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7777-7787 (2023)", + "47. Siam, M., Oreshkin, B.N., Jagersand, M.: Amp: Adaptive masked proxies for few-shot segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5249-5258 (2019)", + "48. Sun, Z., Cao, S., Yang, Y., Kitani, K.M.: Rethinking transformer-based set prediction for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3611-3620 (2021)", + "49. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "50. Wang, W., Zhou, T., Yu, F., Dai, J., Konukoglu, E., Van Gool, L.: Exploring cross-image pixel contrast for semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7303-7313 (2021)", + "51. Wang, Y., Liu, W., Li, G., Ding, J., Hu, D., Li, X.: Prompting segmentation with sound is generalizable audio-visual source localizer. arXiv preprint arXiv:2309.07929 (2023)", + "52. Wu, J., Jiang, Y., Sun, P., Yuan, Z., Luo, P.: Language as queries for referring video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4974-4984 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Class-conditional Prompting Machine (CPM)", + "bbox": [ + 428, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "53. Xu, P., Zhu, X., Clifton, D.A.: Multimodal learning with transformers: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)", + "54. Yang, Q., Nie, X., Li, T., Gao, P., Guo, Y., Zhen, C., Yan, P., Xiang, S.: Cooperation does matter: Exploring multi-order bilateral relations for audio-visual segmentation. arXiv preprint arXiv:2312.06462 (2023)", + "55. Zhang, H., Li, F., Xu, H., Huang, S., Liu, S., Ni, L.M., Zhang, L.: Mp-former: Mask-piloted transformer for image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18074-18083 (2023)", + "56. Zhou, J., Shen, X., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., et al.: Audio-visual segmentation with semantics. arXiv preprint arXiv:2301.13190 (2023)", + "57. Zhou, J., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., Zhong, Y.: Audio-visual segmentation. In: European Conference on Computer Vision. pp. 386-403. Springer (2022)", + "58. Zhou, T., Zhang, M., Zhao, F., Li, J.: Regional semantic contrast and aggregation for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4299-4309 (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 383 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Chen et al.", + "bbox": [ + 271, + 114, + 367, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/CPM_ Class-conditional Prompting Machine for Audio-visual Segmentation/462b4fbe-3dda-4385-88f4-05811a4623a9_model.json b/2024/CPM_ Class-conditional Prompting Machine for Audio-visual Segmentation/462b4fbe-3dda-4385-88f4-05811a4623a9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..40d1e0ceebc01a655a8f72024a40049dde4bda92 --- /dev/null +++ b/2024/CPM_ Class-conditional Prompting Machine for Audio-visual Segmentation/462b4fbe-3dda-4385-88f4-05811a4623a9_model.json @@ -0,0 +1,2260 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.229, + 0.141, + 0.776, + 0.187 + ], + "angle": 0, + "content": "CPM: Class-conditional Prompting Machine for Audio-visual Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.212, + 0.768, + 0.243 + ], + "angle": 0, + "content": "Yuanhong Chen1, Chong Wang1, Yuyuan Liu1, Hu Wang2, and Gustavo Carneiro3" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.255, + 0.772, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\) Australian Institute for Machine Learning, University of Adelaide, Australia" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.269, + 0.772, + 0.282 + ], + "angle": 0, + "content": "\\(^{2}\\) Mohamed bin Zayed University of Artificial Intelligence, United Arab Emirates" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.282, + 0.74, + 0.311 + ], + "angle": 0, + "content": "3 Centre for Vision, Speech and Signal Processing, University of Surrey yuanhong.chen@adelaide.edu.au" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.348, + 0.741, + 0.584 + ], + "angle": 0, + "content": "Abstract. Audio-visual segmentation (AVS) is an emerging task that aims to accurately segment sounding objects based on audio-visual cues. The success of AVS learning systems depends on the effectiveness of cross-modal interaction. Such a requirement can be naturally fulfilled by leveraging transformer-based segmentation architecture due to its inherent ability to capture long-range dependencies and flexibility in handling different modalities. However, the inherent training issues of transformer-based methods, such as the low efficacy of cross-attention and unstable bipartite matching, can be amplified in AVS, particularly when the learned audio query does not provide a clear semantic clue. In this paper, we address these two issues with the new Class-conditional Prompting Machine (CPM). CPM improves the bipartite matching with a learning strategy combining class-agnostic queries with class-conditional queries. The efficacy of cross-modal attention is upgraded with new learning objectives for the audio, visual and joint modalities. We conduct experiments on AVS benchmarks, demonstrating that our method achieves state-of-the-art (SOTA) segmentation accuracy4." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.597, + 0.741, + 0.625 + ], + "angle": 0, + "content": "Keywords: Audio-visual Learning \\(\\cdot\\) Segmentation \\(\\cdot\\) Multi-modal Learning" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.651, + 0.377, + 0.667 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.682, + 0.788, + 0.803 + ], + "angle": 0, + "content": "The recognition and integration of auditory and visual data are fundamental to human cognitive processes, playing a critical role in facilitating meaningful communication [44]. Audio-visual segmentation (AVS) is an emerging cross-modal reasoning task that mimics such cognitive processes, aiming to localize visual objects based on audio-visual cues. AVS has many applications, such as the automatic localisation and identification of sounding objects to improve the accessibility of videos for the blind and visually impaired [34]. A major challenge in AVS is achieving effective cross-modal interaction between sound and visual" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "This project is supported by the Australian Research Council (ARC) through grant FT190100525." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.127 + ], + "angle": 0, + "content": "Y. Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.146, + 0.737, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.402 + ], + "angle": 0, + "content": "Fig. 1: Comparing conventional AVS methods [7, 14] with our CPM approach, CPM inherits the class-agnostic query from transformer-based methods and integrates class-conditional prompts sampled from the learned joint-modal data distribution to achieve three objectives: 1) learn disentangled audio partitioning, 2) facilitate semantic-guided object identification, and 3) promote more explicit audio-visual contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.788, + 0.553 + ], + "angle": 0, + "content": "objects [46]. Many AVS methods [7, 37, 57] generally adopt a traditional per-pixel classification framework [8, 52], utilising early fusion strategies (e.g., cross-attention fusion [49]) together with an FCN decoder [4, 35] to make predictions. Such per-pixel design has achieved good performance, but it tends to underutilise the audio data due to its lower informativeness compared with the visual data [7, 46]. Contrastive learning can mitigate this issue by matching informative audio-visual pairs [7]. However, another critical limitation of the per-pixel design is its failure to capture instance-level visual information, resulting in inconsistent segmentation predictions within or between frames in a video sequence [52]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.765 + ], + "angle": 0, + "content": "These two issues have been addressed by transformer-based AVS methods designed to capture instance-level information and to rely on more effective contrastive learning [14, 24, 28, 30, 31, 52, 54]. Nevertheless, these AVS approaches still show slow convergence and relatively poor accuracy [23, 33, 55], primarily attributed to the low efficacy of the cross-attention [48] and the unstable bipartite matching [23]. Solutions for these problems are based on integrated masked-attention [8], which focuses on features around predicted segments, and anchor de-noising [23] to reconstruct the noisy ground-truth bounding box detection. However, these solutions may not work well in AVS because of the weak constraint provided by global audio features that contain a mixture of sound sources [14, 26, 30] resulting in more instability during training. It is worth noting that all methods above have the common issue of relying on class-agnostic prompts that provide little guidance to the bipartite matching process, thereby reducing training efficacy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In this paper, we introduce the Class-conditional Prompting Machine (CPM), an audio-visual segmentation training approach that leverages class-conditional prompts to enhance bipartite matching stability and improve cross-modal attention efficacy. To enhance bipartite matching, we introduce a novel learning method, combining class-agnostic queries [8,14,26] with class-conditional queries," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.429, + 0.115, + 0.731, + 0.13 + ], + "angle": 0, + "content": "Class-conditional Prompting Machine (CPM)" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.358 + ], + "angle": 0, + "content": "sampled from our iteratively updated generative model of class-specific embeddings, with the former queries matched to ground-truth labels using the Hungarian Algorithm [1], followed by individual processing of class-conditional queries in two modalities. To improve cross-modal attention efficacy, new learning objectives are proposed for both audio and visual modalities. In the audio modality (audio conditional prompting, denoted as ACP), the original spectrogram is corrupted with auditory off-the-screen noise and class-conditional queries are employed to reconstruct the original spectrogram, while in the visual modality (visual conditional prompting, denoted as VCP), noisy class-conditional queries sampled from our generative model are used to probe semantically similar content in the image space. To further upgrade the performance of cross-modal attention, we introduce a new prompting-based audio-visual contrastive learning (PCL) task, guided by class-specific queries to densely constrain the cross-modal representations. To summarise, our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.369, + 0.785, + 0.413 + ], + "angle": 0, + "content": "- A new AVS training approach to enhance bipartite matching stability and enhance the efficacy of cross-modal attention. Our core innovation lies in the development of a Class-conditional Prompting Machine (CPM)." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.415, + 0.785, + 0.459 + ], + "angle": 0, + "content": "- To improve the bipartite matching, we propose a new AVS learning strategy that combines class-agnostic queries with class-conditional queries, sampled from our iteratively updated generative model of class-specific embeddings." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.461, + 0.785, + 0.58 + ], + "angle": 0, + "content": "- The efficacy of cross-modal attention is upgraded with new learning objectives for the audio, visual and joint modalities. For audio, we present ACP that perturbs the original spectrogram and uses class-conditional queries to reconstruct the spectrogram; for visual, we introduce VCP that explores noisy class-conditional queries sampled from our generative model to probe the corresponding semantic in visual feature map. To enhance the cross-modal attention efficacy further, we propose PCL, consisting of a new prompting-based joint (audio-visual) contrastive learning task." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.369, + 0.785, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.788, + 0.698 + ], + "angle": 0, + "content": "We firstly show the effectiveness of our CPM model through rigorous evaluation on established benchmarks such as AVSBench-Objects [57] and AVSBench-Semantics [56]. Furthermore, we extend our evaluation by including VPO synthetic benchmarks [7], aiming to enhance our comprehension of AVS methods' capacity to capture audio-visual correlations. Our findings across these benchmarks consistently demonstrate that our approach yields better classification accuracy compared to existing methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.721, + 0.398, + 0.737 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.754, + 0.521, + 0.768 + ], + "angle": 0, + "content": "2.1 Transformer-based architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Transformer-based methods have shown promising performance in detection [1] and segmentation [8,9] benchmarks. The fundamental concept is to leverage the object query to probe image features from the output of transformer encoders and bipartite graph matching to perform set-based box/mask prediction. It is" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "Y. Chen et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.359 + ], + "angle": 0, + "content": "also evidenced that such a framework can benefit multi-modal learning due to its attention mechanism and flexibility in data modelling [53]. Despite its successful application in various domains [1, 8], such frameworks have also been reported to exhibit a poorer convergence rate compared to traditional CNN-based methods [8,23,33]. In exploring the reasons behind the poor convergence, previous studies have emphasized enhancing the interpretability of the learnable query [33, 39] (often referring to them as anchor points/boxes). Alternatively, the utilisation of denoising methods [23, 55] has also been adopted to facilitate the learning of bounding box offset [23] and improve the utilisation of the transformer decoder layers [55]. Furthermore, masked attention has proven effective in enhancing both convergence rate and model performance [8]. The successful implementation of denoising methods and masked attention mechanisms has inspired us to devise a mitigation strategy for bipartite matching that promotes improved cross-modal understanding within the cross-attention process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.37, + 0.542, + 0.387 + ], + "angle": 0, + "content": "2.2 Audio-visual Segmentation (AVS)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.789, + 0.759 + ], + "angle": 0, + "content": "Audio-visual Segmentation is a dense classification task or detection of sounding visual objects in videos, using image sequences and audio cues. Zhou et al. [57] introduced the AVSBench-Object and AVSBench-Semantics benchmarks [56], which enable the evaluations of single-source and multi-source AVS tasks for salient and multi-class object segmentation. To address concerns related to the high annotation cost and dataset diversity, Chen et al. [7] introduced a cost-effective strategy to build a relatively unbiased AVS dataset, named Visual Post-production (VPO). Mainstream AVS methods use audio as the reference query [7, 14, 19, 24, 29–32] or prompt [43, 51]. For example, some methods adopt MaskFormer [9] or segment anything model (SAM) [22] to perform image segmentation using audio queries or encoded audio prompts and cross-attention layers. These methods benefit from the attention mechanism's ability and mask-classification's features to capture long-range dependencies and enhance image segmentation ability [9], spatial-temporal reasoning [24] and task-related features [14, 24, 29, 30, 37, 57]. The adaptation of the Maskformer-based framework for AVS relies on methods to encourage the audio-visual semantic alignment [24, 30, 31, 54]. Such strategy mitigates the poor audio semantic information caused by modality imbalance [5] and learns disentangled multi-modal representations [26, 54] via vector quantization [15] and bidirectional attention mechanism. Even though the query-based transformer architecture [1] has shown great success in semantic segmentation tasks, certain weaknesses such as low efficacy of cross-attention and unstable bipartite matching process [23] are still only partially addressed. These two issues may be exacerbated in AVS due to the poor audio semantic information." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.769, + 0.548, + 0.785 + ], + "angle": 0, + "content": "2.3 Audio-visual Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Contrastive Learning has shown promising results in audio-visual learning (AVL) methods [2, 7, 18, 40, 41]. These methods bring together augmented representations from the same instance as positives while separating representation pairs" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.429, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Class-conditional Prompting Machine (CPM)" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.329 + ], + "angle": 0, + "content": "from different instances as negatives within a batch. A reported issue with current AVL contrastive learning is its reliance on self-supervision [6] to connect audio and visual representations of the same class. To overcome this issue, CAVP [7] propose a supervised contrastive learning [21, 25, 50] method that mines informative contrastive pairs from arbitrary audio-visual pairings to constrain the learning of audio-visual embeddings. Nevertheless, these previous approaches predominantly leverage global audio representations, thereby limiting the model's capacity to discern individual audio sources. Consequently, this constraint reduces the model's effectiveness in scenarios involving multiple sound sources. In our work, we employ class-specific queries to retrieve the corresponding audio representations, which facilitates a more explicit form of contrastive learning between audio and visual modalities." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.351, + 0.331, + 0.367 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.789, + 0.489 + ], + "angle": 0, + "content": "We denote a multi-class audio-visual dataset as \\(\\mathcal{D} = \\{(\\mathbf{a}_i, \\mathbf{x}_i, \\mathbf{y}_i, \\mathbf{t}_i))\\}_{i=1}^{|\\mathcal{D}|}\\), where \\(\\mathbf{x}_i \\in \\mathcal{X} \\subset \\mathbb{R}^{H \\times W \\times 3}\\) is an RGB image with resolution \\(H \\times W\\), \\(\\mathbf{a} \\in \\mathcal{A} \\subset \\mathbb{R}^{T \\times F}\\) denotes the magnitude spectrogram of the audio data with \\(T\\) time and \\(F\\) frequency bins, \\(\\mathbf{y}_i \\in \\mathcal{Y} \\subset \\{0,1\\}^{H \\times W \\times C}\\) denotes the pixel-level ground truth for the \\(C\\) classes (the background class is included in these \\(C\\) classes), and \\(\\mathbf{t}_i \\in \\mathcal{Y} \\subset \\{0,1\\}^C\\) is a multi-label ground truth audio annotation. We use Mask2former [8] as the segmentation framework." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.497, + 0.613, + 0.514 + ], + "angle": 0, + "content": "3.1 Preliminaries using Class-agnostic Queries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.789, + 0.811 + ], + "angle": 0, + "content": "Like other Maskformer-based methods [14,30,31,52], we aim to learn the parameters \\(\\theta \\in \\Theta\\) for the model \\(f_{\\theta}:\\mathcal{X}\\times \\mathcal{A}\\to [0,1]^{H\\times W\\times C}\\), which comprises the image and audio backbones that extract features with \\(\\mathbf{u}_a = f_\\gamma (\\mathbf{a})\\) and \\(\\mathbf{u}_v = f_\\phi (\\mathbf{x})\\), respectively, where \\(\\gamma ,\\phi \\in \\theta\\), and \\(\\mathbf{u}_a,\\mathbf{u}_v\\in \\mathcal{U}\\), with \\(\\mathcal{U}\\) denoting a unified feature space. A set of learnable query features (comprising the object query feature and positional query embeddings) are defined as the joint-modal output embeddings similar to the Perceiver model [20]. We define the class-agnostic query feature as \\(\\mathbf{q}\\in \\mathcal{Q}\\subset \\mathbb{R}^{N\\times D_q}\\), where \\(N\\) denotes the number of class-agnostic queries, and \\(D_{q}\\) represents the dimensionality of the feature space. As depicted in Fig. 2, given \\(\\mathbf{q}\\), we aim to group pixels with matched semantic information from \\(\\mathbf{u}_a\\) and \\(\\mathbf{u}_v\\) through consecutive transformer decoder layers [8,49] and generate \\(N\\) mask embeddings \\(\\tilde{\\mathbf{q}}\\in \\mathcal{Q}\\) and pixel embeddings \\(\\tilde{\\mathbf{u}}_v\\in \\mathbb{R}^{H\\times W\\times D_q}\\). Then, the model independently predicts the embeddings into \\(N\\) set of class predictions (via Softmax(·)) and mask predictions (via Sigmoid(·)) denoted as \\(\\mathcal{E}_{\\mathrm{pred}} = \\left\\{\\left(\\mathbf{m}_i,\\mathbf{p}_i\\right)\\right\\}_{i = 1}^N\\), where \\(\\mathbf{m}\\in \\mathcal{M}\\subset \\{0,1\\}^{H\\times W}\\) and \\(\\mathbf{p}\\in \\mathcal{P}\\subset \\{0,1\\}^C\\). We denote the ground-truth set derived from the training set \\(\\mathcal{D}\\) with \\(\\mathcal{E}_{\\mathrm{gt}} = \\left\\{\\left(\\mathbf{m}_i^{\\mathrm{gt}},\\mathbf{p}_i^{\\mathrm{gt}}\\right)\\right\\}_{i = 1}^{N^{\\mathrm{gt}}}\\). During training, we use the Hungarian algorithm to perform optimal matching between \\(\\mathcal{E}_{\\mathrm{pred}}\\) and \\(\\mathcal{E}_{\\mathrm{gt}}\\) [1,8] to facilitate the label assignment. Since \\(|\\mathcal{E}_{\\mathrm{pred}}| > |\\mathcal{E}_{\\mathrm{gt}}|\\), we pad \\(\\mathcal{E}_{\\mathrm{gt}}\\) with no-object class \\(\\varnothing\\) to achieve one-to-one matching [8]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.789, + 0.842 + ], + "angle": 0, + "content": "The loss \\(\\ell\\) to train the model \\(f_{\\theta}(.)\\) includes a cross-entropy query classification loss \\(\\ell_{ce}\\), and a binary mask loss \\(\\ell_{mask} = \\ell_{focal} + \\ell_{dice}\\), which combines a focal loss" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.127 + ], + "angle": 0, + "content": "Y. Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.146, + 0.74, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.362, + 0.788, + 0.542 + ], + "angle": 0, + "content": "Fig. 2: Illustration of our CPM method. Starting with a scene with a mixture of sound sources including Male, Female and Guitar, the training alternates between the use of learnable class-agnostic queries, and queries sampled from a list of class-specific query features from the GMM, denoted as \\(\\mathbf{z}^{\\mathrm{male}}\\), \\(\\mathbf{z}^{\\mathrm{female}}\\) and \\(\\mathbf{z}^{\\mathrm{guitar}}\\). The overall training objective is composed of three learning tasks: 1) in audio conditional prompting (ACP), we aim to use \\(\\mathbf{z}^{\\mathrm{male}}\\), \\(\\mathbf{z}^{\\mathrm{female}}\\) and \\(\\mathbf{z}^{\\mathrm{guitar}}\\) to recover the original magnitude spectrogram \\(\\mathbf{a}_i\\) from the noise spectrogram (i.e., \\(\\mathbf{a}_i + \\mathbf{a}_j\\)) that is corrupted by another Dog audio signal \\(\\mathbf{a}_j\\); 2) a visual conditional prompting (VCP) that aim to probe the corresponding pixels w.r.t to the class-specific query features; and 3) a contrastive learning task that target to densely constrain the audio and visual representations. For training, both the CPM Workflow, indicated by the orange arrow, and the Class-agnostic Workflow, marked by the black arrow, are utilized. However, only the Class-agnostic Workflow is used for inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.56, + 0.788, + 0.637 + ], + "angle": 0, + "content": "\\(\\ell_{focal}\\) and a dice loss \\(\\ell_{dice}\\) [8,14,30]. The overall training loss for the class-agnostic query feature is defined as \\(\\ell_{agn} = \\ell_{ce} + \\ell_{mask}\\). During testing, \\(\\mathbf{p}\\) and \\(\\mathbf{m}\\) are merged via the multiplication \\(\\arg \\max_{i:c_i\\neq \\emptyset}\\mathbf{p}(c_i)\\cdot \\mathbf{m}_i\\), where \\(c_{i}\\) is the class label with maximum likelihood \\(c_{i} = \\arg \\max_{c\\in \\{1,\\dots,C,\\emptyset \\}}\\mathbf{p}_{i}(c)\\) for each probability-mask pair indexed by \\(i\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.648, + 0.637, + 0.664 + ], + "angle": 0, + "content": "3.2 Class-conditional Prompting Machine (CPM)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "The proposed CPM training builds upon the class-agnostic training method from Sec. 3.1 by introducing a new training stage based on class-conditional prompting. This allows the probing of both the magnitude spectrogram and the image feature map, aiming to mitigate unstable training issues and improve cross-attention efficacy. In some cases, the prompts can be manually crafted (i.e., text label, bounding box) based on domain knowledge or specific task requirements [22]. Prompts can also be automatically learned to form a fixed set of prototypical embeddings for each class [42]. However, it is impossible to manually define class conditional prompts in high-dimensional latent spaces, and using a limited-size set of learned prompts may fail to capture the comprehensive distribution of class-specific prompts. Hence, it is desirable that these" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.429, + 0.115, + 0.731, + 0.13 + ], + "angle": 0, + "content": "Class-conditional Prompting Machine (CPM)" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.253 + ], + "angle": 0, + "content": "class-specific prompts can be sampled from a generative model [3,27]. that comprehensively represents the respective class. We adopt the Gaussian Mixture Models (GMMs) as such generative model [45], which improves the intra-class variability and increases robustness to class imbalances when compared to the alternative approaches mentioned before. Before delving into the methodology of the CPM, we first introduce the generation process of the class-conditional query features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.26, + 0.785, + 0.322 + ], + "angle": 0, + "content": "Class Conditional Distribution Modelling (CCDM) Instead of constructing a posterior \\( p(c|\\tilde{\\mathbf{q}}) \\) with Softmax classifier for the mask embeddings \\( \\tilde{\\mathbf{q}} \\), the generative classifier employs the Bayes rule for label prediction, estimating the class-conditional distribution \\( p(\\tilde{\\mathbf{q}}|c) \\) alongside the class prior \\( p(c) \\) [27], with:" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.33, + 0.785, + 0.363 + ], + "angle": 0, + "content": "\\[\np (c | \\tilde {\\mathbf {q}}) = \\frac {p (\\tilde {\\mathbf {q}} | c) p (c)}{\\sum_ {c ^ {\\prime}} p (\\tilde {\\mathbf {q}} | c ^ {\\prime}) p \\left(c ^ {\\prime}\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.785, + 0.536 + ], + "angle": 0, + "content": "where the class probability prior \\( p(c) \\) is uniform. To enable GMM modelling in Maskformer architecture [8, 9], we replace the \\( \\mathbf{p} = \\operatorname{Softmax}(\\tilde{\\mathbf{q}}) \\) activation function with \\( f_{\\mathrm{GMM}}(\\tilde{\\mathbf{q}}) = p(c|\\tilde{\\mathbf{q}}) \\) defined in (1) which maps the input mask embeddings \\( \\tilde{\\mathbf{q}} \\) to probability data density over \\( C \\) classes. Assuming that after the Hungarian matching between the predictions set \\( \\{(\\mathbf{m}_i,\\mathbf{p}_i)\\}_{i=1}^N \\) and \\( \\mathbf{y} \\), we get a new one-to-one matched ground-truth set \\( \\tilde{\\mathbf{y}} \\) w.r.t each prediction by padding the \\( \\mathbf{y} \\) with no-object class \\( \\varnothing \\). Subsequently, we can form the GMM training dataset \\( \\mathcal{F} = \\{(\\tilde{\\mathbf{q}}_n,\\tilde{\\mathbf{y}}_n)\\}_{n=1}^F \\) by pairing the mask embeddings with the assigned label. In our method, the goal of GMM is to model the data distribution of the joint-modal mask embedding \\( \\tilde{\\mathbf{q}} \\) for each class \\( C \\) in the \\( D_q \\)-dimensional space by employing a weighted mixture of \\( M \\) multivariate Gaussians, defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.248, + 0.546, + 0.785, + 0.587 + ], + "angle": 0, + "content": "\\[\np (\\tilde {\\mathbf {q}} | c) = \\sum_ {m = 1} ^ {M} p (m | c; \\boldsymbol {\\pi} _ {c}) p (\\tilde {\\mathbf {q}} | c, m; \\boldsymbol {\\mu} _ {c}, \\boldsymbol {\\Sigma} _ {c}) = \\sum_ {m = 1} ^ {M} \\boldsymbol {\\pi} _ {c m} \\mathcal {N} (\\tilde {\\mathbf {q}}; \\boldsymbol {\\mu} _ {c m}, \\boldsymbol {\\Sigma} _ {c m}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.784, + 0.669 + ], + "angle": 0, + "content": "where \\(\\sum_{m}\\pi_{cm} = 1\\) represents the mixing coefficients, \\(\\pmb{\\mu}_{cm}\\in \\mathbb{R}^{D_q}\\) denotes the mean vector, and \\(\\Sigma_{cm}\\in \\mathbb{R}^{D_q\\times D_q}\\) is the covariance matrix. The optimisation of the GMM parameters is performed by the Expectation Maximisation (EM) algorithm [10]. In the E-step, we iterate through \\(\\mathcal{F}\\), computing the responsibilities using the current parameters of the GMM for each class \\(c\\):" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.678, + 0.785, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\gamma_ {c, n} ^ {(t)} (m) = \\frac {\\pi_ {c m} ^ {(t - 1)} \\mathcal {N} \\left(\\tilde {\\mathbf {q}} _ {n} ; \\boldsymbol {\\mu} _ {c m} ^ {(t - 1)} , \\boldsymbol {\\Sigma} _ {c m} ^ {(t - 1)}\\right)}{\\sum_ {m ^ {\\prime} = 1} ^ {M} \\pi_ {c m ^ {\\prime}} ^ {(t - 1)} \\mathcal {N} \\left(\\tilde {\\mathbf {q}} _ {n} ; \\boldsymbol {\\mu} _ {c m ^ {\\prime}} ^ {(t - 1)} , \\boldsymbol {\\Sigma} _ {c m ^ {\\prime}} ^ {(t - 1)}\\right)}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.724, + 0.783, + 0.74 + ], + "angle": 0, + "content": "We re-estimate the parameters using the calculated responsibility in the M-step." + }, + { + "type": "equation", + "bbox": [ + 0.305, + 0.761, + 0.785, + 0.809 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} _ {c m} ^ {(t)} = \\frac {1}{N _ {c m} ^ {(t)}} \\sum_ {(\\tilde {\\mathbf {q}} _ {n}, \\tilde {\\mathbf {y}} _ {n}) \\in \\mathcal {F} _ {c}} \\gamma_ {c, n} ^ {(t)} (m) \\tilde {\\mathbf {q}}, \\quad \\boldsymbol {\\pi} _ {c m} ^ {(t)} = \\frac {N _ {c m} ^ {(t)}}{| \\mathcal {F} |}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.304, + 0.807, + 0.703, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} _ {c m} ^ {(t)} = \\frac {1}{N _ {c m} ^ {(t)}} \\sum_ {(\\tilde {\\mathbf {q}} _ {n}, \\tilde {\\mathbf {y}} _ {n}) \\in \\mathcal {F} _ {c}} \\gamma_ {c, n} ^ {(t)} (m) (\\tilde {\\mathbf {q}} _ {n} - \\boldsymbol {\\mu} _ {c m} ^ {(t)}) (\\tilde {\\mathbf {q}} _ {n} - \\boldsymbol {\\mu} _ {c m} ^ {(t)}) ^ {\\top},\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.127 + ], + "angle": 0, + "content": "Y. Chen et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.145, + 0.784, + 0.194 + ], + "angle": 0, + "content": "where \\(N_{cm}^{(t)} = \\sum_{(\\tilde{\\mathbf{q}}_n,\\tilde{\\mathbf{y}}_n)\\in \\mathcal{F}_c}\\gamma_{c,n}^{(t)}(m)\\). We maintain an external memory bank containing GMM training samples, defined above with the set \\(\\mathcal{F}\\), and the GMM parameters are robustly estimated with momentum [27]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.205, + 0.785, + 0.434 + ], + "angle": 0, + "content": "Audio Conditional Prompting (ACP) Motivated by the concept of \"mix-and-separate\" [13, 18], we design the ACP process to improve cross-attention interaction for the dense audio feature representations. We consider an off-the-screen noise dataset \\(\\mathcal{D}_r = \\{(\\mathbf{a}_j)\\}_{j=1}^{|D_r|}\\). For each training iteration, we sample a clean audio sample \\(\\mathbf{a}_i\\) from the \\(\\mathcal{D}\\) and an off-the-screen noisy audio sample \\(\\mathbf{a}_j\\) from \\(\\mathcal{D}_r\\) to build the mixture of magnitude spectrogram \\(\\mathbf{a}_p = \\mathbf{a}_i + \\mathbf{a}_j\\). The whole ACP process is illustrated in the first section of Fig. 2. Before delving into the audio recovery process, we first sample a set of class-conditional prompts \\(\\mathbf{z}^k \\sim \\{\\mathcal{N}(\\boldsymbol{\\mu}_k, \\boldsymbol{\\Sigma}_k)\\}_{k \\in \\mathcal{K}}\\) (where \\(\\mathcal{K} = \\{k | \\mathbf{t}_i(k) = 1\\}\\) represent the indices for ground truth labels) via the GMM model from the last iteration according to the target semantic classes \\(\\mathbf{t}_i\\) that we want to recover. The major process of the audio recovery process shares a similar concept with the visual part described in Sec. 3.1; the difference is that we use the joint-modal mask embeddings \\(\\tilde{\\mathbf{z}}^k\\) to search the semantically similar sound source on the decoded dense audio feature map \\(\\tilde{\\mathbf{u}}_a\\) and generate their corresponding mask predictions \\(\\{\\mathbf{m}_k^a | k \\in \\mathcal{K}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.787, + 0.57 + ], + "angle": 0, + "content": "The learning objective of the ACP process is to encourage the consistency between the predicted masks \\(\\mathbf{m}_k^a\\) with the derived ground-truth spectrogram ratio mask \\(\\frac{\\mathbf{a}_i}{\\mathbf{a}_p}\\). The AVS datasets generally do not provide the separated sound source data for each semantic class (i.e., in Fig. 2, we do not have the ground-truth audio data of Male, Female and Guitar), hence we adopted a summation operation over the predicted masks to integrate all mixed sound sources. Finally, a Sigmoid activation function \\(\\sigma(\\cdot)\\) is applied to constrain the prediction into a binary mask. We denoted the mean squared error (MSE) loss of the ACP process as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.571, + 0.785, + 0.611 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\mathrm {A C P}} = \\left\\| \\sigma \\left(\\sum_ {k \\in \\mathcal {K}} \\mathbf {m} _ {k} ^ {a}\\right) - \\frac {\\mathbf {a} _ {i}}{\\mathbf {a} _ {p}} \\right\\| _ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.631, + 0.788, + 0.784 + ], + "angle": 0, + "content": "Visual Conditional Prompting (VCP) The VCP module is processed simultaneously with ACP. The design of VCP aims to bypass bipartite matching via the generated class-conditional prompts \\(\\mathbf{z}^k\\) to ease the model training as it can provide a more stable learning target for each query feature to mitigate the instability brought by the bipartite matching with the class-agnostic queries. Given a set of class-conditional prompts derived from the ground-truth image labels, our training objective of VCP is to correctly segment the corresponding image regions as well as successfully classify these prompts after the consecutive transformer decoder layers. The loss function is similar to \\(\\ell_{\\text{agn}}\\) in Sec. 3.1, where we replace \\(\\mathbf{q}\\) with \\(\\mathbf{z}^k\\), denoted as \\(\\ell_{\\text{VCP}} = \\ell_{ce} + \\ell_{mask}\\) (6)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Prompting-based Contrastive Learning (PCL) The ability to learn discriminative feature representation is critical for the audio-visual system. One limitation of the previous audio-visual contrastive learning methods [7,36] is that" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.429, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Class-conditional Prompting Machine (CPM)" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.313 + ], + "angle": 0, + "content": "they can only leverage the global audio representation due to the lack of class conditional prompts for class-specific feature disentanglement. By taking advantage of class-conditional distribution modelling, we can overcome this limitation by utilising the predicted spectrogram saliency mask \\(\\mathbf{s}_k^a = \\sigma (\\mathbf{m}_k^a)\\) and its associated class label \\(k\\) of each sound source, denoted as \\(\\{\\mathbf{s}_k^a |\\mathbf{s}_k^a\\in \\mathcal{M}\\subseteq [0,1]^T\\times F,k\\in \\mathcal{K}\\}\\). To disentangle the class-specific representations for the audio feature map, we iteratively apply the masked average pooling (MAP) [47, 58] to \\(\\mathbf{u}_a\\) based on \\(|\\mathcal{K}|\\) saliency masks \\(\\mathbf{s}_k^a\\) for the feature extraction process. We first threshold \\(\\mathbf{s}_k^a\\) and convert it to a binary mask via \\(\\tilde{\\mathbf{s}}_k^a = \\mathbb{1}(\\mathbf{s}_k^a >\\bar{\\mathbf{s}}_k^a)\\), where \\(\\bar{\\mathbf{s}}_k^a\\) represent the mean value of the saliency mask. Then, for the \\(k\\)-th category, we extract its region-level mean feature \\(\\mathbf{f}(k)\\in \\mathbb{R}^{D_q}\\) by MAP defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.325, + 0.786, + 0.36 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} (k) = \\frac {\\sum_ {\\psi \\in \\Psi} \\tilde {\\mathbf {s}} _ {k} ^ {a} \\mathbf {u} _ {a} (\\psi)}{\\sum_ {\\psi} \\tilde {\\mathbf {s}} _ {k} ^ {a}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.785, + 0.463 + ], + "angle": 0, + "content": "where \\(\\varPsi\\) is the lattice of size \\(T\\times F\\). We combine the \\(\\mathbf{f}(k)\\) with the ground-truth class label \\(k\\) to form the anchor set \\(\\mathcal{E}_{\\mathrm{anch}} = \\{(\\mathbf{f}(k),k)|\\mathbf{f}(k)\\in \\mathcal{U},k\\in \\mathcal{K}\\}\\), with \\(\\mathcal{U}\\) denoting the audio and visual feature spaces defined in Sec. 3.1. Since the pixel-level label \\(\\mathbf{y}\\) is available, we can directly form the visual contrastive set \\(\\mathcal{E}_{\\mathrm{cont}} = \\{(\\mathbf{u}_v(\\omega),\\mathbf{y}_v(\\omega))|\\mathbf{u}_v(\\omega)\\in \\mathcal{U},\\mathbf{y}(\\omega)\\in \\mathcal{V}\\}\\). Hence, we can derive the positive and negative sets as:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.477, + 0.785, + 0.499 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {P} (\\mathbf {u} _ {v} (\\omega)) = \\left\\{\\mathbf {u} _ {v} (\\omega) \\mid \\mathbf {u} _ {v} (\\omega) \\in \\mathcal {U}, \\mathbf {y} (\\omega , k) = 1 \\right\\}, \\\\ \\mathcal {U} (v, v) = \\left\\{v: (v) \\mid v: (v) \\in \\mathcal {U}, (v - 1) = 0 \\right\\}. \\end{array} \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.494, + 0.662, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\mathcal {N} (\\mathbf {u} _ {v} (\\omega)) = \\left\\{\\mathbf {u} _ {v} (\\omega) | \\mathbf {u} _ {v} (\\omega) \\in \\mathcal {U}, \\mathbf {y} (\\omega , k) = 0 \\right\\}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.784, + 0.567 + ], + "angle": 0, + "content": "Adopting the supervised InfoNCE [21] as the objective function to pull the anchor \\(\\mathbf{f} \\in \\mathcal{E}_{\\mathrm{anch}}\\) and respective positive visual features closer while repelling anchors and their negative visual features, we define the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.579, + 0.785, + 0.612 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\mathrm {P C L}} (\\mathbf {f}) = \\frac {1}{| \\mathcal {P} \\left(\\mathbf {u} _ {v} (\\omega)\\right) |} \\sum_ {\\mathbf {f} _ {p} \\in \\mathcal {P} \\left(\\mathbf {u} _ {v} (\\omega)\\right)} - \\log \\frac {\\exp \\left(\\mathbf {f} \\cdot \\mathbf {f} _ {p} / \\tau\\right)}{\\exp \\left(\\mathbf {f} \\cdot \\mathbf {f} _ {p} / \\tau\\right) + \\sum_ {\\mathbf {f} _ {n} \\in \\mathcal {N} \\left(\\mathbf {u} _ {v} (\\omega)\\right)} \\exp \\left(\\mathbf {f} \\cdot \\mathbf {f} _ {n} / \\tau\\right)}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.784, + 0.669 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}\\) is an anchor feature, and \\(\\tau\\) is the temperature hyperparameter. The combination of all sub-objective \\(\\ell_{\\mathrm{ACP}}\\), \\(\\ell_{\\mathrm{VCP}}\\) and \\(\\ell_{\\mathrm{PCL}}\\) form the CPM loss \\(\\ell_{\\mathrm{CPM}} = \\ell_{\\mathrm{ACP}} + \\ell_{\\mathrm{VCP}} + \\ell_{\\mathrm{PCL}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.686, + 0.784, + 0.716 + ], + "angle": 0, + "content": "1. Overall training The overall training objective is \\(\\ell = \\ell_{\\text{agn}} + \\lambda \\ell_{\\text{CPM}}\\), where \\(\\lambda\\) is the weight coefficient." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.728, + 0.375, + 0.745 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.754, + 0.457, + 0.769 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Evaluation Protocols We utilize standard evaluation protocols from AVSBench datasets [56, 57] for single-source (SS) and multi-source (MS) scenarios with binary labels, as well as for AVSBench-Semantics with multi-class labels. Image sizes are standardized to \\(224 \\times 224\\) for fair comparison. Additionally, we" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "Y. Chen et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Quantitative (mIoU, \\( F_{\\beta} \\)) audio-visual segmentation results (\\%) for the AVS-Bench test sets [56, 57] (resized to \\( 224 \\times 224 \\)) with ResNet50 [17] backbone. Best results in bold, \\( 2^{nd} \\) best underlined. Improvements against the \\( 2^{nd} \\) best are in the last row." + }, + { + "type": "table", + "bbox": [ + 0.261, + 0.188, + 0.737, + 0.384 + ], + "angle": 0, + "content": "
Eval. MethodsMethodAVSBench-Obj.(SS)AVSBench-Obj.(MS)AVSBench-S
mIoU ↑Fβ ↑mIoU ↑Fβ ↑mIoU ↑Fβ ↑
Per-image [56,57]TPAVI [57]72.7984.8047.8857.8020.1825.20
AVSBG [16]74.1385.4044.9556.80--
ECMVAE [37]76.3386.5048.6960.70--
DiffusionAVS [36]75.8086.9049.7762.10--
CATR [24]74.8086.6052.8065.30--
AuTR [31]75.0085.2049.4061.20--
AQFormer [19]77.0086.4055.7066.90--
AVSegFormer [14]76.5484.8049.5362.8024.9329.30
AVSC [29]77.0285.2449.5861.51--
BAVS [30]77.9685.2950.2362.3724.6829.63
AVSAC [5]76.9086.9553.9565.8125.4329.71
QSD [26]77.6086.0059.6063.5046.60-
COMBO [54]81.7090.1054.5066.6033.3037.30
Per-dataset [12]AVSegFormer* [14]79.8091.4153.2864.7641.6455.05
CAVP [7]85.7792.8662.3973.6244.7057.76
CPM87.6493.5365.2275.2248.3964.01
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.789, + 0.579 + ], + "angle": 0, + "content": "adopt the original image resolution for testing, following [7], to demonstrate optimal model performance. Evaluation is extended to the Visual Post-production (VPO) benchmark [7] for challenging cases. We employ mean Intersection over Union (mIoU) [11] and \\( F_{\\beta} \\) score with \\( \\beta^2 = 0.3 \\) [38, 57] to assess segmentation quality, precision, and recall performance at pixel level. Models are stratified into CNN-based pre-pixel classification and transformer-based mask classification to demonstrate architectural capabilities. Official training splits are used for AVSBench datasets [56, 57] and VPO [7], with results reported on the respective testing sets. Training is performed on the entire AVSBench-Semantics dataset, and testing is conducted on subsets as well as the entire testing set to demonstrate partitioned model performance. For further details on training and inference, please refer to the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.789, + 0.678 + ], + "angle": 0, + "content": "Results We collected the experimental results from the existing benchmark [7] and updated it with recent works [5,26,54]. We modified the AVSegFormer [14] with Mask2former [8] and denoted it as AVSegFormer* in the tables to encourage a fair comparison with our method. Please note that Tab. 1 includes two evaluation protocols. We employ standard semantic segmentation protocols to compute both mIoU and \\( F_{\\beta} \\) same as PascalVOC, initially mentioned in [57]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.682, + 0.676, + 0.697 + ], + "angle": 0, + "content": "4.2 Performance on Low-resolution AVSBench Videos" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.843 + ], + "angle": 0, + "content": "We adopt established methodologies [7, 30, 31, 37] for conducting performance evaluations on down-sampled image benchmarks, such as AVSBench-Objects [57], which includes single-source (SS) and multi-source (MS) splits with binary annotations, as well as the AVSBench-Semantics dataset [56], which contains multiclass annotations. We compare the performance of state-of-the-art (SOTA) methods with our CPM in Tab. 1 using mIoU and \\( F_{\\beta} \\). The results demonstrate that our model surpasses the second-best CAVP [7] in terms of mIoU by \\( 1.87\\% \\) on AVSBench-Object (SS) [57], \\( 2.83\\% \\) on AVSBench-Object (MS) [57] and \\( 1.79\\% \\) on AVSBench-Semantics [57] using the ResNet-50 [17] backbone." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.429, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Class-conditional Prompting Machine (CPM)" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.2 + ], + "angle": 0, + "content": "Table 2: Quantitative (mIoU, \\( F_{\\beta} \\)) audio-visual segmentation results (in %) for the AVSBench-Semantic (AVSS) test sets [56] (original resolution) with ResNet50 [17] backbone. Best results in bold, \\( 2^{nd} \\) best underlined. Improvements against the \\( 2^{nd} \\) best are in the last row." + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.201, + 0.723, + 0.297 + ], + "angle": 0, + "content": "
D-ResNet50 [17]MethodAVSS (SS)AVSS (MS)AVSS
mIoU ↑Fβ ↑mIoU ↑Fβ ↑mIoU ↑Fβ ↑
Per-pixel ClassificationTPAVI [57]42.1061.4626.3340.9943.3959.24
CAVP [7]56.9169.1538.6152.9250.7564.57
TransformerAVSegFormer [14]46.2559.7627.2141.3841.4856.21
AVSegFormer* [14]50.5263.7531.4042.8145.8059.16
CPM61.7172.9443.1156.2857.2570.54
ImprovementCPM+4.80+3.79+4.50+3.36+6.50+5.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.298, + 0.786, + 0.338 + ], + "angle": 0, + "content": "Table 3: Quantitative (mIoU, \\( F_{\\beta} \\)) audio-visual segmentation results (in %) for the VPO test sets (original resolution) with ResNet50 [17] backbone. Best results in bold, \\( 2^{nd} \\) best underlined. Improvements against the \\( 2^{nd} \\) best are in the last row." + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.34, + 0.723, + 0.437 + ], + "angle": 0, + "content": "
D-ResNet50 [17]MethodVPO (SS)VPO (MS)VPO (MSMI)
mIoU ↑Fβ ↑mIoU ↑Fβ ↑mIoU ↑Fβ ↑
Per-pixel ClassificationTPAVI [57]52.7569.5454.3071.9551.7368.85
CAVP62.3178.4664.3178.9260.3675.60
TransformerAVSegFormer [14]57.5573.0358.3374.2854.2270.39
AVSegFormer* [14]60.5174.8162.9177.3356.2472.67
CPM67.0979.8865.9179.9060.5575.58
ImprovementsCPM+4.78+1.42+1.60+0.98+0.19-0.02
" + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.45, + 0.621, + 0.466 + ], + "angle": 0, + "content": "4.3 Performance on Original AVSBench Videos" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.77 + ], + "angle": 0, + "content": "The benchmark mentioned above adopted low-resolution benchmarks using significantly resized input images (from 720p to \\(224 \\times 224\\)). While this simplifies model training, disregarding the original image aspect ratio can lead to a degradation in model performance, which is not advisable for segmentation tasks. Therefore, we follow the evaluation method outlined in [7] for training and testing on the raw AVSBench videos. This includes using random resized crops for training and performing frame-by-frame evaluation during testing. The initial published version of AVSBench [57] only offers single-source and multisource partitioning. However, this partitioning was not conducted in the later version [56], resulting in some cases being overlooked. We re-organised the previous SS and MS with the newly added video data, resulting in 1278 single-source videos (AVSS-SS) and 276 multi-source videos (AVSS-MS). We compare the results with mIoU and \\(F_{\\beta}\\) to demonstrate the comprehensive model performance in Tab. 2. Our method shows a significant mIoU improvement of \\(4.80\\%\\) on AVSS-SS, \\(4.50\\%\\) on AVSS-MS and \\(+6.50\\%\\) on the entire AVSS dataset. To further demonstrate the effectiveness of our method, we show a visualisation of 6-second video clip in Fig. 5 that displays a qualitative comparison between TPAVI, AVSegFormer, CAVP and our CPM. Our method can successfully approximate the ground truth segmentation of the target sound source within a group of other semantic objects." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.788, + 0.51, + 0.803 + ], + "angle": 0, + "content": "4.4 Performance on VPO Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We also compare the model performance on the VPO benchmark [7], equipped with synthesized stereo audios. We adopted ResNet-50 [17] backbone for all" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.127 + ], + "angle": 0, + "content": "Y. Chen et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.232, + 0.145, + 0.77, + 0.159 + ], + "angle": 0, + "content": "Table 4: Ablation study of the model components on AVSBench-Semantics [31]." + }, + { + "type": "table", + "bbox": [ + 0.244, + 0.16, + 0.747, + 0.256 + ], + "angle": 0, + "content": "
MethodMetrics
BaselineCCDMVCPACPPCLmIoU ↑Fβ ↑
✓53.0465.29
✓✓54.1266.65
✓✓✓55.7968.17
✓✓✓55.0768.03
✓✓✓✓56.4169.07
✓✓✓✓✓57.2570.54
" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.257, + 0.498, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.333, + 0.445, + 0.343 + ], + "angle": 0, + "content": "(a) Visual comparison." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.259, + 0.741, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.333, + 0.712, + 0.344 + ], + "angle": 0, + "content": "(b) Coverage score comparison." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.344, + 0.785, + 0.371 + ], + "angle": 0, + "content": "Fig. 3: Qualitative (3a) and quantitative (3b) comparisons between model components in a multi-source scenario (i.e., male singing, female singing and guitar)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.788, + 0.555 + ], + "angle": 0, + "content": "three subsets, and the results are shown in Tab. 3. To facilitate stereo audio encoding, we adopt the approach outlined in [7] by adjusting the number of input channels in the first layer to 2. Our approach surpasses the SOTA method CAVP [7] by \\(4.78\\%\\) and \\(1.60\\%\\) in terms of mIoU on VPO (SS) and VPO (MS), respectively. However, we observe a marginal improvement of \\(0.19\\%\\) on the VPO (MSMI) setup compared with the SS and MS subsets. The possible explanation for this phenomenon may stem from the overly simplistic integration of stereo audio within the transformer architecture, as the foundational AVS transformer architectures [14, 24, 26, 30] did not account for a distinct positional encoding scheme tailored for stereo data. We will explore the design of a dedicated stereo AVS transformer architecture as part of our future work." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.567, + 0.39, + 0.582 + ], + "angle": 0, + "content": "4.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.789, + 0.753 + ], + "angle": 0, + "content": "Ablation of Key Components We first perform the key components analysis of CPM on AVSBench-Semantics [56] in Tab. 4. The baseline is AVSegformer* [14] in the 1st row. We replace the Softmax classifier in baseline with CCDM defined in (1), and observe an mIoU improvement of \\(+1.08\\%\\). By integrating the model with VCP (3rd row), using training loss in (6), and ACP (4th row), using loss in (5), separately, we achieve an improvement of \\(+1.26\\%\\) and \\(+1.39\\%\\), respectively. Subsequently, when we apply both ACP and VCP methods (5th row) the mIoU performance improves \\(+2.04\\%\\) compared to the 2nd row. The final row displays the complete CPM method, incorporating the dense contrastive learning approach we introduced, which leads to an additional mIoU improvement of \\(+1.09\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Ablation of Bipartite Matching Stability To study the stability of the bipartite matching process, we design an entropy-based stability score STS to quantify such performance on AVSBench-Semantics [56]. Intuitively, the STS measures the average assignment consistency after bipartite matching across all the classes. During the training process, we collect the assigned label \\(\\tilde{\\mathbf{y}}_i \\in\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.429, + 0.114, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Class-conditional Prompting Machine (CPM)" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.183, + 0.479, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.287, + 0.498, + 0.316 + ], + "angle": 0, + "content": "Fig.4: Matching stability (STS \\(\\downarrow\\) ) comparison on AVSBench-Semantics [31]" + }, + { + "type": "table_caption", + "bbox": [ + 0.497, + 0.173, + 0.779, + 0.202 + ], + "angle": 0, + "content": "Table 5: Ablation of GMM modelling on AVSBench-Semantics [31]." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.203, + 0.744, + 0.305 + ], + "angle": 0, + "content": "
MethodmIoU ↑Fβ ↓
Point Rep.55.9668.58
Dist Rep.57.2570.54
MmIoUFβ
155.8368.68
357.2570.54
556.3169.04
756.0568.59
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.788, + 0.526 + ], + "angle": 0, + "content": "\\(\\{0, \\dots, C\\}^N\\) to the class-agnostic query \\(\\tilde{\\mathbf{q}}_n\\) produced by the Hungarian algorithm for each training image, resulting in \\(\\mathbf{R} = [\\tilde{\\mathbf{y}}_0 \\oplus \\tilde{\\mathbf{y}}_1 \\oplus, \\dots, \\oplus \\tilde{\\mathbf{y}}_D]\\), where \\(\\oplus\\) represents the concatenation operator and \\(\\mathbf{R}\\) has size \\(D \\times N\\). We then use the indicator function to extract the assignment information over all semantic classes. We denote the resulting set as \\(\\{\\mathbf{s}_0, \\mathbf{s}_1, \\dots, \\mathbf{s}_c\\}\\), where \\(\\mathbf{s}_c(n) = \\sum_{d=1}^{D} \\mathbb{1}(\\mathbf{R}(d,n) = c)\\) for \\(n \\in \\{1, \\dots, N\\}\\), which forms \\(\\mathbf{s}_c \\in [0,1]^N\\) after normalising it to be a probability distribution. Finally, we calculate the stability score \\(\\mathsf{STS} = \\frac{1}{C} \\sum_{c=1}^{C} H(\\mathbf{s}_c)\\), where \\(H(\\cdot)\\) computes the entropy. We compare our CPM with AVSegformer* over 100 training epochs as illustrated in Fig. 4. Our findings reveal that our CPM consistently improves the stability score (showing lower average entropy) throughout the training process in comparison to AVSegformer*, thereby validating the efficacy of our approach in enhancing bipartite matching stability. Please refer to the Supplementary Material for the Pseudo-code of STS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Ablation of Audio Contribution While experimenting with the baseline method mentioned above, we empirically observed that the classification after the audio transformer decoder (i.e., \\( f^{\\mathrm{TD - A}} \\) in Fig. 2) has low accuracy, with noisy predictions. We also found that these noisy predictions are progressively refined by the vision transformer decoder layer based solely on visual clues. Such an observation illustrates that in some hard cases, we may fail to extract useful semantic information from audio, leading to the model being overly reliant on the visual content, which may cause erroneous testing predictions, as shown in Fig. 3a-(i), (ii), (iii). To monitor the amount of valid information retained by the query following interaction with the audio branch, we introduce a classification coverage score. This score is computed by comparing the predictions generated after passing through the audio transformer decoder with the ground-truth class label, using the Jaccard index. We show qualitative results in Fig. 3a for the designated scenario and quantitative results in Fig. 3b for the AVSBench-Semantics testing set. The comparison starts with AVSegFormer as the baseline method with progressive addition of our proposed modules, similar to Tab. 4. Notice that each CPM sub-module contributes to refining key semantic information from the audio modality, as evidenced by the segmentation mask predictions and the improvement in Jaccard indexes. This shows the effectiveness of CPM." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.127 + ], + "angle": 0, + "content": "Y. Chen et al." + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.133, + 0.726, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.299, + 0.785, + 0.34 + ], + "angle": 0, + "content": "Fig. 5: Qualitative audio-visual segmentation results on AVSBench-Semantics [56] by TPAVI [57], AVSegFormer [14], CAVP [7] and our CPM, which can be compared with the ground truth (GT) Ambulance of the first row." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.54 + ], + "angle": 0, + "content": "Ablation of Distribution Representation Table 5 provides a study of the GMM data distribution representation on AVSBench-Semantics [56]. The first row (Point Rep.) replaces the GMM module with a linear SoftMax for the classification output, introducing a learnable \\( C \\times D_q \\) feature map (i.e., \\( D_q \\)-dimensional feature embeddings for \\( C \\) semantic classes for the audio and visual prompting task). The second row (Dist Rep.) shows our approach with the CCDM module, showing that the utilisation of the distribution modelling leads to considerable improvements of \\( 1.29\\% \\) and \\( 1.96\\% \\) on mIoU and \\( F_{\\beta} \\) respectively. The next rows of Tab. 5 show a hyper-parameter analysis on the number of GMM components (from \\( M = 1 \\) to \\( M = 7 \\) components). Results reveal that 3 GMM components enable the best fitting for the data distribution, leading to the top performance with a maximum improvement of \\( 1.42\\% \\) and \\( 1.95\\% \\) on mIoU and \\( F_{\\beta} \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.55, + 0.51, + 0.566 + ], + "angle": 0, + "content": "5 Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.748 + ], + "angle": 0, + "content": "We introduced CPM, a new audio-visual training method designed for the Maskformer based framework to enhance bipartite matching stability and improve the efficacy of cross-modal attention for audio-visual segmentation. We proposed a class-conditional prompting learning strategy that combines class-agnostic queries with class-conditional queries, sampled from our iteratively updated generative model. The generated class-conditional queries are utilised to probe both the magnitude spectrogram and the image feature map, aiming to remove off-the-screen noise and bypass bipartite matching to produce a more stable learning process. Lastly, we extend the class-conditional queries to a new prompting-based audio-visual contrastive learning to explicitly constrain the cross-modal representations. SOTA results on AVS benchmarks suggest that CPM can be a valuable resource for future AVS research." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Limitations and future work. We acknowledge that the current adaptation of stereo audio into the transformer-based method has limitations, as it encodes positional and semantic information jointly within the transformer block. This contrasts with the typical approach of separately encoding these two types of information in the transformer-based framework. Our future work will concentrate on optimizing this framework through the integration of spatial reasoning." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.429, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Class-conditional Prompting Machine (CPM)" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.223 + ], + "angle": 0, + "content": "1. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.223, + 0.785, + 0.263 + ], + "angle": 0, + "content": "2. Chen, H., Xie, W., Afouras, T., Nagrani, A., Vedaldi, A., Zisserman, A.: Localizing visual sounds the hard way. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16867-16876 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.264, + 0.785, + 0.305 + ], + "angle": 0, + "content": "3. Chen, J., Lu, J., Zhu, X., Zhang, L.: Generative semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7111-7120 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.307, + 0.785, + 0.348 + ], + "angle": 0, + "content": "4. Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 801-818 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.349, + 0.785, + 0.39 + ], + "angle": 0, + "content": "5. Chen, T., Tan, Z., Gong, T., Chu, Q., Wu, Y., Liu, B., Lu, L., Ye, J., Yu, N.: Bootstrapping audio-visual segmentation by strengthening audio cues. arXiv preprint arXiv:2402.02327 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.785, + 0.432 + ], + "angle": 0, + "content": "6. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.785, + 0.46 + ], + "angle": 0, + "content": "7. Chen, Y., Liu, Y., Wang, H., Liu, F., Wang, C., Carneiro, G.: A closer look at audio-visual semantic segmentation. arXiv e-prints pp. arXiv-2304 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.785, + 0.516 + ], + "angle": 0, + "content": "8. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290–1299 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.517, + 0.785, + 0.558 + ], + "angle": 0, + "content": "9. Cheng, B., Schwing, A., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems 34, 17864-17875 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.559, + 0.785, + 0.601 + ], + "angle": 0, + "content": "0. Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society: series B (methodological) 39(1), 1-22 (1977)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.602, + 0.785, + 0.642 + ], + "angle": 0, + "content": "1. Everingham, M., Eslami, S.A., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes challenge: A retrospective. International journal of computer vision 111, 98-136 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.643, + 0.785, + 0.685 + ], + "angle": 0, + "content": "2. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. International journal of computer vision 88, 303-338 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.686, + 0.785, + 0.727 + ], + "angle": 0, + "content": "3. Gao, R., Grauman, K.: Co-separating sounds of visual objects. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3879-3888 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.728, + 0.785, + 0.755 + ], + "angle": 0, + "content": "4. Gao, S., Chen, Z., Chen, G., Wang, W., Lu, T.: Avsegformer: Audio-visual segmentation with transformer. arXiv preprint arXiv:2307.01146 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.71, + 0.77 + ], + "angle": 0, + "content": "5. Gray, R.: Vector quantization. IEEE Assp Magazine 1(2), 4-29 (1984)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.785, + 0.798 + ], + "angle": 0, + "content": "6. Hao, D., Mao, Y., He, B., Han, X., Dai, Y., Zhong, Y.: Improving audio-visual segmentation with bidirectional generation. arXiv preprint arXiv:2308.08288 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.128 + ], + "angle": 0, + "content": "Y. Chen et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "18. Hu, X., Chen, Z., Owens, A.: Mix and localize: Localizing sound sources in mixtures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10483-10492 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.231 + ], + "angle": 0, + "content": "19. Huang, S., Li, H., Wang, Y., Zhu, H., Dai, J., Han, J., Rong, W., Liu, S.: Discovering sounding objects by audio queries for audio visual segmentation. arXiv preprint arXiv:2309.09501 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "20. Jaegle, A., Borgeaud, S., Alayrac, J.B., Doersch, C., Ionescu, C., Ding, D., Koppula, S., Zoran, D., Brock, A., Shelhamer, E., et al.: Perceiver io: A general architecture for structured inputs & outputs. arXiv preprint arXiv:2107.14795 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "21. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in neural information processing systems 33, 18661-18673 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "22. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.397 + ], + "angle": 0, + "content": "23. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13619-13627 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "24. Li, K., Yang, Z., Chen, L., Yang, Y., Xun, J.: Catr: Combinatorial-dependence audio-queried transformer for audio-visual video segmentation. arXiv preprint arXiv:2309.09709 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.493 + ], + "angle": 0, + "content": "25. Li, T., Cao, P., Yuan, Y., Fan, L., Yang, Y., Feris, R.S., Indyk, P., Katabi, D.: Targeted supervised contrastive learning for long-tailed recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6918-6928 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.535 + ], + "angle": 0, + "content": "26. Li, X., Wang, J., Xu, X., Peng, X., Singh, R., Lu, Y., Raj, B.: Towards robust audiovisual segmentation in complex environments with quantization-based semantic decomposition. arXiv preprint arXiv:2310.00132 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "27. Liang, C., Wang, W., Miao, J., Yang, Y.: Gmmseg: Gaussian mixture based generative semantic segmentation models. Advances in Neural Information Processing Systems 35, 31360-31375 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.618 + ], + "angle": 0, + "content": "28. Liu, C., Ding, H., Jiang, X.: Gres: Generalized referring expression segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23592-23601 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.646 + ], + "angle": 0, + "content": "29. Liu, C., Li, P., Qi, X., Zhang, H., Li, L., Wang, D., Yu, X.: Audio-visual segmentation by exploring cross-modal mutual semantics (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "30. Liu, C., Li, P., Zhang, H., Li, L., Huang, Z., Wang, D., Yu, X.: Bavs: Bootstrapping audio-visual segmentation by integrating foundation knowledge. arXiv preprint arXiv:2308.10175 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.729 + ], + "angle": 0, + "content": "31. Liu, J., Ju, C., Ma, C., Wang, Y., Wang, Y., Zhang, Y.: Audio-aware query-enhanced transformer for audio-visual segmentation. arXiv preprint arXiv:2307.13236 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.757 + ], + "angle": 0, + "content": "32. Liu, J., Wang, Y., Ju, C., Zhang, Y., Xie, W.: Annotation-free audio-visual segmentation. arXiv preprint arXiv:2305.11019 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "33. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "34. Liu, X., Carrington, P., Chen, X., Pavel, A.: What makes videos accessible to blind and visually impaired people? In: Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. pp. 1-14 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.429, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Class-conditional Prompting Machine (CPM)" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "35. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "36. Mao, Y., Zhang, J., Xiang, M., Lv, Y., Zhong, Y., Dai, Y.: Contrastive conditional latent diffusion for audio-visual segmentation. arXiv preprint arXiv:2307.16579 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.232, + 0.787, + 0.273 + ], + "angle": 0, + "content": "37. Mao, Y., Zhang, J., Xiang, M., Zhong, Y., Dai, Y.: Multimodal variational autoencoder based audio-visual segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 954–965 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "38. Martin, D.R., Fowlkes, C.C., Malik, J.: Learning to detect natural image boundaries using local brightness, color, and texture cues. IEEE transactions on pattern analysis and machine intelligence 26(5), 530-549 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.315, + 0.787, + 0.357 + ], + "angle": 0, + "content": "39. Meng, D., Chen, X., Fan, Z., Zeng, G., Li, H., Yuan, Y., Sun, L., Wang, J.: Conditional detr for fast training convergence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3651-3660 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.357, + 0.787, + 0.384 + ], + "angle": 0, + "content": "40. Mo, S., Morgado, P.: A closer look at weakly-supervised audio-visual source localization. arXiv preprint arXiv:2209.09634 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.384, + 0.787, + 0.425 + ], + "angle": 0, + "content": "41. Mo, S., Morgado, P.: Localizing visual sounds the easy way. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII. pp. 218-234. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.425, + 0.787, + 0.467 + ], + "angle": 0, + "content": "42. Mo, S., Tian, Y.: Audio-visual grouping network for sound localization from mixtures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10565–10574 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.467, + 0.787, + 0.495 + ], + "angle": 0, + "content": "43. Mo, S., Tian, Y.: Av-sam: Segment anything model meets audio-visual localization and segmentation. arXiv preprint arXiv:2305.01836 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.495, + 0.787, + 0.522 + ], + "angle": 0, + "content": "44. Murray, M.M., Wallace, M.T.: The neural bases of multisensory processes. Frontiers in Neuroscience (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.522, + 0.787, + 0.55 + ], + "angle": 0, + "content": "45. Reynolds, D.A., et al.: Gaussian mixture models. Encyclopedia of biometrics 741(659-663) (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.55, + 0.787, + 0.592 + ], + "angle": 0, + "content": "46. Senocak, A., Ryu, H., Kim, J., Oh, T.H., Pfister, H., Chung, J.S.: Sound source localization is all about cross-modal alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7777-7787 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.592, + 0.787, + 0.633 + ], + "angle": 0, + "content": "47. Siam, M., Oreshkin, B.N., Jagersand, M.: Amp: Adaptive masked proxies for few-shot segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5249-5258 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.633, + 0.787, + 0.675 + ], + "angle": 0, + "content": "48. Sun, Z., Cao, S., Yang, Y., Kitani, K.M.: Rethinking transformer-based set prediction for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3611-3620 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.675, + 0.787, + 0.715 + ], + "angle": 0, + "content": "49. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.716, + 0.787, + 0.758 + ], + "angle": 0, + "content": "50. Wang, W., Zhou, T., Yu, F., Dai, J., Konukoglu, E., Van Gool, L.: Exploring cross-image pixel contrast for semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7303-7313 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "51. Wang, Y., Liu, W., Li, G., Ding, J., Hu, D., Li, X.: Prompting segmentation with sound is generalizable audio-visual source localizer. arXiv preprint arXiv:2309.07929 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "52. Wu, J., Jiang, Y., Sun, P., Yuan, Z., Luo, P.: Language as queries for referring video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4974-4984 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.368, + 0.127 + ], + "angle": 0, + "content": "Y. Chen et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "53. Xu, P., Zhu, X., Clifton, D.A.: Multimodal learning with transformers: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.217 + ], + "angle": 0, + "content": "54. Yang, Q., Nie, X., Li, T., Gao, P., Guo, Y., Zhen, C., Yan, P., Xiang, S.: Cooperation does matter: Exploring multi-order bilateral relations for audio-visual segmentation. arXiv preprint arXiv:2312.06462 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "55. Zhang, H., Li, F., Xu, H., Huang, S., Liu, S., Ni, L.M., Zhang, L.: Mp-former: Mask-piloted transformer for image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18074-18083 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.3 + ], + "angle": 0, + "content": "56. Zhou, J., Shen, X., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., et al.: Audio-visual segmentation with semantics. arXiv preprint arXiv:2301.13190 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "57. Zhou, J., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., Zhong, Y.: Audio-visual segmentation. In: European Conference on Computer Vision. pp. 386-403. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.385 + ], + "angle": 0, + "content": "58. Zhou, T., Zhang, M., Zhao, F., Li, J.: Regional semantic contrast and aggregation for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4299-4309 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.385 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CPM_ Class-conditional Prompting Machine for Audio-visual Segmentation/462b4fbe-3dda-4385-88f4-05811a4623a9_origin.pdf b/2024/CPM_ Class-conditional Prompting Machine for Audio-visual Segmentation/462b4fbe-3dda-4385-88f4-05811a4623a9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..91ab6dcf6706e5f88ddcaa665cd9a2a42b89055f --- /dev/null +++ b/2024/CPM_ Class-conditional Prompting Machine for Audio-visual Segmentation/462b4fbe-3dda-4385-88f4-05811a4623a9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb59c8308e466f6d46acc84ce85cc34a8dd0c0c78c94f5162b00db600c8b013a +size 13838326 diff --git a/2024/CPM_ Class-conditional Prompting Machine for Audio-visual Segmentation/full.md b/2024/CPM_ Class-conditional Prompting Machine for Audio-visual Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b60e1abfbb1a5d9085962e832ade1ed0cce8d956 --- /dev/null +++ b/2024/CPM_ Class-conditional Prompting Machine for Audio-visual Segmentation/full.md @@ -0,0 +1,280 @@ +# CPM: Class-conditional Prompting Machine for Audio-visual Segmentation + +Yuanhong Chen1, Chong Wang1, Yuyuan Liu1, Hu Wang2, and Gustavo Carneiro3 + +$^{1}$ Australian Institute for Machine Learning, University of Adelaide, Australia + +$^{2}$ Mohamed bin Zayed University of Artificial Intelligence, United Arab Emirates + +3 Centre for Vision, Speech and Signal Processing, University of Surrey yuanhong.chen@adelaide.edu.au + +Abstract. Audio-visual segmentation (AVS) is an emerging task that aims to accurately segment sounding objects based on audio-visual cues. The success of AVS learning systems depends on the effectiveness of cross-modal interaction. Such a requirement can be naturally fulfilled by leveraging transformer-based segmentation architecture due to its inherent ability to capture long-range dependencies and flexibility in handling different modalities. However, the inherent training issues of transformer-based methods, such as the low efficacy of cross-attention and unstable bipartite matching, can be amplified in AVS, particularly when the learned audio query does not provide a clear semantic clue. In this paper, we address these two issues with the new Class-conditional Prompting Machine (CPM). CPM improves the bipartite matching with a learning strategy combining class-agnostic queries with class-conditional queries. The efficacy of cross-modal attention is upgraded with new learning objectives for the audio, visual and joint modalities. We conduct experiments on AVS benchmarks, demonstrating that our method achieves state-of-the-art (SOTA) segmentation accuracy4. + +Keywords: Audio-visual Learning $\cdot$ Segmentation $\cdot$ Multi-modal Learning + +# 1 Introduction + +The recognition and integration of auditory and visual data are fundamental to human cognitive processes, playing a critical role in facilitating meaningful communication [44]. Audio-visual segmentation (AVS) is an emerging cross-modal reasoning task that mimics such cognitive processes, aiming to localize visual objects based on audio-visual cues. AVS has many applications, such as the automatic localisation and identification of sounding objects to improve the accessibility of videos for the blind and visually impaired [34]. A major challenge in AVS is achieving effective cross-modal interaction between sound and visual + +![](images/6fa6c9fc6131d3e7717ec93b0fa998a611d80393b2f55380e0e67d2b8777a19d.jpg) +Fig. 1: Comparing conventional AVS methods [7, 14] with our CPM approach, CPM inherits the class-agnostic query from transformer-based methods and integrates class-conditional prompts sampled from the learned joint-modal data distribution to achieve three objectives: 1) learn disentangled audio partitioning, 2) facilitate semantic-guided object identification, and 3) promote more explicit audio-visual contrastive learning. + +objects [46]. Many AVS methods [7, 37, 57] generally adopt a traditional per-pixel classification framework [8, 52], utilising early fusion strategies (e.g., cross-attention fusion [49]) together with an FCN decoder [4, 35] to make predictions. Such per-pixel design has achieved good performance, but it tends to underutilise the audio data due to its lower informativeness compared with the visual data [7, 46]. Contrastive learning can mitigate this issue by matching informative audio-visual pairs [7]. However, another critical limitation of the per-pixel design is its failure to capture instance-level visual information, resulting in inconsistent segmentation predictions within or between frames in a video sequence [52]. + +These two issues have been addressed by transformer-based AVS methods designed to capture instance-level information and to rely on more effective contrastive learning [14, 24, 28, 30, 31, 52, 54]. Nevertheless, these AVS approaches still show slow convergence and relatively poor accuracy [23, 33, 55], primarily attributed to the low efficacy of the cross-attention [48] and the unstable bipartite matching [23]. Solutions for these problems are based on integrated masked-attention [8], which focuses on features around predicted segments, and anchor de-noising [23] to reconstruct the noisy ground-truth bounding box detection. However, these solutions may not work well in AVS because of the weak constraint provided by global audio features that contain a mixture of sound sources [14, 26, 30] resulting in more instability during training. It is worth noting that all methods above have the common issue of relying on class-agnostic prompts that provide little guidance to the bipartite matching process, thereby reducing training efficacy. + +In this paper, we introduce the Class-conditional Prompting Machine (CPM), an audio-visual segmentation training approach that leverages class-conditional prompts to enhance bipartite matching stability and improve cross-modal attention efficacy. To enhance bipartite matching, we introduce a novel learning method, combining class-agnostic queries [8,14,26] with class-conditional queries, + +sampled from our iteratively updated generative model of class-specific embeddings, with the former queries matched to ground-truth labels using the Hungarian Algorithm [1], followed by individual processing of class-conditional queries in two modalities. To improve cross-modal attention efficacy, new learning objectives are proposed for both audio and visual modalities. In the audio modality (audio conditional prompting, denoted as ACP), the original spectrogram is corrupted with auditory off-the-screen noise and class-conditional queries are employed to reconstruct the original spectrogram, while in the visual modality (visual conditional prompting, denoted as VCP), noisy class-conditional queries sampled from our generative model are used to probe semantically similar content in the image space. To further upgrade the performance of cross-modal attention, we introduce a new prompting-based audio-visual contrastive learning (PCL) task, guided by class-specific queries to densely constrain the cross-modal representations. To summarise, our main contributions are: + +- A new AVS training approach to enhance bipartite matching stability and enhance the efficacy of cross-modal attention. Our core innovation lies in the development of a Class-conditional Prompting Machine (CPM). +- To improve the bipartite matching, we propose a new AVS learning strategy that combines class-agnostic queries with class-conditional queries, sampled from our iteratively updated generative model of class-specific embeddings. +- The efficacy of cross-modal attention is upgraded with new learning objectives for the audio, visual and joint modalities. For audio, we present ACP that perturbs the original spectrogram and uses class-conditional queries to reconstruct the spectrogram; for visual, we introduce VCP that explores noisy class-conditional queries sampled from our generative model to probe the corresponding semantic in visual feature map. To enhance the cross-modal attention efficacy further, we propose PCL, consisting of a new prompting-based joint (audio-visual) contrastive learning task. + +We firstly show the effectiveness of our CPM model through rigorous evaluation on established benchmarks such as AVSBench-Objects [57] and AVSBench-Semantics [56]. Furthermore, we extend our evaluation by including VPO synthetic benchmarks [7], aiming to enhance our comprehension of AVS methods' capacity to capture audio-visual correlations. Our findings across these benchmarks consistently demonstrate that our approach yields better classification accuracy compared to existing methods. + +# 2 Related Works + +# 2.1 Transformer-based architecture + +Transformer-based methods have shown promising performance in detection [1] and segmentation [8,9] benchmarks. The fundamental concept is to leverage the object query to probe image features from the output of transformer encoders and bipartite graph matching to perform set-based box/mask prediction. It is + +also evidenced that such a framework can benefit multi-modal learning due to its attention mechanism and flexibility in data modelling [53]. Despite its successful application in various domains [1, 8], such frameworks have also been reported to exhibit a poorer convergence rate compared to traditional CNN-based methods [8,23,33]. In exploring the reasons behind the poor convergence, previous studies have emphasized enhancing the interpretability of the learnable query [33, 39] (often referring to them as anchor points/boxes). Alternatively, the utilisation of denoising methods [23, 55] has also been adopted to facilitate the learning of bounding box offset [23] and improve the utilisation of the transformer decoder layers [55]. Furthermore, masked attention has proven effective in enhancing both convergence rate and model performance [8]. The successful implementation of denoising methods and masked attention mechanisms has inspired us to devise a mitigation strategy for bipartite matching that promotes improved cross-modal understanding within the cross-attention process. + +# 2.2 Audio-visual Segmentation (AVS) + +Audio-visual Segmentation is a dense classification task or detection of sounding visual objects in videos, using image sequences and audio cues. Zhou et al. [57] introduced the AVSBench-Object and AVSBench-Semantics benchmarks [56], which enable the evaluations of single-source and multi-source AVS tasks for salient and multi-class object segmentation. To address concerns related to the high annotation cost and dataset diversity, Chen et al. [7] introduced a cost-effective strategy to build a relatively unbiased AVS dataset, named Visual Post-production (VPO). Mainstream AVS methods use audio as the reference query [7, 14, 19, 24, 29–32] or prompt [43, 51]. For example, some methods adopt MaskFormer [9] or segment anything model (SAM) [22] to perform image segmentation using audio queries or encoded audio prompts and cross-attention layers. These methods benefit from the attention mechanism's ability and mask-classification's features to capture long-range dependencies and enhance image segmentation ability [9], spatial-temporal reasoning [24] and task-related features [14, 24, 29, 30, 37, 57]. The adaptation of the Maskformer-based framework for AVS relies on methods to encourage the audio-visual semantic alignment [24, 30, 31, 54]. Such strategy mitigates the poor audio semantic information caused by modality imbalance [5] and learns disentangled multi-modal representations [26, 54] via vector quantization [15] and bidirectional attention mechanism. Even though the query-based transformer architecture [1] has shown great success in semantic segmentation tasks, certain weaknesses such as low efficacy of cross-attention and unstable bipartite matching process [23] are still only partially addressed. These two issues may be exacerbated in AVS due to the poor audio semantic information. + +# 2.3 Audio-visual Contrastive Learning + +Contrastive Learning has shown promising results in audio-visual learning (AVL) methods [2, 7, 18, 40, 41]. These methods bring together augmented representations from the same instance as positives while separating representation pairs + +from different instances as negatives within a batch. A reported issue with current AVL contrastive learning is its reliance on self-supervision [6] to connect audio and visual representations of the same class. To overcome this issue, CAVP [7] propose a supervised contrastive learning [21, 25, 50] method that mines informative contrastive pairs from arbitrary audio-visual pairings to constrain the learning of audio-visual embeddings. Nevertheless, these previous approaches predominantly leverage global audio representations, thereby limiting the model's capacity to discern individual audio sources. Consequently, this constraint reduces the model's effectiveness in scenarios involving multiple sound sources. In our work, we employ class-specific queries to retrieve the corresponding audio representations, which facilitates a more explicit form of contrastive learning between audio and visual modalities. + +# 3 Method + +We denote a multi-class audio-visual dataset as $\mathcal{D} = \{(\mathbf{a}_i, \mathbf{x}_i, \mathbf{y}_i, \mathbf{t}_i))\}_{i=1}^{|\mathcal{D}|}$ , where $\mathbf{x}_i \in \mathcal{X} \subset \mathbb{R}^{H \times W \times 3}$ is an RGB image with resolution $H \times W$ , $\mathbf{a} \in \mathcal{A} \subset \mathbb{R}^{T \times F}$ denotes the magnitude spectrogram of the audio data with $T$ time and $F$ frequency bins, $\mathbf{y}_i \in \mathcal{Y} \subset \{0,1\}^{H \times W \times C}$ denotes the pixel-level ground truth for the $C$ classes (the background class is included in these $C$ classes), and $\mathbf{t}_i \in \mathcal{Y} \subset \{0,1\}^C$ is a multi-label ground truth audio annotation. We use Mask2former [8] as the segmentation framework. + +# 3.1 Preliminaries using Class-agnostic Queries + +Like other Maskformer-based methods [14,30,31,52], we aim to learn the parameters $\theta \in \Theta$ for the model $f_{\theta}:\mathcal{X}\times \mathcal{A}\to [0,1]^{H\times W\times C}$ , which comprises the image and audio backbones that extract features with $\mathbf{u}_a = f_\gamma (\mathbf{a})$ and $\mathbf{u}_v = f_\phi (\mathbf{x})$ , respectively, where $\gamma ,\phi \in \theta$ , and $\mathbf{u}_a,\mathbf{u}_v\in \mathcal{U}$ , with $\mathcal{U}$ denoting a unified feature space. A set of learnable query features (comprising the object query feature and positional query embeddings) are defined as the joint-modal output embeddings similar to the Perceiver model [20]. We define the class-agnostic query feature as $\mathbf{q}\in \mathcal{Q}\subset \mathbb{R}^{N\times D_q}$ , where $N$ denotes the number of class-agnostic queries, and $D_{q}$ represents the dimensionality of the feature space. As depicted in Fig. 2, given $\mathbf{q}$ , we aim to group pixels with matched semantic information from $\mathbf{u}_a$ and $\mathbf{u}_v$ through consecutive transformer decoder layers [8,49] and generate $N$ mask embeddings $\tilde{\mathbf{q}}\in \mathcal{Q}$ and pixel embeddings $\tilde{\mathbf{u}}_v\in \mathbb{R}^{H\times W\times D_q}$ . Then, the model independently predicts the embeddings into $N$ set of class predictions (via Softmax(·)) and mask predictions (via Sigmoid(·)) denoted as $\mathcal{E}_{\mathrm{pred}} = \left\{\left(\mathbf{m}_i,\mathbf{p}_i\right)\right\}_{i = 1}^N$ , where $\mathbf{m}\in \mathcal{M}\subset \{0,1\}^{H\times W}$ and $\mathbf{p}\in \mathcal{P}\subset \{0,1\}^C$ . We denote the ground-truth set derived from the training set $\mathcal{D}$ with $\mathcal{E}_{\mathrm{gt}} = \left\{\left(\mathbf{m}_i^{\mathrm{gt}},\mathbf{p}_i^{\mathrm{gt}}\right)\right\}_{i = 1}^{N^{\mathrm{gt}}}$ . During training, we use the Hungarian algorithm to perform optimal matching between $\mathcal{E}_{\mathrm{pred}}$ and $\mathcal{E}_{\mathrm{gt}}$ [1,8] to facilitate the label assignment. Since $|\mathcal{E}_{\mathrm{pred}}| > |\mathcal{E}_{\mathrm{gt}}|$ , we pad $\mathcal{E}_{\mathrm{gt}}$ with no-object class $\varnothing$ to achieve one-to-one matching [8]. + +The loss $\ell$ to train the model $f_{\theta}(.)$ includes a cross-entropy query classification loss $\ell_{ce}$ , and a binary mask loss $\ell_{mask} = \ell_{focal} + \ell_{dice}$ , which combines a focal loss + +![](images/81a0499895e82667d6bdd7d30964ddf0da46833ba0f91263495d8710a773d9dc.jpg) +Fig. 2: Illustration of our CPM method. Starting with a scene with a mixture of sound sources including Male, Female and Guitar, the training alternates between the use of learnable class-agnostic queries, and queries sampled from a list of class-specific query features from the GMM, denoted as $\mathbf{z}^{\mathrm{male}}$ , $\mathbf{z}^{\mathrm{female}}$ and $\mathbf{z}^{\mathrm{guitar}}$ . The overall training objective is composed of three learning tasks: 1) in audio conditional prompting (ACP), we aim to use $\mathbf{z}^{\mathrm{male}}$ , $\mathbf{z}^{\mathrm{female}}$ and $\mathbf{z}^{\mathrm{guitar}}$ to recover the original magnitude spectrogram $\mathbf{a}_i$ from the noise spectrogram (i.e., $\mathbf{a}_i + \mathbf{a}_j$ ) that is corrupted by another Dog audio signal $\mathbf{a}_j$ ; 2) a visual conditional prompting (VCP) that aim to probe the corresponding pixels w.r.t to the class-specific query features; and 3) a contrastive learning task that target to densely constrain the audio and visual representations. For training, both the CPM Workflow, indicated by the orange arrow, and the Class-agnostic Workflow, marked by the black arrow, are utilized. However, only the Class-agnostic Workflow is used for inference. + +$\ell_{focal}$ and a dice loss $\ell_{dice}$ [8,14,30]. The overall training loss for the class-agnostic query feature is defined as $\ell_{agn} = \ell_{ce} + \ell_{mask}$ . During testing, $\mathbf{p}$ and $\mathbf{m}$ are merged via the multiplication $\arg \max_{i:c_i\neq \emptyset}\mathbf{p}(c_i)\cdot \mathbf{m}_i$ , where $c_{i}$ is the class label with maximum likelihood $c_{i} = \arg \max_{c\in \{1,\dots,C,\emptyset \}}\mathbf{p}_{i}(c)$ for each probability-mask pair indexed by $i$ . + +# 3.2 Class-conditional Prompting Machine (CPM) + +The proposed CPM training builds upon the class-agnostic training method from Sec. 3.1 by introducing a new training stage based on class-conditional prompting. This allows the probing of both the magnitude spectrogram and the image feature map, aiming to mitigate unstable training issues and improve cross-attention efficacy. In some cases, the prompts can be manually crafted (i.e., text label, bounding box) based on domain knowledge or specific task requirements [22]. Prompts can also be automatically learned to form a fixed set of prototypical embeddings for each class [42]. However, it is impossible to manually define class conditional prompts in high-dimensional latent spaces, and using a limited-size set of learned prompts may fail to capture the comprehensive distribution of class-specific prompts. Hence, it is desirable that these + +class-specific prompts can be sampled from a generative model [3,27]. that comprehensively represents the respective class. We adopt the Gaussian Mixture Models (GMMs) as such generative model [45], which improves the intra-class variability and increases robustness to class imbalances when compared to the alternative approaches mentioned before. Before delving into the methodology of the CPM, we first introduce the generation process of the class-conditional query features. + +Class Conditional Distribution Modelling (CCDM) Instead of constructing a posterior $p(c|\tilde{\mathbf{q}})$ with Softmax classifier for the mask embeddings $\tilde{\mathbf{q}}$ , the generative classifier employs the Bayes rule for label prediction, estimating the class-conditional distribution $p(\tilde{\mathbf{q}}|c)$ alongside the class prior $p(c)$ [27], with: + +$$ +p (c | \tilde {\mathbf {q}}) = \frac {p (\tilde {\mathbf {q}} | c) p (c)}{\sum_ {c ^ {\prime}} p (\tilde {\mathbf {q}} | c ^ {\prime}) p \left(c ^ {\prime}\right)}, \tag {1} +$$ + +where the class probability prior $p(c)$ is uniform. To enable GMM modelling in Maskformer architecture [8, 9], we replace the $\mathbf{p} = \operatorname{Softmax}(\tilde{\mathbf{q}})$ activation function with $f_{\mathrm{GMM}}(\tilde{\mathbf{q}}) = p(c|\tilde{\mathbf{q}})$ defined in (1) which maps the input mask embeddings $\tilde{\mathbf{q}}$ to probability data density over $C$ classes. Assuming that after the Hungarian matching between the predictions set $\{(\mathbf{m}_i,\mathbf{p}_i)\}_{i=1}^N$ and $\mathbf{y}$ , we get a new one-to-one matched ground-truth set $\tilde{\mathbf{y}}$ w.r.t each prediction by padding the $\mathbf{y}$ with no-object class $\varnothing$ . Subsequently, we can form the GMM training dataset $\mathcal{F} = \{(\tilde{\mathbf{q}}_n,\tilde{\mathbf{y}}_n)\}_{n=1}^F$ by pairing the mask embeddings with the assigned label. In our method, the goal of GMM is to model the data distribution of the joint-modal mask embedding $\tilde{\mathbf{q}}$ for each class $C$ in the $D_q$ -dimensional space by employing a weighted mixture of $M$ multivariate Gaussians, defined as follows: + +$$ +p (\tilde {\mathbf {q}} | c) = \sum_ {m = 1} ^ {M} p (m | c; \boldsymbol {\pi} _ {c}) p (\tilde {\mathbf {q}} | c, m; \boldsymbol {\mu} _ {c}, \boldsymbol {\Sigma} _ {c}) = \sum_ {m = 1} ^ {M} \boldsymbol {\pi} _ {c m} \mathcal {N} (\tilde {\mathbf {q}}; \boldsymbol {\mu} _ {c m}, \boldsymbol {\Sigma} _ {c m}), \tag {2} +$$ + +where $\sum_{m}\pi_{cm} = 1$ represents the mixing coefficients, $\pmb{\mu}_{cm}\in \mathbb{R}^{D_q}$ denotes the mean vector, and $\Sigma_{cm}\in \mathbb{R}^{D_q\times D_q}$ is the covariance matrix. The optimisation of the GMM parameters is performed by the Expectation Maximisation (EM) algorithm [10]. In the E-step, we iterate through $\mathcal{F}$ , computing the responsibilities using the current parameters of the GMM for each class $c$ : + +$$ +\gamma_ {c, n} ^ {(t)} (m) = \frac {\pi_ {c m} ^ {(t - 1)} \mathcal {N} \left(\tilde {\mathbf {q}} _ {n} ; \boldsymbol {\mu} _ {c m} ^ {(t - 1)} , \boldsymbol {\Sigma} _ {c m} ^ {(t - 1)}\right)}{\sum_ {m ^ {\prime} = 1} ^ {M} \pi_ {c m ^ {\prime}} ^ {(t - 1)} \mathcal {N} \left(\tilde {\mathbf {q}} _ {n} ; \boldsymbol {\mu} _ {c m ^ {\prime}} ^ {(t - 1)} , \boldsymbol {\Sigma} _ {c m ^ {\prime}} ^ {(t - 1)}\right)}. \tag {3} +$$ + +We re-estimate the parameters using the calculated responsibility in the M-step. + +$$ +\boldsymbol {\mu} _ {c m} ^ {(t)} = \frac {1}{N _ {c m} ^ {(t)}} \sum_ {(\tilde {\mathbf {q}} _ {n}, \tilde {\mathbf {y}} _ {n}) \in \mathcal {F} _ {c}} \gamma_ {c, n} ^ {(t)} (m) \tilde {\mathbf {q}}, \quad \boldsymbol {\pi} _ {c m} ^ {(t)} = \frac {N _ {c m} ^ {(t)}}{| \mathcal {F} |}, \tag {4} +$$ + +$$ +\boldsymbol {\Sigma} _ {c m} ^ {(t)} = \frac {1}{N _ {c m} ^ {(t)}} \sum_ {(\tilde {\mathbf {q}} _ {n}, \tilde {\mathbf {y}} _ {n}) \in \mathcal {F} _ {c}} \gamma_ {c, n} ^ {(t)} (m) (\tilde {\mathbf {q}} _ {n} - \boldsymbol {\mu} _ {c m} ^ {(t)}) (\tilde {\mathbf {q}} _ {n} - \boldsymbol {\mu} _ {c m} ^ {(t)}) ^ {\top}, +$$ + +where $N_{cm}^{(t)} = \sum_{(\tilde{\mathbf{q}}_n,\tilde{\mathbf{y}}_n)\in \mathcal{F}_c}\gamma_{c,n}^{(t)}(m)$ . We maintain an external memory bank containing GMM training samples, defined above with the set $\mathcal{F}$ , and the GMM parameters are robustly estimated with momentum [27]. + +Audio Conditional Prompting (ACP) Motivated by the concept of "mix-and-separate" [13, 18], we design the ACP process to improve cross-attention interaction for the dense audio feature representations. We consider an off-the-screen noise dataset $\mathcal{D}_r = \{(\mathbf{a}_j)\}_{j=1}^{|D_r|}$ . For each training iteration, we sample a clean audio sample $\mathbf{a}_i$ from the $\mathcal{D}$ and an off-the-screen noisy audio sample $\mathbf{a}_j$ from $\mathcal{D}_r$ to build the mixture of magnitude spectrogram $\mathbf{a}_p = \mathbf{a}_i + \mathbf{a}_j$ . The whole ACP process is illustrated in the first section of Fig. 2. Before delving into the audio recovery process, we first sample a set of class-conditional prompts $\mathbf{z}^k \sim \{\mathcal{N}(\boldsymbol{\mu}_k, \boldsymbol{\Sigma}_k)\}_{k \in \mathcal{K}}$ (where $\mathcal{K} = \{k | \mathbf{t}_i(k) = 1\}$ represent the indices for ground truth labels) via the GMM model from the last iteration according to the target semantic classes $\mathbf{t}_i$ that we want to recover. The major process of the audio recovery process shares a similar concept with the visual part described in Sec. 3.1; the difference is that we use the joint-modal mask embeddings $\tilde{\mathbf{z}}^k$ to search the semantically similar sound source on the decoded dense audio feature map $\tilde{\mathbf{u}}_a$ and generate their corresponding mask predictions $\{\mathbf{m}_k^a | k \in \mathcal{K}\}$ . + +The learning objective of the ACP process is to encourage the consistency between the predicted masks $\mathbf{m}_k^a$ with the derived ground-truth spectrogram ratio mask $\frac{\mathbf{a}_i}{\mathbf{a}_p}$ . The AVS datasets generally do not provide the separated sound source data for each semantic class (i.e., in Fig. 2, we do not have the ground-truth audio data of Male, Female and Guitar), hence we adopted a summation operation over the predicted masks to integrate all mixed sound sources. Finally, a Sigmoid activation function $\sigma(\cdot)$ is applied to constrain the prediction into a binary mask. We denoted the mean squared error (MSE) loss of the ACP process as follows: + +$$ +\ell_ {\mathrm {A C P}} = \left\| \sigma \left(\sum_ {k \in \mathcal {K}} \mathbf {m} _ {k} ^ {a}\right) - \frac {\mathbf {a} _ {i}}{\mathbf {a} _ {p}} \right\| _ {2}. \tag {5} +$$ + +Visual Conditional Prompting (VCP) The VCP module is processed simultaneously with ACP. The design of VCP aims to bypass bipartite matching via the generated class-conditional prompts $\mathbf{z}^k$ to ease the model training as it can provide a more stable learning target for each query feature to mitigate the instability brought by the bipartite matching with the class-agnostic queries. Given a set of class-conditional prompts derived from the ground-truth image labels, our training objective of VCP is to correctly segment the corresponding image regions as well as successfully classify these prompts after the consecutive transformer decoder layers. The loss function is similar to $\ell_{\text{agn}}$ in Sec. 3.1, where we replace $\mathbf{q}$ with $\mathbf{z}^k$ , denoted as $\ell_{\text{VCP}} = \ell_{ce} + \ell_{mask}$ (6). + +Prompting-based Contrastive Learning (PCL) The ability to learn discriminative feature representation is critical for the audio-visual system. One limitation of the previous audio-visual contrastive learning methods [7,36] is that + +they can only leverage the global audio representation due to the lack of class conditional prompts for class-specific feature disentanglement. By taking advantage of class-conditional distribution modelling, we can overcome this limitation by utilising the predicted spectrogram saliency mask $\mathbf{s}_k^a = \sigma (\mathbf{m}_k^a)$ and its associated class label $k$ of each sound source, denoted as $\{\mathbf{s}_k^a |\mathbf{s}_k^a\in \mathcal{M}\subseteq [0,1]^T\times F,k\in \mathcal{K}\}$ . To disentangle the class-specific representations for the audio feature map, we iteratively apply the masked average pooling (MAP) [47, 58] to $\mathbf{u}_a$ based on $|\mathcal{K}|$ saliency masks $\mathbf{s}_k^a$ for the feature extraction process. We first threshold $\mathbf{s}_k^a$ and convert it to a binary mask via $\tilde{\mathbf{s}}_k^a = \mathbb{1}(\mathbf{s}_k^a >\bar{\mathbf{s}}_k^a)$ , where $\bar{\mathbf{s}}_k^a$ represent the mean value of the saliency mask. Then, for the $k$ -th category, we extract its region-level mean feature $\mathbf{f}(k)\in \mathbb{R}^{D_q}$ by MAP defined as follows: + +$$ +\mathbf {f} (k) = \frac {\sum_ {\psi \in \Psi} \tilde {\mathbf {s}} _ {k} ^ {a} \mathbf {u} _ {a} (\psi)}{\sum_ {\psi} \tilde {\mathbf {s}} _ {k} ^ {a}}, \tag {7} +$$ + +where $\varPsi$ is the lattice of size $T\times F$ . We combine the $\mathbf{f}(k)$ with the ground-truth class label $k$ to form the anchor set $\mathcal{E}_{\mathrm{anch}} = \{(\mathbf{f}(k),k)|\mathbf{f}(k)\in \mathcal{U},k\in \mathcal{K}\}$ , with $\mathcal{U}$ denoting the audio and visual feature spaces defined in Sec. 3.1. Since the pixel-level label $\mathbf{y}$ is available, we can directly form the visual contrastive set $\mathcal{E}_{\mathrm{cont}} = \{(\mathbf{u}_v(\omega),\mathbf{y}_v(\omega))|\mathbf{u}_v(\omega)\in \mathcal{U},\mathbf{y}(\omega)\in \mathcal{V}\}$ . Hence, we can derive the positive and negative sets as: + +$$ +\begin{array}{l} \mathcal {P} (\mathbf {u} _ {v} (\omega)) = \left\{\mathbf {u} _ {v} (\omega) \mid \mathbf {u} _ {v} (\omega) \in \mathcal {U}, \mathbf {y} (\omega , k) = 1 \right\}, \\ \mathcal {U} (v, v) = \left\{v: (v) \mid v: (v) \in \mathcal {U}, (v - 1) = 0 \right\}. \end{array} \tag {8} +$$ + +$$ +\mathcal {N} (\mathbf {u} _ {v} (\omega)) = \left\{\mathbf {u} _ {v} (\omega) | \mathbf {u} _ {v} (\omega) \in \mathcal {U}, \mathbf {y} (\omega , k) = 0 \right\}. +$$ + +Adopting the supervised InfoNCE [21] as the objective function to pull the anchor $\mathbf{f} \in \mathcal{E}_{\mathrm{anch}}$ and respective positive visual features closer while repelling anchors and their negative visual features, we define the following loss: + +$$ +\ell_ {\mathrm {P C L}} (\mathbf {f}) = \frac {1}{| \mathcal {P} \left(\mathbf {u} _ {v} (\omega)\right) |} \sum_ {\mathbf {f} _ {p} \in \mathcal {P} \left(\mathbf {u} _ {v} (\omega)\right)} - \log \frac {\exp \left(\mathbf {f} \cdot \mathbf {f} _ {p} / \tau\right)}{\exp \left(\mathbf {f} \cdot \mathbf {f} _ {p} / \tau\right) + \sum_ {\mathbf {f} _ {n} \in \mathcal {N} \left(\mathbf {u} _ {v} (\omega)\right)} \exp \left(\mathbf {f} \cdot \mathbf {f} _ {n} / \tau\right)}, \tag {9} +$$ + +where $\mathbf{f}$ is an anchor feature, and $\tau$ is the temperature hyperparameter. The combination of all sub-objective $\ell_{\mathrm{ACP}}$ , $\ell_{\mathrm{VCP}}$ and $\ell_{\mathrm{PCL}}$ form the CPM loss $\ell_{\mathrm{CPM}} = \ell_{\mathrm{ACP}} + \ell_{\mathrm{VCP}} + \ell_{\mathrm{PCL}}$ . + +1. Overall training The overall training objective is $\ell = \ell_{\text{agn}} + \lambda \ell_{\text{CPM}}$ , where $\lambda$ is the weight coefficient. + +# 4 Experiments + +# 4.1 Implementation Details + +Evaluation Protocols We utilize standard evaluation protocols from AVSBench datasets [56, 57] for single-source (SS) and multi-source (MS) scenarios with binary labels, as well as for AVSBench-Semantics with multi-class labels. Image sizes are standardized to $224 \times 224$ for fair comparison. Additionally, we + +Table 1: Quantitative (mIoU, $F_{\beta}$ ) audio-visual segmentation results (\%) for the AVS-Bench test sets [56, 57] (resized to $224 \times 224$ ) with ResNet50 [17] backbone. Best results in bold, $2^{nd}$ best underlined. Improvements against the $2^{nd}$ best are in the last row. + +
Eval. MethodsMethodAVSBench-Obj.(SS)AVSBench-Obj.(MS)AVSBench-S
mIoU ↑Fβ ↑mIoU ↑Fβ ↑mIoU ↑Fβ ↑
Per-image [56,57]TPAVI [57]72.7984.8047.8857.8020.1825.20
AVSBG [16]74.1385.4044.9556.80--
ECMVAE [37]76.3386.5048.6960.70--
DiffusionAVS [36]75.8086.9049.7762.10--
CATR [24]74.8086.6052.8065.30--
AuTR [31]75.0085.2049.4061.20--
AQFormer [19]77.0086.4055.7066.90--
AVSegFormer [14]76.5484.8049.5362.8024.9329.30
AVSC [29]77.0285.2449.5861.51--
BAVS [30]77.9685.2950.2362.3724.6829.63
AVSAC [5]76.9086.9553.9565.8125.4329.71
QSD [26]77.6086.0059.6063.5046.60-
COMBO [54]81.7090.1054.5066.6033.3037.30
Per-dataset [12]AVSegFormer* [14]79.8091.4153.2864.7641.6455.05
CAVP [7]85.7792.8662.3973.6244.7057.76
CPM87.6493.5365.2275.2248.3964.01
+ +adopt the original image resolution for testing, following [7], to demonstrate optimal model performance. Evaluation is extended to the Visual Post-production (VPO) benchmark [7] for challenging cases. We employ mean Intersection over Union (mIoU) [11] and $F_{\beta}$ score with $\beta^2 = 0.3$ [38, 57] to assess segmentation quality, precision, and recall performance at pixel level. Models are stratified into CNN-based pre-pixel classification and transformer-based mask classification to demonstrate architectural capabilities. Official training splits are used for AVSBench datasets [56, 57] and VPO [7], with results reported on the respective testing sets. Training is performed on the entire AVSBench-Semantics dataset, and testing is conducted on subsets as well as the entire testing set to demonstrate partitioned model performance. For further details on training and inference, please refer to the Supplementary Material. + +Results We collected the experimental results from the existing benchmark [7] and updated it with recent works [5,26,54]. We modified the AVSegFormer [14] with Mask2former [8] and denoted it as AVSegFormer* in the tables to encourage a fair comparison with our method. Please note that Tab. 1 includes two evaluation protocols. We employ standard semantic segmentation protocols to compute both mIoU and $F_{\beta}$ same as PascalVOC, initially mentioned in [57]. + +# 4.2 Performance on Low-resolution AVSBench Videos + +We adopt established methodologies [7, 30, 31, 37] for conducting performance evaluations on down-sampled image benchmarks, such as AVSBench-Objects [57], which includes single-source (SS) and multi-source (MS) splits with binary annotations, as well as the AVSBench-Semantics dataset [56], which contains multiclass annotations. We compare the performance of state-of-the-art (SOTA) methods with our CPM in Tab. 1 using mIoU and $F_{\beta}$ . The results demonstrate that our model surpasses the second-best CAVP [7] in terms of mIoU by $1.87\%$ on AVSBench-Object (SS) [57], $2.83\%$ on AVSBench-Object (MS) [57] and $1.79\%$ on AVSBench-Semantics [57] using the ResNet-50 [17] backbone. + +Table 2: Quantitative (mIoU, $F_{\beta}$ ) audio-visual segmentation results (in %) for the AVSBench-Semantic (AVSS) test sets [56] (original resolution) with ResNet50 [17] backbone. Best results in bold, $2^{nd}$ best underlined. Improvements against the $2^{nd}$ best are in the last row. + +
D-ResNet50 [17]MethodAVSS (SS)AVSS (MS)AVSS
mIoU ↑Fβ ↑mIoU ↑Fβ ↑mIoU ↑Fβ ↑
Per-pixel ClassificationTPAVI [57]42.1061.4626.3340.9943.3959.24
CAVP [7]56.9169.1538.6152.9250.7564.57
TransformerAVSegFormer [14]46.2559.7627.2141.3841.4856.21
AVSegFormer* [14]50.5263.7531.4042.8145.8059.16
CPM61.7172.9443.1156.2857.2570.54
ImprovementCPM+4.80+3.79+4.50+3.36+6.50+5.97
+ +Table 3: Quantitative (mIoU, $F_{\beta}$ ) audio-visual segmentation results (in %) for the VPO test sets (original resolution) with ResNet50 [17] backbone. Best results in bold, $2^{nd}$ best underlined. Improvements against the $2^{nd}$ best are in the last row. + +
D-ResNet50 [17]MethodVPO (SS)VPO (MS)VPO (MSMI)
mIoU ↑Fβ ↑mIoU ↑Fβ ↑mIoU ↑Fβ ↑
Per-pixel ClassificationTPAVI [57]52.7569.5454.3071.9551.7368.85
CAVP62.3178.4664.3178.9260.3675.60
TransformerAVSegFormer [14]57.5573.0358.3374.2854.2270.39
AVSegFormer* [14]60.5174.8162.9177.3356.2472.67
CPM67.0979.8865.9179.9060.5575.58
ImprovementsCPM+4.78+1.42+1.60+0.98+0.19-0.02
+ +# 4.3 Performance on Original AVSBench Videos + +The benchmark mentioned above adopted low-resolution benchmarks using significantly resized input images (from 720p to $224 \times 224$ ). While this simplifies model training, disregarding the original image aspect ratio can lead to a degradation in model performance, which is not advisable for segmentation tasks. Therefore, we follow the evaluation method outlined in [7] for training and testing on the raw AVSBench videos. This includes using random resized crops for training and performing frame-by-frame evaluation during testing. The initial published version of AVSBench [57] only offers single-source and multisource partitioning. However, this partitioning was not conducted in the later version [56], resulting in some cases being overlooked. We re-organised the previous SS and MS with the newly added video data, resulting in 1278 single-source videos (AVSS-SS) and 276 multi-source videos (AVSS-MS). We compare the results with mIoU and $F_{\beta}$ to demonstrate the comprehensive model performance in Tab. 2. Our method shows a significant mIoU improvement of $4.80\%$ on AVSS-SS, $4.50\%$ on AVSS-MS and $+6.50\%$ on the entire AVSS dataset. To further demonstrate the effectiveness of our method, we show a visualisation of 6-second video clip in Fig. 5 that displays a qualitative comparison between TPAVI, AVSegFormer, CAVP and our CPM. Our method can successfully approximate the ground truth segmentation of the target sound source within a group of other semantic objects. + +# 4.4 Performance on VPO Dataset + +We also compare the model performance on the VPO benchmark [7], equipped with synthesized stereo audios. We adopted ResNet-50 [17] backbone for all + +Table 4: Ablation study of the model components on AVSBench-Semantics [31]. + +
MethodMetrics
BaselineCCDMVCPACPPCLmIoU ↑Fβ ↑
✓53.0465.29
✓✓54.1266.65
✓✓✓55.7968.17
✓✓✓55.0768.03
✓✓✓✓56.4169.07
✓✓✓✓✓57.2570.54
+ +![](images/26a18b65d97af0e551be591b6aa13fb28b79e7304bbee5c7e72fc29a7292de34.jpg) +(a) Visual comparison. + +![](images/6846e96eea8630f3724ef00e3f1716a43504c81c0cd83464be8a5e18e380199e.jpg) +(b) Coverage score comparison. +Fig. 3: Qualitative (3a) and quantitative (3b) comparisons between model components in a multi-source scenario (i.e., male singing, female singing and guitar) + +three subsets, and the results are shown in Tab. 3. To facilitate stereo audio encoding, we adopt the approach outlined in [7] by adjusting the number of input channels in the first layer to 2. Our approach surpasses the SOTA method CAVP [7] by $4.78\%$ and $1.60\%$ in terms of mIoU on VPO (SS) and VPO (MS), respectively. However, we observe a marginal improvement of $0.19\%$ on the VPO (MSMI) setup compared with the SS and MS subsets. The possible explanation for this phenomenon may stem from the overly simplistic integration of stereo audio within the transformer architecture, as the foundational AVS transformer architectures [14, 24, 26, 30] did not account for a distinct positional encoding scheme tailored for stereo data. We will explore the design of a dedicated stereo AVS transformer architecture as part of our future work. + +# 4.5 Ablation Study + +Ablation of Key Components We first perform the key components analysis of CPM on AVSBench-Semantics [56] in Tab. 4. The baseline is AVSegformer* [14] in the 1st row. We replace the Softmax classifier in baseline with CCDM defined in (1), and observe an mIoU improvement of $+1.08\%$ . By integrating the model with VCP (3rd row), using training loss in (6), and ACP (4th row), using loss in (5), separately, we achieve an improvement of $+1.26\%$ and $+1.39\%$ , respectively. Subsequently, when we apply both ACP and VCP methods (5th row) the mIoU performance improves $+2.04\%$ compared to the 2nd row. The final row displays the complete CPM method, incorporating the dense contrastive learning approach we introduced, which leads to an additional mIoU improvement of $+1.09\%$ . + +Ablation of Bipartite Matching Stability To study the stability of the bipartite matching process, we design an entropy-based stability score STS to quantify such performance on AVSBench-Semantics [56]. Intuitively, the STS measures the average assignment consistency after bipartite matching across all the classes. During the training process, we collect the assigned label $\tilde{\mathbf{y}}_i \in$ + +![](images/c8f2d72823cf817151f3c48a18aca44174943092ad494ff965295b1b42bcdf80.jpg) +Fig.4: Matching stability (STS $\downarrow$ ) comparison on AVSBench-Semantics [31] + +Table 5: Ablation of GMM modelling on AVSBench-Semantics [31]. + +
MethodmIoU ↑Fβ ↓
Point Rep.55.9668.58
Dist Rep.57.2570.54
MmIoUFβ
155.8368.68
357.2570.54
556.3169.04
756.0568.59
+ +$\{0, \dots, C\}^N$ to the class-agnostic query $\tilde{\mathbf{q}}_n$ produced by the Hungarian algorithm for each training image, resulting in $\mathbf{R} = [\tilde{\mathbf{y}}_0 \oplus \tilde{\mathbf{y}}_1 \oplus, \dots, \oplus \tilde{\mathbf{y}}_D]$ , where $\oplus$ represents the concatenation operator and $\mathbf{R}$ has size $D \times N$ . We then use the indicator function to extract the assignment information over all semantic classes. We denote the resulting set as $\{\mathbf{s}_0, \mathbf{s}_1, \dots, \mathbf{s}_c\}$ , where $\mathbf{s}_c(n) = \sum_{d=1}^{D} \mathbb{1}(\mathbf{R}(d,n) = c)$ for $n \in \{1, \dots, N\}$ , which forms $\mathbf{s}_c \in [0,1]^N$ after normalising it to be a probability distribution. Finally, we calculate the stability score $\mathsf{STS} = \frac{1}{C} \sum_{c=1}^{C} H(\mathbf{s}_c)$ , where $H(\cdot)$ computes the entropy. We compare our CPM with AVSegformer* over 100 training epochs as illustrated in Fig. 4. Our findings reveal that our CPM consistently improves the stability score (showing lower average entropy) throughout the training process in comparison to AVSegformer*, thereby validating the efficacy of our approach in enhancing bipartite matching stability. Please refer to the Supplementary Material for the Pseudo-code of STS. + +Ablation of Audio Contribution While experimenting with the baseline method mentioned above, we empirically observed that the classification after the audio transformer decoder (i.e., $f^{\mathrm{TD - A}}$ in Fig. 2) has low accuracy, with noisy predictions. We also found that these noisy predictions are progressively refined by the vision transformer decoder layer based solely on visual clues. Such an observation illustrates that in some hard cases, we may fail to extract useful semantic information from audio, leading to the model being overly reliant on the visual content, which may cause erroneous testing predictions, as shown in Fig. 3a-(i), (ii), (iii). To monitor the amount of valid information retained by the query following interaction with the audio branch, we introduce a classification coverage score. This score is computed by comparing the predictions generated after passing through the audio transformer decoder with the ground-truth class label, using the Jaccard index. We show qualitative results in Fig. 3a for the designated scenario and quantitative results in Fig. 3b for the AVSBench-Semantics testing set. The comparison starts with AVSegFormer as the baseline method with progressive addition of our proposed modules, similar to Tab. 4. Notice that each CPM sub-module contributes to refining key semantic information from the audio modality, as evidenced by the segmentation mask predictions and the improvement in Jaccard indexes. This shows the effectiveness of CPM. + +![](images/215ea39162a9a98b1b8b6cd2fd660eb079218ae58c13dd8d6b376358c38c42bf.jpg) +Fig. 5: Qualitative audio-visual segmentation results on AVSBench-Semantics [56] by TPAVI [57], AVSegFormer [14], CAVP [7] and our CPM, which can be compared with the ground truth (GT) Ambulance of the first row. + +Ablation of Distribution Representation Table 5 provides a study of the GMM data distribution representation on AVSBench-Semantics [56]. The first row (Point Rep.) replaces the GMM module with a linear SoftMax for the classification output, introducing a learnable $C \times D_q$ feature map (i.e., $D_q$ -dimensional feature embeddings for $C$ semantic classes for the audio and visual prompting task). The second row (Dist Rep.) shows our approach with the CCDM module, showing that the utilisation of the distribution modelling leads to considerable improvements of $1.29\%$ and $1.96\%$ on mIoU and $F_{\beta}$ respectively. The next rows of Tab. 5 show a hyper-parameter analysis on the number of GMM components (from $M = 1$ to $M = 7$ components). Results reveal that 3 GMM components enable the best fitting for the data distribution, leading to the top performance with a maximum improvement of $1.42\%$ and $1.95\%$ on mIoU and $F_{\beta}$ . + +# 5 Discussion and Conclusion + +We introduced CPM, a new audio-visual training method designed for the Maskformer based framework to enhance bipartite matching stability and improve the efficacy of cross-modal attention for audio-visual segmentation. We proposed a class-conditional prompting learning strategy that combines class-agnostic queries with class-conditional queries, sampled from our iteratively updated generative model. The generated class-conditional queries are utilised to probe both the magnitude spectrogram and the image feature map, aiming to remove off-the-screen noise and bypass bipartite matching to produce a more stable learning process. Lastly, we extend the class-conditional queries to a new prompting-based audio-visual contrastive learning to explicitly constrain the cross-modal representations. SOTA results on AVS benchmarks suggest that CPM can be a valuable resource for future AVS research. + +Limitations and future work. We acknowledge that the current adaptation of stereo audio into the transformer-based method has limitations, as it encodes positional and semantic information jointly within the transformer block. This contrasts with the typical approach of separately encoding these two types of information in the transformer-based framework. Our future work will concentrate on optimizing this framework through the integration of spatial reasoning. + +# References + +1. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020) +2. Chen, H., Xie, W., Afouras, T., Nagrani, A., Vedaldi, A., Zisserman, A.: Localizing visual sounds the hard way. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16867-16876 (2021) +3. Chen, J., Lu, J., Zhu, X., Zhang, L.: Generative semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7111-7120 (2023) +4. Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 801-818 (2018) +5. Chen, T., Tan, Z., Gong, T., Chu, Q., Wu, Y., Liu, B., Lu, L., Ye, J., Yu, N.: Bootstrapping audio-visual segmentation by strengthening audio cues. arXiv preprint arXiv:2402.02327 (2024) +6. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020) +7. Chen, Y., Liu, Y., Wang, H., Liu, F., Wang, C., Carneiro, G.: A closer look at audio-visual semantic segmentation. arXiv e-prints pp. arXiv-2304 (2023) +8. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290–1299 (2022) +9. Cheng, B., Schwing, A., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems 34, 17864-17875 (2021) +0. Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society: series B (methodological) 39(1), 1-22 (1977) +1. Everingham, M., Eslami, S.A., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes challenge: A retrospective. International journal of computer vision 111, 98-136 (2015) +2. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. International journal of computer vision 88, 303-338 (2010) +3. Gao, R., Grauman, K.: Co-separating sounds of visual objects. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3879-3888 (2019) +4. Gao, S., Chen, Z., Chen, G., Wang, W., Lu, T.: Avsegformer: Audio-visual segmentation with transformer. arXiv preprint arXiv:2307.01146 (2023) +5. Gray, R.: Vector quantization. IEEE Assp Magazine 1(2), 4-29 (1984) +6. Hao, D., Mao, Y., He, B., Han, X., Dai, Y., Zhong, Y.: Improving audio-visual segmentation with bidirectional generation. arXiv preprint arXiv:2308.08288 (2023) +7. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) + +18. Hu, X., Chen, Z., Owens, A.: Mix and localize: Localizing sound sources in mixtures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10483-10492 (2022) +19. Huang, S., Li, H., Wang, Y., Zhu, H., Dai, J., Han, J., Rong, W., Liu, S.: Discovering sounding objects by audio queries for audio visual segmentation. arXiv preprint arXiv:2309.09501 (2023) +20. Jaegle, A., Borgeaud, S., Alayrac, J.B., Doersch, C., Ionescu, C., Ding, D., Koppula, S., Zoran, D., Brock, A., Shelhamer, E., et al.: Perceiver io: A general architecture for structured inputs & outputs. arXiv preprint arXiv:2107.14795 (2021) +21. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in neural information processing systems 33, 18661-18673 (2020) +22. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) +23. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13619-13627 (2022) +24. Li, K., Yang, Z., Chen, L., Yang, Y., Xun, J.: Catr: Combinatorial-dependence audio-queried transformer for audio-visual video segmentation. arXiv preprint arXiv:2309.09709 (2023) +25. Li, T., Cao, P., Yuan, Y., Fan, L., Yang, Y., Feris, R.S., Indyk, P., Katabi, D.: Targeted supervised contrastive learning for long-tailed recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6918-6928 (2022) +26. Li, X., Wang, J., Xu, X., Peng, X., Singh, R., Lu, Y., Raj, B.: Towards robust audiovisual segmentation in complex environments with quantization-based semantic decomposition. arXiv preprint arXiv:2310.00132 (2023) +27. Liang, C., Wang, W., Miao, J., Yang, Y.: Gmmseg: Gaussian mixture based generative semantic segmentation models. Advances in Neural Information Processing Systems 35, 31360-31375 (2022) +28. Liu, C., Ding, H., Jiang, X.: Gres: Generalized referring expression segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23592-23601 (2023) +29. Liu, C., Li, P., Qi, X., Zhang, H., Li, L., Wang, D., Yu, X.: Audio-visual segmentation by exploring cross-modal mutual semantics (2023) +30. Liu, C., Li, P., Zhang, H., Li, L., Huang, Z., Wang, D., Yu, X.: Bavs: Bootstrapping audio-visual segmentation by integrating foundation knowledge. arXiv preprint arXiv:2308.10175 (2023) +31. Liu, J., Ju, C., Ma, C., Wang, Y., Wang, Y., Zhang, Y.: Audio-aware query-enhanced transformer for audio-visual segmentation. arXiv preprint arXiv:2307.13236 (2023) +32. Liu, J., Wang, Y., Ju, C., Zhang, Y., Xie, W.: Annotation-free audio-visual segmentation. arXiv preprint arXiv:2305.11019 (2023) +33. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329 (2022) +34. Liu, X., Carrington, P., Chen, X., Pavel, A.: What makes videos accessible to blind and visually impaired people? In: Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. pp. 1-14 (2021) + +35. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015) +36. Mao, Y., Zhang, J., Xiang, M., Lv, Y., Zhong, Y., Dai, Y.: Contrastive conditional latent diffusion for audio-visual segmentation. arXiv preprint arXiv:2307.16579 (2023) +37. Mao, Y., Zhang, J., Xiang, M., Zhong, Y., Dai, Y.: Multimodal variational autoencoder based audio-visual segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 954–965 (2023) +38. Martin, D.R., Fowlkes, C.C., Malik, J.: Learning to detect natural image boundaries using local brightness, color, and texture cues. IEEE transactions on pattern analysis and machine intelligence 26(5), 530-549 (2004) +39. Meng, D., Chen, X., Fan, Z., Zeng, G., Li, H., Yuan, Y., Sun, L., Wang, J.: Conditional detr for fast training convergence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3651-3660 (2021) +40. Mo, S., Morgado, P.: A closer look at weakly-supervised audio-visual source localization. arXiv preprint arXiv:2209.09634 (2022) +41. Mo, S., Morgado, P.: Localizing visual sounds the easy way. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII. pp. 218-234. Springer (2022) +42. Mo, S., Tian, Y.: Audio-visual grouping network for sound localization from mixtures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10565–10574 (2023) +43. Mo, S., Tian, Y.: Av-sam: Segment anything model meets audio-visual localization and segmentation. arXiv preprint arXiv:2305.01836 (2023) +44. Murray, M.M., Wallace, M.T.: The neural bases of multisensory processes. Frontiers in Neuroscience (2011) +45. Reynolds, D.A., et al.: Gaussian mixture models. Encyclopedia of biometrics 741(659-663) (2009) +46. Senocak, A., Ryu, H., Kim, J., Oh, T.H., Pfister, H., Chung, J.S.: Sound source localization is all about cross-modal alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7777-7787 (2023) +47. Siam, M., Oreshkin, B.N., Jagersand, M.: Amp: Adaptive masked proxies for few-shot segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5249-5258 (2019) +48. Sun, Z., Cao, S., Yang, Y., Kitani, K.M.: Rethinking transformer-based set prediction for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3611-3620 (2021) +49. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +50. Wang, W., Zhou, T., Yu, F., Dai, J., Konukoglu, E., Van Gool, L.: Exploring cross-image pixel contrast for semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7303-7313 (2021) +51. Wang, Y., Liu, W., Li, G., Ding, J., Hu, D., Li, X.: Prompting segmentation with sound is generalizable audio-visual source localizer. arXiv preprint arXiv:2309.07929 (2023) +52. Wu, J., Jiang, Y., Sun, P., Yuan, Z., Luo, P.: Language as queries for referring video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4974-4984 (2022) + +53. Xu, P., Zhu, X., Clifton, D.A.: Multimodal learning with transformers: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) +54. Yang, Q., Nie, X., Li, T., Gao, P., Guo, Y., Zhen, C., Yan, P., Xiang, S.: Cooperation does matter: Exploring multi-order bilateral relations for audio-visual segmentation. arXiv preprint arXiv:2312.06462 (2023) +55. Zhang, H., Li, F., Xu, H., Huang, S., Liu, S., Ni, L.M., Zhang, L.: Mp-former: Mask-piloted transformer for image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18074-18083 (2023) +56. Zhou, J., Shen, X., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., et al.: Audio-visual segmentation with semantics. arXiv preprint arXiv:2301.13190 (2023) +57. Zhou, J., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., Zhong, Y.: Audio-visual segmentation. In: European Conference on Computer Vision. pp. 386-403. Springer (2022) +58. Zhou, T., Zhang, M., Zhao, F., Li, J.: Regional semantic contrast and aggregation for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4299-4309 (2022) \ No newline at end of file diff --git a/2024/CPM_ Class-conditional Prompting Machine for Audio-visual Segmentation/images.zip b/2024/CPM_ Class-conditional Prompting Machine for Audio-visual Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2a65289c108c67765e2ea4f2935784d10e5a56f8 --- /dev/null +++ b/2024/CPM_ Class-conditional Prompting Machine for Audio-visual Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f7e48dced94ba68a9340e2d1a6f2f2d741deedac7d9630540094e837d75ffe2 +size 509539 diff --git a/2024/CPM_ Class-conditional Prompting Machine for Audio-visual Segmentation/layout.json b/2024/CPM_ Class-conditional Prompting Machine for Audio-visual Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b483d36578f52589cb57058d085343383f8d49f6 --- /dev/null +++ b/2024/CPM_ Class-conditional Prompting Machine for Audio-visual Segmentation/layout.json @@ -0,0 +1,10752 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "text", + "content": "CPM: Class-conditional Prompting Machine for Audio-visual Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 143, + 167, + 470, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 167, + 470, + 192 + ], + "spans": [ + { + "bbox": [ + 143, + 167, + 470, + 192 + ], + "type": "text", + "content": "Yuanhong Chen1, Chong Wang1, Yuyuan Liu1, Hu Wang2, and Gustavo Carneiro3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 140, + 201, + 472, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 201, + 472, + 213 + ], + "spans": [ + { + "bbox": [ + 140, + 201, + 472, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 201, + 472, + 213 + ], + "type": "text", + "content": " Australian Institute for Machine Learning, University of Adelaide, Australia" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 213, + 472, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 213, + 472, + 223 + ], + "spans": [ + { + "bbox": [ + 140, + 213, + 472, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 213, + 472, + 223 + ], + "type": "text", + "content": " Mohamed bin Zayed University of Artificial Intelligence, United Arab Emirates" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 223, + 452, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 223, + 452, + 246 + ], + "spans": [ + { + "bbox": [ + 160, + 223, + 452, + 246 + ], + "type": "text", + "content": "3 Centre for Vision, Speech and Signal Processing, University of Surrey yuanhong.chen@adelaide.edu.au" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 275, + 453, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 275, + 453, + 462 + ], + "spans": [ + { + "bbox": [ + 160, + 275, + 453, + 462 + ], + "type": "text", + "content": "Abstract. Audio-visual segmentation (AVS) is an emerging task that aims to accurately segment sounding objects based on audio-visual cues. The success of AVS learning systems depends on the effectiveness of cross-modal interaction. Such a requirement can be naturally fulfilled by leveraging transformer-based segmentation architecture due to its inherent ability to capture long-range dependencies and flexibility in handling different modalities. However, the inherent training issues of transformer-based methods, such as the low efficacy of cross-attention and unstable bipartite matching, can be amplified in AVS, particularly when the learned audio query does not provide a clear semantic clue. In this paper, we address these two issues with the new Class-conditional Prompting Machine (CPM). CPM improves the bipartite matching with a learning strategy combining class-agnostic queries with class-conditional queries. The efficacy of cross-modal attention is upgraded with new learning objectives for the audio, visual and joint modalities. We conduct experiments on AVS benchmarks, demonstrating that our method achieves state-of-the-art (SOTA) segmentation accuracy4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 472, + 453, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 472, + 453, + 495 + ], + "spans": [ + { + "bbox": [ + 159, + 472, + 453, + 495 + ], + "type": "text", + "content": "Keywords: Audio-visual Learning " + }, + { + "bbox": [ + 159, + 472, + 453, + 495 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 472, + 453, + 495 + ], + "type": "text", + "content": " Segmentation " + }, + { + "bbox": [ + 159, + 472, + 453, + 495 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 472, + 453, + 495 + ], + "type": "text", + "content": " Multi-modal Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 515, + 230, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 230, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 230, + 528 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 540, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 482, + 635 + ], + "type": "text", + "content": "The recognition and integration of auditory and visual data are fundamental to human cognitive processes, playing a critical role in facilitating meaningful communication [44]. Audio-visual segmentation (AVS) is an emerging cross-modal reasoning task that mimics such cognitive processes, aiming to localize visual objects based on audio-visual cues. AVS has many applications, such as the automatic localisation and identification of sounding objects to improve the accessibility of videos for the blind and visually impaired [34]. A major challenge in AVS is achieving effective cross-modal interaction between sound and visual" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 665 + ], + "type": "text", + "content": "This project is supported by the Australian Research Council (ARC) through grant FT190100525." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 160, + 115, + 451, + 258 + ], + "blocks": [ + { + "bbox": [ + 160, + 115, + 451, + 258 + ], + "lines": [ + { + "bbox": [ + 160, + 115, + 451, + 258 + ], + "spans": [ + { + "bbox": [ + 160, + 115, + 451, + 258 + ], + "type": "image", + "image_path": "6fa6c9fc6131d3e7717ec93b0fa998a611d80393b2f55380e0e67d2b8777a19d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 261, + 482, + 318 + ], + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 318 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 318 + ], + "type": "text", + "content": "Fig. 1: Comparing conventional AVS methods [7, 14] with our CPM approach, CPM inherits the class-agnostic query from transformer-based methods and integrates class-conditional prompts sampled from the learned joint-modal data distribution to achieve three objectives: 1) learn disentangled audio partitioning, 2) facilitate semantic-guided object identification, and 3) promote more explicit audio-visual contrastive learning." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 330, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 437 + ], + "type": "text", + "content": "objects [46]. Many AVS methods [7, 37, 57] generally adopt a traditional per-pixel classification framework [8, 52], utilising early fusion strategies (e.g., cross-attention fusion [49]) together with an FCN decoder [4, 35] to make predictions. Such per-pixel design has achieved good performance, but it tends to underutilise the audio data due to its lower informativeness compared with the visual data [7, 46]. Contrastive learning can mitigate this issue by matching informative audio-visual pairs [7]. However, another critical limitation of the per-pixel design is its failure to capture instance-level visual information, resulting in inconsistent segmentation predictions within or between frames in a video sequence [52]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 438, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 605 + ], + "type": "text", + "content": "These two issues have been addressed by transformer-based AVS methods designed to capture instance-level information and to rely on more effective contrastive learning [14, 24, 28, 30, 31, 52, 54]. Nevertheless, these AVS approaches still show slow convergence and relatively poor accuracy [23, 33, 55], primarily attributed to the low efficacy of the cross-attention [48] and the unstable bipartite matching [23]. Solutions for these problems are based on integrated masked-attention [8], which focuses on features around predicted segments, and anchor de-noising [23] to reconstruct the noisy ground-truth bounding box detection. However, these solutions may not work well in AVS because of the weak constraint provided by global audio features that contain a mixture of sound sources [14, 26, 30] resulting in more instability during training. It is worth noting that all methods above have the common issue of relying on class-agnostic prompts that provide little guidance to the bipartite matching process, thereby reducing training efficacy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we introduce the Class-conditional Prompting Machine (CPM), an audio-visual segmentation training approach that leverages class-conditional prompts to enhance bipartite matching stability and improve cross-modal attention efficacy. To enhance bipartite matching, we introduce a novel learning method, combining class-agnostic queries [8,14,26] with class-conditional queries," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "Y. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "content": "sampled from our iteratively updated generative model of class-specific embeddings, with the former queries matched to ground-truth labels using the Hungarian Algorithm [1], followed by individual processing of class-conditional queries in two modalities. To improve cross-modal attention efficacy, new learning objectives are proposed for both audio and visual modalities. In the audio modality (audio conditional prompting, denoted as ACP), the original spectrogram is corrupted with auditory off-the-screen noise and class-conditional queries are employed to reconstruct the original spectrogram, while in the visual modality (visual conditional prompting, denoted as VCP), noisy class-conditional queries sampled from our generative model are used to probe semantically similar content in the image space. To further upgrade the performance of cross-modal attention, we introduce a new prompting-based audio-visual contrastive learning (PCL) task, guided by class-specific queries to densely constrain the cross-modal representations. To summarise, our main contributions are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 292, + 480, + 459 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 292, + 480, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 292, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 292, + 480, + 327 + ], + "type": "text", + "content": "- A new AVS training approach to enhance bipartite matching stability and enhance the efficacy of cross-modal attention. Our core innovation lies in the development of a Class-conditional Prompting Machine (CPM)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 328, + 480, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 328, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 328, + 480, + 363 + ], + "type": "text", + "content": "- To improve the bipartite matching, we propose a new AVS learning strategy that combines class-agnostic queries with class-conditional queries, sampled from our iteratively updated generative model of class-specific embeddings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 365, + 480, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 459 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 459 + ], + "type": "text", + "content": "- The efficacy of cross-modal attention is upgraded with new learning objectives for the audio, visual and joint modalities. For audio, we present ACP that perturbs the original spectrogram and uses class-conditional queries to reconstruct the spectrogram; for visual, we introduce VCP that explores noisy class-conditional queries sampled from our generative model to probe the corresponding semantic in visual feature map. To enhance the cross-modal attention efficacy further, we propose PCL, consisting of a new prompting-based joint (audio-visual) contrastive learning task." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "text", + "content": "We firstly show the effectiveness of our CPM model through rigorous evaluation on established benchmarks such as AVSBench-Objects [57] and AVSBench-Semantics [56]. Furthermore, we extend our evaluation by including VPO synthetic benchmarks [7], aiming to enhance our comprehension of AVS methods' capacity to capture audio-visual correlations. Our findings across these benchmarks consistently demonstrate that our approach yields better classification accuracy compared to existing methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 571, + 243, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 571, + 243, + 583 + ], + "spans": [ + { + "bbox": [ + 132, + 571, + 243, + 583 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 597, + 318, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 318, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 318, + 608 + ], + "type": "text", + "content": "2.1 Transformer-based architecture" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Transformer-based methods have shown promising performance in detection [1] and segmentation [8,9] benchmarks. The fundamental concept is to leverage the object query to probe image features from the output of transformer encoders and bipartite graph matching to perform set-based box/mask prediction. It is" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 262, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 262, + 91, + 447, + 102 + ], + "type": "text", + "content": "Class-conditional Prompting Machine (CPM)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": "also evidenced that such a framework can benefit multi-modal learning due to its attention mechanism and flexibility in data modelling [53]. Despite its successful application in various domains [1, 8], such frameworks have also been reported to exhibit a poorer convergence rate compared to traditional CNN-based methods [8,23,33]. In exploring the reasons behind the poor convergence, previous studies have emphasized enhancing the interpretability of the learnable query [33, 39] (often referring to them as anchor points/boxes). Alternatively, the utilisation of denoising methods [23, 55] has also been adopted to facilitate the learning of bounding box offset [23] and improve the utilisation of the transformer decoder layers [55]. Furthermore, masked attention has proven effective in enhancing both convergence rate and model performance [8]. The successful implementation of denoising methods and masked attention mechanisms has inspired us to devise a mitigation strategy for bipartite matching that promotes improved cross-modal understanding within the cross-attention process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 293, + 331, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 331, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 331, + 306 + ], + "type": "text", + "content": "2.2 Audio-visual Segmentation (AVS)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 313, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 482, + 601 + ], + "type": "text", + "content": "Audio-visual Segmentation is a dense classification task or detection of sounding visual objects in videos, using image sequences and audio cues. Zhou et al. [57] introduced the AVSBench-Object and AVSBench-Semantics benchmarks [56], which enable the evaluations of single-source and multi-source AVS tasks for salient and multi-class object segmentation. To address concerns related to the high annotation cost and dataset diversity, Chen et al. [7] introduced a cost-effective strategy to build a relatively unbiased AVS dataset, named Visual Post-production (VPO). Mainstream AVS methods use audio as the reference query [7, 14, 19, 24, 29–32] or prompt [43, 51]. For example, some methods adopt MaskFormer [9] or segment anything model (SAM) [22] to perform image segmentation using audio queries or encoded audio prompts and cross-attention layers. These methods benefit from the attention mechanism's ability and mask-classification's features to capture long-range dependencies and enhance image segmentation ability [9], spatial-temporal reasoning [24] and task-related features [14, 24, 29, 30, 37, 57]. The adaptation of the Maskformer-based framework for AVS relies on methods to encourage the audio-visual semantic alignment [24, 30, 31, 54]. Such strategy mitigates the poor audio semantic information caused by modality imbalance [5] and learns disentangled multi-modal representations [26, 54] via vector quantization [15] and bidirectional attention mechanism. Even though the query-based transformer architecture [1] has shown great success in semantic segmentation tasks, certain weaknesses such as low efficacy of cross-attention and unstable bipartite matching process [23] are still only partially addressed. These two issues may be exacerbated in AVS due to the poor audio semantic information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 609, + 335, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 335, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 335, + 621 + ], + "type": "text", + "content": "2.3 Audio-visual Contrastive Learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Contrastive Learning has shown promising results in audio-visual learning (AVL) methods [2, 7, 18, 40, 41]. These methods bring together augmented representations from the same instance as positives while separating representation pairs" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "Y. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": "from different instances as negatives within a batch. A reported issue with current AVL contrastive learning is its reliance on self-supervision [6] to connect audio and visual representations of the same class. To overcome this issue, CAVP [7] propose a supervised contrastive learning [21, 25, 50] method that mines informative contrastive pairs from arbitrary audio-visual pairings to constrain the learning of audio-visual embeddings. Nevertheless, these previous approaches predominantly leverage global audio representations, thereby limiting the model's capacity to discern individual audio sources. Consequently, this constraint reduces the model's effectiveness in scenarios involving multiple sound sources. In our work, we employ class-specific queries to retrieve the corresponding audio representations, which facilitates a more explicit form of contrastive learning between audio and visual modalities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 277, + 202, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 202, + 290 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 202, + 290 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 300, + 482, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 387 + ], + "type": "text", + "content": "We denote a multi-class audio-visual dataset as " + }, + { + "bbox": [ + 130, + 300, + 482, + 387 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(\\mathbf{a}_i, \\mathbf{x}_i, \\mathbf{y}_i, \\mathbf{t}_i))\\}_{i=1}^{|\\mathcal{D}|}" + }, + { + "bbox": [ + 130, + 300, + 482, + 387 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 300, + 482, + 387 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i \\in \\mathcal{X} \\subset \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 300, + 482, + 387 + ], + "type": "text", + "content": " is an RGB image with resolution " + }, + { + "bbox": [ + 130, + 300, + 482, + 387 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 130, + 300, + 482, + 387 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 300, + 482, + 387 + ], + "type": "inline_equation", + "content": "\\mathbf{a} \\in \\mathcal{A} \\subset \\mathbb{R}^{T \\times F}" + }, + { + "bbox": [ + 130, + 300, + 482, + 387 + ], + "type": "text", + "content": " denotes the magnitude spectrogram of the audio data with " + }, + { + "bbox": [ + 130, + 300, + 482, + 387 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 300, + 482, + 387 + ], + "type": "text", + "content": " time and " + }, + { + "bbox": [ + 130, + 300, + 482, + 387 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 300, + 482, + 387 + ], + "type": "text", + "content": " frequency bins, " + }, + { + "bbox": [ + 130, + 300, + 482, + 387 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i \\in \\mathcal{Y} \\subset \\{0,1\\}^{H \\times W \\times C}" + }, + { + "bbox": [ + 130, + 300, + 482, + 387 + ], + "type": "text", + "content": " denotes the pixel-level ground truth for the " + }, + { + "bbox": [ + 130, + 300, + 482, + 387 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 300, + 482, + 387 + ], + "type": "text", + "content": " classes (the background class is included in these " + }, + { + "bbox": [ + 130, + 300, + 482, + 387 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 300, + 482, + 387 + ], + "type": "text", + "content": " classes), and " + }, + { + "bbox": [ + 130, + 300, + 482, + 387 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i \\in \\mathcal{Y} \\subset \\{0,1\\}^C" + }, + { + "bbox": [ + 130, + 300, + 482, + 387 + ], + "type": "text", + "content": " is a multi-label ground truth audio annotation. We use Mask2former [8] as the segmentation framework." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 393, + 375, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 393, + 375, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 393, + 375, + 407 + ], + "type": "text", + "content": "3.1 Preliminaries using Class-agnostic Queries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": "Like other Maskformer-based methods [14,30,31,52], we aim to learn the parameters " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\theta \\in \\Theta" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": " for the model " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "f_{\\theta}:\\mathcal{X}\\times \\mathcal{A}\\to [0,1]^{H\\times W\\times C}" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": ", which comprises the image and audio backbones that extract features with " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_a = f_\\gamma (\\mathbf{a})" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_v = f_\\phi (\\mathbf{x})" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": ", respectively, where " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\gamma ,\\phi \\in \\theta" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_a,\\mathbf{u}_v\\in \\mathcal{U}" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": " denoting a unified feature space. A set of learnable query features (comprising the object query feature and positional query embeddings) are defined as the joint-modal output embeddings similar to the Perceiver model [20]. We define the class-agnostic query feature as " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{q}\\in \\mathcal{Q}\\subset \\mathbb{R}^{N\\times D_q}" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": " denotes the number of class-agnostic queries, and " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "D_{q}" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": " represents the dimensionality of the feature space. As depicted in Fig. 2, given " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": ", we aim to group pixels with matched semantic information from " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_a" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_v" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": " through consecutive transformer decoder layers [8,49] and generate " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": " mask embeddings " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{q}}\\in \\mathcal{Q}" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": " and pixel embeddings " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{u}}_v\\in \\mathbb{R}^{H\\times W\\times D_q}" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": ". Then, the model independently predicts the embeddings into " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": " set of class predictions (via Softmax(·)) and mask predictions (via Sigmoid(·)) denoted as " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{pred}} = \\left\\{\\left(\\mathbf{m}_i,\\mathbf{p}_i\\right)\\right\\}_{i = 1}^N" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{m}\\in \\mathcal{M}\\subset \\{0,1\\}^{H\\times W}" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{p}\\in \\mathcal{P}\\subset \\{0,1\\}^C" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": ". We denote the ground-truth set derived from the training set " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{gt}} = \\left\\{\\left(\\mathbf{m}_i^{\\mathrm{gt}},\\mathbf{p}_i^{\\mathrm{gt}}\\right)\\right\\}_{i = 1}^{N^{\\mathrm{gt}}}" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": ". During training, we use the Hungarian algorithm to perform optimal matching between " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{pred}}" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{gt}}" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": " [1,8] to facilitate the label assignment. Since " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "|\\mathcal{E}_{\\mathrm{pred}}| > |\\mathcal{E}_{\\mathrm{gt}}|" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": ", we pad " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{gt}}" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": " with no-object class " + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 130, + 412, + 482, + 642 + ], + "type": "text", + "content": " to achieve one-to-one matching [8]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 642, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 482, + 666 + ], + "type": "text", + "content": "The loss " + }, + { + "bbox": [ + 130, + 642, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 130, + 642, + 482, + 666 + ], + "type": "text", + "content": " to train the model " + }, + { + "bbox": [ + 130, + 642, + 482, + 666 + ], + "type": "inline_equation", + "content": "f_{\\theta}(.)" + }, + { + "bbox": [ + 130, + 642, + 482, + 666 + ], + "type": "text", + "content": " includes a cross-entropy query classification loss " + }, + { + "bbox": [ + 130, + 642, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\ell_{ce}" + }, + { + "bbox": [ + 130, + 642, + 482, + 666 + ], + "type": "text", + "content": ", and a binary mask loss " + }, + { + "bbox": [ + 130, + 642, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\ell_{mask} = \\ell_{focal} + \\ell_{dice}" + }, + { + "bbox": [ + 130, + 642, + 482, + 666 + ], + "type": "text", + "content": ", which combines a focal loss" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 262, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 262, + 90, + 447, + 103 + ], + "type": "text", + "content": "Class-conditional Prompting Machine (CPM)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 160, + 115, + 452, + 277 + ], + "blocks": [ + { + "bbox": [ + 160, + 115, + 452, + 277 + ], + "lines": [ + { + "bbox": [ + 160, + 115, + 452, + 277 + ], + "spans": [ + { + "bbox": [ + 160, + 115, + 452, + 277 + ], + "type": "image", + "image_path": "81a0499895e82667d6bdd7d30964ddf0da46833ba0f91263495d8710a773d9dc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 286, + 482, + 429 + ], + "lines": [ + { + "bbox": [ + 130, + 286, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 482, + 429 + ], + "type": "text", + "content": "Fig. 2: Illustration of our CPM method. Starting with a scene with a mixture of sound sources including Male, Female and Guitar, the training alternates between the use of learnable class-agnostic queries, and queries sampled from a list of class-specific query features from the GMM, denoted as " + }, + { + "bbox": [ + 130, + 286, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{male}}" + }, + { + "bbox": [ + 130, + 286, + 482, + 429 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 286, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{female}}" + }, + { + "bbox": [ + 130, + 286, + 482, + 429 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 286, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{guitar}}" + }, + { + "bbox": [ + 130, + 286, + 482, + 429 + ], + "type": "text", + "content": ". The overall training objective is composed of three learning tasks: 1) in audio conditional prompting (ACP), we aim to use " + }, + { + "bbox": [ + 130, + 286, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{male}}" + }, + { + "bbox": [ + 130, + 286, + 482, + 429 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 286, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{female}}" + }, + { + "bbox": [ + 130, + 286, + 482, + 429 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 286, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{guitar}}" + }, + { + "bbox": [ + 130, + 286, + 482, + 429 + ], + "type": "text", + "content": " to recover the original magnitude spectrogram " + }, + { + "bbox": [ + 130, + 286, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i" + }, + { + "bbox": [ + 130, + 286, + 482, + 429 + ], + "type": "text", + "content": " from the noise spectrogram (i.e., " + }, + { + "bbox": [ + 130, + 286, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i + \\mathbf{a}_j" + }, + { + "bbox": [ + 130, + 286, + 482, + 429 + ], + "type": "text", + "content": ") that is corrupted by another Dog audio signal " + }, + { + "bbox": [ + 130, + 286, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_j" + }, + { + "bbox": [ + 130, + 286, + 482, + 429 + ], + "type": "text", + "content": "; 2) a visual conditional prompting (VCP) that aim to probe the corresponding pixels w.r.t to the class-specific query features; and 3) a contrastive learning task that target to densely constrain the audio and visual representations. For training, both the CPM Workflow, indicated by the orange arrow, and the Class-agnostic Workflow, marked by the black arrow, are utilized. However, only the Class-agnostic Workflow is used for inference." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 443, + 482, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 443, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 443, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\ell_{focal}" + }, + { + "bbox": [ + 130, + 443, + 482, + 504 + ], + "type": "text", + "content": " and a dice loss " + }, + { + "bbox": [ + 130, + 443, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\ell_{dice}" + }, + { + "bbox": [ + 130, + 443, + 482, + 504 + ], + "type": "text", + "content": " [8,14,30]. The overall training loss for the class-agnostic query feature is defined as " + }, + { + "bbox": [ + 130, + 443, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\ell_{agn} = \\ell_{ce} + \\ell_{mask}" + }, + { + "bbox": [ + 130, + 443, + 482, + 504 + ], + "type": "text", + "content": ". During testing, " + }, + { + "bbox": [ + 130, + 443, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 443, + 482, + 504 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 443, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 130, + 443, + 482, + 504 + ], + "type": "text", + "content": " are merged via the multiplication " + }, + { + "bbox": [ + 130, + 443, + 482, + 504 + ], + "type": "inline_equation", + "content": "\\arg \\max_{i:c_i\\neq \\emptyset}\\mathbf{p}(c_i)\\cdot \\mathbf{m}_i" + }, + { + "bbox": [ + 130, + 443, + 482, + 504 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 443, + 482, + 504 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 443, + 482, + 504 + ], + "type": "text", + "content": " is the class label with maximum likelihood " + }, + { + "bbox": [ + 130, + 443, + 482, + 504 + ], + "type": "inline_equation", + "content": "c_{i} = \\arg \\max_{c\\in \\{1,\\dots,C,\\emptyset \\}}\\mathbf{p}_{i}(c)" + }, + { + "bbox": [ + 130, + 443, + 482, + 504 + ], + "type": "text", + "content": " for each probability-mask pair indexed by " + }, + { + "bbox": [ + 130, + 443, + 482, + 504 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 443, + 482, + 504 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 513, + 389, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 513, + 389, + 525 + ], + "spans": [ + { + "bbox": [ + 131, + 513, + 389, + 525 + ], + "type": "text", + "content": "3.2 Class-conditional Prompting Machine (CPM)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "The proposed CPM training builds upon the class-agnostic training method from Sec. 3.1 by introducing a new training stage based on class-conditional prompting. This allows the probing of both the magnitude spectrogram and the image feature map, aiming to mitigate unstable training issues and improve cross-attention efficacy. In some cases, the prompts can be manually crafted (i.e., text label, bounding box) based on domain knowledge or specific task requirements [22]. Prompts can also be automatically learned to form a fixed set of prototypical embeddings for each class [42]. However, it is impossible to manually define class conditional prompts in high-dimensional latent spaces, and using a limited-size set of learned prompts may fail to capture the comprehensive distribution of class-specific prompts. Hence, it is desirable that these" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "Y. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": "class-specific prompts can be sampled from a generative model [3,27]. that comprehensively represents the respective class. We adopt the Gaussian Mixture Models (GMMs) as such generative model [45], which improves the intra-class variability and increases robustness to class imbalances when compared to the alternative approaches mentioned before. Before delving into the methodology of the CPM, we first introduce the generation process of the class-conditional query features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 205, + 480, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 480, + 255 + ], + "type": "text", + "content": "Class Conditional Distribution Modelling (CCDM) Instead of constructing a posterior " + }, + { + "bbox": [ + 130, + 205, + 480, + 255 + ], + "type": "inline_equation", + "content": "p(c|\\tilde{\\mathbf{q}})" + }, + { + "bbox": [ + 130, + 205, + 480, + 255 + ], + "type": "text", + "content": " with Softmax classifier for the mask embeddings " + }, + { + "bbox": [ + 130, + 205, + 480, + 255 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{q}}" + }, + { + "bbox": [ + 130, + 205, + 480, + 255 + ], + "type": "text", + "content": ", the generative classifier employs the Bayes rule for label prediction, estimating the class-conditional distribution " + }, + { + "bbox": [ + 130, + 205, + 480, + 255 + ], + "type": "inline_equation", + "content": "p(\\tilde{\\mathbf{q}}|c)" + }, + { + "bbox": [ + 130, + 205, + 480, + 255 + ], + "type": "text", + "content": " alongside the class prior " + }, + { + "bbox": [ + 130, + 205, + 480, + 255 + ], + "type": "inline_equation", + "content": "p(c)" + }, + { + "bbox": [ + 130, + 205, + 480, + 255 + ], + "type": "text", + "content": " [27], with:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 250, + 261, + 480, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 261, + 480, + 287 + ], + "spans": [ + { + "bbox": [ + 250, + 261, + 480, + 287 + ], + "type": "interline_equation", + "content": "p (c | \\tilde {\\mathbf {q}}) = \\frac {p (\\tilde {\\mathbf {q}} | c) p (c)}{\\sum_ {c ^ {\\prime}} p (\\tilde {\\mathbf {q}} | c ^ {\\prime}) p \\left(c ^ {\\prime}\\right)}, \\tag {1}", + "image_path": "37cfd85a75d188562154236386e7aee74a86023eeddbe6111e90d357b7181fb3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "text", + "content": "where the class probability prior " + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "inline_equation", + "content": "p(c)" + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "text", + "content": " is uniform. To enable GMM modelling in Maskformer architecture [8, 9], we replace the " + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = \\operatorname{Softmax}(\\tilde{\\mathbf{q}})" + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "text", + "content": " activation function with " + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{GMM}}(\\tilde{\\mathbf{q}}) = p(c|\\tilde{\\mathbf{q}})" + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "text", + "content": " defined in (1) which maps the input mask embeddings " + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{q}}" + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "text", + "content": " to probability data density over " + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "text", + "content": " classes. Assuming that after the Hungarian matching between the predictions set " + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "inline_equation", + "content": "\\{(\\mathbf{m}_i,\\mathbf{p}_i)\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "text", + "content": ", we get a new one-to-one matched ground-truth set " + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{y}}" + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "text", + "content": " w.r.t each prediction by padding the " + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "text", + "content": " with no-object class " + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "text", + "content": ". Subsequently, we can form the GMM training dataset " + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = \\{(\\tilde{\\mathbf{q}}_n,\\tilde{\\mathbf{y}}_n)\\}_{n=1}^F" + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "text", + "content": " by pairing the mask embeddings with the assigned label. In our method, the goal of GMM is to model the data distribution of the joint-modal mask embedding " + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{q}}" + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "text", + "content": " for each class " + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "inline_equation", + "content": "D_q" + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "text", + "content": "-dimensional space by employing a weighted mixture of " + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 293, + 480, + 424 + ], + "type": "text", + "content": " multivariate Gaussians, defined as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 432, + 480, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 432, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 151, + 432, + 480, + 464 + ], + "type": "interline_equation", + "content": "p (\\tilde {\\mathbf {q}} | c) = \\sum_ {m = 1} ^ {M} p (m | c; \\boldsymbol {\\pi} _ {c}) p (\\tilde {\\mathbf {q}} | c, m; \\boldsymbol {\\mu} _ {c}, \\boldsymbol {\\Sigma} _ {c}) = \\sum_ {m = 1} ^ {M} \\boldsymbol {\\pi} _ {c m} \\mathcal {N} (\\tilde {\\mathbf {q}}; \\boldsymbol {\\mu} _ {c m}, \\boldsymbol {\\Sigma} _ {c m}), \\tag {2}", + "image_path": "98ca5cbf902a16b192ff01cdc064ffb0ec3e0312f5ac915e01b602a8e5f55bda.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 470, + 479, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 479, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 479, + 529 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 470, + 479, + 529 + ], + "type": "inline_equation", + "content": "\\sum_{m}\\pi_{cm} = 1" + }, + { + "bbox": [ + 130, + 470, + 479, + 529 + ], + "type": "text", + "content": " represents the mixing coefficients, " + }, + { + "bbox": [ + 130, + 470, + 479, + 529 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{cm}\\in \\mathbb{R}^{D_q}" + }, + { + "bbox": [ + 130, + 470, + 479, + 529 + ], + "type": "text", + "content": " denotes the mean vector, and " + }, + { + "bbox": [ + 130, + 470, + 479, + 529 + ], + "type": "inline_equation", + "content": "\\Sigma_{cm}\\in \\mathbb{R}^{D_q\\times D_q}" + }, + { + "bbox": [ + 130, + 470, + 479, + 529 + ], + "type": "text", + "content": " is the covariance matrix. The optimisation of the GMM parameters is performed by the Expectation Maximisation (EM) algorithm [10]. In the E-step, we iterate through " + }, + { + "bbox": [ + 130, + 470, + 479, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 470, + 479, + 529 + ], + "type": "text", + "content": ", computing the responsibilities using the current parameters of the GMM for each class " + }, + { + "bbox": [ + 130, + 470, + 479, + 529 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 470, + 479, + 529 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 205, + 536, + 480, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 536, + 480, + 569 + ], + "spans": [ + { + "bbox": [ + 205, + 536, + 480, + 569 + ], + "type": "interline_equation", + "content": "\\gamma_ {c, n} ^ {(t)} (m) = \\frac {\\pi_ {c m} ^ {(t - 1)} \\mathcal {N} \\left(\\tilde {\\mathbf {q}} _ {n} ; \\boldsymbol {\\mu} _ {c m} ^ {(t - 1)} , \\boldsymbol {\\Sigma} _ {c m} ^ {(t - 1)}\\right)}{\\sum_ {m ^ {\\prime} = 1} ^ {M} \\pi_ {c m ^ {\\prime}} ^ {(t - 1)} \\mathcal {N} \\left(\\tilde {\\mathbf {q}} _ {n} ; \\boldsymbol {\\mu} _ {c m ^ {\\prime}} ^ {(t - 1)} , \\boldsymbol {\\Sigma} _ {c m ^ {\\prime}} ^ {(t - 1)}\\right)}. \\tag {3}", + "image_path": "f07c22051ec60ac3704032b98de2d542138b5ab4ae5e86641cb61464d0d40c3c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 573, + 479, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 573, + 479, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 573, + 479, + 586 + ], + "type": "text", + "content": "We re-estimate the parameters using the calculated responsibility in the M-step." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 186, + 602, + 480, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 602, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 186, + 602, + 480, + 640 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} _ {c m} ^ {(t)} = \\frac {1}{N _ {c m} ^ {(t)}} \\sum_ {(\\tilde {\\mathbf {q}} _ {n}, \\tilde {\\mathbf {y}} _ {n}) \\in \\mathcal {F} _ {c}} \\gamma_ {c, n} ^ {(t)} (m) \\tilde {\\mathbf {q}}, \\quad \\boldsymbol {\\pi} _ {c m} ^ {(t)} = \\frac {N _ {c m} ^ {(t)}}{| \\mathcal {F} |}, \\tag {4}", + "image_path": "847a62c80ef888036653a4d1ad5e3ecbd58dba924c4430adc6baa16dfdb46d1b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 186, + 639, + 430, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 639, + 430, + 667 + ], + "spans": [ + { + "bbox": [ + 186, + 639, + 430, + 667 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} _ {c m} ^ {(t)} = \\frac {1}{N _ {c m} ^ {(t)}} \\sum_ {(\\tilde {\\mathbf {q}} _ {n}, \\tilde {\\mathbf {y}} _ {n}) \\in \\mathcal {F} _ {c}} \\gamma_ {c, n} ^ {(t)} (m) (\\tilde {\\mathbf {q}} _ {n} - \\boldsymbol {\\mu} _ {c m} ^ {(t)}) (\\tilde {\\mathbf {q}} _ {n} - \\boldsymbol {\\mu} _ {c m} ^ {(t)}) ^ {\\top},", + "image_path": "41e71f68e6e2818bf5839cc8aafc88ab337a84481928e4e4fddd6f93b78b8fdb.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 262, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 262, + 91, + 447, + 102 + ], + "type": "text", + "content": "Class-conditional Prompting Machine (CPM)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 479, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 479, + 153 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 479, + 153 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 114, + 479, + 153 + ], + "type": "inline_equation", + "content": "N_{cm}^{(t)} = \\sum_{(\\tilde{\\mathbf{q}}_n,\\tilde{\\mathbf{y}}_n)\\in \\mathcal{F}_c}\\gamma_{c,n}^{(t)}(m)" + }, + { + "bbox": [ + 130, + 114, + 479, + 153 + ], + "type": "text", + "content": ". We maintain an external memory bank containing GMM training samples, defined above with the set " + }, + { + "bbox": [ + 130, + 114, + 479, + 153 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 114, + 479, + 153 + ], + "type": "text", + "content": ", and the GMM parameters are robustly estimated with momentum [27]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "text", + "content": "Audio Conditional Prompting (ACP) Motivated by the concept of \"mix-and-separate\" [13, 18], we design the ACP process to improve cross-attention interaction for the dense audio feature representations. We consider an off-the-screen noise dataset " + }, + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r = \\{(\\mathbf{a}_j)\\}_{j=1}^{|D_r|}" + }, + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "text", + "content": ". For each training iteration, we sample a clean audio sample " + }, + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i" + }, + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "text", + "content": " and an off-the-screen noisy audio sample " + }, + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_j" + }, + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "text", + "content": " to build the mixture of magnitude spectrogram " + }, + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_p = \\mathbf{a}_i + \\mathbf{a}_j" + }, + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "text", + "content": ". The whole ACP process is illustrated in the first section of Fig. 2. Before delving into the audio recovery process, we first sample a set of class-conditional prompts " + }, + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^k \\sim \\{\\mathcal{N}(\\boldsymbol{\\mu}_k, \\boldsymbol{\\Sigma}_k)\\}_{k \\in \\mathcal{K}}" + }, + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "inline_equation", + "content": "\\mathcal{K} = \\{k | \\mathbf{t}_i(k) = 1\\}" + }, + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "text", + "content": " represent the indices for ground truth labels) via the GMM model from the last iteration according to the target semantic classes " + }, + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i" + }, + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "text", + "content": " that we want to recover. The major process of the audio recovery process shares a similar concept with the visual part described in Sec. 3.1; the difference is that we use the joint-modal mask embeddings " + }, + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{z}}^k" + }, + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "text", + "content": " to search the semantically similar sound source on the decoded dense audio feature map " + }, + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{u}}_a" + }, + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "text", + "content": " and generate their corresponding mask predictions " + }, + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{m}_k^a | k \\in \\mathcal{K}\\}" + }, + { + "bbox": [ + 130, + 162, + 480, + 343 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 344, + 481, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 481, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 481, + 451 + ], + "type": "text", + "content": "The learning objective of the ACP process is to encourage the consistency between the predicted masks " + }, + { + "bbox": [ + 130, + 344, + 481, + 451 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_k^a" + }, + { + "bbox": [ + 130, + 344, + 481, + 451 + ], + "type": "text", + "content": " with the derived ground-truth spectrogram ratio mask " + }, + { + "bbox": [ + 130, + 344, + 481, + 451 + ], + "type": "inline_equation", + "content": "\\frac{\\mathbf{a}_i}{\\mathbf{a}_p}" + }, + { + "bbox": [ + 130, + 344, + 481, + 451 + ], + "type": "text", + "content": ". The AVS datasets generally do not provide the separated sound source data for each semantic class (i.e., in Fig. 2, we do not have the ground-truth audio data of Male, Female and Guitar), hence we adopted a summation operation over the predicted masks to integrate all mixed sound sources. Finally, a Sigmoid activation function " + }, + { + "bbox": [ + 130, + 344, + 481, + 451 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 130, + 344, + 481, + 451 + ], + "type": "text", + "content": " is applied to constrain the prediction into a binary mask. We denoted the mean squared error (MSE) loss of the ACP process as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 240, + 452, + 480, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 452, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 240, + 452, + 480, + 483 + ], + "type": "interline_equation", + "content": "\\ell_ {\\mathrm {A C P}} = \\left\\| \\sigma \\left(\\sum_ {k \\in \\mathcal {K}} \\mathbf {m} _ {k} ^ {a}\\right) - \\frac {\\mathbf {a} _ {i}}{\\mathbf {a} _ {p}} \\right\\| _ {2}. \\tag {5}", + "image_path": "3380966cd4bfc53d1f685015fea39ffa9ccbe681b1200c6ddbd9873b596febdb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 499, + 482, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 499, + 482, + 620 + ], + "spans": [ + { + "bbox": [ + 130, + 499, + 482, + 620 + ], + "type": "text", + "content": "Visual Conditional Prompting (VCP) The VCP module is processed simultaneously with ACP. The design of VCP aims to bypass bipartite matching via the generated class-conditional prompts " + }, + { + "bbox": [ + 130, + 499, + 482, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^k" + }, + { + "bbox": [ + 130, + 499, + 482, + 620 + ], + "type": "text", + "content": " to ease the model training as it can provide a more stable learning target for each query feature to mitigate the instability brought by the bipartite matching with the class-agnostic queries. Given a set of class-conditional prompts derived from the ground-truth image labels, our training objective of VCP is to correctly segment the corresponding image regions as well as successfully classify these prompts after the consecutive transformer decoder layers. The loss function is similar to " + }, + { + "bbox": [ + 130, + 499, + 482, + 620 + ], + "type": "inline_equation", + "content": "\\ell_{\\text{agn}}" + }, + { + "bbox": [ + 130, + 499, + 482, + 620 + ], + "type": "text", + "content": " in Sec. 3.1, where we replace " + }, + { + "bbox": [ + 130, + 499, + 482, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 130, + 499, + 482, + 620 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 499, + 482, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^k" + }, + { + "bbox": [ + 130, + 499, + 482, + 620 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 130, + 499, + 482, + 620 + ], + "type": "inline_equation", + "content": "\\ell_{\\text{VCP}} = \\ell_{ce} + \\ell_{mask}" + }, + { + "bbox": [ + 130, + 499, + 482, + 620 + ], + "type": "text", + "content": " (6)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Prompting-based Contrastive Learning (PCL) The ability to learn discriminative feature representation is critical for the audio-visual system. One limitation of the previous audio-visual contrastive learning methods [7,36] is that" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "Y. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "text", + "content": "they can only leverage the global audio representation due to the lack of class conditional prompts for class-specific feature disentanglement. By taking advantage of class-conditional distribution modelling, we can overcome this limitation by utilising the predicted spectrogram saliency mask " + }, + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_k^a = \\sigma (\\mathbf{m}_k^a)" + }, + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "text", + "content": " and its associated class label " + }, + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "text", + "content": " of each sound source, denoted as " + }, + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{s}_k^a |\\mathbf{s}_k^a\\in \\mathcal{M}\\subseteq [0,1]^T\\times F,k\\in \\mathcal{K}\\}" + }, + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "text", + "content": ". To disentangle the class-specific representations for the audio feature map, we iteratively apply the masked average pooling (MAP) [47, 58] to " + }, + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_a" + }, + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "inline_equation", + "content": "|\\mathcal{K}|" + }, + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "text", + "content": " saliency masks " + }, + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_k^a" + }, + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "text", + "content": " for the feature extraction process. We first threshold " + }, + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_k^a" + }, + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "text", + "content": " and convert it to a binary mask via " + }, + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{s}}_k^a = \\mathbb{1}(\\mathbf{s}_k^a >\\bar{\\mathbf{s}}_k^a)" + }, + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{s}}_k^a" + }, + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "text", + "content": " represent the mean value of the saliency mask. Then, for the " + }, + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "text", + "content": "-th category, we extract its region-level mean feature " + }, + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{f}(k)\\in \\mathbb{R}^{D_q}" + }, + { + "bbox": [ + 130, + 116, + 479, + 247 + ], + "type": "text", + "content": " by MAP defined as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 255, + 257, + 481, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 257, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 255, + 257, + 481, + 285 + ], + "type": "interline_equation", + "content": "\\mathbf {f} (k) = \\frac {\\sum_ {\\psi \\in \\Psi} \\tilde {\\mathbf {s}} _ {k} ^ {a} \\mathbf {u} _ {a} (\\psi)}{\\sum_ {\\psi} \\tilde {\\mathbf {s}} _ {k} ^ {a}}, \\tag {7}", + "image_path": "55cf60398db184f96796c70e1fcf4ff59e714efaa6fdc7fbbfc12b862832edf6.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 294, + 480, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 480, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 480, + 366 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 294, + 480, + 366 + ], + "type": "inline_equation", + "content": "\\varPsi" + }, + { + "bbox": [ + 130, + 294, + 480, + 366 + ], + "type": "text", + "content": " is the lattice of size " + }, + { + "bbox": [ + 130, + 294, + 480, + 366 + ], + "type": "inline_equation", + "content": "T\\times F" + }, + { + "bbox": [ + 130, + 294, + 480, + 366 + ], + "type": "text", + "content": ". We combine the " + }, + { + "bbox": [ + 130, + 294, + 480, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{f}(k)" + }, + { + "bbox": [ + 130, + 294, + 480, + 366 + ], + "type": "text", + "content": " with the ground-truth class label " + }, + { + "bbox": [ + 130, + 294, + 480, + 366 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 294, + 480, + 366 + ], + "type": "text", + "content": " to form the anchor set " + }, + { + "bbox": [ + 130, + 294, + 480, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{anch}} = \\{(\\mathbf{f}(k),k)|\\mathbf{f}(k)\\in \\mathcal{U},k\\in \\mathcal{K}\\}" + }, + { + "bbox": [ + 130, + 294, + 480, + 366 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 294, + 480, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 130, + 294, + 480, + 366 + ], + "type": "text", + "content": " denoting the audio and visual feature spaces defined in Sec. 3.1. Since the pixel-level label " + }, + { + "bbox": [ + 130, + 294, + 480, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 130, + 294, + 480, + 366 + ], + "type": "text", + "content": " is available, we can directly form the visual contrastive set " + }, + { + "bbox": [ + 130, + 294, + 480, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{cont}} = \\{(\\mathbf{u}_v(\\omega),\\mathbf{y}_v(\\omega))|\\mathbf{u}_v(\\omega)\\in \\mathcal{U},\\mathbf{y}(\\omega)\\in \\mathcal{V}\\}" + }, + { + "bbox": [ + 130, + 294, + 480, + 366 + ], + "type": "text", + "content": ". Hence, we can derive the positive and negative sets as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 211, + 377, + 480, + 395 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 377, + 480, + 395 + ], + "spans": [ + { + "bbox": [ + 211, + 377, + 480, + 395 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {P} (\\mathbf {u} _ {v} (\\omega)) = \\left\\{\\mathbf {u} _ {v} (\\omega) \\mid \\mathbf {u} _ {v} (\\omega) \\in \\mathcal {U}, \\mathbf {y} (\\omega , k) = 1 \\right\\}, \\\\ \\mathcal {U} (v, v) = \\left\\{v: (v) \\mid v: (v) \\in \\mathcal {U}, (v - 1) = 0 \\right\\}. \\end{array} \\tag {8}", + "image_path": "f1402bdba4e4e4798d9f11f46deed3dc2cc8865f73b67dcb82334ec4481c1c55.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 209, + 391, + 405, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 391, + 405, + 403 + ], + "spans": [ + { + "bbox": [ + 209, + 391, + 405, + 403 + ], + "type": "interline_equation", + "content": "\\mathcal {N} (\\mathbf {u} _ {v} (\\omega)) = \\left\\{\\mathbf {u} _ {v} (\\omega) | \\mathbf {u} _ {v} (\\omega) \\in \\mathcal {U}, \\mathbf {y} (\\omega , k) = 0 \\right\\}.", + "image_path": "09e77a280d32d44a1ad272aa01391c1b3f3a1c9879941f3c78bafbf8786afc78.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 413, + 479, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 479, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 479, + 449 + ], + "type": "text", + "content": "Adopting the supervised InfoNCE [21] as the objective function to pull the anchor " + }, + { + "bbox": [ + 130, + 413, + 479, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{f} \\in \\mathcal{E}_{\\mathrm{anch}}" + }, + { + "bbox": [ + 130, + 413, + 479, + 449 + ], + "type": "text", + "content": " and respective positive visual features closer while repelling anchors and their negative visual features, we define the following loss:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 144, + 458, + 480, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 458, + 480, + 484 + ], + "spans": [ + { + "bbox": [ + 144, + 458, + 480, + 484 + ], + "type": "interline_equation", + "content": "\\ell_ {\\mathrm {P C L}} (\\mathbf {f}) = \\frac {1}{| \\mathcal {P} \\left(\\mathbf {u} _ {v} (\\omega)\\right) |} \\sum_ {\\mathbf {f} _ {p} \\in \\mathcal {P} \\left(\\mathbf {u} _ {v} (\\omega)\\right)} - \\log \\frac {\\exp \\left(\\mathbf {f} \\cdot \\mathbf {f} _ {p} / \\tau\\right)}{\\exp \\left(\\mathbf {f} \\cdot \\mathbf {f} _ {p} / \\tau\\right) + \\sum_ {\\mathbf {f} _ {n} \\in \\mathcal {N} \\left(\\mathbf {u} _ {v} (\\omega)\\right)} \\exp \\left(\\mathbf {f} \\cdot \\mathbf {f} _ {n} / \\tau\\right)}, \\tag {9}", + "image_path": "ebaffd30a00af4423febb7c073dd4f6a5d2958cfd8249f655c70e446dc59e890.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 494, + 479, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 479, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 479, + 529 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 494, + 479, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 130, + 494, + 479, + 529 + ], + "type": "text", + "content": " is an anchor feature, and " + }, + { + "bbox": [ + 130, + 494, + 479, + 529 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 494, + 479, + 529 + ], + "type": "text", + "content": " is the temperature hyperparameter. The combination of all sub-objective " + }, + { + "bbox": [ + 130, + 494, + 479, + 529 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{ACP}}" + }, + { + "bbox": [ + 130, + 494, + 479, + 529 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 494, + 479, + 529 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{VCP}}" + }, + { + "bbox": [ + 130, + 494, + 479, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 494, + 479, + 529 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{PCL}}" + }, + { + "bbox": [ + 130, + 494, + 479, + 529 + ], + "type": "text", + "content": " form the CPM loss " + }, + { + "bbox": [ + 130, + 494, + 479, + 529 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{CPM}} = \\ell_{\\mathrm{ACP}} + \\ell_{\\mathrm{VCP}} + \\ell_{\\mathrm{PCL}}" + }, + { + "bbox": [ + 130, + 494, + 479, + 529 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 543, + 479, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 543, + 479, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 543, + 479, + 567 + ], + "type": "text", + "content": "1. Overall training The overall training objective is " + }, + { + "bbox": [ + 130, + 543, + 479, + 567 + ], + "type": "inline_equation", + "content": "\\ell = \\ell_{\\text{agn}} + \\lambda \\ell_{\\text{CPM}}" + }, + { + "bbox": [ + 130, + 543, + 479, + 567 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 543, + 479, + 567 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 543, + 479, + 567 + ], + "type": "text", + "content": " is the weight coefficient." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 576, + 229, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 576, + 229, + 590 + ], + "spans": [ + { + "bbox": [ + 131, + 576, + 229, + 590 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 597, + 279, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 597, + 279, + 609 + ], + "spans": [ + { + "bbox": [ + 131, + 597, + 279, + 609 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": "Evaluation Protocols We utilize standard evaluation protocols from AVSBench datasets [56, 57] for single-source (SS) and multi-source (MS) scenarios with binary labels, as well as for AVSBench-Semantics with multi-class labels. Image sizes are standardized to " + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": " for fair comparison. Additionally, we" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 262, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 262, + 91, + 447, + 102 + ], + "type": "text", + "content": "Class-conditional Prompting Machine (CPM)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 159, + 148, + 451, + 304 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Quantitative (mIoU, " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": ") audio-visual segmentation results (\\%) for the AVS-Bench test sets [56, 57] (resized to " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": ") with ResNet50 [17] backbone. Best results in bold, " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " best underlined. Improvements against the " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " best are in the last row." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 148, + 451, + 304 + ], + "lines": [ + { + "bbox": [ + 159, + 148, + 451, + 304 + ], + "spans": [ + { + "bbox": [ + 159, + 148, + 451, + 304 + ], + "type": "table", + "html": "
Eval. MethodsMethodAVSBench-Obj.(SS)AVSBench-Obj.(MS)AVSBench-S
mIoU ↑Fβ ↑mIoU ↑Fβ ↑mIoU ↑Fβ ↑
Per-image [56,57]TPAVI [57]72.7984.8047.8857.8020.1825.20
AVSBG [16]74.1385.4044.9556.80--
ECMVAE [37]76.3386.5048.6960.70--
DiffusionAVS [36]75.8086.9049.7762.10--
CATR [24]74.8086.6052.8065.30--
AuTR [31]75.0085.2049.4061.20--
AQFormer [19]77.0086.4055.7066.90--
AVSegFormer [14]76.5484.8049.5362.8024.9329.30
AVSC [29]77.0285.2449.5861.51--
BAVS [30]77.9685.2950.2362.3724.6829.63
AVSAC [5]76.9086.9553.9565.8125.4329.71
QSD [26]77.6086.0059.6063.5046.60-
COMBO [54]81.7090.1054.5066.6033.3037.30
Per-dataset [12]AVSegFormer* [14]79.8091.4153.2864.7641.6455.05
CAVP [7]85.7792.8662.3973.6244.7057.76
CPM87.6493.5365.2275.2248.3964.01
", + "image_path": "77b2707f40699657f856fb1e078d3b5f65358d400e0053ca9681b41592eca0a3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 314, + 482, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 458 + ], + "type": "text", + "content": "adopt the original image resolution for testing, following [7], to demonstrate optimal model performance. Evaluation is extended to the Visual Post-production (VPO) benchmark [7] for challenging cases. We employ mean Intersection over Union (mIoU) [11] and " + }, + { + "bbox": [ + 130, + 314, + 482, + 458 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 130, + 314, + 482, + 458 + ], + "type": "text", + "content": " score with " + }, + { + "bbox": [ + 130, + 314, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\beta^2 = 0.3" + }, + { + "bbox": [ + 130, + 314, + 482, + 458 + ], + "type": "text", + "content": " [38, 57] to assess segmentation quality, precision, and recall performance at pixel level. Models are stratified into CNN-based pre-pixel classification and transformer-based mask classification to demonstrate architectural capabilities. Official training splits are used for AVSBench datasets [56, 57] and VPO [7], with results reported on the respective testing sets. Training is performed on the entire AVSBench-Semantics dataset, and testing is conducted on subsets as well as the entire testing set to demonstrate partitioned model performance. For further details on training and inference, please refer to the Supplementary Material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "text", + "content": "Results We collected the experimental results from the existing benchmark [7] and updated it with recent works [5,26,54]. We modified the AVSegFormer [14] with Mask2former [8] and denoted it as AVSegFormer* in the tables to encourage a fair comparison with our method. Please note that Tab. 1 includes two evaluation protocols. We employ standard semantic segmentation protocols to compute both mIoU and " + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "text", + "content": " same as PascalVOC, initially mentioned in [57]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 540, + 413, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 540, + 413, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 540, + 413, + 552 + ], + "type": "text", + "content": "4.2 Performance on Low-resolution AVSBench Videos" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 667 + ], + "type": "text", + "content": "We adopt established methodologies [7, 30, 31, 37] for conducting performance evaluations on down-sampled image benchmarks, such as AVSBench-Objects [57], which includes single-source (SS) and multi-source (MS) splits with binary annotations, as well as the AVSBench-Semantics dataset [56], which contains multiclass annotations. We compare the performance of state-of-the-art (SOTA) methods with our CPM in Tab. 1 using mIoU and " + }, + { + "bbox": [ + 130, + 557, + 482, + 667 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 130, + 557, + 482, + 667 + ], + "type": "text", + "content": ". The results demonstrate that our model surpasses the second-best CAVP [7] in terms of mIoU by " + }, + { + "bbox": [ + 130, + 557, + 482, + 667 + ], + "type": "inline_equation", + "content": "1.87\\%" + }, + { + "bbox": [ + 130, + 557, + 482, + 667 + ], + "type": "text", + "content": " on AVSBench-Object (SS) [57], " + }, + { + "bbox": [ + 130, + 557, + 482, + 667 + ], + "type": "inline_equation", + "content": "2.83\\%" + }, + { + "bbox": [ + 130, + 557, + 482, + 667 + ], + "type": "text", + "content": " on AVSBench-Object (MS) [57] and " + }, + { + "bbox": [ + 130, + 557, + 482, + 667 + ], + "type": "inline_equation", + "content": "1.79\\%" + }, + { + "bbox": [ + 130, + 557, + 482, + 667 + ], + "type": "text", + "content": " on AVSBench-Semantics [57] using the ResNet-50 [17] backbone." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "Y. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 168, + 159, + 442, + 235 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "text", + "content": "Table 2: Quantitative (mIoU, " + }, + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "text", + "content": ") audio-visual segmentation results (in %) for the AVSBench-Semantic (AVSS) test sets [56] (original resolution) with ResNet50 [17] backbone. Best results in bold, " + }, + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "text", + "content": " best underlined. Improvements against the " + }, + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "text", + "content": " best are in the last row." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 168, + 159, + 442, + 235 + ], + "lines": [ + { + "bbox": [ + 168, + 159, + 442, + 235 + ], + "spans": [ + { + "bbox": [ + 168, + 159, + 442, + 235 + ], + "type": "table", + "html": "
D-ResNet50 [17]MethodAVSS (SS)AVSS (MS)AVSS
mIoU ↑Fβ ↑mIoU ↑Fβ ↑mIoU ↑Fβ ↑
Per-pixel ClassificationTPAVI [57]42.1061.4626.3340.9943.3959.24
CAVP [7]56.9169.1538.6152.9250.7564.57
TransformerAVSegFormer [14]46.2559.7627.2141.3841.4856.21
AVSegFormer* [14]50.5263.7531.4042.8145.8059.16
CPM61.7172.9443.1156.2857.2570.54
ImprovementCPM+4.80+3.79+4.50+3.36+6.50+5.97
", + "image_path": "d7c01bd7e5822334db778a916bb4b0eb49ba56821a42f3e046b7c5745bfc97da.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 168, + 269, + 442, + 346 + ], + "blocks": [ + { + "bbox": [ + 130, + 236, + 481, + 267 + ], + "lines": [ + { + "bbox": [ + 130, + 236, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 481, + 267 + ], + "type": "text", + "content": "Table 3: Quantitative (mIoU, " + }, + { + "bbox": [ + 130, + 236, + 481, + 267 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 130, + 236, + 481, + 267 + ], + "type": "text", + "content": ") audio-visual segmentation results (in %) for the VPO test sets (original resolution) with ResNet50 [17] backbone. Best results in bold, " + }, + { + "bbox": [ + 130, + 236, + 481, + 267 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 130, + 236, + 481, + 267 + ], + "type": "text", + "content": " best underlined. Improvements against the " + }, + { + "bbox": [ + 130, + 236, + 481, + 267 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 130, + 236, + 481, + 267 + ], + "type": "text", + "content": " best are in the last row." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 168, + 269, + 442, + 346 + ], + "lines": [ + { + "bbox": [ + 168, + 269, + 442, + 346 + ], + "spans": [ + { + "bbox": [ + 168, + 269, + 442, + 346 + ], + "type": "table", + "html": "
D-ResNet50 [17]MethodVPO (SS)VPO (MS)VPO (MSMI)
mIoU ↑Fβ ↑mIoU ↑Fβ ↑mIoU ↑Fβ ↑
Per-pixel ClassificationTPAVI [57]52.7569.5454.3071.9551.7368.85
CAVP62.3178.4664.3178.9260.3675.60
TransformerAVSegFormer [14]57.5573.0358.3374.2854.2270.39
AVSegFormer* [14]60.5174.8162.9177.3356.2472.67
CPM67.0979.8865.9179.9060.5575.58
ImprovementsCPM+4.78+1.42+1.60+0.98+0.19-0.02
", + "image_path": "5f27222fcf904db2238c12d81bccaffec47182983902d76aaab931e669b2069a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 356, + 380, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 380, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 380, + 369 + ], + "type": "text", + "content": "4.3 Performance on Original AVSBench Videos" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 369, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 609 + ], + "type": "text", + "content": "The benchmark mentioned above adopted low-resolution benchmarks using significantly resized input images (from 720p to " + }, + { + "bbox": [ + 130, + 369, + 482, + 609 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 369, + 482, + 609 + ], + "type": "text", + "content": "). While this simplifies model training, disregarding the original image aspect ratio can lead to a degradation in model performance, which is not advisable for segmentation tasks. Therefore, we follow the evaluation method outlined in [7] for training and testing on the raw AVSBench videos. This includes using random resized crops for training and performing frame-by-frame evaluation during testing. The initial published version of AVSBench [57] only offers single-source and multisource partitioning. However, this partitioning was not conducted in the later version [56], resulting in some cases being overlooked. We re-organised the previous SS and MS with the newly added video data, resulting in 1278 single-source videos (AVSS-SS) and 276 multi-source videos (AVSS-MS). We compare the results with mIoU and " + }, + { + "bbox": [ + 130, + 369, + 482, + 609 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 130, + 369, + 482, + 609 + ], + "type": "text", + "content": " to demonstrate the comprehensive model performance in Tab. 2. Our method shows a significant mIoU improvement of " + }, + { + "bbox": [ + 130, + 369, + 482, + 609 + ], + "type": "inline_equation", + "content": "4.80\\%" + }, + { + "bbox": [ + 130, + 369, + 482, + 609 + ], + "type": "text", + "content": " on AVSS-SS, " + }, + { + "bbox": [ + 130, + 369, + 482, + 609 + ], + "type": "inline_equation", + "content": "4.50\\%" + }, + { + "bbox": [ + 130, + 369, + 482, + 609 + ], + "type": "text", + "content": " on AVSS-MS and " + }, + { + "bbox": [ + 130, + 369, + 482, + 609 + ], + "type": "inline_equation", + "content": "+6.50\\%" + }, + { + "bbox": [ + 130, + 369, + 482, + 609 + ], + "type": "text", + "content": " on the entire AVSS dataset. To further demonstrate the effectiveness of our method, we show a visualisation of 6-second video clip in Fig. 5 that displays a qualitative comparison between TPAVI, AVSegFormer, CAVP and our CPM. Our method can successfully approximate the ground truth segmentation of the target sound source within a group of other semantic objects." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 624, + 312, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 624, + 312, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 624, + 312, + 635 + ], + "type": "text", + "content": "4.4 Performance on VPO Dataset" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "We also compare the model performance on the VPO benchmark [7], equipped with synthesized stereo audios. We adopted ResNet-50 [17] backbone for all" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 262, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 262, + 91, + 447, + 102 + ], + "type": "text", + "content": "Class-conditional Prompting Machine (CPM)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 149, + 126, + 457, + 202 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 471, + 125 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 471, + 125 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 471, + 125 + ], + "type": "text", + "content": "Table 4: Ablation study of the model components on AVSBench-Semantics [31]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 126, + 457, + 202 + ], + "lines": [ + { + "bbox": [ + 149, + 126, + 457, + 202 + ], + "spans": [ + { + "bbox": [ + 149, + 126, + 457, + 202 + ], + "type": "table", + "html": "
MethodMetrics
BaselineCCDMVCPACPPCLmIoU ↑Fβ ↑
✓53.0465.29
✓✓54.1266.65
✓✓✓55.7968.17
✓✓✓55.0768.03
✓✓✓✓56.4169.07
✓✓✓✓✓57.2570.54
", + "image_path": "f1aecbcb4a60bc22c9f039aeafe13eaa23b9ac8c1bd9d2ff4b23258e507271a8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 157, + 203, + 304, + 262 + ], + "blocks": [ + { + "bbox": [ + 157, + 203, + 304, + 262 + ], + "lines": [ + { + "bbox": [ + 157, + 203, + 304, + 262 + ], + "spans": [ + { + "bbox": [ + 157, + 203, + 304, + 262 + ], + "type": "image", + "image_path": "26a18b65d97af0e551be591b6aa13fb28b79e7304bbee5c7e72fc29a7292de34.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 263, + 272, + 271 + ], + "lines": [ + { + "bbox": [ + 190, + 263, + 272, + 271 + ], + "spans": [ + { + "bbox": [ + 190, + 263, + 272, + 271 + ], + "type": "text", + "content": "(a) Visual comparison." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 205, + 453, + 262 + ], + "blocks": [ + { + "bbox": [ + 306, + 205, + 453, + 262 + ], + "lines": [ + { + "bbox": [ + 306, + 205, + 453, + 262 + ], + "spans": [ + { + "bbox": [ + 306, + 205, + 453, + 262 + ], + "type": "image", + "image_path": "6846e96eea8630f3724ef00e3f1716a43504c81c0cd83464be8a5e18e380199e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 263, + 435, + 272 + ], + "lines": [ + { + "bbox": [ + 323, + 263, + 435, + 272 + ], + "spans": [ + { + "bbox": [ + 323, + 263, + 435, + 272 + ], + "type": "text", + "content": "(b) Coverage score comparison." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 272, + 480, + 293 + ], + "lines": [ + { + "bbox": [ + 130, + 272, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 480, + 293 + ], + "type": "text", + "content": "Fig. 3: Qualitative (3a) and quantitative (3b) comparisons between model components in a multi-source scenario (i.e., male singing, female singing and guitar)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 308, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 439 + ], + "type": "text", + "content": "three subsets, and the results are shown in Tab. 3. To facilitate stereo audio encoding, we adopt the approach outlined in [7] by adjusting the number of input channels in the first layer to 2. Our approach surpasses the SOTA method CAVP [7] by " + }, + { + "bbox": [ + 130, + 308, + 482, + 439 + ], + "type": "inline_equation", + "content": "4.78\\%" + }, + { + "bbox": [ + 130, + 308, + 482, + 439 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 308, + 482, + 439 + ], + "type": "inline_equation", + "content": "1.60\\%" + }, + { + "bbox": [ + 130, + 308, + 482, + 439 + ], + "type": "text", + "content": " in terms of mIoU on VPO (SS) and VPO (MS), respectively. However, we observe a marginal improvement of " + }, + { + "bbox": [ + 130, + 308, + 482, + 439 + ], + "type": "inline_equation", + "content": "0.19\\%" + }, + { + "bbox": [ + 130, + 308, + 482, + 439 + ], + "type": "text", + "content": " on the VPO (MSMI) setup compared with the SS and MS subsets. The possible explanation for this phenomenon may stem from the overly simplistic integration of stereo audio within the transformer architecture, as the foundational AVS transformer architectures [14, 24, 26, 30] did not account for a distinct positional encoding scheme tailored for stereo data. We will explore the design of a dedicated stereo AVS transformer architecture as part of our future work." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 449, + 238, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 238, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 238, + 460 + ], + "type": "text", + "content": "4.5 Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 464, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 596 + ], + "type": "text", + "content": "Ablation of Key Components We first perform the key components analysis of CPM on AVSBench-Semantics [56] in Tab. 4. The baseline is AVSegformer* [14] in the 1st row. We replace the Softmax classifier in baseline with CCDM defined in (1), and observe an mIoU improvement of " + }, + { + "bbox": [ + 130, + 464, + 482, + 596 + ], + "type": "inline_equation", + "content": "+1.08\\%" + }, + { + "bbox": [ + 130, + 464, + 482, + 596 + ], + "type": "text", + "content": ". By integrating the model with VCP (3rd row), using training loss in (6), and ACP (4th row), using loss in (5), separately, we achieve an improvement of " + }, + { + "bbox": [ + 130, + 464, + 482, + 596 + ], + "type": "inline_equation", + "content": "+1.26\\%" + }, + { + "bbox": [ + 130, + 464, + 482, + 596 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 464, + 482, + 596 + ], + "type": "inline_equation", + "content": "+1.39\\%" + }, + { + "bbox": [ + 130, + 464, + 482, + 596 + ], + "type": "text", + "content": ", respectively. Subsequently, when we apply both ACP and VCP methods (5th row) the mIoU performance improves " + }, + { + "bbox": [ + 130, + 464, + 482, + 596 + ], + "type": "inline_equation", + "content": "+2.04\\%" + }, + { + "bbox": [ + 130, + 464, + 482, + 596 + ], + "type": "text", + "content": " compared to the 2nd row. The final row displays the complete CPM method, incorporating the dense contrastive learning approach we introduced, which leads to an additional mIoU improvement of " + }, + { + "bbox": [ + 130, + 464, + 482, + 596 + ], + "type": "inline_equation", + "content": "+1.09\\%" + }, + { + "bbox": [ + 130, + 464, + 482, + 596 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Ablation of Bipartite Matching Stability To study the stability of the bipartite matching process, we design an entropy-based stability score STS to quantify such performance on AVSBench-Semantics [56]. Intuitively, the STS measures the average assignment consistency after bipartite matching across all the classes. During the training process, we collect the assigned label " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{y}}_i \\in" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "Y. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 144, + 293, + 226 + ], + "blocks": [ + { + "bbox": [ + 141, + 144, + 293, + 226 + ], + "lines": [ + { + "bbox": [ + 141, + 144, + 293, + 226 + ], + "spans": [ + { + "bbox": [ + 141, + 144, + 293, + 226 + ], + "type": "image", + "image_path": "c8f2d72823cf817151f3c48a18aca44174943092ad494ff965295b1b42bcdf80.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 227, + 304, + 250 + ], + "lines": [ + { + "bbox": [ + 132, + 227, + 304, + 250 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 304, + 250 + ], + "type": "text", + "content": "Fig.4: Matching stability (STS " + }, + { + "bbox": [ + 132, + 227, + 304, + 250 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 132, + 227, + 304, + 250 + ], + "type": "text", + "content": " ) comparison on AVSBench-Semantics [31]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 319, + 160, + 455, + 241 + ], + "blocks": [ + { + "bbox": [ + 304, + 137, + 476, + 159 + ], + "lines": [ + { + "bbox": [ + 304, + 137, + 476, + 159 + ], + "spans": [ + { + "bbox": [ + 304, + 137, + 476, + 159 + ], + "type": "text", + "content": "Table 5: Ablation of GMM modelling on AVSBench-Semantics [31]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 160, + 455, + 241 + ], + "lines": [ + { + "bbox": [ + 319, + 160, + 455, + 241 + ], + "spans": [ + { + "bbox": [ + 319, + 160, + 455, + 241 + ], + "type": "table", + "html": "
MethodmIoU ↑Fβ ↓
Point Rep.55.9668.58
Dist Rep.57.2570.54
MmIoUFβ
155.8368.68
357.2570.54
556.3169.04
756.0568.59
", + "image_path": "a13f8acf23366f7b33e25c0d55285bb020f275b621e931d26ec058ee45c1b788.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\{0, \\dots, C\\}^N" + }, + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "type": "text", + "content": " to the class-agnostic query " + }, + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{q}}_n" + }, + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "type": "text", + "content": " produced by the Hungarian algorithm for each training image, resulting in " + }, + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{R} = [\\tilde{\\mathbf{y}}_0 \\oplus \\tilde{\\mathbf{y}}_1 \\oplus, \\dots, \\oplus \\tilde{\\mathbf{y}}_D]" + }, + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "type": "text", + "content": " represents the concatenation operator and " + }, + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "type": "text", + "content": " has size " + }, + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "type": "inline_equation", + "content": "D \\times N" + }, + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "type": "text", + "content": ". We then use the indicator function to extract the assignment information over all semantic classes. We denote the resulting set as " + }, + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{s}_0, \\mathbf{s}_1, \\dots, \\mathbf{s}_c\\}" + }, + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_c(n) = \\sum_{d=1}^{D} \\mathbb{1}(\\mathbf{R}(d,n) = c)" + }, + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "type": "inline_equation", + "content": "n \\in \\{1, \\dots, N\\}" + }, + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "type": "text", + "content": ", which forms " + }, + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_c \\in [0,1]^N" + }, + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "type": "text", + "content": " after normalising it to be a probability distribution. Finally, we calculate the stability score " + }, + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "type": "inline_equation", + "content": "\\mathsf{STS} = \\frac{1}{C} \\sum_{c=1}^{C} H(\\mathbf{s}_c)" + }, + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "type": "inline_equation", + "content": "H(\\cdot)" + }, + { + "bbox": [ + 130, + 258, + 482, + 416 + ], + "type": "text", + "content": " computes the entropy. We compare our CPM with AVSegformer* over 100 training epochs as illustrated in Fig. 4. Our findings reveal that our CPM consistently improves the stability score (showing lower average entropy) throughout the training process in comparison to AVSegformer*, thereby validating the efficacy of our approach in enhancing bipartite matching stability. Please refer to the Supplementary Material for the Pseudo-code of STS." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": "Ablation of Audio Contribution While experimenting with the baseline method mentioned above, we empirically observed that the classification after the audio transformer decoder (i.e., " + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "inline_equation", + "content": "f^{\\mathrm{TD - A}}" + }, + { + "bbox": [ + 130, + 437, + 482, + 666 + ], + "type": "text", + "content": " in Fig. 2) has low accuracy, with noisy predictions. We also found that these noisy predictions are progressively refined by the vision transformer decoder layer based solely on visual clues. Such an observation illustrates that in some hard cases, we may fail to extract useful semantic information from audio, leading to the model being overly reliant on the visual content, which may cause erroneous testing predictions, as shown in Fig. 3a-(i), (ii), (iii). To monitor the amount of valid information retained by the query following interaction with the audio branch, we introduce a classification coverage score. This score is computed by comparing the predictions generated after passing through the audio transformer decoder with the ground-truth class label, using the Jaccard index. We show qualitative results in Fig. 3a for the designated scenario and quantitative results in Fig. 3b for the AVSBench-Semantics testing set. The comparison starts with AVSegFormer as the baseline method with progressive addition of our proposed modules, similar to Tab. 4. Notice that each CPM sub-module contributes to refining key semantic information from the audio modality, as evidenced by the segmentation mask predictions and the improvement in Jaccard indexes. This shows the effectiveness of CPM." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 262, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 262, + 90, + 447, + 102 + ], + "type": "text", + "content": "Class-conditional Prompting Machine (CPM)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 105, + 444, + 236 + ], + "blocks": [ + { + "bbox": [ + 167, + 105, + 444, + 236 + ], + "lines": [ + { + "bbox": [ + 167, + 105, + 444, + 236 + ], + "spans": [ + { + "bbox": [ + 167, + 105, + 444, + 236 + ], + "type": "image", + "image_path": "215ea39162a9a98b1b8b6cd2fd660eb079218ae58c13dd8d6b376358c38c42bf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 236, + 480, + 269 + ], + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 269 + ], + "type": "text", + "content": "Fig. 5: Qualitative audio-visual segmentation results on AVSBench-Semantics [56] by TPAVI [57], AVSegFormer [14], CAVP [7] and our CPM, which can be compared with the ground truth (GT) Ambulance of the first row." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 283, + 482, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 427 + ], + "type": "text", + "content": "Ablation of Distribution Representation Table 5 provides a study of the GMM data distribution representation on AVSBench-Semantics [56]. The first row (Point Rep.) replaces the GMM module with a linear SoftMax for the classification output, introducing a learnable " + }, + { + "bbox": [ + 130, + 283, + 482, + 427 + ], + "type": "inline_equation", + "content": "C \\times D_q" + }, + { + "bbox": [ + 130, + 283, + 482, + 427 + ], + "type": "text", + "content": " feature map (i.e., " + }, + { + "bbox": [ + 130, + 283, + 482, + 427 + ], + "type": "inline_equation", + "content": "D_q" + }, + { + "bbox": [ + 130, + 283, + 482, + 427 + ], + "type": "text", + "content": "-dimensional feature embeddings for " + }, + { + "bbox": [ + 130, + 283, + 482, + 427 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 283, + 482, + 427 + ], + "type": "text", + "content": " semantic classes for the audio and visual prompting task). The second row (Dist Rep.) shows our approach with the CCDM module, showing that the utilisation of the distribution modelling leads to considerable improvements of " + }, + { + "bbox": [ + 130, + 283, + 482, + 427 + ], + "type": "inline_equation", + "content": "1.29\\%" + }, + { + "bbox": [ + 130, + 283, + 482, + 427 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 283, + 482, + 427 + ], + "type": "inline_equation", + "content": "1.96\\%" + }, + { + "bbox": [ + 130, + 283, + 482, + 427 + ], + "type": "text", + "content": " on mIoU and " + }, + { + "bbox": [ + 130, + 283, + 482, + 427 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 130, + 283, + 482, + 427 + ], + "type": "text", + "content": " respectively. The next rows of Tab. 5 show a hyper-parameter analysis on the number of GMM components (from " + }, + { + "bbox": [ + 130, + 283, + 482, + 427 + ], + "type": "inline_equation", + "content": "M = 1" + }, + { + "bbox": [ + 130, + 283, + 482, + 427 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 283, + 482, + 427 + ], + "type": "inline_equation", + "content": "M = 7" + }, + { + "bbox": [ + 130, + 283, + 482, + 427 + ], + "type": "text", + "content": " components). Results reveal that 3 GMM components enable the best fitting for the data distribution, leading to the top performance with a maximum improvement of " + }, + { + "bbox": [ + 130, + 283, + 482, + 427 + ], + "type": "inline_equation", + "content": "1.42\\%" + }, + { + "bbox": [ + 130, + 283, + 482, + 427 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 283, + 482, + 427 + ], + "type": "inline_equation", + "content": "1.95\\%" + }, + { + "bbox": [ + 130, + 283, + 482, + 427 + ], + "type": "text", + "content": " on mIoU and " + }, + { + "bbox": [ + 130, + 283, + 482, + 427 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 130, + 283, + 482, + 427 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 435, + 312, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 312, + 448 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 312, + 448 + ], + "type": "text", + "content": "5 Discussion and Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 450, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 592 + ], + "type": "text", + "content": "We introduced CPM, a new audio-visual training method designed for the Maskformer based framework to enhance bipartite matching stability and improve the efficacy of cross-modal attention for audio-visual segmentation. We proposed a class-conditional prompting learning strategy that combines class-agnostic queries with class-conditional queries, sampled from our iteratively updated generative model. The generated class-conditional queries are utilised to probe both the magnitude spectrogram and the image feature map, aiming to remove off-the-screen noise and bypass bipartite matching to produce a more stable learning process. Lastly, we extend the class-conditional queries to a new prompting-based audio-visual contrastive learning to explicitly constrain the cross-modal representations. SOTA results on AVS benchmarks suggest that CPM can be a valuable resource for future AVS research." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Limitations and future work. We acknowledge that the current adaptation of stereo audio into the transformer-based method has limitations, as it encodes positional and semantic information jointly within the transformer block. This contrasts with the typical approach of separately encoding these two types of information in the transformer-based framework. Our future work will concentrate on optimizing this framework through the integration of spatial reasoning." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "Y. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 176 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 176 + ], + "type": "text", + "content": "1. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 176, + 480, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 480, + 208 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 480, + 208 + ], + "type": "text", + "content": "2. Chen, H., Xie, W., Afouras, T., Nagrani, A., Vedaldi, A., Zisserman, A.: Localizing visual sounds the hard way. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16867-16876 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 209, + 480, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 480, + 241 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 480, + 241 + ], + "type": "text", + "content": "3. Chen, J., Lu, J., Zhu, X., Zhang, L.: Generative semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7111-7120 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 243, + 480, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 480, + 275 + ], + "type": "text", + "content": "4. Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 801-818 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 276, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 276, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 276, + 480, + 308 + ], + "type": "text", + "content": "5. Chen, T., Tan, Z., Gong, T., Chu, Q., Wu, Y., Liu, B., Lu, L., Ye, J., Yu, N.: Bootstrapping audio-visual segmentation by strengthening audio cues. arXiv preprint arXiv:2402.02327 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "type": "text", + "content": "6. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 342, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 480, + 364 + ], + "type": "text", + "content": "7. Chen, Y., Liu, Y., Wang, H., Liu, F., Wang, C., Carneiro, G.: A closer look at audio-visual semantic segmentation. arXiv e-prints pp. arXiv-2304 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 365, + 480, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 408 + ], + "type": "text", + "content": "8. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290–1299 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 409, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 409, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 138, + 409, + 480, + 441 + ], + "type": "text", + "content": "9. Cheng, B., Schwing, A., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems 34, 17864-17875 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 442, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 442, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 442, + 480, + 475 + ], + "type": "text", + "content": "0. Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society: series B (methodological) 39(1), 1-22 (1977)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 476, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 476, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 138, + 476, + 480, + 508 + ], + "type": "text", + "content": "1. Everingham, M., Eslami, S.A., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes challenge: A retrospective. International journal of computer vision 111, 98-136 (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 509, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 509, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 138, + 509, + 480, + 542 + ], + "type": "text", + "content": "2. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. International journal of computer vision 88, 303-338 (2010)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 543, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 543, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 138, + 543, + 480, + 575 + ], + "type": "text", + "content": "3. Gao, R., Grauman, K.: Co-separating sounds of visual objects. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3879-3888 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 576, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 576, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 576, + 480, + 597 + ], + "type": "text", + "content": "4. Gao, S., Chen, Z., Chen, G., Wang, W., Lu, T.: Avsegformer: Audio-visual segmentation with transformer. arXiv preprint arXiv:2307.01146 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 599, + 434, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 434, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 434, + 609 + ], + "type": "text", + "content": "5. Gray, R.: Vector quantization. IEEE Assp Magazine 1(2), 4-29 (1984)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 610, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 480, + 632 + ], + "type": "text", + "content": "6. Hao, D., Mao, Y., He, B., Han, X., Dai, Y., Zhong, Y.: Improving audio-visual segmentation with bidirectional generation. arXiv preprint arXiv:2308.08288 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "7. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 262, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 262, + 91, + 447, + 102 + ], + "type": "text", + "content": "Class-conditional Prompting Machine (CPM)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "18. Hu, X., Chen, Z., Owens, A.: Mix and localize: Localizing sound sources in mixtures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10483-10492 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "text", + "content": "19. Huang, S., Li, H., Wang, Y., Zhu, H., Dai, J., Han, J., Rong, W., Liu, S.: Discovering sounding objects by audio queries for audio visual segmentation. arXiv preprint arXiv:2309.09501 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "20. Jaegle, A., Borgeaud, S., Alayrac, J.B., Doersch, C., Ionescu, C., Ding, D., Koppula, S., Zoran, D., Brock, A., Shelhamer, E., et al.: Perceiver io: A general architecture for structured inputs & outputs. arXiv preprint arXiv:2107.14795 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "text", + "content": "21. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in neural information processing systems 33, 18661-18673 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "22. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "type": "text", + "content": "23. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13619-13627 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "24. Li, K., Yang, Z., Chen, L., Yang, Y., Xun, J.: Catr: Combinatorial-dependence audio-queried transformer for audio-visual video segmentation. arXiv preprint arXiv:2309.09709 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 390 + ], + "type": "text", + "content": "25. Li, T., Cao, P., Yuan, Y., Fan, L., Yang, Y., Feris, R.S., Indyk, P., Katabi, D.: Targeted supervised contrastive learning for long-tailed recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6918-6928 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "text", + "content": "26. Li, X., Wang, J., Xu, X., Peng, X., Singh, R., Lu, Y., Raj, B.: Towards robust audiovisual segmentation in complex environments with quantization-based semantic decomposition. arXiv preprint arXiv:2310.00132 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "27. Liang, C., Wang, W., Miao, J., Yang, Y.: Gmmseg: Gaussian mixture based generative semantic segmentation models. Advances in Neural Information Processing Systems 35, 31360-31375 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "text", + "content": "28. Liu, C., Ding, H., Jiang, X.: Gres: Generalized referring expression segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23592-23601 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "text", + "content": "29. Liu, C., Li, P., Qi, X., Zhang, H., Li, L., Wang, D., Yu, X.: Audio-visual segmentation by exploring cross-modal mutual semantics (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "30. Liu, C., Li, P., Zhang, H., Li, L., Huang, Z., Wang, D., Yu, X.: Bavs: Bootstrapping audio-visual segmentation by integrating foundation knowledge. arXiv preprint arXiv:2308.10175 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "type": "text", + "content": "31. Liu, J., Ju, C., Ma, C., Wang, Y., Wang, Y., Zhang, Y.: Audio-aware query-enhanced transformer for audio-visual segmentation. arXiv preprint arXiv:2307.13236 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "text", + "content": "32. Liu, J., Wang, Y., Ju, C., Zhang, Y., Xie, W.: Annotation-free audio-visual segmentation. arXiv preprint arXiv:2305.11019 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "33. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "34. Liu, X., Carrington, P., Chen, X., Pavel, A.: What makes videos accessible to blind and visually impaired people? In: Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. pp. 1-14 (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 101 + ], + "type": "text", + "content": "Y. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "35. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "text", + "content": "36. Mao, Y., Zhang, J., Xiang, M., Lv, Y., Zhong, Y., Dai, Y.: Contrastive conditional latent diffusion for audio-visual segmentation. arXiv preprint arXiv:2307.16579 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "text", + "content": "37. Mao, Y., Zhang, J., Xiang, M., Zhong, Y., Dai, Y.: Multimodal variational autoencoder based audio-visual segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 954–965 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "38. Martin, D.R., Fowlkes, C.C., Malik, J.: Learning to detect natural image boundaries using local brightness, color, and texture cues. IEEE transactions on pattern analysis and machine intelligence 26(5), 530-549 (2004)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 282 + ], + "type": "text", + "content": "39. Meng, D., Chen, X., Fan, Z., Zeng, G., Li, H., Yuan, Y., Sun, L., Wang, J.: Conditional detr for fast training convergence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3651-3660 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 304 + ], + "type": "text", + "content": "40. Mo, S., Morgado, P.: A closer look at weakly-supervised audio-visual source localization. arXiv preprint arXiv:2209.09634 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 336 + ], + "type": "text", + "content": "41. Mo, S., Morgado, P.: Localizing visual sounds the easy way. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVII. pp. 218-234. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "42. Mo, S., Tian, Y.: Audio-visual grouping network for sound localization from mixtures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10565–10574 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 392 + ], + "type": "text", + "content": "43. Mo, S., Tian, Y.: Av-sam: Segment anything model meets audio-visual localization and segmentation. arXiv preprint arXiv:2305.01836 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 413 + ], + "type": "text", + "content": "44. Murray, M.M., Wallace, M.T.: The neural bases of multisensory processes. Frontiers in Neuroscience (2011)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 435 + ], + "type": "text", + "content": "45. Reynolds, D.A., et al.: Gaussian mixture models. Encyclopedia of biometrics 741(659-663) (2009)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "46. Senocak, A., Ryu, H., Kim, J., Oh, T.H., Pfister, H., Chung, J.S.: Sound source localization is all about cross-modal alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7777-7787 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "text", + "content": "47. Siam, M., Oreshkin, B.N., Jagersand, M.: Amp: Adaptive masked proxies for few-shot segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5249-5258 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 534 + ], + "type": "text", + "content": "48. Sun, Z., Cao, S., Yang, Y., Kitani, K.M.: Rethinking transformer-based set prediction for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3611-3620 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "text", + "content": "49. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "text", + "content": "50. Wang, W., Zhou, T., Yu, F., Dai, J., Konukoglu, E., Van Gool, L.: Exploring cross-image pixel contrast for semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7303-7313 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "51. Wang, Y., Liu, W., Li, G., Ding, J., Hu, D., Li, X.: Prompting segmentation with sound is generalizable audio-visual source localizer. arXiv preprint arXiv:2309.07929 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "52. Wu, J., Jiang, Y., Sun, P., Yuan, Z., Luo, P.: Language as queries for referring video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4974-4984 (2022)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 262, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 262, + 91, + 447, + 102 + ], + "type": "text", + "content": "Class-conditional Prompting Machine (CPM)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 304 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "53. Xu, P., Zhu, X., Clifton, D.A.: Multimodal learning with transformers: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "type": "text", + "content": "54. Yang, Q., Nie, X., Li, T., Gao, P., Guo, Y., Zhen, C., Yan, P., Xiang, S.: Cooperation does matter: Exploring multi-order bilateral relations for audio-visual segmentation. arXiv preprint arXiv:2312.06462 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "text", + "content": "55. Zhang, H., Li, F., Xu, H., Huang, S., Liu, S., Ni, L.M., Zhang, L.: Mp-former: Mask-piloted transformer for image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18074-18083 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "56. Zhou, J., Shen, X., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., et al.: Audio-visual segmentation with semantics. arXiv preprint arXiv:2301.13190 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "57. Zhou, J., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., Zhong, Y.: Audio-visual segmentation. In: European Conference on Computer Vision. pp. 386-403. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 304 + ], + "type": "text", + "content": "58. Zhou, T., Zhang, M., Zhao, F., Li, J.: Regional semantic contrast and aggregation for weakly supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4299-4309 (2022)" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 100 + ], + "type": "text", + "content": "Y. Chen et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CPT-VR_ Improving Surface Rendering via Closest Point Transform with View-Reflection Appearance/255265d5-d4ab-4d8b-8210-adf9209a31be_content_list.json b/2024/CPT-VR_ Improving Surface Rendering via Closest Point Transform with View-Reflection Appearance/255265d5-d4ab-4d8b-8210-adf9209a31be_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7a01e17878e4fb0c02ec6582eacee1101d4f3ee1 --- /dev/null +++ b/2024/CPT-VR_ Improving Surface Rendering via Closest Point Transform with View-Reflection Appearance/255265d5-d4ab-4d8b-8210-adf9209a31be_content_list.json @@ -0,0 +1,2035 @@ +[ + { + "type": "text", + "text": "CPT-VR: Improving Surface Rendering via Closest Point Transform with View-Reflection Appearance", + "text_level": 1, + "bbox": [ + 238, + 140, + 764, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhipeng Hu $^{1,3}$ , Yongqiang Zhang $^{1,3,*}$ , Chen Liu $^{2}$ , Lincheng Li $^{3,*}$ , Sida Peng $^{3}$ , Xiaowei Zhou $^{3}$ , Changjie Fan $^{1}$ , and Xin Yu $^{2}$", + "bbox": [ + 225, + 233, + 776, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ NetEase Fuxi AI Lab, Hangzhou, China {zhangyongqiang02,zphu,lilincheng,fanchangjie}@corp.netease.com", + "bbox": [ + 256, + 276, + 743, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 The University of Queensland, Queensland, Australia", + "bbox": [ + 316, + 305, + 684, + 318 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "chen.liu7@uqconnect.edu.au, xin.yu@uq.edu.au", + "bbox": [ + 331, + 320, + 671, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Zhejiang University, Hangzhou, China", + "bbox": [ + 367, + 333, + 635, + 345 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{pengsida,xwzhou}@zju.edu.cn", + "bbox": [ + 393, + 348, + 609, + 361 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Differentiable surface rendering has significantly advanced 3D reconstruction. Existing surface rendering methods assume that the local surface is planar, and thus employ linear approximation based on the Singed Distance Field (SDF) values to predict the point on the surface. However, this assumption overlooks the inherently irregular and non-planar nature of object surfaces in the real world. Consequently, the approximate points tend to deviate from the zero-level set, affecting the fidelity of the reconstructed shape. In this paper, we propose a novel surface rendering method termed CPT-VR, which leverages the Closet Point Transform (CPT) and View and Reflection direction vectors to enhance the quality of reconstruction. Specifically, leveraging the physical property of CPT that accurately projects points near the surface onto the zero-level set, we correct the deviated points, thus achieving an accurate geometry representation. Based on our accurate geometry representation, incorporating the reflection vector into our method can facilitate the appearance modeling of specular regions. Moreover, to enable our method to no longer be dependent on any prior knowledge of the background, we present a background model to learn the background appearance. Compared to previous state-of-the-art methods, CPT-VR achieves better surface reconstruction quality, even for cases with complex structures and specular highlights.", + "bbox": [ + 259, + 391, + 738, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Surface Rendering $\\cdot$ 3D Reconstruction $\\cdot$ Differentiable Rendering $\\cdot$ Closest Point Transform", + "bbox": [ + 259, + 695, + 738, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 744, + 374, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Differentiable rendering plays a crucial role in 3D content creation, especially showing greater promise in multi-view 3D reconstruction and text-to-3D generation. Recent advancements have utilized differentiable rendering for attaining", + "bbox": [ + 212, + 772, + 785, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding authors.", + "bbox": [ + 230, + 825, + 403, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e7a287aca156de308693b5480b9aa7127e69a8ee7c9698378b1fddc5d3ed5121.jpg", + "image_caption": [ + "(a) IDR" + ], + "image_footnote": [], + "bbox": [ + 225, + 162, + 392, + 243 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/723c411c6796cdef18ca737e0debe2f4a918269a926ad53d9149b74c335df7e5.jpg", + "image_caption": [ + "$S_0$ : zero-level set", + "v: the view ray", + "$\\mathbf{x}^{\\prime}$ : the linear projection along the ray $\\mathbf{v}$" + ], + "image_footnote": [], + "bbox": [ + 233, + 247, + 364, + 299 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aa034c72e6a9d6884d0a6d31ff401f50e92f4fa256aa890020c409ef469a57c5.jpg", + "image_caption": [ + "(b)DMTet" + ], + "image_footnote": [], + "bbox": [ + 406, + 162, + 573, + 243 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/927bf82a786c33bc9f3df6e502fe3e3b105a47b992f2d59cca3832fcce6d54d7.jpg", + "image_caption": [ + "x: the sampled point by sphere tracing $d$ : the SDF value of x, n: the normal vector of x", + "$S_0$ : zero-level set", + "$\\mathbf{x}_i, \\mathbf{x}_j$ : vertices of an edge where the SDF", + "changes sign", + "$d_{i},d_{j}$ :SDF values of $\\mathbf{x}_i,\\mathbf{x}_j$", + "$\\mathbf{x}_i^{\\prime},\\mathbf{x}_j^{\\prime}$ : local deformed vertices" + ], + "image_footnote": [], + "bbox": [ + 411, + 243, + 566, + 292 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/80b2f5e2512588da28f2dc9ddee4384a6911a09974e5e2cf6944d77a7262f118.jpg", + "image_caption": [ + "(c) CPT-Ours" + ], + "image_footnote": [], + "bbox": [ + 594, + 162, + 763, + 243 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c5abb21e4df1684c16d3070c03b02aa390d3103a04d06826b26328c452e82f9c.jpg", + "image_caption": [ + "$\\mathbf{x}_{ij}^{\\prime}$ : linear interpolation of $\\mathbf{x}_i^{\\prime},\\mathbf{x}_j^{\\prime}$ by $d_{i},d_{j}$", + "$S_0$ : zero-level set", + "$\\mathbf{x}_i, \\mathbf{x}_j$ : vertices of an edge where the SDF changes sign", + "$d_{i},d_{j},d_{ij}$ : SDF values, $\\mathbf{n}_{ij}$ : the normal vector of $\\mathbf{x}_{ij}$", + "$\\mathbf{x}_{ij}$ : linear interpolation of $\\mathbf{x}_i, \\mathbf{x}_j$ by $d_i, d_j$", + "$\\mathbf{x}_{ij}^{\\prime}$ : the projected vertex of $\\mathbf{x}_{ij}$ by CPT", + "Fig. 1: Illustration of the surface point sampling methods and their corresponding construction results. (a) IDR [34] first employs a sphere tracing method to identify the sampled point $x$ and then leverages the SDF value as well as the normal vector to obtain the approximate point $x'$ . (b) DMTet [19] detects surfaces through the observation of sign changes in the SDF at the vertices of a tetrahedron. Then, it employs linear interpolation to approximate the points $x_{ij}'$ on the zero-level set. Compared with them, our method (c) utilizes CPT to project the interpolated points $x_{ij}$ on the zero-level set, and thus attain the transformed points $x_{ij}'$ . The approximate points $x'$ of IDR and $x_{ij}'$ of DMTet are deviated from the zero-level set, while $x_{ij}'$ obtained from our method exactly on the object surface. This implies our method is able to obtain a more accurate geometry representation and gradient backpropagation." + ], + "image_footnote": [], + "bbox": [ + 594, + 244, + 741, + 304 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the Singed Distance Field (SDF) and object appearance. These methods can be divided into two types: volume rendering and surface rendering. In volume rendering, SDF is first converted into density values that are then rendered via the NeRF framework [17, 38]. While these methods consistently generate reliable geometry reconstruction, the dense points sampling along the ray increases the rendering time cost. Conversely, surface rendering methods only sample one point along the ray, markedly optimizing rendering speeds.", + "bbox": [ + 212, + 518, + 787, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing surface rendering methods, such as IDR (Implicit Differentiable Renderer) [34] and DMTet (Deep Marching Tetrahedra) [19] achieve computational efficiency by assuming the object surface can be considered linear or at least can be approximated as linear within a local scope. Based on this assumption, they employ linear interpolation to determine the points at specific locations on the surface. As shown in Fig. 1, IDR employs sphere tracing [7] to identify points close to the surface along a ray and then relies on the SDF values and the normal vector to calculate the intersection points between rays and the surface. DMTet identifies the surfaces by observing changes in the sign of the SDF at the vertices of a tetrahedron and then leverages linear interpolation to approximate the surface points. However, as demonstrated in Fig. 1, their approximate points always deviate from the object surface (i.e., zero-level set), since the object surface is mostly irregular and non-planar. Hence, modeling geometry representations based on these deviated points would further lead to inaccurate gradient back-", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Hu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "propagation, resulting in the reconstructed surfaces losing a lot of details, as illustrated by the reconstruction results in Fig. 1 (a) and (b).", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the above problem, we introduce a novel surface rendering approach, termed CPT-VR, which is grounded in the Closet Point Transform algorithm and augmented with View-Reflection appearance modeling. In a nutshell, in the geometry representation modeling process, we conduct the Closet Point Transform (CPT) for deviated points to enhance the geometry representation. In the appearance modeling process, we intend to incorporate the View and Reflection vectors and thus attain better appearance representation for cases with specular highlights. To the best of our knowledge, we are the first attempt to correct the deviated points via CPT in the surface rendering task.", + "bbox": [ + 212, + 180, + 785, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To be specific, in the geometry representation modeling process, we first devise a neural SDF network to compute SDF values for the entire 3D space. Based on the SDF values, we employ the Marching Cubes [14] to attain the vertices close to the surface of the object. As suggested in Fig. 1 (c), most of the obtained vertices are deviated from the object's surface rather than on it. Hence, we leverage CPT to calculate the closest point on the object surface for each deviated vertex, thus obtaining accurate and differentiable vertices. Subsequently, we utilize a feature-based rasterizer to render this triangle mesh in each view. In this fashion, we obtain the corresponding 3D position, geometry feature, and normal vector of each pixel in an image.", + "bbox": [ + 212, + 319, + 787, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the appearance representation modeling process, we focus on improving the geometry quality for specular highlights. Existing methods such as IDR [34] and NeuS2 [25] rely solely on the view direction to capture the surface information of objects, which cannot handle the object with highlights well. Ref-NeRF [21] demonstrates that replacing the view direction with the reflection direction can mitigate the impact of highlights. However, since the reflection vector calculated based on the deviated points is inaccurate, directly incorporating these reflection vectors into existing surface rendering methods will lead to significant backpropagation errors. Consequently, the model will not only be unable to handle the specular highlights but will also diminish its capability to reconstruct non-specular areas. In contrast, benefiting from our accurate point sampling on the zero-level set based on CPT, incorporating view-reflection vectors in our framework can facilitate the appearance modeling of specular regions.", + "bbox": [ + 212, + 474, + 787, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, we found that existing surface rendering methods [1,22,34] require prior background knowledge, such as foreground masks, or the background image, to precisely process and render foreground objects, which constrains their flexibility and applicability in various scenes. To solve this problem, we present a background model to obtain the background appearance representation. Then we attain the final appearance modeling by blending the foreground and background appearance. Extensive experiments conducted on DTU [9] and BlendedMVS [31] demonstrate that our method achieves state-of-the-art geometry quality and surface rendering. Additionally, we also illustrate that our method is effective and reliable in cases with complex structures and specular highlights. In summary, our contributions are three-fold:", + "bbox": [ + 212, + 672, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Improving Surface Rendering via CPT with View-Reflection Appearance", + "bbox": [ + 246, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a novel differentiable surface rendering method CPT-VR which shows superior performance in reconstructing the detailed structure of objects. With the CPT algorithm, we can address the approximate point deviation issue caused by linear interpolation.", + "- We incorporate the view and reflection vectors to model the foreground appearance representation, facilitating our method to be robust against specular highlights.", + "- We develop a simple background model to generate the background appearance representation, enabling our method to no longer be dependent on any prior background knowledge." + ], + "bbox": [ + 225, + 146, + 782, + 300 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 309, + 387, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Surface Rendering with SDF. IDR [34] introduces a system that learns surfaces in a self-supervised manner. It first identifies points close to the surface via sphere tracing and then finds approximate intersection points on the surface by projecting them along the view ray. Although IDR can learn the 3D shape and appearance from images, it relies heavily on accurate masks. To mitigate this issue, MVsDF [37] supervises the SDF network with depth maps generated by Vis-MVSNET [36]. It also uses relaxed masks generated by probability maps to learn a surface indicator. RegSDF [35] further introduces the Hessian regularization and the minimal surface constraint. Some works [1, 22] use specialized reparameterization techniques and design weighting functions along a ray based on SDF values, thus they can propagate gradients to intermediate points from sphere tracing. These reparameterization-based methods are similar to methods based on volume rendering, but they still need a given background: either black or a pre-defined environment map.", + "bbox": [ + 212, + 340, + 787, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rasterization-based Rendering. Recently, there have been attempts to utilize the differentiable rasterizer for surface rendering. NVDIFFREC [18] and NVDIFFRECmc [8] adapt DMTet [19] to extract a mesh from an implicit SDF and then render the mesh by the differentiable rasterizer Nvdiffrast [10]. While these methods are efficient, their reliance on the linear approximation in DMTet limits their ability to recover shapes with complex structures. Meanwhile, NIE [16] propagates neural implicit surfaces according to a flow field from a certain energy function. ENS [23] combines neural deformation fields with rasterization rendering, allowing deform an initial shape into a target surface progressively. Moreover, NDS [29] and FastMESH [39] focus on optimizing a rough mesh derived from priors through differentiable rasterization.", + "bbox": [ + 212, + 553, + 787, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Volume Rendering of SDF. Several methods [2, 20, 24, 32] have been developed for learning SDF through volume rendering. VolSDF [32] and NeuS [24] leverage Laplace or Logistic cumulative distribution function of SDF to represent the volume density and sample dense points along a ray based on SDF values. Based on these methods, some works [2, 3, 12, 26–28, 40, 41] aim to further enhance the geometry details. Among them, Geo-Neus [3] and TUVR [40] improve SDF accuracy by introducing additional supervision from point clouds. On the training efficiency front, Voxurf [30] and Vox-Surf [11] speed up the training process", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Hu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "through voxel-based representations. NeuS2 [25] speeds up NeuS [24] by simplifying the second-order derivatives calculation of the SDF network constructed based on a multi-resolution hash encoding.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hybrid Volume-rasterization Rendering. Recently, some rendering methods have been developed based on a hybrid volume-rasterization way. BakedSDF [33] first learns the SDF and appearance of a scene based on VolSDF [32] and transforms the SDF into a mesh. Then the transformed mesh is refined using a rasterizer pipeline. VMesh [6] takes a comprehensive approach by initially training a combination of a neural SDF and a density field through volume rendering. Then the mesh undergoes further refinement by DMTet [19] and a differentiable rasterizer [10], ensuring accurate surface representation.", + "bbox": [ + 212, + 191, + 784, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 335, + 330, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 366, + 372, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Signed Distance Field (SDF) is a mathematical function $f$ that measures the shortest distance from a point $\\mathbf{x}$ to an object surface, indicating whether the point is outside (positive distance) or inside (negative distance) the surface. The surface is identified by the zero-level set $S$ , where SDF equals to zero, providing a concise description of the object's surface:", + "bbox": [ + 212, + 390, + 784, + 465 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} = \\left\\{\\mathbf {x} \\in \\mathbb {R} ^ {3} | f (\\mathbf {x}) = 0 \\right\\}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 478, + 784, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Neural SDF denotes a neural network that models SDFs for 3D objects. In this work, we employ SDFnet which leverages the Multilayer Perceptrons (MLPs) enhanced with a multi-resolution hash encoding to model SDF [25]:", + "bbox": [ + 212, + 503, + 782, + 551 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n(d, \\mathbf {z}) = f _ {\\theta} (\\mathbf {x}, h _ {\\Omega} (\\mathbf {x})). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 561, + 784, + 577 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $d$ is the predicted SDF value, $\\mathbf{z}$ signifies the predicted geometry feature, $f_{\\theta}$ denotes an MLP parameterized by weights $\\theta$ , and $h_{\\Omega}(\\mathbf{x})$ is the multi-resolution hash encoding of the input point $\\mathbf{x}$ . Moreover, the normal vector $\\mathbf{n}$ of $\\mathbf{x}$ is derived as the normalized gradient of the implicit surface:", + "bbox": [ + 212, + 588, + 784, + 648 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {n} = \\nabla_ {\\mathbf {x}} d / \\| \\nabla_ {\\mathbf {x}} d \\|. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 439, + 662, + 784, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To ensure training stability, our SDFnet employs an optimization schedule that progressively incorporates the level of detail and the numerical gradients [12].", + "bbox": [ + 212, + 686, + 782, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Geometry Representation Modeling", + "text_level": 1, + "bbox": [ + 215, + 739, + 558, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Geometry Representation via SDF. As suggested in Fig. 2, we first leverage our SDFnet to compute the SDF values for the entire 3D space. Based on SDF values, we utilize the isosurface algorithm Marching Cubes to extract and represent the triangle mesh $\\mathcal{G} = (\\mathcal{V},\\mathcal{F})$ , where $\\mathcal{V}$ and $\\mathcal{F}$ indicate the vertices and faces respectively. However, this method assumes that the surface changes", + "bbox": [ + 212, + 763, + 784, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Improving Surface Rendering via CPT with View-Reflection Appearance", + "bbox": [ + 246, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/26875512fec8bc781bf49e5624a3f02723133c86a23d48a02078e4b3b5828188.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 152, + 774, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5dc2c214efe8bfac2d39fd250aa4ab30e32109455e5872013c11439d5411f6e5.jpg", + "image_caption": [ + "Fig. 2: Illustration of our proposed method CPT-VR. In (a), we first extract the mesh via Marching Cubes, which does not involve gradient propagation. Then we leverage CPT to project the approximate points on the zero-level set (b). Subsequently, we render the transformed mesh via a rasterizer in (c). Specifically, we incorporate the view-reflection vectors into the foreground appearance modeling stage via RGBNet. Additionally, we employ the Background Model to attain the background information. Finally, we integrate the foreground and background appearance to generate the rendering results. We adopt four supervisions to optimize these neural networks." + ], + "image_footnote": [], + "bbox": [ + 225, + 236, + 774, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "linearly inside the voxel, overlooking the complex convex structure of the object surface. Hence, the linearly interpolated vertices tend to deviate from the zero-level set and the triangle mesh may not accurately reflect the complex structure of the object surface.", + "bbox": [ + 212, + 500, + 787, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Closest Point Transform. To project these deviated vertices onto the zero-level set, we devise the closet set transform (CPT) to SDF. Specifically, CPT first utilizes the norm vector $\\mathbf{n}$ within the signed distance field to identify the projected direction of a deviated point. It then moves the deviated point along the opposite direction of $\\mathbf{n}$ by a distance $d$ equal to its SDF value. Mathematically, for a vertex $\\mathbf{x} \\in \\mathcal{V}$ , this transformation process is expressed as follows:", + "bbox": [ + 212, + 561, + 787, + 652 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} ^ {\\prime} = \\mathbf {x} - d \\mathbf {n}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 452, + 669, + 785, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After that, we obtain the transformed mesh $\\mathcal{G}' = (\\mathcal{V}',\\mathcal{F})$", + "bbox": [ + 215, + 693, + 629, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Feature-based Rasterisation", + "text_level": 1, + "bbox": [ + 215, + 734, + 493, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After obtaining the transformed mesh, we first employ a feature-based differentiable rasterizer Nvdiffrast [10] to render triangle meshes in each view. Specifically, as suggested in Fig. 2 (c), for a pixel $p(x,y)$ , its projected primitive triangle face within the mesh is represented as $(\\mathbf{x}_0',\\mathbf{x}_1',\\mathbf{x}_2')$ . The rasterizer computes the barycentric coordinates $u,v$ , depth $D_{p}$ , and visible mask $M_{p}$ for this pixel as", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Hu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "follows:", + "bbox": [ + 215, + 146, + 272, + 159 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n(u, v, D _ {p}, M _ {p}) = \\mathcal {R} \\left(\\mathcal {G} ^ {\\prime}, p (x, y), K, R, \\mathbf {t}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 162, + 784, + 178 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{R}$ denotes the rasterizer, $\\mathcal{G}'$ represents the transformed mesh, and $K$ and $R, \\mathbf{t}$ are the camera's intrinsic and extrinsic parameters, respectively.", + "bbox": [ + 215, + 183, + 785, + 213 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Subsequently, each projected vertex $\\mathbf{x}'$ from the set $\\mathcal{V}'$ is input into SDFnet. SDFnet outputs $(d',\\mathbf{z}',\\mathbf{n}')$ , encapsulating the distance, geometric feature, and the normal vector associated with each vertex. These outputs are then interpolated for a pixel $p(x,y)$ based on its barycentric coordinates $u,v$ :", + "bbox": [ + 214, + 214, + 782, + 273 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {p} = u \\mathbf {x} _ {0} ^ {\\prime} + v \\mathbf {x} _ {1} ^ {\\prime} + (1 - u - v) \\mathbf {x} _ {2} ^ {\\prime},\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 287, + 617, + 301 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {p} = u \\mathbf {z} _ {0} ^ {\\prime} + v \\mathbf {z} _ {1} ^ {\\prime} + (1 - u - v) \\mathbf {z} _ {2} ^ {\\prime}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 303, + 784, + 316 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {n} _ {p} = u \\mathbf {n} _ {0} ^ {\\prime} + v \\mathbf {n} _ {1} ^ {\\prime} + (1 - u - v) \\mathbf {n} _ {2} ^ {\\prime},\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 319, + 619, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $(\\mathbf{x_p},\\mathbf{z_p},\\mathbf{n_p})$ denote the interpolated features.", + "bbox": [ + 215, + 343, + 584, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Appearance Representation Modeling", + "text_level": 1, + "bbox": [ + 215, + 380, + 573, + 395 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this work, we divide the appearance modeling process into two parts:", + "bbox": [ + 215, + 405, + 733, + 419 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Foreground Appearance Modeling based on View-Reflection Vectors. To better capture the appearance of objects with specular highlights, we incorporate the view direction and reflection direction into the foreground modeling process. Specifically, we employ the RGBnet $g_{\\gamma}^{f}$ , a Multilayer Perceptron (MLP) with weights $\\gamma$ to determine the foreground color. RGBnet takes the interpolated 3D features $(\\mathbf{x_p}, \\mathbf{z_p}, \\mathbf{n_p})$ , view direction $\\mathbf{v}_p$ , and reflection direction $\\mathbf{r}_p$ as input, and outputs the RGB color $\\mathbf{c}_p^f$ for the foreground. The process is formulated as:", + "bbox": [ + 215, + 419, + 782, + 526 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {p} ^ {f} = g _ {\\gamma} ^ {f} \\left(\\mathbf {x} _ {\\mathbf {p}}, \\mathbf {z} _ {\\mathbf {p}}, \\mathbf {n} _ {\\mathbf {p}}, Y (\\mathbf {v} _ {\\mathbf {p}}), Y (\\mathbf {r} _ {\\mathbf {p}})\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 379, + 539, + 784, + 558 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $Y(\\cdot)$ denotes Spherical Harmonics encoding. Here, the reflection direction $\\mathbf{r_p}$ is defined as the mirror vector of the view direction $\\mathbf{v_p}$ relative to the normal vector $\\mathbf{n_p}$ , calculated as:", + "bbox": [ + 215, + 561, + 782, + 608 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {r} _ {\\mathbf {p}} = 2 \\left(- \\mathbf {v} _ {\\mathbf {p}} \\cdot \\mathbf {n} _ {\\mathbf {p}}\\right) \\mathbf {n} _ {\\mathbf {p}} + \\mathbf {v} _ {\\mathbf {p}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 621, + 784, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Spherical harmonics are orthogonal bases on the unit sphere, and their mathematical definition and physical properties make them ideal for encoding direction frequencies compared to 3D space (xyz) frequencies.", + "bbox": [ + 215, + 646, + 782, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Background Appearance Modeling. To enable the model to reconstruct objects without any background priors, we present a one-point background model composed of a Multilayer Perceptron (MLP) $g_{\\pi}^{b}$ with weights $\\pi$ , alongside an additional multi-resolution hash encoding $h_{\\Omega}^{b}$ . The modeling process of the background appearance $\\mathbf{c}_p^b$ is formulated as:", + "bbox": [ + 214, + 691, + 782, + 768 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {p} ^ {b} = g _ {\\pi} ^ {b} \\left(\\mathbf {x} _ {p} ^ {b}, h _ {\\Omega} ^ {b} \\left(\\mathbf {x} _ {p} ^ {b}\\right), Y \\left(\\mathbf {v} _ {p}\\right)\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 780, + 784, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_p^b$ represents the background sampling point, determined by the camera location $\\mathbf{x}_o$ , the veiw direction $\\mathbf{v}_p$ and the far distance $F_{p}$ of the sampling", + "bbox": [ + 215, + 809, + 782, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Improving Surface Rendering via CPT with View-Reflection Appearance", + "bbox": [ + 246, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "boundary. $\\mathbf{x}_p^b$ is mathematically expressed as:", + "bbox": [ + 215, + 145, + 544, + 164 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {p} ^ {b} = \\mathbf {x} _ {o} + 2 F _ {p} \\mathbf {v} _ {p}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 169, + 785, + 189 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Following NeuS [24], the foreground is normalized within a unit sphere centered at $\\mathbf{0}$ . As indicated in Fig. 3, the far distance $F_{p}$ is calculated as:", + "bbox": [ + 214, + 193, + 782, + 224 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nm = - \\mathbf {x} _ {o} \\cdot \\mathbf {v} _ {p}; F _ {p} = m + 1. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 231, + 785, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This approach facilitates a more versatile background representation, compensating for the absence of masks.", + "bbox": [ + 214, + 253, + 782, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Appearance Integration. Finally, we integrate the background appearance $\\mathbf{c}_p^b$ and foreground appearance $\\mathbf{c}_p^f$ to attain the rendered color, defined as:", + "bbox": [ + 214, + 284, + 517, + 345 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {p} = M _ {p} \\mathbf {c} _ {p} ^ {f} + (1 - M _ {p}) \\mathbf {c} _ {p} ^ {b}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 256, + 349, + 513, + 371 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall, our method enables surface rendering of objects without any known background information, offering greater flexibility and adaptability.", + "bbox": [ + 214, + 375, + 514, + 436 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ba50007b33ae024c00c2145e20c314139d4244190f6134a11800937f98841ac7.jpg", + "image_caption": [ + "Fig.3: Illustration of the one-point background sampling." + ], + "image_footnote": [], + "bbox": [ + 527, + 280, + 787, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Training Objectives", + "text_level": 1, + "bbox": [ + 215, + 453, + 426, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Color loss. We optimize the CPT-VR model by ensuring the colors it generates match real image colors as closely as possible. This is done by calculating the L1 error, which is the average difference between the colors $\\hat{\\mathbf{c}}_i$ produced by the model and the ground-truth colors $\\mathbf{c}_i$ for all pixels $P$ :", + "bbox": [ + 214, + 474, + 782, + 536 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C} = \\frac {1}{P} \\sum_ {i} ^ {P} \\| \\hat {\\mathbf {c}} _ {i} - \\mathbf {c} _ {i} \\| _ {1}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 541, + 785, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Regularization of SDFnet. A well-defined neural SDF exhibits Lipschitz continuity [13] and its gradient adheres to the Eikonal equation [4], ensuring smoothness and stability of the generated field. A valid SDF field holds the property that the normal of a point near the surface and the normal of its projected point remain consistent. For effective closest point transform, we directly constrain our SDFnet based on this normal consistency property [15]. Specifically, it aligns the normal of a deviated point $\\mathbf{n}$ with the normal $\\mathbf{n}'$ of its projected point, enhancing stability for surface rendering via the closest point transform of SDF. Mathematically, it is represented by:", + "bbox": [ + 214, + 587, + 787, + 724 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {n c} = 1 - \\mathbf {n} \\cdot \\mathbf {n} ^ {\\prime}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 439, + 729, + 785, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mask Loss. If the mask of the foreground region is available, we can further refine the rendering effect of CPT-VR by including a mask loss. The mask constraint is formulated as:", + "bbox": [ + 214, + 753, + 782, + 797 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {M} = \\frac {1}{P} \\sum_ {i} ^ {P} \\left(\\hat {M} _ {i} - M _ {i}\\right) ^ {2}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 801, + 785, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Hu et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where the $\\hat{M}_i$ and the $M_{i}$ represent the predicted mask and the ground truth mask, respectively.", + "bbox": [ + 212, + 145, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Depth loss. Inspired by previous works [37] and [35], we utilize depth maps generated by the pre-trained network CasMVSNet [5] to guide our model. We minimize the squared difference (MSE) between the rendered depths $(\\hat{D}_i)$ and those from the pre-trained network $(D_i)$ , weighted by the confidence $prob_i$ from CasMVSNet, as follows:", + "bbox": [ + 212, + 176, + 784, + 252 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D} = \\frac {1}{P} \\sum_ {i} ^ {P} p r o b _ {i} \\left(\\hat {D} _ {i} - D _ {i}\\right) ^ {2}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 262, + 784, + 301 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where the confidence ranges from 0 to 1, and depths considered unreliable are given a confidence of 0.", + "bbox": [ + 212, + 309, + 782, + 338 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Total loss. Our final metric combines all these aspects, adjusting their importance with specific weights $\\lambda$ :", + "bbox": [ + 212, + 339, + 782, + 369 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {r g b} \\mathcal {L} _ {r g b} + \\lambda_ {n c} \\mathcal {L} _ {n c} + \\lambda_ {M} \\mathcal {L} _ {M} + \\lambda_ {D} \\mathcal {L} _ {D}. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 351, + 380, + 784, + 395 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Here, $\\lambda_{rgb}$ , $\\lambda_{nc}$ , $\\lambda_M$ , and $\\lambda_D$ are empirically set to 3, 0.01, 1, and 1, respectively.", + "bbox": [ + 214, + 402, + 779, + 420 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 440, + 375, + 458 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 214, + 469, + 444, + 484 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation details. For the DTU dataset, we train CPT-VR for 10,000 iterations with a batch size of one image, targeting a Marching Cubes resolution of $256^3$ . In each training batch, we render one view at a resolution of 800 $\\times$ 600 pixels. Note that, when we only utilize depth loss and color loss for supervision, we employ a higher Marching Cubes resolution of $512^3$ and render images at $1600 \\times 1200$ pixels. For BlendedMVS, CPT-VR is trained for 20,000 iterations with a batch size of one image and a Marching Cubes resolution of $256^3$ , rendering images at $768 \\times 576$ pixels. Moreover, the initial learning rate of our model is set at 0.005, with a final epoch decay factor of 0.1 for DTU and 0.005 for BlendedMVS. More details about the architecture and the isosurface configuration are provided in the supplementary material.", + "bbox": [ + 212, + 491, + 784, + 657 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. All experiments are conducted on DTU [9] and BlendedMVS [31]. Objects in DTU are collected in a controlled lab setting, with ground truth data that includes images, masks, and point clouds. For our training and evaluation, we leverage 15 objects identified by IDR with challenges such as specular highlights, delicate structures, and areas hidden from view. Meanwhile, BlendedMVS encompasses a diverse range of objects and scenes captured in natural settings. From this dataset, we select 7 difficult objects as NeuS [24] and the 'Bull' (with specular highlights) to further assess the model performance in reconstructing objects against complex backgrounds.", + "bbox": [ + 212, + 657, + 784, + 794 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 214, + 814, + 460, + 829 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Improving Surface Rendering via CPT with View-Reflection Appearance", + "bbox": [ + 246, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Analysis on DTU. We benchmark our method against the state-of-the-art methods on the DTU dataset. DTU dataset contains 128 scans captured in the laboratory. In this work, we evaluate the surface mesh using the same set of scans as in [34,37]. As denoted in Tab. 2, our method consistently outperforms other surface rendering methods on most scans. Note that, our method achieves the lowest", + "bbox": [ + 212, + 146, + 491, + 311 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "value in mean Chamfer Distance (0.59), which is $16.9\\%$ lower than the second-best surface rendering method RegSDF (0.71). Since SDFnet in our method has a similar structure to it in NeuS2, we also compare NeuS2 and recent volume rendering methods [26, 41] with CPT-VR. As demonstrated in Tab. 2, whether the two methods incorporate background information from the object mask $(\\mathrm{w} / \\mathrm{M})$ or use our proposed one-point background model $(\\mathrm{w} / \\mathrm{BG1})$ to attain background information, our method surpasses NeuS2 on the vast majority of ScanIDs.", + "bbox": [ + 212, + 311, + 787, + 419 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 1: Comparison of the average rendering time on DTU. The rendered image resolution of all methods is $1600 \\times 1200$ , and the resolution of Marching Cubes is $512^{3}$ . The rendering time of our method is significantly lower than that of others.", + "bbox": [ + 501, + 136, + 787, + 220 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/9733f11fd782c4f65e25275e97947de2095f62c5b1e84b289f6509199817ed1f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodRendering Time (s)
IDR27.233
NeuS2 (w /M)1.979
Ours (w/ M)0.057
Ours (w/ BG1)0.104
", + "bbox": [ + 529, + 220, + 756, + 306 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 2: Quantitative results on DTU dataset. Here, we evaluate all methods on Chamfer Distance (CD). The lower the CD value, the better the performance. For a fair comparison, we also present the results of the volume rendering method NeuS2 since our SDFnet has a similar structure to it. \"w/ M\" means providing the prior background knowledge by masks, while \"w/ BG1\" means employing our one-point background model without any prior background knowledge. The best and second-best results are highlighted by different colors. The results illustrated that CPT-VR attains better performance among most ScanIDs.", + "bbox": [ + 212, + 439, + 787, + 555 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/cb6c3a5708e2f893f400baa119114ceb9dc8f15203adcc22cdb1adc09a4394fe.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ScanID243740556365698397105106110114118122Mean
DMTet [19]1.841.731.050.611.661.370.681.283.821.070.911.150.370.540.551.24
IDR [34]1.631.870.630.481.040.790.771.331.160.760.670.900.420.510.530.90
MVSDF [37]0.831.760.880.441.110.900.751.261.021.350.870.840.340.470.460.89
FastMESH [39]0.651.480.570.401.480.770.560.860.840.940.720.810.520.490.540.77
RegSDF [35]0.591.410.630.421.340.620.590.890.911.020.600.590.290.400.380.71
HF-NeuS [26] (w/ M)1.111.280.610.470.970.680.621.340.910.730.531.820.380.540.510.83
LOD-NeuS [41] (w/ M)0.650.910.370.481.050.870.821.220.950.690.561.300.420.580.570.76
NeuS2 [25] (w/ M)0.560.760.490.370.920.710.761.221.080.630.590.890.400.480.550.70
NeuS2 [25] (w/ BG1)0.690.860.830.341.050.680.651.051.120.670.581.410.370.500.520.75
Ours (w/ M)0.550.720.350.380.850.590.540.810.830.670.530.590.320.380.370.56
Ours (w/ BG1)0.430.730.360.320.930.610.610.891.020.680.480.730.320.370.370.59
", + "bbox": [ + 215, + 560, + 802, + 734 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Moreover, we also measure the rendered results with PSNR to the ground truths. As indicated in Tab. 3, with the introduction of background information (BG1: $\\pmb{x}$ ), our method achieves the highest PSNR values on most ScanIDs. Additionally, it can be observed that when using the one-point background model, both NeuS2 and CPT-VR are able to achieve similar or better performance compared to leveraging background information from masks (BG1: $\\checkmark$ ).", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Hu et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e3ed2bd4c727a0763f2888e06485bd2ec24570969cb2046824255c7c74e804f6.jpg", + "table_caption": [ + "Table 3: Quantitative comparison of 2D view synthesis on DTU dataset. We measure all methods based on the metric PSNR. A higher PSNR indicates better synthesis quality. Here, B1 represents whether to introduce extra background knowledge into the method. $\\times$ means to leverage the background information from masks, while $\\checkmark$ indicates utilizing our one-point background model without introducing any background information. We color code the best and second-best results. Compared to NeuS2, our method achieves much better appearance reconstruction results." + ], + "table_footnote": [], + "table_body": "
ScanIDBG1243740556365698397105106110114118122Mean
NeuS2✗28.4427.1429.7029.6731.7527.8324.8431.2426.8630.5726.0528.9328.9827.8232.4828.82
NeuS2✓27.7124.9927.5128.8920.5530.1429.0127.7026.1631.0933.1030.6930.4832.7734.1028.99
Ours✗28.1527.3128.5429.5234.0429.8329.0035.9229.3233.0829.4531.1630.7730.5038.3831.00
Ours✓27.3025.5926.8827.7730.9130.0629.2828.0526.6828.9232.0330.1829.0234.0234.0229.38
", + "bbox": [ + 215, + 243, + 797, + 321 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/e2a1825a603f9b9eb28897618583d4c663f5d693f7a75285a9aa0eaa181d8b38.jpg", + "table_caption": [ + "Table 4: Quantitative comparison on BlendedMVS dataset. We measure the method on the metric Chamfer Distance (CD). The lower CD values indicate better performance. For a fair comparison, we adopt our one-point background model to NeuS2. Compared to NeuS2, CPT-VR attains better geometry reconstruction results among most of the categories." + ], + "table_footnote": [], + "table_body": "
MethodBearClockDogDurianManSculptureBullStoneMean
NeuS20.380.470.950.470.460.400.650.380.52
Ours0.350.450.680.480.460.390.450.380.45
", + "bbox": [ + 225, + 402, + 774, + 455 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Analysis on BlendedMVS. As illustrated in Tab. 4, we compare our method with NeuS2 on the eight selected categories with complex backgrounds. Since the scales are unknown in BlendedMVS, we scale the bounding box of each object to 100 units to measure their performance on Chamfer Distance (CD). CPT-VR achieves the best performance in the categories of 'Bear', 'Clock', 'Dog', 'Sculpture', and 'Bull', with the CD value for 'Man' and 'Stone' being on par with NeuS2, and only 0.01 behind NeuS2 for 'Durian'. Moreover, CPT-VR outperforms NeuS2 by $14.8\\%$ . These results demonstrate the advantage of our method in reconstructing objects within complex environments.", + "bbox": [ + 212, + 474, + 787, + 609 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Analysis of Method Efficiency. We evaluate the rendering time of IDR, NeuS2, and CPT-VR on DTU. The results in Tab. 1 suggest the rendering time of IDR (24.233s) and NeuS2 (1.97s) is lengthy. Compared with them, our method achieves real-time rendering under the cases without introducing background priors being about 19 times faster than NeuS2. This is primarily because NeuS2 utilizes the dense sampling method in the volume rendering process.", + "bbox": [ + 212, + 609, + 787, + 702 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 214, + 707, + 447, + 722 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Impact of Specular Highlights. Fig. 4 illustrates the qualitative results on the DTU dataset. We compare our CPT-VR(w/M) model with IDR [34], DMTet [19], MVsDF [37], and NeuS2 [25]. As shown by the close-up regions in Fig. 4, the specular regions constructed by other methods exhibit obvious indentations, while our method can construct these regions with high quality.", + "bbox": [ + 212, + 733, + 787, + 808 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Impact of Complex Backgrounds. Fig. 5 demonstrates the visual results of reconstructing objects with complex backgrounds. Since existing surface render-", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Improving Surface Rendering via CPT with View-Reflection Appearance", + "bbox": [ + 246, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "ing methods cannot handle the cases without any background prior, we specifically evaluate NeuS2 and our method on BlendedMVS. As the elliptical area demonstrated, our method can effectively model the geometry of foreground objects without introducing background information, particularly the details on the surface of objects. Moreover, the rendering results of our method closely resemble the input image.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/59960e7ac11b8e536f954b9cbb78f2ccc369a13f0cee48625ae01bd4e4627fcc.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparisons on objects in DTU with specular highlights. The close-ups in yellow highlight the specific regions for comparison. The geometry reconstruction results demonstrate that our method is robust against specular highlights." + ], + "image_footnote": [], + "bbox": [ + 246, + 263, + 756, + 373 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/594bbbeb7067c1830b4433ee94e0d1d2fa430c9536d6c0267e63678028ea101f.jpg", + "image_caption": [ + "Fig.5: Qualitative results of objects with complex background on BlendedMVS. We outline the close-up regions for detailed comparison. It is evident that without the mask guidance, our method also exhibits strong capabilities in modeling foreground objects, especially for the finer-grained surface structures." + ], + "image_footnote": [], + "bbox": [ + 250, + 488, + 759, + 670 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Impact of the Sampling Point Amount. Tab. 5 suggests the impact of varying the number of sampling points and their distances. The optimal result is achieved when sampling one point at $2F_{p}$ , and the $CD$ is 0.66. Conversely, sampling either too closely (at $F_{p}$ ) or too distantly (at $1000F_{p}$ ) causes the sampled points to stray outside the desired background range, leading to suboptimal", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Hu et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "results with CD values of 0.68 and 0.73, respectively. Additionally, increasing the number of sampling points in our approach tends to accentuate background details, thereby negatively impacting the accuracy of foreground object modeling. More background points would render the model to prioritize background fitting over the foreground, thus generating inferior results.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 237, + 390, + 252 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As shown in Tab. 6, we set five settings to demonstrate the superiority of our method design. From the results of CPT-V-M, CPT-R-M, and CPT-VR-M, it is evident that introducing view-ray directions brings greater benefits to geometry reconstruction than introducing the two vectors alone. Meanwhile, the comparable $CD$ values of CPT-VR-M and CPT-VR-BG1 demonstrate that the one-point background model can largely eliminate the negative impact of complex backgrounds in images on rendering the target object, enabling the model to", + "bbox": [ + 212, + 267, + 493, + 492 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 5: Impact of the Number of Background Points. Here, Num. indicates the sampling amount of points. Pos. is the sampling positions of points along the ray direction. $F_{p}$ is defined in Sec. 3.4. All settings are evaluated on Chamfer Distance (CD). Sampling one point at $2F_{p}$ attains the best performance.", + "bbox": [ + 501, + 285, + 787, + 396 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/e725b8325e6401dd117c2c5b92696d09982234508fc87d07492264af4642ffb9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodNum.Pos.CD↓
CPT-VR22Fp, 1000Fp0.74
CPT-VR2Fp, 2Fp0.67
CPT-VR11000Fp0.73
CPT-VR1Fp0.68
CPT-VR12Fp0.66
", + "bbox": [ + 527, + 398, + 759, + 479 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "no longer be dependent on any background knowledge. Moreover, the results of CPT-VR-BG1 and CPT-VR-BG1-D also verify that introducing depth supervision can further boost the reconstruction quality of our method.", + "bbox": [ + 212, + 494, + 787, + 540 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3b474c590a27cd876841f7b07c5181b6372e93c3af2fbd865472b83149b077e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 299, + 561, + 424, + 643 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/cd4d922a9c87d431a84ed5e55367b084b0ecf1ab2916371c62e8588c1eb141a4.jpg", + "image_caption": [ + "Fig.6: Qualitative results of ablation studies on DTU. The close-ups in red highlight the specific regions for comparison. The content in the bracket is the Chamfer Distance, indicating the reconstruction quality. Incorporating the reflection vector enhances the ability of the model to simulate the appearance representation on specular regions. In addition, when utilizing the background model to attain the background information, introducing depth supervision greatly improves the reconstruction quality." + ], + "image_footnote": [], + "bbox": [ + 308, + 646, + 423, + 728 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5e81aaf8d80130cf46e63e2a718e0b07b2d9929bec0a21d9ce83de1d3c78b2ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 437, + 561, + 557, + 642 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/cb4ba2cecde4e49a0b71a9e1e5f252a169b7614ff6c704d9b862fbaee26d9440.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 646, + 557, + 728 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d0cb822b8e330a8debf845bdfa826f12d9fbf8d5c8cdcd7d4784f4e776c6a67d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 575, + 563, + 697, + 643 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9e5be864a994b94fdcd4c0842cbbb2df6565090fb3590b1f23cdb48821a071e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 648, + 697, + 728 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Improving Surface Rendering via CPT with View-Reflection Appearance", + "bbox": [ + 246, + 114, + 732, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/effcc41dc21fd21d7bfdb79d0c2a473f43874a3a8dbc2dc498802f211df1cef8.jpg", + "table_caption": [ + "Table 6: Ablation study of our design choices. We evaluate all designs on Chamfer Distance (CD). Results demonstrate the effectiveness of all our design choices and the importance of using accurate geometry representation when adding the reflection vector for appearance modeling." + ], + "table_footnote": [], + "table_body": "
SettingsDesign ChoicesCD↓
View VectorReflection VectorBackground ModelLCLMLD
CPT-V-M✓✓✓0.72
CPT-R-M✓✓✓0.83
CPT-VR-M✓✓✓✓0.65
CPT-VR-BG1✓✓✓✓0.66
CPT-VR-BG1-D✓✓✓✓✓0.59
", + "bbox": [ + 241, + 205, + 758, + 320 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "As illustrated in Sec. 1, we argue that only when the surface sampling points are accurate can the introduction of the reflection vector achieve better appearance modeling of specular regions. To further demonstrate that, we incorporate the reflection vector into the surface rendering method DMTet [19] and IDR [34]. Tab. 7 indicates that introducing the reflection vector reduces the geometry reconstruction ability of DMTet ( $CD$ rises from 1.24 to 1.63) and IDR ( $CD$ rises from 0.90 to 1.68). Conversely, the sampling surface points of our method with CPT are more accurate. The accurate points facilitate the modeling of the reflection vector, avoiding a greater backpropagation error. Thus, incorporating the reflection vector into our method enables our model more robust against specular highlights. Fig. 6 also demonstrates the superiority of our design choices.", + "bbox": [ + 212, + 330, + 787, + 496 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/de96f119f9cf158f4b643514fbfc7f8e0cbbaa114eb5e46bb72fe1d46aafdeb1.jpg", + "table_caption": [ + "Table 7: Impact of view-reflection vectors. All methods are evaluated with masks on Chamfer Distance (CD). With the reflection vector, the performance of our method has significantly improved, while DMTet and IDR become worse." + ], + "table_footnote": [], + "table_body": "
MethodIDRIDRDMTetDMTetOursOursOurs
View Vector (V)✓✓✓✓✓✓
Reflection Vector (R)✓✓✓✓
CD↓0.901.681.241.630.720.830.65
", + "bbox": [ + 232, + 561, + 766, + 619 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 636, + 369, + 652 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we present a novel surface rendering method CPT-VR. Based on the CPT algorithm, we correct the deviated points approximated by linear interpolation, enhancing the geometric accuracy. Meanwhile, benefiting from the accurate geometry representation, we incorporate view-reflection vectors into the appearance molding process, which enables our method against the specular highlights. Moreover, we devise a background model to enable our model to handle cases with complex backgrounds without any background knowledge. Experiments on the popular 3D reconstruction datasets demonstrate the superiority of our method in surface rendering, especially for cases with complex structures and specular highlights. We hope that our designs can provide some insights for future works.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Hu et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bangaru, S.P., Gharbi, M., Luan, F., Li, T.M., Sunkavalli, K., Hasan, M., Bi, S., Xu, Z., Bernstein, G., Durand, F.: Differentiable rendering of neural sdfs through reparameterization. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)", + "2. Darmon, F., Bascle, B., Devaux, J.C., Monasse, P., Aubry, M.: Improving neural implicit surfaces geometry with patch warping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6260-6269 (2022)", + "3. Fu, Q., Xu, Q., Ong, Y.S., Tao, W.: Geo-neus: geometry-consistent neural implicit surfaces learning for multi-view reconstruction. arXiv preprint arXiv:2205.15848 (2022)", + "4. Gropp, A., Yariv, L., Haim, N., Atzmon, M., Lipman, Y.: Implicit geometric regularization for learning shapes. In: International Conference on Machine Learning. pp. 3789-3799. PMLR (2020)", + "5. Gu, X., Fan, Z., Zhu, S., Dai, Z., Tan, F., Tan, P.: Cascade cost volume for high-resolution multi-view stereo and stereo matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2495-2504 (2020)", + "6. Guo, Y.C., Cao, Y.P., Wang, C., He, Y., Shan, Y., Qie, X., Zhang, S.H.: Vmesh: Hybrid volume-mesh representation for efficient view synthesis. arXiv preprint arXiv:2303.16184 (2023)", + "7. Hart, J.C.: Sphere tracing: A geometric method for the antiaised ray tracing of implicit surfaces. The Visual Computer 12(10), 527-545 (1996)", + "8. Hasselgren, J., Hofmann, N., Munkberg, J.: Shape, light, and material decomposition from images using monte carlo rendering and denoising. Advances in Neural Information Processing Systems 35, 22856-22869 (2022)", + "9. Jensen, R., Dahl, A., Vogiatzis, G., Tola, E., Aanaes, H.: Large scale multi-view stereopsis evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 406-413 (2014)", + "0. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG) 39(6), 1-14 (2020)", + "1. Li, H., Yang, X., Zhai, H., Liu, Y., Bao, H., Zhang, G.: Vox-surf: Voxel-based implicit surface representation. IEEE Transactions on Visualization and Computer Graphics (2022)", + "2. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023)", + "3. Liu, H.T.D., Williams, F., Jacobson, A., Fidler, S., Litany, O.: Learning smooth neural functions via lipschitz regularization. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-13 (2022)", + "4. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics 21(4), 163-169 (1987)", + "5. Ma, B., Zhou, J., Liu, Y.S., Han, Z.: Towards better gradient consistency for neural signed distance functions via level set alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17724-17734 (2023)", + "6. Mehta, I., Chandraker, M., Ramamoorthi, R.: A level set theory for neural implicit evolution under explicit flows. In: European Conference on Computer Vision. pp. 711-729. Springer (2022)" + ], + "bbox": [ + 225, + 175, + 784, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Improving Surface Rendering via CPT with View-Reflection Appearance", + "bbox": [ + 246, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I. pp. 405-421 (2020)", + "18. Munkberg, J., Hasselgren, J., Shen, T., Gao, J., Chen, W., Evans, A., Müller, T., Fidler, S.: Extracting triangular 3d models, materials, and lighting from images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8280-8290 (2022)", + "19. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems 34, 6087-6101 (2021)", + "20. Sun, J., Chen, X., Wang, Q., Li, Z., Averbuch-Elor, H., Zhou, X., Snavely, N.: Neural 3d reconstruction in the wild. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-9 (2022)", + "21. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5481-5490. IEEE (2022)", + "22. Vicini, D., Speierer, S., Jakob, W.: Differentiable signed distance function rendering. ACM Transactions on Graphics (TOG) 41(4), 1-18 (2022)", + "23. Walker, T., Mariotti, O., Vaxman, A., Bilen, H.: Explicit neural surfaces: Learning continuous geometry with deformation fields. arXiv preprint arXiv:2306.02956 (2023)", + "24. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems 34, 27171-27183 (2021)", + "25. Wang, Y., Han, Q., Habermann, M., Daniilidis, K., Theobalt, C., Liu, L.: Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3295-3306 (2023)", + "26. Wang, Y., Skorokhodov, I., Wonka, P.: Hf-neus: Improved surface reconstruction using high-frequency details. Advances in Neural Information Processing Systems 35, 1966-1978 (2022)", + "27. Wang, Y., Skorokhodov, I., Wonka, P.: Improved surface reconstruction using high-frequency details. arXiv preprint arXiv:2206.07850 (2022)", + "28. Wang, Y., Skorokhodov, I., Wonka, P.: Pet-neus: Positional encoding tri-planes for neural surfaces. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12598-12607 (2023)", + "29. Worchel, M., Diaz, R., Hu, W., Schreer, O., Feldmann, I., Eisert, P.: Multiview mesh reconstruction with neural deferred shading. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6187-6197 (2022)", + "30. Wu, T., Wang, J., Pan, X., Xu, X., Theobalt, C., Liu, Z., Lin, D.: Voxurf: Voxel-based efficient and accurate neural surface reconstruction. arXiv preprint arXiv:2208.12697 (2022)", + "31. Yao, Y., Luo, Z., Li, S., Zhang, J., Ren, Y., Zhou, L., Fang, T., Quan, L.: Blended-mvs: A large-scale dataset for generalized multi-view stereo networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1790-1799 (2020)", + "32. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Hu et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Yariv, L., Hedman, P., Reiser, C., Verbin, D., Srinivasan, P.P., Szeliski, R., Barron, J.T., Mildenhall, B.: Bakedsdf: Meshing neural sdfs for real-time view synthesis. arXiv preprint arXiv:2302.14859 (2023)", + "34. Yariv, L., Kasten, Y., Moran, D., Galun, M., Atzmon, M., Ronen, B., Lipman, Y.: Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems 33, 2492-2502 (2020)", + "35. Zhang, J., Yao, Y., Li, S., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Critical regularizations for neural surface reconstruction in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6270-6279 (2022)", + "36. Zhang, J., Yao, Y., Li, S., Luo, Z., Fang, T.: Visibility-aware multi-view stereo network. British Machine Vision Conference (BMVC) (2020)", + "37. Zhang, J., Yao, Y., Quan, L.: Learning signed distance field for multi-view surface reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6525-6534 (2021)", + "38. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020)", + "39. Zhang, Y., Zhu, J., Lin, L.: Fastmesh: Fast surface reconstruction by hexagonal mesh-based neural rendering. arXiv preprint arXiv:2305.17858 (2023)", + "40. Zhang, Y., Hu, Z., Wu, H., Zhao, M., Li, L., Zou, Z., Fan, C.: Towards unbiased volume rendering of neural implicit surfaces with geometry priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4359-4368 (2023)", + "41. Zhuang, Y., Zhang, Q., Feng, Y., Zhu, H., Yao, Y., Li, X., Cao, Y.P., Shan, Y., Cao, X.: Anti-aliased neural implicit surfaces with encoding level of detail. In: SIGGRAPH Asia 2023 Conference Papers. pp. 1-10 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 507 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Improving Surface Rendering via CPT with View-Reflection Appearance", + "bbox": [ + 246, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/CPT-VR_ Improving Surface Rendering via Closest Point Transform with View-Reflection Appearance/255265d5-d4ab-4d8b-8210-adf9209a31be_model.json b/2024/CPT-VR_ Improving Surface Rendering via Closest Point Transform with View-Reflection Appearance/255265d5-d4ab-4d8b-8210-adf9209a31be_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7fe8b8aba0836cb4f5b205a25ffb896987c1d867 --- /dev/null +++ b/2024/CPT-VR_ Improving Surface Rendering via Closest Point Transform with View-Reflection Appearance/255265d5-d4ab-4d8b-8210-adf9209a31be_model.json @@ -0,0 +1,2676 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.24, + 0.141, + 0.765, + 0.21 + ], + "angle": 0, + "content": "CPT-VR: Improving Surface Rendering via Closest Point Transform with View-Reflection Appearance" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.234, + 0.777, + 0.266 + ], + "angle": 0, + "content": "Zhipeng Hu\\(^{1,3}\\), Yongqiang Zhang\\(^{1,3,*}\\), Chen Liu\\(^{2}\\), Lincheng Li\\(^{3,*}\\), Sida Peng\\(^{3}\\), Xiaowei Zhou\\(^{3}\\), Changjie Fan\\(^{1}\\), and Xin Yu\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.277, + 0.745, + 0.307 + ], + "angle": 0, + "content": "\\(^{1}\\)NetEase Fuxi AI Lab, Hangzhou, China {zhangyongqiang02,zphu,lilincheng,fanchangjie}@corp.netease.com" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.306, + 0.686, + 0.319 + ], + "angle": 0, + "content": "2 The University of Queensland, Queensland, Australia" + }, + { + "type": "text", + "bbox": [ + 0.332, + 0.321, + 0.672, + 0.334 + ], + "angle": 0, + "content": "chen.liu7@uqconnect.edu.au, xin.yu@uq.edu.au" + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.334, + 0.636, + 0.347 + ], + "angle": 0, + "content": "3 Zhejiang University, Hangzhou, China" + }, + { + "type": "text", + "bbox": [ + 0.394, + 0.349, + 0.61, + 0.362 + ], + "angle": 0, + "content": "{pengsida,xwzhou}@zju.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.392, + 0.74, + 0.683 + ], + "angle": 0, + "content": "Abstract. Differentiable surface rendering has significantly advanced 3D reconstruction. Existing surface rendering methods assume that the local surface is planar, and thus employ linear approximation based on the Singed Distance Field (SDF) values to predict the point on the surface. However, this assumption overlooks the inherently irregular and non-planar nature of object surfaces in the real world. Consequently, the approximate points tend to deviate from the zero-level set, affecting the fidelity of the reconstructed shape. In this paper, we propose a novel surface rendering method termed CPT-VR, which leverages the Closet Point Transform (CPT) and View and Reflection direction vectors to enhance the quality of reconstruction. Specifically, leveraging the physical property of CPT that accurately projects points near the surface onto the zero-level set, we correct the deviated points, thus achieving an accurate geometry representation. Based on our accurate geometry representation, incorporating the reflection vector into our method can facilitate the appearance modeling of specular regions. Moreover, to enable our method to no longer be dependent on any prior knowledge of the background, we present a background model to learn the background appearance. Compared to previous state-of-the-art methods, CPT-VR achieves better surface reconstruction quality, even for cases with complex structures and specular highlights." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.696, + 0.74, + 0.723 + ], + "angle": 0, + "content": "Keywords: Surface Rendering \\(\\cdot\\) 3D Reconstruction \\(\\cdot\\) Differentiable Rendering \\(\\cdot\\) Closest Point Transform" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.745, + 0.375, + 0.76 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.773, + 0.787, + 0.819 + ], + "angle": 0, + "content": "Differentiable rendering plays a crucial role in 3D content creation, especially showing greater promise in multi-view 3D reconstruction and text-to-3D generation. Recent advancements have utilized differentiable rendering for attaining" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.405, + 0.84 + ], + "angle": 0, + "content": "* Corresponding authors." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Z. Hu et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.149, + 0.324, + 0.162 + ], + "angle": 0, + "content": "(a) IDR" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.163, + 0.393, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.248, + 0.365, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.276, + 0.303, + 0.343, + 0.311 + ], + "angle": 0, + "content": "\\(S_0\\): zero-level set" + }, + { + "type": "image_caption", + "bbox": [ + 0.28, + 0.312, + 0.339, + 0.319 + ], + "angle": 0, + "content": "v: the view ray" + }, + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.32, + 0.4, + 0.337 + ], + "angle": 0, + "content": "x: the sampled point by sphere tracing \\(d\\): the SDF value of x, n: the normal vector of x" + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.337, + 0.385, + 0.345 + ], + "angle": 0, + "content": "\\(\\mathbf{x}^{\\prime}\\) : the linear projection along the ray \\(\\mathbf{v}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.149, + 0.523, + 0.162 + ], + "angle": 0, + "content": "(b)DMTet" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.163, + 0.575, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.413, + 0.244, + 0.567, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.296, + 0.517, + 0.305 + ], + "angle": 0, + "content": "\\(S_0\\): zero-level set" + }, + { + "type": "image_caption", + "bbox": [ + 0.407, + 0.305, + 0.562, + 0.313 + ], + "angle": 0, + "content": "\\(\\mathbf{x}_i, \\mathbf{x}_j\\): vertices of an edge where the SDF" + }, + { + "type": "image_caption", + "bbox": [ + 0.46, + 0.313, + 0.509, + 0.322 + ], + "angle": 0, + "content": "changes sign" + }, + { + "type": "image_caption", + "bbox": [ + 0.433, + 0.323, + 0.533, + 0.331 + ], + "angle": 0, + "content": "\\(d_{i},d_{j}\\) :SDF values of \\(\\mathbf{x}_i,\\mathbf{x}_j\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.428, + 0.331, + 0.542, + 0.339 + ], + "angle": 0, + "content": "\\(\\mathbf{x}_i^{\\prime},\\mathbf{x}_j^{\\prime}\\) : local deformed vertices" + }, + { + "type": "image_caption", + "bbox": [ + 0.407, + 0.34, + 0.561, + 0.35 + ], + "angle": 0, + "content": "\\(\\mathbf{x}_{ij}^{\\prime}\\) : linear interpolation of \\(\\mathbf{x}_i^{\\prime},\\mathbf{x}_j^{\\prime}\\) by \\(d_{i},d_{j}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.15, + 0.728, + 0.162 + ], + "angle": 0, + "content": "(c) CPT-Ours" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.163, + 0.764, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.246, + 0.742, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.305, + 0.715, + 0.312 + ], + "angle": 0, + "content": "\\(S_0\\) : zero-level set" + }, + { + "type": "image_caption", + "bbox": [ + 0.579, + 0.313, + 0.784, + 0.322 + ], + "angle": 0, + "content": "\\(\\mathbf{x}_i, \\mathbf{x}_j\\): vertices of an edge where the SDF changes sign" + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.322, + 0.779, + 0.33 + ], + "angle": 0, + "content": "\\(d_{i},d_{j},d_{ij}\\) : SDF values, \\(\\mathbf{n}_{ij}\\) : the normal vector of \\(\\mathbf{x}_{ij}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.33, + 0.758, + 0.339 + ], + "angle": 0, + "content": "\\(\\mathbf{x}_{ij}\\): linear interpolation of \\(\\mathbf{x}_i, \\mathbf{x}_j\\) by \\(d_i, d_j\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.339, + 0.756, + 0.349 + ], + "angle": 0, + "content": "\\(\\mathbf{x}_{ij}^{\\prime}\\) : the projected vertex of \\(\\mathbf{x}_{ij}\\) by CPT" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.514 + ], + "angle": 0, + "content": "Fig. 1: Illustration of the surface point sampling methods and their corresponding construction results. (a) IDR [34] first employs a sphere tracing method to identify the sampled point \\( x \\) and then leverages the SDF value as well as the normal vector to obtain the approximate point \\( x' \\). (b) DMTet [19] detects surfaces through the observation of sign changes in the SDF at the vertices of a tetrahedron. Then, it employs linear interpolation to approximate the points \\( x_{ij}' \\) on the zero-level set. Compared with them, our method (c) utilizes CPT to project the interpolated points \\( x_{ij} \\) on the zero-level set, and thus attain the transformed points \\( x_{ij}' \\). The approximate points \\( x' \\) of IDR and \\( x_{ij}' \\) of DMTet are deviated from the zero-level set, while \\( x_{ij}' \\) obtained from our method exactly on the object surface. This implies our method is able to obtain a more accurate geometry representation and gradient backpropagation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.788, + 0.625 + ], + "angle": 0, + "content": "the Singed Distance Field (SDF) and object appearance. These methods can be divided into two types: volume rendering and surface rendering. In volume rendering, SDF is first converted into density values that are then rendered via the NeRF framework [17, 38]. While these methods consistently generate reliable geometry reconstruction, the dense points sampling along the ray increases the rendering time cost. Conversely, surface rendering methods only sample one point along the ray, markedly optimizing rendering speeds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Existing surface rendering methods, such as IDR (Implicit Differentiable Renderer) [34] and DMTet (Deep Marching Tetrahedra) [19] achieve computational efficiency by assuming the object surface can be considered linear or at least can be approximated as linear within a local scope. Based on this assumption, they employ linear interpolation to determine the points at specific locations on the surface. As shown in Fig. 1, IDR employs sphere tracing [7] to identify points close to the surface along a ray and then relies on the SDF values and the normal vector to calculate the intersection points between rays and the surface. DMTet identifies the surfaces by observing changes in the sign of the SDF at the vertices of a tetrahedron and then leverages linear interpolation to approximate the surface points. However, as demonstrated in Fig. 1, their approximate points always deviate from the object surface (i.e., zero-level set), since the object surface is mostly irregular and non-planar. Hence, modeling geometry representations based on these deviated points would further lead to inaccurate gradient back-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.247, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Improving Surface Rendering via CPT with View-Reflection Appearance" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "propagation, resulting in the reconstructed surfaces losing a lot of details, as illustrated by the reconstruction results in Fig. 1 (a) and (b)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.787, + 0.317 + ], + "angle": 0, + "content": "To address the above problem, we introduce a novel surface rendering approach, termed CPT-VR, which is grounded in the Closet Point Transform algorithm and augmented with View-Reflection appearance modeling. In a nutshell, in the geometry representation modeling process, we conduct the Closet Point Transform (CPT) for deviated points to enhance the geometry representation. In the appearance modeling process, we intend to incorporate the View and Reflection vectors and thus attain better appearance representation for cases with specular highlights. To the best of our knowledge, we are the first attempt to correct the deviated points via CPT in the surface rendering task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.32, + 0.788, + 0.472 + ], + "angle": 0, + "content": "To be specific, in the geometry representation modeling process, we first devise a neural SDF network to compute SDF values for the entire 3D space. Based on the SDF values, we employ the Marching Cubes [14] to attain the vertices close to the surface of the object. As suggested in Fig. 1 (c), most of the obtained vertices are deviated from the object's surface rather than on it. Hence, we leverage CPT to calculate the closest point on the object surface for each deviated vertex, thus obtaining accurate and differentiable vertices. Subsequently, we utilize a feature-based rasterizer to render this triangle mesh in each view. In this fashion, we obtain the corresponding 3D position, geometry feature, and normal vector of each pixel in an image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.788, + 0.672 + ], + "angle": 0, + "content": "In the appearance representation modeling process, we focus on improving the geometry quality for specular highlights. Existing methods such as IDR [34] and NeuS2 [25] rely solely on the view direction to capture the surface information of objects, which cannot handle the object with highlights well. Ref-NeRF [21] demonstrates that replacing the view direction with the reflection direction can mitigate the impact of highlights. However, since the reflection vector calculated based on the deviated points is inaccurate, directly incorporating these reflection vectors into existing surface rendering methods will lead to significant backpropagation errors. Consequently, the model will not only be unable to handle the specular highlights but will also diminish its capability to reconstruct non-specular areas. In contrast, benefiting from our accurate point sampling on the zero-level set based on CPT, incorporating view-reflection vectors in our framework can facilitate the appearance modeling of specular regions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Moreover, we found that existing surface rendering methods [1,22,34] require prior background knowledge, such as foreground masks, or the background image, to precisely process and render foreground objects, which constrains their flexibility and applicability in various scenes. To solve this problem, we present a background model to obtain the background appearance representation. Then we attain the final appearance modeling by blending the foreground and background appearance. Extensive experiments conducted on DTU [9] and BlendedMVS [31] demonstrate that our method achieves state-of-the-art geometry quality and surface rendering. Additionally, we also illustrate that our method is effective and reliable in cases with complex structures and specular highlights. In summary, our contributions are three-fold:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Z. Hu et al." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.147, + 0.784, + 0.206 + ], + "angle": 0, + "content": "- We present a novel differentiable surface rendering method CPT-VR which shows superior performance in reconstructing the detailed structure of objects. With the CPT algorithm, we can address the approximate point deviation issue caused by linear interpolation." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.21, + 0.784, + 0.253 + ], + "angle": 0, + "content": "- We incorporate the view and reflection vectors to model the foreground appearance representation, facilitating our method to be robust against specular highlights." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.258, + 0.784, + 0.301 + ], + "angle": 0, + "content": "- We develop a simple background model to generate the background appearance representation, enabling our method to no longer be dependent on any prior background knowledge." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.784, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.31, + 0.388, + 0.326 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.552 + ], + "angle": 0, + "content": "Surface Rendering with SDF. IDR [34] introduces a system that learns surfaces in a self-supervised manner. It first identifies points close to the surface via sphere tracing and then finds approximate intersection points on the surface by projecting them along the view ray. Although IDR can learn the 3D shape and appearance from images, it relies heavily on accurate masks. To mitigate this issue, MVsDF [37] supervises the SDF network with depth maps generated by Vis-MVSNET [36]. It also uses relaxed masks generated by probability maps to learn a surface indicator. RegSDF [35] further introduces the Hessian regularization and the minimal surface constraint. Some works [1, 22] use specialized reparameterization techniques and design weighting functions along a ray based on SDF values, thus they can propagate gradients to intermediate points from sphere tracing. These reparameterization-based methods are similar to methods based on volume rendering, but they still need a given background: either black or a pre-defined environment map." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Rasterization-based Rendering. Recently, there have been attempts to utilize the differentiable rasterizer for surface rendering. NVDIFFREC [18] and NVDIFFRECmc [8] adapt DMTet [19] to extract a mesh from an implicit SDF and then render the mesh by the differentiable rasterizer Nvdiffrast [10]. While these methods are efficient, their reliance on the linear approximation in DMTet limits their ability to recover shapes with complex structures. Meanwhile, NIE [16] propagates neural implicit surfaces according to a flow field from a certain energy function. ENS [23] combines neural deformation fields with rasterization rendering, allowing deform an initial shape into a target surface progressively. Moreover, NDS [29] and FastMESH [39] focus on optimizing a rough mesh derived from priors through differentiable rasterization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Volume Rendering of SDF. Several methods [2, 20, 24, 32] have been developed for learning SDF through volume rendering. VolSDF [32] and NeuS [24] leverage Laplace or Logistic cumulative distribution function of SDF to represent the volume density and sample dense points along a ray based on SDF values. Based on these methods, some works [2, 3, 12, 26–28, 40, 41] aim to further enhance the geometry details. Among them, Geo-Neus [3] and TUVR [40] improve SDF accuracy by introducing additional supervision from point clouds. On the training efficiency front, Voxurf [30] and Vox-Surf [11] speed up the training process" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.247, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Improving Surface Rendering via CPT with View-Reflection Appearance" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "through voxel-based representations. NeuS2 [25] speeds up NeuS [24] by simplifying the second-order derivatives calculation of the SDF network constructed based on a multi-resolution hash encoding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.314 + ], + "angle": 0, + "content": "Hybrid Volume-rasterization Rendering. Recently, some rendering methods have been developed based on a hybrid volume-rasterization way. BakedSDF [33] first learns the SDF and appearance of a scene based on VolSDF [32] and transforms the SDF into a mesh. Then the transformed mesh is refined using a rasterizer pipeline. VMesh [6] takes a comprehensive approach by initially training a combination of a neural SDF and a density field through volume rendering. Then the mesh undergoes further refinement by DMTet [19] and a differentiable rasterizer [10], ensuring accurate surface representation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.336, + 0.331, + 0.351 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.367, + 0.373, + 0.381 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.785, + 0.467 + ], + "angle": 0, + "content": "Signed Distance Field (SDF) is a mathematical function \\( f \\) that measures the shortest distance from a point \\( \\mathbf{x} \\) to an object surface, indicating whether the point is outside (positive distance) or inside (negative distance) the surface. The surface is identified by the zero-level set \\( S \\), where SDF equals to zero, providing a concise description of the object's surface:" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.479, + 0.785, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} = \\left\\{\\mathbf {x} \\in \\mathbb {R} ^ {3} | f (\\mathbf {x}) = 0 \\right\\}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.784, + 0.552 + ], + "angle": 0, + "content": "Neural SDF denotes a neural network that models SDFs for 3D objects. In this work, we employ SDFnet which leverages the Multilayer Perceptrons (MLPs) enhanced with a multi-resolution hash encoding to model SDF [25]:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.563, + 0.785, + 0.578 + ], + "angle": 0, + "content": "\\[\n(d, \\mathbf {z}) = f _ {\\theta} (\\mathbf {x}, h _ {\\Omega} (\\mathbf {x})). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.589, + 0.785, + 0.65 + ], + "angle": 0, + "content": "Here, \\(d\\) is the predicted SDF value, \\(\\mathbf{z}\\) signifies the predicted geometry feature, \\(f_{\\theta}\\) denotes an MLP parameterized by weights \\(\\theta\\), and \\(h_{\\Omega}(\\mathbf{x})\\) is the multi-resolution hash encoding of the input point \\(\\mathbf{x}\\). Moreover, the normal vector \\(\\mathbf{n}\\) of \\(\\mathbf{x}\\) is derived as the normalized gradient of the implicit surface:" + }, + { + "type": "equation", + "bbox": [ + 0.441, + 0.663, + 0.785, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\mathbf {n} = \\nabla_ {\\mathbf {x}} d / \\| \\nabla_ {\\mathbf {x}} d \\|. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.784, + 0.719 + ], + "angle": 0, + "content": "To ensure training stability, our SDFnet employs an optimization schedule that progressively incorporates the level of detail and the numerical gradients [12]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.741, + 0.559, + 0.756 + ], + "angle": 0, + "content": "3.2 Geometry Representation Modeling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.841 + ], + "angle": 0, + "content": "Geometry Representation via SDF. As suggested in Fig. 2, we first leverage our SDFnet to compute the SDF values for the entire 3D space. Based on SDF values, we utilize the isosurface algorithm Marching Cubes to extract and represent the triangle mesh \\(\\mathcal{G} = (\\mathcal{V},\\mathcal{F})\\), where \\(\\mathcal{V}\\) and \\(\\mathcal{F}\\) indicate the vertices and faces respectively. However, this method assumes that the surface changes" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Z. Hu et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.154, + 0.775, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.237, + 0.775, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.381, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Fig. 2: Illustration of our proposed method CPT-VR. In (a), we first extract the mesh via Marching Cubes, which does not involve gradient propagation. Then we leverage CPT to project the approximate points on the zero-level set (b). Subsequently, we render the transformed mesh via a rasterizer in (c). Specifically, we incorporate the view-reflection vectors into the foreground appearance modeling stage via RGBNet. Additionally, we employ the Background Model to attain the background information. Finally, we integrate the foreground and background appearance to generate the rendering results. We adopt four supervisions to optimize these neural networks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.788, + 0.56 + ], + "angle": 0, + "content": "linearly inside the voxel, overlooking the complex convex structure of the object surface. Hence, the linearly interpolated vertices tend to deviate from the zero-level set and the triangle mesh may not accurately reflect the complex structure of the object surface." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.562, + 0.788, + 0.653 + ], + "angle": 0, + "content": "Closest Point Transform. To project these deviated vertices onto the zero-level set, we devise the closet set transform (CPT) to SDF. Specifically, CPT first utilizes the norm vector \\(\\mathbf{n}\\) within the signed distance field to identify the projected direction of a deviated point. It then moves the deviated point along the opposite direction of \\(\\mathbf{n}\\) by a distance \\(d\\) equal to its SDF value. Mathematically, for a vertex \\(\\mathbf{x} \\in \\mathcal{V}\\), this transformation process is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.453, + 0.67, + 0.786, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} ^ {\\prime} = \\mathbf {x} - d \\mathbf {n}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.694, + 0.63, + 0.711 + ], + "angle": 0, + "content": "After that, we obtain the transformed mesh \\(\\mathcal{G}' = (\\mathcal{V}',\\mathcal{F})\\)" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.736, + 0.494, + 0.751 + ], + "angle": 0, + "content": "3.3 Feature-based Rasterisation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "After obtaining the transformed mesh, we first employ a feature-based differentiable rasterizer Nvdiffrast [10] to render triangle meshes in each view. Specifically, as suggested in Fig. 2 (c), for a pixel \\( p(x,y) \\), its projected primitive triangle face within the mesh is represented as \\( (\\mathbf{x}_0',\\mathbf{x}_1',\\mathbf{x}_2') \\). The rasterizer computes the barycentric coordinates \\( u,v \\), depth \\( D_{p} \\), and visible mask \\( M_{p} \\) for this pixel as" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.247, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Improving Surface Rendering via CPT with View-Reflection Appearance" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.274, + 0.16 + ], + "angle": 0, + "content": "follows:" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.163, + 0.785, + 0.179 + ], + "angle": 0, + "content": "\\[\n(u, v, D _ {p}, M _ {p}) = \\mathcal {R} \\left(\\mathcal {G} ^ {\\prime}, p (x, y), K, R, \\mathbf {t}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.184, + 0.786, + 0.214 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}\\) denotes the rasterizer, \\(\\mathcal{G}'\\) represents the transformed mesh, and \\(K\\) and \\(R, \\mathbf{t}\\) are the camera's intrinsic and extrinsic parameters, respectively." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.215, + 0.784, + 0.275 + ], + "angle": 0, + "content": "Subsequently, each projected vertex \\(\\mathbf{x}'\\) from the set \\(\\mathcal{V}'\\) is input into SDFnet. SDFnet outputs \\((d',\\mathbf{z}',\\mathbf{n}')\\), encapsulating the distance, geometric feature, and the normal vector associated with each vertex. These outputs are then interpolated for a pixel \\(p(x,y)\\) based on its barycentric coordinates \\(u,v\\):" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.288, + 0.619, + 0.303 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {p} = u \\mathbf {x} _ {0} ^ {\\prime} + v \\mathbf {x} _ {1} ^ {\\prime} + (1 - u - v) \\mathbf {x} _ {2} ^ {\\prime},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.304, + 0.785, + 0.318 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {p} = u \\mathbf {z} _ {0} ^ {\\prime} + v \\mathbf {z} _ {1} ^ {\\prime} + (1 - u - v) \\mathbf {z} _ {2} ^ {\\prime}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.32, + 0.62, + 0.334 + ], + "angle": 0, + "content": "\\[\n\\mathbf {n} _ {p} = u \\mathbf {n} _ {0} ^ {\\prime} + v \\mathbf {n} _ {1} ^ {\\prime} + (1 - u - v) \\mathbf {n} _ {2} ^ {\\prime},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.344, + 0.586, + 0.361 + ], + "angle": 0, + "content": "where \\((\\mathbf{x_p},\\mathbf{z_p},\\mathbf{n_p})\\) denote the interpolated features." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.381, + 0.574, + 0.396 + ], + "angle": 0, + "content": "3.4 Appearance Representation Modeling" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.406, + 0.734, + 0.42 + ], + "angle": 0, + "content": "In this work, we divide the appearance modeling process into two parts:" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.42, + 0.784, + 0.527 + ], + "angle": 0, + "content": "Foreground Appearance Modeling based on View-Reflection Vectors. To better capture the appearance of objects with specular highlights, we incorporate the view direction and reflection direction into the foreground modeling process. Specifically, we employ the RGBnet \\( g_{\\gamma}^{f} \\), a Multilayer Perceptron (MLP) with weights \\( \\gamma \\) to determine the foreground color. RGBnet takes the interpolated 3D features \\( (\\mathbf{x_p}, \\mathbf{z_p}, \\mathbf{n_p}) \\), view direction \\( \\mathbf{v}_p \\), and reflection direction \\( \\mathbf{r}_p \\) as input, and outputs the RGB color \\( \\mathbf{c}_p^f \\) for the foreground. The process is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.38, + 0.54, + 0.785, + 0.559 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {p} ^ {f} = g _ {\\gamma} ^ {f} \\left(\\mathbf {x} _ {\\mathbf {p}}, \\mathbf {z} _ {\\mathbf {p}}, \\mathbf {n} _ {\\mathbf {p}}, Y (\\mathbf {v} _ {\\mathbf {p}}), Y (\\mathbf {r} _ {\\mathbf {p}})\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.563, + 0.784, + 0.609 + ], + "angle": 0, + "content": "where \\( Y(\\cdot) \\) denotes Spherical Harmonics encoding. Here, the reflection direction \\( \\mathbf{r_p} \\) is defined as the mirror vector of the view direction \\( \\mathbf{v_p} \\) relative to the normal vector \\( \\mathbf{n_p} \\), calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.622, + 0.785, + 0.637 + ], + "angle": 0, + "content": "\\[\n\\mathbf {r} _ {\\mathbf {p}} = 2 \\left(- \\mathbf {v} _ {\\mathbf {p}} \\cdot \\mathbf {n} _ {\\mathbf {p}}\\right) \\mathbf {n} _ {\\mathbf {p}} + \\mathbf {v} _ {\\mathbf {p}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.647, + 0.784, + 0.692 + ], + "angle": 0, + "content": "Spherical harmonics are orthogonal bases on the unit sphere, and their mathematical definition and physical properties make them ideal for encoding direction frequencies compared to 3D space (xyz) frequencies." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.693, + 0.784, + 0.77 + ], + "angle": 0, + "content": "Background Appearance Modeling. To enable the model to reconstruct objects without any background priors, we present a one-point background model composed of a Multilayer Perceptron (MLP) \\( g_{\\pi}^{b} \\) with weights \\( \\pi \\), alongside an additional multi-resolution hash encoding \\( h_{\\Omega}^{b} \\). The modeling process of the background appearance \\( \\mathbf{c}_p^b \\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.781, + 0.785, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {p} ^ {b} = g _ {\\pi} ^ {b} \\left(\\mathbf {x} _ {p} ^ {b}, h _ {\\Omega} ^ {b} \\left(\\mathbf {x} _ {p} ^ {b}\\right), Y \\left(\\mathbf {v} _ {p}\\right)\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.784, + 0.842 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_p^b\\) represents the background sampling point, determined by the camera location \\(\\mathbf{x}_o\\), the veiw direction \\(\\mathbf{v}_p\\) and the far distance \\(F_{p}\\) of the sampling" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "Z. Hu et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.146, + 0.545, + 0.165 + ], + "angle": 0, + "content": "boundary. \\(\\mathbf{x}_p^b\\) is mathematically expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.436, + 0.17, + 0.786, + 0.19 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {p} ^ {b} = \\mathbf {x} _ {o} + 2 F _ {p} \\mathbf {v} _ {p}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.194, + 0.784, + 0.226 + ], + "angle": 0, + "content": "Following NeuS [24], the foreground is normalized within a unit sphere centered at \\(\\mathbf{0}\\). As indicated in Fig. 3, the far distance \\(F_{p}\\) is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.404, + 0.232, + 0.786, + 0.248 + ], + "angle": 0, + "content": "\\[\nm = - \\mathbf {x} _ {o} \\cdot \\mathbf {v} _ {p}; F _ {p} = m + 1. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.254, + 0.784, + 0.283 + ], + "angle": 0, + "content": "This approach facilitates a more versatile background representation, compensating for the absence of masks." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.285, + 0.518, + 0.346 + ], + "angle": 0, + "content": "Appearance Integration. Finally, we integrate the background appearance \\(\\mathbf{c}_p^b\\) and foreground appearance \\(\\mathbf{c}_p^f\\) to attain the rendered color, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.258, + 0.351, + 0.514, + 0.372 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {p} = M _ {p} \\mathbf {c} _ {p} ^ {f} + (1 - M _ {p}) \\mathbf {c} _ {p} ^ {b}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.376, + 0.516, + 0.437 + ], + "angle": 0, + "content": "Overall, our method enables surface rendering of objects without any known background information, offering greater flexibility and adaptability." + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.281, + 0.789, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.388, + 0.786, + 0.418 + ], + "angle": 0, + "content": "Fig.3: Illustration of the one-point background sampling." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.454, + 0.427, + 0.47 + ], + "angle": 0, + "content": "3.5 Training Objectives" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.475, + 0.784, + 0.537 + ], + "angle": 0, + "content": "Color loss. We optimize the CPT-VR model by ensuring the colors it generates match real image colors as closely as possible. This is done by calculating the L1 error, which is the average difference between the colors \\(\\hat{\\mathbf{c}}_i\\) produced by the model and the ground-truth colors \\(\\mathbf{c}_i\\) for all pixels \\(P\\):" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.542, + 0.786, + 0.582 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C} = \\frac {1}{P} \\sum_ {i} ^ {P} \\| \\hat {\\mathbf {c}} _ {i} - \\mathbf {c} _ {i} \\| _ {1}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.588, + 0.788, + 0.725 + ], + "angle": 0, + "content": "Regularization of SDFnet. A well-defined neural SDF exhibits Lipschitz continuity [13] and its gradient adheres to the Eikonal equation [4], ensuring smoothness and stability of the generated field. A valid SDF field holds the property that the normal of a point near the surface and the normal of its projected point remain consistent. For effective closest point transform, we directly constrain our SDFnet based on this normal consistency property [15]. Specifically, it aligns the normal of a deviated point \\(\\mathbf{n}\\) with the normal \\(\\mathbf{n}'\\) of its projected point, enhancing stability for surface rendering via the closest point transform of SDF. Mathematically, it is represented by:" + }, + { + "type": "equation", + "bbox": [ + 0.44, + 0.731, + 0.786, + 0.747 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {n c} = 1 - \\mathbf {n} \\cdot \\mathbf {n} ^ {\\prime}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.754, + 0.784, + 0.799 + ], + "angle": 0, + "content": "Mask Loss. If the mask of the foreground region is available, we can further refine the rendering effect of CPT-VR by including a mask loss. The mask constraint is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.409, + 0.802, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {M} = \\frac {1}{P} \\sum_ {i} ^ {P} \\left(\\hat {M} _ {i} - M _ {i}\\right) ^ {2}, \\tag {15}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.247, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Improving Surface Rendering via CPT with View-Reflection Appearance" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.177 + ], + "angle": 0, + "content": "where the \\(\\hat{M}_i\\) and the \\(M_{i}\\) represent the predicted mask and the ground truth mask, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.253 + ], + "angle": 0, + "content": "Depth loss. Inspired by previous works [37] and [35], we utilize depth maps generated by the pre-trained network CasMVSNet [5] to guide our model. We minimize the squared difference (MSE) between the rendered depths \\((\\hat{D}_i)\\) and those from the pre-trained network \\((D_i)\\), weighted by the confidence \\(prob_i\\) from CasMVSNet, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.398, + 0.263, + 0.785, + 0.302 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D} = \\frac {1}{P} \\sum_ {i} ^ {P} p r o b _ {i} \\left(\\hat {D} _ {i} - D _ {i}\\right) ^ {2}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.31, + 0.784, + 0.339 + ], + "angle": 0, + "content": "where the confidence ranges from 0 to 1, and depths considered unreliable are given a confidence of 0." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.784, + 0.37 + ], + "angle": 0, + "content": "Total loss. Our final metric combines all these aspects, adjusting their importance with specific weights \\(\\lambda\\):" + }, + { + "type": "equation", + "bbox": [ + 0.352, + 0.381, + 0.785, + 0.396 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {r g b} \\mathcal {L} _ {r g b} + \\lambda_ {n c} \\mathcal {L} _ {n c} + \\lambda_ {M} \\mathcal {L} _ {M} + \\lambda_ {D} \\mathcal {L} _ {D}. \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.404, + 0.781, + 0.421 + ], + "angle": 0, + "content": "Here, \\(\\lambda_{rgb}\\), \\(\\lambda_{nc}\\), \\(\\lambda_M\\), and \\(\\lambda_D\\) are empirically set to 3, 0.01, 1, and 1, respectively." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.441, + 0.376, + 0.459 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.47, + 0.446, + 0.485 + ], + "angle": 0, + "content": "4.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.785, + 0.658 + ], + "angle": 0, + "content": "Implementation details. For the DTU dataset, we train CPT-VR for 10,000 iterations with a batch size of one image, targeting a Marching Cubes resolution of \\(256^3\\). In each training batch, we render one view at a resolution of 800 \\(\\times\\) 600 pixels. Note that, when we only utilize depth loss and color loss for supervision, we employ a higher Marching Cubes resolution of \\(512^3\\) and render images at \\(1600 \\times 1200\\) pixels. For BlendedMVS, CPT-VR is trained for 20,000 iterations with a batch size of one image and a Marching Cubes resolution of \\(256^3\\), rendering images at \\(768 \\times 576\\) pixels. Moreover, the initial learning rate of our model is set at 0.005, with a final epoch decay factor of 0.1 for DTU and 0.005 for BlendedMVS. More details about the architecture and the isosurface configuration are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.785, + 0.795 + ], + "angle": 0, + "content": "Datasets. All experiments are conducted on DTU [9] and BlendedMVS [31]. Objects in DTU are collected in a controlled lab setting, with ground truth data that includes images, masks, and point clouds. For our training and evaluation, we leverage 15 objects identified by IDR with challenges such as specular highlights, delicate structures, and areas hidden from view. Meanwhile, BlendedMVS encompasses a diverse range of objects and scenes captured in natural settings. From this dataset, we select 7 difficult objects as NeuS [24] and the 'Bull' (with specular highlights) to further assess the model performance in reconstructing objects against complex backgrounds." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.815, + 0.461, + 0.83 + ], + "angle": 0, + "content": "4.2 Quantitative Evaluation" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "Z. Hu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.493, + 0.312 + ], + "angle": 0, + "content": "Analysis on DTU. We benchmark our method against the state-of-the-art methods on the DTU dataset. DTU dataset contains 128 scans captured in the laboratory. In this work, we evaluate the surface mesh using the same set of scans as in [34,37]. As denoted in Tab. 2, our method consistently outperforms other surface rendering methods on most scans. Note that, our method achieves the lowest" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.42 + ], + "angle": 0, + "content": "value in mean Chamfer Distance (0.59), which is \\(16.9\\%\\) lower than the second-best surface rendering method RegSDF (0.71). Since SDFnet in our method has a similar structure to it in NeuS2, we also compare NeuS2 and recent volume rendering methods [26, 41] with CPT-VR. As demonstrated in Tab. 2, whether the two methods incorporate background information from the object mask \\((\\mathrm{w} / \\mathrm{M})\\) or use our proposed one-point background model \\((\\mathrm{w} / \\mathrm{BG1})\\) to attain background information, our method surpasses NeuS2 on the vast majority of ScanIDs." + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.137, + 0.788, + 0.221 + ], + "angle": 0, + "content": "Table 1: Comparison of the average rendering time on DTU. The rendered image resolution of all methods is \\(1600 \\times 1200\\), and the resolution of Marching Cubes is \\(512^{3}\\). The rendering time of our method is significantly lower than that of others." + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.222, + 0.758, + 0.307 + ], + "angle": 0, + "content": "
MethodRendering Time (s)
IDR27.233
NeuS2 (w /M)1.979
Ours (w/ M)0.057
Ours (w/ BG1)0.104
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.788, + 0.556 + ], + "angle": 0, + "content": "Table 2: Quantitative results on DTU dataset. Here, we evaluate all methods on Chamfer Distance (CD). The lower the CD value, the better the performance. For a fair comparison, we also present the results of the volume rendering method NeuS2 since our SDFnet has a similar structure to it. \"w/ M\" means providing the prior background knowledge by masks, while \"w/ BG1\" means employing our one-point background model without any prior background knowledge. The best and second-best results are highlighted by different colors. The results illustrated that CPT-VR attains better performance among most ScanIDs." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.561, + 0.803, + 0.735 + ], + "angle": 0, + "content": "
ScanID243740556365698397105106110114118122Mean
DMTet [19]1.841.731.050.611.661.370.681.283.821.070.911.150.370.540.551.24
IDR [34]1.631.870.630.481.040.790.771.331.160.760.670.900.420.510.530.90
MVSDF [37]0.831.760.880.441.110.900.751.261.021.350.870.840.340.470.460.89
FastMESH [39]0.651.480.570.401.480.770.560.860.840.940.720.810.520.490.540.77
RegSDF [35]0.591.410.630.421.340.620.590.890.911.020.600.590.290.400.380.71
HF-NeuS [26] (w/ M)1.111.280.610.470.970.680.621.340.910.730.531.820.380.540.510.83
LOD-NeuS [41] (w/ M)0.650.910.370.481.050.870.821.220.950.690.561.300.420.580.570.76
NeuS2 [25] (w/ M)0.560.760.490.370.920.710.761.221.080.630.590.890.400.480.550.70
NeuS2 [25] (w/ BG1)0.690.860.830.341.050.680.651.051.120.670.581.410.370.500.520.75
Ours (w/ M)0.550.720.350.380.850.590.540.810.830.670.530.590.320.380.370.56
Ours (w/ BG1)0.430.730.360.320.930.610.610.891.020.680.480.730.320.370.370.59
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Moreover, we also measure the rendered results with PSNR to the ground truths. As indicated in Tab. 3, with the introduction of background information (BG1: \\(\\pmb{x}\\)), our method achieves the highest PSNR values on most ScanIDs. Additionally, it can be observed that when using the one-point background model, both NeuS2 and CPT-VR are able to achieve similar or better performance compared to leveraging background information from masks (BG1: \\(\\checkmark\\))." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.247, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Improving Surface Rendering via CPT with View-Reflection Appearance" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.139, + 0.788, + 0.236 + ], + "angle": 0, + "content": "Table 3: Quantitative comparison of 2D view synthesis on DTU dataset. We measure all methods based on the metric PSNR. A higher PSNR indicates better synthesis quality. Here, B1 represents whether to introduce extra background knowledge into the method. \\(\\times\\) means to leverage the background information from masks, while \\(\\checkmark\\) indicates utilizing our one-point background model without introducing any background information. We color code the best and second-best results. Compared to NeuS2, our method achieves much better appearance reconstruction results." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.244, + 0.798, + 0.322 + ], + "angle": 0, + "content": "
ScanIDBG1243740556365698397105106110114118122Mean
NeuS2✗28.4427.1429.7029.6731.7527.8324.8431.2426.8630.5726.0528.9328.9827.8232.4828.82
NeuS2✓27.7124.9927.5128.8920.5530.1429.0127.7026.1631.0933.1030.6930.4832.7734.1028.99
Ours✗28.1527.3128.5429.5234.0429.8329.0035.9229.3233.0829.4531.1630.7730.5038.3831.00
Ours✓27.3025.5926.8827.7730.9130.0629.2828.0526.6828.9232.0330.1829.0234.0234.0229.38
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.329, + 0.788, + 0.398 + ], + "angle": 0, + "content": "Table 4: Quantitative comparison on BlendedMVS dataset. We measure the method on the metric Chamfer Distance (CD). The lower CD values indicate better performance. For a fair comparison, we adopt our one-point background model to NeuS2. Compared to NeuS2, CPT-VR attains better geometry reconstruction results among most of the categories." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.403, + 0.776, + 0.457 + ], + "angle": 0, + "content": "
MethodBearClockDogDurianManSculptureBullStoneMean
NeuS20.380.470.950.470.460.400.650.380.52
Ours0.350.450.680.480.460.390.450.380.45
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.788, + 0.61 + ], + "angle": 0, + "content": "Analysis on BlendedMVS. As illustrated in Tab. 4, we compare our method with NeuS2 on the eight selected categories with complex backgrounds. Since the scales are unknown in BlendedMVS, we scale the bounding box of each object to 100 units to measure their performance on Chamfer Distance (CD). CPT-VR achieves the best performance in the categories of 'Bear', 'Clock', 'Dog', 'Sculpture', and 'Bull', with the CD value for 'Man' and 'Stone' being on par with NeuS2, and only 0.01 behind NeuS2 for 'Durian'. Moreover, CPT-VR outperforms NeuS2 by \\(14.8\\%\\). These results demonstrate the advantage of our method in reconstructing objects within complex environments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.788, + 0.703 + ], + "angle": 0, + "content": "Analysis of Method Efficiency. We evaluate the rendering time of IDR, NeuS2, and CPT-VR on DTU. The results in Tab. 1 suggest the rendering time of IDR (24.233s) and NeuS2 (1.97s) is lengthy. Compared with them, our method achieves real-time rendering under the cases without introducing background priors being about 19 times faster than NeuS2. This is primarily because NeuS2 utilizes the dense sampling method in the volume rendering process." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.708, + 0.449, + 0.723 + ], + "angle": 0, + "content": "4.3 Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Impact of Specular Highlights. Fig. 4 illustrates the qualitative results on the DTU dataset. We compare our CPT-VR(w/M) model with IDR [34], DMTet [19], MVsDF [37], and NeuS2 [25]. As shown by the close-up regions in Fig. 4, the specular regions constructed by other methods exhibit obvious indentations, while our method can construct these regions with high quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Impact of Complex Backgrounds. Fig. 5 demonstrates the visual results of reconstructing objects with complex backgrounds. Since existing surface render-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "Z. Hu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "ing methods cannot handle the cases without any background prior, we specifically evaluate NeuS2 and our method on BlendedMVS. As the elliptical area demonstrated, our method can effectively model the geometry of foreground objects without introducing background information, particularly the details on the surface of objects. Moreover, the rendering results of our method closely resemble the input image." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.265, + 0.758, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.388, + 0.788, + 0.431 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparisons on objects in DTU with specular highlights. The close-ups in yellow highlight the specific regions for comparison. The geometry reconstruction results demonstrate that our method is robust against specular highlights." + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.489, + 0.761, + 0.671 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.679, + 0.788, + 0.737 + ], + "angle": 0, + "content": "Fig.5: Qualitative results of objects with complex background on BlendedMVS. We outline the close-up regions for detailed comparison. It is evident that without the mask guidance, our method also exhibits strong capabilities in modeling foreground objects, especially for the finer-grained surface structures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Impact of the Sampling Point Amount. Tab. 5 suggests the impact of varying the number of sampling points and their distances. The optimal result is achieved when sampling one point at \\(2F_{p}\\), and the \\(CD\\) is 0.66. Conversely, sampling either too closely (at \\(F_{p}\\)) or too distantly (at \\(1000F_{p}\\)) causes the sampled points to stray outside the desired background range, leading to suboptimal" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.247, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Improving Surface Rendering via CPT with View-Reflection Appearance" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "results with CD values of 0.68 and 0.73, respectively. Additionally, increasing the number of sampling points in our approach tends to accentuate background details, thereby negatively impacting the accuracy of foreground object modeling. More background points would render the model to prioritize background fitting over the foreground, thus generating inferior results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.238, + 0.391, + 0.253 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.494, + 0.493 + ], + "angle": 0, + "content": "As shown in Tab. 6, we set five settings to demonstrate the superiority of our method design. From the results of CPT-V-M, CPT-R-M, and CPT-VR-M, it is evident that introducing view-ray directions brings greater benefits to geometry reconstruction than introducing the two vectors alone. Meanwhile, the comparable \\(CD\\) values of CPT-VR-M and CPT-VR-BG1 demonstrate that the one-point background model can largely eliminate the negative impact of complex backgrounds in images on rendering the target object, enabling the model to" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.286, + 0.789, + 0.397 + ], + "angle": 0, + "content": "Table 5: Impact of the Number of Background Points. Here, Num. indicates the sampling amount of points. Pos. is the sampling positions of points along the ray direction. \\(F_{p}\\) is defined in Sec. 3.4. All settings are evaluated on Chamfer Distance (CD). Sampling one point at \\(2F_{p}\\) attains the best performance." + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.399, + 0.76, + 0.481 + ], + "angle": 0, + "content": "
MethodNum.Pos.CD↓
CPT-VR22Fp, 1000Fp0.74
CPT-VR2Fp, 2Fp0.67
CPT-VR11000Fp0.73
CPT-VR1Fp0.68
CPT-VR12Fp0.66
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.789, + 0.541 + ], + "angle": 0, + "content": "no longer be dependent on any background knowledge. Moreover, the results of CPT-VR-BG1 and CPT-VR-BG1-D also verify that introducing depth supervision can further boost the reconstruction quality of our method." + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.563, + 0.425, + 0.644 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.647, + 0.424, + 0.729 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.439, + 0.563, + 0.558, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.647, + 0.558, + 0.729 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.564, + 0.698, + 0.644 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.579, + 0.649, + 0.699, + 0.729 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.741, + 0.788, + 0.825 + ], + "angle": 0, + "content": "Fig.6: Qualitative results of ablation studies on DTU. The close-ups in red highlight the specific regions for comparison. The content in the bracket is the Chamfer Distance, indicating the reconstruction quality. Incorporating the reflection vector enhances the ability of the model to simulate the appearance representation on specular regions. In addition, when utilizing the background model to attain the background information, introducing depth supervision greatly improves the reconstruction quality." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "Z. Hu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 6: Ablation study of our design choices. We evaluate all designs on Chamfer Distance (CD). Results demonstrate the effectiveness of all our design choices and the importance of using accurate geometry representation when adding the reflection vector for appearance modeling." + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.206, + 0.759, + 0.321 + ], + "angle": 0, + "content": "
SettingsDesign ChoicesCD↓
View VectorReflection VectorBackground ModelLCLMLD
CPT-V-M✓✓✓0.72
CPT-R-M✓✓✓0.83
CPT-VR-M✓✓✓✓0.65
CPT-VR-BG1✓✓✓✓0.66
CPT-VR-BG1-D✓✓✓✓✓0.59
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.331, + 0.788, + 0.497 + ], + "angle": 0, + "content": "As illustrated in Sec. 1, we argue that only when the surface sampling points are accurate can the introduction of the reflection vector achieve better appearance modeling of specular regions. To further demonstrate that, we incorporate the reflection vector into the surface rendering method DMTet [19] and IDR [34]. Tab. 7 indicates that introducing the reflection vector reduces the geometry reconstruction ability of DMTet (\\(CD\\) rises from 1.24 to 1.63) and IDR (\\(CD\\) rises from 0.90 to 1.68). Conversely, the sampling surface points of our method with CPT are more accurate. The accurate points facilitate the modeling of the reflection vector, avoiding a greater backpropagation error. Thus, incorporating the reflection vector into our method enables our model more robust against specular highlights. Fig. 6 also demonstrates the superiority of our design choices." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.513, + 0.788, + 0.556 + ], + "angle": 0, + "content": "Table 7: Impact of view-reflection vectors. All methods are evaluated with masks on Chamfer Distance (CD). With the reflection vector, the performance of our method has significantly improved, while DMTet and IDR become worse." + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.562, + 0.767, + 0.62 + ], + "angle": 0, + "content": "
MethodIDRIDRDMTetDMTetOursOursOurs
View Vector (V)✓✓✓✓✓✓
Reflection Vector (R)✓✓✓✓
CD↓0.901.681.241.630.720.830.65
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.637, + 0.37, + 0.654 + ], + "angle": 0, + "content": "5 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we present a novel surface rendering method CPT-VR. Based on the CPT algorithm, we correct the deviated points approximated by linear interpolation, enhancing the geometric accuracy. Meanwhile, benefiting from the accurate geometry representation, we incorporate view-reflection vectors into the appearance molding process, which enables our method against the specular highlights. Moreover, we devise a background model to enable our model to handle cases with complex backgrounds without any background knowledge. Experiments on the popular 3D reconstruction datasets demonstrate the superiority of our method in surface rendering, especially for cases with complex structures and specular highlights. We hope that our designs can provide some insights for future works." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.247, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Improving Surface Rendering via CPT with View-Reflection Appearance" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.785, + 0.218 + ], + "angle": 0, + "content": "1. Bangaru, S.P., Gharbi, M., Luan, F., Li, T.M., Sunkavalli, K., Hasan, M., Bi, S., Xu, Z., Bernstein, G., Durand, F.: Differentiable rendering of neural sdfs through reparameterization. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.26 + ], + "angle": 0, + "content": "2. Darmon, F., Bascle, B., Devaux, J.C., Monasse, P., Aubry, M.: Improving neural implicit surfaces geometry with patch warping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6260-6269 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.3 + ], + "angle": 0, + "content": "3. Fu, Q., Xu, Q., Ong, Y.S., Tao, W.: Geo-neus: geometry-consistent neural implicit surfaces learning for multi-view reconstruction. arXiv preprint arXiv:2205.15848 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.301, + 0.785, + 0.342 + ], + "angle": 0, + "content": "4. Gropp, A., Yariv, L., Haim, N., Atzmon, M., Lipman, Y.: Implicit geometric regularization for learning shapes. In: International Conference on Machine Learning. pp. 3789-3799. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.397 + ], + "angle": 0, + "content": "5. Gu, X., Fan, Z., Zhu, S., Dai, Z., Tan, F., Tan, P.: Cascade cost volume for high-resolution multi-view stereo and stereo matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2495-2504 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.398, + 0.785, + 0.44 + ], + "angle": 0, + "content": "6. Guo, Y.C., Cao, Y.P., Wang, C., He, Y., Shan, Y., Qie, X., Zhang, S.H.: Vmesh: Hybrid volume-mesh representation for efficient view synthesis. arXiv preprint arXiv:2303.16184 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.785, + 0.468 + ], + "angle": 0, + "content": "7. Hart, J.C.: Sphere tracing: A geometric method for the antiaised ray tracing of implicit surfaces. The Visual Computer 12(10), 527-545 (1996)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.785, + 0.509 + ], + "angle": 0, + "content": "8. Hasselgren, J., Hofmann, N., Munkberg, J.: Shape, light, and material decomposition from images using monte carlo rendering and denoising. Advances in Neural Information Processing Systems 35, 22856-22869 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "9. Jensen, R., Dahl, A., Vogiatzis, G., Tola, E., Aanaes, H.: Large scale multi-view stereopsis evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 406-413 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.55, + 0.785, + 0.592 + ], + "angle": 0, + "content": "0. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG) 39(6), 1-14 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.785, + 0.633 + ], + "angle": 0, + "content": "1. Li, H., Yang, X., Zhai, H., Liu, Y., Bao, H., Zhang, G.: Vox-surf: Voxel-based implicit surface representation. IEEE Transactions on Visualization and Computer Graphics (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.785, + 0.688 + ], + "angle": 0, + "content": "2. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.688, + 0.785, + 0.73 + ], + "angle": 0, + "content": "3. Liu, H.T.D., Williams, F., Jacobson, A., Fidler, S., Litany, O.: Learning smooth neural functions via lipschitz regularization. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-13 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.785, + 0.758 + ], + "angle": 0, + "content": "4. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics 21(4), 163-169 (1987)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "5. Ma, B., Zhou, J., Liu, Y.S., Han, Z.: Towards better gradient consistency for neural signed distance functions via level set alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17724-17734 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "6. Mehta, I., Chandraker, M., Ramamoorthi, R.: A level set theory for neural implicit evolution under explicit flows. In: European Conference on Computer Vision. pp. 711-729. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.176, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "Z. Hu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "17. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I. pp. 405-421 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.787, + 0.258 + ], + "angle": 0, + "content": "18. Munkberg, J., Hasselgren, J., Shen, T., Gao, J., Chen, W., Evans, A., Müller, T., Fidler, S.: Extracting triangular 3d models, materials, and lighting from images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8280-8290 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.258, + 0.787, + 0.298 + ], + "angle": 0, + "content": "19. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems 34, 6087-6101 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.787, + 0.338 + ], + "angle": 0, + "content": "20. Sun, J., Chen, X., Wang, Q., Li, Z., Averbuch-Elor, H., Zhou, X., Snavely, N.: Neural 3d reconstruction in the wild. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-9 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.787, + 0.393 + ], + "angle": 0, + "content": "21. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5481-5490. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.787, + 0.42 + ], + "angle": 0, + "content": "22. Vicini, D., Speierer, S., Jakob, W.: Differentiable signed distance function rendering. ACM Transactions on Graphics (TOG) 41(4), 1-18 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.787, + 0.46 + ], + "angle": 0, + "content": "23. Walker, T., Mariotti, O., Vaxman, A., Bilen, H.: Explicit neural surfaces: Learning continuous geometry with deformation fields. arXiv preprint arXiv:2306.02956 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.787, + 0.501 + ], + "angle": 0, + "content": "24. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems 34, 27171-27183 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.787, + 0.555 + ], + "angle": 0, + "content": "25. Wang, Y., Han, Q., Habermann, M., Daniilidis, K., Theobalt, C., Liu, L.: Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3295-3306 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.787, + 0.596 + ], + "angle": 0, + "content": "26. Wang, Y., Skorokhodov, I., Wonka, P.: Hf-neus: Improved surface reconstruction using high-frequency details. Advances in Neural Information Processing Systems 35, 1966-1978 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.787, + 0.623 + ], + "angle": 0, + "content": "27. Wang, Y., Skorokhodov, I., Wonka, P.: Improved surface reconstruction using high-frequency details. arXiv preprint arXiv:2206.07850 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.787, + 0.663 + ], + "angle": 0, + "content": "28. Wang, Y., Skorokhodov, I., Wonka, P.: Pet-neus: Positional encoding tri-planes for neural surfaces. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12598-12607 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.663, + 0.787, + 0.717 + ], + "angle": 0, + "content": "29. Worchel, M., Diaz, R., Hu, W., Schreer, O., Feldmann, I., Eisert, P.: Multiview mesh reconstruction with neural deferred shading. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6187-6197 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.787, + 0.759 + ], + "angle": 0, + "content": "30. Wu, T., Wang, J., Pan, X., Xu, X., Theobalt, C., Liu, Z., Lin, D.: Voxurf: Voxel-based efficient and accurate neural surface reconstruction. arXiv preprint arXiv:2208.12697 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.813 + ], + "angle": 0, + "content": "31. Yao, Y., Luo, Z., Li, S., Zhang, J., Ren, Y., Zhou, L., Fang, T., Quan, L.: Blended-mvs: A large-scale dataset for generalized multi-view stereo networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1790-1799 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "32. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.248, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Improving Surface Rendering via CPT with View-Reflection Appearance" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "33. Yariv, L., Hedman, P., Reiser, C., Verbin, D., Srinivasan, P.P., Szeliski, R., Barron, J.T., Mildenhall, B.: Bakedsdf: Meshing neural sdfs for real-time view synthesis. arXiv preprint arXiv:2302.14859 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.231 + ], + "angle": 0, + "content": "34. Yariv, L., Kasten, Y., Moran, D., Galun, M., Atzmon, M., Ronen, B., Lipman, Y.: Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems 33, 2492-2502 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.785, + 0.286 + ], + "angle": 0, + "content": "35. Zhang, J., Yao, Y., Li, S., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Critical regularizations for neural surface reconstruction in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6270-6279 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "36. Zhang, J., Yao, Y., Li, S., Luo, Z., Fang, T.: Visibility-aware multi-view stereo network. British Machine Vision Conference (BMVC) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.357 + ], + "angle": 0, + "content": "37. Zhang, J., Yao, Y., Quan, L.: Learning signed distance field for multi-view surface reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6525-6534 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.384 + ], + "angle": 0, + "content": "38. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "39. Zhang, Y., Zhu, J., Lin, L.: Fastmesh: Fast surface reconstruction by hexagonal mesh-based neural rendering. arXiv preprint arXiv:2305.17858 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.466 + ], + "angle": 0, + "content": "40. Zhang, Y., Hu, Z., Wu, H., Zhao, M., Li, L., Zou, Z., Fan, C.: Towards unbiased volume rendering of neural implicit surfaces with geometry priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4359-4368 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "41. Zhuang, Y., Zhang, Q., Feng, Y., Zhu, H., Yao, Y., Li, X., Cao, Y.P., Shan, Y., Cao, X.: Anti-aliased neural implicit surfaces with encoding level of detail. In: SIGGRAPH Asia 2023 Conference Papers. pp. 1-10 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.508 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CPT-VR_ Improving Surface Rendering via Closest Point Transform with View-Reflection Appearance/255265d5-d4ab-4d8b-8210-adf9209a31be_origin.pdf b/2024/CPT-VR_ Improving Surface Rendering via Closest Point Transform with View-Reflection Appearance/255265d5-d4ab-4d8b-8210-adf9209a31be_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..831ec0d12698005abcef4f417ff3ebf421a8b9fa --- /dev/null +++ b/2024/CPT-VR_ Improving Surface Rendering via Closest Point Transform with View-Reflection Appearance/255265d5-d4ab-4d8b-8210-adf9209a31be_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7aee02cda9738366c680f6741a19e01890ecf85ba779e60fdf0f1d40cf9bda41 +size 2516416 diff --git a/2024/CPT-VR_ Improving Surface Rendering via Closest Point Transform with View-Reflection Appearance/full.md b/2024/CPT-VR_ Improving Surface Rendering via Closest Point Transform with View-Reflection Appearance/full.md new file mode 100644 index 0000000000000000000000000000000000000000..95466d0f74b6209b0b7fabb9bf0635a1798993aa --- /dev/null +++ b/2024/CPT-VR_ Improving Surface Rendering via Closest Point Transform with View-Reflection Appearance/full.md @@ -0,0 +1,379 @@ +# CPT-VR: Improving Surface Rendering via Closest Point Transform with View-Reflection Appearance + +Zhipeng Hu $^{1,3}$ , Yongqiang Zhang $^{1,3,*}$ , Chen Liu $^{2}$ , Lincheng Li $^{3,*}$ , Sida Peng $^{3}$ , Xiaowei Zhou $^{3}$ , Changjie Fan $^{1}$ , and Xin Yu $^{2}$ + +$^{1}$ NetEase Fuxi AI Lab, Hangzhou, China {zhangyongqiang02,zphu,lilincheng,fanchangjie}@corp.netease.com + +2 The University of Queensland, Queensland, Australia + +chen.liu7@uqconnect.edu.au, xin.yu@uq.edu.au + +3 Zhejiang University, Hangzhou, China + +{pengsida,xwzhou}@zju.edu.cn + +Abstract. Differentiable surface rendering has significantly advanced 3D reconstruction. Existing surface rendering methods assume that the local surface is planar, and thus employ linear approximation based on the Singed Distance Field (SDF) values to predict the point on the surface. However, this assumption overlooks the inherently irregular and non-planar nature of object surfaces in the real world. Consequently, the approximate points tend to deviate from the zero-level set, affecting the fidelity of the reconstructed shape. In this paper, we propose a novel surface rendering method termed CPT-VR, which leverages the Closet Point Transform (CPT) and View and Reflection direction vectors to enhance the quality of reconstruction. Specifically, leveraging the physical property of CPT that accurately projects points near the surface onto the zero-level set, we correct the deviated points, thus achieving an accurate geometry representation. Based on our accurate geometry representation, incorporating the reflection vector into our method can facilitate the appearance modeling of specular regions. Moreover, to enable our method to no longer be dependent on any prior knowledge of the background, we present a background model to learn the background appearance. Compared to previous state-of-the-art methods, CPT-VR achieves better surface reconstruction quality, even for cases with complex structures and specular highlights. + +Keywords: Surface Rendering $\cdot$ 3D Reconstruction $\cdot$ Differentiable Rendering $\cdot$ Closest Point Transform + +# 1 Introduction + +Differentiable rendering plays a crucial role in 3D content creation, especially showing greater promise in multi-view 3D reconstruction and text-to-3D generation. Recent advancements have utilized differentiable rendering for attaining + +![](images/e7a287aca156de308693b5480b9aa7127e69a8ee7c9698378b1fddc5d3ed5121.jpg) +(a) IDR + +![](images/723c411c6796cdef18ca737e0debe2f4a918269a926ad53d9149b74c335df7e5.jpg) +$S_0$ : zero-level set +v: the view ray +$\mathbf{x}^{\prime}$ : the linear projection along the ray $\mathbf{v}$ + +![](images/aa034c72e6a9d6884d0a6d31ff401f50e92f4fa256aa890020c409ef469a57c5.jpg) +(b)DMTet + +![](images/927bf82a786c33bc9f3df6e502fe3e3b105a47b992f2d59cca3832fcce6d54d7.jpg) +x: the sampled point by sphere tracing $d$ : the SDF value of x, n: the normal vector of x +$S_0$ : zero-level set +$\mathbf{x}_i, \mathbf{x}_j$ : vertices of an edge where the SDF +changes sign +$d_{i},d_{j}$ :SDF values of $\mathbf{x}_i,\mathbf{x}_j$ +$\mathbf{x}_i^{\prime},\mathbf{x}_j^{\prime}$ : local deformed vertices + +![](images/80b2f5e2512588da28f2dc9ddee4384a6911a09974e5e2cf6944d77a7262f118.jpg) +(c) CPT-Ours + +![](images/c5abb21e4df1684c16d3070c03b02aa390d3103a04d06826b26328c452e82f9c.jpg) +$\mathbf{x}_{ij}^{\prime}$ : linear interpolation of $\mathbf{x}_i^{\prime},\mathbf{x}_j^{\prime}$ by $d_{i},d_{j}$ +$S_0$ : zero-level set +$\mathbf{x}_i, \mathbf{x}_j$ : vertices of an edge where the SDF changes sign +$d_{i},d_{j},d_{ij}$ : SDF values, $\mathbf{n}_{ij}$ : the normal vector of $\mathbf{x}_{ij}$ +$\mathbf{x}_{ij}$ : linear interpolation of $\mathbf{x}_i, \mathbf{x}_j$ by $d_i, d_j$ +$\mathbf{x}_{ij}^{\prime}$ : the projected vertex of $\mathbf{x}_{ij}$ by CPT +Fig. 1: Illustration of the surface point sampling methods and their corresponding construction results. (a) IDR [34] first employs a sphere tracing method to identify the sampled point $x$ and then leverages the SDF value as well as the normal vector to obtain the approximate point $x'$ . (b) DMTet [19] detects surfaces through the observation of sign changes in the SDF at the vertices of a tetrahedron. Then, it employs linear interpolation to approximate the points $x_{ij}'$ on the zero-level set. Compared with them, our method (c) utilizes CPT to project the interpolated points $x_{ij}$ on the zero-level set, and thus attain the transformed points $x_{ij}'$ . The approximate points $x'$ of IDR and $x_{ij}'$ of DMTet are deviated from the zero-level set, while $x_{ij}'$ obtained from our method exactly on the object surface. This implies our method is able to obtain a more accurate geometry representation and gradient backpropagation. + +the Singed Distance Field (SDF) and object appearance. These methods can be divided into two types: volume rendering and surface rendering. In volume rendering, SDF is first converted into density values that are then rendered via the NeRF framework [17, 38]. While these methods consistently generate reliable geometry reconstruction, the dense points sampling along the ray increases the rendering time cost. Conversely, surface rendering methods only sample one point along the ray, markedly optimizing rendering speeds. + +Existing surface rendering methods, such as IDR (Implicit Differentiable Renderer) [34] and DMTet (Deep Marching Tetrahedra) [19] achieve computational efficiency by assuming the object surface can be considered linear or at least can be approximated as linear within a local scope. Based on this assumption, they employ linear interpolation to determine the points at specific locations on the surface. As shown in Fig. 1, IDR employs sphere tracing [7] to identify points close to the surface along a ray and then relies on the SDF values and the normal vector to calculate the intersection points between rays and the surface. DMTet identifies the surfaces by observing changes in the sign of the SDF at the vertices of a tetrahedron and then leverages linear interpolation to approximate the surface points. However, as demonstrated in Fig. 1, their approximate points always deviate from the object surface (i.e., zero-level set), since the object surface is mostly irregular and non-planar. Hence, modeling geometry representations based on these deviated points would further lead to inaccurate gradient back- + +propagation, resulting in the reconstructed surfaces losing a lot of details, as illustrated by the reconstruction results in Fig. 1 (a) and (b). + +To address the above problem, we introduce a novel surface rendering approach, termed CPT-VR, which is grounded in the Closet Point Transform algorithm and augmented with View-Reflection appearance modeling. In a nutshell, in the geometry representation modeling process, we conduct the Closet Point Transform (CPT) for deviated points to enhance the geometry representation. In the appearance modeling process, we intend to incorporate the View and Reflection vectors and thus attain better appearance representation for cases with specular highlights. To the best of our knowledge, we are the first attempt to correct the deviated points via CPT in the surface rendering task. + +To be specific, in the geometry representation modeling process, we first devise a neural SDF network to compute SDF values for the entire 3D space. Based on the SDF values, we employ the Marching Cubes [14] to attain the vertices close to the surface of the object. As suggested in Fig. 1 (c), most of the obtained vertices are deviated from the object's surface rather than on it. Hence, we leverage CPT to calculate the closest point on the object surface for each deviated vertex, thus obtaining accurate and differentiable vertices. Subsequently, we utilize a feature-based rasterizer to render this triangle mesh in each view. In this fashion, we obtain the corresponding 3D position, geometry feature, and normal vector of each pixel in an image. + +In the appearance representation modeling process, we focus on improving the geometry quality for specular highlights. Existing methods such as IDR [34] and NeuS2 [25] rely solely on the view direction to capture the surface information of objects, which cannot handle the object with highlights well. Ref-NeRF [21] demonstrates that replacing the view direction with the reflection direction can mitigate the impact of highlights. However, since the reflection vector calculated based on the deviated points is inaccurate, directly incorporating these reflection vectors into existing surface rendering methods will lead to significant backpropagation errors. Consequently, the model will not only be unable to handle the specular highlights but will also diminish its capability to reconstruct non-specular areas. In contrast, benefiting from our accurate point sampling on the zero-level set based on CPT, incorporating view-reflection vectors in our framework can facilitate the appearance modeling of specular regions. + +Moreover, we found that existing surface rendering methods [1,22,34] require prior background knowledge, such as foreground masks, or the background image, to precisely process and render foreground objects, which constrains their flexibility and applicability in various scenes. To solve this problem, we present a background model to obtain the background appearance representation. Then we attain the final appearance modeling by blending the foreground and background appearance. Extensive experiments conducted on DTU [9] and BlendedMVS [31] demonstrate that our method achieves state-of-the-art geometry quality and surface rendering. Additionally, we also illustrate that our method is effective and reliable in cases with complex structures and specular highlights. In summary, our contributions are three-fold: + +- We present a novel differentiable surface rendering method CPT-VR which shows superior performance in reconstructing the detailed structure of objects. With the CPT algorithm, we can address the approximate point deviation issue caused by linear interpolation. +- We incorporate the view and reflection vectors to model the foreground appearance representation, facilitating our method to be robust against specular highlights. +- We develop a simple background model to generate the background appearance representation, enabling our method to no longer be dependent on any prior background knowledge. + +# 2 Related Work + +Surface Rendering with SDF. IDR [34] introduces a system that learns surfaces in a self-supervised manner. It first identifies points close to the surface via sphere tracing and then finds approximate intersection points on the surface by projecting them along the view ray. Although IDR can learn the 3D shape and appearance from images, it relies heavily on accurate masks. To mitigate this issue, MVsDF [37] supervises the SDF network with depth maps generated by Vis-MVSNET [36]. It also uses relaxed masks generated by probability maps to learn a surface indicator. RegSDF [35] further introduces the Hessian regularization and the minimal surface constraint. Some works [1, 22] use specialized reparameterization techniques and design weighting functions along a ray based on SDF values, thus they can propagate gradients to intermediate points from sphere tracing. These reparameterization-based methods are similar to methods based on volume rendering, but they still need a given background: either black or a pre-defined environment map. + +Rasterization-based Rendering. Recently, there have been attempts to utilize the differentiable rasterizer for surface rendering. NVDIFFREC [18] and NVDIFFRECmc [8] adapt DMTet [19] to extract a mesh from an implicit SDF and then render the mesh by the differentiable rasterizer Nvdiffrast [10]. While these methods are efficient, their reliance on the linear approximation in DMTet limits their ability to recover shapes with complex structures. Meanwhile, NIE [16] propagates neural implicit surfaces according to a flow field from a certain energy function. ENS [23] combines neural deformation fields with rasterization rendering, allowing deform an initial shape into a target surface progressively. Moreover, NDS [29] and FastMESH [39] focus on optimizing a rough mesh derived from priors through differentiable rasterization. + +Volume Rendering of SDF. Several methods [2, 20, 24, 32] have been developed for learning SDF through volume rendering. VolSDF [32] and NeuS [24] leverage Laplace or Logistic cumulative distribution function of SDF to represent the volume density and sample dense points along a ray based on SDF values. Based on these methods, some works [2, 3, 12, 26–28, 40, 41] aim to further enhance the geometry details. Among them, Geo-Neus [3] and TUVR [40] improve SDF accuracy by introducing additional supervision from point clouds. On the training efficiency front, Voxurf [30] and Vox-Surf [11] speed up the training process + +through voxel-based representations. NeuS2 [25] speeds up NeuS [24] by simplifying the second-order derivatives calculation of the SDF network constructed based on a multi-resolution hash encoding. + +Hybrid Volume-rasterization Rendering. Recently, some rendering methods have been developed based on a hybrid volume-rasterization way. BakedSDF [33] first learns the SDF and appearance of a scene based on VolSDF [32] and transforms the SDF into a mesh. Then the transformed mesh is refined using a rasterizer pipeline. VMesh [6] takes a comprehensive approach by initially training a combination of a neural SDF and a density field through volume rendering. Then the mesh undergoes further refinement by DMTet [19] and a differentiable rasterizer [10], ensuring accurate surface representation. + +# 3 Method + +# 3.1 Preliminaries + +Signed Distance Field (SDF) is a mathematical function $f$ that measures the shortest distance from a point $\mathbf{x}$ to an object surface, indicating whether the point is outside (positive distance) or inside (negative distance) the surface. The surface is identified by the zero-level set $S$ , where SDF equals to zero, providing a concise description of the object's surface: + +$$ +\mathcal {S} = \left\{\mathbf {x} \in \mathbb {R} ^ {3} | f (\mathbf {x}) = 0 \right\}. \tag {1} +$$ + +Neural SDF denotes a neural network that models SDFs for 3D objects. In this work, we employ SDFnet which leverages the Multilayer Perceptrons (MLPs) enhanced with a multi-resolution hash encoding to model SDF [25]: + +$$ +(d, \mathbf {z}) = f _ {\theta} (\mathbf {x}, h _ {\Omega} (\mathbf {x})). \tag {2} +$$ + +Here, $d$ is the predicted SDF value, $\mathbf{z}$ signifies the predicted geometry feature, $f_{\theta}$ denotes an MLP parameterized by weights $\theta$ , and $h_{\Omega}(\mathbf{x})$ is the multi-resolution hash encoding of the input point $\mathbf{x}$ . Moreover, the normal vector $\mathbf{n}$ of $\mathbf{x}$ is derived as the normalized gradient of the implicit surface: + +$$ +\mathbf {n} = \nabla_ {\mathbf {x}} d / \| \nabla_ {\mathbf {x}} d \|. \tag {3} +$$ + +To ensure training stability, our SDFnet employs an optimization schedule that progressively incorporates the level of detail and the numerical gradients [12]. + +# 3.2 Geometry Representation Modeling + +Geometry Representation via SDF. As suggested in Fig. 2, we first leverage our SDFnet to compute the SDF values for the entire 3D space. Based on SDF values, we utilize the isosurface algorithm Marching Cubes to extract and represent the triangle mesh $\mathcal{G} = (\mathcal{V},\mathcal{F})$ , where $\mathcal{V}$ and $\mathcal{F}$ indicate the vertices and faces respectively. However, this method assumes that the surface changes + +![](images/26875512fec8bc781bf49e5624a3f02723133c86a23d48a02078e4b3b5828188.jpg) + +![](images/5dc2c214efe8bfac2d39fd250aa4ab30e32109455e5872013c11439d5411f6e5.jpg) +Fig. 2: Illustration of our proposed method CPT-VR. In (a), we first extract the mesh via Marching Cubes, which does not involve gradient propagation. Then we leverage CPT to project the approximate points on the zero-level set (b). Subsequently, we render the transformed mesh via a rasterizer in (c). Specifically, we incorporate the view-reflection vectors into the foreground appearance modeling stage via RGBNet. Additionally, we employ the Background Model to attain the background information. Finally, we integrate the foreground and background appearance to generate the rendering results. We adopt four supervisions to optimize these neural networks. + +linearly inside the voxel, overlooking the complex convex structure of the object surface. Hence, the linearly interpolated vertices tend to deviate from the zero-level set and the triangle mesh may not accurately reflect the complex structure of the object surface. + +Closest Point Transform. To project these deviated vertices onto the zero-level set, we devise the closet set transform (CPT) to SDF. Specifically, CPT first utilizes the norm vector $\mathbf{n}$ within the signed distance field to identify the projected direction of a deviated point. It then moves the deviated point along the opposite direction of $\mathbf{n}$ by a distance $d$ equal to its SDF value. Mathematically, for a vertex $\mathbf{x} \in \mathcal{V}$ , this transformation process is expressed as follows: + +$$ +\mathbf {x} ^ {\prime} = \mathbf {x} - d \mathbf {n}. \tag {4} +$$ + +After that, we obtain the transformed mesh $\mathcal{G}' = (\mathcal{V}',\mathcal{F})$ + +# 3.3 Feature-based Rasterisation + +After obtaining the transformed mesh, we first employ a feature-based differentiable rasterizer Nvdiffrast [10] to render triangle meshes in each view. Specifically, as suggested in Fig. 2 (c), for a pixel $p(x,y)$ , its projected primitive triangle face within the mesh is represented as $(\mathbf{x}_0',\mathbf{x}_1',\mathbf{x}_2')$ . The rasterizer computes the barycentric coordinates $u,v$ , depth $D_{p}$ , and visible mask $M_{p}$ for this pixel as + +follows: + +$$ +(u, v, D _ {p}, M _ {p}) = \mathcal {R} \left(\mathcal {G} ^ {\prime}, p (x, y), K, R, \mathbf {t}\right), \tag {5} +$$ + +where $\mathcal{R}$ denotes the rasterizer, $\mathcal{G}'$ represents the transformed mesh, and $K$ and $R, \mathbf{t}$ are the camera's intrinsic and extrinsic parameters, respectively. + +Subsequently, each projected vertex $\mathbf{x}'$ from the set $\mathcal{V}'$ is input into SDFnet. SDFnet outputs $(d',\mathbf{z}',\mathbf{n}')$ , encapsulating the distance, geometric feature, and the normal vector associated with each vertex. These outputs are then interpolated for a pixel $p(x,y)$ based on its barycentric coordinates $u,v$ : + +$$ +\mathbf {x} _ {p} = u \mathbf {x} _ {0} ^ {\prime} + v \mathbf {x} _ {1} ^ {\prime} + (1 - u - v) \mathbf {x} _ {2} ^ {\prime}, +$$ + +$$ +\mathbf {z} _ {p} = u \mathbf {z} _ {0} ^ {\prime} + v \mathbf {z} _ {1} ^ {\prime} + (1 - u - v) \mathbf {z} _ {2} ^ {\prime}, \tag {6} +$$ + +$$ +\mathbf {n} _ {p} = u \mathbf {n} _ {0} ^ {\prime} + v \mathbf {n} _ {1} ^ {\prime} + (1 - u - v) \mathbf {n} _ {2} ^ {\prime}, +$$ + +where $(\mathbf{x_p},\mathbf{z_p},\mathbf{n_p})$ denote the interpolated features. + +# 3.4 Appearance Representation Modeling + +In this work, we divide the appearance modeling process into two parts: + +Foreground Appearance Modeling based on View-Reflection Vectors. To better capture the appearance of objects with specular highlights, we incorporate the view direction and reflection direction into the foreground modeling process. Specifically, we employ the RGBnet $g_{\gamma}^{f}$ , a Multilayer Perceptron (MLP) with weights $\gamma$ to determine the foreground color. RGBnet takes the interpolated 3D features $(\mathbf{x_p}, \mathbf{z_p}, \mathbf{n_p})$ , view direction $\mathbf{v}_p$ , and reflection direction $\mathbf{r}_p$ as input, and outputs the RGB color $\mathbf{c}_p^f$ for the foreground. The process is formulated as: + +$$ +\mathbf {c} _ {p} ^ {f} = g _ {\gamma} ^ {f} \left(\mathbf {x} _ {\mathbf {p}}, \mathbf {z} _ {\mathbf {p}}, \mathbf {n} _ {\mathbf {p}}, Y (\mathbf {v} _ {\mathbf {p}}), Y (\mathbf {r} _ {\mathbf {p}})\right), \tag {7} +$$ + +where $Y(\cdot)$ denotes Spherical Harmonics encoding. Here, the reflection direction $\mathbf{r_p}$ is defined as the mirror vector of the view direction $\mathbf{v_p}$ relative to the normal vector $\mathbf{n_p}$ , calculated as: + +$$ +\mathbf {r} _ {\mathbf {p}} = 2 \left(- \mathbf {v} _ {\mathbf {p}} \cdot \mathbf {n} _ {\mathbf {p}}\right) \mathbf {n} _ {\mathbf {p}} + \mathbf {v} _ {\mathbf {p}}. \tag {8} +$$ + +Spherical harmonics are orthogonal bases on the unit sphere, and their mathematical definition and physical properties make them ideal for encoding direction frequencies compared to 3D space (xyz) frequencies. + +Background Appearance Modeling. To enable the model to reconstruct objects without any background priors, we present a one-point background model composed of a Multilayer Perceptron (MLP) $g_{\pi}^{b}$ with weights $\pi$ , alongside an additional multi-resolution hash encoding $h_{\Omega}^{b}$ . The modeling process of the background appearance $\mathbf{c}_p^b$ is formulated as: + +$$ +\mathbf {c} _ {p} ^ {b} = g _ {\pi} ^ {b} \left(\mathbf {x} _ {p} ^ {b}, h _ {\Omega} ^ {b} \left(\mathbf {x} _ {p} ^ {b}\right), Y \left(\mathbf {v} _ {p}\right)\right), \tag {9} +$$ + +where $\mathbf{x}_p^b$ represents the background sampling point, determined by the camera location $\mathbf{x}_o$ , the veiw direction $\mathbf{v}_p$ and the far distance $F_{p}$ of the sampling + +boundary. $\mathbf{x}_p^b$ is mathematically expressed as: + +$$ +\mathbf {x} _ {p} ^ {b} = \mathbf {x} _ {o} + 2 F _ {p} \mathbf {v} _ {p}. \tag {10} +$$ + +Following NeuS [24], the foreground is normalized within a unit sphere centered at $\mathbf{0}$ . As indicated in Fig. 3, the far distance $F_{p}$ is calculated as: + +$$ +m = - \mathbf {x} _ {o} \cdot \mathbf {v} _ {p}; F _ {p} = m + 1. \tag {11} +$$ + +This approach facilitates a more versatile background representation, compensating for the absence of masks. + +Appearance Integration. Finally, we integrate the background appearance $\mathbf{c}_p^b$ and foreground appearance $\mathbf{c}_p^f$ to attain the rendered color, defined as: + +$$ +\mathbf {c} _ {p} = M _ {p} \mathbf {c} _ {p} ^ {f} + (1 - M _ {p}) \mathbf {c} _ {p} ^ {b}. \tag {12} +$$ + +Overall, our method enables surface rendering of objects without any known background information, offering greater flexibility and adaptability. + +![](images/ba50007b33ae024c00c2145e20c314139d4244190f6134a11800937f98841ac7.jpg) +Fig.3: Illustration of the one-point background sampling. + +# 3.5 Training Objectives + +Color loss. We optimize the CPT-VR model by ensuring the colors it generates match real image colors as closely as possible. This is done by calculating the L1 error, which is the average difference between the colors $\hat{\mathbf{c}}_i$ produced by the model and the ground-truth colors $\mathbf{c}_i$ for all pixels $P$ : + +$$ +\mathcal {L} _ {C} = \frac {1}{P} \sum_ {i} ^ {P} \| \hat {\mathbf {c}} _ {i} - \mathbf {c} _ {i} \| _ {1}. \tag {13} +$$ + +Regularization of SDFnet. A well-defined neural SDF exhibits Lipschitz continuity [13] and its gradient adheres to the Eikonal equation [4], ensuring smoothness and stability of the generated field. A valid SDF field holds the property that the normal of a point near the surface and the normal of its projected point remain consistent. For effective closest point transform, we directly constrain our SDFnet based on this normal consistency property [15]. Specifically, it aligns the normal of a deviated point $\mathbf{n}$ with the normal $\mathbf{n}'$ of its projected point, enhancing stability for surface rendering via the closest point transform of SDF. Mathematically, it is represented by: + +$$ +\mathcal {L} _ {n c} = 1 - \mathbf {n} \cdot \mathbf {n} ^ {\prime}. \tag {14} +$$ + +Mask Loss. If the mask of the foreground region is available, we can further refine the rendering effect of CPT-VR by including a mask loss. The mask constraint is formulated as: + +$$ +\mathcal {L} _ {M} = \frac {1}{P} \sum_ {i} ^ {P} \left(\hat {M} _ {i} - M _ {i}\right) ^ {2}, \tag {15} +$$ + +where the $\hat{M}_i$ and the $M_{i}$ represent the predicted mask and the ground truth mask, respectively. + +Depth loss. Inspired by previous works [37] and [35], we utilize depth maps generated by the pre-trained network CasMVSNet [5] to guide our model. We minimize the squared difference (MSE) between the rendered depths $(\hat{D}_i)$ and those from the pre-trained network $(D_i)$ , weighted by the confidence $prob_i$ from CasMVSNet, as follows: + +$$ +\mathcal {L} _ {D} = \frac {1}{P} \sum_ {i} ^ {P} p r o b _ {i} \left(\hat {D} _ {i} - D _ {i}\right) ^ {2}, \tag {16} +$$ + +where the confidence ranges from 0 to 1, and depths considered unreliable are given a confidence of 0. + +Total loss. Our final metric combines all these aspects, adjusting their importance with specific weights $\lambda$ : + +$$ +\mathcal {L} = \lambda_ {r g b} \mathcal {L} _ {r g b} + \lambda_ {n c} \mathcal {L} _ {n c} + \lambda_ {M} \mathcal {L} _ {M} + \lambda_ {D} \mathcal {L} _ {D}. \tag {17} +$$ + +Here, $\lambda_{rgb}$ , $\lambda_{nc}$ , $\lambda_M$ , and $\lambda_D$ are empirically set to 3, 0.01, 1, and 1, respectively. + +# 4 Experiments + +# 4.1 Experimental Settings + +Implementation details. For the DTU dataset, we train CPT-VR for 10,000 iterations with a batch size of one image, targeting a Marching Cubes resolution of $256^3$ . In each training batch, we render one view at a resolution of 800 $\times$ 600 pixels. Note that, when we only utilize depth loss and color loss for supervision, we employ a higher Marching Cubes resolution of $512^3$ and render images at $1600 \times 1200$ pixels. For BlendedMVS, CPT-VR is trained for 20,000 iterations with a batch size of one image and a Marching Cubes resolution of $256^3$ , rendering images at $768 \times 576$ pixels. Moreover, the initial learning rate of our model is set at 0.005, with a final epoch decay factor of 0.1 for DTU and 0.005 for BlendedMVS. More details about the architecture and the isosurface configuration are provided in the supplementary material. + +Datasets. All experiments are conducted on DTU [9] and BlendedMVS [31]. Objects in DTU are collected in a controlled lab setting, with ground truth data that includes images, masks, and point clouds. For our training and evaluation, we leverage 15 objects identified by IDR with challenges such as specular highlights, delicate structures, and areas hidden from view. Meanwhile, BlendedMVS encompasses a diverse range of objects and scenes captured in natural settings. From this dataset, we select 7 difficult objects as NeuS [24] and the 'Bull' (with specular highlights) to further assess the model performance in reconstructing objects against complex backgrounds. + +# 4.2 Quantitative Evaluation + +Analysis on DTU. We benchmark our method against the state-of-the-art methods on the DTU dataset. DTU dataset contains 128 scans captured in the laboratory. In this work, we evaluate the surface mesh using the same set of scans as in [34,37]. As denoted in Tab. 2, our method consistently outperforms other surface rendering methods on most scans. Note that, our method achieves the lowest + +value in mean Chamfer Distance (0.59), which is $16.9\%$ lower than the second-best surface rendering method RegSDF (0.71). Since SDFnet in our method has a similar structure to it in NeuS2, we also compare NeuS2 and recent volume rendering methods [26, 41] with CPT-VR. As demonstrated in Tab. 2, whether the two methods incorporate background information from the object mask $(\mathrm{w} / \mathrm{M})$ or use our proposed one-point background model $(\mathrm{w} / \mathrm{BG1})$ to attain background information, our method surpasses NeuS2 on the vast majority of ScanIDs. + +Table 1: Comparison of the average rendering time on DTU. The rendered image resolution of all methods is $1600 \times 1200$ , and the resolution of Marching Cubes is $512^{3}$ . The rendering time of our method is significantly lower than that of others. + +
MethodRendering Time (s)
IDR27.233
NeuS2 (w /M)1.979
Ours (w/ M)0.057
Ours (w/ BG1)0.104
+ +Table 2: Quantitative results on DTU dataset. Here, we evaluate all methods on Chamfer Distance (CD). The lower the CD value, the better the performance. For a fair comparison, we also present the results of the volume rendering method NeuS2 since our SDFnet has a similar structure to it. "w/ M" means providing the prior background knowledge by masks, while "w/ BG1" means employing our one-point background model without any prior background knowledge. The best and second-best results are highlighted by different colors. The results illustrated that CPT-VR attains better performance among most ScanIDs. + +
ScanID243740556365698397105106110114118122Mean
DMTet [19]1.841.731.050.611.661.370.681.283.821.070.911.150.370.540.551.24
IDR [34]1.631.870.630.481.040.790.771.331.160.760.670.900.420.510.530.90
MVSDF [37]0.831.760.880.441.110.900.751.261.021.350.870.840.340.470.460.89
FastMESH [39]0.651.480.570.401.480.770.560.860.840.940.720.810.520.490.540.77
RegSDF [35]0.591.410.630.421.340.620.590.890.911.020.600.590.290.400.380.71
HF-NeuS [26] (w/ M)1.111.280.610.470.970.680.621.340.910.730.531.820.380.540.510.83
LOD-NeuS [41] (w/ M)0.650.910.370.481.050.870.821.220.950.690.561.300.420.580.570.76
NeuS2 [25] (w/ M)0.560.760.490.370.920.710.761.221.080.630.590.890.400.480.550.70
NeuS2 [25] (w/ BG1)0.690.860.830.341.050.680.651.051.120.670.581.410.370.500.520.75
Ours (w/ M)0.550.720.350.380.850.590.540.810.830.670.530.590.320.380.370.56
Ours (w/ BG1)0.430.730.360.320.930.610.610.891.020.680.480.730.320.370.370.59
+ +Moreover, we also measure the rendered results with PSNR to the ground truths. As indicated in Tab. 3, with the introduction of background information (BG1: $\pmb{x}$ ), our method achieves the highest PSNR values on most ScanIDs. Additionally, it can be observed that when using the one-point background model, both NeuS2 and CPT-VR are able to achieve similar or better performance compared to leveraging background information from masks (BG1: $\checkmark$ ). + +Table 3: Quantitative comparison of 2D view synthesis on DTU dataset. We measure all methods based on the metric PSNR. A higher PSNR indicates better synthesis quality. Here, B1 represents whether to introduce extra background knowledge into the method. $\times$ means to leverage the background information from masks, while $\checkmark$ indicates utilizing our one-point background model without introducing any background information. We color code the best and second-best results. Compared to NeuS2, our method achieves much better appearance reconstruction results. + +
ScanIDBG1243740556365698397105106110114118122Mean
NeuS2✗28.4427.1429.7029.6731.7527.8324.8431.2426.8630.5726.0528.9328.9827.8232.4828.82
NeuS2✓27.7124.9927.5128.8920.5530.1429.0127.7026.1631.0933.1030.6930.4832.7734.1028.99
Ours✗28.1527.3128.5429.5234.0429.8329.0035.9229.3233.0829.4531.1630.7730.5038.3831.00
Ours✓27.3025.5926.8827.7730.9130.0629.2828.0526.6828.9232.0330.1829.0234.0234.0229.38
+ +Table 4: Quantitative comparison on BlendedMVS dataset. We measure the method on the metric Chamfer Distance (CD). The lower CD values indicate better performance. For a fair comparison, we adopt our one-point background model to NeuS2. Compared to NeuS2, CPT-VR attains better geometry reconstruction results among most of the categories. + +
MethodBearClockDogDurianManSculptureBullStoneMean
NeuS20.380.470.950.470.460.400.650.380.52
Ours0.350.450.680.480.460.390.450.380.45
+ +Analysis on BlendedMVS. As illustrated in Tab. 4, we compare our method with NeuS2 on the eight selected categories with complex backgrounds. Since the scales are unknown in BlendedMVS, we scale the bounding box of each object to 100 units to measure their performance on Chamfer Distance (CD). CPT-VR achieves the best performance in the categories of 'Bear', 'Clock', 'Dog', 'Sculpture', and 'Bull', with the CD value for 'Man' and 'Stone' being on par with NeuS2, and only 0.01 behind NeuS2 for 'Durian'. Moreover, CPT-VR outperforms NeuS2 by $14.8\%$ . These results demonstrate the advantage of our method in reconstructing objects within complex environments. + +Analysis of Method Efficiency. We evaluate the rendering time of IDR, NeuS2, and CPT-VR on DTU. The results in Tab. 1 suggest the rendering time of IDR (24.233s) and NeuS2 (1.97s) is lengthy. Compared with them, our method achieves real-time rendering under the cases without introducing background priors being about 19 times faster than NeuS2. This is primarily because NeuS2 utilizes the dense sampling method in the volume rendering process. + +# 4.3 Qualitative Evaluation + +Impact of Specular Highlights. Fig. 4 illustrates the qualitative results on the DTU dataset. We compare our CPT-VR(w/M) model with IDR [34], DMTet [19], MVsDF [37], and NeuS2 [25]. As shown by the close-up regions in Fig. 4, the specular regions constructed by other methods exhibit obvious indentations, while our method can construct these regions with high quality. + +Impact of Complex Backgrounds. Fig. 5 demonstrates the visual results of reconstructing objects with complex backgrounds. Since existing surface render- + +ing methods cannot handle the cases without any background prior, we specifically evaluate NeuS2 and our method on BlendedMVS. As the elliptical area demonstrated, our method can effectively model the geometry of foreground objects without introducing background information, particularly the details on the surface of objects. Moreover, the rendering results of our method closely resemble the input image. + +![](images/59960e7ac11b8e536f954b9cbb78f2ccc369a13f0cee48625ae01bd4e4627fcc.jpg) +Fig. 4: Qualitative comparisons on objects in DTU with specular highlights. The close-ups in yellow highlight the specific regions for comparison. The geometry reconstruction results demonstrate that our method is robust against specular highlights. + +![](images/594bbbeb7067c1830b4433ee94e0d1d2fa430c9536d6c0267e63678028ea101f.jpg) +Fig.5: Qualitative results of objects with complex background on BlendedMVS. We outline the close-up regions for detailed comparison. It is evident that without the mask guidance, our method also exhibits strong capabilities in modeling foreground objects, especially for the finer-grained surface structures. + +Impact of the Sampling Point Amount. Tab. 5 suggests the impact of varying the number of sampling points and their distances. The optimal result is achieved when sampling one point at $2F_{p}$ , and the $CD$ is 0.66. Conversely, sampling either too closely (at $F_{p}$ ) or too distantly (at $1000F_{p}$ ) causes the sampled points to stray outside the desired background range, leading to suboptimal + +results with CD values of 0.68 and 0.73, respectively. Additionally, increasing the number of sampling points in our approach tends to accentuate background details, thereby negatively impacting the accuracy of foreground object modeling. More background points would render the model to prioritize background fitting over the foreground, thus generating inferior results. + +# 4.4 Ablation Study + +As shown in Tab. 6, we set five settings to demonstrate the superiority of our method design. From the results of CPT-V-M, CPT-R-M, and CPT-VR-M, it is evident that introducing view-ray directions brings greater benefits to geometry reconstruction than introducing the two vectors alone. Meanwhile, the comparable $CD$ values of CPT-VR-M and CPT-VR-BG1 demonstrate that the one-point background model can largely eliminate the negative impact of complex backgrounds in images on rendering the target object, enabling the model to + +Table 5: Impact of the Number of Background Points. Here, Num. indicates the sampling amount of points. Pos. is the sampling positions of points along the ray direction. $F_{p}$ is defined in Sec. 3.4. All settings are evaluated on Chamfer Distance (CD). Sampling one point at $2F_{p}$ attains the best performance. + +
MethodNum.Pos.CD↓
CPT-VR22Fp, 1000Fp0.74
CPT-VR2Fp, 2Fp0.67
CPT-VR11000Fp0.73
CPT-VR1Fp0.68
CPT-VR12Fp0.66
+ +no longer be dependent on any background knowledge. Moreover, the results of CPT-VR-BG1 and CPT-VR-BG1-D also verify that introducing depth supervision can further boost the reconstruction quality of our method. + +![](images/3b474c590a27cd876841f7b07c5181b6372e93c3af2fbd865472b83149b077e7.jpg) + +![](images/cd4d922a9c87d431a84ed5e55367b084b0ecf1ab2916371c62e8588c1eb141a4.jpg) +Fig.6: Qualitative results of ablation studies on DTU. The close-ups in red highlight the specific regions for comparison. The content in the bracket is the Chamfer Distance, indicating the reconstruction quality. Incorporating the reflection vector enhances the ability of the model to simulate the appearance representation on specular regions. In addition, when utilizing the background model to attain the background information, introducing depth supervision greatly improves the reconstruction quality. + +![](images/5e81aaf8d80130cf46e63e2a718e0b07b2d9929bec0a21d9ce83de1d3c78b2ca.jpg) + +![](images/cb4ba2cecde4e49a0b71a9e1e5f252a169b7614ff6c704d9b862fbaee26d9440.jpg) + +![](images/d0cb822b8e330a8debf845bdfa826f12d9fbf8d5c8cdcd7d4784f4e776c6a67d.jpg) + +![](images/9e5be864a994b94fdcd4c0842cbbb2df6565090fb3590b1f23cdb48821a071e5.jpg) + +Table 6: Ablation study of our design choices. We evaluate all designs on Chamfer Distance (CD). Results demonstrate the effectiveness of all our design choices and the importance of using accurate geometry representation when adding the reflection vector for appearance modeling. + +
SettingsDesign ChoicesCD↓
View VectorReflection VectorBackground ModelLCLMLD
CPT-V-M✓✓✓0.72
CPT-R-M✓✓✓0.83
CPT-VR-M✓✓✓✓0.65
CPT-VR-BG1✓✓✓✓0.66
CPT-VR-BG1-D✓✓✓✓✓0.59
+ +As illustrated in Sec. 1, we argue that only when the surface sampling points are accurate can the introduction of the reflection vector achieve better appearance modeling of specular regions. To further demonstrate that, we incorporate the reflection vector into the surface rendering method DMTet [19] and IDR [34]. Tab. 7 indicates that introducing the reflection vector reduces the geometry reconstruction ability of DMTet ( $CD$ rises from 1.24 to 1.63) and IDR ( $CD$ rises from 0.90 to 1.68). Conversely, the sampling surface points of our method with CPT are more accurate. The accurate points facilitate the modeling of the reflection vector, avoiding a greater backpropagation error. Thus, incorporating the reflection vector into our method enables our model more robust against specular highlights. Fig. 6 also demonstrates the superiority of our design choices. + +Table 7: Impact of view-reflection vectors. All methods are evaluated with masks on Chamfer Distance (CD). With the reflection vector, the performance of our method has significantly improved, while DMTet and IDR become worse. + +
MethodIDRIDRDMTetDMTetOursOursOurs
View Vector (V)✓✓✓✓✓✓
Reflection Vector (R)✓✓✓✓
CD↓0.901.681.241.630.720.830.65
+ +# 5 Conclusions + +In this paper, we present a novel surface rendering method CPT-VR. Based on the CPT algorithm, we correct the deviated points approximated by linear interpolation, enhancing the geometric accuracy. Meanwhile, benefiting from the accurate geometry representation, we incorporate view-reflection vectors into the appearance molding process, which enables our method against the specular highlights. Moreover, we devise a background model to enable our model to handle cases with complex backgrounds without any background knowledge. Experiments on the popular 3D reconstruction datasets demonstrate the superiority of our method in surface rendering, especially for cases with complex structures and specular highlights. We hope that our designs can provide some insights for future works. + +# References + +1. Bangaru, S.P., Gharbi, M., Luan, F., Li, T.M., Sunkavalli, K., Hasan, M., Bi, S., Xu, Z., Bernstein, G., Durand, F.: Differentiable rendering of neural sdfs through reparameterization. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022) +2. Darmon, F., Bascle, B., Devaux, J.C., Monasse, P., Aubry, M.: Improving neural implicit surfaces geometry with patch warping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6260-6269 (2022) +3. Fu, Q., Xu, Q., Ong, Y.S., Tao, W.: Geo-neus: geometry-consistent neural implicit surfaces learning for multi-view reconstruction. arXiv preprint arXiv:2205.15848 (2022) +4. Gropp, A., Yariv, L., Haim, N., Atzmon, M., Lipman, Y.: Implicit geometric regularization for learning shapes. In: International Conference on Machine Learning. pp. 3789-3799. PMLR (2020) +5. Gu, X., Fan, Z., Zhu, S., Dai, Z., Tan, F., Tan, P.: Cascade cost volume for high-resolution multi-view stereo and stereo matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2495-2504 (2020) +6. Guo, Y.C., Cao, Y.P., Wang, C., He, Y., Shan, Y., Qie, X., Zhang, S.H.: Vmesh: Hybrid volume-mesh representation for efficient view synthesis. arXiv preprint arXiv:2303.16184 (2023) +7. Hart, J.C.: Sphere tracing: A geometric method for the antiaised ray tracing of implicit surfaces. The Visual Computer 12(10), 527-545 (1996) +8. Hasselgren, J., Hofmann, N., Munkberg, J.: Shape, light, and material decomposition from images using monte carlo rendering and denoising. Advances in Neural Information Processing Systems 35, 22856-22869 (2022) +9. Jensen, R., Dahl, A., Vogiatzis, G., Tola, E., Aanaes, H.: Large scale multi-view stereopsis evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 406-413 (2014) +0. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG) 39(6), 1-14 (2020) +1. Li, H., Yang, X., Zhai, H., Liu, Y., Bao, H., Zhang, G.: Vox-surf: Voxel-based implicit surface representation. IEEE Transactions on Visualization and Computer Graphics (2022) +2. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023) +3. Liu, H.T.D., Williams, F., Jacobson, A., Fidler, S., Litany, O.: Learning smooth neural functions via lipschitz regularization. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-13 (2022) +4. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics 21(4), 163-169 (1987) +5. Ma, B., Zhou, J., Liu, Y.S., Han, Z.: Towards better gradient consistency for neural signed distance functions via level set alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17724-17734 (2023) +6. Mehta, I., Chandraker, M., Ramamoorthi, R.: A level set theory for neural implicit evolution under explicit flows. In: European Conference on Computer Vision. pp. 711-729. Springer (2022) + +17. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I. pp. 405-421 (2020) +18. Munkberg, J., Hasselgren, J., Shen, T., Gao, J., Chen, W., Evans, A., Müller, T., Fidler, S.: Extracting triangular 3d models, materials, and lighting from images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8280-8290 (2022) +19. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems 34, 6087-6101 (2021) +20. Sun, J., Chen, X., Wang, Q., Li, Z., Averbuch-Elor, H., Zhou, X., Snavely, N.: Neural 3d reconstruction in the wild. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-9 (2022) +21. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5481-5490. IEEE (2022) +22. Vicini, D., Speierer, S., Jakob, W.: Differentiable signed distance function rendering. ACM Transactions on Graphics (TOG) 41(4), 1-18 (2022) +23. Walker, T., Mariotti, O., Vaxman, A., Bilen, H.: Explicit neural surfaces: Learning continuous geometry with deformation fields. arXiv preprint arXiv:2306.02956 (2023) +24. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems 34, 27171-27183 (2021) +25. Wang, Y., Han, Q., Habermann, M., Daniilidis, K., Theobalt, C., Liu, L.: Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3295-3306 (2023) +26. Wang, Y., Skorokhodov, I., Wonka, P.: Hf-neus: Improved surface reconstruction using high-frequency details. Advances in Neural Information Processing Systems 35, 1966-1978 (2022) +27. Wang, Y., Skorokhodov, I., Wonka, P.: Improved surface reconstruction using high-frequency details. arXiv preprint arXiv:2206.07850 (2022) +28. Wang, Y., Skorokhodov, I., Wonka, P.: Pet-neus: Positional encoding tri-planes for neural surfaces. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12598-12607 (2023) +29. Worchel, M., Diaz, R., Hu, W., Schreer, O., Feldmann, I., Eisert, P.: Multiview mesh reconstruction with neural deferred shading. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6187-6197 (2022) +30. Wu, T., Wang, J., Pan, X., Xu, X., Theobalt, C., Liu, Z., Lin, D.: Voxurf: Voxel-based efficient and accurate neural surface reconstruction. arXiv preprint arXiv:2208.12697 (2022) +31. Yao, Y., Luo, Z., Li, S., Zhang, J., Ren, Y., Zhou, L., Fang, T., Quan, L.: Blended-mvs: A large-scale dataset for generalized multi-view stereo networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1790-1799 (2020) +32. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021) + +33. Yariv, L., Hedman, P., Reiser, C., Verbin, D., Srinivasan, P.P., Szeliski, R., Barron, J.T., Mildenhall, B.: Bakedsdf: Meshing neural sdfs for real-time view synthesis. arXiv preprint arXiv:2302.14859 (2023) +34. Yariv, L., Kasten, Y., Moran, D., Galun, M., Atzmon, M., Ronen, B., Lipman, Y.: Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems 33, 2492-2502 (2020) +35. Zhang, J., Yao, Y., Li, S., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Critical regularizations for neural surface reconstruction in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6270-6279 (2022) +36. Zhang, J., Yao, Y., Li, S., Luo, Z., Fang, T.: Visibility-aware multi-view stereo network. British Machine Vision Conference (BMVC) (2020) +37. Zhang, J., Yao, Y., Quan, L.: Learning signed distance field for multi-view surface reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6525-6534 (2021) +38. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020) +39. Zhang, Y., Zhu, J., Lin, L.: Fastmesh: Fast surface reconstruction by hexagonal mesh-based neural rendering. arXiv preprint arXiv:2305.17858 (2023) +40. Zhang, Y., Hu, Z., Wu, H., Zhao, M., Li, L., Zou, Z., Fan, C.: Towards unbiased volume rendering of neural implicit surfaces with geometry priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4359-4368 (2023) +41. Zhuang, Y., Zhang, Q., Feng, Y., Zhu, H., Yao, Y., Li, X., Cao, Y.P., Shan, Y., Cao, X.: Anti-aliased neural implicit surfaces with encoding level of detail. In: SIGGRAPH Asia 2023 Conference Papers. pp. 1-10 (2023) \ No newline at end of file diff --git a/2024/CPT-VR_ Improving Surface Rendering via Closest Point Transform with View-Reflection Appearance/images.zip b/2024/CPT-VR_ Improving Surface Rendering via Closest Point Transform with View-Reflection Appearance/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..478da52fb48ef140b15639b94e62b8e8463d9810 --- /dev/null +++ b/2024/CPT-VR_ Improving Surface Rendering via Closest Point Transform with View-Reflection Appearance/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3dedcf62077e719a95cdde45337626840da7c552d7bdfd382c784897d1a76143 +size 579371 diff --git a/2024/CPT-VR_ Improving Surface Rendering via Closest Point Transform with View-Reflection Appearance/layout.json b/2024/CPT-VR_ Improving Surface Rendering via Closest Point Transform with View-Reflection Appearance/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..82f0da1635573ab92902c5352614b3ee53db7702 --- /dev/null +++ b/2024/CPT-VR_ Improving Surface Rendering via Closest Point Transform with View-Reflection Appearance/layout.json @@ -0,0 +1,10931 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 111, + 468, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 111, + 468, + 166 + ], + "spans": [ + { + "bbox": [ + 146, + 111, + 468, + 166 + ], + "type": "text", + "content": "CPT-VR: Improving Surface Rendering via Closest Point Transform with View-Reflection Appearance" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 185, + 475, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 185, + 475, + 210 + ], + "spans": [ + { + "bbox": [ + 138, + 185, + 475, + 210 + ], + "type": "text", + "content": "Zhipeng Hu" + }, + { + "bbox": [ + 138, + 185, + 475, + 210 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 138, + 185, + 475, + 210 + ], + "type": "text", + "content": ", Yongqiang Zhang" + }, + { + "bbox": [ + 138, + 185, + 475, + 210 + ], + "type": "inline_equation", + "content": "^{1,3,*}" + }, + { + "bbox": [ + 138, + 185, + 475, + 210 + ], + "type": "text", + "content": ", Chen Liu" + }, + { + "bbox": [ + 138, + 185, + 475, + 210 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 185, + 475, + 210 + ], + "type": "text", + "content": ", Lincheng Li" + }, + { + "bbox": [ + 138, + 185, + 475, + 210 + ], + "type": "inline_equation", + "content": "^{3,*}" + }, + { + "bbox": [ + 138, + 185, + 475, + 210 + ], + "type": "text", + "content": ", Sida Peng" + }, + { + "bbox": [ + 138, + 185, + 475, + 210 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 138, + 185, + 475, + 210 + ], + "type": "text", + "content": ", Xiaowei Zhou" + }, + { + "bbox": [ + 138, + 185, + 475, + 210 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 138, + 185, + 475, + 210 + ], + "type": "text", + "content": ", Changjie Fan" + }, + { + "bbox": [ + 138, + 185, + 475, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 185, + 475, + 210 + ], + "type": "text", + "content": ", and Xin Yu" + }, + { + "bbox": [ + 138, + 185, + 475, + 210 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 157, + 219, + 455, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 219, + 455, + 243 + ], + "spans": [ + { + "bbox": [ + 157, + 219, + 455, + 243 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 219, + 455, + 243 + ], + "type": "text", + "content": "NetEase Fuxi AI Lab, Hangzhou, China {zhangyongqiang02,zphu,lilincheng,fanchangjie}@corp.netease.com" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 194, + 242, + 419, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 242, + 419, + 252 + ], + "spans": [ + { + "bbox": [ + 194, + 242, + 419, + 252 + ], + "type": "text", + "content": "2 The University of Queensland, Queensland, Australia" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 203, + 254, + 411, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 254, + 411, + 264 + ], + "spans": [ + { + "bbox": [ + 203, + 254, + 411, + 264 + ], + "type": "text", + "content": "chen.liu7@uqconnect.edu.au, xin.yu@uq.edu.au" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 225, + 264, + 389, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 264, + 389, + 274 + ], + "spans": [ + { + "bbox": [ + 225, + 264, + 389, + 274 + ], + "type": "text", + "content": "3 Zhejiang University, Hangzhou, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 241, + 276, + 373, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 276, + 373, + 286 + ], + "spans": [ + { + "bbox": [ + 241, + 276, + 373, + 286 + ], + "type": "text", + "content": "{pengsida,xwzhou}@zju.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 310, + 452, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 310, + 452, + 540 + ], + "spans": [ + { + "bbox": [ + 159, + 310, + 452, + 540 + ], + "type": "text", + "content": "Abstract. Differentiable surface rendering has significantly advanced 3D reconstruction. Existing surface rendering methods assume that the local surface is planar, and thus employ linear approximation based on the Singed Distance Field (SDF) values to predict the point on the surface. However, this assumption overlooks the inherently irregular and non-planar nature of object surfaces in the real world. Consequently, the approximate points tend to deviate from the zero-level set, affecting the fidelity of the reconstructed shape. In this paper, we propose a novel surface rendering method termed CPT-VR, which leverages the Closet Point Transform (CPT) and View and Reflection direction vectors to enhance the quality of reconstruction. Specifically, leveraging the physical property of CPT that accurately projects points near the surface onto the zero-level set, we correct the deviated points, thus achieving an accurate geometry representation. Based on our accurate geometry representation, incorporating the reflection vector into our method can facilitate the appearance modeling of specular regions. Moreover, to enable our method to no longer be dependent on any prior knowledge of the background, we present a background model to learn the background appearance. Compared to previous state-of-the-art methods, CPT-VR achieves better surface reconstruction quality, even for cases with complex structures and specular highlights." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 551, + 452, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 551, + 452, + 572 + ], + "spans": [ + { + "bbox": [ + 159, + 551, + 452, + 572 + ], + "type": "text", + "content": "Keywords: Surface Rendering " + }, + { + "bbox": [ + 159, + 551, + 452, + 572 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 551, + 452, + 572 + ], + "type": "text", + "content": " 3D Reconstruction " + }, + { + "bbox": [ + 159, + 551, + 452, + 572 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 551, + 452, + 572 + ], + "type": "text", + "content": " Differentiable Rendering " + }, + { + "bbox": [ + 159, + 551, + 452, + 572 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 551, + 452, + 572 + ], + "type": "text", + "content": " Closest Point Transform" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 590, + 229, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 229, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 229, + 601 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 612, + 481, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 612, + 481, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 612, + 481, + 648 + ], + "type": "text", + "content": "Differentiable rendering plays a crucial role in 3D content creation, especially showing greater promise in multi-view 3D reconstruction and text-to-3D generation. Recent advancements have utilized differentiable rendering for attaining" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 247, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 247, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 247, + 665 + ], + "type": "text", + "content": "* Corresponding authors." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 129, + 240, + 193 + ], + "blocks": [ + { + "bbox": [ + 173, + 118, + 198, + 128 + ], + "lines": [ + { + "bbox": [ + 173, + 118, + 198, + 128 + ], + "spans": [ + { + "bbox": [ + 173, + 118, + 198, + 128 + ], + "type": "text", + "content": "(a) IDR" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 129, + 240, + 193 + ], + "lines": [ + { + "bbox": [ + 138, + 129, + 240, + 193 + ], + "spans": [ + { + "bbox": [ + 138, + 129, + 240, + 193 + ], + "type": "image", + "image_path": "e7a287aca156de308693b5480b9aa7127e69a8ee7c9698378b1fddc5d3ed5121.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 143, + 196, + 223, + 237 + ], + "blocks": [ + { + "bbox": [ + 143, + 196, + 223, + 237 + ], + "lines": [ + { + "bbox": [ + 143, + 196, + 223, + 237 + ], + "spans": [ + { + "bbox": [ + 143, + 196, + 223, + 237 + ], + "type": "image", + "image_path": "723c411c6796cdef18ca737e0debe2f4a918269a926ad53d9149b74c335df7e5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 239, + 209, + 246 + ], + "lines": [ + { + "bbox": [ + 168, + 239, + 209, + 246 + ], + "spans": [ + { + "bbox": [ + 168, + 239, + 209, + 246 + ], + "type": "inline_equation", + "content": "S_0" + }, + { + "bbox": [ + 168, + 239, + 209, + 246 + ], + "type": "text", + "content": ": zero-level set" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 171, + 247, + 207, + 252 + ], + "lines": [ + { + "bbox": [ + 171, + 247, + 207, + 252 + ], + "spans": [ + { + "bbox": [ + 171, + 247, + 207, + 252 + ], + "type": "text", + "content": "v: the view ray" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 143, + 266, + 235, + 273 + ], + "lines": [ + { + "bbox": [ + 143, + 266, + 235, + 273 + ], + "spans": [ + { + "bbox": [ + 143, + 266, + 235, + 273 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\prime}" + }, + { + "bbox": [ + 143, + 266, + 235, + 273 + ], + "type": "text", + "content": " : the linear projection along the ray " + }, + { + "bbox": [ + 143, + 266, + 235, + 273 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 249, + 129, + 351, + 193 + ], + "blocks": [ + { + "bbox": [ + 284, + 118, + 320, + 128 + ], + "lines": [ + { + "bbox": [ + 284, + 118, + 320, + 128 + ], + "spans": [ + { + "bbox": [ + 284, + 118, + 320, + 128 + ], + "type": "text", + "content": "(b)DMTet" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 249, + 129, + 351, + 193 + ], + "lines": [ + { + "bbox": [ + 249, + 129, + 351, + 193 + ], + "spans": [ + { + "bbox": [ + 249, + 129, + 351, + 193 + ], + "type": "image", + "image_path": "aa034c72e6a9d6884d0a6d31ff401f50e92f4fa256aa890020c409ef469a57c5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 252, + 193, + 347, + 232 + ], + "blocks": [ + { + "bbox": [ + 134, + 253, + 244, + 266 + ], + "lines": [ + { + "bbox": [ + 134, + 253, + 244, + 266 + ], + "spans": [ + { + "bbox": [ + 134, + 253, + 244, + 266 + ], + "type": "text", + "content": "x: the sampled point by sphere tracing " + }, + { + "bbox": [ + 134, + 253, + 244, + 266 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 134, + 253, + 244, + 266 + ], + "type": "text", + "content": ": the SDF value of x, n: the normal vector of x" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 252, + 193, + 347, + 232 + ], + "lines": [ + { + "bbox": [ + 252, + 193, + 347, + 232 + ], + "spans": [ + { + "bbox": [ + 252, + 193, + 347, + 232 + ], + "type": "image", + "image_path": "927bf82a786c33bc9f3df6e502fe3e3b105a47b992f2d59cca3832fcce6d54d7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 234, + 316, + 241 + ], + "lines": [ + { + "bbox": [ + 276, + 234, + 316, + 241 + ], + "spans": [ + { + "bbox": [ + 276, + 234, + 316, + 241 + ], + "type": "inline_equation", + "content": "S_0" + }, + { + "bbox": [ + 276, + 234, + 316, + 241 + ], + "type": "text", + "content": ": zero-level set" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 249, + 241, + 343, + 247 + ], + "lines": [ + { + "bbox": [ + 249, + 241, + 343, + 247 + ], + "spans": [ + { + "bbox": [ + 249, + 241, + 343, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i, \\mathbf{x}_j" + }, + { + "bbox": [ + 249, + 241, + 343, + 247 + ], + "type": "text", + "content": ": vertices of an edge where the SDF" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 281, + 247, + 311, + 255 + ], + "lines": [ + { + "bbox": [ + 281, + 247, + 311, + 255 + ], + "spans": [ + { + "bbox": [ + 281, + 247, + 311, + 255 + ], + "type": "text", + "content": "changes sign" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 264, + 255, + 326, + 262 + ], + "lines": [ + { + "bbox": [ + 264, + 255, + 326, + 262 + ], + "spans": [ + { + "bbox": [ + 264, + 255, + 326, + 262 + ], + "type": "inline_equation", + "content": "d_{i},d_{j}" + }, + { + "bbox": [ + 264, + 255, + 326, + 262 + ], + "type": "text", + "content": " :SDF values of " + }, + { + "bbox": [ + 264, + 255, + 326, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i,\\mathbf{x}_j" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 261, + 262, + 331, + 268 + ], + "lines": [ + { + "bbox": [ + 261, + 262, + 331, + 268 + ], + "spans": [ + { + "bbox": [ + 261, + 262, + 331, + 268 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^{\\prime},\\mathbf{x}_j^{\\prime}" + }, + { + "bbox": [ + 261, + 262, + 331, + 268 + ], + "type": "text", + "content": " : local deformed vertices" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 364, + 129, + 467, + 193 + ], + "blocks": [ + { + "bbox": [ + 399, + 118, + 445, + 128 + ], + "lines": [ + { + "bbox": [ + 399, + 118, + 445, + 128 + ], + "spans": [ + { + "bbox": [ + 399, + 118, + 445, + 128 + ], + "type": "text", + "content": "(c) CPT-Ours" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 364, + 129, + 467, + 193 + ], + "lines": [ + { + "bbox": [ + 364, + 129, + 467, + 193 + ], + "spans": [ + { + "bbox": [ + 364, + 129, + 467, + 193 + ], + "type": "image", + "image_path": "80b2f5e2512588da28f2dc9ddee4384a6911a09974e5e2cf6944d77a7262f118.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 364, + 194, + 454, + 241 + ], + "blocks": [ + { + "bbox": [ + 249, + 269, + 343, + 277 + ], + "lines": [ + { + "bbox": [ + 249, + 269, + 343, + 277 + ], + "spans": [ + { + "bbox": [ + 249, + 269, + 343, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{ij}^{\\prime}" + }, + { + "bbox": [ + 249, + 269, + 343, + 277 + ], + "type": "text", + "content": " : linear interpolation of " + }, + { + "bbox": [ + 249, + 269, + 343, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^{\\prime},\\mathbf{x}_j^{\\prime}" + }, + { + "bbox": [ + 249, + 269, + 343, + 277 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 249, + 269, + 343, + 277 + ], + "type": "inline_equation", + "content": "d_{i},d_{j}" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 364, + 194, + 454, + 241 + ], + "lines": [ + { + "bbox": [ + 364, + 194, + 454, + 241 + ], + "spans": [ + { + "bbox": [ + 364, + 194, + 454, + 241 + ], + "type": "image", + "image_path": "c5abb21e4df1684c16d3070c03b02aa390d3103a04d06826b26328c452e82f9c.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 241, + 437, + 247 + ], + "lines": [ + { + "bbox": [ + 396, + 241, + 437, + 247 + ], + "spans": [ + { + "bbox": [ + 396, + 241, + 437, + 247 + ], + "type": "inline_equation", + "content": "S_0" + }, + { + "bbox": [ + 396, + 241, + 437, + 247 + ], + "type": "text", + "content": " : zero-level set" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 354, + 247, + 479, + 255 + ], + "lines": [ + { + "bbox": [ + 354, + 247, + 479, + 255 + ], + "spans": [ + { + "bbox": [ + 354, + 247, + 479, + 255 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i, \\mathbf{x}_j" + }, + { + "bbox": [ + 354, + 247, + 479, + 255 + ], + "type": "text", + "content": ": vertices of an edge where the SDF changes sign" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 357, + 255, + 476, + 261 + ], + "lines": [ + { + "bbox": [ + 357, + 255, + 476, + 261 + ], + "spans": [ + { + "bbox": [ + 357, + 255, + 476, + 261 + ], + "type": "inline_equation", + "content": "d_{i},d_{j},d_{ij}" + }, + { + "bbox": [ + 357, + 255, + 476, + 261 + ], + "type": "text", + "content": " : SDF values, " + }, + { + "bbox": [ + 357, + 255, + 476, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_{ij}" + }, + { + "bbox": [ + 357, + 255, + 476, + 261 + ], + "type": "text", + "content": " : the normal vector of " + }, + { + "bbox": [ + 357, + 255, + 476, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{ij}" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 371, + 261, + 463, + 268 + ], + "lines": [ + { + "bbox": [ + 371, + 261, + 463, + 268 + ], + "spans": [ + { + "bbox": [ + 371, + 261, + 463, + 268 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{ij}" + }, + { + "bbox": [ + 371, + 261, + 463, + 268 + ], + "type": "text", + "content": ": linear interpolation of " + }, + { + "bbox": [ + 371, + 261, + 463, + 268 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i, \\mathbf{x}_j" + }, + { + "bbox": [ + 371, + 261, + 463, + 268 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 371, + 261, + 463, + 268 + ], + "type": "inline_equation", + "content": "d_i, d_j" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 371, + 268, + 462, + 276 + ], + "lines": [ + { + "bbox": [ + 371, + 268, + 462, + 276 + ], + "spans": [ + { + "bbox": [ + 371, + 268, + 462, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{ij}^{\\prime}" + }, + { + "bbox": [ + 371, + 268, + 462, + 276 + ], + "type": "text", + "content": " : the projected vertex of " + }, + { + "bbox": [ + 371, + 268, + 462, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{ij}" + }, + { + "bbox": [ + 371, + 268, + 462, + 276 + ], + "type": "text", + "content": " by CPT" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 282, + 482, + 407 + ], + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 407 + ], + "type": "text", + "content": "Fig. 1: Illustration of the surface point sampling methods and their corresponding construction results. (a) IDR [34] first employs a sphere tracing method to identify the sampled point " + }, + { + "bbox": [ + 130, + 282, + 482, + 407 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 282, + 482, + 407 + ], + "type": "text", + "content": " and then leverages the SDF value as well as the normal vector to obtain the approximate point " + }, + { + "bbox": [ + 130, + 282, + 482, + 407 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 130, + 282, + 482, + 407 + ], + "type": "text", + "content": ". (b) DMTet [19] detects surfaces through the observation of sign changes in the SDF at the vertices of a tetrahedron. Then, it employs linear interpolation to approximate the points " + }, + { + "bbox": [ + 130, + 282, + 482, + 407 + ], + "type": "inline_equation", + "content": "x_{ij}'" + }, + { + "bbox": [ + 130, + 282, + 482, + 407 + ], + "type": "text", + "content": " on the zero-level set. Compared with them, our method (c) utilizes CPT to project the interpolated points " + }, + { + "bbox": [ + 130, + 282, + 482, + 407 + ], + "type": "inline_equation", + "content": "x_{ij}" + }, + { + "bbox": [ + 130, + 282, + 482, + 407 + ], + "type": "text", + "content": " on the zero-level set, and thus attain the transformed points " + }, + { + "bbox": [ + 130, + 282, + 482, + 407 + ], + "type": "inline_equation", + "content": "x_{ij}'" + }, + { + "bbox": [ + 130, + 282, + 482, + 407 + ], + "type": "text", + "content": ". The approximate points " + }, + { + "bbox": [ + 130, + 282, + 482, + 407 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 130, + 282, + 482, + 407 + ], + "type": "text", + "content": " of IDR and " + }, + { + "bbox": [ + 130, + 282, + 482, + 407 + ], + "type": "inline_equation", + "content": "x_{ij}'" + }, + { + "bbox": [ + 130, + 282, + 482, + 407 + ], + "type": "text", + "content": " of DMTet are deviated from the zero-level set, while " + }, + { + "bbox": [ + 130, + 282, + 482, + 407 + ], + "type": "inline_equation", + "content": "x_{ij}'" + }, + { + "bbox": [ + 130, + 282, + 482, + 407 + ], + "type": "text", + "content": " obtained from our method exactly on the object surface. This implies our method is able to obtain a more accurate geometry representation and gradient backpropagation." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 411, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 495 + ], + "type": "text", + "content": "the Singed Distance Field (SDF) and object appearance. These methods can be divided into two types: volume rendering and surface rendering. In volume rendering, SDF is first converted into density values that are then rendered via the NeRF framework [17, 38]. While these methods consistently generate reliable geometry reconstruction, the dense points sampling along the ray increases the rendering time cost. Conversely, surface rendering methods only sample one point along the ray, markedly optimizing rendering speeds." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Existing surface rendering methods, such as IDR (Implicit Differentiable Renderer) [34] and DMTet (Deep Marching Tetrahedra) [19] achieve computational efficiency by assuming the object surface can be considered linear or at least can be approximated as linear within a local scope. Based on this assumption, they employ linear interpolation to determine the points at specific locations on the surface. As shown in Fig. 1, IDR employs sphere tracing [7] to identify points close to the surface along a ray and then relies on the SDF values and the normal vector to calculate the intersection points between rays and the surface. DMTet identifies the surfaces by observing changes in the sign of the SDF at the vertices of a tetrahedron and then leverages linear interpolation to approximate the surface points. However, as demonstrated in Fig. 1, their approximate points always deviate from the object surface (i.e., zero-level set), since the object surface is mostly irregular and non-planar. Hence, modeling geometry representations based on these deviated points would further lead to inaccurate gradient back-" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Z. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "propagation, resulting in the reconstructed surfaces losing a lot of details, as illustrated by the reconstruction results in Fig. 1 (a) and (b)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 481, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 481, + 251 + ], + "type": "text", + "content": "To address the above problem, we introduce a novel surface rendering approach, termed CPT-VR, which is grounded in the Closet Point Transform algorithm and augmented with View-Reflection appearance modeling. In a nutshell, in the geometry representation modeling process, we conduct the Closet Point Transform (CPT) for deviated points to enhance the geometry representation. In the appearance modeling process, we intend to incorporate the View and Reflection vectors and thus attain better appearance representation for cases with specular highlights. To the best of our knowledge, we are the first attempt to correct the deviated points via CPT in the surface rendering task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 253, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 253, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 482, + 373 + ], + "type": "text", + "content": "To be specific, in the geometry representation modeling process, we first devise a neural SDF network to compute SDF values for the entire 3D space. Based on the SDF values, we employ the Marching Cubes [14] to attain the vertices close to the surface of the object. As suggested in Fig. 1 (c), most of the obtained vertices are deviated from the object's surface rather than on it. Hence, we leverage CPT to calculate the closest point on the object surface for each deviated vertex, thus obtaining accurate and differentiable vertices. Subsequently, we utilize a feature-based rasterizer to render this triangle mesh in each view. In this fashion, we obtain the corresponding 3D position, geometry feature, and normal vector of each pixel in an image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 376, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 532 + ], + "type": "text", + "content": "In the appearance representation modeling process, we focus on improving the geometry quality for specular highlights. Existing methods such as IDR [34] and NeuS2 [25] rely solely on the view direction to capture the surface information of objects, which cannot handle the object with highlights well. Ref-NeRF [21] demonstrates that replacing the view direction with the reflection direction can mitigate the impact of highlights. However, since the reflection vector calculated based on the deviated points is inaccurate, directly incorporating these reflection vectors into existing surface rendering methods will lead to significant backpropagation errors. Consequently, the model will not only be unable to handle the specular highlights but will also diminish its capability to reconstruct non-specular areas. In contrast, benefiting from our accurate point sampling on the zero-level set based on CPT, incorporating view-reflection vectors in our framework can facilitate the appearance modeling of specular regions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": "Moreover, we found that existing surface rendering methods [1,22,34] require prior background knowledge, such as foreground masks, or the background image, to precisely process and render foreground objects, which constrains their flexibility and applicability in various scenes. To solve this problem, we present a background model to obtain the background appearance representation. Then we attain the final appearance modeling by blending the foreground and background appearance. Extensive experiments conducted on DTU [9] and BlendedMVS [31] demonstrate that our method achieves state-of-the-art geometry quality and surface rendering. Additionally, we also illustrate that our method is effective and reliable in cases with complex structures and specular highlights. In summary, our contributions are three-fold:" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "text", + "content": "Improving Surface Rendering via CPT with View-Reflection Appearance" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 479, + 238 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 138, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 479, + 163 + ], + "type": "text", + "content": "- We present a novel differentiable surface rendering method CPT-VR which shows superior performance in reconstructing the detailed structure of objects. With the CPT algorithm, we can address the approximate point deviation issue caused by linear interpolation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 166, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 166, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 138, + 166, + 479, + 200 + ], + "type": "text", + "content": "- We incorporate the view and reflection vectors to model the foreground appearance representation, facilitating our method to be robust against specular highlights." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 204, + 479, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 204, + 479, + 238 + ], + "spans": [ + { + "bbox": [ + 138, + 204, + 479, + 238 + ], + "type": "text", + "content": "- We develop a simple background model to generate the background appearance representation, enabling our method to no longer be dependent on any prior background knowledge." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 245, + 237, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 245, + 237, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 245, + 237, + 258 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 270, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 437 + ], + "type": "text", + "content": "Surface Rendering with SDF. IDR [34] introduces a system that learns surfaces in a self-supervised manner. It first identifies points close to the surface via sphere tracing and then finds approximate intersection points on the surface by projecting them along the view ray. Although IDR can learn the 3D shape and appearance from images, it relies heavily on accurate masks. To mitigate this issue, MVsDF [37] supervises the SDF network with depth maps generated by Vis-MVSNET [36]. It also uses relaxed masks generated by probability maps to learn a surface indicator. RegSDF [35] further introduces the Hessian regularization and the minimal surface constraint. Some works [1, 22] use specialized reparameterization techniques and design weighting functions along a ray based on SDF values, thus they can propagate gradients to intermediate points from sphere tracing. These reparameterization-based methods are similar to methods based on volume rendering, but they still need a given background: either black or a pre-defined environment map." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 438, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 568 + ], + "type": "text", + "content": "Rasterization-based Rendering. Recently, there have been attempts to utilize the differentiable rasterizer for surface rendering. NVDIFFREC [18] and NVDIFFRECmc [8] adapt DMTet [19] to extract a mesh from an implicit SDF and then render the mesh by the differentiable rasterizer Nvdiffrast [10]. While these methods are efficient, their reliance on the linear approximation in DMTet limits their ability to recover shapes with complex structures. Meanwhile, NIE [16] propagates neural implicit surfaces according to a flow field from a certain energy function. ENS [23] combines neural deformation fields with rasterization rendering, allowing deform an initial shape into a target surface progressively. Moreover, NDS [29] and FastMESH [39] focus on optimizing a rough mesh derived from priors through differentiable rasterization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Volume Rendering of SDF. Several methods [2, 20, 24, 32] have been developed for learning SDF through volume rendering. VolSDF [32] and NeuS [24] leverage Laplace or Logistic cumulative distribution function of SDF to represent the volume density and sample dense points along a ray based on SDF values. Based on these methods, some works [2, 3, 12, 26–28, 40, 41] aim to further enhance the geometry details. Among them, Geo-Neus [3] and TUVR [40] improve SDF accuracy by introducing additional supervision from point clouds. On the training efficiency front, Voxurf [30] and Vox-Surf [11] speed up the training process" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Z. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "through voxel-based representations. NeuS2 [25] speeds up NeuS [24] by simplifying the second-order derivatives calculation of the SDF network constructed based on a multi-resolution hash encoding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 248 + ], + "type": "text", + "content": "Hybrid Volume-rasterization Rendering. Recently, some rendering methods have been developed based on a hybrid volume-rasterization way. BakedSDF [33] first learns the SDF and appearance of a scene based on VolSDF [32] and transforms the SDF into a mesh. Then the transformed mesh is refined using a rasterizer pipeline. VMesh [6] takes a comprehensive approach by initially training a combination of a neural SDF and a density field through volume rendering. Then the mesh undergoes further refinement by DMTet [19] and a differentiable rasterizer [10], ensuring accurate surface representation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 266, + 202, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 266, + 202, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 266, + 202, + 277 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 290, + 228, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 290, + 228, + 301 + ], + "spans": [ + { + "bbox": [ + 132, + 290, + 228, + 301 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 309, + 480, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 480, + 369 + ], + "type": "text", + "content": "Signed Distance Field (SDF) is a mathematical function " + }, + { + "bbox": [ + 130, + 309, + 480, + 369 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 309, + 480, + 369 + ], + "type": "text", + "content": " that measures the shortest distance from a point " + }, + { + "bbox": [ + 130, + 309, + 480, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 309, + 480, + 369 + ], + "type": "text", + "content": " to an object surface, indicating whether the point is outside (positive distance) or inside (negative distance) the surface. The surface is identified by the zero-level set " + }, + { + "bbox": [ + 130, + 309, + 480, + 369 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 309, + 480, + 369 + ], + "type": "text", + "content": ", where SDF equals to zero, providing a concise description of the object's surface:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 255, + 379, + 480, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 379, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 255, + 379, + 480, + 392 + ], + "type": "interline_equation", + "content": "\\mathcal {S} = \\left\\{\\mathbf {x} \\in \\mathbb {R} ^ {3} | f (\\mathbf {x}) = 0 \\right\\}. \\tag {1}", + "image_path": "dda2d939317f0f85bcdf04049cbc7f467946308299a727a48e8466d294c10d53.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 399, + 479, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 479, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 479, + 437 + ], + "type": "text", + "content": "Neural SDF denotes a neural network that models SDFs for 3D objects. In this work, we employ SDFnet which leverages the Multilayer Perceptrons (MLPs) enhanced with a multi-resolution hash encoding to model SDF [25]:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 261, + 445, + 480, + 457 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 445, + 480, + 457 + ], + "spans": [ + { + "bbox": [ + 261, + 445, + 480, + 457 + ], + "type": "interline_equation", + "content": "(d, \\mathbf {z}) = f _ {\\theta} (\\mathbf {x}, h _ {\\Omega} (\\mathbf {x})). \\tag {2}", + "image_path": "9e1815ab77b4da205482f095ed4a2c8a806596f0680307c4d2a5df4a83841f9d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 466, + 480, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 466, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 466, + 480, + 514 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 466, + 480, + 514 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 466, + 480, + 514 + ], + "type": "text", + "content": " is the predicted SDF value, " + }, + { + "bbox": [ + 130, + 466, + 480, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 466, + 480, + 514 + ], + "type": "text", + "content": " signifies the predicted geometry feature, " + }, + { + "bbox": [ + 130, + 466, + 480, + 514 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 130, + 466, + 480, + 514 + ], + "type": "text", + "content": " denotes an MLP parameterized by weights " + }, + { + "bbox": [ + 130, + 466, + 480, + 514 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 466, + 480, + 514 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 466, + 480, + 514 + ], + "type": "inline_equation", + "content": "h_{\\Omega}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 466, + 480, + 514 + ], + "type": "text", + "content": " is the multi-resolution hash encoding of the input point " + }, + { + "bbox": [ + 130, + 466, + 480, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 466, + 480, + 514 + ], + "type": "text", + "content": ". Moreover, the normal vector " + }, + { + "bbox": [ + 130, + 466, + 480, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 466, + 480, + 514 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 466, + 480, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 466, + 480, + 514 + ], + "type": "text", + "content": " is derived as the normalized gradient of the implicit surface:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 269, + 525, + 480, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 525, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 269, + 525, + 480, + 536 + ], + "type": "interline_equation", + "content": "\\mathbf {n} = \\nabla_ {\\mathbf {x}} d / \\| \\nabla_ {\\mathbf {x}} d \\|. \\tag {3}", + "image_path": "a5fa404e27796c34d07676c0c31defabe9189b52b8822cee2f70e5f7a033f745.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 544, + 479, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 479, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 479, + 569 + ], + "type": "text", + "content": "To ensure training stability, our SDFnet employs an optimization schedule that progressively incorporates the level of detail and the numerical gradients [12]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 586, + 342, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 342, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 342, + 598 + ], + "type": "text", + "content": "3.2 Geometry Representation Modeling" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "content": "Geometry Representation via SDF. As suggested in Fig. 2, we first leverage our SDFnet to compute the SDF values for the entire 3D space. Based on SDF values, we utilize the isosurface algorithm Marching Cubes to extract and represent the triangle mesh " + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = (\\mathcal{V},\\mathcal{F})" + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 605, + 480, + 666 + ], + "type": "text", + "content": " indicate the vertices and faces respectively. However, this method assumes that the surface changes" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 448, + 103 + ], + "type": "text", + "content": "Improving Surface Rendering via CPT with View-Reflection Appearance" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 121, + 474, + 186 + ], + "blocks": [ + { + "bbox": [ + 138, + 121, + 474, + 186 + ], + "lines": [ + { + "bbox": [ + 138, + 121, + 474, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 121, + 474, + 186 + ], + "type": "image", + "image_path": "26875512fec8bc781bf49e5624a3f02723133c86a23d48a02078e4b3b5828188.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 138, + 187, + 474, + 295 + ], + "blocks": [ + { + "bbox": [ + 138, + 187, + 474, + 295 + ], + "lines": [ + { + "bbox": [ + 138, + 187, + 474, + 295 + ], + "spans": [ + { + "bbox": [ + 138, + 187, + 474, + 295 + ], + "type": "image", + "image_path": "5dc2c214efe8bfac2d39fd250aa4ab30e32109455e5872013c11439d5411f6e5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 301, + 482, + 390 + ], + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 390 + ], + "type": "text", + "content": "Fig. 2: Illustration of our proposed method CPT-VR. In (a), we first extract the mesh via Marching Cubes, which does not involve gradient propagation. Then we leverage CPT to project the approximate points on the zero-level set (b). Subsequently, we render the transformed mesh via a rasterizer in (c). Specifically, we incorporate the view-reflection vectors into the foreground appearance modeling stage via RGBNet. Additionally, we employ the Background Model to attain the background information. Finally, we integrate the foreground and background appearance to generate the rendering results. We adopt four supervisions to optimize these neural networks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 396, + 482, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 443 + ], + "type": "text", + "content": "linearly inside the voxel, overlooking the complex convex structure of the object surface. Hence, the linearly interpolated vertices tend to deviate from the zero-level set and the triangle mesh may not accurately reflect the complex structure of the object surface." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 445, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 517 + ], + "type": "text", + "content": "Closest Point Transform. To project these deviated vertices onto the zero-level set, we devise the closet set transform (CPT) to SDF. Specifically, CPT first utilizes the norm vector " + }, + { + "bbox": [ + 130, + 445, + 482, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 445, + 482, + 517 + ], + "type": "text", + "content": " within the signed distance field to identify the projected direction of a deviated point. It then moves the deviated point along the opposite direction of " + }, + { + "bbox": [ + 130, + 445, + 482, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 445, + 482, + 517 + ], + "type": "text", + "content": " by a distance " + }, + { + "bbox": [ + 130, + 445, + 482, + 517 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 445, + 482, + 517 + ], + "type": "text", + "content": " equal to its SDF value. Mathematically, for a vertex " + }, + { + "bbox": [ + 130, + 445, + 482, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathcal{V}" + }, + { + "bbox": [ + 130, + 445, + 482, + 517 + ], + "type": "text", + "content": ", this transformation process is expressed as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 277, + 530, + 481, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 530, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 277, + 530, + 481, + 544 + ], + "type": "interline_equation", + "content": "\\mathbf {x} ^ {\\prime} = \\mathbf {x} - d \\mathbf {n}. \\tag {4}", + "image_path": "e7db0baf8d908d2e61e569fd06fc5995b3d926359275c883148f39a78fd36ad4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 549, + 385, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 549, + 385, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 549, + 385, + 563 + ], + "type": "text", + "content": "After that, we obtain the transformed mesh " + }, + { + "bbox": [ + 132, + 549, + 385, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{G}' = (\\mathcal{V}',\\mathcal{F})" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 582, + 302, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 582, + 302, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 582, + 302, + 594 + ], + "type": "text", + "content": "3.3 Feature-based Rasterisation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "After obtaining the transformed mesh, we first employ a feature-based differentiable rasterizer Nvdiffrast [10] to render triangle meshes in each view. Specifically, as suggested in Fig. 2 (c), for a pixel " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "p(x,y)" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ", its projected primitive triangle face within the mesh is represented as " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_0',\\mathbf{x}_1',\\mathbf{x}_2')" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ". The rasterizer computes the barycentric coordinates " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "u,v" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ", depth " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "D_{p}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ", and visible mask " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "M_{p}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " for this pixel as" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Z. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 167, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 167, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 167, + 126 + ], + "type": "text", + "content": "follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 224, + 129, + 480, + 141 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 129, + 480, + 141 + ], + "spans": [ + { + "bbox": [ + 224, + 129, + 480, + 141 + ], + "type": "interline_equation", + "content": "(u, v, D _ {p}, M _ {p}) = \\mathcal {R} \\left(\\mathcal {G} ^ {\\prime}, p (x, y), K, R, \\mathbf {t}\\right), \\tag {5}", + "image_path": "47f95f9dd2abab262336d1065f3c29af9bd393115a7d56c52f0db637ba3b752a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 145, + 481, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 145, + 481, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 145, + 481, + 169 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 145, + 481, + 169 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 132, + 145, + 481, + 169 + ], + "type": "text", + "content": " denotes the rasterizer, " + }, + { + "bbox": [ + 132, + 145, + 481, + 169 + ], + "type": "inline_equation", + "content": "\\mathcal{G}'" + }, + { + "bbox": [ + 132, + 145, + 481, + 169 + ], + "type": "text", + "content": " represents the transformed mesh, and " + }, + { + "bbox": [ + 132, + 145, + 481, + 169 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 132, + 145, + 481, + 169 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 145, + 481, + 169 + ], + "type": "inline_equation", + "content": "R, \\mathbf{t}" + }, + { + "bbox": [ + 132, + 145, + 481, + 169 + ], + "type": "text", + "content": " are the camera's intrinsic and extrinsic parameters, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 170, + 479, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 170, + 479, + 217 + ], + "spans": [ + { + "bbox": [ + 131, + 170, + 479, + 217 + ], + "type": "text", + "content": "Subsequently, each projected vertex " + }, + { + "bbox": [ + 131, + 170, + 479, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 131, + 170, + 479, + 217 + ], + "type": "text", + "content": " from the set " + }, + { + "bbox": [ + 131, + 170, + 479, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{V}'" + }, + { + "bbox": [ + 131, + 170, + 479, + 217 + ], + "type": "text", + "content": " is input into SDFnet. SDFnet outputs " + }, + { + "bbox": [ + 131, + 170, + 479, + 217 + ], + "type": "inline_equation", + "content": "(d',\\mathbf{z}',\\mathbf{n}')" + }, + { + "bbox": [ + 131, + 170, + 479, + 217 + ], + "type": "text", + "content": ", encapsulating the distance, geometric feature, and the normal vector associated with each vertex. These outputs are then interpolated for a pixel " + }, + { + "bbox": [ + 131, + 170, + 479, + 217 + ], + "type": "inline_equation", + "content": "p(x,y)" + }, + { + "bbox": [ + 131, + 170, + 479, + 217 + ], + "type": "text", + "content": " based on its barycentric coordinates " + }, + { + "bbox": [ + 131, + 170, + 479, + 217 + ], + "type": "inline_equation", + "content": "u,v" + }, + { + "bbox": [ + 131, + 170, + 479, + 217 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 236, + 228, + 378, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 228, + 378, + 239 + ], + "spans": [ + { + "bbox": [ + 236, + 228, + 378, + 239 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {p} = u \\mathbf {x} _ {0} ^ {\\prime} + v \\mathbf {x} _ {1} ^ {\\prime} + (1 - u - v) \\mathbf {x} _ {2} ^ {\\prime},", + "image_path": "5af0148065563d2846b24fb63837816efdabe72960aef8e48c17a5e32c97c6ad.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 236, + 240, + 480, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 240, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 236, + 240, + 480, + 251 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {p} = u \\mathbf {z} _ {0} ^ {\\prime} + v \\mathbf {z} _ {1} ^ {\\prime} + (1 - u - v) \\mathbf {z} _ {2} ^ {\\prime}, \\tag {6}", + "image_path": "12561a069b121c41763a4d4ccdc653dd09880f7af189c707aa3010c62a0ee537.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 236, + 253, + 379, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 253, + 379, + 264 + ], + "spans": [ + { + "bbox": [ + 236, + 253, + 379, + 264 + ], + "type": "interline_equation", + "content": "\\mathbf {n} _ {p} = u \\mathbf {n} _ {0} ^ {\\prime} + v \\mathbf {n} _ {1} ^ {\\prime} + (1 - u - v) \\mathbf {n} _ {2} ^ {\\prime},", + "image_path": "bcce957aab122fe21a27eb86ff1d2fdc624abaf7b90d0911370260b3639ef5f7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 272, + 358, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 272, + 358, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 272, + 358, + 285 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 272, + 358, + 285 + ], + "type": "inline_equation", + "content": "(\\mathbf{x_p},\\mathbf{z_p},\\mathbf{n_p})" + }, + { + "bbox": [ + 132, + 272, + 358, + 285 + ], + "type": "text", + "content": " denote the interpolated features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 301, + 351, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 301, + 351, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 301, + 351, + 313 + ], + "type": "text", + "content": "3.4 Appearance Representation Modeling" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 321, + 449, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 449, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 449, + 332 + ], + "type": "text", + "content": "In this work, we divide the appearance modeling process into two parts:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 332, + 479, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 479, + 417 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 479, + 417 + ], + "type": "text", + "content": "Foreground Appearance Modeling based on View-Reflection Vectors. To better capture the appearance of objects with specular highlights, we incorporate the view direction and reflection direction into the foreground modeling process. Specifically, we employ the RGBnet " + }, + { + "bbox": [ + 132, + 332, + 479, + 417 + ], + "type": "inline_equation", + "content": "g_{\\gamma}^{f}" + }, + { + "bbox": [ + 132, + 332, + 479, + 417 + ], + "type": "text", + "content": ", a Multilayer Perceptron (MLP) with weights " + }, + { + "bbox": [ + 132, + 332, + 479, + 417 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 132, + 332, + 479, + 417 + ], + "type": "text", + "content": " to determine the foreground color. RGBnet takes the interpolated 3D features " + }, + { + "bbox": [ + 132, + 332, + 479, + 417 + ], + "type": "inline_equation", + "content": "(\\mathbf{x_p}, \\mathbf{z_p}, \\mathbf{n_p})" + }, + { + "bbox": [ + 132, + 332, + 479, + 417 + ], + "type": "text", + "content": ", view direction " + }, + { + "bbox": [ + 132, + 332, + 479, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_p" + }, + { + "bbox": [ + 132, + 332, + 479, + 417 + ], + "type": "text", + "content": ", and reflection direction " + }, + { + "bbox": [ + 132, + 332, + 479, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_p" + }, + { + "bbox": [ + 132, + 332, + 479, + 417 + ], + "type": "text", + "content": " as input, and outputs the RGB color " + }, + { + "bbox": [ + 132, + 332, + 479, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_p^f" + }, + { + "bbox": [ + 132, + 332, + 479, + 417 + ], + "type": "text", + "content": " for the foreground. The process is formulated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 232, + 427, + 480, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 427, + 480, + 442 + ], + "spans": [ + { + "bbox": [ + 232, + 427, + 480, + 442 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {p} ^ {f} = g _ {\\gamma} ^ {f} \\left(\\mathbf {x} _ {\\mathbf {p}}, \\mathbf {z} _ {\\mathbf {p}}, \\mathbf {n} _ {\\mathbf {p}}, Y (\\mathbf {v} _ {\\mathbf {p}}), Y (\\mathbf {r} _ {\\mathbf {p}})\\right), \\tag {7}", + "image_path": "03c78bfc631f5d464da18a20479d8ae7385e1fa2be4ad77c736c41836058be3c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 445, + 479, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 445, + 479, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 445, + 479, + 482 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 445, + 479, + 482 + ], + "type": "inline_equation", + "content": "Y(\\cdot)" + }, + { + "bbox": [ + 132, + 445, + 479, + 482 + ], + "type": "text", + "content": " denotes Spherical Harmonics encoding. Here, the reflection direction " + }, + { + "bbox": [ + 132, + 445, + 479, + 482 + ], + "type": "inline_equation", + "content": "\\mathbf{r_p}" + }, + { + "bbox": [ + 132, + 445, + 479, + 482 + ], + "type": "text", + "content": " is defined as the mirror vector of the view direction " + }, + { + "bbox": [ + 132, + 445, + 479, + 482 + ], + "type": "inline_equation", + "content": "\\mathbf{v_p}" + }, + { + "bbox": [ + 132, + 445, + 479, + 482 + ], + "type": "text", + "content": " relative to the normal vector " + }, + { + "bbox": [ + 132, + 445, + 479, + 482 + ], + "type": "inline_equation", + "content": "\\mathbf{n_p}" + }, + { + "bbox": [ + 132, + 445, + 479, + 482 + ], + "type": "text", + "content": ", calculated as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 250, + 492, + 480, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 492, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 250, + 492, + 480, + 504 + ], + "type": "interline_equation", + "content": "\\mathbf {r} _ {\\mathbf {p}} = 2 \\left(- \\mathbf {v} _ {\\mathbf {p}} \\cdot \\mathbf {n} _ {\\mathbf {p}}\\right) \\mathbf {n} _ {\\mathbf {p}} + \\mathbf {v} _ {\\mathbf {p}}. \\tag {8}", + "image_path": "894fd3f03bdba0e4d662d50bc3823210cd6c5f64fd7647291a4a5ac3b99f6fb2.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 512, + 479, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 479, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 479, + 548 + ], + "type": "text", + "content": "Spherical harmonics are orthogonal bases on the unit sphere, and their mathematical definition and physical properties make them ideal for encoding direction frequencies compared to 3D space (xyz) frequencies." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 548, + 479, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 479, + 609 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 479, + 609 + ], + "type": "text", + "content": "Background Appearance Modeling. To enable the model to reconstruct objects without any background priors, we present a one-point background model composed of a Multilayer Perceptron (MLP) " + }, + { + "bbox": [ + 131, + 548, + 479, + 609 + ], + "type": "inline_equation", + "content": "g_{\\pi}^{b}" + }, + { + "bbox": [ + 131, + 548, + 479, + 609 + ], + "type": "text", + "content": " with weights " + }, + { + "bbox": [ + 131, + 548, + 479, + 609 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 131, + 548, + 479, + 609 + ], + "type": "text", + "content": ", alongside an additional multi-resolution hash encoding " + }, + { + "bbox": [ + 131, + 548, + 479, + 609 + ], + "type": "inline_equation", + "content": "h_{\\Omega}^{b}" + }, + { + "bbox": [ + 131, + 548, + 479, + 609 + ], + "type": "text", + "content": ". The modeling process of the background appearance " + }, + { + "bbox": [ + 131, + 548, + 479, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_p^b" + }, + { + "bbox": [ + 131, + 548, + 479, + 609 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 246, + 618, + 480, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 618, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 246, + 618, + 480, + 632 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {p} ^ {b} = g _ {\\pi} ^ {b} \\left(\\mathbf {x} _ {p} ^ {b}, h _ {\\Omega} ^ {b} \\left(\\mathbf {x} _ {p} ^ {b}\\right), Y \\left(\\mathbf {v} _ {p}\\right)\\right), \\tag {9}", + "image_path": "8da49765dfab164bb4f39b164d5035ac2cdbf781dc627f62a2f93372b4e35fe0.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 641, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 479, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 641, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_p^b" + }, + { + "bbox": [ + 132, + 641, + 479, + 666 + ], + "type": "text", + "content": " represents the background sampling point, determined by the camera location " + }, + { + "bbox": [ + 132, + 641, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_o" + }, + { + "bbox": [ + 132, + 641, + 479, + 666 + ], + "type": "text", + "content": ", the veiw direction " + }, + { + "bbox": [ + 132, + 641, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_p" + }, + { + "bbox": [ + 132, + 641, + 479, + 666 + ], + "type": "text", + "content": " and the far distance " + }, + { + "bbox": [ + 132, + 641, + 479, + 666 + ], + "type": "inline_equation", + "content": "F_{p}" + }, + { + "bbox": [ + 132, + 641, + 479, + 666 + ], + "type": "text", + "content": " of the sampling" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 448, + 103 + ], + "type": "text", + "content": "Improving Surface Rendering via CPT with View-Reflection Appearance" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 333, + 130 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 333, + 130 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 333, + 130 + ], + "type": "text", + "content": "boundary. " + }, + { + "bbox": [ + 132, + 115, + 333, + 130 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_p^b" + }, + { + "bbox": [ + 132, + 115, + 333, + 130 + ], + "type": "text", + "content": " is mathematically expressed as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 266, + 134, + 481, + 150 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 134, + 481, + 150 + ], + "spans": [ + { + "bbox": [ + 266, + 134, + 481, + 150 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {p} ^ {b} = \\mathbf {x} _ {o} + 2 F _ {p} \\mathbf {v} _ {p}. \\tag {10}", + "image_path": "e34592c0f7d12b221c1374514099bb6b959dc8427a10c724673a6dd26d878db0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 153, + 479, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 153, + 479, + 178 + ], + "spans": [ + { + "bbox": [ + 131, + 153, + 479, + 178 + ], + "type": "text", + "content": "Following NeuS [24], the foreground is normalized within a unit sphere centered at " + }, + { + "bbox": [ + 131, + 153, + 479, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{0}" + }, + { + "bbox": [ + 131, + 153, + 479, + 178 + ], + "type": "text", + "content": ". As indicated in Fig. 3, the far distance " + }, + { + "bbox": [ + 131, + 153, + 479, + 178 + ], + "type": "inline_equation", + "content": "F_{p}" + }, + { + "bbox": [ + 131, + 153, + 479, + 178 + ], + "type": "text", + "content": " is calculated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 247, + 183, + 481, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 183, + 481, + 196 + ], + "spans": [ + { + "bbox": [ + 247, + 183, + 481, + 196 + ], + "type": "interline_equation", + "content": "m = - \\mathbf {x} _ {o} \\cdot \\mathbf {v} _ {p}; F _ {p} = m + 1. \\tag {11}", + "image_path": "2ceaa206d6bcf1a7d1350b852eaea9bf472c42184e1839260f598a7d7877e22d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 201, + 479, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 201, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 131, + 201, + 479, + 224 + ], + "type": "text", + "content": "This approach facilitates a more versatile background representation, compensating for the absence of masks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 225, + 317, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 225, + 317, + 274 + ], + "spans": [ + { + "bbox": [ + 131, + 225, + 317, + 274 + ], + "type": "text", + "content": "Appearance Integration. Finally, we integrate the background appearance " + }, + { + "bbox": [ + 131, + 225, + 317, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_p^b" + }, + { + "bbox": [ + 131, + 225, + 317, + 274 + ], + "type": "text", + "content": " and foreground appearance " + }, + { + "bbox": [ + 131, + 225, + 317, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_p^f" + }, + { + "bbox": [ + 131, + 225, + 317, + 274 + ], + "type": "text", + "content": " to attain the rendered color, defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 157, + 277, + 314, + 294 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 277, + 314, + 294 + ], + "spans": [ + { + "bbox": [ + 157, + 277, + 314, + 294 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {p} = M _ {p} \\mathbf {c} _ {p} ^ {f} + (1 - M _ {p}) \\mathbf {c} _ {p} ^ {b}. \\tag {12}", + "image_path": "bb87de8ab98632c4117f4c8ec9e1e89e62a57432407159419cbd64625b59dcae.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 297, + 315, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 297, + 315, + 346 + ], + "spans": [ + { + "bbox": [ + 131, + 297, + 315, + 346 + ], + "type": "text", + "content": "Overall, our method enables surface rendering of objects without any known background information, offering greater flexibility and adaptability." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 323, + 222, + 482, + 300 + ], + "blocks": [ + { + "bbox": [ + 323, + 222, + 482, + 300 + ], + "lines": [ + { + "bbox": [ + 323, + 222, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 323, + 222, + 482, + 300 + ], + "type": "image", + "image_path": "ba50007b33ae024c00c2145e20c314139d4244190f6134a11800937f98841ac7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 307, + 481, + 331 + ], + "lines": [ + { + "bbox": [ + 321, + 307, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 321, + 307, + 481, + 331 + ], + "type": "text", + "content": "Fig.3: Illustration of the one-point background sampling." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 359, + 261, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 359, + 261, + 372 + ], + "spans": [ + { + "bbox": [ + 132, + 359, + 261, + 372 + ], + "type": "text", + "content": "3.5 Training Objectives" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 376, + 479, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 376, + 479, + 425 + ], + "spans": [ + { + "bbox": [ + 131, + 376, + 479, + 425 + ], + "type": "text", + "content": "Color loss. We optimize the CPT-VR model by ensuring the colors it generates match real image colors as closely as possible. This is done by calculating the L1 error, which is the average difference between the colors " + }, + { + "bbox": [ + 131, + 376, + 479, + 425 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{c}}_i" + }, + { + "bbox": [ + 131, + 376, + 479, + 425 + ], + "type": "text", + "content": " produced by the model and the ground-truth colors " + }, + { + "bbox": [ + 131, + 376, + 479, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 131, + 376, + 479, + 425 + ], + "type": "text", + "content": " for all pixels " + }, + { + "bbox": [ + 131, + 376, + 479, + 425 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 131, + 376, + 479, + 425 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 255, + 429, + 481, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 429, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 255, + 429, + 481, + 460 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C} = \\frac {1}{P} \\sum_ {i} ^ {P} \\| \\hat {\\mathbf {c}} _ {i} - \\mathbf {c} _ {i} \\| _ {1}. \\tag {13}", + "image_path": "17b4f708ec38d56f0003b2bbf868f208ef9839413e7e2e19e871e3ca6fc59b14.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 465, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 465, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 465, + 482, + 574 + ], + "type": "text", + "content": "Regularization of SDFnet. A well-defined neural SDF exhibits Lipschitz continuity [13] and its gradient adheres to the Eikonal equation [4], ensuring smoothness and stability of the generated field. A valid SDF field holds the property that the normal of a point near the surface and the normal of its projected point remain consistent. For effective closest point transform, we directly constrain our SDFnet based on this normal consistency property [15]. Specifically, it aligns the normal of a deviated point " + }, + { + "bbox": [ + 131, + 465, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 131, + 465, + 482, + 574 + ], + "type": "text", + "content": " with the normal " + }, + { + "bbox": [ + 131, + 465, + 482, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{n}'" + }, + { + "bbox": [ + 131, + 465, + 482, + 574 + ], + "type": "text", + "content": " of its projected point, enhancing stability for surface rendering via the closest point transform of SDF. Mathematically, it is represented by:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 269, + 578, + 481, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 578, + 481, + 591 + ], + "spans": [ + { + "bbox": [ + 269, + 578, + 481, + 591 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {n c} = 1 - \\mathbf {n} \\cdot \\mathbf {n} ^ {\\prime}. \\tag {14}", + "image_path": "8092312bbd31864a7f0033458996e68b93a732198496bc497f0b0f1fa59a8c2a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 597, + 479, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 597, + 479, + 632 + ], + "spans": [ + { + "bbox": [ + 131, + 597, + 479, + 632 + ], + "type": "text", + "content": "Mask Loss. If the mask of the foreground region is available, we can further refine the rendering effect of CPT-VR by including a mask loss. The mask constraint is formulated as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 250, + 635, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 635, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 250, + 635, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {M} = \\frac {1}{P} \\sum_ {i} ^ {P} \\left(\\hat {M} _ {i} - M _ {i}\\right) ^ {2}, \\tag {15}", + "image_path": "5929d49d66a5421191cfa204ec7c8b475536f51027b1a5fd58ae86311dc2596c.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "Z. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\hat{M}_i" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": " represent the predicted mask and the ground truth mask, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": "Depth loss. Inspired by previous works [37] and [35], we utilize depth maps generated by the pre-trained network CasMVSNet [5] to guide our model. We minimize the squared difference (MSE) between the rendered depths " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "(\\hat{D}_i)" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": " and those from the pre-trained network " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "(D_i)" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": ", weighted by the confidence " + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "inline_equation", + "content": "prob_i" + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": " from CasMVSNet, as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 243, + 208, + 480, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 208, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 243, + 208, + 480, + 239 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D} = \\frac {1}{P} \\sum_ {i} ^ {P} p r o b _ {i} \\left(\\hat {D} _ {i} - D _ {i}\\right) ^ {2}, \\tag {16}", + "image_path": "e0214b27aec380b226a3747111927762490eaeba1775d352cfa6331a6165abbe.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 245, + 479, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 245, + 479, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 479, + 268 + ], + "type": "text", + "content": "where the confidence ranges from 0 to 1, and depths considered unreliable are given a confidence of 0." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 269, + 479, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 479, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 479, + 293 + ], + "type": "text", + "content": "Total loss. Our final metric combines all these aspects, adjusting their importance with specific weights " + }, + { + "bbox": [ + 130, + 269, + 479, + 293 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 269, + 479, + 293 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 215, + 301, + 480, + 313 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 301, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 215, + 301, + 480, + 313 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {r g b} \\mathcal {L} _ {r g b} + \\lambda_ {n c} \\mathcal {L} _ {n c} + \\lambda_ {M} \\mathcal {L} _ {M} + \\lambda_ {D} \\mathcal {L} _ {D}. \\tag {17}", + "image_path": "2bcc7ef337e5eae3ae3fccffbc6abaa25438df29996d16dad4eeee72a462bd6a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 319, + 477, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 319, + 477, + 333 + ], + "spans": [ + { + "bbox": [ + 131, + 319, + 477, + 333 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 319, + 477, + 333 + ], + "type": "inline_equation", + "content": "\\lambda_{rgb}" + }, + { + "bbox": [ + 131, + 319, + 477, + 333 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 319, + 477, + 333 + ], + "type": "inline_equation", + "content": "\\lambda_{nc}" + }, + { + "bbox": [ + 131, + 319, + 477, + 333 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 319, + 477, + 333 + ], + "type": "inline_equation", + "content": "\\lambda_M" + }, + { + "bbox": [ + 131, + 319, + 477, + 333 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 319, + 477, + 333 + ], + "type": "inline_equation", + "content": "\\lambda_D" + }, + { + "bbox": [ + 131, + 319, + 477, + 333 + ], + "type": "text", + "content": " are empirically set to 3, 0.01, 1, and 1, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 349, + 230, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 349, + 230, + 363 + ], + "spans": [ + { + "bbox": [ + 131, + 349, + 230, + 363 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 372, + 272, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 372, + 272, + 384 + ], + "spans": [ + { + "bbox": [ + 131, + 372, + 272, + 384 + ], + "type": "text", + "content": "4.1 Experimental Settings" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 389, + 480, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 480, + 521 + ], + "type": "text", + "content": "Implementation details. For the DTU dataset, we train CPT-VR for 10,000 iterations with a batch size of one image, targeting a Marching Cubes resolution of " + }, + { + "bbox": [ + 130, + 389, + 480, + 521 + ], + "type": "inline_equation", + "content": "256^3" + }, + { + "bbox": [ + 130, + 389, + 480, + 521 + ], + "type": "text", + "content": ". In each training batch, we render one view at a resolution of 800 " + }, + { + "bbox": [ + 130, + 389, + 480, + 521 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 130, + 389, + 480, + 521 + ], + "type": "text", + "content": " 600 pixels. Note that, when we only utilize depth loss and color loss for supervision, we employ a higher Marching Cubes resolution of " + }, + { + "bbox": [ + 130, + 389, + 480, + 521 + ], + "type": "inline_equation", + "content": "512^3" + }, + { + "bbox": [ + 130, + 389, + 480, + 521 + ], + "type": "text", + "content": " and render images at " + }, + { + "bbox": [ + 130, + 389, + 480, + 521 + ], + "type": "inline_equation", + "content": "1600 \\times 1200" + }, + { + "bbox": [ + 130, + 389, + 480, + 521 + ], + "type": "text", + "content": " pixels. For BlendedMVS, CPT-VR is trained for 20,000 iterations with a batch size of one image and a Marching Cubes resolution of " + }, + { + "bbox": [ + 130, + 389, + 480, + 521 + ], + "type": "inline_equation", + "content": "256^3" + }, + { + "bbox": [ + 130, + 389, + 480, + 521 + ], + "type": "text", + "content": ", rendering images at " + }, + { + "bbox": [ + 130, + 389, + 480, + 521 + ], + "type": "inline_equation", + "content": "768 \\times 576" + }, + { + "bbox": [ + 130, + 389, + 480, + 521 + ], + "type": "text", + "content": " pixels. Moreover, the initial learning rate of our model is set at 0.005, with a final epoch decay factor of 0.1 for DTU and 0.005 for BlendedMVS. More details about the architecture and the isosurface configuration are provided in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 521, + 480, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 480, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 480, + 629 + ], + "type": "text", + "content": "Datasets. All experiments are conducted on DTU [9] and BlendedMVS [31]. Objects in DTU are collected in a controlled lab setting, with ground truth data that includes images, masks, and point clouds. For our training and evaluation, we leverage 15 objects identified by IDR with challenges such as specular highlights, delicate structures, and areas hidden from view. Meanwhile, BlendedMVS encompasses a diverse range of objects and scenes captured in natural settings. From this dataset, we select 7 difficult objects as NeuS [24] and the 'Bull' (with specular highlights) to further assess the model performance in reconstructing objects against complex backgrounds." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 645, + 282, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 645, + 282, + 657 + ], + "spans": [ + { + "bbox": [ + 131, + 645, + 282, + 657 + ], + "type": "text", + "content": "4.2 Quantitative Evaluation" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 448, + 103 + ], + "type": "text", + "content": "Improving Surface Rendering via CPT with View-Reflection Appearance" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 301, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 301, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 301, + 247 + ], + "type": "text", + "content": "Analysis on DTU. We benchmark our method against the state-of-the-art methods on the DTU dataset. DTU dataset contains 128 scans captured in the laboratory. In this work, we evaluate the surface mesh using the same set of scans as in [34,37]. As denoted in Tab. 2, our method consistently outperforms other surface rendering methods on most scans. Note that, our method achieves the lowest" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 247, + 482, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 332 + ], + "type": "text", + "content": "value in mean Chamfer Distance (0.59), which is " + }, + { + "bbox": [ + 130, + 247, + 482, + 332 + ], + "type": "inline_equation", + "content": "16.9\\%" + }, + { + "bbox": [ + 130, + 247, + 482, + 332 + ], + "type": "text", + "content": " lower than the second-best surface rendering method RegSDF (0.71). Since SDFnet in our method has a similar structure to it in NeuS2, we also compare NeuS2 and recent volume rendering methods [26, 41] with CPT-VR. As demonstrated in Tab. 2, whether the two methods incorporate background information from the object mask " + }, + { + "bbox": [ + 130, + 247, + 482, + 332 + ], + "type": "inline_equation", + "content": "(\\mathrm{w} / \\mathrm{M})" + }, + { + "bbox": [ + 130, + 247, + 482, + 332 + ], + "type": "text", + "content": " or use our proposed one-point background model " + }, + { + "bbox": [ + 130, + 247, + 482, + 332 + ], + "type": "inline_equation", + "content": "(\\mathrm{w} / \\mathrm{BG1})" + }, + { + "bbox": [ + 130, + 247, + 482, + 332 + ], + "type": "text", + "content": " to attain background information, our method surpasses NeuS2 on the vast majority of ScanIDs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 307, + 108, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 482, + 175 + ], + "type": "text", + "content": "Table 1: Comparison of the average rendering time on DTU. The rendered image resolution of all methods is " + }, + { + "bbox": [ + 307, + 108, + 482, + 175 + ], + "type": "inline_equation", + "content": "1600 \\times 1200" + }, + { + "bbox": [ + 307, + 108, + 482, + 175 + ], + "type": "text", + "content": ", and the resolution of Marching Cubes is " + }, + { + "bbox": [ + 307, + 108, + 482, + 175 + ], + "type": "inline_equation", + "content": "512^{3}" + }, + { + "bbox": [ + 307, + 108, + 482, + 175 + ], + "type": "text", + "content": ". The rendering time of our method is significantly lower than that of others." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 324, + 175, + 463, + 243 + ], + "blocks": [ + { + "bbox": [ + 324, + 175, + 463, + 243 + ], + "lines": [ + { + "bbox": [ + 324, + 175, + 463, + 243 + ], + "spans": [ + { + "bbox": [ + 324, + 175, + 463, + 243 + ], + "type": "table", + "html": "
MethodRendering Time (s)
IDR27.233
NeuS2 (w /M)1.979
Ours (w/ M)0.057
Ours (w/ BG1)0.104
", + "image_path": "9733f11fd782c4f65e25275e97947de2095f62c5b1e84b289f6509199817ed1f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 348, + 482, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 482, + 440 + ], + "type": "text", + "content": "Table 2: Quantitative results on DTU dataset. Here, we evaluate all methods on Chamfer Distance (CD). The lower the CD value, the better the performance. For a fair comparison, we also present the results of the volume rendering method NeuS2 since our SDFnet has a similar structure to it. \"w/ M\" means providing the prior background knowledge by masks, while \"w/ BG1\" means employing our one-point background model without any prior background knowledge. The best and second-best results are highlighted by different colors. The results illustrated that CPT-VR attains better performance among most ScanIDs." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 132, + 444, + 491, + 582 + ], + "blocks": [ + { + "bbox": [ + 132, + 444, + 491, + 582 + ], + "lines": [ + { + "bbox": [ + 132, + 444, + 491, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 444, + 491, + 582 + ], + "type": "table", + "html": "
ScanID243740556365698397105106110114118122Mean
DMTet [19]1.841.731.050.611.661.370.681.283.821.070.911.150.370.540.551.24
IDR [34]1.631.870.630.481.040.790.771.331.160.760.670.900.420.510.530.90
MVSDF [37]0.831.760.880.441.110.900.751.261.021.350.870.840.340.470.460.89
FastMESH [39]0.651.480.570.401.480.770.560.860.840.940.720.810.520.490.540.77
RegSDF [35]0.591.410.630.421.340.620.590.890.911.020.600.590.290.400.380.71
HF-NeuS [26] (w/ M)1.111.280.610.470.970.680.621.340.910.730.531.820.380.540.510.83
LOD-NeuS [41] (w/ M)0.650.910.370.481.050.870.821.220.950.690.561.300.420.580.570.76
NeuS2 [25] (w/ M)0.560.760.490.370.920.710.761.221.080.630.590.890.400.480.550.70
NeuS2 [25] (w/ BG1)0.690.860.830.341.050.680.651.051.120.670.581.410.370.500.520.75
Ours (w/ M)0.550.720.350.380.850.590.540.810.830.670.530.590.320.380.370.56
Ours (w/ BG1)0.430.730.360.320.930.610.610.891.020.680.480.730.320.370.370.59
", + "image_path": "cb6c3a5708e2f893f400baa119114ceb9dc8f15203adcc22cdb1adc09a4394fe.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Moreover, we also measure the rendered results with PSNR to the ground truths. As indicated in Tab. 3, with the introduction of background information (BG1: " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "), our method achieves the highest PSNR values on most ScanIDs. Additionally, it can be observed that when using the one-point background model, both NeuS2 and CPT-VR are able to achieve similar or better performance compared to leveraging background information from masks (BG1: " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "Z. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 193, + 488, + 255 + ], + "blocks": [ + { + "bbox": [ + 130, + 110, + 482, + 186 + ], + "lines": [ + { + "bbox": [ + 130, + 110, + 482, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 110, + 482, + 186 + ], + "type": "text", + "content": "Table 3: Quantitative comparison of 2D view synthesis on DTU dataset. We measure all methods based on the metric PSNR. A higher PSNR indicates better synthesis quality. Here, B1 represents whether to introduce extra background knowledge into the method. " + }, + { + "bbox": [ + 130, + 110, + 482, + 186 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 130, + 110, + 482, + 186 + ], + "type": "text", + "content": " means to leverage the background information from masks, while " + }, + { + "bbox": [ + 130, + 110, + 482, + 186 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 130, + 110, + 482, + 186 + ], + "type": "text", + "content": " indicates utilizing our one-point background model without introducing any background information. We color code the best and second-best results. Compared to NeuS2, our method achieves much better appearance reconstruction results." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 193, + 488, + 255 + ], + "lines": [ + { + "bbox": [ + 132, + 193, + 488, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 488, + 255 + ], + "type": "table", + "html": "
ScanIDBG1243740556365698397105106110114118122Mean
NeuS2✗28.4427.1429.7029.6731.7527.8324.8431.2426.8630.5726.0528.9328.9827.8232.4828.82
NeuS2✓27.7124.9927.5128.8920.5530.1429.0127.7026.1631.0933.1030.6930.4832.7734.1028.99
Ours✗28.1527.3128.5429.5234.0429.8329.0035.9229.3233.0829.4531.1630.7730.5038.3831.00
Ours✓27.3025.5926.8827.7730.9130.0629.2828.0526.6828.9232.0330.1829.0234.0234.0229.38
", + "image_path": "e3ed2bd4c727a0763f2888e06485bd2ec24570969cb2046824255c7c74e804f6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 138, + 319, + 474, + 361 + ], + "blocks": [ + { + "bbox": [ + 130, + 260, + 482, + 315 + ], + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 315 + ], + "type": "text", + "content": "Table 4: Quantitative comparison on BlendedMVS dataset. We measure the method on the metric Chamfer Distance (CD). The lower CD values indicate better performance. For a fair comparison, we adopt our one-point background model to NeuS2. Compared to NeuS2, CPT-VR attains better geometry reconstruction results among most of the categories." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 319, + 474, + 361 + ], + "lines": [ + { + "bbox": [ + 138, + 319, + 474, + 361 + ], + "spans": [ + { + "bbox": [ + 138, + 319, + 474, + 361 + ], + "type": "table", + "html": "
MethodBearClockDogDurianManSculptureBullStoneMean
NeuS20.380.470.950.470.460.400.650.380.52
Ours0.350.450.680.480.460.390.450.380.45
", + "image_path": "e2a1825a603f9b9eb28897618583d4c663f5d693f7a75285a9aa0eaa181d8b38.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 376, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 483 + ], + "type": "text", + "content": "Analysis on BlendedMVS. As illustrated in Tab. 4, we compare our method with NeuS2 on the eight selected categories with complex backgrounds. Since the scales are unknown in BlendedMVS, we scale the bounding box of each object to 100 units to measure their performance on Chamfer Distance (CD). CPT-VR achieves the best performance in the categories of 'Bear', 'Clock', 'Dog', 'Sculpture', and 'Bull', with the CD value for 'Man' and 'Stone' being on par with NeuS2, and only 0.01 behind NeuS2 for 'Durian'. Moreover, CPT-VR outperforms NeuS2 by " + }, + { + "bbox": [ + 130, + 376, + 482, + 483 + ], + "type": "inline_equation", + "content": "14.8\\%" + }, + { + "bbox": [ + 130, + 376, + 482, + 483 + ], + "type": "text", + "content": ". These results demonstrate the advantage of our method in reconstructing objects within complex environments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 483, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 556 + ], + "type": "text", + "content": "Analysis of Method Efficiency. We evaluate the rendering time of IDR, NeuS2, and CPT-VR on DTU. The results in Tab. 1 suggest the rendering time of IDR (24.233s) and NeuS2 (1.97s) is lengthy. Compared with them, our method achieves real-time rendering under the cases without introducing background priors being about 19 times faster than NeuS2. This is primarily because NeuS2 utilizes the dense sampling method in the volume rendering process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 560, + 274, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 560, + 274, + 572 + ], + "spans": [ + { + "bbox": [ + 131, + 560, + 274, + 572 + ], + "type": "text", + "content": "4.3 Qualitative Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 640 + ], + "type": "text", + "content": "Impact of Specular Highlights. Fig. 4 illustrates the qualitative results on the DTU dataset. We compare our CPT-VR(w/M) model with IDR [34], DMTet [19], MVsDF [37], and NeuS2 [25]. As shown by the close-up regions in Fig. 4, the specular regions constructed by other methods exhibit obvious indentations, while our method can construct these regions with high quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "Impact of Complex Backgrounds. Fig. 5 demonstrates the visual results of reconstructing objects with complex backgrounds. Since existing surface render-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "text", + "content": "Improving Surface Rendering via CPT with View-Reflection Appearance" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "ing methods cannot handle the cases without any background prior, we specifically evaluate NeuS2 and our method on BlendedMVS. As the elliptical area demonstrated, our method can effectively model the geometry of foreground objects without introducing background information, particularly the details on the surface of objects. Moreover, the rendering results of our method closely resemble the input image." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 151, + 209, + 463, + 296 + ], + "blocks": [ + { + "bbox": [ + 151, + 209, + 463, + 296 + ], + "lines": [ + { + "bbox": [ + 151, + 209, + 463, + 296 + ], + "spans": [ + { + "bbox": [ + 151, + 209, + 463, + 296 + ], + "type": "image", + "image_path": "59960e7ac11b8e536f954b9cbb78f2ccc369a13f0cee48625ae01bd4e4627fcc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 307, + 482, + 341 + ], + "lines": [ + { + "bbox": [ + 131, + 307, + 482, + 341 + ], + "spans": [ + { + "bbox": [ + 131, + 307, + 482, + 341 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparisons on objects in DTU with specular highlights. The close-ups in yellow highlight the specific regions for comparison. The geometry reconstruction results demonstrate that our method is robust against specular highlights." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 153, + 387, + 465, + 531 + ], + "blocks": [ + { + "bbox": [ + 153, + 387, + 465, + 531 + ], + "lines": [ + { + "bbox": [ + 153, + 387, + 465, + 531 + ], + "spans": [ + { + "bbox": [ + 153, + 387, + 465, + 531 + ], + "type": "image", + "image_path": "594bbbeb7067c1830b4433ee94e0d1d2fa430c9536d6c0267e63678028ea101f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 537, + 482, + 583 + ], + "lines": [ + { + "bbox": [ + 130, + 537, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 537, + 482, + 583 + ], + "type": "text", + "content": "Fig.5: Qualitative results of objects with complex background on BlendedMVS. We outline the close-up regions for detailed comparison. It is evident that without the mask guidance, our method also exhibits strong capabilities in modeling foreground objects, especially for the finer-grained surface structures." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Impact of the Sampling Point Amount. Tab. 5 suggests the impact of varying the number of sampling points and their distances. The optimal result is achieved when sampling one point at " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "2F_{p}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ", and the " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "CD" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " is 0.66. Conversely, sampling either too closely (at " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "F_{p}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ") or too distantly (at " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "1000F_{p}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ") causes the sampled points to stray outside the desired background range, leading to suboptimal" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "Z. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "results with CD values of 0.68 and 0.73, respectively. Additionally, increasing the number of sampling points in our approach tends to accentuate background details, thereby negatively impacting the accuracy of foreground object modeling. More background points would render the model to prioritize background fitting over the foreground, thus generating inferior results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 188, + 239, + 200 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 188, + 239, + 200 + ], + "spans": [ + { + "bbox": [ + 132, + 188, + 239, + 200 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 212, + 302, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 302, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 302, + 390 + ], + "type": "text", + "content": "As shown in Tab. 6, we set five settings to demonstrate the superiority of our method design. From the results of CPT-V-M, CPT-R-M, and CPT-VR-M, it is evident that introducing view-ray directions brings greater benefits to geometry reconstruction than introducing the two vectors alone. Meanwhile, the comparable " + }, + { + "bbox": [ + 130, + 212, + 302, + 390 + ], + "type": "inline_equation", + "content": "CD" + }, + { + "bbox": [ + 130, + 212, + 302, + 390 + ], + "type": "text", + "content": " values of CPT-VR-M and CPT-VR-BG1 demonstrate that the one-point background model can largely eliminate the negative impact of complex backgrounds in images on rendering the target object, enabling the model to" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 307, + 226, + 482, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 226, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 307, + 226, + 482, + 314 + ], + "type": "text", + "content": "Table 5: Impact of the Number of Background Points. Here, Num. indicates the sampling amount of points. Pos. is the sampling positions of points along the ray direction. " + }, + { + "bbox": [ + 307, + 226, + 482, + 314 + ], + "type": "inline_equation", + "content": "F_{p}" + }, + { + "bbox": [ + 307, + 226, + 482, + 314 + ], + "type": "text", + "content": " is defined in Sec. 3.4. All settings are evaluated on Chamfer Distance (CD). Sampling one point at " + }, + { + "bbox": [ + 307, + 226, + 482, + 314 + ], + "type": "inline_equation", + "content": "2F_{p}" + }, + { + "bbox": [ + 307, + 226, + 482, + 314 + ], + "type": "text", + "content": " attains the best performance." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 323, + 316, + 465, + 380 + ], + "blocks": [ + { + "bbox": [ + 323, + 316, + 465, + 380 + ], + "lines": [ + { + "bbox": [ + 323, + 316, + 465, + 380 + ], + "spans": [ + { + "bbox": [ + 323, + 316, + 465, + 380 + ], + "type": "table", + "html": "
MethodNum.Pos.CD↓
CPT-VR22Fp, 1000Fp0.74
CPT-VR2Fp, 2Fp0.67
CPT-VR11000Fp0.73
CPT-VR1Fp0.68
CPT-VR12Fp0.66
", + "image_path": "e725b8325e6401dd117c2c5b92696d09982234508fc87d07492264af4642ffb9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 392, + 482, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 428 + ], + "type": "text", + "content": "no longer be dependent on any background knowledge. Moreover, the results of CPT-VR-BG1 and CPT-VR-BG1-D also verify that introducing depth supervision can further boost the reconstruction quality of our method." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 183, + 445, + 260, + 510 + ], + "blocks": [ + { + "bbox": [ + 183, + 445, + 260, + 510 + ], + "lines": [ + { + "bbox": [ + 183, + 445, + 260, + 510 + ], + "spans": [ + { + "bbox": [ + 183, + 445, + 260, + 510 + ], + "type": "image", + "image_path": "3b474c590a27cd876841f7b07c5181b6372e93c3af2fbd865472b83149b077e7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 189, + 512, + 259, + 577 + ], + "blocks": [ + { + "bbox": [ + 189, + 512, + 259, + 577 + ], + "lines": [ + { + "bbox": [ + 189, + 512, + 259, + 577 + ], + "spans": [ + { + "bbox": [ + 189, + 512, + 259, + 577 + ], + "type": "image", + "image_path": "cd4d922a9c87d431a84ed5e55367b084b0ecf1ab2916371c62e8588c1eb141a4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 586, + 482, + 653 + ], + "lines": [ + { + "bbox": [ + 130, + 586, + 482, + 653 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 482, + 653 + ], + "type": "text", + "content": "Fig.6: Qualitative results of ablation studies on DTU. The close-ups in red highlight the specific regions for comparison. The content in the bracket is the Chamfer Distance, indicating the reconstruction quality. Incorporating the reflection vector enhances the ability of the model to simulate the appearance representation on specular regions. In addition, when utilizing the background model to attain the background information, introducing depth supervision greatly improves the reconstruction quality." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 268, + 445, + 341, + 509 + ], + "blocks": [ + { + "bbox": [ + 268, + 445, + 341, + 509 + ], + "lines": [ + { + "bbox": [ + 268, + 445, + 341, + 509 + ], + "spans": [ + { + "bbox": [ + 268, + 445, + 341, + 509 + ], + "type": "image", + "image_path": "5e81aaf8d80130cf46e63e2a718e0b07b2d9929bec0a21d9ce83de1d3c78b2ca.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 269, + 512, + 341, + 577 + ], + "blocks": [ + { + "bbox": [ + 269, + 512, + 341, + 577 + ], + "lines": [ + { + "bbox": [ + 269, + 512, + 341, + 577 + ], + "spans": [ + { + "bbox": [ + 269, + 512, + 341, + 577 + ], + "type": "image", + "image_path": "cb4ba2cecde4e49a0b71a9e1e5f252a169b7614ff6c704d9b862fbaee26d9440.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 352, + 446, + 427, + 510 + ], + "blocks": [ + { + "bbox": [ + 352, + 446, + 427, + 510 + ], + "lines": [ + { + "bbox": [ + 352, + 446, + 427, + 510 + ], + "spans": [ + { + "bbox": [ + 352, + 446, + 427, + 510 + ], + "type": "image", + "image_path": "d0cb822b8e330a8debf845bdfa826f12d9fbf8d5c8cdcd7d4784f4e776c6a67d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 354, + 514, + 427, + 577 + ], + "blocks": [ + { + "bbox": [ + 354, + 514, + 427, + 577 + ], + "lines": [ + { + "bbox": [ + 354, + 514, + 427, + 577 + ], + "spans": [ + { + "bbox": [ + 354, + 514, + 427, + 577 + ], + "type": "image", + "image_path": "9e5be864a994b94fdcd4c0842cbbb2df6565090fb3590b1f23cdb48821a071e5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 448, + 103 + ], + "type": "text", + "content": "Improving Surface Rendering via CPT with View-Reflection Appearance" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 148, + 163, + 464, + 254 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 6: Ablation study of our design choices. We evaluate all designs on Chamfer Distance (CD). Results demonstrate the effectiveness of all our design choices and the importance of using accurate geometry representation when adding the reflection vector for appearance modeling." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 163, + 464, + 254 + ], + "lines": [ + { + "bbox": [ + 148, + 163, + 464, + 254 + ], + "spans": [ + { + "bbox": [ + 148, + 163, + 464, + 254 + ], + "type": "table", + "html": "
SettingsDesign ChoicesCD↓
View VectorReflection VectorBackground ModelLCLMLD
CPT-V-M✓✓✓0.72
CPT-R-M✓✓✓0.83
CPT-VR-M✓✓✓✓0.65
CPT-VR-BG1✓✓✓✓0.66
CPT-VR-BG1-D✓✓✓✓✓0.59
", + "image_path": "effcc41dc21fd21d7bfdb79d0c2a473f43874a3a8dbc2dc498802f211df1cef8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 262, + 482, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 393 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 393 + ], + "type": "text", + "content": "As illustrated in Sec. 1, we argue that only when the surface sampling points are accurate can the introduction of the reflection vector achieve better appearance modeling of specular regions. To further demonstrate that, we incorporate the reflection vector into the surface rendering method DMTet [19] and IDR [34]. Tab. 7 indicates that introducing the reflection vector reduces the geometry reconstruction ability of DMTet (" + }, + { + "bbox": [ + 130, + 262, + 482, + 393 + ], + "type": "inline_equation", + "content": "CD" + }, + { + "bbox": [ + 130, + 262, + 482, + 393 + ], + "type": "text", + "content": " rises from 1.24 to 1.63) and IDR (" + }, + { + "bbox": [ + 130, + 262, + 482, + 393 + ], + "type": "inline_equation", + "content": "CD" + }, + { + "bbox": [ + 130, + 262, + 482, + 393 + ], + "type": "text", + "content": " rises from 0.90 to 1.68). Conversely, the sampling surface points of our method with CPT are more accurate. The accurate points facilitate the modeling of the reflection vector, avoiding a greater backpropagation error. Thus, incorporating the reflection vector into our method enables our model more robust against specular highlights. Fig. 6 also demonstrates the superiority of our design choices." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 142, + 445, + 469, + 491 + ], + "blocks": [ + { + "bbox": [ + 130, + 406, + 482, + 440 + ], + "lines": [ + { + "bbox": [ + 130, + 406, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 406, + 482, + 440 + ], + "type": "text", + "content": "Table 7: Impact of view-reflection vectors. All methods are evaluated with masks on Chamfer Distance (CD). With the reflection vector, the performance of our method has significantly improved, while DMTet and IDR become worse." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 142, + 445, + 469, + 491 + ], + "lines": [ + { + "bbox": [ + 142, + 445, + 469, + 491 + ], + "spans": [ + { + "bbox": [ + 142, + 445, + 469, + 491 + ], + "type": "table", + "html": "
MethodIDRIDRDMTetDMTetOursOursOurs
View Vector (V)✓✓✓✓✓✓
Reflection Vector (R)✓✓✓✓
CD↓0.901.681.241.630.720.830.65
", + "image_path": "de96f119f9cf158f4b643514fbfc7f8e0cbbaa114eb5e46bb72fe1d46aafdeb1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 504, + 226, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 226, + 517 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 226, + 517 + ], + "type": "text", + "content": "5 Conclusions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we present a novel surface rendering method CPT-VR. Based on the CPT algorithm, we correct the deviated points approximated by linear interpolation, enhancing the geometric accuracy. Meanwhile, benefiting from the accurate geometry representation, we incorporate view-reflection vectors into the appearance molding process, which enables our method against the specular highlights. Moreover, we devise a background model to enable our model to handle cases with complex backgrounds without any background knowledge. Experiments on the popular 3D reconstruction datasets demonstrate the superiority of our method in surface rendering, especially for cases with complex structures and specular highlights. We hope that our designs can provide some insights for future works." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "Z. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 139, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 139, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 480, + 172 + ], + "type": "text", + "content": "1. Bangaru, S.P., Gharbi, M., Luan, F., Li, T.M., Sunkavalli, K., Hasan, M., Bi, S., Xu, Z., Bernstein, G., Durand, F.: Differentiable rendering of neural sdfs through reparameterization. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "text", + "content": "2. Darmon, F., Bascle, B., Devaux, J.C., Monasse, P., Aubry, M.: Improving neural implicit surfaces geometry with patch warping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6260-6269 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 206, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 480, + 237 + ], + "type": "text", + "content": "3. Fu, Q., Xu, Q., Ong, Y.S., Tao, W.: Geo-neus: geometry-consistent neural implicit surfaces learning for multi-view reconstruction. arXiv preprint arXiv:2205.15848 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "type": "text", + "content": "4. Gropp, A., Yariv, L., Haim, N., Atzmon, M., Lipman, Y.: Implicit geometric regularization for learning shapes. In: International Conference on Machine Learning. pp. 3789-3799. PMLR (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 271, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 314 + ], + "type": "text", + "content": "5. Gu, X., Fan, Z., Zhu, S., Dai, Z., Tan, F., Tan, P.: Cascade cost volume for high-resolution multi-view stereo and stereo matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2495-2504 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 315, + 480, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 315, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 315, + 480, + 348 + ], + "type": "text", + "content": "6. Guo, Y.C., Cao, Y.P., Wang, C., He, Y., Shan, Y., Qie, X., Zhang, S.H.: Vmesh: Hybrid volume-mesh representation for efficient view synthesis. arXiv preprint arXiv:2303.16184 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 348, + 480, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 480, + 370 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 480, + 370 + ], + "type": "text", + "content": "7. Hart, J.C.: Sphere tracing: A geometric method for the antiaised ray tracing of implicit surfaces. The Visual Computer 12(10), 527-545 (1996)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 370, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 480, + 403 + ], + "type": "text", + "content": "8. Hasselgren, J., Hofmann, N., Munkberg, J.: Shape, light, and material decomposition from images using monte carlo rendering and denoising. Advances in Neural Information Processing Systems 35, 22856-22869 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "text", + "content": "9. Jensen, R., Dahl, A., Vogiatzis, G., Tola, E., Aanaes, H.: Large scale multi-view stereopsis evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 406-413 (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "type": "text", + "content": "0. Laine, S., Hellsten, J., Karras, T., Seol, Y., Lehtinen, J., Aila, T.: Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG) 39(6), 1-14 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "type": "text", + "content": "1. Li, H., Yang, X., Zhai, H., Liu, Y., Bao, H., Zhang, G.: Vox-surf: Voxel-based implicit surface representation. IEEE Transactions on Visualization and Computer Graphics (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 501, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 480, + 544 + ], + "type": "text", + "content": "2. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 544, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 578 + ], + "type": "text", + "content": "3. Liu, H.T.D., Williams, F., Jacobson, A., Fidler, S., Litany, O.: Learning smooth neural functions via lipschitz regularization. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-13 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 578, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 600 + ], + "type": "text", + "content": "4. Lorensen, W.E., Cline, H.E.: Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics 21(4), 163-169 (1987)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "text", + "content": "5. Ma, B., Zhou, J., Liu, Y.S., Han, Z.: Towards better gradient consistency for neural signed distance functions via level set alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17724-17734 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 666 + ], + "type": "text", + "content": "6. Mehta, I., Chandraker, M., Ramamoorthi, R.: A level set theory for neural implicit evolution under explicit flows. In: European Conference on Computer Vision. pp. 711-729. Springer (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "text", + "content": "Improving Surface Rendering via CPT with View-Reflection Appearance" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "17. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I. pp. 405-421 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 481, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 481, + 204 + ], + "type": "text", + "content": "18. Munkberg, J., Hasselgren, J., Shen, T., Gao, J., Chen, W., Evans, A., Müller, T., Fidler, S.: Extracting triangular 3d models, materials, and lighting from images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8280-8290 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 204, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 204, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 204, + 481, + 236 + ], + "type": "text", + "content": "19. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems 34, 6087-6101 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "type": "text", + "content": "20. Sun, J., Chen, X., Wang, Q., Li, Z., Averbuch-Elor, H., Zhou, X., Snavely, N.: Neural 3d reconstruction in the wild. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-9 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 268, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 481, + 311 + ], + "type": "text", + "content": "21. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5481-5490. IEEE (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 311, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 481, + 332 + ], + "type": "text", + "content": "22. Vicini, D., Speierer, S., Jakob, W.: Differentiable signed distance function rendering. ACM Transactions on Graphics (TOG) 41(4), 1-18 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 332, + 481, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 481, + 364 + ], + "type": "text", + "content": "23. Walker, T., Mariotti, O., Vaxman, A., Bilen, H.: Explicit neural surfaces: Learning continuous geometry with deformation fields. arXiv preprint arXiv:2306.02956 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 481, + 396 + ], + "type": "text", + "content": "24. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems 34, 27171-27183 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 396, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 439 + ], + "type": "text", + "content": "25. Wang, Y., Han, Q., Habermann, M., Daniilidis, K., Theobalt, C., Liu, L.: Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3295-3306 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "type": "text", + "content": "26. Wang, Y., Skorokhodov, I., Wonka, P.: Hf-neus: Improved surface reconstruction using high-frequency details. Advances in Neural Information Processing Systems 35, 1966-1978 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 472, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 493 + ], + "type": "text", + "content": "27. Wang, Y., Skorokhodov, I., Wonka, P.: Improved surface reconstruction using high-frequency details. arXiv preprint arXiv:2206.07850 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "type": "text", + "content": "28. Wang, Y., Skorokhodov, I., Wonka, P.: Pet-neus: Positional encoding tri-planes for neural surfaces. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12598-12607 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 525, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 567 + ], + "type": "text", + "content": "29. Worchel, M., Diaz, R., Hu, W., Schreer, O., Feldmann, I., Eisert, P.: Multiview mesh reconstruction with neural deferred shading. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6187-6197 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "text", + "content": "30. Wu, T., Wang, J., Pan, X., Xu, X., Theobalt, C., Liu, Z., Lin, D.: Voxurf: Voxel-based efficient and accurate neural surface reconstruction. arXiv preprint arXiv:2208.12697 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 601, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 643 + ], + "type": "text", + "content": "31. Yao, Y., Luo, Z., Li, S., Zhang, J., Ren, Y., Zhou, L., Fang, T., Quan, L.: Blended-mvs: A large-scale dataset for generalized multi-view stereo networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1790-1799 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "32. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems 34, 4805-4815 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "Z. Hu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 402 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "33. Yariv, L., Hedman, P., Reiser, C., Verbin, D., Srinivasan, P.P., Szeliski, R., Barron, J.T., Mildenhall, B.: Bakedsdf: Meshing neural sdfs for real-time view synthesis. arXiv preprint arXiv:2302.14859 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "text", + "content": "34. Yariv, L., Kasten, Y., Moran, D., Galun, M., Atzmon, M., Ronen, B., Lipman, Y.: Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems 33, 2492-2502 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 226 + ], + "type": "text", + "content": "35. Zhang, J., Yao, Y., Li, S., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Critical regularizations for neural surface reconstruction in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6270-6279 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "36. Zhang, J., Yao, Y., Li, S., Luo, Z., Fang, T.: Visibility-aware multi-view stereo network. British Machine Vision Conference (BMVC) (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 282 + ], + "type": "text", + "content": "37. Zhang, J., Yao, Y., Quan, L.: Learning signed distance field for multi-view surface reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6525-6534 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "type": "text", + "content": "38. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "39. Zhang, Y., Zhu, J., Lin, L.: Fastmesh: Fast surface reconstruction by hexagonal mesh-based neural rendering. arXiv preprint arXiv:2305.17858 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 369 + ], + "type": "text", + "content": "40. Zhang, Y., Hu, Z., Wu, H., Zhao, M., Li, L., Zou, Z., Fan, C.: Towards unbiased volume rendering of neural implicit surfaces with geometry priors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4359-4368 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "41. Zhuang, Y., Zhang, Q., Feng, Y., Zhu, H., Yao, Y., Li, X., Cao, Y.P., Shan, Y., Cao, X.: Anti-aliased neural implicit surfaces with encoding level of detail. In: SIGGRAPH Asia 2023 Conference Papers. pp. 1-10 (2023)" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "text", + "content": "Improving Surface Rendering via CPT with View-Reflection Appearance" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CRM_ Single Image to 3D Textured Mesh with Convolutional Reconstruction Model/9e4b61b2-e1dd-4f99-9ab9-4fe8376f780a_content_list.json b/2024/CRM_ Single Image to 3D Textured Mesh with Convolutional Reconstruction Model/9e4b61b2-e1dd-4f99-9ab9-4fe8376f780a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8c7eb56b3aeae30e2b98e842def628db6e13e85d --- /dev/null +++ b/2024/CRM_ Single Image to 3D Textured Mesh with Convolutional Reconstruction Model/9e4b61b2-e1dd-4f99-9ab9-4fe8376f780a_content_list.json @@ -0,0 +1,2099 @@ +[ + { + "type": "text", + "text": "CRM: Single Image to 3D Textured Mesh with Convolutional Reconstruction Model", + "text_level": 1, + "bbox": [ + 232, + 140, + 772, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhengyi Wang $^{1,2}$ , Yikai Wang $^{1,2}$ , Yifei Chen $^{1}$ , Chendong Xiang $^{1,2}$ , Shuo Chen $^{1}$ , Dajiang Yu $^{1}$ , Chongxuan Li $^{3}$ , Hang Su $^{1}$ and Jun Zhu $^{\\star 1,2}$", + "bbox": [ + 238, + 210, + 761, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Dept. of Comp. Sci. & Tech., BNrist Center, Tsinghua-Bosch Joint ML Center, Tsinghua University;", + "bbox": [ + 227, + 253, + 772, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 ShengShu, Beijing, China;", + "bbox": [ + 405, + 282, + 594, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Gaoling School of Artificial Intelligence, Renmin University of China, Beijing Key Laboratory of Big Data Management and Analysis Methods, Beijing, China.", + "bbox": [ + 222, + 296, + 779, + 325 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/179d3eff26b50d33ee74c72b3ffbf2ba60d92125c82f0cba79fa432c8c9aaf6e.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 254, + 388, + 313, + 443 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/df48604738fb65a7d7389664d381d7ac4cdc7c8a18a0d60a735423725ee47550.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 449, + 303, + 500 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/798007dac4cff8fbbb561379e256148586715340dfc6f87669e84557b791aa4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 258, + 503, + 300, + 551 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/591f7aa234c38dffcd2fbb7179ba047e740ecbd855bd15b4d4539f817afe9bb8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 561, + 315, + 606 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/76c809462deddf1df94da004a029e25690492d60bf777d8ecb1fb570ed58d38b.jpg", + "image_caption": [ + "Fig. 1: CRM generates high-fidelity textured mesh from single image in 10 seconds." + ], + "image_footnote": [], + "bbox": [ + 253, + 613, + 308, + 661 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e8682e31401391693f271d493b90e151c2237e0622046c56b19359da91421dfe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 390, + 413, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f699a043da85542a61278aac2dea1ac2c32e84720e953c62cf419b5ca66ab373.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 446, + 411, + 501 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f2f2a3ebc059277458dcee7aa5e010e10c9d74ad98bfed40b9f9ec17d4f422ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 503, + 413, + 554 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f6ddc99930a0d96d2ecd56c02476009fbbc226d08f5a4aabb8b7a3ed3c5eef83.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 558, + 413, + 613 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e00f04b10c425cc88c3872edf2f0db16ce42e17e07d14d23c543be825541dd16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 614, + 405, + 664 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6d6cddfd4e2a9d79488994b74a58427941189f98e50d01485658b8c0e722c5b1.jpg", + "image_caption": [ + "Generated Textured Mesh" + ], + "image_footnote": [], + "bbox": [ + 426, + 390, + 483, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/64cdf860448bffcc6f38c258baa22100d0348a80fcc1977095c0602734c3ee36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 446, + 483, + 501 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bb52f3d47de4c2c2908ab641a8a694257ed1204a2c85fb4431cb7cd3d886c947.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 503, + 483, + 554 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ca39cd3d253d95c4592937f21037fce576f4e2eda993f7891737929ee6a20172.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 558, + 483, + 613 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c4c87c5e1ef94b66d4389518a8081d10be6e3505cee42f7a5ddb143d9aba1697.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 614, + 475, + 664 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5e02daf5ff5a6f19fd1a7212f8dd098f1209e7268dd7adf6f6ce286974153efd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 490, + 390, + 627, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6d45e051e2e78d6762db3e8fd7c444d362a5a86a70207ee6a2779832ae41ca3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 490, + 444, + 627, + 501 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4b730102c951ebaf3517452d75773288fce30dc434ee92d0d9b0babf14b0e8ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 503, + 627, + 554 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aab8a59f0ea80db81f3403de012c53c6c843aa70ef619e39adde9423a596e774.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 558, + 627, + 613 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d837c50dcf46eace8551e3fea5f9a8e965ee41e63095ecd4c70cd71680ce0323.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 614, + 627, + 662 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4ebd870be2e7710cf158b9ee04d44e8ec6c603917c26408e5611c098ad516dbd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 390, + 699, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8e6a6e2902f982c95b7bcea635b0d732db15a2ff96451c4a760714b68325a3e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 445, + 699, + 501 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/50b9c42502a45048091168e9e8c0b68fb4b9a9bcf6310053e702cc2b3776b8ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 503, + 692, + 554 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/baa1ba49c0e39769ee921893dc80b48a3dadf28721b7b3d0fbfbdf0932c52a00.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 558, + 697, + 613 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/76802b0a9327190810c277f429bbc10a3a0a6fa9504d70a174e830cba6538c46.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 619, + 687, + 665 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/76dc3a4df519da49d09aad4c668e6491fa94e96fcef91587514f86d7e0f46875.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 390, + 766, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c3ca2a83f79ee6844dd382a93e472905415ec64b976460977b6943258594e8e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 445, + 767, + 501 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e3bd3ffc3ddef8ec997dc4332d5d7c40c9c9c9873e34d126198e327b919277dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 506, + 761, + 555 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fb32a0150f49ab7c1a3d53b1c5c89c8da46aa86fdbc43d80bceaaa985afff56b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 558, + 767, + 613 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/416d4c6de8015f60630874f0a601921f3a3cf048ff8f50413205256586da69fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 619, + 758, + 665 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Feed-forward 3D generative models like the Large Reconstruction Model (LRM) [18] have demonstrated exceptional generation speed. However, the transformer-based methods do not leverage the geometric priors of the triplane component in their architecture, often leading to sub-optimal quality given the limited size of 3D data and", + "bbox": [ + 259, + 746, + 740, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* The Corresponding author.", + "bbox": [ + 218, + 825, + 415, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "slow training. In this work, we present the Convolutional Reconstruction Model (CRM), a high-fidelity feed-forward single image-to-3D generative model. Recognizing the limitations posed by sparse 3D data, we highlight the necessity of integrating geometric priors into network design. CRM builds on the key observation that the visualization of triplane exhibits spatial correspondence of six orthographic images. First, it generates six orthographic view images from a single input image, then feeds these images into a convolutional U-Net, leveraging its strong pixel-level alignment capabilities and significant bandwidth to create a high-resolution triplane. CRM further employs Flexicubes as geometric representation, facilitating direct end-to-end optimization on textured meshes. Overall, our model delivers a high-fidelity textured mesh from an image in just 10 seconds, without any test-time optimization.", + "bbox": [ + 259, + 147, + 743, + 328 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keywords: 3D Generation $\\cdot$ Textured Mesh $\\cdot$ Diffusion Models", + "bbox": [ + 261, + 340, + 691, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 383, + 375, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, generative models have witnessed significant advancements, largely attributed to the fast growth in data size. Transformers [56], in particular, have achieved high-performance results across various domains including language [3], image [1,38] and video generation [2]. However, the domain of 3D generation presents unique challenges. Unlike the abundance of other modal's data, 3D data is comparatively scarce. The creation of 3D data requires specialized expertise and considerable time, leading to a situation where the largest 3D datasets, namely Objaverse [12,13], only contain millions of 3D content, much smaller than image datasets like Laion [46] which contains 5 billion images.", + "bbox": [ + 212, + 416, + 785, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite this, recent developments have introduced some transformer-based methods [18, 21, 60, 64, 71] like LRM [18] for creating 3D content from single or multi-view images in a feed-forward manner. Among these models, the triplane has emerged as a popular component due to its efficiency in generating high-resolution 3D results with minimal memory cost. However, reliance on transformer-based networks for generating triplane patches has not utilized the geometric priors inherent to the triplane concept, leading to sub-optimal results in terms of quality and fidelity, and long training time.", + "bbox": [ + 212, + 551, + 785, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above challenges, this paper presents a new Convolutional Reconstruction Model (CRM) with high generation quality as well as fast training. Given the limited amount of 3D contents, CRM builds on a key hypothesis that it is beneficial to explore geometric priors in architecture design. Namely, we observe from the visualization of triplane [5,6,48] that triplane exhibits spatial correspondence of input six orthographic images, as shown in Fig. 2. The silhouette and texture of the input images have a natural alignment with the triplane structure. This motivates us to (1) use six orthographic images as input images to reconstruct 3D contents, which align well as the triplane feature, instead of other arbitrarily chosen poses; (2) use a U-Net convolutional network to map the input images to a rolled-out triplane by exploring the strong pixel-level alignment", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "between the input and output. Furthermore, the significant bandwidth capacity of our U-Net enables a direct transformation of the six orthographic images into the triplane, yielding highly detailed outcomes. Besides, we also add Canonical Coordinate Map (CCM) to the reconstruction network, a novel addition that enriches the model's understanding of spatial relations and geometry.", + "bbox": [ + 212, + 146, + 784, + 222 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the task of 3D generation from a single image, as the six orthographic images and CCMs are not directly available, we train a multi-view diffusion model conditioned on the input image to generate the six orthographic images and another diffusion model to generate the CCMs conditioned on the generated six orthographic images. Both diffusion models are trained on a filtered version of the Objaverse dataset [13]. To further enhance quality and robustness, we implement training improvements for the multi-view diffusion models, including Zero-SNR [26], random resizing, and contour augmentation.", + "bbox": [ + 212, + 222, + 785, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, as directly optimizing high-quality textured meshes is challenging, we adopt Flexicubes [47] as the geometry representation to falcitate gradient-based mesh optimization. This is unlike previous works [18, 71] that use alternative 3D representation like NeRF [36] or Gaussian Splatting [19]. Such methods often involve extra procedure steps to obtain textured meshes [52], although they can produce detailed visualizations. With our designs, we are able to train CRM with textured mesh as the final output in an end-to-end manner, and our approach has a more straightforward inference pipeline and better mesh quality. Overall, our method can generate high-fidelity textured mesh within 10 seconds, as shown in Fig. 1.", + "bbox": [ + 212, + 343, + 787, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 515, + 395, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Score Distillation for 3D Generation", + "text_level": 1, + "bbox": [ + 214, + 545, + 560, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DreamFusion [39] proposes a technique called Score Distillation Sampling (SDS) (also known as Score Jacobian Chaining [57]). It utilizes large scale image diffusion models [44,45] to iteratively refine 3D models to align with specific prompts or images. Thus it can generate 3D content without training on 3D dataset. Along this line, ProlificDreamer [63] proposes Variational Score Distillation (VSD), a principled variational framework which greatly mitigates the over-saturation problems in SDS and improves the diversity. Zero123 [30], MVDream [49], ImageDream [59] and many others [22,41,42] further improve the results and mitigate the multi-face problems using diffusion models fine-tuned on 3D data. [33,40] explore amortized score distillation. Many other works [7-9,11,20,23-25,27,51,53,55,62,65,70] improve the results a lot, in either speed or quality. However, methods based on score distillation usually take from minutes to hours to generate single object, which is computationally expensive.", + "bbox": [ + 212, + 568, + 787, + 765 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 3D Generation with Sparse View Reconstruction", + "text_level": 1, + "bbox": [ + 214, + 786, + 665, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several approaches aim to generate multi-view consistent images and then create 3D contents using sparse views reconstruction. For example, SyncDreamer [31]", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Convolutional Reconstruction Model", + "bbox": [ + 485, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/56c9a154283ec62e3d80cc0acf3f444ab7884443cddd051d3f7966e931aa5396.jpg", + "image_caption": [ + "Fig. 2: One of our key motivation is that the triplane shares a strong spatial alignment with the input six orthographic images. (a) The six orthographic images of a input shape. (b) The six orthographic CCMs. (c) The triplane (mean value of all channels) output by our U-Net, which spatially aligns with the input images. (d) The textured mesh output by our convolutional reconstruction model." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 784, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "generates multi-view consistent images and then uses NeuS [58] for reconstruction. Wonder3D [32] improves the results with cross-domain diffusion. Direct2.5 [34] improves the results with 2.5D diffusion. However, one common issue of these methods is that they need test-time optimization for reconstruction with sparse views, which may lead to extra computing and compromise the final quality", + "bbox": [ + 212, + 479, + 802, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Feed-forward 3D Generative Models", + "text_level": 1, + "bbox": [ + 215, + 598, + 563, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Some works try to generate 3D objects using a feed forward model [4, 10, 15, 16, 66, 69]. Feed-forward methods demonstrate significantly faster generation speeds compared to the two types of methods mentioned above. Recently there are some works trained on larger 3D dataset Objaverse [13]. One-2-3-45 [29] generates multi-view images and then feed the images into a network to get the 3D object. LRM series [18, 21, 60, 64] improve the quality of generated results with a transformer-based architecture. TGS [71] and LGM [52] use Gaussian Splitting [19] as the geometry representation. There are also many other works [28, 35, 54, 68] that improve the results with different techniques. Despite these advancements, there remains room for improvement in the network architecture or geometry representation. Our approach utilizes a network with a strategically designed architecture and an end-to-end training approach producing meshes directly as the final output.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/940ce2ff2c999b5c6d44e215b2a5fa46a46dd256259cc3a5012d957982f2a0b2.jpg", + "image_caption": [ + "Fig. 3: Overall pipeline of our method. The input image is fed into a multi-view image diffusion model to generate six orthographic images. Then another diffusion model is used to generate the CCMs conditioned on the six images. The six images along with the CCMs are send into CRM to reconstruct the final textured mesh. The whole inference process takes around 10 seconds on an A800 GPU. *The 4 seconds include the U-Net forward (less than 0.1s), querying surface points for UV texture and file I/O." + ], + "image_footnote": [], + "bbox": [ + 222, + 145, + 782, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 385, + 330, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we illustrate the detailed design of our method (shown in Fig. 3). Given a single input image, our model first utilizes multi-view diffusion models (Sec. 3.1) to generate six orthographic images and the canonical coordinates maps (CCMs). Then we develop our convolutional reconstruction model (CRM, Sec. 3.2) to reconstruct 3D textured mesh from the images and CCMs.", + "bbox": [ + 212, + 421, + 787, + 497 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Multi-view Diffusion Model", + "text_level": 1, + "bbox": [ + 214, + 523, + 491, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first explain the design of the multi-view diffusion model to generate six orthographic view images from a single input image. Instead of training from scratch, which is typically extremely expensive, we initialize the diffusion models using the checkpoint of ImageDream [59], a high-performance diffusion model for the task of multi-view images generation from a single image. The original ImageDream supports 4 views generation. We expand it to include 6 views by adding two more perspectives (up and down). We use another diffusion model which is conditioned on the generated six views to generate the canonical coordinate map. The conditional RGB image is concatenated with the noisy canonical coordinate map. It is also initialized from ImageDream checkpoint. Both the diffusion models are fine-tuned on the Objaverse [13] dataset.", + "bbox": [ + 212, + 551, + 787, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further improve the quality and robustness of our results, we introduce several enhancements: (1) Zero-SNR Training. We use the zero-SNR trick as mentioned in [26]. This can alleviate the problem resulting from the discrepancy between the initial Gaussian noise during sampling and the noisiest training sample. (2) Random Resizing. A naive implementation would make the model tends to generate objects that occupy the entire image. To mitigate this, we randomly resize the objects when training. (3) Contour Augmentation. We find that the model tends to predict the backview color largely relying on the", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Convolutional Reconstruction Model", + "bbox": [ + 485, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "contour of the input view. To make the model insensitive to the contour, we randomly change the contour color during training.", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Convolutional Reconstruction Model", + "text_level": 1, + "bbox": [ + 215, + 199, + 565, + 214 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We now move to introduce the detailed architecture of the convolutional reconstruction model (CRM). As outlined in Fig. 4, given the input six images and CCMs, a convolutional U-Net is used to map the input images along with the CCMs to a rolled-out triplane. Then the rolled-out triplane is reshaped into the triplane. Small multi-layer perceptions (MLPs) are used to decode the triplane features into SDF values, texture color and Flexicubes parameters. Lastly, these values are used to get texture mesh by dual marching cubes. Below, we explain the key components of CRM in detail.", + "bbox": [ + 215, + 224, + 785, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Triplane Representation We choose triplane as the 3D representation, because it can achieve high resolution 3D results with 2D computation consumption. It projects each query grid cell to axis-aligned orthogonal planes $(xy, xz,$ and $yz$ planes) and then aggregates the feature from each planes. Then the feature is decoded by 3 tiny MLPs with 2 hidden layers to get the SDF values along with deformation, color and Flexicubes weights, respectively. Further, to avoid the unnecessary entanglements of different planes, we use rolled-out triplane [61].", + "bbox": [ + 215, + 369, + 785, + 476 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Canonical Coordinates Map (CCM) We also add CCM as input [22], which contains extra geometry information. This is different from previous works, which typically use pure RGB images as the input to predict the 3D object [18]. Using pure RGB images make it extremely hard to predict the correct geometry, and sometime the geometry degrades (details in Sec. 4.3). Formally, CCM is the coordinates of each point in canonical space. It contains 3 channels whose values are within $[0,1]$ , representing the coordinates in the canonical space.", + "bbox": [ + 215, + 498, + 785, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "UNet-based Convolutional Network Our key insight is that the triplane is spatially aligned with the input six orthographic images and CCMs, as shown in Fig. 2. To match the rolled-out triplane, the six images and CCMs are arranged in a similar way. We render the six images and CCMs at a resolution of $256 \\times 256$ . They are split into two groups, with each group holding three images. These images in the four groups are then combined to create four larger images, each with a resolution of $256 \\times 768$ , allowing for spatial alignment. By concatenating these four groups, we form a 12-channel input. Next, a convolutional U-Net processes this input to produce the output triplane.", + "bbox": [ + 215, + 627, + 785, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared to transformer-based methods [18, 21, 64, 71], our U-shape design has a larger bandwidth in preserving the input information, leading to highly detailed triplane features and finally elaborate textured meshes. Moreover, the convolutional network fully utilizes the geometry prior of the spatial correspondence of triplanes and input six orthographic images, which greatly fasten the", + "bbox": [ + 215, + 765, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d4214eb88ff384be362cfc0c4e72a8822208ce107a45485ce0c668c79f6ad6b9.jpg", + "image_caption": [ + "Fig. 4: Architecture along with training pipeline of CRM. We render the 3D mesh into six orthographic images and CCMs. Then the images and CCMs are concatenated and fed into the U-Net. The output triplane is decoded by small MLP networks to form the feature grid of Flexicubes, then textured mesh is get by dual marching cubes. During training, we render the color images, depth maps and masks from GT mesh and reconstructed mesh for supervision." + ], + "image_footnote": [], + "bbox": [ + 223, + 146, + 764, + 363 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "convergence and stabilize the training. Our model can get reasonable reconstruction results at very early stage of training (around 20 minutes of training from scratch). Also, our model can be trained with a much smaller batch size 32 (compared to transformer-based LRM that uses a batch size of 1024), which makes that all of our experiments can be conducted on an 8-GPU-cards machine. The overall training cost of our reconstruction model is only $1/8$ than LRM. More details are shown in the experiments (see Sec. 4.1).", + "bbox": [ + 212, + 487, + 787, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Flexicubes Geometry Previous generic 3D generation methods mostly adopt NeRF [36] or Gaussian splatting [19] as the geometry representation, which relies on extra procedures like Marching Cubes (MC) to extract the iso-surface, suffering from topological ambiguities and struggling to represent high-fidelity geometric details. In this work, we use Flexicubes [47] as our geometry representation. It can get meshes from the features on the grid by dual marching cubes [37] during training. The features include SDF values, deformation and weights. The texture is obtained by querying the color at the surface. Flexicubes enables us to train our reconstruction model with textured mesh as the final output in an end-to-end manner.", + "bbox": [ + 212, + 618, + 787, + 770 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Loss Function Finally, to train our CRM model, we use a combination of MSE loss $\\mathcal{L}_{\\mathrm{MSE}}$ and LPIPS loss [67] $\\mathcal{L}_{\\mathrm{LPIPS}}$ for the rendered images for texture, similar as LRM [18]. To further enhance the geometry, we also include the depth map", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Convolutional Reconstruction Model", + "bbox": [ + 485, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/800904ce077e7454446324a5ce216fd08d3262868c689bdda543685ba5769ce6.jpg", + "image_caption": [ + "Fig. 5: Qualitative comparison with baselines. Our models generate high-fidelity results with better geometry and texture." + ], + "image_footnote": [], + "bbox": [ + 225, + 146, + 784, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and mask for supervision [16]. The overall loss function is", + "bbox": [ + 212, + 573, + 632, + 589 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {\\mathrm {M S E}} (\\boldsymbol {x}, \\boldsymbol {x} ^ {\\mathrm {G T}}) + \\lambda_ {\\mathrm {L P I P S}} \\mathcal {L} _ {\\mathrm {L P I P S}} (\\boldsymbol {x}, \\boldsymbol {x} ^ {\\mathrm {G T}}) \\\\ + \\lambda_ {\\text {d e p t h}} \\mathcal {L} _ {\\text {M S E}} \\left(\\boldsymbol {x} _ {\\text {d e p t h}}, \\boldsymbol {x} _ {\\text {d e p t h}} ^ {\\mathrm {G T}}\\right) + \\lambda_ {\\text {m a s k}} \\mathcal {L} _ {\\text {M S E}} \\left(\\boldsymbol {x} _ {\\text {m a s k}}, \\boldsymbol {x} _ {\\text {m a s k}} ^ {\\mathrm {G T}}\\right) + \\lambda_ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}}, \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 267, + 598, + 785, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\pmb{x}$ , $\\pmb{x}_{\\mathrm{depth}}$ and $\\pmb{x}_{\\mathrm{sil}}$ represent the RGB image, depth map and mask rendered from the reconstruction textured mesh, respectively. And $\\pmb{x}^{\\mathrm{GT}}$ , $\\pmb{x}_{\\mathrm{depth}}^{\\mathrm{GT}}$ and $\\pmb{x}_{\\mathrm{mask}}^{\\mathrm{GT}}$ are rendered from the ground truth textured mesh. $\\mathcal{L}_{\\mathrm{reg}}$ is the mesh-quality regularizers introduced in Flexicubes [47]. $\\lambda_{\\mathrm{LPIPS}}$ , $\\lambda_{\\mathrm{depth}}$ , $\\lambda_{\\mathrm{mask}}$ and $\\lambda_{\\mathrm{reg}}$ are the coefficients that balance each loss.", + "bbox": [ + 212, + 643, + 787, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 739, + 377, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Experimental Setting", + "text_level": 1, + "bbox": [ + 214, + 770, + 441, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dataset We filter the Objaverse [13] dataset, removing scene-level objects and low quality meshes, and get around 376k valid high quality objects as the training set. We reuse the rendered images from SyncDreamer [31] which contain", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "16 images per shape at the resolution of $256 \\times 256$ , and additionally render 6 orthographic images and CCM with the same lighting and resolution.", + "bbox": [ + 214, + 146, + 785, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Network Architecture The reconstruction model contains around 300M parameters. The U-Net contains [64, 128, 128, 256, 256, 512, 512] channels, with attention blocks at resolution [32, 16, 8]. We set the Flexicubes grid size as 80.", + "bbox": [ + 214, + 194, + 785, + 241 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details The reconstruction model was trained on 8 NVIDIA A800 80GB GPU cards for 6 days with $110k$ iterations. The model was trained with batch size 32 (32 shapes per iteration). At each iteration, we randomly sampled 8 views among the total 16 images for each shape for supervision. We used the Adam optimizer with learning rate $1e - 4$ . The coefficients that balancing each loss were set as $\\lambda_{\\mathrm{LPIPS}} = 0.1$ , $\\lambda_{\\mathrm{depth}} = 0.5$ , $\\lambda_{\\mathrm{mask}} = 0.5$ and $\\lambda_{\\mathrm{reg}} = 0.005$ . To enhance robustness against minor inconsistency in the generated multi-view images, we introduced small Gaussian noise to the inputs in both training and inference.", + "bbox": [ + 212, + 257, + 787, + 391 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The diffusion models for both six orthographic images and CCMs were trained on 8 NVIDIA A800 80GB GPU cards for 2 days with $10k$ iterations. The gradient accumulation is set as 12 steps, yielding a total batch size of 1536. We used the Adam optimizer with learning rate $5e - 5$ . During sampling both diffusions were sampled with 50 steps using DDIM [50].", + "bbox": [ + 212, + 393, + 787, + 469 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Comparison with baselines", + "text_level": 1, + "bbox": [ + 214, + 487, + 482, + 501 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Qualitative Results To validate the effectiveness of our method, we qualitatively compare our results with previous works including Wonder3d [32], SyncDreamer [31], Magic123 [41], One-2-3-45 [29] and OpenLRM [17]. Since LRM [18] is not open-sourced, we use OpenLRM [17], an open-sourced implementation of LRM for comparisons. For the other baselines, we use their official codes and checkpoints. As for the input images for testing, we choose two from GSO [14] dataset, one downloaded from web and one generated by text-to-image diffusion model. The results are shown in Fig. 5. It can be seen from the figure that our method generates 3D textured meshes with better texture and geometry than all other baselines. This is because our reconstruction model fully utilizes the spatial alignment of input six orthographic images and output triplane. Also, our model can generate with only 10 seconds, much faster than most of the baselines. Our method is trained in an end-to-end manner with textured mesh as final output, thus avoiding a time-consuming post-processing for converting to mesh as in [52].", + "bbox": [ + 212, + 507, + 787, + 734 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Additionally, we visualize the generated meshes comparing to the previous work LRM [18] and a concurrent work LGM [52]. Since LRM is not open-sourced, we use the meshes from their project page. The results are shown in Fig. 6. It can be seen from the figure that our results have better texture. Our method also has smoother geometry than LRM and better geometry details than LGM.", + "bbox": [ + 212, + 734, + 785, + 809 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In Fig. 7, we show more results of the high-fidelity textured meshes generated from single image by our method.", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Convolutional Reconstruction Model", + "bbox": [ + 485, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/89ea23725eeb3f89f3193c423cb2e43247383cd53dbdc7321f2e955a7ef6389f.jpg", + "image_caption": [ + "Fig. 6: Qualitative comparison with LRM [18] and LGM [52]. Our models generate high-fidelity results with detailed texture and smooth geometry." + ], + "image_footnote": [], + "bbox": [ + 225, + 143, + 785, + 458 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Quantitative Results In line with previous studies [32], we evaluate our method using the Google Scanned Objects (GSO) dataset [14] which is not included in our training dataset. We randomly choose 30 shapes and render a single image with size of $256 \\times 256$ as input for evaluation. To ensure the generated mesh accurately aligned with the ground truth mesh, we carefully adjust their pose and scale them to fit within the $[-0.5, 0.5]$ box. For mesh geometry evaluation, we report Chamfer Distance (CD), Column IoU and F-Score (with a threshold of 0.05, following One-2-3-45 [29]), which measure the geometry similarity between the reconstructed mesh and ground truth mesh. The results are shown in Table 1. It can be seen from the table that our method outperforms all of the baselines, which demonstrates the effectiveness of our method for geometry quality.", + "bbox": [ + 212, + 544, + 787, + 727 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Furthermore, for evaluating mesh texture, we render 24 images at $512 \\times 512$ resolution at elevation angles of 0, 15, and 30 degrees, for the generated meshes and ground-truth meshes respectively. For each elevation, the 8 images are evenly distributed around a full 360-degree rotation. Then we assessed them using several metrics: PSNR, SSIM, LPIPS and Clip-Similarity, which measure the resemblance in appearance between the reconstructed mesh and the original ground truth mesh. The results are shown in Table 2. It shows that the generated", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/71497811b6dc9337e1c15445273e7a297d4a88dfe94ddc81167a9d4c8dea33ff.jpg", + "image_caption": [ + "Fig. 7: More Results of the generated results of our method from a single image." + ], + "image_footnote": [], + "bbox": [ + 228, + 146, + 777, + 513 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "textured meshes by our model surpass those of all baselines in appearance, which demonstrates the effectiveness of our method for texture quality.", + "bbox": [ + 212, + 573, + 784, + 603 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Additionally, we carry out experiments to evaluate the effectiveness of our single image to multi-view diffusion models. We use PSNR, SSIM, and LPIPS to measure the similarity between the generated multi-view images and ground truth multi-view images. For this analysis, we use four views of the generated images (left, right, front, back) from each model under comparison, setting the background color to grey (value 128). We compare with SyncDreamer [31] and Wonder3D [32]. The outcomes of this evaluation are documented in Table 3. It can be seen from the table that our method outperforms all the baselines.", + "bbox": [ + 212, + 604, + 787, + 726 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Ablation Study and Analysis", + "text_level": 1, + "bbox": [ + 215, + 750, + 501, + 766 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Reconstruction Results on Early Training Stage An advantage of our CRM is that it is easy to train. In fact, we find that CRM starts to show reasonable results at very early stage of training. The results are shown in Fig. 8. The results are good even with barely 280 iterations (only 20 minutes of training).", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Convolutional Reconstruction Model", + "bbox": [ + 485, + 114, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/09ec059f55aa499e6ba78b4ae4c820a84eed1eebdafd23babce5f8e0c2aa9f91.jpg", + "table_caption": [ + "Table 1: Quantitative comparison for the geometry quality between our method and baselines for single image to 3D textured mesh generation. We report the metrics of Chamfer Distance, Column IoU and F-score on GSO dataset." + ], + "table_footnote": [], + "table_body": "
MethodChamfer Dist.↓ Vol. IoU↑ F-Sco. (%)↑
One-2-3-45 [29]0.01720.446372.19
SyncDreamer [31]0.01400.390075.74
Wonder3D [32]0.01860.439876.75
Magic123 [41]0.01880.371460.66
TGS [71]0.01720.298265.17
OpenLRM [17,18]0.01680.377463.22
LGM [52]0.01170.468568.69
Ours0.00940.613179.38
", + "bbox": [ + 310, + 198, + 692, + 340 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/a45b13e4cf8c34edd0f32f95ef862c186acf35988628e22f5f6949b10bab31bf.jpg", + "table_caption": [ + "Table 2: Quantitative comparison for the texture quality between our method and baselines for single image to 3D textured mesh generation. We report the metric of PSNR, SSIM, LPIPS and Clip [43]-Similarity on GSO dataset." + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑LPIPS↓Clip-Sim↑
One-2-3-45 [29]13.930.80840.262579.83
SyncDreamer [31]14.000.81650.259182.76
Wonder3D [32]13.310.81210.255483.70
Magic123 [41]12.690.79840.244285.16
OpenLRM [17,18]14.300.82940.227684.20
LGM [71]13.280.79460.256085.20
Ours16.220.83810.214387.55
", + "bbox": [ + 316, + 407, + 686, + 536 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conjecture that the fast convergence results from the strong geometry prior in our architecture design.", + "bbox": [ + 215, + 565, + 785, + 595 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Training Time of CRM In Fig. 9 we compare the training cost between our method (reconstruction model only) and two baselines, LRM [18] and LGM [52]. We measure the training cost by the training days multiplying the amount of used NVIDIA A100/A800 GPU cards. It can be seen that our model takes much smaller training time than the two baselines. This is because our model utilizes the spatial correspondence between input six orthographic images/CCMs and triplanes, which serves as a strong prior that makes the training easier.", + "bbox": [ + 215, + 618, + 787, + 724 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Importance of Input CCM We examine the importance of the CCMs that are concatenated to the input images. To compare, we train a reconstruction model that takes only the six RGB images as input, without CCMs. The results are shown in Fig. 10. It can be seen that the results of the geometry degrade a lot without CCM input. This is because CCM provides important geometry information for the model, especially when the geometry is complex.", + "bbox": [ + 215, + 747, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/897d0152144cd4c7d8ec18e1ab84dc2a38e4f1db9777fde1b50a5ecff542d82b.jpg", + "table_caption": [ + "Table 3: Quantitative comparison between our method and baselines for novel view synthesis of the multi-view diffusion model. We report the metric of PSNR, SSIM and LPIPS on GSO dataset." + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑LPIPS↓
SyncDreamer [31]20.300.78040.2932
Wonder3D [32]23.760.81270.2210
Ours29.360.87210.1354
", + "bbox": [ + 354, + 198, + 645, + 271 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5631a7b4a9a4de8da88578f23a96c9c2a6081801c8c582a822d054c81189a900.jpg", + "image_caption": [ + "(a) 0 iters (Initialization)" + ], + "image_footnote": [], + "bbox": [ + 253, + 308, + 339, + 339 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ed9850342b38e8734023597fdd40bc18b788b18c1f82275e993a9a15f9cb4b1e.jpg", + "image_caption": [ + "(b) 280 iters $(\\sim 20$ min training)" + ], + "image_footnote": [], + "bbox": [ + 354, + 292, + 468, + 345 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7f82071071640cd1ea11c97faff85820c3335d8b66ddf33e61fbc189970c6d53.jpg", + "image_caption": [ + "(c) 4k iters $(\\sim 5$ hours training)" + ], + "image_footnote": [], + "bbox": [ + 236, + 378, + 346, + 431 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b12c8e77990ff262098ce165e36edb196ebcc6dfb9dbd0747fedf1e6927fc518.jpg", + "image_caption": [ + "(d) Ground Truth Mesh" + ], + "image_footnote": [], + "bbox": [ + 380, + 378, + 436, + 430 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7a69169d99d1f4e864126c8739ec686448a7d636802dd089a704018c5ec6828a.jpg", + "image_caption": [ + "Fig. 9: Training cost comparison. Our model require much less computation cost than other baselines." + ], + "image_footnote": [], + "bbox": [ + 522, + 289, + 764, + 465 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4f84525782f831285f21ca25454032ed4c91a8f7f355c96a108f402d002df564.jpg", + "image_caption": [ + "Fig. 8: Reconstruction results on unseen samples during early stage of training." + ], + "image_footnote": [], + "bbox": [ + 235, + 544, + 429, + 614 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/42a32e64756660ebbcffd2abe9bfffc54a4d6e7639058983810447857084e3b0.jpg", + "image_caption": [ + "(a) w/o CCM Input" + ], + "image_footnote": [], + "bbox": [ + 246, + 619, + 326, + 681 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "", + "bbox": [ + 339, + 547, + 428, + 614 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/bfbfcdff8787aa365202dce2f181624248a4dc94cab01336d8fbd59207a2c55d.jpg", + "image_caption": [ + "Fig. 10: The CCM concatenated to the input images is beneficial for our model. (a) Without providing CCM, the model outputs a geometry which is reasonable, but not very good. (b) The shape reconstructed using our full model with CCM input, with a much better geometry. (c) Ground truth mesh rendered from the same pose." + ], + "image_footnote": [], + "bbox": [ + 339, + 619, + 416, + 681 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/08abccfcb2adfce265eb95dcc235f806e2f74626c4f17ae2c5acb987bb2a075e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 547, + 635, + 614 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2a9ed6b6c784b4c32e4a2fd3d7b4f9eea934e2d8ff5233f08826588c172e1532.jpg", + "image_caption": [ + "(b) w/ CCM Input" + ], + "image_footnote": [], + "bbox": [ + 454, + 619, + 531, + 681 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/fe5d1ed6f6d9bd3b09d62e90f1e6e4557e0b3c26f9deb2f7aae3afece14607ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 544, + 743, + 613 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/23679090a5b355d39a3aa067a1be0e679d45e297ab83cc1c66944c0d93f6e6a7.jpg", + "image_caption": [ + "(c) Ground Truth Mesh" + ], + "image_footnote": [], + "bbox": [ + 658, + 619, + 736, + 681 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Design of Multi-view Diffusion Here we examine the effectiveness of the design of the multi-view diffusion models. Starting from the baseline that naively", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Convolutional Reconstruction Model", + "bbox": [ + 485, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b4a26437682d78009383ccf04b905936ffe9cef3d3b2f1bf50f99db56518666c.jpg", + "table_caption": [ + "Table 4: Ablation study on the design of multi-view diffusion on novel view synthesis. We report the metrics of PSNR, SSIM and LPIPS on GSO dataset." + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑ SSIM↑ LPIPS↓
ImageDream (6 view)28.990.85650.1497
+ Zero-SNR29.130.85980.1498
+ Random Resizing29.360.87210.1354
+ Contour Augmentation28.920.86810.1444
", + "bbox": [ + 328, + 184, + 671, + 271 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3cbaaf3456ced12b33b53d51aabdd26a26ba0547ad725e472d55ed3ad830fd55.jpg", + "image_caption": [ + "(a) Input Image" + ], + "image_footnote": [], + "bbox": [ + 222, + 286, + 310, + 354 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/be846a4b9118f82ecddaabd6ae6cf3b61073f55ee599ada869aca97bc4ee4fad.jpg", + "image_caption": [ + "(b) Segmentation Result" + ], + "image_footnote": [], + "bbox": [ + 339, + 286, + 475, + 354 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/afaa77feb293f97da71e6e6aabfa1855817c73d8bff5b59a60282e105e9906d3.jpg", + "image_caption": [ + "Fig. 11: Demonstration of contour augmentation. (a) Given an input image, (b) off-the-shelf segmentation model sometimes provides imperfect results. (c) Without contour augmentation, the predicted backview color is sensitive to the contour. (d) With contour augmentation, the model predicts reasonable result. (e) We demonstrate how we augment the input image during training." + ], + "image_footnote": [], + "bbox": [ + 519, + 287, + 573, + 352 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0324a6d9884bb8b93e9653c54a7aefa89ac3b7fd2efcde330926b09383942e44.jpg", + "image_caption": [ + "(c) w/o Contour Augmentation" + ], + "image_footnote": [], + "bbox": [ + 612, + 287, + 665, + 352 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/cff429cc48c431105176294e822c9176b6af397b6db115d0cd3455c6fd8de114.jpg", + "image_caption": [ + "(d) w/ Contour Augmentation", + "(e) Contour Augmented Image" + ], + "image_footnote": [], + "bbox": [ + 712, + 291, + 759, + 347 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "fine-tunes the pre-trained ImageDream model with 2 additional views, we sequentially add the proposed techniques on the training. We examine the results on a subset of GSO, comparing the similarity of the generated novel view images with the ground truth images using PSNR, SSIM and LPIPS metrics. The results are shown in table 4. It can be seen that both the Zero-SNR trick and random resizing are beneficial. Note that the contour augmentation does not improve the quantitative metrics. However, we find that this trick makes the model more robust to in the wild input images (Fig. 11).", + "bbox": [ + 212, + 500, + 787, + 623 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 650, + 359, + 666 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we present a convolutional reconstruction model (CRM) for creating high-quality 3D models from a single image. Our approach effectively utilizes the spatial relationship between input images and the output triplane, leading to improved textured meshes, with significantly less training cost compared to previous transformer-based methods [18]. The model operates on an end-to-end training basis, directly outputting textured meshes. Overall, our method can produce detailed textured meshes in just 10 seconds.", + "bbox": [ + 212, + 686, + 787, + 792 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Potential Negative Impact. Similar to many other generated models, our CRM may be used to generate malicious or fake 3D contents, which may need additional caution.", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by Natural Science Foundation of China (No.s 62350080, 62076147), Tsinghua Institute for Guo Qiang, and the High Performance Computing Center, Tsinghua University. J.Z is also supported by the New Cornerstone Science Foundation through the XPlorer Prize.", + "bbox": [ + 212, + 175, + 789, + 236 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 258, + 321, + 273 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bao, F., Nie, S., Xue, K., Cao, Y., Li, C., Su, H., Zhu, J.: All are worth words: A vit backbone for diffusion models. In: CVPR (2023)", + "2. Brooks, T., Peebles, B., Homes, C., DePue, W., Guo, Y., Jing, L., Schnurr, D., Taylor, J., Luhman, T., Luhman, E., Ng, C., Wang, R., Ramesh, A.: Video generation models as world simulators (2024), https://openai.com/research/video-generation-models-as-world-simulators", + "3. Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D.M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners (2020)", + "4. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)", + "5. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision. pp. 333-350. Springer (2022)", + "6. Chen, H., Gu, J., Chen, A., Tian, W., Tu, Z., Liu, L., Su, H.: Single-stage diffusion nef: A unified approach to 3d generation and reconstruction. arXiv preprint arXiv:2304.06714 (2023)", + "7. Chen, R., Chen, Y., Jiao, N., Jia, K.: *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation*. arXiv preprint arXiv:2303.13873 (2023)", + "8. Chen, Y., Zhang, C., Yang, X., Cai, Z., Yu, G., Yang, L., Lin, G.: It3d: Improved text-to-3d generation with explicit view synthesis (2023)", + "9. Chen, Z., Wang, F., Liu, H.: Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585 (2023)", + "10. Cheng, Y.C., Lee, H.Y., Tuyakov, S., Schwing, A., Gui, L.: SDFusion: Multimodal 3d shape completion, reconstruction, and generation. In: CVPR (2023)", + "11. Decatur, D., Lang, I., Aberman, K., Hanocka, R.: 3d paintbrush: Local stylization of 3d shapes with cascaded score distillation. arXiv preprint arXiv:2311.09571 (2023)", + "12. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objaverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects. Advances in Neural Information Processing Systems 36 (2024)", + "13. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objaverse: A universe of annotated 3d objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13142-13153 (2023)" + ], + "bbox": [ + 218, + 287, + 785, + 838 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Convolutional Reconstruction Model", + "bbox": [ + 485, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Downs, L., Francis, A., Koenig, N., Kinman, B., Hickman, R., Reymann, K., McHugh, T.B., Vanhoucke, V.: Google scanned objects: A high-quality dataset of 3d scanned household items. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 2553-2560. IEEE (2022)", + "15. Gao, J., Shen, T., Wang, Z., Chen, W., Yin, K., Li, D., Litany, O., Gojcic, Z., Fidler, S.: Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems 35, 31841-31854 (2022)", + "16. Gupta, A., Xiong, W., Nie, Y., Jones, I., Oğuz, B.: 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371 (2023)", + "17. He, Z., Wang, T.: Openlrm: Open-source large reconstruction models. https://github.com/3DTopia/OpenLRM (2023)", + "18. Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D., Liu, F., Sunkavalli, K., Bui, T., Tan, H.: Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400 (2023)", + "19. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)", + "20. Kim, S., Lee, K., Choi, J.S., Jeong, J., Sohn, K., Shin, J.: Collaborative score distillation for consistent visual synthesis (2023)", + "21. Li, J., Tan, H., Zhang, K., Xu, Z., Luan, F., Xu, Y., Hong, Y., Sunkavalli, K., Shakhnarovich, G., Bi, S.: Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214 (2023)", + "22. Li, W., Chen, R., Chen, X., Tan, P.: Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596 (2023)", + "23. Li, Y., Dou, Y., Shi, Y., Lei, Y., Chen, X., Zhang, Y., Zhou, P., Ni, B.: Focaldreamer: Text-driven 3d editing via focal-fusion assembly (2023)", + "24. Liang, Y., Yang, X., Lin, J., Li, H., Xu, X., Chen, Y.: Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching (2023)", + "25. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300–309 (2023)", + "26. Lin, S., Liu, B., Li, J., Yang, X.: Common diffusion noise schedules and sample steps are flawed. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5404-5411 (2024)", + "27. Liu, F., Wu, D., Wei, Y., Rao, Y., Duan, Y.: Sherpa3d: Boosting high-fidelity text-to-3d generation via coarse 3d prior (2023)", + "28. Liu, M., Shi, R., Chen, L., Zhang, Z., Xu, C., Wei, X., Chen, H., Zeng, C., Gu, J., Su, H.: One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion. arXiv preprint arXiv:2311.07885 (2023)", + "29. Liu, M., Xu, C., Jin, H., Chen, L., Xu, Z., Su, H., et al.: One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928 (2023)", + "30. Liu, R., Wu, R., Hoorick, B.V., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object (2023)", + "31. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453 (2023)", + "32. Long, X., Guo, Y.C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.H., Habermann, M., Theobalt, C., et al.: Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008 (2023)" + ], + "bbox": [ + 217, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Lorraine, J., Xie, K., Zeng, X., Lin, C.H., Takikawa, T., Sharp, N., Lin, T.Y., Liu, M.Y., Fidler, S., Lucas, J.: Att3d: Amortized text-to-3d object synthesis (2023)", + "34. Lu, Y., Zhang, J., Li, S., Fang, T., McKinnon, D., Tsin, Y., Quan, L., Cao, X., Yao, Y.: Direct2.5: Diverse text-to-3d generation via multi-view 2.5d diffusion (2023)", + "35. Mercier, A., Nakhli, R., Reddy, M., Yasarla, R., Cai, H., Porikli, F., Berger, G.: Hexagen3d: Stablediffusion is just one step away from fast and diverse text-to-3d generation. arXiv preprint arXiv:2401.07727 (2024)", + "36. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)", + "37. Nielson, G.M.: Dual marching cubes. In: IEEE visualization 2004. pp. 489-496. IEEE (2004)", + "38. Peebles, W., Xie, S.: Scalable diffusion models with transformers (2023)", + "39. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)", + "40. Qian, G., Cao, J., Siarohin, A., Kant, Y., Wang, C., Vasilkovsky, M., Lee, H.Y., Fang, Y., Skorokhodov, I., Zhuang, P., Gilitschenski, I., Ren, J., Ghanem, B., Aberman, K., Tulyakov, S.: Atom: Amortized text-to-mesh using 2d diffusion (2024)", + "41. Qian, G., Mai, J., Hamdi, A., Ren, J., Siarohin, A., Li, B., Lee, H.Y., Skorokhodov, I., Wonka, P., Tulyakov, S., Ghanem, B.: Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors (2023)", + "42. Qiu, L., Chen, G., Gu, X., Zuo, Q., Xu, M., Wu, Y., Yuan, W., Dong, Z., Bo, L., Han, X.: Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d (2023)", + "43. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "44. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "45. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)", + "46. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: Laion-5b: An open large-scale dataset for training next generation image-text models (2022)", + "47. Shen, T., Munkberg, J., Hasselgren, J., Yin, K., Wang, Z., Chen, W., Gojcic, Z., Fidler, S., Sharp, N., Gao, J.: Flexible isosurface extraction for gradient-based mesh optimization. ACM Transactions on Graphics (TOG) 42(4), 1-16 (2023)", + "48. Shi, R., Wei, X., Wang, C., Su, H.: Zerorf: Fast sparse view 360 $\\{\\backslash \\mathrm{deg}\\}$ reconstruction with zero pretraining. arXiv preprint arXiv:2312.09249 (2023)", + "49. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023)", + "50. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)", + "51. Sun, J., Zhang, B., Shao, R., Wang, L., Liu, W., Xie, Z., Liu, Y.: Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior (2023)" + ], + "bbox": [ + 212, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Convolutional Reconstruction Model", + "bbox": [ + 485, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "52. Tang, J., Chen, Z., Chen, X., Wang, T., Zeng, G., Liu, Z.: Lgm: Large multi-view gaussian model for high-resolution 3d content creation (2024)", + "53. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)", + "54. Tochilkin, D., Pankratz, D., Liu, Z., Huang, Z., Letts, A., Li, Y., Liang, D., Laforte, C., Jampani, V., Cao, Y.P.: Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151 (2024)", + "55. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439 (2023)", + "56. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need (2023)", + "57. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation (2022)", + "58. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction (2023)", + "59. Wang, P., Shi, Y.: Imaginedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201 (2023)", + "60. Wang, P., Tan, H., Bi, S., Xu, Y., Luan, F., Sunkavalli, K., Wang, W., Xu, Z., Zhang, K.: Pf-lrm: Pose-free large reconstruction model for joint pose and shape prediction. arXiv preprint arXiv:2311.12024 (2023)", + "61. Wang, T., Zhang, B., Zhang, T., Gu, S., Bao, J., Baltrusaitis, T., Shen, J., Chen, D., Wen, F., Chen, Q., et al.: Rodin: A generative model for sculpting 3d digital avatars using diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4563-4573 (2023)", + "62. Wang, X., Wang, Y., Ye, J., Wang, Z., Sun, F., Liu, P., Wang, L., Sun, K., Wang, X., He, B.: Animatabledreamer: Text-guided non-rigid 3d model generation and reconstruction with canonical score distillation", + "63. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=ppJuFSOAnM", + "64. Xu, Y., Tan, H., Luan, F., Bi, S., Wang, P., Li, J., Shi, Z., Sunkavalli, K., Wetzstein, G., Xu, Z., et al.: Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model. arXiv preprint arXiv:2311.09217 (2023)", + "65. Yu, X., Guo, Y.C., Li, Y., Liang, D., Zhang, S.H., Qi, X.: Text-to-3d with classifier score distillation (2023)", + "66. Zeng, X., Vahdat, A., Williams, F., Gojcic, Z., Litany, O., Fidler, S., Kreis, K.: Lion: Latent point diffusion models for 3d shape generation (2022)", + "67. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric (2018)", + "68. Zheng, X.Y., Pan, H., Guo, Y.X., Tong, X., Liu, Y.: Mvd $^2$ : Efficient multiview 3d reconstruction for multiview diffusion (2024)", + "69. Zheng, X.Y., Pan, H., Wang, P.S., Tong, X., Liu, Y., Shum, H.Y.: Locally attentional sdf diffusion for controllable 3d shape generation. arXiv preprint arXiv:2305.04461 (2023)", + "70. Zhu, J., Zhuang, P.: Hifa: High-fidelity text-to-3d generation with advanced diffusion guidance (2023)", + "71. Zou, Z.X., Yu, Z., Guo, Y.C., Li, Y., Liang, D., Cao, Y.P., Zhang, S.H.: Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. arXiv preprint arXiv:2312.09147 (2023)" + ], + "bbox": [ + 215, + 147, + 784, + 832 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/CRM_ Single Image to 3D Textured Mesh with Convolutional Reconstruction Model/9e4b61b2-e1dd-4f99-9ab9-4fe8376f780a_model.json b/2024/CRM_ Single Image to 3D Textured Mesh with Convolutional Reconstruction Model/9e4b61b2-e1dd-4f99-9ab9-4fe8376f780a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c5a30fffdbe558e5dbcfa8e85b5a8915a9d3534f --- /dev/null +++ b/2024/CRM_ Single Image to 3D Textured Mesh with Convolutional Reconstruction Model/9e4b61b2-e1dd-4f99-9ab9-4fe8376f780a_model.json @@ -0,0 +1,2964 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.233, + 0.141, + 0.773, + 0.185 + ], + "angle": 0, + "content": "CRM: Single Image to 3D Textured Mesh with Convolutional Reconstruction Model" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.212, + 0.763, + 0.245 + ], + "angle": 0, + "content": "Zhengyi Wang\\(^{1,2}\\), Yikai Wang\\(^{1,2}\\), Yifei Chen\\(^{1}\\), Chendong Xiang\\(^{1,2}\\), Shuo Chen\\(^{1}\\), Dajiang Yu\\(^{1}\\), Chongxuan Li\\(^{3}\\), Hang Su\\(^{1}\\) and Jun Zhu\\(^{\\star 1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.255, + 0.774, + 0.283 + ], + "angle": 0, + "content": "\\(^{1}\\) Dept. of Comp. Sci. & Tech., BNrist Center, Tsinghua-Bosch Joint ML Center, Tsinghua University;" + }, + { + "type": "text", + "bbox": [ + 0.406, + 0.283, + 0.596, + 0.298 + ], + "angle": 0, + "content": "2 ShengShu, Beijing, China;" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.297, + 0.78, + 0.326 + ], + "angle": 0, + "content": "\\(^{3}\\) Gaoling School of Artificial Intelligence, Renmin University of China, Beijing Key Laboratory of Big Data Management and Analysis Methods, Beijing, China." + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.366, + 0.321, + 0.381 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.39, + 0.315, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.45, + 0.304, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.505, + 0.301, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.562, + 0.316, + 0.607 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.614, + 0.309, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.477, + 0.366, + 0.644, + 0.38 + ], + "angle": 0, + "content": "Generated Textured Mesh" + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.391, + 0.414, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.447, + 0.413, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.505, + 0.414, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.559, + 0.414, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.616, + 0.406, + 0.665 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.391, + 0.485, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.447, + 0.485, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.505, + 0.485, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.559, + 0.485, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.616, + 0.476, + 0.665 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.391, + 0.628, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.445, + 0.628, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.505, + 0.628, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.559, + 0.628, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.616, + 0.628, + 0.664 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.391, + 0.7, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.446, + 0.7, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.505, + 0.693, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.559, + 0.698, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.621, + 0.688, + 0.666 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.391, + 0.767, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.707, + 0.446, + 0.769, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.507, + 0.762, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.559, + 0.768, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.621, + 0.759, + 0.666 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.676, + 0.786, + 0.691 + ], + "angle": 0, + "content": "Fig. 1: CRM generates high-fidelity textured mesh from single image in 10 seconds." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.747, + 0.741, + 0.817 + ], + "angle": 0, + "content": "Abstract. Feed-forward 3D generative models like the Large Reconstruction Model (LRM) [18] have demonstrated exceptional generation speed. However, the transformer-based methods do not leverage the geometric priors of the triplane component in their architecture, often leading to sub-optimal quality given the limited size of 3D data and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.416, + 0.84 + ], + "angle": 0, + "content": "* The Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.148, + 0.744, + 0.329 + ], + "angle": 0, + "content": "slow training. In this work, we present the Convolutional Reconstruction Model (CRM), a high-fidelity feed-forward single image-to-3D generative model. Recognizing the limitations posed by sparse 3D data, we highlight the necessity of integrating geometric priors into network design. CRM builds on the key observation that the visualization of triplane exhibits spatial correspondence of six orthographic images. First, it generates six orthographic view images from a single input image, then feeds these images into a convolutional U-Net, leveraging its strong pixel-level alignment capabilities and significant bandwidth to create a high-resolution triplane. CRM further employs Flexicubes as geometric representation, facilitating direct end-to-end optimization on textured meshes. Overall, our model delivers a high-fidelity textured mesh from an image in just 10 seconds, without any test-time optimization." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.342, + 0.692, + 0.356 + ], + "angle": 0, + "content": "Keywords: 3D Generation \\(\\cdot\\) Textured Mesh \\(\\cdot\\) Diffusion Models" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.384, + 0.377, + 0.399 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.787, + 0.552 + ], + "angle": 0, + "content": "In recent years, generative models have witnessed significant advancements, largely attributed to the fast growth in data size. Transformers [56], in particular, have achieved high-performance results across various domains including language [3], image [1,38] and video generation [2]. However, the domain of 3D generation presents unique challenges. Unlike the abundance of other modal's data, 3D data is comparatively scarce. The creation of 3D data requires specialized expertise and considerable time, leading to a situation where the largest 3D datasets, namely Objaverse [12,13], only contain millions of 3D content, much smaller than image datasets like Laion [46] which contains 5 billion images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.787, + 0.674 + ], + "angle": 0, + "content": "Despite this, recent developments have introduced some transformer-based methods [18, 21, 60, 64, 71] like LRM [18] for creating 3D content from single or multi-view images in a feed-forward manner. Among these models, the triplane has emerged as a popular component due to its efficiency in generating high-resolution 3D results with minimal memory cost. However, reliance on transformer-based networks for generating triplane patches has not utilized the geometric priors inherent to the triplane concept, leading to sub-optimal results in terms of quality and fidelity, and long training time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To address the above challenges, this paper presents a new Convolutional Reconstruction Model (CRM) with high generation quality as well as fast training. Given the limited amount of 3D contents, CRM builds on a key hypothesis that it is beneficial to explore geometric priors in architecture design. Namely, we observe from the visualization of triplane [5,6,48] that triplane exhibits spatial correspondence of input six orthographic images, as shown in Fig. 2. The silhouette and texture of the input images have a natural alignment with the triplane structure. This motivates us to (1) use six orthographic images as input images to reconstruct 3D contents, which align well as the triplane feature, instead of other arbitrarily chosen poses; (2) use a U-Net convolutional network to map the input images to a rolled-out triplane by exploring the strong pixel-level alignment" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.486, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Convolutional Reconstruction Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.223 + ], + "angle": 0, + "content": "between the input and output. Furthermore, the significant bandwidth capacity of our U-Net enables a direct transformation of the six orthographic images into the triplane, yielding highly detailed outcomes. Besides, we also add Canonical Coordinate Map (CCM) to the reconstruction network, a novel addition that enriches the model's understanding of spatial relations and geometry." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.787, + 0.344 + ], + "angle": 0, + "content": "For the task of 3D generation from a single image, as the six orthographic images and CCMs are not directly available, we train a multi-view diffusion model conditioned on the input image to generate the six orthographic images and another diffusion model to generate the CCMs conditioned on the generated six orthographic images. Both diffusion models are trained on a filtered version of the Objaverse dataset [13]. To further enhance quality and robustness, we implement training improvements for the multi-view diffusion models, including Zero-SNR [26], random resizing, and contour augmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.496 + ], + "angle": 0, + "content": "Finally, as directly optimizing high-quality textured meshes is challenging, we adopt Flexicubes [47] as the geometry representation to falcitate gradient-based mesh optimization. This is unlike previous works [18, 71] that use alternative 3D representation like NeRF [36] or Gaussian Splatting [19]. Such methods often involve extra procedure steps to obtain textured meshes [52], although they can produce detailed visualizations. With our designs, we are able to train CRM with textured mesh as the final output in an end-to-end manner, and our approach has a more straightforward inference pipeline and better mesh quality. Overall, our method can generate high-fidelity textured mesh within 10 seconds, as shown in Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.516, + 0.396, + 0.531 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.546, + 0.562, + 0.56 + ], + "angle": 0, + "content": "2.1 Score Distillation for 3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.766 + ], + "angle": 0, + "content": "DreamFusion [39] proposes a technique called Score Distillation Sampling (SDS) (also known as Score Jacobian Chaining [57]). It utilizes large scale image diffusion models [44,45] to iteratively refine 3D models to align with specific prompts or images. Thus it can generate 3D content without training on 3D dataset. Along this line, ProlificDreamer [63] proposes Variational Score Distillation (VSD), a principled variational framework which greatly mitigates the over-saturation problems in SDS and improves the diversity. Zero123 [30], MVDream [49], ImageDream [59] and many others [22,41,42] further improve the results and mitigate the multi-face problems using diffusion models fine-tuned on 3D data. [33,40] explore amortized score distillation. Many other works [7-9,11,20,23-25,27,51,53,55,62,65,70] improve the results a lot, in either speed or quality. However, methods based on score distillation usually take from minutes to hours to generate single object, which is computationally expensive." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.787, + 0.666, + 0.803 + ], + "angle": 0, + "content": "2.2 3D Generation with Sparse View Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Several approaches aim to generate multi-view consistent images and then create 3D contents using sparse views reconstruction. For example, SyncDreamer [31]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.785, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.373, + 0.788, + 0.444 + ], + "angle": 0, + "content": "Fig. 2: One of our key motivation is that the triplane shares a strong spatial alignment with the input six orthographic images. (a) The six orthographic images of a input shape. (b) The six orthographic CCMs. (c) The triplane (mean value of all channels) output by our U-Net, which spatially aligns with the input images. (d) The textured mesh output by our convolutional reconstruction model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.803, + 0.56 + ], + "angle": 0, + "content": "generates multi-view consistent images and then uses NeuS [58] for reconstruction. Wonder3D [32] improves the results with cross-domain diffusion. Direct2.5 [34] improves the results with 2.5D diffusion. However, one common issue of these methods is that they need test-time optimization for reconstruction with sparse views, which may lead to extra computing and compromise the final quality" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.599, + 0.564, + 0.615 + ], + "angle": 0, + "content": "2.3 Feed-forward 3D Generative Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Some works try to generate 3D objects using a feed forward model [4, 10, 15, 16, 66, 69]. Feed-forward methods demonstrate significantly faster generation speeds compared to the two types of methods mentioned above. Recently there are some works trained on larger 3D dataset Objaverse [13]. One-2-3-45 [29] generates multi-view images and then feed the images into a network to get the 3D object. LRM series [18, 21, 60, 64] improve the quality of generated results with a transformer-based architecture. TGS [71] and LGM [52] use Gaussian Splitting [19] as the geometry representation. There are also many other works [28, 35, 54, 68] that improve the results with different techniques. Despite these advancements, there remains room for improvement in the network architecture or geometry representation. Our approach utilizes a network with a strategically designed architecture and an end-to-end training approach producing meshes directly as the final output." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.486, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Convolutional Reconstruction Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.146, + 0.784, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.357 + ], + "angle": 0, + "content": "Fig. 3: Overall pipeline of our method. The input image is fed into a multi-view image diffusion model to generate six orthographic images. Then another diffusion model is used to generate the CCMs conditioned on the six images. The six images along with the CCMs are send into CRM to reconstruct the final textured mesh. The whole inference process takes around 10 seconds on an A800 GPU. *The 4 seconds include the U-Net forward (less than 0.1s), querying surface points for UV texture and file I/O." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.386, + 0.331, + 0.401 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.788, + 0.498 + ], + "angle": 0, + "content": "In this section, we illustrate the detailed design of our method (shown in Fig. 3). Given a single input image, our model first utilizes multi-view diffusion models (Sec. 3.1) to generate six orthographic images and the canonical coordinates maps (CCMs). Then we develop our convolutional reconstruction model (CRM, Sec. 3.2) to reconstruct 3D textured mesh from the images and CCMs." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.524, + 0.492, + 0.538 + ], + "angle": 0, + "content": "3.1 Multi-view Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.719 + ], + "angle": 0, + "content": "We first explain the design of the multi-view diffusion model to generate six orthographic view images from a single input image. Instead of training from scratch, which is typically extremely expensive, we initialize the diffusion models using the checkpoint of ImageDream [59], a high-performance diffusion model for the task of multi-view images generation from a single image. The original ImageDream supports 4 views generation. We expand it to include 6 views by adding two more perspectives (up and down). We use another diffusion model which is conditioned on the generated six views to generate the canonical coordinate map. The conditional RGB image is concatenated with the noisy canonical coordinate map. It is also initialized from ImageDream checkpoint. Both the diffusion models are fine-tuned on the Objaverse [13] dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To further improve the quality and robustness of our results, we introduce several enhancements: (1) Zero-SNR Training. We use the zero-SNR trick as mentioned in [26]. This can alleviate the problem resulting from the discrepancy between the initial Gaussian noise during sampling and the noisiest training sample. (2) Random Resizing. A naive implementation would make the model tends to generate objects that occupy the entire image. To mitigate this, we randomly resize the objects when training. (3) Contour Augmentation. We find that the model tends to predict the backview color largely relying on the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "contour of the input view. To make the model insensitive to the contour, we randomly change the contour color during training." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.2, + 0.566, + 0.215 + ], + "angle": 0, + "content": "3.2 Convolutional Reconstruction Model" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.226, + 0.787, + 0.348 + ], + "angle": 0, + "content": "We now move to introduce the detailed architecture of the convolutional reconstruction model (CRM). As outlined in Fig. 4, given the input six images and CCMs, a convolutional U-Net is used to map the input images along with the CCMs to a rolled-out triplane. Then the rolled-out triplane is reshaped into the triplane. Small multi-layer perceptions (MLPs) are used to decode the triplane features into SDF values, texture color and Flexicubes parameters. Lastly, these values are used to get texture mesh by dual marching cubes. Below, we explain the key components of CRM in detail." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.477 + ], + "angle": 0, + "content": "Triplane Representation We choose triplane as the 3D representation, because it can achieve high resolution 3D results with 2D computation consumption. It projects each query grid cell to axis-aligned orthogonal planes \\((xy, xz,\\) and \\(yz\\) planes) and then aggregates the feature from each planes. Then the feature is decoded by 3 tiny MLPs with 2 hidden layers to get the SDF values along with deformation, color and Flexicubes weights, respectively. Further, to avoid the unnecessary entanglements of different planes, we use rolled-out triplane [61]." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.499, + 0.787, + 0.606 + ], + "angle": 0, + "content": "Canonical Coordinates Map (CCM) We also add CCM as input [22], which contains extra geometry information. This is different from previous works, which typically use pure RGB images as the input to predict the 3D object [18]. Using pure RGB images make it extremely hard to predict the correct geometry, and sometime the geometry degrades (details in Sec. 4.3). Formally, CCM is the coordinates of each point in canonical space. It contains 3 channels whose values are within \\([0,1]\\), representing the coordinates in the canonical space." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.628, + 0.787, + 0.764 + ], + "angle": 0, + "content": "UNet-based Convolutional Network Our key insight is that the triplane is spatially aligned with the input six orthographic images and CCMs, as shown in Fig. 2. To match the rolled-out triplane, the six images and CCMs are arranged in a similar way. We render the six images and CCMs at a resolution of \\(256 \\times 256\\). They are split into two groups, with each group holding three images. These images in the four groups are then combined to create four larger images, each with a resolution of \\(256 \\times 768\\), allowing for spatial alignment. By concatenating these four groups, we form a 12-channel input. Next, a convolutional U-Net processes this input to produce the output triplane." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.766, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Compared to transformer-based methods [18, 21, 64, 71], our U-shape design has a larger bandwidth in preserving the input information, leading to highly detailed triplane features and finally elaborate textured meshes. Moreover, the convolutional network fully utilizes the geometry prior of the spatial correspondence of triplanes and input six orthographic images, which greatly fasten the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.486, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Convolutional Reconstruction Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.147, + 0.766, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.373, + 0.788, + 0.457 + ], + "angle": 0, + "content": "Fig. 4: Architecture along with training pipeline of CRM. We render the 3D mesh into six orthographic images and CCMs. Then the images and CCMs are concatenated and fed into the U-Net. The output triplane is decoded by small MLP networks to form the feature grid of Flexicubes, then textured mesh is get by dual marching cubes. During training, we render the color images, depth maps and masks from GT mesh and reconstructed mesh for supervision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.596 + ], + "angle": 0, + "content": "convergence and stabilize the training. Our model can get reasonable reconstruction results at very early stage of training (around 20 minutes of training from scratch). Also, our model can be trained with a much smaller batch size 32 (compared to transformer-based LRM that uses a batch size of 1024), which makes that all of our experiments can be conducted on an 8-GPU-cards machine. The overall training cost of our reconstruction model is only \\(1/8\\) than LRM. More details are shown in the experiments (see Sec. 4.1)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.771 + ], + "angle": 0, + "content": "Flexicubes Geometry Previous generic 3D generation methods mostly adopt NeRF [36] or Gaussian splatting [19] as the geometry representation, which relies on extra procedures like Marching Cubes (MC) to extract the iso-surface, suffering from topological ambiguities and struggling to represent high-fidelity geometric details. In this work, we use Flexicubes [47] as our geometry representation. It can get meshes from the features on the grid by dual marching cubes [37] during training. The features include SDF values, deformation and weights. The texture is obtained by querying the color at the surface. Flexicubes enables us to train our reconstruction model with textured mesh as the final output in an end-to-end manner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Loss Function Finally, to train our CRM model, we use a combination of MSE loss \\(\\mathcal{L}_{\\mathrm{MSE}}\\) and LPIPS loss [67] \\(\\mathcal{L}_{\\mathrm{LPIPS}}\\) for the rendered images for texture, similar as LRM [18]. To further enhance the geometry, we also include the depth map" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.517, + 0.788, + 0.547 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparison with baselines. Our models generate high-fidelity results with better geometry and texture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.574, + 0.633, + 0.59 + ], + "angle": 0, + "content": "and mask for supervision [16]. The overall loss function is" + }, + { + "type": "equation", + "bbox": [ + 0.268, + 0.599, + 0.786, + 0.636 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {\\mathrm {M S E}} (\\boldsymbol {x}, \\boldsymbol {x} ^ {\\mathrm {G T}}) + \\lambda_ {\\mathrm {L P I P S}} \\mathcal {L} _ {\\mathrm {L P I P S}} (\\boldsymbol {x}, \\boldsymbol {x} ^ {\\mathrm {G T}}) \\\\ + \\lambda_ {\\text {d e p t h}} \\mathcal {L} _ {\\text {M S E}} \\left(\\boldsymbol {x} _ {\\text {d e p t h}}, \\boldsymbol {x} _ {\\text {d e p t h}} ^ {\\mathrm {G T}}\\right) + \\lambda_ {\\text {m a s k}} \\mathcal {L} _ {\\text {M S E}} \\left(\\boldsymbol {x} _ {\\text {m a s k}}, \\boldsymbol {x} _ {\\text {m a s k}} ^ {\\mathrm {G T}}\\right) + \\lambda_ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}}, \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.719 + ], + "angle": 0, + "content": "where \\(\\pmb{x}\\), \\(\\pmb{x}_{\\mathrm{depth}}\\) and \\(\\pmb{x}_{\\mathrm{sil}}\\) represent the RGB image, depth map and mask rendered from the reconstruction textured mesh, respectively. And \\(\\pmb{x}^{\\mathrm{GT}}\\), \\(\\pmb{x}_{\\mathrm{depth}}^{\\mathrm{GT}}\\) and \\(\\pmb{x}_{\\mathrm{mask}}^{\\mathrm{GT}}\\) are rendered from the ground truth textured mesh. \\(\\mathcal{L}_{\\mathrm{reg}}\\) is the mesh-quality regularizers introduced in Flexicubes [47]. \\(\\lambda_{\\mathrm{LPIPS}}\\), \\(\\lambda_{\\mathrm{depth}}\\), \\(\\lambda_{\\mathrm{mask}}\\) and \\(\\lambda_{\\mathrm{reg}}\\) are the coefficients that balance each loss." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.741, + 0.378, + 0.759 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.771, + 0.442, + 0.789 + ], + "angle": 0, + "content": "4.1 Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Dataset We filter the Objaverse [13] dataset, removing scene-level objects and low quality meshes, and get around 376k valid high quality objects as the training set. We reuse the rendered images from SyncDreamer [31] which contain" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.486, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Convolutional Reconstruction Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.786, + 0.178 + ], + "angle": 0, + "content": "16 images per shape at the resolution of \\(256 \\times 256\\), and additionally render 6 orthographic images and CCM with the same lighting and resolution." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.195, + 0.787, + 0.242 + ], + "angle": 0, + "content": "Network Architecture The reconstruction model contains around 300M parameters. The U-Net contains [64, 128, 128, 256, 256, 512, 512] channels, with attention blocks at resolution [32, 16, 8]. We set the Flexicubes grid size as 80." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.258, + 0.788, + 0.392 + ], + "angle": 0, + "content": "Implementation Details The reconstruction model was trained on 8 NVIDIA A800 80GB GPU cards for 6 days with \\(110k\\) iterations. The model was trained with batch size 32 (32 shapes per iteration). At each iteration, we randomly sampled 8 views among the total 16 images for each shape for supervision. We used the Adam optimizer with learning rate \\(1e - 4\\). The coefficients that balancing each loss were set as \\(\\lambda_{\\mathrm{LPIPS}} = 0.1\\), \\(\\lambda_{\\mathrm{depth}} = 0.5\\), \\(\\lambda_{\\mathrm{mask}} = 0.5\\) and \\(\\lambda_{\\mathrm{reg}} = 0.005\\). To enhance robustness against minor inconsistency in the generated multi-view images, we introduced small Gaussian noise to the inputs in both training and inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.788, + 0.47 + ], + "angle": 0, + "content": "The diffusion models for both six orthographic images and CCMs were trained on 8 NVIDIA A800 80GB GPU cards for 2 days with \\(10k\\) iterations. The gradient accumulation is set as 12 steps, yielding a total batch size of 1536. We used the Adam optimizer with learning rate \\(5e - 5\\). During sampling both diffusions were sampled with 50 steps using DDIM [50]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.488, + 0.483, + 0.502 + ], + "angle": 0, + "content": "4.2 Comparison with baselines" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Qualitative Results To validate the effectiveness of our method, we qualitatively compare our results with previous works including Wonder3d [32], SyncDreamer [31], Magic123 [41], One-2-3-45 [29] and OpenLRM [17]. Since LRM [18] is not open-sourced, we use OpenLRM [17], an open-sourced implementation of LRM for comparisons. For the other baselines, we use their official codes and checkpoints. As for the input images for testing, we choose two from GSO [14] dataset, one downloaded from web and one generated by text-to-image diffusion model. The results are shown in Fig. 5. It can be seen from the figure that our method generates 3D textured meshes with better texture and geometry than all other baselines. This is because our reconstruction model fully utilizes the spatial alignment of input six orthographic images and output triplane. Also, our model can generate with only 10 seconds, much faster than most of the baselines. Our method is trained in an end-to-end manner with textured mesh as final output, thus avoiding a time-consuming post-processing for converting to mesh as in [52]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.81 + ], + "angle": 0, + "content": "Additionally, we visualize the generated meshes comparing to the previous work LRM [18] and a concurrent work LGM [52]. Since LRM is not open-sourced, we use the meshes from their project page. The results are shown in Fig. 6. It can be seen from the figure that our results have better texture. Our method also has smoother geometry than LRM and better geometry details than LGM." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In Fig. 7, we show more results of the high-fidelity textured meshes generated from single image by our method." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.145, + 0.787, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.47, + 0.788, + 0.501 + ], + "angle": 0, + "content": "Fig. 6: Qualitative comparison with LRM [18] and LGM [52]. Our models generate high-fidelity results with detailed texture and smooth geometry." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.545, + 0.788, + 0.728 + ], + "angle": 0, + "content": "Quantitative Results In line with previous studies [32], we evaluate our method using the Google Scanned Objects (GSO) dataset [14] which is not included in our training dataset. We randomly choose 30 shapes and render a single image with size of \\(256 \\times 256\\) as input for evaluation. To ensure the generated mesh accurately aligned with the ground truth mesh, we carefully adjust their pose and scale them to fit within the \\([-0.5, 0.5]\\) box. For mesh geometry evaluation, we report Chamfer Distance (CD), Column IoU and F-Score (with a threshold of 0.05, following One-2-3-45 [29]), which measure the geometry similarity between the reconstructed mesh and ground truth mesh. The results are shown in Table 1. It can be seen from the table that our method outperforms all of the baselines, which demonstrates the effectiveness of our method for geometry quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Furthermore, for evaluating mesh texture, we render 24 images at \\(512 \\times 512\\) resolution at elevation angles of 0, 15, and 30 degrees, for the generated meshes and ground-truth meshes respectively. For each elevation, the 8 images are evenly distributed around a full 360-degree rotation. Then we assessed them using several metrics: PSNR, SSIM, LPIPS and Clip-Similarity, which measure the resemblance in appearance between the reconstructed mesh and the original ground truth mesh. The results are shown in Table 2. It shows that the generated" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.486, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Convolutional Reconstruction Model" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.147, + 0.779, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.528, + 0.771, + 0.543 + ], + "angle": 0, + "content": "Fig. 7: More Results of the generated results of our method from a single image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.574, + 0.785, + 0.604 + ], + "angle": 0, + "content": "textured meshes by our model surpass those of all baselines in appearance, which demonstrates the effectiveness of our method for texture quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.605, + 0.788, + 0.727 + ], + "angle": 0, + "content": "Additionally, we carry out experiments to evaluate the effectiveness of our single image to multi-view diffusion models. We use PSNR, SSIM, and LPIPS to measure the similarity between the generated multi-view images and ground truth multi-view images. For this analysis, we use four views of the generated images (left, right, front, back) from each model under comparison, setting the background color to grey (value 128). We compare with SyncDreamer [31] and Wonder3D [32]. The outcomes of this evaluation are documented in Table 3. It can be seen from the table that our method outperforms all the baselines." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.751, + 0.502, + 0.767 + ], + "angle": 0, + "content": "4.3 Ablation Study and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Reconstruction Results on Early Training Stage An advantage of our CRM is that it is easy to train. In fact, we find that CRM starts to show reasonable results at very early stage of training. The results are shown in Fig. 8. The results are good even with barely 280 iterations (only 20 minutes of training)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison for the geometry quality between our method and baselines for single image to 3D textured mesh generation. We report the metrics of Chamfer Distance, Column IoU and F-score on GSO dataset." + }, + { + "type": "table", + "bbox": [ + 0.311, + 0.199, + 0.693, + 0.341 + ], + "angle": 0, + "content": "
MethodChamfer Dist.↓ Vol. IoU↑ F-Sco. (%)↑
One-2-3-45 [29]0.01720.446372.19
SyncDreamer [31]0.01400.390075.74
Wonder3D [32]0.01860.439876.75
Magic123 [41]0.01880.371460.66
TGS [71]0.01720.298265.17
OpenLRM [17,18]0.01680.377463.22
LGM [52]0.01170.468568.69
Ours0.00940.613179.38
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.355, + 0.788, + 0.397 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison for the texture quality between our method and baselines for single image to 3D textured mesh generation. We report the metric of PSNR, SSIM, LPIPS and Clip [43]-Similarity on GSO dataset." + }, + { + "type": "table", + "bbox": [ + 0.317, + 0.409, + 0.687, + 0.537 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑LPIPS↓Clip-Sim↑
One-2-3-45 [29]13.930.80840.262579.83
SyncDreamer [31]14.000.81650.259182.76
Wonder3D [32]13.310.81210.255483.70
Magic123 [41]12.690.79840.244285.16
OpenLRM [17,18]14.300.82940.227684.20
LGM [71]13.280.79460.256085.20
Ours16.220.83810.214387.55
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.566, + 0.787, + 0.597 + ], + "angle": 0, + "content": "We conjecture that the fast convergence results from the strong geometry prior in our architecture design." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.619, + 0.788, + 0.726 + ], + "angle": 0, + "content": "Training Time of CRM In Fig. 9 we compare the training cost between our method (reconstruction model only) and two baselines, LRM [18] and LGM [52]. We measure the training cost by the training days multiplying the amount of used NVIDIA A100/A800 GPU cards. It can be seen that our model takes much smaller training time than the two baselines. This is because our model utilizes the spatial correspondence between input six orthographic images/CCMs and triplanes, which serves as a strong prior that makes the training easier." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.748, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Importance of Input CCM We examine the importance of the CCMs that are concatenated to the input images. To compare, we train a reconstruction model that takes only the six RGB images as input, without CCMs. The results are shown in Fig. 10. It can be seen that the results of the geometry degrade a lot without CCM input. This is because CCM provides important geometry information for the model, especially when the geometry is complex." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.486, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Convolutional Reconstruction Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 3: Quantitative comparison between our method and baselines for novel view synthesis of the multi-view diffusion model. We report the metric of PSNR, SSIM and LPIPS on GSO dataset." + }, + { + "type": "table", + "bbox": [ + 0.356, + 0.199, + 0.646, + 0.272 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑LPIPS↓
SyncDreamer [31]20.300.78040.2932
Wonder3D [32]23.760.81270.2210
Ours29.360.87210.1354
" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.309, + 0.341, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.351, + 0.333, + 0.375 + ], + "angle": 0, + "content": "(a) 0 iters (Initialization)" + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.294, + 0.47, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.351, + 0.471, + 0.376 + ], + "angle": 0, + "content": "(b) 280 iters \\((\\sim 20\\) min training)" + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.38, + 0.348, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.438, + 0.347, + 0.464 + ], + "angle": 0, + "content": "(c) 4k iters \\((\\sim 5\\) hours training)" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.38, + 0.437, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.364, + 0.437, + 0.462, + 0.46 + ], + "angle": 0, + "content": "(d) Ground Truth Mesh" + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.479, + 0.49, + 0.521 + ], + "angle": 0, + "content": "Fig. 8: Reconstruction results on unseen samples during early stage of training." + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.29, + 0.766, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.479, + 0.776, + 0.52 + ], + "angle": 0, + "content": "Fig. 9: Training cost comparison. Our model require much less computation cost than other baselines." + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.545, + 0.43, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.621, + 0.328, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.69, + 0.387, + 0.704 + ], + "angle": 0, + "content": "(a) w/o CCM Input" + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.548, + 0.429, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.621, + 0.418, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.449, + 0.548, + 0.636, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.62, + 0.532, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.479, + 0.69, + 0.591, + 0.704 + ], + "angle": 0, + "content": "(b) w/ CCM Input" + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.545, + 0.745, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.659, + 0.621, + 0.737, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.69, + 0.767, + 0.704 + ], + "angle": 0, + "content": "(c) Ground Truth Mesh" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.719, + 0.788, + 0.775 + ], + "angle": 0, + "content": "Fig. 10: The CCM concatenated to the input images is beneficial for our model. (a) Without providing CCM, the model outputs a geometry which is reasonable, but not very good. (b) The shape reconstructed using our full model with CCM input, with a much better geometry. (c) Ground truth mesh rendered from the same pose." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Design of Multi-view Diffusion Here we examine the effectiveness of the design of the multi-view diffusion models. Starting from the baseline that naively" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 4: Ablation study on the design of multi-view diffusion on novel view synthesis. We report the metrics of PSNR, SSIM and LPIPS on GSO dataset." + }, + { + "type": "table", + "bbox": [ + 0.33, + 0.185, + 0.673, + 0.272 + ], + "angle": 0, + "content": "
MethodPSNR↑ SSIM↑ LPIPS↓
ImageDream (6 view)28.990.85650.1497
+ Zero-SNR29.130.85980.1498
+ Random Resizing29.360.87210.1354
+ Contour Augmentation28.920.86810.1444
" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.287, + 0.311, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.365, + 0.305, + 0.377 + ], + "angle": 0, + "content": "(a) Input Image" + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.287, + 0.477, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.365, + 0.461, + 0.377 + ], + "angle": 0, + "content": "(b) Segmentation Result" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.289, + 0.574, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.366, + 0.584, + 0.388 + ], + "angle": 0, + "content": "(c) w/o Contour Augmentation" + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.289, + 0.666, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.602, + 0.365, + 0.676, + 0.388 + ], + "angle": 0, + "content": "(d) w/ Contour Augmentation" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.292, + 0.761, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.689, + 0.366, + 0.779, + 0.387 + ], + "angle": 0, + "content": "(e) Contour Augmented Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.399, + 0.789, + 0.469 + ], + "angle": 0, + "content": "Fig. 11: Demonstration of contour augmentation. (a) Given an input image, (b) off-the-shelf segmentation model sometimes provides imperfect results. (c) Without contour augmentation, the predicted backview color is sensitive to the contour. (d) With contour augmentation, the model predicts reasonable result. (e) We demonstrate how we augment the input image during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.789, + 0.624 + ], + "angle": 0, + "content": "fine-tunes the pre-trained ImageDream model with 2 additional views, we sequentially add the proposed techniques on the training. We examine the results on a subset of GSO, comparing the similarity of the generated novel view images with the ground truth images using PSNR, SSIM and LPIPS metrics. The results are shown in table 4. It can be seen that both the Zero-SNR trick and random resizing are beneficial. Note that the contour augmentation does not improve the quantitative metrics. However, we find that this trick makes the model more robust to in the wild input images (Fig. 11)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.651, + 0.36, + 0.667 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.788, + 0.794 + ], + "angle": 0, + "content": "In this work, we present a convolutional reconstruction model (CRM) for creating high-quality 3D models from a single image. Our approach effectively utilizes the spatial relationship between input images and the output triplane, leading to improved textured meshes, with significantly less training cost compared to previous transformer-based methods [18]. The model operates on an end-to-end training basis, directly outputting textured meshes. Overall, our method can produce detailed textured meshes in just 10 seconds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Potential Negative Impact. Similar to many other generated models, our CRM may be used to generate malicious or fake 3D contents, which may need additional caution." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.486, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Convolutional Reconstruction Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.79, + 0.237 + ], + "angle": 0, + "content": "This work was supported by Natural Science Foundation of China (No.s 62350080, 62076147), Tsinghua Institute for Guo Qiang, and the High Performance Computing Center, Tsinghua University. J.Z is also supported by the New Cornerstone Science Foundation through the XPlorer Prize." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.259, + 0.323, + 0.275 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.315 + ], + "angle": 0, + "content": "1. Bao, F., Nie, S., Xue, K., Cao, Y., Li, C., Su, H., Zhu, J.: All are worth words: A vit backbone for diffusion models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.317, + 0.785, + 0.37 + ], + "angle": 0, + "content": "2. Brooks, T., Peebles, B., Homes, C., DePue, W., Guo, Y., Jing, L., Schnurr, D., Taylor, J., Luhman, T., Luhman, E., Ng, C., Wang, R., Ramesh, A.: Video generation models as world simulators (2024), https://openai.com/research/video-generation-models-as-world-simulators" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.373, + 0.785, + 0.453 + ], + "angle": 0, + "content": "3. Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D.M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.455, + 0.785, + 0.509 + ], + "angle": 0, + "content": "4. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.511, + 0.785, + 0.536 + ], + "angle": 0, + "content": "5. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision. pp. 333-350. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.538, + 0.785, + 0.578 + ], + "angle": 0, + "content": "6. Chen, H., Gu, J., Chen, A., Tian, W., Tu, Z., Liu, L., Su, H.: Single-stage diffusion nef: A unified approach to 3d generation and reconstruction. arXiv preprint arXiv:2304.06714 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.58, + 0.785, + 0.619 + ], + "angle": 0, + "content": "7. Chen, R., Chen, Y., Jiao, N., Jia, K.: *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation*. arXiv preprint arXiv:2303.13873 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.621, + 0.785, + 0.646 + ], + "angle": 0, + "content": "8. Chen, Y., Zhang, C., Yang, X., Cai, Z., Yu, G., Yang, L., Lin, G.: It3d: Improved text-to-3d generation with explicit view synthesis (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.649, + 0.785, + 0.673 + ], + "angle": 0, + "content": "9. Chen, Z., Wang, F., Liu, H.: Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.676, + 0.785, + 0.701 + ], + "angle": 0, + "content": "10. Cheng, Y.C., Lee, H.Y., Tuyakov, S., Schwing, A., Gui, L.: SDFusion: Multimodal 3d shape completion, reconstruction, and generation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.704, + 0.785, + 0.742 + ], + "angle": 0, + "content": "11. Decatur, D., Lang, I., Aberman, K., Hanocka, R.: 3d paintbrush: Local stylization of 3d shapes with cascaded score distillation. arXiv preprint arXiv:2311.09571 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.744, + 0.785, + 0.784 + ], + "angle": 0, + "content": "12. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objaverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.786, + 0.787, + 0.839 + ], + "angle": 0, + "content": "13. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objaverse: A universe of annotated 3d objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13142-13153 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.289, + 0.787, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "14. Downs, L., Francis, A., Koenig, N., Kinman, B., Hickman, R., Reymann, K., McHugh, T.B., Vanhoucke, V.: Google scanned objects: A high-quality dataset of 3d scanned household items. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 2553-2560. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.787, + 0.258 + ], + "angle": 0, + "content": "15. Gao, J., Shen, T., Wang, Z., Chen, W., Yin, K., Li, D., Litany, O., Gojcic, Z., Fidler, S.: Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems 35, 31841-31854 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.258, + 0.787, + 0.285 + ], + "angle": 0, + "content": "16. Gupta, A., Xiong, W., Nie, Y., Jones, I., Oğuz, B.: 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.285, + 0.787, + 0.312 + ], + "angle": 0, + "content": "17. He, Z., Wang, T.: Openlrm: Open-source large reconstruction models. https://github.com/3DTopia/OpenLRM (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.312, + 0.787, + 0.353 + ], + "angle": 0, + "content": "18. Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D., Liu, F., Sunkavalli, K., Bui, T., Tan, H.: Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.353, + 0.787, + 0.38 + ], + "angle": 0, + "content": "19. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.38, + 0.787, + 0.406 + ], + "angle": 0, + "content": "20. Kim, S., Lee, K., Choi, J.S., Jeong, J., Sohn, K., Shin, J.: Collaborative score distillation for consistent visual synthesis (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.406, + 0.787, + 0.447 + ], + "angle": 0, + "content": "21. Li, J., Tan, H., Zhang, K., Xu, Z., Luan, F., Xu, Y., Hong, Y., Sunkavalli, K., Shakhnarovich, G., Bi, S.: Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.447, + 0.787, + 0.474 + ], + "angle": 0, + "content": "22. Li, W., Chen, R., Chen, X., Tan, P.: Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.474, + 0.787, + 0.501 + ], + "angle": 0, + "content": "23. Li, Y., Dou, Y., Shi, Y., Lei, Y., Chen, X., Zhang, Y., Zhou, P., Ni, B.: Focaldreamer: Text-driven 3d editing via focal-fusion assembly (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.501, + 0.787, + 0.528 + ], + "angle": 0, + "content": "24. Liang, Y., Yang, X., Lin, J., Li, H., Xu, X., Chen, Y.: Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.528, + 0.787, + 0.582 + ], + "angle": 0, + "content": "25. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300–309 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.582, + 0.787, + 0.623 + ], + "angle": 0, + "content": "26. Lin, S., Liu, B., Li, J., Yang, X.: Common diffusion noise schedules and sample steps are flawed. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5404-5411 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.623, + 0.787, + 0.651 + ], + "angle": 0, + "content": "27. Liu, F., Wu, D., Wei, Y., Rao, Y., Duan, Y.: Sherpa3d: Boosting high-fidelity text-to-3d generation via coarse 3d prior (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.651, + 0.787, + 0.691 + ], + "angle": 0, + "content": "28. Liu, M., Shi, R., Chen, L., Zhang, Z., Xu, C., Wei, X., Chen, H., Zeng, C., Gu, J., Su, H.: One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion. arXiv preprint arXiv:2311.07885 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.691, + 0.787, + 0.732 + ], + "angle": 0, + "content": "29. Liu, M., Xu, C., Jin, H., Chen, L., Xu, Z., Su, H., et al.: One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.732, + 0.787, + 0.759 + ], + "angle": 0, + "content": "30. Liu, R., Wu, R., Hoorick, B.V., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.759, + 0.787, + 0.799 + ], + "angle": 0, + "content": "31. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "32. Long, X., Guo, Y.C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.H., Habermann, M., Theobalt, C., et al.: Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.486, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Convolutional Reconstruction Model" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "33. Lorraine, J., Xie, K., Zeng, X., Lin, C.H., Takikawa, T., Sharp, N., Lin, T.Y., Liu, M.Y., Fidler, S., Lucas, J.: Att3d: Amortized text-to-3d object synthesis (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "34. Lu, Y., Zhang, J., Li, S., Fang, T., McKinnon, D., Tsin, Y., Quan, L., Cao, X., Yao, Y.: Direct2.5: Diverse text-to-3d generation via multi-view 2.5d diffusion (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.204, + 0.785, + 0.245 + ], + "angle": 0, + "content": "35. Mercier, A., Nakhli, R., Reddy, M., Yasarla, R., Cai, H., Porikli, F., Berger, G.: Hexagen3d: Stablediffusion is just one step away from fast and diverse text-to-3d generation. arXiv preprint arXiv:2401.07727 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "36. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "37. Nielson, G.M.: Dual marching cubes. In: IEEE visualization 2004. pp. 489-496. IEEE (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.315, + 0.722, + 0.328 + ], + "angle": 0, + "content": "38. Peebles, W., Xie, S.: Scalable diffusion models with transformers (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.329, + 0.785, + 0.355 + ], + "angle": 0, + "content": "39. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.356, + 0.785, + 0.397 + ], + "angle": 0, + "content": "40. Qian, G., Cao, J., Siarohin, A., Kant, Y., Wang, C., Vasilkovsky, M., Lee, H.Y., Fang, Y., Skorokhodov, I., Zhuang, P., Gilitschenski, I., Ren, J., Ghanem, B., Aberman, K., Tulyakov, S.: Atom: Amortized text-to-mesh using 2d diffusion (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "41. Qian, G., Mai, J., Hamdi, A., Ren, J., Siarohin, A., Li, B., Lee, H.Y., Skorokhodov, I., Wonka, P., Tulyakov, S., Ghanem, B.: Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "42. Qiu, L., Chen, G., Gu, X., Zuo, Q., Xu, M., Wu, Y., Yuan, W., Dong, Z., Bo, L., Han, X.: Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.481, + 0.785, + 0.535 + ], + "angle": 0, + "content": "43. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "44. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.578, + 0.785, + 0.632 + ], + "angle": 0, + "content": "45. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.633, + 0.785, + 0.687 + ], + "angle": 0, + "content": "46. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: Laion-5b: An open large-scale dataset for training next generation image-text models (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.688, + 0.785, + 0.73 + ], + "angle": 0, + "content": "47. Shen, T., Munkberg, J., Hasselgren, J., Yin, K., Wang, Z., Chen, W., Gojcic, Z., Fidler, S., Sharp, N., Gao, J.: Flexible isosurface extraction for gradient-based mesh optimization. ACM Transactions on Graphics (TOG) 42(4), 1-16 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.73, + 0.785, + 0.757 + ], + "angle": 0, + "content": "48. Shi, R., Wei, X., Wang, C., Su, H.: Zerorf: Fast sparse view 360 \\(\\{\\backslash \\mathrm{deg}\\}\\) reconstruction with zero pretraining. arXiv preprint arXiv:2312.09249 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "49. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.786, + 0.785, + 0.812 + ], + "angle": 0, + "content": "50. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "51. Sun, J., Zhang, B., Shao, R., Wang, L., Liu, W., Xie, Z., Liu, Y.: Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "52. Tang, J., Chen, Z., Chen, X., Wang, T., Zeng, G., Liu, Z.: Lgm: Large multi-view gaussian model for high-resolution 3d content creation (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "53. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.242 + ], + "angle": 0, + "content": "54. Tochilkin, D., Pankratz, D., Liu, Z., Huang, Z., Letts, A., Li, Y., Liang, D., Laforte, C., Jampani, V., Cao, Y.P.: Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.785, + 0.282 + ], + "angle": 0, + "content": "55. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.283, + 0.785, + 0.31 + ], + "angle": 0, + "content": "56. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.31, + 0.785, + 0.336 + ], + "angle": 0, + "content": "57. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.336, + 0.785, + 0.363 + ], + "angle": 0, + "content": "58. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.364, + 0.785, + 0.389 + ], + "angle": 0, + "content": "59. Wang, P., Shi, Y.: Imaginedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.39, + 0.785, + 0.43 + ], + "angle": 0, + "content": "60. Wang, P., Tan, H., Bi, S., Xu, Y., Luan, F., Sunkavalli, K., Wang, W., Xu, Z., Zhang, K.: Pf-lrm: Pose-free large reconstruction model for joint pose and shape prediction. arXiv preprint arXiv:2311.12024 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.43, + 0.785, + 0.484 + ], + "angle": 0, + "content": "61. Wang, T., Zhang, B., Zhang, T., Gu, S., Bao, J., Baltrusaitis, T., Shen, J., Chen, D., Wen, F., Chen, Q., et al.: Rodin: A generative model for sculpting 3d digital avatars using diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4563-4573 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.484, + 0.785, + 0.523 + ], + "angle": 0, + "content": "62. Wang, X., Wang, Y., Ye, J., Wang, Z., Sun, F., Liu, P., Wang, L., Sun, K., Wang, X., He, B.: Animatabledreamer: Text-guided non-rigid 3d model generation and reconstruction with canonical score distillation" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.525, + 0.785, + 0.579 + ], + "angle": 0, + "content": "63. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=ppJuFSOAnM" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.58, + 0.785, + 0.619 + ], + "angle": 0, + "content": "64. Xu, Y., Tan, H., Luan, F., Bi, S., Wang, P., Li, J., Shi, Z., Sunkavalli, K., Wetzstein, G., Xu, Z., et al.: Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model. arXiv preprint arXiv:2311.09217 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.645 + ], + "angle": 0, + "content": "65. Yu, X., Guo, Y.C., Li, Y., Liang, D., Zhang, S.H., Qi, X.: Text-to-3d with classifier score distillation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.646, + 0.785, + 0.672 + ], + "angle": 0, + "content": "66. Zeng, X., Vahdat, A., Williams, F., Gojcic, Z., Litany, O., Fidler, S., Kreis, K.: Lion: Latent point diffusion models for 3d shape generation (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.673, + 0.785, + 0.699 + ], + "angle": 0, + "content": "67. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.699, + 0.785, + 0.725 + ], + "angle": 0, + "content": "68. Zheng, X.Y., Pan, H., Guo, Y.X., Tong, X., Liu, Y.: Mvd\\(^2\\): Efficient multiview 3d reconstruction for multiview diffusion (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.726, + 0.785, + 0.765 + ], + "angle": 0, + "content": "69. Zheng, X.Y., Pan, H., Wang, P.S., Tong, X., Liu, Y., Shum, H.Y.: Locally attentional sdf diffusion for controllable 3d shape generation. arXiv preprint arXiv:2305.04461 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.766, + 0.785, + 0.792 + ], + "angle": 0, + "content": "70. Zhu, J., Zhuang, P.: Hifa: High-fidelity text-to-3d generation with advanced diffusion guidance (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.793, + 0.785, + 0.833 + ], + "angle": 0, + "content": "71. Zou, Z.X., Yu, Z., Guo, Y.C., Li, Y., Liang, D., Cao, Y.P., Zhang, S.H.: Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. arXiv preprint arXiv:2312.09147 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.833 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CRM_ Single Image to 3D Textured Mesh with Convolutional Reconstruction Model/9e4b61b2-e1dd-4f99-9ab9-4fe8376f780a_origin.pdf b/2024/CRM_ Single Image to 3D Textured Mesh with Convolutional Reconstruction Model/9e4b61b2-e1dd-4f99-9ab9-4fe8376f780a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..80faf527f893136bed2e57a2885ae6ec7088fe96 --- /dev/null +++ b/2024/CRM_ Single Image to 3D Textured Mesh with Convolutional Reconstruction Model/9e4b61b2-e1dd-4f99-9ab9-4fe8376f780a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad958801cb67b410e8b712d0cbbdb51b3089ea2f1321c905f51ac8217a7c60af +size 5348247 diff --git a/2024/CRM_ Single Image to 3D Textured Mesh with Convolutional Reconstruction Model/full.md b/2024/CRM_ Single Image to 3D Textured Mesh with Convolutional Reconstruction Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3f8c256cc358a6103a3819a117f668bd5a7a83ca --- /dev/null +++ b/2024/CRM_ Single Image to 3D Textured Mesh with Convolutional Reconstruction Model/full.md @@ -0,0 +1,361 @@ +# CRM: Single Image to 3D Textured Mesh with Convolutional Reconstruction Model + +Zhengyi Wang $^{1,2}$ , Yikai Wang $^{1,2}$ , Yifei Chen $^{1}$ , Chendong Xiang $^{1,2}$ , Shuo Chen $^{1}$ , Dajiang Yu $^{1}$ , Chongxuan Li $^{3}$ , Hang Su $^{1}$ and Jun Zhu $^{\star 1,2}$ + +$^{1}$ Dept. of Comp. Sci. & Tech., BNrist Center, Tsinghua-Bosch Joint ML Center, Tsinghua University; + +2 ShengShu, Beijing, China; + +$^{3}$ Gaoling School of Artificial Intelligence, Renmin University of China, Beijing Key Laboratory of Big Data Management and Analysis Methods, Beijing, China. + +![](images/179d3eff26b50d33ee74c72b3ffbf2ba60d92125c82f0cba79fa432c8c9aaf6e.jpg) +Input Image + +![](images/df48604738fb65a7d7389664d381d7ac4cdc7c8a18a0d60a735423725ee47550.jpg) + +![](images/798007dac4cff8fbbb561379e256148586715340dfc6f87669e84557b791aa4b.jpg) + +![](images/591f7aa234c38dffcd2fbb7179ba047e740ecbd855bd15b4d4539f817afe9bb8.jpg) + +![](images/76c809462deddf1df94da004a029e25690492d60bf777d8ecb1fb570ed58d38b.jpg) +Fig. 1: CRM generates high-fidelity textured mesh from single image in 10 seconds. + +![](images/e8682e31401391693f271d493b90e151c2237e0622046c56b19359da91421dfe.jpg) + +![](images/f699a043da85542a61278aac2dea1ac2c32e84720e953c62cf419b5ca66ab373.jpg) + +![](images/f2f2a3ebc059277458dcee7aa5e010e10c9d74ad98bfed40b9f9ec17d4f422ac.jpg) + +![](images/f6ddc99930a0d96d2ecd56c02476009fbbc226d08f5a4aabb8b7a3ed3c5eef83.jpg) + +![](images/e00f04b10c425cc88c3872edf2f0db16ce42e17e07d14d23c543be825541dd16.jpg) + +![](images/6d6cddfd4e2a9d79488994b74a58427941189f98e50d01485658b8c0e722c5b1.jpg) +Generated Textured Mesh + +![](images/64cdf860448bffcc6f38c258baa22100d0348a80fcc1977095c0602734c3ee36.jpg) + +![](images/bb52f3d47de4c2c2908ab641a8a694257ed1204a2c85fb4431cb7cd3d886c947.jpg) + +![](images/ca39cd3d253d95c4592937f21037fce576f4e2eda993f7891737929ee6a20172.jpg) + +![](images/c4c87c5e1ef94b66d4389518a8081d10be6e3505cee42f7a5ddb143d9aba1697.jpg) + +![](images/5e02daf5ff5a6f19fd1a7212f8dd098f1209e7268dd7adf6f6ce286974153efd.jpg) + +![](images/6d45e051e2e78d6762db3e8fd7c444d362a5a86a70207ee6a2779832ae41ca3e.jpg) + +![](images/4b730102c951ebaf3517452d75773288fce30dc434ee92d0d9b0babf14b0e8ce.jpg) + +![](images/aab8a59f0ea80db81f3403de012c53c6c843aa70ef619e39adde9423a596e774.jpg) + +![](images/d837c50dcf46eace8551e3fea5f9a8e965ee41e63095ecd4c70cd71680ce0323.jpg) + +![](images/4ebd870be2e7710cf158b9ee04d44e8ec6c603917c26408e5611c098ad516dbd.jpg) + +![](images/8e6a6e2902f982c95b7bcea635b0d732db15a2ff96451c4a760714b68325a3e9.jpg) + +![](images/50b9c42502a45048091168e9e8c0b68fb4b9a9bcf6310053e702cc2b3776b8ef.jpg) + +![](images/baa1ba49c0e39769ee921893dc80b48a3dadf28721b7b3d0fbfbdf0932c52a00.jpg) + +![](images/76802b0a9327190810c277f429bbc10a3a0a6fa9504d70a174e830cba6538c46.jpg) + +![](images/76dc3a4df519da49d09aad4c668e6491fa94e96fcef91587514f86d7e0f46875.jpg) + +![](images/c3ca2a83f79ee6844dd382a93e472905415ec64b976460977b6943258594e8e1.jpg) + +![](images/e3bd3ffc3ddef8ec997dc4332d5d7c40c9c9c9873e34d126198e327b919277dd.jpg) + +![](images/fb32a0150f49ab7c1a3d53b1c5c89c8da46aa86fdbc43d80bceaaa985afff56b.jpg) + +![](images/416d4c6de8015f60630874f0a601921f3a3cf048ff8f50413205256586da69fd.jpg) + +Abstract. Feed-forward 3D generative models like the Large Reconstruction Model (LRM) [18] have demonstrated exceptional generation speed. However, the transformer-based methods do not leverage the geometric priors of the triplane component in their architecture, often leading to sub-optimal quality given the limited size of 3D data and + +slow training. In this work, we present the Convolutional Reconstruction Model (CRM), a high-fidelity feed-forward single image-to-3D generative model. Recognizing the limitations posed by sparse 3D data, we highlight the necessity of integrating geometric priors into network design. CRM builds on the key observation that the visualization of triplane exhibits spatial correspondence of six orthographic images. First, it generates six orthographic view images from a single input image, then feeds these images into a convolutional U-Net, leveraging its strong pixel-level alignment capabilities and significant bandwidth to create a high-resolution triplane. CRM further employs Flexicubes as geometric representation, facilitating direct end-to-end optimization on textured meshes. Overall, our model delivers a high-fidelity textured mesh from an image in just 10 seconds, without any test-time optimization. + +Keywords: 3D Generation $\cdot$ Textured Mesh $\cdot$ Diffusion Models + +# 1 Introduction + +In recent years, generative models have witnessed significant advancements, largely attributed to the fast growth in data size. Transformers [56], in particular, have achieved high-performance results across various domains including language [3], image [1,38] and video generation [2]. However, the domain of 3D generation presents unique challenges. Unlike the abundance of other modal's data, 3D data is comparatively scarce. The creation of 3D data requires specialized expertise and considerable time, leading to a situation where the largest 3D datasets, namely Objaverse [12,13], only contain millions of 3D content, much smaller than image datasets like Laion [46] which contains 5 billion images. + +Despite this, recent developments have introduced some transformer-based methods [18, 21, 60, 64, 71] like LRM [18] for creating 3D content from single or multi-view images in a feed-forward manner. Among these models, the triplane has emerged as a popular component due to its efficiency in generating high-resolution 3D results with minimal memory cost. However, reliance on transformer-based networks for generating triplane patches has not utilized the geometric priors inherent to the triplane concept, leading to sub-optimal results in terms of quality and fidelity, and long training time. + +To address the above challenges, this paper presents a new Convolutional Reconstruction Model (CRM) with high generation quality as well as fast training. Given the limited amount of 3D contents, CRM builds on a key hypothesis that it is beneficial to explore geometric priors in architecture design. Namely, we observe from the visualization of triplane [5,6,48] that triplane exhibits spatial correspondence of input six orthographic images, as shown in Fig. 2. The silhouette and texture of the input images have a natural alignment with the triplane structure. This motivates us to (1) use six orthographic images as input images to reconstruct 3D contents, which align well as the triplane feature, instead of other arbitrarily chosen poses; (2) use a U-Net convolutional network to map the input images to a rolled-out triplane by exploring the strong pixel-level alignment + +between the input and output. Furthermore, the significant bandwidth capacity of our U-Net enables a direct transformation of the six orthographic images into the triplane, yielding highly detailed outcomes. Besides, we also add Canonical Coordinate Map (CCM) to the reconstruction network, a novel addition that enriches the model's understanding of spatial relations and geometry. + +For the task of 3D generation from a single image, as the six orthographic images and CCMs are not directly available, we train a multi-view diffusion model conditioned on the input image to generate the six orthographic images and another diffusion model to generate the CCMs conditioned on the generated six orthographic images. Both diffusion models are trained on a filtered version of the Objaverse dataset [13]. To further enhance quality and robustness, we implement training improvements for the multi-view diffusion models, including Zero-SNR [26], random resizing, and contour augmentation. + +Finally, as directly optimizing high-quality textured meshes is challenging, we adopt Flexicubes [47] as the geometry representation to falcitate gradient-based mesh optimization. This is unlike previous works [18, 71] that use alternative 3D representation like NeRF [36] or Gaussian Splatting [19]. Such methods often involve extra procedure steps to obtain textured meshes [52], although they can produce detailed visualizations. With our designs, we are able to train CRM with textured mesh as the final output in an end-to-end manner, and our approach has a more straightforward inference pipeline and better mesh quality. Overall, our method can generate high-fidelity textured mesh within 10 seconds, as shown in Fig. 1. + +# 2 Related Works + +# 2.1 Score Distillation for 3D Generation + +DreamFusion [39] proposes a technique called Score Distillation Sampling (SDS) (also known as Score Jacobian Chaining [57]). It utilizes large scale image diffusion models [44,45] to iteratively refine 3D models to align with specific prompts or images. Thus it can generate 3D content without training on 3D dataset. Along this line, ProlificDreamer [63] proposes Variational Score Distillation (VSD), a principled variational framework which greatly mitigates the over-saturation problems in SDS and improves the diversity. Zero123 [30], MVDream [49], ImageDream [59] and many others [22,41,42] further improve the results and mitigate the multi-face problems using diffusion models fine-tuned on 3D data. [33,40] explore amortized score distillation. Many other works [7-9,11,20,23-25,27,51,53,55,62,65,70] improve the results a lot, in either speed or quality. However, methods based on score distillation usually take from minutes to hours to generate single object, which is computationally expensive. + +# 2.2 3D Generation with Sparse View Reconstruction + +Several approaches aim to generate multi-view consistent images and then create 3D contents using sparse views reconstruction. For example, SyncDreamer [31] + +![](images/56c9a154283ec62e3d80cc0acf3f444ab7884443cddd051d3f7966e931aa5396.jpg) +Fig. 2: One of our key motivation is that the triplane shares a strong spatial alignment with the input six orthographic images. (a) The six orthographic images of a input shape. (b) The six orthographic CCMs. (c) The triplane (mean value of all channels) output by our U-Net, which spatially aligns with the input images. (d) The textured mesh output by our convolutional reconstruction model. + +generates multi-view consistent images and then uses NeuS [58] for reconstruction. Wonder3D [32] improves the results with cross-domain diffusion. Direct2.5 [34] improves the results with 2.5D diffusion. However, one common issue of these methods is that they need test-time optimization for reconstruction with sparse views, which may lead to extra computing and compromise the final quality + +# 2.3 Feed-forward 3D Generative Models + +Some works try to generate 3D objects using a feed forward model [4, 10, 15, 16, 66, 69]. Feed-forward methods demonstrate significantly faster generation speeds compared to the two types of methods mentioned above. Recently there are some works trained on larger 3D dataset Objaverse [13]. One-2-3-45 [29] generates multi-view images and then feed the images into a network to get the 3D object. LRM series [18, 21, 60, 64] improve the quality of generated results with a transformer-based architecture. TGS [71] and LGM [52] use Gaussian Splitting [19] as the geometry representation. There are also many other works [28, 35, 54, 68] that improve the results with different techniques. Despite these advancements, there remains room for improvement in the network architecture or geometry representation. Our approach utilizes a network with a strategically designed architecture and an end-to-end training approach producing meshes directly as the final output. + +![](images/940ce2ff2c999b5c6d44e215b2a5fa46a46dd256259cc3a5012d957982f2a0b2.jpg) +Fig. 3: Overall pipeline of our method. The input image is fed into a multi-view image diffusion model to generate six orthographic images. Then another diffusion model is used to generate the CCMs conditioned on the six images. The six images along with the CCMs are send into CRM to reconstruct the final textured mesh. The whole inference process takes around 10 seconds on an A800 GPU. *The 4 seconds include the U-Net forward (less than 0.1s), querying surface points for UV texture and file I/O. + +# 3 Method + +In this section, we illustrate the detailed design of our method (shown in Fig. 3). Given a single input image, our model first utilizes multi-view diffusion models (Sec. 3.1) to generate six orthographic images and the canonical coordinates maps (CCMs). Then we develop our convolutional reconstruction model (CRM, Sec. 3.2) to reconstruct 3D textured mesh from the images and CCMs. + +# 3.1 Multi-view Diffusion Model + +We first explain the design of the multi-view diffusion model to generate six orthographic view images from a single input image. Instead of training from scratch, which is typically extremely expensive, we initialize the diffusion models using the checkpoint of ImageDream [59], a high-performance diffusion model for the task of multi-view images generation from a single image. The original ImageDream supports 4 views generation. We expand it to include 6 views by adding two more perspectives (up and down). We use another diffusion model which is conditioned on the generated six views to generate the canonical coordinate map. The conditional RGB image is concatenated with the noisy canonical coordinate map. It is also initialized from ImageDream checkpoint. Both the diffusion models are fine-tuned on the Objaverse [13] dataset. + +To further improve the quality and robustness of our results, we introduce several enhancements: (1) Zero-SNR Training. We use the zero-SNR trick as mentioned in [26]. This can alleviate the problem resulting from the discrepancy between the initial Gaussian noise during sampling and the noisiest training sample. (2) Random Resizing. A naive implementation would make the model tends to generate objects that occupy the entire image. To mitigate this, we randomly resize the objects when training. (3) Contour Augmentation. We find that the model tends to predict the backview color largely relying on the + +contour of the input view. To make the model insensitive to the contour, we randomly change the contour color during training. + +# 3.2 Convolutional Reconstruction Model + +We now move to introduce the detailed architecture of the convolutional reconstruction model (CRM). As outlined in Fig. 4, given the input six images and CCMs, a convolutional U-Net is used to map the input images along with the CCMs to a rolled-out triplane. Then the rolled-out triplane is reshaped into the triplane. Small multi-layer perceptions (MLPs) are used to decode the triplane features into SDF values, texture color and Flexicubes parameters. Lastly, these values are used to get texture mesh by dual marching cubes. Below, we explain the key components of CRM in detail. + +Triplane Representation We choose triplane as the 3D representation, because it can achieve high resolution 3D results with 2D computation consumption. It projects each query grid cell to axis-aligned orthogonal planes $(xy, xz,$ and $yz$ planes) and then aggregates the feature from each planes. Then the feature is decoded by 3 tiny MLPs with 2 hidden layers to get the SDF values along with deformation, color and Flexicubes weights, respectively. Further, to avoid the unnecessary entanglements of different planes, we use rolled-out triplane [61]. + +Canonical Coordinates Map (CCM) We also add CCM as input [22], which contains extra geometry information. This is different from previous works, which typically use pure RGB images as the input to predict the 3D object [18]. Using pure RGB images make it extremely hard to predict the correct geometry, and sometime the geometry degrades (details in Sec. 4.3). Formally, CCM is the coordinates of each point in canonical space. It contains 3 channels whose values are within $[0,1]$ , representing the coordinates in the canonical space. + +UNet-based Convolutional Network Our key insight is that the triplane is spatially aligned with the input six orthographic images and CCMs, as shown in Fig. 2. To match the rolled-out triplane, the six images and CCMs are arranged in a similar way. We render the six images and CCMs at a resolution of $256 \times 256$ . They are split into two groups, with each group holding three images. These images in the four groups are then combined to create four larger images, each with a resolution of $256 \times 768$ , allowing for spatial alignment. By concatenating these four groups, we form a 12-channel input. Next, a convolutional U-Net processes this input to produce the output triplane. + +Compared to transformer-based methods [18, 21, 64, 71], our U-shape design has a larger bandwidth in preserving the input information, leading to highly detailed triplane features and finally elaborate textured meshes. Moreover, the convolutional network fully utilizes the geometry prior of the spatial correspondence of triplanes and input six orthographic images, which greatly fasten the + +![](images/d4214eb88ff384be362cfc0c4e72a8822208ce107a45485ce0c668c79f6ad6b9.jpg) +Fig. 4: Architecture along with training pipeline of CRM. We render the 3D mesh into six orthographic images and CCMs. Then the images and CCMs are concatenated and fed into the U-Net. The output triplane is decoded by small MLP networks to form the feature grid of Flexicubes, then textured mesh is get by dual marching cubes. During training, we render the color images, depth maps and masks from GT mesh and reconstructed mesh for supervision. + +convergence and stabilize the training. Our model can get reasonable reconstruction results at very early stage of training (around 20 minutes of training from scratch). Also, our model can be trained with a much smaller batch size 32 (compared to transformer-based LRM that uses a batch size of 1024), which makes that all of our experiments can be conducted on an 8-GPU-cards machine. The overall training cost of our reconstruction model is only $1/8$ than LRM. More details are shown in the experiments (see Sec. 4.1). + +Flexicubes Geometry Previous generic 3D generation methods mostly adopt NeRF [36] or Gaussian splatting [19] as the geometry representation, which relies on extra procedures like Marching Cubes (MC) to extract the iso-surface, suffering from topological ambiguities and struggling to represent high-fidelity geometric details. In this work, we use Flexicubes [47] as our geometry representation. It can get meshes from the features on the grid by dual marching cubes [37] during training. The features include SDF values, deformation and weights. The texture is obtained by querying the color at the surface. Flexicubes enables us to train our reconstruction model with textured mesh as the final output in an end-to-end manner. + +Loss Function Finally, to train our CRM model, we use a combination of MSE loss $\mathcal{L}_{\mathrm{MSE}}$ and LPIPS loss [67] $\mathcal{L}_{\mathrm{LPIPS}}$ for the rendered images for texture, similar as LRM [18]. To further enhance the geometry, we also include the depth map + +![](images/800904ce077e7454446324a5ce216fd08d3262868c689bdda543685ba5769ce6.jpg) +Fig. 5: Qualitative comparison with baselines. Our models generate high-fidelity results with better geometry and texture. + +and mask for supervision [16]. The overall loss function is + +$$ +\begin{array}{l} \mathcal {L} = \mathcal {L} _ {\mathrm {M S E}} (\boldsymbol {x}, \boldsymbol {x} ^ {\mathrm {G T}}) + \lambda_ {\mathrm {L P I P S}} \mathcal {L} _ {\mathrm {L P I P S}} (\boldsymbol {x}, \boldsymbol {x} ^ {\mathrm {G T}}) \\ + \lambda_ {\text {d e p t h}} \mathcal {L} _ {\text {M S E}} \left(\boldsymbol {x} _ {\text {d e p t h}}, \boldsymbol {x} _ {\text {d e p t h}} ^ {\mathrm {G T}}\right) + \lambda_ {\text {m a s k}} \mathcal {L} _ {\text {M S E}} \left(\boldsymbol {x} _ {\text {m a s k}}, \boldsymbol {x} _ {\text {m a s k}} ^ {\mathrm {G T}}\right) + \lambda_ {\text {r e g}} \mathcal {L} _ {\text {r e g}}, \tag {1} \\ \end{array} +$$ + +where $\pmb{x}$ , $\pmb{x}_{\mathrm{depth}}$ and $\pmb{x}_{\mathrm{sil}}$ represent the RGB image, depth map and mask rendered from the reconstruction textured mesh, respectively. And $\pmb{x}^{\mathrm{GT}}$ , $\pmb{x}_{\mathrm{depth}}^{\mathrm{GT}}$ and $\pmb{x}_{\mathrm{mask}}^{\mathrm{GT}}$ are rendered from the ground truth textured mesh. $\mathcal{L}_{\mathrm{reg}}$ is the mesh-quality regularizers introduced in Flexicubes [47]. $\lambda_{\mathrm{LPIPS}}$ , $\lambda_{\mathrm{depth}}$ , $\lambda_{\mathrm{mask}}$ and $\lambda_{\mathrm{reg}}$ are the coefficients that balance each loss. + +# 4 Experiments + +# 4.1 Experimental Setting + +Dataset We filter the Objaverse [13] dataset, removing scene-level objects and low quality meshes, and get around 376k valid high quality objects as the training set. We reuse the rendered images from SyncDreamer [31] which contain + +16 images per shape at the resolution of $256 \times 256$ , and additionally render 6 orthographic images and CCM with the same lighting and resolution. + +Network Architecture The reconstruction model contains around 300M parameters. The U-Net contains [64, 128, 128, 256, 256, 512, 512] channels, with attention blocks at resolution [32, 16, 8]. We set the Flexicubes grid size as 80. + +Implementation Details The reconstruction model was trained on 8 NVIDIA A800 80GB GPU cards for 6 days with $110k$ iterations. The model was trained with batch size 32 (32 shapes per iteration). At each iteration, we randomly sampled 8 views among the total 16 images for each shape for supervision. We used the Adam optimizer with learning rate $1e - 4$ . The coefficients that balancing each loss were set as $\lambda_{\mathrm{LPIPS}} = 0.1$ , $\lambda_{\mathrm{depth}} = 0.5$ , $\lambda_{\mathrm{mask}} = 0.5$ and $\lambda_{\mathrm{reg}} = 0.005$ . To enhance robustness against minor inconsistency in the generated multi-view images, we introduced small Gaussian noise to the inputs in both training and inference. + +The diffusion models for both six orthographic images and CCMs were trained on 8 NVIDIA A800 80GB GPU cards for 2 days with $10k$ iterations. The gradient accumulation is set as 12 steps, yielding a total batch size of 1536. We used the Adam optimizer with learning rate $5e - 5$ . During sampling both diffusions were sampled with 50 steps using DDIM [50]. + +# 4.2 Comparison with baselines + +Qualitative Results To validate the effectiveness of our method, we qualitatively compare our results with previous works including Wonder3d [32], SyncDreamer [31], Magic123 [41], One-2-3-45 [29] and OpenLRM [17]. Since LRM [18] is not open-sourced, we use OpenLRM [17], an open-sourced implementation of LRM for comparisons. For the other baselines, we use their official codes and checkpoints. As for the input images for testing, we choose two from GSO [14] dataset, one downloaded from web and one generated by text-to-image diffusion model. The results are shown in Fig. 5. It can be seen from the figure that our method generates 3D textured meshes with better texture and geometry than all other baselines. This is because our reconstruction model fully utilizes the spatial alignment of input six orthographic images and output triplane. Also, our model can generate with only 10 seconds, much faster than most of the baselines. Our method is trained in an end-to-end manner with textured mesh as final output, thus avoiding a time-consuming post-processing for converting to mesh as in [52]. + +Additionally, we visualize the generated meshes comparing to the previous work LRM [18] and a concurrent work LGM [52]. Since LRM is not open-sourced, we use the meshes from their project page. The results are shown in Fig. 6. It can be seen from the figure that our results have better texture. Our method also has smoother geometry than LRM and better geometry details than LGM. + +In Fig. 7, we show more results of the high-fidelity textured meshes generated from single image by our method. + +![](images/89ea23725eeb3f89f3193c423cb2e43247383cd53dbdc7321f2e955a7ef6389f.jpg) +Fig. 6: Qualitative comparison with LRM [18] and LGM [52]. Our models generate high-fidelity results with detailed texture and smooth geometry. + +Quantitative Results In line with previous studies [32], we evaluate our method using the Google Scanned Objects (GSO) dataset [14] which is not included in our training dataset. We randomly choose 30 shapes and render a single image with size of $256 \times 256$ as input for evaluation. To ensure the generated mesh accurately aligned with the ground truth mesh, we carefully adjust their pose and scale them to fit within the $[-0.5, 0.5]$ box. For mesh geometry evaluation, we report Chamfer Distance (CD), Column IoU and F-Score (with a threshold of 0.05, following One-2-3-45 [29]), which measure the geometry similarity between the reconstructed mesh and ground truth mesh. The results are shown in Table 1. It can be seen from the table that our method outperforms all of the baselines, which demonstrates the effectiveness of our method for geometry quality. + +Furthermore, for evaluating mesh texture, we render 24 images at $512 \times 512$ resolution at elevation angles of 0, 15, and 30 degrees, for the generated meshes and ground-truth meshes respectively. For each elevation, the 8 images are evenly distributed around a full 360-degree rotation. Then we assessed them using several metrics: PSNR, SSIM, LPIPS and Clip-Similarity, which measure the resemblance in appearance between the reconstructed mesh and the original ground truth mesh. The results are shown in Table 2. It shows that the generated + +![](images/71497811b6dc9337e1c15445273e7a297d4a88dfe94ddc81167a9d4c8dea33ff.jpg) +Fig. 7: More Results of the generated results of our method from a single image. + +textured meshes by our model surpass those of all baselines in appearance, which demonstrates the effectiveness of our method for texture quality. + +Additionally, we carry out experiments to evaluate the effectiveness of our single image to multi-view diffusion models. We use PSNR, SSIM, and LPIPS to measure the similarity between the generated multi-view images and ground truth multi-view images. For this analysis, we use four views of the generated images (left, right, front, back) from each model under comparison, setting the background color to grey (value 128). We compare with SyncDreamer [31] and Wonder3D [32]. The outcomes of this evaluation are documented in Table 3. It can be seen from the table that our method outperforms all the baselines. + +# 4.3 Ablation Study and Analysis + +Reconstruction Results on Early Training Stage An advantage of our CRM is that it is easy to train. In fact, we find that CRM starts to show reasonable results at very early stage of training. The results are shown in Fig. 8. The results are good even with barely 280 iterations (only 20 minutes of training). + +Table 1: Quantitative comparison for the geometry quality between our method and baselines for single image to 3D textured mesh generation. We report the metrics of Chamfer Distance, Column IoU and F-score on GSO dataset. + +
MethodChamfer Dist.↓ Vol. IoU↑ F-Sco. (%)↑
One-2-3-45 [29]0.01720.446372.19
SyncDreamer [31]0.01400.390075.74
Wonder3D [32]0.01860.439876.75
Magic123 [41]0.01880.371460.66
TGS [71]0.01720.298265.17
OpenLRM [17,18]0.01680.377463.22
LGM [52]0.01170.468568.69
Ours0.00940.613179.38
+ +Table 2: Quantitative comparison for the texture quality between our method and baselines for single image to 3D textured mesh generation. We report the metric of PSNR, SSIM, LPIPS and Clip [43]-Similarity on GSO dataset. + +
MethodPSNR↑SSIM↑LPIPS↓Clip-Sim↑
One-2-3-45 [29]13.930.80840.262579.83
SyncDreamer [31]14.000.81650.259182.76
Wonder3D [32]13.310.81210.255483.70
Magic123 [41]12.690.79840.244285.16
OpenLRM [17,18]14.300.82940.227684.20
LGM [71]13.280.79460.256085.20
Ours16.220.83810.214387.55
+ +We conjecture that the fast convergence results from the strong geometry prior in our architecture design. + +Training Time of CRM In Fig. 9 we compare the training cost between our method (reconstruction model only) and two baselines, LRM [18] and LGM [52]. We measure the training cost by the training days multiplying the amount of used NVIDIA A100/A800 GPU cards. It can be seen that our model takes much smaller training time than the two baselines. This is because our model utilizes the spatial correspondence between input six orthographic images/CCMs and triplanes, which serves as a strong prior that makes the training easier. + +Importance of Input CCM We examine the importance of the CCMs that are concatenated to the input images. To compare, we train a reconstruction model that takes only the six RGB images as input, without CCMs. The results are shown in Fig. 10. It can be seen that the results of the geometry degrade a lot without CCM input. This is because CCM provides important geometry information for the model, especially when the geometry is complex. + +Table 3: Quantitative comparison between our method and baselines for novel view synthesis of the multi-view diffusion model. We report the metric of PSNR, SSIM and LPIPS on GSO dataset. + +
MethodPSNR↑SSIM↑LPIPS↓
SyncDreamer [31]20.300.78040.2932
Wonder3D [32]23.760.81270.2210
Ours29.360.87210.1354
+ +![](images/5631a7b4a9a4de8da88578f23a96c9c2a6081801c8c582a822d054c81189a900.jpg) +(a) 0 iters (Initialization) + +![](images/ed9850342b38e8734023597fdd40bc18b788b18c1f82275e993a9a15f9cb4b1e.jpg) +(b) 280 iters $(\sim 20$ min training) + +![](images/7f82071071640cd1ea11c97faff85820c3335d8b66ddf33e61fbc189970c6d53.jpg) +(c) 4k iters $(\sim 5$ hours training) + +![](images/b12c8e77990ff262098ce165e36edb196ebcc6dfb9dbd0747fedf1e6927fc518.jpg) +(d) Ground Truth Mesh + +![](images/7a69169d99d1f4e864126c8739ec686448a7d636802dd089a704018c5ec6828a.jpg) +Fig. 9: Training cost comparison. Our model require much less computation cost than other baselines. + +![](images/4f84525782f831285f21ca25454032ed4c91a8f7f355c96a108f402d002df564.jpg) +Fig. 8: Reconstruction results on unseen samples during early stage of training. + +![](images/42a32e64756660ebbcffd2abe9bfffc54a4d6e7639058983810447857084e3b0.jpg) +(a) w/o CCM Input + +![](images/bfbfcdff8787aa365202dce2f181624248a4dc94cab01336d8fbd59207a2c55d.jpg) +Fig. 10: The CCM concatenated to the input images is beneficial for our model. (a) Without providing CCM, the model outputs a geometry which is reasonable, but not very good. (b) The shape reconstructed using our full model with CCM input, with a much better geometry. (c) Ground truth mesh rendered from the same pose. + +![](images/08abccfcb2adfce265eb95dcc235f806e2f74626c4f17ae2c5acb987bb2a075e.jpg) + +![](images/2a9ed6b6c784b4c32e4a2fd3d7b4f9eea934e2d8ff5233f08826588c172e1532.jpg) +(b) w/ CCM Input + +![](images/fe5d1ed6f6d9bd3b09d62e90f1e6e4557e0b3c26f9deb2f7aae3afece14607ef.jpg) + +![](images/23679090a5b355d39a3aa067a1be0e679d45e297ab83cc1c66944c0d93f6e6a7.jpg) +(c) Ground Truth Mesh + +Design of Multi-view Diffusion Here we examine the effectiveness of the design of the multi-view diffusion models. Starting from the baseline that naively + +Table 4: Ablation study on the design of multi-view diffusion on novel view synthesis. We report the metrics of PSNR, SSIM and LPIPS on GSO dataset. + +
MethodPSNR↑ SSIM↑ LPIPS↓
ImageDream (6 view)28.990.85650.1497
+ Zero-SNR29.130.85980.1498
+ Random Resizing29.360.87210.1354
+ Contour Augmentation28.920.86810.1444
+ +![](images/3cbaaf3456ced12b33b53d51aabdd26a26ba0547ad725e472d55ed3ad830fd55.jpg) +(a) Input Image + +![](images/be846a4b9118f82ecddaabd6ae6cf3b61073f55ee599ada869aca97bc4ee4fad.jpg) +(b) Segmentation Result + +![](images/afaa77feb293f97da71e6e6aabfa1855817c73d8bff5b59a60282e105e9906d3.jpg) +Fig. 11: Demonstration of contour augmentation. (a) Given an input image, (b) off-the-shelf segmentation model sometimes provides imperfect results. (c) Without contour augmentation, the predicted backview color is sensitive to the contour. (d) With contour augmentation, the model predicts reasonable result. (e) We demonstrate how we augment the input image during training. + +![](images/0324a6d9884bb8b93e9653c54a7aefa89ac3b7fd2efcde330926b09383942e44.jpg) +(c) w/o Contour Augmentation + +![](images/cff429cc48c431105176294e822c9176b6af397b6db115d0cd3455c6fd8de114.jpg) +(d) w/ Contour Augmentation +(e) Contour Augmented Image + +fine-tunes the pre-trained ImageDream model with 2 additional views, we sequentially add the proposed techniques on the training. We examine the results on a subset of GSO, comparing the similarity of the generated novel view images with the ground truth images using PSNR, SSIM and LPIPS metrics. The results are shown in table 4. It can be seen that both the Zero-SNR trick and random resizing are beneficial. Note that the contour augmentation does not improve the quantitative metrics. However, we find that this trick makes the model more robust to in the wild input images (Fig. 11). + +# 5 Conclusion + +In this work, we present a convolutional reconstruction model (CRM) for creating high-quality 3D models from a single image. Our approach effectively utilizes the spatial relationship between input images and the output triplane, leading to improved textured meshes, with significantly less training cost compared to previous transformer-based methods [18]. The model operates on an end-to-end training basis, directly outputting textured meshes. Overall, our method can produce detailed textured meshes in just 10 seconds. + +Potential Negative Impact. Similar to many other generated models, our CRM may be used to generate malicious or fake 3D contents, which may need additional caution. + +# Acknowledgements + +This work was supported by Natural Science Foundation of China (No.s 62350080, 62076147), Tsinghua Institute for Guo Qiang, and the High Performance Computing Center, Tsinghua University. J.Z is also supported by the New Cornerstone Science Foundation through the XPlorer Prize. + +# References + +1. Bao, F., Nie, S., Xue, K., Cao, Y., Li, C., Su, H., Zhu, J.: All are worth words: A vit backbone for diffusion models. In: CVPR (2023) +2. Brooks, T., Peebles, B., Homes, C., DePue, W., Guo, Y., Jing, L., Schnurr, D., Taylor, J., Luhman, T., Luhman, E., Ng, C., Wang, R., Ramesh, A.: Video generation models as world simulators (2024), https://openai.com/research/video-generation-models-as-world-simulators +3. Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D.M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners (2020) +4. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022) +5. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision. pp. 333-350. Springer (2022) +6. Chen, H., Gu, J., Chen, A., Tian, W., Tu, Z., Liu, L., Su, H.: Single-stage diffusion nef: A unified approach to 3d generation and reconstruction. arXiv preprint arXiv:2304.06714 (2023) +7. Chen, R., Chen, Y., Jiao, N., Jia, K.: *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation*. arXiv preprint arXiv:2303.13873 (2023) +8. Chen, Y., Zhang, C., Yang, X., Cai, Z., Yu, G., Yang, L., Lin, G.: It3d: Improved text-to-3d generation with explicit view synthesis (2023) +9. Chen, Z., Wang, F., Liu, H.: Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585 (2023) +10. Cheng, Y.C., Lee, H.Y., Tuyakov, S., Schwing, A., Gui, L.: SDFusion: Multimodal 3d shape completion, reconstruction, and generation. In: CVPR (2023) +11. Decatur, D., Lang, I., Aberman, K., Hanocka, R.: 3d paintbrush: Local stylization of 3d shapes with cascaded score distillation. arXiv preprint arXiv:2311.09571 (2023) +12. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objaverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. Advances in Neural Information Processing Systems 36 (2024) +13. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objaverse: A universe of annotated 3d objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13142-13153 (2023) + +14. Downs, L., Francis, A., Koenig, N., Kinman, B., Hickman, R., Reymann, K., McHugh, T.B., Vanhoucke, V.: Google scanned objects: A high-quality dataset of 3d scanned household items. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 2553-2560. IEEE (2022) +15. Gao, J., Shen, T., Wang, Z., Chen, W., Yin, K., Li, D., Litany, O., Gojcic, Z., Fidler, S.: Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems 35, 31841-31854 (2022) +16. Gupta, A., Xiong, W., Nie, Y., Jones, I., Oğuz, B.: 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371 (2023) +17. He, Z., Wang, T.: Openlrm: Open-source large reconstruction models. https://github.com/3DTopia/OpenLRM (2023) +18. Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D., Liu, F., Sunkavalli, K., Bui, T., Tan, H.: Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400 (2023) +19. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023) +20. Kim, S., Lee, K., Choi, J.S., Jeong, J., Sohn, K., Shin, J.: Collaborative score distillation for consistent visual synthesis (2023) +21. Li, J., Tan, H., Zhang, K., Xu, Z., Luan, F., Xu, Y., Hong, Y., Sunkavalli, K., Shakhnarovich, G., Bi, S.: Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214 (2023) +22. Li, W., Chen, R., Chen, X., Tan, P.: Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596 (2023) +23. Li, Y., Dou, Y., Shi, Y., Lei, Y., Chen, X., Zhang, Y., Zhou, P., Ni, B.: Focaldreamer: Text-driven 3d editing via focal-fusion assembly (2023) +24. Liang, Y., Yang, X., Lin, J., Li, H., Xu, X., Chen, Y.: Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching (2023) +25. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300–309 (2023) +26. Lin, S., Liu, B., Li, J., Yang, X.: Common diffusion noise schedules and sample steps are flawed. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5404-5411 (2024) +27. Liu, F., Wu, D., Wei, Y., Rao, Y., Duan, Y.: Sherpa3d: Boosting high-fidelity text-to-3d generation via coarse 3d prior (2023) +28. Liu, M., Shi, R., Chen, L., Zhang, Z., Xu, C., Wei, X., Chen, H., Zeng, C., Gu, J., Su, H.: One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion. arXiv preprint arXiv:2311.07885 (2023) +29. Liu, M., Xu, C., Jin, H., Chen, L., Xu, Z., Su, H., et al.: One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928 (2023) +30. Liu, R., Wu, R., Hoorick, B.V., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object (2023) +31. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453 (2023) +32. Long, X., Guo, Y.C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.H., Habermann, M., Theobalt, C., et al.: Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008 (2023) + +33. Lorraine, J., Xie, K., Zeng, X., Lin, C.H., Takikawa, T., Sharp, N., Lin, T.Y., Liu, M.Y., Fidler, S., Lucas, J.: Att3d: Amortized text-to-3d object synthesis (2023) +34. Lu, Y., Zhang, J., Li, S., Fang, T., McKinnon, D., Tsin, Y., Quan, L., Cao, X., Yao, Y.: Direct2.5: Diverse text-to-3d generation via multi-view 2.5d diffusion (2023) +35. Mercier, A., Nakhli, R., Reddy, M., Yasarla, R., Cai, H., Porikli, F., Berger, G.: Hexagen3d: Stablediffusion is just one step away from fast and diverse text-to-3d generation. arXiv preprint arXiv:2401.07727 (2024) +36. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) +37. Nielson, G.M.: Dual marching cubes. In: IEEE visualization 2004. pp. 489-496. IEEE (2004) +38. Peebles, W., Xie, S.: Scalable diffusion models with transformers (2023) +39. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022) +40. Qian, G., Cao, J., Siarohin, A., Kant, Y., Wang, C., Vasilkovsky, M., Lee, H.Y., Fang, Y., Skorokhodov, I., Zhuang, P., Gilitschenski, I., Ren, J., Ghanem, B., Aberman, K., Tulyakov, S.: Atom: Amortized text-to-mesh using 2d diffusion (2024) +41. Qian, G., Mai, J., Hamdi, A., Ren, J., Siarohin, A., Li, B., Lee, H.Y., Skorokhodov, I., Wonka, P., Tulyakov, S., Ghanem, B.: Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors (2023) +42. Qiu, L., Chen, G., Gu, X., Zuo, Q., Xu, M., Wu, Y., Yuan, W., Dong, Z., Bo, L., Han, X.: Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d (2023) +43. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +44. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +45. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) +46. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: Laion-5b: An open large-scale dataset for training next generation image-text models (2022) +47. Shen, T., Munkberg, J., Hasselgren, J., Yin, K., Wang, Z., Chen, W., Gojcic, Z., Fidler, S., Sharp, N., Gao, J.: Flexible isosurface extraction for gradient-based mesh optimization. ACM Transactions on Graphics (TOG) 42(4), 1-16 (2023) +48. Shi, R., Wei, X., Wang, C., Su, H.: Zerorf: Fast sparse view 360 $\{\backslash \mathrm{deg}\}$ reconstruction with zero pretraining. arXiv preprint arXiv:2312.09249 (2023) +49. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023) +50. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) +51. Sun, J., Zhang, B., Shao, R., Wang, L., Liu, W., Xie, Z., Liu, Y.: Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior (2023) + +52. Tang, J., Chen, Z., Chen, X., Wang, T., Zeng, G., Liu, Z.: Lgm: Large multi-view gaussian model for high-resolution 3d content creation (2024) +53. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023) +54. Tochilkin, D., Pankratz, D., Liu, Z., Huang, Z., Letts, A., Li, Y., Liang, D., Laforte, C., Jampani, V., Cao, Y.P.: Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151 (2024) +55. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439 (2023) +56. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need (2023) +57. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation (2022) +58. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction (2023) +59. Wang, P., Shi, Y.: Imaginedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201 (2023) +60. Wang, P., Tan, H., Bi, S., Xu, Y., Luan, F., Sunkavalli, K., Wang, W., Xu, Z., Zhang, K.: Pf-lrm: Pose-free large reconstruction model for joint pose and shape prediction. arXiv preprint arXiv:2311.12024 (2023) +61. Wang, T., Zhang, B., Zhang, T., Gu, S., Bao, J., Baltrusaitis, T., Shen, J., Chen, D., Wen, F., Chen, Q., et al.: Rodin: A generative model for sculpting 3d digital avatars using diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4563-4573 (2023) +62. Wang, X., Wang, Y., Ye, J., Wang, Z., Sun, F., Liu, P., Wang, L., Sun, K., Wang, X., He, B.: Animatabledreamer: Text-guided non-rigid 3d model generation and reconstruction with canonical score distillation +63. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=ppJuFSOAnM +64. Xu, Y., Tan, H., Luan, F., Bi, S., Wang, P., Li, J., Shi, Z., Sunkavalli, K., Wetzstein, G., Xu, Z., et al.: Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model. arXiv preprint arXiv:2311.09217 (2023) +65. Yu, X., Guo, Y.C., Li, Y., Liang, D., Zhang, S.H., Qi, X.: Text-to-3d with classifier score distillation (2023) +66. Zeng, X., Vahdat, A., Williams, F., Gojcic, Z., Litany, O., Fidler, S., Kreis, K.: Lion: Latent point diffusion models for 3d shape generation (2022) +67. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric (2018) +68. Zheng, X.Y., Pan, H., Guo, Y.X., Tong, X., Liu, Y.: Mvd $^2$ : Efficient multiview 3d reconstruction for multiview diffusion (2024) +69. Zheng, X.Y., Pan, H., Wang, P.S., Tong, X., Liu, Y., Shum, H.Y.: Locally attentional sdf diffusion for controllable 3d shape generation. arXiv preprint arXiv:2305.04461 (2023) +70. Zhu, J., Zhuang, P.: Hifa: High-fidelity text-to-3d generation with advanced diffusion guidance (2023) +71. Zou, Z.X., Yu, Z., Guo, Y.C., Li, Y., Liang, D., Cao, Y.P., Zhang, S.H.: Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. arXiv preprint arXiv:2312.09147 (2023) \ No newline at end of file diff --git a/2024/CRM_ Single Image to 3D Textured Mesh with Convolutional Reconstruction Model/images.zip b/2024/CRM_ Single Image to 3D Textured Mesh with Convolutional Reconstruction Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f250965b98d695328d1e6f5ea987455da683ccce --- /dev/null +++ b/2024/CRM_ Single Image to 3D Textured Mesh with Convolutional Reconstruction Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b3473f60af0a651e5ecce46cd2399a86c3c6d36c877cdccc193fd3a64ecca985 +size 723627 diff --git a/2024/CRM_ Single Image to 3D Textured Mesh with Convolutional Reconstruction Model/layout.json b/2024/CRM_ Single Image to 3D Textured Mesh with Convolutional Reconstruction Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f2b511e8287faf50cbc4642c65a25b5e38c6c83a --- /dev/null +++ b/2024/CRM_ Single Image to 3D Textured Mesh with Convolutional Reconstruction Model/layout.json @@ -0,0 +1,10545 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 142, + 111, + 473, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 111, + 473, + 146 + ], + "spans": [ + { + "bbox": [ + 142, + 111, + 473, + 146 + ], + "type": "text", + "content": "CRM: Single Image to 3D Textured Mesh with Convolutional Reconstruction Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "spans": [ + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "text", + "content": "Zhengyi Wang" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "text", + "content": ", Yikai Wang" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "text", + "content": ", Yifei Chen" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "text", + "content": ", Chendong Xiang" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "text", + "content": ", Shuo Chen" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "text", + "content": ", Dajiang Yu" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "text", + "content": ", Chongxuan Li" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "text", + "content": ", Hang Su" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "text", + "content": " and Jun Zhu" + }, + { + "bbox": [ + 146, + 167, + 466, + 194 + ], + "type": "inline_equation", + "content": "^{\\star 1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 139, + 201, + 473, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 201, + 473, + 224 + ], + "spans": [ + { + "bbox": [ + 139, + 201, + 473, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 139, + 201, + 473, + 224 + ], + "type": "text", + "content": " Dept. of Comp. Sci. & Tech., BNrist Center, Tsinghua-Bosch Joint ML Center, Tsinghua University;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 248, + 224, + 364, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 224, + 364, + 236 + ], + "spans": [ + { + "bbox": [ + 248, + 224, + 364, + 236 + ], + "type": "text", + "content": "2 ShengShu, Beijing, China;" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 235, + 477, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 235, + 477, + 258 + ], + "spans": [ + { + "bbox": [ + 136, + 235, + 477, + 258 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 136, + 235, + 477, + 258 + ], + "type": "text", + "content": " Gaoling School of Artificial Intelligence, Renmin University of China, Beijing Key Laboratory of Big Data Management and Analysis Methods, Beijing, China." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 156, + 308, + 192, + 351 + ], + "blocks": [ + { + "bbox": [ + 147, + 289, + 196, + 301 + ], + "lines": [ + { + "bbox": [ + 147, + 289, + 196, + 301 + ], + "spans": [ + { + "bbox": [ + 147, + 289, + 196, + 301 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 156, + 308, + 192, + 351 + ], + "lines": [ + { + "bbox": [ + 156, + 308, + 192, + 351 + ], + "spans": [ + { + "bbox": [ + 156, + 308, + 192, + 351 + ], + "type": "image", + "image_path": "179d3eff26b50d33ee74c72b3ffbf2ba60d92125c82f0cba79fa432c8c9aaf6e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 160, + 356, + 186, + 396 + ], + "blocks": [ + { + "bbox": [ + 160, + 356, + 186, + 396 + ], + "lines": [ + { + "bbox": [ + 160, + 356, + 186, + 396 + ], + "spans": [ + { + "bbox": [ + 160, + 356, + 186, + 396 + ], + "type": "image", + "image_path": "df48604738fb65a7d7389664d381d7ac4cdc7c8a18a0d60a735423725ee47550.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 158, + 399, + 184, + 437 + ], + "blocks": [ + { + "bbox": [ + 158, + 399, + 184, + 437 + ], + "lines": [ + { + "bbox": [ + 158, + 399, + 184, + 437 + ], + "spans": [ + { + "bbox": [ + 158, + 399, + 184, + 437 + ], + "type": "image", + "image_path": "798007dac4cff8fbbb561379e256148586715340dfc6f87669e84557b791aa4b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 151, + 445, + 193, + 480 + ], + "blocks": [ + { + "bbox": [ + 151, + 445, + 193, + 480 + ], + "lines": [ + { + "bbox": [ + 151, + 445, + 193, + 480 + ], + "spans": [ + { + "bbox": [ + 151, + 445, + 193, + 480 + ], + "type": "image", + "image_path": "591f7aa234c38dffcd2fbb7179ba047e740ecbd855bd15b4d4539f817afe9bb8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 155, + 486, + 189, + 524 + ], + "blocks": [ + { + "bbox": [ + 155, + 486, + 189, + 524 + ], + "lines": [ + { + "bbox": [ + 155, + 486, + 189, + 524 + ], + "spans": [ + { + "bbox": [ + 155, + 486, + 189, + 524 + ], + "type": "image", + "image_path": "76c809462deddf1df94da004a029e25690492d60bf777d8ecb1fb570ed58d38b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 535, + 481, + 547 + ], + "lines": [ + { + "bbox": [ + 132, + 535, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 535, + 481, + 547 + ], + "type": "text", + "content": "Fig. 1: CRM generates high-fidelity textured mesh from single image in 10 seconds." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 214, + 309, + 253, + 350 + ], + "blocks": [ + { + "bbox": [ + 214, + 309, + 253, + 350 + ], + "lines": [ + { + "bbox": [ + 214, + 309, + 253, + 350 + ], + "spans": [ + { + "bbox": [ + 214, + 309, + 253, + 350 + ], + "type": "image", + "image_path": "e8682e31401391693f271d493b90e151c2237e0622046c56b19359da91421dfe.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 220, + 354, + 252, + 397 + ], + "blocks": [ + { + "bbox": [ + 220, + 354, + 252, + 397 + ], + "lines": [ + { + "bbox": [ + 220, + 354, + 252, + 397 + ], + "spans": [ + { + "bbox": [ + 220, + 354, + 252, + 397 + ], + "type": "image", + "image_path": "f699a043da85542a61278aac2dea1ac2c32e84720e953c62cf419b5ca66ab373.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 217, + 399, + 253, + 439 + ], + "blocks": [ + { + "bbox": [ + 217, + 399, + 253, + 439 + ], + "lines": [ + { + "bbox": [ + 217, + 399, + 253, + 439 + ], + "spans": [ + { + "bbox": [ + 217, + 399, + 253, + 439 + ], + "type": "image", + "image_path": "f2f2a3ebc059277458dcee7aa5e010e10c9d74ad98bfed40b9f9ec17d4f422ac.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 211, + 442, + 253, + 486 + ], + "blocks": [ + { + "bbox": [ + 211, + 442, + 253, + 486 + ], + "lines": [ + { + "bbox": [ + 211, + 442, + 253, + 486 + ], + "spans": [ + { + "bbox": [ + 211, + 442, + 253, + 486 + ], + "type": "image", + "image_path": "f6ddc99930a0d96d2ecd56c02476009fbbc226d08f5a4aabb8b7a3ed3c5eef83.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 217, + 487, + 248, + 526 + ], + "blocks": [ + { + "bbox": [ + 217, + 487, + 248, + 526 + ], + "lines": [ + { + "bbox": [ + 217, + 487, + 248, + 526 + ], + "spans": [ + { + "bbox": [ + 217, + 487, + 248, + 526 + ], + "type": "image", + "image_path": "e00f04b10c425cc88c3872edf2f0db16ce42e17e07d14d23c543be825541dd16.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 261, + 309, + 296, + 350 + ], + "blocks": [ + { + "bbox": [ + 291, + 289, + 394, + 300 + ], + "lines": [ + { + "bbox": [ + 291, + 289, + 394, + 300 + ], + "spans": [ + { + "bbox": [ + 291, + 289, + 394, + 300 + ], + "type": "text", + "content": "Generated Textured Mesh" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 261, + 309, + 296, + 350 + ], + "lines": [ + { + "bbox": [ + 261, + 309, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 261, + 309, + 296, + 350 + ], + "type": "image", + "image_path": "6d6cddfd4e2a9d79488994b74a58427941189f98e50d01485658b8c0e722c5b1.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 263, + 354, + 296, + 397 + ], + "blocks": [ + { + "bbox": [ + 263, + 354, + 296, + 397 + ], + "lines": [ + { + "bbox": [ + 263, + 354, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 263, + 354, + 296, + 397 + ], + "type": "image", + "image_path": "64cdf860448bffcc6f38c258baa22100d0348a80fcc1977095c0602734c3ee36.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 260, + 399, + 296, + 439 + ], + "blocks": [ + { + "bbox": [ + 260, + 399, + 296, + 439 + ], + "lines": [ + { + "bbox": [ + 260, + 399, + 296, + 439 + ], + "spans": [ + { + "bbox": [ + 260, + 399, + 296, + 439 + ], + "type": "image", + "image_path": "bb52f3d47de4c2c2908ab641a8a694257ed1204a2c85fb4431cb7cd3d886c947.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 257, + 442, + 296, + 486 + ], + "blocks": [ + { + "bbox": [ + 257, + 442, + 296, + 486 + ], + "lines": [ + { + "bbox": [ + 257, + 442, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 257, + 442, + 296, + 486 + ], + "type": "image", + "image_path": "ca39cd3d253d95c4592937f21037fce576f4e2eda993f7891737929ee6a20172.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 260, + 487, + 291, + 526 + ], + "blocks": [ + { + "bbox": [ + 260, + 487, + 291, + 526 + ], + "lines": [ + { + "bbox": [ + 260, + 487, + 291, + 526 + ], + "spans": [ + { + "bbox": [ + 260, + 487, + 291, + 526 + ], + "type": "image", + "image_path": "c4c87c5e1ef94b66d4389518a8081d10be6e3505cee42f7a5ddb143d9aba1697.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 300, + 309, + 384, + 350 + ], + "blocks": [ + { + "bbox": [ + 300, + 309, + 384, + 350 + ], + "lines": [ + { + "bbox": [ + 300, + 309, + 384, + 350 + ], + "spans": [ + { + "bbox": [ + 300, + 309, + 384, + 350 + ], + "type": "image", + "image_path": "5e02daf5ff5a6f19fd1a7212f8dd098f1209e7268dd7adf6f6ce286974153efd.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 300, + 352, + 384, + 397 + ], + "blocks": [ + { + "bbox": [ + 300, + 352, + 384, + 397 + ], + "lines": [ + { + "bbox": [ + 300, + 352, + 384, + 397 + ], + "spans": [ + { + "bbox": [ + 300, + 352, + 384, + 397 + ], + "type": "image", + "image_path": "6d45e051e2e78d6762db3e8fd7c444d362a5a86a70207ee6a2779832ae41ca3e.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 299, + 399, + 384, + 439 + ], + "blocks": [ + { + "bbox": [ + 299, + 399, + 384, + 439 + ], + "lines": [ + { + "bbox": [ + 299, + 399, + 384, + 439 + ], + "spans": [ + { + "bbox": [ + 299, + 399, + 384, + 439 + ], + "type": "image", + "image_path": "4b730102c951ebaf3517452d75773288fce30dc434ee92d0d9b0babf14b0e8ce.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 298, + 442, + 384, + 486 + ], + "blocks": [ + { + "bbox": [ + 298, + 442, + 384, + 486 + ], + "lines": [ + { + "bbox": [ + 298, + 442, + 384, + 486 + ], + "spans": [ + { + "bbox": [ + 298, + 442, + 384, + 486 + ], + "type": "image", + "image_path": "aab8a59f0ea80db81f3403de012c53c6c843aa70ef619e39adde9423a596e774.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 303, + 487, + 384, + 525 + ], + "blocks": [ + { + "bbox": [ + 303, + 487, + 384, + 525 + ], + "lines": [ + { + "bbox": [ + 303, + 487, + 384, + 525 + ], + "spans": [ + { + "bbox": [ + 303, + 487, + 384, + 525 + ], + "type": "image", + "image_path": "d837c50dcf46eace8551e3fea5f9a8e965ee41e63095ecd4c70cd71680ce0323.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 392, + 309, + 428, + 350 + ], + "blocks": [ + { + "bbox": [ + 392, + 309, + 428, + 350 + ], + "lines": [ + { + "bbox": [ + 392, + 309, + 428, + 350 + ], + "spans": [ + { + "bbox": [ + 392, + 309, + 428, + 350 + ], + "type": "image", + "image_path": "4ebd870be2e7710cf158b9ee04d44e8ec6c603917c26408e5611c098ad516dbd.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 390, + 353, + 428, + 397 + ], + "blocks": [ + { + "bbox": [ + 390, + 353, + 428, + 397 + ], + "lines": [ + { + "bbox": [ + 390, + 353, + 428, + 397 + ], + "spans": [ + { + "bbox": [ + 390, + 353, + 428, + 397 + ], + "type": "image", + "image_path": "8e6a6e2902f982c95b7bcea635b0d732db15a2ff96451c4a760714b68325a3e9.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 395, + 399, + 424, + 439 + ], + "blocks": [ + { + "bbox": [ + 395, + 399, + 424, + 439 + ], + "lines": [ + { + "bbox": [ + 395, + 399, + 424, + 439 + ], + "spans": [ + { + "bbox": [ + 395, + 399, + 424, + 439 + ], + "type": "image", + "image_path": "50b9c42502a45048091168e9e8c0b68fb4b9a9bcf6310053e702cc2b3776b8ef.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 386, + 442, + 427, + 486 + ], + "blocks": [ + { + "bbox": [ + 386, + 442, + 427, + 486 + ], + "lines": [ + { + "bbox": [ + 386, + 442, + 427, + 486 + ], + "spans": [ + { + "bbox": [ + 386, + 442, + 427, + 486 + ], + "type": "image", + "image_path": "baa1ba49c0e39769ee921893dc80b48a3dadf28721b7b3d0fbfbdf0932c52a00.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 398, + 491, + 421, + 527 + ], + "blocks": [ + { + "bbox": [ + 398, + 491, + 421, + 527 + ], + "lines": [ + { + "bbox": [ + 398, + 491, + 421, + 527 + ], + "spans": [ + { + "bbox": [ + 398, + 491, + 421, + 527 + ], + "type": "image", + "image_path": "76802b0a9327190810c277f429bbc10a3a0a6fa9504d70a174e830cba6538c46.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 433, + 309, + 469, + 350 + ], + "blocks": [ + { + "bbox": [ + 433, + 309, + 469, + 350 + ], + "lines": [ + { + "bbox": [ + 433, + 309, + 469, + 350 + ], + "spans": [ + { + "bbox": [ + 433, + 309, + 469, + 350 + ], + "type": "image", + "image_path": "76dc3a4df519da49d09aad4c668e6491fa94e96fcef91587514f86d7e0f46875.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 432, + 353, + 470, + 397 + ], + "blocks": [ + { + "bbox": [ + 432, + 353, + 470, + 397 + ], + "lines": [ + { + "bbox": [ + 432, + 353, + 470, + 397 + ], + "spans": [ + { + "bbox": [ + 432, + 353, + 470, + 397 + ], + "type": "image", + "image_path": "c3ca2a83f79ee6844dd382a93e472905415ec64b976460977b6943258594e8e1.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 440, + 401, + 466, + 440 + ], + "blocks": [ + { + "bbox": [ + 440, + 401, + 466, + 440 + ], + "lines": [ + { + "bbox": [ + 440, + 401, + 466, + 440 + ], + "spans": [ + { + "bbox": [ + 440, + 401, + 466, + 440 + ], + "type": "image", + "image_path": "e3bd3ffc3ddef8ec997dc4332d5d7c40c9c9c9873e34d126198e327b919277dd.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 430, + 442, + 470, + 486 + ], + "blocks": [ + { + "bbox": [ + 430, + 442, + 470, + 486 + ], + "lines": [ + { + "bbox": [ + 430, + 442, + 470, + 486 + ], + "spans": [ + { + "bbox": [ + 430, + 442, + 470, + 486 + ], + "type": "image", + "image_path": "fb32a0150f49ab7c1a3d53b1c5c89c8da46aa86fdbc43d80bceaaa985afff56b.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 441, + 491, + 464, + 527 + ], + "blocks": [ + { + "bbox": [ + 441, + 491, + 464, + 527 + ], + "lines": [ + { + "bbox": [ + 441, + 491, + 464, + 527 + ], + "spans": [ + { + "bbox": [ + 441, + 491, + 464, + 527 + ], + "type": "image", + "image_path": "416d4c6de8015f60630874f0a601921f3a3cf048ff8f50413205256586da69fd.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "bbox": [ + 159, + 591, + 453, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 591, + 453, + 647 + ], + "spans": [ + { + "bbox": [ + 159, + 591, + 453, + 647 + ], + "type": "text", + "content": "Abstract. Feed-forward 3D generative models like the Large Reconstruction Model (LRM) [18] have demonstrated exceptional generation speed. However, the transformer-based methods do not leverage the geometric priors of the triplane component in their architecture, often leading to sub-optimal quality given the limited size of 3D data and" + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 654, + 254, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 254, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 254, + 665 + ], + "type": "text", + "content": "* The Corresponding author." + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 117, + 455, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 117, + 455, + 260 + ], + "spans": [ + { + "bbox": [ + 159, + 117, + 455, + 260 + ], + "type": "text", + "content": "slow training. In this work, we present the Convolutional Reconstruction Model (CRM), a high-fidelity feed-forward single image-to-3D generative model. Recognizing the limitations posed by sparse 3D data, we highlight the necessity of integrating geometric priors into network design. CRM builds on the key observation that the visualization of triplane exhibits spatial correspondence of six orthographic images. First, it generates six orthographic view images from a single input image, then feeds these images into a convolutional U-Net, leveraging its strong pixel-level alignment capabilities and significant bandwidth to create a high-resolution triplane. CRM further employs Flexicubes as geometric representation, facilitating direct end-to-end optimization on textured meshes. Overall, our model delivers a high-fidelity textured mesh from an image in just 10 seconds, without any test-time optimization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 270, + 423, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 270, + 423, + 281 + ], + "spans": [ + { + "bbox": [ + 160, + 270, + 423, + 281 + ], + "type": "text", + "content": "Keywords: 3D Generation " + }, + { + "bbox": [ + 160, + 270, + 423, + 281 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 270, + 423, + 281 + ], + "type": "text", + "content": " Textured Mesh " + }, + { + "bbox": [ + 160, + 270, + 423, + 281 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 270, + 423, + 281 + ], + "type": "text", + "content": " Diffusion Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 304, + 230, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 230, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 230, + 316 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 330, + 481, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 481, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 481, + 437 + ], + "type": "text", + "content": "In recent years, generative models have witnessed significant advancements, largely attributed to the fast growth in data size. Transformers [56], in particular, have achieved high-performance results across various domains including language [3], image [1,38] and video generation [2]. However, the domain of 3D generation presents unique challenges. Unlike the abundance of other modal's data, 3D data is comparatively scarce. The creation of 3D data requires specialized expertise and considerable time, leading to a situation where the largest 3D datasets, namely Objaverse [12,13], only contain millions of 3D content, much smaller than image datasets like Laion [46] which contains 5 billion images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 481, + 533 + ], + "type": "text", + "content": "Despite this, recent developments have introduced some transformer-based methods [18, 21, 60, 64, 71] like LRM [18] for creating 3D content from single or multi-view images in a feed-forward manner. Among these models, the triplane has emerged as a popular component due to its efficiency in generating high-resolution 3D results with minimal memory cost. However, reliance on transformer-based networks for generating triplane patches has not utilized the geometric priors inherent to the triplane concept, leading to sub-optimal results in terms of quality and fidelity, and long training time." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "To address the above challenges, this paper presents a new Convolutional Reconstruction Model (CRM) with high generation quality as well as fast training. Given the limited amount of 3D contents, CRM builds on a key hypothesis that it is beneficial to explore geometric priors in architecture design. Namely, we observe from the visualization of triplane [5,6,48] that triplane exhibits spatial correspondence of input six orthographic images, as shown in Fig. 2. The silhouette and texture of the input images have a natural alignment with the triplane structure. This motivates us to (1) use six orthographic images as input images to reconstruct 3D contents, which align well as the triplane feature, instead of other arbitrarily chosen poses; (2) use a U-Net convolutional network to map the input images to a rolled-out triplane by exploring the strong pixel-level alignment" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 176 + ], + "type": "text", + "content": "between the input and output. Furthermore, the significant bandwidth capacity of our U-Net enables a direct transformation of the six orthographic images into the triplane, yielding highly detailed outcomes. Besides, we also add Canonical Coordinate Map (CCM) to the reconstruction network, a novel addition that enriches the model's understanding of spatial relations and geometry." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 481, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 481, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 481, + 272 + ], + "type": "text", + "content": "For the task of 3D generation from a single image, as the six orthographic images and CCMs are not directly available, we train a multi-view diffusion model conditioned on the input image to generate the six orthographic images and another diffusion model to generate the CCMs conditioned on the generated six orthographic images. Both diffusion models are trained on a filtered version of the Objaverse dataset [13]. To further enhance quality and robustness, we implement training improvements for the multi-view diffusion models, including Zero-SNR [26], random resizing, and contour augmentation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 272, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 392 + ], + "type": "text", + "content": "Finally, as directly optimizing high-quality textured meshes is challenging, we adopt Flexicubes [47] as the geometry representation to falcitate gradient-based mesh optimization. This is unlike previous works [18, 71] that use alternative 3D representation like NeRF [36] or Gaussian Splatting [19]. Such methods often involve extra procedure steps to obtain textured meshes [52], although they can produce detailed visualizations. With our designs, we are able to train CRM with textured mesh as the final output in an end-to-end manner, and our approach has a more straightforward inference pipeline and better mesh quality. Overall, our method can generate high-fidelity textured mesh within 10 seconds, as shown in Fig. 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 408, + 242, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 242, + 420 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 242, + 420 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 432, + 343, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 432, + 343, + 443 + ], + "spans": [ + { + "bbox": [ + 131, + 432, + 343, + 443 + ], + "type": "text", + "content": "2.1 Score Distillation for 3D Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 450, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 606 + ], + "type": "text", + "content": "DreamFusion [39] proposes a technique called Score Distillation Sampling (SDS) (also known as Score Jacobian Chaining [57]). It utilizes large scale image diffusion models [44,45] to iteratively refine 3D models to align with specific prompts or images. Thus it can generate 3D content without training on 3D dataset. Along this line, ProlificDreamer [63] proposes Variational Score Distillation (VSD), a principled variational framework which greatly mitigates the over-saturation problems in SDS and improves the diversity. Zero123 [30], MVDream [49], ImageDream [59] and many others [22,41,42] further improve the results and mitigate the multi-face problems using diffusion models fine-tuned on 3D data. [33,40] explore amortized score distillation. Many other works [7-9,11,20,23-25,27,51,53,55,62,65,70] improve the results a lot, in either speed or quality. However, methods based on score distillation usually take from minutes to hours to generate single object, which is computationally expensive." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 623, + 407, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 623, + 407, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 623, + 407, + 635 + ], + "type": "text", + "content": "2.2 3D Generation with Sparse View Reconstruction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Several approaches aim to generate multi-view consistent images and then create 3D contents using sparse views reconstruction. For example, SyncDreamer [31]" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 297, + 91, + 447, + 101 + ], + "type": "text", + "content": "Convolutional Reconstruction Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 480, + 287 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 480, + 287 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 480, + 287 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 480, + 287 + ], + "type": "image", + "image_path": "56c9a154283ec62e3d80cc0acf3f444ab7884443cddd051d3f7966e931aa5396.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 295, + 482, + 351 + ], + "lines": [ + { + "bbox": [ + 130, + 295, + 482, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 482, + 351 + ], + "type": "text", + "content": "Fig. 2: One of our key motivation is that the triplane shares a strong spatial alignment with the input six orthographic images. (a) The six orthographic images of a input shape. (b) The six orthographic CCMs. (c) The triplane (mean value of all channels) output by our U-Net, which spatially aligns with the input images. (d) The textured mesh output by our convolutional reconstruction model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 380, + 491, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 491, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 491, + 443 + ], + "type": "text", + "content": "generates multi-view consistent images and then uses NeuS [58] for reconstruction. Wonder3D [32] improves the results with cross-domain diffusion. Direct2.5 [34] improves the results with 2.5D diffusion. However, one common issue of these methods is that they need test-time optimization for reconstruction with sparse views, which may lead to extra computing and compromise the final quality" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 474, + 345, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 474, + 345, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 474, + 345, + 487 + ], + "type": "text", + "content": "2.3 Feed-forward 3D Generative Models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Some works try to generate 3D objects using a feed forward model [4, 10, 15, 16, 66, 69]. Feed-forward methods demonstrate significantly faster generation speeds compared to the two types of methods mentioned above. Recently there are some works trained on larger 3D dataset Objaverse [13]. One-2-3-45 [29] generates multi-view images and then feed the images into a network to get the 3D object. LRM series [18, 21, 60, 64] improve the quality of generated results with a transformer-based architecture. TGS [71] and LGM [52] use Gaussian Splitting [19] as the geometry representation. There are also many other works [28, 35, 54, 68] that improve the results with different techniques. Despite these advancements, there remains room for improvement in the network architecture or geometry representation. Our approach utilizes a network with a strategically designed architecture and an end-to-end training approach producing meshes directly as the final output." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 115, + 479, + 205 + ], + "blocks": [ + { + "bbox": [ + 136, + 115, + 479, + 205 + ], + "lines": [ + { + "bbox": [ + 136, + 115, + 479, + 205 + ], + "spans": [ + { + "bbox": [ + 136, + 115, + 479, + 205 + ], + "type": "image", + "image_path": "940ce2ff2c999b5c6d44e215b2a5fa46a46dd256259cc3a5012d957982f2a0b2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 215, + 482, + 282 + ], + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 282 + ], + "type": "text", + "content": "Fig. 3: Overall pipeline of our method. The input image is fed into a multi-view image diffusion model to generate six orthographic images. Then another diffusion model is used to generate the CCMs conditioned on the six images. The six images along with the CCMs are send into CRM to reconstruct the final textured mesh. The whole inference process takes around 10 seconds on an A800 GPU. *The 4 seconds include the U-Net forward (less than 0.1s), querying surface points for UV texture and file I/O." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 305, + 202, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 305, + 202, + 317 + ], + "spans": [ + { + "bbox": [ + 132, + 305, + 202, + 317 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 334, + 482, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 482, + 394 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 482, + 394 + ], + "type": "text", + "content": "In this section, we illustrate the detailed design of our method (shown in Fig. 3). Given a single input image, our model first utilizes multi-view diffusion models (Sec. 3.1) to generate six orthographic images and the canonical coordinates maps (CCMs). Then we develop our convolutional reconstruction model (CRM, Sec. 3.2) to reconstruct 3D textured mesh from the images and CCMs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 415, + 301, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 415, + 301, + 426 + ], + "spans": [ + { + "bbox": [ + 131, + 415, + 301, + 426 + ], + "type": "text", + "content": "3.1 Multi-view Diffusion Model" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 569 + ], + "type": "text", + "content": "We first explain the design of the multi-view diffusion model to generate six orthographic view images from a single input image. Instead of training from scratch, which is typically extremely expensive, we initialize the diffusion models using the checkpoint of ImageDream [59], a high-performance diffusion model for the task of multi-view images generation from a single image. The original ImageDream supports 4 views generation. We expand it to include 6 views by adding two more perspectives (up and down). We use another diffusion model which is conditioned on the generated six views to generate the canonical coordinate map. The conditional RGB image is concatenated with the noisy canonical coordinate map. It is also initialized from ImageDream checkpoint. Both the diffusion models are fine-tuned on the Objaverse [13] dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "To further improve the quality and robustness of our results, we introduce several enhancements: (1) Zero-SNR Training. We use the zero-SNR trick as mentioned in [26]. This can alleviate the problem resulting from the discrepancy between the initial Gaussian noise during sampling and the noisiest training sample. (2) Random Resizing. A naive implementation would make the model tends to generate objects that occupy the entire image. To mitigate this, we randomly resize the objects when training. (3) Contour Augmentation. We find that the model tends to predict the backview color largely relying on the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 297, + 91, + 447, + 101 + ], + "type": "text", + "content": "Convolutional Reconstruction Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "contour of the input view. To make the model insensitive to the contour, we randomly change the contour color during training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 158, + 346, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 158, + 346, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 158, + 346, + 170 + ], + "type": "text", + "content": "3.2 Convolutional Reconstruction Model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 178, + 481, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 178, + 481, + 275 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 481, + 275 + ], + "type": "text", + "content": "We now move to introduce the detailed architecture of the convolutional reconstruction model (CRM). As outlined in Fig. 4, given the input six images and CCMs, a convolutional U-Net is used to map the input images along with the CCMs to a rolled-out triplane. Then the rolled-out triplane is reshaped into the triplane. Small multi-layer perceptions (MLPs) are used to decode the triplane features into SDF values, texture color and Flexicubes parameters. Lastly, these values are used to get texture mesh by dual marching cubes. Below, we explain the key components of CRM in detail." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 293, + 481, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 377 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 377 + ], + "type": "text", + "content": "Triplane Representation We choose triplane as the 3D representation, because it can achieve high resolution 3D results with 2D computation consumption. It projects each query grid cell to axis-aligned orthogonal planes " + }, + { + "bbox": [ + 132, + 293, + 481, + 377 + ], + "type": "inline_equation", + "content": "(xy, xz," + }, + { + "bbox": [ + 132, + 293, + 481, + 377 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 293, + 481, + 377 + ], + "type": "inline_equation", + "content": "yz" + }, + { + "bbox": [ + 132, + 293, + 481, + 377 + ], + "type": "text", + "content": " planes) and then aggregates the feature from each planes. Then the feature is decoded by 3 tiny MLPs with 2 hidden layers to get the SDF values along with deformation, color and Flexicubes weights, respectively. Further, to avoid the unnecessary entanglements of different planes, we use rolled-out triplane [61]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 395, + 481, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 395, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 395, + 481, + 479 + ], + "type": "text", + "content": "Canonical Coordinates Map (CCM) We also add CCM as input [22], which contains extra geometry information. This is different from previous works, which typically use pure RGB images as the input to predict the 3D object [18]. Using pure RGB images make it extremely hard to predict the correct geometry, and sometime the geometry degrades (details in Sec. 4.3). Formally, CCM is the coordinates of each point in canonical space. It contains 3 channels whose values are within " + }, + { + "bbox": [ + 132, + 395, + 481, + 479 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 132, + 395, + 481, + 479 + ], + "type": "text", + "content": ", representing the coordinates in the canonical space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 497, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 497, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 497, + 481, + 605 + ], + "type": "text", + "content": "UNet-based Convolutional Network Our key insight is that the triplane is spatially aligned with the input six orthographic images and CCMs, as shown in Fig. 2. To match the rolled-out triplane, the six images and CCMs are arranged in a similar way. We render the six images and CCMs at a resolution of " + }, + { + "bbox": [ + 132, + 497, + 481, + 605 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 132, + 497, + 481, + 605 + ], + "type": "text", + "content": ". They are split into two groups, with each group holding three images. These images in the four groups are then combined to create four larger images, each with a resolution of " + }, + { + "bbox": [ + 132, + 497, + 481, + 605 + ], + "type": "inline_equation", + "content": "256 \\times 768" + }, + { + "bbox": [ + 132, + 497, + 481, + 605 + ], + "type": "text", + "content": ", allowing for spatial alignment. By concatenating these four groups, we form a 12-channel input. Next, a convolutional U-Net processes this input to produce the output triplane." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 606, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 606, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 606, + 481, + 665 + ], + "type": "text", + "content": "Compared to transformer-based methods [18, 21, 64, 71], our U-shape design has a larger bandwidth in preserving the input information, leading to highly detailed triplane features and finally elaborate textured meshes. Moreover, the convolutional network fully utilizes the geometry prior of the spatial correspondence of triplanes and input six orthographic images, which greatly fasten the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 116, + 468, + 288 + ], + "blocks": [ + { + "bbox": [ + 137, + 116, + 468, + 288 + ], + "lines": [ + { + "bbox": [ + 137, + 116, + 468, + 288 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 468, + 288 + ], + "type": "image", + "image_path": "d4214eb88ff384be362cfc0c4e72a8822208ce107a45485ce0c668c79f6ad6b9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 482, + 361 + ], + "type": "text", + "content": "Fig. 4: Architecture along with training pipeline of CRM. We render the 3D mesh into six orthographic images and CCMs. Then the images and CCMs are concatenated and fed into the U-Net. The output triplane is decoded by small MLP networks to form the feature grid of Flexicubes, then textured mesh is get by dual marching cubes. During training, we render the color images, depth maps and masks from GT mesh and reconstructed mesh for supervision." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 386, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 472 + ], + "type": "text", + "content": "convergence and stabilize the training. Our model can get reasonable reconstruction results at very early stage of training (around 20 minutes of training from scratch). Also, our model can be trained with a much smaller batch size 32 (compared to transformer-based LRM that uses a batch size of 1024), which makes that all of our experiments can be conducted on an 8-GPU-cards machine. The overall training cost of our reconstruction model is only " + }, + { + "bbox": [ + 130, + 386, + 482, + 472 + ], + "type": "inline_equation", + "content": "1/8" + }, + { + "bbox": [ + 130, + 386, + 482, + 472 + ], + "type": "text", + "content": " than LRM. More details are shown in the experiments (see Sec. 4.1)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 490, + 482, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 610 + ], + "type": "text", + "content": "Flexicubes Geometry Previous generic 3D generation methods mostly adopt NeRF [36] or Gaussian splatting [19] as the geometry representation, which relies on extra procedures like Marching Cubes (MC) to extract the iso-surface, suffering from topological ambiguities and struggling to represent high-fidelity geometric details. In this work, we use Flexicubes [47] as our geometry representation. It can get meshes from the features on the grid by dual marching cubes [37] during training. The features include SDF values, deformation and weights. The texture is obtained by querying the color at the surface. Flexicubes enables us to train our reconstruction model with textured mesh as the final output in an end-to-end manner." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Loss Function Finally, to train our CRM model, we use a combination of MSE loss " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MSE}}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " and LPIPS loss [67] " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LPIPS}}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " for the rendered images for texture, similar as LRM [18]. To further enhance the geometry, we also include the depth map" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 297, + 91, + 447, + 101 + ], + "type": "text", + "content": "Convolutional Reconstruction Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 116, + 480, + 398 + ], + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 398 + ], + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 398 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 398 + ], + "type": "image", + "image_path": "800904ce077e7454446324a5ce216fd08d3262868c689bdda543685ba5769ce6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 409, + 482, + 433 + ], + "lines": [ + { + "bbox": [ + 130, + 409, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 482, + 433 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparison with baselines. Our models generate high-fidelity results with better geometry and texture." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 454, + 387, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 454, + 387, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 454, + 387, + 467 + ], + "type": "text", + "content": "and mask for supervision [16]. The overall loss function is" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 164, + 474, + 481, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 474, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 164, + 474, + 481, + 503 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {\\mathrm {M S E}} (\\boldsymbol {x}, \\boldsymbol {x} ^ {\\mathrm {G T}}) + \\lambda_ {\\mathrm {L P I P S}} \\mathcal {L} _ {\\mathrm {L P I P S}} (\\boldsymbol {x}, \\boldsymbol {x} ^ {\\mathrm {G T}}) \\\\ + \\lambda_ {\\text {d e p t h}} \\mathcal {L} _ {\\text {M S E}} \\left(\\boldsymbol {x} _ {\\text {d e p t h}}, \\boldsymbol {x} _ {\\text {d e p t h}} ^ {\\mathrm {G T}}\\right) + \\lambda_ {\\text {m a s k}} \\mathcal {L} _ {\\text {M S E}} \\left(\\boldsymbol {x} _ {\\text {m a s k}}, \\boldsymbol {x} _ {\\text {m a s k}} ^ {\\mathrm {G T}}\\right) + \\lambda_ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}}, \\tag {1} \\\\ \\end{array}", + "image_path": "ef3adec233f97559fd490b9921791963eb921d942043accef5d8e18ad2081fe5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\mathrm{depth}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\mathrm{sil}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "text", + "content": " represent the RGB image, depth map and mask rendered from the reconstruction textured mesh, respectively. And " + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\pmb{x}^{\\mathrm{GT}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\mathrm{depth}}^{\\mathrm{GT}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\mathrm{mask}}^{\\mathrm{GT}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "text", + "content": " are rendered from the ground truth textured mesh. " + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "text", + "content": " is the mesh-quality regularizers introduced in Flexicubes [47]. " + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{LPIPS}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{depth}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{mask}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{reg}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 569 + ], + "type": "text", + "content": " are the coefficients that balance each loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 586, + 231, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 231, + 601 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 231, + 601 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 610, + 270, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 610, + 270, + 624 + ], + "spans": [ + { + "bbox": [ + 131, + 610, + 270, + 624 + ], + "type": "text", + "content": "4.1 Experimental Setting" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Dataset We filter the Objaverse [13] dataset, removing scene-level objects and low quality meshes, and get around 376k valid high quality objects as the training set. We reuse the rendered images from SyncDreamer [31] which contain" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 481, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 481, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 481, + 140 + ], + "type": "text", + "content": "16 images per shape at the resolution of " + }, + { + "bbox": [ + 131, + 116, + 481, + 140 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 131, + 116, + 481, + 140 + ], + "type": "text", + "content": ", and additionally render 6 orthographic images and CCM with the same lighting and resolution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 154, + 481, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 154, + 481, + 191 + ], + "spans": [ + { + "bbox": [ + 131, + 154, + 481, + 191 + ], + "type": "text", + "content": "Network Architecture The reconstruction model contains around 300M parameters. The U-Net contains [64, 128, 128, 256, 256, 512, 512] channels, with attention blocks at resolution [32, 16, 8]. We set the Flexicubes grid size as 80." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 204, + 482, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 204, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 482, + 310 + ], + "type": "text", + "content": "Implementation Details The reconstruction model was trained on 8 NVIDIA A800 80GB GPU cards for 6 days with " + }, + { + "bbox": [ + 130, + 204, + 482, + 310 + ], + "type": "inline_equation", + "content": "110k" + }, + { + "bbox": [ + 130, + 204, + 482, + 310 + ], + "type": "text", + "content": " iterations. The model was trained with batch size 32 (32 shapes per iteration). At each iteration, we randomly sampled 8 views among the total 16 images for each shape for supervision. We used the Adam optimizer with learning rate " + }, + { + "bbox": [ + 130, + 204, + 482, + 310 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 130, + 204, + 482, + 310 + ], + "type": "text", + "content": ". The coefficients that balancing each loss were set as " + }, + { + "bbox": [ + 130, + 204, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{LPIPS}} = 0.1" + }, + { + "bbox": [ + 130, + 204, + 482, + 310 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 204, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{depth}} = 0.5" + }, + { + "bbox": [ + 130, + 204, + 482, + 310 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 204, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{mask}} = 0.5" + }, + { + "bbox": [ + 130, + 204, + 482, + 310 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 204, + 482, + 310 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{reg}} = 0.005" + }, + { + "bbox": [ + 130, + 204, + 482, + 310 + ], + "type": "text", + "content": ". To enhance robustness against minor inconsistency in the generated multi-view images, we introduced small Gaussian noise to the inputs in both training and inference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 312, + 482, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 372 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 372 + ], + "type": "text", + "content": "The diffusion models for both six orthographic images and CCMs were trained on 8 NVIDIA A800 80GB GPU cards for 2 days with " + }, + { + "bbox": [ + 130, + 312, + 482, + 372 + ], + "type": "inline_equation", + "content": "10k" + }, + { + "bbox": [ + 130, + 312, + 482, + 372 + ], + "type": "text", + "content": " iterations. The gradient accumulation is set as 12 steps, yielding a total batch size of 1536. We used the Adam optimizer with learning rate " + }, + { + "bbox": [ + 130, + 312, + 482, + 372 + ], + "type": "inline_equation", + "content": "5e - 5" + }, + { + "bbox": [ + 130, + 312, + 482, + 372 + ], + "type": "text", + "content": ". During sampling both diffusions were sampled with 50 steps using DDIM [50]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 386, + 295, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 386, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 131, + 386, + 295, + 397 + ], + "type": "text", + "content": "4.2 Comparison with baselines" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 402, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 582 + ], + "type": "text", + "content": "Qualitative Results To validate the effectiveness of our method, we qualitatively compare our results with previous works including Wonder3d [32], SyncDreamer [31], Magic123 [41], One-2-3-45 [29] and OpenLRM [17]. Since LRM [18] is not open-sourced, we use OpenLRM [17], an open-sourced implementation of LRM for comparisons. For the other baselines, we use their official codes and checkpoints. As for the input images for testing, we choose two from GSO [14] dataset, one downloaded from web and one generated by text-to-image diffusion model. The results are shown in Fig. 5. It can be seen from the figure that our method generates 3D textured meshes with better texture and geometry than all other baselines. This is because our reconstruction model fully utilizes the spatial alignment of input six orthographic images and output triplane. Also, our model can generate with only 10 seconds, much faster than most of the baselines. Our method is trained in an end-to-end manner with textured mesh as final output, thus avoiding a time-consuming post-processing for converting to mesh as in [52]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 641 + ], + "type": "text", + "content": "Additionally, we visualize the generated meshes comparing to the previous work LRM [18] and a concurrent work LGM [52]. Since LRM is not open-sourced, we use the meshes from their project page. The results are shown in Fig. 6. It can be seen from the figure that our results have better texture. Our method also has smoother geometry than LRM and better geometry details than LGM." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "In Fig. 7, we show more results of the high-fidelity textured meshes generated from single image by our method." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 297, + 91, + 447, + 101 + ], + "type": "text", + "content": "Convolutional Reconstruction Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 481, + 363 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 481, + 363 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 481, + 363 + ], + "type": "image", + "image_path": "89ea23725eeb3f89f3193c423cb2e43247383cd53dbdc7321f2e955a7ef6389f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 372, + 482, + 396 + ], + "lines": [ + { + "bbox": [ + 130, + 372, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 482, + 396 + ], + "type": "text", + "content": "Fig. 6: Qualitative comparison with LRM [18] and LGM [52]. Our models generate high-fidelity results with detailed texture and smooth geometry." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 431, + 482, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 431, + 482, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 431, + 482, + 576 + ], + "type": "text", + "content": "Quantitative Results In line with previous studies [32], we evaluate our method using the Google Scanned Objects (GSO) dataset [14] which is not included in our training dataset. We randomly choose 30 shapes and render a single image with size of " + }, + { + "bbox": [ + 130, + 431, + 482, + 576 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 431, + 482, + 576 + ], + "type": "text", + "content": " as input for evaluation. To ensure the generated mesh accurately aligned with the ground truth mesh, we carefully adjust their pose and scale them to fit within the " + }, + { + "bbox": [ + 130, + 431, + 482, + 576 + ], + "type": "inline_equation", + "content": "[-0.5, 0.5]" + }, + { + "bbox": [ + 130, + 431, + 482, + 576 + ], + "type": "text", + "content": " box. For mesh geometry evaluation, we report Chamfer Distance (CD), Column IoU and F-Score (with a threshold of 0.05, following One-2-3-45 [29]), which measure the geometry similarity between the reconstructed mesh and ground truth mesh. The results are shown in Table 1. It can be seen from the table that our method outperforms all of the baselines, which demonstrates the effectiveness of our method for geometry quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Furthermore, for evaluating mesh texture, we render 24 images at " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " resolution at elevation angles of 0, 15, and 30 degrees, for the generated meshes and ground-truth meshes respectively. For each elevation, the 8 images are evenly distributed around a full 360-degree rotation. Then we assessed them using several metrics: PSNR, SSIM, LPIPS and Clip-Similarity, which measure the resemblance in appearance between the reconstructed mesh and the original ground truth mesh. The results are shown in Table 2. It shows that the generated" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 116, + 476, + 407 + ], + "blocks": [ + { + "bbox": [ + 140, + 116, + 476, + 407 + ], + "lines": [ + { + "bbox": [ + 140, + 116, + 476, + 407 + ], + "spans": [ + { + "bbox": [ + 140, + 116, + 476, + 407 + ], + "type": "image", + "image_path": "71497811b6dc9337e1c15445273e7a297d4a88dfe94ddc81167a9d4c8dea33ff.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 418, + 471, + 430 + ], + "lines": [ + { + "bbox": [ + 140, + 418, + 471, + 430 + ], + "spans": [ + { + "bbox": [ + 140, + 418, + 471, + 430 + ], + "type": "text", + "content": "Fig. 7: More Results of the generated results of our method from a single image." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 454, + 480, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 454, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 454, + 480, + 478 + ], + "type": "text", + "content": "textured meshes by our model surpass those of all baselines in appearance, which demonstrates the effectiveness of our method for texture quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 479, + 482, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 575 + ], + "type": "text", + "content": "Additionally, we carry out experiments to evaluate the effectiveness of our single image to multi-view diffusion models. We use PSNR, SSIM, and LPIPS to measure the similarity between the generated multi-view images and ground truth multi-view images. For this analysis, we use four views of the generated images (left, right, front, back) from each model under comparison, setting the background color to grey (value 128). We compare with SyncDreamer [31] and Wonder3D [32]. The outcomes of this evaluation are documented in Table 3. It can be seen from the table that our method outperforms all the baselines." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 594, + 307, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 307, + 607 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 307, + 607 + ], + "type": "text", + "content": "4.3 Ablation Study and Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Reconstruction Results on Early Training Stage An advantage of our CRM is that it is easy to train. In fact, we find that CRM starts to show reasonable results at very early stage of training. The results are shown in Fig. 8. The results are good even with barely 280 iterations (only 20 minutes of training)." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 297, + 91, + 447, + 101 + ], + "type": "text", + "content": "Convolutional Reconstruction Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 190, + 157, + 424, + 270 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Quantitative comparison for the geometry quality between our method and baselines for single image to 3D textured mesh generation. We report the metrics of Chamfer Distance, Column IoU and F-score on GSO dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 190, + 157, + 424, + 270 + ], + "lines": [ + { + "bbox": [ + 190, + 157, + 424, + 270 + ], + "spans": [ + { + "bbox": [ + 190, + 157, + 424, + 270 + ], + "type": "table", + "html": "
MethodChamfer Dist.↓ Vol. IoU↑ F-Sco. (%)↑
One-2-3-45 [29]0.01720.446372.19
SyncDreamer [31]0.01400.390075.74
Wonder3D [32]0.01860.439876.75
Magic123 [41]0.01880.371460.66
TGS [71]0.01720.298265.17
OpenLRM [17,18]0.01680.377463.22
LGM [52]0.01170.468568.69
Ours0.00940.613179.38
", + "image_path": "09ec059f55aa499e6ba78b4ae4c820a84eed1eebdafd23babce5f8e0c2aa9f91.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 194, + 323, + 420, + 425 + ], + "blocks": [ + { + "bbox": [ + 132, + 281, + 482, + 314 + ], + "lines": [ + { + "bbox": [ + 132, + 281, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 482, + 314 + ], + "type": "text", + "content": "Table 2: Quantitative comparison for the texture quality between our method and baselines for single image to 3D textured mesh generation. We report the metric of PSNR, SSIM, LPIPS and Clip [43]-Similarity on GSO dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 194, + 323, + 420, + 425 + ], + "lines": [ + { + "bbox": [ + 194, + 323, + 420, + 425 + ], + "spans": [ + { + "bbox": [ + 194, + 323, + 420, + 425 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑LPIPS↓Clip-Sim↑
One-2-3-45 [29]13.930.80840.262579.83
SyncDreamer [31]14.000.81650.259182.76
Wonder3D [32]13.310.81210.255483.70
Magic123 [41]12.690.79840.244285.16
OpenLRM [17,18]14.300.82940.227684.20
LGM [71]13.280.79460.256085.20
Ours16.220.83810.214387.55
", + "image_path": "a45b13e4cf8c34edd0f32f95ef862c186acf35988628e22f5f6949b10bab31bf.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 448, + 481, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 448, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 448, + 481, + 472 + ], + "type": "text", + "content": "We conjecture that the fast convergence results from the strong geometry prior in our architecture design." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 490, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 482, + 574 + ], + "type": "text", + "content": "Training Time of CRM In Fig. 9 we compare the training cost between our method (reconstruction model only) and two baselines, LRM [18] and LGM [52]. We measure the training cost by the training days multiplying the amount of used NVIDIA A100/A800 GPU cards. It can be seen that our model takes much smaller training time than the two baselines. This is because our model utilizes the spatial correspondence between input six orthographic images/CCMs and triplanes, which serves as a strong prior that makes the training easier." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 592, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 592, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 592, + 482, + 666 + ], + "type": "text", + "content": "Importance of Input CCM We examine the importance of the CCMs that are concatenated to the input images. To compare, we train a reconstruction model that takes only the six RGB images as input, without CCMs. The results are shown in Fig. 10. It can be seen that the results of the geometry degrade a lot without CCM input. This is because CCM provides important geometry information for the model, especially when the geometry is complex." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 217, + 157, + 395, + 215 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 3: Quantitative comparison between our method and baselines for novel view synthesis of the multi-view diffusion model. We report the metric of PSNR, SSIM and LPIPS on GSO dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 217, + 157, + 395, + 215 + ], + "lines": [ + { + "bbox": [ + 217, + 157, + 395, + 215 + ], + "spans": [ + { + "bbox": [ + 217, + 157, + 395, + 215 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑LPIPS↓
SyncDreamer [31]20.300.78040.2932
Wonder3D [32]23.760.81270.2210
Ours29.360.87210.1354
", + "image_path": "897d0152144cd4c7d8ec18e1ab84dc2a38e4f1db9777fde1b50a5ecff542d82b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 155, + 244, + 208, + 269 + ], + "blocks": [ + { + "bbox": [ + 155, + 244, + 208, + 269 + ], + "lines": [ + { + "bbox": [ + 155, + 244, + 208, + 269 + ], + "spans": [ + { + "bbox": [ + 155, + 244, + 208, + 269 + ], + "type": "image", + "image_path": "5631a7b4a9a4de8da88578f23a96c9c2a6081801c8c582a822d054c81189a900.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 277, + 203, + 297 + ], + "lines": [ + { + "bbox": [ + 154, + 277, + 203, + 297 + ], + "spans": [ + { + "bbox": [ + 154, + 277, + 203, + 297 + ], + "type": "text", + "content": "(a) 0 iters (Initialization)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 217, + 232, + 287, + 274 + ], + "blocks": [ + { + "bbox": [ + 217, + 232, + 287, + 274 + ], + "lines": [ + { + "bbox": [ + 217, + 232, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 217, + 232, + 287, + 274 + ], + "type": "image", + "image_path": "ed9850342b38e8734023597fdd40bc18b788b18c1f82275e993a9a15f9cb4b1e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 221, + 277, + 288, + 297 + ], + "lines": [ + { + "bbox": [ + 221, + 277, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 221, + 277, + 288, + 297 + ], + "type": "text", + "content": "(b) 280 iters " + }, + { + "bbox": [ + 221, + 277, + 288, + 297 + ], + "type": "inline_equation", + "content": "(\\sim 20" + }, + { + "bbox": [ + 221, + 277, + 288, + 297 + ], + "type": "text", + "content": " min training)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 145, + 300, + 212, + 342 + ], + "blocks": [ + { + "bbox": [ + 145, + 300, + 212, + 342 + ], + "lines": [ + { + "bbox": [ + 145, + 300, + 212, + 342 + ], + "spans": [ + { + "bbox": [ + 145, + 300, + 212, + 342 + ], + "type": "image", + "image_path": "7f82071071640cd1ea11c97faff85820c3335d8b66ddf33e61fbc189970c6d53.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 346, + 212, + 367 + ], + "lines": [ + { + "bbox": [ + 144, + 346, + 212, + 367 + ], + "spans": [ + { + "bbox": [ + 144, + 346, + 212, + 367 + ], + "type": "text", + "content": "(c) 4k iters " + }, + { + "bbox": [ + 144, + 346, + 212, + 367 + ], + "type": "inline_equation", + "content": "(\\sim 5" + }, + { + "bbox": [ + 144, + 346, + 212, + 367 + ], + "type": "text", + "content": " hours training)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 233, + 300, + 267, + 341 + ], + "blocks": [ + { + "bbox": [ + 233, + 300, + 267, + 341 + ], + "lines": [ + { + "bbox": [ + 233, + 300, + 267, + 341 + ], + "spans": [ + { + "bbox": [ + 233, + 300, + 267, + 341 + ], + "type": "image", + "image_path": "b12c8e77990ff262098ce165e36edb196ebcc6dfb9dbd0747fedf1e6927fc518.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 346, + 282, + 364 + ], + "lines": [ + { + "bbox": [ + 222, + 346, + 282, + 364 + ], + "spans": [ + { + "bbox": [ + 222, + 346, + 282, + 364 + ], + "type": "text", + "content": "(d) Ground Truth Mesh" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 320, + 229, + 468, + 369 + ], + "blocks": [ + { + "bbox": [ + 320, + 229, + 468, + 369 + ], + "lines": [ + { + "bbox": [ + 320, + 229, + 468, + 369 + ], + "spans": [ + { + "bbox": [ + 320, + 229, + 468, + 369 + ], + "type": "image", + "image_path": "7a69169d99d1f4e864126c8739ec686448a7d636802dd089a704018c5ec6828a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 379, + 474, + 411 + ], + "lines": [ + { + "bbox": [ + 314, + 379, + 474, + 411 + ], + "spans": [ + { + "bbox": [ + 314, + 379, + 474, + 411 + ], + "type": "text", + "content": "Fig. 9: Training cost comparison. Our model require much less computation cost than other baselines." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 144, + 431, + 263, + 487 + ], + "blocks": [ + { + "bbox": [ + 138, + 379, + 299, + 412 + ], + "lines": [ + { + "bbox": [ + 138, + 379, + 299, + 412 + ], + "spans": [ + { + "bbox": [ + 138, + 379, + 299, + 412 + ], + "type": "text", + "content": "Fig. 8: Reconstruction results on unseen samples during early stage of training." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 144, + 431, + 263, + 487 + ], + "lines": [ + { + "bbox": [ + 144, + 431, + 263, + 487 + ], + "spans": [ + { + "bbox": [ + 144, + 431, + 263, + 487 + ], + "type": "image", + "image_path": "4f84525782f831285f21ca25454032ed4c91a8f7f355c96a108f402d002df564.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 151, + 491, + 200, + 540 + ], + "blocks": [ + { + "bbox": [ + 151, + 491, + 200, + 540 + ], + "lines": [ + { + "bbox": [ + 151, + 491, + 200, + 540 + ], + "spans": [ + { + "bbox": [ + 151, + 491, + 200, + 540 + ], + "type": "image", + "image_path": "42a32e64756660ebbcffd2abe9bfffc54a4d6e7639058983810447857084e3b0.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 546, + 236, + 557 + ], + "lines": [ + { + "bbox": [ + 163, + 546, + 236, + 557 + ], + "spans": [ + { + "bbox": [ + 163, + 546, + 236, + 557 + ], + "type": "text", + "content": "(a) w/o CCM Input" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 208, + 434, + 262, + 487 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 434, + 262, + 487 + ], + "spans": [ + { + "bbox": [ + 208, + 434, + 262, + 487 + ], + "type": "image", + "image_path": "b9a847461a31f8d484454da2c89d9435a6608d63836a42215c14c3cbe54bec11.jpg" + } + ] + } + ], + "index": 18, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 208, + 491, + 255, + 540 + ], + "blocks": [ + { + "bbox": [ + 208, + 491, + 255, + 540 + ], + "lines": [ + { + "bbox": [ + 208, + 491, + 255, + 540 + ], + "spans": [ + { + "bbox": [ + 208, + 491, + 255, + 540 + ], + "type": "image", + "image_path": "bfbfcdff8787aa365202dce2f181624248a4dc94cab01336d8fbd59207a2c55d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 569, + 482, + 613 + ], + "lines": [ + { + "bbox": [ + 131, + 569, + 482, + 613 + ], + "spans": [ + { + "bbox": [ + 131, + 569, + 482, + 613 + ], + "type": "text", + "content": "Fig. 10: The CCM concatenated to the input images is beneficial for our model. (a) Without providing CCM, the model outputs a geometry which is reasonable, but not very good. (b) The shape reconstructed using our full model with CCM input, with a much better geometry. (c) Ground truth mesh rendered from the same pose." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 274, + 434, + 389, + 487 + ], + "blocks": [ + { + "bbox": [ + 274, + 434, + 389, + 487 + ], + "lines": [ + { + "bbox": [ + 274, + 434, + 389, + 487 + ], + "spans": [ + { + "bbox": [ + 274, + 434, + 389, + 487 + ], + "type": "image", + "image_path": "08abccfcb2adfce265eb95dcc235f806e2f74626c4f17ae2c5acb987bb2a075e.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 278, + 491, + 325, + 540 + ], + "blocks": [ + { + "bbox": [ + 278, + 491, + 325, + 540 + ], + "lines": [ + { + "bbox": [ + 278, + 491, + 325, + 540 + ], + "spans": [ + { + "bbox": [ + 278, + 491, + 325, + 540 + ], + "type": "image", + "image_path": "2a9ed6b6c784b4c32e4a2fd3d7b4f9eea934e2d8ff5233f08826588c172e1532.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 293, + 546, + 361, + 557 + ], + "lines": [ + { + "bbox": [ + 293, + 546, + 361, + 557 + ], + "spans": [ + { + "bbox": [ + 293, + 546, + 361, + 557 + ], + "type": "text", + "content": "(b) w/ CCM Input" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 397, + 431, + 455, + 486 + ], + "blocks": [ + { + "bbox": [ + 397, + 431, + 455, + 486 + ], + "lines": [ + { + "bbox": [ + 397, + 431, + 455, + 486 + ], + "spans": [ + { + "bbox": [ + 397, + 431, + 455, + 486 + ], + "type": "image", + "image_path": "fe5d1ed6f6d9bd3b09d62e90f1e6e4557e0b3c26f9deb2f7aae3afece14607ef.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 403, + 491, + 451, + 540 + ], + "blocks": [ + { + "bbox": [ + 403, + 491, + 451, + 540 + ], + "lines": [ + { + "bbox": [ + 403, + 491, + 451, + 540 + ], + "spans": [ + { + "bbox": [ + 403, + 491, + 451, + 540 + ], + "type": "image", + "image_path": "23679090a5b355d39a3aa067a1be0e679d45e297ab83cc1c66944c0d93f6e6a7.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 382, + 546, + 469, + 557 + ], + "lines": [ + { + "bbox": [ + 382, + 546, + 469, + 557 + ], + "spans": [ + { + "bbox": [ + 382, + 546, + 469, + 557 + ], + "type": "text", + "content": "(c) Ground Truth Mesh" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Design of Multi-view Diffusion Here we examine the effectiveness of the design of the multi-view diffusion models. Starting from the baseline that naively" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 297, + 91, + 447, + 101 + ], + "type": "text", + "content": "Convolutional Reconstruction Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 201, + 146, + 411, + 215 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 4: Ablation study on the design of multi-view diffusion on novel view synthesis. We report the metrics of PSNR, SSIM and LPIPS on GSO dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 201, + 146, + 411, + 215 + ], + "lines": [ + { + "bbox": [ + 201, + 146, + 411, + 215 + ], + "spans": [ + { + "bbox": [ + 201, + 146, + 411, + 215 + ], + "type": "table", + "html": "
MethodPSNR↑ SSIM↑ LPIPS↓
ImageDream (6 view)28.990.85650.1497
+ Zero-SNR29.130.85980.1498
+ Random Resizing29.360.87210.1354
+ Contour Augmentation28.920.86810.1444
", + "image_path": "b4a26437682d78009383ccf04b905936ffe9cef3d3b2f1bf50f99db56518666c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 136, + 227, + 190, + 281 + ], + "blocks": [ + { + "bbox": [ + 136, + 227, + 190, + 281 + ], + "lines": [ + { + "bbox": [ + 136, + 227, + 190, + 281 + ], + "spans": [ + { + "bbox": [ + 136, + 227, + 190, + 281 + ], + "type": "image", + "image_path": "3cbaaf3456ced12b33b53d51aabdd26a26ba0547ad725e472d55ed3ad830fd55.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 139, + 289, + 186, + 298 + ], + "lines": [ + { + "bbox": [ + 139, + 289, + 186, + 298 + ], + "spans": [ + { + "bbox": [ + 139, + 289, + 186, + 298 + ], + "type": "text", + "content": "(a) Input Image" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 208, + 227, + 291, + 281 + ], + "blocks": [ + { + "bbox": [ + 208, + 227, + 291, + 281 + ], + "lines": [ + { + "bbox": [ + 208, + 227, + 291, + 281 + ], + "spans": [ + { + "bbox": [ + 208, + 227, + 291, + 281 + ], + "type": "image", + "image_path": "be846a4b9118f82ecddaabd6ae6cf3b61073f55ee599ada869aca97bc4ee4fad.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 289, + 282, + 298 + ], + "lines": [ + { + "bbox": [ + 209, + 289, + 282, + 298 + ], + "spans": [ + { + "bbox": [ + 209, + 289, + 282, + 298 + ], + "type": "text", + "content": "(b) Segmentation Result" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 228, + 351, + 279 + ], + "blocks": [ + { + "bbox": [ + 318, + 228, + 351, + 279 + ], + "lines": [ + { + "bbox": [ + 318, + 228, + 351, + 279 + ], + "spans": [ + { + "bbox": [ + 318, + 228, + 351, + 279 + ], + "type": "image", + "image_path": "afaa77feb293f97da71e6e6aabfa1855817c73d8bff5b59a60282e105e9906d3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 316, + 482, + 371 + ], + "lines": [ + { + "bbox": [ + 131, + 316, + 482, + 371 + ], + "spans": [ + { + "bbox": [ + 131, + 316, + 482, + 371 + ], + "type": "text", + "content": "Fig. 11: Demonstration of contour augmentation. (a) Given an input image, (b) off-the-shelf segmentation model sometimes provides imperfect results. (c) Without contour augmentation, the predicted backview color is sensitive to the contour. (d) With contour augmentation, the model predicts reasonable result. (e) We demonstrate how we augment the input image during training." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 375, + 228, + 407, + 279 + ], + "blocks": [ + { + "bbox": [ + 309, + 289, + 357, + 307 + ], + "lines": [ + { + "bbox": [ + 309, + 289, + 357, + 307 + ], + "spans": [ + { + "bbox": [ + 309, + 289, + 357, + 307 + ], + "type": "text", + "content": "(c) w/o Contour Augmentation" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 375, + 228, + 407, + 279 + ], + "lines": [ + { + "bbox": [ + 375, + 228, + 407, + 279 + ], + "spans": [ + { + "bbox": [ + 375, + 228, + 407, + 279 + ], + "type": "image", + "image_path": "0324a6d9884bb8b93e9653c54a7aefa89ac3b7fd2efcde330926b09383942e44.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 436, + 231, + 465, + 275 + ], + "blocks": [ + { + "bbox": [ + 368, + 289, + 413, + 307 + ], + "lines": [ + { + "bbox": [ + 368, + 289, + 413, + 307 + ], + "spans": [ + { + "bbox": [ + 368, + 289, + 413, + 307 + ], + "type": "text", + "content": "(d) w/ Contour Augmentation" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 436, + 231, + 465, + 275 + ], + "lines": [ + { + "bbox": [ + 436, + 231, + 465, + 275 + ], + "spans": [ + { + "bbox": [ + 436, + 231, + 465, + 275 + ], + "type": "image", + "image_path": "cff429cc48c431105176294e822c9176b6af397b6db115d0cd3455c6fd8de114.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 289, + 476, + 306 + ], + "lines": [ + { + "bbox": [ + 421, + 289, + 476, + 306 + ], + "spans": [ + { + "bbox": [ + 421, + 289, + 476, + 306 + ], + "type": "text", + "content": "(e) Contour Augmented Image" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 396, + 482, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 494 + ], + "type": "text", + "content": "fine-tunes the pre-trained ImageDream model with 2 additional views, we sequentially add the proposed techniques on the training. We examine the results on a subset of GSO, comparing the similarity of the generated novel view images with the ground truth images using PSNR, SSIM and LPIPS metrics. The results are shown in table 4. It can be seen that both the Zero-SNR trick and random resizing are beneficial. Note that the contour augmentation does not improve the quantitative metrics. However, we find that this trick makes the model more robust to in the wild input images (Fig. 11)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 515, + 220, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 220, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 220, + 528 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 544, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 628 + ], + "type": "text", + "content": "In this work, we present a convolutional reconstruction model (CRM) for creating high-quality 3D models from a single image. Our approach effectively utilizes the spatial relationship between input images and the output triplane, leading to improved textured meshes, with significantly less training cost compared to previous transformer-based methods [18]. The model operates on an end-to-end training basis, directly outputting textured meshes. Overall, our method can produce detailed textured meshes in just 10 seconds." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Potential Negative Impact. Similar to many other generated models, our CRM may be used to generate malicious or fake 3D contents, which may need additional caution." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 483, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 483, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 483, + 187 + ], + "type": "text", + "content": "This work was supported by Natural Science Foundation of China (No.s 62350080, 62076147), Tsinghua Institute for Guo Qiang, and the High Performance Computing Center, Tsinghua University. J.Z is also supported by the New Cornerstone Science Foundation through the XPlorer Prize." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 228, + 481, + 664 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 228, + 480, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 249 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 249 + ], + "type": "text", + "content": "1. Bao, F., Nie, S., Xue, K., Cao, Y., Li, C., Su, H., Zhu, J.: All are worth words: A vit backbone for diffusion models. In: CVPR (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 293 + ], + "type": "text", + "content": "2. Brooks, T., Peebles, B., Homes, C., DePue, W., Guo, Y., Jing, L., Schnurr, D., Taylor, J., Luhman, T., Luhman, E., Ng, C., Wang, R., Ramesh, A.: Video generation models as world simulators (2024), https://openai.com/research/video-generation-models-as-world-simulators" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 295, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 295, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 295, + 480, + 358 + ], + "type": "text", + "content": "3. Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D.M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 360, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 360, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 360, + 480, + 403 + ], + "type": "text", + "content": "4. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., De Mello, S., Gallo, O., Guibas, L.J., Tremblay, J., Khamis, S., et al.: Efficient geometry-aware 3d generative adversarial networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16123-16133 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 404, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 404, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 404, + 480, + 424 + ], + "type": "text", + "content": "5. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: European Conference on Computer Vision. pp. 333-350. Springer (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 426, + 480, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 480, + 457 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 480, + 457 + ], + "type": "text", + "content": "6. Chen, H., Gu, J., Chen, A., Tian, W., Tu, Z., Liu, L., Su, H.: Single-stage diffusion nef: A unified approach to 3d generation and reconstruction. arXiv preprint arXiv:2304.06714 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 459, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 480, + 490 + ], + "type": "text", + "content": "7. Chen, R., Chen, Y., Jiao, N., Jia, K.: *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation*. arXiv preprint arXiv:2303.13873 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 511 + ], + "type": "text", + "content": "8. Chen, Y., Zhang, C., Yang, X., Cai, Z., Yu, G., Yang, L., Lin, G.: It3d: Improved text-to-3d generation with explicit view synthesis (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 514, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 533 + ], + "type": "text", + "content": "9. Chen, Z., Wang, F., Liu, H.: Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 135, + 535, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 535, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 135, + 535, + 480, + 555 + ], + "type": "text", + "content": "10. Cheng, Y.C., Lee, H.Y., Tuyakov, S., Schwing, A., Gui, L.: SDFusion: Multimodal 3d shape completion, reconstruction, and generation. In: CVPR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 557, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 557, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 134, + 557, + 480, + 587 + ], + "type": "text", + "content": "11. Decatur, D., Lang, I., Aberman, K., Hanocka, R.: 3d paintbrush: Local stylization of 3d shapes with cascaded score distillation. arXiv preprint arXiv:2311.09571 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 589, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 589, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 134, + 589, + 480, + 620 + ], + "type": "text", + "content": "12. Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objaverse-xl: A universe of " + }, + { + "bbox": [ + 134, + 589, + 480, + 620 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 134, + 589, + 480, + 620 + ], + "type": "text", + "content": " objects. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 622, + 481, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 481, + 664 + ], + "type": "text", + "content": "13. Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objaverse: A universe of annotated 3d objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13142-13153 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 297, + 91, + 447, + 101 + ], + "type": "text", + "content": "Convolutional Reconstruction Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "14. Downs, L., Francis, A., Koenig, N., Kinman, B., Hickman, R., Reymann, K., McHugh, T.B., Vanhoucke, V.: Google scanned objects: A high-quality dataset of 3d scanned household items. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 2553-2560. IEEE (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 481, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 481, + 204 + ], + "type": "text", + "content": "15. Gao, J., Shen, T., Wang, Z., Chen, W., Yin, K., Li, D., Litany, O., Gojcic, Z., Fidler, S.: Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems 35, 31841-31854 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 204, + 481, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 204, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 133, + 204, + 481, + 225 + ], + "type": "text", + "content": "16. Gupta, A., Xiong, W., Nie, Y., Jones, I., Oğuz, B.: 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 225, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 225, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 133, + 225, + 481, + 247 + ], + "type": "text", + "content": "17. He, Z., Wang, T.: Openlrm: Open-source large reconstruction models. https://github.com/3DTopia/OpenLRM (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 247, + 481, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 247, + 481, + 279 + ], + "spans": [ + { + "bbox": [ + 133, + 247, + 481, + 279 + ], + "type": "text", + "content": "18. Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D., Liu, F., Sunkavalli, K., Bui, T., Tan, H.: Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 279, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 279, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 133, + 279, + 481, + 300 + ], + "type": "text", + "content": "19. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 300, + 481, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 300, + 481, + 321 + ], + "spans": [ + { + "bbox": [ + 133, + 300, + 481, + 321 + ], + "type": "text", + "content": "20. Kim, S., Lee, K., Choi, J.S., Jeong, J., Sohn, K., Shin, J.: Collaborative score distillation for consistent visual synthesis (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 321, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 321, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 133, + 321, + 481, + 354 + ], + "type": "text", + "content": "21. Li, J., Tan, H., Zhang, K., Xu, Z., Luan, F., Xu, Y., Hong, Y., Sunkavalli, K., Shakhnarovich, G., Bi, S.: Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 354, + 481, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 354, + 481, + 375 + ], + "spans": [ + { + "bbox": [ + 133, + 354, + 481, + 375 + ], + "type": "text", + "content": "22. Li, W., Chen, R., Chen, X., Tan, P.: Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 375, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 375, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 133, + 375, + 481, + 396 + ], + "type": "text", + "content": "23. Li, Y., Dou, Y., Shi, Y., Lei, Y., Chen, X., Zhang, Y., Zhou, P., Ni, B.: Focaldreamer: Text-driven 3d editing via focal-fusion assembly (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 396, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 396, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 133, + 396, + 481, + 418 + ], + "type": "text", + "content": "24. Liang, Y., Yang, X., Lin, J., Li, H., Xu, X., Chen, Y.: Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 418, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 418, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 133, + 418, + 481, + 460 + ], + "type": "text", + "content": "25. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300–309 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 460, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 460, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 133, + 460, + 481, + 493 + ], + "type": "text", + "content": "26. Lin, S., Liu, B., Li, J., Yang, X.: Common diffusion noise schedules and sample steps are flawed. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5404-5411 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 493, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 493, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 133, + 493, + 481, + 515 + ], + "type": "text", + "content": "27. Liu, F., Wu, D., Wei, Y., Rao, Y., Duan, Y.: Sherpa3d: Boosting high-fidelity text-to-3d generation via coarse 3d prior (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 515, + 481, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 515, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 133, + 515, + 481, + 547 + ], + "type": "text", + "content": "28. Liu, M., Shi, R., Chen, L., Zhang, Z., Xu, C., Wei, X., Chen, H., Zeng, C., Gu, J., Su, H.: One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion. arXiv preprint arXiv:2311.07885 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 547, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 547, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 133, + 547, + 481, + 579 + ], + "type": "text", + "content": "29. Liu, M., Xu, C., Jin, H., Chen, L., Xu, Z., Su, H., et al.: One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 579, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 579, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 133, + 579, + 481, + 601 + ], + "type": "text", + "content": "30. Liu, R., Wu, R., Hoorick, B.V., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 601, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 601, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 601, + 481, + 632 + ], + "type": "text", + "content": "31. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "type": "text", + "content": "32. Long, X., Guo, Y.C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.H., Habermann, M., Theobalt, C., et al.: Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008 (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "33. Lorraine, J., Xie, K., Zeng, X., Lin, C.H., Takikawa, T., Sharp, N., Lin, T.Y., Liu, M.Y., Fidler, S., Lucas, J.: Att3d: Amortized text-to-3d object synthesis (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 160 + ], + "type": "text", + "content": "34. Lu, Y., Zhang, J., Li, S., Fang, T., McKinnon, D., Tsin, Y., Quan, L., Cao, X., Yao, Y.: Direct2.5: Diverse text-to-3d generation via multi-view 2.5d diffusion (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 480, + 194 + ], + "type": "text", + "content": "35. Mercier, A., Nakhli, R., Reddy, M., Yasarla, R., Cai, H., Porikli, F., Berger, G.: Hexagen3d: Stablediffusion is just one step away from fast and diverse text-to-3d generation. arXiv preprint arXiv:2401.07727 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "36. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "37. Nielson, G.M.: Dual marching cubes. In: IEEE visualization 2004. pp. 489-496. IEEE (2004)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 441, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 441, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 441, + 259 + ], + "type": "text", + "content": "38. Peebles, W., Xie, S.: Scalable diffusion models with transformers (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "type": "text", + "content": "39. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "type": "text", + "content": "40. Qian, G., Cao, J., Siarohin, A., Kant, Y., Wang, C., Vasilkovsky, M., Lee, H.Y., Fang, Y., Skorokhodov, I., Zhuang, P., Gilitschenski, I., Ren, J., Ghanem, B., Aberman, K., Tulyakov, S.: Atom: Amortized text-to-mesh using 2d diffusion (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "41. Qian, G., Mai, J., Hamdi, A., Ren, J., Siarohin, A., Li, B., Lee, H.Y., Skorokhodov, I., Wonka, P., Tulyakov, S., Ghanem, B.: Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "42. Qiu, L., Chen, G., Gu, X., Zuo, Q., Xu, M., Wu, Y., Yuan, W., Dong, Z., Bo, L., Han, X.: Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 423 + ], + "type": "text", + "content": "43. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "44. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 457, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 500 + ], + "type": "text", + "content": "45. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 501, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 544 + ], + "type": "text", + "content": "46. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: Laion-5b: An open large-scale dataset for training next generation image-text models (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 480, + 578 + ], + "type": "text", + "content": "47. Shen, T., Munkberg, J., Hasselgren, J., Yin, K., Wang, Z., Chen, W., Gojcic, Z., Fidler, S., Sharp, N., Gao, J.: Flexible isosurface extraction for gradient-based mesh optimization. ACM Transactions on Graphics (TOG) 42(4), 1-16 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "text", + "content": "48. Shi, R., Wei, X., Wang, C., Su, H.: Zerorf: Fast sparse view 360 " + }, + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "inline_equation", + "content": "\\{\\backslash \\mathrm{deg}\\}" + }, + { + "bbox": [ + 132, + 578, + 480, + 599 + ], + "type": "text", + "content": " reconstruction with zero pretraining. arXiv preprint arXiv:2312.09249 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "text", + "content": "49. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "50. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "51. Sun, J., Zhang, B., Shao, R., Wang, L., Liu, W., Xie, Z., Liu, Y.: Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 297, + 91, + 447, + 100 + ], + "type": "text", + "content": "Convolutional Reconstruction Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 659 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "52. Tang, J., Chen, Z., Chen, X., Wang, T., Zeng, G., Liu, Z.: Lgm: Large multi-view gaussian model for high-resolution 3d content creation (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "text", + "content": "53. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 191 + ], + "type": "text", + "content": "54. Tochilkin, D., Pankratz, D., Liu, Z., Huang, Z., Letts, A., Li, Y., Liang, D., Laforte, C., Jampani, V., Cao, Y.P.: Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 192, + 480, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 480, + 223 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 480, + 223 + ], + "type": "text", + "content": "55. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 245 + ], + "type": "text", + "content": "56. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 245, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 245, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 245, + 480, + 266 + ], + "type": "text", + "content": "57. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 266, + 480, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 266, + 480, + 287 + ], + "spans": [ + { + "bbox": [ + 132, + 266, + 480, + 287 + ], + "type": "text", + "content": "58. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 288, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 288, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 288, + 480, + 308 + ], + "type": "text", + "content": "59. Wang, P., Shi, Y.: Imaginedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 308, + 480, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 480, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 480, + 340 + ], + "type": "text", + "content": "60. Wang, P., Tan, H., Bi, S., Xu, Y., Luan, F., Sunkavalli, K., Wang, W., Xu, Z., Zhang, K.: Pf-lrm: Pose-free large reconstruction model for joint pose and shape prediction. arXiv preprint arXiv:2311.12024 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 340, + 480, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 340, + 480, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 340, + 480, + 383 + ], + "type": "text", + "content": "61. Wang, T., Zhang, B., Zhang, T., Gu, S., Bao, J., Baltrusaitis, T., Shen, J., Chen, D., Wen, F., Chen, Q., et al.: Rodin: A generative model for sculpting 3d digital avatars using diffusion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4563-4573 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 383, + 480, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 383, + 480, + 414 + ], + "spans": [ + { + "bbox": [ + 132, + 383, + 480, + 414 + ], + "type": "text", + "content": "62. Wang, X., Wang, Y., Ye, J., Wang, Z., Sun, F., Liu, P., Wang, L., Sun, K., Wang, X., He, B.: Animatabledreamer: Text-guided non-rigid 3d model generation and reconstruction with canonical score distillation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 415, + 480, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 415, + 480, + 458 + ], + "spans": [ + { + "bbox": [ + 132, + 415, + 480, + 458 + ], + "type": "text", + "content": "63. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=ppJuFSOAnM" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 459, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 459, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 459, + 480, + 490 + ], + "type": "text", + "content": "64. Xu, Y., Tan, H., Luan, F., Bi, S., Wang, P., Li, J., Shi, Z., Sunkavalli, K., Wetzstein, G., Xu, Z., et al.: Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model. arXiv preprint arXiv:2311.09217 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 491, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 510 + ], + "type": "text", + "content": "65. Yu, X., Guo, Y.C., Li, Y., Liang, D., Zhang, S.H., Qi, X.: Text-to-3d with classifier score distillation (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 511, + 480, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 480, + 532 + ], + "type": "text", + "content": "66. Zeng, X., Vahdat, A., Williams, F., Gojcic, Z., Litany, O., Fidler, S., Kreis, K.: Lion: Latent point diffusion models for 3d shape generation (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 533, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 480, + 553 + ], + "type": "text", + "content": "67. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 553, + 480, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 553, + 480, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 553, + 480, + 574 + ], + "type": "text", + "content": "68. Zheng, X.Y., Pan, H., Guo, Y.X., Tong, X., Liu, Y.: Mvd" + }, + { + "bbox": [ + 132, + 553, + 480, + 574 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 132, + 553, + 480, + 574 + ], + "type": "text", + "content": ": Efficient multiview 3d reconstruction for multiview diffusion (2024)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 574, + 480, + 605 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 480, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 480, + 605 + ], + "type": "text", + "content": "69. Zheng, X.Y., Pan, H., Wang, P.S., Tong, X., Liu, Y., Shum, H.Y.: Locally attentional sdf diffusion for controllable 3d shape generation. arXiv preprint arXiv:2305.04461 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 606, + 480, + 627 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 606, + 480, + 627 + ], + "spans": [ + { + "bbox": [ + 132, + 606, + 480, + 627 + ], + "type": "text", + "content": "70. Zhu, J., Zhuang, P.: Hifa: High-fidelity text-to-3d generation with advanced diffusion guidance (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 628, + 480, + 659 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 628, + 480, + 659 + ], + "spans": [ + { + "bbox": [ + 132, + 628, + 480, + 659 + ], + "type": "text", + "content": "71. Zou, Z.X., Yu, Z., Guo, Y.C., Li, Y., Liang, D., Cao, Y.P., Zhang, S.H.: Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. arXiv preprint arXiv:2312.09147 (2023)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CSOT_ Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection/4b43039d-68da-4e38-a8cf-39348329a3bc_content_list.json b/2024/CSOT_ Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection/4b43039d-68da-4e38-a8cf-39348329a3bc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..acb2bcd6638388335ea68df22bdbab5b50acbf1d --- /dev/null +++ b/2024/CSOT_ Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection/4b43039d-68da-4e38-a8cf-39348329a3bc_content_list.json @@ -0,0 +1,1531 @@ +[ + { + "type": "text", + "text": "CSOT: Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection", + "text_level": 1, + "bbox": [ + 259, + 141, + 743, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinglin Zhan $^{1}$ , Tiejun Liu $^{1}$ , Rengang Li $^{1}$ , Zhaoxiang Zhang $^{2}$ , and Yuntao Chen $^{3\\star}$", + "bbox": [ + 266, + 210, + 736, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 IEIT Systems", + "bbox": [ + 393, + 253, + 557, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Institute of Automation, CAS", + "bbox": [ + 393, + 268, + 609, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Centre for Artificial Intelligence and Robotics, HKISI, CAS {zhanjinglin, liutj, lirg}@ieisystem.com, zhaoxiang.zhang@ia.ac.cn, chenyuntao08@gmail.com", + "bbox": [ + 295, + 281, + 705, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Large-scale 3D bounding box annotation is crucial for LiDAR object detection but comes at a high cost. Semi-supervised object detection (SSOD) offers promising solutions to leverage unannotated data, but the predominant pseudo-labeling approach requires careful hyperparameter tuning for training on noisy teacher labels. In this work, we propose a Cross-Scan Object Transfer (CSOT) paradigm for LiDAR SSOD. Central to our approach is HotspotNet, a transformer-based network that predicts possible placement locations and the object-place fitness scores for inserting annotated objects into unlabeled scans in a semantic coherence manner. Based on HotspotNet, CSOT successfully enables object copy-paste in LiDAR SSOD for the first time. To train object detectors on partially annotated scans generated by CSOT, we adopt a spatial-aware classification loss throughout our partial supervision to handle false negative issues caused by treating all unlabeled objects as background. We conduct extensive experiments to verify the efficacy and generality of our method. Compared to other state-of-the-art label-efficient methods used in LiDAR detection, our approach requires the least amount of annotation while achieves the best detector. Using only $1\\%$ of the labeled data on the Waymo dataset, our semi-supervised detector achieves performance on par with the fully supervised baseline. Similarly, on the nuScenes dataset, our semi-supervised CenterPoint reaches $99\\%$ of the fully supervised model's detection performance in terms of NDS score, while using just $5\\%$ of the labeled data. Code is released at https://github.com/JinglinZhan/CSOT", + "bbox": [ + 261, + 358, + 743, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Semi-Supervised Learning $\\cdot$ LiDAR Object Detection $\\cdot$ Autonomous Driving", + "bbox": [ + 261, + 703, + 740, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 756, + 375, + 771 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large-scale annotated datasets have catalyzed the success of LiDAR object detection, but at a significant cost in terms of money and time [25,33]. For instance,", + "bbox": [ + 212, + 786, + 785, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 218, + 825, + 385, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a0cfdbec5a5b1e466502e6b36e1b8f1e6949a579beae5cfb01e3f3b675310284.jpg", + "image_caption": [ + "Fig. 1: Preliminary experiments of CSOT pipeline for LiDAR SSOD. (a) Compare CSOT implemented based on different cross-scan object transfer methods and display the NDS differences. (b) Visualize prediction differences among various teacher model that used to generate pseudo labels for self-training." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 759, + 359 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "annotating the 12 million 3D boxes in the Waymo Open [38] dataset may have consumed millions of dollars and tens of thousands of human hours. In contrast, autonomous driving fleets capture thousands, and even millions, of hours of unannotated sensor data every day. The huge gap between the speed of manual labeling and that of data accumulation makes labeling a major slowdown. This calls for a more scalable approach to leverage the abundant data collected by autonomous fleets and continually advance object detection capabilities.", + "bbox": [ + 212, + 459, + 787, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semi-Supervised Learning (SSL) presents a promising way for reducing annotation demands by leveraging both limited labeled data and vast quantities of unlabeled data during training. Various SSL methods have been proposed, including pseudo-labeling (a.k.a. self-training) [23], co-training [2], consistency regularization [20] and generative modeling [19]. Among these SSL approaches, pseudo-labeling has gained significant traction, particularly in the field of semi-supervised object detection (SSOD) [4, 12, 31, 37, 39, 40, 42, 52, 62, 65]. However, generating reliable pseudo-labels without human intervention remains a significant challenge. Noisy pseudo-labels can negatively impact model performance, leading to suboptimal results. To address this, researchers have proposed various techniques, including sophisticated post-processing mechanisms [4, 5, 26, 28, 40, 52, 60], iterative training procedures [39, 45, 49], and noise-tolerant detectors [12, 65] and loss objectives [31, 32, 62, 64]. Nevertheless, these approaches often involve complex pipelines and require careful parameter tuning, making them inefficient and difficult to adapt to new datasets.", + "bbox": [ + 212, + 566, + 787, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast with noisy pseudo-labeling, we leverage a novel semi-supervised learning approach for 3D object detection that transfers annotated objects to generate noise-free training data. Specifically, we transfer point clouds and bound", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Zhan et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing boxes of annotated objects to unannotated LiDAR scans to create partially labeled data. Object transfer via copy-paste is a commonly employed data augmentation [6,7,16,47,51] in LiDAR object detection. But naively applying object copy-paste [11, 15, 47] to semi-supervised LiDAR object detection yields suboptimal results. As shown in Fig. 1, SSOD built upon vanilla object transfer does improve performance over the baseline, but combining it with self-training results in a negative net effect on detection performance compared to using self-training alone. The key insight is that object copy-paste without careful contextual consistency maintenance improves detector performance at the cost of generating a large number of false positives. Although the commonly used object detection metric mAP is not sensitive to false positives if the recall improves, the self-training pipeline accumulates these false positives over iterations.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to achieve coherent object copy-paste, we propose Hotspot Network, a learning-based approach to insert objects with strong contextual consistency, enabling object-transfer-based semi-supervised LiDAR object detection to achieve its full potential. HotspotNet takes unannotated LiDAR scans as inputs, along with point clouds and bounding boxes of annotated objects. It then predicts possible placement locations for these objects and assigns scores to each location, indicating the likelihood of the object fitting well within the context of the scanned environment. This allows us introducing noise-free supervision signal into unlabeled point clouds and use these newly created partially-labeled data to optimize detector performance. Based on HotspotNet, we design a novel Cross-Scan Object Transfer(CSOT) pipeline for LiDAR SSOD. CSOT enjoys precise partial annotations that maintain great contextual coherence with unannotated scans rather than noisy pseudo labels for training detectors on unlabeled data. It's important to note that training detectors on partially annotated point clouds, as opposed to fully labeled ones, can lead to a significant issue: unlabeled objects being miscategorized as background, resulting in a high number of false negatives. To mitigate this problem, we propose a spatial-aware classification loss and employ it throughout our partial supervision. Our CSOT paradigm features a streamlined training pipeline and requires minimal hyper-parameter tuning, enabling it to quickly generalize across various datasets and detectors. Moreover, this approach is orthogonal and complementary to commonly used self-training techniques. We conducted extensive experiments on the Waymo [38] and nuScenes [3] datasets to confirm the effectiveness of our methods.", + "bbox": [ + 212, + 329, + 787, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions are summarized below:", + "bbox": [ + 238, + 676, + 553, + 690 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a Hotspot Network for instructing object placement in a semantic coherence manner, and employ an asymmetric supervision for training HotspotNet on both annotated and unannotated data.", + "- Based on HotspotNet, we design a novel Cross-Scan Object Transfer paradigm for LiDAR semi-supervised object detection. To the best of our knowledge, this is the first successful attempt of object copy-paste in LiDAR SSOD.", + "- We conduct extensive experiments on nuScenes and Waymo datasets to verify the efficacy and generality of our method. Our approach achieves competitive performance compared to fully-supervised baselines while using 20x" + ], + "bbox": [ + 225, + 703, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection", + "bbox": [ + 238, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to 100x fewer annotations, significantly outperforming other state-of-the-art label-efficient approaches in LiDAR object detection.", + "bbox": [ + 241, + 146, + 784, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 195, + 388, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Semi-Supervised Object Detection", + "text_level": 1, + "bbox": [ + 215, + 222, + 545, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unsupervised learning eliminates data annotation but significantly degrades model performance [48, 55, 61]. Semi-supervised methods maintain low annotation costs meanwhile greatly bridges the gap to fully-supervised approaches [43]. As a prior work, STAC [37] proposes a pseudo labeling SSOD paradigm and catalyzes extensive follow-on researches. It trains a teacher with all annotated data and generates pseudo labels for supervising the student on unlabeled data. Producing reliable pseudo labels is critical in this paradigm and numerous methods have been explored. 3D IoUMatch [40] utilizes the estimated 3D IoU to filter poorly localized predictions. LabelMatch [4] leverages adaptive label-distribution-aware confidence thresholds for generating unbiased pseudo labels. Proficient Teacher [52] employs clustering-based box voting for improving the precision of final predictions. HSSDA [29] designs a dynamic dual-threshold strategy to generate more reasonable supervision for training the student model. Besides various filter mechanism, many other approaches, such as iterative training procedures [39, 62], noise-tolerant detectors [12, 65] and a variety of loss objectives with consistency or reweighting terms [31, 42, 62], are also proposed for alleviating negative effects caused by noisy pseudo labels. Despite some progress has been made, unreliable pseudo-labels still remain an open challenge. The sophisticated pipelines with extensive manual threshold tuning also impede these approaches' quick adaptation to new datasets.", + "bbox": [ + 212, + 242, + 787, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Copy-Paste For LiDAR Object Detection", + "text_level": 1, + "bbox": [ + 215, + 561, + 602, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Copy-paste method is developed by Jaderberg et.al. for text recognition [18], then extended by Dwibedi et.al. [9] and Yun et.al. [57] for 2D object detection. Yan et.al. [47] introduces copy-paste into LiDAR object detection and develops the well known ground-truth augmentation approach (so called GT-Aug), which is extensively adopted in many later studies [1,6,16,17,21,51,54,66]. Since then, numerous approaches have been proposed to optimize GT-Aug. Part-Aware [8] and Shape-Aware [63] GT-Aug divide objects into partitions and apply different augmentation methods to each partition. Pattern-Aware [14] GT-Aug downsamples the points of objects to create a new one with farther distance. PointCut-Mix [22] replaces part of the sample with shape-preserved subsets from another one. Some researchers notice placement issues in GT-Aug and develop hand-craft valid maps to restrict the insertion of objects [11,15]. Up to now, researches on copy-paste is primarily focused on exploring and developing its capabilities for augmenting data to support fully-supervised learning tasks. Ghiasi et.al. introduce copy-paste into SSL [13] for augmenting both pseudo labeled and supervised labeled images. The above trial still limited in the field of data augmentation and the copy-paste technique remains largely untapped.", + "bbox": [ + 212, + 583, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Zhan et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/57f1dfe65a9d740e6d0734a6d3dcc14625ba67651a748d8abde37c147147fbf8.jpg", + "image_caption": [ + "Fig. 2: The overall architecture of HotspotNet. The Object Encoders extract features from object point cloud $X_{obj,pc}$ and bounding box meta data $X_{obj,meta}$ . The scene encoder featurize the scene point cloud and cross-attends to object features $\\mathcal{F}_{\\mathrm{obj,pc}}$ and $\\mathcal{F}_{\\mathrm{obj,meta}}$ . The location decoder predicts the placeable locations and their scores." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 763, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 435, + 330, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We provide details of Cross-Scan Object Transfer SSOD paradigm in this section. We present the overall architecture of HotspotNet in Sec. 3.1, and describe the Asymmetric Supervision for training HotspotNet in Sec. 3.2. The pipeline of our newly proposed SSOD paradigm, CSOT, is elaborated in Sec. 3.3.", + "bbox": [ + 212, + 464, + 787, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Hotspot Network", + "text_level": 1, + "bbox": [ + 215, + 545, + 405, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The key to LiDAR copy-paste is identifying semantically coherent locations within background scans to paste objects. Placing objects at unreasonable locations, such as inside walls, floating in sky or with weird pose, will cause a large amount of false positives and thus deteriorate the detector performance.", + "bbox": [ + 212, + 568, + 785, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose a learning-based approach for object placement that provides sufficient supervision signal and maintain a high signal-to-noise ratio. Specifically, we employ a Hotspot Network to predict the placeable location $\\mathbf{b}$ and the object-place fitness scores $o$ for inserting object in a semantic coherence manner.", + "bbox": [ + 212, + 628, + 785, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We abbreviate Hotspot Network as HotspotNet and visualize the overall architecture in Fig. 2. The HotspotNet comprises object encoders for object point cloud $X_{\\mathrm{obj,pc}}$ and meta data $X_{\\mathrm{obj,meta}}$ , respectively. The scene encoder featurizes the scene point cloud and cross-attends to object features $\\mathcal{F}_{\\mathrm{obj,pc}}$ and $\\mathcal{F}_{\\mathrm{obj,meta}}$ . A location decoder predicts the placeable location $\\mathbf{b} = \\{x,y,z,w,l,h,\\theta ,c\\}$ and the location fitness score $o$ .", + "bbox": [ + 212, + 689, + 785, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Object Encoder We process the object point cloud $X_{\\mathrm{obj,pc}}$ with object point cloud encoders like PointNet [34] or PointNeXt [35] and obtain object features $\\mathcal{F}_{\\mathrm{obj,pc}}$ . We encode the bounding box meta $X_{\\mathrm{obj,meta}}$ with appropriate meta feature encoders like Fourier embedding for boxes, and CLIP or one-hot embedding", + "bbox": [ + 214, + 779, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection", + "bbox": [ + 238, + 113, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for object classes. We then transform the initial box embedding with a MLP to obtain bounding box meta feature $\\mathcal{F}_{\\mathrm{obj,meta}}$ .", + "bbox": [ + 212, + 146, + 782, + 178 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\mathrm {o b j}, \\mathrm {p c}} = 0 \\text {b j e c t P o i n t C l o u d E n c o d e r} \\left(X _ {\\mathrm {o b j}, \\mathrm {p c}}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 277, + 191, + 784, + 207 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\text {o b j , m e t a}} = \\operatorname {M L P} \\left(\\text {F o u r i e r E m b e d} \\left(X _ {\\text {o b j , b o x}}\\right) \\mid \\mid \\text {C l a s s E m b e d} \\left(X _ {\\text {o b j , c l a s s}}\\right)\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 261, + 210, + 784, + 226 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scene Encoder Given a scene LiDAR scan $X_{\\mathrm{scene}}$ , we leverage off-the-shelf point cloud encoders like CenterPoint [54] or SST [10] for extracting features $\\mathcal{F}_{\\mathrm{scene}}$ . We use spatial cross-attention in the first convolution block of each stage of the scene point cloud encoder. This allows the scene encoder aggregating features from both the object point clouds and the box metadata for assessing the object fitness of each spatial location.", + "bbox": [ + 212, + 238, + 782, + 328 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Location Decoder We randomly initialize a set of learnable queries $\\mathbf{q}_0$ . These queries are combine with position embeddings to obtain location queries $\\mathbf{q}$ . Our location decoder leverages both scene features and object features for predicting possible inserting locations and scores.", + "bbox": [ + 212, + 330, + 782, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Training HotspotNet with Asymmetric Supervision", + "text_level": 1, + "bbox": [ + 214, + 416, + 684, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The HotspotNet consists of a location predictor and a location scorer that work jointly to suggest object placements. The location predictor outputs candidate object positions, while the scorer assigns fitness scores to evaluate how semantically consistent each candidate is. We employ a multi-task learning approach that features Asymmetric Supervision for optimizing different predictions. The proposed losses supervise the location predictor and scorer to accurately position objects within a new environment.", + "bbox": [ + 212, + 445, + 782, + 550 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To train the location predictor, we utilize the annotated LiDAR scans. For each annotated scan, we randomly extract an object point cloud from the original point clouds along with its corresponding bounding box. The regressor is then supervised with an $\\ell_1$ loss between the predicted location $\\mathbf{b}$ and the extracted bounding box $\\mathbf{b}_{\\mathrm{cut - out}}$ . We constrain object distance and perspective from the ego car in Eq. 3 to preserve perspective and point density. We employ dropout regularization on the bounding box meta data to prevent the location predictor from directly copying the meta data to the prediction output. It encourages the predictor to learn meaningful features from the point clouds itself, rather than relying solely on the provided meta data.", + "bbox": [ + 212, + 551, + 784, + 703 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {l o c}} (\\mathbf {b}) = \\left\\| \\mathbf {b} - \\mathbf {b} _ {\\text {c u t - o u t}} \\right\\| _ {1} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 718, + 784, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To train the location scorer, we utilize an asymmetric supervision approach for positive and negative samples. Positive samples are obtained by matching location queries with the extracted object point clouds. These positive samples encourage the scorer to assign high scores to queries that accurately position objects within the scan.", + "bbox": [ + 212, + 747, + 782, + 821 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s c o r e}} ^ {p} (o) = - \\log o \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 436, + 825, + 784, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Zhan et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To avoid treating all queries that do not match the extracted objects as negative samples, a careful selection process is employed. Instead of using all non-matching queries, negative samples are chosen based on queries that have low fitness scores, denoted by $s$ . This approach ensures that the scorer learns to differentiate between suitable and unsuitable object placements, focusing on the most informative negative samples.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s c o r e}} ^ {n} (o) = - \\left(1 - \\mathbb {1} \\left[ s _ {p} \\cdot s _ {o} \\cdot s _ {c} > \\tau \\right]\\right) \\log (1 - o) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 250, + 785, + 267 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The fitness scores $s_p$ , $s_o$ , and $s_c$ are used to measure the accuracy and coherence of an object's position, orientation, and category within the scene. The position score $s_p = \\exp(-(\\mathbf{b}_{xyz} \\cdot \\mathbf{n}_g + d_g) / \\sigma_p^2)^1$ penalizes the predicted location $\\mathbf{b}_{xyz}$ deviating from the local ground plane specified by $(\\mathbf{n}_g, d_g)$ . Local ground planes here could be obtained via non-learning ground detection methods like Patch-Works++ [24]. Similarly, the orientation score $s_o = \\cos((\\mathbf{b}_\\theta - \\hat{\\mathbf{b}}_\\theta) / 2)$ encourages consistency between the heading of the new object location and the predominant heading $\\hat{\\mathbf{b}}_\\theta$ of objects in the same category within the scene. Finally, the category compatible score $s_c = \\mathbb{1}[\\mathbf{b}_c \\in \\mathcal{C}]$ prevents pasting object of category that is not in the current scene (e.g. bicycle on highway). Notably, these scores can be computed not only between the extracted object and its original scan but also between the extracted object and other unannotated LiDAR scans, in which $\\hat{\\mathbf{b}}_\\theta$ can be simplified as ego orientation and $s_c$ can be ignored directly. This adaptability opens up the possibility of learning object placement scoring from a more extensive dataset through unsupervised learning, which assigns positive and negative samples solely based on the fitness scores.", + "bbox": [ + 212, + 277, + 787, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Combining these asymmetric losses and carefully selecting positive and negative samples, our HotspotNet learns to accurately position objects within a new scan while maintaining semantic coherence.", + "bbox": [ + 212, + 518, + 787, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Cross-Scan Object Transfer for LiDAR-based SSOD", + "text_level": 1, + "bbox": [ + 214, + 587, + 689, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Under the guidance of HotspotNet (described in Sec. 3.1), we inject precise supervision signals into unannotated point clouds and train detectors on these partially-labeled data. Fig. 3 illustrates the differences between our Cross-Scan Object Transfer SSOD paradigm and the commonly used self-training technique. Our method bypasses the iterative self-training procedures, cumbersome pseudolabel filtering mechanisms, and complex loss objective designs.", + "bbox": [ + 212, + 612, + 787, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Partially Supervised Object Detection Our approach for introducing supervision into unlabeled LiDAR scans is straightforward. We denote all objects in labeled point clouds as $\\{\\mathbf{O}^L\\}$ , and all unlabeled scans as $\\{\\mathbf{S}^U\\}$ . For an unlabeled LiDAR scan $\\mathbf{S}^U$ , we randomly sample a subset of $\\{\\mathbf{O}^L\\}$ and follow the instruction of HotspotNet to inject $\\{\\mathbf{O}^L\\}$ into $\\mathbf{S}^U$ in a semantically coherent manner. Since inserted objects come with ground-truth classes, poses and dimensions,", + "bbox": [ + 212, + 724, + 787, + 818 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection", + "bbox": [ + 238, + 113, + 730, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$^{1}$ $\\mathbf{b}_{\\square}$ denotes the $\\square$ component for location prediction $\\mathbf{b}$ .", + "bbox": [ + 217, + 824, + 602, + 840 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/85c682fbbcaea53f860e6d584522858f6ebb12f03a2ccf8115d128d833ec7bcd.jpg", + "image_caption": [ + "Fig. 3: (a) Common-used pseudo-labeling SSOD paradigm. (b) Cross-Scan Object Transfer SSOD paradigm, which bypasses the iterative self-training procedures, cumbersome pseudo-labels filter mechanisms and complex loss objectives designs." + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 771, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "these newly created scans allow for supervision. However, it is important to note that the above synthesized point clouds are only partially annotated, as we lack information about the original unlabeled objects in these unannotated scans. Learning a detector from partially labeled data is known as Sparsely Annotated Object Detection (SAOD) [30, 41]. The differences of training a detector from fully and partially annotated data will be discussed in the next part.", + "bbox": [ + 212, + 481, + 787, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Spatial-Aware Classification Loss An object detector generally includes a classifier to differentiate foreground objects from background clutter, and a localizer to predict the accurate location, dimension and orientation of an object. The localizer is typically supervised by regressing the offset from current voxel location to actual bounding boxes of foreground objects, so it can be naturally supervised by new point clouds generated via cross-scan object transfer. However, the classifier requires both positive and negative samples for learning a decision boundary. Once again, the inserted objects could provide positive samples for the classifier but we can not treat all other voxels as background in SAOD just like vanilla object detection. This is due to the existence of the so-called hard negatives as reported in other SAOD works [41].", + "bbox": [ + 212, + 597, + 787, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our approach, hard negatives refer to the original unlabeled objects intrinsically present in the unannotated LiDAR scans. Training the classifier in our semi-supervised detector with these hard negatives will result in an unacceptable number of false negatives. Instead of resorting to sophisticated re-weighting/ recalibration strategies [44,59] or teacher-student siamese networks [41], we simply", + "bbox": [ + 212, + 765, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Zhan et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "choose negative samples in a spatial-aware manner to avoid sampling unlabeled objects as negatives. We propose a spatial-aware classification loss for our partial supervision as Eq.6:", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {c l s}} ^ {\\mathrm {n}} (\\mathbf {x}) = \\mathbb {1} \\left[ \\left(\\min _ {i} \\| \\mathbf {x} - \\mathbf {c} _ {i} \\| _ {2}\\right) \\leq \\sigma \\right] \\cdot L _ {\\mathrm {c l s}} ^ {\\mathrm {n}} (\\mathbf {x}) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 199, + 785, + 219 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $L_{\\mathrm{cls}}^{\\mathrm{n}}(\\mathbf{x})$ denotes focal loss for negative samples at the heatmap location $\\mathbf{x}$ . The indicator function $\\mathbb{1}(\\cdot)$ equals 1 when the $\\ell 2$ distance between a location $\\mathbf{x}$ and the center location $\\mathbf{c}$ of any foreground object is less than a threshold $\\sigma$ , and 0 otherwise. We set $\\sigma$ as twice of the Gaussian radius used for positive samples assignment in CenterPoint [54] for unlabeled scans, and infinity for fully annotated data. An immediate drawback of our spatial-aware classification loss is that there are too few negative samples for the classifier. To alleviate this issue, we randomly sample $1\\%$ samples in the heatmap predictions as negative samples, and set the corresponding value of $\\mathbb{1}(\\cdot)$ as 1 to make the classification boundary between positive and negative samples more clear.", + "bbox": [ + 212, + 224, + 787, + 377 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Synergy with Self-Training Techniques A deeper analysis of the proposed CSOT SSOD paradigm reveals that the primary mechanism driving improved detection performance is the extrapolation of labeled object detection to novel contextual environments. This stands in contrast to self-training methods, which learn from pseudo-labeled objects in unannotated scans.", + "bbox": [ + 212, + 393, + 782, + 468 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As a result, the proposed CSOT pipeline for semi-supervised LiDAR object detection is orthogonal to, yet complementary with, existing self-training techniques [37], allowing for additional performance gains. Deliberately designing a pseudo-labeling SSOD framework is beyond the scope of this paper. To verify the complementary effect of our approach when combined with self-training techniques, we implement a basic teacher-student paradigm. We employ a single universal confidence score threshold $\\tau_{\\mathrm{pseudo}}$ , which is invariant across categories, to filter low-confidence predictions. We then utilize a fully-supervised loss, as in CenterPoint [54], for optimizing the student detector.", + "bbox": [ + 212, + 469, + 787, + 607 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 215, + 625, + 366, + 641 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The proposed methods are verified in this section. We show implementation details in Sec. 4.1 and provide all experimental results in Sec. 4.2 and Sec. 4.3. More ablations and discussions are shown in our supplementary materials.", + "bbox": [ + 212, + 650, + 782, + 696 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 713, + 426, + 729 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets We validate our method on two widely used autonomous driving datasets: nuScenes [3] and Waymo [38]. The nuScenes dataset has 700 training, 150 validation, and 150 test sequences. Official nuScenes detection metrics are NuScenes Detection Score (NDS) and mean Average Precision (mAP). The Waymo Open Dataset consists of 798 training and 202 validation sequences. Official Waymo detection metrics are mean Average Precision (mAP) and mean Average Precision weighted by Heading (mAPH).", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection", + "bbox": [ + 238, + 114, + 730, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training Details We use one-stage CenterPoint [54] as our baseline detector to validate the effectiveness of our methods. Voxel sizes are set as $[0.075\\mathrm{m}, 0.075\\mathrm{m}, 0.2\\mathrm{m}]$ for nuScenes and $[0.1\\mathrm{m}, 0.1\\mathrm{m}, 0.15\\mathrm{m}]$ for Waymo. The training epochs for one cycle of learning procedure, including fully-supervised learning, partially-supervised learning and self-training, are set as 20 for nuScenes and 7 for Waymo. We follow many previous work [6,51,54,66] to adopt Adam optimizer with one-cycle learning rate policy. Labeled data are uniformly sampled from the entire dataset. Then objects and their annotations are extracted to build the ground-truth database for injecting precise supervision signals into unlabeled scans. We finetune partially-supervised detectors on labeled scans with several epochs. We set a score threshold of 0.3 for nuScenes and 0.6 for Waymo to filter low-quality pseudo labels, and verify the orthogonality of our methods with self-training.", + "bbox": [ + 212, + 146, + 787, + 329 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Benchmarks", + "text_level": 1, + "bbox": [ + 215, + 351, + 364, + 364 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Benchmarks on nuScenes dataset We compare our methods against other label-efficient LiDAR detection approaches on the nuScenes dataset. As shown in Tab. 1, our methods obtain the significant performance improvements over previous works. For the sake of fairness, we choose the fully supervised detector trained by Yin et.al. [54] (the designer of CenterPoint backbone) as the baseline detector, and list the results achieved from other label-efficient LiDAR object detection approaches for comparison [27, 43, 56, 58]. Without leveraging models pre-trained on external datasets and using only $5\\%$ of the annotations, we achieve competitive detector which reaches $97.2\\%$ of the fully-supervised performance on mAP and $98.9\\%$ on NDS. We significantly shrink the mAP and NDS gaps between the baseline detector and the one trained with $20\\mathrm{x}$ less annotations from around $10\\sim 20$ to merely 1.6 and 0.7 respectively. Our semi-supervised detector, trained with only $5\\%$ manual annotations, obtains over $90\\%$ detection performance on 8 of the 10 categories defined in nuScenes dataset, and over $95\\%$ detection performance on 6 categories.", + "bbox": [ + 212, + 378, + 787, + 604 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Benchmarks on Waymo dataset We extend our methods on Waymo dataset and achieve similar outstanding results. we choose the fully supervised detector trained by Shi et.al. [36], Yan et.al. [47], Fan et.al. [10] and Yin et.al. [54] as the baseline detector, and list the results achieved from other label-efficient LiDAR object detection approaches for comparison [29, 46, 50, 52, 53, 56]. As shown in Tab. 2, our methods are valid on both anchor-based (SECOND) and anchor-free (CenterPoint) detectors. Taking CenterPoint as an example, with merely $1\\%$ annotations, we achieve a competitive detector which matches fully supervised performance on vehicle and pedestrian, and exceeds it on cyclist. The $100\\mathrm{x}$ reduction in annotations leads to only a minor 1.8 APH decrease for L2 Vehicle, a negligible 0.4 APH drop for L2 Pedestrian, and boosts the APH of L2 Cyclist by 4.1. The abnormal enhanced detection performance on Cyclist should be attributed to our learning-based approach for cross-scan object transfer. Instructed by the proposed HotspotNet, we paste objects in a semantic coherent", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Zhan et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/d803aba2d072488424b08e9c437d06f25d33a7e59bddcf6ea4e4e0b6fe5d9ce8.jpg", + "table_caption": [ + "Table 1: Compare with other labeled efficient LiDAR object detection approaches on nuScenes Dataset validation split. Performances of the baseline detector reported by its original designer [54] are marked with gray backgrounds." + ], + "table_footnote": [], + "table_body": "
MethodRefAnnosExternalmAPNDSCarTruckC.V.BusTrailerBarrierMot.Byc.Ped.T.C.
BaselineCVPR 2021 [54]100%58.065.584.651.017.560.253.270.953.728.783.476.7
GCC3DICCV 2021 [27]5%41.146.8——————————
SceneFlowECCV 2022 [58]5%36.048.3——————————
SSDA3DAAAI 2023 [43]5%✓——76.2—————————
AD-PTNeurIPS 2023 [56]5%✓45.053.078.943.811.155.221.255.139.017.872.355.4
Ours—5%56.464.881.753.516.558.632.866.060.047.882.065.5
", + "bbox": [ + 218, + 198, + 784, + 275 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/bf4377f933e2e2eed69183592041761d241a69cc54f3b57c5edd8fdc09ffd648.jpg", + "table_caption": [ + "Table 2: Compare with other labeled efficient LiDAR object detection approaches on Waymo Dataset validation split. Performances of the baseline detector reported by its original designers [10,36,47,54] are mark with gray backgrounds." + ], + "table_footnote": [], + "table_body": "
DetectorMethodRefAnnoxExtVehiclePedestrianCyclistAll L2 AP
L1L2L1L2L1L2
APAPHAPAPHAPAPHAPAPHAPAPHAPAPH
PV-RCNNBaselineCVPR 2020 [36]100%77.576.969.068.475.065.766.057.667.866.465.464.063.3
PV-RCNNHSSDACVPR 2023 [29]1%56.453.849.747.340.120.933.517.529.120.927.920.028.3
PV-RCNNMIXSUPICLR 2024 [50]1%———55.5———52.0———62.356.6
SECONDBaselineSensors 2018 [47]100%72.371.763.963.368.758.260.751.360.659.358.357.157.2
SECONDMIXSUPICLR 2024 [50]10%———55.0———49.6———58.154.2
SECONDProficientTeacherECCV 2022 [52]5%——53.052.5——50.338.7——49.946.045.7
SECONDAD-PTNeurIPS 2023 [56]3%✓——60.559.9——54.945.8——50.849.751.8
SECONDProposalContrastECCV 2022 [53]1%——37.636.9——39.731.7——37.735.734.8
SECONDOurs—1%69.468.961.260.864.060.755.152.361.160.458.958.257.1
SSTBaselineCVPR 2022 [10]100%74.273.865.565.178.769.670.061.770.769.668.066.964.6
SSTMIXSUPICLR 2024 [50]10%———59.1———60.0———63.160.7
SSTMV-JARCVPR 2023 [46]5%——56.556.0——57.747.7——37.436.346.7
CenterPointBaselineCVPR 2021 [54]100%76.676.068.968.479.073.471.065.872.171.069.568.567.6
CenterPointMIXSUPICLR 2024 [50]10%———61.8———57.7———67.562.3
CenterPointAD-PTNeurIPS 2023 [56]3%✓——60.459.8——60.654.0——62.761.658.5
CenterPointOurs—1%75.174.767.066.675.972.968.265.475.975.173.472.668.2
", + "bbox": [ + 220, + 340, + 782, + 507 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "manner, which enables our CSOT SSOD to reach its full potential and meanwhile can be used as a powerful data augmentation for fully-supervised detector. Related ablation results are discussed in Sec. 4.3. Notably, under the most stringent constraints with minimal labeled point clouds and without leveraging any external pre-training, our approach stands alone as the only label-efficient technique which is able to achieve equivalence detection performance to the baseline detector on the average L2 APH of all categories defined in Waymo dataset.", + "bbox": [ + 212, + 536, + 787, + 643 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Ablations", + "text_level": 1, + "bbox": [ + 215, + 664, + 343, + 678 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ablate training stage on nuScenes dataset. Our semi-supervised detectors trained with 20x fewer nuScenes annotations are compared to both fully supervised and pseudo-labeling semi-supervised models. Results in Tab. 3 indicate that our newly proposed CSOT is a valid label-efficient approach, and it almost completely maintains the performance gains of self-training when used together. The fully supervised detector trained with $5\\%$ labeled scans shows substantial performance degradation compared to that trained on the entire labeled dataset, with a mAP decline from 59.5 to 43.8 and a NDS decrease from 66.0 to 56.5. As shown in Lines 3-5 of Tab. 3, self-training technique is able to boost the mAP from 43.8 to 51.0 and NDS from 56.5 to 61.0. Our CSOT, which allows", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection", + "bbox": [ + 238, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/ba2bd1af06fa71b7a599f3362493800f65ba680963c8ecdb279d875d742aff07.jpg", + "table_caption": [ + "Table 3: Compare with fully supervised and pseudo-labeling semi-supervised detectors. All results are achieved based on our implementation code of CenterPoint and evaluated on nuScenes dataset validation split." + ], + "table_footnote": [], + "table_body": "
AnnosScansCSOTSelf-TrainingmAPNDSCarTruckC.V.BusTrailerBarrierMot.Byc.Ped.T.C.
1100%100%59.566.083.756.618.766.934.867.163.750.484.568.7
25%5%43.856.575.839.712.344.120.556.936.318.476.756.9
35%100%✓51.061.080.149.114.655.626.762.345.430.881.963.1
45%100%✓50.461.079.646.411.750.125.659.152.137.280.662.0
55%100%✓✓56.464.881.753.516.558.632.866.060.047.882.065.5
", + "bbox": [ + 218, + 199, + 784, + 266 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/8e70eb88e4899ecc4525c6d503673f221fc568b04b544e2981155b2c38971296.jpg", + "table_caption": [ + "Table 4: Compare with fully supervised and pseudo-labeling semi-supervised detectors. All results are achieved based on our implementation code of CenterPoint and evaluated on Waymo dataset validation split (mean L1/L2 AP/APH over all categories defined in Waymo dataset)." + ], + "table_footnote": [], + "table_body": "
AnnosScansCSOTSelf-TrainingObject TransferL1 AP/APHL2 AP/APHmAP/mAPH
1100%100%—74.4 / 72.768.6 / 67.071.5 / 69.9
1a100%100%GT-Aug74.7 / 73.168.8 / 67.371.8 / 70.2
1b100%100%Ours77.5 / 76.071.6 / 70.274.6 / 73.1
21%1%—68.3 / 66.662.3 / 60.865.3 / 63.7
31%100%✓—71.9 / 70.465.8 / 64.568.9 / 67.5
41%100%✓Ours71.4 / 69.865.4 / 63.968.4 / 66.9
51%100%✓✓Ours75.7 / 74.269.5 / 68.272.6 / 71.2
", + "bbox": [ + 220, + 349, + 782, + 434 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "supervision on unannotated point clouds by converting unlabeled scans into partially labeled scans via learning-based copy-paste, achieves a similar detection performance (with a mAP of 50.4 and a NDS of 61.0). It is noteworthy that iterative employment of self-training techniques does not lead to sustained gains in detection performance (see supplementary material), while incorporating just one cycle of self-training on our partially supervised detector is able to further optimize the mAP from 50.4 to 56.4 and NDS from 61.0 to 64.8.", + "bbox": [ + 212, + 465, + 787, + 571 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablate training stage on Waymo dataset We extend the above ablations on Waymo dataset and summarize all results in Tab. 4, which further validate the effectiveness of CSOT, as well as its orthogonality with the common used pseudo-labeling techniques. $100\\mathrm{x}$ fewer labeled point clouds lead to a degradation of detection performance, reducing the mean AP and APH from 71.5, 69.9 to 65.3, 63.7. Our Cross-Scan Object Transfer paradigm brings a boost of around 3 APH for both LEVEL_1 and LEVEL_2 objects. CSOT is proved to be complimentary to self-training, and the performance gains can be completely maintained. As shown in the Line 5 of Tab. 4, the performance of detector trained with $1\\%$ annotations can be finally optimized to match even surpass the fully-supervised baseline model. The abnormal enhanced detection performance with $100\\mathrm{x}$ reduced labeled data should be attributed to the learning-based copy-paste operation employed in this work. As shown in Lines 1a-1b of Tab. 4, replacing the commonly used copy-paste augmentation (GT-Aug) with our learning based approaches boosts the fully-supervised detector with an increase of $2.8~\\mathrm{mAP}$ . Notably, even regarding the fully-supervised detector augmented via our object", + "bbox": [ + 212, + 598, + 787, + 843 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Zhan et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/107a38255d1306e835e526aba177ea89dcd5b0d1ef1ea3f7f8e44821fbcf9a47.jpg", + "table_caption": [ + "Table 5: Ablate our learning-based cross-scan object transfer which is instructed by HotspotNet. All results are evaluated on nuScenes dataset validation split." + ], + "table_footnote": [], + "table_body": "
(a) Compare our learning-based cross-scan object transfer (instructed by HotspotNet) with other similar methods.
AnnosCSOTObj TransferSelfObj Appearance rαLocal Context (x,y,z)Global Context rsclsrule /learnedmAP NDS
15%——————43.8 56.5
25%✓——————51.0 61.0
35%✓GT-Aug [47]✓✓46.0 57.7
45%✓CA-Aug [15]✓✓✓rule46.0 58.0
55%✓LiDAR-Aug [11]✓✓✓rule46.2 57.9
65%✓Ours✓✓✓✓✓learned50.4 61.0
6a5%✓Ours✓✓✓✓✓rule49.7 60.6
75%✓GT-Aug [47]✓✓49.7 60.5
85%✓CA-Aug [15]✓✓✓rule49.7 60.7
95%✓LiDAR-Aug [11]✓✓✓rule49.7 60.2
105%✓Ours✓✓✓✓✓learned56.4 64.8
10a5%✓Ours✓✓✓✓✓rule53.6 63.3
(b) Ablate contribution of each fitness score defined in Ln-score(o) (Eq. 5).
AnnosCSOTSelfposition sporientation socategory scmAP NDS
115%✓✓49.7 60.5
125%✓✓✓52.1 62.6
135%✓✓✓49.4 60.8
13a5%✓✓✓52.2 62.3
13b5%✓✓✓✓55.1 64.3
145%✓✓✓51.6 62.0
155%✓✓✓✓✓56.4 64.8
", + "bbox": [ + 218, + 185, + 782, + 441 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "transfer implementation as the baseline, reducing annotations by $100\\mathrm{x}$ causes only a minimal degradation in performance (remains $97.3\\% / 97.4\\%$ mAP/mAPH of our fully-supervised detector).", + "bbox": [ + 212, + 470, + 784, + 516 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablate our learning-based object transfer. We compare our learning-based object transfer which instructed by the proposed HotspotNet with other similar techniques in Tab. 5(a), and ablate negative sample strategy employed for calculating $\\mathcal{L}_{\\mathrm{score}}^n (o)$ (Eq. 5) in Tab. 5(b). As shown in Lines 3-10 of Tab. 5, the proposed learning-based object transfer is able to maximize the benefits of CSOT. It also stands alone in its ability to seamlessly integrate with self-training technique without compromising their performance gains. CSOT built upon our learning-based object transfer enhances the detection performance with a considerable $6.6\\mathrm{mAP}$ and 4.5 NDS. The improvement surpasses the one built upon previous similar approaches (with a margin of about $4\\mathrm{mAP}$ and 3 NDS). The poor partially-supervised detector achieved from other object transfer approaches produce enormous false positive predictions (visualized in Fig. 1) and greatly influence the downstream self-training. Replacing our learning-based object transfer with previous methods reduces the mAP gains of self-training from 6.0 to around 3.5, and degrades the final mAP from 56.4 to 49.7. The necessity of learning-based object transfer are verified in Tab. 5, Lines 6-6a,10-10a. Learn-based approach enables better generalization and can help reduce noise, which introduces an increased final mAP of 2.8. We further ablate negative sample strategy employed for calculating $\\mathcal{L}_{\\mathrm{score}}^n (o)$ and list results in Lines 11-15 of Tab. 5. Pocation and category provide the most important contextual informa", + "bbox": [ + 212, + 537, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection", + "bbox": [ + 238, + 113, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b5faac4a2c5b301fdc9871036eaafa02f2b85c02a9438d9d4a2b00bd1fb51ce8.jpg", + "table_caption": [ + "Table 6: Ablate robustness of our methods across various network architectures. All results are evaluated on nuScenes dataset validation split." + ], + "table_footnote": [], + "table_body": "
AnnosDetectorEncoderAnchor-baseAnchor-freeVoxel SizeCSOTSelfmAPNDS
1100%CenterPointVFE✓(0.075,0.075,0.2)59.566.0
25%CenterPointVFE✓(0.075,0.075,0.2)43.856.5
35%CenterPointVFE✓(0.075,0.075,0.2)✓51.061.0
45%CenterPointVFE✓(0.075,0.075,0.2)✓50.461.0
55%CenterPointVFE✓(0.075,0.075,0.2)✓✓56.464.8
6100%CenterPointPFE✓(0.1,0.1,8.0)49.260.1
75%CenterPointPFE✓(0.1,0.1,8.0)30.248.3
85%CenterPointPFE✓(0.1,0.1,8.0)✓37.853.1
95%CenterPointPFE✓(0.1,0.1,8.0)✓40.054.9
105%CenterPointPFE✓(0.1,0.1,8.0)✓✓45.057.9
11100%SECONDVFE✓(0.1,0.1,0.2)55.464.7
125%SECONDVFE✓(0.1,0.1,0.2)38.453.2
135%SECONDVFE✓(0.1,0.1,0.2)✓45.258.1
145%SECONDVFE✓(0.1,0.1,0.2)✓45.459.1
155%SECONDVFE✓(0.1,0.1,0.2)✓✓48.661.3
", + "bbox": [ + 220, + 185, + 782, + 356 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "tion, orientation provides additional cues to narrow down the possible identity. Specifically, combining orientation constraint with category consistency improve mAP by 0.6 (Tab. 5, Lines 13a and 14). Removing orientation constraint from all reduces mAP by 1.3 (Tab. 5, Lines 13b and 15).", + "bbox": [ + 215, + 383, + 784, + 445 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Extend our method to different detectors. We confirm the generality of our methods on various detectors, which utilizes diverse point cloud feature encoder or target assigner strategy. All results are listed in Tab. 6. For an anchor-free detector CenterPoint with a pillar-based feature encoder, the proposed CSOT bring a boost of $9.8\\mathrm{mAP}$ and 6.6 NDS, which outperforms the self-training technique by $2.2\\mathrm{mAP}$ and 1.8 NDS. Using our partially-supervised detector as the teacher model to generate pseudo-labels for unlabeled point clouds, and then conduct one-cycle of self-training. The performance of semi-supervised detector is ultimately reaches $45.0\\mathrm{mAP}$ and 57.9 NDS. The mAP gap introduced by $20x$ fewer annotations is narrowed from 19.0 to 4.2. Experiments on anchor-based detector (SECOND) also shows similar results. The effectiveness of CSOT and the additive effect of combining it with self-training techniques are clearly demonstrated in Lines 11-15 of Tab. 6.", + "bbox": [ + 215, + 467, + 785, + 662 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 217, + 686, + 356, + 703 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we propose a novel Cross-Scan Object Transfer paradigm for LiDAR semi-supervised object detection. It features a HotspotNet for transferring objects semantic coherence manner, which reduces false positive predictions and enables the proposed CSOT to reach its full potential. To the best of our knowledge, this work is the first successful attempt of object transfer in LiDAR SSOD. The proposed CSOT paradigm is also orthogonal to the self-training techniques. We achieve competitive detectors on par with fully-supervised baselines while using 20x and 100x reduced annotations on nuScenes and Waymo dataset.", + "bbox": [ + 215, + 718, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Zhan et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the InnoHK funding.", + "bbox": [ + 215, + 178, + 576, + 194 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 219, + 321, + 234 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In: CVPR. pp. 1090-1099 (2022)", + "2. Blum, A., Mitchell, T.: Combining labeled and unlabeled data with co-training. In: COLT. pp. 92-100 (1998)", + "3. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR. pp. 11621-11631 (2020)", + "4. Chen, B., Chen, W., Yang, S., Xuan, Y., Song, J., Xie, D., Pu, S., Song, M., Zhuang, Y.: Label matching semi-supervised object detection. In: CVPR. pp. 14381-14390 (2022)", + "5. Chen, B., Li, P., Chen, X., Wang, B., Zhang, L., Hua, X.S.: Dense learning based semi-supervised object detection. In: CVPR. pp. 4815-4824 (2022)", + "6. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Largekernel3d: Scaling up kernels in 3d sparse cnns. In: CVPR. pp. 13488-13498 (2023)", + "7. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In: CVPR. pp. 21674-21683 (2023)", + "8. Choi, J., Song, Y., Kwak, N.: Part-aware data augmentation for 3d object detection in point cloud. In: IROS. pp. 3391-3397 (2021)", + "9. Dwibedi, D., Misra, I., Hebert, M.: Cut, paste and learn: Surprisingly easy synthesis for instance detection. In: ICCV. pp. 1301-1310 (2017)", + "10. Fan, L., Pang, Z., Zhang, T., Wang, Y.X., Zhao, H., Wang, F., Wang, N., Zhang, Z.: Embracing single stride 3d object detector with sparse transformer. In: CVPR. pp. 8458-8468 (2022)", + "1. Fang, J., Zuo, X., Zhou, D., Jin, S., Wang, S., Zhang, L.: Lidar-aug: A general rendering-based augmentation framework for 3d object detection. In: CVPR. pp. 4710-4720 (2021)", + "2. Gao, J., Wang, J., Dai, S., Li, L.J., Nevatia, R.: Note-rcnn: Noise tolerant ensemble rcnn for semi-supervised object detection. In: ICCV. pp. 9508-9517 (2019)", + "3. Ghiasi, G., Cui, Y., Srinivas, A., Qian, R., Lin, T.Y., Cubuk, E.D., Le, Q.V., Zoph, B.: Simple copy-paste is a strong data augmentation method for instance segmentation. In: CVPR. pp. 2918-2928 (2021)", + "4. Hu, J.S.K., Waslander, S.L.: Pattern-aware data augmentation for lidar 3d object detection. In: IEEE International Intelligent Transportation Systems Conference (ITSC). pp. 2703-2710 (2021)", + "5. Hu, X., Duan, Z., Ma, J.: Context-aware data augmentation for lidar 3d object detection. In: ICIP. pp. 11-15 (2023)", + "6. Hu, Y., Ding, Z., Ge, R., Shao, W., Huang, L., Li, K., Liu, Q.: Afdetv2: Rethinking the necessity of the second stage for object detection from point clouds. In: AAAI. pp. 969-979 (2022)", + "7. Huang, D., Chen, Y., Ding, Y., Liao, J., Liu, J., Wu, K., Nie, Q., Liu, Y., Wang, C., Li, Z.: Rethinking dimensionality reduction in grid-based 3d object detection. arXiv preprint arXiv:2209.09464 (2022)" + ], + "bbox": [ + 225, + 252, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection", + "bbox": [ + 238, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Jaderberg, M., Simonyan, K., Vedaldi, A., Zisserman, A.: Synthetic data and artificial neural networks for natural scene text recognition. arXiv preprint arXiv:1406.2227 (2014)", + "19. Kingma, D.P., Rezende, D.J., Mohamed, S., Welling, M.: Semi-supervised learning with deep generative models. In: Adv. Neural Inform. Process. Syst. (2014)", + "20. Laine, S., Aila, T.: Temporal ensembling for semi-supervised learning. In: ICLR (2017)", + "21. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J.: Pointpillars: Fast encoders for object detection from point clouds. In: CVPR. pp. 12697-12705 (2019)", + "22. Lee, D., Lee, J., Lee, J., Lee, H., Lee, M., Woo, S., Lee, S.: Regularization strategy for point cloud via rigidly mixed sample. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 15900-15909 (2021)", + "23. Lee, D.H.: Pseudo-label : The simple and efficient semi-supervised learning method for deep neural networks. In: ICML. p. 896 (2013)", + "24. Lee, S., Lim, H., Myung, H.: Patchwork++: Fast and robust ground segmentation solving partial under-segmentation using 3d point cloud. In: IROS. pp. 13276-13283 (2022)", + "25. Li, E., Wang, S., Li, C., Li, D., Wu, X., Hao, Q.: Sustech points: A portable 3d point cloud interactive annotation platform system. IEEE Intelligent Vehicles Symposium (IV) pp. 1108-1115 (2020)", + "26. Li, H., Wu, Z., Shrivastava, A., Davis, L.S.: Rethinking pseudo labels for semi-supervised object detection. In: AAAI. pp. 1314-1322 (2022)", + "27. Liang, H., Jiang, C., Feng, D., Chen, X., Xu, H., Liang, X., Zhang, W., Li, Z., Gool, L.V.: Exploring geometry-aware contrast and clustering harmonization for self-supervised 3d object detection. In: ICCV. pp. 3293-3302 (2021)", + "28. Liu, C., Zhang, W., Lin, X., Zhang, W., Tan, X., Han, J., Li, X., Ding, E., Wang, J.: Ambiguity-resistant semi-supervised learning for dense object detection. In: CVPR. pp. 15579-15588 (2023)", + "29. Liu, C., Gao, C., Liu, F., Li, P., Meng, D., Gao, X.: Hierarchical supervision and shuffle data augmentation for 3d semi-supervised object detection. In: CVPR. pp. 23819-23828 (2023)", + "30. Liu, C., Gao, C., Liu, F., Liu, J., Meng, D., Gao, X.: Ss3d: Sparsely-supervised 3d object detection from point cloud. In: CVPR. pp. 8428-8437 (2022)", + "31. Liu, Y.C., Ma, C.Y., He, Z., Kuo, C.W., Chen, K., Zhang, P., Wu, B., Kira, Z., Vajda, P.: Unbiased teacher for semi-supervised object detection. In: ICLR (2021)", + "32. Liu, Y.C., Ma, C.Y., Kira, Z.: Unbiased teacher v2: Semi-supervised object detection for anchor-free and anchor-based detectors. In: CVPR. pp. 9819-9828 (2022)", + "33. Meng, Q., Wang, W., Zhou, T., Shen, J., Jia, Y., Gool, L.V.: Towards a weakly supervised framework for 3d point cloud object detection and annotation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 44(8), 4454-4468 (2021)", + "34. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR. pp. 652-660 (2017)", + "35. Qian, G., Li, Y., Peng, H., Mai, J., Hammoud, H.A.A.K., Elhoseiny, M., Ghanem, B.: Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In: Adv. Neural Inform. Process. Syst. pp. 23192-23204 (2022)", + "36. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In: CVPR. pp. 10529–10538 (2020)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Zhan et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Sohn, K., Zhang, Z., Li, C.L., Zhang, H., Lee, C.Y., Pfister, T.: A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757 (2020)", + "38. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhao, S., Cheng, S., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: CVPR. pp. 2446-2454 (2020)", + "39. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In: Adv. Neural Inform. Process. Syst. (2017)", + "40. Wang, H., Cong, Y., Litany, O., Gao, Y., Guibas, L.J.: 3dioumatch: Leveraging iou prediction for semi-supervised 3d object detection. In: CVPR. pp. 14615-14624 (2021)", + "41. Wang, T., Yang, T., Cao, J., Zhang, X.: Co-mining: Self-supervised learning for sparsely annotated object detection. In: AAAI. pp. 2800-2808 (2021)", + "42. Wang, X., Yang, X., Zhang, S., Li, Y., Feng, L., Fang, S., Lyu, C., Chen, K., Zhang, W.: Consistent-teacher: Towards reducing inconsistent pseudo-targets in semi-supervised object detection. In: CVPR. pp. 3240-3249 (2023)", + "43. Wang, Y., Yin, J., Li, W., Frossard, P., Yang, R., Shen, J.: Ssda3d: Semi-supervised domain adaptation for 3d object detection from point cloud. In: AAAI. pp. 2707-2715 (2023)", + "44. Wu, Z., Bodla, N., Singh, B., Najibi, M., Chellappa, R., Davis, L.S.: Soft sampling for robust object detection. arXiv preprint arXiv:1806.06986 (2018)", + "45. Xu, M., Zhang, Z., Hu, H., Wang, J., Wang, L., Wei, F., Bai, X., Liu, Z.: End-to-end semi-supervised object detection with soft teacher. In: ICCV. pp. 3060-3069 (2021)", + "46. Xu, R., Wang, T., Zhang, W., Chen, R., Cao, J., Pang, J., Lin, D.: Mv-jar: Masked voxel jigsaw and reconstruction for lidar-based self-supervised pre-training. In: CVPR. pp. 13445-13454 (2023)", + "47. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)", + "48. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d: Self-training for unsupervised domain adaptation on 3d object detection. In: CVPR. pp. 10368-10378 (2021)", + "49. Yang, Q., Wei, X., Wang, B., Hua, X.S., Zhang, L.: Interactive self-training with mean teachers for semi-supervised object detection. In: CVPR. pp. 5941-5950 (2021)", + "50. Yang, Y., Fan, L., Zhang, Z.: Mixsup: Mixed-grained supervision for labeled efficient lidar-based 3d object detection. In: ICLR (2024)", + "51. Ye, D., Zhou, Z., Chen, W., Xie, Y., Wang, Y., Wang, P., Foroosh, H.: Lidar-multinet: Towards a unified multi-task network for lidar perception. In: AAAI. pp. 3231-3240 (2023)", + "52. Yin, J., Fang, J., Zhou, D., Zhang, L., Xu, C.Z., Shen, J., Wang, W.: Semi-supervised 3d object detection with proficient teachers. In: ECCV. pp. 727-743 (2022)", + "53. Yin, J., Zhou, D., Zhang, L., Fang, J., Xu, C.Z., Shen, J., Wang, W.: Proposal-contrast: Unsupervised pre-training for lidar-based 3d object detection. In: ECCV. pp. 17-33 (2022)", + "54. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: CVPR. pp. 11784-11793 (2021)" + ], + "bbox": [ + 215, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection", + "bbox": [ + 238, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "55. You, Y., Luo, K., Phoo, C.P., Chao, W.L., Sun, W., Hariharan, B., Campbell, M., Weinberger, K.Q.: Learning to detect mobile objects from lidar scans without labels. In: CVPR. pp. 1130-1140 (2022)", + "56. Yuan, J., Zhang, B., Yan, X., Chen, T., Shi, B., Li, Y., Qiao, Y.: Ad-pt: Autonomous driving pre-training with large-scale point cloud dataset. In: Adv. Neural Inform. Process. Syst. p. 36 (2023)", + "57. Yun, S., Han, D., Oh, S.J., Chun, S., Cho, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: CVPR. pp. 6023-6032 (2019)", + "58. Yurtsever, E., Erçelik, E., Liu, M., Yang, Z., Zhang, H., Topçam, P., Listl, M., Çayli, Y.K., Knoll, A.: 3d object detection with a self-supervised lidar scene flow backbone. In: ECCV. pp. 247-265 (2022)", + "59. Zhang, H., Chen, F., Shen, Z., Hao, Q., Zhu, C., Savvides, M.: Solving missing-annotation object detection with background recalibration loss. In: ICASSP. pp. 1888-1892 (2020)", + "60. Zhang, L., Sun, Y., Wei, W.: Mind the gap: Polishing pseudo labels for accurate semi-supervised object detection. In: AAAI. pp. 3463-3471 (2023)", + "61. Zhang, L., Yang, A.J., Xiong, Y., Casas, S., Yang, B., Ren, M., Urtasun, R.: Towards unsupervised object detection from lidar point clouds. In: CVPR. pp. 9317-9328 (2023)", + "62. Zhao, N., Chua, T.S., Lee, G.H.: Sess: Self-ensembling semi-supervised 3d object detection. In: CVPR. pp. 11079-11087 (2020)", + "63. Zheng, W., Tang, W., Jiang, L., Fu, C.W.: Se-ssd: Self-ensembling single-stage object detector from point cloud. In: CVPR. pp. 14494–14503 (2021)", + "64. Zhou, H., Ge, Z., Liu, S., Mao, W., Li, Z., Yu, H., Sun, J.: Dense teacher: Dense pseudo-labels for semi-supervised object detection. In: ECCV. pp. 35-50 (2022)", + "65. Zhou, Q., Yu, C., Wang, Z., Qian, Q., Li, H.: Instant-teaching: An end-to-end semi-supervised object detection framework. In: CVPR. pp. 4081-4090 (2021)", + "66. Zhu, B., Jiang, Z., Zhou, X., Li, Z., Yu, G.: Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492 (2019)" + ], + "bbox": [ + 215, + 146, + 784, + 563 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "J. Zhan et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/CSOT_ Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection/4b43039d-68da-4e38-a8cf-39348329a3bc_model.json b/2024/CSOT_ Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection/4b43039d-68da-4e38-a8cf-39348329a3bc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8314f093b281555943e51abfb06352bb2629874c --- /dev/null +++ b/2024/CSOT_ Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection/4b43039d-68da-4e38-a8cf-39348329a3bc_model.json @@ -0,0 +1,2282 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.261, + 0.142, + 0.744, + 0.187 + ], + "angle": 0, + "content": "CSOT: Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.267, + 0.212, + 0.737, + 0.243 + ], + "angle": 0, + "content": "Jinglin Zhan\\(^{1}\\), Tiejun Liu\\(^{1}\\), Rengang Li\\(^{1}\\), Zhaoxiang Zhang\\(^{2}\\), and Yuntao Chen\\(^{3\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.394, + 0.255, + 0.558, + 0.269 + ], + "angle": 0, + "content": "1 IEIT Systems" + }, + { + "type": "text", + "bbox": [ + 0.394, + 0.269, + 0.61, + 0.282 + ], + "angle": 0, + "content": "\\(^{2}\\) Institute of Automation, CAS" + }, + { + "type": "text", + "bbox": [ + 0.296, + 0.282, + 0.707, + 0.326 + ], + "angle": 0, + "content": "3 Centre for Artificial Intelligence and Robotics, HKISI, CAS {zhanjinglin, liutj, lirg}@ieisystem.com, zhaoxiang.zhang@ia.ac.cn, chenyuntao08@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.359, + 0.744, + 0.691 + ], + "angle": 0, + "content": "Abstract. Large-scale 3D bounding box annotation is crucial for LiDAR object detection but comes at a high cost. Semi-supervised object detection (SSOD) offers promising solutions to leverage unannotated data, but the predominant pseudo-labeling approach requires careful hyperparameter tuning for training on noisy teacher labels. In this work, we propose a Cross-Scan Object Transfer (CSOT) paradigm for LiDAR SSOD. Central to our approach is HotspotNet, a transformer-based network that predicts possible placement locations and the object-place fitness scores for inserting annotated objects into unlabeled scans in a semantic coherence manner. Based on HotspotNet, CSOT successfully enables object copy-paste in LiDAR SSOD for the first time. To train object detectors on partially annotated scans generated by CSOT, we adopt a spatial-aware classification loss throughout our partial supervision to handle false negative issues caused by treating all unlabeled objects as background. We conduct extensive experiments to verify the efficacy and generality of our method. Compared to other state-of-the-art label-efficient methods used in LiDAR detection, our approach requires the least amount of annotation while achieves the best detector. Using only \\(1\\%\\) of the labeled data on the Waymo dataset, our semi-supervised detector achieves performance on par with the fully supervised baseline. Similarly, on the nuScenes dataset, our semi-supervised CenterPoint reaches \\(99\\%\\) of the fully supervised model's detection performance in terms of NDS score, while using just \\(5\\%\\) of the labeled data. Code is released at https://github.com/JinglinZhan/CSOT" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.704, + 0.741, + 0.733 + ], + "angle": 0, + "content": "Keywords: Semi-Supervised Learning \\(\\cdot\\) LiDAR Object Detection \\(\\cdot\\) Autonomous Driving" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.757, + 0.377, + 0.772 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.787, + 0.787, + 0.817 + ], + "angle": 0, + "content": "Large-scale annotated datasets have catalyzed the success of LiDAR object detection, but at a significant cost in terms of money and time [25,33]. For instance," + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.386, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.127 + ], + "angle": 0, + "content": "J. Zhan et al." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.145, + 0.761, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.372, + 0.788, + 0.43 + ], + "angle": 0, + "content": "Fig. 1: Preliminary experiments of CSOT pipeline for LiDAR SSOD. (a) Compare CSOT implemented based on different cross-scan object transfer methods and display the NDS differences. (b) Visualize prediction differences among various teacher model that used to generate pseudo labels for self-training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.788, + 0.567 + ], + "angle": 0, + "content": "annotating the 12 million 3D boxes in the Waymo Open [38] dataset may have consumed millions of dollars and tens of thousands of human hours. In contrast, autonomous driving fleets capture thousands, and even millions, of hours of unannotated sensor data every day. The huge gap between the speed of manual labeling and that of data accumulation makes labeling a major slowdown. This calls for a more scalable approach to leverage the abundant data collected by autonomous fleets and continually advance object detection capabilities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.793 + ], + "angle": 0, + "content": "Semi-Supervised Learning (SSL) presents a promising way for reducing annotation demands by leveraging both limited labeled data and vast quantities of unlabeled data during training. Various SSL methods have been proposed, including pseudo-labeling (a.k.a. self-training) [23], co-training [2], consistency regularization [20] and generative modeling [19]. Among these SSL approaches, pseudo-labeling has gained significant traction, particularly in the field of semi-supervised object detection (SSOD) [4, 12, 31, 37, 39, 40, 42, 52, 62, 65]. However, generating reliable pseudo-labels without human intervention remains a significant challenge. Noisy pseudo-labels can negatively impact model performance, leading to suboptimal results. To address this, researchers have proposed various techniques, including sophisticated post-processing mechanisms [4, 5, 26, 28, 40, 52, 60], iterative training procedures [39, 45, 49], and noise-tolerant detectors [12, 65] and loss objectives [31, 32, 62, 64]. Nevertheless, these approaches often involve complex pipelines and require careful parameter tuning, making them inefficient and difficult to adapt to new datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In contrast with noisy pseudo-labeling, we leverage a novel semi-supervised learning approach for 3D object detection that transfers annotated objects to generate noise-free training data. Specifically, we transfer point clouds and bound" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "ing boxes of annotated objects to unannotated LiDAR scans to create partially labeled data. Object transfer via copy-paste is a commonly employed data augmentation [6,7,16,47,51] in LiDAR object detection. But naively applying object copy-paste [11, 15, 47] to semi-supervised LiDAR object detection yields suboptimal results. As shown in Fig. 1, SSOD built upon vanilla object transfer does improve performance over the baseline, but combining it with self-training results in a negative net effect on detection performance compared to using self-training alone. The key insight is that object copy-paste without careful contextual consistency maintenance improves detector performance at the cost of generating a large number of false positives. Although the commonly used object detection metric mAP is not sensitive to false positives if the recall improves, the self-training pipeline accumulates these false positives over iterations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.675 + ], + "angle": 0, + "content": "In order to achieve coherent object copy-paste, we propose Hotspot Network, a learning-based approach to insert objects with strong contextual consistency, enabling object-transfer-based semi-supervised LiDAR object detection to achieve its full potential. HotspotNet takes unannotated LiDAR scans as inputs, along with point clouds and bounding boxes of annotated objects. It then predicts possible placement locations for these objects and assigns scores to each location, indicating the likelihood of the object fitting well within the context of the scanned environment. This allows us introducing noise-free supervision signal into unlabeled point clouds and use these newly created partially-labeled data to optimize detector performance. Based on HotspotNet, we design a novel Cross-Scan Object Transfer(CSOT) pipeline for LiDAR SSOD. CSOT enjoys precise partial annotations that maintain great contextual coherence with unannotated scans rather than noisy pseudo labels for training detectors on unlabeled data. It's important to note that training detectors on partially annotated point clouds, as opposed to fully labeled ones, can lead to a significant issue: unlabeled objects being miscategorized as background, resulting in a high number of false negatives. To mitigate this problem, we propose a spatial-aware classification loss and employ it throughout our partial supervision. Our CSOT paradigm features a streamlined training pipeline and requires minimal hyper-parameter tuning, enabling it to quickly generalize across various datasets and detectors. Moreover, this approach is orthogonal and complementary to commonly used self-training techniques. We conducted extensive experiments on the Waymo [38] and nuScenes [3] datasets to confirm the effectiveness of our methods." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.677, + 0.555, + 0.691 + ], + "angle": 0, + "content": "Our contributions are summarized below:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.704, + 0.785, + 0.747 + ], + "angle": 0, + "content": "- We propose a Hotspot Network for instructing object placement in a semantic coherence manner, and employ an asymmetric supervision for training HotspotNet on both annotated and unannotated data." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.749, + 0.788, + 0.794 + ], + "angle": 0, + "content": "- Based on HotspotNet, we design a novel Cross-Scan Object Transfer paradigm for LiDAR semi-supervised object detection. To the best of our knowledge, this is the first successful attempt of object copy-paste in LiDAR SSOD." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.796, + 0.785, + 0.84 + ], + "angle": 0, + "content": "- We conduct extensive experiments on nuScenes and Waymo datasets to verify the efficacy and generality of our method. Our approach achieves competitive performance compared to fully-supervised baselines while using 20x" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.704, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "J. Zhan et al." + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "to 100x fewer annotations, significantly outperforming other state-of-the-art label-efficient approaches in LiDAR object detection." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.196, + 0.389, + 0.213 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.223, + 0.547, + 0.239 + ], + "angle": 0, + "content": "2.1 Semi-Supervised Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.243, + 0.789, + 0.546 + ], + "angle": 0, + "content": "Unsupervised learning eliminates data annotation but significantly degrades model performance [48, 55, 61]. Semi-supervised methods maintain low annotation costs meanwhile greatly bridges the gap to fully-supervised approaches [43]. As a prior work, STAC [37] proposes a pseudo labeling SSOD paradigm and catalyzes extensive follow-on researches. It trains a teacher with all annotated data and generates pseudo labels for supervising the student on unlabeled data. Producing reliable pseudo labels is critical in this paradigm and numerous methods have been explored. 3D IoUMatch [40] utilizes the estimated 3D IoU to filter poorly localized predictions. LabelMatch [4] leverages adaptive label-distribution-aware confidence thresholds for generating unbiased pseudo labels. Proficient Teacher [52] employs clustering-based box voting for improving the precision of final predictions. HSSDA [29] designs a dynamic dual-threshold strategy to generate more reasonable supervision for training the student model. Besides various filter mechanism, many other approaches, such as iterative training procedures [39, 62], noise-tolerant detectors [12, 65] and a variety of loss objectives with consistency or reweighting terms [31, 42, 62], are also proposed for alleviating negative effects caused by noisy pseudo labels. Despite some progress has been made, unreliable pseudo-labels still remain an open challenge. The sophisticated pipelines with extensive manual threshold tuning also impede these approaches' quick adaptation to new datasets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.563, + 0.604, + 0.579 + ], + "angle": 0, + "content": "2.2 Copy-Paste For LiDAR Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Copy-paste method is developed by Jaderberg et.al. for text recognition [18], then extended by Dwibedi et.al. [9] and Yun et.al. [57] for 2D object detection. Yan et.al. [47] introduces copy-paste into LiDAR object detection and develops the well known ground-truth augmentation approach (so called GT-Aug), which is extensively adopted in many later studies [1,6,16,17,21,51,54,66]. Since then, numerous approaches have been proposed to optimize GT-Aug. Part-Aware [8] and Shape-Aware [63] GT-Aug divide objects into partitions and apply different augmentation methods to each partition. Pattern-Aware [14] GT-Aug downsamples the points of objects to create a new one with farther distance. PointCut-Mix [22] replaces part of the sample with shape-preserved subsets from another one. Some researchers notice placement issues in GT-Aug and develop hand-craft valid maps to restrict the insertion of objects [11,15]. Up to now, researches on copy-paste is primarily focused on exploring and developing its capabilities for augmenting data to support fully-supervised learning tasks. Ghiasi et.al. introduce copy-paste into SSL [13] for augmenting both pseudo labeled and supervised labeled images. The above trial still limited in the field of data augmentation and the copy-paste technique remains largely untapped." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.145, + 0.764, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.354, + 0.788, + 0.412 + ], + "angle": 0, + "content": "Fig. 2: The overall architecture of HotspotNet. The Object Encoders extract features from object point cloud \\( X_{obj,pc} \\) and bounding box meta data \\( X_{obj,meta} \\). The scene encoder featurize the scene point cloud and cross-attends to object features \\( \\mathcal{F}_{\\mathrm{obj,pc}} \\) and \\( \\mathcal{F}_{\\mathrm{obj,meta}} \\). The location decoder predicts the placeable locations and their scores." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.436, + 0.331, + 0.451 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.788, + 0.527 + ], + "angle": 0, + "content": "We provide details of Cross-Scan Object Transfer SSOD paradigm in this section. We present the overall architecture of HotspotNet in Sec. 3.1, and describe the Asymmetric Supervision for training HotspotNet in Sec. 3.2. The pipeline of our newly proposed SSOD paradigm, CSOT, is elaborated in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.546, + 0.406, + 0.561 + ], + "angle": 0, + "content": "3.1 Hotspot Network" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.787, + 0.629 + ], + "angle": 0, + "content": "The key to LiDAR copy-paste is identifying semantically coherent locations within background scans to paste objects. Placing objects at unreasonable locations, such as inside walls, floating in sky or with weird pose, will cause a large amount of false positives and thus deteriorate the detector performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.787, + 0.69 + ], + "angle": 0, + "content": "We propose a learning-based approach for object placement that provides sufficient supervision signal and maintain a high signal-to-noise ratio. Specifically, we employ a Hotspot Network to predict the placeable location \\(\\mathbf{b}\\) and the object-place fitness scores \\(o\\) for inserting object in a semantic coherence manner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.787, + 0.779 + ], + "angle": 0, + "content": "We abbreviate Hotspot Network as HotspotNet and visualize the overall architecture in Fig. 2. The HotspotNet comprises object encoders for object point cloud \\( X_{\\mathrm{obj,pc}} \\) and meta data \\( X_{\\mathrm{obj,meta}} \\), respectively. The scene encoder featurizes the scene point cloud and cross-attends to object features \\( \\mathcal{F}_{\\mathrm{obj,pc}} \\) and \\( \\mathcal{F}_{\\mathrm{obj,meta}} \\). A location decoder predicts the placeable location \\( \\mathbf{b} = \\{x,y,z,w,l,h,\\theta ,c\\} \\) and the location fitness score \\( o \\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Object Encoder We process the object point cloud \\( X_{\\mathrm{obj,pc}} \\) with object point cloud encoders like PointNet [34] or PointNeXt [35] and obtain object features \\( \\mathcal{F}_{\\mathrm{obj,pc}} \\). We encode the bounding box meta \\( X_{\\mathrm{obj,meta}} \\) with appropriate meta feature encoders like Fourier embedding for boxes, and CLIP or one-hot embedding" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "J. Zhan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.179 + ], + "angle": 0, + "content": "for object classes. We then transform the initial box embedding with a MLP to obtain bounding box meta feature \\(\\mathcal{F}_{\\mathrm{obj,meta}}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.279, + 0.192, + 0.785, + 0.208 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\mathrm {o b j}, \\mathrm {p c}} = 0 \\text {b j e c t P o i n t C l o u d E n c o d e r} \\left(X _ {\\mathrm {o b j}, \\mathrm {p c}}\\right) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.263, + 0.211, + 0.785, + 0.227 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\text {o b j , m e t a}} = \\operatorname {M L P} \\left(\\text {F o u r i e r E m b e d} \\left(X _ {\\text {o b j , b o x}}\\right) \\mid \\mid \\text {C l a s s E m b e d} \\left(X _ {\\text {o b j , c l a s s}}\\right)\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.784, + 0.329 + ], + "angle": 0, + "content": "Scene Encoder Given a scene LiDAR scan \\( X_{\\mathrm{scene}} \\), we leverage off-the-shelf point cloud encoders like CenterPoint [54] or SST [10] for extracting features \\( \\mathcal{F}_{\\mathrm{scene}} \\). We use spatial cross-attention in the first convolution block of each stage of the scene point cloud encoder. This allows the scene encoder aggregating features from both the object point clouds and the box metadata for assessing the object fitness of each spatial location." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.331, + 0.784, + 0.391 + ], + "angle": 0, + "content": "Location Decoder We randomly initialize a set of learnable queries \\(\\mathbf{q}_0\\). These queries are combine with position embeddings to obtain location queries \\(\\mathbf{q}\\). Our location decoder leverages both scene features and object features for predicting possible inserting locations and scores." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.417, + 0.686, + 0.433 + ], + "angle": 0, + "content": "3.2 Training HotspotNet with Asymmetric Supervision" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.784, + 0.551 + ], + "angle": 0, + "content": "The HotspotNet consists of a location predictor and a location scorer that work jointly to suggest object placements. The location predictor outputs candidate object positions, while the scorer assigns fitness scores to evaluate how semantically consistent each candidate is. We employ a multi-task learning approach that features Asymmetric Supervision for optimizing different predictions. The proposed losses supervise the location predictor and scorer to accurately position objects within a new environment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.785, + 0.704 + ], + "angle": 0, + "content": "To train the location predictor, we utilize the annotated LiDAR scans. For each annotated scan, we randomly extract an object point cloud from the original point clouds along with its corresponding bounding box. The regressor is then supervised with an \\(\\ell_1\\) loss between the predicted location \\(\\mathbf{b}\\) and the extracted bounding box \\(\\mathbf{b}_{\\mathrm{cut - out}}\\). We constrain object distance and perspective from the ego car in Eq. 3 to preserve perspective and point density. We employ dropout regularization on the bounding box meta data to prevent the location predictor from directly copying the meta data to the prediction output. It encourages the predictor to learn meaningful features from the point clouds itself, rather than relying solely on the provided meta data." + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.719, + 0.785, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {l o c}} (\\mathbf {b}) = \\left\\| \\mathbf {b} - \\mathbf {b} _ {\\text {c u t - o u t}} \\right\\| _ {1} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.748, + 0.784, + 0.822 + ], + "angle": 0, + "content": "To train the location scorer, we utilize an asymmetric supervision approach for positive and negative samples. Positive samples are obtained by matching location queries with the extracted object point clouds. These positive samples encourage the scorer to assign high scores to queries that accurately position objects within the scan." + }, + { + "type": "equation", + "bbox": [ + 0.437, + 0.826, + 0.785, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s c o r e}} ^ {p} (o) = - \\log o \\tag {4}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "To avoid treating all queries that do not match the extracted objects as negative samples, a careful selection process is employed. Instead of using all non-matching queries, negative samples are chosen based on queries that have low fitness scores, denoted by \\( s \\). This approach ensures that the scorer learns to differentiate between suitable and unsuitable object placements, focusing on the most informative negative samples." + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.25, + 0.786, + 0.268 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s c o r e}} ^ {n} (o) = - \\left(1 - \\mathbb {1} \\left[ s _ {p} \\cdot s _ {o} \\cdot s _ {c} > \\tau \\right]\\right) \\log (1 - o) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.278, + 0.789, + 0.519 + ], + "angle": 0, + "content": "The fitness scores \\( s_p \\), \\( s_o \\), and \\( s_c \\) are used to measure the accuracy and coherence of an object's position, orientation, and category within the scene. The position score \\( s_p = \\exp(-(\\mathbf{b}_{xyz} \\cdot \\mathbf{n}_g + d_g) / \\sigma_p^2)^1 \\) penalizes the predicted location \\( \\mathbf{b}_{xyz} \\) deviating from the local ground plane specified by \\( (\\mathbf{n}_g, d_g) \\). Local ground planes here could be obtained via non-learning ground detection methods like Patch-Works++ [24]. Similarly, the orientation score \\( s_o = \\cos((\\mathbf{b}_\\theta - \\hat{\\mathbf{b}}_\\theta) / 2) \\) encourages consistency between the heading of the new object location and the predominant heading \\( \\hat{\\mathbf{b}}_\\theta \\) of objects in the same category within the scene. Finally, the category compatible score \\( s_c = \\mathbb{1}[\\mathbf{b}_c \\in \\mathcal{C}] \\) prevents pasting object of category that is not in the current scene (e.g. bicycle on highway). Notably, these scores can be computed not only between the extracted object and its original scan but also between the extracted object and other unannotated LiDAR scans, in which \\( \\hat{\\mathbf{b}}_\\theta \\) can be simplified as ego orientation and \\( s_c \\) can be ignored directly. This adaptability opens up the possibility of learning object placement scoring from a more extensive dataset through unsupervised learning, which assigns positive and negative samples solely based on the fitness scores." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.788, + 0.566 + ], + "angle": 0, + "content": "Combining these asymmetric losses and carefully selecting positive and negative samples, our HotspotNet learns to accurately position objects within a new scan while maintaining semantic coherence." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.588, + 0.691, + 0.604 + ], + "angle": 0, + "content": "3.3 Cross-Scan Object Transfer for LiDAR-based SSOD" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.705 + ], + "angle": 0, + "content": "Under the guidance of HotspotNet (described in Sec. 3.1), we inject precise supervision signals into unannotated point clouds and train detectors on these partially-labeled data. Fig. 3 illustrates the differences between our Cross-Scan Object Transfer SSOD paradigm and the commonly used self-training technique. Our method bypasses the iterative self-training procedures, cumbersome pseudolabel filtering mechanisms, and complex loss objective designs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.726, + 0.788, + 0.819 + ], + "angle": 0, + "content": "Partially Supervised Object Detection Our approach for introducing supervision into unlabeled LiDAR scans is straightforward. We denote all objects in labeled point clouds as \\(\\{\\mathbf{O}^L\\}\\), and all unlabeled scans as \\(\\{\\mathbf{S}^U\\}\\). For an unlabeled LiDAR scan \\(\\mathbf{S}^U\\), we randomly sample a subset of \\(\\{\\mathbf{O}^L\\}\\) and follow the instruction of HotspotNet to inject \\(\\{\\mathbf{O}^L\\}\\) into \\(\\mathbf{S}^U\\) in a semantically coherent manner. Since inserted objects come with ground-truth classes, poses and dimensions," + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.603, + 0.841 + ], + "angle": 0, + "content": "\\(^{1}\\) \\(\\mathbf{b}_{\\square}\\) denotes the \\(\\square\\) component for location prediction \\(\\mathbf{b}\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.229, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "J. Zhan et al." + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.145, + 0.772, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.408, + 0.788, + 0.451 + ], + "angle": 0, + "content": "Fig. 3: (a) Common-used pseudo-labeling SSOD paradigm. (b) Cross-Scan Object Transfer SSOD paradigm, which bypasses the iterative self-training procedures, cumbersome pseudo-labels filter mechanisms and complex loss objectives designs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.788, + 0.573 + ], + "angle": 0, + "content": "these newly created scans allow for supervision. However, it is important to note that the above synthesized point clouds are only partially annotated, as we lack information about the original unlabeled objects in these unannotated scans. Learning a detector from partially labeled data is known as Sparsely Annotated Object Detection (SAOD) [30, 41]. The differences of training a detector from fully and partially annotated data will be discussed in the next part." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Spatial-Aware Classification Loss An object detector generally includes a classifier to differentiate foreground objects from background clutter, and a localizer to predict the accurate location, dimension and orientation of an object. The localizer is typically supervised by regressing the offset from current voxel location to actual bounding boxes of foreground objects, so it can be naturally supervised by new point clouds generated via cross-scan object transfer. However, the classifier requires both positive and negative samples for learning a decision boundary. Once again, the inserted objects could provide positive samples for the classifier but we can not treat all other voxels as background in SAOD just like vanilla object detection. This is due to the existence of the so-called hard negatives as reported in other SAOD works [41]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In our approach, hard negatives refer to the original unlabeled objects intrinsically present in the unannotated LiDAR scans. Training the classifier in our semi-supervised detector with these hard negatives will result in an unacceptable number of false negatives. Instead of resorting to sophisticated re-weighting/ recalibration strategies [44,59] or teacher-student siamese networks [41], we simply" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "choose negative samples in a spatial-aware manner to avoid sampling unlabeled objects as negatives. We propose a spatial-aware classification loss for our partial supervision as Eq.6:" + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.2, + 0.786, + 0.22 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {c l s}} ^ {\\mathrm {n}} (\\mathbf {x}) = \\mathbb {1} \\left[ \\left(\\min _ {i} \\| \\mathbf {x} - \\mathbf {c} _ {i} \\| _ {2}\\right) \\leq \\sigma \\right] \\cdot L _ {\\mathrm {c l s}} ^ {\\mathrm {n}} (\\mathbf {x}) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.226, + 0.788, + 0.378 + ], + "angle": 0, + "content": "where \\( L_{\\mathrm{cls}}^{\\mathrm{n}}(\\mathbf{x}) \\) denotes focal loss for negative samples at the heatmap location \\( \\mathbf{x} \\). The indicator function \\( \\mathbb{1}(\\cdot) \\) equals 1 when the \\( \\ell 2 \\) distance between a location \\( \\mathbf{x} \\) and the center location \\( \\mathbf{c} \\) of any foreground object is less than a threshold \\( \\sigma \\), and 0 otherwise. We set \\( \\sigma \\) as twice of the Gaussian radius used for positive samples assignment in CenterPoint [54] for unlabeled scans, and infinity for fully annotated data. An immediate drawback of our spatial-aware classification loss is that there are too few negative samples for the classifier. To alleviate this issue, we randomly sample \\( 1\\% \\) samples in the heatmap predictions as negative samples, and set the corresponding value of \\( \\mathbb{1}(\\cdot) \\) as 1 to make the classification boundary between positive and negative samples more clear." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.784, + 0.469 + ], + "angle": 0, + "content": "Synergy with Self-Training Techniques A deeper analysis of the proposed CSOT SSOD paradigm reveals that the primary mechanism driving improved detection performance is the extrapolation of labeled object detection to novel contextual environments. This stands in contrast to self-training methods, which learn from pseudo-labeled objects in unannotated scans." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.47, + 0.788, + 0.608 + ], + "angle": 0, + "content": "As a result, the proposed CSOT pipeline for semi-supervised LiDAR object detection is orthogonal to, yet complementary with, existing self-training techniques [37], allowing for additional performance gains. Deliberately designing a pseudo-labeling SSOD framework is beyond the scope of this paper. To verify the complementary effect of our approach when combined with self-training techniques, we implement a basic teacher-student paradigm. We employ a single universal confidence score threshold \\(\\tau_{\\mathrm{pseudo}}\\), which is invariant across categories, to filter low-confidence predictions. We then utilize a fully-supervised loss, as in CenterPoint [54], for optimizing the student detector." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.625, + 0.367, + 0.642 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.651, + 0.784, + 0.698 + ], + "angle": 0, + "content": "The proposed methods are verified in this section. We show implementation details in Sec. 4.1 and provide all experimental results in Sec. 4.2 and Sec. 4.3. More ablations and discussions are shown in our supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.714, + 0.428, + 0.73 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Datasets We validate our method on two widely used autonomous driving datasets: nuScenes [3] and Waymo [38]. The nuScenes dataset has 700 training, 150 validation, and 150 test sequences. Official nuScenes detection metrics are NuScenes Detection Score (NDS) and mean Average Precision (mAP). The Waymo Open Dataset consists of 798 training and 202 validation sequences. Official Waymo detection metrics are mean Average Precision (mAP) and mean Average Precision weighted by Heading (mAPH)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "J. Zhan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.33 + ], + "angle": 0, + "content": "Training Details We use one-stage CenterPoint [54] as our baseline detector to validate the effectiveness of our methods. Voxel sizes are set as \\([0.075\\mathrm{m}, 0.075\\mathrm{m}, 0.2\\mathrm{m}]\\) for nuScenes and \\([0.1\\mathrm{m}, 0.1\\mathrm{m}, 0.15\\mathrm{m}]\\) for Waymo. The training epochs for one cycle of learning procedure, including fully-supervised learning, partially-supervised learning and self-training, are set as 20 for nuScenes and 7 for Waymo. We follow many previous work [6,51,54,66] to adopt Adam optimizer with one-cycle learning rate policy. Labeled data are uniformly sampled from the entire dataset. Then objects and their annotations are extracted to build the ground-truth database for injecting precise supervision signals into unlabeled scans. We finetune partially-supervised detectors on labeled scans with several epochs. We set a score threshold of 0.3 for nuScenes and 0.6 for Waymo to filter low-quality pseudo labels, and verify the orthogonality of our methods with self-training." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.352, + 0.365, + 0.366 + ], + "angle": 0, + "content": "4.2 Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.789, + 0.606 + ], + "angle": 0, + "content": "Benchmarks on nuScenes dataset We compare our methods against other label-efficient LiDAR detection approaches on the nuScenes dataset. As shown in Tab. 1, our methods obtain the significant performance improvements over previous works. For the sake of fairness, we choose the fully supervised detector trained by Yin et.al. [54] (the designer of CenterPoint backbone) as the baseline detector, and list the results achieved from other label-efficient LiDAR object detection approaches for comparison [27, 43, 56, 58]. Without leveraging models pre-trained on external datasets and using only \\(5\\%\\) of the annotations, we achieve competitive detector which reaches \\(97.2\\%\\) of the fully-supervised performance on mAP and \\(98.9\\%\\) on NDS. We significantly shrink the mAP and NDS gaps between the baseline detector and the one trained with \\(20\\mathrm{x}\\) less annotations from around \\(10\\sim 20\\) to merely 1.6 and 0.7 respectively. Our semi-supervised detector, trained with only \\(5\\%\\) manual annotations, obtains over \\(90\\%\\) detection performance on 8 of the 10 categories defined in nuScenes dataset, and over \\(95\\%\\) detection performance on 6 categories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Benchmarks on Waymo dataset We extend our methods on Waymo dataset and achieve similar outstanding results. we choose the fully supervised detector trained by Shi et.al. [36], Yan et.al. [47], Fan et.al. [10] and Yin et.al. [54] as the baseline detector, and list the results achieved from other label-efficient LiDAR object detection approaches for comparison [29, 46, 50, 52, 53, 56]. As shown in Tab. 2, our methods are valid on both anchor-based (SECOND) and anchor-free (CenterPoint) detectors. Taking CenterPoint as an example, with merely \\(1\\%\\) annotations, we achieve a competitive detector which matches fully supervised performance on vehicle and pedestrian, and exceeds it on cyclist. The \\(100\\mathrm{x}\\) reduction in annotations leads to only a minor 1.8 APH decrease for L2 Vehicle, a negligible 0.4 APH drop for L2 Pedestrian, and boosts the APH of L2 Cyclist by 4.1. The abnormal enhanced detection performance on Cyclist should be attributed to our learning-based approach for cross-scan object transfer. Instructed by the proposed HotspotNet, we paste objects in a semantic coherent" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Compare with other labeled efficient LiDAR object detection approaches on nuScenes Dataset validation split. Performances of the baseline detector reported by its original designer [54] are marked with gray backgrounds." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.199, + 0.785, + 0.276 + ], + "angle": 0, + "content": "
MethodRefAnnosExternalmAPNDSCarTruckC.V.BusTrailerBarrierMot.Byc.Ped.T.C.
BaselineCVPR 2021 [54]100%58.065.584.651.017.560.253.270.953.728.783.476.7
GCC3DICCV 2021 [27]5%41.146.8——————————
SceneFlowECCV 2022 [58]5%36.048.3——————————
SSDA3DAAAI 2023 [43]5%✓——76.2—————————
AD-PTNeurIPS 2023 [56]5%✓45.053.078.943.811.155.221.255.139.017.872.355.4
Ours—5%56.464.881.753.516.558.632.866.060.047.882.065.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.288, + 0.788, + 0.331 + ], + "angle": 0, + "content": "Table 2: Compare with other labeled efficient LiDAR object detection approaches on Waymo Dataset validation split. Performances of the baseline detector reported by its original designers [10,36,47,54] are mark with gray backgrounds." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.342, + 0.784, + 0.508 + ], + "angle": 0, + "content": "
DetectorMethodRefAnnoxExtVehiclePedestrianCyclistAll L2 AP
L1L2L1L2L1L2
APAPHAPAPHAPAPHAPAPHAPAPHAPAPH
PV-RCNNBaselineCVPR 2020 [36]100%77.576.969.068.475.065.766.057.667.866.465.464.063.3
PV-RCNNHSSDACVPR 2023 [29]1%56.453.849.747.340.120.933.517.529.120.927.920.028.3
PV-RCNNMIXSUPICLR 2024 [50]1%———55.5———52.0———62.356.6
SECONDBaselineSensors 2018 [47]100%72.371.763.963.368.758.260.751.360.659.358.357.157.2
SECONDMIXSUPICLR 2024 [50]10%———55.0———49.6———58.154.2
SECONDProficientTeacherECCV 2022 [52]5%——53.052.5——50.338.7——49.946.045.7
SECONDAD-PTNeurIPS 2023 [56]3%✓——60.559.9——54.945.8——50.849.751.8
SECONDProposalContrastECCV 2022 [53]1%——37.636.9——39.731.7——37.735.734.8
SECONDOurs—1%69.468.961.260.864.060.755.152.361.160.458.958.257.1
SSTBaselineCVPR 2022 [10]100%74.273.865.565.178.769.670.061.770.769.668.066.964.6
SSTMIXSUPICLR 2024 [50]10%———59.1———60.0———63.160.7
SSTMV-JARCVPR 2023 [46]5%——56.556.0——57.747.7——37.436.346.7
CenterPointBaselineCVPR 2021 [54]100%76.676.068.968.479.073.471.065.872.171.069.568.567.6
CenterPointMIXSUPICLR 2024 [50]10%———61.8———57.7———67.562.3
CenterPointAD-PTNeurIPS 2023 [56]3%✓——60.459.8——60.654.0——62.761.658.5
CenterPointOurs—1%75.174.767.066.675.972.968.265.475.975.173.472.668.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.644 + ], + "angle": 0, + "content": "manner, which enables our CSOT SSOD to reach its full potential and meanwhile can be used as a powerful data augmentation for fully-supervised detector. Related ablation results are discussed in Sec. 4.3. Notably, under the most stringent constraints with minimal labeled point clouds and without leveraging any external pre-training, our approach stands alone as the only label-efficient technique which is able to achieve equivalence detection performance to the baseline detector on the average L2 APH of all categories defined in Waymo dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.665, + 0.344, + 0.679 + ], + "angle": 0, + "content": "4.3 Ablations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Ablate training stage on nuScenes dataset. Our semi-supervised detectors trained with 20x fewer nuScenes annotations are compared to both fully supervised and pseudo-labeling semi-supervised models. Results in Tab. 3 indicate that our newly proposed CSOT is a valid label-efficient approach, and it almost completely maintains the performance gains of self-training when used together. The fully supervised detector trained with \\(5\\%\\) labeled scans shows substantial performance degradation compared to that trained on the entire labeled dataset, with a mAP decline from 59.5 to 43.8 and a NDS decrease from 66.0 to 56.5. As shown in Lines 3-5 of Tab. 3, self-training technique is able to boost the mAP from 43.8 to 51.0 and NDS from 56.5 to 61.0. Our CSOT, which allows" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "J. Zhan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 3: Compare with fully supervised and pseudo-labeling semi-supervised detectors. All results are achieved based on our implementation code of CenterPoint and evaluated on nuScenes dataset validation split." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.2, + 0.785, + 0.267 + ], + "angle": 0, + "content": "
AnnosScansCSOTSelf-TrainingmAPNDSCarTruckC.V.BusTrailerBarrierMot.Byc.Ped.T.C.
1100%100%59.566.083.756.618.766.934.867.163.750.484.568.7
25%5%43.856.575.839.712.344.120.556.936.318.476.756.9
35%100%✓51.061.080.149.114.655.626.762.345.430.881.963.1
45%100%✓50.461.079.646.411.750.125.659.152.137.280.662.0
55%100%✓✓56.464.881.753.516.558.632.866.060.047.882.065.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.282, + 0.788, + 0.338 + ], + "angle": 0, + "content": "Table 4: Compare with fully supervised and pseudo-labeling semi-supervised detectors. All results are achieved based on our implementation code of CenterPoint and evaluated on Waymo dataset validation split (mean L1/L2 AP/APH over all categories defined in Waymo dataset)." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.35, + 0.784, + 0.435 + ], + "angle": 0, + "content": "
AnnosScansCSOTSelf-TrainingObject TransferL1 AP/APHL2 AP/APHmAP/mAPH
1100%100%—74.4 / 72.768.6 / 67.071.5 / 69.9
1a100%100%GT-Aug74.7 / 73.168.8 / 67.371.8 / 70.2
1b100%100%Ours77.5 / 76.071.6 / 70.274.6 / 73.1
21%1%—68.3 / 66.662.3 / 60.865.3 / 63.7
31%100%✓—71.9 / 70.465.8 / 64.568.9 / 67.5
41%100%✓Ours71.4 / 69.865.4 / 63.968.4 / 66.9
51%100%✓✓Ours75.7 / 74.269.5 / 68.272.6 / 71.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.788, + 0.573 + ], + "angle": 0, + "content": "supervision on unannotated point clouds by converting unlabeled scans into partially labeled scans via learning-based copy-paste, achieves a similar detection performance (with a mAP of 50.4 and a NDS of 61.0). It is noteworthy that iterative employment of self-training techniques does not lead to sustained gains in detection performance (see supplementary material), while incorporating just one cycle of self-training on our partially supervised detector is able to further optimize the mAP from 50.4 to 56.4 and NDS from 61.0 to 64.8." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Ablate training stage on Waymo dataset We extend the above ablations on Waymo dataset and summarize all results in Tab. 4, which further validate the effectiveness of CSOT, as well as its orthogonality with the common used pseudo-labeling techniques. \\(100\\mathrm{x}\\) fewer labeled point clouds lead to a degradation of detection performance, reducing the mean AP and APH from 71.5, 69.9 to 65.3, 63.7. Our Cross-Scan Object Transfer paradigm brings a boost of around 3 APH for both LEVEL_1 and LEVEL_2 objects. CSOT is proved to be complimentary to self-training, and the performance gains can be completely maintained. As shown in the Line 5 of Tab. 4, the performance of detector trained with \\(1\\%\\) annotations can be finally optimized to match even surpass the fully-supervised baseline model. The abnormal enhanced detection performance with \\(100\\mathrm{x}\\) reduced labeled data should be attributed to the learning-based copy-paste operation employed in this work. As shown in Lines 1a-1b of Tab. 4, replacing the commonly used copy-paste augmentation (GT-Aug) with our learning based approaches boosts the fully-supervised detector with an increase of \\(2.8~\\mathrm{mAP}\\). Notably, even regarding the fully-supervised detector augmented via our object" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 5: Ablate our learning-based cross-scan object transfer which is instructed by HotspotNet. All results are evaluated on nuScenes dataset validation split." + }, + { + "type": "table", + "bbox": [ + 0.219, + 0.186, + 0.784, + 0.443 + ], + "angle": 0, + "content": "
(a) Compare our learning-based cross-scan object transfer (instructed by HotspotNet) with other similar methods.
AnnosCSOTObj TransferSelfObj Appearance rαLocal Context (x,y,z)Global Context rsclsrule /learnedmAP NDS
15%——————43.8 56.5
25%✓——————51.0 61.0
35%✓GT-Aug [47]✓✓46.0 57.7
45%✓CA-Aug [15]✓✓✓rule46.0 58.0
55%✓LiDAR-Aug [11]✓✓✓rule46.2 57.9
65%✓Ours✓✓✓✓✓learned50.4 61.0
6a5%✓Ours✓✓✓✓✓rule49.7 60.6
75%✓GT-Aug [47]✓✓49.7 60.5
85%✓CA-Aug [15]✓✓✓rule49.7 60.7
95%✓LiDAR-Aug [11]✓✓✓rule49.7 60.2
105%✓Ours✓✓✓✓✓learned56.4 64.8
10a5%✓Ours✓✓✓✓✓rule53.6 63.3
(b) Ablate contribution of each fitness score defined in Ln-score(o) (Eq. 5).
AnnosCSOTSelfposition sporientation socategory scmAP NDS
115%✓✓49.7 60.5
125%✓✓✓52.1 62.6
135%✓✓✓49.4 60.8
13a5%✓✓✓52.2 62.3
13b5%✓✓✓✓55.1 64.3
145%✓✓✓51.6 62.0
155%✓✓✓✓✓56.4 64.8
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.471, + 0.785, + 0.517 + ], + "angle": 0, + "content": "transfer implementation as the baseline, reducing annotations by \\(100\\mathrm{x}\\) causes only a minimal degradation in performance (remains \\(97.3\\% / 97.4\\%\\) mAP/mAPH of our fully-supervised detector)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Ablate our learning-based object transfer. We compare our learning-based object transfer which instructed by the proposed HotspotNet with other similar techniques in Tab. 5(a), and ablate negative sample strategy employed for calculating \\(\\mathcal{L}_{\\mathrm{score}}^n (o)\\) (Eq. 5) in Tab. 5(b). As shown in Lines 3-10 of Tab. 5, the proposed learning-based object transfer is able to maximize the benefits of CSOT. It also stands alone in its ability to seamlessly integrate with self-training technique without compromising their performance gains. CSOT built upon our learning-based object transfer enhances the detection performance with a considerable \\(6.6\\mathrm{mAP}\\) and 4.5 NDS. The improvement surpasses the one built upon previous similar approaches (with a margin of about \\(4\\mathrm{mAP}\\) and 3 NDS). The poor partially-supervised detector achieved from other object transfer approaches produce enormous false positive predictions (visualized in Fig. 1) and greatly influence the downstream self-training. Replacing our learning-based object transfer with previous methods reduces the mAP gains of self-training from 6.0 to around 3.5, and degrades the final mAP from 56.4 to 49.7. The necessity of learning-based object transfer are verified in Tab. 5, Lines 6-6a,10-10a. Learn-based approach enables better generalization and can help reduce noise, which introduces an increased final mAP of 2.8. We further ablate negative sample strategy employed for calculating \\(\\mathcal{L}_{\\mathrm{score}}^n (o)\\) and list results in Lines 11-15 of Tab. 5. Pocation and category provide the most important contextual informa" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "J. Zhan et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 6: Ablate robustness of our methods across various network architectures. All results are evaluated on nuScenes dataset validation split." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.186, + 0.784, + 0.357 + ], + "angle": 0, + "content": "
AnnosDetectorEncoderAnchor-baseAnchor-freeVoxel SizeCSOTSelfmAPNDS
1100%CenterPointVFE✓(0.075,0.075,0.2)59.566.0
25%CenterPointVFE✓(0.075,0.075,0.2)43.856.5
35%CenterPointVFE✓(0.075,0.075,0.2)✓51.061.0
45%CenterPointVFE✓(0.075,0.075,0.2)✓50.461.0
55%CenterPointVFE✓(0.075,0.075,0.2)✓✓56.464.8
6100%CenterPointPFE✓(0.1,0.1,8.0)49.260.1
75%CenterPointPFE✓(0.1,0.1,8.0)30.248.3
85%CenterPointPFE✓(0.1,0.1,8.0)✓37.853.1
95%CenterPointPFE✓(0.1,0.1,8.0)✓40.054.9
105%CenterPointPFE✓(0.1,0.1,8.0)✓✓45.057.9
11100%SECONDVFE✓(0.1,0.1,0.2)55.464.7
125%SECONDVFE✓(0.1,0.1,0.2)38.453.2
135%SECONDVFE✓(0.1,0.1,0.2)✓45.258.1
145%SECONDVFE✓(0.1,0.1,0.2)✓45.459.1
155%SECONDVFE✓(0.1,0.1,0.2)✓✓48.661.3
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.385, + 0.785, + 0.446 + ], + "angle": 0, + "content": "tion, orientation provides additional cues to narrow down the possible identity. Specifically, combining orientation constraint with category consistency improve mAP by 0.6 (Tab. 5, Lines 13a and 14). Removing orientation constraint from all reduces mAP by 1.3 (Tab. 5, Lines 13b and 15)." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.468, + 0.786, + 0.664 + ], + "angle": 0, + "content": "Extend our method to different detectors. We confirm the generality of our methods on various detectors, which utilizes diverse point cloud feature encoder or target assigner strategy. All results are listed in Tab. 6. For an anchor-free detector CenterPoint with a pillar-based feature encoder, the proposed CSOT bring a boost of \\(9.8\\mathrm{mAP}\\) and 6.6 NDS, which outperforms the self-training technique by \\(2.2\\mathrm{mAP}\\) and 1.8 NDS. Using our partially-supervised detector as the teacher model to generate pseudo-labels for unlabeled point clouds, and then conduct one-cycle of self-training. The performance of semi-supervised detector is ultimately reaches \\(45.0\\mathrm{mAP}\\) and 57.9 NDS. The mAP gap introduced by \\(20x\\) fewer annotations is narrowed from 19.0 to 4.2. Experiments on anchor-based detector (SECOND) also shows similar results. The effectiveness of CSOT and the additive effect of combining it with self-training techniques are clearly demonstrated in Lines 11-15 of Tab. 6." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.688, + 0.357, + 0.704 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.719, + 0.786, + 0.84 + ], + "angle": 0, + "content": "In this work, we propose a novel Cross-Scan Object Transfer paradigm for LiDAR semi-supervised object detection. It features a HotspotNet for transferring objects semantic coherence manner, which reduces false positive predictions and enables the proposed CSOT to reach its full potential. To the best of our knowledge, this work is the first successful attempt of object transfer in LiDAR SSOD. The proposed CSOT paradigm is also orthogonal to the self-training techniques. We achieve competitive detectors on par with fully-supervised baselines while using 20x and 100x reduced annotations on nuScenes and Waymo dataset." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.179, + 0.577, + 0.195 + ], + "angle": 0, + "content": "This work was supported by the InnoHK funding." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.22, + 0.323, + 0.236 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.253, + 0.784, + 0.294 + ], + "angle": 0, + "content": "1. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In: CVPR. pp. 1090-1099 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.295, + 0.785, + 0.322 + ], + "angle": 0, + "content": "2. Blum, A., Mitchell, T.: Combining labeled and unlabeled data with co-training. In: COLT. pp. 92-100 (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.785, + 0.351 + ], + "angle": 0, + "content": "3. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR. pp. 11621-11631 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.352, + 0.785, + 0.392 + ], + "angle": 0, + "content": "4. Chen, B., Chen, W., Yang, S., Xuan, Y., Song, J., Xie, D., Pu, S., Song, M., Zhuang, Y.: Label matching semi-supervised object detection. In: CVPR. pp. 14381-14390 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.394, + 0.785, + 0.42 + ], + "angle": 0, + "content": "5. Chen, B., Li, P., Chen, X., Wang, B., Zhang, L., Hua, X.S.: Dense learning based semi-supervised object detection. In: CVPR. pp. 4815-4824 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.422, + 0.785, + 0.448 + ], + "angle": 0, + "content": "6. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Largekernel3d: Scaling up kernels in 3d sparse cnns. In: CVPR. pp. 13488-13498 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.449, + 0.785, + 0.476 + ], + "angle": 0, + "content": "7. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In: CVPR. pp. 21674-21683 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.478, + 0.785, + 0.505 + ], + "angle": 0, + "content": "8. Choi, J., Song, Y., Kwak, N.: Part-aware data augmentation for 3d object detection in point cloud. In: IROS. pp. 3391-3397 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.506, + 0.785, + 0.533 + ], + "angle": 0, + "content": "9. Dwibedi, D., Misra, I., Hebert, M.: Cut, paste and learn: Surprisingly easy synthesis for instance detection. In: ICCV. pp. 1301-1310 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.534, + 0.785, + 0.574 + ], + "angle": 0, + "content": "10. Fan, L., Pang, Z., Zhang, T., Wang, Y.X., Zhao, H., Wang, F., Wang, N., Zhang, Z.: Embracing single stride 3d object detector with sparse transformer. In: CVPR. pp. 8458-8468 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.576, + 0.785, + 0.616 + ], + "angle": 0, + "content": "1. Fang, J., Zuo, X., Zhou, D., Jin, S., Wang, S., Zhang, L.: Lidar-aug: A general rendering-based augmentation framework for 3d object detection. In: CVPR. pp. 4710-4720 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.617, + 0.785, + 0.644 + ], + "angle": 0, + "content": "2. Gao, J., Wang, J., Dai, S., Li, L.J., Nevatia, R.: Note-rcnn: Noise tolerant ensemble rcnn for semi-supervised object detection. In: ICCV. pp. 9508-9517 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.645, + 0.785, + 0.686 + ], + "angle": 0, + "content": "3. Ghiasi, G., Cui, Y., Srinivas, A., Qian, R., Lin, T.Y., Cubuk, E.D., Le, Q.V., Zoph, B.: Simple copy-paste is a strong data augmentation method for instance segmentation. In: CVPR. pp. 2918-2928 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.688, + 0.785, + 0.728 + ], + "angle": 0, + "content": "4. Hu, J.S.K., Waslander, S.L.: Pattern-aware data augmentation for lidar 3d object detection. In: IEEE International Intelligent Transportation Systems Conference (ITSC). pp. 2703-2710 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.785, + 0.756 + ], + "angle": 0, + "content": "5. Hu, X., Duan, Z., Ma, J.: Context-aware data augmentation for lidar 3d object detection. In: ICIP. pp. 11-15 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "6. Hu, Y., Ding, Z., Ge, R., Shao, W., Huang, L., Li, K., Liu, Q.: Afdetv2: Rethinking the necessity of the second stage for object detection from point clouds. In: AAAI. pp. 969-979 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Huang, D., Chen, Y., Ding, Y., Liao, J., Liu, J., Wu, K., Nie, Q., Liu, Y., Wang, C., Li, Z.: Rethinking dimensionality reduction in grid-based 3d object detection. arXiv preprint arXiv:2209.09464 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.253, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "J. Zhan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "18. Jaderberg, M., Simonyan, K., Vedaldi, A., Zisserman, A.: Synthetic data and artificial neural networks for natural scene text recognition. arXiv preprint arXiv:1406.2227 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "19. Kingma, D.P., Rezende, D.J., Mohamed, S., Welling, M.: Semi-supervised learning with deep generative models. In: Adv. Neural Inform. Process. Syst. (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.219, + 0.785, + 0.247 + ], + "angle": 0, + "content": "20. Laine, S., Aila, T.: Temporal ensembling for semi-supervised learning. In: ICLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.248, + 0.785, + 0.275 + ], + "angle": 0, + "content": "21. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J.: Pointpillars: Fast encoders for object detection from point clouds. In: CVPR. pp. 12697-12705 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.785, + 0.331 + ], + "angle": 0, + "content": "22. Lee, D., Lee, J., Lee, J., Lee, H., Lee, M., Woo, S., Lee, S.: Regularization strategy for point cloud via rigidly mixed sample. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 15900-15909 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.785, + 0.36 + ], + "angle": 0, + "content": "23. Lee, D.H.: Pseudo-label : The simple and efficient semi-supervised learning method for deep neural networks. In: ICML. p. 896 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.785, + 0.401 + ], + "angle": 0, + "content": "24. Lee, S., Lim, H., Myung, H.: Patchwork++: Fast and robust ground segmentation solving partial under-segmentation using 3d point cloud. In: IROS. pp. 13276-13283 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.402, + 0.785, + 0.444 + ], + "angle": 0, + "content": "25. Li, E., Wang, S., Li, C., Li, D., Wu, X., Hao, Q.: Sustech points: A portable 3d point cloud interactive annotation platform system. IEEE Intelligent Vehicles Symposium (IV) pp. 1108-1115 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.785, + 0.472 + ], + "angle": 0, + "content": "26. Li, H., Wu, Z., Shrivastava, A., Davis, L.S.: Rethinking pseudo labels for semi-supervised object detection. In: AAAI. pp. 1314-1322 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.515 + ], + "angle": 0, + "content": "27. Liang, H., Jiang, C., Feng, D., Chen, X., Xu, H., Liang, X., Zhang, W., Li, Z., Gool, L.V.: Exploring geometry-aware contrast and clustering harmonization for self-supervised 3d object detection. In: ICCV. pp. 3293-3302 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.785, + 0.558 + ], + "angle": 0, + "content": "28. Liu, C., Zhang, W., Lin, X., Zhang, W., Tan, X., Han, J., Li, X., Ding, E., Wang, J.: Ambiguity-resistant semi-supervised learning for dense object detection. In: CVPR. pp. 15579-15588 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.785, + 0.599 + ], + "angle": 0, + "content": "29. Liu, C., Gao, C., Liu, F., Li, P., Meng, D., Gao, X.: Hierarchical supervision and shuffle data augmentation for 3d semi-supervised object detection. In: CVPR. pp. 23819-23828 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.785, + 0.628 + ], + "angle": 0, + "content": "30. Liu, C., Gao, C., Liu, F., Liu, J., Meng, D., Gao, X.: Ss3d: Sparsely-supervised 3d object detection from point cloud. In: CVPR. pp. 8428-8437 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.785, + 0.657 + ], + "angle": 0, + "content": "31. Liu, Y.C., Ma, C.Y., He, Z., Kuo, C.W., Chen, K., Zhang, P., Wu, B., Kira, Z., Vajda, P.: Unbiased teacher for semi-supervised object detection. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.785, + 0.685 + ], + "angle": 0, + "content": "32. Liu, Y.C., Ma, C.Y., Kira, Z.: Unbiased teacher v2: Semi-supervised object detection for anchor-free and anchor-based detectors. In: CVPR. pp. 9819-9828 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.785, + 0.741 + ], + "angle": 0, + "content": "33. Meng, Q., Wang, W., Zhou, T., Shen, J., Jia, Y., Gool, L.V.: Towards a weakly supervised framework for 3d point cloud object detection and annotation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 44(8), 4454-4468 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.785, + 0.77 + ], + "angle": 0, + "content": "34. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR. pp. 652-660 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.785, + 0.812 + ], + "angle": 0, + "content": "35. Qian, G., Li, Y., Peng, H., Mai, J., Hammoud, H.A.A.K., Elhoseiny, M., Ghanem, B.: Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In: Adv. Neural Inform. Process. Syst. pp. 23192-23204 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "36. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In: CVPR. pp. 10529–10538 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "37. Sohn, K., Zhang, Z., Li, C.L., Zhang, H., Lee, C.Y., Pfister, T.: A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.259 + ], + "angle": 0, + "content": "38. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhao, S., Cheng, S., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: CVPR. pp. 2446-2454 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.788, + 0.3 + ], + "angle": 0, + "content": "39. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In: Adv. Neural Inform. Process. Syst. (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.788, + 0.342 + ], + "angle": 0, + "content": "40. Wang, H., Cong, Y., Litany, O., Gao, Y., Guibas, L.J.: 3dioumatch: Leveraging iou prediction for semi-supervised 3d object detection. In: CVPR. pp. 14615-14624 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.788, + 0.37 + ], + "angle": 0, + "content": "41. Wang, T., Yang, T., Cao, J., Zhang, X.: Co-mining: Self-supervised learning for sparsely annotated object detection. In: AAAI. pp. 2800-2808 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.788, + 0.411 + ], + "angle": 0, + "content": "42. Wang, X., Yang, X., Zhang, S., Li, Y., Feng, L., Fang, S., Lyu, C., Chen, K., Zhang, W.: Consistent-teacher: Towards reducing inconsistent pseudo-targets in semi-supervised object detection. In: CVPR. pp. 3240-3249 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.788, + 0.452 + ], + "angle": 0, + "content": "43. Wang, Y., Yin, J., Li, W., Frossard, P., Yang, R., Shen, J.: Ssda3d: Semi-supervised domain adaptation for 3d object detection from point cloud. In: AAAI. pp. 2707-2715 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.788, + 0.479 + ], + "angle": 0, + "content": "44. Wu, Z., Bodla, N., Singh, B., Najibi, M., Chellappa, R., Davis, L.S.: Soft sampling for robust object detection. arXiv preprint arXiv:1806.06986 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.788, + 0.522 + ], + "angle": 0, + "content": "45. Xu, M., Zhang, Z., Hu, H., Wang, J., Wang, L., Wei, F., Bai, X., Liu, Z.: End-to-end semi-supervised object detection with soft teacher. In: ICCV. pp. 3060-3069 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.788, + 0.563 + ], + "angle": 0, + "content": "46. Xu, R., Wang, T., Zhang, W., Chen, R., Cao, J., Pang, J., Lin, D.: Mv-jar: Masked voxel jigsaw and reconstruction for lidar-based self-supervised pre-training. In: CVPR. pp. 13445-13454 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.788, + 0.59 + ], + "angle": 0, + "content": "47. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.788, + 0.619 + ], + "angle": 0, + "content": "48. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d: Self-training for unsupervised domain adaptation on 3d object detection. In: CVPR. pp. 10368-10378 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.788, + 0.66 + ], + "angle": 0, + "content": "49. Yang, Q., Wei, X., Wang, B., Hua, X.S., Zhang, L.: Interactive self-training with mean teachers for semi-supervised object detection. In: CVPR. pp. 5941-5950 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.788, + 0.688 + ], + "angle": 0, + "content": "50. Yang, Y., Fan, L., Zhang, Z.: Mixsup: Mixed-grained supervision for labeled efficient lidar-based 3d object detection. In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.788, + 0.729 + ], + "angle": 0, + "content": "51. Ye, D., Zhou, Z., Chen, W., Xie, Y., Wang, Y., Wang, P., Foroosh, H.: Lidar-multinet: Towards a unified multi-task network for lidar perception. In: AAAI. pp. 3231-3240 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.788, + 0.771 + ], + "angle": 0, + "content": "52. Yin, J., Fang, J., Zhou, D., Zhang, L., Xu, C.Z., Shen, J., Wang, W.: Semi-supervised 3d object detection with proficient teachers. In: ECCV. pp. 727-743 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.788, + 0.813 + ], + "angle": 0, + "content": "53. Yin, J., Zhou, D., Zhang, L., Fang, J., Xu, C.Z., Shen, J., Wang, W.: Proposal-contrast: Unsupervised pre-training for lidar-based 3d object detection. In: ECCV. pp. 17-33 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.788, + 0.84 + ], + "angle": 0, + "content": "54. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: CVPR. pp. 11784-11793 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "J. Zhan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "55. You, Y., Luo, K., Phoo, C.P., Chao, W.L., Sun, W., Hariharan, B., Campbell, M., Weinberger, K.Q.: Learning to detect mobile objects from lidar scans without labels. In: CVPR. pp. 1130-1140 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "56. Yuan, J., Zhang, B., Yan, X., Chen, T., Shi, B., Li, Y., Qiao, Y.: Ad-pt: Autonomous driving pre-training with large-scale point cloud dataset. In: Adv. Neural Inform. Process. Syst. p. 36 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "57. Yun, S., Han, D., Oh, S.J., Chun, S., Cho, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: CVPR. pp. 6023-6032 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "58. Yurtsever, E., Erçelik, E., Liu, M., Yang, Z., Zhang, H., Topçam, P., Listl, M., Çayli, Y.K., Knoll, A.: 3d object detection with a self-supervised lidar scene flow backbone. In: ECCV. pp. 247-265 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "59. Zhang, H., Chen, F., Shen, Z., Hao, Q., Zhu, C., Savvides, M.: Solving missing-annotation object detection with background recalibration loss. In: ICASSP. pp. 1888-1892 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.384 + ], + "angle": 0, + "content": "60. Zhang, L., Sun, Y., Wei, W.: Mind the gap: Polishing pseudo labels for accurate semi-supervised object detection. In: AAAI. pp. 3463-3471 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "61. Zhang, L., Yang, A.J., Xiong, Y., Casas, S., Yang, B., Ren, M., Urtasun, R.: Towards unsupervised object detection from lidar point clouds. In: CVPR. pp. 9317-9328 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "62. Zhao, N., Chua, T.S., Lee, G.H.: Sess: Self-ensembling semi-supervised 3d object detection. In: CVPR. pp. 11079-11087 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.48 + ], + "angle": 0, + "content": "63. Zheng, W., Tang, W., Jiang, L., Fu, C.W.: Se-ssd: Self-ensembling single-stage object detector from point cloud. In: CVPR. pp. 14494–14503 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "64. Zhou, H., Ge, Z., Liu, S., Mao, W., Li, Z., Yu, H., Sun, J.: Dense teacher: Dense pseudo-labels for semi-supervised object detection. In: ECCV. pp. 35-50 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "65. Zhou, Q., Yu, C., Wang, Z., Qian, Q., Li, H.: Instant-teaching: An end-to-end semi-supervised object detection framework. In: CVPR. pp. 4081-4090 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.564 + ], + "angle": 0, + "content": "66. Zhu, B., Jiang, Z., Zhou, X., Li, Z., Yu, G.: Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.564 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CSOT_ Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection/4b43039d-68da-4e38-a8cf-39348329a3bc_origin.pdf b/2024/CSOT_ Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection/4b43039d-68da-4e38-a8cf-39348329a3bc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..70d92c50fc23a2ecb4f72843a8bd9a18873edcbb --- /dev/null +++ b/2024/CSOT_ Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection/4b43039d-68da-4e38-a8cf-39348329a3bc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7b581b1d0f156face043101957c85e3712b24e12d383ea2b255b83d6da24e04 +size 859275 diff --git a/2024/CSOT_ Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection/full.md b/2024/CSOT_ Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..635893a12d4cbc6c510b0f7718194cb3d5222471 --- /dev/null +++ b/2024/CSOT_ Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection/full.md @@ -0,0 +1,270 @@ +# CSOT: Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection + +Jinglin Zhan $^{1}$ , Tiejun Liu $^{1}$ , Rengang Li $^{1}$ , Zhaoxiang Zhang $^{2}$ , and Yuntao Chen $^{3\star}$ + +1 IEIT Systems + +$^{2}$ Institute of Automation, CAS + +3 Centre for Artificial Intelligence and Robotics, HKISI, CAS {zhanjinglin, liutj, lirg}@ieisystem.com, zhaoxiang.zhang@ia.ac.cn, chenyuntao08@gmail.com + +Abstract. Large-scale 3D bounding box annotation is crucial for LiDAR object detection but comes at a high cost. Semi-supervised object detection (SSOD) offers promising solutions to leverage unannotated data, but the predominant pseudo-labeling approach requires careful hyperparameter tuning for training on noisy teacher labels. In this work, we propose a Cross-Scan Object Transfer (CSOT) paradigm for LiDAR SSOD. Central to our approach is HotspotNet, a transformer-based network that predicts possible placement locations and the object-place fitness scores for inserting annotated objects into unlabeled scans in a semantic coherence manner. Based on HotspotNet, CSOT successfully enables object copy-paste in LiDAR SSOD for the first time. To train object detectors on partially annotated scans generated by CSOT, we adopt a spatial-aware classification loss throughout our partial supervision to handle false negative issues caused by treating all unlabeled objects as background. We conduct extensive experiments to verify the efficacy and generality of our method. Compared to other state-of-the-art label-efficient methods used in LiDAR detection, our approach requires the least amount of annotation while achieves the best detector. Using only $1\%$ of the labeled data on the Waymo dataset, our semi-supervised detector achieves performance on par with the fully supervised baseline. Similarly, on the nuScenes dataset, our semi-supervised CenterPoint reaches $99\%$ of the fully supervised model's detection performance in terms of NDS score, while using just $5\%$ of the labeled data. Code is released at https://github.com/JinglinZhan/CSOT + +Keywords: Semi-Supervised Learning $\cdot$ LiDAR Object Detection $\cdot$ Autonomous Driving + +# 1 Introduction + +Large-scale annotated datasets have catalyzed the success of LiDAR object detection, but at a significant cost in terms of money and time [25,33]. For instance, + +![](images/a0cfdbec5a5b1e466502e6b36e1b8f1e6949a579beae5cfb01e3f3b675310284.jpg) +Fig. 1: Preliminary experiments of CSOT pipeline for LiDAR SSOD. (a) Compare CSOT implemented based on different cross-scan object transfer methods and display the NDS differences. (b) Visualize prediction differences among various teacher model that used to generate pseudo labels for self-training. + +annotating the 12 million 3D boxes in the Waymo Open [38] dataset may have consumed millions of dollars and tens of thousands of human hours. In contrast, autonomous driving fleets capture thousands, and even millions, of hours of unannotated sensor data every day. The huge gap between the speed of manual labeling and that of data accumulation makes labeling a major slowdown. This calls for a more scalable approach to leverage the abundant data collected by autonomous fleets and continually advance object detection capabilities. + +Semi-Supervised Learning (SSL) presents a promising way for reducing annotation demands by leveraging both limited labeled data and vast quantities of unlabeled data during training. Various SSL methods have been proposed, including pseudo-labeling (a.k.a. self-training) [23], co-training [2], consistency regularization [20] and generative modeling [19]. Among these SSL approaches, pseudo-labeling has gained significant traction, particularly in the field of semi-supervised object detection (SSOD) [4, 12, 31, 37, 39, 40, 42, 52, 62, 65]. However, generating reliable pseudo-labels without human intervention remains a significant challenge. Noisy pseudo-labels can negatively impact model performance, leading to suboptimal results. To address this, researchers have proposed various techniques, including sophisticated post-processing mechanisms [4, 5, 26, 28, 40, 52, 60], iterative training procedures [39, 45, 49], and noise-tolerant detectors [12, 65] and loss objectives [31, 32, 62, 64]. Nevertheless, these approaches often involve complex pipelines and require careful parameter tuning, making them inefficient and difficult to adapt to new datasets. + +In contrast with noisy pseudo-labeling, we leverage a novel semi-supervised learning approach for 3D object detection that transfers annotated objects to generate noise-free training data. Specifically, we transfer point clouds and bound + +ing boxes of annotated objects to unannotated LiDAR scans to create partially labeled data. Object transfer via copy-paste is a commonly employed data augmentation [6,7,16,47,51] in LiDAR object detection. But naively applying object copy-paste [11, 15, 47] to semi-supervised LiDAR object detection yields suboptimal results. As shown in Fig. 1, SSOD built upon vanilla object transfer does improve performance over the baseline, but combining it with self-training results in a negative net effect on detection performance compared to using self-training alone. The key insight is that object copy-paste without careful contextual consistency maintenance improves detector performance at the cost of generating a large number of false positives. Although the commonly used object detection metric mAP is not sensitive to false positives if the recall improves, the self-training pipeline accumulates these false positives over iterations. + +In order to achieve coherent object copy-paste, we propose Hotspot Network, a learning-based approach to insert objects with strong contextual consistency, enabling object-transfer-based semi-supervised LiDAR object detection to achieve its full potential. HotspotNet takes unannotated LiDAR scans as inputs, along with point clouds and bounding boxes of annotated objects. It then predicts possible placement locations for these objects and assigns scores to each location, indicating the likelihood of the object fitting well within the context of the scanned environment. This allows us introducing noise-free supervision signal into unlabeled point clouds and use these newly created partially-labeled data to optimize detector performance. Based on HotspotNet, we design a novel Cross-Scan Object Transfer(CSOT) pipeline for LiDAR SSOD. CSOT enjoys precise partial annotations that maintain great contextual coherence with unannotated scans rather than noisy pseudo labels for training detectors on unlabeled data. It's important to note that training detectors on partially annotated point clouds, as opposed to fully labeled ones, can lead to a significant issue: unlabeled objects being miscategorized as background, resulting in a high number of false negatives. To mitigate this problem, we propose a spatial-aware classification loss and employ it throughout our partial supervision. Our CSOT paradigm features a streamlined training pipeline and requires minimal hyper-parameter tuning, enabling it to quickly generalize across various datasets and detectors. Moreover, this approach is orthogonal and complementary to commonly used self-training techniques. We conducted extensive experiments on the Waymo [38] and nuScenes [3] datasets to confirm the effectiveness of our methods. + +Our contributions are summarized below: + +- We propose a Hotspot Network for instructing object placement in a semantic coherence manner, and employ an asymmetric supervision for training HotspotNet on both annotated and unannotated data. +- Based on HotspotNet, we design a novel Cross-Scan Object Transfer paradigm for LiDAR semi-supervised object detection. To the best of our knowledge, this is the first successful attempt of object copy-paste in LiDAR SSOD. +- We conduct extensive experiments on nuScenes and Waymo datasets to verify the efficacy and generality of our method. Our approach achieves competitive performance compared to fully-supervised baselines while using 20x + +to 100x fewer annotations, significantly outperforming other state-of-the-art label-efficient approaches in LiDAR object detection. + +# 2 Related Work + +# 2.1 Semi-Supervised Object Detection + +Unsupervised learning eliminates data annotation but significantly degrades model performance [48, 55, 61]. Semi-supervised methods maintain low annotation costs meanwhile greatly bridges the gap to fully-supervised approaches [43]. As a prior work, STAC [37] proposes a pseudo labeling SSOD paradigm and catalyzes extensive follow-on researches. It trains a teacher with all annotated data and generates pseudo labels for supervising the student on unlabeled data. Producing reliable pseudo labels is critical in this paradigm and numerous methods have been explored. 3D IoUMatch [40] utilizes the estimated 3D IoU to filter poorly localized predictions. LabelMatch [4] leverages adaptive label-distribution-aware confidence thresholds for generating unbiased pseudo labels. Proficient Teacher [52] employs clustering-based box voting for improving the precision of final predictions. HSSDA [29] designs a dynamic dual-threshold strategy to generate more reasonable supervision for training the student model. Besides various filter mechanism, many other approaches, such as iterative training procedures [39, 62], noise-tolerant detectors [12, 65] and a variety of loss objectives with consistency or reweighting terms [31, 42, 62], are also proposed for alleviating negative effects caused by noisy pseudo labels. Despite some progress has been made, unreliable pseudo-labels still remain an open challenge. The sophisticated pipelines with extensive manual threshold tuning also impede these approaches' quick adaptation to new datasets. + +# 2.2 Copy-Paste For LiDAR Object Detection + +Copy-paste method is developed by Jaderberg et.al. for text recognition [18], then extended by Dwibedi et.al. [9] and Yun et.al. [57] for 2D object detection. Yan et.al. [47] introduces copy-paste into LiDAR object detection and develops the well known ground-truth augmentation approach (so called GT-Aug), which is extensively adopted in many later studies [1,6,16,17,21,51,54,66]. Since then, numerous approaches have been proposed to optimize GT-Aug. Part-Aware [8] and Shape-Aware [63] GT-Aug divide objects into partitions and apply different augmentation methods to each partition. Pattern-Aware [14] GT-Aug downsamples the points of objects to create a new one with farther distance. PointCut-Mix [22] replaces part of the sample with shape-preserved subsets from another one. Some researchers notice placement issues in GT-Aug and develop hand-craft valid maps to restrict the insertion of objects [11,15]. Up to now, researches on copy-paste is primarily focused on exploring and developing its capabilities for augmenting data to support fully-supervised learning tasks. Ghiasi et.al. introduce copy-paste into SSL [13] for augmenting both pseudo labeled and supervised labeled images. The above trial still limited in the field of data augmentation and the copy-paste technique remains largely untapped. + +![](images/57f1dfe65a9d740e6d0734a6d3dcc14625ba67651a748d8abde37c147147fbf8.jpg) +Fig. 2: The overall architecture of HotspotNet. The Object Encoders extract features from object point cloud $X_{obj,pc}$ and bounding box meta data $X_{obj,meta}$ . The scene encoder featurize the scene point cloud and cross-attends to object features $\mathcal{F}_{\mathrm{obj,pc}}$ and $\mathcal{F}_{\mathrm{obj,meta}}$ . The location decoder predicts the placeable locations and their scores. + +# 3 Method + +We provide details of Cross-Scan Object Transfer SSOD paradigm in this section. We present the overall architecture of HotspotNet in Sec. 3.1, and describe the Asymmetric Supervision for training HotspotNet in Sec. 3.2. The pipeline of our newly proposed SSOD paradigm, CSOT, is elaborated in Sec. 3.3. + +# 3.1 Hotspot Network + +The key to LiDAR copy-paste is identifying semantically coherent locations within background scans to paste objects. Placing objects at unreasonable locations, such as inside walls, floating in sky or with weird pose, will cause a large amount of false positives and thus deteriorate the detector performance. + +We propose a learning-based approach for object placement that provides sufficient supervision signal and maintain a high signal-to-noise ratio. Specifically, we employ a Hotspot Network to predict the placeable location $\mathbf{b}$ and the object-place fitness scores $o$ for inserting object in a semantic coherence manner. + +We abbreviate Hotspot Network as HotspotNet and visualize the overall architecture in Fig. 2. The HotspotNet comprises object encoders for object point cloud $X_{\mathrm{obj,pc}}$ and meta data $X_{\mathrm{obj,meta}}$ , respectively. The scene encoder featurizes the scene point cloud and cross-attends to object features $\mathcal{F}_{\mathrm{obj,pc}}$ and $\mathcal{F}_{\mathrm{obj,meta}}$ . A location decoder predicts the placeable location $\mathbf{b} = \{x,y,z,w,l,h,\theta ,c\}$ and the location fitness score $o$ . + +Object Encoder We process the object point cloud $X_{\mathrm{obj,pc}}$ with object point cloud encoders like PointNet [34] or PointNeXt [35] and obtain object features $\mathcal{F}_{\mathrm{obj,pc}}$ . We encode the bounding box meta $X_{\mathrm{obj,meta}}$ with appropriate meta feature encoders like Fourier embedding for boxes, and CLIP or one-hot embedding + +for object classes. We then transform the initial box embedding with a MLP to obtain bounding box meta feature $\mathcal{F}_{\mathrm{obj,meta}}$ . + +$$ +\mathcal {F} _ {\mathrm {o b j}, \mathrm {p c}} = 0 \text {b j e c t P o i n t C l o u d E n c o d e r} \left(X _ {\mathrm {o b j}, \mathrm {p c}}\right) \tag {1} +$$ + +$$ +\mathcal {F} _ {\text {o b j , m e t a}} = \operatorname {M L P} \left(\text {F o u r i e r E m b e d} \left(X _ {\text {o b j , b o x}}\right) \mid \mid \text {C l a s s E m b e d} \left(X _ {\text {o b j , c l a s s}}\right)\right) \tag {2} +$$ + +Scene Encoder Given a scene LiDAR scan $X_{\mathrm{scene}}$ , we leverage off-the-shelf point cloud encoders like CenterPoint [54] or SST [10] for extracting features $\mathcal{F}_{\mathrm{scene}}$ . We use spatial cross-attention in the first convolution block of each stage of the scene point cloud encoder. This allows the scene encoder aggregating features from both the object point clouds and the box metadata for assessing the object fitness of each spatial location. + +Location Decoder We randomly initialize a set of learnable queries $\mathbf{q}_0$ . These queries are combine with position embeddings to obtain location queries $\mathbf{q}$ . Our location decoder leverages both scene features and object features for predicting possible inserting locations and scores. + +# 3.2 Training HotspotNet with Asymmetric Supervision + +The HotspotNet consists of a location predictor and a location scorer that work jointly to suggest object placements. The location predictor outputs candidate object positions, while the scorer assigns fitness scores to evaluate how semantically consistent each candidate is. We employ a multi-task learning approach that features Asymmetric Supervision for optimizing different predictions. The proposed losses supervise the location predictor and scorer to accurately position objects within a new environment. + +To train the location predictor, we utilize the annotated LiDAR scans. For each annotated scan, we randomly extract an object point cloud from the original point clouds along with its corresponding bounding box. The regressor is then supervised with an $\ell_1$ loss between the predicted location $\mathbf{b}$ and the extracted bounding box $\mathbf{b}_{\mathrm{cut - out}}$ . We constrain object distance and perspective from the ego car in Eq. 3 to preserve perspective and point density. We employ dropout regularization on the bounding box meta data to prevent the location predictor from directly copying the meta data to the prediction output. It encourages the predictor to learn meaningful features from the point clouds itself, rather than relying solely on the provided meta data. + +$$ +\mathcal {L} _ {\mathrm {l o c}} (\mathbf {b}) = \left\| \mathbf {b} - \mathbf {b} _ {\text {c u t - o u t}} \right\| _ {1} \tag {3} +$$ + +To train the location scorer, we utilize an asymmetric supervision approach for positive and negative samples. Positive samples are obtained by matching location queries with the extracted object point clouds. These positive samples encourage the scorer to assign high scores to queries that accurately position objects within the scan. + +$$ +\mathcal {L} _ {\text {s c o r e}} ^ {p} (o) = - \log o \tag {4} +$$ + +To avoid treating all queries that do not match the extracted objects as negative samples, a careful selection process is employed. Instead of using all non-matching queries, negative samples are chosen based on queries that have low fitness scores, denoted by $s$ . This approach ensures that the scorer learns to differentiate between suitable and unsuitable object placements, focusing on the most informative negative samples. + +$$ +\mathcal {L} _ {\mathrm {s c o r e}} ^ {n} (o) = - \left(1 - \mathbb {1} \left[ s _ {p} \cdot s _ {o} \cdot s _ {c} > \tau \right]\right) \log (1 - o) \tag {5} +$$ + +The fitness scores $s_p$ , $s_o$ , and $s_c$ are used to measure the accuracy and coherence of an object's position, orientation, and category within the scene. The position score $s_p = \exp(-(\mathbf{b}_{xyz} \cdot \mathbf{n}_g + d_g) / \sigma_p^2)^1$ penalizes the predicted location $\mathbf{b}_{xyz}$ deviating from the local ground plane specified by $(\mathbf{n}_g, d_g)$ . Local ground planes here could be obtained via non-learning ground detection methods like Patch-Works++ [24]. Similarly, the orientation score $s_o = \cos((\mathbf{b}_\theta - \hat{\mathbf{b}}_\theta) / 2)$ encourages consistency between the heading of the new object location and the predominant heading $\hat{\mathbf{b}}_\theta$ of objects in the same category within the scene. Finally, the category compatible score $s_c = \mathbb{1}[\mathbf{b}_c \in \mathcal{C}]$ prevents pasting object of category that is not in the current scene (e.g. bicycle on highway). Notably, these scores can be computed not only between the extracted object and its original scan but also between the extracted object and other unannotated LiDAR scans, in which $\hat{\mathbf{b}}_\theta$ can be simplified as ego orientation and $s_c$ can be ignored directly. This adaptability opens up the possibility of learning object placement scoring from a more extensive dataset through unsupervised learning, which assigns positive and negative samples solely based on the fitness scores. + +Combining these asymmetric losses and carefully selecting positive and negative samples, our HotspotNet learns to accurately position objects within a new scan while maintaining semantic coherence. + +# 3.3 Cross-Scan Object Transfer for LiDAR-based SSOD + +Under the guidance of HotspotNet (described in Sec. 3.1), we inject precise supervision signals into unannotated point clouds and train detectors on these partially-labeled data. Fig. 3 illustrates the differences between our Cross-Scan Object Transfer SSOD paradigm and the commonly used self-training technique. Our method bypasses the iterative self-training procedures, cumbersome pseudolabel filtering mechanisms, and complex loss objective designs. + +Partially Supervised Object Detection Our approach for introducing supervision into unlabeled LiDAR scans is straightforward. We denote all objects in labeled point clouds as $\{\mathbf{O}^L\}$ , and all unlabeled scans as $\{\mathbf{S}^U\}$ . For an unlabeled LiDAR scan $\mathbf{S}^U$ , we randomly sample a subset of $\{\mathbf{O}^L\}$ and follow the instruction of HotspotNet to inject $\{\mathbf{O}^L\}$ into $\mathbf{S}^U$ in a semantically coherent manner. Since inserted objects come with ground-truth classes, poses and dimensions, + +![](images/85c682fbbcaea53f860e6d584522858f6ebb12f03a2ccf8115d128d833ec7bcd.jpg) +Fig. 3: (a) Common-used pseudo-labeling SSOD paradigm. (b) Cross-Scan Object Transfer SSOD paradigm, which bypasses the iterative self-training procedures, cumbersome pseudo-labels filter mechanisms and complex loss objectives designs. + +these newly created scans allow for supervision. However, it is important to note that the above synthesized point clouds are only partially annotated, as we lack information about the original unlabeled objects in these unannotated scans. Learning a detector from partially labeled data is known as Sparsely Annotated Object Detection (SAOD) [30, 41]. The differences of training a detector from fully and partially annotated data will be discussed in the next part. + +Spatial-Aware Classification Loss An object detector generally includes a classifier to differentiate foreground objects from background clutter, and a localizer to predict the accurate location, dimension and orientation of an object. The localizer is typically supervised by regressing the offset from current voxel location to actual bounding boxes of foreground objects, so it can be naturally supervised by new point clouds generated via cross-scan object transfer. However, the classifier requires both positive and negative samples for learning a decision boundary. Once again, the inserted objects could provide positive samples for the classifier but we can not treat all other voxels as background in SAOD just like vanilla object detection. This is due to the existence of the so-called hard negatives as reported in other SAOD works [41]. + +In our approach, hard negatives refer to the original unlabeled objects intrinsically present in the unannotated LiDAR scans. Training the classifier in our semi-supervised detector with these hard negatives will result in an unacceptable number of false negatives. Instead of resorting to sophisticated re-weighting/ recalibration strategies [44,59] or teacher-student siamese networks [41], we simply + +choose negative samples in a spatial-aware manner to avoid sampling unlabeled objects as negatives. We propose a spatial-aware classification loss for our partial supervision as Eq.6: + +$$ +L _ {\mathrm {c l s}} ^ {\mathrm {n}} (\mathbf {x}) = \mathbb {1} \left[ \left(\min _ {i} \| \mathbf {x} - \mathbf {c} _ {i} \| _ {2}\right) \leq \sigma \right] \cdot L _ {\mathrm {c l s}} ^ {\mathrm {n}} (\mathbf {x}) \tag {6} +$$ + +where $L_{\mathrm{cls}}^{\mathrm{n}}(\mathbf{x})$ denotes focal loss for negative samples at the heatmap location $\mathbf{x}$ . The indicator function $\mathbb{1}(\cdot)$ equals 1 when the $\ell 2$ distance between a location $\mathbf{x}$ and the center location $\mathbf{c}$ of any foreground object is less than a threshold $\sigma$ , and 0 otherwise. We set $\sigma$ as twice of the Gaussian radius used for positive samples assignment in CenterPoint [54] for unlabeled scans, and infinity for fully annotated data. An immediate drawback of our spatial-aware classification loss is that there are too few negative samples for the classifier. To alleviate this issue, we randomly sample $1\%$ samples in the heatmap predictions as negative samples, and set the corresponding value of $\mathbb{1}(\cdot)$ as 1 to make the classification boundary between positive and negative samples more clear. + +Synergy with Self-Training Techniques A deeper analysis of the proposed CSOT SSOD paradigm reveals that the primary mechanism driving improved detection performance is the extrapolation of labeled object detection to novel contextual environments. This stands in contrast to self-training methods, which learn from pseudo-labeled objects in unannotated scans. + +As a result, the proposed CSOT pipeline for semi-supervised LiDAR object detection is orthogonal to, yet complementary with, existing self-training techniques [37], allowing for additional performance gains. Deliberately designing a pseudo-labeling SSOD framework is beyond the scope of this paper. To verify the complementary effect of our approach when combined with self-training techniques, we implement a basic teacher-student paradigm. We employ a single universal confidence score threshold $\tau_{\mathrm{pseudo}}$ , which is invariant across categories, to filter low-confidence predictions. We then utilize a fully-supervised loss, as in CenterPoint [54], for optimizing the student detector. + +# 4 Experiment + +The proposed methods are verified in this section. We show implementation details in Sec. 4.1 and provide all experimental results in Sec. 4.2 and Sec. 4.3. More ablations and discussions are shown in our supplementary materials. + +# 4.1 Experimental Setup + +Datasets We validate our method on two widely used autonomous driving datasets: nuScenes [3] and Waymo [38]. The nuScenes dataset has 700 training, 150 validation, and 150 test sequences. Official nuScenes detection metrics are NuScenes Detection Score (NDS) and mean Average Precision (mAP). The Waymo Open Dataset consists of 798 training and 202 validation sequences. Official Waymo detection metrics are mean Average Precision (mAP) and mean Average Precision weighted by Heading (mAPH). + +Training Details We use one-stage CenterPoint [54] as our baseline detector to validate the effectiveness of our methods. Voxel sizes are set as $[0.075\mathrm{m}, 0.075\mathrm{m}, 0.2\mathrm{m}]$ for nuScenes and $[0.1\mathrm{m}, 0.1\mathrm{m}, 0.15\mathrm{m}]$ for Waymo. The training epochs for one cycle of learning procedure, including fully-supervised learning, partially-supervised learning and self-training, are set as 20 for nuScenes and 7 for Waymo. We follow many previous work [6,51,54,66] to adopt Adam optimizer with one-cycle learning rate policy. Labeled data are uniformly sampled from the entire dataset. Then objects and their annotations are extracted to build the ground-truth database for injecting precise supervision signals into unlabeled scans. We finetune partially-supervised detectors on labeled scans with several epochs. We set a score threshold of 0.3 for nuScenes and 0.6 for Waymo to filter low-quality pseudo labels, and verify the orthogonality of our methods with self-training. + +# 4.2 Benchmarks + +Benchmarks on nuScenes dataset We compare our methods against other label-efficient LiDAR detection approaches on the nuScenes dataset. As shown in Tab. 1, our methods obtain the significant performance improvements over previous works. For the sake of fairness, we choose the fully supervised detector trained by Yin et.al. [54] (the designer of CenterPoint backbone) as the baseline detector, and list the results achieved from other label-efficient LiDAR object detection approaches for comparison [27, 43, 56, 58]. Without leveraging models pre-trained on external datasets and using only $5\%$ of the annotations, we achieve competitive detector which reaches $97.2\%$ of the fully-supervised performance on mAP and $98.9\%$ on NDS. We significantly shrink the mAP and NDS gaps between the baseline detector and the one trained with $20\mathrm{x}$ less annotations from around $10\sim 20$ to merely 1.6 and 0.7 respectively. Our semi-supervised detector, trained with only $5\%$ manual annotations, obtains over $90\%$ detection performance on 8 of the 10 categories defined in nuScenes dataset, and over $95\%$ detection performance on 6 categories. + +Benchmarks on Waymo dataset We extend our methods on Waymo dataset and achieve similar outstanding results. we choose the fully supervised detector trained by Shi et.al. [36], Yan et.al. [47], Fan et.al. [10] and Yin et.al. [54] as the baseline detector, and list the results achieved from other label-efficient LiDAR object detection approaches for comparison [29, 46, 50, 52, 53, 56]. As shown in Tab. 2, our methods are valid on both anchor-based (SECOND) and anchor-free (CenterPoint) detectors. Taking CenterPoint as an example, with merely $1\%$ annotations, we achieve a competitive detector which matches fully supervised performance on vehicle and pedestrian, and exceeds it on cyclist. The $100\mathrm{x}$ reduction in annotations leads to only a minor 1.8 APH decrease for L2 Vehicle, a negligible 0.4 APH drop for L2 Pedestrian, and boosts the APH of L2 Cyclist by 4.1. The abnormal enhanced detection performance on Cyclist should be attributed to our learning-based approach for cross-scan object transfer. Instructed by the proposed HotspotNet, we paste objects in a semantic coherent + +Table 1: Compare with other labeled efficient LiDAR object detection approaches on nuScenes Dataset validation split. Performances of the baseline detector reported by its original designer [54] are marked with gray backgrounds. + +
MethodRefAnnosExternalmAPNDSCarTruckC.V.BusTrailerBarrierMot.Byc.Ped.T.C.
BaselineCVPR 2021 [54]100%58.065.584.651.017.560.253.270.953.728.783.476.7
GCC3DICCV 2021 [27]5%41.146.8——————————
SceneFlowECCV 2022 [58]5%36.048.3——————————
SSDA3DAAAI 2023 [43]5%✓——76.2—————————
AD-PTNeurIPS 2023 [56]5%✓45.053.078.943.811.155.221.255.139.017.872.355.4
Ours—5%56.464.881.753.516.558.632.866.060.047.882.065.5
+ +Table 2: Compare with other labeled efficient LiDAR object detection approaches on Waymo Dataset validation split. Performances of the baseline detector reported by its original designers [10,36,47,54] are mark with gray backgrounds. + +
DetectorMethodRefAnnoxExtVehiclePedestrianCyclistAll L2 AP
L1L2L1L2L1L2
APAPHAPAPHAPAPHAPAPHAPAPHAPAPH
PV-RCNNBaselineCVPR 2020 [36]100%77.576.969.068.475.065.766.057.667.866.465.464.063.3
PV-RCNNHSSDACVPR 2023 [29]1%56.453.849.747.340.120.933.517.529.120.927.920.028.3
PV-RCNNMIXSUPICLR 2024 [50]1%———55.5———52.0———62.356.6
SECONDBaselineSensors 2018 [47]100%72.371.763.963.368.758.260.751.360.659.358.357.157.2
SECONDMIXSUPICLR 2024 [50]10%———55.0———49.6———58.154.2
SECONDProficientTeacherECCV 2022 [52]5%——53.052.5——50.338.7——49.946.045.7
SECONDAD-PTNeurIPS 2023 [56]3%✓——60.559.9——54.945.8——50.849.751.8
SECONDProposalContrastECCV 2022 [53]1%——37.636.9——39.731.7——37.735.734.8
SECONDOurs—1%69.468.961.260.864.060.755.152.361.160.458.958.257.1
SSTBaselineCVPR 2022 [10]100%74.273.865.565.178.769.670.061.770.769.668.066.964.6
SSTMIXSUPICLR 2024 [50]10%———59.1———60.0———63.160.7
SSTMV-JARCVPR 2023 [46]5%——56.556.0——57.747.7——37.436.346.7
CenterPointBaselineCVPR 2021 [54]100%76.676.068.968.479.073.471.065.872.171.069.568.567.6
CenterPointMIXSUPICLR 2024 [50]10%———61.8———57.7———67.562.3
CenterPointAD-PTNeurIPS 2023 [56]3%✓——60.459.8——60.654.0——62.761.658.5
CenterPointOurs—1%75.174.767.066.675.972.968.265.475.975.173.472.668.2
+ +manner, which enables our CSOT SSOD to reach its full potential and meanwhile can be used as a powerful data augmentation for fully-supervised detector. Related ablation results are discussed in Sec. 4.3. Notably, under the most stringent constraints with minimal labeled point clouds and without leveraging any external pre-training, our approach stands alone as the only label-efficient technique which is able to achieve equivalence detection performance to the baseline detector on the average L2 APH of all categories defined in Waymo dataset. + +# 4.3 Ablations + +Ablate training stage on nuScenes dataset. Our semi-supervised detectors trained with 20x fewer nuScenes annotations are compared to both fully supervised and pseudo-labeling semi-supervised models. Results in Tab. 3 indicate that our newly proposed CSOT is a valid label-efficient approach, and it almost completely maintains the performance gains of self-training when used together. The fully supervised detector trained with $5\%$ labeled scans shows substantial performance degradation compared to that trained on the entire labeled dataset, with a mAP decline from 59.5 to 43.8 and a NDS decrease from 66.0 to 56.5. As shown in Lines 3-5 of Tab. 3, self-training technique is able to boost the mAP from 43.8 to 51.0 and NDS from 56.5 to 61.0. Our CSOT, which allows + +Table 3: Compare with fully supervised and pseudo-labeling semi-supervised detectors. All results are achieved based on our implementation code of CenterPoint and evaluated on nuScenes dataset validation split. + +
AnnosScansCSOTSelf-TrainingmAPNDSCarTruckC.V.BusTrailerBarrierMot.Byc.Ped.T.C.
1100%100%59.566.083.756.618.766.934.867.163.750.484.568.7
25%5%43.856.575.839.712.344.120.556.936.318.476.756.9
35%100%✓51.061.080.149.114.655.626.762.345.430.881.963.1
45%100%✓50.461.079.646.411.750.125.659.152.137.280.662.0
55%100%✓✓56.464.881.753.516.558.632.866.060.047.882.065.5
+ +Table 4: Compare with fully supervised and pseudo-labeling semi-supervised detectors. All results are achieved based on our implementation code of CenterPoint and evaluated on Waymo dataset validation split (mean L1/L2 AP/APH over all categories defined in Waymo dataset). + +
AnnosScansCSOTSelf-TrainingObject TransferL1 AP/APHL2 AP/APHmAP/mAPH
1100%100%—74.4 / 72.768.6 / 67.071.5 / 69.9
1a100%100%GT-Aug74.7 / 73.168.8 / 67.371.8 / 70.2
1b100%100%Ours77.5 / 76.071.6 / 70.274.6 / 73.1
21%1%—68.3 / 66.662.3 / 60.865.3 / 63.7
31%100%✓—71.9 / 70.465.8 / 64.568.9 / 67.5
41%100%✓Ours71.4 / 69.865.4 / 63.968.4 / 66.9
51%100%✓✓Ours75.7 / 74.269.5 / 68.272.6 / 71.2
+ +supervision on unannotated point clouds by converting unlabeled scans into partially labeled scans via learning-based copy-paste, achieves a similar detection performance (with a mAP of 50.4 and a NDS of 61.0). It is noteworthy that iterative employment of self-training techniques does not lead to sustained gains in detection performance (see supplementary material), while incorporating just one cycle of self-training on our partially supervised detector is able to further optimize the mAP from 50.4 to 56.4 and NDS from 61.0 to 64.8. + +Ablate training stage on Waymo dataset We extend the above ablations on Waymo dataset and summarize all results in Tab. 4, which further validate the effectiveness of CSOT, as well as its orthogonality with the common used pseudo-labeling techniques. $100\mathrm{x}$ fewer labeled point clouds lead to a degradation of detection performance, reducing the mean AP and APH from 71.5, 69.9 to 65.3, 63.7. Our Cross-Scan Object Transfer paradigm brings a boost of around 3 APH for both LEVEL_1 and LEVEL_2 objects. CSOT is proved to be complimentary to self-training, and the performance gains can be completely maintained. As shown in the Line 5 of Tab. 4, the performance of detector trained with $1\%$ annotations can be finally optimized to match even surpass the fully-supervised baseline model. The abnormal enhanced detection performance with $100\mathrm{x}$ reduced labeled data should be attributed to the learning-based copy-paste operation employed in this work. As shown in Lines 1a-1b of Tab. 4, replacing the commonly used copy-paste augmentation (GT-Aug) with our learning based approaches boosts the fully-supervised detector with an increase of $2.8~\mathrm{mAP}$ . Notably, even regarding the fully-supervised detector augmented via our object + +Table 5: Ablate our learning-based cross-scan object transfer which is instructed by HotspotNet. All results are evaluated on nuScenes dataset validation split. + +
(a) Compare our learning-based cross-scan object transfer (instructed by HotspotNet) with other similar methods.
AnnosCSOTObj TransferSelfObj Appearance rαLocal Context (x,y,z)Global Context rsclsrule /learnedmAP NDS
15%——————43.8 56.5
25%✓——————51.0 61.0
35%✓GT-Aug [47]✓✓46.0 57.7
45%✓CA-Aug [15]✓✓✓rule46.0 58.0
55%✓LiDAR-Aug [11]✓✓✓rule46.2 57.9
65%✓Ours✓✓✓✓✓learned50.4 61.0
6a5%✓Ours✓✓✓✓✓rule49.7 60.6
75%✓GT-Aug [47]✓✓49.7 60.5
85%✓CA-Aug [15]✓✓✓rule49.7 60.7
95%✓LiDAR-Aug [11]✓✓✓rule49.7 60.2
105%✓Ours✓✓✓✓✓learned56.4 64.8
10a5%✓Ours✓✓✓✓✓rule53.6 63.3
(b) Ablate contribution of each fitness score defined in Ln-score(o) (Eq. 5).
AnnosCSOTSelfposition sporientation socategory scmAP NDS
115%✓✓49.7 60.5
125%✓✓✓52.1 62.6
135%✓✓✓49.4 60.8
13a5%✓✓✓52.2 62.3
13b5%✓✓✓✓55.1 64.3
145%✓✓✓51.6 62.0
155%✓✓✓✓✓56.4 64.8
+ +transfer implementation as the baseline, reducing annotations by $100\mathrm{x}$ causes only a minimal degradation in performance (remains $97.3\% / 97.4\%$ mAP/mAPH of our fully-supervised detector). + +Ablate our learning-based object transfer. We compare our learning-based object transfer which instructed by the proposed HotspotNet with other similar techniques in Tab. 5(a), and ablate negative sample strategy employed for calculating $\mathcal{L}_{\mathrm{score}}^n (o)$ (Eq. 5) in Tab. 5(b). As shown in Lines 3-10 of Tab. 5, the proposed learning-based object transfer is able to maximize the benefits of CSOT. It also stands alone in its ability to seamlessly integrate with self-training technique without compromising their performance gains. CSOT built upon our learning-based object transfer enhances the detection performance with a considerable $6.6\mathrm{mAP}$ and 4.5 NDS. The improvement surpasses the one built upon previous similar approaches (with a margin of about $4\mathrm{mAP}$ and 3 NDS). The poor partially-supervised detector achieved from other object transfer approaches produce enormous false positive predictions (visualized in Fig. 1) and greatly influence the downstream self-training. Replacing our learning-based object transfer with previous methods reduces the mAP gains of self-training from 6.0 to around 3.5, and degrades the final mAP from 56.4 to 49.7. The necessity of learning-based object transfer are verified in Tab. 5, Lines 6-6a,10-10a. Learn-based approach enables better generalization and can help reduce noise, which introduces an increased final mAP of 2.8. We further ablate negative sample strategy employed for calculating $\mathcal{L}_{\mathrm{score}}^n (o)$ and list results in Lines 11-15 of Tab. 5. Pocation and category provide the most important contextual informa + +Table 6: Ablate robustness of our methods across various network architectures. All results are evaluated on nuScenes dataset validation split. + +
AnnosDetectorEncoderAnchor-baseAnchor-freeVoxel SizeCSOTSelfmAPNDS
1100%CenterPointVFE✓(0.075,0.075,0.2)59.566.0
25%CenterPointVFE✓(0.075,0.075,0.2)43.856.5
35%CenterPointVFE✓(0.075,0.075,0.2)✓51.061.0
45%CenterPointVFE✓(0.075,0.075,0.2)✓50.461.0
55%CenterPointVFE✓(0.075,0.075,0.2)✓✓56.464.8
6100%CenterPointPFE✓(0.1,0.1,8.0)49.260.1
75%CenterPointPFE✓(0.1,0.1,8.0)30.248.3
85%CenterPointPFE✓(0.1,0.1,8.0)✓37.853.1
95%CenterPointPFE✓(0.1,0.1,8.0)✓40.054.9
105%CenterPointPFE✓(0.1,0.1,8.0)✓✓45.057.9
11100%SECONDVFE✓(0.1,0.1,0.2)55.464.7
125%SECONDVFE✓(0.1,0.1,0.2)38.453.2
135%SECONDVFE✓(0.1,0.1,0.2)✓45.258.1
145%SECONDVFE✓(0.1,0.1,0.2)✓45.459.1
155%SECONDVFE✓(0.1,0.1,0.2)✓✓48.661.3
+ +tion, orientation provides additional cues to narrow down the possible identity. Specifically, combining orientation constraint with category consistency improve mAP by 0.6 (Tab. 5, Lines 13a and 14). Removing orientation constraint from all reduces mAP by 1.3 (Tab. 5, Lines 13b and 15). + +Extend our method to different detectors. We confirm the generality of our methods on various detectors, which utilizes diverse point cloud feature encoder or target assigner strategy. All results are listed in Tab. 6. For an anchor-free detector CenterPoint with a pillar-based feature encoder, the proposed CSOT bring a boost of $9.8\mathrm{mAP}$ and 6.6 NDS, which outperforms the self-training technique by $2.2\mathrm{mAP}$ and 1.8 NDS. Using our partially-supervised detector as the teacher model to generate pseudo-labels for unlabeled point clouds, and then conduct one-cycle of self-training. The performance of semi-supervised detector is ultimately reaches $45.0\mathrm{mAP}$ and 57.9 NDS. The mAP gap introduced by $20x$ fewer annotations is narrowed from 19.0 to 4.2. Experiments on anchor-based detector (SECOND) also shows similar results. The effectiveness of CSOT and the additive effect of combining it with self-training techniques are clearly demonstrated in Lines 11-15 of Tab. 6. + +# 5 Conclusion + +In this work, we propose a novel Cross-Scan Object Transfer paradigm for LiDAR semi-supervised object detection. It features a HotspotNet for transferring objects semantic coherence manner, which reduces false positive predictions and enables the proposed CSOT to reach its full potential. To the best of our knowledge, this work is the first successful attempt of object transfer in LiDAR SSOD. The proposed CSOT paradigm is also orthogonal to the self-training techniques. We achieve competitive detectors on par with fully-supervised baselines while using 20x and 100x reduced annotations on nuScenes and Waymo dataset. + +# Acknowledgements + +This work was supported by the InnoHK funding. + +# References + +1. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In: CVPR. pp. 1090-1099 (2022) +2. Blum, A., Mitchell, T.: Combining labeled and unlabeled data with co-training. In: COLT. pp. 92-100 (1998) +3. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR. pp. 11621-11631 (2020) +4. Chen, B., Chen, W., Yang, S., Xuan, Y., Song, J., Xie, D., Pu, S., Song, M., Zhuang, Y.: Label matching semi-supervised object detection. In: CVPR. pp. 14381-14390 (2022) +5. Chen, B., Li, P., Chen, X., Wang, B., Zhang, L., Hua, X.S.: Dense learning based semi-supervised object detection. In: CVPR. pp. 4815-4824 (2022) +6. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Largekernel3d: Scaling up kernels in 3d sparse cnns. In: CVPR. pp. 13488-13498 (2023) +7. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In: CVPR. pp. 21674-21683 (2023) +8. Choi, J., Song, Y., Kwak, N.: Part-aware data augmentation for 3d object detection in point cloud. In: IROS. pp. 3391-3397 (2021) +9. Dwibedi, D., Misra, I., Hebert, M.: Cut, paste and learn: Surprisingly easy synthesis for instance detection. In: ICCV. pp. 1301-1310 (2017) +10. Fan, L., Pang, Z., Zhang, T., Wang, Y.X., Zhao, H., Wang, F., Wang, N., Zhang, Z.: Embracing single stride 3d object detector with sparse transformer. In: CVPR. pp. 8458-8468 (2022) +1. Fang, J., Zuo, X., Zhou, D., Jin, S., Wang, S., Zhang, L.: Lidar-aug: A general rendering-based augmentation framework for 3d object detection. In: CVPR. pp. 4710-4720 (2021) +2. Gao, J., Wang, J., Dai, S., Li, L.J., Nevatia, R.: Note-rcnn: Noise tolerant ensemble rcnn for semi-supervised object detection. In: ICCV. pp. 9508-9517 (2019) +3. Ghiasi, G., Cui, Y., Srinivas, A., Qian, R., Lin, T.Y., Cubuk, E.D., Le, Q.V., Zoph, B.: Simple copy-paste is a strong data augmentation method for instance segmentation. In: CVPR. pp. 2918-2928 (2021) +4. Hu, J.S.K., Waslander, S.L.: Pattern-aware data augmentation for lidar 3d object detection. In: IEEE International Intelligent Transportation Systems Conference (ITSC). pp. 2703-2710 (2021) +5. Hu, X., Duan, Z., Ma, J.: Context-aware data augmentation for lidar 3d object detection. In: ICIP. pp. 11-15 (2023) +6. Hu, Y., Ding, Z., Ge, R., Shao, W., Huang, L., Li, K., Liu, Q.: Afdetv2: Rethinking the necessity of the second stage for object detection from point clouds. In: AAAI. pp. 969-979 (2022) +7. Huang, D., Chen, Y., Ding, Y., Liao, J., Liu, J., Wu, K., Nie, Q., Liu, Y., Wang, C., Li, Z.: Rethinking dimensionality reduction in grid-based 3d object detection. arXiv preprint arXiv:2209.09464 (2022) + +18. Jaderberg, M., Simonyan, K., Vedaldi, A., Zisserman, A.: Synthetic data and artificial neural networks for natural scene text recognition. arXiv preprint arXiv:1406.2227 (2014) +19. Kingma, D.P., Rezende, D.J., Mohamed, S., Welling, M.: Semi-supervised learning with deep generative models. In: Adv. Neural Inform. Process. Syst. (2014) +20. Laine, S., Aila, T.: Temporal ensembling for semi-supervised learning. In: ICLR (2017) +21. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J.: Pointpillars: Fast encoders for object detection from point clouds. In: CVPR. pp. 12697-12705 (2019) +22. Lee, D., Lee, J., Lee, J., Lee, H., Lee, M., Woo, S., Lee, S.: Regularization strategy for point cloud via rigidly mixed sample. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 15900-15909 (2021) +23. Lee, D.H.: Pseudo-label : The simple and efficient semi-supervised learning method for deep neural networks. In: ICML. p. 896 (2013) +24. Lee, S., Lim, H., Myung, H.: Patchwork++: Fast and robust ground segmentation solving partial under-segmentation using 3d point cloud. In: IROS. pp. 13276-13283 (2022) +25. Li, E., Wang, S., Li, C., Li, D., Wu, X., Hao, Q.: Sustech points: A portable 3d point cloud interactive annotation platform system. IEEE Intelligent Vehicles Symposium (IV) pp. 1108-1115 (2020) +26. Li, H., Wu, Z., Shrivastava, A., Davis, L.S.: Rethinking pseudo labels for semi-supervised object detection. In: AAAI. pp. 1314-1322 (2022) +27. Liang, H., Jiang, C., Feng, D., Chen, X., Xu, H., Liang, X., Zhang, W., Li, Z., Gool, L.V.: Exploring geometry-aware contrast and clustering harmonization for self-supervised 3d object detection. In: ICCV. pp. 3293-3302 (2021) +28. Liu, C., Zhang, W., Lin, X., Zhang, W., Tan, X., Han, J., Li, X., Ding, E., Wang, J.: Ambiguity-resistant semi-supervised learning for dense object detection. In: CVPR. pp. 15579-15588 (2023) +29. Liu, C., Gao, C., Liu, F., Li, P., Meng, D., Gao, X.: Hierarchical supervision and shuffle data augmentation for 3d semi-supervised object detection. In: CVPR. pp. 23819-23828 (2023) +30. Liu, C., Gao, C., Liu, F., Liu, J., Meng, D., Gao, X.: Ss3d: Sparsely-supervised 3d object detection from point cloud. In: CVPR. pp. 8428-8437 (2022) +31. Liu, Y.C., Ma, C.Y., He, Z., Kuo, C.W., Chen, K., Zhang, P., Wu, B., Kira, Z., Vajda, P.: Unbiased teacher for semi-supervised object detection. In: ICLR (2021) +32. Liu, Y.C., Ma, C.Y., Kira, Z.: Unbiased teacher v2: Semi-supervised object detection for anchor-free and anchor-based detectors. In: CVPR. pp. 9819-9828 (2022) +33. Meng, Q., Wang, W., Zhou, T., Shen, J., Jia, Y., Gool, L.V.: Towards a weakly supervised framework for 3d point cloud object detection and annotation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 44(8), 4454-4468 (2021) +34. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR. pp. 652-660 (2017) +35. Qian, G., Li, Y., Peng, H., Mai, J., Hammoud, H.A.A.K., Elhoseiny, M., Ghanem, B.: Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In: Adv. Neural Inform. Process. Syst. pp. 23192-23204 (2022) +36. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In: CVPR. pp. 10529–10538 (2020) + +37. Sohn, K., Zhang, Z., Li, C.L., Zhang, H., Lee, C.Y., Pfister, T.: A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757 (2020) +38. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhao, S., Cheng, S., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: CVPR. pp. 2446-2454 (2020) +39. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In: Adv. Neural Inform. Process. Syst. (2017) +40. Wang, H., Cong, Y., Litany, O., Gao, Y., Guibas, L.J.: 3dioumatch: Leveraging iou prediction for semi-supervised 3d object detection. In: CVPR. pp. 14615-14624 (2021) +41. Wang, T., Yang, T., Cao, J., Zhang, X.: Co-mining: Self-supervised learning for sparsely annotated object detection. In: AAAI. pp. 2800-2808 (2021) +42. Wang, X., Yang, X., Zhang, S., Li, Y., Feng, L., Fang, S., Lyu, C., Chen, K., Zhang, W.: Consistent-teacher: Towards reducing inconsistent pseudo-targets in semi-supervised object detection. In: CVPR. pp. 3240-3249 (2023) +43. Wang, Y., Yin, J., Li, W., Frossard, P., Yang, R., Shen, J.: Ssda3d: Semi-supervised domain adaptation for 3d object detection from point cloud. In: AAAI. pp. 2707-2715 (2023) +44. Wu, Z., Bodla, N., Singh, B., Najibi, M., Chellappa, R., Davis, L.S.: Soft sampling for robust object detection. arXiv preprint arXiv:1806.06986 (2018) +45. Xu, M., Zhang, Z., Hu, H., Wang, J., Wang, L., Wei, F., Bai, X., Liu, Z.: End-to-end semi-supervised object detection with soft teacher. In: ICCV. pp. 3060-3069 (2021) +46. Xu, R., Wang, T., Zhang, W., Chen, R., Cao, J., Pang, J., Lin, D.: Mv-jar: Masked voxel jigsaw and reconstruction for lidar-based self-supervised pre-training. In: CVPR. pp. 13445-13454 (2023) +47. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018) +48. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d: Self-training for unsupervised domain adaptation on 3d object detection. In: CVPR. pp. 10368-10378 (2021) +49. Yang, Q., Wei, X., Wang, B., Hua, X.S., Zhang, L.: Interactive self-training with mean teachers for semi-supervised object detection. In: CVPR. pp. 5941-5950 (2021) +50. Yang, Y., Fan, L., Zhang, Z.: Mixsup: Mixed-grained supervision for labeled efficient lidar-based 3d object detection. In: ICLR (2024) +51. Ye, D., Zhou, Z., Chen, W., Xie, Y., Wang, Y., Wang, P., Foroosh, H.: Lidar-multinet: Towards a unified multi-task network for lidar perception. In: AAAI. pp. 3231-3240 (2023) +52. Yin, J., Fang, J., Zhou, D., Zhang, L., Xu, C.Z., Shen, J., Wang, W.: Semi-supervised 3d object detection with proficient teachers. In: ECCV. pp. 727-743 (2022) +53. Yin, J., Zhou, D., Zhang, L., Fang, J., Xu, C.Z., Shen, J., Wang, W.: Proposal-contrast: Unsupervised pre-training for lidar-based 3d object detection. In: ECCV. pp. 17-33 (2022) +54. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: CVPR. pp. 11784-11793 (2021) + +55. You, Y., Luo, K., Phoo, C.P., Chao, W.L., Sun, W., Hariharan, B., Campbell, M., Weinberger, K.Q.: Learning to detect mobile objects from lidar scans without labels. In: CVPR. pp. 1130-1140 (2022) +56. Yuan, J., Zhang, B., Yan, X., Chen, T., Shi, B., Li, Y., Qiao, Y.: Ad-pt: Autonomous driving pre-training with large-scale point cloud dataset. In: Adv. Neural Inform. Process. Syst. p. 36 (2023) +57. Yun, S., Han, D., Oh, S.J., Chun, S., Cho, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: CVPR. pp. 6023-6032 (2019) +58. Yurtsever, E., Erçelik, E., Liu, M., Yang, Z., Zhang, H., Topçam, P., Listl, M., Çayli, Y.K., Knoll, A.: 3d object detection with a self-supervised lidar scene flow backbone. In: ECCV. pp. 247-265 (2022) +59. Zhang, H., Chen, F., Shen, Z., Hao, Q., Zhu, C., Savvides, M.: Solving missing-annotation object detection with background recalibration loss. In: ICASSP. pp. 1888-1892 (2020) +60. Zhang, L., Sun, Y., Wei, W.: Mind the gap: Polishing pseudo labels for accurate semi-supervised object detection. In: AAAI. pp. 3463-3471 (2023) +61. Zhang, L., Yang, A.J., Xiong, Y., Casas, S., Yang, B., Ren, M., Urtasun, R.: Towards unsupervised object detection from lidar point clouds. In: CVPR. pp. 9317-9328 (2023) +62. Zhao, N., Chua, T.S., Lee, G.H.: Sess: Self-ensembling semi-supervised 3d object detection. In: CVPR. pp. 11079-11087 (2020) +63. Zheng, W., Tang, W., Jiang, L., Fu, C.W.: Se-ssd: Self-ensembling single-stage object detector from point cloud. In: CVPR. pp. 14494–14503 (2021) +64. Zhou, H., Ge, Z., Liu, S., Mao, W., Li, Z., Yu, H., Sun, J.: Dense teacher: Dense pseudo-labels for semi-supervised object detection. In: ECCV. pp. 35-50 (2022) +65. Zhou, Q., Yu, C., Wang, Z., Qian, Q., Li, H.: Instant-teaching: An end-to-end semi-supervised object detection framework. In: CVPR. pp. 4081-4090 (2021) +66. Zhu, B., Jiang, Z., Zhou, X., Li, Z., Yu, G.: Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492 (2019) \ No newline at end of file diff --git a/2024/CSOT_ Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection/images.zip b/2024/CSOT_ Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..25efa0f8be478a5cb18c02f5af60b7ce4ff37d08 --- /dev/null +++ b/2024/CSOT_ Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f5c2abe3e10bf5f737782d0304c611bf124b6e836f087bf4310e436a934a8f8 +size 574718 diff --git a/2024/CSOT_ Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection/layout.json b/2024/CSOT_ Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e7bce4940025db7d16637cc8d8bcee1567ba1b2f --- /dev/null +++ b/2024/CSOT_ Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection/layout.json @@ -0,0 +1,8801 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 112, + 455, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 112, + 455, + 148 + ], + "spans": [ + { + "bbox": [ + 159, + 112, + 455, + 148 + ], + "type": "text", + "content": "CSOT: Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 163, + 167, + 451, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 163, + 167, + 451, + 192 + ], + "spans": [ + { + "bbox": [ + 163, + 167, + 451, + 192 + ], + "type": "text", + "content": "Jinglin Zhan" + }, + { + "bbox": [ + 163, + 167, + 451, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 163, + 167, + 451, + 192 + ], + "type": "text", + "content": ", Tiejun Liu" + }, + { + "bbox": [ + 163, + 167, + 451, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 163, + 167, + 451, + 192 + ], + "type": "text", + "content": ", Rengang Li" + }, + { + "bbox": [ + 163, + 167, + 451, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 163, + 167, + 451, + 192 + ], + "type": "text", + "content": ", Zhaoxiang Zhang" + }, + { + "bbox": [ + 163, + 167, + 451, + 192 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 163, + 167, + 451, + 192 + ], + "type": "text", + "content": ", and Yuntao Chen" + }, + { + "bbox": [ + 163, + 167, + 451, + 192 + ], + "type": "inline_equation", + "content": "^{3\\star}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 241, + 201, + 341, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 201, + 341, + 213 + ], + "spans": [ + { + "bbox": [ + 241, + 201, + 341, + 213 + ], + "type": "text", + "content": "1 IEIT Systems" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 241, + 213, + 373, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 213, + 373, + 223 + ], + "spans": [ + { + "bbox": [ + 241, + 213, + 373, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 241, + 213, + 373, + 223 + ], + "type": "text", + "content": " Institute of Automation, CAS" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 181, + 223, + 432, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 223, + 432, + 258 + ], + "spans": [ + { + "bbox": [ + 181, + 223, + 432, + 258 + ], + "type": "text", + "content": "3 Centre for Artificial Intelligence and Robotics, HKISI, CAS {zhanjinglin, liutj, lirg}@ieisystem.com, zhaoxiang.zhang@ia.ac.cn, chenyuntao08@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 284, + 455, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 284, + 455, + 547 + ], + "spans": [ + { + "bbox": [ + 160, + 284, + 455, + 547 + ], + "type": "text", + "content": "Abstract. Large-scale 3D bounding box annotation is crucial for LiDAR object detection but comes at a high cost. Semi-supervised object detection (SSOD) offers promising solutions to leverage unannotated data, but the predominant pseudo-labeling approach requires careful hyperparameter tuning for training on noisy teacher labels. In this work, we propose a Cross-Scan Object Transfer (CSOT) paradigm for LiDAR SSOD. Central to our approach is HotspotNet, a transformer-based network that predicts possible placement locations and the object-place fitness scores for inserting annotated objects into unlabeled scans in a semantic coherence manner. Based on HotspotNet, CSOT successfully enables object copy-paste in LiDAR SSOD for the first time. To train object detectors on partially annotated scans generated by CSOT, we adopt a spatial-aware classification loss throughout our partial supervision to handle false negative issues caused by treating all unlabeled objects as background. We conduct extensive experiments to verify the efficacy and generality of our method. Compared to other state-of-the-art label-efficient methods used in LiDAR detection, our approach requires the least amount of annotation while achieves the best detector. Using only " + }, + { + "bbox": [ + 160, + 284, + 455, + 547 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 160, + 284, + 455, + 547 + ], + "type": "text", + "content": " of the labeled data on the Waymo dataset, our semi-supervised detector achieves performance on par with the fully supervised baseline. Similarly, on the nuScenes dataset, our semi-supervised CenterPoint reaches " + }, + { + "bbox": [ + 160, + 284, + 455, + 547 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 160, + 284, + 455, + 547 + ], + "type": "text", + "content": " of the fully supervised model's detection performance in terms of NDS score, while using just " + }, + { + "bbox": [ + 160, + 284, + 455, + 547 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 160, + 284, + 455, + 547 + ], + "type": "text", + "content": " of the labeled data. Code is released at https://github.com/JinglinZhan/CSOT" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 557, + 453, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 557, + 453, + 580 + ], + "spans": [ + { + "bbox": [ + 160, + 557, + 453, + 580 + ], + "type": "text", + "content": "Keywords: Semi-Supervised Learning " + }, + { + "bbox": [ + 160, + 557, + 453, + 580 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 557, + 453, + 580 + ], + "type": "text", + "content": " LiDAR Object Detection " + }, + { + "bbox": [ + 160, + 557, + 453, + 580 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 557, + 453, + 580 + ], + "type": "text", + "content": " Autonomous Driving" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 599, + 230, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 230, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 230, + 611 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 623, + 481, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 623, + 481, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 623, + 481, + 647 + ], + "type": "text", + "content": "Large-scale annotated datasets have catalyzed the success of LiDAR object detection, but at a significant cost in terms of money and time [25,33]. For instance," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 654, + 236, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 236, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 236, + 665 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 465, + 285 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 465, + 285 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 465, + 285 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 465, + 285 + ], + "type": "image", + "image_path": "a0cfdbec5a5b1e466502e6b36e1b8f1e6949a579beae5cfb01e3f3b675310284.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 294, + 482, + 340 + ], + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 340 + ], + "type": "text", + "content": "Fig. 1: Preliminary experiments of CSOT pipeline for LiDAR SSOD. (a) Compare CSOT implemented based on different cross-scan object transfer methods and display the NDS differences. (b) Visualize prediction differences among various teacher model that used to generate pseudo labels for self-training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 364, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 449 + ], + "type": "text", + "content": "annotating the 12 million 3D boxes in the Waymo Open [38] dataset may have consumed millions of dollars and tens of thousands of human hours. In contrast, autonomous driving fleets capture thousands, and even millions, of hours of unannotated sensor data every day. The huge gap between the speed of manual labeling and that of data accumulation makes labeling a major slowdown. This calls for a more scalable approach to leverage the abundant data collected by autonomous fleets and continually advance object detection capabilities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 628 + ], + "type": "text", + "content": "Semi-Supervised Learning (SSL) presents a promising way for reducing annotation demands by leveraging both limited labeled data and vast quantities of unlabeled data during training. Various SSL methods have been proposed, including pseudo-labeling (a.k.a. self-training) [23], co-training [2], consistency regularization [20] and generative modeling [19]. Among these SSL approaches, pseudo-labeling has gained significant traction, particularly in the field of semi-supervised object detection (SSOD) [4, 12, 31, 37, 39, 40, 42, 52, 62, 65]. However, generating reliable pseudo-labels without human intervention remains a significant challenge. Noisy pseudo-labels can negatively impact model performance, leading to suboptimal results. To address this, researchers have proposed various techniques, including sophisticated post-processing mechanisms [4, 5, 26, 28, 40, 52, 60], iterative training procedures [39, 45, 49], and noise-tolerant detectors [12, 65] and loss objectives [31, 32, 62, 64]. Nevertheless, these approaches often involve complex pipelines and require careful parameter tuning, making them inefficient and difficult to adapt to new datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "In contrast with noisy pseudo-labeling, we leverage a novel semi-supervised learning approach for 3D object detection that transfers annotated objects to generate noise-free training data. Specifically, we transfer point clouds and bound" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "J. Zhan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "ing boxes of annotated objects to unannotated LiDAR scans to create partially labeled data. Object transfer via copy-paste is a commonly employed data augmentation [6,7,16,47,51] in LiDAR object detection. But naively applying object copy-paste [11, 15, 47] to semi-supervised LiDAR object detection yields suboptimal results. As shown in Fig. 1, SSOD built upon vanilla object transfer does improve performance over the baseline, but combining it with self-training results in a negative net effect on detection performance compared to using self-training alone. The key insight is that object copy-paste without careful contextual consistency maintenance improves detector performance at the cost of generating a large number of false positives. Although the commonly used object detection metric mAP is not sensitive to false positives if the recall improves, the self-training pipeline accumulates these false positives over iterations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 261, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 534 + ], + "type": "text", + "content": "In order to achieve coherent object copy-paste, we propose Hotspot Network, a learning-based approach to insert objects with strong contextual consistency, enabling object-transfer-based semi-supervised LiDAR object detection to achieve its full potential. HotspotNet takes unannotated LiDAR scans as inputs, along with point clouds and bounding boxes of annotated objects. It then predicts possible placement locations for these objects and assigns scores to each location, indicating the likelihood of the object fitting well within the context of the scanned environment. This allows us introducing noise-free supervision signal into unlabeled point clouds and use these newly created partially-labeled data to optimize detector performance. Based on HotspotNet, we design a novel Cross-Scan Object Transfer(CSOT) pipeline for LiDAR SSOD. CSOT enjoys precise partial annotations that maintain great contextual coherence with unannotated scans rather than noisy pseudo labels for training detectors on unlabeled data. It's important to note that training detectors on partially annotated point clouds, as opposed to fully labeled ones, can lead to a significant issue: unlabeled objects being miscategorized as background, resulting in a high number of false negatives. To mitigate this problem, we propose a spatial-aware classification loss and employ it throughout our partial supervision. Our CSOT paradigm features a streamlined training pipeline and requires minimal hyper-parameter tuning, enabling it to quickly generalize across various datasets and detectors. Moreover, this approach is orthogonal and complementary to commonly used self-training techniques. We conducted extensive experiments on the Waymo [38] and nuScenes [3] datasets to confirm the effectiveness of our methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 536, + 339, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 536, + 339, + 547 + ], + "spans": [ + { + "bbox": [ + 146, + 536, + 339, + 547 + ], + "type": "text", + "content": "Our contributions are summarized below:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 557, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 557, + 480, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 557, + 480, + 591 + ], + "spans": [ + { + "bbox": [ + 138, + 557, + 480, + 591 + ], + "type": "text", + "content": "- We propose a Hotspot Network for instructing object placement in a semantic coherence manner, and employ an asymmetric supervision for training HotspotNet on both annotated and unannotated data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 593, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 593, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 138, + 593, + 482, + 628 + ], + "type": "text", + "content": "- Based on HotspotNet, we design a novel Cross-Scan Object Transfer paradigm for LiDAR semi-supervised object detection. To the best of our knowledge, this is the first successful attempt of object copy-paste in LiDAR SSOD." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 630, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 630, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 630, + 480, + 665 + ], + "type": "text", + "content": "- We conduct extensive experiments on nuScenes and Waymo datasets to verify the efficacy and generality of our method. Our approach achieves competitive performance compared to fully-supervised baselines while using 20x" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "text", + "content": "Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 148, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 148, + 116, + 480, + 140 + ], + "type": "text", + "content": "to 100x fewer annotations, significantly outperforming other state-of-the-art label-efficient approaches in LiDAR object detection." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 155, + 238, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 155, + 238, + 168 + ], + "spans": [ + { + "bbox": [ + 132, + 155, + 238, + 168 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 176, + 334, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 176, + 334, + 189 + ], + "spans": [ + { + "bbox": [ + 132, + 176, + 334, + 189 + ], + "type": "text", + "content": "2.1 Semi-Supervised Object Detection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 192, + 482, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 192, + 482, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 192, + 482, + 432 + ], + "type": "text", + "content": "Unsupervised learning eliminates data annotation but significantly degrades model performance [48, 55, 61]. Semi-supervised methods maintain low annotation costs meanwhile greatly bridges the gap to fully-supervised approaches [43]. As a prior work, STAC [37] proposes a pseudo labeling SSOD paradigm and catalyzes extensive follow-on researches. It trains a teacher with all annotated data and generates pseudo labels for supervising the student on unlabeled data. Producing reliable pseudo labels is critical in this paradigm and numerous methods have been explored. 3D IoUMatch [40] utilizes the estimated 3D IoU to filter poorly localized predictions. LabelMatch [4] leverages adaptive label-distribution-aware confidence thresholds for generating unbiased pseudo labels. Proficient Teacher [52] employs clustering-based box voting for improving the precision of final predictions. HSSDA [29] designs a dynamic dual-threshold strategy to generate more reasonable supervision for training the student model. Besides various filter mechanism, many other approaches, such as iterative training procedures [39, 62], noise-tolerant detectors [12, 65] and a variety of loss objectives with consistency or reweighting terms [31, 42, 62], are also proposed for alleviating negative effects caused by noisy pseudo labels. Despite some progress has been made, unreliable pseudo-labels still remain an open challenge. The sophisticated pipelines with extensive manual threshold tuning also impede these approaches' quick adaptation to new datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 445, + 369, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 445, + 369, + 458 + ], + "spans": [ + { + "bbox": [ + 132, + 445, + 369, + 458 + ], + "type": "text", + "content": "2.2 Copy-Paste For LiDAR Object Detection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": "Copy-paste method is developed by Jaderberg et.al. for text recognition [18], then extended by Dwibedi et.al. [9] and Yun et.al. [57] for 2D object detection. Yan et.al. [47] introduces copy-paste into LiDAR object detection and develops the well known ground-truth augmentation approach (so called GT-Aug), which is extensively adopted in many later studies [1,6,16,17,21,51,54,66]. Since then, numerous approaches have been proposed to optimize GT-Aug. Part-Aware [8] and Shape-Aware [63] GT-Aug divide objects into partitions and apply different augmentation methods to each partition. Pattern-Aware [14] GT-Aug downsamples the points of objects to create a new one with farther distance. PointCut-Mix [22] replaces part of the sample with shape-preserved subsets from another one. Some researchers notice placement issues in GT-Aug and develop hand-craft valid maps to restrict the insertion of objects [11,15]. Up to now, researches on copy-paste is primarily focused on exploring and developing its capabilities for augmenting data to support fully-supervised learning tasks. Ghiasi et.al. introduce copy-paste into SSL [13] for augmenting both pseudo labeled and supervised labeled images. The above trial still limited in the field of data augmentation and the copy-paste technique remains largely untapped." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "J. Zhan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 467, + 273 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 467, + 273 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 467, + 273 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 467, + 273 + ], + "type": "image", + "image_path": "57f1dfe65a9d740e6d0734a6d3dcc14625ba67651a748d8abde37c147147fbf8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 280, + 482, + 326 + ], + "lines": [ + { + "bbox": [ + 130, + 280, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 482, + 326 + ], + "type": "text", + "content": "Fig. 2: The overall architecture of HotspotNet. The Object Encoders extract features from object point cloud " + }, + { + "bbox": [ + 130, + 280, + 482, + 326 + ], + "type": "inline_equation", + "content": "X_{obj,pc}" + }, + { + "bbox": [ + 130, + 280, + 482, + 326 + ], + "type": "text", + "content": " and bounding box meta data " + }, + { + "bbox": [ + 130, + 280, + 482, + 326 + ], + "type": "inline_equation", + "content": "X_{obj,meta}" + }, + { + "bbox": [ + 130, + 280, + 482, + 326 + ], + "type": "text", + "content": ". The scene encoder featurize the scene point cloud and cross-attends to object features " + }, + { + "bbox": [ + 130, + 280, + 482, + 326 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{obj,pc}}" + }, + { + "bbox": [ + 130, + 280, + 482, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 280, + 482, + 326 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{obj,meta}}" + }, + { + "bbox": [ + 130, + 280, + 482, + 326 + ], + "type": "text", + "content": ". The location decoder predicts the placeable locations and their scores." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 345, + 202, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 345, + 202, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 345, + 202, + 357 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 368, + 482, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 417 + ], + "type": "text", + "content": "We provide details of Cross-Scan Object Transfer SSOD paradigm in this section. We present the overall architecture of HotspotNet in Sec. 3.1, and describe the Asymmetric Supervision for training HotspotNet in Sec. 3.2. The pipeline of our newly proposed SSOD paradigm, CSOT, is elaborated in Sec. 3.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 432, + 248, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 432, + 248, + 444 + ], + "spans": [ + { + "bbox": [ + 132, + 432, + 248, + 444 + ], + "type": "text", + "content": "3.1 Hotspot Network" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 450, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 481, + 498 + ], + "type": "text", + "content": "The key to LiDAR copy-paste is identifying semantically coherent locations within background scans to paste objects. Placing objects at unreasonable locations, such as inside walls, floating in sky or with weird pose, will cause a large amount of false positives and thus deteriorate the detector performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 481, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 546 + ], + "type": "text", + "content": "We propose a learning-based approach for object placement that provides sufficient supervision signal and maintain a high signal-to-noise ratio. Specifically, we employ a Hotspot Network to predict the placeable location " + }, + { + "bbox": [ + 130, + 498, + 481, + 546 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 130, + 498, + 481, + 546 + ], + "type": "text", + "content": " and the object-place fitness scores " + }, + { + "bbox": [ + 130, + 498, + 481, + 546 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 130, + 498, + 481, + 546 + ], + "type": "text", + "content": " for inserting object in a semantic coherence manner." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 546, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 481, + 616 + ], + "type": "text", + "content": "We abbreviate Hotspot Network as HotspotNet and visualize the overall architecture in Fig. 2. The HotspotNet comprises object encoders for object point cloud " + }, + { + "bbox": [ + 130, + 546, + 481, + 616 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{obj,pc}}" + }, + { + "bbox": [ + 130, + 546, + 481, + 616 + ], + "type": "text", + "content": " and meta data " + }, + { + "bbox": [ + 130, + 546, + 481, + 616 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{obj,meta}}" + }, + { + "bbox": [ + 130, + 546, + 481, + 616 + ], + "type": "text", + "content": ", respectively. The scene encoder featurizes the scene point cloud and cross-attends to object features " + }, + { + "bbox": [ + 130, + 546, + 481, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{obj,pc}}" + }, + { + "bbox": [ + 130, + 546, + 481, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 546, + 481, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{obj,meta}}" + }, + { + "bbox": [ + 130, + 546, + 481, + 616 + ], + "type": "text", + "content": ". A location decoder predicts the placeable location " + }, + { + "bbox": [ + 130, + 546, + 481, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{b} = \\{x,y,z,w,l,h,\\theta ,c\\}" + }, + { + "bbox": [ + 130, + 546, + 481, + 616 + ], + "type": "text", + "content": " and the location fitness score " + }, + { + "bbox": [ + 130, + 546, + 481, + 616 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 130, + 546, + 481, + 616 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "content": "Object Encoder We process the object point cloud " + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{obj,pc}}" + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "content": " with object point cloud encoders like PointNet [34] or PointNeXt [35] and obtain object features " + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{obj,pc}}" + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "content": ". We encode the bounding box meta " + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{obj,meta}}" + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "content": " with appropriate meta feature encoders like Fourier embedding for boxes, and CLIP or one-hot embedding" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 90, + 447, + 102 + ], + "type": "text", + "content": "Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 141 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 141 + ], + "type": "text", + "content": "for object classes. We then transform the initial box embedding with a MLP to obtain bounding box meta feature " + }, + { + "bbox": [ + 130, + 116, + 479, + 141 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{obj,meta}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 141 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 170, + 152, + 480, + 164 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 152, + 480, + 164 + ], + "spans": [ + { + "bbox": [ + 170, + 152, + 480, + 164 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\mathrm {o b j}, \\mathrm {p c}} = 0 \\text {b j e c t P o i n t C l o u d E n c o d e r} \\left(X _ {\\mathrm {o b j}, \\mathrm {p c}}\\right) \\tag {1}", + "image_path": "db3e94940175bfd688dba2fd0da06d636da67a4cec5f713147f2e27851cda641.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 167, + 480, + 179 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 167, + 480, + 179 + ], + "spans": [ + { + "bbox": [ + 160, + 167, + 480, + 179 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\text {o b j , m e t a}} = \\operatorname {M L P} \\left(\\text {F o u r i e r E m b e d} \\left(X _ {\\text {o b j , b o x}}\\right) \\mid \\mid \\text {C l a s s E m b e d} \\left(X _ {\\text {o b j , c l a s s}}\\right)\\right) \\tag {2}", + "image_path": "a25ae44701f45bc5caed0e9b15f8ee3759c38ff804fed0839f94a9602032bdc0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 189, + 479, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 479, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 479, + 260 + ], + "type": "text", + "content": "Scene Encoder Given a scene LiDAR scan " + }, + { + "bbox": [ + 130, + 189, + 479, + 260 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{scene}}" + }, + { + "bbox": [ + 130, + 189, + 479, + 260 + ], + "type": "text", + "content": ", we leverage off-the-shelf point cloud encoders like CenterPoint [54] or SST [10] for extracting features " + }, + { + "bbox": [ + 130, + 189, + 479, + 260 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{scene}}" + }, + { + "bbox": [ + 130, + 189, + 479, + 260 + ], + "type": "text", + "content": ". We use spatial cross-attention in the first convolution block of each stage of the scene point cloud encoder. This allows the scene encoder aggregating features from both the object point clouds and the box metadata for assessing the object fitness of each spatial location." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 262, + 479, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 479, + 309 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 479, + 309 + ], + "type": "text", + "content": "Location Decoder We randomly initialize a set of learnable queries " + }, + { + "bbox": [ + 130, + 262, + 479, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_0" + }, + { + "bbox": [ + 130, + 262, + 479, + 309 + ], + "type": "text", + "content": ". These queries are combine with position embeddings to obtain location queries " + }, + { + "bbox": [ + 130, + 262, + 479, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 130, + 262, + 479, + 309 + ], + "type": "text", + "content": ". Our location decoder leverages both scene features and object features for predicting possible inserting locations and scores." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 330, + 419, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 330, + 419, + 342 + ], + "spans": [ + { + "bbox": [ + 131, + 330, + 419, + 342 + ], + "type": "text", + "content": "3.2 Training HotspotNet with Asymmetric Supervision" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 353, + 479, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 479, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 479, + 436 + ], + "type": "text", + "content": "The HotspotNet consists of a location predictor and a location scorer that work jointly to suggest object placements. The location predictor outputs candidate object positions, while the scorer assigns fitness scores to evaluate how semantically consistent each candidate is. We employ a multi-task learning approach that features Asymmetric Supervision for optimizing different predictions. The proposed losses supervise the location predictor and scorer to accurately position objects within a new environment." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 437, + 480, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 480, + 557 + ], + "type": "text", + "content": "To train the location predictor, we utilize the annotated LiDAR scans. For each annotated scan, we randomly extract an object point cloud from the original point clouds along with its corresponding bounding box. The regressor is then supervised with an " + }, + { + "bbox": [ + 130, + 437, + 480, + 557 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 130, + 437, + 480, + 557 + ], + "type": "text", + "content": " loss between the predicted location " + }, + { + "bbox": [ + 130, + 437, + 480, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 130, + 437, + 480, + 557 + ], + "type": "text", + "content": " and the extracted bounding box " + }, + { + "bbox": [ + 130, + 437, + 480, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_{\\mathrm{cut - out}}" + }, + { + "bbox": [ + 130, + 437, + 480, + 557 + ], + "type": "text", + "content": ". We constrain object distance and perspective from the ego car in Eq. 3 to preserve perspective and point density. We employ dropout regularization on the bounding box meta data to prevent the location predictor from directly copying the meta data to the prediction output. It encourages the predictor to learn meaningful features from the point clouds itself, rather than relying solely on the provided meta data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 252, + 569, + 480, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 569, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 252, + 569, + 480, + 581 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {l o c}} (\\mathbf {b}) = \\left\\| \\mathbf {b} - \\mathbf {b} _ {\\text {c u t - o u t}} \\right\\| _ {1} \\tag {3}", + "image_path": "54350d984e9877b4e9dc24f66d8c2c319df03813df1969f7786053f7feea0918.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 592, + 479, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 592, + 479, + 651 + ], + "spans": [ + { + "bbox": [ + 130, + 592, + 479, + 651 + ], + "type": "text", + "content": "To train the location scorer, we utilize an asymmetric supervision approach for positive and negative samples. Positive samples are obtained by matching location queries with the extracted object point clouds. These positive samples encourage the scorer to assign high scores to queries that accurately position objects within the scan." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 267, + 654, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 654, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 267, + 654, + 480, + 666 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s c o r e}} ^ {p} (o) = - \\log o \\tag {4}", + "image_path": "6620029840aba26e04f92ae8b601fe8921a9944567a5325f49def887ef63ecb7.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "J. Zhan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "To avoid treating all queries that do not match the extracted objects as negative samples, a careful selection process is employed. Instead of using all non-matching queries, negative samples are chosen based on queries that have low fitness scores, denoted by " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ". This approach ensures that the scorer learns to differentiate between suitable and unsuitable object placements, focusing on the most informative negative samples." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 207, + 198, + 481, + 212 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 198, + 481, + 212 + ], + "spans": [ + { + "bbox": [ + 207, + 198, + 481, + 212 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s c o r e}} ^ {n} (o) = - \\left(1 - \\mathbb {1} \\left[ s _ {p} \\cdot s _ {o} \\cdot s _ {c} > \\tau \\right]\\right) \\log (1 - o) \\tag {5}", + "image_path": "fbd10e774059df6d262befa47b9d218d4ba625f48f1dd3c03a0b029d3004bd62.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 220, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 482, + 411 + ], + "type": "text", + "content": "The fitness scores " + }, + { + "bbox": [ + 130, + 220, + 482, + 411 + ], + "type": "inline_equation", + "content": "s_p" + }, + { + "bbox": [ + 130, + 220, + 482, + 411 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 220, + 482, + 411 + ], + "type": "inline_equation", + "content": "s_o" + }, + { + "bbox": [ + 130, + 220, + 482, + 411 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 220, + 482, + 411 + ], + "type": "inline_equation", + "content": "s_c" + }, + { + "bbox": [ + 130, + 220, + 482, + 411 + ], + "type": "text", + "content": " are used to measure the accuracy and coherence of an object's position, orientation, and category within the scene. The position score " + }, + { + "bbox": [ + 130, + 220, + 482, + 411 + ], + "type": "inline_equation", + "content": "s_p = \\exp(-(\\mathbf{b}_{xyz} \\cdot \\mathbf{n}_g + d_g) / \\sigma_p^2)^1" + }, + { + "bbox": [ + 130, + 220, + 482, + 411 + ], + "type": "text", + "content": " penalizes the predicted location " + }, + { + "bbox": [ + 130, + 220, + 482, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_{xyz}" + }, + { + "bbox": [ + 130, + 220, + 482, + 411 + ], + "type": "text", + "content": " deviating from the local ground plane specified by " + }, + { + "bbox": [ + 130, + 220, + 482, + 411 + ], + "type": "inline_equation", + "content": "(\\mathbf{n}_g, d_g)" + }, + { + "bbox": [ + 130, + 220, + 482, + 411 + ], + "type": "text", + "content": ". Local ground planes here could be obtained via non-learning ground detection methods like Patch-Works++ [24]. Similarly, the orientation score " + }, + { + "bbox": [ + 130, + 220, + 482, + 411 + ], + "type": "inline_equation", + "content": "s_o = \\cos((\\mathbf{b}_\\theta - \\hat{\\mathbf{b}}_\\theta) / 2)" + }, + { + "bbox": [ + 130, + 220, + 482, + 411 + ], + "type": "text", + "content": " encourages consistency between the heading of the new object location and the predominant heading " + }, + { + "bbox": [ + 130, + 220, + 482, + 411 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{b}}_\\theta" + }, + { + "bbox": [ + 130, + 220, + 482, + 411 + ], + "type": "text", + "content": " of objects in the same category within the scene. Finally, the category compatible score " + }, + { + "bbox": [ + 130, + 220, + 482, + 411 + ], + "type": "inline_equation", + "content": "s_c = \\mathbb{1}[\\mathbf{b}_c \\in \\mathcal{C}]" + }, + { + "bbox": [ + 130, + 220, + 482, + 411 + ], + "type": "text", + "content": " prevents pasting object of category that is not in the current scene (e.g. bicycle on highway). Notably, these scores can be computed not only between the extracted object and its original scan but also between the extracted object and other unannotated LiDAR scans, in which " + }, + { + "bbox": [ + 130, + 220, + 482, + 411 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{b}}_\\theta" + }, + { + "bbox": [ + 130, + 220, + 482, + 411 + ], + "type": "text", + "content": " can be simplified as ego orientation and " + }, + { + "bbox": [ + 130, + 220, + 482, + 411 + ], + "type": "inline_equation", + "content": "s_c" + }, + { + "bbox": [ + 130, + 220, + 482, + 411 + ], + "type": "text", + "content": " can be ignored directly. This adaptability opens up the possibility of learning object placement scoring from a more extensive dataset through unsupervised learning, which assigns positive and negative samples solely based on the fitness scores." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 411, + 482, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 448 + ], + "type": "text", + "content": "Combining these asymmetric losses and carefully selecting positive and negative samples, our HotspotNet learns to accurately position objects within a new scan while maintaining semantic coherence." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 465, + 422, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 465, + 422, + 478 + ], + "spans": [ + { + "bbox": [ + 131, + 465, + 422, + 478 + ], + "type": "text", + "content": "3.3 Cross-Scan Object Transfer for LiDAR-based SSOD" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 485, + 482, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 558 + ], + "type": "text", + "content": "Under the guidance of HotspotNet (described in Sec. 3.1), we inject precise supervision signals into unannotated point clouds and train detectors on these partially-labeled data. Fig. 3 illustrates the differences between our Cross-Scan Object Transfer SSOD paradigm and the commonly used self-training technique. Our method bypasses the iterative self-training procedures, cumbersome pseudolabel filtering mechanisms, and complex loss objective designs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 574, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 482, + 648 + ], + "type": "text", + "content": "Partially Supervised Object Detection Our approach for introducing supervision into unlabeled LiDAR scans is straightforward. We denote all objects in labeled point clouds as " + }, + { + "bbox": [ + 130, + 574, + 482, + 648 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{O}^L\\}" + }, + { + "bbox": [ + 130, + 574, + 482, + 648 + ], + "type": "text", + "content": ", and all unlabeled scans as " + }, + { + "bbox": [ + 130, + 574, + 482, + 648 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{S}^U\\}" + }, + { + "bbox": [ + 130, + 574, + 482, + 648 + ], + "type": "text", + "content": ". For an unlabeled LiDAR scan " + }, + { + "bbox": [ + 130, + 574, + 482, + 648 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^U" + }, + { + "bbox": [ + 130, + 574, + 482, + 648 + ], + "type": "text", + "content": ", we randomly sample a subset of " + }, + { + "bbox": [ + 130, + 574, + 482, + 648 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{O}^L\\}" + }, + { + "bbox": [ + 130, + 574, + 482, + 648 + ], + "type": "text", + "content": " and follow the instruction of HotspotNet to inject " + }, + { + "bbox": [ + 130, + 574, + 482, + 648 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{O}^L\\}" + }, + { + "bbox": [ + 130, + 574, + 482, + 648 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 574, + 482, + 648 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^U" + }, + { + "bbox": [ + 130, + 574, + 482, + 648 + ], + "type": "text", + "content": " in a semantically coherent manner. Since inserted objects come with ground-truth classes, poses and dimensions," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 146, + 90, + 447, + 103 + ], + "type": "text", + "content": "Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 369, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 369, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 369, + 666 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 653, + 369, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_{\\square}" + }, + { + "bbox": [ + 133, + 653, + 369, + 666 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 133, + 653, + 369, + 666 + ], + "type": "inline_equation", + "content": "\\square" + }, + { + "bbox": [ + 133, + 653, + 369, + 666 + ], + "type": "text", + "content": " component for location prediction " + }, + { + "bbox": [ + 133, + 653, + 369, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 133, + 653, + 369, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 472, + 314 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 472, + 314 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 472, + 314 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 472, + 314 + ], + "type": "image", + "image_path": "85c682fbbcaea53f860e6d584522858f6ebb12f03a2ccf8115d128d833ec7bcd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 323, + 482, + 357 + ], + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 357 + ], + "type": "text", + "content": "Fig. 3: (a) Common-used pseudo-labeling SSOD paradigm. (b) Cross-Scan Object Transfer SSOD paradigm, which bypasses the iterative self-training procedures, cumbersome pseudo-labels filter mechanisms and complex loss objectives designs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 381, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 482, + 453 + ], + "type": "text", + "content": "these newly created scans allow for supervision. However, it is important to note that the above synthesized point clouds are only partially annotated, as we lack information about the original unlabeled objects in these unannotated scans. Learning a detector from partially labeled data is known as Sparsely Annotated Object Detection (SAOD) [30, 41]. The differences of training a detector from fully and partially annotated data will be discussed in the next part." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 473, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 605 + ], + "type": "text", + "content": "Spatial-Aware Classification Loss An object detector generally includes a classifier to differentiate foreground objects from background clutter, and a localizer to predict the accurate location, dimension and orientation of an object. The localizer is typically supervised by regressing the offset from current voxel location to actual bounding boxes of foreground objects, so it can be naturally supervised by new point clouds generated via cross-scan object transfer. However, the classifier requires both positive and negative samples for learning a decision boundary. Once again, the inserted objects could provide positive samples for the classifier but we can not treat all other voxels as background in SAOD just like vanilla object detection. This is due to the existence of the so-called hard negatives as reported in other SAOD works [41]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": "In our approach, hard negatives refer to the original unlabeled objects intrinsically present in the unannotated LiDAR scans. Training the classifier in our semi-supervised detector with these hard negatives will result in an unacceptable number of false negatives. Instead of resorting to sophisticated re-weighting/ recalibration strategies [44,59] or teacher-student siamese networks [41], we simply" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "J. Zhan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "choose negative samples in a spatial-aware manner to avoid sampling unlabeled objects as negatives. We propose a spatial-aware classification loss for our partial supervision as Eq.6:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 218, + 158, + 481, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 158, + 481, + 174 + ], + "spans": [ + { + "bbox": [ + 218, + 158, + 481, + 174 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {c l s}} ^ {\\mathrm {n}} (\\mathbf {x}) = \\mathbb {1} \\left[ \\left(\\min _ {i} \\| \\mathbf {x} - \\mathbf {c} _ {i} \\| _ {2}\\right) \\leq \\sigma \\right] \\cdot L _ {\\mathrm {c l s}} ^ {\\mathrm {n}} (\\mathbf {x}) \\tag {6}", + "image_path": "faa878a2171941b73ce54d7a2c093684f18fc2a6f5eb54fafe97eb1f4a214185.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 178, + 482, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 178, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 482, + 299 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 178, + 482, + 299 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{cls}}^{\\mathrm{n}}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 178, + 482, + 299 + ], + "type": "text", + "content": " denotes focal loss for negative samples at the heatmap location " + }, + { + "bbox": [ + 130, + 178, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 178, + 482, + 299 + ], + "type": "text", + "content": ". The indicator function " + }, + { + "bbox": [ + 130, + 178, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot)" + }, + { + "bbox": [ + 130, + 178, + 482, + 299 + ], + "type": "text", + "content": " equals 1 when the " + }, + { + "bbox": [ + 130, + 178, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\ell 2" + }, + { + "bbox": [ + 130, + 178, + 482, + 299 + ], + "type": "text", + "content": " distance between a location " + }, + { + "bbox": [ + 130, + 178, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 178, + 482, + 299 + ], + "type": "text", + "content": " and the center location " + }, + { + "bbox": [ + 130, + 178, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 178, + 482, + 299 + ], + "type": "text", + "content": " of any foreground object is less than a threshold " + }, + { + "bbox": [ + 130, + 178, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 178, + 482, + 299 + ], + "type": "text", + "content": ", and 0 otherwise. We set " + }, + { + "bbox": [ + 130, + 178, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 178, + 482, + 299 + ], + "type": "text", + "content": " as twice of the Gaussian radius used for positive samples assignment in CenterPoint [54] for unlabeled scans, and infinity for fully annotated data. An immediate drawback of our spatial-aware classification loss is that there are too few negative samples for the classifier. To alleviate this issue, we randomly sample " + }, + { + "bbox": [ + 130, + 178, + 482, + 299 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 178, + 482, + 299 + ], + "type": "text", + "content": " samples in the heatmap predictions as negative samples, and set the corresponding value of " + }, + { + "bbox": [ + 130, + 178, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot)" + }, + { + "bbox": [ + 130, + 178, + 482, + 299 + ], + "type": "text", + "content": " as 1 to make the classification boundary between positive and negative samples more clear." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 312, + 479, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 479, + 371 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 479, + 371 + ], + "type": "text", + "content": "Synergy with Self-Training Techniques A deeper analysis of the proposed CSOT SSOD paradigm reveals that the primary mechanism driving improved detection performance is the extrapolation of labeled object detection to novel contextual environments. This stands in contrast to self-training methods, which learn from pseudo-labeled objects in unannotated scans." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 372, + 482, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 372, + 482, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 482, + 481 + ], + "type": "text", + "content": "As a result, the proposed CSOT pipeline for semi-supervised LiDAR object detection is orthogonal to, yet complementary with, existing self-training techniques [37], allowing for additional performance gains. Deliberately designing a pseudo-labeling SSOD framework is beyond the scope of this paper. To verify the complementary effect of our approach when combined with self-training techniques, we implement a basic teacher-student paradigm. We employ a single universal confidence score threshold " + }, + { + "bbox": [ + 130, + 372, + 482, + 481 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{pseudo}}" + }, + { + "bbox": [ + 130, + 372, + 482, + 481 + ], + "type": "text", + "content": ", which is invariant across categories, to filter low-confidence predictions. We then utilize a fully-supervised loss, as in CenterPoint [54], for optimizing the student detector." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 495, + 224, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 224, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 224, + 508 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 515, + 479, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 479, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 479, + 552 + ], + "type": "text", + "content": "The proposed methods are verified in this section. We show implementation details in Sec. 4.1 and provide all experimental results in Sec. 4.2 and Sec. 4.3. More ablations and discussions are shown in our supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 565, + 261, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 261, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 261, + 578 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Datasets We validate our method on two widely used autonomous driving datasets: nuScenes [3] and Waymo [38]. The nuScenes dataset has 700 training, 150 validation, and 150 test sequences. Official nuScenes detection metrics are NuScenes Detection Score (NDS) and mean Average Precision (mAP). The Waymo Open Dataset consists of 798 training and 202 validation sequences. Official Waymo detection metrics are mean Average Precision (mAP) and mean Average Precision weighted by Heading (mAPH)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 146, + 91, + 447, + 103 + ], + "type": "text", + "content": "Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "text", + "content": "Training Details We use one-stage CenterPoint [54] as our baseline detector to validate the effectiveness of our methods. Voxel sizes are set as " + }, + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "inline_equation", + "content": "[0.075\\mathrm{m}, 0.075\\mathrm{m}, 0.2\\mathrm{m}]" + }, + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "text", + "content": " for nuScenes and " + }, + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "inline_equation", + "content": "[0.1\\mathrm{m}, 0.1\\mathrm{m}, 0.15\\mathrm{m}]" + }, + { + "bbox": [ + 130, + 116, + 482, + 261 + ], + "type": "text", + "content": " for Waymo. The training epochs for one cycle of learning procedure, including fully-supervised learning, partially-supervised learning and self-training, are set as 20 for nuScenes and 7 for Waymo. We follow many previous work [6,51,54,66] to adopt Adam optimizer with one-cycle learning rate policy. Labeled data are uniformly sampled from the entire dataset. Then objects and their annotations are extracted to build the ground-truth database for injecting precise supervision signals into unlabeled scans. We finetune partially-supervised detectors on labeled scans with several epochs. We set a score threshold of 0.3 for nuScenes and 0.6 for Waymo to filter low-quality pseudo labels, and verify the orthogonality of our methods with self-training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 278, + 223, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 223, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 223, + 289 + ], + "type": "text", + "content": "4.2 Benchmarks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 300, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 479 + ], + "type": "text", + "content": "Benchmarks on nuScenes dataset We compare our methods against other label-efficient LiDAR detection approaches on the nuScenes dataset. As shown in Tab. 1, our methods obtain the significant performance improvements over previous works. For the sake of fairness, we choose the fully supervised detector trained by Yin et.al. [54] (the designer of CenterPoint backbone) as the baseline detector, and list the results achieved from other label-efficient LiDAR object detection approaches for comparison [27, 43, 56, 58]. Without leveraging models pre-trained on external datasets and using only " + }, + { + "bbox": [ + 130, + 300, + 482, + 479 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 300, + 482, + 479 + ], + "type": "text", + "content": " of the annotations, we achieve competitive detector which reaches " + }, + { + "bbox": [ + 130, + 300, + 482, + 479 + ], + "type": "inline_equation", + "content": "97.2\\%" + }, + { + "bbox": [ + 130, + 300, + 482, + 479 + ], + "type": "text", + "content": " of the fully-supervised performance on mAP and " + }, + { + "bbox": [ + 130, + 300, + 482, + 479 + ], + "type": "inline_equation", + "content": "98.9\\%" + }, + { + "bbox": [ + 130, + 300, + 482, + 479 + ], + "type": "text", + "content": " on NDS. We significantly shrink the mAP and NDS gaps between the baseline detector and the one trained with " + }, + { + "bbox": [ + 130, + 300, + 482, + 479 + ], + "type": "inline_equation", + "content": "20\\mathrm{x}" + }, + { + "bbox": [ + 130, + 300, + 482, + 479 + ], + "type": "text", + "content": " less annotations from around " + }, + { + "bbox": [ + 130, + 300, + 482, + 479 + ], + "type": "inline_equation", + "content": "10\\sim 20" + }, + { + "bbox": [ + 130, + 300, + 482, + 479 + ], + "type": "text", + "content": " to merely 1.6 and 0.7 respectively. Our semi-supervised detector, trained with only " + }, + { + "bbox": [ + 130, + 300, + 482, + 479 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 300, + 482, + 479 + ], + "type": "text", + "content": " manual annotations, obtains over " + }, + { + "bbox": [ + 130, + 300, + 482, + 479 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 130, + 300, + 482, + 479 + ], + "type": "text", + "content": " detection performance on 8 of the 10 categories defined in nuScenes dataset, and over " + }, + { + "bbox": [ + 130, + 300, + 482, + 479 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 300, + 482, + 479 + ], + "type": "text", + "content": " detection performance on 6 categories." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Benchmarks on Waymo dataset We extend our methods on Waymo dataset and achieve similar outstanding results. we choose the fully supervised detector trained by Shi et.al. [36], Yan et.al. [47], Fan et.al. [10] and Yin et.al. [54] as the baseline detector, and list the results achieved from other label-efficient LiDAR object detection approaches for comparison [29, 46, 50, 52, 53, 56]. As shown in Tab. 2, our methods are valid on both anchor-based (SECOND) and anchor-free (CenterPoint) detectors. Taking CenterPoint as an example, with merely " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " annotations, we achieve a competitive detector which matches fully supervised performance on vehicle and pedestrian, and exceeds it on cyclist. The " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "100\\mathrm{x}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " reduction in annotations leads to only a minor 1.8 APH decrease for L2 Vehicle, a negligible 0.4 APH drop for L2 Pedestrian, and boosts the APH of L2 Cyclist by 4.1. The abnormal enhanced detection performance on Cyclist should be attributed to our learning-based approach for cross-scan object transfer. Instructed by the proposed HotspotNet, we paste objects in a semantic coherent" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "J. Zhan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 157, + 480, + 218 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Compare with other labeled efficient LiDAR object detection approaches on nuScenes Dataset validation split. Performances of the baseline detector reported by its original designer [54] are marked with gray backgrounds." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 157, + 480, + 218 + ], + "lines": [ + { + "bbox": [ + 134, + 157, + 480, + 218 + ], + "spans": [ + { + "bbox": [ + 134, + 157, + 480, + 218 + ], + "type": "table", + "html": "
MethodRefAnnosExternalmAPNDSCarTruckC.V.BusTrailerBarrierMot.Byc.Ped.T.C.
BaselineCVPR 2021 [54]100%58.065.584.651.017.560.253.270.953.728.783.476.7
GCC3DICCV 2021 [27]5%41.146.8——————————
SceneFlowECCV 2022 [58]5%36.048.3——————————
SSDA3DAAAI 2023 [43]5%✓——76.2—————————
AD-PTNeurIPS 2023 [56]5%✓45.053.078.943.811.155.221.255.139.017.872.355.4
Ours—5%56.464.881.753.516.558.632.866.060.047.882.065.5
", + "image_path": "d803aba2d072488424b08e9c437d06f25d33a7e59bddcf6ea4e4e0b6fe5d9ce8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 270, + 479, + 402 + ], + "blocks": [ + { + "bbox": [ + 132, + 228, + 482, + 262 + ], + "lines": [ + { + "bbox": [ + 132, + 228, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 482, + 262 + ], + "type": "text", + "content": "Table 2: Compare with other labeled efficient LiDAR object detection approaches on Waymo Dataset validation split. Performances of the baseline detector reported by its original designers [10,36,47,54] are mark with gray backgrounds." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 270, + 479, + 402 + ], + "lines": [ + { + "bbox": [ + 135, + 270, + 479, + 402 + ], + "spans": [ + { + "bbox": [ + 135, + 270, + 479, + 402 + ], + "type": "table", + "html": "
DetectorMethodRefAnnoxExtVehiclePedestrianCyclistAll L2 AP
L1L2L1L2L1L2
APAPHAPAPHAPAPHAPAPHAPAPHAPAPH
PV-RCNNBaselineCVPR 2020 [36]100%77.576.969.068.475.065.766.057.667.866.465.464.063.3
PV-RCNNHSSDACVPR 2023 [29]1%56.453.849.747.340.120.933.517.529.120.927.920.028.3
PV-RCNNMIXSUPICLR 2024 [50]1%———55.5———52.0———62.356.6
SECONDBaselineSensors 2018 [47]100%72.371.763.963.368.758.260.751.360.659.358.357.157.2
SECONDMIXSUPICLR 2024 [50]10%———55.0———49.6———58.154.2
SECONDProficientTeacherECCV 2022 [52]5%——53.052.5——50.338.7——49.946.045.7
SECONDAD-PTNeurIPS 2023 [56]3%✓——60.559.9——54.945.8——50.849.751.8
SECONDProposalContrastECCV 2022 [53]1%——37.636.9——39.731.7——37.735.734.8
SECONDOurs—1%69.468.961.260.864.060.755.152.361.160.458.958.257.1
SSTBaselineCVPR 2022 [10]100%74.273.865.565.178.769.670.061.770.769.668.066.964.6
SSTMIXSUPICLR 2024 [50]10%———59.1———60.0———63.160.7
SSTMV-JARCVPR 2023 [46]5%——56.556.0——57.747.7——37.436.346.7
CenterPointBaselineCVPR 2021 [54]100%76.676.068.968.479.073.471.065.872.171.069.568.567.6
CenterPointMIXSUPICLR 2024 [50]10%———61.8———57.7———67.562.3
CenterPointAD-PTNeurIPS 2023 [56]3%✓——60.459.8——60.654.0——62.761.658.5
CenterPointOurs—1%75.174.767.066.675.972.968.265.475.975.173.472.668.2
", + "image_path": "bf4377f933e2e2eed69183592041761d241a69cc54f3b57c5edd8fdc09ffd648.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 425, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 510 + ], + "type": "text", + "content": "manner, which enables our CSOT SSOD to reach its full potential and meanwhile can be used as a powerful data augmentation for fully-supervised detector. Related ablation results are discussed in Sec. 4.3. Notably, under the most stringent constraints with minimal labeled point clouds and without leveraging any external pre-training, our approach stands alone as the only label-efficient technique which is able to achieve equivalence detection performance to the baseline detector on the average L2 APH of all categories defined in Waymo dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 526, + 210, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 526, + 210, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 210, + 537 + ], + "type": "text", + "content": "4.3 Ablations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Ablate training stage on nuScenes dataset. Our semi-supervised detectors trained with 20x fewer nuScenes annotations are compared to both fully supervised and pseudo-labeling semi-supervised models. Results in Tab. 3 indicate that our newly proposed CSOT is a valid label-efficient approach, and it almost completely maintains the performance gains of self-training when used together. The fully supervised detector trained with " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " labeled scans shows substantial performance degradation compared to that trained on the entire labeled dataset, with a mAP decline from 59.5 to 43.8 and a NDS decrease from 66.0 to 56.5. As shown in Lines 3-5 of Tab. 3, self-training technique is able to boost the mAP from 43.8 to 51.0 and NDS from 56.5 to 61.0. Our CSOT, which allows" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "text", + "content": "Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 158, + 480, + 211 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 3: Compare with fully supervised and pseudo-labeling semi-supervised detectors. All results are achieved based on our implementation code of CenterPoint and evaluated on nuScenes dataset validation split." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 158, + 480, + 211 + ], + "lines": [ + { + "bbox": [ + 134, + 158, + 480, + 211 + ], + "spans": [ + { + "bbox": [ + 134, + 158, + 480, + 211 + ], + "type": "table", + "html": "
AnnosScansCSOTSelf-TrainingmAPNDSCarTruckC.V.BusTrailerBarrierMot.Byc.Ped.T.C.
1100%100%59.566.083.756.618.766.934.867.163.750.484.568.7
25%5%43.856.575.839.712.344.120.556.936.318.476.756.9
35%100%✓51.061.080.149.114.655.626.762.345.430.881.963.1
45%100%✓50.461.079.646.411.750.125.659.152.137.280.662.0
55%100%✓✓56.464.881.753.516.558.632.866.060.047.882.065.5
", + "image_path": "ba2bd1af06fa71b7a599f3362493800f65ba680963c8ecdb279d875d742aff07.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 277, + 479, + 344 + ], + "blocks": [ + { + "bbox": [ + 132, + 223, + 482, + 267 + ], + "lines": [ + { + "bbox": [ + 132, + 223, + 482, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 223, + 482, + 267 + ], + "type": "text", + "content": "Table 4: Compare with fully supervised and pseudo-labeling semi-supervised detectors. All results are achieved based on our implementation code of CenterPoint and evaluated on Waymo dataset validation split (mean L1/L2 AP/APH over all categories defined in Waymo dataset)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 277, + 479, + 344 + ], + "lines": [ + { + "bbox": [ + 135, + 277, + 479, + 344 + ], + "spans": [ + { + "bbox": [ + 135, + 277, + 479, + 344 + ], + "type": "table", + "html": "
AnnosScansCSOTSelf-TrainingObject TransferL1 AP/APHL2 AP/APHmAP/mAPH
1100%100%—74.4 / 72.768.6 / 67.071.5 / 69.9
1a100%100%GT-Aug74.7 / 73.168.8 / 67.371.8 / 70.2
1b100%100%Ours77.5 / 76.071.6 / 70.274.6 / 73.1
21%1%—68.3 / 66.662.3 / 60.865.3 / 63.7
31%100%✓—71.9 / 70.465.8 / 64.568.9 / 67.5
41%100%✓Ours71.4 / 69.865.4 / 63.968.4 / 66.9
51%100%✓✓Ours75.7 / 74.269.5 / 68.272.6 / 71.2
", + "image_path": "8e70eb88e4899ecc4525c6d503673f221fc568b04b544e2981155b2c38971296.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 369, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 453 + ], + "type": "text", + "content": "supervision on unannotated point clouds by converting unlabeled scans into partially labeled scans via learning-based copy-paste, achieves a similar detection performance (with a mAP of 50.4 and a NDS of 61.0). It is noteworthy that iterative employment of self-training techniques does not lead to sustained gains in detection performance (see supplementary material), while incorporating just one cycle of self-training on our partially supervised detector is able to further optimize the mAP from 50.4 to 56.4 and NDS from 61.0 to 64.8." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": "Ablate training stage on Waymo dataset We extend the above ablations on Waymo dataset and summarize all results in Tab. 4, which further validate the effectiveness of CSOT, as well as its orthogonality with the common used pseudo-labeling techniques. " + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "inline_equation", + "content": "100\\mathrm{x}" + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": " fewer labeled point clouds lead to a degradation of detection performance, reducing the mean AP and APH from 71.5, 69.9 to 65.3, 63.7. Our Cross-Scan Object Transfer paradigm brings a boost of around 3 APH for both LEVEL_1 and LEVEL_2 objects. CSOT is proved to be complimentary to self-training, and the performance gains can be completely maintained. As shown in the Line 5 of Tab. 4, the performance of detector trained with " + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": " annotations can be finally optimized to match even surpass the fully-supervised baseline model. The abnormal enhanced detection performance with " + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "inline_equation", + "content": "100\\mathrm{x}" + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": " reduced labeled data should be attributed to the learning-based copy-paste operation employed in this work. As shown in Lines 1a-1b of Tab. 4, replacing the commonly used copy-paste augmentation (GT-Aug) with our learning based approaches boosts the fully-supervised detector with an increase of " + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "inline_equation", + "content": "2.8~\\mathrm{mAP}" + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": ". Notably, even regarding the fully-supervised detector augmented via our object" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "J. Zhan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 147, + 479, + 350 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 5: Ablate our learning-based cross-scan object transfer which is instructed by HotspotNet. All results are evaluated on nuScenes dataset validation split." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 147, + 479, + 350 + ], + "lines": [ + { + "bbox": [ + 134, + 147, + 479, + 350 + ], + "spans": [ + { + "bbox": [ + 134, + 147, + 479, + 350 + ], + "type": "table", + "html": "
(a) Compare our learning-based cross-scan object transfer (instructed by HotspotNet) with other similar methods.
AnnosCSOTObj TransferSelfObj Appearance rαLocal Context (x,y,z)Global Context rsclsrule /learnedmAP NDS
15%——————43.8 56.5
25%✓——————51.0 61.0
35%✓GT-Aug [47]✓✓46.0 57.7
45%✓CA-Aug [15]✓✓✓rule46.0 58.0
55%✓LiDAR-Aug [11]✓✓✓rule46.2 57.9
65%✓Ours✓✓✓✓✓learned50.4 61.0
6a5%✓Ours✓✓✓✓✓rule49.7 60.6
75%✓GT-Aug [47]✓✓49.7 60.5
85%✓CA-Aug [15]✓✓✓rule49.7 60.7
95%✓LiDAR-Aug [11]✓✓✓rule49.7 60.2
105%✓Ours✓✓✓✓✓learned56.4 64.8
10a5%✓Ours✓✓✓✓✓rule53.6 63.3
(b) Ablate contribution of each fitness score defined in Ln-score(o) (Eq. 5).
AnnosCSOTSelfposition sporientation socategory scmAP NDS
115%✓✓49.7 60.5
125%✓✓✓52.1 62.6
135%✓✓✓49.4 60.8
13a5%✓✓✓52.2 62.3
13b5%✓✓✓✓55.1 64.3
145%✓✓✓51.6 62.0
155%✓✓✓✓✓56.4 64.8
", + "image_path": "107a38255d1306e835e526aba177ea89dcd5b0d1ef1ea3f7f8e44821fbcf9a47.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 373, + 480, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 480, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 480, + 409 + ], + "type": "text", + "content": "transfer implementation as the baseline, reducing annotations by " + }, + { + "bbox": [ + 130, + 373, + 480, + 409 + ], + "type": "inline_equation", + "content": "100\\mathrm{x}" + }, + { + "bbox": [ + 130, + 373, + 480, + 409 + ], + "type": "text", + "content": " causes only a minimal degradation in performance (remains " + }, + { + "bbox": [ + 130, + 373, + 480, + 409 + ], + "type": "inline_equation", + "content": "97.3\\% / 97.4\\%" + }, + { + "bbox": [ + 130, + 373, + 480, + 409 + ], + "type": "text", + "content": " mAP/mAPH of our fully-supervised detector)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": "Ablate our learning-based object transfer. We compare our learning-based object transfer which instructed by the proposed HotspotNet with other similar techniques in Tab. 5(a), and ablate negative sample strategy employed for calculating " + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{score}}^n (o)" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": " (Eq. 5) in Tab. 5(b). As shown in Lines 3-10 of Tab. 5, the proposed learning-based object transfer is able to maximize the benefits of CSOT. It also stands alone in its ability to seamlessly integrate with self-training technique without compromising their performance gains. CSOT built upon our learning-based object transfer enhances the detection performance with a considerable " + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "inline_equation", + "content": "6.6\\mathrm{mAP}" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": " and 4.5 NDS. The improvement surpasses the one built upon previous similar approaches (with a margin of about " + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "inline_equation", + "content": "4\\mathrm{mAP}" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": " and 3 NDS). The poor partially-supervised detector achieved from other object transfer approaches produce enormous false positive predictions (visualized in Fig. 1) and greatly influence the downstream self-training. Replacing our learning-based object transfer with previous methods reduces the mAP gains of self-training from 6.0 to around 3.5, and degrades the final mAP from 56.4 to 49.7. The necessity of learning-based object transfer are verified in Tab. 5, Lines 6-6a,10-10a. Learn-based approach enables better generalization and can help reduce noise, which introduces an increased final mAP of 2.8. We further ablate negative sample strategy employed for calculating " + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{score}}^n (o)" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": " and list results in Lines 11-15 of Tab. 5. Pocation and category provide the most important contextual informa" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 90, + 447, + 102 + ], + "type": "text", + "content": "Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 147, + 479, + 282 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 6: Ablate robustness of our methods across various network architectures. All results are evaluated on nuScenes dataset validation split." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 147, + 479, + 282 + ], + "lines": [ + { + "bbox": [ + 135, + 147, + 479, + 282 + ], + "spans": [ + { + "bbox": [ + 135, + 147, + 479, + 282 + ], + "type": "table", + "html": "
AnnosDetectorEncoderAnchor-baseAnchor-freeVoxel SizeCSOTSelfmAPNDS
1100%CenterPointVFE✓(0.075,0.075,0.2)59.566.0
25%CenterPointVFE✓(0.075,0.075,0.2)43.856.5
35%CenterPointVFE✓(0.075,0.075,0.2)✓51.061.0
45%CenterPointVFE✓(0.075,0.075,0.2)✓50.461.0
55%CenterPointVFE✓(0.075,0.075,0.2)✓✓56.464.8
6100%CenterPointPFE✓(0.1,0.1,8.0)49.260.1
75%CenterPointPFE✓(0.1,0.1,8.0)30.248.3
85%CenterPointPFE✓(0.1,0.1,8.0)✓37.853.1
95%CenterPointPFE✓(0.1,0.1,8.0)✓40.054.9
105%CenterPointPFE✓(0.1,0.1,8.0)✓✓45.057.9
11100%SECONDVFE✓(0.1,0.1,0.2)55.464.7
125%SECONDVFE✓(0.1,0.1,0.2)38.453.2
135%SECONDVFE✓(0.1,0.1,0.2)✓45.258.1
145%SECONDVFE✓(0.1,0.1,0.2)✓45.459.1
155%SECONDVFE✓(0.1,0.1,0.2)✓✓48.661.3
", + "image_path": "b5faac4a2c5b301fdc9871036eaafa02f2b85c02a9438d9d4a2b00bd1fb51ce8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 304, + 480, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 353 + ], + "type": "text", + "content": "tion, orientation provides additional cues to narrow down the possible identity. Specifically, combining orientation constraint with category consistency improve mAP by 0.6 (Tab. 5, Lines 13a and 14). Removing orientation constraint from all reduces mAP by 1.3 (Tab. 5, Lines 13b and 15)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 370, + 481, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 481, + 525 + ], + "type": "text", + "content": "Extend our method to different detectors. We confirm the generality of our methods on various detectors, which utilizes diverse point cloud feature encoder or target assigner strategy. All results are listed in Tab. 6. For an anchor-free detector CenterPoint with a pillar-based feature encoder, the proposed CSOT bring a boost of " + }, + { + "bbox": [ + 132, + 370, + 481, + 525 + ], + "type": "inline_equation", + "content": "9.8\\mathrm{mAP}" + }, + { + "bbox": [ + 132, + 370, + 481, + 525 + ], + "type": "text", + "content": " and 6.6 NDS, which outperforms the self-training technique by " + }, + { + "bbox": [ + 132, + 370, + 481, + 525 + ], + "type": "inline_equation", + "content": "2.2\\mathrm{mAP}" + }, + { + "bbox": [ + 132, + 370, + 481, + 525 + ], + "type": "text", + "content": " and 1.8 NDS. Using our partially-supervised detector as the teacher model to generate pseudo-labels for unlabeled point clouds, and then conduct one-cycle of self-training. The performance of semi-supervised detector is ultimately reaches " + }, + { + "bbox": [ + 132, + 370, + 481, + 525 + ], + "type": "inline_equation", + "content": "45.0\\mathrm{mAP}" + }, + { + "bbox": [ + 132, + 370, + 481, + 525 + ], + "type": "text", + "content": " and 57.9 NDS. The mAP gap introduced by " + }, + { + "bbox": [ + 132, + 370, + 481, + 525 + ], + "type": "inline_equation", + "content": "20x" + }, + { + "bbox": [ + 132, + 370, + 481, + 525 + ], + "type": "text", + "content": " fewer annotations is narrowed from 19.0 to 4.2. Experiments on anchor-based detector (SECOND) also shows similar results. The effectiveness of CSOT and the additive effect of combining it with self-training techniques are clearly demonstrated in Lines 11-15 of Tab. 6." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 544, + 218, + 557 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 544, + 218, + 557 + ], + "spans": [ + { + "bbox": [ + 133, + 544, + 218, + 557 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 569, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 481, + 665 + ], + "type": "text", + "content": "In this work, we propose a novel Cross-Scan Object Transfer paradigm for LiDAR semi-supervised object detection. It features a HotspotNet for transferring objects semantic coherence manner, which reduces false positive predictions and enables the proposed CSOT to reach its full potential. To the best of our knowledge, this work is the first successful attempt of object transfer in LiDAR SSOD. The proposed CSOT paradigm is also orthogonal to the self-training techniques. We achieve competitive detectors on par with fully-supervised baselines while using 20x and 100x reduced annotations on nuScenes and Waymo dataset." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "J. Zhan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 353, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 353, + 154 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 353, + 154 + ], + "type": "text", + "content": "This work was supported by the InnoHK funding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 174, + 197, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 174, + 197, + 186 + ], + "spans": [ + { + "bbox": [ + 132, + 174, + 197, + 186 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 200, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 200, + 479, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 200, + 479, + 232 + ], + "spans": [ + { + "bbox": [ + 138, + 200, + 479, + 232 + ], + "type": "text", + "content": "1. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In: CVPR. pp. 1090-1099 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 233, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 233, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 233, + 480, + 255 + ], + "type": "text", + "content": "2. Blum, A., Mitchell, T.: Combining labeled and unlabeled data with co-training. In: COLT. pp. 92-100 (1998)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 256, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 480, + 277 + ], + "type": "text", + "content": "3. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR. pp. 11621-11631 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 278, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 278, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 138, + 278, + 480, + 310 + ], + "type": "text", + "content": "4. Chen, B., Chen, W., Yang, S., Xuan, Y., Song, J., Xie, D., Pu, S., Song, M., Zhuang, Y.: Label matching semi-supervised object detection. In: CVPR. pp. 14381-14390 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 312, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 480, + 332 + ], + "type": "text", + "content": "5. Chen, B., Li, P., Chen, X., Wang, B., Zhang, L., Hua, X.S.: Dense learning based semi-supervised object detection. In: CVPR. pp. 4815-4824 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 334, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 334, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 334, + 480, + 354 + ], + "type": "text", + "content": "6. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Largekernel3d: Scaling up kernels in 3d sparse cnns. In: CVPR. pp. 13488-13498 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 355, + 480, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 355, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 355, + 480, + 376 + ], + "type": "text", + "content": "7. Chen, Y., Liu, J., Zhang, X., Qi, X., Jia, J.: Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In: CVPR. pp. 21674-21683 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 378, + 480, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 378, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 378, + 480, + 399 + ], + "type": "text", + "content": "8. Choi, J., Song, Y., Kwak, N.: Part-aware data augmentation for 3d object detection in point cloud. In: IROS. pp. 3391-3397 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 400, + 480, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 400, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 400, + 480, + 422 + ], + "type": "text", + "content": "9. Dwibedi, D., Misra, I., Hebert, M.: Cut, paste and learn: Surprisingly easy synthesis for instance detection. In: ICCV. pp. 1301-1310 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 422, + 480, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 480, + 454 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 480, + 454 + ], + "type": "text", + "content": "10. Fan, L., Pang, Z., Zhang, T., Wang, Y.X., Zhao, H., Wang, F., Wang, N., Zhang, Z.: Embracing single stride 3d object detector with sparse transformer. In: CVPR. pp. 8458-8468 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 456, + 480, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 480, + 487 + ], + "type": "text", + "content": "1. Fang, J., Zuo, X., Zhou, D., Jin, S., Wang, S., Zhang, L.: Lidar-aug: A general rendering-based augmentation framework for 3d object detection. In: CVPR. pp. 4710-4720 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 488, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 488, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 488, + 480, + 510 + ], + "type": "text", + "content": "2. Gao, J., Wang, J., Dai, S., Li, L.J., Nevatia, R.: Note-rcnn: Noise tolerant ensemble rcnn for semi-supervised object detection. In: ICCV. pp. 9508-9517 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 510, + 480, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 480, + 543 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 480, + 543 + ], + "type": "text", + "content": "3. Ghiasi, G., Cui, Y., Srinivas, A., Qian, R., Lin, T.Y., Cubuk, E.D., Le, Q.V., Zoph, B.: Simple copy-paste is a strong data augmentation method for instance segmentation. In: CVPR. pp. 2918-2928 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 544, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 576 + ], + "type": "text", + "content": "4. Hu, J.S.K., Waslander, S.L.: Pattern-aware data augmentation for lidar 3d object detection. In: IEEE International Intelligent Transportation Systems Conference (ITSC). pp. 2703-2710 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 577, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 480, + 598 + ], + "type": "text", + "content": "5. Hu, X., Duan, Z., Ma, J.: Context-aware data augmentation for lidar 3d object detection. In: ICIP. pp. 11-15 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 632 + ], + "type": "text", + "content": "6. Hu, Y., Ding, Z., Ge, R., Shao, W., Huang, L., Li, K., Liu, Q.: Afdetv2: Rethinking the necessity of the second stage for object detection from point clouds. In: AAAI. pp. 969-979 (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "7. Huang, D., Chen, Y., Ding, Y., Liao, J., Liu, J., Wu, K., Nie, Q., Liu, Y., Wang, C., Li, Z.: Rethinking dimensionality reduction in grid-based 3d object detection. arXiv preprint arXiv:2209.09464 (2022)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "text", + "content": "Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "18. Jaderberg, M., Simonyan, K., Vedaldi, A., Zisserman, A.: Synthetic data and artificial neural networks for natural scene text recognition. arXiv preprint arXiv:1406.2227 (2014)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "text", + "content": "19. Kingma, D.P., Rezende, D.J., Mohamed, S., Welling, M.: Semi-supervised learning with deep generative models. In: Adv. Neural Inform. Process. Syst. (2014)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 173, + 480, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 173, + 480, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 480, + 195 + ], + "type": "text", + "content": "20. Laine, S., Aila, T.: Temporal ensembling for semi-supervised learning. In: ICLR (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 196, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 480, + 217 + ], + "type": "text", + "content": "21. Lang, A.H., Vora, S., Caesar, H., Zhou, L., Yang, J.: Pointpillars: Fast encoders for object detection from point clouds. In: CVPR. pp. 12697-12705 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 218, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 480, + 262 + ], + "type": "text", + "content": "22. Lee, D., Lee, J., Lee, J., Lee, H., Lee, M., Woo, S., Lee, S.: Regularization strategy for point cloud via rigidly mixed sample. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 15900-15909 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 262, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 480, + 285 + ], + "type": "text", + "content": "23. Lee, D.H.: Pseudo-label : The simple and efficient semi-supervised learning method for deep neural networks. In: ICML. p. 896 (2013)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 285, + 480, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 317 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 317 + ], + "type": "text", + "content": "24. Lee, S., Lim, H., Myung, H.: Patchwork++: Fast and robust ground segmentation solving partial under-segmentation using 3d point cloud. In: IROS. pp. 13276-13283 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 318, + 480, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 318, + 480, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 480, + 351 + ], + "type": "text", + "content": "25. Li, E., Wang, S., Li, C., Li, D., Wu, X., Hao, Q.: Sustech points: A portable 3d point cloud interactive annotation platform system. IEEE Intelligent Vehicles Symposium (IV) pp. 1108-1115 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 352, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 480, + 373 + ], + "type": "text", + "content": "26. Li, H., Wu, Z., Shrivastava, A., Davis, L.S.: Rethinking pseudo labels for semi-supervised object detection. In: AAAI. pp. 1314-1322 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "text", + "content": "27. Liang, H., Jiang, C., Feng, D., Chen, X., Xu, H., Liang, X., Zhang, W., Li, Z., Gool, L.V.: Exploring geometry-aware contrast and clustering harmonization for self-supervised 3d object detection. In: ICCV. pp. 3293-3302 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 408, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 480, + 441 + ], + "type": "text", + "content": "28. Liu, C., Zhang, W., Lin, X., Zhang, W., Tan, X., Han, J., Li, X., Ding, E., Wang, J.: Ambiguity-resistant semi-supervised learning for dense object detection. In: CVPR. pp. 15579-15588 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 442, + 480, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 480, + 474 + ], + "type": "text", + "content": "29. Liu, C., Gao, C., Liu, F., Li, P., Meng, D., Gao, X.: Hierarchical supervision and shuffle data augmentation for 3d semi-supervised object detection. In: CVPR. pp. 23819-23828 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 475, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 480, + 497 + ], + "type": "text", + "content": "30. Liu, C., Gao, C., Liu, F., Liu, J., Meng, D., Gao, X.: Ss3d: Sparsely-supervised 3d object detection from point cloud. In: CVPR. pp. 8428-8437 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 498, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 520 + ], + "type": "text", + "content": "31. Liu, Y.C., Ma, C.Y., He, Z., Kuo, C.W., Chen, K., Zhang, P., Wu, B., Kira, Z., Vajda, P.: Unbiased teacher for semi-supervised object detection. In: ICLR (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 521, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 480, + 542 + ], + "type": "text", + "content": "32. Liu, Y.C., Ma, C.Y., Kira, Z.: Unbiased teacher v2: Semi-supervised object detection for anchor-free and anchor-based detectors. In: CVPR. pp. 9819-9828 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 543, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 480, + 586 + ], + "type": "text", + "content": "33. Meng, Q., Wang, W., Zhou, T., Shen, J., Jia, Y., Gool, L.V.: Towards a weakly supervised framework for 3d point cloud object detection and annotation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 44(8), 4454-4468 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 587, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 480, + 609 + ], + "type": "text", + "content": "34. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR. pp. 652-660 (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 610, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 480, + 643 + ], + "type": "text", + "content": "35. Qian, G., Li, Y., Peng, H., Mai, J., Hammoud, H.A.A.K., Elhoseiny, M., Ghanem, B.: Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In: Adv. Neural Inform. Process. Syst. pp. 23192-23204 (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "36. Shi, S., Guo, C., Jiang, L., Wang, Z., Shi, J., Wang, X., Li, H.: Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In: CVPR. pp. 10529–10538 (2020)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "J. Zhan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "37. Sohn, K., Zhang, Z., Li, C.L., Zhang, H., Lee, C.Y., Pfister, T.: A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 205 + ], + "type": "text", + "content": "38. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhao, S., Cheng, S., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: CVPR. pp. 2446-2454 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 482, + 237 + ], + "type": "text", + "content": "39. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In: Adv. Neural Inform. Process. Syst. (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 238, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 482, + 270 + ], + "type": "text", + "content": "40. Wang, H., Cong, Y., Litany, O., Gao, Y., Guibas, L.J.: 3dioumatch: Leveraging iou prediction for semi-supervised 3d object detection. In: CVPR. pp. 14615-14624 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 482, + 293 + ], + "type": "text", + "content": "41. Wang, T., Yang, T., Cao, J., Zhang, X.: Co-mining: Self-supervised learning for sparsely annotated object detection. In: AAAI. pp. 2800-2808 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 482, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 482, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 482, + 325 + ], + "type": "text", + "content": "42. Wang, X., Yang, X., Zhang, S., Li, Y., Feng, L., Fang, S., Lyu, C., Chen, K., Zhang, W.: Consistent-teacher: Towards reducing inconsistent pseudo-targets in semi-supervised object detection. In: CVPR. pp. 3240-3249 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 482, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 482, + 357 + ], + "type": "text", + "content": "43. Wang, Y., Yin, J., Li, W., Frossard, P., Yang, R., Shen, J.: Ssda3d: Semi-supervised domain adaptation for 3d object detection from point cloud. In: AAAI. pp. 2707-2715 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 482, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 482, + 379 + ], + "type": "text", + "content": "44. Wu, Z., Bodla, N., Singh, B., Najibi, M., Chellappa, R., Davis, L.S.: Soft sampling for robust object detection. arXiv preprint arXiv:1806.06986 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 482, + 413 + ], + "type": "text", + "content": "45. Xu, M., Zhang, Z., Hu, H., Wang, J., Wang, L., Wei, F., Bai, X., Liu, Z.: End-to-end semi-supervised object detection with soft teacher. In: ICCV. pp. 3060-3069 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 414, + 482, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 482, + 445 + ], + "type": "text", + "content": "46. Xu, R., Wang, T., Zhang, W., Chen, R., Cao, J., Pang, J., Lin, D.: Mv-jar: Masked voxel jigsaw and reconstruction for lidar-based self-supervised pre-training. In: CVPR. pp. 13445-13454 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 447, + 482, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 482, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 482, + 467 + ], + "type": "text", + "content": "47. Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sensors 18(10), 3337 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 482, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 482, + 490 + ], + "type": "text", + "content": "48. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d: Self-training for unsupervised domain adaptation on 3d object detection. In: CVPR. pp. 10368-10378 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 491, + 482, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 482, + 522 + ], + "type": "text", + "content": "49. Yang, Q., Wei, X., Wang, B., Hua, X.S., Zhang, L.: Interactive self-training with mean teachers for semi-supervised object detection. In: CVPR. pp. 5941-5950 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 523, + 482, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 482, + 544 + ], + "type": "text", + "content": "50. Yang, Y., Fan, L., Zhang, Z.: Mixsup: Mixed-grained supervision for labeled efficient lidar-based 3d object detection. In: ICLR (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 545, + 482, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 482, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 482, + 577 + ], + "type": "text", + "content": "51. Ye, D., Zhou, Z., Chen, W., Xie, Y., Wang, Y., Wang, P., Foroosh, H.: Lidar-multinet: Towards a unified multi-task network for lidar perception. In: AAAI. pp. 3231-3240 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 578, + 482, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 482, + 610 + ], + "type": "text", + "content": "52. Yin, J., Fang, J., Zhou, D., Zhang, L., Xu, C.Z., Shen, J., Wang, W.: Semi-supervised 3d object detection with proficient teachers. In: ECCV. pp. 727-743 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 611, + 482, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 482, + 643 + ], + "type": "text", + "content": "53. Yin, J., Zhou, D., Zhang, L., Fang, J., Xu, C.Z., Shen, J., Wang, W.: Proposal-contrast: Unsupervised pre-training for lidar-based 3d object detection. In: ECCV. pp. 17-33 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 644, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 482, + 665 + ], + "type": "text", + "content": "54. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: CVPR. pp. 11784-11793 (2021)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "text", + "content": "Cross-Scan Object Transfer for Semi-Supervised LiDAR Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 446 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "55. You, Y., Luo, K., Phoo, C.P., Chao, W.L., Sun, W., Hariharan, B., Campbell, M., Weinberger, K.Q.: Learning to detect mobile objects from lidar scans without labels. In: CVPR. pp. 1130-1140 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "56. Yuan, J., Zhang, B., Yan, X., Chen, T., Shi, B., Li, Y., Qiao, Y.: Ad-pt: Autonomous driving pre-training with large-scale point cloud dataset. In: Adv. Neural Inform. Process. Syst. p. 36 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "57. Yun, S., Han, D., Oh, S.J., Chun, S., Cho, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: CVPR. pp. 6023-6032 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "text", + "content": "58. Yurtsever, E., Erçelik, E., Liu, M., Yang, Z., Zhang, H., Topçam, P., Listl, M., Çayli, Y.K., Knoll, A.: 3d object detection with a self-supervised lidar scene flow backbone. In: ECCV. pp. 247-265 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "59. Zhang, H., Chen, F., Shen, Z., Hao, Q., Zhu, C., Savvides, M.: Solving missing-annotation object detection with background recalibration loss. In: ICASSP. pp. 1888-1892 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 304 + ], + "type": "text", + "content": "60. Zhang, L., Sun, Y., Wei, W.: Mind the gap: Polishing pseudo labels for accurate semi-supervised object detection. In: AAAI. pp. 3463-3471 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "61. Zhang, L., Yang, A.J., Xiong, Y., Casas, S., Yang, B., Ren, M., Urtasun, R.: Towards unsupervised object detection from lidar point clouds. In: CVPR. pp. 9317-9328 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "62. Zhao, N., Chua, T.S., Lee, G.H.: Sess: Self-ensembling semi-supervised 3d object detection. In: CVPR. pp. 11079-11087 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 380 + ], + "type": "text", + "content": "63. Zheng, W., Tang, W., Jiang, L., Fu, C.W.: Se-ssd: Self-ensembling single-stage object detector from point cloud. In: CVPR. pp. 14494–14503 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "64. Zhou, H., Ge, Z., Liu, S., Mao, W., Li, Z., Yu, H., Sun, J.: Dense teacher: Dense pseudo-labels for semi-supervised object detection. In: ECCV. pp. 35-50 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "text", + "content": "65. Zhou, Q., Yu, C., Wang, Z., Qian, Q., Li, H.: Instant-teaching: An end-to-end semi-supervised object detection framework. In: CVPR. pp. 4081-4090 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 424, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 446 + ], + "type": "text", + "content": "66. Zhu, B., Jiang, Z., Zhou, X., Li, Z., Yu, G.: Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492 (2019)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "J. Zhan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CTRLorALTer_ Conditional LoRAdapter for Efficient 0-Shot Control & Altering of T2I Models/d556ef6f-c8cf-4800-af84-90cf37eef09f_content_list.json b/2024/CTRLorALTer_ Conditional LoRAdapter for Efficient 0-Shot Control & Altering of T2I Models/d556ef6f-c8cf-4800-af84-90cf37eef09f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7c6be13652e72f4a6b4278d887ed21019ec12d4d --- /dev/null +++ b/2024/CTRLorALTer_ Conditional LoRAdapter for Efficient 0-Shot Control & Altering of T2I Models/d556ef6f-c8cf-4800-af84-90cf37eef09f_content_list.json @@ -0,0 +1,1799 @@ +[ + { + "type": "text", + "text": "CTRLorALTer: Conditional LoRAapter for Efficient 0-Shot Control & Altering of T2I Models", + "text_level": 1, + "bbox": [ + 217, + 145, + 785, + 189 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nick Stracke1, Stefan Andreas Baumann1, Joshua Susskind2, Miguel Angel Bautista2, and Björn Ommer1", + "bbox": [ + 282, + 215, + 720, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "CompVis @ LMU Munich, MCML", + "bbox": [ + 375, + 257, + 624, + 271 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Apple", + "bbox": [ + 473, + 272, + 529, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Text-to-image generative models have become a prominent and powerful tool that excels at generating high-resolution realistic images. However, guiding the generative process of these models to take into account detailed forms of conditioning reflecting style and/or structure information remains an open problem. In this paper, we present LoRAAdapter, an approach that unifies both style and structure conditioning under the same formulation using a novel conditional LoRA block that enables zero-shot control. LoRAAdapter is an efficient and powerful approach to condition text-to-image diffusion models, which enables fine-grained control conditioning during generation and outperforms recent state-of-the-art approaches.", + "bbox": [ + 261, + 324, + 738, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Project page: compvis.github.io/LoRAAdapter/", + "bbox": [ + 261, + 477, + 581, + 489 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: T2I Models $\\cdot$ Diffusion Models $\\cdot$ Controllable Generation", + "bbox": [ + 261, + 503, + 725, + 517 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 566, + 374, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-image models have become a foundational component of Computer Vision, especially with the introduction of large-scale diffusion models like Stable Diffusion (SD) [33], DALL-E 2 [31], Imagen [37], RAPHAEL [44] and, eDiff-I [2]. These models have empowered users to create realistic images from textual prompts, although crafting effective prompts is a complex task in itself due to the nuanced prompt engineering required [40] to effectively condition the image generation process. To tackle this problem, prior approaches such as SD Image Variations [1], Stable unCLIP [33], and Composer [15], have attempted to perform conditioning using image prompts to enhance the expressivity of the conditioning mechanism giving users more flexibility during generation. In practice, this family of approaches requires either training from scratch or fully fine-tuning large text-conditioned model weights, incurring high computational costs and requiring large datasets.", + "bbox": [ + 217, + 597, + 785, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To reduce training cost and time, recent approaches like ControlNet [47], T2I-Adapter [24], Uni-ControlNet [50] or IP-Adapters [45] have focused on designing adapters that keep the base model frozen and only introduce a limit number of", + "bbox": [ + 215, + 795, + 785, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f47f35ce19fa3fb2627583f25816ce11b85a6c4d814339b6739cd38db5fc8cde.jpg", + "image_caption": [ + "Fig. 1: LoRAAdapter allows structure and style control of the image generation process of text-to-image models in a zero-shot manner. Our approach enables powerful fine-grained and efficient unified control over both structure and style conditioning using conditional LoRA blocks." + ], + "image_footnote": [], + "bbox": [ + 259, + 150, + 743, + 366 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "new layers. This not only reduces the number of trainable parameters but also allows training on much smaller datasets because the frozen base model cannot suffer from catastrophic forgetting. These adapters are typically handcrafted to a specific model architecture and conditioning modality. They can be roughly assigned into one of two groups:", + "bbox": [ + 212, + 462, + 787, + 539 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Adapters for pixel-level fine-grained conditioning over the generated image. This is often referred to as structural or local conditioning since it allows influencing the generated image on a per-pixel level by e.g. providing a depth map and effectively setting the layout of the image [7,17,24,29,47,50]", + "- Adapters for style or global conditioning perform tasks such as image-to-image translation similar to unCLIP [15, 19, 33, 43]. Here, the focus is not on fine-grained control but instead on generating an image similar to the prompt image, where similarity is defined by sharing the same style or semantics." + ], + "bbox": [ + 223, + 545, + 787, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Empirically, adapters always favor one type of conditioning over the other. For example, both Uni-ControlNet [50] and T2I-Adapter [24] have attempted style conditioning but their performance is not comparable to leading style approaches [45]. Conversely, style adapters require ControlNet or T2I-Adapter for structure conditioning. However, ControlNet is a large adapter that copies the entire encoder of Stable Diffusion's U-Net which significantly increases compute and thus inference time. Our goal is to find a unified approach that is capable of both structure and style conditioning, providing complete control over the generated image. Formulating such a unified approach for conditioning on global controls like style and on local controls like structure, in an efficient and generic manner remains a key open problem.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Stracke et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this challenging problem, we introduce LoRAAdapter (see Fig. 1), a unified approach for incorporating images to condition both the style and structure of the generated images. LoRAAdapter takes inspiration from Low-Rank-Adaptations (LoRAs) [14], which have shown remarkable performance for various tasks in diffusion models, such as distillation or learning specific concepts [23,35]. In particular, LoRAs low-rank property naturally regularizes the conditioning while offering unparalleled flexibility as any layer can be adapted in an architecture-agnostic way. Thus far, however, LoRAs have been implemented as fixed adaptations, being independent of the input. This means layer adaptation cannot change at test time for a given conditioning and, therefore, cannot be used for zero-shot generalization.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LoRAAdapter is a novel approach to adding conditional information to LoRAs, enabling zero-shot generalization and making them applicable for both structure and style and possibly many other conditioning types. LoRAAdapter is compact and efficient, e.g., optimizing 16M parameters vs the 22M of IP-Adapters [45] or 361M of ControlNet [47], while outperforming recent adapter approaches [45] and even approaches that train models from scratch (see Tab. 1). LoRAAdapter trains LoRA [14] modules for different blocks across the base network and is efficient both during training and inference. Our contributions are summarized as follows:", + "bbox": [ + 212, + 311, + 787, + 434 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose LoRAAdapter, a generic approach to train conditional LoRAs that is agnostic to model architecture and conditioning modality.", + "- We implement LoRAAdapter for Stable Diffusion, offering a unified conditioning mechanism for both style and structure, enabling zero-shot conditioning.", + "- We show the effectiveness of this approach, outperforming dedicated adapters for either style and structure on various metrics." + ], + "bbox": [ + 225, + 446, + 782, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 561, + 387, + 578 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now review existing adapter approaches for style and structure and discuss the current role of LoRAs in diffusion models.", + "bbox": [ + 212, + 594, + 785, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Structure Adapters", + "text_level": 1, + "bbox": [ + 214, + 647, + 423, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Structure adapters are specifically designed to operate on various types of local conditioning modalities such as Depth, HED [41], Canny Edges [4], Scribbles, or key poses [5]. The goal of these adapters is to spatially align the generated image with the provided condition. One of the most prominent approaches to tackle this problem is ControlNet [47] for Stable Diffusion. ControlNet creates a copy of the U-Net's encoder and combines the skip connections additively with the original skip connections, which then get fed into the original decoder. While obtaining great performance, ControlNet adds a lot of computational overhead as the forward pass includes a second encoder which can increase inference speed by up to $50\\%$ . Another consequence of the large capacity is that ControlNet tends to interpret the structure map directly. This can be desirable when sampling", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "2. RELATED WORK", + "bbox": [ + 576, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "images without other conditioning via the text prompt or other adapters, but can also lead to entanglement of structure and style.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following a similar idea, T2I-Adapter [24] also modifies skip connections while using a much smaller encoder network. To extend conditioning to multiple modalities, Uni-ControlNet [50] was proposed as an alternative solution to having to train multiple separate and modality-specific ControlNets. Uni-ControlNet trains a single model for multiple conditions by concatenating the structure maps along the channel dimension and using the concatenated tensor as input. Uni-ControlNet uses SPADE [25] to combine skip-connections instead of the simple addition used by ControlNet. SCEdit [17] proposed a similar but more efficient solution for conditioning on multiple modalities by integrating a small tuner network (SC-Tuner) between skip-connections which removes the need for a separate encoder network akin to ControlNet.", + "bbox": [ + 212, + 176, + 787, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Style Adapters", + "text_level": 1, + "bbox": [ + 214, + 364, + 385, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Style adapters are used as an alternative to unCLIP [33] like models, which are directly trained to invert CLIP image embeddings. The main advantage of using a style adapter as opposed to an unCLIP model is that the original text conditioning is preserved, enabling additional control over the generated images. Additionally, all recently released large open-source diffusion models only offered text conditioning, making an adapter approach necessary [27, 28].", + "bbox": [ + 212, + 387, + 787, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "IP-Adapter [45] adds a decoupled cross-attention layer that operates on four image tokens instead of text tokens. The resulting activations of this decoupled cross-attention layer are scaled and added to the original cross-attention activations. This results in a lightweight adapter that integrates well with potential text conditioning and generates images faithful to the image prompt. In contrast, SeeCoder [42] trains an encoder with 2D spatial embedding to convert image tokens to text tokens and completely replaces the standard text tokens with their tokens.", + "bbox": [ + 212, + 478, + 787, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Structure adapters have also attempted to include style conditioning. UniControlNet [50] includes a separate mapping network to convert a pooled CLIP image token to four CLIP text tokens, which they concatenate to the original text embeddings. ControlNet-shuffle [47] shuffles an RGB image and uses that as conditioning similar to its structure conditioning. However, the performance of these approaches does not compare favorably to dedicated style adapters.", + "bbox": [ + 212, + 599, + 787, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 LoRAs in Diffusion Models", + "text_level": 1, + "bbox": [ + 214, + 710, + 486, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While LoRAs were initially introduced as an efficient alternative to fine-tuning large language models [14], they were quickly adopted for diffusion models [36] as a relatively efficient and cheap option to adjust their generation capabilities. Combining LoRAs with methods such as Dreambooth [34] performs on par with a full fine-tuning approach while only introducing a fraction of trainable parameters. LoRAs are commonly trained on a small set of images with the goal of capturing a specific style instead of a subject. ZipLoRA [38] proposes a merging algorithm to", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Stracke et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "combine multiple LoRAs gracefully by minimizing the cosine similarity of LoRAs that adapt the same weight matrix. LyCORIS [46] trained an entire library of specific LoRAs in diffusion models with various benefits. [23] showed that LoRAs are even powerful enough for more complex objectives such as consistency distillation. Finally, Concept Sliders [8] use LoRAs to train disentangled edit directions to edit images. They scale the LoRA in the negative and positive directions to either increase or decrease a given attribute.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 276, + 380, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Diffusion", + "text_level": 1, + "bbox": [ + 215, + 308, + 336, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Diffusion models represent a category of generative models characterized by a two-stage process: a forward diffusion that incrementally introduces Gaussian noise over a series of $T$ steps, and a reverse denoising process that reconstructs samples from this noise. In text-to-image models, diffusion can be directed by supplementary inputs like text descriptions. The primary training objective, $\\epsilon_{\\theta}$ , aimed at noise prediction, simplifies the variational lower bound as follows:", + "bbox": [ + 212, + 334, + 787, + 426 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {s i m p l e}} = \\mathbb {E} _ {\\boldsymbol {x} _ {0}, \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}), \\boldsymbol {c}, t} \\| \\boldsymbol {\\epsilon} - \\epsilon_ {\\theta} \\big (\\underbrace {\\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}} _ {\\boldsymbol {x} _ {t}}, \\boldsymbol {c}, t \\big) \\| ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 287, + 436, + 787, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{x}_0$ is the original data influenced by condition $\\pmb{c}$ , $t$ is the diffusion timestep, and $\\pmb{x}_t$ represents the data at step $t$ , with $\\bar{\\alpha}_t$ dictating the noise level.", + "bbox": [ + 212, + 478, + 787, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During sampling, an initial pure Gaussian latent $\\pmb{x}_T$ is iteratively denoised by the diffusion model to obtain a sample $\\pmb{x}_0$ from the modeled (conditional) distribution. Conditional models often employ classifier guidance to fine-tune the balance between accuracy and diversity, with classifier-free guidance [12] serving as a preferred alternative that merges conditional and unconditional noise predictions during training and sampling:", + "bbox": [ + 212, + 508, + 787, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {\\epsilon}} _ {\\theta} \\left(\\boldsymbol {x} _ {t}, \\boldsymbol {c}, t\\right) = w \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {x} _ {t}, \\boldsymbol {c}, t\\right) + (1 - w) \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {x} _ {t}, t\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 613, + 785, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w$ modulates the adherence to condition $c$ . This technique is pivotal for text-to-image models, enhancing the alignment between generated images and text prompts.", + "bbox": [ + 212, + 640, + 787, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Low-Rank Adaptation (LoRA)", + "text_level": 1, + "bbox": [ + 215, + 708, + 514, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Low-Rank Adaptation (LoRA) [14] is a method for finetuning pre-trained foundation models by keeping the original weight matrices frozen and instead adding a new set of low-rank weight matrices. While this method was originally proposed in the context of large language models, it has been recently adopted in the domain of diffusion models [6,8,23].", + "bbox": [ + 212, + 733, + 787, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, the original weight matrix of a layer in the base model $W_0 \\in \\mathbb{R}^{d \\times k}$ is kept frozen and is adapted by a low-rank matrix $W + \\Delta W = W + BA$ with", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "3. PRELIMINARIES", + "bbox": [ + 581, + 114, + 730, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$B \\in \\mathbb{R}^{d \\times r}$ and $A \\in \\mathbb{R}^{r \\times k}$ . Crucially, $r$ is chosen following $r \\ll \\min(d, k)$ . The forward pass for the adapted layer is then:", + "bbox": [ + 215, + 145, + 787, + 176 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nh = W _ {0} x + \\Delta W x = W _ {0} x + B A x. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 379, + 184, + 787, + 200 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Typically, $B$ is zero-initialized, so the LoRA does not affect the model initially. This yields a generic approach, as we can freely choose the rank $r$ and the weights according to the given task. This happens on a per-layer basis, i.e., if we adapt $n$ layers, we will have $nA$ and $B$ matrices.", + "bbox": [ + 212, + 208, + 787, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 215, + 287, + 330, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "There are two kinds of approaches for conditioning text-to-image diffusion models: style or global conditioning [42, 45], and structure or local conditioning [17, 24, 29, 50]. These approaches are specifically tailored to a specific conditioning objective and therefore missing a holistic approach that can efficiently learn both style and structure. We now describe LoRAAdapter, an efficient, holistic, and customizable method for introducing both style and structure conditioning into text-to-image diffusion models.", + "bbox": [ + 212, + 316, + 787, + 422 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/096c5c3ab9630367a314d2ecf880751abe6c2e22ac93da39af88171be02d1a26.jpg", + "image_caption": [ + "Fig. 2: Overview of the proposed conditional LoRA block. The original weight matrix $W_0^{(i)}$ is frozen while all other layers are trained. $\\phi$ is an affine transformation that operates on the low-dimensional embedding $A^{(i)}x$ and introduces the conditioning. The local mapper network $m_L^{(i)}$ predicts the scale and shift parameters $\\beta, \\gamma$ for the affine transformation. Typically, we set $m_L^{(i)}$ to be a small network. If complex transformations are required to map the conditioning $c$ , this happens in $m_S$ since it is shared across all adapted layers." + ], + "image_footnote": [], + "bbox": [ + 357, + 452, + 647, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Conditional LoRAs", + "text_level": 1, + "bbox": [ + 215, + 787, + 423, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LoRAs have successfully enabled adaptation of diffusion models across a range of architectures [6,23,35] to steer their generation process. In this setting, LoRAs", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Stracke et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "have been considered as free parameters and optimized on a training set of images exhibiting a specific characteristic.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Instead of learning a single LoRA to apply one specific modification, e.g., changing the style of the generated image to a painting or learning to represent one specific subject [34,35], we propose a LoRA-based conditioning mechanism whose behavior changes based on conditioning provided at inference time, enabling zero-shot generalization. While previous methods depend on specific aspects of the architectures of the models they are applied to [45, 47], using LoRAs for conditioning enables us to efficiently adapt foundation models for conditional tasks in an architecture-agnostic way. Our LoRA-based conditioning mechanism is shown in Fig. 2.", + "bbox": [ + 212, + 176, + 785, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Starting from the standard LoRA setup as in Eq. (3), we propose implementing conditioning of the LoRA by applying a transformation $\\phi(Ax|m(c))$ to the low-dimensional intermediate embedding in the LoRA that introduces conditional behavior based on the conditioning $c$ :", + "bbox": [ + 212, + 313, + 785, + 372 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nh = W _ {0} x + \\Delta W x = W _ {0} x + B \\phi (A x | m (c)). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 383, + 785, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This limits the conditioning to apply only to a low-rank subspace of the original vector space the weight matrix operates in. This is crucial for the generalization capabilities of LoRAs, as it introduces regularization and thus limits the adaptation to only focus on relevant aspects instead of spurious correlations from the dataset it was trained on, which results in more efficient training. By introducing our conditional adaptation of the model's behavior in such a low-rank space, our method can also benefit from those advantages.", + "bbox": [ + 212, + 409, + 785, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our framework for conditional LoRAs, we can theoretically choose $\\phi(Ax|m(c))$ to be arbitrary, non-linear transformations, but we find that a simple affine transformation is sufficiently expressive for a wide range of applications:", + "bbox": [ + 212, + 513, + 795, + 559 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\phi (A x \\mid m (c)) = \\gamma_ {\\phi} (m (c)) \\odot A x + \\beta_ {\\phi} (m (c)), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 571, + 785, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with $\\odot$ denoting an elementwise (Hadamard) product and $\\gamma$ and $\\beta$ referring to the scale and shift factors.", + "bbox": [ + 212, + 595, + 785, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To predict the $\\gamma$ and $\\beta$ from the conditioning $c$ , we utilize a mapping network. This is generally a small neural network that we separate into a shared part $m_{S}(\\cdot)$ and a layer-specific part $m_{L}^{(i)}(\\cdot)$ that is individually learned for each adapted layer. For both parts, we generally use very small neural networks, which is in stark contrast to other works such as ControlNet [47] that uses a full copy of the U-Net's encoder (see Tab. 1). For brevity, we will always refer to both parts simply as $m(\\cdot)$ .", + "bbox": [ + 212, + 626, + 785, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This general formulation directly enables both local and global conditioning, as it can be applied to the convolutional and attention layers in a standard diffusion model. For local conditioning, the mapping network can supply spatially dependent LoRA modulations. At the same time, it can also be used to supply global conditioning by introducing non-spatially dependent information, such as by adapting the cross-attention projections. Concurrent works [18, 21, 48] also introduced improvements to the basic LoRA architecture. As LoRAAdapter does", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "4. METHOD", + "bbox": [ + 635, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "not depend on a specific implementation of LoRAs but just on the low-rank intermediate sub-space, these improvements can be directly transferred to our model.", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 A Unified Conditioning Approach", + "text_level": 1, + "bbox": [ + 214, + 215, + 542, + 231 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we provide implementation details for applying our conditional LoRAs to the layer types commonly used in large-scale diffusion models.", + "bbox": [ + 212, + 241, + 785, + 272 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6973291b2a25e51354fbde26ccae2dc3e154d4d904d75f6f26b72638460445a7.jpg", + "image_caption": [ + "(a) In attention layers, we apply an affine transformation on the $K$ and $V$ projection matrix per channel for every token." + ], + "image_footnote": [], + "bbox": [ + 215, + 297, + 478, + 353 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/19b06692c5434f81e10f540612515df71cd7da8d80dc071b2b9afd8793819a09.jpg", + "image_caption": [ + "(b) To add structural conditioning to convolutional layers, we perform an affine transformation for every element similar to [25].", + "Fig. 3: Visualization of implementations of our conditional LoRAs for specific layers." + ], + "image_footnote": [], + "bbox": [ + 531, + 303, + 781, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Attention Layers Many diffusion models, such as Stable Diffusion $\\{1,2,\\mathrm{XL}\\}$ [28,33] or Imagen [37] include self- and cross-attention layers that operate on image (and text) tokens, with transformer-based diffusion models such as [26] even purely relying on them for spatial computations. The output $h^\\prime$ of the attention layers is defined as", + "bbox": [ + 212, + 478, + 787, + 554 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nh ^ {\\prime} = \\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) V \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 564, + 787, + 601 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with the projections $Q = hW_{q}$ , $K = hW_{k}$ or $K = c_{t}W_{k}$ , and $V = hW_{v}$ or $V = c_{t}W_{v}$ , depending on whether it is self- or cross-attention, and the text tokens $c_{t}$ . Here, we apply our conditional LoRA on the linear projections $W_{k}$ and $W_{v}$ based on Equation 4. We show a diagram of conditional LoRAs for attention layers in Fig. 3a.", + "bbox": [ + 212, + 609, + 787, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Convolutional Layers For convolutional layers, which form the main backbone of many diffusion models [27,33], we adapt the convolutional layer", + "bbox": [ + 212, + 708, + 785, + 739 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nh = K _ {0} \\star x, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 455, + 751, + 787, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with the 2D convolution kernel $K$ having input channel count $ch_{i}$ and output channel count $ch_{o}$ , in a similar manner as linear projections. However, in this setting we work on convolution kernels instead of matrices: we perform one convolution $K_{A}$ with the same kernel size, stride, and padding as the original", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Stracke et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ca5a72631ba41370222f12e19d8f6a14ac38c0c282823a8c20f954c659a2eee0.jpg", + "table_caption": [ + "Table 1: Quantitative comparison for style conditioning of our proposed LoRAAdapter with other methods on the COCO validation set with four samples for every image. The best results are in bold (adapted from [45])." + ], + "table_footnote": [], + "table_body": "
Style MethodReusable to custom modelsNative structure controlMultimodal promptsTrainable parametersCLIP-T ↑CLIP-I ↑
Training from scratch
Open unCLIPXXX893M0.6080.858
Kandinsky-2-1XXX1229M0.5990.855
Versatile DiffusionXX✓860M0.5870.830
Fine-tunining from text-to-image model
SD Image VariationsXXX860M0.5480.760
SD unCLIPXXX870M0.5840.810
Adapters
Uni-ControlNet (Global Control)✓✓✓47M0.5060.736
T2I-Adapter (Style)✓✓✓39M0.4850.648
ControlNet Shuffle✓✓✓361M0.4210.616
IP-Adapter✓X✓22M0.5880.828
LoRAAdapter (ours)✓✓✓16M0.6370.831
LoRAAdapter SDXL (ours)✓✓✓103M0.6490.849
", + "bbox": [ + 220, + 198, + 787, + 406 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "convolution $K_{0}$ , which reduces the channel count from $ch_{i}$ to the rank $r$ of the LoRA while keeping all else equal. Then, we apply our LoRA conditioning $\\phi(\\cdot|m(c))$ in this bottlenecked space and expand to the output channel count $ch_{o}$ using a pointwise convolution $K_{B}$ :", + "bbox": [ + 212, + 436, + 787, + 498 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nh = K _ {0} \\star x + K _ {B} \\star \\phi \\left(K _ {A} \\star x | m (c)\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 511, + 787, + 527 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A visual diagram of conditional LoRAs for convolutional layers is shown in Fig. 3b.", + "bbox": [ + 212, + 540, + 787, + 570 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 214, + 598, + 375, + 616 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 214, + 633, + 428, + 648 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Data We train LoRAdapter on a 40 million samples subset of COYO-700M [3] that only contains images with a short side of at least 512 pixels. Larger images are downscaled and center-cropped to 512 pixels. For text-conditioning, we use the standard prompts provided by the COYO-700M dataset.", + "bbox": [ + 212, + 662, + 787, + 723 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training and Implementation Details All experiments are based on Stable Diffusion 1.5 [33] unless noted otherwise and adapted with a conditional LoRA according to the given task. We train our adapter for 50,000 steps on 32 A100 GPUs with a global batch size of 256 and a learning rate of 0.0001. We use AdamW [22] as an optimizer and drop each conditioning with a probability of 0.05. This includes the original text conditioning.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "5. EXPERIMENTS", + "bbox": [ + 593, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/5c3b814cd534f4f9fac50b64ec730792feb230c6eda4958d73d2bada52960839.jpg", + "table_caption": [ + "Table 2: Comparison of a smaller and a larger version of our models vs. ControlNet." + ], + "table_footnote": [], + "table_body": "
ModelParamsMSE-d ↓FID ↓LPIPS ↓
ControlNet361M17.9817.6440.622
Uni-ControlNet361M17.5616.4530.601
T2I-Adapter39M24.7817.9410.636
LoRApapter-A (Ours)17M17.2716.8490.599
LoRApapter-B (Ours)32M15.3415.6700.572
", + "bbox": [ + 305, + 170, + 692, + 272 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For style conditioning, we experiment with both CLIP ViT-L/14 [30] by OpenAI and CLIP ViT-H/14 by OpenClip [16] as the image encoder but found that the ViT-H/14 tends to produce more visually appealing results. We still use ViT-L/14 for some ablation experiments as it results in a smaller model. Similar to [45], we use a single linear layer followed by layer normalization for the shared mapping network $m_S$ as it outperformed much larger mapping networks and showed faster convergence. This maps the pooled CLIP image token $c_{img}$ , keeping its original dimensionality $d_{img}$ . The local mapping network $m_l^{(i)}$ consists of two separate linear layers that predict $\\beta$ and $\\gamma$ in the correct dimensionality $r$ of the LoRA embedding. In total, the shared mapping network only has 1M parameters.", + "bbox": [ + 212, + 300, + 787, + 455 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For structure conditioning, the shared mapping network is a series of convolutional layers that spatially align the dimensions of the conditioning images with the dimensions that Stable Diffusion's U-Net expects which sums up to only 1.3M parameters. As the U-Net includes several down- and upsampling blocks, the mapping network outputs feature maps at various resolutions that match the specific U-Net blocks. Each individual LoRA only gets the single feature map that matches the spatial dimensionality of the U-Net block that it operates in. This feature map is then further processed with a single convolutional layer as a local mapping to align the number of channels with the rank $r$ of the LoRA embedding.", + "bbox": [ + 212, + 472, + 787, + 625 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Inference We sample with 50 DDIM [39] steps and apply classifier-free guidance (CFG) [13] with a scale of 7.5 to the prompt and style conditioning. We do not apply any CFG for structure conditioning. As with all LoRAs, we can also influence the strength of our LoRAAdapter by adjusting the LoRA scale $\\lambda$ , i.e. $h = W_0 + \\lambda BAx$ . We set $\\lambda = 1$ unless otherwise noted.", + "bbox": [ + 212, + 648, + 789, + 726 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 215, + 751, + 460, + 766 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To have a fair comparison with previous approaches, we follow previous literature on adapters [24,29,45] and evaluate our model on the validation set of COCO2017 [20] which contains 5,000 images. As every image comes with multiple prompts, we randomly select one prompt and use that for sampling.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Stracke et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Style Following [45], we sample four images for a given image and text prompt pair. To measure the effectiveness of the new style conditioning, we use CLIP ViT-L/14 to obtain the image embeddings of the generated and the prompt image and compute the cosine similarity between them (CLIP-I). In addition, we quantify how well the generated image still follows the text prompt by calculating the CLIP score [10] between the text prompt and the generated image (CLIP-T). In Tab. 1 we show results across different approaches that either train models from scratch or fine-tune existing models, as well as compare with recent state-of-the-art adapters.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/35c648de9a945bb5b524673399e74a904f952dd81ea1686cba5ba9c687264dbe.jpg", + "image_caption": [ + "Image Conditioning", + "Fig. 4: Samples from our method with style conditioning compared against other methods. We used an empty prompt and only conditioned on the image. We generally perform on par with IP-Adapter and outperform it on some samples. Note that the third image from the left is less degraded, and the third image from the right captures the mane of the horse better." + ], + "image_footnote": [], + "bbox": [ + 235, + 321, + 766, + 551 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "LoRAAdapter obtains state-of-the-art performance on CLIP-I and CLIP-T scores across all adapters, while also being more efficient (approx $27\\%$ reduction in the number of trainable parameters compared with IP-Adapter [45]). We attribute this boost in performance to our conditional LoRA approach which can efficiently capture detailed style conditioning in a more effective manner than previous adapter approaches. Interestingly, we find that LoRAAdapter even outperforms methods that train large models from scratch like unCLIP on CLIP-T score, which further supports our unified conditioning approach.", + "bbox": [ + 212, + 664, + 787, + 787 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Structure For evaluating structure control we follow the settings in UniControlNet [50] and sample a single image per condition. To evaluate how", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "5. EXPERIMENTS", + "bbox": [ + 593, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/fca4aba6f86f6b2ca31196e60cacf7011bc45d55cf135c5734c72beb7f1d13e3.jpg", + "image_caption": [ + "Prompt" + ], + "image_footnote": [], + "bbox": [ + 254, + 157, + 750, + 198 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/91e8b013dc4d4ac6c3bdddb7a64c94b452d5441b0572077f961169a354d40600.jpg", + "image_caption": [ + "ControlNet" + ], + "image_footnote": [], + "bbox": [ + 254, + 212, + 750, + 253 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/550fa223819a407c4315fb0a2f8686ab98644279da79abbae678e30dff1d1bf1.jpg", + "image_caption": [ + "T2I Adapter" + ], + "image_footnote": [], + "bbox": [ + 254, + 268, + 750, + 309 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/878531e1d79adca1b83b27f63bb45bf63699faa24594ccbdd76e15072ac13dea.jpg", + "image_caption": [ + "Uni-ControlNet" + ], + "image_footnote": [], + "bbox": [ + 254, + 323, + 750, + 364 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6aada045b0f7398da0c4a6bf76fcaf218ed7a8d8d5a4326fe84484669de68941.jpg", + "image_caption": [ + "Ours", + "Fig. 5: Samples from our method with structural conditioning compared against other methods. Note that for our method, especially compared with T2I Adapter, the details of the images are substantially more closely aligned with the depth prompt (see, e.g., the lamp in the background of the living room scene and the side table's legs, or the salad on the pizza)" + ], + "image_footnote": [], + "bbox": [ + 254, + 378, + 750, + 419 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "closely the generated image is following the provided structure map, we measure the cycle consistency between the predicted structure map of the original image $c = \\mathrm{encoder}(x)$ and the predicted structure map of the generated image $\\hat{c} = \\mathrm{encoder}(\\hat{x})$ . The exact metric used to measure the discrepancy depends on the type but $c$ and $\\hat{c}$ are always normalized to [0, 1] before any metrics computation. The differences in depth maps are measured by the Mean Squared Error between $c$ and $\\hat{c}$ (MSE-d). The SSIM (Structural Similarity) is used for HED conditioning. Additionally, we also compute the Learned Perceptual Image Patch Similarity (LPIPS) [49] to implicitly measure the fidelity of the control as well as the Frechet Inception Distance (FID) [11], which compares the distributions of intermediate features of a pre-trained network between generated and original images.", + "bbox": [ + 212, + 546, + 787, + 713 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 2 shows results across metrics for different approaches. In particular, adapters that focus on structure conditioning tend to use more parameters than ones used for style therefore we trained two versions of LoRAAdapter. One with 17M (roughly matching the number of parameters used for style conditioning) and 32M parameters, which is still less than T2I-Adapter [24]. LoRAAdapter obtains state-of-the-art performance across the board, even when reducing parameters by half compared to T2I-Adapter. These results show that LoRAAdapter is a very efficient approach for structure conditioning which can be scaled to obtain", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Stracke et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e5f552d84f1b0fd94fa6885b0c1d1c610c652093d5ab7caf34a0bf54fbb2c5dd.jpg", + "image_caption": [ + "Fig. 6: We show how the choice of the LoRA rank affects the generation quality. The rank refers to the cross-attention models in Table 3. Even the smallest model with only 2M parameters can meaningfully capture the content of the image albeit with minor losses in quality." + ], + "image_footnote": [], + "bbox": [ + 230, + 156, + 475, + 297 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/6e35920f2e8b9842e537706df2240c086104f8435d56be07ca5319cc86f217d7.jpg", + "table_caption": [ + "Table 3: Evaluation of different global conditional LoRA configurations. All models were trained using CLIP VIT-L/14 as an image encoder and evaluated on 5000 samples. Adapting the Cross-Attention layers yields the best performance. Even very low-rank LoRA conditioning can still produce good performance." + ], + "table_footnote": [], + "table_body": "
LayerRank ParametersCLIP-ICLIP-T
Cross Attention20821M0.8370.631
Cross Attention12813M0.8350.634
Cross Attention647M0.8290.637
Cross Attention324M0.8160.642
Cross Attention162M0.7990.649
Self Attention12813M0.8280.639
Cross & Self Attention12826M0.8420.629
Cross & Self Attention6413M0.8370.636
Convolutional12820M0.8060.647
", + "bbox": [ + 514, + 268, + 787, + 381 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "high-performant models. We refer the reader to the Appendix for additional results and comparisons.", + "bbox": [ + 212, + 440, + 784, + 470 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Qualitative Comparison", + "text_level": 1, + "bbox": [ + 214, + 501, + 457, + 517 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To show qualitative results for both structure and style conditioning we take images used for conditioning by previous approaches while also adding some conditioning images selected by us. Results are shown in Fig. 4 where we can see that LoRAAdapter better captures the style of the conditioning image compared with previous adapter approaches. We show additional comparison with more models on an extended set of samples in the Appendix.", + "bbox": [ + 212, + 534, + 787, + 625 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We also show samples for structure conditioning on depth maps predicted by MiDaS [32] in Fig 5. LoRAAdapter shows nice sample quality and superior adherence to the fine-grained depth structure compared to ControlNet or T2I-Adapter. We attribute this improvement to our choice of directly adapting the convolutions in the U-Net as opposed to an indirect adaption via skip connections as previous approaches have proposed [7, 24].", + "bbox": [ + 212, + 626, + 787, + 717 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Ablation experiments", + "text_level": 1, + "bbox": [ + 214, + 746, + 441, + 762 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To showcase the modularity of our approach, we train multiple configurations per condition and assess their performance. We study the influence of the LoRAs rank as well as the specific layer choice, which is further discussed for each conditioning type.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "5. EXPERIMENTS", + "bbox": [ + 593, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4ec3335ba28531de687e35d2bc3c692d39cedeaf7881bbd7ad14620ef5697bf3.jpg", + "image_caption": [ + "Fig. 7: Examples of using a style and structure LoRAAdapter jointly. We show both reconstruction tasks (first, second, and third columns) and cases where the style and structure image do not match perfectly (fourth, fifth, and sixth columns)." + ], + "image_footnote": [], + "bbox": [ + 279, + 142, + 723, + 309 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Rank and Layer choice As the modularity of our approach is a key differentiator from other methods, we perform an analysis of how the choice of adapted layers and the overall rank influence the performance in the context of style conditioning (Table 3). Overall the cross-attention layers perform best and show very good performance even with a very small rank. Convolutional and self-attention layers perform worse but may be better suited for other tasks such as conditioning on the style of an image instead of its semantics.", + "bbox": [ + 212, + 393, + 789, + 501 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Structure We adapt the convolutional layers and vary in what blocks we adapt them. Stable Diffusion 1.5 contains four upsampling and downsampling blocks and one middle block. Each block itself contains two ResNet [9] blocks with two convolutional layers each. We analyze two different configurations: Configuration A only adapts the first layer in every block, totaling nine adapted convolutional layers. Configuration B adapts the first convolutional layer in every ResNet block, i.e. 17 adapted layers in total (Table 2). Finally, we show examples of jointly using a style and structure conditioning on Fig. 7. We refer the reader to the Appendix for additional results on prompt combinations.", + "bbox": [ + 212, + 518, + 789, + 656 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 674, + 359, + 691 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper we have introduced LoRAAdapter, an approach to condition the image generation process of foundational text-to-image models. LoRAAdapter is an powerful and efficient approach for conditioning that unifies structure and style control using a conditional LoRA architecture that enables zero-shot generalization. Our approach enables fine-grained control over both structure and style obtaining state-of-the-art results and outperforming recent approaches while also optimizing a more compact number of parameters. LoRAAdapter presents progress towards acquiring fine-grained control over text-to-image models in an effective manner.", + "bbox": [ + 212, + 703, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Stracke et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This project has been supported by the German Federal Ministry for Economic Affairs and Climate Action within the project \"NXT GEN AI METHODS - Generative Methoden für Perzeption, Prädiktion und Planung\", the German Research Foundation (DFG) project 421703927, and the bidt project KLIMA-MEMES. The authors gratefully acknowledge the Gauss Center for Supercomputing for providing compute through the NIC on JUWELS at JSC and the HPC resources supplied by the Erlangen National High Performance Computing Center (NHR@FAU funded by DFG).", + "bbox": [ + 212, + 174, + 787, + 297 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 316, + 321, + 333 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. https://huggingface.co/lambdaabs/sd-image-variations-diffusers", + "2. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)", + "3. Byeon, M., Park, B., Kim, H., Lee, S., Baek, W., Kim, S.: Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset (2022)", + "4. Canny, J.: A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence PAMI-8(6), 679-698 (1986). https://doi.org/10.1109/TPAMI.1986.4767851", + "5. Cao, Z., Simon, T., Wei, S.E., Sheikh, Y.: Realtime multi-person 2d pose estimation using part affinity fields. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7291-7299 (2017)", + "6. Cheng, J., Xie, P., Xia, X., Li, J., Wu, J., Ren, Y., Li, H., Xiao, X., Zheng, M., Fu, L.: Resadapter: Domain consistent resolution adapter for diffusion models (2024)", + "7. Denis Zavadski, J.F.F., Rother, C.: Controlnet-xs: Designing an efficient and effective architecture for controlling text-to-image diffusion models (2023)", + "8. Gandikota, R., Materzynska, J., Zhou, T., Torralba, A., Bau, D.: Concept sliders: Lora adaptors for precise control in diffusion models (2023)", + "9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "0. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021)", + "11. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)", + "12. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)", + "13. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)", + "14. Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2021)", + "15. Huang, L., Chen, D., Liu, Y., Shen, Y., Zhao, D., Zhou, J.: Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint arXiv:2302.09778 (2023)" + ], + "bbox": [ + 218, + 347, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "6. CONCLUSION", + "bbox": [ + 602, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip. https://github.com/mlfoundations/open Clip (2021)", + "17. Jiang, Z., Mao, C., Pan, Y., Han, Z., Zhang, J.: Scedit: Efficient and controllable image diffusion generation via skip connection editing. arXiv preprint arXiv:2312.11392 (2023)", + "18. Kopiczko, D.J., Blankevoort, T., Asano, Y.M.: Vera: Vector-based random matrix adaptation (2024)", + "19. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22511-22521 (June 2023)", + "20. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "21. Liu, S.Y., Wang, C.Y., Yin, H., Molchanov, P., Wang, Y.C.F., Cheng, K.T., Chen, M.H.: Dora: Weight-decomposed low-rank adaptation (2024)", + "22. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)", + "23. Luo, S., Tan, Y., Patil, S., Gu, D., von Platen, P., Passos, A., Huang, L., Li, J., Zhao, H.: Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556 (2023)", + "24. Mou, C., Wang, X., Xie, L., Wu, Y., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023)", + "25. Park, T., Liu, M.Y., Wang, T.C., Zhu, J.Y.: Semantic image synthesis with spatially-adaptive normalization. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2332-2341. IEEE (2019)", + "26. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4195-4205 (2023)", + "27. Pernias, P., Rampas, D., Richter, M.L., Pal, C., Aubreville, M.: Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=gU58d5QeGv", + "28. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis (2023)", + "29. Qin, C., Zhang, S., Yu, N., Feng, Y., Yang, X., Zhou, Y., Wang, H., Niebles, J.C., Xiong, C., Savarese, S., et al.: Unicontrol: A unified diffusion model for controllable visual generation in the wild. arXiv preprint arXiv:2305.11147 (2023)", + "30. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "31. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + ], + "bbox": [ + 215, + 146, + 787, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Stracke et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence 44(3), 1623-1637 (2020)", + "33. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "34. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)", + "35. Ruiz, N., Li, Y., Jampani, V., Wei, W., Hou, T., Pritch, Y., Wadhwa, N., Rubinstein, M., Aberman, K.: Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949 (2023)", + "36. Ryu, S.: Low-rank adaptation for fast text-to-image diffusion fine-tuning, https://github.com/cloneofsimo/lora", + "37. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)", + "38. Shah, V., Ruiz, N., Cole, F., Lu, E., Lazebnik, S., Li, Y., Jampani, V.:Ziplora: Any subject in any style by effectively merging loras (2023)", + "39. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)", + "40. Witteveen, S., Andrews, M.: Investigating prompt engineering in diffusion models. arXiv preprint arXiv:2211.15462 (2022)", + "41. Xie, S., Tu, Z.: Holistically-nested edge detection. In: Proceedings of the IEEE international conference on computer vision. pp. 1395-1403 (2015)", + "42. Xu, X., Guo, J., Wang, Z., Huang, G., Essa, I., Shi, H.: Prompt-free diffusion: Taking\" text\" out of text-to-image diffusion models. arXiv preprint arXiv:2305.16223 (2023)", + "43. Xu, X., Wang, Z., Zhang, E., Wang, K., Shi, H.: Versatile diffusion: Text, images and variations all in one diffusion model. arXiv preprint arXiv:2211.08332 (2022)", + "44. Xue, Z., Song, G., Guo, Q., Liu, B., Zong, Z., Liu, Y., Luo, P.: Raphael: Text-to-image generation via large mixture of diffusion paths. arXiv preprint arXiv:2305.18295 (2023)", + "45. Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models (2023)", + "46. YEH, S.Y., Hsieh, Y.G., Gao, Z., Yang, B.B.W., Oh, G., Gong, Y.: Navigating text-to-image customization: From lyCORIS fine-tuning to model evaluation. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=wfzXa8e783", + "47. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models (2023)", + "48. Zhang, Q., Chen, M., Bukharin, A., Karampatziakis, N., He, P., Cheng, Y., Chen, W., Zhao, T.: Adalora: Adaptive budget allocation for parameter-efficient fine-tuning (2023)", + "49. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "6. CONCLUSION", + "bbox": [ + 602, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "ref_text", + "text": "50. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Unictrlnet: All-in-one control to text-to-image diffusion models. Advances in Neural Information Processing Systems (2023)", + "bbox": [ + 217, + 146, + 787, + 189 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Stracke et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/CTRLorALTer_ Conditional LoRAdapter for Efficient 0-Shot Control & Altering of T2I Models/d556ef6f-c8cf-4800-af84-90cf37eef09f_model.json b/2024/CTRLorALTer_ Conditional LoRAdapter for Efficient 0-Shot Control & Altering of T2I Models/d556ef6f-c8cf-4800-af84-90cf37eef09f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..64b3c9327b8394ad36b5e7a9667d4d7b37d8cd13 --- /dev/null +++ b/2024/CTRLorALTer_ Conditional LoRAdapter for Efficient 0-Shot Control & Altering of T2I Models/d556ef6f-c8cf-4800-af84-90cf37eef09f_model.json @@ -0,0 +1,2469 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.146, + 0.786, + 0.19 + ], + "angle": 0, + "content": "CTRLorALTer: Conditional LoRAapter for Efficient 0-Shot Control & Altering of T2I Models" + }, + { + "type": "text", + "bbox": [ + 0.283, + 0.216, + 0.721, + 0.248 + ], + "angle": 0, + "content": "Nick Stracke1, Stefan Andreas Baumann1, Joshua Susskind2, Miguel Angel Bautista2, and Björn Ommer1" + }, + { + "type": "text", + "bbox": [ + 0.377, + 0.258, + 0.625, + 0.272 + ], + "angle": 0, + "content": "CompVis @ LMU Munich, MCML" + }, + { + "type": "text", + "bbox": [ + 0.474, + 0.273, + 0.531, + 0.287 + ], + "angle": 0, + "content": "2 Apple" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.325, + 0.74, + 0.476 + ], + "angle": 0, + "content": "Abstract. Text-to-image generative models have become a prominent and powerful tool that excels at generating high-resolution realistic images. However, guiding the generative process of these models to take into account detailed forms of conditioning reflecting style and/or structure information remains an open problem. In this paper, we present LoRAAdapter, an approach that unifies both style and structure conditioning under the same formulation using a novel conditional LoRA block that enables zero-shot control. LoRAAdapter is an efficient and powerful approach to condition text-to-image diffusion models, which enables fine-grained control conditioning during generation and outperforms recent state-of-the-art approaches." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.478, + 0.582, + 0.491 + ], + "angle": 0, + "content": "Project page: compvis.github.io/LoRAAdapter/" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.504, + 0.727, + 0.518 + ], + "angle": 0, + "content": "Keywords: T2I Models \\(\\cdot\\) Diffusion Models \\(\\cdot\\) Controllable Generation" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.567, + 0.375, + 0.582 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.598, + 0.787, + 0.794 + ], + "angle": 0, + "content": "Text-to-image models have become a foundational component of Computer Vision, especially with the introduction of large-scale diffusion models like Stable Diffusion (SD) [33], DALL-E 2 [31], Imagen [37], RAPHAEL [44] and, eDiff-I [2]. These models have empowered users to create realistic images from textual prompts, although crafting effective prompts is a complex task in itself due to the nuanced prompt engineering required [40] to effectively condition the image generation process. To tackle this problem, prior approaches such as SD Image Variations [1], Stable unCLIP [33], and Composer [15], have attempted to perform conditioning using image prompts to enhance the expressivity of the conditioning mechanism giving users more flexibility during generation. In practice, this family of approaches requires either training from scratch or fully fine-tuning large text-conditioned model weights, incurring high computational costs and requiring large datasets." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To reduce training cost and time, recent approaches like ControlNet [47], T2I-Adapter [24], Uni-ControlNet [50] or IP-Adapters [45] have focused on designing adapters that keep the base model frozen and only introduce a limit number of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Stracke et al." + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.151, + 0.745, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.38, + 0.79, + 0.437 + ], + "angle": 0, + "content": "Fig. 1: LoRAAdapter allows structure and style control of the image generation process of text-to-image models in a zero-shot manner. Our approach enables powerful fine-grained and efficient unified control over both structure and style conditioning using conditional LoRA blocks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.54 + ], + "angle": 0, + "content": "new layers. This not only reduces the number of trainable parameters but also allows training on much smaller datasets because the frozen base model cannot suffer from catastrophic forgetting. These adapters are typically handcrafted to a specific model architecture and conditioning modality. They can be roughly assigned into one of two groups:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.546, + 0.788, + 0.606 + ], + "angle": 0, + "content": "- Adapters for pixel-level fine-grained conditioning over the generated image. This is often referred to as structural or local conditioning since it allows influencing the generated image on a per-pixel level by e.g. providing a depth map and effectively setting the layout of the image [7,17,24,29,47,50]" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.606, + 0.788, + 0.667 + ], + "angle": 0, + "content": "- Adapters for style or global conditioning perform tasks such as image-to-image translation similar to unCLIP [15, 19, 33, 43]. Here, the focus is not on fine-grained control but instead on generating an image similar to the prompt image, where similarity is defined by sharing the same style or semantics." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.546, + 0.788, + 0.667 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Empirically, adapters always favor one type of conditioning over the other. For example, both Uni-ControlNet [50] and T2I-Adapter [24] have attempted style conditioning but their performance is not comparable to leading style approaches [45]. Conversely, style adapters require ControlNet or T2I-Adapter for structure conditioning. However, ControlNet is a large adapter that copies the entire encoder of Stable Diffusion's U-Net which significantly increases compute and thus inference time. Our goal is to find a unified approach that is capable of both structure and style conditioning, providing complete control over the generated image. Formulating such a unified approach for conditioning on global controls like style and on local controls like structure, in an efficient and generic manner remains a key open problem." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.578, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "2. RELATED WORK" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.312 + ], + "angle": 0, + "content": "To address this challenging problem, we introduce LoRAAdapter (see Fig. 1), a unified approach for incorporating images to condition both the style and structure of the generated images. LoRAAdapter takes inspiration from Low-Rank-Adaptations (LoRAs) [14], which have shown remarkable performance for various tasks in diffusion models, such as distillation or learning specific concepts [23,35]. In particular, LoRAs low-rank property naturally regularizes the conditioning while offering unparalleled flexibility as any layer can be adapted in an architecture-agnostic way. Thus far, however, LoRAs have been implemented as fixed adaptations, being independent of the input. This means layer adaptation cannot change at test time for a given conditioning and, therefore, cannot be used for zero-shot generalization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.435 + ], + "angle": 0, + "content": "LoRAAdapter is a novel approach to adding conditional information to LoRAs, enabling zero-shot generalization and making them applicable for both structure and style and possibly many other conditioning types. LoRAAdapter is compact and efficient, e.g., optimizing 16M parameters vs the 22M of IP-Adapters [45] or 361M of ControlNet [47], while outperforming recent adapter approaches [45] and even approaches that train models from scratch (see Tab. 1). LoRAAdapter trains LoRA [14] modules for different blocks across the base network and is efficient both during training and inference. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.447, + 0.784, + 0.475 + ], + "angle": 0, + "content": "- We propose LoRAAdapter, a generic approach to train conditional LoRAs that is agnostic to model architecture and conditioning modality." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.478, + 0.784, + 0.505 + ], + "angle": 0, + "content": "- We implement LoRAAdapter for Stable Diffusion, offering a unified conditioning mechanism for both style and structure, enabling zero-shot conditioning." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.509, + 0.784, + 0.535 + ], + "angle": 0, + "content": "- We show the effectiveness of this approach, outperforming dedicated adapters for either style and structure on various metrics." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.447, + 0.784, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.562, + 0.388, + 0.579 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.787, + 0.625 + ], + "angle": 0, + "content": "We now review existing adapter approaches for style and structure and discuss the current role of LoRAs in diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.648, + 0.424, + 0.664 + ], + "angle": 0, + "content": "2.1 Structure Adapters" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Structure adapters are specifically designed to operate on various types of local conditioning modalities such as Depth, HED [41], Canny Edges [4], Scribbles, or key poses [5]. The goal of these adapters is to spatially align the generated image with the provided condition. One of the most prominent approaches to tackle this problem is ControlNet [47] for Stable Diffusion. ControlNet creates a copy of the U-Net's encoder and combines the skip connections additively with the original skip connections, which then get fed into the original decoder. While obtaining great performance, ControlNet adds a lot of computational overhead as the forward pass includes a second encoder which can increase inference speed by up to \\(50\\%\\). Another consequence of the large capacity is that ControlNet tends to interpret the structure map directly. This can be desirable when sampling" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "Stracke et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "images without other conditioning via the text prompt or other adapters, but can also lead to entanglement of structure and style." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.344 + ], + "angle": 0, + "content": "Following a similar idea, T2I-Adapter [24] also modifies skip connections while using a much smaller encoder network. To extend conditioning to multiple modalities, Uni-ControlNet [50] was proposed as an alternative solution to having to train multiple separate and modality-specific ControlNets. Uni-ControlNet trains a single model for multiple conditions by concatenating the structure maps along the channel dimension and using the concatenated tensor as input. Uni-ControlNet uses SPADE [25] to combine skip-connections instead of the simple addition used by ControlNet. SCEdit [17] proposed a similar but more efficient solution for conditioning on multiple modalities by integrating a small tuner network (SC-Tuner) between skip-connections which removes the need for a separate encoder network akin to ControlNet." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.365, + 0.386, + 0.38 + ], + "angle": 0, + "content": "2.2 Style Adapters" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.788, + 0.478 + ], + "angle": 0, + "content": "Style adapters are used as an alternative to unCLIP [33] like models, which are directly trained to invert CLIP image embeddings. The main advantage of using a style adapter as opposed to an unCLIP model is that the original text conditioning is preserved, enabling additional control over the generated images. Additionally, all recently released large open-source diffusion models only offered text conditioning, making an adapter approach necessary [27, 28]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.788, + 0.597 + ], + "angle": 0, + "content": "IP-Adapter [45] adds a decoupled cross-attention layer that operates on four image tokens instead of text tokens. The resulting activations of this decoupled cross-attention layer are scaled and added to the original cross-attention activations. This results in a lightweight adapter that integrates well with potential text conditioning and generates images faithful to the image prompt. In contrast, SeeCoder [42] trains an encoder with 2D spatial embedding to convert image tokens to text tokens and completely replaces the standard text tokens with their tokens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.6, + 0.788, + 0.691 + ], + "angle": 0, + "content": "Structure adapters have also attempted to include style conditioning. UniControlNet [50] includes a separate mapping network to convert a pooled CLIP image token to four CLIP text tokens, which they concatenate to the original text embeddings. ControlNet-shuffle [47] shuffles an RGB image and uses that as conditioning similar to its structure conditioning. However, the performance of these approaches does not compare favorably to dedicated style adapters." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.711, + 0.488, + 0.725 + ], + "angle": 0, + "content": "2.3 LoRAs in Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.842 + ], + "angle": 0, + "content": "While LoRAs were initially introduced as an efficient alternative to fine-tuning large language models [14], they were quickly adopted for diffusion models [36] as a relatively efficient and cheap option to adjust their generation capabilities. Combining LoRAs with methods such as Dreambooth [34] performs on par with a full fine-tuning approach while only introducing a fraction of trainable parameters. LoRAs are commonly trained on a small set of images with the goal of capturing a specific style instead of a subject. ZipLoRA [38] proposes a merging algorithm to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.582, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "3. PRELIMINARIES" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "combine multiple LoRAs gracefully by minimizing the cosine similarity of LoRAs that adapt the same weight matrix. LyCORIS [46] trained an entire library of specific LoRAs in diffusion models with various benefits. [23] showed that LoRAs are even powerful enough for more complex objectives such as consistency distillation. Finally, Concept Sliders [8] use LoRAs to train disentangled edit directions to edit images. They scale the LoRA in the negative and positive directions to either increase or decrease a given attribute." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.277, + 0.382, + 0.293 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.309, + 0.338, + 0.323 + ], + "angle": 0, + "content": "3.1 Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.788, + 0.427 + ], + "angle": 0, + "content": "Diffusion models represent a category of generative models characterized by a two-stage process: a forward diffusion that incrementally introduces Gaussian noise over a series of \\(T\\) steps, and a reverse denoising process that reconstructs samples from this noise. In text-to-image models, diffusion can be directed by supplementary inputs like text descriptions. The primary training objective, \\(\\epsilon_{\\theta}\\), aimed at noise prediction, simplifies the variational lower bound as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.289, + 0.437, + 0.788, + 0.472 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {s i m p l e}} = \\mathbb {E} _ {\\boldsymbol {x} _ {0}, \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}), \\boldsymbol {c}, t} \\| \\boldsymbol {\\epsilon} - \\epsilon_ {\\theta} \\big (\\underbrace {\\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}} _ {\\boldsymbol {x} _ {t}}, \\boldsymbol {c}, t \\big) \\| ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.788, + 0.509 + ], + "angle": 0, + "content": "where \\(\\pmb{x}_0\\) is the original data influenced by condition \\(\\pmb{c}\\), \\(t\\) is the diffusion timestep, and \\(\\pmb{x}_t\\) represents the data at step \\(t\\), with \\(\\bar{\\alpha}_t\\) dictating the noise level." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.6 + ], + "angle": 0, + "content": "During sampling, an initial pure Gaussian latent \\( \\pmb{x}_T \\) is iteratively denoised by the diffusion model to obtain a sample \\( \\pmb{x}_0 \\) from the modeled (conditional) distribution. Conditional models often employ classifier guidance to fine-tune the balance between accuracy and diversity, with classifier-free guidance [12] serving as a preferred alternative that merges conditional and unconditional noise predictions during training and sampling:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.614, + 0.787, + 0.629 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {\\epsilon}} _ {\\theta} \\left(\\boldsymbol {x} _ {t}, \\boldsymbol {c}, t\\right) = w \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {x} _ {t}, \\boldsymbol {c}, t\\right) + (1 - w) \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {x} _ {t}, t\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.788, + 0.686 + ], + "angle": 0, + "content": "where \\( w \\) modulates the adherence to condition \\( c \\). This technique is pivotal for text-to-image models, enhancing the alignment between generated images and text prompts." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.709, + 0.515, + 0.725 + ], + "angle": 0, + "content": "3.2 Low-Rank Adaptation (LoRA)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Low-Rank Adaptation (LoRA) [14] is a method for finetuning pre-trained foundation models by keeping the original weight matrices frozen and instead adding a new set of low-rank weight matrices. While this method was originally proposed in the context of large language models, it has been recently adopted in the domain of diffusion models [6,8,23]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Specifically, the original weight matrix of a layer in the base model \\( W_0 \\in \\mathbb{R}^{d \\times k} \\) is kept frozen and is adapted by a low-rank matrix \\( W + \\Delta W = W + BA \\) with" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "Stracke et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.146, + 0.788, + 0.178 + ], + "angle": 0, + "content": "\\(B \\in \\mathbb{R}^{d \\times r}\\) and \\(A \\in \\mathbb{R}^{r \\times k}\\). Crucially, \\(r\\) is chosen following \\(r \\ll \\min(d, k)\\). The forward pass for the adapted layer is then:" + }, + { + "type": "equation", + "bbox": [ + 0.38, + 0.185, + 0.788, + 0.202 + ], + "angle": 0, + "content": "\\[\nh = W _ {0} x + \\Delta W x = W _ {0} x + B A x. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.209, + 0.789, + 0.269 + ], + "angle": 0, + "content": "Typically, \\( B \\) is zero-initialized, so the LoRA does not affect the model initially. This yields a generic approach, as we can freely choose the rank \\( r \\) and the weights according to the given task. This happens on a per-layer basis, i.e., if we adapt \\( n \\) layers, we will have \\( nA \\) and \\( B \\) matrices." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.289, + 0.331, + 0.306 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.318, + 0.789, + 0.424 + ], + "angle": 0, + "content": "There are two kinds of approaches for conditioning text-to-image diffusion models: style or global conditioning [42, 45], and structure or local conditioning [17, 24, 29, 50]. These approaches are specifically tailored to a specific conditioning objective and therefore missing a holistic approach that can efficiently learn both style and structure. We now describe LoRAAdapter, an efficient, holistic, and customizable method for introducing both style and structure conditioning into text-to-image diffusion models." + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.453, + 0.648, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.638, + 0.789, + 0.74 + ], + "angle": 0, + "content": "Fig. 2: Overview of the proposed conditional LoRA block. The original weight matrix \\( W_0^{(i)} \\) is frozen while all other layers are trained. \\( \\phi \\) is an affine transformation that operates on the low-dimensional embedding \\( A^{(i)}x \\) and introduces the conditioning. The local mapper network \\( m_L^{(i)} \\) predicts the scale and shift parameters \\( \\beta, \\gamma \\) for the affine transformation. Typically, we set \\( m_L^{(i)} \\) to be a small network. If complex transformations are required to map the conditioning \\( c \\), this happens in \\( m_S \\) since it is shared across all adapted layers." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.424, + 0.803 + ], + "angle": 0, + "content": "4.1 Conditional LoRAs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "LoRAs have successfully enabled adaptation of diffusion models across a range of architectures [6,23,35] to steer their generation process. In this setting, LoRAs" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.637, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "4. METHOD" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "have been considered as free parameters and optimized on a training set of images exhibiting a specific characteristic." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.312 + ], + "angle": 0, + "content": "Instead of learning a single LoRA to apply one specific modification, e.g., changing the style of the generated image to a painting or learning to represent one specific subject [34,35], we propose a LoRA-based conditioning mechanism whose behavior changes based on conditioning provided at inference time, enabling zero-shot generalization. While previous methods depend on specific aspects of the architectures of the models they are applied to [45, 47], using LoRAs for conditioning enables us to efficiently adapt foundation models for conditional tasks in an architecture-agnostic way. Our LoRA-based conditioning mechanism is shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.787, + 0.373 + ], + "angle": 0, + "content": "Starting from the standard LoRA setup as in Eq. (3), we propose implementing conditioning of the LoRA by applying a transformation \\(\\phi(Ax|m(c))\\) to the low-dimensional intermediate embedding in the LoRA that introduces conditional behavior based on the conditioning \\(c\\):" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.385, + 0.787, + 0.4 + ], + "angle": 0, + "content": "\\[\nh = W _ {0} x + \\Delta W x = W _ {0} x + B \\phi (A x | m (c)). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.787, + 0.514 + ], + "angle": 0, + "content": "This limits the conditioning to apply only to a low-rank subspace of the original vector space the weight matrix operates in. This is crucial for the generalization capabilities of LoRAs, as it introduces regularization and thus limits the adaptation to only focus on relevant aspects instead of spurious correlations from the dataset it was trained on, which results in more efficient training. By introducing our conditional adaptation of the model's behavior in such a low-rank space, our method can also benefit from those advantages." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.796, + 0.56 + ], + "angle": 0, + "content": "In our framework for conditional LoRAs, we can theoretically choose \\(\\phi(Ax|m(c))\\) to be arbitrary, non-linear transformations, but we find that a simple affine transformation is sufficiently expressive for a wide range of applications:" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.572, + 0.787, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\phi (A x \\mid m (c)) = \\gamma_ {\\phi} (m (c)) \\odot A x + \\beta_ {\\phi} (m (c)), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.786, + 0.625 + ], + "angle": 0, + "content": "with \\(\\odot\\) denoting an elementwise (Hadamard) product and \\(\\gamma\\) and \\(\\beta\\) referring to the scale and shift factors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.787, + 0.734 + ], + "angle": 0, + "content": "To predict the \\(\\gamma\\) and \\(\\beta\\) from the conditioning \\(c\\), we utilize a mapping network. This is generally a small neural network that we separate into a shared part \\(m_{S}(\\cdot)\\) and a layer-specific part \\(m_{L}^{(i)}(\\cdot)\\) that is individually learned for each adapted layer. For both parts, we generally use very small neural networks, which is in stark contrast to other works such as ControlNet [47] that uses a full copy of the U-Net's encoder (see Tab. 1). For brevity, we will always refer to both parts simply as \\(m(\\cdot)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "This general formulation directly enables both local and global conditioning, as it can be applied to the convolutional and attention layers in a standard diffusion model. For local conditioning, the mapping network can supply spatially dependent LoRA modulations. At the same time, it can also be used to supply global conditioning by introducing non-spatially dependent information, such as by adapting the cross-attention projections. Concurrent works [18, 21, 48] also introduced improvements to the basic LoRA architecture. As LoRAAdapter does" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "Stracke et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "not depend on a specific implementation of LoRAs but just on the low-rank intermediate sub-space, these improvements can be directly transferred to our model." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.216, + 0.544, + 0.232 + ], + "angle": 0, + "content": "4.2 A Unified Conditioning Approach" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.242, + 0.786, + 0.273 + ], + "angle": 0, + "content": "In this section, we provide implementation details for applying our conditional LoRAs to the layer types commonly used in large-scale diffusion models." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.298, + 0.479, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.367, + 0.478, + 0.4 + ], + "angle": 0, + "content": "(a) In attention layers, we apply an affine transformation on the \\(K\\) and \\(V\\) projection matrix per channel for every token." + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.304, + 0.782, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.367, + 0.789, + 0.4 + ], + "angle": 0, + "content": "(b) To add structural conditioning to convolutional layers, we perform an affine transformation for every element similar to [25]." + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.411, + 0.784, + 0.425 + ], + "angle": 0, + "content": "Fig. 3: Visualization of implementations of our conditional LoRAs for specific layers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.789, + 0.555 + ], + "angle": 0, + "content": "Attention Layers Many diffusion models, such as Stable Diffusion \\(\\{1,2,\\mathrm{XL}\\}\\) [28,33] or Imagen [37] include self- and cross-attention layers that operate on image (and text) tokens, with transformer-based diffusion models such as [26] even purely relying on them for spatial computations. The output \\(h^\\prime\\) of the attention layers is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.329, + 0.565, + 0.788, + 0.602 + ], + "angle": 0, + "content": "\\[\nh ^ {\\prime} = \\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) V \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.61, + 0.788, + 0.687 + ], + "angle": 0, + "content": "with the projections \\( Q = hW_{q} \\), \\( K = hW_{k} \\) or \\( K = c_{t}W_{k} \\), and \\( V = hW_{v} \\) or \\( V = c_{t}W_{v} \\), depending on whether it is self- or cross-attention, and the text tokens \\( c_{t} \\). Here, we apply our conditional LoRA on the linear projections \\( W_{k} \\) and \\( W_{v} \\) based on Equation 4. We show a diagram of conditional LoRAs for attention layers in Fig. 3a." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.709, + 0.787, + 0.74 + ], + "angle": 0, + "content": "Convolutional Layers For convolutional layers, which form the main backbone of many diffusion models [27,33], we adapt the convolutional layer" + }, + { + "type": "equation", + "bbox": [ + 0.457, + 0.752, + 0.788, + 0.769 + ], + "angle": 0, + "content": "\\[\nh = K _ {0} \\star x, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "with the 2D convolution kernel \\(K\\) having input channel count \\(ch_{i}\\) and output channel count \\(ch_{o}\\), in a similar manner as linear projections. However, in this setting we work on convolution kernels instead of matrices: we perform one convolution \\(K_{A}\\) with the same kernel size, stride, and padding as the original" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.594, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "5. EXPERIMENTS" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.188 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison for style conditioning of our proposed LoRAAdapter with other methods on the COCO validation set with four samples for every image. The best results are in bold (adapted from [45])." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.199, + 0.789, + 0.407 + ], + "angle": 0, + "content": "
Style MethodReusable to custom modelsNative structure controlMultimodal promptsTrainable parametersCLIP-T ↑CLIP-I ↑
Training from scratch
Open unCLIPXXX893M0.6080.858
Kandinsky-2-1XXX1229M0.5990.855
Versatile DiffusionXX✓860M0.5870.830
Fine-tunining from text-to-image model
SD Image VariationsXXX860M0.5480.760
SD unCLIPXXX870M0.5840.810
Adapters
Uni-ControlNet (Global Control)✓✓✓47M0.5060.736
T2I-Adapter (Style)✓✓✓39M0.4850.648
ControlNet Shuffle✓✓✓361M0.4210.616
IP-Adapter✓X✓22M0.5880.828
LoRAAdapter (ours)✓✓✓16M0.6370.831
LoRAAdapter SDXL (ours)✓✓✓103M0.6490.849
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.788, + 0.499 + ], + "angle": 0, + "content": "convolution \\( K_{0} \\), which reduces the channel count from \\( ch_{i} \\) to the rank \\( r \\) of the LoRA while keeping all else equal. Then, we apply our LoRA conditioning \\( \\phi(\\cdot|m(c)) \\) in this bottlenecked space and expand to the output channel count \\( ch_{o} \\) using a pointwise convolution \\( K_{B} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.372, + 0.512, + 0.788, + 0.529 + ], + "angle": 0, + "content": "\\[\nh = K _ {0} \\star x + K _ {B} \\star \\phi \\left(K _ {A} \\star x | m (c)\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.541, + 0.788, + 0.571 + ], + "angle": 0, + "content": "A visual diagram of conditional LoRAs for convolutional layers is shown in Fig. 3b." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.599, + 0.377, + 0.617 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.634, + 0.429, + 0.65 + ], + "angle": 0, + "content": "5.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.663, + 0.788, + 0.724 + ], + "angle": 0, + "content": "Data We train LoRAdapter on a 40 million samples subset of COYO-700M [3] that only contains images with a short side of at least 512 pixels. Larger images are downscaled and center-cropped to 512 pixels. For text-conditioning, we use the standard prompts provided by the COYO-700M dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Training and Implementation Details All experiments are based on Stable Diffusion 1.5 [33] unless noted otherwise and adapted with a conditional LoRA according to the given task. We train our adapter for 50,000 steps on 32 A100 GPUs with a global batch size of 256 and a learning rate of 0.0001. We use AdamW [22] as an optimizer and drop each conditioning with a probability of 0.05. This includes the original text conditioning." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Stracke et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.784, + 0.16 + ], + "angle": 0, + "content": "Table 2: Comparison of a smaller and a larger version of our models vs. ControlNet." + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.171, + 0.694, + 0.273 + ], + "angle": 0, + "content": "
ModelParamsMSE-d ↓FID ↓LPIPS ↓
ControlNet361M17.9817.6440.622
Uni-ControlNet361M17.5616.4530.601
T2I-Adapter39M24.7817.9410.636
LoRApapter-A (Ours)17M17.2716.8490.599
LoRApapter-B (Ours)32M15.3415.6700.572
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.301, + 0.788, + 0.456 + ], + "angle": 0, + "content": "For style conditioning, we experiment with both CLIP ViT-L/14 [30] by OpenAI and CLIP ViT-H/14 by OpenClip [16] as the image encoder but found that the ViT-H/14 tends to produce more visually appealing results. We still use ViT-L/14 for some ablation experiments as it results in a smaller model. Similar to [45], we use a single linear layer followed by layer normalization for the shared mapping network \\( m_S \\) as it outperformed much larger mapping networks and showed faster convergence. This maps the pooled CLIP image token \\( c_{img} \\), keeping its original dimensionality \\( d_{img} \\). The local mapping network \\( m_l^{(i)} \\) consists of two separate linear layers that predict \\( \\beta \\) and \\( \\gamma \\) in the correct dimensionality \\( r \\) of the LoRA embedding. In total, the shared mapping network only has 1M parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.789, + 0.625 + ], + "angle": 0, + "content": "For structure conditioning, the shared mapping network is a series of convolutional layers that spatially align the dimensions of the conditioning images with the dimensions that Stable Diffusion's U-Net expects which sums up to only 1.3M parameters. As the U-Net includes several down- and upsampling blocks, the mapping network outputs feature maps at various resolutions that match the specific U-Net blocks. Each individual LoRA only gets the single feature map that matches the spatial dimensionality of the U-Net block that it operates in. This feature map is then further processed with a single convolutional layer as a local mapping to align the number of channels with the rank \\( r \\) of the LoRA embedding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.65, + 0.79, + 0.727 + ], + "angle": 0, + "content": "Inference We sample with 50 DDIM [39] steps and apply classifier-free guidance (CFG) [13] with a scale of 7.5 to the prompt and style conditioning. We do not apply any CFG for structure conditioning. As with all LoRAs, we can also influence the strength of our LoRAAdapter by adjusting the LoRA scale \\(\\lambda\\), i.e. \\(h = W_0 + \\lambda BAx\\). We set \\(\\lambda = 1\\) unless otherwise noted." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.752, + 0.461, + 0.767 + ], + "angle": 0, + "content": "5.2 Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To have a fair comparison with previous approaches, we follow previous literature on adapters [24,29,45] and evaluate our model on the validation set of COCO2017 [20] which contains 5,000 images. As every image comes with multiple prompts, we randomly select one prompt and use that for sampling." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.594, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "5. EXPERIMENTS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "Style Following [45], we sample four images for a given image and text prompt pair. To measure the effectiveness of the new style conditioning, we use CLIP ViT-L/14 to obtain the image embeddings of the generated and the prompt image and compute the cosine similarity between them (CLIP-I). In addition, we quantify how well the generated image still follows the text prompt by calculating the CLIP score [10] between the text prompt and the generated image (CLIP-T). In Tab. 1 we show results across different approaches that either train models from scratch or fine-tune existing models, as well as compare with recent state-of-the-art adapters." + }, + { + "type": "image_caption", + "bbox": [ + 0.437, + 0.311, + 0.568, + 0.324 + ], + "angle": 0, + "content": "Image Conditioning" + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.323, + 0.767, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.621 + ], + "angle": 0, + "content": "Fig. 4: Samples from our method with style conditioning compared against other methods. We used an empty prompt and only conditioned on the image. We generally perform on par with IP-Adapter and outperform it on some samples. Note that the third image from the left is less degraded, and the third image from the right captures the mane of the horse better." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.665, + 0.788, + 0.788 + ], + "angle": 0, + "content": "LoRAAdapter obtains state-of-the-art performance on CLIP-I and CLIP-T scores across all adapters, while also being more efficient (approx \\(27\\%\\) reduction in the number of trainable parameters compared with IP-Adapter [45]). We attribute this boost in performance to our conditional LoRA approach which can efficiently capture detailed style conditioning in a more effective manner than previous adapter approaches. Interestingly, we find that LoRAAdapter even outperforms methods that train large models from scratch like unCLIP on CLIP-T score, which further supports our unified conditioning approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Structure For evaluating structure control we follow the settings in UniControlNet [50] and sample a single image per condition. To evaluate how" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Stracke et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.478, + 0.146, + 0.526, + 0.156 + ], + "angle": 0, + "content": "Prompt" + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.158, + 0.75, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.201, + 0.536, + 0.211 + ], + "angle": 0, + "content": "ControlNet" + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.213, + 0.75, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.463, + 0.256, + 0.54, + 0.268 + ], + "angle": 0, + "content": "T2I Adapter" + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.269, + 0.75, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.312, + 0.549, + 0.323 + ], + "angle": 0, + "content": "Uni-ControlNet" + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.324, + 0.75, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.487, + 0.368, + 0.517, + 0.378 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.379, + 0.75, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.422, + 0.789, + 0.493 + ], + "angle": 0, + "content": "Fig. 5: Samples from our method with structural conditioning compared against other methods. Note that for our method, especially compared with T2I Adapter, the details of the images are substantially more closely aligned with the depth prompt (see, e.g., the lamp in the background of the living room scene and the side table's legs, or the salad on the pizza)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.789, + 0.714 + ], + "angle": 0, + "content": "closely the generated image is following the provided structure map, we measure the cycle consistency between the predicted structure map of the original image \\( c = \\mathrm{encoder}(x) \\) and the predicted structure map of the generated image \\( \\hat{c} = \\mathrm{encoder}(\\hat{x}) \\). The exact metric used to measure the discrepancy depends on the type but \\( c \\) and \\( \\hat{c} \\) are always normalized to [0, 1] before any metrics computation. The differences in depth maps are measured by the Mean Squared Error between \\( c \\) and \\( \\hat{c} \\) (MSE-d). The SSIM (Structural Similarity) is used for HED conditioning. Additionally, we also compute the Learned Perceptual Image Patch Similarity (LPIPS) [49] to implicitly measure the fidelity of the control as well as the Frechet Inception Distance (FID) [11], which compares the distributions of intermediate features of a pre-trained network between generated and original images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Table 2 shows results across metrics for different approaches. In particular, adapters that focus on structure conditioning tend to use more parameters than ones used for style therefore we trained two versions of LoRAAdapter. One with 17M (roughly matching the number of parameters used for style conditioning) and 32M parameters, which is still less than T2I-Adapter [24]. LoRAAdapter obtains state-of-the-art performance across the board, even when reducing parameters by half compared to T2I-Adapter. These results show that LoRAAdapter is a very efficient approach for structure conditioning which can be scaled to obtain" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.594, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "5. EXPERIMENTS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.157, + 0.476, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.311, + 0.496, + 0.409 + ], + "angle": 0, + "content": "Fig. 6: We show how the choice of the LoRA rank affects the generation quality. The rank refers to the cross-attention models in Table 3. Even the smallest model with only 2M parameters can meaningfully capture the content of the image albeit with minor losses in quality." + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.157, + 0.789, + 0.268 + ], + "angle": 0, + "content": "Table 3: Evaluation of different global conditional LoRA configurations. All models were trained using CLIP VIT-L/14 as an image encoder and evaluated on 5000 samples. Adapting the Cross-Attention layers yields the best performance. Even very low-rank LoRA conditioning can still produce good performance." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.27, + 0.789, + 0.382 + ], + "angle": 0, + "content": "
LayerRank ParametersCLIP-ICLIP-T
Cross Attention20821M0.8370.631
Cross Attention12813M0.8350.634
Cross Attention647M0.8290.637
Cross Attention324M0.8160.642
Cross Attention162M0.7990.649
Self Attention12813M0.8280.639
Cross & Self Attention12826M0.8420.629
Cross & Self Attention6413M0.8370.636
Convolutional12820M0.8060.647
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.785, + 0.472 + ], + "angle": 0, + "content": "high-performant models. We refer the reader to the Appendix for additional results and comparisons." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.502, + 0.459, + 0.518 + ], + "angle": 0, + "content": "5.3 Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.625 + ], + "angle": 0, + "content": "To show qualitative results for both structure and style conditioning we take images used for conditioning by previous approaches while also adding some conditioning images selected by us. Results are shown in Fig. 4 where we can see that LoRAAdapter better captures the style of the conditioning image compared with previous adapter approaches. We show additional comparison with more models on an extended set of samples in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.788, + 0.718 + ], + "angle": 0, + "content": "We also show samples for structure conditioning on depth maps predicted by MiDaS [32] in Fig 5. LoRAAdapter shows nice sample quality and superior adherence to the fine-grained depth structure compared to ControlNet or T2I-Adapter. We attribute this improvement to our choice of directly adapting the convolutions in the U-Net as opposed to an indirect adaption via skip connections as previous approaches have proposed [7, 24]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.747, + 0.442, + 0.763 + ], + "angle": 0, + "content": "5.4 Ablation experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To showcase the modularity of our approach, we train multiple configurations per condition and assess their performance. We study the influence of the LoRAs rank as well as the specific layer choice, which is further discussed for each conditioning type." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Stracke et al." + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.143, + 0.724, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.312, + 0.788, + 0.356 + ], + "angle": 0, + "content": "Fig. 7: Examples of using a style and structure LoRAAdapter jointly. We show both reconstruction tasks (first, second, and third columns) and cases where the style and structure image do not match perfectly (fourth, fifth, and sixth columns)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.79, + 0.502 + ], + "angle": 0, + "content": "Rank and Layer choice As the modularity of our approach is a key differentiator from other methods, we perform an analysis of how the choice of adapted layers and the overall rank influence the performance in the context of style conditioning (Table 3). Overall the cross-attention layers perform best and show very good performance even with a very small rank. Convolutional and self-attention layers perform worse but may be better suited for other tasks such as conditioning on the style of an image instead of its semantics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.79, + 0.657 + ], + "angle": 0, + "content": "Structure We adapt the convolutional layers and vary in what blocks we adapt them. Stable Diffusion 1.5 contains four upsampling and downsampling blocks and one middle block. Each block itself contains two ResNet [9] blocks with two convolutional layers each. We analyze two different configurations: Configuration A only adapts the first layer in every block, totaling nine adapted convolutional layers. Configuration B adapts the first convolutional layer in every ResNet block, i.e. 17 adapted layers in total (Table 2). Finally, we show examples of jointly using a style and structure conditioning on Fig. 7. We refer the reader to the Appendix for additional results on prompt combinations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.675, + 0.36, + 0.693 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.84 + ], + "angle": 0, + "content": "In this paper we have introduced LoRAAdapter, an approach to condition the image generation process of foundational text-to-image models. LoRAAdapter is an powerful and efficient approach for conditioning that unifies structure and style control using a conditional LoRA architecture that enables zero-shot generalization. Our approach enables fine-grained control over both structure and style obtaining state-of-the-art results and outperforming recent approaches while also optimizing a more compact number of parameters. LoRAAdapter presents progress towards acquiring fine-grained control over text-to-image models in an effective manner." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.604, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "6. CONCLUSION" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.298 + ], + "angle": 0, + "content": "This project has been supported by the German Federal Ministry for Economic Affairs and Climate Action within the project \"NXT GEN AI METHODS - Generative Methoden für Perzeption, Prädiktion und Planung\", the German Research Foundation (DFG) project 421703927, and the bidt project KLIMA-MEMES. The authors gratefully acknowledge the Gauss Center for Supercomputing for providing compute through the NIC on JUWELS at JSC and the HPC resources supplied by the Erlangen National High Performance Computing Center (NHR@FAU funded by DFG)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.318, + 0.323, + 0.334 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.348, + 0.674, + 0.361 + ], + "angle": 0, + "content": "1. https://huggingface.co/lambdaabs/sd-image-variations-diffusers" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.362, + 0.786, + 0.403 + ], + "angle": 0, + "content": "2. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.404, + 0.786, + 0.431 + ], + "angle": 0, + "content": "3. Byeon, M., Park, B., Kim, H., Lee, S., Baek, W., Kim, S.: Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.431, + 0.787, + 0.47 + ], + "angle": 0, + "content": "4. Canny, J.: A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence PAMI-8(6), 679-698 (1986). https://doi.org/10.1109/TPAMI.1986.4767851" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.471, + 0.786, + 0.512 + ], + "angle": 0, + "content": "5. Cao, Z., Simon, T., Wei, S.E., Sheikh, Y.: Realtime multi-person 2d pose estimation using part affinity fields. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7291-7299 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.512, + 0.787, + 0.539 + ], + "angle": 0, + "content": "6. Cheng, J., Xie, P., Xia, X., Li, J., Wu, J., Ren, Y., Li, H., Xiao, X., Zheng, M., Fu, L.: Resadapter: Domain consistent resolution adapter for diffusion models (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.54, + 0.786, + 0.566 + ], + "angle": 0, + "content": "7. Denis Zavadski, J.F.F., Rother, C.: Controlnet-xs: Designing an efficient and effective architecture for controlling text-to-image diffusion models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.567, + 0.786, + 0.594 + ], + "angle": 0, + "content": "8. Gandikota, R., Materzynska, J., Zhou, T., Torralba, A., Bau, D.: Concept sliders: Lora adaptors for precise control in diffusion models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.595, + 0.787, + 0.635 + ], + "angle": 0, + "content": "9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.636, + 0.786, + 0.662 + ], + "angle": 0, + "content": "0. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.663, + 0.786, + 0.703 + ], + "angle": 0, + "content": "11. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.704, + 0.786, + 0.731 + ], + "angle": 0, + "content": "12. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.731, + 0.786, + 0.757 + ], + "angle": 0, + "content": "13. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.786, + 0.799 + ], + "angle": 0, + "content": "14. Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "15. Huang, L., Chen, D., Liu, Y., Shen, Y., Zhao, D., Zhou, J.: Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint arXiv:2302.09778 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.348, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Stracke et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.788, + 0.19 + ], + "angle": 0, + "content": "16. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip. https://github.com/mlfoundations/open Clip (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "17. Jiang, Z., Mao, C., Pan, Y., Han, Z., Zhang, J.: Scedit: Efficient and controllable image diffusion generation via skip connection editing. arXiv preprint arXiv:2312.11392 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.787, + 0.26 + ], + "angle": 0, + "content": "18. Kopiczko, D.J., Blankevoort, T., Asano, Y.M.: Vera: Vector-based random matrix adaptation (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.787, + 0.317 + ], + "angle": 0, + "content": "19. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22511-22521 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.787, + 0.374 + ], + "angle": 0, + "content": "20. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.374, + 0.787, + 0.402 + ], + "angle": 0, + "content": "21. Liu, S.Y., Wang, C.Y., Yin, H., Molchanov, P., Wang, Y.C.F., Cheng, K.T., Chen, M.H.: Dora: Weight-decomposed low-rank adaptation (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.402, + 0.787, + 0.431 + ], + "angle": 0, + "content": "22. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.787, + 0.472 + ], + "angle": 0, + "content": "23. Luo, S., Tan, Y., Patil, S., Gu, D., von Platen, P., Passos, A., Huang, L., Li, J., Zhao, H.: Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.515 + ], + "angle": 0, + "content": "24. Mou, C., Wang, X., Xie, L., Wu, Y., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.787, + 0.558 + ], + "angle": 0, + "content": "25. Park, T., Liu, M.Y., Wang, T.C., Zhu, J.Y.: Semantic image synthesis with spatially-adaptive normalization. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2332-2341. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.787, + 0.6 + ], + "angle": 0, + "content": "26. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4195-4205 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.787, + 0.657 + ], + "angle": 0, + "content": "27. Pernias, P., Rampas, D., Richter, M.L., Pal, C., Aubreville, M.: Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=gU58d5QeGv" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.787, + 0.699 + ], + "angle": 0, + "content": "28. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.787, + 0.741 + ], + "angle": 0, + "content": "29. Qin, C., Zhang, S., Yu, N., Feng, Y., Yang, X., Zhou, Y., Wang, H., Niebles, J.C., Xiong, C., Savarese, S., et al.: Unicontrol: A unified diffusion model for controllable visual generation in the wild. arXiv preprint arXiv:2305.11147 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.787, + 0.798 + ], + "angle": 0, + "content": "30. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "31. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.604, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "6. CONCLUSION" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "32. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence 44(3), 1623-1637 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "33. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.3 + ], + "angle": 0, + "content": "34. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.787, + 0.342 + ], + "angle": 0, + "content": "35. Ruiz, N., Li, Y., Jampani, V., Wei, W., Hou, T., Pritch, Y., Wadhwa, N., Rubinstein, M., Aberman, K.: Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.37 + ], + "angle": 0, + "content": "36. Ryu, S.: Low-rank adaptation for fast text-to-image diffusion fine-tuning, https://github.com/cloneofsimo/lora" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.424 + ], + "angle": 0, + "content": "37. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.452 + ], + "angle": 0, + "content": "38. Shah, V., Ruiz, N., Cole, F., Lu, E., Lazebnik, S., Li, Y., Jampani, V.:Ziplora: Any subject in any style by effectively merging loras (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.479 + ], + "angle": 0, + "content": "39. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.787, + 0.508 + ], + "angle": 0, + "content": "40. Witteveen, S., Andrews, M.: Investigating prompt engineering in diffusion models. arXiv preprint arXiv:2211.15462 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.536 + ], + "angle": 0, + "content": "41. Xie, S., Tu, Z.: Holistically-nested edge detection. In: Proceedings of the IEEE international conference on computer vision. pp. 1395-1403 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.787, + 0.577 + ], + "angle": 0, + "content": "42. Xu, X., Guo, J., Wang, Z., Huang, G., Essa, I., Shi, H.: Prompt-free diffusion: Taking\" text\" out of text-to-image diffusion models. arXiv preprint arXiv:2305.16223 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.605 + ], + "angle": 0, + "content": "43. Xu, X., Wang, Z., Zhang, E., Wang, K., Shi, H.: Versatile diffusion: Text, images and variations all in one diffusion model. arXiv preprint arXiv:2211.08332 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.646 + ], + "angle": 0, + "content": "44. Xue, Z., Song, G., Guo, Q., Liu, B., Zong, Z., Liu, Y., Luo, P.: Raphael: Text-to-image generation via large mixture of diffusion paths. arXiv preprint arXiv:2305.18295 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.787, + 0.674 + ], + "angle": 0, + "content": "45. Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.729 + ], + "angle": 0, + "content": "46. YEH, S.Y., Hsieh, Y.G., Gao, Z., Yang, B.B.W., Oh, G., Gong, Y.: Navigating text-to-image customization: From lyCORIS fine-tuning to model evaluation. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=wfzXa8e783" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.787, + 0.757 + ], + "angle": 0, + "content": "47. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "48. Zhang, Q., Chen, M., Bukharin, A., Karampatziakis, N., He, P., Cheng, Y., Chen, W., Zhao, T.: Adalora: Adaptive budget allocation for parameter-efficient fine-tuning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "49. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.361, + 0.127 + ], + "angle": 0, + "content": "Stracke et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.788, + 0.19 + ], + "angle": 0, + "content": "50. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Unictrlnet: All-in-one control to text-to-image diffusion models. Advances in Neural Information Processing Systems (2023)" + } + ] +] \ No newline at end of file diff --git a/2024/CTRLorALTer_ Conditional LoRAdapter for Efficient 0-Shot Control & Altering of T2I Models/d556ef6f-c8cf-4800-af84-90cf37eef09f_origin.pdf b/2024/CTRLorALTer_ Conditional LoRAdapter for Efficient 0-Shot Control & Altering of T2I Models/d556ef6f-c8cf-4800-af84-90cf37eef09f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..77afed267896bea3823f015ee091ac76f3220c2a --- /dev/null +++ b/2024/CTRLorALTer_ Conditional LoRAdapter for Efficient 0-Shot Control & Altering of T2I Models/d556ef6f-c8cf-4800-af84-90cf37eef09f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:805e9f40efb704fb707d63fbae56731792b8d427abe1444610b6a20daf68d82e +size 9557797 diff --git a/2024/CTRLorALTer_ Conditional LoRAdapter for Efficient 0-Shot Control & Altering of T2I Models/full.md b/2024/CTRLorALTer_ Conditional LoRAdapter for Efficient 0-Shot Control & Altering of T2I Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5629c7e6515a833e5ad6119be465bcac69a02fbc --- /dev/null +++ b/2024/CTRLorALTer_ Conditional LoRAdapter for Efficient 0-Shot Control & Altering of T2I Models/full.md @@ -0,0 +1,315 @@ +# CTRLorALTer: Conditional LoRAapter for Efficient 0-Shot Control & Altering of T2I Models + +Nick Stracke1, Stefan Andreas Baumann1, Joshua Susskind2, Miguel Angel Bautista2, and Björn Ommer1 + +CompVis @ LMU Munich, MCML + +2 Apple + +Abstract. Text-to-image generative models have become a prominent and powerful tool that excels at generating high-resolution realistic images. However, guiding the generative process of these models to take into account detailed forms of conditioning reflecting style and/or structure information remains an open problem. In this paper, we present LoRAAdapter, an approach that unifies both style and structure conditioning under the same formulation using a novel conditional LoRA block that enables zero-shot control. LoRAAdapter is an efficient and powerful approach to condition text-to-image diffusion models, which enables fine-grained control conditioning during generation and outperforms recent state-of-the-art approaches. + +Project page: compvis.github.io/LoRAAdapter/ + +Keywords: T2I Models $\cdot$ Diffusion Models $\cdot$ Controllable Generation + +# 1 Introduction + +Text-to-image models have become a foundational component of Computer Vision, especially with the introduction of large-scale diffusion models like Stable Diffusion (SD) [33], DALL-E 2 [31], Imagen [37], RAPHAEL [44] and, eDiff-I [2]. These models have empowered users to create realistic images from textual prompts, although crafting effective prompts is a complex task in itself due to the nuanced prompt engineering required [40] to effectively condition the image generation process. To tackle this problem, prior approaches such as SD Image Variations [1], Stable unCLIP [33], and Composer [15], have attempted to perform conditioning using image prompts to enhance the expressivity of the conditioning mechanism giving users more flexibility during generation. In practice, this family of approaches requires either training from scratch or fully fine-tuning large text-conditioned model weights, incurring high computational costs and requiring large datasets. + +To reduce training cost and time, recent approaches like ControlNet [47], T2I-Adapter [24], Uni-ControlNet [50] or IP-Adapters [45] have focused on designing adapters that keep the base model frozen and only introduce a limit number of + +![](images/f47f35ce19fa3fb2627583f25816ce11b85a6c4d814339b6739cd38db5fc8cde.jpg) +Fig. 1: LoRAAdapter allows structure and style control of the image generation process of text-to-image models in a zero-shot manner. Our approach enables powerful fine-grained and efficient unified control over both structure and style conditioning using conditional LoRA blocks. + +new layers. This not only reduces the number of trainable parameters but also allows training on much smaller datasets because the frozen base model cannot suffer from catastrophic forgetting. These adapters are typically handcrafted to a specific model architecture and conditioning modality. They can be roughly assigned into one of two groups: + +- Adapters for pixel-level fine-grained conditioning over the generated image. This is often referred to as structural or local conditioning since it allows influencing the generated image on a per-pixel level by e.g. providing a depth map and effectively setting the layout of the image [7,17,24,29,47,50] +- Adapters for style or global conditioning perform tasks such as image-to-image translation similar to unCLIP [15, 19, 33, 43]. Here, the focus is not on fine-grained control but instead on generating an image similar to the prompt image, where similarity is defined by sharing the same style or semantics. + +Empirically, adapters always favor one type of conditioning over the other. For example, both Uni-ControlNet [50] and T2I-Adapter [24] have attempted style conditioning but their performance is not comparable to leading style approaches [45]. Conversely, style adapters require ControlNet or T2I-Adapter for structure conditioning. However, ControlNet is a large adapter that copies the entire encoder of Stable Diffusion's U-Net which significantly increases compute and thus inference time. Our goal is to find a unified approach that is capable of both structure and style conditioning, providing complete control over the generated image. Formulating such a unified approach for conditioning on global controls like style and on local controls like structure, in an efficient and generic manner remains a key open problem. + +To address this challenging problem, we introduce LoRAAdapter (see Fig. 1), a unified approach for incorporating images to condition both the style and structure of the generated images. LoRAAdapter takes inspiration from Low-Rank-Adaptations (LoRAs) [14], which have shown remarkable performance for various tasks in diffusion models, such as distillation or learning specific concepts [23,35]. In particular, LoRAs low-rank property naturally regularizes the conditioning while offering unparalleled flexibility as any layer can be adapted in an architecture-agnostic way. Thus far, however, LoRAs have been implemented as fixed adaptations, being independent of the input. This means layer adaptation cannot change at test time for a given conditioning and, therefore, cannot be used for zero-shot generalization. + +LoRAAdapter is a novel approach to adding conditional information to LoRAs, enabling zero-shot generalization and making them applicable for both structure and style and possibly many other conditioning types. LoRAAdapter is compact and efficient, e.g., optimizing 16M parameters vs the 22M of IP-Adapters [45] or 361M of ControlNet [47], while outperforming recent adapter approaches [45] and even approaches that train models from scratch (see Tab. 1). LoRAAdapter trains LoRA [14] modules for different blocks across the base network and is efficient both during training and inference. Our contributions are summarized as follows: + +- We propose LoRAAdapter, a generic approach to train conditional LoRAs that is agnostic to model architecture and conditioning modality. +- We implement LoRAAdapter for Stable Diffusion, offering a unified conditioning mechanism for both style and structure, enabling zero-shot conditioning. +- We show the effectiveness of this approach, outperforming dedicated adapters for either style and structure on various metrics. + +# 2 Related Work + +We now review existing adapter approaches for style and structure and discuss the current role of LoRAs in diffusion models. + +# 2.1 Structure Adapters + +Structure adapters are specifically designed to operate on various types of local conditioning modalities such as Depth, HED [41], Canny Edges [4], Scribbles, or key poses [5]. The goal of these adapters is to spatially align the generated image with the provided condition. One of the most prominent approaches to tackle this problem is ControlNet [47] for Stable Diffusion. ControlNet creates a copy of the U-Net's encoder and combines the skip connections additively with the original skip connections, which then get fed into the original decoder. While obtaining great performance, ControlNet adds a lot of computational overhead as the forward pass includes a second encoder which can increase inference speed by up to $50\%$ . Another consequence of the large capacity is that ControlNet tends to interpret the structure map directly. This can be desirable when sampling + +images without other conditioning via the text prompt or other adapters, but can also lead to entanglement of structure and style. + +Following a similar idea, T2I-Adapter [24] also modifies skip connections while using a much smaller encoder network. To extend conditioning to multiple modalities, Uni-ControlNet [50] was proposed as an alternative solution to having to train multiple separate and modality-specific ControlNets. Uni-ControlNet trains a single model for multiple conditions by concatenating the structure maps along the channel dimension and using the concatenated tensor as input. Uni-ControlNet uses SPADE [25] to combine skip-connections instead of the simple addition used by ControlNet. SCEdit [17] proposed a similar but more efficient solution for conditioning on multiple modalities by integrating a small tuner network (SC-Tuner) between skip-connections which removes the need for a separate encoder network akin to ControlNet. + +# 2.2 Style Adapters + +Style adapters are used as an alternative to unCLIP [33] like models, which are directly trained to invert CLIP image embeddings. The main advantage of using a style adapter as opposed to an unCLIP model is that the original text conditioning is preserved, enabling additional control over the generated images. Additionally, all recently released large open-source diffusion models only offered text conditioning, making an adapter approach necessary [27, 28]. + +IP-Adapter [45] adds a decoupled cross-attention layer that operates on four image tokens instead of text tokens. The resulting activations of this decoupled cross-attention layer are scaled and added to the original cross-attention activations. This results in a lightweight adapter that integrates well with potential text conditioning and generates images faithful to the image prompt. In contrast, SeeCoder [42] trains an encoder with 2D spatial embedding to convert image tokens to text tokens and completely replaces the standard text tokens with their tokens. + +Structure adapters have also attempted to include style conditioning. UniControlNet [50] includes a separate mapping network to convert a pooled CLIP image token to four CLIP text tokens, which they concatenate to the original text embeddings. ControlNet-shuffle [47] shuffles an RGB image and uses that as conditioning similar to its structure conditioning. However, the performance of these approaches does not compare favorably to dedicated style adapters. + +# 2.3 LoRAs in Diffusion Models + +While LoRAs were initially introduced as an efficient alternative to fine-tuning large language models [14], they were quickly adopted for diffusion models [36] as a relatively efficient and cheap option to adjust their generation capabilities. Combining LoRAs with methods such as Dreambooth [34] performs on par with a full fine-tuning approach while only introducing a fraction of trainable parameters. LoRAs are commonly trained on a small set of images with the goal of capturing a specific style instead of a subject. ZipLoRA [38] proposes a merging algorithm to + +combine multiple LoRAs gracefully by minimizing the cosine similarity of LoRAs that adapt the same weight matrix. LyCORIS [46] trained an entire library of specific LoRAs in diffusion models with various benefits. [23] showed that LoRAs are even powerful enough for more complex objectives such as consistency distillation. Finally, Concept Sliders [8] use LoRAs to train disentangled edit directions to edit images. They scale the LoRA in the negative and positive directions to either increase or decrease a given attribute. + +# 3 Preliminaries + +# 3.1 Diffusion + +Diffusion models represent a category of generative models characterized by a two-stage process: a forward diffusion that incrementally introduces Gaussian noise over a series of $T$ steps, and a reverse denoising process that reconstructs samples from this noise. In text-to-image models, diffusion can be directed by supplementary inputs like text descriptions. The primary training objective, $\epsilon_{\theta}$ , aimed at noise prediction, simplifies the variational lower bound as follows: + +$$ +L _ {\mathrm {s i m p l e}} = \mathbb {E} _ {\boldsymbol {x} _ {0}, \epsilon \sim \mathcal {N} (\mathbf {0}, \mathbf {I}), \boldsymbol {c}, t} \| \boldsymbol {\epsilon} - \epsilon_ {\theta} \big (\underbrace {\sqrt {\bar {\alpha} _ {t}} \boldsymbol {x} _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \boldsymbol {\epsilon}} _ {\boldsymbol {x} _ {t}}, \boldsymbol {c}, t \big) \| ^ {2}, \tag {1} +$$ + +where $\pmb{x}_0$ is the original data influenced by condition $\pmb{c}$ , $t$ is the diffusion timestep, and $\pmb{x}_t$ represents the data at step $t$ , with $\bar{\alpha}_t$ dictating the noise level. + +During sampling, an initial pure Gaussian latent $\pmb{x}_T$ is iteratively denoised by the diffusion model to obtain a sample $\pmb{x}_0$ from the modeled (conditional) distribution. Conditional models often employ classifier guidance to fine-tune the balance between accuracy and diversity, with classifier-free guidance [12] serving as a preferred alternative that merges conditional and unconditional noise predictions during training and sampling: + +$$ +\hat {\boldsymbol {\epsilon}} _ {\theta} \left(\boldsymbol {x} _ {t}, \boldsymbol {c}, t\right) = w \boldsymbol {\epsilon} _ {\theta} \left(\boldsymbol {x} _ {t}, \boldsymbol {c}, t\right) + (1 - w) \boldsymbol {\epsilon} _ {\theta} \left(\boldsymbol {x} _ {t}, t\right), \tag {2} +$$ + +where $w$ modulates the adherence to condition $c$ . This technique is pivotal for text-to-image models, enhancing the alignment between generated images and text prompts. + +# 3.2 Low-Rank Adaptation (LoRA) + +Low-Rank Adaptation (LoRA) [14] is a method for finetuning pre-trained foundation models by keeping the original weight matrices frozen and instead adding a new set of low-rank weight matrices. While this method was originally proposed in the context of large language models, it has been recently adopted in the domain of diffusion models [6,8,23]. + +Specifically, the original weight matrix of a layer in the base model $W_0 \in \mathbb{R}^{d \times k}$ is kept frozen and is adapted by a low-rank matrix $W + \Delta W = W + BA$ with + +$B \in \mathbb{R}^{d \times r}$ and $A \in \mathbb{R}^{r \times k}$ . Crucially, $r$ is chosen following $r \ll \min(d, k)$ . The forward pass for the adapted layer is then: + +$$ +h = W _ {0} x + \Delta W x = W _ {0} x + B A x. \tag {3} +$$ + +Typically, $B$ is zero-initialized, so the LoRA does not affect the model initially. This yields a generic approach, as we can freely choose the rank $r$ and the weights according to the given task. This happens on a per-layer basis, i.e., if we adapt $n$ layers, we will have $nA$ and $B$ matrices. + +# 4 Method + +There are two kinds of approaches for conditioning text-to-image diffusion models: style or global conditioning [42, 45], and structure or local conditioning [17, 24, 29, 50]. These approaches are specifically tailored to a specific conditioning objective and therefore missing a holistic approach that can efficiently learn both style and structure. We now describe LoRAAdapter, an efficient, holistic, and customizable method for introducing both style and structure conditioning into text-to-image diffusion models. + +![](images/096c5c3ab9630367a314d2ecf880751abe6c2e22ac93da39af88171be02d1a26.jpg) +Fig. 2: Overview of the proposed conditional LoRA block. The original weight matrix $W_0^{(i)}$ is frozen while all other layers are trained. $\phi$ is an affine transformation that operates on the low-dimensional embedding $A^{(i)}x$ and introduces the conditioning. The local mapper network $m_L^{(i)}$ predicts the scale and shift parameters $\beta, \gamma$ for the affine transformation. Typically, we set $m_L^{(i)}$ to be a small network. If complex transformations are required to map the conditioning $c$ , this happens in $m_S$ since it is shared across all adapted layers. + +# 4.1 Conditional LoRAs + +LoRAs have successfully enabled adaptation of diffusion models across a range of architectures [6,23,35] to steer their generation process. In this setting, LoRAs + +have been considered as free parameters and optimized on a training set of images exhibiting a specific characteristic. + +Instead of learning a single LoRA to apply one specific modification, e.g., changing the style of the generated image to a painting or learning to represent one specific subject [34,35], we propose a LoRA-based conditioning mechanism whose behavior changes based on conditioning provided at inference time, enabling zero-shot generalization. While previous methods depend on specific aspects of the architectures of the models they are applied to [45, 47], using LoRAs for conditioning enables us to efficiently adapt foundation models for conditional tasks in an architecture-agnostic way. Our LoRA-based conditioning mechanism is shown in Fig. 2. + +Starting from the standard LoRA setup as in Eq. (3), we propose implementing conditioning of the LoRA by applying a transformation $\phi(Ax|m(c))$ to the low-dimensional intermediate embedding in the LoRA that introduces conditional behavior based on the conditioning $c$ : + +$$ +h = W _ {0} x + \Delta W x = W _ {0} x + B \phi (A x | m (c)). \tag {4} +$$ + +This limits the conditioning to apply only to a low-rank subspace of the original vector space the weight matrix operates in. This is crucial for the generalization capabilities of LoRAs, as it introduces regularization and thus limits the adaptation to only focus on relevant aspects instead of spurious correlations from the dataset it was trained on, which results in more efficient training. By introducing our conditional adaptation of the model's behavior in such a low-rank space, our method can also benefit from those advantages. + +In our framework for conditional LoRAs, we can theoretically choose $\phi(Ax|m(c))$ to be arbitrary, non-linear transformations, but we find that a simple affine transformation is sufficiently expressive for a wide range of applications: + +$$ +\phi (A x \mid m (c)) = \gamma_ {\phi} (m (c)) \odot A x + \beta_ {\phi} (m (c)), \tag {5} +$$ + +with $\odot$ denoting an elementwise (Hadamard) product and $\gamma$ and $\beta$ referring to the scale and shift factors. + +To predict the $\gamma$ and $\beta$ from the conditioning $c$ , we utilize a mapping network. This is generally a small neural network that we separate into a shared part $m_{S}(\cdot)$ and a layer-specific part $m_{L}^{(i)}(\cdot)$ that is individually learned for each adapted layer. For both parts, we generally use very small neural networks, which is in stark contrast to other works such as ControlNet [47] that uses a full copy of the U-Net's encoder (see Tab. 1). For brevity, we will always refer to both parts simply as $m(\cdot)$ . + +This general formulation directly enables both local and global conditioning, as it can be applied to the convolutional and attention layers in a standard diffusion model. For local conditioning, the mapping network can supply spatially dependent LoRA modulations. At the same time, it can also be used to supply global conditioning by introducing non-spatially dependent information, such as by adapting the cross-attention projections. Concurrent works [18, 21, 48] also introduced improvements to the basic LoRA architecture. As LoRAAdapter does + +not depend on a specific implementation of LoRAs but just on the low-rank intermediate sub-space, these improvements can be directly transferred to our model. + +# 4.2 A Unified Conditioning Approach + +In this section, we provide implementation details for applying our conditional LoRAs to the layer types commonly used in large-scale diffusion models. + +![](images/6973291b2a25e51354fbde26ccae2dc3e154d4d904d75f6f26b72638460445a7.jpg) +(a) In attention layers, we apply an affine transformation on the $K$ and $V$ projection matrix per channel for every token. + +![](images/19b06692c5434f81e10f540612515df71cd7da8d80dc071b2b9afd8793819a09.jpg) +(b) To add structural conditioning to convolutional layers, we perform an affine transformation for every element similar to [25]. +Fig. 3: Visualization of implementations of our conditional LoRAs for specific layers. + +Attention Layers Many diffusion models, such as Stable Diffusion $\{1,2,\mathrm{XL}\}$ [28,33] or Imagen [37] include self- and cross-attention layers that operate on image (and text) tokens, with transformer-based diffusion models such as [26] even purely relying on them for spatial computations. The output $h^\prime$ of the attention layers is defined as + +$$ +h ^ {\prime} = \operatorname {A t t e n t i o n} (Q, K, V) = \operatorname {S o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d}}\right) V \tag {6} +$$ + +with the projections $Q = hW_{q}$ , $K = hW_{k}$ or $K = c_{t}W_{k}$ , and $V = hW_{v}$ or $V = c_{t}W_{v}$ , depending on whether it is self- or cross-attention, and the text tokens $c_{t}$ . Here, we apply our conditional LoRA on the linear projections $W_{k}$ and $W_{v}$ based on Equation 4. We show a diagram of conditional LoRAs for attention layers in Fig. 3a. + +Convolutional Layers For convolutional layers, which form the main backbone of many diffusion models [27,33], we adapt the convolutional layer + +$$ +h = K _ {0} \star x, \tag {7} +$$ + +with the 2D convolution kernel $K$ having input channel count $ch_{i}$ and output channel count $ch_{o}$ , in a similar manner as linear projections. However, in this setting we work on convolution kernels instead of matrices: we perform one convolution $K_{A}$ with the same kernel size, stride, and padding as the original + +Table 1: Quantitative comparison for style conditioning of our proposed LoRAAdapter with other methods on the COCO validation set with four samples for every image. The best results are in bold (adapted from [45]). + +
Style MethodReusable to custom modelsNative structure controlMultimodal promptsTrainable parametersCLIP-T ↑CLIP-I ↑
Training from scratch
Open unCLIPXXX893M0.6080.858
Kandinsky-2-1XXX1229M0.5990.855
Versatile DiffusionXX✓860M0.5870.830
Fine-tunining from text-to-image model
SD Image VariationsXXX860M0.5480.760
SD unCLIPXXX870M0.5840.810
Adapters
Uni-ControlNet (Global Control)✓✓✓47M0.5060.736
T2I-Adapter (Style)✓✓✓39M0.4850.648
ControlNet Shuffle✓✓✓361M0.4210.616
IP-Adapter✓X✓22M0.5880.828
LoRAAdapter (ours)✓✓✓16M0.6370.831
LoRAAdapter SDXL (ours)✓✓✓103M0.6490.849
+ +convolution $K_{0}$ , which reduces the channel count from $ch_{i}$ to the rank $r$ of the LoRA while keeping all else equal. Then, we apply our LoRA conditioning $\phi(\cdot|m(c))$ in this bottlenecked space and expand to the output channel count $ch_{o}$ using a pointwise convolution $K_{B}$ : + +$$ +h = K _ {0} \star x + K _ {B} \star \phi \left(K _ {A} \star x | m (c)\right). \tag {8} +$$ + +A visual diagram of conditional LoRAs for convolutional layers is shown in Fig. 3b. + +# 5 Experiments + +# 5.1 Experimental Setup + +Data We train LoRAdapter on a 40 million samples subset of COYO-700M [3] that only contains images with a short side of at least 512 pixels. Larger images are downscaled and center-cropped to 512 pixels. For text-conditioning, we use the standard prompts provided by the COYO-700M dataset. + +Training and Implementation Details All experiments are based on Stable Diffusion 1.5 [33] unless noted otherwise and adapted with a conditional LoRA according to the given task. We train our adapter for 50,000 steps on 32 A100 GPUs with a global batch size of 256 and a learning rate of 0.0001. We use AdamW [22] as an optimizer and drop each conditioning with a probability of 0.05. This includes the original text conditioning. + +Table 2: Comparison of a smaller and a larger version of our models vs. ControlNet. + +
ModelParamsMSE-d ↓FID ↓LPIPS ↓
ControlNet361M17.9817.6440.622
Uni-ControlNet361M17.5616.4530.601
T2I-Adapter39M24.7817.9410.636
LoRApapter-A (Ours)17M17.2716.8490.599
LoRApapter-B (Ours)32M15.3415.6700.572
+ +For style conditioning, we experiment with both CLIP ViT-L/14 [30] by OpenAI and CLIP ViT-H/14 by OpenClip [16] as the image encoder but found that the ViT-H/14 tends to produce more visually appealing results. We still use ViT-L/14 for some ablation experiments as it results in a smaller model. Similar to [45], we use a single linear layer followed by layer normalization for the shared mapping network $m_S$ as it outperformed much larger mapping networks and showed faster convergence. This maps the pooled CLIP image token $c_{img}$ , keeping its original dimensionality $d_{img}$ . The local mapping network $m_l^{(i)}$ consists of two separate linear layers that predict $\beta$ and $\gamma$ in the correct dimensionality $r$ of the LoRA embedding. In total, the shared mapping network only has 1M parameters. + +For structure conditioning, the shared mapping network is a series of convolutional layers that spatially align the dimensions of the conditioning images with the dimensions that Stable Diffusion's U-Net expects which sums up to only 1.3M parameters. As the U-Net includes several down- and upsampling blocks, the mapping network outputs feature maps at various resolutions that match the specific U-Net blocks. Each individual LoRA only gets the single feature map that matches the spatial dimensionality of the U-Net block that it operates in. This feature map is then further processed with a single convolutional layer as a local mapping to align the number of channels with the rank $r$ of the LoRA embedding. + +Inference We sample with 50 DDIM [39] steps and apply classifier-free guidance (CFG) [13] with a scale of 7.5 to the prompt and style conditioning. We do not apply any CFG for structure conditioning. As with all LoRAs, we can also influence the strength of our LoRAAdapter by adjusting the LoRA scale $\lambda$ , i.e. $h = W_0 + \lambda BAx$ . We set $\lambda = 1$ unless otherwise noted. + +# 5.2 Quantitative Evaluation + +To have a fair comparison with previous approaches, we follow previous literature on adapters [24,29,45] and evaluate our model on the validation set of COCO2017 [20] which contains 5,000 images. As every image comes with multiple prompts, we randomly select one prompt and use that for sampling. + +Style Following [45], we sample four images for a given image and text prompt pair. To measure the effectiveness of the new style conditioning, we use CLIP ViT-L/14 to obtain the image embeddings of the generated and the prompt image and compute the cosine similarity between them (CLIP-I). In addition, we quantify how well the generated image still follows the text prompt by calculating the CLIP score [10] between the text prompt and the generated image (CLIP-T). In Tab. 1 we show results across different approaches that either train models from scratch or fine-tune existing models, as well as compare with recent state-of-the-art adapters. + +![](images/35c648de9a945bb5b524673399e74a904f952dd81ea1686cba5ba9c687264dbe.jpg) +Image Conditioning +Fig. 4: Samples from our method with style conditioning compared against other methods. We used an empty prompt and only conditioned on the image. We generally perform on par with IP-Adapter and outperform it on some samples. Note that the third image from the left is less degraded, and the third image from the right captures the mane of the horse better. + +LoRAAdapter obtains state-of-the-art performance on CLIP-I and CLIP-T scores across all adapters, while also being more efficient (approx $27\%$ reduction in the number of trainable parameters compared with IP-Adapter [45]). We attribute this boost in performance to our conditional LoRA approach which can efficiently capture detailed style conditioning in a more effective manner than previous adapter approaches. Interestingly, we find that LoRAAdapter even outperforms methods that train large models from scratch like unCLIP on CLIP-T score, which further supports our unified conditioning approach. + +Structure For evaluating structure control we follow the settings in UniControlNet [50] and sample a single image per condition. To evaluate how + +![](images/fca4aba6f86f6b2ca31196e60cacf7011bc45d55cf135c5734c72beb7f1d13e3.jpg) +Prompt + +![](images/91e8b013dc4d4ac6c3bdddb7a64c94b452d5441b0572077f961169a354d40600.jpg) +ControlNet + +![](images/550fa223819a407c4315fb0a2f8686ab98644279da79abbae678e30dff1d1bf1.jpg) +T2I Adapter + +![](images/878531e1d79adca1b83b27f63bb45bf63699faa24594ccbdd76e15072ac13dea.jpg) +Uni-ControlNet + +![](images/6aada045b0f7398da0c4a6bf76fcaf218ed7a8d8d5a4326fe84484669de68941.jpg) +Ours +Fig. 5: Samples from our method with structural conditioning compared against other methods. Note that for our method, especially compared with T2I Adapter, the details of the images are substantially more closely aligned with the depth prompt (see, e.g., the lamp in the background of the living room scene and the side table's legs, or the salad on the pizza) + +closely the generated image is following the provided structure map, we measure the cycle consistency between the predicted structure map of the original image $c = \mathrm{encoder}(x)$ and the predicted structure map of the generated image $\hat{c} = \mathrm{encoder}(\hat{x})$ . The exact metric used to measure the discrepancy depends on the type but $c$ and $\hat{c}$ are always normalized to [0, 1] before any metrics computation. The differences in depth maps are measured by the Mean Squared Error between $c$ and $\hat{c}$ (MSE-d). The SSIM (Structural Similarity) is used for HED conditioning. Additionally, we also compute the Learned Perceptual Image Patch Similarity (LPIPS) [49] to implicitly measure the fidelity of the control as well as the Frechet Inception Distance (FID) [11], which compares the distributions of intermediate features of a pre-trained network between generated and original images. + +Table 2 shows results across metrics for different approaches. In particular, adapters that focus on structure conditioning tend to use more parameters than ones used for style therefore we trained two versions of LoRAAdapter. One with 17M (roughly matching the number of parameters used for style conditioning) and 32M parameters, which is still less than T2I-Adapter [24]. LoRAAdapter obtains state-of-the-art performance across the board, even when reducing parameters by half compared to T2I-Adapter. These results show that LoRAAdapter is a very efficient approach for structure conditioning which can be scaled to obtain + +![](images/e5f552d84f1b0fd94fa6885b0c1d1c610c652093d5ab7caf34a0bf54fbb2c5dd.jpg) +Fig. 6: We show how the choice of the LoRA rank affects the generation quality. The rank refers to the cross-attention models in Table 3. Even the smallest model with only 2M parameters can meaningfully capture the content of the image albeit with minor losses in quality. + +Table 3: Evaluation of different global conditional LoRA configurations. All models were trained using CLIP VIT-L/14 as an image encoder and evaluated on 5000 samples. Adapting the Cross-Attention layers yields the best performance. Even very low-rank LoRA conditioning can still produce good performance. + +
LayerRank ParametersCLIP-ICLIP-T
Cross Attention20821M0.8370.631
Cross Attention12813M0.8350.634
Cross Attention647M0.8290.637
Cross Attention324M0.8160.642
Cross Attention162M0.7990.649
Self Attention12813M0.8280.639
Cross & Self Attention12826M0.8420.629
Cross & Self Attention6413M0.8370.636
Convolutional12820M0.8060.647
+ +high-performant models. We refer the reader to the Appendix for additional results and comparisons. + +# 5.3 Qualitative Comparison + +To show qualitative results for both structure and style conditioning we take images used for conditioning by previous approaches while also adding some conditioning images selected by us. Results are shown in Fig. 4 where we can see that LoRAAdapter better captures the style of the conditioning image compared with previous adapter approaches. We show additional comparison with more models on an extended set of samples in the Appendix. + +We also show samples for structure conditioning on depth maps predicted by MiDaS [32] in Fig 5. LoRAAdapter shows nice sample quality and superior adherence to the fine-grained depth structure compared to ControlNet or T2I-Adapter. We attribute this improvement to our choice of directly adapting the convolutions in the U-Net as opposed to an indirect adaption via skip connections as previous approaches have proposed [7, 24]. + +# 5.4 Ablation experiments + +To showcase the modularity of our approach, we train multiple configurations per condition and assess their performance. We study the influence of the LoRAs rank as well as the specific layer choice, which is further discussed for each conditioning type. + +![](images/4ec3335ba28531de687e35d2bc3c692d39cedeaf7881bbd7ad14620ef5697bf3.jpg) +Fig. 7: Examples of using a style and structure LoRAAdapter jointly. We show both reconstruction tasks (first, second, and third columns) and cases where the style and structure image do not match perfectly (fourth, fifth, and sixth columns). + +Rank and Layer choice As the modularity of our approach is a key differentiator from other methods, we perform an analysis of how the choice of adapted layers and the overall rank influence the performance in the context of style conditioning (Table 3). Overall the cross-attention layers perform best and show very good performance even with a very small rank. Convolutional and self-attention layers perform worse but may be better suited for other tasks such as conditioning on the style of an image instead of its semantics. + +Structure We adapt the convolutional layers and vary in what blocks we adapt them. Stable Diffusion 1.5 contains four upsampling and downsampling blocks and one middle block. Each block itself contains two ResNet [9] blocks with two convolutional layers each. We analyze two different configurations: Configuration A only adapts the first layer in every block, totaling nine adapted convolutional layers. Configuration B adapts the first convolutional layer in every ResNet block, i.e. 17 adapted layers in total (Table 2). Finally, we show examples of jointly using a style and structure conditioning on Fig. 7. We refer the reader to the Appendix for additional results on prompt combinations. + +# 6 Conclusion + +In this paper we have introduced LoRAAdapter, an approach to condition the image generation process of foundational text-to-image models. LoRAAdapter is an powerful and efficient approach for conditioning that unifies structure and style control using a conditional LoRA architecture that enables zero-shot generalization. Our approach enables fine-grained control over both structure and style obtaining state-of-the-art results and outperforming recent approaches while also optimizing a more compact number of parameters. LoRAAdapter presents progress towards acquiring fine-grained control over text-to-image models in an effective manner. + +# Acknowledgments + +This project has been supported by the German Federal Ministry for Economic Affairs and Climate Action within the project "NXT GEN AI METHODS - Generative Methoden für Perzeption, Prädiktion und Planung", the German Research Foundation (DFG) project 421703927, and the bidt project KLIMA-MEMES. The authors gratefully acknowledge the Gauss Center for Supercomputing for providing compute through the NIC on JUWELS at JSC and the HPC resources supplied by the Erlangen National High Performance Computing Center (NHR@FAU funded by DFG). + +# References + +1. https://huggingface.co/lambdaabs/sd-image-variations-diffusers +2. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022) +3. Byeon, M., Park, B., Kim, H., Lee, S., Baek, W., Kim, S.: Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset (2022) +4. Canny, J.: A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence PAMI-8(6), 679-698 (1986). https://doi.org/10.1109/TPAMI.1986.4767851 +5. Cao, Z., Simon, T., Wei, S.E., Sheikh, Y.: Realtime multi-person 2d pose estimation using part affinity fields. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7291-7299 (2017) +6. Cheng, J., Xie, P., Xia, X., Li, J., Wu, J., Ren, Y., Li, H., Xiao, X., Zheng, M., Fu, L.: Resadapter: Domain consistent resolution adapter for diffusion models (2024) +7. Denis Zavadski, J.F.F., Rother, C.: Controlnet-xs: Designing an efficient and effective architecture for controlling text-to-image diffusion models (2023) +8. Gandikota, R., Materzynska, J., Zhou, T., Torralba, A., Bau, D.: Concept sliders: Lora adaptors for precise control in diffusion models (2023) +9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +0. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021) +11. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017) +12. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) +13. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) +14. Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2021) +15. Huang, L., Chen, D., Liu, Y., Shen, Y., Zhao, D., Zhou, J.: Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint arXiv:2302.09778 (2023) + +16. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip. https://github.com/mlfoundations/open Clip (2021) +17. Jiang, Z., Mao, C., Pan, Y., Han, Z., Zhang, J.: Scedit: Efficient and controllable image diffusion generation via skip connection editing. arXiv preprint arXiv:2312.11392 (2023) +18. Kopiczko, D.J., Blankevoort, T., Asano, Y.M.: Vera: Vector-based random matrix adaptation (2024) +19. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22511-22521 (June 2023) +20. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +21. Liu, S.Y., Wang, C.Y., Yin, H., Molchanov, P., Wang, Y.C.F., Cheng, K.T., Chen, M.H.: Dora: Weight-decomposed low-rank adaptation (2024) +22. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) +23. Luo, S., Tan, Y., Patil, S., Gu, D., von Platen, P., Passos, A., Huang, L., Li, J., Zhao, H.: Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556 (2023) +24. Mou, C., Wang, X., Xie, L., Wu, Y., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023) +25. Park, T., Liu, M.Y., Wang, T.C., Zhu, J.Y.: Semantic image synthesis with spatially-adaptive normalization. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2332-2341. IEEE (2019) +26. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4195-4205 (2023) +27. Pernias, P., Rampas, D., Richter, M.L., Pal, C., Aubreville, M.: Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=gU58d5QeGv +28. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis (2023) +29. Qin, C., Zhang, S., Yu, N., Feng, Y., Yang, X., Zhou, Y., Wang, H., Niebles, J.C., Xiong, C., Savarese, S., et al.: Unicontrol: A unified diffusion model for controllable visual generation in the wild. arXiv preprint arXiv:2305.11147 (2023) +30. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +31. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022) + +32. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence 44(3), 1623-1637 (2020) +33. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +34. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023) +35. Ruiz, N., Li, Y., Jampani, V., Wei, W., Hou, T., Pritch, Y., Wadhwa, N., Rubinstein, M., Aberman, K.: Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949 (2023) +36. Ryu, S.: Low-rank adaptation for fast text-to-image diffusion fine-tuning, https://github.com/cloneofsimo/lora +37. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) +38. Shah, V., Ruiz, N., Cole, F., Lu, E., Lazebnik, S., Li, Y., Jampani, V.:Ziplora: Any subject in any style by effectively merging loras (2023) +39. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) +40. Witteveen, S., Andrews, M.: Investigating prompt engineering in diffusion models. arXiv preprint arXiv:2211.15462 (2022) +41. Xie, S., Tu, Z.: Holistically-nested edge detection. In: Proceedings of the IEEE international conference on computer vision. pp. 1395-1403 (2015) +42. Xu, X., Guo, J., Wang, Z., Huang, G., Essa, I., Shi, H.: Prompt-free diffusion: Taking" text" out of text-to-image diffusion models. arXiv preprint arXiv:2305.16223 (2023) +43. Xu, X., Wang, Z., Zhang, E., Wang, K., Shi, H.: Versatile diffusion: Text, images and variations all in one diffusion model. arXiv preprint arXiv:2211.08332 (2022) +44. Xue, Z., Song, G., Guo, Q., Liu, B., Zong, Z., Liu, Y., Luo, P.: Raphael: Text-to-image generation via large mixture of diffusion paths. arXiv preprint arXiv:2305.18295 (2023) +45. Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models (2023) +46. YEH, S.Y., Hsieh, Y.G., Gao, Z., Yang, B.B.W., Oh, G., Gong, Y.: Navigating text-to-image customization: From lyCORIS fine-tuning to model evaluation. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=wfzXa8e783 +47. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models (2023) +48. Zhang, Q., Chen, M., Bukharin, A., Karampatziakis, N., He, P., Cheng, Y., Chen, W., Zhao, T.: Adalora: Adaptive budget allocation for parameter-efficient fine-tuning (2023) +49. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018) + +50. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Unictrlnet: All-in-one control to text-to-image diffusion models. Advances in Neural Information Processing Systems (2023) \ No newline at end of file diff --git a/2024/CTRLorALTer_ Conditional LoRAdapter for Efficient 0-Shot Control & Altering of T2I Models/images.zip b/2024/CTRLorALTer_ Conditional LoRAdapter for Efficient 0-Shot Control & Altering of T2I Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..322ee7ee31cd46973a9e27749dd81c9fe0d9ca70 --- /dev/null +++ b/2024/CTRLorALTer_ Conditional LoRAdapter for Efficient 0-Shot Control & Altering of T2I Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:865c99a0447e22d5a46b2dc09c91e46ef68e4c493c0c7bc6be6ef8477237ac28 +size 531650 diff --git a/2024/CTRLorALTer_ Conditional LoRAdapter for Efficient 0-Shot Control & Altering of T2I Models/layout.json b/2024/CTRLorALTer_ Conditional LoRAdapter for Efficient 0-Shot Control & Altering of T2I Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..19917c00c0073a3d84e0c60deda3193d50faeba8 --- /dev/null +++ b/2024/CTRLorALTer_ Conditional LoRAdapter for Efficient 0-Shot Control & Altering of T2I Models/layout.json @@ -0,0 +1,9302 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 481, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 481, + 150 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 481, + 150 + ], + "type": "text", + "content": "CTRLorALTer: Conditional LoRAapter for Efficient 0-Shot Control & Altering of T2I Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 173, + 171, + 441, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 171, + 441, + 196 + ], + "spans": [ + { + "bbox": [ + 173, + 171, + 441, + 196 + ], + "type": "text", + "content": "Nick Stracke1, Stefan Andreas Baumann1, Joshua Susskind2, Miguel Angel Bautista2, and Björn Ommer1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 230, + 204, + 382, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 204, + 382, + 215 + ], + "spans": [ + { + "bbox": [ + 230, + 204, + 382, + 215 + ], + "type": "text", + "content": "CompVis @ LMU Munich, MCML" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 290, + 216, + 324, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 216, + 324, + 227 + ], + "spans": [ + { + "bbox": [ + 290, + 216, + 324, + 227 + ], + "type": "text", + "content": "2 Apple" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 257, + 452, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 257, + 452, + 376 + ], + "spans": [ + { + "bbox": [ + 160, + 257, + 452, + 376 + ], + "type": "text", + "content": "Abstract. Text-to-image generative models have become a prominent and powerful tool that excels at generating high-resolution realistic images. However, guiding the generative process of these models to take into account detailed forms of conditioning reflecting style and/or structure information remains an open problem. In this paper, we present LoRAAdapter, an approach that unifies both style and structure conditioning under the same formulation using a novel conditional LoRA block that enables zero-shot control. LoRAAdapter is an efficient and powerful approach to condition text-to-image diffusion models, which enables fine-grained control conditioning during generation and outperforms recent state-of-the-art approaches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 378, + 356, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 378, + 356, + 388 + ], + "spans": [ + { + "bbox": [ + 160, + 378, + 356, + 388 + ], + "type": "text", + "content": "Project page: compvis.github.io/LoRAAdapter/" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 399, + 444, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 399, + 444, + 410 + ], + "spans": [ + { + "bbox": [ + 160, + 399, + 444, + 410 + ], + "type": "text", + "content": "Keywords: T2I Models " + }, + { + "bbox": [ + 160, + 399, + 444, + 410 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 399, + 444, + 410 + ], + "type": "text", + "content": " Diffusion Models " + }, + { + "bbox": [ + 160, + 399, + 444, + 410 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 399, + 444, + 410 + ], + "type": "text", + "content": " Controllable Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 449, + 229, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 449, + 229, + 460 + ], + "spans": [ + { + "bbox": [ + 133, + 449, + 229, + 460 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 473, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 473, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 133, + 473, + 481, + 628 + ], + "type": "text", + "content": "Text-to-image models have become a foundational component of Computer Vision, especially with the introduction of large-scale diffusion models like Stable Diffusion (SD) [33], DALL-E 2 [31], Imagen [37], RAPHAEL [44] and, eDiff-I [2]. These models have empowered users to create realistic images from textual prompts, although crafting effective prompts is a complex task in itself due to the nuanced prompt engineering required [40] to effectively condition the image generation process. To tackle this problem, prior approaches such as SD Image Variations [1], Stable unCLIP [33], and Composer [15], have attempted to perform conditioning using image prompts to enhance the expressivity of the conditioning mechanism giving users more flexibility during generation. In practice, this family of approaches requires either training from scratch or fully fine-tuning large text-conditioned model weights, incurring high computational costs and requiring large datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 630, + 481, + 665 + ], + "type": "text", + "content": "To reduce training cost and time, recent approaches like ControlNet [47], T2I-Adapter [24], Uni-ControlNet [50] or IP-Adapters [45] have focused on designing adapters that keep the base model frozen and only introduce a limit number of" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 159, + 119, + 455, + 290 + ], + "blocks": [ + { + "bbox": [ + 159, + 119, + 455, + 290 + ], + "lines": [ + { + "bbox": [ + 159, + 119, + 455, + 290 + ], + "spans": [ + { + "bbox": [ + 159, + 119, + 455, + 290 + ], + "type": "image", + "image_path": "f47f35ce19fa3fb2627583f25816ce11b85a6c4d814339b6739cd38db5fc8cde.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 300, + 483, + 346 + ], + "lines": [ + { + "bbox": [ + 130, + 300, + 483, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 483, + 346 + ], + "type": "text", + "content": "Fig. 1: LoRAAdapter allows structure and style control of the image generation process of text-to-image models in a zero-shot manner. Our approach enables powerful fine-grained and efficient unified control over both structure and style conditioning using conditional LoRA blocks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 366, + 482, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 427 + ], + "type": "text", + "content": "new layers. This not only reduces the number of trainable parameters but also allows training on much smaller datasets because the frozen base model cannot suffer from catastrophic forgetting. These adapters are typically handcrafted to a specific model architecture and conditioning modality. They can be roughly assigned into one of two groups:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 432, + 482, + 528 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 137, + 432, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 432, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 137, + 432, + 482, + 479 + ], + "type": "text", + "content": "- Adapters for pixel-level fine-grained conditioning over the generated image. This is often referred to as structural or local conditioning since it allows influencing the generated image on a per-pixel level by e.g. providing a depth map and effectively setting the layout of the image [7,17,24,29,47,50]" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 479, + 482, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 479, + 482, + 528 + ], + "spans": [ + { + "bbox": [ + 137, + 479, + 482, + 528 + ], + "type": "text", + "content": "- Adapters for style or global conditioning perform tasks such as image-to-image translation similar to unCLIP [15, 19, 33, 43]. Here, the focus is not on fine-grained control but instead on generating an image similar to the prompt image, where similarity is defined by sharing the same style or semantics." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Empirically, adapters always favor one type of conditioning over the other. For example, both Uni-ControlNet [50] and T2I-Adapter [24] have attempted style conditioning but their performance is not comparable to leading style approaches [45]. Conversely, style adapters require ControlNet or T2I-Adapter for structure conditioning. However, ControlNet is a large adapter that copies the entire encoder of Stable Diffusion's U-Net which significantly increases compute and thus inference time. Our goal is to find a unified approach that is capable of both structure and style conditioning, providing complete control over the generated image. Formulating such a unified approach for conditioning on global controls like style and on local controls like structure, in an efficient and generic manner remains a key open problem." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Stracke et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "To address this challenging problem, we introduce LoRAAdapter (see Fig. 1), a unified approach for incorporating images to condition both the style and structure of the generated images. LoRAAdapter takes inspiration from Low-Rank-Adaptations (LoRAs) [14], which have shown remarkable performance for various tasks in diffusion models, such as distillation or learning specific concepts [23,35]. In particular, LoRAs low-rank property naturally regularizes the conditioning while offering unparalleled flexibility as any layer can be adapted in an architecture-agnostic way. Thus far, however, LoRAs have been implemented as fixed adaptations, being independent of the input. This means layer adaptation cannot change at test time for a given conditioning and, therefore, cannot be used for zero-shot generalization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 247, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 344 + ], + "type": "text", + "content": "LoRAAdapter is a novel approach to adding conditional information to LoRAs, enabling zero-shot generalization and making them applicable for both structure and style and possibly many other conditioning types. LoRAAdapter is compact and efficient, e.g., optimizing 16M parameters vs the 22M of IP-Adapters [45] or 361M of ControlNet [47], while outperforming recent adapter approaches [45] and even approaches that train models from scratch (see Tab. 1). LoRAAdapter trains LoRA [14] modules for different blocks across the base network and is efficient both during training and inference. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 354, + 479, + 423 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 354, + 479, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 479, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 479, + 376 + ], + "type": "text", + "content": "- We propose LoRAAdapter, a generic approach to train conditional LoRAs that is agnostic to model architecture and conditioning modality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 378, + 479, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 378, + 479, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 378, + 479, + 399 + ], + "type": "text", + "content": "- We implement LoRAAdapter for Stable Diffusion, offering a unified conditioning mechanism for both style and structure, enabling zero-shot conditioning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 403, + 479, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 479, + 423 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 479, + 423 + ], + "type": "text", + "content": "- We show the effectiveness of this approach, outperforming dedicated adapters for either style and structure on various metrics." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 445, + 237, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 445, + 237, + 458 + ], + "spans": [ + { + "bbox": [ + 132, + 445, + 237, + 458 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 471, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 481, + 495 + ], + "type": "text", + "content": "We now review existing adapter approaches for style and structure and discuss the current role of LoRAs in diffusion models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 513, + 259, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 513, + 259, + 525 + ], + "spans": [ + { + "bbox": [ + 131, + 513, + 259, + 525 + ], + "type": "text", + "content": "2.1 Structure Adapters" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Structure adapters are specifically designed to operate on various types of local conditioning modalities such as Depth, HED [41], Canny Edges [4], Scribbles, or key poses [5]. The goal of these adapters is to spatially align the generated image with the provided condition. One of the most prominent approaches to tackle this problem is ControlNet [47] for Stable Diffusion. ControlNet creates a copy of the U-Net's encoder and combines the skip connections additively with the original skip connections, which then get fed into the original decoder. While obtaining great performance, ControlNet adds a lot of computational overhead as the forward pass includes a second encoder which can increase inference speed by up to " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ". Another consequence of the large capacity is that ControlNet tends to interpret the structure map directly. This can be desirable when sampling" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 353, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 353, + 91, + 447, + 100 + ], + "type": "text", + "content": "2. RELATED WORK" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "images without other conditioning via the text prompt or other adapters, but can also lead to entanglement of structure and style." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 272 + ], + "type": "text", + "content": "Following a similar idea, T2I-Adapter [24] also modifies skip connections while using a much smaller encoder network. To extend conditioning to multiple modalities, Uni-ControlNet [50] was proposed as an alternative solution to having to train multiple separate and modality-specific ControlNets. Uni-ControlNet trains a single model for multiple conditions by concatenating the structure maps along the channel dimension and using the concatenated tensor as input. Uni-ControlNet uses SPADE [25] to combine skip-connections instead of the simple addition used by ControlNet. SCEdit [17] proposed a similar but more efficient solution for conditioning on multiple modalities by integrating a small tuner network (SC-Tuner) between skip-connections which removes the need for a separate encoder network akin to ControlNet." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 289, + 236, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 289, + 236, + 300 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 236, + 300 + ], + "type": "text", + "content": "2.2 Style Adapters" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 307, + 482, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 482, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 482, + 378 + ], + "type": "text", + "content": "Style adapters are used as an alternative to unCLIP [33] like models, which are directly trained to invert CLIP image embeddings. The main advantage of using a style adapter as opposed to an unCLIP model is that the original text conditioning is preserved, enabling additional control over the generated images. Additionally, all recently released large open-source diffusion models only offered text conditioning, making an adapter approach necessary [27, 28]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 379, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 472 + ], + "type": "text", + "content": "IP-Adapter [45] adds a decoupled cross-attention layer that operates on four image tokens instead of text tokens. The resulting activations of this decoupled cross-attention layer are scaled and added to the original cross-attention activations. This results in a lightweight adapter that integrates well with potential text conditioning and generates images faithful to the image prompt. In contrast, SeeCoder [42] trains an encoder with 2D spatial embedding to convert image tokens to text tokens and completely replaces the standard text tokens with their tokens." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 475, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 547 + ], + "type": "text", + "content": "Structure adapters have also attempted to include style conditioning. UniControlNet [50] includes a separate mapping network to convert a pooled CLIP image token to four CLIP text tokens, which they concatenate to the original text embeddings. ControlNet-shuffle [47] shuffles an RGB image and uses that as conditioning similar to its structure conditioning. However, the performance of these approaches does not compare favorably to dedicated style adapters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 563, + 298, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 563, + 298, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 563, + 298, + 574 + ], + "type": "text", + "content": "2.3 LoRAs in Diffusion Models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "While LoRAs were initially introduced as an efficient alternative to fine-tuning large language models [14], they were quickly adopted for diffusion models [36] as a relatively efficient and cheap option to adjust their generation capabilities. Combining LoRAs with methods such as Dreambooth [34] performs on par with a full fine-tuning approach while only introducing a fraction of trainable parameters. LoRAs are commonly trained on a small set of images with the goal of capturing a specific style instead of a subject. ZipLoRA [38] proposes a merging algorithm to" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "Stracke et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "combine multiple LoRAs gracefully by minimizing the cosine similarity of LoRAs that adapt the same weight matrix. LyCORIS [46] trained an entire library of specific LoRAs in diffusion models with various benefits. [23] showed that LoRAs are even powerful enough for more complex objectives such as consistency distillation. Finally, Concept Sliders [8] use LoRAs to train disentangled edit directions to edit images. They scale the LoRA in the negative and positive directions to either increase or decrease a given attribute." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 219, + 233, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 219, + 233, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 219, + 233, + 232 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 244, + 206, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 244, + 206, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 244, + 206, + 255 + ], + "type": "text", + "content": "3.1 Diffusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 265, + 482, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 482, + 338 + ], + "type": "text", + "content": "Diffusion models represent a category of generative models characterized by a two-stage process: a forward diffusion that incrementally introduces Gaussian noise over a series of " + }, + { + "bbox": [ + 130, + 265, + 482, + 338 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 265, + 482, + 338 + ], + "type": "text", + "content": " steps, and a reverse denoising process that reconstructs samples from this noise. In text-to-image models, diffusion can be directed by supplementary inputs like text descriptions. The primary training objective, " + }, + { + "bbox": [ + 130, + 265, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 130, + 265, + 482, + 338 + ], + "type": "text", + "content": ", aimed at noise prediction, simplifies the variational lower bound as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 176, + 346, + 482, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 346, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 176, + 346, + 482, + 373 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {s i m p l e}} = \\mathbb {E} _ {\\boldsymbol {x} _ {0}, \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}), \\boldsymbol {c}, t} \\| \\boldsymbol {\\epsilon} - \\epsilon_ {\\theta} \\big (\\underbrace {\\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}} _ {\\boldsymbol {x} _ {t}}, \\boldsymbol {c}, t \\big) \\| ^ {2}, \\tag {1}", + "image_path": "e70730658f014f40512f5b2f1623e523bba300a33cb8b0de6e5150e73d8aac7b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 379, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 403 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 379, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 130, + 379, + 482, + 403 + ], + "type": "text", + "content": " is the original data influenced by condition " + }, + { + "bbox": [ + 130, + 379, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\pmb{c}" + }, + { + "bbox": [ + 130, + 379, + 482, + 403 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 379, + 482, + 403 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 379, + 482, + 403 + ], + "type": "text", + "content": " is the diffusion timestep, and " + }, + { + "bbox": [ + 130, + 379, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 130, + 379, + 482, + 403 + ], + "type": "text", + "content": " represents the data at step " + }, + { + "bbox": [ + 130, + 379, + 482, + 403 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 379, + 482, + 403 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 379, + 482, + 403 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 130, + 379, + 482, + 403 + ], + "type": "text", + "content": " dictating the noise level." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "content": "During sampling, an initial pure Gaussian latent " + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\pmb{x}_T" + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "content": " is iteratively denoised by the diffusion model to obtain a sample " + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "content": " from the modeled (conditional) distribution. Conditional models often employ classifier guidance to fine-tune the balance between accuracy and diversity, with classifier-free guidance [12] serving as a preferred alternative that merges conditional and unconditional noise predictions during training and sampling:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 211, + 486, + 481, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 486, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 211, + 486, + 481, + 498 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {\\epsilon}} _ {\\theta} \\left(\\boldsymbol {x} _ {t}, \\boldsymbol {c}, t\\right) = w \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {x} _ {t}, \\boldsymbol {c}, t\\right) + (1 - w) \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {x} _ {t}, t\\right), \\tag {2}", + "image_path": "5c793e59e3a9d089b1b654beee756cc4dcb94d50467811a6e6b4f54fa58f1260.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 507, + 482, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 482, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 482, + 543 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 507, + 482, + 543 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 507, + 482, + 543 + ], + "type": "text", + "content": " modulates the adherence to condition " + }, + { + "bbox": [ + 130, + 507, + 482, + 543 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 507, + 482, + 543 + ], + "type": "text", + "content": ". This technique is pivotal for text-to-image models, enhancing the alignment between generated images and text prompts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 561, + 315, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 561, + 315, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 561, + 315, + 574 + ], + "type": "text", + "content": "3.2 Low-Rank Adaptation (LoRA)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "text", + "content": "Low-Rank Adaptation (LoRA) [14] is a method for finetuning pre-trained foundation models by keeping the original weight matrices frozen and instead adding a new set of low-rank weight matrices. While this method was originally proposed in the context of large language models, it has been recently adopted in the domain of diffusion models [6,8,23]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "Specifically, the original weight matrix of a layer in the base model " + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "inline_equation", + "content": "W_0 \\in \\mathbb{R}^{d \\times k}" + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": " is kept frozen and is adapted by a low-rank matrix " + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "inline_equation", + "content": "W + \\Delta W = W + BA" + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": " with" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 356, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 356, + 91, + 447, + 101 + ], + "type": "text", + "content": "3. PRELIMINARIES" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 482, + 140 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{d \\times r}" + }, + { + "bbox": [ + 132, + 115, + 482, + 140 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 115, + 482, + 140 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{r \\times k}" + }, + { + "bbox": [ + 132, + 115, + 482, + 140 + ], + "type": "text", + "content": ". Crucially, " + }, + { + "bbox": [ + 132, + 115, + 482, + 140 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 132, + 115, + 482, + 140 + ], + "type": "text", + "content": " is chosen following " + }, + { + "bbox": [ + 132, + 115, + 482, + 140 + ], + "type": "inline_equation", + "content": "r \\ll \\min(d, k)" + }, + { + "bbox": [ + 132, + 115, + 482, + 140 + ], + "type": "text", + "content": ". The forward pass for the adapted layer is then:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 232, + 146, + 482, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 146, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 232, + 146, + 482, + 159 + ], + "type": "interline_equation", + "content": "h = W _ {0} x + \\Delta W x = W _ {0} x + B A x. \\tag {3}", + "image_path": "c550ba3b9d5eca339fae65d5504a3fbe682ce8dce8ae836faa2717c1bbe4c2fe.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 165, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 165, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 165, + 482, + 213 + ], + "type": "text", + "content": "Typically, " + }, + { + "bbox": [ + 130, + 165, + 482, + 213 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 165, + 482, + 213 + ], + "type": "text", + "content": " is zero-initialized, so the LoRA does not affect the model initially. This yields a generic approach, as we can freely choose the rank " + }, + { + "bbox": [ + 130, + 165, + 482, + 213 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 165, + 482, + 213 + ], + "type": "text", + "content": " and the weights according to the given task. This happens on a per-layer basis, i.e., if we adapt " + }, + { + "bbox": [ + 130, + 165, + 482, + 213 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 165, + 482, + 213 + ], + "type": "text", + "content": " layers, we will have " + }, + { + "bbox": [ + 130, + 165, + 482, + 213 + ], + "type": "inline_equation", + "content": "nA" + }, + { + "bbox": [ + 130, + 165, + 482, + 213 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 165, + 482, + 213 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 165, + 482, + 213 + ], + "type": "text", + "content": " matrices." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 202, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 202, + 242 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 202, + 242 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 251, + 482, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 335 + ], + "type": "text", + "content": "There are two kinds of approaches for conditioning text-to-image diffusion models: style or global conditioning [42, 45], and structure or local conditioning [17, 24, 29, 50]. These approaches are specifically tailored to a specific conditioning objective and therefore missing a holistic approach that can efficiently learn both style and structure. We now describe LoRAAdapter, an efficient, holistic, and customizable method for introducing both style and structure conditioning into text-to-image diffusion models." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 219, + 358, + 396, + 492 + ], + "blocks": [ + { + "bbox": [ + 219, + 358, + 396, + 492 + ], + "lines": [ + { + "bbox": [ + 219, + 358, + 396, + 492 + ], + "spans": [ + { + "bbox": [ + 219, + 358, + 396, + 492 + ], + "type": "image", + "image_path": "096c5c3ab9630367a314d2ecf880751abe6c2e22ac93da39af88171be02d1a26.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 505, + 482, + 586 + ], + "lines": [ + { + "bbox": [ + 130, + 505, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 505, + 482, + 586 + ], + "type": "text", + "content": "Fig. 2: Overview of the proposed conditional LoRA block. The original weight matrix " + }, + { + "bbox": [ + 130, + 505, + 482, + 586 + ], + "type": "inline_equation", + "content": "W_0^{(i)}" + }, + { + "bbox": [ + 130, + 505, + 482, + 586 + ], + "type": "text", + "content": " is frozen while all other layers are trained. " + }, + { + "bbox": [ + 130, + 505, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 505, + 482, + 586 + ], + "type": "text", + "content": " is an affine transformation that operates on the low-dimensional embedding " + }, + { + "bbox": [ + 130, + 505, + 482, + 586 + ], + "type": "inline_equation", + "content": "A^{(i)}x" + }, + { + "bbox": [ + 130, + 505, + 482, + 586 + ], + "type": "text", + "content": " and introduces the conditioning. The local mapper network " + }, + { + "bbox": [ + 130, + 505, + 482, + 586 + ], + "type": "inline_equation", + "content": "m_L^{(i)}" + }, + { + "bbox": [ + 130, + 505, + 482, + 586 + ], + "type": "text", + "content": " predicts the scale and shift parameters " + }, + { + "bbox": [ + 130, + 505, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\beta, \\gamma" + }, + { + "bbox": [ + 130, + 505, + 482, + 586 + ], + "type": "text", + "content": " for the affine transformation. Typically, we set " + }, + { + "bbox": [ + 130, + 505, + 482, + 586 + ], + "type": "inline_equation", + "content": "m_L^{(i)}" + }, + { + "bbox": [ + 130, + 505, + 482, + 586 + ], + "type": "text", + "content": " to be a small network. If complex transformations are required to map the conditioning " + }, + { + "bbox": [ + 130, + 505, + 482, + 586 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 505, + 482, + 586 + ], + "type": "text", + "content": ", this happens in " + }, + { + "bbox": [ + 130, + 505, + 482, + 586 + ], + "type": "inline_equation", + "content": "m_S" + }, + { + "bbox": [ + 130, + 505, + 482, + 586 + ], + "type": "text", + "content": " since it is shared across all adapted layers." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 624, + 259, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 259, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 259, + 635 + ], + "type": "text", + "content": "4.1 Conditional LoRAs" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "LoRAs have successfully enabled adaptation of diffusion models across a range of architectures [6,23,35] to steer their generation process. In this setting, LoRAs" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "Stracke et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "have been considered as free parameters and optimized on a training set of images exhibiting a specific characteristic." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 247 + ], + "type": "text", + "content": "Instead of learning a single LoRA to apply one specific modification, e.g., changing the style of the generated image to a painting or learning to represent one specific subject [34,35], we propose a LoRA-based conditioning mechanism whose behavior changes based on conditioning provided at inference time, enabling zero-shot generalization. While previous methods depend on specific aspects of the architectures of the models they are applied to [45, 47], using LoRAs for conditioning enables us to efficiently adapt foundation models for conditional tasks in an architecture-agnostic way. Our LoRA-based conditioning mechanism is shown in Fig. 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 248, + 481, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 481, + 295 + ], + "type": "text", + "content": "Starting from the standard LoRA setup as in Eq. (3), we propose implementing conditioning of the LoRA by applying a transformation " + }, + { + "bbox": [ + 130, + 248, + 481, + 295 + ], + "type": "inline_equation", + "content": "\\phi(Ax|m(c))" + }, + { + "bbox": [ + 130, + 248, + 481, + 295 + ], + "type": "text", + "content": " to the low-dimensional intermediate embedding in the LoRA that introduces conditional behavior based on the conditioning " + }, + { + "bbox": [ + 130, + 248, + 481, + 295 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 248, + 481, + 295 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 216, + 304, + 481, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 304, + 481, + 316 + ], + "spans": [ + { + "bbox": [ + 216, + 304, + 481, + 316 + ], + "type": "interline_equation", + "content": "h = W _ {0} x + \\Delta W x = W _ {0} x + B \\phi (A x | m (c)). \\tag {4}", + "image_path": "6668ce0e8ae174be1e708a89c12dd6358f04c80a56776b566534a597f555f9a6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 324, + 481, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 481, + 407 + ], + "type": "text", + "content": "This limits the conditioning to apply only to a low-rank subspace of the original vector space the weight matrix operates in. This is crucial for the generalization capabilities of LoRAs, as it introduces regularization and thus limits the adaptation to only focus on relevant aspects instead of spurious correlations from the dataset it was trained on, which results in more efficient training. By introducing our conditional adaptation of the model's behavior in such a low-rank space, our method can also benefit from those advantages." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 407, + 487, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 487, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 487, + 443 + ], + "type": "text", + "content": "In our framework for conditional LoRAs, we can theoretically choose " + }, + { + "bbox": [ + 130, + 407, + 487, + 443 + ], + "type": "inline_equation", + "content": "\\phi(Ax|m(c))" + }, + { + "bbox": [ + 130, + 407, + 487, + 443 + ], + "type": "text", + "content": " to be arbitrary, non-linear transformations, but we find that a simple affine transformation is sufficiently expressive for a wide range of applications:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 216, + 453, + 481, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 453, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 216, + 453, + 481, + 464 + ], + "type": "interline_equation", + "content": "\\phi (A x \\mid m (c)) = \\gamma_ {\\phi} (m (c)) \\odot A x + \\beta_ {\\phi} (m (c)), \\tag {5}", + "image_path": "bdb54112b56d586173bb9bb9708c6cac72f0e667ae5ba9cd4b477e5cca8dfc0a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 472, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 481, + 495 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 130, + 472, + 481, + 495 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 472, + 481, + 495 + ], + "type": "text", + "content": " denoting an elementwise (Hadamard) product and " + }, + { + "bbox": [ + 130, + 472, + 481, + 495 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 472, + 481, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 472, + 481, + 495 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 472, + 481, + 495 + ], + "type": "text", + "content": " referring to the scale and shift factors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 496, + 481, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 481, + 581 + ], + "type": "text", + "content": "To predict the " + }, + { + "bbox": [ + 130, + 496, + 481, + 581 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 496, + 481, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 496, + 481, + 581 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 496, + 481, + 581 + ], + "type": "text", + "content": " from the conditioning " + }, + { + "bbox": [ + 130, + 496, + 481, + 581 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 496, + 481, + 581 + ], + "type": "text", + "content": ", we utilize a mapping network. This is generally a small neural network that we separate into a shared part " + }, + { + "bbox": [ + 130, + 496, + 481, + 581 + ], + "type": "inline_equation", + "content": "m_{S}(\\cdot)" + }, + { + "bbox": [ + 130, + 496, + 481, + 581 + ], + "type": "text", + "content": " and a layer-specific part " + }, + { + "bbox": [ + 130, + 496, + 481, + 581 + ], + "type": "inline_equation", + "content": "m_{L}^{(i)}(\\cdot)" + }, + { + "bbox": [ + 130, + 496, + 481, + 581 + ], + "type": "text", + "content": " that is individually learned for each adapted layer. For both parts, we generally use very small neural networks, which is in stark contrast to other works such as ControlNet [47] that uses a full copy of the U-Net's encoder (see Tab. 1). For brevity, we will always refer to both parts simply as " + }, + { + "bbox": [ + 130, + 496, + 481, + 581 + ], + "type": "inline_equation", + "content": "m(\\cdot)" + }, + { + "bbox": [ + 130, + 496, + 481, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "This general formulation directly enables both local and global conditioning, as it can be applied to the convolutional and attention layers in a standard diffusion model. For local conditioning, the mapping network can supply spatially dependent LoRA modulations. At the same time, it can also be used to supply global conditioning by introducing non-spatially dependent information, such as by adapting the cross-attention projections. Concurrent works [18, 21, 48] also introduced improvements to the basic LoRA architecture. As LoRAAdapter does" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 389, + 91, + 447, + 100 + ], + "type": "text", + "content": "4. METHOD" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "not depend on a specific implementation of LoRAs but just on the low-rank intermediate sub-space, these improvements can be directly transferred to our model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 171, + 332, + 183 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 171, + 332, + 183 + ], + "spans": [ + { + "bbox": [ + 131, + 171, + 332, + 183 + ], + "type": "text", + "content": "4.2 A Unified Conditioning Approach" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 191, + 481, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 191, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 481, + 216 + ], + "type": "text", + "content": "In this section, we provide implementation details for applying our conditional LoRAs to the layer types commonly used in large-scale diffusion models." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 132, + 236, + 293, + 280 + ], + "blocks": [ + { + "bbox": [ + 132, + 236, + 293, + 280 + ], + "lines": [ + { + "bbox": [ + 132, + 236, + 293, + 280 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 293, + 280 + ], + "type": "image", + "image_path": "6973291b2a25e51354fbde26ccae2dc3e154d4d904d75f6f26b72638460445a7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 290, + 292, + 316 + ], + "lines": [ + { + "bbox": [ + 131, + 290, + 292, + 316 + ], + "spans": [ + { + "bbox": [ + 131, + 290, + 292, + 316 + ], + "type": "text", + "content": "(a) In attention layers, we apply an affine transformation on the " + }, + { + "bbox": [ + 131, + 290, + 292, + 316 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 290, + 292, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 290, + 292, + 316 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 131, + 290, + 292, + 316 + ], + "type": "text", + "content": " projection matrix per channel for every token." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 325, + 240, + 478, + 289 + ], + "blocks": [ + { + "bbox": [ + 325, + 240, + 478, + 289 + ], + "lines": [ + { + "bbox": [ + 325, + 240, + 478, + 289 + ], + "spans": [ + { + "bbox": [ + 325, + 240, + 478, + 289 + ], + "type": "image", + "image_path": "19b06692c5434f81e10f540612515df71cd7da8d80dc071b2b9afd8793819a09.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 290, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 321, + 290, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 321, + 290, + 482, + 316 + ], + "type": "text", + "content": "(b) To add structural conditioning to convolutional layers, we perform an affine transformation for every element similar to [25]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 325, + 479, + 336 + ], + "lines": [ + { + "bbox": [ + 132, + 325, + 479, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 479, + 336 + ], + "type": "text", + "content": "Fig. 3: Visualization of implementations of our conditional LoRAs for specific layers." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 379, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 439 + ], + "type": "text", + "content": "Attention Layers Many diffusion models, such as Stable Diffusion " + }, + { + "bbox": [ + 130, + 379, + 482, + 439 + ], + "type": "inline_equation", + "content": "\\{1,2,\\mathrm{XL}\\}" + }, + { + "bbox": [ + 130, + 379, + 482, + 439 + ], + "type": "text", + "content": " [28,33] or Imagen [37] include self- and cross-attention layers that operate on image (and text) tokens, with transformer-based diffusion models such as [26] even purely relying on them for spatial computations. The output " + }, + { + "bbox": [ + 130, + 379, + 482, + 439 + ], + "type": "inline_equation", + "content": "h^\\prime" + }, + { + "bbox": [ + 130, + 379, + 482, + 439 + ], + "type": "text", + "content": " of the attention layers is defined as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 201, + 447, + 482, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 447, + 482, + 476 + ], + "spans": [ + { + "bbox": [ + 201, + 447, + 482, + 476 + ], + "type": "interline_equation", + "content": "h ^ {\\prime} = \\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) V \\tag {6}", + "image_path": "4d79c770de18ef442ea7b54425a74fbf776b555b7d794953722e51e7f5ec29d5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "text", + "content": "with the projections " + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "inline_equation", + "content": "Q = hW_{q}" + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "inline_equation", + "content": "K = hW_{k}" + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "inline_equation", + "content": "K = c_{t}W_{k}" + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "inline_equation", + "content": "V = hW_{v}" + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "inline_equation", + "content": "V = c_{t}W_{v}" + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "text", + "content": ", depending on whether it is self- or cross-attention, and the text tokens " + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "text", + "content": ". Here, we apply our conditional LoRA on the linear projections " + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "inline_equation", + "content": "W_{k}" + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "inline_equation", + "content": "W_{v}" + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "text", + "content": " based on Equation 4. We show a diagram of conditional LoRAs for attention layers in Fig. 3a." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 561, + 481, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 561, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 561, + 481, + 586 + ], + "type": "text", + "content": "Convolutional Layers For convolutional layers, which form the main backbone of many diffusion models [27,33], we adapt the convolutional layer" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 279, + 595, + 482, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 595, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 279, + 595, + 482, + 609 + ], + "type": "interline_equation", + "content": "h = K _ {0} \\star x, \\tag {7}", + "image_path": "ecf0e61cc05e26e872d7198809b2935ec3a8a0811f588d92bd0972cc0b77e79a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "with the 2D convolution kernel " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " having input channel count " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "ch_{i}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " and output channel count " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "ch_{o}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", in a similar manner as linear projections. However, in this setting we work on convolution kernels instead of matrices: we perform one convolution " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "K_{A}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " with the same kernel size, stride, and padding as the original" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "Stracke et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 157, + 482, + 322 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Quantitative comparison for style conditioning of our proposed LoRAAdapter with other methods on the COCO validation set with four samples for every image. The best results are in bold (adapted from [45])." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 157, + 482, + 322 + ], + "lines": [ + { + "bbox": [ + 135, + 157, + 482, + 322 + ], + "spans": [ + { + "bbox": [ + 135, + 157, + 482, + 322 + ], + "type": "table", + "html": "
Style MethodReusable to custom modelsNative structure controlMultimodal promptsTrainable parametersCLIP-T ↑CLIP-I ↑
Training from scratch
Open unCLIPXXX893M0.6080.858
Kandinsky-2-1XXX1229M0.5990.855
Versatile DiffusionXX✓860M0.5870.830
Fine-tunining from text-to-image model
SD Image VariationsXXX860M0.5480.760
SD unCLIPXXX870M0.5840.810
Adapters
Uni-ControlNet (Global Control)✓✓✓47M0.5060.736
T2I-Adapter (Style)✓✓✓39M0.4850.648
ControlNet Shuffle✓✓✓361M0.4210.616
IP-Adapter✓X✓22M0.5880.828
LoRAAdapter (ours)✓✓✓16M0.6370.831
LoRAAdapter SDXL (ours)✓✓✓103M0.6490.849
", + "image_path": "ca5a72631ba41370222f12e19d8f6a14ac38c0c282823a8c20f954c659a2eee0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 346, + 482, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 395 + ], + "type": "text", + "content": "convolution " + }, + { + "bbox": [ + 130, + 346, + 482, + 395 + ], + "type": "inline_equation", + "content": "K_{0}" + }, + { + "bbox": [ + 130, + 346, + 482, + 395 + ], + "type": "text", + "content": ", which reduces the channel count from " + }, + { + "bbox": [ + 130, + 346, + 482, + 395 + ], + "type": "inline_equation", + "content": "ch_{i}" + }, + { + "bbox": [ + 130, + 346, + 482, + 395 + ], + "type": "text", + "content": " to the rank " + }, + { + "bbox": [ + 130, + 346, + 482, + 395 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 346, + 482, + 395 + ], + "type": "text", + "content": " of the LoRA while keeping all else equal. Then, we apply our LoRA conditioning " + }, + { + "bbox": [ + 130, + 346, + 482, + 395 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot|m(c))" + }, + { + "bbox": [ + 130, + 346, + 482, + 395 + ], + "type": "text", + "content": " in this bottlenecked space and expand to the output channel count " + }, + { + "bbox": [ + 130, + 346, + 482, + 395 + ], + "type": "inline_equation", + "content": "ch_{o}" + }, + { + "bbox": [ + 130, + 346, + 482, + 395 + ], + "type": "text", + "content": " using a pointwise convolution " + }, + { + "bbox": [ + 130, + 346, + 482, + 395 + ], + "type": "inline_equation", + "content": "K_{B}" + }, + { + "bbox": [ + 130, + 346, + 482, + 395 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 227, + 405, + 482, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 405, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 227, + 405, + 482, + 418 + ], + "type": "interline_equation", + "content": "h = K _ {0} \\star x + K _ {B} \\star \\phi \\left(K _ {A} \\star x | m (c)\\right). \\tag {8}", + "image_path": "0e7717776206b42ef730b5a6be4f6e38bd5671f868a9a0db620876d4396af720.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 428, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 428, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 428, + 482, + 452 + ], + "type": "text", + "content": "A visual diagram of conditional LoRAs for convolutional layers is shown in Fig. 3b." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 474, + 230, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 474, + 230, + 488 + ], + "spans": [ + { + "bbox": [ + 131, + 474, + 230, + 488 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 502, + 262, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 262, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 262, + 514 + ], + "type": "text", + "content": "5.1 Experimental Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 525, + 482, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 573 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 573 + ], + "type": "text", + "content": "Data We train LoRAdapter on a 40 million samples subset of COYO-700M [3] that only contains images with a short side of at least 512 pixels. Larger images are downscaled and center-cropped to 512 pixels. For text-conditioning, we use the standard prompts provided by the COYO-700M dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Training and Implementation Details All experiments are based on Stable Diffusion 1.5 [33] unless noted otherwise and adapted with a conditional LoRA according to the given task. We train our adapter for 50,000 steps on 32 A100 GPUs with a global batch size of 256 and a learning rate of 0.0001. We use AdamW [22] as an optimizer and drop each conditioning with a probability of 0.05. This includes the original text conditioning." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "type": "text", + "content": "5. EXPERIMENTS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 187, + 135, + 424, + 216 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 126 + ], + "type": "text", + "content": "Table 2: Comparison of a smaller and a larger version of our models vs. ControlNet." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 135, + 424, + 216 + ], + "lines": [ + { + "bbox": [ + 187, + 135, + 424, + 216 + ], + "spans": [ + { + "bbox": [ + 187, + 135, + 424, + 216 + ], + "type": "table", + "html": "
ModelParamsMSE-d ↓FID ↓LPIPS ↓
ControlNet361M17.9817.6440.622
Uni-ControlNet361M17.5616.4530.601
T2I-Adapter39M24.7817.9410.636
LoRApapter-A (Ours)17M17.2716.8490.599
LoRApapter-B (Ours)32M15.3415.6700.572
", + "image_path": "5c3b814cd534f4f9fac50b64ec730792feb230c6eda4958d73d2bada52960839.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 238, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 238, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 482, + 361 + ], + "type": "text", + "content": "For style conditioning, we experiment with both CLIP ViT-L/14 [30] by OpenAI and CLIP ViT-H/14 by OpenClip [16] as the image encoder but found that the ViT-H/14 tends to produce more visually appealing results. We still use ViT-L/14 for some ablation experiments as it results in a smaller model. Similar to [45], we use a single linear layer followed by layer normalization for the shared mapping network " + }, + { + "bbox": [ + 130, + 238, + 482, + 361 + ], + "type": "inline_equation", + "content": "m_S" + }, + { + "bbox": [ + 130, + 238, + 482, + 361 + ], + "type": "text", + "content": " as it outperformed much larger mapping networks and showed faster convergence. This maps the pooled CLIP image token " + }, + { + "bbox": [ + 130, + 238, + 482, + 361 + ], + "type": "inline_equation", + "content": "c_{img}" + }, + { + "bbox": [ + 130, + 238, + 482, + 361 + ], + "type": "text", + "content": ", keeping its original dimensionality " + }, + { + "bbox": [ + 130, + 238, + 482, + 361 + ], + "type": "inline_equation", + "content": "d_{img}" + }, + { + "bbox": [ + 130, + 238, + 482, + 361 + ], + "type": "text", + "content": ". The local mapping network " + }, + { + "bbox": [ + 130, + 238, + 482, + 361 + ], + "type": "inline_equation", + "content": "m_l^{(i)}" + }, + { + "bbox": [ + 130, + 238, + 482, + 361 + ], + "type": "text", + "content": " consists of two separate linear layers that predict " + }, + { + "bbox": [ + 130, + 238, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 238, + 482, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 238, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 238, + 482, + 361 + ], + "type": "text", + "content": " in the correct dimensionality " + }, + { + "bbox": [ + 130, + 238, + 482, + 361 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 238, + 482, + 361 + ], + "type": "text", + "content": " of the LoRA embedding. In total, the shared mapping network only has 1M parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 374, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 495 + ], + "type": "text", + "content": "For structure conditioning, the shared mapping network is a series of convolutional layers that spatially align the dimensions of the conditioning images with the dimensions that Stable Diffusion's U-Net expects which sums up to only 1.3M parameters. As the U-Net includes several down- and upsampling blocks, the mapping network outputs feature maps at various resolutions that match the specific U-Net blocks. Each individual LoRA only gets the single feature map that matches the spatial dimensionality of the U-Net block that it operates in. This feature map is then further processed with a single convolutional layer as a local mapping to align the number of channels with the rank " + }, + { + "bbox": [ + 130, + 374, + 482, + 495 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 374, + 482, + 495 + ], + "type": "text", + "content": " of the LoRA embedding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 514, + 483, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 483, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 483, + 575 + ], + "type": "text", + "content": "Inference We sample with 50 DDIM [39] steps and apply classifier-free guidance (CFG) [13] with a scale of 7.5 to the prompt and style conditioning. We do not apply any CFG for structure conditioning. As with all LoRAs, we can also influence the strength of our LoRAAdapter by adjusting the LoRA scale " + }, + { + "bbox": [ + 130, + 514, + 483, + 575 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 514, + 483, + 575 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 130, + 514, + 483, + 575 + ], + "type": "inline_equation", + "content": "h = W_0 + \\lambda BAx" + }, + { + "bbox": [ + 130, + 514, + 483, + 575 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 130, + 514, + 483, + 575 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + }, + { + "bbox": [ + 130, + 514, + 483, + 575 + ], + "type": "text", + "content": " unless otherwise noted." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 595, + 282, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 595, + 282, + 607 + ], + "spans": [ + { + "bbox": [ + 132, + 595, + 282, + 607 + ], + "type": "text", + "content": "5.2 Quantitative Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "To have a fair comparison with previous approaches, we follow previous literature on adapters [24,29,45] and evaluate our model on the validation set of COCO2017 [20] which contains 5,000 images. As every image comes with multiple prompts, we randomly select one prompt and use that for sampling." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Stracke et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "Style Following [45], we sample four images for a given image and text prompt pair. To measure the effectiveness of the new style conditioning, we use CLIP ViT-L/14 to obtain the image embeddings of the generated and the prompt image and compute the cosine similarity between them (CLIP-I). In addition, we quantify how well the generated image still follows the text prompt by calculating the CLIP score [10] between the text prompt and the generated image (CLIP-T). In Tab. 1 we show results across different approaches that either train models from scratch or fine-tune existing models, as well as compare with recent state-of-the-art adapters." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 144, + 255, + 469, + 437 + ], + "blocks": [ + { + "bbox": [ + 267, + 246, + 347, + 256 + ], + "lines": [ + { + "bbox": [ + 267, + 246, + 347, + 256 + ], + "spans": [ + { + "bbox": [ + 267, + 246, + 347, + 256 + ], + "type": "text", + "content": "Image Conditioning" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 144, + 255, + 469, + 437 + ], + "lines": [ + { + "bbox": [ + 144, + 255, + 469, + 437 + ], + "spans": [ + { + "bbox": [ + 144, + 255, + 469, + 437 + ], + "type": "image", + "image_path": "35c648de9a945bb5b524673399e74a904f952dd81ea1686cba5ba9c687264dbe.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 437, + 482, + 491 + ], + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 491 + ], + "type": "text", + "content": "Fig. 4: Samples from our method with style conditioning compared against other methods. We used an empty prompt and only conditioned on the image. We generally perform on par with IP-Adapter and outperform it on some samples. Note that the third image from the left is less degraded, and the third image from the right captures the mane of the horse better." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 526, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 526, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 526, + 482, + 624 + ], + "type": "text", + "content": "LoRAAdapter obtains state-of-the-art performance on CLIP-I and CLIP-T scores across all adapters, while also being more efficient (approx " + }, + { + "bbox": [ + 130, + 526, + 482, + 624 + ], + "type": "inline_equation", + "content": "27\\%" + }, + { + "bbox": [ + 130, + 526, + 482, + 624 + ], + "type": "text", + "content": " reduction in the number of trainable parameters compared with IP-Adapter [45]). We attribute this boost in performance to our conditional LoRA approach which can efficiently capture detailed style conditioning in a more effective manner than previous adapter approaches. Interestingly, we find that LoRAAdapter even outperforms methods that train large models from scratch like unCLIP on CLIP-T score, which further supports our unified conditioning approach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Structure For evaluating structure control we follow the settings in UniControlNet [50] and sample a single image per condition. To evaluate how" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "type": "text", + "content": "5. EXPERIMENTS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 156, + 125, + 459, + 157 + ], + "blocks": [ + { + "bbox": [ + 292, + 115, + 321, + 123 + ], + "lines": [ + { + "bbox": [ + 292, + 115, + 321, + 123 + ], + "spans": [ + { + "bbox": [ + 292, + 115, + 321, + 123 + ], + "type": "text", + "content": "Prompt" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 156, + 125, + 459, + 157 + ], + "lines": [ + { + "bbox": [ + 156, + 125, + 459, + 157 + ], + "spans": [ + { + "bbox": [ + 156, + 125, + 459, + 157 + ], + "type": "image", + "image_path": "fca4aba6f86f6b2ca31196e60cacf7011bc45d55cf135c5734c72beb7f1d13e3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 156, + 168, + 459, + 201 + ], + "blocks": [ + { + "bbox": [ + 286, + 159, + 328, + 167 + ], + "lines": [ + { + "bbox": [ + 286, + 159, + 328, + 167 + ], + "spans": [ + { + "bbox": [ + 286, + 159, + 328, + 167 + ], + "type": "text", + "content": "ControlNet" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 156, + 168, + 459, + 201 + ], + "lines": [ + { + "bbox": [ + 156, + 168, + 459, + 201 + ], + "spans": [ + { + "bbox": [ + 156, + 168, + 459, + 201 + ], + "type": "image", + "image_path": "91e8b013dc4d4ac6c3bdddb7a64c94b452d5441b0572077f961169a354d40600.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 156, + 213, + 459, + 245 + ], + "blocks": [ + { + "bbox": [ + 283, + 202, + 330, + 212 + ], + "lines": [ + { + "bbox": [ + 283, + 202, + 330, + 212 + ], + "spans": [ + { + "bbox": [ + 283, + 202, + 330, + 212 + ], + "type": "text", + "content": "T2I Adapter" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 156, + 213, + 459, + 245 + ], + "lines": [ + { + "bbox": [ + 156, + 213, + 459, + 245 + ], + "spans": [ + { + "bbox": [ + 156, + 213, + 459, + 245 + ], + "type": "image", + "image_path": "550fa223819a407c4315fb0a2f8686ab98644279da79abbae678e30dff1d1bf1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 156, + 256, + 459, + 289 + ], + "blocks": [ + { + "bbox": [ + 278, + 247, + 335, + 255 + ], + "lines": [ + { + "bbox": [ + 278, + 247, + 335, + 255 + ], + "spans": [ + { + "bbox": [ + 278, + 247, + 335, + 255 + ], + "type": "text", + "content": "Uni-ControlNet" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 156, + 256, + 459, + 289 + ], + "lines": [ + { + "bbox": [ + 156, + 256, + 459, + 289 + ], + "spans": [ + { + "bbox": [ + 156, + 256, + 459, + 289 + ], + "type": "image", + "image_path": "878531e1d79adca1b83b27f63bb45bf63699faa24594ccbdd76e15072ac13dea.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 156, + 300, + 459, + 332 + ], + "blocks": [ + { + "bbox": [ + 298, + 291, + 316, + 299 + ], + "lines": [ + { + "bbox": [ + 298, + 291, + 316, + 299 + ], + "spans": [ + { + "bbox": [ + 298, + 291, + 316, + 299 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 156, + 300, + 459, + 332 + ], + "lines": [ + { + "bbox": [ + 156, + 300, + 459, + 332 + ], + "spans": [ + { + "bbox": [ + 156, + 300, + 459, + 332 + ], + "type": "image", + "image_path": "6aada045b0f7398da0c4a6bf76fcaf218ed7a8d8d5a4326fe84484669de68941.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 334, + 482, + 390 + ], + "lines": [ + { + "bbox": [ + 130, + 334, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 482, + 390 + ], + "type": "text", + "content": "Fig. 5: Samples from our method with structural conditioning compared against other methods. Note that for our method, especially compared with T2I Adapter, the details of the images are substantially more closely aligned with the depth prompt (see, e.g., the lamp in the background of the living room scene and the side table's legs, or the salad on the pizza)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 433, + 482, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 482, + 565 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 482, + 565 + ], + "type": "text", + "content": "closely the generated image is following the provided structure map, we measure the cycle consistency between the predicted structure map of the original image " + }, + { + "bbox": [ + 130, + 433, + 482, + 565 + ], + "type": "inline_equation", + "content": "c = \\mathrm{encoder}(x)" + }, + { + "bbox": [ + 130, + 433, + 482, + 565 + ], + "type": "text", + "content": " and the predicted structure map of the generated image " + }, + { + "bbox": [ + 130, + 433, + 482, + 565 + ], + "type": "inline_equation", + "content": "\\hat{c} = \\mathrm{encoder}(\\hat{x})" + }, + { + "bbox": [ + 130, + 433, + 482, + 565 + ], + "type": "text", + "content": ". The exact metric used to measure the discrepancy depends on the type but " + }, + { + "bbox": [ + 130, + 433, + 482, + 565 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 433, + 482, + 565 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 433, + 482, + 565 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 130, + 433, + 482, + 565 + ], + "type": "text", + "content": " are always normalized to [0, 1] before any metrics computation. The differences in depth maps are measured by the Mean Squared Error between " + }, + { + "bbox": [ + 130, + 433, + 482, + 565 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 433, + 482, + 565 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 433, + 482, + 565 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 130, + 433, + 482, + 565 + ], + "type": "text", + "content": " (MSE-d). The SSIM (Structural Similarity) is used for HED conditioning. Additionally, we also compute the Learned Perceptual Image Patch Similarity (LPIPS) [49] to implicitly measure the fidelity of the control as well as the Frechet Inception Distance (FID) [11], which compares the distributions of intermediate features of a pre-trained network between generated and original images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Table 2 shows results across metrics for different approaches. In particular, adapters that focus on structure conditioning tend to use more parameters than ones used for style therefore we trained two versions of LoRAAdapter. One with 17M (roughly matching the number of parameters used for style conditioning) and 32M parameters, which is still less than T2I-Adapter [24]. LoRAAdapter obtains state-of-the-art performance across the board, even when reducing parameters by half compared to T2I-Adapter. These results show that LoRAAdapter is a very efficient approach for structure conditioning which can be scaled to obtain" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Stracke et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 124, + 291, + 236 + ], + "blocks": [ + { + "bbox": [ + 141, + 124, + 291, + 236 + ], + "lines": [ + { + "bbox": [ + 141, + 124, + 291, + 236 + ], + "spans": [ + { + "bbox": [ + 141, + 124, + 291, + 236 + ], + "type": "image", + "image_path": "e5f552d84f1b0fd94fa6885b0c1d1c610c652093d5ab7caf34a0bf54fbb2c5dd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 246, + 303, + 323 + ], + "lines": [ + { + "bbox": [ + 131, + 246, + 303, + 323 + ], + "spans": [ + { + "bbox": [ + 131, + 246, + 303, + 323 + ], + "type": "text", + "content": "Fig. 6: We show how the choice of the LoRA rank affects the generation quality. The rank refers to the cross-attention models in Table 3. Even the smallest model with only 2M parameters can meaningfully capture the content of the image albeit with minor losses in quality." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 315, + 213, + 482, + 302 + ], + "blocks": [ + { + "bbox": [ + 311, + 124, + 482, + 212 + ], + "lines": [ + { + "bbox": [ + 311, + 124, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 311, + 124, + 482, + 212 + ], + "type": "text", + "content": "Table 3: Evaluation of different global conditional LoRA configurations. All models were trained using CLIP VIT-L/14 as an image encoder and evaluated on 5000 samples. Adapting the Cross-Attention layers yields the best performance. Even very low-rank LoRA conditioning can still produce good performance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 213, + 482, + 302 + ], + "lines": [ + { + "bbox": [ + 315, + 213, + 482, + 302 + ], + "spans": [ + { + "bbox": [ + 315, + 213, + 482, + 302 + ], + "type": "table", + "html": "
LayerRank ParametersCLIP-ICLIP-T
Cross Attention20821M0.8370.631
Cross Attention12813M0.8350.634
Cross Attention647M0.8290.637
Cross Attention324M0.8160.642
Cross Attention162M0.7990.649
Self Attention12813M0.8280.639
Cross & Self Attention12826M0.8420.629
Cross & Self Attention6413M0.8370.636
Convolutional12820M0.8060.647
", + "image_path": "6e35920f2e8b9842e537706df2240c086104f8435d56be07ca5319cc86f217d7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 349, + 480, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 480, + 373 + ], + "type": "text", + "content": "high-performant models. We refer the reader to the Appendix for additional results and comparisons." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 397, + 280, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 397, + 280, + 410 + ], + "spans": [ + { + "bbox": [ + 131, + 397, + 280, + 410 + ], + "type": "text", + "content": "5.3 Qualitative Comparison" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 423, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 495 + ], + "type": "text", + "content": "To show qualitative results for both structure and style conditioning we take images used for conditioning by previous approaches while also adding some conditioning images selected by us. Results are shown in Fig. 4 where we can see that LoRAAdapter better captures the style of the conditioning image compared with previous adapter approaches. We show additional comparison with more models on an extended set of samples in the Appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 496, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 482, + 568 + ], + "type": "text", + "content": "We also show samples for structure conditioning on depth maps predicted by MiDaS [32] in Fig 5. LoRAAdapter shows nice sample quality and superior adherence to the fine-grained depth structure compared to ControlNet or T2I-Adapter. We attribute this improvement to our choice of directly adapting the convolutions in the U-Net as opposed to an indirect adaption via skip connections as previous approaches have proposed [7, 24]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 591, + 270, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 591, + 270, + 604 + ], + "spans": [ + { + "bbox": [ + 131, + 591, + 270, + 604 + ], + "type": "text", + "content": "5.4 Ablation experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "To showcase the modularity of our approach, we train multiple configurations per condition and assess their performance. We study the influence of the LoRAs rank as well as the specific layer choice, which is further discussed for each conditioning type." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 363, + 91, + 447, + 100 + ], + "type": "text", + "content": "5. EXPERIMENTS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 171, + 113, + 443, + 245 + ], + "blocks": [ + { + "bbox": [ + 171, + 113, + 443, + 245 + ], + "lines": [ + { + "bbox": [ + 171, + 113, + 443, + 245 + ], + "spans": [ + { + "bbox": [ + 171, + 113, + 443, + 245 + ], + "type": "image", + "image_path": "4ec3335ba28531de687e35d2bc3c692d39cedeaf7881bbd7ad14620ef5697bf3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 247, + 482, + 281 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 281 + ], + "type": "text", + "content": "Fig. 7: Examples of using a style and structure LoRAAdapter jointly. We show both reconstruction tasks (first, second, and third columns) and cases where the style and structure image do not match perfectly (fourth, fifth, and sixth columns)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 312, + 483, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 483, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 483, + 397 + ], + "type": "text", + "content": "Rank and Layer choice As the modularity of our approach is a key differentiator from other methods, we perform an analysis of how the choice of adapted layers and the overall rank influence the performance in the context of style conditioning (Table 3). Overall the cross-attention layers perform best and show very good performance even with a very small rank. Convolutional and self-attention layers perform worse but may be better suited for other tasks such as conditioning on the style of an image instead of its semantics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 411, + 483, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 483, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 483, + 520 + ], + "type": "text", + "content": "Structure We adapt the convolutional layers and vary in what blocks we adapt them. Stable Diffusion 1.5 contains four upsampling and downsampling blocks and one middle block. Each block itself contains two ResNet [9] blocks with two convolutional layers each. We analyze two different configurations: Configuration A only adapts the first layer in every block, totaling nine adapted convolutional layers. Configuration B adapts the first convolutional layer in every ResNet block, i.e. 17 adapted layers in total (Table 2). Finally, we show examples of jointly using a style and structure conditioning on Fig. 7. We refer the reader to the Appendix for additional results on prompt combinations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 534, + 220, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 220, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 220, + 548 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": "In this paper we have introduced LoRAAdapter, an approach to condition the image generation process of foundational text-to-image models. LoRAAdapter is an powerful and efficient approach for conditioning that unifies structure and style control using a conditional LoRA architecture that enables zero-shot generalization. Our approach enables fine-grained control over both structure and style obtaining state-of-the-art results and outperforming recent approaches while also optimizing a more compact number of parameters. LoRAAdapter presents progress towards acquiring fine-grained control over text-to-image models in an effective manner." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Stracke et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 236 + ], + "type": "text", + "content": "This project has been supported by the German Federal Ministry for Economic Affairs and Climate Action within the project \"NXT GEN AI METHODS - Generative Methoden für Perzeption, Prädiktion und Planung\", the German Research Foundation (DFG) project 421703927, and the bidt project KLIMA-MEMES. The authors gratefully acknowledge the Gauss Center for Supercomputing for providing compute through the NIC on JUWELS at JSC and the HPC resources supplied by the Erlangen National High Performance Computing Center (NHR@FAU funded by DFG)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 251, + 197, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 251, + 197, + 264 + ], + "spans": [ + { + "bbox": [ + 133, + 251, + 197, + 264 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 275, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 275, + 412, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 275, + 412, + 285 + ], + "spans": [ + { + "bbox": [ + 138, + 275, + 412, + 285 + ], + "type": "text", + "content": "1. https://huggingface.co/lambdaabs/sd-image-variations-diffusers" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 286, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 286, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 286, + 481, + 319 + ], + "type": "text", + "content": "2. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 319, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 319, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 138, + 319, + 481, + 341 + ], + "type": "text", + "content": "3. Byeon, M., Park, B., Kim, H., Lee, S., Baek, W., Kim, S.: Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 341, + 481, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 341, + 481, + 372 + ], + "spans": [ + { + "bbox": [ + 138, + 341, + 481, + 372 + ], + "type": "text", + "content": "4. Canny, J.: A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence PAMI-8(6), 679-698 (1986). https://doi.org/10.1109/TPAMI.1986.4767851" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 373, + 481, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 373, + 481, + 405 + ], + "spans": [ + { + "bbox": [ + 138, + 373, + 481, + 405 + ], + "type": "text", + "content": "5. Cao, Z., Simon, T., Wei, S.E., Sheikh, Y.: Realtime multi-person 2d pose estimation using part affinity fields. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7291-7299 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 405, + 481, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 405, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 138, + 405, + 481, + 426 + ], + "type": "text", + "content": "6. Cheng, J., Xie, P., Xia, X., Li, J., Wu, J., Ren, Y., Li, H., Xiao, X., Zheng, M., Fu, L.: Resadapter: Domain consistent resolution adapter for diffusion models (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 427, + 481, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 427, + 481, + 448 + ], + "spans": [ + { + "bbox": [ + 138, + 427, + 481, + 448 + ], + "type": "text", + "content": "7. Denis Zavadski, J.F.F., Rother, C.: Controlnet-xs: Designing an efficient and effective architecture for controlling text-to-image diffusion models (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 449, + 481, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 481, + 470 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 481, + 470 + ], + "type": "text", + "content": "8. Gandikota, R., Materzynska, J., Zhou, T., Torralba, A., Bau, D.: Concept sliders: Lora adaptors for precise control in diffusion models (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 471, + 481, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 471, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 471, + 481, + 502 + ], + "type": "text", + "content": "9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 503, + 481, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 503, + 481, + 524 + ], + "spans": [ + { + "bbox": [ + 138, + 503, + 481, + 524 + ], + "type": "text", + "content": "0. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 525, + 481, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 525, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 134, + 525, + 481, + 556 + ], + "type": "text", + "content": "11. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 557, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 557, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 134, + 557, + 481, + 578 + ], + "type": "text", + "content": "12. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 578, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 578, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 134, + 578, + 481, + 599 + ], + "type": "text", + "content": "13. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 481, + 632 + ], + "type": "text", + "content": "14. Hu, E.J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "text", + "content": "15. Huang, L., Chen, D., Liu, Y., Shen, Y., Zhao, D., Zhou, J.: Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint arXiv:2302.09778 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 369, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 369, + 91, + 447, + 100 + ], + "type": "text", + "content": "6. CONCLUSION" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 482, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 482, + 150 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 482, + 150 + ], + "type": "text", + "content": "16. Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip. https://github.com/mlfoundations/open Clip (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 482, + 183 + ], + "type": "text", + "content": "17. Jiang, Z., Mao, C., Pan, Y., Han, Z., Zhang, J.: Scedit: Efficient and controllable image diffusion generation via skip connection editing. arXiv preprint arXiv:2312.11392 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 205 + ], + "type": "text", + "content": "18. Kopiczko, D.J., Blankevoort, T., Asano, Y.M.: Vera: Vector-based random matrix adaptation (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 206, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 481, + 251 + ], + "type": "text", + "content": "19. Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22511-22521 (June 2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 296 + ], + "type": "text", + "content": "20. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 296, + 481, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 296, + 481, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 296, + 481, + 318 + ], + "type": "text", + "content": "21. Liu, S.Y., Wang, C.Y., Yin, H., Molchanov, P., Wang, Y.C.F., Cheng, K.T., Chen, M.H.: Dora: Weight-decomposed low-rank adaptation (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 318, + 481, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 318, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 481, + 341 + ], + "type": "text", + "content": "22. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "type": "text", + "content": "23. Luo, S., Tan, Y., Patil, S., Gu, D., von Platen, P., Passos, A., Huang, L., Li, J., Zhao, H.: Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "text", + "content": "24. Mou, C., Wang, X., Xie, L., Wu, Y., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 441 + ], + "type": "text", + "content": "25. Park, T., Liu, M.Y., Wang, T.C., Zhu, J.Y.: Semantic image synthesis with spatially-adaptive normalization. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2332-2341. IEEE (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 475 + ], + "type": "text", + "content": "26. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4195-4205 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 475, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 520 + ], + "type": "text", + "content": "27. Pernias, P., Rampas, D., Richter, M.L., Pal, C., Aubreville, M.: Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=gU58d5QeGv" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "type": "text", + "content": "28. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 554, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 481, + 586 + ], + "type": "text", + "content": "29. Qin, C., Zhang, S., Yu, N., Feng, Y., Yang, X., Zhou, Y., Wang, H., Niebles, J.C., Xiong, C., Savarese, S., et al.: Unicontrol: A unified diffusion model for controllable visual generation in the wild. arXiv preprint arXiv:2305.11147 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 587, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 481, + 632 + ], + "type": "text", + "content": "30. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "31. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Stracke et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "32. Ranftl, R., Lasinger, K., Hafner, D., Schindler, K., Koltun, V.: Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence 44(3), 1623-1637 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "text", + "content": "33. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "type": "text", + "content": "34. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "text", + "content": "35. Ruiz, N., Li, Y., Jampani, V., Wei, W., Hou, T., Pritch, Y., Wadhwa, N., Rubinstein, M., Aberman, K.: Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "text", + "content": "36. Ryu, S.: Low-rank adaptation for fast text-to-image diffusion fine-tuning, https://github.com/cloneofsimo/lora" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "type": "text", + "content": "37. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "text", + "content": "38. Shah, V., Ruiz, N., Cole, F., Lu, E., Lazebnik, S., Li, Y., Jampani, V.:Ziplora: Any subject in any style by effectively merging loras (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "text", + "content": "39. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "text", + "content": "40. Witteveen, S., Andrews, M.: Investigating prompt engineering in diffusion models. arXiv preprint arXiv:2211.15462 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "text", + "content": "41. Xie, S., Tu, Z.: Holistically-nested edge detection. In: Proceedings of the IEEE international conference on computer vision. pp. 1395-1403 (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 425, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 481, + 456 + ], + "type": "text", + "content": "42. Xu, X., Guo, J., Wang, Z., Huang, G., Essa, I., Shi, H.: Prompt-free diffusion: Taking\" text\" out of text-to-image diffusion models. arXiv preprint arXiv:2305.16223 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "text", + "content": "43. Xu, X., Wang, Z., Zhang, E., Wang, K., Shi, H.: Versatile diffusion: Text, images and variations all in one diffusion model. arXiv preprint arXiv:2211.08332 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 511 + ], + "type": "text", + "content": "44. Xue, Z., Song, G., Guo, Q., Liu, B., Zong, Z., Liu, Y., Luo, P.: Raphael: Text-to-image generation via large mixture of diffusion paths. arXiv preprint arXiv:2305.18295 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "text", + "content": "45. Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 534, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 577 + ], + "type": "text", + "content": "46. YEH, S.Y., Hsieh, Y.G., Gao, Z., Yang, B.B.W., Oh, G., Gong, Y.: Navigating text-to-image customization: From lyCORIS fine-tuning to model evaluation. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=wfzXa8e783" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 577, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 481, + 599 + ], + "type": "text", + "content": "47. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "48. Zhang, Q., Chen, M., Bukharin, A., Karampatziakis, N., He, P., Cheng, Y., Chen, W., Zhao, T.: Adalora: Adaptive budget allocation for parameter-efficient fine-tuning (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "49. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 369, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 369, + 91, + 447, + 100 + ], + "type": "text", + "content": "6. CONCLUSION" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 482, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 482, + 150 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 482, + 150 + ], + "type": "text", + "content": "50. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Unictrlnet: All-in-one control to text-to-image diffusion models. Advances in Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "Stracke et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CVT-Occ_ Cost Volume Temporal Fusion for 3D Occupancy Prediction/d0af9ebe-697b-4a4d-bc71-225f5f0fd15a_content_list.json b/2024/CVT-Occ_ Cost Volume Temporal Fusion for 3D Occupancy Prediction/d0af9ebe-697b-4a4d-bc71-225f5f0fd15a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c7eaa1fb241ae9ec53d585776e7ee949c9491dd2 --- /dev/null +++ b/2024/CVT-Occ_ Cost Volume Temporal Fusion for 3D Occupancy Prediction/d0af9ebe-697b-4a4d-bc71-225f5f0fd15a_content_list.json @@ -0,0 +1,1601 @@ +[ + { + "type": "text", + "text": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction", + "text_level": 1, + "bbox": [ + 267, + 140, + 736, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhangchen Ye $^{1*}$ , Tao Jiang $^{1,2*}$ , Chenfeng Xu $^{3}$ , Yiming Li $^{4}$ , and Hang Zhao $^{1,2,5\\boxtimes}$", + "bbox": [ + 269, + 210, + 730, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ IIIS, Tsinghua University $^{2}$ Shanghai AI Lab $^{3}$ UC Berkeley $^{4}$ New York University $^{5}$ Shanghai Qi Zhi Institute", + "bbox": [ + 282, + 253, + 720, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Vision-based 3D occupancy prediction is significantly challenged by the inherent limitations of monocular vision in depth estimation. This paper introduces CVT-Occ, a novel approach that leverages temporal fusion through the geometric correspondence of voxels over time to improve the accuracy of 3D occupancy predictions. By sampling points along the line of sight of each voxel and integrating the features of these points from historical frames, we construct a cost volume feature map that refines current volume features for improved prediction outcomes. Our method takes advantage of parallax cues from historical observations and employs a data-driven approach to learn the cost volume. We validate the effectiveness of CVT-Occ through rigorous experiments on the Occ3D-Waymo dataset, where it outperforms state-of-the-art methods in 3D occupancy prediction with minimal additional computational cost. The code is released at https://github.com/Tsinghua-MARS-Lab/CVT-Occ.", + "bbox": [ + 261, + 323, + 738, + 531 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D Semantic Occupancy Prediction $\\cdot$ Temporal Fusion", + "bbox": [ + 261, + 544, + 712, + 559 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 585, + 375, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-based 3D semantic occupancy prediction is rapidly evolving in the domain of 3D perception, driven by its critical applications in autonomous driving, robotics, and augmented reality. The task aims to estimate the occupancy state and semantic label of every voxel within a 3D space from visual inputs [11,29, 30,40].", + "bbox": [ + 212, + 619, + 785, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite its critical importance, 3D occupancy prediction presents significant challenges. When relying solely on monocular vision, these challenges are particularly pronounced due to the inherent ambiguity in estimating depth from a single image. While stereo vision has been proposed as a solution to augment depth estimation accuracy [14], its application remains limited in practice. It is impractical to use stereo cameras for widespread applications in autonomous vehicles and robotic systems because of the requirement for extensive calibration", + "bbox": [ + 212, + 696, + 787, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*: Equal contribution.", + "bbox": [ + 230, + 811, + 380, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\boxtimes$ : Corresponding to: hangzhao@mail.tsinghua.edu.cn", + "bbox": [ + 230, + 824, + 591, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d6581b4708145846a5456243e9aecf867d65af2190d1ff92e992bfe758335dc0.jpg", + "image_caption": [ + "Fig. 1: Comparison of Temporal Fusion Methods. Illustrated are four key approaches: (1) Temporal Self-Attention [20], leveraging attention mechanisms for temporal integration; (2) Wrap and Concat [8,35,37], combining features across frames and fusing them through convolution; (3) Cost Volume Construction in image space [25], constructing cost volume from image input of different frames and leveraging plane-sweep volumes for depth map generation; and (4) Our Proposed Method, which involves constructing a temporal cost volume in 3D space to enhance feature refinement. In the figure, $\\odot$ and $\\otimes$ represent coordinate alignment and element-wise product, accordingly." + ], + "image_footnote": [], + "bbox": [ + 233, + 150, + 772, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and re-calibration. An alternative and more promising approach is the employment of multi-view temporal fusion, which can utilize the extended multi-view baselines available over time to enhance 3D perception tasks.", + "bbox": [ + 212, + 435, + 784, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advancements in visual 3D object detection have shown the potential of incorporating temporal observations to bolster detection performance [8, 20, 25, 35, 37]. In our research, we have identified and categorized the emerging temporal fusion methods into three paradigms, as depicted in Figure 1.", + "bbox": [ + 212, + 484, + 787, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The first two paradigms are categorized as warp-based methods. These methodologies primarily involve the alignment of Bird's Eye View (BEV) feature maps across multiple temporal instances, utilizing relative camera poses obtained from Inertial Measurement Units (IMUs). Subsequently, the aligned features undergo integration using two primary approaches. The first approach employs self-attention mechanisms, as depicted in Figure 1(a). A representative method utilizing this approach is BEVFormer [20]. Alternatively, the second approach integrates aligned features through concatenation, as illustrated in Figure 1(b). Representative methods employing this approach include BEVDet4D [8], BEVFormerv2 [37], and PanoOcc [35]. These methods emphasize feature fusion using concatenation followed by convolutional operations. Despite their potential, these methods primarily leverage temporal information in an implicit manner, lacking a robust understanding of temporal geometry. As a result, they fall short in fully exploiting the geometric constraints inherent in 3D space.", + "bbox": [ + 212, + 547, + 787, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, cost volume-based methods, exemplified by (c) in Figure 1 and represented by SOLOFusion [25], draw inspiration from stereo matching techniques. Constructing a cost volume from images captured from different viewpoints in the temporal sequence enables the utilization of geometric constraints to obtain depth-aware features. However, for multi-view vision tasks, the sheer", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Ye et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "number of image pairs can lead to significant computational overhead when fusing longer time span temporal information.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recognizing the limitations of these approaches, we propose a novel paradigm, as illustrated in (d) in Figure 1. In this paper we introduce CVT-Occ, an innovative temporal fusion method designed to capitalize on the geometric correspondences of voxels over time, thereby enhancing occupancy prediction accuracy. Our approach involves sampling points along the line of sight for each voxel - defined as the line connecting the voxel to the optical center of the camera - and identifying the corresponding 3D locations of these points in historical frames based on relative camera poses. We then sample the features of these points in historical frames and integrate them with the current voxel features to construct a cost volume feature map. This map is subsequently utilized to refine the features of the current volume, thereby improving occupancy prediction.", + "bbox": [ + 212, + 176, + 785, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compared with paradigms (a) and (b) in Figure 1, our proposed CVT-Occ explicitly utilizes parallax cues to refine the depth of 3D voxel. In contrast to paradigm (c), our approach avoids per-image-pair cost volume computation, achieving superior performance with minimal additional computational overhead. CVT-Occ distinguishes itself by leveraging the inherent parallax information in historical observations and employing a data-driven approach to learning the cost volume. Our experiments on the Occ3D-Waymo dataset demonstrate that CVT-Occ achieves state-of-the-art 3D occupancy prediction performance.", + "bbox": [ + 212, + 343, + 787, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 482, + 387, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 3D Occupancy Prediction", + "text_level": 1, + "bbox": [ + 214, + 508, + 473, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of 3D occupancy prediction is to estimate the occupancy of each quantized voxel in 3D space. This task is originally rooted in Occupancy Grid Mapping (OGM) [24, 28] for mobile robot navigation, where the robot is equipped with range sensors (e.g., LiDAR) and navigates in static environments. Recent works transit to a more generic scenario: they make the occupancy prediction with visual systems in dynamic environments. MonoScene [1] reconstructs the 3D scene from sparse prediction with LiDAR points. VoxFormer [16] makes dense voxel predictions leveraging semantic labels and depth estimation from monocular RGB images. TPVFormer [11] proposes a tri-perspective view decomposition method for efficient 3D occupancy prediction. OccFormer [40] presents a dual-path transformer network to effectively process the 3D volume for semantic occupancy prediction. Besides, these advancements have led to new benchmarks such as visual 3D occupancy prediction [29, 30, 32] and semantic scene completion [15]. We emphasize that our proposed method is a simple and plug-and-play module that seamlessly integrates with existing occupancy prediction pipelines, significantly improving their performance.", + "bbox": [ + 212, + 530, + 787, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Temporal Fusion for 3D Perception", + "text_level": 1, + "bbox": [ + 214, + 789, + 553, + 804 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Leveraging temporal information for occupancy prediction is a natural strategy since it provides sufficient spatial information for building geometric repre", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sentations [25, 36]. A common practice of temporal fusion in visual 3D object detection is BEV feature alignment. It fuses temporal information by warping past BEV features to the current time according to camera view transformation from relative camera poses at different time steps [8, 20]. It has been shown that short-term frames can significantly help improve the 3D detection performance. However, we find that minor improvements are brought up when more frames come in with simply warping and fusion. Recent work UniFusion [27] proposes a new long temporal fusion method. It creates \"virtual views\" for the temporal features as if they are present in the current time and directly access all useful history features. However, this method suffers from intensive computation costs when the number of frames increases. SOLOFusion [25] is a more efficient and effective model that balances the resolution and time stamps, and transforms the temporal fusion problem into a temporal multi-view stereo problem.", + "bbox": [ + 212, + 146, + 787, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Stereo Matching and Multi-View Stereo", + "text_level": 1, + "bbox": [ + 214, + 366, + 591, + 382 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stereo matching is rooted in constructing 3D cost volumes from 2D image features [3, 33, 39] and predicting the depth map from the cost volume. Recent works further advance it by proposing correlation-based cost volumes like GC-Net [2]. Meanwhile, multi-view stereo methods (5,6,10,12,31,38)) leverage plane-sweep volumes for depth map generation. A recent trend of using stereo matching [4, 17, 19] focused on 3D perception in autonomous driving. While existing work in stereo matching often utilizes stereo image pairs, this format is unsuitable for handling multi-view and multi-temporal inputs in autonomous driving scenarios. Additionally, these approaches require generating a Plane-Sweep Volume for each image pair, leading to inefficiencies. Furthermore, recent advancements like OccDepth [23] demonstrate the potential of implicitly learning correlations between stereo images to improve 3D depth-aware feature fusion. Our proposed Temporal Cost Volume method, designed specifically for multi-view and multi-temporal data, offers a more effective and efficient solution. The core innovation of the CVT module lies in its ability to construct a comprehensive 3D representation of the scene by aggregating information over time.", + "bbox": [ + 212, + 393, + 787, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 214, + 659, + 380, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Problem Setup", + "text_level": 1, + "bbox": [ + 214, + 691, + 385, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given only RGB images as inputs, the model is designed to predict a dense semantic scene within a specified volume. Specifically, we utilize as input the current timestamp $t$ along with previous images, denoted by $\\mathbf{I} = \\{\\mathbf{I}_t^i,\\mathbf{I}_{t - 1}^i,\\ldots \\}_{i = 1}^L$ , where $L$ represents the number of cameras. Our output is defined as a voxel grid $Y_{t}\\in \\{c_{0},c_{1},\\dots ,c_{M}\\}^{H\\times W\\times Z}$ , situated in the coordinate frame of the ego-vehicle at timestamp $t$ . Each voxel within this grid is either empty (indicated by $c_{0}$ ) or occupied by a specific semantic class among $\\{c_1,c_2,\\dots ,c_M\\}$ . Here, $M$ denotes the total number of classes of interest, while $H$ , $W$ , and $Z$ represent the length,", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Ye et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c5eb4489ef60b37b601397dd601a292a3cf98d2ba454d1138f9781ddb533bace.jpg", + "image_caption": [ + "Fig. 2: Overall Architecture of CVT-Occ. The image backbone extracts multiscale features from multi-view images, which are transformed into 3D volume features denoted as $\\mathbf{V} \\in \\mathbb{R}^{H \\times W \\times Z \\times C}$ . The Cost Volume Temporal Module samples points along the line of sight within the current volume and projects them onto $K - 1$ historical frames, resulting in $K \\times N$ 3D volume features. These features are concatenated to construct cost volume features $\\mathbf{F} \\in \\mathbb{R}^{H \\times W \\times Z \\times (K \\times N) \\times C}$ . Convolution layers are then applied to generate weights $\\mathbf{W} \\in \\mathbb{R}^{H \\times W \\times Z}$ , refining the depth of 3D voxel. Finally, an occupancy decoder produces 3D semantic occupancy predictions. In the figure, $(\\text{\\textcircled{A}})$ , $(\\text{\\textcircled{C}})$ , and $\\otimes$ represent coordinate alignment, concatenation, and element-wise product, respectively." + ], + "image_footnote": [], + "bbox": [ + 238, + 148, + 756, + 392 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "width, and height of the grid, respectively. The primary goal is to train a neural network $\\Theta(\\cdot)$ to produce a semantic voxel grid $Y_{t} = \\Theta(\\mathbf{I})$ that approximates the ground truth $\\hat{Y}_{t}$ . It is important to note that the model must learn both 3D geometry and semantics solely from vision-based inputs, without the aid of depth measurements from LiDAR. Inferring 3D occupancy from 2D images presents a significant challenge and necessitates the design of efficient methods to leverage geometric constraints for learning accurate geometry.", + "bbox": [ + 212, + 578, + 787, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Overall Architecture", + "text_level": 1, + "bbox": [ + 215, + 707, + 431, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we depict the comprehensive architecture of CVT-Occ, depicted in Figure 2. Our framework processes multi-frame, multi-view images to first extract multi-scale features through an image backbone. Subsequently, these features from image space are transformed into BEV space features, which are refined by a BEV encoder to generate 3D volumetric representations. There is a rich corpus of research focused on the transition from image space to BEV features. One line of works follows the lifting paradigm proposed in LSS [26];", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction", + "bbox": [ + 256, + 113, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "they explicitly predict a depth map and lift multi-view image features onto the BEV plane [8, 9, 17, 18, 25]. Another line of works inherits the spirit of querying from 3D to 2D in DETR3D [34]; they employ learnable queries to extract information from image features by cross-attention mechanism [13, 20, 21]. It is crucial to underscore that our proposed Cost Volume Temporal (CVT) module is versatile and compatible across different strategies for feature transformation from image to volume space. For our experimental validation, we employ BEV-Former [20] as the foundation to generate 3D volume features. These features are subsequently refined and enhanced through the CVT module, demonstrating the efficacy and adaptability of our approach in leveraging temporal and spatial dynamics for enriched 3D scene comprehension. Finally an occupancy decoder is used to generate final voxel prediction results.", + "bbox": [ + 212, + 146, + 787, + 328 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Cost Volume Temporal Module", + "text_level": 1, + "bbox": [ + 215, + 351, + 522, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since direct depth information is not available for each pixel, transforming image features into 3D space introduces ambiguity; for instance, a single pixel may correspond to multiple voxels along the line of sight. To address this challenge, we propose the Cost Volume Temporal module, which leverages temporal data to infer depth information and resolve this ambiguity. Specifically, we construct 3D cost volume features using both historical and current BEV features. These cost volume features are then used to derive learned weights, which are subsequently employed to refine the current BEV features.", + "bbox": [ + 212, + 378, + 787, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Volume Features. We pre-define the volume $\\mathbf{V} \\in \\mathbb{R}^{H \\times W \\times Z}$ , where $H, W, Z$ represent the spatial dimensions of the volume space. Each voxel in the volume space corresponds to a cubic space in the real world with a size of $s$ meters. By default, the center of the volume corresponds to the position of the ego car. For each voxel $v = (i, j, k)$ , its corresponding 3D location $p = (x, y, z)$ can be calculated by Eqn. 1:", + "bbox": [ + 212, + 500, + 787, + 590 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nx = \\left(i - \\frac {W}{2}\\right) \\times s, y = \\left(j - \\frac {H}{2}\\right) \\times s, z = \\left(k - \\frac {Z}{2}\\right) \\times s (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 284, + 601, + 785, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The BEV feature is $\\mathbf{F}_t^{\\mathrm{bev}}\\in \\mathbb{R}^{H\\times W\\times E}$ , where $t$ represents the timestamp and $E$ is the embedding dimensions. We reshape the BEV feature to volume space as the input of the CVT module, producing $\\mathbf{V}_t\\in \\mathbb{R}^{H\\times W\\times Z\\times C}$ .", + "bbox": [ + 212, + 646, + 787, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Construct Cost Volume Feature. Given the ambiguity along the line of sight, for each point $\\mathbf{p_t}$ , we sample several additional points $\\{\\mathbf{p_t^i}\\}_{i=1}^N$ within the current volume. Specifically, we calculate the sight direction $\\mathbf{d_t} \\in \\mathbb{R}^3$ , which is the vector from the center of the volume to the point $\\mathbf{p_t}$ , and then sample points using specific strides $\\{n^i\\}_{i=1}^N$ as shown in Eqn. 2:", + "bbox": [ + 212, + 691, + 787, + 768 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {\\mathbf {t}} ^ {\\mathbf {i}} = \\mathbf {p} _ {\\mathbf {t}} + \\mathbf {d} _ {\\mathbf {t}} \\times \\mathbf {n} ^ {\\mathbf {i}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 437, + 780, + 785, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since these points project to the same pixel in the image space, points along the same line tend to have similar features. To accurately distinguish the correct", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Ye et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "position corresponding to the pixel, we use historical BEV features to obtain complementary information. Here comes the core insight of our proposed Cost Volume Temporal Module. Projecting these points into the historical coordinate frame ensures they are no longer in the same line of sight. This parallax provides additional information from the historical BEV features, which helps to reduce depth ambiguity in the current frame. Projection matrix $\\mathbf{P}_{\\mathbf{t}} \\in \\mathbb{R}^{4 \\times 4}$ can transform points from the coordinate frame of ego-vehicle to the global coordinate frame. Therefore the points $\\{\\mathbf{p}_{\\mathbf{t}}^i\\}_{i=1}^N$ are projected onto $K - 1$ historical frames through the projection matrix as described in Eqn. 3:", + "bbox": [ + 212, + 146, + 782, + 282 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {\\mathbf {t} - \\mathbf {k}} ^ {\\mathrm {i}} = \\mathbf {p} _ {\\mathbf {t}} ^ {\\mathrm {i}} \\mathbf {P} _ {\\mathbf {t}} \\mathbf {P} _ {\\mathbf {t} - \\mathbf {k}} ^ {- 1} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 436, + 295, + 784, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, each point $\\mathbf{p}_{\\mathbf{t} - \\mathbf{k}}^{\\mathbf{i}}$ is converted to voxel coordinates, which is the inverse process of Eqn. 1, and bilinear interpolation is used to sample features from corresponding BEV feature maps. The final cost volume feature map is $\\mathbf{F} \\in \\mathbb{R}^{H \\times W \\times Z \\times (K \\times N) \\times C}$ , where $C$ represents the channels of the BEV features.", + "bbox": [ + 214, + 325, + 782, + 387 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Volume Features Refinement", + "text_level": 1, + "bbox": [ + 214, + 412, + 500, + 426 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {V} _ {\\mathrm {o c c}} = \\mathbf {V} _ {t} \\otimes \\operatorname {S i g m o i d} (\\operatorname {C o n v} (\\mathbf {F})) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 444, + 784, + 459 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As described in the Eqn. 4, the constructed cost volume features $\\mathbf{F}$ are processed through a sequence of convolutional layers, yielding an output weight $\\mathbf{W} \\in \\mathbb{R}^{H \\times W \\times Z}$ . A Sigmoid activation function then normalizes the output weights $\\mathbf{W}$ to the range [0,1]. These weights are directly supervised according to the voxel occupancy state: weights corresponding to occupied voxels are encouraged to reach 1, while those for unoccupied voxels are guided towards 0. The symbol $\\otimes$ in the Eqn. 4 denotes element-wise product of the original volume features at timestamp $t$ with the learned weights $\\mathbf{W}$ . This supervised learning approach produces an occupancy-aware volume feature map $\\mathbf{V}_{\\mathrm{occ}} \\in \\mathbb{R}^{H \\times W \\times Z \\times C}$ . The goal of the learned weights is to diminish the influence of voxel features in incorrectly activated regions due to depth ambiguity and augment the features of correctly identified voxels simultaneously.", + "bbox": [ + 212, + 468, + 787, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.5 Occupancy Decoder", + "text_level": 1, + "bbox": [ + 214, + 675, + 426, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Upon the refinement of voxel features $\\mathbf{V}_{\\mathrm{occ}}$ , our model employs a series of deconvolution layers to transform it into occupancy features. The occupancy features are projected onto the output space, yielding $\\mathbf{X} \\in \\mathbb{R}^{H \\times W \\times Z \\times M}$ . The projection is formulated to map the occupancy features into a discrete set of semantic class predictions for each voxel within the grid. Consequently, the final occupancy output $\\mathbf{Y} \\in \\{c_0, c_1, \\ldots, c_M\\}^{H \\times W \\times Z}$ is generated, where each voxel is assigned one of $M$ semantic labels, including an unoccupied state represented by $c_0$ . This step effectively translates the enriched voxel features into a semantically segmented 3D occupancy map.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.6 Training Loss", + "text_level": 1, + "bbox": [ + 215, + 146, + 374, + 161 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our architecture integrates the CVT module and 3D occupancy prediction within an end-to-end trainable framework. We train the overall network with a multitask loss as", + "bbox": [ + 212, + 167, + 784, + 210 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {L o s s} = \\mathcal {L} _ {\\text {o c c}} + \\lambda \\mathcal {L} _ {\\text {c v t}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 429, + 214, + 784, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{occ}}$ denotes the occupancy prediction loss, and $\\mathcal{L}_{\\mathrm{cvt}}$ represents the cost volume temporal loss. The coefficient $\\lambda$ serves as a balancing factor between these two loss components.", + "bbox": [ + 212, + 232, + 782, + 277 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For the loss of occupancy prediction $\\mathcal{L}_{\\mathrm{occ}}$ , we adopt the cross-entropy loss, which is widely used in semantic occupancy prediction task. $\\mathcal{L}_{\\mathrm{occ}} = \\sum_{c}w_{c}\\mathcal{L}(g_{c},p_{c})$ where $g_{c}$ and $p_c$ represent the ground truth label and the prediction result for the $c$ -th semantic class, respectively. Class-specific weighting $w_{c}$ mitigates class imbalance by inversely correlating with class frequency, as adopted from Occ3D [29].", + "bbox": [ + 212, + 277, + 797, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To enhance the network's performance, the CVT module receives direct supervision. The loss attributed to the CVT module, $\\mathcal{L}_{\\mathrm{cvt}}$ , employs a binary cross-entropy framework as shown in Eqn. 6:", + "bbox": [ + 212, + 368, + 784, + 414 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c v t}} = - \\sum_ {j = 1} ^ {J} \\hat {y} _ {j, 0} \\log \\left(1 - w _ {j}\\right) + \\hat {y} _ {j, 1} \\log w _ {j} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 421, + 784, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $j$ indexes voxels in the grid, with $J = H \\times W \\times Z$ . $w_{k}$ is the $j$ -th element in the output weight $\\mathbf{W} \\in \\mathbb{R}^{H \\times W \\times Z}$ . Specifically, a binary label $\\hat{y}_{j,0} = 1$ signifies the voxel's unoccupied status (i.e., its semantic classification is $c_{0}$ ), while $\\hat{y}_{j,1} = 1$ indicates the voxel is occupied in ground truth. This direct supervision on the CVT module not only improves performance but also underpins the importance of accurately learning temporal and spatial features from the input data.", + "bbox": [ + 212, + 468, + 787, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 215, + 579, + 366, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Dataset", + "text_level": 1, + "bbox": [ + 215, + 607, + 326, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Occ3D-Waymo. We conduct experiments on the Occ3D-Waymo [29] dataset, which provides dense semantic labels for 3D occupancy grids. Occ3D-Waymo consists of 798 training scenes and 202 validation scenes. On the Occ3D-Waymo dataset, the range for the $x$ and $y$ axis as $[-40m, 40m]$ , and for the $z$ axis as $[-1m, 5.4m]$ . The voxel grid size is $(0.4m, 0.4m, 0.4m)$ , resulting in a resolution of $(200 \\times 200 \\times 16)$ for $(H, W, Z)$ . While the semantic labels encompass 16 categories, including \"Free\" class. Besides, it also provides visibility masks for LiDAR and camera modality.", + "bbox": [ + 212, + 628, + 787, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation Metrics. In our assessment of 3D semantic occupancy prediction capabilities, we utilize the IoU of each class to evaluate 3D semantic occupancy prediction. As Occ3D-Waymo provides camera visibility masks, only voxels within the camera's visible region are evaluated. The mean Intersection over Union (mIoU) is computed over 14 non-Free classes, excluding the \"Motorcycle\" class due to its insufficient voxel numbers.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Ye et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Experimental Settings", + "text_level": 1, + "bbox": [ + 215, + 146, + 444, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Architecture Detail. Our architecture leverages the same image backbone and image neck as Occ3D [29] for image feature extraction. The transformation of image features to BEV space involves a four-layered view transformation layers, each layer comprising a sequence of cross-attention layers, normalization layers, feed-forward layers, and normalization layers. This BEV encoder generates a BEV feature representation $\\mathbf{F}^{\\mathrm{bev}}\\in \\mathbb{R}^{200\\times 200\\times 256}$ , which is further reshaped to volume features $\\mathbf{V}\\in \\mathbb{R}^{200\\times 200\\times 16\\times 16}$ . The selected frame configuration includes a time interval of $0.5s$ and a sequence of 7 frames, covering a total time span of $3s$ , which ensures a comprehensive temporal analysis. For the CVT module, by default, it samples 9 points within each volume feature across 7 frames to capture a broad spectrum of temporal and spatial variations. Consequently, this configuration enables the CVT module to generate a cost volume feature map $\\mathbf{F}\\in \\mathbb{R}^{200\\times 200\\times 16\\times (7\\times 9)\\times 16}$ . The occupancy decoder utilizes a series of deconvolution layers to transform the volume features into occupancy features, which is then followed by two additional convolution blocks responsible for generating the final 3D semantic occupancy predictions.", + "bbox": [ + 212, + 172, + 787, + 412 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training. During training, we utilized the AdamW [22] optimizer for 8 epochs and full dataset each epoch. Training was conducted on 8 NVIDIA A100 GPUs with a batch size of 1 per GPU. Learning Rate Scheduler Policy is the cosine annealing schedule with an initial learning rate of $4 \\times 10^{-4}$ . The dimensions for the input images were standardized to $960 \\times 640$ pixels. It is noteworthy that we abstained from employing any data augmentation techniques in order to ensure a fair and unbiased comparison among the models.", + "bbox": [ + 212, + 412, + 787, + 518 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Main Results", + "text_level": 1, + "bbox": [ + 215, + 542, + 372, + 556 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As shown in Figure 1, We categorize previous temporal fusion methods into three types: (a) Attention-based methods, which adopt temporal self-attention to recurrently fuse the history BEV information; (b) WrapConcat-based methods, involving the alignment of BEV feature maps across different time steps; and (c) Plane-Sweep-based methods, which construct a cost volume for image pairs to utilize geometric constraints. Additionally, we introduce a novel approach, CVT-Occ, that utilizes geometric correspondences of 3D voxels over time to improve occupancy prediction accuracy. For comparative analysis, we extended some mainstream BEV models for the 3D semantic occupancy prediction task. We choose BEVFormer [20], BEVFormer-WarpConcat [37], and SOLOFusion [25] as representative methods for categories (a), (b), and (c), respectively.", + "bbox": [ + 212, + 568, + 787, + 734 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For both BEVFormer [20] and SOLOFusion [25], we have replaced their original transformer decoders for object detection with a 3D occupancy prediction decoder. Additionally, for fairness, we omit the depth supervision module when reproducing SOLOFusion [25]. For BEVFormer-WrapConcat, we adopt a commonly used fusion method known as the Wrap and Concatenate strategy, as seen in BEVFormerv2 [37] and PanoOcc [35]. This strategy involves 3D coordinate alignment in BEV space based on the transformation matrix between historical", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/89270818402e32a0ff00390c0a68e4a64093c34a0fc980f7cfe0bf0d73e4e9f4.jpg", + "table_caption": [ + "Table 1: 3D Occupancy Prediction Results on Occ3D-Waymo. This comparison showcases the performance of various principal temporal fusion strategies alongside CVT-Occ, within the same architectural framework and identical frame specifications (number and interval), ensuring a direct and fair comparison across all evaluated methods. The leading performance is highlighted in bold." + ], + "table_footnote": [], + "table_body": "
MethodmIoUGoVehiclePedestrianSignBicyclistTraffic LightPoleCons. ConeBicycleBuildingVegetationTree TrunkRoadWalkable
BEVFormer-w/o TSA23.877.5034.5421.079.6920.9611.4811.4814.0614.5123.1421.828.5778.4556.89
BEVFormer [20]24.587.1836.0621.009.7620.2312.6114.5214.7016.0623.9822.509.3979.1157.04
SOLOFusion [25]24.734.9732.4518.2810.3317.148.0717.8316.2319.331.4928.9816.9370.9553.28
BEVFormer-WrapConcat25.076.236.1720.959.5620.5812.8216.2414.3116.7825.1423.5612.8179.0456.83
CVT-Occ(ours)27.377.4441.023.9311.9220.8112.0718.0316.8821.3729.427.4214.6779.1259.09
", + "bbox": [ + 217, + 226, + 785, + 330 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "and current frames. Subsequently, it concatenates previous BEV features with current BEV features along the channel dimension and applies residual blocks for temporal fusion. We incorporate the residual block proposed by PanoOcc [35] to create the BEVFormer-WrapConcat method. In the \"BEVFormer-w/o TSA\" configuration, we maintain the original architecture of BEVFormer [20] but exclude the Temporal Self-Attention(TSA) layers, thereby eliminating temporal attention mechanisms. In Table 1, we present the evaluation results for 3D occupancy prediction on the Occ3D-Waymo validation set. To ensure a fair comparison, all methods were implemented using the same network architecture and training strategy.", + "bbox": [ + 215, + 357, + 785, + 508 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparing \"BEVFormer-w/o TSA\" with temporal fusion methods in Table 1, All temporal fusion methods demonstrate improved mIoU compared to \"BEVFormer-w/o TSA\". It becomes evident that temporal fusion plays a crucial role in the 3D occupancy prediction task.", + "bbox": [ + 215, + 508, + 785, + 568 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Maintaining consistent frame specifications (number and interval), we explore the temporal fusion ability of different mainstream temporal fusion methods. Our method achieves state-of-the-art performance in terms of IoU and demonstrates significant superiority over other approaches across nearly all classes. As shown in Table 1, CVT-Occ exhibits superior performance compared to representative models of existing temporal fusion methods (a), (b), and (c) in Figure 1, achieving a notable $2.8\\%$ IoU improvement over the baseline BEVFormer [20]. Delving into the details, CVT-Occ shows improvements in IoU across almost all classes. Specifically, for \"Vehicle\", \"Bicycle\", \"Building\", \"Vegetation\", and \"Tree Trunk\", our method achieves more than a $4\\%$ increase in IoU compared to the baseline model BEVFormer [20]. This substantial enhancement across nearly all categories underscores that a comprehensive understanding of temporal geometry can lead to significant performance gains.", + "bbox": [ + 215, + 569, + 785, + 765 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 785, + 387, + 801 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Time Span. The time span is defined as the difference between the timestamp of the earliest historical frame and that of the current frame, which plays a", + "bbox": [ + 215, + 809, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Ye et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/09bb5546cf7c14383681a19b72d1756114845b22c81cc3979e8b29af04634eec.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ExpFrameIntervalCVT supervisionmIoU
(a)30.1s✓25.15
(b)30.3s✓25.86
(c)70.1s✓25.80
(d)70.3s✓26.46
(e)70.5s✓27.37
(f)70.5s26.81
", + "bbox": [ + 215, + 143, + 480, + 296 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 2: Ablation Experiments. We evaluate CVT-Occ under different frame specifications and CVT supervision.", + "bbox": [ + 215, + 305, + 475, + 359 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/89af4c91687cb496213c5e43b741749ba08159e4cc5d847e2eab456beb651ab3.jpg", + "image_caption": [ + "Fig. 3: Experiments of Different Time Spans." + ], + "image_footnote": [], + "bbox": [ + 506, + 150, + 784, + 316 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "A crucial role in the temporal fusion analysis. We conduct experiments to evaluate our method across various configurations of total frame numbers and sampled frame intervals, as summarized in Table 2.", + "bbox": [ + 212, + 395, + 782, + 439 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In experiments (a) and (c), we compared the impact of different total frame numbers. The results demonstrate that a longer sequence of historical frames leads to improved performance. Similarly, we observed that a larger time interval between adjacent frames also contributes to enhanced performance in experiments (c), (d), and (e). These consistent findings underscore the significance of a longer time span across the entire temporal sequence for better model performance. CVT-Occ relies on disparities between frames to mitigate depth ambiguity, where larger intervals and frame numbers contribute to an extended time span, amplifying the temporal parallax. This elucidates the observed increase in mIoU with a longer time span.", + "bbox": [ + 212, + 441, + 787, + 594 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In addition, comparing experiments (b) and (c), where the total time span remains the same 0.6s, highlights the critical importance of the total time span even when the frame interval and number of historical frames varies between these two experiments. This finding further emphasizes that the impact of our proposed method arises from leveraging the parallax of the entire temporal sequence.", + "bbox": [ + 212, + 595, + 787, + 686 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ability of Long Temporal Fusion. Figure 3 illustrates the comparison between BEVFormer [20] and CVT-Occ across varying total time differences. CVT-Occ demonstrates improved performance with increasing temporal queue lengths, whereas BEVFormer struggles to effectively leverage information from long-range historical data.", + "bbox": [ + 212, + 686, + 787, + 762 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The limitation of BEVFormer's long temporal fusion capability has been highlighted in previous researches, such as UniFusion [27] and VideoBEV [7]. Based on results of Figure 3, we propose that BEVFormer fails to benefit from longer time spans due to its recurrent temporal fusion process. This process restricts the direct access to information from distant historical frames, regardless", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction", + "bbox": [ + 256, + 113, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/21b01a6d4a503c4f3777692428849bfe28ac1108c5d7816ed26e60ebd47a9c08.jpg", + "table_caption": [ + "Table 3: Comparison Results under Different Conditions. Here are three sets of experiments. (a) Binary Classification: this task focuses on geometric occupancy prediction, treating all object classes as a single \"Non-Free\" class. (b) BEV Range: Along x-axis, [0,20] means the range $[-20m,0m]\\cup [0,20m]$ and [20,40] means the range $[-40m, - 20m]\\cup [20, - 40m]$ . (c) Ego Vehicle Speed: The 202 scenes for evaluation is partitioned into 2 parts based on the ego-vehicle speed. We evaluate all temporal fusion methods on them separately." + ], + "table_footnote": [], + "table_body": "
ExpmIoUNon-FreeFree[0,20][20,40]FastSlow
BEVFormer [20]24.5855.7486.5427.0921.2424.7324.22
BEVFormer-WrapConcat25.0756.5786.9628.2121.1626.2724.44
CVT-Occ27.3762.9590.4630.8522.6727.926.79
", + "bbox": [ + 220, + 253, + 782, + 325 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "of the temporal sequence length. In contrast, the cost volume feature $\\mathbf{F}$ can directly integrate information from all historical frames. The CVT module utilizes parallax correlation explicitly to refine volume features, leading to a more robust understanding of temporal information. This design overcomes the shortcomings observed in BEVFormer [20], resulting in enhanced long-term temporal fusion capability.", + "bbox": [ + 212, + 354, + 782, + 444 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "CVT Supervision. In our pursuit to improve 3D semantic occupancy prediction, we go beyond traditional loss functions by adding extra supervision to the CVT module output. Experiments (e) and (f) in Table 2 show the performance comparison of CVT-Occ with and without the CVT loss. Incorporating the CVT loss results in a significant increase in mIoU, demonstrating its effectiveness through experimental results. This enhancement validates our design approach, highlighting its ability to enhance the accuracy and reliability of semantic occupancy predictions in 3D.", + "bbox": [ + 212, + 444, + 784, + 565 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In conclusion, our ablation study confirms the critical importance of longer time spans and effective temporal fusion in enhancing CVT-Occ's performance for 3D semantic occupancy prediction. Additionally, integrating supplementary supervision further improves model accuracy, highlighting our approach's effectiveness in advancing temporal parallax-based scene understanding.", + "bbox": [ + 212, + 565, + 785, + 642 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Analysis of Factors", + "text_level": 1, + "bbox": [ + 214, + 662, + 419, + 679 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The core insight of our method lies in leveraging parallax between different frames to enhance the accuracy of 3D voxel depth queries. In this subsection, we conduct several evaluations under various conditions to validate this concept. These experiments include binary classification, different BEV ranges, and varying ego vehicle speeds.", + "bbox": [ + 212, + 688, + 785, + 763 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Binary Classification. Within the framework of 3D semantic occupancy prediction, we aggregate all 15 other object classes into a single \"Non-Free\" class. This adjustment removes the influence of semantic information, allowing us to focus purely on depth accuracy. We then compare our approach against leading temporal fusion methods. The results in Table 3 demonstrate that our method", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Ye et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "provides better predictions on object geometry. This outcome aligns closely with our initial design concept of leveraging parallax for refined depth prediction.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "BEV Range. In the Occ3D-Waymo dataset, the BEV range spans from $-40m$ to $40m$ along the x-axis (i.e., the ego car's direction of movement). We divide this range into two parts: $[-20m, 20m]$ and $[-40m, 20m] \\cup [20m, 40m]$ . Table 3 exhibits the evaluation results. All methods perform better in the near region, which is expected since objects closer to the vehicle are more likely to appear in the overlap region of history and current frames. However, our method exhibits significantly greater improvement in the near region (+3.76) compared to the far region (+1.43) compared to BEVFormer [20]. The larger parallax of nearer objects between different frames contributes to our method's superior prediction capability.", + "bbox": [ + 212, + 176, + 785, + 328 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ego Vehicle Speed. Moreover, the speed of the vehicle also plays a crucial role in determining the magnitude of parallax. We further divide the dataset into two roughly equal parts based on the speed of ego vehicle. Evaluating CVT-Occ separately on these parts yields the results shown in Table 3. Regardless of whether BEVFormer [20] or CVT-Occ, both perform better in fast-moving scenarios. Our method demonstrates much greater improvement under the fast condition $(+3.17)$ than slow one $(+2.57)$ compared to the baseline BEVFormer. This can be attributed to the larger parallax between history and current frames when the ego car is moving rapidly.", + "bbox": [ + 212, + 329, + 785, + 465 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To sum up, these results underscore the significantly enhanced performance of CVT-Occ compared to the baseline model BEVFormer [20] under a wide range of conditions. The evaluations demonstrate that our method can refine the depth estimation of 3D voxel queries more accurately, with a notable advantage in the near region and fast scenes due to the greater parallax effect.", + "bbox": [ + 212, + 465, + 785, + 541 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.6 Visualization Results", + "text_level": 1, + "bbox": [ + 214, + 566, + 434, + 580 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Figure 4 offers a qualitative comparison of CVT-Occ against the baseline method BEVFormer [20] on the Occ3D-Waymo dataset. To ensure clarity in visualization, a dataset-specific mask is applied to exclude regions not within the camera's field of view. These visual representations clearly illustrate the superiority of CVT-Occ over competing approaches. BEVFormer encounters difficulties with depth perception, leading to issues such as atmospheric distortion and blurring of objects. CVT-Occ marks a remarkable improvements in the accurate depiction of buildings and vegetation, thus providing a more precise and reliable interpretation of the environment.", + "bbox": [ + 212, + 595, + 785, + 731 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 214, + 758, + 508, + 773 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We have presented CVT-Occ, a novel approach that significantly enhances the accuracy of 3D occupancy predictions by leveraging temporal fusion and geometric correspondence across time. Unlike traditional methods that rely on", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction", + "bbox": [ + 256, + 113, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/318bdeb9e37f1482b6c0ac38c29f2b6386a83f20123b249808ad5983bc3710a3.jpg", + "image_caption": [ + "Fig. 4: Qualitative Results. CVT-Occ exhibits superior performance in predicting the occupancy of vegetation and buildings." + ], + "image_footnote": [], + "bbox": [ + 253, + 145, + 754, + 599 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "monocular or stereo vision, our method utilizes multi-view temporal fusion, effectively incorporating historical observations to exploit the parallax effect and improve depth estimation. Our innovative approach, which constructs a cost volume feature map by sampling and integrating features across temporal frames, demonstrates superior performance on the Occ3D-Waymo dataset, surpassing current state-of-the-art models in terms of accuracy while maintaining a low computational overhead. CVT-Occ also opens avenues for future research on temporal fusion for 3D perception, e.g., applying the cost volume module in other tasks like 3D reconstruction, and other applications like robotics and virtual reality.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Ye et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments. This work is supported by the National Key R&D Program of China (2022ZD0161700).", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 202, + 321, + 218 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Cao, A.Q., de Charette, R.: Monoscene: Monocular 3d semantic scene completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3991-4001 (June 2022)", + "2. Cao, Y., Xu, J., Lin, S., Wei, F., Hu, H.: Gcnet: Non-local networks meet squeeze-excitation networks and beyond. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops (Oct 2019)", + "3. Chang, J.R., Chen, Y.S.: Pyramid stereo matching network. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018)", + "4. Chen, Y., Liu, S., Shen, X., Jia, J.: Dsgn: Deep stereo geometry network for 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)", + "5. Cheng, S., Xu, Z., Zhu, S., Li, Z., Li, L.E., Ramamoorthi, R., Su, H.: Deep stereo using adaptive thin volume representation with uncertainty awareness. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)", + "6. Gu, X., Fan, Z., Zhu, S., Dai, Z., Tan, F., Tan, P.: Cascade cost volume for high-resolution multi-view stereo and stereo matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)", + "7. Han, C., Yang, J., Sun, J., Ge, Z., Dong, R., Zhou, H., Mao, W., Peng, Y., Zhang, X.: Exploring recurrent long-term temporal fusion for multi-view 3d perception. IEEE Robotics and Automation Letters 9(7), 6544-6551 (2024)", + "8. Huang, J., Huang, G.: Bevdet4d: Exploit temporal cues in multi-camera 3d object detection (2022)", + "9. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view (2022)", + "10. Huang, P.H., Matzen, K., Kopf, J., Ahuja, N., Huang, J.B.: Deepmvs: Learning multi-view stereopsis. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018)", + "1. Huang, Y., Zheng, W., Zhang, Y., Zhou, J., Lu, J.: Tri-perspective view for vision-based 3d semantic occupancy prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9223-9232 (June 2023)", + "2. Im, S., Jeon, H.G., Lin, S., Kweon, I.S.: DPSNet: End-to-end deep plane sweep stereo. In: International Conference on Learning Representations (2019)", + "3. Jiang, Y., Zhang, L., Miao, Z., Zhu, X., Gao, J., Hu, W., Jiang, Y.G.: Polarformer: Multi-camera 3d object detection with polar transformer. Proceedings of the AAAI Conference on Artificial Intelligence 37(1), 1042-1050 (Jun 2023)", + "4. Li, H., Li, Z., Akmandor, N.U., Jiang, H., Wang, Y., Padir, T.: Stereovoxelnet: Real-time obstacle detection based on occupancy voxels from a stereo camera using deep neural networks. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 4826-4833 (2023)" + ], + "bbox": [ + 225, + 237, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Li, Y., Li, S., Liu, X., Gong, M., Li, K., Chen, N., Wang, Z., Li, Z., Jiang, T., Yu, F., Wang, Y., Zhao, H., Yu, Z., Feng, C.: Sscbench: Monocular 3d semantic scene completion benchmark in street views (2023)", + "16. Li, Y., Yu, Z., Choy, C., Xiao, C., Alvarez, J.M., Fidler, S., Feng, C., Anandkumar, A.: Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9087-9098 (June 2023)", + "17. Li, Y., Bao, H., Ge, Z., Yang, J., Sun, J., Li, Z.: Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. Proceedings of the AAAI Conference on Artificial Intelligence 37(2), 1486-1494 (Jun 2023)", + "18. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. Proceedings of the AAAI Conference on Artificial Intelligence 37(2), 1477-1485 (Jun 2023)", + "19. Li, Y., Yang, J., Sun, J., Bao, H., Ge, Z., Xiao, L.: Bevstereo++: Accurate depth estimation in multi-view 3d object detection via dynamic temporal stereo (2023)", + "20. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 1-18 (2022)", + "21. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X.: Petrov2: A unified framework for 3d perception from multi-camera images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3262-3272 (October 2023)", + "22. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2019)", + "23. Miao, R., Liu, W., Chen, M., Gong, Z., Xu, W., Hu, C., Zhou, S.: Occdepth: A depth-aware method for 3d semantic scene completion (2023)", + "24. Moravec, H., Elfes, A.: High resolution maps from wide angle sonar. In: Proceedings. 1985 IEEE International Conference on Robotics and Automation. vol. 2, pp. 116-121 (1985)", + "25. Park, J., Xu, C., Yang, S., Keutzer, K., Kitani, K.M., Tomizuka, M., Zhan, W.: Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. In: The Eleventh International Conference on Learning Representations (2023)", + "26. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020. pp. 194-210. Springer International Publishing, Cham (2020)", + "27. Qin, Z., Chen, J., Chen, C., Chen, X., Li, X.: Unifusion: Unified multi-view fusion transformer for spatial-temporal representation in bird's-eye-view. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 8690-8699 (October 2023)", + "28. Thrun, S.: Probabilistic robotics. Communications of the ACM pp. 52-57 (Mar 2002)", + "29. Tian, X., Jiang, T., Yun, L., Mao, Y., Yang, H., Wang, Y., Wang, Y., Zhao, H.: Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 64318-64330. Curran Associates, Inc. (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Ye et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Tong, W., Sima, C., Wang, T., Chen, L., Wu, S., Deng, H., Gu, Y., Lu, L., Luo, P., Lin, D., Li, H.: Scene as occupancy. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 8406-8415 (October 2023)", + "31. Wang, K., Shen, S.: Mvdepthnet: Real-time multiview depth estimation neural network. In: 2018 International Conference on 3D Vision (3DV) (Sep 2018)", + "32. Wang, X., Zhu, Z., Xu, W., Zhang, Y., Wei, Y., Chi, X., Ye, Y., Du, D., Lu, J., Wang, X.: Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 17850-17859 (October 2023)", + "33. Wang, Y., Lai, Z., Huang, G., Wang, B.H., van der Maaten, L., Campbell, M., Weinberger, K.Q.: Anytime stereo image depth estimation on mobile devices. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 5893-5900 (2019)", + "34. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Faust, A., Hsu, D., Neumann, G. (eds.) Proceedings of the 5th Conference on Robot Learning. Proceedings of Machine Learning Research, vol. 164, pp. 180-191. PMLR (08-11 Nov 2022)", + "35. Wang, Y., Chen, Y., Liao, X., Fan, L., Zhang, Z.: Panoocc: Unified occupancy representation for camera-based 3d panoptic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17158-17168 (June 2024)", + "36. Xu, C., Wu, B., Hou, J., Tsai, S., Li, R., Wang, J., Zhan, W., He, Z., Vajda, P., Keutzer, K., Tomizuka, M.: Nerf-det: Learning geometry-aware volumetric representation for multi-view 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 23320-23330 (October 2023)", + "37. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., Zhou, J., Dai, J.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17830-17839 (June 2023)", + "38. Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstructured multi-view stereo. In: Proceedings of the European Conference on Computer Vision (ECCV) (September 2018)", + "39. Zhang, F., Prisacariu, V., Yang, R., Torr, P.H.: Ga-net: Guided aggregation net for end-to-end stereo matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)", + "40. Zhang, Y., Zhu, Z., Du, D.: Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9433-9443 (October 2023)" + ], + "bbox": [ + 215, + 146, + 784, + 715 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction", + "bbox": [ + 256, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/CVT-Occ_ Cost Volume Temporal Fusion for 3D Occupancy Prediction/d0af9ebe-697b-4a4d-bc71-225f5f0fd15a_model.json b/2024/CVT-Occ_ Cost Volume Temporal Fusion for 3D Occupancy Prediction/d0af9ebe-697b-4a4d-bc71-225f5f0fd15a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..204a5bf60b5b7c97aea5312ea38da061df6252d6 --- /dev/null +++ b/2024/CVT-Occ_ Cost Volume Temporal Fusion for 3D Occupancy Prediction/d0af9ebe-697b-4a4d-bc71-225f5f0fd15a_model.json @@ -0,0 +1,2038 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.268, + 0.141, + 0.737, + 0.187 + ], + "angle": 0, + "content": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction" + }, + { + "type": "text", + "bbox": [ + 0.271, + 0.212, + 0.732, + 0.245 + ], + "angle": 0, + "content": "Zhangchen Ye\\(^{1*}\\), Tao Jiang\\(^{1,2*}\\), Chenfeng Xu\\(^{3}\\), Yiming Li\\(^{4}\\), and Hang Zhao\\(^{1,2,5\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.283, + 0.255, + 0.721, + 0.285 + ], + "angle": 0, + "content": "\\(^{1}\\)IIIS, Tsinghua University \\(^{2}\\)Shanghai AI Lab \\(^{3}\\)UC Berkeley \\(^{4}\\)New York University \\(^{5}\\)Shanghai Qi Zhi Institute" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.324, + 0.74, + 0.532 + ], + "angle": 0, + "content": "Abstract. Vision-based 3D occupancy prediction is significantly challenged by the inherent limitations of monocular vision in depth estimation. This paper introduces CVT-Occ, a novel approach that leverages temporal fusion through the geometric correspondence of voxels over time to improve the accuracy of 3D occupancy predictions. By sampling points along the line of sight of each voxel and integrating the features of these points from historical frames, we construct a cost volume feature map that refines current volume features for improved prediction outcomes. Our method takes advantage of parallax cues from historical observations and employs a data-driven approach to learn the cost volume. We validate the effectiveness of CVT-Occ through rigorous experiments on the Occ3D-Waymo dataset, where it outperforms state-of-the-art methods in 3D occupancy prediction with minimal additional computational cost. The code is released at https://github.com/Tsinghua-MARS-Lab/CVT-Occ." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.545, + 0.714, + 0.56 + ], + "angle": 0, + "content": "Keywords: 3D Semantic Occupancy Prediction \\(\\cdot\\) Temporal Fusion" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.587, + 0.377, + 0.603 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.62, + 0.787, + 0.695 + ], + "angle": 0, + "content": "Vision-based 3D semantic occupancy prediction is rapidly evolving in the domain of 3D perception, driven by its critical applications in autonomous driving, robotics, and augmented reality. The task aims to estimate the occupancy state and semantic label of every voxel within a 3D space from visual inputs [11,29, 30,40]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.788, + 0.803 + ], + "angle": 0, + "content": "Despite its critical importance, 3D occupancy prediction presents significant challenges. When relying solely on monocular vision, these challenges are particularly pronounced due to the inherent ambiguity in estimating depth from a single image. While stereo vision has been proposed as a solution to augment depth estimation accuracy [14], its application remains limited in practice. It is impractical to use stereo cameras for widespread applications in autonomous vehicles and robotic systems because of the requirement for extensive calibration" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.812, + 0.382, + 0.825 + ], + "angle": 0, + "content": "*: Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.825, + 0.592, + 0.84 + ], + "angle": 0, + "content": "\\(\\boxtimes\\) : Corresponding to: hangzhao@mail.tsinghua.edu.cn" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.127 + ], + "angle": 0, + "content": "Z. Ye et al." + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.151, + 0.774, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.398 + ], + "angle": 0, + "content": "Fig. 1: Comparison of Temporal Fusion Methods. Illustrated are four key approaches: (1) Temporal Self-Attention [20], leveraging attention mechanisms for temporal integration; (2) Wrap and Concat [8,35,37], combining features across frames and fusing them through convolution; (3) Cost Volume Construction in image space [25], constructing cost volume from image input of different frames and leveraging plane-sweep volumes for depth map generation; and (4) Our Proposed Method, which involves constructing a temporal cost volume in 3D space to enhance feature refinement. In the figure, \\(\\odot\\) and \\(\\otimes\\) represent coordinate alignment and element-wise product, accordingly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.785, + 0.481 + ], + "angle": 0, + "content": "and re-calibration. An alternative and more promising approach is the employment of multi-view temporal fusion, which can utilize the extended multi-view baselines available over time to enhance 3D perception tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.788, + 0.547 + ], + "angle": 0, + "content": "Recent advancements in visual 3D object detection have shown the potential of incorporating temporal observations to bolster detection performance [8, 20, 25, 35, 37]. In our research, we have identified and categorized the emerging temporal fusion methods into three paradigms, as depicted in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.789, + 0.761 + ], + "angle": 0, + "content": "The first two paradigms are categorized as warp-based methods. These methodologies primarily involve the alignment of Bird's Eye View (BEV) feature maps across multiple temporal instances, utilizing relative camera poses obtained from Inertial Measurement Units (IMUs). Subsequently, the aligned features undergo integration using two primary approaches. The first approach employs self-attention mechanisms, as depicted in Figure 1(a). A representative method utilizing this approach is BEVFormer [20]. Alternatively, the second approach integrates aligned features through concatenation, as illustrated in Figure 1(b). Representative methods employing this approach include BEVDet4D [8], BEVFormerv2 [37], and PanoOcc [35]. These methods emphasize feature fusion using concatenation followed by convolutional operations. Despite their potential, these methods primarily leverage temporal information in an implicit manner, lacking a robust understanding of temporal geometry. As a result, they fall short in fully exploiting the geometric constraints inherent in 3D space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In contrast, cost volume-based methods, exemplified by (c) in Figure 1 and represented by SOLOFusion [25], draw inspiration from stereo matching techniques. Constructing a cost volume from images captured from different viewpoints in the temporal sequence enables the utilization of geometric constraints to obtain depth-aware features. However, for multi-view vision tasks, the sheer" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.258, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "number of image pairs can lead to significant computational overhead when fusing longer time span temporal information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.344 + ], + "angle": 0, + "content": "Recognizing the limitations of these approaches, we propose a novel paradigm, as illustrated in (d) in Figure 1. In this paper we introduce CVT-Occ, an innovative temporal fusion method designed to capitalize on the geometric correspondences of voxels over time, thereby enhancing occupancy prediction accuracy. Our approach involves sampling points along the line of sight for each voxel - defined as the line connecting the voxel to the optical center of the camera - and identifying the corresponding 3D locations of these points in historical frames based on relative camera poses. We then sample the features of these points in historical frames and integrate them with the current voxel features to construct a cost volume feature map. This map is subsequently utilized to refine the features of the current volume, thereby improving occupancy prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.465 + ], + "angle": 0, + "content": "Compared with paradigms (a) and (b) in Figure 1, our proposed CVT-Occ explicitly utilizes parallax cues to refine the depth of 3D voxel. In contrast to paradigm (c), our approach avoids per-image-pair cost volume computation, achieving superior performance with minimal additional computational overhead. CVT-Occ distinguishes itself by leveraging the inherent parallax information in historical observations and employing a data-driven approach to learning the cost volume. Our experiments on the Occ3D-Waymo dataset demonstrate that CVT-Occ achieves state-of-the-art 3D occupancy prediction performance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.483, + 0.388, + 0.499 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.51, + 0.475, + 0.526 + ], + "angle": 0, + "content": "2.1 3D Occupancy Prediction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.788, + 0.773 + ], + "angle": 0, + "content": "The goal of 3D occupancy prediction is to estimate the occupancy of each quantized voxel in 3D space. This task is originally rooted in Occupancy Grid Mapping (OGM) [24, 28] for mobile robot navigation, where the robot is equipped with range sensors (e.g., LiDAR) and navigates in static environments. Recent works transit to a more generic scenario: they make the occupancy prediction with visual systems in dynamic environments. MonoScene [1] reconstructs the 3D scene from sparse prediction with LiDAR points. VoxFormer [16] makes dense voxel predictions leveraging semantic labels and depth estimation from monocular RGB images. TPVFormer [11] proposes a tri-perspective view decomposition method for efficient 3D occupancy prediction. OccFormer [40] presents a dual-path transformer network to effectively process the 3D volume for semantic occupancy prediction. Besides, these advancements have led to new benchmarks such as visual 3D occupancy prediction [29, 30, 32] and semantic scene completion [15]. We emphasize that our proposed method is a simple and plug-and-play module that seamlessly integrates with existing occupancy prediction pipelines, significantly improving their performance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.79, + 0.555, + 0.805 + ], + "angle": 0, + "content": "2.2 Temporal Fusion for 3D Perception" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Leveraging temporal information for occupancy prediction is a natural strategy since it provides sufficient spatial information for building geometric repre" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "Z. Ye et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.345 + ], + "angle": 0, + "content": "sentations [25, 36]. A common practice of temporal fusion in visual 3D object detection is BEV feature alignment. It fuses temporal information by warping past BEV features to the current time according to camera view transformation from relative camera poses at different time steps [8, 20]. It has been shown that short-term frames can significantly help improve the 3D detection performance. However, we find that minor improvements are brought up when more frames come in with simply warping and fusion. Recent work UniFusion [27] proposes a new long temporal fusion method. It creates \"virtual views\" for the temporal features as if they are present in the current time and directly access all useful history features. However, this method suffers from intensive computation costs when the number of frames increases. SOLOFusion [25] is a more efficient and effective model that balances the resolution and time stamps, and transforms the temporal fusion problem into a temporal multi-view stereo problem." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.367, + 0.593, + 0.383 + ], + "angle": 0, + "content": "2.3 Stereo Matching and Multi-View Stereo" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.789, + 0.635 + ], + "angle": 0, + "content": "Stereo matching is rooted in constructing 3D cost volumes from 2D image features [3, 33, 39] and predicting the depth map from the cost volume. Recent works further advance it by proposing correlation-based cost volumes like GC-Net [2]. Meanwhile, multi-view stereo methods (5,6,10,12,31,38)) leverage plane-sweep volumes for depth map generation. A recent trend of using stereo matching [4, 17, 19] focused on 3D perception in autonomous driving. While existing work in stereo matching often utilizes stereo image pairs, this format is unsuitable for handling multi-view and multi-temporal inputs in autonomous driving scenarios. Additionally, these approaches require generating a Plane-Sweep Volume for each image pair, leading to inefficiencies. Furthermore, recent advancements like OccDepth [23] demonstrate the potential of implicitly learning correlations between stereo images to improve 3D depth-aware feature fusion. Our proposed Temporal Cost Volume method, designed specifically for multi-view and multi-temporal data, offers a more effective and efficient solution. The core innovation of the CVT module lies in its ability to construct a comprehensive 3D representation of the scene by aggregating information over time." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.66, + 0.381, + 0.679 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.693, + 0.387, + 0.71 + ], + "angle": 0, + "content": "3.1 Problem Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Given only RGB images as inputs, the model is designed to predict a dense semantic scene within a specified volume. Specifically, we utilize as input the current timestamp \\( t \\) along with previous images, denoted by \\( \\mathbf{I} = \\{\\mathbf{I}_t^i,\\mathbf{I}_{t - 1}^i,\\ldots \\}_{i = 1}^L \\), where \\( L \\) represents the number of cameras. Our output is defined as a voxel grid \\( Y_{t}\\in \\{c_{0},c_{1},\\dots ,c_{M}\\}^{H\\times W\\times Z} \\), situated in the coordinate frame of the ego-vehicle at timestamp \\( t \\). Each voxel within this grid is either empty (indicated by \\( c_{0} \\)) or occupied by a specific semantic class among \\( \\{c_1,c_2,\\dots ,c_M\\} \\). Here, \\( M \\) denotes the total number of classes of interest, while \\( H \\), \\( W \\), and \\( Z \\) represent the length," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.258, + 0.114, + 0.731, + 0.129 + ], + "angle": 0, + "content": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.239, + 0.15, + 0.758, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.409, + 0.788, + 0.548 + ], + "angle": 0, + "content": "Fig. 2: Overall Architecture of CVT-Occ. The image backbone extracts multiscale features from multi-view images, which are transformed into 3D volume features denoted as \\(\\mathbf{V} \\in \\mathbb{R}^{H \\times W \\times Z \\times C}\\). The Cost Volume Temporal Module samples points along the line of sight within the current volume and projects them onto \\(K - 1\\) historical frames, resulting in \\(K \\times N\\) 3D volume features. These features are concatenated to construct cost volume features \\(\\mathbf{F} \\in \\mathbb{R}^{H \\times W \\times Z \\times (K \\times N) \\times C}\\). Convolution layers are then applied to generate weights \\(\\mathbf{W} \\in \\mathbb{R}^{H \\times W \\times Z}\\), refining the depth of 3D voxel. Finally, an occupancy decoder produces 3D semantic occupancy predictions. In the figure, \\((\\text{\\textcircled{A}})\\), \\((\\text{\\textcircled{C}})\\), and \\(\\otimes\\) represent coordinate alignment, concatenation, and element-wise product, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.788, + 0.685 + ], + "angle": 0, + "content": "width, and height of the grid, respectively. The primary goal is to train a neural network \\(\\Theta(\\cdot)\\) to produce a semantic voxel grid \\(Y_{t} = \\Theta(\\mathbf{I})\\) that approximates the ground truth \\(\\hat{Y}_{t}\\). It is important to note that the model must learn both 3D geometry and semantics solely from vision-based inputs, without the aid of depth measurements from LiDAR. Inferring 3D occupancy from 2D images presents a significant challenge and necessitates the design of efficient methods to leverage geometric constraints for learning accurate geometry." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.708, + 0.433, + 0.723 + ], + "angle": 0, + "content": "3.2 Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In this section, we depict the comprehensive architecture of CVT-Occ, depicted in Figure 2. Our framework processes multi-frame, multi-view images to first extract multi-scale features through an image backbone. Subsequently, these features from image space are transformed into BEV space features, which are refined by a BEV encoder to generate 3D volumetric representations. There is a rich corpus of research focused on the transition from image space to BEV features. One line of works follows the lifting paradigm proposed in LSS [26];" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "Z. Ye et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.329 + ], + "angle": 0, + "content": "they explicitly predict a depth map and lift multi-view image features onto the BEV plane [8, 9, 17, 18, 25]. Another line of works inherits the spirit of querying from 3D to 2D in DETR3D [34]; they employ learnable queries to extract information from image features by cross-attention mechanism [13, 20, 21]. It is crucial to underscore that our proposed Cost Volume Temporal (CVT) module is versatile and compatible across different strategies for feature transformation from image to volume space. For our experimental validation, we employ BEV-Former [20] as the foundation to generate 3D volume features. These features are subsequently refined and enhanced through the CVT module, demonstrating the efficacy and adaptability of our approach in leveraging temporal and spatial dynamics for enriched 3D scene comprehension. Finally an occupancy decoder is used to generate final voxel prediction results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.352, + 0.523, + 0.368 + ], + "angle": 0, + "content": "3.3 Cost Volume Temporal Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.788, + 0.499 + ], + "angle": 0, + "content": "Since direct depth information is not available for each pixel, transforming image features into 3D space introduces ambiguity; for instance, a single pixel may correspond to multiple voxels along the line of sight. To address this challenge, we propose the Cost Volume Temporal module, which leverages temporal data to infer depth information and resolve this ambiguity. Specifically, we construct 3D cost volume features using both historical and current BEV features. These cost volume features are then used to derive learned weights, which are subsequently employed to refine the current BEV features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.5, + 0.788, + 0.591 + ], + "angle": 0, + "content": "3D Volume Features. We pre-define the volume \\(\\mathbf{V} \\in \\mathbb{R}^{H \\times W \\times Z}\\), where \\(H, W, Z\\) represent the spatial dimensions of the volume space. Each voxel in the volume space corresponds to a cubic space in the real world with a size of \\(s\\) meters. By default, the center of the volume corresponds to the position of the ego car. For each voxel \\(v = (i, j, k)\\), its corresponding 3D location \\(p = (x, y, z)\\) can be calculated by Eqn. 1:" + }, + { + "type": "equation", + "bbox": [ + 0.285, + 0.602, + 0.787, + 0.637 + ], + "angle": 0, + "content": "\\[\nx = \\left(i - \\frac {W}{2}\\right) \\times s, y = \\left(j - \\frac {H}{2}\\right) \\times s, z = \\left(k - \\frac {Z}{2}\\right) \\times s (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.647, + 0.788, + 0.693 + ], + "angle": 0, + "content": "The BEV feature is \\(\\mathbf{F}_t^{\\mathrm{bev}}\\in \\mathbb{R}^{H\\times W\\times E}\\), where \\(t\\) represents the timestamp and \\(E\\) is the embedding dimensions. We reshape the BEV feature to volume space as the input of the CVT module, producing \\(\\mathbf{V}_t\\in \\mathbb{R}^{H\\times W\\times Z\\times C}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.693, + 0.788, + 0.77 + ], + "angle": 0, + "content": "Construct Cost Volume Feature. Given the ambiguity along the line of sight, for each point \\(\\mathbf{p_t}\\), we sample several additional points \\(\\{\\mathbf{p_t^i}\\}_{i=1}^N\\) within the current volume. Specifically, we calculate the sight direction \\(\\mathbf{d_t} \\in \\mathbb{R}^3\\), which is the vector from the center of the volume to the point \\(\\mathbf{p_t}\\), and then sample points using specific strides \\(\\{n^i\\}_{i=1}^N\\) as shown in Eqn. 2:" + }, + { + "type": "equation", + "bbox": [ + 0.438, + 0.781, + 0.786, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {\\mathbf {t}} ^ {\\mathbf {i}} = \\mathbf {p} _ {\\mathbf {t}} + \\mathbf {d} _ {\\mathbf {t}} \\times \\mathbf {n} ^ {\\mathbf {i}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Since these points project to the same pixel in the image space, points along the same line tend to have similar features. To accurately distinguish the correct" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.258, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.283 + ], + "angle": 0, + "content": "position corresponding to the pixel, we use historical BEV features to obtain complementary information. Here comes the core insight of our proposed Cost Volume Temporal Module. Projecting these points into the historical coordinate frame ensures they are no longer in the same line of sight. This parallax provides additional information from the historical BEV features, which helps to reduce depth ambiguity in the current frame. Projection matrix \\(\\mathbf{P}_{\\mathbf{t}} \\in \\mathbb{R}^{4 \\times 4}\\) can transform points from the coordinate frame of ego-vehicle to the global coordinate frame. Therefore the points \\(\\{\\mathbf{p}_{\\mathbf{t}}^i\\}_{i=1}^N\\) are projected onto \\(K - 1\\) historical frames through the projection matrix as described in Eqn. 3:" + }, + { + "type": "equation", + "bbox": [ + 0.437, + 0.296, + 0.785, + 0.314 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {\\mathbf {t} - \\mathbf {k}} ^ {\\mathrm {i}} = \\mathbf {p} _ {\\mathbf {t}} ^ {\\mathrm {i}} \\mathbf {P} _ {\\mathbf {t}} \\mathbf {P} _ {\\mathbf {t} - \\mathbf {k}} ^ {- 1} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.327, + 0.784, + 0.388 + ], + "angle": 0, + "content": "Finally, each point \\(\\mathbf{p}_{\\mathbf{t} - \\mathbf{k}}^{\\mathbf{i}}\\) is converted to voxel coordinates, which is the inverse process of Eqn. 1, and bilinear interpolation is used to sample features from corresponding BEV feature maps. The final cost volume feature map is \\(\\mathbf{F} \\in \\mathbb{R}^{H \\times W \\times Z \\times (K \\times N) \\times C}\\), where \\(C\\) represents the channels of the BEV features." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.414, + 0.501, + 0.427 + ], + "angle": 0, + "content": "3.4 Volume Features Refinement" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.445, + 0.785, + 0.46 + ], + "angle": 0, + "content": "\\[\n\\mathbf {V} _ {\\mathrm {o c c}} = \\mathbf {V} _ {t} \\otimes \\operatorname {S i g m o i d} (\\operatorname {C o n v} (\\mathbf {F})) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.788, + 0.651 + ], + "angle": 0, + "content": "As described in the Eqn. 4, the constructed cost volume features \\(\\mathbf{F}\\) are processed through a sequence of convolutional layers, yielding an output weight \\(\\mathbf{W} \\in \\mathbb{R}^{H \\times W \\times Z}\\). A Sigmoid activation function then normalizes the output weights \\(\\mathbf{W}\\) to the range [0,1]. These weights are directly supervised according to the voxel occupancy state: weights corresponding to occupied voxels are encouraged to reach 1, while those for unoccupied voxels are guided towards 0. The symbol \\(\\otimes\\) in the Eqn. 4 denotes element-wise product of the original volume features at timestamp \\(t\\) with the learned weights \\(\\mathbf{W}\\). This supervised learning approach produces an occupancy-aware volume feature map \\(\\mathbf{V}_{\\mathrm{occ}} \\in \\mathbb{R}^{H \\times W \\times Z \\times C}\\). The goal of the learned weights is to diminish the influence of voxel features in incorrectly activated regions due to depth ambiguity and augment the features of correctly identified voxels simultaneously." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.676, + 0.427, + 0.691 + ], + "angle": 0, + "content": "3.5 Occupancy Decoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Upon the refinement of voxel features \\(\\mathbf{V}_{\\mathrm{occ}}\\), our model employs a series of deconvolution layers to transform it into occupancy features. The occupancy features are projected onto the output space, yielding \\(\\mathbf{X} \\in \\mathbb{R}^{H \\times W \\times Z \\times M}\\). The projection is formulated to map the occupancy features into a discrete set of semantic class predictions for each voxel within the grid. Consequently, the final occupancy output \\(\\mathbf{Y} \\in \\{c_0, c_1, \\ldots, c_M\\}^{H \\times W \\times Z}\\) is generated, where each voxel is assigned one of \\(M\\) semantic labels, including an unoccupied state represented by \\(c_0\\). This step effectively translates the enriched voxel features into a semantically segmented 3D occupancy map." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "Z. Ye et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.375, + 0.162 + ], + "angle": 0, + "content": "3.6 Training Loss" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.168, + 0.785, + 0.211 + ], + "angle": 0, + "content": "Our architecture integrates the CVT module and 3D occupancy prediction within an end-to-end trainable framework. We train the overall network with a multitask loss as" + }, + { + "type": "equation", + "bbox": [ + 0.43, + 0.215, + 0.785, + 0.23 + ], + "angle": 0, + "content": "\\[\n\\operatorname {L o s s} = \\mathcal {L} _ {\\text {o c c}} + \\lambda \\mathcal {L} _ {\\text {c v t}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.233, + 0.784, + 0.278 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{occ}}\\) denotes the occupancy prediction loss, and \\(\\mathcal{L}_{\\mathrm{cvt}}\\) represents the cost volume temporal loss. The coefficient \\(\\lambda\\) serves as a balancing factor between these two loss components." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.279, + 0.798, + 0.368 + ], + "angle": 0, + "content": "For the loss of occupancy prediction \\(\\mathcal{L}_{\\mathrm{occ}}\\), we adopt the cross-entropy loss, which is widely used in semantic occupancy prediction task. \\(\\mathcal{L}_{\\mathrm{occ}} = \\sum_{c}w_{c}\\mathcal{L}(g_{c},p_{c})\\) where \\(g_{c}\\) and \\(p_c\\) represent the ground truth label and the prediction result for the \\(c\\)-th semantic class, respectively. Class-specific weighting \\(w_{c}\\) mitigates class imbalance by inversely correlating with class frequency, as adopted from Occ3D [29]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.785, + 0.415 + ], + "angle": 0, + "content": "To enhance the network's performance, the CVT module receives direct supervision. The loss attributed to the CVT module, \\(\\mathcal{L}_{\\mathrm{cvt}}\\), employs a binary cross-entropy framework as shown in Eqn. 6:" + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.422, + 0.785, + 0.463 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c v t}} = - \\sum_ {j = 1} ^ {J} \\hat {y} _ {j, 0} \\log \\left(1 - w _ {j}\\right) + \\hat {y} _ {j, 1} \\log w _ {j} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.788, + 0.561 + ], + "angle": 0, + "content": "where \\( j \\) indexes voxels in the grid, with \\( J = H \\times W \\times Z \\). \\( w_{k} \\) is the \\( j \\)-th element in the output weight \\( \\mathbf{W} \\in \\mathbb{R}^{H \\times W \\times Z} \\). Specifically, a binary label \\( \\hat{y}_{j,0} = 1 \\) signifies the voxel's unoccupied status (i.e., its semantic classification is \\( c_{0} \\)), while \\( \\hat{y}_{j,1} = 1 \\) indicates the voxel is occupied in ground truth. This direct supervision on the CVT module not only improves performance but also underpins the importance of accurately learning temporal and spatial features from the input data." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.58, + 0.367, + 0.597 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.608, + 0.327, + 0.621 + ], + "angle": 0, + "content": "4.1 Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Occ3D-Waymo. We conduct experiments on the Occ3D-Waymo [29] dataset, which provides dense semantic labels for 3D occupancy grids. Occ3D-Waymo consists of 798 training scenes and 202 validation scenes. On the Occ3D-Waymo dataset, the range for the \\( x \\) and \\( y \\) axis as \\([-40m, 40m]\\), and for the \\( z \\) axis as \\([-1m, 5.4m]\\). The voxel grid size is \\((0.4m, 0.4m, 0.4m)\\), resulting in a resolution of \\((200 \\times 200 \\times 16)\\) for \\((H, W, Z)\\). While the semantic labels encompass 16 categories, including \"Free\" class. Besides, it also provides visibility masks for LiDAR and camera modality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Evaluation Metrics. In our assessment of 3D semantic occupancy prediction capabilities, we utilize the IoU of each class to evaluate 3D semantic occupancy prediction. As Occ3D-Waymo provides camera visibility masks, only voxels within the camera's visible region are evaluated. The mean Intersection over Union (mIoU) is computed over 14 non-Free classes, excluding the \"Motorcycle\" class due to its insufficient voxel numbers." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.258, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.446, + 0.163 + ], + "angle": 0, + "content": "4.2 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.414 + ], + "angle": 0, + "content": "Architecture Detail. Our architecture leverages the same image backbone and image neck as Occ3D [29] for image feature extraction. The transformation of image features to BEV space involves a four-layered view transformation layers, each layer comprising a sequence of cross-attention layers, normalization layers, feed-forward layers, and normalization layers. This BEV encoder generates a BEV feature representation \\(\\mathbf{F}^{\\mathrm{bev}}\\in \\mathbb{R}^{200\\times 200\\times 256}\\), which is further reshaped to volume features \\(\\mathbf{V}\\in \\mathbb{R}^{200\\times 200\\times 16\\times 16}\\). The selected frame configuration includes a time interval of \\(0.5s\\) and a sequence of 7 frames, covering a total time span of \\(3s\\), which ensures a comprehensive temporal analysis. For the CVT module, by default, it samples 9 points within each volume feature across 7 frames to capture a broad spectrum of temporal and spatial variations. Consequently, this configuration enables the CVT module to generate a cost volume feature map \\(\\mathbf{F}\\in \\mathbb{R}^{200\\times 200\\times 16\\times (7\\times 9)\\times 16}\\). The occupancy decoder utilizes a series of deconvolution layers to transform the volume features into occupancy features, which is then followed by two additional convolution blocks responsible for generating the final 3D semantic occupancy predictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.788, + 0.52 + ], + "angle": 0, + "content": "Training. During training, we utilized the AdamW [22] optimizer for 8 epochs and full dataset each epoch. Training was conducted on 8 NVIDIA A100 GPUs with a batch size of 1 per GPU. Learning Rate Scheduler Policy is the cosine annealing schedule with an initial learning rate of \\(4 \\times 10^{-4}\\). The dimensions for the input images were standardized to \\(960 \\times 640\\) pixels. It is noteworthy that we abstained from employing any data augmentation techniques in order to ensure a fair and unbiased comparison among the models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.543, + 0.373, + 0.557 + ], + "angle": 0, + "content": "4.3 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.735 + ], + "angle": 0, + "content": "As shown in Figure 1, We categorize previous temporal fusion methods into three types: (a) Attention-based methods, which adopt temporal self-attention to recurrently fuse the history BEV information; (b) WrapConcat-based methods, involving the alignment of BEV feature maps across different time steps; and (c) Plane-Sweep-based methods, which construct a cost volume for image pairs to utilize geometric constraints. Additionally, we introduce a novel approach, CVT-Occ, that utilizes geometric correspondences of 3D voxels over time to improve occupancy prediction accuracy. For comparative analysis, we extended some mainstream BEV models for the 3D semantic occupancy prediction task. We choose BEVFormer [20], BEVFormer-WarpConcat [37], and SOLOFusion [25] as representative methods for categories (a), (b), and (c), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "For both BEVFormer [20] and SOLOFusion [25], we have replaced their original transformer decoders for object detection with a 3D occupancy prediction decoder. Additionally, for fairness, we omit the depth supervision module when reproducing SOLOFusion [25]. For BEVFormer-WrapConcat, we adopt a commonly used fusion method known as the Wrap and Concatenate strategy, as seen in BEVFormerv2 [37] and PanoOcc [35]. This strategy involves 3D coordinate alignment in BEV space based on the transformation matrix between historical" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.127 + ], + "angle": 0, + "content": "Z. Ye et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.787, + 0.215 + ], + "angle": 0, + "content": "Table 1: 3D Occupancy Prediction Results on Occ3D-Waymo. This comparison showcases the performance of various principal temporal fusion strategies alongside CVT-Occ, within the same architectural framework and identical frame specifications (number and interval), ensuring a direct and fair comparison across all evaluated methods. The leading performance is highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.227, + 0.787, + 0.332 + ], + "angle": 0, + "content": "
MethodmIoUGoVehiclePedestrianSignBicyclistTraffic LightPoleCons. ConeBicycleBuildingVegetationTree TrunkRoadWalkable
BEVFormer-w/o TSA23.877.5034.5421.079.6920.9611.4811.4814.0614.5123.1421.828.5778.4556.89
BEVFormer [20]24.587.1836.0621.009.7620.2312.6114.5214.7016.0623.9822.509.3979.1157.04
SOLOFusion [25]24.734.9732.4518.2810.3317.148.0717.8316.2319.331.4928.9816.9370.9553.28
BEVFormer-WrapConcat25.076.236.1720.959.5620.5812.8216.2414.3116.7825.1423.5612.8179.0456.83
CVT-Occ(ours)27.377.4441.023.9311.9220.8112.0718.0316.8821.3729.427.4214.6779.1259.09
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.358, + 0.787, + 0.509 + ], + "angle": 0, + "content": "and current frames. Subsequently, it concatenates previous BEV features with current BEV features along the channel dimension and applies residual blocks for temporal fusion. We incorporate the residual block proposed by PanoOcc [35] to create the BEVFormer-WrapConcat method. In the \"BEVFormer-w/o TSA\" configuration, we maintain the original architecture of BEVFormer [20] but exclude the Temporal Self-Attention(TSA) layers, thereby eliminating temporal attention mechanisms. In Table 1, we present the evaluation results for 3D occupancy prediction on the Occ3D-Waymo validation set. To ensure a fair comparison, all methods were implemented using the same network architecture and training strategy." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.509, + 0.787, + 0.569 + ], + "angle": 0, + "content": "Comparing \"BEVFormer-w/o TSA\" with temporal fusion methods in Table 1, All temporal fusion methods demonstrate improved mIoU compared to \"BEVFormer-w/o TSA\". It becomes evident that temporal fusion plays a crucial role in the 3D occupancy prediction task." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.57, + 0.787, + 0.766 + ], + "angle": 0, + "content": "Maintaining consistent frame specifications (number and interval), we explore the temporal fusion ability of different mainstream temporal fusion methods. Our method achieves state-of-the-art performance in terms of IoU and demonstrates significant superiority over other approaches across nearly all classes. As shown in Table 1, CVT-Occ exhibits superior performance compared to representative models of existing temporal fusion methods (a), (b), and (c) in Figure 1, achieving a notable \\(2.8\\%\\) IoU improvement over the baseline BEVFormer [20]. Delving into the details, CVT-Occ shows improvements in IoU across almost all classes. Specifically, for \"Vehicle\", \"Bicycle\", \"Building\", \"Vegetation\", and \"Tree Trunk\", our method achieves more than a \\(4\\%\\) increase in IoU compared to the baseline model BEVFormer [20]. This substantial enhancement across nearly all categories underscores that a comprehensive understanding of temporal geometry can lead to significant performance gains." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.786, + 0.388, + 0.802 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Time Span. The time span is defined as the difference between the timestamp of the earliest historical frame and that of the current frame, which plays a" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.258, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.145, + 0.481, + 0.297 + ], + "angle": 0, + "content": "
ExpFrameIntervalCVT supervisionmIoU
(a)30.1s✓25.15
(b)30.3s✓25.86
(c)70.1s✓25.80
(d)70.3s✓26.46
(e)70.5s✓27.37
(f)70.5s26.81
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.306, + 0.477, + 0.361 + ], + "angle": 0, + "content": "Table 2: Ablation Experiments. We evaluate CVT-Occ under different frame specifications and CVT supervision." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.151, + 0.785, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.334, + 0.788, + 0.364 + ], + "angle": 0, + "content": "Fig. 3: Experiments of Different Time Spans." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.784, + 0.44 + ], + "angle": 0, + "content": "A crucial role in the temporal fusion analysis. We conduct experiments to evaluate our method across various configurations of total frame numbers and sampled frame intervals, as summarized in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.595 + ], + "angle": 0, + "content": "In experiments (a) and (c), we compared the impact of different total frame numbers. The results demonstrate that a longer sequence of historical frames leads to improved performance. Similarly, we observed that a larger time interval between adjacent frames also contributes to enhanced performance in experiments (c), (d), and (e). These consistent findings underscore the significance of a longer time span across the entire temporal sequence for better model performance. CVT-Occ relies on disparities between frames to mitigate depth ambiguity, where larger intervals and frame numbers contribute to an extended time span, amplifying the temporal parallax. This elucidates the observed increase in mIoU with a longer time span." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.687 + ], + "angle": 0, + "content": "In addition, comparing experiments (b) and (c), where the total time span remains the same 0.6s, highlights the critical importance of the total time span even when the frame interval and number of historical frames varies between these two experiments. This finding further emphasizes that the impact of our proposed method arises from leveraging the parallax of the entire temporal sequence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Ability of Long Temporal Fusion. Figure 3 illustrates the comparison between BEVFormer [20] and CVT-Occ across varying total time differences. CVT-Occ demonstrates improved performance with increasing temporal queue lengths, whereas BEVFormer struggles to effectively leverage information from long-range historical data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "The limitation of BEVFormer's long temporal fusion capability has been highlighted in previous researches, such as UniFusion [27] and VideoBEV [7]. Based on results of Figure 3, we propose that BEVFormer fails to benefit from longer time spans due to its recurrent temporal fusion process. This process restricts the direct access to information from distant historical frames, regardless" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "Z. Ye et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.244 + ], + "angle": 0, + "content": "Table 3: Comparison Results under Different Conditions. Here are three sets of experiments. (a) Binary Classification: this task focuses on geometric occupancy prediction, treating all object classes as a single \"Non-Free\" class. (b) BEV Range: Along x-axis, [0,20] means the range \\([-20m,0m]\\cup [0,20m]\\) and [20,40] means the range \\([-40m, - 20m]\\cup [20, - 40m]\\). (c) Ego Vehicle Speed: The 202 scenes for evaluation is partitioned into 2 parts based on the ego-vehicle speed. We evaluate all temporal fusion methods on them separately." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.254, + 0.784, + 0.327 + ], + "angle": 0, + "content": "
ExpmIoUNon-FreeFree[0,20][20,40]FastSlow
BEVFormer [20]24.5855.7486.5427.0921.2424.7324.22
BEVFormer-WrapConcat25.0756.5786.9628.2121.1626.2724.44
CVT-Occ27.3762.9590.4630.8522.6727.926.79
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.784, + 0.445 + ], + "angle": 0, + "content": "of the temporal sequence length. In contrast, the cost volume feature \\(\\mathbf{F}\\) can directly integrate information from all historical frames. The CVT module utilizes parallax correlation explicitly to refine volume features, leading to a more robust understanding of temporal information. This design overcomes the shortcomings observed in BEVFormer [20], resulting in enhanced long-term temporal fusion capability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.785, + 0.566 + ], + "angle": 0, + "content": "CVT Supervision. In our pursuit to improve 3D semantic occupancy prediction, we go beyond traditional loss functions by adding extra supervision to the CVT module output. Experiments (e) and (f) in Table 2 show the performance comparison of CVT-Occ with and without the CVT loss. Incorporating the CVT loss results in a significant increase in mIoU, demonstrating its effectiveness through experimental results. This enhancement validates our design approach, highlighting its ability to enhance the accuracy and reliability of semantic occupancy predictions in 3D." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.787, + 0.643 + ], + "angle": 0, + "content": "In conclusion, our ablation study confirms the critical importance of longer time spans and effective temporal fusion in enhancing CVT-Occ's performance for 3D semantic occupancy prediction. Additionally, integrating supplementary supervision further improves model accuracy, highlighting our approach's effectiveness in advancing temporal parallax-based scene understanding." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.664, + 0.421, + 0.68 + ], + "angle": 0, + "content": "4.5 Analysis of Factors" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.764 + ], + "angle": 0, + "content": "The core insight of our method lies in leveraging parallax between different frames to enhance the accuracy of 3D voxel depth queries. In this subsection, we conduct several evaluations under various conditions to validate this concept. These experiments include binary classification, different BEV ranges, and varying ego vehicle speeds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Binary Classification. Within the framework of 3D semantic occupancy prediction, we aggregate all 15 other object classes into a single \"Non-Free\" class. This adjustment removes the influence of semantic information, allowing us to focus purely on depth accuracy. We then compare our approach against leading temporal fusion methods. The results in Table 3 demonstrate that our method" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.258, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "provides better predictions on object geometry. This outcome aligns closely with our initial design concept of leveraging parallax for refined depth prediction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.329 + ], + "angle": 0, + "content": "BEV Range. In the Occ3D-Waymo dataset, the BEV range spans from \\(-40m\\) to \\(40m\\) along the x-axis (i.e., the ego car's direction of movement). We divide this range into two parts: \\([-20m, 20m]\\) and \\([-40m, 20m] \\cup [20m, 40m]\\). Table 3 exhibits the evaluation results. All methods perform better in the near region, which is expected since objects closer to the vehicle are more likely to appear in the overlap region of history and current frames. However, our method exhibits significantly greater improvement in the near region (+3.76) compared to the far region (+1.43) compared to BEVFormer [20]. The larger parallax of nearer objects between different frames contributes to our method's superior prediction capability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.787, + 0.466 + ], + "angle": 0, + "content": "Ego Vehicle Speed. Moreover, the speed of the vehicle also plays a crucial role in determining the magnitude of parallax. We further divide the dataset into two roughly equal parts based on the speed of ego vehicle. Evaluating CVT-Occ separately on these parts yields the results shown in Table 3. Regardless of whether BEVFormer [20] or CVT-Occ, both perform better in fast-moving scenarios. Our method demonstrates much greater improvement under the fast condition \\((+3.17)\\) than slow one \\((+2.57)\\) compared to the baseline BEVFormer. This can be attributed to the larger parallax between history and current frames when the ego car is moving rapidly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.787, + 0.542 + ], + "angle": 0, + "content": "To sum up, these results underscore the significantly enhanced performance of CVT-Occ compared to the baseline model BEVFormer [20] under a wide range of conditions. The evaluations demonstrate that our method can refine the depth estimation of 3D voxel queries more accurately, with a notable advantage in the near region and fast scenes due to the greater parallax effect." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.568, + 0.436, + 0.582 + ], + "angle": 0, + "content": "4.6 Visualization Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.787, + 0.732 + ], + "angle": 0, + "content": "Figure 4 offers a qualitative comparison of CVT-Occ against the baseline method BEVFormer [20] on the Occ3D-Waymo dataset. To ensure clarity in visualization, a dataset-specific mask is applied to exclude regions not within the camera's field of view. These visual representations clearly illustrate the superiority of CVT-Occ over competing approaches. BEVFormer encounters difficulties with depth perception, leading to issues such as atmospheric distortion and blurring of objects. CVT-Occ marks a remarkable improvements in the accurate depiction of buildings and vegetation, thus providing a more precise and reliable interpretation of the environment." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.76, + 0.509, + 0.775 + ], + "angle": 0, + "content": "5 Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We have presented CVT-Occ, a novel approach that significantly enhances the accuracy of 3D occupancy predictions by leveraging temporal fusion and geometric correspondence across time. Unlike traditional methods that rely on" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.127 + ], + "angle": 0, + "content": "Z. Ye et al." + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.146, + 0.755, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.628, + 0.788, + 0.657 + ], + "angle": 0, + "content": "Fig. 4: Qualitative Results. CVT-Occ exhibits superior performance in predicting the occupancy of vegetation and buildings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "monocular or stereo vision, our method utilizes multi-view temporal fusion, effectively incorporating historical observations to exploit the parallax effect and improve depth estimation. Our innovative approach, which constructs a cost volume feature map by sampling and integrating features across temporal frames, demonstrates superior performance on the Occ3D-Waymo dataset, surpassing current state-of-the-art models in terms of accuracy while maintaining a low computational overhead. CVT-Occ also opens avenues for future research on temporal fusion for 3D perception, e.g., applying the cost volume module in other tasks like 3D reconstruction, and other applications like robotics and virtual reality." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.258, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "Acknowledgments. This work is supported by the National Key R&D Program of China (2022ZD0161700)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.203, + 0.323, + 0.219 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.238, + 0.785, + 0.279 + ], + "angle": 0, + "content": "1. Cao, A.Q., de Charette, R.: Monoscene: Monocular 3d semantic scene completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3991-4001 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.281, + 0.785, + 0.322 + ], + "angle": 0, + "content": "2. Cao, Y., Xu, J., Lin, S., Wei, F., Hu, H.: Gcnet: Non-local networks meet squeeze-excitation networks and beyond. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops (Oct 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.323, + 0.785, + 0.363 + ], + "angle": 0, + "content": "3. Chang, J.R., Chen, Y.S.: Pyramid stereo matching network. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.365, + 0.785, + 0.405 + ], + "angle": 0, + "content": "4. Chen, Y., Liu, S., Shen, X., Jia, J.: Dsgn: Deep stereo geometry network for 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.407, + 0.785, + 0.462 + ], + "angle": 0, + "content": "5. Cheng, S., Xu, Z., Zhu, S., Li, Z., Li, L.E., Ramamoorthi, R., Su, H.: Deep stereo using adaptive thin volume representation with uncertainty awareness. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.463, + 0.785, + 0.517 + ], + "angle": 0, + "content": "6. Gu, X., Fan, Z., Zhu, S., Dai, Z., Tan, F., Tan, P.: Cascade cost volume for high-resolution multi-view stereo and stereo matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.519, + 0.785, + 0.559 + ], + "angle": 0, + "content": "7. Han, C., Yang, J., Sun, J., Ge, Z., Dong, R., Zhou, H., Mao, W., Peng, Y., Zhang, X.: Exploring recurrent long-term temporal fusion for multi-view 3d perception. IEEE Robotics and Automation Letters 9(7), 6544-6551 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.561, + 0.785, + 0.587 + ], + "angle": 0, + "content": "8. Huang, J., Huang, G.: Bevdet4d: Exploit temporal cues in multi-camera 3d object detection (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.589, + 0.785, + 0.616 + ], + "angle": 0, + "content": "9. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.617, + 0.785, + 0.657 + ], + "angle": 0, + "content": "10. Huang, P.H., Matzen, K., Kopf, J., Ahuja, N., Huang, J.B.: Deepmvs: Learning multi-view stereopsis. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.785, + 0.713 + ], + "angle": 0, + "content": "1. Huang, Y., Zheng, W., Zhang, Y., Zhou, J., Lu, J.: Tri-perspective view for vision-based 3d semantic occupancy prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9223-9232 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.715, + 0.785, + 0.743 + ], + "angle": 0, + "content": "2. Im, S., Jeon, H.G., Lin, S., Kweon, I.S.: DPSNet: End-to-end deep plane sweep stereo. In: International Conference on Learning Representations (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.784 + ], + "angle": 0, + "content": "3. Jiang, Y., Zhang, L., Miao, Z., Zhu, X., Gao, J., Hu, W., Jiang, Y.G.: Polarformer: Multi-camera 3d object detection with polar transformer. Proceedings of the AAAI Conference on Artificial Intelligence 37(1), 1042-1050 (Jun 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "4. Li, H., Li, Z., Akmandor, N.U., Jiang, H., Wang, Y., Padir, T.: Stereovoxelnet: Real-time obstacle detection based on occupancy voxels from a stereo camera using deep neural networks. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 4826-4833 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.238, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.127 + ], + "angle": 0, + "content": "Z. Ye et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "15. Li, Y., Li, S., Liu, X., Gong, M., Li, K., Chen, N., Wang, Z., Li, Z., Jiang, T., Yu, F., Wang, Y., Zhao, H., Yu, Z., Feng, C.: Sscbench: Monocular 3d semantic scene completion benchmark in street views (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.247 + ], + "angle": 0, + "content": "16. Li, Y., Yu, Z., Choy, C., Xiao, C., Alvarez, J.M., Fidler, S., Feng, C., Anandkumar, A.: Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9087-9098 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.247, + 0.787, + 0.288 + ], + "angle": 0, + "content": "17. Li, Y., Bao, H., Ge, Z., Yang, J., Sun, J., Li, Z.: Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. Proceedings of the AAAI Conference on Artificial Intelligence 37(2), 1486-1494 (Jun 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.289, + 0.787, + 0.331 + ], + "angle": 0, + "content": "18. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. Proceedings of the AAAI Conference on Artificial Intelligence 37(2), 1477-1485 (Jun 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.787, + 0.36 + ], + "angle": 0, + "content": "19. Li, Y., Yang, J., Sun, J., Bao, H., Ge, Z., Xiao, L.: Bevstereo++: Accurate depth estimation in multi-view 3d object detection via dynamic temporal stereo (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.787, + 0.416 + ], + "angle": 0, + "content": "20. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 1-18 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.787, + 0.472 + ], + "angle": 0, + "content": "21. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X.: Petrov2: A unified framework for 3d perception from multi-camera images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3262-3272 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.502 + ], + "angle": 0, + "content": "22. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.787, + 0.531 + ], + "angle": 0, + "content": "23. Miao, R., Liu, W., Chen, M., Gong, Z., Xu, W., Hu, C., Zhou, S.: Occdepth: A depth-aware method for 3d semantic scene completion (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.531, + 0.787, + 0.572 + ], + "angle": 0, + "content": "24. Moravec, H., Elfes, A.: High resolution maps from wide angle sonar. In: Proceedings. 1985 IEEE International Conference on Robotics and Automation. vol. 2, pp. 116-121 (1985)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.787, + 0.628 + ], + "angle": 0, + "content": "25. Park, J., Xu, C., Yang, S., Keutzer, K., Kitani, K.M., Tomizuka, M., Zhan, W.: Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. In: The Eleventh International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.63, + 0.787, + 0.685 + ], + "angle": 0, + "content": "26. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020. pp. 194-210. Springer International Publishing, Cham (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.787, + 0.741 + ], + "angle": 0, + "content": "27. Qin, Z., Chen, J., Chen, C., Chen, X., Li, X.: Unifusion: Unified multi-view fusion transformer for spatial-temporal representation in bird's-eye-view. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 8690-8699 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.787, + 0.771 + ], + "angle": 0, + "content": "28. Thrun, S.: Probabilistic robotics. Communications of the ACM pp. 52-57 (Mar 2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.787, + 0.841 + ], + "angle": 0, + "content": "29. Tian, X., Jiang, T., Yun, L., Mao, Y., Yang, H., Wang, Y., Wang, Y., Zhao, H.: Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 64318-64330. Curran Associates, Inc. (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.258, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "30. Tong, W., Sima, C., Wang, T., Chen, L., Wu, S., Deng, H., Gu, Y., Lu, L., Luo, P., Lin, D., Li, H.: Scene as occupancy. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 8406-8415 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "31. Wang, K., Shen, S.: Mvdepthnet: Real-time multiview depth estimation neural network. In: 2018 International Conference on 3D Vision (3DV) (Sep 2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.272 + ], + "angle": 0, + "content": "32. Wang, X., Zhu, Z., Xu, W., Zhang, Y., Wei, Y., Chi, X., Ye, Y., Du, D., Lu, J., Wang, X.: Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 17850-17859 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.327 + ], + "angle": 0, + "content": "33. Wang, Y., Lai, Z., Huang, G., Wang, B.H., van der Maaten, L., Campbell, M., Weinberger, K.Q.: Anytime stereo image depth estimation on mobile devices. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 5893-5900 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.396 + ], + "angle": 0, + "content": "34. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Faust, A., Hsu, D., Neumann, G. (eds.) Proceedings of the 5th Conference on Robot Learning. Proceedings of Machine Learning Research, vol. 164, pp. 180-191. PMLR (08-11 Nov 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.453 + ], + "angle": 0, + "content": "35. Wang, Y., Chen, Y., Liao, X., Fan, L., Zhang, Z.: Panoocc: Unified occupancy representation for camera-based 3d panoptic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17158-17168 (June 2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.522 + ], + "angle": 0, + "content": "36. Xu, C., Wu, B., Hou, J., Tsai, S., Li, R., Wang, J., Zhan, W., He, Z., Vajda, P., Keutzer, K., Tomizuka, M.: Nerf-det: Learning geometry-aware volumetric representation for multi-view 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 23320-23330 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.591 + ], + "angle": 0, + "content": "37. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., Zhou, J., Dai, J.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17830-17839 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.632 + ], + "angle": 0, + "content": "38. Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstructured multi-view stereo. In: Proceedings of the European Conference on Computer Vision (ECCV) (September 2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "39. Zhang, F., Prisacariu, V., Yang, R., Torr, P.H.: Ga-net: Guided aggregation net for end-to-end stereo matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.785, + 0.716 + ], + "angle": 0, + "content": "40. Zhang, Y., Zhu, Z., Du, D.: Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9433-9443 (October 2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.716 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CVT-Occ_ Cost Volume Temporal Fusion for 3D Occupancy Prediction/d0af9ebe-697b-4a4d-bc71-225f5f0fd15a_origin.pdf b/2024/CVT-Occ_ Cost Volume Temporal Fusion for 3D Occupancy Prediction/d0af9ebe-697b-4a4d-bc71-225f5f0fd15a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0e8a28c6cd9415b3ea50062516ea3e680ffdeedb --- /dev/null +++ b/2024/CVT-Occ_ Cost Volume Temporal Fusion for 3D Occupancy Prediction/d0af9ebe-697b-4a4d-bc71-225f5f0fd15a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82c4657ece740e51f42224de3a36759eda276b92a90252fd82bb2d3615589958 +size 3889567 diff --git a/2024/CVT-Occ_ Cost Volume Temporal Fusion for 3D Occupancy Prediction/full.md b/2024/CVT-Occ_ Cost Volume Temporal Fusion for 3D Occupancy Prediction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..14aa08eac06abfaa77144f9b1c6718e954f0dc48 --- /dev/null +++ b/2024/CVT-Occ_ Cost Volume Temporal Fusion for 3D Occupancy Prediction/full.md @@ -0,0 +1,260 @@ +# CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction + +Zhangchen Ye $^{1*}$ , Tao Jiang $^{1,2*}$ , Chenfeng Xu $^{3}$ , Yiming Li $^{4}$ , and Hang Zhao $^{1,2,5\boxtimes}$ + +$^{1}$ IIIS, Tsinghua University $^{2}$ Shanghai AI Lab $^{3}$ UC Berkeley $^{4}$ New York University $^{5}$ Shanghai Qi Zhi Institute + +Abstract. Vision-based 3D occupancy prediction is significantly challenged by the inherent limitations of monocular vision in depth estimation. This paper introduces CVT-Occ, a novel approach that leverages temporal fusion through the geometric correspondence of voxels over time to improve the accuracy of 3D occupancy predictions. By sampling points along the line of sight of each voxel and integrating the features of these points from historical frames, we construct a cost volume feature map that refines current volume features for improved prediction outcomes. Our method takes advantage of parallax cues from historical observations and employs a data-driven approach to learn the cost volume. We validate the effectiveness of CVT-Occ through rigorous experiments on the Occ3D-Waymo dataset, where it outperforms state-of-the-art methods in 3D occupancy prediction with minimal additional computational cost. The code is released at https://github.com/Tsinghua-MARS-Lab/CVT-Occ. + +Keywords: 3D Semantic Occupancy Prediction $\cdot$ Temporal Fusion + +# 1 Introduction + +Vision-based 3D semantic occupancy prediction is rapidly evolving in the domain of 3D perception, driven by its critical applications in autonomous driving, robotics, and augmented reality. The task aims to estimate the occupancy state and semantic label of every voxel within a 3D space from visual inputs [11,29, 30,40]. + +Despite its critical importance, 3D occupancy prediction presents significant challenges. When relying solely on monocular vision, these challenges are particularly pronounced due to the inherent ambiguity in estimating depth from a single image. While stereo vision has been proposed as a solution to augment depth estimation accuracy [14], its application remains limited in practice. It is impractical to use stereo cameras for widespread applications in autonomous vehicles and robotic systems because of the requirement for extensive calibration + +![](images/d6581b4708145846a5456243e9aecf867d65af2190d1ff92e992bfe758335dc0.jpg) +Fig. 1: Comparison of Temporal Fusion Methods. Illustrated are four key approaches: (1) Temporal Self-Attention [20], leveraging attention mechanisms for temporal integration; (2) Wrap and Concat [8,35,37], combining features across frames and fusing them through convolution; (3) Cost Volume Construction in image space [25], constructing cost volume from image input of different frames and leveraging plane-sweep volumes for depth map generation; and (4) Our Proposed Method, which involves constructing a temporal cost volume in 3D space to enhance feature refinement. In the figure, $\odot$ and $\otimes$ represent coordinate alignment and element-wise product, accordingly. + +and re-calibration. An alternative and more promising approach is the employment of multi-view temporal fusion, which can utilize the extended multi-view baselines available over time to enhance 3D perception tasks. + +Recent advancements in visual 3D object detection have shown the potential of incorporating temporal observations to bolster detection performance [8, 20, 25, 35, 37]. In our research, we have identified and categorized the emerging temporal fusion methods into three paradigms, as depicted in Figure 1. + +The first two paradigms are categorized as warp-based methods. These methodologies primarily involve the alignment of Bird's Eye View (BEV) feature maps across multiple temporal instances, utilizing relative camera poses obtained from Inertial Measurement Units (IMUs). Subsequently, the aligned features undergo integration using two primary approaches. The first approach employs self-attention mechanisms, as depicted in Figure 1(a). A representative method utilizing this approach is BEVFormer [20]. Alternatively, the second approach integrates aligned features through concatenation, as illustrated in Figure 1(b). Representative methods employing this approach include BEVDet4D [8], BEVFormerv2 [37], and PanoOcc [35]. These methods emphasize feature fusion using concatenation followed by convolutional operations. Despite their potential, these methods primarily leverage temporal information in an implicit manner, lacking a robust understanding of temporal geometry. As a result, they fall short in fully exploiting the geometric constraints inherent in 3D space. + +In contrast, cost volume-based methods, exemplified by (c) in Figure 1 and represented by SOLOFusion [25], draw inspiration from stereo matching techniques. Constructing a cost volume from images captured from different viewpoints in the temporal sequence enables the utilization of geometric constraints to obtain depth-aware features. However, for multi-view vision tasks, the sheer + +number of image pairs can lead to significant computational overhead when fusing longer time span temporal information. + +Recognizing the limitations of these approaches, we propose a novel paradigm, as illustrated in (d) in Figure 1. In this paper we introduce CVT-Occ, an innovative temporal fusion method designed to capitalize on the geometric correspondences of voxels over time, thereby enhancing occupancy prediction accuracy. Our approach involves sampling points along the line of sight for each voxel - defined as the line connecting the voxel to the optical center of the camera - and identifying the corresponding 3D locations of these points in historical frames based on relative camera poses. We then sample the features of these points in historical frames and integrate them with the current voxel features to construct a cost volume feature map. This map is subsequently utilized to refine the features of the current volume, thereby improving occupancy prediction. + +Compared with paradigms (a) and (b) in Figure 1, our proposed CVT-Occ explicitly utilizes parallax cues to refine the depth of 3D voxel. In contrast to paradigm (c), our approach avoids per-image-pair cost volume computation, achieving superior performance with minimal additional computational overhead. CVT-Occ distinguishes itself by leveraging the inherent parallax information in historical observations and employing a data-driven approach to learning the cost volume. Our experiments on the Occ3D-Waymo dataset demonstrate that CVT-Occ achieves state-of-the-art 3D occupancy prediction performance. + +# 2 Related Work + +# 2.1 3D Occupancy Prediction + +The goal of 3D occupancy prediction is to estimate the occupancy of each quantized voxel in 3D space. This task is originally rooted in Occupancy Grid Mapping (OGM) [24, 28] for mobile robot navigation, where the robot is equipped with range sensors (e.g., LiDAR) and navigates in static environments. Recent works transit to a more generic scenario: they make the occupancy prediction with visual systems in dynamic environments. MonoScene [1] reconstructs the 3D scene from sparse prediction with LiDAR points. VoxFormer [16] makes dense voxel predictions leveraging semantic labels and depth estimation from monocular RGB images. TPVFormer [11] proposes a tri-perspective view decomposition method for efficient 3D occupancy prediction. OccFormer [40] presents a dual-path transformer network to effectively process the 3D volume for semantic occupancy prediction. Besides, these advancements have led to new benchmarks such as visual 3D occupancy prediction [29, 30, 32] and semantic scene completion [15]. We emphasize that our proposed method is a simple and plug-and-play module that seamlessly integrates with existing occupancy prediction pipelines, significantly improving their performance. + +# 2.2 Temporal Fusion for 3D Perception + +Leveraging temporal information for occupancy prediction is a natural strategy since it provides sufficient spatial information for building geometric repre + +sentations [25, 36]. A common practice of temporal fusion in visual 3D object detection is BEV feature alignment. It fuses temporal information by warping past BEV features to the current time according to camera view transformation from relative camera poses at different time steps [8, 20]. It has been shown that short-term frames can significantly help improve the 3D detection performance. However, we find that minor improvements are brought up when more frames come in with simply warping and fusion. Recent work UniFusion [27] proposes a new long temporal fusion method. It creates "virtual views" for the temporal features as if they are present in the current time and directly access all useful history features. However, this method suffers from intensive computation costs when the number of frames increases. SOLOFusion [25] is a more efficient and effective model that balances the resolution and time stamps, and transforms the temporal fusion problem into a temporal multi-view stereo problem. + +# 2.3 Stereo Matching and Multi-View Stereo + +Stereo matching is rooted in constructing 3D cost volumes from 2D image features [3, 33, 39] and predicting the depth map from the cost volume. Recent works further advance it by proposing correlation-based cost volumes like GC-Net [2]. Meanwhile, multi-view stereo methods (5,6,10,12,31,38)) leverage plane-sweep volumes for depth map generation. A recent trend of using stereo matching [4, 17, 19] focused on 3D perception in autonomous driving. While existing work in stereo matching often utilizes stereo image pairs, this format is unsuitable for handling multi-view and multi-temporal inputs in autonomous driving scenarios. Additionally, these approaches require generating a Plane-Sweep Volume for each image pair, leading to inefficiencies. Furthermore, recent advancements like OccDepth [23] demonstrate the potential of implicitly learning correlations between stereo images to improve 3D depth-aware feature fusion. Our proposed Temporal Cost Volume method, designed specifically for multi-view and multi-temporal data, offers a more effective and efficient solution. The core innovation of the CVT module lies in its ability to construct a comprehensive 3D representation of the scene by aggregating information over time. + +# 3 Methodology + +# 3.1 Problem Setup + +Given only RGB images as inputs, the model is designed to predict a dense semantic scene within a specified volume. Specifically, we utilize as input the current timestamp $t$ along with previous images, denoted by $\mathbf{I} = \{\mathbf{I}_t^i,\mathbf{I}_{t - 1}^i,\ldots \}_{i = 1}^L$ , where $L$ represents the number of cameras. Our output is defined as a voxel grid $Y_{t}\in \{c_{0},c_{1},\dots ,c_{M}\}^{H\times W\times Z}$ , situated in the coordinate frame of the ego-vehicle at timestamp $t$ . Each voxel within this grid is either empty (indicated by $c_{0}$ ) or occupied by a specific semantic class among $\{c_1,c_2,\dots ,c_M\}$ . Here, $M$ denotes the total number of classes of interest, while $H$ , $W$ , and $Z$ represent the length, + +![](images/c5eb4489ef60b37b601397dd601a292a3cf98d2ba454d1138f9781ddb533bace.jpg) +Fig. 2: Overall Architecture of CVT-Occ. The image backbone extracts multiscale features from multi-view images, which are transformed into 3D volume features denoted as $\mathbf{V} \in \mathbb{R}^{H \times W \times Z \times C}$ . The Cost Volume Temporal Module samples points along the line of sight within the current volume and projects them onto $K - 1$ historical frames, resulting in $K \times N$ 3D volume features. These features are concatenated to construct cost volume features $\mathbf{F} \in \mathbb{R}^{H \times W \times Z \times (K \times N) \times C}$ . Convolution layers are then applied to generate weights $\mathbf{W} \in \mathbb{R}^{H \times W \times Z}$ , refining the depth of 3D voxel. Finally, an occupancy decoder produces 3D semantic occupancy predictions. In the figure, $(\text{\textcircled{A}})$ , $(\text{\textcircled{C}})$ , and $\otimes$ represent coordinate alignment, concatenation, and element-wise product, respectively. + +width, and height of the grid, respectively. The primary goal is to train a neural network $\Theta(\cdot)$ to produce a semantic voxel grid $Y_{t} = \Theta(\mathbf{I})$ that approximates the ground truth $\hat{Y}_{t}$ . It is important to note that the model must learn both 3D geometry and semantics solely from vision-based inputs, without the aid of depth measurements from LiDAR. Inferring 3D occupancy from 2D images presents a significant challenge and necessitates the design of efficient methods to leverage geometric constraints for learning accurate geometry. + +# 3.2 Overall Architecture + +In this section, we depict the comprehensive architecture of CVT-Occ, depicted in Figure 2. Our framework processes multi-frame, multi-view images to first extract multi-scale features through an image backbone. Subsequently, these features from image space are transformed into BEV space features, which are refined by a BEV encoder to generate 3D volumetric representations. There is a rich corpus of research focused on the transition from image space to BEV features. One line of works follows the lifting paradigm proposed in LSS [26]; + +they explicitly predict a depth map and lift multi-view image features onto the BEV plane [8, 9, 17, 18, 25]. Another line of works inherits the spirit of querying from 3D to 2D in DETR3D [34]; they employ learnable queries to extract information from image features by cross-attention mechanism [13, 20, 21]. It is crucial to underscore that our proposed Cost Volume Temporal (CVT) module is versatile and compatible across different strategies for feature transformation from image to volume space. For our experimental validation, we employ BEV-Former [20] as the foundation to generate 3D volume features. These features are subsequently refined and enhanced through the CVT module, demonstrating the efficacy and adaptability of our approach in leveraging temporal and spatial dynamics for enriched 3D scene comprehension. Finally an occupancy decoder is used to generate final voxel prediction results. + +# 3.3 Cost Volume Temporal Module + +Since direct depth information is not available for each pixel, transforming image features into 3D space introduces ambiguity; for instance, a single pixel may correspond to multiple voxels along the line of sight. To address this challenge, we propose the Cost Volume Temporal module, which leverages temporal data to infer depth information and resolve this ambiguity. Specifically, we construct 3D cost volume features using both historical and current BEV features. These cost volume features are then used to derive learned weights, which are subsequently employed to refine the current BEV features. + +3D Volume Features. We pre-define the volume $\mathbf{V} \in \mathbb{R}^{H \times W \times Z}$ , where $H, W, Z$ represent the spatial dimensions of the volume space. Each voxel in the volume space corresponds to a cubic space in the real world with a size of $s$ meters. By default, the center of the volume corresponds to the position of the ego car. For each voxel $v = (i, j, k)$ , its corresponding 3D location $p = (x, y, z)$ can be calculated by Eqn. 1: + +$$ +x = \left(i - \frac {W}{2}\right) \times s, y = \left(j - \frac {H}{2}\right) \times s, z = \left(k - \frac {Z}{2}\right) \times s (1) +$$ + +The BEV feature is $\mathbf{F}_t^{\mathrm{bev}}\in \mathbb{R}^{H\times W\times E}$ , where $t$ represents the timestamp and $E$ is the embedding dimensions. We reshape the BEV feature to volume space as the input of the CVT module, producing $\mathbf{V}_t\in \mathbb{R}^{H\times W\times Z\times C}$ . + +Construct Cost Volume Feature. Given the ambiguity along the line of sight, for each point $\mathbf{p_t}$ , we sample several additional points $\{\mathbf{p_t^i}\}_{i=1}^N$ within the current volume. Specifically, we calculate the sight direction $\mathbf{d_t} \in \mathbb{R}^3$ , which is the vector from the center of the volume to the point $\mathbf{p_t}$ , and then sample points using specific strides $\{n^i\}_{i=1}^N$ as shown in Eqn. 2: + +$$ +\mathbf {p} _ {\mathbf {t}} ^ {\mathbf {i}} = \mathbf {p} _ {\mathbf {t}} + \mathbf {d} _ {\mathbf {t}} \times \mathbf {n} ^ {\mathbf {i}} \tag {2} +$$ + +Since these points project to the same pixel in the image space, points along the same line tend to have similar features. To accurately distinguish the correct + +position corresponding to the pixel, we use historical BEV features to obtain complementary information. Here comes the core insight of our proposed Cost Volume Temporal Module. Projecting these points into the historical coordinate frame ensures they are no longer in the same line of sight. This parallax provides additional information from the historical BEV features, which helps to reduce depth ambiguity in the current frame. Projection matrix $\mathbf{P}_{\mathbf{t}} \in \mathbb{R}^{4 \times 4}$ can transform points from the coordinate frame of ego-vehicle to the global coordinate frame. Therefore the points $\{\mathbf{p}_{\mathbf{t}}^i\}_{i=1}^N$ are projected onto $K - 1$ historical frames through the projection matrix as described in Eqn. 3: + +$$ +\mathbf {p} _ {\mathbf {t} - \mathbf {k}} ^ {\mathrm {i}} = \mathbf {p} _ {\mathbf {t}} ^ {\mathrm {i}} \mathbf {P} _ {\mathbf {t}} \mathbf {P} _ {\mathbf {t} - \mathbf {k}} ^ {- 1} \tag {3} +$$ + +Finally, each point $\mathbf{p}_{\mathbf{t} - \mathbf{k}}^{\mathbf{i}}$ is converted to voxel coordinates, which is the inverse process of Eqn. 1, and bilinear interpolation is used to sample features from corresponding BEV feature maps. The final cost volume feature map is $\mathbf{F} \in \mathbb{R}^{H \times W \times Z \times (K \times N) \times C}$ , where $C$ represents the channels of the BEV features. + +# 3.4 Volume Features Refinement + +$$ +\mathbf {V} _ {\mathrm {o c c}} = \mathbf {V} _ {t} \otimes \operatorname {S i g m o i d} (\operatorname {C o n v} (\mathbf {F})) \tag {4} +$$ + +As described in the Eqn. 4, the constructed cost volume features $\mathbf{F}$ are processed through a sequence of convolutional layers, yielding an output weight $\mathbf{W} \in \mathbb{R}^{H \times W \times Z}$ . A Sigmoid activation function then normalizes the output weights $\mathbf{W}$ to the range [0,1]. These weights are directly supervised according to the voxel occupancy state: weights corresponding to occupied voxels are encouraged to reach 1, while those for unoccupied voxels are guided towards 0. The symbol $\otimes$ in the Eqn. 4 denotes element-wise product of the original volume features at timestamp $t$ with the learned weights $\mathbf{W}$ . This supervised learning approach produces an occupancy-aware volume feature map $\mathbf{V}_{\mathrm{occ}} \in \mathbb{R}^{H \times W \times Z \times C}$ . The goal of the learned weights is to diminish the influence of voxel features in incorrectly activated regions due to depth ambiguity and augment the features of correctly identified voxels simultaneously. + +# 3.5 Occupancy Decoder + +Upon the refinement of voxel features $\mathbf{V}_{\mathrm{occ}}$ , our model employs a series of deconvolution layers to transform it into occupancy features. The occupancy features are projected onto the output space, yielding $\mathbf{X} \in \mathbb{R}^{H \times W \times Z \times M}$ . The projection is formulated to map the occupancy features into a discrete set of semantic class predictions for each voxel within the grid. Consequently, the final occupancy output $\mathbf{Y} \in \{c_0, c_1, \ldots, c_M\}^{H \times W \times Z}$ is generated, where each voxel is assigned one of $M$ semantic labels, including an unoccupied state represented by $c_0$ . This step effectively translates the enriched voxel features into a semantically segmented 3D occupancy map. + +# 3.6 Training Loss + +Our architecture integrates the CVT module and 3D occupancy prediction within an end-to-end trainable framework. We train the overall network with a multitask loss as + +$$ +\operatorname {L o s s} = \mathcal {L} _ {\text {o c c}} + \lambda \mathcal {L} _ {\text {c v t}} \tag {5} +$$ + +where $\mathcal{L}_{\mathrm{occ}}$ denotes the occupancy prediction loss, and $\mathcal{L}_{\mathrm{cvt}}$ represents the cost volume temporal loss. The coefficient $\lambda$ serves as a balancing factor between these two loss components. + +For the loss of occupancy prediction $\mathcal{L}_{\mathrm{occ}}$ , we adopt the cross-entropy loss, which is widely used in semantic occupancy prediction task. $\mathcal{L}_{\mathrm{occ}} = \sum_{c}w_{c}\mathcal{L}(g_{c},p_{c})$ where $g_{c}$ and $p_c$ represent the ground truth label and the prediction result for the $c$ -th semantic class, respectively. Class-specific weighting $w_{c}$ mitigates class imbalance by inversely correlating with class frequency, as adopted from Occ3D [29]. + +To enhance the network's performance, the CVT module receives direct supervision. The loss attributed to the CVT module, $\mathcal{L}_{\mathrm{cvt}}$ , employs a binary cross-entropy framework as shown in Eqn. 6: + +$$ +\mathcal {L} _ {\mathrm {c v t}} = - \sum_ {j = 1} ^ {J} \hat {y} _ {j, 0} \log \left(1 - w _ {j}\right) + \hat {y} _ {j, 1} \log w _ {j} \tag {6} +$$ + +where $j$ indexes voxels in the grid, with $J = H \times W \times Z$ . $w_{k}$ is the $j$ -th element in the output weight $\mathbf{W} \in \mathbb{R}^{H \times W \times Z}$ . Specifically, a binary label $\hat{y}_{j,0} = 1$ signifies the voxel's unoccupied status (i.e., its semantic classification is $c_{0}$ ), while $\hat{y}_{j,1} = 1$ indicates the voxel is occupied in ground truth. This direct supervision on the CVT module not only improves performance but also underpins the importance of accurately learning temporal and spatial features from the input data. + +# 4 Experiment + +# 4.1 Dataset + +Occ3D-Waymo. We conduct experiments on the Occ3D-Waymo [29] dataset, which provides dense semantic labels for 3D occupancy grids. Occ3D-Waymo consists of 798 training scenes and 202 validation scenes. On the Occ3D-Waymo dataset, the range for the $x$ and $y$ axis as $[-40m, 40m]$ , and for the $z$ axis as $[-1m, 5.4m]$ . The voxel grid size is $(0.4m, 0.4m, 0.4m)$ , resulting in a resolution of $(200 \times 200 \times 16)$ for $(H, W, Z)$ . While the semantic labels encompass 16 categories, including "Free" class. Besides, it also provides visibility masks for LiDAR and camera modality. + +Evaluation Metrics. In our assessment of 3D semantic occupancy prediction capabilities, we utilize the IoU of each class to evaluate 3D semantic occupancy prediction. As Occ3D-Waymo provides camera visibility masks, only voxels within the camera's visible region are evaluated. The mean Intersection over Union (mIoU) is computed over 14 non-Free classes, excluding the "Motorcycle" class due to its insufficient voxel numbers. + +# 4.2 Experimental Settings + +Architecture Detail. Our architecture leverages the same image backbone and image neck as Occ3D [29] for image feature extraction. The transformation of image features to BEV space involves a four-layered view transformation layers, each layer comprising a sequence of cross-attention layers, normalization layers, feed-forward layers, and normalization layers. This BEV encoder generates a BEV feature representation $\mathbf{F}^{\mathrm{bev}}\in \mathbb{R}^{200\times 200\times 256}$ , which is further reshaped to volume features $\mathbf{V}\in \mathbb{R}^{200\times 200\times 16\times 16}$ . The selected frame configuration includes a time interval of $0.5s$ and a sequence of 7 frames, covering a total time span of $3s$ , which ensures a comprehensive temporal analysis. For the CVT module, by default, it samples 9 points within each volume feature across 7 frames to capture a broad spectrum of temporal and spatial variations. Consequently, this configuration enables the CVT module to generate a cost volume feature map $\mathbf{F}\in \mathbb{R}^{200\times 200\times 16\times (7\times 9)\times 16}$ . The occupancy decoder utilizes a series of deconvolution layers to transform the volume features into occupancy features, which is then followed by two additional convolution blocks responsible for generating the final 3D semantic occupancy predictions. + +Training. During training, we utilized the AdamW [22] optimizer for 8 epochs and full dataset each epoch. Training was conducted on 8 NVIDIA A100 GPUs with a batch size of 1 per GPU. Learning Rate Scheduler Policy is the cosine annealing schedule with an initial learning rate of $4 \times 10^{-4}$ . The dimensions for the input images were standardized to $960 \times 640$ pixels. It is noteworthy that we abstained from employing any data augmentation techniques in order to ensure a fair and unbiased comparison among the models. + +# 4.3 Main Results + +As shown in Figure 1, We categorize previous temporal fusion methods into three types: (a) Attention-based methods, which adopt temporal self-attention to recurrently fuse the history BEV information; (b) WrapConcat-based methods, involving the alignment of BEV feature maps across different time steps; and (c) Plane-Sweep-based methods, which construct a cost volume for image pairs to utilize geometric constraints. Additionally, we introduce a novel approach, CVT-Occ, that utilizes geometric correspondences of 3D voxels over time to improve occupancy prediction accuracy. For comparative analysis, we extended some mainstream BEV models for the 3D semantic occupancy prediction task. We choose BEVFormer [20], BEVFormer-WarpConcat [37], and SOLOFusion [25] as representative methods for categories (a), (b), and (c), respectively. + +For both BEVFormer [20] and SOLOFusion [25], we have replaced their original transformer decoders for object detection with a 3D occupancy prediction decoder. Additionally, for fairness, we omit the depth supervision module when reproducing SOLOFusion [25]. For BEVFormer-WrapConcat, we adopt a commonly used fusion method known as the Wrap and Concatenate strategy, as seen in BEVFormerv2 [37] and PanoOcc [35]. This strategy involves 3D coordinate alignment in BEV space based on the transformation matrix between historical + +Table 1: 3D Occupancy Prediction Results on Occ3D-Waymo. This comparison showcases the performance of various principal temporal fusion strategies alongside CVT-Occ, within the same architectural framework and identical frame specifications (number and interval), ensuring a direct and fair comparison across all evaluated methods. The leading performance is highlighted in bold. + +
MethodmIoUGoVehiclePedestrianSignBicyclistTraffic LightPoleCons. ConeBicycleBuildingVegetationTree TrunkRoadWalkable
BEVFormer-w/o TSA23.877.5034.5421.079.6920.9611.4811.4814.0614.5123.1421.828.5778.4556.89
BEVFormer [20]24.587.1836.0621.009.7620.2312.6114.5214.7016.0623.9822.509.3979.1157.04
SOLOFusion [25]24.734.9732.4518.2810.3317.148.0717.8316.2319.331.4928.9816.9370.9553.28
BEVFormer-WrapConcat25.076.236.1720.959.5620.5812.8216.2414.3116.7825.1423.5612.8179.0456.83
CVT-Occ(ours)27.377.4441.023.9311.9220.8112.0718.0316.8821.3729.427.4214.6779.1259.09
+ +and current frames. Subsequently, it concatenates previous BEV features with current BEV features along the channel dimension and applies residual blocks for temporal fusion. We incorporate the residual block proposed by PanoOcc [35] to create the BEVFormer-WrapConcat method. In the "BEVFormer-w/o TSA" configuration, we maintain the original architecture of BEVFormer [20] but exclude the Temporal Self-Attention(TSA) layers, thereby eliminating temporal attention mechanisms. In Table 1, we present the evaluation results for 3D occupancy prediction on the Occ3D-Waymo validation set. To ensure a fair comparison, all methods were implemented using the same network architecture and training strategy. + +Comparing "BEVFormer-w/o TSA" with temporal fusion methods in Table 1, All temporal fusion methods demonstrate improved mIoU compared to "BEVFormer-w/o TSA". It becomes evident that temporal fusion plays a crucial role in the 3D occupancy prediction task. + +Maintaining consistent frame specifications (number and interval), we explore the temporal fusion ability of different mainstream temporal fusion methods. Our method achieves state-of-the-art performance in terms of IoU and demonstrates significant superiority over other approaches across nearly all classes. As shown in Table 1, CVT-Occ exhibits superior performance compared to representative models of existing temporal fusion methods (a), (b), and (c) in Figure 1, achieving a notable $2.8\%$ IoU improvement over the baseline BEVFormer [20]. Delving into the details, CVT-Occ shows improvements in IoU across almost all classes. Specifically, for "Vehicle", "Bicycle", "Building", "Vegetation", and "Tree Trunk", our method achieves more than a $4\%$ increase in IoU compared to the baseline model BEVFormer [20]. This substantial enhancement across nearly all categories underscores that a comprehensive understanding of temporal geometry can lead to significant performance gains. + +# 4.4 Ablation Study + +Time Span. The time span is defined as the difference between the timestamp of the earliest historical frame and that of the current frame, which plays a + +
ExpFrameIntervalCVT supervisionmIoU
(a)30.1s✓25.15
(b)30.3s✓25.86
(c)70.1s✓25.80
(d)70.3s✓26.46
(e)70.5s✓27.37
(f)70.5s26.81
+ +Table 2: Ablation Experiments. We evaluate CVT-Occ under different frame specifications and CVT supervision. + +![](images/89af4c91687cb496213c5e43b741749ba08159e4cc5d847e2eab456beb651ab3.jpg) +Fig. 3: Experiments of Different Time Spans. + +A crucial role in the temporal fusion analysis. We conduct experiments to evaluate our method across various configurations of total frame numbers and sampled frame intervals, as summarized in Table 2. + +In experiments (a) and (c), we compared the impact of different total frame numbers. The results demonstrate that a longer sequence of historical frames leads to improved performance. Similarly, we observed that a larger time interval between adjacent frames also contributes to enhanced performance in experiments (c), (d), and (e). These consistent findings underscore the significance of a longer time span across the entire temporal sequence for better model performance. CVT-Occ relies on disparities between frames to mitigate depth ambiguity, where larger intervals and frame numbers contribute to an extended time span, amplifying the temporal parallax. This elucidates the observed increase in mIoU with a longer time span. + +In addition, comparing experiments (b) and (c), where the total time span remains the same 0.6s, highlights the critical importance of the total time span even when the frame interval and number of historical frames varies between these two experiments. This finding further emphasizes that the impact of our proposed method arises from leveraging the parallax of the entire temporal sequence. + +Ability of Long Temporal Fusion. Figure 3 illustrates the comparison between BEVFormer [20] and CVT-Occ across varying total time differences. CVT-Occ demonstrates improved performance with increasing temporal queue lengths, whereas BEVFormer struggles to effectively leverage information from long-range historical data. + +The limitation of BEVFormer's long temporal fusion capability has been highlighted in previous researches, such as UniFusion [27] and VideoBEV [7]. Based on results of Figure 3, we propose that BEVFormer fails to benefit from longer time spans due to its recurrent temporal fusion process. This process restricts the direct access to information from distant historical frames, regardless + +Table 3: Comparison Results under Different Conditions. Here are three sets of experiments. (a) Binary Classification: this task focuses on geometric occupancy prediction, treating all object classes as a single "Non-Free" class. (b) BEV Range: Along x-axis, [0,20] means the range $[-20m,0m]\cup [0,20m]$ and [20,40] means the range $[-40m, - 20m]\cup [20, - 40m]$ . (c) Ego Vehicle Speed: The 202 scenes for evaluation is partitioned into 2 parts based on the ego-vehicle speed. We evaluate all temporal fusion methods on them separately. + +
ExpmIoUNon-FreeFree[0,20][20,40]FastSlow
BEVFormer [20]24.5855.7486.5427.0921.2424.7324.22
BEVFormer-WrapConcat25.0756.5786.9628.2121.1626.2724.44
CVT-Occ27.3762.9590.4630.8522.6727.926.79
+ +of the temporal sequence length. In contrast, the cost volume feature $\mathbf{F}$ can directly integrate information from all historical frames. The CVT module utilizes parallax correlation explicitly to refine volume features, leading to a more robust understanding of temporal information. This design overcomes the shortcomings observed in BEVFormer [20], resulting in enhanced long-term temporal fusion capability. + +CVT Supervision. In our pursuit to improve 3D semantic occupancy prediction, we go beyond traditional loss functions by adding extra supervision to the CVT module output. Experiments (e) and (f) in Table 2 show the performance comparison of CVT-Occ with and without the CVT loss. Incorporating the CVT loss results in a significant increase in mIoU, demonstrating its effectiveness through experimental results. This enhancement validates our design approach, highlighting its ability to enhance the accuracy and reliability of semantic occupancy predictions in 3D. + +In conclusion, our ablation study confirms the critical importance of longer time spans and effective temporal fusion in enhancing CVT-Occ's performance for 3D semantic occupancy prediction. Additionally, integrating supplementary supervision further improves model accuracy, highlighting our approach's effectiveness in advancing temporal parallax-based scene understanding. + +# 4.5 Analysis of Factors + +The core insight of our method lies in leveraging parallax between different frames to enhance the accuracy of 3D voxel depth queries. In this subsection, we conduct several evaluations under various conditions to validate this concept. These experiments include binary classification, different BEV ranges, and varying ego vehicle speeds. + +Binary Classification. Within the framework of 3D semantic occupancy prediction, we aggregate all 15 other object classes into a single "Non-Free" class. This adjustment removes the influence of semantic information, allowing us to focus purely on depth accuracy. We then compare our approach against leading temporal fusion methods. The results in Table 3 demonstrate that our method + +provides better predictions on object geometry. This outcome aligns closely with our initial design concept of leveraging parallax for refined depth prediction. + +BEV Range. In the Occ3D-Waymo dataset, the BEV range spans from $-40m$ to $40m$ along the x-axis (i.e., the ego car's direction of movement). We divide this range into two parts: $[-20m, 20m]$ and $[-40m, 20m] \cup [20m, 40m]$ . Table 3 exhibits the evaluation results. All methods perform better in the near region, which is expected since objects closer to the vehicle are more likely to appear in the overlap region of history and current frames. However, our method exhibits significantly greater improvement in the near region (+3.76) compared to the far region (+1.43) compared to BEVFormer [20]. The larger parallax of nearer objects between different frames contributes to our method's superior prediction capability. + +Ego Vehicle Speed. Moreover, the speed of the vehicle also plays a crucial role in determining the magnitude of parallax. We further divide the dataset into two roughly equal parts based on the speed of ego vehicle. Evaluating CVT-Occ separately on these parts yields the results shown in Table 3. Regardless of whether BEVFormer [20] or CVT-Occ, both perform better in fast-moving scenarios. Our method demonstrates much greater improvement under the fast condition $(+3.17)$ than slow one $(+2.57)$ compared to the baseline BEVFormer. This can be attributed to the larger parallax between history and current frames when the ego car is moving rapidly. + +To sum up, these results underscore the significantly enhanced performance of CVT-Occ compared to the baseline model BEVFormer [20] under a wide range of conditions. The evaluations demonstrate that our method can refine the depth estimation of 3D voxel queries more accurately, with a notable advantage in the near region and fast scenes due to the greater parallax effect. + +# 4.6 Visualization Results + +Figure 4 offers a qualitative comparison of CVT-Occ against the baseline method BEVFormer [20] on the Occ3D-Waymo dataset. To ensure clarity in visualization, a dataset-specific mask is applied to exclude regions not within the camera's field of view. These visual representations clearly illustrate the superiority of CVT-Occ over competing approaches. BEVFormer encounters difficulties with depth perception, leading to issues such as atmospheric distortion and blurring of objects. CVT-Occ marks a remarkable improvements in the accurate depiction of buildings and vegetation, thus providing a more precise and reliable interpretation of the environment. + +# 5 Discussion and Conclusion + +We have presented CVT-Occ, a novel approach that significantly enhances the accuracy of 3D occupancy predictions by leveraging temporal fusion and geometric correspondence across time. Unlike traditional methods that rely on + +![](images/318bdeb9e37f1482b6c0ac38c29f2b6386a83f20123b249808ad5983bc3710a3.jpg) +Fig. 4: Qualitative Results. CVT-Occ exhibits superior performance in predicting the occupancy of vegetation and buildings. + +monocular or stereo vision, our method utilizes multi-view temporal fusion, effectively incorporating historical observations to exploit the parallax effect and improve depth estimation. Our innovative approach, which constructs a cost volume feature map by sampling and integrating features across temporal frames, demonstrates superior performance on the Occ3D-Waymo dataset, surpassing current state-of-the-art models in terms of accuracy while maintaining a low computational overhead. CVT-Occ also opens avenues for future research on temporal fusion for 3D perception, e.g., applying the cost volume module in other tasks like 3D reconstruction, and other applications like robotics and virtual reality. + +Acknowledgments. This work is supported by the National Key R&D Program of China (2022ZD0161700). + +# References + +1. Cao, A.Q., de Charette, R.: Monoscene: Monocular 3d semantic scene completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3991-4001 (June 2022) +2. Cao, Y., Xu, J., Lin, S., Wei, F., Hu, H.: Gcnet: Non-local networks meet squeeze-excitation networks and beyond. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops (Oct 2019) +3. Chang, J.R., Chen, Y.S.: Pyramid stereo matching network. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018) +4. Chen, Y., Liu, S., Shen, X., Jia, J.: Dsgn: Deep stereo geometry network for 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020) +5. Cheng, S., Xu, Z., Zhu, S., Li, Z., Li, L.E., Ramamoorthi, R., Su, H.: Deep stereo using adaptive thin volume representation with uncertainty awareness. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020) +6. Gu, X., Fan, Z., Zhu, S., Dai, Z., Tan, F., Tan, P.: Cascade cost volume for high-resolution multi-view stereo and stereo matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020) +7. Han, C., Yang, J., Sun, J., Ge, Z., Dong, R., Zhou, H., Mao, W., Peng, Y., Zhang, X.: Exploring recurrent long-term temporal fusion for multi-view 3d perception. IEEE Robotics and Automation Letters 9(7), 6544-6551 (2024) +8. Huang, J., Huang, G.: Bevdet4d: Exploit temporal cues in multi-camera 3d object detection (2022) +9. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view (2022) +10. Huang, P.H., Matzen, K., Kopf, J., Ahuja, N., Huang, J.B.: Deepmvs: Learning multi-view stereopsis. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018) +1. Huang, Y., Zheng, W., Zhang, Y., Zhou, J., Lu, J.: Tri-perspective view for vision-based 3d semantic occupancy prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9223-9232 (June 2023) +2. Im, S., Jeon, H.G., Lin, S., Kweon, I.S.: DPSNet: End-to-end deep plane sweep stereo. In: International Conference on Learning Representations (2019) +3. Jiang, Y., Zhang, L., Miao, Z., Zhu, X., Gao, J., Hu, W., Jiang, Y.G.: Polarformer: Multi-camera 3d object detection with polar transformer. Proceedings of the AAAI Conference on Artificial Intelligence 37(1), 1042-1050 (Jun 2023) +4. Li, H., Li, Z., Akmandor, N.U., Jiang, H., Wang, Y., Padir, T.: Stereovoxelnet: Real-time obstacle detection based on occupancy voxels from a stereo camera using deep neural networks. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 4826-4833 (2023) + +15. Li, Y., Li, S., Liu, X., Gong, M., Li, K., Chen, N., Wang, Z., Li, Z., Jiang, T., Yu, F., Wang, Y., Zhao, H., Yu, Z., Feng, C.: Sscbench: Monocular 3d semantic scene completion benchmark in street views (2023) +16. Li, Y., Yu, Z., Choy, C., Xiao, C., Alvarez, J.M., Fidler, S., Feng, C., Anandkumar, A.: Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9087-9098 (June 2023) +17. Li, Y., Bao, H., Ge, Z., Yang, J., Sun, J., Li, Z.: Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. Proceedings of the AAAI Conference on Artificial Intelligence 37(2), 1486-1494 (Jun 2023) +18. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. Proceedings of the AAAI Conference on Artificial Intelligence 37(2), 1477-1485 (Jun 2023) +19. Li, Y., Yang, J., Sun, J., Bao, H., Ge, Z., Xiao, L.: Bevstereo++: Accurate depth estimation in multi-view 3d object detection via dynamic temporal stereo (2023) +20. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 1-18 (2022) +21. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X.: Petrov2: A unified framework for 3d perception from multi-camera images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3262-3272 (October 2023) +22. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2019) +23. Miao, R., Liu, W., Chen, M., Gong, Z., Xu, W., Hu, C., Zhou, S.: Occdepth: A depth-aware method for 3d semantic scene completion (2023) +24. Moravec, H., Elfes, A.: High resolution maps from wide angle sonar. In: Proceedings. 1985 IEEE International Conference on Robotics and Automation. vol. 2, pp. 116-121 (1985) +25. Park, J., Xu, C., Yang, S., Keutzer, K., Kitani, K.M., Tomizuka, M., Zhan, W.: Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. In: The Eleventh International Conference on Learning Representations (2023) +26. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020. pp. 194-210. Springer International Publishing, Cham (2020) +27. Qin, Z., Chen, J., Chen, C., Chen, X., Li, X.: Unifusion: Unified multi-view fusion transformer for spatial-temporal representation in bird's-eye-view. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 8690-8699 (October 2023) +28. Thrun, S.: Probabilistic robotics. Communications of the ACM pp. 52-57 (Mar 2002) +29. Tian, X., Jiang, T., Yun, L., Mao, Y., Yang, H., Wang, Y., Wang, Y., Zhao, H.: Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 64318-64330. Curran Associates, Inc. (2023) + +30. Tong, W., Sima, C., Wang, T., Chen, L., Wu, S., Deng, H., Gu, Y., Lu, L., Luo, P., Lin, D., Li, H.: Scene as occupancy. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 8406-8415 (October 2023) +31. Wang, K., Shen, S.: Mvdepthnet: Real-time multiview depth estimation neural network. In: 2018 International Conference on 3D Vision (3DV) (Sep 2018) +32. Wang, X., Zhu, Z., Xu, W., Zhang, Y., Wei, Y., Chi, X., Ye, Y., Du, D., Lu, J., Wang, X.: Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 17850-17859 (October 2023) +33. Wang, Y., Lai, Z., Huang, G., Wang, B.H., van der Maaten, L., Campbell, M., Weinberger, K.Q.: Anytime stereo image depth estimation on mobile devices. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 5893-5900 (2019) +34. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Faust, A., Hsu, D., Neumann, G. (eds.) Proceedings of the 5th Conference on Robot Learning. Proceedings of Machine Learning Research, vol. 164, pp. 180-191. PMLR (08-11 Nov 2022) +35. Wang, Y., Chen, Y., Liao, X., Fan, L., Zhang, Z.: Panoocc: Unified occupancy representation for camera-based 3d panoptic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17158-17168 (June 2024) +36. Xu, C., Wu, B., Hou, J., Tsai, S., Li, R., Wang, J., Zhan, W., He, Z., Vajda, P., Keutzer, K., Tomizuka, M.: Nerf-det: Learning geometry-aware volumetric representation for multi-view 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 23320-23330 (October 2023) +37. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., Zhou, J., Dai, J.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17830-17839 (June 2023) +38. Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstructured multi-view stereo. In: Proceedings of the European Conference on Computer Vision (ECCV) (September 2018) +39. Zhang, F., Prisacariu, V., Yang, R., Torr, P.H.: Ga-net: Guided aggregation net for end-to-end stereo matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019) +40. Zhang, Y., Zhu, Z., Du, D.: Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9433-9443 (October 2023) \ No newline at end of file diff --git a/2024/CVT-Occ_ Cost Volume Temporal Fusion for 3D Occupancy Prediction/images.zip b/2024/CVT-Occ_ Cost Volume Temporal Fusion for 3D Occupancy Prediction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6fd48bc8094e9f1aabef042bbabac66672178931 --- /dev/null +++ b/2024/CVT-Occ_ Cost Volume Temporal Fusion for 3D Occupancy Prediction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac480b18a5580aad4f3b3eb023de6d58b7948f3ca428afc87413f63acee628aa +size 379953 diff --git a/2024/CVT-Occ_ Cost Volume Temporal Fusion for 3D Occupancy Prediction/layout.json b/2024/CVT-Occ_ Cost Volume Temporal Fusion for 3D Occupancy Prediction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..69854dceadde8658eab60df275969aebeb8bc008 --- /dev/null +++ b/2024/CVT-Occ_ Cost Volume Temporal Fusion for 3D Occupancy Prediction/layout.json @@ -0,0 +1,8445 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 164, + 111, + 451, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 111, + 451, + 148 + ], + "spans": [ + { + "bbox": [ + 164, + 111, + 451, + 148 + ], + "type": "text", + "content": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 167, + 447, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 167, + 447, + 194 + ], + "spans": [ + { + "bbox": [ + 165, + 167, + 447, + 194 + ], + "type": "text", + "content": "Zhangchen Ye" + }, + { + "bbox": [ + 165, + 167, + 447, + 194 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 165, + 167, + 447, + 194 + ], + "type": "text", + "content": ", Tao Jiang" + }, + { + "bbox": [ + 165, + 167, + 447, + 194 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 165, + 167, + 447, + 194 + ], + "type": "text", + "content": ", Chenfeng Xu" + }, + { + "bbox": [ + 165, + 167, + 447, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 165, + 167, + 447, + 194 + ], + "type": "text", + "content": ", Yiming Li" + }, + { + "bbox": [ + 165, + 167, + 447, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 165, + 167, + 447, + 194 + ], + "type": "text", + "content": ", and Hang Zhao" + }, + { + "bbox": [ + 165, + 167, + 447, + 194 + ], + "type": "inline_equation", + "content": "^{1,2,5\\boxtimes}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 173, + 201, + 441, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 201, + 441, + 225 + ], + "spans": [ + { + "bbox": [ + 173, + 201, + 441, + 225 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 173, + 201, + 441, + 225 + ], + "type": "text", + "content": "IIIS, Tsinghua University " + }, + { + "bbox": [ + 173, + 201, + 441, + 225 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 173, + 201, + 441, + 225 + ], + "type": "text", + "content": "Shanghai AI Lab " + }, + { + "bbox": [ + 173, + 201, + 441, + 225 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 173, + 201, + 441, + 225 + ], + "type": "text", + "content": "UC Berkeley " + }, + { + "bbox": [ + 173, + 201, + 441, + 225 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 173, + 201, + 441, + 225 + ], + "type": "text", + "content": "New York University " + }, + { + "bbox": [ + 173, + 201, + 441, + 225 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 173, + 201, + 441, + 225 + ], + "type": "text", + "content": "Shanghai Qi Zhi Institute" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 256, + 452, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 256, + 452, + 421 + ], + "spans": [ + { + "bbox": [ + 160, + 256, + 452, + 421 + ], + "type": "text", + "content": "Abstract. Vision-based 3D occupancy prediction is significantly challenged by the inherent limitations of monocular vision in depth estimation. This paper introduces CVT-Occ, a novel approach that leverages temporal fusion through the geometric correspondence of voxels over time to improve the accuracy of 3D occupancy predictions. By sampling points along the line of sight of each voxel and integrating the features of these points from historical frames, we construct a cost volume feature map that refines current volume features for improved prediction outcomes. Our method takes advantage of parallax cues from historical observations and employs a data-driven approach to learn the cost volume. We validate the effectiveness of CVT-Occ through rigorous experiments on the Occ3D-Waymo dataset, where it outperforms state-of-the-art methods in 3D occupancy prediction with minimal additional computational cost. The code is released at https://github.com/Tsinghua-MARS-Lab/CVT-Occ." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 431, + 436, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 431, + 436, + 443 + ], + "spans": [ + { + "bbox": [ + 160, + 431, + 436, + 443 + ], + "type": "text", + "content": "Keywords: 3D Semantic Occupancy Prediction " + }, + { + "bbox": [ + 160, + 431, + 436, + 443 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 431, + 436, + 443 + ], + "type": "text", + "content": " Temporal Fusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 464, + 230, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 230, + 477 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 230, + 477 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 491, + 481, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 481, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 481, + 550 + ], + "type": "text", + "content": "Vision-based 3D semantic occupancy prediction is rapidly evolving in the domain of 3D perception, driven by its critical applications in autonomous driving, robotics, and augmented reality. The task aims to estimate the occupancy state and semantic label of every voxel within a 3D space from visual inputs [11,29, 30,40]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 552, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 482, + 635 + ], + "type": "text", + "content": "Despite its critical importance, 3D occupancy prediction presents significant challenges. When relying solely on monocular vision, these challenges are particularly pronounced due to the inherent ambiguity in estimating depth from a single image. While stereo vision has been proposed as a solution to augment depth estimation accuracy [14], its application remains limited in practice. It is impractical to use stereo cameras for widespread applications in autonomous vehicles and robotic systems because of the requirement for extensive calibration" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 643, + 233, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 643, + 233, + 653 + ], + "spans": [ + { + "bbox": [ + 141, + 643, + 233, + 653 + ], + "type": "text", + "content": "*: Equal contribution." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 653, + 362, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 653, + 362, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 653, + 362, + 665 + ], + "type": "inline_equation", + "content": "\\boxtimes" + }, + { + "bbox": [ + 141, + 653, + 362, + 665 + ], + "type": "text", + "content": " : Corresponding to: hangzhao@mail.tsinghua.edu.cn" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 119, + 473, + 203 + ], + "blocks": [ + { + "bbox": [ + 143, + 119, + 473, + 203 + ], + "lines": [ + { + "bbox": [ + 143, + 119, + 473, + 203 + ], + "spans": [ + { + "bbox": [ + 143, + 119, + 473, + 203 + ], + "type": "image", + "image_path": "d6581b4708145846a5456243e9aecf867d65af2190d1ff92e992bfe758335dc0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 215, + 482, + 315 + ], + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 315 + ], + "type": "text", + "content": "Fig. 1: Comparison of Temporal Fusion Methods. Illustrated are four key approaches: (1) Temporal Self-Attention [20], leveraging attention mechanisms for temporal integration; (2) Wrap and Concat [8,35,37], combining features across frames and fusing them through convolution; (3) Cost Volume Construction in image space [25], constructing cost volume from image input of different frames and leveraging plane-sweep volumes for depth map generation; and (4) Our Proposed Method, which involves constructing a temporal cost volume in 3D space to enhance feature refinement. In the figure, " + }, + { + "bbox": [ + 130, + 215, + 482, + 315 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 215, + 482, + 315 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 215, + 482, + 315 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 130, + 215, + 482, + 315 + ], + "type": "text", + "content": " represent coordinate alignment and element-wise product, accordingly." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 345, + 480, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 480, + 380 + ], + "type": "text", + "content": "and re-calibration. An alternative and more promising approach is the employment of multi-view temporal fusion, which can utilize the extended multi-view baselines available over time to enhance 3D perception tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 384, + 482, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 433 + ], + "type": "text", + "content": "Recent advancements in visual 3D object detection have shown the potential of incorporating temporal observations to bolster detection performance [8, 20, 25, 35, 37]. In our research, we have identified and categorized the emerging temporal fusion methods into three paradigms, as depicted in Figure 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 434, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 602 + ], + "type": "text", + "content": "The first two paradigms are categorized as warp-based methods. These methodologies primarily involve the alignment of Bird's Eye View (BEV) feature maps across multiple temporal instances, utilizing relative camera poses obtained from Inertial Measurement Units (IMUs). Subsequently, the aligned features undergo integration using two primary approaches. The first approach employs self-attention mechanisms, as depicted in Figure 1(a). A representative method utilizing this approach is BEVFormer [20]. Alternatively, the second approach integrates aligned features through concatenation, as illustrated in Figure 1(b). Representative methods employing this approach include BEVDet4D [8], BEVFormerv2 [37], and PanoOcc [35]. These methods emphasize feature fusion using concatenation followed by convolutional operations. Despite their potential, these methods primarily leverage temporal information in an implicit manner, lacking a robust understanding of temporal geometry. As a result, they fall short in fully exploiting the geometric constraints inherent in 3D space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "In contrast, cost volume-based methods, exemplified by (c) in Figure 1 and represented by SOLOFusion [25], draw inspiration from stereo matching techniques. Constructing a cost volume from images captured from different viewpoints in the temporal sequence enables the utilization of geometric constraints to obtain depth-aware features. However, for multi-view vision tasks, the sheer" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Z. Ye et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "number of image pairs can lead to significant computational overhead when fusing longer time span temporal information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 272 + ], + "type": "text", + "content": "Recognizing the limitations of these approaches, we propose a novel paradigm, as illustrated in (d) in Figure 1. In this paper we introduce CVT-Occ, an innovative temporal fusion method designed to capitalize on the geometric correspondences of voxels over time, thereby enhancing occupancy prediction accuracy. Our approach involves sampling points along the line of sight for each voxel - defined as the line connecting the voxel to the optical center of the camera - and identifying the corresponding 3D locations of these points in historical frames based on relative camera poses. We then sample the features of these points in historical frames and integrate them with the current voxel features to construct a cost volume feature map. This map is subsequently utilized to refine the features of the current volume, thereby improving occupancy prediction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 272, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 368 + ], + "type": "text", + "content": "Compared with paradigms (a) and (b) in Figure 1, our proposed CVT-Occ explicitly utilizes parallax cues to refine the depth of 3D voxel. In contrast to paradigm (c), our approach avoids per-image-pair cost volume computation, achieving superior performance with minimal additional computational overhead. CVT-Occ distinguishes itself by leveraging the inherent parallax information in historical observations and employing a data-driven approach to learning the cost volume. Our experiments on the Occ3D-Waymo dataset demonstrate that CVT-Occ achieves state-of-the-art 3D occupancy prediction performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 382, + 237, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 382, + 237, + 395 + ], + "spans": [ + { + "bbox": [ + 131, + 382, + 237, + 395 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 403, + 290, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 403, + 290, + 416 + ], + "spans": [ + { + "bbox": [ + 131, + 403, + 290, + 416 + ], + "type": "text", + "content": "2.1 3D Occupancy Prediction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 482, + 612 + ], + "type": "text", + "content": "The goal of 3D occupancy prediction is to estimate the occupancy of each quantized voxel in 3D space. This task is originally rooted in Occupancy Grid Mapping (OGM) [24, 28] for mobile robot navigation, where the robot is equipped with range sensors (e.g., LiDAR) and navigates in static environments. Recent works transit to a more generic scenario: they make the occupancy prediction with visual systems in dynamic environments. MonoScene [1] reconstructs the 3D scene from sparse prediction with LiDAR points. VoxFormer [16] makes dense voxel predictions leveraging semantic labels and depth estimation from monocular RGB images. TPVFormer [11] proposes a tri-perspective view decomposition method for efficient 3D occupancy prediction. OccFormer [40] presents a dual-path transformer network to effectively process the 3D volume for semantic occupancy prediction. Besides, these advancements have led to new benchmarks such as visual 3D occupancy prediction [29, 30, 32] and semantic scene completion [15]. We emphasize that our proposed method is a simple and plug-and-play module that seamlessly integrates with existing occupancy prediction pipelines, significantly improving their performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 625, + 339, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 625, + 339, + 637 + ], + "spans": [ + { + "bbox": [ + 131, + 625, + 339, + 637 + ], + "type": "text", + "content": "2.2 Temporal Fusion for 3D Perception" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Leveraging temporal information for occupancy prediction is a natural strategy since it provides sufficient spatial information for building geometric repre" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": "sentations [25, 36]. A common practice of temporal fusion in visual 3D object detection is BEV feature alignment. It fuses temporal information by warping past BEV features to the current time according to camera view transformation from relative camera poses at different time steps [8, 20]. It has been shown that short-term frames can significantly help improve the 3D detection performance. However, we find that minor improvements are brought up when more frames come in with simply warping and fusion. Recent work UniFusion [27] proposes a new long temporal fusion method. It creates \"virtual views\" for the temporal features as if they are present in the current time and directly access all useful history features. However, this method suffers from intensive computation costs when the number of frames increases. SOLOFusion [25] is a more efficient and effective model that balances the resolution and time stamps, and transforms the temporal fusion problem into a temporal multi-view stereo problem." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 290, + 362, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 290, + 362, + 303 + ], + "spans": [ + { + "bbox": [ + 131, + 290, + 362, + 303 + ], + "type": "text", + "content": "2.3 Stereo Matching and Multi-View Stereo" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 312, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 502 + ], + "type": "text", + "content": "Stereo matching is rooted in constructing 3D cost volumes from 2D image features [3, 33, 39] and predicting the depth map from the cost volume. Recent works further advance it by proposing correlation-based cost volumes like GC-Net [2]. Meanwhile, multi-view stereo methods (5,6,10,12,31,38)) leverage plane-sweep volumes for depth map generation. A recent trend of using stereo matching [4, 17, 19] focused on 3D perception in autonomous driving. While existing work in stereo matching often utilizes stereo image pairs, this format is unsuitable for handling multi-view and multi-temporal inputs in autonomous driving scenarios. Additionally, these approaches require generating a Plane-Sweep Volume for each image pair, leading to inefficiencies. Furthermore, recent advancements like OccDepth [23] demonstrate the potential of implicitly learning correlations between stereo images to improve 3D depth-aware feature fusion. Our proposed Temporal Cost Volume method, designed specifically for multi-view and multi-temporal data, offers a more effective and efficient solution. The core innovation of the CVT module lies in its ability to construct a comprehensive 3D representation of the scene by aggregating information over time." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 522, + 233, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 522, + 233, + 537 + ], + "spans": [ + { + "bbox": [ + 131, + 522, + 233, + 537 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 548, + 236, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 236, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 236, + 562 + ], + "type": "text", + "content": "3.1 Problem Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Given only RGB images as inputs, the model is designed to predict a dense semantic scene within a specified volume. Specifically, we utilize as input the current timestamp " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " along with previous images, denoted by " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{I} = \\{\\mathbf{I}_t^i,\\mathbf{I}_{t - 1}^i,\\ldots \\}_{i = 1}^L" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " represents the number of cameras. Our output is defined as a voxel grid " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "Y_{t}\\in \\{c_{0},c_{1},\\dots ,c_{M}\\}^{H\\times W\\times Z}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", situated in the coordinate frame of the ego-vehicle at timestamp " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ". Each voxel within this grid is either empty (indicated by " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "c_{0}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ") or occupied by a specific semantic class among " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\{c_1,c_2,\\dots ,c_M\\}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " denotes the total number of classes of interest, while " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " represent the length," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Z. Ye et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 146, + 118, + 463, + 311 + ], + "blocks": [ + { + "bbox": [ + 146, + 118, + 463, + 311 + ], + "lines": [ + { + "bbox": [ + 146, + 118, + 463, + 311 + ], + "spans": [ + { + "bbox": [ + 146, + 118, + 463, + 311 + ], + "type": "image", + "image_path": "c5eb4489ef60b37b601397dd601a292a3cf98d2ba454d1138f9781ddb533bace.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 323, + 482, + 434 + ], + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 434 + ], + "type": "text", + "content": "Fig. 2: Overall Architecture of CVT-Occ. The image backbone extracts multiscale features from multi-view images, which are transformed into 3D volume features denoted as " + }, + { + "bbox": [ + 130, + 323, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in \\mathbb{R}^{H \\times W \\times Z \\times C}" + }, + { + "bbox": [ + 130, + 323, + 482, + 434 + ], + "type": "text", + "content": ". The Cost Volume Temporal Module samples points along the line of sight within the current volume and projects them onto " + }, + { + "bbox": [ + 130, + 323, + 482, + 434 + ], + "type": "inline_equation", + "content": "K - 1" + }, + { + "bbox": [ + 130, + 323, + 482, + 434 + ], + "type": "text", + "content": " historical frames, resulting in " + }, + { + "bbox": [ + 130, + 323, + 482, + 434 + ], + "type": "inline_equation", + "content": "K \\times N" + }, + { + "bbox": [ + 130, + 323, + 482, + 434 + ], + "type": "text", + "content": " 3D volume features. These features are concatenated to construct cost volume features " + }, + { + "bbox": [ + 130, + 323, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\in \\mathbb{R}^{H \\times W \\times Z \\times (K \\times N) \\times C}" + }, + { + "bbox": [ + 130, + 323, + 482, + 434 + ], + "type": "text", + "content": ". Convolution layers are then applied to generate weights " + }, + { + "bbox": [ + 130, + 323, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{H \\times W \\times Z}" + }, + { + "bbox": [ + 130, + 323, + 482, + 434 + ], + "type": "text", + "content": ", refining the depth of 3D voxel. Finally, an occupancy decoder produces 3D semantic occupancy predictions. In the figure, " + }, + { + "bbox": [ + 130, + 323, + 482, + 434 + ], + "type": "inline_equation", + "content": "(\\text{\\textcircled{A}})" + }, + { + "bbox": [ + 130, + 323, + 482, + 434 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 323, + 482, + 434 + ], + "type": "inline_equation", + "content": "(\\text{\\textcircled{C}})" + }, + { + "bbox": [ + 130, + 323, + 482, + 434 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 323, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 130, + 323, + 482, + 434 + ], + "type": "text", + "content": " represent coordinate alignment, concatenation, and element-wise product, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 458, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 482, + 542 + ], + "type": "text", + "content": "width, and height of the grid, respectively. The primary goal is to train a neural network " + }, + { + "bbox": [ + 130, + 458, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\Theta(\\cdot)" + }, + { + "bbox": [ + 130, + 458, + 482, + 542 + ], + "type": "text", + "content": " to produce a semantic voxel grid " + }, + { + "bbox": [ + 130, + 458, + 482, + 542 + ], + "type": "inline_equation", + "content": "Y_{t} = \\Theta(\\mathbf{I})" + }, + { + "bbox": [ + 130, + 458, + 482, + 542 + ], + "type": "text", + "content": " that approximates the ground truth " + }, + { + "bbox": [ + 130, + 458, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\hat{Y}_{t}" + }, + { + "bbox": [ + 130, + 458, + 482, + 542 + ], + "type": "text", + "content": ". It is important to note that the model must learn both 3D geometry and semantics solely from vision-based inputs, without the aid of depth measurements from LiDAR. Inferring 3D occupancy from 2D images presents a significant challenge and necessitates the design of efficient methods to leverage geometric constraints for learning accurate geometry." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 560, + 264, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 560, + 264, + 572 + ], + "spans": [ + { + "bbox": [ + 132, + 560, + 264, + 572 + ], + "type": "text", + "content": "3.2 Overall Architecture" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "In this section, we depict the comprehensive architecture of CVT-Occ, depicted in Figure 2. Our framework processes multi-frame, multi-view images to first extract multi-scale features through an image backbone. Subsequently, these features from image space are transformed into BEV space features, which are refined by a BEV encoder to generate 3D volumetric representations. There is a rich corpus of research focused on the transition from image space to BEV features. One line of works follows the lifting paradigm proposed in LSS [26];" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 90, + 447, + 102 + ], + "type": "text", + "content": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": "they explicitly predict a depth map and lift multi-view image features onto the BEV plane [8, 9, 17, 18, 25]. Another line of works inherits the spirit of querying from 3D to 2D in DETR3D [34]; they employ learnable queries to extract information from image features by cross-attention mechanism [13, 20, 21]. It is crucial to underscore that our proposed Cost Volume Temporal (CVT) module is versatile and compatible across different strategies for feature transformation from image to volume space. For our experimental validation, we employ BEV-Former [20] as the foundation to generate 3D volume features. These features are subsequently refined and enhanced through the CVT module, demonstrating the efficacy and adaptability of our approach in leveraging temporal and spatial dynamics for enriched 3D scene comprehension. Finally an occupancy decoder is used to generate final voxel prediction results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 278, + 320, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 320, + 291 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 320, + 291 + ], + "type": "text", + "content": "3.3 Cost Volume Temporal Module" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 300, + 482, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 395 + ], + "type": "text", + "content": "Since direct depth information is not available for each pixel, transforming image features into 3D space introduces ambiguity; for instance, a single pixel may correspond to multiple voxels along the line of sight. To address this challenge, we propose the Cost Volume Temporal module, which leverages temporal data to infer depth information and resolve this ambiguity. Specifically, we construct 3D cost volume features using both historical and current BEV features. These cost volume features are then used to derive learned weights, which are subsequently employed to refine the current BEV features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "content": "3D Volume Features. We pre-define the volume " + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in \\mathbb{R}^{H \\times W \\times Z}" + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "inline_equation", + "content": "H, W, Z" + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "content": " represent the spatial dimensions of the volume space. Each voxel in the volume space corresponds to a cubic space in the real world with a size of " + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "content": " meters. By default, the center of the volume corresponds to the position of the ego car. For each voxel " + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "inline_equation", + "content": "v = (i, j, k)" + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "content": ", its corresponding 3D location " + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "inline_equation", + "content": "p = (x, y, z)" + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "content": " can be calculated by Eqn. 1:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 174, + 476, + 481, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 476, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 174, + 476, + 481, + 504 + ], + "type": "interline_equation", + "content": "x = \\left(i - \\frac {W}{2}\\right) \\times s, y = \\left(j - \\frac {H}{2}\\right) \\times s, z = \\left(k - \\frac {Z}{2}\\right) \\times s (1)", + "image_path": "b0fc97d38dc10972f87257bdd36204520949b89c56e0c05a249455cb3cc69e62.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 512, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 482, + 548 + ], + "type": "text", + "content": "The BEV feature is " + }, + { + "bbox": [ + 130, + 512, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t^{\\mathrm{bev}}\\in \\mathbb{R}^{H\\times W\\times E}" + }, + { + "bbox": [ + 130, + 512, + 482, + 548 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 512, + 482, + 548 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 512, + 482, + 548 + ], + "type": "text", + "content": " represents the timestamp and " + }, + { + "bbox": [ + 130, + 512, + 482, + 548 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 512, + 482, + 548 + ], + "type": "text", + "content": " is the embedding dimensions. We reshape the BEV feature to volume space as the input of the CVT module, producing " + }, + { + "bbox": [ + 130, + 512, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_t\\in \\mathbb{R}^{H\\times W\\times Z\\times C}" + }, + { + "bbox": [ + 130, + 512, + 482, + 548 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "text", + "content": "Construct Cost Volume Feature. Given the ambiguity along the line of sight, for each point " + }, + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{p_t}" + }, + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "text", + "content": ", we sample several additional points " + }, + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p_t^i}\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "text", + "content": " within the current volume. Specifically, we calculate the sight direction " + }, + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{d_t} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "text", + "content": ", which is the vector from the center of the volume to the point " + }, + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{p_t}" + }, + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "text", + "content": ", and then sample points using specific strides " + }, + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\{n^i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 548, + 482, + 609 + ], + "type": "text", + "content": " as shown in Eqn. 2:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 268, + 618, + 481, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 618, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 268, + 618, + 481, + 633 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {\\mathbf {t}} ^ {\\mathbf {i}} = \\mathbf {p} _ {\\mathbf {t}} + \\mathbf {d} _ {\\mathbf {t}} \\times \\mathbf {n} ^ {\\mathbf {i}} \\tag {2}", + "image_path": "9a8f80f56393143e2c038f9cb5159dca0030af3b56a52d2bd34d44546a3cf6ed.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Since these points project to the same pixel in the image space, points along the same line tend to have similar features. To accurately distinguish the correct" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Z. Ye et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": "position corresponding to the pixel, we use historical BEV features to obtain complementary information. Here comes the core insight of our proposed Cost Volume Temporal Module. Projecting these points into the historical coordinate frame ensures they are no longer in the same line of sight. This parallax provides additional information from the historical BEV features, which helps to reduce depth ambiguity in the current frame. Projection matrix " + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathbf{t}} \\in \\mathbb{R}^{4 \\times 4}" + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": " can transform points from the coordinate frame of ego-vehicle to the global coordinate frame. Therefore the points " + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}_{\\mathbf{t}}^i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": " are projected onto " + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "inline_equation", + "content": "K - 1" + }, + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": " historical frames through the projection matrix as described in Eqn. 3:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 267, + 234, + 480, + 248 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 234, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 267, + 234, + 480, + 248 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {\\mathbf {t} - \\mathbf {k}} ^ {\\mathrm {i}} = \\mathbf {p} _ {\\mathbf {t}} ^ {\\mathrm {i}} \\mathbf {P} _ {\\mathbf {t}} \\mathbf {P} _ {\\mathbf {t} - \\mathbf {k}} ^ {- 1} \\tag {3}", + "image_path": "0ab861225d0d91898d83203bc21edb446de61194348d9799a2e5be5ddc51dafa.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 258, + 479, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 258, + 479, + 307 + ], + "spans": [ + { + "bbox": [ + 131, + 258, + 479, + 307 + ], + "type": "text", + "content": "Finally, each point " + }, + { + "bbox": [ + 131, + 258, + 479, + 307 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{\\mathbf{t} - \\mathbf{k}}^{\\mathbf{i}}" + }, + { + "bbox": [ + 131, + 258, + 479, + 307 + ], + "type": "text", + "content": " is converted to voxel coordinates, which is the inverse process of Eqn. 1, and bilinear interpolation is used to sample features from corresponding BEV feature maps. The final cost volume feature map is " + }, + { + "bbox": [ + 131, + 258, + 479, + 307 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\in \\mathbb{R}^{H \\times W \\times Z \\times (K \\times N) \\times C}" + }, + { + "bbox": [ + 131, + 258, + 479, + 307 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 258, + 479, + 307 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 131, + 258, + 479, + 307 + ], + "type": "text", + "content": " represents the channels of the BEV features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 327, + 306, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 327, + 306, + 338 + ], + "spans": [ + { + "bbox": [ + 131, + 327, + 306, + 338 + ], + "type": "text", + "content": "3.4 Volume Features Refinement" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 238, + 352, + 480, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 352, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 238, + 352, + 480, + 364 + ], + "type": "interline_equation", + "content": "\\mathbf {V} _ {\\mathrm {o c c}} = \\mathbf {V} _ {t} \\otimes \\operatorname {S i g m o i d} (\\operatorname {C o n v} (\\mathbf {F})) \\tag {4}", + "image_path": "9e517e6694c954427109e47467693bf3e87f8198503b9480598c59aea8873901.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 371, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 482, + 515 + ], + "type": "text", + "content": "As described in the Eqn. 4, the constructed cost volume features " + }, + { + "bbox": [ + 130, + 371, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 130, + 371, + 482, + 515 + ], + "type": "text", + "content": " are processed through a sequence of convolutional layers, yielding an output weight " + }, + { + "bbox": [ + 130, + 371, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{H \\times W \\times Z}" + }, + { + "bbox": [ + 130, + 371, + 482, + 515 + ], + "type": "text", + "content": ". A Sigmoid activation function then normalizes the output weights " + }, + { + "bbox": [ + 130, + 371, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 130, + 371, + 482, + 515 + ], + "type": "text", + "content": " to the range [0,1]. These weights are directly supervised according to the voxel occupancy state: weights corresponding to occupied voxels are encouraged to reach 1, while those for unoccupied voxels are guided towards 0. The symbol " + }, + { + "bbox": [ + 130, + 371, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 130, + 371, + 482, + 515 + ], + "type": "text", + "content": " in the Eqn. 4 denotes element-wise product of the original volume features at timestamp " + }, + { + "bbox": [ + 130, + 371, + 482, + 515 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 371, + 482, + 515 + ], + "type": "text", + "content": " with the learned weights " + }, + { + "bbox": [ + 130, + 371, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 130, + 371, + 482, + 515 + ], + "type": "text", + "content": ". This supervised learning approach produces an occupancy-aware volume feature map " + }, + { + "bbox": [ + 130, + 371, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{\\mathrm{occ}} \\in \\mathbb{R}^{H \\times W \\times Z \\times C}" + }, + { + "bbox": [ + 130, + 371, + 482, + 515 + ], + "type": "text", + "content": ". The goal of the learned weights is to diminish the influence of voxel features in incorrectly activated regions due to depth ambiguity and augment the features of correctly identified voxels simultaneously." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 535, + 261, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 535, + 261, + 547 + ], + "spans": [ + { + "bbox": [ + 131, + 535, + 261, + 547 + ], + "type": "text", + "content": "3.5 Occupancy Decoder" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Upon the refinement of voxel features " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{\\mathrm{occ}}" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": ", our model employs a series of deconvolution layers to transform it into occupancy features. The occupancy features are projected onto the output space, yielding " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{H \\times W \\times Z \\times M}" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": ". The projection is formulated to map the occupancy features into a discrete set of semantic class predictions for each voxel within the grid. Consequently, the final occupancy output " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} \\in \\{c_0, c_1, \\ldots, c_M\\}^{H \\times W \\times Z}" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " is generated, where each voxel is assigned one of " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " semantic labels, including an unoccupied state represented by " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "c_0" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": ". This step effectively translates the enriched voxel features into a semantically segmented 3D occupancy map." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 229, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 229, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 229, + 128 + ], + "type": "text", + "content": "3.6 Training Loss" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 480, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 480, + 167 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 480, + 167 + ], + "type": "text", + "content": "Our architecture integrates the CVT module and 3D occupancy prediction within an end-to-end trainable framework. We train the overall network with a multitask loss as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 263, + 170, + 480, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 170, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 263, + 170, + 480, + 182 + ], + "type": "interline_equation", + "content": "\\operatorname {L o s s} = \\mathcal {L} _ {\\text {o c c}} + \\lambda \\mathcal {L} _ {\\text {c v t}} \\tag {5}", + "image_path": "cc16f2b6242dc78cd0c6149ecb1373d7462524be2df32c021d168043a744542c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 184, + 479, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 184, + 479, + 220 + ], + "spans": [ + { + "bbox": [ + 130, + 184, + 479, + 220 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 184, + 479, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{occ}}" + }, + { + "bbox": [ + 130, + 184, + 479, + 220 + ], + "type": "text", + "content": " denotes the occupancy prediction loss, and " + }, + { + "bbox": [ + 130, + 184, + 479, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cvt}}" + }, + { + "bbox": [ + 130, + 184, + 479, + 220 + ], + "type": "text", + "content": " represents the cost volume temporal loss. The coefficient " + }, + { + "bbox": [ + 130, + 184, + 479, + 220 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 184, + 479, + 220 + ], + "type": "text", + "content": " serves as a balancing factor between these two loss components." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 220, + 488, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 488, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 488, + 291 + ], + "type": "text", + "content": "For the loss of occupancy prediction " + }, + { + "bbox": [ + 130, + 220, + 488, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{occ}}" + }, + { + "bbox": [ + 130, + 220, + 488, + 291 + ], + "type": "text", + "content": ", we adopt the cross-entropy loss, which is widely used in semantic occupancy prediction task. " + }, + { + "bbox": [ + 130, + 220, + 488, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{occ}} = \\sum_{c}w_{c}\\mathcal{L}(g_{c},p_{c})" + }, + { + "bbox": [ + 130, + 220, + 488, + 291 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 220, + 488, + 291 + ], + "type": "inline_equation", + "content": "g_{c}" + }, + { + "bbox": [ + 130, + 220, + 488, + 291 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 220, + 488, + 291 + ], + "type": "inline_equation", + "content": "p_c" + }, + { + "bbox": [ + 130, + 220, + 488, + 291 + ], + "type": "text", + "content": " represent the ground truth label and the prediction result for the " + }, + { + "bbox": [ + 130, + 220, + 488, + 291 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 220, + 488, + 291 + ], + "type": "text", + "content": "-th semantic class, respectively. Class-specific weighting " + }, + { + "bbox": [ + 130, + 220, + 488, + 291 + ], + "type": "inline_equation", + "content": "w_{c}" + }, + { + "bbox": [ + 130, + 220, + 488, + 291 + ], + "type": "text", + "content": " mitigates class imbalance by inversely correlating with class frequency, as adopted from Occ3D [29]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 292, + 480, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 480, + 328 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 480, + 328 + ], + "type": "text", + "content": "To enhance the network's performance, the CVT module receives direct supervision. The loss attributed to the CVT module, " + }, + { + "bbox": [ + 130, + 292, + 480, + 328 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cvt}}" + }, + { + "bbox": [ + 130, + 292, + 480, + 328 + ], + "type": "text", + "content": ", employs a binary cross-entropy framework as shown in Eqn. 6:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 218, + 334, + 480, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 334, + 480, + 366 + ], + "spans": [ + { + "bbox": [ + 218, + 334, + 480, + 366 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c v t}} = - \\sum_ {j = 1} ^ {J} \\hat {y} _ {j, 0} \\log \\left(1 - w _ {j}\\right) + \\hat {y} _ {j, 1} \\log w _ {j} \\tag {6}", + "image_path": "37bf4674cd265954cde90b1fe62e4196036daf1badcd5866a33235152bfa1c36.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 371, + 482, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 482, + 444 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 482, + 444 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 371, + 482, + 444 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 371, + 482, + 444 + ], + "type": "text", + "content": " indexes voxels in the grid, with " + }, + { + "bbox": [ + 130, + 371, + 482, + 444 + ], + "type": "inline_equation", + "content": "J = H \\times W \\times Z" + }, + { + "bbox": [ + 130, + 371, + 482, + 444 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 371, + 482, + 444 + ], + "type": "inline_equation", + "content": "w_{k}" + }, + { + "bbox": [ + 130, + 371, + 482, + 444 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 371, + 482, + 444 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 371, + 482, + 444 + ], + "type": "text", + "content": "-th element in the output weight " + }, + { + "bbox": [ + 130, + 371, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{H \\times W \\times Z}" + }, + { + "bbox": [ + 130, + 371, + 482, + 444 + ], + "type": "text", + "content": ". Specifically, a binary label " + }, + { + "bbox": [ + 130, + 371, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\hat{y}_{j,0} = 1" + }, + { + "bbox": [ + 130, + 371, + 482, + 444 + ], + "type": "text", + "content": " signifies the voxel's unoccupied status (i.e., its semantic classification is " + }, + { + "bbox": [ + 130, + 371, + 482, + 444 + ], + "type": "inline_equation", + "content": "c_{0}" + }, + { + "bbox": [ + 130, + 371, + 482, + 444 + ], + "type": "text", + "content": "), while " + }, + { + "bbox": [ + 130, + 371, + 482, + 444 + ], + "type": "inline_equation", + "content": "\\hat{y}_{j,1} = 1" + }, + { + "bbox": [ + 130, + 371, + 482, + 444 + ], + "type": "text", + "content": " indicates the voxel is occupied in ground truth. This direct supervision on the CVT module not only improves performance but also underpins the importance of accurately learning temporal and spatial features from the input data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 459, + 224, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 459, + 224, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 459, + 224, + 472 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 481, + 200, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 481, + 200, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 481, + 200, + 491 + ], + "type": "text", + "content": "4.1 Dataset" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 498, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 593 + ], + "type": "text", + "content": "Occ3D-Waymo. We conduct experiments on the Occ3D-Waymo [29] dataset, which provides dense semantic labels for 3D occupancy grids. Occ3D-Waymo consists of 798 training scenes and 202 validation scenes. On the Occ3D-Waymo dataset, the range for the " + }, + { + "bbox": [ + 130, + 498, + 482, + 593 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 498, + 482, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 498, + 482, + 593 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 498, + 482, + 593 + ], + "type": "text", + "content": " axis as " + }, + { + "bbox": [ + 130, + 498, + 482, + 593 + ], + "type": "inline_equation", + "content": "[-40m, 40m]" + }, + { + "bbox": [ + 130, + 498, + 482, + 593 + ], + "type": "text", + "content": ", and for the " + }, + { + "bbox": [ + 130, + 498, + 482, + 593 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 498, + 482, + 593 + ], + "type": "text", + "content": " axis as " + }, + { + "bbox": [ + 130, + 498, + 482, + 593 + ], + "type": "inline_equation", + "content": "[-1m, 5.4m]" + }, + { + "bbox": [ + 130, + 498, + 482, + 593 + ], + "type": "text", + "content": ". The voxel grid size is " + }, + { + "bbox": [ + 130, + 498, + 482, + 593 + ], + "type": "inline_equation", + "content": "(0.4m, 0.4m, 0.4m)" + }, + { + "bbox": [ + 130, + 498, + 482, + 593 + ], + "type": "text", + "content": ", resulting in a resolution of " + }, + { + "bbox": [ + 130, + 498, + 482, + 593 + ], + "type": "inline_equation", + "content": "(200 \\times 200 \\times 16)" + }, + { + "bbox": [ + 130, + 498, + 482, + 593 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 498, + 482, + 593 + ], + "type": "inline_equation", + "content": "(H, W, Z)" + }, + { + "bbox": [ + 130, + 498, + 482, + 593 + ], + "type": "text", + "content": ". While the semantic labels encompass 16 categories, including \"Free\" class. Besides, it also provides visibility masks for LiDAR and camera modality." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Evaluation Metrics. In our assessment of 3D semantic occupancy prediction capabilities, we utilize the IoU of each class to evaluate 3D semantic occupancy prediction. As Occ3D-Waymo provides camera visibility masks, only voxels within the camera's visible region are evaluated. The mean Intersection over Union (mIoU) is computed over 14 non-Free classes, excluding the \"Motorcycle\" class due to its insufficient voxel numbers." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Z. Ye et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 272, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 272, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 272, + 129 + ], + "type": "text", + "content": "4.2 Experimental Settings" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 327 + ], + "type": "text", + "content": "Architecture Detail. Our architecture leverages the same image backbone and image neck as Occ3D [29] for image feature extraction. The transformation of image features to BEV space involves a four-layered view transformation layers, each layer comprising a sequence of cross-attention layers, normalization layers, feed-forward layers, and normalization layers. This BEV encoder generates a BEV feature representation " + }, + { + "bbox": [ + 130, + 137, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{\\mathrm{bev}}\\in \\mathbb{R}^{200\\times 200\\times 256}" + }, + { + "bbox": [ + 130, + 137, + 482, + 327 + ], + "type": "text", + "content": ", which is further reshaped to volume features " + }, + { + "bbox": [ + 130, + 137, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{V}\\in \\mathbb{R}^{200\\times 200\\times 16\\times 16}" + }, + { + "bbox": [ + 130, + 137, + 482, + 327 + ], + "type": "text", + "content": ". The selected frame configuration includes a time interval of " + }, + { + "bbox": [ + 130, + 137, + 482, + 327 + ], + "type": "inline_equation", + "content": "0.5s" + }, + { + "bbox": [ + 130, + 137, + 482, + 327 + ], + "type": "text", + "content": " and a sequence of 7 frames, covering a total time span of " + }, + { + "bbox": [ + 130, + 137, + 482, + 327 + ], + "type": "inline_equation", + "content": "3s" + }, + { + "bbox": [ + 130, + 137, + 482, + 327 + ], + "type": "text", + "content": ", which ensures a comprehensive temporal analysis. For the CVT module, by default, it samples 9 points within each volume feature across 7 frames to capture a broad spectrum of temporal and spatial variations. Consequently, this configuration enables the CVT module to generate a cost volume feature map " + }, + { + "bbox": [ + 130, + 137, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{F}\\in \\mathbb{R}^{200\\times 200\\times 16\\times (7\\times 9)\\times 16}" + }, + { + "bbox": [ + 130, + 137, + 482, + 327 + ], + "type": "text", + "content": ". The occupancy decoder utilizes a series of deconvolution layers to transform the volume features into occupancy features, which is then followed by two additional convolution blocks responsible for generating the final 3D semantic occupancy predictions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 327, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 411 + ], + "type": "text", + "content": "Training. During training, we utilized the AdamW [22] optimizer for 8 epochs and full dataset each epoch. Training was conducted on 8 NVIDIA A100 GPUs with a batch size of 1 per GPU. Learning Rate Scheduler Policy is the cosine annealing schedule with an initial learning rate of " + }, + { + "bbox": [ + 130, + 327, + 482, + 411 + ], + "type": "inline_equation", + "content": "4 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 327, + 482, + 411 + ], + "type": "text", + "content": ". The dimensions for the input images were standardized to " + }, + { + "bbox": [ + 130, + 327, + 482, + 411 + ], + "type": "inline_equation", + "content": "960 \\times 640" + }, + { + "bbox": [ + 130, + 327, + 482, + 411 + ], + "type": "text", + "content": " pixels. It is noteworthy that we abstained from employing any data augmentation techniques in order to ensure a fair and unbiased comparison among the models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 430, + 228, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 228, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 228, + 441 + ], + "type": "text", + "content": "4.3 Main Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 450, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 582 + ], + "type": "text", + "content": "As shown in Figure 1, We categorize previous temporal fusion methods into three types: (a) Attention-based methods, which adopt temporal self-attention to recurrently fuse the history BEV information; (b) WrapConcat-based methods, involving the alignment of BEV feature maps across different time steps; and (c) Plane-Sweep-based methods, which construct a cost volume for image pairs to utilize geometric constraints. Additionally, we introduce a novel approach, CVT-Occ, that utilizes geometric correspondences of 3D voxels over time to improve occupancy prediction accuracy. For comparative analysis, we extended some mainstream BEV models for the 3D semantic occupancy prediction task. We choose BEVFormer [20], BEVFormer-WarpConcat [37], and SOLOFusion [25] as representative methods for categories (a), (b), and (c), respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "For both BEVFormer [20] and SOLOFusion [25], we have replaced their original transformer decoders for object detection with a 3D occupancy prediction decoder. Additionally, for fairness, we omit the depth supervision module when reproducing SOLOFusion [25]. For BEVFormer-WrapConcat, we adopt a commonly used fusion method known as the Wrap and Concatenate strategy, as seen in BEVFormerv2 [37] and PanoOcc [35]. This strategy involves 3D coordinate alignment in BEV space based on the transformation matrix between historical" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 179, + 481, + 262 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 170 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 170 + ], + "type": "text", + "content": "Table 1: 3D Occupancy Prediction Results on Occ3D-Waymo. This comparison showcases the performance of various principal temporal fusion strategies alongside CVT-Occ, within the same architectural framework and identical frame specifications (number and interval), ensuring a direct and fair comparison across all evaluated methods. The leading performance is highlighted in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 179, + 481, + 262 + ], + "lines": [ + { + "bbox": [ + 133, + 179, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 133, + 179, + 481, + 262 + ], + "type": "table", + "html": "
MethodmIoUGoVehiclePedestrianSignBicyclistTraffic LightPoleCons. ConeBicycleBuildingVegetationTree TrunkRoadWalkable
BEVFormer-w/o TSA23.877.5034.5421.079.6920.9611.4811.4814.0614.5123.1421.828.5778.4556.89
BEVFormer [20]24.587.1836.0621.009.7620.2312.6114.5214.7016.0623.9822.509.3979.1157.04
SOLOFusion [25]24.734.9732.4518.2810.3317.148.0717.8316.2319.331.4928.9816.9370.9553.28
BEVFormer-WrapConcat25.076.236.1720.959.5620.5812.8216.2414.3116.7825.1423.5612.8179.0456.83
CVT-Occ(ours)27.377.4441.023.9311.9220.8112.0718.0316.8821.3729.427.4214.6779.1259.09
", + "image_path": "89270818402e32a0ff00390c0a68e4a64093c34a0fc980f7cfe0bf0d73e4e9f4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 283, + 481, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 283, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 283, + 481, + 403 + ], + "type": "text", + "content": "and current frames. Subsequently, it concatenates previous BEV features with current BEV features along the channel dimension and applies residual blocks for temporal fusion. We incorporate the residual block proposed by PanoOcc [35] to create the BEVFormer-WrapConcat method. In the \"BEVFormer-w/o TSA\" configuration, we maintain the original architecture of BEVFormer [20] but exclude the Temporal Self-Attention(TSA) layers, thereby eliminating temporal attention mechanisms. In Table 1, we present the evaluation results for 3D occupancy prediction on the Occ3D-Waymo validation set. To ensure a fair comparison, all methods were implemented using the same network architecture and training strategy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 403, + 481, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 450 + ], + "type": "text", + "content": "Comparing \"BEVFormer-w/o TSA\" with temporal fusion methods in Table 1, All temporal fusion methods demonstrate improved mIoU compared to \"BEVFormer-w/o TSA\". It becomes evident that temporal fusion plays a crucial role in the 3D occupancy prediction task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 451, + 481, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 451, + 481, + 606 + ], + "spans": [ + { + "bbox": [ + 132, + 451, + 481, + 606 + ], + "type": "text", + "content": "Maintaining consistent frame specifications (number and interval), we explore the temporal fusion ability of different mainstream temporal fusion methods. Our method achieves state-of-the-art performance in terms of IoU and demonstrates significant superiority over other approaches across nearly all classes. As shown in Table 1, CVT-Occ exhibits superior performance compared to representative models of existing temporal fusion methods (a), (b), and (c) in Figure 1, achieving a notable " + }, + { + "bbox": [ + 132, + 451, + 481, + 606 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 132, + 451, + 481, + 606 + ], + "type": "text", + "content": " IoU improvement over the baseline BEVFormer [20]. Delving into the details, CVT-Occ shows improvements in IoU across almost all classes. Specifically, for \"Vehicle\", \"Bicycle\", \"Building\", \"Vegetation\", and \"Tree Trunk\", our method achieves more than a " + }, + { + "bbox": [ + 132, + 451, + 481, + 606 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 132, + 451, + 481, + 606 + ], + "type": "text", + "content": " increase in IoU compared to the baseline model BEVFormer [20]. This substantial enhancement across nearly all categories underscores that a comprehensive understanding of temporal geometry can lead to significant performance gains." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 622, + 237, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 237, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 237, + 635 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "content": "Time Span. The time span is defined as the difference between the timestamp of the earliest historical frame and that of the current frame, which plays a" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Z. Ye et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 114, + 294, + 235 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 294, + 235 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 294, + 235 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 294, + 235 + ], + "type": "table", + "html": "
ExpFrameIntervalCVT supervisionmIoU
(a)30.1s✓25.15
(b)30.3s✓25.86
(c)70.1s✓25.80
(d)70.3s✓26.46
(e)70.5s✓27.37
(f)70.5s26.81
", + "image_path": "09bb5546cf7c14383681a19b72d1756114845b22c81cc3979e8b29af04634eec.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 242, + 291, + 285 + ], + "lines": [ + { + "bbox": [ + 132, + 242, + 291, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 242, + 291, + 285 + ], + "type": "text", + "content": "Table 2: Ablation Experiments. We evaluate CVT-Occ under different frame specifications and CVT supervision." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 310, + 119, + 480, + 251 + ], + "blocks": [ + { + "bbox": [ + 310, + 119, + 480, + 251 + ], + "lines": [ + { + "bbox": [ + 310, + 119, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 310, + 119, + 480, + 251 + ], + "type": "image", + "image_path": "89af4c91687cb496213c5e43b741749ba08159e4cc5d847e2eab456beb651ab3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 264, + 482, + 288 + ], + "lines": [ + { + "bbox": [ + 304, + 264, + 482, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 482, + 288 + ], + "type": "text", + "content": "Fig. 3: Experiments of Different Time Spans." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 313, + 479, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 479, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 479, + 348 + ], + "type": "text", + "content": "A crucial role in the temporal fusion analysis. We conduct experiments to evaluate our method across various configurations of total frame numbers and sampled frame intervals, as summarized in Table 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 350, + 482, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 471 + ], + "type": "text", + "content": "In experiments (a) and (c), we compared the impact of different total frame numbers. The results demonstrate that a longer sequence of historical frames leads to improved performance. Similarly, we observed that a larger time interval between adjacent frames also contributes to enhanced performance in experiments (c), (d), and (e). These consistent findings underscore the significance of a longer time span across the entire temporal sequence for better model performance. CVT-Occ relies on disparities between frames to mitigate depth ambiguity, where larger intervals and frame numbers contribute to an extended time span, amplifying the temporal parallax. This elucidates the observed increase in mIoU with a longer time span." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 544 + ], + "type": "text", + "content": "In addition, comparing experiments (b) and (c), where the total time span remains the same 0.6s, highlights the critical importance of the total time span even when the frame interval and number of historical frames varies between these two experiments. This finding further emphasizes that the impact of our proposed method arises from leveraging the parallax of the entire temporal sequence." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 544, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 604 + ], + "type": "text", + "content": "Ability of Long Temporal Fusion. Figure 3 illustrates the comparison between BEVFormer [20] and CVT-Occ across varying total time differences. CVT-Occ demonstrates improved performance with increasing temporal queue lengths, whereas BEVFormer struggles to effectively leverage information from long-range historical data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "The limitation of BEVFormer's long temporal fusion capability has been highlighted in previous researches, such as UniFusion [27] and VideoBEV [7]. Based on results of Figure 3, we propose that BEVFormer fails to benefit from longer time spans due to its recurrent temporal fusion process. This process restricts the direct access to information from distant historical frames, regardless" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 90, + 447, + 102 + ], + "type": "text", + "content": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 201, + 479, + 258 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "text", + "content": "Table 3: Comparison Results under Different Conditions. Here are three sets of experiments. (a) Binary Classification: this task focuses on geometric occupancy prediction, treating all object classes as a single \"Non-Free\" class. (b) BEV Range: Along x-axis, [0,20] means the range " + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "inline_equation", + "content": "[-20m,0m]\\cup [0,20m]" + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "text", + "content": " and [20,40] means the range " + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "inline_equation", + "content": "[-40m, - 20m]\\cup [20, - 40m]" + }, + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "text", + "content": ". (c) Ego Vehicle Speed: The 202 scenes for evaluation is partitioned into 2 parts based on the ego-vehicle speed. We evaluate all temporal fusion methods on them separately." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 201, + 479, + 258 + ], + "lines": [ + { + "bbox": [ + 135, + 201, + 479, + 258 + ], + "spans": [ + { + "bbox": [ + 135, + 201, + 479, + 258 + ], + "type": "table", + "html": "
ExpmIoUNon-FreeFree[0,20][20,40]FastSlow
BEVFormer [20]24.5855.7486.5427.0921.2424.7324.22
BEVFormer-WrapConcat25.0756.5786.9628.2121.1626.2724.44
CVT-Occ27.3762.9590.4630.8522.6727.926.79
", + "image_path": "21b01a6d4a503c4f3777692428849bfe28ac1108c5d7816ed26e60ebd47a9c08.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 281, + 479, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 479, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 479, + 352 + ], + "type": "text", + "content": "of the temporal sequence length. In contrast, the cost volume feature " + }, + { + "bbox": [ + 130, + 281, + 479, + 352 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 130, + 281, + 479, + 352 + ], + "type": "text", + "content": " can directly integrate information from all historical frames. The CVT module utilizes parallax correlation explicitly to refine volume features, leading to a more robust understanding of temporal information. This design overcomes the shortcomings observed in BEVFormer [20], resulting in enhanced long-term temporal fusion capability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 352, + 480, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 480, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 480, + 448 + ], + "type": "text", + "content": "CVT Supervision. In our pursuit to improve 3D semantic occupancy prediction, we go beyond traditional loss functions by adding extra supervision to the CVT module output. Experiments (e) and (f) in Table 2 show the performance comparison of CVT-Occ with and without the CVT loss. Incorporating the CVT loss results in a significant increase in mIoU, demonstrating its effectiveness through experimental results. This enhancement validates our design approach, highlighting its ability to enhance the accuracy and reliability of semantic occupancy predictions in 3D." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 448, + 481, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 481, + 509 + ], + "type": "text", + "content": "In conclusion, our ablation study confirms the critical importance of longer time spans and effective temporal fusion in enhancing CVT-Occ's performance for 3D semantic occupancy prediction. Additionally, integrating supplementary supervision further improves model accuracy, highlighting our approach's effectiveness in advancing temporal parallax-based scene understanding." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 525, + 257, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 525, + 257, + 538 + ], + "spans": [ + { + "bbox": [ + 131, + 525, + 257, + 538 + ], + "type": "text", + "content": "4.5 Analysis of Factors" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 605 + ], + "type": "text", + "content": "The core insight of our method lies in leveraging parallax between different frames to enhance the accuracy of 3D voxel depth queries. In this subsection, we conduct several evaluations under various conditions to validate this concept. These experiments include binary classification, different BEV ranges, and varying ego vehicle speeds." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Binary Classification. Within the framework of 3D semantic occupancy prediction, we aggregate all 15 other object classes into a single \"Non-Free\" class. This adjustment removes the influence of semantic information, allowing us to focus purely on depth accuracy. We then compare our approach against leading temporal fusion methods. The results in Table 3 demonstrate that our method" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Z. Ye et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "provides better predictions on object geometry. This outcome aligns closely with our initial design concept of leveraging parallax for refined depth prediction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 260 + ], + "type": "text", + "content": "BEV Range. In the Occ3D-Waymo dataset, the BEV range spans from " + }, + { + "bbox": [ + 130, + 140, + 481, + 260 + ], + "type": "inline_equation", + "content": "-40m" + }, + { + "bbox": [ + 130, + 140, + 481, + 260 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 140, + 481, + 260 + ], + "type": "inline_equation", + "content": "40m" + }, + { + "bbox": [ + 130, + 140, + 481, + 260 + ], + "type": "text", + "content": " along the x-axis (i.e., the ego car's direction of movement). We divide this range into two parts: " + }, + { + "bbox": [ + 130, + 140, + 481, + 260 + ], + "type": "inline_equation", + "content": "[-20m, 20m]" + }, + { + "bbox": [ + 130, + 140, + 481, + 260 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 140, + 481, + 260 + ], + "type": "inline_equation", + "content": "[-40m, 20m] \\cup [20m, 40m]" + }, + { + "bbox": [ + 130, + 140, + 481, + 260 + ], + "type": "text", + "content": ". Table 3 exhibits the evaluation results. All methods perform better in the near region, which is expected since objects closer to the vehicle are more likely to appear in the overlap region of history and current frames. However, our method exhibits significantly greater improvement in the near region (+3.76) compared to the far region (+1.43) compared to BEVFormer [20]. The larger parallax of nearer objects between different frames contributes to our method's superior prediction capability." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 261, + 481, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 481, + 369 + ], + "type": "text", + "content": "Ego Vehicle Speed. Moreover, the speed of the vehicle also plays a crucial role in determining the magnitude of parallax. We further divide the dataset into two roughly equal parts based on the speed of ego vehicle. Evaluating CVT-Occ separately on these parts yields the results shown in Table 3. Regardless of whether BEVFormer [20] or CVT-Occ, both perform better in fast-moving scenarios. Our method demonstrates much greater improvement under the fast condition " + }, + { + "bbox": [ + 130, + 261, + 481, + 369 + ], + "type": "inline_equation", + "content": "(+3.17)" + }, + { + "bbox": [ + 130, + 261, + 481, + 369 + ], + "type": "text", + "content": " than slow one " + }, + { + "bbox": [ + 130, + 261, + 481, + 369 + ], + "type": "inline_equation", + "content": "(+2.57)" + }, + { + "bbox": [ + 130, + 261, + 481, + 369 + ], + "type": "text", + "content": " compared to the baseline BEVFormer. This can be attributed to the larger parallax between history and current frames when the ego car is moving rapidly." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 369, + 481, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 481, + 429 + ], + "type": "text", + "content": "To sum up, these results underscore the significantly enhanced performance of CVT-Occ compared to the baseline model BEVFormer [20] under a wide range of conditions. The evaluations demonstrate that our method can refine the depth estimation of 3D voxel queries more accurately, with a notable advantage in the near region and fast scenes due to the greater parallax effect." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 449, + 266, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 449, + 266, + 460 + ], + "spans": [ + { + "bbox": [ + 131, + 449, + 266, + 460 + ], + "type": "text", + "content": "4.6 Visualization Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 472, + 481, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 481, + 579 + ], + "type": "text", + "content": "Figure 4 offers a qualitative comparison of CVT-Occ against the baseline method BEVFormer [20] on the Occ3D-Waymo dataset. To ensure clarity in visualization, a dataset-specific mask is applied to exclude regions not within the camera's field of view. These visual representations clearly illustrate the superiority of CVT-Occ over competing approaches. BEVFormer encounters difficulties with depth perception, leading to issues such as atmospheric distortion and blurring of objects. CVT-Occ marks a remarkable improvements in the accurate depiction of buildings and vegetation, thus providing a more precise and reliable interpretation of the environment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 601, + 311, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 601, + 311, + 613 + ], + "spans": [ + { + "bbox": [ + 131, + 601, + 311, + 613 + ], + "type": "text", + "content": "5 Discussion and Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "We have presented CVT-Occ, a novel approach that significantly enhances the accuracy of 3D occupancy predictions by leveraging temporal fusion and geometric correspondence across time. Unlike traditional methods that rely on" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 90, + 447, + 102 + ], + "type": "text", + "content": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 155, + 115, + 462, + 475 + ], + "blocks": [ + { + "bbox": [ + 155, + 115, + 462, + 475 + ], + "lines": [ + { + "bbox": [ + 155, + 115, + 462, + 475 + ], + "spans": [ + { + "bbox": [ + 155, + 115, + 462, + 475 + ], + "type": "image", + "image_path": "318bdeb9e37f1482b6c0ac38c29f2b6386a83f20123b249808ad5983bc3710a3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 497, + 482, + 520 + ], + "lines": [ + { + "bbox": [ + 132, + 497, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 497, + 482, + 520 + ], + "type": "text", + "content": "Fig. 4: Qualitative Results. CVT-Occ exhibits superior performance in predicting the occupancy of vegetation and buildings." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "monocular or stereo vision, our method utilizes multi-view temporal fusion, effectively incorporating historical observations to exploit the parallax effect and improve depth estimation. Our innovative approach, which constructs a cost volume feature map by sampling and integrating features across temporal frames, demonstrates superior performance on the Occ3D-Waymo dataset, surpassing current state-of-the-art models in terms of accuracy while maintaining a low computational overhead. CVT-Occ also opens avenues for future research on temporal fusion for 3D perception, e.g., applying the cost volume module in other tasks like 3D reconstruction, and other applications like robotics and virtual reality." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Z. Ye et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "Acknowledgments. This work is supported by the National Key R&D Program of China (2022ZD0161700)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 197, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 197, + 173 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 197, + 173 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 188, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 188, + 480, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 188, + 480, + 220 + ], + "spans": [ + { + "bbox": [ + 138, + 188, + 480, + 220 + ], + "type": "text", + "content": "1. Cao, A.Q., de Charette, R.: Monoscene: Monocular 3d semantic scene completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3991-4001 (June 2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 222, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 222, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 222, + 480, + 255 + ], + "type": "text", + "content": "2. Cao, Y., Xu, J., Lin, S., Wei, F., Hu, H.: Gcnet: Non-local networks meet squeeze-excitation networks and beyond. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops (Oct 2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 255, + 480, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 255, + 480, + 287 + ], + "spans": [ + { + "bbox": [ + 138, + 255, + 480, + 287 + ], + "type": "text", + "content": "3. Chang, J.R., Chen, Y.S.: Pyramid stereo matching network. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 289, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 480, + 320 + ], + "type": "text", + "content": "4. Chen, Y., Liu, S., Shen, X., Jia, J.: Dsgn: Deep stereo geometry network for 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 322, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 322, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 322, + 480, + 365 + ], + "type": "text", + "content": "5. Cheng, S., Xu, Z., Zhu, S., Li, Z., Li, L.E., Ramamoorthi, R., Su, H.: Deep stereo using adaptive thin volume representation with uncertainty awareness. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 366, + 480, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 366, + 480, + 409 + ], + "spans": [ + { + "bbox": [ + 138, + 366, + 480, + 409 + ], + "type": "text", + "content": "6. Gu, X., Fan, Z., Zhu, S., Dai, Z., Tan, F., Tan, P.: Cascade cost volume for high-resolution multi-view stereo and stereo matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 411, + 480, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 480, + 442 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 480, + 442 + ], + "type": "text", + "content": "7. Han, C., Yang, J., Sun, J., Ge, Z., Dong, R., Zhou, H., Mao, W., Peng, Y., Zhang, X.: Exploring recurrent long-term temporal fusion for multi-view 3d perception. IEEE Robotics and Automation Letters 9(7), 6544-6551 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 444, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 444, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 444, + 480, + 464 + ], + "type": "text", + "content": "8. Huang, J., Huang, G.: Bevdet4d: Exploit temporal cues in multi-camera 3d object detection (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 466, + 480, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 466, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 466, + 480, + 487 + ], + "type": "text", + "content": "9. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 488, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 488, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 488, + 480, + 520 + ], + "type": "text", + "content": "10. Huang, P.H., Matzen, K., Kopf, J., Ahuja, N., Huang, J.B.: Deepmvs: Learning multi-view stereopsis. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 522, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 480, + 564 + ], + "type": "text", + "content": "1. Huang, Y., Zheng, W., Zhang, Y., Zhou, J., Lu, J.: Tri-perspective view for vision-based 3d semantic occupancy prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9223-9232 (June 2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 566, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 566, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 566, + 480, + 588 + ], + "type": "text", + "content": "2. Im, S., Jeon, H.G., Lin, S., Kweon, I.S.: DPSNet: End-to-end deep plane sweep stereo. In: International Conference on Learning Representations (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 589, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 620 + ], + "type": "text", + "content": "3. Jiang, Y., Zhang, L., Miao, Z., Zhu, X., Gao, J., Hu, W., Jiang, Y.G.: Polarformer: Multi-camera 3d object detection with polar transformer. Proceedings of the AAAI Conference on Artificial Intelligence 37(1), 1042-1050 (Jun 2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "text", + "content": "4. Li, H., Li, Z., Akmandor, N.U., Jiang, H., Wang, Y., Padir, T.: Stereovoxelnet: Real-time obstacle detection based on occupancy voxels from a stereo camera using deep neural networks. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 4826-4833 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "15. Li, Y., Li, S., Liu, X., Gong, M., Li, K., Chen, N., Wang, Z., Li, Z., Jiang, T., Yu, F., Wang, Y., Zhao, H., Yu, Z., Feng, C.: Sscbench: Monocular 3d semantic scene completion benchmark in street views (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 195 + ], + "type": "text", + "content": "16. Li, Y., Yu, Z., Choy, C., Xiao, C., Alvarez, J.M., Fidler, S., Feng, C., Anandkumar, A.: Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9087-9098 (June 2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "type": "text", + "content": "17. Li, Y., Bao, H., Ge, Z., Yang, J., Sun, J., Li, Z.: Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. Proceedings of the AAAI Conference on Artificial Intelligence 37(2), 1486-1494 (Jun 2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 481, + 262 + ], + "type": "text", + "content": "18. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. Proceedings of the AAAI Conference on Artificial Intelligence 37(2), 1477-1485 (Jun 2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 262, + 481, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 481, + 285 + ], + "type": "text", + "content": "19. Li, Y., Yang, J., Sun, J., Bao, H., Ge, Z., Xiao, L.: Bevstereo++: Accurate depth estimation in multi-view 3d object detection via dynamic temporal stereo (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 285, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 329 + ], + "type": "text", + "content": "20. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 1-18 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 330, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 481, + 373 + ], + "type": "text", + "content": "21. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X.: Petrov2: A unified framework for 3d perception from multi-camera images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3262-3272 (October 2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 374, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 397 + ], + "type": "text", + "content": "22. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 397, + 481, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 420 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 420 + ], + "type": "text", + "content": "23. Miao, R., Liu, W., Chen, M., Gong, Z., Xu, W., Hu, C., Zhou, S.: Occdepth: A depth-aware method for 3d semantic scene completion (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 420, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 481, + 453 + ], + "type": "text", + "content": "24. Moravec, H., Elfes, A.: High resolution maps from wide angle sonar. In: Proceedings. 1985 IEEE International Conference on Robotics and Automation. vol. 2, pp. 116-121 (1985)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 453, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 497 + ], + "type": "text", + "content": "25. Park, J., Xu, C., Yang, S., Keutzer, K., Kitani, K.M., Tomizuka, M., Zhan, W.: Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. In: The Eleventh International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 498, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 542 + ], + "type": "text", + "content": "26. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020. pp. 194-210. Springer International Publishing, Cham (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 543, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 481, + 586 + ], + "type": "text", + "content": "27. Qin, Z., Chen, J., Chen, C., Chen, X., Li, X.: Unifusion: Unified multi-view fusion transformer for spatial-temporal representation in bird's-eye-view. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 8690-8699 (October 2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 587, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 481, + 610 + ], + "type": "text", + "content": "28. Thrun, S.: Probabilistic robotics. Communications of the ACM pp. 52-57 (Mar 2002)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 610, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 666 + ], + "type": "text", + "content": "29. Tian, X., Jiang, T., Yun, L., Mao, Y., Yang, H., Wang, Y., Wang, Y., Zhao, H.: Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 64318-64330. Curran Associates, Inc. (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Z. Ye et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 567 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "30. Tong, W., Sima, C., Wang, T., Chen, L., Wu, S., Deng, H., Gu, Y., Lu, L., Luo, P., Lin, D., Li, H.: Scene as occupancy. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 8406-8415 (October 2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "text", + "content": "31. Wang, K., Shen, S.: Mvdepthnet: Real-time multiview depth estimation neural network. In: 2018 International Conference on 3D Vision (3DV) (Sep 2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 215 + ], + "type": "text", + "content": "32. Wang, X., Zhu, Z., Xu, W., Zhang, Y., Wei, Y., Chi, X., Ye, Y., Du, D., Lu, J., Wang, X.: Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 17850-17859 (October 2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "type": "text", + "content": "33. Wang, Y., Lai, Z., Huang, G., Wang, B.H., van der Maaten, L., Campbell, M., Weinberger, K.Q.: Anytime stereo image depth estimation on mobile devices. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 5893-5900 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 480, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 313 + ], + "type": "text", + "content": "34. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Faust, A., Hsu, D., Neumann, G. (eds.) Proceedings of the 5th Conference on Robot Learning. Proceedings of Machine Learning Research, vol. 164, pp. 180-191. PMLR (08-11 Nov 2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 315, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 358 + ], + "type": "text", + "content": "35. Wang, Y., Chen, Y., Liao, X., Fan, L., Zhang, Z.: Panoocc: Unified occupancy representation for camera-based 3d panoptic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17158-17168 (June 2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 358, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 413 + ], + "type": "text", + "content": "36. Xu, C., Wu, B., Hou, J., Tsai, S., Li, R., Wang, J., Zhan, W., He, Z., Vajda, P., Keutzer, K., Tomizuka, M.: Nerf-det: Learning geometry-aware volumetric representation for multi-view 3d object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 23320-23330 (October 2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 413, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 468 + ], + "type": "text", + "content": "37. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., Zhou, J., Dai, J.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17830-17839 (June 2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "text", + "content": "38. Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstructured multi-view stereo. In: Proceedings of the European Conference on Computer Vision (ECCV) (September 2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "text", + "content": "39. Zhang, F., Prisacariu, V., Yang, R., Torr, P.H.: Ga-net: Guided aggregation net for end-to-end stereo matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 533, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 480, + 567 + ], + "type": "text", + "content": "40. Zhang, Y., Zhu, Z., Du, D.: Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9433-9443 (October 2023)" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 157, + 91, + 447, + 102 + ], + "type": "text", + "content": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CadVLM_ Bridging Language and Vision in the Generation of Parametric CAD Sketches/7ca1dca8-97a9-4690-a292-e4d8a7eacc0d_content_list.json b/2024/CadVLM_ Bridging Language and Vision in the Generation of Parametric CAD Sketches/7ca1dca8-97a9-4690-a292-e4d8a7eacc0d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2d13b4637d634ca603ca46d38114b24c6812fd82 --- /dev/null +++ b/2024/CadVLM_ Bridging Language and Vision in the Generation of Parametric CAD Sketches/7ca1dca8-97a9-4690-a292-e4d8a7eacc0d_content_list.json @@ -0,0 +1,1624 @@ +[ + { + "type": "text", + "text": "CadVLM: Bridging Language and Vision in the Generation of Parametric CAD Sketches", + "text_level": 1, + "bbox": [ + 228, + 140, + 774, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sifan Wu1, Amir Hosein Khasahmadi2, Mor Katz2, Pradeep Kumar Jayaraman2, Yewen Pu2, Karl Willis2, and Bang Liu1", + "bbox": [ + 251, + 210, + 750, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Université de Montréal & MILA, Montreal, Canada \n2 Autodesk AI Lab, Toronto, Canada", + "bbox": [ + 323, + 253, + 676, + 282 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b76ecdde05298a4355611c5a2772e7e571b5f9ab6b6028dd1b2470845351ffb0.jpg", + "image_caption": [ + "Fig. 1: For the CAD autocompletion task, our multi-modal CadVLM model (b) receives partial CAD entities as both image and text input (a) and generates the remaining sketch entities as output (c). The complete sketch, with optional predicted constraints, can then be used in CAD software (d) to form 3D shapes (e). More description about the primitive values in sketch text are in the Appendix." + ], + "image_footnote": [], + "bbox": [ + 218, + 313, + 776, + 438 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Parametric Computer-Aided Design (CAD) is central to contemporary mechanical design. However, it encounters challenges in achieving precise parametric sketch modeling and lacks practical evaluation metrics suitable for mechanical design. We harness the capabilities of pre-trained foundation models, renowned for their successes in natural language processing and computer vision, to develop generative models specifically for CAD. These models are adept at understanding complex geometries and design reasoning, a crucial advancement in CAD technology. In this paper, we propose CadVLM, an end-to-end vision language model for CAD generation. Our approach involves adapting pre-trained foundation models to manipulate engineering sketches effectively, integrating both sketch primitive sequences and sketch images. Extensive experiments demonstrate superior performance on multiple CAD sketch generation tasks such as CAD autocomblection, CAD autoconstraint, and image conditional generation. To our knowledge, this is the first instance of a multimodal Large Language Model (LLM) being successfully applied to parametric CAD generation, representing a pioneering step in the field of computer-aided mechanical design.", + "bbox": [ + 259, + 563, + 743, + 815 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: CAD Representation $\\cdot$ Vision language model", + "bbox": [ + 261, + 825, + 648, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 143, + 374, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Parametric Computer-Aided Design (CAD) is crucial in mechanical engineering, playing a key role in developing modern objects such as lamps, cars, and buildings. Its importance lies in its ability to accurately depict complex designs, essential for the integrity and functionality of engineered products interacting with humans and the environment. CAD sketches, consisting of basic 2D shapes, are foundational for parametric CAD modeling [3]. As illustrated in Fig [1], these sketches comprise primitive geometric elements (e.g., lines, arcs, points) defined by various relationships (e.g., coincident, parallel, tangent). However, the required precision for 3D construction in parametric CAD can be both challenging and time-consuming [15]. Consequently, there is a significant push to improve CAD generation to simplify and accelerate the manual design process. Leveraging deep learning to identify patterns in CAD sketches could automate the completion of geometry and constraints in engineering drawings [22,23,28].", + "bbox": [ + 212, + 175, + 787, + 372 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Integrating deep learning into CAD offers a significant opportunity to transform the field. Recent research [10,14,19,22,23,29,31] has concentrated on the generative modeling of engineering sketches, particularly focusing on Transformer-based architectures [26]. These approaches often utilize graph structures [22] or attention mechanisms [23] to process sketch information. However, the effectiveness of these methods is limited by the availability of data and a lack of commonsense understanding, which restricts their capacity to perform complex geometric reasoning in engineering sketches. Additionally, rendered sketch images (such as in Fig. 1(a) and 4) have great potential to represent geometry patterns from image modality, which could be complementary for parametric sketches language. While none of current CAD generative models take sketch image into consideration. Our research presents a novel Multimodal Transformer-based Generative model for CAD generation tasks, pioneering in its use of both language and image for sketch modeling. By merging these modalities, we aim to increase the accuracy of CAD generation, setting a new benchmark by simultaneously modeling text-style sketch parametric and image-style sketch renderings.", + "bbox": [ + 212, + 373, + 789, + 612 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large Language Models (LLMs) have gained widespread success for their powerful commonsense reasoning and problem-solving abilities in a variety of domains, such as natural language processing, mathematics, and medical sciences [1,5,7,18,24,25,33]. The capabilities of LLMs align well with engineering sketches, which can be understood as sequences representing points and connections. Commonsense knowledge in pre-trained LLM can also offer valuable hints for sketch patterns. Furthermore, combining sketch images with textual descriptions can greatly improve the performance of language models in CAD sketch generative models. In the realm of parametric CAD design, which requires a thorough understanding of complex interactions(e.g. parallel, coincident, symmetry), the advantages of LLMs become even more pronounced.", + "bbox": [ + 212, + 613, + 787, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we leverage the capabilities of pre-trained LLMs to develop a generative model tailored for parametric CAD, addressing the limitations of the existing methods. We introduce CadVLM, a CAD Vision Language Model, featuring a multimodal encoder-decoder framework. CadVLM stands out by pro", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. Wu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cessing engineering sketches both linguistically, through token sequences representing sketch geometry and constraints, and visually, through the use of rendered sketch images. To the best of our knowledge, CadVLM is the first generative model to utilize pre-trained vision and language models for CAD sketches, which is capable of addressing different CAD representation learning tasks such as CAD autocompletion, CAD autoconstraint, and image conditional generation tasks. Our contributions are threefold:", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose CadVLM, a novel CAD Vision-Language Model, for simultaneously modeling engineering sketches as both language and images. As we know, we are the first to combine visual and textual data for CAD generative models.", + "- We introduce three novel evaluation metrics—Entity Accuracy, Sketch Accuracy, and CAD F1 score—to quantitatively assess the quality of generated CAD sketches.", + "- CadVLM demonstrates superior performance on the SketchGraphs dataset [22] in both CAD autocompletion and autoconstraint tasks." + ], + "bbox": [ + 225, + 260, + 794, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 416, + 387, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision Language Pre-Training. The great success of large language models (LLMs) such as BERT [11] and GPT [2] has drawn significant attention within the field of natural language processing and code generation domain. CodeT5+ [27] is a family of open-source encoder-decoder LLMs for code tasks, ranging from 220 million to 16 Billion parameters. Beyond single modality LLMs, multimodal foundation models have also exhibited enhanced performance in various vision-and-language tasks [6,12,20]. Depending on downstream tasks, Various pre-training objectives have also been proposed like image-text contrastive learning [20], image-text matching [12], and (masked) language modeling [12]. Most recently, developing domain-specific multimodality LLMs has gained momentum such as MedBLIP [4] and TableGPT [33].", + "bbox": [ + 212, + 446, + 787, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CAD Sketch Generation 2D engineering sketches are fundamental in CAD for designing and manufacturing mechanical parts. The recent availability of extensive engineering sketch datasets has facilitated the development of generative models that enhance traditional CAD workflows. The SketchGraphs [22] dataset, featuring 15 million parametric CAD sketches with detailed geometric constraint graphs, showcases primitive interrelations. The paper also introduces a baseline model using autoregressive message-passing networks for constraint graph generation, but it lacks primitive coordinate output, relying solely on constraint graphs and the Onshape solver for sketch configuration, thus limiting its utility in CAD autocomblection tasks. Ganin et al. [8] proposed another large-scale CAD dataset containing over 4.7M parametric sketches from the Onshape public repository. However, limited by sketch format, this dataset has not been widely researched in the realm of CAD generative models.", + "bbox": [ + 212, + 613, + 787, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another line of work for generative modeling of CAD sketches is based on Transformer [23,28-30]. Vitruvion [23] models both primitives and constraints,", + "bbox": [ + 214, + 809, + 785, + 842 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Abbreviated paper title", + "bbox": [ + 571, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "but its training on Transformer does not inherently include geometric knowledge. Furthermore, Vitruvion's major limitation lies in its lack of quantitative comparison in terms of CAD generation quality. While CurveGen [28] tackles the problem of engineering sketch generation without considering sketch constraints like us, it struggles to generalize to few-shot samples. Both Vitruvion and CurveGen face limitations in their model performance, making them challenging to deploy in practical CAD engineering applications. Recently, there has been some research about CAD construction sequences generation [30,31]. However, these works cannot model sketch primitives to autocomplete parametric CAD sketches. We consider Vitruvion, Deepcad, and SketchGraphs as the baselines for CAD Autocompletion task and CAD Autoconstraint task since they are current state-of-the-art models using transformer architectures. Different from existing work, our method utilizes sketch image modality and commonsense knowledge in pre-trained foundation models, providing both sketch sequence textual and image information for sketch generation.", + "bbox": [ + 212, + 146, + 787, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Background", + "text_level": 1, + "bbox": [ + 215, + 398, + 369, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Engineering Sketch Definition. In the field of parametric CAD, engineering sketches are fundamental, serving as detailed blueprints for designing a range of objects from small components to large structures. These sketches are essential for converting conceptual ideas into precise, manufacturable designs. To ensure a standardized representation, sketches of varying dimensions, spanning from millimeters to meters, are uniformly positioned within a bounding box of 1-meter width at the origin. Each sketch consists of two key sequences: a sequence of primitives, $S$ , and a sequence of constraints, $C$ .", + "bbox": [ + 212, + 431, + 787, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We adopt the normalization and quantization methods for primitive parameters as outlined by [23]. A sketch primitive is expressed as $S = (e_1, e_2, \\dots, e_m)$ , where $m$ represents the number of entities in the sketch. Each entity $e_i$ is characterized by positional parameters $(p_1, \\dots, p_k)$ , where $p$ represents the normalized coordinates of points that constitute the entity. Sketches typically include three types of entities—lines, arcs, and circles. As illustrated in Figure 11, all sketches are centrally positioned and normalized within a range of [1,64]. Lines, shown in blue, are depicted using two points: start and end point coordinates. Curves, shown in green, are characterized by three points: the start, the end, and a mid-point on the curve. Circles, illustrated in red, are portrayed with four uniformly distributed points on the circle's circumference. This format allows for the interpretation of each sketch as a sequence of tokens. Constraints are similarly represented as primitives in $\\mathcal{C} = (c_1, \\dots, c_n)$ , with each constraint entity $c$ being defined by its type and the reference indices of parameters to which the primitives must conform.", + "bbox": [ + 212, + 551, + 787, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CAD AutoCompletion Task. In addressing the existing CAD AutoCompletion task, which aims to automate routine CAD design procedures, we propose a new autoregressive generative model. This task involves completing a partially given sketch to form a fully realized design. The mathematical formulation of", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. Wu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fb06fae5aa28f00a34061e4f935404b39b30c03ba9aad4b5e3105e1e54fda047.jpg", + "image_caption": [ + "Fig. 2: CadVLM model architecture and training objectives for CAD autoconstraint. We propose a unified vision-language model capable of outputting both autocompleted sketch images and sketch primitives. The inputs are paired with incomplete sketch primitives and images, which are encoded by pre-trained text encoder and image encoder correspondingly. Aligned by image-text contrastive(ITC) loss, the concatenation of image embedding and text embedding then will be input to the text-grounded image decoder and image-grounded text decoder. The image decoder is trained with image decoding loss between autocompleted sketch images and ground-truth sketch images. The text decoder is trained with a language modeling(LM) loss to generate autocompleted primitive." + ], + "image_footnote": [], + "bbox": [ + 218, + 147, + 782, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the task is:", + "bbox": [ + 215, + 525, + 297, + 537 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\Theta) = - \\log p _ {\\Theta} (\\mathcal {S} | \\mathcal {S} _ {p}) = - \\sum_ {i = j} ^ {m} \\log p _ {\\Theta} \\left(e _ {i: m} \\mid e _ {1: i}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 539, + 784, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $m$ is the total number of entities in the complete sketch $S$ , and $S_{p} = (e_{1},\\dots,e_{i})$ is a partial sequence containing $20\\%$ to $80\\%$ of the primitive entities from $S$ . Our model's goal is to optimize the log-likelihood of completing the sketch $S$ based on these partial sequences.", + "bbox": [ + 214, + 580, + 787, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We employ open-source natural language pre-trained Transformer models for our generative model, with a particular emphasis on T5 models [21]. Additionally, we explore finetuning with GPT-3.5 [2]. Our results indicate significant cross-domain knowledge transfer capabilities in these pre-trained models, showcasing their effectiveness in the context of CAD design.", + "bbox": [ + 212, + 646, + 787, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CAD Autoconstrain Task. The CAD autoconstrain task aims to autoregressively generate a set of constraints based on given geometry primitives. We model the constraint model as", + "bbox": [ + 214, + 726, + 787, + 770 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\phi) = - \\log p _ {\\Phi} (\\mathcal {C} | \\mathcal {S}) = - \\sum_ {i = 1} ^ {n} \\log p _ {\\Phi} \\left(C _ {j: n} \\mid C _ {1: j}, S\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 771, + 785, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n$ is the number of constraints, and each $C_i$ is represented by a tuple of tokens indicating its constraint type and corresponding primitive entity param", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Abbreviated paper title", + "bbox": [ + 571, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "eters. Like autocompletion task, the autoconstrain model is trained to maximize the log-likelihood of generated tokens with respect to the ground truth.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Aligned with [23], we focus on 13 constraint types, which are coincident, parallel, and so on. Further details about these constraints are provided in the Supplementary.", + "bbox": [ + 212, + 176, + 782, + 220 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. To quantitatively assess CAD generation models, we introduce three CAD-specific metrics: Sketch Accuracy, Entity Accuracy, and CAD F1.", + "bbox": [ + 212, + 222, + 782, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sketch Accuracy measures the probability of correctly generating the entire remaining sketch in the test set. Entity Accuracy evaluates the likelihood of correctly generating at least one entity in the ground truth. They are calculated as:", + "bbox": [ + 212, + 267, + 784, + 324 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\text {S k e t c h A c c u r a c y} = \\frac {N _ {s}}{N}, \\quad \\text {E n t i t y A c c u r a c y} = \\frac {N _ {e}}{N}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 325, + 782, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $N_{s}$ is the number of correctly generated sketches in the test set, $N_{e}$ is the number of generated sketches that match at least one entity with the ground truth sketch, and $N$ is the total number of sketches in the test set. CAD F1 is a harmonic mean of precision and recall for each sketch, averaged over the dataset:", + "bbox": [ + 212, + 359, + 782, + 419 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nC A D F 1 = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {2 \\times \\text {s k e t c h p r e c i s i o n} \\times \\text {s k e t c h r e c a l l}}{(\\text {s k e t c h p r e c i s i o n} + \\text {s k e t c h r e c a l l})}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 431, + 782, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where sketch precision $= \\frac{n_c}{n_p}$ , sketch recall $= \\frac{n_c}{m}$ , $n_c$ is the count of correct entities in the completed sketch, $n_p$ is the total number of entities predicted, and $m$ is the number of entities in the original sketch (ground truth).", + "bbox": [ + 212, + 470, + 782, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 CadVLM: A Vision Language Model for CAD Generation", + "text_level": 1, + "bbox": [ + 212, + 542, + 705, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we delve into the specifics of our CAD Vision Language Model (CadVLM), employing the CAD autocompletion task as an illustrative example. In our approach, the sequence of sketch primitives $S$ is interpreted akin to language text, while the corresponding visual representation of the sketch is treated as image $I$ . It should be noted that the text-grounded image decoder is not utilized in the CAD auto-constraint task.", + "bbox": [ + 212, + 593, + 782, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Network Framework", + "text_level": 1, + "bbox": [ + 214, + 708, + 429, + 722 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CadVLM is structured as an asymmetrical encoder-decoder architecture, featuring a two-stream encoder, a text decoder based on a language model, and a lightweight image decoder for constructing the autocompleted image. This design is depicted in Figure 2. The two-stream encoder is tailored to process dual-modal inputs: a vision sub-encoder for the sketch image $I$ and a text sub-encoder for the prefix sequence of sketch primitives $S_{p}$ . For text decoding, we employ a causal language transformer model, which autoregressively generates", + "bbox": [ + 212, + 733, + 782, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. Wu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the output, i.e., the autocompleted sketch. Finally, the image decoder utilizes the latent representation tokens to construct the visual representation of the autocompleted sketch.", + "bbox": [ + 212, + 146, + 784, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Vision Encoding Stream. To encode partial sketch images and harness the capabilities of an existing large vision autoencoder model, we employ a pretrained vision transformer (ViT) [9] for extracting vision features. To bridge the gap between CAD images and natural images typically processed by ViT, we fine-tune the model on the SketchGraphs dataset. This fine-tuning involves a one-epoch image reconstruction task, with a $75\\%$ masking ratio, specifically tailored for CAD sketch data. Figure 3 illustrates the image reconstruction outcomes using both the original pre-trained ViT-MAE and the version fine-tuned on the SketchGraphs dataset. After finetuning, ViT-MAE shows enhanced representational performance for CAD sketches.", + "bbox": [ + 212, + 194, + 787, + 345 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e783fb245e3cd97204a1943891292cce70129cf25e18c9c916a63bf6a8f6580c.jpg", + "image_caption": [ + "Fig.3: CAD image reconstruction results by ViTMAE. (a) Input sketches, (b) masked sketches with the mask ratio of $75\\%$ , reconstructed results using a (c) pretrained and (d) further fine-tuned ViTMAE." + ], + "image_footnote": [], + "bbox": [ + 225, + 378, + 496, + 553 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "However, it's important to note that while some reconstructed sketches are notably accurate, they do not fully meet the requirements for parametric CAD engineering due to the vagueness and lack of parameterization in the entities. To bring image and text features to the same embedding space, we integrate a down-sampling layer and a lightweight projection module atop the vision encoder.", + "bbox": [ + 511, + 348, + 787, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Text Encoding Stream. Treating the sketch primitive sequence as textual data, we first tokenize this sequence to derive a series of text tokens $T = t_{1}, t_{2}, \\ldots, t_{l}$ . These tokens are then processed using the encoder of a pre-trained language model, effectively encoding the sketch information in a textual format.", + "bbox": [ + 511, + 517, + 787, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Text-grounded Vision Decoding Stream. This stream is dedicated", + "bbox": [ + 513, + 655, + 787, + 685 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to constructing the autocompleted pixels, mapping the fused image and text representation to the pixel-level sketch image. The input for the vision decoder comprises the fused image and text embeddings, augmented with adaptively masked token patches to keep the same patch size as the input image. This construction process ensures that the vision embedding is grounded by textual information and also enhances the text representation. Note that for CAD autoconstraint task, we don't reconstruct the image since the input sketches are complete sketch images already.", + "bbox": [ + 212, + 686, + 787, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Image-grounded Text Decoding Stream. In alignment with Equations 1 and 2, we conceptualize autocompletion as a conditional text generation process.", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Abbreviated paper title", + "bbox": [ + 571, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Within this framework, the combined image and text embeddings serve as inputs to the decoder of a language model. This decoder is capable of autoregressively generating the required output. Notably, the parameters of this language model are initialized from a pre-trained model, which itself has undergone extensive pre-training on vast datasets of code [27], enhancing its effectiveness in this context.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Training Objectives", + "text_level": 1, + "bbox": [ + 215, + 262, + 426, + 277 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Figure 2, we jointly optimize three distinct objectives during training: one focused on understanding and two on generation.", + "bbox": [ + 212, + 290, + 785, + 320 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Image-Text Contrastive Loss (ITC) is employed to activate the unimodal encoders, following the approach in [12]. This loss function aligns the visual and textual feature spaces by ensuring that positive image-text pairs have similar representations while contrasting with negative pairs. Such a contrastive approach has been instrumental in enhancing multimodal model training [13,20], not only improving the alignment of visual and language features but also accelerating the convergence during training. The image-text contrastive loss is defined as the cross-entropy between the similarity $p$ of image(I) text(T) embedding pairs and ground-truth one-hot similarity $y$ (array from [0,N], N is the number of samples), formulated as:", + "bbox": [ + 212, + 321, + 787, + 472 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\np = \\frac {\\exp (s (I , T))}{\\sum_ {i , j} ^ {N} \\exp (s (I _ {i} , T _ {j}))}, \\mathcal {L} _ {I T C} = \\text {C r o s s - E n t r o p y} (y, p). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 484, + 785, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Image Decoding Loss (IDL) is responsible for constructing autocompletion sketch images. This process involves translating the fused visual and textual information into pixel values for each patch. IDL employs a mean squared error (MSE) loss, similarly to the approach in ViTMAE [9], comparing the autocompleted sketch images with their ground-truth counterparts at the pixel level. The Image decoding loss can be formulated as:", + "bbox": [ + 214, + 549, + 787, + 642 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I D L} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(Y _ {i} - \\hat {Y} _ {i}\\right) ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 654, + 785, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $n$ is the number of total pixels, $Y_{i}$ is the ground-truth pixel value of $i$ -th pixel, and $\\hat{Y}_i$ is the output pixel value correspondingly. For CAD autoconstraint tasks, since the input is a full sketch image, we ignore IDL during the training of CAD autoconstraint models.", + "bbox": [ + 212, + 703, + 787, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Language Modeling Loss (LM) is designed to generate text-like autocompletion sketches, focusing on constraint sequences and utilizing the image-grounded text decoder. It optimizes a cross-entropy loss to maximize the likelihood described in Equation ①. LM endows the model with the capability to transform fused multimodal information into coherent autocompletion sketch", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. Wu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sequences. For the CAD autoconstraint task, we use Equation 2 as the language modeling loss.", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Therefore, the full training objective of CadVLM is:", + "bbox": [ + 238, + 176, + 617, + 191 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {I T C} + \\mathcal {L} _ {L M} + \\mathcal {L} _ {I D L}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 207, + 785, + 222 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "During inference, given incomplete sketch image and primitive pairs, we only evaluate the generated text and use that for further mechanical applications.", + "bbox": [ + 212, + 229, + 782, + 260 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 214, + 282, + 375, + 301 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conduct extensive experiments and ablation studies on CadVLM for the CAD autocompletion and autoconstraint tasks. Quantitative assessment is conducted by measuring Entity Accuracy, Sketch Accuracy, and CAD F1.", + "bbox": [ + 212, + 315, + 782, + 359 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Experimental settings", + "text_level": 1, + "bbox": [ + 214, + 383, + 442, + 398 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dataset We use the SketchGraphs [22] dataset for all our experiments, consisting of 15 million CAD sketches collected from Onshape. Referring to [23], we use the filtered version and further deduplicate the dataset. After deduplicated, the training, validation, and test set have 626,236, 22,031, 21,979 unique sketches correspondingly. The filtered sketches are restricted to sketches comprised of arcs, circles, and lines. To ensure a fair comparison, we normalize each sketch as in [23]; specifically the quantized range of numerical tokens in the primitive sequence is [1, 64].", + "bbox": [ + 212, + 407, + 784, + 527 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baselines We evaluate several ablation and conditional variants of CadVLM, as well as three SOTA CAD generative model baselines: Vitruvion [23], Deepcad [29] and SketchGraphs [22]. Since SketchGraphs does not incorporate continuous primitive parameters such as coordinate value, we only compare with the CAD autoconstraint task. And DeepCAD only considers parametrized command sequences, which are incompatible with CAD constraints. We primarily evaluate its performance on CAD autocompletion tasks.", + "bbox": [ + 212, + 529, + 784, + 633 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Settings For the image encoder and decoder, we choose the encoder and decoder of state-of-the-art pre-trained ViT-MAE-base from [9]. Following [9] we set the patch size and stride both as 32. We render all sketch images with size of $224 \\times 224 \\times 3$ . For the text encoder and text decoder, we use the encoder and decoder of pre-trained CodeT5+ [27] correspondingly, which is an instruction-tuned code LLM with 770M parameters trained on various code datasets. We train the autocompletion model and autoconstraint model both for 30 epochs, using the Adam optimizer with decoupled weight decay regularization [17], with the learning rate set according to the CosineAnnealingLR [16] scheduler. The initial learning rate is set to $3e - 4$ , with batch size 32. All our experiments are run using 8 NVIDIA A100 GPUs, and our models take between 30-40 hours to train. In the CAD Autocompletion experiments, we randomly sample $20\\% -80\\%$", + "bbox": [ + 212, + 635, + 785, + 816 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Abbreviated paper title", + "bbox": [ + 571, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "3 https://www.onshape.com", + "bbox": [ + 217, + 823, + 411, + 839 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/71facc16a8f0f9ebf2533a6e898c45b689acc633060a28f6b1db10746ab9c394.jpg", + "table_caption": [ + "Table 1: Quantitative sketch generative model comparison results for CAD autocompletion task. Ske-Acc means the probability of generating the correct full remaining sketch and Ent-Acc means the probability of generating at least one entity of the ground truth. Bold shows the best results of each metric. For all metrics the higher the better." + ], + "table_footnote": [], + "table_body": "
ModelSke-AccEnt-AccCAD-F1
Vitruvion3.0%40.7%11.3%
Deepcad1.2%23.1%8.9%
ChatGPT-15%6.8%41.3%21.2%
ChatGPT-7%7.6%43.8%22.2%
ChatGPT-1%6.5%43.7%22.5%
CadVLM-Text21.2%66.7%42.8%
CadVLM-w/o-IDL22.6%68.4%44.2%
CadVLM-w/o-ITC22.8%67.6%43.9%
CadVLM-w/o-IDL&ITC23.0%67.7%44.1%
CadVLM23.8%68.3%45.2%
", + "bbox": [ + 325, + 226, + 674, + 390 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/1f5b17cd0bd8d9226804305d089bc3ef773691c94a3630e8c3cbf58944d8015f.jpg", + "table_caption": [ + "Table 2: Quantitative sketch generative model comparison results for CAD autoconstraint task. For all metrics higher is better." + ], + "table_footnote": [], + "table_body": "
ModelSke-AccEnt-AccCAD-F1
SketchGraphs0.48%93.0%42.0%
Vitruvion10.7%94.6%71.6%
CadVLM-Text11.5%95.0%72.6%
CadVLM-w/o-ITC12.1%95.4%74.5%
CadVLM11.3%94.9%73.8%
", + "bbox": [ + 344, + 441, + 655, + 537 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "of sketch entities as input and aim to predict the remaining entities from the ground truth sketch.", + "bbox": [ + 212, + 566, + 785, + 597 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Quantitative comparison with baseline sketch generative models", + "text_level": 1, + "bbox": [ + 212, + 619, + 785, + 635 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 1 reports results on the CAD Autocompletion task. We find that Vitruvion achieves $3\\%$ on Sketch Accuracy and $40.7\\%$ on Entity Accuracy, meaning it can accurately predict $3\\%$ of sketches and correctly identifies at least one entity in $40.7\\%$ of sketches. Deepcad performs worst since it only has 4 layers of encoder and decoder, makes it hard to represent complex CAD sketches. Our CadVLM outperforms Virruvion on Sketch Accuracy, Entity Accuracy, and CAD-F1 by $20.8\\%$ , $28.3\\%$ , and $39.9\\%$ respectively. These results demonstrate the effectiveness of pre-trained foundational models and image modality. A closer examination of the qualitative results in Figure 4 reveals that Vitruvion produces many invalid results missing symmetric curves or lines. Sketches from CadVLM are better in terms of quality with more complex shapes and stronger symmetry.", + "bbox": [ + 212, + 643, + 787, + 809 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "ChatGPT Series Model Finetune To evaluate the ability of a general LLM on the CAD autocompletion task, we fine-tuned ChatGPT on three subsets", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. Wu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/6fa3e0379ea4f351096e7c56ee38ffb02db6dec4adc16eb015e6c7fe2feb043e.jpg", + "image_caption": [ + "Fig. 4: Comparative Analysis of Autocompletion in CAD Design. Top row: Random samples from the SketchGraphs test dataset. Second row: Initial input entities serving as the primer for the autocompletion task. Third and fourth row: Autocompletion results produced by Vitruvion and CadVLM." + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 782, + 296 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "of $15\\%$ , $7\\%$ , and $1\\%$ of the SketchGraphs dataset. The results indicate that the ChatGPT-series outperforms Vitruvion on all three metrics, especially on Sketch Accuracy and CAD-F1. However, there were no significant performance improvements when finetuning on larger portions of the training dataset. One possible reason is we only finetune ChatGPT for less than 5 epochs due to training costs, making it hard for ChatGPT to adapt to CAD data.", + "bbox": [ + 212, + 393, + 787, + 486 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "CadVLM variants To better understand the source of performance gains, we conducted a comparative analysis of different variants of our proposed CadVLM: CadVLM-text, CadVLM-w/o-ITC, CadVLM-w/o-IDL&ITC, and CadVLM w/o-IDL. CadVLM-text relies solely on the sketch primitive sequence text as its input, optimizing for language modeling loss. In contrast, CadVLM-w/o-IDL utilizes ITC and LM loss for its training objectives, CadVLM-w/o-ITC focuses on IDL and LM, and CadVLM-w/o-IDL&ITC prioritizes only the LM objective. The findings reveal that CadVLM-Text underperforms compared to all other CadVLM variants across all three metrics, which underlines the value of the sketch image modality in providing diverse geometric information alongside sequence text.", + "bbox": [ + 212, + 488, + 790, + 652 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Interestingly, CadVLM-w/o-IDL emerges as the superior performer among the CadVLM variants, highlighting the significance of aligning image and text embeddings in multimodal modeling. Incorporating Image Decoding Loss (IDL) further enhances the alignment between image and text embeddings, leading CadVLM to show a notable improvement in all three metrics.", + "bbox": [ + 212, + 656, + 787, + 731 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For the CAD Autoconstraint task, as shown in Table 2, SketchGraphs struggles to predict the whole sketch correctly, with only $0.48\\%$ Sketch Accuracy. This suggests that relying solely on graph structures of sketches is insufficient for handling long sequences of sketch primitives. In contrast, Vitruvion achieves $10.7\\%$ on Sketch Accuracy, showing the necessity to model primitives. For the CAD Autoconstraint task it is less challenging to predict at least one constraint correctly but more challenging to accurately predict all constraints, easily caused", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Abbreviated paper title", + "bbox": [ + 571, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7b6a304b5cd51cb43d1faab4160f0cb561893f81784c2c80f1b338c5e815da9f.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 403, + 255 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/105a66c6699583199a1cfd7681410c6dd78e0e9b628d477664e19c86ffa84cb5.jpg", + "image_caption": [ + "(b)", + "Fig. 5: Effect of input entity ratio to (a) Entity Accuracy, (b) Sketch Accuracy, and (c) CAD F1." + ], + "image_footnote": [], + "bbox": [ + 408, + 143, + 594, + 255 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7840ffc800e1712acb1ca19b97e73bfe705811115cb05d3fd5384d80cfee2d9f.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 598, + 143, + 784, + 255 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "by potential data labeling inconsistencies. As is well-known, minor discrepancies in primitive values can drastically alter constraints, such as changing them from parallel to coincident.", + "bbox": [ + 212, + 339, + 782, + 383 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The CadVLM series model exceeds the performance of both SketchGraphs and Vitruvion across all metrics. CadVLM-Text surpasses CadVLM on Ske-Acc and Ent-Acc since the pretraining task of ViTMAE has a great discrepancy with CAD autoconstraint tasks. Therefore aligning the image and text representation directly might not be an optimal way for CAD autoconstraint task.Notably, CadVLM-w/o-ITC emerges as the top performer within the CadVLM series. This may be attributed to its more substantial text decoder and utilization of multimodal information, which enables the model to more effectively integrate knowledge from both primitives and images to inform geometric information.", + "bbox": [ + 212, + 385, + 784, + 521 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/da36c649555c8ca72f2a5df89b445bb03b0ba114367712f74fb0ea2d7315e918.jpg", + "table_caption": [ + "Table 3: Quantitative results for image condi- Table 4: Effect of adding the inductioned primitive generation. Bold represents tive bias of entity-level embeddings the best results of all metrics. to the sketch sequence." + ], + "table_footnote": [], + "table_body": "
Conditioned -ImageModelEnt-AccCAD-F1
PreciseVitruvion83.8%35.8%
CadVLM98.0%62.7%
Hand-drawnVitruvion84.2%38.4%
CadVLM99.4%60.1%
Noisy hand-drawnVitruvion83.2%35.8%
CadVLM94.4%56.5%
", + "bbox": [ + 251, + 584, + 529, + 705 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/97eae3def93c4d7d69d52b261ef73458b81980fa623df491d04cd8fd8395019e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelCadVLM w/-Inductive -TextBias
Ent-Acc66.7%67.5%
Ske-Acc21.2%22.2%
CAD-F142.8%43.7%
", + "bbox": [ + 552, + 584, + 772, + 681 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 726, + 390, + 742 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conducted several ablation studies to assess the performance of CadVLM on image-conditioned primitive generation tasks and investigate the effectiveness of each component within CadVLM.", + "bbox": [ + 212, + 750, + 787, + 792 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Image Conditional Generation. Following [23], we conduct a series of experiments to evaluate the effectiveness of CadVLM modeling primitive sequences conditioned on the images of sketches. As designers may begin their", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. Wu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "designs with hand-drawn sketches before transitioning to CAD programs, accurately inferring parametric CAD from images or scans can dramatically reduce the effort of translating from rough paper sketches to the CAD model.", + "bbox": [ + 212, + 146, + 784, + 191 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "When conditioning only on sketch images, the text encoder of CadVLM is ignored. We only use the image encoder, image decoder, and text decoder. The overall model is trained to maximize the probability of the ground truth primitives portrayed in the image, conditioned on a sequence of patch embeddings.", + "bbox": [ + 212, + 194, + 785, + 255 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The results of Table 3 show three kinds of sketch image rendering input to imitate different settings in the real world: precise renderings, rendering from the hand-drawn simulator of [23], and renderings with random affine augmentations of the hand-drawn simulator. As both Vitruvion and CadVLM can seldom predict the sketch exactly correctly, we ignore Sketch Accuracy in our image conditional generation experiments. The results show that CadVLM outperforms Vitruvion on both Entity Accuracy and CAD F1. Vitruvion performs best on hand-drawn images with $83.8\\%$ Entity Accuracy and $38.4\\%$ CAD F1. While CadVLM achieves almost $99.4\\%$ Entity Accuracy and $62.7\\%$ CAD F1. Conditioning on precise rendered images achieves the best performance.", + "bbox": [ + 212, + 257, + 787, + 409 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Entity Level Modelling. As described earlier, each CAD sketch is comprised of multiple entities. We considered adding this inductive bias to the model by processing each entity separately and then combining their embeddings with the initial sequence of all sketch tokens. For this purpose, we first pretend a special token $\\langle \\text{ENTITY} \\rangle$ to each of the entity input sequences. Subsequently, the entities from all the sketches in the batch are fed into the model's encoder in parallel. Next, the embeddings of each entity are gathered for each sketch. For each sketch, we concatenate these embeddings with the original sequence tokens using another special token $\\langle \\text{TOKEN} \\rangle$ as a delimiter for the second part of the sequence. The decoder remains unchanged. As shown in Table 4, incorporating this inductive bias increases computation and memory demands while only marginally enhancing the performance.", + "bbox": [ + 212, + 411, + 787, + 592 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Impact of different input entity ratios for CAD autocomblection. To explore the impact of different input entity ratios for CAD autocomblection, we further test CAD autocomblection results with different input entity ratios: $20\\%$ , $40\\%$ , $60\\%$ , and $80\\%$ . Compared with the experiments in Sec 5.4, we specify the sampled ratio of input entities. As shown in Figure 5, CadVLM outperforms Vitruvion on all input entity ratios. In particular, CadVLM achieves $73.8\\%$ Entity Accuracy, $49.5\\%$ Sketch Accuracy, and $62.7\\%$ CAD F1 when given $80\\%$ of entities as input. For the hardest setting when given only $20\\%$ of entities as input, CadVLM can still achieve $43.9\\%$ Entity Accuracy and $16.9\\%$ CAD-F1, which proves that CadVLM has great potential to be used in real world settings.", + "bbox": [ + 212, + 595, + 787, + 747 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effect of using different language models for text encoder and decoder. To better compare the performance of different text encoder-decoder, we experiment using different pre-trained foundation models as text encoder and decoder. To eliminate the influence of the image encoder and decoder, we only consider the CadVLM-Text variant. For the choice of text encoder and decoder, we also compare the performance of different Encoder-Decoder pre-trained", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Abbreviated paper title", + "bbox": [ + 571, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/eb803a6f3d0f2c516d9af618956509a1500a66da2d1246d08c4230b1a770cdd9.jpg", + "table_caption": [ + "Table 5: Effect of using different architectures for CadVLM-Text. CodeT5+ 770M is the encoder-decoder for text we use in our main experiments. CodeT5+\\* is the result of training from scratch rather than pre-trained." + ], + "table_footnote": [], + "table_body": "
Model#ParameterEnt-AccSke-AccCAD-F1
CodeT5+220M66.5%21.5%42.9%
CodeT5+*220M65.0%18.9%40.3%
CodeT5+770M66.7%21.2%42.8%
ByT5-base580M66.9%22.0%43.1%
ByT5-large1.2B66.8%20.6%41.9%
T5-base223M66.2%20.2%41.7%
T5-large783M66.8%21.2%42.6%
", + "bbox": [ + 326, + 198, + 671, + 319 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "language models on CAD autocombpletion. The T5 family of models are among the most recognized and widely used encoder-decoder models in the open-source community. Therefore, we experimented with T5 [21], ByT5 [32] and CodeT5+ [27]. The results are shown in Table 5. The effect of model size is clear for T5 and CodeT5+ but not ByT5. Also, The cost-benefit of solely making the model bigger becomes less favorable over time as the improvements in performance start to plateau. Another observation is that, because of the nature of the data which is closer to a code sequence than natural language, CodeT5+, which is fine-tuned on code, exhibits better performance.", + "bbox": [ + 212, + 349, + 787, + 487 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We also trained CodeT5+ from scratch to examine the impact of pre-training on natural language and code. As evidenced in Table5, training from scratch falls short in all three metrics. Additionally, training from scratch converged much slower than finetuning. Despite the fact that our representation of sketches as code is relatively unique and not found in natural language, the rules and token embeddings learned from text are very helpful for domain adaptation.", + "bbox": [ + 212, + 487, + 787, + 578 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 599, + 359, + 617 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this study, we introduce CadVLM, an innovative end-to-end vision-language model tailored for modeling parametric CAD sketches, which are pivotal in modern mechanical design. CadVLM harnesses the commonsense knowledge and reasoning abilities inherent in large language models, enabling it to adeptly handle the complex geometric reasoning required for CAD sketches. Our experimental results demonstrate that CadVLM outperforms existing state-of-the-art models in both autocompletion and autoconstraint tasks, highlighting the potent enhancement that pre-trained models bring to CAD construction.", + "bbox": [ + 212, + 632, + 787, + 753 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 777, + 403, + 795 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work is supported by the Canada CIFAR AI Chair Program and the Canada NSERC Discovery Grant (RGPIN-2021-03115).", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. Wu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Azerbayev, Z., Schoelkopf, H., Paster, K., Santos, M.D., McAleer, S., Jiang, A.Q., Deng, J., Biderman, S., Welleck, S.: Llemma: An open language model for mathematics. arXiv preprint arXiv:2310.10631 (2023)", + "2. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)", + "3. Camba, J.D., Contero, M., Company, P.: Parametric cad modeling: An analysis of strategies for design reusability. Computer-Aided Design 74, 18-31 (2016)", + "4. Chen, Q., Hu, X., Wang, Z., Hong, Y.: Medblip: Bootstrapping language-image pre-training from 3d medical images and texts. arXiv preprint arXiv:2305.10799 (2023)", + "5. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., et al.: Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794 (2022)", + "6. Chen, Y., Zhang, S., Han, B., He, T., Li, B.: Camml: Context-aware multimodal learner for large models. arXiv preprint arXiv:2401.03149 (2024)", + "7. Floridi, L., Chiriatti, M.: Gpt-3: Its nature, scope, limits, and consequences. Minds and Machines 30, 681-694 (2020)", + "8. Ganin, Y., Bartunov, S., Li, Y., Keller, E., Saliceti, S.: Computer-aided design as language (2021)", + "9. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)", + "0. Jones, B.T., Hu, M., Kodnongbua, M., Kim, V.G., Schulz, A.: Self-supervised representation learning for cad. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21327-21336 (2023)", + "1. Kenton, J.D.M.W.C., Toutanova, L.K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of naacL-HLT. vol. 1, p. 2 (2019)", + "2. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)", + "3. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems 34, 9694-9705 (2021)", + "4. Li, P., Guo, J., Zhang, X., Yan, D.M.: Secad-net: Self-supervised cad reconstruction by learning sketch-extrude operations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16826 (2023)", + "5. Lindsay, A., Paterson, A., Graham, I.: Identifying and quantifying inefficiencies within industrial parametric cad models. In: Advances in Manufacturing Technology XXXII: Proceedings of the 16th International Conference on Manufacturing Research. vol. 8, p. 227 (2018)", + "6. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)", + "7. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)", + "8. Luo, H., Sun, Q., Xu, C., Zhao, P., Lou, J., Tao, C., Geng, X., Lin, Q., Chen, S., Zhang, D.: Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct. arXiv preprint arXiv:2308.09583 (2023)" + ], + "bbox": [ + 225, + 172, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Abbreviated paper title", + "bbox": [ + 571, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Para, W., Bhat, S., Guerrero, P., Kelly, T., Mitra, N., Guibas, L.J., Wonka, P.: Sketchgen: Generating constrained cad sketches. Advances in Neural Information Processing Systems 34, 5077-5088 (2021)", + "20. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "21. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)", + "22. Seff, A., Ovadia, Y., Zhou, W., Adams, R.P.: Sketchgraphs: A large-scale dataset for modeling relational geometry in computer-aided design. arXiv preprint arXiv:2007.08506 (2020)", + "23. Seff, A., Zhou, W., Richardson, N., Adams, R.P.: Vitruvion: A generative model of parametric cad sketches. arXiv preprint arXiv:2109.14124 (2021)", + "24. Singhal, K., Azizi, S., Tu, T., Mahdavi, S.S., Wei, J., Chung, H.W., Scales, N., Tanwani, A., Cole-Lewis, H., Pfohl, S., et al.: Large language models encode clinical knowledge. arXiv preprint arXiv:2212.13138 (2022)", + "25. Singhal, K., Tu, T., Gottweis, J., Sayres, R., Wulczyn, E., Hou, L., Clark, K., Pfohl, S., Cole-Lewis, H., Neal, D., et al.: Towards expert-level medical question answering with large language models. arXiv preprint arXiv:2305.09617 (2023)", + "26. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "27. Wang, Y., Le, H., Gotmare, A.D., Bui, N.D., Li, J., Hoi, S.C.: Codet5+: Open code large language models for code understanding and generation. arXiv preprint arXiv:2305.07922 (2023)", + "28. Willis, K.D., Jayaraman, P.K., Lambourne, J.G., Chu, H., Pu, Y.: Engineering sketch generation for computer-aided design. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2105–2114 (2021)", + "29. Wu, R., Xiao, C., Zheng, C.: Deepcad: A deep generative network for computer-aided design models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6772-6782 (2021)", + "30. Xu, X., Jayaraman, P.K., Lambourne, J.G., Willis, K.D., Furukawa, Y.: Hierarchical neural coding for controllable cad model generation. arXiv preprint arXiv:2307.00149 (2023)", + "31. Xu, X., Willis, K.D., Lambourne, J.G., Cheng, C.Y., Jayaraman, P.K., Furukawa, Y.: Skexgen: Autoregressive generation of cad construction sequences with disentangled codebooks. In: International Conference on Machine Learning. pp. 24698-24724. PMLR (2022)", + "32. Xue, L., Barua, A., Constant, N., Al-Rfou, R., Narang, S., Kale, M., Roberts, A., Raffel, C.: Byt5: Towards a token-free future with pre-trained byte-to-byte models. Transactions of the Association for Computational Linguistics 10, 291-306 (2022)", + "33. Zha, L., Zhou, J., Li, L., Wang, R., Huang, Q., Yang, S., Yuan, J., Su, C., Li, X., Su, A., et al.: Tablegpt: Towards unifying tables, nature language and commands into one gpt. arXiv preprint arXiv:2307.08674 (2023)" + ], + "bbox": [ + 215, + 147, + 784, + 784 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. Wu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/2024/CadVLM_ Bridging Language and Vision in the Generation of Parametric CAD Sketches/7ca1dca8-97a9-4690-a292-e4d8a7eacc0d_model.json b/2024/CadVLM_ Bridging Language and Vision in the Generation of Parametric CAD Sketches/7ca1dca8-97a9-4690-a292-e4d8a7eacc0d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bd4b7d46674b9881c67c9730a1101ec77929889e --- /dev/null +++ b/2024/CadVLM_ Bridging Language and Vision in the Generation of Parametric CAD Sketches/7ca1dca8-97a9-4690-a292-e4d8a7eacc0d_model.json @@ -0,0 +1,2069 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.23, + 0.141, + 0.776, + 0.187 + ], + "angle": 0, + "content": "CadVLM: Bridging Language and Vision in the Generation of Parametric CAD Sketches" + }, + { + "type": "text", + "bbox": [ + 0.252, + 0.212, + 0.751, + 0.246 + ], + "angle": 0, + "content": "Sifan Wu1, Amir Hosein Khasahmadi2, Mor Katz2, Pradeep Kumar Jayaraman2, Yewen Pu2, Karl Willis2, and Bang Liu1" + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.255, + 0.678, + 0.283 + ], + "angle": 0, + "content": "1 Université de Montréal & MILA, Montreal, Canada \n2 Autodesk AI Lab, Toronto, Canada" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.314, + 0.777, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.518 + ], + "angle": 0, + "content": "Fig. 1: For the CAD autocompletion task, our multi-modal CadVLM model (b) receives partial CAD entities as both image and text input (a) and generates the remaining sketch entities as output (c). The complete sketch, with optional predicted constraints, can then be used in CAD software (d) to form 3D shapes (e). More description about the primitive values in sketch text are in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.564, + 0.744, + 0.816 + ], + "angle": 0, + "content": "Abstract. Parametric Computer-Aided Design (CAD) is central to contemporary mechanical design. However, it encounters challenges in achieving precise parametric sketch modeling and lacks practical evaluation metrics suitable for mechanical design. We harness the capabilities of pre-trained foundation models, renowned for their successes in natural language processing and computer vision, to develop generative models specifically for CAD. These models are adept at understanding complex geometries and design reasoning, a crucial advancement in CAD technology. In this paper, we propose CadVLM, an end-to-end vision language model for CAD generation. Our approach involves adapting pre-trained foundation models to manipulate engineering sketches effectively, integrating both sketch primitive sequences and sketch images. Extensive experiments demonstrate superior performance on multiple CAD sketch generation tasks such as CAD autocomblection, CAD autoconstraint, and image conditional generation. To our knowledge, this is the first instance of a multimodal Large Language Model (LLM) being successfully applied to parametric CAD generation, representing a pioneering step in the field of computer-aided mechanical design." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.826, + 0.649, + 0.84 + ], + "angle": 0, + "content": "Keywords: CAD Representation \\(\\cdot\\) Vision language model" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "S. Wu et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.375, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.373 + ], + "angle": 0, + "content": "Parametric Computer-Aided Design (CAD) is crucial in mechanical engineering, playing a key role in developing modern objects such as lamps, cars, and buildings. Its importance lies in its ability to accurately depict complex designs, essential for the integrity and functionality of engineered products interacting with humans and the environment. CAD sketches, consisting of basic 2D shapes, are foundational for parametric CAD modeling [3]. As illustrated in Fig [1], these sketches comprise primitive geometric elements (e.g., lines, arcs, points) defined by various relationships (e.g., coincident, parallel, tangent). However, the required precision for 3D construction in parametric CAD can be both challenging and time-consuming [15]. Consequently, there is a significant push to improve CAD generation to simplify and accelerate the manual design process. Leveraging deep learning to identify patterns in CAD sketches could automate the completion of geometry and constraints in engineering drawings [22,23,28]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.79, + 0.613 + ], + "angle": 0, + "content": "Integrating deep learning into CAD offers a significant opportunity to transform the field. Recent research [10,14,19,22,23,29,31] has concentrated on the generative modeling of engineering sketches, particularly focusing on Transformer-based architectures [26]. These approaches often utilize graph structures [22] or attention mechanisms [23] to process sketch information. However, the effectiveness of these methods is limited by the availability of data and a lack of commonsense understanding, which restricts their capacity to perform complex geometric reasoning in engineering sketches. Additionally, rendered sketch images (such as in Fig. 1(a) and 4) have great potential to represent geometry patterns from image modality, which could be complementary for parametric sketches language. While none of current CAD generative models take sketch image into consideration. Our research presents a novel Multimodal Transformer-based Generative model for CAD generation tasks, pioneering in its use of both language and image for sketch modeling. By merging these modalities, we aim to increase the accuracy of CAD generation, setting a new benchmark by simultaneously modeling text-style sketch parametric and image-style sketch renderings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Large Language Models (LLMs) have gained widespread success for their powerful commonsense reasoning and problem-solving abilities in a variety of domains, such as natural language processing, mathematics, and medical sciences [1,5,7,18,24,25,33]. The capabilities of LLMs align well with engineering sketches, which can be understood as sequences representing points and connections. Commonsense knowledge in pre-trained LLM can also offer valuable hints for sketch patterns. Furthermore, combining sketch images with textual descriptions can greatly improve the performance of language models in CAD sketch generative models. In the realm of parametric CAD design, which requires a thorough understanding of complex interactions(e.g. parallel, coincident, symmetry), the advantages of LLMs become even more pronounced." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this work, we leverage the capabilities of pre-trained LLMs to develop a generative model tailored for parametric CAD, addressing the limitations of the existing methods. We introduce CadVLM, a CAD Vision Language Model, featuring a multimodal encoder-decoder framework. CadVLM stands out by pro" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.572, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Abbreviated paper title" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "cessing engineering sketches both linguistically, through token sequences representing sketch geometry and constraints, and visually, through the use of rendered sketch images. To the best of our knowledge, CadVLM is the first generative model to utilize pre-trained vision and language models for CAD sketches, which is capable of addressing different CAD representation learning tasks such as CAD autocompletion, CAD autoconstraint, and image conditional generation tasks. Our contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.261, + 0.786, + 0.318 + ], + "angle": 0, + "content": "- We propose CadVLM, a novel CAD Vision-Language Model, for simultaneously modeling engineering sketches as both language and images. As we know, we are the first to combine visual and textual data for CAD generative models." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.321, + 0.786, + 0.362 + ], + "angle": 0, + "content": "- We introduce three novel evaluation metrics—Entity Accuracy, Sketch Accuracy, and CAD F1 score—to quantitatively assess the quality of generated CAD sketches." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.365, + 0.795, + 0.394 + ], + "angle": 0, + "content": "- CadVLM demonstrates superior performance on the SketchGraphs dataset [22] in both CAD autocompletion and autoconstraint tasks." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.261, + 0.795, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.417, + 0.388, + 0.433 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.613 + ], + "angle": 0, + "content": "Vision Language Pre-Training. The great success of large language models (LLMs) such as BERT [11] and GPT [2] has drawn significant attention within the field of natural language processing and code generation domain. CodeT5+ [27] is a family of open-source encoder-decoder LLMs for code tasks, ranging from 220 million to 16 Billion parameters. Beyond single modality LLMs, multimodal foundation models have also exhibited enhanced performance in various vision-and-language tasks [6,12,20]. Depending on downstream tasks, Various pre-training objectives have also been proposed like image-text contrastive learning [20], image-text matching [12], and (masked) language modeling [12]. Most recently, developing domain-specific multimodality LLMs has gained momentum such as MedBLIP [4] and TableGPT [33]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.809 + ], + "angle": 0, + "content": "CAD Sketch Generation 2D engineering sketches are fundamental in CAD for designing and manufacturing mechanical parts. The recent availability of extensive engineering sketch datasets has facilitated the development of generative models that enhance traditional CAD workflows. The SketchGraphs [22] dataset, featuring 15 million parametric CAD sketches with detailed geometric constraint graphs, showcases primitive interrelations. The paper also introduces a baseline model using autoregressive message-passing networks for constraint graph generation, but it lacks primitive coordinate output, relying solely on constraint graphs and the Onshape solver for sketch configuration, thus limiting its utility in CAD autocomblection tasks. Ganin et al. [8] proposed another large-scale CAD dataset containing over 4.7M parametric sketches from the Onshape public repository. However, limited by sketch format, this dataset has not been widely researched in the realm of CAD generative models." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.843 + ], + "angle": 0, + "content": "Another line of work for generative modeling of CAD sketches is based on Transformer [23,28-30]. Vitruvion [23] models both primitives and constraints," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "S. Wu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.374 + ], + "angle": 0, + "content": "but its training on Transformer does not inherently include geometric knowledge. Furthermore, Vitruvion's major limitation lies in its lack of quantitative comparison in terms of CAD generation quality. While CurveGen [28] tackles the problem of engineering sketch generation without considering sketch constraints like us, it struggles to generalize to few-shot samples. Both Vitruvion and CurveGen face limitations in their model performance, making them challenging to deploy in practical CAD engineering applications. Recently, there has been some research about CAD construction sequences generation [30,31]. However, these works cannot model sketch primitives to autocomplete parametric CAD sketches. We consider Vitruvion, Deepcad, and SketchGraphs as the baselines for CAD Autocompletion task and CAD Autoconstraint task since they are current state-of-the-art models using transformer architectures. Different from existing work, our method utilizes sketch image modality and commonsense knowledge in pre-trained foundation models, providing both sketch sequence textual and image information for sketch generation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.399, + 0.37, + 0.416 + ], + "angle": 0, + "content": "3 Background" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.552 + ], + "angle": 0, + "content": "Engineering Sketch Definition. In the field of parametric CAD, engineering sketches are fundamental, serving as detailed blueprints for designing a range of objects from small components to large structures. These sketches are essential for converting conceptual ideas into precise, manufacturable designs. To ensure a standardized representation, sketches of varying dimensions, spanning from millimeters to meters, are uniformly positioned within a bounding box of 1-meter width at the origin. Each sketch consists of two key sequences: a sequence of primitives, \\( S \\), and a sequence of constraints, \\( C \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.78 + ], + "angle": 0, + "content": "We adopt the normalization and quantization methods for primitive parameters as outlined by [23]. A sketch primitive is expressed as \\( S = (e_1, e_2, \\dots, e_m) \\), where \\( m \\) represents the number of entities in the sketch. Each entity \\( e_i \\) is characterized by positional parameters \\( (p_1, \\dots, p_k) \\), where \\( p \\) represents the normalized coordinates of points that constitute the entity. Sketches typically include three types of entities—lines, arcs, and circles. As illustrated in Figure 11, all sketches are centrally positioned and normalized within a range of [1,64]. Lines, shown in blue, are depicted using two points: start and end point coordinates. Curves, shown in green, are characterized by three points: the start, the end, and a mid-point on the curve. Circles, illustrated in red, are portrayed with four uniformly distributed points on the circle's circumference. This format allows for the interpretation of each sketch as a sequence of tokens. Constraints are similarly represented as primitives in \\( \\mathcal{C} = (c_1, \\dots, c_n) \\), with each constraint entity \\( c \\) being defined by its type and the reference indices of parameters to which the primitives must conform." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "CAD AutoCompletion Task. In addressing the existing CAD AutoCompletion task, which aims to automate routine CAD design procedures, we propose a new autoregressive generative model. This task involves completing a partially given sketch to form a fully realized design. The mathematical formulation of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.573, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Abbreviated paper title" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.148, + 0.784, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.347, + 0.788, + 0.485 + ], + "angle": 0, + "content": "Fig. 2: CadVLM model architecture and training objectives for CAD autoconstraint. We propose a unified vision-language model capable of outputting both autocompleted sketch images and sketch primitives. The inputs are paired with incomplete sketch primitives and images, which are encoded by pre-trained text encoder and image encoder correspondingly. Aligned by image-text contrastive(ITC) loss, the concatenation of image embedding and text embedding then will be input to the text-grounded image decoder and image-grounded text decoder. The image decoder is trained with image decoding loss between autocompleted sketch images and ground-truth sketch images. The text decoder is trained with a language modeling(LM) loss to generate autocompleted primitive." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.526, + 0.299, + 0.539 + ], + "angle": 0, + "content": "the task is:" + }, + { + "type": "equation", + "bbox": [ + 0.328, + 0.54, + 0.785, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\Theta) = - \\log p _ {\\Theta} (\\mathcal {S} | \\mathcal {S} _ {p}) = - \\sum_ {i = j} ^ {m} \\log p _ {\\Theta} \\left(e _ {i: m} \\mid e _ {1: i}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.581, + 0.788, + 0.642 + ], + "angle": 0, + "content": "Here, \\( m \\) is the total number of entities in the complete sketch \\( S \\), and \\( S_{p} = (e_{1},\\dots,e_{i}) \\) is a partial sequence containing \\( 20\\% \\) to \\( 80\\% \\) of the primitive entities from \\( S \\). Our model's goal is to optimize the log-likelihood of completing the sketch \\( S \\) based on these partial sequences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.647, + 0.788, + 0.723 + ], + "angle": 0, + "content": "We employ open-source natural language pre-trained Transformer models for our generative model, with a particular emphasis on T5 models [21]. Additionally, we explore finetuning with GPT-3.5 [2]. Our results indicate significant cross-domain knowledge transfer capabilities in these pre-trained models, showcasing their effectiveness in the context of CAD design." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.727, + 0.788, + 0.771 + ], + "angle": 0, + "content": "CAD Autoconstrain Task. The CAD autoconstrain task aims to autoregressively generate a set of constraints based on given geometry primitives. We model the constraint model as" + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.772, + 0.786, + 0.809 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\phi) = - \\log p _ {\\Phi} (\\mathcal {C} | \\mathcal {S}) = - \\sum_ {i = 1} ^ {n} \\log p _ {\\Phi} \\left(C _ {j: n} \\mid C _ {1: j}, S\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "where \\( n \\) is the number of constraints, and each \\( C_i \\) is represented by a tuple of tokens indicating its constraint type and corresponding primitive entity param" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "S. Wu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "eters. Like autocompletion task, the autoconstrain model is trained to maximize the log-likelihood of generated tokens with respect to the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.784, + 0.221 + ], + "angle": 0, + "content": "Aligned with [23], we focus on 13 constraint types, which are coincident, parallel, and so on. Further details about these constraints are provided in the Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.784, + 0.265 + ], + "angle": 0, + "content": "Evaluation Metrics. To quantitatively assess CAD generation models, we introduce three CAD-specific metrics: Sketch Accuracy, Entity Accuracy, and CAD F1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.785, + 0.325 + ], + "angle": 0, + "content": "Sketch Accuracy measures the probability of correctly generating the entire remaining sketch in the test set. Entity Accuracy evaluates the likelihood of correctly generating at least one entity in the ground truth. They are calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.328, + 0.326, + 0.784, + 0.354 + ], + "angle": 0, + "content": "\\[\n\\text {S k e t c h A c c u r a c y} = \\frac {N _ {s}}{N}, \\quad \\text {E n t i t y A c c u r a c y} = \\frac {N _ {e}}{N}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.784, + 0.42 + ], + "angle": 0, + "content": "where \\( N_{s} \\) is the number of correctly generated sketches in the test set, \\( N_{e} \\) is the number of generated sketches that match at least one entity with the ground truth sketch, and \\( N \\) is the total number of sketches in the test set. CAD F1 is a harmonic mean of precision and recall for each sketch, averaged over the dataset:" + }, + { + "type": "equation", + "bbox": [ + 0.325, + 0.433, + 0.784, + 0.468 + ], + "angle": 0, + "content": "\\[\nC A D F 1 = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {2 \\times \\text {s k e t c h p r e c i s i o n} \\times \\text {s k e t c h r e c a l l}}{(\\text {s k e t c h p r e c i s i o n} + \\text {s k e t c h r e c a l l})}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.784, + 0.519 + ], + "angle": 0, + "content": "where sketch precision \\( = \\frac{n_c}{n_p} \\), sketch recall \\( = \\frac{n_c}{m} \\), \\( n_c \\) is the count of correct entities in the completed sketch, \\( n_p \\) is the total number of entities predicted, and \\( m \\) is the number of entities in the original sketch (ground truth)." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.543, + 0.707, + 0.576 + ], + "angle": 0, + "content": "4 CadVLM: A Vision Language Model for CAD Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.784, + 0.683 + ], + "angle": 0, + "content": "In this section, we delve into the specifics of our CAD Vision Language Model (CadVLM), employing the CAD autocompletion task as an illustrative example. In our approach, the sequence of sketch primitives \\( S \\) is interpreted akin to language text, while the corresponding visual representation of the sketch is treated as image \\( I \\). It should be noted that the text-grounded image decoder is not utilized in the CAD auto-constraint task." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.709, + 0.431, + 0.723 + ], + "angle": 0, + "content": "4.1 Network Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.784, + 0.84 + ], + "angle": 0, + "content": "CadVLM is structured as an asymmetrical encoder-decoder architecture, featuring a two-stream encoder, a text decoder based on a language model, and a lightweight image decoder for constructing the autocompleted image. This design is depicted in Figure 2. The two-stream encoder is tailored to process dual-modal inputs: a vision sub-encoder for the sketch image \\( I \\) and a text sub-encoder for the prefix sequence of sketch primitives \\( S_{p} \\). For text decoding, we employ a causal language transformer model, which autoregressively generates" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.572, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Abbreviated paper title" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.193 + ], + "angle": 0, + "content": "the output, i.e., the autocompleted sketch. Finally, the image decoder utilizes the latent representation tokens to construct the visual representation of the autocompleted sketch." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.195, + 0.788, + 0.346 + ], + "angle": 0, + "content": "Vision Encoding Stream. To encode partial sketch images and harness the capabilities of an existing large vision autoencoder model, we employ a pretrained vision transformer (ViT) [9] for extracting vision features. To bridge the gap between CAD images and natural images typically processed by ViT, we fine-tune the model on the SketchGraphs dataset. This fine-tuning involves a one-epoch image reconstruction task, with a \\(75\\%\\) masking ratio, specifically tailored for CAD sketch data. Figure 3 illustrates the image reconstruction outcomes using both the original pre-trained ViT-MAE and the version fine-tuned on the SketchGraphs dataset. After finetuning, ViT-MAE shows enhanced representational performance for CAD sketches." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.379, + 0.498, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.564, + 0.504, + 0.646 + ], + "angle": 0, + "content": "Fig.3: CAD image reconstruction results by ViTMAE. (a) Input sketches, (b) masked sketches with the mask ratio of \\(75\\%\\), reconstructed results using a (c) pretrained and (d) further fine-tuned ViTMAE." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.349, + 0.789, + 0.515 + ], + "angle": 0, + "content": "However, it's important to note that while some reconstructed sketches are notably accurate, they do not fully meet the requirements for parametric CAD engineering due to the vagueness and lack of parameterization in the entities. To bring image and text features to the same embedding space, we integrate a down-sampling layer and a lightweight projection module atop the vision encoder." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.518, + 0.789, + 0.653 + ], + "angle": 0, + "content": "Text Encoding Stream. Treating the sketch primitive sequence as textual data, we first tokenize this sequence to derive a series of text tokens \\( T = t_{1}, t_{2}, \\ldots, t_{l} \\). These tokens are then processed using the encoder of a pre-trained language model, effectively encoding the sketch information in a textual format." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.656, + 0.788, + 0.686 + ], + "angle": 0, + "content": "Text-grounded Vision Decoding Stream. This stream is dedicated" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.788, + 0.807 + ], + "angle": 0, + "content": "to constructing the autocompleted pixels, mapping the fused image and text representation to the pixel-level sketch image. The input for the vision decoder comprises the fused image and text embeddings, augmented with adaptively masked token patches to keep the same patch size as the input image. This construction process ensures that the vision embedding is grounded by textual information and also enhances the text representation. Note that for CAD autoconstraint task, we don't reconstruct the image since the input sketches are complete sketch images already." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Image-grounded Text Decoding Stream. In alignment with Equations 1 and 2, we conceptualize autocompletion as a conditional text generation process." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "S. Wu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Within this framework, the combined image and text embeddings serve as inputs to the decoder of a language model. This decoder is capable of autoregressively generating the required output. Notably, the parameters of this language model are initialized from a pre-trained model, which itself has undergone extensive pre-training on vast datasets of code [27], enhancing its effectiveness in this context." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.263, + 0.427, + 0.279 + ], + "angle": 0, + "content": "4.2 Training Objectives" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.291, + 0.786, + 0.321 + ], + "angle": 0, + "content": "As shown in Figure 2, we jointly optimize three distinct objectives during training: one focused on understanding and two on generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.322, + 0.788, + 0.473 + ], + "angle": 0, + "content": "Image-Text Contrastive Loss (ITC) is employed to activate the unimodal encoders, following the approach in [12]. This loss function aligns the visual and textual feature spaces by ensuring that positive image-text pairs have similar representations while contrasting with negative pairs. Such a contrastive approach has been instrumental in enhancing multimodal model training [13,20], not only improving the alignment of visual and language features but also accelerating the convergence during training. The image-text contrastive loss is defined as the cross-entropy between the similarity \\( p \\) of image(I) text(T) embedding pairs and ground-truth one-hot similarity \\( y \\) (array from [0,N], N is the number of samples), formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.485, + 0.787, + 0.539 + ], + "angle": 0, + "content": "\\[\np = \\frac {\\exp (s (I , T))}{\\sum_ {i , j} ^ {N} \\exp (s (I _ {i} , T _ {j}))}, \\mathcal {L} _ {I T C} = \\text {C r o s s - E n t r o p y} (y, p). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.55, + 0.788, + 0.643 + ], + "angle": 0, + "content": "Image Decoding Loss (IDL) is responsible for constructing autocompletion sketch images. This process involves translating the fused visual and textual information into pixel values for each patch. IDL employs a mean squared error (MSE) loss, similarly to the approach in ViTMAE [9], comparing the autocompleted sketch images with their ground-truth counterparts at the pixel level. The Image decoding loss can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.414, + 0.655, + 0.787, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I D L} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(Y _ {i} - \\hat {Y} _ {i}\\right) ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.763 + ], + "angle": 0, + "content": "where \\( n \\) is the number of total pixels, \\( Y_{i} \\) is the ground-truth pixel value of \\( i \\)-th pixel, and \\( \\hat{Y}_i \\) is the output pixel value correspondingly. For CAD autoconstraint tasks, since the input is a full sketch image, we ignore IDL during the training of CAD autoconstraint models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Language Modeling Loss (LM) is designed to generate text-like autocompletion sketches, focusing on constraint sequences and utilizing the image-grounded text decoder. It optimizes a cross-entropy loss to maximize the likelihood described in Equation ①. LM endows the model with the capability to transform fused multimodal information into coherent autocompletion sketch" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.572, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Abbreviated paper title" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "sequences. For the CAD autoconstraint task, we use Equation 2 as the language modeling loss." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.178, + 0.618, + 0.193 + ], + "angle": 0, + "content": "Therefore, the full training objective of CadVLM is:" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.208, + 0.786, + 0.223 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {I T C} + \\mathcal {L} _ {L M} + \\mathcal {L} _ {I D L}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.23, + 0.784, + 0.261 + ], + "angle": 0, + "content": "During inference, given incomplete sketch image and primitive pairs, we only evaluate the generated text and use that for further mechanical applications." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.284, + 0.376, + 0.302 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.316, + 0.784, + 0.361 + ], + "angle": 0, + "content": "We conduct extensive experiments and ablation studies on CadVLM for the CAD autocompletion and autoconstraint tasks. Quantitative assessment is conducted by measuring Entity Accuracy, Sketch Accuracy, and CAD F1." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.384, + 0.444, + 0.399 + ], + "angle": 0, + "content": "5.1 Experimental settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.409, + 0.785, + 0.528 + ], + "angle": 0, + "content": "Dataset We use the SketchGraphs [22] dataset for all our experiments, consisting of 15 million CAD sketches collected from Onshape. Referring to [23], we use the filtered version and further deduplicate the dataset. After deduplicated, the training, validation, and test set have 626,236, 22,031, 21,979 unique sketches correspondingly. The filtered sketches are restricted to sketches comprised of arcs, circles, and lines. To ensure a fair comparison, we normalize each sketch as in [23]; specifically the quantized range of numerical tokens in the primitive sequence is [1, 64]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.53, + 0.785, + 0.634 + ], + "angle": 0, + "content": "Baselines We evaluate several ablation and conditional variants of CadVLM, as well as three SOTA CAD generative model baselines: Vitruvion [23], Deepcad [29] and SketchGraphs [22]. Since SketchGraphs does not incorporate continuous primitive parameters such as coordinate value, we only compare with the CAD autoconstraint task. And DeepCAD only considers parametrized command sequences, which are incompatible with CAD constraints. We primarily evaluate its performance on CAD autocompletion tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.786, + 0.817 + ], + "angle": 0, + "content": "Settings For the image encoder and decoder, we choose the encoder and decoder of state-of-the-art pre-trained ViT-MAE-base from [9]. Following [9] we set the patch size and stride both as 32. We render all sketch images with size of \\(224 \\times 224 \\times 3\\). For the text encoder and text decoder, we use the encoder and decoder of pre-trained CodeT5+ [27] correspondingly, which is an instruction-tuned code LLM with 770M parameters trained on various code datasets. We train the autocompletion model and autoconstraint model both for 30 epochs, using the Adam optimizer with decoupled weight decay regularization [17], with the learning rate set according to the CosineAnnealingLR [16] scheduler. The initial learning rate is set to \\(3e - 4\\), with batch size 32. All our experiments are run using 8 NVIDIA A100 GPUs, and our models take between 30-40 hours to train. In the CAD Autocompletion experiments, we randomly sample \\(20\\% -80\\%\\)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.824, + 0.412, + 0.84 + ], + "angle": 0, + "content": "3 https://www.onshape.com" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "S. Wu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.214 + ], + "angle": 0, + "content": "Table 1: Quantitative sketch generative model comparison results for CAD autocompletion task. Ske-Acc means the probability of generating the correct full remaining sketch and Ent-Acc means the probability of generating at least one entity of the ground truth. Bold shows the best results of each metric. For all metrics the higher the better." + }, + { + "type": "table", + "bbox": [ + 0.326, + 0.227, + 0.675, + 0.391 + ], + "angle": 0, + "content": "
ModelSke-AccEnt-AccCAD-F1
Vitruvion3.0%40.7%11.3%
Deepcad1.2%23.1%8.9%
ChatGPT-15%6.8%41.3%21.2%
ChatGPT-7%7.6%43.8%22.2%
ChatGPT-1%6.5%43.7%22.5%
CadVLM-Text21.2%66.7%42.8%
CadVLM-w/o-IDL22.6%68.4%44.2%
CadVLM-w/o-ITC22.8%67.6%43.9%
CadVLM-w/o-IDL&ITC23.0%67.7%44.1%
CadVLM23.8%68.3%45.2%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.402, + 0.787, + 0.431 + ], + "angle": 0, + "content": "Table 2: Quantitative sketch generative model comparison results for CAD autoconstraint task. For all metrics higher is better." + }, + { + "type": "table", + "bbox": [ + 0.345, + 0.443, + 0.656, + 0.538 + ], + "angle": 0, + "content": "
ModelSke-AccEnt-AccCAD-F1
SketchGraphs0.48%93.0%42.0%
Vitruvion10.7%94.6%71.6%
CadVLM-Text11.5%95.0%72.6%
CadVLM-w/o-ITC12.1%95.4%74.5%
CadVLM11.3%94.9%73.8%
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.787, + 0.598 + ], + "angle": 0, + "content": "of sketch entities as input and aim to predict the remaining entities from the ground truth sketch." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.62, + 0.787, + 0.636 + ], + "angle": 0, + "content": "5.2 Quantitative comparison with baseline sketch generative models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Table 1 reports results on the CAD Autocompletion task. We find that Vitruvion achieves \\(3\\%\\) on Sketch Accuracy and \\(40.7\\%\\) on Entity Accuracy, meaning it can accurately predict \\(3\\%\\) of sketches and correctly identifies at least one entity in \\(40.7\\%\\) of sketches. Deepcad performs worst since it only has 4 layers of encoder and decoder, makes it hard to represent complex CAD sketches. Our CadVLM outperforms Virruvion on Sketch Accuracy, Entity Accuracy, and CAD-F1 by \\(20.8\\%\\), \\(28.3\\%\\), and \\(39.9\\%\\) respectively. These results demonstrate the effectiveness of pre-trained foundational models and image modality. A closer examination of the qualitative results in Figure 4 reveals that Vitruvion produces many invalid results missing symmetric curves or lines. Sketches from CadVLM are better in terms of quality with more complex shapes and stronger symmetry." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "ChatGPT Series Model Finetune To evaluate the ability of a general LLM on the CAD autocompletion task, we fine-tuned ChatGPT on three subsets" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.572, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Abbreviated paper title" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.784, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.306, + 0.788, + 0.361 + ], + "angle": 0, + "content": "Fig. 4: Comparative Analysis of Autocompletion in CAD Design. Top row: Random samples from the SketchGraphs test dataset. Second row: Initial input entities serving as the primer for the autocompletion task. Third and fourth row: Autocompletion results produced by Vitruvion and CadVLM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.789, + 0.487 + ], + "angle": 0, + "content": "of \\(15\\%\\), \\(7\\%\\), and \\(1\\%\\) of the SketchGraphs dataset. The results indicate that the ChatGPT-series outperforms Vitruvion on all three metrics, especially on Sketch Accuracy and CAD-F1. However, there were no significant performance improvements when finetuning on larger portions of the training dataset. One possible reason is we only finetune ChatGPT for less than 5 epochs due to training costs, making it hard for ChatGPT to adapt to CAD data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.791, + 0.653 + ], + "angle": 0, + "content": "CadVLM variants To better understand the source of performance gains, we conducted a comparative analysis of different variants of our proposed CadVLM: CadVLM-text, CadVLM-w/o-ITC, CadVLM-w/o-IDL&ITC, and CadVLM w/o-IDL. CadVLM-text relies solely on the sketch primitive sequence text as its input, optimizing for language modeling loss. In contrast, CadVLM-w/o-IDL utilizes ITC and LM loss for its training objectives, CadVLM-w/o-ITC focuses on IDL and LM, and CadVLM-w/o-IDL&ITC prioritizes only the LM objective. The findings reveal that CadVLM-Text underperforms compared to all other CadVLM variants across all three metrics, which underlines the value of the sketch image modality in providing diverse geometric information alongside sequence text." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.657, + 0.789, + 0.732 + ], + "angle": 0, + "content": "Interestingly, CadVLM-w/o-IDL emerges as the superior performer among the CadVLM variants, highlighting the significance of aligning image and text embeddings in multimodal modeling. Incorporating Image Decoding Loss (IDL) further enhances the alignment between image and text embeddings, leading CadVLM to show a notable improvement in all three metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.84 + ], + "angle": 0, + "content": "For the CAD Autoconstraint task, as shown in Table 2, SketchGraphs struggles to predict the whole sketch correctly, with only \\(0.48\\%\\) Sketch Accuracy. This suggests that relying solely on graph structures of sketches is insufficient for handling long sequences of sketch primitives. In contrast, Vitruvion achieves \\(10.7\\%\\) on Sketch Accuracy, showing the necessity to model primitives. For the CAD Autoconstraint task it is less challenging to predict at least one constraint correctly but more challenging to accurately predict all constraints, easily caused" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "S. Wu et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.404, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.302, + 0.26, + 0.322, + 0.274 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.145, + 0.596, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.491, + 0.26, + 0.513, + 0.275 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.145, + 0.785, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.261, + 0.702, + 0.275 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.285, + 0.789, + 0.314 + ], + "angle": 0, + "content": "Fig. 5: Effect of input entity ratio to (a) Entity Accuracy, (b) Sketch Accuracy, and (c) CAD F1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.784, + 0.385 + ], + "angle": 0, + "content": "by potential data labeling inconsistencies. As is well-known, minor discrepancies in primitive values can drastically alter constraints, such as changing them from parallel to coincident." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.785, + 0.522 + ], + "angle": 0, + "content": "The CadVLM series model exceeds the performance of both SketchGraphs and Vitruvion across all metrics. CadVLM-Text surpasses CadVLM on Ske-Acc and Ent-Acc since the pretraining task of ViTMAE has a great discrepancy with CAD autoconstraint tasks. Therefore aligning the image and text representation directly might not be an optimal way for CAD autoconstraint task.Notably, CadVLM-w/o-ITC emerges as the top performer within the CadVLM series. This may be attributed to its more substantial text decoder and utilization of multimodal information, which enables the model to more effectively integrate knowledge from both primitives and images to inform geometric information." + }, + { + "type": "table_caption", + "bbox": [ + 0.239, + 0.532, + 0.788, + 0.575 + ], + "angle": 0, + "content": "Table 3: Quantitative results for image condi- Table 4: Effect of adding the inductioned primitive generation. Bold represents tive bias of entity-level embeddings the best results of all metrics. to the sketch sequence." + }, + { + "type": "table", + "bbox": [ + 0.253, + 0.585, + 0.53, + 0.707 + ], + "angle": 0, + "content": "
Conditioned -ImageModelEnt-AccCAD-F1
PreciseVitruvion83.8%35.8%
CadVLM98.0%62.7%
Hand-drawnVitruvion84.2%38.4%
CadVLM99.4%60.1%
Noisy hand-drawnVitruvion83.2%35.8%
CadVLM94.4%56.5%
" + }, + { + "type": "table", + "bbox": [ + 0.553, + 0.585, + 0.774, + 0.683 + ], + "angle": 0, + "content": "
ModelCadVLM w/-Inductive -TextBias
Ent-Acc66.7%67.5%
Ske-Acc21.2%22.2%
CAD-F142.8%43.7%
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.727, + 0.391, + 0.743 + ], + "angle": 0, + "content": "5.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.794 + ], + "angle": 0, + "content": "We conducted several ablation studies to assess the performance of CadVLM on image-conditioned primitive generation tasks and investigate the effectiveness of each component within CadVLM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Image Conditional Generation. Following [23], we conduct a series of experiments to evaluate the effectiveness of CadVLM modeling primitive sequences conditioned on the images of sketches. As designers may begin their" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.572, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Abbreviated paper title" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.193 + ], + "angle": 0, + "content": "designs with hand-drawn sketches before transitioning to CAD programs, accurately inferring parametric CAD from images or scans can dramatically reduce the effort of translating from rough paper sketches to the CAD model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.195, + 0.787, + 0.256 + ], + "angle": 0, + "content": "When conditioning only on sketch images, the text encoder of CadVLM is ignored. We only use the image encoder, image decoder, and text decoder. The overall model is trained to maximize the probability of the ground truth primitives portrayed in the image, conditioned on a sequence of patch embeddings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.258, + 0.788, + 0.41 + ], + "angle": 0, + "content": "The results of Table 3 show three kinds of sketch image rendering input to imitate different settings in the real world: precise renderings, rendering from the hand-drawn simulator of [23], and renderings with random affine augmentations of the hand-drawn simulator. As both Vitruvion and CadVLM can seldom predict the sketch exactly correctly, we ignore Sketch Accuracy in our image conditional generation experiments. The results show that CadVLM outperforms Vitruvion on both Entity Accuracy and CAD F1. Vitruvion performs best on hand-drawn images with \\(83.8\\%\\) Entity Accuracy and \\(38.4\\%\\) CAD F1. While CadVLM achieves almost \\(99.4\\%\\) Entity Accuracy and \\(62.7\\%\\) CAD F1. Conditioning on precise rendered images achieves the best performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.788, + 0.593 + ], + "angle": 0, + "content": "Entity Level Modelling. As described earlier, each CAD sketch is comprised of multiple entities. We considered adding this inductive bias to the model by processing each entity separately and then combining their embeddings with the initial sequence of all sketch tokens. For this purpose, we first pretend a special token \\(\\langle \\text{ENTITY} \\rangle\\) to each of the entity input sequences. Subsequently, the entities from all the sketches in the batch are fed into the model's encoder in parallel. Next, the embeddings of each entity are gathered for each sketch. For each sketch, we concatenate these embeddings with the original sequence tokens using another special token \\(\\langle \\text{TOKEN} \\rangle\\) as a delimiter for the second part of the sequence. The decoder remains unchanged. As shown in Table 4, incorporating this inductive bias increases computation and memory demands while only marginally enhancing the performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.748 + ], + "angle": 0, + "content": "Impact of different input entity ratios for CAD autocomblection. To explore the impact of different input entity ratios for CAD autocomblection, we further test CAD autocomblection results with different input entity ratios: \\(20\\%\\), \\(40\\%\\), \\(60\\%\\), and \\(80\\%\\). Compared with the experiments in Sec 5.4, we specify the sampled ratio of input entities. As shown in Figure 5, CadVLM outperforms Vitruvion on all input entity ratios. In particular, CadVLM achieves \\(73.8\\%\\) Entity Accuracy, \\(49.5\\%\\) Sketch Accuracy, and \\(62.7\\%\\) CAD F1 when given \\(80\\%\\) of entities as input. For the hardest setting when given only \\(20\\%\\) of entities as input, CadVLM can still achieve \\(43.9\\%\\) Entity Accuracy and \\(16.9\\%\\) CAD-F1, which proves that CadVLM has great potential to be used in real world settings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Effect of using different language models for text encoder and decoder. To better compare the performance of different text encoder-decoder, we experiment using different pre-trained foundation models as text encoder and decoder. To eliminate the influence of the image encoder and decoder, we only consider the CadVLM-Text variant. For the choice of text encoder and decoder, we also compare the performance of different Encoder-Decoder pre-trained" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "S. Wu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 5: Effect of using different architectures for CadVLM-Text. CodeT5+ 770M is the encoder-decoder for text we use in our main experiments. CodeT5+\\* is the result of training from scratch rather than pre-trained." + }, + { + "type": "table", + "bbox": [ + 0.328, + 0.199, + 0.672, + 0.32 + ], + "angle": 0, + "content": "
Model#ParameterEnt-AccSke-AccCAD-F1
CodeT5+220M66.5%21.5%42.9%
CodeT5+*220M65.0%18.9%40.3%
CodeT5+770M66.7%21.2%42.8%
ByT5-base580M66.9%22.0%43.1%
ByT5-large1.2B66.8%20.6%41.9%
T5-base223M66.2%20.2%41.7%
T5-large783M66.8%21.2%42.6%
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.788, + 0.488 + ], + "angle": 0, + "content": "language models on CAD autocombpletion. The T5 family of models are among the most recognized and widely used encoder-decoder models in the open-source community. Therefore, we experimented with T5 [21], ByT5 [32] and CodeT5+ [27]. The results are shown in Table 5. The effect of model size is clear for T5 and CodeT5+ but not ByT5. Also, The cost-benefit of solely making the model bigger becomes less favorable over time as the improvements in performance start to plateau. Another observation is that, because of the nature of the data which is closer to a code sequence than natural language, CodeT5+, which is fine-tuned on code, exhibits better performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.579 + ], + "angle": 0, + "content": "We also trained CodeT5+ from scratch to examine the impact of pre-training on natural language and code. As evidenced in Table5, training from scratch falls short in all three metrics. Additionally, training from scratch converged much slower than finetuning. Despite the fact that our representation of sketches as code is relatively unique and not found in natural language, the rules and token embeddings learned from text are very helpful for domain adaptation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.601, + 0.36, + 0.618 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.755 + ], + "angle": 0, + "content": "In this study, we introduce CadVLM, an innovative end-to-end vision-language model tailored for modeling parametric CAD sketches, which are pivotal in modern mechanical design. CadVLM harnesses the commonsense knowledge and reasoning abilities inherent in large language models, enabling it to adeptly handle the complex geometric reasoning required for CAD sketches. Our experimental results demonstrate that CadVLM outperforms existing state-of-the-art models in both autocompletion and autoconstraint tasks, highlighting the potent enhancement that pre-trained models bring to CAD construction." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.778, + 0.405, + 0.796 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "This work is supported by the Canada CIFAR AI Chair Program and the Canada NSERC Discovery Grant (RGPIN-2021-03115)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.573, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Abbreviated paper title" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.787, + 0.215 + ], + "angle": 0, + "content": "1. Azerbayev, Z., Schoelkopf, H., Paster, K., Santos, M.D., McAleer, S., Jiang, A.Q., Deng, J., Biderman, S., Welleck, S.: Llemma: An open language model for mathematics. arXiv preprint arXiv:2310.10631 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.215, + 0.787, + 0.257 + ], + "angle": 0, + "content": "2. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.257, + 0.787, + 0.283 + ], + "angle": 0, + "content": "3. Camba, J.D., Contero, M., Company, P.: Parametric cad modeling: An analysis of strategies for design reusability. Computer-Aided Design 74, 18-31 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.283, + 0.787, + 0.323 + ], + "angle": 0, + "content": "4. Chen, Q., Hu, X., Wang, Z., Hong, Y.: Medblip: Bootstrapping language-image pre-training from 3d medical images and texts. arXiv preprint arXiv:2305.10799 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.787, + 0.365 + ], + "angle": 0, + "content": "5. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., et al.: Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.365, + 0.787, + 0.392 + ], + "angle": 0, + "content": "6. Chen, Y., Zhang, S., Han, B., He, T., Li, B.: Camml: Context-aware multimodal learner for large models. arXiv preprint arXiv:2401.03149 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.392, + 0.787, + 0.418 + ], + "angle": 0, + "content": "7. Floridi, L., Chiriatti, M.: Gpt-3: Its nature, scope, limits, and consequences. Minds and Machines 30, 681-694 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.419, + 0.787, + 0.446 + ], + "angle": 0, + "content": "8. Ganin, Y., Bartunov, S., Li, Y., Keller, E., Saliceti, S.: Computer-aided design as language (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.787, + 0.487 + ], + "angle": 0, + "content": "9. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.487, + 0.787, + 0.528 + ], + "angle": 0, + "content": "0. Jones, B.T., Hu, M., Kodnongbua, M., Kim, V.G., Schulz, A.: Self-supervised representation learning for cad. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21327-21336 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.528, + 0.787, + 0.568 + ], + "angle": 0, + "content": "1. Kenton, J.D.M.W.C., Toutanova, L.K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of naacL-HLT. vol. 1, p. 2 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.568, + 0.787, + 0.609 + ], + "angle": 0, + "content": "2. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.609, + 0.787, + 0.65 + ], + "angle": 0, + "content": "3. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems 34, 9694-9705 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.65, + 0.787, + 0.691 + ], + "angle": 0, + "content": "4. Li, P., Guo, J., Zhang, X., Yan, D.M.: Secad-net: Self-supervised cad reconstruction by learning sketch-extrude operations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16826 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.787, + 0.746 + ], + "angle": 0, + "content": "5. Lindsay, A., Paterson, A., Graham, I.: Identifying and quantifying inefficiencies within industrial parametric cad models. In: Advances in Manufacturing Technology XXXII: Proceedings of the 16th International Conference on Manufacturing Research. vol. 8, p. 227 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.746, + 0.787, + 0.772 + ], + "angle": 0, + "content": "6. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.787, + 0.8 + ], + "angle": 0, + "content": "7. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "8. Luo, H., Sun, Q., Xu, C., Zhao, P., Lou, J., Tao, C., Geng, X., Lin, Q., Chen, S., Zhang, D.: Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct. arXiv preprint arXiv:2308.09583 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "S. Wu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "19. Para, W., Bhat, S., Guerrero, P., Kelly, T., Mitra, N., Guibas, L.J., Wonka, P.: Sketchgen: Generating constrained cad sketches. Advances in Neural Information Processing Systems 34, 5077-5088 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.244 + ], + "angle": 0, + "content": "20. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.286 + ], + "angle": 0, + "content": "21. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "22. Seff, A., Ovadia, Y., Zhou, W., Adams, R.P.: Sketchgraphs: A large-scale dataset for modeling relational geometry in computer-aided design. arXiv preprint arXiv:2007.08506 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.355 + ], + "angle": 0, + "content": "23. Seff, A., Zhou, W., Richardson, N., Adams, R.P.: Vitruvion: A generative model of parametric cad sketches. arXiv preprint arXiv:2109.14124 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.397 + ], + "angle": 0, + "content": "24. Singhal, K., Azizi, S., Tu, T., Mahdavi, S.S., Wei, J., Chung, H.W., Scales, N., Tanwani, A., Cole-Lewis, H., Pfohl, S., et al.: Large language models encode clinical knowledge. arXiv preprint arXiv:2212.13138 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "25. Singhal, K., Tu, T., Gottweis, J., Sayres, R., Wulczyn, E., Hou, L., Clark, K., Pfohl, S., Cole-Lewis, H., Neal, D., et al.: Towards expert-level medical question answering with large language models. arXiv preprint arXiv:2305.09617 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "26. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "27. Wang, Y., Le, H., Gotmare, A.D., Bui, N.D., Li, J., Hoi, S.C.: Codet5+: Open code large language models for code understanding and generation. arXiv preprint arXiv:2305.07922 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "28. Willis, K.D., Jayaraman, P.K., Lambourne, J.G., Chu, H., Pu, Y.: Engineering sketch generation for computer-aided design. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2105–2114 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.605 + ], + "angle": 0, + "content": "29. Wu, R., Xiao, C., Zheng, C.: Deepcad: A deep generative network for computer-aided design models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6772-6782 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "30. Xu, X., Jayaraman, P.K., Lambourne, J.G., Willis, K.D., Furukawa, Y.: Hierarchical neural coding for controllable cad model generation. arXiv preprint arXiv:2307.00149 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.646, + 0.785, + 0.702 + ], + "angle": 0, + "content": "31. Xu, X., Willis, K.D., Lambourne, J.G., Cheng, C.Y., Jayaraman, P.K., Furukawa, Y.: Skexgen: Autoregressive generation of cad construction sequences with disentangled codebooks. In: International Conference on Machine Learning. pp. 24698-24724. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.743 + ], + "angle": 0, + "content": "32. Xue, L., Barua, A., Constant, N., Al-Rfou, R., Narang, S., Kale, M., Roberts, A., Raffel, C.: Byt5: Towards a token-free future with pre-trained byte-to-byte models. Transactions of the Association for Computational Linguistics 10, 291-306 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.785 + ], + "angle": 0, + "content": "33. Zha, L., Zhou, J., Li, L., Wang, R., Huang, Q., Yang, S., Yuan, J., Su, C., Li, X., Su, A., et al.: Tablegpt: Towards unifying tables, nature language and commands into one gpt. arXiv preprint arXiv:2307.08674 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.785 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CadVLM_ Bridging Language and Vision in the Generation of Parametric CAD Sketches/7ca1dca8-97a9-4690-a292-e4d8a7eacc0d_origin.pdf b/2024/CadVLM_ Bridging Language and Vision in the Generation of Parametric CAD Sketches/7ca1dca8-97a9-4690-a292-e4d8a7eacc0d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ab32e8ef06a2ed147ec43372d618d668421e82e5 --- /dev/null +++ b/2024/CadVLM_ Bridging Language and Vision in the Generation of Parametric CAD Sketches/7ca1dca8-97a9-4690-a292-e4d8a7eacc0d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83684935c81736ea17e29bf7f1cd260a401cbc69759cdd4f7f84b7f660f2eb0b +size 3008402 diff --git a/2024/CadVLM_ Bridging Language and Vision in the Generation of Parametric CAD Sketches/full.md b/2024/CadVLM_ Bridging Language and Vision in the Generation of Parametric CAD Sketches/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ac5fffbd182e408f1a707f54f33339385b90b48a --- /dev/null +++ b/2024/CadVLM_ Bridging Language and Vision in the Generation of Parametric CAD Sketches/full.md @@ -0,0 +1,275 @@ +# CadVLM: Bridging Language and Vision in the Generation of Parametric CAD Sketches + +Sifan Wu1, Amir Hosein Khasahmadi2, Mor Katz2, Pradeep Kumar Jayaraman2, Yewen Pu2, Karl Willis2, and Bang Liu1 + +1 Université de Montréal & MILA, Montreal, Canada +2 Autodesk AI Lab, Toronto, Canada + +![](images/b76ecdde05298a4355611c5a2772e7e571b5f9ab6b6028dd1b2470845351ffb0.jpg) +Fig. 1: For the CAD autocompletion task, our multi-modal CadVLM model (b) receives partial CAD entities as both image and text input (a) and generates the remaining sketch entities as output (c). The complete sketch, with optional predicted constraints, can then be used in CAD software (d) to form 3D shapes (e). More description about the primitive values in sketch text are in the Appendix. + +Abstract. Parametric Computer-Aided Design (CAD) is central to contemporary mechanical design. However, it encounters challenges in achieving precise parametric sketch modeling and lacks practical evaluation metrics suitable for mechanical design. We harness the capabilities of pre-trained foundation models, renowned for their successes in natural language processing and computer vision, to develop generative models specifically for CAD. These models are adept at understanding complex geometries and design reasoning, a crucial advancement in CAD technology. In this paper, we propose CadVLM, an end-to-end vision language model for CAD generation. Our approach involves adapting pre-trained foundation models to manipulate engineering sketches effectively, integrating both sketch primitive sequences and sketch images. Extensive experiments demonstrate superior performance on multiple CAD sketch generation tasks such as CAD autocomblection, CAD autoconstraint, and image conditional generation. To our knowledge, this is the first instance of a multimodal Large Language Model (LLM) being successfully applied to parametric CAD generation, representing a pioneering step in the field of computer-aided mechanical design. + +Keywords: CAD Representation $\cdot$ Vision language model + +# 1 Introduction + +Parametric Computer-Aided Design (CAD) is crucial in mechanical engineering, playing a key role in developing modern objects such as lamps, cars, and buildings. Its importance lies in its ability to accurately depict complex designs, essential for the integrity and functionality of engineered products interacting with humans and the environment. CAD sketches, consisting of basic 2D shapes, are foundational for parametric CAD modeling [3]. As illustrated in Fig [1], these sketches comprise primitive geometric elements (e.g., lines, arcs, points) defined by various relationships (e.g., coincident, parallel, tangent). However, the required precision for 3D construction in parametric CAD can be both challenging and time-consuming [15]. Consequently, there is a significant push to improve CAD generation to simplify and accelerate the manual design process. Leveraging deep learning to identify patterns in CAD sketches could automate the completion of geometry and constraints in engineering drawings [22,23,28]. + +Integrating deep learning into CAD offers a significant opportunity to transform the field. Recent research [10,14,19,22,23,29,31] has concentrated on the generative modeling of engineering sketches, particularly focusing on Transformer-based architectures [26]. These approaches often utilize graph structures [22] or attention mechanisms [23] to process sketch information. However, the effectiveness of these methods is limited by the availability of data and a lack of commonsense understanding, which restricts their capacity to perform complex geometric reasoning in engineering sketches. Additionally, rendered sketch images (such as in Fig. 1(a) and 4) have great potential to represent geometry patterns from image modality, which could be complementary for parametric sketches language. While none of current CAD generative models take sketch image into consideration. Our research presents a novel Multimodal Transformer-based Generative model for CAD generation tasks, pioneering in its use of both language and image for sketch modeling. By merging these modalities, we aim to increase the accuracy of CAD generation, setting a new benchmark by simultaneously modeling text-style sketch parametric and image-style sketch renderings. + +Large Language Models (LLMs) have gained widespread success for their powerful commonsense reasoning and problem-solving abilities in a variety of domains, such as natural language processing, mathematics, and medical sciences [1,5,7,18,24,25,33]. The capabilities of LLMs align well with engineering sketches, which can be understood as sequences representing points and connections. Commonsense knowledge in pre-trained LLM can also offer valuable hints for sketch patterns. Furthermore, combining sketch images with textual descriptions can greatly improve the performance of language models in CAD sketch generative models. In the realm of parametric CAD design, which requires a thorough understanding of complex interactions(e.g. parallel, coincident, symmetry), the advantages of LLMs become even more pronounced. + +In this work, we leverage the capabilities of pre-trained LLMs to develop a generative model tailored for parametric CAD, addressing the limitations of the existing methods. We introduce CadVLM, a CAD Vision Language Model, featuring a multimodal encoder-decoder framework. CadVLM stands out by pro + +cessing engineering sketches both linguistically, through token sequences representing sketch geometry and constraints, and visually, through the use of rendered sketch images. To the best of our knowledge, CadVLM is the first generative model to utilize pre-trained vision and language models for CAD sketches, which is capable of addressing different CAD representation learning tasks such as CAD autocompletion, CAD autoconstraint, and image conditional generation tasks. Our contributions are threefold: + +- We propose CadVLM, a novel CAD Vision-Language Model, for simultaneously modeling engineering sketches as both language and images. As we know, we are the first to combine visual and textual data for CAD generative models. +- We introduce three novel evaluation metrics—Entity Accuracy, Sketch Accuracy, and CAD F1 score—to quantitatively assess the quality of generated CAD sketches. +- CadVLM demonstrates superior performance on the SketchGraphs dataset [22] in both CAD autocompletion and autoconstraint tasks. + +# 2 Related Work + +Vision Language Pre-Training. The great success of large language models (LLMs) such as BERT [11] and GPT [2] has drawn significant attention within the field of natural language processing and code generation domain. CodeT5+ [27] is a family of open-source encoder-decoder LLMs for code tasks, ranging from 220 million to 16 Billion parameters. Beyond single modality LLMs, multimodal foundation models have also exhibited enhanced performance in various vision-and-language tasks [6,12,20]. Depending on downstream tasks, Various pre-training objectives have also been proposed like image-text contrastive learning [20], image-text matching [12], and (masked) language modeling [12]. Most recently, developing domain-specific multimodality LLMs has gained momentum such as MedBLIP [4] and TableGPT [33]. + +CAD Sketch Generation 2D engineering sketches are fundamental in CAD for designing and manufacturing mechanical parts. The recent availability of extensive engineering sketch datasets has facilitated the development of generative models that enhance traditional CAD workflows. The SketchGraphs [22] dataset, featuring 15 million parametric CAD sketches with detailed geometric constraint graphs, showcases primitive interrelations. The paper also introduces a baseline model using autoregressive message-passing networks for constraint graph generation, but it lacks primitive coordinate output, relying solely on constraint graphs and the Onshape solver for sketch configuration, thus limiting its utility in CAD autocomblection tasks. Ganin et al. [8] proposed another large-scale CAD dataset containing over 4.7M parametric sketches from the Onshape public repository. However, limited by sketch format, this dataset has not been widely researched in the realm of CAD generative models. + +Another line of work for generative modeling of CAD sketches is based on Transformer [23,28-30]. Vitruvion [23] models both primitives and constraints, + +but its training on Transformer does not inherently include geometric knowledge. Furthermore, Vitruvion's major limitation lies in its lack of quantitative comparison in terms of CAD generation quality. While CurveGen [28] tackles the problem of engineering sketch generation without considering sketch constraints like us, it struggles to generalize to few-shot samples. Both Vitruvion and CurveGen face limitations in their model performance, making them challenging to deploy in practical CAD engineering applications. Recently, there has been some research about CAD construction sequences generation [30,31]. However, these works cannot model sketch primitives to autocomplete parametric CAD sketches. We consider Vitruvion, Deepcad, and SketchGraphs as the baselines for CAD Autocompletion task and CAD Autoconstraint task since they are current state-of-the-art models using transformer architectures. Different from existing work, our method utilizes sketch image modality and commonsense knowledge in pre-trained foundation models, providing both sketch sequence textual and image information for sketch generation. + +# 3 Background + +Engineering Sketch Definition. In the field of parametric CAD, engineering sketches are fundamental, serving as detailed blueprints for designing a range of objects from small components to large structures. These sketches are essential for converting conceptual ideas into precise, manufacturable designs. To ensure a standardized representation, sketches of varying dimensions, spanning from millimeters to meters, are uniformly positioned within a bounding box of 1-meter width at the origin. Each sketch consists of two key sequences: a sequence of primitives, $S$ , and a sequence of constraints, $C$ . + +We adopt the normalization and quantization methods for primitive parameters as outlined by [23]. A sketch primitive is expressed as $S = (e_1, e_2, \dots, e_m)$ , where $m$ represents the number of entities in the sketch. Each entity $e_i$ is characterized by positional parameters $(p_1, \dots, p_k)$ , where $p$ represents the normalized coordinates of points that constitute the entity. Sketches typically include three types of entities—lines, arcs, and circles. As illustrated in Figure 11, all sketches are centrally positioned and normalized within a range of [1,64]. Lines, shown in blue, are depicted using two points: start and end point coordinates. Curves, shown in green, are characterized by three points: the start, the end, and a mid-point on the curve. Circles, illustrated in red, are portrayed with four uniformly distributed points on the circle's circumference. This format allows for the interpretation of each sketch as a sequence of tokens. Constraints are similarly represented as primitives in $\mathcal{C} = (c_1, \dots, c_n)$ , with each constraint entity $c$ being defined by its type and the reference indices of parameters to which the primitives must conform. + +CAD AutoCompletion Task. In addressing the existing CAD AutoCompletion task, which aims to automate routine CAD design procedures, we propose a new autoregressive generative model. This task involves completing a partially given sketch to form a fully realized design. The mathematical formulation of + +![](images/fb06fae5aa28f00a34061e4f935404b39b30c03ba9aad4b5e3105e1e54fda047.jpg) +Fig. 2: CadVLM model architecture and training objectives for CAD autoconstraint. We propose a unified vision-language model capable of outputting both autocompleted sketch images and sketch primitives. The inputs are paired with incomplete sketch primitives and images, which are encoded by pre-trained text encoder and image encoder correspondingly. Aligned by image-text contrastive(ITC) loss, the concatenation of image embedding and text embedding then will be input to the text-grounded image decoder and image-grounded text decoder. The image decoder is trained with image decoding loss between autocompleted sketch images and ground-truth sketch images. The text decoder is trained with a language modeling(LM) loss to generate autocompleted primitive. + +the task is: + +$$ +\mathcal {L} (\Theta) = - \log p _ {\Theta} (\mathcal {S} | \mathcal {S} _ {p}) = - \sum_ {i = j} ^ {m} \log p _ {\Theta} \left(e _ {i: m} \mid e _ {1: i}\right). \tag {1} +$$ + +Here, $m$ is the total number of entities in the complete sketch $S$ , and $S_{p} = (e_{1},\dots,e_{i})$ is a partial sequence containing $20\%$ to $80\%$ of the primitive entities from $S$ . Our model's goal is to optimize the log-likelihood of completing the sketch $S$ based on these partial sequences. + +We employ open-source natural language pre-trained Transformer models for our generative model, with a particular emphasis on T5 models [21]. Additionally, we explore finetuning with GPT-3.5 [2]. Our results indicate significant cross-domain knowledge transfer capabilities in these pre-trained models, showcasing their effectiveness in the context of CAD design. + +CAD Autoconstrain Task. The CAD autoconstrain task aims to autoregressively generate a set of constraints based on given geometry primitives. We model the constraint model as + +$$ +\mathcal {L} (\phi) = - \log p _ {\Phi} (\mathcal {C} | \mathcal {S}) = - \sum_ {i = 1} ^ {n} \log p _ {\Phi} \left(C _ {j: n} \mid C _ {1: j}, S\right), \tag {2} +$$ + +where $n$ is the number of constraints, and each $C_i$ is represented by a tuple of tokens indicating its constraint type and corresponding primitive entity param + +eters. Like autocompletion task, the autoconstrain model is trained to maximize the log-likelihood of generated tokens with respect to the ground truth. + +Aligned with [23], we focus on 13 constraint types, which are coincident, parallel, and so on. Further details about these constraints are provided in the Supplementary. + +Evaluation Metrics. To quantitatively assess CAD generation models, we introduce three CAD-specific metrics: Sketch Accuracy, Entity Accuracy, and CAD F1. + +Sketch Accuracy measures the probability of correctly generating the entire remaining sketch in the test set. Entity Accuracy evaluates the likelihood of correctly generating at least one entity in the ground truth. They are calculated as: + +$$ +\text {S k e t c h A c c u r a c y} = \frac {N _ {s}}{N}, \quad \text {E n t i t y A c c u r a c y} = \frac {N _ {e}}{N}, \tag {3} +$$ + +where $N_{s}$ is the number of correctly generated sketches in the test set, $N_{e}$ is the number of generated sketches that match at least one entity with the ground truth sketch, and $N$ is the total number of sketches in the test set. CAD F1 is a harmonic mean of precision and recall for each sketch, averaged over the dataset: + +$$ +C A D F 1 = \frac {1}{N} \sum_ {i = 1} ^ {N} \frac {2 \times \text {s k e t c h p r e c i s i o n} \times \text {s k e t c h r e c a l l}}{(\text {s k e t c h p r e c i s i o n} + \text {s k e t c h r e c a l l})}, \tag {4} +$$ + +where sketch precision $= \frac{n_c}{n_p}$ , sketch recall $= \frac{n_c}{m}$ , $n_c$ is the count of correct entities in the completed sketch, $n_p$ is the total number of entities predicted, and $m$ is the number of entities in the original sketch (ground truth). + +# 4 CadVLM: A Vision Language Model for CAD Generation + +In this section, we delve into the specifics of our CAD Vision Language Model (CadVLM), employing the CAD autocompletion task as an illustrative example. In our approach, the sequence of sketch primitives $S$ is interpreted akin to language text, while the corresponding visual representation of the sketch is treated as image $I$ . It should be noted that the text-grounded image decoder is not utilized in the CAD auto-constraint task. + +# 4.1 Network Framework + +CadVLM is structured as an asymmetrical encoder-decoder architecture, featuring a two-stream encoder, a text decoder based on a language model, and a lightweight image decoder for constructing the autocompleted image. This design is depicted in Figure 2. The two-stream encoder is tailored to process dual-modal inputs: a vision sub-encoder for the sketch image $I$ and a text sub-encoder for the prefix sequence of sketch primitives $S_{p}$ . For text decoding, we employ a causal language transformer model, which autoregressively generates + +the output, i.e., the autocompleted sketch. Finally, the image decoder utilizes the latent representation tokens to construct the visual representation of the autocompleted sketch. + +Vision Encoding Stream. To encode partial sketch images and harness the capabilities of an existing large vision autoencoder model, we employ a pretrained vision transformer (ViT) [9] for extracting vision features. To bridge the gap between CAD images and natural images typically processed by ViT, we fine-tune the model on the SketchGraphs dataset. This fine-tuning involves a one-epoch image reconstruction task, with a $75\%$ masking ratio, specifically tailored for CAD sketch data. Figure 3 illustrates the image reconstruction outcomes using both the original pre-trained ViT-MAE and the version fine-tuned on the SketchGraphs dataset. After finetuning, ViT-MAE shows enhanced representational performance for CAD sketches. + +![](images/e783fb245e3cd97204a1943891292cce70129cf25e18c9c916a63bf6a8f6580c.jpg) +Fig.3: CAD image reconstruction results by ViTMAE. (a) Input sketches, (b) masked sketches with the mask ratio of $75\%$ , reconstructed results using a (c) pretrained and (d) further fine-tuned ViTMAE. + +However, it's important to note that while some reconstructed sketches are notably accurate, they do not fully meet the requirements for parametric CAD engineering due to the vagueness and lack of parameterization in the entities. To bring image and text features to the same embedding space, we integrate a down-sampling layer and a lightweight projection module atop the vision encoder. + +Text Encoding Stream. Treating the sketch primitive sequence as textual data, we first tokenize this sequence to derive a series of text tokens $T = t_{1}, t_{2}, \ldots, t_{l}$ . These tokens are then processed using the encoder of a pre-trained language model, effectively encoding the sketch information in a textual format. + +Text-grounded Vision Decoding Stream. This stream is dedicated + +to constructing the autocompleted pixels, mapping the fused image and text representation to the pixel-level sketch image. The input for the vision decoder comprises the fused image and text embeddings, augmented with adaptively masked token patches to keep the same patch size as the input image. This construction process ensures that the vision embedding is grounded by textual information and also enhances the text representation. Note that for CAD autoconstraint task, we don't reconstruct the image since the input sketches are complete sketch images already. + +Image-grounded Text Decoding Stream. In alignment with Equations 1 and 2, we conceptualize autocompletion as a conditional text generation process. + +Within this framework, the combined image and text embeddings serve as inputs to the decoder of a language model. This decoder is capable of autoregressively generating the required output. Notably, the parameters of this language model are initialized from a pre-trained model, which itself has undergone extensive pre-training on vast datasets of code [27], enhancing its effectiveness in this context. + +# 4.2 Training Objectives + +As shown in Figure 2, we jointly optimize three distinct objectives during training: one focused on understanding and two on generation. + +Image-Text Contrastive Loss (ITC) is employed to activate the unimodal encoders, following the approach in [12]. This loss function aligns the visual and textual feature spaces by ensuring that positive image-text pairs have similar representations while contrasting with negative pairs. Such a contrastive approach has been instrumental in enhancing multimodal model training [13,20], not only improving the alignment of visual and language features but also accelerating the convergence during training. The image-text contrastive loss is defined as the cross-entropy between the similarity $p$ of image(I) text(T) embedding pairs and ground-truth one-hot similarity $y$ (array from [0,N], N is the number of samples), formulated as: + +$$ +p = \frac {\exp (s (I , T))}{\sum_ {i , j} ^ {N} \exp (s (I _ {i} , T _ {j}))}, \mathcal {L} _ {I T C} = \text {C r o s s - E n t r o p y} (y, p). \tag {5} +$$ + +Image Decoding Loss (IDL) is responsible for constructing autocompletion sketch images. This process involves translating the fused visual and textual information into pixel values for each patch. IDL employs a mean squared error (MSE) loss, similarly to the approach in ViTMAE [9], comparing the autocompleted sketch images with their ground-truth counterparts at the pixel level. The Image decoding loss can be formulated as: + +$$ +\mathcal {L} _ {I D L} = \frac {1}{n} \sum_ {i = 1} ^ {n} \left(Y _ {i} - \hat {Y} _ {i}\right) ^ {2}, \tag {6} +$$ + +where $n$ is the number of total pixels, $Y_{i}$ is the ground-truth pixel value of $i$ -th pixel, and $\hat{Y}_i$ is the output pixel value correspondingly. For CAD autoconstraint tasks, since the input is a full sketch image, we ignore IDL during the training of CAD autoconstraint models. + +Language Modeling Loss (LM) is designed to generate text-like autocompletion sketches, focusing on constraint sequences and utilizing the image-grounded text decoder. It optimizes a cross-entropy loss to maximize the likelihood described in Equation ①. LM endows the model with the capability to transform fused multimodal information into coherent autocompletion sketch + +sequences. For the CAD autoconstraint task, we use Equation 2 as the language modeling loss. + +Therefore, the full training objective of CadVLM is: + +$$ +\mathcal {L} = \mathcal {L} _ {I T C} + \mathcal {L} _ {L M} + \mathcal {L} _ {I D L}. \tag {7} +$$ + +During inference, given incomplete sketch image and primitive pairs, we only evaluate the generated text and use that for further mechanical applications. + +# 5 Experiments + +We conduct extensive experiments and ablation studies on CadVLM for the CAD autocompletion and autoconstraint tasks. Quantitative assessment is conducted by measuring Entity Accuracy, Sketch Accuracy, and CAD F1. + +# 5.1 Experimental settings + +Dataset We use the SketchGraphs [22] dataset for all our experiments, consisting of 15 million CAD sketches collected from Onshape. Referring to [23], we use the filtered version and further deduplicate the dataset. After deduplicated, the training, validation, and test set have 626,236, 22,031, 21,979 unique sketches correspondingly. The filtered sketches are restricted to sketches comprised of arcs, circles, and lines. To ensure a fair comparison, we normalize each sketch as in [23]; specifically the quantized range of numerical tokens in the primitive sequence is [1, 64]. + +Baselines We evaluate several ablation and conditional variants of CadVLM, as well as three SOTA CAD generative model baselines: Vitruvion [23], Deepcad [29] and SketchGraphs [22]. Since SketchGraphs does not incorporate continuous primitive parameters such as coordinate value, we only compare with the CAD autoconstraint task. And DeepCAD only considers parametrized command sequences, which are incompatible with CAD constraints. We primarily evaluate its performance on CAD autocompletion tasks. + +Settings For the image encoder and decoder, we choose the encoder and decoder of state-of-the-art pre-trained ViT-MAE-base from [9]. Following [9] we set the patch size and stride both as 32. We render all sketch images with size of $224 \times 224 \times 3$ . For the text encoder and text decoder, we use the encoder and decoder of pre-trained CodeT5+ [27] correspondingly, which is an instruction-tuned code LLM with 770M parameters trained on various code datasets. We train the autocompletion model and autoconstraint model both for 30 epochs, using the Adam optimizer with decoupled weight decay regularization [17], with the learning rate set according to the CosineAnnealingLR [16] scheduler. The initial learning rate is set to $3e - 4$ , with batch size 32. All our experiments are run using 8 NVIDIA A100 GPUs, and our models take between 30-40 hours to train. In the CAD Autocompletion experiments, we randomly sample $20\% -80\%$ + +Table 1: Quantitative sketch generative model comparison results for CAD autocompletion task. Ske-Acc means the probability of generating the correct full remaining sketch and Ent-Acc means the probability of generating at least one entity of the ground truth. Bold shows the best results of each metric. For all metrics the higher the better. + +
ModelSke-AccEnt-AccCAD-F1
Vitruvion3.0%40.7%11.3%
Deepcad1.2%23.1%8.9%
ChatGPT-15%6.8%41.3%21.2%
ChatGPT-7%7.6%43.8%22.2%
ChatGPT-1%6.5%43.7%22.5%
CadVLM-Text21.2%66.7%42.8%
CadVLM-w/o-IDL22.6%68.4%44.2%
CadVLM-w/o-ITC22.8%67.6%43.9%
CadVLM-w/o-IDL&ITC23.0%67.7%44.1%
CadVLM23.8%68.3%45.2%
+ +Table 2: Quantitative sketch generative model comparison results for CAD autoconstraint task. For all metrics higher is better. + +
ModelSke-AccEnt-AccCAD-F1
SketchGraphs0.48%93.0%42.0%
Vitruvion10.7%94.6%71.6%
CadVLM-Text11.5%95.0%72.6%
CadVLM-w/o-ITC12.1%95.4%74.5%
CadVLM11.3%94.9%73.8%
+ +of sketch entities as input and aim to predict the remaining entities from the ground truth sketch. + +# 5.2 Quantitative comparison with baseline sketch generative models + +Table 1 reports results on the CAD Autocompletion task. We find that Vitruvion achieves $3\%$ on Sketch Accuracy and $40.7\%$ on Entity Accuracy, meaning it can accurately predict $3\%$ of sketches and correctly identifies at least one entity in $40.7\%$ of sketches. Deepcad performs worst since it only has 4 layers of encoder and decoder, makes it hard to represent complex CAD sketches. Our CadVLM outperforms Virruvion on Sketch Accuracy, Entity Accuracy, and CAD-F1 by $20.8\%$ , $28.3\%$ , and $39.9\%$ respectively. These results demonstrate the effectiveness of pre-trained foundational models and image modality. A closer examination of the qualitative results in Figure 4 reveals that Vitruvion produces many invalid results missing symmetric curves or lines. Sketches from CadVLM are better in terms of quality with more complex shapes and stronger symmetry. + +ChatGPT Series Model Finetune To evaluate the ability of a general LLM on the CAD autocompletion task, we fine-tuned ChatGPT on three subsets + +![](images/6fa3e0379ea4f351096e7c56ee38ffb02db6dec4adc16eb015e6c7fe2feb043e.jpg) +Fig. 4: Comparative Analysis of Autocompletion in CAD Design. Top row: Random samples from the SketchGraphs test dataset. Second row: Initial input entities serving as the primer for the autocompletion task. Third and fourth row: Autocompletion results produced by Vitruvion and CadVLM. + +of $15\%$ , $7\%$ , and $1\%$ of the SketchGraphs dataset. The results indicate that the ChatGPT-series outperforms Vitruvion on all three metrics, especially on Sketch Accuracy and CAD-F1. However, there were no significant performance improvements when finetuning on larger portions of the training dataset. One possible reason is we only finetune ChatGPT for less than 5 epochs due to training costs, making it hard for ChatGPT to adapt to CAD data. + +CadVLM variants To better understand the source of performance gains, we conducted a comparative analysis of different variants of our proposed CadVLM: CadVLM-text, CadVLM-w/o-ITC, CadVLM-w/o-IDL&ITC, and CadVLM w/o-IDL. CadVLM-text relies solely on the sketch primitive sequence text as its input, optimizing for language modeling loss. In contrast, CadVLM-w/o-IDL utilizes ITC and LM loss for its training objectives, CadVLM-w/o-ITC focuses on IDL and LM, and CadVLM-w/o-IDL&ITC prioritizes only the LM objective. The findings reveal that CadVLM-Text underperforms compared to all other CadVLM variants across all three metrics, which underlines the value of the sketch image modality in providing diverse geometric information alongside sequence text. + +Interestingly, CadVLM-w/o-IDL emerges as the superior performer among the CadVLM variants, highlighting the significance of aligning image and text embeddings in multimodal modeling. Incorporating Image Decoding Loss (IDL) further enhances the alignment between image and text embeddings, leading CadVLM to show a notable improvement in all three metrics. + +For the CAD Autoconstraint task, as shown in Table 2, SketchGraphs struggles to predict the whole sketch correctly, with only $0.48\%$ Sketch Accuracy. This suggests that relying solely on graph structures of sketches is insufficient for handling long sequences of sketch primitives. In contrast, Vitruvion achieves $10.7\%$ on Sketch Accuracy, showing the necessity to model primitives. For the CAD Autoconstraint task it is less challenging to predict at least one constraint correctly but more challenging to accurately predict all constraints, easily caused + +![](images/7b6a304b5cd51cb43d1faab4160f0cb561893f81784c2c80f1b338c5e815da9f.jpg) +(a) + +![](images/105a66c6699583199a1cfd7681410c6dd78e0e9b628d477664e19c86ffa84cb5.jpg) +(b) +Fig. 5: Effect of input entity ratio to (a) Entity Accuracy, (b) Sketch Accuracy, and (c) CAD F1. + +![](images/7840ffc800e1712acb1ca19b97e73bfe705811115cb05d3fd5384d80cfee2d9f.jpg) +(c) + +by potential data labeling inconsistencies. As is well-known, minor discrepancies in primitive values can drastically alter constraints, such as changing them from parallel to coincident. + +The CadVLM series model exceeds the performance of both SketchGraphs and Vitruvion across all metrics. CadVLM-Text surpasses CadVLM on Ske-Acc and Ent-Acc since the pretraining task of ViTMAE has a great discrepancy with CAD autoconstraint tasks. Therefore aligning the image and text representation directly might not be an optimal way for CAD autoconstraint task.Notably, CadVLM-w/o-ITC emerges as the top performer within the CadVLM series. This may be attributed to its more substantial text decoder and utilization of multimodal information, which enables the model to more effectively integrate knowledge from both primitives and images to inform geometric information. + +Table 3: Quantitative results for image condi- Table 4: Effect of adding the inductioned primitive generation. Bold represents tive bias of entity-level embeddings the best results of all metrics. to the sketch sequence. + +
Conditioned -ImageModelEnt-AccCAD-F1
PreciseVitruvion83.8%35.8%
CadVLM98.0%62.7%
Hand-drawnVitruvion84.2%38.4%
CadVLM99.4%60.1%
Noisy hand-drawnVitruvion83.2%35.8%
CadVLM94.4%56.5%
+ +
ModelCadVLM w/-Inductive -TextBias
Ent-Acc66.7%67.5%
Ske-Acc21.2%22.2%
CAD-F142.8%43.7%
+ +# 5.3 Ablation Study + +We conducted several ablation studies to assess the performance of CadVLM on image-conditioned primitive generation tasks and investigate the effectiveness of each component within CadVLM. + +Image Conditional Generation. Following [23], we conduct a series of experiments to evaluate the effectiveness of CadVLM modeling primitive sequences conditioned on the images of sketches. As designers may begin their + +designs with hand-drawn sketches before transitioning to CAD programs, accurately inferring parametric CAD from images or scans can dramatically reduce the effort of translating from rough paper sketches to the CAD model. + +When conditioning only on sketch images, the text encoder of CadVLM is ignored. We only use the image encoder, image decoder, and text decoder. The overall model is trained to maximize the probability of the ground truth primitives portrayed in the image, conditioned on a sequence of patch embeddings. + +The results of Table 3 show three kinds of sketch image rendering input to imitate different settings in the real world: precise renderings, rendering from the hand-drawn simulator of [23], and renderings with random affine augmentations of the hand-drawn simulator. As both Vitruvion and CadVLM can seldom predict the sketch exactly correctly, we ignore Sketch Accuracy in our image conditional generation experiments. The results show that CadVLM outperforms Vitruvion on both Entity Accuracy and CAD F1. Vitruvion performs best on hand-drawn images with $83.8\%$ Entity Accuracy and $38.4\%$ CAD F1. While CadVLM achieves almost $99.4\%$ Entity Accuracy and $62.7\%$ CAD F1. Conditioning on precise rendered images achieves the best performance. + +Entity Level Modelling. As described earlier, each CAD sketch is comprised of multiple entities. We considered adding this inductive bias to the model by processing each entity separately and then combining their embeddings with the initial sequence of all sketch tokens. For this purpose, we first pretend a special token $\langle \text{ENTITY} \rangle$ to each of the entity input sequences. Subsequently, the entities from all the sketches in the batch are fed into the model's encoder in parallel. Next, the embeddings of each entity are gathered for each sketch. For each sketch, we concatenate these embeddings with the original sequence tokens using another special token $\langle \text{TOKEN} \rangle$ as a delimiter for the second part of the sequence. The decoder remains unchanged. As shown in Table 4, incorporating this inductive bias increases computation and memory demands while only marginally enhancing the performance. + +Impact of different input entity ratios for CAD autocomblection. To explore the impact of different input entity ratios for CAD autocomblection, we further test CAD autocomblection results with different input entity ratios: $20\%$ , $40\%$ , $60\%$ , and $80\%$ . Compared with the experiments in Sec 5.4, we specify the sampled ratio of input entities. As shown in Figure 5, CadVLM outperforms Vitruvion on all input entity ratios. In particular, CadVLM achieves $73.8\%$ Entity Accuracy, $49.5\%$ Sketch Accuracy, and $62.7\%$ CAD F1 when given $80\%$ of entities as input. For the hardest setting when given only $20\%$ of entities as input, CadVLM can still achieve $43.9\%$ Entity Accuracy and $16.9\%$ CAD-F1, which proves that CadVLM has great potential to be used in real world settings. + +Effect of using different language models for text encoder and decoder. To better compare the performance of different text encoder-decoder, we experiment using different pre-trained foundation models as text encoder and decoder. To eliminate the influence of the image encoder and decoder, we only consider the CadVLM-Text variant. For the choice of text encoder and decoder, we also compare the performance of different Encoder-Decoder pre-trained + +Table 5: Effect of using different architectures for CadVLM-Text. CodeT5+ 770M is the encoder-decoder for text we use in our main experiments. CodeT5+\* is the result of training from scratch rather than pre-trained. + +
Model#ParameterEnt-AccSke-AccCAD-F1
CodeT5+220M66.5%21.5%42.9%
CodeT5+*220M65.0%18.9%40.3%
CodeT5+770M66.7%21.2%42.8%
ByT5-base580M66.9%22.0%43.1%
ByT5-large1.2B66.8%20.6%41.9%
T5-base223M66.2%20.2%41.7%
T5-large783M66.8%21.2%42.6%
+ +language models on CAD autocombpletion. The T5 family of models are among the most recognized and widely used encoder-decoder models in the open-source community. Therefore, we experimented with T5 [21], ByT5 [32] and CodeT5+ [27]. The results are shown in Table 5. The effect of model size is clear for T5 and CodeT5+ but not ByT5. Also, The cost-benefit of solely making the model bigger becomes less favorable over time as the improvements in performance start to plateau. Another observation is that, because of the nature of the data which is closer to a code sequence than natural language, CodeT5+, which is fine-tuned on code, exhibits better performance. + +We also trained CodeT5+ from scratch to examine the impact of pre-training on natural language and code. As evidenced in Table5, training from scratch falls short in all three metrics. Additionally, training from scratch converged much slower than finetuning. Despite the fact that our representation of sketches as code is relatively unique and not found in natural language, the rules and token embeddings learned from text are very helpful for domain adaptation. + +# 6 Conclusion + +In this study, we introduce CadVLM, an innovative end-to-end vision-language model tailored for modeling parametric CAD sketches, which are pivotal in modern mechanical design. CadVLM harnesses the commonsense knowledge and reasoning abilities inherent in large language models, enabling it to adeptly handle the complex geometric reasoning required for CAD sketches. Our experimental results demonstrate that CadVLM outperforms existing state-of-the-art models in both autocompletion and autoconstraint tasks, highlighting the potent enhancement that pre-trained models bring to CAD construction. + +# Acknowledgements + +This work is supported by the Canada CIFAR AI Chair Program and the Canada NSERC Discovery Grant (RGPIN-2021-03115). + +# References + +1. Azerbayev, Z., Schoelkopf, H., Paster, K., Santos, M.D., McAleer, S., Jiang, A.Q., Deng, J., Biderman, S., Welleck, S.: Llemma: An open language model for mathematics. arXiv preprint arXiv:2310.10631 (2023) +2. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) +3. Camba, J.D., Contero, M., Company, P.: Parametric cad modeling: An analysis of strategies for design reusability. Computer-Aided Design 74, 18-31 (2016) +4. Chen, Q., Hu, X., Wang, Z., Hong, Y.: Medblip: Bootstrapping language-image pre-training from 3d medical images and texts. arXiv preprint arXiv:2305.10799 (2023) +5. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., et al.: Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794 (2022) +6. Chen, Y., Zhang, S., Han, B., He, T., Li, B.: Camml: Context-aware multimodal learner for large models. arXiv preprint arXiv:2401.03149 (2024) +7. Floridi, L., Chiriatti, M.: Gpt-3: Its nature, scope, limits, and consequences. Minds and Machines 30, 681-694 (2020) +8. Ganin, Y., Bartunov, S., Li, Y., Keller, E., Saliceti, S.: Computer-aided design as language (2021) +9. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022) +0. Jones, B.T., Hu, M., Kodnongbua, M., Kim, V.G., Schulz, A.: Self-supervised representation learning for cad. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21327-21336 (2023) +1. Kenton, J.D.M.W.C., Toutanova, L.K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of naacL-HLT. vol. 1, p. 2 (2019) +2. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022) +3. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems 34, 9694-9705 (2021) +4. Li, P., Guo, J., Zhang, X., Yan, D.M.: Secad-net: Self-supervised cad reconstruction by learning sketch-extrude operations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16826 (2023) +5. Lindsay, A., Paterson, A., Graham, I.: Identifying and quantifying inefficiencies within industrial parametric cad models. In: Advances in Manufacturing Technology XXXII: Proceedings of the 16th International Conference on Manufacturing Research. vol. 8, p. 227 (2018) +6. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) +7. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) +8. Luo, H., Sun, Q., Xu, C., Zhao, P., Lou, J., Tao, C., Geng, X., Lin, Q., Chen, S., Zhang, D.: Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct. arXiv preprint arXiv:2308.09583 (2023) + +19. Para, W., Bhat, S., Guerrero, P., Kelly, T., Mitra, N., Guibas, L.J., Wonka, P.: Sketchgen: Generating constrained cad sketches. Advances in Neural Information Processing Systems 34, 5077-5088 (2021) +20. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +21. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020) +22. Seff, A., Ovadia, Y., Zhou, W., Adams, R.P.: Sketchgraphs: A large-scale dataset for modeling relational geometry in computer-aided design. arXiv preprint arXiv:2007.08506 (2020) +23. Seff, A., Zhou, W., Richardson, N., Adams, R.P.: Vitruvion: A generative model of parametric cad sketches. arXiv preprint arXiv:2109.14124 (2021) +24. Singhal, K., Azizi, S., Tu, T., Mahdavi, S.S., Wei, J., Chung, H.W., Scales, N., Tanwani, A., Cole-Lewis, H., Pfohl, S., et al.: Large language models encode clinical knowledge. arXiv preprint arXiv:2212.13138 (2022) +25. Singhal, K., Tu, T., Gottweis, J., Sayres, R., Wulczyn, E., Hou, L., Clark, K., Pfohl, S., Cole-Lewis, H., Neal, D., et al.: Towards expert-level medical question answering with large language models. arXiv preprint arXiv:2305.09617 (2023) +26. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +27. Wang, Y., Le, H., Gotmare, A.D., Bui, N.D., Li, J., Hoi, S.C.: Codet5+: Open code large language models for code understanding and generation. arXiv preprint arXiv:2305.07922 (2023) +28. Willis, K.D., Jayaraman, P.K., Lambourne, J.G., Chu, H., Pu, Y.: Engineering sketch generation for computer-aided design. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2105–2114 (2021) +29. Wu, R., Xiao, C., Zheng, C.: Deepcad: A deep generative network for computer-aided design models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6772-6782 (2021) +30. Xu, X., Jayaraman, P.K., Lambourne, J.G., Willis, K.D., Furukawa, Y.: Hierarchical neural coding for controllable cad model generation. arXiv preprint arXiv:2307.00149 (2023) +31. Xu, X., Willis, K.D., Lambourne, J.G., Cheng, C.Y., Jayaraman, P.K., Furukawa, Y.: Skexgen: Autoregressive generation of cad construction sequences with disentangled codebooks. In: International Conference on Machine Learning. pp. 24698-24724. PMLR (2022) +32. Xue, L., Barua, A., Constant, N., Al-Rfou, R., Narang, S., Kale, M., Roberts, A., Raffel, C.: Byt5: Towards a token-free future with pre-trained byte-to-byte models. Transactions of the Association for Computational Linguistics 10, 291-306 (2022) +33. Zha, L., Zhou, J., Li, L., Wang, R., Huang, Q., Yang, S., Yuan, J., Su, C., Li, X., Su, A., et al.: Tablegpt: Towards unifying tables, nature language and commands into one gpt. arXiv preprint arXiv:2307.08674 (2023) \ No newline at end of file diff --git a/2024/CadVLM_ Bridging Language and Vision in the Generation of Parametric CAD Sketches/images.zip b/2024/CadVLM_ Bridging Language and Vision in the Generation of Parametric CAD Sketches/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5b92cad3da580878e23d7a9e3c12e1b411b69d74 --- /dev/null +++ b/2024/CadVLM_ Bridging Language and Vision in the Generation of Parametric CAD Sketches/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3aca8f1abede7ebf01f856f4d044a7de979de877414fa9ece077633890bbfa2b +size 382394 diff --git a/2024/CadVLM_ Bridging Language and Vision in the Generation of Parametric CAD Sketches/layout.json b/2024/CadVLM_ Bridging Language and Vision in the Generation of Parametric CAD Sketches/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c02cde71367af70e0dd1bea3d6be613aef78d2fc --- /dev/null +++ b/2024/CadVLM_ Bridging Language and Vision in the Generation of Parametric CAD Sketches/layout.json @@ -0,0 +1,7815 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "text", + "content": "CadVLM: Bridging Language and Vision in the Generation of Parametric CAD Sketches" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 154, + 167, + 459, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 167, + 459, + 194 + ], + "spans": [ + { + "bbox": [ + 154, + 167, + 459, + 194 + ], + "type": "text", + "content": "Sifan Wu1, Amir Hosein Khasahmadi2, Mor Katz2, Pradeep Kumar Jayaraman2, Yewen Pu2, Karl Willis2, and Bang Liu1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 201, + 414, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 201, + 414, + 224 + ], + "spans": [ + { + "bbox": [ + 198, + 201, + 414, + 224 + ], + "type": "text", + "content": "1 Université de Montréal & MILA, Montreal, Canada \n2 Autodesk AI Lab, Toronto, Canada" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 248, + 475, + 347 + ], + "blocks": [ + { + "bbox": [ + 134, + 248, + 475, + 347 + ], + "lines": [ + { + "bbox": [ + 134, + 248, + 475, + 347 + ], + "spans": [ + { + "bbox": [ + 134, + 248, + 475, + 347 + ], + "type": "image", + "image_path": "b76ecdde05298a4355611c5a2772e7e571b5f9ab6b6028dd1b2470845351ffb0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 354, + 482, + 410 + ], + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 410 + ], + "type": "text", + "content": "Fig. 1: For the CAD autocompletion task, our multi-modal CadVLM model (b) receives partial CAD entities as both image and text input (a) and generates the remaining sketch entities as output (c). The complete sketch, with optional predicted constraints, can then be used in CAD software (d) to form 3D shapes (e). More description about the primitive values in sketch text are in the Appendix." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 446, + 455, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 446, + 455, + 646 + ], + "spans": [ + { + "bbox": [ + 159, + 446, + 455, + 646 + ], + "type": "text", + "content": "Abstract. Parametric Computer-Aided Design (CAD) is central to contemporary mechanical design. However, it encounters challenges in achieving precise parametric sketch modeling and lacks practical evaluation metrics suitable for mechanical design. We harness the capabilities of pre-trained foundation models, renowned for their successes in natural language processing and computer vision, to develop generative models specifically for CAD. These models are adept at understanding complex geometries and design reasoning, a crucial advancement in CAD technology. In this paper, we propose CadVLM, an end-to-end vision language model for CAD generation. Our approach involves adapting pre-trained foundation models to manipulate engineering sketches effectively, integrating both sketch primitive sequences and sketch images. Extensive experiments demonstrate superior performance on multiple CAD sketch generation tasks such as CAD autocomblection, CAD autoconstraint, and image conditional generation. To our knowledge, this is the first instance of a multimodal Large Language Model (LLM) being successfully applied to parametric CAD generation, representing a pioneering step in the field of computer-aided mechanical design." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 654, + 397, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 654, + 397, + 665 + ], + "spans": [ + { + "bbox": [ + 160, + 654, + 397, + 665 + ], + "type": "text", + "content": "Keywords: CAD Representation " + }, + { + "bbox": [ + 160, + 654, + 397, + 665 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 654, + 397, + 665 + ], + "type": "text", + "content": " Vision language model" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 295 + ], + "type": "text", + "content": "Parametric Computer-Aided Design (CAD) is crucial in mechanical engineering, playing a key role in developing modern objects such as lamps, cars, and buildings. Its importance lies in its ability to accurately depict complex designs, essential for the integrity and functionality of engineered products interacting with humans and the environment. CAD sketches, consisting of basic 2D shapes, are foundational for parametric CAD modeling [3]. As illustrated in Fig [1], these sketches comprise primitive geometric elements (e.g., lines, arcs, points) defined by various relationships (e.g., coincident, parallel, tangent). However, the required precision for 3D construction in parametric CAD can be both challenging and time-consuming [15]. Consequently, there is a significant push to improve CAD generation to simplify and accelerate the manual design process. Leveraging deep learning to identify patterns in CAD sketches could automate the completion of geometry and constraints in engineering drawings [22,23,28]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 296, + 483, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 483, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 483, + 485 + ], + "type": "text", + "content": "Integrating deep learning into CAD offers a significant opportunity to transform the field. Recent research [10,14,19,22,23,29,31] has concentrated on the generative modeling of engineering sketches, particularly focusing on Transformer-based architectures [26]. These approaches often utilize graph structures [22] or attention mechanisms [23] to process sketch information. However, the effectiveness of these methods is limited by the availability of data and a lack of commonsense understanding, which restricts their capacity to perform complex geometric reasoning in engineering sketches. Additionally, rendered sketch images (such as in Fig. 1(a) and 4) have great potential to represent geometry patterns from image modality, which could be complementary for parametric sketches language. While none of current CAD generative models take sketch image into consideration. Our research presents a novel Multimodal Transformer-based Generative model for CAD generation tasks, pioneering in its use of both language and image for sketch modeling. By merging these modalities, we aim to increase the accuracy of CAD generation, setting a new benchmark by simultaneously modeling text-style sketch parametric and image-style sketch renderings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 486, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 617 + ], + "type": "text", + "content": "Large Language Models (LLMs) have gained widespread success for their powerful commonsense reasoning and problem-solving abilities in a variety of domains, such as natural language processing, mathematics, and medical sciences [1,5,7,18,24,25,33]. The capabilities of LLMs align well with engineering sketches, which can be understood as sequences representing points and connections. Commonsense knowledge in pre-trained LLM can also offer valuable hints for sketch patterns. Furthermore, combining sketch images with textual descriptions can greatly improve the performance of language models in CAD sketch generative models. In the realm of parametric CAD design, which requires a thorough understanding of complex interactions(e.g. parallel, coincident, symmetry), the advantages of LLMs become even more pronounced." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "In this work, we leverage the capabilities of pre-trained LLMs to develop a generative model tailored for parametric CAD, addressing the limitations of the existing methods. We introduce CadVLM, a CAD Vision Language Model, featuring a multimodal encoder-decoder framework. CadVLM stands out by pro" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "S. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "cessing engineering sketches both linguistically, through token sequences representing sketch geometry and constraints, and visually, through the use of rendered sketch images. To the best of our knowledge, CadVLM is the first generative model to utilize pre-trained vision and language models for CAD sketches, which is capable of addressing different CAD representation learning tasks such as CAD autocompletion, CAD autoconstraint, and image conditional generation tasks. Our contributions are threefold:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 206, + 486, + 312 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 206, + 481, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 481, + 251 + ], + "type": "text", + "content": "- We propose CadVLM, a novel CAD Vision-Language Model, for simultaneously modeling engineering sketches as both language and images. As we know, we are the first to combine visual and textual data for CAD generative models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 254, + 481, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 254, + 481, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 254, + 481, + 286 + ], + "type": "text", + "content": "- We introduce three novel evaluation metrics—Entity Accuracy, Sketch Accuracy, and CAD F1 score—to quantitatively assess the quality of generated CAD sketches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 289, + 486, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 486, + 312 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 486, + 312 + ], + "type": "text", + "content": "- CadVLM demonstrates superior performance on the SketchGraphs dataset [22] in both CAD autocompletion and autoconstraint tasks." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 330, + 237, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 237, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 237, + 342 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 354, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 485 + ], + "type": "text", + "content": "Vision Language Pre-Training. The great success of large language models (LLMs) such as BERT [11] and GPT [2] has drawn significant attention within the field of natural language processing and code generation domain. CodeT5+ [27] is a family of open-source encoder-decoder LLMs for code tasks, ranging from 220 million to 16 Billion parameters. Beyond single modality LLMs, multimodal foundation models have also exhibited enhanced performance in various vision-and-language tasks [6,12,20]. Depending on downstream tasks, Various pre-training objectives have also been proposed like image-text contrastive learning [20], image-text matching [12], and (masked) language modeling [12]. Most recently, developing domain-specific multimodality LLMs has gained momentum such as MedBLIP [4] and TableGPT [33]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 486, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 640 + ], + "type": "text", + "content": "CAD Sketch Generation 2D engineering sketches are fundamental in CAD for designing and manufacturing mechanical parts. The recent availability of extensive engineering sketch datasets has facilitated the development of generative models that enhance traditional CAD workflows. The SketchGraphs [22] dataset, featuring 15 million parametric CAD sketches with detailed geometric constraint graphs, showcases primitive interrelations. The paper also introduces a baseline model using autoregressive message-passing networks for constraint graph generation, but it lacks primitive coordinate output, relying solely on constraint graphs and the Onshape solver for sketch configuration, thus limiting its utility in CAD autocomblection tasks. Ganin et al. [8] proposed another large-scale CAD dataset containing over 4.7M parametric sketches from the Onshape public repository. However, limited by sketch format, this dataset has not been widely researched in the realm of CAD generative models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 641, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 667 + ], + "type": "text", + "content": "Another line of work for generative modeling of CAD sketches is based on Transformer [23,28-30]. Vitruvion [23] models both primitives and constraints," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "text", + "content": "Abbreviated paper title" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 296 + ], + "type": "text", + "content": "but its training on Transformer does not inherently include geometric knowledge. Furthermore, Vitruvion's major limitation lies in its lack of quantitative comparison in terms of CAD generation quality. While CurveGen [28] tackles the problem of engineering sketch generation without considering sketch constraints like us, it struggles to generalize to few-shot samples. Both Vitruvion and CurveGen face limitations in their model performance, making them challenging to deploy in practical CAD engineering applications. Recently, there has been some research about CAD construction sequences generation [30,31]. However, these works cannot model sketch primitives to autocomplete parametric CAD sketches. We consider Vitruvion, Deepcad, and SketchGraphs as the baselines for CAD Autocompletion task and CAD Autoconstraint task since they are current state-of-the-art models using transformer architectures. Different from existing work, our method utilizes sketch image modality and commonsense knowledge in pre-trained foundation models, providing both sketch sequence textual and image information for sketch generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 316, + 226, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 316, + 226, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 226, + 329 + ], + "type": "text", + "content": "3 Background" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 342, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 437 + ], + "type": "text", + "content": "Engineering Sketch Definition. In the field of parametric CAD, engineering sketches are fundamental, serving as detailed blueprints for designing a range of objects from small components to large structures. These sketches are essential for converting conceptual ideas into precise, manufacturable designs. To ensure a standardized representation, sketches of varying dimensions, spanning from millimeters to meters, are uniformly positioned within a bounding box of 1-meter width at the origin. Each sketch consists of two key sequences: a sequence of primitives, " + }, + { + "bbox": [ + 130, + 342, + 482, + 437 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 342, + 482, + 437 + ], + "type": "text", + "content": ", and a sequence of constraints, " + }, + { + "bbox": [ + 130, + 342, + 482, + 437 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 342, + 482, + 437 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "text", + "content": "We adopt the normalization and quantization methods for primitive parameters as outlined by [23]. A sketch primitive is expressed as " + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "inline_equation", + "content": "S = (e_1, e_2, \\dots, e_m)" + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "text", + "content": " represents the number of entities in the sketch. Each entity " + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "text", + "content": " is characterized by positional parameters " + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "inline_equation", + "content": "(p_1, \\dots, p_k)" + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "text", + "content": " represents the normalized coordinates of points that constitute the entity. Sketches typically include three types of entities—lines, arcs, and circles. As illustrated in Figure 11, all sketches are centrally positioned and normalized within a range of [1,64]. Lines, shown in blue, are depicted using two points: start and end point coordinates. Curves, shown in green, are characterized by three points: the start, the end, and a mid-point on the curve. Circles, illustrated in red, are portrayed with four uniformly distributed points on the circle's circumference. This format allows for the interpretation of each sketch as a sequence of tokens. Constraints are similarly represented as primitives in " + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = (c_1, \\dots, c_n)" + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "text", + "content": ", with each constraint entity " + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 437, + 482, + 617 + ], + "type": "text", + "content": " being defined by its type and the reference indices of parameters to which the primitives must conform." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "CAD AutoCompletion Task. In addressing the existing CAD AutoCompletion task, which aims to automate routine CAD design procedures, we propose a new autoregressive generative model. This task involves completing a partially given sketch to form a fully realized design. The mathematical formulation of" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "S. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 117, + 479, + 266 + ], + "blocks": [ + { + "bbox": [ + 134, + 117, + 479, + 266 + ], + "lines": [ + { + "bbox": [ + 134, + 117, + 479, + 266 + ], + "spans": [ + { + "bbox": [ + 134, + 117, + 479, + 266 + ], + "type": "image", + "image_path": "fb06fae5aa28f00a34061e4f935404b39b30c03ba9aad4b5e3105e1e54fda047.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 274, + 482, + 384 + ], + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 384 + ], + "type": "text", + "content": "Fig. 2: CadVLM model architecture and training objectives for CAD autoconstraint. We propose a unified vision-language model capable of outputting both autocompleted sketch images and sketch primitives. The inputs are paired with incomplete sketch primitives and images, which are encoded by pre-trained text encoder and image encoder correspondingly. Aligned by image-text contrastive(ITC) loss, the concatenation of image embedding and text embedding then will be input to the text-grounded image decoder and image-grounded text decoder. The image decoder is trained with image decoding loss between autocompleted sketch images and ground-truth sketch images. The text decoder is trained with a language modeling(LM) loss to generate autocompleted primitive." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 416, + 182, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 416, + 182, + 426 + ], + "spans": [ + { + "bbox": [ + 132, + 416, + 182, + 426 + ], + "type": "text", + "content": "the task is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 200, + 427, + 480, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 427, + 480, + 459 + ], + "spans": [ + { + "bbox": [ + 200, + 427, + 480, + 459 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\Theta) = - \\log p _ {\\Theta} (\\mathcal {S} | \\mathcal {S} _ {p}) = - \\sum_ {i = j} ^ {m} \\log p _ {\\Theta} \\left(e _ {i: m} \\mid e _ {1: i}\\right). \\tag {1}", + "image_path": "f62c70ab1c157df9ce614015df3112eb36a0232ccf75fb0c5007befd88671ca6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 460, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 460, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 131, + 460, + 482, + 508 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 460, + 482, + 508 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 131, + 460, + 482, + 508 + ], + "type": "text", + "content": " is the total number of entities in the complete sketch " + }, + { + "bbox": [ + 131, + 460, + 482, + 508 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 131, + 460, + 482, + 508 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 460, + 482, + 508 + ], + "type": "inline_equation", + "content": "S_{p} = (e_{1},\\dots,e_{i})" + }, + { + "bbox": [ + 131, + 460, + 482, + 508 + ], + "type": "text", + "content": " is a partial sequence containing " + }, + { + "bbox": [ + 131, + 460, + 482, + 508 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 131, + 460, + 482, + 508 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 131, + 460, + 482, + 508 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 131, + 460, + 482, + 508 + ], + "type": "text", + "content": " of the primitive entities from " + }, + { + "bbox": [ + 131, + 460, + 482, + 508 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 131, + 460, + 482, + 508 + ], + "type": "text", + "content": ". Our model's goal is to optimize the log-likelihood of completing the sketch " + }, + { + "bbox": [ + 131, + 460, + 482, + 508 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 131, + 460, + 482, + 508 + ], + "type": "text", + "content": " based on these partial sequences." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 512, + 482, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 482, + 572 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 482, + 572 + ], + "type": "text", + "content": "We employ open-source natural language pre-trained Transformer models for our generative model, with a particular emphasis on T5 models [21]. Additionally, we explore finetuning with GPT-3.5 [2]. Our results indicate significant cross-domain knowledge transfer capabilities in these pre-trained models, showcasing their effectiveness in the context of CAD design." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 575, + 482, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 575, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 131, + 575, + 482, + 610 + ], + "type": "text", + "content": "CAD Autoconstrain Task. The CAD autoconstrain task aims to autoregressively generate a set of constraints based on given geometry primitives. We model the constraint model as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 195, + 611, + 481, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 611, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 195, + 611, + 481, + 640 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\phi) = - \\log p _ {\\Phi} (\\mathcal {C} | \\mathcal {S}) = - \\sum_ {i = 1} ^ {n} \\log p _ {\\Phi} \\left(C _ {j: n} \\mid C _ {1: j}, S\\right), \\tag {2}", + "image_path": "f1a87ce9205b7368c667fe3c9a7c2f1c5b8be3b6a06c72c5b4d2126668807867.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": " is the number of constraints, and each " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": " is represented by a tuple of tokens indicating its constraint type and corresponding primitive entity param" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "text", + "content": "Abbreviated paper title" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "eters. Like autocompletion task, the autoconstrain model is trained to maximize the log-likelihood of generated tokens with respect to the ground truth." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 479, + 175 + ], + "type": "text", + "content": "Aligned with [23], we focus on 13 constraint types, which are coincident, parallel, and so on. Further details about these constraints are provided in the Supplementary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 176, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 479, + 209 + ], + "type": "text", + "content": "Evaluation Metrics. To quantitatively assess CAD generation models, we introduce three CAD-specific metrics: Sketch Accuracy, Entity Accuracy, and CAD F1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 212, + 480, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 480, + 257 + ], + "type": "text", + "content": "Sketch Accuracy measures the probability of correctly generating the entire remaining sketch in the test set. Entity Accuracy evaluates the likelihood of correctly generating at least one entity in the ground truth. They are calculated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 200, + 258, + 479, + 280 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 258, + 479, + 280 + ], + "spans": [ + { + "bbox": [ + 200, + 258, + 479, + 280 + ], + "type": "interline_equation", + "content": "\\text {S k e t c h A c c u r a c y} = \\frac {N _ {s}}{N}, \\quad \\text {E n t i t y A c c u r a c y} = \\frac {N _ {e}}{N}, \\tag {3}", + "image_path": "385a298e5c6d041623b421cf91dbf65f4d01a1b522213f1698d4662501003422.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 285, + 479, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 479, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 479, + 332 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 285, + 479, + 332 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 130, + 285, + 479, + 332 + ], + "type": "text", + "content": " is the number of correctly generated sketches in the test set, " + }, + { + "bbox": [ + 130, + 285, + 479, + 332 + ], + "type": "inline_equation", + "content": "N_{e}" + }, + { + "bbox": [ + 130, + 285, + 479, + 332 + ], + "type": "text", + "content": " is the number of generated sketches that match at least one entity with the ground truth sketch, and " + }, + { + "bbox": [ + 130, + 285, + 479, + 332 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 285, + 479, + 332 + ], + "type": "text", + "content": " is the total number of sketches in the test set. CAD F1 is a harmonic mean of precision and recall for each sketch, averaged over the dataset:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 198, + 342, + 479, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 342, + 479, + 370 + ], + "spans": [ + { + "bbox": [ + 198, + 342, + 479, + 370 + ], + "type": "interline_equation", + "content": "C A D F 1 = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {2 \\times \\text {s k e t c h p r e c i s i o n} \\times \\text {s k e t c h r e c a l l}}{(\\text {s k e t c h p r e c i s i o n} + \\text {s k e t c h r e c a l l})}, \\tag {4}", + "image_path": "7ee416313a48c09acd3b7c446efd6310e30ecb44ce37dfe85fc300eb65d29526.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 373, + 479, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 479, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 479, + 411 + ], + "type": "text", + "content": "where sketch precision " + }, + { + "bbox": [ + 130, + 373, + 479, + 411 + ], + "type": "inline_equation", + "content": "= \\frac{n_c}{n_p}" + }, + { + "bbox": [ + 130, + 373, + 479, + 411 + ], + "type": "text", + "content": ", sketch recall " + }, + { + "bbox": [ + 130, + 373, + 479, + 411 + ], + "type": "inline_equation", + "content": "= \\frac{n_c}{m}" + }, + { + "bbox": [ + 130, + 373, + 479, + 411 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 373, + 479, + 411 + ], + "type": "inline_equation", + "content": "n_c" + }, + { + "bbox": [ + 130, + 373, + 479, + 411 + ], + "type": "text", + "content": " is the count of correct entities in the completed sketch, " + }, + { + "bbox": [ + 130, + 373, + 479, + 411 + ], + "type": "inline_equation", + "content": "n_p" + }, + { + "bbox": [ + 130, + 373, + 479, + 411 + ], + "type": "text", + "content": " is the total number of entities predicted, and " + }, + { + "bbox": [ + 130, + 373, + 479, + 411 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 373, + 479, + 411 + ], + "type": "text", + "content": " is the number of entities in the original sketch (ground truth)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 430, + 432, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 432, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 432, + 456 + ], + "type": "text", + "content": "4 CadVLM: A Vision Language Model for CAD Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 470, + 479, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 479, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 479, + 540 + ], + "type": "text", + "content": "In this section, we delve into the specifics of our CAD Vision Language Model (CadVLM), employing the CAD autocompletion task as an illustrative example. In our approach, the sequence of sketch primitives " + }, + { + "bbox": [ + 130, + 470, + 479, + 540 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 470, + 479, + 540 + ], + "type": "text", + "content": " is interpreted akin to language text, while the corresponding visual representation of the sketch is treated as image " + }, + { + "bbox": [ + 130, + 470, + 479, + 540 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 470, + 479, + 540 + ], + "type": "text", + "content": ". It should be noted that the text-grounded image decoder is not utilized in the CAD auto-constraint task." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 561, + 263, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 561, + 263, + 572 + ], + "spans": [ + { + "bbox": [ + 131, + 561, + 263, + 572 + ], + "type": "text", + "content": "4.1 Network Framework" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "content": "CadVLM is structured as an asymmetrical encoder-decoder architecture, featuring a two-stream encoder, a text decoder based on a language model, and a lightweight image decoder for constructing the autocompleted image. This design is depicted in Figure 2. The two-stream encoder is tailored to process dual-modal inputs: a vision sub-encoder for the sketch image " + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "content": " and a text sub-encoder for the prefix sequence of sketch primitives " + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "inline_equation", + "content": "S_{p}" + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "content": ". For text decoding, we employ a causal language transformer model, which autoregressively generates" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "S. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": "the output, i.e., the autocompleted sketch. Finally, the image decoder utilizes the latent representation tokens to construct the visual representation of the autocompleted sketch." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 154, + 482, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 154, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 154, + 482, + 274 + ], + "type": "text", + "content": "Vision Encoding Stream. To encode partial sketch images and harness the capabilities of an existing large vision autoencoder model, we employ a pretrained vision transformer (ViT) [9] for extracting vision features. To bridge the gap between CAD images and natural images typically processed by ViT, we fine-tune the model on the SketchGraphs dataset. This fine-tuning involves a one-epoch image reconstruction task, with a " + }, + { + "bbox": [ + 130, + 154, + 482, + 274 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 130, + 154, + 482, + 274 + ], + "type": "text", + "content": " masking ratio, specifically tailored for CAD sketch data. Figure 3 illustrates the image reconstruction outcomes using both the original pre-trained ViT-MAE and the version fine-tuned on the SketchGraphs dataset. After finetuning, ViT-MAE shows enhanced representational performance for CAD sketches." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 138, + 300, + 304, + 438 + ], + "blocks": [ + { + "bbox": [ + 138, + 300, + 304, + 438 + ], + "lines": [ + { + "bbox": [ + 138, + 300, + 304, + 438 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 304, + 438 + ], + "type": "image", + "image_path": "e783fb245e3cd97204a1943891292cce70129cf25e18c9c916a63bf6a8f6580c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 446, + 308, + 511 + ], + "lines": [ + { + "bbox": [ + 131, + 446, + 308, + 511 + ], + "spans": [ + { + "bbox": [ + 131, + 446, + 308, + 511 + ], + "type": "text", + "content": "Fig.3: CAD image reconstruction results by ViTMAE. (a) Input sketches, (b) masked sketches with the mask ratio of " + }, + { + "bbox": [ + 131, + 446, + 308, + 511 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 131, + 446, + 308, + 511 + ], + "type": "text", + "content": ", reconstructed results using a (c) pretrained and (d) further fine-tuned ViTMAE." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 276, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 482, + 407 + ], + "type": "text", + "content": "However, it's important to note that while some reconstructed sketches are notably accurate, they do not fully meet the requirements for parametric CAD engineering due to the vagueness and lack of parameterization in the entities. To bring image and text features to the same embedding space, we integrate a down-sampling layer and a lightweight projection module atop the vision encoder." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 410, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 482, + 517 + ], + "type": "text", + "content": "Text Encoding Stream. Treating the sketch primitive sequence as textual data, we first tokenize this sequence to derive a series of text tokens " + }, + { + "bbox": [ + 313, + 410, + 482, + 517 + ], + "type": "inline_equation", + "content": "T = t_{1}, t_{2}, \\ldots, t_{l}" + }, + { + "bbox": [ + 313, + 410, + 482, + 517 + ], + "type": "text", + "content": ". These tokens are then processed using the encoder of a pre-trained language model, effectively encoding the sketch information in a textual format." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 519, + 482, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 519, + 482, + 543 + ], + "spans": [ + { + "bbox": [ + 314, + 519, + 482, + 543 + ], + "type": "text", + "content": "Text-grounded Vision Decoding Stream. This stream is dedicated" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 544, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 639 + ], + "type": "text", + "content": "to constructing the autocompleted pixels, mapping the fused image and text representation to the pixel-level sketch image. The input for the vision decoder comprises the fused image and text embeddings, augmented with adaptively masked token patches to keep the same patch size as the input image. This construction process ensures that the vision embedding is grounded by textual information and also enhances the text representation. Note that for CAD autoconstraint task, we don't reconstruct the image since the input sketches are complete sketch images already." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "Image-grounded Text Decoding Stream. In alignment with Equations 1 and 2, we conceptualize autocompletion as a conditional text generation process." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "text", + "content": "Abbreviated paper title" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "Within this framework, the combined image and text embeddings serve as inputs to the decoder of a language model. This decoder is capable of autoregressively generating the required output. Notably, the parameters of this language model are initialized from a pre-trained model, which itself has undergone extensive pre-training on vast datasets of code [27], enhancing its effectiveness in this context." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 208, + 261, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 208, + 261, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 208, + 261, + 220 + ], + "type": "text", + "content": "4.2 Training Objectives" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 230, + 481, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 230, + 481, + 254 + ], + "spans": [ + { + "bbox": [ + 130, + 230, + 481, + 254 + ], + "type": "text", + "content": "As shown in Figure 2, we jointly optimize three distinct objectives during training: one focused on understanding and two on generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 255, + 482, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 482, + 374 + ], + "type": "text", + "content": "Image-Text Contrastive Loss (ITC) is employed to activate the unimodal encoders, following the approach in [12]. This loss function aligns the visual and textual feature spaces by ensuring that positive image-text pairs have similar representations while contrasting with negative pairs. Such a contrastive approach has been instrumental in enhancing multimodal model training [13,20], not only improving the alignment of visual and language features but also accelerating the convergence during training. The image-text contrastive loss is defined as the cross-entropy between the similarity " + }, + { + "bbox": [ + 130, + 255, + 482, + 374 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 255, + 482, + 374 + ], + "type": "text", + "content": " of image(I) text(T) embedding pairs and ground-truth one-hot similarity " + }, + { + "bbox": [ + 130, + 255, + 482, + 374 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 255, + 482, + 374 + ], + "type": "text", + "content": " (array from [0,N], N is the number of samples), formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 195, + 384, + 481, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 384, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 195, + 384, + 481, + 426 + ], + "type": "interline_equation", + "content": "p = \\frac {\\exp (s (I , T))}{\\sum_ {i , j} ^ {N} \\exp (s (I _ {i} , T _ {j}))}, \\mathcal {L} _ {I T C} = \\text {C r o s s - E n t r o p y} (y, p). \\tag {5}", + "image_path": "8fc7b3b1a85ce26ed6b0fbfbf504f3731ecdd7f01f82ef4100f1adfd86059074.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 435, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 435, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 131, + 435, + 482, + 509 + ], + "type": "text", + "content": "Image Decoding Loss (IDL) is responsible for constructing autocompletion sketch images. This process involves translating the fused visual and textual information into pixel values for each patch. IDL employs a mean squared error (MSE) loss, similarly to the approach in ViTMAE [9], comparing the autocompleted sketch images with their ground-truth counterparts at the pixel level. The Image decoding loss can be formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 253, + 518, + 481, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 518, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 253, + 518, + 481, + 548 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I D L} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(Y _ {i} - \\hat {Y} _ {i}\\right) ^ {2}, \\tag {6}", + "image_path": "dc9c001a7bbf82290703655c6c6db94307a71ef25c66938ffcbbb4f58d67fa24.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 557, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 557, + 482, + 604 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 557, + 482, + 604 + ], + "type": "text", + "content": " is the number of total pixels, " + }, + { + "bbox": [ + 130, + 557, + 482, + 604 + ], + "type": "inline_equation", + "content": "Y_{i}" + }, + { + "bbox": [ + 130, + 557, + 482, + 604 + ], + "type": "text", + "content": " is the ground-truth pixel value of " + }, + { + "bbox": [ + 130, + 557, + 482, + 604 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 557, + 482, + 604 + ], + "type": "text", + "content": "-th pixel, and " + }, + { + "bbox": [ + 130, + 557, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\hat{Y}_i" + }, + { + "bbox": [ + 130, + 557, + 482, + 604 + ], + "type": "text", + "content": " is the output pixel value correspondingly. For CAD autoconstraint tasks, since the input is a full sketch image, we ignore IDL during the training of CAD autoconstraint models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Language Modeling Loss (LM) is designed to generate text-like autocompletion sketches, focusing on constraint sequences and utilizing the image-grounded text decoder. It optimizes a cross-entropy loss to maximize the likelihood described in Equation ①. LM endows the model with the capability to transform fused multimodal information into coherent autocompletion sketch" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "S. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "sequences. For the CAD autoconstraint task, we use Equation 2 as the language modeling loss." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 140, + 378, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 140, + 378, + 152 + ], + "spans": [ + { + "bbox": [ + 146, + 140, + 378, + 152 + ], + "type": "text", + "content": "Therefore, the full training objective of CadVLM is:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 249, + 164, + 481, + 176 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 164, + 481, + 176 + ], + "spans": [ + { + "bbox": [ + 249, + 164, + 481, + 176 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {I T C} + \\mathcal {L} _ {L M} + \\mathcal {L} _ {I D L}. \\tag {7}", + "image_path": "9cd98220d94ca5b5c191dad27670da148aa60227e9d1efbc2b44c8a24c0845b7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 182, + 479, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 479, + 206 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 479, + 206 + ], + "type": "text", + "content": "During inference, given incomplete sketch image and primitive pairs, we only evaluate the generated text and use that for further mechanical applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 224, + 230, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 224, + 230, + 239 + ], + "spans": [ + { + "bbox": [ + 131, + 224, + 230, + 239 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 250, + 479, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 250, + 479, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 479, + 285 + ], + "type": "text", + "content": "We conduct extensive experiments and ablation studies on CadVLM for the CAD autocompletion and autoconstraint tasks. Quantitative assessment is conducted by measuring Entity Accuracy, Sketch Accuracy, and CAD F1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 304, + 271, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 304, + 271, + 316 + ], + "spans": [ + { + "bbox": [ + 131, + 304, + 271, + 316 + ], + "type": "text", + "content": "5.1 Experimental settings" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 323, + 480, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 480, + 418 + ], + "type": "text", + "content": "Dataset We use the SketchGraphs [22] dataset for all our experiments, consisting of 15 million CAD sketches collected from Onshape. Referring to [23], we use the filtered version and further deduplicate the dataset. After deduplicated, the training, validation, and test set have 626,236, 22,031, 21,979 unique sketches correspondingly. The filtered sketches are restricted to sketches comprised of arcs, circles, and lines. To ensure a fair comparison, we normalize each sketch as in [23]; specifically the quantized range of numerical tokens in the primitive sequence is [1, 64]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 419, + 480, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 480, + 502 + ], + "type": "text", + "content": "Baselines We evaluate several ablation and conditional variants of CadVLM, as well as three SOTA CAD generative model baselines: Vitruvion [23], Deepcad [29] and SketchGraphs [22]. Since SketchGraphs does not incorporate continuous primitive parameters such as coordinate value, we only compare with the CAD autoconstraint task. And DeepCAD only considers parametrized command sequences, which are incompatible with CAD constraints. We primarily evaluate its performance on CAD autocompletion tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 503, + 481, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 481, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 481, + 647 + ], + "type": "text", + "content": "Settings For the image encoder and decoder, we choose the encoder and decoder of state-of-the-art pre-trained ViT-MAE-base from [9]. Following [9] we set the patch size and stride both as 32. We render all sketch images with size of " + }, + { + "bbox": [ + 130, + 503, + 481, + 647 + ], + "type": "inline_equation", + "content": "224 \\times 224 \\times 3" + }, + { + "bbox": [ + 130, + 503, + 481, + 647 + ], + "type": "text", + "content": ". For the text encoder and text decoder, we use the encoder and decoder of pre-trained CodeT5+ [27] correspondingly, which is an instruction-tuned code LLM with 770M parameters trained on various code datasets. We train the autocompletion model and autoconstraint model both for 30 epochs, using the Adam optimizer with decoupled weight decay regularization [17], with the learning rate set according to the CosineAnnealingLR [16] scheduler. The initial learning rate is set to " + }, + { + "bbox": [ + 130, + 503, + 481, + 647 + ], + "type": "inline_equation", + "content": "3e - 4" + }, + { + "bbox": [ + 130, + 503, + 481, + 647 + ], + "type": "text", + "content": ", with batch size 32. All our experiments are run using 8 NVIDIA A100 GPUs, and our models take between 30-40 hours to train. In the CAD Autocompletion experiments, we randomly sample " + }, + { + "bbox": [ + 130, + 503, + 481, + 647 + ], + "type": "inline_equation", + "content": "20\\% -80\\%" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "text", + "content": "Abbreviated paper title" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 652, + 252, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 652, + 252, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 652, + 252, + 665 + ], + "type": "text", + "content": "3 https://www.onshape.com" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 199, + 179, + 413, + 309 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": "Table 1: Quantitative sketch generative model comparison results for CAD autocompletion task. Ske-Acc means the probability of generating the correct full remaining sketch and Ent-Acc means the probability of generating at least one entity of the ground truth. Bold shows the best results of each metric. For all metrics the higher the better." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 199, + 179, + 413, + 309 + ], + "lines": [ + { + "bbox": [ + 199, + 179, + 413, + 309 + ], + "spans": [ + { + "bbox": [ + 199, + 179, + 413, + 309 + ], + "type": "table", + "html": "
ModelSke-AccEnt-AccCAD-F1
Vitruvion3.0%40.7%11.3%
Deepcad1.2%23.1%8.9%
ChatGPT-15%6.8%41.3%21.2%
ChatGPT-7%7.6%43.8%22.2%
ChatGPT-1%6.5%43.7%22.5%
CadVLM-Text21.2%66.7%42.8%
CadVLM-w/o-IDL22.6%68.4%44.2%
CadVLM-w/o-ITC22.8%67.6%43.9%
CadVLM-w/o-IDL&ITC23.0%67.7%44.1%
CadVLM23.8%68.3%45.2%
", + "image_path": "71facc16a8f0f9ebf2533a6e898c45b689acc633060a28f6b1db10746ab9c394.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 211, + 350, + 401, + 426 + ], + "blocks": [ + { + "bbox": [ + 130, + 318, + 481, + 341 + ], + "lines": [ + { + "bbox": [ + 130, + 318, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 481, + 341 + ], + "type": "text", + "content": "Table 2: Quantitative sketch generative model comparison results for CAD autoconstraint task. For all metrics higher is better." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 211, + 350, + 401, + 426 + ], + "lines": [ + { + "bbox": [ + 211, + 350, + 401, + 426 + ], + "spans": [ + { + "bbox": [ + 211, + 350, + 401, + 426 + ], + "type": "table", + "html": "
ModelSke-AccEnt-AccCAD-F1
SketchGraphs0.48%93.0%42.0%
Vitruvion10.7%94.6%71.6%
CadVLM-Text11.5%95.0%72.6%
CadVLM-w/o-ITC12.1%95.4%74.5%
CadVLM11.3%94.9%73.8%
", + "image_path": "1f5b17cd0bd8d9226804305d089bc3ef773691c94a3630e8c3cbf58944d8015f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 481, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 481, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 481, + 473 + ], + "type": "text", + "content": "of sketch entities as input and aim to predict the remaining entities from the ground truth sketch." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 491, + 481, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 481, + 503 + ], + "type": "text", + "content": "5.2 Quantitative comparison with baseline sketch generative models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "content": "Table 1 reports results on the CAD Autocompletion task. We find that Vitruvion achieves " + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "content": " on Sketch Accuracy and " + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "inline_equation", + "content": "40.7\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "content": " on Entity Accuracy, meaning it can accurately predict " + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "content": " of sketches and correctly identifies at least one entity in " + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "inline_equation", + "content": "40.7\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "content": " of sketches. Deepcad performs worst since it only has 4 layers of encoder and decoder, makes it hard to represent complex CAD sketches. Our CadVLM outperforms Virruvion on Sketch Accuracy, Entity Accuracy, and CAD-F1 by " + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "inline_equation", + "content": "20.8\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "inline_equation", + "content": "28.3\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "inline_equation", + "content": "39.9\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "content": " respectively. These results demonstrate the effectiveness of pre-trained foundational models and image modality. A closer examination of the qualitative results in Figure 4 reveals that Vitruvion produces many invalid results missing symmetric curves or lines. Sketches from CadVLM are better in terms of quality with more complex shapes and stronger symmetry." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "ChatGPT Series Model Finetune To evaluate the ability of a general LLM on the CAD autocompletion task, we fine-tuned ChatGPT on three subsets" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "S. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 479, + 235 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 479, + 235 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 479, + 235 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 479, + 235 + ], + "type": "image", + "image_path": "6fa3e0379ea4f351096e7c56ee38ffb02db6dec4adc16eb015e6c7fe2feb043e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 242, + 482, + 285 + ], + "lines": [ + { + "bbox": [ + 130, + 242, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 482, + 285 + ], + "type": "text", + "content": "Fig. 4: Comparative Analysis of Autocompletion in CAD Design. Top row: Random samples from the SketchGraphs test dataset. Second row: Initial input entities serving as the primer for the autocompletion task. Third and fourth row: Autocompletion results produced by Vitruvion and CadVLM." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 312, + 482, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 385 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 130, + 312, + 482, + 385 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 130, + 312, + 482, + 385 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 312, + 482, + 385 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 130, + 312, + 482, + 385 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 312, + 482, + 385 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 312, + 482, + 385 + ], + "type": "text", + "content": " of the SketchGraphs dataset. The results indicate that the ChatGPT-series outperforms Vitruvion on all three metrics, especially on Sketch Accuracy and CAD-F1. However, there were no significant performance improvements when finetuning on larger portions of the training dataset. One possible reason is we only finetune ChatGPT for less than 5 epochs due to training costs, making it hard for ChatGPT to adapt to CAD data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 387, + 484, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 484, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 484, + 517 + ], + "type": "text", + "content": "CadVLM variants To better understand the source of performance gains, we conducted a comparative analysis of different variants of our proposed CadVLM: CadVLM-text, CadVLM-w/o-ITC, CadVLM-w/o-IDL&ITC, and CadVLM w/o-IDL. CadVLM-text relies solely on the sketch primitive sequence text as its input, optimizing for language modeling loss. In contrast, CadVLM-w/o-IDL utilizes ITC and LM loss for its training objectives, CadVLM-w/o-ITC focuses on IDL and LM, and CadVLM-w/o-IDL&ITC prioritizes only the LM objective. The findings reveal that CadVLM-Text underperforms compared to all other CadVLM variants across all three metrics, which underlines the value of the sketch image modality in providing diverse geometric information alongside sequence text." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 520, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 482, + 579 + ], + "type": "text", + "content": "Interestingly, CadVLM-w/o-IDL emerges as the superior performer among the CadVLM variants, highlighting the significance of aligning image and text embeddings in multimodal modeling. Incorporating Image Decoding Loss (IDL) further enhances the alignment between image and text embeddings, leading CadVLM to show a notable improvement in all three metrics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "For the CAD Autoconstraint task, as shown in Table 2, SketchGraphs struggles to predict the whole sketch correctly, with only " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "0.48\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " Sketch Accuracy. This suggests that relying solely on graph structures of sketches is insufficient for handling long sequences of sketch primitives. In contrast, Vitruvion achieves " + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "inline_equation", + "content": "10.7\\%" + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": " on Sketch Accuracy, showing the necessity to model primitives. For the CAD Autoconstraint task it is less challenging to predict at least one constraint correctly but more challenging to accurately predict all constraints, easily caused" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "text", + "content": "Abbreviated paper title" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 247, + 202 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 247, + 202 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 247, + 202 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 247, + 202 + ], + "type": "image", + "image_path": "7b6a304b5cd51cb43d1faab4160f0cb561893f81784c2c80f1b338c5e815da9f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 205, + 197, + 217 + ], + "lines": [ + { + "bbox": [ + 184, + 205, + 197, + 217 + ], + "spans": [ + { + "bbox": [ + 184, + 205, + 197, + 217 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 250, + 114, + 364, + 202 + ], + "blocks": [ + { + "bbox": [ + 250, + 114, + 364, + 202 + ], + "lines": [ + { + "bbox": [ + 250, + 114, + 364, + 202 + ], + "spans": [ + { + "bbox": [ + 250, + 114, + 364, + 202 + ], + "type": "image", + "image_path": "105a66c6699583199a1cfd7681410c6dd78e0e9b628d477664e19c86ffa84cb5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 300, + 205, + 313, + 217 + ], + "lines": [ + { + "bbox": [ + 300, + 205, + 313, + 217 + ], + "spans": [ + { + "bbox": [ + 300, + 205, + 313, + 217 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 225, + 482, + 248 + ], + "lines": [ + { + "bbox": [ + 131, + 225, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 131, + 225, + 482, + 248 + ], + "type": "text", + "content": "Fig. 5: Effect of input entity ratio to (a) Entity Accuracy, (b) Sketch Accuracy, and (c) CAD F1." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 366, + 114, + 480, + 202 + ], + "blocks": [ + { + "bbox": [ + 366, + 114, + 480, + 202 + ], + "lines": [ + { + "bbox": [ + 366, + 114, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 366, + 114, + 480, + 202 + ], + "type": "image", + "image_path": "7840ffc800e1712acb1ca19b97e73bfe705811115cb05d3fd5384d80cfee2d9f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 206, + 429, + 217 + ], + "lines": [ + { + "bbox": [ + 417, + 206, + 429, + 217 + ], + "spans": [ + { + "bbox": [ + 417, + 206, + 429, + 217 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 269, + 479, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 479, + 304 + ], + "type": "text", + "content": "by potential data labeling inconsistencies. As is well-known, minor discrepancies in primitive values can drastically alter constraints, such as changing them from parallel to coincident." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 305, + 480, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 480, + 413 + ], + "type": "text", + "content": "The CadVLM series model exceeds the performance of both SketchGraphs and Vitruvion across all metrics. CadVLM-Text surpasses CadVLM on Ske-Acc and Ent-Acc since the pretraining task of ViTMAE has a great discrepancy with CAD autoconstraint tasks. Therefore aligning the image and text representation directly might not be an optimal way for CAD autoconstraint task.Notably, CadVLM-w/o-ITC emerges as the top performer within the CadVLM series. This may be attributed to its more substantial text decoder and utilization of multimodal information, which enables the model to more effectively integrate knowledge from both primitives and images to inform geometric information." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 154, + 463, + 324, + 559 + ], + "blocks": [ + { + "bbox": [ + 146, + 421, + 482, + 455 + ], + "lines": [ + { + "bbox": [ + 146, + 421, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 146, + 421, + 482, + 455 + ], + "type": "text", + "content": "Table 3: Quantitative results for image condi- Table 4: Effect of adding the inductioned primitive generation. Bold represents tive bias of entity-level embeddings the best results of all metrics. to the sketch sequence." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 154, + 463, + 324, + 559 + ], + "lines": [ + { + "bbox": [ + 154, + 463, + 324, + 559 + ], + "spans": [ + { + "bbox": [ + 154, + 463, + 324, + 559 + ], + "type": "table", + "html": "
Conditioned -ImageModelEnt-AccCAD-F1
PreciseVitruvion83.8%35.8%
CadVLM98.0%62.7%
Hand-drawnVitruvion84.2%38.4%
CadVLM99.4%60.1%
Noisy hand-drawnVitruvion83.2%35.8%
CadVLM94.4%56.5%
", + "image_path": "da36c649555c8ca72f2a5df89b445bb03b0ba114367712f74fb0ea2d7315e918.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 338, + 463, + 473, + 540 + ], + "blocks": [ + { + "bbox": [ + 338, + 463, + 473, + 540 + ], + "lines": [ + { + "bbox": [ + 338, + 463, + 473, + 540 + ], + "spans": [ + { + "bbox": [ + 338, + 463, + 473, + 540 + ], + "type": "table", + "html": "
ModelCadVLM w/-Inductive -TextBias
Ent-Acc66.7%67.5%
Ske-Acc21.2%22.2%
CAD-F142.8%43.7%
", + "image_path": "97eae3def93c4d7d69d52b261ef73458b81980fa623df491d04cd8fd8395019e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 575, + 239, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 575, + 239, + 588 + ], + "spans": [ + { + "bbox": [ + 131, + 575, + 239, + 588 + ], + "type": "text", + "content": "5.3 Ablation Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 594, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 628 + ], + "type": "text", + "content": "We conducted several ablation studies to assess the performance of CadVLM on image-conditioned primitive generation tasks and investigate the effectiveness of each component within CadVLM." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Image Conditional Generation. Following [23], we conduct a series of experiments to evaluate the effectiveness of CadVLM modeling primitive sequences conditioned on the images of sketches. As designers may begin their" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "S. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 152 + ], + "type": "text", + "content": "designs with hand-drawn sketches before transitioning to CAD programs, accurately inferring parametric CAD from images or scans can dramatically reduce the effort of translating from rough paper sketches to the CAD model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 154, + 481, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 154, + 481, + 202 + ], + "spans": [ + { + "bbox": [ + 130, + 154, + 481, + 202 + ], + "type": "text", + "content": "When conditioning only on sketch images, the text encoder of CadVLM is ignored. We only use the image encoder, image decoder, and text decoder. The overall model is trained to maximize the probability of the ground truth primitives portrayed in the image, conditioned on a sequence of patch embeddings." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 204, + 482, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 204, + 482, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 482, + 324 + ], + "type": "text", + "content": "The results of Table 3 show three kinds of sketch image rendering input to imitate different settings in the real world: precise renderings, rendering from the hand-drawn simulator of [23], and renderings with random affine augmentations of the hand-drawn simulator. As both Vitruvion and CadVLM can seldom predict the sketch exactly correctly, we ignore Sketch Accuracy in our image conditional generation experiments. The results show that CadVLM outperforms Vitruvion on both Entity Accuracy and CAD F1. Vitruvion performs best on hand-drawn images with " + }, + { + "bbox": [ + 130, + 204, + 482, + 324 + ], + "type": "inline_equation", + "content": "83.8\\%" + }, + { + "bbox": [ + 130, + 204, + 482, + 324 + ], + "type": "text", + "content": " Entity Accuracy and " + }, + { + "bbox": [ + 130, + 204, + 482, + 324 + ], + "type": "inline_equation", + "content": "38.4\\%" + }, + { + "bbox": [ + 130, + 204, + 482, + 324 + ], + "type": "text", + "content": " CAD F1. While CadVLM achieves almost " + }, + { + "bbox": [ + 130, + 204, + 482, + 324 + ], + "type": "inline_equation", + "content": "99.4\\%" + }, + { + "bbox": [ + 130, + 204, + 482, + 324 + ], + "type": "text", + "content": " Entity Accuracy and " + }, + { + "bbox": [ + 130, + 204, + 482, + 324 + ], + "type": "inline_equation", + "content": "62.7\\%" + }, + { + "bbox": [ + 130, + 204, + 482, + 324 + ], + "type": "text", + "content": " CAD F1. Conditioning on precise rendered images achieves the best performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 326, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 482, + 469 + ], + "type": "text", + "content": "Entity Level Modelling. As described earlier, each CAD sketch is comprised of multiple entities. We considered adding this inductive bias to the model by processing each entity separately and then combining their embeddings with the initial sequence of all sketch tokens. For this purpose, we first pretend a special token " + }, + { + "bbox": [ + 130, + 326, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\langle \\text{ENTITY} \\rangle" + }, + { + "bbox": [ + 130, + 326, + 482, + 469 + ], + "type": "text", + "content": " to each of the entity input sequences. Subsequently, the entities from all the sketches in the batch are fed into the model's encoder in parallel. Next, the embeddings of each entity are gathered for each sketch. For each sketch, we concatenate these embeddings with the original sequence tokens using another special token " + }, + { + "bbox": [ + 130, + 326, + 482, + 469 + ], + "type": "inline_equation", + "content": "\\langle \\text{TOKEN} \\rangle" + }, + { + "bbox": [ + 130, + 326, + 482, + 469 + ], + "type": "text", + "content": " as a delimiter for the second part of the sequence. The decoder remains unchanged. As shown in Table 4, incorporating this inductive bias increases computation and memory demands while only marginally enhancing the performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "text", + "content": "Impact of different input entity ratios for CAD autocomblection. To explore the impact of different input entity ratios for CAD autocomblection, we further test CAD autocomblection results with different input entity ratios: " + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "text", + "content": ". Compared with the experiments in Sec 5.4, we specify the sampled ratio of input entities. As shown in Figure 5, CadVLM outperforms Vitruvion on all input entity ratios. In particular, CadVLM achieves " + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "inline_equation", + "content": "73.8\\%" + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "text", + "content": " Entity Accuracy, " + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "inline_equation", + "content": "49.5\\%" + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "text", + "content": " Sketch Accuracy, and " + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "inline_equation", + "content": "62.7\\%" + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "text", + "content": " CAD F1 when given " + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "text", + "content": " of entities as input. For the hardest setting when given only " + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "text", + "content": " of entities as input, CadVLM can still achieve " + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "inline_equation", + "content": "43.9\\%" + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "text", + "content": " Entity Accuracy and " + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "inline_equation", + "content": "16.9\\%" + }, + { + "bbox": [ + 130, + 472, + 482, + 592 + ], + "type": "text", + "content": " CAD-F1, which proves that CadVLM has great potential to be used in real world settings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Effect of using different language models for text encoder and decoder. To better compare the performance of different text encoder-decoder, we experiment using different pre-trained foundation models as text encoder and decoder. To eliminate the influence of the image encoder and decoder, we only consider the CadVLM-Text variant. For the choice of text encoder and decoder, we also compare the performance of different Encoder-Decoder pre-trained" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "text", + "content": "Abbreviated paper title" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 200, + 157, + 411, + 253 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 5: Effect of using different architectures for CadVLM-Text. CodeT5+ 770M is the encoder-decoder for text we use in our main experiments. CodeT5+\\* is the result of training from scratch rather than pre-trained." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 200, + 157, + 411, + 253 + ], + "lines": [ + { + "bbox": [ + 200, + 157, + 411, + 253 + ], + "spans": [ + { + "bbox": [ + 200, + 157, + 411, + 253 + ], + "type": "table", + "html": "
Model#ParameterEnt-AccSke-AccCAD-F1
CodeT5+220M66.5%21.5%42.9%
CodeT5+*220M65.0%18.9%40.3%
CodeT5+770M66.7%21.2%42.8%
ByT5-base580M66.9%22.0%43.1%
ByT5-large1.2B66.8%20.6%41.9%
T5-base223M66.2%20.2%41.7%
T5-large783M66.8%21.2%42.6%
", + "image_path": "eb803a6f3d0f2c516d9af618956509a1500a66da2d1246d08c4230b1a770cdd9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 277, + 482, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 386 + ], + "type": "text", + "content": "language models on CAD autocombpletion. The T5 family of models are among the most recognized and widely used encoder-decoder models in the open-source community. Therefore, we experimented with T5 [21], ByT5 [32] and CodeT5+ [27]. The results are shown in Table 5. The effect of model size is clear for T5 and CodeT5+ but not ByT5. Also, The cost-benefit of solely making the model bigger becomes less favorable over time as the improvements in performance start to plateau. Another observation is that, because of the nature of the data which is closer to a code sequence than natural language, CodeT5+, which is fine-tuned on code, exhibits better performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 386, + 482, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 458 + ], + "type": "text", + "content": "We also trained CodeT5+ from scratch to examine the impact of pre-training on natural language and code. As evidenced in Table5, training from scratch falls short in all three metrics. Additionally, training from scratch converged much slower than finetuning. Despite the fact that our representation of sketches as code is relatively unique and not found in natural language, the rules and token embeddings learned from text are very helpful for domain adaptation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 475, + 220, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 475, + 220, + 489 + ], + "spans": [ + { + "bbox": [ + 131, + 475, + 220, + 489 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 501, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 597 + ], + "type": "text", + "content": "In this study, we introduce CadVLM, an innovative end-to-end vision-language model tailored for modeling parametric CAD sketches, which are pivotal in modern mechanical design. CadVLM harnesses the commonsense knowledge and reasoning abilities inherent in large language models, enabling it to adeptly handle the complex geometric reasoning required for CAD sketches. Our experimental results demonstrate that CadVLM outperforms existing state-of-the-art models in both autocompletion and autoconstraint tasks, highlighting the potent enhancement that pre-trained models bring to CAD construction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 616, + 247, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 616, + 247, + 630 + ], + "spans": [ + { + "bbox": [ + 132, + 616, + 247, + 630 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "This work is supported by the Canada CIFAR AI Chair Program and the Canada NSERC Discovery Grant (RGPIN-2021-03115)." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "S. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 137, + 481, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 481, + 170 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 481, + 170 + ], + "type": "text", + "content": "1. Azerbayev, Z., Schoelkopf, H., Paster, K., Santos, M.D., McAleer, S., Jiang, A.Q., Deng, J., Biderman, S., Welleck, S.: Llemma: An open language model for mathematics. arXiv preprint arXiv:2310.10631 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 170, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 170, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 138, + 170, + 481, + 203 + ], + "type": "text", + "content": "2. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 203, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 203, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 203, + 481, + 224 + ], + "type": "text", + "content": "3. Camba, J.D., Contero, M., Company, P.: Parametric cad modeling: An analysis of strategies for design reusability. Computer-Aided Design 74, 18-31 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 224, + 481, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 481, + 255 + ], + "type": "text", + "content": "4. Chen, Q., Hu, X., Wang, Z., Hong, Y.: Medblip: Bootstrapping language-image pre-training from 3d medical images and texts. arXiv preprint arXiv:2305.10799 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 256, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 481, + 289 + ], + "type": "text", + "content": "5. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A., Padlewski, P., Salz, D., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., et al.: Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 289, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 481, + 310 + ], + "type": "text", + "content": "6. Chen, Y., Zhang, S., Han, B., He, T., Li, B.: Camml: Context-aware multimodal learner for large models. arXiv preprint arXiv:2401.03149 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 310, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 310, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 310, + 481, + 331 + ], + "type": "text", + "content": "7. Floridi, L., Chiriatti, M.: Gpt-3: Its nature, scope, limits, and consequences. Minds and Machines 30, 681-694 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 331, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 481, + 353 + ], + "type": "text", + "content": "8. Ganin, Y., Bartunov, S., Li, Y., Keller, E., Saliceti, S.: Computer-aided design as language (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 353, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 481, + 385 + ], + "type": "text", + "content": "9. He, K., Chen, X., Xie, S., Li, Y., Dólár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 385, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 385, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 385, + 481, + 418 + ], + "type": "text", + "content": "0. Jones, B.T., Hu, M., Kodnongbua, M., Kim, V.G., Schulz, A.: Self-supervised representation learning for cad. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21327-21336 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 418, + 481, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 481, + 449 + ], + "type": "text", + "content": "1. Kenton, J.D.M.W.C., Toutanova, L.K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proceedings of naacL-HLT. vol. 1, p. 2 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 449, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 481, + 482 + ], + "type": "text", + "content": "2. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 482, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 482, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 482, + 481, + 514 + ], + "type": "text", + "content": "3. Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems 34, 9694-9705 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 514, + 481, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 481, + 547 + ], + "type": "text", + "content": "4. Li, P., Guo, J., Zhang, X., Yan, D.M.: Secad-net: Self-supervised cad reconstruction by learning sketch-extrude operations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16826 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 547, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 481, + 590 + ], + "type": "text", + "content": "5. Lindsay, A., Paterson, A., Graham, I.: Identifying and quantifying inefficiencies within industrial parametric cad models. In: Advances in Manufacturing Technology XXXII: Proceedings of the 16th International Conference on Manufacturing Research. vol. 8, p. 227 (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 590, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 481, + 611 + ], + "type": "text", + "content": "6. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 611, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 633 + ], + "type": "text", + "content": "7. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "text", + "content": "8. Luo, H., Sun, Q., Xu, C., Zhao, P., Lou, J., Tao, C., Geng, X., Lin, Q., Chen, S., Zhang, D.: Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct. arXiv preprint arXiv:2308.09583 (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "text", + "content": "Abbreviated paper title" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 621 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "19. Para, W., Bhat, S., Guerrero, P., Kelly, T., Mitra, N., Guibas, L.J., Wonka, P.: Sketchgen: Generating constrained cad sketches. Advances in Neural Information Processing Systems 34, 5077-5088 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 193 + ], + "type": "text", + "content": "20. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "21. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "22. Seff, A., Ovadia, Y., Zhou, W., Adams, R.P.: Sketchgraphs: A large-scale dataset for modeling relational geometry in computer-aided design. arXiv preprint arXiv:2007.08506 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "23. Seff, A., Zhou, W., Richardson, N., Adams, R.P.: Vitruvion: A generative model of parametric cad sketches. arXiv preprint arXiv:2109.14124 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "type": "text", + "content": "24. Singhal, K., Azizi, S., Tu, T., Mahdavi, S.S., Wei, J., Chung, H.W., Scales, N., Tanwani, A., Cole-Lewis, H., Pfohl, S., et al.: Large language models encode clinical knowledge. arXiv preprint arXiv:2212.13138 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "25. Singhal, K., Tu, T., Gottweis, J., Sayres, R., Wulczyn, E., Hou, L., Clark, K., Pfohl, S., Cole-Lewis, H., Neal, D., et al.: Towards expert-level medical question answering with large language models. arXiv preprint arXiv:2305.09617 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "26. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "27. Wang, Y., Le, H., Gotmare, A.D., Bui, N.D., Li, J., Hoi, S.C.: Codet5+: Open code large language models for code understanding and generation. arXiv preprint arXiv:2305.07922 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "text", + "content": "28. Willis, K.D., Jayaraman, P.K., Lambourne, J.G., Chu, H., Pu, Y.: Engineering sketch generation for computer-aided design. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2105–2114 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "text", + "content": "29. Wu, R., Xiao, C., Zheng, C.: Deepcad: A deep generative network for computer-aided design models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6772-6782 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "text", + "content": "30. Xu, X., Jayaraman, P.K., Lambourne, J.G., Willis, K.D., Furukawa, Y.: Hierarchical neural coding for controllable cad model generation. arXiv preprint arXiv:2307.00149 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 511, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 480, + 555 + ], + "type": "text", + "content": "31. Xu, X., Willis, K.D., Lambourne, J.G., Cheng, C.Y., Jayaraman, P.K., Furukawa, Y.: Skexgen: Autoregressive generation of cad construction sequences with disentangled codebooks. In: International Conference on Machine Learning. pp. 24698-24724. PMLR (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "text", + "content": "32. Xue, L., Barua, A., Constant, N., Al-Rfou, R., Narang, S., Kale, M., Roberts, A., Raffel, C.: Byt5: Towards a token-free future with pre-trained byte-to-byte models. Transactions of the Association for Computational Linguistics 10, 291-306 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 621 + ], + "type": "text", + "content": "33. Zha, L., Zhou, J., Li, L., Wang, R., Huang, Q., Yang, S., Yuan, J., Su, C., Li, X., Su, A., et al.: Tablegpt: Towards unifying tables, nature language and commands into one gpt. arXiv preprint arXiv:2307.08674 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "S. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CaesarNeRF_ Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering/df1f2576-4362-497d-ae9b-cdaf433e83ec_content_list.json b/2024/CaesarNeRF_ Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering/df1f2576-4362-497d-ae9b-cdaf433e83ec_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0450ef1527a98e168b7fe7bf5650f64adbe34729 --- /dev/null +++ b/2024/CaesarNeRF_ Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering/df1f2576-4362-497d-ae9b-cdaf433e83ec_content_list.json @@ -0,0 +1,2078 @@ +[ + { + "type": "text", + "text": "CaesarNeRF: Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering", + "text_level": 1, + "bbox": [ + 228, + 140, + 776, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haidong Zhu $^{1,*}$", + "bbox": [ + 228, + 212, + 336, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tianyu Ding $^{2,\\ast,\\dagger}$ Ram Nevatia", + "bbox": [ + 372, + 213, + 495, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tianyi Chen $^{2}$ , Ning Liang $^{2,\\dagger}$", + "bbox": [ + 540, + 213, + 630, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ilya Zharkov2", + "bbox": [ + 676, + 213, + 772, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Southern California \n2Microsoft \n{haidongz, nevatia}@usc.edu", + "bbox": [ + 334, + 255, + 666, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{tianyuding,tiachen,zharkov,lulian}@microsoft.com", + "bbox": [ + 285, + 285, + 718, + 297 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8e8110b63c99922a92cd84756438c1580ba201780eb229980ac533d92538a796.jpg", + "image_caption": [ + "Fig.1: Novel view synthesis for novel scenes using ONE reference view on Shiny [67], LLFF [42], and MVImgNet [77] (top to bottom). Each triplet of images corresponds to the results from GNT [60] (left), CaesarNeRF (middle) and groundtruth (right)." + ], + "image_footnote": [], + "bbox": [ + 217, + 324, + 785, + 492 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Generalizability and few-shot learning are key challenges in Neural Radiance Fields (NeRF), often due to the lack of a holistic understanding in pixel-level rendering. We introduce CaesarNeRF, an end-to-end approach that leverages scene-level CAlibratEd SemAntic Representation along with pixel-level representations to advance few-shot, generalizable neural rendering, facilitating a holistic understanding without compromising high-quality details. CaesarNeRF explicitly models pose differences of reference views to combine scene-level semantic representations, providing a calibrated holistic understanding. This calibration process aligns various viewpoints with precise location and is further enhanced by sequential refinement to capture varying details. Extensive experiments on public datasets, including LLFF, Shiny, mip-NeRF 360, and MVImgNet, show that CaesarNeRF delivers state-of-the-art performance across varying numbers of reference views, proving effective even with a single reference image.", + "bbox": [ + 259, + 575, + 743, + 756 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Few-view rendering $\\cdot$ Generalizable NeRF $\\cdot$ Neural rendering", + "bbox": [ + 261, + 784, + 702, + 799 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{*}$ Equal contribution. $\\dagger$ Corresponding author. This work was done when Haidong Zhu was an intern at Microsoft.", + "bbox": [ + 232, + 811, + 640, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 143, + 356, + 161 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Rendering a scene from a novel camera position is essential in view synthesis [5,11,62]. The recent advancement of Neural Radiance Field (NeRF) [44] has shown impressive results in creating photo-realistic images from novel viewpoints. However, conventional NeRF methods are either typically scene-specific, necessitating retraining for novel scenes [15, 16, 44, 67, 74], or require a large number of reference views as input for generalizing to novel scenarios [6, 58, 60, 76]. These constraints highlight the complexity of the few-shot generalizable neural rendering, which aims to render unseen scenes from novel viewpoints with a limited number of reference images.", + "bbox": [ + 215, + 178, + 785, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generalizing NeRF to novel scenes often involves using pixel-level feature embeddings encoded from input images, as seen in existing methods [59, 76]. These methods adapt NeRF to novel scenes by separating the scene representation from the model through an image encoder. However, relying solely on pixel-level features has its drawbacks: it requires highly precise epipolar geometry and often overlooks occlusion in complex scenes. Moreover, employing pixel-level features ignores the inherent interconnections within objects in the scene, treating the prediction of each pixel independently. Prior attempts to utilize scene-level representation either suffers from style mismatches [35] during scene-wide rendering or have been limited to specific object categories [23, 39, 69]. With few input reference under generalizable settings, these limitations become more pronounced, exacerbating the ambiguity in predictions due to biases from camera viewpoints. Although recent diffusion-based models [18, 56] attempt to address these issues through generative approaches, they struggle to effectively use the input views as contextual references for specific scenes.", + "bbox": [ + 215, + 299, + 785, + 510 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We present CaesarNeRF, a method that advances the generalizability of NeRF by incorporating calibrated semantic representation. This enables rendering from novel viewpoints using as few as one input reference view, as depicted in Figure 1. Our approach combines semantic scene-level representation with per-pixel features, enhancing consistency across different views of the same scene. The encoder-generated scene-level representations capture both semantic features and biases linked to specific camera poses. When reference views are limited, these biases can introduce uncertainty in the rendered images. To counter this, CaesarNeRF integrates camera pose transformations into the semantic representation, hence the term calibrated. By isolating pose-specific information from the scene-level representation, our model harmonizes features across input views, mitigating view-specific biases and, in turn, reducing ambiguity. In addition, CaesarNeRF introduces a sequential refinement process, which equips the model with varying levels of detail needed to enhance the semantic features. Extensive experiments on datasets such as LLFF [42], Shiny [67], mip-NeRF 360 [4], and the newly released MVImgNet [77] demonstrate that CaesarNeRF outperforms current state-of-the-art methods with limited reference views available, proving effective in generalizable settings with as few as one reference view. The project can be found here.", + "bbox": [ + 215, + 511, + 785, + 767 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 240, + 768, + 539, + 782 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We introduce CaesarNeRF, which utilizes scene-level calibrated semantic representation to achieve few-shot, generalizable neural rendering. This innovation leads to coherent and high-quality renderings.", + "bbox": [ + 225, + 794, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H. Zhu and T. Ding et al.", + "bbox": [ + 271, + 114, + 421, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We integrate semantic scene context with pixel-level details, in contrast to existing methods that rely solely on pixel-level features. We also address view-specific biases by modeling camera pose transformations and enhance the scene understanding through the sequential refinement of semantic features.", + "bbox": [ + 225, + 146, + 785, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We demonstrate through extensive experiments that CaesarNeRF consistently outperforms state-of-the-art generalizable NeRF methods across a variety of datasets. Furthermore, integrating the Caesar pipeline into other baseline methods leads to consistent performance gains, highlighting its effectiveness and adaptability.", + "bbox": [ + 225, + 207, + 785, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 292, + 362, + 308 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Radiance Field (NeRF) implicitly captures the density and appearance of points within a scene or object [43, 44] and enables rendering from novel camera positions. In recent years, NeRF has witnessed improvements in a wide range of applications, such as photo-realistic novel view synthesis for large-scale scenes [40,68,80], dynamic scene decomposition and deformation [24,30,34,46-48,50,84,85], occupancy or depth estimation [66,71,83], scene generation and editing [1,21,31,33,41,49,72,73,79], and so on. Despite these advances, most methods still rely on the original NeRF and require retraining or fine-tuning for novel scenes not covered in the training data.", + "bbox": [ + 212, + 325, + 787, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generalizable NeRF aims to adapt a single NeRF model to multiple scenes by separating the scene representation from the model. This field has seen notable advancements, with efforts focused on avoiding the need for retraining [8,20,38,64,65,75,76]. PixelNeRF [76] and GRF [59] pioneered the application of an image encoder to transform images into per-pixel features, with NeRF functioning as a decoder for predicting density and color from these features. MVSNeRF [6] introduces the use of a cost volume from MVSNet [75] to encode 3-D features from multiple views. Recognizing the intrinsic connection between points along a ray, IBRNet [64] employs self-attention to enhance point density predictions. Transformer-based [61] networks like GNT [10,60], GeoNeRF [25], and GPNR [58] are explored as alternatives to volume rendering, concentrating on pixel and patch-level representations. Additionally, InsertNeRF [2] utilizes hypernet modules to adapt parameters for novel scenes efficiently.", + "bbox": [ + 212, + 446, + 787, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These methods primarily depend on image encoders to extract pixel-aligned features from reference views. As a result, many of them lack a comprehensive understanding of the entire scene. Furthermore, with few reference views, the features become intertwined with view-specific details, compromising the quality of the rendering results.", + "bbox": [ + 212, + 628, + 787, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Few-shot Neural Radiance Field aims to render novel views using a limited number of reference images. To this end, various methods have been developed, incorporating additional information such as normalization-flow [78], semantic constraints [17, 22], depth cues [13, 52], geometry consistency [3, 29, 45, 63, 70], and frequency content [74]. Others [6, 9] emphasize pretraining on large-scale datasets.", + "bbox": [ + 212, + 688, + 787, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While these methods offer reasonable reconstructions with few inputs, they typically still require training or fine-tuning for specific scenes. Moreover, these methods usually require at least three reference images. With fewer than three, view-specific biases lead to ambiguity, complicating the rendering. Diffusion-based [12, 35, 57, 82] and other generative methods [27, 86] have been explored for single-view synthesis or gen-", + "bbox": [ + 212, + 765, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "erative rendering, yet they are mostly limited to single-object rendering and generally fall short for complex scenes, which often result in a style change.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CaesarNeRF confronts the above challenges by leveraging calibrated semantic representations that exploit scene geometry and variations in camera viewpoints. As a result, CaesarNeRF overcomes the limitations of pixel-level features and reduces dependency on external data or extensive pretraining, delivering high-quality renderings in few-shot and generalizable settings.", + "bbox": [ + 212, + 176, + 784, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 The proposed method", + "text_level": 1, + "bbox": [ + 214, + 275, + 433, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first outline the general framework of existing generalizable NeRF in Section 3.1. Then, we present our proposed CaesarNeRF, as illustrated in Figure 2. This model integrates elements of semantic representation, calibration, and sequential refinement, detailed in Section 3.2, 3.3, and 3.4, respectively. The training objective is given in 3.5.", + "bbox": [ + 212, + 306, + 782, + 368 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 NeRF and generalizable NeRF", + "text_level": 1, + "bbox": [ + 214, + 388, + 470, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Neural Radiance Field (NeRF) [43, 44] aims to render 3D scenes by predicting both the density and RGB values at points where light rays intersect the radiance field. For a query point $\\pmb{x} \\in \\mathbb{R}^3$ and a viewing direction $\\pmb{d}$ on the unit sphere $\\mathbb{S}^2$ in 3D space, the NeRF model $\\mathcal{F}$ is defined as:", + "bbox": [ + 212, + 414, + 784, + 473 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma , \\boldsymbol {c} = \\mathcal {F} (\\boldsymbol {x}, \\boldsymbol {d}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 444, + 486, + 785, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\sigma \\in \\mathbb{R}$ and $c\\in \\mathbb{R}^3$ denote the density and the RGB values, respectively. After computing these values for a collection of discretized points along each ray, volume rendering techniques are employed to calculate the final RGB values for each pixel, thus reconstructing the image.", + "bbox": [ + 212, + 513, + 784, + 573 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, traditional NeRF models $\\mathcal{F}$ are limited by their requirement for scene-specific training, making it unsuitable for generalizing to novel scenes. To overcome this, generalizable NeRF models, denoted by $\\mathcal{F}_G$ , are designed to render images of novel scenes without per-scene training. Given $N$ reference images $\\{I_n\\}_{n=1}^N$ , an encoder-based generalizable NeRF model $\\mathcal{F}_G$ decouples the object representation from the original NeRF by using an encoder to extract per-pixel feature maps $\\{F_n\\}_{n=1}^N$ from the input images. To synthesize a pixel associated with a point $x$ along a ray in direction $d$ , it projects $\\{F_n\\}_{n=1}^N$ from nearby views and aggregates this multi-view pixel-level information using techniques such as average pooling [76] or cost volumes [6]. This results in a fused feature embedding $\\widetilde{\\mathcal{F}}$ , allowing $\\mathcal{F}_G$ to predict density $\\sigma$ and RGB values $c$ for each point along the ray, as expressed by:", + "bbox": [ + 212, + 574, + 790, + 739 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma , \\boldsymbol {c} = \\mathcal {F} _ {\\boldsymbol {G}} (\\boldsymbol {x}, \\boldsymbol {d}, \\tilde {\\boldsymbol {F}}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 750, + 785, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our method, we adopt the recently introduced fully attention-based generalizable NeRF method, GNT [60], as both the backbone and the baseline. GNT shares a similar paradigm with (2) but employs transformers [61] to aggregate pixel-level features into $\\widetilde{F}$ . It uses a view transformer to fuse projected pixel-level features from reference", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H. Zhu and T. Ding et al.", + "bbox": [ + 271, + 114, + 421, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/52879611db002d2c8b11c42bf8c0ec659f19185a257aad77a0e6e7b4e624d1c2.jpg", + "image_caption": [ + "Fig. 2: Overview of CaesarNeRF. CaesarNeRF employs a shared encoder to capture two types of features from input views, including scene-level semantic representation $\\{S_n\\}$ and pixel-level feature representation $\\{F_n\\}$ . We use the same encoder for both the scene-level semantic representation and the pixel-level embeddings. Following calibration and aggregation of $\\{S_n\\}$ from various views, we concatenate it with the pixel-level fused feature, processed by the view transformer. Subsequent use of the ray-transformer, coupled with sequential refinement, enables us to render the final RGB values for each pixel in the target view. The output features serve as the input for the next stage, indicated by matching line colors." + ], + "image_footnote": [], + "bbox": [ + 243, + 142, + 763, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "views, and a ray transformer to combine features from different points along a ray, eliminating the need for volume rendering. Further details about GNT can be found in [60]. We also demonstrate that our approach can be extended to other generalizable NeRF models, as discusses in Section 4.2.", + "bbox": [ + 212, + 465, + 787, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Scene-level semantic representation", + "text_level": 1, + "bbox": [ + 215, + 550, + 504, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Both encoder-based generalizable NeRF models [6, 59, 76] and their attention-based counterparts [58, 60] mainly rely on pixel-level feature representations. While effective, this approach restricts their capability for a holistic scene understanding, especially when reference views are scarce. This limitation also exacerbates challenges in resolving depth ambiguities between points along the rays, a problem that becomes more pronounced with fewer reference views.", + "bbox": [ + 212, + 578, + 787, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address these challenges, we introduce semantic representations aimed at enriching the scene-level understanding. We utilize a shared CNN encoder and apply a Global Average Pooling (GAP) to its $C$ -dimensional output feature map, generating $N$ global feature vectors $\\{S_n\\}_{n=1}^N$ corresponding to each input view. These feature vectors are then averaged to form a unified scene-level representation $\\mathbf{S}$ , i.e.,", + "bbox": [ + 212, + 669, + 787, + 744 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {S} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\boldsymbol {S} _ {n} \\in \\mathbb {R} ^ {C}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 421, + 757, + 785, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In GNT [60], which uses a view transformer to aggregate pixel-level features into an $L$ -dimensional vector $\\bar{F}$ , we extend this by concatenating $\\bar{F}$ with $S$ to construct a", + "bbox": [ + 215, + 809, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering", + "bbox": [ + 241, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8f7b9c12a59b68e826371afd7cffc3a1352efbfc3eaf5d12ae4cd257823bd8a8.jpg", + "image_caption": [ + "(a) \"L: green, R: red\"" + ], + "image_footnote": [], + "bbox": [ + 313, + 148, + 478, + 243 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/04577ff17977f6c7da53b81e333c8c379aa5ebe0bd6a6e937a1f70b07343c26f.jpg", + "image_caption": [ + "(b) \"L: red, R: green\"", + "Fig. 3: An illustration of conflicting semantic meanings from multiple viewpoints of the same object. When observing the cup from distinct angles, the features extracted after pooling retain spatial information but are inconsistent in the scene-level semantic understanding, leading to conflicts across various reference images after aggregation." + ], + "image_footnote": [], + "bbox": [ + 493, + 150, + 656, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "global-local embedding $\\pmb{E}$ , as formulated by:", + "bbox": [ + 214, + 369, + 517, + 383 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {E} = \\operatorname {C o n c a t} (\\widetilde {\\boldsymbol {F}}, \\boldsymbol {S}) \\in \\mathbb {R} ^ {L + C}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 395, + 785, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This combined embedding $\\pmb{E}$ is then subjected to the standard self-attention mechanism used in GNT [60]. This approach enables the scene-level semantic representation $(S)$ to integrate with per-point features $(\\widetilde{\\pmb{F}})$ , offering a more nuanced understanding at both levels. It also allows each point to selectively draw from the scene-level information. To maintain dimensional consistency across the input and output layers of multiple transformer modules, we employ a two-layer MLP to project the enhanced features back to the original dimension $L$ of the per-point embedding $\\widetilde{\\pmb{F}}$ .", + "bbox": [ + 212, + 425, + 787, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Calibration of semantic representation", + "text_level": 1, + "bbox": [ + 214, + 556, + 526, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The integration of the scene-level semantic representation $S$ , generated through simple averaging of global feature vectors as in (3), improves rendering quality. However, this approach has limitations when dealing with multiple views. As illustrated in Figure 3, viewing the same object from distinct angles may retain spatial attributes but can lead to conflicting semantic meanings. Merely averaging these global feature vectors without accounting for camera positions can result in a distorted scene-level understanding.", + "bbox": [ + 212, + 582, + 787, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To mitigate this inconsistency, we propose a semantic calibration technique using feature rotation. This adjustment aligns the semantic representation across different camera poses. Our inspiration comes from the use of camera pose projection in computing the fused pixel-level feature $\\widetilde{F}$ and is further motivated by [51], which demonstrates that explicit rotation operations in feature spaces are feasible. Unlike point clouds in [51] that inherently lack a defined canonical orientation, NeRF explicitly encodes differences between camera viewpoints, thereby enabling precise calibration between the reference and target images.", + "bbox": [ + 212, + 674, + 787, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Building on this observation, we calculate calibrated semantic representations $\\{\\widetilde{S}_n\\}_{n=1}^N$ from the $N$ original semantic representations $\\{S_n\\}_{n=1}^N$ derived from the reference views. We accomplish this by leveraging their respective rotation matrices $\\{T_n\\}_{n=1}^N$", + "bbox": [ + 212, + 792, + 807, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H. Zhu and T. Ding et al.", + "bbox": [ + 271, + 114, + 421, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8c28005788ec4022f614351b3ccdc65476c63ba2e63fa8cd2167983135731c61.jpg", + "image_caption": [ + "Fig. 4: Visualization of decoded feature maps for \"orchid\" in LLFF dataset, produced by ray transformers [60] at different stages. From left to right, the transformer stages increase in depth." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 406, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c551c24a3094aff2d59889d370cfc1e79cdc141e0a8aba3afef905b181d96f99.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 143, + 596, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bccc6953b678e6b0d380c9ca486b8b45a81d7086611ef6f84020c1be1b1903a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 145, + 784, + 253 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to model the rotational variations between each input view and the target view. The alignment of the original semantic features is performed as follows:", + "bbox": [ + 212, + 321, + 782, + 351 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\boldsymbol {S}} _ {n} = \\boldsymbol {\\mathcal {P}} \\left(\\boldsymbol {T} _ {n} \\cdot \\boldsymbol {\\mathcal {P}} ^ {- 1} \\left(\\boldsymbol {S} _ {n}\\right)\\right), \\text {w h e r e} \\boldsymbol {T} _ {n} = \\boldsymbol {T} _ {\\text {o u t}} ^ {\\mathrm {w} 2 \\mathrm {c}} \\cdot \\boldsymbol {T} _ {n} ^ {\\mathrm {c} 2 \\mathrm {w}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 323, + 358, + 785, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, $T_{n}^{\\mathrm{c2w}}$ is the inverse of the extrinsic matrix used for $I_{n}$ , and $T_{\\mathrm{out}}^{\\mathrm{w2c}}$ is the extrinsic matrix for the target view. $\\mathcal{P}(\\cdot)$ and $\\mathcal{P}^{-1}(\\cdot)$ are the flattening and inverse flattening operations, which reshape the feature to a 1D vector of shape 1-by- $C$ and a 2D matrix of shape 3-by- $\\frac{C}{3}$ , respectively.", + "bbox": [ + 212, + 386, + 782, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that for the extrinsic matrix, we consider only the top-left $3 \\times 3$ submatrix that accounts for rotation. Using GAP to condense feature maps of various sizes into a 1-by- $C$ feature vector eliminates the need for scaling parameters in the semantic representation. As a result, modeling the intrinsic matrix is unnecessary, assuming no skewing, making our approach adaptable to different camera configurations.", + "bbox": [ + 212, + 446, + 782, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With the calibrated semantic features $\\{\\widetilde{S}_n\\}_{n=1}^N$ for each reference view, we average these, similar to (3), to obtain the calibrated scene-level semantic representation $\\widetilde{S}$ , i.e.,", + "bbox": [ + 212, + 522, + 782, + 554 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\boldsymbol {S}} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\widetilde {\\boldsymbol {S}} _ {n} \\in \\mathbb {R} ^ {C}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 421, + 565, + 785, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, akin to (4), we concatenate the pixel-level fused feature $\\widetilde{F}$ with the calibrated scene-level semantic representation $\\widetilde{S}$ to form the final global-local embedding $\\widetilde{E}$ :", + "bbox": [ + 212, + 612, + 782, + 643 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\boldsymbol {E}} = \\operatorname {C o n c a t} (\\widetilde {\\boldsymbol {F}}, \\widetilde {\\boldsymbol {S}}) \\in \\mathbb {R} ^ {L + C}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 648, + 785, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This unified embedding then feeds into ray transformers, passing through standard self-attention mechanisms. In the original GNT [60], multiple view transformers and ray transformers are stacked alternately for sequential feature processing. The last ray transformer integrates features from multiple points along a ray to yield the final RGB value. We denote the corresponding feature representations at stage $k$ as $\\widetilde{F}^{(k)}$ and $\\widetilde{E}^{(k)}$ . Notably, the calibrated semantic representation $\\widetilde{S}$ remains constant across these stages.", + "bbox": [ + 212, + 676, + 782, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Sequential refinement", + "text_level": 1, + "bbox": [ + 214, + 787, + 408, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While leveraging $\\widetilde{S}$ improves consistency, a single, uniform $\\widetilde{S}$ may not be adequate for deeper layers that demand more nuanced details. In fact, we find that deeper transform-", + "bbox": [ + 212, + 806, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering", + "bbox": [ + 241, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ers capture finer details compared to shallower ones, as shown in Figure 4. To address this limitation, we introduce a sequential semantic feature refinement module that progressively enriches features at each stage. Specifically, we learn the residual $\\pmb{\\Delta}^{(k)}$ to update $\\widetilde{\\pmb{S}}$ at each stage $k$ as follows:", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\boldsymbol {S}} ^ {(k + 1)} \\leftarrow \\widetilde {\\boldsymbol {S}} ^ {(k)} + \\boldsymbol {\\Delta} ^ {(k)}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 214, + 785, + 233 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, $\\pmb{\\Delta}^{(k)}$ is calculated by first performing specialized cross-attentions between $\\widetilde{\\pmb{S}}^{(k)}$ and the original, uncalibrated per-frame semantic features $\\{S_n\\}_{n=1}^N$ (see Figure 2), followed by their summation. Our goal is to fuse information from different source views to enrich the scene-level semantic representation with features from each reference frame. With this sequential refinement, we combine $\\widetilde{\\pmb{S}}^{(k)}$ with $\\widetilde{\\pmb{F}}^{(k)}$ at each stage, yielding a stage-specific global-local embedding $\\widetilde{\\pmb{E}}^{(k)}$ , which completes our approach.", + "bbox": [ + 212, + 244, + 782, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Discussion. In scenarios with few reference views, especially when limited to just one, the primary issue is inaccurate depth estimation, resulting in depth ambiguity [12]. This compromises the quality of images when rendered from novel viewpoints. Despite this, essential visual information generally remains accurate across different camera poses. Incorporating our proposed scene-level representation improves the understanding of the overall scene layout [7], distinguishing our approach from existing generalizable NeRF models that predict pixels individually. The advantage of our approach is its holistic view; the semantic representation enriches per-pixel predictions by providing broader context. This semantic constraint ensures that fewer abrupt changes between adjacent points. Consequently, it leads to more reliable depth estimations, making the images rendered from limited reference views more plausible.", + "bbox": [ + 212, + 337, + 787, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Training objectives", + "text_level": 1, + "bbox": [ + 214, + 523, + 392, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During training, we employ three different loss functions:", + "bbox": [ + 214, + 547, + 596, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MSE loss. The Mean Square Error (MSE) loss is the standard photometric loss used in NeRF [43]. It computes the MSE between the actual and predicted pixel values.", + "bbox": [ + 212, + 563, + 785, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Central loss. Since we project the view-level features from different input camera poses to the shared target view, we introduce a central loss to ensure frame-wise calibrated semantic features $\\{\\widetilde{S}_n\\}_{n=1}^N$ are consistent when projected onto the same target view, which is defined as:", + "bbox": [ + 212, + 594, + 785, + 654 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c e n t r a l}} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\left\\| \\widetilde {\\boldsymbol {S}} _ {n} - \\widetilde {\\boldsymbol {S}} \\right\\| _ {1}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 662, + 785, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Point-wise perceptual loss. During the rendering of a batch of pixels in a target view, we inpaint the ground-truth image by replacing the corresponding pixels with the predicted ones. Then, a perceptual loss [26] is computed between the inpainted image and the target image to guide the training process at the whole-image level.", + "bbox": [ + 214, + 712, + 785, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The final loss function is formulated as follows:", + "bbox": [ + 238, + 773, + 558, + 786 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {M S E}} + \\lambda_ {1} \\mathcal {L} _ {\\text {c e n t r a l}} + \\lambda_ {2} \\mathcal {L} _ {\\text {p e r c}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 797, + 785, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Empirically, we set $\\lambda_1 = 1$ and $\\lambda_{2} = 0.001$ , following [32].", + "bbox": [ + 214, + 824, + 609, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H. Zhu and T. Ding et al.", + "bbox": [ + 271, + 114, + 421, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5c5dcf8c6dfd7bacea48c9c40cee78701af1adee7faa63022c4f5ceb314571ad.jpg", + "table_caption": [ + "Table 1: Results for generalizable scene rendering on LLFF with few reference views. GeoNeRF, MatchNeRF, and MVSNeRF necessitate variance as input, defaulting to 0 for single-image cases, hence their results are not included for 1-view scenarios." + ], + "table_footnote": [], + "table_body": "
Method1 reference view2 reference views3 reference views
PSNR (↑)LPIPS (↓)SSIM (↑)PSNR (↑)LPIPS (↓)SSIM (↑)PSNR (↑)LPIPS (↓)SSIM (↑)
PixelNeRF [76]9.320.8980.26411.230.7660.28211.240.6710.486
GPNR [58]15.910.5270.40018.790.3800.57521.570.2880.695
NeuRay [38]16.180.5840.39317.710.3360.64618.260.3100.672
GeoNeRF [25]---18.760.4730.50023.400.2460.766
MatchNeRF [8]---21.080.2720.68922.300.2340.731
MVSNeRF [6]---19.150.3360.70419.840.3140.729
IBRNet [64]16.850.5420.50721.250.3330.68523.000.2620.752
GNT [60]16.570.5000.42420.880.2510.69123.210.1780.782
Ours18.310.4350.52121.940.2240.73623.450.1760.794
", + "bbox": [ + 217, + 198, + 785, + 323 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 349, + 356, + 369 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we first discuss our experimental setups and other details required for our experiments in Section 4.1, followed by our results for different experimental settings, along with ablation studies and analysis in Section 4.2.", + "bbox": [ + 214, + 387, + 785, + 433 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental setups", + "text_level": 1, + "bbox": [ + 215, + 460, + 401, + 476 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. Firstly, following [60], we construct our training data from both synthetic and real data. This collection includes scanned models from Google Scanned Objects [14], RealEstate10K [81], and handheld phone captures [64]. For evaluation, we utilize real data encompassing complex scenes from sources such as LLFF [42], Shiny [67], and mip-NeRF 360 [4]. Additionally, we train and test our model using the recently released MVImgNet dataset [77]. We adhere to the official split, focusing on examples from the containers category, and select 2,500 scenes for training. During inference, we choose 100 scenes, using their first images as target views and the spatially nearest images as references. Since MVImgNet does not provide camera poses, we utilize COLMAP [54, 55] to deduce the camera positions within these scenes.", + "bbox": [ + 214, + 489, + 787, + 641 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation details. CaesarNeRF is built upon GNT [60], for which we maintain the same configuration, setting the ray and view transformers stack number $(K)$ to 8 for generalizable setting and 4 for single-scene setting. The feature encoder extracts bottleneck features, applies GAP, and then uses a fully connected (FC) layer to reduce the input dimension $C$ to 96. Training involves 500,000 iterations using the Adam optimizer [28], with learning rate set at 0.001 for the feature encoder and 0.0005 for CaesarNeRF, halving them every 100,000 iterations. Each iteration samples 4,096 rays from a single scene. In line with [60], we randomly choose between 8 to 10 reference views for training, and 3 to 7 views when using the MVImgNet [77].", + "bbox": [ + 214, + 642, + 787, + 777 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baseline methods. We compare CaesarNeRF with several state-of-the-art methods suited for generalizable NeRF applications, including earlier works such as MVS-NeRF [6], PixelNeRF [76], and IBRNet [64], alongside more recent ones, including GPNR [58], NeuRay [38], GNT [60], GeoNeRF [25] and MatchNeRF [8].", + "bbox": [ + 214, + 779, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/687323f487c0adf1d018ae1072e1869a4fcd2e883061f3bc5ada3e103545f12c.jpg", + "table_caption": [ + "Table 2: Results for generalizable scene rendering on Shiny with few reference views." + ], + "table_footnote": [], + "table_body": "
Method1 reference view2 reference views3 reference views
PSNR (↑)LPIPS (↓)SSIM (↑)PSNR (↑)LPIPS (↓)SSIM (↑)PSNR (↑)LPIPS (↓)SSIM (↑)
MatchNeRF [8]---20.280.2780.63620.770.2490.672
MVSNeRF [6]---17.250.4160.57718.550.3430.645
IBRNet [64]14.930.6250.40118.400.4000.59521.960.2810.710
GNT [60]15.990.5480.40020.420.3270.61722.470.2470.720
Ours17.570.4670.47221.470.2930.65222.740.2410.723
", + "bbox": [ + 217, + 170, + 784, + 256 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/f85e95249daeadc8b36fea3de9e9c1c727a77061a3471c1dcc94f2c62bdb64ed.jpg", + "table_caption": [ + "Table 3: Results for generalizable scene rendering on mip-NeRF 360 with few reference views." + ], + "table_footnote": [], + "table_body": "
Method1 reference view2 reference views3 reference views
PSNR (↑)LPIPS (↓)SSIM (↑)PSNR (↑)LPIPS (↓)SSIM (↑)PSNR (↑)LPIPS (↓)SSIM (↑)
MatchNeRF [8]---17.000.5660.39217.260.5510.407
MVSNeRF [6]---14.230.6810.36614.290.6740.406
IBRNet [64]14.120.6820.28316.240.6180.36017.700.5550.420
GNT [60]13.480.6300.31415.210.5590.37015.590.5380.395
Ours15.200.5920.35017.050.5380.40317.550.5120.430
", + "bbox": [ + 217, + 292, + 784, + 378 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Results and analysis", + "text_level": 1, + "bbox": [ + 215, + 405, + 398, + 420 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We compare results in two settings: a generalizable setting, where the model is trained on multiple scenes without fine-tuning during inference for both few and all reference view cases, and a single-scene setting where the model is trained and evaluated on just one scene. Following these comparisons, we conduct ablation studies and test the generalizability of our method with other state-of-the-art approaches.", + "bbox": [ + 212, + 429, + 785, + 503 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Generalizable rendering. In the generalizable setting, we adopt two training strategies. First, we train the model on multiple datasets as described in Section 4.1 and evaluate on LLFF [42], Shiny [67] and mip-NeRF 360 [4] datasets. In addition, the model is trained and tested on the MVImgNet [77] for object-centric generalizability.", + "bbox": [ + 212, + 505, + 785, + 565 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(a) LLFF, Shiny, and mip-NeRF 360. The results for few-reference view scenarios on these datasets are shown in Tables 1, 2 and 3, respectively. Methods like MatchNeRF [8], MVSNeRF [75], and GeoNeRF [25] require at least two reference views. On the LLFF dataset, all methods experience a performance decline as the number of views decreases. CaesarNeRF, however, consistently outperforms others across varying reference view numbers, with the performance gap becoming more significant with fewer views. For example, with 3 views, while IBRNet [64] and GNT [60] have comparable PSNRs, CaesarNeRF demonstrates a more substantial lead in LPIPS and SSIM metrics.", + "bbox": [ + 212, + 565, + 785, + 685 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Similar patterns are observed on the Shiny [67] and mip-NeRF 360 [4] datasets. We apply the highest-performing methods from the LLFF evaluations and report the results", + "bbox": [ + 212, + 685, + 785, + 715 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/12b60e7ee87f5e7b5db63d6046800fa24fdc977238f4990bbd69f3f253055863.jpg", + "table_caption": [ + "Table 4: Results on MVImgNet across varying numbers of reference views. 'C.' represents the use of calibration before averaging." + ], + "table_footnote": [], + "table_body": "
Method1 reference view2 reference views3 reference views4 reference views5 reference views
PSNR (†)LPIPS (↓)SSIM (†)PSNR (†)LPIPS (↓)SSIM (†)PSNR (†)LPIPS (↓)SSIM (†)PSNR (†)LPIPS (↓)SSIM (†)PSNR (†)LPIPS (↓)SSIM (†)
IBRNet19.140.4580.59524.380.2660.81825.530.2030.85825.990.1900.86726.120.1880.867
GNT22.220.4330.67826.940.2360.85027.410.2060.87027.510.1970.87527.510.1940.876
Ours w/oC.23.610.3710.71826.340.2740.81727.100.2280.85027.300.2100.86227.340.2030.865
Ours24.280.3340.74727.340.2150.85627.820.1900.87527.920.1810.88127.920.1790.882
", + "bbox": [ + 217, + 781, + 784, + 838 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H. Zhu and T. Ding et al.", + "bbox": [ + 271, + 114, + 419, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/696656a1160f30f0004505714c3681d5d520d23c6c930899fbec31aadfca793b.jpg", + "table_caption": [ + "Table 5: Results of per-scene optimization on LLFF, in comparison with state-of-the-art methods." + ], + "table_footnote": [], + "table_body": "
MethodLLFF [42]NeRF [44]NeX [67]GNT [60]Ours
PSNR (↑)23.2726.5027.2627.2427.64
LPIPS (↓)0.2120.2500.1790.0870.081
SSIM (↑)0.7980.8110.9040.8890.904
", + "bbox": [ + 261, + 170, + 741, + 229 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/1f5e288988be0eef1d7146547258ff5c32c3c003b3fb5d2e366708854c818db9.jpg", + "table_caption": [ + "Table 6: Results on LLFF for few-shot generalization after adapting Caesar to other baselines." + ], + "table_footnote": [], + "table_body": "
Method1 reference view2 reference views3 reference views
PSNR (↑)SSIM (↑)LPIPS (↓)PSNR (↑)SSIM (↑)LPIPS (↓)PSNR (↑)SSIM (↑)LPIPS (↓)
MatchNeRF [8]---20.590.7750.27622.430.8050.244
+ Caesar---21.550.7820.26822.980.8240.242
IBRNet [64]16.850.5070.54221.250.6850.33323.000.7520.262
+ Caesar17.760.5430.50022.390.7400.27523.670.7720.242
", + "bbox": [ + 232, + 268, + 769, + 347 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "for those that produce satisfactory outcomes with few reference views. CaesarNeRF maintains superior performance throughout. Notably, for complex datasets like mipNeRF 360 [4], which have sparse camera inputs, the quality of rendered images generally decreases with fewer available reference views. Nonetheless, CaesarNeRF shows the most robust performance compared to the other methods.", + "bbox": [ + 212, + 375, + 785, + 450 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "(b) MVImgNet. We extend our comparison of CaesarNeRF with GNT [60] and IBRNet [64] on the MVImgNet dataset, focusing on object-centric scenes, as shown in Table 4. We examine a variant of CaesarNeRF where semantic calibration is substituted with simple feature averaging from multiple frames. While the performance of all methods improves with more views, CaesarNeRF consistently outperforms GNT and IBRNet. Notably, CaesarNeRF with feature averaging surpasses GNT in 1-view case but lags with additional views, implying that the absence of calibration leads to ambiguities when rendering from multiple views.", + "bbox": [ + 212, + 450, + 787, + 571 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Per-scene optimization. Beyond the multi-scene generalizable setting, we demonstrate per-scene optimization results in Table 5. We calculate the average performance over 8 categories from the LLFF dataset [42]. CaesarNeRF consistently outperforms", + "bbox": [ + 212, + 571, + 787, + 617 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/b57f162c0320554496d88143968ead42d70864083a644364d70ab2e1e85c13c7.jpg", + "table_caption": [ + "Table 7: Ablations on the semantic representation length $\\mathbf{R}$ , sequential refinement (Seq.) and calibration (Cali.). 'Ext.' denotes the extension of per-pixel feature to a length of 64 in GNT." + ], + "table_footnote": [], + "table_body": "
Model VariationsPSNR (↑)LPIPS (↓)SSIM (↑)
Rlen.Seq.Cali.
(Baseline GNT)20.930.1850.731
Ext.20.850.1730.735
+3221.430.1520.763
+6421.490.1490.766
+9621.460.1500.766
+12821.490.1470.763
+96✓21.530.1460.770
+96✓21.510.1470.769
+96✓✓21.670.1390.781
", + "bbox": [ + 303, + 684, + 697, + 837 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering", + "bbox": [ + 241, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/858b3b8516b308fe17dc8e4c054e24dd6a64817d339484ac52acb726d2c85097.jpg", + "image_caption": [ + "(b) Using two images as reference view.", + "Fig. 5: Comparative visualization of our proposed method against other state-of-the-art methods." + ], + "image_footnote": [], + "bbox": [ + 217, + 148, + 785, + 441 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "nearly all state-of-the-art methods in the comparison, across all three metrics, showing a significant improvement over our baseline method, GNT [60].", + "bbox": [ + 212, + 532, + 784, + 561 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Adaptability. To test the adaptability of our Caesar pipeline, we apply it to two other state-of-the-art methods that use view transformers, namely MatchNeRF [8] and IBRNet [64]. We demonstrate in Table 6 that our enhancements in scene-level semantic understanding significantly boost the performance of these methods across all metrics. This indicates that the Caesar framework is not only beneficial in our CaesarNeRF, which is based on GNT [60], but can also be a versatile addition to other NeRF pipelines with view transformers to aggregate different input views.", + "bbox": [ + 212, + 565, + 787, + 671 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation analysis. We conduct ablation studies on the \"orchid\" scene from the LLFF dataset, with findings detailed in Table 7. Testing variations in representation and the impact of the sequential refinement and calibration modules, we find that increasing the latent size in GNT yields marginal benefits. However, incorporating even a modest semantic representation size distinctly improves results. The length of the semantic representation has a minimal impact on quality. Our ablation studies indicate that while sequential refinement and calibration each offer slight performance gains, their combined effect is most significant. In a single-scene context, semantic information is effectively embedded within the representation, making the benefits of individual modules subtler. Together, however, they provide a framework where sequential refinement can leverage calibrated features for deeper insights.", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H. Zhu and T. Ding et al.", + "bbox": [ + 271, + 114, + 421, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5e8f225582df2f2400796b948c79fb2f54544d41e8e0fe57e7fbb97ba084cb56.jpg", + "image_caption": [ + "Fig. 6: Depth estimation prediction using one reference view (first row) and two reference views (second row) as input from LLFF comparing CaesarNeRF with GNT." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 266 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/935270915a1cd7e5f145edcda7312b0c103867507a8548dcf4724c575cce0874.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 321, + 321, + 387 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/04c266c0e809b999741573b06e9286bba04568365e4aa0b2d0af948652131fea.jpg", + "image_caption": [ + "fern" + ], + "image_footnote": [], + "bbox": [ + 228, + 390, + 323, + 455 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c7f95fcadbe2b730ae942a8335f8543ceb5de0247b4698070d2882aeef12411e.jpg", + "image_caption": [ + "chesstable (0.45)" + ], + "image_footnote": [], + "bbox": [ + 334, + 321, + 428, + 387 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d7cbc2bf633ea45fcd2dab003a8f042a898b44f8595dd902e61b08a1c02ea3d4.jpg", + "image_caption": [ + "playground (0.46)" + ], + "image_footnote": [], + "bbox": [ + 334, + 401, + 428, + 455 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8b6c91324e0aa73e803c36ed7535388ae3d6e9cbb7637f03af1ff0901b88215a.jpg", + "image_caption": [ + "colorfountain (0.50)" + ], + "image_footnote": [], + "bbox": [ + 442, + 321, + 553, + 387 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a538e770a173f7d57c111ae225805b73394531243d0153e86fa675b6362c1a7e.jpg", + "image_caption": [ + "bikes (0.50)", + "Fig.7: Largest and smallest distances for two examples from LLFF test split when matching with training scenes. Numbers $(\\times 10^{-2})$ denote the L-2 distance to the source image within the semantic feature space." + ], + "image_footnote": [], + "bbox": [ + 450, + 401, + 544, + 455 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4afed97bc09049d0a839f5b6c8cf217b7f6cf9bdda1cc4c03c6ee454d42eb1ba.jpg", + "image_caption": [ + "fenceflower (1.91)" + ], + "image_footnote": [], + "bbox": [ + 570, + 321, + 660, + 387 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3e87864d5f85b0a1adb736dca13d4289acb30644d97beda0529ccbed111e0708.jpg", + "image_caption": [ + "peppers (2.02)" + ], + "image_footnote": [], + "bbox": [ + 568, + 401, + 661, + 455 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9e7836e5d80788e8f2ee9da19d594ab64ce94f4184259f32ff306735b7b5656d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 321, + 781, + 387 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c07c7dcd68b3781bef9cb67469986fab91e8a6c05910b020cb23a813ba757f3c.jpg", + "image_caption": [ + "pond (1.80)", + "apples (2.01)" + ], + "image_footnote": [], + "bbox": [ + 681, + 400, + 774, + 455 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Visualizations. We present our visualization results in Figure 5, where we compare our method with others using one or two views from the LLFF dataset. Additional visual comparisons are provided in the supplementary materials. These visualizations highlight that in scenarios with few views, our method significantly surpasses other generalizable NeRF models, particularly excelling when only a single view is available. CaesarNeRF demonstrates rendering with sharper boundaries and more distinct objects.", + "bbox": [ + 212, + 566, + 787, + 657 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Depth estimation. We extend our evaluation to depth prediction within the LLFF [42] dataset, focusing on challenges presented by few reference views, such as scenarios with just one or two images. In a comparison between CaesarNeRF and GNT [60], we observe in Figure 6 that GNT struggles to accurately capture the relative positions of objects when reference images are sparse. For instance, with only a single view of a flower, CaesarNeRF precisely indicates the flower's proximity to the camera compared to the leaves in the background, a distinction that GNT fails to make. Furthermore, the depth estimations provided by CaesarNeRF are consistently more reliable. In the horn example involving two views, CaesarNeRF offers better boundary delineation, showing particular strength in handling reflective surfaces such as grass in the background.", + "bbox": [ + 212, + 657, + 789, + 809 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Semantic analysis for $\\tilde{S}$ . To assess whether the calibrated semantic representation $\\tilde{S}$ truly captures semantic details of the scene, we analyze the highest and lowest", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering", + "bbox": [ + 241, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c47997a4afe51bbe83335e1b3d5cb5d3584e4b767cedb3d24cc4346a96259a24.jpg", + "image_caption": [ + "(a) leaves" + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 359, + 223 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f25a670ca80735b3c0d74ca3048f82dc6e715a6996e9e3285d0333f93be315e4.jpg", + "image_caption": [ + "Fig.8: Synthetic results for two examples from LLFF [42], “leaves” and “orchid”, using zero 1-to-3 [35] with one reference image as input and 2 degrees of vertical shift. The left image of each pair is the input, and the right one is the output of zero 1-to-3 [35]." + ], + "image_footnote": [], + "bbox": [ + 361, + 143, + 500, + 223 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/acb39d00ad69cd230a42b3ce357f84f418a987ab62c3dd9c20e1cce0df09bdf0.jpg", + "image_caption": [ + "(b) orchid" + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 638, + 223 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/7d2b2c139693508b5ec4cc740792990e53171d1bd735753f62a0e66adbdba032.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 143, + 779, + 223 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "response based on the L-2 distance between features from two different scenes, interpreting smaller distances as greater similarity. We consider the first image from each category in the LLFF dataset as a reference and extract the scene-level representation for the ten closest views to this reference image using CaesarNeRF.", + "bbox": [ + 212, + 321, + 782, + 382 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In Figure 7, we present two examples, \"room\" and \"fortress\", from the LLFF dataset, representing the first image of these categories alongside images corresponding to the top-2 highest and lowest responses. This analysis reveals that the scene-level representation predominantly emphasizes structural information and objects of similar categories. For instance, with the source image \"room\", the highest responses correlate with images featuring table-like structures, indicating an object-centric focus. In contrast, the lowest responses include images of flowers or ponds found in open areas, which diverge significantly from the structural and object content in the source images.", + "bbox": [ + 212, + 383, + 784, + 503 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Comparison with generative methods. Single-view scenarios are often addressed by generative methods [35-37] that employ diffusion models [19, 53]. While these models can produce reasonable results for object-centric renderings, they struggle with scene-level renderings from novel viewpoints. We show two LLFF examples in Figure 8 using zero 1-to-3 [35], where the left image is the input, and the right one is the output. Images rendered with zero 1-to-3 [35] suffer from style difference. Unlike NeRF-based approaches that reconstruct images from observed pixels, generative models synthesize an entire image from its semantic representation but may not maintain the style.", + "bbox": [ + 212, + 503, + 785, + 625 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion and limitation", + "text_level": 1, + "bbox": [ + 214, + 645, + 465, + 660 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce CaesarNeRF, a few-shot and generalizable NeRF pipeline that combines scene-level semantic with per-pixel feature representations, aiding in rendering from novel camera positions with limited reference views. We calibrate the semantic representations across different input views and employ a sequential refinement network to offer distinct semantic representations at various levels. Our method has been extensively evaluated on a broad range of datasets, exhibiting state-of-the-art performance in both generalizable and single-scene settings.", + "bbox": [ + 212, + 672, + 784, + 777 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations and potential negative impact. Instead of a generative approach, CaesarNeRF relies on NeRF's scheme, using input images to render the target view. This approach restricts its ability to render parts of the scene that are not present in the reference images. CaesarNeRF may create fake images for authentication.", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H. Zhu and T. Ding et al.", + "bbox": [ + 271, + 114, + 421, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 310, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bao, C., Zhang, Y., Yang, B., Fan, T., Yang, Z., Bao, H., Zhang, G., Cui, Z.: Sine: Semantic-driven image-based nerf editing with prior-guided editing field. In: CVPR. pp. 20919-20929 (2023)", + "2. Bao, Y., Ding, T., Huo, J., Li, W., Li, Y., Gao, Y.: Insertnerf: Instilling generalizability into nef with hypernet modules. ICLR (2024)", + "3. Bao, Y., Li, Y., Huo, J., Ding, T., Liang, X., Li, W., Gao, Y.: Where and how: Mitigating confusion in neural radiance fields from sparse inputs. arXiv preprint arXiv:2308.02908 (2023)", + "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In: CVPR. pp. 5470–5479 (2022)", + "5. Buehler, C., Bosse, M., McMillan, L., Gortler, S., Cohen, M.: Unstructured lumigraph rendering. In: SIGGRAPH. pp. 425-432 (2001)", + "6. Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In: ICCV. pp. 14124-14133 (2021)", + "7. Chen, W., Xu, X., Jia, J., Luo, H., Wang, Y., Wang, F., Jin, R., Sun, X.: Beyond appearance: a semantic controllable self-supervised learning framework for human-centric visual tasks. In: CVPR (2023)", + "8. Chen, Y., Xu, H., Wu, Q., Zheng, C., Cham, T.J., Cai, J.: Explicit correspondence matching for generalizable neural radiance fields. arXiv preprint arXiv:2304.12294 (2023)", + "9. Chibane, J., Bansal, A., Lazova, V., Pons-Moll, G.: Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In: CVPR. pp. 7911-7920 (2021)", + "0. Cong, W., Liang, H., Wang, P., Fan, Z., Chen, T., Varma, M., Wang, Y., Wang, Z.: Enhancing nerf akin to enhancing llms: Generalizable nerf transformer with mixture-of-view-experts. In: ICCV (2023)", + "1. Debevec, P.E., Taylor, C.J., Malik, J.: Modeling and rendering architecture from photographs: A hybrid geometry-and image-based approach. In: SIGGRAPH. pp. 11-20 (1996)", + "2. Deng, C., Jiang, C., Qi, C.R., Yan, X., Zhou, Y., Guibas, L., Anguelov, D.: Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In: CVPR. pp. 20637-20647 (2023)", + "3. Deng, K., Liu, A., Zhu, J.Y., Ramanan, D.: Depth-supervised nerf: Fewer views and faster training for free. In: CVPR. pp. 12882-12891 (2022)", + "4. Downs, L., Francis, A., Koenig, N., Kinman, B., Hickman, R., Reymann, K., McHugh, T.B., Vanhoucke, V.: Google scanned objects: A high-quality dataset of 3d scanned household items. In: ICRA. pp. 2553-2560 (2022)", + "5. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: CVPR. pp. 12479-12488 (2023)", + "6. Fu, Y., Misra, I., Wang, X.: Multiplane nef-supervised disentanglement of depth and camera pose from videos. In: ICML (2022)", + "7. Gao, Y., Cao, Y.P., Shan, Y.: Surfelnerf: Neural surfel radiance fields for online photorealistic reconstruction of indoor scenes. In: CVPR. pp. 108-118 (2023)", + "8. Gu, J., Liu, L., Wang, P., Theobalt, C.: Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985 (2021)", + "9. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS pp. 6840-6851 (2020)", + "20. Irshad, M.Z., Zakharov, S., Liu, K., Guizilini, V., Kollar, T., Gaidon, A., Kira, Z., Ambrus, R.: Neo 360: Neural fields for sparse view synthesis of outdoor scenes. In: ICCV. pp. 9187-9198 (2023)", + "21. Jain, A., Mildenhall, B., Barron, J.T., Abbeel, P., Poole, B.: Zero-shot text-guided object generation with dream fields. In: CVPR. pp. 867-876 (2022)" + ], + "bbox": [ + 225, + 176, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 769, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "22. Jain, A., Tancik, M., Abbeel, P.: Putting nerf on a diet: Semantically consistent few-shot view synthesis. In: ICCV. pp. 5885-5894 (2021)", + "23. Jang, W., Agapito, L.: Codenerf: Disentangled neural radiance fields for object categories. In: ICCV. pp. 12949-12958 (2021)", + "24. Jiang, Y., Hedman, P., Mildenhall, B., Xu, D., Barron, J.T., Wang, Z., Xue, T.: Alignerf: High-fidelity neural radiance fields via alignment-aware training. In: CVPR. pp. 46-55 (2023)", + "25. Johari, M.M., Lepoittevin, Y., Fleuret, F.: Geonerf: Generalizing nerf with geometry priors. In: CVPR. pp. 18365-18375 (2022)", + "26. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and superresolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016)", + "27. Kania, A., Kasymov, A., Zieba, M., Spurek, P.: Hypernerfgan: Hypernetwork approach to 3d nerf gan. arXiv preprint arXiv:2301.11631 (2023)", + "28. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "29. Kwak, M., Song, J., Kim, S.: Geconerf: Few-shot neural radiance fields via geometric consistency. arXiv preprint arXiv:2301.10941 (2023)", + "30. Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: Dynibar: Neural dynamic image-based rendering. In: CVPR. pp. 4273-4284 (2023)", + "31. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: CVPR. pp. 300–309 (2023)", + "32. Lin, H., Peng, S., Xu, Z., Yan, Y., Shuai, Q., Bao, H., Zhou, X.: Efficient neural radiance fields for interactive free-viewpoint video. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)", + "33. Lin, Y., Bai, H., Li, S., Lu, H., Lin, X., Xiong, H., Wang, L.: Componerf: Text-guided multi-object compositional nef with editable 3d scene layout. arXiv preprint arXiv:2303.13843 (2023)", + "34. Liu, L., Habermann, M., Rudnev, V., Sarkar, K., Gu, J., Theobalt, C.: Neural actor: Neural free-view synthesis of human actors with pose control. TOC 40(6), 1-16 (2021)", + "35. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: ICCV (2023)", + "36. Liu, X., Kao, S.h., Chen, J., Tai, Y.W., Tang, C.K.: Deceptive-nerf: Enhancing nerf reconstruction using pseudo-observations from diffusion models. arXiv preprint arXiv:2305.15171 (2023)", + "37. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453 (2023)", + "38. Liu, Y., Peng, S., Liu, L., Wang, Q., Wang, P., Theobalt, C., Zhou, X., Wang, W.: Neural rays for occlusion-aware image-based rendering. In: CVPR. pp. 7824-7833 (2022)", + "39. Mariotti, O., Mac Aodha, O., Bilen, H.: Viewnerf: Unsupervised viewpoint estimation using category-level neural radiance fields. arXiv preprint arXiv:2212.00436 (2022)", + "40. Martin-Brualla, R., Radwan, N., Sajjadi, M.S., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: CVPR. pp. 7210-7219 (2021)", + "41. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: CVPR. pp. 12663–12673 (2023)", + "42. Mildenhall, B., Srinivasan, P.P., Ortiz-Cayon, R., Kalantari, N.K., Ramamoorthi, R., Ng, R., Kar, A.: Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. TOG 38(4), 1-14 (2019)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H. Zhu and T. Ding et al.", + "bbox": [ + 271, + 114, + 421, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)", + "44. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)", + "45. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: CVPR. pp. 5480-5490 (2022)", + "46. Noguchi, A., Sun, X., Lin, S., Harada, T.: Neural articulated radiance field. In: ICCV. pp. 5762-5772 (2021)", + "47. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: ICCV. pp. 5865-5874 (2021)", + "48. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable neural radiance fields for modeling dynamic human bodies. In: ICCV. pp. 14314-14323 (2021)", + "49. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)", + "50. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: CVPR. pp. 10318-10327 (2021)", + "51. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR. pp. 652-660 (2017)", + "52. Roessle, B., Barron, J.T., Mildenhall, B., Srinivasan, P.P., Nießner, M.: Dense depth priors for neural radiance fields from sparse input views. In: CVPR. pp. 12892-12901 (2022)", + "53. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)", + "54. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: CVPR (2016)", + "55. Schonberger, J.L., Zheng, E., Pollefeys, M., Frahm, J.M.: Pixelwise view selection for unstructured multi-view stereo. In: ECCV (2016)", + "56. Schwarz, K., Liao, Y., Niemeyer, M., Geiger, A.: Graf: Generative radiance fields for 3d-aware image synthesis. NeurIPS 33, 20154-20166 (2020)", + "57. Shue, J.R., Chan, E.R., Po, R., Ankner, Z., Wu, J., Wetzstein, G.: 3d neural field generation using triplane diffusion. In: CVPR. pp. 20875-20886 (2023)", + "58. Suhail, M., Esteves, C., Sigal, L., Makadia, A.: Generalizable patch-based neural rendering. In: ECCV. pp. 156-174 (2022)", + "59. Trevithick, A., Yang, B.: Grf: Learning a general radiance field for 3d representation and rendering. In: ICCV. pp. 15182-15192 (2021)", + "60. Varma, M., Wang, P., Chen, X., Chen, T., Venugopalan, S., Wang, Z.: Is attention all that nerf needs? In: ICLR (2023)", + "61. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017)", + "62. Waechter, M., Moehrle, N., Goesele, M.: Let there be color! large-scale texturing of 3d reconstructions. In: ECCV. pp. 836-850 (2014)", + "63. Wang, G., Chen, Z., Loy, C.C., Liu, Z.: Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In: ICCV (2023)", + "64. Wang, Q., Wang, Z., Genova, K., Srinivasan, P.P., Zhou, H., Barron, J.T., Martin-Brualla, R., Snavely, N., Funkhouser, T.: Ibrnet: Learning multi-view image-based rendering. In: CVPR. pp. 4690-4699 (2021)", + "65. Wang, T., Zhang, B., Zhang, T., Gu, S., Bao, J., Baltrusaitis, T., Shen, J., Chen, D., Wen, F., Chen, Q., et al.: Rodin: A generative model for sculpting 3d digital avatars using diffusion. In: CVPR. pp. 4563-4573 (2023)", + "66. Wei, Y., Liu, S., Rao, Y., Zhao, W., Lu, J., Zhou, J.: Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In: CVPR. pp. 5610-5619 (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering", + "bbox": [ + 241, + 114, + 732, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "67. Wizadwongsa, S., Phongthawee, P., Yenphraphai, J., Suwajanakorn, S.: Nex: Real-time view synthesis with neural basis expansion. In: CVPR. pp. 8534-8543 (2021)", + "68. Xiangli, Y., Xu, L., Pan, X., Zhao, N., Rao, A., Theobalt, C., Dai, B., Lin, D.: Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In: ECCV (2022)", + "69. Xie, C., Park, K., Martin-Brualla, R., Brown, M.: Fig-nerf: Figure-ground neural radiance fields for 3d object category modelling. In: 3DV. pp. 962–971 (2021)", + "70. Xu, D., Jiang, Y., Wang, P., Fan, Z., Shi, H., Wang, Z.: Sinnerf: Training neural radiance fields on complex scenes from a single image. In: ECCV. pp. 736-753 (2022)", + "71. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: CVPR. pp. 5438-5448 (2022)", + "72. Yang, B., Zhang, Y., Xu, Y., Li, Y., Zhou, H., Bao, H., Zhang, G., Cui, Z.: Learning object-compositional neural radiance field for editable scene rendering. In: ICCV. pp. 13779-13788 (2021)", + "73. Yang, H., Hong, L., Li, A., Hu, T., Li, Z., Lee, G.H., Wang, L.: Contranerf: Generalizable neural radiance fields for synthetic-to-real novel view synthesis via contrastive learning. In: CVPR. pp. 16508-16517 (2023)", + "74. Yang, J., Pavone, M., Wang, Y.: Freenerf: Improving few-shot neural rendering with free frequency regularization. In: CVPR. pp. 8254-8263 (2023)", + "75. Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstructured multiview stereo. In: ECCV. pp. 767-783 (2018)", + "76. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: CVPR. pp. 4578-4587 (2021)", + "77. Yu, X., Xu, M., Zhang, Y., Liu, H., Ye, C., Wu, Y., Yan, Z., Zhu, C., Xiong, Z., Liang, T., et al.: Mvimgnet: A large-scale dataset of multi-view images. In: CVPR. pp. 9150-9161 (2023)", + "78. Zhang, J., Yang, G., Tulsiani, S., Ramanan, D.: Ners: Neural reflectance surfaces for sparse-view 3d reconstruction in the wild. NeurIPS 34, 29835-29847 (2021)", + "79. Zhang, J., Li, X., Wan, Z., Wang, C., Liao, J.: Text2nerf: Text-driven 3d scene generation with neural radiance fields. arXiv preprint arXiv:2305.11588 (2023)", + "80. Zhenxing, M., Xu, D.: Switch-nerf: Learning scene decomposition with mixture of experts for large-scale neural radiance fields. In: ICLR (2023)", + "81. Zhou, T., Tucker, R., Flynn, J., Fyffe, G., Snavely, N.: Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817 (2018)", + "82. Zhou, Z., Tulsiani, S.: Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In: CVPR. pp. 12588-12597 (2023)", + "83. Zhu, H., Sun, Y., Liu, C., Xia, L., Luo, J., Qiao, N., Nevatia, R., Kuo, C.H.: Multimodal neural radiance field. In: ICRA. pp. 9393-9399 (2023)", + "84. Zhu, H., Zheng, Z., Zheng, W., Nevatia, R.: Cat-nerf: Constancy-aware tx2former for dynamic body modeling. In: CVPRW. pp. 6618-6627 (2023)", + "85. Zhuang, Y., Zhu, H., Sun, X., Cao, X.: Mofanerf: Morphable facial neural radiance field. In: ECCV. pp. 268-285 (2022)", + "86. Zimny, D., Trzeciński, T., Spurek, P.: Points2nerf: Generating neural radiance fields from 3d point cloud. arXiv preprint arXiv:2206.01290 (2022)" + ], + "bbox": [ + 215, + 147, + 784, + 742 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "H. Zhu and T. Ding et al.", + "bbox": [ + 271, + 114, + 421, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/CaesarNeRF_ Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering/df1f2576-4362-497d-ae9b-cdaf433e83ec_model.json b/2024/CaesarNeRF_ Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering/df1f2576-4362-497d-ae9b-cdaf433e83ec_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f75192549f4a1df7ee200e469086e202d7352648 --- /dev/null +++ b/2024/CaesarNeRF_ Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering/df1f2576-4362-497d-ae9b-cdaf433e83ec_model.json @@ -0,0 +1,3140 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.229, + 0.141, + 0.777, + 0.187 + ], + "angle": 0, + "content": "CaesarNeRF: Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.213, + 0.338, + 0.23 + ], + "angle": 0, + "content": "Haidong Zhu\\(^{1,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.374, + 0.214, + 0.496, + 0.245 + ], + "angle": 0, + "content": "Tianyu Ding\\(^{2,\\ast,\\dagger}\\) Ram Nevatia" + }, + { + "type": "text", + "bbox": [ + 0.542, + 0.214, + 0.632, + 0.246 + ], + "angle": 0, + "content": "Tianyi Chen\\(^{2}\\), Ning Liang\\(^{2,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.678, + 0.214, + 0.773, + 0.23 + ], + "angle": 0, + "content": "Ilya Zharkov2" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.256, + 0.668, + 0.285 + ], + "angle": 0, + "content": "1University of Southern California \n2Microsoft \n{haidongz, nevatia}@usc.edu" + }, + { + "type": "text", + "bbox": [ + 0.287, + 0.286, + 0.72, + 0.299 + ], + "angle": 0, + "content": "{tianyuding,tiachen,zharkov,lulian}@microsoft.com" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.325, + 0.787, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.502, + 0.788, + 0.545 + ], + "angle": 0, + "content": "Fig.1: Novel view synthesis for novel scenes using ONE reference view on Shiny [67], LLFF [42], and MVImgNet [77] (top to bottom). Each triplet of images corresponds to the results from GNT [60] (left), CaesarNeRF (middle) and groundtruth (right)." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.576, + 0.744, + 0.757 + ], + "angle": 0, + "content": "Abstract. Generalizability and few-shot learning are key challenges in Neural Radiance Fields (NeRF), often due to the lack of a holistic understanding in pixel-level rendering. We introduce CaesarNeRF, an end-to-end approach that leverages scene-level CAlibratEd SemAntic Representation along with pixel-level representations to advance few-shot, generalizable neural rendering, facilitating a holistic understanding without compromising high-quality details. CaesarNeRF explicitly models pose differences of reference views to combine scene-level semantic representations, providing a calibrated holistic understanding. This calibration process aligns various viewpoints with precise location and is further enhanced by sequential refinement to capture varying details. Extensive experiments on public datasets, including LLFF, Shiny, mip-NeRF 360, and MVImgNet, show that CaesarNeRF delivers state-of-the-art performance across varying numbers of reference views, proving effective even with a single reference image." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.785, + 0.703, + 0.8 + ], + "angle": 0, + "content": "Keywords: Few-view rendering \\(\\cdot\\) Generalizable NeRF \\(\\cdot\\) Neural rendering" + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.812, + 0.641, + 0.84 + ], + "angle": 0, + "content": "\\(^{*}\\)Equal contribution. \\(\\dagger\\)Corresponding author. This work was done when Haidong Zhu was an intern at Microsoft." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.129 + ], + "angle": 0, + "content": "H. Zhu and T. Ding et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.357, + 0.162 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.179, + 0.787, + 0.299 + ], + "angle": 0, + "content": "Rendering a scene from a novel camera position is essential in view synthesis [5,11,62]. The recent advancement of Neural Radiance Field (NeRF) [44] has shown impressive results in creating photo-realistic images from novel viewpoints. However, conventional NeRF methods are either typically scene-specific, necessitating retraining for novel scenes [15, 16, 44, 67, 74], or require a large number of reference views as input for generalizing to novel scenarios [6, 58, 60, 76]. These constraints highlight the complexity of the few-shot generalizable neural rendering, which aims to render unseen scenes from novel viewpoints with a limited number of reference images." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.511 + ], + "angle": 0, + "content": "Generalizing NeRF to novel scenes often involves using pixel-level feature embeddings encoded from input images, as seen in existing methods [59, 76]. These methods adapt NeRF to novel scenes by separating the scene representation from the model through an image encoder. However, relying solely on pixel-level features has its drawbacks: it requires highly precise epipolar geometry and often overlooks occlusion in complex scenes. Moreover, employing pixel-level features ignores the inherent interconnections within objects in the scene, treating the prediction of each pixel independently. Prior attempts to utilize scene-level representation either suffers from style mismatches [35] during scene-wide rendering or have been limited to specific object categories [23, 39, 69]. With few input reference under generalizable settings, these limitations become more pronounced, exacerbating the ambiguity in predictions due to biases from camera viewpoints. Although recent diffusion-based models [18, 56] attempt to address these issues through generative approaches, they struggle to effectively use the input views as contextual references for specific scenes." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.512, + 0.787, + 0.768 + ], + "angle": 0, + "content": "We present CaesarNeRF, a method that advances the generalizability of NeRF by incorporating calibrated semantic representation. This enables rendering from novel viewpoints using as few as one input reference view, as depicted in Figure 1. Our approach combines semantic scene-level representation with per-pixel features, enhancing consistency across different views of the same scene. The encoder-generated scene-level representations capture both semantic features and biases linked to specific camera poses. When reference views are limited, these biases can introduce uncertainty in the rendered images. To counter this, CaesarNeRF integrates camera pose transformations into the semantic representation, hence the term calibrated. By isolating pose-specific information from the scene-level representation, our model harmonizes features across input views, mitigating view-specific biases and, in turn, reducing ambiguity. In addition, CaesarNeRF introduces a sequential refinement process, which equips the model with varying levels of detail needed to enhance the semantic features. Extensive experiments on datasets such as LLFF [42], Shiny [67], mip-NeRF 360 [4], and the newly released MVImgNet [77] demonstrate that CaesarNeRF outperforms current state-of-the-art methods with limited reference views available, proving effective in generalizable settings with as few as one reference view. The project can be found here." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.769, + 0.54, + 0.783 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- We introduce CaesarNeRF, which utilizes scene-level calibrated semantic representation to achieve few-shot, generalizable neural rendering. This innovation leads to coherent and high-quality renderings." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.147, + 0.786, + 0.207 + ], + "angle": 0, + "content": "- We integrate semantic scene context with pixel-level details, in contrast to existing methods that rely solely on pixel-level features. We also address view-specific biases by modeling camera pose transformations and enhance the scene understanding through the sequential refinement of semantic features." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.208, + 0.786, + 0.268 + ], + "angle": 0, + "content": "- We demonstrate through extensive experiments that CaesarNeRF consistently outperforms state-of-the-art generalizable NeRF methods across a variety of datasets. Furthermore, integrating the Caesar pipeline into other baseline methods leads to consistent performance gains, highlighting its effectiveness and adaptability." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.293, + 0.364, + 0.309 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.788, + 0.446 + ], + "angle": 0, + "content": "Neural Radiance Field (NeRF) implicitly captures the density and appearance of points within a scene or object [43, 44] and enables rendering from novel camera positions. In recent years, NeRF has witnessed improvements in a wide range of applications, such as photo-realistic novel view synthesis for large-scale scenes [40,68,80], dynamic scene decomposition and deformation [24,30,34,46-48,50,84,85], occupancy or depth estimation [66,71,83], scene generation and editing [1,21,31,33,41,49,72,73,79], and so on. Despite these advances, most methods still rely on the original NeRF and require retraining or fine-tuning for novel scenes not covered in the training data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.628 + ], + "angle": 0, + "content": "Generalizable NeRF aims to adapt a single NeRF model to multiple scenes by separating the scene representation from the model. This field has seen notable advancements, with efforts focused on avoiding the need for retraining [8,20,38,64,65,75,76]. PixelNeRF [76] and GRF [59] pioneered the application of an image encoder to transform images into per-pixel features, with NeRF functioning as a decoder for predicting density and color from these features. MVSNeRF [6] introduces the use of a cost volume from MVSNet [75] to encode 3-D features from multiple views. Recognizing the intrinsic connection between points along a ray, IBRNet [64] employs self-attention to enhance point density predictions. Transformer-based [61] networks like GNT [10,60], GeoNeRF [25], and GPNR [58] are explored as alternatives to volume rendering, concentrating on pixel and patch-level representations. Additionally, InsertNeRF [2] utilizes hypernet modules to adapt parameters for novel scenes efficiently." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.689 + ], + "angle": 0, + "content": "These methods primarily depend on image encoders to extract pixel-aligned features from reference views. As a result, many of them lack a comprehensive understanding of the entire scene. Furthermore, with few reference views, the features become intertwined with view-specific details, compromising the quality of the rendering results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Few-shot Neural Radiance Field aims to render novel views using a limited number of reference images. To this end, various methods have been developed, incorporating additional information such as normalization-flow [78], semantic constraints [17, 22], depth cues [13, 52], geometry consistency [3, 29, 45, 63, 70], and frequency content [74]. Others [6, 9] emphasize pretraining on large-scale datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.84 + ], + "angle": 0, + "content": "While these methods offer reasonable reconstructions with few inputs, they typically still require training or fine-tuning for specific scenes. Moreover, these methods usually require at least three reference images. With fewer than three, view-specific biases lead to ambiguity, complicating the rendering. Diffusion-based [12, 35, 57, 82] and other generative methods [27, 86] have been explored for single-view synthesis or gen-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.129 + ], + "angle": 0, + "content": "H. Zhu and T. Ding et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "erative rendering, yet they are mostly limited to single-object rendering and generally fall short for complex scenes, which often result in a style change." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.253 + ], + "angle": 0, + "content": "CaesarNeRF confronts the above challenges by leveraging calibrated semantic representations that exploit scene geometry and variations in camera viewpoints. As a result, CaesarNeRF overcomes the limitations of pixel-level features and reduces dependency on external data or extensive pretraining, delivering high-quality renderings in few-shot and generalizable settings." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.276, + 0.434, + 0.294 + ], + "angle": 0, + "content": "3 The proposed method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.784, + 0.369 + ], + "angle": 0, + "content": "We first outline the general framework of existing generalizable NeRF in Section 3.1. Then, we present our proposed CaesarNeRF, as illustrated in Figure 2. This model integrates elements of semantic representation, calibration, and sequential refinement, detailed in Section 3.2, 3.3, and 3.4, respectively. The training objective is given in 3.5." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.39, + 0.471, + 0.405 + ], + "angle": 0, + "content": "3.1 NeRF and generalizable NeRF" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.785, + 0.474 + ], + "angle": 0, + "content": "Neural Radiance Field (NeRF) [43, 44] aims to render 3D scenes by predicting both the density and RGB values at points where light rays intersect the radiance field. For a query point \\( \\pmb{x} \\in \\mathbb{R}^3 \\) and a viewing direction \\( \\pmb{d} \\) on the unit sphere \\( \\mathbb{S}^2 \\) in 3D space, the NeRF model \\( \\mathcal{F} \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.445, + 0.487, + 0.787, + 0.503 + ], + "angle": 0, + "content": "\\[\n\\sigma , \\boldsymbol {c} = \\mathcal {F} (\\boldsymbol {x}, \\boldsymbol {d}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.514, + 0.785, + 0.574 + ], + "angle": 0, + "content": "Here, \\(\\sigma \\in \\mathbb{R}\\) and \\(c\\in \\mathbb{R}^3\\) denote the density and the RGB values, respectively. After computing these values for a collection of discretized points along each ray, volume rendering techniques are employed to calculate the final RGB values for each pixel, thus reconstructing the image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.791, + 0.741 + ], + "angle": 0, + "content": "However, traditional NeRF models \\(\\mathcal{F}\\) are limited by their requirement for scene-specific training, making it unsuitable for generalizing to novel scenes. To overcome this, generalizable NeRF models, denoted by \\(\\mathcal{F}_G\\), are designed to render images of novel scenes without per-scene training. Given \\(N\\) reference images \\(\\{I_n\\}_{n=1}^N\\), an encoder-based generalizable NeRF model \\(\\mathcal{F}_G\\) decouples the object representation from the original NeRF by using an encoder to extract per-pixel feature maps \\(\\{F_n\\}_{n=1}^N\\) from the input images. To synthesize a pixel associated with a point \\(x\\) along a ray in direction \\(d\\), it projects \\(\\{F_n\\}_{n=1}^N\\) from nearby views and aggregates this multi-view pixel-level information using techniques such as average pooling [76] or cost volumes [6]. This results in a fused feature embedding \\(\\widetilde{\\mathcal{F}}\\), allowing \\(\\mathcal{F}_G\\) to predict density \\(\\sigma\\) and RGB values \\(c\\) for each point along the ray, as expressed by:" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.751, + 0.787, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\sigma , \\boldsymbol {c} = \\mathcal {F} _ {\\boldsymbol {G}} (\\boldsymbol {x}, \\boldsymbol {d}, \\tilde {\\boldsymbol {F}}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In our method, we adopt the recently introduced fully attention-based generalizable NeRF method, GNT [60], as both the backbone and the baseline. GNT shares a similar paradigm with (2) but employs transformers [61] to aggregate pixel-level features into \\(\\widetilde{F}\\). It uses a view transformer to fuse projected pixel-level features from reference" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.143, + 0.764, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.323, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Fig. 2: Overview of CaesarNeRF. CaesarNeRF employs a shared encoder to capture two types of features from input views, including scene-level semantic representation \\(\\{S_n\\}\\) and pixel-level feature representation \\(\\{F_n\\}\\). We use the same encoder for both the scene-level semantic representation and the pixel-level embeddings. Following calibration and aggregation of \\(\\{S_n\\}\\) from various views, we concatenate it with the pixel-level fused feature, processed by the view transformer. Subsequent use of the ray-transformer, coupled with sequential refinement, enables us to render the final RGB values for each pixel in the target view. The output features serve as the input for the next stage, indicated by matching line colors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.788, + 0.526 + ], + "angle": 0, + "content": "views, and a ray transformer to combine features from different points along a ray, eliminating the need for volume rendering. Further details about GNT can be found in [60]. We also demonstrate that our approach can be extended to other generalizable NeRF models, as discusses in Section 4.2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.551, + 0.505, + 0.567 + ], + "angle": 0, + "content": "3.2 Scene-level semantic representation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.788, + 0.669 + ], + "angle": 0, + "content": "Both encoder-based generalizable NeRF models [6, 59, 76] and their attention-based counterparts [58, 60] mainly rely on pixel-level feature representations. While effective, this approach restricts their capability for a holistic scene understanding, especially when reference views are scarce. This limitation also exacerbates challenges in resolving depth ambiguities between points along the rays, a problem that becomes more pronounced with fewer reference views." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.67, + 0.788, + 0.746 + ], + "angle": 0, + "content": "To address these challenges, we introduce semantic representations aimed at enriching the scene-level understanding. We utilize a shared CNN encoder and apply a Global Average Pooling (GAP) to its \\(C\\)-dimensional output feature map, generating \\(N\\) global feature vectors \\(\\{S_n\\}_{n=1}^N\\) corresponding to each input view. These feature vectors are then averaged to form a unified scene-level representation \\(\\mathbf{S}\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.423, + 0.758, + 0.787, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {S} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\boldsymbol {S} _ {n} \\in \\mathbb {R} ^ {C}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In GNT [60], which uses a view transformer to aggregate pixel-level features into an \\(L\\)-dimensional vector \\(\\bar{F}\\), we extend this by concatenating \\(\\bar{F}\\) with \\(S\\) to construct a" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.129 + ], + "angle": 0, + "content": "H. Zhu and T. Ding et al." + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.149, + 0.48, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.261, + 0.441, + 0.272 + ], + "angle": 0, + "content": "(a) \"L: green, R: red\"" + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.151, + 0.657, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.261, + 0.621, + 0.272 + ], + "angle": 0, + "content": "(b) \"L: red, R: green\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.34 + ], + "angle": 0, + "content": "Fig. 3: An illustration of conflicting semantic meanings from multiple viewpoints of the same object. When observing the cup from distinct angles, the features extracted after pooling retain spatial information but are inconsistent in the scene-level semantic understanding, leading to conflicts across various reference images after aggregation." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.37, + 0.518, + 0.385 + ], + "angle": 0, + "content": "global-local embedding \\( \\pmb{E} \\), as formulated by:" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.396, + 0.787, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {E} = \\operatorname {C o n c a t} (\\widetilde {\\boldsymbol {F}}, \\boldsymbol {S}) \\in \\mathbb {R} ^ {L + C}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.426, + 0.788, + 0.533 + ], + "angle": 0, + "content": "This combined embedding \\(\\pmb{E}\\) is then subjected to the standard self-attention mechanism used in GNT [60]. This approach enables the scene-level semantic representation \\((S)\\) to integrate with per-point features \\((\\widetilde{\\pmb{F}})\\), offering a more nuanced understanding at both levels. It also allows each point to selectively draw from the scene-level information. To maintain dimensional consistency across the input and output layers of multiple transformer modules, we employ a two-layer MLP to project the enhanced features back to the original dimension \\(L\\) of the per-point embedding \\(\\widetilde{\\pmb{F}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.557, + 0.527, + 0.572 + ], + "angle": 0, + "content": "3.3 Calibration of semantic representation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.674 + ], + "angle": 0, + "content": "The integration of the scene-level semantic representation \\( S \\), generated through simple averaging of global feature vectors as in (3), improves rendering quality. However, this approach has limitations when dealing with multiple views. As illustrated in Figure 3, viewing the same object from distinct angles may retain spatial attributes but can lead to conflicting semantic meanings. Merely averaging these global feature vectors without accounting for camera positions can result in a distorted scene-level understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.795 + ], + "angle": 0, + "content": "To mitigate this inconsistency, we propose a semantic calibration technique using feature rotation. This adjustment aligns the semantic representation across different camera poses. Our inspiration comes from the use of camera pose projection in computing the fused pixel-level feature \\(\\widetilde{F}\\) and is further motivated by [51], which demonstrates that explicit rotation operations in feature spaces are feasible. Unlike point clouds in [51] that inherently lack a defined canonical orientation, NeRF explicitly encodes differences between camera viewpoints, thereby enabling precise calibration between the reference and target images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.794, + 0.808, + 0.843 + ], + "angle": 0, + "content": "Building on this observation, we calculate calibrated semantic representations \\(\\{\\widetilde{S}_n\\}_{n=1}^N\\) from the \\(N\\) original semantic representations \\(\\{S_n\\}_{n=1}^N\\) derived from the reference views. We accomplish this by leveraging their respective rotation matrices \\(\\{T_n\\}_{n=1}^N\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.407, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.145, + 0.597, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.146, + 0.785, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.266, + 0.785, + 0.295 + ], + "angle": 0, + "content": "Fig. 4: Visualization of decoded feature maps for \"orchid\" in LLFF dataset, produced by ray transformers [60] at different stages. From left to right, the transformer stages increase in depth." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.322, + 0.784, + 0.352 + ], + "angle": 0, + "content": "to model the rotational variations between each input view and the target view. The alignment of the original semantic features is performed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.324, + 0.359, + 0.786, + 0.379 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\boldsymbol {S}} _ {n} = \\boldsymbol {\\mathcal {P}} \\left(\\boldsymbol {T} _ {n} \\cdot \\boldsymbol {\\mathcal {P}} ^ {- 1} \\left(\\boldsymbol {S} _ {n}\\right)\\right), \\text {w h e r e} \\boldsymbol {T} _ {n} = \\boldsymbol {T} _ {\\text {o u t}} ^ {\\mathrm {w} 2 \\mathrm {c}} \\cdot \\boldsymbol {T} _ {n} ^ {\\mathrm {c} 2 \\mathrm {w}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.784, + 0.447 + ], + "angle": 0, + "content": "Here, \\( T_{n}^{\\mathrm{c2w}} \\) is the inverse of the extrinsic matrix used for \\( I_{n} \\), and \\( T_{\\mathrm{out}}^{\\mathrm{w2c}} \\) is the extrinsic matrix for the target view. \\( \\mathcal{P}(\\cdot) \\) and \\( \\mathcal{P}^{-1}(\\cdot) \\) are the flattening and inverse flattening operations, which reshape the feature to a 1D vector of shape 1-by- \\( C \\) and a 2D matrix of shape 3-by- \\( \\frac{C}{3} \\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.784, + 0.522 + ], + "angle": 0, + "content": "Note that for the extrinsic matrix, we consider only the top-left \\(3 \\times 3\\) submatrix that accounts for rotation. Using GAP to condense feature maps of various sizes into a 1-by-\\(C\\) feature vector eliminates the need for scaling parameters in the semantic representation. As a result, modeling the intrinsic matrix is unnecessary, assuming no skewing, making our approach adaptable to different camera configurations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.784, + 0.555 + ], + "angle": 0, + "content": "With the calibrated semantic features \\(\\{\\widetilde{S}_n\\}_{n=1}^N\\) for each reference view, we average these, similar to (3), to obtain the calibrated scene-level semantic representation \\(\\widetilde{S}\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.423, + 0.566, + 0.786, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\boldsymbol {S}} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\widetilde {\\boldsymbol {S}} _ {n} \\in \\mathbb {R} ^ {C}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.784, + 0.644 + ], + "angle": 0, + "content": "Finally, akin to (4), we concatenate the pixel-level fused feature \\(\\widetilde{F}\\) with the calibrated scene-level semantic representation \\(\\widetilde{S}\\) to form the final global-local embedding \\(\\widetilde{E}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.65, + 0.786, + 0.669 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\boldsymbol {E}} = \\operatorname {C o n c a t} (\\widetilde {\\boldsymbol {F}}, \\widetilde {\\boldsymbol {S}}) \\in \\mathbb {R} ^ {L + C}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.677, + 0.784, + 0.768 + ], + "angle": 0, + "content": "This unified embedding then feeds into ray transformers, passing through standard self-attention mechanisms. In the original GNT [60], multiple view transformers and ray transformers are stacked alternately for sequential feature processing. The last ray transformer integrates features from multiple points along a ray to yield the final RGB value. We denote the corresponding feature representations at stage \\( k \\) as \\( \\widetilde{F}^{(k)} \\) and \\( \\widetilde{E}^{(k)} \\). Notably, the calibrated semantic representation \\( \\widetilde{S} \\) remains constant across these stages." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.788, + 0.41, + 0.803 + ], + "angle": 0, + "content": "3.4 Sequential refinement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.808, + 0.784, + 0.84 + ], + "angle": 0, + "content": "While leveraging \\(\\widetilde{S}\\) improves consistency, a single, uniform \\(\\widetilde{S}\\) may not be adequate for deeper layers that demand more nuanced details. In fact, we find that deeper transform-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.129 + ], + "angle": 0, + "content": "H. Zhu and T. Ding et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "ers capture finer details compared to shallower ones, as shown in Figure 4. To address this limitation, we introduce a sequential semantic feature refinement module that progressively enriches features at each stage. Specifically, we learn the residual \\(\\pmb{\\Delta}^{(k)}\\) to update \\(\\widetilde{\\pmb{S}}\\) at each stage \\(k\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.42, + 0.215, + 0.787, + 0.234 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\boldsymbol {S}} ^ {(k + 1)} \\leftarrow \\widetilde {\\boldsymbol {S}} ^ {(k)} + \\boldsymbol {\\Delta} ^ {(k)}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.246, + 0.784, + 0.337 + ], + "angle": 0, + "content": "Here, \\(\\pmb{\\Delta}^{(k)}\\) is calculated by first performing specialized cross-attentions between \\(\\widetilde{\\pmb{S}}^{(k)}\\) and the original, uncalibrated per-frame semantic features \\(\\{S_n\\}_{n=1}^N\\) (see Figure 2), followed by their summation. Our goal is to fuse information from different source views to enrich the scene-level semantic representation with features from each reference frame. With this sequential refinement, we combine \\(\\widetilde{\\pmb{S}}^{(k)}\\) with \\(\\widetilde{\\pmb{F}}^{(k)}\\) at each stage, yielding a stage-specific global-local embedding \\(\\widetilde{\\pmb{E}}^{(k)}\\), which completes our approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.338, + 0.788, + 0.504 + ], + "angle": 0, + "content": "Discussion. In scenarios with few reference views, especially when limited to just one, the primary issue is inaccurate depth estimation, resulting in depth ambiguity [12]. This compromises the quality of images when rendered from novel viewpoints. Despite this, essential visual information generally remains accurate across different camera poses. Incorporating our proposed scene-level representation improves the understanding of the overall scene layout [7], distinguishing our approach from existing generalizable NeRF models that predict pixels individually. The advantage of our approach is its holistic view; the semantic representation enriches per-pixel predictions by providing broader context. This semantic constraint ensures that fewer abrupt changes between adjacent points. Consequently, it leads to more reliable depth estimations, making the images rendered from limited reference views more plausible." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.525, + 0.393, + 0.54 + ], + "angle": 0, + "content": "3.5 Training objectives" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.549, + 0.597, + 0.563 + ], + "angle": 0, + "content": "During training, we employ three different loss functions:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.787, + 0.594 + ], + "angle": 0, + "content": "MSE loss. The Mean Square Error (MSE) loss is the standard photometric loss used in NeRF [43]. It computes the MSE between the actual and predicted pixel values." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.787, + 0.655 + ], + "angle": 0, + "content": "Central loss. Since we project the view-level features from different input camera poses to the shared target view, we introduce a central loss to ensure frame-wise calibrated semantic features \\(\\{\\widetilde{S}_n\\}_{n=1}^N\\) are consistent when projected onto the same target view, which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.663, + 0.787, + 0.705 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c e n t r a l}} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\left\\| \\widetilde {\\boldsymbol {S}} _ {n} - \\widetilde {\\boldsymbol {S}} \\right\\| _ {1}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.713, + 0.787, + 0.773 + ], + "angle": 0, + "content": "Point-wise perceptual loss. During the rendering of a batch of pixels in a target view, we inpaint the ground-truth image by replacing the corresponding pixels with the predicted ones. Then, a perceptual loss [26] is computed between the inpainted image and the target image to guide the training process at the whole-image level." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.774, + 0.559, + 0.787 + ], + "angle": 0, + "content": "The final loss function is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.799, + 0.787, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {M S E}} + \\lambda_ {1} \\mathcal {L} _ {\\text {c e n t r a l}} + \\lambda_ {2} \\mathcal {L} _ {\\text {p e r c}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.825, + 0.61, + 0.841 + ], + "angle": 0, + "content": "Empirically, we set \\(\\lambda_1 = 1\\) and \\(\\lambda_{2} = 0.001\\), following [32]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Results for generalizable scene rendering on LLFF with few reference views. GeoNeRF, MatchNeRF, and MVSNeRF necessitate variance as input, defaulting to 0 for single-image cases, hence their results are not included for 1-view scenarios." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.199, + 0.787, + 0.324 + ], + "angle": 0, + "content": "
Method1 reference view2 reference views3 reference views
PSNR (↑)LPIPS (↓)SSIM (↑)PSNR (↑)LPIPS (↓)SSIM (↑)PSNR (↑)LPIPS (↓)SSIM (↑)
PixelNeRF [76]9.320.8980.26411.230.7660.28211.240.6710.486
GPNR [58]15.910.5270.40018.790.3800.57521.570.2880.695
NeuRay [38]16.180.5840.39317.710.3360.64618.260.3100.672
GeoNeRF [25]---18.760.4730.50023.400.2460.766
MatchNeRF [8]---21.080.2720.68922.300.2340.731
MVSNeRF [6]---19.150.3360.70419.840.3140.729
IBRNet [64]16.850.5420.50721.250.3330.68523.000.2620.752
GNT [60]16.570.5000.42420.880.2510.69123.210.1780.782
Ours18.310.4350.52121.940.2240.73623.450.1760.794
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.351, + 0.357, + 0.37 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.388, + 0.787, + 0.434 + ], + "angle": 0, + "content": "In this section, we first discuss our experimental setups and other details required for our experiments in Section 4.1, followed by our results for different experimental settings, along with ablation studies and analysis in Section 4.2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.461, + 0.402, + 0.477 + ], + "angle": 0, + "content": "4.1 Experimental setups" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.491, + 0.788, + 0.642 + ], + "angle": 0, + "content": "Datasets. Firstly, following [60], we construct our training data from both synthetic and real data. This collection includes scanned models from Google Scanned Objects [14], RealEstate10K [81], and handheld phone captures [64]. For evaluation, we utilize real data encompassing complex scenes from sources such as LLFF [42], Shiny [67], and mip-NeRF 360 [4]. Additionally, we train and test our model using the recently released MVImgNet dataset [77]. We adhere to the official split, focusing on examples from the containers category, and select 2,500 scenes for training. During inference, we choose 100 scenes, using their first images as target views and the spatially nearest images as references. Since MVImgNet does not provide camera poses, we utilize COLMAP [54, 55] to deduce the camera positions within these scenes." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.643, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Implementation details. CaesarNeRF is built upon GNT [60], for which we maintain the same configuration, setting the ray and view transformers stack number \\((K)\\) to 8 for generalizable setting and 4 for single-scene setting. The feature encoder extracts bottleneck features, applies GAP, and then uses a fully connected (FC) layer to reduce the input dimension \\(C\\) to 96. Training involves 500,000 iterations using the Adam optimizer [28], with learning rate set at 0.001 for the feature encoder and 0.0005 for CaesarNeRF, halving them every 100,000 iterations. Each iteration samples 4,096 rays from a single scene. In line with [60], we randomly choose between 8 to 10 reference views for training, and 3 to 7 views when using the MVImgNet [77]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Baseline methods. We compare CaesarNeRF with several state-of-the-art methods suited for generalizable NeRF applications, including earlier works such as MVS-NeRF [6], PixelNeRF [76], and IBRNet [64], alongside more recent ones, including GPNR [58], NeuRay [38], GNT [60], GeoNeRF [25] and MatchNeRF [8]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.421, + 0.129 + ], + "angle": 0, + "content": "H. Zhu and T. Ding et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.245, + 0.145, + 0.756, + 0.16 + ], + "angle": 0, + "content": "Table 2: Results for generalizable scene rendering on Shiny with few reference views." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.171, + 0.785, + 0.257 + ], + "angle": 0, + "content": "
Method1 reference view2 reference views3 reference views
PSNR (↑)LPIPS (↓)SSIM (↑)PSNR (↑)LPIPS (↓)SSIM (↑)PSNR (↑)LPIPS (↓)SSIM (↑)
MatchNeRF [8]---20.280.2780.63620.770.2490.672
MVSNeRF [6]---17.250.4160.57718.550.3430.645
IBRNet [64]14.930.6250.40118.400.4000.59521.960.2810.710
GNT [60]15.990.5480.40020.420.3270.61722.470.2470.720
Ours17.570.4670.47221.470.2930.65222.740.2410.723
" + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.268, + 0.784, + 0.283 + ], + "angle": 0, + "content": "Table 3: Results for generalizable scene rendering on mip-NeRF 360 with few reference views." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.294, + 0.785, + 0.38 + ], + "angle": 0, + "content": "
Method1 reference view2 reference views3 reference views
PSNR (↑)LPIPS (↓)SSIM (↑)PSNR (↑)LPIPS (↓)SSIM (↑)PSNR (↑)LPIPS (↓)SSIM (↑)
MatchNeRF [8]---17.000.5660.39217.260.5510.407
MVSNeRF [6]---14.230.6810.36614.290.6740.406
IBRNet [64]14.120.6820.28316.240.6180.36017.700.5550.420
GNT [60]13.480.6300.31415.210.5590.37015.590.5380.395
Ours15.200.5920.35017.050.5380.40317.550.5120.430
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.406, + 0.4, + 0.421 + ], + "angle": 0, + "content": "4.2 Results and analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.787, + 0.505 + ], + "angle": 0, + "content": "We compare results in two settings: a generalizable setting, where the model is trained on multiple scenes without fine-tuning during inference for both few and all reference view cases, and a single-scene setting where the model is trained and evaluated on just one scene. Following these comparisons, we conduct ablation studies and test the generalizability of our method with other state-of-the-art approaches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.787, + 0.566 + ], + "angle": 0, + "content": "Generalizable rendering. In the generalizable setting, we adopt two training strategies. First, we train the model on multiple datasets as described in Section 4.1 and evaluate on LLFF [42], Shiny [67] and mip-NeRF 360 [4] datasets. In addition, the model is trained and tested on the MVImgNet [77] for object-centric generalizability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.787, + 0.686 + ], + "angle": 0, + "content": "(a) LLFF, Shiny, and mip-NeRF 360. The results for few-reference view scenarios on these datasets are shown in Tables 1, 2 and 3, respectively. Methods like MatchNeRF [8], MVSNeRF [75], and GeoNeRF [25] require at least two reference views. On the LLFF dataset, all methods experience a performance decline as the number of views decreases. CaesarNeRF, however, consistently outperforms others across varying reference view numbers, with the performance gap becoming more significant with fewer views. For example, with 3 views, while IBRNet [64] and GNT [60] have comparable PSNRs, CaesarNeRF demonstrates a more substantial lead in LPIPS and SSIM metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.686, + 0.787, + 0.717 + ], + "angle": 0, + "content": "Similar patterns are observed on the Shiny [67] and mip-NeRF 360 [4] datasets. We apply the highest-performing methods from the LLFF evaluations and report the results" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.742, + 0.787, + 0.771 + ], + "angle": 0, + "content": "Table 4: Results on MVImgNet across varying numbers of reference views. 'C.' represents the use of calibration before averaging." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.782, + 0.785, + 0.839 + ], + "angle": 0, + "content": "
Method1 reference view2 reference views3 reference views4 reference views5 reference views
PSNR (†)LPIPS (↓)SSIM (†)PSNR (†)LPIPS (↓)SSIM (†)PSNR (†)LPIPS (↓)SSIM (†)PSNR (†)LPIPS (↓)SSIM (†)PSNR (†)LPIPS (↓)SSIM (†)
IBRNet19.140.4580.59524.380.2660.81825.530.2030.85825.990.1900.86726.120.1880.867
GNT22.220.4330.67826.940.2360.85027.410.2060.87027.510.1970.87527.510.1940.876
Ours w/oC.23.610.3710.71826.340.2740.81727.100.2280.85027.300.2100.86227.340.2030.865
Ours24.280.3340.74727.340.2150.85627.820.1900.87527.920.1810.88127.920.1790.882
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.784, + 0.159 + ], + "angle": 0, + "content": "Table 5: Results of per-scene optimization on LLFF, in comparison with state-of-the-art methods." + }, + { + "type": "table", + "bbox": [ + 0.262, + 0.171, + 0.742, + 0.23 + ], + "angle": 0, + "content": "
MethodLLFF [42]NeRF [44]NeX [67]GNT [60]Ours
PSNR (↑)23.2726.5027.2627.2427.64
LPIPS (↓)0.2120.2500.1790.0870.081
SSIM (↑)0.7980.8110.9040.8890.904
" + }, + { + "type": "table_caption", + "bbox": [ + 0.223, + 0.244, + 0.779, + 0.259 + ], + "angle": 0, + "content": "Table 6: Results on LLFF for few-shot generalization after adapting Caesar to other baselines." + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.27, + 0.77, + 0.348 + ], + "angle": 0, + "content": "
Method1 reference view2 reference views3 reference views
PSNR (↑)SSIM (↑)LPIPS (↓)PSNR (↑)SSIM (↑)LPIPS (↓)PSNR (↑)SSIM (↑)LPIPS (↓)
MatchNeRF [8]---20.590.7750.27622.430.8050.244
+ Caesar---21.550.7820.26822.980.8240.242
IBRNet [64]16.850.5070.54221.250.6850.33323.000.7520.262
+ Caesar17.760.5430.50022.390.7400.27523.670.7720.242
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.787, + 0.451 + ], + "angle": 0, + "content": "for those that produce satisfactory outcomes with few reference views. CaesarNeRF maintains superior performance throughout. Notably, for complex datasets like mipNeRF 360 [4], which have sparse camera inputs, the quality of rendered images generally decreases with fewer available reference views. Nonetheless, CaesarNeRF shows the most robust performance compared to the other methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.452, + 0.788, + 0.572 + ], + "angle": 0, + "content": "(b) MVImgNet. We extend our comparison of CaesarNeRF with GNT [60] and IBRNet [64] on the MVImgNet dataset, focusing on object-centric scenes, as shown in Table 4. We examine a variant of CaesarNeRF where semantic calibration is substituted with simple feature averaging from multiple frames. While the performance of all methods improves with more views, CaesarNeRF consistently outperforms GNT and IBRNet. Notably, CaesarNeRF with feature averaging surpasses GNT in 1-view case but lags with additional views, implying that the absence of calibration leads to ambiguities when rendering from multiple views." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.573, + 0.788, + 0.618 + ], + "angle": 0, + "content": "Per-scene optimization. Beyond the multi-scene generalizable setting, we demonstrate per-scene optimization results in Table 5. We calculate the average performance over 8 categories from the LLFF dataset [42]. CaesarNeRF consistently outperforms" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.645, + 0.788, + 0.673 + ], + "angle": 0, + "content": "Table 7: Ablations on the semantic representation length \\( \\mathbf{R} \\), sequential refinement (Seq.) and calibration (Cali.). 'Ext.' denotes the extension of per-pixel feature to a length of 64 in GNT." + }, + { + "type": "table", + "bbox": [ + 0.305, + 0.685, + 0.699, + 0.838 + ], + "angle": 0, + "content": "
Model VariationsPSNR (↑)LPIPS (↓)SSIM (↑)
Rlen.Seq.Cali.
(Baseline GNT)20.930.1850.731
Ext.20.850.1730.735
+3221.430.1520.763
+6421.490.1490.766
+9621.460.1500.766
+12821.490.1470.763
+96✓21.530.1460.770
+96✓21.510.1470.769
+96✓✓21.670.1390.781
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.129 + ], + "angle": 0, + "content": "H. Zhu and T. Ding et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.149, + 0.787, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.445, + 0.621, + 0.457 + ], + "angle": 0, + "content": "(b) Using two images as reference view." + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.469, + 0.785, + 0.484 + ], + "angle": 0, + "content": "Fig. 5: Comparative visualization of our proposed method against other state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.785, + 0.563 + ], + "angle": 0, + "content": "nearly all state-of-the-art methods in the comparison, across all three metrics, showing a significant improvement over our baseline method, GNT [60]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.672 + ], + "angle": 0, + "content": "Adaptability. To test the adaptability of our Caesar pipeline, we apply it to two other state-of-the-art methods that use view transformers, namely MatchNeRF [8] and IBRNet [64]. We demonstrate in Table 6 that our enhancements in scene-level semantic understanding significantly boost the performance of these methods across all metrics. This indicates that the Caesar framework is not only beneficial in our CaesarNeRF, which is based on GNT [60], but can also be a versatile addition to other NeRF pipelines with view transformers to aggregate different input views." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Ablation analysis. We conduct ablation studies on the \"orchid\" scene from the LLFF dataset, with findings detailed in Table 7. Testing variations in representation and the impact of the sequential refinement and calibration modules, we find that increasing the latent size in GNT yields marginal benefits. However, incorporating even a modest semantic representation size distinctly improves results. The length of the semantic representation has a minimal impact on quality. Our ablation studies indicate that while sequential refinement and calibration each offer slight performance gains, their combined effect is most significant. In a single-scene context, semantic information is effectively embedded within the representation, making the benefits of individual modules subtler. Together, however, they provide a framework where sequential refinement can leverage calibrated features for deeper insights." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.243, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.277, + 0.785, + 0.304 + ], + "angle": 0, + "content": "Fig. 6: Depth estimation prediction using one reference view (first row) and two reference views (second row) as input from LLFF comparing CaesarNeRF with GNT." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.323, + 0.323, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.391, + 0.324, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.459, + 0.286, + 0.468 + ], + "angle": 0, + "content": "fern" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.323, + 0.429, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.339, + 0.391, + 0.425, + 0.4 + ], + "angle": 0, + "content": "chesstable (0.45)" + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.402, + 0.429, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.337, + 0.457, + 0.428, + 0.469 + ], + "angle": 0, + "content": "playground (0.46)" + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.323, + 0.555, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.448, + 0.391, + 0.548, + 0.4 + ], + "angle": 0, + "content": "colorfountain (0.50)" + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.402, + 0.545, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.457, + 0.529, + 0.469 + ], + "angle": 0, + "content": "bikes (0.50)" + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.323, + 0.661, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.391, + 0.662, + 0.4 + ], + "angle": 0, + "content": "fenceflower (1.91)" + }, + { + "type": "image", + "bbox": [ + 0.569, + 0.402, + 0.662, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.579, + 0.457, + 0.652, + 0.469 + ], + "angle": 0, + "content": "peppers (2.02)" + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.322, + 0.782, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.391, + 0.759, + 0.4 + ], + "angle": 0, + "content": "pond (1.80)" + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.401, + 0.776, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.457, + 0.762, + 0.469 + ], + "angle": 0, + "content": "apples (2.01)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.481, + 0.788, + 0.523 + ], + "angle": 0, + "content": "Fig.7: Largest and smallest distances for two examples from LLFF test split when matching with training scenes. Numbers \\((\\times 10^{-2})\\) denote the L-2 distance to the source image within the semantic feature space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.658 + ], + "angle": 0, + "content": "Visualizations. We present our visualization results in Figure 5, where we compare our method with others using one or two views from the LLFF dataset. Additional visual comparisons are provided in the supplementary materials. These visualizations highlight that in scenarios with few views, our method significantly surpasses other generalizable NeRF models, particularly excelling when only a single view is available. CaesarNeRF demonstrates rendering with sharper boundaries and more distinct objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.79, + 0.81 + ], + "angle": 0, + "content": "Depth estimation. We extend our evaluation to depth prediction within the LLFF [42] dataset, focusing on challenges presented by few reference views, such as scenarios with just one or two images. In a comparison between CaesarNeRF and GNT [60], we observe in Figure 6 that GNT struggles to accurately capture the relative positions of objects when reference images are sparse. For instance, with only a single view of a flower, CaesarNeRF precisely indicates the flower's proximity to the camera compared to the leaves in the background, a distinction that GNT fails to make. Furthermore, the depth estimations provided by CaesarNeRF are consistently more reliable. In the horn example involving two views, CaesarNeRF offers better boundary delineation, showing particular strength in handling reflective surfaces such as grass in the background." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Semantic analysis for \\(\\tilde{S}\\). To assess whether the calibrated semantic representation \\(\\tilde{S}\\) truly captures semantic details of the scene, we analyze the highest and lowest" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.129 + ], + "angle": 0, + "content": "H. Zhu and T. Ding et al." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.144, + 0.361, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.144, + 0.501, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.144, + 0.639, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.144, + 0.78, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.228, + 0.393, + 0.241 + ], + "angle": 0, + "content": "(a) leaves" + }, + { + "type": "image_caption", + "bbox": [ + 0.61, + 0.228, + 0.673, + 0.241 + ], + "angle": 0, + "content": "(b) orchid" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.254, + 0.788, + 0.296 + ], + "angle": 0, + "content": "Fig.8: Synthetic results for two examples from LLFF [42], “leaves” and “orchid”, using zero 1-to-3 [35] with one reference image as input and 2 degrees of vertical shift. The left image of each pair is the input, and the right one is the output of zero 1-to-3 [35]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.323, + 0.784, + 0.383 + ], + "angle": 0, + "content": "response based on the L-2 distance between features from two different scenes, interpreting smaller distances as greater similarity. We consider the first image from each category in the LLFF dataset as a reference and extract the scene-level representation for the ten closest views to this reference image using CaesarNeRF." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.785, + 0.504 + ], + "angle": 0, + "content": "In Figure 7, we present two examples, \"room\" and \"fortress\", from the LLFF dataset, representing the first image of these categories alongside images corresponding to the top-2 highest and lowest responses. This analysis reveals that the scene-level representation predominantly emphasizes structural information and objects of similar categories. For instance, with the source image \"room\", the highest responses correlate with images featuring table-like structures, indicating an object-centric focus. In contrast, the lowest responses include images of flowers or ponds found in open areas, which diverge significantly from the structural and object content in the source images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.787, + 0.625 + ], + "angle": 0, + "content": "Comparison with generative methods. Single-view scenarios are often addressed by generative methods [35-37] that employ diffusion models [19, 53]. While these models can produce reasonable results for object-centric renderings, they struggle with scene-level renderings from novel viewpoints. We show two LLFF examples in Figure 8 using zero 1-to-3 [35], where the left image is the input, and the right one is the output. Images rendered with zero 1-to-3 [35] suffer from style difference. Unlike NeRF-based approaches that reconstruct images from observed pixels, generative models synthesize an entire image from its semantic representation but may not maintain the style." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.646, + 0.467, + 0.661 + ], + "angle": 0, + "content": "5 Conclusion and limitation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.785, + 0.779 + ], + "angle": 0, + "content": "In this paper, we introduce CaesarNeRF, a few-shot and generalizable NeRF pipeline that combines scene-level semantic with per-pixel feature representations, aiding in rendering from novel camera positions with limited reference views. We calibrate the semantic representations across different input views and employ a sequential refinement network to offer distinct semantic representations at various levels. Our method has been extensively evaluated on a broad range of datasets, exhibiting state-of-the-art performance in both generalizable and single-scene settings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Limitations and potential negative impact. Instead of a generative approach, CaesarNeRF relies on NeRF's scheme, using input images to render the target view. This approach restricts its ability to render parts of the scene that are not present in the reference images. CaesarNeRF may create fake images for authentication." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.312, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.786, + 0.217 + ], + "angle": 0, + "content": "1. Bao, C., Zhang, Y., Yang, B., Fan, T., Yang, Z., Bao, H., Zhang, G., Cui, Z.: Sine: Semantic-driven image-based nerf editing with prior-guided editing field. In: CVPR. pp. 20919-20929 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.786, + 0.245 + ], + "angle": 0, + "content": "2. Bao, Y., Ding, T., Huo, J., Li, W., Li, Y., Gao, Y.: Insertnerf: Instilling generalizability into nef with hypernet modules. ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.247, + 0.786, + 0.273 + ], + "angle": 0, + "content": "3. Bao, Y., Li, Y., Huo, J., Ding, T., Liang, X., Li, W., Gao, Y.: Where and how: Mitigating confusion in neural radiance fields from sparse inputs. arXiv preprint arXiv:2308.02908 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.786, + 0.3 + ], + "angle": 0, + "content": "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In: CVPR. pp. 5470–5479 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.302, + 0.786, + 0.328 + ], + "angle": 0, + "content": "5. Buehler, C., Bosse, M., McMillan, L., Gortler, S., Cohen, M.: Unstructured lumigraph rendering. In: SIGGRAPH. pp. 425-432 (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.786, + 0.356 + ], + "angle": 0, + "content": "6. Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In: ICCV. pp. 14124-14133 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.358, + 0.786, + 0.397 + ], + "angle": 0, + "content": "7. Chen, W., Xu, X., Jia, J., Luo, H., Wang, Y., Wang, F., Jin, R., Sun, X.: Beyond appearance: a semantic controllable self-supervised learning framework for human-centric visual tasks. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.786, + 0.425 + ], + "angle": 0, + "content": "8. Chen, Y., Xu, H., Wu, Q., Zheng, C., Cham, T.J., Cai, J.: Explicit correspondence matching for generalizable neural radiance fields. arXiv preprint arXiv:2304.12294 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.427, + 0.786, + 0.452 + ], + "angle": 0, + "content": "9. Chibane, J., Bansal, A., Lazova, V., Pons-Moll, G.: Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In: CVPR. pp. 7911-7920 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.454, + 0.786, + 0.493 + ], + "angle": 0, + "content": "0. Cong, W., Liang, H., Wang, P., Fan, Z., Chen, T., Varma, M., Wang, Y., Wang, Z.: Enhancing nerf akin to enhancing llms: Generalizable nerf transformer with mixture-of-view-experts. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.496, + 0.786, + 0.522 + ], + "angle": 0, + "content": "1. Debevec, P.E., Taylor, C.J., Malik, J.: Modeling and rendering architecture from photographs: A hybrid geometry-and image-based approach. In: SIGGRAPH. pp. 11-20 (1996)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.524, + 0.786, + 0.562 + ], + "angle": 0, + "content": "2. Deng, C., Jiang, C., Qi, C.R., Yan, X., Zhou, Y., Guibas, L., Anguelov, D.: Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In: CVPR. pp. 20637-20647 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.786, + 0.591 + ], + "angle": 0, + "content": "3. Deng, K., Liu, A., Zhu, J.Y., Ramanan, D.: Depth-supervised nerf: Fewer views and faster training for free. In: CVPR. pp. 12882-12891 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.593, + 0.786, + 0.632 + ], + "angle": 0, + "content": "4. Downs, L., Francis, A., Koenig, N., Kinman, B., Hickman, R., Reymann, K., McHugh, T.B., Vanhoucke, V.: Google scanned objects: A high-quality dataset of 3d scanned household items. In: ICRA. pp. 2553-2560 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.786, + 0.66 + ], + "angle": 0, + "content": "5. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: CVPR. pp. 12479-12488 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.662, + 0.786, + 0.687 + ], + "angle": 0, + "content": "6. Fu, Y., Misra, I., Wang, X.: Multiplane nef-supervised disentanglement of depth and camera pose from videos. In: ICML (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.786, + 0.715 + ], + "angle": 0, + "content": "7. Gao, Y., Cao, Y.P., Shan, Y.: Surfelnerf: Neural surfel radiance fields for online photorealistic reconstruction of indoor scenes. In: CVPR. pp. 108-118 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.786, + 0.743 + ], + "angle": 0, + "content": "8. Gu, J., Liu, L., Wang, P., Theobalt, C.: Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.786, + 0.769 + ], + "angle": 0, + "content": "9. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS pp. 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.786, + 0.811 + ], + "angle": 0, + "content": "20. Irshad, M.Z., Zakharov, S., Liu, K., Guizilini, V., Kollar, T., Gaidon, A., Kira, Z., Ambrus, R.: Neo 360: Neural fields for sparse view synthesis of outdoor scenes. In: ICCV. pp. 9187-9198 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "21. Jain, A., Mildenhall, B., Barron, J.T., Abbeel, P., Poole, B.: Zero-shot text-guided object generation with dream fields. In: CVPR. pp. 867-876 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.129 + ], + "angle": 0, + "content": "H. Zhu and T. Ding et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "22. Jain, A., Tancik, M., Abbeel, P.: Putting nerf on a diet: Semantically consistent few-shot view synthesis. In: ICCV. pp. 5885-5894 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "23. Jang, W., Agapito, L.: Codenerf: Disentangled neural radiance fields for object categories. In: ICCV. pp. 12949-12958 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "24. Jiang, Y., Hedman, P., Mildenhall, B., Xu, D., Barron, J.T., Wang, Z., Xue, T.: Alignerf: High-fidelity neural radiance fields via alignment-aware training. In: CVPR. pp. 46-55 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.258 + ], + "angle": 0, + "content": "25. Johari, M.M., Lepoittevin, Y., Fleuret, F.: Geonerf: Generalizing nerf with geometry priors. In: CVPR. pp. 18365-18375 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "26. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and superresolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.301, + 0.785, + 0.327 + ], + "angle": 0, + "content": "27. Kania, A., Kasymov, A., Zieba, M., Spurek, P.: Hypernerfgan: Hypernetwork approach to 3d nerf gan. arXiv preprint arXiv:2301.11631 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.329, + 0.785, + 0.355 + ], + "angle": 0, + "content": "28. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.357, + 0.785, + 0.383 + ], + "angle": 0, + "content": "29. Kwak, M., Song, J., Kim, S.: Geconerf: Few-shot neural radiance fields via geometric consistency. arXiv preprint arXiv:2301.10941 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "30. Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: Dynibar: Neural dynamic image-based rendering. In: CVPR. pp. 4273-4284 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.412, + 0.785, + 0.451 + ], + "angle": 0, + "content": "31. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: CVPR. pp. 300–309 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.453, + 0.785, + 0.493 + ], + "angle": 0, + "content": "32. Lin, H., Peng, S., Xu, Z., Yan, Y., Shuai, Q., Bao, H., Zhou, X.: Efficient neural radiance fields for interactive free-viewpoint video. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.495, + 0.785, + 0.535 + ], + "angle": 0, + "content": "33. Lin, Y., Bai, H., Li, S., Lu, H., Lin, X., Xiong, H., Wang, L.: Componerf: Text-guided multi-object compositional nef with editable 3d scene layout. arXiv preprint arXiv:2303.13843 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.536, + 0.785, + 0.563 + ], + "angle": 0, + "content": "34. Liu, L., Habermann, M., Rudnev, V., Sarkar, K., Gu, J., Theobalt, C.: Neural actor: Neural free-view synthesis of human actors with pose control. TOC 40(6), 1-16 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.565, + 0.785, + 0.591 + ], + "angle": 0, + "content": "35. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.592, + 0.785, + 0.631 + ], + "angle": 0, + "content": "36. Liu, X., Kao, S.h., Chen, J., Tai, Y.W., Tang, C.K.: Deceptive-nerf: Enhancing nerf reconstruction using pseudo-observations from diffusion models. arXiv preprint arXiv:2305.15171 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.633, + 0.785, + 0.673 + ], + "angle": 0, + "content": "37. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "38. Liu, Y., Peng, S., Liu, L., Wang, Q., Wang, P., Theobalt, C., Zhou, X., Wang, W.: Neural rays for occlusion-aware image-based rendering. In: CVPR. pp. 7824-7833 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.703, + 0.785, + 0.729 + ], + "angle": 0, + "content": "39. Mariotti, O., Mac Aodha, O., Bilen, H.: Viewnerf: Unsupervised viewpoint estimation using category-level neural radiance fields. arXiv preprint arXiv:2212.00436 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.73, + 0.785, + 0.77 + ], + "angle": 0, + "content": "40. Martin-Brualla, R., Radwan, N., Sajjadi, M.S., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: CVPR. pp. 7210-7219 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.772, + 0.785, + 0.798 + ], + "angle": 0, + "content": "41. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: CVPR. pp. 12663–12673 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "42. Mildenhall, B., Srinivasan, P.P., Ortiz-Cayon, R., Kalantari, N.K., Ramamoorthi, R., Ng, R., Kar, A.: Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. TOG 38(4), 1-14 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "43. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.787, + 0.215 + ], + "angle": 0, + "content": "44. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.785, + 0.257 + ], + "angle": 0, + "content": "45. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: CVPR. pp. 5480-5490 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.258, + 0.785, + 0.284 + ], + "angle": 0, + "content": "46. Noguchi, A., Sun, X., Lin, S., Harada, T.: Neural articulated radiance field. In: ICCV. pp. 5762-5772 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.311 + ], + "angle": 0, + "content": "47. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: ICCV. pp. 5865-5874 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.312, + 0.785, + 0.339 + ], + "angle": 0, + "content": "48. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable neural radiance fields for modeling dynamic human bodies. In: ICCV. pp. 14314-14323 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.34, + 0.785, + 0.366 + ], + "angle": 0, + "content": "49. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.367, + 0.785, + 0.393 + ], + "angle": 0, + "content": "50. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: CVPR. pp. 10318-10327 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.394, + 0.785, + 0.42 + ], + "angle": 0, + "content": "51. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR. pp. 652-660 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.421, + 0.785, + 0.447 + ], + "angle": 0, + "content": "52. Roessle, B., Barron, J.T., Mildenhall, B., Srinivasan, P.P., Nießner, M.: Dense depth priors for neural radiance fields from sparse input views. In: CVPR. pp. 12892-12901 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.448, + 0.785, + 0.474 + ], + "angle": 0, + "content": "53. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.475, + 0.733, + 0.487 + ], + "angle": 0, + "content": "54. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.514 + ], + "angle": 0, + "content": "55. Schonberger, J.L., Zheng, E., Pollefeys, M., Frahm, J.M.: Pixelwise view selection for unstructured multi-view stereo. In: ECCV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.785, + 0.541 + ], + "angle": 0, + "content": "56. Schwarz, K., Liao, Y., Niemeyer, M., Geiger, A.: Graf: Generative radiance fields for 3d-aware image synthesis. NeurIPS 33, 20154-20166 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.568 + ], + "angle": 0, + "content": "57. Shue, J.R., Chan, E.R., Po, R., Ankner, Z., Wu, J., Wetzstein, G.: 3d neural field generation using triplane diffusion. In: CVPR. pp. 20875-20886 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.595 + ], + "angle": 0, + "content": "58. Suhail, M., Esteves, C., Sigal, L., Makadia, A.: Generalizable patch-based neural rendering. In: ECCV. pp. 156-174 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.597, + 0.785, + 0.623 + ], + "angle": 0, + "content": "59. Trevithick, A., Yang, B.: Grf: Learning a general radiance field for 3d representation and rendering. In: ICCV. pp. 15182-15192 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.624, + 0.785, + 0.649 + ], + "angle": 0, + "content": "60. Varma, M., Wang, P., Chen, X., Chen, T., Venugopalan, S., Wang, Z.: Is attention all that nerf needs? In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.785, + 0.677 + ], + "angle": 0, + "content": "61. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.703 + ], + "angle": 0, + "content": "62. Waechter, M., Moehrle, N., Goesele, M.: Let there be color! large-scale texturing of 3d reconstructions. In: ECCV. pp. 836-850 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.785, + 0.731 + ], + "angle": 0, + "content": "63. Wang, G., Chen, Z., Loy, C.C., Liu, Z.: Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.785, + 0.771 + ], + "angle": 0, + "content": "64. Wang, Q., Wang, Z., Genova, K., Srinivasan, P.P., Zhou, H., Barron, J.T., Martin-Brualla, R., Snavely, N., Funkhouser, T.: Ibrnet: Learning multi-view image-based rendering. In: CVPR. pp. 4690-4699 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.813 + ], + "angle": 0, + "content": "65. Wang, T., Zhang, B., Zhang, T., Gu, S., Bao, J., Baltrusaitis, T., Shen, J., Chen, D., Wen, F., Chen, Q., et al.: Rodin: A generative model for sculpting 3d digital avatars using diffusion. In: CVPR. pp. 4563-4573 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.787, + 0.84 + ], + "angle": 0, + "content": "66. Wei, Y., Liu, S., Rao, Y., Zhao, W., Lu, J., Zhou, J.: Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In: CVPR. pp. 5610-5619 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.422, + 0.129 + ], + "angle": 0, + "content": "H. Zhu and T. Ding et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "67. Wizadwongsa, S., Phongthawee, P., Yenphraphai, J., Suwajanakorn, S.: Nex: Real-time view synthesis with neural basis expansion. In: CVPR. pp. 8534-8543 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "68. Xiangli, Y., Xu, L., Pan, X., Zhao, N., Rao, A., Theobalt, C., Dai, B., Lin, D.: Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "69. Xie, C., Park, K., Martin-Brualla, R., Brown, M.: Fig-nerf: Figure-ground neural radiance fields for 3d object category modelling. In: 3DV. pp. 962–971 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "70. Xu, D., Jiang, Y., Wang, P., Fan, Z., Shi, H., Wang, Z.: Sinnerf: Training neural radiance fields on complex scenes from a single image. In: ECCV. pp. 736-753 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "71. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: CVPR. pp. 5438-5448 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "72. Yang, B., Zhang, Y., Xu, Y., Li, Y., Zhou, H., Bao, H., Zhang, G., Cui, Z.: Learning object-compositional neural radiance field for editable scene rendering. In: ICCV. pp. 13779-13788 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "73. Yang, H., Hong, L., Li, A., Hu, T., Li, Z., Lee, G.H., Wang, L.: Contranerf: Generalizable neural radiance fields for synthetic-to-real novel view synthesis via contrastive learning. In: CVPR. pp. 16508-16517 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "74. Yang, J., Pavone, M., Wang, Y.: Freenerf: Improving few-shot neural rendering with free frequency regularization. In: CVPR. pp. 8254-8263 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "75. Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstructured multiview stereo. In: ECCV. pp. 767-783 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "76. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: CVPR. pp. 4578-4587 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.493 + ], + "angle": 0, + "content": "77. Yu, X., Xu, M., Zhang, Y., Liu, H., Ye, C., Wu, Y., Yan, Z., Zhu, C., Xiong, Z., Liang, T., et al.: Mvimgnet: A large-scale dataset of multi-view images. In: CVPR. pp. 9150-9161 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.521 + ], + "angle": 0, + "content": "78. Zhang, J., Yang, G., Tulsiani, S., Ramanan, D.: Ners: Neural reflectance surfaces for sparse-view 3d reconstruction in the wild. NeurIPS 34, 29835-29847 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.549 + ], + "angle": 0, + "content": "79. Zhang, J., Li, X., Wan, Z., Wang, C., Liao, J.: Text2nerf: Text-driven 3d scene generation with neural radiance fields. arXiv preprint arXiv:2305.11588 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "80. Zhenxing, M., Xu, D.: Switch-nerf: Learning scene decomposition with mixture of experts for large-scale neural radiance fields. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.605 + ], + "angle": 0, + "content": "81. Zhou, T., Tucker, R., Flynn, J., Fyffe, G., Snavely, N.: Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.632 + ], + "angle": 0, + "content": "82. Zhou, Z., Tulsiani, S.: Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In: CVPR. pp. 12588-12597 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.66 + ], + "angle": 0, + "content": "83. Zhu, H., Sun, Y., Liu, C., Xia, L., Luo, J., Qiao, N., Nevatia, R., Kuo, C.H.: Multimodal neural radiance field. In: ICRA. pp. 9393-9399 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.688 + ], + "angle": 0, + "content": "84. Zhu, H., Zheng, Z., Zheng, W., Nevatia, R.: Cat-nerf: Constancy-aware tx2former for dynamic body modeling. In: CVPRW. pp. 6618-6627 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "85. Zhuang, Y., Zhu, H., Sun, X., Cao, X.: Mofanerf: Morphable facial neural radiance field. In: ECCV. pp. 268-285 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.785, + 0.743 + ], + "angle": 0, + "content": "86. Zimny, D., Trzeciński, T., Spurek, P.: Points2nerf: Generating neural radiance fields from 3d point cloud. arXiv preprint arXiv:2206.01290 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.743 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CaesarNeRF_ Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering/df1f2576-4362-497d-ae9b-cdaf433e83ec_origin.pdf b/2024/CaesarNeRF_ Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering/df1f2576-4362-497d-ae9b-cdaf433e83ec_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8a577ed4ba6a32a6925f19a707d5660b4a6cc149 --- /dev/null +++ b/2024/CaesarNeRF_ Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering/df1f2576-4362-497d-ae9b-cdaf433e83ec_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:86ae0938738ea887731e1003b0c4709c4029d162889079a9393c8dcda33949cd +size 10084923 diff --git a/2024/CaesarNeRF_ Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering/full.md b/2024/CaesarNeRF_ Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e22d0a0319bd4cce6aca7c6657309667db7327a9 --- /dev/null +++ b/2024/CaesarNeRF_ Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering/full.md @@ -0,0 +1,403 @@ +# CaesarNeRF: Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering + +Haidong Zhu $^{1,*}$ + +Tianyu Ding $^{2,\ast,\dagger}$ Ram Nevatia + +Tianyi Chen $^{2}$ , Ning Liang $^{2,\dagger}$ + +Ilya Zharkov2 + +1University of Southern California +2Microsoft +{haidongz, nevatia}@usc.edu + +{tianyuding,tiachen,zharkov,lulian}@microsoft.com + +![](images/8e8110b63c99922a92cd84756438c1580ba201780eb229980ac533d92538a796.jpg) +Fig.1: Novel view synthesis for novel scenes using ONE reference view on Shiny [67], LLFF [42], and MVImgNet [77] (top to bottom). Each triplet of images corresponds to the results from GNT [60] (left), CaesarNeRF (middle) and groundtruth (right). + +Abstract. Generalizability and few-shot learning are key challenges in Neural Radiance Fields (NeRF), often due to the lack of a holistic understanding in pixel-level rendering. We introduce CaesarNeRF, an end-to-end approach that leverages scene-level CAlibratEd SemAntic Representation along with pixel-level representations to advance few-shot, generalizable neural rendering, facilitating a holistic understanding without compromising high-quality details. CaesarNeRF explicitly models pose differences of reference views to combine scene-level semantic representations, providing a calibrated holistic understanding. This calibration process aligns various viewpoints with precise location and is further enhanced by sequential refinement to capture varying details. Extensive experiments on public datasets, including LLFF, Shiny, mip-NeRF 360, and MVImgNet, show that CaesarNeRF delivers state-of-the-art performance across varying numbers of reference views, proving effective even with a single reference image. + +Keywords: Few-view rendering $\cdot$ Generalizable NeRF $\cdot$ Neural rendering + +# 1 Introduction + +Rendering a scene from a novel camera position is essential in view synthesis [5,11,62]. The recent advancement of Neural Radiance Field (NeRF) [44] has shown impressive results in creating photo-realistic images from novel viewpoints. However, conventional NeRF methods are either typically scene-specific, necessitating retraining for novel scenes [15, 16, 44, 67, 74], or require a large number of reference views as input for generalizing to novel scenarios [6, 58, 60, 76]. These constraints highlight the complexity of the few-shot generalizable neural rendering, which aims to render unseen scenes from novel viewpoints with a limited number of reference images. + +Generalizing NeRF to novel scenes often involves using pixel-level feature embeddings encoded from input images, as seen in existing methods [59, 76]. These methods adapt NeRF to novel scenes by separating the scene representation from the model through an image encoder. However, relying solely on pixel-level features has its drawbacks: it requires highly precise epipolar geometry and often overlooks occlusion in complex scenes. Moreover, employing pixel-level features ignores the inherent interconnections within objects in the scene, treating the prediction of each pixel independently. Prior attempts to utilize scene-level representation either suffers from style mismatches [35] during scene-wide rendering or have been limited to specific object categories [23, 39, 69]. With few input reference under generalizable settings, these limitations become more pronounced, exacerbating the ambiguity in predictions due to biases from camera viewpoints. Although recent diffusion-based models [18, 56] attempt to address these issues through generative approaches, they struggle to effectively use the input views as contextual references for specific scenes. + +We present CaesarNeRF, a method that advances the generalizability of NeRF by incorporating calibrated semantic representation. This enables rendering from novel viewpoints using as few as one input reference view, as depicted in Figure 1. Our approach combines semantic scene-level representation with per-pixel features, enhancing consistency across different views of the same scene. The encoder-generated scene-level representations capture both semantic features and biases linked to specific camera poses. When reference views are limited, these biases can introduce uncertainty in the rendered images. To counter this, CaesarNeRF integrates camera pose transformations into the semantic representation, hence the term calibrated. By isolating pose-specific information from the scene-level representation, our model harmonizes features across input views, mitigating view-specific biases and, in turn, reducing ambiguity. In addition, CaesarNeRF introduces a sequential refinement process, which equips the model with varying levels of detail needed to enhance the semantic features. Extensive experiments on datasets such as LLFF [42], Shiny [67], mip-NeRF 360 [4], and the newly released MVImgNet [77] demonstrate that CaesarNeRF outperforms current state-of-the-art methods with limited reference views available, proving effective in generalizable settings with as few as one reference view. The project can be found here. + +In summary, our contributions are as follows: + +- We introduce CaesarNeRF, which utilizes scene-level calibrated semantic representation to achieve few-shot, generalizable neural rendering. This innovation leads to coherent and high-quality renderings. + +- We integrate semantic scene context with pixel-level details, in contrast to existing methods that rely solely on pixel-level features. We also address view-specific biases by modeling camera pose transformations and enhance the scene understanding through the sequential refinement of semantic features. + +- We demonstrate through extensive experiments that CaesarNeRF consistently outperforms state-of-the-art generalizable NeRF methods across a variety of datasets. Furthermore, integrating the Caesar pipeline into other baseline methods leads to consistent performance gains, highlighting its effectiveness and adaptability. + +# 2 Related work + +Neural Radiance Field (NeRF) implicitly captures the density and appearance of points within a scene or object [43, 44] and enables rendering from novel camera positions. In recent years, NeRF has witnessed improvements in a wide range of applications, such as photo-realistic novel view synthesis for large-scale scenes [40,68,80], dynamic scene decomposition and deformation [24,30,34,46-48,50,84,85], occupancy or depth estimation [66,71,83], scene generation and editing [1,21,31,33,41,49,72,73,79], and so on. Despite these advances, most methods still rely on the original NeRF and require retraining or fine-tuning for novel scenes not covered in the training data. + +Generalizable NeRF aims to adapt a single NeRF model to multiple scenes by separating the scene representation from the model. This field has seen notable advancements, with efforts focused on avoiding the need for retraining [8,20,38,64,65,75,76]. PixelNeRF [76] and GRF [59] pioneered the application of an image encoder to transform images into per-pixel features, with NeRF functioning as a decoder for predicting density and color from these features. MVSNeRF [6] introduces the use of a cost volume from MVSNet [75] to encode 3-D features from multiple views. Recognizing the intrinsic connection between points along a ray, IBRNet [64] employs self-attention to enhance point density predictions. Transformer-based [61] networks like GNT [10,60], GeoNeRF [25], and GPNR [58] are explored as alternatives to volume rendering, concentrating on pixel and patch-level representations. Additionally, InsertNeRF [2] utilizes hypernet modules to adapt parameters for novel scenes efficiently. + +These methods primarily depend on image encoders to extract pixel-aligned features from reference views. As a result, many of them lack a comprehensive understanding of the entire scene. Furthermore, with few reference views, the features become intertwined with view-specific details, compromising the quality of the rendering results. + +Few-shot Neural Radiance Field aims to render novel views using a limited number of reference images. To this end, various methods have been developed, incorporating additional information such as normalization-flow [78], semantic constraints [17, 22], depth cues [13, 52], geometry consistency [3, 29, 45, 63, 70], and frequency content [74]. Others [6, 9] emphasize pretraining on large-scale datasets. + +While these methods offer reasonable reconstructions with few inputs, they typically still require training or fine-tuning for specific scenes. Moreover, these methods usually require at least three reference images. With fewer than three, view-specific biases lead to ambiguity, complicating the rendering. Diffusion-based [12, 35, 57, 82] and other generative methods [27, 86] have been explored for single-view synthesis or gen- + +erative rendering, yet they are mostly limited to single-object rendering and generally fall short for complex scenes, which often result in a style change. + +CaesarNeRF confronts the above challenges by leveraging calibrated semantic representations that exploit scene geometry and variations in camera viewpoints. As a result, CaesarNeRF overcomes the limitations of pixel-level features and reduces dependency on external data or extensive pretraining, delivering high-quality renderings in few-shot and generalizable settings. + +# 3 The proposed method + +We first outline the general framework of existing generalizable NeRF in Section 3.1. Then, we present our proposed CaesarNeRF, as illustrated in Figure 2. This model integrates elements of semantic representation, calibration, and sequential refinement, detailed in Section 3.2, 3.3, and 3.4, respectively. The training objective is given in 3.5. + +# 3.1 NeRF and generalizable NeRF + +Neural Radiance Field (NeRF) [43, 44] aims to render 3D scenes by predicting both the density and RGB values at points where light rays intersect the radiance field. For a query point $\pmb{x} \in \mathbb{R}^3$ and a viewing direction $\pmb{d}$ on the unit sphere $\mathbb{S}^2$ in 3D space, the NeRF model $\mathcal{F}$ is defined as: + +$$ +\sigma , \boldsymbol {c} = \mathcal {F} (\boldsymbol {x}, \boldsymbol {d}). \tag {1} +$$ + +Here, $\sigma \in \mathbb{R}$ and $c\in \mathbb{R}^3$ denote the density and the RGB values, respectively. After computing these values for a collection of discretized points along each ray, volume rendering techniques are employed to calculate the final RGB values for each pixel, thus reconstructing the image. + +However, traditional NeRF models $\mathcal{F}$ are limited by their requirement for scene-specific training, making it unsuitable for generalizing to novel scenes. To overcome this, generalizable NeRF models, denoted by $\mathcal{F}_G$ , are designed to render images of novel scenes without per-scene training. Given $N$ reference images $\{I_n\}_{n=1}^N$ , an encoder-based generalizable NeRF model $\mathcal{F}_G$ decouples the object representation from the original NeRF by using an encoder to extract per-pixel feature maps $\{F_n\}_{n=1}^N$ from the input images. To synthesize a pixel associated with a point $x$ along a ray in direction $d$ , it projects $\{F_n\}_{n=1}^N$ from nearby views and aggregates this multi-view pixel-level information using techniques such as average pooling [76] or cost volumes [6]. This results in a fused feature embedding $\widetilde{\mathcal{F}}$ , allowing $\mathcal{F}_G$ to predict density $\sigma$ and RGB values $c$ for each point along the ray, as expressed by: + +$$ +\sigma , \boldsymbol {c} = \mathcal {F} _ {\boldsymbol {G}} (\boldsymbol {x}, \boldsymbol {d}, \tilde {\boldsymbol {F}}). \tag {2} +$$ + +In our method, we adopt the recently introduced fully attention-based generalizable NeRF method, GNT [60], as both the backbone and the baseline. GNT shares a similar paradigm with (2) but employs transformers [61] to aggregate pixel-level features into $\widetilde{F}$ . It uses a view transformer to fuse projected pixel-level features from reference + +![](images/52879611db002d2c8b11c42bf8c0ec659f19185a257aad77a0e6e7b4e624d1c2.jpg) +Fig. 2: Overview of CaesarNeRF. CaesarNeRF employs a shared encoder to capture two types of features from input views, including scene-level semantic representation $\{S_n\}$ and pixel-level feature representation $\{F_n\}$ . We use the same encoder for both the scene-level semantic representation and the pixel-level embeddings. Following calibration and aggregation of $\{S_n\}$ from various views, we concatenate it with the pixel-level fused feature, processed by the view transformer. Subsequent use of the ray-transformer, coupled with sequential refinement, enables us to render the final RGB values for each pixel in the target view. The output features serve as the input for the next stage, indicated by matching line colors. + +views, and a ray transformer to combine features from different points along a ray, eliminating the need for volume rendering. Further details about GNT can be found in [60]. We also demonstrate that our approach can be extended to other generalizable NeRF models, as discusses in Section 4.2. + +# 3.2 Scene-level semantic representation + +Both encoder-based generalizable NeRF models [6, 59, 76] and their attention-based counterparts [58, 60] mainly rely on pixel-level feature representations. While effective, this approach restricts their capability for a holistic scene understanding, especially when reference views are scarce. This limitation also exacerbates challenges in resolving depth ambiguities between points along the rays, a problem that becomes more pronounced with fewer reference views. + +To address these challenges, we introduce semantic representations aimed at enriching the scene-level understanding. We utilize a shared CNN encoder and apply a Global Average Pooling (GAP) to its $C$ -dimensional output feature map, generating $N$ global feature vectors $\{S_n\}_{n=1}^N$ corresponding to each input view. These feature vectors are then averaged to form a unified scene-level representation $\mathbf{S}$ , i.e., + +$$ +\boldsymbol {S} = \frac {1}{N} \sum_ {n = 1} ^ {N} \boldsymbol {S} _ {n} \in \mathbb {R} ^ {C}. \tag {3} +$$ + +In GNT [60], which uses a view transformer to aggregate pixel-level features into an $L$ -dimensional vector $\bar{F}$ , we extend this by concatenating $\bar{F}$ with $S$ to construct a + +![](images/8f7b9c12a59b68e826371afd7cffc3a1352efbfc3eaf5d12ae4cd257823bd8a8.jpg) +(a) "L: green, R: red" + +![](images/04577ff17977f6c7da53b81e333c8c379aa5ebe0bd6a6e937a1f70b07343c26f.jpg) +(b) "L: red, R: green" +Fig. 3: An illustration of conflicting semantic meanings from multiple viewpoints of the same object. When observing the cup from distinct angles, the features extracted after pooling retain spatial information but are inconsistent in the scene-level semantic understanding, leading to conflicts across various reference images after aggregation. + +global-local embedding $\pmb{E}$ , as formulated by: + +$$ +\boldsymbol {E} = \operatorname {C o n c a t} (\widetilde {\boldsymbol {F}}, \boldsymbol {S}) \in \mathbb {R} ^ {L + C}. \tag {4} +$$ + +This combined embedding $\pmb{E}$ is then subjected to the standard self-attention mechanism used in GNT [60]. This approach enables the scene-level semantic representation $(S)$ to integrate with per-point features $(\widetilde{\pmb{F}})$ , offering a more nuanced understanding at both levels. It also allows each point to selectively draw from the scene-level information. To maintain dimensional consistency across the input and output layers of multiple transformer modules, we employ a two-layer MLP to project the enhanced features back to the original dimension $L$ of the per-point embedding $\widetilde{\pmb{F}}$ . + +# 3.3 Calibration of semantic representation + +The integration of the scene-level semantic representation $S$ , generated through simple averaging of global feature vectors as in (3), improves rendering quality. However, this approach has limitations when dealing with multiple views. As illustrated in Figure 3, viewing the same object from distinct angles may retain spatial attributes but can lead to conflicting semantic meanings. Merely averaging these global feature vectors without accounting for camera positions can result in a distorted scene-level understanding. + +To mitigate this inconsistency, we propose a semantic calibration technique using feature rotation. This adjustment aligns the semantic representation across different camera poses. Our inspiration comes from the use of camera pose projection in computing the fused pixel-level feature $\widetilde{F}$ and is further motivated by [51], which demonstrates that explicit rotation operations in feature spaces are feasible. Unlike point clouds in [51] that inherently lack a defined canonical orientation, NeRF explicitly encodes differences between camera viewpoints, thereby enabling precise calibration between the reference and target images. + +Building on this observation, we calculate calibrated semantic representations $\{\widetilde{S}_n\}_{n=1}^N$ from the $N$ original semantic representations $\{S_n\}_{n=1}^N$ derived from the reference views. We accomplish this by leveraging their respective rotation matrices $\{T_n\}_{n=1}^N$ + +![](images/8c28005788ec4022f614351b3ccdc65476c63ba2e63fa8cd2167983135731c61.jpg) +Fig. 4: Visualization of decoded feature maps for "orchid" in LLFF dataset, produced by ray transformers [60] at different stages. From left to right, the transformer stages increase in depth. + +![](images/c551c24a3094aff2d59889d370cfc1e79cdc141e0a8aba3afef905b181d96f99.jpg) + +![](images/bccc6953b678e6b0d380c9ca486b8b45a81d7086611ef6f84020c1be1b1903a6.jpg) + +to model the rotational variations between each input view and the target view. The alignment of the original semantic features is performed as follows: + +$$ +\widetilde {\boldsymbol {S}} _ {n} = \boldsymbol {\mathcal {P}} \left(\boldsymbol {T} _ {n} \cdot \boldsymbol {\mathcal {P}} ^ {- 1} \left(\boldsymbol {S} _ {n}\right)\right), \text {w h e r e} \boldsymbol {T} _ {n} = \boldsymbol {T} _ {\text {o u t}} ^ {\mathrm {w} 2 \mathrm {c}} \cdot \boldsymbol {T} _ {n} ^ {\mathrm {c} 2 \mathrm {w}}. \tag {5} +$$ + +Here, $T_{n}^{\mathrm{c2w}}$ is the inverse of the extrinsic matrix used for $I_{n}$ , and $T_{\mathrm{out}}^{\mathrm{w2c}}$ is the extrinsic matrix for the target view. $\mathcal{P}(\cdot)$ and $\mathcal{P}^{-1}(\cdot)$ are the flattening and inverse flattening operations, which reshape the feature to a 1D vector of shape 1-by- $C$ and a 2D matrix of shape 3-by- $\frac{C}{3}$ , respectively. + +Note that for the extrinsic matrix, we consider only the top-left $3 \times 3$ submatrix that accounts for rotation. Using GAP to condense feature maps of various sizes into a 1-by- $C$ feature vector eliminates the need for scaling parameters in the semantic representation. As a result, modeling the intrinsic matrix is unnecessary, assuming no skewing, making our approach adaptable to different camera configurations. + +With the calibrated semantic features $\{\widetilde{S}_n\}_{n=1}^N$ for each reference view, we average these, similar to (3), to obtain the calibrated scene-level semantic representation $\widetilde{S}$ , i.e., + +$$ +\widetilde {\boldsymbol {S}} = \frac {1}{N} \sum_ {n = 1} ^ {N} \widetilde {\boldsymbol {S}} _ {n} \in \mathbb {R} ^ {C}. \tag {6} +$$ + +Finally, akin to (4), we concatenate the pixel-level fused feature $\widetilde{F}$ with the calibrated scene-level semantic representation $\widetilde{S}$ to form the final global-local embedding $\widetilde{E}$ : + +$$ +\widetilde {\boldsymbol {E}} = \operatorname {C o n c a t} (\widetilde {\boldsymbol {F}}, \widetilde {\boldsymbol {S}}) \in \mathbb {R} ^ {L + C}. \tag {7} +$$ + +This unified embedding then feeds into ray transformers, passing through standard self-attention mechanisms. In the original GNT [60], multiple view transformers and ray transformers are stacked alternately for sequential feature processing. The last ray transformer integrates features from multiple points along a ray to yield the final RGB value. We denote the corresponding feature representations at stage $k$ as $\widetilde{F}^{(k)}$ and $\widetilde{E}^{(k)}$ . Notably, the calibrated semantic representation $\widetilde{S}$ remains constant across these stages. + +# 3.4 Sequential refinement + +While leveraging $\widetilde{S}$ improves consistency, a single, uniform $\widetilde{S}$ may not be adequate for deeper layers that demand more nuanced details. In fact, we find that deeper transform- + +ers capture finer details compared to shallower ones, as shown in Figure 4. To address this limitation, we introduce a sequential semantic feature refinement module that progressively enriches features at each stage. Specifically, we learn the residual $\pmb{\Delta}^{(k)}$ to update $\widetilde{\pmb{S}}$ at each stage $k$ as follows: + +$$ +\widetilde {\boldsymbol {S}} ^ {(k + 1)} \leftarrow \widetilde {\boldsymbol {S}} ^ {(k)} + \boldsymbol {\Delta} ^ {(k)}. \tag {8} +$$ + +Here, $\pmb{\Delta}^{(k)}$ is calculated by first performing specialized cross-attentions between $\widetilde{\pmb{S}}^{(k)}$ and the original, uncalibrated per-frame semantic features $\{S_n\}_{n=1}^N$ (see Figure 2), followed by their summation. Our goal is to fuse information from different source views to enrich the scene-level semantic representation with features from each reference frame. With this sequential refinement, we combine $\widetilde{\pmb{S}}^{(k)}$ with $\widetilde{\pmb{F}}^{(k)}$ at each stage, yielding a stage-specific global-local embedding $\widetilde{\pmb{E}}^{(k)}$ , which completes our approach. + +Discussion. In scenarios with few reference views, especially when limited to just one, the primary issue is inaccurate depth estimation, resulting in depth ambiguity [12]. This compromises the quality of images when rendered from novel viewpoints. Despite this, essential visual information generally remains accurate across different camera poses. Incorporating our proposed scene-level representation improves the understanding of the overall scene layout [7], distinguishing our approach from existing generalizable NeRF models that predict pixels individually. The advantage of our approach is its holistic view; the semantic representation enriches per-pixel predictions by providing broader context. This semantic constraint ensures that fewer abrupt changes between adjacent points. Consequently, it leads to more reliable depth estimations, making the images rendered from limited reference views more plausible. + +# 3.5 Training objectives + +During training, we employ three different loss functions: + +MSE loss. The Mean Square Error (MSE) loss is the standard photometric loss used in NeRF [43]. It computes the MSE between the actual and predicted pixel values. + +Central loss. Since we project the view-level features from different input camera poses to the shared target view, we introduce a central loss to ensure frame-wise calibrated semantic features $\{\widetilde{S}_n\}_{n=1}^N$ are consistent when projected onto the same target view, which is defined as: + +$$ +\mathcal {L} _ {\text {c e n t r a l}} = \frac {1}{N} \sum_ {n = 1} ^ {N} \left\| \widetilde {\boldsymbol {S}} _ {n} - \widetilde {\boldsymbol {S}} \right\| _ {1}. \tag {9} +$$ + +Point-wise perceptual loss. During the rendering of a batch of pixels in a target view, we inpaint the ground-truth image by replacing the corresponding pixels with the predicted ones. Then, a perceptual loss [26] is computed between the inpainted image and the target image to guide the training process at the whole-image level. + +The final loss function is formulated as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {M S E}} + \lambda_ {1} \mathcal {L} _ {\text {c e n t r a l}} + \lambda_ {2} \mathcal {L} _ {\text {p e r c}}. \tag {10} +$$ + +Empirically, we set $\lambda_1 = 1$ and $\lambda_{2} = 0.001$ , following [32]. + +Table 1: Results for generalizable scene rendering on LLFF with few reference views. GeoNeRF, MatchNeRF, and MVSNeRF necessitate variance as input, defaulting to 0 for single-image cases, hence their results are not included for 1-view scenarios. + +
Method1 reference view2 reference views3 reference views
PSNR (↑)LPIPS (↓)SSIM (↑)PSNR (↑)LPIPS (↓)SSIM (↑)PSNR (↑)LPIPS (↓)SSIM (↑)
PixelNeRF [76]9.320.8980.26411.230.7660.28211.240.6710.486
GPNR [58]15.910.5270.40018.790.3800.57521.570.2880.695
NeuRay [38]16.180.5840.39317.710.3360.64618.260.3100.672
GeoNeRF [25]---18.760.4730.50023.400.2460.766
MatchNeRF [8]---21.080.2720.68922.300.2340.731
MVSNeRF [6]---19.150.3360.70419.840.3140.729
IBRNet [64]16.850.5420.50721.250.3330.68523.000.2620.752
GNT [60]16.570.5000.42420.880.2510.69123.210.1780.782
Ours18.310.4350.52121.940.2240.73623.450.1760.794
+ +# 4 Experiments + +In this section, we first discuss our experimental setups and other details required for our experiments in Section 4.1, followed by our results for different experimental settings, along with ablation studies and analysis in Section 4.2. + +# 4.1 Experimental setups + +Datasets. Firstly, following [60], we construct our training data from both synthetic and real data. This collection includes scanned models from Google Scanned Objects [14], RealEstate10K [81], and handheld phone captures [64]. For evaluation, we utilize real data encompassing complex scenes from sources such as LLFF [42], Shiny [67], and mip-NeRF 360 [4]. Additionally, we train and test our model using the recently released MVImgNet dataset [77]. We adhere to the official split, focusing on examples from the containers category, and select 2,500 scenes for training. During inference, we choose 100 scenes, using their first images as target views and the spatially nearest images as references. Since MVImgNet does not provide camera poses, we utilize COLMAP [54, 55] to deduce the camera positions within these scenes. + +Implementation details. CaesarNeRF is built upon GNT [60], for which we maintain the same configuration, setting the ray and view transformers stack number $(K)$ to 8 for generalizable setting and 4 for single-scene setting. The feature encoder extracts bottleneck features, applies GAP, and then uses a fully connected (FC) layer to reduce the input dimension $C$ to 96. Training involves 500,000 iterations using the Adam optimizer [28], with learning rate set at 0.001 for the feature encoder and 0.0005 for CaesarNeRF, halving them every 100,000 iterations. Each iteration samples 4,096 rays from a single scene. In line with [60], we randomly choose between 8 to 10 reference views for training, and 3 to 7 views when using the MVImgNet [77]. + +Baseline methods. We compare CaesarNeRF with several state-of-the-art methods suited for generalizable NeRF applications, including earlier works such as MVS-NeRF [6], PixelNeRF [76], and IBRNet [64], alongside more recent ones, including GPNR [58], NeuRay [38], GNT [60], GeoNeRF [25] and MatchNeRF [8]. + +Table 2: Results for generalizable scene rendering on Shiny with few reference views. + +
Method1 reference view2 reference views3 reference views
PSNR (↑)LPIPS (↓)SSIM (↑)PSNR (↑)LPIPS (↓)SSIM (↑)PSNR (↑)LPIPS (↓)SSIM (↑)
MatchNeRF [8]---20.280.2780.63620.770.2490.672
MVSNeRF [6]---17.250.4160.57718.550.3430.645
IBRNet [64]14.930.6250.40118.400.4000.59521.960.2810.710
GNT [60]15.990.5480.40020.420.3270.61722.470.2470.720
Ours17.570.4670.47221.470.2930.65222.740.2410.723
+ +Table 3: Results for generalizable scene rendering on mip-NeRF 360 with few reference views. + +
Method1 reference view2 reference views3 reference views
PSNR (↑)LPIPS (↓)SSIM (↑)PSNR (↑)LPIPS (↓)SSIM (↑)PSNR (↑)LPIPS (↓)SSIM (↑)
MatchNeRF [8]---17.000.5660.39217.260.5510.407
MVSNeRF [6]---14.230.6810.36614.290.6740.406
IBRNet [64]14.120.6820.28316.240.6180.36017.700.5550.420
GNT [60]13.480.6300.31415.210.5590.37015.590.5380.395
Ours15.200.5920.35017.050.5380.40317.550.5120.430
+ +# 4.2 Results and analysis + +We compare results in two settings: a generalizable setting, where the model is trained on multiple scenes without fine-tuning during inference for both few and all reference view cases, and a single-scene setting where the model is trained and evaluated on just one scene. Following these comparisons, we conduct ablation studies and test the generalizability of our method with other state-of-the-art approaches. + +Generalizable rendering. In the generalizable setting, we adopt two training strategies. First, we train the model on multiple datasets as described in Section 4.1 and evaluate on LLFF [42], Shiny [67] and mip-NeRF 360 [4] datasets. In addition, the model is trained and tested on the MVImgNet [77] for object-centric generalizability. + +(a) LLFF, Shiny, and mip-NeRF 360. The results for few-reference view scenarios on these datasets are shown in Tables 1, 2 and 3, respectively. Methods like MatchNeRF [8], MVSNeRF [75], and GeoNeRF [25] require at least two reference views. On the LLFF dataset, all methods experience a performance decline as the number of views decreases. CaesarNeRF, however, consistently outperforms others across varying reference view numbers, with the performance gap becoming more significant with fewer views. For example, with 3 views, while IBRNet [64] and GNT [60] have comparable PSNRs, CaesarNeRF demonstrates a more substantial lead in LPIPS and SSIM metrics. + +Similar patterns are observed on the Shiny [67] and mip-NeRF 360 [4] datasets. We apply the highest-performing methods from the LLFF evaluations and report the results + +Table 4: Results on MVImgNet across varying numbers of reference views. 'C.' represents the use of calibration before averaging. + +
Method1 reference view2 reference views3 reference views4 reference views5 reference views
PSNR (†)LPIPS (↓)SSIM (†)PSNR (†)LPIPS (↓)SSIM (†)PSNR (†)LPIPS (↓)SSIM (†)PSNR (†)LPIPS (↓)SSIM (†)PSNR (†)LPIPS (↓)SSIM (†)
IBRNet19.140.4580.59524.380.2660.81825.530.2030.85825.990.1900.86726.120.1880.867
GNT22.220.4330.67826.940.2360.85027.410.2060.87027.510.1970.87527.510.1940.876
Ours w/oC.23.610.3710.71826.340.2740.81727.100.2280.85027.300.2100.86227.340.2030.865
Ours24.280.3340.74727.340.2150.85627.820.1900.87527.920.1810.88127.920.1790.882
+ +Table 5: Results of per-scene optimization on LLFF, in comparison with state-of-the-art methods. + +
MethodLLFF [42]NeRF [44]NeX [67]GNT [60]Ours
PSNR (↑)23.2726.5027.2627.2427.64
LPIPS (↓)0.2120.2500.1790.0870.081
SSIM (↑)0.7980.8110.9040.8890.904
+ +Table 6: Results on LLFF for few-shot generalization after adapting Caesar to other baselines. + +
Method1 reference view2 reference views3 reference views
PSNR (↑)SSIM (↑)LPIPS (↓)PSNR (↑)SSIM (↑)LPIPS (↓)PSNR (↑)SSIM (↑)LPIPS (↓)
MatchNeRF [8]---20.590.7750.27622.430.8050.244
+ Caesar---21.550.7820.26822.980.8240.242
IBRNet [64]16.850.5070.54221.250.6850.33323.000.7520.262
+ Caesar17.760.5430.50022.390.7400.27523.670.7720.242
+ +for those that produce satisfactory outcomes with few reference views. CaesarNeRF maintains superior performance throughout. Notably, for complex datasets like mipNeRF 360 [4], which have sparse camera inputs, the quality of rendered images generally decreases with fewer available reference views. Nonetheless, CaesarNeRF shows the most robust performance compared to the other methods. + +(b) MVImgNet. We extend our comparison of CaesarNeRF with GNT [60] and IBRNet [64] on the MVImgNet dataset, focusing on object-centric scenes, as shown in Table 4. We examine a variant of CaesarNeRF where semantic calibration is substituted with simple feature averaging from multiple frames. While the performance of all methods improves with more views, CaesarNeRF consistently outperforms GNT and IBRNet. Notably, CaesarNeRF with feature averaging surpasses GNT in 1-view case but lags with additional views, implying that the absence of calibration leads to ambiguities when rendering from multiple views. + +Per-scene optimization. Beyond the multi-scene generalizable setting, we demonstrate per-scene optimization results in Table 5. We calculate the average performance over 8 categories from the LLFF dataset [42]. CaesarNeRF consistently outperforms + +Table 7: Ablations on the semantic representation length $\mathbf{R}$ , sequential refinement (Seq.) and calibration (Cali.). 'Ext.' denotes the extension of per-pixel feature to a length of 64 in GNT. + +
Model VariationsPSNR (↑)LPIPS (↓)SSIM (↑)
Rlen.Seq.Cali.
(Baseline GNT)20.930.1850.731
Ext.20.850.1730.735
+3221.430.1520.763
+6421.490.1490.766
+9621.460.1500.766
+12821.490.1470.763
+96✓21.530.1460.770
+96✓21.510.1470.769
+96✓✓21.670.1390.781
+ +![](images/858b3b8516b308fe17dc8e4c054e24dd6a64817d339484ac52acb726d2c85097.jpg) +(b) Using two images as reference view. +Fig. 5: Comparative visualization of our proposed method against other state-of-the-art methods. + +nearly all state-of-the-art methods in the comparison, across all three metrics, showing a significant improvement over our baseline method, GNT [60]. + +Adaptability. To test the adaptability of our Caesar pipeline, we apply it to two other state-of-the-art methods that use view transformers, namely MatchNeRF [8] and IBRNet [64]. We demonstrate in Table 6 that our enhancements in scene-level semantic understanding significantly boost the performance of these methods across all metrics. This indicates that the Caesar framework is not only beneficial in our CaesarNeRF, which is based on GNT [60], but can also be a versatile addition to other NeRF pipelines with view transformers to aggregate different input views. + +Ablation analysis. We conduct ablation studies on the "orchid" scene from the LLFF dataset, with findings detailed in Table 7. Testing variations in representation and the impact of the sequential refinement and calibration modules, we find that increasing the latent size in GNT yields marginal benefits. However, incorporating even a modest semantic representation size distinctly improves results. The length of the semantic representation has a minimal impact on quality. Our ablation studies indicate that while sequential refinement and calibration each offer slight performance gains, their combined effect is most significant. In a single-scene context, semantic information is effectively embedded within the representation, making the benefits of individual modules subtler. Together, however, they provide a framework where sequential refinement can leverage calibrated features for deeper insights. + +![](images/5e8f225582df2f2400796b948c79fb2f54544d41e8e0fe57e7fbb97ba084cb56.jpg) +Fig. 6: Depth estimation prediction using one reference view (first row) and two reference views (second row) as input from LLFF comparing CaesarNeRF with GNT. + +![](images/935270915a1cd7e5f145edcda7312b0c103867507a8548dcf4724c575cce0874.jpg) + +![](images/04c266c0e809b999741573b06e9286bba04568365e4aa0b2d0af948652131fea.jpg) +fern + +![](images/c7f95fcadbe2b730ae942a8335f8543ceb5de0247b4698070d2882aeef12411e.jpg) +chesstable (0.45) + +![](images/d7cbc2bf633ea45fcd2dab003a8f042a898b44f8595dd902e61b08a1c02ea3d4.jpg) +playground (0.46) + +![](images/8b6c91324e0aa73e803c36ed7535388ae3d6e9cbb7637f03af1ff0901b88215a.jpg) +colorfountain (0.50) + +![](images/a538e770a173f7d57c111ae225805b73394531243d0153e86fa675b6362c1a7e.jpg) +bikes (0.50) +Fig.7: Largest and smallest distances for two examples from LLFF test split when matching with training scenes. Numbers $(\times 10^{-2})$ denote the L-2 distance to the source image within the semantic feature space. + +![](images/4afed97bc09049d0a839f5b6c8cf217b7f6cf9bdda1cc4c03c6ee454d42eb1ba.jpg) +fenceflower (1.91) + +![](images/3e87864d5f85b0a1adb736dca13d4289acb30644d97beda0529ccbed111e0708.jpg) +peppers (2.02) + +![](images/9e7836e5d80788e8f2ee9da19d594ab64ce94f4184259f32ff306735b7b5656d.jpg) + +![](images/c07c7dcd68b3781bef9cb67469986fab91e8a6c05910b020cb23a813ba757f3c.jpg) +pond (1.80) +apples (2.01) + +Visualizations. We present our visualization results in Figure 5, where we compare our method with others using one or two views from the LLFF dataset. Additional visual comparisons are provided in the supplementary materials. These visualizations highlight that in scenarios with few views, our method significantly surpasses other generalizable NeRF models, particularly excelling when only a single view is available. CaesarNeRF demonstrates rendering with sharper boundaries and more distinct objects. + +Depth estimation. We extend our evaluation to depth prediction within the LLFF [42] dataset, focusing on challenges presented by few reference views, such as scenarios with just one or two images. In a comparison between CaesarNeRF and GNT [60], we observe in Figure 6 that GNT struggles to accurately capture the relative positions of objects when reference images are sparse. For instance, with only a single view of a flower, CaesarNeRF precisely indicates the flower's proximity to the camera compared to the leaves in the background, a distinction that GNT fails to make. Furthermore, the depth estimations provided by CaesarNeRF are consistently more reliable. In the horn example involving two views, CaesarNeRF offers better boundary delineation, showing particular strength in handling reflective surfaces such as grass in the background. + +Semantic analysis for $\tilde{S}$ . To assess whether the calibrated semantic representation $\tilde{S}$ truly captures semantic details of the scene, we analyze the highest and lowest + +![](images/c47997a4afe51bbe83335e1b3d5cb5d3584e4b767cedb3d24cc4346a96259a24.jpg) +(a) leaves + +![](images/f25a670ca80735b3c0d74ca3048f82dc6e715a6996e9e3285d0333f93be315e4.jpg) +Fig.8: Synthetic results for two examples from LLFF [42], “leaves” and “orchid”, using zero 1-to-3 [35] with one reference image as input and 2 degrees of vertical shift. The left image of each pair is the input, and the right one is the output of zero 1-to-3 [35]. + +![](images/acb39d00ad69cd230a42b3ce357f84f418a987ab62c3dd9c20e1cce0df09bdf0.jpg) +(b) orchid + +![](images/7d2b2c139693508b5ec4cc740792990e53171d1bd735753f62a0e66adbdba032.jpg) + +response based on the L-2 distance between features from two different scenes, interpreting smaller distances as greater similarity. We consider the first image from each category in the LLFF dataset as a reference and extract the scene-level representation for the ten closest views to this reference image using CaesarNeRF. + +In Figure 7, we present two examples, "room" and "fortress", from the LLFF dataset, representing the first image of these categories alongside images corresponding to the top-2 highest and lowest responses. This analysis reveals that the scene-level representation predominantly emphasizes structural information and objects of similar categories. For instance, with the source image "room", the highest responses correlate with images featuring table-like structures, indicating an object-centric focus. In contrast, the lowest responses include images of flowers or ponds found in open areas, which diverge significantly from the structural and object content in the source images. + +Comparison with generative methods. Single-view scenarios are often addressed by generative methods [35-37] that employ diffusion models [19, 53]. While these models can produce reasonable results for object-centric renderings, they struggle with scene-level renderings from novel viewpoints. We show two LLFF examples in Figure 8 using zero 1-to-3 [35], where the left image is the input, and the right one is the output. Images rendered with zero 1-to-3 [35] suffer from style difference. Unlike NeRF-based approaches that reconstruct images from observed pixels, generative models synthesize an entire image from its semantic representation but may not maintain the style. + +# 5 Conclusion and limitation + +In this paper, we introduce CaesarNeRF, a few-shot and generalizable NeRF pipeline that combines scene-level semantic with per-pixel feature representations, aiding in rendering from novel camera positions with limited reference views. We calibrate the semantic representations across different input views and employ a sequential refinement network to offer distinct semantic representations at various levels. Our method has been extensively evaluated on a broad range of datasets, exhibiting state-of-the-art performance in both generalizable and single-scene settings. + +Limitations and potential negative impact. Instead of a generative approach, CaesarNeRF relies on NeRF's scheme, using input images to render the target view. This approach restricts its ability to render parts of the scene that are not present in the reference images. CaesarNeRF may create fake images for authentication. + +# References + +1. Bao, C., Zhang, Y., Yang, B., Fan, T., Yang, Z., Bao, H., Zhang, G., Cui, Z.: Sine: Semantic-driven image-based nerf editing with prior-guided editing field. In: CVPR. pp. 20919-20929 (2023) +2. Bao, Y., Ding, T., Huo, J., Li, W., Li, Y., Gao, Y.: Insertnerf: Instilling generalizability into nef with hypernet modules. ICLR (2024) +3. Bao, Y., Li, Y., Huo, J., Ding, T., Liang, X., Li, W., Gao, Y.: Where and how: Mitigating confusion in neural radiance fields from sparse inputs. arXiv preprint arXiv:2308.02908 (2023) +4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In: CVPR. pp. 5470–5479 (2022) +5. Buehler, C., Bosse, M., McMillan, L., Gortler, S., Cohen, M.: Unstructured lumigraph rendering. In: SIGGRAPH. pp. 425-432 (2001) +6. Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In: ICCV. pp. 14124-14133 (2021) +7. Chen, W., Xu, X., Jia, J., Luo, H., Wang, Y., Wang, F., Jin, R., Sun, X.: Beyond appearance: a semantic controllable self-supervised learning framework for human-centric visual tasks. In: CVPR (2023) +8. Chen, Y., Xu, H., Wu, Q., Zheng, C., Cham, T.J., Cai, J.: Explicit correspondence matching for generalizable neural radiance fields. arXiv preprint arXiv:2304.12294 (2023) +9. Chibane, J., Bansal, A., Lazova, V., Pons-Moll, G.: Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In: CVPR. pp. 7911-7920 (2021) +0. Cong, W., Liang, H., Wang, P., Fan, Z., Chen, T., Varma, M., Wang, Y., Wang, Z.: Enhancing nerf akin to enhancing llms: Generalizable nerf transformer with mixture-of-view-experts. In: ICCV (2023) +1. Debevec, P.E., Taylor, C.J., Malik, J.: Modeling and rendering architecture from photographs: A hybrid geometry-and image-based approach. In: SIGGRAPH. pp. 11-20 (1996) +2. Deng, C., Jiang, C., Qi, C.R., Yan, X., Zhou, Y., Guibas, L., Anguelov, D.: Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In: CVPR. pp. 20637-20647 (2023) +3. Deng, K., Liu, A., Zhu, J.Y., Ramanan, D.: Depth-supervised nerf: Fewer views and faster training for free. In: CVPR. pp. 12882-12891 (2022) +4. Downs, L., Francis, A., Koenig, N., Kinman, B., Hickman, R., Reymann, K., McHugh, T.B., Vanhoucke, V.: Google scanned objects: A high-quality dataset of 3d scanned household items. In: ICRA. pp. 2553-2560 (2022) +5. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: CVPR. pp. 12479-12488 (2023) +6. Fu, Y., Misra, I., Wang, X.: Multiplane nef-supervised disentanglement of depth and camera pose from videos. In: ICML (2022) +7. Gao, Y., Cao, Y.P., Shan, Y.: Surfelnerf: Neural surfel radiance fields for online photorealistic reconstruction of indoor scenes. In: CVPR. pp. 108-118 (2023) +8. Gu, J., Liu, L., Wang, P., Theobalt, C.: Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985 (2021) +9. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS pp. 6840-6851 (2020) +20. Irshad, M.Z., Zakharov, S., Liu, K., Guizilini, V., Kollar, T., Gaidon, A., Kira, Z., Ambrus, R.: Neo 360: Neural fields for sparse view synthesis of outdoor scenes. In: ICCV. pp. 9187-9198 (2023) +21. Jain, A., Mildenhall, B., Barron, J.T., Abbeel, P., Poole, B.: Zero-shot text-guided object generation with dream fields. In: CVPR. pp. 867-876 (2022) + +22. Jain, A., Tancik, M., Abbeel, P.: Putting nerf on a diet: Semantically consistent few-shot view synthesis. In: ICCV. pp. 5885-5894 (2021) +23. Jang, W., Agapito, L.: Codenerf: Disentangled neural radiance fields for object categories. In: ICCV. pp. 12949-12958 (2021) +24. Jiang, Y., Hedman, P., Mildenhall, B., Xu, D., Barron, J.T., Wang, Z., Xue, T.: Alignerf: High-fidelity neural radiance fields via alignment-aware training. In: CVPR. pp. 46-55 (2023) +25. Johari, M.M., Lepoittevin, Y., Fleuret, F.: Geonerf: Generalizing nerf with geometry priors. In: CVPR. pp. 18365-18375 (2022) +26. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and superresolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016) +27. Kania, A., Kasymov, A., Zieba, M., Spurek, P.: Hypernerfgan: Hypernetwork approach to 3d nerf gan. arXiv preprint arXiv:2301.11631 (2023) +28. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +29. Kwak, M., Song, J., Kim, S.: Geconerf: Few-shot neural radiance fields via geometric consistency. arXiv preprint arXiv:2301.10941 (2023) +30. Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: Dynibar: Neural dynamic image-based rendering. In: CVPR. pp. 4273-4284 (2023) +31. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: CVPR. pp. 300–309 (2023) +32. Lin, H., Peng, S., Xu, Z., Yan, Y., Shuai, Q., Bao, H., Zhou, X.: Efficient neural radiance fields for interactive free-viewpoint video. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022) +33. Lin, Y., Bai, H., Li, S., Lu, H., Lin, X., Xiong, H., Wang, L.: Componerf: Text-guided multi-object compositional nef with editable 3d scene layout. arXiv preprint arXiv:2303.13843 (2023) +34. Liu, L., Habermann, M., Rudnev, V., Sarkar, K., Gu, J., Theobalt, C.: Neural actor: Neural free-view synthesis of human actors with pose control. TOC 40(6), 1-16 (2021) +35. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: ICCV (2023) +36. Liu, X., Kao, S.h., Chen, J., Tai, Y.W., Tang, C.K.: Deceptive-nerf: Enhancing nerf reconstruction using pseudo-observations from diffusion models. arXiv preprint arXiv:2305.15171 (2023) +37. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453 (2023) +38. Liu, Y., Peng, S., Liu, L., Wang, Q., Wang, P., Theobalt, C., Zhou, X., Wang, W.: Neural rays for occlusion-aware image-based rendering. In: CVPR. pp. 7824-7833 (2022) +39. Mariotti, O., Mac Aodha, O., Bilen, H.: Viewnerf: Unsupervised viewpoint estimation using category-level neural radiance fields. arXiv preprint arXiv:2212.00436 (2022) +40. Martin-Brualla, R., Radwan, N., Sajjadi, M.S., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: CVPR. pp. 7210-7219 (2021) +41. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: CVPR. pp. 12663–12673 (2023) +42. Mildenhall, B., Srinivasan, P.P., Ortiz-Cayon, R., Kalantari, N.K., Ramamoorthi, R., Ng, R., Kar, A.: Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. TOG 38(4), 1-14 (2019) + +43. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) +44. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) +45. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: CVPR. pp. 5480-5490 (2022) +46. Noguchi, A., Sun, X., Lin, S., Harada, T.: Neural articulated radiance field. In: ICCV. pp. 5762-5772 (2021) +47. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: ICCV. pp. 5865-5874 (2021) +48. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable neural radiance fields for modeling dynamic human bodies. In: ICCV. pp. 14314-14323 (2021) +49. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022) +50. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: CVPR. pp. 10318-10327 (2021) +51. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR. pp. 652-660 (2017) +52. Roessle, B., Barron, J.T., Mildenhall, B., Srinivasan, P.P., Nießner, M.: Dense depth priors for neural radiance fields from sparse input views. In: CVPR. pp. 12892-12901 (2022) +53. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022) +54. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: CVPR (2016) +55. Schonberger, J.L., Zheng, E., Pollefeys, M., Frahm, J.M.: Pixelwise view selection for unstructured multi-view stereo. In: ECCV (2016) +56. Schwarz, K., Liao, Y., Niemeyer, M., Geiger, A.: Graf: Generative radiance fields for 3d-aware image synthesis. NeurIPS 33, 20154-20166 (2020) +57. Shue, J.R., Chan, E.R., Po, R., Ankner, Z., Wu, J., Wetzstein, G.: 3d neural field generation using triplane diffusion. In: CVPR. pp. 20875-20886 (2023) +58. Suhail, M., Esteves, C., Sigal, L., Makadia, A.: Generalizable patch-based neural rendering. In: ECCV. pp. 156-174 (2022) +59. Trevithick, A., Yang, B.: Grf: Learning a general radiance field for 3d representation and rendering. In: ICCV. pp. 15182-15192 (2021) +60. Varma, M., Wang, P., Chen, X., Chen, T., Venugopalan, S., Wang, Z.: Is attention all that nerf needs? In: ICLR (2023) +61. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017) +62. Waechter, M., Moehrle, N., Goesele, M.: Let there be color! large-scale texturing of 3d reconstructions. In: ECCV. pp. 836-850 (2014) +63. Wang, G., Chen, Z., Loy, C.C., Liu, Z.: Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In: ICCV (2023) +64. Wang, Q., Wang, Z., Genova, K., Srinivasan, P.P., Zhou, H., Barron, J.T., Martin-Brualla, R., Snavely, N., Funkhouser, T.: Ibrnet: Learning multi-view image-based rendering. In: CVPR. pp. 4690-4699 (2021) +65. Wang, T., Zhang, B., Zhang, T., Gu, S., Bao, J., Baltrusaitis, T., Shen, J., Chen, D., Wen, F., Chen, Q., et al.: Rodin: A generative model for sculpting 3d digital avatars using diffusion. In: CVPR. pp. 4563-4573 (2023) +66. Wei, Y., Liu, S., Rao, Y., Zhao, W., Lu, J., Zhou, J.: Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In: CVPR. pp. 5610-5619 (2021) + +67. Wizadwongsa, S., Phongthawee, P., Yenphraphai, J., Suwajanakorn, S.: Nex: Real-time view synthesis with neural basis expansion. In: CVPR. pp. 8534-8543 (2021) +68. Xiangli, Y., Xu, L., Pan, X., Zhao, N., Rao, A., Theobalt, C., Dai, B., Lin, D.: Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In: ECCV (2022) +69. Xie, C., Park, K., Martin-Brualla, R., Brown, M.: Fig-nerf: Figure-ground neural radiance fields for 3d object category modelling. In: 3DV. pp. 962–971 (2021) +70. Xu, D., Jiang, Y., Wang, P., Fan, Z., Shi, H., Wang, Z.: Sinnerf: Training neural radiance fields on complex scenes from a single image. In: ECCV. pp. 736-753 (2022) +71. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: CVPR. pp. 5438-5448 (2022) +72. Yang, B., Zhang, Y., Xu, Y., Li, Y., Zhou, H., Bao, H., Zhang, G., Cui, Z.: Learning object-compositional neural radiance field for editable scene rendering. In: ICCV. pp. 13779-13788 (2021) +73. Yang, H., Hong, L., Li, A., Hu, T., Li, Z., Lee, G.H., Wang, L.: Contranerf: Generalizable neural radiance fields for synthetic-to-real novel view synthesis via contrastive learning. In: CVPR. pp. 16508-16517 (2023) +74. Yang, J., Pavone, M., Wang, Y.: Freenerf: Improving few-shot neural rendering with free frequency regularization. In: CVPR. pp. 8254-8263 (2023) +75. Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstructured multiview stereo. In: ECCV. pp. 767-783 (2018) +76. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: CVPR. pp. 4578-4587 (2021) +77. Yu, X., Xu, M., Zhang, Y., Liu, H., Ye, C., Wu, Y., Yan, Z., Zhu, C., Xiong, Z., Liang, T., et al.: Mvimgnet: A large-scale dataset of multi-view images. In: CVPR. pp. 9150-9161 (2023) +78. Zhang, J., Yang, G., Tulsiani, S., Ramanan, D.: Ners: Neural reflectance surfaces for sparse-view 3d reconstruction in the wild. NeurIPS 34, 29835-29847 (2021) +79. Zhang, J., Li, X., Wan, Z., Wang, C., Liao, J.: Text2nerf: Text-driven 3d scene generation with neural radiance fields. arXiv preprint arXiv:2305.11588 (2023) +80. Zhenxing, M., Xu, D.: Switch-nerf: Learning scene decomposition with mixture of experts for large-scale neural radiance fields. In: ICLR (2023) +81. Zhou, T., Tucker, R., Flynn, J., Fyffe, G., Snavely, N.: Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817 (2018) +82. Zhou, Z., Tulsiani, S.: Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In: CVPR. pp. 12588-12597 (2023) +83. Zhu, H., Sun, Y., Liu, C., Xia, L., Luo, J., Qiao, N., Nevatia, R., Kuo, C.H.: Multimodal neural radiance field. In: ICRA. pp. 9393-9399 (2023) +84. Zhu, H., Zheng, Z., Zheng, W., Nevatia, R.: Cat-nerf: Constancy-aware tx2former for dynamic body modeling. In: CVPRW. pp. 6618-6627 (2023) +85. Zhuang, Y., Zhu, H., Sun, X., Cao, X.: Mofanerf: Morphable facial neural radiance field. In: ECCV. pp. 268-285 (2022) +86. Zimny, D., Trzeciński, T., Spurek, P.: Points2nerf: Generating neural radiance fields from 3d point cloud. arXiv preprint arXiv:2206.01290 (2022) \ No newline at end of file diff --git a/2024/CaesarNeRF_ Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering/images.zip b/2024/CaesarNeRF_ Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4e7959edc05427e08ae2b5356cb5e9f9cc226015 --- /dev/null +++ b/2024/CaesarNeRF_ Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81dd81edad5cece7ca9583de5c7c656f3c063e76a1f91fdb0b9d3a9b7184e56d +size 778647 diff --git a/2024/CaesarNeRF_ Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering/layout.json b/2024/CaesarNeRF_ Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..303f48c9b27882246bfa116c38bb3ab854c59356 --- /dev/null +++ b/2024/CaesarNeRF_ Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering/layout.json @@ -0,0 +1,11551 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 475, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 475, + 148 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 475, + 148 + ], + "type": "text", + "content": "CaesarNeRF: Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 168, + 206, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 168, + 206, + 182 + ], + "spans": [ + { + "bbox": [ + 140, + 168, + 206, + 182 + ], + "type": "text", + "content": "Haidong Zhu" + }, + { + "bbox": [ + 140, + 168, + 206, + 182 + ], + "type": "inline_equation", + "content": "^{1,*}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 228, + 169, + 303, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 169, + 303, + 194 + ], + "spans": [ + { + "bbox": [ + 228, + 169, + 303, + 194 + ], + "type": "text", + "content": "Tianyu Ding" + }, + { + "bbox": [ + 228, + 169, + 303, + 194 + ], + "type": "inline_equation", + "content": "^{2,\\ast,\\dagger}" + }, + { + "bbox": [ + 228, + 169, + 303, + 194 + ], + "type": "text", + "content": " Ram Nevatia" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 331, + 169, + 386, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 169, + 386, + 194 + ], + "spans": [ + { + "bbox": [ + 331, + 169, + 386, + 194 + ], + "type": "text", + "content": "Tianyi Chen" + }, + { + "bbox": [ + 331, + 169, + 386, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 331, + 169, + 386, + 194 + ], + "type": "text", + "content": ", Ning Liang" + }, + { + "bbox": [ + 331, + 169, + 386, + 194 + ], + "type": "inline_equation", + "content": "^{2,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 414, + 169, + 473, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 169, + 473, + 182 + ], + "spans": [ + { + "bbox": [ + 414, + 169, + 473, + 182 + ], + "type": "text", + "content": "Ilya Zharkov2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 205, + 202, + 408, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 202, + 408, + 225 + ], + "spans": [ + { + "bbox": [ + 205, + 202, + 408, + 225 + ], + "type": "text", + "content": "1University of Southern California \n2Microsoft \n{haidongz, nevatia}@usc.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 175, + 226, + 440, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 226, + 440, + 236 + ], + "spans": [ + { + "bbox": [ + 175, + 226, + 440, + 236 + ], + "type": "text", + "content": "{tianyuding,tiachen,zharkov,lulian}@microsoft.com" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 133, + 257, + 481, + 390 + ], + "blocks": [ + { + "bbox": [ + 133, + 257, + 481, + 390 + ], + "lines": [ + { + "bbox": [ + 133, + 257, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 133, + 257, + 481, + 390 + ], + "type": "image", + "image_path": "8e8110b63c99922a92cd84756438c1580ba201780eb229980ac533d92538a796.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 397, + 482, + 431 + ], + "lines": [ + { + "bbox": [ + 132, + 397, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 482, + 431 + ], + "type": "text", + "content": "Fig.1: Novel view synthesis for novel scenes using ONE reference view on Shiny [67], LLFF [42], and MVImgNet [77] (top to bottom). Each triplet of images corresponds to the results from GNT [60] (left), CaesarNeRF (middle) and groundtruth (right)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 456, + 455, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 456, + 455, + 599 + ], + "spans": [ + { + "bbox": [ + 159, + 456, + 455, + 599 + ], + "type": "text", + "content": "Abstract. Generalizability and few-shot learning are key challenges in Neural Radiance Fields (NeRF), often due to the lack of a holistic understanding in pixel-level rendering. We introduce CaesarNeRF, an end-to-end approach that leverages scene-level CAlibratEd SemAntic Representation along with pixel-level representations to advance few-shot, generalizable neural rendering, facilitating a holistic understanding without compromising high-quality details. CaesarNeRF explicitly models pose differences of reference views to combine scene-level semantic representations, providing a calibrated holistic understanding. This calibration process aligns various viewpoints with precise location and is further enhanced by sequential refinement to capture varying details. Extensive experiments on public datasets, including LLFF, Shiny, mip-NeRF 360, and MVImgNet, show that CaesarNeRF delivers state-of-the-art performance across varying numbers of reference views, proving effective even with a single reference image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 160, + 621, + 430, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 621, + 430, + 633 + ], + "spans": [ + { + "bbox": [ + 160, + 621, + 430, + 633 + ], + "type": "text", + "content": "Keywords: Few-view rendering " + }, + { + "bbox": [ + 160, + 621, + 430, + 633 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 621, + 430, + 633 + ], + "type": "text", + "content": " Generalizable NeRF " + }, + { + "bbox": [ + 160, + 621, + 430, + 633 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 621, + 430, + 633 + ], + "type": "text", + "content": " Neural rendering" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 643, + 392, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 643, + 392, + 665 + ], + "spans": [ + { + "bbox": [ + 142, + 643, + 392, + 665 + ], + "type": "inline_equation", + "content": "^{*}" + }, + { + "bbox": [ + 142, + 643, + 392, + 665 + ], + "type": "text", + "content": "Equal contribution. " + }, + { + "bbox": [ + 142, + 643, + 392, + 665 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 142, + 643, + 392, + 665 + ], + "type": "text", + "content": "Corresponding author. This work was done when Haidong Zhu was an intern at Microsoft." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 218, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 218, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 218, + 128 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 481, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 481, + 236 + ], + "type": "text", + "content": "Rendering a scene from a novel camera position is essential in view synthesis [5,11,62]. The recent advancement of Neural Radiance Field (NeRF) [44] has shown impressive results in creating photo-realistic images from novel viewpoints. However, conventional NeRF methods are either typically scene-specific, necessitating retraining for novel scenes [15, 16, 44, 67, 74], or require a large number of reference views as input for generalizing to novel scenarios [6, 58, 60, 76]. These constraints highlight the complexity of the few-shot generalizable neural rendering, which aims to render unseen scenes from novel viewpoints with a limited number of reference images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 237, + 481, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 404 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 404 + ], + "type": "text", + "content": "Generalizing NeRF to novel scenes often involves using pixel-level feature embeddings encoded from input images, as seen in existing methods [59, 76]. These methods adapt NeRF to novel scenes by separating the scene representation from the model through an image encoder. However, relying solely on pixel-level features has its drawbacks: it requires highly precise epipolar geometry and often overlooks occlusion in complex scenes. Moreover, employing pixel-level features ignores the inherent interconnections within objects in the scene, treating the prediction of each pixel independently. Prior attempts to utilize scene-level representation either suffers from style mismatches [35] during scene-wide rendering or have been limited to specific object categories [23, 39, 69]. With few input reference under generalizable settings, these limitations become more pronounced, exacerbating the ambiguity in predictions due to biases from camera viewpoints. Although recent diffusion-based models [18, 56] attempt to address these issues through generative approaches, they struggle to effectively use the input views as contextual references for specific scenes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 405, + 481, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 405, + 481, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 405, + 481, + 608 + ], + "type": "text", + "content": "We present CaesarNeRF, a method that advances the generalizability of NeRF by incorporating calibrated semantic representation. This enables rendering from novel viewpoints using as few as one input reference view, as depicted in Figure 1. Our approach combines semantic scene-level representation with per-pixel features, enhancing consistency across different views of the same scene. The encoder-generated scene-level representations capture both semantic features and biases linked to specific camera poses. When reference views are limited, these biases can introduce uncertainty in the rendered images. To counter this, CaesarNeRF integrates camera pose transformations into the semantic representation, hence the term calibrated. By isolating pose-specific information from the scene-level representation, our model harmonizes features across input views, mitigating view-specific biases and, in turn, reducing ambiguity. In addition, CaesarNeRF introduces a sequential refinement process, which equips the model with varying levels of detail needed to enhance the semantic features. Extensive experiments on datasets such as LLFF [42], Shiny [67], mip-NeRF 360 [4], and the newly released MVImgNet [77] demonstrate that CaesarNeRF outperforms current state-of-the-art methods with limited reference views available, proving effective in generalizable settings with as few as one reference view. The project can be found here." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 147, + 609, + 330, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 609, + 330, + 620 + ], + "spans": [ + { + "bbox": [ + 147, + 609, + 330, + 620 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 629, + 481, + 665 + ], + "type": "text", + "content": "- We introduce CaesarNeRF, which utilizes scene-level calibrated semantic representation to achieve few-shot, generalizable neural rendering. This innovation leads to coherent and high-quality renderings." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "text", + "content": "H. Zhu and T. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 481, + 163 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 481, + 163 + ], + "type": "text", + "content": "- We integrate semantic scene context with pixel-level details, in contrast to existing methods that rely solely on pixel-level features. We also address view-specific biases by modeling camera pose transformations and enhance the scene understanding through the sequential refinement of semantic features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 164, + 481, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 164, + 481, + 212 + ], + "spans": [ + { + "bbox": [ + 138, + 164, + 481, + 212 + ], + "type": "text", + "content": "- We demonstrate through extensive experiments that CaesarNeRF consistently outperforms state-of-the-art generalizable NeRF methods across a variety of datasets. Furthermore, integrating the Caesar pipeline into other baseline methods leads to consistent performance gains, highlighting its effectiveness and adaptability." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 232, + 222, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 232, + 222, + 244 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 222, + 244 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 258, + 482, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 353 + ], + "type": "text", + "content": "Neural Radiance Field (NeRF) implicitly captures the density and appearance of points within a scene or object [43, 44] and enables rendering from novel camera positions. In recent years, NeRF has witnessed improvements in a wide range of applications, such as photo-realistic novel view synthesis for large-scale scenes [40,68,80], dynamic scene decomposition and deformation [24,30,34,46-48,50,84,85], occupancy or depth estimation [66,71,83], scene generation and editing [1,21,31,33,41,49,72,73,79], and so on. Despite these advances, most methods still rely on the original NeRF and require retraining or fine-tuning for novel scenes not covered in the training data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 354, + 482, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 497 + ], + "type": "text", + "content": "Generalizable NeRF aims to adapt a single NeRF model to multiple scenes by separating the scene representation from the model. This field has seen notable advancements, with efforts focused on avoiding the need for retraining [8,20,38,64,65,75,76]. PixelNeRF [76] and GRF [59] pioneered the application of an image encoder to transform images into per-pixel features, with NeRF functioning as a decoder for predicting density and color from these features. MVSNeRF [6] introduces the use of a cost volume from MVSNet [75] to encode 3-D features from multiple views. Recognizing the intrinsic connection between points along a ray, IBRNet [64] employs self-attention to enhance point density predictions. Transformer-based [61] networks like GNT [10,60], GeoNeRF [25], and GPNR [58] are explored as alternatives to volume rendering, concentrating on pixel and patch-level representations. Additionally, InsertNeRF [2] utilizes hypernet modules to adapt parameters for novel scenes efficiently." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 545 + ], + "type": "text", + "content": "These methods primarily depend on image encoders to extract pixel-aligned features from reference views. As a result, many of them lack a comprehensive understanding of the entire scene. Furthermore, with few reference views, the features become intertwined with view-specific details, compromising the quality of the rendering results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "type": "text", + "content": "Few-shot Neural Radiance Field aims to render novel views using a limited number of reference images. To this end, various methods have been developed, incorporating additional information such as normalization-flow [78], semantic constraints [17, 22], depth cues [13, 52], geometry consistency [3, 29, 45, 63, 70], and frequency content [74]. Others [6, 9] emphasize pretraining on large-scale datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": "While these methods offer reasonable reconstructions with few inputs, they typically still require training or fine-tuning for specific scenes. Moreover, these methods usually require at least three reference images. With fewer than three, view-specific biases lead to ambiguity, complicating the rendering. Diffusion-based [12, 35, 57, 82] and other generative methods [27, 86] have been explored for single-view synthesis or gen-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "erative rendering, yet they are mostly limited to single-object rendering and generally fall short for complex scenes, which often result in a style change." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": "CaesarNeRF confronts the above challenges by leveraging calibrated semantic representations that exploit scene geometry and variations in camera viewpoints. As a result, CaesarNeRF overcomes the limitations of pixel-level features and reduces dependency on external data or extensive pretraining, delivering high-quality renderings in few-shot and generalizable settings." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 218, + 265, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 218, + 265, + 232 + ], + "spans": [ + { + "bbox": [ + 131, + 218, + 265, + 232 + ], + "type": "text", + "content": "3 The proposed method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 243, + 479, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 479, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 479, + 292 + ], + "type": "text", + "content": "We first outline the general framework of existing generalizable NeRF in Section 3.1. Then, we present our proposed CaesarNeRF, as illustrated in Figure 2. This model integrates elements of semantic representation, calibration, and sequential refinement, detailed in Section 3.2, 3.3, and 3.4, respectively. The training objective is given in 3.5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 308, + 288, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 308, + 288, + 320 + ], + "spans": [ + { + "bbox": [ + 131, + 308, + 288, + 320 + ], + "type": "text", + "content": "3.1 NeRF and generalizable NeRF" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 328, + 480, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 480, + 375 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 480, + 375 + ], + "type": "text", + "content": "Neural Radiance Field (NeRF) [43, 44] aims to render 3D scenes by predicting both the density and RGB values at points where light rays intersect the radiance field. For a query point " + }, + { + "bbox": [ + 130, + 328, + 480, + 375 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 328, + 480, + 375 + ], + "type": "text", + "content": " and a viewing direction " + }, + { + "bbox": [ + 130, + 328, + 480, + 375 + ], + "type": "inline_equation", + "content": "\\pmb{d}" + }, + { + "bbox": [ + 130, + 328, + 480, + 375 + ], + "type": "text", + "content": " on the unit sphere " + }, + { + "bbox": [ + 130, + 328, + 480, + 375 + ], + "type": "inline_equation", + "content": "\\mathbb{S}^2" + }, + { + "bbox": [ + 130, + 328, + 480, + 375 + ], + "type": "text", + "content": " in 3D space, the NeRF model " + }, + { + "bbox": [ + 130, + 328, + 480, + 375 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 328, + 480, + 375 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 272, + 385, + 481, + 398 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 385, + 481, + 398 + ], + "spans": [ + { + "bbox": [ + 272, + 385, + 481, + 398 + ], + "type": "interline_equation", + "content": "\\sigma , \\boldsymbol {c} = \\mathcal {F} (\\boldsymbol {x}, \\boldsymbol {d}). \\tag {1}", + "image_path": "ab2f629e861edd1991fa73a096cc489e16b36daabc69e6401adfc9177f2a7911.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 407, + 480, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 480, + 454 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 480, + 454 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 407, + 480, + 454 + ], + "type": "inline_equation", + "content": "\\sigma \\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 407, + 480, + 454 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 407, + 480, + 454 + ], + "type": "inline_equation", + "content": "c\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 407, + 480, + 454 + ], + "type": "text", + "content": " denote the density and the RGB values, respectively. After computing these values for a collection of discretized points along each ray, volume rendering techniques are employed to calculate the final RGB values for each pixel, thus reconstructing the image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "text", + "content": "However, traditional NeRF models " + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "text", + "content": " are limited by their requirement for scene-specific training, making it unsuitable for generalizing to novel scenes. To overcome this, generalizable NeRF models, denoted by " + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_G" + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "text", + "content": ", are designed to render images of novel scenes without per-scene training. Given " + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "text", + "content": " reference images " + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "inline_equation", + "content": "\\{I_n\\}_{n=1}^N" + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "text", + "content": ", an encoder-based generalizable NeRF model " + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_G" + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "text", + "content": " decouples the object representation from the original NeRF by using an encoder to extract per-pixel feature maps " + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "inline_equation", + "content": "\\{F_n\\}_{n=1}^N" + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "text", + "content": " from the input images. To synthesize a pixel associated with a point " + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "text", + "content": " along a ray in direction " + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "text", + "content": ", it projects " + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "inline_equation", + "content": "\\{F_n\\}_{n=1}^N" + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "text", + "content": " from nearby views and aggregates this multi-view pixel-level information using techniques such as average pooling [76] or cost volumes [6]. This results in a fused feature embedding " + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{F}}" + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "text", + "content": ", allowing " + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_G" + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "text", + "content": " to predict density " + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "text", + "content": " and RGB values " + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 455, + 484, + 586 + ], + "type": "text", + "content": " for each point along the ray, as expressed by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 261, + 594, + 481, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 594, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 261, + 594, + 481, + 609 + ], + "type": "interline_equation", + "content": "\\sigma , \\boldsymbol {c} = \\mathcal {F} _ {\\boldsymbol {G}} (\\boldsymbol {x}, \\boldsymbol {d}, \\tilde {\\boldsymbol {F}}). \\tag {2}", + "image_path": "ad2c18563142122b9e6483b365ff8f03c094a3dd9a7902913683e59d5486a3fa.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "In our method, we adopt the recently introduced fully attention-based generalizable NeRF method, GNT [60], as both the backbone and the baseline. GNT shares a similar paradigm with (2) but employs transformers [61] to aggregate pixel-level features into " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\widetilde{F}" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": ". It uses a view transformer to fuse projected pixel-level features from reference" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "text", + "content": "H. Zhu and T. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 113, + 467, + 248 + ], + "blocks": [ + { + "bbox": [ + 149, + 113, + 467, + 248 + ], + "lines": [ + { + "bbox": [ + 149, + 113, + 467, + 248 + ], + "spans": [ + { + "bbox": [ + 149, + 113, + 467, + 248 + ], + "type": "image", + "image_path": "52879611db002d2c8b11c42bf8c0ec659f19185a257aad77a0e6e7b4e624d1c2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 255, + 482, + 344 + ], + "lines": [ + { + "bbox": [ + 130, + 255, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 482, + 344 + ], + "type": "text", + "content": "Fig. 2: Overview of CaesarNeRF. CaesarNeRF employs a shared encoder to capture two types of features from input views, including scene-level semantic representation " + }, + { + "bbox": [ + 130, + 255, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\{S_n\\}" + }, + { + "bbox": [ + 130, + 255, + 482, + 344 + ], + "type": "text", + "content": " and pixel-level feature representation " + }, + { + "bbox": [ + 130, + 255, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\{F_n\\}" + }, + { + "bbox": [ + 130, + 255, + 482, + 344 + ], + "type": "text", + "content": ". We use the same encoder for both the scene-level semantic representation and the pixel-level embeddings. Following calibration and aggregation of " + }, + { + "bbox": [ + 130, + 255, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\{S_n\\}" + }, + { + "bbox": [ + 130, + 255, + 482, + 344 + ], + "type": "text", + "content": " from various views, we concatenate it with the pixel-level fused feature, processed by the view transformer. Subsequent use of the ray-transformer, coupled with sequential refinement, enables us to render the final RGB values for each pixel in the target view. The output features serve as the input for the next stage, indicated by matching line colors." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 369, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 416 + ], + "type": "text", + "content": "views, and a ray transformer to combine features from different points along a ray, eliminating the need for volume rendering. Further details about GNT can be found in [60]. We also demonstrate that our approach can be extended to other generalizable NeRF models, as discusses in Section 4.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 436, + 309, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 309, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 309, + 449 + ], + "type": "text", + "content": "3.2 Scene-level semantic representation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 458, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 482, + 529 + ], + "type": "text", + "content": "Both encoder-based generalizable NeRF models [6, 59, 76] and their attention-based counterparts [58, 60] mainly rely on pixel-level feature representations. While effective, this approach restricts their capability for a holistic scene understanding, especially when reference views are scarce. This limitation also exacerbates challenges in resolving depth ambiguities between points along the rays, a problem that becomes more pronounced with fewer reference views." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 530, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 530, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 530, + 482, + 590 + ], + "type": "text", + "content": "To address these challenges, we introduce semantic representations aimed at enriching the scene-level understanding. We utilize a shared CNN encoder and apply a Global Average Pooling (GAP) to its " + }, + { + "bbox": [ + 130, + 530, + 482, + 590 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 530, + 482, + 590 + ], + "type": "text", + "content": "-dimensional output feature map, generating " + }, + { + "bbox": [ + 130, + 530, + 482, + 590 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 530, + 482, + 590 + ], + "type": "text", + "content": " global feature vectors " + }, + { + "bbox": [ + 130, + 530, + 482, + 590 + ], + "type": "inline_equation", + "content": "\\{S_n\\}_{n=1}^N" + }, + { + "bbox": [ + 130, + 530, + 482, + 590 + ], + "type": "text", + "content": " corresponding to each input view. These feature vectors are then averaged to form a unified scene-level representation " + }, + { + "bbox": [ + 130, + 530, + 482, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 130, + 530, + 482, + 590 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 258, + 600, + 481, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 600, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 258, + 600, + 481, + 633 + ], + "type": "interline_equation", + "content": "\\boldsymbol {S} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\boldsymbol {S} _ {n} \\in \\mathbb {R} ^ {C}. \\tag {3}", + "image_path": "fa962bd93121e56a6d027bc6f6fc6e1471cb94436396820266b4c46639ef59a1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "type": "text", + "content": "In GNT [60], which uses a view transformer to aggregate pixel-level features into an " + }, + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "type": "text", + "content": "-dimensional vector " + }, + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\bar{F}" + }, + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "type": "text", + "content": ", we extend this by concatenating " + }, + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\bar{F}" + }, + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "type": "text", + "content": " to construct a" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 448, + 103 + ], + "type": "text", + "content": "Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 192, + 118, + 293, + 193 + ], + "blocks": [ + { + "bbox": [ + 192, + 118, + 293, + 193 + ], + "lines": [ + { + "bbox": [ + 192, + 118, + 293, + 193 + ], + "spans": [ + { + "bbox": [ + 192, + 118, + 293, + 193 + ], + "type": "image", + "image_path": "8f7b9c12a59b68e826371afd7cffc3a1352efbfc3eaf5d12ae4cd257823bd8a8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 206, + 269, + 215 + ], + "lines": [ + { + "bbox": [ + 214, + 206, + 269, + 215 + ], + "spans": [ + { + "bbox": [ + 214, + 206, + 269, + 215 + ], + "type": "text", + "content": "(a) \"L: green, R: red\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 302, + 119, + 402, + 192 + ], + "blocks": [ + { + "bbox": [ + 302, + 119, + 402, + 192 + ], + "lines": [ + { + "bbox": [ + 302, + 119, + 402, + 192 + ], + "spans": [ + { + "bbox": [ + 302, + 119, + 402, + 192 + ], + "type": "image", + "image_path": "04577ff17977f6c7da53b81e333c8c379aa5ebe0bd6a6e937a1f70b07343c26f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 206, + 380, + 215 + ], + "lines": [ + { + "bbox": [ + 323, + 206, + 380, + 215 + ], + "spans": [ + { + "bbox": [ + 323, + 206, + 380, + 215 + ], + "type": "text", + "content": "(b) \"L: red, R: green\"" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 224, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 269 + ], + "type": "text", + "content": "Fig. 3: An illustration of conflicting semantic meanings from multiple viewpoints of the same object. When observing the cup from distinct angles, the features extracted after pooling retain spatial information but are inconsistent in the scene-level semantic understanding, leading to conflicts across various reference images after aggregation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 293, + 317, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 293, + 317, + 304 + ], + "spans": [ + { + "bbox": [ + 131, + 293, + 317, + 304 + ], + "type": "text", + "content": "global-local embedding " + }, + { + "bbox": [ + 131, + 293, + 317, + 304 + ], + "type": "inline_equation", + "content": "\\pmb{E}" + }, + { + "bbox": [ + 131, + 293, + 317, + 304 + ], + "type": "text", + "content": ", as formulated by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 246, + 313, + 481, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 313, + 481, + 328 + ], + "spans": [ + { + "bbox": [ + 246, + 313, + 481, + 328 + ], + "type": "interline_equation", + "content": "\\boldsymbol {E} = \\operatorname {C o n c a t} (\\widetilde {\\boldsymbol {F}}, \\boldsymbol {S}) \\in \\mathbb {R} ^ {L + C}. \\tag {4}", + "image_path": "3ae9ce51237f6027bdcc3c89f7ed87aa19a439aca5d6d64cd001ba9935859a79.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 337, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 337, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 482, + 422 + ], + "type": "text", + "content": "This combined embedding " + }, + { + "bbox": [ + 130, + 337, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\pmb{E}" + }, + { + "bbox": [ + 130, + 337, + 482, + 422 + ], + "type": "text", + "content": " is then subjected to the standard self-attention mechanism used in GNT [60]. This approach enables the scene-level semantic representation " + }, + { + "bbox": [ + 130, + 337, + 482, + 422 + ], + "type": "inline_equation", + "content": "(S)" + }, + { + "bbox": [ + 130, + 337, + 482, + 422 + ], + "type": "text", + "content": " to integrate with per-point features " + }, + { + "bbox": [ + 130, + 337, + 482, + 422 + ], + "type": "inline_equation", + "content": "(\\widetilde{\\pmb{F}})" + }, + { + "bbox": [ + 130, + 337, + 482, + 422 + ], + "type": "text", + "content": ", offering a more nuanced understanding at both levels. It also allows each point to selectively draw from the scene-level information. To maintain dimensional consistency across the input and output layers of multiple transformer modules, we employ a two-layer MLP to project the enhanced features back to the original dimension " + }, + { + "bbox": [ + 130, + 337, + 482, + 422 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 337, + 482, + 422 + ], + "type": "text", + "content": " of the per-point embedding " + }, + { + "bbox": [ + 130, + 337, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{F}}" + }, + { + "bbox": [ + 130, + 337, + 482, + 422 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 441, + 322, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 441, + 322, + 453 + ], + "spans": [ + { + "bbox": [ + 131, + 441, + 322, + 453 + ], + "type": "text", + "content": "3.3 Calibration of semantic representation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 461, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 533 + ], + "type": "text", + "content": "The integration of the scene-level semantic representation " + }, + { + "bbox": [ + 130, + 461, + 482, + 533 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 461, + 482, + 533 + ], + "type": "text", + "content": ", generated through simple averaging of global feature vectors as in (3), improves rendering quality. However, this approach has limitations when dealing with multiple views. As illustrated in Figure 3, viewing the same object from distinct angles may retain spatial attributes but can lead to conflicting semantic meanings. Merely averaging these global feature vectors without accounting for camera positions can result in a distorted scene-level understanding." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 534, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 629 + ], + "type": "text", + "content": "To mitigate this inconsistency, we propose a semantic calibration technique using feature rotation. This adjustment aligns the semantic representation across different camera poses. Our inspiration comes from the use of camera pose projection in computing the fused pixel-level feature " + }, + { + "bbox": [ + 130, + 534, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\widetilde{F}" + }, + { + "bbox": [ + 130, + 534, + 482, + 629 + ], + "type": "text", + "content": " and is further motivated by [51], which demonstrates that explicit rotation operations in feature spaces are feasible. Unlike point clouds in [51] that inherently lack a defined canonical orientation, NeRF explicitly encodes differences between camera viewpoints, thereby enabling precise calibration between the reference and target images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 628, + 494, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 628, + 494, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 494, + 667 + ], + "type": "text", + "content": "Building on this observation, we calculate calibrated semantic representations " + }, + { + "bbox": [ + 130, + 628, + 494, + 667 + ], + "type": "inline_equation", + "content": "\\{\\widetilde{S}_n\\}_{n=1}^N" + }, + { + "bbox": [ + 130, + 628, + 494, + 667 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 130, + 628, + 494, + 667 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 628, + 494, + 667 + ], + "type": "text", + "content": " original semantic representations " + }, + { + "bbox": [ + 130, + 628, + 494, + 667 + ], + "type": "inline_equation", + "content": "\\{S_n\\}_{n=1}^N" + }, + { + "bbox": [ + 130, + 628, + 494, + 667 + ], + "type": "text", + "content": " derived from the reference views. We accomplish this by leveraging their respective rotation matrices " + }, + { + "bbox": [ + 130, + 628, + 494, + 667 + ], + "type": "inline_equation", + "content": "\\{T_n\\}_{n=1}^N" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "text", + "content": "H. Zhu and T. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 249, + 201 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 249, + 201 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 249, + 201 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 249, + 201 + ], + "type": "image", + "image_path": "8c28005788ec4022f614351b3ccdc65476c63ba2e63fa8cd2167983135731c61.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 210, + 480, + 233 + ], + "lines": [ + { + "bbox": [ + 131, + 210, + 480, + 233 + ], + "spans": [ + { + "bbox": [ + 131, + 210, + 480, + 233 + ], + "type": "text", + "content": "Fig. 4: Visualization of decoded feature maps for \"orchid\" in LLFF dataset, produced by ray transformers [60] at different stages. From left to right, the transformer stages increase in depth." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 249, + 114, + 365, + 201 + ], + "blocks": [ + { + "bbox": [ + 249, + 114, + 365, + 201 + ], + "lines": [ + { + "bbox": [ + 249, + 114, + 365, + 201 + ], + "spans": [ + { + "bbox": [ + 249, + 114, + 365, + 201 + ], + "type": "image", + "image_path": "c551c24a3094aff2d59889d370cfc1e79cdc141e0a8aba3afef905b181d96f99.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 365, + 115, + 480, + 201 + ], + "blocks": [ + { + "bbox": [ + 365, + 115, + 480, + 201 + ], + "lines": [ + { + "bbox": [ + 365, + 115, + 480, + 201 + ], + "spans": [ + { + "bbox": [ + 365, + 115, + 480, + 201 + ], + "type": "image", + "image_path": "bccc6953b678e6b0d380c9ca486b8b45a81d7086611ef6f84020c1be1b1903a6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 255, + 479, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 479, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 479, + 278 + ], + "type": "text", + "content": "to model the rotational variations between each input view and the target view. The alignment of the original semantic features is performed as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 198, + 284, + 481, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 284, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 198, + 284, + 481, + 300 + ], + "type": "interline_equation", + "content": "\\widetilde {\\boldsymbol {S}} _ {n} = \\boldsymbol {\\mathcal {P}} \\left(\\boldsymbol {T} _ {n} \\cdot \\boldsymbol {\\mathcal {P}} ^ {- 1} \\left(\\boldsymbol {S} _ {n}\\right)\\right), \\text {w h e r e} \\boldsymbol {T} _ {n} = \\boldsymbol {T} _ {\\text {o u t}} ^ {\\mathrm {w} 2 \\mathrm {c}} \\cdot \\boldsymbol {T} _ {n} ^ {\\mathrm {c} 2 \\mathrm {w}}. \\tag {5}", + "image_path": "e1ae10e7e960a86786574fcfaa8e8b48994511598fa97de1b181941957c035a5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 306, + 479, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 479, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 479, + 354 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 306, + 479, + 354 + ], + "type": "inline_equation", + "content": "T_{n}^{\\mathrm{c2w}}" + }, + { + "bbox": [ + 130, + 306, + 479, + 354 + ], + "type": "text", + "content": " is the inverse of the extrinsic matrix used for " + }, + { + "bbox": [ + 130, + 306, + 479, + 354 + ], + "type": "inline_equation", + "content": "I_{n}" + }, + { + "bbox": [ + 130, + 306, + 479, + 354 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 306, + 479, + 354 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{out}}^{\\mathrm{w2c}}" + }, + { + "bbox": [ + 130, + 306, + 479, + 354 + ], + "type": "text", + "content": " is the extrinsic matrix for the target view. " + }, + { + "bbox": [ + 130, + 306, + 479, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\cdot)" + }, + { + "bbox": [ + 130, + 306, + 479, + 354 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 306, + 479, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{-1}(\\cdot)" + }, + { + "bbox": [ + 130, + 306, + 479, + 354 + ], + "type": "text", + "content": " are the flattening and inverse flattening operations, which reshape the feature to a 1D vector of shape 1-by- " + }, + { + "bbox": [ + 130, + 306, + 479, + 354 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 306, + 479, + 354 + ], + "type": "text", + "content": " and a 2D matrix of shape 3-by- " + }, + { + "bbox": [ + 130, + 306, + 479, + 354 + ], + "type": "inline_equation", + "content": "\\frac{C}{3}" + }, + { + "bbox": [ + 130, + 306, + 479, + 354 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 354, + 479, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 479, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 479, + 413 + ], + "type": "text", + "content": "Note that for the extrinsic matrix, we consider only the top-left " + }, + { + "bbox": [ + 130, + 354, + 479, + 413 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 354, + 479, + 413 + ], + "type": "text", + "content": " submatrix that accounts for rotation. Using GAP to condense feature maps of various sizes into a 1-by-" + }, + { + "bbox": [ + 130, + 354, + 479, + 413 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 354, + 479, + 413 + ], + "type": "text", + "content": " feature vector eliminates the need for scaling parameters in the semantic representation. As a result, modeling the intrinsic matrix is unnecessary, assuming no skewing, making our approach adaptable to different camera configurations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 414, + 479, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 479, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 479, + 439 + ], + "type": "text", + "content": "With the calibrated semantic features " + }, + { + "bbox": [ + 130, + 414, + 479, + 439 + ], + "type": "inline_equation", + "content": "\\{\\widetilde{S}_n\\}_{n=1}^N" + }, + { + "bbox": [ + 130, + 414, + 479, + 439 + ], + "type": "text", + "content": " for each reference view, we average these, similar to (3), to obtain the calibrated scene-level semantic representation " + }, + { + "bbox": [ + 130, + 414, + 479, + 439 + ], + "type": "inline_equation", + "content": "\\widetilde{S}" + }, + { + "bbox": [ + 130, + 414, + 479, + 439 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 258, + 448, + 481, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 448, + 481, + 480 + ], + "spans": [ + { + "bbox": [ + 258, + 448, + 481, + 480 + ], + "type": "interline_equation", + "content": "\\widetilde {\\boldsymbol {S}} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\widetilde {\\boldsymbol {S}} _ {n} \\in \\mathbb {R} ^ {C}. \\tag {6}", + "image_path": "f44fe169382980c9b85963a10c0f6ceb97e186aaab262e0462a968797c4a2824.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 485, + 479, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 479, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 479, + 510 + ], + "type": "text", + "content": "Finally, akin to (4), we concatenate the pixel-level fused feature " + }, + { + "bbox": [ + 130, + 485, + 479, + 510 + ], + "type": "inline_equation", + "content": "\\widetilde{F}" + }, + { + "bbox": [ + 130, + 485, + 479, + 510 + ], + "type": "text", + "content": " with the calibrated scene-level semantic representation " + }, + { + "bbox": [ + 130, + 485, + 479, + 510 + ], + "type": "inline_equation", + "content": "\\widetilde{S}" + }, + { + "bbox": [ + 130, + 485, + 479, + 510 + ], + "type": "text", + "content": " to form the final global-local embedding " + }, + { + "bbox": [ + 130, + 485, + 479, + 510 + ], + "type": "inline_equation", + "content": "\\widetilde{E}" + }, + { + "bbox": [ + 130, + 485, + 479, + 510 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 246, + 514, + 481, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 514, + 481, + 529 + ], + "spans": [ + { + "bbox": [ + 246, + 514, + 481, + 529 + ], + "type": "interline_equation", + "content": "\\widetilde {\\boldsymbol {E}} = \\operatorname {C o n c a t} (\\widetilde {\\boldsymbol {F}}, \\widetilde {\\boldsymbol {S}}) \\in \\mathbb {R} ^ {L + C}. \\tag {7}", + "image_path": "d2efbcf8adbff27a0e7b0449fb0f02500e0f81297614f79e43d5e046f5d697f6.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 536, + 479, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 479, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 479, + 608 + ], + "type": "text", + "content": "This unified embedding then feeds into ray transformers, passing through standard self-attention mechanisms. In the original GNT [60], multiple view transformers and ray transformers are stacked alternately for sequential feature processing. The last ray transformer integrates features from multiple points along a ray to yield the final RGB value. We denote the corresponding feature representations at stage " + }, + { + "bbox": [ + 130, + 536, + 479, + 608 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 536, + 479, + 608 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 536, + 479, + 608 + ], + "type": "inline_equation", + "content": "\\widetilde{F}^{(k)}" + }, + { + "bbox": [ + 130, + 536, + 479, + 608 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 536, + 479, + 608 + ], + "type": "inline_equation", + "content": "\\widetilde{E}^{(k)}" + }, + { + "bbox": [ + 130, + 536, + 479, + 608 + ], + "type": "text", + "content": ". Notably, the calibrated semantic representation " + }, + { + "bbox": [ + 130, + 536, + 479, + 608 + ], + "type": "inline_equation", + "content": "\\widetilde{S}" + }, + { + "bbox": [ + 130, + 536, + 479, + 608 + ], + "type": "text", + "content": " remains constant across these stages." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 624, + 250, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 250, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 250, + 635 + ], + "type": "text", + "content": "3.4 Sequential refinement" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 639, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 639, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 639, + 479, + 665 + ], + "type": "text", + "content": "While leveraging " + }, + { + "bbox": [ + 130, + 639, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\widetilde{S}" + }, + { + "bbox": [ + 130, + 639, + 479, + 665 + ], + "type": "text", + "content": " improves consistency, a single, uniform " + }, + { + "bbox": [ + 130, + 639, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\widetilde{S}" + }, + { + "bbox": [ + 130, + 639, + 479, + 665 + ], + "type": "text", + "content": " may not be adequate for deeper layers that demand more nuanced details. In fact, we find that deeper transform-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 448, + 102 + ], + "type": "text", + "content": "Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "ers capture finer details compared to shallower ones, as shown in Figure 4. To address this limitation, we introduce a sequential semantic feature refinement module that progressively enriches features at each stage. Specifically, we learn the residual " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\pmb{\\Delta}^{(k)}" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " to update " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{S}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " at each stage " + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 257, + 170, + 481, + 185 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 170, + 481, + 185 + ], + "spans": [ + { + "bbox": [ + 257, + 170, + 481, + 185 + ], + "type": "interline_equation", + "content": "\\widetilde {\\boldsymbol {S}} ^ {(k + 1)} \\leftarrow \\widetilde {\\boldsymbol {S}} ^ {(k)} + \\boldsymbol {\\Delta} ^ {(k)}. \\tag {8}", + "image_path": "75a3a8edaf82c10913798477b58e25d6657cc464eba4cefab6a27862c6b18fa7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 479, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 479, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 479, + 266 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 194, + 479, + 266 + ], + "type": "inline_equation", + "content": "\\pmb{\\Delta}^{(k)}" + }, + { + "bbox": [ + 130, + 194, + 479, + 266 + ], + "type": "text", + "content": " is calculated by first performing specialized cross-attentions between " + }, + { + "bbox": [ + 130, + 194, + 479, + 266 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{S}}^{(k)}" + }, + { + "bbox": [ + 130, + 194, + 479, + 266 + ], + "type": "text", + "content": " and the original, uncalibrated per-frame semantic features " + }, + { + "bbox": [ + 130, + 194, + 479, + 266 + ], + "type": "inline_equation", + "content": "\\{S_n\\}_{n=1}^N" + }, + { + "bbox": [ + 130, + 194, + 479, + 266 + ], + "type": "text", + "content": " (see Figure 2), followed by their summation. Our goal is to fuse information from different source views to enrich the scene-level semantic representation with features from each reference frame. With this sequential refinement, we combine " + }, + { + "bbox": [ + 130, + 194, + 479, + 266 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{S}}^{(k)}" + }, + { + "bbox": [ + 130, + 194, + 479, + 266 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 194, + 479, + 266 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{F}}^{(k)}" + }, + { + "bbox": [ + 130, + 194, + 479, + 266 + ], + "type": "text", + "content": " at each stage, yielding a stage-specific global-local embedding " + }, + { + "bbox": [ + 130, + 194, + 479, + 266 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{E}}^{(k)}" + }, + { + "bbox": [ + 130, + 194, + 479, + 266 + ], + "type": "text", + "content": ", which completes our approach." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 267, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 482, + 399 + ], + "type": "text", + "content": "Discussion. In scenarios with few reference views, especially when limited to just one, the primary issue is inaccurate depth estimation, resulting in depth ambiguity [12]. This compromises the quality of images when rendered from novel viewpoints. Despite this, essential visual information generally remains accurate across different camera poses. Incorporating our proposed scene-level representation improves the understanding of the overall scene layout [7], distinguishing our approach from existing generalizable NeRF models that predict pixels individually. The advantage of our approach is its holistic view; the semantic representation enriches per-pixel predictions by providing broader context. This semantic constraint ensures that fewer abrupt changes between adjacent points. Consequently, it leads to more reliable depth estimations, making the images rendered from limited reference views more plausible." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 415, + 240, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 415, + 240, + 427 + ], + "spans": [ + { + "bbox": [ + 131, + 415, + 240, + 427 + ], + "type": "text", + "content": "3.5 Training objectives" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 434, + 365, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 434, + 365, + 445 + ], + "spans": [ + { + "bbox": [ + 131, + 434, + 365, + 445 + ], + "type": "text", + "content": "During training, we employ three different loss functions:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 446, + 481, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 481, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 481, + 470 + ], + "type": "text", + "content": "MSE loss. The Mean Square Error (MSE) loss is the standard photometric loss used in NeRF [43]. It computes the MSE between the actual and predicted pixel values." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 471, + 481, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 481, + 518 + ], + "type": "text", + "content": "Central loss. Since we project the view-level features from different input camera poses to the shared target view, we introduce a central loss to ensure frame-wise calibrated semantic features " + }, + { + "bbox": [ + 130, + 471, + 481, + 518 + ], + "type": "inline_equation", + "content": "\\{\\widetilde{S}_n\\}_{n=1}^N" + }, + { + "bbox": [ + 130, + 471, + 481, + 518 + ], + "type": "text", + "content": " are consistent when projected onto the same target view, which is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 244, + 525, + 481, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 525, + 481, + 558 + ], + "spans": [ + { + "bbox": [ + 244, + 525, + 481, + 558 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c e n t r a l}} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\left\\| \\widetilde {\\boldsymbol {S}} _ {n} - \\widetilde {\\boldsymbol {S}} \\right\\| _ {1}. \\tag {9}", + "image_path": "0bf6f1baa18c7f0c9b87092b2ecea9e887a6c3dc4aa9dc3e4762bd9315854b50.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 564, + 481, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 564, + 481, + 612 + ], + "spans": [ + { + "bbox": [ + 131, + 564, + 481, + 612 + ], + "type": "text", + "content": "Point-wise perceptual loss. During the rendering of a batch of pixels in a target view, we inpaint the ground-truth image by replacing the corresponding pixels with the predicted ones. Then, a perceptual loss [26] is computed between the inpainted image and the target image to guide the training process at the whole-image level." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 146, + 613, + 342, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 613, + 342, + 623 + ], + "spans": [ + { + "bbox": [ + 146, + 613, + 342, + 623 + ], + "type": "text", + "content": "The final loss function is formulated as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 237, + 632, + 481, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 632, + 481, + 646 + ], + "spans": [ + { + "bbox": [ + 237, + 632, + 481, + 646 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {M S E}} + \\lambda_ {1} \\mathcal {L} _ {\\text {c e n t r a l}} + \\lambda_ {2} \\mathcal {L} _ {\\text {p e r c}}. \\tag {10}", + "image_path": "482edd69d1b8bca96e58402d7925cd19b135062924ee978d37e76719d5ca56dc.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 653, + 373, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 653, + 373, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 653, + 373, + 666 + ], + "type": "text", + "content": "Empirically, we set " + }, + { + "bbox": [ + 131, + 653, + 373, + 666 + ], + "type": "inline_equation", + "content": "\\lambda_1 = 1" + }, + { + "bbox": [ + 131, + 653, + 373, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 653, + 373, + 666 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 0.001" + }, + { + "bbox": [ + 131, + 653, + 373, + 666 + ], + "type": "text", + "content": ", following [32]." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "text", + "content": "H. Zhu and T. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 157, + 481, + 256 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Results for generalizable scene rendering on LLFF with few reference views. GeoNeRF, MatchNeRF, and MVSNeRF necessitate variance as input, defaulting to 0 for single-image cases, hence their results are not included for 1-view scenarios." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 157, + 481, + 256 + ], + "lines": [ + { + "bbox": [ + 133, + 157, + 481, + 256 + ], + "spans": [ + { + "bbox": [ + 133, + 157, + 481, + 256 + ], + "type": "table", + "html": "
Method1 reference view2 reference views3 reference views
PSNR (↑)LPIPS (↓)SSIM (↑)PSNR (↑)LPIPS (↓)SSIM (↑)PSNR (↑)LPIPS (↓)SSIM (↑)
PixelNeRF [76]9.320.8980.26411.230.7660.28211.240.6710.486
GPNR [58]15.910.5270.40018.790.3800.57521.570.2880.695
NeuRay [38]16.180.5840.39317.710.3360.64618.260.3100.672
GeoNeRF [25]---18.760.4730.50023.400.2460.766
MatchNeRF [8]---21.080.2720.68922.300.2340.731
MVSNeRF [6]---19.150.3360.70419.840.3140.729
IBRNet [64]16.850.5420.50721.250.3330.68523.000.2620.752
GNT [60]16.570.5000.42420.880.2510.69123.210.1780.782
Ours18.310.4350.52121.940.2240.73623.450.1760.794
", + "image_path": "5c5dcf8c6dfd7bacea48c9c40cee78701af1adee7faa63022c4f5ceb314571ad.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 277, + 218, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 277, + 218, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 277, + 218, + 293 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 307, + 481, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 307, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 131, + 307, + 481, + 343 + ], + "type": "text", + "content": "In this section, we first discuss our experimental setups and other details required for our experiments in Section 4.1, followed by our results for different experimental settings, along with ablation studies and analysis in Section 4.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 365, + 246, + 377 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 246, + 377 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 246, + 377 + ], + "type": "text", + "content": "4.1 Experimental setups" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 388, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 388, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 131, + 388, + 482, + 508 + ], + "type": "text", + "content": "Datasets. Firstly, following [60], we construct our training data from both synthetic and real data. This collection includes scanned models from Google Scanned Objects [14], RealEstate10K [81], and handheld phone captures [64]. For evaluation, we utilize real data encompassing complex scenes from sources such as LLFF [42], Shiny [67], and mip-NeRF 360 [4]. Additionally, we train and test our model using the recently released MVImgNet dataset [77]. We adhere to the official split, focusing on examples from the containers category, and select 2,500 scenes for training. During inference, we choose 100 scenes, using their first images as target views and the spatially nearest images as references. Since MVImgNet does not provide camera poses, we utilize COLMAP [54, 55] to deduce the camera positions within these scenes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 509, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 509, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 131, + 509, + 482, + 616 + ], + "type": "text", + "content": "Implementation details. CaesarNeRF is built upon GNT [60], for which we maintain the same configuration, setting the ray and view transformers stack number " + }, + { + "bbox": [ + 131, + 509, + 482, + 616 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 131, + 509, + 482, + 616 + ], + "type": "text", + "content": " to 8 for generalizable setting and 4 for single-scene setting. The feature encoder extracts bottleneck features, applies GAP, and then uses a fully connected (FC) layer to reduce the input dimension " + }, + { + "bbox": [ + 131, + 509, + 482, + 616 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 131, + 509, + 482, + 616 + ], + "type": "text", + "content": " to 96. Training involves 500,000 iterations using the Adam optimizer [28], with learning rate set at 0.001 for the feature encoder and 0.0005 for CaesarNeRF, halving them every 100,000 iterations. Each iteration samples 4,096 rays from a single scene. In line with [60], we randomly choose between 8 to 10 reference views for training, and 3 to 7 views when using the MVImgNet [77]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "content": "Baseline methods. We compare CaesarNeRF with several state-of-the-art methods suited for generalizable NeRF applications, including earlier works such as MVS-NeRF [6], PixelNeRF [76], and IBRNet [64], alongside more recent ones, including GPNR [58], NeuRay [38], GNT [60], GeoNeRF [25] and MatchNeRF [8]." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 135, + 480, + 203 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 462, + 126 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 462, + 126 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 462, + 126 + ], + "type": "text", + "content": "Table 2: Results for generalizable scene rendering on Shiny with few reference views." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 135, + 480, + 203 + ], + "lines": [ + { + "bbox": [ + 133, + 135, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 135, + 480, + 203 + ], + "type": "table", + "html": "
Method1 reference view2 reference views3 reference views
PSNR (↑)LPIPS (↓)SSIM (↑)PSNR (↑)LPIPS (↓)SSIM (↑)PSNR (↑)LPIPS (↓)SSIM (↑)
MatchNeRF [8]---20.280.2780.63620.770.2490.672
MVSNeRF [6]---17.250.4160.57718.550.3430.645
IBRNet [64]14.930.6250.40118.400.4000.59521.960.2810.710
GNT [60]15.990.5480.40020.420.3270.61722.470.2470.720
Ours17.570.4670.47221.470.2930.65222.740.2410.723
", + "image_path": "687323f487c0adf1d018ae1072e1869a4fcd2e883061f3bc5ada3e103545f12c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 232, + 480, + 300 + ], + "blocks": [ + { + "bbox": [ + 133, + 212, + 479, + 224 + ], + "lines": [ + { + "bbox": [ + 133, + 212, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 212, + 479, + 224 + ], + "type": "text", + "content": "Table 3: Results for generalizable scene rendering on mip-NeRF 360 with few reference views." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 232, + 480, + 300 + ], + "lines": [ + { + "bbox": [ + 133, + 232, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 133, + 232, + 480, + 300 + ], + "type": "table", + "html": "
Method1 reference view2 reference views3 reference views
PSNR (↑)LPIPS (↓)SSIM (↑)PSNR (↑)LPIPS (↓)SSIM (↑)PSNR (↑)LPIPS (↓)SSIM (↑)
MatchNeRF [8]---17.000.5660.39217.260.5510.407
MVSNeRF [6]---14.230.6810.36614.290.6740.406
IBRNet [64]14.120.6820.28316.240.6180.36017.700.5550.420
GNT [60]13.480.6300.31415.210.5590.37015.590.5380.395
Ours15.200.5920.35017.050.5380.40317.550.5120.430
", + "image_path": "f85e95249daeadc8b36fea3de9e9c1c727a77061a3471c1dcc94f2c62bdb64ed.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 321, + 244, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 244, + 333 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 244, + 333 + ], + "type": "text", + "content": "4.2 Results and analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 340, + 481, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 481, + 399 + ], + "type": "text", + "content": "We compare results in two settings: a generalizable setting, where the model is trained on multiple scenes without fine-tuning during inference for both few and all reference view cases, and a single-scene setting where the model is trained and evaluated on just one scene. Following these comparisons, we conduct ablation studies and test the generalizability of our method with other state-of-the-art approaches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 400, + 481, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 481, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 481, + 448 + ], + "type": "text", + "content": "Generalizable rendering. In the generalizable setting, we adopt two training strategies. First, we train the model on multiple datasets as described in Section 4.1 and evaluate on LLFF [42], Shiny [67] and mip-NeRF 360 [4] datasets. In addition, the model is trained and tested on the MVImgNet [77] for object-centric generalizability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 448, + 481, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 481, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 481, + 543 + ], + "type": "text", + "content": "(a) LLFF, Shiny, and mip-NeRF 360. The results for few-reference view scenarios on these datasets are shown in Tables 1, 2 and 3, respectively. Methods like MatchNeRF [8], MVSNeRF [75], and GeoNeRF [25] require at least two reference views. On the LLFF dataset, all methods experience a performance decline as the number of views decreases. CaesarNeRF, however, consistently outperforms others across varying reference view numbers, with the performance gap becoming more significant with fewer views. For example, with 3 views, while IBRNet [64] and GNT [60] have comparable PSNRs, CaesarNeRF demonstrates a more substantial lead in LPIPS and SSIM metrics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 543, + 481, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 543, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 543, + 481, + 567 + ], + "type": "text", + "content": "Similar patterns are observed on the Shiny [67] and mip-NeRF 360 [4] datasets. We apply the highest-performing methods from the LLFF evaluations and report the results" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 133, + 619, + 480, + 664 + ], + "blocks": [ + { + "bbox": [ + 130, + 587, + 481, + 610 + ], + "lines": [ + { + "bbox": [ + 130, + 587, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 587, + 481, + 610 + ], + "type": "text", + "content": "Table 4: Results on MVImgNet across varying numbers of reference views. 'C.' represents the use of calibration before averaging." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 619, + 480, + 664 + ], + "lines": [ + { + "bbox": [ + 133, + 619, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 133, + 619, + 480, + 664 + ], + "type": "table", + "html": "
Method1 reference view2 reference views3 reference views4 reference views5 reference views
PSNR (†)LPIPS (↓)SSIM (†)PSNR (†)LPIPS (↓)SSIM (†)PSNR (†)LPIPS (↓)SSIM (†)PSNR (†)LPIPS (↓)SSIM (†)PSNR (†)LPIPS (↓)SSIM (†)
IBRNet19.140.4580.59524.380.2660.81825.530.2030.85825.990.1900.86726.120.1880.867
GNT22.220.4330.67826.940.2360.85027.410.2060.87027.510.1970.87527.510.1940.876
Ours w/oC.23.610.3710.71826.340.2740.81727.100.2280.85027.300.2100.86227.340.2030.865
Ours24.280.3340.74727.340.2150.85627.820.1900.87527.920.1810.88127.920.1790.882
", + "image_path": "12b60e7ee87f5e7b5db63d6046800fa24fdc977238f4990bbd69f3f253055863.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 257, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 257, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 257, + 102 + ], + "type": "text", + "content": "H. Zhu and T. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 160, + 135, + 454, + 182 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 125 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 125 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 125 + ], + "type": "text", + "content": "Table 5: Results of per-scene optimization on LLFF, in comparison with state-of-the-art methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 160, + 135, + 454, + 182 + ], + "lines": [ + { + "bbox": [ + 160, + 135, + 454, + 182 + ], + "spans": [ + { + "bbox": [ + 160, + 135, + 454, + 182 + ], + "type": "table", + "html": "
MethodLLFF [42]NeRF [44]NeX [67]GNT [60]Ours
PSNR (↑)23.2726.5027.2627.2427.64
LPIPS (↓)0.2120.2500.1790.0870.081
SSIM (↑)0.7980.8110.9040.8890.904
", + "image_path": "696656a1160f30f0004505714c3681d5d520d23c6c930899fbec31aadfca793b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 142, + 213, + 471, + 275 + ], + "blocks": [ + { + "bbox": [ + 136, + 193, + 476, + 205 + ], + "lines": [ + { + "bbox": [ + 136, + 193, + 476, + 205 + ], + "spans": [ + { + "bbox": [ + 136, + 193, + 476, + 205 + ], + "type": "text", + "content": "Table 6: Results on LLFF for few-shot generalization after adapting Caesar to other baselines." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 142, + 213, + 471, + 275 + ], + "lines": [ + { + "bbox": [ + 142, + 213, + 471, + 275 + ], + "spans": [ + { + "bbox": [ + 142, + 213, + 471, + 275 + ], + "type": "table", + "html": "
Method1 reference view2 reference views3 reference views
PSNR (↑)SSIM (↑)LPIPS (↓)PSNR (↑)SSIM (↑)LPIPS (↓)PSNR (↑)SSIM (↑)LPIPS (↓)
MatchNeRF [8]---20.590.7750.27622.430.8050.244
+ Caesar---21.550.7820.26822.980.8240.242
IBRNet [64]16.850.5070.54221.250.6850.33323.000.7520.262
+ Caesar17.760.5430.50022.390.7400.27523.670.7720.242
", + "image_path": "1f5e288988be0eef1d7146547258ff5c32c3c003b3fb5d2e366708854c818db9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 297, + 481, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 481, + 357 + ], + "type": "text", + "content": "for those that produce satisfactory outcomes with few reference views. CaesarNeRF maintains superior performance throughout. Notably, for complex datasets like mipNeRF 360 [4], which have sparse camera inputs, the quality of rendered images generally decreases with fewer available reference views. Nonetheless, CaesarNeRF shows the most robust performance compared to the other methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 357, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 453 + ], + "type": "text", + "content": "(b) MVImgNet. We extend our comparison of CaesarNeRF with GNT [60] and IBRNet [64] on the MVImgNet dataset, focusing on object-centric scenes, as shown in Table 4. We examine a variant of CaesarNeRF where semantic calibration is substituted with simple feature averaging from multiple frames. While the performance of all methods improves with more views, CaesarNeRF consistently outperforms GNT and IBRNet. Notably, CaesarNeRF with feature averaging surpasses GNT in 1-view case but lags with additional views, implying that the absence of calibration leads to ambiguities when rendering from multiple views." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 453, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 489 + ], + "type": "text", + "content": "Per-scene optimization. Beyond the multi-scene generalizable setting, we demonstrate per-scene optimization results in Table 5. We calculate the average performance over 8 categories from the LLFF dataset [42]. CaesarNeRF consistently outperforms" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 186, + 542, + 427, + 663 + ], + "blocks": [ + { + "bbox": [ + 130, + 510, + 482, + 533 + ], + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 533 + ], + "type": "text", + "content": "Table 7: Ablations on the semantic representation length " + }, + { + "bbox": [ + 130, + 510, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 130, + 510, + 482, + 533 + ], + "type": "text", + "content": ", sequential refinement (Seq.) and calibration (Cali.). 'Ext.' denotes the extension of per-pixel feature to a length of 64 in GNT." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 186, + 542, + 427, + 663 + ], + "lines": [ + { + "bbox": [ + 186, + 542, + 427, + 663 + ], + "spans": [ + { + "bbox": [ + 186, + 542, + 427, + 663 + ], + "type": "table", + "html": "
Model VariationsPSNR (↑)LPIPS (↓)SSIM (↑)
Rlen.Seq.Cali.
(Baseline GNT)20.930.1850.731
Ext.20.850.1730.735
+3221.430.1520.763
+6421.490.1490.766
+9621.460.1500.766
+12821.490.1470.763
+96✓21.530.1460.770
+96✓21.510.1470.769
+96✓✓21.670.1390.781
", + "image_path": "b57f162c0320554496d88143968ead42d70864083a644364d70ab2e1e85c13c7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 448, + 102 + ], + "type": "text", + "content": "Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 118, + 481, + 350 + ], + "blocks": [ + { + "bbox": [ + 133, + 118, + 481, + 350 + ], + "lines": [ + { + "bbox": [ + 133, + 118, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 133, + 118, + 481, + 350 + ], + "type": "image", + "image_path": "858b3b8516b308fe17dc8e4c054e24dd6a64817d339484ac52acb726d2c85097.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 352, + 380, + 361 + ], + "lines": [ + { + "bbox": [ + 234, + 352, + 380, + 361 + ], + "spans": [ + { + "bbox": [ + 234, + 352, + 380, + 361 + ], + "type": "text", + "content": "(b) Using two images as reference view." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 371, + 480, + 383 + ], + "lines": [ + { + "bbox": [ + 132, + 371, + 480, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 371, + 480, + 383 + ], + "type": "text", + "content": "Fig. 5: Comparative visualization of our proposed method against other state-of-the-art methods." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 422, + 480, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 480, + 445 + ], + "type": "text", + "content": "nearly all state-of-the-art methods in the comparison, across all three metrics, showing a significant improvement over our baseline method, GNT [60]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "text", + "content": "Adaptability. To test the adaptability of our Caesar pipeline, we apply it to two other state-of-the-art methods that use view transformers, namely MatchNeRF [8] and IBRNet [64]. We demonstrate in Table 6 that our enhancements in scene-level semantic understanding significantly boost the performance of these methods across all metrics. This indicates that the Caesar framework is not only beneficial in our CaesarNeRF, which is based on GNT [60], but can also be a versatile addition to other NeRF pipelines with view transformers to aggregate different input views." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "Ablation analysis. We conduct ablation studies on the \"orchid\" scene from the LLFF dataset, with findings detailed in Table 7. Testing variations in representation and the impact of the sequential refinement and calibration modules, we find that increasing the latent size in GNT yields marginal benefits. However, incorporating even a modest semantic representation size distinctly improves results. The length of the semantic representation has a minimal impact on quality. Our ablation studies indicate that while sequential refinement and calibration each offer slight performance gains, their combined effect is most significant. In a single-scene context, semantic information is effectively embedded within the representation, making the benefits of individual modules subtler. Together, however, they provide a framework where sequential refinement can leverage calibrated features for deeper insights." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "text", + "content": "H. Zhu and T. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 211 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 211 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 211 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 211 + ], + "type": "image", + "image_path": "5e8f225582df2f2400796b948c79fb2f54544d41e8e0fe57e7fbb97ba084cb56.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 219, + 480, + 240 + ], + "lines": [ + { + "bbox": [ + 132, + 219, + 480, + 240 + ], + "spans": [ + { + "bbox": [ + 132, + 219, + 480, + 240 + ], + "type": "text", + "content": "Fig. 6: Depth estimation prediction using one reference view (first row) and two reference views (second row) as input from LLFF comparing CaesarNeRF with GNT." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 140, + 255, + 197, + 307 + ], + "blocks": [ + { + "bbox": [ + 140, + 255, + 197, + 307 + ], + "lines": [ + { + "bbox": [ + 140, + 255, + 197, + 307 + ], + "spans": [ + { + "bbox": [ + 140, + 255, + 197, + 307 + ], + "type": "image", + "image_path": "935270915a1cd7e5f145edcda7312b0c103867507a8548dcf4724c575cce0874.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 140, + 309, + 198, + 361 + ], + "blocks": [ + { + "bbox": [ + 140, + 309, + 198, + 361 + ], + "lines": [ + { + "bbox": [ + 140, + 309, + 198, + 361 + ], + "spans": [ + { + "bbox": [ + 140, + 309, + 198, + 361 + ], + "type": "image", + "image_path": "04c266c0e809b999741573b06e9286bba04568365e4aa0b2d0af948652131fea.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 363, + 175, + 370 + ], + "lines": [ + { + "bbox": [ + 162, + 363, + 175, + 370 + ], + "spans": [ + { + "bbox": [ + 162, + 363, + 175, + 370 + ], + "type": "text", + "content": "fern" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 205, + 255, + 262, + 307 + ], + "blocks": [ + { + "bbox": [ + 205, + 255, + 262, + 307 + ], + "lines": [ + { + "bbox": [ + 205, + 255, + 262, + 307 + ], + "spans": [ + { + "bbox": [ + 205, + 255, + 262, + 307 + ], + "type": "image", + "image_path": "c7f95fcadbe2b730ae942a8335f8543ceb5de0247b4698070d2882aeef12411e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 207, + 309, + 260, + 316 + ], + "lines": [ + { + "bbox": [ + 207, + 309, + 260, + 316 + ], + "spans": [ + { + "bbox": [ + 207, + 309, + 260, + 316 + ], + "type": "text", + "content": "chesstable (0.45)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 205, + 318, + 262, + 361 + ], + "blocks": [ + { + "bbox": [ + 205, + 318, + 262, + 361 + ], + "lines": [ + { + "bbox": [ + 205, + 318, + 262, + 361 + ], + "spans": [ + { + "bbox": [ + 205, + 318, + 262, + 361 + ], + "type": "image", + "image_path": "d7cbc2bf633ea45fcd2dab003a8f042a898b44f8595dd902e61b08a1c02ea3d4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 361, + 261, + 371 + ], + "lines": [ + { + "bbox": [ + 206, + 361, + 261, + 371 + ], + "spans": [ + { + "bbox": [ + 206, + 361, + 261, + 371 + ], + "type": "text", + "content": "playground (0.46)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 271, + 255, + 339, + 307 + ], + "blocks": [ + { + "bbox": [ + 271, + 255, + 339, + 307 + ], + "lines": [ + { + "bbox": [ + 271, + 255, + 339, + 307 + ], + "spans": [ + { + "bbox": [ + 271, + 255, + 339, + 307 + ], + "type": "image", + "image_path": "8b6c91324e0aa73e803c36ed7535388ae3d6e9cbb7637f03af1ff0901b88215a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 274, + 309, + 335, + 316 + ], + "lines": [ + { + "bbox": [ + 274, + 309, + 335, + 316 + ], + "spans": [ + { + "bbox": [ + 274, + 309, + 335, + 316 + ], + "type": "text", + "content": "colorfountain (0.50)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 276, + 318, + 333, + 361 + ], + "blocks": [ + { + "bbox": [ + 276, + 318, + 333, + 361 + ], + "lines": [ + { + "bbox": [ + 276, + 318, + 333, + 361 + ], + "spans": [ + { + "bbox": [ + 276, + 318, + 333, + 361 + ], + "type": "image", + "image_path": "a538e770a173f7d57c111ae225805b73394531243d0153e86fa675b6362c1a7e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 286, + 361, + 323, + 371 + ], + "lines": [ + { + "bbox": [ + 286, + 361, + 323, + 371 + ], + "spans": [ + { + "bbox": [ + 286, + 361, + 323, + 371 + ], + "type": "text", + "content": "bikes (0.50)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 380, + 482, + 414 + ], + "lines": [ + { + "bbox": [ + 131, + 380, + 482, + 414 + ], + "spans": [ + { + "bbox": [ + 131, + 380, + 482, + 414 + ], + "type": "text", + "content": "Fig.7: Largest and smallest distances for two examples from LLFF test split when matching with training scenes. Numbers " + }, + { + "bbox": [ + 131, + 380, + 482, + 414 + ], + "type": "inline_equation", + "content": "(\\times 10^{-2})" + }, + { + "bbox": [ + 131, + 380, + 482, + 414 + ], + "type": "text", + "content": " denote the L-2 distance to the source image within the semantic feature space." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 349, + 255, + 404, + 307 + ], + "blocks": [ + { + "bbox": [ + 349, + 255, + 404, + 307 + ], + "lines": [ + { + "bbox": [ + 349, + 255, + 404, + 307 + ], + "spans": [ + { + "bbox": [ + 349, + 255, + 404, + 307 + ], + "type": "image", + "image_path": "4afed97bc09049d0a839f5b6c8cf217b7f6cf9bdda1cc4c03c6ee454d42eb1ba.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 309, + 405, + 316 + ], + "lines": [ + { + "bbox": [ + 348, + 309, + 405, + 316 + ], + "spans": [ + { + "bbox": [ + 348, + 309, + 405, + 316 + ], + "type": "text", + "content": "fenceflower (1.91)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 348, + 318, + 405, + 361 + ], + "blocks": [ + { + "bbox": [ + 348, + 318, + 405, + 361 + ], + "lines": [ + { + "bbox": [ + 348, + 318, + 405, + 361 + ], + "spans": [ + { + "bbox": [ + 348, + 318, + 405, + 361 + ], + "type": "image", + "image_path": "3e87864d5f85b0a1adb736dca13d4289acb30644d97beda0529ccbed111e0708.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 361, + 399, + 371 + ], + "lines": [ + { + "bbox": [ + 354, + 361, + 399, + 371 + ], + "spans": [ + { + "bbox": [ + 354, + 361, + 399, + 371 + ], + "type": "text", + "content": "peppers (2.02)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 413, + 255, + 478, + 307 + ], + "blocks": [ + { + "bbox": [ + 413, + 255, + 478, + 307 + ], + "lines": [ + { + "bbox": [ + 413, + 255, + 478, + 307 + ], + "spans": [ + { + "bbox": [ + 413, + 255, + 478, + 307 + ], + "type": "image", + "image_path": "9e7836e5d80788e8f2ee9da19d594ab64ce94f4184259f32ff306735b7b5656d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 417, + 317, + 474, + 361 + ], + "blocks": [ + { + "bbox": [ + 427, + 309, + 464, + 316 + ], + "lines": [ + { + "bbox": [ + 427, + 309, + 464, + 316 + ], + "spans": [ + { + "bbox": [ + 427, + 309, + 464, + 316 + ], + "type": "text", + "content": "pond (1.80)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 317, + 474, + 361 + ], + "lines": [ + { + "bbox": [ + 417, + 317, + 474, + 361 + ], + "spans": [ + { + "bbox": [ + 417, + 317, + 474, + 361 + ], + "type": "image", + "image_path": "c07c7dcd68b3781bef9cb67469986fab91e8a6c05910b020cb23a813ba757f3c.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 361, + 466, + 371 + ], + "lines": [ + { + "bbox": [ + 425, + 361, + 466, + 371 + ], + "spans": [ + { + "bbox": [ + 425, + 361, + 466, + 371 + ], + "type": "text", + "content": "apples (2.01)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 521 + ], + "type": "text", + "content": "Visualizations. We present our visualization results in Figure 5, where we compare our method with others using one or two views from the LLFF dataset. Additional visual comparisons are provided in the supplementary materials. These visualizations highlight that in scenarios with few views, our method significantly surpasses other generalizable NeRF models, particularly excelling when only a single view is available. CaesarNeRF demonstrates rendering with sharper boundaries and more distinct objects." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 130, + 521, + 483, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 483, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 483, + 641 + ], + "type": "text", + "content": "Depth estimation. We extend our evaluation to depth prediction within the LLFF [42] dataset, focusing on challenges presented by few reference views, such as scenarios with just one or two images. In a comparison between CaesarNeRF and GNT [60], we observe in Figure 6 that GNT struggles to accurately capture the relative positions of objects when reference images are sparse. For instance, with only a single view of a flower, CaesarNeRF precisely indicates the flower's proximity to the camera compared to the leaves in the background, a distinction that GNT fails to make. Furthermore, the depth estimations provided by CaesarNeRF are consistently more reliable. In the horn example involving two views, CaesarNeRF offers better boundary delineation, showing particular strength in handling reflective surfaces such as grass in the background." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Semantic analysis for " + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\tilde{S}" + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": ". To assess whether the calibrated semantic representation " + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\tilde{S}" + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": " truly captures semantic details of the scene, we analyze the highest and lowest" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 448, + 102 + ], + "type": "text", + "content": "Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 220, + 177 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 220, + 177 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 220, + 177 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 220, + 177 + ], + "type": "image", + "image_path": "c47997a4afe51bbe83335e1b3d5cb5d3584e4b767cedb3d24cc4346a96259a24.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 180, + 240, + 190 + ], + "lines": [ + { + "bbox": [ + 203, + 180, + 240, + 190 + ], + "spans": [ + { + "bbox": [ + 203, + 180, + 240, + 190 + ], + "type": "text", + "content": "(a) leaves" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 221, + 114, + 306, + 177 + ], + "blocks": [ + { + "bbox": [ + 221, + 114, + 306, + 177 + ], + "lines": [ + { + "bbox": [ + 221, + 114, + 306, + 177 + ], + "spans": [ + { + "bbox": [ + 221, + 114, + 306, + 177 + ], + "type": "image", + "image_path": "f25a670ca80735b3c0d74ca3048f82dc6e715a6996e9e3285d0333f93be315e4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 201, + 482, + 234 + ], + "lines": [ + { + "bbox": [ + 131, + 201, + 482, + 234 + ], + "spans": [ + { + "bbox": [ + 131, + 201, + 482, + 234 + ], + "type": "text", + "content": "Fig.8: Synthetic results for two examples from LLFF [42], “leaves” and “orchid”, using zero 1-to-3 [35] with one reference image as input and 2 degrees of vertical shift. The left image of each pair is the input, and the right one is the output of zero 1-to-3 [35]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 391, + 177 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 391, + 177 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 391, + 177 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 391, + 177 + ], + "type": "image", + "image_path": "acb39d00ad69cd230a42b3ce357f84f418a987ab62c3dd9c20e1cce0df09bdf0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 180, + 411, + 190 + ], + "lines": [ + { + "bbox": [ + 373, + 180, + 411, + 190 + ], + "spans": [ + { + "bbox": [ + 373, + 180, + 411, + 190 + ], + "type": "text", + "content": "(b) orchid" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 392, + 114, + 477, + 177 + ], + "blocks": [ + { + "bbox": [ + 392, + 114, + 477, + 177 + ], + "lines": [ + { + "bbox": [ + 392, + 114, + 477, + 177 + ], + "spans": [ + { + "bbox": [ + 392, + 114, + 477, + 177 + ], + "type": "image", + "image_path": "7d2b2c139693508b5ec4cc740792990e53171d1bd735753f62a0e66adbdba032.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 255, + 479, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 479, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 479, + 303 + ], + "type": "text", + "content": "response based on the L-2 distance between features from two different scenes, interpreting smaller distances as greater similarity. We consider the first image from each category in the LLFF dataset as a reference and extract the scene-level representation for the ten closest views to this reference image using CaesarNeRF." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 304, + 480, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 480, + 399 + ], + "type": "text", + "content": "In Figure 7, we present two examples, \"room\" and \"fortress\", from the LLFF dataset, representing the first image of these categories alongside images corresponding to the top-2 highest and lowest responses. This analysis reveals that the scene-level representation predominantly emphasizes structural information and objects of similar categories. For instance, with the source image \"room\", the highest responses correlate with images featuring table-like structures, indicating an object-centric focus. In contrast, the lowest responses include images of flowers or ponds found in open areas, which diverge significantly from the structural and object content in the source images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 399, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 481, + 495 + ], + "type": "text", + "content": "Comparison with generative methods. Single-view scenarios are often addressed by generative methods [35-37] that employ diffusion models [19, 53]. While these models can produce reasonable results for object-centric renderings, they struggle with scene-level renderings from novel viewpoints. We show two LLFF examples in Figure 8 using zero 1-to-3 [35], where the left image is the input, and the right one is the output. Images rendered with zero 1-to-3 [35] suffer from style difference. Unlike NeRF-based approaches that reconstruct images from observed pixels, generative models synthesize an entire image from its semantic representation but may not maintain the style." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 511, + 285, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 511, + 285, + 523 + ], + "spans": [ + { + "bbox": [ + 131, + 511, + 285, + 523 + ], + "type": "text", + "content": "5 Conclusion and limitation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 533, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 480, + 616 + ], + "type": "text", + "content": "In this paper, we introduce CaesarNeRF, a few-shot and generalizable NeRF pipeline that combines scene-level semantic with per-pixel feature representations, aiding in rendering from novel camera positions with limited reference views. We calibrate the semantic representations across different input views and employ a sequential refinement network to offer distinct semantic representations at various levels. Our method has been extensively evaluated on a broad range of datasets, exhibiting state-of-the-art performance in both generalizable and single-scene settings." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "Limitations and potential negative impact. Instead of a generative approach, CaesarNeRF relies on NeRF's scheme, using input images to render the target view. This approach restricts its ability to render parts of the scene that are not present in the reference images. CaesarNeRF may create fake images for authentication." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "text", + "content": "H. Zhu and T. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 138, + 140, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 481, + 171 + ], + "type": "text", + "content": "1. Bao, C., Zhang, Y., Yang, B., Fan, T., Yang, Z., Bao, H., Zhang, G., Cui, Z.: Sine: Semantic-driven image-based nerf editing with prior-guided editing field. In: CVPR. pp. 20919-20929 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 481, + 194 + ], + "type": "text", + "content": "2. Bao, Y., Ding, T., Huo, J., Li, W., Li, Y., Gao, Y.: Insertnerf: Instilling generalizability into nef with hypernet modules. ICLR (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 195, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 195, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 138, + 195, + 481, + 216 + ], + "type": "text", + "content": "3. Bao, Y., Li, Y., Huo, J., Ding, T., Liang, X., Li, W., Gao, Y.: Where and how: Mitigating confusion in neural radiance fields from sparse inputs. arXiv preprint arXiv:2308.02908 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 217, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 481, + 237 + ], + "type": "text", + "content": "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In: CVPR. pp. 5470–5479 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 239, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 481, + 259 + ], + "type": "text", + "content": "5. Buehler, C., Bosse, M., McMillan, L., Gortler, S., Cohen, M.: Unstructured lumigraph rendering. In: SIGGRAPH. pp. 425-432 (2001)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 261, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 481, + 281 + ], + "type": "text", + "content": "6. Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In: ICCV. pp. 14124-14133 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 283, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 283, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 138, + 283, + 481, + 314 + ], + "type": "text", + "content": "7. Chen, W., Xu, X., Jia, J., Luo, H., Wang, Y., Wang, F., Jin, R., Sun, X.: Beyond appearance: a semantic controllable self-supervised learning framework for human-centric visual tasks. In: CVPR (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 316, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 481, + 336 + ], + "type": "text", + "content": "8. Chen, Y., Xu, H., Wu, Q., Zheng, C., Cham, T.J., Cai, J.: Explicit correspondence matching for generalizable neural radiance fields. arXiv preprint arXiv:2304.12294 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 338, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 338, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 338, + 481, + 357 + ], + "type": "text", + "content": "9. Chibane, J., Bansal, A., Lazova, V., Pons-Moll, G.: Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In: CVPR. pp. 7911-7920 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 359, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 359, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 138, + 359, + 481, + 390 + ], + "type": "text", + "content": "0. Cong, W., Liang, H., Wang, P., Fan, Z., Chen, T., Varma, M., Wang, Y., Wang, Z.: Enhancing nerf akin to enhancing llms: Generalizable nerf transformer with mixture-of-view-experts. In: ICCV (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 392, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 481, + 413 + ], + "type": "text", + "content": "1. Debevec, P.E., Taylor, C.J., Malik, J.: Modeling and rendering architecture from photographs: A hybrid geometry-and image-based approach. In: SIGGRAPH. pp. 11-20 (1996)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 415, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 481, + 445 + ], + "type": "text", + "content": "2. Deng, C., Jiang, C., Qi, C.R., Yan, X., Zhou, Y., Guibas, L., Anguelov, D.: Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In: CVPR. pp. 20637-20647 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 447, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 481, + 468 + ], + "type": "text", + "content": "3. Deng, K., Liu, A., Zhu, J.Y., Ramanan, D.: Depth-supervised nerf: Fewer views and faster training for free. In: CVPR. pp. 12882-12891 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 469, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 469, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 138, + 469, + 481, + 500 + ], + "type": "text", + "content": "4. Downs, L., Francis, A., Koenig, N., Kinman, B., Hickman, R., Reymann, K., McHugh, T.B., Vanhoucke, V.: Google scanned objects: A high-quality dataset of 3d scanned household items. In: ICRA. pp. 2553-2560 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 502, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 481, + 522 + ], + "type": "text", + "content": "5. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-planes: Explicit radiance fields in space, time, and appearance. In: CVPR. pp. 12479-12488 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 524, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 524, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 524, + 481, + 544 + ], + "type": "text", + "content": "6. Fu, Y., Misra, I., Wang, X.: Multiplane nef-supervised disentanglement of depth and camera pose from videos. In: ICML (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 545, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 481, + 566 + ], + "type": "text", + "content": "7. Gao, Y., Cao, Y.P., Shan, Y.: Surfelnerf: Neural surfel radiance fields for online photorealistic reconstruction of indoor scenes. In: CVPR. pp. 108-118 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 567, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 588 + ], + "type": "text", + "content": "8. Gu, J., Liu, L., Wang, P., Theobalt, C.: Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985 (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 590, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 481, + 609 + ], + "type": "text", + "content": "9. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. NeurIPS pp. 6840-6851 (2020)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 611, + 481, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 642 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 642 + ], + "type": "text", + "content": "20. Irshad, M.Z., Zakharov, S., Liu, K., Guizilini, V., Kollar, T., Gaidon, A., Kira, Z., Ambrus, R.: Neo 360: Neural fields for sparse view synthesis of outdoor scenes. In: ICCV. pp. 9187-9198 (2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "21. Jain, A., Mildenhall, B., Barron, J.T., Abbeel, P., Poole, B.: Zero-shot text-guided object generation with dream fields. In: CVPR. pp. 867-876 (2022)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "22. Jain, A., Tancik, M., Abbeel, P.: Putting nerf on a diet: Semantically consistent few-shot view synthesis. In: ICCV. pp. 5885-5894 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "23. Jang, W., Agapito, L.: Codenerf: Disentangled neural radiance fields for object categories. In: ICCV. pp. 12949-12958 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "24. Jiang, Y., Hedman, P., Mildenhall, B., Xu, D., Barron, J.T., Wang, Z., Xue, T.: Alignerf: High-fidelity neural radiance fields via alignment-aware training. In: CVPR. pp. 46-55 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 183, + 480, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 204 + ], + "type": "text", + "content": "25. Johari, M.M., Lepoittevin, Y., Fleuret, F.: Geonerf: Generalizing nerf with geometry priors. In: CVPR. pp. 18365-18375 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "text", + "content": "26. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and superresolution. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. pp. 694-711. Springer (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 238, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 238, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 480, + 258 + ], + "type": "text", + "content": "27. Kania, A., Kasymov, A., Zieba, M., Spurek, P.: Hypernerfgan: Hypernetwork approach to 3d nerf gan. arXiv preprint arXiv:2301.11631 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 260, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 260, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 133, + 260, + 480, + 281 + ], + "type": "text", + "content": "28. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 282, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 282, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 133, + 282, + 480, + 303 + ], + "type": "text", + "content": "29. Kwak, M., Song, J., Kim, S.: Geconerf: Few-shot neural radiance fields via geometric consistency. arXiv preprint arXiv:2301.10941 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 133, + 304, + 480, + 325 + ], + "type": "text", + "content": "30. Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: Dynibar: Neural dynamic image-based rendering. In: CVPR. pp. 4273-4284 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 133, + 326, + 480, + 357 + ], + "type": "text", + "content": "31. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: CVPR. pp. 300–309 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 358, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 358, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 134, + 358, + 480, + 390 + ], + "type": "text", + "content": "32. Lin, H., Peng, S., Xu, Z., Yan, Y., Shuai, Q., Bao, H., Zhou, X.: Efficient neural radiance fields for interactive free-viewpoint video. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 134, + 392, + 480, + 423 + ], + "type": "text", + "content": "33. Lin, Y., Bai, H., Li, S., Lu, H., Lin, X., Xiong, H., Wang, L.: Componerf: Text-guided multi-object compositional nef with editable 3d scene layout. arXiv preprint arXiv:2303.13843 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 424, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 424, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 134, + 424, + 480, + 445 + ], + "type": "text", + "content": "34. Liu, L., Habermann, M., Rudnev, V., Sarkar, K., Gu, J., Theobalt, C.: Neural actor: Neural free-view synthesis of human actors with pose control. TOC 40(6), 1-16 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 134, + 447, + 480, + 468 + ], + "type": "text", + "content": "35. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: ICCV (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 468, + 480, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 468, + 480, + 499 + ], + "spans": [ + { + "bbox": [ + 134, + 468, + 480, + 499 + ], + "type": "text", + "content": "36. Liu, X., Kao, S.h., Chen, J., Tai, Y.W., Tang, C.K.: Deceptive-nerf: Enhancing nerf reconstruction using pseudo-observations from diffusion models. arXiv preprint arXiv:2305.15171 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 134, + 501, + 480, + 533 + ], + "type": "text", + "content": "37. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 134, + 534, + 480, + 555 + ], + "type": "text", + "content": "38. Liu, Y., Peng, S., Liu, L., Wang, Q., Wang, P., Theobalt, C., Zhou, X., Wang, W.: Neural rays for occlusion-aware image-based rendering. In: CVPR. pp. 7824-7833 (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 556, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 556, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 134, + 556, + 480, + 577 + ], + "type": "text", + "content": "39. Mariotti, O., Mac Aodha, O., Bilen, H.: Viewnerf: Unsupervised viewpoint estimation using category-level neural radiance fields. arXiv preprint arXiv:2212.00436 (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 134, + 578, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 578, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 134, + 578, + 480, + 609 + ], + "type": "text", + "content": "40. Martin-Brualla, R., Radwan, N., Sajjadi, M.S., Barron, J.T., Dosovitskiy, A., Duckworth, D.: Nerf in the wild: Neural radiance fields for unconstrained photo collections. In: CVPR. pp. 7210-7219 (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 134, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 611, + 480, + 632 + ], + "type": "text", + "content": "41. Metzer, G., Richardson, E., Patashnik, O., Giryes, R., Cohen-Or, D.: Latent-nerf for shape-guided generation of 3d shapes and textures. In: CVPR. pp. 12663–12673 (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "42. Mildenhall, B., Srinivasan, P.P., Ortiz-Cayon, R., Kalantari, N.K., Ramamoorthi, R., Ng, R., Kar, A.: Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. TOG 38(4), 1-14 (2019)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "text", + "content": "H. Zhu and T. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "43. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 481, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 481, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 481, + 170 + ], + "type": "text", + "content": "44. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 203 + ], + "type": "text", + "content": "45. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S., Geiger, A., Radwan, N.: Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In: CVPR. pp. 5480-5490 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 204, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 204, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 204, + 480, + 224 + ], + "type": "text", + "content": "46. Noguchi, A., Sun, X., Lin, S., Harada, T.: Neural articulated radiance field. In: ICCV. pp. 5762-5772 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 225, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 246 + ], + "type": "text", + "content": "47. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: ICCV. pp. 5865-5874 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 247, + 480, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 480, + 268 + ], + "type": "text", + "content": "48. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable neural radiance fields for modeling dynamic human bodies. In: ICCV. pp. 14314-14323 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 269, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 269, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 269, + 480, + 289 + ], + "type": "text", + "content": "49. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 290, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 290, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 290, + 480, + 311 + ], + "type": "text", + "content": "50. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: CVPR. pp. 10318-10327 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 312, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 480, + 332 + ], + "type": "text", + "content": "51. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR. pp. 652-660 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 333, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 480, + 354 + ], + "type": "text", + "content": "52. Roessle, B., Barron, J.T., Mildenhall, B., Srinivasan, P.P., Nießner, M.: Dense depth priors for neural radiance fields from sparse input views. In: CVPR. pp. 12892-12901 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 354, + 480, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 375 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 375 + ], + "type": "text", + "content": "53. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 376, + 448, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 448, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 448, + 385 + ], + "type": "text", + "content": "54. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: CVPR (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 407 + ], + "type": "text", + "content": "55. Schonberger, J.L., Zheng, E., Pollefeys, M., Frahm, J.M.: Pixelwise view selection for unstructured multi-view stereo. In: ECCV (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 407, + 480, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 428 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 428 + ], + "type": "text", + "content": "56. Schwarz, K., Liao, Y., Niemeyer, M., Geiger, A.: Graf: Generative radiance fields for 3d-aware image synthesis. NeurIPS 33, 20154-20166 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 429, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 449 + ], + "type": "text", + "content": "57. Shue, J.R., Chan, E.R., Po, R., Ankner, Z., Wu, J., Wetzstein, G.: 3d neural field generation using triplane diffusion. In: CVPR. pp. 20875-20886 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 450, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 471 + ], + "type": "text", + "content": "58. Suhail, M., Esteves, C., Sigal, L., Makadia, A.: Generalizable patch-based neural rendering. In: ECCV. pp. 156-174 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 493 + ], + "type": "text", + "content": "59. Trevithick, A., Yang, B.: Grf: Learning a general radiance field for 3d representation and rendering. In: ICCV. pp. 15182-15192 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 494, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 480, + 514 + ], + "type": "text", + "content": "60. Varma, M., Wang, P., Chen, X., Chen, T., Venugopalan, S., Wang, Z.: Is attention all that nerf needs? In: ICLR (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 514, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 536 + ], + "type": "text", + "content": "61. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 536, + 480, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 556 + ], + "type": "text", + "content": "62. Waechter, M., Moehrle, N., Goesele, M.: Let there be color! large-scale texturing of 3d reconstructions. In: ECCV. pp. 836-850 (2014)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 557, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 480, + 578 + ], + "type": "text", + "content": "63. Wang, G., Chen, Z., Loy, C.C., Liu, Z.: Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In: ICCV (2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 579, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 480, + 610 + ], + "type": "text", + "content": "64. Wang, Q., Wang, Z., Genova, K., Srinivasan, P.P., Zhou, H., Barron, J.T., Martin-Brualla, R., Snavely, N., Funkhouser, T.: Ibrnet: Learning multi-view image-based rendering. In: CVPR. pp. 4690-4699 (2021)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "65. Wang, T., Zhang, B., Zhang, T., Gu, S., Bao, J., Baltrusaitis, T., Shen, J., Chen, D., Wen, F., Chen, Q., et al.: Rodin: A generative model for sculpting 3d digital avatars using diffusion. In: CVPR. pp. 4563-4573 (2023)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "text", + "content": "66. Wei, Y., Liu, S., Rao, Y., Zhao, W., Lu, J., Zhou, J.: Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In: CVPR. pp. 5610-5619 (2021)" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 448, + 102 + ], + "type": "text", + "content": "Calibrated Semantic Representation for Few-Shot Generalizable Neural Rendering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 588 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "67. Wizadwongsa, S., Phongthawee, P., Yenphraphai, J., Suwajanakorn, S.: Nex: Real-time view synthesis with neural basis expansion. In: CVPR. pp. 8534-8543 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "68. Xiangli, Y., Xu, L., Pan, X., Zhao, N., Rao, A., Theobalt, C., Dai, B., Lin, D.: Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In: ECCV (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "69. Xie, C., Park, K., Martin-Brualla, R., Brown, M.: Fig-nerf: Figure-ground neural radiance fields for 3d object category modelling. In: 3DV. pp. 962–971 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 205 + ], + "type": "text", + "content": "70. Xu, D., Jiang, Y., Wang, P., Fan, Z., Shi, H., Wang, Z.: Sinnerf: Training neural radiance fields on complex scenes from a single image. In: ECCV. pp. 736-753 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "71. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-nerf: Point-based neural radiance fields. In: CVPR. pp. 5438-5448 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "72. Yang, B., Zhang, Y., Xu, Y., Li, Y., Zhou, H., Bao, H., Zhang, G., Cui, Z.: Learning object-compositional neural radiance field for editable scene rendering. In: ICCV. pp. 13779-13788 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "73. Yang, H., Hong, L., Li, A., Hu, T., Li, Z., Lee, G.H., Wang, L.: Contranerf: Generalizable neural radiance fields for synthetic-to-real novel view synthesis via contrastive learning. In: CVPR. pp. 16508-16517 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "74. Yang, J., Pavone, M., Wang, Y.: Freenerf: Improving few-shot neural rendering with free frequency regularization. In: CVPR. pp. 8254-8263 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "75. Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstructured multiview stereo. In: ECCV. pp. 767-783 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "76. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: CVPR. pp. 4578-4587 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "text", + "content": "77. Yu, X., Xu, M., Zhang, Y., Liu, H., Ye, C., Wu, Y., Yan, Z., Zhu, C., Xiong, Z., Liang, T., et al.: Mvimgnet: A large-scale dataset of multi-view images. In: CVPR. pp. 9150-9161 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 391, + 480, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 412 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 412 + ], + "type": "text", + "content": "78. Zhang, J., Yang, G., Tulsiani, S., Ramanan, D.: Ners: Neural reflectance surfaces for sparse-view 3d reconstruction in the wild. NeurIPS 34, 29835-29847 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 413, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 434 + ], + "type": "text", + "content": "79. Zhang, J., Li, X., Wan, Z., Wang, C., Liao, J.: Text2nerf: Text-driven 3d scene generation with neural radiance fields. arXiv preprint arXiv:2305.11588 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "text", + "content": "80. Zhenxing, M., Xu, D.: Switch-nerf: Learning scene decomposition with mixture of experts for large-scale neural radiance fields. In: ICLR (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "text", + "content": "81. Zhou, T., Tucker, R., Flynn, J., Fyffe, G., Snavely, N.: Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "type": "text", + "content": "82. Zhou, Z., Tulsiani, S.: Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In: CVPR. pp. 12588-12597 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "text", + "content": "83. Zhu, H., Sun, Y., Liu, C., Xia, L., Luo, J., Qiao, N., Nevatia, R., Kuo, C.H.: Multimodal neural radiance field. In: ICRA. pp. 9393-9399 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "type": "text", + "content": "84. Zhu, H., Zheng, Z., Zheng, W., Nevatia, R.: Cat-nerf: Constancy-aware tx2former for dynamic body modeling. In: CVPRW. pp. 6618-6627 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 567 + ], + "type": "text", + "content": "85. Zhuang, Y., Zhu, H., Sun, X., Cao, X.: Mofanerf: Morphable facial neural radiance field. In: ECCV. pp. 268-285 (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "text", + "content": "86. Zimny, D., Trzeciński, T., Spurek, P.: Points2nerf: Generating neural radiance fields from 3d point cloud. arXiv preprint arXiv:2206.01290 (2022)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 258, + 102 + ], + "type": "text", + "content": "H. Zhu and T. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Caltech Aerial RGB-Thermal Dataset in the Wild/377537ff-2360-4d17-b986-8337083e74f6_content_list.json b/2024/Caltech Aerial RGB-Thermal Dataset in the Wild/377537ff-2360-4d17-b986-8337083e74f6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f029ecd174274c79ba61abc38caba02f4eae70e4 --- /dev/null +++ b/2024/Caltech Aerial RGB-Thermal Dataset in the Wild/377537ff-2360-4d17-b986-8337083e74f6_content_list.json @@ -0,0 +1,1951 @@ +[ + { + "type": "text", + "text": "Caltech Aerial RGB-Thermal Dataset in the Wild", + "text_level": 1, + "bbox": [ + 217, + 141, + 785, + 161 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Connor Lee*, Matthew Anderson*, Nikhil Ranganathan, Xingxing Zuo, Kevin Do, Georgia Gkioxari, and Soon-Jo Chung", + "bbox": [ + 243, + 189, + 756, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "California Institute of Technology", + "bbox": [ + 387, + 232, + 614, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{cee, matta, nrangana, zuox, kdo, georgia, sjchung}@caltech.edu", + "bbox": [ + 251, + 247, + 750, + 260 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We present the first publicly-available RGB-thermal dataset designed for aerial robotics operating in natural environments. Our dataset captures a variety of terrain across the United States, including rivers, lakes, coastlines, deserts, and forests, and consists of synchronized RGB, thermal, global positioning, and inertial data. We provide semantic segmentation annotations for 10 classes commonly encountered in natural settings in order to drive the development of perception algorithms robust to adverse weather and nighttime conditions. Using this dataset, we propose new and challenging benchmarks for thermal and RGB-thermal (RGB-T) semantic segmentation, RGB-T image translation, and motion tracking. We present extensive results using state-of-the-art methods and highlight the challenges posed by temporal and geographical domain shifts in our data. The dataset and accompanying code is available at https://github.com/aerorobotics/caltech-aerial-rgbt-dataset.", + "bbox": [ + 259, + 294, + 743, + 489 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Robotics $\\cdot$ Semantic segmentation $\\cdot$ RGB-T image translation $\\cdot$ Visual inertial odometry", + "bbox": [ + 261, + 502, + 740, + 530 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c66c1a82a96585488db1264a7a275c3e3743fbc07f341757e581bcdcc184c156.jpg", + "image_caption": [ + "Thermal Dataset Comparisons" + ], + "image_footnote": [], + "bbox": [ + 217, + 575, + 366, + 688 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/acaa29bae9a83e7c9278a033abb99491c6761ad857f96e58994dbd3f761588bd.jpg", + "image_caption": [ + "A. Thermal + RGB-T Semantic Segmentation", + "Aligned RGB-T pairs" + ], + "image_footnote": [], + "bbox": [ + 388, + 575, + 625, + 616 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d092be201dc124929f1cd40740c76ad1ac02d792f663a20c7faf0d7b90b5c546.jpg", + "image_caption": [ + "B. RGB-to-Thermal Image Translation", + "Fig. 1: Left: Our dataset is uniquely designed to improve thermal scene perception for field robots. Right: We provide new benchmarks for thermal-based vision algorithms, including (a) semantic segmentation, (b) image translation, and (c) motion tracking." + ], + "image_footnote": [], + "bbox": [ + 390, + 641, + 584, + 689 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d04c086a5971a11d35af62916fb5bd6da8ce7ff746002449eacdb275aa97f438.jpg", + "image_caption": [ + "C. Motion Tracking" + ], + "image_footnote": [], + "bbox": [ + 630, + 582, + 785, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 761, + 375, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Field robots rely predominantly on visual cameras, lidar, and radar to perceive their surroundings [16, 18]. These sensors provide robust perception capabilities,", + "bbox": [ + 212, + 785, + 785, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* These authors contributed equally to this work.", + "bbox": [ + 217, + 824, + 550, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "but falter in low-light and adverse weather conditions [70]. In contrast, thermal cameras use long-wave infrared wavelengths to capture emitted heat, offering dense, visual information in such conditions where other sensors struggle [19]. Recently, they have been used for multimodal perception in autonomous driving and are increasingly being used to enable nighttime robot autonomy [11, 12, 36, 37, 49, 71, 76, 78]. However, successful integration of thermal cameras requires extensive datasets across diverse settings, preventing widespread adoption in field robotics.", + "bbox": [ + 212, + 146, + 787, + 265 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Current thermal datasets focus on urban settings for autonomous driving applications (Fig. 1). They comprise of thermal-only imagery [36,76] or RGB-Thermal (RGB-T) pairs [17,22,41,71], with some including global positioning (GPS/GNSS) and inertial measurements (IMU) for visual-inertial odometry (VIO) and simultaneous localization and mapping (SLAM) [11,61,78]. While comprehensive, they lack data depicting natural settings like rivers and forests, which are typical operating areas for field robotics that perform coastline mapping and bathymetry [4] or activity monitoring during forest fires [30].", + "bbox": [ + 212, + 268, + 787, + 388 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to the lack of thermal benchmarks for algorithms like semantic segmentation and SLAM, field robots cannot easily operate at night or in adverse conditions, requiring online learning [34, 72] or unsupervised domain adaptation to compensate [20, 32, 71], which still require labeled data for evaluation. As such, enabling thermal perception requires collecting and annotating field-specific data from scratch. However, creating a thermal dataset for field robotics is difficult due to two reasons: First, thermal data is hard to crowdsourced and web-scrape due to the richness of thermal cameras. Hence, acquiring a diverse dataset for algorithm development requires physically going to distinct field locations, increasing time and cost. Second, operating field robots, especially uninhabited aerial vehicles (UAV), usually requires special permits for each location-of-capture.", + "bbox": [ + 212, + 390, + 787, + 555 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present the first dataset targeting thermal scene perception for field robotics. The dataset consists of oblique-facing imagery captured from a UAV, supplemented with image sets captured at ground level, and focuses on littoral settings within various desert, forest, and coastal environments across the United States. We contribute new benchmarks for thermal and RGB-T semantic segmentation, RGB-T image translation, and VIO/SLAM algorithms, with unique challenges characterized by temporal and geographical domain shifts for learning-based methods and feature sparsity for motion-tracking algorithms. Aside from the proposed benchmarks, our dataset can also be used towards other applications like visual place recognition and terrain-relative navigation.", + "bbox": [ + 212, + 556, + 787, + 705 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The paper is organized as follows: Sec. 2 reviews relevant datasets, benchmarks, and algorithms. Sec. 3 describes our dataset details and curation process, and Sec. 4 presents benchmark results. Sec. 5 offers concluding remarks.", + "bbox": [ + 212, + 708, + 785, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 763, + 388, + 780 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Datasets and Benchmarks", + "text_level": 1, + "bbox": [ + 214, + 791, + 477, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thermal/RGB-T benchmarks for urban robotics Current urban thermal and RGB-T datasets primarily focus on autonomous vehicle (AV) technology", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "C. Lee et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/c28f3f6765115e48ce8a0c421d7fe7c592d232d413cb16be8c7f2798c517db91.jpg", + "table_caption": [ + "Table 1: Comparison of datasets captured from aerial platforms. Camera angle of $90^{\\circ}$ is nadir-pointing. See Tab. S1 in supplement for comparison against thermal datasets captured from surface and ground vehicles." + ], + "table_footnote": [], + "table_body": "
DatasetTaskRGBThermalGPSIMUSettingLocation# SamplesCamera pose
VisDrone [83]Obj. Det.✓XXXUrbanChina10,209Variable
HIT-UAV [67], NII-CU [63]Obj. Det.✓✓XXUrbanChina, Japan2,898 - 5,88030°, 90°, 60-130 m
WTT-UAS [30]Obj. Det.X✓✓✓ForestUSA6,95130°, 90°, < 120 m
BIRDSAI [3]Obj. Track.X✓XXSavannasS. Africa62,000Off-nadir, 60-120 m
AeroScapes [48], UAVid [45], [10], [6], Swiss/Okutama [63]Sem. Seg.✓XXXUrbanChina, Japan, Switzerland191 - 326930 - 90°, < 100 m
Semantic Drone Dataset [47]Sem. Seg.✓✓X✓UrbanGermany40090°, < 30m
OursSem. Seg.✓✓✓✓River, Coast, Lake, Mountain, DesertUSA4,19520°, 45°, < 120 m
", + "bbox": [ + 215, + 193, + 784, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and surveillance applications. AV-related datasets cover tasks such as object detection [11, 17, 41], semantic segmentation [22, 36, 71, 76], and localization [11, 61, 78]. Some include data from other spectra, benchmarking not only thermal algorithms but also RGB-T [11, 22, 41, 63, 71] and other multispectral models. In contrast, datasets for surveillance applications use fixed cameras [29] and UAVs [63, 67] to detect personnel, vehicles, and other objects. Despite many urban thermal benchmarks, they cannot be directly used to develop algorithms for robots operating in natural environments due to the domain gap.", + "bbox": [ + 212, + 299, + 784, + 420 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thermal/RGB-T benchmarks for field robotics Datasets for robots operating in non-urban environments remain limited. MassMIND [49] and PST900 [62] benchmark thermal semantic segmentation for maritime and subterranean robotic environments, respectively. WIT-UAS [30] provides UAV-borne thermal images for object detection in wildfire-prone environments. One work [34] proposes an aerial RGB-T dataset for thermal water segmentation in littoral areas but only provides 1272 labels consisting of a single water class. Benchmarks like [3,7] depict natural environments but are targeted towards wildlife tracking. To address this gap, we present an aerial RGB-T dataset which can be used to benchmark multiple tasks, including semantic segmentation, RGB-T image translation, and localization, across diverse natural environment robotic applications. Specifically, we build upon the dataset from [34] by extending the number of semantic classes from 1 to 10 and adding additional RGB-T data captured from air and ground, resulting in a total of 4195 semantic segmentation annotations. In comparison to existing thermal/RGB-T datasets, our dataset can enable nighttime field robotic applications like remote sensing research, geological sampling, precision agriculture, and wilderness search and rescue.", + "bbox": [ + 212, + 426, + 787, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "UAV-based semantic segmentation datasets Aside from [34], existing UAV-captured semantic segmentation datasets [6, 10, 45, 47, 48, 63] cover urban environments (Tab. 1). With the exception of [47] which provides 400 thermal image samples, all of these datasets contain only RGB data for semantic segmentation. While they provide images with relevant view angles, the modality gap and focus on urban content makes them impractical for aerial field robotic applications.", + "bbox": [ + 212, + 689, + 787, + 780 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Algorithms using Thermal Datasets", + "text_level": 1, + "bbox": [ + 214, + 789, + 555, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thermal semantic segmentation Since thermal images are spatially identical to their RGB counterparts, popular RGB semantic segmentation algo", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Caltech Aerial RGB-Thermal Dataset in the Wild", + "bbox": [ + 397, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "rithms [5, 8, 53, 74] can be applied directly. However, some works [36, 51, 76] develop models specifically for thermal imagery, and leverage edge priors to overcome low resolution and blurriness due to thermal crossover [56]. Although specifically designed for thermal, these models only marginally outperform standard RGB segmentation models [51] and their performance in field settings, where edges are less structured, remains untested. Aside from architectural advancements, improvements in thermal semantic segmentation come from domain adaptation (DA) methods [20, 36, 69, 71] that use labeled RGB data with unlabeled thermal data. To provide fair comparisons, we do not provide DA baselines in our benchmarks as performance varies based on the choice of RGB data.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "RGB-T semantic segmentation RGB-T semantic segmentation methods compensate for the weaknesses of each modality via deep feature fusion, data augmentation, and adversarial training. Most approaches [22, 60, 65, 66, 79] utilize encoder-decoder architectures with modality-specific encoders and shared decoders, and show significant improvements over channel-stacked RGB-T inputs. Works like [39, 81] make improvements by integrating RGB and thermal features at multiple stages within the encoders, while [62] eschews RGB-T training pairs by processing RGB-T inputs sequentially. Other methods include random input masking to reduce reliance on a single modality [60] and adversarial training to adapt to different times of day [71].", + "bbox": [ + 212, + 301, + 789, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "RGB-T image translation RGB-T image translation is vital as it may enable thermal training data to be generated from large-scale RGB datasets. Works based on generative adversarial networks (GAN) such as [27,35,42,52,82] can be used without RGB-T image pairs. However, GANs like [28,73] that require paired imagery provide better translations for both visual appeal and domain adaptation [36]. Recent diffusion methods [57,58] have demonstrated remarkable results on RGB imagery, but have not been tested on RGB-T image translation.", + "bbox": [ + 212, + 455, + 787, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thermal VIO and SLAM Low signal-to-noise ratio, reduced contrast, and blurred boundaries in thermal images present challenges for VIO and SLAM. Several studies [2, 13, 14, 24, 26] use indirect methods by tracking keypoints extracted from thermal images for visual constraints and rely on image normalization techniques to improve keypoint extraction. In contrast, [31] uses a direct method to extract and track high-gradient points by minimizing radiometric error. Other works utilize learned local and global features to improve feature tracking [80] and loop closure detection [59], respectively.", + "bbox": [ + 212, + 565, + 787, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 The Caltech Aerial RGB-Thermal Dataset", + "text_level": 1, + "bbox": [ + 214, + 696, + 663, + 713 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Data Acquisition Hardware", + "text_level": 1, + "bbox": [ + 214, + 723, + 490, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To capture this dataset, we developed a custom sensor stack with synchronized RGB-thermal imagery, IMU, and global positioning data (Fig. 2). The sensor stack features a non-radiometric FLIR ADK thermal camera ( $640 \\times 512$ px and $75^{\\circ}$ horizontal FoV, $60\\mathrm{Hz}$ ) flanked by a pair of FLIR Blackfly electro-optical (EO) monochrome and color cameras $^{1}$ ( $960 \\times 600$ px and $75^{\\circ}$ horizontal FoV, $30\\mathrm{Hz}$ ).", + "bbox": [ + 212, + 739, + 787, + 816 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "C. Lee et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1 BFLY-U3-23S6M-C mono/color with Tamron M112FM08 lenses", + "bbox": [ + 217, + 823, + 661, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pose information is provided by a VectorNav VN100 IMU (200 Hz) and a u-blox M8N GPS (5 Hz). All three cameras and the VN100 are hardware synchronized using a dedicated signal generator and are rigidly attached to each other using a stiff, 3D printed mount. A Simply NUC Ruby Mini PC (AMD Ryzen 7 4800U, 32 GB RAM) is used for the compute, and the entire system is interfaced using ROS Noetic. Our cameras and IMU were calibrated following the procedures found in Supplementary Material Sec. S2.2 and were not disassembled from each other between any of the dataset collections.", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bfbf75378943dbb397f1282aab309c5ea349027af3482b5447f284aaca0de067.jpg", + "image_caption": [ + "Fig. 2: (a) The Aurelia X6 hexacopter and sensor stack used to capture our aerial dataset. (b) Geographic distribution of our data collection sites and collection times." + ], + "image_footnote": [], + "bbox": [ + 225, + 276, + 405, + 468 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8c39ed1a4de2c64a999a423b96d24ac4a04788514abf50adee8c5fb6bd87997a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 276, + 776, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The sensor stack is mounted rigidly onboard an Aurelia X6 hexacopter (Fig. 2) without gimbal stabilization at angles between $0^{\\circ}$ (level with the horizon) to $45^{\\circ}$ downwards to provide different viewpoints. When mounted to a UAV, we also log the position and attitude estimates from the UAV at $20\\mathrm{Hz}$ providing improved positioning accuracy from the onboard RTK GPS. Additionally, the sensor stack's modularity allows us to capture datasets on foot by mounting it to a tripod where flight restrictions are in-place.", + "bbox": [ + 212, + 516, + 787, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Data Collection and Processing", + "text_level": 1, + "bbox": [ + 214, + 632, + 522, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data capture We captured 37 aerial and ground trajectories covering lakes, rivers, coastlines, deserts, and mountains from around Southern California, Kentucky, and North Carolina (Fig. 2). Aerial trajectories (18) involve high motion with intermittent hovering, and mostly depict scenes from $40\\mathrm{m}$ altitude. Ground-level trajectories (19) were captured on foot, with 7 captured without any movement for VIO debugging purposes. For more details, see Supplementary Materials Tab. S2 and Sec. S2.1.", + "bbox": [ + 212, + 652, + 787, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thermal image normalization For the baselines presented in this work, we typically normalize 16-bit thermal data by rescaling between the $1^{\\mathrm{st}}$ and $99^{\\mathrm{th}}$ percentile pixel values and follow with contrast limited adaptive histogram equalization (CLAHE). We keep normalized data as floats to minimize information loss, and only convert to 8-bit format when necessary or for visualization.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Caltech Aerial RGB-Thermal Dataset in the Wild", + "bbox": [ + 397, + 114, + 732, + 127 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/762c02a09c796d13493e5be1f26e0262c19945d73ac2f8deb360b8dab5b20328.jpg", + "image_caption": [ + "Semantic Segmentation Classes and Description" + ], + "image_footnote": [], + "bbox": [ + 228, + 157, + 514, + 287 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/026bac4764f8e46486c4665e30fc64344aeeba0305af094380eedf3b0642642b.jpg", + "image_caption": [ + "a", + "b" + ], + "image_footnote": [], + "bbox": [ + 539, + 148, + 776, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dfe213dcb7b6976c69c56621d6fea9ca1dbc70f616b4a21e7c7b5d2eeb685d65.jpg", + "image_caption": [ + "Fig. 3: Semantic segmentation classes in our dataset. The color mapping is used throughout this paper. (a) Hourly distribution of annotated thermal images. (b) Histogram of semantic classes." + ], + "image_footnote": [], + "bbox": [ + 539, + 217, + 764, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Semantic segmentation annotation To aid development of scene perception algorithms for field robots, we manually-annotated a subset of thermal images for semantic segmentation with classes in Fig. 3. To avoid redundancy, images were sampled every 3s when moving and every 20s when still, resulting in 4195 samples. We considered a frame in motion if the distance moved within the past 2 seconds exceeded 0.5 meters, with distances computed using GPS coordinates.", + "bbox": [ + 212, + 349, + 784, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To promote labeling consistency, annotators preread descriptions of each trajectory, which contained a list of classes present, labeled examples provided by the authors, and GPS coordinates (for Google Earth look-up). Annotators were then given an RGB-T image pair with trajectory metadata and asked to label the thermal frame. Annotations were reviewed by authors, with inconsistent or incorrect labels sent back. 3 rounds of annotation were done, with the final and initial set of labels differing by $0.08\\mathrm{mIoU}$ . Classes most frequently re-annotated were bare ground, rocky terrain, road, trees, and vehicles.", + "bbox": [ + 212, + 444, + 787, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "RGB-thermal image alignment To align thermal and RGB image pairs, we stereo rectified the RGB-T image pairs using the camera matrices obtained from calibration (Sec. S2.2) and projected the thermal image into the larger RGB image frame to preserve RGB resolution. As our baseline is small compared to the depth of the scenes we capture, image pairs are near-coregistered after rectification. Due to differences in optics, the field-of-view of the aligned image pairs is narrower compared to that of the thermal camera.", + "bbox": [ + 212, + 571, + 787, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Dataset Splits", + "text_level": 1, + "bbox": [ + 214, + 689, + 380, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "General (benchmark) split This is our primary split for semantic segmentation and image translation benchmarks. We randomly partition the annotated thermal dataset (Sec. 3.2) into train/val/test sets at a 75:12.5:12.5 ratio.", + "bbox": [ + 212, + 710, + 785, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Temporal split To promote studies into the effect of different time-of-day capture of the thermal images, we split the dataset using three time periods: twilight/sunrise (5 AM - 7 AM), daytime (10 AM - 5 PM), and nighttime (7 PM - 4 AM). We randomly partition the daytime images into train/val/test sets using the ratio from above and leave the twilight and nighttime sets for testing.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "C. Lee et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Geographical split The geographical splits are intended to test algorithm adaptability to unseen settings. We provide two splits:", + "bbox": [ + 215, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Terrain-based: This split partitions data into the following terrain categories: aerial river, aerial lake, aerial coast, and ground-captured images. Train/val/test splits are created per category with a 50:15:35 ratio.", + "2. Region-based: This is based on the area of the United States where the data was captured. Distinct regions include California (CA), Kentucky (KY), and North Carolina (NC), with train/val/test created in the same way as above." + ], + "bbox": [ + 220, + 188, + 784, + 279 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 292, + 375, + 309 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We provide baselines for the following benchmarks: thermal semantic segmentation, RGB-T semantic segmentation, RGB-T image translation, and VIO/SLAM. Furthermore, we explore the zero-shot thermal segmentation capabilities of large vision-language models on our dataset and perform further ablation studies.", + "bbox": [ + 212, + 316, + 782, + 377 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d735ba5de4554b9e51e03a6142064d20f32236d745eec9fb142908ab89ceaff9.jpg", + "table_caption": [ + "Table 2: Supervised model performance (mIoU), framerate (FPS at batch size 1), and floating point operations (FLOP) for various semantic segmentation network baselines." + ], + "table_footnote": [ + "$\\dagger$ Thermal-specific model * Frozen encoder" + ], + "table_body": "
ModelBare groundRocky terrainDevelop. struct.RoadShrubsTreesSkyWaterVehiclesPersonStuffObjectAllOrin (GPU)NUC (CPU)FLOPS (G)
FastSCNN [53]0.8070.9070.7280.6310.6940.7760.9560.9650.3840.0580.8080.2210.690124.285.61.0
MobileNet-V3-S 0.75 [25]0.8020.9140.6760.6330.7040.8140.9550.9760.3830.1110.8090.2470.69775.224.03.3
MobileNet-V3-L 0.750.8010.9140.7010.6150.7130.7930.9590.9760.4320.1480.8090.2900.70559.015.54.8
MobileViTV2 0.50 [46]0.8220.9130.7590.6500.6940.7810.9540.9700.2220.0300.8180.1260.68060.110.75.5
EfficientViT-B0 [5]0.8250.9170.7770.6180.7100.7930.9610.9650.4950.1910.8210.3430.72551.25.53.9
EfficientNet-Lite0 [68]0.8190.9080.7690.6010.7090.8050.9620.9680.3800.0790.8180.2290.70047.712.87.2
EfficientNet-Lite20.8250.9200.7650.6140.6980.8050.9580.9710.3860.1180.8200.2520.70638.911.29.4
Segformer-B0 [74]0.8040.9100.6900.6240.6960.7880.9600.9700.1950.0000.8050.0970.66438.35.910.2
Segformer-B10.8140.9090.7730.6030.7130.7990.9600.9650.3660.0000.8170.1830.69030.33.519.9
ResNet18 [23]0.8100.9050.7660.6180.7110.8070.9590.9660.4310.1580.8180.2940.71369.312.422.4
ResNet500.8190.9160.7470.6580.7110.7990.9610.9740.3610.1820.8230.2720.71329.96.145.4
ResNeXt50 [75]0.8250.9190.7730.6530.7090.7940.9640.9760.4360.1370.8270.2870.71923.05.645.5
ConvNext-T [44]0.7990.9090.7190.6080.7060.8080.9610.9700.3630.0030.8100.1830.68525.83.347.4
ConvNext-S0.8100.9130.7710.6030.7180.8100.9650.9640.4920.0480.8190.2700.70919.12.375.0
ConvNext-B0.8100.9210.6970.6110.7060.8030.9630.9720.3480.1120.8110.2300.69414.61.4130.5
ConvNext-B (CLIP) [55]0.8130.9180.6830.6320.7180.8130.9650.9720.5170.1370.8140.3270.71714.61.4130.4
ConvNext-B (CLIP) *0.7730.8870.7130.4870.6560.7560.9460.9510.1580.0190.7710.0890.63514.61.4130.4
DINOv2 [50] (linear head)0.8000.9070.6810.6060.6930.7960.9560.9670.3750.0790.8010.2270.68615.22.1—
DINOv2 (linear head) *0.7050.8440.6190.4190.5580.7250.9220.9200.3560.1580.7140.2570.62315.22.1—
DINOv2 (nonlin. head)0.8100.9160.7080.6350.7000.8190.9590.9730.3990.1190.8150.2590.70415.22.1—
DINOv2 (nonlin. head) *0.7960.9030.7320.6060.6910.7940.9540.9630.4710.1490.8050.3100.70615.22.1—
FTNet† [51]0.7550.8670.6350.5760.6430.6530.7870.9470.2340.0240.7330.1290.61311.11.4100.5
", + "bbox": [ + 215, + 424, + 784, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Thermal Semantic Segmentation", + "text_level": 1, + "bbox": [ + 215, + 636, + 532, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We run multiple baselines on our general split before choosing a specific model to test on the geographic and temporal splits. We go over our baselines before analyzing their performance on our benchmarks.", + "bbox": [ + 212, + 654, + 782, + 699 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines Motivated by robotic applications, we are interested in both inference speed and segmentation performance. Our baselines place an emphasis on lightweight, real-time networks, but also on foundation models like DINOv2 [50] and ConvNext-B (CLIP) [44, 55] to explore any latent multimodal capabilities. We also test a thermal-specific model, FTNet [51], which uses an edge-based loss function to compensate for blurred boundaries in thermal images.", + "bbox": [ + 212, + 703, + 784, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Besides FastSCNN [53], EfficientViT [5], Segformer [74], DINOv2, and FT-Net, all network encoders in Tab. 2 feed into a DeepLabV3+ segmentation head. In particular, DINOv2 was employed with linear and non-linear, multi-scale", + "bbox": [ + 212, + 794, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Caltech Aerial RGB-Thermal Dataset in the Wild", + "bbox": [ + 397, + 114, + 732, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "heads in order to study the generalization capacity of its pretrained features to the thermal modality. All baselines except DINOv2 and ConvNext-B (CLIP) were trained from pretrained ImageNet weights. All networks, besides FTNet, were trained using the cross entropy loss. Further details on baseline implementation and training can be found in Supplementary Material Sec. S3.1.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a2591cf958e101bfcfcee7b7f300ddc5ece351de40c0abfb93ca990a7e49c77d.jpg", + "image_caption": [ + "Fig. 4: Thermal images and semantic segmentation labels from each capture area with inference results from EfficientViT, FastSCNN, and ConvNext-B (CLIP)." + ], + "image_footnote": [], + "bbox": [ + 225, + 232, + 779, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "General benchmark We use the general (random) split (Sec. 3.3) to compare thermal semantic segmentation performance between baselines (Tab. 2). With field robotic applications in mind, we analyze segmentation results in context of GPU and CPU frame rates, which were measured onboard the Nvidia Jetson AGX Orin (GPU) and the Ruby NUC (CPU) embedded platforms, respectively2.", + "bbox": [ + 212, + 507, + 787, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In general, we found that larger models provide marginal benefit over smaller, compute-efficient models. In particular, a 0.027 mIoU difference between the largest (ConvNext-B) and fastest (FastSCNN) baselines comes at the cost of $130 \\times$ more FLOPs. Overall, EfficientViT-B0 performed the best, attaining the highest mIoU (0.725) while performing $50\\mathrm{Hz}$ inference on the Jetson AGX Orin.", + "bbox": [ + 212, + 585, + 787, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While large RGB foundation models show little gain over ImageNet pretraining, their features do extend to the thermal modality. Specifically, a frozen DINOv2 (linear head) outperforms the thermal-specific FT-Net. With a nonlinear head, it matches the performance of other end-to-end trained networks. While this shows the benefit of large-scale RGB pretraining, the choice of RGB pretraining data or learning strategy seems to matter. Notably, the frozen ConvNext-B (CLIP) model gives marginal gain over the frozen DINOv2 (linear head), despite a larger, nonlinear segmentation head. Overall, foundation models do not confer", + "bbox": [ + 212, + 665, + 787, + 787 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "C. Lee et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "2 Models of input size $512 \\times 640$ were converted to ONNX and inference speed was averaged over 50 forward passes after a warm-up period.", + "bbox": [ + 217, + 810, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "significant performance advantages for semantic segmentation on our dataset in order to justify their computational cost.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Overall, all baselines can segment stuff classes well but none excel with the rare object classes: vehicle and person. Future models would likely need to consider weighted loss functions, smarter data augmentation, or domain adaptation techniques in order to improve on object classes. Lastly, we note that the thermal-specific model, FTNet, performs poorly, likely as its edge priors are less effective in unstructured, natural environments compared to urban environments.", + "bbox": [ + 212, + 196, + 784, + 287 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/88dffea47f008d8f3fcc643d7b80116e98c6a3535a63ec3af6e34eadd7692487.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 301, + 777, + 435 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/959545bc483c2744d0a2b4ad7268e29fff5c154fb6de346b036a45c4f21ffd96.jpg", + "image_caption": [ + "Fig. 5: (a) Thermal semantic segmentation failures due to intra-class semantic variations when testing on geographically-partitioned, out-of-domain data. (b) Failures due to photometric variations between day (in-domain) and night (out-of-domain)." + ], + "image_footnote": [], + "bbox": [ + 233, + 436, + 777, + 559 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Geographically-partitioned benchmarks We retrained our best baseline (Tab. 2) on geographically-partitioned sets to evaluate intra-thermal generalization. Given the cost and effort of collecting diverse aerial thermal data of field settings, we aim to determine the minimum amount of annotated data required to effectively deploy. Using EfficientViT-B0, we performed two experiments to study generalization across difference geographic regions and terrain types.", + "bbox": [ + 212, + 625, + 784, + 715 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/443b38ac82f58e0680036f50defabfe7e43e86f1291af699b1facd5b199041bc.jpg", + "table_caption": [ + "Table 3: Thermal semantic segmentation results on geographically-split data", + "(a) Region-based split" + ], + "table_footnote": [], + "table_body": "
Train AreaTest AreaAvg. mIoU
CANCKY
CA0.6660.0840.1930.314
NC0.0840.5080.0310.208
All0.6480.5250.5870.586
", + "bbox": [ + 236, + 763, + 493, + 825 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/6064778701cec0ecfc8495e2ed7abeaab2b7ba5269934382e1447e266218e12e.jpg", + "table_caption": [ + "(b) Terrain-based split" + ], + "table_footnote": [], + "table_body": "
Training TerrainTesting TerrainAvg. mIoU
RiverLakeCoast
River0.6680.1960.1170.327
Lake0.1960.3640.0610.207
Coast0.1270.1190.5220.256
All0.6460.4150.4930.518
All + ground0.6580.4160.5200.532
", + "bbox": [ + 504, + 763, + 761, + 837 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Caltech Aerial RGB-Thermal Dataset in the Wild", + "bbox": [ + 397, + 114, + 732, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In the first experiment, we trained on thermal data from CA and tested on images from KY and NC (Tab. 3a). Despite common classes in each region, results show poor generalization to class variations across geographic areas. We see this again when repeating with the NC split. In the second experiment, we divided our data into three terrain types, aerial river, aerial lake, and aerial coast. Remaining ground-captured images were lumped into a fourth ground category. We trained and tested all pairwise combinations and found poor generalization across terrain types as well (Tab. 3b). Notably, we found low, in-domain performance for the lake setting but saw gains after training on aerial data from other terrain and further improvements after training on ground level images. Overall, our results reiterate the benefit of having training data representative of the testing environment, but also show that gains can be made by including out-of-domain (OOD) data from different terrain and different viewpoints.", + "bbox": [ + 212, + 146, + 787, + 343 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Overall, the geographic benchmark presents a difficult domain adaptation problem, requiring both inter- and intra-modality domain adaptation techniques to overcome. Although such techniques could perform well in this challenge, the results also highlight a need for an even larger and more diverse thermal dataset capturing common field environments than what we presented.", + "bbox": [ + 212, + 349, + 787, + 426 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Temporally-partitioned benchmark Here, we assess EfficientViT-B0's performance against intra-thermal, temporal distribution shift due to thermal inversion. We train the model on images from the daytime and test on images captured at sunrise and at night. The model achieved mIoUs of 0.242 at sunrise, 0.193 at night, and 0.777 on a separate daytime test split. Despite training on daytime images from the same location, the model fails on nighttime scenes, struggling to classify water due to thermal inversion (Fig. 5b). As such, to perform well on this split, future algorithms need to augment datasets to simulate such inversions or intentional collect data capturing such phenomena.", + "bbox": [ + 212, + 436, + 787, + 574 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/9ec779b42a5425fddda1a42c53fb70e90294d2aea5aa83095ad104c8a951bd9f.jpg", + "table_caption": [ + "Table 4: RGB-Thermal semantic segmentation network baseline results (mIoU)." + ], + "table_footnote": [ + "$\\dagger$ DeepLabV3+ (w/ ResNet50 encoder)" + ], + "table_body": "
ModelBare groundRocky terrainDevel. struct.RoadShrubsTreesSkyWaterVehiclesPersonAllFLOPS (G)
RGB Only†0.7900.8730.8080.5590.6640.7520.8910.9670.3990.0000.67045
Thermal Only†0.7660.8350.7890.5360.6540.7490.9070.9710.3990.0000.66145
EAEFNet [39]0.8050.8630.8130.5450.7020.8020.9370.9760.4810.0000.692358
CRM [60]0.8160.9000.8510.6160.6950.7790.9230.9740.3770.2060.714310
CMNeXt [79]0.8300.9000.8610.6070.7180.8080.9330.9800.5600.1900.740148
", + "bbox": [ + 217, + 604, + 782, + 676 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 RGB-T Semantic Segmentation", + "text_level": 1, + "bbox": [ + 215, + 712, + 522, + 729 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We use the paired RGB-T dataset (Sec. 3.2) and partition it using the general split (Sec. 4.1). Recall that RGB and thermal input sizes are $960 \\times 600$ pixels and the thermal FoV is narrower than in the previous benchmarks (Sec. 4.1).", + "bbox": [ + 212, + 737, + 785, + 784 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baselines We test three RGB-T semantic segmentation algorithms achieving state-of-the-art results in urban RGB-T scene segmentation benchmarks: CRM [60], EAEFNet [39], and CMNeXt [79]. To compare against single-modality mod-", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "C. Lee et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/d43441f2e987afa8bcd53dbdfa70daf5c177f73b4a06e1572d760593eb6933af.jpg", + "table_caption": [ + "Table 5: RGB-Thermal image translation results" + ], + "table_footnote": [], + "table_body": "
MethodMethod TypeRGB → Thermal
PSNR ↑SSIM ↑mIoU ↑
UNIT [42]Unpaired GAN12.850.4210.189
MUNIT [27]Unpaired GAN16.080.4590.173
Edge-guided RGB-T [35]Unpaired GAN12.200.4360.225
Pix2Pix [28]Paired GAN20.060.5470.354
Pix2PixHD [73]Paired GAN20.050.5790.379
VQGAN [15]Paired GAN18.090.5400.388
Palette [58]Paired Diffusion11.510.3990.194
", + "bbox": [ + 215, + 166, + 782, + 267 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "els, we train two additional ResNet50/DeepLabV3+ segmentation models on RGB and thermal data accordingly (Sec. 4.1).", + "bbox": [ + 212, + 279, + 782, + 310 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Performance analysis All RGB-T models outperform the single-modality baselines, with CMNeXt outperforming the next best RGB-T model by 0.026 mIoU (Tab. 4). Overall, incorporating RGB imagery into the segmentation pipeline greatly improved the distinction between land-based stuff classes which can be hard to distinguish in thermal due to thermal crossover. However, these improvements come at a high computational cost and would not be suitable for field robotic applications. Excelling in this benchmark would involve improving the segmentation performance of the object classes, possibly by leveraging RGB data in unsupervised domain adaptation training schemes.", + "bbox": [ + 212, + 315, + 784, + 450 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 RGB-Thermal Image Translation", + "text_level": 1, + "bbox": [ + 214, + 460, + 535, + 477 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We use the RGB-T paired dataset from Sec. 4.2 to benchmark image translation algorithms operating in the RGB $\\rightarrow$ T direction.", + "bbox": [ + 212, + 482, + 782, + 512 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Baselines We benchmark GAN methods that need RGB-T pairs (Pix2Pix [28], Pix2PixHD [73], VQGAN [15]) and unpaired methods (UNIT [42], MUNIT [27], [35]) that do not. We also evaluate Palette [58], a paired diffusion approach.", + "bbox": [ + 212, + 518, + 782, + 565 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/a9a61df679ccbf54e3fb41a2d642ff4c809d26829049499726c6365b8fd2beef.jpg", + "image_caption": [ + "Fig. 6: RGB-to-thermal image translation results. Zoom-in to see fine-details." + ], + "image_footnote": [], + "bbox": [ + 225, + 575, + 777, + 714 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Image translation metrics We quantify RGB-T image translation using the Peak-Signal-to-Noise ratio (PSNR) and the Structural Similarity Index (SSIM). In addition, we propose a third metric: the thermal mIoU. This is equivalent to the FCN-score used in RGB image translation works [28]. Instead of an RGB network, our metric uses our EfficientViT thermal segmentation network (Tab. 2) to segment translated thermal images before evaluating with ground truth.", + "bbox": [ + 212, + 750, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Caltech Aerial RGB-Thermal Dataset in the Wild", + "bbox": [ + 397, + 114, + 732, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Performance analysis Paired GANs outperform unpaired GANs and diffusion methods, achieving higher values across all three metrics (Tab. 5). Qualitative assessment (Fig. 6) shows poor translations from unpaired techniques, with most retaining geometric artifacts unique to RGB (ocean waves and shadows) and ignoring relative temperature characteristics. Paired GANs produce results similar to real thermal images but with inconsistent details upon closer inspection. Likewise, the diffusion method produces accurate translations but is not consistent. Overall, RGB-T translation still requires improvements before being able to provide a reliable source of thermal training data from existing RGB datasets. For domain adaptation purposes, we emphasize that maximizing the thermal mIoU score is more important than photometric consistency.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Motion Tracking", + "text_level": 1, + "bbox": [ + 214, + 321, + 401, + 338 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To quantify VIO/SLAM robustness, we select 12 clipped sequences from our dataset ranging in motion tracking difficulty: from urban sports fields (easy) to our natural environments (hard). We normalize thermal images using the $5^{\\text{th}}$ and $95^{\\text{th}}$ percentile pixel values (Sec. 3.2) to enable feature detection.", + "bbox": [ + 212, + 342, + 785, + 402 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/26ad600a345e93697527d5a02c6219e5a65f4e2c4035340411ddd0408f5c6240.jpg", + "table_caption": [ + "Table 6: VIO/SLAM performance (Absolute Trajectory Error [m]) on aerial sequences" + ], + "table_footnote": [], + "table_body": "
MethodTypeModalityNorth FieldCastaic Lake, CADuck, NC
N1N2N3N4N5N6C1C2C3C4D1D2
VINS-FusionSLAMRGB2.8837.5304.0362.0094.0450.8341.555—3.2771.5661.0310.700
VINS-FusionThermal9.45412.269.8542.92611.005.1147.2961.5135.3212.0250.879—
VINS-FusionVIORGB5.4589.5096.5182.1216.0261.2771.555—3.2771.3771.0310.725
VINS-FusionThermal13.8212.8611.992.92613.985.1947.28410.605.3212.0340.879—
Open-VINSVIORGB22.8036.1537.155.165—3.896—3.573——0.4750.700
Open-VINSThermal14.4330.6416.122.258—1.562—5.311——1.073—
Trajectory length (m)1533132612531531104642310651741378880
", + "bbox": [ + 215, + 436, + 784, + 532 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Baselines We evaluate VINS-Fusion [54] (graph optimization-based) and OpenVINS [21] (filtering-based). We test VINS-Fusion in VIO and SLAM modes, with and without loop closure constraints, respectively, and report Absolute Trajectory Error [64] averaged over four runs.", + "bbox": [ + 212, + 547, + 785, + 607 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Performance analysis This benchmark reveals challenges due to fast motion at altitude and periodic feature sparsity in littoral environments. In such scenes, motion tracking frequently failed due to loss of features, requiring sequences to be clipped for quantifiable evaluations (Tab. 6). VINS-Fusion outperformed Open-VINS mainly due to higher feature tracking reliability. VINS-Fusion works better on RGB images than on thermal images in general. However, VINS-Fusion demonstrates superior feature tracking on thermal images compared to RGB on the C2 sequence of Castaic Lake, where the thermal contrast of stone grains on the lake coast was particularly pronounced in the late afternoon. This observation motivates future research on integrating thermal and RGB cameras for motion tracking with enhanced robustness and versatility. Our benchmark presents a difficult challenge for VIO and SLAM algorithms in extreme environments where water and reflections off water surfaces dominate image scenes, providing a unique test bed to invigorate interest in robust feature matching and motion tracking in natural texture-deficient scenarios.", + "bbox": [ + 212, + 613, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "C. Lee et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 7: RGB foundation model performances on thermal imagery", + "bbox": [ + 274, + 151, + 727, + 166 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/0288b5b7fb3c8b43a407e02c3ecb88b593d9a0718cec210ea880fb8c7a18d00d.jpg", + "table_caption": [ + "(a) Instance segmentation via SAM on thermal and augmented color images using coregistered color image segmentations as ground truth" + ], + "table_footnote": [ + "$\\dagger$ refers to AP@0.5::0.95, following MSCOCO [40]" + ], + "table_body": "
ModalityAPallAPsmallAPmedAPlarge
Thermal0.0180.0070.0220.204
Color (grayscale)0.5380.5360.5310.601
Color (col. jitter)0.8000.7920.8010.841
", + "bbox": [ + 215, + 202, + 495, + 257 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b02135979860aa337e2ccd24a9a2591c7fbdbe157d3f595dc3357180f9bd6fdc.jpg", + "table_caption": [ + "(b) Zero-shot semantic segmentation (mIoU) on coregistered color and thermal imagery" + ], + "table_footnote": [ + "Most frequent ground truth class label in a SAM mask assigned as label for entire mask" + ], + "table_body": "
MethodColorColor (gray)Color (jitter)Thermal
SAM [33] + GT ann.‡0.7040.6980.7000.653
Grounded SAM [43]0.3800.3510.3610.193
OV-Seg [38]0.3620.3400.3650.217
ODISE [77]0.5040.4500.4860.232
", + "bbox": [ + 500, + 195, + 779, + 257 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Further Analysis", + "text_level": 1, + "bbox": [ + 215, + 295, + 401, + 310 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Zero-shot foundation models on thermal imagery Current foundation models work well on RGB imagery but have not been benchmarked on thermal, especially in non-urban settings. To evaluate these models on thermal imagery, specifically for zero-shot segmentation and semantic segmentation, we compare RGB and thermal performance using the aligned set (Sec. 3.2) in order to isolate modality as the root cause for performance differences. In the following tests, our metrics do not include the rare object classes (vehicle, person) as they drop metrics to near-uniform values for both modalities, making it difficult to compare.", + "bbox": [ + 212, + 313, + 787, + 434 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d9b6b21b6e513c5b08fe3a5be09c135904f6fb26b560d0765750b59e2ec7805b.jpg", + "image_caption": [ + "Fig. 7: Zero-shot foundation models on RGB-T pairs. Oversegmented SAM outputs differ across modalities but are similar with ground truth semantics added. Semantic segmentation models (prompted with all classes) perform better on RGB imagery." + ], + "image_footnote": [], + "bbox": [ + 235, + 455, + 769, + 625 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We quantify Segment Anything's (SAM) [33] zero-shot segmentation sensitivity to modality change by measuring the spatial alignment of its mask outputs. We apply it to the RGB set and the 8-bit thermal image set (Sec. 3.2) with the default $32 \\times 32$ grid points to segment everything (Fig. 7). We compute the average precision (AP) of the predicted thermal masks, using the predicted RGB masks as ground truth, and compare the thermal AP to the APs of a color-jittered RGB set and a grayscale set. We find that SAM is sensitive to photometric changes, with large performance drops from color-jitter to grayscale to thermal (Tab. 7a).", + "bbox": [ + 212, + 688, + 787, + 809 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Next, we test large vision-language models (Grounded-SAM [43], OV-Seg [38], and ODISE [77]) in zero-shot semantic segmentation. We prompt the models", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Caltech Aerial RGB-Thermal Dataset in the Wild", + "bbox": [ + 397, + 114, + 732, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "with a list of classes in our dataset; generate semantic segmentation masks for RGB, color-jittered RGB, grayscale, and thermal image sets; and compute the mIoU of their outputs using our ground truth annotations. Although RGB, RGB-variants, and thermal modalities all yield poor semantic segmentation performance (Tab. 7b), all RGB variants outperform thermal by at least $0.13\\mathrm{mIoU}$ indicating low generalization to the thermal modality (Fig. 7).", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Finally, we create semantic SAM masks by assigning class labels to SAM instances based on the most frequent ground truth class within an instance (Fig. 7). We find that these masks attain much higher mIoU score (0.653) compared to the other zero-shot methods. Based on the results (Tab. 7), we note that SAM can find object boundaries in the thermal domain reasonably well but can oversegment, resulting in poor AP scores when compared to RGB (which can also oversegment). When taken with ground truth semantics, this behavior is hidden.", + "bbox": [ + 212, + 237, + 787, + 343 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Transfer learning with urban thermal datasets We test if pretraining on thermal urban data can assist transfer to field settings. We sample 40k unlabeled thermal images from [11, 17, 22, 41, 61, 67, 71, 76] and pretrain various networks using Mo", + "bbox": [ + 212, + 347, + 486, + 450 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "CoV2 [9]. We then finetune the pretrained networks on our training set and evaluate on our test set. Overall, we find sparse evidence suggesting any advantage of pretraining with existing urban thermal datasets (Tab. 8). Off-the-shelf ImageNet weights can provide strong performance in the thermal field domain, especially over training from scratch, without the effort and costs of pretraining.", + "bbox": [ + 212, + 452, + 787, + 529 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/10532871c4a315a3608d20adde7c7b71cb612806eead0ad81fa79f95aab7425d.jpg", + "table_caption": [ + "Table 8: Network pretraining methods for downstream thermal segmentation." + ], + "table_footnote": [], + "table_body": "
ModelmIoU# Params. (M)
NoneImageNetTherm. Urban
FastSCNN0.6400.6900.6881.1
EfficientViT0.6870.7250.7144.8
ResNet180.7020.7130.70612.3
ResNet500.6820.7130.72826.7
ConvNext-B0.6250.7170.69789.4
", + "bbox": [ + 498, + 378, + 784, + 441 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 537, + 359, + 555 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We presented the Caltech Aerial RGB-Thermal Dataset, the first dataset tailored towards advancing thermal semantic perception and motion tracking algorithms in natural environments. We created four benchmarks encompassing semantic segmentation, image translation, and motion tracking, and demonstrate challenges faced by current thermal perception and localization algorithms. Semantic segmentation and image translation methods were notably affected by geographical domain shifts, reflecting the wide intra-class and inter-scenery distributions in our dataset, and temporal domain shifts, due to thermal inversion and crossover. Motion tracking in thermal suffered from feature sparsity, a challenge commonly encountered in real-world robotic deployments [1]. Our motion tracking benchmark penalizes current VIO/SLAM algorithms for assuming ideal feature-tracking conditions, serving as a testbed to fuel improvements in this area. Our dataset and benchmarks can be used into the future to help develop algorithms that expand the operating domains of robotics and computer vision.", + "bbox": [ + 212, + 563, + 787, + 776 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 214, + 785, + 403, + 801 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was funded by the Office of Naval Research. We thank K. Koetje, K. Brodie, T. Hesser, T. Almeida, T. Cook, N. Spore, and R. Beach.", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "C. Lee et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ackerman, E.: Blade strike on landing ends mars helicopter's epic journey (Jan 2024), https://spectrum.ieee.org/mars-helicopter-ingenuity-end-mission", + "2. Bloesch, M., Burri, M., Omari, S., Hutter, M., Siegwart, R.: Iterated extended kalman filter based visual-inertial odometry using direct photometric feedback. The International Journal of Robotics Research 36(10), 1053-1072 (2017)", + "3. Bondi, E., Jain, R., Aggrawal, P., Anand, S., Hannaford, R., Kapoor, A., Piavis, J., Shah, S., Joppa, L., Dilkina, B., et al.: Birdsai: A dataset for detection and tracking in aerial thermal infrared videos. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1747-1756 (2020)", + "4. Brodie, K.L., Bruder, B.L., Slocum, R.K., Spore, N.J.: Simultaneous mapping of coastal topography and bathymetry from a lightweight multicamera uas. IEEE Transactions on Geoscience and Remote Sensing 57(9), 6844-6864 (2019). https://doi.org/10.1109/TGRS.2019.2909026", + "5. Cai, H., Gan, C., Han, S.: Efficientvit: Enhanced linear attention for high-resolution low-computation visual recognition. arXiv preprint arXiv:2205.14756 (2022)", + "6. Cai, W., Jin, K., Hou, J., Guo, C., Wu, L., Yang, W.: Vdd: Varied drone dataset for semantic segmentation. arXiv preprint arXiv:2305.13608 (2023)", + "7. Center, A.F.S.: A dataset for machine learning algorithm development from 2010-06-15 to 2010-08-15. NOAA National Centers for Environmental Information (2019)", + "8. Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 801-818 (2018)", + "9. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)", + "0. Chen, Y., Wang, Y., Lu, P., Chen, Y., Wang, G.: Large-scale structure from motion with semantic constraints of aerial images. In: Pattern Recognition and Computer Vision: First Chinese Conference, PRCV 2018, Guangzhou, China, November 23-26, 2018, Proceedings, Part I 1. pp. 347-359. Springer (2018)", + "1. Choi, Y., Kim, N., Hwang, S., Park, K., Yoon, J.S., An, K., Kweon, I.S.: Kaist multi-spectral day/night data set for autonomous and assisted driving. IEEE Transactions on Intelligent Transportation Systems 19(3), 934-948 (2018)", + "2. Deevi, S.A., Lee, C., Gan, L., Nagesh, S., Pandey, G., Chung, S.J.: Rgb-x object detection via scene-specific fusion modules. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 7366-7375 (2024)", + "3. Delaune, J., Hewitt, R., Lytle, L., Sorice, C., Thakker, R., Matthies, L.: Thermal-inertial odometry for autonomous flight throughout the night. In: 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1122-1128. IEEE (2019)", + "4. Doer, C., Trommer, G.F.: Radar visual inertial odometry and radar thermal inertial odometry: Robust navigation even in challenging visual conditions. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 331-338. IEEE (2021)", + "5. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873-12883 (2021)", + "6. Feng, D., Haase-Schütz, C., Rosenbaum, L., Hertlein, H., Glaeser, C., Timm, F., Wiesbeck, W., Dietmayer, K.: Deep multi-modal object detection and semantic" + ], + "bbox": [ + 225, + 169, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Caltech Aerial RGB-Thermal Dataset in the Wild", + "bbox": [ + 397, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems 22(3), 1341-1360 (2020)", + "17. Teledyne fliradas dataset,https://www.flir.com/oem/adas/adas-dataset-form/,Lastaccessed on2023-10-27", + "18. Fountas, S., Mylonas, N., Malounas, I., Rodias, E., Hellmann Santos, C., Pekkeriet, E.: Agricultural robotics for field operations. Sensors 20(9), 2672 (2020)", + "19. Gade, R., Moeslund, T.B.: Thermal cameras and applications: a survey. Machine vision and applications 25, 245-262 (2014)", + "20. Gan, L., Lee, C., Chung, S.J.: Unsupervised rgb-to-thermal domain adaptation via multi-domain attention network. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 6014-6020. IEEE (2023)", + "21. Geneva, P., Eckenhoff, K., Lee, W., Yang, Y., Huang, G.: Openvins: A research platform for visual-inertial estimation. In: 2020 IEEE International Conference on Robotics and Automation (ICRA). pp. 4666-4672. IEEE (2020)", + "22. Ha, Q., Watanabe, K., Karasawa, T., Ushiku, Y., Harada, T.: Mfnet: Towards real-time semantic segmentation for autonomous vehicles with multi-spectral scenes. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 5108-5115. IEEE (2017)", + "23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "24. He, M., Rajkumar, R.R.: Using thermal vision for extended vins-mono to localize vehicles in large-scale outdoor road environments. In: 2021 IEEE Intelligent Vehicles Symposium (IV). pp. 953-960. IEEE (2021)", + "25. Howard, A., Sandler, M., Chu, G., Chen, L.C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., et al.: Searching for mobilenetv3. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1314-1324 (2019)", + "26. Hua, T., Pei, L., Li, T., Wu, Q., Wang, R., Yu, W.: I2-slam: Fusing infrared camera and imu for simultaneous localization and mapping. In: International Conference on Autonomous Unmanned Systems. pp. 2834-2844. Springer (2021)", + "27. Huang, X., Liu, M.Y., Belongie, S., Kautz, J.: Multimodal unsupervised image-to-image translation. In: European Conference on Computer Vision (ECCV) (2018)", + "28. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1125-1134 (2017)", + "29. Jia, X., Zhu, C., Li, M., Tang, W., Zhou, W.: Llvip: A visible-infrared paired dataset for low-light vision. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3496-3504 (2021)", + "30. Jong, A., Yu, M., Dhrafani, D., Kailas, S., Moon, B., Sycara, K., Scherer, S.: Wit-uas: A wildland-fire infrared thermal dataset to detect crew assets from aerial views. In: 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 11464-11471. IEEE (2023)", + "31. Khattak, S., Papachristos, C., Alexis, K.: Keyframe-based direct thermal-inertial odometry. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 3563-3569. IEEE (2019)", + "32. Kim, Y.H., Shin, U., Park, J., Kweon, I.S.: Ms-uda: Multi-spectral unsupervised domain adaptation for thermal image semantic segmentation. IEEE Robotics and Automation Letters 6(4), 6497–6504 (Oct 2021). https://doi.org/10.1109/LRA.2021.3093652" + ], + "bbox": [ + 215, + 147, + 785, + 837 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "C. Lee et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)", + "34. Lee, C., Frennert, J.G., Gan, L., Anderson, M., Chung, S.J.: Online self-supervised thermal water segmentation for aerial vehicles. In: 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 7734-7741. IEEE (2023)", + "35. Lee, D.G., Jeon, M.H., Cho, Y., Kim, A.: Edge-guided multi-domain rgb-to-tir image translation for training vision tasks with challenging labels. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 8291-8298. IEEE (2023)", + "36. Li, C., Xia, W., Yan, Y., Luo, B., Tang, J.: Segmenting objects in day and night: Edge-conditioned cnn for thermal image semantic segmentation. IEEE Transactions on Neural Networks and Learning Systems 32(7), 3069-3082 (2020)", + "37. Li, H., Ma, Y., Huang, Y., Gu, Y., Xu, W., Liu, Y., Zuo, X.: Riders: Radar-infrared depth estimation for robust sensing. arXiv preprint arXiv:2402.02067 (2024)", + "38. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7061-7070 (2023)", + "39. Liang, M., Hu, J., Bao, C., Feng, H., Deng, F., Lam, T.L.: Explicit attention-enhanced fusion for rgb-thermal perception tasks. IEEE Robotics and Automation Letters (2023)", + "40. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "41. Liu, J., Fan, X., Huang, Z., Wu, G., Liu, R., Zhong, W., Luo, Z.: Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5802-5811 (2022)", + "42. Liu, M.Y., Breuel, T., Kautz, J.: Unsupervised image-to-image translation networks. In: Conference on Neural Information Processing Systems (NeurIPS) (2017)", + "43. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)", + "44. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022)", + "45. Lyu, Y., Vosselman, G., Xia, G.S., Yilmaz, A., Yang, M.Y.: Uavid: A semantic segmentation dataset for uav imagery. ISPRS Journal of Photogrammetry and Remote Sensing 165, 108 - 119 (2020). https://doi.org/https://doi.org/10.1016/j.isprsjprs.2020.05.009, http://www.sciencedirect.com/science/article/pii/S0924271620301295", + "46. Mehta, S., Rastegari, M.: Separable self-attention for mobile vision transformers. arXiv preprint arXiv:2206.02680 (2022)", + "47. Mostegel, C., Maurer, M., Heran, N., Pestana Puerta, J., Fraundorfer, F.: Semantic drone dataset (Jan 2019), http://dronedataset.icg.tugraz.at/, Last accessed on 2023-10-27", + "48. Nigam, I., Huang, C., Ramanan, D.: Ensemble knowledge transfer for semantic segmentation. In: 2018 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 1499-1508. IEEE (2018)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Caltech Aerial RGB-Thermal Dataset in the Wild", + "bbox": [ + 397, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. Nirgudkar, S., DeFilippo, M., Sacarny, M., Benjamin, M., Robinette, P.: Massmind: Massachusetts maritime infrared dataset. The International Journal of Robotics Research 42(1-2), 21-32 (2023)", + "50. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Howes, R., Huang, P.Y., Xu, H., Sharma, V., Li, S.W., Galuba, W., Rabbat, M., Assran, M., Ballas, N., Synnaeve, G., Misra, I., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2023)", + "51. Panetta, K., Shreyas Kamath, K.M., Rajeev, S., Agaian, S.S.: Ftnet: Feature transverse network for thermal image semantic segmentation. IEEE Access 9, 145212-145227 (2021). https://doi.org/10.1109/ACCESS.2021.3123066", + "52. Park, T., Efros, A.A., Zhang, R., Zhu, J.Y.: Contrastive learning for unpaired image-to-image translation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16. pp. 319-345. Springer (2020)", + "53. Poudel, R., Liwicki, S., Cipolla, R.: Fast-scnn: Fast semantic segmentation network. In: Sidorov, K., Hicks, Y. (eds.) Proceedings of the British Machine Vision Conference (BMVC). pp. 187.1-187.12. BMVA Press (September 2019). https://doi.org/10.5244/C.33.187, https://dx.doi.org/10.5244/C.33.187", + "54. Qin, T., Li, P., Shen, S.: Vins-mono: A robust and versatile monocular visual-inertial state estimator. IEEE Transactions on Robotics 34(4), 1004-1020 (2018)", + "55. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "56. Retief, S.P., Willers, C., Wheeler, M.: Prediction of thermal crossover based on imaging measurements over the diurnal cycle. In: Geo-Spatial and Temporal Image and Data Exploitation III. vol. 5097, pp. 58-69. SPIE (2003)", + "57. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022)", + "58. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022)", + "59. Saputra, M.R.U., Lu, C.X., de Gusmao, P.P.B., Wang, B., Markham, A., Trigoni, N.: Graph-based thermal-inertial slam with probabilistic neural networks. IEEE Transactions on Robotics 38(3), 1875-1893 (2022)", + "60. Shin, U., Lee, K., Kweon, I.S.: Complementary random masking for rgb-thermal semantic segmentation. In: IEEE International Conference on Robotics and Automation (2024)", + "61. Shin, U., Park, J., Kweon, I.S.: Deep depth estimation from thermal image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1043-1053 (2023)", + "62. Shivakumar, S.S., Rodrigues, N., Zhou, A., Miller, I.D., Kumar, V., Taylor, C.J.: Pst900: Rgb-thermal calibration, dataset and segmentation network. In: 2020 IEEE international conference on robotics and automation (ICRA). pp. 9441-9447. IEEE (2020)", + "63. Speth, S., Goncalves, A., Rigault, B., Suzuki, S., Bouazizi, M., Matsuo, Y., Prendinger, H.: Deep learning with rgb and thermal images onboard a drone for monitoring operations. Journal of Field Robotics 39(6), 840-868 (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "C. Lee et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "64. Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for the evaluation of rgb-d slam systems. In: 2012 IEEE/RSJ international conference on intelligent robots and systems. pp. 573-580. IEEE (2012)", + "65. Sun, Y., Zuo, W., Liu, M.: Rtfnet: RGB-thermal fusion network for semantic segmentation of urban scenes. IEEE Robotics and Automation Letters 4(3), 2576-2583 (2019). https://doi.org/10.1109/LRA.2019.2904733", + "66. Sun, Y., Zuo, W., Yun, P., Wang, H., Liu, M.: Fuseseg: Semantic segmentation of urban scenes based on rgb and thermal data fusion. IEEE Transactions on Automation Science and Engineering 18(3), 1000-1011 (2021). https://doi.org/10.1109/TASE.2020.2993143", + "67. Suo, J., Wang, T., Zhang, X., Chen, H., Zhou, W., Shi, W.: Hit-uav: A high-altitude infrared thermal dataset for unmanned aerial vehicle-based object detection. Scientific Data 10, 227 (2023)", + "68. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105-6114. PMLR (2019)", + "69. Ustun, B., Kaya, A.K., Ayerden, E.C., Altinel, F.: Spectral transfer guided active domain adaptation for thermal imagery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. pp. 449-458 (June 2023)", + "70. Vargas, J., Alsweiss, S., Token, O., Razdan, R., Santos, J.: An overview of autonomous vehicles sensors and their vulnerability to weather conditions. Sensors 21(16), 5397 (2021)", + "71. Vertens, J., Zürn, J., Burgard, W.: Heatnet: Bridging the day-night domain gap in semantic segmentation with thermal images. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8461-8468. IEEE (2020)", + "72. VS, V., Poster, D., You, S., Hu, S., Patel, V.M.: Meta-uda: Unsupervised domain adaptive thermal object detection using meta-learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1412-1423 (January 2022)", + "73. Wang, T.C., Liu, M.Y., Zhu, J.Y., Tao, A., Kautz, J., Catanzaro, B.: High-resolution image synthesis and semantic manipulation with conditional gans. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2018)", + "74. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. In: Neural Information Processing Systems (NeurIPS) (2021)", + "75. Xie, S., Girshick, R., Dollar, P., Tu, Z., He, K.: Aggregated residual transformations for deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1492-1500 (2017)", + "76. Xiong, H., Cai, W., Liu, Q.: Mcnet: Multi-level correction network for thermal image semantic segmentation of nighttime driving scene. Infrared Physics & Technology p. 103628 (2021). https://doi.org/https://doi.org/10.1016/j.infrared.2020.103628", + "77. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2955-2966 (2023)", + "78. Yun, S., Jung, M., Kim, J., Jung, S., Cho, Y., Jeon, M.H., Kim, G., Kim, A.: Stereo: Stereo thermal dataset for research in odometry and mapping. In: 2022" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Caltech Aerial RGB-Thermal Dataset in the Wild", + "bbox": [ + 397, + 114, + 730, + 127 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 3857-3864. IEEE (2022)", + "79. Zhang, J., Liu, R., Shi, H., Yang, K., Reif, S., Peng, K., Fu, H., Wang, K., Stiefelhagen, R.: Delivering arbitrary-modal semantic segmentation. In: CVPR (2023)", + "80. Zhao, S., Wang, P., Zhang, H., Fang, Z., Scherer, S.: Tp-tio: A robust thermal-inertial odometry with deep thermalpoint. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4505-4512. IEEE (2020)", + "81. Zhou, W., Liu, J., Lei, J., Yu, L., Hwang, J.N.: Gmnet: Graded-feature multilabel-learning network for rgb-thermal urban scene semantic segmentation. IEEE Transactions on Image Processing 30, 7790–7802 (2021). https://doi.org/10.1109/TIP.2021.3109518", + "82. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Proceedings of the IEEE international conference on computer vision. pp. 2223-2232 (2017)", + "83. Zhu, P., Wen, L., Du, D., Bian, X., Fan, H., Hu, Q., Ling, H.: Detection and tracking meet drones challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(11), 7380-7399 (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 383 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "C. Lee et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/Caltech Aerial RGB-Thermal Dataset in the Wild/377537ff-2360-4d17-b986-8337083e74f6_model.json b/2024/Caltech Aerial RGB-Thermal Dataset in the Wild/377537ff-2360-4d17-b986-8337083e74f6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..23434c6c419553004bbe3e865116ea50c5066f32 --- /dev/null +++ b/2024/Caltech Aerial RGB-Thermal Dataset in the Wild/377537ff-2360-4d17-b986-8337083e74f6_model.json @@ -0,0 +1,3045 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.142, + 0.787, + 0.162 + ], + "angle": 0, + "content": "Caltech Aerial RGB-Thermal Dataset in the Wild" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.19, + 0.758, + 0.222 + ], + "angle": 0, + "content": "Connor Lee*, Matthew Anderson*, Nikhil Ranganathan, Xingxing Zuo, Kevin Do, Georgia Gkioxari, and Soon-Jo Chung" + }, + { + "type": "text", + "bbox": [ + 0.388, + 0.233, + 0.616, + 0.247 + ], + "angle": 0, + "content": "California Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.248, + 0.75, + 0.261 + ], + "angle": 0, + "content": "{cee, matta, nrangana, zuox, kdo, georgia, sjchung}@caltech.edu" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.295, + 0.745, + 0.49 + ], + "angle": 0, + "content": "Abstract. We present the first publicly-available RGB-thermal dataset designed for aerial robotics operating in natural environments. Our dataset captures a variety of terrain across the United States, including rivers, lakes, coastlines, deserts, and forests, and consists of synchronized RGB, thermal, global positioning, and inertial data. We provide semantic segmentation annotations for 10 classes commonly encountered in natural settings in order to drive the development of perception algorithms robust to adverse weather and nighttime conditions. Using this dataset, we propose new and challenging benchmarks for thermal and RGB-thermal (RGB-T) semantic segmentation, RGB-T image translation, and motion tracking. We present extensive results using state-of-the-art methods and highlight the challenges posed by temporal and geographical domain shifts in our data. The dataset and accompanying code is available at https://github.com/aerorobotics/caltech-aerial-rgbt-dataset." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.503, + 0.741, + 0.531 + ], + "angle": 0, + "content": "Keywords: Robotics \\(\\cdot\\) Semantic segmentation \\(\\cdot\\) RGB-T image translation \\(\\cdot\\) Visual inertial odometry" + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.564, + 0.371, + 0.576 + ], + "angle": 0, + "content": "Thermal Dataset Comparisons" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.577, + 0.367, + 0.689 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.565, + 0.603, + 0.576 + ], + "angle": 0, + "content": "A. Thermal + RGB-T Semantic Segmentation" + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.577, + 0.626, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.429, + 0.617, + 0.499, + 0.623 + ], + "angle": 0, + "content": "Aligned RGB-T pairs" + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.628, + 0.567, + 0.639 + ], + "angle": 0, + "content": "B. RGB-to-Thermal Image Translation" + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.642, + 0.586, + 0.69 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.565, + 0.745, + 0.576 + ], + "angle": 0, + "content": "C. Motion Tracking" + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.583, + 0.787, + 0.688 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.694, + 0.788, + 0.737 + ], + "angle": 0, + "content": "Fig. 1: Left: Our dataset is uniquely designed to improve thermal scene perception for field robots. Right: We provide new benchmarks for thermal-based vision algorithms, including (a) semantic segmentation, (b) image translation, and (c) motion tracking." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.762, + 0.376, + 0.777 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.786, + 0.787, + 0.817 + ], + "angle": 0, + "content": "Field robots rely predominantly on visual cameras, lidar, and radar to perceive their surroundings [16, 18]. These sensors provide robust perception capabilities," + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.551, + 0.84 + ], + "angle": 0, + "content": "* These authors contributed equally to this work." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "C. Lee et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.266 + ], + "angle": 0, + "content": "but falter in low-light and adverse weather conditions [70]. In contrast, thermal cameras use long-wave infrared wavelengths to capture emitted heat, offering dense, visual information in such conditions where other sensors struggle [19]. Recently, they have been used for multimodal perception in autonomous driving and are increasingly being used to enable nighttime robot autonomy [11, 12, 36, 37, 49, 71, 76, 78]. However, successful integration of thermal cameras requires extensive datasets across diverse settings, preventing widespread adoption in field robotics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.788, + 0.389 + ], + "angle": 0, + "content": "Current thermal datasets focus on urban settings for autonomous driving applications (Fig. 1). They comprise of thermal-only imagery [36,76] or RGB-Thermal (RGB-T) pairs [17,22,41,71], with some including global positioning (GPS/GNSS) and inertial measurements (IMU) for visual-inertial odometry (VIO) and simultaneous localization and mapping (SLAM) [11,61,78]. While comprehensive, they lack data depicting natural settings like rivers and forests, which are typical operating areas for field robotics that perform coastline mapping and bathymetry [4] or activity monitoring during forest fires [30]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.788, + 0.556 + ], + "angle": 0, + "content": "Due to the lack of thermal benchmarks for algorithms like semantic segmentation and SLAM, field robots cannot easily operate at night or in adverse conditions, requiring online learning [34, 72] or unsupervised domain adaptation to compensate [20, 32, 71], which still require labeled data for evaluation. As such, enabling thermal perception requires collecting and annotating field-specific data from scratch. However, creating a thermal dataset for field robotics is difficult due to two reasons: First, thermal data is hard to crowdsourced and web-scrape due to the richness of thermal cameras. Hence, acquiring a diverse dataset for algorithm development requires physically going to distinct field locations, increasing time and cost. Second, operating field robots, especially uninhabited aerial vehicles (UAV), usually requires special permits for each location-of-capture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.558, + 0.788, + 0.707 + ], + "angle": 0, + "content": "In this work, we present the first dataset targeting thermal scene perception for field robotics. The dataset consists of oblique-facing imagery captured from a UAV, supplemented with image sets captured at ground level, and focuses on littoral settings within various desert, forest, and coastal environments across the United States. We contribute new benchmarks for thermal and RGB-T semantic segmentation, RGB-T image translation, and VIO/SLAM algorithms, with unique challenges characterized by temporal and geographical domain shifts for learning-based methods and feature sparsity for motion-tracking algorithms. Aside from the proposed benchmarks, our dataset can also be used towards other applications like visual place recognition and terrain-relative navigation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.709, + 0.787, + 0.754 + ], + "angle": 0, + "content": "The paper is organized as follows: Sec. 2 reviews relevant datasets, benchmarks, and algorithms. Sec. 3 describes our dataset details and curation process, and Sec. 4 presents benchmark results. Sec. 5 offers concluding remarks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.765, + 0.39, + 0.781 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.792, + 0.478, + 0.806 + ], + "angle": 0, + "content": "2.1 Datasets and Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Thermal/RGB-T benchmarks for urban robotics Current urban thermal and RGB-T datasets primarily focus on autonomous vehicle (AV) technology" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.398, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Caltech Aerial RGB-Thermal Dataset in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.152, + 0.785, + 0.193 + ], + "angle": 0, + "content": "Table 1: Comparison of datasets captured from aerial platforms. Camera angle of \\(90^{\\circ}\\) is nadir-pointing. See Tab. S1 in supplement for comparison against thermal datasets captured from surface and ground vehicles." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.194, + 0.785, + 0.289 + ], + "angle": 0, + "content": "
DatasetTaskRGBThermalGPSIMUSettingLocation# SamplesCamera pose
VisDrone [83]Obj. Det.✓XXXUrbanChina10,209Variable
HIT-UAV [67], NII-CU [63]Obj. Det.✓✓XXUrbanChina, Japan2,898 - 5,88030°, 90°, 60-130 m
WTT-UAS [30]Obj. Det.X✓✓✓ForestUSA6,95130°, 90°, < 120 m
BIRDSAI [3]Obj. Track.X✓XXSavannasS. Africa62,000Off-nadir, 60-120 m
AeroScapes [48], UAVid [45], [10], [6], Swiss/Okutama [63]Sem. Seg.✓XXXUrbanChina, Japan, Switzerland191 - 326930 - 90°, < 100 m
Semantic Drone Dataset [47]Sem. Seg.✓✓X✓UrbanGermany40090°, < 30m
OursSem. Seg.✓✓✓✓River, Coast, Lake, Mountain, DesertUSA4,19520°, 45°, < 120 m
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.3, + 0.785, + 0.421 + ], + "angle": 0, + "content": "and surveillance applications. AV-related datasets cover tasks such as object detection [11, 17, 41], semantic segmentation [22, 36, 71, 76], and localization [11, 61, 78]. Some include data from other spectra, benchmarking not only thermal algorithms but also RGB-T [11, 22, 41, 63, 71] and other multispectral models. In contrast, datasets for surveillance applications use fixed cameras [29] and UAVs [63, 67] to detect personnel, vehicles, and other objects. Despite many urban thermal benchmarks, they cannot be directly used to develop algorithms for robots operating in natural environments due to the domain gap." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.788, + 0.683 + ], + "angle": 0, + "content": "Thermal/RGB-T benchmarks for field robotics Datasets for robots operating in non-urban environments remain limited. MassMIND [49] and PST900 [62] benchmark thermal semantic segmentation for maritime and subterranean robotic environments, respectively. WIT-UAS [30] provides UAV-borne thermal images for object detection in wildfire-prone environments. One work [34] proposes an aerial RGB-T dataset for thermal water segmentation in littoral areas but only provides 1272 labels consisting of a single water class. Benchmarks like [3,7] depict natural environments but are targeted towards wildlife tracking. To address this gap, we present an aerial RGB-T dataset which can be used to benchmark multiple tasks, including semantic segmentation, RGB-T image translation, and localization, across diverse natural environment robotic applications. Specifically, we build upon the dataset from [34] by extending the number of semantic classes from 1 to 10 and adding additional RGB-T data captured from air and ground, resulting in a total of 4195 semantic segmentation annotations. In comparison to existing thermal/RGB-T datasets, our dataset can enable nighttime field robotic applications like remote sensing research, geological sampling, precision agriculture, and wilderness search and rescue." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.781 + ], + "angle": 0, + "content": "UAV-based semantic segmentation datasets Aside from [34], existing UAV-captured semantic segmentation datasets [6, 10, 45, 47, 48, 63] cover urban environments (Tab. 1). With the exception of [47] which provides 400 thermal image samples, all of these datasets contain only RGB data for semantic segmentation. While they provide images with relevant view angles, the modality gap and focus on urban content makes them impractical for aerial field robotic applications." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.79, + 0.557, + 0.807 + ], + "angle": 0, + "content": "2.2 Algorithms using Thermal Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Thermal semantic segmentation Since thermal images are spatially identical to their RGB counterparts, popular RGB semantic segmentation algo" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "C. Lee et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "rithms [5, 8, 53, 74] can be applied directly. However, some works [36, 51, 76] develop models specifically for thermal imagery, and leverage edge priors to overcome low resolution and blurriness due to thermal crossover [56]. Although specifically designed for thermal, these models only marginally outperform standard RGB segmentation models [51] and their performance in field settings, where edges are less structured, remains untested. Aside from architectural advancements, improvements in thermal semantic segmentation come from domain adaptation (DA) methods [20, 36, 69, 71] that use labeled RGB data with unlabeled thermal data. To provide fair comparisons, we do not provide DA baselines in our benchmarks as performance varies based on the choice of RGB data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.302, + 0.79, + 0.455 + ], + "angle": 0, + "content": "RGB-T semantic segmentation RGB-T semantic segmentation methods compensate for the weaknesses of each modality via deep feature fusion, data augmentation, and adversarial training. Most approaches [22, 60, 65, 66, 79] utilize encoder-decoder architectures with modality-specific encoders and shared decoders, and show significant improvements over channel-stacked RGB-T inputs. Works like [39, 81] make improvements by integrating RGB and thermal features at multiple stages within the encoders, while [62] eschews RGB-T training pairs by processing RGB-T inputs sequentially. Other methods include random input masking to reduce reliance on a single modality [60] and adversarial training to adapt to different times of day [71]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.562 + ], + "angle": 0, + "content": "RGB-T image translation RGB-T image translation is vital as it may enable thermal training data to be generated from large-scale RGB datasets. Works based on generative adversarial networks (GAN) such as [27,35,42,52,82] can be used without RGB-T image pairs. However, GANs like [28,73] that require paired imagery provide better translations for both visual appeal and domain adaptation [36]. Recent diffusion methods [57,58] have demonstrated remarkable results on RGB imagery, but have not been tested on RGB-T image translation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.688 + ], + "angle": 0, + "content": "Thermal VIO and SLAM Low signal-to-noise ratio, reduced contrast, and blurred boundaries in thermal images present challenges for VIO and SLAM. Several studies [2, 13, 14, 24, 26] use indirect methods by tracking keypoints extracted from thermal images for visual constraints and rely on image normalization techniques to improve keypoint extraction. In contrast, [31] uses a direct method to extract and track high-gradient points by minimizing radiometric error. Other works utilize learned local and global features to improve feature tracking [80] and loop closure detection [59], respectively." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.697, + 0.664, + 0.714 + ], + "angle": 0, + "content": "3 The Caltech Aerial RGB-Thermal Dataset" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.724, + 0.491, + 0.739 + ], + "angle": 0, + "content": "3.1 Data Acquisition Hardware" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.741, + 0.788, + 0.817 + ], + "angle": 0, + "content": "To capture this dataset, we developed a custom sensor stack with synchronized RGB-thermal imagery, IMU, and global positioning data (Fig. 2). The sensor stack features a non-radiometric FLIR ADK thermal camera (\\(640 \\times 512\\) px and \\(75^{\\circ}\\) horizontal FoV, \\(60\\mathrm{Hz}\\)) flanked by a pair of FLIR Blackfly electro-optical (EO) monochrome and color cameras\\(^{1}\\) (\\(960 \\times 600\\) px and \\(75^{\\circ}\\) horizontal FoV, \\(30\\mathrm{Hz}\\))." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.824, + 0.663, + 0.84 + ], + "angle": 0, + "content": "1 BFLY-U3-23S6M-C mono/color with Tamron M112FM08 lenses" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.398, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "Caltech Aerial RGB-Thermal Dataset in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "Pose information is provided by a VectorNav VN100 IMU (200 Hz) and a u-blox M8N GPS (5 Hz). All three cameras and the VN100 are hardware synchronized using a dedicated signal generator and are rigidly attached to each other using a stiff, 3D printed mount. A Simply NUC Ruby Mini PC (AMD Ryzen 7 4800U, 32 GB RAM) is used for the compute, and the entire system is interfaced using ROS Noetic. Our cameras and IMU were calibrated following the procedures found in Supplementary Material Sec. S2.2 and were not disassembled from each other between any of the dataset collections." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.277, + 0.406, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.277, + 0.777, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.475, + 0.788, + 0.503 + ], + "angle": 0, + "content": "Fig. 2: (a) The Aurelia X6 hexacopter and sensor stack used to capture our aerial dataset. (b) Geographic distribution of our data collection sites and collection times." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.517, + 0.788, + 0.623 + ], + "angle": 0, + "content": "The sensor stack is mounted rigidly onboard an Aurelia X6 hexacopter (Fig. 2) without gimbal stabilization at angles between \\(0^{\\circ}\\) (level with the horizon) to \\(45^{\\circ}\\) downwards to provide different viewpoints. When mounted to a UAV, we also log the position and attitude estimates from the UAV at \\(20\\mathrm{Hz}\\) providing improved positioning accuracy from the onboard RTK GPS. Additionally, the sensor stack's modularity allows us to capture datasets on foot by mounting it to a tripod where flight restrictions are in-place." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.633, + 0.524, + 0.649 + ], + "angle": 0, + "content": "3.2 Data Collection and Processing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.653, + 0.788, + 0.759 + ], + "angle": 0, + "content": "Data capture We captured 37 aerial and ground trajectories covering lakes, rivers, coastlines, deserts, and mountains from around Southern California, Kentucky, and North Carolina (Fig. 2). Aerial trajectories (18) involve high motion with intermittent hovering, and mostly depict scenes from \\(40\\mathrm{m}\\) altitude. Ground-level trajectories (19) were captured on foot, with 7 captured without any movement for VIO debugging purposes. For more details, see Supplementary Materials Tab. S2 and Sec. S2.1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Thermal image normalization For the baselines presented in this work, we typically normalize 16-bit thermal data by rescaling between the \\(1^{\\mathrm{st}}\\) and \\(99^{\\mathrm{th}}\\) percentile pixel values and follow with contrast limited adaptive histogram equalization (CLAHE). We keep normalized data as floats to minimize information loss, and only convert to 8-bit format when necessary or for visualization." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "C. Lee et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.144, + 0.488, + 0.156 + ], + "angle": 0, + "content": "Semantic Segmentation Classes and Description" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.158, + 0.516, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.149, + 0.539, + 0.161 + ], + "angle": 0, + "content": "a" + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.149, + 0.777, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.214, + 0.539, + 0.224 + ], + "angle": 0, + "content": "b" + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.218, + 0.765, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.295, + 0.787, + 0.337 + ], + "angle": 0, + "content": "Fig. 3: Semantic segmentation classes in our dataset. The color mapping is used throughout this paper. (a) Hourly distribution of annotated thermal images. (b) Histogram of semantic classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.785, + 0.442 + ], + "angle": 0, + "content": "Semantic segmentation annotation To aid development of scene perception algorithms for field robots, we manually-annotated a subset of thermal images for semantic segmentation with classes in Fig. 3. To avoid redundancy, images were sampled every 3s when moving and every 20s when still, resulting in 4195 samples. We considered a frame in motion if the distance moved within the past 2 seconds exceeded 0.5 meters, with distances computed using GPS coordinates." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.567 + ], + "angle": 0, + "content": "To promote labeling consistency, annotators preread descriptions of each trajectory, which contained a list of classes present, labeled examples provided by the authors, and GPS coordinates (for Google Earth look-up). Annotators were then given an RGB-T image pair with trajectory metadata and asked to label the thermal frame. Annotations were reviewed by authors, with inconsistent or incorrect labels sent back. 3 rounds of annotation were done, with the final and initial set of labels differing by \\(0.08\\mathrm{mIoU}\\). Classes most frequently re-annotated were bare ground, rocky terrain, road, trees, and vehicles." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.573, + 0.788, + 0.68 + ], + "angle": 0, + "content": "RGB-thermal image alignment To align thermal and RGB image pairs, we stereo rectified the RGB-T image pairs using the camera matrices obtained from calibration (Sec. S2.2) and projected the thermal image into the larger RGB image frame to preserve RGB resolution. As our baseline is small compared to the depth of the scenes we capture, image pairs are near-coregistered after rectification. Due to differences in optics, the field-of-view of the aligned image pairs is narrower compared to that of the thermal camera." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.69, + 0.381, + 0.706 + ], + "angle": 0, + "content": "3.3 Dataset Splits" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.712, + 0.787, + 0.759 + ], + "angle": 0, + "content": "General (benchmark) split This is our primary split for semantic segmentation and image translation benchmarks. We randomly partition the annotated thermal dataset (Sec. 3.2) into train/val/test sets at a 75:12.5:12.5 ratio." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Temporal split To promote studies into the effect of different time-of-day capture of the thermal images, we split the dataset using three time periods: twilight/sunrise (5 AM - 7 AM), daytime (10 AM - 5 PM), and nighttime (7 PM - 4 AM). We randomly partition the daytime images into train/val/test sets using the ratio from above and leave the twilight and nighttime sets for testing." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.398, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "Caltech Aerial RGB-Thermal Dataset in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "Geographical split The geographical splits are intended to test algorithm adaptability to unseen settings. We provide two splits:" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.189, + 0.784, + 0.234 + ], + "angle": 0, + "content": "1. Terrain-based: This split partitions data into the following terrain categories: aerial river, aerial lake, aerial coast, and ground-captured images. Train/val/test splits are created per category with a 50:15:35 ratio." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.235, + 0.785, + 0.28 + ], + "angle": 0, + "content": "2. Region-based: This is based on the area of the United States where the data was captured. Distinct regions include California (CA), Kentucky (KY), and North Carolina (NC), with train/val/test created in the same way as above." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.189, + 0.785, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.293, + 0.377, + 0.31 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.317, + 0.784, + 0.378 + ], + "angle": 0, + "content": "We provide baselines for the following benchmarks: thermal semantic segmentation, RGB-T semantic segmentation, RGB-T image translation, and VIO/SLAM. Furthermore, we explore the zero-shot thermal segmentation capabilities of large vision-language models on our dataset and perform further ablation studies." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.396, + 0.784, + 0.424 + ], + "angle": 0, + "content": "Table 2: Supervised model performance (mIoU), framerate (FPS at batch size 1), and floating point operations (FLOP) for various semantic segmentation network baselines." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.599 + ], + "angle": 0, + "content": "
ModelBare groundRocky terrainDevelop. struct.RoadShrubsTreesSkyWaterVehiclesPersonStuffObjectAllOrin (GPU)NUC (CPU)FLOPS (G)
FastSCNN [53]0.8070.9070.7280.6310.6940.7760.9560.9650.3840.0580.8080.2210.690124.285.61.0
MobileNet-V3-S 0.75 [25]0.8020.9140.6760.6330.7040.8140.9550.9760.3830.1110.8090.2470.69775.224.03.3
MobileNet-V3-L 0.750.8010.9140.7010.6150.7130.7930.9590.9760.4320.1480.8090.2900.70559.015.54.8
MobileViTV2 0.50 [46]0.8220.9130.7590.6500.6940.7810.9540.9700.2220.0300.8180.1260.68060.110.75.5
EfficientViT-B0 [5]0.8250.9170.7770.6180.7100.7930.9610.9650.4950.1910.8210.3430.72551.25.53.9
EfficientNet-Lite0 [68]0.8190.9080.7690.6010.7090.8050.9620.9680.3800.0790.8180.2290.70047.712.87.2
EfficientNet-Lite20.8250.9200.7650.6140.6980.8050.9580.9710.3860.1180.8200.2520.70638.911.29.4
Segformer-B0 [74]0.8040.9100.6900.6240.6960.7880.9600.9700.1950.0000.8050.0970.66438.35.910.2
Segformer-B10.8140.9090.7730.6030.7130.7990.9600.9650.3660.0000.8170.1830.69030.33.519.9
ResNet18 [23]0.8100.9050.7660.6180.7110.8070.9590.9660.4310.1580.8180.2940.71369.312.422.4
ResNet500.8190.9160.7470.6580.7110.7990.9610.9740.3610.1820.8230.2720.71329.96.145.4
ResNeXt50 [75]0.8250.9190.7730.6530.7090.7940.9640.9760.4360.1370.8270.2870.71923.05.645.5
ConvNext-T [44]0.7990.9090.7190.6080.7060.8080.9610.9700.3630.0030.8100.1830.68525.83.347.4
ConvNext-S0.8100.9130.7710.6030.7180.8100.9650.9640.4920.0480.8190.2700.70919.12.375.0
ConvNext-B0.8100.9210.6970.6110.7060.8030.9630.9720.3480.1120.8110.2300.69414.61.4130.5
ConvNext-B (CLIP) [55]0.8130.9180.6830.6320.7180.8130.9650.9720.5170.1370.8140.3270.71714.61.4130.4
ConvNext-B (CLIP) *0.7730.8870.7130.4870.6560.7560.9460.9510.1580.0190.7710.0890.63514.61.4130.4
DINOv2 [50] (linear head)0.8000.9070.6810.6060.6930.7960.9560.9670.3750.0790.8010.2270.68615.22.1—
DINOv2 (linear head) *0.7050.8440.6190.4190.5580.7250.9220.9200.3560.1580.7140.2570.62315.22.1—
DINOv2 (nonlin. head)0.8100.9160.7080.6350.7000.8190.9590.9730.3990.1190.8150.2590.70415.22.1—
DINOv2 (nonlin. head) *0.7960.9030.7320.6060.6910.7940.9540.9630.4710.1490.8050.3100.70615.22.1—
FTNet† [51]0.7550.8670.6350.5760.6430.6530.7870.9470.2340.0240.7330.1290.61311.11.4100.5
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.371, + 0.603, + 0.633, + 0.614 + ], + "angle": 0, + "content": "\\(\\dagger\\) Thermal-specific model * Frozen encoder" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.637, + 0.533, + 0.651 + ], + "angle": 0, + "content": "4.1 Thermal Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.655, + 0.784, + 0.7 + ], + "angle": 0, + "content": "We run multiple baselines on our general split before choosing a specific model to test on the geographic and temporal splits. We go over our baselines before analyzing their performance on our benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.785, + 0.794 + ], + "angle": 0, + "content": "Baselines Motivated by robotic applications, we are interested in both inference speed and segmentation performance. Our baselines place an emphasis on lightweight, real-time networks, but also on foundation models like DINOv2 [50] and ConvNext-B (CLIP) [44, 55] to explore any latent multimodal capabilities. We also test a thermal-specific model, FTNet [51], which uses an edge-based loss function to compensate for blurred boundaries in thermal images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Besides FastSCNN [53], EfficientViT [5], Segformer [74], DINOv2, and FT-Net, all network encoders in Tab. 2 feed into a DeepLabV3+ segmentation head. In particular, DINOv2 was employed with linear and non-linear, multi-scale" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "C. Lee et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "heads in order to study the generalization capacity of its pretrained features to the thermal modality. All baselines except DINOv2 and ConvNext-B (CLIP) were trained from pretrained ImageNet weights. All networks, besides FTNet, were trained using the cross entropy loss. Further details on baseline implementation and training can be found in Supplementary Material Sec. S3.1." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.233, + 0.78, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.461, + 0.788, + 0.492 + ], + "angle": 0, + "content": "Fig. 4: Thermal images and semantic segmentation labels from each capture area with inference results from EfficientViT, FastSCNN, and ConvNext-B (CLIP)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.584 + ], + "angle": 0, + "content": "General benchmark We use the general (random) split (Sec. 3.3) to compare thermal semantic segmentation performance between baselines (Tab. 2). With field robotic applications in mind, we analyze segmentation results in context of GPU and CPU frame rates, which were measured onboard the Nvidia Jetson AGX Orin (GPU) and the Ruby NUC (CPU) embedded platforms, respectively2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.788, + 0.663 + ], + "angle": 0, + "content": "In general, we found that larger models provide marginal benefit over smaller, compute-efficient models. In particular, a 0.027 mIoU difference between the largest (ConvNext-B) and fastest (FastSCNN) baselines comes at the cost of \\(130 \\times\\) more FLOPs. Overall, EfficientViT-B0 performed the best, attaining the highest mIoU (0.725) while performing \\(50\\mathrm{Hz}\\) inference on the Jetson AGX Orin." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.666, + 0.788, + 0.789 + ], + "angle": 0, + "content": "While large RGB foundation models show little gain over ImageNet pretraining, their features do extend to the thermal modality. Specifically, a frozen DINOv2 (linear head) outperforms the thermal-specific FT-Net. With a nonlinear head, it matches the performance of other end-to-end trained networks. While this shows the benefit of large-scale RGB pretraining, the choice of RGB pretraining data or learning strategy seems to matter. Notably, the frozen ConvNext-B (CLIP) model gives marginal gain over the frozen DINOv2 (linear head), despite a larger, nonlinear segmentation head. Overall, foundation models do not confer" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.788, + 0.842 + ], + "angle": 0, + "content": "2 Models of input size \\(512 \\times 640\\) were converted to ONNX and inference speed was averaged over 50 forward passes after a warm-up period." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.398, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "Caltech Aerial RGB-Thermal Dataset in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "significant performance advantages for semantic segmentation on our dataset in order to justify their computational cost." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.198, + 0.785, + 0.289 + ], + "angle": 0, + "content": "Overall, all baselines can segment stuff classes well but none excel with the rare object classes: vehicle and person. Future models would likely need to consider weighted loss functions, smarter data augmentation, or domain adaptation techniques in order to improve on object classes. Lastly, we note that the thermal-specific model, FTNet, performs poorly, likely as its edge priors are less effective in unstructured, natural environments compared to urban environments." + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.303, + 0.779, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.438, + 0.779, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.565, + 0.785, + 0.607 + ], + "angle": 0, + "content": "Fig. 5: (a) Thermal semantic segmentation failures due to intra-class semantic variations when testing on geographically-partitioned, out-of-domain data. (b) Failures due to photometric variations between day (in-domain) and night (out-of-domain)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.785, + 0.716 + ], + "angle": 0, + "content": "Geographically-partitioned benchmarks We retrained our best baseline (Tab. 2) on geographically-partitioned sets to evaluate intra-thermal generalization. Given the cost and effort of collecting diverse aerial thermal data of field settings, we aim to determine the minimum amount of annotated data required to effectively deploy. Using EfficientViT-B0, we performed two experiments to study generalization across difference geographic regions and terrain types." + }, + { + "type": "table_caption", + "bbox": [ + 0.241, + 0.734, + 0.76, + 0.748 + ], + "angle": 0, + "content": "Table 3: Thermal semantic segmentation results on geographically-split data" + }, + { + "type": "table_caption", + "bbox": [ + 0.303, + 0.751, + 0.433, + 0.763 + ], + "angle": 0, + "content": "(a) Region-based split" + }, + { + "type": "table", + "bbox": [ + 0.237, + 0.764, + 0.494, + 0.827 + ], + "angle": 0, + "content": "
Train AreaTest AreaAvg. mIoU
CANCKY
CA0.6660.0840.1930.314
NC0.0840.5080.0310.208
All0.6480.5250.5870.586
" + }, + { + "type": "table_caption", + "bbox": [ + 0.57, + 0.751, + 0.703, + 0.763 + ], + "angle": 0, + "content": "(b) Terrain-based split" + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.764, + 0.763, + 0.838 + ], + "angle": 0, + "content": "
Training TerrainTesting TerrainAvg. mIoU
RiverLakeCoast
River0.6680.1960.1170.327
Lake0.1960.3640.0610.207
Coast0.1270.1190.5220.256
All0.6460.4150.4930.518
All + ground0.6580.4160.5200.532
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "C. Lee et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "In the first experiment, we trained on thermal data from CA and tested on images from KY and NC (Tab. 3a). Despite common classes in each region, results show poor generalization to class variations across geographic areas. We see this again when repeating with the NC split. In the second experiment, we divided our data into three terrain types, aerial river, aerial lake, and aerial coast. Remaining ground-captured images were lumped into a fourth ground category. We trained and tested all pairwise combinations and found poor generalization across terrain types as well (Tab. 3b). Notably, we found low, in-domain performance for the lake setting but saw gains after training on aerial data from other terrain and further improvements after training on ground level images. Overall, our results reiterate the benefit of having training data representative of the testing environment, but also show that gains can be made by including out-of-domain (OOD) data from different terrain and different viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.788, + 0.428 + ], + "angle": 0, + "content": "Overall, the geographic benchmark presents a difficult domain adaptation problem, requiring both inter- and intra-modality domain adaptation techniques to overcome. Although such techniques could perform well in this challenge, the results also highlight a need for an even larger and more diverse thermal dataset capturing common field environments than what we presented." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.437, + 0.788, + 0.575 + ], + "angle": 0, + "content": "Temporally-partitioned benchmark Here, we assess EfficientViT-B0's performance against intra-thermal, temporal distribution shift due to thermal inversion. We train the model on images from the daytime and test on images captured at sunrise and at night. The model achieved mIoUs of 0.242 at sunrise, 0.193 at night, and 0.777 on a separate daytime test split. Despite training on daytime images from the same location, the model fails on nighttime scenes, struggling to classify water due to thermal inversion (Fig. 5b). As such, to perform well on this split, future algorithms need to augment datasets to simulate such inversions or intentional collect data capturing such phenomena." + }, + { + "type": "table_caption", + "bbox": [ + 0.23, + 0.59, + 0.77, + 0.605 + ], + "angle": 0, + "content": "Table 4: RGB-Thermal semantic segmentation network baseline results (mIoU)." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.606, + 0.784, + 0.678 + ], + "angle": 0, + "content": "
ModelBare groundRocky terrainDevel. struct.RoadShrubsTreesSkyWaterVehiclesPersonAllFLOPS (G)
RGB Only†0.7900.8730.8080.5590.6640.7520.8910.9670.3990.0000.67045
Thermal Only†0.7660.8350.7890.5360.6540.7490.9070.9710.3990.0000.66145
EAEFNet [39]0.8050.8630.8130.5450.7020.8020.9370.9760.4810.0000.692358
CRM [60]0.8160.9000.8510.6160.6950.7790.9230.9740.3770.2060.714310
CMNeXt [79]0.8300.9000.8610.6070.7180.8080.9330.9800.5600.1900.740148
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.388, + 0.679, + 0.615, + 0.693 + ], + "angle": 0, + "content": "\\(\\dagger\\) DeepLabV3+ (w/ ResNet50 encoder)" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.713, + 0.524, + 0.73 + ], + "angle": 0, + "content": "4.2 RGB-T Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.738, + 0.787, + 0.785 + ], + "angle": 0, + "content": "We use the paired RGB-T dataset (Sec. 3.2) and partition it using the general split (Sec. 4.1). Recall that RGB and thermal input sizes are \\(960 \\times 600\\) pixels and the thermal FoV is narrower than in the previous benchmarks (Sec. 4.1)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Baselines We test three RGB-T semantic segmentation algorithms achieving state-of-the-art results in urban RGB-T scene segmentation benchmarks: CRM [60], EAEFNet [39], and CMNeXt [79]. To compare against single-modality mod-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.398, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "Caltech Aerial RGB-Thermal Dataset in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.335, + 0.153, + 0.669, + 0.166 + ], + "angle": 0, + "content": "Table 5: RGB-Thermal image translation results" + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.167, + 0.784, + 0.268 + ], + "angle": 0, + "content": "
MethodMethod TypeRGB → Thermal
PSNR ↑SSIM ↑mIoU ↑
UNIT [42]Unpaired GAN12.850.4210.189
MUNIT [27]Unpaired GAN16.080.4590.173
Edge-guided RGB-T [35]Unpaired GAN12.200.4360.225
Pix2Pix [28]Paired GAN20.060.5470.354
Pix2PixHD [73]Paired GAN20.050.5790.379
VQGAN [15]Paired GAN18.090.5400.388
Palette [58]Paired Diffusion11.510.3990.194
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.28, + 0.784, + 0.311 + ], + "angle": 0, + "content": "els, we train two additional ResNet50/DeepLabV3+ segmentation models on RGB and thermal data accordingly (Sec. 4.1)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.316, + 0.785, + 0.452 + ], + "angle": 0, + "content": "Performance analysis All RGB-T models outperform the single-modality baselines, with CMNeXt outperforming the next best RGB-T model by 0.026 mIoU (Tab. 4). Overall, incorporating RGB imagery into the segmentation pipeline greatly improved the distinction between land-based stuff classes which can be hard to distinguish in thermal due to thermal crossover. However, these improvements come at a high computational cost and would not be suitable for field robotic applications. Excelling in this benchmark would involve improving the segmentation performance of the object classes, possibly by leveraging RGB data in unsupervised domain adaptation training schemes." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.462, + 0.536, + 0.478 + ], + "angle": 0, + "content": "4.3 RGB-Thermal Image Translation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.784, + 0.513 + ], + "angle": 0, + "content": "We use the RGB-T paired dataset from Sec. 4.2 to benchmark image translation algorithms operating in the RGB \\(\\rightarrow\\) T direction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.784, + 0.566 + ], + "angle": 0, + "content": "Baselines We benchmark GAN methods that need RGB-T pairs (Pix2Pix [28], Pix2PixHD [73], VQGAN [15]) and unpaired methods (UNIT [42], MUNIT [27], [35]) that do not. We also evaluate Palette [58], a paired diffusion approach." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.577, + 0.779, + 0.715 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.718, + 0.759, + 0.733 + ], + "angle": 0, + "content": "Fig. 6: RGB-to-thermal image translation results. Zoom-in to see fine-details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Image translation metrics We quantify RGB-T image translation using the Peak-Signal-to-Noise ratio (PSNR) and the Structural Similarity Index (SSIM). In addition, we propose a third metric: the thermal mIoU. This is equivalent to the FCN-score used in RGB image translation works [28]. Instead of an RGB network, our metric uses our EfficientViT thermal segmentation network (Tab. 2) to segment translated thermal images before evaluating with ground truth." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "C. Lee et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Performance analysis Paired GANs outperform unpaired GANs and diffusion methods, achieving higher values across all three metrics (Tab. 5). Qualitative assessment (Fig. 6) shows poor translations from unpaired techniques, with most retaining geometric artifacts unique to RGB (ocean waves and shadows) and ignoring relative temperature characteristics. Paired GANs produce results similar to real thermal images but with inconsistent details upon closer inspection. Likewise, the diffusion method produces accurate translations but is not consistent. Overall, RGB-T translation still requires improvements before being able to provide a reliable source of thermal training data from existing RGB datasets. For domain adaptation purposes, we emphasize that maximizing the thermal mIoU score is more important than photometric consistency." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.323, + 0.403, + 0.339 + ], + "angle": 0, + "content": "4.4 Motion Tracking" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.787, + 0.403 + ], + "angle": 0, + "content": "To quantify VIO/SLAM robustness, we select 12 clipped sequences from our dataset ranging in motion tracking difficulty: from urban sports fields (easy) to our natural environments (hard). We normalize thermal images using the \\(5^{\\text{th}}\\) and \\(95^{\\text{th}}\\) percentile pixel values (Sec. 3.2) to enable feature detection." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.422, + 0.786, + 0.436 + ], + "angle": 0, + "content": "Table 6: VIO/SLAM performance (Absolute Trajectory Error [m]) on aerial sequences" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.437, + 0.785, + 0.533 + ], + "angle": 0, + "content": "
MethodTypeModalityNorth FieldCastaic Lake, CADuck, NC
N1N2N3N4N5N6C1C2C3C4D1D2
VINS-FusionSLAMRGB2.8837.5304.0362.0094.0450.8341.555—3.2771.5661.0310.700
VINS-FusionThermal9.45412.269.8542.92611.005.1147.2961.5135.3212.0250.879—
VINS-FusionVIORGB5.4589.5096.5182.1216.0261.2771.555—3.2771.3771.0310.725
VINS-FusionThermal13.8212.8611.992.92613.985.1947.28410.605.3212.0340.879—
Open-VINSVIORGB22.8036.1537.155.165—3.896—3.573——0.4750.700
Open-VINSThermal14.4330.6416.122.258—1.562—5.311——1.073—
Trajectory length (m)1533132612531531104642310651741378880
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.787, + 0.608 + ], + "angle": 0, + "content": "Baselines We evaluate VINS-Fusion [54] (graph optimization-based) and OpenVINS [21] (filtering-based). We test VINS-Fusion in VIO and SLAM modes, with and without loop closure constraints, respectively, and report Absolute Trajectory Error [64] averaged over four runs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Performance analysis This benchmark reveals challenges due to fast motion at altitude and periodic feature sparsity in littoral environments. In such scenes, motion tracking frequently failed due to loss of features, requiring sequences to be clipped for quantifiable evaluations (Tab. 6). VINS-Fusion outperformed Open-VINS mainly due to higher feature tracking reliability. VINS-Fusion works better on RGB images than on thermal images in general. However, VINS-Fusion demonstrates superior feature tracking on thermal images compared to RGB on the C2 sequence of Castaic Lake, where the thermal contrast of stone grains on the lake coast was particularly pronounced in the late afternoon. This observation motivates future research on integrating thermal and RGB cameras for motion tracking with enhanced robustness and versatility. Our benchmark presents a difficult challenge for VIO and SLAM algorithms in extreme environments where water and reflections off water surfaces dominate image scenes, providing a unique test bed to invigorate interest in robust feature matching and motion tracking in natural texture-deficient scenarios." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.398, + 0.115, + 0.733, + 0.128 + ], + "angle": 0, + "content": "Caltech Aerial RGB-Thermal Dataset in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.275, + 0.152, + 0.728, + 0.167 + ], + "angle": 0, + "content": "Table 7: RGB foundation model performances on thermal imagery" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.17, + 0.497, + 0.202 + ], + "angle": 0, + "content": "(a) Instance segmentation via SAM on thermal and augmented color images using coregistered color image segmentations as ground truth" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.203, + 0.496, + 0.258 + ], + "angle": 0, + "content": "
ModalityAPallAPsmallAPmedAPlarge
Thermal0.0180.0070.0220.204
Color (grayscale)0.5380.5360.5310.601
Color (col. jitter)0.8000.7920.8010.841
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.217, + 0.259, + 0.497, + 0.272 + ], + "angle": 0, + "content": "\\(\\dagger\\) refers to AP@0.5::0.95, following MSCOCO [40]" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.17, + 0.781, + 0.192 + ], + "angle": 0, + "content": "(b) Zero-shot semantic segmentation (mIoU) on coregistered color and thermal imagery" + }, + { + "type": "table", + "bbox": [ + 0.5, + 0.196, + 0.78, + 0.258 + ], + "angle": 0, + "content": "
MethodColorColor (gray)Color (jitter)Thermal
SAM [33] + GT ann.‡0.7040.6980.7000.653
Grounded SAM [43]0.3800.3510.3610.193
OV-Seg [38]0.3620.3400.3650.217
ODISE [77]0.5040.4500.4860.232
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.5, + 0.259, + 0.781, + 0.282 + ], + "angle": 0, + "content": "Most frequent ground truth class label in a SAM mask assigned as label for entire mask" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.296, + 0.402, + 0.311 + ], + "angle": 0, + "content": "4.5 Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Zero-shot foundation models on thermal imagery Current foundation models work well on RGB imagery but have not been benchmarked on thermal, especially in non-urban settings. To evaluate these models on thermal imagery, specifically for zero-shot segmentation and semantic segmentation, we compare RGB and thermal performance using the aligned set (Sec. 3.2) in order to isolate modality as the root cause for performance differences. In the following tests, our metrics do not include the rare object classes (vehicle, person) as they drop metrics to near-uniform values for both modalities, making it difficult to compare." + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.456, + 0.771, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.631, + 0.788, + 0.675 + ], + "angle": 0, + "content": "Fig. 7: Zero-shot foundation models on RGB-T pairs. Oversegmented SAM outputs differ across modalities but are similar with ground truth semantics added. Semantic segmentation models (prompted with all classes) perform better on RGB imagery." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.81 + ], + "angle": 0, + "content": "We quantify Segment Anything's (SAM) [33] zero-shot segmentation sensitivity to modality change by measuring the spatial alignment of its mask outputs. We apply it to the RGB set and the 8-bit thermal image set (Sec. 3.2) with the default \\(32 \\times 32\\) grid points to segment everything (Fig. 7). We compute the average precision (AP) of the predicted thermal masks, using the predicted RGB masks as ground truth, and compare the thermal AP to the APs of a color-jittered RGB set and a grayscale set. We find that SAM is sensitive to photometric changes, with large performance drops from color-jitter to grayscale to thermal (Tab. 7a)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Next, we test large vision-language models (Grounded-SAM [43], OV-Seg [38], and ODISE [77]) in zero-shot semantic segmentation. We prompt the models" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "C. Lee et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "with a list of classes in our dataset; generate semantic segmentation masks for RGB, color-jittered RGB, grayscale, and thermal image sets; and compute the mIoU of their outputs using our ground truth annotations. Although RGB, RGB-variants, and thermal modalities all yield poor semantic segmentation performance (Tab. 7b), all RGB variants outperform thermal by at least \\(0.13\\mathrm{mIoU}\\) indicating low generalization to the thermal modality (Fig. 7)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.344 + ], + "angle": 0, + "content": "Finally, we create semantic SAM masks by assigning class labels to SAM instances based on the most frequent ground truth class within an instance (Fig. 7). We find that these masks attain much higher mIoU score (0.653) compared to the other zero-shot methods. Based on the results (Tab. 7), we note that SAM can find object boundaries in the thermal domain reasonably well but can oversegment, resulting in poor AP scores when compared to RGB (which can also oversegment). When taken with ground truth semantics, this behavior is hidden." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.487, + 0.452 + ], + "angle": 0, + "content": "Transfer learning with urban thermal datasets We test if pretraining on thermal urban data can assist transfer to field settings. We sample 40k unlabeled thermal images from [11, 17, 22, 41, 61, 67, 71, 76] and pretrain various networks using Mo" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.789, + 0.53 + ], + "angle": 0, + "content": "CoV2 [9]. We then finetune the pretrained networks on our training set and evaluate on our test set. Overall, we find sparse evidence suggesting any advantage of pretraining with existing urban thermal datasets (Tab. 8). Off-the-shelf ImageNet weights can provide strong performance in the thermal field domain, especially over training from scratch, without the effort and costs of pretraining." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.351, + 0.786, + 0.379 + ], + "angle": 0, + "content": "Table 8: Network pretraining methods for downstream thermal segmentation." + }, + { + "type": "table", + "bbox": [ + 0.499, + 0.38, + 0.785, + 0.443 + ], + "angle": 0, + "content": "
ModelmIoU# Params. (M)
NoneImageNetTherm. Urban
FastSCNN0.6400.6900.6881.1
EfficientViT0.6870.7250.7144.8
ResNet180.7020.7130.70612.3
ResNet500.6820.7130.72826.7
ConvNext-B0.6250.7170.69789.4
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.539, + 0.36, + 0.556 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.777 + ], + "angle": 0, + "content": "We presented the Caltech Aerial RGB-Thermal Dataset, the first dataset tailored towards advancing thermal semantic perception and motion tracking algorithms in natural environments. We created four benchmarks encompassing semantic segmentation, image translation, and motion tracking, and demonstrate challenges faced by current thermal perception and localization algorithms. Semantic segmentation and image translation methods were notably affected by geographical domain shifts, reflecting the wide intra-class and inter-scenery distributions in our dataset, and temporal domain shifts, due to thermal inversion and crossover. Motion tracking in thermal suffered from feature sparsity, a challenge commonly encountered in real-world robotic deployments [1]. Our motion tracking benchmark penalizes current VIO/SLAM algorithms for assuming ideal feature-tracking conditions, serving as a testbed to fuel improvements in this area. Our dataset and benchmarks can be used into the future to help develop algorithms that expand the operating domains of robotics and computer vision." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.786, + 0.405, + 0.803 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "This work was funded by the Office of Naval Research. We thank K. Koetje, K. Brodie, T. Hesser, T. Almeida, T. Cook, N. Spore, and R. Beach." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.398, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Caltech Aerial RGB-Thermal Dataset in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.17, + 0.785, + 0.198 + ], + "angle": 0, + "content": "1. Ackerman, E.: Blade strike on landing ends mars helicopter's epic journey (Jan 2024), https://spectrum.ieee.org/mars-helicopter-ingenuity-end-mission" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.199, + 0.785, + 0.239 + ], + "angle": 0, + "content": "2. Bloesch, M., Burri, M., Omari, S., Hutter, M., Siegwart, R.: Iterated extended kalman filter based visual-inertial odometry using direct photometric feedback. The International Journal of Robotics Research 36(10), 1053-1072 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.24, + 0.785, + 0.295 + ], + "angle": 0, + "content": "3. Bondi, E., Jain, R., Aggrawal, P., Anand, S., Hannaford, R., Kapoor, A., Piavis, J., Shah, S., Joppa, L., Dilkina, B., et al.: Birdsai: A dataset for detection and tracking in aerial thermal infrared videos. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1747-1756 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.787, + 0.35 + ], + "angle": 0, + "content": "4. Brodie, K.L., Bruder, B.L., Slocum, R.K., Spore, N.J.: Simultaneous mapping of coastal topography and bathymetry from a lightweight multicamera uas. IEEE Transactions on Geoscience and Remote Sensing 57(9), 6844-6864 (2019). https://doi.org/10.1109/TGRS.2019.2909026" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.351, + 0.785, + 0.379 + ], + "angle": 0, + "content": "5. Cai, H., Gan, C., Han, S.: Efficientvit: Enhanced linear attention for high-resolution low-computation visual recognition. arXiv preprint arXiv:2205.14756 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.379, + 0.785, + 0.407 + ], + "angle": 0, + "content": "6. Cai, W., Jin, K., Hou, J., Guo, C., Wu, L., Yang, W.: Vdd: Varied drone dataset for semantic segmentation. arXiv preprint arXiv:2305.13608 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.408, + 0.785, + 0.449 + ], + "angle": 0, + "content": "7. Center, A.F.S.: A dataset for machine learning algorithm development from 2010-06-15 to 2010-08-15. NOAA National Centers for Environmental Information (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.449, + 0.785, + 0.492 + ], + "angle": 0, + "content": "8. Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 801-818 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.492, + 0.785, + 0.519 + ], + "angle": 0, + "content": "9. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.52, + 0.785, + 0.575 + ], + "angle": 0, + "content": "0. Chen, Y., Wang, Y., Lu, P., Chen, Y., Wang, G.: Large-scale structure from motion with semantic constraints of aerial images. In: Pattern Recognition and Computer Vision: First Chinese Conference, PRCV 2018, Guangzhou, China, November 23-26, 2018, Proceedings, Part I 1. pp. 347-359. Springer (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.576, + 0.785, + 0.617 + ], + "angle": 0, + "content": "1. Choi, Y., Kim, N., Hwang, S., Park, K., Yoon, J.S., An, K., Kweon, I.S.: Kaist multi-spectral day/night data set for autonomous and assisted driving. IEEE Transactions on Intelligent Transportation Systems 19(3), 934-948 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.617, + 0.785, + 0.659 + ], + "angle": 0, + "content": "2. Deevi, S.A., Lee, C., Gan, L., Nagesh, S., Pandey, G., Chung, S.J.: Rgb-x object detection via scene-specific fusion modules. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 7366-7375 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.785, + 0.714 + ], + "angle": 0, + "content": "3. Delaune, J., Hewitt, R., Lytle, L., Sorice, C., Thakker, R., Matthies, L.: Thermal-inertial odometry for autonomous flight throughout the night. In: 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1122-1128. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.714, + 0.785, + 0.77 + ], + "angle": 0, + "content": "4. Doer, C., Trommer, G.F.: Radar visual inertial odometry and radar thermal inertial odometry: Robust navigation even in challenging visual conditions. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 331-338. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.785, + 0.812 + ], + "angle": 0, + "content": "5. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873-12883 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.812, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. Feng, D., Haase-Schütz, C., Rosenbaum, L., Hertlein, H., Glaeser, C., Timm, F., Wiesbeck, W., Dietmayer, K.: Deep multi-modal object detection and semantic" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.17, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "C. Lee et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.148, + 0.786, + 0.176 + ], + "angle": 0, + "content": "segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems 22(3), 1341-1360 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "17. Teledyne fliradas dataset,https://www.flir.com/oem/adas/adas-dataset-form/,Lastaccessed on2023-10-27" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.785, + 0.233 + ], + "angle": 0, + "content": "18. Fountas, S., Mylonas, N., Malounas, I., Rodias, E., Hellmann Santos, C., Pekkeriet, E.: Agricultural robotics for field operations. Sensors 20(9), 2672 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.785, + 0.261 + ], + "angle": 0, + "content": "19. Gade, R., Moeslund, T.B.: Thermal cameras and applications: a survey. Machine vision and applications 25, 245-262 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.263, + 0.785, + 0.303 + ], + "angle": 0, + "content": "20. Gan, L., Lee, C., Chung, S.J.: Unsupervised rgb-to-thermal domain adaptation via multi-domain attention network. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 6014-6020. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.305, + 0.785, + 0.346 + ], + "angle": 0, + "content": "21. Geneva, P., Eckenhoff, K., Lee, W., Yang, Y., Huang, G.: Openvins: A research platform for visual-inertial estimation. In: 2020 IEEE International Conference on Robotics and Automation (ICRA). pp. 4666-4672. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.347, + 0.785, + 0.403 + ], + "angle": 0, + "content": "22. Ha, Q., Watanabe, K., Karasawa, T., Ushiku, Y., Harada, T.: Mfnet: Towards real-time semantic segmentation for autonomous vehicles with multi-spectral scenes. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 5108-5115. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.404, + 0.785, + 0.445 + ], + "angle": 0, + "content": "23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.785, + 0.487 + ], + "angle": 0, + "content": "24. He, M., Rajkumar, R.R.: Using thermal vision for extended vins-mono to localize vehicles in large-scale outdoor road environments. In: 2021 IEEE Intelligent Vehicles Symposium (IV). pp. 953-960. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.53 + ], + "angle": 0, + "content": "25. Howard, A., Sandler, M., Chu, G., Chen, L.C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., et al.: Searching for mobilenetv3. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1314-1324 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.531, + 0.785, + 0.572 + ], + "angle": 0, + "content": "26. Hua, T., Pei, L., Li, T., Wu, Q., Wang, R., Yu, W.: I2-slam: Fusing infrared camera and imu for simultaneous localization and mapping. In: International Conference on Autonomous Unmanned Systems. pp. 2834-2844. Springer (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.785, + 0.6 + ], + "angle": 0, + "content": "27. Huang, X., Liu, M.Y., Belongie, S., Kautz, J.: Multimodal unsupervised image-to-image translation. In: European Conference on Computer Vision (ECCV) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.785, + 0.643 + ], + "angle": 0, + "content": "28. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1125-1134 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.644, + 0.785, + 0.685 + ], + "angle": 0, + "content": "29. Jia, X., Zhu, C., Li, M., Tang, W., Zhou, W.: Llvip: A visible-infrared paired dataset for low-light vision. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3496-3504 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.785, + 0.741 + ], + "angle": 0, + "content": "30. Jong, A., Yu, M., Dhrafani, D., Kailas, S., Moon, B., Sycara, K., Scherer, S.: Wit-uas: A wildland-fire infrared thermal dataset to detect crew assets from aerial views. In: 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 11464-11471. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.785, + 0.784 + ], + "angle": 0, + "content": "31. Khattak, S., Papachristos, C., Alexis, K.: Keyframe-based direct thermal-inertial odometry. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 3563-3569. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.838 + ], + "angle": 0, + "content": "32. Kim, Y.H., Shin, U., Park, J., Kweon, I.S.: Ms-uda: Multi-spectral unsupervised domain adaptation for thermal image semantic segmentation. IEEE Robotics and Automation Letters 6(4), 6497–6504 (Oct 2021). https://doi.org/10.1109/LRA.2021.3093652" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.786, + 0.838 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.398, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Caltech Aerial RGB-Thermal Dataset in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "33. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.23 + ], + "angle": 0, + "content": "34. Lee, C., Frennert, J.G., Gan, L., Anderson, M., Chung, S.J.: Online self-supervised thermal water segmentation for aerial vehicles. In: 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 7734-7741. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.23, + 0.788, + 0.285 + ], + "angle": 0, + "content": "35. Lee, D.G., Jeon, M.H., Cho, Y., Kim, A.: Edge-guided multi-domain rgb-to-tir image translation for training vision tasks with challenging labels. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 8291-8298. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.285, + 0.788, + 0.325 + ], + "angle": 0, + "content": "36. Li, C., Xia, W., Yan, Y., Luo, B., Tang, J.: Segmenting objects in day and night: Edge-conditioned cnn for thermal image semantic segmentation. IEEE Transactions on Neural Networks and Learning Systems 32(7), 3069-3082 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.325, + 0.788, + 0.352 + ], + "angle": 0, + "content": "37. Li, H., Ma, Y., Huang, Y., Gu, Y., Xu, W., Liu, Y., Zuo, X.: Riders: Radar-infrared depth estimation for robust sensing. arXiv preprint arXiv:2402.02067 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.352, + 0.788, + 0.405 + ], + "angle": 0, + "content": "38. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7061-7070 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.406, + 0.788, + 0.447 + ], + "angle": 0, + "content": "39. Liang, M., Hu, J., Bao, C., Feng, H., Deng, F., Lam, T.L.: Explicit attention-enhanced fusion for rgb-thermal perception tasks. IEEE Robotics and Automation Letters (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.502 + ], + "angle": 0, + "content": "40. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.502, + 0.788, + 0.556 + ], + "angle": 0, + "content": "41. Liu, J., Fan, X., Huang, Z., Wu, G., Liu, R., Zhong, W., Luo, Z.: Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5802-5811 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.556, + 0.788, + 0.583 + ], + "angle": 0, + "content": "42. Liu, M.Y., Breuel, T., Kautz, J.: Unsupervised image-to-image translation networks. In: Conference on Neural Information Processing Systems (NeurIPS) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.623 + ], + "angle": 0, + "content": "43. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.623, + 0.788, + 0.663 + ], + "angle": 0, + "content": "44. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.663, + 0.788, + 0.732 + ], + "angle": 0, + "content": "45. Lyu, Y., Vosselman, G., Xia, G.S., Yilmaz, A., Yang, M.Y.: Uavid: A semantic segmentation dataset for uav imagery. ISPRS Journal of Photogrammetry and Remote Sensing 165, 108 - 119 (2020). https://doi.org/https://doi.org/10.1016/j.isprsjprs.2020.05.009, http://www.sciencedirect.com/science/article/pii/S0924271620301295" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.732, + 0.788, + 0.759 + ], + "angle": 0, + "content": "46. Mehta, S., Rastegari, M.: Separable self-attention for mobile vision transformers. arXiv preprint arXiv:2206.02680 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.759, + 0.788, + 0.799 + ], + "angle": 0, + "content": "47. Mostegel, C., Maurer, M., Heran, N., Pestana Puerta, J., Fraundorfer, F.: Semantic drone dataset (Jan 2019), http://dronedataset.icg.tugraz.at/, Last accessed on 2023-10-27" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.841 + ], + "angle": 0, + "content": "48. Nigam, I., Huang, C., Ramanan, D.: Ensemble knowledge transfer for semantic segmentation. In: 2018 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 1499-1508. IEEE (2018)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "C. Lee et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "49. Nirgudkar, S., DeFilippo, M., Sacarny, M., Benjamin, M., Robinette, P.: Massmind: Massachusetts maritime infrared dataset. The International Journal of Robotics Research 42(1-2), 21-32 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.787, + 0.257 + ], + "angle": 0, + "content": "50. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Howes, R., Huang, P.Y., Xu, H., Sharma, V., Li, S.W., Galuba, W., Rabbat, M., Assran, M., Ballas, N., Synnaeve, G., Misra, I., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.787, + 0.298 + ], + "angle": 0, + "content": "51. Panetta, K., Shreyas Kamath, K.M., Rajeev, S., Agaian, S.S.: Ftnet: Feature transverse network for thermal image semantic segmentation. IEEE Access 9, 145212-145227 (2021). https://doi.org/10.1109/ACCESS.2021.3123066" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.787, + 0.352 + ], + "angle": 0, + "content": "52. Park, T., Efros, A.A., Zhang, R., Zhu, J.Y.: Contrastive learning for unpaired image-to-image translation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16. pp. 319-345. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.787, + 0.407 + ], + "angle": 0, + "content": "53. Poudel, R., Liwicki, S., Cipolla, R.: Fast-scnn: Fast semantic segmentation network. In: Sidorov, K., Hicks, Y. (eds.) Proceedings of the British Machine Vision Conference (BMVC). pp. 187.1-187.12. BMVA Press (September 2019). https://doi.org/10.5244/C.33.187, https://dx.doi.org/10.5244/C.33.187" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.787, + 0.434 + ], + "angle": 0, + "content": "54. Qin, T., Li, P., Shen, S.: Vins-mono: A robust and versatile monocular visual-inertial state estimator. IEEE Transactions on Robotics 34(4), 1004-1020 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.787, + 0.488 + ], + "angle": 0, + "content": "55. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.528 + ], + "angle": 0, + "content": "56. Retief, S.P., Willers, C., Wheeler, M.: Prediction of thermal crossover based on imaging measurements over the diurnal cycle. In: Geo-Spatial and Temporal Image and Data Exploitation III. vol. 5097, pp. 58-69. SPIE (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.787, + 0.582 + ], + "angle": 0, + "content": "57. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.787, + 0.623 + ], + "angle": 0, + "content": "58. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.787, + 0.663 + ], + "angle": 0, + "content": "59. Saputra, M.R.U., Lu, C.X., de Gusmao, P.P.B., Wang, B., Markham, A., Trigoni, N.: Graph-based thermal-inertial slam with probabilistic neural networks. IEEE Transactions on Robotics 38(3), 1875-1893 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.663, + 0.787, + 0.704 + ], + "angle": 0, + "content": "60. Shin, U., Lee, K., Kweon, I.S.: Complementary random masking for rgb-thermal semantic segmentation. In: IEEE International Conference on Robotics and Automation (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.787, + 0.745 + ], + "angle": 0, + "content": "61. Shin, U., Park, J., Kweon, I.S.: Deep depth estimation from thermal image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1043-1053 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.787, + 0.799 + ], + "angle": 0, + "content": "62. Shivakumar, S.S., Rodrigues, N., Zhou, A., Miller, I.D., Kumar, V., Taylor, C.J.: Pst900: Rgb-thermal calibration, dataset and segmentation network. In: 2020 IEEE international conference on robotics and automation (ICRA). pp. 9441-9447. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "63. Speth, S., Goncalves, A., Rigault, B., Suzuki, S., Bouazizi, M., Matsuo, Y., Prendinger, H.: Deep learning with rgb and thermal images onboard a drone for monitoring operations. Journal of Field Robotics 39(6), 840-868 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.398, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Caltech Aerial RGB-Thermal Dataset in the Wild" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "64. Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for the evaluation of rgb-d slam systems. In: 2012 IEEE/RSJ international conference on intelligent robots and systems. pp. 573-580. IEEE (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "65. Sun, Y., Zuo, W., Liu, M.: Rtfnet: RGB-thermal fusion network for semantic segmentation of urban scenes. IEEE Robotics and Automation Letters 4(3), 2576-2583 (2019). https://doi.org/10.1109/LRA.2019.2904733" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.285 + ], + "angle": 0, + "content": "66. Sun, Y., Zuo, W., Yun, P., Wang, H., Liu, M.: Fuseseg: Semantic segmentation of urban scenes based on rgb and thermal data fusion. IEEE Transactions on Automation Science and Engineering 18(3), 1000-1011 (2021). https://doi.org/10.1109/TASE.2020.2993143" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.287, + 0.788, + 0.327 + ], + "angle": 0, + "content": "67. Suo, J., Wang, T., Zhang, X., Chen, H., Zhou, W., Shi, W.: Hit-uav: A high-altitude infrared thermal dataset for unmanned aerial vehicle-based object detection. Scientific Data 10, 227 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.37 + ], + "angle": 0, + "content": "68. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105-6114. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.424 + ], + "angle": 0, + "content": "69. Ustun, B., Kaya, A.K., Ayerden, E.C., Altinel, F.: Spectral transfer guided active domain adaptation for thermal imagery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. pp. 449-458 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.467 + ], + "angle": 0, + "content": "70. Vargas, J., Alsweiss, S., Token, O., Razdan, R., Santos, J.: An overview of autonomous vehicles sensors and their vulnerability to weather conditions. Sensors 21(16), 5397 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.509 + ], + "angle": 0, + "content": "71. Vertens, J., Zürn, J., Burgard, W.: Heatnet: Bridging the day-night domain gap in semantic segmentation with thermal images. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8461-8468. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.563 + ], + "angle": 0, + "content": "72. VS, V., Poster, D., You, S., Hu, S., Patel, V.M.: Meta-uda: Unsupervised domain adaptive thermal object detection using meta-learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1412-1423 (January 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.618 + ], + "angle": 0, + "content": "73. Wang, T.C., Liu, M.Y., Zhu, J.Y., Tao, A., Kautz, J., Catanzaro, B.: High-resolution image synthesis and semantic manipulation with conditional gans. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.66 + ], + "angle": 0, + "content": "74. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. In: Neural Information Processing Systems (NeurIPS) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.702 + ], + "angle": 0, + "content": "75. Xie, S., Girshick, R., Dollar, P., Tu, Z., He, K.: Aggregated residual transformations for deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1492-1500 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.702, + 0.788, + 0.756 + ], + "angle": 0, + "content": "76. Xiong, H., Cai, W., Liu, Q.: Mcnet: Multi-level correction network for thermal image semantic segmentation of nighttime driving scene. Infrared Physics & Technology p. 103628 (2021). https://doi.org/https://doi.org/10.1016/j.infrared.2020.103628" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.757, + 0.788, + 0.813 + ], + "angle": 0, + "content": "77. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2955-2966 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.813, + 0.788, + 0.84 + ], + "angle": 0, + "content": "78. Yun, S., Jung, M., Kim, J., Jung, S., Cho, Y., Jeon, M.H., Kim, G., Kim, A.: Stereo: Stereo thermal dataset for research in odometry and mapping. In: 2022" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "C. Lee et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.784, + 0.175 + ], + "angle": 0, + "content": "IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 3857-3864. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "79. Zhang, J., Liu, R., Shi, H., Yang, K., Reif, S., Peng, K., Fu, H., Wang, K., Stiefelhagen, R.: Delivering arbitrary-modal semantic segmentation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.245 + ], + "angle": 0, + "content": "80. Zhao, S., Wang, P., Zhang, H., Fang, Z., Scherer, S.: Tp-tio: A robust thermal-inertial odometry with deep thermalpoint. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4505-4512. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.299 + ], + "angle": 0, + "content": "81. Zhou, W., Liu, J., Lei, J., Yu, L., Hwang, J.N.: Gmnet: Graded-feature multilabel-learning network for rgb-thermal urban scene semantic segmentation. IEEE Transactions on Image Processing 30, 7790–7802 (2021). https://doi.org/10.1109/TIP.2021.3109518" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "82. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Proceedings of the IEEE international conference on computer vision. pp. 2223-2232 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.384 + ], + "angle": 0, + "content": "83. Zhu, P., Wen, L., Du, D., Bian, X., Fan, H., Hu, Q., Ling, H.: Detection and tracking meet drones challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(11), 7380-7399 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.384 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Caltech Aerial RGB-Thermal Dataset in the Wild/377537ff-2360-4d17-b986-8337083e74f6_origin.pdf b/2024/Caltech Aerial RGB-Thermal Dataset in the Wild/377537ff-2360-4d17-b986-8337083e74f6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b44952ea084d17664230ba82766a8ae17bd54dd9 --- /dev/null +++ b/2024/Caltech Aerial RGB-Thermal Dataset in the Wild/377537ff-2360-4d17-b986-8337083e74f6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f873f1397a427dd6dabcb19067ee8a95ac1e018d9fbc8282bc0134cf5b20f7d +size 7656260 diff --git a/2024/Caltech Aerial RGB-Thermal Dataset in the Wild/full.md b/2024/Caltech Aerial RGB-Thermal Dataset in the Wild/full.md new file mode 100644 index 0000000000000000000000000000000000000000..508f72247dc80ea07bba04e1cd6ab52f760e0d02 --- /dev/null +++ b/2024/Caltech Aerial RGB-Thermal Dataset in the Wild/full.md @@ -0,0 +1,353 @@ +# Caltech Aerial RGB-Thermal Dataset in the Wild + +Connor Lee*, Matthew Anderson*, Nikhil Ranganathan, Xingxing Zuo, Kevin Do, Georgia Gkioxari, and Soon-Jo Chung + +California Institute of Technology + +{cee, matta, nrangana, zuox, kdo, georgia, sjchung}@caltech.edu + +Abstract. We present the first publicly-available RGB-thermal dataset designed for aerial robotics operating in natural environments. Our dataset captures a variety of terrain across the United States, including rivers, lakes, coastlines, deserts, and forests, and consists of synchronized RGB, thermal, global positioning, and inertial data. We provide semantic segmentation annotations for 10 classes commonly encountered in natural settings in order to drive the development of perception algorithms robust to adverse weather and nighttime conditions. Using this dataset, we propose new and challenging benchmarks for thermal and RGB-thermal (RGB-T) semantic segmentation, RGB-T image translation, and motion tracking. We present extensive results using state-of-the-art methods and highlight the challenges posed by temporal and geographical domain shifts in our data. The dataset and accompanying code is available at https://github.com/aerorobotics/caltech-aerial-rgbt-dataset. + +Keywords: Robotics $\cdot$ Semantic segmentation $\cdot$ RGB-T image translation $\cdot$ Visual inertial odometry + +![](images/c66c1a82a96585488db1264a7a275c3e3743fbc07f341757e581bcdcc184c156.jpg) +Thermal Dataset Comparisons + +![](images/acaa29bae9a83e7c9278a033abb99491c6761ad857f96e58994dbd3f761588bd.jpg) +A. Thermal + RGB-T Semantic Segmentation +Aligned RGB-T pairs + +![](images/d092be201dc124929f1cd40740c76ad1ac02d792f663a20c7faf0d7b90b5c546.jpg) +B. RGB-to-Thermal Image Translation +Fig. 1: Left: Our dataset is uniquely designed to improve thermal scene perception for field robots. Right: We provide new benchmarks for thermal-based vision algorithms, including (a) semantic segmentation, (b) image translation, and (c) motion tracking. + +![](images/d04c086a5971a11d35af62916fb5bd6da8ce7ff746002449eacdb275aa97f438.jpg) +C. Motion Tracking + +# 1 Introduction + +Field robots rely predominantly on visual cameras, lidar, and radar to perceive their surroundings [16, 18]. These sensors provide robust perception capabilities, + +but falter in low-light and adverse weather conditions [70]. In contrast, thermal cameras use long-wave infrared wavelengths to capture emitted heat, offering dense, visual information in such conditions where other sensors struggle [19]. Recently, they have been used for multimodal perception in autonomous driving and are increasingly being used to enable nighttime robot autonomy [11, 12, 36, 37, 49, 71, 76, 78]. However, successful integration of thermal cameras requires extensive datasets across diverse settings, preventing widespread adoption in field robotics. + +Current thermal datasets focus on urban settings for autonomous driving applications (Fig. 1). They comprise of thermal-only imagery [36,76] or RGB-Thermal (RGB-T) pairs [17,22,41,71], with some including global positioning (GPS/GNSS) and inertial measurements (IMU) for visual-inertial odometry (VIO) and simultaneous localization and mapping (SLAM) [11,61,78]. While comprehensive, they lack data depicting natural settings like rivers and forests, which are typical operating areas for field robotics that perform coastline mapping and bathymetry [4] or activity monitoring during forest fires [30]. + +Due to the lack of thermal benchmarks for algorithms like semantic segmentation and SLAM, field robots cannot easily operate at night or in adverse conditions, requiring online learning [34, 72] or unsupervised domain adaptation to compensate [20, 32, 71], which still require labeled data for evaluation. As such, enabling thermal perception requires collecting and annotating field-specific data from scratch. However, creating a thermal dataset for field robotics is difficult due to two reasons: First, thermal data is hard to crowdsourced and web-scrape due to the richness of thermal cameras. Hence, acquiring a diverse dataset for algorithm development requires physically going to distinct field locations, increasing time and cost. Second, operating field robots, especially uninhabited aerial vehicles (UAV), usually requires special permits for each location-of-capture. + +In this work, we present the first dataset targeting thermal scene perception for field robotics. The dataset consists of oblique-facing imagery captured from a UAV, supplemented with image sets captured at ground level, and focuses on littoral settings within various desert, forest, and coastal environments across the United States. We contribute new benchmarks for thermal and RGB-T semantic segmentation, RGB-T image translation, and VIO/SLAM algorithms, with unique challenges characterized by temporal and geographical domain shifts for learning-based methods and feature sparsity for motion-tracking algorithms. Aside from the proposed benchmarks, our dataset can also be used towards other applications like visual place recognition and terrain-relative navigation. + +The paper is organized as follows: Sec. 2 reviews relevant datasets, benchmarks, and algorithms. Sec. 3 describes our dataset details and curation process, and Sec. 4 presents benchmark results. Sec. 5 offers concluding remarks. + +# 2 Related Work + +# 2.1 Datasets and Benchmarks + +Thermal/RGB-T benchmarks for urban robotics Current urban thermal and RGB-T datasets primarily focus on autonomous vehicle (AV) technology + +Table 1: Comparison of datasets captured from aerial platforms. Camera angle of $90^{\circ}$ is nadir-pointing. See Tab. S1 in supplement for comparison against thermal datasets captured from surface and ground vehicles. + +
DatasetTaskRGBThermalGPSIMUSettingLocation# SamplesCamera pose
VisDrone [83]Obj. Det.✓XXXUrbanChina10,209Variable
HIT-UAV [67], NII-CU [63]Obj. Det.✓✓XXUrbanChina, Japan2,898 - 5,88030°, 90°, 60-130 m
WTT-UAS [30]Obj. Det.X✓✓✓ForestUSA6,95130°, 90°, < 120 m
BIRDSAI [3]Obj. Track.X✓XXSavannasS. Africa62,000Off-nadir, 60-120 m
AeroScapes [48], UAVid [45], [10], [6], Swiss/Okutama [63]Sem. Seg.✓XXXUrbanChina, Japan, Switzerland191 - 326930 - 90°, < 100 m
Semantic Drone Dataset [47]Sem. Seg.✓✓X✓UrbanGermany40090°, < 30m
OursSem. Seg.✓✓✓✓River, Coast, Lake, Mountain, DesertUSA4,19520°, 45°, < 120 m
+ +and surveillance applications. AV-related datasets cover tasks such as object detection [11, 17, 41], semantic segmentation [22, 36, 71, 76], and localization [11, 61, 78]. Some include data from other spectra, benchmarking not only thermal algorithms but also RGB-T [11, 22, 41, 63, 71] and other multispectral models. In contrast, datasets for surveillance applications use fixed cameras [29] and UAVs [63, 67] to detect personnel, vehicles, and other objects. Despite many urban thermal benchmarks, they cannot be directly used to develop algorithms for robots operating in natural environments due to the domain gap. + +Thermal/RGB-T benchmarks for field robotics Datasets for robots operating in non-urban environments remain limited. MassMIND [49] and PST900 [62] benchmark thermal semantic segmentation for maritime and subterranean robotic environments, respectively. WIT-UAS [30] provides UAV-borne thermal images for object detection in wildfire-prone environments. One work [34] proposes an aerial RGB-T dataset for thermal water segmentation in littoral areas but only provides 1272 labels consisting of a single water class. Benchmarks like [3,7] depict natural environments but are targeted towards wildlife tracking. To address this gap, we present an aerial RGB-T dataset which can be used to benchmark multiple tasks, including semantic segmentation, RGB-T image translation, and localization, across diverse natural environment robotic applications. Specifically, we build upon the dataset from [34] by extending the number of semantic classes from 1 to 10 and adding additional RGB-T data captured from air and ground, resulting in a total of 4195 semantic segmentation annotations. In comparison to existing thermal/RGB-T datasets, our dataset can enable nighttime field robotic applications like remote sensing research, geological sampling, precision agriculture, and wilderness search and rescue. + +UAV-based semantic segmentation datasets Aside from [34], existing UAV-captured semantic segmentation datasets [6, 10, 45, 47, 48, 63] cover urban environments (Tab. 1). With the exception of [47] which provides 400 thermal image samples, all of these datasets contain only RGB data for semantic segmentation. While they provide images with relevant view angles, the modality gap and focus on urban content makes them impractical for aerial field robotic applications. + +# 2.2 Algorithms using Thermal Datasets + +Thermal semantic segmentation Since thermal images are spatially identical to their RGB counterparts, popular RGB semantic segmentation algo + +rithms [5, 8, 53, 74] can be applied directly. However, some works [36, 51, 76] develop models specifically for thermal imagery, and leverage edge priors to overcome low resolution and blurriness due to thermal crossover [56]. Although specifically designed for thermal, these models only marginally outperform standard RGB segmentation models [51] and their performance in field settings, where edges are less structured, remains untested. Aside from architectural advancements, improvements in thermal semantic segmentation come from domain adaptation (DA) methods [20, 36, 69, 71] that use labeled RGB data with unlabeled thermal data. To provide fair comparisons, we do not provide DA baselines in our benchmarks as performance varies based on the choice of RGB data. + +RGB-T semantic segmentation RGB-T semantic segmentation methods compensate for the weaknesses of each modality via deep feature fusion, data augmentation, and adversarial training. Most approaches [22, 60, 65, 66, 79] utilize encoder-decoder architectures with modality-specific encoders and shared decoders, and show significant improvements over channel-stacked RGB-T inputs. Works like [39, 81] make improvements by integrating RGB and thermal features at multiple stages within the encoders, while [62] eschews RGB-T training pairs by processing RGB-T inputs sequentially. Other methods include random input masking to reduce reliance on a single modality [60] and adversarial training to adapt to different times of day [71]. + +RGB-T image translation RGB-T image translation is vital as it may enable thermal training data to be generated from large-scale RGB datasets. Works based on generative adversarial networks (GAN) such as [27,35,42,52,82] can be used without RGB-T image pairs. However, GANs like [28,73] that require paired imagery provide better translations for both visual appeal and domain adaptation [36]. Recent diffusion methods [57,58] have demonstrated remarkable results on RGB imagery, but have not been tested on RGB-T image translation. + +Thermal VIO and SLAM Low signal-to-noise ratio, reduced contrast, and blurred boundaries in thermal images present challenges for VIO and SLAM. Several studies [2, 13, 14, 24, 26] use indirect methods by tracking keypoints extracted from thermal images for visual constraints and rely on image normalization techniques to improve keypoint extraction. In contrast, [31] uses a direct method to extract and track high-gradient points by minimizing radiometric error. Other works utilize learned local and global features to improve feature tracking [80] and loop closure detection [59], respectively. + +# 3 The Caltech Aerial RGB-Thermal Dataset + +# 3.1 Data Acquisition Hardware + +To capture this dataset, we developed a custom sensor stack with synchronized RGB-thermal imagery, IMU, and global positioning data (Fig. 2). The sensor stack features a non-radiometric FLIR ADK thermal camera ( $640 \times 512$ px and $75^{\circ}$ horizontal FoV, $60\mathrm{Hz}$ ) flanked by a pair of FLIR Blackfly electro-optical (EO) monochrome and color cameras $^{1}$ ( $960 \times 600$ px and $75^{\circ}$ horizontal FoV, $30\mathrm{Hz}$ ). + +Pose information is provided by a VectorNav VN100 IMU (200 Hz) and a u-blox M8N GPS (5 Hz). All three cameras and the VN100 are hardware synchronized using a dedicated signal generator and are rigidly attached to each other using a stiff, 3D printed mount. A Simply NUC Ruby Mini PC (AMD Ryzen 7 4800U, 32 GB RAM) is used for the compute, and the entire system is interfaced using ROS Noetic. Our cameras and IMU were calibrated following the procedures found in Supplementary Material Sec. S2.2 and were not disassembled from each other between any of the dataset collections. + +![](images/bfbf75378943dbb397f1282aab309c5ea349027af3482b5447f284aaca0de067.jpg) +Fig. 2: (a) The Aurelia X6 hexacopter and sensor stack used to capture our aerial dataset. (b) Geographic distribution of our data collection sites and collection times. + +![](images/8c39ed1a4de2c64a999a423b96d24ac4a04788514abf50adee8c5fb6bd87997a.jpg) + +The sensor stack is mounted rigidly onboard an Aurelia X6 hexacopter (Fig. 2) without gimbal stabilization at angles between $0^{\circ}$ (level with the horizon) to $45^{\circ}$ downwards to provide different viewpoints. When mounted to a UAV, we also log the position and attitude estimates from the UAV at $20\mathrm{Hz}$ providing improved positioning accuracy from the onboard RTK GPS. Additionally, the sensor stack's modularity allows us to capture datasets on foot by mounting it to a tripod where flight restrictions are in-place. + +# 3.2 Data Collection and Processing + +Data capture We captured 37 aerial and ground trajectories covering lakes, rivers, coastlines, deserts, and mountains from around Southern California, Kentucky, and North Carolina (Fig. 2). Aerial trajectories (18) involve high motion with intermittent hovering, and mostly depict scenes from $40\mathrm{m}$ altitude. Ground-level trajectories (19) were captured on foot, with 7 captured without any movement for VIO debugging purposes. For more details, see Supplementary Materials Tab. S2 and Sec. S2.1. + +Thermal image normalization For the baselines presented in this work, we typically normalize 16-bit thermal data by rescaling between the $1^{\mathrm{st}}$ and $99^{\mathrm{th}}$ percentile pixel values and follow with contrast limited adaptive histogram equalization (CLAHE). We keep normalized data as floats to minimize information loss, and only convert to 8-bit format when necessary or for visualization. + +![](images/762c02a09c796d13493e5be1f26e0262c19945d73ac2f8deb360b8dab5b20328.jpg) +Semantic Segmentation Classes and Description + +![](images/026bac4764f8e46486c4665e30fc64344aeeba0305af094380eedf3b0642642b.jpg) +a +b + +![](images/dfe213dcb7b6976c69c56621d6fea9ca1dbc70f616b4a21e7c7b5d2eeb685d65.jpg) +Fig. 3: Semantic segmentation classes in our dataset. The color mapping is used throughout this paper. (a) Hourly distribution of annotated thermal images. (b) Histogram of semantic classes. + +Semantic segmentation annotation To aid development of scene perception algorithms for field robots, we manually-annotated a subset of thermal images for semantic segmentation with classes in Fig. 3. To avoid redundancy, images were sampled every 3s when moving and every 20s when still, resulting in 4195 samples. We considered a frame in motion if the distance moved within the past 2 seconds exceeded 0.5 meters, with distances computed using GPS coordinates. + +To promote labeling consistency, annotators preread descriptions of each trajectory, which contained a list of classes present, labeled examples provided by the authors, and GPS coordinates (for Google Earth look-up). Annotators were then given an RGB-T image pair with trajectory metadata and asked to label the thermal frame. Annotations were reviewed by authors, with inconsistent or incorrect labels sent back. 3 rounds of annotation were done, with the final and initial set of labels differing by $0.08\mathrm{mIoU}$ . Classes most frequently re-annotated were bare ground, rocky terrain, road, trees, and vehicles. + +RGB-thermal image alignment To align thermal and RGB image pairs, we stereo rectified the RGB-T image pairs using the camera matrices obtained from calibration (Sec. S2.2) and projected the thermal image into the larger RGB image frame to preserve RGB resolution. As our baseline is small compared to the depth of the scenes we capture, image pairs are near-coregistered after rectification. Due to differences in optics, the field-of-view of the aligned image pairs is narrower compared to that of the thermal camera. + +# 3.3 Dataset Splits + +General (benchmark) split This is our primary split for semantic segmentation and image translation benchmarks. We randomly partition the annotated thermal dataset (Sec. 3.2) into train/val/test sets at a 75:12.5:12.5 ratio. + +Temporal split To promote studies into the effect of different time-of-day capture of the thermal images, we split the dataset using three time periods: twilight/sunrise (5 AM - 7 AM), daytime (10 AM - 5 PM), and nighttime (7 PM - 4 AM). We randomly partition the daytime images into train/val/test sets using the ratio from above and leave the twilight and nighttime sets for testing. + +Geographical split The geographical splits are intended to test algorithm adaptability to unseen settings. We provide two splits: + +1. Terrain-based: This split partitions data into the following terrain categories: aerial river, aerial lake, aerial coast, and ground-captured images. Train/val/test splits are created per category with a 50:15:35 ratio. +2. Region-based: This is based on the area of the United States where the data was captured. Distinct regions include California (CA), Kentucky (KY), and North Carolina (NC), with train/val/test created in the same way as above. + +# 4 Experiments + +We provide baselines for the following benchmarks: thermal semantic segmentation, RGB-T semantic segmentation, RGB-T image translation, and VIO/SLAM. Furthermore, we explore the zero-shot thermal segmentation capabilities of large vision-language models on our dataset and perform further ablation studies. + +Table 2: Supervised model performance (mIoU), framerate (FPS at batch size 1), and floating point operations (FLOP) for various semantic segmentation network baselines. + +
ModelBare groundRocky terrainDevelop. struct.RoadShrubsTreesSkyWaterVehiclesPersonStuffObjectAllOrin (GPU)NUC (CPU)FLOPS (G)
FastSCNN [53]0.8070.9070.7280.6310.6940.7760.9560.9650.3840.0580.8080.2210.690124.285.61.0
MobileNet-V3-S 0.75 [25]0.8020.9140.6760.6330.7040.8140.9550.9760.3830.1110.8090.2470.69775.224.03.3
MobileNet-V3-L 0.750.8010.9140.7010.6150.7130.7930.9590.9760.4320.1480.8090.2900.70559.015.54.8
MobileViTV2 0.50 [46]0.8220.9130.7590.6500.6940.7810.9540.9700.2220.0300.8180.1260.68060.110.75.5
EfficientViT-B0 [5]0.8250.9170.7770.6180.7100.7930.9610.9650.4950.1910.8210.3430.72551.25.53.9
EfficientNet-Lite0 [68]0.8190.9080.7690.6010.7090.8050.9620.9680.3800.0790.8180.2290.70047.712.87.2
EfficientNet-Lite20.8250.9200.7650.6140.6980.8050.9580.9710.3860.1180.8200.2520.70638.911.29.4
Segformer-B0 [74]0.8040.9100.6900.6240.6960.7880.9600.9700.1950.0000.8050.0970.66438.35.910.2
Segformer-B10.8140.9090.7730.6030.7130.7990.9600.9650.3660.0000.8170.1830.69030.33.519.9
ResNet18 [23]0.8100.9050.7660.6180.7110.8070.9590.9660.4310.1580.8180.2940.71369.312.422.4
ResNet500.8190.9160.7470.6580.7110.7990.9610.9740.3610.1820.8230.2720.71329.96.145.4
ResNeXt50 [75]0.8250.9190.7730.6530.7090.7940.9640.9760.4360.1370.8270.2870.71923.05.645.5
ConvNext-T [44]0.7990.9090.7190.6080.7060.8080.9610.9700.3630.0030.8100.1830.68525.83.347.4
ConvNext-S0.8100.9130.7710.6030.7180.8100.9650.9640.4920.0480.8190.2700.70919.12.375.0
ConvNext-B0.8100.9210.6970.6110.7060.8030.9630.9720.3480.1120.8110.2300.69414.61.4130.5
ConvNext-B (CLIP) [55]0.8130.9180.6830.6320.7180.8130.9650.9720.5170.1370.8140.3270.71714.61.4130.4
ConvNext-B (CLIP) *0.7730.8870.7130.4870.6560.7560.9460.9510.1580.0190.7710.0890.63514.61.4130.4
DINOv2 [50] (linear head)0.8000.9070.6810.6060.6930.7960.9560.9670.3750.0790.8010.2270.68615.22.1—
DINOv2 (linear head) *0.7050.8440.6190.4190.5580.7250.9220.9200.3560.1580.7140.2570.62315.22.1—
DINOv2 (nonlin. head)0.8100.9160.7080.6350.7000.8190.9590.9730.3990.1190.8150.2590.70415.22.1—
DINOv2 (nonlin. head) *0.7960.9030.7320.6060.6910.7940.9540.9630.4710.1490.8050.3100.70615.22.1—
FTNet† [51]0.7550.8670.6350.5760.6430.6530.7870.9470.2340.0240.7330.1290.61311.11.4100.5
+ +$\dagger$ Thermal-specific model * Frozen encoder + +# 4.1 Thermal Semantic Segmentation + +We run multiple baselines on our general split before choosing a specific model to test on the geographic and temporal splits. We go over our baselines before analyzing their performance on our benchmarks. + +Baselines Motivated by robotic applications, we are interested in both inference speed and segmentation performance. Our baselines place an emphasis on lightweight, real-time networks, but also on foundation models like DINOv2 [50] and ConvNext-B (CLIP) [44, 55] to explore any latent multimodal capabilities. We also test a thermal-specific model, FTNet [51], which uses an edge-based loss function to compensate for blurred boundaries in thermal images. + +Besides FastSCNN [53], EfficientViT [5], Segformer [74], DINOv2, and FT-Net, all network encoders in Tab. 2 feed into a DeepLabV3+ segmentation head. In particular, DINOv2 was employed with linear and non-linear, multi-scale + +heads in order to study the generalization capacity of its pretrained features to the thermal modality. All baselines except DINOv2 and ConvNext-B (CLIP) were trained from pretrained ImageNet weights. All networks, besides FTNet, were trained using the cross entropy loss. Further details on baseline implementation and training can be found in Supplementary Material Sec. S3.1. + +![](images/a2591cf958e101bfcfcee7b7f300ddc5ece351de40c0abfb93ca990a7e49c77d.jpg) +Fig. 4: Thermal images and semantic segmentation labels from each capture area with inference results from EfficientViT, FastSCNN, and ConvNext-B (CLIP). + +General benchmark We use the general (random) split (Sec. 3.3) to compare thermal semantic segmentation performance between baselines (Tab. 2). With field robotic applications in mind, we analyze segmentation results in context of GPU and CPU frame rates, which were measured onboard the Nvidia Jetson AGX Orin (GPU) and the Ruby NUC (CPU) embedded platforms, respectively2. + +In general, we found that larger models provide marginal benefit over smaller, compute-efficient models. In particular, a 0.027 mIoU difference between the largest (ConvNext-B) and fastest (FastSCNN) baselines comes at the cost of $130 \times$ more FLOPs. Overall, EfficientViT-B0 performed the best, attaining the highest mIoU (0.725) while performing $50\mathrm{Hz}$ inference on the Jetson AGX Orin. + +While large RGB foundation models show little gain over ImageNet pretraining, their features do extend to the thermal modality. Specifically, a frozen DINOv2 (linear head) outperforms the thermal-specific FT-Net. With a nonlinear head, it matches the performance of other end-to-end trained networks. While this shows the benefit of large-scale RGB pretraining, the choice of RGB pretraining data or learning strategy seems to matter. Notably, the frozen ConvNext-B (CLIP) model gives marginal gain over the frozen DINOv2 (linear head), despite a larger, nonlinear segmentation head. Overall, foundation models do not confer + +significant performance advantages for semantic segmentation on our dataset in order to justify their computational cost. + +Overall, all baselines can segment stuff classes well but none excel with the rare object classes: vehicle and person. Future models would likely need to consider weighted loss functions, smarter data augmentation, or domain adaptation techniques in order to improve on object classes. Lastly, we note that the thermal-specific model, FTNet, performs poorly, likely as its edge priors are less effective in unstructured, natural environments compared to urban environments. + +![](images/88dffea47f008d8f3fcc643d7b80116e98c6a3535a63ec3af6e34eadd7692487.jpg) + +![](images/959545bc483c2744d0a2b4ad7268e29fff5c154fb6de346b036a45c4f21ffd96.jpg) +Fig. 5: (a) Thermal semantic segmentation failures due to intra-class semantic variations when testing on geographically-partitioned, out-of-domain data. (b) Failures due to photometric variations between day (in-domain) and night (out-of-domain). + +Geographically-partitioned benchmarks We retrained our best baseline (Tab. 2) on geographically-partitioned sets to evaluate intra-thermal generalization. Given the cost and effort of collecting diverse aerial thermal data of field settings, we aim to determine the minimum amount of annotated data required to effectively deploy. Using EfficientViT-B0, we performed two experiments to study generalization across difference geographic regions and terrain types. + +Table 3: Thermal semantic segmentation results on geographically-split data +(a) Region-based split + +
Train AreaTest AreaAvg. mIoU
CANCKY
CA0.6660.0840.1930.314
NC0.0840.5080.0310.208
All0.6480.5250.5870.586
+ +(b) Terrain-based split + +
Training TerrainTesting TerrainAvg. mIoU
RiverLakeCoast
River0.6680.1960.1170.327
Lake0.1960.3640.0610.207
Coast0.1270.1190.5220.256
All0.6460.4150.4930.518
All + ground0.6580.4160.5200.532
+ +In the first experiment, we trained on thermal data from CA and tested on images from KY and NC (Tab. 3a). Despite common classes in each region, results show poor generalization to class variations across geographic areas. We see this again when repeating with the NC split. In the second experiment, we divided our data into three terrain types, aerial river, aerial lake, and aerial coast. Remaining ground-captured images were lumped into a fourth ground category. We trained and tested all pairwise combinations and found poor generalization across terrain types as well (Tab. 3b). Notably, we found low, in-domain performance for the lake setting but saw gains after training on aerial data from other terrain and further improvements after training on ground level images. Overall, our results reiterate the benefit of having training data representative of the testing environment, but also show that gains can be made by including out-of-domain (OOD) data from different terrain and different viewpoints. + +Overall, the geographic benchmark presents a difficult domain adaptation problem, requiring both inter- and intra-modality domain adaptation techniques to overcome. Although such techniques could perform well in this challenge, the results also highlight a need for an even larger and more diverse thermal dataset capturing common field environments than what we presented. + +Temporally-partitioned benchmark Here, we assess EfficientViT-B0's performance against intra-thermal, temporal distribution shift due to thermal inversion. We train the model on images from the daytime and test on images captured at sunrise and at night. The model achieved mIoUs of 0.242 at sunrise, 0.193 at night, and 0.777 on a separate daytime test split. Despite training on daytime images from the same location, the model fails on nighttime scenes, struggling to classify water due to thermal inversion (Fig. 5b). As such, to perform well on this split, future algorithms need to augment datasets to simulate such inversions or intentional collect data capturing such phenomena. + +Table 4: RGB-Thermal semantic segmentation network baseline results (mIoU). + +
ModelBare groundRocky terrainDevel. struct.RoadShrubsTreesSkyWaterVehiclesPersonAllFLOPS (G)
RGB Only†0.7900.8730.8080.5590.6640.7520.8910.9670.3990.0000.67045
Thermal Only†0.7660.8350.7890.5360.6540.7490.9070.9710.3990.0000.66145
EAEFNet [39]0.8050.8630.8130.5450.7020.8020.9370.9760.4810.0000.692358
CRM [60]0.8160.9000.8510.6160.6950.7790.9230.9740.3770.2060.714310
CMNeXt [79]0.8300.9000.8610.6070.7180.8080.9330.9800.5600.1900.740148
+ +$\dagger$ DeepLabV3+ (w/ ResNet50 encoder) + +# 4.2 RGB-T Semantic Segmentation + +We use the paired RGB-T dataset (Sec. 3.2) and partition it using the general split (Sec. 4.1). Recall that RGB and thermal input sizes are $960 \times 600$ pixels and the thermal FoV is narrower than in the previous benchmarks (Sec. 4.1). + +Baselines We test three RGB-T semantic segmentation algorithms achieving state-of-the-art results in urban RGB-T scene segmentation benchmarks: CRM [60], EAEFNet [39], and CMNeXt [79]. To compare against single-modality mod- + +Table 5: RGB-Thermal image translation results + +
MethodMethod TypeRGB → Thermal
PSNR ↑SSIM ↑mIoU ↑
UNIT [42]Unpaired GAN12.850.4210.189
MUNIT [27]Unpaired GAN16.080.4590.173
Edge-guided RGB-T [35]Unpaired GAN12.200.4360.225
Pix2Pix [28]Paired GAN20.060.5470.354
Pix2PixHD [73]Paired GAN20.050.5790.379
VQGAN [15]Paired GAN18.090.5400.388
Palette [58]Paired Diffusion11.510.3990.194
+ +els, we train two additional ResNet50/DeepLabV3+ segmentation models on RGB and thermal data accordingly (Sec. 4.1). + +Performance analysis All RGB-T models outperform the single-modality baselines, with CMNeXt outperforming the next best RGB-T model by 0.026 mIoU (Tab. 4). Overall, incorporating RGB imagery into the segmentation pipeline greatly improved the distinction between land-based stuff classes which can be hard to distinguish in thermal due to thermal crossover. However, these improvements come at a high computational cost and would not be suitable for field robotic applications. Excelling in this benchmark would involve improving the segmentation performance of the object classes, possibly by leveraging RGB data in unsupervised domain adaptation training schemes. + +# 4.3 RGB-Thermal Image Translation + +We use the RGB-T paired dataset from Sec. 4.2 to benchmark image translation algorithms operating in the RGB $\rightarrow$ T direction. + +Baselines We benchmark GAN methods that need RGB-T pairs (Pix2Pix [28], Pix2PixHD [73], VQGAN [15]) and unpaired methods (UNIT [42], MUNIT [27], [35]) that do not. We also evaluate Palette [58], a paired diffusion approach. + +![](images/a9a61df679ccbf54e3fb41a2d642ff4c809d26829049499726c6365b8fd2beef.jpg) +Fig. 6: RGB-to-thermal image translation results. Zoom-in to see fine-details. + +Image translation metrics We quantify RGB-T image translation using the Peak-Signal-to-Noise ratio (PSNR) and the Structural Similarity Index (SSIM). In addition, we propose a third metric: the thermal mIoU. This is equivalent to the FCN-score used in RGB image translation works [28]. Instead of an RGB network, our metric uses our EfficientViT thermal segmentation network (Tab. 2) to segment translated thermal images before evaluating with ground truth. + +Performance analysis Paired GANs outperform unpaired GANs and diffusion methods, achieving higher values across all three metrics (Tab. 5). Qualitative assessment (Fig. 6) shows poor translations from unpaired techniques, with most retaining geometric artifacts unique to RGB (ocean waves and shadows) and ignoring relative temperature characteristics. Paired GANs produce results similar to real thermal images but with inconsistent details upon closer inspection. Likewise, the diffusion method produces accurate translations but is not consistent. Overall, RGB-T translation still requires improvements before being able to provide a reliable source of thermal training data from existing RGB datasets. For domain adaptation purposes, we emphasize that maximizing the thermal mIoU score is more important than photometric consistency. + +# 4.4 Motion Tracking + +To quantify VIO/SLAM robustness, we select 12 clipped sequences from our dataset ranging in motion tracking difficulty: from urban sports fields (easy) to our natural environments (hard). We normalize thermal images using the $5^{\text{th}}$ and $95^{\text{th}}$ percentile pixel values (Sec. 3.2) to enable feature detection. + +Table 6: VIO/SLAM performance (Absolute Trajectory Error [m]) on aerial sequences + +
MethodTypeModalityNorth FieldCastaic Lake, CADuck, NC
N1N2N3N4N5N6C1C2C3C4D1D2
VINS-FusionSLAMRGB2.8837.5304.0362.0094.0450.8341.555—3.2771.5661.0310.700
VINS-FusionThermal9.45412.269.8542.92611.005.1147.2961.5135.3212.0250.879—
VINS-FusionVIORGB5.4589.5096.5182.1216.0261.2771.555—3.2771.3771.0310.725
VINS-FusionThermal13.8212.8611.992.92613.985.1947.28410.605.3212.0340.879—
Open-VINSVIORGB22.8036.1537.155.165—3.896—3.573——0.4750.700
Open-VINSThermal14.4330.6416.122.258—1.562—5.311——1.073—
Trajectory length (m)1533132612531531104642310651741378880
+ +Baselines We evaluate VINS-Fusion [54] (graph optimization-based) and OpenVINS [21] (filtering-based). We test VINS-Fusion in VIO and SLAM modes, with and without loop closure constraints, respectively, and report Absolute Trajectory Error [64] averaged over four runs. + +Performance analysis This benchmark reveals challenges due to fast motion at altitude and periodic feature sparsity in littoral environments. In such scenes, motion tracking frequently failed due to loss of features, requiring sequences to be clipped for quantifiable evaluations (Tab. 6). VINS-Fusion outperformed Open-VINS mainly due to higher feature tracking reliability. VINS-Fusion works better on RGB images than on thermal images in general. However, VINS-Fusion demonstrates superior feature tracking on thermal images compared to RGB on the C2 sequence of Castaic Lake, where the thermal contrast of stone grains on the lake coast was particularly pronounced in the late afternoon. This observation motivates future research on integrating thermal and RGB cameras for motion tracking with enhanced robustness and versatility. Our benchmark presents a difficult challenge for VIO and SLAM algorithms in extreme environments where water and reflections off water surfaces dominate image scenes, providing a unique test bed to invigorate interest in robust feature matching and motion tracking in natural texture-deficient scenarios. + +Table 7: RGB foundation model performances on thermal imagery + +(a) Instance segmentation via SAM on thermal and augmented color images using coregistered color image segmentations as ground truth + +
ModalityAPallAPsmallAPmedAPlarge
Thermal0.0180.0070.0220.204
Color (grayscale)0.5380.5360.5310.601
Color (col. jitter)0.8000.7920.8010.841
+ +$\dagger$ refers to AP@0.5::0.95, following MSCOCO [40] + +(b) Zero-shot semantic segmentation (mIoU) on coregistered color and thermal imagery + +
MethodColorColor (gray)Color (jitter)Thermal
SAM [33] + GT ann.‡0.7040.6980.7000.653
Grounded SAM [43]0.3800.3510.3610.193
OV-Seg [38]0.3620.3400.3650.217
ODISE [77]0.5040.4500.4860.232
+ +Most frequent ground truth class label in a SAM mask assigned as label for entire mask + +# 4.5 Further Analysis + +Zero-shot foundation models on thermal imagery Current foundation models work well on RGB imagery but have not been benchmarked on thermal, especially in non-urban settings. To evaluate these models on thermal imagery, specifically for zero-shot segmentation and semantic segmentation, we compare RGB and thermal performance using the aligned set (Sec. 3.2) in order to isolate modality as the root cause for performance differences. In the following tests, our metrics do not include the rare object classes (vehicle, person) as they drop metrics to near-uniform values for both modalities, making it difficult to compare. + +![](images/d9b6b21b6e513c5b08fe3a5be09c135904f6fb26b560d0765750b59e2ec7805b.jpg) +Fig. 7: Zero-shot foundation models on RGB-T pairs. Oversegmented SAM outputs differ across modalities but are similar with ground truth semantics added. Semantic segmentation models (prompted with all classes) perform better on RGB imagery. + +We quantify Segment Anything's (SAM) [33] zero-shot segmentation sensitivity to modality change by measuring the spatial alignment of its mask outputs. We apply it to the RGB set and the 8-bit thermal image set (Sec. 3.2) with the default $32 \times 32$ grid points to segment everything (Fig. 7). We compute the average precision (AP) of the predicted thermal masks, using the predicted RGB masks as ground truth, and compare the thermal AP to the APs of a color-jittered RGB set and a grayscale set. We find that SAM is sensitive to photometric changes, with large performance drops from color-jitter to grayscale to thermal (Tab. 7a). + +Next, we test large vision-language models (Grounded-SAM [43], OV-Seg [38], and ODISE [77]) in zero-shot semantic segmentation. We prompt the models + +with a list of classes in our dataset; generate semantic segmentation masks for RGB, color-jittered RGB, grayscale, and thermal image sets; and compute the mIoU of their outputs using our ground truth annotations. Although RGB, RGB-variants, and thermal modalities all yield poor semantic segmentation performance (Tab. 7b), all RGB variants outperform thermal by at least $0.13\mathrm{mIoU}$ indicating low generalization to the thermal modality (Fig. 7). + +Finally, we create semantic SAM masks by assigning class labels to SAM instances based on the most frequent ground truth class within an instance (Fig. 7). We find that these masks attain much higher mIoU score (0.653) compared to the other zero-shot methods. Based on the results (Tab. 7), we note that SAM can find object boundaries in the thermal domain reasonably well but can oversegment, resulting in poor AP scores when compared to RGB (which can also oversegment). When taken with ground truth semantics, this behavior is hidden. + +Transfer learning with urban thermal datasets We test if pretraining on thermal urban data can assist transfer to field settings. We sample 40k unlabeled thermal images from [11, 17, 22, 41, 61, 67, 71, 76] and pretrain various networks using Mo + +CoV2 [9]. We then finetune the pretrained networks on our training set and evaluate on our test set. Overall, we find sparse evidence suggesting any advantage of pretraining with existing urban thermal datasets (Tab. 8). Off-the-shelf ImageNet weights can provide strong performance in the thermal field domain, especially over training from scratch, without the effort and costs of pretraining. + +Table 8: Network pretraining methods for downstream thermal segmentation. + +
ModelmIoU# Params. (M)
NoneImageNetTherm. Urban
FastSCNN0.6400.6900.6881.1
EfficientViT0.6870.7250.7144.8
ResNet180.7020.7130.70612.3
ResNet500.6820.7130.72826.7
ConvNext-B0.6250.7170.69789.4
+ +# 5 Conclusion + +We presented the Caltech Aerial RGB-Thermal Dataset, the first dataset tailored towards advancing thermal semantic perception and motion tracking algorithms in natural environments. We created four benchmarks encompassing semantic segmentation, image translation, and motion tracking, and demonstrate challenges faced by current thermal perception and localization algorithms. Semantic segmentation and image translation methods were notably affected by geographical domain shifts, reflecting the wide intra-class and inter-scenery distributions in our dataset, and temporal domain shifts, due to thermal inversion and crossover. Motion tracking in thermal suffered from feature sparsity, a challenge commonly encountered in real-world robotic deployments [1]. Our motion tracking benchmark penalizes current VIO/SLAM algorithms for assuming ideal feature-tracking conditions, serving as a testbed to fuel improvements in this area. Our dataset and benchmarks can be used into the future to help develop algorithms that expand the operating domains of robotics and computer vision. + +# Acknowledgements + +This work was funded by the Office of Naval Research. We thank K. Koetje, K. Brodie, T. Hesser, T. Almeida, T. Cook, N. Spore, and R. Beach. + +# References + +1. Ackerman, E.: Blade strike on landing ends mars helicopter's epic journey (Jan 2024), https://spectrum.ieee.org/mars-helicopter-ingenuity-end-mission +2. Bloesch, M., Burri, M., Omari, S., Hutter, M., Siegwart, R.: Iterated extended kalman filter based visual-inertial odometry using direct photometric feedback. The International Journal of Robotics Research 36(10), 1053-1072 (2017) +3. Bondi, E., Jain, R., Aggrawal, P., Anand, S., Hannaford, R., Kapoor, A., Piavis, J., Shah, S., Joppa, L., Dilkina, B., et al.: Birdsai: A dataset for detection and tracking in aerial thermal infrared videos. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1747-1756 (2020) +4. Brodie, K.L., Bruder, B.L., Slocum, R.K., Spore, N.J.: Simultaneous mapping of coastal topography and bathymetry from a lightweight multicamera uas. IEEE Transactions on Geoscience and Remote Sensing 57(9), 6844-6864 (2019). https://doi.org/10.1109/TGRS.2019.2909026 +5. Cai, H., Gan, C., Han, S.: Efficientvit: Enhanced linear attention for high-resolution low-computation visual recognition. arXiv preprint arXiv:2205.14756 (2022) +6. Cai, W., Jin, K., Hou, J., Guo, C., Wu, L., Yang, W.: Vdd: Varied drone dataset for semantic segmentation. arXiv preprint arXiv:2305.13608 (2023) +7. Center, A.F.S.: A dataset for machine learning algorithm development from 2010-06-15 to 2010-08-15. NOAA National Centers for Environmental Information (2019) +8. Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 801-818 (2018) +9. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020) +0. Chen, Y., Wang, Y., Lu, P., Chen, Y., Wang, G.: Large-scale structure from motion with semantic constraints of aerial images. In: Pattern Recognition and Computer Vision: First Chinese Conference, PRCV 2018, Guangzhou, China, November 23-26, 2018, Proceedings, Part I 1. pp. 347-359. Springer (2018) +1. Choi, Y., Kim, N., Hwang, S., Park, K., Yoon, J.S., An, K., Kweon, I.S.: Kaist multi-spectral day/night data set for autonomous and assisted driving. IEEE Transactions on Intelligent Transportation Systems 19(3), 934-948 (2018) +2. Deevi, S.A., Lee, C., Gan, L., Nagesh, S., Pandey, G., Chung, S.J.: Rgb-x object detection via scene-specific fusion modules. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 7366-7375 (2024) +3. Delaune, J., Hewitt, R., Lytle, L., Sorice, C., Thakker, R., Matthies, L.: Thermal-inertial odometry for autonomous flight throughout the night. In: 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1122-1128. IEEE (2019) +4. Doer, C., Trommer, G.F.: Radar visual inertial odometry and radar thermal inertial odometry: Robust navigation even in challenging visual conditions. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 331-338. IEEE (2021) +5. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873-12883 (2021) +6. Feng, D., Haase-Schütz, C., Rosenbaum, L., Hertlein, H., Glaeser, C., Timm, F., Wiesbeck, W., Dietmayer, K.: Deep multi-modal object detection and semantic + +segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems 22(3), 1341-1360 (2020) +17. Teledyne fliradas dataset,https://www.flir.com/oem/adas/adas-dataset-form/,Lastaccessed on2023-10-27 +18. Fountas, S., Mylonas, N., Malounas, I., Rodias, E., Hellmann Santos, C., Pekkeriet, E.: Agricultural robotics for field operations. Sensors 20(9), 2672 (2020) +19. Gade, R., Moeslund, T.B.: Thermal cameras and applications: a survey. Machine vision and applications 25, 245-262 (2014) +20. Gan, L., Lee, C., Chung, S.J.: Unsupervised rgb-to-thermal domain adaptation via multi-domain attention network. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 6014-6020. IEEE (2023) +21. Geneva, P., Eckenhoff, K., Lee, W., Yang, Y., Huang, G.: Openvins: A research platform for visual-inertial estimation. In: 2020 IEEE International Conference on Robotics and Automation (ICRA). pp. 4666-4672. IEEE (2020) +22. Ha, Q., Watanabe, K., Karasawa, T., Ushiku, Y., Harada, T.: Mfnet: Towards real-time semantic segmentation for autonomous vehicles with multi-spectral scenes. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 5108-5115. IEEE (2017) +23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +24. He, M., Rajkumar, R.R.: Using thermal vision for extended vins-mono to localize vehicles in large-scale outdoor road environments. In: 2021 IEEE Intelligent Vehicles Symposium (IV). pp. 953-960. IEEE (2021) +25. Howard, A., Sandler, M., Chu, G., Chen, L.C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., et al.: Searching for mobilenetv3. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1314-1324 (2019) +26. Hua, T., Pei, L., Li, T., Wu, Q., Wang, R., Yu, W.: I2-slam: Fusing infrared camera and imu for simultaneous localization and mapping. In: International Conference on Autonomous Unmanned Systems. pp. 2834-2844. Springer (2021) +27. Huang, X., Liu, M.Y., Belongie, S., Kautz, J.: Multimodal unsupervised image-to-image translation. In: European Conference on Computer Vision (ECCV) (2018) +28. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1125-1134 (2017) +29. Jia, X., Zhu, C., Li, M., Tang, W., Zhou, W.: Llvip: A visible-infrared paired dataset for low-light vision. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3496-3504 (2021) +30. Jong, A., Yu, M., Dhrafani, D., Kailas, S., Moon, B., Sycara, K., Scherer, S.: Wit-uas: A wildland-fire infrared thermal dataset to detect crew assets from aerial views. In: 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 11464-11471. IEEE (2023) +31. Khattak, S., Papachristos, C., Alexis, K.: Keyframe-based direct thermal-inertial odometry. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 3563-3569. IEEE (2019) +32. Kim, Y.H., Shin, U., Park, J., Kweon, I.S.: Ms-uda: Multi-spectral unsupervised domain adaptation for thermal image semantic segmentation. IEEE Robotics and Automation Letters 6(4), 6497–6504 (Oct 2021). https://doi.org/10.1109/LRA.2021.3093652 + +33. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023) +34. Lee, C., Frennert, J.G., Gan, L., Anderson, M., Chung, S.J.: Online self-supervised thermal water segmentation for aerial vehicles. In: 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 7734-7741. IEEE (2023) +35. Lee, D.G., Jeon, M.H., Cho, Y., Kim, A.: Edge-guided multi-domain rgb-to-tir image translation for training vision tasks with challenging labels. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 8291-8298. IEEE (2023) +36. Li, C., Xia, W., Yan, Y., Luo, B., Tang, J.: Segmenting objects in day and night: Edge-conditioned cnn for thermal image semantic segmentation. IEEE Transactions on Neural Networks and Learning Systems 32(7), 3069-3082 (2020) +37. Li, H., Ma, Y., Huang, Y., Gu, Y., Xu, W., Liu, Y., Zuo, X.: Riders: Radar-infrared depth estimation for robust sensing. arXiv preprint arXiv:2402.02067 (2024) +38. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7061-7070 (2023) +39. Liang, M., Hu, J., Bao, C., Feng, H., Deng, F., Lam, T.L.: Explicit attention-enhanced fusion for rgb-thermal perception tasks. IEEE Robotics and Automation Letters (2023) +40. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +41. Liu, J., Fan, X., Huang, Z., Wu, G., Liu, R., Zhong, W., Luo, Z.: Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5802-5811 (2022) +42. Liu, M.Y., Breuel, T., Kautz, J.: Unsupervised image-to-image translation networks. In: Conference on Neural Information Processing Systems (NeurIPS) (2017) +43. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023) +44. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022) +45. Lyu, Y., Vosselman, G., Xia, G.S., Yilmaz, A., Yang, M.Y.: Uavid: A semantic segmentation dataset for uav imagery. ISPRS Journal of Photogrammetry and Remote Sensing 165, 108 - 119 (2020). https://doi.org/https://doi.org/10.1016/j.isprsjprs.2020.05.009, http://www.sciencedirect.com/science/article/pii/S0924271620301295 +46. Mehta, S., Rastegari, M.: Separable self-attention for mobile vision transformers. arXiv preprint arXiv:2206.02680 (2022) +47. Mostegel, C., Maurer, M., Heran, N., Pestana Puerta, J., Fraundorfer, F.: Semantic drone dataset (Jan 2019), http://dronedataset.icg.tugraz.at/, Last accessed on 2023-10-27 +48. Nigam, I., Huang, C., Ramanan, D.: Ensemble knowledge transfer for semantic segmentation. In: 2018 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 1499-1508. IEEE (2018) + +49. Nirgudkar, S., DeFilippo, M., Sacarny, M., Benjamin, M., Robinette, P.: Massmind: Massachusetts maritime infrared dataset. The International Journal of Robotics Research 42(1-2), 21-32 (2023) +50. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Howes, R., Huang, P.Y., Xu, H., Sharma, V., Li, S.W., Galuba, W., Rabbat, M., Assran, M., Ballas, N., Synnaeve, G., Misra, I., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2023) +51. Panetta, K., Shreyas Kamath, K.M., Rajeev, S., Agaian, S.S.: Ftnet: Feature transverse network for thermal image semantic segmentation. IEEE Access 9, 145212-145227 (2021). https://doi.org/10.1109/ACCESS.2021.3123066 +52. Park, T., Efros, A.A., Zhang, R., Zhu, J.Y.: Contrastive learning for unpaired image-to-image translation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16. pp. 319-345. Springer (2020) +53. Poudel, R., Liwicki, S., Cipolla, R.: Fast-scnn: Fast semantic segmentation network. In: Sidorov, K., Hicks, Y. (eds.) Proceedings of the British Machine Vision Conference (BMVC). pp. 187.1-187.12. BMVA Press (September 2019). https://doi.org/10.5244/C.33.187, https://dx.doi.org/10.5244/C.33.187 +54. Qin, T., Li, P., Shen, S.: Vins-mono: A robust and versatile monocular visual-inertial state estimator. IEEE Transactions on Robotics 34(4), 1004-1020 (2018) +55. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +56. Retief, S.P., Willers, C., Wheeler, M.: Prediction of thermal crossover based on imaging measurements over the diurnal cycle. In: Geo-Spatial and Temporal Image and Data Exploitation III. vol. 5097, pp. 58-69. SPIE (2003) +57. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022) +58. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022) +59. Saputra, M.R.U., Lu, C.X., de Gusmao, P.P.B., Wang, B., Markham, A., Trigoni, N.: Graph-based thermal-inertial slam with probabilistic neural networks. IEEE Transactions on Robotics 38(3), 1875-1893 (2022) +60. Shin, U., Lee, K., Kweon, I.S.: Complementary random masking for rgb-thermal semantic segmentation. In: IEEE International Conference on Robotics and Automation (2024) +61. Shin, U., Park, J., Kweon, I.S.: Deep depth estimation from thermal image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1043-1053 (2023) +62. Shivakumar, S.S., Rodrigues, N., Zhou, A., Miller, I.D., Kumar, V., Taylor, C.J.: Pst900: Rgb-thermal calibration, dataset and segmentation network. In: 2020 IEEE international conference on robotics and automation (ICRA). pp. 9441-9447. IEEE (2020) +63. Speth, S., Goncalves, A., Rigault, B., Suzuki, S., Bouazizi, M., Matsuo, Y., Prendinger, H.: Deep learning with rgb and thermal images onboard a drone for monitoring operations. Journal of Field Robotics 39(6), 840-868 (2022) + +64. Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for the evaluation of rgb-d slam systems. In: 2012 IEEE/RSJ international conference on intelligent robots and systems. pp. 573-580. IEEE (2012) +65. Sun, Y., Zuo, W., Liu, M.: Rtfnet: RGB-thermal fusion network for semantic segmentation of urban scenes. IEEE Robotics and Automation Letters 4(3), 2576-2583 (2019). https://doi.org/10.1109/LRA.2019.2904733 +66. Sun, Y., Zuo, W., Yun, P., Wang, H., Liu, M.: Fuseseg: Semantic segmentation of urban scenes based on rgb and thermal data fusion. IEEE Transactions on Automation Science and Engineering 18(3), 1000-1011 (2021). https://doi.org/10.1109/TASE.2020.2993143 +67. Suo, J., Wang, T., Zhang, X., Chen, H., Zhou, W., Shi, W.: Hit-uav: A high-altitude infrared thermal dataset for unmanned aerial vehicle-based object detection. Scientific Data 10, 227 (2023) +68. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105-6114. PMLR (2019) +69. Ustun, B., Kaya, A.K., Ayerden, E.C., Altinel, F.: Spectral transfer guided active domain adaptation for thermal imagery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. pp. 449-458 (June 2023) +70. Vargas, J., Alsweiss, S., Token, O., Razdan, R., Santos, J.: An overview of autonomous vehicles sensors and their vulnerability to weather conditions. Sensors 21(16), 5397 (2021) +71. Vertens, J., Zürn, J., Burgard, W.: Heatnet: Bridging the day-night domain gap in semantic segmentation with thermal images. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8461-8468. IEEE (2020) +72. VS, V., Poster, D., You, S., Hu, S., Patel, V.M.: Meta-uda: Unsupervised domain adaptive thermal object detection using meta-learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1412-1423 (January 2022) +73. Wang, T.C., Liu, M.Y., Zhu, J.Y., Tao, A., Kautz, J., Catanzaro, B.: High-resolution image synthesis and semantic manipulation with conditional gans. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2018) +74. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. In: Neural Information Processing Systems (NeurIPS) (2021) +75. Xie, S., Girshick, R., Dollar, P., Tu, Z., He, K.: Aggregated residual transformations for deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1492-1500 (2017) +76. Xiong, H., Cai, W., Liu, Q.: Mcnet: Multi-level correction network for thermal image semantic segmentation of nighttime driving scene. Infrared Physics & Technology p. 103628 (2021). https://doi.org/https://doi.org/10.1016/j.infrared.2020.103628 +77. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2955-2966 (2023) +78. Yun, S., Jung, M., Kim, J., Jung, S., Cho, Y., Jeon, M.H., Kim, G., Kim, A.: Stereo: Stereo thermal dataset for research in odometry and mapping. In: 2022 + +IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 3857-3864. IEEE (2022) +79. Zhang, J., Liu, R., Shi, H., Yang, K., Reif, S., Peng, K., Fu, H., Wang, K., Stiefelhagen, R.: Delivering arbitrary-modal semantic segmentation. In: CVPR (2023) +80. Zhao, S., Wang, P., Zhang, H., Fang, Z., Scherer, S.: Tp-tio: A robust thermal-inertial odometry with deep thermalpoint. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4505-4512. IEEE (2020) +81. Zhou, W., Liu, J., Lei, J., Yu, L., Hwang, J.N.: Gmnet: Graded-feature multilabel-learning network for rgb-thermal urban scene semantic segmentation. IEEE Transactions on Image Processing 30, 7790–7802 (2021). https://doi.org/10.1109/TIP.2021.3109518 +82. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Proceedings of the IEEE international conference on computer vision. pp. 2223-2232 (2017) +83. Zhu, P., Wen, L., Du, D., Bian, X., Fan, H., Hu, Q., Ling, H.: Detection and tracking meet drones challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(11), 7380-7399 (2021) \ No newline at end of file diff --git a/2024/Caltech Aerial RGB-Thermal Dataset in the Wild/images.zip b/2024/Caltech Aerial RGB-Thermal Dataset in the Wild/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..597d9f82b19f0ec71e044f8c0b5eb99d9761ca43 --- /dev/null +++ b/2024/Caltech Aerial RGB-Thermal Dataset in the Wild/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ede26a2f3a2da7da5d0fd6b8e6621d77d47d12f62970ee92a64d1c6d8bc5c0e1 +size 752457 diff --git a/2024/Caltech Aerial RGB-Thermal Dataset in the Wild/layout.json b/2024/Caltech Aerial RGB-Thermal Dataset in the Wild/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..52d33e7977cbd98adaca6c8b19e17271134a6cc7 --- /dev/null +++ b/2024/Caltech Aerial RGB-Thermal Dataset in the Wild/layout.json @@ -0,0 +1,9970 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 112, + 481, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 112, + 481, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 112, + 481, + 128 + ], + "type": "text", + "content": "Caltech Aerial RGB-Thermal Dataset in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 149, + 150, + 463, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 150, + 463, + 175 + ], + "spans": [ + { + "bbox": [ + 149, + 150, + 463, + 175 + ], + "type": "text", + "content": "Connor Lee*, Matthew Anderson*, Nikhil Ranganathan, Xingxing Zuo, Kevin Do, Georgia Gkioxari, and Soon-Jo Chung" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 237, + 184, + 376, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 184, + 376, + 195 + ], + "spans": [ + { + "bbox": [ + 237, + 184, + 376, + 195 + ], + "type": "text", + "content": "California Institute of Technology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 154, + 196, + 459, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 196, + 459, + 206 + ], + "spans": [ + { + "bbox": [ + 154, + 196, + 459, + 206 + ], + "type": "text", + "content": "{cee, matta, nrangana, zuox, kdo, georgia, sjchung}@caltech.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 233, + 455, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 233, + 455, + 388 + ], + "spans": [ + { + "bbox": [ + 159, + 233, + 455, + 388 + ], + "type": "text", + "content": "Abstract. We present the first publicly-available RGB-thermal dataset designed for aerial robotics operating in natural environments. Our dataset captures a variety of terrain across the United States, including rivers, lakes, coastlines, deserts, and forests, and consists of synchronized RGB, thermal, global positioning, and inertial data. We provide semantic segmentation annotations for 10 classes commonly encountered in natural settings in order to drive the development of perception algorithms robust to adverse weather and nighttime conditions. Using this dataset, we propose new and challenging benchmarks for thermal and RGB-thermal (RGB-T) semantic segmentation, RGB-T image translation, and motion tracking. We present extensive results using state-of-the-art methods and highlight the challenges posed by temporal and geographical domain shifts in our data. The dataset and accompanying code is available at https://github.com/aerorobotics/caltech-aerial-rgbt-dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 398, + 453, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 398, + 453, + 420 + ], + "spans": [ + { + "bbox": [ + 160, + 398, + 453, + 420 + ], + "type": "text", + "content": "Keywords: Robotics " + }, + { + "bbox": [ + 160, + 398, + 453, + 420 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 398, + 453, + 420 + ], + "type": "text", + "content": " Semantic segmentation " + }, + { + "bbox": [ + 160, + 398, + 453, + 420 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 398, + 453, + 420 + ], + "type": "text", + "content": " RGB-T image translation " + }, + { + "bbox": [ + 160, + 398, + 453, + 420 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 398, + 453, + 420 + ], + "type": "text", + "content": " Visual inertial odometry" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 133, + 456, + 224, + 545 + ], + "blocks": [ + { + "bbox": [ + 135, + 446, + 227, + 456 + ], + "lines": [ + { + "bbox": [ + 135, + 446, + 227, + 456 + ], + "spans": [ + { + "bbox": [ + 135, + 446, + 227, + 456 + ], + "type": "text", + "content": "Thermal Dataset Comparisons" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 456, + 224, + 545 + ], + "lines": [ + { + "bbox": [ + 133, + 456, + 224, + 545 + ], + "spans": [ + { + "bbox": [ + 133, + 456, + 224, + 545 + ], + "type": "image", + "image_path": "c66c1a82a96585488db1264a7a275c3e3743fbc07f341757e581bcdcc184c156.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 238, + 456, + 383, + 488 + ], + "blocks": [ + { + "bbox": [ + 238, + 447, + 369, + 456 + ], + "lines": [ + { + "bbox": [ + 238, + 447, + 369, + 456 + ], + "spans": [ + { + "bbox": [ + 238, + 447, + 369, + 456 + ], + "type": "text", + "content": "A. Thermal + RGB-T Semantic Segmentation" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 238, + 456, + 383, + 488 + ], + "lines": [ + { + "bbox": [ + 238, + 456, + 383, + 488 + ], + "spans": [ + { + "bbox": [ + 238, + 456, + 383, + 488 + ], + "type": "image", + "image_path": "acaa29bae9a83e7c9278a033abb99491c6761ad857f96e58994dbd3f761588bd.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 262, + 488, + 305, + 493 + ], + "lines": [ + { + "bbox": [ + 262, + 488, + 305, + 493 + ], + "spans": [ + { + "bbox": [ + 262, + 488, + 305, + 493 + ], + "type": "text", + "content": "Aligned RGB-T pairs" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 239, + 508, + 358, + 546 + ], + "blocks": [ + { + "bbox": [ + 238, + 497, + 347, + 506 + ], + "lines": [ + { + "bbox": [ + 238, + 497, + 347, + 506 + ], + "spans": [ + { + "bbox": [ + 238, + 497, + 347, + 506 + ], + "type": "text", + "content": "B. RGB-to-Thermal Image Translation" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 239, + 508, + 358, + 546 + ], + "lines": [ + { + "bbox": [ + 239, + 508, + 358, + 546 + ], + "spans": [ + { + "bbox": [ + 239, + 508, + 358, + 546 + ], + "type": "image", + "image_path": "d092be201dc124929f1cd40740c76ad1ac02d792f663a20c7faf0d7b90b5c546.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 549, + 482, + 583 + ], + "lines": [ + { + "bbox": [ + 130, + 549, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 549, + 482, + 583 + ], + "type": "text", + "content": "Fig. 1: Left: Our dataset is uniquely designed to improve thermal scene perception for field robots. Right: We provide new benchmarks for thermal-based vision algorithms, including (a) semantic segmentation, (b) image translation, and (c) motion tracking." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 386, + 461, + 481, + 544 + ], + "blocks": [ + { + "bbox": [ + 399, + 447, + 455, + 456 + ], + "lines": [ + { + "bbox": [ + 399, + 447, + 455, + 456 + ], + "spans": [ + { + "bbox": [ + 399, + 447, + 455, + 456 + ], + "type": "text", + "content": "C. Motion Tracking" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 386, + 461, + 481, + 544 + ], + "lines": [ + { + "bbox": [ + 386, + 461, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 386, + 461, + 481, + 544 + ], + "type": "image", + "image_path": "d04c086a5971a11d35af62916fb5bd6da8ce7ff746002449eacdb275aa97f438.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 603, + 230, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 603, + 230, + 615 + ], + "spans": [ + { + "bbox": [ + 132, + 603, + 230, + 615 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 622, + 481, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 622, + 481, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 622, + 481, + 647 + ], + "type": "text", + "content": "Field robots rely predominantly on visual cameras, lidar, and radar to perceive their surroundings [16, 18]. These sensors provide robust perception capabilities," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 337, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 337, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 337, + 665 + ], + "type": "text", + "content": "* These authors contributed equally to this work." + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 210 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 210 + ], + "type": "text", + "content": "but falter in low-light and adverse weather conditions [70]. In contrast, thermal cameras use long-wave infrared wavelengths to capture emitted heat, offering dense, visual information in such conditions where other sensors struggle [19]. Recently, they have been used for multimodal perception in autonomous driving and are increasingly being used to enable nighttime robot autonomy [11, 12, 36, 37, 49, 71, 76, 78]. However, successful integration of thermal cameras requires extensive datasets across diverse settings, preventing widespread adoption in field robotics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "type": "text", + "content": "Current thermal datasets focus on urban settings for autonomous driving applications (Fig. 1). They comprise of thermal-only imagery [36,76] or RGB-Thermal (RGB-T) pairs [17,22,41,71], with some including global positioning (GPS/GNSS) and inertial measurements (IMU) for visual-inertial odometry (VIO) and simultaneous localization and mapping (SLAM) [11,61,78]. While comprehensive, they lack data depicting natural settings like rivers and forests, which are typical operating areas for field robotics that perform coastline mapping and bathymetry [4] or activity monitoring during forest fires [30]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 309, + 482, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 482, + 440 + ], + "type": "text", + "content": "Due to the lack of thermal benchmarks for algorithms like semantic segmentation and SLAM, field robots cannot easily operate at night or in adverse conditions, requiring online learning [34, 72] or unsupervised domain adaptation to compensate [20, 32, 71], which still require labeled data for evaluation. As such, enabling thermal perception requires collecting and annotating field-specific data from scratch. However, creating a thermal dataset for field robotics is difficult due to two reasons: First, thermal data is hard to crowdsourced and web-scrape due to the richness of thermal cameras. Hence, acquiring a diverse dataset for algorithm development requires physically going to distinct field locations, increasing time and cost. Second, operating field robots, especially uninhabited aerial vehicles (UAV), usually requires special permits for each location-of-capture." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 441, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 482, + 559 + ], + "type": "text", + "content": "In this work, we present the first dataset targeting thermal scene perception for field robotics. The dataset consists of oblique-facing imagery captured from a UAV, supplemented with image sets captured at ground level, and focuses on littoral settings within various desert, forest, and coastal environments across the United States. We contribute new benchmarks for thermal and RGB-T semantic segmentation, RGB-T image translation, and VIO/SLAM algorithms, with unique challenges characterized by temporal and geographical domain shifts for learning-based methods and feature sparsity for motion-tracking algorithms. Aside from the proposed benchmarks, our dataset can also be used towards other applications like visual place recognition and terrain-relative navigation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 561, + 481, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 561, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 561, + 481, + 597 + ], + "type": "text", + "content": "The paper is organized as follows: Sec. 2 reviews relevant datasets, benchmarks, and algorithms. Sec. 3 describes our dataset details and curation process, and Sec. 4 presents benchmark results. Sec. 5 offers concluding remarks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 605, + 238, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 238, + 618 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 238, + 618 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 627, + 292, + 638 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 627, + 292, + 638 + ], + "spans": [ + { + "bbox": [ + 131, + 627, + 292, + 638 + ], + "type": "text", + "content": "2.1 Datasets and Benchmarks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Thermal/RGB-T benchmarks for urban robotics Current urban thermal and RGB-T datasets primarily focus on autonomous vehicle (AV) technology" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "C. Lee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 153, + 480, + 228 + ], + "blocks": [ + { + "bbox": [ + 132, + 120, + 480, + 152 + ], + "lines": [ + { + "bbox": [ + 132, + 120, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 120, + 480, + 152 + ], + "type": "text", + "content": "Table 1: Comparison of datasets captured from aerial platforms. Camera angle of " + }, + { + "bbox": [ + 132, + 120, + 480, + 152 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 132, + 120, + 480, + 152 + ], + "type": "text", + "content": " is nadir-pointing. See Tab. S1 in supplement for comparison against thermal datasets captured from surface and ground vehicles." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 153, + 480, + 228 + ], + "lines": [ + { + "bbox": [ + 132, + 153, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 153, + 480, + 228 + ], + "type": "table", + "html": "
DatasetTaskRGBThermalGPSIMUSettingLocation# SamplesCamera pose
VisDrone [83]Obj. Det.✓XXXUrbanChina10,209Variable
HIT-UAV [67], NII-CU [63]Obj. Det.✓✓XXUrbanChina, Japan2,898 - 5,88030°, 90°, 60-130 m
WTT-UAS [30]Obj. Det.X✓✓✓ForestUSA6,95130°, 90°, < 120 m
BIRDSAI [3]Obj. Track.X✓XXSavannasS. Africa62,000Off-nadir, 60-120 m
AeroScapes [48], UAVid [45], [10], [6], Swiss/Okutama [63]Sem. Seg.✓XXXUrbanChina, Japan, Switzerland191 - 326930 - 90°, < 100 m
Semantic Drone Dataset [47]Sem. Seg.✓✓X✓UrbanGermany40090°, < 30m
OursSem. Seg.✓✓✓✓River, Coast, Lake, Mountain, DesertUSA4,19520°, 45°, < 120 m
", + "image_path": "c28f3f6765115e48ce8a0c421d7fe7c592d232d413cb16be8c7f2798c517db91.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 237, + 480, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 480, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 480, + 333 + ], + "type": "text", + "content": "and surveillance applications. AV-related datasets cover tasks such as object detection [11, 17, 41], semantic segmentation [22, 36, 71, 76], and localization [11, 61, 78]. Some include data from other spectra, benchmarking not only thermal algorithms but also RGB-T [11, 22, 41, 63, 71] and other multispectral models. In contrast, datasets for surveillance applications use fixed cameras [29] and UAVs [63, 67] to detect personnel, vehicles, and other objects. Despite many urban thermal benchmarks, they cannot be directly used to develop algorithms for robots operating in natural environments due to the domain gap." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 338, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 540 + ], + "type": "text", + "content": "Thermal/RGB-T benchmarks for field robotics Datasets for robots operating in non-urban environments remain limited. MassMIND [49] and PST900 [62] benchmark thermal semantic segmentation for maritime and subterranean robotic environments, respectively. WIT-UAS [30] provides UAV-borne thermal images for object detection in wildfire-prone environments. One work [34] proposes an aerial RGB-T dataset for thermal water segmentation in littoral areas but only provides 1272 labels consisting of a single water class. Benchmarks like [3,7] depict natural environments but are targeted towards wildlife tracking. To address this gap, we present an aerial RGB-T dataset which can be used to benchmark multiple tasks, including semantic segmentation, RGB-T image translation, and localization, across diverse natural environment robotic applications. Specifically, we build upon the dataset from [34] by extending the number of semantic classes from 1 to 10 and adding additional RGB-T data captured from air and ground, resulting in a total of 4195 semantic segmentation annotations. In comparison to existing thermal/RGB-T datasets, our dataset can enable nighttime field robotic applications like remote sensing research, geological sampling, precision agriculture, and wilderness search and rescue." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 546, + 482, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 618 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 618 + ], + "type": "text", + "content": "UAV-based semantic segmentation datasets Aside from [34], existing UAV-captured semantic segmentation datasets [6, 10, 45, 47, 48, 63] cover urban environments (Tab. 1). With the exception of [47] which provides 400 thermal image samples, all of these datasets contain only RGB data for semantic segmentation. While they provide images with relevant view angles, the modality gap and focus on urban content makes them impractical for aerial field robotic applications." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 625, + 340, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 625, + 340, + 639 + ], + "spans": [ + { + "bbox": [ + 131, + 625, + 340, + 639 + ], + "type": "text", + "content": "2.2 Algorithms using Thermal Datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "Thermal semantic segmentation Since thermal images are spatially identical to their RGB counterparts, popular RGB semantic segmentation algo" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 243, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 243, + 91, + 447, + 101 + ], + "type": "text", + "content": "Caltech Aerial RGB-Thermal Dataset in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "rithms [5, 8, 53, 74] can be applied directly. However, some works [36, 51, 76] develop models specifically for thermal imagery, and leverage edge priors to overcome low resolution and blurriness due to thermal crossover [56]. Although specifically designed for thermal, these models only marginally outperform standard RGB segmentation models [51] and their performance in field settings, where edges are less structured, remains untested. Aside from architectural advancements, improvements in thermal semantic segmentation come from domain adaptation (DA) methods [20, 36, 69, 71] that use labeled RGB data with unlabeled thermal data. To provide fair comparisons, we do not provide DA baselines in our benchmarks as performance varies based on the choice of RGB data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 239, + 483, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 483, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 483, + 360 + ], + "type": "text", + "content": "RGB-T semantic segmentation RGB-T semantic segmentation methods compensate for the weaknesses of each modality via deep feature fusion, data augmentation, and adversarial training. Most approaches [22, 60, 65, 66, 79] utilize encoder-decoder architectures with modality-specific encoders and shared decoders, and show significant improvements over channel-stacked RGB-T inputs. Works like [39, 81] make improvements by integrating RGB and thermal features at multiple stages within the encoders, while [62] eschews RGB-T training pairs by processing RGB-T inputs sequentially. Other methods include random input masking to reduce reliance on a single modality [60] and adversarial training to adapt to different times of day [71]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 445 + ], + "type": "text", + "content": "RGB-T image translation RGB-T image translation is vital as it may enable thermal training data to be generated from large-scale RGB datasets. Works based on generative adversarial networks (GAN) such as [27,35,42,52,82] can be used without RGB-T image pairs. However, GANs like [28,73] that require paired imagery provide better translations for both visual appeal and domain adaptation [36]. Recent diffusion methods [57,58] have demonstrated remarkable results on RGB imagery, but have not been tested on RGB-T image translation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "text", + "content": "Thermal VIO and SLAM Low signal-to-noise ratio, reduced contrast, and blurred boundaries in thermal images present challenges for VIO and SLAM. Several studies [2, 13, 14, 24, 26] use indirect methods by tracking keypoints extracted from thermal images for visual constraints and rely on image normalization techniques to improve keypoint extraction. In contrast, [31] uses a direct method to extract and track high-gradient points by minimizing radiometric error. Other works utilize learned local and global features to improve feature tracking [80] and loop closure detection [59], respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 552, + 406, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 552, + 406, + 565 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 406, + 565 + ], + "type": "text", + "content": "3 The Caltech Aerial RGB-Thermal Dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 573, + 300, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 573, + 300, + 585 + ], + "spans": [ + { + "bbox": [ + 131, + 573, + 300, + 585 + ], + "type": "text", + "content": "3.1 Data Acquisition Hardware" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": "To capture this dataset, we developed a custom sensor stack with synchronized RGB-thermal imagery, IMU, and global positioning data (Fig. 2). The sensor stack features a non-radiometric FLIR ADK thermal camera (" + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "inline_equation", + "content": "640 \\times 512" + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": " px and " + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "inline_equation", + "content": "75^{\\circ}" + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": " horizontal FoV, " + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "inline_equation", + "content": "60\\mathrm{Hz}" + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": ") flanked by a pair of FLIR Blackfly electro-optical (EO) monochrome and color cameras" + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "inline_equation", + "content": "960 \\times 600" + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": " px and " + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "inline_equation", + "content": "75^{\\circ}" + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": " horizontal FoV, " + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "inline_equation", + "content": "30\\mathrm{Hz}" + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "C. Lee et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 652, + 405, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 652, + 405, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 652, + 405, + 665 + ], + "type": "text", + "content": "1 BFLY-U3-23S6M-C mono/color with Tamron M112FM08 lenses" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "Pose information is provided by a VectorNav VN100 IMU (200 Hz) and a u-blox M8N GPS (5 Hz). All three cameras and the VN100 are hardware synchronized using a dedicated signal generator and are rigidly attached to each other using a stiff, 3D printed mount. A Simply NUC Ruby Mini PC (AMD Ryzen 7 4800U, 32 GB RAM) is used for the compute, and the entire system is interfaced using ROS Noetic. Our cameras and IMU were calibrated following the procedures found in Supplementary Material Sec. S2.2 and were not disassembled from each other between any of the dataset collections." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 138, + 219, + 248, + 371 + ], + "blocks": [ + { + "bbox": [ + 138, + 219, + 248, + 371 + ], + "lines": [ + { + "bbox": [ + 138, + 219, + 248, + 371 + ], + "spans": [ + { + "bbox": [ + 138, + 219, + 248, + 371 + ], + "type": "image", + "image_path": "bfbf75378943dbb397f1282aab309c5ea349027af3482b5447f284aaca0de067.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 376, + 482, + 398 + ], + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 398 + ], + "type": "text", + "content": "Fig. 2: (a) The Aurelia X6 hexacopter and sensor stack used to capture our aerial dataset. (b) Geographic distribution of our data collection sites and collection times." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 255, + 219, + 475, + 371 + ], + "blocks": [ + { + "bbox": [ + 255, + 219, + 475, + 371 + ], + "lines": [ + { + "bbox": [ + 255, + 219, + 475, + 371 + ], + "spans": [ + { + "bbox": [ + 255, + 219, + 475, + 371 + ], + "type": "image", + "image_path": "8c39ed1a4de2c64a999a423b96d24ac4a04788514abf50adee8c5fb6bd87997a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 409, + 482, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 409, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 482, + 493 + ], + "type": "text", + "content": "The sensor stack is mounted rigidly onboard an Aurelia X6 hexacopter (Fig. 2) without gimbal stabilization at angles between " + }, + { + "bbox": [ + 130, + 409, + 482, + 493 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 130, + 409, + 482, + 493 + ], + "type": "text", + "content": " (level with the horizon) to " + }, + { + "bbox": [ + 130, + 409, + 482, + 493 + ], + "type": "inline_equation", + "content": "45^{\\circ}" + }, + { + "bbox": [ + 130, + 409, + 482, + 493 + ], + "type": "text", + "content": " downwards to provide different viewpoints. When mounted to a UAV, we also log the position and attitude estimates from the UAV at " + }, + { + "bbox": [ + 130, + 409, + 482, + 493 + ], + "type": "inline_equation", + "content": "20\\mathrm{Hz}" + }, + { + "bbox": [ + 130, + 409, + 482, + 493 + ], + "type": "text", + "content": " providing improved positioning accuracy from the onboard RTK GPS. Additionally, the sensor stack's modularity allows us to capture datasets on foot by mounting it to a tripod where flight restrictions are in-place." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 501, + 320, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 501, + 320, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 501, + 320, + 514 + ], + "type": "text", + "content": "3.2 Data Collection and Processing" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 517, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 482, + 601 + ], + "type": "text", + "content": "Data capture We captured 37 aerial and ground trajectories covering lakes, rivers, coastlines, deserts, and mountains from around Southern California, Kentucky, and North Carolina (Fig. 2). Aerial trajectories (18) involve high motion with intermittent hovering, and mostly depict scenes from " + }, + { + "bbox": [ + 130, + 517, + 482, + 601 + ], + "type": "inline_equation", + "content": "40\\mathrm{m}" + }, + { + "bbox": [ + 130, + 517, + 482, + 601 + ], + "type": "text", + "content": " altitude. Ground-level trajectories (19) were captured on foot, with 7 captured without any movement for VIO debugging purposes. For more details, see Supplementary Materials Tab. S2 and Sec. S2.1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Thermal image normalization For the baselines presented in this work, we typically normalize 16-bit thermal data by rescaling between the " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "1^{\\mathrm{st}}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "99^{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " percentile pixel values and follow with contrast limited adaptive histogram equalization (CLAHE). We keep normalized data as floats to minimize information loss, and only convert to 8-bit format when necessary or for visualization." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 243, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 243, + 91, + 448, + 101 + ], + "type": "text", + "content": "Caltech Aerial RGB-Thermal Dataset in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 125, + 315, + 228 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 298, + 123 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 298, + 123 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 298, + 123 + ], + "type": "text", + "content": "Semantic Segmentation Classes and Description" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 140, + 125, + 315, + 228 + ], + "lines": [ + { + "bbox": [ + 140, + 125, + 315, + 228 + ], + "spans": [ + { + "bbox": [ + 140, + 125, + 315, + 228 + ], + "type": "image", + "image_path": "762c02a09c796d13493e5be1f26e0262c19945d73ac2f8deb360b8dab5b20328.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 330, + 118, + 475, + 175 + ], + "blocks": [ + { + "bbox": [ + 322, + 118, + 329, + 127 + ], + "lines": [ + { + "bbox": [ + 322, + 118, + 329, + 127 + ], + "spans": [ + { + "bbox": [ + 322, + 118, + 329, + 127 + ], + "type": "text", + "content": "a" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 330, + 118, + 475, + 175 + ], + "lines": [ + { + "bbox": [ + 330, + 118, + 475, + 175 + ], + "spans": [ + { + "bbox": [ + 330, + 118, + 475, + 175 + ], + "type": "image", + "image_path": "026bac4764f8e46486c4665e30fc64344aeeba0305af094380eedf3b0642642b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 169, + 329, + 177 + ], + "lines": [ + { + "bbox": [ + 323, + 169, + 329, + 177 + ], + "spans": [ + { + "bbox": [ + 323, + 169, + 329, + 177 + ], + "type": "text", + "content": "b" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 330, + 172, + 468, + 228 + ], + "blocks": [ + { + "bbox": [ + 330, + 172, + 468, + 228 + ], + "lines": [ + { + "bbox": [ + 330, + 172, + 468, + 228 + ], + "spans": [ + { + "bbox": [ + 330, + 172, + 468, + 228 + ], + "type": "image", + "image_path": "dfe213dcb7b6976c69c56621d6fea9ca1dbc70f616b4a21e7c7b5d2eeb685d65.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 233, + 481, + 266 + ], + "lines": [ + { + "bbox": [ + 131, + 233, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 131, + 233, + 481, + 266 + ], + "type": "text", + "content": "Fig. 3: Semantic segmentation classes in our dataset. The color mapping is used throughout this paper. (a) Hourly distribution of annotated thermal images. (b) Histogram of semantic classes." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 277, + 480, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 480, + 350 + ], + "type": "text", + "content": "Semantic segmentation annotation To aid development of scene perception algorithms for field robots, we manually-annotated a subset of thermal images for semantic segmentation with classes in Fig. 3. To avoid redundancy, images were sampled every 3s when moving and every 20s when still, resulting in 4195 samples. We considered a frame in motion if the distance moved within the past 2 seconds exceeded 0.5 meters, with distances computed using GPS coordinates." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 352, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 449 + ], + "type": "text", + "content": "To promote labeling consistency, annotators preread descriptions of each trajectory, which contained a list of classes present, labeled examples provided by the authors, and GPS coordinates (for Google Earth look-up). Annotators were then given an RGB-T image pair with trajectory metadata and asked to label the thermal frame. Annotations were reviewed by authors, with inconsistent or incorrect labels sent back. 3 rounds of annotation were done, with the final and initial set of labels differing by " + }, + { + "bbox": [ + 130, + 352, + 482, + 449 + ], + "type": "inline_equation", + "content": "0.08\\mathrm{mIoU}" + }, + { + "bbox": [ + 130, + 352, + 482, + 449 + ], + "type": "text", + "content": ". Classes most frequently re-annotated were bare ground, rocky terrain, road, trees, and vehicles." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 453, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 538 + ], + "type": "text", + "content": "RGB-thermal image alignment To align thermal and RGB image pairs, we stereo rectified the RGB-T image pairs using the camera matrices obtained from calibration (Sec. S2.2) and projected the thermal image into the larger RGB image frame to preserve RGB resolution. As our baseline is small compared to the depth of the scenes we capture, image pairs are near-coregistered after rectification. Due to differences in optics, the field-of-view of the aligned image pairs is narrower compared to that of the thermal camera." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 546, + 233, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 546, + 233, + 559 + ], + "spans": [ + { + "bbox": [ + 131, + 546, + 233, + 559 + ], + "type": "text", + "content": "3.3 Dataset Splits" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 563, + 481, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 481, + 601 + ], + "type": "text", + "content": "General (benchmark) split This is our primary split for semantic segmentation and image translation benchmarks. We randomly partition the annotated thermal dataset (Sec. 3.2) into train/val/test sets at a 75:12.5:12.5 ratio." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Temporal split To promote studies into the effect of different time-of-day capture of the thermal images, we split the dataset using three time periods: twilight/sunrise (5 AM - 7 AM), daytime (10 AM - 5 PM), and nighttime (7 PM - 4 AM). We randomly partition the daytime images into train/val/test sets using the ratio from above and leave the twilight and nighttime sets for testing." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "C. Lee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "content": "Geographical split The geographical splits are intended to test algorithm adaptability to unseen settings. We provide two splits:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 149, + 480, + 221 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 135, + 149, + 479, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 149, + 479, + 185 + ], + "spans": [ + { + "bbox": [ + 135, + 149, + 479, + 185 + ], + "type": "text", + "content": "1. Terrain-based: This split partitions data into the following terrain categories: aerial river, aerial lake, aerial coast, and ground-captured images. Train/val/test splits are created per category with a 50:15:35 ratio." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 186, + 480, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 186, + 480, + 221 + ], + "spans": [ + { + "bbox": [ + 135, + 186, + 480, + 221 + ], + "type": "text", + "content": "2. Region-based: This is based on the area of the United States where the data was captured. Distinct regions include California (CA), Kentucky (KY), and North Carolina (NC), with train/val/test created in the same way as above." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 232, + 230, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 232, + 230, + 245 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 230, + 245 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 251, + 479, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 479, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 479, + 299 + ], + "type": "text", + "content": "We provide baselines for the following benchmarks: thermal semantic segmentation, RGB-T semantic segmentation, RGB-T image translation, and VIO/SLAM. Furthermore, we explore the zero-shot thermal segmentation capabilities of large vision-language models on our dataset and perform further ablation studies." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 132, + 336, + 480, + 474 + ], + "blocks": [ + { + "bbox": [ + 132, + 313, + 479, + 335 + ], + "lines": [ + { + "bbox": [ + 132, + 313, + 479, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 313, + 479, + 335 + ], + "type": "text", + "content": "Table 2: Supervised model performance (mIoU), framerate (FPS at batch size 1), and floating point operations (FLOP) for various semantic segmentation network baselines." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 336, + 480, + 474 + ], + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 474 + ], + "type": "table", + "html": "
ModelBare groundRocky terrainDevelop. struct.RoadShrubsTreesSkyWaterVehiclesPersonStuffObjectAllOrin (GPU)NUC (CPU)FLOPS (G)
FastSCNN [53]0.8070.9070.7280.6310.6940.7760.9560.9650.3840.0580.8080.2210.690124.285.61.0
MobileNet-V3-S 0.75 [25]0.8020.9140.6760.6330.7040.8140.9550.9760.3830.1110.8090.2470.69775.224.03.3
MobileNet-V3-L 0.750.8010.9140.7010.6150.7130.7930.9590.9760.4320.1480.8090.2900.70559.015.54.8
MobileViTV2 0.50 [46]0.8220.9130.7590.6500.6940.7810.9540.9700.2220.0300.8180.1260.68060.110.75.5
EfficientViT-B0 [5]0.8250.9170.7770.6180.7100.7930.9610.9650.4950.1910.8210.3430.72551.25.53.9
EfficientNet-Lite0 [68]0.8190.9080.7690.6010.7090.8050.9620.9680.3800.0790.8180.2290.70047.712.87.2
EfficientNet-Lite20.8250.9200.7650.6140.6980.8050.9580.9710.3860.1180.8200.2520.70638.911.29.4
Segformer-B0 [74]0.8040.9100.6900.6240.6960.7880.9600.9700.1950.0000.8050.0970.66438.35.910.2
Segformer-B10.8140.9090.7730.6030.7130.7990.9600.9650.3660.0000.8170.1830.69030.33.519.9
ResNet18 [23]0.8100.9050.7660.6180.7110.8070.9590.9660.4310.1580.8180.2940.71369.312.422.4
ResNet500.8190.9160.7470.6580.7110.7990.9610.9740.3610.1820.8230.2720.71329.96.145.4
ResNeXt50 [75]0.8250.9190.7730.6530.7090.7940.9640.9760.4360.1370.8270.2870.71923.05.645.5
ConvNext-T [44]0.7990.9090.7190.6080.7060.8080.9610.9700.3630.0030.8100.1830.68525.83.347.4
ConvNext-S0.8100.9130.7710.6030.7180.8100.9650.9640.4920.0480.8190.2700.70919.12.375.0
ConvNext-B0.8100.9210.6970.6110.7060.8030.9630.9720.3480.1120.8110.2300.69414.61.4130.5
ConvNext-B (CLIP) [55]0.8130.9180.6830.6320.7180.8130.9650.9720.5170.1370.8140.3270.71714.61.4130.4
ConvNext-B (CLIP) *0.7730.8870.7130.4870.6560.7560.9460.9510.1580.0190.7710.0890.63514.61.4130.4
DINOv2 [50] (linear head)0.8000.9070.6810.6060.6930.7960.9560.9670.3750.0790.8010.2270.68615.22.1—
DINOv2 (linear head) *0.7050.8440.6190.4190.5580.7250.9220.9200.3560.1580.7140.2570.62315.22.1—
DINOv2 (nonlin. head)0.8100.9160.7080.6350.7000.8190.9590.9730.3990.1190.8150.2590.70415.22.1—
DINOv2 (nonlin. head) *0.7960.9030.7320.6060.6910.7940.9540.9630.4710.1490.8050.3100.70615.22.1—
FTNet† [51]0.7550.8670.6350.5760.6430.6530.7870.9470.2340.0240.7330.1290.61311.11.4100.5
", + "image_path": "d735ba5de4554b9e51e03a6142064d20f32236d745eec9fb142908ab89ceaff9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 227, + 477, + 387, + 486 + ], + "lines": [ + { + "bbox": [ + 227, + 477, + 387, + 486 + ], + "spans": [ + { + "bbox": [ + 227, + 477, + 387, + 486 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 227, + 477, + 387, + 486 + ], + "type": "text", + "content": " Thermal-specific model * Frozen encoder" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 504, + 326, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 326, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 326, + 515 + ], + "type": "text", + "content": "4.1 Thermal Semantic Segmentation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 518, + 479, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 518, + 479, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 518, + 479, + 554 + ], + "type": "text", + "content": "We run multiple baselines on our general split before choosing a specific model to test on the geographic and temporal splits. We go over our baselines before analyzing their performance on our benchmarks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 557, + 480, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 480, + 628 + ], + "type": "text", + "content": "Baselines Motivated by robotic applications, we are interested in both inference speed and segmentation performance. Our baselines place an emphasis on lightweight, real-time networks, but also on foundation models like DINOv2 [50] and ConvNext-B (CLIP) [44, 55] to explore any latent multimodal capabilities. We also test a thermal-specific model, FTNet [51], which uses an edge-based loss function to compensate for blurred boundaries in thermal images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": "Besides FastSCNN [53], EfficientViT [5], Segformer [74], DINOv2, and FT-Net, all network encoders in Tab. 2 feed into a DeepLabV3+ segmentation head. In particular, DINOv2 was employed with linear and non-linear, multi-scale" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 243, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 243, + 91, + 448, + 101 + ], + "type": "text", + "content": "Caltech Aerial RGB-Thermal Dataset in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "heads in order to study the generalization capacity of its pretrained features to the thermal modality. All baselines except DINOv2 and ConvNext-B (CLIP) were trained from pretrained ImageNet weights. All networks, besides FTNet, were trained using the cross entropy loss. Further details on baseline implementation and training can be found in Supplementary Material Sec. S3.1." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 138, + 184, + 477, + 361 + ], + "blocks": [ + { + "bbox": [ + 138, + 184, + 477, + 361 + ], + "lines": [ + { + "bbox": [ + 138, + 184, + 477, + 361 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 477, + 361 + ], + "type": "image", + "image_path": "a2591cf958e101bfcfcee7b7f300ddc5ece351de40c0abfb93ca990a7e49c77d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 365, + 482, + 389 + ], + "lines": [ + { + "bbox": [ + 131, + 365, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 131, + 365, + 482, + 389 + ], + "type": "text", + "content": "Fig. 4: Thermal images and semantic segmentation labels from each capture area with inference results from EfficientViT, FastSCNN, and ConvNext-B (CLIP)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 402, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 462 + ], + "type": "text", + "content": "General benchmark We use the general (random) split (Sec. 3.3) to compare thermal semantic segmentation performance between baselines (Tab. 2). With field robotic applications in mind, we analyze segmentation results in context of GPU and CPU frame rates, which were measured onboard the Nvidia Jetson AGX Orin (GPU) and the Ruby NUC (CPU) embedded platforms, respectively2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "type": "text", + "content": "In general, we found that larger models provide marginal benefit over smaller, compute-efficient models. In particular, a 0.027 mIoU difference between the largest (ConvNext-B) and fastest (FastSCNN) baselines comes at the cost of " + }, + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "type": "inline_equation", + "content": "130 \\times" + }, + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "type": "text", + "content": " more FLOPs. Overall, EfficientViT-B0 performed the best, attaining the highest mIoU (0.725) while performing " + }, + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "type": "inline_equation", + "content": "50\\mathrm{Hz}" + }, + { + "bbox": [ + 130, + 464, + 482, + 525 + ], + "type": "text", + "content": " inference on the Jetson AGX Orin." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 527, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 527, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 527, + 482, + 624 + ], + "type": "text", + "content": "While large RGB foundation models show little gain over ImageNet pretraining, their features do extend to the thermal modality. Specifically, a frozen DINOv2 (linear head) outperforms the thermal-specific FT-Net. With a nonlinear head, it matches the performance of other end-to-end trained networks. While this shows the benefit of large-scale RGB pretraining, the choice of RGB pretraining data or learning strategy seems to matter. Notably, the frozen ConvNext-B (CLIP) model gives marginal gain over the frozen DINOv2 (linear head), despite a larger, nonlinear segmentation head. Overall, foundation models do not confer" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "C. Lee et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 482, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 482, + 666 + ], + "type": "text", + "content": "2 Models of input size " + }, + { + "bbox": [ + 133, + 642, + 482, + 666 + ], + "type": "inline_equation", + "content": "512 \\times 640" + }, + { + "bbox": [ + 133, + 642, + 482, + 666 + ], + "type": "text", + "content": " were converted to ONNX and inference speed was averaged over 50 forward passes after a warm-up period." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "significant performance advantages for semantic segmentation on our dataset in order to justify their computational cost." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 156, + 480, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 156, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 156, + 480, + 228 + ], + "type": "text", + "content": "Overall, all baselines can segment stuff classes well but none excel with the rare object classes: vehicle and person. Future models would likely need to consider weighted loss functions, smarter data augmentation, or domain adaptation techniques in order to improve on object classes. Lastly, we note that the thermal-specific model, FTNet, performs poorly, likely as its edge priors are less effective in unstructured, natural environments compared to urban environments." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 143, + 239, + 476, + 345 + ], + "blocks": [ + { + "bbox": [ + 143, + 239, + 476, + 345 + ], + "lines": [ + { + "bbox": [ + 143, + 239, + 476, + 345 + ], + "spans": [ + { + "bbox": [ + 143, + 239, + 476, + 345 + ], + "type": "image", + "image_path": "88dffea47f008d8f3fcc643d7b80116e98c6a3535a63ec3af6e34eadd7692487.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 143, + 346, + 476, + 443 + ], + "blocks": [ + { + "bbox": [ + 143, + 346, + 476, + 443 + ], + "lines": [ + { + "bbox": [ + 143, + 346, + 476, + 443 + ], + "spans": [ + { + "bbox": [ + 143, + 346, + 476, + 443 + ], + "type": "image", + "image_path": "959545bc483c2744d0a2b4ad7268e29fff5c154fb6de346b036a45c4f21ffd96.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 447, + 480, + 480 + ], + "lines": [ + { + "bbox": [ + 130, + 447, + 480, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 480, + 480 + ], + "type": "text", + "content": "Fig. 5: (a) Thermal semantic segmentation failures due to intra-class semantic variations when testing on geographically-partitioned, out-of-domain data. (b) Failures due to photometric variations between day (in-domain) and night (out-of-domain)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 495, + 480, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 480, + 567 + ], + "type": "text", + "content": "Geographically-partitioned benchmarks We retrained our best baseline (Tab. 2) on geographically-partitioned sets to evaluate intra-thermal generalization. Given the cost and effort of collecting diverse aerial thermal data of field settings, we aim to determine the minimum amount of annotated data required to effectively deploy. Using EfficientViT-B0, we performed two experiments to study generalization across difference geographic regions and terrain types." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 145, + 605, + 302, + 654 + ], + "blocks": [ + { + "bbox": [ + 147, + 581, + 465, + 592 + ], + "lines": [ + { + "bbox": [ + 147, + 581, + 465, + 592 + ], + "spans": [ + { + "bbox": [ + 147, + 581, + 465, + 592 + ], + "type": "text", + "content": "Table 3: Thermal semantic segmentation results on geographically-split data" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 185, + 594, + 264, + 604 + ], + "lines": [ + { + "bbox": [ + 185, + 594, + 264, + 604 + ], + "spans": [ + { + "bbox": [ + 185, + 594, + 264, + 604 + ], + "type": "text", + "content": "(a) Region-based split" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 145, + 605, + 302, + 654 + ], + "lines": [ + { + "bbox": [ + 145, + 605, + 302, + 654 + ], + "spans": [ + { + "bbox": [ + 145, + 605, + 302, + 654 + ], + "type": "table", + "html": "
Train AreaTest AreaAvg. mIoU
CANCKY
CA0.6660.0840.1930.314
NC0.0840.5080.0310.208
All0.6480.5250.5870.586
", + "image_path": "443b38ac82f58e0680036f50defabfe7e43e86f1291af699b1facd5b199041bc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 309, + 605, + 466, + 663 + ], + "blocks": [ + { + "bbox": [ + 348, + 594, + 430, + 604 + ], + "lines": [ + { + "bbox": [ + 348, + 594, + 430, + 604 + ], + "spans": [ + { + "bbox": [ + 348, + 594, + 430, + 604 + ], + "type": "text", + "content": "(b) Terrain-based split" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 605, + 466, + 663 + ], + "lines": [ + { + "bbox": [ + 309, + 605, + 466, + 663 + ], + "spans": [ + { + "bbox": [ + 309, + 605, + 466, + 663 + ], + "type": "table", + "html": "
Training TerrainTesting TerrainAvg. mIoU
RiverLakeCoast
River0.6680.1960.1170.327
Lake0.1960.3640.0610.207
Coast0.1270.1190.5220.256
All0.6460.4150.4930.518
All + ground0.6580.4160.5200.532
", + "image_path": "6064778701cec0ecfc8495e2ed7abeaab2b7ba5269934382e1447e266218e12e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 243, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 243, + 91, + 448, + 101 + ], + "type": "text", + "content": "Caltech Aerial RGB-Thermal Dataset in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "In the first experiment, we trained on thermal data from CA and tested on images from KY and NC (Tab. 3a). Despite common classes in each region, results show poor generalization to class variations across geographic areas. We see this again when repeating with the NC split. In the second experiment, we divided our data into three terrain types, aerial river, aerial lake, and aerial coast. Remaining ground-captured images were lumped into a fourth ground category. We trained and tested all pairwise combinations and found poor generalization across terrain types as well (Tab. 3b). Notably, we found low, in-domain performance for the lake setting but saw gains after training on aerial data from other terrain and further improvements after training on ground level images. Overall, our results reiterate the benefit of having training data representative of the testing environment, but also show that gains can be made by including out-of-domain (OOD) data from different terrain and different viewpoints." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 277, + 482, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 338 + ], + "type": "text", + "content": "Overall, the geographic benchmark presents a difficult domain adaptation problem, requiring both inter- and intra-modality domain adaptation techniques to overcome. Although such techniques could perform well in this challenge, the results also highlight a need for an even larger and more diverse thermal dataset capturing common field environments than what we presented." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 346, + 482, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 455 + ], + "type": "text", + "content": "Temporally-partitioned benchmark Here, we assess EfficientViT-B0's performance against intra-thermal, temporal distribution shift due to thermal inversion. We train the model on images from the daytime and test on images captured at sunrise and at night. The model achieved mIoUs of 0.242 at sunrise, 0.193 at night, and 0.777 on a separate daytime test split. Despite training on daytime images from the same location, the model fails on nighttime scenes, struggling to classify water due to thermal inversion (Fig. 5b). As such, to perform well on this split, future algorithms need to augment datasets to simulate such inversions or intentional collect data capturing such phenomena." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 133, + 479, + 479, + 536 + ], + "blocks": [ + { + "bbox": [ + 140, + 467, + 471, + 479 + ], + "lines": [ + { + "bbox": [ + 140, + 467, + 471, + 479 + ], + "spans": [ + { + "bbox": [ + 140, + 467, + 471, + 479 + ], + "type": "text", + "content": "Table 4: RGB-Thermal semantic segmentation network baseline results (mIoU)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 479, + 479, + 536 + ], + "lines": [ + { + "bbox": [ + 133, + 479, + 479, + 536 + ], + "spans": [ + { + "bbox": [ + 133, + 479, + 479, + 536 + ], + "type": "table", + "html": "
ModelBare groundRocky terrainDevel. struct.RoadShrubsTreesSkyWaterVehiclesPersonAllFLOPS (G)
RGB Only†0.7900.8730.8080.5590.6640.7520.8910.9670.3990.0000.67045
Thermal Only†0.7660.8350.7890.5360.6540.7490.9070.9710.3990.0000.66145
EAEFNet [39]0.8050.8630.8130.5450.7020.8020.9370.9760.4810.0000.692358
CRM [60]0.8160.9000.8510.6160.6950.7790.9230.9740.3770.2060.714310
CMNeXt [79]0.8300.9000.8610.6070.7180.8080.9330.9800.5600.1900.740148
", + "image_path": "9ec779b42a5425fddda1a42c53fb70e90294d2aea5aa83095ad104c8a951bd9f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 237, + 537, + 376, + 548 + ], + "lines": [ + { + "bbox": [ + 237, + 537, + 376, + 548 + ], + "spans": [ + { + "bbox": [ + 237, + 537, + 376, + 548 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 237, + 537, + 376, + 548 + ], + "type": "text", + "content": " DeepLabV3+ (w/ ResNet50 encoder)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 564, + 320, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 564, + 320, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 564, + 320, + 578 + ], + "type": "text", + "content": "4.2 RGB-T Semantic Segmentation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 584, + 481, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 584, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 130, + 584, + 481, + 621 + ], + "type": "text", + "content": "We use the paired RGB-T dataset (Sec. 3.2) and partition it using the general split (Sec. 4.1). Recall that RGB and thermal input sizes are " + }, + { + "bbox": [ + 130, + 584, + 481, + 621 + ], + "type": "inline_equation", + "content": "960 \\times 600" + }, + { + "bbox": [ + 130, + 584, + 481, + 621 + ], + "type": "text", + "content": " pixels and the thermal FoV is narrower than in the previous benchmarks (Sec. 4.1)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Baselines We test three RGB-T semantic segmentation algorithms achieving state-of-the-art results in urban RGB-T scene segmentation benchmarks: CRM [60], EAEFNet [39], and CMNeXt [79]. To compare against single-modality mod-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "C. Lee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 132, + 479, + 212 + ], + "blocks": [ + { + "bbox": [ + 205, + 121, + 409, + 131 + ], + "lines": [ + { + "bbox": [ + 205, + 121, + 409, + 131 + ], + "spans": [ + { + "bbox": [ + 205, + 121, + 409, + 131 + ], + "type": "text", + "content": "Table 5: RGB-Thermal image translation results" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 132, + 479, + 212 + ], + "lines": [ + { + "bbox": [ + 132, + 132, + 479, + 212 + ], + "spans": [ + { + "bbox": [ + 132, + 132, + 479, + 212 + ], + "type": "table", + "html": "
MethodMethod TypeRGB → Thermal
PSNR ↑SSIM ↑mIoU ↑
UNIT [42]Unpaired GAN12.850.4210.189
MUNIT [27]Unpaired GAN16.080.4590.173
Edge-guided RGB-T [35]Unpaired GAN12.200.4360.225
Pix2Pix [28]Paired GAN20.060.5470.354
Pix2PixHD [73]Paired GAN20.050.5790.379
VQGAN [15]Paired GAN18.090.5400.388
Palette [58]Paired Diffusion11.510.3990.194
", + "image_path": "d43441f2e987afa8bcd53dbdfa70daf5c177f73b4a06e1572d760593eb6933af.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 221, + 479, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 221, + 479, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 479, + 246 + ], + "type": "text", + "content": "els, we train two additional ResNet50/DeepLabV3+ segmentation models on RGB and thermal data accordingly (Sec. 4.1)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 480, + 357 + ], + "type": "text", + "content": "Performance analysis All RGB-T models outperform the single-modality baselines, with CMNeXt outperforming the next best RGB-T model by 0.026 mIoU (Tab. 4). Overall, incorporating RGB imagery into the segmentation pipeline greatly improved the distinction between land-based stuff classes which can be hard to distinguish in thermal due to thermal crossover. However, these improvements come at a high computational cost and would not be suitable for field robotic applications. Excelling in this benchmark would involve improving the segmentation performance of the object classes, possibly by leveraging RGB data in unsupervised domain adaptation training schemes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 365, + 328, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 365, + 328, + 378 + ], + "spans": [ + { + "bbox": [ + 131, + 365, + 328, + 378 + ], + "type": "text", + "content": "4.3 RGB-Thermal Image Translation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 382, + 479, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 479, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 479, + 406 + ], + "type": "text", + "content": "We use the RGB-T paired dataset from Sec. 4.2 to benchmark image translation algorithms operating in the RGB " + }, + { + "bbox": [ + 130, + 382, + 479, + 406 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 382, + 479, + 406 + ], + "type": "text", + "content": " T direction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 411, + 479, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 479, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 479, + 448 + ], + "type": "text", + "content": "Baselines We benchmark GAN methods that need RGB-T pairs (Pix2Pix [28], Pix2PixHD [73], VQGAN [15]) and unpaired methods (UNIT [42], MUNIT [27], [35]) that do not. We also evaluate Palette [58], a paired diffusion approach." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 138, + 456, + 476, + 566 + ], + "blocks": [ + { + "bbox": [ + 138, + 456, + 476, + 566 + ], + "lines": [ + { + "bbox": [ + 138, + 456, + 476, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 476, + 566 + ], + "type": "image", + "image_path": "a9a61df679ccbf54e3fb41a2d642ff4c809d26829049499726c6365b8fd2beef.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 568, + 464, + 580 + ], + "lines": [ + { + "bbox": [ + 147, + 568, + 464, + 580 + ], + "spans": [ + { + "bbox": [ + 147, + 568, + 464, + 580 + ], + "type": "text", + "content": "Fig. 6: RGB-to-thermal image translation results. Zoom-in to see fine-details." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "content": "Image translation metrics We quantify RGB-T image translation using the Peak-Signal-to-Noise ratio (PSNR) and the Structural Similarity Index (SSIM). In addition, we propose a third metric: the thermal mIoU. This is equivalent to the FCN-score used in RGB image translation works [28]. Instead of an RGB network, our metric uses our EfficientViT thermal segmentation network (Tab. 2) to segment translated thermal images before evaluating with ground truth." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 243, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 243, + 91, + 448, + 101 + ], + "type": "text", + "content": "Caltech Aerial RGB-Thermal Dataset in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "Performance analysis Paired GANs outperform unpaired GANs and diffusion methods, achieving higher values across all three metrics (Tab. 5). Qualitative assessment (Fig. 6) shows poor translations from unpaired techniques, with most retaining geometric artifacts unique to RGB (ocean waves and shadows) and ignoring relative temperature characteristics. Paired GANs produce results similar to real thermal images but with inconsistent details upon closer inspection. Likewise, the diffusion method produces accurate translations but is not consistent. Overall, RGB-T translation still requires improvements before being able to provide a reliable source of thermal training data from existing RGB datasets. For domain adaptation purposes, we emphasize that maximizing the thermal mIoU score is more important than photometric consistency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 255, + 246, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 255, + 246, + 268 + ], + "spans": [ + { + "bbox": [ + 131, + 255, + 246, + 268 + ], + "type": "text", + "content": "4.4 Motion Tracking" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 271, + 481, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 481, + 319 + ], + "type": "text", + "content": "To quantify VIO/SLAM robustness, we select 12 clipped sequences from our dataset ranging in motion tracking difficulty: from urban sports fields (easy) to our natural environments (hard). We normalize thermal images using the " + }, + { + "bbox": [ + 130, + 271, + 481, + 319 + ], + "type": "inline_equation", + "content": "5^{\\text{th}}" + }, + { + "bbox": [ + 130, + 271, + 481, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 271, + 481, + 319 + ], + "type": "inline_equation", + "content": "95^{\\text{th}}" + }, + { + "bbox": [ + 130, + 271, + 481, + 319 + ], + "type": "text", + "content": " percentile pixel values (Sec. 3.2) to enable feature detection." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 132, + 346, + 480, + 422 + ], + "blocks": [ + { + "bbox": [ + 132, + 334, + 481, + 345 + ], + "lines": [ + { + "bbox": [ + 132, + 334, + 481, + 345 + ], + "spans": [ + { + "bbox": [ + 132, + 334, + 481, + 345 + ], + "type": "text", + "content": "Table 6: VIO/SLAM performance (Absolute Trajectory Error [m]) on aerial sequences" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 346, + 480, + 422 + ], + "lines": [ + { + "bbox": [ + 132, + 346, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 132, + 346, + 480, + 422 + ], + "type": "table", + "html": "
MethodTypeModalityNorth FieldCastaic Lake, CADuck, NC
N1N2N3N4N5N6C1C2C3C4D1D2
VINS-FusionSLAMRGB2.8837.5304.0362.0094.0450.8341.555—3.2771.5661.0310.700
VINS-FusionThermal9.45412.269.8542.92611.005.1147.2961.5135.3212.0250.879—
VINS-FusionVIORGB5.4589.5096.5182.1216.0261.2771.555—3.2771.3771.0310.725
VINS-FusionThermal13.8212.8611.992.92613.985.1947.28410.605.3212.0340.879—
Open-VINSVIORGB22.8036.1537.155.165—3.896—3.573——0.4750.700
Open-VINSThermal14.4330.6416.122.258—1.562—5.311——1.073—
Trajectory length (m)1533132612531531104642310651741378880
", + "image_path": "26ad600a345e93697527d5a02c6219e5a65f4e2c4035340411ddd0408f5c6240.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 434, + 481, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 481, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 481, + 481 + ], + "type": "text", + "content": "Baselines We evaluate VINS-Fusion [54] (graph optimization-based) and OpenVINS [21] (filtering-based). We test VINS-Fusion in VIO and SLAM modes, with and without loop closure constraints, respectively, and report Absolute Trajectory Error [64] averaged over four runs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 486, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 665 + ], + "type": "text", + "content": "Performance analysis This benchmark reveals challenges due to fast motion at altitude and periodic feature sparsity in littoral environments. In such scenes, motion tracking frequently failed due to loss of features, requiring sequences to be clipped for quantifiable evaluations (Tab. 6). VINS-Fusion outperformed Open-VINS mainly due to higher feature tracking reliability. VINS-Fusion works better on RGB images than on thermal images in general. However, VINS-Fusion demonstrates superior feature tracking on thermal images compared to RGB on the C2 sequence of Castaic Lake, where the thermal contrast of stone grains on the lake coast was particularly pronounced in the late afternoon. This observation motivates future research on integrating thermal and RGB cameras for motion tracking with enhanced robustness and versatility. Our benchmark presents a difficult challenge for VIO and SLAM algorithms in extreme environments where water and reflections off water surfaces dominate image scenes, providing a unique test bed to invigorate interest in robust feature matching and motion tracking in natural texture-deficient scenarios." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "C. Lee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 168, + 120, + 445, + 132 + ], + "lines": [ + { + "bbox": [ + 168, + 120, + 445, + 132 + ], + "spans": [ + { + "bbox": [ + 168, + 120, + 445, + 132 + ], + "type": "text", + "content": "Table 7: RGB foundation model performances on thermal imagery" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 132, + 160, + 303, + 204 + ], + "blocks": [ + { + "bbox": [ + 132, + 134, + 304, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 134, + 304, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 134, + 304, + 159 + ], + "type": "text", + "content": "(a) Instance segmentation via SAM on thermal and augmented color images using coregistered color image segmentations as ground truth" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 160, + 303, + 204 + ], + "lines": [ + { + "bbox": [ + 132, + 160, + 303, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 303, + 204 + ], + "type": "table", + "html": "
ModalityAPallAPsmallAPmedAPlarge
Thermal0.0180.0070.0220.204
Color (grayscale)0.5380.5360.5310.601
Color (col. jitter)0.8000.7920.8010.841
", + "image_path": "0288b5b7fb3c8b43a407e02c3ecb88b593d9a0718cec210ea880fb8c7a18d00d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 132, + 205, + 304, + 215 + ], + "lines": [ + { + "bbox": [ + 132, + 205, + 304, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 304, + 215 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 132, + 205, + 304, + 215 + ], + "type": "text", + "content": " refers to AP@0.5::0.95, following MSCOCO [40]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 306, + 155, + 477, + 204 + ], + "blocks": [ + { + "bbox": [ + 306, + 134, + 477, + 152 + ], + "lines": [ + { + "bbox": [ + 306, + 134, + 477, + 152 + ], + "spans": [ + { + "bbox": [ + 306, + 134, + 477, + 152 + ], + "type": "text", + "content": "(b) Zero-shot semantic segmentation (mIoU) on coregistered color and thermal imagery" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 155, + 477, + 204 + ], + "lines": [ + { + "bbox": [ + 306, + 155, + 477, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 155, + 477, + 204 + ], + "type": "table", + "html": "
MethodColorColor (gray)Color (jitter)Thermal
SAM [33] + GT ann.‡0.7040.6980.7000.653
Grounded SAM [43]0.3800.3510.3610.193
OV-Seg [38]0.3620.3400.3650.217
ODISE [77]0.5040.4500.4860.232
", + "image_path": "b02135979860aa337e2ccd24a9a2591c7fbdbe157d3f595dc3357180f9bd6fdc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 306, + 205, + 477, + 223 + ], + "lines": [ + { + "bbox": [ + 306, + 205, + 477, + 223 + ], + "spans": [ + { + "bbox": [ + 306, + 205, + 477, + 223 + ], + "type": "text", + "content": "Most frequent ground truth class label in a SAM mask assigned as label for entire mask" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 234, + 246, + 246 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 234, + 246, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 234, + 246, + 246 + ], + "type": "text", + "content": "4.5 Further Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 248, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 344 + ], + "type": "text", + "content": "Zero-shot foundation models on thermal imagery Current foundation models work well on RGB imagery but have not been benchmarked on thermal, especially in non-urban settings. To evaluate these models on thermal imagery, specifically for zero-shot segmentation and semantic segmentation, we compare RGB and thermal performance using the aligned set (Sec. 3.2) in order to isolate modality as the root cause for performance differences. In the following tests, our metrics do not include the rare object classes (vehicle, person) as they drop metrics to near-uniform values for both modalities, making it difficult to compare." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 144, + 361, + 471, + 495 + ], + "blocks": [ + { + "bbox": [ + 144, + 361, + 471, + 495 + ], + "lines": [ + { + "bbox": [ + 144, + 361, + 471, + 495 + ], + "spans": [ + { + "bbox": [ + 144, + 361, + 471, + 495 + ], + "type": "image", + "image_path": "d9b6b21b6e513c5b08fe3a5be09c135904f6fb26b560d0765750b59e2ec7805b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 499, + 482, + 534 + ], + "lines": [ + { + "bbox": [ + 130, + 499, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 499, + 482, + 534 + ], + "type": "text", + "content": "Fig. 7: Zero-shot foundation models on RGB-T pairs. Oversegmented SAM outputs differ across modalities but are similar with ground truth semantics added. Semantic segmentation models (prompted with all classes) perform better on RGB imagery." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 545, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 641 + ], + "type": "text", + "content": "We quantify Segment Anything's (SAM) [33] zero-shot segmentation sensitivity to modality change by measuring the spatial alignment of its mask outputs. We apply it to the RGB set and the 8-bit thermal image set (Sec. 3.2) with the default " + }, + { + "bbox": [ + 130, + 545, + 482, + 641 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 130, + 545, + 482, + 641 + ], + "type": "text", + "content": " grid points to segment everything (Fig. 7). We compute the average precision (AP) of the predicted thermal masks, using the predicted RGB masks as ground truth, and compare the thermal AP to the APs of a color-jittered RGB set and a grayscale set. We find that SAM is sensitive to photometric changes, with large performance drops from color-jitter to grayscale to thermal (Tab. 7a)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Next, we test large vision-language models (Grounded-SAM [43], OV-Seg [38], and ODISE [77]) in zero-shot semantic segmentation. We prompt the models" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 243, + 91, + 448, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 91, + 448, + 101 + ], + "spans": [ + { + "bbox": [ + 243, + 91, + 448, + 101 + ], + "type": "text", + "content": "Caltech Aerial RGB-Thermal Dataset in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "with a list of classes in our dataset; generate semantic segmentation masks for RGB, color-jittered RGB, grayscale, and thermal image sets; and compute the mIoU of their outputs using our ground truth annotations. Although RGB, RGB-variants, and thermal modalities all yield poor semantic segmentation performance (Tab. 7b), all RGB variants outperform thermal by at least " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "0.13\\mathrm{mIoU}" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": " indicating low generalization to the thermal modality (Fig. 7)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 272 + ], + "type": "text", + "content": "Finally, we create semantic SAM masks by assigning class labels to SAM instances based on the most frequent ground truth class within an instance (Fig. 7). We find that these masks attain much higher mIoU score (0.653) compared to the other zero-shot methods. Based on the results (Tab. 7), we note that SAM can find object boundaries in the thermal domain reasonably well but can oversegment, resulting in poor AP scores when compared to RGB (which can also oversegment). When taken with ground truth semantics, this behavior is hidden." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 275, + 298, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 298, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 298, + 357 + ], + "type": "text", + "content": "Transfer learning with urban thermal datasets We test if pretraining on thermal urban data can assist transfer to field settings. We sample 40k unlabeled thermal images from [11, 17, 22, 41, 61, 67, 71, 76] and pretrain various networks using Mo" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 358, + 482, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 419 + ], + "type": "text", + "content": "CoV2 [9]. We then finetune the pretrained networks on our training set and evaluate on our test set. Overall, we find sparse evidence suggesting any advantage of pretraining with existing urban thermal datasets (Tab. 8). Off-the-shelf ImageNet weights can provide strong performance in the thermal field domain, especially over training from scratch, without the effort and costs of pretraining." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 305, + 300, + 480, + 350 + ], + "blocks": [ + { + "bbox": [ + 304, + 277, + 481, + 300 + ], + "lines": [ + { + "bbox": [ + 304, + 277, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 481, + 300 + ], + "type": "text", + "content": "Table 8: Network pretraining methods for downstream thermal segmentation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 305, + 300, + 480, + 350 + ], + "lines": [ + { + "bbox": [ + 305, + 300, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 305, + 300, + 480, + 350 + ], + "type": "table", + "html": "
ModelmIoU# Params. (M)
NoneImageNetTherm. Urban
FastSCNN0.6400.6900.6881.1
EfficientViT0.6870.7250.7144.8
ResNet180.7020.7130.70612.3
ResNet500.6820.7130.72826.7
ConvNext-B0.6250.7170.69789.4
", + "image_path": "10532871c4a315a3608d20adde7c7b71cb612806eead0ad81fa79f95aab7425d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 426, + 220, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 426, + 220, + 440 + ], + "spans": [ + { + "bbox": [ + 131, + 426, + 220, + 440 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 446, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 615 + ], + "type": "text", + "content": "We presented the Caltech Aerial RGB-Thermal Dataset, the first dataset tailored towards advancing thermal semantic perception and motion tracking algorithms in natural environments. We created four benchmarks encompassing semantic segmentation, image translation, and motion tracking, and demonstrate challenges faced by current thermal perception and localization algorithms. Semantic segmentation and image translation methods were notably affected by geographical domain shifts, reflecting the wide intra-class and inter-scenery distributions in our dataset, and temporal domain shifts, due to thermal inversion and crossover. Motion tracking in thermal suffered from feature sparsity, a challenge commonly encountered in real-world robotic deployments [1]. Our motion tracking benchmark penalizes current VIO/SLAM algorithms for assuming ideal feature-tracking conditions, serving as a testbed to fuel improvements in this area. Our dataset and benchmarks can be used into the future to help develop algorithms that expand the operating domains of robotics and computer vision." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 622, + 247, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 622, + 247, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 622, + 247, + 635 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "This work was funded by the Office of Naval Research. We thank K. Koetje, K. Brodie, T. Hesser, T. Almeida, T. Cook, N. Spore, and R. Beach." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "C. Lee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 134, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 134, + 480, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 134, + 480, + 156 + ], + "spans": [ + { + "bbox": [ + 138, + 134, + 480, + 156 + ], + "type": "text", + "content": "1. Ackerman, E.: Blade strike on landing ends mars helicopter's epic journey (Jan 2024), https://spectrum.ieee.org/mars-helicopter-ingenuity-end-mission" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 157, + 480, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 157, + 480, + 189 + ], + "spans": [ + { + "bbox": [ + 138, + 157, + 480, + 189 + ], + "type": "text", + "content": "2. Bloesch, M., Burri, M., Omari, S., Hutter, M., Siegwart, R.: Iterated extended kalman filter based visual-inertial odometry using direct photometric feedback. The International Journal of Robotics Research 36(10), 1053-1072 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 190, + 480, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 190, + 480, + 233 + ], + "spans": [ + { + "bbox": [ + 138, + 190, + 480, + 233 + ], + "type": "text", + "content": "3. Bondi, E., Jain, R., Aggrawal, P., Anand, S., Hannaford, R., Kapoor, A., Piavis, J., Shah, S., Joppa, L., Dilkina, B., et al.: Birdsai: A dataset for detection and tracking in aerial thermal infrared videos. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1747-1756 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 234, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 481, + 277 + ], + "type": "text", + "content": "4. Brodie, K.L., Bruder, B.L., Slocum, R.K., Spore, N.J.: Simultaneous mapping of coastal topography and bathymetry from a lightweight multicamera uas. IEEE Transactions on Geoscience and Remote Sensing 57(9), 6844-6864 (2019). https://doi.org/10.1109/TGRS.2019.2909026" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 277, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 480, + 300 + ], + "type": "text", + "content": "5. Cai, H., Gan, C., Han, S.: Efficientvit: Enhanced linear attention for high-resolution low-computation visual recognition. arXiv preprint arXiv:2205.14756 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 300, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 480, + 322 + ], + "type": "text", + "content": "6. Cai, W., Jin, K., Hou, J., Guo, C., Wu, L., Yang, W.: Vdd: Varied drone dataset for semantic segmentation. arXiv preprint arXiv:2305.13608 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 323, + 480, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 480, + 355 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 480, + 355 + ], + "type": "text", + "content": "7. Center, A.F.S.: A dataset for machine learning algorithm development from 2010-06-15 to 2010-08-15. NOAA National Centers for Environmental Information (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 355, + 480, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 355, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 138, + 355, + 480, + 389 + ], + "type": "text", + "content": "8. Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 801-818 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 389, + 480, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 389, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 389, + 480, + 411 + ], + "type": "text", + "content": "9. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 411, + 480, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 480, + 455 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 480, + 455 + ], + "type": "text", + "content": "0. Chen, Y., Wang, Y., Lu, P., Chen, Y., Wang, G.: Large-scale structure from motion with semantic constraints of aerial images. In: Pattern Recognition and Computer Vision: First Chinese Conference, PRCV 2018, Guangzhou, China, November 23-26, 2018, Proceedings, Part I 1. pp. 347-359. Springer (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 456, + 480, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 480, + 488 + ], + "type": "text", + "content": "1. Choi, Y., Kim, N., Hwang, S., Park, K., Yoon, J.S., An, K., Kweon, I.S.: Kaist multi-spectral day/night data set for autonomous and assisted driving. IEEE Transactions on Intelligent Transportation Systems 19(3), 934-948 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 488, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 488, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 488, + 480, + 521 + ], + "type": "text", + "content": "2. Deevi, S.A., Lee, C., Gan, L., Nagesh, S., Pandey, G., Chung, S.J.: Rgb-x object detection via scene-specific fusion modules. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 7366-7375 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 521, + 480, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 480, + 565 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 480, + 565 + ], + "type": "text", + "content": "3. Delaune, J., Hewitt, R., Lytle, L., Sorice, C., Thakker, R., Matthies, L.: Thermal-inertial odometry for autonomous flight throughout the night. In: 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1122-1128. IEEE (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 565, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 565, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 565, + 480, + 609 + ], + "type": "text", + "content": "4. Doer, C., Trommer, G.F.: Radar visual inertial odometry and radar thermal inertial odometry: Robust navigation even in challenging visual conditions. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 331-338. IEEE (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "text", + "content": "5. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873-12883 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "6. Feng, D., Haase-Schütz, C., Rosenbaum, L., Hertlein, H., Glaeser, C., Timm, F., Wiesbeck, W., Dietmayer, K.: Deep multi-modal object detection and semantic" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 243, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 243, + 91, + 447, + 101 + ], + "type": "text", + "content": "Caltech Aerial RGB-Thermal Dataset in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 663 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 147, + 117, + 481, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 117, + 481, + 139 + ], + "spans": [ + { + "bbox": [ + 147, + 117, + 481, + 139 + ], + "type": "text", + "content": "segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems 22(3), 1341-1360 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "text", + "content": "17. Teledyne fliradas dataset,https://www.flir.com/oem/adas/adas-dataset-form/,Lastaccessed on2023-10-27" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 162, + 480, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 480, + 184 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 480, + 184 + ], + "type": "text", + "content": "18. Fountas, S., Mylonas, N., Malounas, I., Rodias, E., Hellmann Santos, C., Pekkeriet, E.: Agricultural robotics for field operations. Sensors 20(9), 2672 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 185, + 480, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 480, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 480, + 206 + ], + "type": "text", + "content": "19. Gade, R., Moeslund, T.B.: Thermal cameras and applications: a survey. Machine vision and applications 25, 245-262 (2014)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 208, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 208, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 208, + 480, + 239 + ], + "type": "text", + "content": "20. Gan, L., Lee, C., Chung, S.J.: Unsupervised rgb-to-thermal domain adaptation via multi-domain attention network. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 6014-6020. IEEE (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 241, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 241, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 241, + 480, + 274 + ], + "type": "text", + "content": "21. Geneva, P., Eckenhoff, K., Lee, W., Yang, Y., Huang, G.: Openvins: A research platform for visual-inertial estimation. In: 2020 IEEE International Conference on Robotics and Automation (ICRA). pp. 4666-4672. IEEE (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 274, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 480, + 319 + ], + "type": "text", + "content": "22. Ha, Q., Watanabe, K., Karasawa, T., Ushiku, Y., Harada, T.: Mfnet: Towards real-time semantic segmentation for autonomous vehicles with multi-spectral scenes. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 5108-5115. IEEE (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 319, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 480, + 352 + ], + "type": "text", + "content": "23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 353, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 480, + 385 + ], + "type": "text", + "content": "24. He, M., Rajkumar, R.R.: Using thermal vision for extended vins-mono to localize vehicles in large-scale outdoor road environments. In: 2021 IEEE Intelligent Vehicles Symposium (IV). pp. 953-960. IEEE (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 480, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 419 + ], + "type": "text", + "content": "25. Howard, A., Sandler, M., Chu, G., Chen, L.C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., et al.: Searching for mobilenetv3. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1314-1324 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 420, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 480, + 453 + ], + "type": "text", + "content": "26. Hua, T., Pei, L., Li, T., Wu, Q., Wang, R., Yu, W.: I2-slam: Fusing infrared camera and imu for simultaneous localization and mapping. In: International Conference on Autonomous Unmanned Systems. pp. 2834-2844. Springer (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 453, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 480, + 475 + ], + "type": "text", + "content": "27. Huang, X., Liu, M.Y., Belongie, S., Kautz, J.: Multimodal unsupervised image-to-image translation. In: European Conference on Computer Vision (ECCV) (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 475, + 480, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 480, + 509 + ], + "type": "text", + "content": "28. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1125-1134 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 510, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 480, + 542 + ], + "type": "text", + "content": "29. Jia, X., Zhu, C., Li, M., Tang, W., Zhou, W.: Llvip: A visible-infrared paired dataset for low-light vision. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3496-3504 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 543, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 480, + 586 + ], + "type": "text", + "content": "30. Jong, A., Yu, M., Dhrafani, D., Kailas, S., Moon, B., Sycara, K., Scherer, S.: Wit-uas: A wildland-fire infrared thermal dataset to detect crew assets from aerial views. In: 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 11464-11471. IEEE (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 587, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 480, + 620 + ], + "type": "text", + "content": "31. Khattak, S., Papachristos, C., Alexis, K.: Keyframe-based direct thermal-inertial odometry. In: 2019 International Conference on Robotics and Automation (ICRA). pp. 3563-3569. IEEE (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 621, + 480, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 663 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 663 + ], + "type": "text", + "content": "32. Kim, Y.H., Shin, U., Park, J., Kweon, I.S.: Ms-uda: Multi-spectral unsupervised domain adaptation for thermal image semantic segmentation. IEEE Robotics and Automation Letters 6(4), 6497–6504 (Oct 2021). https://doi.org/10.1109/LRA.2021.3093652" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "C. Lee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "33. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "34. Lee, C., Frennert, J.G., Gan, L., Anderson, M., Chung, S.J.: Online self-supervised thermal water segmentation for aerial vehicles. In: 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 7734-7741. IEEE (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 225 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 225 + ], + "type": "text", + "content": "35. Lee, D.G., Jeon, M.H., Cho, Y., Kim, A.: Edge-guided multi-domain rgb-to-tir image translation for training vision tasks with challenging labels. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 8291-8298. IEEE (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 225, + 482, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 482, + 257 + ], + "type": "text", + "content": "36. Li, C., Xia, W., Yan, Y., Luo, B., Tang, J.: Segmenting objects in day and night: Edge-conditioned cnn for thermal image semantic segmentation. IEEE Transactions on Neural Networks and Learning Systems 32(7), 3069-3082 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 257, + 482, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 278 + ], + "type": "text", + "content": "37. Li, H., Ma, Y., Huang, Y., Gu, Y., Xu, W., Liu, Y., Zuo, X.: Riders: Radar-infrared depth estimation for robust sensing. arXiv preprint arXiv:2402.02067 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 278, + 482, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 320 + ], + "type": "text", + "content": "38. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7061-7070 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 321, + 482, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 354 + ], + "type": "text", + "content": "39. Liang, M., Hu, J., Bao, C., Feng, H., Deng, F., Lam, T.L.: Explicit attention-enhanced fusion for rgb-thermal perception tasks. IEEE Robotics and Automation Letters (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 354, + 482, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 397 + ], + "type": "text", + "content": "40. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 397, + 482, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 482, + 440 + ], + "type": "text", + "content": "41. Liu, J., Fan, X., Huang, Z., Wu, G., Liu, R., Zhong, W., Luo, Z.: Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5802-5811 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 440, + 482, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 482, + 461 + ], + "type": "text", + "content": "42. Liu, M.Y., Breuel, T., Kautz, J.: Unsupervised image-to-image translation networks. In: Conference on Neural Information Processing Systems (NeurIPS) (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 461, + 482, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 493 + ], + "type": "text", + "content": "43. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 493, + 482, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 493, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 493, + 482, + 525 + ], + "type": "text", + "content": "44. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 525, + 482, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 579 + ], + "type": "text", + "content": "45. Lyu, Y., Vosselman, G., Xia, G.S., Yilmaz, A., Yang, M.Y.: Uavid: A semantic segmentation dataset for uav imagery. ISPRS Journal of Photogrammetry and Remote Sensing 165, 108 - 119 (2020). https://doi.org/https://doi.org/10.1016/j.isprsjprs.2020.05.009, http://www.sciencedirect.com/science/article/pii/S0924271620301295" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 579, + 482, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 579, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 579, + 482, + 601 + ], + "type": "text", + "content": "46. Mehta, S., Rastegari, M.: Separable self-attention for mobile vision transformers. arXiv preprint arXiv:2206.02680 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 601, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 482, + 632 + ], + "type": "text", + "content": "47. Mostegel, C., Maurer, M., Heran, N., Pestana Puerta, J., Fraundorfer, F.: Semantic drone dataset (Jan 2019), http://dronedataset.icg.tugraz.at/, Last accessed on 2023-10-27" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "text", + "content": "48. Nigam, I., Huang, C., Ramanan, D.: Ensemble knowledge transfer for semantic segmentation. In: 2018 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 1499-1508. IEEE (2018)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 243, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 243, + 91, + 447, + 101 + ], + "type": "text", + "content": "Caltech Aerial RGB-Thermal Dataset in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "49. Nirgudkar, S., DeFilippo, M., Sacarny, M., Benjamin, M., Robinette, P.: Massmind: Massachusetts maritime infrared dataset. The International Journal of Robotics Research 42(1-2), 21-32 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 481, + 203 + ], + "type": "text", + "content": "50. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Howes, R., Huang, P.Y., Xu, H., Sharma, V., Li, S.W., Galuba, W., Rabbat, M., Assran, M., Ballas, N., Synnaeve, G., Misra, I., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "type": "text", + "content": "51. Panetta, K., Shreyas Kamath, K.M., Rajeev, S., Agaian, S.S.: Ftnet: Feature transverse network for thermal image semantic segmentation. IEEE Access 9, 145212-145227 (2021). https://doi.org/10.1109/ACCESS.2021.3123066" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 236, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 481, + 278 + ], + "type": "text", + "content": "52. Park, T., Efros, A.A., Zhang, R., Zhu, J.Y.: Contrastive learning for unpaired image-to-image translation. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16. pp. 319-345. Springer (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 278, + 481, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 481, + 322 + ], + "type": "text", + "content": "53. Poudel, R., Liwicki, S., Cipolla, R.: Fast-scnn: Fast semantic segmentation network. In: Sidorov, K., Hicks, Y. (eds.) Proceedings of the British Machine Vision Conference (BMVC). pp. 187.1-187.12. BMVA Press (September 2019). https://doi.org/10.5244/C.33.187, https://dx.doi.org/10.5244/C.33.187" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 322, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 481, + 343 + ], + "type": "text", + "content": "54. Qin, T., Li, P., Shen, S.: Vins-mono: A robust and versatile monocular visual-inertial state estimator. IEEE Transactions on Robotics 34(4), 1004-1020 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 343, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 481, + 386 + ], + "type": "text", + "content": "55. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "text", + "content": "56. Retief, S.P., Willers, C., Wheeler, M.: Prediction of thermal crossover based on imaging measurements over the diurnal cycle. In: Geo-Spatial and Temporal Image and Data Exploitation III. vol. 5097, pp. 58-69. SPIE (2003)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 418, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 460 + ], + "type": "text", + "content": "57. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 460, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 481, + 493 + ], + "type": "text", + "content": "58. Sahara, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-10 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "type": "text", + "content": "59. Saputra, M.R.U., Lu, C.X., de Gusmao, P.P.B., Wang, B., Markham, A., Trigoni, N.: Graph-based thermal-inertial slam with probabilistic neural networks. IEEE Transactions on Robotics 38(3), 1875-1893 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "text", + "content": "60. Shin, U., Lee, K., Kweon, I.S.: Complementary random masking for rgb-thermal semantic segmentation. In: IEEE International Conference on Robotics and Automation (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 557, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 481, + 590 + ], + "type": "text", + "content": "61. Shin, U., Park, J., Kweon, I.S.: Deep depth estimation from thermal image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1043-1053 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 590, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 481, + 632 + ], + "type": "text", + "content": "62. Shivakumar, S.S., Rodrigues, N., Zhou, A., Miller, I.D., Kumar, V., Taylor, C.J.: Pst900: Rgb-thermal calibration, dataset and segmentation network. In: 2020 IEEE international conference on robotics and automation (ICRA). pp. 9441-9447. IEEE (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "63. Speth, S., Goncalves, A., Rigault, B., Suzuki, S., Bouazizi, M., Matsuo, Y., Prendinger, H.: Deep learning with rgb and thermal images onboard a drone for monitoring operations. Journal of Field Robotics 39(6), 840-868 (2022)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "C. Lee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "64. Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for the evaluation of rgb-d slam systems. In: 2012 IEEE/RSJ international conference on intelligent robots and systems. pp. 573-580. IEEE (2012)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "65. Sun, Y., Zuo, W., Liu, M.: Rtfnet: RGB-thermal fusion network for semantic segmentation of urban scenes. IEEE Robotics and Automation Letters 4(3), 2576-2583 (2019). https://doi.org/10.1109/LRA.2019.2904733" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 225 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 225 + ], + "type": "text", + "content": "66. Sun, Y., Zuo, W., Yun, P., Wang, H., Liu, M.: Fuseseg: Semantic segmentation of urban scenes based on rgb and thermal data fusion. IEEE Transactions on Automation Science and Engineering 18(3), 1000-1011 (2021). https://doi.org/10.1109/TASE.2020.2993143" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 227, + 482, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 227, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 482, + 258 + ], + "type": "text", + "content": "67. Suo, J., Wang, T., Zhang, X., Chen, H., Zhou, W., Shi, W.: Hit-uav: A high-altitude infrared thermal dataset for unmanned aerial vehicle-based object detection. Scientific Data 10, 227 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "text", + "content": "68. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105-6114. PMLR (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 335 + ], + "type": "text", + "content": "69. Ustun, B., Kaya, A.K., Ayerden, E.C., Altinel, F.: Spectral transfer guided active domain adaptation for thermal imagery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. pp. 449-458 (June 2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "type": "text", + "content": "70. Vargas, J., Alsweiss, S., Token, O., Razdan, R., Santos, J.: An overview of autonomous vehicles sensors and their vulnerability to weather conditions. Sensors 21(16), 5397 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 369, + 482, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 403 + ], + "type": "text", + "content": "71. Vertens, J., Zürn, J., Burgard, W.: Heatnet: Bridging the day-night domain gap in semantic segmentation with thermal images. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8461-8468. IEEE (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 403, + 482, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 445 + ], + "type": "text", + "content": "72. VS, V., Poster, D., You, S., Hu, S., Patel, V.M.: Meta-uda: Unsupervised domain adaptive thermal object detection using meta-learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1412-1423 (January 2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 446, + 482, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 489 + ], + "type": "text", + "content": "73. Wang, T.C., Liu, M.Y., Zhu, J.Y., Tao, A., Kautz, J., Catanzaro, B.: High-resolution image synthesis and semantic manipulation with conditional gans. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 490, + 482, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 522 + ], + "type": "text", + "content": "74. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. In: Neural Information Processing Systems (NeurIPS) (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "type": "text", + "content": "75. Xie, S., Girshick, R., Dollar, P., Tu, Z., He, K.: Aggregated residual transformations for deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1492-1500 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 555, + 482, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 598 + ], + "type": "text", + "content": "76. Xiong, H., Cai, W., Liu, Q.: Mcnet: Multi-level correction network for thermal image semantic segmentation of nighttime driving scene. Infrared Physics & Technology p. 103628 (2021). https://doi.org/https://doi.org/10.1016/j.infrared.2020.103628" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 599, + 482, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 599, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 599, + 482, + 643 + ], + "type": "text", + "content": "77. Xu, J., Liu, S., Vahdat, A., Byeon, W., Wang, X., De Mello, S.: Open-vocabulary panoptic segmentation with text-to-image diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2955-2966 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 643, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 482, + 665 + ], + "type": "text", + "content": "78. Yun, S., Jung, M., Kim, J., Jung, S., Cho, Y., Jeon, M.H., Kim, G., Kim, A.: Stereo: Stereo thermal dataset for research in odometry and mapping. In: 2022" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 243, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 243, + 91, + 447, + 101 + ], + "type": "text", + "content": "Caltech Aerial RGB-Thermal Dataset in the Wild" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 304 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 147, + 116, + 479, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 479, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 479, + 138 + ], + "type": "text", + "content": "IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 3857-3864. IEEE (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "text", + "content": "79. Zhang, J., Liu, R., Shi, H., Yang, K., Reif, S., Peng, K., Fu, H., Wang, K., Stiefelhagen, R.: Delivering arbitrary-modal semantic segmentation. In: CVPR (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 194 + ], + "type": "text", + "content": "80. Zhao, S., Wang, P., Zhang, H., Fang, Z., Scherer, S.: Tp-tio: A robust thermal-inertial odometry with deep thermalpoint. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 4505-4512. IEEE (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 236 + ], + "type": "text", + "content": "81. Zhou, W., Liu, J., Lei, J., Yu, L., Hwang, J.N.: Gmnet: Graded-feature multilabel-learning network for rgb-thermal urban scene semantic segmentation. IEEE Transactions on Image Processing 30, 7790–7802 (2021). https://doi.org/10.1109/TIP.2021.3109518" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "82. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Proceedings of the IEEE international conference on computer vision. pp. 2223-2232 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 304 + ], + "type": "text", + "content": "83. Zhu, P., Wen, L., Du, D., Bian, X., Fan, H., Hu, Q., Ling, H.: Detection and tracking meet drones challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(11), 7380-7399 (2021)" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "C. Lee et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Camera Calibration using a Collimator System/26f230b2-04ba-4e24-b664-739a6c3feb85_content_list.json b/2024/Camera Calibration using a Collimator System/26f230b2-04ba-4e24-b664-739a6c3feb85_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1fa49461ed64af2cce2390a989a20607951b9fa3 --- /dev/null +++ b/2024/Camera Calibration using a Collimator System/26f230b2-04ba-4e24-b664-739a6c3feb85_content_list.json @@ -0,0 +1,2175 @@ +[ + { + "type": "text", + "text": "Camera Calibration using a Collimator System", + "text_level": 1, + "bbox": [ + 232, + 141, + 772, + 165 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shunkun Liang, Banglei Guan, Zhenbao Yu, Pengju Sun, and Yang Shang", + "bbox": [ + 225, + 189, + 779, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "College of Aerospace Science and Engineering, National University of Defense Technology, China.", + "bbox": [ + 241, + 232, + 759, + 261 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Camera calibration is a crucial step in photogrammetry and 3D vision applications. In practical scenarios with a long working distance to cover a wide area, target-based calibration methods become complicated and inflexible due to site limitations. This paper introduces a novel camera calibration method using a collimator system, which can provide a reliable and controllable calibration environment for cameras with varying working distances. Based on the optical geometry of the collimator system, we prove that the relative motion between the target and camera conforms to the spherical motion model, reducing the original 6DOF relative motion to 3DOF pure rotation motion. Furthermore, a closed-form solver for multiple views and a minimal solver for two views are proposed for camera calibration. The performance of our method is evaluated in both synthetic and real-world experiments, which verify the feasibility of calibration using the collimator system and demonstrate that our method is superior to the state-of-the-art methods. Demo code is available at https://github.com/LiangSK98/CollimatorCalibration", + "bbox": [ + 261, + 297, + 743, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Camera calibration $\\cdot$ Collimator system $\\cdot$ Spherical motion $\\cdot$ Closed-form solver $\\cdot$ Minimal solver", + "bbox": [ + 261, + 532, + 740, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 588, + 375, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Geometric camera calibration determines the mapping between 3D rays in space and 2D image pixels. Various vision applications, such as simultaneous localization and mapping(SLAM) [6], structure from motion(SfM) [32], and pose estimation [11, 12], rely on camera parameters as inputs. The close-range calibration methods using a planar target represent the prevailing techniques in 3D vision applications [5, 26, 33, 35, 46]. In some special outdoor scenarios, such as long-range depth estimation [45] and monitoring of wind turbine blades [10], the cameras achieve a long working distance and a large field of view. The traditional target-based calibration methods [3, 37, 46] may face challenges, such as needing large targets, moving or arranging targets over a wide area, and dealing with varying and unstable lighting conditions. While constructing a 3D calibration field can offer precise calibration data [34, 40, 42], its operational complexity and limited adaptability present significant challenges. In addition, the collimator is", + "bbox": [ + 212, + 619, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author. guanbanglei12@nudt.edu.cn", + "bbox": [ + 230, + 825, + 607, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dae56f8f0e73dac9831bd26e8e0e5f4f165909d42b47cc78de68a53334b7e8fd.jpg", + "image_caption": [ + "Fig. 1: Diagram of the collimator-based calibration method. From the inherent optical geometric properties of the collimator, the relative motion between the target and camera can be proved to conform to the spherical motion model." + ], + "image_footnote": [], + "bbox": [ + 217, + 141, + 380, + 273 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7511b3c1e41ee793df3e12e29b90c82418900e776999c27a764f787b99e779ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 142, + 545, + 273 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/537c6d63bd0a53ebb6d285fa51a947c270240e674aaf3028b72b99ffc6f0b0a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 549, + 142, + 785, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "commonly used to generate calibration targets [7]. Traditional collimator-based methods need to measure the direction of collimated rays using a goniometer or theodolite, which makes the calibration process challenging, and measurement errors directly impact the accuracy [41, 43, 44].", + "bbox": [ + 212, + 352, + 782, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we design a novel collimator system with a structured calibration pattern on its reticle, and a novel calibration algorithm based on the collimator images is proposed. Our method only requires the camera to observe the pattern from our collimator system at a few (at least two) different orientations. Fig. 1 illustrates this method. We briefly analyze the geometric optical property of the collimator and illustrate the angle invariance of any point pairs on the pattern. By leveraging angle invariance, we prove that the relative motion between the calibration target and the camera conforms to a spherical motion model. This constraint effectively reduces the original 6DOF general motion to the 3DOF pure rotation motion. Furthermore, we propose a novel calibration method using the collimator system. Compared with the target-based methods, more geometric constraints can be derived from the collimator, leading to higher accuracy. Compared to traditional collimator-based methods, our method is more flexible as it eliminates the need for direction measurement.", + "bbox": [ + 212, + 414, + 784, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our primary contributions can be summarized as follows: (1) We propose a novel camera calibration method using a collimator system. The entire calibration process can be implemented with multiple close-range observations and does not require the known direction of collimated rays. (2) We prove that the relative motion between the calibration target and the camera conforms to the spherical motion model. This constraint effectively reduces the number of motion parameters to be estimated. (3) A closed-form solver for multiple views and a minimal solver for two views are proposed based on the spherical motion constraint. Moreover, the degenerate case of the proposed solvers is proved.", + "bbox": [ + 212, + 625, + 785, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 780, + 388, + 796 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Camera calibration is a crucial subject in 3D vision and continues to be actively researched. The literature on camera calibration introduces a wide range of al", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. Liang, B. Guan et al.", + "bbox": [ + 271, + 114, + 433, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "gorithms and toolboxes [3-5,22,26,33,35,46]. To ensure the highest accuracy for calibration, targets with known structures are usually used for camera calibration. Various calibration targets have been proposed, including 3D targets [37], 2D targets [13,18] and 1D targets [47]. In some practical applications where no calibration target is available, the camera calibration can be performed by establishing the correspondence between images during the movement of the camera, which is known as self-calibration [8,9,16,19,28,31]. Among these options, the target-based close-range calibration method is widely used.", + "bbox": [ + 212, + 146, + 782, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In outdoor long-range measurement applications [10, 45], the camera has a long working distance to measure the long-distance targets. Thus, the close-range calibration methods are impractical due to the site limitation. A commonly employed approach involves constructing a calibration field, which serves as a known structure for the calibration [21, 25, 29, 34, 40, 42]. Oniga et al. [29] lay markers on buildings to calibrate cameras mounted on UAV. Shang et al. [34] proposed a high-precision calibration method by placing several markers near the ground. Additionally, Klaus et al. [23] proposed using stars as distant targets to estimate camera parameters.", + "bbox": [ + 212, + 267, + 784, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The collimator is also commonly used to provide 2D-3D correspondence for camera calibration in the laboratory [2, 20, 41, 44]. Previous collimator-based methods proposed to use goniometer or theodolite to obtain the direction of collimated rays and then estimate the camera parameters [41, 43]. The collimator-based method offers both accuracy and flexibility in the limited space, but a high-precision goniometer or theodolite is expensive. To address this, Yuan et al. [44] designed a multiple pinhole mask for the collimator to generate the collimated rays with known directions. However, the accuracy of calibration is sensitive to machining errors. Bauer et al. [2] proposed a calibration method using a diffractive optical element (DOE). The angle between the laser beam and the camera can then be estimated to determine the camera parameters.", + "bbox": [ + 212, + 402, + 785, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Collimator System", + "text_level": 1, + "bbox": [ + 215, + 593, + 431, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section introduces the designed collimator system and shows that the equivalent relative motion between the calibration target and the camera conforms to a spherical motion model.", + "bbox": [ + 212, + 625, + 782, + 670 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 Scheme of Collimator System", + "text_level": 1, + "bbox": [ + 214, + 693, + 504, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The collimator is an optical instrument that can produce collimated rays and plays an important role when adjusting other optical instruments. A collimator consists of essential components such as a light source, a reticle, and an optical lens, as illustrated in Fig. 2. The reticle is a thin glass located on the focal plane of the collimator. The generation of collimated rays can be regarded as the inverse process of imaging a target at infinity onto the focal plane. Specifically, a point on the lens's focal plane will generate collimated rays, which can be regarded as the target at infinity. Therefore, cameras with varying focal lengths", + "bbox": [ + 212, + 719, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Camera Calibration using a Collimator System", + "bbox": [ + 416, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "can consistently and accurately observe the target through the collimator at a close range.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we design a collimator system for camera calibration. A star-based pattern, proposed by Thomas et al. [33], is attached to the reticle of the collimator as the calibration target. The star-based pattern offers richer gradient information for improving feature extraction accuracy. Similar to the plane-based calibration method [35, 46], the star-based pattern provides 2D-3D correspondences for camera calibration. Our collimator system is portable with a size of about $200\\mathrm{mm} \\times 170\\mathrm{mm} \\times 300\\mathrm{mm}$ . The collimator system is designed to provide a reliable and controllable calibration environment for some special scenarios like inappropriate illumination, limited space, and long-range cameras. However, printed targets face challenges in these scenarios. When calibrating a camera, we only need to capture the target from multiple orientations through the lens of the collimator, which is as easy as the printed targets method.", + "bbox": [ + 212, + 176, + 787, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 Spherical Motion Model", + "text_level": 1, + "bbox": [ + 215, + 378, + 462, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we prove that the relative motion between the calibration target and the camera conforms to the spherical motion model. Firstly, we briefly analyze the geometric property of the collimator. As shown in Fig. 2, let us note two arbitrary points $P_{1}$ and $P_{2}$ on the reticle of the collimator. Rays emitted from $P_{1}$ and $P_{2}$ pass through the collimator lens, generating two parallel beams observed by cameras denoted as $C_1$ and $C_2$ with arbitrary poses. The propagation routes of the two beams are visually distinguished by red and blue lines in Fig. 2. Obviously, the angle between two lines observed by the two cameras is equal, that is, $\\theta_{1}\\equiv \\theta_{2}\\equiv \\theta$ . We can obtain a property that the two rays", + "bbox": [ + 212, + 402, + 488, + 688 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/41ab016be5e74375c3a1a1ead198fed0102ad5984f17748cf004b8ec92ebd71f.jpg", + "image_caption": [ + "Fig. 2: Diagram of collimator system. The geometric property of the collimator system leads to the angle invariance of any point pair on the reticle." + ], + "image_footnote": [], + "bbox": [ + 509, + 430, + 771, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "observed by the cameras always maintain a constant angle, regardless of the poses of observing cameras. We refer to this property as the angle invariance.", + "bbox": [ + 212, + 689, + 784, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we show that this setup, where a camera moves around a collimator system and captures images, is theoretically equivalent to the capture setting where the camera is fixed, but the images are moving around the camera spherically. This can be derived from the angle invariance. For convenience in introducing our derivation process, we treat the camera as a fixed reference coordinate, and assume that the calibration target moves relative to the camera in space. The angle invariance allows us to infer that when the target performs a specific motion in space, the angular distance between any point pair on the target always", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. Liang, B. Guan et al.", + "bbox": [ + 271, + 114, + 431, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "remains constant. The angular distance means the angle between two vectors from the origin (camera optical center) to two points. By arbitrarily selecting a pair of points $\\mathbf{P}_i$ and $\\mathbf{P}_j$ on the target and transforming them, we can establish the following equation", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\angle \\left(\\mathbf {P} _ {i}, \\mathbf {P} _ {j}\\right) = \\angle \\left(\\mathbf {R} \\left(\\mathbf {P} _ {i} + \\mathbf {t}\\right), \\mathbf {R} \\left(\\mathbf {P} _ {j} + \\mathbf {t}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 215, + 785, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{R}$ is rotation matrix and $\\mathbf{t}$ is translation vector. From the rotation invariance [39] that a pair of 3D points are jointly rotated about the origin, their angular distance remains constant. We have $\\angle (\\mathbf{R}(\\mathbf{P}_i + \\mathbf{t}),\\mathbf{R}(\\mathbf{P}_j + \\mathbf{t})) = \\angle (\\mathbf{P}_i + \\mathbf{t},\\mathbf{P}_j + \\mathbf{t})$ . Thus, the angle invariance is maintained when the calibration target is arbitrarily rotated about the origin. To make $\\angle (\\mathbf{P}_i,\\mathbf{P}_j) = \\angle (\\mathbf{P}_i + \\mathbf{t},\\mathbf{P}_j + \\mathbf{t})$ hold for any pair of points on the target, the translation vector must be a $\\mathbf{0}$ vector $(\\mathbf{t} = [0,0,0]^T)$ . The proof is given in the supplementary material. Therefore, the calibration target performs pure rotation motion with respect to the origin (i.e. camera optical center). Such pure rotational motion is a typical spherical motion.", + "bbox": [ + 212, + 239, + 784, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The spherical motion can be visualized in Fig. 1. In the spherical motion model, the calibration target exhibits only rotational motion with a fixed distance relative to the camera. We establish a local coordinate on the target, where the $z'$ -axis aligns with the ray between the camera coordinate system $c$ and calibration target coordinate system $p'$ . Inspired by [38], the relative pose between the calibration target and camera can be mathematically represented as", + "bbox": [ + 212, + 390, + 784, + 481 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {c p ^ {\\prime}} = \\left( \\begin{array}{l l} \\mathbf {R} _ {c p ^ {\\prime}} & \\mathbf {t} _ {c p ^ {\\prime}} \\\\ \\mathbf {0} _ {1 \\times 3} & 1 \\end{array} \\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 489, + 785, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{R}_{cp'} \\in SO(3)$ , $\\mathbf{t}_{cp'} = (0,0,-r)^T$ and $r$ is the radius of spherical motion. We usually establish the origin of the coordinate system $p$ on the upper left corner of the calibration pattern. Therefore, there is a relative translation $\\mathbf{t}_{p'p} = (x,y,0)$ between the coordinate systems $p'$ and $p$ . Furthermore, the transformation matrix between camera $c$ and calibration target $p$ is expressed as follows:", + "bbox": [ + 212, + 527, + 782, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {p c} = \\left[ \\left( \\begin{array}{l l} \\mathbf {R} _ {c p ^ {\\prime}} & \\mathbf {t} _ {c p ^ {\\prime}} \\\\ \\mathbf {0} _ {1 \\times 3} & 1 \\end{array} \\right) \\left( \\begin{array}{l l} \\mathbf {I} _ {3 \\times 3} & \\mathbf {t} _ {p ^ {\\prime} p} \\\\ \\mathbf {0} _ {1 \\times 3} & 1 \\end{array} \\right) \\right] ^ {- 1} = \\left( \\begin{array}{l l} \\mathbf {R} _ {p c} & - \\mathbf {R} _ {p c} \\mathbf {t} _ {c p} \\\\ \\mathbf {0} _ {1 \\times 3} & 1 \\end{array} \\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 289, + 613, + 785, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{R}_{pc} = \\mathbf{R}_{cp'}^T$ and $\\mathbf{t}_{cp} = \\mathbf{t}_{p'p} + \\mathbf{t}_{cp'} = (x,y, -r)^T$ . $\\mathbf{t}_{cp}$ determines the position of the camera in the calibration target coordinate system. Note that $\\mathbf{t}_{cp}$ is a fixed value and has nothing to do with the pose of the calibration target. Given the constraint from the spherical motion, the original 6DOF general motion is limited to a 3DOF pure rotation motion.", + "bbox": [ + 212, + 654, + 782, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Calibration Method", + "text_level": 1, + "bbox": [ + 215, + 750, + 444, + 766 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The spherical motion constraint effectively reduces the number of motion parameters to be estimated. In this section, we show how to calibrate camera parameters using a set of collimator images. A closed-form solver for multiple views and a minimal solver for two views are proposed.", + "bbox": [ + 212, + 779, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Camera Calibration using a Collimator System", + "bbox": [ + 416, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Geometric Constraints", + "text_level": 1, + "bbox": [ + 215, + 146, + 450, + 159 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the well-known pinhole model to model cameras. The mapping from a 3D point $\\mathbf{P} = (X,Y,Z)^T$ to an image point $\\mathbf{p} = (u,v)^T$ is computed as", + "bbox": [ + 212, + 169, + 787, + 200 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ns \\widetilde {\\mathbf {p}} = \\mathbf {K} [ \\mathbf {R} \\mid \\mathbf {t} ] \\widetilde {\\mathbf {P}}, \\text {w i t h} \\mathbf {K} = \\left[ \\begin{array}{l l l} f _ {x} & \\gamma & c _ {x} \\\\ 0 & f _ {y} & c _ {y} \\\\ 0 & 0 & 1 \\end{array} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 208, + 785, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $s$ is an arbitrary scale factor, $(\\mathbf{R},\\mathbf{t})$ is the camera pose with respect to the calibration target coordinate, and $\\mathbf{K}$ is the camera intrinsic matrix. $\\widetilde{\\mathbf{p}} = [u,v,1]^T$ and $\\widetilde{\\mathbf{P}} = [X,Y,Z,1]^T$ denote the homogeneous coordinate forms of points $\\mathbf{p}$ and $\\mathbf{P}$ , respectively. The intrinsic matrix $\\mathbf{K}$ consists of five parameters: the focal length $(f_x,f_y)$ , the principal point $(c_x,c_y)$ and skew factor $\\gamma$ .", + "bbox": [ + 212, + 265, + 782, + 342 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Without loss of generality, we assume the planar target is on the plane $Z = 0$ of the world coordinate system. The transformation between the image plane and calibration target plane can be constrained by a $3 \\times 3$ homography matrix, denoted as $\\mathbf{H}$ . For perspective projection, the homography matrix $\\mathbf{H}$ depends on the image pose $(\\mathbf{R},\\mathbf{t})$ and intrinsic matrix $\\mathbf{K}$ :", + "bbox": [ + 212, + 342, + 784, + 417 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} = \\left[ \\begin{array}{l l l} \\mathbf {h} _ {1} & \\mathbf {h} _ {2} & \\mathbf {h} _ {3} \\end{array} \\right] = \\lambda \\mathbf {K} \\left[ \\begin{array}{l l l} \\mathbf {r} _ {1} & \\mathbf {r} _ {2} & \\mathbf {t} \\end{array} \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 382, + 426, + 784, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda$ is an arbitrary scalar, $\\mathbf{r}_i$ is the $i$ -th columns of $\\mathbf{R}$ . Given an image of the calibration target, the homography matrix $\\mathbf{H}$ can be estimated using the plane point correspondences. Combining Eq. (3) and Eq. (5), we can get", + "bbox": [ + 212, + 453, + 782, + 500 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} = \\left[ \\mathbf {r} _ {1} \\mathbf {r} _ {2} - \\mathbf {R t} _ {c p} \\right] = \\frac {1}{\\lambda} \\mathbf {K} ^ {- 1} \\mathbf {H}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 508, + 784, + 535 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Using the knowledge that $\\mathbf{R}$ is an orthogonal matrix, we have", + "bbox": [ + 215, + 542, + 661, + 558 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} ^ {T} \\mathbf {K} ^ {- T} \\mathbf {K} ^ {- 1} \\mathbf {H} = \\lambda^ {2} \\left[ \\begin{array}{c c c} 1 & 0 & - x \\\\ 0 & 1 & - y \\\\ - x & - y & \\| \\mathbf {t} _ {c p} \\| ^ {2} \\end{array} \\right]. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 566, + 784, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "There are 8 parameters to be estimated, 5 for intrinsic matrix and 3 for $\\mathbf{t}_{cp} = (x,y,-r)^T$ . For each given image, Eq. (7) provides 5 independent constraints. Theoretically, a minimum of two images can completely solve all unknown parameters. In the next two subsections, we give a closed-form solver for $N(N > 2)$ collimator images and a minimal solver for two collimator images.", + "bbox": [ + 214, + 623, + 782, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Closed-form Solver for Multiple Views", + "text_level": 1, + "bbox": [ + 215, + 719, + 578, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When there are more than two collimator images for calibration, we model the problem as a linear system and propose a closed-form solution. Suppose that the homography matrices $\\mathbf{H}_i(i = 1,\\dots,N)$ have been estimated for all images. The transformations between images $i$ and $j$ can be expressed as", + "bbox": [ + 212, + 742, + 782, + 803 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} _ {j} ^ {- 1} \\mathbf {H} _ {i} = \\frac {\\lambda_ {i}}{\\lambda_ {j}} \\mathbf {M} _ {j} ^ {- 1} \\mathbf {M} _ {i}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 421, + 811, + 784, + 843 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. Liang, B. Guan et al.", + "bbox": [ + 271, + 114, + 431, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the matrix $\\mathbf{M}$ of each image, we can calculate the determinant as", + "bbox": [ + 215, + 146, + 715, + 161 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left| \\mathbf {M} \\right| = \\left| \\mathbf {r} _ {1} \\mathbf {r} _ {2} - \\mathbf {R t} _ {c p} \\right| = \\left| \\mathbf {r} _ {1} \\mathbf {r} _ {2} - x \\mathbf {r} _ {1} - y \\mathbf {r} _ {2} + r \\mathbf {r} _ {3} \\right| \\tag {9} \\\\ = - x \\left| \\mathbf {r} _ {1} \\mathbf {r} _ {2} \\mathbf {r} _ {1} \\right| - y \\left| \\mathbf {r} _ {1} \\mathbf {r} _ {2} \\mathbf {r} _ {2} \\right| + r \\left| \\mathbf {r} _ {1} \\mathbf {r} _ {2} \\mathbf {r} _ {3} \\right| = r, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 172, + 784, + 209 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $r$ is a non-zero fixed value. Using the fact that $\\operatorname*{det}(\\mathbf{M}_j^{-1}\\mathbf{M}_i) = 1$ , we have $\\lambda_{ij} = \\lambda_i / \\lambda_j = \\sqrt[3]{|\\mathbf{H}_j^{-1}\\mathbf{H}_i|}$ . Then, we choose the image with the most feature points observed as the base image. For convenience, here the first image is selected as the base image and we get $\\lambda_{i1} = \\sqrt[3]{|\\mathbf{H}_j^{-1}\\mathbf{H}_1|}$ . By inverting Eq. (7) of image $i$ , we get", + "bbox": [ + 214, + 219, + 782, + 311 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} _ {i} ^ {- 1} \\mathbf {K} \\mathbf {K} ^ {T} \\mathbf {H} _ {i} ^ {- T} = \\frac {1}{\\lambda_ {i} ^ {2} r ^ {2}} \\left[ \\begin{array}{c c c} r ^ {2} + x ^ {2} & x y & x \\\\ x y & r ^ {2} + y ^ {2} & y \\\\ x & y & 1 \\end{array} \\right]. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 321, + 782, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Substitute $\\lambda_{i} = \\lambda_{i1}\\lambda_{1}$ into Eq. (10), we get", + "bbox": [ + 215, + 378, + 527, + 395 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} _ {i} ^ {- 1} \\mathbf {K K} ^ {T} \\mathbf {H} _ {i} ^ {- T} = \\frac {1}{\\lambda_ {i 1} ^ {2}} \\mathbf {A}, \\text {w i t h} \\mathbf {A} = \\left[ \\begin{array}{c c c} \\frac {r ^ {2} + x ^ {2}}{(\\lambda_ {1} r) ^ {2}} & \\frac {x y}{(\\lambda_ {1} r) ^ {2}} & \\frac {x}{(\\lambda_ {1} r) ^ {2}} \\\\ \\frac {x y}{(\\lambda_ {1} r) ^ {2}} & \\frac {r ^ {2} + y ^ {2}}{(\\lambda_ {1} r) ^ {2}} & \\frac {y}{(\\lambda_ {1} r) ^ {2}} \\\\ \\frac {x}{(\\lambda_ {1} r) ^ {2}} & \\frac {y}{(\\lambda_ {1} r) ^ {2}} & \\frac {1}{(\\lambda_ {1} r) ^ {2}} \\end{array} \\right]. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 289, + 404, + 782, + 463 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Let us define a 6-vector $\\mathbf{a}$ to represent matrix $\\mathbf{A}$ :", + "bbox": [ + 215, + 472, + 570, + 486 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {a} = \\left(A _ {1 1}, A _ {1 2}, A _ {1 3}, A _ {2 2}, A _ {2 3}, A _ {3 3}\\right) ^ {T},\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 497, + 627, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $A_{mn}$ is the value of $m$ -th row and $n$ -th column of matrix $\\mathbf{A}$ . Let", + "bbox": [ + 215, + 525, + 725, + 539 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} = \\left[ \\begin{array}{c c c} c _ {x} ^ {2} + f _ {x} ^ {2} + \\gamma^ {2} & c _ {x} c _ {y} + f _ {y} \\gamma & c _ {x} \\\\ c _ {x} c _ {y} + f _ {y} \\gamma & c _ {y} ^ {2} + f _ {y} ^ {2} & c _ {y} \\\\ c _ {x} & c _ {y} & 1 \\end{array} \\right]. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 549, + 782, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that $\\mathbf{W}$ is a symmetric matrix and can be defined by a 5-vector:", + "bbox": [ + 215, + 607, + 718, + 622 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} = \\left(W _ {1 1}, W _ {1 2}, W _ {1 3}, W _ {2 2}, W _ {2 3}\\right) ^ {T},\n$$\n", + "text_format": "latex", + "bbox": [ + 379, + 632, + 622, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $W_{mn}$ is the value of $m$ -th row and $n$ -th column of matrix $\\mathbf{W}$ . Obviously, Eq. (11) represents a linear system of equations involving the unknowns $[\\mathbf{w}^T, \\mathbf{a}^T]^T$ . Combining Eqs. (11) and (12), we can get the following equations for each image:", + "bbox": [ + 215, + 659, + 782, + 719 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\mathbf {V} - \\lambda_ {i 1} ^ {2} \\mathbf {I} _ {5 \\times 5} \\right] \\left[ \\begin{array}{l} \\mathbf {w} \\\\ \\mathbf {a} \\end{array} \\right] = \\mathbf {b}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 719, + 782, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with", + "bbox": [ + 215, + 755, + 251, + 768 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {V} = \\left[ \\mathbf {v} _ {1 1}, \\mathbf {v} _ {1 2}, \\mathbf {v} _ {1 3}, \\mathbf {v} _ {2 2}, \\mathbf {v} _ {2 3} \\right] ^ {T},\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 767, + 516, + 782 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {m n} = \\left[ h _ {m 1} h _ {n 1}, h _ {m 1} h _ {n 2} + h _ {m 2} h _ {n 1}, h _ {m 1} h _ {n 3} + h _ {m 3} h _ {n 1}, \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 787, + 689, + 801 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left. h _ {m 2} h _ {n 2}, h _ {m 2} h _ {n 3} + h _ {m 3} h _ {n 2} \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 806, + 555, + 821 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {b} = - \\left[ h _ {1 3} h _ {1 3}, h _ {1 3} h _ {2 3}, h _ {1 3} h _ {3 3}, h _ {2 3} h _ {2 3}, h _ {2 3} h _ {3 3} \\right] ^ {T},\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 825, + 638, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Camera Calibration using a Collimator System", + "bbox": [ + 416, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $h_{mn}$ denotes the value of $m$ -th row and $n$ -th column of matrix $\\mathbf{H}^{-1}$ . Given $N(N > 2)$ collimator images, we can stack $N$ such equations as Eq. (13) together in matrix form:", + "bbox": [ + 215, + 146, + 787, + 188 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} \\left[ \\begin{array}{l} \\mathbf {w} \\\\ \\mathbf {a} \\end{array} \\right] = \\mathbf {b}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 455, + 189, + 785, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{D}$ is a $5N\\times 11$ matrix. The closed-form solution of Eq. (14) is given by", + "bbox": [ + 215, + 223, + 781, + 239 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\begin{array}{l} \\mathbf {w} \\\\ \\mathbf {a} \\end{array} \\right] = \\left(\\mathbf {D} ^ {T} \\mathbf {D}\\right) ^ {- 1} \\mathbf {D} ^ {T} \\mathbf {b}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 247, + 785, + 280 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Once $\\mathbf{w}$ and $\\mathbf{a}$ are estimated, we can calculate all camera intrinsic parameters and optical center $\\mathbf{t}_{cp}$ by decomposing $\\mathbf{W}$ and $\\mathbf{A}$ respectively:", + "bbox": [ + 215, + 286, + 785, + 316 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} c _ {x} = W _ {1 3}, \\quad c _ {y} = W _ {2 3}, \\quad f _ {y} = \\sqrt {W _ {2 2} - c _ {y} ^ {2}}, \\\\ \\gamma = \\left(W _ {1 2} - c _ {x} c _ {y}\\right) / f _ {y}, \\quad f _ {x} = \\sqrt {W _ {1 1} - c _ {x} ^ {2} - \\gamma^ {2}}, \\\\ x = A _ {1 3} / A _ {3 3}, \\quad y = A _ {2 3} / A _ {3 3}, \\\\ r = - \\sqrt {\\left(A _ {1 1} / A _ {3 3}\\right) - \\left(A _ {1 3} / A _ {3 3}\\right) ^ {2}}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 327, + 674, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Subsequently, the rotation matrix of each image can be calculated with Eq. (5):", + "bbox": [ + 215, + 420, + 782, + 436 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {r} _ {1} = (1 / \\lambda) \\mathbf {K} ^ {- 1} \\mathbf {h} _ {1}, \\quad \\mathbf {r} _ {2} = (1 / \\lambda) \\mathbf {K} ^ {- 1} \\mathbf {h} _ {2}, \\quad \\mathbf {r} _ {3} = \\mathbf {r} _ {1} \\times \\mathbf {r} _ {2}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 305, + 450, + 784, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\lambda = (\\| \\mathbf{K}^{-1}\\mathbf{h}_1\\| +\\| \\mathbf{K}^{-1}\\mathbf{h}_2\\|) / 2$ . Finally, a singular value decomposition for re-orthogonalizing the rotation matrix $\\mathbf{R} = [\\mathbf{r}_1,\\mathbf{r}_2,\\mathbf{r}_3]$ is required.", + "bbox": [ + 215, + 470, + 785, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 Minimal Solver for Two Views", + "text_level": 1, + "bbox": [ + 215, + 521, + 511, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When only two collimator images are used for calibration, the rank of the coefficient matrix in Eq. (14) is determined to be 10. Solving rank-deficient equations presents a complex challenge. To address this issue, we construct a nonlinear equation system and propose a novel approach that utilizes the hidden variable technique [14] to estimate the unknowns in the minimum configuration.", + "bbox": [ + 215, + 542, + 785, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Estimation of $\\mathbf{W} = \\mathbf{K}^{-T}\\mathbf{K}^{-1}$ is the core of calibration. Note that $\\mathbf{W}$ is a symmetric matrix, defined by a 6-vector $\\mathbf{w} = (W_{11}, W_{12}, W_{13}, W_{22}, W_{23}, W_{33})^T$ . We can extract the following 5 constraints from Eq. (7)", + "bbox": [ + 215, + 619, + 785, + 665 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {1 2} ^ {T} \\mathbf {w} = 0 \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 672, + 784, + 688 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\mathbf {v} _ {1 1} ^ {T} - \\mathbf {v} _ {2 2} ^ {T}\\right) \\mathbf {w} = 0 \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 460, + 693, + 784, + 709 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {1 3} ^ {T} \\mathbf {w} + x \\mathbf {v} _ {1 1} ^ {T} \\mathbf {w} = 0 \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 449, + 713, + 784, + 729 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {2 3} ^ {T} \\mathbf {w} + y \\mathbf {v} _ {1 1} ^ {T} \\mathbf {w} = 0 \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 450, + 734, + 784, + 751 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {3 3} ^ {T} \\mathbf {w} - \\| \\mathbf {t} _ {c p} \\| ^ {2} \\mathbf {v} _ {1 1} ^ {T} \\mathbf {w} = 0. \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 756, + 784, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\mathbf{v}_{mn}$ can be calculated as following:", + "bbox": [ + 215, + 781, + 473, + 796 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {m n} = \\left[ h _ {1 m} h _ {1 n}, h _ {1 m} h _ {2 n} + h _ {1 n} h _ {2 m}, h _ {1 m} h _ {3 n} + h _ {1 n} h _ {3 m} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 805, + 687, + 819 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\left. h _ {2 m} h _ {2 n}, h _ {2 m} h _ {3 n} + h _ {2 n} h _ {3 m}, h _ {3 m} h _ {3 n} \\right] ^ {T},\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 824, + 635, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. Liang, B. Guan et al.", + "bbox": [ + 269, + 114, + 433, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $h_{mn}$ denotes the value of $m$ -th row and $n$ -th column of matrix $\\mathbf{H}$ . Given only 2 collimator images, we can obtain 10 independent equations for 8 unknowns. The hidden variable technique [14] is employed to solve the above polynomial equations. Combining Eqs. (19) to (21), we have", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\mathbf {v} _ {1 3} ^ {T} + \\mathbf {v} _ {2 3} ^ {T} + \\mathbf {v} _ {3 3} ^ {T}\\right) \\mathbf {w} + c \\mathbf {v} _ {1 1} ^ {T} \\mathbf {w} = 0, \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 214, + 785, + 232 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $c = (x + y - \\| \\mathbf{t}_{cp} \\|^2)$ . We first select constraints Eqs. (17), (18) and (22) to establish a new polynomial equation system. Next, we treat $c$ as a hidden variable, and the complete set of equations can be expressed in matrix form as", + "bbox": [ + 212, + 239, + 782, + 286 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} (c) \\mathbf {w} = 0, \\text {w i t h} \\mathbf {C} (c) = \\left[ \\begin{array}{c} \\mathbf {v} _ {1 2} ^ {T} \\\\ (\\mathbf {v} _ {1 1} ^ {T} - \\mathbf {v} _ {2 2} ^ {T}) \\\\ (\\mathbf {v} _ {1 3} ^ {T} + \\mathbf {v} _ {2 3} ^ {T} + \\mathbf {v} _ {3 3} ^ {T}) + c \\mathbf {v} _ {1 1} ^ {T} \\end{array} \\right], \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 292, + 785, + 340 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where the coefficient matrix $\\mathbf{C}(c)$ contains the hidden variable $c$ . Two images provide 6 equations for solving $c$ , and $\\mathbf{C}(c)$ is a $6 \\times 6$ matrix. Since the equation system $\\mathbf{C}(c)\\mathbf{w} = 0$ must have at least one nontrivial solution for $\\mathbf{w}$ , the determinant of $\\mathbf{C}(c)$ is equal to zero. The equation, $\\operatorname*{det}(\\mathbf{C}(c)) = 0$ , is a polynomial of degree 2 in terms of $c$ , which can be efficiently solved using a straightforward root-finding algorithm. Substituting the known value of $c$ into the constraint equations Eqs. (17), (18) and (22), we obtain six constraint equations about $\\mathbf{w}$ . The least squares method is then employed to solve these linear equations and compute the optimal $\\mathbf{w}$ up to scale. Once $\\mathbf{w}$ is estimated, the intrinsic matrix $\\mathbf{K}$ can be calculated using the Cholesky factorization [15].", + "bbox": [ + 212, + 348, + 784, + 500 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4 Bundle Adjustment", + "text_level": 1, + "bbox": [ + 214, + 518, + 421, + 534 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Bundle adjustment jointly refines the camera intrinsic $\\mathbf{K}$ , lens distortion coefficients $\\mathbf{d}$ , the rotation matrix of the $i$ -th image $\\mathbf{R}_i$ and camera optical center $\\mathbf{t}_{cp}$ . In cases of moderate lens distortion, it is typical to initialize the distortion coefficient to $\\mathbf{0}$ and estimate them in the bundle adjustment step [36]. Given a set of $N$ images, each with $M$ corresponding 2D image points, we establish a cost function aimed at minimizing the re-projection error. The cost function is expressed as follows:", + "bbox": [ + 212, + 540, + 782, + 645 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {X}} \\sum_ {i} ^ {N} \\sum_ {j} ^ {M} \\rho \\left(\\left\\| \\pi \\left(\\mathbf {K}, \\mathbf {d}, \\mathbf {R} _ {i}, \\mathbf {t} _ {c p}, \\mathbf {P} _ {i j}\\right) - \\mathbf {p} _ {i j} \\right\\| ^ {2}\\right), \\tag {24}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 654, + 785, + 696 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathbf{X} = (\\mathbf{K},\\mathbf{d},\\mathbf{R}_i,\\mathbf{t}_{cp})$ is the parameter to be refined. $\\mathbf{P}_{ij}$ is the $j$ -th 3D pattern point of $i$ -th image and corresponds to the 2D image point $\\mathbf{p}_{ij}$ . $\\pi (\\cdot)$ projects the world points to image according to Eq. (4). $\\rho (\\cdot)$ is the Cauchy cost function. As is common, we optimize Eq. (24) with the Levenberg-Marquardt (LM) method [27], which is implemented in some excellent frameworks [1].", + "bbox": [ + 212, + 703, + 782, + 779 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Note that the camera optical center is fixed at $\\mathbf{t}_{cp}$ with the spherical motion constraint. However, for general motion, the camera optical center is free. Therefore, when calibrating the camera using collimator images, the number of motion parameters to be refined is reduced from $6N$ to $3N + 3$ .", + "bbox": [ + 212, + 780, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Camera Calibration using a Collimator System", + "bbox": [ + 416, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.5 Degenerated Configuration", + "text_level": 1, + "bbox": [ + 215, + 146, + 485, + 162 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We prove a degenerate configuration where additional images cannot provide more constraints about the camera intrinsic parameter.", + "bbox": [ + 212, + 172, + 785, + 202 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Proposition 1. If the calibration target rotates only around an axis that is parallel to the $z$ axis and intersects the target plane at $(x,y,0)^T$ , the newly added image cannot provide additional constraints on the camera intrinsic parameters.", + "bbox": [ + 212, + 203, + 785, + 247 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Due to space limitations, the proof is provided in the supplementary material.", + "bbox": [ + 238, + 250, + 785, + 263 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Evaluation and Results", + "text_level": 1, + "bbox": [ + 215, + 287, + 477, + 304 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we first evaluate the robustness and accuracy of the proposed algorithm on synthetic data. Then, we calibrate the camera and evaluate the effectiveness of the proposed method using real collimator images.", + "bbox": [ + 212, + 321, + 785, + 367 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Synthetic Data Experiments", + "text_level": 1, + "bbox": [ + 215, + 390, + 496, + 406 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In the synthetic data experiment, we evaluate the performance of the proposed algorithm. We simulate the process of calibrating predefined camera parameters using synthetic images that conform to the spherical motion model. The proposed is compared with the following algorithms:", + "bbox": [ + 212, + 416, + 785, + 476 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Zhang [46] is a plane-based algorithm and suitable for general 6DOF motion.", + "- Bouguet [3] is a calibration toolbox and suitable for general 6DOF motion. Unlike Zhang, Bouguet computes the focal length using the orthogonal vanishing points constraint.", + "- Hartley [17] is a self-calibration method based on a pure rotating camera, which is specifically suitable for the 3DOF spherical motion model." + ], + "bbox": [ + 212, + 477, + 782, + 565 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We define a virtual camera with the following parameters: $f_{x} = f_{y} = 1000$ , $c_{x} = 542$ , $c_{y} = 478$ , $\\gamma = 0.01$ , radial distortion coefficients $d_{1} = 0.1$ , $d_{2} = -0.2$ , and the image size is set to $1080 \\times 960$ pixels. The calibration target is a planar pattern containing $11 \\times 8$ points, each square size of $30mm \\times 30mm$ . The target maintains a spherical motion with a radius of $700mm$ relative to the camera. The camera optical center in the target coordinate system is $t_{cp} = [150, 105, -700]^{T}mm$ , and the camera's orientation is arbitrary. Given the camera parameters and poses, the planar target points are projected onto images. Finally, zero-mean Gaussian noise with $\\sigma$ standard deviation is added to 2D image points.", + "bbox": [ + 212, + 566, + 794, + 703 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation of Calibration. In this experiment, we seek to evaluate the performance of the algorithms in solving camera parameters. For the four algorithms, we perform two steps of solving initial value and bundle adjustment to obtain camera parameters. Firstly, we compare the performance of the mentioned algorithms as the noise level increases. The noise level varies from 0 to 3.0 pixels, and we use 15 images to calibrate the camera. For each noise level, we conducted 200 trials and calculated the average errors. As shown in Fig. 3, calibration error increases linearly with the noise level. Compared with the other three algorithms, ours shows greater robustness to noise. Additionally, at a noise", + "bbox": [ + 212, + 704, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. Liang, B. Guan et al.", + "bbox": [ + 271, + 114, + 431, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/2e124ae4208e52f24641e2988d76866c8ef6bc7fdb5781abe9eaada7c108dc26.jpg", + "image_caption": [ + "Fig. 3: Comparison of calibration errors with increasing noise level." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 359, + 244 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1c25801bcbb8e6190215e5e34675e904e1525946a08b499cf88fbae47fe9765a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 143, + 496, + 244 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/bc1076d33f01ccb87ffdf2f913b9c964067086737439128326e4dad7b4190a50.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 142, + 638, + 244 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9e7ca5d23c88494883ca0fa5b7f0a0e794cb7a7d8202c7e47660a30cc773620b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 142, + 784, + 244 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d93f5f3c79cb58859b4b3fcf9c38bddd090015ddb8eff8270397ed3568f3e503.jpg", + "image_caption": [ + "Fig. 4: Comparison of calibration errors with the number of images increases." + ], + "image_footnote": [], + "bbox": [ + 215, + 304, + 359, + 404 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e9df0542566cdf33338b02f3c618fdd99fb618ee7483fea2d108057fc1feaaaf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 305, + 496, + 405 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/64074298e4f3bcd9f9414e821b3874e7e72962c09b4337c94b6f621eacd8bd56.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 304, + 640, + 405 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7ab5cbb4daf7b03699e49b69f7f31651c5c52c38a1ddac561eac6757314e7c67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 304, + 784, + 405 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "level of 0.5 pixels, we evaluate the performance with respect to the number of calibration images. The image count ranged from 3 to 30, with 200 trials conducted for each quantity. The results are illustrated in Fig. 4. A significant decrease in error is observed when increasing the number of images from 3 to 5. Both Fig. 3 and Fig. 4 show that our algorithm outperforms other algorithms in robustness and accuracy. We believe that our algorithm benefits from the consideration of both spherical motion and planar homography constraints, along with reducing optimization variables. Due to space limitations, the additional comparison results of the initial value estimation are shown in the supplementary material.", + "bbox": [ + 212, + 481, + 787, + 618 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sensitivity of imperfect spherical motion. The low-cost collimators exhibit an inferior parallelism, which is reflected in the fact that the relative motion between the target and camera is not a perfect spherical motion. This experiment evaluates the sensitivity of initial value estimation with respect to imperfect spherical motion. We add zeros-mean Gaussian noise with standard deviation from $0mm$ to $30mm$ to $\\mathbf{t}_{cp}$ to replicate this imperfection. We use 15 images and add a noise level of 0.5 pixels to the image points. For each noise level, a total of independent 500 trials are conducted and the average errors are computed. The sensitivity analysis results are shown in Fig. 5. Noise has no impact on Bouguet and Zhang, because they do not rely on the spherical motion constraint. As the noise increases, calibration errors of Ours and Hartley increase gradually. Notably, Hartley is significantly influenced by noise, whereas Ours displays superior robustness. Even with noise levels below $15mm$ (equivalent to $2.143\\%$ of the spherical radius), our algorithm remains superior to the alternatives.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Camera Calibration using a Collimator System", + "bbox": [ + 416, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/4a215737a0b6bb2ae96cecbb6dbe73a3c26e43c0929e5dcb1e765036cd3a4a52.jpg", + "image_caption": [ + "Fig. 5: Sensitivity analysis of imperfect spherical motion." + ], + "image_footnote": [], + "bbox": [ + 316, + 143, + 493, + 265 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b10433ef99734c73e1bca81cff90288ef225737bc6e852db743a65f53bb80c69.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 142, + 684, + 266 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/8c16e1b52e515b58e928759256a3f8660620e249a604f07a6054dca67f583f53.jpg", + "table_caption": [ + "Table 1: Comparison of calibration and pose estimation error results (pixels)." + ], + "table_footnote": [], + "table_body": "
NumberZhang [46]Bouguet [3]Hartley [17]Ours
εcalibεevalεcalibεevalεcalibεevalεcalibεeval
20.18420.15450.17390.1518--0.17390.1519
50.23160.12450.22020.1236-0.25770.22030.1236
100.19230.12310.19120.1232-0.19860.19510.1222
150.17750.12210.17180.1232-0.17910.17790.1216
200.17730.12250.17220.1221-0.18710.17780.1207
", + "bbox": [ + 303, + 330, + 696, + 421 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Real Images Experiments", + "text_level": 1, + "bbox": [ + 215, + 450, + 473, + 465 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In real image experiments, we verify the effectiveness of our collimator system for calibration and evaluate the performance of the proposed algorithm.", + "bbox": [ + 212, + 477, + 785, + 506 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Calibration and Accuracy Evaluation. Following the experimental methodology utilized in previous studies [13,30], we design a pose estimation experiment to evaluate the accuracy of camera calibration quantitatively. A 8mm camera captures images from our collimator system and a practical printed pattern. Calibration is performed using the collimator images and a subset of the printed pattern images, while the remaining printed pattern images are used to evaluate the calibration results. After obtaining the camera parameters from the calibration images, the pose of each evaluation image can be determined using absolute pose estimation [24]. In our setup, a quarter of the corner points on the printed pattern are used for pose estimation, and the remaining points are used for evaluation. By re-projecting the evaluation corner points onto the image with the camera parameters and the calculated pose, we can evaluate the calibration quality based on the re-projection error. A total of 200 printed pattern images are used to evaluate the calibration. The calibration uses 2, 5, 10, 15, and 20 images. The re-projection error of the calibration and evaluation images are donated as $\\epsilon_{calib}$ and $\\epsilon_{eval}$ , respectively.", + "bbox": [ + 212, + 507, + 785, + 750 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The comparison results of the four algorithms are shown in Tab. 1. Due to the limited degrees of freedom, our calibrated re-projection error $(\\epsilon_{calib})$ is slightly higher. However, Ours exhibits the lowest re-projection error in evaluation images compared to alternative methods, indicating that our method achieves more precise camera parameters. The Hartley method only uses the correspondence of 2D points in images, which makes it impossible to calculate the re-projection", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. Liang, B. Guan et al.", + "bbox": [ + 269, + 114, + 431, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/bcc5c90d60bb7c14649c13bfd890ce1e7ad8ed32cd593ff76150a7320f263760.jpg", + "table_caption": [ + "Table 2: Comparison of calibration results using 20 images in a real camera." + ], + "table_footnote": [], + "table_body": "
PrintedCollimator
Bouguet [3]Zhang [46]Bouguet [3]Hartley [17]Ours
fx2384.52380.82369.22644.62388.3
fy2386.32379.82368.92646.02387.9
cx1209.81222.11221.11214.61221.0
cy1008.21008.41009.81012.71009.6
k1-0.0843-0.0899-0.0908-0.0754-0.0900
k20.08190.09120.08920.11810.0954
εcalib0.15130.17730.1722-0.1778
εeval0.12250.12250.12210.18710.1207
", + "bbox": [ + 316, + 170, + 679, + 297 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/6841262af48481af4e85a3e1c955aaef82b4ec39fa66d3b2cfdba465bd903206.jpg", + "table_caption": [ + "Table 3: Comparison of epipolar rectification error results (pixels)." + ], + "table_footnote": [], + "table_body": "
NumberZhang [46]Bouguet [3]Hartley [17]Ours
meanstdmeanstdmeanstdmeanstd
20.08460.04060.07700.0220--0.07700.0229
50.08090.03070.07640.02140.08320.03970.07620.0211
100.07730.02340.07580.02130.08540.05090.07570.0208
150.07820.02660.07710.02530.08370.05220.07560.0207
200.07720.02520.07680.02530.08180.03650.07560.0208
", + "bbox": [ + 303, + 339, + 699, + 429 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "error of the calibration images directly. In cases with only two images, Zhang and Bouguet commonly depend on beneficial assumptions, such as $\\gamma = 0$ . Ours achieves comparable evaluation errors without the need for such assumptions.", + "bbox": [ + 212, + 459, + 785, + 503 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Additionally, the Bouguet toolbox is also used to calculate camera parameters from printed pattern images that are not included in the evaluation. We use 20 images for calibration and 200 for evaluation. The estimated camera parameters and error comparison are shown in Tab. 2. The second column is a good reference. The lower evaluation error indicates that our method outperforms the alternative and validates the effectiveness of using the collimator image for calibration. It can be expected that the calibration parameters are not significantly different. Therefore, even a minor improvement holds considerable significance.", + "bbox": [ + 212, + 505, + 785, + 626 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Then, we evaluate the performance of all methods using epipolar rectification error, which has not been optimized. Given the camera parameters from different methods, we calculate the epipolar rectification error (in pixels) for 300 pairs of printed evaluation images. The mean and standard deviation of the errors are shown in Tab. 3. Moreover, the result using Bouguet with 20 printed pattern images is (0.0772, 0.0259) pixels. It is observed that as the number of calibration images increases, the error stabilizes. The results indicate improved accuracy in the collimator image and highlight the beneficial impact of the spherical constraint on accuracy.", + "bbox": [ + 212, + 627, + 785, + 763 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Multiple Cameras Calibration with a Same Collimator. In this experiment, we evaluate our method on multiple cameras with the same collimator. We used the cameras with $12mm$ , $16mm$ , $25mm$ , and $35mm$ focal lengths of lenses. The calibration images are captured from the same collimator, as shown in Fig. 6 and left of Fig. 1. As the focal length increases, the more pixels a single", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Camera Calibration using a Collimator System", + "bbox": [ + 416, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/832534a9e98b421a82ef6735dfe964b146ccf1b579a1712e7798c7ae8f7fe012.jpg", + "image_caption": [ + "Fig. 6: Calibration image sample obtained from the collimator." + ], + "image_footnote": [], + "bbox": [ + 357, + 143, + 643, + 224 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/8cebb14b2c242b2dad97b25f7ad4542d9309b70efd1b50229402f6a2be3961ce.jpg", + "table_caption": [ + "Table 4: Comparison of calibration results with different focal length lens (pixels)." + ], + "table_footnote": [], + "table_body": "
Zhang [46]Bouguet [3]Hartley [17]Ours
12mm0.22730.22250.34040.2174
16mm0.29940.25910.41820.2536
25mm0.31330.30780.41950.2923
35mm0.32890.33480.72720.3118
", + "bbox": [ + 328, + 292, + 669, + 367 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "square occupies. Cameras with different focal lengths can observe the calibration target through our collimator system in a limited space.", + "bbox": [ + 212, + 398, + 782, + 428 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Similar to the evaluation approach in the previous experiment, we segregate all collimator images into calibration and evaluation sets. For each lens, we use 20 images for calibration and 200 for evaluation. The results are listed in Tab. 4. The proposed algorithm demonstrates the lowest re-projection error for lenses with varying focal lengths among the evaluation images. The re-projection error tends to increase with the increase of focal length. We believe that the main reason is that the longer the lens's focal length, the fewer points observed. More importantly, this experiment confirms the effectiveness of our collimator system and the proposed method. Moreover, we provide a spherical motion verification experiment in supplementary material.", + "bbox": [ + 212, + 429, + 784, + 580 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 607, + 357, + 622 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper presents a novel camera calibration method using a designed collimator system. Based on the optical geometry of the collimator system, we prove that the relative motion between the target and the camera conforms to a spherical motion. Furthermore, we propose a closed-form solver for multiple views and a minimal solver for two views for camera calibration. Our collimator-based method effectively combines the advantages of close-range calibration using a collimator and flexibility using a planar calibration target. We validate the effectiveness of using only collimator images to calibrate the camera. Synthetic and real data experiments demonstrate that the calibration accuracy of the proposed method is superior to the state-of-the-art methods.", + "bbox": [ + 212, + 642, + 784, + 792 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations: When calibrating a fisheye camera, the distortion should be paid more attention rather than simply setting it to 0. Future work will focus on cameras with significant distortion.", + "bbox": [ + 212, + 795, + 784, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. Liang, B. Guan et al.", + "bbox": [ + 271, + 114, + 431, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research has been supported by the Hunan Provincial Natural Science Foundation for Excellent Young Scholars under Grant 2023JJ20045, and the National Natural Science Foundation of China under Grant 12372189. We thank Zijie Wang for useful discussion on the writing of the paper.", + "bbox": [ + 212, + 176, + 787, + 239 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 263, + 323, + 277 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Agarwal, S., Mierle, K., Team, T.C.S.: Ceres Solver. https://github.com/ceres-solver/ceres-solver", + "2. Bauer, M., Grießbach, D., Hermerschmidt, A., Krüger, S., Scheele, M., Schischmanow, A.: Geometrical camera calibration with diffractive optical elements. Optics express 16(25), 20241-20248 (2008)", + "3. Bouguet, J.Y.: Camera calibration toolbox for matlab. http://robots.stanford.edu/cs223b04/JeanYvesCalib/index.html (2004)", + "4. Bradski, G.: The opencv library. Dr. Dobb's Journal: Software Tools for the Professional Programmer 25(11), 120-123 (2000)", + "5. Brown, D.C.: Close-range camera calibration. Photogramm. Eng 37(8), 855-866 (1971)", + "6. Campos, C., Elvira, R., Rodríguez, J.J.G., M. Montiel, J.M., D. Tardós, J.: Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics 37(6), 1874-1890 (2021)", + "7. Clarke, T.A., Fryer, J.G.: The development of camera calibration methods and models. The Photogrammetric Record 16(91), 51-66 (1998)", + "8. Faugeras, O.D., Luong, Q.T., Maybank, S.J.: Camera self-calibration: Theory and experiments. In: IEEE European Conference on Computer Vision (ECCV). p. 321-334 (1992)", + "9. Fuhr, G., Jung, C.R.: Camera self-calibration based on nonlinear optimization and applications in surveillance systems. IEEE Transactions on Circuits and Systems for Video Technology 27(5), 1132-1142 (2017)", + "0. Guan, B., Su, Z., Yu, Q., Li, Z., Feng, W., Yang, D., Zhang, D.: Monitoring the blades of a wind turbine by using videogrammetry. Optics and Lasers in Engineering 152, 106901 (2022)", + "1. Guan, B., Zhao, J., Barath, D., Fraundorfer, F.: Minimal solvers for relative pose estimation of multi-camera systems using affine correspondences. International Journal of Computer Vision 131(1), 324-345 (2023)", + "2. Guan, B., Zhao, J., Li, Z., Sun, F., Fraundorfer, F.: Relative pose estimation with a single affine correspondence. IEEE Transactions on Cybernetics 52(10), 10111-10122 (2022)", + "3. Ha, H., Perdoch, M., Alismail, H., Kweon, I.S., Sheikh, Y.: Deltille grids for geometric camera calibration. In: IEEE International Conference on Computer Vision (ICCV). pp. 5354-5362 (2017)", + "4. Hartley, R., Li, H.: An efficient hidden variable approach to minimal-case camera motion estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence 34(12), 2303-2314 (2012)", + "5. Hartley, R., Zisserman, A.: Multiple View Geometry in Computer Vision. Cambridge University Press, 2 edn. (2004)" + ], + "bbox": [ + 225, + 295, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Camera Calibration using a Collimator System", + "bbox": [ + 416, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Hartley, R.I.: Self-calibration from multiple views with a rotating camera. In: European Conference on Computer Vision (ECCV). pp. 471-478 (1994)", + "17. Hartley, R.I.: Self-calibration of stationary cameras. International Journal of Computer Vision 22(1), 5-23 (1997)", + "18. Heikkila, J.: Geometric camera calibration using circular control points. IEEE Transactions on Pattern Analysis and Machine Intelligence 22(10), 1066-1077 (2000)", + "19. Herrera, D., Kannala, C.J., Heikkila, J.: Forget the checkerboard: Practical self-calibration using a planar scene. In: IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 1-9. IEEE (2016)", + "20. Hieronymus, J.: Comparison of methods for geometric camera calibration. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences 39, 595-599 (2012)", + "21. Hu, H., Wei, B., Mei, S., Liang, J., Zhang, Y.: A two-step calibration method for vision measurement with large field of view. IEEE Transactions on Instrumentation and Measurement 71, 1-10 (2022)", + "22. Kannala, J., Brandt, S.S.: A generic camera model and calibration method for conventional, wide-angle, and fish-eye lenses. IEEE Transactions on Pattern Analysis and Machine Intelligence 28(8), 1335-1340 (2006)", + "23. Klaus, A., Bauer, J., Karner, K., Elbischger, P., Perko, R., Bischof, H.: Camera calibration from a single night sky image. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). vol. 1, pp. I-I (2004)", + "24. Kneip, L., Furgale, P.: Opengv: A unified and generalized approach to real-time calibrated geometric vision. In: IEEE International Conference on Robotics and Automation (ICRA). pp. 1-8 (2014)", + "25. Liu, Z., Li, F., Li, X., Zhang, G.: A novel and accurate calibration method for cameras with large field of view using combined small targets. Measurement 64, 1-16 (2015)", + "26. Lochman, Y., Liepieshov, K., Chen, J., Perdoch, M., Zach, C., Pritts, J.: Babel-calib: A universal approach to calibrating central cameras. In: IEEE International Conference on Computer Vision (ICCV). pp. 15253-15262 (2021)", + "27. Moré, J.J.: The levenberg-marquardt algorithm: implementation and theory. In: Numerical analysis: proceedings of the biennial Conference held at Dundee, June 28-July 1, 1977. pp. 105-116 (2006)", + "28. Nister, D., Stewenius, H., Grossmann, E.: Non-parametric self-calibration. In: IEEE International Conference on Computer Vision (ICCV). vol. 1, pp. 120-127 (2005)", + "29. Oniga, V.E., Pfeifer, N., Loghin, A.M.: 3d calibration test-field for digital cameras mounted on unmanned aerial systems (uas). Remote Sensing 10(12), 2017 (2018)", + "30. Peng, S., Sturm, P.: Calibration wizard: A guidance system for camera calibration based on modelling geometric and corner uncertainty. In: IEEE International Conference on Computer Vision (ICCV). pp. 1497-1505 (2019)", + "31. Pollefeys, M., Koch, R., Gool, L.V.: Self-calibration and metric reconstruction in spite of varying and unknown intrinsic camera parameters. International journal of computer vision 32(1), 7-25 (1999)", + "32. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4104-4113 (2016)", + "33. Schöps, T., Larsson, V., Pollefeys, M., Sattler, T.: Why having 10,000 parameters in your camera model is better than twelve. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2532-2541 (2020)" + ], + "bbox": [ + 217, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. Liang, B. Guan et al.", + "bbox": [ + 269, + 114, + 431, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Shang, Y., Sun, X., Yang, X., Wang, X., Yu, Q.: A camera calibration method for large field optical measurement. Optik - International Journal for Light and Electron Optics 124, 6553-6558 (12 2013)", + "35. Sturm, P.F., Maybank, S.J.: On plane-based camera calibration: A general algorithm, singularities, applications. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). vol. 1, pp. 432-437. IEEE (1999)", + "36. Triggs, B., McLauchlan, P.F., Hartley, R.I., Fitzgibbon, A.W.: Bundle adjustment—a modern synthesis. In: Vision Algorithms: Theory and Practice: International Workshop on Vision Algorithms. pp. 298-372. Springer (2000)", + "37. Tsai, R.: A versatile camera calibration technique for high-accuracy 3d machine vision metrology using off-the-shelf tv cameras and lenses. IEEE Journal on Robotics and Automation 3(4), 323-344 (1987)", + "38. Ventura, J.: Structure from motion on a sphere. In: European Conference on Computer Vision (ECCV). vol. 9907, pp. 53-68 (2016)", + "39. Wang, C., Liu, Y., Wang, Y., Li, X., Wang, M.: Efficient and outlier-robust simultaneous pose and correspondence determination by branch-and-bound and transformation decomposition. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(10), 6924-6938 (2022)", + "40. Wang, Z., Wu, Z., Zhen, X., Yang, R., Xi, J., Chen, X.: A two-step calibration method of a largefov binocular stereovision sensor for onsite measurement. Measurement 62, 15-24 (2015)", + "41. Wu, G., Han, B., He, X.: Calibration of geometric parameters of line array ccd camera based on exact measuring angle in lab. Optics and Precision Engineering 15(10), 1628-1632 (2007)", + "42. Xiao, Z., Jin, L., Yu, D., Tang, Z.: A cross-target-based accurate calibration method of binocular stereo systems with large-scale field-of-view. Measurement 43(6), 747-754 (2010)", + "43. Yuan, G., Zheng, L., Ding, Y., Zhang, H., Zhang, X., Liu, X., Sun, J.: A precise calibration method for line scan cameras. IEEE Transactions on Instrumentation and Measurement 70, 1-9 (2021)", + "44. Yuan, G., Zheng, L., Sun, J., Liu, X., Wang, X., Zhang, Z.: Practical calibration method for aerial mapping camera based on multiple pinhole collimator. IEEE Access 8, 39725-39733 (2019)", + "45. Zhang, K., Xie, J., Snavely, N., Chen, Q.: Depth sensing beyond lidar range. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1689-1697 (2020)", + "46. Zhang, Z.: A flexible new technique for camera calibration. IEEE Transactions on Pattern Analysis and Machine Intelligence 22(11), 1330-1334 (2000)", + "47. Zhang, Z.: Camera calibration with one-dimensional objects. IEEE Transactions on Pattern Analysis and Machine Intelligence 26(7), 892-899 (2004)" + ], + "bbox": [ + 212, + 146, + 787, + 702 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Camera Calibration using a Collimator System", + "bbox": [ + 416, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Camera Calibration using a Collimator System/26f230b2-04ba-4e24-b664-739a6c3feb85_model.json b/2024/Camera Calibration using a Collimator System/26f230b2-04ba-4e24-b664-739a6c3feb85_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5cbdb1f520dda261868b71ed963d39196e4e850f --- /dev/null +++ b/2024/Camera Calibration using a Collimator System/26f230b2-04ba-4e24-b664-739a6c3feb85_model.json @@ -0,0 +1,2698 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.233, + 0.142, + 0.773, + 0.166 + ], + "angle": 0, + "content": "Camera Calibration using a Collimator System" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.19, + 0.78, + 0.222 + ], + "angle": 0, + "content": "Shunkun Liang, Banglei Guan, Zhenbao Yu, Pengju Sun, and Yang Shang" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.233, + 0.761, + 0.262 + ], + "angle": 0, + "content": "College of Aerospace Science and Engineering, National University of Defense Technology, China." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.299, + 0.744, + 0.522 + ], + "angle": 0, + "content": "Abstract. Camera calibration is a crucial step in photogrammetry and 3D vision applications. In practical scenarios with a long working distance to cover a wide area, target-based calibration methods become complicated and inflexible due to site limitations. This paper introduces a novel camera calibration method using a collimator system, which can provide a reliable and controllable calibration environment for cameras with varying working distances. Based on the optical geometry of the collimator system, we prove that the relative motion between the target and camera conforms to the spherical motion model, reducing the original 6DOF relative motion to 3DOF pure rotation motion. Furthermore, a closed-form solver for multiple views and a minimal solver for two views are proposed for camera calibration. The performance of our method is evaluated in both synthetic and real-world experiments, which verify the feasibility of calibration using the collimator system and demonstrate that our method is superior to the state-of-the-art methods. Demo code is available at https://github.com/LiangSK98/CollimatorCalibration" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.534, + 0.741, + 0.562 + ], + "angle": 0, + "content": "Keywords: Camera calibration \\(\\cdot\\) Collimator system \\(\\cdot\\) Spherical motion \\(\\cdot\\) Closed-form solver \\(\\cdot\\) Minimal solver" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.589, + 0.377, + 0.604 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.621, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Geometric camera calibration determines the mapping between 3D rays in space and 2D image pixels. Various vision applications, such as simultaneous localization and mapping(SLAM) [6], structure from motion(SfM) [32], and pose estimation [11, 12], rely on camera parameters as inputs. The close-range calibration methods using a planar target represent the prevailing techniques in 3D vision applications [5, 26, 33, 35, 46]. In some special outdoor scenarios, such as long-range depth estimation [45] and monitoring of wind turbine blades [10], the cameras achieve a long working distance and a large field of view. The traditional target-based calibration methods [3, 37, 46] may face challenges, such as needing large targets, moving or arranging targets over a wide area, and dealing with varying and unstable lighting conditions. While constructing a 3D calibration field can offer precise calibration data [34, 40, 42], its operational complexity and limited adaptability present significant challenges. In addition, the collimator is" + }, + { + "type": "page_footnote", + "bbox": [ + 0.231, + 0.827, + 0.608, + 0.84 + ], + "angle": 0, + "content": "Corresponding author. guanbanglei12@nudt.edu.cn" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.434, + 0.129 + ], + "angle": 0, + "content": "S. Liang, B. Guan et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.142, + 0.382, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.143, + 0.547, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.143, + 0.787, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.284, + 0.788, + 0.327 + ], + "angle": 0, + "content": "Fig. 1: Diagram of the collimator-based calibration method. From the inherent optical geometric properties of the collimator, the relative motion between the target and camera can be proved to conform to the spherical motion model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.784, + 0.414 + ], + "angle": 0, + "content": "commonly used to generate calibration targets [7]. Traditional collimator-based methods need to measure the direction of collimated rays using a goniometer or theodolite, which makes the calibration process challenging, and measurement errors directly impact the accuracy [41, 43, 44]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.785, + 0.625 + ], + "angle": 0, + "content": "In this paper, we design a novel collimator system with a structured calibration pattern on its reticle, and a novel calibration algorithm based on the collimator images is proposed. Our method only requires the camera to observe the pattern from our collimator system at a few (at least two) different orientations. Fig. 1 illustrates this method. We briefly analyze the geometric optical property of the collimator and illustrate the angle invariance of any point pairs on the pattern. By leveraging angle invariance, we prove that the relative motion between the calibration target and the camera conforms to a spherical motion model. This constraint effectively reduces the original 6DOF general motion to the 3DOF pure rotation motion. Furthermore, we propose a novel calibration method using the collimator system. Compared with the target-based methods, more geometric constraints can be derived from the collimator, leading to higher accuracy. Compared to traditional collimator-based methods, our method is more flexible as it eliminates the need for direction measurement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.787, + 0.761 + ], + "angle": 0, + "content": "Our primary contributions can be summarized as follows: (1) We propose a novel camera calibration method using a collimator system. The entire calibration process can be implemented with multiple close-range observations and does not require the known direction of collimated rays. (2) We prove that the relative motion between the calibration target and the camera conforms to the spherical motion model. This constraint effectively reduces the number of motion parameters to be estimated. (3) A closed-form solver for multiple views and a minimal solver for two views are proposed based on the spherical motion constraint. Moreover, the degenerate case of the proposed solvers is proved." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.781, + 0.389, + 0.797 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Camera calibration is a crucial subject in 3D vision and continues to be actively researched. The literature on camera calibration introduces a wide range of al" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Camera Calibration using a Collimator System" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.265 + ], + "angle": 0, + "content": "gorithms and toolboxes [3-5,22,26,33,35,46]. To ensure the highest accuracy for calibration, targets with known structures are usually used for camera calibration. Various calibration targets have been proposed, including 3D targets [37], 2D targets [13,18] and 1D targets [47]. In some practical applications where no calibration target is available, the camera calibration can be performed by establishing the correspondence between images during the movement of the camera, which is known as self-calibration [8,9,16,19,28,31]. Among these options, the target-based close-range calibration method is widely used." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.785, + 0.403 + ], + "angle": 0, + "content": "In outdoor long-range measurement applications [10, 45], the camera has a long working distance to measure the long-distance targets. Thus, the close-range calibration methods are impractical due to the site limitation. A commonly employed approach involves constructing a calibration field, which serves as a known structure for the calibration [21, 25, 29, 34, 40, 42]. Oniga et al. [29] lay markers on buildings to calibrate cameras mounted on UAV. Shang et al. [34] proposed a high-precision calibration method by placing several markers near the ground. Additionally, Klaus et al. [23] proposed using stars as distant targets to estimate camera parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.787, + 0.569 + ], + "angle": 0, + "content": "The collimator is also commonly used to provide 2D-3D correspondence for camera calibration in the laboratory [2, 20, 41, 44]. Previous collimator-based methods proposed to use goniometer or theodolite to obtain the direction of collimated rays and then estimate the camera parameters [41, 43]. The collimator-based method offers both accuracy and flexibility in the limited space, but a high-precision goniometer or theodolite is expensive. To address this, Yuan et al. [44] designed a multiple pinhole mask for the collimator to generate the collimated rays with known directions. However, the accuracy of calibration is sensitive to machining errors. Bauer et al. [2] proposed a calibration method using a diffractive optical element (DOE). The angle between the laser beam and the camera can then be estimated to determine the camera parameters." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.594, + 0.433, + 0.612 + ], + "angle": 0, + "content": "3 Collimator System" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.784, + 0.671 + ], + "angle": 0, + "content": "This section introduces the designed collimator system and shows that the equivalent relative motion between the calibration target and the camera conforms to a spherical motion model." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.694, + 0.506, + 0.71 + ], + "angle": 0, + "content": "3.1 Scheme of Collimator System" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.785, + 0.84 + ], + "angle": 0, + "content": "The collimator is an optical instrument that can produce collimated rays and plays an important role when adjusting other optical instruments. A collimator consists of essential components such as a light source, a reticle, and an optical lens, as illustrated in Fig. 2. The reticle is a thin glass located on the focal plane of the collimator. The generation of collimated rays can be regarded as the inverse process of imaging a target at infinity onto the focal plane. Specifically, a point on the lens's focal plane will generate collimated rays, which can be regarded as the target at infinity. Therefore, cameras with varying focal lengths" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.433, + 0.129 + ], + "angle": 0, + "content": "S. Liang, B. Guan et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "can consistently and accurately observe the target through the collimator at a close range." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.359 + ], + "angle": 0, + "content": "In this paper, we design a collimator system for camera calibration. A star-based pattern, proposed by Thomas et al. [33], is attached to the reticle of the collimator as the calibration target. The star-based pattern offers richer gradient information for improving feature extraction accuracy. Similar to the plane-based calibration method [35, 46], the star-based pattern provides 2D-3D correspondences for camera calibration. Our collimator system is portable with a size of about \\(200\\mathrm{mm} \\times 170\\mathrm{mm} \\times 300\\mathrm{mm}\\). The collimator system is designed to provide a reliable and controllable calibration environment for some special scenarios like inappropriate illumination, limited space, and long-range cameras. However, printed targets face challenges in these scenarios. When calibrating a camera, we only need to capture the target from multiple orientations through the lens of the collimator, which is as easy as the printed targets method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.379, + 0.463, + 0.394 + ], + "angle": 0, + "content": "3.2 Spherical Motion Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.489, + 0.689 + ], + "angle": 0, + "content": "In this section, we prove that the relative motion between the calibration target and the camera conforms to the spherical motion model. Firstly, we briefly analyze the geometric property of the collimator. As shown in Fig. 2, let us note two arbitrary points \\( P_{1} \\) and \\( P_{2} \\) on the reticle of the collimator. Rays emitted from \\( P_{1} \\) and \\( P_{2} \\) pass through the collimator lens, generating two parallel beams observed by cameras denoted as \\( C_1 \\) and \\( C_2 \\) with arbitrary poses. The propagation routes of the two beams are visually distinguished by red and blue lines in Fig. 2. Obviously, the angle between two lines observed by the two cameras is equal, that is, \\( \\theta_{1}\\equiv \\theta_{2}\\equiv \\theta \\). We can obtain a property that the two rays" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.431, + 0.772, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.593, + 0.788, + 0.649 + ], + "angle": 0, + "content": "Fig. 2: Diagram of collimator system. The geometric property of the collimator system leads to the angle invariance of any point pair on the reticle." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.785, + 0.72 + ], + "angle": 0, + "content": "observed by the cameras always maintain a constant angle, regardless of the poses of observing cameras. We refer to this property as the angle invariance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Next, we show that this setup, where a camera moves around a collimator system and captures images, is theoretically equivalent to the capture setting where the camera is fixed, but the images are moving around the camera spherically. This can be derived from the angle invariance. For convenience in introducing our derivation process, we treat the camera as a fixed reference coordinate, and assume that the calibration target moves relative to the camera in space. The angle invariance allows us to infer that when the target performs a specific motion in space, the angular distance between any point pair on the target always" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Camera Calibration using a Collimator System" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "remains constant. The angular distance means the angle between two vectors from the origin (camera optical center) to two points. By arbitrarily selecting a pair of points \\(\\mathbf{P}_i\\) and \\(\\mathbf{P}_j\\) on the target and transforming them, we can establish the following equation" + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.217, + 0.786, + 0.233 + ], + "angle": 0, + "content": "\\[\n\\angle \\left(\\mathbf {P} _ {i}, \\mathbf {P} _ {j}\\right) = \\angle \\left(\\mathbf {R} \\left(\\mathbf {P} _ {i} + \\mathbf {t}\\right), \\mathbf {R} \\left(\\mathbf {P} _ {j} + \\mathbf {t}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.24, + 0.785, + 0.39 + ], + "angle": 0, + "content": "where \\(\\mathbf{R}\\) is rotation matrix and \\(\\mathbf{t}\\) is translation vector. From the rotation invariance [39] that a pair of 3D points are jointly rotated about the origin, their angular distance remains constant. We have \\(\\angle (\\mathbf{R}(\\mathbf{P}_i + \\mathbf{t}),\\mathbf{R}(\\mathbf{P}_j + \\mathbf{t})) = \\angle (\\mathbf{P}_i + \\mathbf{t},\\mathbf{P}_j + \\mathbf{t})\\). Thus, the angle invariance is maintained when the calibration target is arbitrarily rotated about the origin. To make \\(\\angle (\\mathbf{P}_i,\\mathbf{P}_j) = \\angle (\\mathbf{P}_i + \\mathbf{t},\\mathbf{P}_j + \\mathbf{t})\\) hold for any pair of points on the target, the translation vector must be a \\(\\mathbf{0}\\) vector \\((\\mathbf{t} = [0,0,0]^T)\\). The proof is given in the supplementary material. Therefore, the calibration target performs pure rotation motion with respect to the origin (i.e. camera optical center). Such pure rotational motion is a typical spherical motion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.785, + 0.482 + ], + "angle": 0, + "content": "The spherical motion can be visualized in Fig. 1. In the spherical motion model, the calibration target exhibits only rotational motion with a fixed distance relative to the camera. We establish a local coordinate on the target, where the \\( z' \\)-axis aligns with the ray between the camera coordinate system \\( c \\) and calibration target coordinate system \\( p' \\). Inspired by [38], the relative pose between the calibration target and camera can be mathematically represented as" + }, + { + "type": "equation", + "bbox": [ + 0.427, + 0.49, + 0.786, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {c p ^ {\\prime}} = \\left( \\begin{array}{l l} \\mathbf {R} _ {c p ^ {\\prime}} & \\mathbf {t} _ {c p ^ {\\prime}} \\\\ \\mathbf {0} _ {1 \\times 3} & 1 \\end{array} \\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.529, + 0.784, + 0.605 + ], + "angle": 0, + "content": "where \\(\\mathbf{R}_{cp'} \\in SO(3)\\), \\(\\mathbf{t}_{cp'} = (0,0,-r)^T\\) and \\(r\\) is the radius of spherical motion. We usually establish the origin of the coordinate system \\(p\\) on the upper left corner of the calibration pattern. Therefore, there is a relative translation \\(\\mathbf{t}_{p'p} = (x,y,0)\\) between the coordinate systems \\(p'\\) and \\(p\\). Furthermore, the transformation matrix between camera \\(c\\) and calibration target \\(p\\) is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.29, + 0.614, + 0.786, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {p c} = \\left[ \\left( \\begin{array}{l l} \\mathbf {R} _ {c p ^ {\\prime}} & \\mathbf {t} _ {c p ^ {\\prime}} \\\\ \\mathbf {0} _ {1 \\times 3} & 1 \\end{array} \\right) \\left( \\begin{array}{l l} \\mathbf {I} _ {3 \\times 3} & \\mathbf {t} _ {p ^ {\\prime} p} \\\\ \\mathbf {0} _ {1 \\times 3} & 1 \\end{array} \\right) \\right] ^ {- 1} = \\left( \\begin{array}{l l} \\mathbf {R} _ {p c} & - \\mathbf {R} _ {p c} \\mathbf {t} _ {c p} \\\\ \\mathbf {0} _ {1 \\times 3} & 1 \\end{array} \\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.655, + 0.784, + 0.731 + ], + "angle": 0, + "content": "where \\(\\mathbf{R}_{pc} = \\mathbf{R}_{cp'}^T\\) and \\(\\mathbf{t}_{cp} = \\mathbf{t}_{p'p} + \\mathbf{t}_{cp'} = (x,y, -r)^T\\). \\(\\mathbf{t}_{cp}\\) determines the position of the camera in the calibration target coordinate system. Note that \\(\\mathbf{t}_{cp}\\) is a fixed value and has nothing to do with the pose of the calibration target. Given the constraint from the spherical motion, the original 6DOF general motion is limited to a 3DOF pure rotation motion." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.751, + 0.445, + 0.767 + ], + "angle": 0, + "content": "4 Calibration Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.84 + ], + "angle": 0, + "content": "The spherical motion constraint effectively reduces the number of motion parameters to be estimated. In this section, we show how to calibrate camera parameters using a set of collimator images. A closed-form solver for multiple views and a minimal solver for two views are proposed." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.432, + 0.129 + ], + "angle": 0, + "content": "S. Liang, B. Guan et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.451, + 0.16 + ], + "angle": 0, + "content": "4.1 Geometric Constraints" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.202 + ], + "angle": 0, + "content": "We use the well-known pinhole model to model cameras. The mapping from a 3D point \\(\\mathbf{P} = (X,Y,Z)^T\\) to an image point \\(\\mathbf{p} = (u,v)^T\\) is computed as" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.209, + 0.786, + 0.258 + ], + "angle": 0, + "content": "\\[\ns \\widetilde {\\mathbf {p}} = \\mathbf {K} [ \\mathbf {R} \\mid \\mathbf {t} ] \\widetilde {\\mathbf {P}}, \\text {w i t h} \\mathbf {K} = \\left[ \\begin{array}{l l l} f _ {x} & \\gamma & c _ {x} \\\\ 0 & f _ {y} & c _ {y} \\\\ 0 & 0 & 1 \\end{array} \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.266, + 0.784, + 0.343 + ], + "angle": 0, + "content": "where \\( s \\) is an arbitrary scale factor, \\( (\\mathbf{R},\\mathbf{t}) \\) is the camera pose with respect to the calibration target coordinate, and \\( \\mathbf{K} \\) is the camera intrinsic matrix. \\( \\widetilde{\\mathbf{p}} = [u,v,1]^T \\) and \\( \\widetilde{\\mathbf{P}} = [X,Y,Z,1]^T \\) denote the homogeneous coordinate forms of points \\( \\mathbf{p} \\) and \\( \\mathbf{P} \\), respectively. The intrinsic matrix \\( \\mathbf{K} \\) consists of five parameters: the focal length \\( (f_x,f_y) \\), the principal point \\( (c_x,c_y) \\) and skew factor \\( \\gamma \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.785, + 0.418 + ], + "angle": 0, + "content": "Without loss of generality, we assume the planar target is on the plane \\( Z = 0 \\) of the world coordinate system. The transformation between the image plane and calibration target plane can be constrained by a \\( 3 \\times 3 \\) homography matrix, denoted as \\( \\mathbf{H} \\). For perspective projection, the homography matrix \\( \\mathbf{H} \\) depends on the image pose \\( (\\mathbf{R},\\mathbf{t}) \\) and intrinsic matrix \\( \\mathbf{K} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.383, + 0.428, + 0.785, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} = \\left[ \\begin{array}{l l l} \\mathbf {h} _ {1} & \\mathbf {h} _ {2} & \\mathbf {h} _ {3} \\end{array} \\right] = \\lambda \\mathbf {K} \\left[ \\begin{array}{l l l} \\mathbf {r} _ {1} & \\mathbf {r} _ {2} & \\mathbf {t} \\end{array} \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.454, + 0.784, + 0.5 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is an arbitrary scalar, \\(\\mathbf{r}_i\\) is the \\(i\\)-th columns of \\(\\mathbf{R}\\). Given an image of the calibration target, the homography matrix \\(\\mathbf{H}\\) can be estimated using the plane point correspondences. Combining Eq. (3) and Eq. (5), we can get" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} = \\left[ \\mathbf {r} _ {1} \\mathbf {r} _ {2} - \\mathbf {R t} _ {c p} \\right] = \\frac {1}{\\lambda} \\mathbf {K} ^ {- 1} \\mathbf {H}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.543, + 0.663, + 0.559 + ], + "angle": 0, + "content": "Using the knowledge that \\(\\mathbf{R}\\) is an orthogonal matrix, we have" + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.568, + 0.785, + 0.616 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} ^ {T} \\mathbf {K} ^ {- T} \\mathbf {K} ^ {- 1} \\mathbf {H} = \\lambda^ {2} \\left[ \\begin{array}{c c c} 1 & 0 & - x \\\\ 0 & 1 & - y \\\\ - x & - y & \\| \\mathbf {t} _ {c p} \\| ^ {2} \\end{array} \\right]. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.624, + 0.784, + 0.7 + ], + "angle": 0, + "content": "There are 8 parameters to be estimated, 5 for intrinsic matrix and 3 for \\(\\mathbf{t}_{cp} = (x,y,-r)^T\\). For each given image, Eq. (7) provides 5 independent constraints. Theoretically, a minimum of two images can completely solve all unknown parameters. In the next two subsections, we give a closed-form solver for \\(N(N > 2)\\) collimator images and a minimal solver for two collimator images." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.72, + 0.58, + 0.736 + ], + "angle": 0, + "content": "4.2 Closed-form Solver for Multiple Views" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.743, + 0.784, + 0.804 + ], + "angle": 0, + "content": "When there are more than two collimator images for calibration, we model the problem as a linear system and propose a closed-form solution. Suppose that the homography matrices \\(\\mathbf{H}_i(i = 1,\\dots,N)\\) have been estimated for all images. The transformations between images \\(i\\) and \\(j\\) can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.423, + 0.813, + 0.785, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} _ {j} ^ {- 1} \\mathbf {H} _ {i} = \\frac {\\lambda_ {i}}{\\lambda_ {j}} \\mathbf {M} _ {j} ^ {- 1} \\mathbf {M} _ {i}. \\tag {8}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Camera Calibration using a Collimator System" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.717, + 0.162 + ], + "angle": 0, + "content": "For the matrix \\(\\mathbf{M}\\) of each image, we can calculate the determinant as" + }, + { + "type": "equation", + "bbox": [ + 0.315, + 0.173, + 0.785, + 0.21 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left| \\mathbf {M} \\right| = \\left| \\mathbf {r} _ {1} \\mathbf {r} _ {2} - \\mathbf {R t} _ {c p} \\right| = \\left| \\mathbf {r} _ {1} \\mathbf {r} _ {2} - x \\mathbf {r} _ {1} - y \\mathbf {r} _ {2} + r \\mathbf {r} _ {3} \\right| \\tag {9} \\\\ = - x \\left| \\mathbf {r} _ {1} \\mathbf {r} _ {2} \\mathbf {r} _ {1} \\right| - y \\left| \\mathbf {r} _ {1} \\mathbf {r} _ {2} \\mathbf {r} _ {2} \\right| + r \\left| \\mathbf {r} _ {1} \\mathbf {r} _ {2} \\mathbf {r} _ {3} \\right| = r, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.22, + 0.784, + 0.313 + ], + "angle": 0, + "content": "where \\(r\\) is a non-zero fixed value. Using the fact that \\(\\operatorname*{det}(\\mathbf{M}_j^{-1}\\mathbf{M}_i) = 1\\), we have \\(\\lambda_{ij} = \\lambda_i / \\lambda_j = \\sqrt[3]{|\\mathbf{H}_j^{-1}\\mathbf{H}_i|}\\). Then, we choose the image with the most feature points observed as the base image. For convenience, here the first image is selected as the base image and we get \\(\\lambda_{i1} = \\sqrt[3]{|\\mathbf{H}_j^{-1}\\mathbf{H}_1|}\\). By inverting Eq. (7) of image \\(i\\), we get" + }, + { + "type": "equation", + "bbox": [ + 0.338, + 0.322, + 0.784, + 0.37 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} _ {i} ^ {- 1} \\mathbf {K} \\mathbf {K} ^ {T} \\mathbf {H} _ {i} ^ {- T} = \\frac {1}{\\lambda_ {i} ^ {2} r ^ {2}} \\left[ \\begin{array}{c c c} r ^ {2} + x ^ {2} & x y & x \\\\ x y & r ^ {2} + y ^ {2} & y \\\\ x & y & 1 \\end{array} \\right]. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.379, + 0.528, + 0.396 + ], + "angle": 0, + "content": "Substitute \\(\\lambda_{i} = \\lambda_{i1}\\lambda_{1}\\) into Eq. (10), we get" + }, + { + "type": "equation", + "bbox": [ + 0.29, + 0.405, + 0.784, + 0.464 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} _ {i} ^ {- 1} \\mathbf {K K} ^ {T} \\mathbf {H} _ {i} ^ {- T} = \\frac {1}{\\lambda_ {i 1} ^ {2}} \\mathbf {A}, \\text {w i t h} \\mathbf {A} = \\left[ \\begin{array}{c c c} \\frac {r ^ {2} + x ^ {2}}{(\\lambda_ {1} r) ^ {2}} & \\frac {x y}{(\\lambda_ {1} r) ^ {2}} & \\frac {x}{(\\lambda_ {1} r) ^ {2}} \\\\ \\frac {x y}{(\\lambda_ {1} r) ^ {2}} & \\frac {r ^ {2} + y ^ {2}}{(\\lambda_ {1} r) ^ {2}} & \\frac {y}{(\\lambda_ {1} r) ^ {2}} \\\\ \\frac {x}{(\\lambda_ {1} r) ^ {2}} & \\frac {y}{(\\lambda_ {1} r) ^ {2}} & \\frac {1}{(\\lambda_ {1} r) ^ {2}} \\end{array} \\right]. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.473, + 0.571, + 0.487 + ], + "angle": 0, + "content": "Let us define a 6-vector \\(\\mathbf{a}\\) to represent matrix \\(\\mathbf{A}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.498, + 0.629, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\mathbf {a} = \\left(A _ {1 1}, A _ {1 2}, A _ {1 3}, A _ {2 2}, A _ {2 3}, A _ {3 3}\\right) ^ {T},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.526, + 0.727, + 0.54 + ], + "angle": 0, + "content": "where \\( A_{mn} \\) is the value of \\( m \\)-th row and \\( n \\)-th column of matrix \\( \\mathbf{A} \\). Let" + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.55, + 0.784, + 0.599 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} = \\left[ \\begin{array}{c c c} c _ {x} ^ {2} + f _ {x} ^ {2} + \\gamma^ {2} & c _ {x} c _ {y} + f _ {y} \\gamma & c _ {x} \\\\ c _ {x} c _ {y} + f _ {y} \\gamma & c _ {y} ^ {2} + f _ {y} ^ {2} & c _ {y} \\\\ c _ {x} & c _ {y} & 1 \\end{array} \\right]. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.608, + 0.72, + 0.623 + ], + "angle": 0, + "content": "Note that \\(\\mathbf{W}\\) is a symmetric matrix and can be defined by a 5-vector:" + }, + { + "type": "equation", + "bbox": [ + 0.38, + 0.633, + 0.623, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} = \\left(W _ {1 1}, W _ {1 2}, W _ {1 3}, W _ {2 2}, W _ {2 3}\\right) ^ {T},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.66, + 0.784, + 0.72 + ], + "angle": 0, + "content": "where \\( W_{mn} \\) is the value of \\( m \\)-th row and \\( n \\)-th column of matrix \\( \\mathbf{W} \\). Obviously, Eq. (11) represents a linear system of equations involving the unknowns \\( [\\mathbf{w}^T, \\mathbf{a}^T]^T \\). Combining Eqs. (11) and (12), we can get the following equations for each image:" + }, + { + "type": "equation", + "bbox": [ + 0.414, + 0.72, + 0.784, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\left[ \\mathbf {V} - \\lambda_ {i 1} ^ {2} \\mathbf {I} _ {5 \\times 5} \\right] \\left[ \\begin{array}{l} \\mathbf {w} \\\\ \\mathbf {a} \\end{array} \\right] = \\mathbf {b}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.756, + 0.252, + 0.77 + ], + "angle": 0, + "content": "with" + }, + { + "type": "equation", + "bbox": [ + 0.309, + 0.768, + 0.517, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\mathbf {V} = \\left[ \\mathbf {v} _ {1 1}, \\mathbf {v} _ {1 2}, \\mathbf {v} _ {1 3}, \\mathbf {v} _ {2 2}, \\mathbf {v} _ {2 3} \\right] ^ {T},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.309, + 0.789, + 0.691, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {m n} = \\left[ h _ {m 1} h _ {n 1}, h _ {m 1} h _ {n 2} + h _ {m 2} h _ {n 1}, h _ {m 1} h _ {n 3} + h _ {m 3} h _ {n 1}, \\right.\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.808, + 0.556, + 0.822 + ], + "angle": 0, + "content": "\\[\n\\left. h _ {m 2} h _ {n 2}, h _ {m 2} h _ {n 3} + h _ {m 3} h _ {n 2} \\right],\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.309, + 0.826, + 0.639, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathbf {b} = - \\left[ h _ {1 3} h _ {1 3}, h _ {1 3} h _ {2 3}, h _ {1 3} h _ {3 3}, h _ {2 3} h _ {2 3}, h _ {2 3} h _ {3 3} \\right] ^ {T},\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.434, + 0.129 + ], + "angle": 0, + "content": "S. Liang, B. Guan et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "where \\( h_{mn} \\) denotes the value of \\( m \\)-th row and \\( n \\)-th column of matrix \\( \\mathbf{H}^{-1} \\). Given \\( N(N > 2) \\) collimator images, we can stack \\( N \\) such equations as Eq. (13) together in matrix form:" + }, + { + "type": "equation", + "bbox": [ + 0.457, + 0.19, + 0.786, + 0.221 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} \\left[ \\begin{array}{l} \\mathbf {w} \\\\ \\mathbf {a} \\end{array} \\right] = \\mathbf {b}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.224, + 0.782, + 0.24 + ], + "angle": 0, + "content": "where \\(\\mathbf{D}\\) is a \\(5N\\times 11\\) matrix. The closed-form solution of Eq. (14) is given by" + }, + { + "type": "equation", + "bbox": [ + 0.421, + 0.248, + 0.786, + 0.281 + ], + "angle": 0, + "content": "\\[\n\\left[ \\begin{array}{l} \\mathbf {w} \\\\ \\mathbf {a} \\end{array} \\right] = \\left(\\mathbf {D} ^ {T} \\mathbf {D}\\right) ^ {- 1} \\mathbf {D} ^ {T} \\mathbf {b}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.287, + 0.787, + 0.318 + ], + "angle": 0, + "content": "Once \\(\\mathbf{w}\\) and \\(\\mathbf{a}\\) are estimated, we can calculate all camera intrinsic parameters and optical center \\(\\mathbf{t}_{cp}\\) by decomposing \\(\\mathbf{W}\\) and \\(\\mathbf{A}\\) respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.33, + 0.328, + 0.675, + 0.414 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} c _ {x} = W _ {1 3}, \\quad c _ {y} = W _ {2 3}, \\quad f _ {y} = \\sqrt {W _ {2 2} - c _ {y} ^ {2}}, \\\\ \\gamma = \\left(W _ {1 2} - c _ {x} c _ {y}\\right) / f _ {y}, \\quad f _ {x} = \\sqrt {W _ {1 1} - c _ {x} ^ {2} - \\gamma^ {2}}, \\\\ x = A _ {1 3} / A _ {3 3}, \\quad y = A _ {2 3} / A _ {3 3}, \\\\ r = - \\sqrt {\\left(A _ {1 1} / A _ {3 3}\\right) - \\left(A _ {1 3} / A _ {3 3}\\right) ^ {2}}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.421, + 0.784, + 0.437 + ], + "angle": 0, + "content": "Subsequently, the rotation matrix of each image can be calculated with Eq. (5):" + }, + { + "type": "equation", + "bbox": [ + 0.307, + 0.452, + 0.785, + 0.467 + ], + "angle": 0, + "content": "\\[\n\\mathbf {r} _ {1} = (1 / \\lambda) \\mathbf {K} ^ {- 1} \\mathbf {h} _ {1}, \\quad \\mathbf {r} _ {2} = (1 / \\lambda) \\mathbf {K} ^ {- 1} \\mathbf {h} _ {2}, \\quad \\mathbf {r} _ {3} = \\mathbf {r} _ {1} \\times \\mathbf {r} _ {2}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.471, + 0.786, + 0.503 + ], + "angle": 0, + "content": "where \\(\\lambda = (\\| \\mathbf{K}^{-1}\\mathbf{h}_1\\| +\\| \\mathbf{K}^{-1}\\mathbf{h}_2\\|) / 2\\). Finally, a singular value decomposition for re-orthogonalizing the rotation matrix \\(\\mathbf{R} = [\\mathbf{r}_1,\\mathbf{r}_2,\\mathbf{r}_3]\\) is required." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.522, + 0.513, + 0.536 + ], + "angle": 0, + "content": "4.3 Minimal Solver for Two Views" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.544, + 0.787, + 0.62 + ], + "angle": 0, + "content": "When only two collimator images are used for calibration, the rank of the coefficient matrix in Eq. (14) is determined to be 10. Solving rank-deficient equations presents a complex challenge. To address this issue, we construct a nonlinear equation system and propose a novel approach that utilizes the hidden variable technique [14] to estimate the unknowns in the minimum configuration." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.62, + 0.787, + 0.666 + ], + "angle": 0, + "content": "Estimation of \\(\\mathbf{W} = \\mathbf{K}^{-T}\\mathbf{K}^{-1}\\) is the core of calibration. Note that \\(\\mathbf{W}\\) is a symmetric matrix, defined by a 6-vector \\(\\mathbf{w} = (W_{11}, W_{12}, W_{13}, W_{22}, W_{23}, W_{33})^T\\). We can extract the following 5 constraints from Eq. (7)" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.673, + 0.785, + 0.689 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {1 2} ^ {T} \\mathbf {w} = 0 \\tag {17}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.462, + 0.694, + 0.785, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\left(\\mathbf {v} _ {1 1} ^ {T} - \\mathbf {v} _ {2 2} ^ {T}\\right) \\mathbf {w} = 0 \\tag {18}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.45, + 0.714, + 0.785, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {1 3} ^ {T} \\mathbf {w} + x \\mathbf {v} _ {1 1} ^ {T} \\mathbf {w} = 0 \\tag {19}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.451, + 0.735, + 0.785, + 0.752 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {2 3} ^ {T} \\mathbf {w} + y \\mathbf {v} _ {1 1} ^ {T} \\mathbf {w} = 0 \\tag {20}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.757, + 0.785, + 0.773 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {3 3} ^ {T} \\mathbf {w} - \\| \\mathbf {t} _ {c p} \\| ^ {2} \\mathbf {v} _ {1 1} ^ {T} \\mathbf {w} = 0. \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.782, + 0.474, + 0.797 + ], + "angle": 0, + "content": "\\(\\mathbf{v}_{mn}\\) can be calculated as following:" + }, + { + "type": "equation", + "bbox": [ + 0.315, + 0.806, + 0.688, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {m n} = \\left[ h _ {1 m} h _ {1 n}, h _ {1 m} h _ {2 n} + h _ {1 n} h _ {2 m}, h _ {1 m} h _ {3 n} + h _ {1 n} h _ {3 m} \\right.\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.825, + 0.636, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\left. h _ {2 m} h _ {2 n}, h _ {2 m} h _ {3 n} + h _ {2 n} h _ {3 m}, h _ {3 m} h _ {3 n} \\right] ^ {T},\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Camera Calibration using a Collimator System" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "where \\( h_{mn} \\) denotes the value of \\( m \\)-th row and \\( n \\)-th column of matrix \\( \\mathbf{H} \\). Given only 2 collimator images, we can obtain 10 independent equations for 8 unknowns. The hidden variable technique [14] is employed to solve the above polynomial equations. Combining Eqs. (19) to (21), we have" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.215, + 0.786, + 0.233 + ], + "angle": 0, + "content": "\\[\n\\left(\\mathbf {v} _ {1 3} ^ {T} + \\mathbf {v} _ {2 3} ^ {T} + \\mathbf {v} _ {3 3} ^ {T}\\right) \\mathbf {w} + c \\mathbf {v} _ {1 1} ^ {T} \\mathbf {w} = 0, \\tag {22}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.24, + 0.784, + 0.287 + ], + "angle": 0, + "content": "where \\( c = (x + y - \\| \\mathbf{t}_{cp} \\|^2) \\). We first select constraints Eqs. (17), (18) and (22) to establish a new polynomial equation system. Next, we treat \\( c \\) as a hidden variable, and the complete set of equations can be expressed in matrix form as" + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.294, + 0.786, + 0.342 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} (c) \\mathbf {w} = 0, \\text {w i t h} \\mathbf {C} (c) = \\left[ \\begin{array}{c} \\mathbf {v} _ {1 2} ^ {T} \\\\ (\\mathbf {v} _ {1 1} ^ {T} - \\mathbf {v} _ {2 2} ^ {T}) \\\\ (\\mathbf {v} _ {1 3} ^ {T} + \\mathbf {v} _ {2 3} ^ {T} + \\mathbf {v} _ {3 3} ^ {T}) + c \\mathbf {v} _ {1 1} ^ {T} \\end{array} \\right], \\tag {23}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.349, + 0.785, + 0.501 + ], + "angle": 0, + "content": "where the coefficient matrix \\(\\mathbf{C}(c)\\) contains the hidden variable \\(c\\). Two images provide 6 equations for solving \\(c\\), and \\(\\mathbf{C}(c)\\) is a \\(6 \\times 6\\) matrix. Since the equation system \\(\\mathbf{C}(c)\\mathbf{w} = 0\\) must have at least one nontrivial solution for \\(\\mathbf{w}\\), the determinant of \\(\\mathbf{C}(c)\\) is equal to zero. The equation, \\(\\operatorname*{det}(\\mathbf{C}(c)) = 0\\), is a polynomial of degree 2 in terms of \\(c\\), which can be efficiently solved using a straightforward root-finding algorithm. Substituting the known value of \\(c\\) into the constraint equations Eqs. (17), (18) and (22), we obtain six constraint equations about \\(\\mathbf{w}\\). The least squares method is then employed to solve these linear equations and compute the optimal \\(\\mathbf{w}\\) up to scale. Once \\(\\mathbf{w}\\) is estimated, the intrinsic matrix \\(\\mathbf{K}\\) can be calculated using the Cholesky factorization [15]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.519, + 0.423, + 0.535 + ], + "angle": 0, + "content": "4.4 Bundle Adjustment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.541, + 0.784, + 0.646 + ], + "angle": 0, + "content": "Bundle adjustment jointly refines the camera intrinsic \\(\\mathbf{K}\\), lens distortion coefficients \\(\\mathbf{d}\\), the rotation matrix of the \\(i\\)-th image \\(\\mathbf{R}_i\\) and camera optical center \\(\\mathbf{t}_{cp}\\). In cases of moderate lens distortion, it is typical to initialize the distortion coefficient to \\(\\mathbf{0}\\) and estimate them in the bundle adjustment step [36]. Given a set of \\(N\\) images, each with \\(M\\) corresponding 2D image points, we establish a cost function aimed at minimizing the re-projection error. The cost function is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.655, + 0.786, + 0.697 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {X}} \\sum_ {i} ^ {N} \\sum_ {j} ^ {M} \\rho \\left(\\left\\| \\pi \\left(\\mathbf {K}, \\mathbf {d}, \\mathbf {R} _ {i}, \\mathbf {t} _ {c p}, \\mathbf {P} _ {i j}\\right) - \\mathbf {p} _ {i j} \\right\\| ^ {2}\\right), \\tag {24}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.784, + 0.78 + ], + "angle": 0, + "content": "where \\(\\mathbf{X} = (\\mathbf{K},\\mathbf{d},\\mathbf{R}_i,\\mathbf{t}_{cp})\\) is the parameter to be refined. \\(\\mathbf{P}_{ij}\\) is the \\(j\\)-th 3D pattern point of \\(i\\)-th image and corresponds to the 2D image point \\(\\mathbf{p}_{ij}\\). \\(\\pi (\\cdot)\\) projects the world points to image according to Eq. (4). \\(\\rho (\\cdot)\\) is the Cauchy cost function. As is common, we optimize Eq. (24) with the Levenberg-Marquardt (LM) method [27], which is implemented in some excellent frameworks [1]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.781, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Note that the camera optical center is fixed at \\(\\mathbf{t}_{cp}\\) with the spherical motion constraint. However, for general motion, the camera optical center is free. Therefore, when calibrating the camera using collimator images, the number of motion parameters to be refined is reduced from \\(6N\\) to \\(3N + 3\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.432, + 0.128 + ], + "angle": 0, + "content": "S. Liang, B. Guan et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.486, + 0.163 + ], + "angle": 0, + "content": "4.5 Degenerated Configuration" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.786, + 0.203 + ], + "angle": 0, + "content": "We prove a degenerate configuration where additional images cannot provide more constraints about the camera intrinsic parameter." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.204, + 0.787, + 0.248 + ], + "angle": 0, + "content": "Proposition 1. If the calibration target rotates only around an axis that is parallel to the \\(z\\) axis and intersects the target plane at \\((x,y,0)^T\\), the newly added image cannot provide additional constraints on the camera intrinsic parameters." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.25, + 0.786, + 0.264 + ], + "angle": 0, + "content": "Due to space limitations, the proof is provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.289, + 0.478, + 0.305 + ], + "angle": 0, + "content": "5 Evaluation and Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.322, + 0.786, + 0.368 + ], + "angle": 0, + "content": "In this section, we first evaluate the robustness and accuracy of the proposed algorithm on synthetic data. Then, we calibrate the camera and evaluate the effectiveness of the proposed method using real collimator images." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.391, + 0.498, + 0.407 + ], + "angle": 0, + "content": "5.1 Synthetic Data Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.786, + 0.477 + ], + "angle": 0, + "content": "In the synthetic data experiment, we evaluate the performance of the proposed algorithm. We simulate the process of calibrating predefined camera parameters using synthetic images that conform to the spherical motion model. The proposed is compared with the following algorithms:" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.478, + 0.781, + 0.492 + ], + "angle": 0, + "content": "- Zhang [46] is a plane-based algorithm and suitable for general 6DOF motion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.784, + 0.536 + ], + "angle": 0, + "content": "- Bouguet [3] is a calibration toolbox and suitable for general 6DOF motion. Unlike Zhang, Bouguet computes the focal length using the orthogonal vanishing points constraint." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.784, + 0.566 + ], + "angle": 0, + "content": "- Hartley [17] is a self-calibration method based on a pure rotating camera, which is specifically suitable for the 3DOF spherical motion model." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.478, + 0.784, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.795, + 0.704 + ], + "angle": 0, + "content": "We define a virtual camera with the following parameters: \\( f_{x} = f_{y} = 1000 \\), \\( c_{x} = 542 \\), \\( c_{y} = 478 \\), \\( \\gamma = 0.01 \\), radial distortion coefficients \\( d_{1} = 0.1 \\), \\( d_{2} = -0.2 \\), and the image size is set to \\( 1080 \\times 960 \\) pixels. The calibration target is a planar pattern containing \\( 11 \\times 8 \\) points, each square size of \\( 30mm \\times 30mm \\). The target maintains a spherical motion with a radius of \\( 700mm \\) relative to the camera. The camera optical center in the target coordinate system is \\( t_{cp} = [150, 105, -700]^{T}mm \\), and the camera's orientation is arbitrary. Given the camera parameters and poses, the planar target points are projected onto images. Finally, zero-mean Gaussian noise with \\( \\sigma \\) standard deviation is added to 2D image points." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Evaluation of Calibration. In this experiment, we seek to evaluate the performance of the algorithms in solving camera parameters. For the four algorithms, we perform two steps of solving initial value and bundle adjustment to obtain camera parameters. Firstly, we compare the performance of the mentioned algorithms as the noise level increases. The noise level varies from 0 to 3.0 pixels, and we use 15 images to calibrate the camera. For each noise level, we conducted 200 trials and calculated the average errors. As shown in Fig. 3, calibration error increases linearly with the noise level. Compared with the other three algorithms, ours shows greater robustness to noise. Additionally, at a noise" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Camera Calibration using a Collimator System" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.144, + 0.36, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.144, + 0.498, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.143, + 0.64, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.143, + 0.785, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.256, + 0.725, + 0.27 + ], + "angle": 0, + "content": "Fig. 3: Comparison of calibration errors with increasing noise level." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.305, + 0.36, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.306, + 0.498, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.305, + 0.641, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.305, + 0.785, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.416, + 0.759, + 0.43 + ], + "angle": 0, + "content": "Fig. 4: Comparison of calibration errors with the number of images increases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.788, + 0.619 + ], + "angle": 0, + "content": "level of 0.5 pixels, we evaluate the performance with respect to the number of calibration images. The image count ranged from 3 to 30, with 200 trials conducted for each quantity. The results are illustrated in Fig. 4. A significant decrease in error is observed when increasing the number of images from 3 to 5. Both Fig. 3 and Fig. 4 show that our algorithm outperforms other algorithms in robustness and accuracy. We believe that our algorithm benefits from the consideration of both spherical motion and planar homography constraints, along with reducing optimization variables. Due to space limitations, the additional comparison results of the initial value estimation are shown in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Sensitivity of imperfect spherical motion. The low-cost collimators exhibit an inferior parallelism, which is reflected in the fact that the relative motion between the target and camera is not a perfect spherical motion. This experiment evaluates the sensitivity of initial value estimation with respect to imperfect spherical motion. We add zeros-mean Gaussian noise with standard deviation from \\(0mm\\) to \\(30mm\\) to \\(\\mathbf{t}_{cp}\\) to replicate this imperfection. We use 15 images and add a noise level of 0.5 pixels to the image points. For each noise level, a total of independent 500 trials are conducted and the average errors are computed. The sensitivity analysis results are shown in Fig. 5. Noise has no impact on Bouguet and Zhang, because they do not rely on the spherical motion constraint. As the noise increases, calibration errors of Ours and Hartley increase gradually. Notably, Hartley is significantly influenced by noise, whereas Ours displays superior robustness. Even with noise levels below \\(15mm\\) (equivalent to \\(2.143\\%\\) of the spherical radius), our algorithm remains superior to the alternatives." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.432, + 0.129 + ], + "angle": 0, + "content": "S. Liang, B. Guan et al." + }, + { + "type": "image", + "bbox": [ + 0.317, + 0.144, + 0.495, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.143, + 0.685, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.277, + 0.694, + 0.292 + ], + "angle": 0, + "content": "Fig. 5: Sensitivity analysis of imperfect spherical motion." + }, + { + "type": "table_caption", + "bbox": [ + 0.24, + 0.306, + 0.761, + 0.322 + ], + "angle": 0, + "content": "Table 1: Comparison of calibration and pose estimation error results (pixels)." + }, + { + "type": "table", + "bbox": [ + 0.304, + 0.332, + 0.697, + 0.422 + ], + "angle": 0, + "content": "
NumberZhang [46]Bouguet [3]Hartley [17]Ours
εcalibεevalεcalibεevalεcalibεevalεcalibεeval
20.18420.15450.17390.1518--0.17390.1519
50.23160.12450.22020.1236-0.25770.22030.1236
100.19230.12310.19120.1232-0.19860.19510.1222
150.17750.12210.17180.1232-0.17910.17790.1216
200.17730.12250.17220.1221-0.18710.17780.1207
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.451, + 0.475, + 0.466 + ], + "angle": 0, + "content": "5.2 Real Images Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.786, + 0.507 + ], + "angle": 0, + "content": "In real image experiments, we verify the effectiveness of our collimator system for calibration and evaluate the performance of the proposed algorithm." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.787, + 0.75 + ], + "angle": 0, + "content": "Calibration and Accuracy Evaluation. Following the experimental methodology utilized in previous studies [13,30], we design a pose estimation experiment to evaluate the accuracy of camera calibration quantitatively. A 8mm camera captures images from our collimator system and a practical printed pattern. Calibration is performed using the collimator images and a subset of the printed pattern images, while the remaining printed pattern images are used to evaluate the calibration results. After obtaining the camera parameters from the calibration images, the pose of each evaluation image can be determined using absolute pose estimation [24]. In our setup, a quarter of the corner points on the printed pattern are used for pose estimation, and the remaining points are used for evaluation. By re-projecting the evaluation corner points onto the image with the camera parameters and the calculated pose, we can evaluate the calibration quality based on the re-projection error. A total of 200 printed pattern images are used to evaluate the calibration. The calibration uses 2, 5, 10, 15, and 20 images. The re-projection error of the calibration and evaluation images are donated as \\(\\epsilon_{calib}\\) and \\(\\epsilon_{eval}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The comparison results of the four algorithms are shown in Tab. 1. Due to the limited degrees of freedom, our calibrated re-projection error \\((\\epsilon_{calib})\\) is slightly higher. However, Ours exhibits the lowest re-projection error in evaluation images compared to alternative methods, indicating that our method achieves more precise camera parameters. The Hartley method only uses the correspondence of 2D points in images, which makes it impossible to calculate the re-projection" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Camera Calibration using a Collimator System" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.244, + 0.145, + 0.757, + 0.16 + ], + "angle": 0, + "content": "Table 2: Comparison of calibration results using 20 images in a real camera." + }, + { + "type": "table", + "bbox": [ + 0.317, + 0.171, + 0.681, + 0.299 + ], + "angle": 0, + "content": "
PrintedCollimator
Bouguet [3]Zhang [46]Bouguet [3]Hartley [17]Ours
fx2384.52380.82369.22644.62388.3
fy2386.32379.82368.92646.02387.9
cx1209.81222.11221.11214.61221.0
cy1008.21008.41009.81012.71009.6
k1-0.0843-0.0899-0.0908-0.0754-0.0900
k20.08190.09120.08920.11810.0954
εcalib0.15130.17730.1722-0.1778
εeval0.12250.12250.12210.18710.1207
" + }, + { + "type": "table_caption", + "bbox": [ + 0.276, + 0.314, + 0.726, + 0.33 + ], + "angle": 0, + "content": "Table 3: Comparison of epipolar rectification error results (pixels)." + }, + { + "type": "table", + "bbox": [ + 0.304, + 0.34, + 0.7, + 0.43 + ], + "angle": 0, + "content": "
NumberZhang [46]Bouguet [3]Hartley [17]Ours
meanstdmeanstdmeanstdmeanstd
20.08460.04060.07700.0220--0.07700.0229
50.08090.03070.07640.02140.08320.03970.07620.0211
100.07730.02340.07580.02130.08540.05090.07570.0208
150.07820.02660.07710.02530.08370.05220.07560.0207
200.07720.02520.07680.02530.08180.03650.07560.0208
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.787, + 0.505 + ], + "angle": 0, + "content": "error of the calibration images directly. In cases with only two images, Zhang and Bouguet commonly depend on beneficial assumptions, such as \\(\\gamma = 0\\). Ours achieves comparable evaluation errors without the need for such assumptions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.787, + 0.627 + ], + "angle": 0, + "content": "Additionally, the Bouguet toolbox is also used to calculate camera parameters from printed pattern images that are not included in the evaluation. We use 20 images for calibration and 200 for evaluation. The estimated camera parameters and error comparison are shown in Tab. 2. The second column is a good reference. The lower evaluation error indicates that our method outperforms the alternative and validates the effectiveness of using the collimator image for calibration. It can be expected that the calibration parameters are not significantly different. Therefore, even a minor improvement holds considerable significance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.787, + 0.764 + ], + "angle": 0, + "content": "Then, we evaluate the performance of all methods using epipolar rectification error, which has not been optimized. Given the camera parameters from different methods, we calculate the epipolar rectification error (in pixels) for 300 pairs of printed evaluation images. The mean and standard deviation of the errors are shown in Tab. 3. Moreover, the result using Bouguet with 20 printed pattern images is (0.0772, 0.0259) pixels. It is observed that as the number of calibration images increases, the error stabilizes. The results indicate improved accuracy in the collimator image and highlight the beneficial impact of the spherical constraint on accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Multiple Cameras Calibration with a Same Collimator. In this experiment, we evaluate our method on multiple cameras with the same collimator. We used the cameras with \\(12mm\\), \\(16mm\\), \\(25mm\\), and \\(35mm\\) focal lengths of lenses. The calibration images are captured from the same collimator, as shown in Fig. 6 and left of Fig. 1. As the focal length increases, the more pixels a single" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.432, + 0.129 + ], + "angle": 0, + "content": "S. Liang, B. Guan et al." + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.144, + 0.645, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.238, + 0.712, + 0.252 + ], + "angle": 0, + "content": "Fig. 6: Calibration image sample obtained from the collimator." + }, + { + "type": "table_caption", + "bbox": [ + 0.22, + 0.268, + 0.781, + 0.283 + ], + "angle": 0, + "content": "Table 4: Comparison of calibration results with different focal length lens (pixels)." + }, + { + "type": "table", + "bbox": [ + 0.329, + 0.294, + 0.67, + 0.368 + ], + "angle": 0, + "content": "
Zhang [46]Bouguet [3]Hartley [17]Ours
12mm0.22730.22250.34040.2174
16mm0.29940.25910.41820.2536
25mm0.31330.30780.41950.2923
35mm0.32890.33480.72720.3118
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.784, + 0.429 + ], + "angle": 0, + "content": "square occupies. Cameras with different focal lengths can observe the calibration target through our collimator system in a limited space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.785, + 0.581 + ], + "angle": 0, + "content": "Similar to the evaluation approach in the previous experiment, we segregate all collimator images into calibration and evaluation sets. For each lens, we use 20 images for calibration and 200 for evaluation. The results are listed in Tab. 4. The proposed algorithm demonstrates the lowest re-projection error for lenses with varying focal lengths among the evaluation images. The re-projection error tends to increase with the increase of focal length. We believe that the main reason is that the longer the lens's focal length, the fewer points observed. More importantly, this experiment confirms the effectiveness of our collimator system and the proposed method. Moreover, we provide a spherical motion verification experiment in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.608, + 0.358, + 0.623 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.785, + 0.793 + ], + "angle": 0, + "content": "This paper presents a novel camera calibration method using a designed collimator system. Based on the optical geometry of the collimator system, we prove that the relative motion between the target and the camera conforms to a spherical motion. Furthermore, we propose a closed-form solver for multiple views and a minimal solver for two views for camera calibration. Our collimator-based method effectively combines the advantages of close-range calibration using a collimator and flexibility using a planar calibration target. We validate the effectiveness of using only collimator images to calibrate the camera. Synthetic and real data experiments demonstrate that the calibration accuracy of the proposed method is superior to the state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Limitations: When calibrating a fisheye camera, the distortion should be paid more attention rather than simply setting it to 0. Future work will focus on cameras with significant distortion." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Camera Calibration using a Collimator System" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.24 + ], + "angle": 0, + "content": "This research has been supported by the Hunan Provincial Natural Science Foundation for Excellent Young Scholars under Grant 2023JJ20045, and the National Natural Science Foundation of China under Grant 12372189. We thank Zijie Wang for useful discussion on the writing of the paper." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.264, + 0.324, + 0.279 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.784, + 0.323 + ], + "angle": 0, + "content": "1. Agarwal, S., Mierle, K., Team, T.C.S.: Ceres Solver. https://github.com/ceres-solver/ceres-solver" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.326, + 0.785, + 0.366 + ], + "angle": 0, + "content": "2. Bauer, M., Grießbach, D., Hermerschmidt, A., Krüger, S., Scheele, M., Schischmanow, A.: Geometrical camera calibration with diffractive optical elements. Optics express 16(25), 20241-20248 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.368, + 0.785, + 0.394 + ], + "angle": 0, + "content": "3. Bouguet, J.Y.: Camera calibration toolbox for matlab. http://robots.stanford.edu/cs223b04/JeanYvesCalib/index.html (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.395, + 0.785, + 0.422 + ], + "angle": 0, + "content": "4. Bradski, G.: The opencv library. Dr. Dobb's Journal: Software Tools for the Professional Programmer 25(11), 120-123 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.423, + 0.785, + 0.449 + ], + "angle": 0, + "content": "5. Brown, D.C.: Close-range camera calibration. Photogramm. Eng 37(8), 855-866 (1971)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.451, + 0.785, + 0.491 + ], + "angle": 0, + "content": "6. Campos, C., Elvira, R., Rodríguez, J.J.G., M. Montiel, J.M., D. Tardós, J.: Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics 37(6), 1874-1890 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.493, + 0.785, + 0.52 + ], + "angle": 0, + "content": "7. Clarke, T.A., Fryer, J.G.: The development of camera calibration methods and models. The Photogrammetric Record 16(91), 51-66 (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.521, + 0.785, + 0.561 + ], + "angle": 0, + "content": "8. Faugeras, O.D., Luong, Q.T., Maybank, S.J.: Camera self-calibration: Theory and experiments. In: IEEE European Conference on Computer Vision (ECCV). p. 321-334 (1992)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.563, + 0.785, + 0.603 + ], + "angle": 0, + "content": "9. Fuhr, G., Jung, C.R.: Camera self-calibration based on nonlinear optimization and applications in surveillance systems. IEEE Transactions on Circuits and Systems for Video Technology 27(5), 1132-1142 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.604, + 0.785, + 0.644 + ], + "angle": 0, + "content": "0. Guan, B., Su, Z., Yu, Q., Li, Z., Feng, W., Yang, D., Zhang, D.: Monitoring the blades of a wind turbine by using videogrammetry. Optics and Lasers in Engineering 152, 106901 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.646, + 0.785, + 0.686 + ], + "angle": 0, + "content": "1. Guan, B., Zhao, J., Barath, D., Fraundorfer, F.: Minimal solvers for relative pose estimation of multi-camera systems using affine correspondences. International Journal of Computer Vision 131(1), 324-345 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.688, + 0.785, + 0.728 + ], + "angle": 0, + "content": "2. Guan, B., Zhao, J., Li, Z., Sun, F., Fraundorfer, F.: Relative pose estimation with a single affine correspondence. IEEE Transactions on Cybernetics 52(10), 10111-10122 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.785, + 0.77 + ], + "angle": 0, + "content": "3. Ha, H., Perdoch, M., Alismail, H., Kweon, I.S., Sheikh, Y.: Deltille grids for geometric camera calibration. In: IEEE International Conference on Computer Vision (ICCV). pp. 5354-5362 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "4. Hartley, R., Li, H.: An efficient hidden variable approach to minimal-case camera motion estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence 34(12), 2303-2314 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Hartley, R., Zisserman, A.: Multiple View Geometry in Computer Vision. Cambridge University Press, 2 edn. (2004)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.296, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.432, + 0.129 + ], + "angle": 0, + "content": "S. Liang, B. Guan et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Hartley, R.I.: Self-calibration from multiple views with a rotating camera. In: European Conference on Computer Vision (ECCV). pp. 471-478 (1994)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "17. Hartley, R.I.: Self-calibration of stationary cameras. International Journal of Computer Vision 22(1), 5-23 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.204, + 0.785, + 0.245 + ], + "angle": 0, + "content": "18. Heikkila, J.: Geometric camera calibration using circular control points. IEEE Transactions on Pattern Analysis and Machine Intelligence 22(10), 1066-1077 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "19. Herrera, D., Kannala, C.J., Heikkila, J.: Forget the checkerboard: Practical self-calibration using a planar scene. In: IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 1-9. IEEE (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "20. Hieronymus, J.: Comparison of methods for geometric camera calibration. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences 39, 595-599 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "21. Hu, H., Wei, B., Mei, S., Liang, J., Zhang, Y.: A two-step calibration method for vision measurement with large field of view. IEEE Transactions on Instrumentation and Measurement 71, 1-10 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "22. Kannala, J., Brandt, S.S.: A generic camera model and calibration method for conventional, wide-angle, and fish-eye lenses. IEEE Transactions on Pattern Analysis and Machine Intelligence 28(8), 1335-1340 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "23. Klaus, A., Bauer, J., Karner, K., Elbischger, P., Perko, R., Bischof, H.: Camera calibration from a single night sky image. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). vol. 1, pp. I-I (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.453, + 0.785, + 0.494 + ], + "angle": 0, + "content": "24. Kneip, L., Furgale, P.: Opengv: A unified and generalized approach to real-time calibrated geometric vision. In: IEEE International Conference on Robotics and Automation (ICRA). pp. 1-8 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.495, + 0.785, + 0.535 + ], + "angle": 0, + "content": "25. Liu, Z., Li, F., Li, X., Zhang, G.: A novel and accurate calibration method for cameras with large field of view using combined small targets. Measurement 64, 1-16 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "26. Lochman, Y., Liepieshov, K., Chen, J., Perdoch, M., Zach, C., Pritts, J.: Babel-calib: A universal approach to calibrating central cameras. In: IEEE International Conference on Computer Vision (ICCV). pp. 15253-15262 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.578, + 0.785, + 0.618 + ], + "angle": 0, + "content": "27. Moré, J.J.: The levenberg-marquardt algorithm: implementation and theory. In: Numerical analysis: proceedings of the biennial Conference held at Dundee, June 28-July 1, 1977. pp. 105-116 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.62, + 0.785, + 0.659 + ], + "angle": 0, + "content": "28. Nister, D., Stewenius, H., Grossmann, E.: Non-parametric self-calibration. In: IEEE International Conference on Computer Vision (ICCV). vol. 1, pp. 120-127 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.661, + 0.785, + 0.688 + ], + "angle": 0, + "content": "29. Oniga, V.E., Pfeifer, N., Loghin, A.M.: 3d calibration test-field for digital cameras mounted on unmanned aerial systems (uas). Remote Sensing 10(12), 2017 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.689, + 0.785, + 0.729 + ], + "angle": 0, + "content": "30. Peng, S., Sturm, P.: Calibration wizard: A guidance system for camera calibration based on modelling geometric and corner uncertainty. In: IEEE International Conference on Computer Vision (ICCV). pp. 1497-1505 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.73, + 0.785, + 0.77 + ], + "angle": 0, + "content": "31. Pollefeys, M., Koch, R., Gool, L.V.: Self-calibration and metric reconstruction in spite of varying and unknown intrinsic camera parameters. International journal of computer vision 32(1), 7-25 (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "32. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4104-4113 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "33. Schöps, T., Larsson, V., Pollefeys, M., Sattler, T.: Why having 10,000 parameters in your camera model is better than twelve. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2532-2541 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Camera Calibration using a Collimator System" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "34. Shang, Y., Sun, X., Yang, X., Wang, X., Yu, Q.: A camera calibration method for large field optical measurement. Optik - International Journal for Light and Electron Optics 124, 6553-6558 (12 2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "35. Sturm, P.F., Maybank, S.J.: On plane-based camera calibration: A general algorithm, singularities, applications. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). vol. 1, pp. 432-437. IEEE (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.273 + ], + "angle": 0, + "content": "36. Triggs, B., McLauchlan, P.F., Hartley, R.I., Fitzgibbon, A.W.: Bundle adjustment—a modern synthesis. In: Vision Algorithms: Theory and Practice: International Workshop on Vision Algorithms. pp. 298-372. Springer (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.314 + ], + "angle": 0, + "content": "37. Tsai, R.: A versatile camera calibration technique for high-accuracy 3d machine vision metrology using off-the-shelf tv cameras and lenses. IEEE Journal on Robotics and Automation 3(4), 323-344 (1987)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.315, + 0.788, + 0.342 + ], + "angle": 0, + "content": "38. Ventura, J.: Structure from motion on a sphere. In: European Conference on Computer Vision (ECCV). vol. 9907, pp. 53-68 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.343, + 0.788, + 0.397 + ], + "angle": 0, + "content": "39. Wang, C., Liu, Y., Wang, Y., Li, X., Wang, M.: Efficient and outlier-robust simultaneous pose and correspondence determination by branch-and-bound and transformation decomposition. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(10), 6924-6938 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.438 + ], + "angle": 0, + "content": "40. Wang, Z., Wu, Z., Zhen, X., Yang, R., Xi, J., Chen, X.: A two-step calibration method of a largefov binocular stereovision sensor for onsite measurement. Measurement 62, 15-24 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.481 + ], + "angle": 0, + "content": "41. Wu, G., Han, B., He, X.: Calibration of geometric parameters of line array ccd camera based on exact measuring angle in lab. Optics and Precision Engineering 15(10), 1628-1632 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.522 + ], + "angle": 0, + "content": "42. Xiao, Z., Jin, L., Yu, D., Tang, Z.: A cross-target-based accurate calibration method of binocular stereo systems with large-scale field-of-view. Measurement 43(6), 747-754 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.563 + ], + "angle": 0, + "content": "43. Yuan, G., Zheng, L., Ding, Y., Zhang, H., Zhang, X., Liu, X., Sun, J.: A precise calibration method for line scan cameras. IEEE Transactions on Instrumentation and Measurement 70, 1-9 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.605 + ], + "angle": 0, + "content": "44. Yuan, G., Zheng, L., Sun, J., Liu, X., Wang, X., Zhang, Z.: Practical calibration method for aerial mapping camera based on multiple pinhole collimator. IEEE Access 8, 39725-39733 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.606, + 0.788, + 0.646 + ], + "angle": 0, + "content": "45. Zhang, K., Xie, J., Snavely, N., Chen, Q.: Depth sensing beyond lidar range. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1689-1697 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.647, + 0.788, + 0.674 + ], + "angle": 0, + "content": "46. Zhang, Z.: A flexible new technique for camera calibration. IEEE Transactions on Pattern Analysis and Machine Intelligence 22(11), 1330-1334 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.703 + ], + "angle": 0, + "content": "47. Zhang, Z.: Camera calibration with one-dimensional objects. IEEE Transactions on Pattern Analysis and Machine Intelligence 26(7), 892-899 (2004)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.703 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Camera Calibration using a Collimator System/26f230b2-04ba-4e24-b664-739a6c3feb85_origin.pdf b/2024/Camera Calibration using a Collimator System/26f230b2-04ba-4e24-b664-739a6c3feb85_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2300723954a545a413d4892463c02d45cc670ff2 --- /dev/null +++ b/2024/Camera Calibration using a Collimator System/26f230b2-04ba-4e24-b664-739a6c3feb85_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59d5817e76a5f54e6befaa264093b8828aaa20d26b016062dfaf53f79765a6ef +size 3513291 diff --git a/2024/Camera Calibration using a Collimator System/full.md b/2024/Camera Calibration using a Collimator System/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0683cdb718d0dd666fec4eeccfb7daada1845f5e --- /dev/null +++ b/2024/Camera Calibration using a Collimator System/full.md @@ -0,0 +1,423 @@ +# Camera Calibration using a Collimator System + +Shunkun Liang, Banglei Guan, Zhenbao Yu, Pengju Sun, and Yang Shang + +College of Aerospace Science and Engineering, National University of Defense Technology, China. + +Abstract. Camera calibration is a crucial step in photogrammetry and 3D vision applications. In practical scenarios with a long working distance to cover a wide area, target-based calibration methods become complicated and inflexible due to site limitations. This paper introduces a novel camera calibration method using a collimator system, which can provide a reliable and controllable calibration environment for cameras with varying working distances. Based on the optical geometry of the collimator system, we prove that the relative motion between the target and camera conforms to the spherical motion model, reducing the original 6DOF relative motion to 3DOF pure rotation motion. Furthermore, a closed-form solver for multiple views and a minimal solver for two views are proposed for camera calibration. The performance of our method is evaluated in both synthetic and real-world experiments, which verify the feasibility of calibration using the collimator system and demonstrate that our method is superior to the state-of-the-art methods. Demo code is available at https://github.com/LiangSK98/CollimatorCalibration + +Keywords: Camera calibration $\cdot$ Collimator system $\cdot$ Spherical motion $\cdot$ Closed-form solver $\cdot$ Minimal solver + +# 1 Introduction + +Geometric camera calibration determines the mapping between 3D rays in space and 2D image pixels. Various vision applications, such as simultaneous localization and mapping(SLAM) [6], structure from motion(SfM) [32], and pose estimation [11, 12], rely on camera parameters as inputs. The close-range calibration methods using a planar target represent the prevailing techniques in 3D vision applications [5, 26, 33, 35, 46]. In some special outdoor scenarios, such as long-range depth estimation [45] and monitoring of wind turbine blades [10], the cameras achieve a long working distance and a large field of view. The traditional target-based calibration methods [3, 37, 46] may face challenges, such as needing large targets, moving or arranging targets over a wide area, and dealing with varying and unstable lighting conditions. While constructing a 3D calibration field can offer precise calibration data [34, 40, 42], its operational complexity and limited adaptability present significant challenges. In addition, the collimator is + +![](images/dae56f8f0e73dac9831bd26e8e0e5f4f165909d42b47cc78de68a53334b7e8fd.jpg) +Fig. 1: Diagram of the collimator-based calibration method. From the inherent optical geometric properties of the collimator, the relative motion between the target and camera can be proved to conform to the spherical motion model. + +![](images/7511b3c1e41ee793df3e12e29b90c82418900e776999c27a764f787b99e779ff.jpg) + +![](images/537c6d63bd0a53ebb6d285fa51a947c270240e674aaf3028b72b99ffc6f0b0a2.jpg) + +commonly used to generate calibration targets [7]. Traditional collimator-based methods need to measure the direction of collimated rays using a goniometer or theodolite, which makes the calibration process challenging, and measurement errors directly impact the accuracy [41, 43, 44]. + +In this paper, we design a novel collimator system with a structured calibration pattern on its reticle, and a novel calibration algorithm based on the collimator images is proposed. Our method only requires the camera to observe the pattern from our collimator system at a few (at least two) different orientations. Fig. 1 illustrates this method. We briefly analyze the geometric optical property of the collimator and illustrate the angle invariance of any point pairs on the pattern. By leveraging angle invariance, we prove that the relative motion between the calibration target and the camera conforms to a spherical motion model. This constraint effectively reduces the original 6DOF general motion to the 3DOF pure rotation motion. Furthermore, we propose a novel calibration method using the collimator system. Compared with the target-based methods, more geometric constraints can be derived from the collimator, leading to higher accuracy. Compared to traditional collimator-based methods, our method is more flexible as it eliminates the need for direction measurement. + +Our primary contributions can be summarized as follows: (1) We propose a novel camera calibration method using a collimator system. The entire calibration process can be implemented with multiple close-range observations and does not require the known direction of collimated rays. (2) We prove that the relative motion between the calibration target and the camera conforms to the spherical motion model. This constraint effectively reduces the number of motion parameters to be estimated. (3) A closed-form solver for multiple views and a minimal solver for two views are proposed based on the spherical motion constraint. Moreover, the degenerate case of the proposed solvers is proved. + +# 2 Related Work + +Camera calibration is a crucial subject in 3D vision and continues to be actively researched. The literature on camera calibration introduces a wide range of al + +gorithms and toolboxes [3-5,22,26,33,35,46]. To ensure the highest accuracy for calibration, targets with known structures are usually used for camera calibration. Various calibration targets have been proposed, including 3D targets [37], 2D targets [13,18] and 1D targets [47]. In some practical applications where no calibration target is available, the camera calibration can be performed by establishing the correspondence between images during the movement of the camera, which is known as self-calibration [8,9,16,19,28,31]. Among these options, the target-based close-range calibration method is widely used. + +In outdoor long-range measurement applications [10, 45], the camera has a long working distance to measure the long-distance targets. Thus, the close-range calibration methods are impractical due to the site limitation. A commonly employed approach involves constructing a calibration field, which serves as a known structure for the calibration [21, 25, 29, 34, 40, 42]. Oniga et al. [29] lay markers on buildings to calibrate cameras mounted on UAV. Shang et al. [34] proposed a high-precision calibration method by placing several markers near the ground. Additionally, Klaus et al. [23] proposed using stars as distant targets to estimate camera parameters. + +The collimator is also commonly used to provide 2D-3D correspondence for camera calibration in the laboratory [2, 20, 41, 44]. Previous collimator-based methods proposed to use goniometer or theodolite to obtain the direction of collimated rays and then estimate the camera parameters [41, 43]. The collimator-based method offers both accuracy and flexibility in the limited space, but a high-precision goniometer or theodolite is expensive. To address this, Yuan et al. [44] designed a multiple pinhole mask for the collimator to generate the collimated rays with known directions. However, the accuracy of calibration is sensitive to machining errors. Bauer et al. [2] proposed a calibration method using a diffractive optical element (DOE). The angle between the laser beam and the camera can then be estimated to determine the camera parameters. + +# 3 Collimator System + +This section introduces the designed collimator system and shows that the equivalent relative motion between the calibration target and the camera conforms to a spherical motion model. + +# 3.1 Scheme of Collimator System + +The collimator is an optical instrument that can produce collimated rays and plays an important role when adjusting other optical instruments. A collimator consists of essential components such as a light source, a reticle, and an optical lens, as illustrated in Fig. 2. The reticle is a thin glass located on the focal plane of the collimator. The generation of collimated rays can be regarded as the inverse process of imaging a target at infinity onto the focal plane. Specifically, a point on the lens's focal plane will generate collimated rays, which can be regarded as the target at infinity. Therefore, cameras with varying focal lengths + +can consistently and accurately observe the target through the collimator at a close range. + +In this paper, we design a collimator system for camera calibration. A star-based pattern, proposed by Thomas et al. [33], is attached to the reticle of the collimator as the calibration target. The star-based pattern offers richer gradient information for improving feature extraction accuracy. Similar to the plane-based calibration method [35, 46], the star-based pattern provides 2D-3D correspondences for camera calibration. Our collimator system is portable with a size of about $200\mathrm{mm} \times 170\mathrm{mm} \times 300\mathrm{mm}$ . The collimator system is designed to provide a reliable and controllable calibration environment for some special scenarios like inappropriate illumination, limited space, and long-range cameras. However, printed targets face challenges in these scenarios. When calibrating a camera, we only need to capture the target from multiple orientations through the lens of the collimator, which is as easy as the printed targets method. + +# 3.2 Spherical Motion Model + +In this section, we prove that the relative motion between the calibration target and the camera conforms to the spherical motion model. Firstly, we briefly analyze the geometric property of the collimator. As shown in Fig. 2, let us note two arbitrary points $P_{1}$ and $P_{2}$ on the reticle of the collimator. Rays emitted from $P_{1}$ and $P_{2}$ pass through the collimator lens, generating two parallel beams observed by cameras denoted as $C_1$ and $C_2$ with arbitrary poses. The propagation routes of the two beams are visually distinguished by red and blue lines in Fig. 2. Obviously, the angle between two lines observed by the two cameras is equal, that is, $\theta_{1}\equiv \theta_{2}\equiv \theta$ . We can obtain a property that the two rays + +![](images/41ab016be5e74375c3a1a1ead198fed0102ad5984f17748cf004b8ec92ebd71f.jpg) +Fig. 2: Diagram of collimator system. The geometric property of the collimator system leads to the angle invariance of any point pair on the reticle. + +observed by the cameras always maintain a constant angle, regardless of the poses of observing cameras. We refer to this property as the angle invariance. + +Next, we show that this setup, where a camera moves around a collimator system and captures images, is theoretically equivalent to the capture setting where the camera is fixed, but the images are moving around the camera spherically. This can be derived from the angle invariance. For convenience in introducing our derivation process, we treat the camera as a fixed reference coordinate, and assume that the calibration target moves relative to the camera in space. The angle invariance allows us to infer that when the target performs a specific motion in space, the angular distance between any point pair on the target always + +remains constant. The angular distance means the angle between two vectors from the origin (camera optical center) to two points. By arbitrarily selecting a pair of points $\mathbf{P}_i$ and $\mathbf{P}_j$ on the target and transforming them, we can establish the following equation + +$$ +\angle \left(\mathbf {P} _ {i}, \mathbf {P} _ {j}\right) = \angle \left(\mathbf {R} \left(\mathbf {P} _ {i} + \mathbf {t}\right), \mathbf {R} \left(\mathbf {P} _ {j} + \mathbf {t}\right)\right), \tag {1} +$$ + +where $\mathbf{R}$ is rotation matrix and $\mathbf{t}$ is translation vector. From the rotation invariance [39] that a pair of 3D points are jointly rotated about the origin, their angular distance remains constant. We have $\angle (\mathbf{R}(\mathbf{P}_i + \mathbf{t}),\mathbf{R}(\mathbf{P}_j + \mathbf{t})) = \angle (\mathbf{P}_i + \mathbf{t},\mathbf{P}_j + \mathbf{t})$ . Thus, the angle invariance is maintained when the calibration target is arbitrarily rotated about the origin. To make $\angle (\mathbf{P}_i,\mathbf{P}_j) = \angle (\mathbf{P}_i + \mathbf{t},\mathbf{P}_j + \mathbf{t})$ hold for any pair of points on the target, the translation vector must be a $\mathbf{0}$ vector $(\mathbf{t} = [0,0,0]^T)$ . The proof is given in the supplementary material. Therefore, the calibration target performs pure rotation motion with respect to the origin (i.e. camera optical center). Such pure rotational motion is a typical spherical motion. + +The spherical motion can be visualized in Fig. 1. In the spherical motion model, the calibration target exhibits only rotational motion with a fixed distance relative to the camera. We establish a local coordinate on the target, where the $z'$ -axis aligns with the ray between the camera coordinate system $c$ and calibration target coordinate system $p'$ . Inspired by [38], the relative pose between the calibration target and camera can be mathematically represented as + +$$ +\mathbf {T} _ {c p ^ {\prime}} = \left( \begin{array}{l l} \mathbf {R} _ {c p ^ {\prime}} & \mathbf {t} _ {c p ^ {\prime}} \\ \mathbf {0} _ {1 \times 3} & 1 \end{array} \right), \tag {2} +$$ + +where $\mathbf{R}_{cp'} \in SO(3)$ , $\mathbf{t}_{cp'} = (0,0,-r)^T$ and $r$ is the radius of spherical motion. We usually establish the origin of the coordinate system $p$ on the upper left corner of the calibration pattern. Therefore, there is a relative translation $\mathbf{t}_{p'p} = (x,y,0)$ between the coordinate systems $p'$ and $p$ . Furthermore, the transformation matrix between camera $c$ and calibration target $p$ is expressed as follows: + +$$ +\mathbf {T} _ {p c} = \left[ \left( \begin{array}{l l} \mathbf {R} _ {c p ^ {\prime}} & \mathbf {t} _ {c p ^ {\prime}} \\ \mathbf {0} _ {1 \times 3} & 1 \end{array} \right) \left( \begin{array}{l l} \mathbf {I} _ {3 \times 3} & \mathbf {t} _ {p ^ {\prime} p} \\ \mathbf {0} _ {1 \times 3} & 1 \end{array} \right) \right] ^ {- 1} = \left( \begin{array}{l l} \mathbf {R} _ {p c} & - \mathbf {R} _ {p c} \mathbf {t} _ {c p} \\ \mathbf {0} _ {1 \times 3} & 1 \end{array} \right), \tag {3} +$$ + +where $\mathbf{R}_{pc} = \mathbf{R}_{cp'}^T$ and $\mathbf{t}_{cp} = \mathbf{t}_{p'p} + \mathbf{t}_{cp'} = (x,y, -r)^T$ . $\mathbf{t}_{cp}$ determines the position of the camera in the calibration target coordinate system. Note that $\mathbf{t}_{cp}$ is a fixed value and has nothing to do with the pose of the calibration target. Given the constraint from the spherical motion, the original 6DOF general motion is limited to a 3DOF pure rotation motion. + +# 4 Calibration Method + +The spherical motion constraint effectively reduces the number of motion parameters to be estimated. In this section, we show how to calibrate camera parameters using a set of collimator images. A closed-form solver for multiple views and a minimal solver for two views are proposed. + +# 4.1 Geometric Constraints + +We use the well-known pinhole model to model cameras. The mapping from a 3D point $\mathbf{P} = (X,Y,Z)^T$ to an image point $\mathbf{p} = (u,v)^T$ is computed as + +$$ +s \widetilde {\mathbf {p}} = \mathbf {K} [ \mathbf {R} \mid \mathbf {t} ] \widetilde {\mathbf {P}}, \text {w i t h} \mathbf {K} = \left[ \begin{array}{l l l} f _ {x} & \gamma & c _ {x} \\ 0 & f _ {y} & c _ {y} \\ 0 & 0 & 1 \end{array} \right], \tag {4} +$$ + +where $s$ is an arbitrary scale factor, $(\mathbf{R},\mathbf{t})$ is the camera pose with respect to the calibration target coordinate, and $\mathbf{K}$ is the camera intrinsic matrix. $\widetilde{\mathbf{p}} = [u,v,1]^T$ and $\widetilde{\mathbf{P}} = [X,Y,Z,1]^T$ denote the homogeneous coordinate forms of points $\mathbf{p}$ and $\mathbf{P}$ , respectively. The intrinsic matrix $\mathbf{K}$ consists of five parameters: the focal length $(f_x,f_y)$ , the principal point $(c_x,c_y)$ and skew factor $\gamma$ . + +Without loss of generality, we assume the planar target is on the plane $Z = 0$ of the world coordinate system. The transformation between the image plane and calibration target plane can be constrained by a $3 \times 3$ homography matrix, denoted as $\mathbf{H}$ . For perspective projection, the homography matrix $\mathbf{H}$ depends on the image pose $(\mathbf{R},\mathbf{t})$ and intrinsic matrix $\mathbf{K}$ : + +$$ +\mathbf {H} = \left[ \begin{array}{l l l} \mathbf {h} _ {1} & \mathbf {h} _ {2} & \mathbf {h} _ {3} \end{array} \right] = \lambda \mathbf {K} \left[ \begin{array}{l l l} \mathbf {r} _ {1} & \mathbf {r} _ {2} & \mathbf {t} \end{array} \right], \tag {5} +$$ + +where $\lambda$ is an arbitrary scalar, $\mathbf{r}_i$ is the $i$ -th columns of $\mathbf{R}$ . Given an image of the calibration target, the homography matrix $\mathbf{H}$ can be estimated using the plane point correspondences. Combining Eq. (3) and Eq. (5), we can get + +$$ +\mathbf {M} = \left[ \mathbf {r} _ {1} \mathbf {r} _ {2} - \mathbf {R t} _ {c p} \right] = \frac {1}{\lambda} \mathbf {K} ^ {- 1} \mathbf {H}. \tag {6} +$$ + +Using the knowledge that $\mathbf{R}$ is an orthogonal matrix, we have + +$$ +\mathbf {H} ^ {T} \mathbf {K} ^ {- T} \mathbf {K} ^ {- 1} \mathbf {H} = \lambda^ {2} \left[ \begin{array}{c c c} 1 & 0 & - x \\ 0 & 1 & - y \\ - x & - y & \| \mathbf {t} _ {c p} \| ^ {2} \end{array} \right]. \tag {7} +$$ + +There are 8 parameters to be estimated, 5 for intrinsic matrix and 3 for $\mathbf{t}_{cp} = (x,y,-r)^T$ . For each given image, Eq. (7) provides 5 independent constraints. Theoretically, a minimum of two images can completely solve all unknown parameters. In the next two subsections, we give a closed-form solver for $N(N > 2)$ collimator images and a minimal solver for two collimator images. + +# 4.2 Closed-form Solver for Multiple Views + +When there are more than two collimator images for calibration, we model the problem as a linear system and propose a closed-form solution. Suppose that the homography matrices $\mathbf{H}_i(i = 1,\dots,N)$ have been estimated for all images. The transformations between images $i$ and $j$ can be expressed as + +$$ +\mathbf {H} _ {j} ^ {- 1} \mathbf {H} _ {i} = \frac {\lambda_ {i}}{\lambda_ {j}} \mathbf {M} _ {j} ^ {- 1} \mathbf {M} _ {i}. \tag {8} +$$ + +For the matrix $\mathbf{M}$ of each image, we can calculate the determinant as + +$$ +\begin{array}{l} \left| \mathbf {M} \right| = \left| \mathbf {r} _ {1} \mathbf {r} _ {2} - \mathbf {R t} _ {c p} \right| = \left| \mathbf {r} _ {1} \mathbf {r} _ {2} - x \mathbf {r} _ {1} - y \mathbf {r} _ {2} + r \mathbf {r} _ {3} \right| \tag {9} \\ = - x \left| \mathbf {r} _ {1} \mathbf {r} _ {2} \mathbf {r} _ {1} \right| - y \left| \mathbf {r} _ {1} \mathbf {r} _ {2} \mathbf {r} _ {2} \right| + r \left| \mathbf {r} _ {1} \mathbf {r} _ {2} \mathbf {r} _ {3} \right| = r, \\ \end{array} +$$ + +where $r$ is a non-zero fixed value. Using the fact that $\operatorname*{det}(\mathbf{M}_j^{-1}\mathbf{M}_i) = 1$ , we have $\lambda_{ij} = \lambda_i / \lambda_j = \sqrt[3]{|\mathbf{H}_j^{-1}\mathbf{H}_i|}$ . Then, we choose the image with the most feature points observed as the base image. For convenience, here the first image is selected as the base image and we get $\lambda_{i1} = \sqrt[3]{|\mathbf{H}_j^{-1}\mathbf{H}_1|}$ . By inverting Eq. (7) of image $i$ , we get + +$$ +\mathbf {H} _ {i} ^ {- 1} \mathbf {K} \mathbf {K} ^ {T} \mathbf {H} _ {i} ^ {- T} = \frac {1}{\lambda_ {i} ^ {2} r ^ {2}} \left[ \begin{array}{c c c} r ^ {2} + x ^ {2} & x y & x \\ x y & r ^ {2} + y ^ {2} & y \\ x & y & 1 \end{array} \right]. \tag {10} +$$ + +Substitute $\lambda_{i} = \lambda_{i1}\lambda_{1}$ into Eq. (10), we get + +$$ +\mathbf {H} _ {i} ^ {- 1} \mathbf {K K} ^ {T} \mathbf {H} _ {i} ^ {- T} = \frac {1}{\lambda_ {i 1} ^ {2}} \mathbf {A}, \text {w i t h} \mathbf {A} = \left[ \begin{array}{c c c} \frac {r ^ {2} + x ^ {2}}{(\lambda_ {1} r) ^ {2}} & \frac {x y}{(\lambda_ {1} r) ^ {2}} & \frac {x}{(\lambda_ {1} r) ^ {2}} \\ \frac {x y}{(\lambda_ {1} r) ^ {2}} & \frac {r ^ {2} + y ^ {2}}{(\lambda_ {1} r) ^ {2}} & \frac {y}{(\lambda_ {1} r) ^ {2}} \\ \frac {x}{(\lambda_ {1} r) ^ {2}} & \frac {y}{(\lambda_ {1} r) ^ {2}} & \frac {1}{(\lambda_ {1} r) ^ {2}} \end{array} \right]. \tag {11} +$$ + +Let us define a 6-vector $\mathbf{a}$ to represent matrix $\mathbf{A}$ : + +$$ +\mathbf {a} = \left(A _ {1 1}, A _ {1 2}, A _ {1 3}, A _ {2 2}, A _ {2 3}, A _ {3 3}\right) ^ {T}, +$$ + +where $A_{mn}$ is the value of $m$ -th row and $n$ -th column of matrix $\mathbf{A}$ . Let + +$$ +\mathbf {W} = \left[ \begin{array}{c c c} c _ {x} ^ {2} + f _ {x} ^ {2} + \gamma^ {2} & c _ {x} c _ {y} + f _ {y} \gamma & c _ {x} \\ c _ {x} c _ {y} + f _ {y} \gamma & c _ {y} ^ {2} + f _ {y} ^ {2} & c _ {y} \\ c _ {x} & c _ {y} & 1 \end{array} \right]. \tag {12} +$$ + +Note that $\mathbf{W}$ is a symmetric matrix and can be defined by a 5-vector: + +$$ +\mathbf {w} = \left(W _ {1 1}, W _ {1 2}, W _ {1 3}, W _ {2 2}, W _ {2 3}\right) ^ {T}, +$$ + +where $W_{mn}$ is the value of $m$ -th row and $n$ -th column of matrix $\mathbf{W}$ . Obviously, Eq. (11) represents a linear system of equations involving the unknowns $[\mathbf{w}^T, \mathbf{a}^T]^T$ . Combining Eqs. (11) and (12), we can get the following equations for each image: + +$$ +\left[ \mathbf {V} - \lambda_ {i 1} ^ {2} \mathbf {I} _ {5 \times 5} \right] \left[ \begin{array}{l} \mathbf {w} \\ \mathbf {a} \end{array} \right] = \mathbf {b}, \tag {13} +$$ + +with + +$$ +\mathbf {V} = \left[ \mathbf {v} _ {1 1}, \mathbf {v} _ {1 2}, \mathbf {v} _ {1 3}, \mathbf {v} _ {2 2}, \mathbf {v} _ {2 3} \right] ^ {T}, +$$ + +$$ +\mathbf {v} _ {m n} = \left[ h _ {m 1} h _ {n 1}, h _ {m 1} h _ {n 2} + h _ {m 2} h _ {n 1}, h _ {m 1} h _ {n 3} + h _ {m 3} h _ {n 1}, \right. +$$ + +$$ +\left. h _ {m 2} h _ {n 2}, h _ {m 2} h _ {n 3} + h _ {m 3} h _ {n 2} \right], +$$ + +$$ +\mathbf {b} = - \left[ h _ {1 3} h _ {1 3}, h _ {1 3} h _ {2 3}, h _ {1 3} h _ {3 3}, h _ {2 3} h _ {2 3}, h _ {2 3} h _ {3 3} \right] ^ {T}, +$$ + +where $h_{mn}$ denotes the value of $m$ -th row and $n$ -th column of matrix $\mathbf{H}^{-1}$ . Given $N(N > 2)$ collimator images, we can stack $N$ such equations as Eq. (13) together in matrix form: + +$$ +\mathbf {D} \left[ \begin{array}{l} \mathbf {w} \\ \mathbf {a} \end{array} \right] = \mathbf {b}, \tag {14} +$$ + +where $\mathbf{D}$ is a $5N\times 11$ matrix. The closed-form solution of Eq. (14) is given by + +$$ +\left[ \begin{array}{l} \mathbf {w} \\ \mathbf {a} \end{array} \right] = \left(\mathbf {D} ^ {T} \mathbf {D}\right) ^ {- 1} \mathbf {D} ^ {T} \mathbf {b}. \tag {15} +$$ + +Once $\mathbf{w}$ and $\mathbf{a}$ are estimated, we can calculate all camera intrinsic parameters and optical center $\mathbf{t}_{cp}$ by decomposing $\mathbf{W}$ and $\mathbf{A}$ respectively: + +$$ +\begin{array}{l} c _ {x} = W _ {1 3}, \quad c _ {y} = W _ {2 3}, \quad f _ {y} = \sqrt {W _ {2 2} - c _ {y} ^ {2}}, \\ \gamma = \left(W _ {1 2} - c _ {x} c _ {y}\right) / f _ {y}, \quad f _ {x} = \sqrt {W _ {1 1} - c _ {x} ^ {2} - \gamma^ {2}}, \\ x = A _ {1 3} / A _ {3 3}, \quad y = A _ {2 3} / A _ {3 3}, \\ r = - \sqrt {\left(A _ {1 1} / A _ {3 3}\right) - \left(A _ {1 3} / A _ {3 3}\right) ^ {2}}. \\ \end{array} +$$ + +Subsequently, the rotation matrix of each image can be calculated with Eq. (5): + +$$ +\mathbf {r} _ {1} = (1 / \lambda) \mathbf {K} ^ {- 1} \mathbf {h} _ {1}, \quad \mathbf {r} _ {2} = (1 / \lambda) \mathbf {K} ^ {- 1} \mathbf {h} _ {2}, \quad \mathbf {r} _ {3} = \mathbf {r} _ {1} \times \mathbf {r} _ {2}, \tag {16} +$$ + +where $\lambda = (\| \mathbf{K}^{-1}\mathbf{h}_1\| +\| \mathbf{K}^{-1}\mathbf{h}_2\|) / 2$ . Finally, a singular value decomposition for re-orthogonalizing the rotation matrix $\mathbf{R} = [\mathbf{r}_1,\mathbf{r}_2,\mathbf{r}_3]$ is required. + +# 4.3 Minimal Solver for Two Views + +When only two collimator images are used for calibration, the rank of the coefficient matrix in Eq. (14) is determined to be 10. Solving rank-deficient equations presents a complex challenge. To address this issue, we construct a nonlinear equation system and propose a novel approach that utilizes the hidden variable technique [14] to estimate the unknowns in the minimum configuration. + +Estimation of $\mathbf{W} = \mathbf{K}^{-T}\mathbf{K}^{-1}$ is the core of calibration. Note that $\mathbf{W}$ is a symmetric matrix, defined by a 6-vector $\mathbf{w} = (W_{11}, W_{12}, W_{13}, W_{22}, W_{23}, W_{33})^T$ . We can extract the following 5 constraints from Eq. (7) + +$$ +\mathbf {v} _ {1 2} ^ {T} \mathbf {w} = 0 \tag {17} +$$ + +$$ +\left(\mathbf {v} _ {1 1} ^ {T} - \mathbf {v} _ {2 2} ^ {T}\right) \mathbf {w} = 0 \tag {18} +$$ + +$$ +\mathbf {v} _ {1 3} ^ {T} \mathbf {w} + x \mathbf {v} _ {1 1} ^ {T} \mathbf {w} = 0 \tag {19} +$$ + +$$ +\mathbf {v} _ {2 3} ^ {T} \mathbf {w} + y \mathbf {v} _ {1 1} ^ {T} \mathbf {w} = 0 \tag {20} +$$ + +$$ +\mathbf {v} _ {3 3} ^ {T} \mathbf {w} - \| \mathbf {t} _ {c p} \| ^ {2} \mathbf {v} _ {1 1} ^ {T} \mathbf {w} = 0. \tag {21} +$$ + +$\mathbf{v}_{mn}$ can be calculated as following: + +$$ +\mathbf {v} _ {m n} = \left[ h _ {1 m} h _ {1 n}, h _ {1 m} h _ {2 n} + h _ {1 n} h _ {2 m}, h _ {1 m} h _ {3 n} + h _ {1 n} h _ {3 m} \right. +$$ + +$$ +\left. h _ {2 m} h _ {2 n}, h _ {2 m} h _ {3 n} + h _ {2 n} h _ {3 m}, h _ {3 m} h _ {3 n} \right] ^ {T}, +$$ + +where $h_{mn}$ denotes the value of $m$ -th row and $n$ -th column of matrix $\mathbf{H}$ . Given only 2 collimator images, we can obtain 10 independent equations for 8 unknowns. The hidden variable technique [14] is employed to solve the above polynomial equations. Combining Eqs. (19) to (21), we have + +$$ +\left(\mathbf {v} _ {1 3} ^ {T} + \mathbf {v} _ {2 3} ^ {T} + \mathbf {v} _ {3 3} ^ {T}\right) \mathbf {w} + c \mathbf {v} _ {1 1} ^ {T} \mathbf {w} = 0, \tag {22} +$$ + +where $c = (x + y - \| \mathbf{t}_{cp} \|^2)$ . We first select constraints Eqs. (17), (18) and (22) to establish a new polynomial equation system. Next, we treat $c$ as a hidden variable, and the complete set of equations can be expressed in matrix form as + +$$ +\mathbf {C} (c) \mathbf {w} = 0, \text {w i t h} \mathbf {C} (c) = \left[ \begin{array}{c} \mathbf {v} _ {1 2} ^ {T} \\ (\mathbf {v} _ {1 1} ^ {T} - \mathbf {v} _ {2 2} ^ {T}) \\ (\mathbf {v} _ {1 3} ^ {T} + \mathbf {v} _ {2 3} ^ {T} + \mathbf {v} _ {3 3} ^ {T}) + c \mathbf {v} _ {1 1} ^ {T} \end{array} \right], \tag {23} +$$ + +where the coefficient matrix $\mathbf{C}(c)$ contains the hidden variable $c$ . Two images provide 6 equations for solving $c$ , and $\mathbf{C}(c)$ is a $6 \times 6$ matrix. Since the equation system $\mathbf{C}(c)\mathbf{w} = 0$ must have at least one nontrivial solution for $\mathbf{w}$ , the determinant of $\mathbf{C}(c)$ is equal to zero. The equation, $\operatorname*{det}(\mathbf{C}(c)) = 0$ , is a polynomial of degree 2 in terms of $c$ , which can be efficiently solved using a straightforward root-finding algorithm. Substituting the known value of $c$ into the constraint equations Eqs. (17), (18) and (22), we obtain six constraint equations about $\mathbf{w}$ . The least squares method is then employed to solve these linear equations and compute the optimal $\mathbf{w}$ up to scale. Once $\mathbf{w}$ is estimated, the intrinsic matrix $\mathbf{K}$ can be calculated using the Cholesky factorization [15]. + +# 4.4 Bundle Adjustment + +Bundle adjustment jointly refines the camera intrinsic $\mathbf{K}$ , lens distortion coefficients $\mathbf{d}$ , the rotation matrix of the $i$ -th image $\mathbf{R}_i$ and camera optical center $\mathbf{t}_{cp}$ . In cases of moderate lens distortion, it is typical to initialize the distortion coefficient to $\mathbf{0}$ and estimate them in the bundle adjustment step [36]. Given a set of $N$ images, each with $M$ corresponding 2D image points, we establish a cost function aimed at minimizing the re-projection error. The cost function is expressed as follows: + +$$ +\min _ {\mathbf {X}} \sum_ {i} ^ {N} \sum_ {j} ^ {M} \rho \left(\left\| \pi \left(\mathbf {K}, \mathbf {d}, \mathbf {R} _ {i}, \mathbf {t} _ {c p}, \mathbf {P} _ {i j}\right) - \mathbf {p} _ {i j} \right\| ^ {2}\right), \tag {24} +$$ + +where $\mathbf{X} = (\mathbf{K},\mathbf{d},\mathbf{R}_i,\mathbf{t}_{cp})$ is the parameter to be refined. $\mathbf{P}_{ij}$ is the $j$ -th 3D pattern point of $i$ -th image and corresponds to the 2D image point $\mathbf{p}_{ij}$ . $\pi (\cdot)$ projects the world points to image according to Eq. (4). $\rho (\cdot)$ is the Cauchy cost function. As is common, we optimize Eq. (24) with the Levenberg-Marquardt (LM) method [27], which is implemented in some excellent frameworks [1]. + +Note that the camera optical center is fixed at $\mathbf{t}_{cp}$ with the spherical motion constraint. However, for general motion, the camera optical center is free. Therefore, when calibrating the camera using collimator images, the number of motion parameters to be refined is reduced from $6N$ to $3N + 3$ . + +# 4.5 Degenerated Configuration + +We prove a degenerate configuration where additional images cannot provide more constraints about the camera intrinsic parameter. + +Proposition 1. If the calibration target rotates only around an axis that is parallel to the $z$ axis and intersects the target plane at $(x,y,0)^T$ , the newly added image cannot provide additional constraints on the camera intrinsic parameters. + +Due to space limitations, the proof is provided in the supplementary material. + +# 5 Evaluation and Results + +In this section, we first evaluate the robustness and accuracy of the proposed algorithm on synthetic data. Then, we calibrate the camera and evaluate the effectiveness of the proposed method using real collimator images. + +# 5.1 Synthetic Data Experiments + +In the synthetic data experiment, we evaluate the performance of the proposed algorithm. We simulate the process of calibrating predefined camera parameters using synthetic images that conform to the spherical motion model. The proposed is compared with the following algorithms: + +- Zhang [46] is a plane-based algorithm and suitable for general 6DOF motion. +- Bouguet [3] is a calibration toolbox and suitable for general 6DOF motion. Unlike Zhang, Bouguet computes the focal length using the orthogonal vanishing points constraint. +- Hartley [17] is a self-calibration method based on a pure rotating camera, which is specifically suitable for the 3DOF spherical motion model. + +We define a virtual camera with the following parameters: $f_{x} = f_{y} = 1000$ , $c_{x} = 542$ , $c_{y} = 478$ , $\gamma = 0.01$ , radial distortion coefficients $d_{1} = 0.1$ , $d_{2} = -0.2$ , and the image size is set to $1080 \times 960$ pixels. The calibration target is a planar pattern containing $11 \times 8$ points, each square size of $30mm \times 30mm$ . The target maintains a spherical motion with a radius of $700mm$ relative to the camera. The camera optical center in the target coordinate system is $t_{cp} = [150, 105, -700]^{T}mm$ , and the camera's orientation is arbitrary. Given the camera parameters and poses, the planar target points are projected onto images. Finally, zero-mean Gaussian noise with $\sigma$ standard deviation is added to 2D image points. + +Evaluation of Calibration. In this experiment, we seek to evaluate the performance of the algorithms in solving camera parameters. For the four algorithms, we perform two steps of solving initial value and bundle adjustment to obtain camera parameters. Firstly, we compare the performance of the mentioned algorithms as the noise level increases. The noise level varies from 0 to 3.0 pixels, and we use 15 images to calibrate the camera. For each noise level, we conducted 200 trials and calculated the average errors. As shown in Fig. 3, calibration error increases linearly with the noise level. Compared with the other three algorithms, ours shows greater robustness to noise. Additionally, at a noise + +![](images/2e124ae4208e52f24641e2988d76866c8ef6bc7fdb5781abe9eaada7c108dc26.jpg) +Fig. 3: Comparison of calibration errors with increasing noise level. + +![](images/1c25801bcbb8e6190215e5e34675e904e1525946a08b499cf88fbae47fe9765a.jpg) + +![](images/bc1076d33f01ccb87ffdf2f913b9c964067086737439128326e4dad7b4190a50.jpg) + +![](images/9e7ca5d23c88494883ca0fa5b7f0a0e794cb7a7d8202c7e47660a30cc773620b.jpg) + +![](images/d93f5f3c79cb58859b4b3fcf9c38bddd090015ddb8eff8270397ed3568f3e503.jpg) +Fig. 4: Comparison of calibration errors with the number of images increases. + +![](images/e9df0542566cdf33338b02f3c618fdd99fb618ee7483fea2d108057fc1feaaaf.jpg) + +![](images/64074298e4f3bcd9f9414e821b3874e7e72962c09b4337c94b6f621eacd8bd56.jpg) + +![](images/7ab5cbb4daf7b03699e49b69f7f31651c5c52c38a1ddac561eac6757314e7c67.jpg) + +level of 0.5 pixels, we evaluate the performance with respect to the number of calibration images. The image count ranged from 3 to 30, with 200 trials conducted for each quantity. The results are illustrated in Fig. 4. A significant decrease in error is observed when increasing the number of images from 3 to 5. Both Fig. 3 and Fig. 4 show that our algorithm outperforms other algorithms in robustness and accuracy. We believe that our algorithm benefits from the consideration of both spherical motion and planar homography constraints, along with reducing optimization variables. Due to space limitations, the additional comparison results of the initial value estimation are shown in the supplementary material. + +Sensitivity of imperfect spherical motion. The low-cost collimators exhibit an inferior parallelism, which is reflected in the fact that the relative motion between the target and camera is not a perfect spherical motion. This experiment evaluates the sensitivity of initial value estimation with respect to imperfect spherical motion. We add zeros-mean Gaussian noise with standard deviation from $0mm$ to $30mm$ to $\mathbf{t}_{cp}$ to replicate this imperfection. We use 15 images and add a noise level of 0.5 pixels to the image points. For each noise level, a total of independent 500 trials are conducted and the average errors are computed. The sensitivity analysis results are shown in Fig. 5. Noise has no impact on Bouguet and Zhang, because they do not rely on the spherical motion constraint. As the noise increases, calibration errors of Ours and Hartley increase gradually. Notably, Hartley is significantly influenced by noise, whereas Ours displays superior robustness. Even with noise levels below $15mm$ (equivalent to $2.143\%$ of the spherical radius), our algorithm remains superior to the alternatives. + +![](images/4a215737a0b6bb2ae96cecbb6dbe73a3c26e43c0929e5dcb1e765036cd3a4a52.jpg) +Fig. 5: Sensitivity analysis of imperfect spherical motion. + +![](images/b10433ef99734c73e1bca81cff90288ef225737bc6e852db743a65f53bb80c69.jpg) + +Table 1: Comparison of calibration and pose estimation error results (pixels). + +
NumberZhang [46]Bouguet [3]Hartley [17]Ours
εcalibεevalεcalibεevalεcalibεevalεcalibεeval
20.18420.15450.17390.1518--0.17390.1519
50.23160.12450.22020.1236-0.25770.22030.1236
100.19230.12310.19120.1232-0.19860.19510.1222
150.17750.12210.17180.1232-0.17910.17790.1216
200.17730.12250.17220.1221-0.18710.17780.1207
+ +# 5.2 Real Images Experiments + +In real image experiments, we verify the effectiveness of our collimator system for calibration and evaluate the performance of the proposed algorithm. + +Calibration and Accuracy Evaluation. Following the experimental methodology utilized in previous studies [13,30], we design a pose estimation experiment to evaluate the accuracy of camera calibration quantitatively. A 8mm camera captures images from our collimator system and a practical printed pattern. Calibration is performed using the collimator images and a subset of the printed pattern images, while the remaining printed pattern images are used to evaluate the calibration results. After obtaining the camera parameters from the calibration images, the pose of each evaluation image can be determined using absolute pose estimation [24]. In our setup, a quarter of the corner points on the printed pattern are used for pose estimation, and the remaining points are used for evaluation. By re-projecting the evaluation corner points onto the image with the camera parameters and the calculated pose, we can evaluate the calibration quality based on the re-projection error. A total of 200 printed pattern images are used to evaluate the calibration. The calibration uses 2, 5, 10, 15, and 20 images. The re-projection error of the calibration and evaluation images are donated as $\epsilon_{calib}$ and $\epsilon_{eval}$ , respectively. + +The comparison results of the four algorithms are shown in Tab. 1. Due to the limited degrees of freedom, our calibrated re-projection error $(\epsilon_{calib})$ is slightly higher. However, Ours exhibits the lowest re-projection error in evaluation images compared to alternative methods, indicating that our method achieves more precise camera parameters. The Hartley method only uses the correspondence of 2D points in images, which makes it impossible to calculate the re-projection + +Table 2: Comparison of calibration results using 20 images in a real camera. + +
PrintedCollimator
Bouguet [3]Zhang [46]Bouguet [3]Hartley [17]Ours
fx2384.52380.82369.22644.62388.3
fy2386.32379.82368.92646.02387.9
cx1209.81222.11221.11214.61221.0
cy1008.21008.41009.81012.71009.6
k1-0.0843-0.0899-0.0908-0.0754-0.0900
k20.08190.09120.08920.11810.0954
εcalib0.15130.17730.1722-0.1778
εeval0.12250.12250.12210.18710.1207
+ +Table 3: Comparison of epipolar rectification error results (pixels). + +
NumberZhang [46]Bouguet [3]Hartley [17]Ours
meanstdmeanstdmeanstdmeanstd
20.08460.04060.07700.0220--0.07700.0229
50.08090.03070.07640.02140.08320.03970.07620.0211
100.07730.02340.07580.02130.08540.05090.07570.0208
150.07820.02660.07710.02530.08370.05220.07560.0207
200.07720.02520.07680.02530.08180.03650.07560.0208
+ +error of the calibration images directly. In cases with only two images, Zhang and Bouguet commonly depend on beneficial assumptions, such as $\gamma = 0$ . Ours achieves comparable evaluation errors without the need for such assumptions. + +Additionally, the Bouguet toolbox is also used to calculate camera parameters from printed pattern images that are not included in the evaluation. We use 20 images for calibration and 200 for evaluation. The estimated camera parameters and error comparison are shown in Tab. 2. The second column is a good reference. The lower evaluation error indicates that our method outperforms the alternative and validates the effectiveness of using the collimator image for calibration. It can be expected that the calibration parameters are not significantly different. Therefore, even a minor improvement holds considerable significance. + +Then, we evaluate the performance of all methods using epipolar rectification error, which has not been optimized. Given the camera parameters from different methods, we calculate the epipolar rectification error (in pixels) for 300 pairs of printed evaluation images. The mean and standard deviation of the errors are shown in Tab. 3. Moreover, the result using Bouguet with 20 printed pattern images is (0.0772, 0.0259) pixels. It is observed that as the number of calibration images increases, the error stabilizes. The results indicate improved accuracy in the collimator image and highlight the beneficial impact of the spherical constraint on accuracy. + +Multiple Cameras Calibration with a Same Collimator. In this experiment, we evaluate our method on multiple cameras with the same collimator. We used the cameras with $12mm$ , $16mm$ , $25mm$ , and $35mm$ focal lengths of lenses. The calibration images are captured from the same collimator, as shown in Fig. 6 and left of Fig. 1. As the focal length increases, the more pixels a single + +![](images/832534a9e98b421a82ef6735dfe964b146ccf1b579a1712e7798c7ae8f7fe012.jpg) +Fig. 6: Calibration image sample obtained from the collimator. + +Table 4: Comparison of calibration results with different focal length lens (pixels). + +
Zhang [46]Bouguet [3]Hartley [17]Ours
12mm0.22730.22250.34040.2174
16mm0.29940.25910.41820.2536
25mm0.31330.30780.41950.2923
35mm0.32890.33480.72720.3118
+ +square occupies. Cameras with different focal lengths can observe the calibration target through our collimator system in a limited space. + +Similar to the evaluation approach in the previous experiment, we segregate all collimator images into calibration and evaluation sets. For each lens, we use 20 images for calibration and 200 for evaluation. The results are listed in Tab. 4. The proposed algorithm demonstrates the lowest re-projection error for lenses with varying focal lengths among the evaluation images. The re-projection error tends to increase with the increase of focal length. We believe that the main reason is that the longer the lens's focal length, the fewer points observed. More importantly, this experiment confirms the effectiveness of our collimator system and the proposed method. Moreover, we provide a spherical motion verification experiment in supplementary material. + +# 6 Conclusion + +This paper presents a novel camera calibration method using a designed collimator system. Based on the optical geometry of the collimator system, we prove that the relative motion between the target and the camera conforms to a spherical motion. Furthermore, we propose a closed-form solver for multiple views and a minimal solver for two views for camera calibration. Our collimator-based method effectively combines the advantages of close-range calibration using a collimator and flexibility using a planar calibration target. We validate the effectiveness of using only collimator images to calibrate the camera. Synthetic and real data experiments demonstrate that the calibration accuracy of the proposed method is superior to the state-of-the-art methods. + +Limitations: When calibrating a fisheye camera, the distortion should be paid more attention rather than simply setting it to 0. Future work will focus on cameras with significant distortion. + +# Acknowledgements + +This research has been supported by the Hunan Provincial Natural Science Foundation for Excellent Young Scholars under Grant 2023JJ20045, and the National Natural Science Foundation of China under Grant 12372189. We thank Zijie Wang for useful discussion on the writing of the paper. + +# References + +1. Agarwal, S., Mierle, K., Team, T.C.S.: Ceres Solver. https://github.com/ceres-solver/ceres-solver +2. Bauer, M., Grießbach, D., Hermerschmidt, A., Krüger, S., Scheele, M., Schischmanow, A.: Geometrical camera calibration with diffractive optical elements. Optics express 16(25), 20241-20248 (2008) +3. Bouguet, J.Y.: Camera calibration toolbox for matlab. http://robots.stanford.edu/cs223b04/JeanYvesCalib/index.html (2004) +4. Bradski, G.: The opencv library. Dr. Dobb's Journal: Software Tools for the Professional Programmer 25(11), 120-123 (2000) +5. Brown, D.C.: Close-range camera calibration. Photogramm. Eng 37(8), 855-866 (1971) +6. Campos, C., Elvira, R., Rodríguez, J.J.G., M. Montiel, J.M., D. Tardós, J.: Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics 37(6), 1874-1890 (2021) +7. Clarke, T.A., Fryer, J.G.: The development of camera calibration methods and models. The Photogrammetric Record 16(91), 51-66 (1998) +8. Faugeras, O.D., Luong, Q.T., Maybank, S.J.: Camera self-calibration: Theory and experiments. In: IEEE European Conference on Computer Vision (ECCV). p. 321-334 (1992) +9. Fuhr, G., Jung, C.R.: Camera self-calibration based on nonlinear optimization and applications in surveillance systems. IEEE Transactions on Circuits and Systems for Video Technology 27(5), 1132-1142 (2017) +0. Guan, B., Su, Z., Yu, Q., Li, Z., Feng, W., Yang, D., Zhang, D.: Monitoring the blades of a wind turbine by using videogrammetry. Optics and Lasers in Engineering 152, 106901 (2022) +1. Guan, B., Zhao, J., Barath, D., Fraundorfer, F.: Minimal solvers for relative pose estimation of multi-camera systems using affine correspondences. International Journal of Computer Vision 131(1), 324-345 (2023) +2. Guan, B., Zhao, J., Li, Z., Sun, F., Fraundorfer, F.: Relative pose estimation with a single affine correspondence. IEEE Transactions on Cybernetics 52(10), 10111-10122 (2022) +3. Ha, H., Perdoch, M., Alismail, H., Kweon, I.S., Sheikh, Y.: Deltille grids for geometric camera calibration. In: IEEE International Conference on Computer Vision (ICCV). pp. 5354-5362 (2017) +4. Hartley, R., Li, H.: An efficient hidden variable approach to minimal-case camera motion estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence 34(12), 2303-2314 (2012) +5. Hartley, R., Zisserman, A.: Multiple View Geometry in Computer Vision. Cambridge University Press, 2 edn. (2004) + +16. Hartley, R.I.: Self-calibration from multiple views with a rotating camera. In: European Conference on Computer Vision (ECCV). pp. 471-478 (1994) +17. Hartley, R.I.: Self-calibration of stationary cameras. International Journal of Computer Vision 22(1), 5-23 (1997) +18. Heikkila, J.: Geometric camera calibration using circular control points. IEEE Transactions on Pattern Analysis and Machine Intelligence 22(10), 1066-1077 (2000) +19. Herrera, D., Kannala, C.J., Heikkila, J.: Forget the checkerboard: Practical self-calibration using a planar scene. In: IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 1-9. IEEE (2016) +20. Hieronymus, J.: Comparison of methods for geometric camera calibration. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences 39, 595-599 (2012) +21. Hu, H., Wei, B., Mei, S., Liang, J., Zhang, Y.: A two-step calibration method for vision measurement with large field of view. IEEE Transactions on Instrumentation and Measurement 71, 1-10 (2022) +22. Kannala, J., Brandt, S.S.: A generic camera model and calibration method for conventional, wide-angle, and fish-eye lenses. IEEE Transactions on Pattern Analysis and Machine Intelligence 28(8), 1335-1340 (2006) +23. Klaus, A., Bauer, J., Karner, K., Elbischger, P., Perko, R., Bischof, H.: Camera calibration from a single night sky image. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). vol. 1, pp. I-I (2004) +24. Kneip, L., Furgale, P.: Opengv: A unified and generalized approach to real-time calibrated geometric vision. In: IEEE International Conference on Robotics and Automation (ICRA). pp. 1-8 (2014) +25. Liu, Z., Li, F., Li, X., Zhang, G.: A novel and accurate calibration method for cameras with large field of view using combined small targets. Measurement 64, 1-16 (2015) +26. Lochman, Y., Liepieshov, K., Chen, J., Perdoch, M., Zach, C., Pritts, J.: Babel-calib: A universal approach to calibrating central cameras. In: IEEE International Conference on Computer Vision (ICCV). pp. 15253-15262 (2021) +27. Moré, J.J.: The levenberg-marquardt algorithm: implementation and theory. In: Numerical analysis: proceedings of the biennial Conference held at Dundee, June 28-July 1, 1977. pp. 105-116 (2006) +28. Nister, D., Stewenius, H., Grossmann, E.: Non-parametric self-calibration. In: IEEE International Conference on Computer Vision (ICCV). vol. 1, pp. 120-127 (2005) +29. Oniga, V.E., Pfeifer, N., Loghin, A.M.: 3d calibration test-field for digital cameras mounted on unmanned aerial systems (uas). Remote Sensing 10(12), 2017 (2018) +30. Peng, S., Sturm, P.: Calibration wizard: A guidance system for camera calibration based on modelling geometric and corner uncertainty. In: IEEE International Conference on Computer Vision (ICCV). pp. 1497-1505 (2019) +31. Pollefeys, M., Koch, R., Gool, L.V.: Self-calibration and metric reconstruction in spite of varying and unknown intrinsic camera parameters. International journal of computer vision 32(1), 7-25 (1999) +32. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4104-4113 (2016) +33. Schöps, T., Larsson, V., Pollefeys, M., Sattler, T.: Why having 10,000 parameters in your camera model is better than twelve. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2532-2541 (2020) + +34. Shang, Y., Sun, X., Yang, X., Wang, X., Yu, Q.: A camera calibration method for large field optical measurement. Optik - International Journal for Light and Electron Optics 124, 6553-6558 (12 2013) +35. Sturm, P.F., Maybank, S.J.: On plane-based camera calibration: A general algorithm, singularities, applications. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). vol. 1, pp. 432-437. IEEE (1999) +36. Triggs, B., McLauchlan, P.F., Hartley, R.I., Fitzgibbon, A.W.: Bundle adjustment—a modern synthesis. In: Vision Algorithms: Theory and Practice: International Workshop on Vision Algorithms. pp. 298-372. Springer (2000) +37. Tsai, R.: A versatile camera calibration technique for high-accuracy 3d machine vision metrology using off-the-shelf tv cameras and lenses. IEEE Journal on Robotics and Automation 3(4), 323-344 (1987) +38. Ventura, J.: Structure from motion on a sphere. In: European Conference on Computer Vision (ECCV). vol. 9907, pp. 53-68 (2016) +39. Wang, C., Liu, Y., Wang, Y., Li, X., Wang, M.: Efficient and outlier-robust simultaneous pose and correspondence determination by branch-and-bound and transformation decomposition. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(10), 6924-6938 (2022) +40. Wang, Z., Wu, Z., Zhen, X., Yang, R., Xi, J., Chen, X.: A two-step calibration method of a largefov binocular stereovision sensor for onsite measurement. Measurement 62, 15-24 (2015) +41. Wu, G., Han, B., He, X.: Calibration of geometric parameters of line array ccd camera based on exact measuring angle in lab. Optics and Precision Engineering 15(10), 1628-1632 (2007) +42. Xiao, Z., Jin, L., Yu, D., Tang, Z.: A cross-target-based accurate calibration method of binocular stereo systems with large-scale field-of-view. Measurement 43(6), 747-754 (2010) +43. Yuan, G., Zheng, L., Ding, Y., Zhang, H., Zhang, X., Liu, X., Sun, J.: A precise calibration method for line scan cameras. IEEE Transactions on Instrumentation and Measurement 70, 1-9 (2021) +44. Yuan, G., Zheng, L., Sun, J., Liu, X., Wang, X., Zhang, Z.: Practical calibration method for aerial mapping camera based on multiple pinhole collimator. IEEE Access 8, 39725-39733 (2019) +45. Zhang, K., Xie, J., Snavely, N., Chen, Q.: Depth sensing beyond lidar range. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1689-1697 (2020) +46. Zhang, Z.: A flexible new technique for camera calibration. IEEE Transactions on Pattern Analysis and Machine Intelligence 22(11), 1330-1334 (2000) +47. Zhang, Z.: Camera calibration with one-dimensional objects. IEEE Transactions on Pattern Analysis and Machine Intelligence 26(7), 892-899 (2004) \ No newline at end of file diff --git a/2024/Camera Calibration using a Collimator System/images.zip b/2024/Camera Calibration using a Collimator System/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4b38610ef4c2108d6dd8899f86b80d0c753d0f4d --- /dev/null +++ b/2024/Camera Calibration using a Collimator System/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56849bfa53811a1133a7358422eb05e44dd113a569f34c20bba9a13682f77ba9 +size 524405 diff --git a/2024/Camera Calibration using a Collimator System/layout.json b/2024/Camera Calibration using a Collimator System/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9663516897027fbc4eb2eabb0eaf7490e2257967 --- /dev/null +++ b/2024/Camera Calibration using a Collimator System/layout.json @@ -0,0 +1,11758 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 142, + 112, + 473, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 112, + 473, + 131 + ], + "spans": [ + { + "bbox": [ + 142, + 112, + 473, + 131 + ], + "type": "text", + "content": "Camera Calibration using a Collimator System" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 150, + 477, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 150, + 477, + 175 + ], + "spans": [ + { + "bbox": [ + 138, + 150, + 477, + 175 + ], + "type": "text", + "content": "Shunkun Liang, Banglei Guan, Zhenbao Yu, Pengju Sun, and Yang Shang" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 148, + 184, + 465, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 184, + 465, + 207 + ], + "spans": [ + { + "bbox": [ + 148, + 184, + 465, + 207 + ], + "type": "text", + "content": "College of Aerospace Science and Engineering, National University of Defense Technology, China." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 236, + 455, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 236, + 455, + 413 + ], + "spans": [ + { + "bbox": [ + 160, + 236, + 455, + 413 + ], + "type": "text", + "content": "Abstract. Camera calibration is a crucial step in photogrammetry and 3D vision applications. In practical scenarios with a long working distance to cover a wide area, target-based calibration methods become complicated and inflexible due to site limitations. This paper introduces a novel camera calibration method using a collimator system, which can provide a reliable and controllable calibration environment for cameras with varying working distances. Based on the optical geometry of the collimator system, we prove that the relative motion between the target and camera conforms to the spherical motion model, reducing the original 6DOF relative motion to 3DOF pure rotation motion. Furthermore, a closed-form solver for multiple views and a minimal solver for two views are proposed for camera calibration. The performance of our method is evaluated in both synthetic and real-world experiments, which verify the feasibility of calibration using the collimator system and demonstrate that our method is superior to the state-of-the-art methods. Demo code is available at https://github.com/LiangSK98/CollimatorCalibration" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 422, + 453, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 422, + 453, + 445 + ], + "spans": [ + { + "bbox": [ + 160, + 422, + 453, + 445 + ], + "type": "text", + "content": "Keywords: Camera calibration " + }, + { + "bbox": [ + 160, + 422, + 453, + 445 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 422, + 453, + 445 + ], + "type": "text", + "content": " Collimator system " + }, + { + "bbox": [ + 160, + 422, + 453, + 445 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 422, + 453, + 445 + ], + "type": "text", + "content": " Spherical motion " + }, + { + "bbox": [ + 160, + 422, + 453, + 445 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 422, + 453, + 445 + ], + "type": "text", + "content": " Closed-form solver " + }, + { + "bbox": [ + 160, + 422, + 453, + 445 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 422, + 453, + 445 + ], + "type": "text", + "content": " Minimal solver" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 466, + 230, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 466, + 230, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 466, + 230, + 478 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 491, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 647 + ], + "type": "text", + "content": "Geometric camera calibration determines the mapping between 3D rays in space and 2D image pixels. Various vision applications, such as simultaneous localization and mapping(SLAM) [6], structure from motion(SfM) [32], and pose estimation [11, 12], rely on camera parameters as inputs. The close-range calibration methods using a planar target represent the prevailing techniques in 3D vision applications [5, 26, 33, 35, 46]. In some special outdoor scenarios, such as long-range depth estimation [45] and monitoring of wind turbine blades [10], the cameras achieve a long working distance and a large field of view. The traditional target-based calibration methods [3, 37, 46] may face challenges, such as needing large targets, moving or arranging targets over a wide area, and dealing with varying and unstable lighting conditions. While constructing a 3D calibration field can offer precise calibration data [34, 40, 42], its operational complexity and limited adaptability present significant challenges. In addition, the collimator is" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 372, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 372, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 372, + 665 + ], + "type": "text", + "content": "Corresponding author. guanbanglei12@nudt.edu.cn" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 112, + 233, + 217 + ], + "blocks": [ + { + "bbox": [ + 133, + 112, + 233, + 217 + ], + "lines": [ + { + "bbox": [ + 133, + 112, + 233, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 112, + 233, + 217 + ], + "type": "image", + "image_path": "dae56f8f0e73dac9831bd26e8e0e5f4f165909d42b47cc78de68a53334b7e8fd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 224, + 482, + 258 + ], + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 258 + ], + "type": "text", + "content": "Fig. 1: Diagram of the collimator-based calibration method. From the inherent optical geometric properties of the collimator, the relative motion between the target and camera can be proved to conform to the spherical motion model." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 238, + 113, + 334, + 217 + ], + "blocks": [ + { + "bbox": [ + 238, + 113, + 334, + 217 + ], + "lines": [ + { + "bbox": [ + 238, + 113, + 334, + 217 + ], + "spans": [ + { + "bbox": [ + 238, + 113, + 334, + 217 + ], + "type": "image", + "image_path": "7511b3c1e41ee793df3e12e29b90c82418900e776999c27a764f787b99e779ff.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 336, + 113, + 481, + 216 + ], + "blocks": [ + { + "bbox": [ + 336, + 113, + 481, + 216 + ], + "lines": [ + { + "bbox": [ + 336, + 113, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 336, + 113, + 481, + 216 + ], + "type": "image", + "image_path": "537c6d63bd0a53ebb6d285fa51a947c270240e674aaf3028b72b99ffc6f0b0a2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 279, + 479, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 479, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 479, + 327 + ], + "type": "text", + "content": "commonly used to generate calibration targets [7]. Traditional collimator-based methods need to measure the direction of collimated rays using a goniometer or theodolite, which makes the calibration process challenging, and measurement errors directly impact the accuracy [41, 43, 44]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 328, + 480, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 480, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 480, + 495 + ], + "type": "text", + "content": "In this paper, we design a novel collimator system with a structured calibration pattern on its reticle, and a novel calibration algorithm based on the collimator images is proposed. Our method only requires the camera to observe the pattern from our collimator system at a few (at least two) different orientations. Fig. 1 illustrates this method. We briefly analyze the geometric optical property of the collimator and illustrate the angle invariance of any point pairs on the pattern. By leveraging angle invariance, we prove that the relative motion between the calibration target and the camera conforms to a spherical motion model. This constraint effectively reduces the original 6DOF general motion to the 3DOF pure rotation motion. Furthermore, we propose a novel calibration method using the collimator system. Compared with the target-based methods, more geometric constraints can be derived from the collimator, leading to higher accuracy. Compared to traditional collimator-based methods, our method is more flexible as it eliminates the need for direction measurement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 495, + 481, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 481, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 481, + 602 + ], + "type": "text", + "content": "Our primary contributions can be summarized as follows: (1) We propose a novel camera calibration method using a collimator system. The entire calibration process can be implemented with multiple close-range observations and does not require the known direction of collimated rays. (2) We prove that the relative motion between the calibration target and the camera conforms to the spherical motion model. This constraint effectively reduces the number of motion parameters to be estimated. (3) A closed-form solver for multiple views and a minimal solver for two views are proposed based on the spherical motion constraint. Moreover, the degenerate case of the proposed solvers is proved." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 618, + 238, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 618, + 238, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 618, + 238, + 631 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "Camera calibration is a crucial subject in 3D vision and continues to be actively researched. The literature on camera calibration introduces a wide range of al" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 265, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 265, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 265, + 102 + ], + "type": "text", + "content": "S. Liang, B. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": "gorithms and toolboxes [3-5,22,26,33,35,46]. To ensure the highest accuracy for calibration, targets with known structures are usually used for camera calibration. Various calibration targets have been proposed, including 3D targets [37], 2D targets [13,18] and 1D targets [47]. In some practical applications where no calibration target is available, the camera calibration can be performed by establishing the correspondence between images during the movement of the camera, which is known as self-calibration [8,9,16,19,28,31]. Among these options, the target-based close-range calibration method is widely used." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 480, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 480, + 319 + ], + "type": "text", + "content": "In outdoor long-range measurement applications [10, 45], the camera has a long working distance to measure the long-distance targets. Thus, the close-range calibration methods are impractical due to the site limitation. A commonly employed approach involves constructing a calibration field, which serves as a known structure for the calibration [21, 25, 29, 34, 40, 42]. Oniga et al. [29] lay markers on buildings to calibrate cameras mounted on UAV. Shang et al. [34] proposed a high-precision calibration method by placing several markers near the ground. Additionally, Klaus et al. [23] proposed using stars as distant targets to estimate camera parameters." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 319, + 481, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 481, + 450 + ], + "type": "text", + "content": "The collimator is also commonly used to provide 2D-3D correspondence for camera calibration in the laboratory [2, 20, 41, 44]. Previous collimator-based methods proposed to use goniometer or theodolite to obtain the direction of collimated rays and then estimate the camera parameters [41, 43]. The collimator-based method offers both accuracy and flexibility in the limited space, but a high-precision goniometer or theodolite is expensive. To address this, Yuan et al. [44] designed a multiple pinhole mask for the collimator to generate the collimated rays with known directions. However, the accuracy of calibration is sensitive to machining errors. Bauer et al. [2] proposed a calibration method using a diffractive optical element (DOE). The angle between the laser beam and the camera can then be estimated to determine the camera parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 470, + 264, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 470, + 264, + 484 + ], + "spans": [ + { + "bbox": [ + 132, + 470, + 264, + 484 + ], + "type": "text", + "content": "3 Collimator System" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 495, + 479, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 479, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 479, + 531 + ], + "type": "text", + "content": "This section introduces the designed collimator system and shows that the equivalent relative motion between the calibration target and the camera conforms to a spherical motion model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 549, + 309, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 549, + 309, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 549, + 309, + 562 + ], + "type": "text", + "content": "3.1 Scheme of Collimator System" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "content": "The collimator is an optical instrument that can produce collimated rays and plays an important role when adjusting other optical instruments. A collimator consists of essential components such as a light source, a reticle, and an optical lens, as illustrated in Fig. 2. The reticle is a thin glass located on the focal plane of the collimator. The generation of collimated rays can be regarded as the inverse process of imaging a target at infinity onto the focal plane. Specifically, a point on the lens's focal plane will generate collimated rays, which can be regarded as the target at infinity. Therefore, cameras with varying focal lengths" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "text", + "content": "Camera Calibration using a Collimator System" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "can consistently and accurately observe the target through the collimator at a close range." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 284 + ], + "type": "text", + "content": "In this paper, we design a collimator system for camera calibration. A star-based pattern, proposed by Thomas et al. [33], is attached to the reticle of the collimator as the calibration target. The star-based pattern offers richer gradient information for improving feature extraction accuracy. Similar to the plane-based calibration method [35, 46], the star-based pattern provides 2D-3D correspondences for camera calibration. Our collimator system is portable with a size of about " + }, + { + "bbox": [ + 130, + 140, + 482, + 284 + ], + "type": "inline_equation", + "content": "200\\mathrm{mm} \\times 170\\mathrm{mm} \\times 300\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 140, + 482, + 284 + ], + "type": "text", + "content": ". The collimator system is designed to provide a reliable and controllable calibration environment for some special scenarios like inappropriate illumination, limited space, and long-range cameras. However, printed targets face challenges in these scenarios. When calibrating a camera, we only need to capture the target from multiple orientations through the lens of the collimator, which is as easy as the printed targets method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 300, + 283, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 283, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 283, + 312 + ], + "type": "text", + "content": "3.2 Spherical Motion Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 299, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 299, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 299, + 545 + ], + "type": "text", + "content": "In this section, we prove that the relative motion between the calibration target and the camera conforms to the spherical motion model. Firstly, we briefly analyze the geometric property of the collimator. As shown in Fig. 2, let us note two arbitrary points " + }, + { + "bbox": [ + 130, + 319, + 299, + 545 + ], + "type": "inline_equation", + "content": "P_{1}" + }, + { + "bbox": [ + 130, + 319, + 299, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 319, + 299, + 545 + ], + "type": "inline_equation", + "content": "P_{2}" + }, + { + "bbox": [ + 130, + 319, + 299, + 545 + ], + "type": "text", + "content": " on the reticle of the collimator. Rays emitted from " + }, + { + "bbox": [ + 130, + 319, + 299, + 545 + ], + "type": "inline_equation", + "content": "P_{1}" + }, + { + "bbox": [ + 130, + 319, + 299, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 319, + 299, + 545 + ], + "type": "inline_equation", + "content": "P_{2}" + }, + { + "bbox": [ + 130, + 319, + 299, + 545 + ], + "type": "text", + "content": " pass through the collimator lens, generating two parallel beams observed by cameras denoted as " + }, + { + "bbox": [ + 130, + 319, + 299, + 545 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 130, + 319, + 299, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 319, + 299, + 545 + ], + "type": "inline_equation", + "content": "C_2" + }, + { + "bbox": [ + 130, + 319, + 299, + 545 + ], + "type": "text", + "content": " with arbitrary poses. The propagation routes of the two beams are visually distinguished by red and blue lines in Fig. 2. Obviously, the angle between two lines observed by the two cameras is equal, that is, " + }, + { + "bbox": [ + 130, + 319, + 299, + 545 + ], + "type": "inline_equation", + "content": "\\theta_{1}\\equiv \\theta_{2}\\equiv \\theta" + }, + { + "bbox": [ + 130, + 319, + 299, + 545 + ], + "type": "text", + "content": ". We can obtain a property that the two rays" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 312, + 341, + 472, + 462 + ], + "blocks": [ + { + "bbox": [ + 312, + 341, + 472, + 462 + ], + "lines": [ + { + "bbox": [ + 312, + 341, + 472, + 462 + ], + "spans": [ + { + "bbox": [ + 312, + 341, + 472, + 462 + ], + "type": "image", + "image_path": "41ab016be5e74375c3a1a1ead198fed0102ad5984f17748cf004b8ec92ebd71f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 469, + 482, + 514 + ], + "lines": [ + { + "bbox": [ + 304, + 469, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 482, + 514 + ], + "type": "text", + "content": "Fig. 2: Diagram of collimator system. The geometric property of the collimator system leads to the angle invariance of any point pair on the reticle." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 546, + 480, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 480, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 480, + 570 + ], + "type": "text", + "content": "observed by the cameras always maintain a constant angle, regardless of the poses of observing cameras. We refer to this property as the angle invariance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Next, we show that this setup, where a camera moves around a collimator system and captures images, is theoretically equivalent to the capture setting where the camera is fixed, but the images are moving around the camera spherically. This can be derived from the angle invariance. For convenience in introducing our derivation process, we treat the camera as a fixed reference coordinate, and assume that the calibration target moves relative to the camera in space. The angle invariance allows us to infer that when the target performs a specific motion in space, the angular distance between any point pair on the target always" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "text", + "content": "S. Liang, B. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "remains constant. The angular distance means the angle between two vectors from the origin (camera optical center) to two points. By arbitrarily selecting a pair of points " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_j" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " on the target and transforming them, we can establish the following equation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 218, + 171, + 481, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 171, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 218, + 171, + 481, + 184 + ], + "type": "interline_equation", + "content": "\\angle \\left(\\mathbf {P} _ {i}, \\mathbf {P} _ {j}\\right) = \\angle \\left(\\mathbf {R} \\left(\\mathbf {P} _ {i} + \\mathbf {t}\\right), \\mathbf {R} \\left(\\mathbf {P} _ {j} + \\mathbf {t}\\right)\\right), \\tag {1}", + "image_path": "19701140bdcd73b9a4989b6cb10bfda656a3e446e796698a6c6228edb2b2982b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 190, + 480, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 480, + 308 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 190, + 480, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 130, + 190, + 480, + 308 + ], + "type": "text", + "content": " is rotation matrix and " + }, + { + "bbox": [ + 130, + 190, + 480, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 130, + 190, + 480, + 308 + ], + "type": "text", + "content": " is translation vector. From the rotation invariance [39] that a pair of 3D points are jointly rotated about the origin, their angular distance remains constant. We have " + }, + { + "bbox": [ + 130, + 190, + 480, + 308 + ], + "type": "inline_equation", + "content": "\\angle (\\mathbf{R}(\\mathbf{P}_i + \\mathbf{t}),\\mathbf{R}(\\mathbf{P}_j + \\mathbf{t})) = \\angle (\\mathbf{P}_i + \\mathbf{t},\\mathbf{P}_j + \\mathbf{t})" + }, + { + "bbox": [ + 130, + 190, + 480, + 308 + ], + "type": "text", + "content": ". Thus, the angle invariance is maintained when the calibration target is arbitrarily rotated about the origin. To make " + }, + { + "bbox": [ + 130, + 190, + 480, + 308 + ], + "type": "inline_equation", + "content": "\\angle (\\mathbf{P}_i,\\mathbf{P}_j) = \\angle (\\mathbf{P}_i + \\mathbf{t},\\mathbf{P}_j + \\mathbf{t})" + }, + { + "bbox": [ + 130, + 190, + 480, + 308 + ], + "type": "text", + "content": " hold for any pair of points on the target, the translation vector must be a " + }, + { + "bbox": [ + 130, + 190, + 480, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{0}" + }, + { + "bbox": [ + 130, + 190, + 480, + 308 + ], + "type": "text", + "content": " vector " + }, + { + "bbox": [ + 130, + 190, + 480, + 308 + ], + "type": "inline_equation", + "content": "(\\mathbf{t} = [0,0,0]^T)" + }, + { + "bbox": [ + 130, + 190, + 480, + 308 + ], + "type": "text", + "content": ". The proof is given in the supplementary material. Therefore, the calibration target performs pure rotation motion with respect to the origin (i.e. camera optical center). Such pure rotational motion is a typical spherical motion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 309, + 480, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 480, + 381 + ], + "type": "text", + "content": "The spherical motion can be visualized in Fig. 1. In the spherical motion model, the calibration target exhibits only rotational motion with a fixed distance relative to the camera. We establish a local coordinate on the target, where the " + }, + { + "bbox": [ + 130, + 309, + 480, + 381 + ], + "type": "inline_equation", + "content": "z'" + }, + { + "bbox": [ + 130, + 309, + 480, + 381 + ], + "type": "text", + "content": "-axis aligns with the ray between the camera coordinate system " + }, + { + "bbox": [ + 130, + 309, + 480, + 381 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 309, + 480, + 381 + ], + "type": "text", + "content": " and calibration target coordinate system " + }, + { + "bbox": [ + 130, + 309, + 480, + 381 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 130, + 309, + 480, + 381 + ], + "type": "text", + "content": ". Inspired by [38], the relative pose between the calibration target and camera can be mathematically represented as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 261, + 388, + 481, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 388, + 481, + 412 + ], + "spans": [ + { + "bbox": [ + 261, + 388, + 481, + 412 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {c p ^ {\\prime}} = \\left( \\begin{array}{l l} \\mathbf {R} _ {c p ^ {\\prime}} & \\mathbf {t} _ {c p ^ {\\prime}} \\\\ \\mathbf {0} _ {1 \\times 3} & 1 \\end{array} \\right), \\tag {2}", + "image_path": "450797cac3e63f89cdc1bdb98e66b6121bbe8673387b29d99f7cfbc44c603a3d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 418, + 479, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 479, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 479, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 418, + 479, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{cp'} \\in SO(3)" + }, + { + "bbox": [ + 130, + 418, + 479, + 479 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 418, + 479, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{cp'} = (0,0,-r)^T" + }, + { + "bbox": [ + 130, + 418, + 479, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 418, + 479, + 479 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 418, + 479, + 479 + ], + "type": "text", + "content": " is the radius of spherical motion. We usually establish the origin of the coordinate system " + }, + { + "bbox": [ + 130, + 418, + 479, + 479 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 418, + 479, + 479 + ], + "type": "text", + "content": " on the upper left corner of the calibration pattern. Therefore, there is a relative translation " + }, + { + "bbox": [ + 130, + 418, + 479, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{p'p} = (x,y,0)" + }, + { + "bbox": [ + 130, + 418, + 479, + 479 + ], + "type": "text", + "content": " between the coordinate systems " + }, + { + "bbox": [ + 130, + 418, + 479, + 479 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 130, + 418, + 479, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 418, + 479, + 479 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 418, + 479, + 479 + ], + "type": "text", + "content": ". Furthermore, the transformation matrix between camera " + }, + { + "bbox": [ + 130, + 418, + 479, + 479 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 418, + 479, + 479 + ], + "type": "text", + "content": " and calibration target " + }, + { + "bbox": [ + 130, + 418, + 479, + 479 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 418, + 479, + 479 + ], + "type": "text", + "content": " is expressed as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 177, + 486, + 481, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 486, + 481, + 513 + ], + "spans": [ + { + "bbox": [ + 177, + 486, + 481, + 513 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {p c} = \\left[ \\left( \\begin{array}{l l} \\mathbf {R} _ {c p ^ {\\prime}} & \\mathbf {t} _ {c p ^ {\\prime}} \\\\ \\mathbf {0} _ {1 \\times 3} & 1 \\end{array} \\right) \\left( \\begin{array}{l l} \\mathbf {I} _ {3 \\times 3} & \\mathbf {t} _ {p ^ {\\prime} p} \\\\ \\mathbf {0} _ {1 \\times 3} & 1 \\end{array} \\right) \\right] ^ {- 1} = \\left( \\begin{array}{l l} \\mathbf {R} _ {p c} & - \\mathbf {R} _ {p c} \\mathbf {t} _ {c p} \\\\ \\mathbf {0} _ {1 \\times 3} & 1 \\end{array} \\right), \\tag {3}", + "image_path": "b25483e7d05637a27d86d82cac53fce5fdf29e7afff1aaf74e54e0e818056e34.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 518, + 479, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 518, + 479, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 518, + 479, + 578 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 518, + 479, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{pc} = \\mathbf{R}_{cp'}^T" + }, + { + "bbox": [ + 130, + 518, + 479, + 578 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 518, + 479, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{cp} = \\mathbf{t}_{p'p} + \\mathbf{t}_{cp'} = (x,y, -r)^T" + }, + { + "bbox": [ + 130, + 518, + 479, + 578 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 518, + 479, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{cp}" + }, + { + "bbox": [ + 130, + 518, + 479, + 578 + ], + "type": "text", + "content": " determines the position of the camera in the calibration target coordinate system. Note that " + }, + { + "bbox": [ + 130, + 518, + 479, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{cp}" + }, + { + "bbox": [ + 130, + 518, + 479, + 578 + ], + "type": "text", + "content": " is a fixed value and has nothing to do with the pose of the calibration target. Given the constraint from the spherical motion, the original 6DOF general motion is limited to a 3DOF pure rotation motion." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 594, + 272, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 272, + 607 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 272, + 607 + ], + "type": "text", + "content": "4 Calibration Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": "The spherical motion constraint effectively reduces the number of motion parameters to be estimated. In this section, we show how to calibrate camera parameters using a set of collimator images. A closed-form solver for multiple views and a minimal solver for two views are proposed." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "text", + "content": "Camera Calibration using a Collimator System" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 276, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 276, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 276, + 126 + ], + "type": "text", + "content": "4.1 Geometric Constraints" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 159 + ], + "type": "text", + "content": "We use the well-known pinhole model to model cameras. The mapping from a 3D point " + }, + { + "bbox": [ + 130, + 134, + 482, + 159 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = (X,Y,Z)^T" + }, + { + "bbox": [ + 130, + 134, + 482, + 159 + ], + "type": "text", + "content": " to an image point " + }, + { + "bbox": [ + 130, + 134, + 482, + 159 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = (u,v)^T" + }, + { + "bbox": [ + 130, + 134, + 482, + 159 + ], + "type": "text", + "content": " is computed as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 219, + 165, + 481, + 204 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 165, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 219, + 165, + 481, + 204 + ], + "type": "interline_equation", + "content": "s \\widetilde {\\mathbf {p}} = \\mathbf {K} [ \\mathbf {R} \\mid \\mathbf {t} ] \\widetilde {\\mathbf {P}}, \\text {w i t h} \\mathbf {K} = \\left[ \\begin{array}{l l l} f _ {x} & \\gamma & c _ {x} \\\\ 0 & f _ {y} & c _ {y} \\\\ 0 & 0 & 1 \\end{array} \\right], \\tag {4}", + "image_path": "c79d3081f7a63e20a100f8633acb8803b0774bff5161179e7da95d78bb3fee4e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 210, + 479, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 210, + 479, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 210, + 479, + 271 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 210, + 479, + 271 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 210, + 479, + 271 + ], + "type": "text", + "content": " is an arbitrary scale factor, " + }, + { + "bbox": [ + 130, + 210, + 479, + 271 + ], + "type": "inline_equation", + "content": "(\\mathbf{R},\\mathbf{t})" + }, + { + "bbox": [ + 130, + 210, + 479, + 271 + ], + "type": "text", + "content": " is the camera pose with respect to the calibration target coordinate, and " + }, + { + "bbox": [ + 130, + 210, + 479, + 271 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 210, + 479, + 271 + ], + "type": "text", + "content": " is the camera intrinsic matrix. " + }, + { + "bbox": [ + 130, + 210, + 479, + 271 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{p}} = [u,v,1]^T" + }, + { + "bbox": [ + 130, + 210, + 479, + 271 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 210, + 479, + 271 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{P}} = [X,Y,Z,1]^T" + }, + { + "bbox": [ + 130, + 210, + 479, + 271 + ], + "type": "text", + "content": " denote the homogeneous coordinate forms of points " + }, + { + "bbox": [ + 130, + 210, + 479, + 271 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 210, + 479, + 271 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 210, + 479, + 271 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 130, + 210, + 479, + 271 + ], + "type": "text", + "content": ", respectively. The intrinsic matrix " + }, + { + "bbox": [ + 130, + 210, + 479, + 271 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 210, + 479, + 271 + ], + "type": "text", + "content": " consists of five parameters: the focal length " + }, + { + "bbox": [ + 130, + 210, + 479, + 271 + ], + "type": "inline_equation", + "content": "(f_x,f_y)" + }, + { + "bbox": [ + 130, + 210, + 479, + 271 + ], + "type": "text", + "content": ", the principal point " + }, + { + "bbox": [ + 130, + 210, + 479, + 271 + ], + "type": "inline_equation", + "content": "(c_x,c_y)" + }, + { + "bbox": [ + 130, + 210, + 479, + 271 + ], + "type": "text", + "content": " and skew factor " + }, + { + "bbox": [ + 130, + 210, + 479, + 271 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 210, + 479, + 271 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 271, + 480, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 480, + 331 + ], + "type": "text", + "content": "Without loss of generality, we assume the planar target is on the plane " + }, + { + "bbox": [ + 130, + 271, + 480, + 331 + ], + "type": "inline_equation", + "content": "Z = 0" + }, + { + "bbox": [ + 130, + 271, + 480, + 331 + ], + "type": "text", + "content": " of the world coordinate system. The transformation between the image plane and calibration target plane can be constrained by a " + }, + { + "bbox": [ + 130, + 271, + 480, + 331 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 271, + 480, + 331 + ], + "type": "text", + "content": " homography matrix, denoted as " + }, + { + "bbox": [ + 130, + 271, + 480, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 130, + 271, + 480, + 331 + ], + "type": "text", + "content": ". For perspective projection, the homography matrix " + }, + { + "bbox": [ + 130, + 271, + 480, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 130, + 271, + 480, + 331 + ], + "type": "text", + "content": " depends on the image pose " + }, + { + "bbox": [ + 130, + 271, + 480, + 331 + ], + "type": "inline_equation", + "content": "(\\mathbf{R},\\mathbf{t})" + }, + { + "bbox": [ + 130, + 271, + 480, + 331 + ], + "type": "text", + "content": " and intrinsic matrix " + }, + { + "bbox": [ + 130, + 271, + 480, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 271, + 480, + 331 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 234, + 338, + 480, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 338, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 234, + 338, + 480, + 353 + ], + "type": "interline_equation", + "content": "\\mathbf {H} = \\left[ \\begin{array}{l l l} \\mathbf {h} _ {1} & \\mathbf {h} _ {2} & \\mathbf {h} _ {3} \\end{array} \\right] = \\lambda \\mathbf {K} \\left[ \\begin{array}{l l l} \\mathbf {r} _ {1} & \\mathbf {r} _ {2} & \\mathbf {t} \\end{array} \\right], \\tag {5}", + "image_path": "7fd2bf2f9a4babb5c42169c46112c4f29b504e62ee6cb9cf92fbe2bd9eb72205.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "text", + "content": " is an arbitrary scalar, " + }, + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_i" + }, + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "text", + "content": "-th columns of " + }, + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "text", + "content": ". Given an image of the calibration target, the homography matrix " + }, + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 130, + 359, + 479, + 396 + ], + "type": "text", + "content": " can be estimated using the plane point correspondences. Combining Eq. (3) and Eq. (5), we can get" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 236, + 403, + 480, + 424 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 236, + 403, + 480, + 424 + ], + "type": "interline_equation", + "content": "\\mathbf {M} = \\left[ \\mathbf {r} _ {1} \\mathbf {r} _ {2} - \\mathbf {R t} _ {c p} \\right] = \\frac {1}{\\lambda} \\mathbf {K} ^ {- 1} \\mathbf {H}. \\tag {6}", + "image_path": "b61f51bab9376108d88780a1208ab897d084278767233baa3eff40b81e97b89e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 430, + 405, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 405, + 442 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 405, + 442 + ], + "type": "text", + "content": "Using the knowledge that " + }, + { + "bbox": [ + 132, + 430, + 405, + 442 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 132, + 430, + 405, + 442 + ], + "type": "text", + "content": " is an orthogonal matrix, we have" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 222, + 449, + 480, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 449, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 222, + 449, + 480, + 487 + ], + "type": "interline_equation", + "content": "\\mathbf {H} ^ {T} \\mathbf {K} ^ {- T} \\mathbf {K} ^ {- 1} \\mathbf {H} = \\lambda^ {2} \\left[ \\begin{array}{c c c} 1 & 0 & - x \\\\ 0 & 1 & - y \\\\ - x & - y & \\| \\mathbf {t} _ {c p} \\| ^ {2} \\end{array} \\right]. \\tag {7}", + "image_path": "81691d6642ceeb513ea1023de990631e9d2a671beec2ea8c75ac2c0870f9f45b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 494, + 479, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 494, + 479, + 554 + ], + "spans": [ + { + "bbox": [ + 131, + 494, + 479, + 554 + ], + "type": "text", + "content": "There are 8 parameters to be estimated, 5 for intrinsic matrix and 3 for " + }, + { + "bbox": [ + 131, + 494, + 479, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{cp} = (x,y,-r)^T" + }, + { + "bbox": [ + 131, + 494, + 479, + 554 + ], + "type": "text", + "content": ". For each given image, Eq. (7) provides 5 independent constraints. Theoretically, a minimum of two images can completely solve all unknown parameters. In the next two subsections, we give a closed-form solver for " + }, + { + "bbox": [ + 131, + 494, + 479, + 554 + ], + "type": "inline_equation", + "content": "N(N > 2)" + }, + { + "bbox": [ + 131, + 494, + 479, + 554 + ], + "type": "text", + "content": " collimator images and a minimal solver for two collimator images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 570, + 354, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 570, + 354, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 570, + 354, + 582 + ], + "type": "text", + "content": "4.2 Closed-form Solver for Multiple Views" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 588, + 479, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 588, + 479, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 588, + 479, + 636 + ], + "type": "text", + "content": "When there are more than two collimator images for calibration, we model the problem as a linear system and propose a closed-form solution. Suppose that the homography matrices " + }, + { + "bbox": [ + 130, + 588, + 479, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_i(i = 1,\\dots,N)" + }, + { + "bbox": [ + 130, + 588, + 479, + 636 + ], + "type": "text", + "content": " have been estimated for all images. The transformations between images " + }, + { + "bbox": [ + 130, + 588, + 479, + 636 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 588, + 479, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 588, + 479, + 636 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 588, + 479, + 636 + ], + "type": "text", + "content": " can be expressed as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 258, + 643, + 480, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 643, + 480, + 668 + ], + "spans": [ + { + "bbox": [ + 258, + 643, + 480, + 668 + ], + "type": "interline_equation", + "content": "\\mathbf {H} _ {j} ^ {- 1} \\mathbf {H} _ {i} = \\frac {\\lambda_ {i}}{\\lambda_ {j}} \\mathbf {M} _ {j} ^ {- 1} \\mathbf {M} _ {i}. \\tag {8}", + "image_path": "9ad8faa6c03b5af795b628fbfe8066e6d8b84c0844ae588f09f9b117384dfe2b.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "text", + "content": "S. Liang, B. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 438, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 438, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 438, + 128 + ], + "type": "text", + "content": "For the matrix " + }, + { + "bbox": [ + 132, + 116, + 438, + 128 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 132, + 116, + 438, + 128 + ], + "type": "text", + "content": " of each image, we can calculate the determinant as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 192, + 137, + 480, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 137, + 480, + 166 + ], + "spans": [ + { + "bbox": [ + 192, + 137, + 480, + 166 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left| \\mathbf {M} \\right| = \\left| \\mathbf {r} _ {1} \\mathbf {r} _ {2} - \\mathbf {R t} _ {c p} \\right| = \\left| \\mathbf {r} _ {1} \\mathbf {r} _ {2} - x \\mathbf {r} _ {1} - y \\mathbf {r} _ {2} + r \\mathbf {r} _ {3} \\right| \\tag {9} \\\\ = - x \\left| \\mathbf {r} _ {1} \\mathbf {r} _ {2} \\mathbf {r} _ {1} \\right| - y \\left| \\mathbf {r} _ {1} \\mathbf {r} _ {2} \\mathbf {r} _ {2} \\right| + r \\left| \\mathbf {r} _ {1} \\mathbf {r} _ {2} \\mathbf {r} _ {3} \\right| = r, \\\\ \\end{array}", + "image_path": "8fab7d83e41279c4cabe4e1a9833b2b0f21541e83285af7622b6c4b85a46dc25.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 174, + 479, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 174, + 479, + 247 + ], + "spans": [ + { + "bbox": [ + 131, + 174, + 479, + 247 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 174, + 479, + 247 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 131, + 174, + 479, + 247 + ], + "type": "text", + "content": " is a non-zero fixed value. Using the fact that " + }, + { + "bbox": [ + 131, + 174, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\operatorname*{det}(\\mathbf{M}_j^{-1}\\mathbf{M}_i) = 1" + }, + { + "bbox": [ + 131, + 174, + 479, + 247 + ], + "type": "text", + "content": ", we have " + }, + { + "bbox": [ + 131, + 174, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\lambda_{ij} = \\lambda_i / \\lambda_j = \\sqrt[3]{|\\mathbf{H}_j^{-1}\\mathbf{H}_i|}" + }, + { + "bbox": [ + 131, + 174, + 479, + 247 + ], + "type": "text", + "content": ". Then, we choose the image with the most feature points observed as the base image. For convenience, here the first image is selected as the base image and we get " + }, + { + "bbox": [ + 131, + 174, + 479, + 247 + ], + "type": "inline_equation", + "content": "\\lambda_{i1} = \\sqrt[3]{|\\mathbf{H}_j^{-1}\\mathbf{H}_1|}" + }, + { + "bbox": [ + 131, + 174, + 479, + 247 + ], + "type": "text", + "content": ". By inverting Eq. (7) of image " + }, + { + "bbox": [ + 131, + 174, + 479, + 247 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 174, + 479, + 247 + ], + "type": "text", + "content": ", we get" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 206, + 255, + 479, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 255, + 479, + 293 + ], + "spans": [ + { + "bbox": [ + 206, + 255, + 479, + 293 + ], + "type": "interline_equation", + "content": "\\mathbf {H} _ {i} ^ {- 1} \\mathbf {K} \\mathbf {K} ^ {T} \\mathbf {H} _ {i} ^ {- T} = \\frac {1}{\\lambda_ {i} ^ {2} r ^ {2}} \\left[ \\begin{array}{c c c} r ^ {2} + x ^ {2} & x y & x \\\\ x y & r ^ {2} + y ^ {2} & y \\\\ x & y & 1 \\end{array} \\right]. \\tag {10}", + "image_path": "f7cd7be19cdabf8d7ad19a54657ad3accb67a5a5f825f91ce159cbd8d9cf48f8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 300, + 323, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 323, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 323, + 313 + ], + "type": "text", + "content": "Substitute " + }, + { + "bbox": [ + 132, + 300, + 323, + 313 + ], + "type": "inline_equation", + "content": "\\lambda_{i} = \\lambda_{i1}\\lambda_{1}" + }, + { + "bbox": [ + 132, + 300, + 323, + 313 + ], + "type": "text", + "content": " into Eq. (10), we get" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 177, + 320, + 479, + 367 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 320, + 479, + 367 + ], + "spans": [ + { + "bbox": [ + 177, + 320, + 479, + 367 + ], + "type": "interline_equation", + "content": "\\mathbf {H} _ {i} ^ {- 1} \\mathbf {K K} ^ {T} \\mathbf {H} _ {i} ^ {- T} = \\frac {1}{\\lambda_ {i 1} ^ {2}} \\mathbf {A}, \\text {w i t h} \\mathbf {A} = \\left[ \\begin{array}{c c c} \\frac {r ^ {2} + x ^ {2}}{(\\lambda_ {1} r) ^ {2}} & \\frac {x y}{(\\lambda_ {1} r) ^ {2}} & \\frac {x}{(\\lambda_ {1} r) ^ {2}} \\\\ \\frac {x y}{(\\lambda_ {1} r) ^ {2}} & \\frac {r ^ {2} + y ^ {2}}{(\\lambda_ {1} r) ^ {2}} & \\frac {y}{(\\lambda_ {1} r) ^ {2}} \\\\ \\frac {x}{(\\lambda_ {1} r) ^ {2}} & \\frac {y}{(\\lambda_ {1} r) ^ {2}} & \\frac {1}{(\\lambda_ {1} r) ^ {2}} \\end{array} \\right]. \\tag {11}", + "image_path": "bef9fabde28a30a84461b781be957136b47de9394609656e661f8ea5c0c38834.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 374, + 349, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 349, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 349, + 385 + ], + "type": "text", + "content": "Let us define a 6-vector " + }, + { + "bbox": [ + 132, + 374, + 349, + 385 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 132, + 374, + 349, + 385 + ], + "type": "text", + "content": " to represent matrix " + }, + { + "bbox": [ + 132, + 374, + 349, + 385 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 132, + 374, + 349, + 385 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 228, + 394, + 384, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 394, + 384, + 407 + ], + "spans": [ + { + "bbox": [ + 228, + 394, + 384, + 407 + ], + "type": "interline_equation", + "content": "\\mathbf {a} = \\left(A _ {1 1}, A _ {1 2}, A _ {1 3}, A _ {2 2}, A _ {2 3}, A _ {3 3}\\right) ^ {T},", + "image_path": "924fdc2f5698619ecde2c1da605e86faa48cb3702074413a332270ee0358f573.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 416, + 444, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 416, + 444, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 416, + 444, + 427 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 416, + 444, + 427 + ], + "type": "inline_equation", + "content": "A_{mn}" + }, + { + "bbox": [ + 132, + 416, + 444, + 427 + ], + "type": "text", + "content": " is the value of " + }, + { + "bbox": [ + 132, + 416, + 444, + 427 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 132, + 416, + 444, + 427 + ], + "type": "text", + "content": "-th row and " + }, + { + "bbox": [ + 132, + 416, + 444, + 427 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 132, + 416, + 444, + 427 + ], + "type": "text", + "content": "-th column of matrix " + }, + { + "bbox": [ + 132, + 416, + 444, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 132, + 416, + 444, + 427 + ], + "type": "text", + "content": ". Let" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 220, + 435, + 479, + 474 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 435, + 479, + 474 + ], + "spans": [ + { + "bbox": [ + 220, + 435, + 479, + 474 + ], + "type": "interline_equation", + "content": "\\mathbf {W} = \\left[ \\begin{array}{c c c} c _ {x} ^ {2} + f _ {x} ^ {2} + \\gamma^ {2} & c _ {x} c _ {y} + f _ {y} \\gamma & c _ {x} \\\\ c _ {x} c _ {y} + f _ {y} \\gamma & c _ {y} ^ {2} + f _ {y} ^ {2} & c _ {y} \\\\ c _ {x} & c _ {y} & 1 \\end{array} \\right]. \\tag {12}", + "image_path": "49529a729c87802683f1b44d9a2a3c6359ad4ce08f9e8048f9bcb23a48658ac2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 481, + 440, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 481, + 440, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 481, + 440, + 493 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 132, + 481, + 440, + 493 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 132, + 481, + 440, + 493 + ], + "type": "text", + "content": " is a symmetric matrix and can be defined by a 5-vector:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 232, + 501, + 381, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 501, + 381, + 514 + ], + "spans": [ + { + "bbox": [ + 232, + 501, + 381, + 514 + ], + "type": "interline_equation", + "content": "\\mathbf {w} = \\left(W _ {1 1}, W _ {1 2}, W _ {1 3}, W _ {2 2}, W _ {2 3}\\right) ^ {T},", + "image_path": "635b2dafecff955f06a3125b61e3f5fdc299f86b09cd8aa0d9f403b9179f8955.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 522, + 479, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 479, + 570 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 479, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 522, + 479, + 570 + ], + "type": "inline_equation", + "content": "W_{mn}" + }, + { + "bbox": [ + 132, + 522, + 479, + 570 + ], + "type": "text", + "content": " is the value of " + }, + { + "bbox": [ + 132, + 522, + 479, + 570 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 132, + 522, + 479, + 570 + ], + "type": "text", + "content": "-th row and " + }, + { + "bbox": [ + 132, + 522, + 479, + 570 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 132, + 522, + 479, + 570 + ], + "type": "text", + "content": "-th column of matrix " + }, + { + "bbox": [ + 132, + 522, + 479, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 132, + 522, + 479, + 570 + ], + "type": "text", + "content": ". Obviously, Eq. (11) represents a linear system of equations involving the unknowns " + }, + { + "bbox": [ + 132, + 522, + 479, + 570 + ], + "type": "inline_equation", + "content": "[\\mathbf{w}^T, \\mathbf{a}^T]^T" + }, + { + "bbox": [ + 132, + 522, + 479, + 570 + ], + "type": "text", + "content": ". Combining Eqs. (11) and (12), we can get the following equations for each image:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 253, + 570, + 479, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 570, + 479, + 594 + ], + "spans": [ + { + "bbox": [ + 253, + 570, + 479, + 594 + ], + "type": "interline_equation", + "content": "\\left[ \\mathbf {V} - \\lambda_ {i 1} ^ {2} \\mathbf {I} _ {5 \\times 5} \\right] \\left[ \\begin{array}{l} \\mathbf {w} \\\\ \\mathbf {a} \\end{array} \\right] = \\mathbf {b}, \\tag {13}", + "image_path": "0221ea10094ce41387488f2246ebadc315e2241d081a311f4fb0cb4c5a2bbfad.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 598, + 154, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 154, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 154, + 609 + ], + "type": "text", + "content": "with" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 189, + 608, + 316, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 608, + 316, + 620 + ], + "spans": [ + { + "bbox": [ + 189, + 608, + 316, + 620 + ], + "type": "interline_equation", + "content": "\\mathbf {V} = \\left[ \\mathbf {v} _ {1 1}, \\mathbf {v} _ {1 2}, \\mathbf {v} _ {1 3}, \\mathbf {v} _ {2 2}, \\mathbf {v} _ {2 3} \\right] ^ {T},", + "image_path": "0b2a9847b0c6489a4cf268148c812f292dd9158bb94d13b34ffb0a6630d9c4ff.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 189, + 624, + 422, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 624, + 422, + 635 + ], + "spans": [ + { + "bbox": [ + 189, + 624, + 422, + 635 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {m n} = \\left[ h _ {m 1} h _ {n 1}, h _ {m 1} h _ {n 2} + h _ {m 2} h _ {n 1}, h _ {m 1} h _ {n 3} + h _ {m 3} h _ {n 1}, \\right.", + "image_path": "cf3e8954be4615568569441ec6847298a637ff4b7ddc8729e88367ee0276520c.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 219, + 639, + 340, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 639, + 340, + 651 + ], + "spans": [ + { + "bbox": [ + 219, + 639, + 340, + 651 + ], + "type": "interline_equation", + "content": "\\left. h _ {m 2} h _ {n 2}, h _ {m 2} h _ {n 3} + h _ {m 3} h _ {n 2} \\right],", + "image_path": "7ce3125fc78cb706aaaa02a743c4fe4d6e5aec6f1b3ca4f7aac891aa6b995a3c.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 189, + 654, + 391, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 654, + 391, + 667 + ], + "spans": [ + { + "bbox": [ + 189, + 654, + 391, + 667 + ], + "type": "interline_equation", + "content": "\\mathbf {b} = - \\left[ h _ {1 3} h _ {1 3}, h _ {1 3} h _ {2 3}, h _ {1 3} h _ {3 3}, h _ {2 3} h _ {2 3}, h _ {2 3} h _ {3 3} \\right] ^ {T},", + "image_path": "2ecdd9c46ce312c52c05483393428b334111517cb240c31add2d09f470cd3fb2.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "text", + "content": "Camera Calibration using a Collimator System" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "inline_equation", + "content": "h_{mn}" + }, + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": " denotes the value of " + }, + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "-th row and " + }, + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "-th column of matrix " + }, + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^{-1}" + }, + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": ". Given " + }, + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "inline_equation", + "content": "N(N > 2)" + }, + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": " collimator images, we can stack " + }, + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": " such equations as Eq. (13) together in matrix form:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 279, + 150, + 481, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 150, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 279, + 150, + 481, + 175 + ], + "type": "interline_equation", + "content": "\\mathbf {D} \\left[ \\begin{array}{l} \\mathbf {w} \\\\ \\mathbf {a} \\end{array} \\right] = \\mathbf {b}, \\tag {14}", + "image_path": "241e853b22b2bf7b97cc2479790739e9f06e3757738e61d09c5f05238ae6d5b2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 177, + 478, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 177, + 478, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 177, + 478, + 190 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 177, + 478, + 190 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 132, + 177, + 478, + 190 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 132, + 177, + 478, + 190 + ], + "type": "inline_equation", + "content": "5N\\times 11" + }, + { + "bbox": [ + 132, + 177, + 478, + 190 + ], + "type": "text", + "content": " matrix. The closed-form solution of Eq. (14) is given by" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 257, + 196, + 481, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 196, + 481, + 222 + ], + "spans": [ + { + "bbox": [ + 257, + 196, + 481, + 222 + ], + "type": "interline_equation", + "content": "\\left[ \\begin{array}{l} \\mathbf {w} \\\\ \\mathbf {a} \\end{array} \\right] = \\left(\\mathbf {D} ^ {T} \\mathbf {D}\\right) ^ {- 1} \\mathbf {D} ^ {T} \\mathbf {b}. \\tag {15}", + "image_path": "7b15d7b285ac9aa82c28f88c3481d12981d96fcca4c6e59c72a7c993c99a8243.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 251 + ], + "type": "text", + "content": "Once " + }, + { + "bbox": [ + 132, + 227, + 481, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 132, + 227, + 481, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 227, + 481, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 132, + 227, + 481, + 251 + ], + "type": "text", + "content": " are estimated, we can calculate all camera intrinsic parameters and optical center " + }, + { + "bbox": [ + 132, + 227, + 481, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{cp}" + }, + { + "bbox": [ + 132, + 227, + 481, + 251 + ], + "type": "text", + "content": " by decomposing " + }, + { + "bbox": [ + 132, + 227, + 481, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 132, + 227, + 481, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 227, + 481, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 132, + 227, + 481, + 251 + ], + "type": "text", + "content": " respectively:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 201, + 259, + 413, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 259, + 413, + 327 + ], + "spans": [ + { + "bbox": [ + 201, + 259, + 413, + 327 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} c _ {x} = W _ {1 3}, \\quad c _ {y} = W _ {2 3}, \\quad f _ {y} = \\sqrt {W _ {2 2} - c _ {y} ^ {2}}, \\\\ \\gamma = \\left(W _ {1 2} - c _ {x} c _ {y}\\right) / f _ {y}, \\quad f _ {x} = \\sqrt {W _ {1 1} - c _ {x} ^ {2} - \\gamma^ {2}}, \\\\ x = A _ {1 3} / A _ {3 3}, \\quad y = A _ {2 3} / A _ {3 3}, \\\\ r = - \\sqrt {\\left(A _ {1 1} / A _ {3 3}\\right) - \\left(A _ {1 3} / A _ {3 3}\\right) ^ {2}}. \\\\ \\end{array}", + "image_path": "ead05a6cb06b74cac24321f817b48b88c93aa7211310f1f631df209a8a062e04.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 333, + 479, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 479, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 479, + 346 + ], + "type": "text", + "content": "Subsequently, the rotation matrix of each image can be calculated with Eq. (5):" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 187, + 357, + 480, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 357, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 187, + 357, + 480, + 369 + ], + "type": "interline_equation", + "content": "\\mathbf {r} _ {1} = (1 / \\lambda) \\mathbf {K} ^ {- 1} \\mathbf {h} _ {1}, \\quad \\mathbf {r} _ {2} = (1 / \\lambda) \\mathbf {K} ^ {- 1} \\mathbf {h} _ {2}, \\quad \\mathbf {r} _ {3} = \\mathbf {r} _ {1} \\times \\mathbf {r} _ {2}, \\tag {16}", + "image_path": "f6d60bcd7cfe12ea9975bfd08dec6d4c87da73a182d548e1fd296f9c41bf0252.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 373, + 481, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 373, + 481, + 398 + ], + "spans": [ + { + "bbox": [ + 132, + 373, + 481, + 398 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 373, + 481, + 398 + ], + "type": "inline_equation", + "content": "\\lambda = (\\| \\mathbf{K}^{-1}\\mathbf{h}_1\\| +\\| \\mathbf{K}^{-1}\\mathbf{h}_2\\|) / 2" + }, + { + "bbox": [ + 132, + 373, + 481, + 398 + ], + "type": "text", + "content": ". Finally, a singular value decomposition for re-orthogonalizing the rotation matrix " + }, + { + "bbox": [ + 132, + 373, + 481, + 398 + ], + "type": "inline_equation", + "content": "\\mathbf{R} = [\\mathbf{r}_1,\\mathbf{r}_2,\\mathbf{r}_3]" + }, + { + "bbox": [ + 132, + 373, + 481, + 398 + ], + "type": "text", + "content": " is required." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 313, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 313, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 313, + 424 + ], + "type": "text", + "content": "4.3 Minimal Solver for Two Views" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 430, + 481, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 491 + ], + "type": "text", + "content": "When only two collimator images are used for calibration, the rank of the coefficient matrix in Eq. (14) is determined to be 10. Solving rank-deficient equations presents a complex challenge. To address this issue, we construct a nonlinear equation system and propose a novel approach that utilizes the hidden variable technique [14] to estimate the unknowns in the minimum configuration." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 491, + 481, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 527 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 527 + ], + "type": "text", + "content": "Estimation of " + }, + { + "bbox": [ + 132, + 491, + 481, + 527 + ], + "type": "inline_equation", + "content": "\\mathbf{W} = \\mathbf{K}^{-T}\\mathbf{K}^{-1}" + }, + { + "bbox": [ + 132, + 491, + 481, + 527 + ], + "type": "text", + "content": " is the core of calibration. Note that " + }, + { + "bbox": [ + 132, + 491, + 481, + 527 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 132, + 491, + 481, + 527 + ], + "type": "text", + "content": " is a symmetric matrix, defined by a 6-vector " + }, + { + "bbox": [ + 132, + 491, + 481, + 527 + ], + "type": "inline_equation", + "content": "\\mathbf{w} = (W_{11}, W_{12}, W_{13}, W_{22}, W_{23}, W_{33})^T" + }, + { + "bbox": [ + 132, + 491, + 481, + 527 + ], + "type": "text", + "content": ". We can extract the following 5 constraints from Eq. (7)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 533, + 480, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 533, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 317, + 533, + 480, + 545 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {1 2} ^ {T} \\mathbf {w} = 0 \\tag {17}", + "image_path": "27bd4c2db4840e0f5b46ec813f6140502c0add7b138985a3f8fa13acde5da670.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 282, + 549, + 480, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 549, + 480, + 562 + ], + "spans": [ + { + "bbox": [ + 282, + 549, + 480, + 562 + ], + "type": "interline_equation", + "content": "\\left(\\mathbf {v} _ {1 1} ^ {T} - \\mathbf {v} _ {2 2} ^ {T}\\right) \\mathbf {w} = 0 \\tag {18}", + "image_path": "71f582cf67c5405ac43cb16770f946c3e0d0554847724d84355d48edf8bb23bd.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 275, + 565, + 480, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 565, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 275, + 565, + 480, + 578 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {1 3} ^ {T} \\mathbf {w} + x \\mathbf {v} _ {1 1} ^ {T} \\mathbf {w} = 0 \\tag {19}", + "image_path": "80215c91616cccc8c3ce3f93899d5a1384e912229f3ccf1b32f51e3add0c9ae1.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 276, + 582, + 480, + 595 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 582, + 480, + 595 + ], + "spans": [ + { + "bbox": [ + 276, + 582, + 480, + 595 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {2 3} ^ {T} \\mathbf {w} + y \\mathbf {v} _ {1 1} ^ {T} \\mathbf {w} = 0 \\tag {20}", + "image_path": "3e16b89d661a8690069d13c5c6f09fcb7e42bc58ac2e80c37d586834caa64e9f.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 254, + 599, + 480, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 599, + 480, + 612 + ], + "spans": [ + { + "bbox": [ + 254, + 599, + 480, + 612 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {3 3} ^ {T} \\mathbf {w} - \\| \\mathbf {t} _ {c p} \\| ^ {2} \\mathbf {v} _ {1 1} ^ {T} \\mathbf {w} = 0. \\tag {21}", + "image_path": "3871c8919f5cc88e5454ae0960452038c9521de80647040628aa1d24a602c628.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 619, + 290, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 619, + 290, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 619, + 290, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{mn}" + }, + { + "bbox": [ + 132, + 619, + 290, + 631 + ], + "type": "text", + "content": " can be calculated as following:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 192, + 638, + 421, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 638, + 421, + 649 + ], + "spans": [ + { + "bbox": [ + 192, + 638, + 421, + 649 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {m n} = \\left[ h _ {1 m} h _ {1 n}, h _ {1 m} h _ {2 n} + h _ {1 n} h _ {2 m}, h _ {1 m} h _ {3 n} + h _ {1 n} h _ {3 m} \\right.", + "image_path": "a2361fc0240fd3a0956e4752bdcf3e15624023ea00140f93f69bc1684333b88a.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 224, + 653, + 389, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 653, + 389, + 666 + ], + "spans": [ + { + "bbox": [ + 224, + 653, + 389, + 666 + ], + "type": "interline_equation", + "content": "\\left. h _ {2 m} h _ {2 n}, h _ {2 m} h _ {3 n} + h _ {2 n} h _ {3 m}, h _ {3 m} h _ {3 n} \\right] ^ {T},", + "image_path": "37fc7ab7d8766aa7ce4817c5a6752d85e445101ded7a223caaa38eca495a82d4.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 265, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 265, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 265, + 102 + ], + "type": "text", + "content": "S. Liang, B. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "h_{mn}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " denotes the value of " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "-th row and " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "-th column of matrix " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": ". Given only 2 collimator images, we can obtain 10 independent equations for 8 unknowns. The hidden variable technique [14] is employed to solve the above polynomial equations. Combining Eqs. (19) to (21), we have" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 235, + 170, + 481, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 170, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 235, + 170, + 481, + 184 + ], + "type": "interline_equation", + "content": "\\left(\\mathbf {v} _ {1 3} ^ {T} + \\mathbf {v} _ {2 3} ^ {T} + \\mathbf {v} _ {3 3} ^ {T}\\right) \\mathbf {w} + c \\mathbf {v} _ {1 1} ^ {T} \\mathbf {w} = 0, \\tag {22}", + "image_path": "92f0b0b8a7d0fa6870eaf2da053417ecee62178abbc5729ecb2db44bc5cac155.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 190, + 479, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 479, + 227 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 479, + 227 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 190, + 479, + 227 + ], + "type": "inline_equation", + "content": "c = (x + y - \\| \\mathbf{t}_{cp} \\|^2)" + }, + { + "bbox": [ + 130, + 190, + 479, + 227 + ], + "type": "text", + "content": ". We first select constraints Eqs. (17), (18) and (22) to establish a new polynomial equation system. Next, we treat " + }, + { + "bbox": [ + 130, + 190, + 479, + 227 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 190, + 479, + 227 + ], + "type": "text", + "content": " as a hidden variable, and the complete set of equations can be expressed in matrix form as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 188, + 232, + 481, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 232, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 188, + 232, + 481, + 270 + ], + "type": "interline_equation", + "content": "\\mathbf {C} (c) \\mathbf {w} = 0, \\text {w i t h} \\mathbf {C} (c) = \\left[ \\begin{array}{c} \\mathbf {v} _ {1 2} ^ {T} \\\\ (\\mathbf {v} _ {1 1} ^ {T} - \\mathbf {v} _ {2 2} ^ {T}) \\\\ (\\mathbf {v} _ {1 3} ^ {T} + \\mathbf {v} _ {2 3} ^ {T} + \\mathbf {v} _ {3 3} ^ {T}) + c \\mathbf {v} _ {1 1} ^ {T} \\end{array} \\right], \\tag {23}", + "image_path": "8c5bca5746cc8828bcb63490d5cf201b5af5dd5915c43e74b14634f49f61f3a6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "text", + "content": "where the coefficient matrix " + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{C}(c)" + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "text", + "content": " contains the hidden variable " + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "text", + "content": ". Two images provide 6 equations for solving " + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{C}(c)" + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "inline_equation", + "content": "6 \\times 6" + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "text", + "content": " matrix. Since the equation system " + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{C}(c)\\mathbf{w} = 0" + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "text", + "content": " must have at least one nontrivial solution for " + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "text", + "content": ", the determinant of " + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{C}(c)" + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "text", + "content": " is equal to zero. The equation, " + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "inline_equation", + "content": "\\operatorname*{det}(\\mathbf{C}(c)) = 0" + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "text", + "content": ", is a polynomial of degree 2 in terms of " + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "text", + "content": ", which can be efficiently solved using a straightforward root-finding algorithm. Substituting the known value of " + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "text", + "content": " into the constraint equations Eqs. (17), (18) and (22), we obtain six constraint equations about " + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "text", + "content": ". The least squares method is then employed to solve these linear equations and compute the optimal " + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "text", + "content": " up to scale. Once " + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "text", + "content": " is estimated, the intrinsic matrix " + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 276, + 480, + 396 + ], + "type": "text", + "content": " can be calculated using the Cholesky factorization [15]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 411, + 258, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 411, + 258, + 423 + ], + "spans": [ + { + "bbox": [ + 131, + 411, + 258, + 423 + ], + "type": "text", + "content": "4.4 Bundle Adjustment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 428, + 479, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 428, + 479, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 428, + 479, + 511 + ], + "type": "text", + "content": "Bundle adjustment jointly refines the camera intrinsic " + }, + { + "bbox": [ + 130, + 428, + 479, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 428, + 479, + 511 + ], + "type": "text", + "content": ", lens distortion coefficients " + }, + { + "bbox": [ + 130, + 428, + 479, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 130, + 428, + 479, + 511 + ], + "type": "text", + "content": ", the rotation matrix of the " + }, + { + "bbox": [ + 130, + 428, + 479, + 511 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 428, + 479, + 511 + ], + "type": "text", + "content": "-th image " + }, + { + "bbox": [ + 130, + 428, + 479, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i" + }, + { + "bbox": [ + 130, + 428, + 479, + 511 + ], + "type": "text", + "content": " and camera optical center " + }, + { + "bbox": [ + 130, + 428, + 479, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{cp}" + }, + { + "bbox": [ + 130, + 428, + 479, + 511 + ], + "type": "text", + "content": ". In cases of moderate lens distortion, it is typical to initialize the distortion coefficient to " + }, + { + "bbox": [ + 130, + 428, + 479, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{0}" + }, + { + "bbox": [ + 130, + 428, + 479, + 511 + ], + "type": "text", + "content": " and estimate them in the bundle adjustment step [36]. Given a set of " + }, + { + "bbox": [ + 130, + 428, + 479, + 511 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 428, + 479, + 511 + ], + "type": "text", + "content": " images, each with " + }, + { + "bbox": [ + 130, + 428, + 479, + 511 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 428, + 479, + 511 + ], + "type": "text", + "content": " corresponding 2D image points, we establish a cost function aimed at minimizing the re-projection error. The cost function is expressed as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 207, + 518, + 481, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 518, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 207, + 518, + 481, + 552 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {X}} \\sum_ {i} ^ {N} \\sum_ {j} ^ {M} \\rho \\left(\\left\\| \\pi \\left(\\mathbf {K}, \\mathbf {d}, \\mathbf {R} _ {i}, \\mathbf {t} _ {c p}, \\mathbf {P} _ {i j}\\right) - \\mathbf {p} _ {i j} \\right\\| ^ {2}\\right), \\tag {24}", + "image_path": "95c38413b3bb15b72fa3c10093636cbb202b65b1d7122ef9c7a0d63149d1fd23.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 557, + 479, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 479, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 479, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 557, + 479, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = (\\mathbf{K},\\mathbf{d},\\mathbf{R}_i,\\mathbf{t}_{cp})" + }, + { + "bbox": [ + 130, + 557, + 479, + 617 + ], + "type": "text", + "content": " is the parameter to be refined. " + }, + { + "bbox": [ + 130, + 557, + 479, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{ij}" + }, + { + "bbox": [ + 130, + 557, + 479, + 617 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 557, + 479, + 617 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 557, + 479, + 617 + ], + "type": "text", + "content": "-th 3D pattern point of " + }, + { + "bbox": [ + 130, + 557, + 479, + 617 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 557, + 479, + 617 + ], + "type": "text", + "content": "-th image and corresponds to the 2D image point " + }, + { + "bbox": [ + 130, + 557, + 479, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{ij}" + }, + { + "bbox": [ + 130, + 557, + 479, + 617 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 557, + 479, + 617 + ], + "type": "inline_equation", + "content": "\\pi (\\cdot)" + }, + { + "bbox": [ + 130, + 557, + 479, + 617 + ], + "type": "text", + "content": " projects the world points to image according to Eq. (4). " + }, + { + "bbox": [ + 130, + 557, + 479, + 617 + ], + "type": "inline_equation", + "content": "\\rho (\\cdot)" + }, + { + "bbox": [ + 130, + 557, + 479, + 617 + ], + "type": "text", + "content": " is the Cauchy cost function. As is common, we optimize Eq. (24) with the Levenberg-Marquardt (LM) method [27], which is implemented in some excellent frameworks [1]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 618, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 618, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 618, + 481, + 665 + ], + "type": "text", + "content": "Note that the camera optical center is fixed at " + }, + { + "bbox": [ + 130, + 618, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{cp}" + }, + { + "bbox": [ + 130, + 618, + 481, + 665 + ], + "type": "text", + "content": " with the spherical motion constraint. However, for general motion, the camera optical center is free. Therefore, when calibrating the camera using collimator images, the number of motion parameters to be refined is reduced from " + }, + { + "bbox": [ + 130, + 618, + 481, + 665 + ], + "type": "inline_equation", + "content": "6N" + }, + { + "bbox": [ + 130, + 618, + 481, + 665 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 618, + 481, + 665 + ], + "type": "inline_equation", + "content": "3N + 3" + }, + { + "bbox": [ + 130, + 618, + 481, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "text", + "content": "Camera Calibration using a Collimator System" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 297, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 297, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 297, + 129 + ], + "type": "text", + "content": "4.5 Degenerated Configuration" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 481, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 481, + 160 + ], + "type": "text", + "content": "We prove a degenerate configuration where additional images cannot provide more constraints about the camera intrinsic parameter." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 161, + 481, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 481, + 196 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 481, + 196 + ], + "type": "text", + "content": "Proposition 1. If the calibration target rotates only around an axis that is parallel to the " + }, + { + "bbox": [ + 130, + 161, + 481, + 196 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 161, + 481, + 196 + ], + "type": "text", + "content": " axis and intersects the target plane at " + }, + { + "bbox": [ + 130, + 161, + 481, + 196 + ], + "type": "inline_equation", + "content": "(x,y,0)^T" + }, + { + "bbox": [ + 130, + 161, + 481, + 196 + ], + "type": "text", + "content": ", the newly added image cannot provide additional constraints on the camera intrinsic parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 198, + 481, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 198, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 146, + 198, + 481, + 209 + ], + "type": "text", + "content": "Due to space limitations, the proof is provided in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 228, + 292, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 292, + 241 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 292, + 241 + ], + "type": "text", + "content": "5 Evaluation and Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 255, + 481, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 481, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 481, + 291 + ], + "type": "text", + "content": "In this section, we first evaluate the robustness and accuracy of the proposed algorithm on synthetic data. Then, we calibrate the camera and evaluate the effectiveness of the proposed method using real collimator images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 309, + 304, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 309, + 304, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 309, + 304, + 322 + ], + "type": "text", + "content": "5.1 Synthetic Data Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 330, + 481, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 481, + 377 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 481, + 377 + ], + "type": "text", + "content": "In the synthetic data experiment, we evaluate the performance of the proposed algorithm. We simulate the process of calibrating predefined camera parameters using synthetic images that conform to the spherical motion model. The proposed is compared with the following algorithms:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 378, + 479, + 448 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 132, + 378, + 477, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 378, + 477, + 389 + ], + "spans": [ + { + "bbox": [ + 132, + 378, + 477, + 389 + ], + "type": "text", + "content": "- Zhang [46] is a plane-based algorithm and suitable for general 6DOF motion." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 390, + 479, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 479, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 479, + 424 + ], + "type": "text", + "content": "- Bouguet [3] is a calibration toolbox and suitable for general 6DOF motion. Unlike Zhang, Bouguet computes the focal length using the orthogonal vanishing points constraint." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 426, + 479, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 479, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 479, + 448 + ], + "type": "text", + "content": "- Hartley [17] is a self-calibration method based on a pure rotating camera, which is specifically suitable for the 3DOF spherical motion model." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "text", + "content": "We define a virtual camera with the following parameters: " + }, + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "inline_equation", + "content": "f_{x} = f_{y} = 1000" + }, + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "inline_equation", + "content": "c_{x} = 542" + }, + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "inline_equation", + "content": "c_{y} = 478" + }, + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "inline_equation", + "content": "\\gamma = 0.01" + }, + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "text", + "content": ", radial distortion coefficients " + }, + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "inline_equation", + "content": "d_{1} = 0.1" + }, + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "inline_equation", + "content": "d_{2} = -0.2" + }, + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "text", + "content": ", and the image size is set to " + }, + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "inline_equation", + "content": "1080 \\times 960" + }, + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "text", + "content": " pixels. The calibration target is a planar pattern containing " + }, + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "inline_equation", + "content": "11 \\times 8" + }, + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "text", + "content": " points, each square size of " + }, + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "inline_equation", + "content": "30mm \\times 30mm" + }, + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "text", + "content": ". The target maintains a spherical motion with a radius of " + }, + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "inline_equation", + "content": "700mm" + }, + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "text", + "content": " relative to the camera. The camera optical center in the target coordinate system is " + }, + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "inline_equation", + "content": "t_{cp} = [150, 105, -700]^{T}mm" + }, + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "text", + "content": ", and the camera's orientation is arbitrary. Given the camera parameters and poses, the planar target points are projected onto images. Finally, zero-mean Gaussian noise with " + }, + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 449, + 486, + 557 + ], + "type": "text", + "content": " standard deviation is added to 2D image points." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": "Evaluation of Calibration. In this experiment, we seek to evaluate the performance of the algorithms in solving camera parameters. For the four algorithms, we perform two steps of solving initial value and bundle adjustment to obtain camera parameters. Firstly, we compare the performance of the mentioned algorithms as the noise level increases. The noise level varies from 0 to 3.0 pixels, and we use 15 images to calibrate the camera. For each noise level, we conducted 200 trials and calculated the average errors. As shown in Fig. 3, calibration error increases linearly with the noise level. Compared with the other three algorithms, ours shows greater robustness to noise. Additionally, at a noise" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 264, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 264, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 264, + 101 + ], + "type": "text", + "content": "S. Liang, B. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 220, + 194 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 220, + 194 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 220, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 220, + 194 + ], + "type": "image", + "image_path": "2e124ae4208e52f24641e2988d76866c8ef6bc7fdb5781abe9eaada7c108dc26.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 202, + 443, + 213 + ], + "lines": [ + { + "bbox": [ + 168, + 202, + 443, + 213 + ], + "spans": [ + { + "bbox": [ + 168, + 202, + 443, + 213 + ], + "type": "text", + "content": "Fig. 3: Comparison of calibration errors with increasing noise level." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 221, + 114, + 304, + 194 + ], + "blocks": [ + { + "bbox": [ + 221, + 114, + 304, + 194 + ], + "lines": [ + { + "bbox": [ + 221, + 114, + 304, + 194 + ], + "spans": [ + { + "bbox": [ + 221, + 114, + 304, + 194 + ], + "type": "image", + "image_path": "1c25801bcbb8e6190215e5e34675e904e1525946a08b499cf88fbae47fe9765a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 306, + 113, + 391, + 194 + ], + "blocks": [ + { + "bbox": [ + 306, + 113, + 391, + 194 + ], + "lines": [ + { + "bbox": [ + 306, + 113, + 391, + 194 + ], + "spans": [ + { + "bbox": [ + 306, + 113, + 391, + 194 + ], + "type": "image", + "image_path": "bc1076d33f01ccb87ffdf2f913b9c964067086737439128326e4dad7b4190a50.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 392, + 113, + 480, + 194 + ], + "blocks": [ + { + "bbox": [ + 392, + 113, + 480, + 194 + ], + "lines": [ + { + "bbox": [ + 392, + 113, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 392, + 113, + 480, + 194 + ], + "type": "image", + "image_path": "9e7ca5d23c88494883ca0fa5b7f0a0e794cb7a7d8202c7e47660a30cc773620b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 132, + 241, + 220, + 320 + ], + "blocks": [ + { + "bbox": [ + 132, + 241, + 220, + 320 + ], + "lines": [ + { + "bbox": [ + 132, + 241, + 220, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 241, + 220, + 320 + ], + "type": "image", + "image_path": "d93f5f3c79cb58859b4b3fcf9c38bddd090015ddb8eff8270397ed3568f3e503.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 329, + 464, + 340 + ], + "lines": [ + { + "bbox": [ + 147, + 329, + 464, + 340 + ], + "spans": [ + { + "bbox": [ + 147, + 329, + 464, + 340 + ], + "type": "text", + "content": "Fig. 4: Comparison of calibration errors with the number of images increases." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 221, + 242, + 304, + 321 + ], + "blocks": [ + { + "bbox": [ + 221, + 242, + 304, + 321 + ], + "lines": [ + { + "bbox": [ + 221, + 242, + 304, + 321 + ], + "spans": [ + { + "bbox": [ + 221, + 242, + 304, + 321 + ], + "type": "image", + "image_path": "e9df0542566cdf33338b02f3c618fdd99fb618ee7483fea2d108057fc1feaaaf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 241, + 392, + 321 + ], + "blocks": [ + { + "bbox": [ + 306, + 241, + 392, + 321 + ], + "lines": [ + { + "bbox": [ + 306, + 241, + 392, + 321 + ], + "spans": [ + { + "bbox": [ + 306, + 241, + 392, + 321 + ], + "type": "image", + "image_path": "64074298e4f3bcd9f9414e821b3874e7e72962c09b4337c94b6f621eacd8bd56.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 392, + 241, + 480, + 321 + ], + "blocks": [ + { + "bbox": [ + 392, + 241, + 480, + 321 + ], + "lines": [ + { + "bbox": [ + 392, + 241, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 392, + 241, + 480, + 321 + ], + "type": "image", + "image_path": "7ab5cbb4daf7b03699e49b69f7f31651c5c52c38a1ddac561eac6757314e7c67.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 381, + 482, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 482, + 490 + ], + "type": "text", + "content": "level of 0.5 pixels, we evaluate the performance with respect to the number of calibration images. The image count ranged from 3 to 30, with 200 trials conducted for each quantity. The results are illustrated in Fig. 4. A significant decrease in error is observed when increasing the number of images from 3 to 5. Both Fig. 3 and Fig. 4 show that our algorithm outperforms other algorithms in robustness and accuracy. We believe that our algorithm benefits from the consideration of both spherical motion and planar homography constraints, along with reducing optimization variables. Due to space limitations, the additional comparison results of the initial value estimation are shown in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Sensitivity of imperfect spherical motion. The low-cost collimators exhibit an inferior parallelism, which is reflected in the fact that the relative motion between the target and camera is not a perfect spherical motion. This experiment evaluates the sensitivity of initial value estimation with respect to imperfect spherical motion. We add zeros-mean Gaussian noise with standard deviation from " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "0mm" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "30mm" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{cp}" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " to replicate this imperfection. We use 15 images and add a noise level of 0.5 pixels to the image points. For each noise level, a total of independent 500 trials are conducted and the average errors are computed. The sensitivity analysis results are shown in Fig. 5. Noise has no impact on Bouguet and Zhang, because they do not rely on the spherical motion constraint. As the noise increases, calibration errors of Ours and Hartley increase gradually. Notably, Hartley is significantly influenced by noise, whereas Ours displays superior robustness. Even with noise levels below " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "15mm" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " (equivalent to " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "2.143\\%" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": " of the spherical radius), our algorithm remains superior to the alternatives." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "text", + "content": "Camera Calibration using a Collimator System" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 194, + 114, + 302, + 210 + ], + "blocks": [ + { + "bbox": [ + 194, + 114, + 302, + 210 + ], + "lines": [ + { + "bbox": [ + 194, + 114, + 302, + 210 + ], + "spans": [ + { + "bbox": [ + 194, + 114, + 302, + 210 + ], + "type": "image", + "image_path": "4a215737a0b6bb2ae96cecbb6dbe73a3c26e43c0929e5dcb1e765036cd3a4a52.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 188, + 219, + 424, + 231 + ], + "lines": [ + { + "bbox": [ + 188, + 219, + 424, + 231 + ], + "spans": [ + { + "bbox": [ + 188, + 219, + 424, + 231 + ], + "type": "text", + "content": "Fig. 5: Sensitivity analysis of imperfect spherical motion." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 113, + 419, + 211 + ], + "blocks": [ + { + "bbox": [ + 309, + 113, + 419, + 211 + ], + "lines": [ + { + "bbox": [ + 309, + 113, + 419, + 211 + ], + "spans": [ + { + "bbox": [ + 309, + 113, + 419, + 211 + ], + "type": "image", + "image_path": "b10433ef99734c73e1bca81cff90288ef225737bc6e852db743a65f53bb80c69.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 186, + 262, + 426, + 334 + ], + "blocks": [ + { + "bbox": [ + 146, + 242, + 465, + 255 + ], + "lines": [ + { + "bbox": [ + 146, + 242, + 465, + 255 + ], + "spans": [ + { + "bbox": [ + 146, + 242, + 465, + 255 + ], + "type": "text", + "content": "Table 1: Comparison of calibration and pose estimation error results (pixels)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 186, + 262, + 426, + 334 + ], + "lines": [ + { + "bbox": [ + 186, + 262, + 426, + 334 + ], + "spans": [ + { + "bbox": [ + 186, + 262, + 426, + 334 + ], + "type": "table", + "html": "
NumberZhang [46]Bouguet [3]Hartley [17]Ours
εcalibεevalεcalibεevalεcalibεevalεcalibεeval
20.18420.15450.17390.1518--0.17390.1519
50.23160.12450.22020.1236-0.25770.22030.1236
100.19230.12310.19120.1232-0.19860.19510.1222
150.17750.12210.17180.1232-0.17910.17790.1216
200.17730.12250.17220.1221-0.18710.17780.1207
", + "image_path": "8c16e1b52e515b58e928759256a3f8660620e249a604f07a6054dca67f583f53.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 357, + 290, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 290, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 290, + 369 + ], + "type": "text", + "content": "5.2 Real Images Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 378, + 481, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 481, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 481, + 401 + ], + "type": "text", + "content": "In real image experiments, we verify the effectiveness of our collimator system for calibration and evaluate the performance of the proposed algorithm." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 402, + 481, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 481, + 594 + ], + "type": "text", + "content": "Calibration and Accuracy Evaluation. Following the experimental methodology utilized in previous studies [13,30], we design a pose estimation experiment to evaluate the accuracy of camera calibration quantitatively. A 8mm camera captures images from our collimator system and a practical printed pattern. Calibration is performed using the collimator images and a subset of the printed pattern images, while the remaining printed pattern images are used to evaluate the calibration results. After obtaining the camera parameters from the calibration images, the pose of each evaluation image can be determined using absolute pose estimation [24]. In our setup, a quarter of the corner points on the printed pattern are used for pose estimation, and the remaining points are used for evaluation. By re-projecting the evaluation corner points onto the image with the camera parameters and the calculated pose, we can evaluate the calibration quality based on the re-projection error. A total of 200 printed pattern images are used to evaluate the calibration. The calibration uses 2, 5, 10, 15, and 20 images. The re-projection error of the calibration and evaluation images are donated as " + }, + { + "bbox": [ + 130, + 402, + 481, + 594 + ], + "type": "inline_equation", + "content": "\\epsilon_{calib}" + }, + { + "bbox": [ + 130, + 402, + 481, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 402, + 481, + 594 + ], + "type": "inline_equation", + "content": "\\epsilon_{eval}" + }, + { + "bbox": [ + 130, + 402, + 481, + 594 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "The comparison results of the four algorithms are shown in Tab. 1. Due to the limited degrees of freedom, our calibrated re-projection error " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\epsilon_{calib})" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " is slightly higher. However, Ours exhibits the lowest re-projection error in evaluation images compared to alternative methods, indicating that our method achieves more precise camera parameters. The Hartley method only uses the correspondence of 2D points in images, which makes it impossible to calculate the re-projection" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 264, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 264, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 264, + 102 + ], + "type": "text", + "content": "S. Liang, B. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 194, + 135, + 416, + 236 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 463, + 126 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 463, + 126 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 463, + 126 + ], + "type": "text", + "content": "Table 2: Comparison of calibration results using 20 images in a real camera." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 194, + 135, + 416, + 236 + ], + "lines": [ + { + "bbox": [ + 194, + 135, + 416, + 236 + ], + "spans": [ + { + "bbox": [ + 194, + 135, + 416, + 236 + ], + "type": "table", + "html": "
PrintedCollimator
Bouguet [3]Zhang [46]Bouguet [3]Hartley [17]Ours
fx2384.52380.82369.22644.62388.3
fy2386.32379.82368.92646.02387.9
cx1209.81222.11221.11214.61221.0
cy1008.21008.41009.81012.71009.6
k1-0.0843-0.0899-0.0908-0.0754-0.0900
k20.08190.09120.08920.11810.0954
εcalib0.15130.17730.1722-0.1778
εeval0.12250.12250.12210.18710.1207
", + "image_path": "bcc5c90d60bb7c14649c13bfd890ce1e7ad8ed32cd593ff76150a7320f263760.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 186, + 269, + 428, + 340 + ], + "blocks": [ + { + "bbox": [ + 168, + 248, + 444, + 261 + ], + "lines": [ + { + "bbox": [ + 168, + 248, + 444, + 261 + ], + "spans": [ + { + "bbox": [ + 168, + 248, + 444, + 261 + ], + "type": "text", + "content": "Table 3: Comparison of epipolar rectification error results (pixels)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 186, + 269, + 428, + 340 + ], + "lines": [ + { + "bbox": [ + 186, + 269, + 428, + 340 + ], + "spans": [ + { + "bbox": [ + 186, + 269, + 428, + 340 + ], + "type": "table", + "html": "
NumberZhang [46]Bouguet [3]Hartley [17]Ours
meanstdmeanstdmeanstdmeanstd
20.08460.04060.07700.0220--0.07700.0229
50.08090.03070.07640.02140.08320.03970.07620.0211
100.07730.02340.07580.02130.08540.05090.07570.0208
150.07820.02660.07710.02530.08370.05220.07560.0207
200.07720.02520.07680.02530.08180.03650.07560.0208
", + "image_path": "6841262af48481af4e85a3e1c955aaef82b4ec39fa66d3b2cfdba465bd903206.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 364, + 481, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 481, + 399 + ], + "type": "text", + "content": "error of the calibration images directly. In cases with only two images, Zhang and Bouguet commonly depend on beneficial assumptions, such as " + }, + { + "bbox": [ + 130, + 364, + 481, + 399 + ], + "type": "inline_equation", + "content": "\\gamma = 0" + }, + { + "bbox": [ + 130, + 364, + 481, + 399 + ], + "type": "text", + "content": ". Ours achieves comparable evaluation errors without the need for such assumptions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 400, + 481, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 481, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 481, + 496 + ], + "type": "text", + "content": "Additionally, the Bouguet toolbox is also used to calculate camera parameters from printed pattern images that are not included in the evaluation. We use 20 images for calibration and 200 for evaluation. The estimated camera parameters and error comparison are shown in Tab. 2. The second column is a good reference. The lower evaluation error indicates that our method outperforms the alternative and validates the effectiveness of using the collimator image for calibration. It can be expected that the calibration parameters are not significantly different. Therefore, even a minor improvement holds considerable significance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 497, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 481, + 605 + ], + "type": "text", + "content": "Then, we evaluate the performance of all methods using epipolar rectification error, which has not been optimized. Given the camera parameters from different methods, we calculate the epipolar rectification error (in pixels) for 300 pairs of printed evaluation images. The mean and standard deviation of the errors are shown in Tab. 3. Moreover, the result using Bouguet with 20 printed pattern images is (0.0772, 0.0259) pixels. It is observed that as the number of calibration images increases, the error stabilizes. The results indicate improved accuracy in the collimator image and highlight the beneficial impact of the spherical constraint on accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "Multiple Cameras Calibration with a Same Collimator. In this experiment, we evaluate our method on multiple cameras with the same collimator. We used the cameras with " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "12mm" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "16mm" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "25mm" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "35mm" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": " focal lengths of lenses. The calibration images are captured from the same collimator, as shown in Fig. 6 and left of Fig. 1. As the focal length increases, the more pixels a single" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "text", + "content": "Camera Calibration using a Collimator System" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 219, + 114, + 394, + 178 + ], + "blocks": [ + { + "bbox": [ + 219, + 114, + 394, + 178 + ], + "lines": [ + { + "bbox": [ + 219, + 114, + 394, + 178 + ], + "spans": [ + { + "bbox": [ + 219, + 114, + 394, + 178 + ], + "type": "image", + "image_path": "832534a9e98b421a82ef6735dfe964b146ccf1b579a1712e7798c7ae8f7fe012.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 188, + 435, + 199 + ], + "lines": [ + { + "bbox": [ + 176, + 188, + 435, + 199 + ], + "spans": [ + { + "bbox": [ + 176, + 188, + 435, + 199 + ], + "type": "text", + "content": "Fig. 6: Calibration image sample obtained from the collimator." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 201, + 232, + 410, + 291 + ], + "blocks": [ + { + "bbox": [ + 134, + 212, + 477, + 224 + ], + "lines": [ + { + "bbox": [ + 134, + 212, + 477, + 224 + ], + "spans": [ + { + "bbox": [ + 134, + 212, + 477, + 224 + ], + "type": "text", + "content": "Table 4: Comparison of calibration results with different focal length lens (pixels)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 201, + 232, + 410, + 291 + ], + "lines": [ + { + "bbox": [ + 201, + 232, + 410, + 291 + ], + "spans": [ + { + "bbox": [ + 201, + 232, + 410, + 291 + ], + "type": "table", + "html": "
Zhang [46]Bouguet [3]Hartley [17]Ours
12mm0.22730.22250.34040.2174
16mm0.29940.25910.41820.2536
25mm0.31330.30780.41950.2923
35mm0.32890.33480.72720.3118
", + "image_path": "8cebb14b2c242b2dad97b25f7ad4542d9309b70efd1b50229402f6a2be3961ce.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 316, + 479, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 479, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 479, + 339 + ], + "type": "text", + "content": "square occupies. Cameras with different focal lengths can observe the calibration target through our collimator system in a limited space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 340, + 480, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 480, + 460 + ], + "type": "text", + "content": "Similar to the evaluation approach in the previous experiment, we segregate all collimator images into calibration and evaluation sets. For each lens, we use 20 images for calibration and 200 for evaluation. The results are listed in Tab. 4. The proposed algorithm demonstrates the lowest re-projection error for lenses with varying focal lengths among the evaluation images. The re-projection error tends to increase with the increase of focal length. We believe that the main reason is that the longer the lens's focal length, the fewer points observed. More importantly, this experiment confirms the effectiveness of our collimator system and the proposed method. Moreover, we provide a spherical motion verification experiment in supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 481, + 219, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 481, + 219, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 481, + 219, + 493 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 509, + 480, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 480, + 628 + ], + "type": "text", + "content": "This paper presents a novel camera calibration method using a designed collimator system. Based on the optical geometry of the collimator system, we prove that the relative motion between the target and the camera conforms to a spherical motion. Furthermore, we propose a closed-form solver for multiple views and a minimal solver for two views for camera calibration. Our collimator-based method effectively combines the advantages of close-range calibration using a collimator and flexibility using a planar calibration target. We validate the effectiveness of using only collimator images to calibrate the camera. Synthetic and real data experiments demonstrate that the calibration accuracy of the proposed method is superior to the state-of-the-art methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 630, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 480, + 665 + ], + "type": "text", + "content": "Limitations: When calibrating a fisheye camera, the distortion should be paid more attention rather than simply setting it to 0. Future work will focus on cameras with significant distortion." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "text", + "content": "S. Liang, B. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 190 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 190 + ], + "type": "text", + "content": "This research has been supported by the Hunan Provincial Natural Science Foundation for Excellent Young Scholars under Grant 2023JJ20045, and the National Natural Science Foundation of China under Grant 12372189. We thank Zijie Wang for useful discussion on the writing of the paper." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 209, + 198, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 209, + 198, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 209, + 198, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 234, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 234, + 479, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 479, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 479, + 255 + ], + "type": "text", + "content": "1. Agarwal, S., Mierle, K., Team, T.C.S.: Ceres Solver. https://github.com/ceres-solver/ceres-solver" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 258, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 258, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 258, + 480, + 289 + ], + "type": "text", + "content": "2. Bauer, M., Grießbach, D., Hermerschmidt, A., Krüger, S., Scheele, M., Schischmanow, A.: Geometrical camera calibration with diffractive optical elements. Optics express 16(25), 20241-20248 (2008)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 291, + 480, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 291, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 138, + 291, + 480, + 312 + ], + "type": "text", + "content": "3. Bouguet, J.Y.: Camera calibration toolbox for matlab. http://robots.stanford.edu/cs223b04/JeanYvesCalib/index.html (2004)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 312, + 480, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 480, + 334 + ], + "type": "text", + "content": "4. Bradski, G.: The opencv library. Dr. Dobb's Journal: Software Tools for the Professional Programmer 25(11), 120-123 (2000)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 335, + 480, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 480, + 355 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 480, + 355 + ], + "type": "text", + "content": "5. Brown, D.C.: Close-range camera calibration. Photogramm. Eng 37(8), 855-866 (1971)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 357, + 480, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 480, + 388 + ], + "type": "text", + "content": "6. Campos, C., Elvira, R., Rodríguez, J.J.G., M. Montiel, J.M., D. Tardós, J.: Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics 37(6), 1874-1890 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 390, + 480, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 390, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 390, + 480, + 411 + ], + "type": "text", + "content": "7. Clarke, T.A., Fryer, J.G.: The development of camera calibration methods and models. The Photogrammetric Record 16(91), 51-66 (1998)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 412, + 480, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 412, + 480, + 444 + ], + "spans": [ + { + "bbox": [ + 138, + 412, + 480, + 444 + ], + "type": "text", + "content": "8. Faugeras, O.D., Luong, Q.T., Maybank, S.J.: Camera self-calibration: Theory and experiments. In: IEEE European Conference on Computer Vision (ECCV). p. 321-334 (1992)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 445, + 480, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 480, + 477 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 480, + 477 + ], + "type": "text", + "content": "9. Fuhr, G., Jung, C.R.: Camera self-calibration based on nonlinear optimization and applications in surveillance systems. IEEE Transactions on Circuits and Systems for Video Technology 27(5), 1132-1142 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 478, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 478, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 478, + 480, + 510 + ], + "type": "text", + "content": "0. Guan, B., Su, Z., Yu, Q., Li, Z., Feng, W., Yang, D., Zhang, D.: Monitoring the blades of a wind turbine by using videogrammetry. Optics and Lasers in Engineering 152, 106901 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 511, + 480, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 511, + 480, + 543 + ], + "spans": [ + { + "bbox": [ + 138, + 511, + 480, + 543 + ], + "type": "text", + "content": "1. Guan, B., Zhao, J., Barath, D., Fraundorfer, F.: Minimal solvers for relative pose estimation of multi-camera systems using affine correspondences. International Journal of Computer Vision 131(1), 324-345 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 544, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 576 + ], + "type": "text", + "content": "2. Guan, B., Zhao, J., Li, Z., Sun, F., Fraundorfer, F.: Relative pose estimation with a single affine correspondence. IEEE Transactions on Cybernetics 52(10), 10111-10122 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 577, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 480, + 609 + ], + "type": "text", + "content": "3. Ha, H., Perdoch, M., Alismail, H., Kweon, I.S., Sheikh, Y.: Deltille grids for geometric camera calibration. In: IEEE International Conference on Computer Vision (ICCV). pp. 5354-5362 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "text", + "content": "4. Hartley, R., Li, H.: An efficient hidden variable approach to minimal-case camera motion estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence 34(12), 2303-2314 (2012)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "5. Hartley, R., Zisserman, A.: Multiple View Geometry in Computer Vision. Cambridge University Press, 2 edn. (2004)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "text", + "content": "Camera Calibration using a Collimator System" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "16. Hartley, R.I.: Self-calibration from multiple views with a rotating camera. In: European Conference on Computer Vision (ECCV). pp. 471-478 (1994)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "text", + "content": "17. Hartley, R.I.: Self-calibration of stationary cameras. International Journal of Computer Vision 22(1), 5-23 (1997)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 161, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 161, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 134, + 161, + 480, + 194 + ], + "type": "text", + "content": "18. Heikkila, J.: Geometric camera calibration using circular control points. IEEE Transactions on Pattern Analysis and Machine Intelligence 22(10), 1066-1077 (2000)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 134, + 194, + 480, + 226 + ], + "type": "text", + "content": "19. Herrera, D., Kannala, C.J., Heikkila, J.: Forget the checkerboard: Practical self-calibration using a planar scene. In: IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 1-9. IEEE (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 133, + 227, + 480, + 258 + ], + "type": "text", + "content": "20. Hieronymus, J.: Comparison of methods for geometric camera calibration. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences 39, 595-599 (2012)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 259, + 480, + 293 + ], + "type": "text", + "content": "21. Hu, H., Wei, B., Mei, S., Liang, J., Zhang, Y.: A two-step calibration method for vision measurement with large field of view. IEEE Transactions on Instrumentation and Measurement 71, 1-10 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 133, + 293, + 480, + 325 + ], + "type": "text", + "content": "22. Kannala, J., Brandt, S.S.: A generic camera model and calibration method for conventional, wide-angle, and fish-eye lenses. IEEE Transactions on Pattern Analysis and Machine Intelligence 28(8), 1335-1340 (2006)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 133, + 326, + 480, + 357 + ], + "type": "text", + "content": "23. Klaus, A., Bauer, J., Karner, K., Elbischger, P., Perko, R., Bischof, H.: Camera calibration from a single night sky image. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). vol. 1, pp. I-I (2004)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 358, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 358, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 133, + 358, + 480, + 391 + ], + "type": "text", + "content": "24. Kneip, L., Furgale, P.: Opengv: A unified and generalized approach to real-time calibrated geometric vision. In: IEEE International Conference on Robotics and Automation (ICRA). pp. 1-8 (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 133, + 392, + 480, + 423 + ], + "type": "text", + "content": "25. Liu, Z., Li, F., Li, X., Zhang, G.: A novel and accurate calibration method for cameras with large field of view using combined small targets. Measurement 64, 1-16 (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 133, + 424, + 480, + 456 + ], + "type": "text", + "content": "26. Lochman, Y., Liepieshov, K., Chen, J., Perdoch, M., Zach, C., Pritts, J.: Babel-calib: A universal approach to calibrating central cameras. In: IEEE International Conference on Computer Vision (ICCV). pp. 15253-15262 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 457, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 457, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 133, + 457, + 480, + 489 + ], + "type": "text", + "content": "27. Moré, J.J.: The levenberg-marquardt algorithm: implementation and theory. In: Numerical analysis: proceedings of the biennial Conference held at Dundee, June 28-July 1, 1977. pp. 105-116 (2006)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 491, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 491, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 133, + 491, + 480, + 521 + ], + "type": "text", + "content": "28. Nister, D., Stewenius, H., Grossmann, E.: Non-parametric self-calibration. In: IEEE International Conference on Computer Vision (ICCV). vol. 1, pp. 120-127 (2005)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 133, + 523, + 480, + 544 + ], + "type": "text", + "content": "29. Oniga, V.E., Pfeifer, N., Loghin, A.M.: 3d calibration test-field for digital cameras mounted on unmanned aerial systems (uas). Remote Sensing 10(12), 2017 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 545, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 545, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 133, + 545, + 480, + 577 + ], + "type": "text", + "content": "30. Peng, S., Sturm, P.: Calibration wizard: A guidance system for camera calibration based on modelling geometric and corner uncertainty. In: IEEE International Conference on Computer Vision (ICCV). pp. 1497-1505 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 578, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 578, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 578, + 480, + 609 + ], + "type": "text", + "content": "31. Pollefeys, M., Koch, R., Gool, L.V.: Self-calibration and metric reconstruction in spite of varying and unknown intrinsic camera parameters. International journal of computer vision 32(1), 7-25 (1999)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 611, + 480, + 632 + ], + "type": "text", + "content": "32. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4104-4113 (2016)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "text", + "content": "33. Schöps, T., Larsson, V., Pollefeys, M., Sattler, T.: Why having 10,000 parameters in your camera model is better than twelve. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2532-2541 (2020)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 264, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 264, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 264, + 102 + ], + "type": "text", + "content": "S. Liang, B. Guan et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 556 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "34. Shang, Y., Sun, X., Yang, X., Wang, X., Yu, Q.: A camera calibration method for large field optical measurement. Optik - International Journal for Light and Electron Optics 124, 6553-6558 (12 2013)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "35. Sturm, P.F., Maybank, S.J.: On plane-based camera calibration: A general algorithm, singularities, applications. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). vol. 1, pp. 432-437. IEEE (1999)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 216 + ], + "type": "text", + "content": "36. Triggs, B., McLauchlan, P.F., Hartley, R.I., Fitzgibbon, A.W.: Bundle adjustment—a modern synthesis. In: Vision Algorithms: Theory and Practice: International Workshop on Vision Algorithms. pp. 298-372. Springer (2000)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 248 + ], + "type": "text", + "content": "37. Tsai, R.: A versatile camera calibration technique for high-accuracy 3d machine vision metrology using off-the-shelf tv cameras and lenses. IEEE Journal on Robotics and Automation 3(4), 323-344 (1987)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 249, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 270 + ], + "type": "text", + "content": "38. Ventura, J.: Structure from motion on a sphere. In: European Conference on Computer Vision (ECCV). vol. 9907, pp. 53-68 (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 271, + 482, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 314 + ], + "type": "text", + "content": "39. Wang, C., Liu, Y., Wang, Y., Li, X., Wang, M.: Efficient and outlier-robust simultaneous pose and correspondence determination by branch-and-bound and transformation decomposition. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(10), 6924-6938 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 315, + 482, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 346 + ], + "type": "text", + "content": "40. Wang, Z., Wu, Z., Zhen, X., Yang, R., Xi, J., Chen, X.: A two-step calibration method of a largefov binocular stereovision sensor for onsite measurement. Measurement 62, 15-24 (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 347, + 482, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 380 + ], + "type": "text", + "content": "41. Wu, G., Han, B., He, X.: Calibration of geometric parameters of line array ccd camera based on exact measuring angle in lab. Optics and Precision Engineering 15(10), 1628-1632 (2007)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "type": "text", + "content": "42. Xiao, Z., Jin, L., Yu, D., Tang, Z.: A cross-target-based accurate calibration method of binocular stereo systems with large-scale field-of-view. Measurement 43(6), 747-754 (2010)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 413, + 482, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 445 + ], + "type": "text", + "content": "43. Yuan, G., Zheng, L., Ding, Y., Zhang, H., Zhang, X., Liu, X., Sun, J.: A precise calibration method for line scan cameras. IEEE Transactions on Instrumentation and Measurement 70, 1-9 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "type": "text", + "content": "44. Yuan, G., Zheng, L., Sun, J., Liu, X., Wang, X., Zhang, Z.: Practical calibration method for aerial mapping camera based on multiple pinhole collimator. IEEE Access 8, 39725-39733 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 479, + 482, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 511 + ], + "type": "text", + "content": "45. Zhang, K., Xie, J., Snavely, N., Chen, Q.: Depth sensing beyond lidar range. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1689-1697 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 512, + 482, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 482, + 533 + ], + "type": "text", + "content": "46. Zhang, Z.: A flexible new technique for camera calibration. IEEE Transactions on Pattern Analysis and Machine Intelligence 22(11), 1330-1334 (2000)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 533, + 482, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 556 + ], + "type": "text", + "content": "47. Zhang, Z.: Camera calibration with one-dimensional objects. IEEE Transactions on Pattern Analysis and Machine Intelligence 26(7), 892-899 (2004)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "text", + "content": "Camera Calibration using a Collimator System" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Camera Height Doesn't Change_ Unsupervised Training for Metric Monocular Road-Scene Depth Estimation/b1560f9a-1722-4ac2-a48d-d70d8859d7c8_content_list.json b/2024/Camera Height Doesn't Change_ Unsupervised Training for Metric Monocular Road-Scene Depth Estimation/b1560f9a-1722-4ac2-a48d-d70d8859d7c8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0ef0cf52ac451dda46cfc8c3bf2d1e6ff611265e --- /dev/null +++ b/2024/Camera Height Doesn't Change_ Unsupervised Training for Metric Monocular Road-Scene Depth Estimation/b1560f9a-1722-4ac2-a48d-d70d8859d7c8_content_list.json @@ -0,0 +1,1777 @@ +[ + { + "type": "text", + "text": "Camera Height Doesn't Change: Unsupervised Training for Metric Monocular Road-Scene Depth Estimation", + "text_level": 1, + "bbox": [ + 243, + 140, + 759, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Genki Kinoshita and Ko Nishino", + "bbox": [ + 370, + 234, + 633, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Graduate School of Informatics, Kyoto University, Kyoto, Japan \ngkinoshita@vision.ist.i.kyoto-u.ac.jp, kon@i.kyoto-u.ac.jp \nhttps://vision.ist.i.kyoto-u.ac.jp/", + "bbox": [ + 274, + 262, + 727, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this paper, we introduce a novel training method for making any monocular depth network learn absolute scale and estimate metric road-scene depth just from regular training data, i.e., driving videos. We refer to this training framework as FUMET. The key idea is to leverage cars found on the road as sources of scale supervision and to incorporate them in network training robustly. FUMET detects and estimates the sizes of cars in a frame and aggregates scale information extracted from them into an estimate of the camera height whose consistency across the entire video sequence is enforced as scale supervision. This realizes robust unsupervised training of any, otherwise scale-oblivious, monocular depth network so that they become not only scale-aware but also metric-accurate without the need for auxiliary sensors and extra supervision. Extensive experiments on the KITTI and the Cityscapes datasets show the effectiveness of FUMET, which achieves state-of-the-art accuracy. We also show that FUMET enables training on mixed datasets of different camera heights, which leads to larger-scale training and better generalization. Metric depth reconstruction is essential in any road-scene visual modeling, and FUMET democratizes its deployment by establishing the means to convert any model into a metric depth estimator.", + "bbox": [ + 261, + 340, + 743, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Unsupervised Monocular Depth Estimation $\\cdot$ Metric Scale $\\cdot$ Camera Height Invariance $\\cdot$ Object Size Prior", + "bbox": [ + 261, + 618, + 740, + 646 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 672, + 375, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Monocular depth estimation is essential for autonomous driving and advanced driver assistance systems (ADAS). It underlies many of their key perceptual tasks including 3D object detection [7,33,39,49], motion planning [31,38,51], and road segmentation [26]. Although supervised methods have achieved impressive accuracy, they suffer from costly ground-truth data collection. This limits their generalizability [34] and also makes continuous learning difficult.", + "bbox": [ + 212, + 703, + 787, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many recent methods [5, 15, 18, 25, 52] leverage self-supervision to scale up training. Most of these methods, however, suffer from scale ambiguity which renders them difficult to use off-the-shelf. To resolve the scale ambiguity and", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "recover in metric scale, the actual scene scale needs to be supervised in some form. Past works have leveraged velocity [16], GPS [3], IMU [41, 53], or camera height [36, 45] for this. These methods require auxiliary sensors to measure these extra quantities in addition to the monocular RGB camera, which significantly limits the ability to scale the size of training data. Arbitrary in-the-wild driving videos cannot be used as they lack such measurements. These self-supervised methods often still require weak scale supervision and cannot be made fully self-supervised as the scale changes for each sequence captured with a different camera.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Road scenes actually have plenty of untapped sources of metric scale. Cars found on the road are one of them as they are rigid objects whose actual sizes do not change and are unique to each make and model. The road is also exactly the region where we need accurate metric depth the most. If we can leverage the sizes of cars to bridge the 3D to 2D projection, we may train a mono-depth network to become metric-accurate. Simply employing a prior on the car size, however, would be too brittle since the metric supervision will be as ambiguous as the accuracy of that prior. The perspective projection and the variations in car makes, years, and colors, lead to a large variance in the size estimates, which will directly affect the mono-depth metric accuracy. In addition, there are no reliable object size priors except for a few introduced as probability densities for other tasks [20,54,57] whose parameters are not shared.", + "bbox": [ + 212, + 286, + 787, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "How can we train an arbitrary mono-depth estimator to become metric-accurate by robustly leveraging prior knowledge about the 3D sizes of cars found in driving videos? We derive fully unsupervised metric depth trainer (FUMET) to answer this question. Our key idea is to aggregate car size cues in the frames and the sequence into a single physical measure, namely the camera height. Regardless of the scene, the camera capturing it is the same for the same sequence. This means that, by transforming the car size cues found in different frames into a camera height estimate, the very fact that it should not change can be used as a supervision of invariance to achieve metric depth reconstruction. We can formulate this as an optimization of the camera height across frames and training epochs. This leads to gradual metric-aware learning that is stable and accurate, which also makes the depth estimates consistent across the sequence.", + "bbox": [ + 212, + 470, + 787, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce a learned size prior (LSP) to estimate the size of each car found in a frame. This prior gives us the vehicle dimensions from its appearance. It is trained with various augmentations on a large-scale dataset without the need for any manual annotation. By comparing these dimension estimates with those computed from the depth estimates, we obtain a per-frame scale factor. The multiplication of this scale estimate with the camera height estimate derived from the estimated depth becomes the camera height estimate of the frame. These estimated camera heights across all frames in a sequence are then consolidated by using their median value which is then used as scale supervision with its weighted moving average at the end of each training epoch.", + "bbox": [ + 212, + 655, + 787, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate the effectiveness of FUMET with extensive experiments using the KITTI and the Cityscapes datasets and compare its accuracy with a number", + "bbox": [ + 214, + 809, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "G.Kinoshita et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of weakly-supervised methods. The results clearly show that FUMET succeeds in making monocular depth estimation methods learn metric scales and achieves the state-of-the-art accuracy without any direct depth and scale supervision. We also demonstrate its capability to train a model on mixed datasets collected with different cameras, i.e., camera heights. This enables larger-scale training and leads to higher generalization. We believe FUMET can serve as a fundamental building block for monocular depth estimation in the wild.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 214, + 273, + 382, + 289 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Self-Supervision", + "text_level": 1, + "bbox": [ + 214, + 304, + 395, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although, as in other tasks, supervised methods achieve the highest accuracy in monocular depth estimation (MDE) [1, 23, 27, 29, 50], they obviously require ground-truth data usually obtained with a LiDAR sensor. Self-supervised MDE methods remove the need for such hardware costs for training. Garg et al. [11] proposed the first approach to self-supervised MDE by predicting per-pixel disparities between a stereo pair. Godard et al. [14] extend this idea by introducing left-right consistency and greatly improve the accuracy. These methods require calibrated stereo data for training which limits their applicability. Zhou et al. [56] introduced a pioneering method for training solely from monocular videos by incorporating a network that estimates the relative camera poses between successive frames. Based on this idea, many MDE methods that can be trained solely with monocular videos have been proposed [5, 15, 18, 25, 52, 55].", + "bbox": [ + 212, + 328, + 787, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite the popularity and improvements of self-supervised MDE, directly deploying those models in autonomous driving or ADAS systems is unrealistic as they only produce relative depth up to an unknown scale. Our FUMET overcomes this issue without the need for any external supervision.", + "bbox": [ + 212, + 508, + 785, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Weak Supervision", + "text_level": 1, + "bbox": [ + 214, + 590, + 411, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several works have tackled scale-aware MDE by adopting weak scale supervision through velocity [16], GPS [3], IMU [41], or camera height [36,45] measurements. Zhang et al. [53] use the combination of IMU, velocity, and gravity to achieve the highest accuracy among these methods. These methods fundamentally rely on such auxiliary measurements in addition to regular RGB images. This makes it impossible to take advantage of the many road-scene videos on the Internet (e.g., YouTube) or casually installed dashcam videos for training. In contrast, FUMET relies on a universal car size prior which is independent of the data collection environments and does not require any such weak supervision; it can be trained solely from monocular video sequences.", + "bbox": [ + 212, + 613, + 787, + 765 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3 Object Size Priors", + "text_level": 1, + "bbox": [ + 214, + 785, + 415, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several works for other tasks leverage priors on object sizes including single view metrology [57], object insertion [20,54], and visual odometry [10,35]. These", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Unsupervised Training for Metric Monocular Road-Scene Depth Estimation", + "bbox": [ + 227, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "methods model the priors as normal distributions or Gaussian mixture models, and the parameters of these distributions are predetermined. When there is a gap between those values and the true ones, which is usually the case in the real world, the difference directly surfaces in erroneous scale estimates.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead of these fixed priors, we introduce a learning-based size prior which we refer to as LSP. This estimates vehicle dimensions based on its appearance, thus can avoid the aforementioned problem. Furthermore, it is trained on a large-scale dataset with various augmentations, so that it can robustly estimate vehicle dimensions from any image.", + "bbox": [ + 212, + 208, + 787, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 310, + 382, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to train a metric-accurate MDE model $\\theta_{\\mathrm{depth}}$ with only self-supervision. This self-supervision is comprised of two parts: one designed for learning metric scale (Sec. 4) and the other dedicated to learning depth geometry. For the latter, we follow Zhou et al. [56] and minimize the reconstruction loss between the current frame $I_{t}$ and the ones synthesized from temporally adjacent frames $I_{s}$ , $(s = t \\pm 1)$ . We employ a pose network $\\theta_{\\mathrm{pose}}$ in addition to $\\theta_{\\mathrm{depth}}$ to estimate the relative camera pose $\\mathbf{T}_{t \\to s} \\in \\mathrm{SE}(3)$ and the current dense depth map $D_{t}$ , respectively,", + "bbox": [ + 212, + 345, + 784, + 465 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {t} = \\theta_ {\\text {d e p t h}} \\left(I _ {t}\\right), \\mathbf {T} _ {t \\rightarrow s} = \\theta_ {\\text {p o s e}} \\left(I _ {t}, I _ {s}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 468, + 785, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then synthesize the image from the current viewpoint", + "bbox": [ + 215, + 492, + 630, + 508 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {s \\rightarrow t} = I _ {s} \\left\\langle \\operatorname {p r o j} \\left(D _ {t}, \\mathbf {T} _ {t \\rightarrow s}, \\mathbf {K}\\right)\\right\\rangle , \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 521, + 784, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\langle \\cdot \\rangle$ is a bilinear sampling operator, $proj(\\cdot)$ outputs the 2D coordinates of $D_{t}$ when projected into the viewpoint of $I_{s}$ , and $\\mathbf{K}$ is the camera intrinsic. Following Godard et al. [15], we use an affine combination of SSIM [42] and L1 loss as the photometric error $pe$ and define the reconstruction loss $\\mathcal{L}_{\\mathrm{rec}}$ as the photometric error between $I_{t}$ and $I_{s\\rightarrow t}$ for the pixels with minimum error across source frames", + "bbox": [ + 212, + 550, + 787, + 641 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c}} = \\min _ {s} p e \\left(I _ {t}, I _ {s \\rightarrow t}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 295, + 656, + 784, + 675 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np e \\left(I _ {t}, I _ {s \\rightarrow t}\\right) = \\frac {\\lambda}{2} \\left(1 - \\operatorname {S S I M} \\left(I _ {t}, I _ {s \\rightarrow t}\\right)\\right) + (1 - \\lambda) \\left| I _ {t} - I _ {s \\rightarrow t} \\right|, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 294, + 680, + 784, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda = 0.85$ . We also employ an edge-aware smoothness loss [14]", + "bbox": [ + 215, + 719, + 702, + 736 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s m}} = \\left| \\partial_ {x} d _ {t} ^ {*} \\right| e ^ {- \\left| \\partial_ {x} I _ {t} \\right|} + \\left| \\partial_ {y} d _ {t} ^ {*} \\right| e ^ {- \\left| \\partial_ {y} I _ {t} \\right|}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 369, + 747, + 784, + 766 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d_{t}^{*} = d_{t} / \\bar{d}_{t}$ is the mean-normalized inverse depth [37] to prevent the estimated depth from shrinking. Furthermore, when computing $\\mathcal{L}_{\\mathrm{rec}}$ we use automasking [15] to mask out stationary pixels which generate incorrect supervisory signals.", + "bbox": [ + 212, + 777, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "G.Kinoshita et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0f6dd62890ff3d85b780da962060a8c8ac07d2a640ae93d3bee87fc2121218aa.jpg", + "image_caption": [ + "Fig. 1: Overview of FUMET. At each training step $n$ , an unscaled camera height $H_{\\mathrm{cam}}^{n,\\tau}$ is computed differentiably from the estimated depth. The previous epoch $\\tau - 1$ provides supervision with a scaled camera height $H_{\\mathrm{cam}}^{*\\tau - 1}$ . To obtain this scaled camera height supervision, Silhouette Projector first computes the object silhouette heights $H_{\\mathrm{obj}}'$ from the depth map. By comparing $H_{\\mathrm{obj}}'$ and the estimated one $H_{\\mathrm{obj}}$ with LSP, per-frame scale factor $s$ is determined and we obtain the scaled camera height $H_{\\mathrm{cam}}^{n,\\tau} = s \\cdot H_{\\mathrm{cam}}'^{\\prime,\\tau}$ . At the end of each epoch, $H_{\\mathrm{cam}}^*$ is optimized across a series of consecutive frames and updated with the weighted moving average.", + "Fig. 1 depicts the overall architecture of FUMET. Please see the caption for an overview of the framework. Let “'” denote “not metrically scaled”." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 787, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 FUMET", + "text_level": 1, + "bbox": [ + 215, + 406, + 334, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Scale-Aware Self-Supervised Learning", + "text_level": 1, + "bbox": [ + 215, + 491, + 571, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In FUMET, we derive a camera height pseudo-supervision from depth estimates. To obtain the per-frame camera height, we first compute the normal vector for each pixel $p_i$ . We consider the 8-neighborhood of $p_i$ and form 8 pairs $\\mathcal{N}(p_i) = \\{(p_{j_0}^i,p_{j_1}^i)\\}_{j = 1}^8$ . Each pair $(p_{j_0}^i,p_{j_1}^i)$ consists of two pixels whose difference vectors w.r.t. $p_i$ are perpendicular to each other and are sorted counterclockwise. We reproject the paired pixels onto 3D space and compute the cross-product of the two difference vectors w.r.t. the reprojected 3D point of $p_i$ . By summing and normalizing all the cross-product vectors of each pair, we obtain the normal vector $\\hat{\\mathbf{n}}(p_i)$", + "bbox": [ + 212, + 516, + 787, + 652 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {n} \\left(p _ {i}\\right) = \\sum_ {\\mathcal {N} \\left(p _ {i}\\right)} \\left(\\phi \\left(p _ {j _ {0}} ^ {i}\\right) - \\phi \\left(p _ {i}\\right)\\right) \\times \\left(\\phi \\left(p _ {j _ {1}} ^ {i}\\right) - \\phi \\left(p _ {i}\\right)\\right), \\quad \\hat {\\mathbf {n}} \\left(p _ {i}\\right) = \\frac {\\mathbf {n} \\left(p _ {i}\\right)}{\\| \\mathbf {n} \\left(p _ {i}\\right) \\|}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 259, + 662, + 785, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\phi (\\cdot)$ denotes perspective reprojection.", + "bbox": [ + 215, + 714, + 535, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For pixels in the road region, the negated inner product between the normal vector and the reprojected 3D point corresponds to the camera height from the ground plane", + "bbox": [ + 212, + 729, + 784, + 773 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH _ {\\mathrm {c a m}} ^ {\\prime} (p _ {i}) = - \\phi (p _ {i}) \\cdot \\hat {\\mathbf {n}} (p _ {i}). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 400, + 773, + 785, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We represent the median of $H_{\\mathrm{cam}}'(p_i)$ as the camera height of the frame", + "bbox": [ + 215, + 797, + 732, + 813 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {H} _ {\\mathrm {c a m}} ^ {\\prime} = \\operatorname {m e d} \\left(\\left\\{H _ {\\mathrm {c a m}} ^ {\\prime} \\left(p _ {i}\\right) \\mid p _ {i} \\in M _ {\\mathrm {r}} \\right\\}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 369, + 823, + 785, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Unsupervised Training for Metric Monocular Road-Scene Depth Estimation", + "bbox": [ + 227, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p_i$ is in the road region mask $M_{\\mathrm{r}}$ obtained with an off-the-shelf semantic segmentation model [22] before training. This camera height $\\tilde{H}_{\\mathrm{cam}}'$ estimate is unscaled and needs to be upgraded to a real-scaled value to impose the invariance loss across frames. By using the Silhouette Projector, we aggregate the scale information from the object height prior into a per-frame scale factor $s$ (see Sec. 4.2) and obtain the scaled camera height as $H_{\\mathrm{cam}} = s \\cdot \\tilde{H}_{\\mathrm{cam}}'$ .", + "bbox": [ + 212, + 146, + 787, + 237 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We leverage the fact that the camera height does not change in a sequence and formulate the invariance into an optimization across frames and training epochs. Thanks to this optimization, we acquire an accurate and stable supervision, which is also consistent across the scene. At the end of the $\\tau$ -th epoch, we use the median of $\\{H_{\\mathrm{cam}}^{n,\\tau}\\}_{n=1}^{N}$ for all frames $N$ as the representative camera height $H_{\\mathrm{cam}}^{\\tau}$ at the $\\tau$ -th epoch. Then, we update the scaled camera height pseudo-invariance $H_{\\mathrm{cam}}^{*}$ with its weighted moving average", + "bbox": [ + 212, + 237, + 787, + 344 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nH _ {\\mathrm {c a m}} ^ {* \\tau} \\leftarrow \\left\\{\\frac {\\tau (\\tau - 1)}{2} \\cdot H _ {\\mathrm {c a m}} ^ {* \\tau - 1} + \\tau \\cdot H _ {\\mathrm {c a m}} ^ {\\tau} \\right\\} / \\frac {\\tau (\\tau + 1)}{2}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 356, + 785, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When training on a dataset containing images captured at multiple camera heights, we create pseudo-labels corresponding to each camera height and optimize them individually.", + "bbox": [ + 212, + 398, + 785, + 443 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To make the network learn the absolute scale and directly produce the scaled depth map, we enforce a loss on the camera height $H_{\\mathrm{cam}}'$ using the pseudo-invariance $H_{\\mathrm{cam}}^{*\\tau - 1}$ computed in the previous epoch", + "bbox": [ + 212, + 444, + 787, + 491 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c a m}} = \\frac {1}{| M _ {\\mathrm {r}} |} \\sum_ {i} M _ {\\mathrm {r}} \\left(p _ {i}\\right) \\cdot \\left| H _ {\\mathrm {c a m}} ^ {\\prime} \\left(p _ {i}\\right) - H _ {\\mathrm {c a m}} ^ {* \\tau - 1} \\right|. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 502, + 785, + 535 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It is worth noting that at inference, the network does not assume camera height invariance in the input sequence; it is only assumed during training. The computational cost at inference is exactly the same as the original MDE model trained without FUMET.", + "bbox": [ + 212, + 547, + 787, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Silhouette Projector", + "text_level": 1, + "bbox": [ + 214, + 631, + 431, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the process of creating a scaled camera height supervision, we need to know the per-frame scale factor. For this, we can consider leveraging an object of known size. This, however, would necessitate extraction of the exact 3D object region which is non-trivial as objects are often occluded or cropped and oriented in various poses.", + "bbox": [ + 212, + 657, + 787, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We introduce Silhouette Projector to robustly estimate this per-frame scale factor. Silhouette Projector leverages two simple facts of road scenes. First, the height of the object silhouette projected on the plane perpendicular to the ground plane does not change regardless of the object pose. Second, we can compute the silhouette height by measuring the distance from the top of the silhouette to the ground plane even if they are partly occluded or truncated, as long as the top is visible and the object is on the ground plane.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "G.Kinoshita et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first reconstruct an object point cloud with the estimated depth map through perspective reprojection. The object region $M_{\\mathrm{obj}}$ in the image can be obtained with an off-the-shelf instance segmentation model [22] before training. We then orthographically project these 3D points onto a plane $P^{\\perp}$ perpendicular to the ground plane, i.e., an arbitrary plane parallel to the road normal vector $\\tilde{\\mathbf{n}}$ . We compute $\\tilde{\\mathbf{n}}$ as the median of the normal vectors $\\hat{\\mathbf{n}}(p_i)$ (see Eq. (6)) in the road region $M_{\\mathrm{r}}$ . We define the farthest distance from the 2D point on $P^{\\perp}$ to the ground plane as the silhouette height. This value is unscaled because we derived this from the estimated unscaled depth map. The ratio between a height prior and its silhouette height corresponds to the scale factor of interest. When there are multiple objects in an image, we use the median of the scale factors each corresponding to each object as the scale factor of the frame for computational simplicity and robustness.", + "bbox": [ + 212, + 146, + 787, + 344 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Outlier Filtering. Some objects cannot be handled well with Silhouette Projector for various reasons including occluded upper parts, not on the ground plane, inaccurate segmented region, and misclassified object category. We introduce an algorithm to detect these as outliers based on geometric plausibility. We first calculate the horizon $\\mathbf{l}$ with the road normal vector $\\tilde{\\mathbf{n}}$ and the camera intrinsics $\\mathbf{K}$ [17]", + "bbox": [ + 212, + 369, + 787, + 460 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {l} = \\mathbf {K} ^ {- \\top} \\tilde {\\mathbf {n}}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 460, + 463, + 784, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on the approximations in [20], we derive the approximate metric-scale object height $\\hat{H}_{\\mathrm{obj}} \\approx \\frac{h_{\\mathrm{obj}}}{h_{\\mathrm{cam}}} H_{\\mathrm{cam}}^{*\\tau - 1}$ , where $h_{\\mathrm{cam}}$ is the farthest distance from the pixel in the object instance mask to $\\mathbf{l}$ , $h_{\\mathrm{obj}}$ is the farthest distance between two points on the object mask on which tangent lines parallel to $\\mathbf{l}$ lie, and $H_{\\mathrm{cam}}^{*\\tau - 1}$ is the pseudo camera height at the $\\tau$ -th epoch derived in Eq. (9). By comparing the relative gap between $\\hat{H}_{\\mathrm{obj}}$ and the object height prior $H_{\\mathrm{obj}}$ , we determine the outliers as", + "bbox": [ + 212, + 489, + 787, + 595 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{k \\left| \\frac {\\left| H _ {\\mathrm {o b j}} ^ {k} - \\hat {H} _ {\\mathrm {o b j}} ^ {k} \\right|}{H _ {\\mathrm {o b j}} ^ {k}} > \\mathcal {T} \\right. \\right\\}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 598, + 784, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In all the experiments, we set the threshold $\\mathcal{T}$ to 0.2.", + "bbox": [ + 214, + 645, + 599, + 659 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Training Losses", + "text_level": 1, + "bbox": [ + 214, + 688, + 390, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Auxiliary Rough Geometric Loss. We experimentally observe that estimated object depths, especially those distant from the camera, often erroneously lean towards the road surface because of the inaccurate supervisory signal from the reconstruction loss in low-texture regions. This tendency makes silhouette heights inaccurate. To mitigate this issue, we introduce an auxiliary rough geometric loss $\\mathcal{L}_{\\mathrm{aux}}$ , which enforces object point clouds to be within a close region. For each inlier object $k$ ( $k = 1 \\dots K$ ), we define $\\mathcal{L}_{\\mathrm{aux}}$ as the gap between the estimated depth map $D$ in the object region $M_{\\mathrm{obj}}$ and the approximate object", + "bbox": [ + 212, + 719, + 787, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Unsupervised Training for Metric Monocular Road-Scene Depth Estimation", + "bbox": [ + 227, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "depth $D_{\\mathrm{aprx}} = \\frac{H_{\\mathrm{obj}}}{h_{\\mathrm{obj}}} f_y$ where $f_{y}$ is the focal length", + "bbox": [ + 212, + 143, + 576, + 165 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a u x}} = \\frac {1}{K} \\sum_ {k} \\frac {1}{| M _ {\\mathrm {o b j}} ^ {k} |} \\sum_ {i} M _ {\\mathrm {o b j}} ^ {k} \\left(p _ {i}\\right) \\cdot \\left| D \\left(p _ {i}\\right) - D _ {\\mathrm {a p r x}} ^ {k} \\right|. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 176, + 785, + 214 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As a by-product, this loss accelerates learning the metric scale.", + "bbox": [ + 215, + 224, + 669, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The $\\mathcal{L}_{\\mathrm{aux}}$ is potentially inaccurate since $D_{\\mathrm{aprx}}$ assumes that each object can be represented by a plane parallel to the image plane, and we observe that it degrades the depth accuracy in the late stage of training. On the other hand, in the early stage, estimated depth is unreliable, so heavily relying on the camera height loss $\\mathcal{L}_{\\mathrm{cam}}$ makes the training unstable. For this, we gradually decrease the weight of $\\mathcal{L}_{\\mathrm{aux}}$ and increase the one for $\\mathcal{L}_{\\mathrm{cam}}$ as the training proceeds in a logarithmic way. To ensure training stability, we stop changing these weights after the $\\tau_{\\mathrm{mid-th}}$ epoch", + "bbox": [ + 212, + 239, + 787, + 359 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {\\mathrm {a u x}} (\\tau) = \\left\\{ \\begin{array}{l l} \\frac {- \\log (\\tau + 1)}{\\log \\left(\\tau_ {\\mathrm {m i d}} + 1\\right)} & (\\tau \\leq \\tau_ {\\mathrm {m i d}}) \\\\ \\epsilon & (\\tau > \\tau_ {\\mathrm {m i d}}) \\end{array} , \\lambda_ {\\mathrm {c a m}} (\\tau) = \\left\\{ \\begin{array}{l l} \\frac {\\log (\\tau + 1)}{\\log \\left(\\tau_ {\\mathrm {m i d}} + 1\\right)} & (\\tau \\leq \\tau_ {\\mathrm {m i d}}) \\\\ 1. 0 & (\\tau > \\tau_ {\\mathrm {m i d}}) \\end{array} , \\right. \\right. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 372, + 785, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\epsilon = 0.005$", + "bbox": [ + 215, + 422, + 336, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The total loss is $\\mathcal{L} = \\alpha \\lambda_{\\mathrm{cam}}\\mathcal{L}_{\\mathrm{cam}} + \\beta \\lambda_{\\mathrm{aux}}\\mathcal{L}_{\\mathrm{aux}} + \\mathcal{L}_{\\mathrm{sm}} + \\mathcal{L}_{\\mathrm{rec}}$ where $\\alpha = 0.01$ and $\\beta = 1.0$ .", + "bbox": [ + 212, + 436, + 784, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 LSP", + "text_level": 1, + "bbox": [ + 215, + 489, + 294, + 506 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our FUMET extracts metric scale from a prior on object heights through the camera height loss $\\mathcal{L}_{\\mathrm{cam}}$ and the auxiliary rough geometric loss $\\mathcal{L}_{\\mathrm{aux}}$ . To ensure that FUMET learns the scale regardless of the particular dataset used for training, a universal prior of object heights, independent of the environment, becomes essential. Cars are objects whose dimensions are explicitly and accurately known. Better yet, they are abundantly located on the road, the key semantic region whose metric depth is essential for many downstream tasks. Some methods for other tasks have modeled the size prior as a probability distribution [20, 57]. The gap between its mean and the true expectation for each environment, however, is directly reflected in the uncertainty of the scale in the supervision. Wang et al. [40] also report that there is a gap in car dimensions among datasets. These facts suggest that modeling the object size prior as a single probability distribution is problematic, especially for man-made objects like cars.", + "bbox": [ + 212, + 522, + 787, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We humans not only possess rough prior knowledge about the vehicle size but can also estimate it more accurately by extracting instance-specific information such as car models from its appearance. Motivated by this intuition, we introduce a learned universal vehicle size prior termed LSP. Given a masked vehicle image, LSP estimates its dimensions.", + "bbox": [ + 212, + 719, + 787, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For training, we use a dataset consisting of a large number of vehicle images and their information such as 3D dimensions. This dataset can be easily collected with web scraping and does not require manual annotations. To enhance", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "G.Kinoshita et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the robustness to various input images, in addition to common color jittering and blur augmentation, we perform augmentations to simulate occlusion and truncation. For occlusion augmentation, we mask out random regions with the same shape as a vehicle or with a beam of random width. For truncation augmentation, we translate the vehicle mask in a random direction. We visualize these in the supplementary material. The abundance of training images and diverse augmentations enable LSP to estimate vehicle dimensions robustly across various datasets.", + "bbox": [ + 212, + 146, + 782, + 265 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Although FUMET requires only a car height prior, we construct LSP to estimate the width and length as well as height. This enhances it to grasp an overall car size and improves the prediction accuracy of car heights.", + "bbox": [ + 212, + 267, + 784, + 313 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6 Experiments", + "text_level": 1, + "bbox": [ + 214, + 334, + 375, + 352 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6.1 FUMET", + "text_level": 1, + "bbox": [ + 214, + 364, + 333, + 378 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluate the accuracy of depth estimators trained on our framework and demonstrate the effectiveness of FUMET as a novel unsupervised metric-scale training method through depth prediction accuracy, its applicability to any network architecture, and its ability to leverage any dataset for training.", + "bbox": [ + 212, + 388, + 784, + 450 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. We implement FUMET with PyTorch [28]. We employ Monodepth2 [15] with ResNet50 encoder pre-trained on ImageNet [30] as the default depth estimator which is one of the most basic architectures in self-supervised MDE. For all experiments unless mentioned otherwise, we train the model for 50 epochs with a batch size of 8 and set $\\tau_{\\mathrm{mid}}$ to 20. We use Adam optimizer [24] with an initial learning rate of $5.0\\times 10^{-5}$ decayed by half every 15 epochs. During training, we perform the following augmentations with $50\\%$ chance in random order: horizontal flips, brightness adjustment $(\\pm 0.2)$ , saturation adjustment $(\\pm 0.2)$ , contrast adjustment $(\\pm 0.2)$ , and hue jitter $(\\pm 0.1)$ . For quantitative evaluation, we follow [9] and compute the seven standard metrics: AbsRel, SqRel, RMSE, $\\mathrm{RMSE}_{\\log}$ , $\\delta < 1.25$ , $\\delta < 1.25^2$ , and $\\delta < 1.25^3$ .", + "bbox": [ + 212, + 470, + 787, + 638 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "KITTI. Our FUMET is independent of network architectures so we experiment on several models with FUMET as well as Monodepth2: Lite-Mono [52] with ResNet18, HR-Depth [25] with ResNet18, and VADepth [45]. We train these models on the KITTI raw dataset [12], which contains 39,810 images for training and 4,424 for validation, and evaluate it with the KITTI Eigen test split [8] (697 images). For training and evaluation, all images are resized to $640 \\times 192$ .", + "bbox": [ + 212, + 657, + 782, + 748 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We show the qualitative results in Fig. 2. The quantitative results shown in Tab. 1 demonstrate that FUMET enables the various architectures to learn metric scale properly. Even when employing the simplest network architecture, i.e., Monodepth2, it outperforms weakly-supervised methods that rely on ground-truth scale labels for training which is not necessary for our method. Similar to VADepth [45], FUMET uses the camera height invariance as supervision though", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Unsupervised Training for Metric Monocular Road-Scene Depth Estimation", + "bbox": [ + 227, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/95fcf9ea43a3540a38e7735e083de082977449474786237020b9ba2732f03e76.jpg", + "table_caption": [ + "Table 1: Quantitative results on the KITTI Eigen test split [8] ( $640 \\times 192$ resolution). In the Supervision column, V, G, and CamH represent velocity, gravity, and camera height, respectively. The Scaling column represents whether to perform the median scaling [56]. We denote the best results in bold and the second best ones with an underline for each block. Note that we train VADepth [45] from scratch using ground-truth camera height (1.65m) provided by KITTI, unlike its original paper. FUMET outperforms weakly-supervised methods regardless of the network architectures and without the need for ground truth supervision. In the rescaled results, every model trained with FUMET achieves better results than its original one." + ], + "table_footnote": [], + "table_body": "
MethodSupervisionScalingError (↓)Accuracy (↑)
AbsRelSqRelRMSERMSElogδ<1.25δ<1.252δ<1.253
G2S [3]GPS-0.1090.8604.8550.1980.8650.9540.980
PackNet-SfM [16]V-0.1110.8294.7880.1990.8640.9540.980
Wagstaff et al. [36]CamH-0.1230.9965.2530.2130.8400.9470.978
VADepth [45]CamH-0.1200.9754.9710.2030.8670.9560.979
DynaDepth [53]IMU+V+G-0.1090.7874.7050.1950.8690.9580.981
Ours--0.1080.7854.7360.1950.8710.9580.981
Lite-Mono [52]+Ours--0.1120.7984.6920.1940.8710.9590.981
HR-Depth [25]+Ours--0.1090.7944.7340.1930.8690.9590.982
VADepth [45]+Ours--0.1080.8094.5720.1850.8830.9630.982
G2S [3]GPS✓0.1120.8944.8520.1920.8770.9580.981
PackNet-SfM [16]V✓0.1110.7854.6010.1890.8780.9600.982
Wagstaff et al. [36]CamH✓0.1170.9524.9890.1970.8670.9570.981
VADepth [45]CamH✓0.1100.9774.8720.1870.8890.9620.981
DynaDepth [53]IMU+V+G✓0.1080.7614.6080.1870.8830.9620.982
Ours-✓0.1060.7594.6020.1840.8870.9630.983
Monodepth2 [15]-✓0.1150.9034.8630.1930.8770.9590.981
+ Ours-✓0.1060.7594.6020.1840.8870.9630.983
HR-Depth [25]-✓0.1090.7924.6320.1850.8840.9620.983
+ Ours-✓0.1080.7594.5840.1840.8850.9630.983
Lite-Mono [52]-✓0.1100.8024.6710.1860.8790.9610.982
+ Ours-✓0.1070.7754.5640.1830.8880.9630.983
VADepth [45]-✓0.1040.7744.5520.1810.8920.9650.982
+ Ours-✓0.1020.7604.4730.1780.8970.9660.983
", + "bbox": [ + 217, + 282, + 782, + 592 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "VADepth requires the ground-truth camera height and employs a complex network architecture. FUMET with Monodepth2 (row 6 in Tab. 1) achieves higher accuracy. This indicates that measuring camera height precisely is difficult and using unreliable ground-truth camera height causes low accuracy in the scale of estimated depth maps.", + "bbox": [ + 212, + 645, + 787, + 720 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In the results showing the accuracy of the estimated depth map scaled with ground-truth scale factors via median scaling [56], it is evident that every model trained with FUMET attains superior results to the one trained without FUMET. This shows that our training framework not only enables the models to learn metric scale but also enhances the geometric accuracy. This improvement owes to the camera height loss $\\mathcal{L}_{\\mathrm{cam}}$ and $\\mathcal{L}_{\\mathrm{aux}}$ , which makes the model respect road regions and object regions, respectively.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "G.Kinoshita et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/1ae246018d255e87deaff77763c04c403f7ba641eb1140f2fafdf86d6dc1d768.jpg", + "image_caption": [ + "Fig. 2: Qualitative comparison on KITTI. In error maps, the larger depth errors are represented in red, smaller ones are in blue. The model trained with our FUMET predicts more accurate depth maps compared to the weakly-supervised methods." + ], + "image_footnote": [], + "bbox": [ + 217, + 161, + 781, + 366 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "FUMET successfully enables a model to learn metric scale, even though the training dataset of KITTI contains many frames with no or few detected cars. We detail this in the supplementary material.", + "bbox": [ + 212, + 458, + 787, + 505 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Cityscapes. To evaluate the ability of FUMET to leverage any dataset for training, we train and evaluate it on the Cityscapes dataset [6]. We follow [43, 46, 48, 56] and collect 69,731 images for training and evaluate on the 1,525 test images. In this dataset, there are multiple sets of camera intrinsics. Since this strongly affects the metric depth estimation [47], we first align the focal lengths of all images to 587.5 by resizing them, and crop them to $512 \\times 192$ resolution. For evaluation, we follow [2, 43] and further crop the estimated image into $416 \\times 128$ . In order to align the number of iterations on KITTI, we train for 29 epochs with the same learning rate decreased by half every 8 epochs, and $\\tau_{\\mathrm{mid}}$ is set to 12.", + "bbox": [ + 212, + 527, + 787, + 664 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tab. 2 shows the quantitative results and FUMET outperforms weakly-supervised methods by wide margins. These methods assume that highly accurate sensor data is available and fail to train scale-aware MDE models on datasets consisting of unreliable sensor data such as Cityscapes. On the other hand, FUMET does not require any sensor measurement other than RGB video, so it can robustly leverage any driving videos for training.", + "bbox": [ + 212, + 665, + 787, + 755 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Mixed datasets. To demonstrate that FUMET can train a model on a mixed set of datasets to attain strong generalizability, we train with 235,341 images collected from the Argoverse2 [44], Lyft [21], A2D2 [13], and DDAD [16] dataset and evaluate it on the KITTI Eigen test split [8]. Both images for training and", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Unsupervised Training for Metric Monocular Road-Scene Depth Estimation", + "bbox": [ + 227, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/e8b0b2119c61b3f15e71387f08c49f20039b50583335c768c484b996aaaa6e5a.jpg", + "table_caption": [ + "Table 2: Quantitative results on Cityscapes [6]. We train all the models from scratch with the same supervisions as the Supervision column in Tab. 1. FUMET significantly outperforms weakly-supervised methods." + ], + "table_footnote": [], + "table_body": "
Method|AbsRelSqRelRMSEδ<1.25
G2S [3]4.156276.1658.890.046
PackNet-SfM [16]0.5046.63914.900.029
VADepth [45]0.3637.11511.950.295
FUMET0.1251.2886.3590.858
", + "bbox": [ + 220, + 239, + 486, + 314 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/aedfdb04c55d4b83ffb2a77a3137ac06788f07234c33cbdd75b37b0b16b1cff2.jpg", + "table_caption": [ + "Table 3: FUMET can be trained on mixed datasets (Argoverse2, Lyft, A2D2, and DDAD) with various camera settings, and attains high generalizability. The results are evaluated on the KITTI Eigen split [8] $(832\\times 512$ resolution)." + ], + "table_footnote": [], + "table_body": "
Training datasetAbsRelSqRelRMSEδ<1.25
KITTI0.1030.6754.7080.903
Mixed0.1130.9165.0090.883
Mixed+KITTI0.0820.6114.3070.923
", + "bbox": [ + 511, + 239, + 784, + 297 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "evaluation are resized and cropped to $832 \\times 512$ resolution. The training images are the sets of video sequences and each sequence is captured with its own unique camera setting. We individually optimize each camera height for each sequence during training. We detail the dataset composition and training settings in the supplementary material.", + "bbox": [ + 212, + 344, + 782, + 419 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "As shown in Tab. 3, the model trained with FUMET on the large-scale mixed dataset achieves comparable accuracy to the one trained on KITTI despite the domain gap and attains high generalizability. The model trained on both the mixed dataset and KITTI outperforms the model trained only on KITTI. This suggests that training on various images enhances accuracy which is the strength of fully self-supervised learning methods.", + "bbox": [ + 212, + 420, + 784, + 508 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We also show that FUMET can leverage in-the-wild videos for training by using camera intrinsics recovered with SLAM, e.g., COLMAP [32] in the supplementary material.", + "bbox": [ + 212, + 511, + 785, + 556 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation Studies on KITTI. To study the effect of the proposed components in FUMET, we conduct ablation studies on the training losses, balancing the loss weights, camera height optimization, and outlier filtering. The results are shown in Tab. 4. Thanks to the accuracy of LSP and the robustness of Silhouette Projector, the camera height loss makes a greater contribution to the depth accuracy than the auxiliary rough geometric loss. By introducing the camera height optimization, the model can learn the scale with more accurate and consistent supervision than leveraging the size prior independently for each frame (rows 1, 2, and 6 in Tab. 4), which leads to improved accuracy. Though balancing the loss weights (Eq. (14)) with either $\\mathcal{L}_{\\mathrm{aux}}$ or $\\mathcal{L}_{\\mathrm{cam}}$ degrades the results due to the insufficient strength of the supervisory signals for metric scale, it contributes to improving accuracy when both $\\mathcal{L}_{\\mathrm{aux}}$ and $\\mathcal{L}_{\\mathrm{cam}}$ are employed.", + "bbox": [ + 212, + 582, + 787, + 763 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Our proposed method can also be used with a pre-trained MDE model in an offline manner. We first train the original Monodepth2 [15] without FUMET, then predict depth maps of training images. With Silhouette Projector and LSP, we obtain per-frame scaled camera height and optimize it across frames with the median operation. We fix this value and use it as the camera height supervision", + "bbox": [ + 212, + 765, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "G.Kinoshita et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/298a50cc890b347e099a232bd8c4b5d570e841bdf446a514df813e2eaf395518.jpg", + "table_caption": [ + "Table 4: Each loss contributes to learning metric scale and the others enhance the depth accuracy. Bal: balancing the loss weights (Eq. (14)). Opt: camera height optimization. Flt: outlier filtering." + ], + "table_footnote": [], + "table_body": "
\\( \\mathcal{L}_{\\text{aux}} \\)\\( \\mathcal{L}_{\\text{cam}} \\)BalOptFltAbsRelSqRelRMSE\\( \\delta < 1.25 \\)
✓---✓0.1250.8875.0550.834
✓-✓-✓0.1490.9505.2190.785
-✓-✓✓0.1160.8754.8740.863
-✓✓✓✓0.2111.3036.1430.585
✓✓-✓✓0.1220.8805.0170.849
✓✓✓-✓0.1150.8224.8110.858
✓✓✓✓-0.1160.8234.9010.860
✓✓✓✓✓0.1080.7854.7360.871
", + "bbox": [ + 217, + 226, + 506, + 334 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/96c41be46c977c401190fed497ae037464701f57429324dd5421a3e2be55c7d9.jpg", + "table_caption": [ + "Table 5: Training with a fine-tuned pseudo camera height achieves the highest accuracy. Online: optimizing the camera height while training. Offline: training with a fixed one computed from depth maps predicted with pre-trained Monodepth2. Fine-tune: training with the fixed one computed offline for 20 epochs, then optimizing online." + ], + "table_footnote": [], + "table_body": "
Deriv. methodAbsRelSqRelRMSEδ<1.25
Online0.1080.8084.7400.870
Offline0.1060.7904.7190.873
Fine-tune0.1050.7824.6820.876
", + "bbox": [ + 522, + 282, + 781, + 339 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "for an untrained Monodepth2 on FUMET. We examine the difference between the proposed camera height optimized online and the fixed one derived offline when we use them as the pseudo-supervision. Note that we also fix $\\lambda_{\\mathrm{cam}}$ because the pseudo-label can be considered sufficiently accurate from the early stage of training. We also run another experiment in which we fix camera height supervision derived offline for the first 20 epochs, which is the same number of training epochs as the original Monodepth2, then unfit the camera height pseudo-label to fine-tune it via online optimization. As shown in Tab. 5, the model supervised with the fixed camera height computed offline achieves slightly better results than the one supervised with the camera height optimized online. This result indicates that our strategy of deriving scaled camera height is highly accurate and fixing it stabilizes the training. The model trained with the fine-tuned camera height achieves the highest accuracy, which shows that our FUMET can attain better accuracy with additional pre-training costs.", + "bbox": [ + 212, + 366, + 784, + 577 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.2 LSP", + "text_level": 1, + "bbox": [ + 215, + 598, + 295, + 611 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We provide implementation details of LSP in the supplementary material.", + "bbox": [ + 215, + 622, + 745, + 636 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Evaluation on KITTI and Cityscapes. We evaluate the generalizability of our LSP with the dimensions of the annotated 3D bounding box on both the KITTI [12] and the Cityscapes [6] datasets in the Car class. We compare the results of LSP to a fixed car height prior (1.59m, following [20]) and MonoCIinIS [19], a 3D object detection model that outputs object dimensions from an image and an instance mask. We train MonoCIinIS from scratch with the KITTI dataset. As an evaluation metric, we compute the absolute relative error for each dimension. When we estimate car dimensions with LSP and MonoCIinIS, we use a ground-truth instance mask for making a car image mask and as input, respectively. For all the experiments in LSP, input vehicle masks are cropped and resized to $300 \\times 300$ resolution. On the KITTI dataset, we extract 3,769 samples for evaluation [4]. On the Cityscapes dataset, we evaluate on the validation split,", + "bbox": [ + 212, + 657, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Unsupervised Training for Metric Monocular Road-Scene Depth Estimation", + "bbox": [ + 227, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/b8b23428d89e4835a8ecd35d9e30ed9ae75738701589483affbd6bd11d12d8be.jpg", + "table_caption": [ + "Table 6: Generalizability on KITTI and Cityscapes. We evaluate LSP on Absolute Relative Error for each dimension (height/width/length), comparing with the fixed car height prior (1.59 m, following [20]) and MonoCInIS [19] trained on KITTI. LSP robustly surpasses other methods in most cases and can predict accurately regardless of the datasets." + ], + "table_footnote": [], + "table_body": "
MethodEasyModerateHardAll
KITTIFixed Height [20]0.065 / - / -0.064 / - / -0.065 / - / -0.065 / - / -
MonoClIns [19]0.069 / 0.065 / 0.0930.112 / 0.106 / 0.1330.268 / 0.263 / 0.2690.159 / 0.155 / 0.177
LSP (only height)0.057 / - / -0.089 / - / -0.106 / - / -0.087 / - / -
LSP0.046 / 0.136 / 0.1300.060 / 0.144 / 0.1390.073 / 0.155 / 0.1500.062 / 0.147 / 0.145
CityscapesFixed Height [20]0.093 / - / -0.093 / - / -0.076 / - / -0.092 / - / -
MonoClIns [19]0.578 / 0.599 / 0.5840.759 / 0.766 / 0.7670.843 / 0.852 / 0.8520.851 / 0.859 / 0.856
LSP (only height)0.063 / - / -0.075 / - / -0.103 / - / -0.096 / - / -
LSP0.056 / 0.069 / 0.0670.068 / 0.068 / 0.0780.069 / 0.059 / 0.0750.080 / 0.064 / 0.077
", + "bbox": [ + 238, + 226, + 764, + 334 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "which has 500 samples. On the KITTI dataset, each object is separated into 3 levels based on detection difficulty: Easy, Moderate, and Hard. Considering the difference in focal lengths between KITTI and Cityscapes, we also separate objects on Cityscapes into the same levels as on KITTI.", + "bbox": [ + 212, + 366, + 782, + 426 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "As shown in Tab. 6, LSP outperforms the other two methods in most cases. Particularly, the accuracy of MonoCInIS drops on Cityscapes since it overfits to the training dataset, i.e., KITTI and cannot address the domain gap. In contrast, thanks to the large training data and various augmentations, our LSP obtains the highest generalizability and robustly estimates dimensions on any dataset. Learning not only car height but also width and length also improves the prediction accuracy as this facilitates estimation of the overall car sizes.", + "bbox": [ + 212, + 429, + 787, + 537 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 569, + 359, + 585 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we proposed FUMET, a novel training framework for any MDE model to learn metric scale. FUMET aggregates the scale information obtained from a learned car size prior into the camera height and uses it as pseudosupervision formulated as optimization across frames and epochs by leveraging its invariance. We showed that FUMET enables monocular depth estimators to embody metric scale regardless of the underlying network architectures and achieves state-of-the-art absolute-scale depth estimation accuracy on both the KITTI and the Cityscapes datasets, and demonstrated that it can leverage any dataset for training. We also proposed LSP, which can be easily trained on a large-scale dataset without manual annotations and robustly estimates vehicle dimensions from their appearances. We showed that the learned size prior is superior to a fixed one and outperforms a similar 3D object detection model. FUMET can be leveraged as a drop-in training scheme for any monocular depth estimation method. We hope it will be a versatile tool for road scene understanding research.", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "G.Kinoshita et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was in part supported by JSPS 20H05951 and 21H04893, JST JP-MJCR20G7 and JPMJAP2305, and RIKEN GRP.", + "bbox": [ + 215, + 178, + 784, + 208 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 233, + 321, + 250 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bhat, S.F., Alhashim, I., Wonka, P.: AdaBins: Depth Estimation Using Adaptive Bins. In: CVPR. pp. 4009-4018 (2021)", + "2. Casser, V., Pirk, S., Mahjourian, R., Angelova, A.: Unsupervised Monocular Depth and Ego-Motion Learning with Structure and Semantics. In: CVPRW. pp. 381-388 (2019)", + "3. Chawla, H., Varma, A., Arani, E., Zonooz, B.: Multimodal Scale Consistency and Awareness for Monocular Self-Supervised Depth Estimation. In: ICRA. pp. 5140-5146 (2021)", + "4. Chen, X., Kundu, K., Zhu, Y., Berneshawi, A.G., Ma, H., Fidler, S., Urtasun, R.: 3D Object Proposals for Accurate Object Class Detection. Adv. Neural Inform. Process. Syst. 28 (2015)", + "5. Chen, X., Li, T.H., Zhang, R., Li, G.: Frequency-Aware Self-Supervised Monocular Depth Estimation. In: WACV. pp. 5808-5817 (2023)", + "6. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The Cityscapes Dataset for Semantic Urban Scene Understanding. In: CVPR (2016)", + "7. Ding, M., Huo, Y., Yi, H., Wang, Z., Shi, J., Lu, Z., Luo, P.: Learning Depth-Guided Convolutions for Monocular 3D Object Detection. In: CVPR. pp. 11669-11678 (2020)", + "8. Eigen, D., Fergus, R.: Predicting Depth, Surface Normals and Semantic Labels with a Common Multi-Scale Convolutional Architecture. In: ICCV. pp. 3213-3223 (2015)", + "9. Eigen, D., Puhrsch, C., Fergus, R.: Depth Map Prediction from a Single Image Using a Multi-Scale Deep Network. In: Adv. Neural Inform. Process. Syst. vol. 2, pp. 2366-2374 (2014)", + "0. Frost, D., Prisacariu, V., Murray, D.: Recovering Stable Scale in Monocular SLAM Using Object-Supplemented Bundle Adjustment. IEEE Transactions on Robotics 34(3), 736-747 (2018). https://doi.org/10.1109/TR0.2018.2820722", + "1. Garg, R., Bg, V.K., Carneiro, G., Reid, I.: Unsupervised CNN for Single View Depth Estimation: Geometry to the Rescue. In: ECCV. pp. 740-756 (2016)", + "2. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision Meets Robotics: The KITTI Dataset. International Journal of Robotics Research (IJRR) 32(11), 1231--1237 (2013). https://doi.org/10.1177/0278364913491297", + "3. Geyer, J., Kassahun, Y., Mahmudi, M., Ricou, X., Durgesh, R., Chung, A.S., Hauswald, L., Pham, V.H., Mühlegg, M., Dorn, S., Fernandez, T., Janicke, M., Mirashi, S., Savani, C., Sturm, M., Vorobiov, O., Oelker, M., Garreis, S., Schuberth, P.: A2D2: Audi Autonomous Driving Dataset. CoRR abs/2004.06320 (2020)", + "4. Godard, C., Mac Aodha, O., Brostow, G.J.: Unsupervised Monocular Depth Estimation with Left-Right Consistency. In: CVPR. pp. 6602-6611 (2017)", + "5. Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging Into Self-Supervised Monocular Depth Estimation. In: ICCV. pp. 3827-3837 (2019)" + ], + "bbox": [ + 225, + 266, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Unsupervised Training for Metric Monocular Road-Scene Depth Estimation", + "bbox": [ + 225, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Guizilini, V., Ambrus, R., Pillai, S., Raventos, A., Gaidon, A.: 3D Packing for Self-Supervised Monocular Depth Estimation. In: CVPR. pp. 2482-2491 (2020)", + "17. Hartley, R., Zisserman, A.: Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edn. (2004)", + "18. He, M., Hui, L., Bian, Y., Ren, J., Xie, J., Yang, J.: RA-Depth: Resolution Adaptive Self-supervised Monocular Depth Estimation. In: ECCV. pp. 565–581 (2022)", + "19. Heylen, J., De Wolf, M., Dawagne, B., Proesmans, M., Van Gool, L., Abbeloos, W., Abdelkawy, H., Reino, D.O.: MonoCInIS: Camera Independent Monocular 3D Object Detection Using Instance Segmentation. In: ICCVW. pp. 923-934 (2021)", + "20. Hoiem, D., Efros, A.A., Hebert, M.: Putting Objects in Perspective. IJCV 80, 3-15 (2008). https://doi.org/10.1007/s11263-008-0137-5", + "21. Houston, J., Zuidhof, G., Bergamini, L., Ye, Y., Chen, L., Jain, A., Omari, S., Iglovikov, V., Ondruska, P.: One Thousand and One Hours: Self-Driving Motion Prediction Dataset. In: Conference on Robot Learning. pp. 409-418 (2021)", + "22. Jain, J., Li, J., Chiu, M.T., Hassani, A., Orlov, N., Shi, H.: OneFormer: One Transformer to Rule Universal Image Segmentation. In: CVPR. pp. 2989-2998 (2023)", + "23. Kim, D., Ga, W., Ahn, P., Joo, D., Chun, S., Kim, J.: Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth. CoRR abs/2201.07436 (2022)", + "24. Kingma, D.P., Ba, J.: Adam: A Method for Stochastic Optimization. CoRR abs/1412.6980 (2014)", + "25. Lyu, X., Liu, L., Wang, M., Kong, X., Liu, L., Liu, Y., Chen, X., Yuan, Y.: HR-Depth: High Resolution Self-Supervised Monocular Depth Estimation. In: AAAI. vol. 35, pp. 2294-2301 (2021)", + "26. Milli, E., Erkent, O., Yilmaz, A.E.: Multi-Modal Multi-Task (3MT) Road Segmentation. IEEE Robotics and Automation Letters 8(9), 5408-5415 (2023). https://doi.org/10.1109/LRA.2023.3295254", + "27. Ning, C., Gan, H.: Trap Attention: Monocular Depth Estimation with Manual Traps. In: CVPR. pp. 5033-5043 (2023)", + "28. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: PyTorch: An Imperative Style, High-Performance Deep Learning Library. In: Adv. Neural Inform. Process. Syst. pp. 8026--8037 (2019)", + "29. Piccinelli, L., Sakaridis, C., Yu, F.: iDisc: Internal Discretization for Monocular Depth Estimation. In: CVPR. pp. 21477-21487 (2023)", + "30. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet: A Large Scale Visual Recognition Challenge. IJCV 115, 211-252 (2015)", + "31. Sadat, A., Casas, S., Ren, M., Wu, X., Dhawan, P., Urtasun, R.: Perceive, Predict, and Plan: Safe Motion Planning Through Interpretable Semantic Representations. In: ECCV. pp. 414-430 (2020)", + "32. Schonberger, J.L., Frahm, J.M.: Structure-From-Motion Revisited. In: CVPR. pp. 4104-4113 (2016)", + "33. Simonelli, A., Bulò, S.R., Porzi, L., Kontschieder, P., Ricci, E.: Are We Missing Confidence in Pseudo-LiDAR Methods for Monocular 3D Object Detection? In: ICCV. pp. 3225-3233 (2021)", + "34. Spencer, J., Russell, C., Hadfield, S., Bowden, R.: Kick Back & Relax: Learning to Reconstruct the World by Watching SlowTV (2023)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "G.Kinoshita et al.", + "bbox": [ + 271, + 114, + 392, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Sucar, E., Hayet, J.B.: Bayesian Scale Estimation for Monocular SLAM Based on Generic Object Detection for Correcting Scale Drift. In: ICRA. pp. 5152-5158 (2018)", + "36. Wagstaff, B., Kelly, J.: Self-Supervised Scale Recovery for Monocular Depth and Egomotion Estimation. In: IROS. pp. 2620-2627 (2021)", + "37. Wang, C., Buenaposada, J.M., Zhu, R., Lucey, S.: Learning Depth from Monocular Videos Using Direct Methods. In: CVPR. pp. 2022-2030 (2018)", + "38. Wang, H., Cai, P., Fan, R., Sun, Y., Liu, M.: End-to-End Interactive Prediction and Planning with Optical Flow Distillation for Autonomous Driving. In: CVPRW. pp. 2229–2238 (2021)", + "39. Wang, Y., Chao, W.L., Garg, D., Hariharan, B., Campbell, M., Weinberger, K.Q.: Pseudo-LiDAR From Visual Depth Estimation: Bridging the Gap in 3D Object Detection for Autonomous Driving. In: CVPR. pp. 8437-8445 (2019)", + "40. Wang, Y., Chen, X., You, Y., Li, L.E., Hariharan, B., Campbell, M., Weinberger, K.Q., Chao, W.L.: Train in Germany, Test in the USA: Making 3D Object Detectors Generalize. In: CVPR. pp. 11710-11720 (2020)", + "41. Wang, Y., Yang, H., Cai, J., Wang, G., Wang, J., Huang, Y.: Unsupervised Learning of Depth and Pose Based on Monocular Camera and Inertial Measurement Unit (IMU). In: ICRA. pp. 10010-10017 (2023)", + "42. Wang, Z., Bovik, A., Sheikh, H., Simoncelli, E.: Image Quality Assessment: From Error Visibility to Structural Similarity. IEEE Transactions on Image Processing 13(4), 600-612 (2004). https://doi.org/10.1109/TIP.2003.819861", + "43. Watson, J., Mac Aodha, O., Prisacariu, V., Brostow, G., Firman, M.: The Temporal Opportunist: Self-Supervised Multi-Frame Monocular Depth. In: CVPR. pp. 1164–1174 (2021)", + "44. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., Ramanan, D., Carr, P., Hays, J.: Argoverse 2: Next Generation Datasets for Self-driving Perception and Forecasting. In: NeurIPS Datasets and Benchmarks (2021)", + "45. Xiang, J., Wang, Y., An, L., Liu, H., Wang, Z., Liu, J.: Visual Attention-Based Self-Supervised Absolute Depth Estimation Using Geometric Priors in Autonomous Driving. IEEE Robotics and Automation Letters 7(4), 11998-12005 (2022). https://doi.org/10.1109/LRA.2022.3210298", + "46. Yang, Z., Wang, P., Wang, Y., Xu, W., Nevatia, R.: LEGO: Learning Edge with Geometry All at Once by Watching Videos. In: CVPR. pp. 225-234 (2018)", + "47. Yin, W., Zhang, C., Chen, H., Cai, Z., Yu, G., Wang, K., Chen, X., Shen, C.: Metric3D: Towards Zero-Shot Metric 3D Prediction from a Single Image. In: ICCV. pp. 9043-9053 (2023)", + "48. Yin, Z., Shi, J.: GeoNet: Unsupervised Learning of Dense Depth, Optical Flow and Camera Pose. In: CVPR. pp. 1983-1992 (2018)", + "49. You, Y., Wang, Y., Chao, W.L., Garg, D., Pleiss, G., Hariharan, B., Campbell, M., Weinberger, K.Q.: Pseudo-LiDAR++: Accurate Depth for 3D Object Detection in Autonomous Driving. In: ICLR (2020)", + "50. Yuan, W., Gu, X., Dai, Z., Zhu, S., Tan, P.: Neural Window Fully-Connected CRFs for Monocular Depth Estimation. In: CVPR. pp. 3916-3925 (2022)", + "51. Zeng, W., Luo, W., Suo, S., Sadat, A., Yang, B., Casas, S., Urtasun, R.: End-To-End Interpretable Neural Motion Planner. In: CVPR. pp. 8652-8661 (2019)", + "52. Zhang, N., Nex, F., Vosselman, G., Kerle, N.: Lite-Mono: A Lightweight CNN and Transformer Architecture for Self-Supervised Monocular Depth Estimation. In: CVPR. pp. 18537-18546 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Unsupervised Training for Metric Monocular Road-Scene Depth Estimation", + "bbox": [ + 227, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "53. Zhang, S., Zhang, J., Tao, D.: Towards Scale-Aware, Robust, and Generalizable Unsupervised Monocular Depth Estimation by Integrating IMU Motion Dynamics. In: ECCV. pp. 143-160 (2022)", + "54. Zhang, S., Li, X., Liu, Y., Fu, H.: Scale-aware Insertion of Virtual Objects in Monocular Videos. In: ISMAR. pp. 36-44 (2020)", + "55. Zhou, H., Greenwood, D., Taylor, S.: Self-Supervised Monocular Depth Estimation with Internal Feature Fusion. In: BMVC (2021)", + "56. Zhou, T., Brown, M., Snavely, N., Lowe, D.G.: Unsupervised Learning of Depth and Ego-Motion from Video. In: CVPR. pp. 6612-6619 (2017)", + "57. Zhu, R., Yang, X., Hold-Geoffroy, Y., Perazzi, F., Eisenmann, J., Sunkavalli, K., Chandraker, M.: Single View Metrology in the Wild. In: ECCV. pp. 316-333 (2020)" + ], + "bbox": [ + 215, + 146, + 784, + 300 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "G.Kinoshita et al.", + "bbox": [ + 271, + 114, + 392, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Camera Height Doesn't Change_ Unsupervised Training for Metric Monocular Road-Scene Depth Estimation/b1560f9a-1722-4ac2-a48d-d70d8859d7c8_model.json b/2024/Camera Height Doesn't Change_ Unsupervised Training for Metric Monocular Road-Scene Depth Estimation/b1560f9a-1722-4ac2-a48d-d70d8859d7c8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..102c477795bc6f71f105a1ac80f0ebea433942b7 --- /dev/null +++ b/2024/Camera Height Doesn't Change_ Unsupervised Training for Metric Monocular Road-Scene Depth Estimation/b1560f9a-1722-4ac2-a48d-d70d8859d7c8_model.json @@ -0,0 +1,2403 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.245, + 0.141, + 0.761, + 0.21 + ], + "angle": 0, + "content": "Camera Height Doesn't Change: Unsupervised Training for Metric Monocular Road-Scene Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.371, + 0.235, + 0.634, + 0.251 + ], + "angle": 0, + "content": "Genki Kinoshita and Ko Nishino" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.263, + 0.728, + 0.306 + ], + "angle": 0, + "content": "Graduate School of Informatics, Kyoto University, Kyoto, Japan \ngkinoshita@vision.ist.i.kyoto-u.ac.jp, kon@i.kyoto-u.ac.jp \nhttps://vision.ist.i.kyoto-u.ac.jp/" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.342, + 0.744, + 0.606 + ], + "angle": 0, + "content": "Abstract. In this paper, we introduce a novel training method for making any monocular depth network learn absolute scale and estimate metric road-scene depth just from regular training data, i.e., driving videos. We refer to this training framework as FUMET. The key idea is to leverage cars found on the road as sources of scale supervision and to incorporate them in network training robustly. FUMET detects and estimates the sizes of cars in a frame and aggregates scale information extracted from them into an estimate of the camera height whose consistency across the entire video sequence is enforced as scale supervision. This realizes robust unsupervised training of any, otherwise scale-oblivious, monocular depth network so that they become not only scale-aware but also metric-accurate without the need for auxiliary sensors and extra supervision. Extensive experiments on the KITTI and the Cityscapes datasets show the effectiveness of FUMET, which achieves state-of-the-art accuracy. We also show that FUMET enables training on mixed datasets of different camera heights, which leads to larger-scale training and better generalization. Metric depth reconstruction is essential in any road-scene visual modeling, and FUMET democratizes its deployment by establishing the means to convert any model into a metric depth estimator." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.619, + 0.741, + 0.647 + ], + "angle": 0, + "content": "Keywords: Unsupervised Monocular Depth Estimation \\(\\cdot\\) Metric Scale \\(\\cdot\\) Camera Height Invariance \\(\\cdot\\) Object Size Prior" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.673, + 0.377, + 0.689 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Monocular depth estimation is essential for autonomous driving and advanced driver assistance systems (ADAS). It underlies many of their key perceptual tasks including 3D object detection [7,33,39,49], motion planning [31,38,51], and road segmentation [26]. Although supervised methods have achieved impressive accuracy, they suffer from costly ground-truth data collection. This limits their generalizability [34] and also makes continuous learning difficult." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Many recent methods [5, 15, 18, 25, 52] leverage self-supervision to scale up training. Most of these methods, however, suffer from scale ambiguity which renders them difficult to use off-the-shelf. To resolve the scale ambiguity and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "G.Kinoshita et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "recover in metric scale, the actual scene scale needs to be supervised in some form. Past works have leveraged velocity [16], GPS [3], IMU [41, 53], or camera height [36, 45] for this. These methods require auxiliary sensors to measure these extra quantities in addition to the monocular RGB camera, which significantly limits the ability to scale the size of training data. Arbitrary in-the-wild driving videos cannot be used as they lack such measurements. These self-supervised methods often still require weak scale supervision and cannot be made fully self-supervised as the scale changes for each sequence captured with a different camera." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.287, + 0.789, + 0.468 + ], + "angle": 0, + "content": "Road scenes actually have plenty of untapped sources of metric scale. Cars found on the road are one of them as they are rigid objects whose actual sizes do not change and are unique to each make and model. The road is also exactly the region where we need accurate metric depth the most. If we can leverage the sizes of cars to bridge the 3D to 2D projection, we may train a mono-depth network to become metric-accurate. Simply employing a prior on the car size, however, would be too brittle since the metric supervision will be as ambiguous as the accuracy of that prior. The perspective projection and the variations in car makes, years, and colors, lead to a large variance in the size estimates, which will directly affect the mono-depth metric accuracy. In addition, there are no reliable object size priors except for a few introduced as probability densities for other tasks [20,54,57] whose parameters are not shared." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.789, + 0.653 + ], + "angle": 0, + "content": "How can we train an arbitrary mono-depth estimator to become metric-accurate by robustly leveraging prior knowledge about the 3D sizes of cars found in driving videos? We derive fully unsupervised metric depth trainer (FUMET) to answer this question. Our key idea is to aggregate car size cues in the frames and the sequence into a single physical measure, namely the camera height. Regardless of the scene, the camera capturing it is the same for the same sequence. This means that, by transforming the car size cues found in different frames into a camera height estimate, the very fact that it should not change can be used as a supervision of invariance to achieve metric depth reconstruction. We can formulate this as an optimization of the camera height across frames and training epochs. This leads to gradual metric-aware learning that is stable and accurate, which also makes the depth estimates consistent across the sequence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.656, + 0.789, + 0.807 + ], + "angle": 0, + "content": "We introduce a learned size prior (LSP) to estimate the size of each car found in a frame. This prior gives us the vehicle dimensions from its appearance. It is trained with various augmentations on a large-scale dataset without the need for any manual annotation. By comparing these dimension estimates with those computed from the depth estimates, we obtain a per-frame scale factor. The multiplication of this scale estimate with the camera height estimate derived from the estimated depth becomes the camera height estimate of the frame. These estimated camera heights across all frames in a sequence are then consolidated by using their median value which is then used as scale supervision with its weighted moving average at the end of each training epoch." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.789, + 0.84 + ], + "angle": 0, + "content": "We evaluate the effectiveness of FUMET with extensive experiments using the KITTI and the Cityscapes datasets and compare its accuracy with a number" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.228, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Unsupervised Training for Metric Monocular Road-Scene Depth Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "of weakly-supervised methods. The results clearly show that FUMET succeeds in making monocular depth estimation methods learn metric scales and achieves the state-of-the-art accuracy without any direct depth and scale supervision. We also demonstrate its capability to train a model on mixed datasets collected with different cameras, i.e., camera heights. This enables larger-scale training and leads to higher generalization. We believe FUMET can serve as a fundamental building block for monocular depth estimation in the wild." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.275, + 0.383, + 0.29 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.305, + 0.396, + 0.32 + ], + "angle": 0, + "content": "2.1 Self-Supervision" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.788, + 0.51 + ], + "angle": 0, + "content": "Although, as in other tasks, supervised methods achieve the highest accuracy in monocular depth estimation (MDE) [1, 23, 27, 29, 50], they obviously require ground-truth data usually obtained with a LiDAR sensor. Self-supervised MDE methods remove the need for such hardware costs for training. Garg et al. [11] proposed the first approach to self-supervised MDE by predicting per-pixel disparities between a stereo pair. Godard et al. [14] extend this idea by introducing left-right consistency and greatly improve the accuracy. These methods require calibrated stereo data for training which limits their applicability. Zhou et al. [56] introduced a pioneering method for training solely from monocular videos by incorporating a network that estimates the relative camera poses between successive frames. Based on this idea, many MDE methods that can be trained solely with monocular videos have been proposed [5, 15, 18, 25, 52, 55]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.787, + 0.571 + ], + "angle": 0, + "content": "Despite the popularity and improvements of self-supervised MDE, directly deploying those models in autonomous driving or ADAS systems is unrealistic as they only produce relative depth up to an unknown scale. Our FUMET overcomes this issue without the need for any external supervision." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.591, + 0.412, + 0.607 + ], + "angle": 0, + "content": "2.2 Weak Supervision" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.766 + ], + "angle": 0, + "content": "Several works have tackled scale-aware MDE by adopting weak scale supervision through velocity [16], GPS [3], IMU [41], or camera height [36,45] measurements. Zhang et al. [53] use the combination of IMU, velocity, and gravity to achieve the highest accuracy among these methods. These methods fundamentally rely on such auxiliary measurements in addition to regular RGB images. This makes it impossible to take advantage of the many road-scene videos on the Internet (e.g., YouTube) or casually installed dashcam videos for training. In contrast, FUMET relies on a universal car size prior which is independent of the data collection environments and does not require any such weak supervision; it can be trained solely from monocular video sequences." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.786, + 0.416, + 0.801 + ], + "angle": 0, + "content": "2.3 Object Size Priors" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Several works for other tasks leverage priors on object sizes including single view metrology [57], object insertion [20,54], and visual odometry [10,35]. These" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "G.Kinoshita et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "methods model the priors as normal distributions or Gaussian mixture models, and the parameters of these distributions are predetermined. When there is a gap between those values and the true ones, which is usually the case in the real world, the difference directly surfaces in erroneous scale estimates." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.209, + 0.788, + 0.284 + ], + "angle": 0, + "content": "Instead of these fixed priors, we introduce a learning-based size prior which we refer to as LSP. This estimates vehicle dimensions based on its appearance, thus can avoid the aforementioned problem. Furthermore, it is trained on a large-scale dataset with various augmentations, so that it can robustly estimate vehicle dimensions from any image." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.311, + 0.383, + 0.327 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.785, + 0.466 + ], + "angle": 0, + "content": "Our goal is to train a metric-accurate MDE model \\(\\theta_{\\mathrm{depth}}\\) with only self-supervision. This self-supervision is comprised of two parts: one designed for learning metric scale (Sec. 4) and the other dedicated to learning depth geometry. For the latter, we follow Zhou et al. [56] and minimize the reconstruction loss between the current frame \\(I_{t}\\) and the ones synthesized from temporally adjacent frames \\(I_{s}\\), \\((s = t \\pm 1)\\). We employ a pose network \\(\\theta_{\\mathrm{pose}}\\) in addition to \\(\\theta_{\\mathrm{depth}}\\) to estimate the relative camera pose \\(\\mathbf{T}_{t \\to s} \\in \\mathrm{SE}(3)\\) and the current dense depth map \\(D_{t}\\), respectively," + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.469, + 0.786, + 0.485 + ], + "angle": 0, + "content": "\\[\nD _ {t} = \\theta_ {\\text {d e p t h}} \\left(I _ {t}\\right), \\mathbf {T} _ {t \\rightarrow s} = \\theta_ {\\text {p o s e}} \\left(I _ {t}, I _ {s}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.493, + 0.632, + 0.509 + ], + "angle": 0, + "content": "We then synthesize the image from the current viewpoint" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.522, + 0.785, + 0.539 + ], + "angle": 0, + "content": "\\[\nI _ {s \\rightarrow t} = I _ {s} \\left\\langle \\operatorname {p r o j} \\left(D _ {t}, \\mathbf {T} _ {t \\rightarrow s}, \\mathbf {K}\\right)\\right\\rangle , \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.642 + ], + "angle": 0, + "content": "where \\(\\langle \\cdot \\rangle\\) is a bilinear sampling operator, \\(proj(\\cdot)\\) outputs the 2D coordinates of \\(D_{t}\\) when projected into the viewpoint of \\(I_{s}\\), and \\(\\mathbf{K}\\) is the camera intrinsic. Following Godard et al. [15], we use an affine combination of SSIM [42] and L1 loss as the photometric error \\(pe\\) and define the reconstruction loss \\(\\mathcal{L}_{\\mathrm{rec}}\\) as the photometric error between \\(I_{t}\\) and \\(I_{s\\rightarrow t}\\) for the pixels with minimum error across source frames" + }, + { + "type": "equation", + "bbox": [ + 0.297, + 0.657, + 0.785, + 0.676 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c}} = \\min _ {s} p e \\left(I _ {t}, I _ {s \\rightarrow t}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.295, + 0.681, + 0.785, + 0.711 + ], + "angle": 0, + "content": "\\[\np e \\left(I _ {t}, I _ {s \\rightarrow t}\\right) = \\frac {\\lambda}{2} \\left(1 - \\operatorname {S S I M} \\left(I _ {t}, I _ {s \\rightarrow t}\\right)\\right) + (1 - \\lambda) \\left| I _ {t} - I _ {s \\rightarrow t} \\right|, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.72, + 0.704, + 0.737 + ], + "angle": 0, + "content": "where \\(\\lambda = 0.85\\). We also employ an edge-aware smoothness loss [14]" + }, + { + "type": "equation", + "bbox": [ + 0.37, + 0.748, + 0.785, + 0.767 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s m}} = \\left| \\partial_ {x} d _ {t} ^ {*} \\right| e ^ {- \\left| \\partial_ {x} I _ {t} \\right|} + \\left| \\partial_ {y} d _ {t} ^ {*} \\right| e ^ {- \\left| \\partial_ {y} I _ {t} \\right|}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.779, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\( d_{t}^{*} = d_{t} / \\bar{d}_{t} \\) is the mean-normalized inverse depth [37] to prevent the estimated depth from shrinking. Furthermore, when computing \\( \\mathcal{L}_{\\mathrm{rec}} \\) we use automasking [15] to mask out stationary pixels which generate incorrect supervisory signals." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.228, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Unsupervised Training for Metric Monocular Road-Scene Depth Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.144, + 0.788, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.267, + 0.789, + 0.379 + ], + "angle": 0, + "content": "Fig. 1: Overview of FUMET. At each training step \\( n \\), an unscaled camera height \\( H_{\\mathrm{cam}}^{n,\\tau} \\) is computed differentiably from the estimated depth. The previous epoch \\( \\tau - 1 \\) provides supervision with a scaled camera height \\( H_{\\mathrm{cam}}^{*\\tau - 1} \\). To obtain this scaled camera height supervision, Silhouette Projector first computes the object silhouette heights \\( H_{\\mathrm{obj}}' \\) from the depth map. By comparing \\( H_{\\mathrm{obj}}' \\) and the estimated one \\( H_{\\mathrm{obj}} \\) with LSP, per-frame scale factor \\( s \\) is determined and we obtain the scaled camera height \\( H_{\\mathrm{cam}}^{n,\\tau} = s \\cdot H_{\\mathrm{cam}}'^{\\prime,\\tau} \\). At the end of each epoch, \\( H_{\\mathrm{cam}}^* \\) is optimized across a series of consecutive frames and updated with the weighted moving average." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.407, + 0.336, + 0.423 + ], + "angle": 0, + "content": "4 FUMET" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.439, + 0.784, + 0.469 + ], + "angle": 0, + "content": "Fig. 1 depicts the overall architecture of FUMET. Please see the caption for an overview of the framework. Let “'” denote “not metrically scaled”." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.492, + 0.573, + 0.508 + ], + "angle": 0, + "content": "4.1 Scale-Aware Self-Supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.517, + 0.788, + 0.653 + ], + "angle": 0, + "content": "In FUMET, we derive a camera height pseudo-supervision from depth estimates. To obtain the per-frame camera height, we first compute the normal vector for each pixel \\( p_i \\). We consider the 8-neighborhood of \\( p_i \\) and form 8 pairs \\( \\mathcal{N}(p_i) = \\{(p_{j_0}^i,p_{j_1}^i)\\}_{j = 1}^8 \\). Each pair \\( (p_{j_0}^i,p_{j_1}^i) \\) consists of two pixels whose difference vectors w.r.t. \\( p_i \\) are perpendicular to each other and are sorted counterclockwise. We reproject the paired pixels onto 3D space and compute the cross-product of the two difference vectors w.r.t. the reprojected 3D point of \\( p_i \\). By summing and normalizing all the cross-product vectors of each pair, we obtain the normal vector \\( \\hat{\\mathbf{n}}(p_i) \\)" + }, + { + "type": "equation", + "bbox": [ + 0.261, + 0.664, + 0.787, + 0.703 + ], + "angle": 0, + "content": "\\[\n\\mathbf {n} \\left(p _ {i}\\right) = \\sum_ {\\mathcal {N} \\left(p _ {i}\\right)} \\left(\\phi \\left(p _ {j _ {0}} ^ {i}\\right) - \\phi \\left(p _ {i}\\right)\\right) \\times \\left(\\phi \\left(p _ {j _ {1}} ^ {i}\\right) - \\phi \\left(p _ {i}\\right)\\right), \\quad \\hat {\\mathbf {n}} \\left(p _ {i}\\right) = \\frac {\\mathbf {n} \\left(p _ {i}\\right)}{\\| \\mathbf {n} \\left(p _ {i}\\right) \\|}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.715, + 0.536, + 0.73 + ], + "angle": 0, + "content": "where \\(\\phi (\\cdot)\\) denotes perspective reprojection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.73, + 0.785, + 0.775 + ], + "angle": 0, + "content": "For pixels in the road region, the negated inner product between the normal vector and the reprojected 3D point corresponds to the camera height from the ground plane" + }, + { + "type": "equation", + "bbox": [ + 0.401, + 0.775, + 0.786, + 0.793 + ], + "angle": 0, + "content": "\\[\nH _ {\\mathrm {c a m}} ^ {\\prime} (p _ {i}) = - \\phi (p _ {i}) \\cdot \\hat {\\mathbf {n}} (p _ {i}). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.798, + 0.733, + 0.814 + ], + "angle": 0, + "content": "We represent the median of \\( H_{\\mathrm{cam}}'(p_i) \\) as the camera height of the frame" + }, + { + "type": "equation", + "bbox": [ + 0.37, + 0.824, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\tilde {H} _ {\\mathrm {c a m}} ^ {\\prime} = \\operatorname {m e d} \\left(\\left\\{H _ {\\mathrm {c a m}} ^ {\\prime} \\left(p _ {i}\\right) \\mid p _ {i} \\in M _ {\\mathrm {r}} \\right\\}\\right), \\tag {8}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "G.Kinoshita et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.238 + ], + "angle": 0, + "content": "where \\( p_i \\) is in the road region mask \\( M_{\\mathrm{r}} \\) obtained with an off-the-shelf semantic segmentation model [22] before training. This camera height \\( \\tilde{H}_{\\mathrm{cam}}' \\) estimate is unscaled and needs to be upgraded to a real-scaled value to impose the invariance loss across frames. By using the Silhouette Projector, we aggregate the scale information from the object height prior into a per-frame scale factor \\( s \\) (see Sec. 4.2) and obtain the scaled camera height as \\( H_{\\mathrm{cam}} = s \\cdot \\tilde{H}_{\\mathrm{cam}}' \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.345 + ], + "angle": 0, + "content": "We leverage the fact that the camera height does not change in a sequence and formulate the invariance into an optimization across frames and training epochs. Thanks to this optimization, we acquire an accurate and stable supervision, which is also consistent across the scene. At the end of the \\(\\tau\\)-th epoch, we use the median of \\(\\{H_{\\mathrm{cam}}^{n,\\tau}\\}_{n=1}^{N}\\) for all frames \\(N\\) as the representative camera height \\(H_{\\mathrm{cam}}^{\\tau}\\) at the \\(\\tau\\)-th epoch. Then, we update the scaled camera height pseudo-invariance \\(H_{\\mathrm{cam}}^{*}\\) with its weighted moving average" + }, + { + "type": "equation", + "bbox": [ + 0.317, + 0.357, + 0.786, + 0.39 + ], + "angle": 0, + "content": "\\[\nH _ {\\mathrm {c a m}} ^ {* \\tau} \\leftarrow \\left\\{\\frac {\\tau (\\tau - 1)}{2} \\cdot H _ {\\mathrm {c a m}} ^ {* \\tau - 1} + \\tau \\cdot H _ {\\mathrm {c a m}} ^ {\\tau} \\right\\} / \\frac {\\tau (\\tau + 1)}{2}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.787, + 0.444 + ], + "angle": 0, + "content": "When training on a dataset containing images captured at multiple camera heights, we create pseudo-labels corresponding to each camera height and optimize them individually." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.492 + ], + "angle": 0, + "content": "To make the network learn the absolute scale and directly produce the scaled depth map, we enforce a loss on the camera height \\( H_{\\mathrm{cam}}' \\) using the pseudo-invariance \\( H_{\\mathrm{cam}}^{*\\tau - 1} \\) computed in the previous epoch" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.503, + 0.786, + 0.536 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c a m}} = \\frac {1}{| M _ {\\mathrm {r}} |} \\sum_ {i} M _ {\\mathrm {r}} \\left(p _ {i}\\right) \\cdot \\left| H _ {\\mathrm {c a m}} ^ {\\prime} \\left(p _ {i}\\right) - H _ {\\mathrm {c a m}} ^ {* \\tau - 1} \\right|. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.609 + ], + "angle": 0, + "content": "It is worth noting that at inference, the network does not assume camera height invariance in the input sequence; it is only assumed during training. The computational cost at inference is exactly the same as the original MDE model trained without FUMET." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.632, + 0.432, + 0.648 + ], + "angle": 0, + "content": "4.2 Silhouette Projector" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.734 + ], + "angle": 0, + "content": "In the process of creating a scaled camera height supervision, we need to know the per-frame scale factor. For this, we can consider leveraging an object of known size. This, however, would necessitate extraction of the exact 3D object region which is non-trivial as objects are often occluded or cropped and oriented in various poses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We introduce Silhouette Projector to robustly estimate this per-frame scale factor. Silhouette Projector leverages two simple facts of road scenes. First, the height of the object silhouette projected on the plane perpendicular to the ground plane does not change regardless of the object pose. Second, we can compute the silhouette height by measuring the distance from the top of the silhouette to the ground plane even if they are partly occluded or truncated, as long as the top is visible and the object is on the ground plane." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.228, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Unsupervised Training for Metric Monocular Road-Scene Depth Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.345 + ], + "angle": 0, + "content": "We first reconstruct an object point cloud with the estimated depth map through perspective reprojection. The object region \\( M_{\\mathrm{obj}} \\) in the image can be obtained with an off-the-shelf instance segmentation model [22] before training. We then orthographically project these 3D points onto a plane \\( P^{\\perp} \\) perpendicular to the ground plane, i.e., an arbitrary plane parallel to the road normal vector \\( \\tilde{\\mathbf{n}} \\). We compute \\( \\tilde{\\mathbf{n}} \\) as the median of the normal vectors \\( \\hat{\\mathbf{n}}(p_i) \\) (see Eq. (6)) in the road region \\( M_{\\mathrm{r}} \\). We define the farthest distance from the 2D point on \\( P^{\\perp} \\) to the ground plane as the silhouette height. This value is unscaled because we derived this from the estimated unscaled depth map. The ratio between a height prior and its silhouette height corresponds to the scale factor of interest. When there are multiple objects in an image, we use the median of the scale factors each corresponding to each object as the scale factor of the frame for computational simplicity and robustness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.462 + ], + "angle": 0, + "content": "Outlier Filtering. Some objects cannot be handled well with Silhouette Projector for various reasons including occluded upper parts, not on the ground plane, inaccurate segmented region, and misclassified object category. We introduce an algorithm to detect these as outliers based on geometric plausibility. We first calculate the horizon \\(\\mathbf{l}\\) with the road normal vector \\(\\tilde{\\mathbf{n}}\\) and the camera intrinsics \\(\\mathbf{K}\\) [17]" + }, + { + "type": "equation", + "bbox": [ + 0.462, + 0.464, + 0.785, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\mathbf {l} = \\mathbf {K} ^ {- \\top} \\tilde {\\mathbf {n}}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.788, + 0.597 + ], + "angle": 0, + "content": "Based on the approximations in [20], we derive the approximate metric-scale object height \\(\\hat{H}_{\\mathrm{obj}} \\approx \\frac{h_{\\mathrm{obj}}}{h_{\\mathrm{cam}}} H_{\\mathrm{cam}}^{*\\tau - 1}\\), where \\(h_{\\mathrm{cam}}\\) is the farthest distance from the pixel in the object instance mask to \\(\\mathbf{l}\\), \\(h_{\\mathrm{obj}}\\) is the farthest distance between two points on the object mask on which tangent lines parallel to \\(\\mathbf{l}\\) lie, and \\(H_{\\mathrm{cam}}^{*\\tau - 1}\\) is the pseudo camera height at the \\(\\tau\\)-th epoch derived in Eq. (9). By comparing the relative gap between \\(\\hat{H}_{\\mathrm{obj}}\\) and the object height prior \\(H_{\\mathrm{obj}}\\), we determine the outliers as" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.599, + 0.785, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\left\\{k \\left| \\frac {\\left| H _ {\\mathrm {o b j}} ^ {k} - \\hat {H} _ {\\mathrm {o b j}} ^ {k} \\right|}{H _ {\\mathrm {o b j}} ^ {k}} > \\mathcal {T} \\right. \\right\\}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.646, + 0.6, + 0.66 + ], + "angle": 0, + "content": "In all the experiments, we set the threshold \\(\\mathcal{T}\\) to 0.2." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.689, + 0.391, + 0.705 + ], + "angle": 0, + "content": "4.3 Training Losses" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Auxiliary Rough Geometric Loss. We experimentally observe that estimated object depths, especially those distant from the camera, often erroneously lean towards the road surface because of the inaccurate supervisory signal from the reconstruction loss in low-texture regions. This tendency makes silhouette heights inaccurate. To mitigate this issue, we introduce an auxiliary rough geometric loss \\(\\mathcal{L}_{\\mathrm{aux}}\\), which enforces object point clouds to be within a close region. For each inlier object \\(k\\) (\\(k = 1 \\dots K\\)), we define \\(\\mathcal{L}_{\\mathrm{aux}}\\) as the gap between the estimated depth map \\(D\\) in the object region \\(M_{\\mathrm{obj}}\\) and the approximate object" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "G.Kinoshita et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.144, + 0.578, + 0.166 + ], + "angle": 0, + "content": "depth \\(D_{\\mathrm{aprx}} = \\frac{H_{\\mathrm{obj}}}{h_{\\mathrm{obj}}} f_y\\) where \\(f_{y}\\) is the focal length" + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.178, + 0.786, + 0.215 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a u x}} = \\frac {1}{K} \\sum_ {k} \\frac {1}{| M _ {\\mathrm {o b j}} ^ {k} |} \\sum_ {i} M _ {\\mathrm {o b j}} ^ {k} \\left(p _ {i}\\right) \\cdot \\left| D \\left(p _ {i}\\right) - D _ {\\mathrm {a p r x}} ^ {k} \\right|. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.225, + 0.67, + 0.24 + ], + "angle": 0, + "content": "As a by-product, this loss accelerates learning the metric scale." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.24, + 0.788, + 0.361 + ], + "angle": 0, + "content": "The \\(\\mathcal{L}_{\\mathrm{aux}}\\) is potentially inaccurate since \\(D_{\\mathrm{aprx}}\\) assumes that each object can be represented by a plane parallel to the image plane, and we observe that it degrades the depth accuracy in the late stage of training. On the other hand, in the early stage, estimated depth is unreliable, so heavily relying on the camera height loss \\(\\mathcal{L}_{\\mathrm{cam}}\\) makes the training unstable. For this, we gradually decrease the weight of \\(\\mathcal{L}_{\\mathrm{aux}}\\) and increase the one for \\(\\mathcal{L}_{\\mathrm{cam}}\\) as the training proceeds in a logarithmic way. To ensure training stability, we stop changing these weights after the \\(\\tau_{\\mathrm{mid-th}}\\) epoch" + }, + { + "type": "equation", + "bbox": [ + 0.231, + 0.373, + 0.786, + 0.414 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {\\mathrm {a u x}} (\\tau) = \\left\\{ \\begin{array}{l l} \\frac {- \\log (\\tau + 1)}{\\log \\left(\\tau_ {\\mathrm {m i d}} + 1\\right)} & (\\tau \\leq \\tau_ {\\mathrm {m i d}}) \\\\ \\epsilon & (\\tau > \\tau_ {\\mathrm {m i d}}) \\end{array} , \\lambda_ {\\mathrm {c a m}} (\\tau) = \\left\\{ \\begin{array}{l l} \\frac {\\log (\\tau + 1)}{\\log \\left(\\tau_ {\\mathrm {m i d}} + 1\\right)} & (\\tau \\leq \\tau_ {\\mathrm {m i d}}) \\\\ 1. 0 & (\\tau > \\tau_ {\\mathrm {m i d}}) \\end{array} , \\right. \\right. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.423, + 0.338, + 0.436 + ], + "angle": 0, + "content": "where \\(\\epsilon = 0.005\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.437, + 0.785, + 0.468 + ], + "angle": 0, + "content": "The total loss is \\(\\mathcal{L} = \\alpha \\lambda_{\\mathrm{cam}}\\mathcal{L}_{\\mathrm{cam}} + \\beta \\lambda_{\\mathrm{aux}}\\mathcal{L}_{\\mathrm{aux}} + \\mathcal{L}_{\\mathrm{sm}} + \\mathcal{L}_{\\mathrm{rec}}\\) where \\(\\alpha = 0.01\\) and \\(\\beta = 1.0\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.491, + 0.295, + 0.507 + ], + "angle": 0, + "content": "5 LSP" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.719 + ], + "angle": 0, + "content": "Our FUMET extracts metric scale from a prior on object heights through the camera height loss \\(\\mathcal{L}_{\\mathrm{cam}}\\) and the auxiliary rough geometric loss \\(\\mathcal{L}_{\\mathrm{aux}}\\). To ensure that FUMET learns the scale regardless of the particular dataset used for training, a universal prior of object heights, independent of the environment, becomes essential. Cars are objects whose dimensions are explicitly and accurately known. Better yet, they are abundantly located on the road, the key semantic region whose metric depth is essential for many downstream tasks. Some methods for other tasks have modeled the size prior as a probability distribution [20, 57]. The gap between its mean and the true expectation for each environment, however, is directly reflected in the uncertainty of the scale in the supervision. Wang et al. [40] also report that there is a gap in car dimensions among datasets. These facts suggest that modeling the object size prior as a single probability distribution is problematic, especially for man-made objects like cars." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.795 + ], + "angle": 0, + "content": "We humans not only possess rough prior knowledge about the vehicle size but can also estimate it more accurately by extracting instance-specific information such as car models from its appearance. Motivated by this intuition, we introduce a learned universal vehicle size prior termed LSP. Given a masked vehicle image, LSP estimates its dimensions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "For training, we use a dataset consisting of a large number of vehicle images and their information such as 3D dimensions. This dataset can be easily collected with web scraping and does not require manual annotations. To enhance" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.228, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Unsupervised Training for Metric Monocular Road-Scene Depth Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.266 + ], + "angle": 0, + "content": "the robustness to various input images, in addition to common color jittering and blur augmentation, we perform augmentations to simulate occlusion and truncation. For occlusion augmentation, we mask out random regions with the same shape as a vehicle or with a beam of random width. For truncation augmentation, we translate the vehicle mask in a random direction. We visualize these in the supplementary material. The abundance of training images and diverse augmentations enable LSP to estimate vehicle dimensions robustly across various datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.785, + 0.314 + ], + "angle": 0, + "content": "Although FUMET requires only a car height prior, we construct LSP to estimate the width and length as well as height. This enhances it to grasp an overall car size and improves the prediction accuracy of car heights." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.335, + 0.376, + 0.353 + ], + "angle": 0, + "content": "6 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.366, + 0.334, + 0.38 + ], + "angle": 0, + "content": "6.1 FUMET" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.785, + 0.451 + ], + "angle": 0, + "content": "We evaluate the accuracy of depth estimators trained on our framework and demonstrate the effectiveness of FUMET as a novel unsupervised metric-scale training method through depth prediction accuracy, its applicability to any network architecture, and its ability to leverage any dataset for training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.471, + 0.788, + 0.639 + ], + "angle": 0, + "content": "Implementation Details. We implement FUMET with PyTorch [28]. We employ Monodepth2 [15] with ResNet50 encoder pre-trained on ImageNet [30] as the default depth estimator which is one of the most basic architectures in self-supervised MDE. For all experiments unless mentioned otherwise, we train the model for 50 epochs with a batch size of 8 and set \\(\\tau_{\\mathrm{mid}}\\) to 20. We use Adam optimizer [24] with an initial learning rate of \\(5.0\\times 10^{-5}\\) decayed by half every 15 epochs. During training, we perform the following augmentations with \\(50\\%\\) chance in random order: horizontal flips, brightness adjustment \\((\\pm 0.2)\\), saturation adjustment \\((\\pm 0.2)\\), contrast adjustment \\((\\pm 0.2)\\), and hue jitter \\((\\pm 0.1)\\). For quantitative evaluation, we follow [9] and compute the seven standard metrics: AbsRel, SqRel, RMSE, \\(\\mathrm{RMSE}_{\\log}\\), \\(\\delta < 1.25\\), \\(\\delta < 1.25^2\\), and \\(\\delta < 1.25^3\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.784, + 0.749 + ], + "angle": 0, + "content": "KITTI. Our FUMET is independent of network architectures so we experiment on several models with FUMET as well as Monodepth2: Lite-Mono [52] with ResNet18, HR-Depth [25] with ResNet18, and VADepth [45]. We train these models on the KITTI raw dataset [12], which contains 39,810 images for training and 4,424 for validation, and evaluate it with the KITTI Eigen test split [8] (697 images). For training and evaluation, all images are resized to \\(640 \\times 192\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We show the qualitative results in Fig. 2. The quantitative results shown in Tab. 1 demonstrate that FUMET enables the various architectures to learn metric scale properly. Even when employing the simplest network architecture, i.e., Monodepth2, it outperforms weakly-supervised methods that rely on ground-truth scale labels for training which is not necessary for our method. Similar to VADepth [45], FUMET uses the camera height invariance as supervision though" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "G.Kinoshita et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.271 + ], + "angle": 0, + "content": "Table 1: Quantitative results on the KITTI Eigen test split [8] (\\(640 \\times 192\\) resolution). In the Supervision column, V, G, and CamH represent velocity, gravity, and camera height, respectively. The Scaling column represents whether to perform the median scaling [56]. We denote the best results in bold and the second best ones with an underline for each block. Note that we train VADepth [45] from scratch using ground-truth camera height (1.65m) provided by KITTI, unlike its original paper. FUMET outperforms weakly-supervised methods regardless of the network architectures and without the need for ground truth supervision. In the rescaled results, every model trained with FUMET achieves better results than its original one." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.283, + 0.784, + 0.593 + ], + "angle": 0, + "content": "
MethodSupervisionScalingError (↓)Accuracy (↑)
AbsRelSqRelRMSERMSElogδ<1.25δ<1.252δ<1.253
G2S [3]GPS-0.1090.8604.8550.1980.8650.9540.980
PackNet-SfM [16]V-0.1110.8294.7880.1990.8640.9540.980
Wagstaff et al. [36]CamH-0.1230.9965.2530.2130.8400.9470.978
VADepth [45]CamH-0.1200.9754.9710.2030.8670.9560.979
DynaDepth [53]IMU+V+G-0.1090.7874.7050.1950.8690.9580.981
Ours--0.1080.7854.7360.1950.8710.9580.981
Lite-Mono [52]+Ours--0.1120.7984.6920.1940.8710.9590.981
HR-Depth [25]+Ours--0.1090.7944.7340.1930.8690.9590.982
VADepth [45]+Ours--0.1080.8094.5720.1850.8830.9630.982
G2S [3]GPS✓0.1120.8944.8520.1920.8770.9580.981
PackNet-SfM [16]V✓0.1110.7854.6010.1890.8780.9600.982
Wagstaff et al. [36]CamH✓0.1170.9524.9890.1970.8670.9570.981
VADepth [45]CamH✓0.1100.9774.8720.1870.8890.9620.981
DynaDepth [53]IMU+V+G✓0.1080.7614.6080.1870.8830.9620.982
Ours-✓0.1060.7594.6020.1840.8870.9630.983
Monodepth2 [15]-✓0.1150.9034.8630.1930.8770.9590.981
+ Ours-✓0.1060.7594.6020.1840.8870.9630.983
HR-Depth [25]-✓0.1090.7924.6320.1850.8840.9620.983
+ Ours-✓0.1080.7594.5840.1840.8850.9630.983
Lite-Mono [52]-✓0.1100.8024.6710.1860.8790.9610.982
+ Ours-✓0.1070.7754.5640.1830.8880.9630.983
VADepth [45]-✓0.1040.7744.5520.1810.8920.9650.982
+ Ours-✓0.1020.7604.4730.1780.8970.9660.983
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.646, + 0.788, + 0.722 + ], + "angle": 0, + "content": "VADepth requires the ground-truth camera height and employs a complex network architecture. FUMET with Monodepth2 (row 6 in Tab. 1) achieves higher accuracy. This indicates that measuring camera height precisely is difficult and using unreliable ground-truth camera height causes low accuracy in the scale of estimated depth maps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In the results showing the accuracy of the estimated depth map scaled with ground-truth scale factors via median scaling [56], it is evident that every model trained with FUMET attains superior results to the one trained without FUMET. This shows that our training framework not only enables the models to learn metric scale but also enhances the geometric accuracy. This improvement owes to the camera height loss \\(\\mathcal{L}_{\\mathrm{cam}}\\) and \\(\\mathcal{L}_{\\mathrm{aux}}\\), which makes the model respect road regions and object regions, respectively." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.228, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Unsupervised Training for Metric Monocular Road-Scene Depth Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.162, + 0.782, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.387, + 0.788, + 0.43 + ], + "angle": 0, + "content": "Fig. 2: Qualitative comparison on KITTI. In error maps, the larger depth errors are represented in red, smaller ones are in blue. The model trained with our FUMET predicts more accurate depth maps compared to the weakly-supervised methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.506 + ], + "angle": 0, + "content": "FUMET successfully enables a model to learn metric scale, even though the training dataset of KITTI contains many frames with no or few detected cars. We detail this in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.529, + 0.788, + 0.665 + ], + "angle": 0, + "content": "Cityscapes. To evaluate the ability of FUMET to leverage any dataset for training, we train and evaluate it on the Cityscapes dataset [6]. We follow [43, 46, 48, 56] and collect 69,731 images for training and evaluate on the 1,525 test images. In this dataset, there are multiple sets of camera intrinsics. Since this strongly affects the metric depth estimation [47], we first align the focal lengths of all images to 587.5 by resizing them, and crop them to \\(512 \\times 192\\) resolution. For evaluation, we follow [2, 43] and further crop the estimated image into \\(416 \\times 128\\). In order to align the number of iterations on KITTI, we train for 29 epochs with the same learning rate decreased by half every 8 epochs, and \\(\\tau_{\\mathrm{mid}}\\) is set to 12." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.666, + 0.788, + 0.756 + ], + "angle": 0, + "content": "Tab. 2 shows the quantitative results and FUMET outperforms weakly-supervised methods by wide margins. These methods assume that highly accurate sensor data is available and fail to train scale-aware MDE models on datasets consisting of unreliable sensor data such as Cityscapes. On the other hand, FUMET does not require any sensor measurement other than RGB video, so it can robustly leverage any driving videos for training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Mixed datasets. To demonstrate that FUMET can train a model on a mixed set of datasets to attain strong generalizability, we train with 235,341 images collected from the Argoverse2 [44], Lyft [21], A2D2 [13], and DDAD [16] dataset and evaluate it on the KITTI Eigen test split [8]. Both images for training and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "G.Kinoshita et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.493, + 0.228 + ], + "angle": 0, + "content": "Table 2: Quantitative results on Cityscapes [6]. We train all the models from scratch with the same supervisions as the Supervision column in Tab. 1. FUMET significantly outperforms weakly-supervised methods." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.241, + 0.487, + 0.315 + ], + "angle": 0, + "content": "
Method|AbsRelSqRelRMSEδ<1.25
G2S [3]4.156276.1658.890.046
PackNet-SfM [16]0.5046.63914.900.029
VADepth [45]0.3637.11511.950.295
FUMET0.1251.2886.3590.858
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.145, + 0.788, + 0.229 + ], + "angle": 0, + "content": "Table 3: FUMET can be trained on mixed datasets (Argoverse2, Lyft, A2D2, and DDAD) with various camera settings, and attains high generalizability. The results are evaluated on the KITTI Eigen split [8] \\((832\\times 512\\) resolution)." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.241, + 0.785, + 0.299 + ], + "angle": 0, + "content": "
Training datasetAbsRelSqRelRMSEδ<1.25
KITTI0.1030.6754.7080.903
Mixed0.1130.9165.0090.883
Mixed+KITTI0.0820.6114.3070.923
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.345, + 0.784, + 0.42 + ], + "angle": 0, + "content": "evaluation are resized and cropped to \\(832 \\times 512\\) resolution. The training images are the sets of video sequences and each sequence is captured with its own unique camera setting. We individually optimize each camera height for each sequence during training. We detail the dataset composition and training settings in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.785, + 0.51 + ], + "angle": 0, + "content": "As shown in Tab. 3, the model trained with FUMET on the large-scale mixed dataset achieves comparable accuracy to the one trained on KITTI despite the domain gap and attains high generalizability. The model trained on both the mixed dataset and KITTI outperforms the model trained only on KITTI. This suggests that training on various images enhances accuracy which is the strength of fully self-supervised learning methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.787, + 0.557 + ], + "angle": 0, + "content": "We also show that FUMET can leverage in-the-wild videos for training by using camera intrinsics recovered with SLAM, e.g., COLMAP [32] in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Ablation Studies on KITTI. To study the effect of the proposed components in FUMET, we conduct ablation studies on the training losses, balancing the loss weights, camera height optimization, and outlier filtering. The results are shown in Tab. 4. Thanks to the accuracy of LSP and the robustness of Silhouette Projector, the camera height loss makes a greater contribution to the depth accuracy than the auxiliary rough geometric loss. By introducing the camera height optimization, the model can learn the scale with more accurate and consistent supervision than leveraging the size prior independently for each frame (rows 1, 2, and 6 in Tab. 4), which leads to improved accuracy. Though balancing the loss weights (Eq. (14)) with either \\(\\mathcal{L}_{\\mathrm{aux}}\\) or \\(\\mathcal{L}_{\\mathrm{cam}}\\) degrades the results due to the insufficient strength of the supervisory signals for metric scale, it contributes to improving accuracy when both \\(\\mathcal{L}_{\\mathrm{aux}}\\) and \\(\\mathcal{L}_{\\mathrm{cam}}\\) are employed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Our proposed method can also be used with a pre-trained MDE model in an offline manner. We first train the original Monodepth2 [15] without FUMET, then predict depth maps of training images. With Silhouette Projector and LSP, we obtain per-frame scaled camera height and optimize it across frames with the median operation. We fix this value and use it as the camera height supervision" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.228, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Unsupervised Training for Metric Monocular Road-Scene Depth Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.508, + 0.214 + ], + "angle": 0, + "content": "Table 4: Each loss contributes to learning metric scale and the others enhance the depth accuracy. Bal: balancing the loss weights (Eq. (14)). Opt: camera height optimization. Flt: outlier filtering." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.227, + 0.508, + 0.335 + ], + "angle": 0, + "content": "
\\( \\mathcal{L}_{\\text{aux}} \\)\\( \\mathcal{L}_{\\text{cam}} \\)BalOptFltAbsRelSqRelRMSE\\( \\delta < 1.25 \\)
✓---✓0.1250.8875.0550.834
✓-✓-✓0.1490.9505.2190.785
-✓-✓✓0.1160.8754.8740.863
-✓✓✓✓0.2111.3036.1430.585
✓✓-✓✓0.1220.8805.0170.849
✓✓✓-✓0.1150.8224.8110.858
✓✓✓✓-0.1160.8234.9010.860
✓✓✓✓✓0.1080.7854.7360.871
" + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.145, + 0.785, + 0.27 + ], + "angle": 0, + "content": "Table 5: Training with a fine-tuned pseudo camera height achieves the highest accuracy. Online: optimizing the camera height while training. Offline: training with a fixed one computed from depth maps predicted with pre-trained Monodepth2. Fine-tune: training with the fixed one computed offline for 20 epochs, then optimizing online." + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.283, + 0.782, + 0.34 + ], + "angle": 0, + "content": "
Deriv. methodAbsRelSqRelRMSEδ<1.25
Online0.1080.8084.7400.870
Offline0.1060.7904.7190.873
Fine-tune0.1050.7824.6820.876
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.785, + 0.578 + ], + "angle": 0, + "content": "for an untrained Monodepth2 on FUMET. We examine the difference between the proposed camera height optimized online and the fixed one derived offline when we use them as the pseudo-supervision. Note that we also fix \\(\\lambda_{\\mathrm{cam}}\\) because the pseudo-label can be considered sufficiently accurate from the early stage of training. We also run another experiment in which we fix camera height supervision derived offline for the first 20 epochs, which is the same number of training epochs as the original Monodepth2, then unfit the camera height pseudo-label to fine-tune it via online optimization. As shown in Tab. 5, the model supervised with the fixed camera height computed offline achieves slightly better results than the one supervised with the camera height optimized online. This result indicates that our strategy of deriving scaled camera height is highly accurate and fixing it stabilizes the training. The model trained with the fine-tuned camera height achieves the highest accuracy, which shows that our FUMET can attain better accuracy with additional pre-training costs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.599, + 0.296, + 0.612 + ], + "angle": 0, + "content": "6.2 LSP" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.623, + 0.746, + 0.637 + ], + "angle": 0, + "content": "We provide implementation details of LSP in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Evaluation on KITTI and Cityscapes. We evaluate the generalizability of our LSP with the dimensions of the annotated 3D bounding box on both the KITTI [12] and the Cityscapes [6] datasets in the Car class. We compare the results of LSP to a fixed car height prior (1.59m, following [20]) and MonoCIinIS [19], a 3D object detection model that outputs object dimensions from an image and an instance mask. We train MonoCIinIS from scratch with the KITTI dataset. As an evaluation metric, we compute the absolute relative error for each dimension. When we estimate car dimensions with LSP and MonoCIinIS, we use a ground-truth instance mask for making a car image mask and as input, respectively. For all the experiments in LSP, input vehicle masks are cropped and resized to \\(300 \\times 300\\) resolution. On the KITTI dataset, we extract 3,769 samples for evaluation [4]. On the Cityscapes dataset, we evaluate on the validation split," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "G.Kinoshita et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.214 + ], + "angle": 0, + "content": "Table 6: Generalizability on KITTI and Cityscapes. We evaluate LSP on Absolute Relative Error for each dimension (height/width/length), comparing with the fixed car height prior (1.59 m, following [20]) and MonoCInIS [19] trained on KITTI. LSP robustly surpasses other methods in most cases and can predict accurately regardless of the datasets." + }, + { + "type": "table", + "bbox": [ + 0.24, + 0.227, + 0.766, + 0.335 + ], + "angle": 0, + "content": "
MethodEasyModerateHardAll
KITTIFixed Height [20]0.065 / - / -0.064 / - / -0.065 / - / -0.065 / - / -
MonoClIns [19]0.069 / 0.065 / 0.0930.112 / 0.106 / 0.1330.268 / 0.263 / 0.2690.159 / 0.155 / 0.177
LSP (only height)0.057 / - / -0.089 / - / -0.106 / - / -0.087 / - / -
LSP0.046 / 0.136 / 0.1300.060 / 0.144 / 0.1390.073 / 0.155 / 0.1500.062 / 0.147 / 0.145
CityscapesFixed Height [20]0.093 / - / -0.093 / - / -0.076 / - / -0.092 / - / -
MonoClIns [19]0.578 / 0.599 / 0.5840.759 / 0.766 / 0.7670.843 / 0.852 / 0.8520.851 / 0.859 / 0.856
LSP (only height)0.063 / - / -0.075 / - / -0.103 / - / -0.096 / - / -
LSP0.056 / 0.069 / 0.0670.068 / 0.068 / 0.0780.069 / 0.059 / 0.0750.080 / 0.064 / 0.077
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.784, + 0.428 + ], + "angle": 0, + "content": "which has 500 samples. On the KITTI dataset, each object is separated into 3 levels based on detection difficulty: Easy, Moderate, and Hard. Considering the difference in focal lengths between KITTI and Cityscapes, we also separate objects on Cityscapes into the same levels as on KITTI." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.538 + ], + "angle": 0, + "content": "As shown in Tab. 6, LSP outperforms the other two methods in most cases. Particularly, the accuracy of MonoCInIS drops on Cityscapes since it overfits to the training dataset, i.e., KITTI and cannot address the domain gap. In contrast, thanks to the large training data and various augmentations, our LSP obtains the highest generalizability and robustly estimates dimensions on any dataset. Learning not only car height but also width and length also improves the prediction accuracy as this facilitates estimation of the overall car sizes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.57, + 0.36, + 0.587 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this work, we proposed FUMET, a novel training framework for any MDE model to learn metric scale. FUMET aggregates the scale information obtained from a learned car size prior into the camera height and uses it as pseudosupervision formulated as optimization across frames and epochs by leveraging its invariance. We showed that FUMET enables monocular depth estimators to embody metric scale regardless of the underlying network architectures and achieves state-of-the-art absolute-scale depth estimation accuracy on both the KITTI and the Cityscapes datasets, and demonstrated that it can leverage any dataset for training. We also proposed LSP, which can be easily trained on a large-scale dataset without manual annotations and robustly estimates vehicle dimensions from their appearances. We showed that the learned size prior is superior to a fixed one and outperforms a similar 3D object detection model. FUMET can be leveraged as a drop-in training scheme for any monocular depth estimation method. We hope it will be a versatile tool for road scene understanding research." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Unsupervised Training for Metric Monocular Road-Scene Depth Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.179, + 0.785, + 0.209 + ], + "angle": 0, + "content": "This work was in part supported by JSPS 20H05951 and 21H04893, JST JP-MJCR20G7 and JPMJAP2305, and RIKEN GRP." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.234, + 0.323, + 0.25 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.267, + 0.785, + 0.295 + ], + "angle": 0, + "content": "1. Bhat, S.F., Alhashim, I., Wonka, P.: AdaBins: Depth Estimation Using Adaptive Bins. In: CVPR. pp. 4009-4018 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.297, + 0.785, + 0.337 + ], + "angle": 0, + "content": "2. Casser, V., Pirk, S., Mahjourian, R., Angelova, A.: Unsupervised Monocular Depth and Ego-Motion Learning with Structure and Semantics. In: CVPRW. pp. 381-388 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.339, + 0.785, + 0.379 + ], + "angle": 0, + "content": "3. Chawla, H., Varma, A., Arani, E., Zonooz, B.: Multimodal Scale Consistency and Awareness for Monocular Self-Supervised Depth Estimation. In: ICRA. pp. 5140-5146 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.381, + 0.785, + 0.421 + ], + "angle": 0, + "content": "4. Chen, X., Kundu, K., Zhu, Y., Berneshawi, A.G., Ma, H., Fidler, S., Urtasun, R.: 3D Object Proposals for Accurate Object Class Detection. Adv. Neural Inform. Process. Syst. 28 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.423, + 0.785, + 0.449 + ], + "angle": 0, + "content": "5. Chen, X., Li, T.H., Zhang, R., Li, G.: Frequency-Aware Self-Supervised Monocular Depth Estimation. In: WACV. pp. 5808-5817 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.451, + 0.785, + 0.491 + ], + "angle": 0, + "content": "6. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The Cityscapes Dataset for Semantic Urban Scene Understanding. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.493, + 0.785, + 0.533 + ], + "angle": 0, + "content": "7. Ding, M., Huo, Y., Yi, H., Wang, Z., Shi, J., Lu, Z., Luo, P.: Learning Depth-Guided Convolutions for Monocular 3D Object Detection. In: CVPR. pp. 11669-11678 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.535, + 0.785, + 0.575 + ], + "angle": 0, + "content": "8. Eigen, D., Fergus, R.: Predicting Depth, Surface Normals and Semantic Labels with a Common Multi-Scale Convolutional Architecture. In: ICCV. pp. 3213-3223 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.577, + 0.785, + 0.617 + ], + "angle": 0, + "content": "9. Eigen, D., Puhrsch, C., Fergus, R.: Depth Map Prediction from a Single Image Using a Multi-Scale Deep Network. In: Adv. Neural Inform. Process. Syst. vol. 2, pp. 2366-2374 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.619, + 0.785, + 0.658 + ], + "angle": 0, + "content": "0. Frost, D., Prisacariu, V., Murray, D.: Recovering Stable Scale in Monocular SLAM Using Object-Supplemented Bundle Adjustment. IEEE Transactions on Robotics 34(3), 736-747 (2018). https://doi.org/10.1109/TR0.2018.2820722" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.785, + 0.687 + ], + "angle": 0, + "content": "1. Garg, R., Bg, V.K., Carneiro, G., Reid, I.: Unsupervised CNN for Single View Depth Estimation: Geometry to the Rescue. In: ECCV. pp. 740-756 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.728 + ], + "angle": 0, + "content": "2. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision Meets Robotics: The KITTI Dataset. International Journal of Robotics Research (IJRR) 32(11), 1231--1237 (2013). https://doi.org/10.1177/0278364913491297" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.785, + 0.784 + ], + "angle": 0, + "content": "3. Geyer, J., Kassahun, Y., Mahmudi, M., Ricou, X., Durgesh, R., Chung, A.S., Hauswald, L., Pham, V.H., Mühlegg, M., Dorn, S., Fernandez, T., Janicke, M., Mirashi, S., Savani, C., Sturm, M., Vorobiov, O., Oelker, M., Garreis, S., Schuberth, P.: A2D2: Audi Autonomous Driving Dataset. CoRR abs/2004.06320 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "4. Godard, C., Mac Aodha, O., Brostow, G.J.: Unsupervised Monocular Depth Estimation with Left-Right Consistency. In: CVPR. pp. 6602-6611 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging Into Self-Supervised Monocular Depth Estimation. In: ICCV. pp. 3827-3837 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.267, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.128 + ], + "angle": 0, + "content": "G.Kinoshita et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Guizilini, V., Ambrus, R., Pillai, S., Raventos, A., Gaidon, A.: 3D Packing for Self-Supervised Monocular Depth Estimation. In: CVPR. pp. 2482-2491 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "17. Hartley, R., Zisserman, A.: Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edn. (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "18. He, M., Hui, L., Bian, Y., Ren, J., Xie, J., Yang, J.: RA-Depth: Resolution Adaptive Self-supervised Monocular Depth Estimation. In: ECCV. pp. 565–581 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "19. Heylen, J., De Wolf, M., Dawagne, B., Proesmans, M., Van Gool, L., Abbeloos, W., Abdelkawy, H., Reino, D.O.: MonoCInIS: Camera Independent Monocular 3D Object Detection Using Instance Segmentation. In: ICCVW. pp. 923-934 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "20. Hoiem, D., Efros, A.A., Hebert, M.: Putting Objects in Perspective. IJCV 80, 3-15 (2008). https://doi.org/10.1007/s11263-008-0137-5" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.301, + 0.785, + 0.342 + ], + "angle": 0, + "content": "21. Houston, J., Zuidhof, G., Bergamini, L., Ye, Y., Chen, L., Jain, A., Omari, S., Iglovikov, V., Ondruska, P.: One Thousand and One Hours: Self-Driving Motion Prediction Dataset. In: Conference on Robot Learning. pp. 409-418 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "22. Jain, J., Li, J., Chiu, M.T., Hassani, A., Orlov, N., Shi, H.: OneFormer: One Transformer to Rule Universal Image Segmentation. In: CVPR. pp. 2989-2998 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "23. Kim, D., Ga, W., Ahn, P., Joo, D., Chun, S., Kim, J.: Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth. CoRR abs/2201.07436 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "24. Kingma, D.P., Ba, J.: Adam: A Method for Stochastic Optimization. CoRR abs/1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.493 + ], + "angle": 0, + "content": "25. Lyu, X., Liu, L., Wang, M., Kong, X., Liu, L., Liu, Y., Chen, X., Yuan, Y.: HR-Depth: High Resolution Self-Supervised Monocular Depth Estimation. In: AAAI. vol. 35, pp. 2294-2301 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.535 + ], + "angle": 0, + "content": "26. Milli, E., Erkent, O., Yilmaz, A.E.: Multi-Modal Multi-Task (3MT) Road Segmentation. IEEE Robotics and Automation Letters 8(9), 5408-5415 (2023). https://doi.org/10.1109/LRA.2023.3295254" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.563 + ], + "angle": 0, + "content": "27. Ning, C., Gan, H.: Trap Attention: Monocular Depth Estimation with Manual Traps. In: CVPR. pp. 5033-5043 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.632 + ], + "angle": 0, + "content": "28. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: PyTorch: An Imperative Style, High-Performance Deep Learning Library. In: Adv. Neural Inform. Process. Syst. pp. 8026--8037 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.66 + ], + "angle": 0, + "content": "29. Piccinelli, L., Sakaridis, C., Yu, F.: iDisc: Internal Discretization for Monocular Depth Estimation. In: CVPR. pp. 21477-21487 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "30. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet: A Large Scale Visual Recognition Challenge. IJCV 115, 211-252 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.742 + ], + "angle": 0, + "content": "31. Sadat, A., Casas, S., Ren, M., Wu, X., Dhawan, P., Urtasun, R.: Perceive, Predict, and Plan: Safe Motion Planning Through Interpretable Semantic Representations. In: ECCV. pp. 414-430 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.771 + ], + "angle": 0, + "content": "32. Schonberger, J.L., Frahm, J.M.: Structure-From-Motion Revisited. In: CVPR. pp. 4104-4113 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "33. Simonelli, A., Bulò, S.R., Porzi, L., Kontschieder, P., Ricci, E.: Are We Missing Confidence in Pseudo-LiDAR Methods for Monocular 3D Object Detection? In: ICCV. pp. 3225-3233 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "34. Spencer, J., Russell, C., Hadfield, S., Bowden, R.: Kick Back & Relax: Learning to Reconstruct the World by Watching SlowTV (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.228, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Unsupervised Training for Metric Monocular Road-Scene Depth Estimation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "35. Sucar, E., Hayet, J.B.: Bayesian Scale Estimation for Monocular SLAM Based on Generic Object Detection for Correcting Scale Drift. In: ICRA. pp. 5152-5158 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "36. Wagstaff, B., Kelly, J.: Self-Supervised Scale Recovery for Monocular Depth and Egomotion Estimation. In: IROS. pp. 2620-2627 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.245 + ], + "angle": 0, + "content": "37. Wang, C., Buenaposada, J.M., Zhu, R., Lucey, S.: Learning Depth from Monocular Videos Using Direct Methods. In: CVPR. pp. 2022-2030 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "38. Wang, H., Cai, P., Fan, R., Sun, Y., Liu, M.: End-to-End Interactive Prediction and Planning with Optical Flow Distillation for Autonomous Driving. In: CVPRW. pp. 2229–2238 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.328 + ], + "angle": 0, + "content": "39. Wang, Y., Chao, W.L., Garg, D., Hariharan, B., Campbell, M., Weinberger, K.Q.: Pseudo-LiDAR From Visual Depth Estimation: Bridging the Gap in 3D Object Detection for Autonomous Driving. In: CVPR. pp. 8437-8445 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.785, + 0.37 + ], + "angle": 0, + "content": "40. Wang, Y., Chen, X., You, Y., Li, L.E., Hariharan, B., Campbell, M., Weinberger, K.Q., Chao, W.L.: Train in Germany, Test in the USA: Making 3D Object Detectors Generalize. In: CVPR. pp. 11710-11720 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.412 + ], + "angle": 0, + "content": "41. Wang, Y., Yang, H., Cai, J., Wang, G., Wang, J., Huang, Y.: Unsupervised Learning of Depth and Pose Based on Monocular Camera and Inertial Measurement Unit (IMU). In: ICRA. pp. 10010-10017 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.453 + ], + "angle": 0, + "content": "42. Wang, Z., Bovik, A., Sheikh, H., Simoncelli, E.: Image Quality Assessment: From Error Visibility to Structural Similarity. IEEE Transactions on Image Processing 13(4), 600-612 (2004). https://doi.org/10.1109/TIP.2003.819861" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.494 + ], + "angle": 0, + "content": "43. Watson, J., Mac Aodha, O., Prisacariu, V., Brostow, G., Firman, M.: The Temporal Opportunist: Self-Supervised Multi-Frame Monocular Depth. In: CVPR. pp. 1164–1174 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.55 + ], + "angle": 0, + "content": "44. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., Ramanan, D., Carr, P., Hays, J.: Argoverse 2: Next Generation Datasets for Self-driving Perception and Forecasting. In: NeurIPS Datasets and Benchmarks (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.605 + ], + "angle": 0, + "content": "45. Xiang, J., Wang, Y., An, L., Liu, H., Wang, Z., Liu, J.: Visual Attention-Based Self-Supervised Absolute Depth Estimation Using Geometric Priors in Autonomous Driving. IEEE Robotics and Automation Letters 7(4), 11998-12005 (2022). https://doi.org/10.1109/LRA.2022.3210298" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.633 + ], + "angle": 0, + "content": "46. Yang, Z., Wang, P., Wang, Y., Xu, W., Nevatia, R.: LEGO: Learning Edge with Geometry All at Once by Watching Videos. In: CVPR. pp. 225-234 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.675 + ], + "angle": 0, + "content": "47. Yin, W., Zhang, C., Chen, H., Cai, Z., Yu, G., Wang, K., Chen, X., Shen, C.: Metric3D: Towards Zero-Shot Metric 3D Prediction from a Single Image. In: ICCV. pp. 9043-9053 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "48. Yin, Z., Shi, J.: GeoNet: Unsupervised Learning of Dense Depth, Optical Flow and Camera Pose. In: CVPR. pp. 1983-1992 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.743 + ], + "angle": 0, + "content": "49. You, Y., Wang, Y., Chao, W.L., Garg, D., Pleiss, G., Hariharan, B., Campbell, M., Weinberger, K.Q.: Pseudo-LiDAR++: Accurate Depth for 3D Object Detection in Autonomous Driving. In: ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.772 + ], + "angle": 0, + "content": "50. Yuan, W., Gu, X., Dai, Z., Zhu, S., Tan, P.: Neural Window Fully-Connected CRFs for Monocular Depth Estimation. In: CVPR. pp. 3916-3925 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "51. Zeng, W., Luo, W., Suo, S., Sadat, A., Yang, B., Casas, S., Urtasun, R.: End-To-End Interpretable Neural Motion Planner. In: CVPR. pp. 8652-8661 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "52. Zhang, N., Nex, F., Vosselman, G., Kerle, N.: Lite-Mono: A Lightweight CNN and Transformer Architecture for Self-Supervised Monocular Depth Estimation. In: CVPR. pp. 18537-18546 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.393, + 0.127 + ], + "angle": 0, + "content": "G.Kinoshita et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "53. Zhang, S., Zhang, J., Tao, D.: Towards Scale-Aware, Robust, and Generalizable Unsupervised Monocular Depth Estimation by Integrating IMU Motion Dynamics. In: ECCV. pp. 143-160 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.217 + ], + "angle": 0, + "content": "54. Zhang, S., Li, X., Liu, Y., Fu, H.: Scale-aware Insertion of Virtual Objects in Monocular Videos. In: ISMAR. pp. 36-44 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "55. Zhou, H., Greenwood, D., Taylor, S.: Self-Supervised Monocular Depth Estimation with Internal Feature Fusion. In: BMVC (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.273 + ], + "angle": 0, + "content": "56. Zhou, T., Brown, M., Snavely, N., Lowe, D.G.: Unsupervised Learning of Depth and Ego-Motion from Video. In: CVPR. pp. 6612-6619 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.301 + ], + "angle": 0, + "content": "57. Zhu, R., Yang, X., Hold-Geoffroy, Y., Perazzi, F., Eisenmann, J., Sunkavalli, K., Chandraker, M.: Single View Metrology in the Wild. In: ECCV. pp. 316-333 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.301 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Camera Height Doesn't Change_ Unsupervised Training for Metric Monocular Road-Scene Depth Estimation/b1560f9a-1722-4ac2-a48d-d70d8859d7c8_origin.pdf b/2024/Camera Height Doesn't Change_ Unsupervised Training for Metric Monocular Road-Scene Depth Estimation/b1560f9a-1722-4ac2-a48d-d70d8859d7c8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0224cd4b851b5f5713fd56e2c1386d8a130083f4 --- /dev/null +++ b/2024/Camera Height Doesn't Change_ Unsupervised Training for Metric Monocular Road-Scene Depth Estimation/b1560f9a-1722-4ac2-a48d-d70d8859d7c8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8542888806ee5a14ffe2802d7ee99a873c14a8d6ba75d31167ed277de8be182b +size 2402187 diff --git a/2024/Camera Height Doesn't Change_ Unsupervised Training for Metric Monocular Road-Scene Depth Estimation/full.md b/2024/Camera Height Doesn't Change_ Unsupervised Training for Metric Monocular Road-Scene Depth Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cb69974bd58ca4d07b83350feb33a05dc97ff040 --- /dev/null +++ b/2024/Camera Height Doesn't Change_ Unsupervised Training for Metric Monocular Road-Scene Depth Estimation/full.md @@ -0,0 +1,327 @@ +# Camera Height Doesn't Change: Unsupervised Training for Metric Monocular Road-Scene Depth Estimation + +Genki Kinoshita and Ko Nishino + +Graduate School of Informatics, Kyoto University, Kyoto, Japan +gkinoshita@vision.ist.i.kyoto-u.ac.jp, kon@i.kyoto-u.ac.jp +https://vision.ist.i.kyoto-u.ac.jp/ + +Abstract. In this paper, we introduce a novel training method for making any monocular depth network learn absolute scale and estimate metric road-scene depth just from regular training data, i.e., driving videos. We refer to this training framework as FUMET. The key idea is to leverage cars found on the road as sources of scale supervision and to incorporate them in network training robustly. FUMET detects and estimates the sizes of cars in a frame and aggregates scale information extracted from them into an estimate of the camera height whose consistency across the entire video sequence is enforced as scale supervision. This realizes robust unsupervised training of any, otherwise scale-oblivious, monocular depth network so that they become not only scale-aware but also metric-accurate without the need for auxiliary sensors and extra supervision. Extensive experiments on the KITTI and the Cityscapes datasets show the effectiveness of FUMET, which achieves state-of-the-art accuracy. We also show that FUMET enables training on mixed datasets of different camera heights, which leads to larger-scale training and better generalization. Metric depth reconstruction is essential in any road-scene visual modeling, and FUMET democratizes its deployment by establishing the means to convert any model into a metric depth estimator. + +Keywords: Unsupervised Monocular Depth Estimation $\cdot$ Metric Scale $\cdot$ Camera Height Invariance $\cdot$ Object Size Prior + +# 1 Introduction + +Monocular depth estimation is essential for autonomous driving and advanced driver assistance systems (ADAS). It underlies many of their key perceptual tasks including 3D object detection [7,33,39,49], motion planning [31,38,51], and road segmentation [26]. Although supervised methods have achieved impressive accuracy, they suffer from costly ground-truth data collection. This limits their generalizability [34] and also makes continuous learning difficult. + +Many recent methods [5, 15, 18, 25, 52] leverage self-supervision to scale up training. Most of these methods, however, suffer from scale ambiguity which renders them difficult to use off-the-shelf. To resolve the scale ambiguity and + +recover in metric scale, the actual scene scale needs to be supervised in some form. Past works have leveraged velocity [16], GPS [3], IMU [41, 53], or camera height [36, 45] for this. These methods require auxiliary sensors to measure these extra quantities in addition to the monocular RGB camera, which significantly limits the ability to scale the size of training data. Arbitrary in-the-wild driving videos cannot be used as they lack such measurements. These self-supervised methods often still require weak scale supervision and cannot be made fully self-supervised as the scale changes for each sequence captured with a different camera. + +Road scenes actually have plenty of untapped sources of metric scale. Cars found on the road are one of them as they are rigid objects whose actual sizes do not change and are unique to each make and model. The road is also exactly the region where we need accurate metric depth the most. If we can leverage the sizes of cars to bridge the 3D to 2D projection, we may train a mono-depth network to become metric-accurate. Simply employing a prior on the car size, however, would be too brittle since the metric supervision will be as ambiguous as the accuracy of that prior. The perspective projection and the variations in car makes, years, and colors, lead to a large variance in the size estimates, which will directly affect the mono-depth metric accuracy. In addition, there are no reliable object size priors except for a few introduced as probability densities for other tasks [20,54,57] whose parameters are not shared. + +How can we train an arbitrary mono-depth estimator to become metric-accurate by robustly leveraging prior knowledge about the 3D sizes of cars found in driving videos? We derive fully unsupervised metric depth trainer (FUMET) to answer this question. Our key idea is to aggregate car size cues in the frames and the sequence into a single physical measure, namely the camera height. Regardless of the scene, the camera capturing it is the same for the same sequence. This means that, by transforming the car size cues found in different frames into a camera height estimate, the very fact that it should not change can be used as a supervision of invariance to achieve metric depth reconstruction. We can formulate this as an optimization of the camera height across frames and training epochs. This leads to gradual metric-aware learning that is stable and accurate, which also makes the depth estimates consistent across the sequence. + +We introduce a learned size prior (LSP) to estimate the size of each car found in a frame. This prior gives us the vehicle dimensions from its appearance. It is trained with various augmentations on a large-scale dataset without the need for any manual annotation. By comparing these dimension estimates with those computed from the depth estimates, we obtain a per-frame scale factor. The multiplication of this scale estimate with the camera height estimate derived from the estimated depth becomes the camera height estimate of the frame. These estimated camera heights across all frames in a sequence are then consolidated by using their median value which is then used as scale supervision with its weighted moving average at the end of each training epoch. + +We evaluate the effectiveness of FUMET with extensive experiments using the KITTI and the Cityscapes datasets and compare its accuracy with a number + +of weakly-supervised methods. The results clearly show that FUMET succeeds in making monocular depth estimation methods learn metric scales and achieves the state-of-the-art accuracy without any direct depth and scale supervision. We also demonstrate its capability to train a model on mixed datasets collected with different cameras, i.e., camera heights. This enables larger-scale training and leads to higher generalization. We believe FUMET can serve as a fundamental building block for monocular depth estimation in the wild. + +# 2 Related work + +# 2.1 Self-Supervision + +Although, as in other tasks, supervised methods achieve the highest accuracy in monocular depth estimation (MDE) [1, 23, 27, 29, 50], they obviously require ground-truth data usually obtained with a LiDAR sensor. Self-supervised MDE methods remove the need for such hardware costs for training. Garg et al. [11] proposed the first approach to self-supervised MDE by predicting per-pixel disparities between a stereo pair. Godard et al. [14] extend this idea by introducing left-right consistency and greatly improve the accuracy. These methods require calibrated stereo data for training which limits their applicability. Zhou et al. [56] introduced a pioneering method for training solely from monocular videos by incorporating a network that estimates the relative camera poses between successive frames. Based on this idea, many MDE methods that can be trained solely with monocular videos have been proposed [5, 15, 18, 25, 52, 55]. + +Despite the popularity and improvements of self-supervised MDE, directly deploying those models in autonomous driving or ADAS systems is unrealistic as they only produce relative depth up to an unknown scale. Our FUMET overcomes this issue without the need for any external supervision. + +# 2.2 Weak Supervision + +Several works have tackled scale-aware MDE by adopting weak scale supervision through velocity [16], GPS [3], IMU [41], or camera height [36,45] measurements. Zhang et al. [53] use the combination of IMU, velocity, and gravity to achieve the highest accuracy among these methods. These methods fundamentally rely on such auxiliary measurements in addition to regular RGB images. This makes it impossible to take advantage of the many road-scene videos on the Internet (e.g., YouTube) or casually installed dashcam videos for training. In contrast, FUMET relies on a universal car size prior which is independent of the data collection environments and does not require any such weak supervision; it can be trained solely from monocular video sequences. + +# 2.3 Object Size Priors + +Several works for other tasks leverage priors on object sizes including single view metrology [57], object insertion [20,54], and visual odometry [10,35]. These + +methods model the priors as normal distributions or Gaussian mixture models, and the parameters of these distributions are predetermined. When there is a gap between those values and the true ones, which is usually the case in the real world, the difference directly surfaces in erroneous scale estimates. + +Instead of these fixed priors, we introduce a learning-based size prior which we refer to as LSP. This estimates vehicle dimensions based on its appearance, thus can avoid the aforementioned problem. Furthermore, it is trained on a large-scale dataset with various augmentations, so that it can robustly estimate vehicle dimensions from any image. + +# 3 Preliminaries + +Our goal is to train a metric-accurate MDE model $\theta_{\mathrm{depth}}$ with only self-supervision. This self-supervision is comprised of two parts: one designed for learning metric scale (Sec. 4) and the other dedicated to learning depth geometry. For the latter, we follow Zhou et al. [56] and minimize the reconstruction loss between the current frame $I_{t}$ and the ones synthesized from temporally adjacent frames $I_{s}$ , $(s = t \pm 1)$ . We employ a pose network $\theta_{\mathrm{pose}}$ in addition to $\theta_{\mathrm{depth}}$ to estimate the relative camera pose $\mathbf{T}_{t \to s} \in \mathrm{SE}(3)$ and the current dense depth map $D_{t}$ , respectively, + +$$ +D _ {t} = \theta_ {\text {d e p t h}} \left(I _ {t}\right), \mathbf {T} _ {t \rightarrow s} = \theta_ {\text {p o s e}} \left(I _ {t}, I _ {s}\right). \tag {1} +$$ + +We then synthesize the image from the current viewpoint + +$$ +I _ {s \rightarrow t} = I _ {s} \left\langle \operatorname {p r o j} \left(D _ {t}, \mathbf {T} _ {t \rightarrow s}, \mathbf {K}\right)\right\rangle , \tag {2} +$$ + +where $\langle \cdot \rangle$ is a bilinear sampling operator, $proj(\cdot)$ outputs the 2D coordinates of $D_{t}$ when projected into the viewpoint of $I_{s}$ , and $\mathbf{K}$ is the camera intrinsic. Following Godard et al. [15], we use an affine combination of SSIM [42] and L1 loss as the photometric error $pe$ and define the reconstruction loss $\mathcal{L}_{\mathrm{rec}}$ as the photometric error between $I_{t}$ and $I_{s\rightarrow t}$ for the pixels with minimum error across source frames + +$$ +\mathcal {L} _ {\text {r e c}} = \min _ {s} p e \left(I _ {t}, I _ {s \rightarrow t}\right), \tag {3} +$$ + +$$ +p e \left(I _ {t}, I _ {s \rightarrow t}\right) = \frac {\lambda}{2} \left(1 - \operatorname {S S I M} \left(I _ {t}, I _ {s \rightarrow t}\right)\right) + (1 - \lambda) \left| I _ {t} - I _ {s \rightarrow t} \right|, \tag {4} +$$ + +where $\lambda = 0.85$ . We also employ an edge-aware smoothness loss [14] + +$$ +\mathcal {L} _ {\mathrm {s m}} = \left| \partial_ {x} d _ {t} ^ {*} \right| e ^ {- \left| \partial_ {x} I _ {t} \right|} + \left| \partial_ {y} d _ {t} ^ {*} \right| e ^ {- \left| \partial_ {y} I _ {t} \right|}, \tag {5} +$$ + +where $d_{t}^{*} = d_{t} / \bar{d}_{t}$ is the mean-normalized inverse depth [37] to prevent the estimated depth from shrinking. Furthermore, when computing $\mathcal{L}_{\mathrm{rec}}$ we use automasking [15] to mask out stationary pixels which generate incorrect supervisory signals. + +![](images/0f6dd62890ff3d85b780da962060a8c8ac07d2a640ae93d3bee87fc2121218aa.jpg) +Fig. 1: Overview of FUMET. At each training step $n$ , an unscaled camera height $H_{\mathrm{cam}}^{n,\tau}$ is computed differentiably from the estimated depth. The previous epoch $\tau - 1$ provides supervision with a scaled camera height $H_{\mathrm{cam}}^{*\tau - 1}$ . To obtain this scaled camera height supervision, Silhouette Projector first computes the object silhouette heights $H_{\mathrm{obj}}'$ from the depth map. By comparing $H_{\mathrm{obj}}'$ and the estimated one $H_{\mathrm{obj}}$ with LSP, per-frame scale factor $s$ is determined and we obtain the scaled camera height $H_{\mathrm{cam}}^{n,\tau} = s \cdot H_{\mathrm{cam}}'^{\prime,\tau}$ . At the end of each epoch, $H_{\mathrm{cam}}^*$ is optimized across a series of consecutive frames and updated with the weighted moving average. +Fig. 1 depicts the overall architecture of FUMET. Please see the caption for an overview of the framework. Let “'” denote “not metrically scaled”. + +# 4 FUMET + +# 4.1 Scale-Aware Self-Supervised Learning + +In FUMET, we derive a camera height pseudo-supervision from depth estimates. To obtain the per-frame camera height, we first compute the normal vector for each pixel $p_i$ . We consider the 8-neighborhood of $p_i$ and form 8 pairs $\mathcal{N}(p_i) = \{(p_{j_0}^i,p_{j_1}^i)\}_{j = 1}^8$ . Each pair $(p_{j_0}^i,p_{j_1}^i)$ consists of two pixels whose difference vectors w.r.t. $p_i$ are perpendicular to each other and are sorted counterclockwise. We reproject the paired pixels onto 3D space and compute the cross-product of the two difference vectors w.r.t. the reprojected 3D point of $p_i$ . By summing and normalizing all the cross-product vectors of each pair, we obtain the normal vector $\hat{\mathbf{n}}(p_i)$ + +$$ +\mathbf {n} \left(p _ {i}\right) = \sum_ {\mathcal {N} \left(p _ {i}\right)} \left(\phi \left(p _ {j _ {0}} ^ {i}\right) - \phi \left(p _ {i}\right)\right) \times \left(\phi \left(p _ {j _ {1}} ^ {i}\right) - \phi \left(p _ {i}\right)\right), \quad \hat {\mathbf {n}} \left(p _ {i}\right) = \frac {\mathbf {n} \left(p _ {i}\right)}{\| \mathbf {n} \left(p _ {i}\right) \|}, \tag {6} +$$ + +where $\phi (\cdot)$ denotes perspective reprojection. + +For pixels in the road region, the negated inner product between the normal vector and the reprojected 3D point corresponds to the camera height from the ground plane + +$$ +H _ {\mathrm {c a m}} ^ {\prime} (p _ {i}) = - \phi (p _ {i}) \cdot \hat {\mathbf {n}} (p _ {i}). \tag {7} +$$ + +We represent the median of $H_{\mathrm{cam}}'(p_i)$ as the camera height of the frame + +$$ +\tilde {H} _ {\mathrm {c a m}} ^ {\prime} = \operatorname {m e d} \left(\left\{H _ {\mathrm {c a m}} ^ {\prime} \left(p _ {i}\right) \mid p _ {i} \in M _ {\mathrm {r}} \right\}\right), \tag {8} +$$ + +where $p_i$ is in the road region mask $M_{\mathrm{r}}$ obtained with an off-the-shelf semantic segmentation model [22] before training. This camera height $\tilde{H}_{\mathrm{cam}}'$ estimate is unscaled and needs to be upgraded to a real-scaled value to impose the invariance loss across frames. By using the Silhouette Projector, we aggregate the scale information from the object height prior into a per-frame scale factor $s$ (see Sec. 4.2) and obtain the scaled camera height as $H_{\mathrm{cam}} = s \cdot \tilde{H}_{\mathrm{cam}}'$ . + +We leverage the fact that the camera height does not change in a sequence and formulate the invariance into an optimization across frames and training epochs. Thanks to this optimization, we acquire an accurate and stable supervision, which is also consistent across the scene. At the end of the $\tau$ -th epoch, we use the median of $\{H_{\mathrm{cam}}^{n,\tau}\}_{n=1}^{N}$ for all frames $N$ as the representative camera height $H_{\mathrm{cam}}^{\tau}$ at the $\tau$ -th epoch. Then, we update the scaled camera height pseudo-invariance $H_{\mathrm{cam}}^{*}$ with its weighted moving average + +$$ +H _ {\mathrm {c a m}} ^ {* \tau} \leftarrow \left\{\frac {\tau (\tau - 1)}{2} \cdot H _ {\mathrm {c a m}} ^ {* \tau - 1} + \tau \cdot H _ {\mathrm {c a m}} ^ {\tau} \right\} / \frac {\tau (\tau + 1)}{2}. \tag {9} +$$ + +When training on a dataset containing images captured at multiple camera heights, we create pseudo-labels corresponding to each camera height and optimize them individually. + +To make the network learn the absolute scale and directly produce the scaled depth map, we enforce a loss on the camera height $H_{\mathrm{cam}}'$ using the pseudo-invariance $H_{\mathrm{cam}}^{*\tau - 1}$ computed in the previous epoch + +$$ +\mathcal {L} _ {\mathrm {c a m}} = \frac {1}{| M _ {\mathrm {r}} |} \sum_ {i} M _ {\mathrm {r}} \left(p _ {i}\right) \cdot \left| H _ {\mathrm {c a m}} ^ {\prime} \left(p _ {i}\right) - H _ {\mathrm {c a m}} ^ {* \tau - 1} \right|. \tag {10} +$$ + +It is worth noting that at inference, the network does not assume camera height invariance in the input sequence; it is only assumed during training. The computational cost at inference is exactly the same as the original MDE model trained without FUMET. + +# 4.2 Silhouette Projector + +In the process of creating a scaled camera height supervision, we need to know the per-frame scale factor. For this, we can consider leveraging an object of known size. This, however, would necessitate extraction of the exact 3D object region which is non-trivial as objects are often occluded or cropped and oriented in various poses. + +We introduce Silhouette Projector to robustly estimate this per-frame scale factor. Silhouette Projector leverages two simple facts of road scenes. First, the height of the object silhouette projected on the plane perpendicular to the ground plane does not change regardless of the object pose. Second, we can compute the silhouette height by measuring the distance from the top of the silhouette to the ground plane even if they are partly occluded or truncated, as long as the top is visible and the object is on the ground plane. + +We first reconstruct an object point cloud with the estimated depth map through perspective reprojection. The object region $M_{\mathrm{obj}}$ in the image can be obtained with an off-the-shelf instance segmentation model [22] before training. We then orthographically project these 3D points onto a plane $P^{\perp}$ perpendicular to the ground plane, i.e., an arbitrary plane parallel to the road normal vector $\tilde{\mathbf{n}}$ . We compute $\tilde{\mathbf{n}}$ as the median of the normal vectors $\hat{\mathbf{n}}(p_i)$ (see Eq. (6)) in the road region $M_{\mathrm{r}}$ . We define the farthest distance from the 2D point on $P^{\perp}$ to the ground plane as the silhouette height. This value is unscaled because we derived this from the estimated unscaled depth map. The ratio between a height prior and its silhouette height corresponds to the scale factor of interest. When there are multiple objects in an image, we use the median of the scale factors each corresponding to each object as the scale factor of the frame for computational simplicity and robustness. + +Outlier Filtering. Some objects cannot be handled well with Silhouette Projector for various reasons including occluded upper parts, not on the ground plane, inaccurate segmented region, and misclassified object category. We introduce an algorithm to detect these as outliers based on geometric plausibility. We first calculate the horizon $\mathbf{l}$ with the road normal vector $\tilde{\mathbf{n}}$ and the camera intrinsics $\mathbf{K}$ [17] + +$$ +\mathbf {l} = \mathbf {K} ^ {- \top} \tilde {\mathbf {n}}. \tag {11} +$$ + +Based on the approximations in [20], we derive the approximate metric-scale object height $\hat{H}_{\mathrm{obj}} \approx \frac{h_{\mathrm{obj}}}{h_{\mathrm{cam}}} H_{\mathrm{cam}}^{*\tau - 1}$ , where $h_{\mathrm{cam}}$ is the farthest distance from the pixel in the object instance mask to $\mathbf{l}$ , $h_{\mathrm{obj}}$ is the farthest distance between two points on the object mask on which tangent lines parallel to $\mathbf{l}$ lie, and $H_{\mathrm{cam}}^{*\tau - 1}$ is the pseudo camera height at the $\tau$ -th epoch derived in Eq. (9). By comparing the relative gap between $\hat{H}_{\mathrm{obj}}$ and the object height prior $H_{\mathrm{obj}}$ , we determine the outliers as + +$$ +\left\{k \left| \frac {\left| H _ {\mathrm {o b j}} ^ {k} - \hat {H} _ {\mathrm {o b j}} ^ {k} \right|}{H _ {\mathrm {o b j}} ^ {k}} > \mathcal {T} \right. \right\}. \tag {12} +$$ + +In all the experiments, we set the threshold $\mathcal{T}$ to 0.2. + +# 4.3 Training Losses + +Auxiliary Rough Geometric Loss. We experimentally observe that estimated object depths, especially those distant from the camera, often erroneously lean towards the road surface because of the inaccurate supervisory signal from the reconstruction loss in low-texture regions. This tendency makes silhouette heights inaccurate. To mitigate this issue, we introduce an auxiliary rough geometric loss $\mathcal{L}_{\mathrm{aux}}$ , which enforces object point clouds to be within a close region. For each inlier object $k$ ( $k = 1 \dots K$ ), we define $\mathcal{L}_{\mathrm{aux}}$ as the gap between the estimated depth map $D$ in the object region $M_{\mathrm{obj}}$ and the approximate object + +depth $D_{\mathrm{aprx}} = \frac{H_{\mathrm{obj}}}{h_{\mathrm{obj}}} f_y$ where $f_{y}$ is the focal length + +$$ +\mathcal {L} _ {\mathrm {a u x}} = \frac {1}{K} \sum_ {k} \frac {1}{| M _ {\mathrm {o b j}} ^ {k} |} \sum_ {i} M _ {\mathrm {o b j}} ^ {k} \left(p _ {i}\right) \cdot \left| D \left(p _ {i}\right) - D _ {\mathrm {a p r x}} ^ {k} \right|. \tag {13} +$$ + +As a by-product, this loss accelerates learning the metric scale. + +The $\mathcal{L}_{\mathrm{aux}}$ is potentially inaccurate since $D_{\mathrm{aprx}}$ assumes that each object can be represented by a plane parallel to the image plane, and we observe that it degrades the depth accuracy in the late stage of training. On the other hand, in the early stage, estimated depth is unreliable, so heavily relying on the camera height loss $\mathcal{L}_{\mathrm{cam}}$ makes the training unstable. For this, we gradually decrease the weight of $\mathcal{L}_{\mathrm{aux}}$ and increase the one for $\mathcal{L}_{\mathrm{cam}}$ as the training proceeds in a logarithmic way. To ensure training stability, we stop changing these weights after the $\tau_{\mathrm{mid-th}}$ epoch + +$$ +\lambda_ {\mathrm {a u x}} (\tau) = \left\{ \begin{array}{l l} \frac {- \log (\tau + 1)}{\log \left(\tau_ {\mathrm {m i d}} + 1\right)} & (\tau \leq \tau_ {\mathrm {m i d}}) \\ \epsilon & (\tau > \tau_ {\mathrm {m i d}}) \end{array} , \lambda_ {\mathrm {c a m}} (\tau) = \left\{ \begin{array}{l l} \frac {\log (\tau + 1)}{\log \left(\tau_ {\mathrm {m i d}} + 1\right)} & (\tau \leq \tau_ {\mathrm {m i d}}) \\ 1. 0 & (\tau > \tau_ {\mathrm {m i d}}) \end{array} , \right. \right. \tag {14} +$$ + +where $\epsilon = 0.005$ + +The total loss is $\mathcal{L} = \alpha \lambda_{\mathrm{cam}}\mathcal{L}_{\mathrm{cam}} + \beta \lambda_{\mathrm{aux}}\mathcal{L}_{\mathrm{aux}} + \mathcal{L}_{\mathrm{sm}} + \mathcal{L}_{\mathrm{rec}}$ where $\alpha = 0.01$ and $\beta = 1.0$ . + +# 5 LSP + +Our FUMET extracts metric scale from a prior on object heights through the camera height loss $\mathcal{L}_{\mathrm{cam}}$ and the auxiliary rough geometric loss $\mathcal{L}_{\mathrm{aux}}$ . To ensure that FUMET learns the scale regardless of the particular dataset used for training, a universal prior of object heights, independent of the environment, becomes essential. Cars are objects whose dimensions are explicitly and accurately known. Better yet, they are abundantly located on the road, the key semantic region whose metric depth is essential for many downstream tasks. Some methods for other tasks have modeled the size prior as a probability distribution [20, 57]. The gap between its mean and the true expectation for each environment, however, is directly reflected in the uncertainty of the scale in the supervision. Wang et al. [40] also report that there is a gap in car dimensions among datasets. These facts suggest that modeling the object size prior as a single probability distribution is problematic, especially for man-made objects like cars. + +We humans not only possess rough prior knowledge about the vehicle size but can also estimate it more accurately by extracting instance-specific information such as car models from its appearance. Motivated by this intuition, we introduce a learned universal vehicle size prior termed LSP. Given a masked vehicle image, LSP estimates its dimensions. + +For training, we use a dataset consisting of a large number of vehicle images and their information such as 3D dimensions. This dataset can be easily collected with web scraping and does not require manual annotations. To enhance + +the robustness to various input images, in addition to common color jittering and blur augmentation, we perform augmentations to simulate occlusion and truncation. For occlusion augmentation, we mask out random regions with the same shape as a vehicle or with a beam of random width. For truncation augmentation, we translate the vehicle mask in a random direction. We visualize these in the supplementary material. The abundance of training images and diverse augmentations enable LSP to estimate vehicle dimensions robustly across various datasets. + +Although FUMET requires only a car height prior, we construct LSP to estimate the width and length as well as height. This enhances it to grasp an overall car size and improves the prediction accuracy of car heights. + +# 6 Experiments + +# 6.1 FUMET + +We evaluate the accuracy of depth estimators trained on our framework and demonstrate the effectiveness of FUMET as a novel unsupervised metric-scale training method through depth prediction accuracy, its applicability to any network architecture, and its ability to leverage any dataset for training. + +Implementation Details. We implement FUMET with PyTorch [28]. We employ Monodepth2 [15] with ResNet50 encoder pre-trained on ImageNet [30] as the default depth estimator which is one of the most basic architectures in self-supervised MDE. For all experiments unless mentioned otherwise, we train the model for 50 epochs with a batch size of 8 and set $\tau_{\mathrm{mid}}$ to 20. We use Adam optimizer [24] with an initial learning rate of $5.0\times 10^{-5}$ decayed by half every 15 epochs. During training, we perform the following augmentations with $50\%$ chance in random order: horizontal flips, brightness adjustment $(\pm 0.2)$ , saturation adjustment $(\pm 0.2)$ , contrast adjustment $(\pm 0.2)$ , and hue jitter $(\pm 0.1)$ . For quantitative evaluation, we follow [9] and compute the seven standard metrics: AbsRel, SqRel, RMSE, $\mathrm{RMSE}_{\log}$ , $\delta < 1.25$ , $\delta < 1.25^2$ , and $\delta < 1.25^3$ . + +KITTI. Our FUMET is independent of network architectures so we experiment on several models with FUMET as well as Monodepth2: Lite-Mono [52] with ResNet18, HR-Depth [25] with ResNet18, and VADepth [45]. We train these models on the KITTI raw dataset [12], which contains 39,810 images for training and 4,424 for validation, and evaluate it with the KITTI Eigen test split [8] (697 images). For training and evaluation, all images are resized to $640 \times 192$ . + +We show the qualitative results in Fig. 2. The quantitative results shown in Tab. 1 demonstrate that FUMET enables the various architectures to learn metric scale properly. Even when employing the simplest network architecture, i.e., Monodepth2, it outperforms weakly-supervised methods that rely on ground-truth scale labels for training which is not necessary for our method. Similar to VADepth [45], FUMET uses the camera height invariance as supervision though + +Table 1: Quantitative results on the KITTI Eigen test split [8] ( $640 \times 192$ resolution). In the Supervision column, V, G, and CamH represent velocity, gravity, and camera height, respectively. The Scaling column represents whether to perform the median scaling [56]. We denote the best results in bold and the second best ones with an underline for each block. Note that we train VADepth [45] from scratch using ground-truth camera height (1.65m) provided by KITTI, unlike its original paper. FUMET outperforms weakly-supervised methods regardless of the network architectures and without the need for ground truth supervision. In the rescaled results, every model trained with FUMET achieves better results than its original one. + +
MethodSupervisionScalingError (↓)Accuracy (↑)
AbsRelSqRelRMSERMSElogδ<1.25δ<1.252δ<1.253
G2S [3]GPS-0.1090.8604.8550.1980.8650.9540.980
PackNet-SfM [16]V-0.1110.8294.7880.1990.8640.9540.980
Wagstaff et al. [36]CamH-0.1230.9965.2530.2130.8400.9470.978
VADepth [45]CamH-0.1200.9754.9710.2030.8670.9560.979
DynaDepth [53]IMU+V+G-0.1090.7874.7050.1950.8690.9580.981
Ours--0.1080.7854.7360.1950.8710.9580.981
Lite-Mono [52]+Ours--0.1120.7984.6920.1940.8710.9590.981
HR-Depth [25]+Ours--0.1090.7944.7340.1930.8690.9590.982
VADepth [45]+Ours--0.1080.8094.5720.1850.8830.9630.982
G2S [3]GPS✓0.1120.8944.8520.1920.8770.9580.981
PackNet-SfM [16]V✓0.1110.7854.6010.1890.8780.9600.982
Wagstaff et al. [36]CamH✓0.1170.9524.9890.1970.8670.9570.981
VADepth [45]CamH✓0.1100.9774.8720.1870.8890.9620.981
DynaDepth [53]IMU+V+G✓0.1080.7614.6080.1870.8830.9620.982
Ours-✓0.1060.7594.6020.1840.8870.9630.983
Monodepth2 [15]-✓0.1150.9034.8630.1930.8770.9590.981
+ Ours-✓0.1060.7594.6020.1840.8870.9630.983
HR-Depth [25]-✓0.1090.7924.6320.1850.8840.9620.983
+ Ours-✓0.1080.7594.5840.1840.8850.9630.983
Lite-Mono [52]-✓0.1100.8024.6710.1860.8790.9610.982
+ Ours-✓0.1070.7754.5640.1830.8880.9630.983
VADepth [45]-✓0.1040.7744.5520.1810.8920.9650.982
+ Ours-✓0.1020.7604.4730.1780.8970.9660.983
+ +VADepth requires the ground-truth camera height and employs a complex network architecture. FUMET with Monodepth2 (row 6 in Tab. 1) achieves higher accuracy. This indicates that measuring camera height precisely is difficult and using unreliable ground-truth camera height causes low accuracy in the scale of estimated depth maps. + +In the results showing the accuracy of the estimated depth map scaled with ground-truth scale factors via median scaling [56], it is evident that every model trained with FUMET attains superior results to the one trained without FUMET. This shows that our training framework not only enables the models to learn metric scale but also enhances the geometric accuracy. This improvement owes to the camera height loss $\mathcal{L}_{\mathrm{cam}}$ and $\mathcal{L}_{\mathrm{aux}}$ , which makes the model respect road regions and object regions, respectively. + +![](images/1ae246018d255e87deaff77763c04c403f7ba641eb1140f2fafdf86d6dc1d768.jpg) +Fig. 2: Qualitative comparison on KITTI. In error maps, the larger depth errors are represented in red, smaller ones are in blue. The model trained with our FUMET predicts more accurate depth maps compared to the weakly-supervised methods. + +FUMET successfully enables a model to learn metric scale, even though the training dataset of KITTI contains many frames with no or few detected cars. We detail this in the supplementary material. + +Cityscapes. To evaluate the ability of FUMET to leverage any dataset for training, we train and evaluate it on the Cityscapes dataset [6]. We follow [43, 46, 48, 56] and collect 69,731 images for training and evaluate on the 1,525 test images. In this dataset, there are multiple sets of camera intrinsics. Since this strongly affects the metric depth estimation [47], we first align the focal lengths of all images to 587.5 by resizing them, and crop them to $512 \times 192$ resolution. For evaluation, we follow [2, 43] and further crop the estimated image into $416 \times 128$ . In order to align the number of iterations on KITTI, we train for 29 epochs with the same learning rate decreased by half every 8 epochs, and $\tau_{\mathrm{mid}}$ is set to 12. + +Tab. 2 shows the quantitative results and FUMET outperforms weakly-supervised methods by wide margins. These methods assume that highly accurate sensor data is available and fail to train scale-aware MDE models on datasets consisting of unreliable sensor data such as Cityscapes. On the other hand, FUMET does not require any sensor measurement other than RGB video, so it can robustly leverage any driving videos for training. + +Mixed datasets. To demonstrate that FUMET can train a model on a mixed set of datasets to attain strong generalizability, we train with 235,341 images collected from the Argoverse2 [44], Lyft [21], A2D2 [13], and DDAD [16] dataset and evaluate it on the KITTI Eigen test split [8]. Both images for training and + +Table 2: Quantitative results on Cityscapes [6]. We train all the models from scratch with the same supervisions as the Supervision column in Tab. 1. FUMET significantly outperforms weakly-supervised methods. + +
Method|AbsRelSqRelRMSEδ<1.25
G2S [3]4.156276.1658.890.046
PackNet-SfM [16]0.5046.63914.900.029
VADepth [45]0.3637.11511.950.295
FUMET0.1251.2886.3590.858
+ +Table 3: FUMET can be trained on mixed datasets (Argoverse2, Lyft, A2D2, and DDAD) with various camera settings, and attains high generalizability. The results are evaluated on the KITTI Eigen split [8] $(832\times 512$ resolution). + +
Training datasetAbsRelSqRelRMSEδ<1.25
KITTI0.1030.6754.7080.903
Mixed0.1130.9165.0090.883
Mixed+KITTI0.0820.6114.3070.923
+ +evaluation are resized and cropped to $832 \times 512$ resolution. The training images are the sets of video sequences and each sequence is captured with its own unique camera setting. We individually optimize each camera height for each sequence during training. We detail the dataset composition and training settings in the supplementary material. + +As shown in Tab. 3, the model trained with FUMET on the large-scale mixed dataset achieves comparable accuracy to the one trained on KITTI despite the domain gap and attains high generalizability. The model trained on both the mixed dataset and KITTI outperforms the model trained only on KITTI. This suggests that training on various images enhances accuracy which is the strength of fully self-supervised learning methods. + +We also show that FUMET can leverage in-the-wild videos for training by using camera intrinsics recovered with SLAM, e.g., COLMAP [32] in the supplementary material. + +Ablation Studies on KITTI. To study the effect of the proposed components in FUMET, we conduct ablation studies on the training losses, balancing the loss weights, camera height optimization, and outlier filtering. The results are shown in Tab. 4. Thanks to the accuracy of LSP and the robustness of Silhouette Projector, the camera height loss makes a greater contribution to the depth accuracy than the auxiliary rough geometric loss. By introducing the camera height optimization, the model can learn the scale with more accurate and consistent supervision than leveraging the size prior independently for each frame (rows 1, 2, and 6 in Tab. 4), which leads to improved accuracy. Though balancing the loss weights (Eq. (14)) with either $\mathcal{L}_{\mathrm{aux}}$ or $\mathcal{L}_{\mathrm{cam}}$ degrades the results due to the insufficient strength of the supervisory signals for metric scale, it contributes to improving accuracy when both $\mathcal{L}_{\mathrm{aux}}$ and $\mathcal{L}_{\mathrm{cam}}$ are employed. + +Our proposed method can also be used with a pre-trained MDE model in an offline manner. We first train the original Monodepth2 [15] without FUMET, then predict depth maps of training images. With Silhouette Projector and LSP, we obtain per-frame scaled camera height and optimize it across frames with the median operation. We fix this value and use it as the camera height supervision + +Table 4: Each loss contributes to learning metric scale and the others enhance the depth accuracy. Bal: balancing the loss weights (Eq. (14)). Opt: camera height optimization. Flt: outlier filtering. + +
\( \mathcal{L}_{\text{aux}} \)\( \mathcal{L}_{\text{cam}} \)BalOptFltAbsRelSqRelRMSE\( \delta < 1.25 \)
✓---✓0.1250.8875.0550.834
✓-✓-✓0.1490.9505.2190.785
-✓-✓✓0.1160.8754.8740.863
-✓✓✓✓0.2111.3036.1430.585
✓✓-✓✓0.1220.8805.0170.849
✓✓✓-✓0.1150.8224.8110.858
✓✓✓✓-0.1160.8234.9010.860
✓✓✓✓✓0.1080.7854.7360.871
+ +Table 5: Training with a fine-tuned pseudo camera height achieves the highest accuracy. Online: optimizing the camera height while training. Offline: training with a fixed one computed from depth maps predicted with pre-trained Monodepth2. Fine-tune: training with the fixed one computed offline for 20 epochs, then optimizing online. + +
Deriv. methodAbsRelSqRelRMSEδ<1.25
Online0.1080.8084.7400.870
Offline0.1060.7904.7190.873
Fine-tune0.1050.7824.6820.876
+ +for an untrained Monodepth2 on FUMET. We examine the difference between the proposed camera height optimized online and the fixed one derived offline when we use them as the pseudo-supervision. Note that we also fix $\lambda_{\mathrm{cam}}$ because the pseudo-label can be considered sufficiently accurate from the early stage of training. We also run another experiment in which we fix camera height supervision derived offline for the first 20 epochs, which is the same number of training epochs as the original Monodepth2, then unfit the camera height pseudo-label to fine-tune it via online optimization. As shown in Tab. 5, the model supervised with the fixed camera height computed offline achieves slightly better results than the one supervised with the camera height optimized online. This result indicates that our strategy of deriving scaled camera height is highly accurate and fixing it stabilizes the training. The model trained with the fine-tuned camera height achieves the highest accuracy, which shows that our FUMET can attain better accuracy with additional pre-training costs. + +# 6.2 LSP + +We provide implementation details of LSP in the supplementary material. + +Evaluation on KITTI and Cityscapes. We evaluate the generalizability of our LSP with the dimensions of the annotated 3D bounding box on both the KITTI [12] and the Cityscapes [6] datasets in the Car class. We compare the results of LSP to a fixed car height prior (1.59m, following [20]) and MonoCIinIS [19], a 3D object detection model that outputs object dimensions from an image and an instance mask. We train MonoCIinIS from scratch with the KITTI dataset. As an evaluation metric, we compute the absolute relative error for each dimension. When we estimate car dimensions with LSP and MonoCIinIS, we use a ground-truth instance mask for making a car image mask and as input, respectively. For all the experiments in LSP, input vehicle masks are cropped and resized to $300 \times 300$ resolution. On the KITTI dataset, we extract 3,769 samples for evaluation [4]. On the Cityscapes dataset, we evaluate on the validation split, + +Table 6: Generalizability on KITTI and Cityscapes. We evaluate LSP on Absolute Relative Error for each dimension (height/width/length), comparing with the fixed car height prior (1.59 m, following [20]) and MonoCInIS [19] trained on KITTI. LSP robustly surpasses other methods in most cases and can predict accurately regardless of the datasets. + +
MethodEasyModerateHardAll
KITTIFixed Height [20]0.065 / - / -0.064 / - / -0.065 / - / -0.065 / - / -
MonoClIns [19]0.069 / 0.065 / 0.0930.112 / 0.106 / 0.1330.268 / 0.263 / 0.2690.159 / 0.155 / 0.177
LSP (only height)0.057 / - / -0.089 / - / -0.106 / - / -0.087 / - / -
LSP0.046 / 0.136 / 0.1300.060 / 0.144 / 0.1390.073 / 0.155 / 0.1500.062 / 0.147 / 0.145
CityscapesFixed Height [20]0.093 / - / -0.093 / - / -0.076 / - / -0.092 / - / -
MonoClIns [19]0.578 / 0.599 / 0.5840.759 / 0.766 / 0.7670.843 / 0.852 / 0.8520.851 / 0.859 / 0.856
LSP (only height)0.063 / - / -0.075 / - / -0.103 / - / -0.096 / - / -
LSP0.056 / 0.069 / 0.0670.068 / 0.068 / 0.0780.069 / 0.059 / 0.0750.080 / 0.064 / 0.077
+ +which has 500 samples. On the KITTI dataset, each object is separated into 3 levels based on detection difficulty: Easy, Moderate, and Hard. Considering the difference in focal lengths between KITTI and Cityscapes, we also separate objects on Cityscapes into the same levels as on KITTI. + +As shown in Tab. 6, LSP outperforms the other two methods in most cases. Particularly, the accuracy of MonoCInIS drops on Cityscapes since it overfits to the training dataset, i.e., KITTI and cannot address the domain gap. In contrast, thanks to the large training data and various augmentations, our LSP obtains the highest generalizability and robustly estimates dimensions on any dataset. Learning not only car height but also width and length also improves the prediction accuracy as this facilitates estimation of the overall car sizes. + +# 7 Conclusion + +In this work, we proposed FUMET, a novel training framework for any MDE model to learn metric scale. FUMET aggregates the scale information obtained from a learned car size prior into the camera height and uses it as pseudosupervision formulated as optimization across frames and epochs by leveraging its invariance. We showed that FUMET enables monocular depth estimators to embody metric scale regardless of the underlying network architectures and achieves state-of-the-art absolute-scale depth estimation accuracy on both the KITTI and the Cityscapes datasets, and demonstrated that it can leverage any dataset for training. We also proposed LSP, which can be easily trained on a large-scale dataset without manual annotations and robustly estimates vehicle dimensions from their appearances. We showed that the learned size prior is superior to a fixed one and outperforms a similar 3D object detection model. FUMET can be leveraged as a drop-in training scheme for any monocular depth estimation method. We hope it will be a versatile tool for road scene understanding research. + +# Acknowledgements + +This work was in part supported by JSPS 20H05951 and 21H04893, JST JP-MJCR20G7 and JPMJAP2305, and RIKEN GRP. + +# References + +1. Bhat, S.F., Alhashim, I., Wonka, P.: AdaBins: Depth Estimation Using Adaptive Bins. In: CVPR. pp. 4009-4018 (2021) +2. Casser, V., Pirk, S., Mahjourian, R., Angelova, A.: Unsupervised Monocular Depth and Ego-Motion Learning with Structure and Semantics. In: CVPRW. pp. 381-388 (2019) +3. Chawla, H., Varma, A., Arani, E., Zonooz, B.: Multimodal Scale Consistency and Awareness for Monocular Self-Supervised Depth Estimation. In: ICRA. pp. 5140-5146 (2021) +4. Chen, X., Kundu, K., Zhu, Y., Berneshawi, A.G., Ma, H., Fidler, S., Urtasun, R.: 3D Object Proposals for Accurate Object Class Detection. Adv. Neural Inform. Process. Syst. 28 (2015) +5. Chen, X., Li, T.H., Zhang, R., Li, G.: Frequency-Aware Self-Supervised Monocular Depth Estimation. In: WACV. pp. 5808-5817 (2023) +6. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The Cityscapes Dataset for Semantic Urban Scene Understanding. In: CVPR (2016) +7. Ding, M., Huo, Y., Yi, H., Wang, Z., Shi, J., Lu, Z., Luo, P.: Learning Depth-Guided Convolutions for Monocular 3D Object Detection. In: CVPR. pp. 11669-11678 (2020) +8. Eigen, D., Fergus, R.: Predicting Depth, Surface Normals and Semantic Labels with a Common Multi-Scale Convolutional Architecture. In: ICCV. pp. 3213-3223 (2015) +9. Eigen, D., Puhrsch, C., Fergus, R.: Depth Map Prediction from a Single Image Using a Multi-Scale Deep Network. In: Adv. Neural Inform. Process. Syst. vol. 2, pp. 2366-2374 (2014) +0. Frost, D., Prisacariu, V., Murray, D.: Recovering Stable Scale in Monocular SLAM Using Object-Supplemented Bundle Adjustment. IEEE Transactions on Robotics 34(3), 736-747 (2018). https://doi.org/10.1109/TR0.2018.2820722 +1. Garg, R., Bg, V.K., Carneiro, G., Reid, I.: Unsupervised CNN for Single View Depth Estimation: Geometry to the Rescue. In: ECCV. pp. 740-756 (2016) +2. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision Meets Robotics: The KITTI Dataset. International Journal of Robotics Research (IJRR) 32(11), 1231--1237 (2013). https://doi.org/10.1177/0278364913491297 +3. Geyer, J., Kassahun, Y., Mahmudi, M., Ricou, X., Durgesh, R., Chung, A.S., Hauswald, L., Pham, V.H., Mühlegg, M., Dorn, S., Fernandez, T., Janicke, M., Mirashi, S., Savani, C., Sturm, M., Vorobiov, O., Oelker, M., Garreis, S., Schuberth, P.: A2D2: Audi Autonomous Driving Dataset. CoRR abs/2004.06320 (2020) +4. Godard, C., Mac Aodha, O., Brostow, G.J.: Unsupervised Monocular Depth Estimation with Left-Right Consistency. In: CVPR. pp. 6602-6611 (2017) +5. Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging Into Self-Supervised Monocular Depth Estimation. In: ICCV. pp. 3827-3837 (2019) + +16. Guizilini, V., Ambrus, R., Pillai, S., Raventos, A., Gaidon, A.: 3D Packing for Self-Supervised Monocular Depth Estimation. In: CVPR. pp. 2482-2491 (2020) +17. Hartley, R., Zisserman, A.: Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edn. (2004) +18. He, M., Hui, L., Bian, Y., Ren, J., Xie, J., Yang, J.: RA-Depth: Resolution Adaptive Self-supervised Monocular Depth Estimation. In: ECCV. pp. 565–581 (2022) +19. Heylen, J., De Wolf, M., Dawagne, B., Proesmans, M., Van Gool, L., Abbeloos, W., Abdelkawy, H., Reino, D.O.: MonoCInIS: Camera Independent Monocular 3D Object Detection Using Instance Segmentation. In: ICCVW. pp. 923-934 (2021) +20. Hoiem, D., Efros, A.A., Hebert, M.: Putting Objects in Perspective. IJCV 80, 3-15 (2008). https://doi.org/10.1007/s11263-008-0137-5 +21. Houston, J., Zuidhof, G., Bergamini, L., Ye, Y., Chen, L., Jain, A., Omari, S., Iglovikov, V., Ondruska, P.: One Thousand and One Hours: Self-Driving Motion Prediction Dataset. In: Conference on Robot Learning. pp. 409-418 (2021) +22. Jain, J., Li, J., Chiu, M.T., Hassani, A., Orlov, N., Shi, H.: OneFormer: One Transformer to Rule Universal Image Segmentation. In: CVPR. pp. 2989-2998 (2023) +23. Kim, D., Ga, W., Ahn, P., Joo, D., Chun, S., Kim, J.: Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth. CoRR abs/2201.07436 (2022) +24. Kingma, D.P., Ba, J.: Adam: A Method for Stochastic Optimization. CoRR abs/1412.6980 (2014) +25. Lyu, X., Liu, L., Wang, M., Kong, X., Liu, L., Liu, Y., Chen, X., Yuan, Y.: HR-Depth: High Resolution Self-Supervised Monocular Depth Estimation. In: AAAI. vol. 35, pp. 2294-2301 (2021) +26. Milli, E., Erkent, O., Yilmaz, A.E.: Multi-Modal Multi-Task (3MT) Road Segmentation. IEEE Robotics and Automation Letters 8(9), 5408-5415 (2023). https://doi.org/10.1109/LRA.2023.3295254 +27. Ning, C., Gan, H.: Trap Attention: Monocular Depth Estimation with Manual Traps. In: CVPR. pp. 5033-5043 (2023) +28. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: PyTorch: An Imperative Style, High-Performance Deep Learning Library. In: Adv. Neural Inform. Process. Syst. pp. 8026--8037 (2019) +29. Piccinelli, L., Sakaridis, C., Yu, F.: iDisc: Internal Discretization for Monocular Depth Estimation. In: CVPR. pp. 21477-21487 (2023) +30. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet: A Large Scale Visual Recognition Challenge. IJCV 115, 211-252 (2015) +31. Sadat, A., Casas, S., Ren, M., Wu, X., Dhawan, P., Urtasun, R.: Perceive, Predict, and Plan: Safe Motion Planning Through Interpretable Semantic Representations. In: ECCV. pp. 414-430 (2020) +32. Schonberger, J.L., Frahm, J.M.: Structure-From-Motion Revisited. In: CVPR. pp. 4104-4113 (2016) +33. Simonelli, A., Bulò, S.R., Porzi, L., Kontschieder, P., Ricci, E.: Are We Missing Confidence in Pseudo-LiDAR Methods for Monocular 3D Object Detection? In: ICCV. pp. 3225-3233 (2021) +34. Spencer, J., Russell, C., Hadfield, S., Bowden, R.: Kick Back & Relax: Learning to Reconstruct the World by Watching SlowTV (2023) + +35. Sucar, E., Hayet, J.B.: Bayesian Scale Estimation for Monocular SLAM Based on Generic Object Detection for Correcting Scale Drift. In: ICRA. pp. 5152-5158 (2018) +36. Wagstaff, B., Kelly, J.: Self-Supervised Scale Recovery for Monocular Depth and Egomotion Estimation. In: IROS. pp. 2620-2627 (2021) +37. Wang, C., Buenaposada, J.M., Zhu, R., Lucey, S.: Learning Depth from Monocular Videos Using Direct Methods. In: CVPR. pp. 2022-2030 (2018) +38. Wang, H., Cai, P., Fan, R., Sun, Y., Liu, M.: End-to-End Interactive Prediction and Planning with Optical Flow Distillation for Autonomous Driving. In: CVPRW. pp. 2229–2238 (2021) +39. Wang, Y., Chao, W.L., Garg, D., Hariharan, B., Campbell, M., Weinberger, K.Q.: Pseudo-LiDAR From Visual Depth Estimation: Bridging the Gap in 3D Object Detection for Autonomous Driving. In: CVPR. pp. 8437-8445 (2019) +40. Wang, Y., Chen, X., You, Y., Li, L.E., Hariharan, B., Campbell, M., Weinberger, K.Q., Chao, W.L.: Train in Germany, Test in the USA: Making 3D Object Detectors Generalize. In: CVPR. pp. 11710-11720 (2020) +41. Wang, Y., Yang, H., Cai, J., Wang, G., Wang, J., Huang, Y.: Unsupervised Learning of Depth and Pose Based on Monocular Camera and Inertial Measurement Unit (IMU). In: ICRA. pp. 10010-10017 (2023) +42. Wang, Z., Bovik, A., Sheikh, H., Simoncelli, E.: Image Quality Assessment: From Error Visibility to Structural Similarity. IEEE Transactions on Image Processing 13(4), 600-612 (2004). https://doi.org/10.1109/TIP.2003.819861 +43. Watson, J., Mac Aodha, O., Prisacariu, V., Brostow, G., Firman, M.: The Temporal Opportunist: Self-Supervised Multi-Frame Monocular Depth. In: CVPR. pp. 1164–1174 (2021) +44. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., Ramanan, D., Carr, P., Hays, J.: Argoverse 2: Next Generation Datasets for Self-driving Perception and Forecasting. In: NeurIPS Datasets and Benchmarks (2021) +45. Xiang, J., Wang, Y., An, L., Liu, H., Wang, Z., Liu, J.: Visual Attention-Based Self-Supervised Absolute Depth Estimation Using Geometric Priors in Autonomous Driving. IEEE Robotics and Automation Letters 7(4), 11998-12005 (2022). https://doi.org/10.1109/LRA.2022.3210298 +46. Yang, Z., Wang, P., Wang, Y., Xu, W., Nevatia, R.: LEGO: Learning Edge with Geometry All at Once by Watching Videos. In: CVPR. pp. 225-234 (2018) +47. Yin, W., Zhang, C., Chen, H., Cai, Z., Yu, G., Wang, K., Chen, X., Shen, C.: Metric3D: Towards Zero-Shot Metric 3D Prediction from a Single Image. In: ICCV. pp. 9043-9053 (2023) +48. Yin, Z., Shi, J.: GeoNet: Unsupervised Learning of Dense Depth, Optical Flow and Camera Pose. In: CVPR. pp. 1983-1992 (2018) +49. You, Y., Wang, Y., Chao, W.L., Garg, D., Pleiss, G., Hariharan, B., Campbell, M., Weinberger, K.Q.: Pseudo-LiDAR++: Accurate Depth for 3D Object Detection in Autonomous Driving. In: ICLR (2020) +50. Yuan, W., Gu, X., Dai, Z., Zhu, S., Tan, P.: Neural Window Fully-Connected CRFs for Monocular Depth Estimation. In: CVPR. pp. 3916-3925 (2022) +51. Zeng, W., Luo, W., Suo, S., Sadat, A., Yang, B., Casas, S., Urtasun, R.: End-To-End Interpretable Neural Motion Planner. In: CVPR. pp. 8652-8661 (2019) +52. Zhang, N., Nex, F., Vosselman, G., Kerle, N.: Lite-Mono: A Lightweight CNN and Transformer Architecture for Self-Supervised Monocular Depth Estimation. In: CVPR. pp. 18537-18546 (2023) + +53. Zhang, S., Zhang, J., Tao, D.: Towards Scale-Aware, Robust, and Generalizable Unsupervised Monocular Depth Estimation by Integrating IMU Motion Dynamics. In: ECCV. pp. 143-160 (2022) +54. Zhang, S., Li, X., Liu, Y., Fu, H.: Scale-aware Insertion of Virtual Objects in Monocular Videos. In: ISMAR. pp. 36-44 (2020) +55. Zhou, H., Greenwood, D., Taylor, S.: Self-Supervised Monocular Depth Estimation with Internal Feature Fusion. In: BMVC (2021) +56. Zhou, T., Brown, M., Snavely, N., Lowe, D.G.: Unsupervised Learning of Depth and Ego-Motion from Video. In: CVPR. pp. 6612-6619 (2017) +57. Zhu, R., Yang, X., Hold-Geoffroy, Y., Perazzi, F., Eisenmann, J., Sunkavalli, K., Chandraker, M.: Single View Metrology in the Wild. In: ECCV. pp. 316-333 (2020) \ No newline at end of file diff --git a/2024/Camera Height Doesn't Change_ Unsupervised Training for Metric Monocular Road-Scene Depth Estimation/images.zip b/2024/Camera Height Doesn't Change_ Unsupervised Training for Metric Monocular Road-Scene Depth Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5c637e5131d1cd7cc46388b6baefb55ca1e28d1c --- /dev/null +++ b/2024/Camera Height Doesn't Change_ Unsupervised Training for Metric Monocular Road-Scene Depth Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:397016d53b060db9746932042796825270f73864dccdfc5e1ae0a3b1c8bd7cdc +size 464150 diff --git a/2024/Camera Height Doesn't Change_ Unsupervised Training for Metric Monocular Road-Scene Depth Estimation/layout.json b/2024/Camera Height Doesn't Change_ Unsupervised Training for Metric Monocular Road-Scene Depth Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..94bdf59edc4ddd33f1c1382945205d5eef219ad5 --- /dev/null +++ b/2024/Camera Height Doesn't Change_ Unsupervised Training for Metric Monocular Road-Scene Depth Estimation/layout.json @@ -0,0 +1,9738 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 111, + 465, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 111, + 465, + 166 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 465, + 166 + ], + "type": "text", + "content": "Camera Height Doesn't Change: Unsupervised Training for Metric Monocular Road-Scene Depth Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 227, + 186, + 388, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 186, + 388, + 198 + ], + "spans": [ + { + "bbox": [ + 227, + 186, + 388, + 198 + ], + "type": "text", + "content": "Genki Kinoshita and Ko Nishino" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 168, + 208, + 445, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 208, + 445, + 242 + ], + "spans": [ + { + "bbox": [ + 168, + 208, + 445, + 242 + ], + "type": "text", + "content": "Graduate School of Informatics, Kyoto University, Kyoto, Japan \ngkinoshita@vision.ist.i.kyoto-u.ac.jp, kon@i.kyoto-u.ac.jp \nhttps://vision.ist.i.kyoto-u.ac.jp/" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 270, + 455, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 270, + 455, + 479 + ], + "spans": [ + { + "bbox": [ + 160, + 270, + 455, + 479 + ], + "type": "text", + "content": "Abstract. In this paper, we introduce a novel training method for making any monocular depth network learn absolute scale and estimate metric road-scene depth just from regular training data, i.e., driving videos. We refer to this training framework as FUMET. The key idea is to leverage cars found on the road as sources of scale supervision and to incorporate them in network training robustly. FUMET detects and estimates the sizes of cars in a frame and aggregates scale information extracted from them into an estimate of the camera height whose consistency across the entire video sequence is enforced as scale supervision. This realizes robust unsupervised training of any, otherwise scale-oblivious, monocular depth network so that they become not only scale-aware but also metric-accurate without the need for auxiliary sensors and extra supervision. Extensive experiments on the KITTI and the Cityscapes datasets show the effectiveness of FUMET, which achieves state-of-the-art accuracy. We also show that FUMET enables training on mixed datasets of different camera heights, which leads to larger-scale training and better generalization. Metric depth reconstruction is essential in any road-scene visual modeling, and FUMET democratizes its deployment by establishing the means to convert any model into a metric depth estimator." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 490, + 453, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 490, + 453, + 512 + ], + "spans": [ + { + "bbox": [ + 160, + 490, + 453, + 512 + ], + "type": "text", + "content": "Keywords: Unsupervised Monocular Depth Estimation " + }, + { + "bbox": [ + 160, + 490, + 453, + 512 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 490, + 453, + 512 + ], + "type": "text", + "content": " Metric Scale " + }, + { + "bbox": [ + 160, + 490, + 453, + 512 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 490, + 453, + 512 + ], + "type": "text", + "content": " Camera Height Invariance " + }, + { + "bbox": [ + 160, + 490, + 453, + 512 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 490, + 453, + 512 + ], + "type": "text", + "content": " Object Size Prior" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 533, + 230, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 230, + 545 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 230, + 545 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "text", + "content": "Monocular depth estimation is essential for autonomous driving and advanced driver assistance systems (ADAS). It underlies many of their key perceptual tasks including 3D object detection [7,33,39,49], motion planning [31,38,51], and road segmentation [26]. Although supervised methods have achieved impressive accuracy, they suffer from costly ground-truth data collection. This limits their generalizability [34] and also makes continuous learning difficult." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "Many recent methods [5, 15, 18, 25, 52] leverage self-supervision to scale up training. Most of these methods, however, suffer from scale ambiguity which renders them difficult to use off-the-shelf. To resolve the scale ambiguity and" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "recover in metric scale, the actual scene scale needs to be supervised in some form. Past works have leveraged velocity [16], GPS [3], IMU [41, 53], or camera height [36, 45] for this. These methods require auxiliary sensors to measure these extra quantities in addition to the monocular RGB camera, which significantly limits the ability to scale the size of training data. Arbitrary in-the-wild driving videos cannot be used as they lack such measurements. These self-supervised methods often still require weak scale supervision and cannot be made fully self-supervised as the scale changes for each sequence captured with a different camera." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 227, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 227, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 482, + 370 + ], + "type": "text", + "content": "Road scenes actually have plenty of untapped sources of metric scale. Cars found on the road are one of them as they are rigid objects whose actual sizes do not change and are unique to each make and model. The road is also exactly the region where we need accurate metric depth the most. If we can leverage the sizes of cars to bridge the 3D to 2D projection, we may train a mono-depth network to become metric-accurate. Simply employing a prior on the car size, however, would be too brittle since the metric supervision will be as ambiguous as the accuracy of that prior. The perspective projection and the variations in car makes, years, and colors, lead to a large variance in the size estimates, which will directly affect the mono-depth metric accuracy. In addition, there are no reliable object size priors except for a few introduced as probability densities for other tasks [20,54,57] whose parameters are not shared." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 373, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 517 + ], + "type": "text", + "content": "How can we train an arbitrary mono-depth estimator to become metric-accurate by robustly leveraging prior knowledge about the 3D sizes of cars found in driving videos? We derive fully unsupervised metric depth trainer (FUMET) to answer this question. Our key idea is to aggregate car size cues in the frames and the sequence into a single physical measure, namely the camera height. Regardless of the scene, the camera capturing it is the same for the same sequence. This means that, by transforming the car size cues found in different frames into a camera height estimate, the very fact that it should not change can be used as a supervision of invariance to achieve metric depth reconstruction. We can formulate this as an optimization of the camera height across frames and training epochs. This leads to gradual metric-aware learning that is stable and accurate, which also makes the depth estimates consistent across the sequence." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 519, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 519, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 482, + 639 + ], + "type": "text", + "content": "We introduce a learned size prior (LSP) to estimate the size of each car found in a frame. This prior gives us the vehicle dimensions from its appearance. It is trained with various augmentations on a large-scale dataset without the need for any manual annotation. By comparing these dimension estimates with those computed from the depth estimates, we obtain a per-frame scale factor. The multiplication of this scale estimate with the camera height estimate derived from the estimated depth becomes the camera height estimate of the frame. These estimated camera heights across all frames in a sequence are then consolidated by using their median value which is then used as scale supervision with its weighted moving average at the end of each training epoch." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "content": "We evaluate the effectiveness of FUMET with extensive experiments using the KITTI and the Cityscapes datasets and compare its accuracy with a number" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "G.Kinoshita et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "of weakly-supervised methods. The results clearly show that FUMET succeeds in making monocular depth estimation methods learn metric scales and achieves the state-of-the-art accuracy without any direct depth and scale supervision. We also demonstrate its capability to train a model on mixed datasets collected with different cameras, i.e., camera heights. This enables larger-scale training and leads to higher generalization. We believe FUMET can serve as a fundamental building block for monocular depth estimation in the wild." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 217, + 234, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 217, + 234, + 229 + ], + "spans": [ + { + "bbox": [ + 131, + 217, + 234, + 229 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 241, + 242, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 241, + 242, + 253 + ], + "spans": [ + { + "bbox": [ + 131, + 241, + 242, + 253 + ], + "type": "text", + "content": "2.1 Self-Supervision" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 260, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 403 + ], + "type": "text", + "content": "Although, as in other tasks, supervised methods achieve the highest accuracy in monocular depth estimation (MDE) [1, 23, 27, 29, 50], they obviously require ground-truth data usually obtained with a LiDAR sensor. Self-supervised MDE methods remove the need for such hardware costs for training. Garg et al. [11] proposed the first approach to self-supervised MDE by predicting per-pixel disparities between a stereo pair. Godard et al. [14] extend this idea by introducing left-right consistency and greatly improve the accuracy. These methods require calibrated stereo data for training which limits their applicability. Zhou et al. [56] introduced a pioneering method for training solely from monocular videos by incorporating a network that estimates the relative camera poses between successive frames. Based on this idea, many MDE methods that can be trained solely with monocular videos have been proposed [5, 15, 18, 25, 52, 55]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 403, + 481, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 481, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 481, + 452 + ], + "type": "text", + "content": "Despite the popularity and improvements of self-supervised MDE, directly deploying those models in autonomous driving or ADAS systems is unrealistic as they only produce relative depth up to an unknown scale. Our FUMET overcomes this issue without the need for any external supervision." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 468, + 252, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 468, + 252, + 480 + ], + "spans": [ + { + "bbox": [ + 131, + 468, + 252, + 480 + ], + "type": "text", + "content": "2.2 Weak Supervision" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 486, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 606 + ], + "type": "text", + "content": "Several works have tackled scale-aware MDE by adopting weak scale supervision through velocity [16], GPS [3], IMU [41], or camera height [36,45] measurements. Zhang et al. [53] use the combination of IMU, velocity, and gravity to achieve the highest accuracy among these methods. These methods fundamentally rely on such auxiliary measurements in addition to regular RGB images. This makes it impossible to take advantage of the many road-scene videos on the Internet (e.g., YouTube) or casually installed dashcam videos for training. In contrast, FUMET relies on a universal car size prior which is independent of the data collection environments and does not require any such weak supervision; it can be trained solely from monocular video sequences." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 622, + 254, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 622, + 254, + 634 + ], + "spans": [ + { + "bbox": [ + 131, + 622, + 254, + 634 + ], + "type": "text", + "content": "2.3 Object Size Priors" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Several works for other tasks leverage priors on object sizes including single view metrology [57], object insertion [20,54], and visual odometry [10,35]. These" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 139, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 139, + 91, + 447, + 102 + ], + "type": "text", + "content": "Unsupervised Training for Metric Monocular Road-Scene Depth Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "methods model the priors as normal distributions or Gaussian mixture models, and the parameters of these distributions are predetermined. When there is a gap between those values and the true ones, which is usually the case in the real world, the difference directly surfaces in erroneous scale estimates." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 165, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 165, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 165, + 482, + 224 + ], + "type": "text", + "content": "Instead of these fixed priors, we introduce a learning-based size prior which we refer to as LSP. This estimates vehicle dimensions based on its appearance, thus can avoid the aforementioned problem. Furthermore, it is trained on a large-scale dataset with various augmentations, so that it can robustly estimate vehicle dimensions from any image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 246, + 234, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 234, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 234, + 258 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 274, + 480, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 480, + 369 + ], + "type": "text", + "content": "Our goal is to train a metric-accurate MDE model " + }, + { + "bbox": [ + 130, + 274, + 480, + 369 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{depth}}" + }, + { + "bbox": [ + 130, + 274, + 480, + 369 + ], + "type": "text", + "content": " with only self-supervision. This self-supervision is comprised of two parts: one designed for learning metric scale (Sec. 4) and the other dedicated to learning depth geometry. For the latter, we follow Zhou et al. [56] and minimize the reconstruction loss between the current frame " + }, + { + "bbox": [ + 130, + 274, + 480, + 369 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 130, + 274, + 480, + 369 + ], + "type": "text", + "content": " and the ones synthesized from temporally adjacent frames " + }, + { + "bbox": [ + 130, + 274, + 480, + 369 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 130, + 274, + 480, + 369 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 274, + 480, + 369 + ], + "type": "inline_equation", + "content": "(s = t \\pm 1)" + }, + { + "bbox": [ + 130, + 274, + 480, + 369 + ], + "type": "text", + "content": ". We employ a pose network " + }, + { + "bbox": [ + 130, + 274, + 480, + 369 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{pose}}" + }, + { + "bbox": [ + 130, + 274, + 480, + 369 + ], + "type": "text", + "content": " in addition to " + }, + { + "bbox": [ + 130, + 274, + 480, + 369 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{depth}}" + }, + { + "bbox": [ + 130, + 274, + 480, + 369 + ], + "type": "text", + "content": " to estimate the relative camera pose " + }, + { + "bbox": [ + 130, + 274, + 480, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{t \\to s} \\in \\mathrm{SE}(3)" + }, + { + "bbox": [ + 130, + 274, + 480, + 369 + ], + "type": "text", + "content": " and the current dense depth map " + }, + { + "bbox": [ + 130, + 274, + 480, + 369 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 130, + 274, + 480, + 369 + ], + "type": "text", + "content": ", respectively," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 225, + 371, + 481, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 371, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 225, + 371, + 481, + 384 + ], + "type": "interline_equation", + "content": "D _ {t} = \\theta_ {\\text {d e p t h}} \\left(I _ {t}\\right), \\mathbf {T} _ {t \\rightarrow s} = \\theta_ {\\text {p o s e}} \\left(I _ {t}, I _ {s}\\right). \\tag {1}", + "image_path": "5769ae352e6eacb7a8ea4a674db6a1f6cc74929198ea6f0b328cf5493b425bf2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 390, + 386, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 390, + 386, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 390, + 386, + 403 + ], + "type": "text", + "content": "We then synthesize the image from the current viewpoint" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 238, + 413, + 480, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 413, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 238, + 413, + 480, + 426 + ], + "type": "interline_equation", + "content": "I _ {s \\rightarrow t} = I _ {s} \\left\\langle \\operatorname {p r o j} \\left(D _ {t}, \\mathbf {T} _ {t \\rightarrow s}, \\mathbf {K}\\right)\\right\\rangle , \\tag {2}", + "image_path": "89b93f156afde5d1a5045a7e2710ff5c36b0f78a3f5373d8c747f606c1ed6ac7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 436, + 482, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 508 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 436, + 482, + 508 + ], + "type": "inline_equation", + "content": "\\langle \\cdot \\rangle" + }, + { + "bbox": [ + 130, + 436, + 482, + 508 + ], + "type": "text", + "content": " is a bilinear sampling operator, " + }, + { + "bbox": [ + 130, + 436, + 482, + 508 + ], + "type": "inline_equation", + "content": "proj(\\cdot)" + }, + { + "bbox": [ + 130, + 436, + 482, + 508 + ], + "type": "text", + "content": " outputs the 2D coordinates of " + }, + { + "bbox": [ + 130, + 436, + 482, + 508 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 130, + 436, + 482, + 508 + ], + "type": "text", + "content": " when projected into the viewpoint of " + }, + { + "bbox": [ + 130, + 436, + 482, + 508 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 130, + 436, + 482, + 508 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 436, + 482, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 436, + 482, + 508 + ], + "type": "text", + "content": " is the camera intrinsic. Following Godard et al. [15], we use an affine combination of SSIM [42] and L1 loss as the photometric error " + }, + { + "bbox": [ + 130, + 436, + 482, + 508 + ], + "type": "inline_equation", + "content": "pe" + }, + { + "bbox": [ + 130, + 436, + 482, + 508 + ], + "type": "text", + "content": " and define the reconstruction loss " + }, + { + "bbox": [ + 130, + 436, + 482, + 508 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 130, + 436, + 482, + 508 + ], + "type": "text", + "content": " as the photometric error between " + }, + { + "bbox": [ + 130, + 436, + 482, + 508 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 130, + 436, + 482, + 508 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 436, + 482, + 508 + ], + "type": "inline_equation", + "content": "I_{s\\rightarrow t}" + }, + { + "bbox": [ + 130, + 436, + 482, + 508 + ], + "type": "text", + "content": " for the pixels with minimum error across source frames" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 181, + 520, + 480, + 535 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 520, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 181, + 520, + 480, + 535 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c}} = \\min _ {s} p e \\left(I _ {t}, I _ {s \\rightarrow t}\\right), \\tag {3}", + "image_path": "a02d6330e6083675dfe237f39e71f349e96935e5c062d27dd632dff68387aed6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 180, + 539, + 480, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 539, + 480, + 563 + ], + "spans": [ + { + "bbox": [ + 180, + 539, + 480, + 563 + ], + "type": "interline_equation", + "content": "p e \\left(I _ {t}, I _ {s \\rightarrow t}\\right) = \\frac {\\lambda}{2} \\left(1 - \\operatorname {S S I M} \\left(I _ {t}, I _ {s \\rightarrow t}\\right)\\right) + (1 - \\lambda) \\left| I _ {t} - I _ {s \\rightarrow t} \\right|, \\tag {4}", + "image_path": "4e9d9024006e6b3a18a4588d1144ca421292e24920a0d8a10095b829a5232d65.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 570, + 430, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 570, + 430, + 583 + ], + "spans": [ + { + "bbox": [ + 132, + 570, + 430, + 583 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 570, + 430, + 583 + ], + "type": "inline_equation", + "content": "\\lambda = 0.85" + }, + { + "bbox": [ + 132, + 570, + 430, + 583 + ], + "type": "text", + "content": ". We also employ an edge-aware smoothness loss [14]" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 226, + 592, + 480, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 592, + 480, + 607 + ], + "spans": [ + { + "bbox": [ + 226, + 592, + 480, + 607 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s m}} = \\left| \\partial_ {x} d _ {t} ^ {*} \\right| e ^ {- \\left| \\partial_ {x} I _ {t} \\right|} + \\left| \\partial_ {y} d _ {t} ^ {*} \\right| e ^ {- \\left| \\partial_ {y} I _ {t} \\right|}, \\tag {5}", + "image_path": "246f8fc9b1e94b5732c65472bc353041641a9b43294dc545d33599aa52ab5c2c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "d_{t}^{*} = d_{t} / \\bar{d}_{t}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " is the mean-normalized inverse depth [37] to prevent the estimated depth from shrinking. Furthermore, when computing " + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 130, + 616, + 482, + 666 + ], + "type": "text", + "content": " we use automasking [15] to mask out stationary pixels which generate incorrect supervisory signals." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "G.Kinoshita et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 482, + 205 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 205 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 205 + ], + "type": "image", + "image_path": "0f6dd62890ff3d85b780da962060a8c8ac07d2a640ae93d3bee87fc2121218aa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 211, + 482, + 300 + ], + "lines": [ + { + "bbox": [ + 130, + 211, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 482, + 300 + ], + "type": "text", + "content": "Fig. 1: Overview of FUMET. At each training step " + }, + { + "bbox": [ + 130, + 211, + 482, + 300 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 211, + 482, + 300 + ], + "type": "text", + "content": ", an unscaled camera height " + }, + { + "bbox": [ + 130, + 211, + 482, + 300 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{cam}}^{n,\\tau}" + }, + { + "bbox": [ + 130, + 211, + 482, + 300 + ], + "type": "text", + "content": " is computed differentiably from the estimated depth. The previous epoch " + }, + { + "bbox": [ + 130, + 211, + 482, + 300 + ], + "type": "inline_equation", + "content": "\\tau - 1" + }, + { + "bbox": [ + 130, + 211, + 482, + 300 + ], + "type": "text", + "content": " provides supervision with a scaled camera height " + }, + { + "bbox": [ + 130, + 211, + 482, + 300 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{cam}}^{*\\tau - 1}" + }, + { + "bbox": [ + 130, + 211, + 482, + 300 + ], + "type": "text", + "content": ". To obtain this scaled camera height supervision, Silhouette Projector first computes the object silhouette heights " + }, + { + "bbox": [ + 130, + 211, + 482, + 300 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{obj}}'" + }, + { + "bbox": [ + 130, + 211, + 482, + 300 + ], + "type": "text", + "content": " from the depth map. By comparing " + }, + { + "bbox": [ + 130, + 211, + 482, + 300 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{obj}}'" + }, + { + "bbox": [ + 130, + 211, + 482, + 300 + ], + "type": "text", + "content": " and the estimated one " + }, + { + "bbox": [ + 130, + 211, + 482, + 300 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{obj}}" + }, + { + "bbox": [ + 130, + 211, + 482, + 300 + ], + "type": "text", + "content": " with LSP, per-frame scale factor " + }, + { + "bbox": [ + 130, + 211, + 482, + 300 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 211, + 482, + 300 + ], + "type": "text", + "content": " is determined and we obtain the scaled camera height " + }, + { + "bbox": [ + 130, + 211, + 482, + 300 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{cam}}^{n,\\tau} = s \\cdot H_{\\mathrm{cam}}'^{\\prime,\\tau}" + }, + { + "bbox": [ + 130, + 211, + 482, + 300 + ], + "type": "text", + "content": ". At the end of each epoch, " + }, + { + "bbox": [ + 130, + 211, + 482, + 300 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{cam}}^*" + }, + { + "bbox": [ + 130, + 211, + 482, + 300 + ], + "type": "text", + "content": " is optimized across a series of consecutive frames and updated with the weighted moving average." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 347, + 479, + 371 + ], + "lines": [ + { + "bbox": [ + 130, + 347, + 479, + 371 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 479, + 371 + ], + "type": "text", + "content": "Fig. 1 depicts the overall architecture of FUMET. Please see the caption for an overview of the framework. Let “'” denote “not metrically scaled”." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 322, + 205, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 205, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 205, + 335 + ], + "type": "text", + "content": "4 FUMET" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 389, + 350, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 389, + 350, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 389, + 350, + 402 + ], + "type": "text", + "content": "4.1 Scale-Aware Self-Supervised Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 409, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 409, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 482, + 517 + ], + "type": "text", + "content": "In FUMET, we derive a camera height pseudo-supervision from depth estimates. To obtain the per-frame camera height, we first compute the normal vector for each pixel " + }, + { + "bbox": [ + 130, + 409, + 482, + 517 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 409, + 482, + 517 + ], + "type": "text", + "content": ". We consider the 8-neighborhood of " + }, + { + "bbox": [ + 130, + 409, + 482, + 517 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 409, + 482, + 517 + ], + "type": "text", + "content": " and form 8 pairs " + }, + { + "bbox": [ + 130, + 409, + 482, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(p_i) = \\{(p_{j_0}^i,p_{j_1}^i)\\}_{j = 1}^8" + }, + { + "bbox": [ + 130, + 409, + 482, + 517 + ], + "type": "text", + "content": ". Each pair " + }, + { + "bbox": [ + 130, + 409, + 482, + 517 + ], + "type": "inline_equation", + "content": "(p_{j_0}^i,p_{j_1}^i)" + }, + { + "bbox": [ + 130, + 409, + 482, + 517 + ], + "type": "text", + "content": " consists of two pixels whose difference vectors w.r.t. " + }, + { + "bbox": [ + 130, + 409, + 482, + 517 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 409, + 482, + 517 + ], + "type": "text", + "content": " are perpendicular to each other and are sorted counterclockwise. We reproject the paired pixels onto 3D space and compute the cross-product of the two difference vectors w.r.t. the reprojected 3D point of " + }, + { + "bbox": [ + 130, + 409, + 482, + 517 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 409, + 482, + 517 + ], + "type": "text", + "content": ". By summing and normalizing all the cross-product vectors of each pair, we obtain the normal vector " + }, + { + "bbox": [ + 130, + 409, + 482, + 517 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}}(p_i)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 525, + 481, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 525, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 159, + 525, + 481, + 556 + ], + "type": "interline_equation", + "content": "\\mathbf {n} \\left(p _ {i}\\right) = \\sum_ {\\mathcal {N} \\left(p _ {i}\\right)} \\left(\\phi \\left(p _ {j _ {0}} ^ {i}\\right) - \\phi \\left(p _ {i}\\right)\\right) \\times \\left(\\phi \\left(p _ {j _ {1}} ^ {i}\\right) - \\phi \\left(p _ {i}\\right)\\right), \\quad \\hat {\\mathbf {n}} \\left(p _ {i}\\right) = \\frac {\\mathbf {n} \\left(p _ {i}\\right)}{\\| \\mathbf {n} \\left(p _ {i}\\right) \\|}, \\tag {6}", + "image_path": "a840f020b2a4343d57cf070d73ae02b1042cc629b2a3b13c4efa0b4b4d7de347.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 566, + 328, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 566, + 328, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 328, + 578 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 566, + 328, + 578 + ], + "type": "inline_equation", + "content": "\\phi (\\cdot)" + }, + { + "bbox": [ + 132, + 566, + 328, + 578 + ], + "type": "text", + "content": " denotes perspective reprojection." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 578, + 480, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 480, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 480, + 613 + ], + "type": "text", + "content": "For pixels in the road region, the negated inner product between the normal vector and the reprojected 3D point corresponds to the camera height from the ground plane" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 245, + 613, + 481, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 613, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 245, + 613, + 481, + 628 + ], + "type": "interline_equation", + "content": "H _ {\\mathrm {c a m}} ^ {\\prime} (p _ {i}) = - \\phi (p _ {i}) \\cdot \\hat {\\mathbf {n}} (p _ {i}). \\tag {7}", + "image_path": "eb02e74efe8fe4df2c0c8045ddf806a38eb14c47965b9b36ddd4237822c065c9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 632, + 448, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 448, + 644 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 448, + 644 + ], + "type": "text", + "content": "We represent the median of " + }, + { + "bbox": [ + 132, + 632, + 448, + 644 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{cam}}'(p_i)" + }, + { + "bbox": [ + 132, + 632, + 448, + 644 + ], + "type": "text", + "content": " as the camera height of the frame" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 226, + 652, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 652, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 226, + 652, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\tilde {H} _ {\\mathrm {c a m}} ^ {\\prime} = \\operatorname {m e d} \\left(\\left\\{H _ {\\mathrm {c a m}} ^ {\\prime} \\left(p _ {i}\\right) \\mid p _ {i} \\in M _ {\\mathrm {r}} \\right\\}\\right), \\tag {8}", + "image_path": "1e9ff777d4561716d8d517efbc9cfa2e3e845d37562b43d37e9a5468858f03b4.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 139, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 139, + 91, + 448, + 102 + ], + "type": "text", + "content": "Unsupervised Training for Metric Monocular Road-Scene Depth Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": " is in the road region mask " + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{r}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": " obtained with an off-the-shelf semantic segmentation model [22] before training. This camera height " + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "\\tilde{H}_{\\mathrm{cam}}'" + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": " estimate is unscaled and needs to be upgraded to a real-scaled value to impose the invariance loss across frames. By using the Silhouette Projector, we aggregate the scale information from the object height prior into a per-frame scale factor " + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": " (see Sec. 4.2) and obtain the scaled camera height as " + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{cam}} = s \\cdot \\tilde{H}_{\\mathrm{cam}}'" + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 273 + ], + "type": "text", + "content": "We leverage the fact that the camera height does not change in a sequence and formulate the invariance into an optimization across frames and training epochs. Thanks to this optimization, we acquire an accurate and stable supervision, which is also consistent across the scene. At the end of the " + }, + { + "bbox": [ + 130, + 188, + 482, + 273 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 188, + 482, + 273 + ], + "type": "text", + "content": "-th epoch, we use the median of " + }, + { + "bbox": [ + 130, + 188, + 482, + 273 + ], + "type": "inline_equation", + "content": "\\{H_{\\mathrm{cam}}^{n,\\tau}\\}_{n=1}^{N}" + }, + { + "bbox": [ + 130, + 188, + 482, + 273 + ], + "type": "text", + "content": " for all frames " + }, + { + "bbox": [ + 130, + 188, + 482, + 273 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 188, + 482, + 273 + ], + "type": "text", + "content": " as the representative camera height " + }, + { + "bbox": [ + 130, + 188, + 482, + 273 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{cam}}^{\\tau}" + }, + { + "bbox": [ + 130, + 188, + 482, + 273 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 130, + 188, + 482, + 273 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 188, + 482, + 273 + ], + "type": "text", + "content": "-th epoch. Then, we update the scaled camera height pseudo-invariance " + }, + { + "bbox": [ + 130, + 188, + 482, + 273 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{cam}}^{*}" + }, + { + "bbox": [ + 130, + 188, + 482, + 273 + ], + "type": "text", + "content": " with its weighted moving average" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 194, + 282, + 481, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 282, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 194, + 282, + 481, + 308 + ], + "type": "interline_equation", + "content": "H _ {\\mathrm {c a m}} ^ {* \\tau} \\leftarrow \\left\\{\\frac {\\tau (\\tau - 1)}{2} \\cdot H _ {\\mathrm {c a m}} ^ {* \\tau - 1} + \\tau \\cdot H _ {\\mathrm {c a m}} ^ {\\tau} \\right\\} / \\frac {\\tau (\\tau + 1)}{2}. \\tag {9}", + "image_path": "6758292f4aec4f27d0c6a3ec1cdc03ccabb382e841a3436de32b1398b74b5331.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 316, + 481, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 481, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 481, + 351 + ], + "type": "text", + "content": "When training on a dataset containing images captured at multiple camera heights, we create pseudo-labels corresponding to each camera height and optimize them individually." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 352, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 389 + ], + "type": "text", + "content": "To make the network learn the absolute scale and directly produce the scaled depth map, we enforce a loss on the camera height " + }, + { + "bbox": [ + 130, + 352, + 482, + 389 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{cam}}'" + }, + { + "bbox": [ + 130, + 352, + 482, + 389 + ], + "type": "text", + "content": " using the pseudo-invariance " + }, + { + "bbox": [ + 130, + 352, + 482, + 389 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{cam}}^{*\\tau - 1}" + }, + { + "bbox": [ + 130, + 352, + 482, + 389 + ], + "type": "text", + "content": " computed in the previous epoch" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 207, + 398, + 481, + 424 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 398, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 207, + 398, + 481, + 424 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c a m}} = \\frac {1}{| M _ {\\mathrm {r}} |} \\sum_ {i} M _ {\\mathrm {r}} \\left(p _ {i}\\right) \\cdot \\left| H _ {\\mathrm {c a m}} ^ {\\prime} \\left(p _ {i}\\right) - H _ {\\mathrm {c a m}} ^ {* \\tau - 1} \\right|. \\tag {10}", + "image_path": "22c1013c86bf7a0bbb4396902c28e29d5ac77c4a00c455418968cbc44b9deea3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 434, + 482, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 482 + ], + "type": "text", + "content": "It is worth noting that at inference, the network does not assume camera height invariance in the input sequence; it is only assumed during training. The computational cost at inference is exactly the same as the original MDE model trained without FUMET." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 500, + 264, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 500, + 264, + 513 + ], + "spans": [ + { + "bbox": [ + 131, + 500, + 264, + 513 + ], + "type": "text", + "content": "4.2 Silhouette Projector" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "text", + "content": "In the process of creating a scaled camera height supervision, we need to know the per-frame scale factor. For this, we can consider leveraging an object of known size. This, however, would necessitate extraction of the exact 3D object region which is non-trivial as objects are often occluded or cropped and oriented in various poses." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "We introduce Silhouette Projector to robustly estimate this per-frame scale factor. Silhouette Projector leverages two simple facts of road scenes. First, the height of the object silhouette projected on the plane perpendicular to the ground plane does not change regardless of the object pose. Second, we can compute the silhouette height by measuring the distance from the top of the silhouette to the ground plane even if they are partly occluded or truncated, as long as the top is visible and the object is on the ground plane." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "G.Kinoshita et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": "We first reconstruct an object point cloud with the estimated depth map through perspective reprojection. The object region " + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{obj}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": " in the image can be obtained with an off-the-shelf instance segmentation model [22] before training. We then orthographically project these 3D points onto a plane " + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "inline_equation", + "content": "P^{\\perp}" + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": " perpendicular to the ground plane, i.e., an arbitrary plane parallel to the road normal vector " + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{n}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": ". We compute " + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{n}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": " as the median of the normal vectors " + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}}(p_i)" + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": " (see Eq. (6)) in the road region " + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{r}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": ". We define the farthest distance from the 2D point on " + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "inline_equation", + "content": "P^{\\perp}" + }, + { + "bbox": [ + 130, + 116, + 482, + 273 + ], + "type": "text", + "content": " to the ground plane as the silhouette height. This value is unscaled because we derived this from the estimated unscaled depth map. The ratio between a height prior and its silhouette height corresponds to the scale factor of interest. When there are multiple objects in an image, we use the median of the scale factors each corresponding to each object as the scale factor of the frame for computational simplicity and robustness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 293, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 365 + ], + "type": "text", + "content": "Outlier Filtering. Some objects cannot be handled well with Silhouette Projector for various reasons including occluded upper parts, not on the ground plane, inaccurate segmented region, and misclassified object category. We introduce an algorithm to detect these as outliers based on geometric plausibility. We first calculate the horizon " + }, + { + "bbox": [ + 130, + 293, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{l}" + }, + { + "bbox": [ + 130, + 293, + 482, + 365 + ], + "type": "text", + "content": " with the road normal vector " + }, + { + "bbox": [ + 130, + 293, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{n}}" + }, + { + "bbox": [ + 130, + 293, + 482, + 365 + ], + "type": "text", + "content": " and the camera intrinsics " + }, + { + "bbox": [ + 130, + 293, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 130, + 293, + 482, + 365 + ], + "type": "text", + "content": " [17]" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 282, + 367, + 480, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 367, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 282, + 367, + 480, + 380 + ], + "type": "interline_equation", + "content": "\\mathbf {l} = \\mathbf {K} ^ {- \\top} \\tilde {\\mathbf {n}}. \\tag {11}", + "image_path": "4e605b8238faae66c7ad4f264f5f4bbd16429e323cf8668afbc32595546f26ee.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "text", + "content": "Based on the approximations in [20], we derive the approximate metric-scale object height " + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\hat{H}_{\\mathrm{obj}} \\approx \\frac{h_{\\mathrm{obj}}}{h_{\\mathrm{cam}}} H_{\\mathrm{cam}}^{*\\tau - 1}" + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{cam}}" + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "text", + "content": " is the farthest distance from the pixel in the object instance mask to " + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{l}" + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{obj}}" + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "text", + "content": " is the farthest distance between two points on the object mask on which tangent lines parallel to " + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{l}" + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "text", + "content": " lie, and " + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{cam}}^{*\\tau - 1}" + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "text", + "content": " is the pseudo camera height at the " + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "text", + "content": "-th epoch derived in Eq. (9). By comparing the relative gap between " + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "inline_equation", + "content": "\\hat{H}_{\\mathrm{obj}}" + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "text", + "content": " and the object height prior " + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{obj}}" + }, + { + "bbox": [ + 130, + 388, + 482, + 472 + ], + "type": "text", + "content": ", we determine the outliers as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 252, + 474, + 480, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 474, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 252, + 474, + 480, + 506 + ], + "type": "interline_equation", + "content": "\\left\\{k \\left| \\frac {\\left| H _ {\\mathrm {o b j}} ^ {k} - \\hat {H} _ {\\mathrm {o b j}} ^ {k} \\right|}{H _ {\\mathrm {o b j}} ^ {k}} > \\mathcal {T} \\right. \\right\\}. \\tag {12}", + "image_path": "b0443a2d659d0560aab40effc714195f079c4c81d67c2f58346e416173e4d6aa.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 511, + 367, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 511, + 367, + 522 + ], + "spans": [ + { + "bbox": [ + 131, + 511, + 367, + 522 + ], + "type": "text", + "content": "In all the experiments, we set the threshold " + }, + { + "bbox": [ + 131, + 511, + 367, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 131, + 511, + 367, + 522 + ], + "type": "text", + "content": " to 0.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 545, + 239, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 545, + 239, + 558 + ], + "spans": [ + { + "bbox": [ + 131, + 545, + 239, + 558 + ], + "type": "text", + "content": "4.3 Training Losses" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": "Auxiliary Rough Geometric Loss. We experimentally observe that estimated object depths, especially those distant from the camera, often erroneously lean towards the road surface because of the inaccurate supervisory signal from the reconstruction loss in low-texture regions. This tendency makes silhouette heights inaccurate. To mitigate this issue, we introduce an auxiliary rough geometric loss " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{aux}}" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": ", which enforces object point clouds to be within a close region. For each inlier object " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "k = 1 \\dots K" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": "), we define " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{aux}}" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " as the gap between the estimated depth map " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " in the object region " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{obj}}" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " and the approximate object" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 139, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 139, + 91, + 447, + 102 + ], + "type": "text", + "content": "Unsupervised Training for Metric Monocular Road-Scene Depth Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 353, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 353, + 131 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 353, + 131 + ], + "type": "text", + "content": "depth " + }, + { + "bbox": [ + 130, + 114, + 353, + 131 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{aprx}} = \\frac{H_{\\mathrm{obj}}}{h_{\\mathrm{obj}}} f_y" + }, + { + "bbox": [ + 130, + 114, + 353, + 131 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 114, + 353, + 131 + ], + "type": "inline_equation", + "content": "f_{y}" + }, + { + "bbox": [ + 130, + 114, + 353, + 131 + ], + "type": "text", + "content": " is the focal length" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 195, + 140, + 481, + 170 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 140, + 481, + 170 + ], + "spans": [ + { + "bbox": [ + 195, + 140, + 481, + 170 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a u x}} = \\frac {1}{K} \\sum_ {k} \\frac {1}{| M _ {\\mathrm {o b j}} ^ {k} |} \\sum_ {i} M _ {\\mathrm {o b j}} ^ {k} \\left(p _ {i}\\right) \\cdot \\left| D \\left(p _ {i}\\right) - D _ {\\mathrm {a p r x}} ^ {k} \\right|. \\tag {13}", + "image_path": "ae7d943748e73eb9586578055f00ccce7017bf4023315d6c828c5ad99c024979.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 178, + 410, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 178, + 410, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 410, + 190 + ], + "type": "text", + "content": "As a by-product, this loss accelerates learning the metric scale." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 190, + 482, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 482, + 285 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 130, + 190, + 482, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{aux}}" + }, + { + "bbox": [ + 130, + 190, + 482, + 285 + ], + "type": "text", + "content": " is potentially inaccurate since " + }, + { + "bbox": [ + 130, + 190, + 482, + 285 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{aprx}}" + }, + { + "bbox": [ + 130, + 190, + 482, + 285 + ], + "type": "text", + "content": " assumes that each object can be represented by a plane parallel to the image plane, and we observe that it degrades the depth accuracy in the late stage of training. On the other hand, in the early stage, estimated depth is unreliable, so heavily relying on the camera height loss " + }, + { + "bbox": [ + 130, + 190, + 482, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cam}}" + }, + { + "bbox": [ + 130, + 190, + 482, + 285 + ], + "type": "text", + "content": " makes the training unstable. For this, we gradually decrease the weight of " + }, + { + "bbox": [ + 130, + 190, + 482, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{aux}}" + }, + { + "bbox": [ + 130, + 190, + 482, + 285 + ], + "type": "text", + "content": " and increase the one for " + }, + { + "bbox": [ + 130, + 190, + 482, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cam}}" + }, + { + "bbox": [ + 130, + 190, + 482, + 285 + ], + "type": "text", + "content": " as the training proceeds in a logarithmic way. To ensure training stability, we stop changing these weights after the " + }, + { + "bbox": [ + 130, + 190, + 482, + 285 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{mid-th}}" + }, + { + "bbox": [ + 130, + 190, + 482, + 285 + ], + "type": "text", + "content": " epoch" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 295, + 481, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 295, + 481, + 327 + ], + "spans": [ + { + "bbox": [ + 141, + 295, + 481, + 327 + ], + "type": "interline_equation", + "content": "\\lambda_ {\\mathrm {a u x}} (\\tau) = \\left\\{ \\begin{array}{l l} \\frac {- \\log (\\tau + 1)}{\\log \\left(\\tau_ {\\mathrm {m i d}} + 1\\right)} & (\\tau \\leq \\tau_ {\\mathrm {m i d}}) \\\\ \\epsilon & (\\tau > \\tau_ {\\mathrm {m i d}}) \\end{array} , \\lambda_ {\\mathrm {c a m}} (\\tau) = \\left\\{ \\begin{array}{l l} \\frac {\\log (\\tau + 1)}{\\log \\left(\\tau_ {\\mathrm {m i d}} + 1\\right)} & (\\tau \\leq \\tau_ {\\mathrm {m i d}}) \\\\ 1. 0 & (\\tau > \\tau_ {\\mathrm {m i d}}) \\end{array} , \\right. \\right. \\tag {14}", + "image_path": "99a87c07e85a94dbcd7e43c9e98b39ab6e21b086c7f5b403692ef8a52c86e053.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 335, + 206, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 335, + 206, + 345 + ], + "spans": [ + { + "bbox": [ + 132, + 335, + 206, + 345 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 335, + 206, + 345 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.005" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 346, + 480, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 480, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 480, + 370 + ], + "type": "text", + "content": "The total loss is " + }, + { + "bbox": [ + 130, + 346, + 480, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\alpha \\lambda_{\\mathrm{cam}}\\mathcal{L}_{\\mathrm{cam}} + \\beta \\lambda_{\\mathrm{aux}}\\mathcal{L}_{\\mathrm{aux}} + \\mathcal{L}_{\\mathrm{sm}} + \\mathcal{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 130, + 346, + 480, + 370 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 346, + 480, + 370 + ], + "type": "inline_equation", + "content": "\\alpha = 0.01" + }, + { + "bbox": [ + 130, + 346, + 480, + 370 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 346, + 480, + 370 + ], + "type": "inline_equation", + "content": "\\beta = 1.0" + }, + { + "bbox": [ + 130, + 346, + 480, + 370 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 388, + 180, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 388, + 180, + 401 + ], + "spans": [ + { + "bbox": [ + 132, + 388, + 180, + 401 + ], + "type": "text", + "content": "5 LSP" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 414, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 569 + ], + "type": "text", + "content": "Our FUMET extracts metric scale from a prior on object heights through the camera height loss " + }, + { + "bbox": [ + 130, + 414, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cam}}" + }, + { + "bbox": [ + 130, + 414, + 482, + 569 + ], + "type": "text", + "content": " and the auxiliary rough geometric loss " + }, + { + "bbox": [ + 130, + 414, + 482, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{aux}}" + }, + { + "bbox": [ + 130, + 414, + 482, + 569 + ], + "type": "text", + "content": ". To ensure that FUMET learns the scale regardless of the particular dataset used for training, a universal prior of object heights, independent of the environment, becomes essential. Cars are objects whose dimensions are explicitly and accurately known. Better yet, they are abundantly located on the road, the key semantic region whose metric depth is essential for many downstream tasks. Some methods for other tasks have modeled the size prior as a probability distribution [20, 57]. The gap between its mean and the true expectation for each environment, however, is directly reflected in the uncertainty of the scale in the supervision. Wang et al. [40] also report that there is a gap in car dimensions among datasets. These facts suggest that modeling the object size prior as a single probability distribution is problematic, especially for man-made objects like cars." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 629 + ], + "type": "text", + "content": "We humans not only possess rough prior knowledge about the vehicle size but can also estimate it more accurately by extracting instance-specific information such as car models from its appearance. Motivated by this intuition, we introduce a learned universal vehicle size prior termed LSP. Given a masked vehicle image, LSP estimates its dimensions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "For training, we use a dataset consisting of a large number of vehicle images and their information such as 3D dimensions. This dataset can be easily collected with web scraping and does not require manual annotations. To enhance" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "G.Kinoshita et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 210 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 210 + ], + "type": "text", + "content": "the robustness to various input images, in addition to common color jittering and blur augmentation, we perform augmentations to simulate occlusion and truncation. For occlusion augmentation, we mask out random regions with the same shape as a vehicle or with a beam of random width. For truncation augmentation, we translate the vehicle mask in a random direction. We visualize these in the supplementary material. The abundance of training images and diverse augmentations enable LSP to estimate vehicle dimensions robustly across various datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 480, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 480, + 248 + ], + "type": "text", + "content": "Although FUMET requires only a car height prior, we construct LSP to estimate the width and length as well as height. This enhances it to grasp an overall car size and improves the prediction accuracy of car heights." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 265, + 230, + 279 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 265, + 230, + 279 + ], + "spans": [ + { + "bbox": [ + 131, + 265, + 230, + 279 + ], + "type": "text", + "content": "6 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 289, + 204, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 289, + 204, + 300 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 204, + 300 + ], + "type": "text", + "content": "6.1 FUMET" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 308, + 480, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 480, + 357 + ], + "type": "text", + "content": "We evaluate the accuracy of depth estimators trained on our framework and demonstrate the effectiveness of FUMET as a novel unsupervised metric-scale training method through depth prediction accuracy, its applicability to any network architecture, and its ability to leverage any dataset for training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "text", + "content": "Implementation Details. We implement FUMET with PyTorch [28]. We employ Monodepth2 [15] with ResNet50 encoder pre-trained on ImageNet [30] as the default depth estimator which is one of the most basic architectures in self-supervised MDE. For all experiments unless mentioned otherwise, we train the model for 50 epochs with a batch size of 8 and set " + }, + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{mid}}" + }, + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "text", + "content": " to 20. We use Adam optimizer [24] with an initial learning rate of " + }, + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "inline_equation", + "content": "5.0\\times 10^{-5}" + }, + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "text", + "content": " decayed by half every 15 epochs. During training, we perform the following augmentations with " + }, + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "text", + "content": " chance in random order: horizontal flips, brightness adjustment " + }, + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "inline_equation", + "content": "(\\pm 0.2)" + }, + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "text", + "content": ", saturation adjustment " + }, + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "inline_equation", + "content": "(\\pm 0.2)" + }, + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "text", + "content": ", contrast adjustment " + }, + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "inline_equation", + "content": "(\\pm 0.2)" + }, + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "text", + "content": ", and hue jitter " + }, + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "inline_equation", + "content": "(\\pm 0.1)" + }, + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "text", + "content": ". For quantitative evaluation, we follow [9] and compute the seven standard metrics: AbsRel, SqRel, RMSE, " + }, + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{RMSE}_{\\log}" + }, + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\delta < 1.25" + }, + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\delta < 1.25^2" + }, + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\delta < 1.25^3" + }, + { + "bbox": [ + 130, + 373, + 482, + 506 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 479, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 479, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 479, + 593 + ], + "type": "text", + "content": "KITTI. Our FUMET is independent of network architectures so we experiment on several models with FUMET as well as Monodepth2: Lite-Mono [52] with ResNet18, HR-Depth [25] with ResNet18, and VADepth [45]. We train these models on the KITTI raw dataset [12], which contains 39,810 images for training and 4,424 for validation, and evaluate it with the KITTI Eigen test split [8] (697 images). For training and evaluation, all images are resized to " + }, + { + "bbox": [ + 130, + 521, + 479, + 593 + ], + "type": "inline_equation", + "content": "640 \\times 192" + }, + { + "bbox": [ + 130, + 521, + 479, + 593 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "We show the qualitative results in Fig. 2. The quantitative results shown in Tab. 1 demonstrate that FUMET enables the various architectures to learn metric scale properly. Even when employing the simplest network architecture, i.e., Monodepth2, it outperforms weakly-supervised methods that rely on ground-truth scale labels for training which is not necessary for our method. Similar to VADepth [45], FUMET uses the camera height invariance as supervision though" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 139, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 139, + 91, + 447, + 102 + ], + "type": "text", + "content": "Unsupervised Training for Metric Monocular Road-Scene Depth Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 224, + 479, + 469 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "text", + "content": "Table 1: Quantitative results on the KITTI Eigen test split [8] (" + }, + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "inline_equation", + "content": "640 \\times 192" + }, + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "text", + "content": " resolution). In the Supervision column, V, G, and CamH represent velocity, gravity, and camera height, respectively. The Scaling column represents whether to perform the median scaling [56]. We denote the best results in bold and the second best ones with an underline for each block. Note that we train VADepth [45] from scratch using ground-truth camera height (1.65m) provided by KITTI, unlike its original paper. FUMET outperforms weakly-supervised methods regardless of the network architectures and without the need for ground truth supervision. In the rescaled results, every model trained with FUMET achieves better results than its original one." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 224, + 479, + 469 + ], + "lines": [ + { + "bbox": [ + 133, + 224, + 479, + 469 + ], + "spans": [ + { + "bbox": [ + 133, + 224, + 479, + 469 + ], + "type": "table", + "html": "
MethodSupervisionScalingError (↓)Accuracy (↑)
AbsRelSqRelRMSERMSElogδ<1.25δ<1.252δ<1.253
G2S [3]GPS-0.1090.8604.8550.1980.8650.9540.980
PackNet-SfM [16]V-0.1110.8294.7880.1990.8640.9540.980
Wagstaff et al. [36]CamH-0.1230.9965.2530.2130.8400.9470.978
VADepth [45]CamH-0.1200.9754.9710.2030.8670.9560.979
DynaDepth [53]IMU+V+G-0.1090.7874.7050.1950.8690.9580.981
Ours--0.1080.7854.7360.1950.8710.9580.981
Lite-Mono [52]+Ours--0.1120.7984.6920.1940.8710.9590.981
HR-Depth [25]+Ours--0.1090.7944.7340.1930.8690.9590.982
VADepth [45]+Ours--0.1080.8094.5720.1850.8830.9630.982
G2S [3]GPS✓0.1120.8944.8520.1920.8770.9580.981
PackNet-SfM [16]V✓0.1110.7854.6010.1890.8780.9600.982
Wagstaff et al. [36]CamH✓0.1170.9524.9890.1970.8670.9570.981
VADepth [45]CamH✓0.1100.9774.8720.1870.8890.9620.981
DynaDepth [53]IMU+V+G✓0.1080.7614.6080.1870.8830.9620.982
Ours-✓0.1060.7594.6020.1840.8870.9630.983
Monodepth2 [15]-✓0.1150.9034.8630.1930.8770.9590.981
+ Ours-✓0.1060.7594.6020.1840.8870.9630.983
HR-Depth [25]-✓0.1090.7924.6320.1850.8840.9620.983
+ Ours-✓0.1080.7594.5840.1840.8850.9630.983
Lite-Mono [52]-✓0.1100.8024.6710.1860.8790.9610.982
+ Ours-✓0.1070.7754.5640.1830.8880.9630.983
VADepth [45]-✓0.1040.7744.5520.1810.8920.9650.982
+ Ours-✓0.1020.7604.4730.1780.8970.9660.983
", + "image_path": "95fcf9ea43a3540a38e7735e083de082977449474786237020b9ba2732f03e76.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 511, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 482, + 571 + ], + "type": "text", + "content": "VADepth requires the ground-truth camera height and employs a complex network architecture. FUMET with Monodepth2 (row 6 in Tab. 1) achieves higher accuracy. This indicates that measuring camera height precisely is difficult and using unreliable ground-truth camera height causes low accuracy in the scale of estimated depth maps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "In the results showing the accuracy of the estimated depth map scaled with ground-truth scale factors via median scaling [56], it is evident that every model trained with FUMET attains superior results to the one trained without FUMET. This shows that our training framework not only enables the models to learn metric scale but also enhances the geometric accuracy. This improvement owes to the camera height loss " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cam}}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{aux}}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", which makes the model respect road regions and object regions, respectively." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "G.Kinoshita et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 128, + 478, + 290 + ], + "blocks": [ + { + "bbox": [ + 133, + 128, + 478, + 290 + ], + "lines": [ + { + "bbox": [ + 133, + 128, + 478, + 290 + ], + "spans": [ + { + "bbox": [ + 133, + 128, + 478, + 290 + ], + "type": "image", + "image_path": "1ae246018d255e87deaff77763c04c403f7ba641eb1140f2fafdf86d6dc1d768.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 306, + 482, + 340 + ], + "lines": [ + { + "bbox": [ + 130, + 306, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 482, + 340 + ], + "type": "text", + "content": "Fig. 2: Qualitative comparison on KITTI. In error maps, the larger depth errors are represented in red, smaller ones are in blue. The model trained with our FUMET predicts more accurate depth maps compared to the weakly-supervised methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 363, + 482, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 400 + ], + "type": "text", + "content": "FUMET successfully enables a model to learn metric scale, even though the training dataset of KITTI contains many frames with no or few detected cars. We detail this in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 418, + 482, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 526 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 526 + ], + "type": "text", + "content": "Cityscapes. To evaluate the ability of FUMET to leverage any dataset for training, we train and evaluate it on the Cityscapes dataset [6]. We follow [43, 46, 48, 56] and collect 69,731 images for training and evaluate on the 1,525 test images. In this dataset, there are multiple sets of camera intrinsics. Since this strongly affects the metric depth estimation [47], we first align the focal lengths of all images to 587.5 by resizing them, and crop them to " + }, + { + "bbox": [ + 130, + 418, + 482, + 526 + ], + "type": "inline_equation", + "content": "512 \\times 192" + }, + { + "bbox": [ + 130, + 418, + 482, + 526 + ], + "type": "text", + "content": " resolution. For evaluation, we follow [2, 43] and further crop the estimated image into " + }, + { + "bbox": [ + 130, + 418, + 482, + 526 + ], + "type": "inline_equation", + "content": "416 \\times 128" + }, + { + "bbox": [ + 130, + 418, + 482, + 526 + ], + "type": "text", + "content": ". In order to align the number of iterations on KITTI, we train for 29 epochs with the same learning rate decreased by half every 8 epochs, and " + }, + { + "bbox": [ + 130, + 418, + 482, + 526 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{mid}}" + }, + { + "bbox": [ + 130, + 418, + 482, + 526 + ], + "type": "text", + "content": " is set to 12." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 527, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 527, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 527, + 482, + 598 + ], + "type": "text", + "content": "Tab. 2 shows the quantitative results and FUMET outperforms weakly-supervised methods by wide margins. These methods assume that highly accurate sensor data is available and fail to train scale-aware MDE models on datasets consisting of unreliable sensor data such as Cityscapes. On the other hand, FUMET does not require any sensor measurement other than RGB video, so it can robustly leverage any driving videos for training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Mixed datasets. To demonstrate that FUMET can train a model on a mixed set of datasets to attain strong generalizability, we train with 235,341 images collected from the Argoverse2 [44], Lyft [21], A2D2 [13], and DDAD [16] dataset and evaluate it on the KITTI Eigen test split [8]. Both images for training and" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 139, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 139, + 91, + 447, + 102 + ], + "type": "text", + "content": "Unsupervised Training for Metric Monocular Road-Scene Depth Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 190, + 298, + 249 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 301, + 180 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 301, + 180 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 301, + 180 + ], + "type": "text", + "content": "Table 2: Quantitative results on Cityscapes [6]. We train all the models from scratch with the same supervisions as the Supervision column in Tab. 1. FUMET significantly outperforms weakly-supervised methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 190, + 298, + 249 + ], + "lines": [ + { + "bbox": [ + 135, + 190, + 298, + 249 + ], + "spans": [ + { + "bbox": [ + 135, + 190, + 298, + 249 + ], + "type": "table", + "html": "
Method|AbsRelSqRelRMSEδ<1.25
G2S [3]4.156276.1658.890.046
PackNet-SfM [16]0.5046.63914.900.029
VADepth [45]0.3637.11511.950.295
FUMET0.1251.2886.3590.858
", + "image_path": "e8b0b2119c61b3f15e71387f08c49f20039b50583335c768c484b996aaaa6e5a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 313, + 190, + 480, + 236 + ], + "blocks": [ + { + "bbox": [ + 311, + 114, + 482, + 181 + ], + "lines": [ + { + "bbox": [ + 311, + 114, + 482, + 181 + ], + "spans": [ + { + "bbox": [ + 311, + 114, + 482, + 181 + ], + "type": "text", + "content": "Table 3: FUMET can be trained on mixed datasets (Argoverse2, Lyft, A2D2, and DDAD) with various camera settings, and attains high generalizability. The results are evaluated on the KITTI Eigen split [8] " + }, + { + "bbox": [ + 311, + 114, + 482, + 181 + ], + "type": "inline_equation", + "content": "(832\\times 512" + }, + { + "bbox": [ + 311, + 114, + 482, + 181 + ], + "type": "text", + "content": " resolution)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 190, + 480, + 236 + ], + "lines": [ + { + "bbox": [ + 313, + 190, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 480, + 236 + ], + "type": "table", + "html": "
Training datasetAbsRelSqRelRMSEδ<1.25
KITTI0.1030.6754.7080.903
Mixed0.1130.9165.0090.883
Mixed+KITTI0.0820.6114.3070.923
", + "image_path": "aedfdb04c55d4b83ffb2a77a3137ac06788f07234c33cbdd75b37b0b16b1cff2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 273, + 479, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 273, + 479, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 273, + 479, + 332 + ], + "type": "text", + "content": "evaluation are resized and cropped to " + }, + { + "bbox": [ + 130, + 273, + 479, + 332 + ], + "type": "inline_equation", + "content": "832 \\times 512" + }, + { + "bbox": [ + 130, + 273, + 479, + 332 + ], + "type": "text", + "content": " resolution. The training images are the sets of video sequences and each sequence is captured with its own unique camera setting. We individually optimize each camera height for each sequence during training. We detail the dataset composition and training settings in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 333, + 480, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 480, + 403 + ], + "type": "text", + "content": "As shown in Tab. 3, the model trained with FUMET on the large-scale mixed dataset achieves comparable accuracy to the one trained on KITTI despite the domain gap and attains high generalizability. The model trained on both the mixed dataset and KITTI outperforms the model trained only on KITTI. This suggests that training on various images enhances accuracy which is the strength of fully self-supervised learning methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 405, + 481, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 481, + 441 + ], + "type": "text", + "content": "We also show that FUMET can leverage in-the-wild videos for training by using camera intrinsics recovered with SLAM, e.g., COLMAP [32] in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 461, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 605 + ], + "type": "text", + "content": "Ablation Studies on KITTI. To study the effect of the proposed components in FUMET, we conduct ablation studies on the training losses, balancing the loss weights, camera height optimization, and outlier filtering. The results are shown in Tab. 4. Thanks to the accuracy of LSP and the robustness of Silhouette Projector, the camera height loss makes a greater contribution to the depth accuracy than the auxiliary rough geometric loss. By introducing the camera height optimization, the model can learn the scale with more accurate and consistent supervision than leveraging the size prior independently for each frame (rows 1, 2, and 6 in Tab. 4), which leads to improved accuracy. Though balancing the loss weights (Eq. (14)) with either " + }, + { + "bbox": [ + 130, + 461, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{aux}}" + }, + { + "bbox": [ + 130, + 461, + 482, + 605 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 461, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cam}}" + }, + { + "bbox": [ + 130, + 461, + 482, + 605 + ], + "type": "text", + "content": " degrades the results due to the insufficient strength of the supervisory signals for metric scale, it contributes to improving accuracy when both " + }, + { + "bbox": [ + 130, + 461, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{aux}}" + }, + { + "bbox": [ + 130, + 461, + 482, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 461, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cam}}" + }, + { + "bbox": [ + 130, + 461, + 482, + 605 + ], + "type": "text", + "content": " are employed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": "Our proposed method can also be used with a pre-trained MDE model in an offline manner. We first train the original Monodepth2 [15] without FUMET, then predict depth maps of training images. With Silhouette Projector and LSP, we obtain per-frame scaled camera height and optimize it across frames with the median operation. We fix this value and use it as the camera height supervision" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "G.Kinoshita et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 179, + 310, + 265 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 310, + 169 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 310, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 310, + 169 + ], + "type": "text", + "content": "Table 4: Each loss contributes to learning metric scale and the others enhance the depth accuracy. Bal: balancing the loss weights (Eq. (14)). Opt: camera height optimization. Flt: outlier filtering." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 179, + 310, + 265 + ], + "lines": [ + { + "bbox": [ + 133, + 179, + 310, + 265 + ], + "spans": [ + { + "bbox": [ + 133, + 179, + 310, + 265 + ], + "type": "table", + "html": "
\\( \\mathcal{L}_{\\text{aux}} \\)\\( \\mathcal{L}_{\\text{cam}} \\)BalOptFltAbsRelSqRelRMSE\\( \\delta < 1.25 \\)
✓---✓0.1250.8875.0550.834
✓-✓-✓0.1490.9505.2190.785
-✓-✓✓0.1160.8754.8740.863
-✓✓✓✓0.2111.3036.1430.585
✓✓-✓✓0.1220.8805.0170.849
✓✓✓-✓0.1150.8224.8110.858
✓✓✓✓-0.1160.8234.9010.860
✓✓✓✓✓0.1080.7854.7360.871
", + "image_path": "298a50cc890b347e099a232bd8c4b5d570e841bdf446a514df813e2eaf395518.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 320, + 224, + 478, + 269 + ], + "blocks": [ + { + "bbox": [ + 318, + 114, + 480, + 213 + ], + "lines": [ + { + "bbox": [ + 318, + 114, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 318, + 114, + 480, + 213 + ], + "type": "text", + "content": "Table 5: Training with a fine-tuned pseudo camera height achieves the highest accuracy. Online: optimizing the camera height while training. Offline: training with a fixed one computed from depth maps predicted with pre-trained Monodepth2. Fine-tune: training with the fixed one computed offline for 20 epochs, then optimizing online." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 224, + 478, + 269 + ], + "lines": [ + { + "bbox": [ + 320, + 224, + 478, + 269 + ], + "spans": [ + { + "bbox": [ + 320, + 224, + 478, + 269 + ], + "type": "table", + "html": "
Deriv. methodAbsRelSqRelRMSEδ<1.25
Online0.1080.8084.7400.870
Offline0.1060.7904.7190.873
Fine-tune0.1050.7824.6820.876
", + "image_path": "96c41be46c977c401190fed497ae037464701f57429324dd5421a3e2be55c7d9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 290, + 480, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 480, + 457 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 480, + 457 + ], + "type": "text", + "content": "for an untrained Monodepth2 on FUMET. We examine the difference between the proposed camera height optimized online and the fixed one derived offline when we use them as the pseudo-supervision. Note that we also fix " + }, + { + "bbox": [ + 130, + 290, + 480, + 457 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{cam}}" + }, + { + "bbox": [ + 130, + 290, + 480, + 457 + ], + "type": "text", + "content": " because the pseudo-label can be considered sufficiently accurate from the early stage of training. We also run another experiment in which we fix camera height supervision derived offline for the first 20 epochs, which is the same number of training epochs as the original Monodepth2, then unfit the camera height pseudo-label to fine-tune it via online optimization. As shown in Tab. 5, the model supervised with the fixed camera height computed offline achieves slightly better results than the one supervised with the camera height optimized online. This result indicates that our strategy of deriving scaled camera height is highly accurate and fixing it stabilizes the training. The model trained with the fine-tuned camera height achieves the highest accuracy, which shows that our FUMET can attain better accuracy with additional pre-training costs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 474, + 181, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 474, + 181, + 484 + ], + "spans": [ + { + "bbox": [ + 132, + 474, + 181, + 484 + ], + "type": "text", + "content": "6.2 LSP" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 493, + 456, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 456, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 456, + 504 + ], + "type": "text", + "content": "We provide implementation details of LSP in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 480, + 665 + ], + "type": "text", + "content": "Evaluation on KITTI and Cityscapes. We evaluate the generalizability of our LSP with the dimensions of the annotated 3D bounding box on both the KITTI [12] and the Cityscapes [6] datasets in the Car class. We compare the results of LSP to a fixed car height prior (1.59m, following [20]) and MonoCIinIS [19], a 3D object detection model that outputs object dimensions from an image and an instance mask. We train MonoCIinIS from scratch with the KITTI dataset. As an evaluation metric, we compute the absolute relative error for each dimension. When we estimate car dimensions with LSP and MonoCIinIS, we use a ground-truth instance mask for making a car image mask and as input, respectively. For all the experiments in LSP, input vehicle masks are cropped and resized to " + }, + { + "bbox": [ + 130, + 521, + 480, + 665 + ], + "type": "inline_equation", + "content": "300 \\times 300" + }, + { + "bbox": [ + 130, + 521, + 480, + 665 + ], + "type": "text", + "content": " resolution. On the KITTI dataset, we extract 3,769 samples for evaluation [4]. On the Cityscapes dataset, we evaluate on the validation split," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 139, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 139, + 91, + 447, + 102 + ], + "type": "text", + "content": "Unsupervised Training for Metric Monocular Road-Scene Depth Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 146, + 179, + 468, + 265 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": "Table 6: Generalizability on KITTI and Cityscapes. We evaluate LSP on Absolute Relative Error for each dimension (height/width/length), comparing with the fixed car height prior (1.59 m, following [20]) and MonoCInIS [19] trained on KITTI. LSP robustly surpasses other methods in most cases and can predict accurately regardless of the datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 146, + 179, + 468, + 265 + ], + "lines": [ + { + "bbox": [ + 146, + 179, + 468, + 265 + ], + "spans": [ + { + "bbox": [ + 146, + 179, + 468, + 265 + ], + "type": "table", + "html": "
MethodEasyModerateHardAll
KITTIFixed Height [20]0.065 / - / -0.064 / - / -0.065 / - / -0.065 / - / -
MonoClIns [19]0.069 / 0.065 / 0.0930.112 / 0.106 / 0.1330.268 / 0.263 / 0.2690.159 / 0.155 / 0.177
LSP (only height)0.057 / - / -0.089 / - / -0.106 / - / -0.087 / - / -
LSP0.046 / 0.136 / 0.1300.060 / 0.144 / 0.1390.073 / 0.155 / 0.1500.062 / 0.147 / 0.145
CityscapesFixed Height [20]0.093 / - / -0.093 / - / -0.076 / - / -0.092 / - / -
MonoClIns [19]0.578 / 0.599 / 0.5840.759 / 0.766 / 0.7670.843 / 0.852 / 0.8520.851 / 0.859 / 0.856
LSP (only height)0.063 / - / -0.075 / - / -0.103 / - / -0.096 / - / -
LSP0.056 / 0.069 / 0.0670.068 / 0.068 / 0.0780.069 / 0.059 / 0.0750.080 / 0.064 / 0.077
", + "image_path": "b8b23428d89e4835a8ecd35d9e30ed9ae75738701589483affbd6bd11d12d8be.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 290, + 479, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 479, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 479, + 338 + ], + "type": "text", + "content": "which has 500 samples. On the KITTI dataset, each object is separated into 3 levels based on detection difficulty: Easy, Moderate, and Hard. Considering the difference in focal lengths between KITTI and Cityscapes, we also separate objects on Cityscapes into the same levels as on KITTI." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 340, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 426 + ], + "type": "text", + "content": "As shown in Tab. 6, LSP outperforms the other two methods in most cases. Particularly, the accuracy of MonoCInIS drops on Cityscapes since it overfits to the training dataset, i.e., KITTI and cannot address the domain gap. In contrast, thanks to the large training data and various augmentations, our LSP obtains the highest generalizability and robustly estimates dimensions on any dataset. Learning not only car height but also width and length also improves the prediction accuracy as this facilitates estimation of the overall car sizes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 451, + 220, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 451, + 220, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 451, + 220, + 464 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "In this work, we proposed FUMET, a novel training framework for any MDE model to learn metric scale. FUMET aggregates the scale information obtained from a learned car size prior into the camera height and uses it as pseudosupervision formulated as optimization across frames and epochs by leveraging its invariance. We showed that FUMET enables monocular depth estimators to embody metric scale regardless of the underlying network architectures and achieves state-of-the-art absolute-scale depth estimation accuracy on both the KITTI and the Cityscapes datasets, and demonstrated that it can leverage any dataset for training. We also proposed LSP, which can be easily trained on a large-scale dataset without manual annotations and robustly estimates vehicle dimensions from their appearances. We showed that the learned size prior is superior to a fixed one and outperforms a similar 3D object detection model. FUMET can be leveraged as a drop-in training scheme for any monocular depth estimation method. We hope it will be a versatile tool for road scene understanding research." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "G.Kinoshita et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 480, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 480, + 165 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 480, + 165 + ], + "type": "text", + "content": "This work was in part supported by JSPS 20H05951 and 21H04893, JST JP-MJCR20G7 and JPMJAP2305, and RIKEN GRP." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "spans": [ + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 211, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 211, + 480, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 211, + 480, + 233 + ], + "spans": [ + { + "bbox": [ + 138, + 211, + 480, + 233 + ], + "type": "text", + "content": "1. Bhat, S.F., Alhashim, I., Wonka, P.: AdaBins: Depth Estimation Using Adaptive Bins. In: CVPR. pp. 4009-4018 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 235, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 235, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 235, + 480, + 266 + ], + "type": "text", + "content": "2. Casser, V., Pirk, S., Mahjourian, R., Angelova, A.: Unsupervised Monocular Depth and Ego-Motion Learning with Structure and Semantics. In: CVPRW. pp. 381-388 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 268, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 268, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 268, + 480, + 300 + ], + "type": "text", + "content": "3. Chawla, H., Varma, A., Arani, E., Zonooz, B.: Multimodal Scale Consistency and Awareness for Monocular Self-Supervised Depth Estimation. In: ICRA. pp. 5140-5146 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 301, + 480, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 301, + 480, + 333 + ], + "spans": [ + { + "bbox": [ + 138, + 301, + 480, + 333 + ], + "type": "text", + "content": "4. Chen, X., Kundu, K., Zhu, Y., Berneshawi, A.G., Ma, H., Fidler, S., Urtasun, R.: 3D Object Proposals for Accurate Object Class Detection. Adv. Neural Inform. Process. Syst. 28 (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 335, + 480, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 480, + 355 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 480, + 355 + ], + "type": "text", + "content": "5. Chen, X., Li, T.H., Zhang, R., Li, G.: Frequency-Aware Self-Supervised Monocular Depth Estimation. In: WACV. pp. 5808-5817 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 357, + 480, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 480, + 388 + ], + "type": "text", + "content": "6. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The Cityscapes Dataset for Semantic Urban Scene Understanding. In: CVPR (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 390, + 480, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 390, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 390, + 480, + 422 + ], + "type": "text", + "content": "7. Ding, M., Huo, Y., Yi, H., Wang, Z., Shi, J., Lu, Z., Luo, P.: Learning Depth-Guided Convolutions for Monocular 3D Object Detection. In: CVPR. pp. 11669-11678 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 423, + 480, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 480, + 455 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 480, + 455 + ], + "type": "text", + "content": "8. Eigen, D., Fergus, R.: Predicting Depth, Surface Normals and Semantic Labels with a Common Multi-Scale Convolutional Architecture. In: ICCV. pp. 3213-3223 (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 456, + 480, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 480, + 488 + ], + "type": "text", + "content": "9. Eigen, D., Puhrsch, C., Fergus, R.: Depth Map Prediction from a Single Image Using a Multi-Scale Deep Network. In: Adv. Neural Inform. Process. Syst. vol. 2, pp. 2366-2374 (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 490, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 490, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 490, + 480, + 521 + ], + "type": "text", + "content": "0. Frost, D., Prisacariu, V., Murray, D.: Recovering Stable Scale in Monocular SLAM Using Object-Supplemented Bundle Adjustment. IEEE Transactions on Robotics 34(3), 736-747 (2018). https://doi.org/10.1109/TR0.2018.2820722" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 522, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 480, + 544 + ], + "type": "text", + "content": "1. Garg, R., Bg, V.K., Carneiro, G., Reid, I.: Unsupervised CNN for Single View Depth Estimation: Geometry to the Rescue. In: ECCV. pp. 740-756 (2016)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 545, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 576 + ], + "type": "text", + "content": "2. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision Meets Robotics: The KITTI Dataset. International Journal of Robotics Research (IJRR) 32(11), 1231--1237 (2013). https://doi.org/10.1177/0278364913491297" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 578, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 620 + ], + "type": "text", + "content": "3. Geyer, J., Kassahun, Y., Mahmudi, M., Ricou, X., Durgesh, R., Chung, A.S., Hauswald, L., Pham, V.H., Mühlegg, M., Dorn, S., Fernandez, T., Janicke, M., Mirashi, S., Savani, C., Sturm, M., Vorobiov, O., Oelker, M., Garreis, S., Schuberth, P.: A2D2: Audi Autonomous Driving Dataset. CoRR abs/2004.06320 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "text", + "content": "4. Godard, C., Mac Aodha, O., Brostow, G.J.: Unsupervised Monocular Depth Estimation with Left-Right Consistency. In: CVPR. pp. 6602-6611 (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "text", + "content": "5. Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging Into Self-Supervised Monocular Depth Estimation. In: ICCV. pp. 3827-3837 (2019)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "text", + "content": "Unsupervised Training for Metric Monocular Road-Scene Depth Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "16. Guizilini, V., Ambrus, R., Pillai, S., Raventos, A., Gaidon, A.: 3D Packing for Self-Supervised Monocular Depth Estimation. In: CVPR. pp. 2482-2491 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "text", + "content": "17. Hartley, R., Zisserman, A.: Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edn. (2004)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "18. He, M., Hui, L., Bian, Y., Ren, J., Xie, J., Yang, J.: RA-Depth: Resolution Adaptive Self-supervised Monocular Depth Estimation. In: ECCV. pp. 565–581 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "type": "text", + "content": "19. Heylen, J., De Wolf, M., Dawagne, B., Proesmans, M., Van Gool, L., Abbeloos, W., Abdelkawy, H., Reino, D.O.: MonoCInIS: Camera Independent Monocular 3D Object Detection Using Instance Segmentation. In: ICCVW. pp. 923-934 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 480, + 237 + ], + "type": "text", + "content": "20. Hoiem, D., Efros, A.A., Hebert, M.: Putting Objects in Perspective. IJCV 80, 3-15 (2008). https://doi.org/10.1007/s11263-008-0137-5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 238, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 238, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 480, + 270 + ], + "type": "text", + "content": "21. Houston, J., Zuidhof, G., Bergamini, L., Ye, Y., Chen, L., Jain, A., Omari, S., Iglovikov, V., Ondruska, P.: One Thousand and One Hours: Self-Driving Motion Prediction Dataset. In: Conference on Robot Learning. pp. 409-418 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "text", + "content": "22. Jain, J., Li, J., Chiu, M.T., Hassani, A., Orlov, N., Shi, H.: OneFormer: One Transformer to Rule Universal Image Segmentation. In: CVPR. pp. 2989-2998 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "23. Kim, D., Ga, W., Ahn, P., Joo, D., Chun, S., Kim, J.: Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth. CoRR abs/2201.07436 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "24. Kingma, D.P., Ba, J.: Adam: A Method for Stochastic Optimization. CoRR abs/1412.6980 (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "text", + "content": "25. Lyu, X., Liu, L., Wang, M., Kong, X., Liu, L., Liu, Y., Chen, X., Yuan, Y.: HR-Depth: High Resolution Self-Supervised Monocular Depth Estimation. In: AAAI. vol. 35, pp. 2294-2301 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 391, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 423 + ], + "type": "text", + "content": "26. Milli, E., Erkent, O., Yilmaz, A.E.: Multi-Modal Multi-Task (3MT) Road Segmentation. IEEE Robotics and Automation Letters 8(9), 5408-5415 (2023). https://doi.org/10.1109/LRA.2023.3295254" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 445 + ], + "type": "text", + "content": "27. Ning, C., Gan, H.: Trap Attention: Monocular Depth Estimation with Manual Traps. In: CVPR. pp. 5033-5043 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 446, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 500 + ], + "type": "text", + "content": "28. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: PyTorch: An Imperative Style, High-Performance Deep Learning Library. In: Adv. Neural Inform. Process. Syst. pp. 8026--8037 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 522 + ], + "type": "text", + "content": "29. Piccinelli, L., Sakaridis, C., Yu, F.: iDisc: Internal Discretization for Monocular Depth Estimation. In: CVPR. pp. 21477-21487 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "text", + "content": "30. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet: A Large Scale Visual Recognition Challenge. IJCV 115, 211-252 (2015)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 587 + ], + "type": "text", + "content": "31. Sadat, A., Casas, S., Ren, M., Wu, X., Dhawan, P., Urtasun, R.: Perceive, Predict, and Plan: Safe Motion Planning Through Interpretable Semantic Representations. In: ECCV. pp. 414-430 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 588, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 610 + ], + "type": "text", + "content": "32. Schonberger, J.L., Frahm, J.M.: Structure-From-Motion Revisited. In: CVPR. pp. 4104-4113 (2016)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "33. Simonelli, A., Bulò, S.R., Porzi, L., Kontschieder, P., Ricci, E.: Are We Missing Confidence in Pseudo-LiDAR Methods for Monocular 3D Object Detection? In: ICCV. pp. 3225-3233 (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "34. Spencer, J., Russell, C., Hadfield, S., Bowden, R.: Kick Back & Relax: Learning to Reconstruct the World by Watching SlowTV (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 101 + ], + "type": "text", + "content": "G.Kinoshita et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "35. Sucar, E., Hayet, J.B.: Bayesian Scale Estimation for Monocular SLAM Based on Generic Object Detection for Correcting Scale Drift. In: ICRA. pp. 5152-5158 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "text", + "content": "36. Wagstaff, B., Kelly, J.: Self-Supervised Scale Recovery for Monocular Depth and Egomotion Estimation. In: IROS. pp. 2620-2627 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 194 + ], + "type": "text", + "content": "37. Wang, C., Buenaposada, J.M., Zhu, R., Lucey, S.: Learning Depth from Monocular Videos Using Direct Methods. In: CVPR. pp. 2022-2030 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "38. Wang, H., Cai, P., Fan, R., Sun, Y., Liu, M.: End-to-End Interactive Prediction and Planning with Optical Flow Distillation for Autonomous Driving. In: CVPRW. pp. 2229–2238 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 259 + ], + "type": "text", + "content": "39. Wang, Y., Chao, W.L., Garg, D., Hariharan, B., Campbell, M., Weinberger, K.Q.: Pseudo-LiDAR From Visual Depth Estimation: Bridging the Gap in 3D Object Detection for Autonomous Driving. In: CVPR. pp. 8437-8445 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 260, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 480, + 293 + ], + "type": "text", + "content": "40. Wang, Y., Chen, X., You, Y., Li, L.E., Hariharan, B., Campbell, M., Weinberger, K.Q., Chao, W.L.: Train in Germany, Test in the USA: Making 3D Object Detectors Generalize. In: CVPR. pp. 11710-11720 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "type": "text", + "content": "41. Wang, Y., Yang, H., Cai, J., Wang, G., Wang, J., Huang, Y.: Unsupervised Learning of Depth and Pose Based on Monocular Camera and Inertial Measurement Unit (IMU). In: ICRA. pp. 10010-10017 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "type": "text", + "content": "42. Wang, Z., Bovik, A., Sheikh, H., Simoncelli, E.: Image Quality Assessment: From Error Visibility to Structural Similarity. IEEE Transactions on Image Processing 13(4), 600-612 (2004). https://doi.org/10.1109/TIP.2003.819861" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "type": "text", + "content": "43. Watson, J., Mac Aodha, O., Prisacariu, V., Brostow, G., Firman, M.: The Temporal Opportunist: Self-Supervised Multi-Frame Monocular Depth. In: CVPR. pp. 1164–1174 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 435 + ], + "type": "text", + "content": "44. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., Ramanan, D., Carr, P., Hays, J.: Argoverse 2: Next Generation Datasets for Self-driving Perception and Forecasting. In: NeurIPS Datasets and Benchmarks (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 479 + ], + "type": "text", + "content": "45. Xiang, J., Wang, Y., An, L., Liu, H., Wang, Z., Liu, J.: Visual Attention-Based Self-Supervised Absolute Depth Estimation Using Geometric Priors in Autonomous Driving. IEEE Robotics and Automation Letters 7(4), 11998-12005 (2022). https://doi.org/10.1109/LRA.2022.3210298" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "text", + "content": "46. Yang, Z., Wang, P., Wang, Y., Xu, W., Nevatia, R.: LEGO: Learning Edge with Geometry All at Once by Watching Videos. In: CVPR. pp. 225-234 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "type": "text", + "content": "47. Yin, W., Zhang, C., Chen, H., Cai, Z., Yu, G., Wang, K., Chen, X., Shen, C.: Metric3D: Towards Zero-Shot Metric 3D Prediction from a Single Image. In: ICCV. pp. 9043-9053 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "text", + "content": "48. Yin, Z., Shi, J.: GeoNet: Unsupervised Learning of Dense Depth, Optical Flow and Camera Pose. In: CVPR. pp. 1983-1992 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "text", + "content": "49. You, Y., Wang, Y., Chao, W.L., Garg, D., Pleiss, G., Hariharan, B., Campbell, M., Weinberger, K.Q.: Pseudo-LiDAR++: Accurate Depth for 3D Object Detection in Autonomous Driving. In: ICLR (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 589, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 611 + ], + "type": "text", + "content": "50. Yuan, W., Gu, X., Dai, Z., Zhu, S., Tan, P.: Neural Window Fully-Connected CRFs for Monocular Depth Estimation. In: CVPR. pp. 3916-3925 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "51. Zeng, W., Luo, W., Suo, S., Sadat, A., Yang, B., Casas, S., Urtasun, R.: End-To-End Interpretable Neural Motion Planner. In: CVPR. pp. 8652-8661 (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "52. Zhang, N., Nex, F., Vosselman, G., Kerle, N.: Lite-Mono: A Lightweight CNN and Transformer Architecture for Self-Supervised Monocular Depth Estimation. In: CVPR. pp. 18537-18546 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 139, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 139, + 91, + 447, + 102 + ], + "type": "text", + "content": "Unsupervised Training for Metric Monocular Road-Scene Depth Estimation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 238 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "53. Zhang, S., Zhang, J., Tao, D.: Towards Scale-Aware, Robust, and Generalizable Unsupervised Monocular Depth Estimation by Integrating IMU Motion Dynamics. In: ECCV. pp. 143-160 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "text", + "content": "54. Zhang, S., Li, X., Liu, Y., Fu, H.: Scale-aware Insertion of Virtual Objects in Monocular Videos. In: ISMAR. pp. 36-44 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "55. Zhou, H., Greenwood, D., Taylor, S.: Self-Supervised Monocular Depth Estimation with Internal Feature Fusion. In: BMVC (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "text", + "content": "56. Zhou, T., Brown, M., Snavely, N., Lowe, D.G.: Unsupervised Learning of Depth and Ego-Motion from Video. In: CVPR. pp. 6612-6619 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 238 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 238 + ], + "type": "text", + "content": "57. Zhu, R., Yang, X., Hold-Geoffroy, Y., Perazzi, F., Eisenmann, J., Sunkavalli, K., Chandraker, M.: Single View Metrology in the Wild. In: ECCV. pp. 316-333 (2020)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 240, + 100 + ], + "type": "text", + "content": "G.Kinoshita et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Camera-LiDAR Cross-modality Gait Recognition/ce94d361-28a2-4f35-a27c-eca979dcca0e_content_list.json b/2024/Camera-LiDAR Cross-modality Gait Recognition/ce94d361-28a2-4f35-a27c-eca979dcca0e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e5af755c26cf32e35ec9a6e287d9c328f2c26f09 --- /dev/null +++ b/2024/Camera-LiDAR Cross-modality Gait Recognition/ce94d361-28a2-4f35-a27c-eca979dcca0e_content_list.json @@ -0,0 +1,1697 @@ +[ + { + "type": "text", + "text": "Camera-LiDAR Cross-modality Gait Recognition", + "text_level": 1, + "bbox": [ + 220, + 141, + 781, + 165 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenxuan Guo $^{1*}$ , Yingping Liang $^{2*}$ , Zhiyu Pan $^{1}$ , Ziheng Xi $^{1}$ , Jianjiang Feng $^{1\\dagger}$ , and Jie Zhou $^{1}$", + "bbox": [ + 223, + 189, + 779, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Department of Automation, BNRist, Tsinghua University", + "bbox": [ + 305, + 232, + 699, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Beijing Institute of Technology", + "bbox": [ + 392, + 247, + 611, + 261 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{gwx22, pzy20, xizh21}@mails.tsinghua.edu.cn liangyingping@bit.edu.cn", + "bbox": [ + 233, + 262, + 769, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jfeng, jzhou}@tsinghua.edu.cn", + "bbox": [ + 387, + 276, + 617, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Gait recognition is a crucial biometric identification technique. Camera-based gait recognition has been widely applied in both research and industrial fields. LiDAR-based gait recognition has also begun to evolve most recently, due to the provision of 3D structural information. However, in certain applications, cameras fail to recognize persons, such as in low-light environments and long-distance recognition scenarios, where LiDARs work well. On the other hand, the deployment cost and complexity of LiDAR systems limit its wider application. Therefore, it is essential to consider cross-modality gait recognition between cameras and LiDARs for a broader range of applications. In this work, we propose the first cross-modality gait recognition framework between Camera and LiDAR, namely CL-Gait. It employs a two-stream network for feature embedding of both modalities. This poses a challenging recognition task due to the inherent matching between 3D and 2D data, exhibiting significant modality discrepancy. To align the feature spaces of the two modalities, i.e., camera silhouettes and LiDAR points, we propose a contrastive pre-training strategy to mitigate modality discrepancy. To make up for the absence of paired camera-LiDAR data for pre-training, we also introduce a strategy for generating data on a large scale. This strategy utilizes monocular depth estimated from single RGB images and virtual cameras to generate pseudo point clouds for contrastive pre-training. Extensive experiments show that the cross-modality gait recognition is very challenging but still contains potential and feasibility with our proposed model and pre-training strategy. To the best of our knowledge, this is the first work to address cross-modality gait recognition. The code and dataset are available at https://github.com/GWxuan/CL-Gait.", + "bbox": [ + 259, + 330, + 743, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Gait recognition $\\cdot$ Cross-modality $\\cdot$ Contrastive pre-training", + "bbox": [ + 261, + 705, + 740, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 748, + 375, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gait recognition is a crucial long-range identification technology without physical contact, which has great advantages in privacy protection and cross-clothing", + "bbox": [ + 212, + 782, + 787, + 815 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. † Corresponding author.", + "bbox": [ + 217, + 824, + 532, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d4fdceced539d9b484f3e35912d492c854116eaa852e8b993eb3495d49d4afa7.jpg", + "image_caption": [ + "(a) Camera-based gait recognition (top), and LiDAR-based gait recognition (bottom)." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 493, + 268 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8bce0b8fd5723c2d202e639839223a620f367be28abaa338faaaa9b9012b1727.jpg", + "image_caption": [ + "(b) Our proposed cross-modality gait recognition between camera and LiDAR.", + "Fig. 1: Overview of single-modality and cross-modality gait recognition. Single-modality gait recognition takes data of one modality as input and searches within a gallery of the same data type. In contrast, cross-modality gait recognition processes two modalities and identify individuals within a gallery of a different modality." + ], + "image_footnote": [], + "bbox": [ + 504, + 143, + 787, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "recognition [2]. It identifies individuals based on behavioral characteristics extracted from walking sequences recorded by sensors. Gait recognition has been extensively applied in fields such as user identification, sports science, and public security. Most current gait recognition research employs cameras as sensors [29]. With the advancement of computer vision technology, camera-based gait recognition continues to achieve new progress [16-19, 22, 32].", + "bbox": [ + 212, + 398, + 782, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While utilizing cameras for gait recognition is cost-effective, it faces limitations in certain applications, such as low-light environments, long-distance recognition, and scenarios requiring precise 3D perception. In low-light conditions, the available information in RGB images significantly diminishes [9], making it challenging to accurately detect pedestrians and segment their silhouettes. Furthermore, due to the decline in image clarity, camera-based systems struggle to distinguish the fine-grained features from a distance [5]. Additionally, accurately capturing the 3D movement of individuals, crucial for comprehensive gait analysis, is beyond the capability of traditional camera systems [30].", + "bbox": [ + 212, + 491, + 784, + 627 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, to address the challenges under low light environment faced by RGB cameras, some researches [21, 23, 25, 40] have focused on RGB-Infrared cross-modality person re-identification (ReID), achieving commendable performance. However, infrared cameras struggle to capture pedestrians wearing insulating or thick clothing. Besides, they are constrained by limited detection distance and resolution, as well as susceptibility to lighting conditions and weather. It also faces the risk of pedestrian privacy and safety, compared with gait recognition. These limitations highlight the need for exploring alternative or complementary technologies that can offer more robust solutions for person identification tasks across a variety of environmental conditions.", + "bbox": [ + 212, + 628, + 785, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, LiDAR technology has been increasingly utilized [35] due to its impressive gains on 3D information. Beyond its mature applications in autonomous driving, LiDAR has also been applied in individual identification areas, such as gait recognition [30] and person re-identification [9] most recently.", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "W. Guo, Y. Liang et al.", + "bbox": [ + 271, + 114, + 431, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through point cloud data, LiDAR can provide precise 3D perception, aiding in the acquisition of an individual's 3D geometric structure. From this, intrinsic characteristics of a person, including height, body shape, and gait, can be extracted. Moreover, LiDARs possess long-distance perception capabilities, unaffected by lighting conditions and complex backgrounds. And it offers enhanced privacy protection than camera [9]. The robustness to environmental variables and the ability to capture detailed biometric data make it an invaluable tool for advanced surveillance and security systems. However, compared to camera systems, LiDAR's deployment complexity and cost are somehow higher, making it unsuitable for large-scale deployment in general scenarios.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Considering the respective advantages of cameras and LiDARs, switching sensors in different scenarios to explore cross-modality gait recognition presents a worthwhile task for investigation. Cross-modality gait recognition offers an option with a broader application scope. Cameras can be utilized in general scenes with normal lighting, while LiDAR can be applied in specific scenarios, offering a complementary tool that enhances overall recognition capabilities.", + "bbox": [ + 212, + 306, + 787, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we present the first study on camera-LiDAR cross-modality gait recognition as shown in Fig. 1. This task poses significant challenges due to the need for matching data between 3D and 2D modalities. Based on our observations, there are substantial differences between point clouds and images. To be specific, for gait recognition, point clouds primarily focus on the 3D positioning of body parts, whereas images capture precise contour information of individuals. To address these challenges and explore the feasibility of cross-modality gait recognition, we propose a novel cross-modality framework, namely CL-Gait.", + "bbox": [ + 212, + 406, + 787, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To extract gait features from both modalities, CL-Gait basically employs a two-stream network. Considering modality differences and data processing consistency, we first project and upsample the point clouds to obtain depth images with the same resolution as silhouettes, and use the projected depth images to train the network. Furthermore, to mitigate modality discrepancy, CL-Gait utilizes a contrastive silhouette-point pre-training approach (CSPP) to align the feature spaces of the two modalities. Pre-training requires aligned camera-LiDAR data under the same view, which is hard to obtain. To this end, we propose a new strategy of generating cross-modality gait data for pre-training. This strategy utilizes advances in monocular depth estimation and is able to generate large-scale dataset for pre-training using only single RGB images.", + "bbox": [ + 212, + 535, + 787, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Extensive experiments have revealed the following three insights: 1) CL-Gait achieves an average rank-1 accuracy of $54.21\\%$ . The results show at least $22.90\\%$ improvement over the baseline models and demonstrate the significant potential of cross-modality gait recognition. 2) Utilizing large-scale depth images generated from point clouds as input is proved superior to using point clouds directly. 3) The strategy of contrastive pre-training on large-scale generated paired data mitigates the modality differences, contributing to performance improvement.", + "bbox": [ + 212, + 710, + 787, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, our main contributions are as follows:", + "bbox": [ + 238, + 825, + 624, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "CL-Gait", + "bbox": [ + 673, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, this is the first work on cross-modality gait recognition. Extensive experiments show the potential of utilizing camera and LiDAR for gait recognition in challenging cross-modality scenarios.", + "- We analyse several network structures and conduct comparable experiments on their effectiveness for cross-modality gait recognition. Based on the two-stream network with better performance, our proposed CL-Gait is capable of matching camera silhouettes and LiDAR points.", + "- We propose a contrastive pre-training method to align the feature spaces of the two modalities. To make up for the absence of paired camera-LiDAR data, we further introduce a large-scale data generation strategy." + ], + "bbox": [ + 225, + 146, + 784, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 321, + 387, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Gait Recognition. Based on the sensor used, exiting gait recognition methods can be categorized into camera-based and LiDAR-based methods.", + "bbox": [ + 212, + 354, + 782, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Camera-based gait recognition has been extensively studied over the past decade. The majority of camera-based methods focus on extracting appearance features directly from images or videos [2, 6], adapting well to resolution reduction and achieving impressive performance [5]. These methods utilize silhouettes [16, 18, 19] or other gait templates [1, 34] for spatial feature extraction and temporal modeling. Additionally, some researchers have explored using estimated underlying structure of human body [15, 17, 32], such as 2D/3D pose and the SMPL model [22], as inputs. While theoretically robust against factors like carrying and clothing, these models often struggle at low resolutions, limiting their practicality in some real-world scenarios [5].", + "bbox": [ + 212, + 386, + 784, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LiDAR-based gait recognition is an emerging field that utilizes precise 3D representations, point clouds, capturing complex motion patterns and the 3D structure of individual gaits. This approach is less susceptible to variations in lighting, clothing, and background, offering promising avenues for accurate gait recognition in diverse conditions. LidarGait [30] stands as a pioneering work in this field, introducing a multimodal gait recognition dataset named SUSTECH1K. This dataset is designed for evaluating the performance of gait recognition based on various sensors and has demonstrated the potential and practicality of LiDAR-based gait recognition. Additionally, some researchers have applied LidAR to the task of person re-identification [9], achieving impressive results.", + "bbox": [ + 212, + 537, + 785, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Camera-based and LiDAR-based methods have been proven practical, each with its unique advantages and disadvantages. Therefore, providing a compromise solution by addressing cross-modality gait recognition represents a valuable research direction, which has not been studied currently.", + "bbox": [ + 212, + 688, + 785, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Contrastive Pre-training. Contrastive pre-training is primarily applied in the field of vision-language models, aimed at aligning the feature spaces of visual and linguistic models. In recent years, inspired by the success of self-supervised learning within intra-modal tasks, researchers have begun to explore pre-training objectives for tasks involving multiple modalities, such as vision and language [37]. The pioneering work by CLIP [27] performs cross-modality", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "W. Guo, Y. Liang et al.", + "bbox": [ + 271, + 113, + 431, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "contrastive pre-training on hundreds of millions of image-text pairs. CLIP [27] can generate a task-agnostic model that achieves surprisingly effective results. Subsequently, ALIGN [14] expands upon CLIP by utilizing a noisy dataset that covers more than a billion image-text pairs, further extending the capabilities of cross-modality pre-training. Contrastive pre-training plays a crucial role in cross-modality tasks. Therefore, we introduce contrastive pre-training into camera-LiDAR cross-modality gait recognition.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "RGB-IR Cross-modality Person ReID. Visible-infrared person ReID addresses the challenge of matching persons across different modalities, specifically between RGB and infrared cameras [36]. This task is particularly significant for scenarios with poor illumination, notably at night. With two primary datasets, SYSU-MM01 [36] and RegDB [24], RGB-IR cross-modality person ReID has been extensively studied, most of which focusing on metric learning, feature learning, and adversarial learning approaches. Metric learning [7, 10, 21, 40] and feature learning [25, 39, 41] methods aim to extract and align multimodal features into a common space for effective comparison, employing strategies like angle-based measurement and Euclidean constraints to bridge the modality gap. Notably, advancements in this area include the development of novel loss functions [21]. Concurrently, Generative Adversarial Networks (GANs) [3, 4, 23] have been instrumental in cross-modality ReID, enabling the transformation of images between RGB and IR while maintaining identity information, and fostering feature learning through adversarial and disentanglement strategies. Despite their potential, GAN-based methods face challenges such as high computational demands and the occasional production of low-quality images that may negatively impact ReID performance.", + "bbox": [ + 212, + 252, + 787, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While RGB-IR person ReID mainly relies on appearance for distinction, which is a short-term feature and may raise privacy concerns. In contrast, cross-modality gait recognition focuses on shape and gait of individuals, which are time-invariant features, and offers better protection of privacy.", + "bbox": [ + 212, + 523, + 785, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 604, + 330, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this work, we propose CL-Gait for cross-modality gait recognition between camera and LiDAR. CL-Gait employs a two-stream network for cross-modality feature embedding. The network utilizes modality-specific modules in the shallow layers and shared modules in the deeper layers, as illustrated in Fig. 2. Besides, CL-Gait adopts a contrastive learning strategy to align the feature spaces of the two modalities, mitigating modality differences, as shown in Fig. 3. The backbone obtained from pre-training is used to initialize the cross-modality embedding network. To facilitate pre-training on large-scale data, we also propose a method for generating cross-modality gait data, as shown in Fig. 4.", + "bbox": [ + 212, + 632, + 787, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Cross-modality Network of CL-Gait", + "text_level": 1, + "bbox": [ + 214, + 787, + 558, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Two-Stream Backbone. The CL-Gait framework processes two distinct data modalities: 2D silhouette sequences and 3D point cloud sequences, optimizing", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "CL-Gait", + "bbox": [ + 673, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e9666d6768e7918a5e7b504bc5974f00b5b163fa618e56ec24222bea9720ade3.jpg", + "image_caption": [ + "Fig. 2: The cross-modality network of our proposed CL-Gait. It employs a two-stream architecture that encodes sequences from two modalities into a consistent feature space. HP stands for horizontal pooling, and TP represents temporal pooling." + ], + "image_footnote": [], + "bbox": [ + 230, + 143, + 774, + 314 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cross-modality feature extracting. Inspired by [30], we first project and interpolate point clouds into depth images. For each modality, we apply modality-specific convolution layers (Conv-S for silhouettes and Conv-P for depth images generated from point clouds) before projecting the features into a shared embedding space, as shown in Fig. 2. Following [36], the two-stream backbone is modified from ResNet-18, with its first layer altered to be the modality-specific convolution layers, and the subsequent layers serving as shared convolution layers. For silhouettes, we define the processing as follows:", + "bbox": [ + 212, + 398, + 787, + 518 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {I} = L _ {S} (I), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 455, + 529, + 785, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $I$ represents the silhouette and $L_{S}$ denotes the silhouette-specific convolution layer, which produces the feature map $F_{I}$ .", + "bbox": [ + 212, + 555, + 784, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Considering the modality differences and the need for consistency in data processing, we first project and interpolate point clouds to generate depths with the same resolution as silhouettes. And the processing is as follows:", + "bbox": [ + 212, + 585, + 785, + 631 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {P} = L _ {P} (\\operatorname {P I} (P)), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 641, + 784, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $P$ denotes the point cloud, PI represents the projection and interpolation process, and $L_{P}$ is the depth-specific convolution layer, which produces the feature map $F_{P}$ for the point cloud.", + "bbox": [ + 212, + 667, + 784, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After extracting modality-specific features $F_{P}$ and $F_{S}$ , we aim to unify these features within a consistent feature space. To achieve this, we employ a shared convolutional layer, denoted as $L_{\\mathrm{sh}}$ , to process both $F_{P}$ and $F_{S}$ :", + "bbox": [ + 212, + 712, + 785, + 758 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF ^ {\\prime} = L _ {\\mathrm {s h}} \\left(F _ {m}\\right), \\text {w h e r e} F _ {m} = F _ {P} \\text {o r} F _ {S}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 768, + 785, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $F^{\\prime}$ is the output feature map of $L_{\\mathrm{sh}}$", + "bbox": [ + 212, + 794, + 527, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Temporal and Spatial Pooling. In our network, we follow commonly used strategies in gait recognition for temporal and spatial aggregation, specifically", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "W. Guo, Y. Liang et al.", + "bbox": [ + 271, + 114, + 431, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "through temporal and horizontal pooling [5]. These techniques are essential for effectively summarizing and interpreting the dynamic and spatial aspects of gait data. Horizontal pooling divides the input feature map horizontally into $N_{H}$ parts, with each part being aggregated into a single feature vector:", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {H P}: \\mathbb {R} ^ {T \\times H \\times W \\times C} \\rightarrow \\mathbb {R} ^ {T \\times N _ {H} \\times C}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 217, + 785, + 234 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $T$ denotes the sequence length, $H$ and $W$ represent the height and width of the feature map, respectively, and $C$ indicates the dimension of the features.", + "bbox": [ + 212, + 244, + 782, + 275 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Temporal pooling allows for the aggregation of features from variable-length sequences, enabling the model to capture the final sequence-level representation:", + "bbox": [ + 212, + 276, + 782, + 306 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {T P}: \\mathbb {R} ^ {T \\times N _ {H} \\times C} \\rightarrow \\mathbb {R} ^ {N _ {H} \\times C}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 320, + 785, + 337 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Therefore, we obtain $N_{H}$ feature vectors, which are further mapped to the metric space using $N_{H}$ independent Multi-Layer Perceptrons. These mapped vectors are then concatenated to form the final sequence-level feature embedding. Training and Inference. The cross-modality network of CL-Gait is trained with cross-modality triplet loss and cross-entropy loss. We have modified the triplet loss to accommodate cross-modality feature learning, where each modality is alternately used as the anchor, while the corresponding positive and negative samples are selected from the other modality, formulated as:", + "bbox": [ + 212, + 343, + 784, + 463 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c r o s s - t r i p l e t}} = \\frac {1}{2} \\left(\\mathcal {L} _ {\\text {t r i p l e t}} \\left(P, I _ {\\text {p o s}}, I _ {\\text {n e g}}\\right) + \\mathcal {L} _ {\\text {t r i p l e t}} \\left(I, P _ {\\text {p o s}}, P _ {\\text {n e g}}\\right)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 281, + 474, + 785, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The training loss is a combination of both losses with a hyperparameter $\\gamma$ :", + "bbox": [ + 214, + 512, + 751, + 527 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {c r o s s - t r i p l e t}} + \\gamma \\mathcal {L} _ {\\mathrm {c e}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 541, + 784, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{ce}}$ is the cross-entropy loss. During inference, the similarity between samples from different modalities is measured using the Euclidean distance.", + "bbox": [ + 212, + 566, + 782, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Contrastive Silhouette-point Pre-training", + "text_level": 1, + "bbox": [ + 214, + 619, + 602, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on our observations in gait recognition tasks, the significant modality difference between 3D point clouds and 2D images could be a critical factor affecting model performance. To be specific, point clouds focus more on the 3D positioning of body parts, whereas images concentrate on the contour information of the individual. It's crucial to establish connections between the distinct information focused on by each modality. Inspired by CLIP [27], we propose a contrastive silhouette-point pre-training (CSPP) strategy to align the feature spaces of both modalities in the convolution-based encoders, as shown in Fig. 3.", + "bbox": [ + 212, + 643, + 784, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The pre-training process does not require identity labels from the samples for supervision. Trained on paired single-view data from aligned camera and LiDAR, the pre-training could make the model focus on learning a robust representation that bridges the gap between the modalities without direct identity-based guidance, and enhance the performance of the cross-modality network.", + "bbox": [ + 212, + 765, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "CL-Gait", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e512b73a4f3d63db19719fb49c70ff0b30d4faa444d5ddae1defc963ad4533a6.jpg", + "image_caption": [ + "Fig. 3: The contrastive silhouette-point pre-training (CSPP) approach of CL-Gait. Paired silhouettes and point clouds are taken as inputs, and the backbone of the cross-modality network is pre-trained with contrastive learning loss to align the feature spaces of the two modalities. The pre-training process does not require identity labels of the samples for supervision." + ], + "image_footnote": [], + "bbox": [ + 236, + 143, + 769, + 281 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pre-training Process. For pre-training, the input shifts from sequence data to numerous pairs of point clouds and silhouettes. These pairs are defined as data collected from the same individual at the same moment, encompassing both modalities. It ensures that the network learns to reconcile the inherent differences between the 3D and 2D representations, focusing on the alignment of features that accurately reflect the same gait patterns across modalities.", + "bbox": [ + 212, + 388, + 782, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The paired data is processed by the two-stream backbone and spatial pooling operations to obtain paired feature embeddings, i.e., $\\mathcal{S} = \\{S_i | i = 1, 2, \\dots, N\\}$ for silhouettes and $\\mathcal{P} = \\{P_i | i = 1, 2, \\dots, N\\}$ for point clouds. They are then utilized to compute the loss for contrastive representation learning.", + "bbox": [ + 212, + 479, + 784, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Contrastive Learning loss. The cornerstone of our pre-training is the contrastive learning loss [27]. Given the paired feature embeddings $S = \\{S_{i} | i = 1, 2, \\dots, N\\}$ and $\\mathcal{P} = \\{P_{i} | i = 1, 2, \\dots, N\\}$ , there are $N \\times N$ possible (silhouettes, point clouds) pairings. CL-Gait is trained to learn a multi-modality embedding space by maximizing the cosine similarity of the feature embeddings of the $N$ real pairs while minimizing the cosine similarity of the embeddings of the $N^2 - N$ incorrect pairings. The cosine similarity matrix is calculated as follows:", + "bbox": [ + 212, + 539, + 787, + 645 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} = N _ {\\mathrm {L} 2} (\\mathcal {S}) \\times N _ {\\mathrm {L} 2} (\\mathcal {P}) ^ {T}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 651, + 784, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{S} \\in \\mathbb{R}^{N \\times D}$ , $\\mathcal{P} \\in \\mathbb{R}^{N \\times D}$ , $\\mathcal{M} \\in \\mathbb{R}^{N \\times N}$ , $N_{\\mathrm{L2}}(\\cdot)$ denotes L2 normalization, and $(\\cdot)^T$ denotes the transpose of the matrix. Then, a symmetric cross entropy loss over these similarity scores is optimized:", + "bbox": [ + 212, + 676, + 787, + 722 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n}} = \\frac {1}{2} \\left(\\mathcal {L} _ {\\mathrm {c e}} (\\mathcal {M}, G) + \\mathcal {L} _ {\\mathrm {c e}} \\left(\\mathcal {M} ^ {T}, G\\right)\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 729, + 784, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $G = \\{1,2,\\dots,N\\}$ . $G$ represents the labels used for calculating the cross-entropy loss, indicating that pairs corresponding to the similarities on the diagonal of matrices $\\mathcal{M}$ and $\\mathcal{M}^T$ are considered positive samples, while all others are treated as negative samples. This approach ensures that the network learns to distinguish between matching and non-matching modality pairs effectively.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "W. Guo, Y. Liang et al.", + "bbox": [ + 271, + 114, + 431, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/16e78f6ce5709cbd6ca4f6dc3707c941667c4b0581186fbdfdfc3d3793683077.jpg", + "image_caption": [ + "Fig. 4: Illustration of paired gait data generation from single RGB images for contrastive pre-training. The quality of the synthesized data is comparable to the real-world data, making it possible to synthesize large-scale pre-training data." + ], + "image_footnote": [], + "bbox": [ + 228, + 141, + 772, + 305 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In the process described above, each frame of point cloud or silhouette, after passing through the pre-trained network, yields a feature embedding of size $(1,C)$ , indicating global pooling is applied to the output feature map of the backbone. For practical application and to align each local feature embedding, horizontal pooling is used to obtain features of size $(N,C)$ for each frame. During the computation of the contrastive learning loss, pairs of features from the same individual, moment, and part across two modalities are considered positive samples, with all others being negative. Through our practice, we have found that adopting a strategy of aligning local features during the pre-training phase is more suitable for gait recognition tasks, leading to performance improvements.", + "bbox": [ + 212, + 385, + 787, + 537 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.3 Generation of Pre-training Gait Data", + "text_level": 1, + "bbox": [ + 214, + 556, + 571, + 571 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Pre-training on real and large-scale data is challenging due to the high cost of acquiring paired RGB and point cloud data. To address this issue, we propose a method for generating pseudo data based on monocular depth estimation. As illustrated in Fig. 4, we use Depth Anything [38] to estimate dense depth $D \\in \\mathbb{R}^{H \\times W}$ from large-scale single RGB images. Then, utilizing a virtual camera with intrinsic $K$ , the estimated depth is mapped to pseudo point clouds:", + "bbox": [ + 214, + 579, + 787, + 670 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n[ P ^ {\\prime}, 1 ] ^ {T} = K ^ {- 1} D [ u, v, 1 ] ^ {T}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 405, + 676, + 785, + 696 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $P' \\in \\mathbb{R}^{H \\times W \\times 3}$ is the point coordinates, $K^{-1}$ indicates the inverse of intrinsic $K$ , and $u, v$ are the image pixel coordinates. The pseudo points $P'$ , after being downsampled by voxel grid into $P_{down}'$ , can be used for contrastive pretraining between camera silhouettes and LiDAR point clouds. To be specific, the downsampled points $P_{down}'$ are projected back into the image pixel coordinate to obtain the depth images for contrastive pre-training. Compared to point clouds, RGB images are less costly to collect and easier to acquire, with many public datasets of pedestrian image already available. Our proposed method makes it feasible to synthesize large-scale data for contrastive pre-training.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "CL-Gait", + "bbox": [ + 673, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 162 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 215, + 176, + 333, + 191 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The SUSTech1K Dataset [30] is the first large-scale LiDAR-based gait recognition dataset collected by a LiDAR sensor and an RGB camera. The dataset contains 25,239 sequences from 1,050 subjects and covers many variations, including visibility, views, occlusions, clothing, carrying, and scenes. It is times-tamped frames for each modality of frames. The first single-modality 3D gait recognition framework is trained on SUSTech1K using point clouds. In our study, we explore new uses for the SUSTech1K dataset. Our approach involves both images and point clouds, which marks the first attempt of cross-modality gait recognition with this dataset. We also introduce a novel pre-training strategy based on paired images and point clouds from the dataset. This strategy utilizes contrastive learning for cross-modality matching between images and point clouds and improves the performance of our proposed cross-modality gait method.", + "bbox": [ + 212, + 202, + 785, + 383 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The HITSZ-VCM Dataset and LIP Dataset are large-scale datasets, which contains a large amount of person videos/images captured by cameras. In detail, the HITSZ-VCM dataset [20] contains 927 valid identities with image sequences. The LIP dataset [8] contains about 50,000 identities but with only one image for each. Totally, there are 251,452 RGB images in the HITSZ-VCM dataset and 50,000 in the LIP dataset. We implement the pre-training strategy on these datasets containing only RGB images, to explore the impact of pre-training with out-of-domain data on this task. Given that there is no point cloud data paired with these images, we use monocular depth estimation to generate pseudo point clouds, as shown in Fig. 4. This greatly expands the amount of data available for pre-training at scale. This also improves the accuracy of cross-modality recognition between 2D and 3D spaces, which marks the effectiveness of our proposed contrastive pre-training strategy.", + "bbox": [ + 212, + 383, + 785, + 580 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 603, + 426, + 618 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details. Following [30], all the camera-based silhouettes and LiDAR-based depth images are aligned and then resized into the resolution of 6464. The total number of iterations is set to 120,000 for pre-training and 60,000 for fine-tuning. The Adam optimizer is used to prevent the issue of gradient vanishing because of low-quality silhouettes. The triplet and cross-entropy loss weights are set to the same. All comparison methods are trained using the same training strategy as LidarGait [30]. The OpenGait codebase [5] is utilized to conduct all experiments.", + "bbox": [ + 212, + 628, + 785, + 748 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Protocol. All experiments are conducted on SUSTech1K. Following LidarGait [30], the dataset is divided into three splits: a training set with 250 identities and 6,011 sequences, a validation set with 6,025 sequences from 250 unseen identities, and a test set with the remaining 550 identities and 13,203 sequences. The SUSTech1K dataset provides gait sequences from multiple viewpoints. Thus, we also adopt the cross-view evaluation protocol [31,42] used in", + "bbox": [ + 212, + 750, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "W. Guo, Y. Liang et al.", + "bbox": [ + 271, + 114, + 431, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "CASIA-B and OUMVLP. During the test, the sequences in normal conditions are grouped into gallery sets, and the sequences in variant conditions are taken as probe sets. To accurately assess the performance of camera-LiDAR cross-modality gait recognition, we evaluate the results using point cloud data as the probe and silhouette data as the gallery, i.e., LiDAR to Camera (L to C). We also evaluate the results using silhouette data as the probe and point cloud data as the gallery, i.e., Camera to LiDAR (C to L).", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Comparative Methods", + "text_level": 1, + "bbox": [ + 215, + 289, + 446, + 305 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We evaluate several commonly used structures for cross-modality modeling as baselines, which basically follows the setting of [36]. The baseline models include one-steam structure, asymmetric FC layer structure, image-point structure, and the proposed two-stream structure. We apply residual block in ResNet-18 [11] as the base convolution block for all the four structures. For them, the loss function are weighted cross-entropy loss and triplet loss [2], which is commonly used and relatively stable. And all of the hyper parameters are kept the same. As for the input of one-steam structure, asymmetric FC layer structure and two-stream structure, the point clouds are converted to three-channel colored depth images and adjusted to the resolution of $64 \\times 64$ with zero padding. We introduce the four structures as following:", + "bbox": [ + 212, + 329, + 787, + 496 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "One-stream Structure is the most commonly used in vision tasks, where all parameters are shared in the whole network. Therefore, we organize silhouettes and depth images obtained from point clouds into the same dimensions before feeding them into the network.", + "bbox": [ + 212, + 498, + 787, + 559 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Asymmetric FC Layer Structure is used in multi-domain tasks, for example, IDR [12] for VIS-NIR face recognition. This structure shares nearly all parameters except for the last FC layer. It assumes that the feature extraction for different modalities can be same and adaptation is achieved in the backbone.", + "bbox": [ + 212, + 561, + 787, + 623 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Image-point Structure directly uses point clouds as input, rather than projected depth images. It employs two types of encoders: an image encoder for the silhouette input and a point cloud encoder for the point cloud input. Under this structure, we implement three commonly used point cloud encoders: GCN [9], PointNet [26], and Point-Transformer [43].", + "bbox": [ + 212, + 625, + 787, + 702 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Two-stream Structure is commonly used in cross-modality matching tasks. It utilizes modality-specific modules in the shallow layers of the network and uniform modules with shared parameters in the deeper layers. Based on the two-stream structure, our proposed CL-Gait is equipped with a horizontal pooling module (HP) and a contrastive silhouette-point pre-training strategy (CSPP). To show the effectiveness of the HP module and CSPP strategy of CL-Gait, we implement three additional networks: 1) a two-stream backbone without CSPP, using global pooling instead of HP; 2) a two-stream backbone without CSPP, using HP; 3) a two-stream backbone with CSPP, using global pooling.", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "CL-Gait", + "bbox": [ + 673, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/cd055f6af92f16b244dffb120d4ed43e8dbe3475d569e723b5f478a452476639.jpg", + "table_caption": [ + "Table 1: Evaluation with different structures on SUSTech1K valid + test set. We use ResNet-18 to extract image features by default. \"L to C\" indicates the results with point clouds as probe and silhouettes as gallery, and \"C to L\" indicates the reverse." + ], + "table_footnote": [ + "i.e., ResNet + CSPP + HP." + ], + "table_body": "
StructureModelL to CC to L
Rank-1Rank-3Rank-5Rank-1Rank-3Rank-5
One-streamResNet2.605.617.9914.9425.9231.87
Asymmetric FCResNet3.087.2310.5216.4329.1335.90
Image-pointGCN [9]12.2025.0133.2314.3028.3836.62
PointNet [26]7.1116.6123.358.4019.2526.90
Point-Transformer [43]13.2526.7235.2915.4230.0438.70
Two-streamResNet30.3948.7757.3135.2654.4262.84
ResNet+CSPP32.1850.8559.2236.5455.7564.11
ResNet+HP45.4162.8969.7142.3059.6767.12
CLGait (ours)*53.2969.5475.5955.1271.2377.31
", + "bbox": [ + 220, + 199, + 777, + 334 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Comparative Results", + "text_level": 1, + "bbox": [ + 215, + 378, + 434, + 395 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Following the cross-view and cross-modality evaluation protocol, we report the comparative results of the methods above in the Tab. 1. From the results, we can obtain the following observations: 1) CL-Gait demonstrates its superiority to all existing structures and methods, primarily due to the feature extraction capabilities of the two-stream structure. Additionally, the use of pre-training effectively mitigates modality discrepancy, further improving its performance. 2) Experiments under the two-stream structure can be considered as an ablation study, highlighting the important roles of CSPP and HP. Adding HP to the ResNet base improves the average rank-1 by approximately $11\\%$ , and further incorporating CSPP increases the average rank-1 by about another $10\\%$ . 3) The performance of the one-stream structure is poor, primarily because the one-stream structure uses an identical network to process both silhouettes and point clouds, failing to address the modality discrepancy between them. 4) Although the Asymmetric FC structure employs different FC layers for the two modalities, its performance remains poor. This indicates that modality-specific adaptations are best applied at the shallower levels of the network. Furthermore, during the training process, the asymmetric FC layers tend to cause the network to overfit quickly. 5) All methods utilizing the two-stream structure outperform those based on the image-point structure. This likely results from the depth images, obtained through projection, having data consistency and smaller modality discrepancy with silhouettes compared to point cloud inputs.", + "bbox": [ + 212, + 402, + 787, + 720 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Discussion", + "text_level": 1, + "bbox": [ + 215, + 739, + 349, + 755 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "3D Geometry Information. To evaluate the effectiveness of depth information for cross-modality gait recognition, we compare another type of data as input for the second modality, e.g., LiDAR silhouettes. LiDAR silhouettes are obtained by range-view projection of point cloud sets and contains no 3D geometry information. The results are shown in Fig. 5. When 3D geometry information is not", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "W. Guo, Y. Liang et al.", + "bbox": [ + 271, + 114, + 431, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8b2dbdb56ab34ac339b7fc996661c87b045cb2b68bce392dd2a453f4c5df972d.jpg", + "image_caption": [ + "Fig. 5: Comparison of different input forms of LiDAR data. The results show that the projected and interpolated depth from point cloud works best for the cross-modality matching. This indicates that the 3D geometry information are essential. For each input form, the best-performing model is utilized." + ], + "image_footnote": [], + "bbox": [ + 243, + 140, + 759, + 263 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "included, the performance of LiDAR silhouettes is much lower. This is because the camera has a much higher resolution than LiDAR, so the silhouettes from the camera can have more details. Integrating depth information can improve average rank-1 accuracy from $40.59\\%$ to $54.21\\%$ , validating the necessity and effectiveness of 3D information for cross-modality gait recognition. One important reason may be that the 3D information contained in the point cloud can compensate for the difference in viewing angles to some extent. Furthermore, we include the performance of direct point cloud input in Fig. 5 to compare with depth images and illustrate the input form of point clouds.", + "bbox": [ + 212, + 364, + 787, + 500 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Contrastive Pre-training Datasets. To make up for the absence of paired camera-LiDAR data, we further propose a large-scale data generation strategy. Camera silhouette and LiDAR point data are generated from real-world RGB images for contrastive pre-training on such paired pseudo data. To valid its performance, we construct training sets from two different datasets, including LIP and HITSZ-VCM. These two datasets are commonly used for camera-based person segmentation and re-identification and contain single still RGB images. To ensure fair comparisons, we also conduct experiments on SUSTech1K training set, which contains real paired point clouds and camera silhouettes.", + "bbox": [ + 212, + 503, + 787, + 638 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As shown in Tab. 2, when the amount of data is similar using data generated from the LIP dataset, the performance of pre-training on generated pseudo data is comparable to that of real data. When pre-trained on a larger scale of synthetic pseudo data generated from the HITSZ-VSM dataset, the accuracy from LiDAR to camera improves and exceeds that on the real data, proving the effectiveness of our data generation strategy for contrastive pre-training. However, the accuracy from camera to LiDAR decreases a little, probably due to the difference between generated point clouds from estimated depth and real points from LiDAR.", + "bbox": [ + 212, + 641, + 787, + 762 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Temporal Modules. In order to aggregate the different frames in the sequence, temporal modules are used to fuse the features of multiple frames. We evaluate the effectiveness of different temporal modules in combination with our proposed CL-Gait, as shown in Tab. 3. We can observe that the simple temporal pooling structure demonstrates superiority over the other learning-based structure. This", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "CL-Gait", + "bbox": [ + 673, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/10d7fc2f656a9996a34c332a4e8896919fe9173842f6ae485f895db0f3f2e0ec.jpg", + "table_caption": [ + "Table 2: Performance with different pre-training datasets. Using estimated depth for pre-training obtains close results compared with using real paired point-image dataset." + ], + "table_footnote": [], + "table_body": "
DatasetData AmountL to CC to L
Rank-1Rank-3Rank-5Rank-1Rank-3Rank-5
SUSTech1K120K51.6068.2274.4356.1272.1577.95
LIP50K51.0068.0973.9955.0771.5977.59
HITSZ-VCM770K53.2969.5475.5955.1271.2377.31
", + "bbox": [ + 274, + 185, + 718, + 252 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/9153d7be42e932b0e84d71af87090bf37c69c1ba177210bf9095947c0acc82a4.jpg", + "table_caption": [ + "Table 3: Performance of CL-Gait with different temporal modules." + ], + "table_footnote": [], + "table_body": "
FeatureL to CC to L
Rank-1Rank-3Rank-5Rank-1Rank-3Rank-5
Temporal Pooling [5]53.2969.5475.5955.1271.2377.31
LSTM [13]37.4055.0160.8939.4756.4162.72
Bi-LSTM [28]39.6355.5561.8339.8857.8263.71
Transformer [33]44.8061.3167.3647.1663.8369.11
", + "bbox": [ + 281, + 291, + 712, + 368 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "result indicates that temporal pooling can better fuse multi-frame features to match domain information more effectively. One reason may be that learning-based approaches such as LSTM and Transformer are more likely to overfit.", + "bbox": [ + 212, + 393, + 785, + 441 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion and Future work", + "text_level": 1, + "bbox": [ + 215, + 459, + 527, + 476 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper presents the first research on cross-modality gait recognition with point clouds and silhouettes from RGB images. Firstly, We propose a cross-modality gait recognition framework, named CL-Gait, that utilizes a two-stream network for feature embedding of different modalities, i.e., camera and LiDAR. Moreover, to mitigate modality discrepancy, we propose a contrastive pre-training strategy along with a large-scale data generation method. It can generate large number of data pairs (silhouette and point cloud) based on monocular depth estimation and pre-train the model in a contrastive manner. Extensive experiments demonstrate that CL-Gait achieves commendable performance in cross-modality evaluation modes, proving the importance of maintaining modality consistency and the effectiveness of contrastive pre-training. This also highlights the potential of cross-modality gait recognition.", + "bbox": [ + 212, + 488, + 787, + 669 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "CL-Gait has shown remarkable results, but we believe there remains significant room for improvement. The modality and resolution differences between point clouds and silhouettes, along with the inherent motion blur in point clouds, may be key factors impacting model performance. Therefore, densifying point clouds or developing point cloud encoders specifically for cross-modality gait recognition could enhance model performance.", + "bbox": [ + 212, + 670, + 787, + 761 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 215, + 781, + 393, + 797 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work is supported by the National Key Research and Development Program of China under Grant (2018AAA0102803).", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "W. Guo, Y. Liang et al.", + "bbox": [ + 271, + 114, + 431, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bobick, A.F., Davis, J.W.: The recognition of human movement using temporal templates. IEEE Transactions on Pattern Analysis and Machine Intelligence 23(3), 257-267 (2001)", + "2. Chao, H., He, Y., Zhang, J., Feng, J.: Gaitset: Regarding gait as a set for cross-view gait recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 8126-8133 (2019)", + "3. Choi, S., Lee, S., Kim, Y., Kim, T., Kim, C.: Hi-cmd: Hierarchical cross-modality disentanglement for visible-infrared person re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10257-10266 (2020)", + "4. Dai, P., Ji, R., Wang, H., Wu, Q., Huang, Y.: Cross-modality person re-identification with generative adversarial training. In: IJCAI. vol. 1, p. 6 (2018)", + "5. Fan, C., Liang, J., Shen, C., Hou, S., Huang, Y., Yu, S.: Opengait: Revisiting gait recognition towards better practicality. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9707-9716 (2023)", + "6. Fan, C., Peng, Y., Cao, C., Liu, X., Hou, S., Chi, J., Huang, Y., Li, Q., He, Z.: Gaitpart: Temporal part-based model for gait recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14225-14233 (2020)", + "7. Feng, Z., Lai, J., Xie, X.: Learning modality-specific representations for visible-infrared person re-identification. IEEE Transactions on Image Processing 29, 579-590 (2019)", + "8. Gong, K., Liang, X., Zhang, D., Shen, X., Lin, L.: Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 932-940 (2017)", + "9. Guo, W., Pan, Z., Liang, Y., Xi, Z., Zhong, Z.C., Feng, J., el al: Lidar-based person re-identification. arXiv preprint arXiv:2312.03033 (2023)", + "0. Hao, Y., Wang, N., Li, J., Gao, X.: Hsme: Hypersphere manifold embedding for visible thermal person re-identification. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 8385-8392 (2019)", + "1. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016)", + "2. He, R., Wu, X., Sun, Z., Tan, T.: Learning invariant deep representation for nir-vis face recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 31 (2017)", + "3. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural computation 9(8), 1735-1780 (1997)", + "4. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., el al: Scaling up visual and vision-language representation learning with noisy text supervision. In: International Conference on Machine Learning. pp. 4904-4916. PMLR (2021)", + "5. Li, X., Makihara, Y., Xu, C., Yagi, Y., Yu, S., Ren, M.: End-to-end model-based gait recognition. In: Proceedings of the Asian Conference on Computer Vision (2020)", + "6. Liang, J., Fan, C., Hou, S., Shen, C., Huang, Y., Yu, S.: Gaitedge: Beyond plain end-to-end gait recognition for better practicality. In: European Conference on Computer Vision. pp. 375-390. Springer (2022)" + ], + "bbox": [ + 225, + 175, + 785, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "CL-Gait", + "bbox": [ + 673, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Liao, R., Yu, S., An, W., Huang, Y.: A model-based gait recognition method with body pose and human prior knowledge. Pattern Recognition 98, 107069 (2020)", + "18. Lin, B., Liu, Y., Zhang, S.: Gaitmask: Mask-based model for gait recognition. In: BMVC. pp. 1-12 (2021)", + "19. Lin, B., Zhang, S., Wang, M., Li, L., Yu, X.: Caitgl: Learning discriminative global-local feature representations for gait recognition. arXiv preprint arXiv:2208.01380 (2022)", + "20. Lin, X., Li, J., Ma, Z., Li, H., Li, S., Xu, K., Lu, G., Zhang, D.: Learning modal-invariant and temporal-memory for video-based visible-infrared person re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20973–20982 (2022)", + "21. Ling, Y., Zhong, Z., Luo, Z., Rota, P., Li, S., Sebe, N.: Class-aware modality mix and center-guided metric learning for visible-thermal person re-identification. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 889-897 (2020)", + "22. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Spl: A skinned multi-person linear model. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 851-866 (2023)", + "23. Lu, Y., Wu, Y., Liu, B., Zhang, T., Li, B., Chu, Q., Yu, N.: Cross-modality person re-identification with shared-specific feature transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13379–13389 (2020)", + "24. Nguyen, D.T., Hong, H.G., Kim, K.W., Park, K.R.: Person recognition system based on a combination of body images from visible light and thermal cameras. Sensors 17(3), 605 (2017)", + "25. Park, H., Lee, S., Lee, J., Ham, B.: Learning by aligning: Visible-infrared person re-identification using cross-modal correspondences. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12046–12055 (2021)", + "26. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3D classification and segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 652-660 (2017)", + "27. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)", + "28. Schuster, M., Paliwal, K.K.: Bidirectional recurrent neural networks. IEEE Transactions on Signal Processing 45(11), 2673-2681 (1997)", + "29. Sepas-Moghaddam, A., Etemad, A.: Deep gait recognition: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(1), 264-284 (2022)", + "30. Shen, C., Fan, C., Wu, W., Wang, R., Huang, G.Q., Yu, S.: Lidargait: Benchmarking 3D gait recognition with point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1054-1063 (2023)", + "31. Takemura, N., Makihara, Y., Muramatsu, D., Echigo, T., Yagi, Y.: Multi-view large population gait dataset and its performance evaluation for cross-view gait recognition. IPSJ Transactions on Computer Vision and Applications 10, 1-14 (2018)", + "32. Teepe, T., Khan, A., Gilg, J., Herzog, F., Hörmann, S., Rigoll, G.: Gaitgraph: Graph convolutional network for skeleton-based gait recognition. In: 2021 IEEE International Conference on Image Processing (ICIP). pp. 2314-2318. IEEE (2021)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "W. Guo, Y. Liang et al.", + "bbox": [ + 271, + 114, + 431, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in Neural Information Processing Systems 30 (2017)", + "34. Wang, C., Zhang, J., Pu, J., Yuan, X., Wang, L.: Chrono-gait image: A novel temporal template for gait recognition. In: Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part I 11. pp. 257-270. Springer (2010)", + "35. Wang, Y., Chao, W.L., Garg, D., Hariharan, B., Campbell, M., Weinberger, K.Q.: Pseudo-lidar from visual depth estimation: Bridging the gap in 3D object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8445-8453 (2019)", + "36. Wu, A., Zheng, W.S., Yu, H.X., Gong, S., Lai, J.: RGB-infrared cross-modality person re-identification. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 5380-5389 (2017)", + "37. Yang, J., Duan, J., Tran, S., Xu, Y., Chanda, S., Chen, L., Zeng, B., Chilimbi, T., Huang, J.: Vision-language pre-training with triple contrastive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15671-15680 (2022)", + "38. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. arXiv preprint arXiv:2401.10891 (2024)", + "39. Ye, M., Lan, X., Leng, Q.: Modality-aware collaborative learning for visible thermal person re-identification. In: Proceedings of the 27th ACM International Conference on Multimedia. pp. 347-355 (2019)", + "40. Ye, M., Lan, X., Leng, Q., Shen, J.: Cross-modality person re-identification via modality-aware collaborative ensemble learning. IEEE Transactions on Image Processing 29, 9387–9399 (2020)", + "41. Ye, M., Shen, J., J. Crandall, D., Shao, L., Luo, J.: Dynamic dual-attentive aggregation learning for visible-infrared person re-identification. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVII 16. pp. 229–247. Springer (2020)", + "42. Yu, S., Tan, D., Tan, T.: A framework for evaluating the effect of view angle, clothing and carrying condition on gait recognition. In: 18th International Conference on Pattern Recognition (ICPR'06). vol. 4, pp. 441-444. IEEE (2006)", + "43. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16259-16268 (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 659 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "CL-Gait", + "bbox": [ + 673, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Camera-LiDAR Cross-modality Gait Recognition/ce94d361-28a2-4f35-a27c-eca979dcca0e_model.json b/2024/Camera-LiDAR Cross-modality Gait Recognition/ce94d361-28a2-4f35-a27c-eca979dcca0e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8311d6910d542c84826e460742aee203ac6735f1 --- /dev/null +++ b/2024/Camera-LiDAR Cross-modality Gait Recognition/ce94d361-28a2-4f35-a27c-eca979dcca0e_model.json @@ -0,0 +1,2236 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.221, + 0.142, + 0.782, + 0.166 + ], + "angle": 0, + "content": "Camera-LiDAR Cross-modality Gait Recognition" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.19, + 0.78, + 0.222 + ], + "angle": 0, + "content": "Wenxuan Guo\\(^{1*}\\), Yingping Liang\\(^{2*}\\), Zhiyu Pan\\(^{1}\\), Ziheng Xi\\(^{1}\\), Jianjiang Feng\\(^{1\\dagger}\\), and Jie Zhou\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.233, + 0.7, + 0.248 + ], + "angle": 0, + "content": "1 Department of Automation, BNRist, Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.393, + 0.248, + 0.612, + 0.262 + ], + "angle": 0, + "content": "2 Beijing Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.263, + 0.77, + 0.276 + ], + "angle": 0, + "content": "{gwx22, pzy20, xizh21}@mails.tsinghua.edu.cn liangyingping@bit.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.388, + 0.277, + 0.618, + 0.289 + ], + "angle": 0, + "content": "{jfeng, jzhou}@tsinghua.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.332, + 0.744, + 0.693 + ], + "angle": 0, + "content": "Abstract. Gait recognition is a crucial biometric identification technique. Camera-based gait recognition has been widely applied in both research and industrial fields. LiDAR-based gait recognition has also begun to evolve most recently, due to the provision of 3D structural information. However, in certain applications, cameras fail to recognize persons, such as in low-light environments and long-distance recognition scenarios, where LiDARs work well. On the other hand, the deployment cost and complexity of LiDAR systems limit its wider application. Therefore, it is essential to consider cross-modality gait recognition between cameras and LiDARs for a broader range of applications. In this work, we propose the first cross-modality gait recognition framework between Camera and LiDAR, namely CL-Gait. It employs a two-stream network for feature embedding of both modalities. This poses a challenging recognition task due to the inherent matching between 3D and 2D data, exhibiting significant modality discrepancy. To align the feature spaces of the two modalities, i.e., camera silhouettes and LiDAR points, we propose a contrastive pre-training strategy to mitigate modality discrepancy. To make up for the absence of paired camera-LiDAR data for pre-training, we also introduce a strategy for generating data on a large scale. This strategy utilizes monocular depth estimated from single RGB images and virtual cameras to generate pseudo point clouds for contrastive pre-training. Extensive experiments show that the cross-modality gait recognition is very challenging but still contains potential and feasibility with our proposed model and pre-training strategy. To the best of our knowledge, this is the first work to address cross-modality gait recognition. The code and dataset are available at https://github.com/GWxuan/CL-Gait." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.706, + 0.741, + 0.721 + ], + "angle": 0, + "content": "Keywords: Gait recognition \\(\\cdot\\) Cross-modality \\(\\cdot\\) Contrastive pre-training" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.749, + 0.376, + 0.765 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.784, + 0.788, + 0.816 + ], + "angle": 0, + "content": "Gait recognition is a crucial long-range identification technology without physical contact, which has great advantages in privacy protection and cross-clothing" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.533, + 0.841 + ], + "angle": 0, + "content": "* Equal contribution. † Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.432, + 0.129 + ], + "angle": 0, + "content": "W. Guo, Y. Liang et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.495, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.275, + 0.493, + 0.297 + ], + "angle": 0, + "content": "(a) Camera-based gait recognition (top), and LiDAR-based gait recognition (bottom)." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.144, + 0.788, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.275, + 0.784, + 0.296 + ], + "angle": 0, + "content": "(b) Our proposed cross-modality gait recognition between camera and LiDAR." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.37 + ], + "angle": 0, + "content": "Fig. 1: Overview of single-modality and cross-modality gait recognition. Single-modality gait recognition takes data of one modality as input and searches within a gallery of the same data type. In contrast, cross-modality gait recognition processes two modalities and identify individuals within a gallery of a different modality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.784, + 0.492 + ], + "angle": 0, + "content": "recognition [2]. It identifies individuals based on behavioral characteristics extracted from walking sequences recorded by sensors. Gait recognition has been extensively applied in fields such as user identification, sports science, and public security. Most current gait recognition research employs cameras as sensors [29]. With the advancement of computer vision technology, camera-based gait recognition continues to achieve new progress [16-19, 22, 32]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.785, + 0.628 + ], + "angle": 0, + "content": "While utilizing cameras for gait recognition is cost-effective, it faces limitations in certain applications, such as low-light environments, long-distance recognition, and scenarios requiring precise 3D perception. In low-light conditions, the available information in RGB images significantly diminishes [9], making it challenging to accurately detect pedestrians and segment their silhouettes. Furthermore, due to the decline in image clarity, camera-based systems struggle to distinguish the fine-grained features from a distance [5]. Additionally, accurately capturing the 3D movement of individuals, crucial for comprehensive gait analysis, is beyond the capability of traditional camera systems [30]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.786, + 0.779 + ], + "angle": 0, + "content": "Recently, to address the challenges under low light environment faced by RGB cameras, some researches [21, 23, 25, 40] have focused on RGB-Infrared cross-modality person re-identification (ReID), achieving commendable performance. However, infrared cameras struggle to capture pedestrians wearing insulating or thick clothing. Besides, they are constrained by limited detection distance and resolution, as well as susceptibility to lighting conditions and weather. It also faces the risk of pedestrian privacy and safety, compared with gait recognition. These limitations highlight the need for exploring alternative or complementary technologies that can offer more robust solutions for person identification tasks across a variety of environmental conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In recent years, LiDAR technology has been increasingly utilized [35] due to its impressive gains on 3D information. Beyond its mature applications in autonomous driving, LiDAR has also been applied in individual identification areas, such as gait recognition [30] and person re-identification [9] most recently." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.674, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CL-Gait" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "Through point cloud data, LiDAR can provide precise 3D perception, aiding in the acquisition of an individual's 3D geometric structure. From this, intrinsic characteristics of a person, including height, body shape, and gait, can be extracted. Moreover, LiDARs possess long-distance perception capabilities, unaffected by lighting conditions and complex backgrounds. And it offers enhanced privacy protection than camera [9]. The robustness to environmental variables and the ability to capture detailed biometric data make it an invaluable tool for advanced surveillance and security systems. However, compared to camera systems, LiDAR's deployment complexity and cost are somehow higher, making it unsuitable for large-scale deployment in general scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.788, + 0.399 + ], + "angle": 0, + "content": "Considering the respective advantages of cameras and LiDARs, switching sensors in different scenarios to explore cross-modality gait recognition presents a worthwhile task for investigation. Cross-modality gait recognition offers an option with a broader application scope. Cameras can be utilized in general scenes with normal lighting, while LiDAR can be applied in specific scenarios, offering a complementary tool that enhances overall recognition capabilities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.788, + 0.528 + ], + "angle": 0, + "content": "In this work, we present the first study on camera-LiDAR cross-modality gait recognition as shown in Fig. 1. This task poses significant challenges due to the need for matching data between 3D and 2D modalities. Based on our observations, there are substantial differences between point clouds and images. To be specific, for gait recognition, point clouds primarily focus on the 3D positioning of body parts, whereas images capture precise contour information of individuals. To address these challenges and explore the feasibility of cross-modality gait recognition, we propose a novel cross-modality framework, namely CL-Gait." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.703 + ], + "angle": 0, + "content": "To extract gait features from both modalities, CL-Gait basically employs a two-stream network. Considering modality differences and data processing consistency, we first project and upsample the point clouds to obtain depth images with the same resolution as silhouettes, and use the projected depth images to train the network. Furthermore, to mitigate modality discrepancy, CL-Gait utilizes a contrastive silhouette-point pre-training approach (CSPP) to align the feature spaces of the two modalities. Pre-training requires aligned camera-LiDAR data under the same view, which is hard to obtain. To this end, we propose a new strategy of generating cross-modality gait data for pre-training. This strategy utilizes advances in monocular depth estimation and is able to generate large-scale dataset for pre-training using only single RGB images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.711, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Extensive experiments have revealed the following three insights: 1) CL-Gait achieves an average rank-1 accuracy of \\(54.21\\%\\). The results show at least \\(22.90\\%\\) improvement over the baseline models and demonstrate the significant potential of cross-modality gait recognition. 2) Utilizing large-scale depth images generated from point clouds as input is proved superior to using point clouds directly. 3) The strategy of contrastive pre-training on large-scale generated paired data mitigates the modality differences, contributing to performance improvement." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.826, + 0.625, + 0.84 + ], + "angle": 0, + "content": "To summarize, our main contributions are as follows:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.114, + 0.432, + 0.129 + ], + "angle": 0, + "content": "W. Guo, Y. Liang et al." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "- To the best of our knowledge, this is the first work on cross-modality gait recognition. Extensive experiments show the potential of utilizing camera and LiDAR for gait recognition in challenging cross-modality scenarios." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.193, + 0.785, + 0.252 + ], + "angle": 0, + "content": "- We analyse several network structures and conduct comparable experiments on their effectiveness for cross-modality gait recognition. Based on the two-stream network with better performance, our proposed CL-Gait is capable of matching camera silhouettes and LiDAR points." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.254, + 0.785, + 0.298 + ], + "angle": 0, + "content": "- We propose a contrastive pre-training method to align the feature spaces of the two modalities. To make up for the absence of paired camera-LiDAR data, we further introduce a large-scale data generation strategy." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.147, + 0.785, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.323, + 0.388, + 0.339 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.356, + 0.784, + 0.385 + ], + "angle": 0, + "content": "Gait Recognition. Based on the sensor used, exiting gait recognition methods can be categorized into camera-based and LiDAR-based methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.785, + 0.538 + ], + "angle": 0, + "content": "Camera-based gait recognition has been extensively studied over the past decade. The majority of camera-based methods focus on extracting appearance features directly from images or videos [2, 6], adapting well to resolution reduction and achieving impressive performance [5]. These methods utilize silhouettes [16, 18, 19] or other gait templates [1, 34] for spatial feature extraction and temporal modeling. Additionally, some researchers have explored using estimated underlying structure of human body [15, 17, 32], such as 2D/3D pose and the SMPL model [22], as inputs. While theoretically robust against factors like carrying and clothing, these models often struggle at low resolutions, limiting their practicality in some real-world scenarios [5]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.786, + 0.688 + ], + "angle": 0, + "content": "LiDAR-based gait recognition is an emerging field that utilizes precise 3D representations, point clouds, capturing complex motion patterns and the 3D structure of individual gaits. This approach is less susceptible to variations in lighting, clothing, and background, offering promising avenues for accurate gait recognition in diverse conditions. LidarGait [30] stands as a pioneering work in this field, introducing a multimodal gait recognition dataset named SUSTECH1K. This dataset is designed for evaluating the performance of gait recognition based on various sensors and has demonstrated the potential and practicality of LiDAR-based gait recognition. Additionally, some researchers have applied LidAR to the task of person re-identification [9], achieving impressive results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.786, + 0.749 + ], + "angle": 0, + "content": "Camera-based and LiDAR-based methods have been proven practical, each with its unique advantages and disadvantages. Therefore, providing a compromise solution by addressing cross-modality gait recognition represents a valuable research direction, which has not been studied currently." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Contrastive Pre-training. Contrastive pre-training is primarily applied in the field of vision-language models, aimed at aligning the feature spaces of visual and linguistic models. In recent years, inspired by the success of self-supervised learning within intra-modal tasks, researchers have begun to explore pre-training objectives for tasks involving multiple modalities, such as vision and language [37]. The pioneering work by CLIP [27] performs cross-modality" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.674, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CL-Gait" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "contrastive pre-training on hundreds of millions of image-text pairs. CLIP [27] can generate a task-agnostic model that achieves surprisingly effective results. Subsequently, ALIGN [14] expands upon CLIP by utilizing a noisy dataset that covers more than a billion image-text pairs, further extending the capabilities of cross-modality pre-training. Contrastive pre-training plays a crucial role in cross-modality tasks. Therefore, we introduce contrastive pre-training into camera-LiDAR cross-modality gait recognition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.525 + ], + "angle": 0, + "content": "RGB-IR Cross-modality Person ReID. Visible-infrared person ReID addresses the challenge of matching persons across different modalities, specifically between RGB and infrared cameras [36]. This task is particularly significant for scenarios with poor illumination, notably at night. With two primary datasets, SYSU-MM01 [36] and RegDB [24], RGB-IR cross-modality person ReID has been extensively studied, most of which focusing on metric learning, feature learning, and adversarial learning approaches. Metric learning [7, 10, 21, 40] and feature learning [25, 39, 41] methods aim to extract and align multimodal features into a common space for effective comparison, employing strategies like angle-based measurement and Euclidean constraints to bridge the modality gap. Notably, advancements in this area include the development of novel loss functions [21]. Concurrently, Generative Adversarial Networks (GANs) [3, 4, 23] have been instrumental in cross-modality ReID, enabling the transformation of images between RGB and IR while maintaining identity information, and fostering feature learning through adversarial and disentanglement strategies. Despite their potential, GAN-based methods face challenges such as high computational demands and the occasional production of low-quality images that may negatively impact ReID performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.525, + 0.787, + 0.586 + ], + "angle": 0, + "content": "While RGB-IR person ReID mainly relies on appearance for distinction, which is a short-term feature and may raise privacy concerns. In contrast, cross-modality gait recognition focuses on shape and gait of individuals, which are time-invariant features, and offers better protection of privacy." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.605, + 0.331, + 0.621 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.77 + ], + "angle": 0, + "content": "In this work, we propose CL-Gait for cross-modality gait recognition between camera and LiDAR. CL-Gait employs a two-stream network for cross-modality feature embedding. The network utilizes modality-specific modules in the shallow layers and shared modules in the deeper layers, as illustrated in Fig. 2. Besides, CL-Gait adopts a contrastive learning strategy to align the feature spaces of the two modalities, mitigating modality differences, as shown in Fig. 3. The backbone obtained from pre-training is used to initialize the cross-modality embedding network. To facilitate pre-training on large-scale data, we also propose a method for generating cross-modality gait data, as shown in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.788, + 0.56, + 0.804 + ], + "angle": 0, + "content": "3.1 Cross-modality Network of CL-Gait" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Two-Stream Backbone. The CL-Gait framework processes two distinct data modalities: 2D silhouette sequences and 3D point cloud sequences, optimizing" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.432, + 0.129 + ], + "angle": 0, + "content": "W. Guo, Y. Liang et al." + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.144, + 0.775, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.328, + 0.788, + 0.371 + ], + "angle": 0, + "content": "Fig. 2: The cross-modality network of our proposed CL-Gait. It employs a two-stream architecture that encodes sequences from two modalities into a consistent feature space. HP stands for horizontal pooling, and TP represents temporal pooling." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.789, + 0.52 + ], + "angle": 0, + "content": "cross-modality feature extracting. Inspired by [30], we first project and interpolate point clouds into depth images. For each modality, we apply modality-specific convolution layers (Conv-S for silhouettes and Conv-P for depth images generated from point clouds) before projecting the features into a shared embedding space, as shown in Fig. 2. Following [36], the two-stream backbone is modified from ResNet-18, with its first layer altered to be the modality-specific convolution layers, and the subsequent layers serving as shared convolution layers. For silhouettes, we define the processing as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.456, + 0.53, + 0.787, + 0.547 + ], + "angle": 0, + "content": "\\[\nF _ {I} = L _ {S} (I), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.785, + 0.586 + ], + "angle": 0, + "content": "where \\(I\\) represents the silhouette and \\(L_{S}\\) denotes the silhouette-specific convolution layer, which produces the feature map \\(F_{I}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.586, + 0.787, + 0.632 + ], + "angle": 0, + "content": "Considering the modality differences and the need for consistency in data processing, we first project and interpolate point clouds to generate depths with the same resolution as silhouettes. And the processing is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.436, + 0.642, + 0.785, + 0.658 + ], + "angle": 0, + "content": "\\[\nF _ {P} = L _ {P} (\\operatorname {P I} (P)), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.668, + 0.785, + 0.713 + ], + "angle": 0, + "content": "where \\( P \\) denotes the point cloud, PI represents the projection and interpolation process, and \\( L_{P} \\) is the depth-specific convolution layer, which produces the feature map \\( F_{P} \\) for the point cloud." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.713, + 0.787, + 0.759 + ], + "angle": 0, + "content": "After extracting modality-specific features \\( F_{P} \\) and \\( F_{S} \\), we aim to unify these features within a consistent feature space. To achieve this, we employ a shared convolutional layer, denoted as \\( L_{\\mathrm{sh}} \\), to process both \\( F_{P} \\) and \\( F_{S} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.769, + 0.787, + 0.786 + ], + "angle": 0, + "content": "\\[\nF ^ {\\prime} = L _ {\\mathrm {s h}} \\left(F _ {m}\\right), \\text {w h e r e} F _ {m} = F _ {P} \\text {o r} F _ {S}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.528, + 0.81 + ], + "angle": 0, + "content": "where \\(F^{\\prime}\\) is the output feature map of \\(L_{\\mathrm{sh}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Temporal and Spatial Pooling. In our network, we follow commonly used strategies in gait recognition for temporal and spatial aggregation, specifically" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CL-Gait" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "through temporal and horizontal pooling [5]. These techniques are essential for effectively summarizing and interpreting the dynamic and spatial aspects of gait data. Horizontal pooling divides the input feature map horizontally into \\( N_{H} \\) parts, with each part being aggregated into a single feature vector:" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.218, + 0.786, + 0.235 + ], + "angle": 0, + "content": "\\[\n\\mathrm {H P}: \\mathbb {R} ^ {T \\times H \\times W \\times C} \\rightarrow \\mathbb {R} ^ {T \\times N _ {H} \\times C}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.246, + 0.784, + 0.276 + ], + "angle": 0, + "content": "where \\( T \\) denotes the sequence length, \\( H \\) and \\( W \\) represent the height and width of the feature map, respectively, and \\( C \\) indicates the dimension of the features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.277, + 0.784, + 0.307 + ], + "angle": 0, + "content": "Temporal pooling allows for the aggregation of features from variable-length sequences, enabling the model to capture the final sequence-level representation:" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.321, + 0.786, + 0.338 + ], + "angle": 0, + "content": "\\[\n\\mathrm {T P}: \\mathbb {R} ^ {T \\times N _ {H} \\times C} \\rightarrow \\mathbb {R} ^ {N _ {H} \\times C}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.785, + 0.464 + ], + "angle": 0, + "content": "Therefore, we obtain \\( N_{H} \\) feature vectors, which are further mapped to the metric space using \\( N_{H} \\) independent Multi-Layer Perceptrons. These mapped vectors are then concatenated to form the final sequence-level feature embedding. Training and Inference. The cross-modality network of CL-Gait is trained with cross-modality triplet loss and cross-entropy loss. We have modified the triplet loss to accommodate cross-modality feature learning, where each modality is alternately used as the anchor, while the corresponding positive and negative samples are selected from the other modality, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.282, + 0.476, + 0.786, + 0.504 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c r o s s - t r i p l e t}} = \\frac {1}{2} \\left(\\mathcal {L} _ {\\text {t r i p l e t}} \\left(P, I _ {\\text {p o s}}, I _ {\\text {n e g}}\\right) + \\mathcal {L} _ {\\text {t r i p l e t}} \\left(I, P _ {\\text {p o s}}, P _ {\\text {n e g}}\\right)\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.513, + 0.753, + 0.528 + ], + "angle": 0, + "content": "The training loss is a combination of both losses with a hyperparameter \\(\\gamma\\):" + }, + { + "type": "equation", + "bbox": [ + 0.418, + 0.542, + 0.785, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {c r o s s - t r i p l e t}} + \\gamma \\mathcal {L} _ {\\mathrm {c e}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.784, + 0.597 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{ce}}\\) is the cross-entropy loss. During inference, the similarity between samples from different modalities is measured using the Euclidean distance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.62, + 0.603, + 0.636 + ], + "angle": 0, + "content": "3.2 Contrastive Silhouette-point Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.785, + 0.764 + ], + "angle": 0, + "content": "Based on our observations in gait recognition tasks, the significant modality difference between 3D point clouds and 2D images could be a critical factor affecting model performance. To be specific, point clouds focus more on the 3D positioning of body parts, whereas images concentrate on the contour information of the individual. It's crucial to establish connections between the distinct information focused on by each modality. Inspired by CLIP [27], we propose a contrastive silhouette-point pre-training (CSPP) strategy to align the feature spaces of both modalities in the convolution-based encoders, as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.784, + 0.84 + ], + "angle": 0, + "content": "The pre-training process does not require identity labels from the samples for supervision. Trained on paired single-view data from aligned camera and LiDAR, the pre-training could make the model focus on learning a robust representation that bridges the gap between the modalities without direct identity-based guidance, and enhance the performance of the cross-modality network." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.432, + 0.129 + ], + "angle": 0, + "content": "W. Guo, Y. Liang et al." + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.144, + 0.77, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.292, + 0.788, + 0.361 + ], + "angle": 0, + "content": "Fig. 3: The contrastive silhouette-point pre-training (CSPP) approach of CL-Gait. Paired silhouettes and point clouds are taken as inputs, and the backbone of the cross-modality network is pre-trained with contrastive learning loss to align the feature spaces of the two modalities. The pre-training process does not require identity labels of the samples for supervision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.784, + 0.478 + ], + "angle": 0, + "content": "Pre-training Process. For pre-training, the input shifts from sequence data to numerous pairs of point clouds and silhouettes. These pairs are defined as data collected from the same individual at the same moment, encompassing both modalities. It ensures that the network learns to reconcile the inherent differences between the 3D and 2D representations, focusing on the alignment of features that accurately reflect the same gait patterns across modalities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.785, + 0.539 + ], + "angle": 0, + "content": "The paired data is processed by the two-stream backbone and spatial pooling operations to obtain paired feature embeddings, i.e., \\(\\mathcal{S} = \\{S_i | i = 1, 2, \\dots, N\\}\\) for silhouettes and \\(\\mathcal{P} = \\{P_i | i = 1, 2, \\dots, N\\}\\) for point clouds. They are then utilized to compute the loss for contrastive representation learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.788, + 0.646 + ], + "angle": 0, + "content": "Contrastive Learning loss. The cornerstone of our pre-training is the contrastive learning loss [27]. Given the paired feature embeddings \\( S = \\{S_{i} | i = 1, 2, \\dots, N\\} \\) and \\( \\mathcal{P} = \\{P_{i} | i = 1, 2, \\dots, N\\} \\), there are \\( N \\times N \\) possible (silhouettes, point clouds) pairings. CL-Gait is trained to learn a multi-modality embedding space by maximizing the cosine similarity of the feature embeddings of the \\( N \\) real pairs while minimizing the cosine similarity of the embeddings of the \\( N^2 - N \\) incorrect pairings. The cosine similarity matrix is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.652, + 0.785, + 0.671 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} = N _ {\\mathrm {L} 2} (\\mathcal {S}) \\times N _ {\\mathrm {L} 2} (\\mathcal {P}) ^ {T}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.677, + 0.788, + 0.723 + ], + "angle": 0, + "content": "where \\(\\mathcal{S} \\in \\mathbb{R}^{N \\times D}\\), \\(\\mathcal{P} \\in \\mathbb{R}^{N \\times D}\\), \\(\\mathcal{M} \\in \\mathbb{R}^{N \\times N}\\), \\(N_{\\mathrm{L2}}(\\cdot)\\) denotes L2 normalization, and \\((\\cdot)^T\\) denotes the transpose of the matrix. Then, a symmetric cross entropy loss over these similarity scores is optimized:" + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.73, + 0.785, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n}} = \\frac {1}{2} \\left(\\mathcal {L} _ {\\mathrm {c e}} (\\mathcal {M}, G) + \\mathcal {L} _ {\\mathrm {c e}} \\left(\\mathcal {M} ^ {T}, G\\right)\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\( G = \\{1,2,\\dots,N\\} \\). \\( G \\) represents the labels used for calculating the cross-entropy loss, indicating that pairs corresponding to the similarities on the diagonal of matrices \\( \\mathcal{M} \\) and \\( \\mathcal{M}^T \\) are considered positive samples, while all others are treated as negative samples. This approach ensures that the network learns to distinguish between matching and non-matching modality pairs effectively." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.674, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CL-Gait" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.142, + 0.773, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.316, + 0.785, + 0.358 + ], + "angle": 0, + "content": "Fig. 4: Illustration of paired gait data generation from single RGB images for contrastive pre-training. The quality of the synthesized data is comparable to the real-world data, making it possible to synthesize large-scale pre-training data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.788, + 0.538 + ], + "angle": 0, + "content": "In the process described above, each frame of point cloud or silhouette, after passing through the pre-trained network, yields a feature embedding of size \\((1,C)\\), indicating global pooling is applied to the output feature map of the backbone. For practical application and to align each local feature embedding, horizontal pooling is used to obtain features of size \\((N,C)\\) for each frame. During the computation of the contrastive learning loss, pairs of features from the same individual, moment, and part across two modalities are considered positive samples, with all others being negative. Through our practice, we have found that adopting a strategy of aligning local features during the pre-training phase is more suitable for gait recognition tasks, leading to performance improvements." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.557, + 0.572, + 0.573 + ], + "angle": 0, + "content": "3.3 Generation of Pre-training Gait Data" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.58, + 0.788, + 0.671 + ], + "angle": 0, + "content": "Pre-training on real and large-scale data is challenging due to the high cost of acquiring paired RGB and point cloud data. To address this issue, we propose a method for generating pseudo data based on monocular depth estimation. As illustrated in Fig. 4, we use Depth Anything [38] to estimate dense depth \\( D \\in \\mathbb{R}^{H \\times W} \\) from large-scale single RGB images. Then, utilizing a virtual camera with intrinsic \\( K \\), the estimated depth is mapped to pseudo point clouds:" + }, + { + "type": "equation", + "bbox": [ + 0.406, + 0.677, + 0.786, + 0.697 + ], + "angle": 0, + "content": "\\[\n[ P ^ {\\prime}, 1 ] ^ {T} = K ^ {- 1} D [ u, v, 1 ] ^ {T}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\( P' \\in \\mathbb{R}^{H \\times W \\times 3} \\) is the point coordinates, \\( K^{-1} \\) indicates the inverse of intrinsic \\( K \\), and \\( u, v \\) are the image pixel coordinates. The pseudo points \\( P' \\), after being downsampled by voxel grid into \\( P_{down}' \\), can be used for contrastive pretraining between camera silhouettes and LiDAR point clouds. To be specific, the downsampled points \\( P_{down}' \\) are projected back into the image pixel coordinate to obtain the depth images for contrastive pre-training. Compared to point clouds, RGB images are less costly to collect and easier to acquire, with many public datasets of pedestrian image already available. Our proposed method makes it feasible to synthesize large-scale data for contrastive pre-training." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.432, + 0.129 + ], + "angle": 0, + "content": "W. Guo, Y. Liang et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.377, + 0.163 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.178, + 0.334, + 0.192 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.203, + 0.787, + 0.384 + ], + "angle": 0, + "content": "The SUSTech1K Dataset [30] is the first large-scale LiDAR-based gait recognition dataset collected by a LiDAR sensor and an RGB camera. The dataset contains 25,239 sequences from 1,050 subjects and covers many variations, including visibility, views, occlusions, clothing, carrying, and scenes. It is times-tamped frames for each modality of frames. The first single-modality 3D gait recognition framework is trained on SUSTech1K using point clouds. In our study, we explore new uses for the SUSTech1K dataset. Our approach involves both images and point clouds, which marks the first attempt of cross-modality gait recognition with this dataset. We also introduce a novel pre-training strategy based on paired images and point clouds from the dataset. This strategy utilizes contrastive learning for cross-modality matching between images and point clouds and improves the performance of our proposed cross-modality gait method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.787, + 0.581 + ], + "angle": 0, + "content": "The HITSZ-VCM Dataset and LIP Dataset are large-scale datasets, which contains a large amount of person videos/images captured by cameras. In detail, the HITSZ-VCM dataset [20] contains 927 valid identities with image sequences. The LIP dataset [8] contains about 50,000 identities but with only one image for each. Totally, there are 251,452 RGB images in the HITSZ-VCM dataset and 50,000 in the LIP dataset. We implement the pre-training strategy on these datasets containing only RGB images, to explore the impact of pre-training with out-of-domain data on this task. Given that there is no point cloud data paired with these images, we use monocular depth estimation to generate pseudo point clouds, as shown in Fig. 4. This greatly expands the amount of data available for pre-training at scale. This also improves the accuracy of cross-modality recognition between 2D and 3D spaces, which marks the effectiveness of our proposed contrastive pre-training strategy." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.604, + 0.427, + 0.619 + ], + "angle": 0, + "content": "4.2 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.787, + 0.749 + ], + "angle": 0, + "content": "Implementation Details. Following [30], all the camera-based silhouettes and LiDAR-based depth images are aligned and then resized into the resolution of 6464. The total number of iterations is set to 120,000 for pre-training and 60,000 for fine-tuning. The Adam optimizer is used to prevent the issue of gradient vanishing because of low-quality silhouettes. The triplet and cross-entropy loss weights are set to the same. All comparison methods are trained using the same training strategy as LidarGait [30]. The OpenGait codebase [5] is utilized to conduct all experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Evaluation Protocol. All experiments are conducted on SUSTech1K. Following LidarGait [30], the dataset is divided into three splits: a training set with 250 identities and 6,011 sequences, a validation set with 6,025 sequences from 250 unseen identities, and a test set with the remaining 550 identities and 13,203 sequences. The SUSTech1K dataset provides gait sequences from multiple viewpoints. Thus, we also adopt the cross-view evaluation protocol [31,42] used in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.674, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CL-Gait" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "CASIA-B and OUMVLP. During the test, the sequences in normal conditions are grouped into gallery sets, and the sequences in variant conditions are taken as probe sets. To accurately assess the performance of camera-LiDAR cross-modality gait recognition, we evaluate the results using point cloud data as the probe and silhouette data as the gallery, i.e., LiDAR to Camera (L to C). We also evaluate the results using silhouette data as the probe and point cloud data as the gallery, i.e., Camera to LiDAR (C to L)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.29, + 0.447, + 0.306 + ], + "angle": 0, + "content": "4.3 Comparative Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.497 + ], + "angle": 0, + "content": "We evaluate several commonly used structures for cross-modality modeling as baselines, which basically follows the setting of [36]. The baseline models include one-steam structure, asymmetric FC layer structure, image-point structure, and the proposed two-stream structure. We apply residual block in ResNet-18 [11] as the base convolution block for all the four structures. For them, the loss function are weighted cross-entropy loss and triplet loss [2], which is commonly used and relatively stable. And all of the hyper parameters are kept the same. As for the input of one-steam structure, asymmetric FC layer structure and two-stream structure, the point clouds are converted to three-channel colored depth images and adjusted to the resolution of \\(64 \\times 64\\) with zero padding. We introduce the four structures as following:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.499, + 0.788, + 0.56 + ], + "angle": 0, + "content": "One-stream Structure is the most commonly used in vision tasks, where all parameters are shared in the whole network. Therefore, we organize silhouettes and depth images obtained from point clouds into the same dimensions before feeding them into the network." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.563, + 0.788, + 0.624 + ], + "angle": 0, + "content": "Asymmetric FC Layer Structure is used in multi-domain tasks, for example, IDR [12] for VIS-NIR face recognition. This structure shares nearly all parameters except for the last FC layer. It assumes that the feature extraction for different modalities can be same and adaptation is achieved in the backbone." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.788, + 0.703 + ], + "angle": 0, + "content": "Image-point Structure directly uses point clouds as input, rather than projected depth images. It employs two types of encoders: an image encoder for the silhouette input and a point cloud encoder for the point cloud input. Under this structure, we implement three commonly used point cloud encoders: GCN [9], PointNet [26], and Point-Transformer [43]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Two-stream Structure is commonly used in cross-modality matching tasks. It utilizes modality-specific modules in the shallow layers of the network and uniform modules with shared parameters in the deeper layers. Based on the two-stream structure, our proposed CL-Gait is equipped with a horizontal pooling module (HP) and a contrastive silhouette-point pre-training strategy (CSPP). To show the effectiveness of the HP module and CSPP strategy of CL-Gait, we implement three additional networks: 1) a two-stream backbone without CSPP, using global pooling instead of HP; 2) a two-stream backbone without CSPP, using HP; 3) a two-stream backbone with CSPP, using global pooling." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.432, + 0.129 + ], + "angle": 0, + "content": "W. Guo, Y. Liang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.188 + ], + "angle": 0, + "content": "Table 1: Evaluation with different structures on SUSTech1K valid + test set. We use ResNet-18 to extract image features by default. \"L to C\" indicates the results with point clouds as probe and silhouettes as gallery, and \"C to L\" indicates the reverse." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.2, + 0.779, + 0.335 + ], + "angle": 0, + "content": "
StructureModelL to CC to L
Rank-1Rank-3Rank-5Rank-1Rank-3Rank-5
One-streamResNet2.605.617.9914.9425.9231.87
Asymmetric FCResNet3.087.2310.5216.4329.1335.90
Image-pointGCN [9]12.2025.0133.2314.3028.3836.62
PointNet [26]7.1116.6123.358.4019.2526.90
Point-Transformer [43]13.2526.7235.2915.4230.0438.70
Two-streamResNet30.3948.7757.3135.2654.4262.84
ResNet+CSPP32.1850.8559.2236.5455.7564.11
ResNet+HP45.4162.8969.7142.3059.6767.12
CLGait (ours)*53.2969.5475.5955.1271.2377.31
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.232, + 0.337, + 0.416, + 0.352 + ], + "angle": 0, + "content": "i.e., ResNet + CSPP + HP." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.38, + 0.436, + 0.396 + ], + "angle": 0, + "content": "4.4 Comparative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.789, + 0.722 + ], + "angle": 0, + "content": "Following the cross-view and cross-modality evaluation protocol, we report the comparative results of the methods above in the Tab. 1. From the results, we can obtain the following observations: 1) CL-Gait demonstrates its superiority to all existing structures and methods, primarily due to the feature extraction capabilities of the two-stream structure. Additionally, the use of pre-training effectively mitigates modality discrepancy, further improving its performance. 2) Experiments under the two-stream structure can be considered as an ablation study, highlighting the important roles of CSPP and HP. Adding HP to the ResNet base improves the average rank-1 by approximately \\(11\\%\\), and further incorporating CSPP increases the average rank-1 by about another \\(10\\%\\). 3) The performance of the one-stream structure is poor, primarily because the one-stream structure uses an identical network to process both silhouettes and point clouds, failing to address the modality discrepancy between them. 4) Although the Asymmetric FC structure employs different FC layers for the two modalities, its performance remains poor. This indicates that modality-specific adaptations are best applied at the shallower levels of the network. Furthermore, during the training process, the asymmetric FC layers tend to cause the network to overfit quickly. 5) All methods utilizing the two-stream structure outperform those based on the image-point structure. This likely results from the depth images, obtained through projection, having data consistency and smaller modality discrepancy with silhouettes compared to point cloud inputs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.741, + 0.351, + 0.756 + ], + "angle": 0, + "content": "4.5 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "3D Geometry Information. To evaluate the effectiveness of depth information for cross-modality gait recognition, we compare another type of data as input for the second modality, e.g., LiDAR silhouettes. LiDAR silhouettes are obtained by range-view projection of point cloud sets and contains no 3D geometry information. The results are shown in Fig. 5. When 3D geometry information is not" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.674, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CL-Gait" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.141, + 0.761, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.275, + 0.788, + 0.331 + ], + "angle": 0, + "content": "Fig. 5: Comparison of different input forms of LiDAR data. The results show that the projected and interpolated depth from point cloud works best for the cross-modality matching. This indicates that the 3D geometry information are essential. For each input form, the best-performing model is utilized." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.501 + ], + "angle": 0, + "content": "included, the performance of LiDAR silhouettes is much lower. This is because the camera has a much higher resolution than LiDAR, so the silhouettes from the camera can have more details. Integrating depth information can improve average rank-1 accuracy from \\(40.59\\%\\) to \\(54.21\\%\\), validating the necessity and effectiveness of 3D information for cross-modality gait recognition. One important reason may be that the 3D information contained in the point cloud can compensate for the difference in viewing angles to some extent. Furthermore, we include the performance of direct point cloud input in Fig. 5 to compare with depth images and illustrate the input form of point clouds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.788, + 0.64 + ], + "angle": 0, + "content": "Contrastive Pre-training Datasets. To make up for the absence of paired camera-LiDAR data, we further propose a large-scale data generation strategy. Camera silhouette and LiDAR point data are generated from real-world RGB images for contrastive pre-training on such paired pseudo data. To valid its performance, we construct training sets from two different datasets, including LIP and HITSZ-VCM. These two datasets are commonly used for camera-based person segmentation and re-identification and contain single still RGB images. To ensure fair comparisons, we also conduct experiments on SUSTech1K training set, which contains real paired point clouds and camera silhouettes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.788, + 0.763 + ], + "angle": 0, + "content": "As shown in Tab. 2, when the amount of data is similar using data generated from the LIP dataset, the performance of pre-training on generated pseudo data is comparable to that of real data. When pre-trained on a larger scale of synthetic pseudo data generated from the HITSZ-VSM dataset, the accuracy from LiDAR to camera improves and exceeds that on the real data, proving the effectiveness of our data generation strategy for contrastive pre-training. However, the accuracy from camera to LiDAR decreases a little, probably due to the difference between generated point clouds from estimated depth and real points from LiDAR." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Temporal Modules. In order to aggregate the different frames in the sequence, temporal modules are used to fuse the features of multiple frames. We evaluate the effectiveness of different temporal modules in combination with our proposed CL-Gait, as shown in Tab. 3. We can observe that the simple temporal pooling structure demonstrates superiority over the other learning-based structure. This" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.432, + 0.129 + ], + "angle": 0, + "content": "W. Guo, Y. Liang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 2: Performance with different pre-training datasets. Using estimated depth for pre-training obtains close results compared with using real paired point-image dataset." + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.186, + 0.72, + 0.253 + ], + "angle": 0, + "content": "
DatasetData AmountL to CC to L
Rank-1Rank-3Rank-5Rank-1Rank-3Rank-5
SUSTech1K120K51.6068.2274.4356.1272.1577.95
LIP50K51.0068.0973.9955.0771.5977.59
HITSZ-VCM770K53.2969.5475.5955.1271.2377.31
" + }, + { + "type": "table_caption", + "bbox": [ + 0.276, + 0.265, + 0.726, + 0.279 + ], + "angle": 0, + "content": "Table 3: Performance of CL-Gait with different temporal modules." + }, + { + "type": "table", + "bbox": [ + 0.282, + 0.292, + 0.713, + 0.369 + ], + "angle": 0, + "content": "
FeatureL to CC to L
Rank-1Rank-3Rank-5Rank-1Rank-3Rank-5
Temporal Pooling [5]53.2969.5475.5955.1271.2377.31
LSTM [13]37.4055.0160.8939.4756.4162.72
Bi-LSTM [28]39.6355.5561.8339.8857.8263.71
Transformer [33]44.8061.3167.3647.1663.8369.11
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.786, + 0.442 + ], + "angle": 0, + "content": "result indicates that temporal pooling can better fuse multi-frame features to match domain information more effectively. One reason may be that learning-based approaches such as LSTM and Transformer are more likely to overfit." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.46, + 0.528, + 0.477 + ], + "angle": 0, + "content": "5 Conclusion and Future work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.788, + 0.67 + ], + "angle": 0, + "content": "This paper presents the first research on cross-modality gait recognition with point clouds and silhouettes from RGB images. Firstly, We propose a cross-modality gait recognition framework, named CL-Gait, that utilizes a two-stream network for feature embedding of different modalities, i.e., camera and LiDAR. Moreover, to mitigate modality discrepancy, we propose a contrastive pre-training strategy along with a large-scale data generation method. It can generate large number of data pairs (silhouette and point cloud) based on monocular depth estimation and pre-train the model in a contrastive manner. Extensive experiments demonstrate that CL-Gait achieves commendable performance in cross-modality evaluation modes, proving the importance of maintaining modality consistency and the effectiveness of contrastive pre-training. This also highlights the potential of cross-modality gait recognition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.788, + 0.762 + ], + "angle": 0, + "content": "CL-Gait has shown remarkable results, but we believe there remains significant room for improvement. The modality and resolution differences between point clouds and silhouettes, along with the inherent motion blur in point clouds, may be key factors impacting model performance. Therefore, densifying point clouds or developing point cloud encoders specifically for cross-modality gait recognition could enhance model performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.782, + 0.395, + 0.799 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "This work is supported by the National Key Research and Development Program of China under Grant (2018AAA0102803)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.674, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CL-Gait" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.786, + 0.218 + ], + "angle": 0, + "content": "1. Bobick, A.F., Davis, J.W.: The recognition of human movement using temporal templates. IEEE Transactions on Pattern Analysis and Machine Intelligence 23(3), 257-267 (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.786, + 0.26 + ], + "angle": 0, + "content": "2. Chao, H., He, Y., Zhang, J., Feng, J.: Gaitset: Regarding gait as a set for cross-view gait recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 8126-8133 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.786, + 0.314 + ], + "angle": 0, + "content": "3. Choi, S., Lee, S., Kim, Y., Kim, T., Kim, C.: Hi-cmd: Hierarchical cross-modality disentanglement for visible-infrared person re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10257-10266 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.315, + 0.786, + 0.343 + ], + "angle": 0, + "content": "4. Dai, P., Ji, R., Wang, H., Wu, Q., Huang, Y.: Cross-modality person re-identification with generative adversarial training. In: IJCAI. vol. 1, p. 6 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.786, + 0.384 + ], + "angle": 0, + "content": "5. Fan, C., Liang, J., Shen, C., Hou, S., Huang, Y., Yu, S.: Opengait: Revisiting gait recognition towards better practicality. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9707-9716 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.384, + 0.786, + 0.439 + ], + "angle": 0, + "content": "6. Fan, C., Peng, Y., Cao, C., Liu, X., Hou, S., Chi, J., Huang, Y., Li, Q., He, Z.: Gaitpart: Temporal part-based model for gait recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14225-14233 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.439, + 0.786, + 0.481 + ], + "angle": 0, + "content": "7. Feng, Z., Lai, J., Xie, X.: Learning modality-specific representations for visible-infrared person re-identification. IEEE Transactions on Image Processing 29, 579-590 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.481, + 0.786, + 0.536 + ], + "angle": 0, + "content": "8. Gong, K., Liang, X., Zhang, D., Shen, X., Lin, L.: Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 932-940 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.536, + 0.786, + 0.564 + ], + "angle": 0, + "content": "9. Guo, W., Pan, Z., Liang, Y., Xi, Z., Zhong, Z.C., Feng, J., el al: Lidar-based person re-identification. arXiv preprint arXiv:2312.03033 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.564, + 0.786, + 0.606 + ], + "angle": 0, + "content": "0. Hao, Y., Wang, N., Li, J., Gao, X.: Hsme: Hypersphere manifold embedding for visible thermal person re-identification. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 8385-8392 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.786, + 0.647 + ], + "angle": 0, + "content": "1. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.786, + 0.688 + ], + "angle": 0, + "content": "2. He, R., Wu, X., Sun, Z., Tan, T.: Learning invariant deep representation for nir-vis face recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 31 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.688, + 0.786, + 0.716 + ], + "angle": 0, + "content": "3. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural computation 9(8), 1735-1780 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.786, + 0.758 + ], + "angle": 0, + "content": "4. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., el al: Scaling up visual and vision-language representation learning with noisy text supervision. In: International Conference on Machine Learning. pp. 4904-4916. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.786, + 0.799 + ], + "angle": 0, + "content": "5. Li, X., Makihara, Y., Xu, C., Yagi, Y., Yu, S., Ren, M.: End-to-end model-based gait recognition. In: Proceedings of the Asian Conference on Computer Vision (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.786, + 0.841 + ], + "angle": 0, + "content": "6. Liang, J., Fan, C., Hou, S., Shen, C., Huang, Y., Yu, S.: Gaitedge: Beyond plain end-to-end gait recognition for better practicality. In: European Conference on Computer Vision. pp. 375-390. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.176, + 0.786, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.432, + 0.129 + ], + "angle": 0, + "content": "W. Guo, Y. Liang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "17. Liao, R., Yu, S., An, W., Huang, Y.: A model-based gait recognition method with body pose and human prior knowledge. Pattern Recognition 98, 107069 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "18. Lin, B., Liu, Y., Zhang, S.: Gaitmask: Mask-based model for gait recognition. In: BMVC. pp. 1-12 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.206, + 0.785, + 0.247 + ], + "angle": 0, + "content": "19. Lin, B., Zhang, S., Wang, M., Li, L., Yu, X.: Caitgl: Learning discriminative global-local feature representations for gait recognition. arXiv preprint arXiv:2208.01380 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.249, + 0.785, + 0.303 + ], + "angle": 0, + "content": "20. Lin, X., Li, J., Ma, Z., Li, H., Li, S., Xu, K., Lu, G., Zhang, D.: Learning modal-invariant and temporal-memory for video-based visible-infrared person re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20973–20982 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.305, + 0.785, + 0.358 + ], + "angle": 0, + "content": "21. Ling, Y., Zhong, Z., Luo, Z., Rota, P., Li, S., Sebe, N.: Class-aware modality mix and center-guided metric learning for visible-thermal person re-identification. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 889-897 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.785, + 0.402 + ], + "angle": 0, + "content": "22. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Spl: A skinned multi-person linear model. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 851-866 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.404, + 0.785, + 0.457 + ], + "angle": 0, + "content": "23. Lu, Y., Wu, Y., Liu, B., Zhang, T., Li, B., Chu, Q., Yu, N.: Cross-modality person re-identification with shared-specific feature transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13379–13389 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.5 + ], + "angle": 0, + "content": "24. Nguyen, D.T., Hong, H.G., Kim, K.W., Park, K.R.: Person recognition system based on a combination of body images from visible light and thermal cameras. Sensors 17(3), 605 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.543 + ], + "angle": 0, + "content": "25. Park, H., Lee, S., Lee, J., Ham, B.: Learning by aligning: Visible-infrared person re-identification using cross-modal correspondences. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12046–12055 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.545, + 0.785, + 0.585 + ], + "angle": 0, + "content": "26. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3D classification and segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 652-660 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.785, + 0.641 + ], + "angle": 0, + "content": "27. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.785, + 0.67 + ], + "angle": 0, + "content": "28. Schuster, M., Paliwal, K.K.: Bidirectional recurrent neural networks. IEEE Transactions on Signal Processing 45(11), 2673-2681 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.785, + 0.699 + ], + "angle": 0, + "content": "29. Sepas-Moghaddam, A., Etemad, A.: Deep gait recognition: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(1), 264-284 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.701, + 0.785, + 0.741 + ], + "angle": 0, + "content": "30. Shen, C., Fan, C., Wu, W., Wang, R., Huang, G.Q., Yu, S.: Lidargait: Benchmarking 3D gait recognition with point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1054-1063 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.797 + ], + "angle": 0, + "content": "31. Takemura, N., Makihara, Y., Muramatsu, D., Echigo, T., Yagi, Y.: Multi-view large population gait dataset and its performance evaluation for cross-view gait recognition. IPSJ Transactions on Computer Vision and Applications 10, 1-14 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "32. Teepe, T., Khan, A., Gilg, J., Herzog, F., Hörmann, S., Rigoll, G.: Gaitgraph: Graph convolutional network for skeleton-based gait recognition. In: 2021 IEEE International Conference on Image Processing (ICIP). pp. 2314-2318. IEEE (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.674, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CL-Gait" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "33. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in Neural Information Processing Systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "34. Wang, C., Zhang, J., Pu, J., Yuan, X., Wang, L.: Chrono-gait image: A novel temporal template for gait recognition. In: Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part I 11. pp. 257-270. Springer (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.3 + ], + "angle": 0, + "content": "35. Wang, Y., Chao, W.L., Garg, D., Hariharan, B., Campbell, M., Weinberger, K.Q.: Pseudo-lidar from visual depth estimation: Bridging the gap in 3D object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8445-8453 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.787, + 0.342 + ], + "angle": 0, + "content": "36. Wu, A., Zheng, W.S., Yu, H.X., Gong, S., Lai, J.: RGB-infrared cross-modality person re-identification. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 5380-5389 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.397 + ], + "angle": 0, + "content": "37. Yang, J., Duan, J., Tran, S., Xu, Y., Chanda, S., Chen, L., Zeng, B., Chilimbi, T., Huang, J.: Vision-language pre-training with triple contrastive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15671-15680 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.438 + ], + "angle": 0, + "content": "38. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. arXiv preprint arXiv:2401.10891 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "39. Ye, M., Lan, X., Leng, Q.: Modality-aware collaborative learning for visible thermal person re-identification. In: Proceedings of the 27th ACM International Conference on Multimedia. pp. 347-355 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "40. Ye, M., Lan, X., Leng, Q., Shen, J.: Cross-modality person re-identification via modality-aware collaborative ensemble learning. IEEE Transactions on Image Processing 29, 9387–9399 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.578 + ], + "angle": 0, + "content": "41. Ye, M., Shen, J., J. Crandall, D., Shao, L., Luo, J.: Dynamic dual-attentive aggregation learning for visible-infrared person re-identification. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVII 16. pp. 229–247. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.619 + ], + "angle": 0, + "content": "42. Yu, S., Tan, D., Tan, T.: A framework for evaluating the effect of view angle, clothing and carrying condition on gait recognition. In: 18th International Conference on Pattern Recognition (ICPR'06). vol. 4, pp. 441-444. IEEE (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.787, + 0.66 + ], + "angle": 0, + "content": "43. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16259-16268 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.66 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Camera-LiDAR Cross-modality Gait Recognition/ce94d361-28a2-4f35-a27c-eca979dcca0e_origin.pdf b/2024/Camera-LiDAR Cross-modality Gait Recognition/ce94d361-28a2-4f35-a27c-eca979dcca0e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b6b707526137c9e8fe2e3f8bf19014a94036a6a1 --- /dev/null +++ b/2024/Camera-LiDAR Cross-modality Gait Recognition/ce94d361-28a2-4f35-a27c-eca979dcca0e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d78396cf17d9d54348ed96b421dca01736b1da8f4f019806aa177d96b085623 +size 1941044 diff --git a/2024/Camera-LiDAR Cross-modality Gait Recognition/full.md b/2024/Camera-LiDAR Cross-modality Gait Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bc7324d3b201d72e3ecdc01f1caa3acfb6ee69ea --- /dev/null +++ b/2024/Camera-LiDAR Cross-modality Gait Recognition/full.md @@ -0,0 +1,296 @@ +# Camera-LiDAR Cross-modality Gait Recognition + +Wenxuan Guo $^{1*}$ , Yingping Liang $^{2*}$ , Zhiyu Pan $^{1}$ , Ziheng Xi $^{1}$ , Jianjiang Feng $^{1\dagger}$ , and Jie Zhou $^{1}$ + +1 Department of Automation, BNRist, Tsinghua University + +2 Beijing Institute of Technology + +{gwx22, pzy20, xizh21}@mails.tsinghua.edu.cn liangyingping@bit.edu.cn + +{jfeng, jzhou}@tsinghua.edu.cn + +Abstract. Gait recognition is a crucial biometric identification technique. Camera-based gait recognition has been widely applied in both research and industrial fields. LiDAR-based gait recognition has also begun to evolve most recently, due to the provision of 3D structural information. However, in certain applications, cameras fail to recognize persons, such as in low-light environments and long-distance recognition scenarios, where LiDARs work well. On the other hand, the deployment cost and complexity of LiDAR systems limit its wider application. Therefore, it is essential to consider cross-modality gait recognition between cameras and LiDARs for a broader range of applications. In this work, we propose the first cross-modality gait recognition framework between Camera and LiDAR, namely CL-Gait. It employs a two-stream network for feature embedding of both modalities. This poses a challenging recognition task due to the inherent matching between 3D and 2D data, exhibiting significant modality discrepancy. To align the feature spaces of the two modalities, i.e., camera silhouettes and LiDAR points, we propose a contrastive pre-training strategy to mitigate modality discrepancy. To make up for the absence of paired camera-LiDAR data for pre-training, we also introduce a strategy for generating data on a large scale. This strategy utilizes monocular depth estimated from single RGB images and virtual cameras to generate pseudo point clouds for contrastive pre-training. Extensive experiments show that the cross-modality gait recognition is very challenging but still contains potential and feasibility with our proposed model and pre-training strategy. To the best of our knowledge, this is the first work to address cross-modality gait recognition. The code and dataset are available at https://github.com/GWxuan/CL-Gait. + +Keywords: Gait recognition $\cdot$ Cross-modality $\cdot$ Contrastive pre-training + +# 1 Introduction + +Gait recognition is a crucial long-range identification technology without physical contact, which has great advantages in privacy protection and cross-clothing + +![](images/d4fdceced539d9b484f3e35912d492c854116eaa852e8b993eb3495d49d4afa7.jpg) +(a) Camera-based gait recognition (top), and LiDAR-based gait recognition (bottom). + +![](images/8bce0b8fd5723c2d202e639839223a620f367be28abaa338faaaa9b9012b1727.jpg) +(b) Our proposed cross-modality gait recognition between camera and LiDAR. +Fig. 1: Overview of single-modality and cross-modality gait recognition. Single-modality gait recognition takes data of one modality as input and searches within a gallery of the same data type. In contrast, cross-modality gait recognition processes two modalities and identify individuals within a gallery of a different modality. + +recognition [2]. It identifies individuals based on behavioral characteristics extracted from walking sequences recorded by sensors. Gait recognition has been extensively applied in fields such as user identification, sports science, and public security. Most current gait recognition research employs cameras as sensors [29]. With the advancement of computer vision technology, camera-based gait recognition continues to achieve new progress [16-19, 22, 32]. + +While utilizing cameras for gait recognition is cost-effective, it faces limitations in certain applications, such as low-light environments, long-distance recognition, and scenarios requiring precise 3D perception. In low-light conditions, the available information in RGB images significantly diminishes [9], making it challenging to accurately detect pedestrians and segment their silhouettes. Furthermore, due to the decline in image clarity, camera-based systems struggle to distinguish the fine-grained features from a distance [5]. Additionally, accurately capturing the 3D movement of individuals, crucial for comprehensive gait analysis, is beyond the capability of traditional camera systems [30]. + +Recently, to address the challenges under low light environment faced by RGB cameras, some researches [21, 23, 25, 40] have focused on RGB-Infrared cross-modality person re-identification (ReID), achieving commendable performance. However, infrared cameras struggle to capture pedestrians wearing insulating or thick clothing. Besides, they are constrained by limited detection distance and resolution, as well as susceptibility to lighting conditions and weather. It also faces the risk of pedestrian privacy and safety, compared with gait recognition. These limitations highlight the need for exploring alternative or complementary technologies that can offer more robust solutions for person identification tasks across a variety of environmental conditions. + +In recent years, LiDAR technology has been increasingly utilized [35] due to its impressive gains on 3D information. Beyond its mature applications in autonomous driving, LiDAR has also been applied in individual identification areas, such as gait recognition [30] and person re-identification [9] most recently. + +Through point cloud data, LiDAR can provide precise 3D perception, aiding in the acquisition of an individual's 3D geometric structure. From this, intrinsic characteristics of a person, including height, body shape, and gait, can be extracted. Moreover, LiDARs possess long-distance perception capabilities, unaffected by lighting conditions and complex backgrounds. And it offers enhanced privacy protection than camera [9]. The robustness to environmental variables and the ability to capture detailed biometric data make it an invaluable tool for advanced surveillance and security systems. However, compared to camera systems, LiDAR's deployment complexity and cost are somehow higher, making it unsuitable for large-scale deployment in general scenarios. + +Considering the respective advantages of cameras and LiDARs, switching sensors in different scenarios to explore cross-modality gait recognition presents a worthwhile task for investigation. Cross-modality gait recognition offers an option with a broader application scope. Cameras can be utilized in general scenes with normal lighting, while LiDAR can be applied in specific scenarios, offering a complementary tool that enhances overall recognition capabilities. + +In this work, we present the first study on camera-LiDAR cross-modality gait recognition as shown in Fig. 1. This task poses significant challenges due to the need for matching data between 3D and 2D modalities. Based on our observations, there are substantial differences between point clouds and images. To be specific, for gait recognition, point clouds primarily focus on the 3D positioning of body parts, whereas images capture precise contour information of individuals. To address these challenges and explore the feasibility of cross-modality gait recognition, we propose a novel cross-modality framework, namely CL-Gait. + +To extract gait features from both modalities, CL-Gait basically employs a two-stream network. Considering modality differences and data processing consistency, we first project and upsample the point clouds to obtain depth images with the same resolution as silhouettes, and use the projected depth images to train the network. Furthermore, to mitigate modality discrepancy, CL-Gait utilizes a contrastive silhouette-point pre-training approach (CSPP) to align the feature spaces of the two modalities. Pre-training requires aligned camera-LiDAR data under the same view, which is hard to obtain. To this end, we propose a new strategy of generating cross-modality gait data for pre-training. This strategy utilizes advances in monocular depth estimation and is able to generate large-scale dataset for pre-training using only single RGB images. + +Extensive experiments have revealed the following three insights: 1) CL-Gait achieves an average rank-1 accuracy of $54.21\%$ . The results show at least $22.90\%$ improvement over the baseline models and demonstrate the significant potential of cross-modality gait recognition. 2) Utilizing large-scale depth images generated from point clouds as input is proved superior to using point clouds directly. 3) The strategy of contrastive pre-training on large-scale generated paired data mitigates the modality differences, contributing to performance improvement. + +To summarize, our main contributions are as follows: + +- To the best of our knowledge, this is the first work on cross-modality gait recognition. Extensive experiments show the potential of utilizing camera and LiDAR for gait recognition in challenging cross-modality scenarios. +- We analyse several network structures and conduct comparable experiments on their effectiveness for cross-modality gait recognition. Based on the two-stream network with better performance, our proposed CL-Gait is capable of matching camera silhouettes and LiDAR points. +- We propose a contrastive pre-training method to align the feature spaces of the two modalities. To make up for the absence of paired camera-LiDAR data, we further introduce a large-scale data generation strategy. + +# 2 Related Work + +Gait Recognition. Based on the sensor used, exiting gait recognition methods can be categorized into camera-based and LiDAR-based methods. + +Camera-based gait recognition has been extensively studied over the past decade. The majority of camera-based methods focus on extracting appearance features directly from images or videos [2, 6], adapting well to resolution reduction and achieving impressive performance [5]. These methods utilize silhouettes [16, 18, 19] or other gait templates [1, 34] for spatial feature extraction and temporal modeling. Additionally, some researchers have explored using estimated underlying structure of human body [15, 17, 32], such as 2D/3D pose and the SMPL model [22], as inputs. While theoretically robust against factors like carrying and clothing, these models often struggle at low resolutions, limiting their practicality in some real-world scenarios [5]. + +LiDAR-based gait recognition is an emerging field that utilizes precise 3D representations, point clouds, capturing complex motion patterns and the 3D structure of individual gaits. This approach is less susceptible to variations in lighting, clothing, and background, offering promising avenues for accurate gait recognition in diverse conditions. LidarGait [30] stands as a pioneering work in this field, introducing a multimodal gait recognition dataset named SUSTECH1K. This dataset is designed for evaluating the performance of gait recognition based on various sensors and has demonstrated the potential and practicality of LiDAR-based gait recognition. Additionally, some researchers have applied LidAR to the task of person re-identification [9], achieving impressive results. + +Camera-based and LiDAR-based methods have been proven practical, each with its unique advantages and disadvantages. Therefore, providing a compromise solution by addressing cross-modality gait recognition represents a valuable research direction, which has not been studied currently. + +Contrastive Pre-training. Contrastive pre-training is primarily applied in the field of vision-language models, aimed at aligning the feature spaces of visual and linguistic models. In recent years, inspired by the success of self-supervised learning within intra-modal tasks, researchers have begun to explore pre-training objectives for tasks involving multiple modalities, such as vision and language [37]. The pioneering work by CLIP [27] performs cross-modality + +contrastive pre-training on hundreds of millions of image-text pairs. CLIP [27] can generate a task-agnostic model that achieves surprisingly effective results. Subsequently, ALIGN [14] expands upon CLIP by utilizing a noisy dataset that covers more than a billion image-text pairs, further extending the capabilities of cross-modality pre-training. Contrastive pre-training plays a crucial role in cross-modality tasks. Therefore, we introduce contrastive pre-training into camera-LiDAR cross-modality gait recognition. + +RGB-IR Cross-modality Person ReID. Visible-infrared person ReID addresses the challenge of matching persons across different modalities, specifically between RGB and infrared cameras [36]. This task is particularly significant for scenarios with poor illumination, notably at night. With two primary datasets, SYSU-MM01 [36] and RegDB [24], RGB-IR cross-modality person ReID has been extensively studied, most of which focusing on metric learning, feature learning, and adversarial learning approaches. Metric learning [7, 10, 21, 40] and feature learning [25, 39, 41] methods aim to extract and align multimodal features into a common space for effective comparison, employing strategies like angle-based measurement and Euclidean constraints to bridge the modality gap. Notably, advancements in this area include the development of novel loss functions [21]. Concurrently, Generative Adversarial Networks (GANs) [3, 4, 23] have been instrumental in cross-modality ReID, enabling the transformation of images between RGB and IR while maintaining identity information, and fostering feature learning through adversarial and disentanglement strategies. Despite their potential, GAN-based methods face challenges such as high computational demands and the occasional production of low-quality images that may negatively impact ReID performance. + +While RGB-IR person ReID mainly relies on appearance for distinction, which is a short-term feature and may raise privacy concerns. In contrast, cross-modality gait recognition focuses on shape and gait of individuals, which are time-invariant features, and offers better protection of privacy. + +# 3 Method + +In this work, we propose CL-Gait for cross-modality gait recognition between camera and LiDAR. CL-Gait employs a two-stream network for cross-modality feature embedding. The network utilizes modality-specific modules in the shallow layers and shared modules in the deeper layers, as illustrated in Fig. 2. Besides, CL-Gait adopts a contrastive learning strategy to align the feature spaces of the two modalities, mitigating modality differences, as shown in Fig. 3. The backbone obtained from pre-training is used to initialize the cross-modality embedding network. To facilitate pre-training on large-scale data, we also propose a method for generating cross-modality gait data, as shown in Fig. 4. + +# 3.1 Cross-modality Network of CL-Gait + +Two-Stream Backbone. The CL-Gait framework processes two distinct data modalities: 2D silhouette sequences and 3D point cloud sequences, optimizing + +![](images/e9666d6768e7918a5e7b504bc5974f00b5b163fa618e56ec24222bea9720ade3.jpg) +Fig. 2: The cross-modality network of our proposed CL-Gait. It employs a two-stream architecture that encodes sequences from two modalities into a consistent feature space. HP stands for horizontal pooling, and TP represents temporal pooling. + +cross-modality feature extracting. Inspired by [30], we first project and interpolate point clouds into depth images. For each modality, we apply modality-specific convolution layers (Conv-S for silhouettes and Conv-P for depth images generated from point clouds) before projecting the features into a shared embedding space, as shown in Fig. 2. Following [36], the two-stream backbone is modified from ResNet-18, with its first layer altered to be the modality-specific convolution layers, and the subsequent layers serving as shared convolution layers. For silhouettes, we define the processing as follows: + +$$ +F _ {I} = L _ {S} (I), \tag {1} +$$ + +where $I$ represents the silhouette and $L_{S}$ denotes the silhouette-specific convolution layer, which produces the feature map $F_{I}$ . + +Considering the modality differences and the need for consistency in data processing, we first project and interpolate point clouds to generate depths with the same resolution as silhouettes. And the processing is as follows: + +$$ +F _ {P} = L _ {P} (\operatorname {P I} (P)), \tag {2} +$$ + +where $P$ denotes the point cloud, PI represents the projection and interpolation process, and $L_{P}$ is the depth-specific convolution layer, which produces the feature map $F_{P}$ for the point cloud. + +After extracting modality-specific features $F_{P}$ and $F_{S}$ , we aim to unify these features within a consistent feature space. To achieve this, we employ a shared convolutional layer, denoted as $L_{\mathrm{sh}}$ , to process both $F_{P}$ and $F_{S}$ : + +$$ +F ^ {\prime} = L _ {\mathrm {s h}} \left(F _ {m}\right), \text {w h e r e} F _ {m} = F _ {P} \text {o r} F _ {S}, \tag {3} +$$ + +where $F^{\prime}$ is the output feature map of $L_{\mathrm{sh}}$ + +Temporal and Spatial Pooling. In our network, we follow commonly used strategies in gait recognition for temporal and spatial aggregation, specifically + +through temporal and horizontal pooling [5]. These techniques are essential for effectively summarizing and interpreting the dynamic and spatial aspects of gait data. Horizontal pooling divides the input feature map horizontally into $N_{H}$ parts, with each part being aggregated into a single feature vector: + +$$ +\mathrm {H P}: \mathbb {R} ^ {T \times H \times W \times C} \rightarrow \mathbb {R} ^ {T \times N _ {H} \times C}, \tag {4} +$$ + +where $T$ denotes the sequence length, $H$ and $W$ represent the height and width of the feature map, respectively, and $C$ indicates the dimension of the features. + +Temporal pooling allows for the aggregation of features from variable-length sequences, enabling the model to capture the final sequence-level representation: + +$$ +\mathrm {T P}: \mathbb {R} ^ {T \times N _ {H} \times C} \rightarrow \mathbb {R} ^ {N _ {H} \times C}. \tag {5} +$$ + +Therefore, we obtain $N_{H}$ feature vectors, which are further mapped to the metric space using $N_{H}$ independent Multi-Layer Perceptrons. These mapped vectors are then concatenated to form the final sequence-level feature embedding. Training and Inference. The cross-modality network of CL-Gait is trained with cross-modality triplet loss and cross-entropy loss. We have modified the triplet loss to accommodate cross-modality feature learning, where each modality is alternately used as the anchor, while the corresponding positive and negative samples are selected from the other modality, formulated as: + +$$ +\mathcal {L} _ {\text {c r o s s - t r i p l e t}} = \frac {1}{2} \left(\mathcal {L} _ {\text {t r i p l e t}} \left(P, I _ {\text {p o s}}, I _ {\text {n e g}}\right) + \mathcal {L} _ {\text {t r i p l e t}} \left(I, P _ {\text {p o s}}, P _ {\text {n e g}}\right)\right). \tag {6} +$$ + +The training loss is a combination of both losses with a hyperparameter $\gamma$ : + +$$ +\mathcal {L} = \mathcal {L} _ {\text {c r o s s - t r i p l e t}} + \gamma \mathcal {L} _ {\mathrm {c e}}, \tag {7} +$$ + +where $\mathcal{L}_{\mathrm{ce}}$ is the cross-entropy loss. During inference, the similarity between samples from different modalities is measured using the Euclidean distance. + +# 3.2 Contrastive Silhouette-point Pre-training + +Based on our observations in gait recognition tasks, the significant modality difference between 3D point clouds and 2D images could be a critical factor affecting model performance. To be specific, point clouds focus more on the 3D positioning of body parts, whereas images concentrate on the contour information of the individual. It's crucial to establish connections between the distinct information focused on by each modality. Inspired by CLIP [27], we propose a contrastive silhouette-point pre-training (CSPP) strategy to align the feature spaces of both modalities in the convolution-based encoders, as shown in Fig. 3. + +The pre-training process does not require identity labels from the samples for supervision. Trained on paired single-view data from aligned camera and LiDAR, the pre-training could make the model focus on learning a robust representation that bridges the gap between the modalities without direct identity-based guidance, and enhance the performance of the cross-modality network. + +![](images/e512b73a4f3d63db19719fb49c70ff0b30d4faa444d5ddae1defc963ad4533a6.jpg) +Fig. 3: The contrastive silhouette-point pre-training (CSPP) approach of CL-Gait. Paired silhouettes and point clouds are taken as inputs, and the backbone of the cross-modality network is pre-trained with contrastive learning loss to align the feature spaces of the two modalities. The pre-training process does not require identity labels of the samples for supervision. + +Pre-training Process. For pre-training, the input shifts from sequence data to numerous pairs of point clouds and silhouettes. These pairs are defined as data collected from the same individual at the same moment, encompassing both modalities. It ensures that the network learns to reconcile the inherent differences between the 3D and 2D representations, focusing on the alignment of features that accurately reflect the same gait patterns across modalities. + +The paired data is processed by the two-stream backbone and spatial pooling operations to obtain paired feature embeddings, i.e., $\mathcal{S} = \{S_i | i = 1, 2, \dots, N\}$ for silhouettes and $\mathcal{P} = \{P_i | i = 1, 2, \dots, N\}$ for point clouds. They are then utilized to compute the loss for contrastive representation learning. + +Contrastive Learning loss. The cornerstone of our pre-training is the contrastive learning loss [27]. Given the paired feature embeddings $S = \{S_{i} | i = 1, 2, \dots, N\}$ and $\mathcal{P} = \{P_{i} | i = 1, 2, \dots, N\}$ , there are $N \times N$ possible (silhouettes, point clouds) pairings. CL-Gait is trained to learn a multi-modality embedding space by maximizing the cosine similarity of the feature embeddings of the $N$ real pairs while minimizing the cosine similarity of the embeddings of the $N^2 - N$ incorrect pairings. The cosine similarity matrix is calculated as follows: + +$$ +\mathcal {M} = N _ {\mathrm {L} 2} (\mathcal {S}) \times N _ {\mathrm {L} 2} (\mathcal {P}) ^ {T}, \tag {8} +$$ + +where $\mathcal{S} \in \mathbb{R}^{N \times D}$ , $\mathcal{P} \in \mathbb{R}^{N \times D}$ , $\mathcal{M} \in \mathbb{R}^{N \times N}$ , $N_{\mathrm{L2}}(\cdot)$ denotes L2 normalization, and $(\cdot)^T$ denotes the transpose of the matrix. Then, a symmetric cross entropy loss over these similarity scores is optimized: + +$$ +\mathcal {L} _ {\text {c o n}} = \frac {1}{2} \left(\mathcal {L} _ {\mathrm {c e}} (\mathcal {M}, G) + \mathcal {L} _ {\mathrm {c e}} \left(\mathcal {M} ^ {T}, G\right)\right), \tag {9} +$$ + +where $G = \{1,2,\dots,N\}$ . $G$ represents the labels used for calculating the cross-entropy loss, indicating that pairs corresponding to the similarities on the diagonal of matrices $\mathcal{M}$ and $\mathcal{M}^T$ are considered positive samples, while all others are treated as negative samples. This approach ensures that the network learns to distinguish between matching and non-matching modality pairs effectively. + +![](images/16e78f6ce5709cbd6ca4f6dc3707c941667c4b0581186fbdfdfc3d3793683077.jpg) +Fig. 4: Illustration of paired gait data generation from single RGB images for contrastive pre-training. The quality of the synthesized data is comparable to the real-world data, making it possible to synthesize large-scale pre-training data. + +In the process described above, each frame of point cloud or silhouette, after passing through the pre-trained network, yields a feature embedding of size $(1,C)$ , indicating global pooling is applied to the output feature map of the backbone. For practical application and to align each local feature embedding, horizontal pooling is used to obtain features of size $(N,C)$ for each frame. During the computation of the contrastive learning loss, pairs of features from the same individual, moment, and part across two modalities are considered positive samples, with all others being negative. Through our practice, we have found that adopting a strategy of aligning local features during the pre-training phase is more suitable for gait recognition tasks, leading to performance improvements. + +# 3.3 Generation of Pre-training Gait Data + +Pre-training on real and large-scale data is challenging due to the high cost of acquiring paired RGB and point cloud data. To address this issue, we propose a method for generating pseudo data based on monocular depth estimation. As illustrated in Fig. 4, we use Depth Anything [38] to estimate dense depth $D \in \mathbb{R}^{H \times W}$ from large-scale single RGB images. Then, utilizing a virtual camera with intrinsic $K$ , the estimated depth is mapped to pseudo point clouds: + +$$ +[ P ^ {\prime}, 1 ] ^ {T} = K ^ {- 1} D [ u, v, 1 ] ^ {T}, \tag {10} +$$ + +where $P' \in \mathbb{R}^{H \times W \times 3}$ is the point coordinates, $K^{-1}$ indicates the inverse of intrinsic $K$ , and $u, v$ are the image pixel coordinates. The pseudo points $P'$ , after being downsampled by voxel grid into $P_{down}'$ , can be used for contrastive pretraining between camera silhouettes and LiDAR point clouds. To be specific, the downsampled points $P_{down}'$ are projected back into the image pixel coordinate to obtain the depth images for contrastive pre-training. Compared to point clouds, RGB images are less costly to collect and easier to acquire, with many public datasets of pedestrian image already available. Our proposed method makes it feasible to synthesize large-scale data for contrastive pre-training. + +# 4 Experiments + +# 4.1 Datasets + +The SUSTech1K Dataset [30] is the first large-scale LiDAR-based gait recognition dataset collected by a LiDAR sensor and an RGB camera. The dataset contains 25,239 sequences from 1,050 subjects and covers many variations, including visibility, views, occlusions, clothing, carrying, and scenes. It is times-tamped frames for each modality of frames. The first single-modality 3D gait recognition framework is trained on SUSTech1K using point clouds. In our study, we explore new uses for the SUSTech1K dataset. Our approach involves both images and point clouds, which marks the first attempt of cross-modality gait recognition with this dataset. We also introduce a novel pre-training strategy based on paired images and point clouds from the dataset. This strategy utilizes contrastive learning for cross-modality matching between images and point clouds and improves the performance of our proposed cross-modality gait method. + +The HITSZ-VCM Dataset and LIP Dataset are large-scale datasets, which contains a large amount of person videos/images captured by cameras. In detail, the HITSZ-VCM dataset [20] contains 927 valid identities with image sequences. The LIP dataset [8] contains about 50,000 identities but with only one image for each. Totally, there are 251,452 RGB images in the HITSZ-VCM dataset and 50,000 in the LIP dataset. We implement the pre-training strategy on these datasets containing only RGB images, to explore the impact of pre-training with out-of-domain data on this task. Given that there is no point cloud data paired with these images, we use monocular depth estimation to generate pseudo point clouds, as shown in Fig. 4. This greatly expands the amount of data available for pre-training at scale. This also improves the accuracy of cross-modality recognition between 2D and 3D spaces, which marks the effectiveness of our proposed contrastive pre-training strategy. + +# 4.2 Experimental Setup + +Implementation Details. Following [30], all the camera-based silhouettes and LiDAR-based depth images are aligned and then resized into the resolution of 6464. The total number of iterations is set to 120,000 for pre-training and 60,000 for fine-tuning. The Adam optimizer is used to prevent the issue of gradient vanishing because of low-quality silhouettes. The triplet and cross-entropy loss weights are set to the same. All comparison methods are trained using the same training strategy as LidarGait [30]. The OpenGait codebase [5] is utilized to conduct all experiments. + +Evaluation Protocol. All experiments are conducted on SUSTech1K. Following LidarGait [30], the dataset is divided into three splits: a training set with 250 identities and 6,011 sequences, a validation set with 6,025 sequences from 250 unseen identities, and a test set with the remaining 550 identities and 13,203 sequences. The SUSTech1K dataset provides gait sequences from multiple viewpoints. Thus, we also adopt the cross-view evaluation protocol [31,42] used in + +CASIA-B and OUMVLP. During the test, the sequences in normal conditions are grouped into gallery sets, and the sequences in variant conditions are taken as probe sets. To accurately assess the performance of camera-LiDAR cross-modality gait recognition, we evaluate the results using point cloud data as the probe and silhouette data as the gallery, i.e., LiDAR to Camera (L to C). We also evaluate the results using silhouette data as the probe and point cloud data as the gallery, i.e., Camera to LiDAR (C to L). + +# 4.3 Comparative Methods + +We evaluate several commonly used structures for cross-modality modeling as baselines, which basically follows the setting of [36]. The baseline models include one-steam structure, asymmetric FC layer structure, image-point structure, and the proposed two-stream structure. We apply residual block in ResNet-18 [11] as the base convolution block for all the four structures. For them, the loss function are weighted cross-entropy loss and triplet loss [2], which is commonly used and relatively stable. And all of the hyper parameters are kept the same. As for the input of one-steam structure, asymmetric FC layer structure and two-stream structure, the point clouds are converted to three-channel colored depth images and adjusted to the resolution of $64 \times 64$ with zero padding. We introduce the four structures as following: + +One-stream Structure is the most commonly used in vision tasks, where all parameters are shared in the whole network. Therefore, we organize silhouettes and depth images obtained from point clouds into the same dimensions before feeding them into the network. + +Asymmetric FC Layer Structure is used in multi-domain tasks, for example, IDR [12] for VIS-NIR face recognition. This structure shares nearly all parameters except for the last FC layer. It assumes that the feature extraction for different modalities can be same and adaptation is achieved in the backbone. + +Image-point Structure directly uses point clouds as input, rather than projected depth images. It employs two types of encoders: an image encoder for the silhouette input and a point cloud encoder for the point cloud input. Under this structure, we implement three commonly used point cloud encoders: GCN [9], PointNet [26], and Point-Transformer [43]. + +Two-stream Structure is commonly used in cross-modality matching tasks. It utilizes modality-specific modules in the shallow layers of the network and uniform modules with shared parameters in the deeper layers. Based on the two-stream structure, our proposed CL-Gait is equipped with a horizontal pooling module (HP) and a contrastive silhouette-point pre-training strategy (CSPP). To show the effectiveness of the HP module and CSPP strategy of CL-Gait, we implement three additional networks: 1) a two-stream backbone without CSPP, using global pooling instead of HP; 2) a two-stream backbone without CSPP, using HP; 3) a two-stream backbone with CSPP, using global pooling. + +Table 1: Evaluation with different structures on SUSTech1K valid + test set. We use ResNet-18 to extract image features by default. "L to C" indicates the results with point clouds as probe and silhouettes as gallery, and "C to L" indicates the reverse. + +
StructureModelL to CC to L
Rank-1Rank-3Rank-5Rank-1Rank-3Rank-5
One-streamResNet2.605.617.9914.9425.9231.87
Asymmetric FCResNet3.087.2310.5216.4329.1335.90
Image-pointGCN [9]12.2025.0133.2314.3028.3836.62
PointNet [26]7.1116.6123.358.4019.2526.90
Point-Transformer [43]13.2526.7235.2915.4230.0438.70
Two-streamResNet30.3948.7757.3135.2654.4262.84
ResNet+CSPP32.1850.8559.2236.5455.7564.11
ResNet+HP45.4162.8969.7142.3059.6767.12
CLGait (ours)*53.2969.5475.5955.1271.2377.31
+ +i.e., ResNet + CSPP + HP. + +# 4.4 Comparative Results + +Following the cross-view and cross-modality evaluation protocol, we report the comparative results of the methods above in the Tab. 1. From the results, we can obtain the following observations: 1) CL-Gait demonstrates its superiority to all existing structures and methods, primarily due to the feature extraction capabilities of the two-stream structure. Additionally, the use of pre-training effectively mitigates modality discrepancy, further improving its performance. 2) Experiments under the two-stream structure can be considered as an ablation study, highlighting the important roles of CSPP and HP. Adding HP to the ResNet base improves the average rank-1 by approximately $11\%$ , and further incorporating CSPP increases the average rank-1 by about another $10\%$ . 3) The performance of the one-stream structure is poor, primarily because the one-stream structure uses an identical network to process both silhouettes and point clouds, failing to address the modality discrepancy between them. 4) Although the Asymmetric FC structure employs different FC layers for the two modalities, its performance remains poor. This indicates that modality-specific adaptations are best applied at the shallower levels of the network. Furthermore, during the training process, the asymmetric FC layers tend to cause the network to overfit quickly. 5) All methods utilizing the two-stream structure outperform those based on the image-point structure. This likely results from the depth images, obtained through projection, having data consistency and smaller modality discrepancy with silhouettes compared to point cloud inputs. + +# 4.5 Discussion + +3D Geometry Information. To evaluate the effectiveness of depth information for cross-modality gait recognition, we compare another type of data as input for the second modality, e.g., LiDAR silhouettes. LiDAR silhouettes are obtained by range-view projection of point cloud sets and contains no 3D geometry information. The results are shown in Fig. 5. When 3D geometry information is not + +![](images/8b2dbdb56ab34ac339b7fc996661c87b045cb2b68bce392dd2a453f4c5df972d.jpg) +Fig. 5: Comparison of different input forms of LiDAR data. The results show that the projected and interpolated depth from point cloud works best for the cross-modality matching. This indicates that the 3D geometry information are essential. For each input form, the best-performing model is utilized. + +included, the performance of LiDAR silhouettes is much lower. This is because the camera has a much higher resolution than LiDAR, so the silhouettes from the camera can have more details. Integrating depth information can improve average rank-1 accuracy from $40.59\%$ to $54.21\%$ , validating the necessity and effectiveness of 3D information for cross-modality gait recognition. One important reason may be that the 3D information contained in the point cloud can compensate for the difference in viewing angles to some extent. Furthermore, we include the performance of direct point cloud input in Fig. 5 to compare with depth images and illustrate the input form of point clouds. + +Contrastive Pre-training Datasets. To make up for the absence of paired camera-LiDAR data, we further propose a large-scale data generation strategy. Camera silhouette and LiDAR point data are generated from real-world RGB images for contrastive pre-training on such paired pseudo data. To valid its performance, we construct training sets from two different datasets, including LIP and HITSZ-VCM. These two datasets are commonly used for camera-based person segmentation and re-identification and contain single still RGB images. To ensure fair comparisons, we also conduct experiments on SUSTech1K training set, which contains real paired point clouds and camera silhouettes. + +As shown in Tab. 2, when the amount of data is similar using data generated from the LIP dataset, the performance of pre-training on generated pseudo data is comparable to that of real data. When pre-trained on a larger scale of synthetic pseudo data generated from the HITSZ-VSM dataset, the accuracy from LiDAR to camera improves and exceeds that on the real data, proving the effectiveness of our data generation strategy for contrastive pre-training. However, the accuracy from camera to LiDAR decreases a little, probably due to the difference between generated point clouds from estimated depth and real points from LiDAR. + +Temporal Modules. In order to aggregate the different frames in the sequence, temporal modules are used to fuse the features of multiple frames. We evaluate the effectiveness of different temporal modules in combination with our proposed CL-Gait, as shown in Tab. 3. We can observe that the simple temporal pooling structure demonstrates superiority over the other learning-based structure. This + +Table 2: Performance with different pre-training datasets. Using estimated depth for pre-training obtains close results compared with using real paired point-image dataset. + +
DatasetData AmountL to CC to L
Rank-1Rank-3Rank-5Rank-1Rank-3Rank-5
SUSTech1K120K51.6068.2274.4356.1272.1577.95
LIP50K51.0068.0973.9955.0771.5977.59
HITSZ-VCM770K53.2969.5475.5955.1271.2377.31
+ +Table 3: Performance of CL-Gait with different temporal modules. + +
FeatureL to CC to L
Rank-1Rank-3Rank-5Rank-1Rank-3Rank-5
Temporal Pooling [5]53.2969.5475.5955.1271.2377.31
LSTM [13]37.4055.0160.8939.4756.4162.72
Bi-LSTM [28]39.6355.5561.8339.8857.8263.71
Transformer [33]44.8061.3167.3647.1663.8369.11
+ +result indicates that temporal pooling can better fuse multi-frame features to match domain information more effectively. One reason may be that learning-based approaches such as LSTM and Transformer are more likely to overfit. + +# 5 Conclusion and Future work + +This paper presents the first research on cross-modality gait recognition with point clouds and silhouettes from RGB images. Firstly, We propose a cross-modality gait recognition framework, named CL-Gait, that utilizes a two-stream network for feature embedding of different modalities, i.e., camera and LiDAR. Moreover, to mitigate modality discrepancy, we propose a contrastive pre-training strategy along with a large-scale data generation method. It can generate large number of data pairs (silhouette and point cloud) based on monocular depth estimation and pre-train the model in a contrastive manner. Extensive experiments demonstrate that CL-Gait achieves commendable performance in cross-modality evaluation modes, proving the importance of maintaining modality consistency and the effectiveness of contrastive pre-training. This also highlights the potential of cross-modality gait recognition. + +CL-Gait has shown remarkable results, but we believe there remains significant room for improvement. The modality and resolution differences between point clouds and silhouettes, along with the inherent motion blur in point clouds, may be key factors impacting model performance. Therefore, densifying point clouds or developing point cloud encoders specifically for cross-modality gait recognition could enhance model performance. + +# Acknowledgement + +This work is supported by the National Key Research and Development Program of China under Grant (2018AAA0102803). + +# References + +1. Bobick, A.F., Davis, J.W.: The recognition of human movement using temporal templates. IEEE Transactions on Pattern Analysis and Machine Intelligence 23(3), 257-267 (2001) +2. Chao, H., He, Y., Zhang, J., Feng, J.: Gaitset: Regarding gait as a set for cross-view gait recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 8126-8133 (2019) +3. Choi, S., Lee, S., Kim, Y., Kim, T., Kim, C.: Hi-cmd: Hierarchical cross-modality disentanglement for visible-infrared person re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10257-10266 (2020) +4. Dai, P., Ji, R., Wang, H., Wu, Q., Huang, Y.: Cross-modality person re-identification with generative adversarial training. In: IJCAI. vol. 1, p. 6 (2018) +5. Fan, C., Liang, J., Shen, C., Hou, S., Huang, Y., Yu, S.: Opengait: Revisiting gait recognition towards better practicality. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9707-9716 (2023) +6. Fan, C., Peng, Y., Cao, C., Liu, X., Hou, S., Chi, J., Huang, Y., Li, Q., He, Z.: Gaitpart: Temporal part-based model for gait recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14225-14233 (2020) +7. Feng, Z., Lai, J., Xie, X.: Learning modality-specific representations for visible-infrared person re-identification. IEEE Transactions on Image Processing 29, 579-590 (2019) +8. Gong, K., Liang, X., Zhang, D., Shen, X., Lin, L.: Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 932-940 (2017) +9. Guo, W., Pan, Z., Liang, Y., Xi, Z., Zhong, Z.C., Feng, J., el al: Lidar-based person re-identification. arXiv preprint arXiv:2312.03033 (2023) +0. Hao, Y., Wang, N., Li, J., Gao, X.: Hsme: Hypersphere manifold embedding for visible thermal person re-identification. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 8385-8392 (2019) +1. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016) +2. He, R., Wu, X., Sun, Z., Tan, T.: Learning invariant deep representation for nir-vis face recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 31 (2017) +3. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural computation 9(8), 1735-1780 (1997) +4. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., el al: Scaling up visual and vision-language representation learning with noisy text supervision. In: International Conference on Machine Learning. pp. 4904-4916. PMLR (2021) +5. Li, X., Makihara, Y., Xu, C., Yagi, Y., Yu, S., Ren, M.: End-to-end model-based gait recognition. In: Proceedings of the Asian Conference on Computer Vision (2020) +6. Liang, J., Fan, C., Hou, S., Shen, C., Huang, Y., Yu, S.: Gaitedge: Beyond plain end-to-end gait recognition for better practicality. In: European Conference on Computer Vision. pp. 375-390. Springer (2022) + +17. Liao, R., Yu, S., An, W., Huang, Y.: A model-based gait recognition method with body pose and human prior knowledge. Pattern Recognition 98, 107069 (2020) +18. Lin, B., Liu, Y., Zhang, S.: Gaitmask: Mask-based model for gait recognition. In: BMVC. pp. 1-12 (2021) +19. Lin, B., Zhang, S., Wang, M., Li, L., Yu, X.: Caitgl: Learning discriminative global-local feature representations for gait recognition. arXiv preprint arXiv:2208.01380 (2022) +20. Lin, X., Li, J., Ma, Z., Li, H., Li, S., Xu, K., Lu, G., Zhang, D.: Learning modal-invariant and temporal-memory for video-based visible-infrared person re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20973–20982 (2022) +21. Ling, Y., Zhong, Z., Luo, Z., Rota, P., Li, S., Sebe, N.: Class-aware modality mix and center-guided metric learning for visible-thermal person re-identification. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 889-897 (2020) +22. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Spl: A skinned multi-person linear model. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 851-866 (2023) +23. Lu, Y., Wu, Y., Liu, B., Zhang, T., Li, B., Chu, Q., Yu, N.: Cross-modality person re-identification with shared-specific feature transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13379–13389 (2020) +24. Nguyen, D.T., Hong, H.G., Kim, K.W., Park, K.R.: Person recognition system based on a combination of body images from visible light and thermal cameras. Sensors 17(3), 605 (2017) +25. Park, H., Lee, S., Lee, J., Ham, B.: Learning by aligning: Visible-infrared person re-identification using cross-modal correspondences. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12046–12055 (2021) +26. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3D classification and segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 652-660 (2017) +27. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021) +28. Schuster, M., Paliwal, K.K.: Bidirectional recurrent neural networks. IEEE Transactions on Signal Processing 45(11), 2673-2681 (1997) +29. Sepas-Moghaddam, A., Etemad, A.: Deep gait recognition: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(1), 264-284 (2022) +30. Shen, C., Fan, C., Wu, W., Wang, R., Huang, G.Q., Yu, S.: Lidargait: Benchmarking 3D gait recognition with point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1054-1063 (2023) +31. Takemura, N., Makihara, Y., Muramatsu, D., Echigo, T., Yagi, Y.: Multi-view large population gait dataset and its performance evaluation for cross-view gait recognition. IPSJ Transactions on Computer Vision and Applications 10, 1-14 (2018) +32. Teepe, T., Khan, A., Gilg, J., Herzog, F., Hörmann, S., Rigoll, G.: Gaitgraph: Graph convolutional network for skeleton-based gait recognition. In: 2021 IEEE International Conference on Image Processing (ICIP). pp. 2314-2318. IEEE (2021) + +33. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in Neural Information Processing Systems 30 (2017) +34. Wang, C., Zhang, J., Pu, J., Yuan, X., Wang, L.: Chrono-gait image: A novel temporal template for gait recognition. In: Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part I 11. pp. 257-270. Springer (2010) +35. Wang, Y., Chao, W.L., Garg, D., Hariharan, B., Campbell, M., Weinberger, K.Q.: Pseudo-lidar from visual depth estimation: Bridging the gap in 3D object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8445-8453 (2019) +36. Wu, A., Zheng, W.S., Yu, H.X., Gong, S., Lai, J.: RGB-infrared cross-modality person re-identification. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 5380-5389 (2017) +37. Yang, J., Duan, J., Tran, S., Xu, Y., Chanda, S., Chen, L., Zeng, B., Chilimbi, T., Huang, J.: Vision-language pre-training with triple contrastive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15671-15680 (2022) +38. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. arXiv preprint arXiv:2401.10891 (2024) +39. Ye, M., Lan, X., Leng, Q.: Modality-aware collaborative learning for visible thermal person re-identification. In: Proceedings of the 27th ACM International Conference on Multimedia. pp. 347-355 (2019) +40. Ye, M., Lan, X., Leng, Q., Shen, J.: Cross-modality person re-identification via modality-aware collaborative ensemble learning. IEEE Transactions on Image Processing 29, 9387–9399 (2020) +41. Ye, M., Shen, J., J. Crandall, D., Shao, L., Luo, J.: Dynamic dual-attentive aggregation learning for visible-infrared person re-identification. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVII 16. pp. 229–247. Springer (2020) +42. Yu, S., Tan, D., Tan, T.: A framework for evaluating the effect of view angle, clothing and carrying condition on gait recognition. In: 18th International Conference on Pattern Recognition (ICPR'06). vol. 4, pp. 441-444. IEEE (2006) +43. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16259-16268 (2021) \ No newline at end of file diff --git a/2024/Camera-LiDAR Cross-modality Gait Recognition/images.zip b/2024/Camera-LiDAR Cross-modality Gait Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f491d7ad9ab65789ab241ac609c1fd55ca68fe3e --- /dev/null +++ b/2024/Camera-LiDAR Cross-modality Gait Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3afd31ddb932c8b694d38de21a32fd6d042bd054e62c764fbe2ca6ccbe47ce12 +size 344603 diff --git a/2024/Camera-LiDAR Cross-modality Gait Recognition/layout.json b/2024/Camera-LiDAR Cross-modality Gait Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6373e3c48352454ab8d7b031299626de2eaed9cd --- /dev/null +++ b/2024/Camera-LiDAR Cross-modality Gait Recognition/layout.json @@ -0,0 +1,8098 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 112, + 478, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 112, + 478, + 131 + ], + "spans": [ + { + "bbox": [ + 135, + 112, + 478, + 131 + ], + "type": "text", + "content": "Camera-LiDAR Cross-modality Gait Recognition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 137, + 150, + 477, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 150, + 477, + 175 + ], + "spans": [ + { + "bbox": [ + 137, + 150, + 477, + 175 + ], + "type": "text", + "content": "Wenxuan Guo" + }, + { + "bbox": [ + 137, + 150, + 477, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 137, + 150, + 477, + 175 + ], + "type": "text", + "content": ", Yingping Liang" + }, + { + "bbox": [ + 137, + 150, + 477, + 175 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 137, + 150, + 477, + 175 + ], + "type": "text", + "content": ", Zhiyu Pan" + }, + { + "bbox": [ + 137, + 150, + 477, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 150, + 477, + 175 + ], + "type": "text", + "content": ", Ziheng Xi" + }, + { + "bbox": [ + 137, + 150, + 477, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 150, + 477, + 175 + ], + "type": "text", + "content": ", Jianjiang Feng" + }, + { + "bbox": [ + 137, + 150, + 477, + 175 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 137, + 150, + 477, + 175 + ], + "type": "text", + "content": ", and Jie Zhou" + }, + { + "bbox": [ + 137, + 150, + 477, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 187, + 184, + 428, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 184, + 428, + 196 + ], + "spans": [ + { + "bbox": [ + 187, + 184, + 428, + 196 + ], + "type": "text", + "content": "1 Department of Automation, BNRist, Tsinghua University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 240, + 196, + 374, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 196, + 374, + 207 + ], + "spans": [ + { + "bbox": [ + 240, + 196, + 374, + 207 + ], + "type": "text", + "content": "2 Beijing Institute of Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 208, + 471, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 208, + 471, + 218 + ], + "spans": [ + { + "bbox": [ + 143, + 208, + 471, + 218 + ], + "type": "text", + "content": "{gwx22, pzy20, xizh21}@mails.tsinghua.edu.cn liangyingping@bit.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 237, + 219, + 378, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 219, + 378, + 228 + ], + "spans": [ + { + "bbox": [ + 237, + 219, + 378, + 228 + ], + "type": "text", + "content": "{jfeng, jzhou}@tsinghua.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 262, + 455, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 262, + 455, + 548 + ], + "spans": [ + { + "bbox": [ + 159, + 262, + 455, + 548 + ], + "type": "text", + "content": "Abstract. Gait recognition is a crucial biometric identification technique. Camera-based gait recognition has been widely applied in both research and industrial fields. LiDAR-based gait recognition has also begun to evolve most recently, due to the provision of 3D structural information. However, in certain applications, cameras fail to recognize persons, such as in low-light environments and long-distance recognition scenarios, where LiDARs work well. On the other hand, the deployment cost and complexity of LiDAR systems limit its wider application. Therefore, it is essential to consider cross-modality gait recognition between cameras and LiDARs for a broader range of applications. In this work, we propose the first cross-modality gait recognition framework between Camera and LiDAR, namely CL-Gait. It employs a two-stream network for feature embedding of both modalities. This poses a challenging recognition task due to the inherent matching between 3D and 2D data, exhibiting significant modality discrepancy. To align the feature spaces of the two modalities, i.e., camera silhouettes and LiDAR points, we propose a contrastive pre-training strategy to mitigate modality discrepancy. To make up for the absence of paired camera-LiDAR data for pre-training, we also introduce a strategy for generating data on a large scale. This strategy utilizes monocular depth estimated from single RGB images and virtual cameras to generate pseudo point clouds for contrastive pre-training. Extensive experiments show that the cross-modality gait recognition is very challenging but still contains potential and feasibility with our proposed model and pre-training strategy. To the best of our knowledge, this is the first work to address cross-modality gait recognition. The code and dataset are available at https://github.com/GWxuan/CL-Gait." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 559, + 453, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 559, + 453, + 571 + ], + "spans": [ + { + "bbox": [ + 160, + 559, + 453, + 571 + ], + "type": "text", + "content": "Keywords: Gait recognition " + }, + { + "bbox": [ + 160, + 559, + 453, + 571 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 559, + 453, + 571 + ], + "type": "text", + "content": " Cross-modality " + }, + { + "bbox": [ + 160, + 559, + 453, + 571 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 559, + 453, + 571 + ], + "type": "text", + "content": " Contrastive pre-training" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 593, + 230, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 593, + 230, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 593, + 230, + 605 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 620, + 482, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 620, + 482, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 620, + 482, + 646 + ], + "type": "text", + "content": "Gait recognition is a crucial long-range identification technology without physical contact, which has great advantages in privacy protection and cross-clothing" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 326, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 326, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 326, + 666 + ], + "type": "text", + "content": "* Equal contribution. † Corresponding author." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 302, + 213 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 302, + 213 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 302, + 213 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 302, + 213 + ], + "type": "image", + "image_path": "d4fdceced539d9b484f3e35912d492c854116eaa852e8b993eb3495d49d4afa7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 217, + 301, + 235 + ], + "lines": [ + { + "bbox": [ + 133, + 217, + 301, + 235 + ], + "spans": [ + { + "bbox": [ + 133, + 217, + 301, + 235 + ], + "type": "text", + "content": "(a) Camera-based gait recognition (top), and LiDAR-based gait recognition (bottom)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 482, + 213 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 482, + 213 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 482, + 213 + ], + "type": "image", + "image_path": "8bce0b8fd5723c2d202e639839223a620f367be28abaa338faaaa9b9012b1727.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 217, + 479, + 234 + ], + "lines": [ + { + "bbox": [ + 312, + 217, + 479, + 234 + ], + "spans": [ + { + "bbox": [ + 312, + 217, + 479, + 234 + ], + "type": "text", + "content": "(b) Our proposed cross-modality gait recognition between camera and LiDAR." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 247, + 482, + 293 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 293 + ], + "type": "text", + "content": "Fig. 1: Overview of single-modality and cross-modality gait recognition. Single-modality gait recognition takes data of one modality as input and searches within a gallery of the same data type. In contrast, cross-modality gait recognition processes two modalities and identify individuals within a gallery of a different modality." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 316, + 479, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 479, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 479, + 389 + ], + "type": "text", + "content": "recognition [2]. It identifies individuals based on behavioral characteristics extracted from walking sequences recorded by sensors. Gait recognition has been extensively applied in fields such as user identification, sports science, and public security. Most current gait recognition research employs cameras as sensors [29]. With the advancement of computer vision technology, camera-based gait recognition continues to achieve new progress [16-19, 22, 32]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 389, + 480, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 480, + 497 + ], + "type": "text", + "content": "While utilizing cameras for gait recognition is cost-effective, it faces limitations in certain applications, such as low-light environments, long-distance recognition, and scenarios requiring precise 3D perception. In low-light conditions, the available information in RGB images significantly diminishes [9], making it challenging to accurately detect pedestrians and segment their silhouettes. Furthermore, due to the decline in image clarity, camera-based systems struggle to distinguish the fine-grained features from a distance [5]. Additionally, accurately capturing the 3D movement of individuals, crucial for comprehensive gait analysis, is beyond the capability of traditional camera systems [30]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 616 + ], + "type": "text", + "content": "Recently, to address the challenges under low light environment faced by RGB cameras, some researches [21, 23, 25, 40] have focused on RGB-Infrared cross-modality person re-identification (ReID), achieving commendable performance. However, infrared cameras struggle to capture pedestrians wearing insulating or thick clothing. Besides, they are constrained by limited detection distance and resolution, as well as susceptibility to lighting conditions and weather. It also faces the risk of pedestrian privacy and safety, compared with gait recognition. These limitations highlight the need for exploring alternative or complementary technologies that can offer more robust solutions for person identification tasks across a variety of environmental conditions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "In recent years, LiDAR technology has been increasingly utilized [35] due to its impressive gains on 3D information. Beyond its mature applications in autonomous driving, LiDAR has also been applied in individual identification areas, such as gait recognition [30] and person re-identification [9] most recently." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "text", + "content": "W. Guo, Y. Liang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "Through point cloud data, LiDAR can provide precise 3D perception, aiding in the acquisition of an individual's 3D geometric structure. From this, intrinsic characteristics of a person, including height, body shape, and gait, can be extracted. Moreover, LiDARs possess long-distance perception capabilities, unaffected by lighting conditions and complex backgrounds. And it offers enhanced privacy protection than camera [9]. The robustness to environmental variables and the ability to capture detailed biometric data make it an invaluable tool for advanced surveillance and security systems. However, compared to camera systems, LiDAR's deployment complexity and cost are somehow higher, making it unsuitable for large-scale deployment in general scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 243, + 482, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 316 + ], + "type": "text", + "content": "Considering the respective advantages of cameras and LiDARs, switching sensors in different scenarios to explore cross-modality gait recognition presents a worthwhile task for investigation. Cross-modality gait recognition offers an option with a broader application scope. Cameras can be utilized in general scenes with normal lighting, while LiDAR can be applied in specific scenarios, offering a complementary tool that enhances overall recognition capabilities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 322, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 482, + 418 + ], + "type": "text", + "content": "In this work, we present the first study on camera-LiDAR cross-modality gait recognition as shown in Fig. 1. This task poses significant challenges due to the need for matching data between 3D and 2D modalities. Based on our observations, there are substantial differences between point clouds and images. To be specific, for gait recognition, point clouds primarily focus on the 3D positioning of body parts, whereas images capture precise contour information of individuals. To address these challenges and explore the feasibility of cross-modality gait recognition, we propose a novel cross-modality framework, namely CL-Gait." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 424, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 556 + ], + "type": "text", + "content": "To extract gait features from both modalities, CL-Gait basically employs a two-stream network. Considering modality differences and data processing consistency, we first project and upsample the point clouds to obtain depth images with the same resolution as silhouettes, and use the projected depth images to train the network. Furthermore, to mitigate modality discrepancy, CL-Gait utilizes a contrastive silhouette-point pre-training approach (CSPP) to align the feature spaces of the two modalities. Pre-training requires aligned camera-LiDAR data under the same view, which is hard to obtain. To this end, we propose a new strategy of generating cross-modality gait data for pre-training. This strategy utilizes advances in monocular depth estimation and is able to generate large-scale dataset for pre-training using only single RGB images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "content": "Extensive experiments have revealed the following three insights: 1) CL-Gait achieves an average rank-1 accuracy of " + }, + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "inline_equation", + "content": "54.21\\%" + }, + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "content": ". The results show at least " + }, + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "inline_equation", + "content": "22.90\\%" + }, + { + "bbox": [ + 130, + 563, + 482, + 647 + ], + "type": "text", + "content": " improvement over the baseline models and demonstrate the significant potential of cross-modality gait recognition. 2) Utilizing large-scale depth images generated from point clouds as input is proved superior to using point clouds directly. 3) The strategy of contrastive pre-training on large-scale generated paired data mitigates the modality differences, contributing to performance improvement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 146, + 654, + 382, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 654, + 382, + 665 + ], + "spans": [ + { + "bbox": [ + 146, + 654, + 382, + 665 + ], + "type": "text", + "content": "To summarize, our main contributions are as follows:" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "text", + "content": "CL-Gait" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 236 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 138, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 479, + 151 + ], + "type": "text", + "content": "- To the best of our knowledge, this is the first work on cross-modality gait recognition. Extensive experiments show the potential of utilizing camera and LiDAR for gait recognition in challenging cross-modality scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 139, + 152, + 480, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 152, + 480, + 199 + ], + "spans": [ + { + "bbox": [ + 139, + 152, + 480, + 199 + ], + "type": "text", + "content": "- We analyse several network structures and conduct comparable experiments on their effectiveness for cross-modality gait recognition. Based on the two-stream network with better performance, our proposed CL-Gait is capable of matching camera silhouettes and LiDAR points." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 201, + 480, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 201, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 139, + 201, + 480, + 236 + ], + "type": "text", + "content": "- We propose a contrastive pre-training method to align the feature spaces of the two modalities. To make up for the absence of paired camera-LiDAR data, we further introduce a large-scale data generation strategy." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 255, + 237, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 237, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 237, + 268 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 281, + 479, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 479, + 304 + ], + "type": "text", + "content": "Gait Recognition. Based on the sensor used, exiting gait recognition methods can be categorized into camera-based and LiDAR-based methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 306, + 480, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 480, + 426 + ], + "type": "text", + "content": "Camera-based gait recognition has been extensively studied over the past decade. The majority of camera-based methods focus on extracting appearance features directly from images or videos [2, 6], adapting well to resolution reduction and achieving impressive performance [5]. These methods utilize silhouettes [16, 18, 19] or other gait templates [1, 34] for spatial feature extraction and temporal modeling. Additionally, some researchers have explored using estimated underlying structure of human body [15, 17, 32], such as 2D/3D pose and the SMPL model [22], as inputs. While theoretically robust against factors like carrying and clothing, these models often struggle at low resolutions, limiting their practicality in some real-world scenarios [5]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 426, + 481, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 481, + 544 + ], + "type": "text", + "content": "LiDAR-based gait recognition is an emerging field that utilizes precise 3D representations, point clouds, capturing complex motion patterns and the 3D structure of individual gaits. This approach is less susceptible to variations in lighting, clothing, and background, offering promising avenues for accurate gait recognition in diverse conditions. LidarGait [30] stands as a pioneering work in this field, introducing a multimodal gait recognition dataset named SUSTECH1K. This dataset is designed for evaluating the performance of gait recognition based on various sensors and has demonstrated the potential and practicality of LiDAR-based gait recognition. Additionally, some researchers have applied LidAR to the task of person re-identification [9], achieving impressive results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 593 + ], + "type": "text", + "content": "Camera-based and LiDAR-based methods have been proven practical, each with its unique advantages and disadvantages. Therefore, providing a compromise solution by addressing cross-modality gait recognition represents a valuable research direction, which has not been studied currently." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "Contrastive Pre-training. Contrastive pre-training is primarily applied in the field of vision-language models, aimed at aligning the feature spaces of visual and linguistic models. In recent years, inspired by the success of self-supervised learning within intra-modal tasks, researchers have begun to explore pre-training objectives for tasks involving multiple modalities, such as vision and language [37]. The pioneering work by CLIP [27] performs cross-modality" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 90, + 264, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 90, + 264, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 90, + 264, + 102 + ], + "type": "text", + "content": "W. Guo, Y. Liang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "contrastive pre-training on hundreds of millions of image-text pairs. CLIP [27] can generate a task-agnostic model that achieves surprisingly effective results. Subsequently, ALIGN [14] expands upon CLIP by utilizing a noisy dataset that covers more than a billion image-text pairs, further extending the capabilities of cross-modality pre-training. Contrastive pre-training plays a crucial role in cross-modality tasks. Therefore, we introduce contrastive pre-training into camera-LiDAR cross-modality gait recognition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 415 + ], + "type": "text", + "content": "RGB-IR Cross-modality Person ReID. Visible-infrared person ReID addresses the challenge of matching persons across different modalities, specifically between RGB and infrared cameras [36]. This task is particularly significant for scenarios with poor illumination, notably at night. With two primary datasets, SYSU-MM01 [36] and RegDB [24], RGB-IR cross-modality person ReID has been extensively studied, most of which focusing on metric learning, feature learning, and adversarial learning approaches. Metric learning [7, 10, 21, 40] and feature learning [25, 39, 41] methods aim to extract and align multimodal features into a common space for effective comparison, employing strategies like angle-based measurement and Euclidean constraints to bridge the modality gap. Notably, advancements in this area include the development of novel loss functions [21]. Concurrently, Generative Adversarial Networks (GANs) [3, 4, 23] have been instrumental in cross-modality ReID, enabling the transformation of images between RGB and IR while maintaining identity information, and fostering feature learning through adversarial and disentanglement strategies. Despite their potential, GAN-based methods face challenges such as high computational demands and the occasional production of low-quality images that may negatively impact ReID performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 415, + 481, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 481, + 464 + ], + "type": "text", + "content": "While RGB-IR person ReID mainly relies on appearance for distinction, which is a short-term feature and may raise privacy concerns. In contrast, cross-modality gait recognition focuses on shape and gait of individuals, which are time-invariant features, and offers better protection of privacy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 479, + 202, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 202, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 202, + 491 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 501, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 609 + ], + "type": "text", + "content": "In this work, we propose CL-Gait for cross-modality gait recognition between camera and LiDAR. CL-Gait employs a two-stream network for cross-modality feature embedding. The network utilizes modality-specific modules in the shallow layers and shared modules in the deeper layers, as illustrated in Fig. 2. Besides, CL-Gait adopts a contrastive learning strategy to align the feature spaces of the two modalities, mitigating modality differences, as shown in Fig. 3. The backbone obtained from pre-training is used to initialize the cross-modality embedding network. To facilitate pre-training on large-scale data, we also propose a method for generating cross-modality gait data, as shown in Fig. 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 624, + 342, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 342, + 636 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 342, + 636 + ], + "type": "text", + "content": "3.1 Cross-modality Network of CL-Gait" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "Two-Stream Backbone. The CL-Gait framework processes two distinct data modalities: 2D silhouette sequences and 3D point cloud sequences, optimizing" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "text", + "content": "CL-Gait" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 114, + 474, + 249 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 474, + 249 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 474, + 249 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 474, + 249 + ], + "type": "image", + "image_path": "e9666d6768e7918a5e7b504bc5974f00b5b163fa618e56ec24222bea9720ade3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 259, + 482, + 293 + ], + "lines": [ + { + "bbox": [ + 131, + 259, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 131, + 259, + 482, + 293 + ], + "type": "text", + "content": "Fig. 2: The cross-modality network of our proposed CL-Gait. It employs a two-stream architecture that encodes sequences from two modalities into a consistent feature space. HP stands for horizontal pooling, and TP represents temporal pooling." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 316, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 411 + ], + "type": "text", + "content": "cross-modality feature extracting. Inspired by [30], we first project and interpolate point clouds into depth images. For each modality, we apply modality-specific convolution layers (Conv-S for silhouettes and Conv-P for depth images generated from point clouds) before projecting the features into a shared embedding space, as shown in Fig. 2. Following [36], the two-stream backbone is modified from ResNet-18, with its first layer altered to be the modality-specific convolution layers, and the subsequent layers serving as shared convolution layers. For silhouettes, we define the processing as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 279, + 419, + 481, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 419, + 481, + 433 + ], + "spans": [ + { + "bbox": [ + 279, + 419, + 481, + 433 + ], + "type": "interline_equation", + "content": "F _ {I} = L _ {S} (I), \\tag {1}", + "image_path": "1b1e9a6808016da4e122ae717ffa19510e657bf8d154aa8ac996737d83bbc05b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 440, + 480, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 480, + 464 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 440, + 480, + 464 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 440, + 480, + 464 + ], + "type": "text", + "content": " represents the silhouette and " + }, + { + "bbox": [ + 130, + 440, + 480, + 464 + ], + "type": "inline_equation", + "content": "L_{S}" + }, + { + "bbox": [ + 130, + 440, + 480, + 464 + ], + "type": "text", + "content": " denotes the silhouette-specific convolution layer, which produces the feature map " + }, + { + "bbox": [ + 130, + 440, + 480, + 464 + ], + "type": "inline_equation", + "content": "F_{I}" + }, + { + "bbox": [ + 130, + 440, + 480, + 464 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 464, + 481, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 481, + 500 + ], + "type": "text", + "content": "Considering the modality differences and the need for consistency in data processing, we first project and interpolate point clouds to generate depths with the same resolution as silhouettes. And the processing is as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 266, + 508, + 480, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 508, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 266, + 508, + 480, + 521 + ], + "type": "interline_equation", + "content": "F _ {P} = L _ {P} (\\operatorname {P I} (P)), \\tag {2}", + "image_path": "fa4ec8a52a460bfd60ae475f40f100bf50ea922075e38f6e8ac502de1237b6c4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 529, + 480, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 480, + 564 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 529, + 480, + 564 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 529, + 480, + 564 + ], + "type": "text", + "content": " denotes the point cloud, PI represents the projection and interpolation process, and " + }, + { + "bbox": [ + 130, + 529, + 480, + 564 + ], + "type": "inline_equation", + "content": "L_{P}" + }, + { + "bbox": [ + 130, + 529, + 480, + 564 + ], + "type": "text", + "content": " is the depth-specific convolution layer, which produces the feature map " + }, + { + "bbox": [ + 130, + 529, + 480, + 564 + ], + "type": "inline_equation", + "content": "F_{P}" + }, + { + "bbox": [ + 130, + 529, + 480, + 564 + ], + "type": "text", + "content": " for the point cloud." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "text", + "content": "After extracting modality-specific features " + }, + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "inline_equation", + "content": "F_{P}" + }, + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "inline_equation", + "content": "F_{S}" + }, + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "text", + "content": ", we aim to unify these features within a consistent feature space. To achieve this, we employ a shared convolutional layer, denoted as " + }, + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{sh}}" + }, + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "text", + "content": ", to process both " + }, + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "inline_equation", + "content": "F_{P}" + }, + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "inline_equation", + "content": "F_{S}" + }, + { + "bbox": [ + 130, + 564, + 481, + 601 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 220, + 609, + 481, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 609, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 220, + 609, + 481, + 622 + ], + "type": "interline_equation", + "content": "F ^ {\\prime} = L _ {\\mathrm {s h}} \\left(F _ {m}\\right), \\text {w h e r e} F _ {m} = F _ {P} \\text {o r} F _ {S}, \\tag {3}", + "image_path": "0fb8b9348f92a3b45759e18a437ccf8ede8816ba5e57225453e0b8aab18c7807.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 323, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 323, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 323, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 629, + 323, + 641 + ], + "type": "inline_equation", + "content": "F^{\\prime}" + }, + { + "bbox": [ + 130, + 629, + 323, + 641 + ], + "type": "text", + "content": " is the output feature map of " + }, + { + "bbox": [ + 130, + 629, + 323, + 641 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{sh}}" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Temporal and Spatial Pooling. In our network, we follow commonly used strategies in gait recognition for temporal and spatial aggregation, specifically" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "text", + "content": "W. Guo, Y. Liang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "through temporal and horizontal pooling [5]. These techniques are essential for effectively summarizing and interpreting the dynamic and spatial aspects of gait data. Horizontal pooling divides the input feature map horizontally into " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "N_{H}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " parts, with each part being aggregated into a single feature vector:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 237, + 172, + 481, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 172, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 237, + 172, + 481, + 186 + ], + "type": "interline_equation", + "content": "\\mathrm {H P}: \\mathbb {R} ^ {T \\times H \\times W \\times C} \\rightarrow \\mathbb {R} ^ {T \\times N _ {H} \\times C}, \\tag {4}", + "image_path": "5459496e79b255ae40cb3a588379265074cb18d56c807ea271ab57195b338779.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 479, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 479, + 218 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 479, + 218 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 194, + 479, + 218 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 194, + 479, + 218 + ], + "type": "text", + "content": " denotes the sequence length, " + }, + { + "bbox": [ + 130, + 194, + 479, + 218 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 194, + 479, + 218 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 194, + 479, + 218 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 194, + 479, + 218 + ], + "type": "text", + "content": " represent the height and width of the feature map, respectively, and " + }, + { + "bbox": [ + 130, + 194, + 479, + 218 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 194, + 479, + 218 + ], + "type": "text", + "content": " indicates the dimension of the features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 219, + 479, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 219, + 479, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 219, + 479, + 243 + ], + "type": "text", + "content": "Temporal pooling allows for the aggregation of features from variable-length sequences, enabling the model to capture the final sequence-level representation:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 247, + 254, + 481, + 267 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 254, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 247, + 254, + 481, + 267 + ], + "type": "interline_equation", + "content": "\\mathrm {T P}: \\mathbb {R} ^ {T \\times N _ {H} \\times C} \\rightarrow \\mathbb {R} ^ {N _ {H} \\times C}. \\tag {5}", + "image_path": "2f9543deef6df1d72789f3653c8c37b941982c8f5530f2f0607e740739046f01.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 272, + 480, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 480, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 480, + 367 + ], + "type": "text", + "content": "Therefore, we obtain " + }, + { + "bbox": [ + 130, + 272, + 480, + 367 + ], + "type": "inline_equation", + "content": "N_{H}" + }, + { + "bbox": [ + 130, + 272, + 480, + 367 + ], + "type": "text", + "content": " feature vectors, which are further mapped to the metric space using " + }, + { + "bbox": [ + 130, + 272, + 480, + 367 + ], + "type": "inline_equation", + "content": "N_{H}" + }, + { + "bbox": [ + 130, + 272, + 480, + 367 + ], + "type": "text", + "content": " independent Multi-Layer Perceptrons. These mapped vectors are then concatenated to form the final sequence-level feature embedding. Training and Inference. The cross-modality network of CL-Gait is trained with cross-modality triplet loss and cross-entropy loss. We have modified the triplet loss to accommodate cross-modality feature learning, where each modality is alternately used as the anchor, while the corresponding positive and negative samples are selected from the other modality, formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 172, + 376, + 481, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 376, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 172, + 376, + 481, + 399 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c r o s s - t r i p l e t}} = \\frac {1}{2} \\left(\\mathcal {L} _ {\\text {t r i p l e t}} \\left(P, I _ {\\text {p o s}}, I _ {\\text {n e g}}\\right) + \\mathcal {L} _ {\\text {t r i p l e t}} \\left(I, P _ {\\text {p o s}}, P _ {\\text {n e g}}\\right)\\right). \\tag {6}", + "image_path": "02f76de1a420cd95f03935b67becada6669dfbcb652d02d5ec49f7ffe23d0cde.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 406, + 460, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 406, + 460, + 418 + ], + "spans": [ + { + "bbox": [ + 131, + 406, + 460, + 418 + ], + "type": "text", + "content": "The training loss is a combination of both losses with a hyperparameter " + }, + { + "bbox": [ + 131, + 406, + 460, + 418 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 406, + 460, + 418 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 255, + 429, + 480, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 429, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 255, + 429, + 480, + 441 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {c r o s s - t r i p l e t}} + \\gamma \\mathcal {L} _ {\\mathrm {c e}}, \\tag {7}", + "image_path": "6224f1479a7f3296c3ded3efb88190ab6dec18ec096788567a2ccb9cec34ab5b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 449, + 479, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 479, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 479, + 472 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 449, + 479, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ce}}" + }, + { + "bbox": [ + 130, + 449, + 479, + 472 + ], + "type": "text", + "content": " is the cross-entropy loss. During inference, the similarity between samples from different modalities is measured using the Euclidean distance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 491, + 369, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 491, + 369, + 503 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 369, + 503 + ], + "type": "text", + "content": "3.2 Contrastive Silhouette-point Pre-training" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 510, + 480, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 480, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 480, + 605 + ], + "type": "text", + "content": "Based on our observations in gait recognition tasks, the significant modality difference between 3D point clouds and 2D images could be a critical factor affecting model performance. To be specific, point clouds focus more on the 3D positioning of body parts, whereas images concentrate on the contour information of the individual. It's crucial to establish connections between the distinct information focused on by each modality. Inspired by CLIP [27], we propose a contrastive silhouette-point pre-training (CSPP) strategy to align the feature spaces of both modalities in the convolution-based encoders, as shown in Fig. 3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "text", + "content": "The pre-training process does not require identity labels from the samples for supervision. Trained on paired single-view data from aligned camera and LiDAR, the pre-training could make the model focus on learning a robust representation that bridges the gap between the modalities without direct identity-based guidance, and enhance the performance of the cross-modality network." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "CL-Gait" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 114, + 471, + 223 + ], + "blocks": [ + { + "bbox": [ + 145, + 114, + 471, + 223 + ], + "lines": [ + { + "bbox": [ + 145, + 114, + 471, + 223 + ], + "spans": [ + { + "bbox": [ + 145, + 114, + 471, + 223 + ], + "type": "image", + "image_path": "e512b73a4f3d63db19719fb49c70ff0b30d4faa444d5ddae1defc963ad4533a6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 231, + 482, + 285 + ], + "lines": [ + { + "bbox": [ + 130, + 231, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 482, + 285 + ], + "type": "text", + "content": "Fig. 3: The contrastive silhouette-point pre-training (CSPP) approach of CL-Gait. Paired silhouettes and point clouds are taken as inputs, and the backbone of the cross-modality network is pre-trained with contrastive learning loss to align the feature spaces of the two modalities. The pre-training process does not require identity labels of the samples for supervision." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 308, + 479, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 479, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 479, + 378 + ], + "type": "text", + "content": "Pre-training Process. For pre-training, the input shifts from sequence data to numerous pairs of point clouds and silhouettes. These pairs are defined as data collected from the same individual at the same moment, encompassing both modalities. It ensures that the network learns to reconcile the inherent differences between the 3D and 2D representations, focusing on the alignment of features that accurately reflect the same gait patterns across modalities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 380, + 480, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 480, + 426 + ], + "type": "text", + "content": "The paired data is processed by the two-stream backbone and spatial pooling operations to obtain paired feature embeddings, i.e., " + }, + { + "bbox": [ + 130, + 380, + 480, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\{S_i | i = 1, 2, \\dots, N\\}" + }, + { + "bbox": [ + 130, + 380, + 480, + 426 + ], + "type": "text", + "content": " for silhouettes and " + }, + { + "bbox": [ + 130, + 380, + 480, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{P_i | i = 1, 2, \\dots, N\\}" + }, + { + "bbox": [ + 130, + 380, + 480, + 426 + ], + "type": "text", + "content": " for point clouds. They are then utilized to compute the loss for contrastive representation learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 427, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 482, + 511 + ], + "type": "text", + "content": "Contrastive Learning loss. The cornerstone of our pre-training is the contrastive learning loss [27]. Given the paired feature embeddings " + }, + { + "bbox": [ + 130, + 427, + 482, + 511 + ], + "type": "inline_equation", + "content": "S = \\{S_{i} | i = 1, 2, \\dots, N\\}" + }, + { + "bbox": [ + 130, + 427, + 482, + 511 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 427, + 482, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{P_{i} | i = 1, 2, \\dots, N\\}" + }, + { + "bbox": [ + 130, + 427, + 482, + 511 + ], + "type": "text", + "content": ", there are " + }, + { + "bbox": [ + 130, + 427, + 482, + 511 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 130, + 427, + 482, + 511 + ], + "type": "text", + "content": " possible (silhouettes, point clouds) pairings. CL-Gait is trained to learn a multi-modality embedding space by maximizing the cosine similarity of the feature embeddings of the " + }, + { + "bbox": [ + 130, + 427, + 482, + 511 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 427, + 482, + 511 + ], + "type": "text", + "content": " real pairs while minimizing the cosine similarity of the embeddings of the " + }, + { + "bbox": [ + 130, + 427, + 482, + 511 + ], + "type": "inline_equation", + "content": "N^2 - N" + }, + { + "bbox": [ + 130, + 427, + 482, + 511 + ], + "type": "text", + "content": " incorrect pairings. The cosine similarity matrix is calculated as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 249, + 516, + 480, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 516, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 249, + 516, + 480, + 531 + ], + "type": "interline_equation", + "content": "\\mathcal {M} = N _ {\\mathrm {L} 2} (\\mathcal {S}) \\times N _ {\\mathrm {L} 2} (\\mathcal {P}) ^ {T}, \\tag {8}", + "image_path": "7a8a818797be294a02b10cd7d7100e2c607f7d61059424d643ff5f011b39df2a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 536, + 482, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 572 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 572 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 536, + 482, + 572 + ], + "type": "inline_equation", + "content": "\\mathcal{S} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 130, + 536, + 482, + 572 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 536, + 482, + 572 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 130, + 536, + 482, + 572 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 536, + 482, + 572 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 130, + 536, + 482, + 572 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 536, + 482, + 572 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{L2}}(\\cdot)" + }, + { + "bbox": [ + 130, + 536, + 482, + 572 + ], + "type": "text", + "content": " denotes L2 normalization, and " + }, + { + "bbox": [ + 130, + 536, + 482, + 572 + ], + "type": "inline_equation", + "content": "(\\cdot)^T" + }, + { + "bbox": [ + 130, + 536, + 482, + 572 + ], + "type": "text", + "content": " denotes the transpose of the matrix. Then, a symmetric cross entropy loss over these similarity scores is optimized:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 220, + 578, + 480, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 578, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 220, + 578, + 480, + 600 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n}} = \\frac {1}{2} \\left(\\mathcal {L} _ {\\mathrm {c e}} (\\mathcal {M}, G) + \\mathcal {L} _ {\\mathrm {c e}} \\left(\\mathcal {M} ^ {T}, G\\right)\\right), \\tag {9}", + "image_path": "5d3c7663937097518edf155d0e83b969b0d012b8430949329d7312ccfc1c3075.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "G = \\{1,2,\\dots,N\\}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " represents the labels used for calculating the cross-entropy loss, indicating that pairs corresponding to the similarities on the diagonal of matrices " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^T" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " are considered positive samples, while all others are treated as negative samples. This approach ensures that the network learns to distinguish between matching and non-matching modality pairs effectively." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "text", + "content": "W. Guo, Y. Liang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 112, + 473, + 242 + ], + "blocks": [ + { + "bbox": [ + 140, + 112, + 473, + 242 + ], + "lines": [ + { + "bbox": [ + 140, + 112, + 473, + 242 + ], + "spans": [ + { + "bbox": [ + 140, + 112, + 473, + 242 + ], + "type": "image", + "image_path": "16e78f6ce5709cbd6ca4f6dc3707c941667c4b0581186fbdfdfc3d3793683077.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 250, + 480, + 283 + ], + "lines": [ + { + "bbox": [ + 131, + 250, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 131, + 250, + 480, + 283 + ], + "type": "text", + "content": "Fig. 4: Illustration of paired gait data generation from single RGB images for contrastive pre-training. The quality of the synthesized data is comparable to the real-world data, making it possible to synthesize large-scale pre-training data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 305, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 482, + 426 + ], + "type": "text", + "content": "In the process described above, each frame of point cloud or silhouette, after passing through the pre-trained network, yields a feature embedding of size " + }, + { + "bbox": [ + 130, + 305, + 482, + 426 + ], + "type": "inline_equation", + "content": "(1,C)" + }, + { + "bbox": [ + 130, + 305, + 482, + 426 + ], + "type": "text", + "content": ", indicating global pooling is applied to the output feature map of the backbone. For practical application and to align each local feature embedding, horizontal pooling is used to obtain features of size " + }, + { + "bbox": [ + 130, + 305, + 482, + 426 + ], + "type": "inline_equation", + "content": "(N,C)" + }, + { + "bbox": [ + 130, + 305, + 482, + 426 + ], + "type": "text", + "content": " for each frame. During the computation of the contrastive learning loss, pairs of features from the same individual, moment, and part across two modalities are considered positive samples, with all others being negative. Through our practice, we have found that adopting a strategy of aligning local features during the pre-training phase is more suitable for gait recognition tasks, leading to performance improvements." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 441, + 350, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 441, + 350, + 453 + ], + "spans": [ + { + "bbox": [ + 131, + 441, + 350, + 453 + ], + "type": "text", + "content": "3.3 Generation of Pre-training Gait Data" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 459, + 482, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 459, + 482, + 531 + ], + "spans": [ + { + "bbox": [ + 131, + 459, + 482, + 531 + ], + "type": "text", + "content": "Pre-training on real and large-scale data is challenging due to the high cost of acquiring paired RGB and point cloud data. To address this issue, we propose a method for generating pseudo data based on monocular depth estimation. As illustrated in Fig. 4, we use Depth Anything [38] to estimate dense depth " + }, + { + "bbox": [ + 131, + 459, + 482, + 531 + ], + "type": "inline_equation", + "content": "D \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 131, + 459, + 482, + 531 + ], + "type": "text", + "content": " from large-scale single RGB images. Then, utilizing a virtual camera with intrinsic " + }, + { + "bbox": [ + 131, + 459, + 482, + 531 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 459, + 482, + 531 + ], + "type": "text", + "content": ", the estimated depth is mapped to pseudo point clouds:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 248, + 536, + 481, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 536, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 248, + 536, + 481, + 552 + ], + "type": "interline_equation", + "content": "[ P ^ {\\prime}, 1 ] ^ {T} = K ^ {- 1} D [ u, v, 1 ] ^ {T}, \\tag {10}", + "image_path": "aeb40e308c22dc7e30a1a64077e80c3376b3823c8db40a63be97fe1797558389.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "P' \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " is the point coordinates, " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "K^{-1}" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " indicates the inverse of intrinsic " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "u, v" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " are the image pixel coordinates. The pseudo points " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "P'" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": ", after being downsampled by voxel grid into " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "P_{down}'" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": ", can be used for contrastive pretraining between camera silhouettes and LiDAR point clouds. To be specific, the downsampled points " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "P_{down}'" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " are projected back into the image pixel coordinate to obtain the depth images for contrastive pre-training. Compared to point clouds, RGB images are less costly to collect and easier to acquire, with many public datasets of pedestrian image already available. Our proposed method makes it feasible to synthesize large-scale data for contrastive pre-training." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "text", + "content": "CL-Gait" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 204, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 204, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 204, + 152 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 481, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 481, + 304 + ], + "type": "text", + "content": "The SUSTech1K Dataset [30] is the first large-scale LiDAR-based gait recognition dataset collected by a LiDAR sensor and an RGB camera. The dataset contains 25,239 sequences from 1,050 subjects and covers many variations, including visibility, views, occlusions, clothing, carrying, and scenes. It is times-tamped frames for each modality of frames. The first single-modality 3D gait recognition framework is trained on SUSTech1K using point clouds. In our study, we explore new uses for the SUSTech1K dataset. Our approach involves both images and point clouds, which marks the first attempt of cross-modality gait recognition with this dataset. We also introduce a novel pre-training strategy based on paired images and point clouds from the dataset. This strategy utilizes contrastive learning for cross-modality matching between images and point clouds and improves the performance of our proposed cross-modality gait method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 304, + 481, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 481, + 460 + ], + "type": "text", + "content": "The HITSZ-VCM Dataset and LIP Dataset are large-scale datasets, which contains a large amount of person videos/images captured by cameras. In detail, the HITSZ-VCM dataset [20] contains 927 valid identities with image sequences. The LIP dataset [8] contains about 50,000 identities but with only one image for each. Totally, there are 251,452 RGB images in the HITSZ-VCM dataset and 50,000 in the LIP dataset. We implement the pre-training strategy on these datasets containing only RGB images, to explore the impact of pre-training with out-of-domain data on this task. Given that there is no point cloud data paired with these images, we use monocular depth estimation to generate pseudo point clouds, as shown in Fig. 4. This greatly expands the amount of data available for pre-training at scale. This also improves the accuracy of cross-modality recognition between 2D and 3D spaces, which marks the effectiveness of our proposed contrastive pre-training strategy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 478, + 261, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 478, + 261, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 478, + 261, + 490 + ], + "type": "text", + "content": "4.2 Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 481, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 593 + ], + "type": "text", + "content": "Implementation Details. Following [30], all the camera-based silhouettes and LiDAR-based depth images are aligned and then resized into the resolution of 6464. The total number of iterations is set to 120,000 for pre-training and 60,000 for fine-tuning. The Adam optimizer is used to prevent the issue of gradient vanishing because of low-quality silhouettes. The triplet and cross-entropy loss weights are set to the same. All comparison methods are trained using the same training strategy as LidarGait [30]. The OpenGait codebase [5] is utilized to conduct all experiments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "content": "Evaluation Protocol. All experiments are conducted on SUSTech1K. Following LidarGait [30], the dataset is divided into three splits: a training set with 250 identities and 6,011 sequences, a validation set with 6,025 sequences from 250 unseen identities, and a test set with the remaining 550 identities and 13,203 sequences. The SUSTech1K dataset provides gait sequences from multiple viewpoints. Thus, we also adopt the cross-view evaluation protocol [31,42] used in" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "text", + "content": "W. Guo, Y. Liang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "CASIA-B and OUMVLP. During the test, the sequences in normal conditions are grouped into gallery sets, and the sequences in variant conditions are taken as probe sets. To accurately assess the performance of camera-LiDAR cross-modality gait recognition, we evaluate the results using point cloud data as the probe and silhouette data as the gallery, i.e., LiDAR to Camera (L to C). We also evaluate the results using silhouette data as the probe and point cloud data as the gallery, i.e., Camera to LiDAR (C to L)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 229, + 273, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 273, + 242 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 273, + 242 + ], + "type": "text", + "content": "4.3 Comparative Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 261, + 482, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 393 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 393 + ], + "type": "text", + "content": "We evaluate several commonly used structures for cross-modality modeling as baselines, which basically follows the setting of [36]. The baseline models include one-steam structure, asymmetric FC layer structure, image-point structure, and the proposed two-stream structure. We apply residual block in ResNet-18 [11] as the base convolution block for all the four structures. For them, the loss function are weighted cross-entropy loss and triplet loss [2], which is commonly used and relatively stable. And all of the hyper parameters are kept the same. As for the input of one-steam structure, asymmetric FC layer structure and two-stream structure, the point clouds are converted to three-channel colored depth images and adjusted to the resolution of " + }, + { + "bbox": [ + 130, + 261, + 482, + 393 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 130, + 261, + 482, + 393 + ], + "type": "text", + "content": " with zero padding. We introduce the four structures as following:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 395, + 482, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 395, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 395, + 482, + 443 + ], + "type": "text", + "content": "One-stream Structure is the most commonly used in vision tasks, where all parameters are shared in the whole network. Therefore, we organize silhouettes and depth images obtained from point clouds into the same dimensions before feeding them into the network." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 445, + 482, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 494 + ], + "type": "text", + "content": "Asymmetric FC Layer Structure is used in multi-domain tasks, for example, IDR [12] for VIS-NIR face recognition. This structure shares nearly all parameters except for the last FC layer. It assumes that the feature extraction for different modalities can be same and adaptation is achieved in the backbone." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 495, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 556 + ], + "type": "text", + "content": "Image-point Structure directly uses point clouds as input, rather than projected depth images. It employs two types of encoders: an image encoder for the silhouette input and a point cloud encoder for the point cloud input. Under this structure, we implement three commonly used point cloud encoders: GCN [9], PointNet [26], and Point-Transformer [43]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Two-stream Structure is commonly used in cross-modality matching tasks. It utilizes modality-specific modules in the shallow layers of the network and uniform modules with shared parameters in the deeper layers. Based on the two-stream structure, our proposed CL-Gait is equipped with a horizontal pooling module (HP) and a contrastive silhouette-point pre-training strategy (CSPP). To show the effectiveness of the HP module and CSPP strategy of CL-Gait, we implement three additional networks: 1) a two-stream backbone without CSPP, using global pooling instead of HP; 2) a two-stream backbone without CSPP, using HP; 3) a two-stream backbone with CSPP, using global pooling." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "text", + "content": "CL-Gait" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 158, + 476, + 265 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Evaluation with different structures on SUSTech1K valid + test set. We use ResNet-18 to extract image features by default. \"L to C\" indicates the results with point clouds as probe and silhouettes as gallery, and \"C to L\" indicates the reverse." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 158, + 476, + 265 + ], + "lines": [ + { + "bbox": [ + 135, + 158, + 476, + 265 + ], + "spans": [ + { + "bbox": [ + 135, + 158, + 476, + 265 + ], + "type": "table", + "html": "
StructureModelL to CC to L
Rank-1Rank-3Rank-5Rank-1Rank-3Rank-5
One-streamResNet2.605.617.9914.9425.9231.87
Asymmetric FCResNet3.087.2310.5216.4329.1335.90
Image-pointGCN [9]12.2025.0133.2314.3028.3836.62
PointNet [26]7.1116.6123.358.4019.2526.90
Point-Transformer [43]13.2526.7235.2915.4230.0438.70
Two-streamResNet30.3948.7757.3135.2654.4262.84
ResNet+CSPP32.1850.8559.2236.5455.7564.11
ResNet+HP45.4162.8969.7142.3059.6767.12
CLGait (ours)*53.2969.5475.5955.1271.2377.31
", + "image_path": "cd055f6af92f16b244dffb120d4ed43e8dbe3475d569e723b5f478a452476639.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 141, + 266, + 254, + 278 + ], + "lines": [ + { + "bbox": [ + 141, + 266, + 254, + 278 + ], + "spans": [ + { + "bbox": [ + 141, + 266, + 254, + 278 + ], + "type": "text", + "content": "i.e., ResNet + CSPP + HP." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 300, + 266, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 266, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 266, + 313 + ], + "type": "text", + "content": "4.4 Comparative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 319, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 571 + ], + "type": "text", + "content": "Following the cross-view and cross-modality evaluation protocol, we report the comparative results of the methods above in the Tab. 1. From the results, we can obtain the following observations: 1) CL-Gait demonstrates its superiority to all existing structures and methods, primarily due to the feature extraction capabilities of the two-stream structure. Additionally, the use of pre-training effectively mitigates modality discrepancy, further improving its performance. 2) Experiments under the two-stream structure can be considered as an ablation study, highlighting the important roles of CSPP and HP. Adding HP to the ResNet base improves the average rank-1 by approximately " + }, + { + "bbox": [ + 130, + 319, + 482, + 571 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 130, + 319, + 482, + 571 + ], + "type": "text", + "content": ", and further incorporating CSPP increases the average rank-1 by about another " + }, + { + "bbox": [ + 130, + 319, + 482, + 571 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 319, + 482, + 571 + ], + "type": "text", + "content": ". 3) The performance of the one-stream structure is poor, primarily because the one-stream structure uses an identical network to process both silhouettes and point clouds, failing to address the modality discrepancy between them. 4) Although the Asymmetric FC structure employs different FC layers for the two modalities, its performance remains poor. This indicates that modality-specific adaptations are best applied at the shallower levels of the network. Furthermore, during the training process, the asymmetric FC layers tend to cause the network to overfit quickly. 5) All methods utilizing the two-stream structure outperform those based on the image-point structure. This likely results from the depth images, obtained through projection, having data consistency and smaller modality discrepancy with silhouettes compared to point cloud inputs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 586, + 214, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 214, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 214, + 598 + ], + "type": "text", + "content": "4.5 Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "3D Geometry Information. To evaluate the effectiveness of depth information for cross-modality gait recognition, we compare another type of data as input for the second modality, e.g., LiDAR silhouettes. LiDAR silhouettes are obtained by range-view projection of point cloud sets and contains no 3D geometry information. The results are shown in Fig. 5. When 3D geometry information is not" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "text", + "content": "W. Guo, Y. Liang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 111, + 465, + 209 + ], + "blocks": [ + { + "bbox": [ + 149, + 111, + 465, + 209 + ], + "lines": [ + { + "bbox": [ + 149, + 111, + 465, + 209 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 465, + 209 + ], + "type": "image", + "image_path": "8b2dbdb56ab34ac339b7fc996661c87b045cb2b68bce392dd2a453f4c5df972d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 217, + 482, + 262 + ], + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 262 + ], + "type": "text", + "content": "Fig. 5: Comparison of different input forms of LiDAR data. The results show that the projected and interpolated depth from point cloud works best for the cross-modality matching. This indicates that the 3D geometry information are essential. For each input form, the best-performing model is utilized." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 289, + 482, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 396 + ], + "type": "text", + "content": "included, the performance of LiDAR silhouettes is much lower. This is because the camera has a much higher resolution than LiDAR, so the silhouettes from the camera can have more details. Integrating depth information can improve average rank-1 accuracy from " + }, + { + "bbox": [ + 130, + 289, + 482, + 396 + ], + "type": "inline_equation", + "content": "40.59\\%" + }, + { + "bbox": [ + 130, + 289, + 482, + 396 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 289, + 482, + 396 + ], + "type": "inline_equation", + "content": "54.21\\%" + }, + { + "bbox": [ + 130, + 289, + 482, + 396 + ], + "type": "text", + "content": ", validating the necessity and effectiveness of 3D information for cross-modality gait recognition. One important reason may be that the 3D information contained in the point cloud can compensate for the difference in viewing angles to some extent. Furthermore, we include the performance of direct point cloud input in Fig. 5 to compare with depth images and illustrate the input form of point clouds." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 399, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 506 + ], + "type": "text", + "content": "Contrastive Pre-training Datasets. To make up for the absence of paired camera-LiDAR data, we further propose a large-scale data generation strategy. Camera silhouette and LiDAR point data are generated from real-world RGB images for contrastive pre-training on such paired pseudo data. To valid its performance, we construct training sets from two different datasets, including LIP and HITSZ-VCM. These two datasets are commonly used for camera-based person segmentation and re-identification and contain single still RGB images. To ensure fair comparisons, we also conduct experiments on SUSTech1K training set, which contains real paired point clouds and camera silhouettes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 508, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 482, + 604 + ], + "type": "text", + "content": "As shown in Tab. 2, when the amount of data is similar using data generated from the LIP dataset, the performance of pre-training on generated pseudo data is comparable to that of real data. When pre-trained on a larger scale of synthetic pseudo data generated from the HITSZ-VSM dataset, the accuracy from LiDAR to camera improves and exceeds that on the real data, proving the effectiveness of our data generation strategy for contrastive pre-training. However, the accuracy from camera to LiDAR decreases a little, probably due to the difference between generated point clouds from estimated depth and real points from LiDAR." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Temporal Modules. In order to aggregate the different frames in the sequence, temporal modules are used to fuse the features of multiple frames. We evaluate the effectiveness of different temporal modules in combination with our proposed CL-Gait, as shown in Tab. 3. We can observe that the simple temporal pooling structure demonstrates superiority over the other learning-based structure. This" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "text", + "content": "CL-Gait" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 168, + 147, + 440, + 200 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 2: Performance with different pre-training datasets. Using estimated depth for pre-training obtains close results compared with using real paired point-image dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 168, + 147, + 440, + 200 + ], + "lines": [ + { + "bbox": [ + 168, + 147, + 440, + 200 + ], + "spans": [ + { + "bbox": [ + 168, + 147, + 440, + 200 + ], + "type": "table", + "html": "
DatasetData AmountL to CC to L
Rank-1Rank-3Rank-5Rank-1Rank-3Rank-5
SUSTech1K120K51.6068.2274.4356.1272.1577.95
LIP50K51.0068.0973.9955.0771.5977.59
HITSZ-VCM770K53.2969.5475.5955.1271.2377.31
", + "image_path": "10d7fc2f656a9996a34c332a4e8896919fe9173842f6ae485f895db0f3f2e0ec.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 172, + 231, + 436, + 292 + ], + "blocks": [ + { + "bbox": [ + 168, + 209, + 444, + 220 + ], + "lines": [ + { + "bbox": [ + 168, + 209, + 444, + 220 + ], + "spans": [ + { + "bbox": [ + 168, + 209, + 444, + 220 + ], + "type": "text", + "content": "Table 3: Performance of CL-Gait with different temporal modules." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 172, + 231, + 436, + 292 + ], + "lines": [ + { + "bbox": [ + 172, + 231, + 436, + 292 + ], + "spans": [ + { + "bbox": [ + 172, + 231, + 436, + 292 + ], + "type": "table", + "html": "
FeatureL to CC to L
Rank-1Rank-3Rank-5Rank-1Rank-3Rank-5
Temporal Pooling [5]53.2969.5475.5955.1271.2377.31
LSTM [13]37.4055.0160.8939.4756.4162.72
Bi-LSTM [28]39.6355.5561.8339.8857.8263.71
Transformer [33]44.8061.3167.3647.1663.8369.11
", + "image_path": "9153d7be42e932b0e84d71af87090bf37c69c1ba177210bf9095947c0acc82a4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 312, + 481, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 481, + 350 + ], + "type": "text", + "content": "result indicates that temporal pooling can better fuse multi-frame features to match domain information more effectively. One reason may be that learning-based approaches such as LSTM and Transformer are more likely to overfit." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 364, + 323, + 377 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 323, + 377 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 323, + 377 + ], + "type": "text", + "content": "5 Conclusion and Future work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 387, + 482, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 530 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 530 + ], + "type": "text", + "content": "This paper presents the first research on cross-modality gait recognition with point clouds and silhouettes from RGB images. Firstly, We propose a cross-modality gait recognition framework, named CL-Gait, that utilizes a two-stream network for feature embedding of different modalities, i.e., camera and LiDAR. Moreover, to mitigate modality discrepancy, we propose a contrastive pre-training strategy along with a large-scale data generation method. It can generate large number of data pairs (silhouette and point cloud) based on monocular depth estimation and pre-train the model in a contrastive manner. Extensive experiments demonstrate that CL-Gait achieves commendable performance in cross-modality evaluation modes, proving the importance of maintaining modality consistency and the effectiveness of contrastive pre-training. This also highlights the potential of cross-modality gait recognition." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 531, + 482, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 482, + 603 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 482, + 603 + ], + "type": "text", + "content": "CL-Gait has shown remarkable results, but we believe there remains significant room for improvement. The modality and resolution differences between point clouds and silhouettes, along with the inherent motion blur in point clouds, may be key factors impacting model performance. Therefore, densifying point clouds or developing point cloud encoders specifically for cross-modality gait recognition could enhance model performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 619, + 241, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 619, + 241, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 619, + 241, + 632 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "This work is supported by the National Key Research and Development Program of China under Grant (2018AAA0102803)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "text", + "content": "W. Guo, Y. Liang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 139, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 139, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 481, + 172 + ], + "type": "text", + "content": "1. Bobick, A.F., Davis, J.W.: The recognition of human movement using temporal templates. IEEE Transactions on Pattern Analysis and Machine Intelligence 23(3), 257-267 (2001)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 481, + 205 + ], + "type": "text", + "content": "2. Chao, H., He, Y., Zhang, J., Feng, J.: Gaitset: Regarding gait as a set for cross-view gait recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 8126-8133 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 205, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 481, + 248 + ], + "type": "text", + "content": "3. Choi, S., Lee, S., Kim, Y., Kim, T., Kim, C.: Hi-cmd: Hierarchical cross-modality disentanglement for visible-infrared person re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10257-10266 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 249, + 481, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 249, + 481, + 271 + ], + "spans": [ + { + "bbox": [ + 138, + 249, + 481, + 271 + ], + "type": "text", + "content": "4. Dai, P., Ji, R., Wang, H., Wu, Q., Huang, Y.: Cross-modality person re-identification with generative adversarial training. In: IJCAI. vol. 1, p. 6 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 271, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 481, + 304 + ], + "type": "text", + "content": "5. Fan, C., Liang, J., Shen, C., Hou, S., Huang, Y., Yu, S.: Opengait: Revisiting gait recognition towards better practicality. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9707-9716 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 304, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 481, + 347 + ], + "type": "text", + "content": "6. Fan, C., Peng, Y., Cao, C., Liu, X., Hou, S., Chi, J., Huang, Y., Li, Q., He, Z.: Gaitpart: Temporal part-based model for gait recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14225-14233 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 347, + 481, + 380 + ], + "type": "text", + "content": "7. Feng, Z., Lai, J., Xie, X.: Learning modality-specific representations for visible-infrared person re-identification. IEEE Transactions on Image Processing 29, 579-590 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 380, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 481, + 424 + ], + "type": "text", + "content": "8. Gong, K., Liang, X., Zhang, D., Shen, X., Lin, L.: Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 932-940 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 424, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 424, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 424, + 481, + 446 + ], + "type": "text", + "content": "9. Guo, W., Pan, Z., Liang, Y., Xi, Z., Zhong, Z.C., Feng, J., el al: Lidar-based person re-identification. arXiv preprint arXiv:2312.03033 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 446, + 481, + 479 + ], + "type": "text", + "content": "0. Hao, Y., Wang, N., Li, J., Gao, X.: Hsme: Hypersphere manifold embedding for visible thermal person re-identification. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 8385-8392 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 479, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 481, + 512 + ], + "type": "text", + "content": "1. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 481, + 544 + ], + "type": "text", + "content": "2. He, R., Wu, X., Sun, Z., Tan, T.: Learning invariant deep representation for nir-vis face recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 31 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 544, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 481, + 567 + ], + "type": "text", + "content": "3. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural computation 9(8), 1735-1780 (1997)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 600 + ], + "type": "text", + "content": "4. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., el al: Scaling up visual and vision-language representation learning with noisy text supervision. In: International Conference on Machine Learning. pp. 4904-4916. PMLR (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 481, + 632 + ], + "type": "text", + "content": "5. Li, X., Makihara, Y., Xu, C., Yagi, Y., Yu, S., Ren, M.: End-to-end model-based gait recognition. In: Proceedings of the Asian Conference on Computer Vision (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 666 + ], + "type": "text", + "content": "6. Liang, J., Fan, C., Hou, S., Shen, C., Huang, Y., Yu, S.: Gaitedge: Beyond plain end-to-end gait recognition for better practicality. In: European Conference on Computer Vision. pp. 375-390. Springer (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "text", + "content": "CL-Gait" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "17. Liao, R., Yu, S., An, W., Huang, Y.: A model-based gait recognition method with body pose and human prior knowledge. Pattern Recognition 98, 107069 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 161 + ], + "type": "text", + "content": "18. Lin, B., Liu, Y., Zhang, S.: Gaitmask: Mask-based model for gait recognition. In: BMVC. pp. 1-12 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 163, + 480, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 163, + 480, + 195 + ], + "spans": [ + { + "bbox": [ + 133, + 163, + 480, + 195 + ], + "type": "text", + "content": "19. Lin, B., Zhang, S., Wang, M., Li, L., Yu, X.: Caitgl: Learning discriminative global-local feature representations for gait recognition. arXiv preprint arXiv:2208.01380 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 197, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 197, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 197, + 480, + 239 + ], + "type": "text", + "content": "20. Lin, X., Li, J., Ma, Z., Li, H., Li, S., Xu, K., Lu, G., Zhang, D.: Learning modal-invariant and temporal-memory for video-based visible-infrared person re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20973–20982 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 241, + 480, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 241, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 132, + 241, + 480, + 283 + ], + "type": "text", + "content": "21. Ling, Y., Zhong, Z., Luo, Z., Rota, P., Li, S., Sebe, N.: Class-aware modality mix and center-guided metric learning for visible-thermal person re-identification. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 889-897 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 285, + 480, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 318 + ], + "type": "text", + "content": "22. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Spl: A skinned multi-person linear model. In: Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pp. 851-866 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 319, + 480, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 480, + 361 + ], + "type": "text", + "content": "23. Lu, Y., Wu, Y., Liu, B., Zhang, T., Li, B., Chu, Q., Yu, N.: Cross-modality person re-identification with shared-specific feature transfer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13379–13389 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "text", + "content": "24. Nguyen, D.T., Hong, H.G., Kim, K.W., Park, K.R.: Person recognition system based on a combination of body images from visible light and thermal cameras. Sensors 17(3), 605 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 397, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 430 + ], + "type": "text", + "content": "25. Park, H., Lee, S., Lee, J., Ham, B.: Learning by aligning: Visible-infrared person re-identification using cross-modal correspondences. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12046–12055 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 431, + 480, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 480, + 463 + ], + "type": "text", + "content": "26. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3D classification and segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 652-660 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 464, + 480, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 480, + 507 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 480, + 507 + ], + "type": "text", + "content": "27. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 509, + 480, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 480, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 480, + 530 + ], + "type": "text", + "content": "28. Schuster, M., Paliwal, K.K.: Bidirectional recurrent neural networks. IEEE Transactions on Signal Processing 45(11), 2673-2681 (1997)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 532, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 480, + 553 + ], + "type": "text", + "content": "29. Sepas-Moghaddam, A., Etemad, A.: Deep gait recognition: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(1), 264-284 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 555, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 586 + ], + "type": "text", + "content": "30. Shen, C., Fan, C., Wu, W., Wang, R., Huang, G.Q., Yu, S.: Lidargait: Benchmarking 3D gait recognition with point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1054-1063 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 588, + 480, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 631 + ], + "type": "text", + "content": "31. Takemura, N., Makihara, Y., Muramatsu, D., Echigo, T., Yagi, Y.: Multi-view large population gait dataset and its performance evaluation for cross-view gait recognition. IPSJ Transactions on Computer Vision and Applications 10, 1-14 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "32. Teepe, T., Khan, A., Gilg, J., Herzog, F., Hörmann, S., Rigoll, G.: Gaitgraph: Graph convolutional network for skeleton-based gait recognition. In: 2021 IEEE International Conference on Image Processing (ICIP). pp. 2314-2318. IEEE (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 264, + 102 + ], + "type": "text", + "content": "W. Guo, Y. Liang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 522 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "33. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in Neural Information Processing Systems 30 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "type": "text", + "content": "34. Wang, C., Zhang, J., Pu, J., Yuan, X., Wang, L.: Chrono-gait image: A novel temporal template for gait recognition. In: Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part I 11. pp. 257-270. Springer (2010)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 237 + ], + "type": "text", + "content": "35. Wang, Y., Chao, W.L., Garg, D., Hariharan, B., Campbell, M., Weinberger, K.Q.: Pseudo-lidar from visual depth estimation: Bridging the gap in 3D object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8445-8453 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "text", + "content": "36. Wu, A., Zheng, W.S., Yu, H.X., Gong, S., Lai, J.: RGB-infrared cross-modality person re-identification. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 5380-5389 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 314 + ], + "type": "text", + "content": "37. Yang, J., Duan, J., Tran, S., Xu, Y., Chanda, S., Chen, L., Zeng, B., Chilimbi, T., Huang, J.: Vision-language pre-training with triple contrastive learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15671-15680 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "text", + "content": "38. Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Unleashing the power of large-scale unlabeled data. arXiv preprint arXiv:2401.10891 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "39. Ye, M., Lan, X., Leng, Q.: Modality-aware collaborative learning for visible thermal person re-identification. In: Proceedings of the 27th ACM International Conference on Multimedia. pp. 347-355 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "40. Ye, M., Lan, X., Leng, Q., Shen, J.: Cross-modality person re-identification via modality-aware collaborative ensemble learning. IEEE Transactions on Image Processing 29, 9387–9399 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 413, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 457 + ], + "type": "text", + "content": "41. Ye, M., Shen, J., J. Crandall, D., Shao, L., Luo, J.: Dynamic dual-attentive aggregation learning for visible-infrared person re-identification. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVII 16. pp. 229–247. Springer (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "text", + "content": "42. Yu, S., Tan, D., Tan, T.: A framework for evaluating the effect of view angle, clothing and carrying condition on gait recognition. In: 18th International Conference on Pattern Recognition (ICPR'06). vol. 4, pp. 441-444. IEEE (2006)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "text", + "content": "43. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16259-16268 (2021)" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 412, + 91, + 447, + 100 + ], + "type": "text", + "content": "CL-Gait" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/CamoTeacher_ Dual-Rotation Consistency Learning for Semi-Supervised Camouflaged Object Detection/e7b0e9c5-9645-43e1-9e04-1cc9804f36e6_content_list.json b/2024/CamoTeacher_ Dual-Rotation Consistency Learning for Semi-Supervised Camouflaged Object Detection/e7b0e9c5-9645-43e1-9e04-1cc9804f36e6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d573698315c7ec3bde38dd818cbe5e69f0abdb67 --- /dev/null +++ b/2024/CamoTeacher_ Dual-Rotation Consistency Learning for Semi-Supervised Camouflaged Object Detection/e7b0e9c5-9645-43e1-9e04-1cc9804f36e6_content_list.json @@ -0,0 +1,2004 @@ +[ + { + "type": "text", + "text": "CamoTeacher: Dual-Rotation Consistency Learning for Semi-Supervised Camouflaged Object Detection", + "text_level": 1, + "bbox": [ + 256, + 140, + 750, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xunfa Lai $^{1}$ , Zhiyu Yang $^{1}$ , Jie Hu $^{2}$ , Shengchuan Zhang $^{1*}$ , Liujuan Cao $^{1}$ , Guannan Jiang $^{2}$ , Zhiyu Wang $^{2}$ , Songan Zhang $^{3}$ , and Rongrong Ji $^{1}$", + "bbox": [ + 241, + 233, + 759, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University", + "bbox": [ + 238, + 276, + 761, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Contemporary Amperex Technology Co. Limited", + "bbox": [ + 331, + 305, + 671, + 318 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Shanghai Jiao Tong University", + "bbox": [ + 392, + 319, + 611, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Existing camouflaged object detection (COD) methods depend heavily on large-scale pixel-level annotations. However, acquiring such annotations is laborious due to the inherent camouflage characteristics of the objects. Semi-supervised learning offers a promising solution to this challenge. Yet, its application in COD is hindered by significant pseudo-label noise, both pixel-level and instance-level. We introduce CamoTeacher, a novel semi-supervised COD framework, utilizing Dual-Rotation Consistency Learning (DRCL) to effectively address these noise issues. Specifically, DRCL minimizes pseudo-label noise by leveraging rotation views' consistency in pixel-level and instance-level. First, it employs Pixel-wise Consistency Learning (PCL) to deal with pixel-level noise by reweighting the different parts within the pseudo-label. Second, Instance-wise Consistency Learning (ICL) is used to adjust weights for pseudo-labels, which handles instance-level noise. Extensive experiments on four COD benchmark datasets demonstrate that the proposed CamoTeacher not only achieves state-of-the-art compared with semi-supervised learning methods, but also rivals established fully-supervised learning methods. Our code will be available soon.", + "bbox": [ + 259, + 364, + 743, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Camouflaged object detection $\\cdot$ Semi-supervised learning", + "bbox": [ + 261, + 628, + 730, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 666, + 375, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Camouflaged object detection (COD) aims to identify objects that blend seamlessly into their environments including animals or artificial entities that possess protective coloring and have the ability to blend into their surroundings, a task complicated by low contrast, similar textures, and blurred boundaries [6,7,9,14, 17,24,43]. Unlike general object detection, COD is challenged by these factors, making detection significantly more difficult. Existing methods in COD heavily depend on extensive pixel-level annotated datasets, the creation of which incurs substantial human effort and cost, thus limiting COD's advancement.", + "bbox": [ + 212, + 696, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding authors", + "bbox": [ + 217, + 824, + 387, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2750ee8e1c66b13332dc42b66a5003268eb0fd34beaf0c10bc638b21a505fc9e.jpg", + "image_caption": [ + "Fig. 1: (a) Illustration of pixel-level and instance-level noise. Pixel-level noise refers to varying degrees of noise in different parts within an instance. Instance-level noise pertains to varying degrees of noise among different instances. (b) Visualization of performance trends on COD10K [7] under different percentages of labeled data." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 754, + 296 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To mitigate that, semi-supervised learning [13, 27, 36, 47, 49] emerges as a promising approach by leveraging both labeled and unlabeled data, yet its application in COD is not straightforward due to complex backgrounds and subtle object boundaries. The efficacy of semi-supervised learning in COD is significantly compromised by the presence of substantial noise in pseudo-labels. Our investigation into pseudo-label noise reveals two primary types: pixel-level noise, indicating variation within a single pseudo-label, and instance-level noise, showing variation across different pseudo-labels. This distinction is critical, as it informs our approach to refining pseudo-label quality for improved model training. (1) Pixel-level noise is characterized by inconsistent labeling within various parts of a pseudo-label. As demonstrated in Fig. 1(a), in the first row, the tail of the gecko is visually more camouflaged than its head. The pseudo-label generated by the SINet [7] is less accurate in its tail region (highlighted by the red box). This observation underscores the inappropriateness of uniformly treating all parts within a pseudo-label. (2) Instance-level noise refers to the variability in noise levels across distinct pseudo-labels. As shown in Fig. 1(a), the pseudo-labels of the third row are less accurate compared to that in the second row because the camouflaged objects in the third row is more difficult to detect. These discrepancies indicate that each pseudo-label contributes variably to model training, emphasizing the need for a nuanced approach to integrating pseudo-label information.", + "bbox": [ + 212, + 398, + 787, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the challenge of evaluating pseudo-label noise without ground truth (GT) for unlabeled data, we introduce two novel strategies based on pixel-wise inconsistency and instance-wise consistency across two rotation views. Specifically, for pixel-level noise, we observe that pixel-wise inconsistency—calculated by comparing pseudo-labels from two rotated views—mirrors the actual error relative to GT, as shown in Fig. 2(a). This relationship, illustrated by the polyline in Fig. 2(b), shows a positive correlation between mean pixel-wise inconsistency and the mean absolute error (MAE) across different parts. Consequently, areas", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "X.Lai et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/63d370dfa8c3be84598bc01d5d5dbc4b1143cf95d613c6f58a3611fef14511db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 147, + 321, + 195 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fabd694cf371834a6a1826740a1d54d3837808358b24c8abf6cfb8e75bcc608c.jpg", + "image_caption": [ + "Image" + ], + "image_footnote": [], + "bbox": [ + 245, + 195, + 321, + 241 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a5b2451f2405bab07b40156a647bba41cd7476d2e9886d789e54bd6e4bebee20.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 148, + 398, + 195 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/af04ce3cddbd74a8e254b7dc4d67dd4183c4d872cfa43eff974d5b8cc58c6a40.jpg", + "image_caption": [ + "GT", + "(a)" + ], + "image_footnote": [], + "bbox": [ + 321, + 195, + 398, + 241 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c4784a760d8bc2b32816fcb1b8219c072aea690d6810ec730a0aef7d1b4b21b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 148, + 475, + 195 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f683e8bc0d370e8f21acf430e32f8b21ef8b19169d28807a5c710bf5e524d592.jpg", + "image_caption": [ + "Real absolute error" + ], + "image_footnote": [], + "bbox": [ + 398, + 195, + 475, + 241 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b77dc97c7ce1f08852d2144c5126918ec180ab0bbdc385da7144288d635e7e7f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 477, + 148, + 552, + 195 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8f379e3f424a96bb1a6c75427f066387208eee2bad31904820091fe32671416f.jpg", + "image_caption": [ + "Pixel-wise" + ], + "image_footnote": [], + "bbox": [ + 477, + 195, + 552, + 241 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1ca49ad23aadb3d715b3a63ce54a1cece65637740934209f1e1cd137208dcedd.jpg", + "image_caption": [ + "(b)", + "Fig. 2: (a) The real absolute error is obtained by subtracting the pseudo-label from the ground truth, reflecting the actual noisy level. On the other hand, pixel-wise inconsistency is derived from the subtraction of pseudo-labels from two rotated views. Visually, pixel-wise inconsistency and real absolute error appear very similar. (b) We define the area within 20 pixels from the edge as the boundary according to GT. We calculate mean pixel-wise inconsistency (MPI) and mean absolute error (MAE) in background, foreground, and boundary. The pseudo-labels are obtained by our proposed model on the test dataset COD10K [7] at a $10\\%$ semi-supervised setting." + ], + "image_footnote": [], + "bbox": [ + 576, + 152, + 756, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with higher pixel-wise inconsistency are more prone to inaccuracies, suggesting a need for their de-emphasis during training. For instance-level noise, we determine that pseudo-labels with greater similarity across rotation views exhibit lower noise levels, as indicated in Fig. 3(a). A positive correlation between instance-wise consistency and the SSIM [39] computed by pseudo-labels and GT further supports this observation, depicted in Fig. 3(b). Thus, pseudo-labels demonstrating higher instance-wise consistency are likely of superior quality and should be prioritized in the learning process.", + "bbox": [ + 212, + 426, + 787, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the insights gained from these observations, we propose a semi-supervised camouflaged object detection framework called CamoTeacher, which incorporates a novel method known as Dual-Rotation Consistency Learning (DRCL). Specifically, DRCL operationalizes its strategy through two core components: Pixel-wise Consistency Learning (PCL) and Instance-wise Consistency Learning (ICL). PCL innovatively assigns variable weights to different parts within a pseudo-label, considering the pixel-wise inconsistency across diverse rotation view. Simultaneously, ICL adjusts the importance of individual pseudo-labels based on their instance-wise consistency, enabling a nuanced, noise-aware training process.", + "bbox": [ + 212, + 550, + 790, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We adopt SINet [7] as the base model to implement CamoTeacher and apply it to more classic COD models, i.e. CNN-based model SINet-v2 [6] and SegMaR [17] and Transformed-based model DTINet [24] and FSPNet [14]. Extensive experiments on four COD benchmark datasets, i.e. CAMO [21], CHAMELEON [33], COD10K [7], and NC4K [25], demonstrate that the proposed CamoTeacher not only achieves state-of-the-art compared with semi-supervised learning methods, but also rivals established fully-supervised learning methods. Specifically, as shown on Fig. 1(b), with only $20\\%$ labeled data, it nearly reached the performance of fully supervised model on COD10K.", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "CamoTeacher", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/eb5606f302aca6f3c031cf325c65857217b138eb00218453a2ee1e95046fde66.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 143, + 321, + 189 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/86512bfcad0262e90612a854a28984157791811722294cc37e5e97a91d44d01e.jpg", + "image_caption": [ + "Image" + ], + "image_footnote": [], + "bbox": [ + 245, + 190, + 321, + 236 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/16473e4f15ed9b7d3cd81140765cee16dc38133a1377a053e9ecad0455ea93c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 143, + 398, + 189 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/eb3e99a5a1d883bb3ba4cc364590f0ad6a2be5f3fbcd30de54af3e545bc7145d.jpg", + "image_caption": [ + "GT", + "(a)" + ], + "image_footnote": [], + "bbox": [ + 321, + 190, + 397, + 236 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bfa004632f7a3d0fae0ac3376c0f01e253a3d31bc818072604a355dd65131740.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 400, + 143, + 475, + 189 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cf497c818a3215c6e17ac635ee5c2eaa4375ee24090dfe2b81b84108771ad664.jpg", + "image_caption": [ + "Pseudo-label" + ], + "image_footnote": [], + "bbox": [ + 400, + 190, + 473, + 236 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/48b89a1690b82f7df298fd50507ebd46e4e6b6e11778e91bb9b236fd99916937.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 477, + 143, + 553, + 189 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/33ae46de3df3f755559a8a3a650bf1764c5d7efef8af9c4a73d8dc59c0b2782c.jpg", + "image_caption": [ + "Pseudo-label" + ], + "image_footnote": [], + "bbox": [ + 477, + 190, + 553, + 236 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f6964705870498f4b41264c2a2ed0ef0c919bd519a923a42178a41c8f7acc88e.jpg", + "image_caption": [ + "(b)", + "Fig. 3: (a) The more similar the pseudo-labels are between two rotated views, the lower the noise level and the higher the quality of the pseudo-labels. For example, in the first row, the pseudo-labels are more similar, indicating a lower noise level compared to the pseudo-labels in the second row. (b) The positive correlation between instance-wise consistency and the SSIM [39] of pseudo-labels and GT." + ], + "image_footnote": [], + "bbox": [ + 565, + 145, + 756, + 248 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 238, + 376, + 591, + 390 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the first end-to-end semi-supervised camouflaged object detection framework CamoTeacher, offering a simple yet effective baseline for future research in semi-supervised camouflaged object detection.", + "- To address the problem of large amounts of noise in pseudo-labels in semi-supervised camouflaged object detection, we proposed the Dual-Rotation Consistency Learning (DRCL), which includes Pixel-wise Consistency Learning (PCL) and Instance-wise Consistency Learning (ICL), allows for the adaptive adjustment of the contribution of pseudo-labels with varying quality, enabling the effective utilization of pseudo-label information.", + "- We conducted extensive experiments on COD benchmark datasets, and achieved significant improvements compared to the fully supervised setting." + ], + "bbox": [ + 225, + 402, + 805, + 569 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related works", + "text_level": 1, + "bbox": [ + 215, + 595, + 390, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Camouflaged Object Detection", + "text_level": 1, + "bbox": [ + 214, + 630, + 516, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Camouflaged Object Detection(COD) is different from traditional visual detection tasks as objects and backgrounds in traditional detection tasks have significant visual distinctions, whereas camouflaged objects are typically concealed within their surrounding environments [6, 7, 14, 17, 24, 29, 43]. In recent years, many models have been proposed on this task, which can be divided into CNN-based and Transformed-based. i) CNN-based models: SINet [7] is a simple yet effective baseline that views camouflage detection as searching and identification process. SINet-v2 [6] significantly improves performance by introducing a neighbor connection decoder and cascaded group-reversal attention design. SegMaR [17] firstly performs initial detection of camouflaged objects and then progressively enlarges the detected regions to achieve precise detection. FEDER [9] achieves impressive results by conducting feature decomposition and", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "X.Lai et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "edge reconstruction. ii) Transformed-based models: DTINet [24] proposes a dual-task interactive transformer to detect accurate position and detailed boundary. CamoFormer [43] presents masked separable attention to separate foreground and background regions. FSPNet [14] designs a feature shrinkage decoder hierarchically decode locality-enhanced neighboring transformer features. These outstanding works have achieved excellent results in COD, but they rely on large-scale pixel-level annotations, which are time-consuming and labor-intensive to obtain.", + "bbox": [ + 212, + 146, + 782, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In response to the reliance of full-supervised COD on expensive pixel-level annotations, some unsupervised/weakly-supervised approaches have emerged [10-12,46]. UCOS-DA [46] formulates the unsupervised camouflaged object segmentation as a source-free unsupervised domain adaptation task and designs a foreground-background-contrastive self-adversarial pipeline. SCOD [11] proposes the first weakly-supervised COD method with relabeled scribbles. WS-SAM [10] uses the provided sparse annotations as prompts for SAM [18] to generate segmentation masks to train the model. These methods reduce annotation costs but often exhibit lower performance compared to fully-supervised methods. Is it possible to leverage existing annotated data without additional labeling while ensuring accuracy? To address this issue, we propose a new benchmark that combines semi-supervised learning with camouflaged object detection.", + "bbox": [ + 212, + 267, + 784, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.2 Semi-Supervised Learning", + "text_level": 1, + "bbox": [ + 214, + 465, + 478, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In traditional supervised learning, training a high-performing model requires a substantial amount of labeled data. However, obtaining labeled data in practical applications is often costly and time-consuming. Consequently, the effective utilization of unlabeled data to improve model performance and generalization has become a critical concern. Semi-supervised learning addresses this challenge by leveraging unlabeled data to enhance the model's learning process [2,8,22,27,28,30,34,36,48], thereby improving its performance and generalization capability. Pseudo-labeling [3,35,37,38,40,42] and consistency regularization [13,19,20,23,41] are popular techniques in the field of semi-supervised learning. These methods have shown great efficacy when combined, producing impressive results.", + "bbox": [ + 212, + 486, + 784, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address the issue of heavy reliance on labeled data in camouflage object detection, we propose the first end-to-end semi-supervised camouflage object detection method. By integrating semi-supervised learning, our approach aims to enhance the performance of camouflage object detection models while significantly reducing the need for annotated data.", + "bbox": [ + 212, + 652, + 784, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 214, + 747, + 380, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Task Formulation", + "text_level": 1, + "bbox": [ + 214, + 773, + 406, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semi-supervised camouflaged object detection aims to leverage a limited amount of annotated data to train a detector capable of identifying objects that seamlessly blend into their surroundings. This task is inherently challenging due to", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "CamoTeacher", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the low contrast between the objects and their background. Given a camouflaged object detection dataset $D$ for training, we represent the labeled subset with $M$ labeled samples as $D_{L} = \\{x_{i}^{l},y_{i}\\}_{i = 1}^{M}$ , and the unlabeled subset with $N$ unlabeled samples as $D_{U} = \\{x_{i}^{u}\\}_{i = 1}^{N}$ , where $x_{i}^{l}$ and $x_{i}^{u}$ represent the input image, $y_{i}$ represent the corresponding annotation mask of labeled data. Typically, $D_{L}$ constitutes a small fraction of the entire dataset $D$ , highlighting the semi-supervised learning scenario with $M \\ll N$ . The emphasis on $M \\ll N$ underscores the challenge and opportunity within semi-supervised learning: to enhance detection capabilities by utilizing the untapped potential of the unlabeled data $D_{U}$ , which far exceeds the labeled subset $D_{L}$ .", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Overall Framework", + "text_level": 1, + "bbox": [ + 215, + 315, + 419, + 329 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 4, we adopt Mean Teacher [36] as preliminary to achieve end-to-end semi-supervised camouflaged object detection framework, which contains two COD models with the same structure, i.e., teacher model and student model, parameterized by $\\Theta_t$ and $\\Theta_s$ , respectively. The teacher model generates pseudolabels, which are then used to optimize the student model. The overall loss function $L$ can be defined as:", + "bbox": [ + 212, + 335, + 787, + 426 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {s} + \\lambda_ {u} L _ {u}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 446, + 441, + 785, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $L_{s}$ and $L_{u}$ denote supervised loss and unsupervised loss respectively, $\\lambda_{u}$ is the weight of unsupervised loss to balance the loss terms. Following the classic COD method, we use binary cross-entropy loss $L_{bce}$ for training.", + "bbox": [ + 212, + 460, + 785, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During the training process, we employ a combination of weak augmentation $\\mathcal{A}^w (\\cdot)$ and strong augmentation $\\mathcal{A}^s (\\cdot)$ strategies. Weak augmentation is applied to labeled data to mitigate overfitting, while unlabeled data undergo various data perturbations under strong augmentation to create diverse perspectives of the same image. The supervised loss $L_{s}$ is defined by:", + "bbox": [ + 212, + 507, + 787, + 580 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} L _ {b c e} \\left(F \\left(\\mathcal {A} ^ {w} \\left(x _ {i} ^ {l}\\right); \\Theta_ {s}\\right), y _ {i}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 593, + 785, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $F(\\mathcal{A}(x_i);\\Theta)$ represents the detection results obtained by the model $\\Theta$ for the $i$ -th image with augmentations $\\mathcal{A}(\\cdot)$ . For unlabeled images, we first apply weak augmentation, denoted as $\\mathcal{A}^w (\\cdot)$ , before forwarding them to the teacher model. This initial step is crucial for generating reliable pseudo-labels $\\widehat{y}_i$ under variations that do not significantly alter the core features of the images. These pseudo-labels serve as a form of soft supervision for the student model. Subsequently, the same images are subjected to strong augmentation, $\\mathcal{A}^s (\\cdot)$ , and then passed to the student model. This process introduces a higher level of variability and complexity, simulating more challenging conditions for the student model to adapt to. The student model generates predictions $p_i$ based on these strongly augmented images, leveraging the pseudo-labels $\\widehat{y}_i$ for guidance in learning from the unlabeled data. It can be formulated as:", + "bbox": [ + 212, + 636, + 787, + 816 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {y _ {i}} = F (\\mathcal {A} ^ {w} (x _ {i} ^ {u}); \\Theta_ {t}), p _ {i} = F (\\mathcal {A} ^ {s} (\\mathcal {A} ^ {w} (x _ {i} ^ {u})); \\Theta_ {s}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 825, + 785, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "X.Lai et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/37f00157404283e719b043fde4c197bbba4387395910e6f53f3ac6fc3cc64896.jpg", + "image_caption": [ + "DRCL: Dual-Rotation Consistency Learning PCL: Pixel-wise Consistency Learning ICL: Instance-wise Consistency Learning" + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 630, + 291 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bd3151ee0ffcb62fbe085cefbf294478d42de739c4e0a78207812d206e4d11e2.jpg", + "image_caption": [ + "Fig. 4: The overall pipeline of our CamoTeacher. CamoTeacher consists of a teacher model and a student model. The teacher model generates pseudo-labels to optimize the student model, while the teacher model is updated through EMA [36] from the student. To mitigate pseudo-labels' noise, we introduced Dual-Rotation Consistency Learning (DRCL), which involves Pixel-wise Consistency Learning (PCL) and Instance-wise Consistency Learning (ICL). DRCL weights the loss function based on pixel-wise inconsistency and instance-wise consistency in pseudo-labels across different rotation views." + ], + "image_footnote": [], + "bbox": [ + 638, + 143, + 784, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hence, the unsupervised loss $L_{u}$ can be expressed as:", + "bbox": [ + 215, + 460, + 599, + 474 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {u} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} L _ {b c e} \\left(p _ {i}, \\widehat {y _ {i}}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 487, + 785, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, the student model is intensively trained using the total loss $L$ , which encapsulates both the supervised and unsupervised learning aspects of the semi-supervised framework. This approach ensures that the student model benefits from both labeled and pseudo-labeled data, enhancing its detection capabilities. Concurrently, the teacher model is methodically updated through the Exponential Moving Average (EMA) mechanism [36], which effectively distills students' knowledge while preventing noise, formulated as:", + "bbox": [ + 212, + 539, + 787, + 645 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\Theta_ {t} \\leftarrow \\eta \\Theta_ {t} + (1 - \\eta) \\Theta_ {s}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 657, + 785, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\eta$ is hyper-parameters denotes the keeping rate.", + "bbox": [ + 215, + 685, + 606, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Dual-Rotation Consistency Learning", + "text_level": 1, + "bbox": [ + 215, + 723, + 563, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Due to the camouflaging nature of the objects, pseudo-labels contain a substantial amount of noise, and directly using them to optimize the student model potentially harm the model's performance. To address this problem, one of the most intuitive possible methods is to set a fixed high threshold to filter high-quality pseudo-labels, but this results in a low recall rate and makes it difficult to fully utilize the supervision information of pseudo-labels. To this end, we propose", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "CamoTeacher", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dual-Rotation Consistency Learning (DRCL) to dynamically adjust the weight of the pseudo-labels to reduce the impact of noise.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We randomly apply two separate rotations to the image $x_{i}$ , which has previously been flipped and random resized, resulting in two distinct rotation views $x_{i}^{r_{1}}$ and $x_{i}^{r_{2}}$ .", + "bbox": [ + 212, + 176, + 784, + 220 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nx _ {i} ^ {r _ {1}} = R \\left(\\mathcal {A} ^ {w} \\left(x _ {i}\\right), \\theta_ {1}\\right), x _ {i} ^ {r _ {2}} = R \\left(\\mathcal {A} ^ {w} \\left(x _ {i}\\right), \\theta_ {2}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 222, + 784, + 238 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $x_{i}^{r} = R(x_{i},\\theta)$ represents rotate the input image $x_{i}$ by $\\theta$ degrees. We input the acquired rotation view into the teacher model to obtain the corresponding prediction, i.e. $\\widehat{y}_i^r = F(x_i^r;\\Theta_t)$ . Subsequently, we apply an opposite rotation of $-\\theta$ to return the prediction to its original horizontal orientation, resulting in $\\widehat{y}_i^{h_1}$ and $\\widehat{y}_i^{h_2}$ , facilitating the calculation of prediction inconsistency under different rotation views, respectively,", + "bbox": [ + 214, + 243, + 785, + 335 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {i} ^ {h _ {1}} = R \\left(\\hat {y} _ {i} ^ {r _ {1}}, - \\theta_ {1}\\right), \\hat {y} _ {i} ^ {h _ {2}} = R \\left(\\hat {y} _ {i} ^ {r _ {2}}, - \\theta_ {2}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 345, + 785, + 363 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Note that rotation introduces black boundary areas, which are not involved in the implement of DRCL.", + "bbox": [ + 212, + 372, + 782, + 401 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since the noise levels vary across different regions of pseudo-labels and among different pseudo-labels, we introduce PCL and ICL dynamically adapt the contributions of different pixels within pseudo-labels and across various pseudo-labels.", + "bbox": [ + 212, + 402, + 784, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pixel-wise Consistency Learning We subtract the horizontal predictions $\\widehat{y}_i^{h_1}$ and $\\widehat{y}_i^{h_2}$ at the pixel level to obtain the pixel-wise inconsistency $\\Delta_i$ ,", + "bbox": [ + 214, + 468, + 782, + 502 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta_ {i} = \\left| \\widehat {y} _ {i} ^ {h _ {1}} - \\widehat {y} _ {i} ^ {h _ {2}} \\right|. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 441, + 512, + 784, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The pixel-wise inconsistency $\\Delta_{i}$ across different views reflects the reliability of pseudo-labels. However, in cases where both rotation views' prediction are close to 0.5, $\\Delta_{i}$ cannot effectively distinguish them. These predictions exhibit high uncertainty, meaning they cannot be definitively classified as foreground or background, and are likely to represent noisy labels. Consequently, it becomes necessary to diminish their influence by lowering their weights. Hence, we calculate the average of horizontal predictions $\\widehat{y}_i^h$ ,", + "bbox": [ + 212, + 539, + 784, + 645 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {y} _ {i} ^ {h} = \\operatorname {a v g} \\left(\\widehat {y} _ {i} ^ {h _ {1}}, \\widehat {y} _ {i} ^ {h _ {2}}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 433, + 655, + 784, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $avg(\\cdot, \\cdot)$ represents calculate the average of two inputs in pixel-level, and the use its L2 distance from 0.5 as a component to adjust the weights.", + "bbox": [ + 212, + 681, + 782, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As a result, we derive the pixel-wise consistency weight $\\omega_{i}^{pc}$ based on the pixel-wise inconsistency among different rotation views, as:", + "bbox": [ + 212, + 712, + 782, + 742 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {i} ^ {p c} = \\left(1 - \\Delta_ {i} ^ {\\alpha}\\right) \\left| \\left| \\hat {y} _ {i} ^ {h} - \\mu \\right| \\right| _ {2} ^ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 753, + 784, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\alpha$ is a hyper-parameter and $\\mu = 0.5$ . This dynamic pixel-wise consistency weight $\\omega_{i}^{pc}$ assigns greater weight to regions with consistent predictions across various rotation views, while regions with inconsistent predictions receive a smaller weight.", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "X.Lai et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In conclusion, we formulate the PCL loss $L_{u}^{PC}$ as:", + "bbox": [ + 238, + 145, + 599, + 162 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {u} ^ {P C} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\omega_ {i} ^ {p c} L _ {b c e} (p _ {i}, \\widehat {y} _ {i} ^ {r _ {1}}) \\\\ = - \\frac {1}{N H W} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {H \\times W} \\omega_ {i, j} ^ {p c} \\left[ \\hat {y} _ {i, j} ^ {r _ {1}} \\log p _ {i, j} \\right. \\tag {11} \\\\ + \\left(1 - \\widehat {y} _ {i, j} ^ {r _ {1}}\\right) \\log \\left(1 - p _ {i, j}\\right) ], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 171, + 785, + 277 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "adaptively adjusts the weight of each pixel to ensure comprehensive supervision of student model while avoiding confirmation bias.", + "bbox": [ + 212, + 282, + 782, + 313 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Instance-wise Consistency Learning The degree of camouflage varies between different images, leading to significant variations in the quality of pseudolabels across images. Treating all pseudo-labels equally would be unreasonable. Unfortunately, for unlabeled images, assessing the quality of pseudo-labels is challenging because there are no GT labels available. We have observed a positive correlation between the instance-wise consistency of two rotation views and the quality of pseudo-labels, as quantified by SSIM [39]. Base on this, we introduce ICL to adjust the contributions of pseudo-labels with varying quality. We represent the instance-wise consistency weight $\\omega_{i}^{ic}$ as follows:", + "bbox": [ + 212, + 333, + 784, + 469 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {i} ^ {i c} = \\left(S S I M \\left(\\widehat {y} _ {i} ^ {h _ {1}}, \\widehat {y} _ {i} ^ {h _ {2}}\\right)\\right) ^ {\\beta}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 476, + 784, + 494 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\beta$ is a hyper-parameter used to adjust the distribution relationship between instance-wise consistency and pseudo-label quality.", + "bbox": [ + 212, + 500, + 782, + 530 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We use intersection-over-union (IoU) loss as the instance-wise restriction, therefore, the ICL loss can be expressed as:", + "bbox": [ + 212, + 531, + 782, + 561 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {u} ^ {I C} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\omega_ {i} ^ {i c} L _ {i o u} (p _ {i}, \\widehat {y} _ {i} ^ {r _ {1}}) \\\\ = \\frac {1}{N H W} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {H \\times W} \\omega_ {i} ^ {i c} \\left(1 - \\frac {p _ {i , j} \\widehat {y} _ {i , j} ^ {r _ {1}}}{p _ {i , j} + \\widehat {y} _ {i , j} ^ {r _ {1}} - p _ {i , j} \\widehat {y} _ {i , j} ^ {r _ {1}}}\\right). \\tag {13} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 569, + 784, + 655 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Therefore, the final total $L$ loss comprises three components: supervised loss $L_{s}$ , PCL loss $L_{u}^{LC}$ , and ICL loss $L_{u}^{GC}$ , which can be expressed as:", + "bbox": [ + 214, + 661, + 782, + 691 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {s} + \\lambda_ {u} ^ {p c} L _ {u} ^ {P C} + \\lambda_ {u} ^ {i c} L _ {u} ^ {I C}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 698, + 784, + 718 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda_u^{pc},\\lambda_u^{ic}$ are the hyper-parameters.", + "bbox": [ + 214, + 723, + 506, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 214, + 758, + 366, + 776 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experiment Settings", + "text_level": 1, + "bbox": [ + 214, + 787, + 431, + 803 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dataset. We evaluate our CamoTeacher on four benchmark datasets, CAMO [21], CHAMELEON [33], COD10K [7], and NC4K [25]. In CAMO, there are", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "CamoTeacher", + "bbox": [ + 637, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/6bb7c3b9de840178cdb1f2f38dd56d12dbadee66319e3f3840f18a3f3694bbbc.jpg", + "table_caption": [ + "Table 1: Experimental results on four COD benchmark datasets includes CAMO, CHAMELEON, COD10K, NC4K. $\\dagger$ denotes using the same data augmentation as semi-supervised training. $\\uparrow$ indicates the higher the score the better, and $\\downarrow$ indicates the lower the better. \"n%\" means that \"n%\" of the data was utilized as labeled data, with the remainder as unlabeled data. As the data partitioning is random, we conducted 5 fold experiments and the results are the average of all 5 folds." + ], + "table_footnote": [], + "table_body": "
CHAMELEON (76)
Methods1% (41)5% (202)10% (404)
\\( S_{\\alpha} \\uparrow \\)\\( F_{\\beta}^{w} \\uparrow \\)\\( E_{\\phi}^{m} \\uparrow \\)\\( E_{\\phi}^{x} \\uparrow \\)\\( F_{\\beta}^{m} \\uparrow \\)\\( M \\downarrow \\)\\( S_{\\alpha} \\uparrow \\)\\( F_{\\beta}^{w} \\uparrow \\)\\( E_{\\phi}^{m} \\uparrow \\)\\( E_{\\phi}^{x} \\uparrow \\)\\( F_{\\beta}^{m} \\uparrow \\)\\( M \\downarrow \\)\\( S_{\\alpha} \\uparrow \\)\\( F_{\\beta}^{\\omega } \\uparrow \\)\\( E_{\\phi}^{m} \\uparrow \\)\\( E_{\\phi}^{x} \\uparrow \\)\\( F_{\\beta}^{m} \\uparrow \\)\\( M \\downarrow \\)
Supervised0.6050.2680.5820.7410.4350.2160.6130.2830.6010.7320.4390.1850.7140.4340.7250.8290.5920.110
Supervised†0.6590.3310.6550.7870.4990.1590.7280.4540.7440.8440.6070.106
Mean Teacher0.5370.1990.4180.6360.2290.2040.6110.3090.5240.7450.3530.1370.6790.4500.6500.8120.5120.102
CamoTeacher0.6520.4760.7140.7620.5580.0930.7290.5870.7850.8220.6560.0700.7560.6170.8130.8510.6840.065
", + "bbox": [ + 218, + 241, + 782, + 306 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/2f22377da969e5568c5f8580d78c0124919c1932e4538b2afbbfc3a6c99dfd89.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CAMO (250)
Methods1% (41)5% (202)10% (404)
\\( S_{\\alpha} \\uparrow \\)\\( F_{w}^{w} \\uparrow \\)\\( E_{m}^{m} \\uparrow \\)\\( E_{\\sigma}^{e} \\uparrow \\)\\( F_{g}^{m} \\uparrow \\)\\( M \\downarrow \\)\\( S_{\\alpha} \\uparrow \\)\\( F_{w}^{w} \\uparrow \\)\\( E_{\\sigma}^{m} \\uparrow \\)\\( E_{\\sigma}^{e} \\uparrow \\)\\( F_{g}^{m} \\uparrow \\)\\( M \\downarrow \\)\\( S_{\\alpha} \\uparrow \\)\\( F_{w}^{w} \\uparrow \\)\\( E_{m}^{m} \\uparrow \\)\\( E_{\\sigma}^{e} \\uparrow \\)\\( F_{g}^{m} \\uparrow \\)\\( M \\downarrow \\)
Supervised0.5730.2710.5410.7040.4020.2510.5920.2980.5810.7040.4320.2160.6520.3900.6470.7600.5250.159
Supervised †0.5960.2980.5800.7170.4420.2350.6100.3160.5950.7200.4500.2050.6670.4090.6660.7790.5460.153
Mean Teacher0.5180.2070.3990.6200.2270.2260.5750.2860.4820.7080.3220.1840.6250.3970.5780.7730.4540.150
CamoTeacher0.6210.4560.6690.7360.5450.1360.6690.5230.7110.7750.6010.1220.7010.5600.7420.7950.6350.112
", + "bbox": [ + 218, + 306, + 782, + 371 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/cd7bd3c03c3653fafce76d490ec32e2f8ab815ed542ed933399bc8d51206186c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
COD10K (2026)
Methods1% (41)5% (202)10% (404)
So ↑Fw↑Em↑Eax↑Eax↑Fm↑M↓So ↑Fw↑Em↑Eax↑Eax↑Fm↑M↓So ↑Fw↑Em↑Eax↑Eax↑Fm↑M↓
Supervised0.5660.1890.5760.7230.3560.2090.5960.2200.6270.7470.4030.1640.6750.3360.7260.8100.5240.095
Supervised †0.5820.2060.6090.7430.3860.1940.6210.2460.6590.7740.4440.1470.6800.3380.7260.8120.5280.097
Mean Teacher0.5460.1680.4410.6330.2260.1610.6210.2720.5550.7320.3430.1070.6830.4040.6660.7990.4820.078
CamoTeacher0.6990.5170.7880.7970.5820.0620.7450.5830.8270.8400.6440.0500.7590.5940.8360.8540.6520.049
", + "bbox": [ + 218, + 372, + 782, + 435 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/bd03aa88fc663650e25f795d6a8ddb0a6b683916d3fa83474f598b486f402eba.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
NC4K (4121)
Methods1% (41)5% (202)10% (404)
\\( S_{\\alpha} \\uparrow \\)\\( E_{m}^{w} \\uparrow \\)\\( E_{m}^{m} \\uparrow \\)\\( E_{e}^{s} \\uparrow \\)\\( F_{m}^{m} \\uparrow \\)\\( M \\downarrow \\)\\( S_{\\alpha} \\uparrow \\)\\( F_{m}^{w} \\uparrow \\)\\( E_{m}^{m} \\uparrow \\)\\( E_{e}^{s} \\uparrow \\)\\( F_{m}^{m} \\uparrow \\)\\( M \\downarrow \\)\\( S_{\\alpha} \\uparrow \\)\\( F_{m}^{w} \\uparrow \\)\\( E_{m}^{m} \\uparrow } \\)\\( E_{e}^{s} \\uparrow \\)\\( F_{m}^{m} \\uparrow \\)\\( M \\downarrow \\)
Supervised0.6180.2910.6040.7520.4740.2130.6490.3320.6600.7730.5240.1790.7270.4620.7540.8330.6430.116
Supervised †0.6380.3130.6390.7680.5100.2010.6760.3620.6890.7980.5620.1630.7350.4710.7600.8370.6500.114
Mean Teacher0.5410.2130.4240.6370.2580.1930.6340.3550.5560.7670.4200.1400.7000.4920.6700.8270.5650.109
CamoTeacher0.7180.5990.7790.8140.6750.0900.7770.6770.8340.8590.7390.0710.7910.6870.8420.8680.7460.068
", + "bbox": [ + 218, + 436, + 782, + 500 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "2500 images, consisting of 1250 camouflage images and 1250 non-camouflage images. CHAMELEON contains 76 manually annotated images. COD10K dataset consists of 5066 camouflage images, 3000 background images, and 1934 non-camouflage images. NC4K is another large-scale COD test dataset that comprises 4,121 images. Based on the data partitioning in previous works [7,29,45], we used 3040 images from COD10K and 1000 images from CAMO as the training set for our experiments. The remaining images from both datasets were used as the test set. During the training process, we followed a semi-supervised segmentation [15,47] data partitioning approach. We sampled $1\\%$ , $5\\%$ , $10\\%$ , $20\\%$ , and $30\\%$ of the training set as labeled data, while the remaining portion was used as unlabeled data.", + "bbox": [ + 212, + 534, + 787, + 699 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Metrics. Referring to previous work [16, 17, 29], we used 6 common evaluation indicators in COD to evaluate our CamoTeacher, including S-measure $(S_{\\alpha})$ [4], weight F-measure $(F_{\\beta}^{w})$ [26], mean E-measure $(E_{\\phi}^{m})$ [5], max E-measure $(E_{\\phi}^{x})$ , mean F-measure $(F_{\\beta}^{m})$ [1], and mean absolute error $(M)$ .", + "bbox": [ + 212, + 700, + 787, + 762 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details. The proposed CamoTeacher is implemented with PyTorch. We adopt SINet [7] as the baseline of COD model. We use an SGD optimizer with a momentum of 0.9 and a polynomial learning-rate decay with an initial value of 0.01 to train the student model. The training epoch is set to 40 with 10 epochs for burn-in phase. And the batch size is 20, with a ratio of", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "X.Lai et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e619d9210d4fb34e3b4a2075a534abae5c9b33a2f819337fdaf7a0ea1590c626.jpg", + "table_caption": [ + "Table 2: Results of CamoTeacher on different COD models with $10\\%$ labeled data on COD10K." + ], + "table_footnote": [], + "table_body": "
ModelSettingSα↑Fβw↑Eφm↑Eφx↑Fβm↑M↓
CNN-Based Models
SINet-v2 [6] (TPAMI2022)supervised0.7320.5440.8190.8350.5860.056
CamoTeacher0.7600.6050.8350.8430.6520.051
SegMaR [17] (CVPR2022)supervised0.7550.5830.8370.8510.6230.050
CamoTeacher0.7750.6290.8560.8640.6750.045
Transformer-Based Models
DTINet [24] (ICPR2022)supervised0.7710.6260.8570.8650.6700.043
CamoTeacher0.7940.6610.8730.8820.6950.040
FSPNet [14] (CVPR2023)supervised0.7890.6220.8230.8670.6750.042
CamoTeacher0.8170.6840.8840.9030.7200.037
", + "bbox": [ + 220, + 184, + 782, + 348 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "1:1 between labeled and unlabeled data, i.e. each batch consists of 10 labeled and 10 unlabeled images. During training and inference, each image is resized to $352 \\times 352$ . The teacher model is updated via EMA with a momentum $\\eta$ of 0.996. Weak data augmentation involves Random Flip and Random Scale, while strong data augmentation entails color space transformations including Identity, Autocontrast, Equalize, Gaussian blur, Contrast, Sharpness, Color, Brightness, Hue, Posterize, Solarize, with a random selection of up to 3 from this list.", + "bbox": [ + 212, + 375, + 787, + 481 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Main Results", + "text_level": 1, + "bbox": [ + 215, + 500, + 372, + 513 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison with Baselines. In Tab. 1, we compared CamoTeacher, the supervised baseline [7], and Mean Teacher [36]. Mean Teacher does not show improvements over the supervised in the semi-supervised setting due to treating all pseudo-labels and regions equally, which does not account for pseudo-labels noise problem. For instance, on the NC4K test set, under the $1\\%$ labeled data, the $S_{\\alpha}$ dropped by $9.7\\%$ compared to supervised. Compared to Mean Teacher, CamoTeacher integrated with DRCL effectively mitigates pseudo-label noise issues and efficiently utilizes pseudo-label information. It demonstrates substantial improvements over both supervised and Mean Teacher across various proportions of labeled data, especially in scenarios with extremely limited. For instance, on $1\\%$ NC4K, CamoTeacher exhibited a substantial increase of $28.6\\%$ in the $F_{\\beta}^{w}$ and a $12.3\\%$ enhancement of $S_{\\alpha}$ . Compared with Mean Teacher, the performance gains by CamoTeacher are more significant, which are $38.6\\%$ and $17.7\\%$ , respectively.", + "bbox": [ + 212, + 522, + 785, + 734 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In Fig. 1(b), we visualize the performance changes of supervised, Mean teacher and our CamoTeacher on COD10K under different labeled data ratios of $1\\%$ , $5\\%$ , $10\\%$ , and further $20\\%$ , $30\\%$ , which intuitively reflects the effectiveness of CamoTeacher. It's evident that metrics like $F_{\\beta}^{w}$ , $F_{\\beta}^{m}$ , $E_{\\phi}^{m}$ , and $M$ surpass $100\\%$ of fully supervised performance with just $5\\%$ labeled data. Additionally, with $30\\%$ labeled data, all metrics significantly exceed $100\\%$ of fully supervised performance.", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "CamoTeacher", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/ff68466bdbca52b35cab1521aef42b40ae339ea72f837f84cae70bb16aa5b0c2.jpg", + "table_caption": [ + "Table 3: Result of CamoTeacher on COD10K with additional data. We use $100\\%$ train dataset as labeled data and test dataset as the unlabeled data. \"U\" represents the unsupervised setting, \"W\" denotes the weakly-supervised setting, where weakly-supervision results from scribble annotations. \"L100%\" indicates using $100\\%$ of the training set as labeled data, \"U(CAMO)\" denotes utilizing CAMO as unlabeled data, and likewise for others." + ], + "table_footnote": [], + "table_body": "
ModelSettingSα↑Fwβ↑Emφ↑Exφ↑Fmβ↑M↓
Unsupervised
SelfMask [31] (CVPRW2022)U0.645 0.440 0.687 0.728 0.478 0.125
FOUND [32] (CVPR2023)U0.670 0.482 0.751 0.753 0.520 0.085
UCOS-DA [46] (ICCVW2023)U0.689 0.513 0.740 0.741 0.546 0.086
Weakly-Supervised
SCWSSOD [44] (AAAI2021)W0.710 0.546 0.805 - - 0.055
SCOD [11](AAAI2023)W0.733 0.576 0.832 0.845 0.633 0.049
WS-SAM [10] (NeurIPS2023)W0.803 - 0.878 - - 0.038
Semi-Supervised
CamoTeacher SINetL100%0.771 0.551 0.806 0.868 0.634 0.051
L100%+U(CAMO)0.779 0.554 0.799 0.875 0.632 0.051
L100%+U(NC4K)0.785 0.574 0.812 0.879 0.642 0.049
L100%+U(CAMO,NC4K)0.787 0.576 0.813 0.881 0.644 0.048
", + "bbox": [ + 220, + 239, + 784, + 463 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Application to Existing COD Methods. We applied the CamoTeacher approach to additional COD models, including CNN-based model SINet-v2 [6] and SegMaR [17] and Transformed-based model DTINet [24] and FSPNet [14], and the corresponding results are presented in Tab. 2. In terms of SINet-v2, CamoTeacher demonstrated a $2.8\\%$ improvement in the $S_{\\alpha}$ metric and a $6.1\\%$ improvement in the $F_{\\beta}^{w}$ metric compared to the supervised baseline. Based on FSPNet, CamoTeacher achieved a $2.8\\%$ enhancement in the $S_{\\alpha}$ metric and a $6.2\\%$ improvement in the $F_{\\beta}^{w}$ metric without utilizing additional discriminative annotation. These outcomes provide robust evidence of the effectiveness of CamoTeacher.", + "bbox": [ + 212, + 491, + 787, + 642 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Addition unlabeled data. In Tab. 3, based on the base model SINet, we train CamoTeacher with $100\\%$ labeled data and check whether additional unlabeled data can further improve the performance. Compared with unsupervised and weakly-supervised methods, it shows the superiority of our method. We use the testset CAMO and NC4K as unlabeled data, and we find that adding more additional data CamoTeacher can make SINet continued to gain, such as $S_{\\alpha}$ metric increased by $1.6\\%$ and $F_{\\beta}^{w}$ metric increased by $2.5\\%$ . We achieved significantly better performance than unsupervised methods [31,32,46] using existing labeled data. Compared with the weakly-supervision methods [10,11,44], we do not need additional annotation for the new unlabeled data, and to a certain extent, as the more unlabeled data, the performance will get better and better. And at the same time, replacing the base model with a stronger model will get better results.", + "bbox": [ + 212, + 643, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "X.Lai et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/f16dfdf701535b66d259cd63b66493e4640a806c5bfd88f504e85919a69802cd.jpg", + "table_caption": [ + "Table 4: The ablation study of Pixel-wise Table 5: The ablation study of various Consistency Learning (PCL) and Instance- weight approach of Pixel-wise Consistency wise Consistency Learning (ICL) with $5\\%$ Learning (PCL) with $5\\%$ labeled data on labeled data on COD10K. COD10K. In this experiment $\\alpha$ is set to 1/4." + ], + "table_footnote": [], + "table_body": "
PCLICL\\( {S}_{\\alpha } \\uparrow \\)\\( {F}_{\\beta }^{w} \\uparrow \\)\\( {E}_{\\phi }^{m} \\uparrow \\)\\( {E}_{\\phi }^{x} \\uparrow \\)\\( {F}_{\\beta }^{m} \\uparrow \\)\\( M \\downarrow \\)
\\( L_{bce} \\)\\( {\\omega }^{pc} \\)\\( L_{iou} \\)\\( {\\omega }^{tc} \\)
✓0.6820.3700.6360.7870.4610.091
✓0.7340.5720.8170.8210.6260.057
✓✓0.7240.5010.7380.8320.5740.058
✓✓0.7420.5850.8280.8340.6390.052
✓✓✓0.7450.5860.8300.8400.6430.051
✓✓✓✓0.7460.5860.8300.8430.6450.050
", + "bbox": [ + 233, + 213, + 483, + 287 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/7ed5a63caa2c204233eefac17a40e347299c68af8c15d891aa40be673bfbdf2c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
weight approach\\( {S}_{\\alpha } \\uparrow {F}_{\\beta }^{w} \\uparrow {E}_{\\phi }^{m} \\uparrow {E}_{\\phi }^{x} \\uparrow {F}_{\\beta }^{m} \\uparrow M \\downarrow \\)
\\( {\\omega }^{pc} = {\\widehat{y}}^{r} \\)0.6980.3930.6840.8040.5000.110
\\( {\\omega }^{pc} = {\\begin{Vmatrix}{\\widehat{y}}^{r} - \\mu \\end{Vmatrix}}_{2}^{2} \\)0.7170.4850.7340.8220.5590.061
\\( {\\omega }^{pc} = 1 - {\\Delta }^{\\alpha } \\)0.7190.4810.7200.8270.5530.062
\\( {\\omega }^{pc} = \\left( {1 - {\\Delta }^{\\alpha }}\\right) {\\widehat{y}}^{r} \\)0.6950.4050.6960.7950.5060.114
\\( {\\omega }^{pc} = \\left( {1 - {\\Delta }^{\\alpha }}\\right) {\\begin{Vmatrix}{\\widehat{y}}^{r} - \\mu \\end{Vmatrix}}_{2}^{2}\\left\\lbrack \\begin{array}{llllllll} 0 & {724} & {0.501} & {0.738} & {0.832} & {0.574} & {0.058} \\end{array}\\right\\rbrack \\)
", + "bbox": [ + 504, + 227, + 779, + 286 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/f20d28e82e3a14ada8b6a2eb6ba06eea93cc88436f82927f61e9826fabf03617.jpg", + "table_caption": [ + "Table 6: The ablation study of hyper- Table 7: The ablation study of hyperparameters $\\alpha$ in Pixel-wise Consistency parameters $\\beta$ in Instance-wise Consistency Learning (PCL) with $5\\%$ labeled data on Learning (ICL) with $5\\%$ labeled data on COD10K. In this experiment $\\beta$ is set to 4, COD10K. In this experiment $\\alpha$ is set to $\\lambda_u^{PC}$ is set to 8, and $\\lambda_u^{IC}$ is set to 0.3. $1/4, \\lambda_u^{PC}$ is set to 8, and $\\lambda_u^{IC}$ is set to 0.3." + ], + "table_footnote": [], + "table_body": "
αSα↑Fwβ↑Eφm↑Eφx↑Fmβ↑M↓
20.7370.5650.8130.8360.6290.052
10.7390.5660.8150.8320.6270.056
1/20.7450.5750.8200.8390.6370.051
1/40.7470.5790.8250.8420.6380.051
1/80.7450.5780.8250.8380.6360.052
", + "bbox": [ + 225, + 388, + 496, + 473 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/ab6fa66203865e46ece1bc287bdf636138dcd5c61d51213b5d487bba22110ef8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
βSα↑Fβw↑Eφm↑Eφx↑Fβm↑M↓
10.7430.5760.8240.8390.6360.051
20.7420.5730.8230.8370.6320.052
40.7470.5790.8250.8420.6380.051
50.7440.5780.8240.8410.6390.051
60.7450.5750.8220.8390.6350.052
", + "bbox": [ + 511, + 388, + 767, + 473 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 508, + 388, + 523 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation on PCL and ICL. We conducted ablation experiments on different combinations of weight and loss functions for PCL and ICL. From Tab. 4, increasing the weights $\\omega^{pc}$ and $\\omega^{ic}$ improves the model's performance. Specifically, with $\\omega^{pc}$ results in a higher $F_{\\beta}^{w}$ score, increasing from 0.370 to 0.501. Similarly, with $\\omega^{ic}$ improved the $F_{\\beta}^{w}$ metric from 0.572 to 0.585. Combining these two weighted loss functions led to a significant performance improvement, with the $F_{\\beta}^{w}$ score increasing from 0.370 to 0.586 (an increase of $21.6\\%$ ). This demonstrates the effectiveness of PCL and ICL in enhancing the model's performance.", + "bbox": [ + 212, + 546, + 787, + 667 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation of weight approach in PCL. The choice of method for obtaining $\\omega^{pc}$ significantly affects the performance of the model. Therefore, we conducted several experiments to select the most effective method for obtaining $\\omega^{pc}$ . From Tab. 5, it can be observed that the best performance is achieved when $\\omega^{pc}$ is obtained using Eq. (10).", + "bbox": [ + 212, + 671, + 787, + 747 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation of hyper-parameters in PCL and ICL. We perform ablation experiments on the hyper-parameters associated with the PCL and ICL modules. As shown in Tab. 6, the optimal results are achieved when setting the hyper-parameter $\\alpha$ , which corresponds to the coefficient $\\omega^{pc}$ , to $1/4$ . Similarly, according to Tab. 7, the best performance is observed when the hyper-parameter $\\beta$ , related to the coefficient $\\omega^{ic}$ , is set to 4.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "CamoTeacher", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3642b2a7ae66b712dc7766f375dc5b354c32de919ada0deaf38816c562fb24bc.jpg", + "image_caption": [ + "Fig. 5: Visualization of predictions in different proportions of labeled data." + ], + "image_footnote": [], + "bbox": [ + 276, + 146, + 723, + 378 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.4 Qualitative Analysis", + "text_level": 1, + "bbox": [ + 215, + 438, + 429, + 453 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Result visualization. We conducte visualizations of CamoTeacher's results under varying proportions of labeled data. The model is trained with $1\\%$ , $5\\%$ , $10\\%$ , $20\\%$ and $30\\%$ labeled data, and we compare its predicted masks with those produced by SINet under fully supervised conditions (as shown in Fig. 5). Notably, even with only $10\\%$ labeled data, CamoTeacher's predicted masks exhibit a remarkable similarity to the fully supervised SINet predictions. In fact, in certain detailed predictions, CamoTeacher even surpasses the performance of fully supervised SINet. These findings highlight the impressive capability of CamoTeacher in leveraging limited labeled data for effective camouflaged object detection.", + "bbox": [ + 212, + 464, + 787, + 616 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 640, + 369, + 656 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we present the first end-to-end semi-supervised camouflaged object detection model called CamoTeacher. To address the issue of substantial noise within pseudo-labels in semi-supervised COD, including both local noise and global noise, we've introduced a novel approach called Dual-Rotation Consistency Learning (DRCL), including Pixel-wise Consistency Learning (PCL) and Instance-wise Consistency Learning (ICL). DRCL assists the model in alleviating noise issues, effectively leveraging pseudo-labels information, so that the model can receive adequate supervision while avoiding confirmation bias. Extensive experiments validate CamoTeacher's superior performance while significantly reducing annotation expenses. We believe that CamoTeacher can serve as a strong baseline for future research on semi-supervised COD.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "X.Lai et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by National Science and Technology Major Project (No. 2022ZD0118201), the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China a (No. U21B2037, No. U22B2051, No. U23A20383, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No.2022J06001).", + "bbox": [ + 212, + 176, + 787, + 282 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 306, + 321, + 321 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Achanta, R., Hemami, S., Estrada, F., Susstrunk, S.: Frequency-tuned salient region detection. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 1597-1604. IEEE (2009)", + "2. Chen, H., Tao, R., Fan, Y., Wang, Y., Wang, J., Schiele, B., Xie, X., Raj, B., Savvides, M.: Softmatch: Addressing the quantity-quality trade-off in semi-supervised learning. arXiv preprint arXiv:2301.10921 (2023)", + "3. Chen, X., Yuan, Y., Zeng, G., Wang, J.: Semi-supervised semantic segmentation with cross pseudo supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2613-2622 (2021)", + "4. Fan, D.P., Cheng, M.M., Liu, Y., Li, T., Borji, A.: Structure-measure: A new way to evaluate foreground maps. In: Proceedings of the IEEE international conference on computer vision. pp. 4548-4557 (2017)", + "5. Fan, D.P., Gong, C., Cao, Y., Ren, B., Cheng, M.M., Borji, A.: Enhanced-alignment measure for binary foreground map evaluation. arXiv preprint arXiv:1805.10421 (2018)", + "6. Fan, D.P., Ji, G.P., Cheng, M.M., Shao, L.: Concealed object detection. IEEE transactions on pattern analysis and machine intelligence 44(10), 6024-6042 (2021)", + "7. Fan, D.P., Ji, G.P., Sun, G., Cheng, M.M., Shen, J., Shao, L.: Camouflaged object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2777-2787 (2020)", + "8. Grandvalet, Y., Bengio, Y.: Semi-supervised learning by entropy minimization. Advances in neural information processing systems 17 (2004)", + "9. He, C., Li, K., Zhang, Y., Tang, L., Zhang, Y., Guo, Z., Li, X.: Camouflaged object detection with feature decomposition and edge reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22046-22055 (2023)", + "10. He, C., Li, K., Zhang, Y., Xu, G., Tang, L., Zhang, Y., Guo, Z., Li, X.: Weakly-supervised concealed object segmentation with sam-based pseudo labeling and multi-scale feature grouping. Advances in Neural Information Processing Systems 36 (2024)", + "11. He, R., Dong, Q., Lin, J., Lau, R.W.: Weakly-supervised camouflaged object detection with scribble annotations. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 781-789 (2023)", + "12. Hu, J., Lin, J., Cai, W., Gong, S.: Relax image-specific prompt requirement in sam: A single generic prompt for segmenting camouflaged objects. arXiv preprint arXiv:2312.07374 (2023)" + ], + "bbox": [ + 218, + 339, + 784, + 838 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "CamoTeacher", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Hu, J., Chen, C., Cao, L., Zhang, S., Shu, A., Jiang, G., Ji, R.: Pseudo-label alignment for semi-supervised instance segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16337-16347 (2023)", + "14. Huang, Z., Dai, H., Xiang, T.Z., Wang, S., Chen, H.X., Qin, J., Xiong, H.: Feature shrinkage pyramid for camouflaged object detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5557-5566 (2023)", + "15. Hung, W.C., Tsai, Y.H., Liou, Y.T., Lin, Y.Y., Yang, M.H.: Adversarial learning for semi-supervised semantic segmentation. arXiv preprint arXiv:1802.07934 (2018)", + "16. Ji, G.P., Fan, D.P., Chou, Y.C., Dai, D., Liniger, A., Van Gool, L.: Deep gradient learning for efficient camouflaged object detection. Machine Intelligence Research 20(1), 92-108 (2023)", + "17. Jia, Q., Yao, S., Liu, Y., Fan, X., Liu, R., Luo, Z.: Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4713-4722 (2022)", + "18. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)", + "19. Lai, X., Tian, Z., Jiang, L., Liu, S., Zhao, H., Wang, L., Jia, J.: Semi-supervised semantic segmentation with directional context-aware consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1205-1214 (2021)", + "20. Laine, S., Aila, T.: Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242 (2016)", + "21. Le, T.N., Nguyen, T.V., Nie, Z., Tran, M.T., Sugimoto, A.: Anabranch network for camouflaged object segmentation. Computer vision and image understanding 184, 45-56 (2019)", + "22. Lee, D.H., et al.: Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In: Workshop on challenges in representation learning, ICML. vol. 3, p. 896. Atlanta (2013)", + "23. Liu, Y., Tian, Y., Chen, Y., Liu, F., Belagiannis, V., Carneiro, G.: Perturbed and strict mean teachers for semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4258-4267 (2022)", + "24. Liu, Z., Zhang, Z., Tan, Y., Wu, W.: Boosting camouflaged object detection with dual-task interactive transformer. In: 2022 26th International Conference on Pattern Recognition (ICPR). pp. 140-146. IEEE (2022)", + "25. Lv, Y., Zhang, J., Dai, Y., Li, A., Liu, B., Barnes, N., Fan, D.P.: Simultaneously localize, segment and rank the camouflaged objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11591-11601 (2021)", + "26. Margolin, R., Zelnik-Manor, L., Tal, A.: How to evaluate foreground maps? In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 248-255 (2014)", + "27. Mi, P., Lin, J., Zhou, Y., Shen, Y., Luo, G., Sun, X., Cao, L., Fu, R., Xu, Q., Ji, R.: Active teacher for semi-supervised object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14482-14491 (2022)", + "28. Oliver, A., Odena, A., Raffel, C.A., Cubuk, E.D., Goodfellow, I.: Realistic evaluation of deep semi-supervised learning algorithms. Advances in neural information processing systems 31 (2018)" + ], + "bbox": [ + 217, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "X.Lai et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Pang, Y., Zhao, X., Xiang, T.Z., Zhang, L., Lu, H.: Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 2160-2170 (2022)", + "30. Roelofs, B., Berthelot, D., Sohn, K., Carlini, N., Kurakin, A.: Adamatch: a unified approach to semi-supervised learning and domain adaptation. In: Proc. of the Int. Conf. on Learning Representations (ICLR) (2022)", + "31. Shin, G., Albanie, S., Xie, W.: Unsupervised salient object detection with spectral cluster voting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3971-3980 (2022)", + "32. Simeoni, O., Sekkat, C., Puy, G., Vobecký, A., Zablocki, É., Pérez, P.: Unsupervised object localization: Observing the background to discover objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3176-3186 (2023)", + "33. Skurowski, P., Abdulameer, H., Błaszczyk, J., Depta, T., Kornacki, A., Koziel, P.: Animal camouflage analysis: Chameleon database. Unpublished manuscript 2(6), 7 (2018)", + "34. Sohn, K., Berthelot, D., Carlini, N., Zhang, Z., Zhang, H., Raffel, C.A., Cubuk, E.D., Kurakin, A., Li, C.L.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems 33, 596-608 (2020)", + "35. Sohn, K., Zhang, Z., Li, C.L., Zhang, H., Lee, C.Y., Pfister, T.: A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757 (2020)", + "36. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems 30 (2017)", + "37. Wang, Y., Wang, H., Shen, Y., Fei, J., Li, W., Jin, G., Wu, L., Zhao, R., Le, X.: Semi-supervised semantic segmentation using unreliable pseudo-labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4248-4257 (2022)", + "38. Wang, Z., Li, Y., Guo, Y., Fang, L., Wang, S.: Data-uncertainty guided multi-phase learning for semi-supervised object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4568-4577 (2021)", + "39. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)", + "40. Xu, M., Zhang, Z., Hu, H., Wang, J., Wang, L., Wei, F., Bai, X., Liu, Z.: End-to-end semi-supervised object detection with soft teacher. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3060-3069 (2021)", + "41. Yang, L., Qi, L., Feng, L., Zhang, W., Shi, Y.: Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7236-7246 (2023)", + "42. Yang, L., Zhuo, W., Qi, L., Shi, Y., Gao, Y.: St++: Make self-training work better for semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4268-4277 (2022)", + "43. Yin, B., Zhang, X., Hou, Q., Sun, B.Y., Fan, D.P., Van Gool, L.: Camoformer: Masked separable attention for camouflaged object detection. arXiv preprint arXiv:2212.06570 (2022)", + "44. Yu, S., Zhang, B., Xiao, J., Lim, E.G.: Structure-consistent weakly supervised salient object detection with local saliency coherence. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 3234-3242 (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "CamoTeacher", + "bbox": [ + 638, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "45. Zhang, M., Xu, S., Piao, Y., Shi, D., Lin, S., Lu, H.: Preynet: Preying on camouflaged objects. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 5323-5332 (2022)", + "46. Zhang, Y., Wu, C.: Unsupervised camouflaged object segmentation as domain adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4334-4344 (2023)", + "47. Zhao, Z., Yang, L., Long, S., Pi, J., Zhou, L., Wang, J.: Augmentation matters: A simple-yet-effective approach to semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11350-11359 (2023)", + "48. Zhao, Z., Zhou, L., Wang, L., Shi, Y., Gao, Y.: Lassl: Label-guided self-training for semi-supervised learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 9208-9216 (2022)", + "49. Zhou, Z.H., Zhou, Z.H.: Semi-supervised learning. Machine Learning pp. 315-341 (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 354 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "X.Lai et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/CamoTeacher_ Dual-Rotation Consistency Learning for Semi-Supervised Camouflaged Object Detection/e7b0e9c5-9645-43e1-9e04-1cc9804f36e6_model.json b/2024/CamoTeacher_ Dual-Rotation Consistency Learning for Semi-Supervised Camouflaged Object Detection/e7b0e9c5-9645-43e1-9e04-1cc9804f36e6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2d4200da668196d510dae889c4aade472c8e7658 --- /dev/null +++ b/2024/CamoTeacher_ Dual-Rotation Consistency Learning for Semi-Supervised Camouflaged Object Detection/e7b0e9c5-9645-43e1-9e04-1cc9804f36e6_model.json @@ -0,0 +1,2656 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.257, + 0.141, + 0.75, + 0.21 + ], + "angle": 0, + "content": "CamoTeacher: Dual-Rotation Consistency Learning for Semi-Supervised Camouflaged Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.234, + 0.761, + 0.268 + ], + "angle": 0, + "content": "Xunfa Lai\\(^{1}\\), Zhiyu Yang\\(^{1}\\), Jie Hu\\(^{2}\\), Shengchuan Zhang\\(^{1*}\\), Liujuan Cao\\(^{1}\\), Guannan Jiang\\(^{2}\\), Zhiyu Wang\\(^{2}\\), Songan Zhang\\(^{3}\\), and Rongrong Ji\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.277, + 0.763, + 0.306 + ], + "angle": 0, + "content": "1 Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University" + }, + { + "type": "text", + "bbox": [ + 0.332, + 0.306, + 0.672, + 0.319 + ], + "angle": 0, + "content": "\\(^{2}\\) Contemporary Amperex Technology Co. Limited" + }, + { + "type": "text", + "bbox": [ + 0.393, + 0.32, + 0.612, + 0.334 + ], + "angle": 0, + "content": "3 Shanghai Jiao Tong University" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.366, + 0.744, + 0.617 + ], + "angle": 0, + "content": "Abstract. Existing camouflaged object detection (COD) methods depend heavily on large-scale pixel-level annotations. However, acquiring such annotations is laborious due to the inherent camouflage characteristics of the objects. Semi-supervised learning offers a promising solution to this challenge. Yet, its application in COD is hindered by significant pseudo-label noise, both pixel-level and instance-level. We introduce CamoTeacher, a novel semi-supervised COD framework, utilizing Dual-Rotation Consistency Learning (DRCL) to effectively address these noise issues. Specifically, DRCL minimizes pseudo-label noise by leveraging rotation views' consistency in pixel-level and instance-level. First, it employs Pixel-wise Consistency Learning (PCL) to deal with pixel-level noise by reweighting the different parts within the pseudo-label. Second, Instance-wise Consistency Learning (ICL) is used to adjust weights for pseudo-labels, which handles instance-level noise. Extensive experiments on four COD benchmark datasets demonstrate that the proposed CamoTeacher not only achieves state-of-the-art compared with semi-supervised learning methods, but also rivals established fully-supervised learning methods. Our code will be available soon." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.629, + 0.731, + 0.645 + ], + "angle": 0, + "content": "Keywords: Camouflaged object detection \\(\\cdot\\) Semi-supervised learning" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.667, + 0.377, + 0.683 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.788, + 0.819 + ], + "angle": 0, + "content": "Camouflaged object detection (COD) aims to identify objects that blend seamlessly into their environments including animals or artificial entities that possess protective coloring and have the ability to blend into their surroundings, a task complicated by low contrast, similar textures, and blurred boundaries [6,7,9,14, 17,24,43]. Unlike general object detection, COD is challenged by these factors, making detection significantly more difficult. Existing methods in COD heavily depend on extensive pixel-level annotated datasets, the creation of which incurs substantial human effort and cost, thus limiting COD's advancement." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.388, + 0.84 + ], + "angle": 0, + "content": "* Corresponding authors" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.127 + ], + "angle": 0, + "content": "X.Lai et al." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.145, + 0.756, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.307, + 0.788, + 0.364 + ], + "angle": 0, + "content": "Fig. 1: (a) Illustration of pixel-level and instance-level noise. Pixel-level noise refers to varying degrees of noise in different parts within an instance. Instance-level noise pertains to varying degrees of noise among different instances. (b) Visualization of performance trends on COD10K [7] under different percentages of labeled data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.789, + 0.717 + ], + "angle": 0, + "content": "To mitigate that, semi-supervised learning [13, 27, 36, 47, 49] emerges as a promising approach by leveraging both labeled and unlabeled data, yet its application in COD is not straightforward due to complex backgrounds and subtle object boundaries. The efficacy of semi-supervised learning in COD is significantly compromised by the presence of substantial noise in pseudo-labels. Our investigation into pseudo-label noise reveals two primary types: pixel-level noise, indicating variation within a single pseudo-label, and instance-level noise, showing variation across different pseudo-labels. This distinction is critical, as it informs our approach to refining pseudo-label quality for improved model training. (1) Pixel-level noise is characterized by inconsistent labeling within various parts of a pseudo-label. As demonstrated in Fig. 1(a), in the first row, the tail of the gecko is visually more camouflaged than its head. The pseudo-label generated by the SINet [7] is less accurate in its tail region (highlighted by the red box). This observation underscores the inappropriateness of uniformly treating all parts within a pseudo-label. (2) Instance-level noise refers to the variability in noise levels across distinct pseudo-labels. As shown in Fig. 1(a), the pseudo-labels of the third row are less accurate compared to that in the second row because the camouflaged objects in the third row is more difficult to detect. These discrepancies indicate that each pseudo-label contributes variably to model training, emphasizing the need for a nuanced approach to integrating pseudo-label information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To tackle the challenge of evaluating pseudo-label noise without ground truth (GT) for unlabeled data, we introduce two novel strategies based on pixel-wise inconsistency and instance-wise consistency across two rotation views. Specifically, for pixel-level noise, we observe that pixel-wise inconsistency—calculated by comparing pseudo-labels from two rotated views—mirrors the actual error relative to GT, as shown in Fig. 2(a). This relationship, illustrated by the polyline in Fig. 2(b), shows a positive correlation between mean pixel-wise inconsistency and the mean absolute error (MAE) across different parts. Consequently, areas" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.639, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CamoTeacher" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.148, + 0.323, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.196, + 0.322, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.27, + 0.243, + 0.299, + 0.253 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.149, + 0.399, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.196, + 0.399, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.243, + 0.371, + 0.252 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.149, + 0.476, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.196, + 0.476, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.243, + 0.477, + 0.252 + ], + "angle": 0, + "content": "Real absolute error" + }, + { + "type": "image", + "bbox": [ + 0.478, + 0.149, + 0.553, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.478, + 0.196, + 0.553, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.485, + 0.243, + 0.543, + 0.252 + ], + "angle": 0, + "content": "Pixel-wise" + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.153, + 0.757, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.255, + 0.665, + 0.268 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.256, + 0.41, + 0.267 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.279, + 0.788, + 0.39 + ], + "angle": 0, + "content": "Fig. 2: (a) The real absolute error is obtained by subtracting the pseudo-label from the ground truth, reflecting the actual noisy level. On the other hand, pixel-wise inconsistency is derived from the subtraction of pseudo-labels from two rotated views. Visually, pixel-wise inconsistency and real absolute error appear very similar. (b) We define the area within 20 pixels from the edge as the boundary according to GT. We calculate mean pixel-wise inconsistency (MPI) and mean absolute error (MAE) in background, foreground, and boundary. The pseudo-labels are obtained by our proposed model on the test dataset COD10K [7] at a \\(10\\%\\) semi-supervised setting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.788, + 0.547 + ], + "angle": 0, + "content": "with higher pixel-wise inconsistency are more prone to inaccuracies, suggesting a need for their de-emphasis during training. For instance-level noise, we determine that pseudo-labels with greater similarity across rotation views exhibit lower noise levels, as indicated in Fig. 3(a). A positive correlation between instance-wise consistency and the SSIM [39] computed by pseudo-labels and GT further supports this observation, depicted in Fig. 3(b). Thus, pseudo-labels demonstrating higher instance-wise consistency are likely of superior quality and should be prioritized in the learning process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.792, + 0.701 + ], + "angle": 0, + "content": "With the insights gained from these observations, we propose a semi-supervised camouflaged object detection framework called CamoTeacher, which incorporates a novel method known as Dual-Rotation Consistency Learning (DRCL). Specifically, DRCL operationalizes its strategy through two core components: Pixel-wise Consistency Learning (PCL) and Instance-wise Consistency Learning (ICL). PCL innovatively assigns variable weights to different parts within a pseudo-label, considering the pixel-wise inconsistency across diverse rotation view. Simultaneously, ICL adjusts the importance of individual pseudo-labels based on their instance-wise consistency, enabling a nuanced, noise-aware training process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We adopt SINet [7] as the base model to implement CamoTeacher and apply it to more classic COD models, i.e. CNN-based model SINet-v2 [6] and SegMaR [17] and Transformed-based model DTINet [24] and FSPNet [14]. Extensive experiments on four COD benchmark datasets, i.e. CAMO [21], CHAMELEON [33], COD10K [7], and NC4K [25], demonstrate that the proposed CamoTeacher not only achieves state-of-the-art compared with semi-supervised learning methods, but also rivals established fully-supervised learning methods. Specifically, as shown on Fig. 1(b), with only \\(20\\%\\) labeled data, it nearly reached the performance of fully supervised model on COD10K." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "X.Lai et al." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.144, + 0.323, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.191, + 0.322, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.27, + 0.239, + 0.3, + 0.249 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.145, + 0.399, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.191, + 0.398, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.239, + 0.369, + 0.248 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.145, + 0.476, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.191, + 0.474, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.239, + 0.467, + 0.25 + ], + "angle": 0, + "content": "Pseudo-label" + }, + { + "type": "image", + "bbox": [ + 0.478, + 0.145, + 0.554, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.478, + 0.191, + 0.554, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.487, + 0.239, + 0.542, + 0.25 + ], + "angle": 0, + "content": "Pseudo-label" + }, + { + "type": "image", + "bbox": [ + 0.566, + 0.146, + 0.757, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.253, + 0.673, + 0.265 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.393, + 0.253, + 0.411, + 0.265 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.276, + 0.788, + 0.346 + ], + "angle": 0, + "content": "Fig. 3: (a) The more similar the pseudo-labels are between two rotated views, the lower the noise level and the higher the quality of the pseudo-labels. For example, in the first row, the pseudo-labels are more similar, indicating a lower noise level compared to the pseudo-labels in the second row. (b) The positive correlation between instance-wise consistency and the SSIM [39] of pseudo-labels and GT." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.377, + 0.592, + 0.391 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.404, + 0.785, + 0.448 + ], + "angle": 0, + "content": "- We introduce the first end-to-end semi-supervised camouflaged object detection framework CamoTeacher, offering a simple yet effective baseline for future research in semi-supervised camouflaged object detection." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.45, + 0.785, + 0.538 + ], + "angle": 0, + "content": "- To address the problem of large amounts of noise in pseudo-labels in semi-supervised camouflaged object detection, we proposed the Dual-Rotation Consistency Learning (DRCL), which includes Pixel-wise Consistency Learning (PCL) and Instance-wise Consistency Learning (ICL), allows for the adaptive adjustment of the contribution of pseudo-labels with varying quality, enabling the effective utilization of pseudo-label information." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.541, + 0.806, + 0.57 + ], + "angle": 0, + "content": "- We conducted extensive experiments on COD benchmark datasets, and achieved significant improvements compared to the fully supervised setting." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.404, + 0.806, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.597, + 0.391, + 0.613 + ], + "angle": 0, + "content": "2 Related works" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.631, + 0.517, + 0.646 + ], + "angle": 0, + "content": "2.1 Camouflaged Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Camouflaged Object Detection(COD) is different from traditional visual detection tasks as objects and backgrounds in traditional detection tasks have significant visual distinctions, whereas camouflaged objects are typically concealed within their surrounding environments [6, 7, 14, 17, 24, 29, 43]. In recent years, many models have been proposed on this task, which can be divided into CNN-based and Transformed-based. i) CNN-based models: SINet [7] is a simple yet effective baseline that views camouflage detection as searching and identification process. SINet-v2 [6] significantly improves performance by introducing a neighbor connection decoder and cascaded group-reversal attention design. SegMaR [17] firstly performs initial detection of camouflaged objects and then progressively enlarges the detected regions to achieve precise detection. FEDER [9] achieves impressive results by conducting feature decomposition and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.639, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CamoTeacher" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.266 + ], + "angle": 0, + "content": "edge reconstruction. ii) Transformed-based models: DTINet [24] proposes a dual-task interactive transformer to detect accurate position and detailed boundary. CamoFormer [43] presents masked separable attention to separate foreground and background regions. FSPNet [14] designs a feature shrinkage decoder hierarchically decode locality-enhanced neighboring transformer features. These outstanding works have achieved excellent results in COD, but they rely on large-scale pixel-level annotations, which are time-consuming and labor-intensive to obtain." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.785, + 0.449 + ], + "angle": 0, + "content": "In response to the reliance of full-supervised COD on expensive pixel-level annotations, some unsupervised/weakly-supervised approaches have emerged [10-12,46]. UCOS-DA [46] formulates the unsupervised camouflaged object segmentation as a source-free unsupervised domain adaptation task and designs a foreground-background-contrastive self-adversarial pipeline. SCOD [11] proposes the first weakly-supervised COD method with relabeled scribbles. WS-SAM [10] uses the provided sparse annotations as prompts for SAM [18] to generate segmentation masks to train the model. These methods reduce annotation costs but often exhibit lower performance compared to fully-supervised methods. Is it possible to leverage existing annotated data without additional labeling while ensuring accuracy? To address this issue, we propose a new benchmark that combines semi-supervised learning with camouflaged object detection." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.467, + 0.479, + 0.483 + ], + "angle": 0, + "content": "2.2 Semi-Supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.785, + 0.652 + ], + "angle": 0, + "content": "In traditional supervised learning, training a high-performing model requires a substantial amount of labeled data. However, obtaining labeled data in practical applications is often costly and time-consuming. Consequently, the effective utilization of unlabeled data to improve model performance and generalization has become a critical concern. Semi-supervised learning addresses this challenge by leveraging unlabeled data to enhance the model's learning process [2,8,22,27,28,30,34,36,48], thereby improving its performance and generalization capability. Pseudo-labeling [3,35,37,38,40,42] and consistency regularization [13,19,20,23,41] are popular techniques in the field of semi-supervised learning. These methods have shown great efficacy when combined, producing impressive results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.653, + 0.785, + 0.729 + ], + "angle": 0, + "content": "To address the issue of heavy reliance on labeled data in camouflage object detection, we propose the first end-to-end semi-supervised camouflage object detection method. By integrating semi-supervised learning, our approach aims to enhance the performance of camouflage object detection models while significantly reducing the need for annotated data." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.748, + 0.381, + 0.765 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.775, + 0.407, + 0.789 + ], + "angle": 0, + "content": "3.1 Task Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Semi-supervised camouflaged object detection aims to leverage a limited amount of annotated data to train a detector capable of identifying objects that seamlessly blend into their surroundings. This task is inherently challenging due to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "X.Lai et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "the low contrast between the objects and their background. Given a camouflaged object detection dataset \\( D \\) for training, we represent the labeled subset with \\( M \\) labeled samples as \\( D_{L} = \\{x_{i}^{l},y_{i}\\}_{i = 1}^{M} \\), and the unlabeled subset with \\( N \\) unlabeled samples as \\( D_{U} = \\{x_{i}^{u}\\}_{i = 1}^{N} \\), where \\( x_{i}^{l} \\) and \\( x_{i}^{u} \\) represent the input image, \\( y_{i} \\) represent the corresponding annotation mask of labeled data. Typically, \\( D_{L} \\) constitutes a small fraction of the entire dataset \\( D \\), highlighting the semi-supervised learning scenario with \\( M \\ll N \\). The emphasis on \\( M \\ll N \\) underscores the challenge and opportunity within semi-supervised learning: to enhance detection capabilities by utilizing the untapped potential of the unlabeled data \\( D_{U} \\), which far exceeds the labeled subset \\( D_{L} \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.316, + 0.421, + 0.33 + ], + "angle": 0, + "content": "3.2 Overall Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.336, + 0.788, + 0.427 + ], + "angle": 0, + "content": "As shown in Fig. 4, we adopt Mean Teacher [36] as preliminary to achieve end-to-end semi-supervised camouflaged object detection framework, which contains two COD models with the same structure, i.e., teacher model and student model, parameterized by \\(\\Theta_t\\) and \\(\\Theta_s\\), respectively. The teacher model generates pseudolabels, which are then used to optimize the student model. The overall loss function \\(L\\) can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.447, + 0.443, + 0.786, + 0.459 + ], + "angle": 0, + "content": "\\[\nL = L _ {s} + \\lambda_ {u} L _ {u}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.786, + 0.507 + ], + "angle": 0, + "content": "where \\( L_{s} \\) and \\( L_{u} \\) denote supervised loss and unsupervised loss respectively, \\( \\lambda_{u} \\) is the weight of unsupervised loss to balance the loss terms. Following the classic COD method, we use binary cross-entropy loss \\( L_{bce} \\) for training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.582 + ], + "angle": 0, + "content": "During the training process, we employ a combination of weak augmentation \\(\\mathcal{A}^w (\\cdot)\\) and strong augmentation \\(\\mathcal{A}^s (\\cdot)\\) strategies. Weak augmentation is applied to labeled data to mitigate overfitting, while unlabeled data undergo various data perturbations under strong augmentation to create diverse perspectives of the same image. The supervised loss \\(L_{s}\\) is defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.372, + 0.594, + 0.787, + 0.634 + ], + "angle": 0, + "content": "\\[\nL _ {s} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} L _ {b c e} \\left(F \\left(\\mathcal {A} ^ {w} \\left(x _ {i} ^ {l}\\right); \\Theta_ {s}\\right), y _ {i}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.637, + 0.788, + 0.818 + ], + "angle": 0, + "content": "where \\( F(\\mathcal{A}(x_i);\\Theta) \\) represents the detection results obtained by the model \\( \\Theta \\) for the \\( i \\)-th image with augmentations \\( \\mathcal{A}(\\cdot) \\). For unlabeled images, we first apply weak augmentation, denoted as \\( \\mathcal{A}^w (\\cdot) \\), before forwarding them to the teacher model. This initial step is crucial for generating reliable pseudo-labels \\( \\widehat{y}_i \\) under variations that do not significantly alter the core features of the images. These pseudo-labels serve as a form of soft supervision for the student model. Subsequently, the same images are subjected to strong augmentation, \\( \\mathcal{A}^s (\\cdot) \\), and then passed to the student model. This process introduces a higher level of variability and complexity, simulating more challenging conditions for the student model to adapt to. The student model generates predictions \\( p_i \\) based on these strongly augmented images, leveraging the pseudo-labels \\( \\widehat{y}_i \\) for guidance in learning from the unlabeled data. It can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.336, + 0.826, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\widehat {y _ {i}} = F (\\mathcal {A} ^ {w} (x _ {i} ^ {u}); \\Theta_ {t}), p _ {i} = F (\\mathcal {A} ^ {s} (\\mathcal {A} ^ {w} (x _ {i} ^ {u})); \\Theta_ {s}). \\tag {3}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.639, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CamoTeacher" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.631, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.295, + 0.626, + 0.304 + ], + "angle": 0, + "content": "DRCL: Dual-Rotation Consistency Learning PCL: Pixel-wise Consistency Learning ICL: Instance-wise Consistency Learning" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.144, + 0.785, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.319, + 0.788, + 0.43 + ], + "angle": 0, + "content": "Fig. 4: The overall pipeline of our CamoTeacher. CamoTeacher consists of a teacher model and a student model. The teacher model generates pseudo-labels to optimize the student model, while the teacher model is updated through EMA [36] from the student. To mitigate pseudo-labels' noise, we introduced Dual-Rotation Consistency Learning (DRCL), which involves Pixel-wise Consistency Learning (PCL) and Instance-wise Consistency Learning (ICL). DRCL weights the loss function based on pixel-wise inconsistency and instance-wise consistency in pseudo-labels across different rotation views." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.461, + 0.6, + 0.476 + ], + "angle": 0, + "content": "Hence, the unsupervised loss \\(L_{u}\\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.488, + 0.787, + 0.529 + ], + "angle": 0, + "content": "\\[\nL _ {u} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} L _ {b c e} \\left(p _ {i}, \\widehat {y _ {i}}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.788, + 0.646 + ], + "angle": 0, + "content": "Finally, the student model is intensively trained using the total loss \\( L \\), which encapsulates both the supervised and unsupervised learning aspects of the semi-supervised framework. This approach ensures that the student model benefits from both labeled and pseudo-labeled data, enhancing its detection capabilities. Concurrently, the teacher model is methodically updated through the Exponential Moving Average (EMA) mechanism [36], which effectively distills students' knowledge while preventing noise, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.658, + 0.786, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\Theta_ {t} \\leftarrow \\eta \\Theta_ {t} + (1 - \\eta) \\Theta_ {s}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.686, + 0.607, + 0.702 + ], + "angle": 0, + "content": "where \\(\\eta\\) is hyper-parameters denotes the keeping rate." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.724, + 0.564, + 0.74 + ], + "angle": 0, + "content": "3.3 Dual-Rotation Consistency Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Due to the camouflaging nature of the objects, pseudo-labels contain a substantial amount of noise, and directly using them to optimize the student model potentially harm the model's performance. To address this problem, one of the most intuitive possible methods is to set a fixed high threshold to filter high-quality pseudo-labels, but this results in a low recall rate and makes it difficult to fully utilize the supervision information of pseudo-labels. To this end, we propose" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "X.Lai et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "Dual-Rotation Consistency Learning (DRCL) to dynamically adjust the weight of the pseudo-labels to reduce the impact of noise." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.222 + ], + "angle": 0, + "content": "We randomly apply two separate rotations to the image \\( x_{i} \\), which has previously been flipped and random resized, resulting in two distinct rotation views \\( x_{i}^{r_{1}} \\) and \\( x_{i}^{r_{2}} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.223, + 0.785, + 0.239 + ], + "angle": 0, + "content": "\\[\nx _ {i} ^ {r _ {1}} = R \\left(\\mathcal {A} ^ {w} \\left(x _ {i}\\right), \\theta_ {1}\\right), x _ {i} ^ {r _ {2}} = R \\left(\\mathcal {A} ^ {w} \\left(x _ {i}\\right), \\theta_ {2}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.244, + 0.787, + 0.336 + ], + "angle": 0, + "content": "where \\( x_{i}^{r} = R(x_{i},\\theta) \\) represents rotate the input image \\( x_{i} \\) by \\( \\theta \\) degrees. We input the acquired rotation view into the teacher model to obtain the corresponding prediction, i.e. \\( \\widehat{y}_i^r = F(x_i^r;\\Theta_t) \\). Subsequently, we apply an opposite rotation of \\( -\\theta \\) to return the prediction to its original horizontal orientation, resulting in \\( \\widehat{y}_i^{h_1} \\) and \\( \\widehat{y}_i^{h_2} \\), facilitating the calculation of prediction inconsistency under different rotation views, respectively," + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.346, + 0.786, + 0.364 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {i} ^ {h _ {1}} = R \\left(\\hat {y} _ {i} ^ {r _ {1}}, - \\theta_ {1}\\right), \\hat {y} _ {i} ^ {h _ {2}} = R \\left(\\hat {y} _ {i} ^ {r _ {2}}, - \\theta_ {2}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.784, + 0.402 + ], + "angle": 0, + "content": "Note that rotation introduces black boundary areas, which are not involved in the implement of DRCL." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.785, + 0.449 + ], + "angle": 0, + "content": "Since the noise levels vary across different regions of pseudo-labels and among different pseudo-labels, we introduce PCL and ICL dynamically adapt the contributions of different pixels within pseudo-labels and across various pseudo-labels." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.469, + 0.784, + 0.503 + ], + "angle": 0, + "content": "Pixel-wise Consistency Learning We subtract the horizontal predictions \\(\\widehat{y}_i^{h_1}\\) and \\(\\widehat{y}_i^{h_2}\\) at the pixel level to obtain the pixel-wise inconsistency \\(\\Delta_i\\)," + }, + { + "type": "equation", + "bbox": [ + 0.442, + 0.513, + 0.785, + 0.53 + ], + "angle": 0, + "content": "\\[\n\\Delta_ {i} = \\left| \\widehat {y} _ {i} ^ {h _ {1}} - \\widehat {y} _ {i} ^ {h _ {2}} \\right|. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.785, + 0.646 + ], + "angle": 0, + "content": "The pixel-wise inconsistency \\(\\Delta_{i}\\) across different views reflects the reliability of pseudo-labels. However, in cases where both rotation views' prediction are close to 0.5, \\(\\Delta_{i}\\) cannot effectively distinguish them. These predictions exhibit high uncertainty, meaning they cannot be definitively classified as foreground or background, and are likely to represent noisy labels. Consequently, it becomes necessary to diminish their influence by lowering their weights. Hence, we calculate the average of horizontal predictions \\(\\widehat{y}_i^h\\)," + }, + { + "type": "equation", + "bbox": [ + 0.434, + 0.656, + 0.785, + 0.673 + ], + "angle": 0, + "content": "\\[\n\\widehat {y} _ {i} ^ {h} = \\operatorname {a v g} \\left(\\widehat {y} _ {i} ^ {h _ {1}}, \\widehat {y} _ {i} ^ {h _ {2}}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.683, + 0.784, + 0.712 + ], + "angle": 0, + "content": "where \\( avg(\\cdot, \\cdot) \\) represents calculate the average of two inputs in pixel-level, and the use its L2 distance from 0.5 as a component to adjust the weights." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.713, + 0.784, + 0.743 + ], + "angle": 0, + "content": "As a result, we derive the pixel-wise consistency weight \\(\\omega_{i}^{pc}\\) based on the pixel-wise inconsistency among different rotation views, as:" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.754, + 0.785, + 0.771 + ], + "angle": 0, + "content": "\\[\n\\omega_ {i} ^ {p c} = \\left(1 - \\Delta_ {i} ^ {\\alpha}\\right) \\left| \\left| \\hat {y} _ {i} ^ {h} - \\mu \\right| \\right| _ {2} ^ {2}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.786, + 0.84 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a hyper-parameter and \\(\\mu = 0.5\\). This dynamic pixel-wise consistency weight \\(\\omega_{i}^{pc}\\) assigns greater weight to regions with consistent predictions across various rotation views, while regions with inconsistent predictions receive a smaller weight." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.638, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CamoTeacher" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.146, + 0.6, + 0.163 + ], + "angle": 0, + "content": "In conclusion, we formulate the PCL loss \\( L_{u}^{PC} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.172, + 0.786, + 0.279 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {u} ^ {P C} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\omega_ {i} ^ {p c} L _ {b c e} (p _ {i}, \\widehat {y} _ {i} ^ {r _ {1}}) \\\\ = - \\frac {1}{N H W} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {H \\times W} \\omega_ {i, j} ^ {p c} \\left[ \\hat {y} _ {i, j} ^ {r _ {1}} \\log p _ {i, j} \\right. \\tag {11} \\\\ + \\left(1 - \\widehat {y} _ {i, j} ^ {r _ {1}}\\right) \\log \\left(1 - p _ {i, j}\\right) ], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.284, + 0.784, + 0.314 + ], + "angle": 0, + "content": "adaptively adjusts the weight of each pixel to ensure comprehensive supervision of student model while avoiding confirmation bias." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.334, + 0.785, + 0.47 + ], + "angle": 0, + "content": "Instance-wise Consistency Learning The degree of camouflage varies between different images, leading to significant variations in the quality of pseudolabels across images. Treating all pseudo-labels equally would be unreasonable. Unfortunately, for unlabeled images, assessing the quality of pseudo-labels is challenging because there are no GT labels available. We have observed a positive correlation between the instance-wise consistency of two rotation views and the quality of pseudo-labels, as quantified by SSIM [39]. Base on this, we introduce ICL to adjust the contributions of pseudo-labels with varying quality. We represent the instance-wise consistency weight \\(\\omega_{i}^{ic}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.477, + 0.785, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\omega_ {i} ^ {i c} = \\left(S S I M \\left(\\widehat {y} _ {i} ^ {h _ {1}}, \\widehat {y} _ {i} ^ {h _ {2}}\\right)\\right) ^ {\\beta}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.784, + 0.531 + ], + "angle": 0, + "content": "where \\(\\beta\\) is a hyper-parameter used to adjust the distribution relationship between instance-wise consistency and pseudo-label quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.532, + 0.784, + 0.562 + ], + "angle": 0, + "content": "We use intersection-over-union (IoU) loss as the instance-wise restriction, therefore, the ICL loss can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.57, + 0.785, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {u} ^ {I C} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\omega_ {i} ^ {i c} L _ {i o u} (p _ {i}, \\widehat {y} _ {i} ^ {r _ {1}}) \\\\ = \\frac {1}{N H W} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {H \\times W} \\omega_ {i} ^ {i c} \\left(1 - \\frac {p _ {i , j} \\widehat {y} _ {i , j} ^ {r _ {1}}}{p _ {i , j} + \\widehat {y} _ {i , j} ^ {r _ {1}} - p _ {i , j} \\widehat {y} _ {i , j} ^ {r _ {1}}}\\right). \\tag {13} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.662, + 0.784, + 0.693 + ], + "angle": 0, + "content": "Therefore, the final total \\( L \\) loss comprises three components: supervised loss \\( L_{s} \\), PCL loss \\( L_{u}^{LC} \\), and ICL loss \\( L_{u}^{GC} \\), which can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.699, + 0.785, + 0.719 + ], + "angle": 0, + "content": "\\[\nL = L _ {s} + \\lambda_ {u} ^ {p c} L _ {u} ^ {P C} + \\lambda_ {u} ^ {i c} L _ {u} ^ {I C}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.724, + 0.508, + 0.74 + ], + "angle": 0, + "content": "where \\(\\lambda_u^{pc},\\lambda_u^{ic}\\) are the hyper-parameters." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.76, + 0.367, + 0.777 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.788, + 0.432, + 0.804 + ], + "angle": 0, + "content": "4.1 Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Dataset. We evaluate our CamoTeacher on four benchmark datasets, CAMO [21], CHAMELEON [33], COD10K [7], and NC4K [25]. In CAMO, there are" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "X.Lai et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 1: Experimental results on four COD benchmark datasets includes CAMO, CHAMELEON, COD10K, NC4K. \\(\\dagger\\) denotes using the same data augmentation as semi-supervised training. \\(\\uparrow\\) indicates the higher the score the better, and \\(\\downarrow\\) indicates the lower the better. \"n%\" means that \"n%\" of the data was utilized as labeled data, with the remainder as unlabeled data. As the data partitioning is random, we conducted 5 fold experiments and the results are the average of all 5 folds." + }, + { + "type": "table", + "bbox": [ + 0.219, + 0.242, + 0.784, + 0.307 + ], + "angle": 0, + "content": "
CHAMELEON (76)
Methods1% (41)5% (202)10% (404)
\\( S_{\\alpha} \\uparrow \\)\\( F_{\\beta}^{w} \\uparrow \\)\\( E_{\\phi}^{m} \\uparrow \\)\\( E_{\\phi}^{x} \\uparrow \\)\\( F_{\\beta}^{m} \\uparrow \\)\\( M \\downarrow \\)\\( S_{\\alpha} \\uparrow \\)\\( F_{\\beta}^{w} \\uparrow \\)\\( E_{\\phi}^{m} \\uparrow \\)\\( E_{\\phi}^{x} \\uparrow \\)\\( F_{\\beta}^{m} \\uparrow \\)\\( M \\downarrow \\)\\( S_{\\alpha} \\uparrow \\)\\( F_{\\beta}^{\\omega } \\uparrow \\)\\( E_{\\phi}^{m} \\uparrow \\)\\( E_{\\phi}^{x} \\uparrow \\)\\( F_{\\beta}^{m} \\uparrow \\)\\( M \\downarrow \\)
Supervised0.6050.2680.5820.7410.4350.2160.6130.2830.6010.7320.4390.1850.7140.4340.7250.8290.5920.110
Supervised†0.6590.3310.6550.7870.4990.1590.7280.4540.7440.8440.6070.106
Mean Teacher0.5370.1990.4180.6360.2290.2040.6110.3090.5240.7450.3530.1370.6790.4500.6500.8120.5120.102
CamoTeacher0.6520.4760.7140.7620.5580.0930.7290.5870.7850.8220.6560.0700.7560.6170.8130.8510.6840.065
" + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.308, + 0.784, + 0.372 + ], + "angle": 0, + "content": "
CAMO (250)
Methods1% (41)5% (202)10% (404)
\\( S_{\\alpha} \\uparrow \\)\\( F_{w}^{w} \\uparrow \\)\\( E_{m}^{m} \\uparrow \\)\\( E_{\\sigma}^{e} \\uparrow \\)\\( F_{g}^{m} \\uparrow \\)\\( M \\downarrow \\)\\( S_{\\alpha} \\uparrow \\)\\( F_{w}^{w} \\uparrow \\)\\( E_{\\sigma}^{m} \\uparrow \\)\\( E_{\\sigma}^{e} \\uparrow \\)\\( F_{g}^{m} \\uparrow \\)\\( M \\downarrow \\)\\( S_{\\alpha} \\uparrow \\)\\( F_{w}^{w} \\uparrow \\)\\( E_{m}^{m} \\uparrow \\)\\( E_{\\sigma}^{e} \\uparrow \\)\\( F_{g}^{m} \\uparrow \\)\\( M \\downarrow \\)
Supervised0.5730.2710.5410.7040.4020.2510.5920.2980.5810.7040.4320.2160.6520.3900.6470.7600.5250.159
Supervised †0.5960.2980.5800.7170.4420.2350.6100.3160.5950.7200.4500.2050.6670.4090.6660.7790.5460.153
Mean Teacher0.5180.2070.3990.6200.2270.2260.5750.2860.4820.7080.3220.1840.6250.3970.5780.7730.4540.150
CamoTeacher0.6210.4560.6690.7360.5450.1360.6690.5230.7110.7750.6010.1220.7010.5600.7420.7950.6350.112
" + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.373, + 0.784, + 0.436 + ], + "angle": 0, + "content": "
COD10K (2026)
Methods1% (41)5% (202)10% (404)
So ↑Fw↑Em↑Eax↑Eax↑Fm↑M↓So ↑Fw↑Em↑Eax↑Eax↑Fm↑M↓So ↑Fw↑Em↑Eax↑Eax↑Fm↑M↓
Supervised0.5660.1890.5760.7230.3560.2090.5960.2200.6270.7470.4030.1640.6750.3360.7260.8100.5240.095
Supervised †0.5820.2060.6090.7430.3860.1940.6210.2460.6590.7740.4440.1470.6800.3380.7260.8120.5280.097
Mean Teacher0.5460.1680.4410.6330.2260.1610.6210.2720.5550.7320.3430.1070.6830.4040.6660.7990.4820.078
CamoTeacher0.6990.5170.7880.7970.5820.0620.7450.5830.8270.8400.6440.0500.7590.5940.8360.8540.6520.049
" + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.438, + 0.784, + 0.501 + ], + "angle": 0, + "content": "
NC4K (4121)
Methods1% (41)5% (202)10% (404)
\\( S_{\\alpha} \\uparrow \\)\\( E_{m}^{w} \\uparrow \\)\\( E_{m}^{m} \\uparrow \\)\\( E_{e}^{s} \\uparrow \\)\\( F_{m}^{m} \\uparrow \\)\\( M \\downarrow \\)\\( S_{\\alpha} \\uparrow \\)\\( F_{m}^{w} \\uparrow \\)\\( E_{m}^{m} \\uparrow \\)\\( E_{e}^{s} \\uparrow \\)\\( F_{m}^{m} \\uparrow \\)\\( M \\downarrow \\)\\( S_{\\alpha} \\uparrow \\)\\( F_{m}^{w} \\uparrow \\)\\( E_{m}^{m} \\uparrow } \\)\\( E_{e}^{s} \\uparrow \\)\\( F_{m}^{m} \\uparrow \\)\\( M \\downarrow \\)
Supervised0.6180.2910.6040.7520.4740.2130.6490.3320.6600.7730.5240.1790.7270.4620.7540.8330.6430.116
Supervised †0.6380.3130.6390.7680.5100.2010.6760.3620.6890.7980.5620.1630.7350.4710.7600.8370.6500.114
Mean Teacher0.5410.2130.4240.6370.2580.1930.6340.3550.5560.7670.4200.1400.7000.4920.6700.8270.5650.109
CamoTeacher0.7180.5990.7790.8140.6750.0900.7770.6770.8340.8590.7390.0710.7910.6870.8420.8680.7460.068
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.7 + ], + "angle": 0, + "content": "2500 images, consisting of 1250 camouflage images and 1250 non-camouflage images. CHAMELEON contains 76 manually annotated images. COD10K dataset consists of 5066 camouflage images, 3000 background images, and 1934 non-camouflage images. NC4K is another large-scale COD test dataset that comprises 4,121 images. Based on the data partitioning in previous works [7,29,45], we used 3040 images from COD10K and 1000 images from CAMO as the training set for our experiments. The remaining images from both datasets were used as the test set. During the training process, we followed a semi-supervised segmentation [15,47] data partitioning approach. We sampled \\(1\\%\\), \\(5\\%\\), \\(10\\%\\), \\(20\\%\\), and \\(30\\%\\) of the training set as labeled data, while the remaining portion was used as unlabeled data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.702, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Evaluation Metrics. Referring to previous work [16, 17, 29], we used 6 common evaluation indicators in COD to evaluate our CamoTeacher, including S-measure \\((S_{\\alpha})\\) [4], weight F-measure \\((F_{\\beta}^{w})\\) [26], mean E-measure \\((E_{\\phi}^{m})\\) [5], max E-measure \\((E_{\\phi}^{x})\\), mean F-measure \\((F_{\\beta}^{m})\\) [1], and mean absolute error \\((M)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Implementation Details. The proposed CamoTeacher is implemented with PyTorch. We adopt SINet [7] as the baseline of COD model. We use an SGD optimizer with a momentum of 0.9 and a polynomial learning-rate decay with an initial value of 0.01 to train the student model. The training epoch is set to 40 with 10 epochs for burn-in phase. And the batch size is 20, with a ratio of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.639, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CamoTeacher" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.144, + 0.784, + 0.172 + ], + "angle": 0, + "content": "Table 2: Results of CamoTeacher on different COD models with \\(10\\%\\) labeled data on COD10K." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.185, + 0.784, + 0.349 + ], + "angle": 0, + "content": "
ModelSettingSα↑Fβw↑Eφm↑Eφx↑Fβm↑M↓
CNN-Based Models
SINet-v2 [6] (TPAMI2022)supervised0.7320.5440.8190.8350.5860.056
CamoTeacher0.7600.6050.8350.8430.6520.051
SegMaR [17] (CVPR2022)supervised0.7550.5830.8370.8510.6230.050
CamoTeacher0.7750.6290.8560.8640.6750.045
Transformer-Based Models
DTINet [24] (ICPR2022)supervised0.7710.6260.8570.8650.6700.043
CamoTeacher0.7940.6610.8730.8820.6950.040
FSPNet [14] (CVPR2023)supervised0.7890.6220.8230.8670.6750.042
CamoTeacher0.8170.6840.8840.9030.7200.037
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.788, + 0.482 + ], + "angle": 0, + "content": "1:1 between labeled and unlabeled data, i.e. each batch consists of 10 labeled and 10 unlabeled images. During training and inference, each image is resized to \\(352 \\times 352\\). The teacher model is updated via EMA with a momentum \\(\\eta\\) of 0.996. Weak data augmentation involves Random Flip and Random Scale, while strong data augmentation entails color space transformations including Identity, Autocontrast, Equalize, Gaussian blur, Contrast, Sharpness, Color, Brightness, Hue, Posterize, Solarize, with a random selection of up to 3 from this list." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.501, + 0.373, + 0.515 + ], + "angle": 0, + "content": "4.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.787, + 0.735 + ], + "angle": 0, + "content": "Comparison with Baselines. In Tab. 1, we compared CamoTeacher, the supervised baseline [7], and Mean Teacher [36]. Mean Teacher does not show improvements over the supervised in the semi-supervised setting due to treating all pseudo-labels and regions equally, which does not account for pseudo-labels noise problem. For instance, on the NC4K test set, under the \\(1\\%\\) labeled data, the \\(S_{\\alpha}\\) dropped by \\(9.7\\%\\) compared to supervised. Compared to Mean Teacher, CamoTeacher integrated with DRCL effectively mitigates pseudo-label noise issues and efficiently utilizes pseudo-label information. It demonstrates substantial improvements over both supervised and Mean Teacher across various proportions of labeled data, especially in scenarios with extremely limited. For instance, on \\(1\\%\\) NC4K, CamoTeacher exhibited a substantial increase of \\(28.6\\%\\) in the \\(F_{\\beta}^{w}\\) and a \\(12.3\\%\\) enhancement of \\(S_{\\alpha}\\). Compared with Mean Teacher, the performance gains by CamoTeacher are more significant, which are \\(38.6\\%\\) and \\(17.7\\%\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In Fig. 1(b), we visualize the performance changes of supervised, Mean teacher and our CamoTeacher on COD10K under different labeled data ratios of \\(1\\%\\), \\(5\\%\\), \\(10\\%\\), and further \\(20\\%\\), \\(30\\%\\), which intuitively reflects the effectiveness of CamoTeacher. It's evident that metrics like \\(F_{\\beta}^{w}\\), \\(F_{\\beta}^{m}\\), \\(E_{\\phi}^{m}\\), and \\(M\\) surpass \\(100\\%\\) of fully supervised performance with just \\(5\\%\\) labeled data. Additionally, with \\(30\\%\\) labeled data, all metrics significantly exceed \\(100\\%\\) of fully supervised performance." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "X.Lai et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 3: Result of CamoTeacher on COD10K with additional data. We use \\(100\\%\\) train dataset as labeled data and test dataset as the unlabeled data. \"U\" represents the unsupervised setting, \"W\" denotes the weakly-supervised setting, where weakly-supervision results from scribble annotations. \"L100%\" indicates using \\(100\\%\\) of the training set as labeled data, \"U(CAMO)\" denotes utilizing CAMO as unlabeled data, and likewise for others." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.24, + 0.785, + 0.464 + ], + "angle": 0, + "content": "
ModelSettingSα↑Fwβ↑Emφ↑Exφ↑Fmβ↑M↓
Unsupervised
SelfMask [31] (CVPRW2022)U0.645 0.440 0.687 0.728 0.478 0.125
FOUND [32] (CVPR2023)U0.670 0.482 0.751 0.753 0.520 0.085
UCOS-DA [46] (ICCVW2023)U0.689 0.513 0.740 0.741 0.546 0.086
Weakly-Supervised
SCWSSOD [44] (AAAI2021)W0.710 0.546 0.805 - - 0.055
SCOD [11](AAAI2023)W0.733 0.576 0.832 0.845 0.633 0.049
WS-SAM [10] (NeurIPS2023)W0.803 - 0.878 - - 0.038
Semi-Supervised
CamoTeacher SINetL100%0.771 0.551 0.806 0.868 0.634 0.051
L100%+U(CAMO)0.779 0.554 0.799 0.875 0.632 0.051
L100%+U(NC4K)0.785 0.574 0.812 0.879 0.642 0.049
L100%+U(CAMO,NC4K)0.787 0.576 0.813 0.881 0.644 0.048
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.643 + ], + "angle": 0, + "content": "Application to Existing COD Methods. We applied the CamoTeacher approach to additional COD models, including CNN-based model SINet-v2 [6] and SegMaR [17] and Transformed-based model DTINet [24] and FSPNet [14], and the corresponding results are presented in Tab. 2. In terms of SINet-v2, CamoTeacher demonstrated a \\(2.8\\%\\) improvement in the \\(S_{\\alpha}\\) metric and a \\(6.1\\%\\) improvement in the \\(F_{\\beta}^{w}\\) metric compared to the supervised baseline. Based on FSPNet, CamoTeacher achieved a \\(2.8\\%\\) enhancement in the \\(S_{\\alpha}\\) metric and a \\(6.2\\%\\) improvement in the \\(F_{\\beta}^{w}\\) metric without utilizing additional discriminative annotation. These outcomes provide robust evidence of the effectiveness of CamoTeacher." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Addition unlabeled data. In Tab. 3, based on the base model SINet, we train CamoTeacher with \\(100\\%\\) labeled data and check whether additional unlabeled data can further improve the performance. Compared with unsupervised and weakly-supervised methods, it shows the superiority of our method. We use the testset CAMO and NC4K as unlabeled data, and we find that adding more additional data CamoTeacher can make SINet continued to gain, such as \\(S_{\\alpha}\\) metric increased by \\(1.6\\%\\) and \\(F_{\\beta}^{w}\\) metric increased by \\(2.5\\%\\). We achieved significantly better performance than unsupervised methods [31,32,46] using existing labeled data. Compared with the weakly-supervision methods [10,11,44], we do not need additional annotation for the new unlabeled data, and to a certain extent, as the more unlabeled data, the performance will get better and better. And at the same time, replacing the base model with a stronger model will get better results." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.639, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CamoTeacher" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.145, + 0.784, + 0.214 + ], + "angle": 0, + "content": "Table 4: The ablation study of Pixel-wise Table 5: The ablation study of various Consistency Learning (PCL) and Instance- weight approach of Pixel-wise Consistency wise Consistency Learning (ICL) with \\(5\\%\\) Learning (PCL) with \\(5\\%\\) labeled data on labeled data on COD10K. COD10K. In this experiment \\(\\alpha\\) is set to 1/4." + }, + { + "type": "table", + "bbox": [ + 0.234, + 0.214, + 0.484, + 0.289 + ], + "angle": 0, + "content": "
PCLICL\\( {S}_{\\alpha } \\uparrow \\)\\( {F}_{\\beta }^{w} \\uparrow \\)\\( {E}_{\\phi }^{m} \\uparrow \\)\\( {E}_{\\phi }^{x} \\uparrow \\)\\( {F}_{\\beta }^{m} \\uparrow \\)\\( M \\downarrow \\)
\\( L_{bce} \\)\\( {\\omega }^{pc} \\)\\( L_{iou} \\)\\( {\\omega }^{tc} \\)
✓0.6820.3700.6360.7870.4610.091
✓0.7340.5720.8170.8210.6260.057
✓✓0.7240.5010.7380.8320.5740.058
✓✓0.7420.5850.8280.8340.6390.052
✓✓✓0.7450.5860.8300.8400.6430.051
✓✓✓✓0.7460.5860.8300.8430.6450.050
" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.228, + 0.781, + 0.287 + ], + "angle": 0, + "content": "
weight approach\\( {S}_{\\alpha } \\uparrow {F}_{\\beta }^{w} \\uparrow {E}_{\\phi }^{m} \\uparrow {E}_{\\phi }^{x} \\uparrow {F}_{\\beta }^{m} \\uparrow M \\downarrow \\)
\\( {\\omega }^{pc} = {\\widehat{y}}^{r} \\)0.6980.3930.6840.8040.5000.110
\\( {\\omega }^{pc} = {\\begin{Vmatrix}{\\widehat{y}}^{r} - \\mu \\end{Vmatrix}}_{2}^{2} \\)0.7170.4850.7340.8220.5590.061
\\( {\\omega }^{pc} = 1 - {\\Delta }^{\\alpha } \\)0.7190.4810.7200.8270.5530.062
\\( {\\omega }^{pc} = \\left( {1 - {\\Delta }^{\\alpha }}\\right) {\\widehat{y}}^{r} \\)0.6950.4050.6960.7950.5060.114
\\( {\\omega }^{pc} = \\left( {1 - {\\Delta }^{\\alpha }}\\right) {\\begin{Vmatrix}{\\widehat{y}}^{r} - \\mu \\end{Vmatrix}}_{2}^{2}\\left\\lbrack \\begin{array}{llllllll} 0 & {724} & {0.501} & {0.738} & {0.832} & {0.574} & {0.058} \\end{array}\\right\\rbrack \\)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.307, + 0.784, + 0.378 + ], + "angle": 0, + "content": "Table 6: The ablation study of hyper- Table 7: The ablation study of hyperparameters \\(\\alpha\\) in Pixel-wise Consistency parameters \\(\\beta\\) in Instance-wise Consistency Learning (PCL) with \\(5\\%\\) labeled data on Learning (ICL) with \\(5\\%\\) labeled data on COD10K. In this experiment \\(\\beta\\) is set to 4, COD10K. In this experiment \\(\\alpha\\) is set to \\(\\lambda_u^{PC}\\) is set to 8, and \\(\\lambda_u^{IC}\\) is set to 0.3. \\(1/4, \\lambda_u^{PC}\\) is set to 8, and \\(\\lambda_u^{IC}\\) is set to 0.3." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.389, + 0.497, + 0.474 + ], + "angle": 0, + "content": "
αSα↑Fwβ↑Eφm↑Eφx↑Fmβ↑M↓
20.7370.5650.8130.8360.6290.052
10.7390.5660.8150.8320.6270.056
1/20.7450.5750.8200.8390.6370.051
1/40.7470.5790.8250.8420.6380.051
1/80.7450.5780.8250.8380.6360.052
" + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.39, + 0.768, + 0.474 + ], + "angle": 0, + "content": "
βSα↑Fβw↑Eφm↑Eφx↑Fβm↑M↓
10.7430.5760.8240.8390.6360.051
20.7420.5730.8230.8370.6320.052
40.7470.5790.8250.8420.6380.051
50.7440.5780.8240.8410.6390.051
60.7450.5750.8220.8390.6350.052
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.509, + 0.389, + 0.525 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.788, + 0.669 + ], + "angle": 0, + "content": "Ablation on PCL and ICL. We conducted ablation experiments on different combinations of weight and loss functions for PCL and ICL. From Tab. 4, increasing the weights \\(\\omega^{pc}\\) and \\(\\omega^{ic}\\) improves the model's performance. Specifically, with \\(\\omega^{pc}\\) results in a higher \\(F_{\\beta}^{w}\\) score, increasing from 0.370 to 0.501. Similarly, with \\(\\omega^{ic}\\) improved the \\(F_{\\beta}^{w}\\) metric from 0.572 to 0.585. Combining these two weighted loss functions led to a significant performance improvement, with the \\(F_{\\beta}^{w}\\) score increasing from 0.370 to 0.586 (an increase of \\(21.6\\%\\)). This demonstrates the effectiveness of PCL and ICL in enhancing the model's performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.672, + 0.788, + 0.748 + ], + "angle": 0, + "content": "Ablation of weight approach in PCL. The choice of method for obtaining \\(\\omega^{pc}\\) significantly affects the performance of the model. Therefore, we conducted several experiments to select the most effective method for obtaining \\(\\omega^{pc}\\). From Tab. 5, it can be observed that the best performance is achieved when \\(\\omega^{pc}\\) is obtained using Eq. (10)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Ablation of hyper-parameters in PCL and ICL. We perform ablation experiments on the hyper-parameters associated with the PCL and ICL modules. As shown in Tab. 6, the optimal results are achieved when setting the hyper-parameter \\(\\alpha\\), which corresponds to the coefficient \\(\\omega^{pc}\\), to \\(1/4\\). Similarly, according to Tab. 7, the best performance is observed when the hyper-parameter \\(\\beta\\), related to the coefficient \\(\\omega^{ic}\\), is set to 4." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "X.Lai et al." + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.147, + 0.725, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.394, + 0.751, + 0.409 + ], + "angle": 0, + "content": "Fig. 5: Visualization of predictions in different proportions of labeled data." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.439, + 0.43, + 0.454 + ], + "angle": 0, + "content": "4.4 Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.788, + 0.617 + ], + "angle": 0, + "content": "Result visualization. We conducte visualizations of CamoTeacher's results under varying proportions of labeled data. The model is trained with \\(1\\%\\), \\(5\\%\\), \\(10\\%\\), \\(20\\%\\) and \\(30\\%\\) labeled data, and we compare its predicted masks with those produced by SINet under fully supervised conditions (as shown in Fig. 5). Notably, even with only \\(10\\%\\) labeled data, CamoTeacher's predicted masks exhibit a remarkable similarity to the fully supervised SINet predictions. In fact, in certain detailed predictions, CamoTeacher even surpasses the performance of fully supervised SINet. These findings highlight the impressive capability of CamoTeacher in leveraging limited labeled data for effective camouflaged object detection." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.641, + 0.37, + 0.657 + ], + "angle": 0, + "content": "5 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we present the first end-to-end semi-supervised camouflaged object detection model called CamoTeacher. To address the issue of substantial noise within pseudo-labels in semi-supervised COD, including both local noise and global noise, we've introduced a novel approach called Dual-Rotation Consistency Learning (DRCL), including Pixel-wise Consistency Learning (PCL) and Instance-wise Consistency Learning (ICL). DRCL assists the model in alleviating noise issues, effectively leveraging pseudo-labels information, so that the model can receive adequate supervision while avoiding confirmation bias. Extensive experiments validate CamoTeacher's superior performance while significantly reducing annotation expenses. We believe that CamoTeacher can serve as a strong baseline for future research on semi-supervised COD." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.639, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CamoTeacher" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.284 + ], + "angle": 0, + "content": "This work was supported by National Science and Technology Major Project (No. 2022ZD0118201), the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China a (No. U21B2037, No. U22B2051, No. U23A20383, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No.2022J06001)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.308, + 0.323, + 0.323 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.34, + 0.785, + 0.38 + ], + "angle": 0, + "content": "1. Achanta, R., Hemami, S., Estrada, F., Susstrunk, S.: Frequency-tuned salient region detection. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 1597-1604. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.383, + 0.785, + 0.422 + ], + "angle": 0, + "content": "2. Chen, H., Tao, R., Fan, Y., Wang, Y., Wang, J., Schiele, B., Xie, X., Raj, B., Savvides, M.: Softmatch: Addressing the quantity-quality trade-off in semi-supervised learning. arXiv preprint arXiv:2301.10921 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.424, + 0.785, + 0.464 + ], + "angle": 0, + "content": "3. Chen, X., Yuan, Y., Zeng, G., Wang, J.: Semi-supervised semantic segmentation with cross pseudo supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2613-2622 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.466, + 0.785, + 0.506 + ], + "angle": 0, + "content": "4. Fan, D.P., Cheng, M.M., Liu, Y., Li, T., Borji, A.: Structure-measure: A new way to evaluate foreground maps. In: Proceedings of the IEEE international conference on computer vision. pp. 4548-4557 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.508, + 0.785, + 0.547 + ], + "angle": 0, + "content": "5. Fan, D.P., Gong, C., Cao, Y., Ren, B., Cheng, M.M., Borji, A.: Enhanced-alignment measure for binary foreground map evaluation. arXiv preprint arXiv:1805.10421 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.549, + 0.785, + 0.575 + ], + "angle": 0, + "content": "6. Fan, D.P., Ji, G.P., Cheng, M.M., Shao, L.: Concealed object detection. IEEE transactions on pattern analysis and machine intelligence 44(10), 6024-6042 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.577, + 0.785, + 0.617 + ], + "angle": 0, + "content": "7. Fan, D.P., Ji, G.P., Sun, G., Cheng, M.M., Shen, J., Shao, L.: Camouflaged object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2777-2787 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.619, + 0.785, + 0.645 + ], + "angle": 0, + "content": "8. Grandvalet, Y., Bengio, Y.: Semi-supervised learning by entropy minimization. Advances in neural information processing systems 17 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.7 + ], + "angle": 0, + "content": "9. He, C., Li, K., Zhang, Y., Tang, L., Zhang, Y., Guo, Z., Li, X.: Camouflaged object detection with feature decomposition and edge reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22046-22055 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.702, + 0.785, + 0.755 + ], + "angle": 0, + "content": "10. He, C., Li, K., Zhang, Y., Xu, G., Tang, L., Zhang, Y., Guo, Z., Li, X.: Weakly-supervised concealed object segmentation with sam-based pseudo labeling and multi-scale feature grouping. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.785, + 0.797 + ], + "angle": 0, + "content": "11. He, R., Dong, Q., Lin, J., Lau, R.W.: Weakly-supervised camouflaged object detection with scribble annotations. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 781-789 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.785, + 0.839 + ], + "angle": 0, + "content": "12. Hu, J., Lin, J., Cai, W., Gong, S.: Relax image-specific prompt requirement in sam: A single generic prompt for segmenting camouflaged objects. arXiv preprint arXiv:2312.07374 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.34, + 0.785, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "X.Lai et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "13. Hu, J., Chen, C., Cao, L., Zhang, S., Shu, A., Jiang, G., Ji, R.: Pseudo-label alignment for semi-supervised instance segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16337-16347 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.244 + ], + "angle": 0, + "content": "14. Huang, Z., Dai, H., Xiang, T.Z., Wang, S., Chen, H.X., Qin, J., Xiong, H.: Feature shrinkage pyramid for camouflaged object detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5557-5566 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.244, + 0.787, + 0.271 + ], + "angle": 0, + "content": "15. Hung, W.C., Tsai, Y.H., Liou, Y.T., Lin, Y.Y., Yang, M.H.: Adversarial learning for semi-supervised semantic segmentation. arXiv preprint arXiv:1802.07934 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.787, + 0.311 + ], + "angle": 0, + "content": "16. Ji, G.P., Fan, D.P., Chou, Y.C., Dai, D., Liniger, A., Van Gool, L.: Deep gradient learning for efficient camouflaged object detection. Machine Intelligence Research 20(1), 92-108 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.311, + 0.787, + 0.352 + ], + "angle": 0, + "content": "17. Jia, Q., Yao, S., Liu, Y., Fan, X., Liu, R., Luo, Z.: Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4713-4722 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.352, + 0.787, + 0.392 + ], + "angle": 0, + "content": "18. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.392, + 0.787, + 0.447 + ], + "angle": 0, + "content": "19. Lai, X., Tian, Z., Jiang, L., Liu, S., Zhao, H., Wang, L., Jia, J.: Semi-supervised semantic segmentation with directional context-aware consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1205-1214 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.447, + 0.787, + 0.473 + ], + "angle": 0, + "content": "20. Laine, S., Aila, T.: Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.473, + 0.787, + 0.514 + ], + "angle": 0, + "content": "21. Le, T.N., Nguyen, T.V., Nie, Z., Tran, M.T., Sugimoto, A.: Anabranch network for camouflaged object segmentation. Computer vision and image understanding 184, 45-56 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.514, + 0.787, + 0.555 + ], + "angle": 0, + "content": "22. Lee, D.H., et al.: Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In: Workshop on challenges in representation learning, ICML. vol. 3, p. 896. Atlanta (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.555, + 0.787, + 0.609 + ], + "angle": 0, + "content": "23. Liu, Y., Tian, Y., Chen, Y., Liu, F., Belagiannis, V., Carneiro, G.: Perturbed and strict mean teachers for semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4258-4267 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.609, + 0.787, + 0.65 + ], + "angle": 0, + "content": "24. Liu, Z., Zhang, Z., Tan, Y., Wu, W.: Boosting camouflaged object detection with dual-task interactive transformer. In: 2022 26th International Conference on Pattern Recognition (ICPR). pp. 140-146. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.65, + 0.787, + 0.704 + ], + "angle": 0, + "content": "25. Lv, Y., Zhang, J., Dai, Y., Li, A., Liu, B., Barnes, N., Fan, D.P.: Simultaneously localize, segment and rank the camouflaged objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11591-11601 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.704, + 0.787, + 0.745 + ], + "angle": 0, + "content": "26. Margolin, R., Zelnik-Manor, L., Tal, A.: How to evaluate foreground maps? In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 248-255 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.745, + 0.787, + 0.799 + ], + "angle": 0, + "content": "27. Mi, P., Lin, J., Zhou, Y., Shen, Y., Luo, G., Sun, X., Cao, L., Fu, R., Xu, Q., Ji, R.: Active teacher for semi-supervised object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14482-14491 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "28. Oliver, A., Odena, A., Raffel, C.A., Cubuk, E.D., Goodfellow, I.: Realistic evaluation of deep semi-supervised learning algorithms. Advances in neural information processing systems 31 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.639, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CamoTeacher" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "29. Pang, Y., Zhao, X., Xiang, T.Z., Zhang, L., Lu, H.: Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 2160-2170 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.229 + ], + "angle": 0, + "content": "30. Roelofs, B., Berthelot, D., Sohn, K., Carlini, N., Kurakin, A.: Adamatch: a unified approach to semi-supervised learning and domain adaptation. In: Proc. of the Int. Conf. on Learning Representations (ICLR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.229, + 0.788, + 0.271 + ], + "angle": 0, + "content": "31. Shin, G., Albanie, S., Xie, W.: Unsupervised salient object detection with spectral cluster voting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3971-3980 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.271, + 0.788, + 0.324 + ], + "angle": 0, + "content": "32. Simeoni, O., Sekkat, C., Puy, G., Vobecký, A., Zablocki, É., Pérez, P.: Unsupervised object localization: Observing the background to discover objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3176-3186 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.324, + 0.788, + 0.365 + ], + "angle": 0, + "content": "33. Skurowski, P., Abdulameer, H., Błaszczyk, J., Depta, T., Kornacki, A., Koziel, P.: Animal camouflage analysis: Chameleon database. Unpublished manuscript 2(6), 7 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.419 + ], + "angle": 0, + "content": "34. Sohn, K., Berthelot, D., Carlini, N., Zhang, Z., Zhang, H., Raffel, C.A., Cubuk, E.D., Kurakin, A., Li, C.L.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems 33, 596-608 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.42, + 0.788, + 0.461 + ], + "angle": 0, + "content": "35. Sohn, K., Zhang, Z., Li, C.L., Zhang, H., Lee, C.Y., Pfister, T.: A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.461, + 0.788, + 0.5 + ], + "angle": 0, + "content": "36. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.501, + 0.788, + 0.555 + ], + "angle": 0, + "content": "37. Wang, Y., Wang, H., Shen, Y., Fei, J., Li, W., Jin, G., Wu, L., Zhao, R., Le, X.: Semi-supervised semantic segmentation using unreliable pseudo-labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4248-4257 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.556, + 0.788, + 0.597 + ], + "angle": 0, + "content": "38. Wang, Z., Li, Y., Guo, Y., Fang, L., Wang, S.: Data-uncertainty guided multi-phase learning for semi-supervised object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4568-4577 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.597, + 0.788, + 0.637 + ], + "angle": 0, + "content": "39. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.637, + 0.788, + 0.678 + ], + "angle": 0, + "content": "40. Xu, M., Zhang, Z., Hu, H., Wang, J., Wang, L., Wei, F., Bai, X., Liu, Z.: End-to-end semi-supervised object detection with soft teacher. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3060-3069 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.678, + 0.788, + 0.718 + ], + "angle": 0, + "content": "41. Yang, L., Qi, L., Feng, L., Zhang, W., Shi, Y.: Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7236-7246 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.759 + ], + "angle": 0, + "content": "42. Yang, L., Zhuo, W., Qi, L., Shi, Y., Gao, Y.: St++: Make self-training work better for semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4268-4277 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.759, + 0.788, + 0.799 + ], + "angle": 0, + "content": "43. Yin, B., Zhang, X., Hou, Q., Sun, B.Y., Fan, D.P., Van Gool, L.: Camoformer: Masked separable attention for camouflaged object detection. arXiv preprint arXiv:2212.06570 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.84 + ], + "angle": 0, + "content": "44. Yu, S., Zhang, B., Xiao, J., Lim, E.G.: Structure-consistent weakly supervised salient object detection with local saliency coherence. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 3234-3242 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.127 + ], + "angle": 0, + "content": "X.Lai et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "45. Zhang, M., Xu, S., Piao, Y., Shi, D., Lin, S., Lu, H.: Preynet: Preying on camouflaged objects. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 5323-5332 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "46. Zhang, Y., Wu, C.: Unsupervised camouflaged object segmentation as domain adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4334-4344 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.286 + ], + "angle": 0, + "content": "47. Zhao, Z., Yang, L., Long, S., Pi, J., Zhou, L., Wang, J.: Augmentation matters: A simple-yet-effective approach to semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11350-11359 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.328 + ], + "angle": 0, + "content": "48. Zhao, Z., Zhou, L., Wang, L., Shi, Y., Gao, Y.: Lassl: Label-guided self-training for semi-supervised learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 9208-9216 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.356 + ], + "angle": 0, + "content": "49. Zhou, Z.H., Zhou, Z.H.: Semi-supervised learning. Machine Learning pp. 315-341 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.356 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/CamoTeacher_ Dual-Rotation Consistency Learning for Semi-Supervised Camouflaged Object Detection/e7b0e9c5-9645-43e1-9e04-1cc9804f36e6_origin.pdf b/2024/CamoTeacher_ Dual-Rotation Consistency Learning for Semi-Supervised Camouflaged Object Detection/e7b0e9c5-9645-43e1-9e04-1cc9804f36e6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0d1914652f786de1774e180e242e56689cc0539e --- /dev/null +++ b/2024/CamoTeacher_ Dual-Rotation Consistency Learning for Semi-Supervised Camouflaged Object Detection/e7b0e9c5-9645-43e1-9e04-1cc9804f36e6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05222c50ec5a40b5132bfd85f3ae0eb068ee3c1d8b0e6d2317ef1b6f887bd50a +size 3282126 diff --git a/2024/CamoTeacher_ Dual-Rotation Consistency Learning for Semi-Supervised Camouflaged Object Detection/full.md b/2024/CamoTeacher_ Dual-Rotation Consistency Learning for Semi-Supervised Camouflaged Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e1f8ff5b442f78b0cda078a1bf8c1093a0b16a5a --- /dev/null +++ b/2024/CamoTeacher_ Dual-Rotation Consistency Learning for Semi-Supervised Camouflaged Object Detection/full.md @@ -0,0 +1,358 @@ +# CamoTeacher: Dual-Rotation Consistency Learning for Semi-Supervised Camouflaged Object Detection + +Xunfa Lai $^{1}$ , Zhiyu Yang $^{1}$ , Jie Hu $^{2}$ , Shengchuan Zhang $^{1*}$ , Liujuan Cao $^{1}$ , Guannan Jiang $^{2}$ , Zhiyu Wang $^{2}$ , Songan Zhang $^{3}$ , and Rongrong Ji $^{1}$ + +1 Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University + +$^{2}$ Contemporary Amperex Technology Co. Limited + +3 Shanghai Jiao Tong University + +Abstract. Existing camouflaged object detection (COD) methods depend heavily on large-scale pixel-level annotations. However, acquiring such annotations is laborious due to the inherent camouflage characteristics of the objects. Semi-supervised learning offers a promising solution to this challenge. Yet, its application in COD is hindered by significant pseudo-label noise, both pixel-level and instance-level. We introduce CamoTeacher, a novel semi-supervised COD framework, utilizing Dual-Rotation Consistency Learning (DRCL) to effectively address these noise issues. Specifically, DRCL minimizes pseudo-label noise by leveraging rotation views' consistency in pixel-level and instance-level. First, it employs Pixel-wise Consistency Learning (PCL) to deal with pixel-level noise by reweighting the different parts within the pseudo-label. Second, Instance-wise Consistency Learning (ICL) is used to adjust weights for pseudo-labels, which handles instance-level noise. Extensive experiments on four COD benchmark datasets demonstrate that the proposed CamoTeacher not only achieves state-of-the-art compared with semi-supervised learning methods, but also rivals established fully-supervised learning methods. Our code will be available soon. + +Keywords: Camouflaged object detection $\cdot$ Semi-supervised learning + +# 1 Introduction + +Camouflaged object detection (COD) aims to identify objects that blend seamlessly into their environments including animals or artificial entities that possess protective coloring and have the ability to blend into their surroundings, a task complicated by low contrast, similar textures, and blurred boundaries [6,7,9,14, 17,24,43]. Unlike general object detection, COD is challenged by these factors, making detection significantly more difficult. Existing methods in COD heavily depend on extensive pixel-level annotated datasets, the creation of which incurs substantial human effort and cost, thus limiting COD's advancement. + +![](images/2750ee8e1c66b13332dc42b66a5003268eb0fd34beaf0c10bc638b21a505fc9e.jpg) +Fig. 1: (a) Illustration of pixel-level and instance-level noise. Pixel-level noise refers to varying degrees of noise in different parts within an instance. Instance-level noise pertains to varying degrees of noise among different instances. (b) Visualization of performance trends on COD10K [7] under different percentages of labeled data. + +To mitigate that, semi-supervised learning [13, 27, 36, 47, 49] emerges as a promising approach by leveraging both labeled and unlabeled data, yet its application in COD is not straightforward due to complex backgrounds and subtle object boundaries. The efficacy of semi-supervised learning in COD is significantly compromised by the presence of substantial noise in pseudo-labels. Our investigation into pseudo-label noise reveals two primary types: pixel-level noise, indicating variation within a single pseudo-label, and instance-level noise, showing variation across different pseudo-labels. This distinction is critical, as it informs our approach to refining pseudo-label quality for improved model training. (1) Pixel-level noise is characterized by inconsistent labeling within various parts of a pseudo-label. As demonstrated in Fig. 1(a), in the first row, the tail of the gecko is visually more camouflaged than its head. The pseudo-label generated by the SINet [7] is less accurate in its tail region (highlighted by the red box). This observation underscores the inappropriateness of uniformly treating all parts within a pseudo-label. (2) Instance-level noise refers to the variability in noise levels across distinct pseudo-labels. As shown in Fig. 1(a), the pseudo-labels of the third row are less accurate compared to that in the second row because the camouflaged objects in the third row is more difficult to detect. These discrepancies indicate that each pseudo-label contributes variably to model training, emphasizing the need for a nuanced approach to integrating pseudo-label information. + +To tackle the challenge of evaluating pseudo-label noise without ground truth (GT) for unlabeled data, we introduce two novel strategies based on pixel-wise inconsistency and instance-wise consistency across two rotation views. Specifically, for pixel-level noise, we observe that pixel-wise inconsistency—calculated by comparing pseudo-labels from two rotated views—mirrors the actual error relative to GT, as shown in Fig. 2(a). This relationship, illustrated by the polyline in Fig. 2(b), shows a positive correlation between mean pixel-wise inconsistency and the mean absolute error (MAE) across different parts. Consequently, areas + +![](images/63d370dfa8c3be84598bc01d5d5dbc4b1143cf95d613c6f58a3611fef14511db.jpg) + +![](images/fabd694cf371834a6a1826740a1d54d3837808358b24c8abf6cfb8e75bcc608c.jpg) +Image + +![](images/a5b2451f2405bab07b40156a647bba41cd7476d2e9886d789e54bd6e4bebee20.jpg) + +![](images/af04ce3cddbd74a8e254b7dc4d67dd4183c4d872cfa43eff974d5b8cc58c6a40.jpg) +GT +(a) + +![](images/c4784a760d8bc2b32816fcb1b8219c072aea690d6810ec730a0aef7d1b4b21b6.jpg) + +![](images/f683e8bc0d370e8f21acf430e32f8b21ef8b19169d28807a5c710bf5e524d592.jpg) +Real absolute error + +![](images/b77dc97c7ce1f08852d2144c5126918ec180ab0bbdc385da7144288d635e7e7f.jpg) + +![](images/8f379e3f424a96bb1a6c75427f066387208eee2bad31904820091fe32671416f.jpg) +Pixel-wise + +![](images/1ca49ad23aadb3d715b3a63ce54a1cece65637740934209f1e1cd137208dcedd.jpg) +(b) +Fig. 2: (a) The real absolute error is obtained by subtracting the pseudo-label from the ground truth, reflecting the actual noisy level. On the other hand, pixel-wise inconsistency is derived from the subtraction of pseudo-labels from two rotated views. Visually, pixel-wise inconsistency and real absolute error appear very similar. (b) We define the area within 20 pixels from the edge as the boundary according to GT. We calculate mean pixel-wise inconsistency (MPI) and mean absolute error (MAE) in background, foreground, and boundary. The pseudo-labels are obtained by our proposed model on the test dataset COD10K [7] at a $10\%$ semi-supervised setting. + +with higher pixel-wise inconsistency are more prone to inaccuracies, suggesting a need for their de-emphasis during training. For instance-level noise, we determine that pseudo-labels with greater similarity across rotation views exhibit lower noise levels, as indicated in Fig. 3(a). A positive correlation between instance-wise consistency and the SSIM [39] computed by pseudo-labels and GT further supports this observation, depicted in Fig. 3(b). Thus, pseudo-labels demonstrating higher instance-wise consistency are likely of superior quality and should be prioritized in the learning process. + +With the insights gained from these observations, we propose a semi-supervised camouflaged object detection framework called CamoTeacher, which incorporates a novel method known as Dual-Rotation Consistency Learning (DRCL). Specifically, DRCL operationalizes its strategy through two core components: Pixel-wise Consistency Learning (PCL) and Instance-wise Consistency Learning (ICL). PCL innovatively assigns variable weights to different parts within a pseudo-label, considering the pixel-wise inconsistency across diverse rotation view. Simultaneously, ICL adjusts the importance of individual pseudo-labels based on their instance-wise consistency, enabling a nuanced, noise-aware training process. + +We adopt SINet [7] as the base model to implement CamoTeacher and apply it to more classic COD models, i.e. CNN-based model SINet-v2 [6] and SegMaR [17] and Transformed-based model DTINet [24] and FSPNet [14]. Extensive experiments on four COD benchmark datasets, i.e. CAMO [21], CHAMELEON [33], COD10K [7], and NC4K [25], demonstrate that the proposed CamoTeacher not only achieves state-of-the-art compared with semi-supervised learning methods, but also rivals established fully-supervised learning methods. Specifically, as shown on Fig. 1(b), with only $20\%$ labeled data, it nearly reached the performance of fully supervised model on COD10K. + +![](images/eb5606f302aca6f3c031cf325c65857217b138eb00218453a2ee1e95046fde66.jpg) + +![](images/86512bfcad0262e90612a854a28984157791811722294cc37e5e97a91d44d01e.jpg) +Image + +![](images/16473e4f15ed9b7d3cd81140765cee16dc38133a1377a053e9ecad0455ea93c6.jpg) + +![](images/eb3e99a5a1d883bb3ba4cc364590f0ad6a2be5f3fbcd30de54af3e545bc7145d.jpg) +GT +(a) + +![](images/bfa004632f7a3d0fae0ac3376c0f01e253a3d31bc818072604a355dd65131740.jpg) + +![](images/cf497c818a3215c6e17ac635ee5c2eaa4375ee24090dfe2b81b84108771ad664.jpg) +Pseudo-label + +![](images/48b89a1690b82f7df298fd50507ebd46e4e6b6e11778e91bb9b236fd99916937.jpg) + +![](images/33ae46de3df3f755559a8a3a650bf1764c5d7efef8af9c4a73d8dc59c0b2782c.jpg) +Pseudo-label + +![](images/f6964705870498f4b41264c2a2ed0ef0c919bd519a923a42178a41c8f7acc88e.jpg) +(b) +Fig. 3: (a) The more similar the pseudo-labels are between two rotated views, the lower the noise level and the higher the quality of the pseudo-labels. For example, in the first row, the pseudo-labels are more similar, indicating a lower noise level compared to the pseudo-labels in the second row. (b) The positive correlation between instance-wise consistency and the SSIM [39] of pseudo-labels and GT. + +Our contributions can be summarized as follows: + +- We introduce the first end-to-end semi-supervised camouflaged object detection framework CamoTeacher, offering a simple yet effective baseline for future research in semi-supervised camouflaged object detection. +- To address the problem of large amounts of noise in pseudo-labels in semi-supervised camouflaged object detection, we proposed the Dual-Rotation Consistency Learning (DRCL), which includes Pixel-wise Consistency Learning (PCL) and Instance-wise Consistency Learning (ICL), allows for the adaptive adjustment of the contribution of pseudo-labels with varying quality, enabling the effective utilization of pseudo-label information. +- We conducted extensive experiments on COD benchmark datasets, and achieved significant improvements compared to the fully supervised setting. + +# 2 Related works + +# 2.1 Camouflaged Object Detection + +Camouflaged Object Detection(COD) is different from traditional visual detection tasks as objects and backgrounds in traditional detection tasks have significant visual distinctions, whereas camouflaged objects are typically concealed within their surrounding environments [6, 7, 14, 17, 24, 29, 43]. In recent years, many models have been proposed on this task, which can be divided into CNN-based and Transformed-based. i) CNN-based models: SINet [7] is a simple yet effective baseline that views camouflage detection as searching and identification process. SINet-v2 [6] significantly improves performance by introducing a neighbor connection decoder and cascaded group-reversal attention design. SegMaR [17] firstly performs initial detection of camouflaged objects and then progressively enlarges the detected regions to achieve precise detection. FEDER [9] achieves impressive results by conducting feature decomposition and + +edge reconstruction. ii) Transformed-based models: DTINet [24] proposes a dual-task interactive transformer to detect accurate position and detailed boundary. CamoFormer [43] presents masked separable attention to separate foreground and background regions. FSPNet [14] designs a feature shrinkage decoder hierarchically decode locality-enhanced neighboring transformer features. These outstanding works have achieved excellent results in COD, but they rely on large-scale pixel-level annotations, which are time-consuming and labor-intensive to obtain. + +In response to the reliance of full-supervised COD on expensive pixel-level annotations, some unsupervised/weakly-supervised approaches have emerged [10-12,46]. UCOS-DA [46] formulates the unsupervised camouflaged object segmentation as a source-free unsupervised domain adaptation task and designs a foreground-background-contrastive self-adversarial pipeline. SCOD [11] proposes the first weakly-supervised COD method with relabeled scribbles. WS-SAM [10] uses the provided sparse annotations as prompts for SAM [18] to generate segmentation masks to train the model. These methods reduce annotation costs but often exhibit lower performance compared to fully-supervised methods. Is it possible to leverage existing annotated data without additional labeling while ensuring accuracy? To address this issue, we propose a new benchmark that combines semi-supervised learning with camouflaged object detection. + +# 2.2 Semi-Supervised Learning + +In traditional supervised learning, training a high-performing model requires a substantial amount of labeled data. However, obtaining labeled data in practical applications is often costly and time-consuming. Consequently, the effective utilization of unlabeled data to improve model performance and generalization has become a critical concern. Semi-supervised learning addresses this challenge by leveraging unlabeled data to enhance the model's learning process [2,8,22,27,28,30,34,36,48], thereby improving its performance and generalization capability. Pseudo-labeling [3,35,37,38,40,42] and consistency regularization [13,19,20,23,41] are popular techniques in the field of semi-supervised learning. These methods have shown great efficacy when combined, producing impressive results. + +To address the issue of heavy reliance on labeled data in camouflage object detection, we propose the first end-to-end semi-supervised camouflage object detection method. By integrating semi-supervised learning, our approach aims to enhance the performance of camouflage object detection models while significantly reducing the need for annotated data. + +# 3 Methodology + +# 3.1 Task Formulation + +Semi-supervised camouflaged object detection aims to leverage a limited amount of annotated data to train a detector capable of identifying objects that seamlessly blend into their surroundings. This task is inherently challenging due to + +the low contrast between the objects and their background. Given a camouflaged object detection dataset $D$ for training, we represent the labeled subset with $M$ labeled samples as $D_{L} = \{x_{i}^{l},y_{i}\}_{i = 1}^{M}$ , and the unlabeled subset with $N$ unlabeled samples as $D_{U} = \{x_{i}^{u}\}_{i = 1}^{N}$ , where $x_{i}^{l}$ and $x_{i}^{u}$ represent the input image, $y_{i}$ represent the corresponding annotation mask of labeled data. Typically, $D_{L}$ constitutes a small fraction of the entire dataset $D$ , highlighting the semi-supervised learning scenario with $M \ll N$ . The emphasis on $M \ll N$ underscores the challenge and opportunity within semi-supervised learning: to enhance detection capabilities by utilizing the untapped potential of the unlabeled data $D_{U}$ , which far exceeds the labeled subset $D_{L}$ . + +# 3.2 Overall Framework + +As shown in Fig. 4, we adopt Mean Teacher [36] as preliminary to achieve end-to-end semi-supervised camouflaged object detection framework, which contains two COD models with the same structure, i.e., teacher model and student model, parameterized by $\Theta_t$ and $\Theta_s$ , respectively. The teacher model generates pseudolabels, which are then used to optimize the student model. The overall loss function $L$ can be defined as: + +$$ +L = L _ {s} + \lambda_ {u} L _ {u}, \tag {1} +$$ + +where $L_{s}$ and $L_{u}$ denote supervised loss and unsupervised loss respectively, $\lambda_{u}$ is the weight of unsupervised loss to balance the loss terms. Following the classic COD method, we use binary cross-entropy loss $L_{bce}$ for training. + +During the training process, we employ a combination of weak augmentation $\mathcal{A}^w (\cdot)$ and strong augmentation $\mathcal{A}^s (\cdot)$ strategies. Weak augmentation is applied to labeled data to mitigate overfitting, while unlabeled data undergo various data perturbations under strong augmentation to create diverse perspectives of the same image. The supervised loss $L_{s}$ is defined by: + +$$ +L _ {s} = \frac {1}{M} \sum_ {i = 1} ^ {M} L _ {b c e} \left(F \left(\mathcal {A} ^ {w} \left(x _ {i} ^ {l}\right); \Theta_ {s}\right), y _ {i}\right), \tag {2} +$$ + +where $F(\mathcal{A}(x_i);\Theta)$ represents the detection results obtained by the model $\Theta$ for the $i$ -th image with augmentations $\mathcal{A}(\cdot)$ . For unlabeled images, we first apply weak augmentation, denoted as $\mathcal{A}^w (\cdot)$ , before forwarding them to the teacher model. This initial step is crucial for generating reliable pseudo-labels $\widehat{y}_i$ under variations that do not significantly alter the core features of the images. These pseudo-labels serve as a form of soft supervision for the student model. Subsequently, the same images are subjected to strong augmentation, $\mathcal{A}^s (\cdot)$ , and then passed to the student model. This process introduces a higher level of variability and complexity, simulating more challenging conditions for the student model to adapt to. The student model generates predictions $p_i$ based on these strongly augmented images, leveraging the pseudo-labels $\widehat{y}_i$ for guidance in learning from the unlabeled data. It can be formulated as: + +$$ +\widehat {y _ {i}} = F (\mathcal {A} ^ {w} (x _ {i} ^ {u}); \Theta_ {t}), p _ {i} = F (\mathcal {A} ^ {s} (\mathcal {A} ^ {w} (x _ {i} ^ {u})); \Theta_ {s}). \tag {3} +$$ + +![](images/37f00157404283e719b043fde4c197bbba4387395910e6f53f3ac6fc3cc64896.jpg) +DRCL: Dual-Rotation Consistency Learning PCL: Pixel-wise Consistency Learning ICL: Instance-wise Consistency Learning + +![](images/bd3151ee0ffcb62fbe085cefbf294478d42de739c4e0a78207812d206e4d11e2.jpg) +Fig. 4: The overall pipeline of our CamoTeacher. CamoTeacher consists of a teacher model and a student model. The teacher model generates pseudo-labels to optimize the student model, while the teacher model is updated through EMA [36] from the student. To mitigate pseudo-labels' noise, we introduced Dual-Rotation Consistency Learning (DRCL), which involves Pixel-wise Consistency Learning (PCL) and Instance-wise Consistency Learning (ICL). DRCL weights the loss function based on pixel-wise inconsistency and instance-wise consistency in pseudo-labels across different rotation views. + +Hence, the unsupervised loss $L_{u}$ can be expressed as: + +$$ +L _ {u} = \frac {1}{N} \sum_ {i = 1} ^ {N} L _ {b c e} \left(p _ {i}, \widehat {y _ {i}}\right). \tag {4} +$$ + +Finally, the student model is intensively trained using the total loss $L$ , which encapsulates both the supervised and unsupervised learning aspects of the semi-supervised framework. This approach ensures that the student model benefits from both labeled and pseudo-labeled data, enhancing its detection capabilities. Concurrently, the teacher model is methodically updated through the Exponential Moving Average (EMA) mechanism [36], which effectively distills students' knowledge while preventing noise, formulated as: + +$$ +\Theta_ {t} \leftarrow \eta \Theta_ {t} + (1 - \eta) \Theta_ {s}, \tag {5} +$$ + +where $\eta$ is hyper-parameters denotes the keeping rate. + +# 3.3 Dual-Rotation Consistency Learning + +Due to the camouflaging nature of the objects, pseudo-labels contain a substantial amount of noise, and directly using them to optimize the student model potentially harm the model's performance. To address this problem, one of the most intuitive possible methods is to set a fixed high threshold to filter high-quality pseudo-labels, but this results in a low recall rate and makes it difficult to fully utilize the supervision information of pseudo-labels. To this end, we propose + +Dual-Rotation Consistency Learning (DRCL) to dynamically adjust the weight of the pseudo-labels to reduce the impact of noise. + +We randomly apply two separate rotations to the image $x_{i}$ , which has previously been flipped and random resized, resulting in two distinct rotation views $x_{i}^{r_{1}}$ and $x_{i}^{r_{2}}$ . + +$$ +x _ {i} ^ {r _ {1}} = R \left(\mathcal {A} ^ {w} \left(x _ {i}\right), \theta_ {1}\right), x _ {i} ^ {r _ {2}} = R \left(\mathcal {A} ^ {w} \left(x _ {i}\right), \theta_ {2}\right), \tag {6} +$$ + +where $x_{i}^{r} = R(x_{i},\theta)$ represents rotate the input image $x_{i}$ by $\theta$ degrees. We input the acquired rotation view into the teacher model to obtain the corresponding prediction, i.e. $\widehat{y}_i^r = F(x_i^r;\Theta_t)$ . Subsequently, we apply an opposite rotation of $-\theta$ to return the prediction to its original horizontal orientation, resulting in $\widehat{y}_i^{h_1}$ and $\widehat{y}_i^{h_2}$ , facilitating the calculation of prediction inconsistency under different rotation views, respectively, + +$$ +\hat {y} _ {i} ^ {h _ {1}} = R \left(\hat {y} _ {i} ^ {r _ {1}}, - \theta_ {1}\right), \hat {y} _ {i} ^ {h _ {2}} = R \left(\hat {y} _ {i} ^ {r _ {2}}, - \theta_ {2}\right). \tag {7} +$$ + +Note that rotation introduces black boundary areas, which are not involved in the implement of DRCL. + +Since the noise levels vary across different regions of pseudo-labels and among different pseudo-labels, we introduce PCL and ICL dynamically adapt the contributions of different pixels within pseudo-labels and across various pseudo-labels. + +Pixel-wise Consistency Learning We subtract the horizontal predictions $\widehat{y}_i^{h_1}$ and $\widehat{y}_i^{h_2}$ at the pixel level to obtain the pixel-wise inconsistency $\Delta_i$ , + +$$ +\Delta_ {i} = \left| \widehat {y} _ {i} ^ {h _ {1}} - \widehat {y} _ {i} ^ {h _ {2}} \right|. \tag {8} +$$ + +The pixel-wise inconsistency $\Delta_{i}$ across different views reflects the reliability of pseudo-labels. However, in cases where both rotation views' prediction are close to 0.5, $\Delta_{i}$ cannot effectively distinguish them. These predictions exhibit high uncertainty, meaning they cannot be definitively classified as foreground or background, and are likely to represent noisy labels. Consequently, it becomes necessary to diminish their influence by lowering their weights. Hence, we calculate the average of horizontal predictions $\widehat{y}_i^h$ , + +$$ +\widehat {y} _ {i} ^ {h} = \operatorname {a v g} \left(\widehat {y} _ {i} ^ {h _ {1}}, \widehat {y} _ {i} ^ {h _ {2}}\right), \tag {9} +$$ + +where $avg(\cdot, \cdot)$ represents calculate the average of two inputs in pixel-level, and the use its L2 distance from 0.5 as a component to adjust the weights. + +As a result, we derive the pixel-wise consistency weight $\omega_{i}^{pc}$ based on the pixel-wise inconsistency among different rotation views, as: + +$$ +\omega_ {i} ^ {p c} = \left(1 - \Delta_ {i} ^ {\alpha}\right) \left| \left| \hat {y} _ {i} ^ {h} - \mu \right| \right| _ {2} ^ {2}, \tag {10} +$$ + +where $\alpha$ is a hyper-parameter and $\mu = 0.5$ . This dynamic pixel-wise consistency weight $\omega_{i}^{pc}$ assigns greater weight to regions with consistent predictions across various rotation views, while regions with inconsistent predictions receive a smaller weight. + +In conclusion, we formulate the PCL loss $L_{u}^{PC}$ as: + +$$ +\begin{array}{l} L _ {u} ^ {P C} = \frac {1}{N} \sum_ {i = 1} ^ {N} \omega_ {i} ^ {p c} L _ {b c e} (p _ {i}, \widehat {y} _ {i} ^ {r _ {1}}) \\ = - \frac {1}{N H W} \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {H \times W} \omega_ {i, j} ^ {p c} \left[ \hat {y} _ {i, j} ^ {r _ {1}} \log p _ {i, j} \right. \tag {11} \\ + \left(1 - \widehat {y} _ {i, j} ^ {r _ {1}}\right) \log \left(1 - p _ {i, j}\right) ], \\ \end{array} +$$ + +adaptively adjusts the weight of each pixel to ensure comprehensive supervision of student model while avoiding confirmation bias. + +Instance-wise Consistency Learning The degree of camouflage varies between different images, leading to significant variations in the quality of pseudolabels across images. Treating all pseudo-labels equally would be unreasonable. Unfortunately, for unlabeled images, assessing the quality of pseudo-labels is challenging because there are no GT labels available. We have observed a positive correlation between the instance-wise consistency of two rotation views and the quality of pseudo-labels, as quantified by SSIM [39]. Base on this, we introduce ICL to adjust the contributions of pseudo-labels with varying quality. We represent the instance-wise consistency weight $\omega_{i}^{ic}$ as follows: + +$$ +\omega_ {i} ^ {i c} = \left(S S I M \left(\widehat {y} _ {i} ^ {h _ {1}}, \widehat {y} _ {i} ^ {h _ {2}}\right)\right) ^ {\beta}, \tag {12} +$$ + +where $\beta$ is a hyper-parameter used to adjust the distribution relationship between instance-wise consistency and pseudo-label quality. + +We use intersection-over-union (IoU) loss as the instance-wise restriction, therefore, the ICL loss can be expressed as: + +$$ +\begin{array}{l} L _ {u} ^ {I C} = \frac {1}{N} \sum_ {i = 1} ^ {N} \omega_ {i} ^ {i c} L _ {i o u} (p _ {i}, \widehat {y} _ {i} ^ {r _ {1}}) \\ = \frac {1}{N H W} \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {H \times W} \omega_ {i} ^ {i c} \left(1 - \frac {p _ {i , j} \widehat {y} _ {i , j} ^ {r _ {1}}}{p _ {i , j} + \widehat {y} _ {i , j} ^ {r _ {1}} - p _ {i , j} \widehat {y} _ {i , j} ^ {r _ {1}}}\right). \tag {13} \\ \end{array} +$$ + +Therefore, the final total $L$ loss comprises three components: supervised loss $L_{s}$ , PCL loss $L_{u}^{LC}$ , and ICL loss $L_{u}^{GC}$ , which can be expressed as: + +$$ +L = L _ {s} + \lambda_ {u} ^ {p c} L _ {u} ^ {P C} + \lambda_ {u} ^ {i c} L _ {u} ^ {I C}, \tag {14} +$$ + +where $\lambda_u^{pc},\lambda_u^{ic}$ are the hyper-parameters. + +# 4 Experiment + +# 4.1 Experiment Settings + +Dataset. We evaluate our CamoTeacher on four benchmark datasets, CAMO [21], CHAMELEON [33], COD10K [7], and NC4K [25]. In CAMO, there are + +Table 1: Experimental results on four COD benchmark datasets includes CAMO, CHAMELEON, COD10K, NC4K. $\dagger$ denotes using the same data augmentation as semi-supervised training. $\uparrow$ indicates the higher the score the better, and $\downarrow$ indicates the lower the better. "n%" means that "n%" of the data was utilized as labeled data, with the remainder as unlabeled data. As the data partitioning is random, we conducted 5 fold experiments and the results are the average of all 5 folds. + +
CHAMELEON (76)
Methods1% (41)5% (202)10% (404)
\( S_{\alpha} \uparrow \)\( F_{\beta}^{w} \uparrow \)\( E_{\phi}^{m} \uparrow \)\( E_{\phi}^{x} \uparrow \)\( F_{\beta}^{m} \uparrow \)\( M \downarrow \)\( S_{\alpha} \uparrow \)\( F_{\beta}^{w} \uparrow \)\( E_{\phi}^{m} \uparrow \)\( E_{\phi}^{x} \uparrow \)\( F_{\beta}^{m} \uparrow \)\( M \downarrow \)\( S_{\alpha} \uparrow \)\( F_{\beta}^{\omega } \uparrow \)\( E_{\phi}^{m} \uparrow \)\( E_{\phi}^{x} \uparrow \)\( F_{\beta}^{m} \uparrow \)\( M \downarrow \)
Supervised0.6050.2680.5820.7410.4350.2160.6130.2830.6010.7320.4390.1850.7140.4340.7250.8290.5920.110
Supervised†0.6590.3310.6550.7870.4990.1590.7280.4540.7440.8440.6070.106
Mean Teacher0.5370.1990.4180.6360.2290.2040.6110.3090.5240.7450.3530.1370.6790.4500.6500.8120.5120.102
CamoTeacher0.6520.4760.7140.7620.5580.0930.7290.5870.7850.8220.6560.0700.7560.6170.8130.8510.6840.065
+ +
CAMO (250)
Methods1% (41)5% (202)10% (404)
\( S_{\alpha} \uparrow \)\( F_{w}^{w} \uparrow \)\( E_{m}^{m} \uparrow \)\( E_{\sigma}^{e} \uparrow \)\( F_{g}^{m} \uparrow \)\( M \downarrow \)\( S_{\alpha} \uparrow \)\( F_{w}^{w} \uparrow \)\( E_{\sigma}^{m} \uparrow \)\( E_{\sigma}^{e} \uparrow \)\( F_{g}^{m} \uparrow \)\( M \downarrow \)\( S_{\alpha} \uparrow \)\( F_{w}^{w} \uparrow \)\( E_{m}^{m} \uparrow \)\( E_{\sigma}^{e} \uparrow \)\( F_{g}^{m} \uparrow \)\( M \downarrow \)
Supervised0.5730.2710.5410.7040.4020.2510.5920.2980.5810.7040.4320.2160.6520.3900.6470.7600.5250.159
Supervised †0.5960.2980.5800.7170.4420.2350.6100.3160.5950.7200.4500.2050.6670.4090.6660.7790.5460.153
Mean Teacher0.5180.2070.3990.6200.2270.2260.5750.2860.4820.7080.3220.1840.6250.3970.5780.7730.4540.150
CamoTeacher0.6210.4560.6690.7360.5450.1360.6690.5230.7110.7750.6010.1220.7010.5600.7420.7950.6350.112
+ +
COD10K (2026)
Methods1% (41)5% (202)10% (404)
So ↑Fw↑Em↑Eax↑Eax↑Fm↑M↓So ↑Fw↑Em↑Eax↑Eax↑Fm↑M↓So ↑Fw↑Em↑Eax↑Eax↑Fm↑M↓
Supervised0.5660.1890.5760.7230.3560.2090.5960.2200.6270.7470.4030.1640.6750.3360.7260.8100.5240.095
Supervised †0.5820.2060.6090.7430.3860.1940.6210.2460.6590.7740.4440.1470.6800.3380.7260.8120.5280.097
Mean Teacher0.5460.1680.4410.6330.2260.1610.6210.2720.5550.7320.3430.1070.6830.4040.6660.7990.4820.078
CamoTeacher0.6990.5170.7880.7970.5820.0620.7450.5830.8270.8400.6440.0500.7590.5940.8360.8540.6520.049
+ +
NC4K (4121)
Methods1% (41)5% (202)10% (404)
\( S_{\alpha} \uparrow \)\( E_{m}^{w} \uparrow \)\( E_{m}^{m} \uparrow \)\( E_{e}^{s} \uparrow \)\( F_{m}^{m} \uparrow \)\( M \downarrow \)\( S_{\alpha} \uparrow \)\( F_{m}^{w} \uparrow \)\( E_{m}^{m} \uparrow \)\( E_{e}^{s} \uparrow \)\( F_{m}^{m} \uparrow \)\( M \downarrow \)\( S_{\alpha} \uparrow \)\( F_{m}^{w} \uparrow \)\( E_{m}^{m} \uparrow } \)\( E_{e}^{s} \uparrow \)\( F_{m}^{m} \uparrow \)\( M \downarrow \)
Supervised0.6180.2910.6040.7520.4740.2130.6490.3320.6600.7730.5240.1790.7270.4620.7540.8330.6430.116
Supervised †0.6380.3130.6390.7680.5100.2010.6760.3620.6890.7980.5620.1630.7350.4710.7600.8370.6500.114
Mean Teacher0.5410.2130.4240.6370.2580.1930.6340.3550.5560.7670.4200.1400.7000.4920.6700.8270.5650.109
CamoTeacher0.7180.5990.7790.8140.6750.0900.7770.6770.8340.8590.7390.0710.7910.6870.8420.8680.7460.068
+ +2500 images, consisting of 1250 camouflage images and 1250 non-camouflage images. CHAMELEON contains 76 manually annotated images. COD10K dataset consists of 5066 camouflage images, 3000 background images, and 1934 non-camouflage images. NC4K is another large-scale COD test dataset that comprises 4,121 images. Based on the data partitioning in previous works [7,29,45], we used 3040 images from COD10K and 1000 images from CAMO as the training set for our experiments. The remaining images from both datasets were used as the test set. During the training process, we followed a semi-supervised segmentation [15,47] data partitioning approach. We sampled $1\%$ , $5\%$ , $10\%$ , $20\%$ , and $30\%$ of the training set as labeled data, while the remaining portion was used as unlabeled data. + +Evaluation Metrics. Referring to previous work [16, 17, 29], we used 6 common evaluation indicators in COD to evaluate our CamoTeacher, including S-measure $(S_{\alpha})$ [4], weight F-measure $(F_{\beta}^{w})$ [26], mean E-measure $(E_{\phi}^{m})$ [5], max E-measure $(E_{\phi}^{x})$ , mean F-measure $(F_{\beta}^{m})$ [1], and mean absolute error $(M)$ . + +Implementation Details. The proposed CamoTeacher is implemented with PyTorch. We adopt SINet [7] as the baseline of COD model. We use an SGD optimizer with a momentum of 0.9 and a polynomial learning-rate decay with an initial value of 0.01 to train the student model. The training epoch is set to 40 with 10 epochs for burn-in phase. And the batch size is 20, with a ratio of + +Table 2: Results of CamoTeacher on different COD models with $10\%$ labeled data on COD10K. + +
ModelSettingSα↑Fβw↑Eφm↑Eφx↑Fβm↑M↓
CNN-Based Models
SINet-v2 [6] (TPAMI2022)supervised0.7320.5440.8190.8350.5860.056
CamoTeacher0.7600.6050.8350.8430.6520.051
SegMaR [17] (CVPR2022)supervised0.7550.5830.8370.8510.6230.050
CamoTeacher0.7750.6290.8560.8640.6750.045
Transformer-Based Models
DTINet [24] (ICPR2022)supervised0.7710.6260.8570.8650.6700.043
CamoTeacher0.7940.6610.8730.8820.6950.040
FSPNet [14] (CVPR2023)supervised0.7890.6220.8230.8670.6750.042
CamoTeacher0.8170.6840.8840.9030.7200.037
+ +1:1 between labeled and unlabeled data, i.e. each batch consists of 10 labeled and 10 unlabeled images. During training and inference, each image is resized to $352 \times 352$ . The teacher model is updated via EMA with a momentum $\eta$ of 0.996. Weak data augmentation involves Random Flip and Random Scale, while strong data augmentation entails color space transformations including Identity, Autocontrast, Equalize, Gaussian blur, Contrast, Sharpness, Color, Brightness, Hue, Posterize, Solarize, with a random selection of up to 3 from this list. + +# 4.2 Main Results + +Comparison with Baselines. In Tab. 1, we compared CamoTeacher, the supervised baseline [7], and Mean Teacher [36]. Mean Teacher does not show improvements over the supervised in the semi-supervised setting due to treating all pseudo-labels and regions equally, which does not account for pseudo-labels noise problem. For instance, on the NC4K test set, under the $1\%$ labeled data, the $S_{\alpha}$ dropped by $9.7\%$ compared to supervised. Compared to Mean Teacher, CamoTeacher integrated with DRCL effectively mitigates pseudo-label noise issues and efficiently utilizes pseudo-label information. It demonstrates substantial improvements over both supervised and Mean Teacher across various proportions of labeled data, especially in scenarios with extremely limited. For instance, on $1\%$ NC4K, CamoTeacher exhibited a substantial increase of $28.6\%$ in the $F_{\beta}^{w}$ and a $12.3\%$ enhancement of $S_{\alpha}$ . Compared with Mean Teacher, the performance gains by CamoTeacher are more significant, which are $38.6\%$ and $17.7\%$ , respectively. + +In Fig. 1(b), we visualize the performance changes of supervised, Mean teacher and our CamoTeacher on COD10K under different labeled data ratios of $1\%$ , $5\%$ , $10\%$ , and further $20\%$ , $30\%$ , which intuitively reflects the effectiveness of CamoTeacher. It's evident that metrics like $F_{\beta}^{w}$ , $F_{\beta}^{m}$ , $E_{\phi}^{m}$ , and $M$ surpass $100\%$ of fully supervised performance with just $5\%$ labeled data. Additionally, with $30\%$ labeled data, all metrics significantly exceed $100\%$ of fully supervised performance. + +Table 3: Result of CamoTeacher on COD10K with additional data. We use $100\%$ train dataset as labeled data and test dataset as the unlabeled data. "U" represents the unsupervised setting, "W" denotes the weakly-supervised setting, where weakly-supervision results from scribble annotations. "L100%" indicates using $100\%$ of the training set as labeled data, "U(CAMO)" denotes utilizing CAMO as unlabeled data, and likewise for others. + +
ModelSettingSα↑Fwβ↑Emφ↑Exφ↑Fmβ↑M↓
Unsupervised
SelfMask [31] (CVPRW2022)U0.645 0.440 0.687 0.728 0.478 0.125
FOUND [32] (CVPR2023)U0.670 0.482 0.751 0.753 0.520 0.085
UCOS-DA [46] (ICCVW2023)U0.689 0.513 0.740 0.741 0.546 0.086
Weakly-Supervised
SCWSSOD [44] (AAAI2021)W0.710 0.546 0.805 - - 0.055
SCOD [11](AAAI2023)W0.733 0.576 0.832 0.845 0.633 0.049
WS-SAM [10] (NeurIPS2023)W0.803 - 0.878 - - 0.038
Semi-Supervised
CamoTeacher SINetL100%0.771 0.551 0.806 0.868 0.634 0.051
L100%+U(CAMO)0.779 0.554 0.799 0.875 0.632 0.051
L100%+U(NC4K)0.785 0.574 0.812 0.879 0.642 0.049
L100%+U(CAMO,NC4K)0.787 0.576 0.813 0.881 0.644 0.048
+ +Application to Existing COD Methods. We applied the CamoTeacher approach to additional COD models, including CNN-based model SINet-v2 [6] and SegMaR [17] and Transformed-based model DTINet [24] and FSPNet [14], and the corresponding results are presented in Tab. 2. In terms of SINet-v2, CamoTeacher demonstrated a $2.8\%$ improvement in the $S_{\alpha}$ metric and a $6.1\%$ improvement in the $F_{\beta}^{w}$ metric compared to the supervised baseline. Based on FSPNet, CamoTeacher achieved a $2.8\%$ enhancement in the $S_{\alpha}$ metric and a $6.2\%$ improvement in the $F_{\beta}^{w}$ metric without utilizing additional discriminative annotation. These outcomes provide robust evidence of the effectiveness of CamoTeacher. + +Addition unlabeled data. In Tab. 3, based on the base model SINet, we train CamoTeacher with $100\%$ labeled data and check whether additional unlabeled data can further improve the performance. Compared with unsupervised and weakly-supervised methods, it shows the superiority of our method. We use the testset CAMO and NC4K as unlabeled data, and we find that adding more additional data CamoTeacher can make SINet continued to gain, such as $S_{\alpha}$ metric increased by $1.6\%$ and $F_{\beta}^{w}$ metric increased by $2.5\%$ . We achieved significantly better performance than unsupervised methods [31,32,46] using existing labeled data. Compared with the weakly-supervision methods [10,11,44], we do not need additional annotation for the new unlabeled data, and to a certain extent, as the more unlabeled data, the performance will get better and better. And at the same time, replacing the base model with a stronger model will get better results. + +Table 4: The ablation study of Pixel-wise Table 5: The ablation study of various Consistency Learning (PCL) and Instance- weight approach of Pixel-wise Consistency wise Consistency Learning (ICL) with $5\%$ Learning (PCL) with $5\%$ labeled data on labeled data on COD10K. COD10K. In this experiment $\alpha$ is set to 1/4. + +
PCLICL\( {S}_{\alpha } \uparrow \)\( {F}_{\beta }^{w} \uparrow \)\( {E}_{\phi }^{m} \uparrow \)\( {E}_{\phi }^{x} \uparrow \)\( {F}_{\beta }^{m} \uparrow \)\( M \downarrow \)
\( L_{bce} \)\( {\omega }^{pc} \)\( L_{iou} \)\( {\omega }^{tc} \)
✓0.6820.3700.6360.7870.4610.091
✓0.7340.5720.8170.8210.6260.057
✓✓0.7240.5010.7380.8320.5740.058
✓✓0.7420.5850.8280.8340.6390.052
✓✓✓0.7450.5860.8300.8400.6430.051
✓✓✓✓0.7460.5860.8300.8430.6450.050
+ +
weight approach\( {S}_{\alpha } \uparrow {F}_{\beta }^{w} \uparrow {E}_{\phi }^{m} \uparrow {E}_{\phi }^{x} \uparrow {F}_{\beta }^{m} \uparrow M \downarrow \)
\( {\omega }^{pc} = {\widehat{y}}^{r} \)0.6980.3930.6840.8040.5000.110
\( {\omega }^{pc} = {\begin{Vmatrix}{\widehat{y}}^{r} - \mu \end{Vmatrix}}_{2}^{2} \)0.7170.4850.7340.8220.5590.061
\( {\omega }^{pc} = 1 - {\Delta }^{\alpha } \)0.7190.4810.7200.8270.5530.062
\( {\omega }^{pc} = \left( {1 - {\Delta }^{\alpha }}\right) {\widehat{y}}^{r} \)0.6950.4050.6960.7950.5060.114
\( {\omega }^{pc} = \left( {1 - {\Delta }^{\alpha }}\right) {\begin{Vmatrix}{\widehat{y}}^{r} - \mu \end{Vmatrix}}_{2}^{2}\left\lbrack \begin{array}{llllllll} 0 & {724} & {0.501} & {0.738} & {0.832} & {0.574} & {0.058} \end{array}\right\rbrack \)
+ +Table 6: The ablation study of hyper- Table 7: The ablation study of hyperparameters $\alpha$ in Pixel-wise Consistency parameters $\beta$ in Instance-wise Consistency Learning (PCL) with $5\%$ labeled data on Learning (ICL) with $5\%$ labeled data on COD10K. In this experiment $\beta$ is set to 4, COD10K. In this experiment $\alpha$ is set to $\lambda_u^{PC}$ is set to 8, and $\lambda_u^{IC}$ is set to 0.3. $1/4, \lambda_u^{PC}$ is set to 8, and $\lambda_u^{IC}$ is set to 0.3. + +
αSα↑Fwβ↑Eφm↑Eφx↑Fmβ↑M↓
20.7370.5650.8130.8360.6290.052
10.7390.5660.8150.8320.6270.056
1/20.7450.5750.8200.8390.6370.051
1/40.7470.5790.8250.8420.6380.051
1/80.7450.5780.8250.8380.6360.052
+ +
βSα↑Fβw↑Eφm↑Eφx↑Fβm↑M↓
10.7430.5760.8240.8390.6360.051
20.7420.5730.8230.8370.6320.052
40.7470.5790.8250.8420.6380.051
50.7440.5780.8240.8410.6390.051
60.7450.5750.8220.8390.6350.052
+ +# 4.3 Ablation Study + +Ablation on PCL and ICL. We conducted ablation experiments on different combinations of weight and loss functions for PCL and ICL. From Tab. 4, increasing the weights $\omega^{pc}$ and $\omega^{ic}$ improves the model's performance. Specifically, with $\omega^{pc}$ results in a higher $F_{\beta}^{w}$ score, increasing from 0.370 to 0.501. Similarly, with $\omega^{ic}$ improved the $F_{\beta}^{w}$ metric from 0.572 to 0.585. Combining these two weighted loss functions led to a significant performance improvement, with the $F_{\beta}^{w}$ score increasing from 0.370 to 0.586 (an increase of $21.6\%$ ). This demonstrates the effectiveness of PCL and ICL in enhancing the model's performance. + +Ablation of weight approach in PCL. The choice of method for obtaining $\omega^{pc}$ significantly affects the performance of the model. Therefore, we conducted several experiments to select the most effective method for obtaining $\omega^{pc}$ . From Tab. 5, it can be observed that the best performance is achieved when $\omega^{pc}$ is obtained using Eq. (10). + +Ablation of hyper-parameters in PCL and ICL. We perform ablation experiments on the hyper-parameters associated with the PCL and ICL modules. As shown in Tab. 6, the optimal results are achieved when setting the hyper-parameter $\alpha$ , which corresponds to the coefficient $\omega^{pc}$ , to $1/4$ . Similarly, according to Tab. 7, the best performance is observed when the hyper-parameter $\beta$ , related to the coefficient $\omega^{ic}$ , is set to 4. + +![](images/3642b2a7ae66b712dc7766f375dc5b354c32de919ada0deaf38816c562fb24bc.jpg) +Fig. 5: Visualization of predictions in different proportions of labeled data. + +# 4.4 Qualitative Analysis + +Result visualization. We conducte visualizations of CamoTeacher's results under varying proportions of labeled data. The model is trained with $1\%$ , $5\%$ , $10\%$ , $20\%$ and $30\%$ labeled data, and we compare its predicted masks with those produced by SINet under fully supervised conditions (as shown in Fig. 5). Notably, even with only $10\%$ labeled data, CamoTeacher's predicted masks exhibit a remarkable similarity to the fully supervised SINet predictions. In fact, in certain detailed predictions, CamoTeacher even surpasses the performance of fully supervised SINet. These findings highlight the impressive capability of CamoTeacher in leveraging limited labeled data for effective camouflaged object detection. + +# 5 Conclusions + +In this paper, we present the first end-to-end semi-supervised camouflaged object detection model called CamoTeacher. To address the issue of substantial noise within pseudo-labels in semi-supervised COD, including both local noise and global noise, we've introduced a novel approach called Dual-Rotation Consistency Learning (DRCL), including Pixel-wise Consistency Learning (PCL) and Instance-wise Consistency Learning (ICL). DRCL assists the model in alleviating noise issues, effectively leveraging pseudo-labels information, so that the model can receive adequate supervision while avoiding confirmation bias. Extensive experiments validate CamoTeacher's superior performance while significantly reducing annotation expenses. We believe that CamoTeacher can serve as a strong baseline for future research on semi-supervised COD. + +# Acknowledgements + +This work was supported by National Science and Technology Major Project (No. 2022ZD0118201), the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China a (No. U21B2037, No. U22B2051, No. U23A20383, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No.2022J06001). + +# References + +1. Achanta, R., Hemami, S., Estrada, F., Susstrunk, S.: Frequency-tuned salient region detection. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 1597-1604. IEEE (2009) +2. Chen, H., Tao, R., Fan, Y., Wang, Y., Wang, J., Schiele, B., Xie, X., Raj, B., Savvides, M.: Softmatch: Addressing the quantity-quality trade-off in semi-supervised learning. arXiv preprint arXiv:2301.10921 (2023) +3. Chen, X., Yuan, Y., Zeng, G., Wang, J.: Semi-supervised semantic segmentation with cross pseudo supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2613-2622 (2021) +4. Fan, D.P., Cheng, M.M., Liu, Y., Li, T., Borji, A.: Structure-measure: A new way to evaluate foreground maps. In: Proceedings of the IEEE international conference on computer vision. pp. 4548-4557 (2017) +5. Fan, D.P., Gong, C., Cao, Y., Ren, B., Cheng, M.M., Borji, A.: Enhanced-alignment measure for binary foreground map evaluation. arXiv preprint arXiv:1805.10421 (2018) +6. Fan, D.P., Ji, G.P., Cheng, M.M., Shao, L.: Concealed object detection. IEEE transactions on pattern analysis and machine intelligence 44(10), 6024-6042 (2021) +7. Fan, D.P., Ji, G.P., Sun, G., Cheng, M.M., Shen, J., Shao, L.: Camouflaged object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2777-2787 (2020) +8. Grandvalet, Y., Bengio, Y.: Semi-supervised learning by entropy minimization. Advances in neural information processing systems 17 (2004) +9. He, C., Li, K., Zhang, Y., Tang, L., Zhang, Y., Guo, Z., Li, X.: Camouflaged object detection with feature decomposition and edge reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22046-22055 (2023) +10. He, C., Li, K., Zhang, Y., Xu, G., Tang, L., Zhang, Y., Guo, Z., Li, X.: Weakly-supervised concealed object segmentation with sam-based pseudo labeling and multi-scale feature grouping. Advances in Neural Information Processing Systems 36 (2024) +11. He, R., Dong, Q., Lin, J., Lau, R.W.: Weakly-supervised camouflaged object detection with scribble annotations. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 781-789 (2023) +12. Hu, J., Lin, J., Cai, W., Gong, S.: Relax image-specific prompt requirement in sam: A single generic prompt for segmenting camouflaged objects. arXiv preprint arXiv:2312.07374 (2023) + +13. Hu, J., Chen, C., Cao, L., Zhang, S., Shu, A., Jiang, G., Ji, R.: Pseudo-label alignment for semi-supervised instance segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16337-16347 (2023) +14. Huang, Z., Dai, H., Xiang, T.Z., Wang, S., Chen, H.X., Qin, J., Xiong, H.: Feature shrinkage pyramid for camouflaged object detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5557-5566 (2023) +15. Hung, W.C., Tsai, Y.H., Liou, Y.T., Lin, Y.Y., Yang, M.H.: Adversarial learning for semi-supervised semantic segmentation. arXiv preprint arXiv:1802.07934 (2018) +16. Ji, G.P., Fan, D.P., Chou, Y.C., Dai, D., Liniger, A., Van Gool, L.: Deep gradient learning for efficient camouflaged object detection. Machine Intelligence Research 20(1), 92-108 (2023) +17. Jia, Q., Yao, S., Liu, Y., Fan, X., Liu, R., Luo, Z.: Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4713-4722 (2022) +18. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) +19. Lai, X., Tian, Z., Jiang, L., Liu, S., Zhao, H., Wang, L., Jia, J.: Semi-supervised semantic segmentation with directional context-aware consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1205-1214 (2021) +20. Laine, S., Aila, T.: Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242 (2016) +21. Le, T.N., Nguyen, T.V., Nie, Z., Tran, M.T., Sugimoto, A.: Anabranch network for camouflaged object segmentation. Computer vision and image understanding 184, 45-56 (2019) +22. Lee, D.H., et al.: Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In: Workshop on challenges in representation learning, ICML. vol. 3, p. 896. Atlanta (2013) +23. Liu, Y., Tian, Y., Chen, Y., Liu, F., Belagiannis, V., Carneiro, G.: Perturbed and strict mean teachers for semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4258-4267 (2022) +24. Liu, Z., Zhang, Z., Tan, Y., Wu, W.: Boosting camouflaged object detection with dual-task interactive transformer. In: 2022 26th International Conference on Pattern Recognition (ICPR). pp. 140-146. IEEE (2022) +25. Lv, Y., Zhang, J., Dai, Y., Li, A., Liu, B., Barnes, N., Fan, D.P.: Simultaneously localize, segment and rank the camouflaged objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11591-11601 (2021) +26. Margolin, R., Zelnik-Manor, L., Tal, A.: How to evaluate foreground maps? In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 248-255 (2014) +27. Mi, P., Lin, J., Zhou, Y., Shen, Y., Luo, G., Sun, X., Cao, L., Fu, R., Xu, Q., Ji, R.: Active teacher for semi-supervised object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14482-14491 (2022) +28. Oliver, A., Odena, A., Raffel, C.A., Cubuk, E.D., Goodfellow, I.: Realistic evaluation of deep semi-supervised learning algorithms. Advances in neural information processing systems 31 (2018) + +29. Pang, Y., Zhao, X., Xiang, T.Z., Zhang, L., Lu, H.: Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 2160-2170 (2022) +30. Roelofs, B., Berthelot, D., Sohn, K., Carlini, N., Kurakin, A.: Adamatch: a unified approach to semi-supervised learning and domain adaptation. In: Proc. of the Int. Conf. on Learning Representations (ICLR) (2022) +31. Shin, G., Albanie, S., Xie, W.: Unsupervised salient object detection with spectral cluster voting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3971-3980 (2022) +32. Simeoni, O., Sekkat, C., Puy, G., Vobecký, A., Zablocki, É., Pérez, P.: Unsupervised object localization: Observing the background to discover objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3176-3186 (2023) +33. Skurowski, P., Abdulameer, H., Błaszczyk, J., Depta, T., Kornacki, A., Koziel, P.: Animal camouflage analysis: Chameleon database. Unpublished manuscript 2(6), 7 (2018) +34. Sohn, K., Berthelot, D., Carlini, N., Zhang, Z., Zhang, H., Raffel, C.A., Cubuk, E.D., Kurakin, A., Li, C.L.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems 33, 596-608 (2020) +35. Sohn, K., Zhang, Z., Li, C.L., Zhang, H., Lee, C.Y., Pfister, T.: A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757 (2020) +36. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems 30 (2017) +37. Wang, Y., Wang, H., Shen, Y., Fei, J., Li, W., Jin, G., Wu, L., Zhao, R., Le, X.: Semi-supervised semantic segmentation using unreliable pseudo-labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4248-4257 (2022) +38. Wang, Z., Li, Y., Guo, Y., Fang, L., Wang, S.: Data-uncertainty guided multi-phase learning for semi-supervised object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4568-4577 (2021) +39. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004) +40. Xu, M., Zhang, Z., Hu, H., Wang, J., Wang, L., Wei, F., Bai, X., Liu, Z.: End-to-end semi-supervised object detection with soft teacher. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3060-3069 (2021) +41. Yang, L., Qi, L., Feng, L., Zhang, W., Shi, Y.: Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7236-7246 (2023) +42. Yang, L., Zhuo, W., Qi, L., Shi, Y., Gao, Y.: St++: Make self-training work better for semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4268-4277 (2022) +43. Yin, B., Zhang, X., Hou, Q., Sun, B.Y., Fan, D.P., Van Gool, L.: Camoformer: Masked separable attention for camouflaged object detection. arXiv preprint arXiv:2212.06570 (2022) +44. Yu, S., Zhang, B., Xiao, J., Lim, E.G.: Structure-consistent weakly supervised salient object detection with local saliency coherence. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 3234-3242 (2021) + +45. Zhang, M., Xu, S., Piao, Y., Shi, D., Lin, S., Lu, H.: Preynet: Preying on camouflaged objects. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 5323-5332 (2022) +46. Zhang, Y., Wu, C.: Unsupervised camouflaged object segmentation as domain adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4334-4344 (2023) +47. Zhao, Z., Yang, L., Long, S., Pi, J., Zhou, L., Wang, J.: Augmentation matters: A simple-yet-effective approach to semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11350-11359 (2023) +48. Zhao, Z., Zhou, L., Wang, L., Shi, Y., Gao, Y.: Lassl: Label-guided self-training for semi-supervised learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 9208-9216 (2022) +49. Zhou, Z.H., Zhou, Z.H.: Semi-supervised learning. Machine Learning pp. 315-341 (2021) \ No newline at end of file diff --git a/2024/CamoTeacher_ Dual-Rotation Consistency Learning for Semi-Supervised Camouflaged Object Detection/images.zip b/2024/CamoTeacher_ Dual-Rotation Consistency Learning for Semi-Supervised Camouflaged Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a0b76726d8317eca57c4b277626dc36d30f4cd22 --- /dev/null +++ b/2024/CamoTeacher_ Dual-Rotation Consistency Learning for Semi-Supervised Camouflaged Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b4094be64d1d998293270d79eb62d8d34622c7c3fdec8917063edf08083eb687 +size 689393 diff --git a/2024/CamoTeacher_ Dual-Rotation Consistency Learning for Semi-Supervised Camouflaged Object Detection/layout.json b/2024/CamoTeacher_ Dual-Rotation Consistency Learning for Semi-Supervised Camouflaged Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8c11337af50d5466415f6f47b2edd321018c6b52 --- /dev/null +++ b/2024/CamoTeacher_ Dual-Rotation Consistency Learning for Semi-Supervised Camouflaged Object Detection/layout.json @@ -0,0 +1,11711 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 157, + 111, + 459, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 111, + 459, + 166 + ], + "spans": [ + { + "bbox": [ + 157, + 111, + 459, + 166 + ], + "type": "text", + "content": "CamoTeacher: Dual-Rotation Consistency Learning for Semi-Supervised Camouflaged Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 148, + 185, + 465, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 185, + 465, + 212 + ], + "spans": [ + { + "bbox": [ + 148, + 185, + 465, + 212 + ], + "type": "text", + "content": "Xunfa Lai" + }, + { + "bbox": [ + 148, + 185, + 465, + 212 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 185, + 465, + 212 + ], + "type": "text", + "content": ", Zhiyu Yang" + }, + { + "bbox": [ + 148, + 185, + 465, + 212 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 185, + 465, + 212 + ], + "type": "text", + "content": ", Jie Hu" + }, + { + "bbox": [ + 148, + 185, + 465, + 212 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 148, + 185, + 465, + 212 + ], + "type": "text", + "content": ", Shengchuan Zhang" + }, + { + "bbox": [ + 148, + 185, + 465, + 212 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 148, + 185, + 465, + 212 + ], + "type": "text", + "content": ", Liujuan Cao" + }, + { + "bbox": [ + 148, + 185, + 465, + 212 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 185, + 465, + 212 + ], + "type": "text", + "content": ", Guannan Jiang" + }, + { + "bbox": [ + 148, + 185, + 465, + 212 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 148, + 185, + 465, + 212 + ], + "type": "text", + "content": ", Zhiyu Wang" + }, + { + "bbox": [ + 148, + 185, + 465, + 212 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 148, + 185, + 465, + 212 + ], + "type": "text", + "content": ", Songan Zhang" + }, + { + "bbox": [ + 148, + 185, + 465, + 212 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 148, + 185, + 465, + 212 + ], + "type": "text", + "content": ", and Rongrong Ji" + }, + { + "bbox": [ + 148, + 185, + 465, + 212 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 146, + 219, + 466, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 219, + 466, + 242 + ], + "spans": [ + { + "bbox": [ + 146, + 219, + 466, + 242 + ], + "type": "text", + "content": "1 Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 203, + 242, + 411, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 242, + 411, + 252 + ], + "spans": [ + { + "bbox": [ + 203, + 242, + 411, + 252 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 203, + 242, + 411, + 252 + ], + "type": "text", + "content": " Contemporary Amperex Technology Co. Limited" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 240, + 253, + 374, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 253, + 374, + 264 + ], + "spans": [ + { + "bbox": [ + 240, + 253, + 374, + 264 + ], + "type": "text", + "content": "3 Shanghai Jiao Tong University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 289, + 455, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 289, + 455, + 488 + ], + "spans": [ + { + "bbox": [ + 159, + 289, + 455, + 488 + ], + "type": "text", + "content": "Abstract. Existing camouflaged object detection (COD) methods depend heavily on large-scale pixel-level annotations. However, acquiring such annotations is laborious due to the inherent camouflage characteristics of the objects. Semi-supervised learning offers a promising solution to this challenge. Yet, its application in COD is hindered by significant pseudo-label noise, both pixel-level and instance-level. We introduce CamoTeacher, a novel semi-supervised COD framework, utilizing Dual-Rotation Consistency Learning (DRCL) to effectively address these noise issues. Specifically, DRCL minimizes pseudo-label noise by leveraging rotation views' consistency in pixel-level and instance-level. First, it employs Pixel-wise Consistency Learning (PCL) to deal with pixel-level noise by reweighting the different parts within the pseudo-label. Second, Instance-wise Consistency Learning (ICL) is used to adjust weights for pseudo-labels, which handles instance-level noise. Extensive experiments on four COD benchmark datasets demonstrate that the proposed CamoTeacher not only achieves state-of-the-art compared with semi-supervised learning methods, but also rivals established fully-supervised learning methods. Our code will be available soon." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 498, + 447, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 498, + 447, + 510 + ], + "spans": [ + { + "bbox": [ + 160, + 498, + 447, + 510 + ], + "type": "text", + "content": "Keywords: Camouflaged object detection " + }, + { + "bbox": [ + 160, + 498, + 447, + 510 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 498, + 447, + 510 + ], + "type": "text", + "content": " Semi-supervised learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 528, + 230, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 528, + 230, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 528, + 230, + 540 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 552, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 482, + 648 + ], + "type": "text", + "content": "Camouflaged object detection (COD) aims to identify objects that blend seamlessly into their environments including animals or artificial entities that possess protective coloring and have the ability to blend into their surroundings, a task complicated by low contrast, similar textures, and blurred boundaries [6,7,9,14, 17,24,43]. Unlike general object detection, COD is challenged by these factors, making detection significantly more difficult. Existing methods in COD heavily depend on extensive pixel-level annotated datasets, the creation of which incurs substantial human effort and cost, thus limiting COD's advancement." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 237, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 237, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 237, + 665 + ], + "type": "text", + "content": "* Corresponding authors" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 462, + 235 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 462, + 235 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 462, + 235 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 462, + 235 + ], + "type": "image", + "image_path": "2750ee8e1c66b13332dc42b66a5003268eb0fd34beaf0c10bc638b21a505fc9e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 243, + 482, + 288 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 288 + ], + "type": "text", + "content": "Fig. 1: (a) Illustration of pixel-level and instance-level noise. Pixel-level noise refers to varying degrees of noise in different parts within an instance. Instance-level noise pertains to varying degrees of noise among different instances. (b) Visualization of performance trends on COD10K [7] under different percentages of labeled data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 316, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 567 + ], + "type": "text", + "content": "To mitigate that, semi-supervised learning [13, 27, 36, 47, 49] emerges as a promising approach by leveraging both labeled and unlabeled data, yet its application in COD is not straightforward due to complex backgrounds and subtle object boundaries. The efficacy of semi-supervised learning in COD is significantly compromised by the presence of substantial noise in pseudo-labels. Our investigation into pseudo-label noise reveals two primary types: pixel-level noise, indicating variation within a single pseudo-label, and instance-level noise, showing variation across different pseudo-labels. This distinction is critical, as it informs our approach to refining pseudo-label quality for improved model training. (1) Pixel-level noise is characterized by inconsistent labeling within various parts of a pseudo-label. As demonstrated in Fig. 1(a), in the first row, the tail of the gecko is visually more camouflaged than its head. The pseudo-label generated by the SINet [7] is less accurate in its tail region (highlighted by the red box). This observation underscores the inappropriateness of uniformly treating all parts within a pseudo-label. (2) Instance-level noise refers to the variability in noise levels across distinct pseudo-labels. As shown in Fig. 1(a), the pseudo-labels of the third row are less accurate compared to that in the second row because the camouflaged objects in the third row is more difficult to detect. These discrepancies indicate that each pseudo-label contributes variably to model training, emphasizing the need for a nuanced approach to integrating pseudo-label information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "To tackle the challenge of evaluating pseudo-label noise without ground truth (GT) for unlabeled data, we introduce two novel strategies based on pixel-wise inconsistency and instance-wise consistency across two rotation views. Specifically, for pixel-level noise, we observe that pixel-wise inconsistency—calculated by comparing pseudo-labels from two rotated views—mirrors the actual error relative to GT, as shown in Fig. 2(a). This relationship, illustrated by the polyline in Fig. 2(b), shows a positive correlation between mean pixel-wise inconsistency and the mean absolute error (MAE) across different parts. Consequently, areas" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "X.Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 117, + 197, + 155 + ], + "blocks": [ + { + "bbox": [ + 149, + 117, + 197, + 155 + ], + "lines": [ + { + "bbox": [ + 149, + 117, + 197, + 155 + ], + "spans": [ + { + "bbox": [ + 149, + 117, + 197, + 155 + ], + "type": "image", + "image_path": "63d370dfa8c3be84598bc01d5d5dbc4b1143cf95d613c6f58a3611fef14511db.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 150, + 155, + 197, + 191 + ], + "blocks": [ + { + "bbox": [ + 150, + 155, + 197, + 191 + ], + "lines": [ + { + "bbox": [ + 150, + 155, + 197, + 191 + ], + "spans": [ + { + "bbox": [ + 150, + 155, + 197, + 191 + ], + "type": "image", + "image_path": "fabd694cf371834a6a1826740a1d54d3837808358b24c8abf6cfb8e75bcc608c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 192, + 182, + 200 + ], + "lines": [ + { + "bbox": [ + 165, + 192, + 182, + 200 + ], + "spans": [ + { + "bbox": [ + 165, + 192, + 182, + 200 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 197, + 118, + 244, + 155 + ], + "blocks": [ + { + "bbox": [ + 197, + 118, + 244, + 155 + ], + "lines": [ + { + "bbox": [ + 197, + 118, + 244, + 155 + ], + "spans": [ + { + "bbox": [ + 197, + 118, + 244, + 155 + ], + "type": "image", + "image_path": "a5b2451f2405bab07b40156a647bba41cd7476d2e9886d789e54bd6e4bebee20.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 197, + 155, + 244, + 191 + ], + "blocks": [ + { + "bbox": [ + 197, + 155, + 244, + 191 + ], + "lines": [ + { + "bbox": [ + 197, + 155, + 244, + 191 + ], + "spans": [ + { + "bbox": [ + 197, + 155, + 244, + 191 + ], + "type": "image", + "image_path": "af04ce3cddbd74a8e254b7dc4d67dd4183c4d872cfa43eff974d5b8cc58c6a40.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 192, + 227, + 199 + ], + "lines": [ + { + "bbox": [ + 215, + 192, + 227, + 199 + ], + "spans": [ + { + "bbox": [ + 215, + 192, + 227, + 199 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 239, + 202, + 250, + 211 + ], + "lines": [ + { + "bbox": [ + 239, + 202, + 250, + 211 + ], + "spans": [ + { + "bbox": [ + 239, + 202, + 250, + 211 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 244, + 118, + 291, + 155 + ], + "blocks": [ + { + "bbox": [ + 244, + 118, + 291, + 155 + ], + "lines": [ + { + "bbox": [ + 244, + 118, + 291, + 155 + ], + "spans": [ + { + "bbox": [ + 244, + 118, + 291, + 155 + ], + "type": "image", + "image_path": "c4784a760d8bc2b32816fcb1b8219c072aea690d6810ec730a0aef7d1b4b21b6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 244, + 155, + 291, + 191 + ], + "blocks": [ + { + "bbox": [ + 244, + 155, + 291, + 191 + ], + "lines": [ + { + "bbox": [ + 244, + 155, + 291, + 191 + ], + "spans": [ + { + "bbox": [ + 244, + 155, + 291, + 191 + ], + "type": "image", + "image_path": "f683e8bc0d370e8f21acf430e32f8b21ef8b19169d28807a5c710bf5e524d592.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 192, + 291, + 199 + ], + "lines": [ + { + "bbox": [ + 244, + 192, + 291, + 199 + ], + "spans": [ + { + "bbox": [ + 244, + 192, + 291, + 199 + ], + "type": "text", + "content": "Real absolute error" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 292, + 118, + 338, + 155 + ], + "blocks": [ + { + "bbox": [ + 292, + 118, + 338, + 155 + ], + "lines": [ + { + "bbox": [ + 292, + 118, + 338, + 155 + ], + "spans": [ + { + "bbox": [ + 292, + 118, + 338, + 155 + ], + "type": "image", + "image_path": "b77dc97c7ce1f08852d2144c5126918ec180ab0bbdc385da7144288d635e7e7f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 292, + 155, + 338, + 191 + ], + "blocks": [ + { + "bbox": [ + 292, + 155, + 338, + 191 + ], + "lines": [ + { + "bbox": [ + 292, + 155, + 338, + 191 + ], + "spans": [ + { + "bbox": [ + 292, + 155, + 338, + 191 + ], + "type": "image", + "image_path": "8f379e3f424a96bb1a6c75427f066387208eee2bad31904820091fe32671416f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 192, + 332, + 199 + ], + "lines": [ + { + "bbox": [ + 296, + 192, + 332, + 199 + ], + "spans": [ + { + "bbox": [ + 296, + 192, + 332, + 199 + ], + "type": "text", + "content": "Pixel-wise" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 353, + 121, + 463, + 194 + ], + "blocks": [ + { + "bbox": [ + 353, + 121, + 463, + 194 + ], + "lines": [ + { + "bbox": [ + 353, + 121, + 463, + 194 + ], + "spans": [ + { + "bbox": [ + 353, + 121, + 463, + 194 + ], + "type": "image", + "image_path": "1ca49ad23aadb3d715b3a63ce54a1cece65637740934209f1e1cd137208dcedd.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 201, + 406, + 212 + ], + "lines": [ + { + "bbox": [ + 394, + 201, + 406, + 212 + ], + "spans": [ + { + "bbox": [ + 394, + 201, + 406, + 212 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 220, + 482, + 308 + ], + "lines": [ + { + "bbox": [ + 130, + 220, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 482, + 308 + ], + "type": "text", + "content": "Fig. 2: (a) The real absolute error is obtained by subtracting the pseudo-label from the ground truth, reflecting the actual noisy level. On the other hand, pixel-wise inconsistency is derived from the subtraction of pseudo-labels from two rotated views. Visually, pixel-wise inconsistency and real absolute error appear very similar. (b) We define the area within 20 pixels from the edge as the boundary according to GT. We calculate mean pixel-wise inconsistency (MPI) and mean absolute error (MAE) in background, foreground, and boundary. The pseudo-labels are obtained by our proposed model on the test dataset COD10K [7] at a " + }, + { + "bbox": [ + 130, + 220, + 482, + 308 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 220, + 482, + 308 + ], + "type": "text", + "content": " semi-supervised setting." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 338, + 482, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 433 + ], + "type": "text", + "content": "with higher pixel-wise inconsistency are more prone to inaccuracies, suggesting a need for their de-emphasis during training. For instance-level noise, we determine that pseudo-labels with greater similarity across rotation views exhibit lower noise levels, as indicated in Fig. 3(a). A positive correlation between instance-wise consistency and the SSIM [39] computed by pseudo-labels and GT further supports this observation, depicted in Fig. 3(b). Thus, pseudo-labels demonstrating higher instance-wise consistency are likely of superior quality and should be prioritized in the learning process." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 436, + 484, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 484, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 484, + 555 + ], + "type": "text", + "content": "With the insights gained from these observations, we propose a semi-supervised camouflaged object detection framework called CamoTeacher, which incorporates a novel method known as Dual-Rotation Consistency Learning (DRCL). Specifically, DRCL operationalizes its strategy through two core components: Pixel-wise Consistency Learning (PCL) and Instance-wise Consistency Learning (ICL). PCL innovatively assigns variable weights to different parts within a pseudo-label, considering the pixel-wise inconsistency across diverse rotation view. Simultaneously, ICL adjusts the importance of individual pseudo-labels based on their instance-wise consistency, enabling a nuanced, noise-aware training process." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "We adopt SINet [7] as the base model to implement CamoTeacher and apply it to more classic COD models, i.e. CNN-based model SINet-v2 [6] and SegMaR [17] and Transformed-based model DTINet [24] and FSPNet [14]. Extensive experiments on four COD benchmark datasets, i.e. CAMO [21], CHAMELEON [33], COD10K [7], and NC4K [25], demonstrate that the proposed CamoTeacher not only achieves state-of-the-art compared with semi-supervised learning methods, but also rivals established fully-supervised learning methods. Specifically, as shown on Fig. 1(b), with only " + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": " labeled data, it nearly reached the performance of fully supervised model on COD10K." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "CamoTeacher" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 197, + 150 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 197, + 150 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 197, + 150 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 197, + 150 + ], + "type": "image", + "image_path": "eb5606f302aca6f3c031cf325c65857217b138eb00218453a2ee1e95046fde66.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 150, + 151, + 197, + 187 + ], + "blocks": [ + { + "bbox": [ + 150, + 151, + 197, + 187 + ], + "lines": [ + { + "bbox": [ + 150, + 151, + 197, + 187 + ], + "spans": [ + { + "bbox": [ + 150, + 151, + 197, + 187 + ], + "type": "image", + "image_path": "86512bfcad0262e90612a854a28984157791811722294cc37e5e97a91d44d01e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 189, + 183, + 197 + ], + "lines": [ + { + "bbox": [ + 165, + 189, + 183, + 197 + ], + "spans": [ + { + "bbox": [ + 165, + 189, + 183, + 197 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 197, + 114, + 244, + 150 + ], + "blocks": [ + { + "bbox": [ + 197, + 114, + 244, + 150 + ], + "lines": [ + { + "bbox": [ + 197, + 114, + 244, + 150 + ], + "spans": [ + { + "bbox": [ + 197, + 114, + 244, + 150 + ], + "type": "image", + "image_path": "16473e4f15ed9b7d3cd81140765cee16dc38133a1377a053e9ecad0455ea93c6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 197, + 151, + 243, + 187 + ], + "blocks": [ + { + "bbox": [ + 197, + 151, + 243, + 187 + ], + "lines": [ + { + "bbox": [ + 197, + 151, + 243, + 187 + ], + "spans": [ + { + "bbox": [ + 197, + 151, + 243, + 187 + ], + "type": "image", + "image_path": "eb3e99a5a1d883bb3ba4cc364590f0ad6a2be5f3fbcd30de54af3e545bc7145d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 189, + 225, + 196 + ], + "lines": [ + { + "bbox": [ + 216, + 189, + 225, + 196 + ], + "spans": [ + { + "bbox": [ + 216, + 189, + 225, + 196 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 240, + 200, + 251, + 209 + ], + "lines": [ + { + "bbox": [ + 240, + 200, + 251, + 209 + ], + "spans": [ + { + "bbox": [ + 240, + 200, + 251, + 209 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 245, + 114, + 291, + 150 + ], + "blocks": [ + { + "bbox": [ + 245, + 114, + 291, + 150 + ], + "lines": [ + { + "bbox": [ + 245, + 114, + 291, + 150 + ], + "spans": [ + { + "bbox": [ + 245, + 114, + 291, + 150 + ], + "type": "image", + "image_path": "bfa004632f7a3d0fae0ac3376c0f01e253a3d31bc818072604a355dd65131740.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 245, + 151, + 290, + 187 + ], + "blocks": [ + { + "bbox": [ + 245, + 151, + 290, + 187 + ], + "lines": [ + { + "bbox": [ + 245, + 151, + 290, + 187 + ], + "spans": [ + { + "bbox": [ + 245, + 151, + 290, + 187 + ], + "type": "image", + "image_path": "cf497c818a3215c6e17ac635ee5c2eaa4375ee24090dfe2b81b84108771ad664.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 189, + 285, + 198 + ], + "lines": [ + { + "bbox": [ + 251, + 189, + 285, + 198 + ], + "spans": [ + { + "bbox": [ + 251, + 189, + 285, + 198 + ], + "type": "text", + "content": "Pseudo-label" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 292, + 114, + 339, + 150 + ], + "blocks": [ + { + "bbox": [ + 292, + 114, + 339, + 150 + ], + "lines": [ + { + "bbox": [ + 292, + 114, + 339, + 150 + ], + "spans": [ + { + "bbox": [ + 292, + 114, + 339, + 150 + ], + "type": "image", + "image_path": "48b89a1690b82f7df298fd50507ebd46e4e6b6e11778e91bb9b236fd99916937.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 292, + 151, + 339, + 187 + ], + "blocks": [ + { + "bbox": [ + 292, + 151, + 339, + 187 + ], + "lines": [ + { + "bbox": [ + 292, + 151, + 339, + 187 + ], + "spans": [ + { + "bbox": [ + 292, + 151, + 339, + 187 + ], + "type": "image", + "image_path": "33ae46de3df3f755559a8a3a650bf1764c5d7efef8af9c4a73d8dc59c0b2782c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 189, + 331, + 198 + ], + "lines": [ + { + "bbox": [ + 298, + 189, + 331, + 198 + ], + "spans": [ + { + "bbox": [ + 298, + 189, + 331, + 198 + ], + "type": "text", + "content": "Pseudo-label" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 346, + 115, + 463, + 197 + ], + "blocks": [ + { + "bbox": [ + 346, + 115, + 463, + 197 + ], + "lines": [ + { + "bbox": [ + 346, + 115, + 463, + 197 + ], + "spans": [ + { + "bbox": [ + 346, + 115, + 463, + 197 + ], + "type": "image", + "image_path": "f6964705870498f4b41264c2a2ed0ef0c919bd519a923a42178a41c8f7acc88e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 200, + 411, + 209 + ], + "lines": [ + { + "bbox": [ + 400, + 200, + 411, + 209 + ], + "spans": [ + { + "bbox": [ + 400, + 200, + 411, + 209 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 218, + 482, + 274 + ], + "lines": [ + { + "bbox": [ + 130, + 218, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 482, + 274 + ], + "type": "text", + "content": "Fig. 3: (a) The more similar the pseudo-labels are between two rotated views, the lower the noise level and the higher the quality of the pseudo-labels. For example, in the first row, the pseudo-labels are more similar, indicating a lower noise level compared to the pseudo-labels in the second row. (b) The positive correlation between instance-wise consistency and the SSIM [39] of pseudo-labels and GT." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 146, + 298, + 362, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 298, + 362, + 309 + ], + "spans": [ + { + "bbox": [ + 146, + 298, + 362, + 309 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 319, + 493, + 451 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 319, + 480, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 319, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 319, + 480, + 354 + ], + "type": "text", + "content": "- We introduce the first end-to-end semi-supervised camouflaged object detection framework CamoTeacher, offering a simple yet effective baseline for future research in semi-supervised camouflaged object detection." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 356, + 480, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 356, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 138, + 356, + 480, + 426 + ], + "type": "text", + "content": "- To address the problem of large amounts of noise in pseudo-labels in semi-supervised camouflaged object detection, we proposed the Dual-Rotation Consistency Learning (DRCL), which includes Pixel-wise Consistency Learning (PCL) and Instance-wise Consistency Learning (ICL), allows for the adaptive adjustment of the contribution of pseudo-labels with varying quality, enabling the effective utilization of pseudo-label information." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 428, + 493, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 428, + 493, + 451 + ], + "spans": [ + { + "bbox": [ + 138, + 428, + 493, + 451 + ], + "type": "text", + "content": "- We conducted extensive experiments on COD benchmark datasets, and achieved significant improvements compared to the fully supervised setting." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 472, + 239, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 239, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 239, + 485 + ], + "type": "text", + "content": "2 Related works" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 131, + 499, + 316, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 499, + 316, + 511 + ], + "spans": [ + { + "bbox": [ + 131, + 499, + 316, + 511 + ], + "type": "text", + "content": "2.1 Camouflaged Object Detection" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Camouflaged Object Detection(COD) is different from traditional visual detection tasks as objects and backgrounds in traditional detection tasks have significant visual distinctions, whereas camouflaged objects are typically concealed within their surrounding environments [6, 7, 14, 17, 24, 29, 43]. In recent years, many models have been proposed on this task, which can be divided into CNN-based and Transformed-based. i) CNN-based models: SINet [7] is a simple yet effective baseline that views camouflage detection as searching and identification process. SINet-v2 [6] significantly improves performance by introducing a neighbor connection decoder and cascaded group-reversal attention design. SegMaR [17] firstly performs initial detection of camouflaged objects and then progressively enlarges the detected regions to achieve precise detection. FEDER [9] achieves impressive results by conducting feature decomposition and" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "X.Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 210 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 210 + ], + "type": "text", + "content": "edge reconstruction. ii) Transformed-based models: DTINet [24] proposes a dual-task interactive transformer to detect accurate position and detailed boundary. CamoFormer [43] presents masked separable attention to separate foreground and background regions. FSPNet [14] designs a feature shrinkage decoder hierarchically decode locality-enhanced neighboring transformer features. These outstanding works have achieved excellent results in COD, but they rely on large-scale pixel-level annotations, which are time-consuming and labor-intensive to obtain." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 480, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 480, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 480, + 355 + ], + "type": "text", + "content": "In response to the reliance of full-supervised COD on expensive pixel-level annotations, some unsupervised/weakly-supervised approaches have emerged [10-12,46]. UCOS-DA [46] formulates the unsupervised camouflaged object segmentation as a source-free unsupervised domain adaptation task and designs a foreground-background-contrastive self-adversarial pipeline. SCOD [11] proposes the first weakly-supervised COD method with relabeled scribbles. WS-SAM [10] uses the provided sparse annotations as prompts for SAM [18] to generate segmentation masks to train the model. These methods reduce annotation costs but often exhibit lower performance compared to fully-supervised methods. Is it possible to leverage existing annotated data without additional labeling while ensuring accuracy? To address this issue, we propose a new benchmark that combines semi-supervised learning with camouflaged object detection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 369, + 293, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 369, + 293, + 382 + ], + "spans": [ + { + "bbox": [ + 131, + 369, + 293, + 382 + ], + "type": "text", + "content": "2.2 Semi-Supervised Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 385, + 480, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 480, + 516 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 480, + 516 + ], + "type": "text", + "content": "In traditional supervised learning, training a high-performing model requires a substantial amount of labeled data. However, obtaining labeled data in practical applications is often costly and time-consuming. Consequently, the effective utilization of unlabeled data to improve model performance and generalization has become a critical concern. Semi-supervised learning addresses this challenge by leveraging unlabeled data to enhance the model's learning process [2,8,22,27,28,30,34,36,48], thereby improving its performance and generalization capability. Pseudo-labeling [3,35,37,38,40,42] and consistency regularization [13,19,20,23,41] are popular techniques in the field of semi-supervised learning. These methods have shown great efficacy when combined, producing impressive results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 517, + 480, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 480, + 577 + ], + "type": "text", + "content": "To address the issue of heavy reliance on labeled data in camouflage object detection, we propose the first end-to-end semi-supervised camouflage object detection method. By integrating semi-supervised learning, our approach aims to enhance the performance of camouflage object detection models while significantly reducing the need for annotated data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 592, + 233, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 592, + 233, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 592, + 233, + 605 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 613, + 249, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 613, + 249, + 624 + ], + "spans": [ + { + "bbox": [ + 131, + 613, + 249, + 624 + ], + "type": "text", + "content": "3.1 Task Formulation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "Semi-supervised camouflaged object detection aims to leverage a limited amount of annotated data to train a detector capable of identifying objects that seamlessly blend into their surroundings. This task is inherently challenging due to" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "CamoTeacher" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "the low contrast between the objects and their background. Given a camouflaged object detection dataset " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " for training, we represent the labeled subset with " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " labeled samples as " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "D_{L} = \\{x_{i}^{l},y_{i}\\}_{i = 1}^{M}" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ", and the unlabeled subset with " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " unlabeled samples as " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "D_{U} = \\{x_{i}^{u}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "x_{i}^{l}" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "x_{i}^{u}" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " represent the input image, " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " represent the corresponding annotation mask of labeled data. Typically, " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "D_{L}" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " constitutes a small fraction of the entire dataset " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ", highlighting the semi-supervised learning scenario with " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "M \\ll N" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ". The emphasis on " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "M \\ll N" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": " underscores the challenge and opportunity within semi-supervised learning: to enhance detection capabilities by utilizing the untapped potential of the unlabeled data " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "D_{U}" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": ", which far exceeds the labeled subset " + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "inline_equation", + "content": "D_{L}" + }, + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 250, + 257, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 250, + 257, + 261 + ], + "spans": [ + { + "bbox": [ + 132, + 250, + 257, + 261 + ], + "type": "text", + "content": "3.2 Overall Framework" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 266, + 482, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 338 + ], + "type": "text", + "content": "As shown in Fig. 4, we adopt Mean Teacher [36] as preliminary to achieve end-to-end semi-supervised camouflaged object detection framework, which contains two COD models with the same structure, i.e., teacher model and student model, parameterized by " + }, + { + "bbox": [ + 130, + 266, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\Theta_t" + }, + { + "bbox": [ + 130, + 266, + 482, + 338 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 266, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\Theta_s" + }, + { + "bbox": [ + 130, + 266, + 482, + 338 + ], + "type": "text", + "content": ", respectively. The teacher model generates pseudolabels, which are then used to optimize the student model. The overall loss function " + }, + { + "bbox": [ + 130, + 266, + 482, + 338 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 266, + 482, + 338 + ], + "type": "text", + "content": " can be defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 273, + 350, + 481, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 350, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 273, + 350, + 481, + 363 + ], + "type": "interline_equation", + "content": "L = L _ {s} + \\lambda_ {u} L _ {u}, \\tag {1}", + "image_path": "968c815ae592b45ab5574d1d5e416cff1ba91c28b41702e15ad76dd23b001094.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 365, + 481, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 481, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 481, + 401 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 365, + 481, + 401 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 130, + 365, + 481, + 401 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 365, + 481, + 401 + ], + "type": "inline_equation", + "content": "L_{u}" + }, + { + "bbox": [ + 130, + 365, + 481, + 401 + ], + "type": "text", + "content": " denote supervised loss and unsupervised loss respectively, " + }, + { + "bbox": [ + 130, + 365, + 481, + 401 + ], + "type": "inline_equation", + "content": "\\lambda_{u}" + }, + { + "bbox": [ + 130, + 365, + 481, + 401 + ], + "type": "text", + "content": " is the weight of unsupervised loss to balance the loss terms. Following the classic COD method, we use binary cross-entropy loss " + }, + { + "bbox": [ + 130, + 365, + 481, + 401 + ], + "type": "inline_equation", + "content": "L_{bce}" + }, + { + "bbox": [ + 130, + 365, + 481, + 401 + ], + "type": "text", + "content": " for training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 402, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 460 + ], + "type": "text", + "content": "During the training process, we employ a combination of weak augmentation " + }, + { + "bbox": [ + 130, + 402, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^w (\\cdot)" + }, + { + "bbox": [ + 130, + 402, + 482, + 460 + ], + "type": "text", + "content": " and strong augmentation " + }, + { + "bbox": [ + 130, + 402, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^s (\\cdot)" + }, + { + "bbox": [ + 130, + 402, + 482, + 460 + ], + "type": "text", + "content": " strategies. Weak augmentation is applied to labeled data to mitigate overfitting, while unlabeled data undergo various data perturbations under strong augmentation to create diverse perspectives of the same image. The supervised loss " + }, + { + "bbox": [ + 130, + 402, + 482, + 460 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 130, + 402, + 482, + 460 + ], + "type": "text", + "content": " is defined by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 227, + 470, + 481, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 470, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 227, + 470, + 481, + 502 + ], + "type": "interline_equation", + "content": "L _ {s} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} L _ {b c e} \\left(F \\left(\\mathcal {A} ^ {w} \\left(x _ {i} ^ {l}\\right); \\Theta_ {s}\\right), y _ {i}\\right), \\tag {2}", + "image_path": "9e01d94c0e3b6f8e7537b1403a4e3e6b39808be1f5e97b56442f699158859bbc.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 504, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 482, + 647 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 504, + 482, + 647 + ], + "type": "inline_equation", + "content": "F(\\mathcal{A}(x_i);\\Theta)" + }, + { + "bbox": [ + 130, + 504, + 482, + 647 + ], + "type": "text", + "content": " represents the detection results obtained by the model " + }, + { + "bbox": [ + 130, + 504, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 504, + 482, + 647 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 130, + 504, + 482, + 647 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 504, + 482, + 647 + ], + "type": "text", + "content": "-th image with augmentations " + }, + { + "bbox": [ + 130, + 504, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(\\cdot)" + }, + { + "bbox": [ + 130, + 504, + 482, + 647 + ], + "type": "text", + "content": ". For unlabeled images, we first apply weak augmentation, denoted as " + }, + { + "bbox": [ + 130, + 504, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^w (\\cdot)" + }, + { + "bbox": [ + 130, + 504, + 482, + 647 + ], + "type": "text", + "content": ", before forwarding them to the teacher model. This initial step is crucial for generating reliable pseudo-labels " + }, + { + "bbox": [ + 130, + 504, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\widehat{y}_i" + }, + { + "bbox": [ + 130, + 504, + 482, + 647 + ], + "type": "text", + "content": " under variations that do not significantly alter the core features of the images. These pseudo-labels serve as a form of soft supervision for the student model. Subsequently, the same images are subjected to strong augmentation, " + }, + { + "bbox": [ + 130, + 504, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^s (\\cdot)" + }, + { + "bbox": [ + 130, + 504, + 482, + 647 + ], + "type": "text", + "content": ", and then passed to the student model. This process introduces a higher level of variability and complexity, simulating more challenging conditions for the student model to adapt to. The student model generates predictions " + }, + { + "bbox": [ + 130, + 504, + 482, + 647 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 504, + 482, + 647 + ], + "type": "text", + "content": " based on these strongly augmented images, leveraging the pseudo-labels " + }, + { + "bbox": [ + 130, + 504, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\widehat{y}_i" + }, + { + "bbox": [ + 130, + 504, + 482, + 647 + ], + "type": "text", + "content": " for guidance in learning from the unlabeled data. It can be formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 205, + 654, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 654, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 205, + 654, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\widehat {y _ {i}} = F (\\mathcal {A} ^ {w} (x _ {i} ^ {u}); \\Theta_ {t}), p _ {i} = F (\\mathcal {A} ^ {s} (\\mathcal {A} ^ {w} (x _ {i} ^ {u})); \\Theta_ {s}). \\tag {3}", + "image_path": "fbd9b244d2f409e4e61cf3bd04151b986553b845868be829022d3f235a3e539d.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "X.Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 386, + 231 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 386, + 231 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 386, + 231 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 386, + 231 + ], + "type": "image", + "image_path": "37f00157404283e719b043fde4c197bbba4387395910e6f53f3ac6fc3cc64896.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 233, + 383, + 240 + ], + "lines": [ + { + "bbox": [ + 136, + 233, + 383, + 240 + ], + "spans": [ + { + "bbox": [ + 136, + 233, + 383, + 240 + ], + "type": "text", + "content": "DRCL: Dual-Rotation Consistency Learning PCL: Pixel-wise Consistency Learning ICL: Instance-wise Consistency Learning" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 391, + 114, + 480, + 236 + ], + "blocks": [ + { + "bbox": [ + 391, + 114, + 480, + 236 + ], + "lines": [ + { + "bbox": [ + 391, + 114, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 391, + 114, + 480, + 236 + ], + "type": "image", + "image_path": "bd3151ee0ffcb62fbe085cefbf294478d42de739c4e0a78207812d206e4d11e2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 252, + 482, + 340 + ], + "lines": [ + { + "bbox": [ + 130, + 252, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 252, + 482, + 340 + ], + "type": "text", + "content": "Fig. 4: The overall pipeline of our CamoTeacher. CamoTeacher consists of a teacher model and a student model. The teacher model generates pseudo-labels to optimize the student model, while the teacher model is updated through EMA [36] from the student. To mitigate pseudo-labels' noise, we introduced Dual-Rotation Consistency Learning (DRCL), which involves Pixel-wise Consistency Learning (PCL) and Instance-wise Consistency Learning (ICL). DRCL weights the loss function based on pixel-wise inconsistency and instance-wise consistency in pseudo-labels across different rotation views." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 365, + 367, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 367, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 367, + 376 + ], + "type": "text", + "content": "Hence, the unsupervised loss " + }, + { + "bbox": [ + 132, + 365, + 367, + 376 + ], + "type": "inline_equation", + "content": "L_{u}" + }, + { + "bbox": [ + 132, + 365, + 367, + 376 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 252, + 386, + 481, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 252, + 386, + 481, + 418 + ], + "type": "interline_equation", + "content": "L _ {u} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} L _ {b c e} \\left(p _ {i}, \\widehat {y _ {i}}\\right). \\tag {4}", + "image_path": "21931cd7f6f2766bea5c72817819a069107535821956513d22c820ae8f5b5573.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 427, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 482, + 511 + ], + "type": "text", + "content": "Finally, the student model is intensively trained using the total loss " + }, + { + "bbox": [ + 130, + 427, + 482, + 511 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 427, + 482, + 511 + ], + "type": "text", + "content": ", which encapsulates both the supervised and unsupervised learning aspects of the semi-supervised framework. This approach ensures that the student model benefits from both labeled and pseudo-labeled data, enhancing its detection capabilities. Concurrently, the teacher model is methodically updated through the Exponential Moving Average (EMA) mechanism [36], which effectively distills students' knowledge while preventing noise, formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 254, + 521, + 481, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 521, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 254, + 521, + 481, + 534 + ], + "type": "interline_equation", + "content": "\\Theta_ {t} \\leftarrow \\eta \\Theta_ {t} + (1 - \\eta) \\Theta_ {s}, \\tag {5}", + "image_path": "09ec3eba683a9351db713adc0a5ed041a1e22876a7aa2e39a9dc63616737c31f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 543, + 371, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 371, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 371, + 555 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 543, + 371, + 555 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 132, + 543, + 371, + 555 + ], + "type": "text", + "content": " is hyper-parameters denotes the keeping rate." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 573, + 345, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 573, + 345, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 573, + 345, + 586 + ], + "type": "text", + "content": "3.3 Dual-Rotation Consistency Learning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Due to the camouflaging nature of the objects, pseudo-labels contain a substantial amount of noise, and directly using them to optimize the student model potentially harm the model's performance. To address this problem, one of the most intuitive possible methods is to set a fixed high threshold to filter high-quality pseudo-labels, but this results in a low recall rate and makes it difficult to fully utilize the supervision information of pseudo-labels. To this end, we propose" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "CamoTeacher" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "Dual-Rotation Consistency Learning (DRCL) to dynamically adjust the weight of the pseudo-labels to reduce the impact of noise." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 175 + ], + "type": "text", + "content": "We randomly apply two separate rotations to the image " + }, + { + "bbox": [ + 130, + 140, + 480, + 175 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 140, + 480, + 175 + ], + "type": "text", + "content": ", which has previously been flipped and random resized, resulting in two distinct rotation views " + }, + { + "bbox": [ + 130, + 140, + 480, + 175 + ], + "type": "inline_equation", + "content": "x_{i}^{r_{1}}" + }, + { + "bbox": [ + 130, + 140, + 480, + 175 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 140, + 480, + 175 + ], + "type": "inline_equation", + "content": "x_{i}^{r_{2}}" + }, + { + "bbox": [ + 130, + 140, + 480, + 175 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 214, + 176, + 480, + 189 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 176, + 480, + 189 + ], + "spans": [ + { + "bbox": [ + 214, + 176, + 480, + 189 + ], + "type": "interline_equation", + "content": "x _ {i} ^ {r _ {1}} = R \\left(\\mathcal {A} ^ {w} \\left(x _ {i}\\right), \\theta_ {1}\\right), x _ {i} ^ {r _ {2}} = R \\left(\\mathcal {A} ^ {w} \\left(x _ {i}\\right), \\theta_ {2}\\right), \\tag {6}", + "image_path": "b09055d31094b489994082d25164ad9e18b2f08930990cf5017c82f02e1ff070.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 193, + 481, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 193, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 131, + 193, + 481, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 193, + 481, + 266 + ], + "type": "inline_equation", + "content": "x_{i}^{r} = R(x_{i},\\theta)" + }, + { + "bbox": [ + 131, + 193, + 481, + 266 + ], + "type": "text", + "content": " represents rotate the input image " + }, + { + "bbox": [ + 131, + 193, + 481, + 266 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 131, + 193, + 481, + 266 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 131, + 193, + 481, + 266 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 193, + 481, + 266 + ], + "type": "text", + "content": " degrees. We input the acquired rotation view into the teacher model to obtain the corresponding prediction, i.e. " + }, + { + "bbox": [ + 131, + 193, + 481, + 266 + ], + "type": "inline_equation", + "content": "\\widehat{y}_i^r = F(x_i^r;\\Theta_t)" + }, + { + "bbox": [ + 131, + 193, + 481, + 266 + ], + "type": "text", + "content": ". Subsequently, we apply an opposite rotation of " + }, + { + "bbox": [ + 131, + 193, + 481, + 266 + ], + "type": "inline_equation", + "content": "-\\theta" + }, + { + "bbox": [ + 131, + 193, + 481, + 266 + ], + "type": "text", + "content": " to return the prediction to its original horizontal orientation, resulting in " + }, + { + "bbox": [ + 131, + 193, + 481, + 266 + ], + "type": "inline_equation", + "content": "\\widehat{y}_i^{h_1}" + }, + { + "bbox": [ + 131, + 193, + 481, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 193, + 481, + 266 + ], + "type": "inline_equation", + "content": "\\widehat{y}_i^{h_2}" + }, + { + "bbox": [ + 131, + 193, + 481, + 266 + ], + "type": "text", + "content": ", facilitating the calculation of prediction inconsistency under different rotation views, respectively," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 224, + 274, + 481, + 288 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 274, + 481, + 288 + ], + "spans": [ + { + "bbox": [ + 224, + 274, + 481, + 288 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {i} ^ {h _ {1}} = R \\left(\\hat {y} _ {i} ^ {r _ {1}}, - \\theta_ {1}\\right), \\hat {y} _ {i} ^ {h _ {2}} = R \\left(\\hat {y} _ {i} ^ {r _ {2}}, - \\theta_ {2}\\right). \\tag {7}", + "image_path": "d96ffc1087c7630f682f0840d9be29b5a95926cac2063037520d4870466508a3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 295, + 479, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 479, + 318 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 479, + 318 + ], + "type": "text", + "content": "Note that rotation introduces black boundary areas, which are not involved in the implement of DRCL." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 319, + 480, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 480, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 480, + 355 + ], + "type": "text", + "content": "Since the noise levels vary across different regions of pseudo-labels and among different pseudo-labels, we introduce PCL and ICL dynamically adapt the contributions of different pixels within pseudo-labels and across various pseudo-labels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 371, + 479, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 371, + 479, + 398 + ], + "spans": [ + { + "bbox": [ + 131, + 371, + 479, + 398 + ], + "type": "text", + "content": "Pixel-wise Consistency Learning We subtract the horizontal predictions " + }, + { + "bbox": [ + 131, + 371, + 479, + 398 + ], + "type": "inline_equation", + "content": "\\widehat{y}_i^{h_1}" + }, + { + "bbox": [ + 131, + 371, + 479, + 398 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 371, + 479, + 398 + ], + "type": "inline_equation", + "content": "\\widehat{y}_i^{h_2}" + }, + { + "bbox": [ + 131, + 371, + 479, + 398 + ], + "type": "text", + "content": " at the pixel level to obtain the pixel-wise inconsistency " + }, + { + "bbox": [ + 131, + 371, + 479, + 398 + ], + "type": "inline_equation", + "content": "\\Delta_i" + }, + { + "bbox": [ + 131, + 371, + 479, + 398 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 270, + 406, + 480, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 406, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 270, + 406, + 480, + 419 + ], + "type": "interline_equation", + "content": "\\Delta_ {i} = \\left| \\widehat {y} _ {i} ^ {h _ {1}} - \\widehat {y} _ {i} ^ {h _ {2}} \\right|. \\tag {8}", + "image_path": "6a63ae0436627044730c800c52c23986aafee00359d5afb07418f094aa23feb3.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 427, + 480, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 480, + 511 + ], + "type": "text", + "content": "The pixel-wise inconsistency " + }, + { + "bbox": [ + 130, + 427, + 480, + 511 + ], + "type": "inline_equation", + "content": "\\Delta_{i}" + }, + { + "bbox": [ + 130, + 427, + 480, + 511 + ], + "type": "text", + "content": " across different views reflects the reliability of pseudo-labels. However, in cases where both rotation views' prediction are close to 0.5, " + }, + { + "bbox": [ + 130, + 427, + 480, + 511 + ], + "type": "inline_equation", + "content": "\\Delta_{i}" + }, + { + "bbox": [ + 130, + 427, + 480, + 511 + ], + "type": "text", + "content": " cannot effectively distinguish them. These predictions exhibit high uncertainty, meaning they cannot be definitively classified as foreground or background, and are likely to represent noisy labels. Consequently, it becomes necessary to diminish their influence by lowering their weights. Hence, we calculate the average of horizontal predictions " + }, + { + "bbox": [ + 130, + 427, + 480, + 511 + ], + "type": "inline_equation", + "content": "\\widehat{y}_i^h" + }, + { + "bbox": [ + 130, + 427, + 480, + 511 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 265, + 519, + 480, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 519, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 265, + 519, + 480, + 533 + ], + "type": "interline_equation", + "content": "\\widehat {y} _ {i} ^ {h} = \\operatorname {a v g} \\left(\\widehat {y} _ {i} ^ {h _ {1}}, \\widehat {y} _ {i} ^ {h _ {2}}\\right), \\tag {9}", + "image_path": "fb5b0519a051d88f6080a4b39cf8156cc89b8aec1c4edafe7c755f6e5fde1838.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 540, + 479, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 479, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 479, + 563 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 540, + 479, + 563 + ], + "type": "inline_equation", + "content": "avg(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 540, + 479, + 563 + ], + "type": "text", + "content": " represents calculate the average of two inputs in pixel-level, and the use its L2 distance from 0.5 as a component to adjust the weights." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 564, + 479, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 564, + 479, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 564, + 479, + 588 + ], + "type": "text", + "content": "As a result, we derive the pixel-wise consistency weight " + }, + { + "bbox": [ + 130, + 564, + 479, + 588 + ], + "type": "inline_equation", + "content": "\\omega_{i}^{pc}" + }, + { + "bbox": [ + 130, + 564, + 479, + 588 + ], + "type": "text", + "content": " based on the pixel-wise inconsistency among different rotation views, as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 249, + 597, + 480, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 597, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 249, + 597, + 480, + 610 + ], + "type": "interline_equation", + "content": "\\omega_ {i} ^ {p c} = \\left(1 - \\Delta_ {i} ^ {\\alpha}\\right) \\left| \\left| \\hat {y} _ {i} ^ {h} - \\mu \\right| \\right| _ {2} ^ {2}, \\tag {10}", + "image_path": "6e9c3cfc9c5a5cb8982d023fc32fdc68bc6d1cb7f0a17fffe1d7041708195db3.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": " is a hyper-parameter and " + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mu = 0.5" + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": ". This dynamic pixel-wise consistency weight " + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\omega_{i}^{pc}" + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": " assigns greater weight to regions with consistent predictions across various rotation views, while regions with inconsistent predictions receive a smaller weight." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "X.Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 146, + 115, + 367, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 115, + 367, + 129 + ], + "spans": [ + { + "bbox": [ + 146, + 115, + 367, + 129 + ], + "type": "text", + "content": "In conclusion, we formulate the PCL loss " + }, + { + "bbox": [ + 146, + 115, + 367, + 129 + ], + "type": "inline_equation", + "content": "L_{u}^{PC}" + }, + { + "bbox": [ + 146, + 115, + 367, + 129 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 212, + 136, + 481, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 136, + 481, + 220 + ], + "spans": [ + { + "bbox": [ + 212, + 136, + 481, + 220 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {u} ^ {P C} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\omega_ {i} ^ {p c} L _ {b c e} (p _ {i}, \\widehat {y} _ {i} ^ {r _ {1}}) \\\\ = - \\frac {1}{N H W} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {H \\times W} \\omega_ {i, j} ^ {p c} \\left[ \\hat {y} _ {i, j} ^ {r _ {1}} \\log p _ {i, j} \\right. \\tag {11} \\\\ + \\left(1 - \\widehat {y} _ {i, j} ^ {r _ {1}}\\right) \\log \\left(1 - p _ {i, j}\\right) ], \\\\ \\end{array}", + "image_path": "a089146332cf62e51cd7e8e034a3c091af60008eabcbfdb5820f98e278d405f3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 479, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 479, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 479, + 248 + ], + "type": "text", + "content": "adaptively adjusts the weight of each pixel to ensure comprehensive supervision of student model while avoiding confirmation bias." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 264, + 480, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 264, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 480, + 372 + ], + "type": "text", + "content": "Instance-wise Consistency Learning The degree of camouflage varies between different images, leading to significant variations in the quality of pseudolabels across images. Treating all pseudo-labels equally would be unreasonable. Unfortunately, for unlabeled images, assessing the quality of pseudo-labels is challenging because there are no GT labels available. We have observed a positive correlation between the instance-wise consistency of two rotation views and the quality of pseudo-labels, as quantified by SSIM [39]. Base on this, we introduce ICL to adjust the contributions of pseudo-labels with varying quality. We represent the instance-wise consistency weight " + }, + { + "bbox": [ + 130, + 264, + 480, + 372 + ], + "type": "inline_equation", + "content": "\\omega_{i}^{ic}" + }, + { + "bbox": [ + 130, + 264, + 480, + 372 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 249, + 377, + 480, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 377, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 249, + 377, + 480, + 392 + ], + "type": "interline_equation", + "content": "\\omega_ {i} ^ {i c} = \\left(S S I M \\left(\\widehat {y} _ {i} ^ {h _ {1}}, \\widehat {y} _ {i} ^ {h _ {2}}\\right)\\right) ^ {\\beta}, \\tag {12}", + "image_path": "1a3c27dd9f09848267e171497e5c498d351384a1927bf6596f674ff1ccc240cc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 396, + 479, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 479, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 479, + 420 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 396, + 479, + 420 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 396, + 479, + 420 + ], + "type": "text", + "content": " is a hyper-parameter used to adjust the distribution relationship between instance-wise consistency and pseudo-label quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 421, + 479, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 421, + 479, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 421, + 479, + 445 + ], + "type": "text", + "content": "We use intersection-over-union (IoU) loss as the instance-wise restriction, therefore, the ICL loss can be expressed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 188, + 451, + 480, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 451, + 480, + 519 + ], + "spans": [ + { + "bbox": [ + 188, + 451, + 480, + 519 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {u} ^ {I C} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\omega_ {i} ^ {i c} L _ {i o u} (p _ {i}, \\widehat {y} _ {i} ^ {r _ {1}}) \\\\ = \\frac {1}{N H W} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {H \\times W} \\omega_ {i} ^ {i c} \\left(1 - \\frac {p _ {i , j} \\widehat {y} _ {i , j} ^ {r _ {1}}}{p _ {i , j} + \\widehat {y} _ {i , j} ^ {r _ {1}} - p _ {i , j} \\widehat {y} _ {i , j} ^ {r _ {1}}}\\right). \\tag {13} \\\\ \\end{array}", + "image_path": "45ad02b7a0d1d41dce59a0659cdf74a89fa9d1a50cf704b885a189f01a1c9d18.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 524, + 479, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 524, + 479, + 548 + ], + "spans": [ + { + "bbox": [ + 131, + 524, + 479, + 548 + ], + "type": "text", + "content": "Therefore, the final total " + }, + { + "bbox": [ + 131, + 524, + 479, + 548 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 131, + 524, + 479, + 548 + ], + "type": "text", + "content": " loss comprises three components: supervised loss " + }, + { + "bbox": [ + 131, + 524, + 479, + 548 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 131, + 524, + 479, + 548 + ], + "type": "text", + "content": ", PCL loss " + }, + { + "bbox": [ + 131, + 524, + 479, + 548 + ], + "type": "inline_equation", + "content": "L_{u}^{LC}" + }, + { + "bbox": [ + 131, + 524, + 479, + 548 + ], + "type": "text", + "content": ", and ICL loss " + }, + { + "bbox": [ + 131, + 524, + 479, + 548 + ], + "type": "inline_equation", + "content": "L_{u}^{GC}" + }, + { + "bbox": [ + 131, + 524, + 479, + 548 + ], + "type": "text", + "content": ", which can be expressed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 246, + 553, + 480, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 553, + 480, + 569 + ], + "spans": [ + { + "bbox": [ + 246, + 553, + 480, + 569 + ], + "type": "interline_equation", + "content": "L = L _ {s} + \\lambda_ {u} ^ {p c} L _ {u} ^ {P C} + \\lambda_ {u} ^ {i c} L _ {u} ^ {I C}, \\tag {14}", + "image_path": "de4083006b5be286041d61192d155c549cf6d62a6bce0e67410ba9434f6a93bb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 573, + 310, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 573, + 310, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 573, + 310, + 586 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 573, + 310, + 586 + ], + "type": "inline_equation", + "content": "\\lambda_u^{pc},\\lambda_u^{ic}" + }, + { + "bbox": [ + 131, + 573, + 310, + 586 + ], + "type": "text", + "content": " are the hyper-parameters." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 601, + 224, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 601, + 224, + 615 + ], + "spans": [ + { + "bbox": [ + 131, + 601, + 224, + 615 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 624, + 264, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 264, + 636 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 264, + 636 + ], + "type": "text", + "content": "4.1 Experiment Settings" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Dataset. We evaluate our CamoTeacher on four benchmark datasets, CAMO [21], CHAMELEON [33], COD10K [7], and NC4K [25]. In CAMO, there are" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 390, + 91, + 447, + 100 + ], + "type": "text", + "content": "CamoTeacher" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 191, + 479, + 243 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 1: Experimental results on four COD benchmark datasets includes CAMO, CHAMELEON, COD10K, NC4K. " + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": " denotes using the same data augmentation as semi-supervised training. " + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": " indicates the higher the score the better, and " + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": " indicates the lower the better. \"n%\" means that \"n%\" of the data was utilized as labeled data, with the remainder as unlabeled data. As the data partitioning is random, we conducted 5 fold experiments and the results are the average of all 5 folds." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 191, + 479, + 243 + ], + "lines": [ + { + "bbox": [ + 134, + 191, + 479, + 243 + ], + "spans": [ + { + "bbox": [ + 134, + 191, + 479, + 243 + ], + "type": "table", + "html": "
CHAMELEON (76)
Methods1% (41)5% (202)10% (404)
\\( S_{\\alpha} \\uparrow \\)\\( F_{\\beta}^{w} \\uparrow \\)\\( E_{\\phi}^{m} \\uparrow \\)\\( E_{\\phi}^{x} \\uparrow \\)\\( F_{\\beta}^{m} \\uparrow \\)\\( M \\downarrow \\)\\( S_{\\alpha} \\uparrow \\)\\( F_{\\beta}^{w} \\uparrow \\)\\( E_{\\phi}^{m} \\uparrow \\)\\( E_{\\phi}^{x} \\uparrow \\)\\( F_{\\beta}^{m} \\uparrow \\)\\( M \\downarrow \\)\\( S_{\\alpha} \\uparrow \\)\\( F_{\\beta}^{\\omega } \\uparrow \\)\\( E_{\\phi}^{m} \\uparrow \\)\\( E_{\\phi}^{x} \\uparrow \\)\\( F_{\\beta}^{m} \\uparrow \\)\\( M \\downarrow \\)
Supervised0.6050.2680.5820.7410.4350.2160.6130.2830.6010.7320.4390.1850.7140.4340.7250.8290.5920.110
Supervised†0.6590.3310.6550.7870.4990.1590.7280.4540.7440.8440.6070.106
Mean Teacher0.5370.1990.4180.6360.2290.2040.6110.3090.5240.7450.3530.1370.6790.4500.6500.8120.5120.102
CamoTeacher0.6520.4760.7140.7620.5580.0930.7290.5870.7850.8220.6560.0700.7560.6170.8130.8510.6840.065
", + "image_path": "6bb7c3b9de840178cdb1f2f38dd56d12dbadee66319e3f3840f18a3f3694bbbc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 134, + 243, + 479, + 294 + ], + "blocks": [ + { + "bbox": [ + 134, + 243, + 479, + 294 + ], + "lines": [ + { + "bbox": [ + 134, + 243, + 479, + 294 + ], + "spans": [ + { + "bbox": [ + 134, + 243, + 479, + 294 + ], + "type": "table", + "html": "
CAMO (250)
Methods1% (41)5% (202)10% (404)
\\( S_{\\alpha} \\uparrow \\)\\( F_{w}^{w} \\uparrow \\)\\( E_{m}^{m} \\uparrow \\)\\( E_{\\sigma}^{e} \\uparrow \\)\\( F_{g}^{m} \\uparrow \\)\\( M \\downarrow \\)\\( S_{\\alpha} \\uparrow \\)\\( F_{w}^{w} \\uparrow \\)\\( E_{\\sigma}^{m} \\uparrow \\)\\( E_{\\sigma}^{e} \\uparrow \\)\\( F_{g}^{m} \\uparrow \\)\\( M \\downarrow \\)\\( S_{\\alpha} \\uparrow \\)\\( F_{w}^{w} \\uparrow \\)\\( E_{m}^{m} \\uparrow \\)\\( E_{\\sigma}^{e} \\uparrow \\)\\( F_{g}^{m} \\uparrow \\)\\( M \\downarrow \\)
Supervised0.5730.2710.5410.7040.4020.2510.5920.2980.5810.7040.4320.2160.6520.3900.6470.7600.5250.159
Supervised †0.5960.2980.5800.7170.4420.2350.6100.3160.5950.7200.4500.2050.6670.4090.6660.7790.5460.153
Mean Teacher0.5180.2070.3990.6200.2270.2260.5750.2860.4820.7080.3220.1840.6250.3970.5780.7730.4540.150
CamoTeacher0.6210.4560.6690.7360.5450.1360.6690.5230.7110.7750.6010.1220.7010.5600.7420.7950.6350.112
", + "image_path": "2f22377da969e5568c5f8580d78c0124919c1932e4538b2afbbfc3a6c99dfd89.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 134, + 295, + 479, + 345 + ], + "blocks": [ + { + "bbox": [ + 134, + 295, + 479, + 345 + ], + "lines": [ + { + "bbox": [ + 134, + 295, + 479, + 345 + ], + "spans": [ + { + "bbox": [ + 134, + 295, + 479, + 345 + ], + "type": "table", + "html": "
COD10K (2026)
Methods1% (41)5% (202)10% (404)
So ↑Fw↑Em↑Eax↑Eax↑Fm↑M↓So ↑Fw↑Em↑Eax↑Eax↑Fm↑M↓So ↑Fw↑Em↑Eax↑Eax↑Fm↑M↓
Supervised0.5660.1890.5760.7230.3560.2090.5960.2200.6270.7470.4030.1640.6750.3360.7260.8100.5240.095
Supervised †0.5820.2060.6090.7430.3860.1940.6210.2460.6590.7740.4440.1470.6800.3380.7260.8120.5280.097
Mean Teacher0.5460.1680.4410.6330.2260.1610.6210.2720.5550.7320.3430.1070.6830.4040.6660.7990.4820.078
CamoTeacher0.6990.5170.7880.7970.5820.0620.7450.5830.8270.8400.6440.0500.7590.5940.8360.8540.6520.049
", + "image_path": "cd7bd3c03c3653fafce76d490ec32e2f8ab815ed542ed933399bc8d51206186c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 134, + 346, + 479, + 396 + ], + "blocks": [ + { + "bbox": [ + 134, + 346, + 479, + 396 + ], + "lines": [ + { + "bbox": [ + 134, + 346, + 479, + 396 + ], + "spans": [ + { + "bbox": [ + 134, + 346, + 479, + 396 + ], + "type": "table", + "html": "
NC4K (4121)
Methods1% (41)5% (202)10% (404)
\\( S_{\\alpha} \\uparrow \\)\\( E_{m}^{w} \\uparrow \\)\\( E_{m}^{m} \\uparrow \\)\\( E_{e}^{s} \\uparrow \\)\\( F_{m}^{m} \\uparrow \\)\\( M \\downarrow \\)\\( S_{\\alpha} \\uparrow \\)\\( F_{m}^{w} \\uparrow \\)\\( E_{m}^{m} \\uparrow \\)\\( E_{e}^{s} \\uparrow \\)\\( F_{m}^{m} \\uparrow \\)\\( M \\downarrow \\)\\( S_{\\alpha} \\uparrow \\)\\( F_{m}^{w} \\uparrow \\)\\( E_{m}^{m} \\uparrow } \\)\\( E_{e}^{s} \\uparrow \\)\\( F_{m}^{m} \\uparrow \\)\\( M \\downarrow \\)
Supervised0.6180.2910.6040.7520.4740.2130.6490.3320.6600.7730.5240.1790.7270.4620.7540.8330.6430.116
Supervised †0.6380.3130.6390.7680.5100.2010.6760.3620.6890.7980.5620.1630.7350.4710.7600.8370.6500.114
Mean Teacher0.5410.2130.4240.6370.2580.1930.6340.3550.5560.7670.4200.1400.7000.4920.6700.8270.5650.109
CamoTeacher0.7180.5990.7790.8140.6750.0900.7770.6770.8340.8590.7390.0710.7910.6870.8420.8680.7460.068
", + "image_path": "bd03aa88fc663650e25f795d6a8ddb0a6b683916d3fa83474f598b486f402eba.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 423, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 554 + ], + "type": "text", + "content": "2500 images, consisting of 1250 camouflage images and 1250 non-camouflage images. CHAMELEON contains 76 manually annotated images. COD10K dataset consists of 5066 camouflage images, 3000 background images, and 1934 non-camouflage images. NC4K is another large-scale COD test dataset that comprises 4,121 images. Based on the data partitioning in previous works [7,29,45], we used 3040 images from COD10K and 1000 images from CAMO as the training set for our experiments. The remaining images from both datasets were used as the test set. During the training process, we followed a semi-supervised segmentation [15,47] data partitioning approach. We sampled " + }, + { + "bbox": [ + 130, + 423, + 482, + 554 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 423, + 482, + 554 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 423, + 482, + 554 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 423, + 482, + 554 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 423, + 482, + 554 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 423, + 482, + 554 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 423, + 482, + 554 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 423, + 482, + 554 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 423, + 482, + 554 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 423, + 482, + 554 + ], + "type": "text", + "content": " of the training set as labeled data, while the remaining portion was used as unlabeled data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "content": "Evaluation Metrics. Referring to previous work [16, 17, 29], we used 6 common evaluation indicators in COD to evaluate our CamoTeacher, including S-measure " + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "inline_equation", + "content": "(S_{\\alpha})" + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "content": " [4], weight F-measure " + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "inline_equation", + "content": "(F_{\\beta}^{w})" + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "content": " [26], mean E-measure " + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "inline_equation", + "content": "(E_{\\phi}^{m})" + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "content": " [5], max E-measure " + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "inline_equation", + "content": "(E_{\\phi}^{x})" + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "content": ", mean F-measure " + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "inline_equation", + "content": "(F_{\\beta}^{m})" + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "content": " [1], and mean absolute error " + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "inline_equation", + "content": "(M)" + }, + { + "bbox": [ + 130, + 555, + 482, + 604 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Implementation Details. The proposed CamoTeacher is implemented with PyTorch. We adopt SINet [7] as the baseline of COD model. We use an SGD optimizer with a momentum of 0.9 and a polynomial learning-rate decay with an initial value of 0.01 to train the student model. The training epoch is set to 40 with 10 epochs for burn-in phase. And the batch size is 20, with a ratio of" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "X.Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 146, + 479, + 276 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 136 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 136 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 136 + ], + "type": "text", + "content": "Table 2: Results of CamoTeacher on different COD models with " + }, + { + "bbox": [ + 132, + 114, + 479, + 136 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 132, + 114, + 479, + 136 + ], + "type": "text", + "content": " labeled data on COD10K." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 146, + 479, + 276 + ], + "lines": [ + { + "bbox": [ + 135, + 146, + 479, + 276 + ], + "spans": [ + { + "bbox": [ + 135, + 146, + 479, + 276 + ], + "type": "table", + "html": "
ModelSettingSα↑Fβw↑Eφm↑Eφx↑Fβm↑M↓
CNN-Based Models
SINet-v2 [6] (TPAMI2022)supervised0.7320.5440.8190.8350.5860.056
CamoTeacher0.7600.6050.8350.8430.6520.051
SegMaR [17] (CVPR2022)supervised0.7550.5830.8370.8510.6230.050
CamoTeacher0.7750.6290.8560.8640.6750.045
Transformer-Based Models
DTINet [24] (ICPR2022)supervised0.7710.6260.8570.8650.6700.043
CamoTeacher0.7940.6610.8730.8820.6950.040
FSPNet [14] (CVPR2023)supervised0.7890.6220.8230.8670.6750.042
CamoTeacher0.8170.6840.8840.9030.7200.037
", + "image_path": "e619d9210d4fb34e3b4a2075a534abae5c9b33a2f819337fdaf7a0ea1590c626.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 297, + 482, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 381 + ], + "type": "text", + "content": "1:1 between labeled and unlabeled data, i.e. each batch consists of 10 labeled and 10 unlabeled images. During training and inference, each image is resized to " + }, + { + "bbox": [ + 130, + 297, + 482, + 381 + ], + "type": "inline_equation", + "content": "352 \\times 352" + }, + { + "bbox": [ + 130, + 297, + 482, + 381 + ], + "type": "text", + "content": ". The teacher model is updated via EMA with a momentum " + }, + { + "bbox": [ + 130, + 297, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 297, + 482, + 381 + ], + "type": "text", + "content": " of 0.996. Weak data augmentation involves Random Flip and Random Scale, while strong data augmentation entails color space transformations including Identity, Autocontrast, Equalize, Gaussian blur, Contrast, Sharpness, Color, Brightness, Hue, Posterize, Solarize, with a random selection of up to 3 from this list." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 396, + 228, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 228, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 228, + 407 + ], + "type": "text", + "content": "4.2 Main Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 414, + 481, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 481, + 582 + ], + "type": "text", + "content": "Comparison with Baselines. In Tab. 1, we compared CamoTeacher, the supervised baseline [7], and Mean Teacher [36]. Mean Teacher does not show improvements over the supervised in the semi-supervised setting due to treating all pseudo-labels and regions equally, which does not account for pseudo-labels noise problem. For instance, on the NC4K test set, under the " + }, + { + "bbox": [ + 130, + 414, + 481, + 582 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 414, + 481, + 582 + ], + "type": "text", + "content": " labeled data, the " + }, + { + "bbox": [ + 130, + 414, + 481, + 582 + ], + "type": "inline_equation", + "content": "S_{\\alpha}" + }, + { + "bbox": [ + 130, + 414, + 481, + 582 + ], + "type": "text", + "content": " dropped by " + }, + { + "bbox": [ + 130, + 414, + 481, + 582 + ], + "type": "inline_equation", + "content": "9.7\\%" + }, + { + "bbox": [ + 130, + 414, + 481, + 582 + ], + "type": "text", + "content": " compared to supervised. Compared to Mean Teacher, CamoTeacher integrated with DRCL effectively mitigates pseudo-label noise issues and efficiently utilizes pseudo-label information. It demonstrates substantial improvements over both supervised and Mean Teacher across various proportions of labeled data, especially in scenarios with extremely limited. For instance, on " + }, + { + "bbox": [ + 130, + 414, + 481, + 582 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 414, + 481, + 582 + ], + "type": "text", + "content": " NC4K, CamoTeacher exhibited a substantial increase of " + }, + { + "bbox": [ + 130, + 414, + 481, + 582 + ], + "type": "inline_equation", + "content": "28.6\\%" + }, + { + "bbox": [ + 130, + 414, + 481, + 582 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 130, + 414, + 481, + 582 + ], + "type": "inline_equation", + "content": "F_{\\beta}^{w}" + }, + { + "bbox": [ + 130, + 414, + 481, + 582 + ], + "type": "text", + "content": " and a " + }, + { + "bbox": [ + 130, + 414, + 481, + 582 + ], + "type": "inline_equation", + "content": "12.3\\%" + }, + { + "bbox": [ + 130, + 414, + 481, + 582 + ], + "type": "text", + "content": " enhancement of " + }, + { + "bbox": [ + 130, + 414, + 481, + 582 + ], + "type": "inline_equation", + "content": "S_{\\alpha}" + }, + { + "bbox": [ + 130, + 414, + 481, + 582 + ], + "type": "text", + "content": ". Compared with Mean Teacher, the performance gains by CamoTeacher are more significant, which are " + }, + { + "bbox": [ + 130, + 414, + 481, + 582 + ], + "type": "inline_equation", + "content": "38.6\\%" + }, + { + "bbox": [ + 130, + 414, + 481, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 414, + 481, + 582 + ], + "type": "inline_equation", + "content": "17.7\\%" + }, + { + "bbox": [ + 130, + 414, + 481, + 582 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "In Fig. 1(b), we visualize the performance changes of supervised, Mean teacher and our CamoTeacher on COD10K under different labeled data ratios of " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": ", and further " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": ", which intuitively reflects the effectiveness of CamoTeacher. It's evident that metrics like " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "F_{\\beta}^{w}" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "F_{\\beta}^{m}" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "E_{\\phi}^{m}" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": " surpass " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": " of fully supervised performance with just " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": " labeled data. Additionally, with " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": " labeled data, all metrics significantly exceed " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": " of fully supervised performance." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "CamoTeacher" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 190, + 480, + 367 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 3: Result of CamoTeacher on COD10K with additional data. We use " + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": " train dataset as labeled data and test dataset as the unlabeled data. \"U\" represents the unsupervised setting, \"W\" denotes the weakly-supervised setting, where weakly-supervision results from scribble annotations. \"L100%\" indicates using " + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": " of the training set as labeled data, \"U(CAMO)\" denotes utilizing CAMO as unlabeled data, and likewise for others." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 190, + 480, + 367 + ], + "lines": [ + { + "bbox": [ + 135, + 190, + 480, + 367 + ], + "spans": [ + { + "bbox": [ + 135, + 190, + 480, + 367 + ], + "type": "table", + "html": "
ModelSettingSα↑Fwβ↑Emφ↑Exφ↑Fmβ↑M↓
Unsupervised
SelfMask [31] (CVPRW2022)U0.645 0.440 0.687 0.728 0.478 0.125
FOUND [32] (CVPR2023)U0.670 0.482 0.751 0.753 0.520 0.085
UCOS-DA [46] (ICCVW2023)U0.689 0.513 0.740 0.741 0.546 0.086
Weakly-Supervised
SCWSSOD [44] (AAAI2021)W0.710 0.546 0.805 - - 0.055
SCOD [11](AAAI2023)W0.733 0.576 0.832 0.845 0.633 0.049
WS-SAM [10] (NeurIPS2023)W0.803 - 0.878 - - 0.038
Semi-Supervised
CamoTeacher SINetL100%0.771 0.551 0.806 0.868 0.634 0.051
L100%+U(CAMO)0.779 0.554 0.799 0.875 0.632 0.051
L100%+U(NC4K)0.785 0.574 0.812 0.879 0.642 0.049
L100%+U(CAMO,NC4K)0.787 0.576 0.813 0.881 0.644 0.048
", + "image_path": "ff68466bdbca52b35cab1521aef42b40ae339ea72f837f84cae70bb16aa5b0c2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "text", + "content": "Application to Existing COD Methods. We applied the CamoTeacher approach to additional COD models, including CNN-based model SINet-v2 [6] and SegMaR [17] and Transformed-based model DTINet [24] and FSPNet [14], and the corresponding results are presented in Tab. 2. In terms of SINet-v2, CamoTeacher demonstrated a " + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "text", + "content": " improvement in the " + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "inline_equation", + "content": "S_{\\alpha}" + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "text", + "content": " metric and a " + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "inline_equation", + "content": "6.1\\%" + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "text", + "content": " improvement in the " + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "inline_equation", + "content": "F_{\\beta}^{w}" + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "text", + "content": " metric compared to the supervised baseline. Based on FSPNet, CamoTeacher achieved a " + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "text", + "content": " enhancement in the " + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "inline_equation", + "content": "S_{\\alpha}" + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "text", + "content": " metric and a " + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "inline_equation", + "content": "6.2\\%" + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "text", + "content": " improvement in the " + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "inline_equation", + "content": "F_{\\beta}^{w}" + }, + { + "bbox": [ + 130, + 389, + 482, + 509 + ], + "type": "text", + "content": " metric without utilizing additional discriminative annotation. These outcomes provide robust evidence of the effectiveness of CamoTeacher." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": "Addition unlabeled data. In Tab. 3, based on the base model SINet, we train CamoTeacher with " + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": " labeled data and check whether additional unlabeled data can further improve the performance. Compared with unsupervised and weakly-supervised methods, it shows the superiority of our method. We use the testset CAMO and NC4K as unlabeled data, and we find that adding more additional data CamoTeacher can make SINet continued to gain, such as " + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "S_{\\alpha}" + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": " metric increased by " + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "F_{\\beta}^{w}" + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": " metric increased by " + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 665 + ], + "type": "text", + "content": ". We achieved significantly better performance than unsupervised methods [31,32,46] using existing labeled data. Compared with the weakly-supervision methods [10,11,44], we do not need additional annotation for the new unlabeled data, and to a certain extent, as the more unlabeled data, the performance will get better and better. And at the same time, replacing the base model with a stronger model will get better results." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "X.Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 169, + 296, + 228 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 479, + 169 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 479, + 169 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 479, + 169 + ], + "type": "text", + "content": "Table 4: The ablation study of Pixel-wise Table 5: The ablation study of various Consistency Learning (PCL) and Instance- weight approach of Pixel-wise Consistency wise Consistency Learning (ICL) with " + }, + { + "bbox": [ + 133, + 114, + 479, + 169 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 133, + 114, + 479, + 169 + ], + "type": "text", + "content": " Learning (PCL) with " + }, + { + "bbox": [ + 133, + 114, + 479, + 169 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 133, + 114, + 479, + 169 + ], + "type": "text", + "content": " labeled data on labeled data on COD10K. COD10K. In this experiment " + }, + { + "bbox": [ + 133, + 114, + 479, + 169 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 133, + 114, + 479, + 169 + ], + "type": "text", + "content": " is set to 1/4." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 169, + 296, + 228 + ], + "lines": [ + { + "bbox": [ + 143, + 169, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 143, + 169, + 296, + 228 + ], + "type": "table", + "html": "
PCLICL\\( {S}_{\\alpha } \\uparrow \\)\\( {F}_{\\beta }^{w} \\uparrow \\)\\( {E}_{\\phi }^{m} \\uparrow \\)\\( {E}_{\\phi }^{x} \\uparrow \\)\\( {F}_{\\beta }^{m} \\uparrow \\)\\( M \\downarrow \\)
\\( L_{bce} \\)\\( {\\omega }^{pc} \\)\\( L_{iou} \\)\\( {\\omega }^{tc} \\)
✓0.6820.3700.6360.7870.4610.091
✓0.7340.5720.8170.8210.6260.057
✓✓0.7240.5010.7380.8320.5740.058
✓✓0.7420.5850.8280.8340.6390.052
✓✓✓0.7450.5860.8300.8400.6430.051
✓✓✓✓0.7460.5860.8300.8430.6450.050
", + "image_path": "f16dfdf701535b66d259cd63b66493e4640a806c5bfd88f504e85919a69802cd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 309, + 180, + 477, + 227 + ], + "blocks": [ + { + "bbox": [ + 309, + 180, + 477, + 227 + ], + "lines": [ + { + "bbox": [ + 309, + 180, + 477, + 227 + ], + "spans": [ + { + "bbox": [ + 309, + 180, + 477, + 227 + ], + "type": "table", + "html": "
weight approach\\( {S}_{\\alpha } \\uparrow {F}_{\\beta }^{w} \\uparrow {E}_{\\phi }^{m} \\uparrow {E}_{\\phi }^{x} \\uparrow {F}_{\\beta }^{m} \\uparrow M \\downarrow \\)
\\( {\\omega }^{pc} = {\\widehat{y}}^{r} \\)0.6980.3930.6840.8040.5000.110
\\( {\\omega }^{pc} = {\\begin{Vmatrix}{\\widehat{y}}^{r} - \\mu \\end{Vmatrix}}_{2}^{2} \\)0.7170.4850.7340.8220.5590.061
\\( {\\omega }^{pc} = 1 - {\\Delta }^{\\alpha } \\)0.7190.4810.7200.8270.5530.062
\\( {\\omega }^{pc} = \\left( {1 - {\\Delta }^{\\alpha }}\\right) {\\widehat{y}}^{r} \\)0.6950.4050.6960.7950.5060.114
\\( {\\omega }^{pc} = \\left( {1 - {\\Delta }^{\\alpha }}\\right) {\\begin{Vmatrix}{\\widehat{y}}^{r} - \\mu \\end{Vmatrix}}_{2}^{2}\\left\\lbrack \\begin{array}{llllllll} 0 & {724} & {0.501} & {0.738} & {0.832} & {0.574} & {0.058} \\end{array}\\right\\rbrack \\)
", + "image_path": "7ed5a63caa2c204233eefac17a40e347299c68af8c15d891aa40be673bfbdf2c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 138, + 308, + 304, + 375 + ], + "blocks": [ + { + "bbox": [ + 133, + 243, + 479, + 299 + ], + "lines": [ + { + "bbox": [ + 133, + 243, + 479, + 299 + ], + "spans": [ + { + "bbox": [ + 133, + 243, + 479, + 299 + ], + "type": "text", + "content": "Table 6: The ablation study of hyper- Table 7: The ablation study of hyperparameters " + }, + { + "bbox": [ + 133, + 243, + 479, + 299 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 133, + 243, + 479, + 299 + ], + "type": "text", + "content": " in Pixel-wise Consistency parameters " + }, + { + "bbox": [ + 133, + 243, + 479, + 299 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 133, + 243, + 479, + 299 + ], + "type": "text", + "content": " in Instance-wise Consistency Learning (PCL) with " + }, + { + "bbox": [ + 133, + 243, + 479, + 299 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 133, + 243, + 479, + 299 + ], + "type": "text", + "content": " labeled data on Learning (ICL) with " + }, + { + "bbox": [ + 133, + 243, + 479, + 299 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 133, + 243, + 479, + 299 + ], + "type": "text", + "content": " labeled data on COD10K. In this experiment " + }, + { + "bbox": [ + 133, + 243, + 479, + 299 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 133, + 243, + 479, + 299 + ], + "type": "text", + "content": " is set to 4, COD10K. In this experiment " + }, + { + "bbox": [ + 133, + 243, + 479, + 299 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 133, + 243, + 479, + 299 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 133, + 243, + 479, + 299 + ], + "type": "inline_equation", + "content": "\\lambda_u^{PC}" + }, + { + "bbox": [ + 133, + 243, + 479, + 299 + ], + "type": "text", + "content": " is set to 8, and " + }, + { + "bbox": [ + 133, + 243, + 479, + 299 + ], + "type": "inline_equation", + "content": "\\lambda_u^{IC}" + }, + { + "bbox": [ + 133, + 243, + 479, + 299 + ], + "type": "text", + "content": " is set to 0.3. " + }, + { + "bbox": [ + 133, + 243, + 479, + 299 + ], + "type": "inline_equation", + "content": "1/4, \\lambda_u^{PC}" + }, + { + "bbox": [ + 133, + 243, + 479, + 299 + ], + "type": "text", + "content": " is set to 8, and " + }, + { + "bbox": [ + 133, + 243, + 479, + 299 + ], + "type": "inline_equation", + "content": "\\lambda_u^{IC}" + }, + { + "bbox": [ + 133, + 243, + 479, + 299 + ], + "type": "text", + "content": " is set to 0.3." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 308, + 304, + 375 + ], + "lines": [ + { + "bbox": [ + 138, + 308, + 304, + 375 + ], + "spans": [ + { + "bbox": [ + 138, + 308, + 304, + 375 + ], + "type": "table", + "html": "
αSα↑Fwβ↑Eφm↑Eφx↑Fmβ↑M↓
20.7370.5650.8130.8360.6290.052
10.7390.5660.8150.8320.6270.056
1/20.7450.5750.8200.8390.6370.051
1/40.7470.5790.8250.8420.6380.051
1/80.7450.5780.8250.8380.6360.052
", + "image_path": "f20d28e82e3a14ada8b6a2eb6ba06eea93cc88436f82927f61e9826fabf03617.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 313, + 308, + 470, + 375 + ], + "blocks": [ + { + "bbox": [ + 313, + 308, + 470, + 375 + ], + "lines": [ + { + "bbox": [ + 313, + 308, + 470, + 375 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 470, + 375 + ], + "type": "table", + "html": "
βSα↑Fβw↑Eφm↑Eφx↑Fβm↑M↓
10.7430.5760.8240.8390.6360.051
20.7420.5730.8230.8370.6320.052
40.7470.5790.8250.8420.6380.051
50.7440.5780.8240.8410.6390.051
60.7450.5750.8220.8390.6350.052
", + "image_path": "ab6fa66203865e46ece1bc287bdf636138dcd5c61d51213b5d487bba22110ef8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 403, + 238, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 238, + 415 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 238, + 415 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 433, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 482, + 529 + ], + "type": "text", + "content": "Ablation on PCL and ICL. We conducted ablation experiments on different combinations of weight and loss functions for PCL and ICL. From Tab. 4, increasing the weights " + }, + { + "bbox": [ + 130, + 433, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\omega^{pc}" + }, + { + "bbox": [ + 130, + 433, + 482, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 433, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\omega^{ic}" + }, + { + "bbox": [ + 130, + 433, + 482, + 529 + ], + "type": "text", + "content": " improves the model's performance. Specifically, with " + }, + { + "bbox": [ + 130, + 433, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\omega^{pc}" + }, + { + "bbox": [ + 130, + 433, + 482, + 529 + ], + "type": "text", + "content": " results in a higher " + }, + { + "bbox": [ + 130, + 433, + 482, + 529 + ], + "type": "inline_equation", + "content": "F_{\\beta}^{w}" + }, + { + "bbox": [ + 130, + 433, + 482, + 529 + ], + "type": "text", + "content": " score, increasing from 0.370 to 0.501. Similarly, with " + }, + { + "bbox": [ + 130, + 433, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\omega^{ic}" + }, + { + "bbox": [ + 130, + 433, + 482, + 529 + ], + "type": "text", + "content": " improved the " + }, + { + "bbox": [ + 130, + 433, + 482, + 529 + ], + "type": "inline_equation", + "content": "F_{\\beta}^{w}" + }, + { + "bbox": [ + 130, + 433, + 482, + 529 + ], + "type": "text", + "content": " metric from 0.572 to 0.585. Combining these two weighted loss functions led to a significant performance improvement, with the " + }, + { + "bbox": [ + 130, + 433, + 482, + 529 + ], + "type": "inline_equation", + "content": "F_{\\beta}^{w}" + }, + { + "bbox": [ + 130, + 433, + 482, + 529 + ], + "type": "text", + "content": " score increasing from 0.370 to 0.586 (an increase of " + }, + { + "bbox": [ + 130, + 433, + 482, + 529 + ], + "type": "inline_equation", + "content": "21.6\\%" + }, + { + "bbox": [ + 130, + 433, + 482, + 529 + ], + "type": "text", + "content": "). This demonstrates the effectiveness of PCL and ICL in enhancing the model's performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 532, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 532, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 532, + 482, + 592 + ], + "type": "text", + "content": "Ablation of weight approach in PCL. The choice of method for obtaining " + }, + { + "bbox": [ + 130, + 532, + 482, + 592 + ], + "type": "inline_equation", + "content": "\\omega^{pc}" + }, + { + "bbox": [ + 130, + 532, + 482, + 592 + ], + "type": "text", + "content": " significantly affects the performance of the model. Therefore, we conducted several experiments to select the most effective method for obtaining " + }, + { + "bbox": [ + 130, + 532, + 482, + 592 + ], + "type": "inline_equation", + "content": "\\omega^{pc}" + }, + { + "bbox": [ + 130, + 532, + 482, + 592 + ], + "type": "text", + "content": ". From Tab. 5, it can be observed that the best performance is achieved when " + }, + { + "bbox": [ + 130, + 532, + 482, + 592 + ], + "type": "inline_equation", + "content": "\\omega^{pc}" + }, + { + "bbox": [ + 130, + 532, + 482, + 592 + ], + "type": "text", + "content": " is obtained using Eq. (10)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Ablation of hyper-parameters in PCL and ICL. We perform ablation experiments on the hyper-parameters associated with the PCL and ICL modules. As shown in Tab. 6, the optimal results are achieved when setting the hyper-parameter " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", which corresponds to the coefficient " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\omega^{pc}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", to " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "1/4" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". Similarly, according to Tab. 7, the best performance is observed when the hyper-parameter " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", related to the coefficient " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\omega^{ic}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", is set to 4." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "CamoTeacher" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 169, + 116, + 443, + 300 + ], + "blocks": [ + { + "bbox": [ + 169, + 116, + 443, + 300 + ], + "lines": [ + { + "bbox": [ + 169, + 116, + 443, + 300 + ], + "spans": [ + { + "bbox": [ + 169, + 116, + 443, + 300 + ], + "type": "image", + "image_path": "3642b2a7ae66b712dc7766f375dc5b354c32de919ada0deaf38816c562fb24bc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 312, + 459, + 323 + ], + "lines": [ + { + "bbox": [ + 153, + 312, + 459, + 323 + ], + "spans": [ + { + "bbox": [ + 153, + 312, + 459, + 323 + ], + "type": "text", + "content": "Fig. 5: Visualization of predictions in different proportions of labeled data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 347, + 263, + 359 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 263, + 359 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 263, + 359 + ], + "type": "text", + "content": "4.4 Qualitative Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 368, + 482, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 488 + ], + "type": "text", + "content": "Result visualization. We conducte visualizations of CamoTeacher's results under varying proportions of labeled data. The model is trained with " + }, + { + "bbox": [ + 130, + 368, + 482, + 488 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 368, + 482, + 488 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 368, + 482, + 488 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 368, + 482, + 488 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 368, + 482, + 488 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 368, + 482, + 488 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 368, + 482, + 488 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 368, + 482, + 488 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 368, + 482, + 488 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 368, + 482, + 488 + ], + "type": "text", + "content": " labeled data, and we compare its predicted masks with those produced by SINet under fully supervised conditions (as shown in Fig. 5). Notably, even with only " + }, + { + "bbox": [ + 130, + 368, + 482, + 488 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 368, + 482, + 488 + ], + "type": "text", + "content": " labeled data, CamoTeacher's predicted masks exhibit a remarkable similarity to the fully supervised SINet predictions. In fact, in certain detailed predictions, CamoTeacher even surpasses the performance of fully supervised SINet. These findings highlight the impressive capability of CamoTeacher in leveraging limited labeled data for effective camouflaged object detection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 507, + 226, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 507, + 226, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 507, + 226, + 520 + ], + "type": "text", + "content": "5 Conclusions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we present the first end-to-end semi-supervised camouflaged object detection model called CamoTeacher. To address the issue of substantial noise within pseudo-labels in semi-supervised COD, including both local noise and global noise, we've introduced a novel approach called Dual-Rotation Consistency Learning (DRCL), including Pixel-wise Consistency Learning (PCL) and Instance-wise Consistency Learning (ICL). DRCL assists the model in alleviating noise issues, effectively leveraging pseudo-labels information, so that the model can receive adequate supervision while avoiding confirmation bias. Extensive experiments validate CamoTeacher's superior performance while significantly reducing annotation expenses. We believe that CamoTeacher can serve as a strong baseline for future research on semi-supervised COD." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "X.Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": "This work was supported by National Science and Technology Major Project (No. 2022ZD0118201), the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China a (No. U21B2037, No. U22B2051, No. U23A20383, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No.2022J06001)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 243, + 197, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 243, + 197, + 255 + ], + "spans": [ + { + "bbox": [ + 133, + 243, + 197, + 255 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 269, + 480, + 664 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 269, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 480, + 300 + ], + "type": "text", + "content": "1. Achanta, R., Hemami, S., Estrada, F., Susstrunk, S.: Frequency-tuned salient region detection. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 1597-1604. IEEE (2009)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 303, + 480, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 303, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 303, + 480, + 334 + ], + "type": "text", + "content": "2. Chen, H., Tao, R., Fan, Y., Wang, Y., Wang, J., Schiele, B., Xie, X., Raj, B., Savvides, M.: Softmatch: Addressing the quantity-quality trade-off in semi-supervised learning. arXiv preprint arXiv:2301.10921 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 335, + 480, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 480, + 367 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 480, + 367 + ], + "type": "text", + "content": "3. Chen, X., Yuan, Y., Zeng, G., Wang, J.: Semi-supervised semantic segmentation with cross pseudo supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2613-2622 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 369, + 480, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 369, + 480, + 400 + ], + "spans": [ + { + "bbox": [ + 138, + 369, + 480, + 400 + ], + "type": "text", + "content": "4. Fan, D.P., Cheng, M.M., Liu, Y., Li, T., Borji, A.: Structure-measure: A new way to evaluate foreground maps. In: Proceedings of the IEEE international conference on computer vision. pp. 4548-4557 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 402, + 480, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 402, + 480, + 433 + ], + "spans": [ + { + "bbox": [ + 138, + 402, + 480, + 433 + ], + "type": "text", + "content": "5. Fan, D.P., Gong, C., Cao, Y., Ren, B., Cheng, M.M., Borji, A.: Enhanced-alignment measure for binary foreground map evaluation. arXiv preprint arXiv:1805.10421 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 434, + 480, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 480, + 455 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 480, + 455 + ], + "type": "text", + "content": "6. Fan, D.P., Ji, G.P., Cheng, M.M., Shao, L.: Concealed object detection. IEEE transactions on pattern analysis and machine intelligence 44(10), 6024-6042 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 456, + 480, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 480, + 488 + ], + "type": "text", + "content": "7. Fan, D.P., Ji, G.P., Sun, G., Cheng, M.M., Shen, J., Shao, L.: Camouflaged object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2777-2787 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 490, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 490, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 490, + 480, + 510 + ], + "type": "text", + "content": "8. Grandvalet, Y., Bengio, Y.: Semi-supervised learning by entropy minimization. Advances in neural information processing systems 17 (2004)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 512, + 480, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 554 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 554 + ], + "type": "text", + "content": "9. He, C., Li, K., Zhang, Y., Tang, L., Zhang, Y., Guo, Z., Li, X.: Camouflaged object detection with feature decomposition and edge reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22046-22055 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 555, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 555, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 134, + 555, + 480, + 597 + ], + "type": "text", + "content": "10. He, C., Li, K., Zhang, Y., Xu, G., Tang, L., Zhang, Y., Guo, Z., Li, X.: Weakly-supervised concealed object segmentation with sam-based pseudo labeling and multi-scale feature grouping. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 600, + 480, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 480, + 631 + ], + "type": "text", + "content": "11. He, R., Dong, Q., Lin, J., Lau, R.W.: Weakly-supervised camouflaged object detection with scribble annotations. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 781-789 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 632, + 480, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 480, + 664 + ], + "type": "text", + "content": "12. Hu, J., Lin, J., Cai, W., Gong, S.: Relax image-specific prompt requirement in sam: A single generic prompt for segmenting camouflaged objects. arXiv preprint arXiv:2312.07374 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "CamoTeacher" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "13. Hu, J., Chen, C., Cao, L., Zhang, S., Shu, A., Jiang, G., Ji, R.: Pseudo-label alignment for semi-supervised instance segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16337-16347 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 193 + ], + "type": "text", + "content": "14. Huang, Z., Dai, H., Xiang, T.Z., Wang, S., Chen, H.X., Qin, J., Xiong, H.: Feature shrinkage pyramid for camouflaged object detection with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5557-5566 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 193, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 193, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 481, + 214 + ], + "type": "text", + "content": "15. Hung, W.C., Tsai, Y.H., Liou, Y.T., Lin, Y.Y., Yang, M.H.: Adversarial learning for semi-supervised semantic segmentation. arXiv preprint arXiv:1802.07934 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 481, + 246 + ], + "type": "text", + "content": "16. Ji, G.P., Fan, D.P., Chou, Y.C., Dai, D., Liniger, A., Van Gool, L.: Deep gradient learning for efficient camouflaged object detection. Machine Intelligence Research 20(1), 92-108 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 246, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 246, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 133, + 246, + 481, + 278 + ], + "type": "text", + "content": "17. Jia, Q., Yao, S., Liu, Y., Fan, X., Liu, R., Luo, Z.: Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4713-4722 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 278, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 278, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 133, + 278, + 481, + 310 + ], + "type": "text", + "content": "18. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 310, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 310, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 133, + 310, + 481, + 354 + ], + "type": "text", + "content": "19. Lai, X., Tian, Z., Jiang, L., Liu, S., Zhao, H., Wang, L., Jia, J.: Semi-supervised semantic segmentation with directional context-aware consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1205-1214 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 354, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 354, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 133, + 354, + 481, + 374 + ], + "type": "text", + "content": "20. Laine, S., Aila, T.: Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 133, + 374, + 481, + 407 + ], + "type": "text", + "content": "21. Le, T.N., Nguyen, T.V., Nie, Z., Tran, M.T., Sugimoto, A.: Anabranch network for camouflaged object segmentation. Computer vision and image understanding 184, 45-56 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 407, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 407, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 133, + 407, + 481, + 439 + ], + "type": "text", + "content": "22. Lee, D.H., et al.: Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In: Workshop on challenges in representation learning, ICML. vol. 3, p. 896. Atlanta (2013)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 439, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 439, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 133, + 439, + 481, + 482 + ], + "type": "text", + "content": "23. Liu, Y., Tian, Y., Chen, Y., Liu, F., Belagiannis, V., Carneiro, G.: Perturbed and strict mean teachers for semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4258-4267 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 482, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 482, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 133, + 482, + 481, + 514 + ], + "type": "text", + "content": "24. Liu, Z., Zhang, Z., Tan, Y., Wu, W.: Boosting camouflaged object detection with dual-task interactive transformer. In: 2022 26th International Conference on Pattern Recognition (ICPR). pp. 140-146. IEEE (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 514, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 514, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 133, + 514, + 481, + 557 + ], + "type": "text", + "content": "25. Lv, Y., Zhang, J., Dai, Y., Li, A., Liu, B., Barnes, N., Fan, D.P.: Simultaneously localize, segment and rank the camouflaged objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11591-11601 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 557, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 557, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 133, + 557, + 481, + 590 + ], + "type": "text", + "content": "26. Margolin, R., Zelnik-Manor, L., Tal, A.: How to evaluate foreground maps? In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 248-255 (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 590, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 590, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 590, + 481, + 632 + ], + "type": "text", + "content": "27. Mi, P., Lin, J., Zhou, Y., Shen, Y., Luo, G., Sun, X., Cao, L., Fu, R., Xu, Q., Ji, R.: Active teacher for semi-supervised object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14482-14491 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 481, + 666 + ], + "type": "text", + "content": "28. Oliver, A., Odena, A., Raffel, C.A., Cubuk, E.D., Goodfellow, I.: Realistic evaluation of deep semi-supervised learning algorithms. Advances in neural information processing systems 31 (2018)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "X.Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "29. Pang, Y., Zhao, X., Xiang, T.Z., Zhang, L., Lu, H.: Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 2160-2170 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "type": "text", + "content": "30. Roelofs, B., Berthelot, D., Sohn, K., Carlini, N., Kurakin, A.: Adamatch: a unified approach to semi-supervised learning and domain adaptation. In: Proc. of the Int. Conf. on Learning Representations (ICLR) (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 181, + 482, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 181, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 181, + 482, + 214 + ], + "type": "text", + "content": "31. Shin, G., Albanie, S., Xie, W.: Unsupervised salient object detection with spectral cluster voting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3971-3980 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 214, + 482, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 256 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 256 + ], + "type": "text", + "content": "32. Simeoni, O., Sekkat, C., Puy, G., Vobecký, A., Zablocki, É., Pérez, P.: Unsupervised object localization: Observing the background to discover objects. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3176-3186 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 256, + 482, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 256, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 482, + 289 + ], + "type": "text", + "content": "33. Skurowski, P., Abdulameer, H., Błaszczyk, J., Depta, T., Kornacki, A., Koziel, P.: Animal camouflage analysis: Chameleon database. Unpublished manuscript 2(6), 7 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 289, + 482, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 331 + ], + "type": "text", + "content": "34. Sohn, K., Berthelot, D., Carlini, N., Zhang, Z., Zhang, H., Raffel, C.A., Cubuk, E.D., Kurakin, A., Li, C.L.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems 33, 596-608 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 332, + 482, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 365 + ], + "type": "text", + "content": "35. Sohn, K., Zhang, Z., Li, C.L., Zhang, H., Lee, C.Y., Pfister, T.: A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 365, + 482, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 396 + ], + "type": "text", + "content": "36. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 396, + 482, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 439 + ], + "type": "text", + "content": "37. Wang, Y., Wang, H., Shen, Y., Fei, J., Li, W., Jin, G., Wu, L., Zhao, R., Le, X.: Semi-supervised semantic segmentation using unreliable pseudo-labels. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4248-4257 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 440, + 482, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 482, + 472 + ], + "type": "text", + "content": "38. Wang, Z., Li, Y., Guo, Y., Fang, L., Wang, S.: Data-uncertainty guided multi-phase learning for semi-supervised object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4568-4577 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 472, + 482, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 504 + ], + "type": "text", + "content": "39. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 504, + 482, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 482, + 536 + ], + "type": "text", + "content": "40. Xu, M., Zhang, Z., Hu, H., Wang, J., Wang, L., Wei, F., Bai, X., Liu, Z.: End-to-end semi-supervised object detection with soft teacher. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3060-3069 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 568 + ], + "type": "text", + "content": "41. Yang, L., Qi, L., Feng, L., Zhang, W., Shi, Y.: Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7236-7246 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 601 + ], + "type": "text", + "content": "42. Yang, L., Zhuo, W., Qi, L., Shi, Y., Gao, Y.: St++: Make self-training work better for semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4268-4277 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 601, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 482, + 632 + ], + "type": "text", + "content": "43. Yin, B., Zhang, X., Hou, Q., Sun, B.Y., Fan, D.P., Van Gool, L.: Camoformer: Masked separable attention for camouflaged object detection. arXiv preprint arXiv:2212.06570 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "type": "text", + "content": "44. Yu, S., Zhang, B., Xiao, J., Lim, E.G.: Structure-consistent weakly supervised salient object detection with local saliency coherence. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 3234-3242 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 391, + 91, + 447, + 100 + ], + "type": "text", + "content": "CamoTeacher" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 281 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "45. Zhang, M., Xu, S., Piao, Y., Shi, D., Lin, S., Lu, H.: Preynet: Preying on camouflaged objects. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 5323-5332 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 183 + ], + "type": "text", + "content": "46. Zhang, Y., Wu, C.: Unsupervised camouflaged object segmentation as domain adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4334-4344 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "type": "text", + "content": "47. Zhao, Z., Yang, L., Long, S., Pi, J., Zhou, L., Wang, J.: Augmentation matters: A simple-yet-effective approach to semi-supervised semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11350-11359 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "text", + "content": "48. Zhao, Z., Zhou, L., Wang, L., Shi, Y., Gao, Y.: Lassl: Label-guided self-training for semi-supervised learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 9208-9216 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "text", + "content": "49. Zhou, Z.H., Zhou, Z.H.: Semi-supervised learning. Machine Learning pp. 315-341 (2021)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "X.Lai et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Can OOD Object Detectors Learn from Foundation Models_/6d5daec6-da83-48aa-809b-d1c5b6b09141_content_list.json b/2024/Can OOD Object Detectors Learn from Foundation Models_/6d5daec6-da83-48aa-809b-d1c5b6b09141_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..197157d29355415f93313ba5fd2edf6a7c97e37a --- /dev/null +++ b/2024/Can OOD Object Detectors Learn from Foundation Models_/6d5daec6-da83-48aa-809b-d1c5b6b09141_content_list.json @@ -0,0 +1,1575 @@ +[ + { + "type": "text", + "text": "Can OOD Object Detectors Learn from Foundation Models?", + "text_level": 1, + "bbox": [ + 274, + 140, + 727, + 183 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiahui Liu, Xin Wen, Shizhen Zhao, Yingxian Chen, and Xiaojuan Qi*", + "bbox": [ + 246, + 212, + 754, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The University of Hong Kong {liujh,wenxin,zhaosz,chenyx,xjqi}@eee.hku.hk", + "bbox": [ + 330, + 239, + 673, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Out-of-distribution (OOD) object detection is a challenging task due to the absence of open-set OOD data. Inspired by recent advancements in text-to-image generative models, such as Stable Diffusion, we study the potential of generative models trained on large-scale open-set data to synthesize OOD samples, thereby enhancing OOD object detection. We introduce SyncOOD, a simple data curation method that capitalizes on the capabilities of large foundation models to automatically extract meaningful OOD data from text-to-image generative models. This offers the model access to open-world knowledge encapsulated within off-the-shelf foundation models. The synthetic OOD samples are then employed to augment the training of a lightweight, plug-and-play OOD detector, thus effectively optimizing the in-distribution (ID)/OOD decision boundaries. Extensive experiments across multiple benchmarks demonstrate that SyncOOD significantly outperforms existing methods, establishing new state-of-the-art performance with minimal synthetic data usage. The project is available at https://github.com/CVMI-Lab/SyncOOD.", + "bbox": [ + 261, + 308, + 738, + 544 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: OOD object detection $\\cdot$ Synthetic data $\\cdot$ Open-world data", + "bbox": [ + 261, + 558, + 736, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 599, + 375, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modern object detectors, trained on closed-set data, have achieved remarkable success. However, they often incorrectly and confidently classify out-of-distribution (OOD) object categories as in-distribution (ID) categories in open-world applications [10], raising concerns about their reliability for deployment. To enhance the trustworthiness of object detection, researchers have studied the OOD object detection task, which aims to identify and flag unknown or novel objects as distinct from ID ones [3,42].", + "bbox": [ + 212, + 633, + 787, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The vulnerability of these models to OOD samples stems from their lack of awareness of the unknown open data distribution during training. Consequently, synthesizing OOD samples for model learning has emerged as a major research direction for this task [15, 59-61]. Most existing studies [15, 60, 61] concentrate on generating OOD objects in the latent space of an object detection model", + "bbox": [ + 212, + 739, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* corresponding author", + "bbox": [ + 217, + 825, + 377, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7e55f29f4617ab4ee325d9ec919e7a071bb49f89a2bd311b222457a5741a3ab2.jpg", + "image_caption": [ + "Fig. 1: Our pipeline replaces ID objects with semantic-novel yet visual-similar objects for scene-level OOD object synthesis. Middle left: The concepts are imagined by an LLM to ensure semantic separability and rationality, and reformed as text prompts for controllable in-painting using Stable Diffusion. Middle right: During training, only visually similar OOD objects are adopted based on instance-level feature similarity to the original object. A lightweight binary classifier is optimized for OOD detection, and other parts of the detector are kept unchanged." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 785, + 414 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "trained on ID data. These synthesized samples are then used to optimize the decision boundary between ID and OOD data. Alternative approaches include directly synthesizing images by injecting adversarial noise [59] or identifying OOD instances from video data [14]. While these methods have yielded promising results, they remain limited to a closed-set setting, where the latent space for synthesizing outliers or the data is derived from a closed-set data distribution. Consequently, they may be biased towards the ID dataset, leading to suboptimal performance. Besides, effectively handling the unknown may appear unattainable when the unknown is never fully exploitable. Beyond that, is it possible to learn them from massive open-world data knowledge condensed in foundation models?", + "bbox": [ + 212, + 534, + 787, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we investigate text-to-image generation models trained on a large amount of open-set data, which have demonstrated a superior ability to capture the distribution of data across a wide range of visual concepts, in order to synthesize novel data samples for enhancing OOD detection. Nonetheless, automatically extracting meaningful data from generative models for OOD object detection remains challenging due to the extensive vocabulary space to be explored, the complex scene-level synthesis problem, the need for object-level annotations, and potential distractions from contextual cues.", + "bbox": [ + 212, + 686, + 787, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce SyncOOD, an automatic data curation process that leverages foundation models as tools to harvest meaningful data from text-to-image gen-", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "eration models for OOD object detection (see Figure 1). The process is based on two key observations: 1) Hard OOD samples that are close to the ID data contribute more to learning a better OOD detector, and 2) Context may become a distracting cue for OOD object detection tasks, leading to bias towards contexts. With these observations in mind, the outlier synthesis process is formulated as box-conditioned image inpainting, and driven by Stable Diffusion [49] for high-quality controllable editing. The concepts to replace with are imagined by a Large Language Model (LLM) [1] with the aim of semantic novelty, $^{1}$ and the associated bounding box is further refined with SAM [30]. Automated by foundation models, this data collection pipeline requires minimal human labor, while producing high-quality OOD data.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In comparison to existing methods for OOD object detection, our core insight is to broaden the model's exposure to a more extensive range of open-set data and circumvent dataset biases by tapping into the open-world knowledge found in off-the-shelf foundation models. Utilizing generative models also provides us with control over the context of synthesized images and the data distribution. Our comprehensive experiments demonstrate superior performance, emphasizing the untapped potential of text-image-generation models in the context of OOD object detection. Our key contributions are summarized as follows:", + "bbox": [ + 212, + 313, + 787, + 434 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We investigate and unlock the potential of text-to-image generative models trained on large-scale open-set data for synthesizing OOD objects in object detection tasks.", + "- We introduce an automated data curation process for obtaining controllable, annotated scene-level synthetic OOD images for OOD object detection, which utilizes LLMs for novel concept discovery and visual foundation models for data annotation and filtering.", + "- We discover that maintaining ID/OOD image context consistency and obtaining more accurate OOD annotation bounding boxes are crucial for synthesized data to be effective in OOD object detection.", + "- Comprehensive experiments on multiple benchmarks demonstrate the effectiveness of our method, as we significantly outperform existing state-of-the-art approaches while using minimal synthetic data." + ], + "bbox": [ + 225, + 445, + 784, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 662, + 387, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "OOD Object Detection For detecting OOD objects in scene-level images, unlike earlier works that constrain ID samples to a hypothetical distribution [12], it has become a recent trend to explicitly synthesize the outlier data, and incorporate them into training to adjust models' decision boundaries. However, due to the complexity of scene-level images, all previous works bypassed photorealistic outlier synthesis in pixel space, and worked on generating outliers from the model's latent space [15, 60, 61], adversarial attack [59], or utilizing video data in the wild [14]. In the former, outliers can be sampled from the latent", + "bbox": [ + 212, + 694, + 787, + 816 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Can OOD Object Detectors Learn from Foundation Models?", + "bbox": [ + 325, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1 Concepts overlapping with the test data are removed to avoid information leakage.", + "bbox": [ + 217, + 824, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "space using a simple Gaussian prior [15], or more advanced generative models like VAE [61] or diffusion model [60]. Yet their upperbound are commonly limited by the quality of the latent space, and synthesized samples lack interpretability. Despite this, adversarial samples [59] lack semantic diversity, and auxiliary pseudo supervision from videos [14] introduces additional requirements to the setting. Unlike all above, our method applies LLM for OOD concept sampling, and Stable Diffusion for controllable image editing. This decouples the sampling and generation processes, elevates both parts to an unprecedented level, and achieves photo-realistic scene-level OOD image synthesis for the first time.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Open-world Object Detection The fact that real-world applications require object detectors the ability to tackle open categories is also considered in open-world object detection. The focus of this field includes generalization to domain shifts [57], incremental learning of novel classes [27, 39, 43], and zero-shot classification of open-vocabularies [17, 46]. Meanwhile, some works [18, 28] require to distinguish known objects well, be able to detect unknown objects, and finally be able to incrementally learn new objects. These works provide different perspectives on facing the challenges of real-world applications, which complement OOD object detection as a joint effort, but are out of the scope of this paper.", + "bbox": [ + 212, + 297, + 787, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "OOD Image Classification Earlier paradigms for OOD image classification either post hoc adjust models' confidence score at the testing phase, or apply regularization at models' training phase. The former line mainly focuses on the design of score functions, including confidence-based [2, 22, 35], energy-based [38, 56, 62], distance-based [34, 47, 50, 51], gradient-based [26], and approximating Bayesian [7, 10, 19, 40, 41] methods. The latter line of work includes regularizing models to produce lower confidence [23, 33], higher energy [29, 38], or directly shaping latent representations [12]. While outlier synthesis has shown to be effective by [15, 54], these are still generated in the latent space, and a parallel line study utilizing natural images [11, 23, 29] from the wild. Recently, photo-realistic outlier synthesis was first achieved by [13] with help from a text-conditioned diffusion model [49]. However, it is not readily applicable for object detection due to the complexity of scene-level images and the requirement for object-level annotations. In contrast, our work studies under the detection setting and requires outlier synthesis at scene level.", + "bbox": [ + 212, + 446, + 787, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Foundation Models The evolution of large language models (LLMs) began with training on web-scale datasets [9,45], leading to increasingly powerful foundation models [4, 6] capable of harnessing vast open-world data. Notable advancements include models [1] that interact with users and perform complex tasks like question answering, significantly broadening access to global knowledge. In image generation, diffusion models [24, 49] offer robust capabilities in synthesizing realistic content for applications such as image synthesis and inpainting. Additionally, segmentation foundation models like SAM [30] represent a leap forward in precise image segmentation, benefiting from extensive data training. Foundation models provide diverse data, which provides unlimited po", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/599d0ae175dbc499dcd032372e9eec73fbf99c0109a180a9b04be407c86659eb.jpg", + "image_caption": [ + "Fig. 2: Detailed illustration of our outlier synthesis pipeline. It comprises (a) Instructing an LLM to imagine semantic-novel concepts given ID objects, (b) Editing the selected regions to the expected concepts via prompt-conditioned image inpainting using Stable Diffusion, and (c) Refining the bounding boxes of edited objects using SAM." + ], + "image_footnote": [], + "bbox": [ + 218, + 146, + 781, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tential for learning open-world knowledge from these models [55], where the effectiveness of data [21, 53, 63] is also crucial to the downstreaming tasks.", + "bbox": [ + 212, + 378, + 787, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 431, + 330, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Preliminary For OOD object detection, the training set contains only ID scene-level images $\\mathbf{x}^{\\mathrm{id}}$ with ID objects, annotation bounding boxes $\\mathbf{b}^{\\mathrm{id}}$ , and semantic labels $y$ , denoted as $\\mathcal{D}_{\\mathrm{id}} = \\left\\{(\\mathbf{x}^{\\mathrm{id}}, \\mathbf{b}^{\\mathrm{id}}, \\mathbf{y}^{\\mathrm{id}})\\right\\}$ . The labels of ID objects always belong to a close set with $K$ categories, denoted as $\\mathbf{y}^{\\mathrm{id}} \\in \\mathcal{Y}_{\\mathrm{id}}$ and $\\mathcal{Y}_{\\mathrm{id}} = \\{y_1^{\\mathrm{id}}, y_2^{\\mathrm{id}}, \\dots, y_K^{\\mathrm{id}}\\}$ . During inference, for each proposed object from an input scene-level image, it is required to identify whether its category belongs to $\\mathcal{Y}_{\\mathrm{id}}$ or not.", + "bbox": [ + 212, + 462, + 787, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overview As illustrated in Fig. 1, our outlier synthesis pipeline consists of (1) synthesizing a set of effective photo-realistic scene-level OOD images $\\mathbf{x}^{\\mathrm{edit}}$ , denoted as $\\mathcal{D}_{\\mathrm{edit}} = \\left\\{(\\mathbf{x}^{\\mathrm{edit}},\\mathbf{b}^{\\mathrm{edit}})\\right\\}$ , which contains novel objects and corresponding annotation boxes $\\mathbf{b}^{\\mathrm{edit}}$ based on region-level editing from $\\mathcal{D}_{\\mathrm{id}}$ in a fully automated, labor-free way; and (2) select and use the efficient synthetic data to provide pseudo-OOD supervisions for training OOD object detector together with the ID samples in the training set. Further design of the pipeline requires answering the following questions: (1) how to distill the open-set knowledge embedded in foundation models to scene-level OOD data and (2) how to utilize the synthesized data to regularize the decision boundary and facilitate OOD object detection. We discuss them accordingly in Sec. 3.1 and Sec. 3.2.", + "bbox": [ + 212, + 571, + 787, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Synthesizing Semantic-novel Objects in Scene Images", + "text_level": 1, + "bbox": [ + 212, + 768, + 702, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Imagining Novel Concepts from ID objects As shown in Fig. 2(a), based on the ID labels $\\mathcal{V}_{\\mathrm{id}}$ in training set $\\mathcal{D}_{\\mathrm{id}}$ , we consider that novel concepts that are different from ID categories can be potential candidates for generating OOD", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Can OOD Object Detectors Learn from Foundation Models?", + "bbox": [ + 325, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "objects. The next is to discover novel concepts that offer hard OOD samples sharing high visual similarity with ID samples and being contextually compatible with the scene context for object detection. Rather than relying on human labor to investigate all potential candidates, we leverage the vast knowledge and reasoning capabilities of LLM, GPT-4 [1] to check the visual similarity and contextual compatibility. This allows us to associate ID objects and promote the conceptualization of possible novel objects to replace existing ID objects through the use of a prompt with in-context examples [4] as:", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here is a list containing several objects $\\mathcal{V}_{\\mathrm{id}}$ . Now, if I provide you an object name, you should return to me objects that are similar to the usage scenario and volume of the provided object but are not in the previous object list. For example, if I give you the word: person, you should respond and only respond: 'mannequin', 'sculpture', 'scarecrows', 'doll', 'puppet'.", + "bbox": [ + 240, + 284, + 764, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With its robust logical foundation and rich knowledge, the LLM envisions a collection of novel objects for each ID object label, denoted as $\\mathcal{Y}_{\\mathrm{novel}}$ , while maintaining the semantic separability between imagined objects and ID objects. We empirically find one in-context example that is sufficient for us to discover novel concepts. For each ID label $i$ , we discover $M$ novel concepts using LLM $\\mathbf{y}_i^{\\mathrm{novel}}$ of $M$ concepts.", + "bbox": [ + 212, + 376, + 787, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Editing Objects on Selected Regions With the discovered novel concepts $\\mathcal{V}_{\\mathrm{novel}} = \\{\\mathbf{y}_1^{\\mathrm{novel}},\\mathbf{y}_2^{\\mathrm{novel}},\\dots,\\mathbf{y}_K^{\\mathrm{novel}}\\}$ , the next step is to use them as prompts for the text-to-image generation model to generate an image. To generate a new image with novel concepts $y_{j}\\in \\mathbf{y}_{i}^{\\mathrm{novel}}$ , we choose to replace existing ID objects in existing images with label $y_{i}^{\\mathrm{id}}$ instead of finding new locations or generating one image from scratch. By doing so, we can ensure context compatibility and eliminate distractions from the scene context as it is preserved. As illustrated in Fig. 2 (b), we use Stable-Diffusion-Inpainting [49], denoted as $\\mathrm{SDI}(\\cdot)$ , to perform region-level editing on ID images. The ID object is denoted as $\\mathbf{x}^{\\mathrm{id}}$ , with its corresponding annotation box $\\mathbf{b}^{\\mathrm{id}}$ serving as the editing mask, and the associated imagined novel concept $\\mathbf{y}^{\\mathrm{novel}}$ are provided as inputs to the SDI, which is one of the most successful models for conditional image generation and editing. Thus, an edited image $\\mathbf{x}^{\\mathrm{edit}}$ containing a novel object is obtained as:", + "bbox": [ + 212, + 481, + 787, + 679 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} ^ {\\text {e d i t}} = \\operatorname {S D I} \\left(\\mathbf {x} ^ {\\text {i d}}, \\mathbf {b} ^ {\\text {i d}}, \\mathbf {y} ^ {\\text {n o v e l}}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 693, + 784, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Refining Annotation Boxes of Novel Objects Due to the randomness in diffusion models, the attributes of edited objects, such as their quality, volume, and localization, may not match the original object box. To address this issue, as depicted in Fig. 2(c), we design an efficient and effective refiner based on SAM [30] to obtain refined accurate bounding boxes on novel objects. First, for an edited image $\\mathbf{x}^{\\mathrm{edit}}$ with the editing mask $\\mathbf{b}^{\\mathrm{id}}$ , we use a padding area extended from $\\mathbf{b}^{\\mathrm{id}}$ as the prompt and employ SAM to output the instance mask with", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "highest confidence $\\mathbf{m}^{\\mathrm{SAM}}$ for the novel object in the area:", + "bbox": [ + 214, + 145, + 630, + 162 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {m} ^ {\\text {S A M}} = \\operatorname {S A M} \\left(\\mathbf {x} ^ {\\text {e d i t}}; \\text {p a d d i n g} \\left(\\mathbf {b} ^ {\\text {i d}}, e\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 166, + 785, + 185 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $e$ represents the range of padding. Then, we convert obtained masks $\\mathbf{m}^{\\mathrm{SAM}}$ to boxes $\\mathbf{b}^{\\mathrm{SAM}}$ , and calculate IoU between $\\mathbf{b}^{\\mathrm{SAM}}$ and the corresponding $\\mathbf{b}^{\\mathrm{id}}$ to filter out novel objects that vary highly in scale:", + "bbox": [ + 212, + 189, + 785, + 236 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mathbf {b} ^ {\\text {e d i t}} \\right\\} = \\left\\{\\mathbf {b} ^ {\\text {S A M}} \\mid \\operatorname {I o U} \\left(\\mathbf {b} ^ {\\text {S A M}}, \\mathbf {b} ^ {\\text {i d}}\\right) > \\gamma \\right\\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 241, + 785, + 267 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\gamma$ denotes a threshold value on IoU. It ensures a high enough recall rate to rule out the instability of Stable Diffusion and SAM and uncontrollable localization of the edited objects. Thus we obtain the synthetic outlier data $\\mathcal{D}_{\\mathrm{edit}}$ as illustrated in Fig. 1.", + "bbox": [ + 214, + 270, + 787, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Mining Hard OOD Samples and Model Training", + "text_level": 1, + "bbox": [ + 214, + 349, + 661, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mining Hard OOD Objects with High Visual Similarities for Training We consider the novel objects that are most likely to be confused with the corresponding ID objects by the object detector as the most effective ones. We thus aim to find synthetic OOD samples that are most easily confused as ID to participate in training based on pairwise similarity in the latent space of the pre-trained object detector. For each novel object with bounding box $\\mathbf{b}^{\\mathrm{edit}}$ in the synthetic data $\\mathcal{D}_{\\mathrm{edit}}$ , we construct it with the corresponding original ID object with its bounding box as a pair: $\\left\\{(\\mathbf{b}^{\\mathrm{edit}}, \\mathbf{x}^{\\mathrm{edit}}), (\\mathbf{b}^{\\mathrm{id}}, \\mathbf{x}^{\\mathrm{id}})\\right\\}$ . For an off-the-shelf object detector, denoted by $\\mathcal{F}_{\\mathrm{det}}$ , we extract latent features, $\\mathbf{z}^{\\mathrm{edit}}$ and $\\mathbf{z}^{\\mathrm{id}}$ , for each pair:", + "bbox": [ + 212, + 369, + 787, + 527 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} ^ {\\text {e d i t}}, \\mathbf {z} ^ {\\text {i d}} = \\mathcal {F} _ {\\det } \\left(\\mathbf {b} ^ {\\text {e d i t}}; \\mathbf {x} ^ {\\text {e d i t}}\\right), \\mathcal {F} _ {\\det } \\left(\\mathbf {b} ^ {\\text {i d}}; \\mathbf {x} ^ {\\text {i d}}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 527, + 785, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The most effective novel objects are those with visual patterns that can be easily mistaken for their corresponding ID objects by an object detector. Therefore, we filter these novel objects based on their similarity to provide pseudo-OD supervision:", + "bbox": [ + 212, + 547, + 787, + 609 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mathbf {z} ^ {\\mathrm {o o d}} \\right\\} = \\left\\{\\mathbf {z} ^ {\\mathrm {e d i t}} \\big | \\epsilon_ {l o w} < \\mathrm {s i m} (\\mathbf {z} ^ {\\mathrm {e d i t}}, \\mathbf {z} ^ {\\mathrm {i d}}) < \\epsilon_ {u p} \\right\\}, \\qquad \\qquad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 614, + 785, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the similarities are computed between latent object features of edit-ID pairs. Here $\\mathrm{sim}(\\cdot)$ denotes cosine similarity calculating and $\\epsilon_{low},\\epsilon_{up}$ stand for the lower/upper similarity thresholds.", + "bbox": [ + 212, + 637, + 787, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Optimizing ID/OOD Decision Boundary with Synthetic Samples Once we have obtained the ID and synthetic OOD objects, we employ a lightweight MLP, denoted as $\\mathcal{F}_{\\mathrm{ood}}$ , as the OOD detector optimized with a bi-classify loss:", + "bbox": [ + 214, + 689, + 787, + 737 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {o o d}} = \\mathbb {E} _ {\\mathbf {z} \\sim \\mathbf {z} ^ {\\mathrm {i d}}} \\left[ - \\log \\frac {1}{1 + \\exp^ {- \\mathcal {F} _ {\\mathrm {o o d}} (\\mathbf {z})}} \\right] + \\mathbb {E} _ {\\mathbf {z} \\sim \\mathbf {z} ^ {\\mathrm {o o d}}} \\left[ - \\log \\frac {\\exp^ {- \\mathcal {F} _ {\\mathrm {o o d}} (\\mathbf {z})}}{1 + \\exp^ {- \\mathcal {F} _ {\\mathrm {o o d}} (\\mathbf {z})}} \\right]. (6)\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 741, + 785, + 775 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The aforementioned design ensures both semantic separability and pattern similarity for the chosen synthetic samples. As a result, our proposed method elegantly optimizes the decision boundary using only a limited number of samples. It is further demonstrated and validated in the following experiments.", + "bbox": [ + 214, + 779, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Can OOD Object Detectors Learn from Foundation Models?", + "bbox": [ + 325, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets Following OOD object detection setting [15], we use the PASCAL-VOC [16] and Berkeley DeepDrive (BDD-100K) [65] as the ID training datasets, which consist of 20 and 10 ID categories, respectively. Meanwhile, we evaluate the performance of our approach on two OOD datasets: MS-COCO [36] and OpenImages [32] respectively. Categories from the OOD datasets that overlapped with the ID datasets are removed to guarantee the absence of ID categories. We report ID categories, OOD categories, and texts used in driving image synthesis in the supplementary material.", + "bbox": [ + 212, + 172, + 787, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Metrics We primarily focus on reporting the FPR95 which represents the false positive rate of OOD samples when the true positive rate of ID samples is at $95\\%$ and lower is better. FPR95 is widely used to assess the OOD object detection performance [12, 15, 59-61]. Additionally, we present the Area Under the Receiver Operating Characteristic Curve (AUROC, higher is better) that is widely utilized to evaluate binary classification problems. Since we only train a plugin MLP on top of existing object detectors for OOD detection, ID performance, e.g., mean Average Precision (mAP) is unchanged and thus omitted.", + "bbox": [ + 212, + 304, + 787, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation Details For our synthetic data, in order to maintain the experimental setting of OOD object detection and avoid leaking prior knowledge of foundation models, we remove all imagined novel objects that have the same or similar meaning as the ground truth OOD data categories. For model training, we follow the architectures of the baseline method [15, 59] to use a Faster R-CNN [48] as the base object detector with ResNet-50 [20] firstly. Then we trained a simple and lightweight 3-layer MLP. We follow [59] to extract multi-level features as training samples. ID samples are extracted from ID images and OOD samples are extracted from selected synthesized images with OOD bounding boxes. For training on the PASCAL-VOC dataset, we employ a learning rate of 1e-4, while for the BDD-100K dataset, we use a learning rate of 5e-5. Both training processes utilize a momentum of 0.9, a dropout rate of 0.5, and a batch size of 32. All training is conducted on GeForce RTX 3090 GPUs.", + "bbox": [ + 212, + 434, + 787, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Main Results on OOD Object Detection", + "text_level": 1, + "bbox": [ + 215, + 650, + 594, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate the performance of the proposed method on different challenging benchmarks and obtain notable results (see Tab. 1). As the first work to introduce synthetic scene-level natural images as OOD samples, we incorporate our data-centric method to two off-the-shelf object detectors [15,48], achieving new state-of-the-art performance in OOD object detection.", + "bbox": [ + 212, + 672, + 787, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Compared with previous methods, we present comprehensive and substantial performance improvements on FPR95. The encouraging outcomes clearly show that our synthetic data offers superior OOD supervision and are well-suited for forming a precise decision boundary between ID and OOD samples as illustrated in Fig. 1, which significantly reduces the interference caused by contextual information when optimizing the decision boundary.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a3bec877584ded66721edb2a10216232d27424e350049a8aba4ce76d6dbca775.jpg", + "table_caption": [ + "Table 1: Comparing on varied ID (PASCAL VOC [16], BDD-100K [65]) and OOD (MS-COCO [36], OpenImages [32]) datasets, our method significantly outperforms other methods on different metrics and achieves SOTA performance on OOD object detection. Our method is validated on two different existing object detectors, Faster R-CNN [48] and VOS [15] (denoted as Faster R-CNN + Ours and VOS + Ours respectively). (Top results are shown in bold.)" + ], + "table_footnote": [], + "table_body": "
MethodID:PASCAL-VOCID:BDD-100K
MS-COCOOpenImagesMS-COCOOpenImages
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
MSP [22]70.9983.4573.1381.9180.9475.8779.0477.38
ODIN [35]59.8282.2063.1482.5962.8574.4458.9276.61
Mahalanobis [34]96.4659.2596.2757.4257.6684.9260.1686.88
Energy score [38]56.8983.6958.6982.9860.0677.4854.7979.60
Gram matrices [50]62.7579.8867.4277.6260.9374.9377.5559.38
Generalized ODIN [25]58.5783.1270.2879.2357.2785.2250.1787.18
CSI [52]59.9181.8357.4182.9547.1084.0937.0687.99
GAN-synthesis [33]60.9383.6759.9782.6757.0378.8250.6181.25
SIREN-KNN [12]47.4589.6750.3888.80----
VOS [15]47.5388.7051.3385.2344.2786.8735.5488.52
SR-VAE [61]42.1790.2846.2687.8932.2390.6921.8193.55
DFDD [60]41.3490.7944.5288.6530.7190.7422.6792.48
SAFE [59]47.4080.3020.0692.2832.5688.9616.0494.64
Faster R-CNN + Ours36.4486.5213.3495.3722.6795.4412.9696.26
VOS + Ours34.9787.9011.2596.9623.0994.3214.1296.41
", + "bbox": [ + 217, + 236, + 785, + 469 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/e07a92a120f1f8f957865bb4e8b2b27e093375ed5265df1befa3647ea7965ef1.jpg", + "table_caption": [ + "Table 2: Ablation on the number of our synthetic data in training. Taking PASCAL-VOC as the ID dataset, we perform seven groups of random sampling with different numbers in the synthetic dataset to extract features as OOD samples, evaluate and report the performance on the MS-COCO/OpenImages datasets." + ], + "table_footnote": [], + "table_body": "
#Sample14k12k10k8k6k4k2k
FPR95↓36.70/12.9636.27/13.0137.31/13.2536.53/13.3036.70/12.9636.44/13.3437.82/13.87
AUROC↑86.65/95.5486.68/95.5586.64/95.5186.69/95.5286.75/95.5686.52/95.4486.03/95.18
", + "bbox": [ + 218, + 545, + 785, + 595 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Bridged by our synthetic data, foundation models' extensive knowledge and powerful logic about novel concepts are effectively injected into our model through novel concept imagining and region-level editing. Furthermore, powered by the similarity-based filter, our synthetic data proves to be highly effective. Compared with SAFE [59] which uses a similar framework, we only use around $25\\%$ (on PASCAL-VOC) and $20\\%$ (on BDD-100K) of auxiliary data to achieve a significant performance improvement. Further analysis is presented in Sec. 4.2.", + "bbox": [ + 212, + 613, + 787, + 722 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 739, + 390, + 757 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Number of Training Samples We conduct an extensive ablation study on the quantity of synthetic data utilized, as illustrated in Tab. 2. We employ seven sets of synthetic data with varying quantities as OOD samples, using PASCAL-VOC as the ID dataset. Features are extracted and the OOD detector is trained based on the same Faster R-CNN checkpoint for each set. It is noteworthy that", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Can OOD Object Detectors Learn from Foundation Models?", + "bbox": [ + 325, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/c5a3bb8172076a8da0dcc94c6ff5a3fee82c6017c7c3ff1a6bb83ffb02422731.jpg", + "table_caption": [ + "Table 3: We study the impact of associating varying numbers of imagined novel objects with each ID object. Taking PASCAL-VOC as the ID dataset, we report the performance on MS-COCO/OpenImages datasets." + ], + "table_footnote": [], + "table_body": "
#Sample345678
FPR95↓36.96/13.5837.13/13.1537.31/12.8236.87/13.2537.91/13.5337.13/13.15
AUROC↑86.56/95.5486.63/95.5186.46/95.4786.51/95.3786.43/95.3586.44/95.56
", + "bbox": [ + 225, + 195, + 785, + 252 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/8b32d372cb797d2b57b7000c9f7d676823ab2ac2b933b6b52cc21f037e6acb34.jpg", + "table_caption": [ + "Table 4: We randomly sample the same numbers of OOD features as the main experiment instead of using the feature filter (denoted as w/o filter), and evaluate on multiple datasets. The obtained results demonstrate the effectiveness of the proposed data filter." + ], + "table_footnote": [], + "table_body": "
MethodID:PASCAL-VOCID:BDD-100K
MS-COCOOpenImagesMS-COCOOpenImages
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
w/o filter39.2985.4613.6895.2225.4593.1715.3295.83
Ours36.4486.5213.3495.3722.6795.4412.9696.26
", + "bbox": [ + 220, + 330, + 785, + 419 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "as the number of samples decreases from 14k to 2k, the performance does not deteriorate but rather maintains stable and superior results. This highlights our method's data efficiency (SAFE used 16k samples). Combined with our feature similarity-based filtering strategy as in Sec. 3.2, a small number of high-quality OOD samples with visual similarity directly promotes the optimization of precise decision boundaries to achieve stable improvements.", + "bbox": [ + 212, + 438, + 787, + 527 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Meanwhile, we use the same detector checkpoint to assess the parallel baseline, SAFE [59], and get the performance of 50.86/23.60 on FPR95 and 78.15/91.42 on AUROC. SAFE augments about 16k images and extracts more than 100k instance-level features as OOD samples. In contrast, our approach utilizes fewer synthetic images and extracts only one instance-level feature from the edited novel object in each synthetic image as an OOD sample, resulting in significantly superior performance compared to SAFE [59].", + "bbox": [ + 212, + 529, + 787, + 636 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Number of Concepts to Imagine We employ in-context learning to guide LLM in associating new objects for driving image editing. For each ID object, the LLM connects a steady stream of novel objects. We further explore the impact of the number of corresponding novel objects for each ID object on data performance. We randomly sample different numbers of novel objects from LLM's responses for each ID object. As shown in Tab. 3, the performance remains stable despite changes in the number of concepts, further highlighting the stability and robustness of our synthetic data.", + "bbox": [ + 212, + 648, + 787, + 770 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "SAM-based Refiner As mentioned in Sec. 3.1, we propose to utilize SAM-based refiner to correct the bounding boxes of novel objects to obtain higher-quality instance-level OOD features. Therefore, we comparatively remove the proposed refiner and directly used the corresponding editing masks as bound-", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/fe7cb7fe213e39e1e75e33c1560abc363efc2324b6b698df79d021f04c9be644.jpg", + "image_caption": [ + "Fig. 3: We show cases on six intervals of feature similarity (consistent with Eq. (5), indicated at the bottom of the figure). The first line contains the corresponding initial images, the second line contains the synthetic images with the corresponding boxes of the novel objects (yellow boxes), and the third line contains the difference heat maps of the latent feature maps extracted from the above image pairs (superimposed on the corresponding synthetic images, denoted as Diff-map)." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 785, + 448 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "ing boxes to extract OOD features for training. Taking PASCAL-VOC as the ID dataset, after removing the refiner, we obtain 39.55/13.72 of FPR95 and 85.94/95.37 of AUROC on MS-COCO/OpenImages datasets, which is better than previous methods but worse than the results (Faster R-CNN + Ours in Tab. 1) when using the refiner. This proves that OOD supervision signals contained in the synthetic data are already extracted and achieve good results under the localization of the fuzzy boxes, but more precise boxes mean higher quality features. More demos and analyses of the SAM-based refiner are shown in the supplementary material.", + "bbox": [ + 212, + 561, + 787, + 699 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Similarity-based Filter The filter is designed to incorporate the most useful data into training, and avoid unnecessary noise. The design is reflected in two aspects: on one hand, the outlier object should process similar visual patterns to the original object, thus being confusing and can facilitate learning; on the other hand, over-high similarity may indicate failures of the editing process (e.g., when the concept is not an object). These considerations are applied as thresholding on pairwise cosine similarity between object features, as in Eq. (5). As shown in Tab. 4, this filter brings a notable improvement across benchmarks.", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Can OOD Object Detectors Learn from Foundation Models?", + "bbox": [ + 326, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/bd7e1af740465d1a34743812f4612eca5f921cdce2c9f0ca830b39b6e3a19c27.jpg", + "table_caption": [ + "Table 5: Comparing varied images as OOD samples for training, we first show some synthetic object-centric images generated by Stable Diffusion (left side). Then with PASCAL-VOC as ID dataset, we report the results obtained by using synthetic object-centric images (denoted as object-centric images) and scene-level images with novel objects but without bounding boxes (denoted as scene-level w/o boxes) as OOD samples to participate in training." + ], + "table_footnote": [], + "table_body": "
Object-centric Images DataMS-COCOOpenImages
FPR95↓ AUROC↑FPR95↓ AUROC↑
object-centric images51.9981.4820.7093.85
scene-level w/o boxes48.0182.3818.6193.44
Ours36.4486.5213.3495.37
", + "bbox": [ + 220, + 233, + 787, + 319 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To provide more insights on the choices of filtering thresholds, we display some cases in different intervals of feature similarity in Fig. 3. We show the synthetic images, the corresponding initial images, and the difference between feature maps (denoted as Diff-map), respectively. The Diff-maps prove that the edited area is sensitively attended to by the model. But for images with extremely high similarity $(>0.9)$ , they always contain some editing failures and blurs. As illustrated on the top of the first column in Fig. 3, it is not intuitively evident what the ship has been edited into (the target object is a raft). Besides, as the similarity upperbound decreases, we progressively obtain more realistic and reasonable images. But note that when the similarity is excessively low, as seen in the last column of Fig. 3, the objects are edited into the corresponding text or an unnatural object, leading to image distortion. This strongly supports the idea that the quality and usability of edited images are closely connected to visual similarity. More cases and analyses are presented in the supplementary material.", + "bbox": [ + 212, + 330, + 787, + 542 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Discussions on Outlier Synthesis", + "text_level": 1, + "bbox": [ + 215, + 561, + 529, + 575 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Scene-level Editing Matters Through regional-level editing, we replace the ID object with a novel object with a bounding box and ensure consistent context information. However, some simpler methods based on foundation models also achieve the acquisition and use of OOD data. For example, Dream-OOD [13] uses well-trained text-conditional space and diffusion model to synthesize realistic object-centric data for promoting OOD image classification. Similarly, keeping other settings unchanged, we use our novel concepts to drive Stable-Diffusion instead of Stable-Diffusion-Inpainting [49] to synthesize novel images, which are also processed and filtered by our proposed refiner and filter (some synthesized images are shown in Tab. 5), thereby participating in the training as OOD supervision. However, as shown in Tab. 5 (object-centric images), the synthetic novel object-centric images do not aid in training and result in poor performance, even though they possess high visual quality. This clearly validates our decision to edit scene-level images rather than composing new ones, and highlights the significance of maintaining contextual consistency.", + "bbox": [ + 212, + 583, + 787, + 809 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Additionally, we examine the possibility of using the edited scene-level image as a whole (ignoring the boxes) as OOD samples in the training process. The", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a748bb0d397ed352e10074919107cb8a6c4546aec7ae6d4bd06cf5eeeef22710.jpg", + "image_caption": [ + "Fig. 4: We edit the context of the synthetic data (in blue box) so that the images contain novel objects and novel context (in orange box). Then we calculate the similarity between the instance-level feature of the corresponding objects from all synthetic images and the instance-level feature in the initial image (left, the bird). The similarities and the corresponding difference maps are shown in the figure." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 782, + 296 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "results, as depicted in Tab. 5 (scene-level w/o boxes), are significantly inferior compared to our method's performance. This demonstrates that controllable bounding boxes are indispensable in this task.", + "bbox": [ + 212, + 393, + 787, + 441 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Context Consistency Matters Given the importance of scene-level synthesis as discussed in the previous paragraph, we study the factors that make scene-level editing indispensable, and find context consistency to be a crucial one. Besides calculating the similarity between ID/OOD object pairs before/after box-conditioned editing, we also try further editing parts of the background of the already edited images, and also calculate its object similarity with the initial object. As shown in Fig. 4, even small editing on parts of the background (out of the object boxes) can make foreground objects 'look' notably different as perceived by the detector. This highlights the importance of keeping the context unchanged when synthesizing outlier samples, in that if the context is changed, the model easily identifies the object as OOD and cannot break the context bias.", + "bbox": [ + 212, + 450, + 789, + 617 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Discussion", + "text_level": 1, + "bbox": [ + 215, + 638, + 354, + 654 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "What Type of OOD Data Matters? We are the first to explore how to edit scene-level images to include novel categories, which contain annotation boxes and ensure context consistency, facilitating OOD object detection. The achieved state-of-the-art performance (Tab. 1) benefits from the optimization of decision boundaries driven by high-quality OOD features. Our exploration demonstrates that annotation boxes and context consistency are particularly important for synthesizing high-quality OOD instances. On the one hand, high-quality annotation boxes provide the possibility to extract high-quality instance-level features from scene-level images, while unrefined boxes (Sec. 4.2) or discarded boxes (Tab. 5) will have a negative impact on performance. On the other hand, context consistency ensures the most effective OOD features are not interfered by different contexts and selected for utilizing (Fig. 4).", + "bbox": [ + 212, + 666, + 787, + 848 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Can OOD Object Detectors Learn from Foundation Models?", + "bbox": [ + 326, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "How to Synthesize Suitable OOD Data? We are the first to build an automatic, transparent, and low-cost pipeline (Sec. 3.1) for synthesizing scene-level images containing novel objects with annotation boxes and context consistency. It benefits object detectors' robustness and reliability to unseen data and sets up clear state-of-the-art on multiple OOD object detection benchmarks. Specifically, we organically combine and cleverly use different foundation models [1, 30, 49] (Sec. 3.1) to distill open-world knowledge and inpaint the existing scenes for simulating real OOD scenarios. In addition, our design takes into account the instability of the current foundation models and can release better potential performance in the future development of foundation models.", + "bbox": [ + 215, + 145, + 785, + 297 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "How to Select Suitable OOD Data? We are the first to explicitly decouple OOD data synthesis and selection. On the one hand, we ensure the separability of the synthetic objects in semantic concepts through open-world knowledge provided by LLMs (Sec. 3.1). On the other hand, we ensure the similarity of ID and OOD objects in visual patterns (Sec. 3.1), thereby optimizing the precise decision boundary. This line of thinking has the potential to facilitate more open-world solutions.", + "bbox": [ + 215, + 311, + 785, + 417 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Broader Impacts Beyond showcasing engineering success via effectively combining specific foundation models, our work uncovers the untapped potential of the text-to-image generative models and visual foundation models in pushing forward the OOD object detection task to effectively leverage the off-the-shelf open-world data knowledge [5,31,66]. More importantly, our work establishes a bridge between OOD object detection and the latest advancements in deep learning, enabling it to benefit from ongoing developments, go beyond isolated academic practice, and resolve practical challenges in open-world applications. Meanwhile, automating novel data generation and curation will inspire more tasks in more modalities, such as in visual-language [44,58] and 3D vision [8,37,64].", + "bbox": [ + 215, + 433, + 785, + 585 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 608, + 357, + 625 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we investigate improving OOD object detection by distilling open-world data knowledge from text-to-image generative models. We develop an automatic and cost-effective data curation pipeline, SyncOOD, that leverages foundation models as tools to obtain meaningful open-set data from generative models. Through extensive studies, we discover that object boxes and context consistency of the generated data contribute to the improvement of OOD object detection performance. Our comprehensive experiments demonstrate that SyncOOD not only advances the state-of-the-art in OOD object detection but also emphasizes the untapped potential of utilizing large-scale generative models for enhancing the robustness of machine learning systems in open-world settings. As an initial exploration in leveraging foundation models for OOD object detection, we hope our promising results encourage further research in advancing this area in the future.", + "bbox": [ + 215, + 640, + 785, + 835 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work has been supported by Hong Kong Research Grant Council - Early Career Scheme (Grant No. 27209621), General Research Fund Scheme (Grant No. 17202422), Theme-based Research (Grant No. T45-701/22-R) and RGC Matching Fund Scheme (RMGS). Part of the described research work is conducted in the JC STEM Lab of Robotics for Soft Materials funded by The Hong Kong Jockey Club Charities Trust. We would like to thank Chirui Chang, Xiaoyang Lyu, Haoru Tan, and Xiuzhe Wu for their insightful discussions.", + "bbox": [ + 212, + 179, + 787, + 285 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 311, + 321, + 328 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: GPT-4 technical report. arXiv preprint arXiv:2303.08774 (2023)", + "2. Bendale, A., Boult, T.E.: Towards open set deep networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1563-1572 (2016)", + "3. Bishop, C.M.: Novelty detection and neural network validation. IEE Proceedings-Vision, Image and Signal processing 141(4), 217-222 (1994)", + "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in Neural Information Processing Systems 33, 1877-1901 (2020)", + "5. Chang, C., Liu, Z., Lyu, X., Qi, X.: What matters in detecting ai-generated videos like sora? arXiv preprint arXiv:2406.19568 (2024)", + "6. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. Journal of Machine Learning Research 24(240), 1-113 (2023)", + "7. Deepshikha, K., Yelleni, S.H., Srijith, P., Mohan, C.K.: Monte Carlo dropblock for modelling uncertainty in object detection. arXiv preprint arXiv:2108.03614 (2021)", + "8. Deng, W., Ding, R., Yang, J., Liu, J., Li, Y., Qi, X., Ngai, E.: Can 3d vision-language models truly understand natural language? arXiv preprint arXiv:2403.14760 (2024)", + "9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)", + "10. Dhamija, A., Gunther, M., Ventura, J., Boult, T.: The overlooked elephant of object detection: Open set. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1021-1030 (2020)", + "11. Du, X., Fang, Z., Diakonikolas, I., Li, Y.: How does unlabeled data provably help out-of-distribution detection? arXiv preprint arXiv:2402.03502 (2024)", + "12. Du, X., Gozum, G., Ming, Y., Li, Y.: SIREN: Shaping representations for detecting out-of-distribution objects. Advances in Neural Information Processing Systems 35, 20434-20449 (2022)", + "13. Du, X., Sun, Y., Zhu, J., Li, Y.: Dream the impossible: Outlier imagination with diffusion models. Advances in Neural Information Processing Systems 36 (2024)" + ], + "bbox": [ + 218, + 347, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Can OOD Object Detectors Learn from Foundation Models?", + "bbox": [ + 325, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Du, X., Wang, X., Gozum, G., Li, Y.: Unknown-aware object detection: Learning what you don't know from videos in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13678-13688 (2022)", + "15. Du, X., Wang, Z., Cai, M., Li, Y.: VOS: Learning what you don't know by virtual outlier synthesis. arXiv preprint arXiv:2202.01197 (2022)", + "16. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes (VOC) challenge. International journal of computer vision 88, 303-338 (2010)", + "17. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: International Conference on Learning Representations (2022)", + "18. Gupta, A., Narayan, S., Joseph, K., Khan, S., Khan, F.S., Shah, M.: Ow-detr: Open-world detection transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9235-9244 (2022)", + "19. Hall, D., Dayoub, F., Skinner, J., Zhang, H., Miller, D., Corke, P., Carneiro, G., Angelova, A., Sünderhauf, N.: Probabilistic object detection: Definition and evaluation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1031-1040 (2020)", + "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016)", + "21. He, R., Sun, S., Yu, X., Xue, C., Zhang, W., Torr, P., Bai, S., Qi, X.: Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574 (2022)", + "22. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. In: International Conference on Learning Representations (2017)", + "23. Hendrycks, D., Mazeika, M., Dietterich, T.: Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606 (2018)", + "24. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)", + "25. Hsu, Y.C., Shen, Y., Jin, H., Kira, Z.: Generalized ODIN: Detecting out-of-distribution image without learning from out-of-distribution data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10951-10960 (2020)", + "26. Huang, R., Geng, A., Li, Y.: On the importance of gradients for detecting distributional shifts in the wild. Advances in Neural Information Processing Systems 34, 677-689 (2021)", + "27. Joseph, K., Khan, S., Khan, F.S., Balasubramanian, V.N.: Towards open world object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5830-5840 (2021)", + "28. Joseph, K., Khan, S., Khan, F.S., Balasubramanian, V.N.: Towards open world object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5830-5840 (2021)", + "29. Katz-Samuels, J., Nakhleh, J.B., Nowak, R., Li, Y.: Training ood detectors in their natural habitats. In: International Conference on Machine Learning. pp. 10848-10865. PMLR (2022)", + "30. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Kong, S., Ramanan, D.: Opengan: Open-set recognition via open data generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 813-822 (2021)", + "32. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision 128(7), 1956-1981 (2020)", + "33. Lee, K., Lee, H., Lee, K., Shin, J.: Training confidence-calibrated classifiers for detecting out-of-distribution samples. In: International Conference on Learning Representations (2018)", + "34. Lee, K., Lee, K., Lee, H., Shin, J.: A simple unified framework for detecting out-of-distribution samples and adversarial attacks. Advances in Neural Information Processing Systems 31 (2018)", + "35. Liang, S., Li, Y., Srikant, R.: Enhancing the reliability of out-of-distribution image detection in neural networks. In: International Conference on Learning Representations, ICLR 2018 (2018)", + "36. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "37. Liu, J., Chang, C., Liu, J., Wu, X., Ma, L., Qi, X.: Mars3d: A plug-and-play motion-aware model for semantic segmentation on multi-scan 3d point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9372-9381 (2023)", + "38. Liu, W., Wang, X., Owens, J., Li, Y.: Energy-based out-of-distribution detection. Advances in Neural Information Processing Systems 33, 21464-21475 (2020)", + "39. Liu, X., Yang, H., Ravichandran, A., Bhotika, R., Soatto, S.: Continual universal object detection. arXiv preprint arXiv:2002.05347 (2020)", + "40. Miller, D., Dayoub, F., Milford, M., Sünderhauf, N.: Evaluating merging strategies for sampling-based uncertainty techniques in object detection. In: International Conference on Robotics and Automation (ICRA). pp. 2348-2354 (2019)", + "41. Miller, D., Nicholson, L., Dayoub, F., Sünderhauf, N.: Dropout sampling for robust object detection in open-set conditions. In: IEEE International Conference on Robotics and Automation (ICRA). pp. 3243-3249 (2018)", + "42. Nguyen, A., Yosinski, J., Clune, J.: Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 427-436 (2015)", + "43. Perez-Rua, J.M., Zhu, X., Hospedales, T.M., Xiang, T.: Incremental few-shot object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13846-13855 (2020)", + "44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "45. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)", + "46. Rahman, S., Khan, S.H., Porikli, F.: Zero-shot object detection: Joint recognition and localization of novel concepts. International Journal of Computer Vision 128, 2979-2999 (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Can OOD Object Detectors Learn from Foundation Models?", + "bbox": [ + 326, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "47. Ren, J., Fort, S., Liu, J., Roy, A.G., Padhy, S., Lakshminarayanan, B.: A simple fix to Mahalanobis distance for improving near-ood detection. arXiv preprint arXiv:2106.09022 (2021)", + "48. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems 28 (2015)", + "49. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "50. Sastry, C.S., Oore, S.: Detecting out-of-distribution examples with gram matrices. In: International Conference on Machine Learning. pp. 8491-8501. PMLR (2020)", + "51. Sun, Y., Ming, Y., Zhu, X., Li, Y.: Out-of-distribution detection with deep nearest neighbors. In: International Conference on Machine Learning. pp. 20827-20840. PMLR (2022)", + "52. Tack, J., Mo, S., Jeong, J., Shin, J.: CSI: Novelty detection via contrastive learning on distributionally shifted instances. Advances in Neural Information Processing Systems 33, 11839-11852 (2020)", + "53. Tan, H., Wu, S., Du, F., Chen, Y., Wang, Z., Wang, F., Qi, X.: Data pruning via moving-one-sample-out. Advances in Neural Information Processing Systems (2023)", + "54. Tao, L., Du, X., Zhu, X., Li, Y.: Non-parametric outlier synthesis. arXiv preprint arXiv:2303.02966 (2023)", + "55. Tian, Y., Fan, L., Chen, K., Katabi, D., Krishnan, D., Isola, P.: Learning vision from models rivals learning vision from data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15887-15898 (2024)", + "56. Wang, H., Liu, W., Bocchieri, A., Li, Y.: Can multi-label classification networks know what they don't know? Advances in Neural Information Processing Systems 34, 29074-29087 (2021)", + "57. Wang, X., Huang, T.E., Liu, B., Yu, F., Wang, X., Gonzalez, J.E., Darrell, T.: Robust object detection via instance-level temporal cycle confusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9143-9152 (2021)", + "58. Wen, X., Zhao, B., Chen, Y., Pang, J., Qi, X.: Generalization beyond data imbalance: A controlled study on clip for transferable insights. arXiv preprint arXiv:2405.21070 (2024)", + "59. Wilson, S., Fischer, T., Dayoub, F., Miller, D., Sunderhauf, N.: SAFE: Sensitivity-aware features for out-of-distribution object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23565-23576 (2023)", + "60. Wu, A., Chen, D., Deng, C.: Deep feature deblurring diffusion for detecting out-of-distribution objects. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13381-13391 (2023)", + "61. Wu, A., Deng, C.: Discriminating known from unknown objects via structure-enhanced recurrent variational autoencoder. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23956-23965 (2023)", + "62. Wu, Q., Chen, Y., Yang, C., Yan, J.: Energy-based out-of-distribution detection for graph neural networks. arXiv preprint arXiv:2302.02914 (2023)", + "63. Wu, S., Tan, H., Tian, Z., Chen, Y., Qi, X., Jia, J.: Saco loss: Sample-wise affinity consistency for vision-language pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 27358-27369 (2024)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "64. Wu, X., Dai, P., Deng, W., Chen, H., Wu, Y., Cao, Y.P., Shan, Y., Qi, X.: Clnerf: continual learning of neural radiance fields for evolving scene representation. Advances in Neural Information Processing Systems 36 (2024)", + "65. Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F., Madhavan, V., Darrell, T.: BDD100K: A diverse driving dataset for heterogeneous multitask learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2636-2645 (2020)", + "66. Zheng, H., Wang, Q., Fang, Z., Xia, X., Liu, F., Liu, T., Han, B.: Out-of-distribution detection learning with unreliable out-of-distribution sources. Advances in Neural Information Processing Systems 36, 72110-72123 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 285 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Can OOD Object Detectors Learn from Foundation Models?", + "bbox": [ + 326, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Can OOD Object Detectors Learn from Foundation Models_/6d5daec6-da83-48aa-809b-d1c5b6b09141_model.json b/2024/Can OOD Object Detectors Learn from Foundation Models_/6d5daec6-da83-48aa-809b-d1c5b6b09141_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f2ddec88fa5e5d89898b558b2793d39c5f93deed --- /dev/null +++ b/2024/Can OOD Object Detectors Learn from Foundation Models_/6d5daec6-da83-48aa-809b-d1c5b6b09141_model.json @@ -0,0 +1,2339 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.276, + 0.141, + 0.728, + 0.184 + ], + "angle": 0, + "content": "Can OOD Object Detectors Learn from Foundation Models?" + }, + { + "type": "text", + "bbox": [ + 0.247, + 0.213, + 0.756, + 0.229 + ], + "angle": 0, + "content": "Jiahui Liu, Xin Wen, Shizhen Zhao, Yingxian Chen, and Xiaojuan Qi*" + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.24, + 0.674, + 0.27 + ], + "angle": 0, + "content": "The University of Hong Kong {liujh,wenxin,zhaosz,chenyx,xjqi}@eee.hku.hk" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.309, + 0.74, + 0.545 + ], + "angle": 0, + "content": "Abstract. Out-of-distribution (OOD) object detection is a challenging task due to the absence of open-set OOD data. Inspired by recent advancements in text-to-image generative models, such as Stable Diffusion, we study the potential of generative models trained on large-scale open-set data to synthesize OOD samples, thereby enhancing OOD object detection. We introduce SyncOOD, a simple data curation method that capitalizes on the capabilities of large foundation models to automatically extract meaningful OOD data from text-to-image generative models. This offers the model access to open-world knowledge encapsulated within off-the-shelf foundation models. The synthetic OOD samples are then employed to augment the training of a lightweight, plug-and-play OOD detector, thus effectively optimizing the in-distribution (ID)/OOD decision boundaries. Extensive experiments across multiple benchmarks demonstrate that SyncOOD significantly outperforms existing methods, establishing new state-of-the-art performance with minimal synthetic data usage. The project is available at https://github.com/CVMI-Lab/SyncOOD." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.559, + 0.738, + 0.573 + ], + "angle": 0, + "content": "Keywords: OOD object detection \\(\\cdot\\) Synthetic data \\(\\cdot\\) Open-world data" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.601, + 0.377, + 0.617 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.634, + 0.788, + 0.739 + ], + "angle": 0, + "content": "Modern object detectors, trained on closed-set data, have achieved remarkable success. However, they often incorrectly and confidently classify out-of-distribution (OOD) object categories as in-distribution (ID) categories in open-world applications [10], raising concerns about their reliability for deployment. To enhance the trustworthiness of object detection, researchers have studied the OOD object detection task, which aims to identify and flag unknown or novel objects as distinct from ID ones [3,42]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.74, + 0.788, + 0.817 + ], + "angle": 0, + "content": "The vulnerability of these models to OOD samples stems from their lack of awareness of the unknown open data distribution during training. Consequently, synthesizing OOD samples for model learning has emerged as a major research direction for this task [15, 59-61]. Most existing studies [15, 60, 61] concentrate on generating OOD objects in the latent space of an object detection model" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.826, + 0.379, + 0.84 + ], + "angle": 0, + "content": "* corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.787, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.422, + 0.788, + 0.52 + ], + "angle": 0, + "content": "Fig. 1: Our pipeline replaces ID objects with semantic-novel yet visual-similar objects for scene-level OOD object synthesis. Middle left: The concepts are imagined by an LLM to ensure semantic separability and rationality, and reformed as text prompts for controllable in-painting using Stable Diffusion. Middle right: During training, only visually similar OOD objects are adopted based on instance-level feature similarity to the original object. A lightweight binary classifier is optimized for OOD detection, and other parts of the detector are kept unchanged." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.687 + ], + "angle": 0, + "content": "trained on ID data. These synthesized samples are then used to optimize the decision boundary between ID and OOD data. Alternative approaches include directly synthesizing images by injecting adversarial noise [59] or identifying OOD instances from video data [14]. While these methods have yielded promising results, they remain limited to a closed-set setting, where the latent space for synthesizing outliers or the data is derived from a closed-set data distribution. Consequently, they may be biased towards the ID dataset, leading to suboptimal performance. Besides, effectively handling the unknown may appear unattainable when the unknown is never fully exploitable. Beyond that, is it possible to learn them from massive open-world data knowledge condensed in foundation models?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.788, + 0.809 + ], + "angle": 0, + "content": "To this end, we investigate text-to-image generation models trained on a large amount of open-set data, which have demonstrated a superior ability to capture the distribution of data across a wide range of visual concepts, in order to synthesize novel data samples for enhancing OOD detection. Nonetheless, automatically extracting meaningful data from generative models for OOD object detection remains challenging due to the extensive vocabulary space to be explored, the complex scene-level synthesis problem, the need for object-level annotations, and potential distractions from contextual cues." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We introduce SyncOOD, an automatic data curation process that leverages foundation models as tools to harvest meaningful data from text-to-image gen-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.326, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Can OOD Object Detectors Learn from Foundation Models?" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.313 + ], + "angle": 0, + "content": "eration models for OOD object detection (see Figure 1). The process is based on two key observations: 1) Hard OOD samples that are close to the ID data contribute more to learning a better OOD detector, and 2) Context may become a distracting cue for OOD object detection tasks, leading to bias towards contexts. With these observations in mind, the outlier synthesis process is formulated as box-conditioned image inpainting, and driven by Stable Diffusion [49] for high-quality controllable editing. The concepts to replace with are imagined by a Large Language Model (LLM) [1] with the aim of semantic novelty,\\(^{1}\\) and the associated bounding box is further refined with SAM [30]. Automated by foundation models, this data collection pipeline requires minimal human labor, while producing high-quality OOD data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.435 + ], + "angle": 0, + "content": "In comparison to existing methods for OOD object detection, our core insight is to broaden the model's exposure to a more extensive range of open-set data and circumvent dataset biases by tapping into the open-world knowledge found in off-the-shelf foundation models. Utilizing generative models also provides us with control over the context of synthesized images and the data distribution. Our comprehensive experiments demonstrate superior performance, emphasizing the untapped potential of text-image-generation models in the context of OOD object detection. Our key contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.446, + 0.784, + 0.489 + ], + "angle": 0, + "content": "- We investigate and unlock the potential of text-to-image generative models trained on large-scale open-set data for synthesizing OOD objects in object detection tasks." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.491, + 0.785, + 0.55 + ], + "angle": 0, + "content": "- We introduce an automated data curation process for obtaining controllable, annotated scene-level synthetic OOD images for OOD object detection, which utilizes LLMs for novel concept discovery and visual foundation models for data annotation and filtering." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.551, + 0.785, + 0.596 + ], + "angle": 0, + "content": "- We discover that maintaining ID/OOD image context consistency and obtaining more accurate OOD annotation bounding boxes are crucial for synthesized data to be effective in OOD object detection." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.597, + 0.785, + 0.641 + ], + "angle": 0, + "content": "- Comprehensive experiments on multiple benchmarks demonstrate the effectiveness of our method, as we significantly outperform existing state-of-the-art approaches while using minimal synthetic data." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.446, + 0.785, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.664, + 0.388, + 0.68 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.695, + 0.788, + 0.817 + ], + "angle": 0, + "content": "OOD Object Detection For detecting OOD objects in scene-level images, unlike earlier works that constrain ID samples to a hypothetical distribution [12], it has become a recent trend to explicitly synthesize the outlier data, and incorporate them into training to adjust models' decision boundaries. However, due to the complexity of scene-level images, all previous works bypassed photorealistic outlier synthesis in pixel space, and worked on generating outliers from the model's latent space [15, 60, 61], adversarial attack [59], or utilizing video data in the wild [14]. In the former, outliers can be sampled from the latent" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.783, + 0.84 + ], + "angle": 0, + "content": "1 Concepts overlapping with the test data are removed to avoid information leakage." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.128 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "space using a simple Gaussian prior [15], or more advanced generative models like VAE [61] or diffusion model [60]. Yet their upperbound are commonly limited by the quality of the latent space, and synthesized samples lack interpretability. Despite this, adversarial samples [59] lack semantic diversity, and auxiliary pseudo supervision from videos [14] introduces additional requirements to the setting. Unlike all above, our method applies LLM for OOD concept sampling, and Stable Diffusion for controllable image editing. This decouples the sampling and generation processes, elevates both parts to an unprecedented level, and achieves photo-realistic scene-level OOD image synthesis for the first time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Open-world Object Detection The fact that real-world applications require object detectors the ability to tackle open categories is also considered in open-world object detection. The focus of this field includes generalization to domain shifts [57], incremental learning of novel classes [27, 39, 43], and zero-shot classification of open-vocabularies [17, 46]. Meanwhile, some works [18, 28] require to distinguish known objects well, be able to detect unknown objects, and finally be able to incrementally learn new objects. These works provide different perspectives on facing the challenges of real-world applications, which complement OOD object detection as a joint effort, but are out of the scope of this paper." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.675 + ], + "angle": 0, + "content": "OOD Image Classification Earlier paradigms for OOD image classification either post hoc adjust models' confidence score at the testing phase, or apply regularization at models' training phase. The former line mainly focuses on the design of score functions, including confidence-based [2, 22, 35], energy-based [38, 56, 62], distance-based [34, 47, 50, 51], gradient-based [26], and approximating Bayesian [7, 10, 19, 40, 41] methods. The latter line of work includes regularizing models to produce lower confidence [23, 33], higher energy [29, 38], or directly shaping latent representations [12]. While outlier synthesis has shown to be effective by [15, 54], these are still generated in the latent space, and a parallel line study utilizing natural images [11, 23, 29] from the wild. Recently, photo-realistic outlier synthesis was first achieved by [13] with help from a text-conditioned diffusion model [49]. However, it is not readily applicable for object detection due to the complexity of scene-level images and the requirement for object-level annotations. In contrast, our work studies under the detection setting and requires outlier synthesis at scene level." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Foundation Models The evolution of large language models (LLMs) began with training on web-scale datasets [9,45], leading to increasingly powerful foundation models [4, 6] capable of harnessing vast open-world data. Notable advancements include models [1] that interact with users and perform complex tasks like question answering, significantly broadening access to global knowledge. In image generation, diffusion models [24, 49] offer robust capabilities in synthesizing realistic content for applications such as image synthesis and inpainting. Additionally, segmentation foundation models like SAM [30] represent a leap forward in precise image segmentation, benefiting from extensive data training. Foundation models provide diverse data, which provides unlimited po" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.326, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Can OOD Object Detectors Learn from Foundation Models?" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.147, + 0.782, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.293, + 0.788, + 0.351 + ], + "angle": 0, + "content": "Fig. 2: Detailed illustration of our outlier synthesis pipeline. It comprises (a) Instructing an LLM to imagine semantic-novel concepts given ID objects, (b) Editing the selected regions to the expected concepts via prompt-conditioned image inpainting using Stable Diffusion, and (c) Refining the bounding boxes of edited objects using SAM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.788, + 0.41 + ], + "angle": 0, + "content": "tential for learning open-world knowledge from these models [55], where the effectiveness of data [21, 53, 63] is also crucial to the downstreaming tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.432, + 0.331, + 0.449 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.563 + ], + "angle": 0, + "content": "Preliminary For OOD object detection, the training set contains only ID scene-level images \\(\\mathbf{x}^{\\mathrm{id}}\\) with ID objects, annotation bounding boxes \\(\\mathbf{b}^{\\mathrm{id}}\\), and semantic labels \\(y\\), denoted as \\(\\mathcal{D}_{\\mathrm{id}} = \\left\\{(\\mathbf{x}^{\\mathrm{id}}, \\mathbf{b}^{\\mathrm{id}}, \\mathbf{y}^{\\mathrm{id}})\\right\\}\\). The labels of ID objects always belong to a close set with \\(K\\) categories, denoted as \\(\\mathbf{y}^{\\mathrm{id}} \\in \\mathcal{Y}_{\\mathrm{id}}\\) and \\(\\mathcal{Y}_{\\mathrm{id}} = \\{y_1^{\\mathrm{id}}, y_2^{\\mathrm{id}}, \\dots, y_K^{\\mathrm{id}}\\}\\). During inference, for each proposed object from an input scene-level image, it is required to identify whether its category belongs to \\(\\mathcal{Y}_{\\mathrm{id}}\\) or not." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.573, + 0.789, + 0.75 + ], + "angle": 0, + "content": "Overview As illustrated in Fig. 1, our outlier synthesis pipeline consists of (1) synthesizing a set of effective photo-realistic scene-level OOD images \\(\\mathbf{x}^{\\mathrm{edit}}\\), denoted as \\(\\mathcal{D}_{\\mathrm{edit}} = \\left\\{(\\mathbf{x}^{\\mathrm{edit}},\\mathbf{b}^{\\mathrm{edit}})\\right\\}\\), which contains novel objects and corresponding annotation boxes \\(\\mathbf{b}^{\\mathrm{edit}}\\) based on region-level editing from \\(\\mathcal{D}_{\\mathrm{id}}\\) in a fully automated, labor-free way; and (2) select and use the efficient synthetic data to provide pseudo-OOD supervisions for training OOD object detector together with the ID samples in the training set. Further design of the pipeline requires answering the following questions: (1) how to distill the open-set knowledge embedded in foundation models to scene-level OOD data and (2) how to utilize the synthesized data to regularize the decision boundary and facilitate OOD object detection. We discuss them accordingly in Sec. 3.1 and Sec. 3.2." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.77, + 0.703, + 0.787 + ], + "angle": 0, + "content": "3.1 Synthesizing Semantic-novel Objects in Scene Images" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Imagining Novel Concepts from ID objects As shown in Fig. 2(a), based on the ID labels \\(\\mathcal{V}_{\\mathrm{id}}\\) in training set \\(\\mathcal{D}_{\\mathrm{id}}\\), we consider that novel concepts that are different from ID categories can be potential candidates for generating OOD" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.128 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.27 + ], + "angle": 0, + "content": "objects. The next is to discover novel concepts that offer hard OOD samples sharing high visual similarity with ID samples and being contextually compatible with the scene context for object detection. Rather than relying on human labor to investigate all potential candidates, we leverage the vast knowledge and reasoning capabilities of LLM, GPT-4 [1] to check the visual similarity and contextual compatibility. This allows us to associate ID objects and promote the conceptualization of possible novel objects to replace existing ID objects through the use of a prompt with in-context examples [4] as:" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.285, + 0.765, + 0.358 + ], + "angle": 0, + "content": "Here is a list containing several objects \\(\\mathcal{V}_{\\mathrm{id}}\\). Now, if I provide you an object name, you should return to me objects that are similar to the usage scenario and volume of the provided object but are not in the previous object list. For example, if I give you the word: person, you should respond and only respond: 'mannequin', 'sculpture', 'scarecrows', 'doll', 'puppet'." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.789, + 0.468 + ], + "angle": 0, + "content": "With its robust logical foundation and rich knowledge, the LLM envisions a collection of novel objects for each ID object label, denoted as \\(\\mathcal{Y}_{\\mathrm{novel}}\\), while maintaining the semantic separability between imagined objects and ID objects. We empirically find one in-context example that is sufficient for us to discover novel concepts. For each ID label \\(i\\), we discover \\(M\\) novel concepts using LLM \\(\\mathbf{y}_i^{\\mathrm{novel}}\\) of \\(M\\) concepts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.789, + 0.68 + ], + "angle": 0, + "content": "Editing Objects on Selected Regions With the discovered novel concepts \\(\\mathcal{V}_{\\mathrm{novel}} = \\{\\mathbf{y}_1^{\\mathrm{novel}},\\mathbf{y}_2^{\\mathrm{novel}},\\dots,\\mathbf{y}_K^{\\mathrm{novel}}\\}\\), the next step is to use them as prompts for the text-to-image generation model to generate an image. To generate a new image with novel concepts \\(y_{j}\\in \\mathbf{y}_{i}^{\\mathrm{novel}}\\), we choose to replace existing ID objects in existing images with label \\(y_{i}^{\\mathrm{id}}\\) instead of finding new locations or generating one image from scratch. By doing so, we can ensure context compatibility and eliminate distractions from the scene context as it is preserved. As illustrated in Fig. 2 (b), we use Stable-Diffusion-Inpainting [49], denoted as \\(\\mathrm{SDI}(\\cdot)\\), to perform region-level editing on ID images. The ID object is denoted as \\(\\mathbf{x}^{\\mathrm{id}}\\), with its corresponding annotation box \\(\\mathbf{b}^{\\mathrm{id}}\\) serving as the editing mask, and the associated imagined novel concept \\(\\mathbf{y}^{\\mathrm{novel}}\\) are provided as inputs to the SDI, which is one of the most successful models for conditional image generation and editing. Thus, an edited image \\(\\mathbf{x}^{\\mathrm{edit}}\\) containing a novel object is obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.694, + 0.785, + 0.712 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} ^ {\\text {e d i t}} = \\operatorname {S D I} \\left(\\mathbf {x} ^ {\\text {i d}}, \\mathbf {b} ^ {\\text {i d}}, \\mathbf {y} ^ {\\text {n o v e l}}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Refining Annotation Boxes of Novel Objects Due to the randomness in diffusion models, the attributes of edited objects, such as their quality, volume, and localization, may not match the original object box. To address this issue, as depicted in Fig. 2(c), we design an efficient and effective refiner based on SAM [30] to obtain refined accurate bounding boxes on novel objects. First, for an edited image \\(\\mathbf{x}^{\\mathrm{edit}}\\) with the editing mask \\(\\mathbf{b}^{\\mathrm{id}}\\), we use a padding area extended from \\(\\mathbf{b}^{\\mathrm{id}}\\) as the prompt and employ SAM to output the instance mask with" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.326, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Can OOD Object Detectors Learn from Foundation Models?" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.146, + 0.632, + 0.163 + ], + "angle": 0, + "content": "highest confidence \\(\\mathbf{m}^{\\mathrm{SAM}}\\) for the novel object in the area:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.167, + 0.786, + 0.186 + ], + "angle": 0, + "content": "\\[\n\\mathbf {m} ^ {\\text {S A M}} = \\operatorname {S A M} \\left(\\mathbf {x} ^ {\\text {e d i t}}; \\text {p a d d i n g} \\left(\\mathbf {b} ^ {\\text {i d}}, e\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.237 + ], + "angle": 0, + "content": "where \\( e \\) represents the range of padding. Then, we convert obtained masks \\( \\mathbf{m}^{\\mathrm{SAM}} \\) to boxes \\( \\mathbf{b}^{\\mathrm{SAM}} \\), and calculate IoU between \\( \\mathbf{b}^{\\mathrm{SAM}} \\) and the corresponding \\( \\mathbf{b}^{\\mathrm{id}} \\) to filter out novel objects that vary highly in scale:" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.242, + 0.786, + 0.268 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mathbf {b} ^ {\\text {e d i t}} \\right\\} = \\left\\{\\mathbf {b} ^ {\\text {S A M}} \\mid \\operatorname {I o U} \\left(\\mathbf {b} ^ {\\text {S A M}}, \\mathbf {b} ^ {\\text {i d}}\\right) > \\gamma \\right\\}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.271, + 0.788, + 0.334 + ], + "angle": 0, + "content": "where \\(\\gamma\\) denotes a threshold value on IoU. It ensures a high enough recall rate to rule out the instability of Stable Diffusion and SAM and uncontrollable localization of the edited objects. Thus we obtain the synthetic outlier data \\(\\mathcal{D}_{\\mathrm{edit}}\\) as illustrated in Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.35, + 0.662, + 0.366 + ], + "angle": 0, + "content": "3.2 Mining Hard OOD Samples and Model Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.789, + 0.528 + ], + "angle": 0, + "content": "Mining Hard OOD Objects with High Visual Similarities for Training We consider the novel objects that are most likely to be confused with the corresponding ID objects by the object detector as the most effective ones. We thus aim to find synthetic OOD samples that are most easily confused as ID to participate in training based on pairwise similarity in the latent space of the pre-trained object detector. For each novel object with bounding box \\(\\mathbf{b}^{\\mathrm{edit}}\\) in the synthetic data \\(\\mathcal{D}_{\\mathrm{edit}}\\), we construct it with the corresponding original ID object with its bounding box as a pair: \\(\\left\\{(\\mathbf{b}^{\\mathrm{edit}}, \\mathbf{x}^{\\mathrm{edit}}), (\\mathbf{b}^{\\mathrm{id}}, \\mathbf{x}^{\\mathrm{id}})\\right\\}\\). For an off-the-shelf object detector, denoted by \\(\\mathcal{F}_{\\mathrm{det}}\\), we extract latent features, \\(\\mathbf{z}^{\\mathrm{edit}}\\) and \\(\\mathbf{z}^{\\mathrm{id}}\\), for each pair:" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.528, + 0.786, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} ^ {\\text {e d i t}}, \\mathbf {z} ^ {\\text {i d}} = \\mathcal {F} _ {\\det } \\left(\\mathbf {b} ^ {\\text {e d i t}}; \\mathbf {x} ^ {\\text {e d i t}}\\right), \\mathcal {F} _ {\\det } \\left(\\mathbf {b} ^ {\\text {i d}}; \\mathbf {x} ^ {\\text {i d}}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.788, + 0.61 + ], + "angle": 0, + "content": "The most effective novel objects are those with visual patterns that can be easily mistaken for their corresponding ID objects by an object detector. Therefore, we filter these novel objects based on their similarity to provide pseudo-OD supervision:" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.615, + 0.786, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mathbf {z} ^ {\\mathrm {o o d}} \\right\\} = \\left\\{\\mathbf {z} ^ {\\mathrm {e d i t}} \\big | \\epsilon_ {l o w} < \\mathrm {s i m} (\\mathbf {z} ^ {\\mathrm {e d i t}}, \\mathbf {z} ^ {\\mathrm {i d}}) < \\epsilon_ {u p} \\right\\}, \\qquad \\qquad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.638, + 0.788, + 0.684 + ], + "angle": 0, + "content": "where the similarities are computed between latent object features of edit-ID pairs. Here \\(\\mathrm{sim}(\\cdot)\\) denotes cosine similarity calculating and \\(\\epsilon_{low},\\epsilon_{up}\\) stand for the lower/upper similarity thresholds." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.69, + 0.788, + 0.738 + ], + "angle": 0, + "content": "Optimizing ID/OOD Decision Boundary with Synthetic Samples Once we have obtained the ID and synthetic OOD objects, we employ a lightweight MLP, denoted as \\(\\mathcal{F}_{\\mathrm{ood}}\\), as the OOD detector optimized with a bi-classify loss:" + }, + { + "type": "equation", + "bbox": [ + 0.229, + 0.742, + 0.786, + 0.776 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {o o d}} = \\mathbb {E} _ {\\mathbf {z} \\sim \\mathbf {z} ^ {\\mathrm {i d}}} \\left[ - \\log \\frac {1}{1 + \\exp^ {- \\mathcal {F} _ {\\mathrm {o o d}} (\\mathbf {z})}} \\right] + \\mathbb {E} _ {\\mathbf {z} \\sim \\mathbf {z} ^ {\\mathrm {o o d}}} \\left[ - \\log \\frac {\\exp^ {- \\mathcal {F} _ {\\mathrm {o o d}} (\\mathbf {z})}}{1 + \\exp^ {- \\mathcal {F} _ {\\mathrm {o o d}} (\\mathbf {z})}} \\right]. (6)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The aforementioned design ensures both semantic separability and pattern similarity for the chosen synthetic samples. As a result, our proposed method elegantly optimizes the decision boundary using only a limited number of samples. It is further demonstrated and validated in the following experiments." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.376, + 0.163 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.788, + 0.298 + ], + "angle": 0, + "content": "Datasets Following OOD object detection setting [15], we use the PASCAL-VOC [16] and Berkeley DeepDrive (BDD-100K) [65] as the ID training datasets, which consist of 20 and 10 ID categories, respectively. Meanwhile, we evaluate the performance of our approach on two OOD datasets: MS-COCO [36] and OpenImages [32] respectively. Categories from the OOD datasets that overlapped with the ID datasets are removed to guarantee the absence of ID categories. We report ID categories, OOD categories, and texts used in driving image synthesis in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.305, + 0.788, + 0.427 + ], + "angle": 0, + "content": "Metrics We primarily focus on reporting the FPR95 which represents the false positive rate of OOD samples when the true positive rate of ID samples is at \\(95\\%\\) and lower is better. FPR95 is widely used to assess the OOD object detection performance [12, 15, 59-61]. Additionally, we present the Area Under the Receiver Operating Characteristic Curve (AUROC, higher is better) that is widely utilized to evaluate binary classification problems. Since we only train a plugin MLP on top of existing object detectors for OOD detection, ID performance, e.g., mean Average Precision (mAP) is unchanged and thus omitted." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.788, + 0.632 + ], + "angle": 0, + "content": "Implementation Details For our synthetic data, in order to maintain the experimental setting of OOD object detection and avoid leaking prior knowledge of foundation models, we remove all imagined novel objects that have the same or similar meaning as the ground truth OOD data categories. For model training, we follow the architectures of the baseline method [15, 59] to use a Faster R-CNN [48] as the base object detector with ResNet-50 [20] firstly. Then we trained a simple and lightweight 3-layer MLP. We follow [59] to extract multi-level features as training samples. ID samples are extracted from ID images and OOD samples are extracted from selected synthesized images with OOD bounding boxes. For training on the PASCAL-VOC dataset, we employ a learning rate of 1e-4, while for the BDD-100K dataset, we use a learning rate of 5e-5. Both training processes utilize a momentum of 0.9, a dropout rate of 0.5, and a batch size of 32. All training is conducted on GeForce RTX 3090 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.651, + 0.595, + 0.667 + ], + "angle": 0, + "content": "4.1 Main Results on OOD Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.749 + ], + "angle": 0, + "content": "We evaluate the performance of the proposed method on different challenging benchmarks and obtain notable results (see Tab. 1). As the first work to introduce synthetic scene-level natural images as OOD samples, we incorporate our data-centric method to two off-the-shelf object detectors [15,48], achieving new state-of-the-art performance in OOD object detection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Compared with previous methods, we present comprehensive and substantial performance improvements on FPR95. The encouraging outcomes clearly show that our synthetic data offers superior OOD supervision and are well-suited for forming a precise decision boundary between ID and OOD samples as illustrated in Fig. 1, which significantly reduces the interference caused by contextual information when optimizing the decision boundary." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.326, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Can OOD Object Detectors Learn from Foundation Models?" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.144, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 1: Comparing on varied ID (PASCAL VOC [16], BDD-100K [65]) and OOD (MS-COCO [36], OpenImages [32]) datasets, our method significantly outperforms other methods on different metrics and achieves SOTA performance on OOD object detection. Our method is validated on two different existing object detectors, Faster R-CNN [48] and VOS [15] (denoted as Faster R-CNN + Ours and VOS + Ours respectively). (Top results are shown in bold.)" + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.237, + 0.787, + 0.47 + ], + "angle": 0, + "content": "
MethodID:PASCAL-VOCID:BDD-100K
MS-COCOOpenImagesMS-COCOOpenImages
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
MSP [22]70.9983.4573.1381.9180.9475.8779.0477.38
ODIN [35]59.8282.2063.1482.5962.8574.4458.9276.61
Mahalanobis [34]96.4659.2596.2757.4257.6684.9260.1686.88
Energy score [38]56.8983.6958.6982.9860.0677.4854.7979.60
Gram matrices [50]62.7579.8867.4277.6260.9374.9377.5559.38
Generalized ODIN [25]58.5783.1270.2879.2357.2785.2250.1787.18
CSI [52]59.9181.8357.4182.9547.1084.0937.0687.99
GAN-synthesis [33]60.9383.6759.9782.6757.0378.8250.6181.25
SIREN-KNN [12]47.4589.6750.3888.80----
VOS [15]47.5388.7051.3385.2344.2786.8735.5488.52
SR-VAE [61]42.1790.2846.2687.8932.2390.6921.8193.55
DFDD [60]41.3490.7944.5288.6530.7190.7422.6792.48
SAFE [59]47.4080.3020.0692.2832.5688.9616.0494.64
Faster R-CNN + Ours36.4486.5213.3495.3722.6795.4412.9696.26
VOS + Ours34.9787.9011.2596.9623.0994.3214.1296.41
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.539 + ], + "angle": 0, + "content": "Table 2: Ablation on the number of our synthetic data in training. Taking PASCAL-VOC as the ID dataset, we perform seven groups of random sampling with different numbers in the synthetic dataset to extract features as OOD samples, evaluate and report the performance on the MS-COCO/OpenImages datasets." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.546, + 0.787, + 0.597 + ], + "angle": 0, + "content": "
#Sample14k12k10k8k6k4k2k
FPR95↓36.70/12.9636.27/13.0137.31/13.2536.53/13.3036.70/12.9636.44/13.3437.82/13.87
AUROC↑86.65/95.5486.68/95.5586.64/95.5186.69/95.5286.75/95.5686.52/95.4486.03/95.18
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.723 + ], + "angle": 0, + "content": "Bridged by our synthetic data, foundation models' extensive knowledge and powerful logic about novel concepts are effectively injected into our model through novel concept imagining and region-level editing. Furthermore, powered by the similarity-based filter, our synthetic data proves to be highly effective. Compared with SAFE [59] which uses a similar framework, we only use around \\(25\\%\\) (on PASCAL-VOC) and \\(20\\%\\) (on BDD-100K) of auxiliary data to achieve a significant performance improvement. Further analysis is presented in Sec. 4.2." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.741, + 0.391, + 0.758 + ], + "angle": 0, + "content": "4.2 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Number of Training Samples We conduct an extensive ablation study on the quantity of synthetic data utilized, as illustrated in Tab. 2. We employ seven sets of synthetic data with varying quantities as OOD samples, using PASCAL-VOC as the ID dataset. Features are extracted and the OOD detector is trained based on the same Faster R-CNN checkpoint for each set. It is noteworthy that" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.188 + ], + "angle": 0, + "content": "Table 3: We study the impact of associating varying numbers of imagined novel objects with each ID object. Taking PASCAL-VOC as the ID dataset, we report the performance on MS-COCO/OpenImages datasets." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.196, + 0.787, + 0.253 + ], + "angle": 0, + "content": "
#Sample345678
FPR95↓36.96/13.5837.13/13.1537.31/12.8236.87/13.2537.91/13.5337.13/13.15
AUROC↑86.56/95.5486.63/95.5186.46/95.4786.51/95.3786.43/95.3586.44/95.56
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.266, + 0.788, + 0.321 + ], + "angle": 0, + "content": "Table 4: We randomly sample the same numbers of OOD features as the main experiment instead of using the feature filter (denoted as w/o filter), and evaluate on multiple datasets. The obtained results demonstrate the effectiveness of the proposed data filter." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.331, + 0.787, + 0.42 + ], + "angle": 0, + "content": "
MethodID:PASCAL-VOCID:BDD-100K
MS-COCOOpenImagesMS-COCOOpenImages
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
w/o filter39.2985.4613.6895.2225.4593.1715.3295.83
Ours36.4486.5213.3495.3722.6795.4412.9696.26
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.529 + ], + "angle": 0, + "content": "as the number of samples decreases from 14k to 2k, the performance does not deteriorate but rather maintains stable and superior results. This highlights our method's data efficiency (SAFE used 16k samples). Combined with our feature similarity-based filtering strategy as in Sec. 3.2, a small number of high-quality OOD samples with visual similarity directly promotes the optimization of precise decision boundaries to achieve stable improvements." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.53, + 0.788, + 0.637 + ], + "angle": 0, + "content": "Meanwhile, we use the same detector checkpoint to assess the parallel baseline, SAFE [59], and get the performance of 50.86/23.60 on FPR95 and 78.15/91.42 on AUROC. SAFE augments about 16k images and extracts more than 100k instance-level features as OOD samples. In contrast, our approach utilizes fewer synthetic images and extracts only one instance-level feature from the edited novel object in each synthetic image as an OOD sample, resulting in significantly superior performance compared to SAFE [59]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.649, + 0.788, + 0.771 + ], + "angle": 0, + "content": "Number of Concepts to Imagine We employ in-context learning to guide LLM in associating new objects for driving image editing. For each ID object, the LLM connects a steady stream of novel objects. We further explore the impact of the number of corresponding novel objects for each ID object on data performance. We randomly sample different numbers of novel objects from LLM's responses for each ID object. As shown in Tab. 3, the performance remains stable despite changes in the number of concepts, further highlighting the stability and robustness of our synthetic data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "SAM-based Refiner As mentioned in Sec. 3.1, we propose to utilize SAM-based refiner to correct the bounding boxes of novel objects to obtain higher-quality instance-level OOD features. Therefore, we comparatively remove the proposed refiner and directly used the corresponding editing masks as bound-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.327, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Can OOD Object Detectors Learn from Foundation Models?" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.144, + 0.787, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.458, + 0.788, + 0.542 + ], + "angle": 0, + "content": "Fig. 3: We show cases on six intervals of feature similarity (consistent with Eq. (5), indicated at the bottom of the figure). The first line contains the corresponding initial images, the second line contains the synthetic images with the corresponding boxes of the novel objects (yellow boxes), and the third line contains the difference heat maps of the latent feature maps extracted from the above image pairs (superimposed on the corresponding synthetic images, denoted as Diff-map)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.562, + 0.788, + 0.7 + ], + "angle": 0, + "content": "ing boxes to extract OOD features for training. Taking PASCAL-VOC as the ID dataset, after removing the refiner, we obtain 39.55/13.72 of FPR95 and 85.94/95.37 of AUROC on MS-COCO/OpenImages datasets, which is better than previous methods but worse than the results (Faster R-CNN + Ours in Tab. 1) when using the refiner. This proves that OOD supervision signals contained in the synthetic data are already extracted and achieve good results under the localization of the fuzzy boxes, but more precise boxes mean higher quality features. More demos and analyses of the SAM-based refiner are shown in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Similarity-based Filter The filter is designed to incorporate the most useful data into training, and avoid unnecessary noise. The design is reflected in two aspects: on one hand, the outlier object should process similar visual patterns to the original object, thus being confusing and can facilitate learning; on the other hand, over-high similarity may indicate failures of the editing process (e.g., when the concept is not an object). These considerations are applied as thresholding on pairwise cosine similarity between object features, as in Eq. (5). As shown in Tab. 4, this filter brings a notable improvement across benchmarks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 5: Comparing varied images as OOD samples for training, we first show some synthetic object-centric images generated by Stable Diffusion (left side). Then with PASCAL-VOC as ID dataset, we report the results obtained by using synthetic object-centric images (denoted as object-centric images) and scene-level images with novel objects but without bounding boxes (denoted as scene-level w/o boxes) as OOD samples to participate in training." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.234, + 0.788, + 0.32 + ], + "angle": 0, + "content": "
Object-centric Images DataMS-COCOOpenImages
FPR95↓ AUROC↑FPR95↓ AUROC↑
object-centric images51.9981.4820.7093.85
scene-level w/o boxes48.0182.3818.6193.44
Ours36.4486.5213.3495.37
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.331, + 0.789, + 0.543 + ], + "angle": 0, + "content": "To provide more insights on the choices of filtering thresholds, we display some cases in different intervals of feature similarity in Fig. 3. We show the synthetic images, the corresponding initial images, and the difference between feature maps (denoted as Diff-map), respectively. The Diff-maps prove that the edited area is sensitively attended to by the model. But for images with extremely high similarity \\((>0.9)\\), they always contain some editing failures and blurs. As illustrated on the top of the first column in Fig. 3, it is not intuitively evident what the ship has been edited into (the target object is a raft). Besides, as the similarity upperbound decreases, we progressively obtain more realistic and reasonable images. But note that when the similarity is excessively low, as seen in the last column of Fig. 3, the objects are edited into the corresponding text or an unnatural object, leading to image distortion. This strongly supports the idea that the quality and usability of edited images are closely connected to visual similarity. More cases and analyses are presented in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.562, + 0.53, + 0.577 + ], + "angle": 0, + "content": "4.3 Discussions on Outlier Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.81 + ], + "angle": 0, + "content": "Scene-level Editing Matters Through regional-level editing, we replace the ID object with a novel object with a bounding box and ensure consistent context information. However, some simpler methods based on foundation models also achieve the acquisition and use of OOD data. For example, Dream-OOD [13] uses well-trained text-conditional space and diffusion model to synthesize realistic object-centric data for promoting OOD image classification. Similarly, keeping other settings unchanged, we use our novel concepts to drive Stable-Diffusion instead of Stable-Diffusion-Inpainting [49] to synthesize novel images, which are also processed and filtered by our proposed refiner and filter (some synthesized images are shown in Tab. 5), thereby participating in the training as OOD supervision. However, as shown in Tab. 5 (object-centric images), the synthetic novel object-centric images do not aid in training and result in poor performance, even though they possess high visual quality. This clearly validates our decision to edit scene-level images rather than composing new ones, and highlights the significance of maintaining contextual consistency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Additionally, we examine the possibility of using the edited scene-level image as a whole (ignoring the boxes) as OOD samples in the training process. The" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.327, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Can OOD Object Detectors Learn from Foundation Models?" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.144, + 0.784, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.311, + 0.788, + 0.382 + ], + "angle": 0, + "content": "Fig. 4: We edit the context of the synthetic data (in blue box) so that the images contain novel objects and novel context (in orange box). Then we calculate the similarity between the instance-level feature of the corresponding objects from all synthetic images and the instance-level feature in the initial image (left, the bird). The similarities and the corresponding difference maps are shown in the figure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.395, + 0.789, + 0.442 + ], + "angle": 0, + "content": "results, as depicted in Tab. 5 (scene-level w/o boxes), are significantly inferior compared to our method's performance. This demonstrates that controllable bounding boxes are indispensable in this task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.451, + 0.79, + 0.618 + ], + "angle": 0, + "content": "Context Consistency Matters Given the importance of scene-level synthesis as discussed in the previous paragraph, we study the factors that make scene-level editing indispensable, and find context consistency to be a crucial one. Besides calculating the similarity between ID/OOD object pairs before/after box-conditioned editing, we also try further editing parts of the background of the already edited images, and also calculate its object similarity with the initial object. As shown in Fig. 4, even small editing on parts of the background (out of the object boxes) can make foreground objects 'look' notably different as perceived by the detector. This highlights the importance of keeping the context unchanged when synthesizing outlier samples, in that if the context is changed, the model easily identifies the object as OOD and cannot break the context bias." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.64, + 0.355, + 0.655 + ], + "angle": 0, + "content": "5 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.789, + 0.849 + ], + "angle": 0, + "content": "What Type of OOD Data Matters? We are the first to explore how to edit scene-level images to include novel categories, which contain annotation boxes and ensure context consistency, facilitating OOD object detection. The achieved state-of-the-art performance (Tab. 1) benefits from the optimization of decision boundaries driven by high-quality OOD features. Our exploration demonstrates that annotation boxes and context consistency are particularly important for synthesizing high-quality OOD instances. On the one hand, high-quality annotation boxes provide the possibility to extract high-quality instance-level features from scene-level images, while unrefined boxes (Sec. 4.2) or discarded boxes (Tab. 5) will have a negative impact on performance. On the other hand, context consistency ensures the most effective OOD features are not interfered by different contexts and selected for utilizing (Fig. 4)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.146, + 0.787, + 0.298 + ], + "angle": 0, + "content": "How to Synthesize Suitable OOD Data? We are the first to build an automatic, transparent, and low-cost pipeline (Sec. 3.1) for synthesizing scene-level images containing novel objects with annotation boxes and context consistency. It benefits object detectors' robustness and reliability to unseen data and sets up clear state-of-the-art on multiple OOD object detection benchmarks. Specifically, we organically combine and cleverly use different foundation models [1, 30, 49] (Sec. 3.1) to distill open-world knowledge and inpaint the existing scenes for simulating real OOD scenarios. In addition, our design takes into account the instability of the current foundation models and can release better potential performance in the future development of foundation models." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.313, + 0.787, + 0.419 + ], + "angle": 0, + "content": "How to Select Suitable OOD Data? We are the first to explicitly decouple OOD data synthesis and selection. On the one hand, we ensure the separability of the synthetic objects in semantic concepts through open-world knowledge provided by LLMs (Sec. 3.1). On the other hand, we ensure the similarity of ID and OOD objects in visual patterns (Sec. 3.1), thereby optimizing the precise decision boundary. This line of thinking has the potential to facilitate more open-world solutions." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.434, + 0.787, + 0.587 + ], + "angle": 0, + "content": "Broader Impacts Beyond showcasing engineering success via effectively combining specific foundation models, our work uncovers the untapped potential of the text-to-image generative models and visual foundation models in pushing forward the OOD object detection task to effectively leverage the off-the-shelf open-world data knowledge [5,31,66]. More importantly, our work establishes a bridge between OOD object detection and the latest advancements in deep learning, enabling it to benefit from ongoing developments, go beyond isolated academic practice, and resolve practical challenges in open-world applications. Meanwhile, automating novel data generation and curation will inspire more tasks in more modalities, such as in visual-language [44,58] and 3D vision [8,37,64]." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.609, + 0.359, + 0.625 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.641, + 0.787, + 0.837 + ], + "angle": 0, + "content": "In this paper, we investigate improving OOD object detection by distilling open-world data knowledge from text-to-image generative models. We develop an automatic and cost-effective data curation pipeline, SyncOOD, that leverages foundation models as tools to obtain meaningful open-set data from generative models. Through extensive studies, we discover that object boxes and context consistency of the generated data contribute to the improvement of OOD object detection performance. Our comprehensive experiments demonstrate that SyncOOD not only advances the state-of-the-art in OOD object detection but also emphasizes the untapped potential of utilizing large-scale generative models for enhancing the robustness of machine learning systems in open-world settings. As an initial exploration in leveraging foundation models for OOD object detection, we hope our promising results encourage further research in advancing this area in the future." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.326, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Can OOD Object Detectors Learn from Foundation Models?" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.18, + 0.788, + 0.286 + ], + "angle": 0, + "content": "This work has been supported by Hong Kong Research Grant Council - Early Career Scheme (Grant No. 27209621), General Research Fund Scheme (Grant No. 17202422), Theme-based Research (Grant No. T45-701/22-R) and RGC Matching Fund Scheme (RMGS). Part of the described research work is conducted in the JC STEM Lab of Robotics for Soft Materials funded by The Hong Kong Jockey Club Charities Trust. We would like to thank Chirui Chang, Xiaoyang Lyu, Haoru Tan, and Xiuzhe Wu for their insightful discussions." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.313, + 0.323, + 0.329 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.348, + 0.785, + 0.389 + ], + "angle": 0, + "content": "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: GPT-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.391, + 0.785, + 0.432 + ], + "angle": 0, + "content": "2. Bendale, A., Boult, T.E.: Towards open set deep networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1563-1572 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.46 + ], + "angle": 0, + "content": "3. Bishop, C.M.: Novelty detection and neural network validation. IEE Proceedings-Vision, Image and Signal processing 141(4), 217-222 (1994)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.462, + 0.785, + 0.502 + ], + "angle": 0, + "content": "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in Neural Information Processing Systems 33, 1877-1901 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.504, + 0.785, + 0.53 + ], + "angle": 0, + "content": "5. Chang, C., Liu, Z., Lyu, X., Qi, X.: What matters in detecting ai-generated videos like sora? arXiv preprint arXiv:2406.19568 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.532, + 0.785, + 0.585 + ], + "angle": 0, + "content": "6. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. Journal of Machine Learning Research 24(240), 1-113 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.588, + 0.785, + 0.614 + ], + "angle": 0, + "content": "7. Deepshikha, K., Yelleni, S.H., Srijith, P., Mohan, C.K.: Monte Carlo dropblock for modelling uncertainty in object detection. arXiv preprint arXiv:2108.03614 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.616, + 0.785, + 0.656 + ], + "angle": 0, + "content": "8. Deng, W., Ding, R., Yang, J., Liu, J., Li, Y., Qi, X., Ngai, E.: Can 3d vision-language models truly understand natural language? arXiv preprint arXiv:2403.14760 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.785, + 0.698 + ], + "angle": 0, + "content": "9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.701, + 0.785, + 0.741 + ], + "angle": 0, + "content": "10. Dhamija, A., Gunther, M., Ventura, J., Boult, T.: The overlooked elephant of object detection: Open set. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1021-1030 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.743, + 0.785, + 0.769 + ], + "angle": 0, + "content": "11. Du, X., Fang, Z., Diakonikolas, I., Li, Y.: How does unlabeled data provably help out-of-distribution detection? arXiv preprint arXiv:2402.03502 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.771, + 0.785, + 0.811 + ], + "angle": 0, + "content": "12. Du, X., Gozum, G., Ming, Y., Li, Y.: SIREN: Shaping representations for detecting out-of-distribution objects. Advances in Neural Information Processing Systems 35, 20434-20449 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. Du, X., Sun, Y., Zhu, J., Li, Y.: Dream the impossible: Outlier imagination with diffusion models. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.348, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "14. Du, X., Wang, X., Gozum, G., Li, Y.: Unknown-aware object detection: Learning what you don't know from videos in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13678-13688 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.217 + ], + "angle": 0, + "content": "15. Du, X., Wang, Z., Cai, M., Li, Y.: VOS: Learning what you don't know by virtual outlier synthesis. arXiv preprint arXiv:2202.01197 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.258 + ], + "angle": 0, + "content": "16. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes (VOC) challenge. International journal of computer vision 88, 303-338 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.259, + 0.785, + 0.3 + ], + "angle": 0, + "content": "17. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "18. Gupta, A., Narayan, S., Joseph, K., Khan, S., Khan, F.S., Shah, M.: Ow-detr: Open-world detection transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9235-9244 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.342, + 0.785, + 0.397 + ], + "angle": 0, + "content": "19. Hall, D., Dayoub, F., Skinner, J., Zhang, H., Miller, D., Corke, P., Carneiro, G., Angelova, A., Sünderhauf, N.: Probabilistic object detection: Definition and evaluation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1031-1040 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "21. He, R., Sun, S., Yu, X., Xue, C., Zhang, W., Torr, P., Bai, S., Qi, X.: Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "22. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. In: International Conference on Learning Representations (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.55 + ], + "angle": 0, + "content": "23. Hendrycks, D., Mazeika, M., Dietterich, T.: Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "24. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.578, + 0.785, + 0.632 + ], + "angle": 0, + "content": "25. Hsu, Y.C., Shen, Y., Jin, H., Kira, Z.: Generalized ODIN: Detecting out-of-distribution image without learning from out-of-distribution data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10951-10960 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "26. Huang, R., Geng, A., Li, Y.: On the importance of gradients for detecting distributional shifts in the wild. Advances in Neural Information Processing Systems 34, 677-689 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.785, + 0.715 + ], + "angle": 0, + "content": "27. Joseph, K., Khan, S., Khan, F.S., Balasubramanian, V.N.: Towards open world object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5830-5840 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.758 + ], + "angle": 0, + "content": "28. Joseph, K., Khan, S., Khan, F.S., Balasubramanian, V.N.: Towards open world object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5830-5840 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "29. Katz-Samuels, J., Nakhleh, J.B., Nowak, R., Li, Y.: Training ood detectors in their natural habitats. In: International Conference on Machine Learning. pp. 10848-10865. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "30. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.327, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Can OOD Object Detectors Learn from Foundation Models?" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.19 + ], + "angle": 0, + "content": "31. Kong, S., Ramanan, D.: Opengan: Open-set recognition via open data generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 813-822 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.191, + 0.787, + 0.245 + ], + "angle": 0, + "content": "32. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision 128(7), 1956-1981 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.247, + 0.787, + 0.288 + ], + "angle": 0, + "content": "33. Lee, K., Lee, H., Lee, K., Shin, J.: Training confidence-calibrated classifiers for detecting out-of-distribution samples. In: International Conference on Learning Representations (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.787, + 0.331 + ], + "angle": 0, + "content": "34. Lee, K., Lee, K., Lee, H., Shin, J.: A simple unified framework for detecting out-of-distribution samples and adversarial attacks. Advances in Neural Information Processing Systems 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.787, + 0.374 + ], + "angle": 0, + "content": "35. Liang, S., Li, Y., Srikant, R.: Enhancing the reliability of out-of-distribution image detection in neural networks. In: International Conference on Learning Representations, ICLR 2018 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.787, + 0.43 + ], + "angle": 0, + "content": "36. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.787, + 0.486 + ], + "angle": 0, + "content": "37. Liu, J., Chang, C., Liu, J., Wu, X., Ma, L., Qi, X.: Mars3d: A plug-and-play motion-aware model for semantic segmentation on multi-scan 3d point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9372-9381 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.487, + 0.787, + 0.514 + ], + "angle": 0, + "content": "38. Liu, W., Wang, X., Owens, J., Li, Y.: Energy-based out-of-distribution detection. Advances in Neural Information Processing Systems 33, 21464-21475 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.787, + 0.543 + ], + "angle": 0, + "content": "39. Liu, X., Yang, H., Ravichandran, A., Bhotika, R., Soatto, S.: Continual universal object detection. arXiv preprint arXiv:2002.05347 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.545, + 0.787, + 0.585 + ], + "angle": 0, + "content": "40. Miller, D., Dayoub, F., Milford, M., Sünderhauf, N.: Evaluating merging strategies for sampling-based uncertainty techniques in object detection. In: International Conference on Robotics and Automation (ICRA). pp. 2348-2354 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.787, + 0.628 + ], + "angle": 0, + "content": "41. Miller, D., Nicholson, L., Dayoub, F., Sünderhauf, N.: Dropout sampling for robust object detection in open-set conditions. In: IEEE International Conference on Robotics and Automation (ICRA). pp. 3243-3249 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.787, + 0.67 + ], + "angle": 0, + "content": "42. Nguyen, A., Yosinski, J., Clune, J.: Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 427-436 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.787, + 0.712 + ], + "angle": 0, + "content": "43. Perez-Rua, J.M., Zhu, X., Hospedales, T.M., Xiang, T.: Incremental few-shot object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13846-13855 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.787, + 0.769 + ], + "angle": 0, + "content": "44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.77, + 0.787, + 0.798 + ], + "angle": 0, + "content": "45. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "46. Rahman, S., Khan, S.H., Porikli, F.: Zero-shot object detection: Joint recognition and localization of novel concepts. International Journal of Computer Vision 128, 2979-2999 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "47. Ren, J., Fort, S., Liu, J., Roy, A.G., Padhy, S., Lakshminarayanan, B.: A simple fix to Mahalanobis distance for improving near-ood detection. arXiv preprint arXiv:2106.09022 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.191, + 0.785, + 0.232 + ], + "angle": 0, + "content": "48. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems 28 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.785, + 0.274 + ], + "angle": 0, + "content": "49. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.785, + 0.303 + ], + "angle": 0, + "content": "50. Sastry, C.S., Oore, S.: Detecting out-of-distribution examples with gram matrices. In: International Conference on Machine Learning. pp. 8491-8501. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.785, + 0.345 + ], + "angle": 0, + "content": "51. Sun, Y., Ming, Y., Zhu, X., Li, Y.: Out-of-distribution detection with deep nearest neighbors. In: International Conference on Machine Learning. pp. 20827-20840. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.347, + 0.785, + 0.387 + ], + "angle": 0, + "content": "52. Tack, J., Mo, S., Jeong, J., Shin, J.: CSI: Novelty detection via contrastive learning on distributionally shifted instances. Advances in Neural Information Processing Systems 33, 11839-11852 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.785, + 0.43 + ], + "angle": 0, + "content": "53. Tan, H., Wu, S., Du, F., Chen, Y., Wang, Z., Wang, F., Qi, X.: Data pruning via moving-one-sample-out. Advances in Neural Information Processing Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.432, + 0.785, + 0.459 + ], + "angle": 0, + "content": "54. Tao, L., Du, X., Zhu, X., Li, Y.: Non-parametric outlier synthesis. arXiv preprint arXiv:2303.02966 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.501 + ], + "angle": 0, + "content": "55. Tian, Y., Fan, L., Chen, K., Katabi, D., Krishnan, D., Isola, P.: Learning vision from models rivals learning vision from data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15887-15898 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.543 + ], + "angle": 0, + "content": "56. Wang, H., Liu, W., Bocchieri, A., Li, Y.: Can multi-label classification networks know what they don't know? Advances in Neural Information Processing Systems 34, 29074-29087 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.545, + 0.785, + 0.599 + ], + "angle": 0, + "content": "57. Wang, X., Huang, T.E., Liu, B., Yu, F., Wang, X., Gonzalez, J.E., Darrell, T.: Robust object detection via instance-level temporal cycle confusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9143-9152 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.785, + 0.642 + ], + "angle": 0, + "content": "58. Wen, X., Zhao, B., Chen, Y., Pang, J., Qi, X.: Generalization beyond data imbalance: A controlled study on clip for transferable insights. arXiv preprint arXiv:2405.21070 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.785, + 0.685 + ], + "angle": 0, + "content": "59. Wilson, S., Fischer, T., Dayoub, F., Miller, D., Sunderhauf, N.: SAFE: Sensitivity-aware features for out-of-distribution object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23565-23576 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.785, + 0.727 + ], + "angle": 0, + "content": "60. Wu, A., Chen, D., Deng, C.: Deep feature deblurring diffusion for detecting out-of-distribution objects. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13381-13391 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.785, + 0.769 + ], + "angle": 0, + "content": "61. Wu, A., Deng, C.: Discriminating known from unknown objects via structure-enhanced recurrent variational autoencoder. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23956-23965 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.77, + 0.785, + 0.798 + ], + "angle": 0, + "content": "62. Wu, Q., Chen, Y., Yang, C., Yan, J.: Energy-based out-of-distribution detection for graph neural networks. arXiv preprint arXiv:2302.02914 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "63. Wu, S., Tan, H., Tian, Z., Chen, Y., Qi, X., Jia, J.: Saco loss: Sample-wise affinity consistency for vision-language pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 27358-27369 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.327, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Can OOD Object Detectors Learn from Foundation Models?" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "64. Wu, X., Dai, P., Deng, W., Chen, H., Wu, Y., Cao, Y.P., Shan, Y., Qi, X.: Clnerf: continual learning of neural radiance fields for evolving scene representation. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "65. Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F., Madhavan, V., Darrell, T.: BDD100K: A diverse driving dataset for heterogeneous multitask learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2636-2645 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "66. Zheng, H., Wang, Q., Fang, Z., Xia, X., Liu, F., Liu, T., Han, B.: Out-of-distribution detection learning with unreliable out-of-distribution sources. Advances in Neural Information Processing Systems 36, 72110-72123 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.286 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Can OOD Object Detectors Learn from Foundation Models_/6d5daec6-da83-48aa-809b-d1c5b6b09141_origin.pdf b/2024/Can OOD Object Detectors Learn from Foundation Models_/6d5daec6-da83-48aa-809b-d1c5b6b09141_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..09f92e0e30c77d4a9a1316691d66a133713449a2 --- /dev/null +++ b/2024/Can OOD Object Detectors Learn from Foundation Models_/6d5daec6-da83-48aa-809b-d1c5b6b09141_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fbf96329a2616187ee7e06b14164987644faef185e370ab9fadd4a36a01b9cb4 +size 5719887 diff --git a/2024/Can OOD Object Detectors Learn from Foundation Models_/full.md b/2024/Can OOD Object Detectors Learn from Foundation Models_/full.md new file mode 100644 index 0000000000000000000000000000000000000000..63e82811fa65d2638fc1332d09715e0f28e38afe --- /dev/null +++ b/2024/Can OOD Object Detectors Learn from Foundation Models_/full.md @@ -0,0 +1,273 @@ +# Can OOD Object Detectors Learn from Foundation Models? + +Jiahui Liu, Xin Wen, Shizhen Zhao, Yingxian Chen, and Xiaojuan Qi* + +The University of Hong Kong {liujh,wenxin,zhaosz,chenyx,xjqi}@eee.hku.hk + +Abstract. Out-of-distribution (OOD) object detection is a challenging task due to the absence of open-set OOD data. Inspired by recent advancements in text-to-image generative models, such as Stable Diffusion, we study the potential of generative models trained on large-scale open-set data to synthesize OOD samples, thereby enhancing OOD object detection. We introduce SyncOOD, a simple data curation method that capitalizes on the capabilities of large foundation models to automatically extract meaningful OOD data from text-to-image generative models. This offers the model access to open-world knowledge encapsulated within off-the-shelf foundation models. The synthetic OOD samples are then employed to augment the training of a lightweight, plug-and-play OOD detector, thus effectively optimizing the in-distribution (ID)/OOD decision boundaries. Extensive experiments across multiple benchmarks demonstrate that SyncOOD significantly outperforms existing methods, establishing new state-of-the-art performance with minimal synthetic data usage. The project is available at https://github.com/CVMI-Lab/SyncOOD. + +Keywords: OOD object detection $\cdot$ Synthetic data $\cdot$ Open-world data + +# 1 Introduction + +Modern object detectors, trained on closed-set data, have achieved remarkable success. However, they often incorrectly and confidently classify out-of-distribution (OOD) object categories as in-distribution (ID) categories in open-world applications [10], raising concerns about their reliability for deployment. To enhance the trustworthiness of object detection, researchers have studied the OOD object detection task, which aims to identify and flag unknown or novel objects as distinct from ID ones [3,42]. + +The vulnerability of these models to OOD samples stems from their lack of awareness of the unknown open data distribution during training. Consequently, synthesizing OOD samples for model learning has emerged as a major research direction for this task [15, 59-61]. Most existing studies [15, 60, 61] concentrate on generating OOD objects in the latent space of an object detection model + +![](images/7e55f29f4617ab4ee325d9ec919e7a071bb49f89a2bd311b222457a5741a3ab2.jpg) +Fig. 1: Our pipeline replaces ID objects with semantic-novel yet visual-similar objects for scene-level OOD object synthesis. Middle left: The concepts are imagined by an LLM to ensure semantic separability and rationality, and reformed as text prompts for controllable in-painting using Stable Diffusion. Middle right: During training, only visually similar OOD objects are adopted based on instance-level feature similarity to the original object. A lightweight binary classifier is optimized for OOD detection, and other parts of the detector are kept unchanged. + +trained on ID data. These synthesized samples are then used to optimize the decision boundary between ID and OOD data. Alternative approaches include directly synthesizing images by injecting adversarial noise [59] or identifying OOD instances from video data [14]. While these methods have yielded promising results, they remain limited to a closed-set setting, where the latent space for synthesizing outliers or the data is derived from a closed-set data distribution. Consequently, they may be biased towards the ID dataset, leading to suboptimal performance. Besides, effectively handling the unknown may appear unattainable when the unknown is never fully exploitable. Beyond that, is it possible to learn them from massive open-world data knowledge condensed in foundation models? + +To this end, we investigate text-to-image generation models trained on a large amount of open-set data, which have demonstrated a superior ability to capture the distribution of data across a wide range of visual concepts, in order to synthesize novel data samples for enhancing OOD detection. Nonetheless, automatically extracting meaningful data from generative models for OOD object detection remains challenging due to the extensive vocabulary space to be explored, the complex scene-level synthesis problem, the need for object-level annotations, and potential distractions from contextual cues. + +We introduce SyncOOD, an automatic data curation process that leverages foundation models as tools to harvest meaningful data from text-to-image gen- + +eration models for OOD object detection (see Figure 1). The process is based on two key observations: 1) Hard OOD samples that are close to the ID data contribute more to learning a better OOD detector, and 2) Context may become a distracting cue for OOD object detection tasks, leading to bias towards contexts. With these observations in mind, the outlier synthesis process is formulated as box-conditioned image inpainting, and driven by Stable Diffusion [49] for high-quality controllable editing. The concepts to replace with are imagined by a Large Language Model (LLM) [1] with the aim of semantic novelty, $^{1}$ and the associated bounding box is further refined with SAM [30]. Automated by foundation models, this data collection pipeline requires minimal human labor, while producing high-quality OOD data. + +In comparison to existing methods for OOD object detection, our core insight is to broaden the model's exposure to a more extensive range of open-set data and circumvent dataset biases by tapping into the open-world knowledge found in off-the-shelf foundation models. Utilizing generative models also provides us with control over the context of synthesized images and the data distribution. Our comprehensive experiments demonstrate superior performance, emphasizing the untapped potential of text-image-generation models in the context of OOD object detection. Our key contributions are summarized as follows: + +- We investigate and unlock the potential of text-to-image generative models trained on large-scale open-set data for synthesizing OOD objects in object detection tasks. +- We introduce an automated data curation process for obtaining controllable, annotated scene-level synthetic OOD images for OOD object detection, which utilizes LLMs for novel concept discovery and visual foundation models for data annotation and filtering. +- We discover that maintaining ID/OOD image context consistency and obtaining more accurate OOD annotation bounding boxes are crucial for synthesized data to be effective in OOD object detection. +- Comprehensive experiments on multiple benchmarks demonstrate the effectiveness of our method, as we significantly outperform existing state-of-the-art approaches while using minimal synthetic data. + +# 2 Related Work + +OOD Object Detection For detecting OOD objects in scene-level images, unlike earlier works that constrain ID samples to a hypothetical distribution [12], it has become a recent trend to explicitly synthesize the outlier data, and incorporate them into training to adjust models' decision boundaries. However, due to the complexity of scene-level images, all previous works bypassed photorealistic outlier synthesis in pixel space, and worked on generating outliers from the model's latent space [15, 60, 61], adversarial attack [59], or utilizing video data in the wild [14]. In the former, outliers can be sampled from the latent + +space using a simple Gaussian prior [15], or more advanced generative models like VAE [61] or diffusion model [60]. Yet their upperbound are commonly limited by the quality of the latent space, and synthesized samples lack interpretability. Despite this, adversarial samples [59] lack semantic diversity, and auxiliary pseudo supervision from videos [14] introduces additional requirements to the setting. Unlike all above, our method applies LLM for OOD concept sampling, and Stable Diffusion for controllable image editing. This decouples the sampling and generation processes, elevates both parts to an unprecedented level, and achieves photo-realistic scene-level OOD image synthesis for the first time. + +Open-world Object Detection The fact that real-world applications require object detectors the ability to tackle open categories is also considered in open-world object detection. The focus of this field includes generalization to domain shifts [57], incremental learning of novel classes [27, 39, 43], and zero-shot classification of open-vocabularies [17, 46]. Meanwhile, some works [18, 28] require to distinguish known objects well, be able to detect unknown objects, and finally be able to incrementally learn new objects. These works provide different perspectives on facing the challenges of real-world applications, which complement OOD object detection as a joint effort, but are out of the scope of this paper. + +OOD Image Classification Earlier paradigms for OOD image classification either post hoc adjust models' confidence score at the testing phase, or apply regularization at models' training phase. The former line mainly focuses on the design of score functions, including confidence-based [2, 22, 35], energy-based [38, 56, 62], distance-based [34, 47, 50, 51], gradient-based [26], and approximating Bayesian [7, 10, 19, 40, 41] methods. The latter line of work includes regularizing models to produce lower confidence [23, 33], higher energy [29, 38], or directly shaping latent representations [12]. While outlier synthesis has shown to be effective by [15, 54], these are still generated in the latent space, and a parallel line study utilizing natural images [11, 23, 29] from the wild. Recently, photo-realistic outlier synthesis was first achieved by [13] with help from a text-conditioned diffusion model [49]. However, it is not readily applicable for object detection due to the complexity of scene-level images and the requirement for object-level annotations. In contrast, our work studies under the detection setting and requires outlier synthesis at scene level. + +Foundation Models The evolution of large language models (LLMs) began with training on web-scale datasets [9,45], leading to increasingly powerful foundation models [4, 6] capable of harnessing vast open-world data. Notable advancements include models [1] that interact with users and perform complex tasks like question answering, significantly broadening access to global knowledge. In image generation, diffusion models [24, 49] offer robust capabilities in synthesizing realistic content for applications such as image synthesis and inpainting. Additionally, segmentation foundation models like SAM [30] represent a leap forward in precise image segmentation, benefiting from extensive data training. Foundation models provide diverse data, which provides unlimited po + +![](images/599d0ae175dbc499dcd032372e9eec73fbf99c0109a180a9b04be407c86659eb.jpg) +Fig. 2: Detailed illustration of our outlier synthesis pipeline. It comprises (a) Instructing an LLM to imagine semantic-novel concepts given ID objects, (b) Editing the selected regions to the expected concepts via prompt-conditioned image inpainting using Stable Diffusion, and (c) Refining the bounding boxes of edited objects using SAM. + +tential for learning open-world knowledge from these models [55], where the effectiveness of data [21, 53, 63] is also crucial to the downstreaming tasks. + +# 3 Method + +Preliminary For OOD object detection, the training set contains only ID scene-level images $\mathbf{x}^{\mathrm{id}}$ with ID objects, annotation bounding boxes $\mathbf{b}^{\mathrm{id}}$ , and semantic labels $y$ , denoted as $\mathcal{D}_{\mathrm{id}} = \left\{(\mathbf{x}^{\mathrm{id}}, \mathbf{b}^{\mathrm{id}}, \mathbf{y}^{\mathrm{id}})\right\}$ . The labels of ID objects always belong to a close set with $K$ categories, denoted as $\mathbf{y}^{\mathrm{id}} \in \mathcal{Y}_{\mathrm{id}}$ and $\mathcal{Y}_{\mathrm{id}} = \{y_1^{\mathrm{id}}, y_2^{\mathrm{id}}, \dots, y_K^{\mathrm{id}}\}$ . During inference, for each proposed object from an input scene-level image, it is required to identify whether its category belongs to $\mathcal{Y}_{\mathrm{id}}$ or not. + +Overview As illustrated in Fig. 1, our outlier synthesis pipeline consists of (1) synthesizing a set of effective photo-realistic scene-level OOD images $\mathbf{x}^{\mathrm{edit}}$ , denoted as $\mathcal{D}_{\mathrm{edit}} = \left\{(\mathbf{x}^{\mathrm{edit}},\mathbf{b}^{\mathrm{edit}})\right\}$ , which contains novel objects and corresponding annotation boxes $\mathbf{b}^{\mathrm{edit}}$ based on region-level editing from $\mathcal{D}_{\mathrm{id}}$ in a fully automated, labor-free way; and (2) select and use the efficient synthetic data to provide pseudo-OOD supervisions for training OOD object detector together with the ID samples in the training set. Further design of the pipeline requires answering the following questions: (1) how to distill the open-set knowledge embedded in foundation models to scene-level OOD data and (2) how to utilize the synthesized data to regularize the decision boundary and facilitate OOD object detection. We discuss them accordingly in Sec. 3.1 and Sec. 3.2. + +# 3.1 Synthesizing Semantic-novel Objects in Scene Images + +Imagining Novel Concepts from ID objects As shown in Fig. 2(a), based on the ID labels $\mathcal{V}_{\mathrm{id}}$ in training set $\mathcal{D}_{\mathrm{id}}$ , we consider that novel concepts that are different from ID categories can be potential candidates for generating OOD + +objects. The next is to discover novel concepts that offer hard OOD samples sharing high visual similarity with ID samples and being contextually compatible with the scene context for object detection. Rather than relying on human labor to investigate all potential candidates, we leverage the vast knowledge and reasoning capabilities of LLM, GPT-4 [1] to check the visual similarity and contextual compatibility. This allows us to associate ID objects and promote the conceptualization of possible novel objects to replace existing ID objects through the use of a prompt with in-context examples [4] as: + +Here is a list containing several objects $\mathcal{V}_{\mathrm{id}}$ . Now, if I provide you an object name, you should return to me objects that are similar to the usage scenario and volume of the provided object but are not in the previous object list. For example, if I give you the word: person, you should respond and only respond: 'mannequin', 'sculpture', 'scarecrows', 'doll', 'puppet'. + +With its robust logical foundation and rich knowledge, the LLM envisions a collection of novel objects for each ID object label, denoted as $\mathcal{Y}_{\mathrm{novel}}$ , while maintaining the semantic separability between imagined objects and ID objects. We empirically find one in-context example that is sufficient for us to discover novel concepts. For each ID label $i$ , we discover $M$ novel concepts using LLM $\mathbf{y}_i^{\mathrm{novel}}$ of $M$ concepts. + +Editing Objects on Selected Regions With the discovered novel concepts $\mathcal{V}_{\mathrm{novel}} = \{\mathbf{y}_1^{\mathrm{novel}},\mathbf{y}_2^{\mathrm{novel}},\dots,\mathbf{y}_K^{\mathrm{novel}}\}$ , the next step is to use them as prompts for the text-to-image generation model to generate an image. To generate a new image with novel concepts $y_{j}\in \mathbf{y}_{i}^{\mathrm{novel}}$ , we choose to replace existing ID objects in existing images with label $y_{i}^{\mathrm{id}}$ instead of finding new locations or generating one image from scratch. By doing so, we can ensure context compatibility and eliminate distractions from the scene context as it is preserved. As illustrated in Fig. 2 (b), we use Stable-Diffusion-Inpainting [49], denoted as $\mathrm{SDI}(\cdot)$ , to perform region-level editing on ID images. The ID object is denoted as $\mathbf{x}^{\mathrm{id}}$ , with its corresponding annotation box $\mathbf{b}^{\mathrm{id}}$ serving as the editing mask, and the associated imagined novel concept $\mathbf{y}^{\mathrm{novel}}$ are provided as inputs to the SDI, which is one of the most successful models for conditional image generation and editing. Thus, an edited image $\mathbf{x}^{\mathrm{edit}}$ containing a novel object is obtained as: + +$$ +\mathbf {x} ^ {\text {e d i t}} = \operatorname {S D I} \left(\mathbf {x} ^ {\text {i d}}, \mathbf {b} ^ {\text {i d}}, \mathbf {y} ^ {\text {n o v e l}}\right). \tag {1} +$$ + +Refining Annotation Boxes of Novel Objects Due to the randomness in diffusion models, the attributes of edited objects, such as their quality, volume, and localization, may not match the original object box. To address this issue, as depicted in Fig. 2(c), we design an efficient and effective refiner based on SAM [30] to obtain refined accurate bounding boxes on novel objects. First, for an edited image $\mathbf{x}^{\mathrm{edit}}$ with the editing mask $\mathbf{b}^{\mathrm{id}}$ , we use a padding area extended from $\mathbf{b}^{\mathrm{id}}$ as the prompt and employ SAM to output the instance mask with + +highest confidence $\mathbf{m}^{\mathrm{SAM}}$ for the novel object in the area: + +$$ +\mathbf {m} ^ {\text {S A M}} = \operatorname {S A M} \left(\mathbf {x} ^ {\text {e d i t}}; \text {p a d d i n g} \left(\mathbf {b} ^ {\text {i d}}, e\right)\right), \tag {2} +$$ + +where $e$ represents the range of padding. Then, we convert obtained masks $\mathbf{m}^{\mathrm{SAM}}$ to boxes $\mathbf{b}^{\mathrm{SAM}}$ , and calculate IoU between $\mathbf{b}^{\mathrm{SAM}}$ and the corresponding $\mathbf{b}^{\mathrm{id}}$ to filter out novel objects that vary highly in scale: + +$$ +\left\{\mathbf {b} ^ {\text {e d i t}} \right\} = \left\{\mathbf {b} ^ {\text {S A M}} \mid \operatorname {I o U} \left(\mathbf {b} ^ {\text {S A M}}, \mathbf {b} ^ {\text {i d}}\right) > \gamma \right\}, \tag {3} +$$ + +where $\gamma$ denotes a threshold value on IoU. It ensures a high enough recall rate to rule out the instability of Stable Diffusion and SAM and uncontrollable localization of the edited objects. Thus we obtain the synthetic outlier data $\mathcal{D}_{\mathrm{edit}}$ as illustrated in Fig. 1. + +# 3.2 Mining Hard OOD Samples and Model Training + +Mining Hard OOD Objects with High Visual Similarities for Training We consider the novel objects that are most likely to be confused with the corresponding ID objects by the object detector as the most effective ones. We thus aim to find synthetic OOD samples that are most easily confused as ID to participate in training based on pairwise similarity in the latent space of the pre-trained object detector. For each novel object with bounding box $\mathbf{b}^{\mathrm{edit}}$ in the synthetic data $\mathcal{D}_{\mathrm{edit}}$ , we construct it with the corresponding original ID object with its bounding box as a pair: $\left\{(\mathbf{b}^{\mathrm{edit}}, \mathbf{x}^{\mathrm{edit}}), (\mathbf{b}^{\mathrm{id}}, \mathbf{x}^{\mathrm{id}})\right\}$ . For an off-the-shelf object detector, denoted by $\mathcal{F}_{\mathrm{det}}$ , we extract latent features, $\mathbf{z}^{\mathrm{edit}}$ and $\mathbf{z}^{\mathrm{id}}$ , for each pair: + +$$ +\mathbf {z} ^ {\text {e d i t}}, \mathbf {z} ^ {\text {i d}} = \mathcal {F} _ {\det } \left(\mathbf {b} ^ {\text {e d i t}}; \mathbf {x} ^ {\text {e d i t}}\right), \mathcal {F} _ {\det } \left(\mathbf {b} ^ {\text {i d}}; \mathbf {x} ^ {\text {i d}}\right). \tag {4} +$$ + +The most effective novel objects are those with visual patterns that can be easily mistaken for their corresponding ID objects by an object detector. Therefore, we filter these novel objects based on their similarity to provide pseudo-OD supervision: + +$$ +\left\{\mathbf {z} ^ {\mathrm {o o d}} \right\} = \left\{\mathbf {z} ^ {\mathrm {e d i t}} \big | \epsilon_ {l o w} < \mathrm {s i m} (\mathbf {z} ^ {\mathrm {e d i t}}, \mathbf {z} ^ {\mathrm {i d}}) < \epsilon_ {u p} \right\}, \qquad \qquad (5) +$$ + +where the similarities are computed between latent object features of edit-ID pairs. Here $\mathrm{sim}(\cdot)$ denotes cosine similarity calculating and $\epsilon_{low},\epsilon_{up}$ stand for the lower/upper similarity thresholds. + +Optimizing ID/OOD Decision Boundary with Synthetic Samples Once we have obtained the ID and synthetic OOD objects, we employ a lightweight MLP, denoted as $\mathcal{F}_{\mathrm{ood}}$ , as the OOD detector optimized with a bi-classify loss: + +$$ +\mathcal {L} _ {\mathrm {o o d}} = \mathbb {E} _ {\mathbf {z} \sim \mathbf {z} ^ {\mathrm {i d}}} \left[ - \log \frac {1}{1 + \exp^ {- \mathcal {F} _ {\mathrm {o o d}} (\mathbf {z})}} \right] + \mathbb {E} _ {\mathbf {z} \sim \mathbf {z} ^ {\mathrm {o o d}}} \left[ - \log \frac {\exp^ {- \mathcal {F} _ {\mathrm {o o d}} (\mathbf {z})}}{1 + \exp^ {- \mathcal {F} _ {\mathrm {o o d}} (\mathbf {z})}} \right]. (6) +$$ + +The aforementioned design ensures both semantic separability and pattern similarity for the chosen synthetic samples. As a result, our proposed method elegantly optimizes the decision boundary using only a limited number of samples. It is further demonstrated and validated in the following experiments. + +# 4 Experiments + +Datasets Following OOD object detection setting [15], we use the PASCAL-VOC [16] and Berkeley DeepDrive (BDD-100K) [65] as the ID training datasets, which consist of 20 and 10 ID categories, respectively. Meanwhile, we evaluate the performance of our approach on two OOD datasets: MS-COCO [36] and OpenImages [32] respectively. Categories from the OOD datasets that overlapped with the ID datasets are removed to guarantee the absence of ID categories. We report ID categories, OOD categories, and texts used in driving image synthesis in the supplementary material. + +Metrics We primarily focus on reporting the FPR95 which represents the false positive rate of OOD samples when the true positive rate of ID samples is at $95\%$ and lower is better. FPR95 is widely used to assess the OOD object detection performance [12, 15, 59-61]. Additionally, we present the Area Under the Receiver Operating Characteristic Curve (AUROC, higher is better) that is widely utilized to evaluate binary classification problems. Since we only train a plugin MLP on top of existing object detectors for OOD detection, ID performance, e.g., mean Average Precision (mAP) is unchanged and thus omitted. + +Implementation Details For our synthetic data, in order to maintain the experimental setting of OOD object detection and avoid leaking prior knowledge of foundation models, we remove all imagined novel objects that have the same or similar meaning as the ground truth OOD data categories. For model training, we follow the architectures of the baseline method [15, 59] to use a Faster R-CNN [48] as the base object detector with ResNet-50 [20] firstly. Then we trained a simple and lightweight 3-layer MLP. We follow [59] to extract multi-level features as training samples. ID samples are extracted from ID images and OOD samples are extracted from selected synthesized images with OOD bounding boxes. For training on the PASCAL-VOC dataset, we employ a learning rate of 1e-4, while for the BDD-100K dataset, we use a learning rate of 5e-5. Both training processes utilize a momentum of 0.9, a dropout rate of 0.5, and a batch size of 32. All training is conducted on GeForce RTX 3090 GPUs. + +# 4.1 Main Results on OOD Object Detection + +We evaluate the performance of the proposed method on different challenging benchmarks and obtain notable results (see Tab. 1). As the first work to introduce synthetic scene-level natural images as OOD samples, we incorporate our data-centric method to two off-the-shelf object detectors [15,48], achieving new state-of-the-art performance in OOD object detection. + +Compared with previous methods, we present comprehensive and substantial performance improvements on FPR95. The encouraging outcomes clearly show that our synthetic data offers superior OOD supervision and are well-suited for forming a precise decision boundary between ID and OOD samples as illustrated in Fig. 1, which significantly reduces the interference caused by contextual information when optimizing the decision boundary. + +Table 1: Comparing on varied ID (PASCAL VOC [16], BDD-100K [65]) and OOD (MS-COCO [36], OpenImages [32]) datasets, our method significantly outperforms other methods on different metrics and achieves SOTA performance on OOD object detection. Our method is validated on two different existing object detectors, Faster R-CNN [48] and VOS [15] (denoted as Faster R-CNN + Ours and VOS + Ours respectively). (Top results are shown in bold.) + +
MethodID:PASCAL-VOCID:BDD-100K
MS-COCOOpenImagesMS-COCOOpenImages
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
MSP [22]70.9983.4573.1381.9180.9475.8779.0477.38
ODIN [35]59.8282.2063.1482.5962.8574.4458.9276.61
Mahalanobis [34]96.4659.2596.2757.4257.6684.9260.1686.88
Energy score [38]56.8983.6958.6982.9860.0677.4854.7979.60
Gram matrices [50]62.7579.8867.4277.6260.9374.9377.5559.38
Generalized ODIN [25]58.5783.1270.2879.2357.2785.2250.1787.18
CSI [52]59.9181.8357.4182.9547.1084.0937.0687.99
GAN-synthesis [33]60.9383.6759.9782.6757.0378.8250.6181.25
SIREN-KNN [12]47.4589.6750.3888.80----
VOS [15]47.5388.7051.3385.2344.2786.8735.5488.52
SR-VAE [61]42.1790.2846.2687.8932.2390.6921.8193.55
DFDD [60]41.3490.7944.5288.6530.7190.7422.6792.48
SAFE [59]47.4080.3020.0692.2832.5688.9616.0494.64
Faster R-CNN + Ours36.4486.5213.3495.3722.6795.4412.9696.26
VOS + Ours34.9787.9011.2596.9623.0994.3214.1296.41
+ +Table 2: Ablation on the number of our synthetic data in training. Taking PASCAL-VOC as the ID dataset, we perform seven groups of random sampling with different numbers in the synthetic dataset to extract features as OOD samples, evaluate and report the performance on the MS-COCO/OpenImages datasets. + +
#Sample14k12k10k8k6k4k2k
FPR95↓36.70/12.9636.27/13.0137.31/13.2536.53/13.3036.70/12.9636.44/13.3437.82/13.87
AUROC↑86.65/95.5486.68/95.5586.64/95.5186.69/95.5286.75/95.5686.52/95.4486.03/95.18
+ +Bridged by our synthetic data, foundation models' extensive knowledge and powerful logic about novel concepts are effectively injected into our model through novel concept imagining and region-level editing. Furthermore, powered by the similarity-based filter, our synthetic data proves to be highly effective. Compared with SAFE [59] which uses a similar framework, we only use around $25\%$ (on PASCAL-VOC) and $20\%$ (on BDD-100K) of auxiliary data to achieve a significant performance improvement. Further analysis is presented in Sec. 4.2. + +# 4.2 Ablation Study + +Number of Training Samples We conduct an extensive ablation study on the quantity of synthetic data utilized, as illustrated in Tab. 2. We employ seven sets of synthetic data with varying quantities as OOD samples, using PASCAL-VOC as the ID dataset. Features are extracted and the OOD detector is trained based on the same Faster R-CNN checkpoint for each set. It is noteworthy that + +Table 3: We study the impact of associating varying numbers of imagined novel objects with each ID object. Taking PASCAL-VOC as the ID dataset, we report the performance on MS-COCO/OpenImages datasets. + +
#Sample345678
FPR95↓36.96/13.5837.13/13.1537.31/12.8236.87/13.2537.91/13.5337.13/13.15
AUROC↑86.56/95.5486.63/95.5186.46/95.4786.51/95.3786.43/95.3586.44/95.56
+ +Table 4: We randomly sample the same numbers of OOD features as the main experiment instead of using the feature filter (denoted as w/o filter), and evaluate on multiple datasets. The obtained results demonstrate the effectiveness of the proposed data filter. + +
MethodID:PASCAL-VOCID:BDD-100K
MS-COCOOpenImagesMS-COCOOpenImages
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
w/o filter39.2985.4613.6895.2225.4593.1715.3295.83
Ours36.4486.5213.3495.3722.6795.4412.9696.26
+ +as the number of samples decreases from 14k to 2k, the performance does not deteriorate but rather maintains stable and superior results. This highlights our method's data efficiency (SAFE used 16k samples). Combined with our feature similarity-based filtering strategy as in Sec. 3.2, a small number of high-quality OOD samples with visual similarity directly promotes the optimization of precise decision boundaries to achieve stable improvements. + +Meanwhile, we use the same detector checkpoint to assess the parallel baseline, SAFE [59], and get the performance of 50.86/23.60 on FPR95 and 78.15/91.42 on AUROC. SAFE augments about 16k images and extracts more than 100k instance-level features as OOD samples. In contrast, our approach utilizes fewer synthetic images and extracts only one instance-level feature from the edited novel object in each synthetic image as an OOD sample, resulting in significantly superior performance compared to SAFE [59]. + +Number of Concepts to Imagine We employ in-context learning to guide LLM in associating new objects for driving image editing. For each ID object, the LLM connects a steady stream of novel objects. We further explore the impact of the number of corresponding novel objects for each ID object on data performance. We randomly sample different numbers of novel objects from LLM's responses for each ID object. As shown in Tab. 3, the performance remains stable despite changes in the number of concepts, further highlighting the stability and robustness of our synthetic data. + +SAM-based Refiner As mentioned in Sec. 3.1, we propose to utilize SAM-based refiner to correct the bounding boxes of novel objects to obtain higher-quality instance-level OOD features. Therefore, we comparatively remove the proposed refiner and directly used the corresponding editing masks as bound- + +![](images/fe7cb7fe213e39e1e75e33c1560abc363efc2324b6b698df79d021f04c9be644.jpg) +Fig. 3: We show cases on six intervals of feature similarity (consistent with Eq. (5), indicated at the bottom of the figure). The first line contains the corresponding initial images, the second line contains the synthetic images with the corresponding boxes of the novel objects (yellow boxes), and the third line contains the difference heat maps of the latent feature maps extracted from the above image pairs (superimposed on the corresponding synthetic images, denoted as Diff-map). + +ing boxes to extract OOD features for training. Taking PASCAL-VOC as the ID dataset, after removing the refiner, we obtain 39.55/13.72 of FPR95 and 85.94/95.37 of AUROC on MS-COCO/OpenImages datasets, which is better than previous methods but worse than the results (Faster R-CNN + Ours in Tab. 1) when using the refiner. This proves that OOD supervision signals contained in the synthetic data are already extracted and achieve good results under the localization of the fuzzy boxes, but more precise boxes mean higher quality features. More demos and analyses of the SAM-based refiner are shown in the supplementary material. + +Similarity-based Filter The filter is designed to incorporate the most useful data into training, and avoid unnecessary noise. The design is reflected in two aspects: on one hand, the outlier object should process similar visual patterns to the original object, thus being confusing and can facilitate learning; on the other hand, over-high similarity may indicate failures of the editing process (e.g., when the concept is not an object). These considerations are applied as thresholding on pairwise cosine similarity between object features, as in Eq. (5). As shown in Tab. 4, this filter brings a notable improvement across benchmarks. + +Table 5: Comparing varied images as OOD samples for training, we first show some synthetic object-centric images generated by Stable Diffusion (left side). Then with PASCAL-VOC as ID dataset, we report the results obtained by using synthetic object-centric images (denoted as object-centric images) and scene-level images with novel objects but without bounding boxes (denoted as scene-level w/o boxes) as OOD samples to participate in training. + +
Object-centric Images DataMS-COCOOpenImages
FPR95↓ AUROC↑FPR95↓ AUROC↑
object-centric images51.9981.4820.7093.85
scene-level w/o boxes48.0182.3818.6193.44
Ours36.4486.5213.3495.37
+ +To provide more insights on the choices of filtering thresholds, we display some cases in different intervals of feature similarity in Fig. 3. We show the synthetic images, the corresponding initial images, and the difference between feature maps (denoted as Diff-map), respectively. The Diff-maps prove that the edited area is sensitively attended to by the model. But for images with extremely high similarity $(>0.9)$ , they always contain some editing failures and blurs. As illustrated on the top of the first column in Fig. 3, it is not intuitively evident what the ship has been edited into (the target object is a raft). Besides, as the similarity upperbound decreases, we progressively obtain more realistic and reasonable images. But note that when the similarity is excessively low, as seen in the last column of Fig. 3, the objects are edited into the corresponding text or an unnatural object, leading to image distortion. This strongly supports the idea that the quality and usability of edited images are closely connected to visual similarity. More cases and analyses are presented in the supplementary material. + +# 4.3 Discussions on Outlier Synthesis + +Scene-level Editing Matters Through regional-level editing, we replace the ID object with a novel object with a bounding box and ensure consistent context information. However, some simpler methods based on foundation models also achieve the acquisition and use of OOD data. For example, Dream-OOD [13] uses well-trained text-conditional space and diffusion model to synthesize realistic object-centric data for promoting OOD image classification. Similarly, keeping other settings unchanged, we use our novel concepts to drive Stable-Diffusion instead of Stable-Diffusion-Inpainting [49] to synthesize novel images, which are also processed and filtered by our proposed refiner and filter (some synthesized images are shown in Tab. 5), thereby participating in the training as OOD supervision. However, as shown in Tab. 5 (object-centric images), the synthetic novel object-centric images do not aid in training and result in poor performance, even though they possess high visual quality. This clearly validates our decision to edit scene-level images rather than composing new ones, and highlights the significance of maintaining contextual consistency. + +Additionally, we examine the possibility of using the edited scene-level image as a whole (ignoring the boxes) as OOD samples in the training process. The + +![](images/a748bb0d397ed352e10074919107cb8a6c4546aec7ae6d4bd06cf5eeeef22710.jpg) +Fig. 4: We edit the context of the synthetic data (in blue box) so that the images contain novel objects and novel context (in orange box). Then we calculate the similarity between the instance-level feature of the corresponding objects from all synthetic images and the instance-level feature in the initial image (left, the bird). The similarities and the corresponding difference maps are shown in the figure. + +results, as depicted in Tab. 5 (scene-level w/o boxes), are significantly inferior compared to our method's performance. This demonstrates that controllable bounding boxes are indispensable in this task. + +Context Consistency Matters Given the importance of scene-level synthesis as discussed in the previous paragraph, we study the factors that make scene-level editing indispensable, and find context consistency to be a crucial one. Besides calculating the similarity between ID/OOD object pairs before/after box-conditioned editing, we also try further editing parts of the background of the already edited images, and also calculate its object similarity with the initial object. As shown in Fig. 4, even small editing on parts of the background (out of the object boxes) can make foreground objects 'look' notably different as perceived by the detector. This highlights the importance of keeping the context unchanged when synthesizing outlier samples, in that if the context is changed, the model easily identifies the object as OOD and cannot break the context bias. + +# 5 Discussion + +What Type of OOD Data Matters? We are the first to explore how to edit scene-level images to include novel categories, which contain annotation boxes and ensure context consistency, facilitating OOD object detection. The achieved state-of-the-art performance (Tab. 1) benefits from the optimization of decision boundaries driven by high-quality OOD features. Our exploration demonstrates that annotation boxes and context consistency are particularly important for synthesizing high-quality OOD instances. On the one hand, high-quality annotation boxes provide the possibility to extract high-quality instance-level features from scene-level images, while unrefined boxes (Sec. 4.2) or discarded boxes (Tab. 5) will have a negative impact on performance. On the other hand, context consistency ensures the most effective OOD features are not interfered by different contexts and selected for utilizing (Fig. 4). + +How to Synthesize Suitable OOD Data? We are the first to build an automatic, transparent, and low-cost pipeline (Sec. 3.1) for synthesizing scene-level images containing novel objects with annotation boxes and context consistency. It benefits object detectors' robustness and reliability to unseen data and sets up clear state-of-the-art on multiple OOD object detection benchmarks. Specifically, we organically combine and cleverly use different foundation models [1, 30, 49] (Sec. 3.1) to distill open-world knowledge and inpaint the existing scenes for simulating real OOD scenarios. In addition, our design takes into account the instability of the current foundation models and can release better potential performance in the future development of foundation models. + +How to Select Suitable OOD Data? We are the first to explicitly decouple OOD data synthesis and selection. On the one hand, we ensure the separability of the synthetic objects in semantic concepts through open-world knowledge provided by LLMs (Sec. 3.1). On the other hand, we ensure the similarity of ID and OOD objects in visual patterns (Sec. 3.1), thereby optimizing the precise decision boundary. This line of thinking has the potential to facilitate more open-world solutions. + +Broader Impacts Beyond showcasing engineering success via effectively combining specific foundation models, our work uncovers the untapped potential of the text-to-image generative models and visual foundation models in pushing forward the OOD object detection task to effectively leverage the off-the-shelf open-world data knowledge [5,31,66]. More importantly, our work establishes a bridge between OOD object detection and the latest advancements in deep learning, enabling it to benefit from ongoing developments, go beyond isolated academic practice, and resolve practical challenges in open-world applications. Meanwhile, automating novel data generation and curation will inspire more tasks in more modalities, such as in visual-language [44,58] and 3D vision [8,37,64]. + +# 6 Conclusion + +In this paper, we investigate improving OOD object detection by distilling open-world data knowledge from text-to-image generative models. We develop an automatic and cost-effective data curation pipeline, SyncOOD, that leverages foundation models as tools to obtain meaningful open-set data from generative models. Through extensive studies, we discover that object boxes and context consistency of the generated data contribute to the improvement of OOD object detection performance. Our comprehensive experiments demonstrate that SyncOOD not only advances the state-of-the-art in OOD object detection but also emphasizes the untapped potential of utilizing large-scale generative models for enhancing the robustness of machine learning systems in open-world settings. As an initial exploration in leveraging foundation models for OOD object detection, we hope our promising results encourage further research in advancing this area in the future. + +# Acknowledgments + +This work has been supported by Hong Kong Research Grant Council - Early Career Scheme (Grant No. 27209621), General Research Fund Scheme (Grant No. 17202422), Theme-based Research (Grant No. T45-701/22-R) and RGC Matching Fund Scheme (RMGS). Part of the described research work is conducted in the JC STEM Lab of Robotics for Soft Materials funded by The Hong Kong Jockey Club Charities Trust. We would like to thank Chirui Chang, Xiaoyang Lyu, Haoru Tan, and Xiuzhe Wu for their insightful discussions. + +# References + +1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: GPT-4 technical report. arXiv preprint arXiv:2303.08774 (2023) +2. Bendale, A., Boult, T.E.: Towards open set deep networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1563-1572 (2016) +3. Bishop, C.M.: Novelty detection and neural network validation. IEE Proceedings-Vision, Image and Signal processing 141(4), 217-222 (1994) +4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in Neural Information Processing Systems 33, 1877-1901 (2020) +5. Chang, C., Liu, Z., Lyu, X., Qi, X.: What matters in detecting ai-generated videos like sora? arXiv preprint arXiv:2406.19568 (2024) +6. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. Journal of Machine Learning Research 24(240), 1-113 (2023) +7. Deepshikha, K., Yelleni, S.H., Srijith, P., Mohan, C.K.: Monte Carlo dropblock for modelling uncertainty in object detection. arXiv preprint arXiv:2108.03614 (2021) +8. Deng, W., Ding, R., Yang, J., Liu, J., Li, Y., Qi, X., Ngai, E.: Can 3d vision-language models truly understand natural language? arXiv preprint arXiv:2403.14760 (2024) +9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) +10. Dhamija, A., Gunther, M., Ventura, J., Boult, T.: The overlooked elephant of object detection: Open set. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1021-1030 (2020) +11. Du, X., Fang, Z., Diakonikolas, I., Li, Y.: How does unlabeled data provably help out-of-distribution detection? arXiv preprint arXiv:2402.03502 (2024) +12. Du, X., Gozum, G., Ming, Y., Li, Y.: SIREN: Shaping representations for detecting out-of-distribution objects. Advances in Neural Information Processing Systems 35, 20434-20449 (2022) +13. Du, X., Sun, Y., Zhu, J., Li, Y.: Dream the impossible: Outlier imagination with diffusion models. Advances in Neural Information Processing Systems 36 (2024) + +14. Du, X., Wang, X., Gozum, G., Li, Y.: Unknown-aware object detection: Learning what you don't know from videos in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13678-13688 (2022) +15. Du, X., Wang, Z., Cai, M., Li, Y.: VOS: Learning what you don't know by virtual outlier synthesis. arXiv preprint arXiv:2202.01197 (2022) +16. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes (VOC) challenge. International journal of computer vision 88, 303-338 (2010) +17. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: International Conference on Learning Representations (2022) +18. Gupta, A., Narayan, S., Joseph, K., Khan, S., Khan, F.S., Shah, M.: Ow-detr: Open-world detection transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9235-9244 (2022) +19. Hall, D., Dayoub, F., Skinner, J., Zhang, H., Miller, D., Corke, P., Carneiro, G., Angelova, A., Sünderhauf, N.: Probabilistic object detection: Definition and evaluation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1031-1040 (2020) +20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016) +21. He, R., Sun, S., Yu, X., Xue, C., Zhang, W., Torr, P., Bai, S., Qi, X.: Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574 (2022) +22. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. In: International Conference on Learning Representations (2017) +23. Hendrycks, D., Mazeika, M., Dietterich, T.: Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606 (2018) +24. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020) +25. Hsu, Y.C., Shen, Y., Jin, H., Kira, Z.: Generalized ODIN: Detecting out-of-distribution image without learning from out-of-distribution data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10951-10960 (2020) +26. Huang, R., Geng, A., Li, Y.: On the importance of gradients for detecting distributional shifts in the wild. Advances in Neural Information Processing Systems 34, 677-689 (2021) +27. Joseph, K., Khan, S., Khan, F.S., Balasubramanian, V.N.: Towards open world object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5830-5840 (2021) +28. Joseph, K., Khan, S., Khan, F.S., Balasubramanian, V.N.: Towards open world object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5830-5840 (2021) +29. Katz-Samuels, J., Nakhleh, J.B., Nowak, R., Li, Y.: Training ood detectors in their natural habitats. In: International Conference on Machine Learning. pp. 10848-10865. PMLR (2022) +30. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) + +31. Kong, S., Ramanan, D.: Opengan: Open-set recognition via open data generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 813-822 (2021) +32. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision 128(7), 1956-1981 (2020) +33. Lee, K., Lee, H., Lee, K., Shin, J.: Training confidence-calibrated classifiers for detecting out-of-distribution samples. In: International Conference on Learning Representations (2018) +34. Lee, K., Lee, K., Lee, H., Shin, J.: A simple unified framework for detecting out-of-distribution samples and adversarial attacks. Advances in Neural Information Processing Systems 31 (2018) +35. Liang, S., Li, Y., Srikant, R.: Enhancing the reliability of out-of-distribution image detection in neural networks. In: International Conference on Learning Representations, ICLR 2018 (2018) +36. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +37. Liu, J., Chang, C., Liu, J., Wu, X., Ma, L., Qi, X.: Mars3d: A plug-and-play motion-aware model for semantic segmentation on multi-scan 3d point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9372-9381 (2023) +38. Liu, W., Wang, X., Owens, J., Li, Y.: Energy-based out-of-distribution detection. Advances in Neural Information Processing Systems 33, 21464-21475 (2020) +39. Liu, X., Yang, H., Ravichandran, A., Bhotika, R., Soatto, S.: Continual universal object detection. arXiv preprint arXiv:2002.05347 (2020) +40. Miller, D., Dayoub, F., Milford, M., Sünderhauf, N.: Evaluating merging strategies for sampling-based uncertainty techniques in object detection. In: International Conference on Robotics and Automation (ICRA). pp. 2348-2354 (2019) +41. Miller, D., Nicholson, L., Dayoub, F., Sünderhauf, N.: Dropout sampling for robust object detection in open-set conditions. In: IEEE International Conference on Robotics and Automation (ICRA). pp. 3243-3249 (2018) +42. Nguyen, A., Yosinski, J., Clune, J.: Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 427-436 (2015) +43. Perez-Rua, J.M., Zhu, X., Hospedales, T.M., Xiang, T.: Incremental few-shot object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13846-13855 (2020) +44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +45. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019) +46. Rahman, S., Khan, S.H., Porikli, F.: Zero-shot object detection: Joint recognition and localization of novel concepts. International Journal of Computer Vision 128, 2979-2999 (2020) + +47. Ren, J., Fort, S., Liu, J., Roy, A.G., Padhy, S., Lakshminarayanan, B.: A simple fix to Mahalanobis distance for improving near-ood detection. arXiv preprint arXiv:2106.09022 (2021) +48. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems 28 (2015) +49. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +50. Sastry, C.S., Oore, S.: Detecting out-of-distribution examples with gram matrices. In: International Conference on Machine Learning. pp. 8491-8501. PMLR (2020) +51. Sun, Y., Ming, Y., Zhu, X., Li, Y.: Out-of-distribution detection with deep nearest neighbors. In: International Conference on Machine Learning. pp. 20827-20840. PMLR (2022) +52. Tack, J., Mo, S., Jeong, J., Shin, J.: CSI: Novelty detection via contrastive learning on distributionally shifted instances. Advances in Neural Information Processing Systems 33, 11839-11852 (2020) +53. Tan, H., Wu, S., Du, F., Chen, Y., Wang, Z., Wang, F., Qi, X.: Data pruning via moving-one-sample-out. Advances in Neural Information Processing Systems (2023) +54. Tao, L., Du, X., Zhu, X., Li, Y.: Non-parametric outlier synthesis. arXiv preprint arXiv:2303.02966 (2023) +55. Tian, Y., Fan, L., Chen, K., Katabi, D., Krishnan, D., Isola, P.: Learning vision from models rivals learning vision from data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15887-15898 (2024) +56. Wang, H., Liu, W., Bocchieri, A., Li, Y.: Can multi-label classification networks know what they don't know? Advances in Neural Information Processing Systems 34, 29074-29087 (2021) +57. Wang, X., Huang, T.E., Liu, B., Yu, F., Wang, X., Gonzalez, J.E., Darrell, T.: Robust object detection via instance-level temporal cycle confusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9143-9152 (2021) +58. Wen, X., Zhao, B., Chen, Y., Pang, J., Qi, X.: Generalization beyond data imbalance: A controlled study on clip for transferable insights. arXiv preprint arXiv:2405.21070 (2024) +59. Wilson, S., Fischer, T., Dayoub, F., Miller, D., Sunderhauf, N.: SAFE: Sensitivity-aware features for out-of-distribution object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23565-23576 (2023) +60. Wu, A., Chen, D., Deng, C.: Deep feature deblurring diffusion for detecting out-of-distribution objects. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13381-13391 (2023) +61. Wu, A., Deng, C.: Discriminating known from unknown objects via structure-enhanced recurrent variational autoencoder. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23956-23965 (2023) +62. Wu, Q., Chen, Y., Yang, C., Yan, J.: Energy-based out-of-distribution detection for graph neural networks. arXiv preprint arXiv:2302.02914 (2023) +63. Wu, S., Tan, H., Tian, Z., Chen, Y., Qi, X., Jia, J.: Saco loss: Sample-wise affinity consistency for vision-language pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 27358-27369 (2024) + +64. Wu, X., Dai, P., Deng, W., Chen, H., Wu, Y., Cao, Y.P., Shan, Y., Qi, X.: Clnerf: continual learning of neural radiance fields for evolving scene representation. Advances in Neural Information Processing Systems 36 (2024) +65. Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F., Madhavan, V., Darrell, T.: BDD100K: A diverse driving dataset for heterogeneous multitask learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2636-2645 (2020) +66. Zheng, H., Wang, Q., Fang, Z., Xia, X., Liu, F., Liu, T., Han, B.: Out-of-distribution detection learning with unreliable out-of-distribution sources. Advances in Neural Information Processing Systems 36, 72110-72123 (2023) \ No newline at end of file diff --git a/2024/Can OOD Object Detectors Learn from Foundation Models_/images.zip b/2024/Can OOD Object Detectors Learn from Foundation Models_/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1e32304fe8d100694f53227de78149acbe1fa5ff --- /dev/null +++ b/2024/Can OOD Object Detectors Learn from Foundation Models_/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cfaef29842de2573b022f9e5c7f458cc8f18f47e30b3c5641228de55787505fe +size 612536 diff --git a/2024/Can OOD Object Detectors Learn from Foundation Models_/layout.json b/2024/Can OOD Object Detectors Learn from Foundation Models_/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..00be933556a88e2154a35f3ce9807a9daceeff91 --- /dev/null +++ b/2024/Can OOD Object Detectors Learn from Foundation Models_/layout.json @@ -0,0 +1,8217 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 168, + 111, + 445, + 145 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 111, + 445, + 145 + ], + "spans": [ + { + "bbox": [ + 168, + 111, + 445, + 145 + ], + "type": "text", + "content": "Can OOD Object Detectors Learn from Foundation Models?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 151, + 168, + 462, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 168, + 462, + 181 + ], + "spans": [ + { + "bbox": [ + 151, + 168, + 462, + 181 + ], + "type": "text", + "content": "Jiahui Liu, Xin Wen, Shizhen Zhao, Yingxian Chen, and Xiaojuan Qi*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 202, + 190, + 412, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 190, + 412, + 213 + ], + "spans": [ + { + "bbox": [ + 202, + 190, + 412, + 213 + ], + "type": "text", + "content": "The University of Hong Kong {liujh,wenxin,zhaosz,chenyx,xjqi}@eee.hku.hk" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 244, + 452, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 244, + 452, + 431 + ], + "spans": [ + { + "bbox": [ + 160, + 244, + 452, + 431 + ], + "type": "text", + "content": "Abstract. Out-of-distribution (OOD) object detection is a challenging task due to the absence of open-set OOD data. Inspired by recent advancements in text-to-image generative models, such as Stable Diffusion, we study the potential of generative models trained on large-scale open-set data to synthesize OOD samples, thereby enhancing OOD object detection. We introduce SyncOOD, a simple data curation method that capitalizes on the capabilities of large foundation models to automatically extract meaningful OOD data from text-to-image generative models. This offers the model access to open-world knowledge encapsulated within off-the-shelf foundation models. The synthetic OOD samples are then employed to augment the training of a lightweight, plug-and-play OOD detector, thus effectively optimizing the in-distribution (ID)/OOD decision boundaries. Extensive experiments across multiple benchmarks demonstrate that SyncOOD significantly outperforms existing methods, establishing new state-of-the-art performance with minimal synthetic data usage. The project is available at https://github.com/CVMI-Lab/SyncOOD." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 442, + 451, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 442, + 451, + 453 + ], + "spans": [ + { + "bbox": [ + 160, + 442, + 451, + 453 + ], + "type": "text", + "content": "Keywords: OOD object detection " + }, + { + "bbox": [ + 160, + 442, + 451, + 453 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 442, + 451, + 453 + ], + "type": "text", + "content": " Synthetic data " + }, + { + "bbox": [ + 160, + 442, + 451, + 453 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 442, + 451, + 453 + ], + "type": "text", + "content": " Open-world data" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 475, + 230, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 475, + 230, + 488 + ], + "spans": [ + { + "bbox": [ + 133, + 475, + 230, + 488 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 502, + 482, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 585 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 585 + ], + "type": "text", + "content": "Modern object detectors, trained on closed-set data, have achieved remarkable success. However, they often incorrectly and confidently classify out-of-distribution (OOD) object categories as in-distribution (ID) categories in open-world applications [10], raising concerns about their reliability for deployment. To enhance the trustworthiness of object detection, researchers have studied the OOD object detection task, which aims to identify and flag unknown or novel objects as distinct from ID ones [3,42]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": "The vulnerability of these models to OOD samples stems from their lack of awareness of the unknown open data distribution during training. Consequently, synthesizing OOD samples for model learning has emerged as a major research direction for this task [15, 59-61]. Most existing studies [15, 60, 61] concentrate on generating OOD objects in the latent space of an object detection model" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 654, + 231, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 654, + 231, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 654, + 231, + 665 + ], + "type": "text", + "content": "* corresponding author" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 481, + 328 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 481, + 328 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 481, + 328 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 481, + 328 + ], + "type": "image", + "image_path": "7e55f29f4617ab4ee325d9ec919e7a071bb49f89a2bd311b222457a5741a3ab2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 334, + 482, + 411 + ], + "lines": [ + { + "bbox": [ + 130, + 334, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 482, + 411 + ], + "type": "text", + "content": "Fig. 1: Our pipeline replaces ID objects with semantic-novel yet visual-similar objects for scene-level OOD object synthesis. Middle left: The concepts are imagined by an LLM to ensure semantic separability and rationality, and reformed as text prompts for controllable in-painting using Stable Diffusion. Middle right: During training, only visually similar OOD objects are adopted based on instance-level feature similarity to the original object. A lightweight binary classifier is optimized for OOD detection, and other parts of the detector are kept unchanged." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 544 + ], + "type": "text", + "content": "trained on ID data. These synthesized samples are then used to optimize the decision boundary between ID and OOD data. Alternative approaches include directly synthesizing images by injecting adversarial noise [59] or identifying OOD instances from video data [14]. While these methods have yielded promising results, they remain limited to a closed-set setting, where the latent space for synthesizing outliers or the data is derived from a closed-set data distribution. Consequently, they may be biased towards the ID dataset, leading to suboptimal performance. Besides, effectively handling the unknown may appear unattainable when the unknown is never fully exploitable. Beyond that, is it possible to learn them from massive open-world data knowledge condensed in foundation models?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "content": "To this end, we investigate text-to-image generation models trained on a large amount of open-set data, which have demonstrated a superior ability to capture the distribution of data across a wide range of visual concepts, in order to synthesize novel data samples for enhancing OOD detection. Nonetheless, automatically extracting meaningful data from generative models for OOD object detection remains challenging due to the extensive vocabulary space to be explored, the complex scene-level synthesis problem, the need for object-level annotations, and potential distractions from contextual cues." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "We introduce SyncOOD, an automatic data curation process that leverages foundation models as tools to harvest meaningful data from text-to-image gen-" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "eration models for OOD object detection (see Figure 1). The process is based on two key observations: 1) Hard OOD samples that are close to the ID data contribute more to learning a better OOD detector, and 2) Context may become a distracting cue for OOD object detection tasks, leading to bias towards contexts. With these observations in mind, the outlier synthesis process is formulated as box-conditioned image inpainting, and driven by Stable Diffusion [49] for high-quality controllable editing. The concepts to replace with are imagined by a Large Language Model (LLM) [1] with the aim of semantic novelty," + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": " and the associated bounding box is further refined with SAM [30]. Automated by foundation models, this data collection pipeline requires minimal human labor, while producing high-quality OOD data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 248, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 344 + ], + "type": "text", + "content": "In comparison to existing methods for OOD object detection, our core insight is to broaden the model's exposure to a more extensive range of open-set data and circumvent dataset biases by tapping into the open-world knowledge found in off-the-shelf foundation models. Utilizing generative models also provides us with control over the context of synthesized images and the data distribution. Our comprehensive experiments demonstrate superior performance, emphasizing the untapped potential of text-image-generation models in the context of OOD object detection. Our key contributions are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 353, + 480, + 507 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 353, + 479, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 479, + 387 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 479, + 387 + ], + "type": "text", + "content": "- We investigate and unlock the potential of text-to-image generative models trained on large-scale open-set data for synthesizing OOD objects in object detection tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 388, + 480, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 480, + 435 + ], + "type": "text", + "content": "- We introduce an automated data curation process for obtaining controllable, annotated scene-level synthetic OOD images for OOD object detection, which utilizes LLMs for novel concept discovery and visual foundation models for data annotation and filtering." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 436, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 436, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 138, + 436, + 480, + 472 + ], + "type": "text", + "content": "- We discover that maintaining ID/OOD image context consistency and obtaining more accurate OOD annotation bounding boxes are crucial for synthesized data to be effective in OOD object detection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 472, + 480, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 480, + 507 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 480, + 507 + ], + "type": "text", + "content": "- Comprehensive experiments on multiple benchmarks demonstrate the effectiveness of our method, as we significantly outperform existing state-of-the-art approaches while using minimal synthetic data." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 525, + 237, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 237, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 237, + 538 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "text", + "content": "OOD Object Detection For detecting OOD objects in scene-level images, unlike earlier works that constrain ID samples to a hypothetical distribution [12], it has become a recent trend to explicitly synthesize the outlier data, and incorporate them into training to adjust models' decision boundaries. However, due to the complexity of scene-level images, all previous works bypassed photorealistic outlier synthesis in pixel space, and worked on generating outliers from the model's latent space [15, 60, 61], adversarial attack [59], or utilizing video data in the wild [14]. In the former, outliers can be sampled from the latent" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 199, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 199, + 91, + 448, + 102 + ], + "type": "text", + "content": "Can OOD Object Detectors Learn from Foundation Models?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 479, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 479, + 665 + ], + "type": "text", + "content": "1 Concepts overlapping with the test data are removed to avoid information leakage." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "space using a simple Gaussian prior [15], or more advanced generative models like VAE [61] or diffusion model [60]. Yet their upperbound are commonly limited by the quality of the latent space, and synthesized samples lack interpretability. Despite this, adversarial samples [59] lack semantic diversity, and auxiliary pseudo supervision from videos [14] introduces additional requirements to the setting. Unlike all above, our method applies LLM for OOD concept sampling, and Stable Diffusion for controllable image editing. This decouples the sampling and generation processes, elevates both parts to an unprecedented level, and achieves photo-realistic scene-level OOD image synthesis for the first time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 344 + ], + "type": "text", + "content": "Open-world Object Detection The fact that real-world applications require object detectors the ability to tackle open categories is also considered in open-world object detection. The focus of this field includes generalization to domain shifts [57], incremental learning of novel classes [27, 39, 43], and zero-shot classification of open-vocabularies [17, 46]. Meanwhile, some works [18, 28] require to distinguish known objects well, be able to detect unknown objects, and finally be able to incrementally learn new objects. These works provide different perspectives on facing the challenges of real-world applications, which complement OOD object detection as a joint effort, but are out of the scope of this paper." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 354, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 534 + ], + "type": "text", + "content": "OOD Image Classification Earlier paradigms for OOD image classification either post hoc adjust models' confidence score at the testing phase, or apply regularization at models' training phase. The former line mainly focuses on the design of score functions, including confidence-based [2, 22, 35], energy-based [38, 56, 62], distance-based [34, 47, 50, 51], gradient-based [26], and approximating Bayesian [7, 10, 19, 40, 41] methods. The latter line of work includes regularizing models to produce lower confidence [23, 33], higher energy [29, 38], or directly shaping latent representations [12]. While outlier synthesis has shown to be effective by [15, 54], these are still generated in the latent space, and a parallel line study utilizing natural images [11, 23, 29] from the wild. Recently, photo-realistic outlier synthesis was first achieved by [13] with help from a text-conditioned diffusion model [49]. However, it is not readily applicable for object detection due to the complexity of scene-level images and the requirement for object-level annotations. In contrast, our work studies under the detection setting and requires outlier synthesis at scene level." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Foundation Models The evolution of large language models (LLMs) began with training on web-scale datasets [9,45], leading to increasingly powerful foundation models [4, 6] capable of harnessing vast open-world data. Notable advancements include models [1] that interact with users and perform complex tasks like question answering, significantly broadening access to global knowledge. In image generation, diffusion models [24, 49] offer robust capabilities in synthesizing realistic content for applications such as image synthesis and inpainting. Additionally, segmentation foundation models like SAM [30] represent a leap forward in precise image segmentation, benefiting from extensive data training. Foundation models provide diverse data, which provides unlimited po" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 101 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 116, + 478, + 223 + ], + "blocks": [ + { + "bbox": [ + 134, + 116, + 478, + 223 + ], + "lines": [ + { + "bbox": [ + 134, + 116, + 478, + 223 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 478, + 223 + ], + "type": "image", + "image_path": "599d0ae175dbc499dcd032372e9eec73fbf99c0109a180a9b04be407c86659eb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 232, + 482, + 277 + ], + "lines": [ + { + "bbox": [ + 130, + 232, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 482, + 277 + ], + "type": "text", + "content": "Fig. 2: Detailed illustration of our outlier synthesis pipeline. It comprises (a) Instructing an LLM to imagine semantic-novel concepts given ID objects, (b) Editing the selected regions to the expected concepts via prompt-conditioned image inpainting using Stable Diffusion, and (c) Refining the bounding boxes of edited objects using SAM." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 300, + 482, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 324 + ], + "type": "text", + "content": "tential for learning open-world knowledge from these models [55], where the effectiveness of data [21, 53, 63] is also crucial to the downstreaming tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 342, + 202, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 202, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 202, + 355 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 366, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 445 + ], + "type": "text", + "content": "Preliminary For OOD object detection, the training set contains only ID scene-level images " + }, + { + "bbox": [ + 130, + 366, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{id}}" + }, + { + "bbox": [ + 130, + 366, + 482, + 445 + ], + "type": "text", + "content": " with ID objects, annotation bounding boxes " + }, + { + "bbox": [ + 130, + 366, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^{\\mathrm{id}}" + }, + { + "bbox": [ + 130, + 366, + 482, + 445 + ], + "type": "text", + "content": ", and semantic labels " + }, + { + "bbox": [ + 130, + 366, + 482, + 445 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 366, + 482, + 445 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 130, + 366, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{id}} = \\left\\{(\\mathbf{x}^{\\mathrm{id}}, \\mathbf{b}^{\\mathrm{id}}, \\mathbf{y}^{\\mathrm{id}})\\right\\}" + }, + { + "bbox": [ + 130, + 366, + 482, + 445 + ], + "type": "text", + "content": ". The labels of ID objects always belong to a close set with " + }, + { + "bbox": [ + 130, + 366, + 482, + 445 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 366, + 482, + 445 + ], + "type": "text", + "content": " categories, denoted as " + }, + { + "bbox": [ + 130, + 366, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^{\\mathrm{id}} \\in \\mathcal{Y}_{\\mathrm{id}}" + }, + { + "bbox": [ + 130, + 366, + 482, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 366, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_{\\mathrm{id}} = \\{y_1^{\\mathrm{id}}, y_2^{\\mathrm{id}}, \\dots, y_K^{\\mathrm{id}}\\}" + }, + { + "bbox": [ + 130, + 366, + 482, + 445 + ], + "type": "text", + "content": ". During inference, for each proposed object from an input scene-level image, it is required to identify whether its category belongs to " + }, + { + "bbox": [ + 130, + 366, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_{\\mathrm{id}}" + }, + { + "bbox": [ + 130, + 366, + 482, + 445 + ], + "type": "text", + "content": " or not." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 453, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 594 + ], + "type": "text", + "content": "Overview As illustrated in Fig. 1, our outlier synthesis pipeline consists of (1) synthesizing a set of effective photo-realistic scene-level OOD images " + }, + { + "bbox": [ + 130, + 453, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{edit}}" + }, + { + "bbox": [ + 130, + 453, + 482, + 594 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 130, + 453, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{edit}} = \\left\\{(\\mathbf{x}^{\\mathrm{edit}},\\mathbf{b}^{\\mathrm{edit}})\\right\\}" + }, + { + "bbox": [ + 130, + 453, + 482, + 594 + ], + "type": "text", + "content": ", which contains novel objects and corresponding annotation boxes " + }, + { + "bbox": [ + 130, + 453, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^{\\mathrm{edit}}" + }, + { + "bbox": [ + 130, + 453, + 482, + 594 + ], + "type": "text", + "content": " based on region-level editing from " + }, + { + "bbox": [ + 130, + 453, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{id}}" + }, + { + "bbox": [ + 130, + 453, + 482, + 594 + ], + "type": "text", + "content": " in a fully automated, labor-free way; and (2) select and use the efficient synthetic data to provide pseudo-OOD supervisions for training OOD object detector together with the ID samples in the training set. Further design of the pipeline requires answering the following questions: (1) how to distill the open-set knowledge embedded in foundation models to scene-level OOD data and (2) how to utilize the synthesized data to regularize the decision boundary and facilitate OOD object detection. We discuss them accordingly in Sec. 3.1 and Sec. 3.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 609, + 430, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 609, + 430, + 623 + ], + "spans": [ + { + "bbox": [ + 130, + 609, + 430, + 623 + ], + "type": "text", + "content": "3.1 Synthesizing Semantic-novel Objects in Scene Images" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Imagining Novel Concepts from ID objects As shown in Fig. 2(a), based on the ID labels " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathrm{id}}" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " in training set " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{id}}" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": ", we consider that novel concepts that are different from ID categories can be potential candidates for generating OOD" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 199, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 199, + 91, + 448, + 102 + ], + "type": "text", + "content": "Can OOD Object Detectors Learn from Foundation Models?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "objects. The next is to discover novel concepts that offer hard OOD samples sharing high visual similarity with ID samples and being contextually compatible with the scene context for object detection. Rather than relying on human labor to investigate all potential candidates, we leverage the vast knowledge and reasoning capabilities of LLM, GPT-4 [1] to check the visual similarity and contextual compatibility. This allows us to associate ID objects and promote the conceptualization of possible novel objects to replace existing ID objects through the use of a prompt with in-context examples [4] as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 225, + 468, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 225, + 468, + 283 + ], + "spans": [ + { + "bbox": [ + 147, + 225, + 468, + 283 + ], + "type": "text", + "content": "Here is a list containing several objects " + }, + { + "bbox": [ + 147, + 225, + 468, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathrm{id}}" + }, + { + "bbox": [ + 147, + 225, + 468, + 283 + ], + "type": "text", + "content": ". Now, if I provide you an object name, you should return to me objects that are similar to the usage scenario and volume of the provided object but are not in the previous object list. For example, if I give you the word: person, you should respond and only respond: 'mannequin', 'sculpture', 'scarecrows', 'doll', 'puppet'." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 298, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 482, + 370 + ], + "type": "text", + "content": "With its robust logical foundation and rich knowledge, the LLM envisions a collection of novel objects for each ID object label, denoted as " + }, + { + "bbox": [ + 130, + 298, + 482, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_{\\mathrm{novel}}" + }, + { + "bbox": [ + 130, + 298, + 482, + 370 + ], + "type": "text", + "content": ", while maintaining the semantic separability between imagined objects and ID objects. We empirically find one in-context example that is sufficient for us to discover novel concepts. For each ID label " + }, + { + "bbox": [ + 130, + 298, + 482, + 370 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 298, + 482, + 370 + ], + "type": "text", + "content": ", we discover " + }, + { + "bbox": [ + 130, + 298, + 482, + 370 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 298, + 482, + 370 + ], + "type": "text", + "content": " novel concepts using LLM " + }, + { + "bbox": [ + 130, + 298, + 482, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i^{\\mathrm{novel}}" + }, + { + "bbox": [ + 130, + 298, + 482, + 370 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 298, + 482, + 370 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 298, + 482, + 370 + ], + "type": "text", + "content": " concepts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 381, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 482, + 538 + ], + "type": "text", + "content": "Editing Objects on Selected Regions With the discovered novel concepts " + }, + { + "bbox": [ + 130, + 381, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathrm{novel}} = \\{\\mathbf{y}_1^{\\mathrm{novel}},\\mathbf{y}_2^{\\mathrm{novel}},\\dots,\\mathbf{y}_K^{\\mathrm{novel}}\\}" + }, + { + "bbox": [ + 130, + 381, + 482, + 538 + ], + "type": "text", + "content": ", the next step is to use them as prompts for the text-to-image generation model to generate an image. To generate a new image with novel concepts " + }, + { + "bbox": [ + 130, + 381, + 482, + 538 + ], + "type": "inline_equation", + "content": "y_{j}\\in \\mathbf{y}_{i}^{\\mathrm{novel}}" + }, + { + "bbox": [ + 130, + 381, + 482, + 538 + ], + "type": "text", + "content": ", we choose to replace existing ID objects in existing images with label " + }, + { + "bbox": [ + 130, + 381, + 482, + 538 + ], + "type": "inline_equation", + "content": "y_{i}^{\\mathrm{id}}" + }, + { + "bbox": [ + 130, + 381, + 482, + 538 + ], + "type": "text", + "content": " instead of finding new locations or generating one image from scratch. By doing so, we can ensure context compatibility and eliminate distractions from the scene context as it is preserved. As illustrated in Fig. 2 (b), we use Stable-Diffusion-Inpainting [49], denoted as " + }, + { + "bbox": [ + 130, + 381, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\mathrm{SDI}(\\cdot)" + }, + { + "bbox": [ + 130, + 381, + 482, + 538 + ], + "type": "text", + "content": ", to perform region-level editing on ID images. The ID object is denoted as " + }, + { + "bbox": [ + 130, + 381, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{id}}" + }, + { + "bbox": [ + 130, + 381, + 482, + 538 + ], + "type": "text", + "content": ", with its corresponding annotation box " + }, + { + "bbox": [ + 130, + 381, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^{\\mathrm{id}}" + }, + { + "bbox": [ + 130, + 381, + 482, + 538 + ], + "type": "text", + "content": " serving as the editing mask, and the associated imagined novel concept " + }, + { + "bbox": [ + 130, + 381, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^{\\mathrm{novel}}" + }, + { + "bbox": [ + 130, + 381, + 482, + 538 + ], + "type": "text", + "content": " are provided as inputs to the SDI, which is one of the most successful models for conditional image generation and editing. Thus, an edited image " + }, + { + "bbox": [ + 130, + 381, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{edit}}" + }, + { + "bbox": [ + 130, + 381, + 482, + 538 + ], + "type": "text", + "content": " containing a novel object is obtained as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 246, + 549, + 480, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 549, + 480, + 563 + ], + "spans": [ + { + "bbox": [ + 246, + 549, + 480, + 563 + ], + "type": "interline_equation", + "content": "\\mathbf {x} ^ {\\text {e d i t}} = \\operatorname {S D I} \\left(\\mathbf {x} ^ {\\text {i d}}, \\mathbf {b} ^ {\\text {i d}}, \\mathbf {y} ^ {\\text {n o v e l}}\\right). \\tag {1}", + "image_path": "66c765369542dccb8b4815350b029e6728112c346022486aced9514e03b6a40b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Refining Annotation Boxes of Novel Objects Due to the randomness in diffusion models, the attributes of edited objects, such as their quality, volume, and localization, may not match the original object box. To address this issue, as depicted in Fig. 2(c), we design an efficient and effective refiner based on SAM [30] to obtain refined accurate bounding boxes on novel objects. First, for an edited image " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{edit}}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " with the editing mask " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^{\\mathrm{id}}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ", we use a padding area extended from " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^{\\mathrm{id}}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " as the prompt and employ SAM to output the instance mask with" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 101 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 115, + 386, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 115, + 386, + 129 + ], + "spans": [ + { + "bbox": [ + 131, + 115, + 386, + 129 + ], + "type": "text", + "content": "highest confidence " + }, + { + "bbox": [ + 131, + 115, + 386, + 129 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^{\\mathrm{SAM}}" + }, + { + "bbox": [ + 131, + 115, + 386, + 129 + ], + "type": "text", + "content": " for the novel object in the area:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 225, + 132, + 481, + 147 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 132, + 481, + 147 + ], + "spans": [ + { + "bbox": [ + 225, + 132, + 481, + 147 + ], + "type": "interline_equation", + "content": "\\mathbf {m} ^ {\\text {S A M}} = \\operatorname {S A M} \\left(\\mathbf {x} ^ {\\text {e d i t}}; \\text {p a d d i n g} \\left(\\mathbf {b} ^ {\\text {i d}}, e\\right)\\right), \\tag {2}", + "image_path": "2f80dfa53a5b003e91721acccd08b5c1fdbfca728691a3cc15ef38c51b2511de.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 150, + 481, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 187 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 150, + 481, + 187 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 130, + 150, + 481, + 187 + ], + "type": "text", + "content": " represents the range of padding. Then, we convert obtained masks " + }, + { + "bbox": [ + 130, + 150, + 481, + 187 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^{\\mathrm{SAM}}" + }, + { + "bbox": [ + 130, + 150, + 481, + 187 + ], + "type": "text", + "content": " to boxes " + }, + { + "bbox": [ + 130, + 150, + 481, + 187 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^{\\mathrm{SAM}}" + }, + { + "bbox": [ + 130, + 150, + 481, + 187 + ], + "type": "text", + "content": ", and calculate IoU between " + }, + { + "bbox": [ + 130, + 150, + 481, + 187 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^{\\mathrm{SAM}}" + }, + { + "bbox": [ + 130, + 150, + 481, + 187 + ], + "type": "text", + "content": " and the corresponding " + }, + { + "bbox": [ + 130, + 150, + 481, + 187 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^{\\mathrm{id}}" + }, + { + "bbox": [ + 130, + 150, + 481, + 187 + ], + "type": "text", + "content": " to filter out novel objects that vary highly in scale:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 217, + 191, + 481, + 212 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 191, + 481, + 212 + ], + "spans": [ + { + "bbox": [ + 217, + 191, + 481, + 212 + ], + "type": "interline_equation", + "content": "\\left\\{\\mathbf {b} ^ {\\text {e d i t}} \\right\\} = \\left\\{\\mathbf {b} ^ {\\text {S A M}} \\mid \\operatorname {I o U} \\left(\\mathbf {b} ^ {\\text {S A M}}, \\mathbf {b} ^ {\\text {i d}}\\right) > \\gamma \\right\\}, \\tag {3}", + "image_path": "f913fafc2eeb7d5684a35b3e4288533219a993fe5d8ce4d760cbc5eedebdbfc7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 214, + 482, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 214, + 482, + 264 + ], + "spans": [ + { + "bbox": [ + 131, + 214, + 482, + 264 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 214, + 482, + 264 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 214, + 482, + 264 + ], + "type": "text", + "content": " denotes a threshold value on IoU. It ensures a high enough recall rate to rule out the instability of Stable Diffusion and SAM and uncontrollable localization of the edited objects. Thus we obtain the synthetic outlier data " + }, + { + "bbox": [ + 131, + 214, + 482, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{edit}}" + }, + { + "bbox": [ + 131, + 214, + 482, + 264 + ], + "type": "text", + "content": " as illustrated in Fig. 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 277, + 405, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 277, + 405, + 289 + ], + "spans": [ + { + "bbox": [ + 131, + 277, + 405, + 289 + ], + "type": "text", + "content": "3.2 Mining Hard OOD Samples and Model Training" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 293, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 418 + ], + "type": "text", + "content": "Mining Hard OOD Objects with High Visual Similarities for Training We consider the novel objects that are most likely to be confused with the corresponding ID objects by the object detector as the most effective ones. We thus aim to find synthetic OOD samples that are most easily confused as ID to participate in training based on pairwise similarity in the latent space of the pre-trained object detector. For each novel object with bounding box " + }, + { + "bbox": [ + 130, + 293, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^{\\mathrm{edit}}" + }, + { + "bbox": [ + 130, + 293, + 482, + 418 + ], + "type": "text", + "content": " in the synthetic data " + }, + { + "bbox": [ + 130, + 293, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{edit}}" + }, + { + "bbox": [ + 130, + 293, + 482, + 418 + ], + "type": "text", + "content": ", we construct it with the corresponding original ID object with its bounding box as a pair: " + }, + { + "bbox": [ + 130, + 293, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\left\\{(\\mathbf{b}^{\\mathrm{edit}}, \\mathbf{x}^{\\mathrm{edit}}), (\\mathbf{b}^{\\mathrm{id}}, \\mathbf{x}^{\\mathrm{id}})\\right\\}" + }, + { + "bbox": [ + 130, + 293, + 482, + 418 + ], + "type": "text", + "content": ". For an off-the-shelf object detector, denoted by " + }, + { + "bbox": [ + 130, + 293, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{det}}" + }, + { + "bbox": [ + 130, + 293, + 482, + 418 + ], + "type": "text", + "content": ", we extract latent features, " + }, + { + "bbox": [ + 130, + 293, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{edit}}" + }, + { + "bbox": [ + 130, + 293, + 482, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 293, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{id}}" + }, + { + "bbox": [ + 130, + 293, + 482, + 418 + ], + "type": "text", + "content": ", for each pair:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 214, + 418, + 481, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 418, + 481, + 432 + ], + "spans": [ + { + "bbox": [ + 214, + 418, + 481, + 432 + ], + "type": "interline_equation", + "content": "\\mathbf {z} ^ {\\text {e d i t}}, \\mathbf {z} ^ {\\text {i d}} = \\mathcal {F} _ {\\det } \\left(\\mathbf {b} ^ {\\text {e d i t}}; \\mathbf {x} ^ {\\text {e d i t}}\\right), \\mathcal {F} _ {\\det } \\left(\\mathbf {b} ^ {\\text {i d}}; \\mathbf {x} ^ {\\text {i d}}\\right). \\tag {4}", + "image_path": "0483e6d3abca831497505d97badd6fb1752328a3001408d216fd84361ce284a8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 434, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 483 + ], + "type": "text", + "content": "The most effective novel objects are those with visual patterns that can be easily mistaken for their corresponding ID objects by an object detector. Therefore, we filter these novel objects based on their similarity to provide pseudo-OD supervision:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 207, + 487, + 481, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 487, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 207, + 487, + 481, + 502 + ], + "type": "interline_equation", + "content": "\\left\\{\\mathbf {z} ^ {\\mathrm {o o d}} \\right\\} = \\left\\{\\mathbf {z} ^ {\\mathrm {e d i t}} \\big | \\epsilon_ {l o w} < \\mathrm {s i m} (\\mathbf {z} ^ {\\mathrm {e d i t}}, \\mathbf {z} ^ {\\mathrm {i d}}) < \\epsilon_ {u p} \\right\\}, \\qquad \\qquad (5)", + "image_path": "a31287ef844ba5fbb9d3a9422bfb44502a47f5fee70d97148f921f364ad16902.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 505, + 482, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 505, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 505, + 482, + 541 + ], + "type": "text", + "content": "where the similarities are computed between latent object features of edit-ID pairs. Here " + }, + { + "bbox": [ + 130, + 505, + 482, + 541 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(\\cdot)" + }, + { + "bbox": [ + 130, + 505, + 482, + 541 + ], + "type": "text", + "content": " denotes cosine similarity calculating and " + }, + { + "bbox": [ + 130, + 505, + 482, + 541 + ], + "type": "inline_equation", + "content": "\\epsilon_{low},\\epsilon_{up}" + }, + { + "bbox": [ + 130, + 505, + 482, + 541 + ], + "type": "text", + "content": " stand for the lower/upper similarity thresholds." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 546, + 482, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 546, + 482, + 584 + ], + "spans": [ + { + "bbox": [ + 131, + 546, + 482, + 584 + ], + "type": "text", + "content": "Optimizing ID/OOD Decision Boundary with Synthetic Samples Once we have obtained the ID and synthetic OOD objects, we employ a lightweight MLP, denoted as " + }, + { + "bbox": [ + 131, + 546, + 482, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{ood}}" + }, + { + "bbox": [ + 131, + 546, + 482, + 584 + ], + "type": "text", + "content": ", as the OOD detector optimized with a bi-classify loss:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 140, + 587, + 481, + 614 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 587, + 481, + 614 + ], + "spans": [ + { + "bbox": [ + 140, + 587, + 481, + 614 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {o o d}} = \\mathbb {E} _ {\\mathbf {z} \\sim \\mathbf {z} ^ {\\mathrm {i d}}} \\left[ - \\log \\frac {1}{1 + \\exp^ {- \\mathcal {F} _ {\\mathrm {o o d}} (\\mathbf {z})}} \\right] + \\mathbb {E} _ {\\mathbf {z} \\sim \\mathbf {z} ^ {\\mathrm {o o d}}} \\left[ - \\log \\frac {\\exp^ {- \\mathcal {F} _ {\\mathrm {o o d}} (\\mathbf {z})}}{1 + \\exp^ {- \\mathcal {F} _ {\\mathrm {o o d}} (\\mathbf {z})}} \\right]. (6)", + "image_path": "7cd4c339cc99ebd09d4fd72f2c6f344fc5b0976943c7c11a0d7261e7139f0565.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 482, + 666 + ], + "type": "text", + "content": "The aforementioned design ensures both semantic separability and pattern similarity for the chosen synthetic samples. As a result, our proposed method elegantly optimizes the decision boundary using only a limited number of samples. It is further demonstrated and validated in the following experiments." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 199, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 199, + 91, + 448, + 102 + ], + "type": "text", + "content": "Can OOD Object Detectors Learn from Foundation Models?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 236 + ], + "type": "text", + "content": "Datasets Following OOD object detection setting [15], we use the PASCAL-VOC [16] and Berkeley DeepDrive (BDD-100K) [65] as the ID training datasets, which consist of 20 and 10 ID categories, respectively. Meanwhile, we evaluate the performance of our approach on two OOD datasets: MS-COCO [36] and OpenImages [32] respectively. Categories from the OOD datasets that overlapped with the ID datasets are removed to guarantee the absence of ID categories. We report ID categories, OOD categories, and texts used in driving image synthesis in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 241, + 482, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 241, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 482, + 338 + ], + "type": "text", + "content": "Metrics We primarily focus on reporting the FPR95 which represents the false positive rate of OOD samples when the true positive rate of ID samples is at " + }, + { + "bbox": [ + 130, + 241, + 482, + 338 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 241, + 482, + 338 + ], + "type": "text", + "content": " and lower is better. FPR95 is widely used to assess the OOD object detection performance [12, 15, 59-61]. Additionally, we present the Area Under the Receiver Operating Characteristic Curve (AUROC, higher is better) that is widely utilized to evaluate binary classification problems. Since we only train a plugin MLP on top of existing object detectors for OOD detection, ID performance, e.g., mean Average Precision (mAP) is unchanged and thus omitted." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 344, + 482, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 482, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 482, + 500 + ], + "type": "text", + "content": "Implementation Details For our synthetic data, in order to maintain the experimental setting of OOD object detection and avoid leaking prior knowledge of foundation models, we remove all imagined novel objects that have the same or similar meaning as the ground truth OOD data categories. For model training, we follow the architectures of the baseline method [15, 59] to use a Faster R-CNN [48] as the base object detector with ResNet-50 [20] firstly. Then we trained a simple and lightweight 3-layer MLP. We follow [59] to extract multi-level features as training samples. ID samples are extracted from ID images and OOD samples are extracted from selected synthesized images with OOD bounding boxes. For training on the PASCAL-VOC dataset, we employ a learning rate of 1e-4, while for the BDD-100K dataset, we use a learning rate of 5e-5. Both training processes utilize a momentum of 0.9, a dropout rate of 0.5, and a batch size of 32. All training is conducted on GeForce RTX 3090 GPUs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 515, + 364, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 364, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 364, + 528 + ], + "type": "text", + "content": "4.1 Main Results on OOD Object Detection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "content": "We evaluate the performance of the proposed method on different challenging benchmarks and obtain notable results (see Tab. 1). As the first work to introduce synthetic scene-level natural images as OOD samples, we incorporate our data-centric method to two off-the-shelf object detectors [15,48], achieving new state-of-the-art performance in OOD object detection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Compared with previous methods, we present comprehensive and substantial performance improvements on FPR95. The encouraging outcomes clearly show that our synthetic data offers superior OOD supervision and are well-suited for forming a precise decision boundary between ID and OOD samples as illustrated in Fig. 1, which significantly reduces the interference caused by contextual information when optimizing the decision boundary." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 187, + 481, + 372 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 1: Comparing on varied ID (PASCAL VOC [16], BDD-100K [65]) and OOD (MS-COCO [36], OpenImages [32]) datasets, our method significantly outperforms other methods on different metrics and achieves SOTA performance on OOD object detection. Our method is validated on two different existing object detectors, Faster R-CNN [48] and VOS [15] (denoted as Faster R-CNN + Ours and VOS + Ours respectively). (Top results are shown in bold.)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 187, + 481, + 372 + ], + "lines": [ + { + "bbox": [ + 133, + 187, + 481, + 372 + ], + "spans": [ + { + "bbox": [ + 133, + 187, + 481, + 372 + ], + "type": "table", + "html": "
MethodID:PASCAL-VOCID:BDD-100K
MS-COCOOpenImagesMS-COCOOpenImages
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
MSP [22]70.9983.4573.1381.9180.9475.8779.0477.38
ODIN [35]59.8282.2063.1482.5962.8574.4458.9276.61
Mahalanobis [34]96.4659.2596.2757.4257.6684.9260.1686.88
Energy score [38]56.8983.6958.6982.9860.0677.4854.7979.60
Gram matrices [50]62.7579.8867.4277.6260.9374.9377.5559.38
Generalized ODIN [25]58.5783.1270.2879.2357.2785.2250.1787.18
CSI [52]59.9181.8357.4182.9547.1084.0937.0687.99
GAN-synthesis [33]60.9383.6759.9782.6757.0378.8250.6181.25
SIREN-KNN [12]47.4589.6750.3888.80----
VOS [15]47.5388.7051.3385.2344.2786.8735.5488.52
SR-VAE [61]42.1790.2846.2687.8932.2390.6921.8193.55
DFDD [60]41.3490.7944.5288.6530.7190.7422.6792.48
SAFE [59]47.4080.3020.0692.2832.5688.9616.0494.64
Faster R-CNN + Ours36.4486.5213.3495.3722.6795.4412.9696.26
VOS + Ours34.9787.9011.2596.9623.0994.3214.1296.41
", + "image_path": "a3bec877584ded66721edb2a10216232d27424e350049a8aba4ce76d6dbca775.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 134, + 432, + 481, + 472 + ], + "blocks": [ + { + "bbox": [ + 130, + 380, + 482, + 426 + ], + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 426 + ], + "type": "text", + "content": "Table 2: Ablation on the number of our synthetic data in training. Taking PASCAL-VOC as the ID dataset, we perform seven groups of random sampling with different numbers in the synthetic dataset to extract features as OOD samples, evaluate and report the performance on the MS-COCO/OpenImages datasets." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 432, + 481, + 472 + ], + "lines": [ + { + "bbox": [ + 134, + 432, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 134, + 432, + 481, + 472 + ], + "type": "table", + "html": "
#Sample14k12k10k8k6k4k2k
FPR95↓36.70/12.9636.27/13.0137.31/13.2536.53/13.3036.70/12.9636.44/13.3437.82/13.87
AUROC↑86.65/95.5486.68/95.5586.64/95.5186.69/95.5286.75/95.5686.52/95.4486.03/95.18
", + "image_path": "e07a92a120f1f8f957865bb4e8b2b27e093375ed5265df1befa3647ea7965ef1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 482, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 572 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 572 + ], + "type": "text", + "content": "Bridged by our synthetic data, foundation models' extensive knowledge and powerful logic about novel concepts are effectively injected into our model through novel concept imagining and region-level editing. Furthermore, powered by the similarity-based filter, our synthetic data proves to be highly effective. Compared with SAFE [59] which uses a similar framework, we only use around " + }, + { + "bbox": [ + 130, + 486, + 482, + 572 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 572 + ], + "type": "text", + "content": " (on PASCAL-VOC) and " + }, + { + "bbox": [ + 130, + 486, + 482, + 572 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 572 + ], + "type": "text", + "content": " (on BDD-100K) of auxiliary data to achieve a significant performance improvement. Further analysis is presented in Sec. 4.2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 586, + 239, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 239, + 600 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 239, + 600 + ], + "type": "text", + "content": "4.2 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Number of Training Samples We conduct an extensive ablation study on the quantity of synthetic data utilized, as illustrated in Tab. 2. We employ seven sets of synthetic data with varying quantities as OOD samples, using PASCAL-VOC as the ID dataset. Features are extracted and the OOD detector is trained based on the same Faster R-CNN checkpoint for each set. It is noteworthy that" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "type": "text", + "content": "Can OOD Object Detectors Learn from Foundation Models?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 155, + 481, + 200 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 3: We study the impact of associating varying numbers of imagined novel objects with each ID object. Taking PASCAL-VOC as the ID dataset, we report the performance on MS-COCO/OpenImages datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 155, + 481, + 200 + ], + "lines": [ + { + "bbox": [ + 138, + 155, + 481, + 200 + ], + "spans": [ + { + "bbox": [ + 138, + 155, + 481, + 200 + ], + "type": "table", + "html": "
#Sample345678
FPR95↓36.96/13.5837.13/13.1537.31/12.8236.87/13.2537.91/13.5337.13/13.15
AUROC↑86.56/95.5486.63/95.5186.46/95.4786.51/95.3786.43/95.3586.44/95.56
", + "image_path": "c5a3bb8172076a8da0dcc94c6ff5a3fee82c6017c7c3ff1a6bb83ffb02422731.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 262, + 481, + 332 + ], + "blocks": [ + { + "bbox": [ + 130, + 210, + 482, + 254 + ], + "lines": [ + { + "bbox": [ + 130, + 210, + 482, + 254 + ], + "spans": [ + { + "bbox": [ + 130, + 210, + 482, + 254 + ], + "type": "text", + "content": "Table 4: We randomly sample the same numbers of OOD features as the main experiment instead of using the feature filter (denoted as w/o filter), and evaluate on multiple datasets. The obtained results demonstrate the effectiveness of the proposed data filter." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 262, + 481, + 332 + ], + "lines": [ + { + "bbox": [ + 135, + 262, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 135, + 262, + 481, + 332 + ], + "type": "table", + "html": "
MethodID:PASCAL-VOCID:BDD-100K
MS-COCOOpenImagesMS-COCOOpenImages
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
w/o filter39.2985.4613.6895.2225.4593.1715.3295.83
Ours36.4486.5213.3495.3722.6795.4412.9696.26
", + "image_path": "8b32d372cb797d2b57b7000c9f7d676823ab2ac2b933b6b52cc21f037e6acb34.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 347, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 418 + ], + "type": "text", + "content": "as the number of samples decreases from 14k to 2k, the performance does not deteriorate but rather maintains stable and superior results. This highlights our method's data efficiency (SAFE used 16k samples). Combined with our feature similarity-based filtering strategy as in Sec. 3.2, a small number of high-quality OOD samples with visual similarity directly promotes the optimization of precise decision boundaries to achieve stable improvements." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 419, + 482, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 482, + 504 + ], + "type": "text", + "content": "Meanwhile, we use the same detector checkpoint to assess the parallel baseline, SAFE [59], and get the performance of 50.86/23.60 on FPR95 and 78.15/91.42 on AUROC. SAFE augments about 16k images and extracts more than 100k instance-level features as OOD samples. In contrast, our approach utilizes fewer synthetic images and extracts only one instance-level feature from the edited novel object in each synthetic image as an OOD sample, resulting in significantly superior performance compared to SAFE [59]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 514, + 482, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 610 + ], + "type": "text", + "content": "Number of Concepts to Imagine We employ in-context learning to guide LLM in associating new objects for driving image editing. For each ID object, the LLM connects a steady stream of novel objects. We further explore the impact of the number of corresponding novel objects for each ID object on data performance. We randomly sample different numbers of novel objects from LLM's responses for each ID object. As shown in Tab. 3, the performance remains stable despite changes in the number of concepts, further highlighting the stability and robustness of our synthetic data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "SAM-based Refiner As mentioned in Sec. 3.1, we propose to utilize SAM-based refiner to correct the bounding boxes of novel objects to obtain higher-quality instance-level OOD features. Therefore, we comparatively remove the proposed refiner and directly used the corresponding editing masks as bound-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 481, + 355 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 355 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 355 + ], + "type": "image", + "image_path": "fe7cb7fe213e39e1e75e33c1560abc363efc2324b6b698df79d021f04c9be644.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 362, + 482, + 429 + ], + "lines": [ + { + "bbox": [ + 130, + 362, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 362, + 482, + 429 + ], + "type": "text", + "content": "Fig. 3: We show cases on six intervals of feature similarity (consistent with Eq. (5), indicated at the bottom of the figure). The first line contains the corresponding initial images, the second line contains the synthetic images with the corresponding boxes of the novel objects (yellow boxes), and the third line contains the difference heat maps of the latent feature maps extracted from the above image pairs (superimposed on the corresponding synthetic images, denoted as Diff-map)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 445, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 554 + ], + "type": "text", + "content": "ing boxes to extract OOD features for training. Taking PASCAL-VOC as the ID dataset, after removing the refiner, we obtain 39.55/13.72 of FPR95 and 85.94/95.37 of AUROC on MS-COCO/OpenImages datasets, which is better than previous methods but worse than the results (Faster R-CNN + Ours in Tab. 1) when using the refiner. This proves that OOD supervision signals contained in the synthetic data are already extracted and achieve good results under the localization of the fuzzy boxes, but more precise boxes mean higher quality features. More demos and analyses of the SAM-based refiner are shown in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Similarity-based Filter The filter is designed to incorporate the most useful data into training, and avoid unnecessary noise. The design is reflected in two aspects: on one hand, the outlier object should process similar visual patterns to the original object, thus being confusing and can facilitate learning; on the other hand, over-high similarity may indicate failures of the editing process (e.g., when the concept is not an object). These considerations are applied as thresholding on pairwise cosine similarity between object features, as in Eq. (5). As shown in Tab. 4, this filter brings a notable improvement across benchmarks." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "type": "text", + "content": "Can OOD Object Detectors Learn from Foundation Models?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 185, + 482, + 253 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 5: Comparing varied images as OOD samples for training, we first show some synthetic object-centric images generated by Stable Diffusion (left side). Then with PASCAL-VOC as ID dataset, we report the results obtained by using synthetic object-centric images (denoted as object-centric images) and scene-level images with novel objects but without bounding boxes (denoted as scene-level w/o boxes) as OOD samples to participate in training." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 185, + 482, + 253 + ], + "lines": [ + { + "bbox": [ + 135, + 185, + 482, + 253 + ], + "spans": [ + { + "bbox": [ + 135, + 185, + 482, + 253 + ], + "type": "table", + "html": "
Object-centric Images DataMS-COCOOpenImages
FPR95↓ AUROC↑FPR95↓ AUROC↑
object-centric images51.9981.4820.7093.85
scene-level w/o boxes48.0182.3818.6193.44
Ours36.4486.5213.3495.37
", + "image_path": "bd7e1af740465d1a34743812f4612eca5f921cdce2c9f0ca830b39b6e3a19c27.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 262, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 430 + ], + "type": "text", + "content": "To provide more insights on the choices of filtering thresholds, we display some cases in different intervals of feature similarity in Fig. 3. We show the synthetic images, the corresponding initial images, and the difference between feature maps (denoted as Diff-map), respectively. The Diff-maps prove that the edited area is sensitively attended to by the model. But for images with extremely high similarity " + }, + { + "bbox": [ + 130, + 262, + 482, + 430 + ], + "type": "inline_equation", + "content": "(>0.9)" + }, + { + "bbox": [ + 130, + 262, + 482, + 430 + ], + "type": "text", + "content": ", they always contain some editing failures and blurs. As illustrated on the top of the first column in Fig. 3, it is not intuitively evident what the ship has been edited into (the target object is a raft). Besides, as the similarity upperbound decreases, we progressively obtain more realistic and reasonable images. But note that when the similarity is excessively low, as seen in the last column of Fig. 3, the objects are edited into the corresponding text or an unnatural object, leading to image distortion. This strongly supports the idea that the quality and usability of edited images are closely connected to visual similarity. More cases and analyses are presented in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 445, + 324, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 445, + 324, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 445, + 324, + 456 + ], + "type": "text", + "content": "4.3 Discussions on Outlier Synthesis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 641 + ], + "type": "text", + "content": "Scene-level Editing Matters Through regional-level editing, we replace the ID object with a novel object with a bounding box and ensure consistent context information. However, some simpler methods based on foundation models also achieve the acquisition and use of OOD data. For example, Dream-OOD [13] uses well-trained text-conditional space and diffusion model to synthesize realistic object-centric data for promoting OOD image classification. Similarly, keeping other settings unchanged, we use our novel concepts to drive Stable-Diffusion instead of Stable-Diffusion-Inpainting [49] to synthesize novel images, which are also processed and filtered by our proposed refiner and filter (some synthesized images are shown in Tab. 5), thereby participating in the training as OOD supervision. However, as shown in Tab. 5 (object-centric images), the synthetic novel object-centric images do not aid in training and result in poor performance, even though they possess high visual quality. This clearly validates our decision to edit scene-level images rather than composing new ones, and highlights the significance of maintaining contextual consistency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "Additionally, we examine the possibility of using the edited scene-level image as a whole (ignoring the boxes) as OOD samples in the training process. The" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 479, + 235 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 479, + 235 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 479, + 235 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 479, + 235 + ], + "type": "image", + "image_path": "a748bb0d397ed352e10074919107cb8a6c4546aec7ae6d4bd06cf5eeeef22710.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 246, + 482, + 302 + ], + "lines": [ + { + "bbox": [ + 131, + 246, + 482, + 302 + ], + "spans": [ + { + "bbox": [ + 131, + 246, + 482, + 302 + ], + "type": "text", + "content": "Fig. 4: We edit the context of the synthetic data (in blue box) so that the images contain novel objects and novel context (in orange box). Then we calculate the similarity between the instance-level feature of the corresponding objects from all synthetic images and the instance-level feature in the initial image (left, the bird). The similarities and the corresponding difference maps are shown in the figure." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 312, + 482, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 350 + ], + "type": "text", + "content": "results, as depicted in Tab. 5 (scene-level w/o boxes), are significantly inferior compared to our method's performance. This demonstrates that controllable bounding boxes are indispensable in this task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 357, + 483, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 483, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 483, + 489 + ], + "type": "text", + "content": "Context Consistency Matters Given the importance of scene-level synthesis as discussed in the previous paragraph, we study the factors that make scene-level editing indispensable, and find context consistency to be a crucial one. Besides calculating the similarity between ID/OOD object pairs before/after box-conditioned editing, we also try further editing parts of the background of the already edited images, and also calculate its object similarity with the initial object. As shown in Fig. 4, even small editing on parts of the background (out of the object boxes) can make foreground objects 'look' notably different as perceived by the detector. This highlights the importance of keeping the context unchanged when synthesizing outlier samples, in that if the context is changed, the model easily identifies the object as OOD and cannot break the context bias." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 506, + 217, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 506, + 217, + 518 + ], + "spans": [ + { + "bbox": [ + 132, + 506, + 217, + 518 + ], + "type": "text", + "content": "5 Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 528, + 482, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 482, + 672 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 482, + 672 + ], + "type": "text", + "content": "What Type of OOD Data Matters? We are the first to explore how to edit scene-level images to include novel categories, which contain annotation boxes and ensure context consistency, facilitating OOD object detection. The achieved state-of-the-art performance (Tab. 1) benefits from the optimization of decision boundaries driven by high-quality OOD features. Our exploration demonstrates that annotation boxes and context consistency are particularly important for synthesizing high-quality OOD instances. On the one hand, high-quality annotation boxes provide the possibility to extract high-quality instance-level features from scene-level images, while unrefined boxes (Sec. 4.2) or discarded boxes (Tab. 5) will have a negative impact on performance. On the other hand, context consistency ensures the most effective OOD features are not interfered by different contexts and selected for utilizing (Fig. 4)." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "type": "text", + "content": "Can OOD Object Detectors Learn from Foundation Models?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 481, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 481, + 236 + ], + "type": "text", + "content": "How to Synthesize Suitable OOD Data? We are the first to build an automatic, transparent, and low-cost pipeline (Sec. 3.1) for synthesizing scene-level images containing novel objects with annotation boxes and context consistency. It benefits object detectors' robustness and reliability to unseen data and sets up clear state-of-the-art on multiple OOD object detection benchmarks. Specifically, we organically combine and cleverly use different foundation models [1, 30, 49] (Sec. 3.1) to distill open-world knowledge and inpaint the existing scenes for simulating real OOD scenarios. In addition, our design takes into account the instability of the current foundation models and can release better potential performance in the future development of foundation models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 247, + 481, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 481, + 331 + ], + "type": "text", + "content": "How to Select Suitable OOD Data? We are the first to explicitly decouple OOD data synthesis and selection. On the one hand, we ensure the separability of the synthetic objects in semantic concepts through open-world knowledge provided by LLMs (Sec. 3.1). On the other hand, we ensure the similarity of ID and OOD objects in visual patterns (Sec. 3.1), thereby optimizing the precise decision boundary. This line of thinking has the potential to facilitate more open-world solutions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 343, + 481, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 481, + 464 + ], + "type": "text", + "content": "Broader Impacts Beyond showcasing engineering success via effectively combining specific foundation models, our work uncovers the untapped potential of the text-to-image generative models and visual foundation models in pushing forward the OOD object detection task to effectively leverage the off-the-shelf open-world data knowledge [5,31,66]. More importantly, our work establishes a bridge between OOD object detection and the latest advancements in deep learning, enabling it to benefit from ongoing developments, go beyond isolated academic practice, and resolve practical challenges in open-world applications. Meanwhile, automating novel data generation and curation will inspire more tasks in more modalities, such as in visual-language [44,58] and 3D vision [8,37,64]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 482, + 219, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 219, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 219, + 495 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 507, + 481, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 507, + 481, + 662 + ], + "spans": [ + { + "bbox": [ + 132, + 507, + 481, + 662 + ], + "type": "text", + "content": "In this paper, we investigate improving OOD object detection by distilling open-world data knowledge from text-to-image generative models. We develop an automatic and cost-effective data curation pipeline, SyncOOD, that leverages foundation models as tools to obtain meaningful open-set data from generative models. Through extensive studies, we discover that object boxes and context consistency of the generated data contribute to the improvement of OOD object detection performance. Our comprehensive experiments demonstrate that SyncOOD not only advances the state-of-the-art in OOD object detection but also emphasizes the untapped potential of utilizing large-scale generative models for enhancing the robustness of machine learning systems in open-world settings. As an initial exploration in leveraging foundation models for OOD object detection, we hope our promising results encourage further research in advancing this area in the future." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 482, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 482, + 226 + ], + "type": "text", + "content": "This work has been supported by Hong Kong Research Grant Council - Early Career Scheme (Grant No. 27209621), General Research Fund Scheme (Grant No. 17202422), Theme-based Research (Grant No. T45-701/22-R) and RGC Matching Fund Scheme (RMGS). Part of the described research work is conducted in the JC STEM Lab of Robotics for Soft Materials funded by The Hong Kong Jockey Club Charities Trust. We would like to thank Chirui Chang, Xiaoyang Lyu, Haoru Tan, and Xiuzhe Wu for their insightful discussions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 247, + 197, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 247, + 197, + 260 + ], + "spans": [ + { + "bbox": [ + 133, + 247, + 197, + 260 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 275, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 275, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 275, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 275, + 480, + 308 + ], + "type": "text", + "content": "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: GPT-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 309, + 480, + 342 + ], + "type": "text", + "content": "2. Bendale, A., Boult, T.E.: Towards open set deep networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1563-1572 (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 343, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 364 + ], + "type": "text", + "content": "3. Bishop, C.M.: Novelty detection and neural network validation. IEE Proceedings-Vision, Image and Signal processing 141(4), 217-222 (1994)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 365, + 480, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 397 + ], + "type": "text", + "content": "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in Neural Information Processing Systems 33, 1877-1901 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 399, + 480, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 480, + 419 + ], + "type": "text", + "content": "5. Chang, C., Liu, Z., Lyu, X., Qi, X.: What matters in detecting ai-generated videos like sora? arXiv preprint arXiv:2406.19568 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 421, + 480, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 421, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 138, + 421, + 480, + 463 + ], + "type": "text", + "content": "6. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. Journal of Machine Learning Research 24(240), 1-113 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 465, + 480, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 465, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 138, + 465, + 480, + 486 + ], + "type": "text", + "content": "7. Deepshikha, K., Yelleni, S.H., Srijith, P., Mohan, C.K.: Monte Carlo dropblock for modelling uncertainty in object detection. arXiv preprint arXiv:2108.03614 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 487, + 480, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 480, + 519 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 480, + 519 + ], + "type": "text", + "content": "8. Deng, W., Ding, R., Yang, J., Liu, J., Li, Y., Qi, X., Ngai, E.: Can 3d vision-language models truly understand natural language? arXiv preprint arXiv:2403.14760 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 521, + 480, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 480, + 552 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 480, + 552 + ], + "type": "text", + "content": "9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 555, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 555, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 134, + 555, + 480, + 586 + ], + "type": "text", + "content": "10. Dhamija, A., Gunther, M., Ventura, J., Boult, T.: The overlooked elephant of object detection: Open set. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1021-1030 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 588, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 588, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 134, + 588, + 480, + 609 + ], + "type": "text", + "content": "11. Du, X., Fang, Z., Diakonikolas, I., Li, Y.: How does unlabeled data provably help out-of-distribution detection? arXiv preprint arXiv:2402.03502 (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 610, + 480, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 610, + 480, + 642 + ], + "spans": [ + { + "bbox": [ + 134, + 610, + 480, + 642 + ], + "type": "text", + "content": "12. Du, X., Gozum, G., Ming, Y., Li, Y.: SIREN: Shaping representations for detecting out-of-distribution objects. Advances in Neural Information Processing Systems 35, 20434-20449 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "text", + "content": "13. Du, X., Sun, Y., Zhu, J., Li, Y.: Dream the impossible: Outlier imagination with diffusion models. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 199, + 91, + 447, + 102 + ], + "type": "text", + "content": "Can OOD Object Detectors Learn from Foundation Models?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "14. Du, X., Wang, X., Gozum, G., Li, Y.: Unknown-aware object detection: Learning what you don't know from videos in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13678-13688 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 171 + ], + "type": "text", + "content": "15. Du, X., Wang, Z., Cai, M., Li, Y.: VOS: Learning what you don't know by virtual outlier synthesis. arXiv preprint arXiv:2202.01197 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 204 + ], + "type": "text", + "content": "16. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes (VOC) challenge. International journal of computer vision 88, 303-338 (2010)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "text", + "content": "17. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 237, + 480, + 270 + ], + "type": "text", + "content": "18. Gupta, A., Narayan, S., Joseph, K., Khan, S., Khan, F.S., Shah, M.: Ow-detr: Open-world detection transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9235-9244 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 270, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 270, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 133, + 270, + 480, + 314 + ], + "type": "text", + "content": "19. Hall, D., Dayoub, F., Skinner, J., Zhang, H., Miller, D., Corke, P., Carneiro, G., Angelova, A., Sünderhauf, N.: Probabilistic object detection: Definition and evaluation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1031-1040 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "21. He, R., Sun, S., Yu, X., Xue, C., Zhang, W., Torr, P., Bai, S., Qi, X.: Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "22. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. In: International Conference on Learning Representations (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "type": "text", + "content": "23. Hendrycks, D., Mazeika, M., Dietterich, T.: Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 456 + ], + "type": "text", + "content": "24. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 457, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 457, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 133, + 457, + 480, + 500 + ], + "type": "text", + "content": "25. Hsu, Y.C., Shen, Y., Jin, H., Kira, Z.: Generalized ODIN: Detecting out-of-distribution image without learning from out-of-distribution data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10951-10960 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "text", + "content": "26. Huang, R., Geng, A., Li, Y.: On the importance of gradients for detecting distributional shifts in the wild. Advances in Neural Information Processing Systems 34, 677-689 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 480, + 566 + ], + "type": "text", + "content": "27. Joseph, K., Khan, S., Khan, F.S., Balasubramanian, V.N.: Towards open world object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5830-5840 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "text", + "content": "28. Joseph, K., Khan, S., Khan, F.S., Balasubramanian, V.N.: Towards open world object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5830-5840 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "29. Katz-Samuels, J., Nakhleh, J.B., Nowak, R., Li, Y.: Training ood detectors in their natural habitats. In: International Conference on Machine Learning. pp. 10848-10865. PMLR (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "30. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "text", + "content": "31. Kong, S., Ramanan, D.: Opengan: Open-set recognition via open data generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 813-822 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 151, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 151, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 151, + 481, + 194 + ], + "type": "text", + "content": "32. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Kamali, S., Popov, S., Malloci, M., Kolesnikov, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision 128(7), 1956-1981 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "type": "text", + "content": "33. Lee, K., Lee, H., Lee, K., Shin, J.: Training confidence-calibrated classifiers for detecting out-of-distribution samples. In: International Conference on Learning Representations (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "text", + "content": "34. Lee, K., Lee, K., Lee, H., Shin, J.: A simple unified framework for detecting out-of-distribution samples and adversarial attacks. Advances in Neural Information Processing Systems 31 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "type": "text", + "content": "35. Liang, S., Li, Y., Srikant, R.: Enhancing the reliability of out-of-distribution image detection in neural networks. In: International Conference on Learning Representations, ICLR 2018 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 481, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 340 + ], + "type": "text", + "content": "36. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 341, + 481, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 481, + 384 + ], + "type": "text", + "content": "37. Liu, J., Chang, C., Liu, J., Wu, X., Ma, L., Qi, X.: Mars3d: A plug-and-play motion-aware model for semantic segmentation on multi-scan 3d point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9372-9381 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 385, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 481, + 407 + ], + "type": "text", + "content": "38. Liu, W., Wang, X., Owens, J., Li, Y.: Energy-based out-of-distribution detection. Advances in Neural Information Processing Systems 33, 21464-21475 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 408, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 481, + 430 + ], + "type": "text", + "content": "39. Liu, X., Yang, H., Ravichandran, A., Bhotika, R., Soatto, S.: Continual universal object detection. arXiv preprint arXiv:2002.05347 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 431, + 481, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 481, + 463 + ], + "type": "text", + "content": "40. Miller, D., Dayoub, F., Milford, M., Sünderhauf, N.: Evaluating merging strategies for sampling-based uncertainty techniques in object detection. In: International Conference on Robotics and Automation (ICRA). pp. 2348-2354 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "type": "text", + "content": "41. Miller, D., Nicholson, L., Dayoub, F., Sünderhauf, N.: Dropout sampling for robust object detection in open-set conditions. In: IEEE International Conference on Robotics and Automation (ICRA). pp. 3243-3249 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "type": "text", + "content": "42. Nguyen, A., Yosinski, J., Clune, J.: Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 427-436 (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 532, + 481, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 481, + 563 + ], + "type": "text", + "content": "43. Perez-Rua, J.M., Zhu, X., Hospedales, T.M., Xiang, T.: Incremental few-shot object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13846-13855 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 565, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 481, + 609 + ], + "type": "text", + "content": "44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 609, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 481, + 632 + ], + "type": "text", + "content": "45. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "46. Rahman, S., Khan, S.H., Porikli, F.: Zero-shot object detection: Joint recognition and localization of novel concepts. International Journal of Computer Vision 128, 2979-2999 (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 447, + 101 + ], + "type": "text", + "content": "Can OOD Object Detectors Learn from Foundation Models?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "47. Ren, J., Fort, S., Liu, J., Roy, A.G., Padhy, S., Lakshminarayanan, B.: A simple fix to Mahalanobis distance for improving near-ood detection. arXiv preprint arXiv:2106.09022 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 151, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 151, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 151, + 480, + 183 + ], + "type": "text", + "content": "48. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems 28 (2015)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "type": "text", + "content": "49. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 480, + 239 + ], + "type": "text", + "content": "50. Sastry, C.S., Oore, S.: Detecting out-of-distribution examples with gram matrices. In: International Conference on Machine Learning. pp. 8491-8501. PMLR (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 480, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 480, + 273 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 480, + 273 + ], + "type": "text", + "content": "51. Sun, Y., Ming, Y., Zhu, X., Li, Y.: Out-of-distribution detection with deep nearest neighbors. In: International Conference on Machine Learning. pp. 20827-20840. PMLR (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 480, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 480, + 306 + ], + "type": "text", + "content": "52. Tack, J., Mo, S., Jeong, J., Shin, J.: CSI: Novelty detection via contrastive learning on distributionally shifted instances. Advances in Neural Information Processing Systems 33, 11839-11852 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 480, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 480, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 480, + 340 + ], + "type": "text", + "content": "53. Tan, H., Wu, S., Du, F., Chen, Y., Wang, Z., Wang, F., Qi, X.: Data pruning via moving-one-sample-out. Advances in Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 342, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 363 + ], + "type": "text", + "content": "54. Tao, L., Du, X., Zhu, X., Li, Y.: Non-parametric outlier synthesis. arXiv preprint arXiv:2303.02966 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "text", + "content": "55. Tian, Y., Fan, L., Chen, K., Katabi, D., Krishnan, D., Isola, P.: Learning vision from models rivals learning vision from data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15887-15898 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 397, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 430 + ], + "type": "text", + "content": "56. Wang, H., Liu, W., Bocchieri, A., Li, Y.: Can multi-label classification networks know what they don't know? Advances in Neural Information Processing Systems 34, 29074-29087 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 431, + 480, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 480, + 474 + ], + "type": "text", + "content": "57. Wang, X., Huang, T.E., Liu, B., Yu, F., Wang, X., Gonzalez, J.E., Darrell, T.: Robust object detection via instance-level temporal cycle confusion. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9143-9152 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 475, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 480, + 508 + ], + "type": "text", + "content": "58. Wen, X., Zhao, B., Chen, Y., Pang, J., Qi, X.: Generalization beyond data imbalance: A controlled study on clip for transferable insights. arXiv preprint arXiv:2405.21070 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 509, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 480, + 542 + ], + "type": "text", + "content": "59. Wilson, S., Fischer, T., Dayoub, F., Miller, D., Sunderhauf, N.: SAFE: Sensitivity-aware features for out-of-distribution object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23565-23576 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 543, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 480, + 575 + ], + "type": "text", + "content": "60. Wu, A., Chen, D., Deng, C.: Deep feature deblurring diffusion for detecting out-of-distribution objects. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13381-13391 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 576, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 480, + 609 + ], + "type": "text", + "content": "61. Wu, A., Deng, C.: Discriminating known from unknown objects via structure-enhanced recurrent variational autoencoder. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23956-23965 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 609, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 480, + 632 + ], + "type": "text", + "content": "62. Wu, Q., Chen, Y., Yang, C., Yan, J.: Energy-based out-of-distribution detection for graph neural networks. arXiv preprint arXiv:2302.02914 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "63. Wu, S., Tan, H., Tian, Z., Chen, Y., Qi, X., Jia, J.: Saco loss: Sample-wise affinity consistency for vision-language pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 27358-27369 (2024)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 226 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "64. Wu, X., Dai, P., Deng, W., Chen, H., Wu, Y., Cao, Y.P., Shan, Y., Qi, X.: Clnerf: continual learning of neural radiance fields for evolving scene representation. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 194 + ], + "type": "text", + "content": "65. Yu, F., Chen, H., Wang, X., Xian, W., Chen, Y., Liu, F., Madhavan, V., Darrell, T.: BDD100K: A diverse driving dataset for heterogeneous multitask learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2636-2645 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "66. Zheng, H., Wang, Q., Fang, Z., Xia, X., Liu, F., Liu, T., Han, B.: Out-of-distribution detection learning with unreliable out-of-distribution sources. Advances in Neural Information Processing Systems 36, 72110-72123 (2023)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "text", + "content": "Can OOD Object Detectors Learn from Foundation Models?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Can Textual Semantics Mitigate Sounding Object Segmentation Preference_/88d8a449-254f-4c3c-bdf6-605afa985fb1_content_list.json b/2024/Can Textual Semantics Mitigate Sounding Object Segmentation Preference_/88d8a449-254f-4c3c-bdf6-605afa985fb1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2f842530e0890d4ce888fca4120fed4865418e7b --- /dev/null +++ b/2024/Can Textual Semantics Mitigate Sounding Object Segmentation Preference_/88d8a449-254f-4c3c-bdf6-605afa985fb1_content_list.json @@ -0,0 +1,1846 @@ +[ + { + "type": "text", + "text": "Can Textual Semantics Mitigate Sounding Object Segmentation Preference?", + "text_level": 1, + "bbox": [ + 217, + 140, + 785, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yaoting Wang $^{1\\dagger}$ , Peiwen Sun $^{2\\dagger}$ , Yuanchao Li $^{3}$ , Honggang Zhang $^{2}$ , and Di Hu $^{\\text{空}1,4}$", + "bbox": [ + 222, + 214, + 781, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Gaoling School of Artificial Intelligence, Renmin University of China, China", + "bbox": [ + 241, + 257, + 759, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yaoting.wang@outlook.com", + "bbox": [ + 406, + 273, + 596, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dihu@ruc.edu.cn", + "bbox": [ + 441, + 287, + 560, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Beijing University of Posts and Telecommunications, Beijing, China", + "bbox": [ + 267, + 300, + 732, + 314 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{sunpeiwen,zhhg}@bupt.edu.cn", + "bbox": [ + 393, + 315, + 609, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 University of Edinburgh, Scotland, UK", + "bbox": [ + 364, + 328, + 635, + 340 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yuanchao.li@ed.ac.uk", + "bbox": [ + 421, + 343, + 581, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Engineering Research Center of Next-Generation Search and Recommendation", + "bbox": [ + 233, + 354, + 767, + 369 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. The Audio-Visual Segmentation (AVS) task aims to segment sounding objects in the visual space using audio cues. However, in this work, it is recognized that previous AVS methods show a heavy reliance on detrimental segmentation preferences related to audible objects, rather than precise audio guidance. We argue that the primary reason is that audio lacks robust semantics compared to vision, especially in multi-source sounding scenes, resulting in weak audio guidance over the visual space. Motivated by the fact that text modality is well explored and contains rich abstract semantics, we propose leveraging text cues from the visual scene to enhance audio guidance with the semantics inherent in text. Our approach begins by obtaining scene descriptions through an off-the-shelf image captioner and prompting a frozen large language model to deduce potential sounding objects as text cues. Subsequently, we introduce a novel semantics-driven audio modeling module with a dynamic mask to integrate audio features with text cues, leading to representative sounding object features. These features not only encompass audio cues but also possess vivid semantics, providing clearer guidance in the visual space. Experimental results on AVS benchmarks validate that our method exhibits enhanced sensitivity to audio when aided by text cues, achieving highly competitive performance on all three subsets. Project page: https://github.com/GeWu-Lab/Sounding-Object-Segmentation-Preference", + "bbox": [ + 259, + 438, + 743, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Audio-Visual Segmentation $\\cdot$ Textual Guidance $\\cdot$ Multi-modal learning", + "bbox": [ + 259, + 760, + 740, + 789 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Equal contribution.", + "bbox": [ + 230, + 810, + 377, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\mathbf{\\Pi}^{\\text{三}}$ Corresponding author.", + "bbox": [ + 230, + 824, + 398, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/654b534457e0ff9452c91ebdf4e1d512b7c36414ac7bfb9b2ff7b9260d0c95d8.jpg", + "image_caption": [ + "Fig. 1: The previous methods (left) achieve satisfactory results in receiving normal audio input. However, even when the sound is completely silent, they still segment vast pixels to represent the guitar regarding the segmentation preference of audible objects built during the training. In contrast, our approach (right) utilizes a frozen LLM to reason from scene descriptions that the male is not playing the guitar, as his hands are off the strings. Guided by this semantic information, our method produces more precise segmentation with finer audio guidance." + ], + "image_footnote": [], + "bbox": [ + 246, + 143, + 754, + 386 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 526, + 375, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In human perception, visual and auditory senses are closely related, and audio can provide vision with additional and dynamic scene information to enhance visual understanding. In the context of audio-visual understanding tasks, Audio-Visual Localization (AVL) enables the localization of sounding objects within visual scenes using audio references [15, 36]. However, the coarse localization of sounding objects is no longer sufficient to meet the practical demands of complex scene understanding, such as in autonomous driving [12, 49] and augmented reality [33]. To address this challenge, tasks are gradually shifting from bounding-box or rough heat-map localization to finer-grained pixel-level segmented localization, which is known as Audio-Visual Segmentation (AVS).", + "bbox": [ + 212, + 566, + 787, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The existing works on AVS can be broadly categorized into fusion-based methods [8, 16, 20, 21, 29, 47] and prompt-based methods [30, 34, 40]. The former primarily focuses on localizing sounding objects by fusing audio and visual features, while the latter emphasizes constructing effective audio prompts for the visual foundation model. However, most of these methods fail to establish compact audio-visual correlations to achieve clear audio guidance in the visual space. As depicted in the left of Fig. 1, previous methods could achieve satisfactory results with normal audio input. Nevertheless, when the audio is silent,", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "these methods still segment many pixels to represent potential sounding objects based on the segmentation preferences of audible objects1.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, from the experiment with manually muted audio input, we observe that most of the behavior of AVS baselines depends on the segmentation preference established during the training process rather than relying on reliable audio-visual correlation. For instance, these models tend to take a shortcut by segmenting out the muted guitars because the audio guidance is weak, and they simply learned with the easy-to-learn visual feature alone (i.e., shortcut learning) that guitars are often associated with sound-emitting during the training phase. The unreliable audio-visual correlation can be attributed to two main factors. Firstly, the scarcity of training data for AVS is a significant challenge due to the demanding nature of pixel-level annotation. Secondly, the audio modality itself presents inherent complexity and ambiguity [16, 21, 29], especially in scenarios involving multiple audio sources that may be intertwined.", + "bbox": [ + 212, + 178, + 787, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the challenges discussed above, our approach aims to build a robust audio-visual correlation by integrating the text modality, which inherently provides meaningful and robust semantic information about objects and their interactions. Firstly, in the Scene Describing phase, we utilize an off-the-shelf image captioner (LLaVA-1.5 [28]) to generate detailed dense captions for scene understanding. Then, in the Sounding Objects Reasoning phase, we instruct the frozen large language models (LLMs, e.g., LLaMA2 [39]) as a text cues capturer to collect potential sounding objects from the scene description as text cues. Notably, we introduce an elaborate few-shot prompt template to guide the text cues capturer to reason with Chain-of-Thought (CoT) instructions. After that, in the Semantics-Driven Audio Modeling (SeDAM) module, we project the audio feature into latent audio features. However, instead of directly using these latent features to interact with visual features for mask decoding, we leverage these latent features to interact with text cues. Specifically, the latent features are attended to and collected to form sounding object features with semantics provided by the text cues through the crossmodal transformer. We further propose the Prompting Mask Queries with Semantics (PMQS) module to introduce the sounding object features into pre-trained mask queries for later segmentation. Finally, we use simple but effective bottleneck adapters in the Audio-Prompted Decoding phase for better segmentation quality. The experimental results demonstrate that our proposed method improves audio-visual grounding by incorporating guidance from both audio and text cues.", + "bbox": [ + 212, + 359, + 787, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Experiments in Sec. 4.4 reveal the advantage of our method, as verified by its increased sensitivity to the changes in audio input compared to networks without enhancing audio guidance with text cues. In summary, our contributions are threefold:", + "bbox": [ + 212, + 691, + 787, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We recognize the behavior of AVS model is influenced by the preference segmentation phenomenon, which can not be solved easily by previous audio-", + "bbox": [ + 225, + 768, + 787, + 800 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Sounding Object Segmentation Preference", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1 To be clear, we define audible objects as objects capable of producing sound, while sounding objects are defined as objects that are currently producing sound.", + "bbox": [ + 217, + 810, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "visual based methods. A novel strategy introduces textual semantics as the bridge into the AVS task, establishing a better audio-visual correlation.", + "bbox": [ + 240, + 146, + 784, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- We introduce the SeDAM module to form sounding object features with audio and text cues. We also suggest a PMQS module to prompt the pretrained mask queries with these sounding object features for sounding object segmentation using a visual foundation model.", + "bbox": [ + 225, + 176, + 782, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Our experiments show that our method can be more sensitive to the audio guidance, manifested in its high sensitivity to the changes in audio input.", + "bbox": [ + 225, + 238, + 782, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 292, + 395, + 308 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Audio-Visual Localization and Segmentation", + "text_level": 1, + "bbox": [ + 215, + 325, + 627, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Traditional AVL tasks [15,35,36] predict the positions of sounding objects using bounding-box or coarse heat-maps through unsupervised learning of audio-visual correlation. In recent years, driven by the growing demand for more precise localization in industries such as autonomous driving [12,49] and augmented reality [33], the AVL task has evolved towards a finer-grained AVS task that localizes sounding objects at a pixel-level. The existing AVS works can be broadly categorized into fusion-based [8, 16, 20, 21, 29, 43, 47] and prompt-based [30, 32, 40] The pioneering AVS work [47] is fusion-based and adopts a multi-stage strategy to integrate audio with multi-scale visual features. Liu et al. [29] address the issues of inadequate fusion of audio-visual features with an audio-aware query-enhanced transformer (AuTR). In comparison, AVSegFormer [8] directly decodes the fused feature with the audio query. Wang et al. [40] innovatively prompt the visual foundation model with audio input, harnessing the abundant visual before achieving generalizable AVS in zero-shot and few-shot scenarios. However, these methods have not effectively established a reliable audio-visual correlation, preventing the models from perceiving robust audio guidance. On the contrary, our experiments in Sec. 4.4 also reveal that some AVS baselines' competitive performance may stem from segmentation preferences formed during the training.", + "bbox": [ + 212, + 351, + 787, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Text Aided Scene Understanding", + "text_level": 1, + "bbox": [ + 215, + 647, + 535, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Before the explosion of LLM, textual semantic information had been widely used for visual understanding tasks. Zhan et al. [38] and Sharma et al. [37] improve Visual Question Answering by using the extra semantic information of image captions, while Hur and Park [17] utilize the image captioner to help zero-shot image classification. Wang et al. [41] incorporating text with audio and visual cues to solve reference AVS task.", + "bbox": [ + 212, + 672, + 785, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In recent years, the community has witnessed an emergent interest in strong off-the-shelf capabilities in open-world visual understanding [24-26, 45]. Notable examples, such as GPT4(V) [44] and LLaVA [27] have showcased remarkable linguistic and visual capabilities in zero-shot scene understanding [7, 42] without any training requirement. Innovatively, in the field of AVS, BAVS [23] replaces", + "bbox": [ + 212, + 765, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3b7384491ed32f0e29fa98610f0ed5b3dcd66374660b90b82c87554356c724c4.jpg", + "image_caption": [ + "Fig. 2: Overall pipeline of the proposed TeSO method. We utilize an off-the-shelf image captioner for dense scene describing and employ a frozen LLM to reason out potential sounding objects as text cues. These semantic text cues are then aggregated with audio features in the SeDAM module to form the sounding object features. Subsequently, we introduce the sounding object features into pre-trained mask queries in the PMQS module. Finally, we use adapters to tune the visual-only mask decoder for AVS in the Audio-prompted Decoding phase. \"MSDA\" is the multi-scale deformable attention proposed by Zhu et al. [48]." + ], + "image_footnote": [], + "bbox": [ + 235, + 143, + 767, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "audio input with textual audio tags from a large pre-trained vision (UniDiffuser [2]) and audio (BEATs [3]) foundation model, establishing a reliable AVS system with foundation knowledge. However, due to the significant gap in class distribution between the pre-training dataset and the AVS dataset, as well as the presence of multi-source scenarios in the AVS task, the extracted audio tags are not ideal to provide accurate semantics. In contrast, we utilize the text modality with robust semantic features to enhance rather than replace the audio guidance and address the segmentation preference issue.", + "bbox": [ + 212, + 481, + 787, + 603 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Text-guided Sounding Object Segmentation", + "text_level": 1, + "bbox": [ + 214, + 626, + 674, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 2, we introduce our Text-guided Sounding Object Segmentation (TeSO) method in this section. Our experiments about muted audio input imply that audio does not exert reasonable guidance over the visual space for some AVS baselines, resulting in the models relying more on the segmentation preferences established during the training process. In essence, they learn which objects are most likely audible during training and take shortcuts during inference. To address the above issue, our method aims to enhance the audio-visual correlation by leveraging the text modality, which inherently possesses robust semantic information, to obtain finer-grained audio guidance. We begin by acquiring detailed scene descriptions through an off-the-shelf image captioner. Subsequently, a frozen LLM works as the text cues capturer, collecting potential sounding objects as text cues from the scene descriptions with CoT instructions.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Sounding Object Segmentation Preference", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we introduce a novel SeDAM module with a dynamic mask to seamlessly integrate audio features and text semantics through a crossmodal transformer, providing finer audio guidance with text semantic cues.", + "bbox": [ + 212, + 146, + 787, + 193 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Multimodal Representation", + "text_level": 1, + "bbox": [ + 214, + 213, + 491, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual: Following previous works [40, 47], we sample frames from the video at 1-second intervals. Mask2Former [4] serves as our visual foundation model, and we extract visual features $F_V \\in \\mathbb{R}^{d_V \\times H \\times W}$ from a pre-trained Swin transformer [6]. We use simple two-layer MLP adapters to tune the visual encoder for better visual representations.", + "bbox": [ + 212, + 237, + 785, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Audio: Similar to the video processing, we split the audio into clips at 1-second intervals. We extract audio features $F_{As} = \\{F_{A_1}, F_{A_2}, \\dots, F_{A_T}\\} \\in \\mathbb{R}^{T \\times d_A}$ using VGGish [9, 13], where $T$ represents the length of the audio in seconds, matching the number of video frames. Therefore, each video frame corresponds to $F_{A} = F_{As}[i]$ . We freeze the pre-trained parameters for a better comparison with previous works.", + "bbox": [ + 214, + 313, + 785, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text: For each video frame, we complement the captured text cues into a sentence using a template (\"This is a $\\{\\_ \\}$ .\") to extract text features $F_{T} \\in \\mathbb{R}^{N_{T} \\times d_{T}}$ with ImageBind [10], where $N_{T}$ is the number of text cues.", + "bbox": [ + 214, + 404, + 785, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Text Cues Generation", + "text_level": 1, + "bbox": [ + 214, + 470, + 444, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To obtain more precise text cues, we use a two-stage inference flow. Firstly, we employ an off-the-shelf image captioner to generate dense scene descriptions. Secondly, we ask a text cues capturer to collect potential sounding objects from the generated descriptions.", + "bbox": [ + 212, + 494, + 785, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scene Describing. As shown in the top-left of Fig. 3, in this phase, we employ the commonly used and straightforward prompt: \"Please carefully understand the image content and provide as rich a description as possible.\" This instructs the image captioner to generate a detailed description of the visual scene, a task commonly referred to as dense captioning [18]. This process enables the image captioner to produce comprehensive and informative descriptions that not only describe the objects like \"a human and guitar in the image\" but also capture their interactions and relationships.", + "bbox": [ + 212, + 575, + 787, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sounding Object Collecting. To extract potential sounding objects from the generated dense descriptions, we design a task-specific CoT prompt template with few-shot demonstrations, as illustrated in Fig. 3. Specifically, we carefully craft examples that cover various scenarios and provide reasoning approaches for each example. The process begins by identifying the objects present in the dense scene descriptions, followed by step-by-step common-sense reasoning according to the CoT instructions, as depicted in the lower right of Fig. 3. During the CoT reasoning, the text cues capturer first categorizes the sample objects and", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9b78b8aefc443ced04e43ae181a90ca207178e1f8a63d5cdb0ba66a8549c3e3e.jpg", + "image_caption": [ + "Fig. 3: \"P.S.O\" stands for \"Potential Sounding Object\". A frozen LLM reasoner works as a text cues capturer by considering the interaction of audible objects. For instance, a person may sing a song while playing the guitar, but he would not sing along with a saxophone. In contrast, a noun parser simply captures any nouns that are present." + ], + "image_footnote": [], + "bbox": [ + 318, + 143, + 454, + 348 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9eeb3064ba0d82e908aa542331c861dda184f5a60a1efcf17b9011ca2e9e3a82.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 454, + 142, + 684, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "assesses whether objects in that category are likely making sounds. Subsequently, the capturer analyzes which group of potential sounding objects can produce sounds simultaneously. Finally, the group with higher confidence is captured as the text cues of length $N_{T}$ and encoded in $F_{T}$ for further guidance. The right part of Fig. 3 also illustrates why we use the frozen LLM as a text cues capturer instead of using a naive noun parser. It is important to note that the use of LLM is solely for generating high-quality text cues, the LLM itself is not engaged in the training process for a fairness concern.", + "bbox": [ + 212, + 449, + 787, + 569 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Semantic-Driven Audio Modeling", + "text_level": 1, + "bbox": [ + 215, + 595, + 539, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After obtaining potential sounding objects, we can enhance the semantics of audio features with these text cues in this SeDAM module. To allow semantic text cues to capture different aspects of audio features $F_{A}$ , we project $F_{A}$ into distinct latent features $F_{A_{L}} = \\{a_{i}\\}_{i=1}^{N_{L}} \\in \\mathbb{R}^{N_{L} \\times d_{A_{L}}}$ , where $N_{L}$ is the number of latent features. It is worth noting that, we incorporate an auxiliary loss term $\\mathcal{L}_{infoNCE}$ aims at ensuring the distinction among latent audio features. This way, we project audio features into the latent space, where each unique $a_{i}$ represents a latent audio feature to be captured by semantic text cues, forming clear and loud sounding object features with robust semantics for finer guidance.", + "bbox": [ + 212, + 625, + 787, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To incorporate the semantics of text cues, we employ text cues as the query to attend and combine the relevant latent features through crossmodal multi-head attention, formulated as:", + "bbox": [ + 212, + 761, + 787, + 806 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\widetilde {A} _ {L}} = \\operatorname {M u l t i H e a d} \\left(F _ {T}, F _ {A _ {L}}, F _ {A _ {L}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 374, + 824, + 785, + 843 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Sounding Object Segmentation Preference", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $F_{\\widetilde{A}_L} \\in \\mathbb{R}^{N_T \\times d_A}$ is the attended latent audio features and will be combined with text cues $F_T$ as follows:", + "bbox": [ + 212, + 145, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nF _ {A _ {T}} = F C \\left(L N \\left(F _ {T} + F _ {\\widetilde {A} _ {L}}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 191, + 785, + 212 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $F_{A_T} \\in \\mathbb{R}^{N_T \\times d_A}$ is the combined feature with both audio information and semantic text cues, and $FC$ and $LN$ for linear layer and layer normalization. Note that the LLM reasoner may generate a different number of text cues, and we either pad with a zero matrix or truncate to $N_T$ . It is noteworthy that we propose task-specific CoT instructions to prompt the text cues capturer to infer more accurate potential sounding objects as the text cues. Additionally, we also introduce a dynamic mask based on attention score to mitigate the potential noisy information.", + "bbox": [ + 212, + 218, + 784, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We first compute the average pooling attention scores $\\zeta \\in \\mathbb{R}^{N_T}$ for text cues on the attention head and key dimension:", + "bbox": [ + 212, + 340, + 782, + 371 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\xi = \\operatorname {P o o l} \\left(\\frac {F _ {T} F _ {A _ {L}} ^ {\\mathrm {T}}}{\\sqrt {d _ {A _ {L}}}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 382, + 785, + 424 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\sqrt{d_{A_L}}$ is the scaling factor for dot production. Then we form the attention mask $\\mathcal{M} = \\{m_1, m_2, m_3, \\dots, m_{N_T}\\}$ as follows:", + "bbox": [ + 212, + 436, + 782, + 468 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {i} = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} \\xi_ {i} > 0 \\\\ - \\infty & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 479, + 785, + 520 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{M} \\in \\mathbb{R}^{N_T}$ , will be used as a key padding mask for the PMQS module in the next section to ensure that the attention weight of redundant text cues approaches zero.", + "bbox": [ + 212, + 531, + 782, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Prompting Mask Queries with Semantics", + "text_level": 1, + "bbox": [ + 214, + 601, + 599, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Current segmentation foundation models [4, 5, 19] adopt transformer-based decoder with mask queries as the mainstream approach. To prompt such foundation models, we propose the PMQS module, aiming to introduce semantic information to mask query features $F_{Q} = \\{q_{1}, q_{2}, \\dots, q_{N_{Q}}\\}$ , where $F_{Q} \\in \\mathbb{R}^{N_{Q} \\times d_{V}}$ , and $N_{Q}$ is the number of mask queries in crossmodal attention. We formulate this updated expression as follows:", + "bbox": [ + 214, + 628, + 782, + 719 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {c a} \\left(F _ {Q}, F _ {A _ {T}}, F _ {A _ {T}}\\right) = S o f t m a x \\left(\\frac {F _ {Q} F _ {A _ {T}} ^ {\\mathrm {T}}}{\\sqrt {d _ {A _ {T}}}} + \\mathcal {M}\\right) F _ {A _ {T}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 731, + 785, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{M}$ is added to the attention score to mask the redundant text cues for each query. Then we obtain the updated mask query features $\\widetilde{F}_Q\\in \\mathbb{R}^{N_Q\\times d_V}$ as:", + "bbox": [ + 212, + 777, + 782, + 809 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {F} _ {Q} = F _ {Q} + \\phi_ {c a} \\left(F _ {Q}, F _ {A _ {T}}, F _ {A _ {T}}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 823, + 785, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Audio-Prompted Decoding", + "text_level": 1, + "bbox": [ + 215, + 146, + 483, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As shown on the right side of Fig. 2, the updated mask query features $\\widetilde{F}_Q$ are forwarded into the transformer-based decoder, generating mask queries $M_Q \\in \\mathbb{R}^{N_Q \\times H \\times W}$ and class queries $C_Q \\in \\mathbb{R}^{N_Q \\times N_C}$ , where $N_C$ is the number of class queries pre-defined in the segmentation foundation model. Subsequently, we can obtain the final segmentation $M_{pred} \\in \\mathbb{R}^{N_C \\times H \\times W}$ :", + "bbox": [ + 212, + 170, + 782, + 247 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nM _ {p r e d} = \\operatorname {E i n s u m} \\left(C _ {Q}, \\bar {F} _ {Q}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 260, + 785, + 277 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\text{Einsum}(\\cdot)$ is the Einstein Summation Convention.", + "bbox": [ + 214, + 287, + 629, + 303 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Notably, to enhance the quality of the segmentation, we use two-layer bottle-neck MLP adapters [14] to lightly fine-tune the output projection of the following attentions: a) Swin Window Attention (W-MSA) [31], b) Deformable Attention in pixel decoder [48] and c) Masked Attention in transformer decoder [4].", + "bbox": [ + 212, + 304, + 782, + 364 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.6 Learning Objectives", + "text_level": 1, + "bbox": [ + 214, + 386, + 428, + 402 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Segmentation Loss. During the model training process, we employ the binary cross-entropy loss and dice loss to optimize the mask quality:", + "bbox": [ + 214, + 411, + 782, + 441 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {b c e} \\cdot \\mathcal {L} _ {b c e} + \\lambda_ {d i c e} \\cdot \\mathcal {L} _ {d i c e}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 455, + 784, + 469 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Further, we adopt the mask classification loss [5] to compose the final segmentation loss:", + "bbox": [ + 214, + 481, + 782, + 508 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s e g} = \\mathcal {L} _ {m a s k} + \\lambda_ {c l s} \\cdot \\mathcal {L} _ {c l s}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 513, + 784, + 527 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Latent Component Loss. This loss term aims to explicitly guide the model to learn distinct representations for each latent audio component, helping prevent redundancy and overlapping information across the latent components.", + "bbox": [ + 214, + 534, + 782, + 580 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {i n f o N C E}} = - \\log \\left(\\frac {\\exp \\left(\\operatorname {s i m} \\left(a , a ^ {+}\\right)\\right)}{\\sum_ {i = 1} ^ {N _ {L}} \\exp \\left(\\operatorname {s i m} \\left(a , a _ {i}\\right)\\right)}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 590, + 784, + 631 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $a^+$ and $sim(\\cdot)$ denote the positive example and the cosine similarity.", + "bbox": [ + 214, + 641, + 754, + 656 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Total Loss The final loss function is composed of the weighted sum of the two kinds of losses above:", + "bbox": [ + 214, + 657, + 782, + 685 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {s e g}} + \\lambda_ {\\text {i n f o N C E}} \\cdot \\mathcal {L} _ {\\text {i n f o N C E}}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 700, + 784, + 715 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 737, + 375, + 753 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 768, + 455, + 784 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dataset. Our proposed method is evaluated on the AVS Benchmarks [47], which contains three subsets. Firstly, the single-source subset (V1S) contains 4932 videos over 23 categories, covering various sounds, such as humans, animals,", + "bbox": [ + 212, + 794, + 782, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Sounding Object Segmentation Preference", + "bbox": [ + 447, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "vehicles, and musical instruments. For videos in the training split of this subset, only the first sampled frame is annotated. Secondly, the multi-source subset (V1M) contains 424 videos that include two or more categories from V1S, and all sounding objects are visible in the frames. Finally, the AVSBench-semantic (AVSS) subset, an extension of V1S and V1M, contains 12,356 videos. These newly collected videos are trimmed to 10 seconds, differing from the 5 seconds in V1S and V1M.", + "bbox": [ + 212, + 146, + 787, + 251 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Setting. We conduct training and evaluation using the VGGish backbone pretrained on Youtube-8M [1] and Swin-base Transformer backbone pretrained on semantic-ADE20K [46] by Mask2Former [4]. The number $N$ of the crossmodal transformer layers is set to 4, and the parameters $\\lambda_{bce},\\lambda_{dice},\\lambda_{cls},\\lambda_{infoNCE}$ in the loss are set to 5,5,2,1. The AdamW optimizer is adopted with a learning rate of 1e-4 for the visual encoder adapters and 1e-3 for other learnable parameters, and the training epoch is roughly set to 60.", + "bbox": [ + 212, + 252, + 787, + 357 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Metrics. To conduct a comprehensive evaluation of our model, we carry out tests using mean Intersection over Union (mIoU) and F-score as the performance metrics, following previous works [8, 47].", + "bbox": [ + 212, + 358, + 787, + 405 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Comparison Results", + "text_level": 1, + "bbox": [ + 214, + 424, + 429, + 439 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "When compared to methods that do not incorporate text as additional semantic information, our method, which leverages captioning and reasoning, demonstrates strong competitiveness. As illustrated in Sec. 4.2, our method achieves comparable results across all V1S, V1M, and AVSS-binary subsets, showcasing average performance improvements of up to $1.25\\%$ in mIoU and $2.7\\%$ in F-score, respectively. Furthermore, qualitative analysis indicates that our predicted masks exhibit superior quality, as demonstrated in sub-figure (1) of Fig. 4 (more examples in supplementary materials).", + "bbox": [ + 212, + 446, + 787, + 568 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "It is necessary to mention that our proposed TeSO does not show significant improvements on V1S over certain existing methods, such as BAVS [23], which also utilizes text semantic information. Apart from marginal effects, this phenomenon can also be attributed to the singular sound source and simplistic nature of the data scenarios present in V1S. On V1M, however, our performance significantly surpasses that of other methods, including BAVS.", + "bbox": [ + 212, + 569, + 787, + 657 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In addition to the AVS task, we also conduct experiments on the Audio-Visual Semantic Segmentation (AVSS) task, as shown in Tab. 2. Our model exceeds the best performance in previous models by $5.37\\%$ on mIoU, which further demonstrates that our model is capable of establishing a better audio-visual correlation.", + "bbox": [ + 212, + 659, + 787, + 734 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Bottlenecks and Fairness", + "text_level": 1, + "bbox": [ + 214, + 755, + 467, + 768 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Our method (TeSO) collects text cues from dense image descriptions at $\\mathrm{REC}_{\\#1}$ $(80.41\\%)$ and a filter with a dynamic mask at $\\mathrm{REC}_{\\#2}$ $(80.95\\%)$ , indicating a robust startup for the text-guided AVS framework, in comparison with another two-stage model, BAVS.", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/fc507692e3d3bdf620ed78ac8dca7706b0dbbf5a718d386e804495a857c21969.jpg", + "table_caption": [ + "Table 1: Performance on AVS-Benchmarks. The underscore is used to indicate suboptimal results. BAVS uses large pre-trained visual and audio foundation models to replace audio input with text tag. In comparison, TeSO (ours) enhances rather than replaces audio guidance based on the robust semantics of text. It is evident that our method has a good F-score result while ensuring mIoU." + ], + "table_footnote": [ + "$\\dagger$ : To make a fair comparison, the results of CATR here are without supplemented annotation of the training set.", + ": BEATs is a stronger audio backbone, trained on AudioSet dataset with 90M parameters. In comparison, VGGish is trained on Youtube-8M with 70M parameters." + ], + "table_body": "
MethodAudio-backboneVisual-backboneV1SV1MAVSS-binary
mIoU(%)F-scoremIoU(%)F-scoremIoU(%)F-score
AVSBench [47]VGGishPVT-v278.700.87954.000.64562.450.756
AVSegFormer [8]VGGishPVT-v282.060.89958.360.69364.340.759
AVSC [22]VGGishPVT-v281.290.88659.500.657--
AVS-BG [11]VGGishPVT-v281.710.90455.100.668--
AQFormer [16]VGGishPVT-v281.600.89461.100.721--
†CATR [20]VGGishPVT-v281.400.89659.000.700--
AV-SAM [34]ResNet18ViT-Base40.470.566----
SAMA [30]VGGishViT-Huge81.530.88663.140.691--
GAVS [40]VGGishViT-Base80.060.90263.700.77467.700.788
AuTR [29]VGGishSwin-Base80.400.89156.200.672--
MUTR [43]VGGishVideo-Swin-Base81.600.89764.000.735--
BAVS [23]‡BEATsSwin-Base82.680.89859.630.65955.450.640
TeSO (ours)VGGishSwin-base83.270.93366.020.80168.530.813
", + "bbox": [ + 225, + 196, + 772, + 363 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/0a21a4be6dabfc189cea795ffc70a0c44acd6ef62119d25db1c2cb6f230ec6e2.jpg", + "table_caption": [ + "Table 2: Performance on AVSS dataset. Our method showcases a substantial improvement in the semantic subset, achieved by leveraging robust text semantics to enhance audio guidance." + ], + "table_footnote": [], + "table_body": "
MethodBackboneAVSS
mIoU(%)F-Score
AVSBenchPVT-v229.800.352
CATRPVT-v232.800.385
BAVSSwin-base33.590.375
TeSO (Ours)Swin-base38.960.451
", + "bbox": [ + 227, + 505, + 470, + 582 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/d6459ea934d376eb43a55aea806421fa31f72c862f0b7f615b89540294d8c873.jpg", + "table_caption": [ + "Table 3: Potential sounding objects recognition (REC) performance at different model stages and the corresponding segmentation results. Experiments are conducted on AVS-V1M test set." + ], + "table_footnote": [], + "table_body": "
MethodREC(#1)AP(%)REC(#2)AP(%)SEGmIoU(%)
BAVS24.2279.6959.63
TeSOp(ours)77.9678.3363.87
TeSOn(ours)77.8278.2763.31
TeSO(ours)80.4180.9566.02
", + "bbox": [ + 524, + 505, + 772, + 582 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In addition, we develop several variants of TeSO to investigate its performance sources and bottlenecks, as well as its fairness in comparison with other methods. $\\mathrm{TeSO}_p$ denotes the variant using one-shot prompts, while $\\mathrm{TeSO}_n$ denotes the variant employing a noun parser instead of a frozen LLM for text cues collection. As shown in Tab. 3, the results demonstrate that both the number (shots) of demonstrations and the method of text cues collection significantly impact the model's ability to accurately identify audio semantics and perform segmentation (SEG). Thus, utilizing few-shot prompts and off-the-shelf foundation models are both essential for ensuring the acquisition of high-quality text cues (i.e., the bottleneck).", + "bbox": [ + 212, + 612, + 785, + 763 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "However, we would like to emphasize that, even without using LLM to capture text cues, our method variant $(\\mathrm{TeSO}_n)$ still fairly achieves competitive performance (63.31% mIoU) compared with BAVS (59.63% mIoU) and state-of-the-art method GAVS (63.70% mIoU) in the complex multi-source scenario (AVS-V1M), and significantly outperforms other methods. This suggests that", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Sounding Object Segmentation Preference", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/73bdf3b3cef44a48d43bb36ffc69215e1df2f2266cc88f1ce07113188e464339.jpg", + "table_caption": [ + "Table 4: Impact on mIoU and F-score by muted audio or noise-only audio on AVS-V1M test set. TeSO (ours) exhibits the most intensive attenuation up to $48.8\\%$ . It is noteworthy that the other models still generate \"satisfactory\" results even in silent or purely noisy environments. This observation suggests that the models might rely more on the segmentation preferences formed during the training rather than truly learning efficient audio guidance." + ], + "table_footnote": [], + "table_body": "
MethodMuteWGN-10dBWGN-40dB
mIoU% (Δm%)F-score (Δf%)mIoU% (Δm%)F-score (Δf%)mIoU% (Δm%)F-score (Δf%)mIoU% (Δm%)
AVS-Bench49.15 (6.6)0.622 (4.9)46.76 (11.1)0.618 (5.5)47.46 (9.8)0.610 (6.7)
GAVS42.53 (29.2)0.695 (8.4)41.51 (34.8)0.692 (10.6)41.87 (30.3)0.696 (8.4)
BAVS45.06 (26.1)0.644 (5.8)43.27 (27.4)0.601 (8.8)42.93 (31.6)0.595 (9.7)
TeSO (ours)37.89 (42.6)0.671 (16.2)34.20 (49.6)0.661 (17.5)34.74 (47.4)0.656 (18.1)
", + "bbox": [ + 238, + 239, + 759, + 305 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "leveraging text-enhanced audio for visual control with robust semantics is a viable and promising direction.", + "bbox": [ + 212, + 330, + 785, + 359 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Audio Control", + "text_level": 1, + "bbox": [ + 214, + 380, + 382, + 395 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Based on our observations, we notice a distinct shape of the predicted mask in sub-figure (2) of Fig. 4 when some popular AVS baselines are used for inference. Interestingly, this shape remains distinguishable even when the inference is conducted with muted audio. This suggests that the network can predict rough masks without relying on audio input, indicating that the audio-visual grounding is not effectively guided by audio cues. As a result, this undesirable phenomenon can be attributed to the segmentation preference, which is a result of these models lacking proper audio control.", + "bbox": [ + 212, + 402, + 785, + 523 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To evaluate the influence of audio control on segmentation, various models are subjected to inference using muted audio and noise-only audio (at 10 dB and 40 dB) instead of the corresponding audio signals. It should be noted that the visual features remained unchanged throughout the evaluation. As shown in Tab. 4, the disparity $\\Delta_{m}$ (and $\\Delta_{f}$ ) refers to the difference in mIoU (and F-score) values before and after the audio substitution is applied as an evaluation metric for the sensitivity of audio changes. A larger drop in mIoU and F-score indicates a greater degree of audio control over the visual space. In addition to demonstrating strong segmentation capabilities in the AVS task, TeSO (ours) exhibits the highest level of attenuation, reaching up to $47.4\\%$ , when subjected to muted and noise-only audio. Additionally, qualitative analysis revealed that the predicted masks remained predominantly blank, as depicted in sub-figure (2) of Fig. 4 (more examples in supplementary materials). This observation indicates that our model has better audio guidance over the visual space, enabling it to play a crucial role in audio-visual grounding.", + "bbox": [ + 212, + 525, + 787, + 752 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 771, + 390, + 787 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "As shown in Tab. 5, we conduct ablation experiments targeting the designed TeSO modules and the guidance from different modalities. To explore the guiding ability, we progressively remove either the audio or text guidance for TeSO.", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/bf36008771978a66a07336716644a2120701f1bf9601010fb270afffcc490b7a.jpg", + "image_caption": [ + "Fig. 4: Examples of the impact of normal audio input and all-mute audio on popular methods. In normal scenarios, our method shows better masks than previous methods. In all-mute scenarios, our approach exhibits strong sensitivity towards audio inputs, as it is capable of generating blank masks for silent audio clips." + ], + "image_footnote": [], + "bbox": [ + 272, + 143, + 727, + 362 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The experiments indicate that control from both audio and text inputs for visual space can establish a strong audio-visual correlation. Furthermore, by utilizing our proposed TeSO, we achieve enhanced audio guidance and reach a $66.02\\%$ mIoU while incorporating precise semantic information from the text modality. Moreover, through the utilization of our proposed TeSO method that incorporates robust semantic information from the text modality, we achieved improved audio guidance and attained a mIoU of $66.02\\%$ .", + "bbox": [ + 212, + 465, + 787, + 571 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As shown in Tab. 6, we also implement a baseline that directly filters the inferred semantic labels during the model evaluation process by intersecting them with the potential sounding object collected by the frozen LLM capturer. This implies that no text cues are utilized to enhance the audio guidance for better", + "bbox": [ + 212, + 573, + 787, + 633 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/0be921ca0e094cf0aa9ead2e0d306f75dcf676a3f972b67f753205c4bf274e88.jpg", + "table_caption": [ + "Table 5: Ablation study on AVS-V1M. Our method benefits from using both audio and text as the guiding information for sounding objects." + ], + "table_footnote": [], + "table_body": "
ControlmIoU(%)F-score
TeSO66.020.801
(-) LinfoNCE65.340.793
(-) Dynamic Mask64.780.781
(-) PMQS62.900.774
(-) SeDAM62.220.760
(-) Text62.280.768
(-) Audio42.110.692
(-) Text & Audio37.350.656
", + "bbox": [ + 243, + 733, + 444, + 838 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/15822cf2b8f68ad588b581b5fe3fcf80ae8cf19d73a2780cb991ff503d55adaa.jpg", + "table_caption": [ + "Table 6: Comparison of our proposed baseline and TeSO. The baseline method employs LLM to filter the predicted results but does not enhance the audio guidance with text cues. Experiments are conducted on AVS-V1M test set." + ], + "table_footnote": [], + "table_body": "
Normal mIoU%Mute mIoU% (Δm%)WGN-40dB mIoU% (Δm%)
Baseline64.1543.37 (32.4)43.09 (32.8)
TeSO66.0237.89 (42.6)34.74 (47.4)
", + "bbox": [ + 506, + 760, + 774, + 810 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Sounding Object Segmentation Preference", + "bbox": [ + 449, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "control over the visual space. As expected, the sensitivity $(\\Delta_{m})$ of model performance to changes in audio input has decreased compared with TeSO, indicating the effectiveness of our method.", + "bbox": [ + 215, + 146, + 784, + 191 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Discussion", + "text_level": 1, + "bbox": [ + 215, + 220, + 354, + 236 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Noisy text cues. Besides dense image captions for collecting semantic text cues, we also constructed a dynamic mask in Eq. (4) to further select effective semantics based on audio correspondence. We use audio recognition results to determine the accuracy of text cues. Please refer to Tab. 3 to compare the audio recognition results based on text cues collection stage $(\\mathrm{REC}_{\\#1})$ and dynamic mask filtering stage $(\\mathrm{REC}_{\\#2})$ . Our approach achieves higher audio recognition performance, compared with using the audio foundation model in BAVS. This indicates the reliability and accuracy of our strategy for obtaining text cues.", + "bbox": [ + 215, + 257, + 785, + 378 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Prompt with CoT. As shown in the Fig. 3, we utilize task-specific prompt templates with multiple scenario demonstrations and CoT instructions to assist the frozen LLM in collecting more accurate potential sounding objects as text cues. For the detailed prompt templates with CoT instructions and related experimental information, please refer to the Prompt Template part and Text Assistance part in the supplementary materials.", + "bbox": [ + 215, + 380, + 785, + 469 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Versatility. We explore an alternative visual foundation model like the Segment Anything Model [19] for segmentation. Conducting experiments using a ViT-based backbone with minimal adjustments, our model achieves an impressive $65.36\\%$ mIoU on V1M, still getting comparable performance. This demonstrates the versatility and effectiveness of our model. Additional experiment details and analysis can be found in the supplementary materials.", + "bbox": [ + 215, + 470, + 785, + 561 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 590, + 357, + 606 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This study represents a pioneering effort to incorporate text cues into the AVS task. Unlike previous approaches that solely rely on audio assistance, our TeSO method establishes a stronger audio-visual correlation by harnessing the robust semantics unique to the text modality By integrating CoT instructions and common-sense reasoning, we enhance the ability to capture potential sounding objects from detailed scene descriptions, thereby optimizing the utilization of text cues. Furthermore, we have developed a dynamic mask based on attention scores to effectively handle noisy information present in the text cues. Our comprehensive experiments highlight the effectiveness of our approach, demonstrating comparative performance across all AVS Benchmarks. In particular, more observation experiments validate that enhancing audio with text cues grants superior audio guidance. In summary, our work provides a new direction for building reliable audio-visual correlation with textual semantics, and we hope that it can help the community construct more robust audio-visual systems.", + "bbox": [ + 215, + 628, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research was supported by National Natural Science Foundation of China (NO.62106272), and Public Computing Cloud, Renmin University of China.", + "bbox": [ + 215, + 178, + 784, + 209 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 233, + 321, + 250 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abu-El-Haija, S., Kothari, N., Lee, J., Natsev, P., Toderici, G., Varadarajan, B., Vijayanarasimhan, S.: Youtube-8m: A large-scale video classification benchmark. arXiv preprint arXiv:1609.08675 (2016)", + "2. Bao, F., Nie, S., Xue, K., Li, C., Pu, S., Wang, Y., Yue, G., Cao, Y., Su, H., Zhu, J.: One transformer fits all distributions in multi-modal diffusion at scale. arXiv preprint arXiv:2303.06555 (2023)", + "3. Chen, S., Wu, Y., Wang, C., Liu, S., Tompkins, D., Chen, Z., Wei, F.: Beats: Audio pre-training with acoustic tokenizers. arXiv preprint arXiv:2212.09058 (2022)", + "4. Cheng, B., Choudhuri, A., Misra, I., Kirillov, A., Girdhar, R., Schwing, A.G.: Mask2former for video instance segmentation (2021)", + "5. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290-1299 (2022)", + "6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "7. Espejel, J.L., Ettifouri, E.H., Alassan, M.S.Y., Chouham, E.M., Dadhane, W.: Gpt-3.5, gpt-4, or bard? evaluating llms reasoning ability in zero-shot setting and performance boosting through prompts. Natural Language Processing Journal 5, 100032 (2023)", + "8. Gao, S., Chen, Z., Chen, G., Wang, W., Lu, T.: Avsegformer: Audio-visual segmentation with transformer (2023)", + "9. Gemmeke, J.F., Ellis, D.P., Freedman, D., Jansen, A., Lawrence, W., Moore, R.C., Plakal, M., Ritter, M.: Audio set: An ontology and human-labeled dataset for audio events. In: 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP). pp. 776-780. IEEE (2017)", + "0. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15180-15190 (2023)", + "1. Hao, D., Mao, Y., He, B., Han, X., Dai, Y., Zhong, Y.: Improving audio-visual segmentation with bidirectional generation. arXiv preprint arXiv:2308.08288 (2023)", + "2. Van der Heiden, R.M., Janssen, C.P., Donker, S.F., Hardeman, L.E., Mans, K., Kenemans, J.L.: Susceptibility to audio signals during autonomous driving. PloS one 13(8), e0201963 (2018)", + "3. Hershey, S., Chaudhuri, S., Ellis, D.P., Gemmeke, J.F., Jansen, A., Moore, R.C., Plakal, M., Platt, D., Saurous, R.A., Seybold, B., et al.: Cnn architectures for large-scale audio classification. In: 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP). pp. 131-135. IEEE (2017)" + ], + "bbox": [ + 225, + 266, + 785, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Sounding Object Segmentation Preference", + "bbox": [ + 447, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019)", + "15. Hu, D., Wei, Y., Qian, R., Lin, W., Song, R., Wen, J.R.: Class-aware sounding objects localization via audiovisual correspondence. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9844-9859 (2021)", + "16. Huang, S., Li, H., Wang, Y., Zhu, H., Dai, J., Han, J., Rong, W., Liu, S.: Discovering sounding objects by audio queries for audio visual segmentation (2023)", + "17. Hur, C., Park, H.: Zero-shot image classification with rectified embedding vectors using a caption generator. Applied Sciences 13(12), 7071 (2023)", + "18. Johnson, J., Karpathy, A., Fei-Fei, L.: Densecap: Fully convolutional localization networks for dense captioning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4565-4574 (2016)", + "19. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)", + "20. Li, K., Yang, Z., Chen, L., Yang, Y., Xun, J.: Catr: Combinatorial-dependence audio-queried transformer for audio-visual video segmentation. arXiv preprint arXiv:2309.09709 (2023)", + "21. Ling, Y., Li, Y., Gan, Z., Zhang, J., Chi, M., Wang, Y.: Hear to segment: Unmixing the audio to guide the semantic segmentation (2023)", + "22. Liu, C., Li, P., Qi, X., Zhang, H., Li, L., Wang, D., Yu, X.: Audio-visual segmentation by exploring cross-modal mutual semantics (2023)", + "23. Liu, C., Li, P., Zhang, H., Li, L., Huang, Z., Wang, D., Yu, X.: Bavs: Bootstrapping audio-visual segmentation by integrating foundation knowledge. arXiv preprint arXiv:2308.10175 (2023)", + "24. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023)", + "25. Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: Llava-last: Improved reasoning,OCR,and world knowledge (January 2024),https://llava-v1.github.io/blog/2024-01-30-llava-last/", + "26. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023)", + "27. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)", + "28. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024)", + "29. Liu, J., Ju, C., Ma, C., Wang, Y., Wang, Y., Zhang, Y.: Audio-aware query-enhanced transformer for audio-visual segmentation (2023)", + "30. Liu, J., Wang, Y., Ju, C., Zhang, Y., Xie, W.: Annotation-free audio-visual segmentation. arXiv preprint arXiv:2305.11019 (2023)", + "31. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)", + "32. Ma, J., Sun, P., Wang, Y., Hu, D.: Stepping stones: A progressive training strategy for audio-visual semantic segmentation. IEEE European Conference on Computer Vision (ECCV) (2024)", + "33. Majumder, S., Al-Halah, Z., Grauman, K.: Move2hear: Active audio-visual source separation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 275–285 (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Mo, S., Tian, Y.: Av-sam: Segment anything model meets audio-visual localization and segmentation. arXiv preprint arXiv:2305.01836 (2023)", + "35. Park, S., Senocak, A., Chung, J.S.: Marginnce: Robust sound localization with a negative margin. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)", + "36. Senocak, A., Oh, T.H., Kim, J., Yang, M.H., Kweon, I.S.: Learning to localize sound source in visual scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4358-4366 (2018)", + "37. Sharma, H., Jalal, A.S.: Image captioning improved visual question answering. Multimedia tools and applications 81(24), 34775-34796 (2022)", + "38. Shi, Z., Zhou, X., Qiu, X., Zhu, X.: Improving image captioning with better use of captions. arXiv preprint arXiv:2006.11807 (2020)", + "39. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D., Blecher, L., Ferrer, C.C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I., Korenev, A., Koura, P.S., Lachaux, M.A., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A., Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X.E., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models (2023)", + "40. Wang, Y., Liu, W., Li, G., Ding, J., Hu, D., Li, X.: Prompting segmentation with sound is generalizable audio-visual source localizer. arXiv preprint arXiv:2309.07929 (2023)", + "41. Wang, Y., Sun, P., Zhou, D., Li, G., Zhang, H., Hu, D.: Ref-avs: Refer and segment objects in audio-visual scenes. IEEE European Conference on Computer Vision (ECCV) (2024)", + "42. Wu, W., Yao, H., Zhang, M., Song, Y., Ouyang, W., Wang, J.: Gpt4vis: What can gpt-4 do for zero-shot visual recognition? arXiv preprint arXiv:2311.15732 (2023)", + "43. Yan, S., Zhang, R., Guo, Z., Chen, W., Zhang, W., Li, H., Qiao, Y., He, Z., Gao, P.: Referred by multi-modality: A unified temporal transformer for video object segmentation. arXiv preprint arXiv:2305.16318 (2023)", + "44. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of Imms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9(1), 1 (2023)", + "45. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding (2023)", + "46. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127(3), 302-321 (2019)", + "47. Zhou, J., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., Zhong, Y.: Audio-visual segmentation. In: European Conference on Computer Vision (2022)", + "48. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)", + "49. Zürn, J., Burgard, W.: Self-supervised moving vehicle detection from audio-visual cues (2022)" + ], + "bbox": [ + 212, + 146, + 785, + 830 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Sounding Object Segmentation Preference", + "bbox": [ + 447, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Can Textual Semantics Mitigate Sounding Object Segmentation Preference_/88d8a449-254f-4c3c-bdf6-605afa985fb1_model.json b/2024/Can Textual Semantics Mitigate Sounding Object Segmentation Preference_/88d8a449-254f-4c3c-bdf6-605afa985fb1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..642d8b7a19fa2d0f10ee7bd0c1f057776e3add91 --- /dev/null +++ b/2024/Can Textual Semantics Mitigate Sounding Object Segmentation Preference_/88d8a449-254f-4c3c-bdf6-605afa985fb1_model.json @@ -0,0 +1,2423 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.141, + 0.787, + 0.187 + ], + "angle": 0, + "content": "Can Textual Semantics Mitigate Sounding Object Segmentation Preference?" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.215, + 0.782, + 0.247 + ], + "angle": 0, + "content": "Yaoting Wang\\(^{1\\dagger}\\), Peiwen Sun\\(^{2\\dagger}\\), Yuanchao Li\\(^{3}\\), Honggang Zhang\\(^{2}\\), and Di Hu\\(^{\\text{空}1,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.258, + 0.761, + 0.273 + ], + "angle": 0, + "content": "1 Gaoling School of Artificial Intelligence, Renmin University of China, China" + }, + { + "type": "text", + "bbox": [ + 0.407, + 0.275, + 0.597, + 0.288 + ], + "angle": 0, + "content": "yaoting.wang@outlook.com" + }, + { + "type": "text", + "bbox": [ + 0.442, + 0.289, + 0.562, + 0.3 + ], + "angle": 0, + "content": "dihu@ruc.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.301, + 0.733, + 0.315 + ], + "angle": 0, + "content": "\\(^{2}\\) Beijing University of Posts and Telecommunications, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.394, + 0.316, + 0.61, + 0.329 + ], + "angle": 0, + "content": "{sunpeiwen,zhhg}@bupt.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.365, + 0.329, + 0.637, + 0.342 + ], + "angle": 0, + "content": "3 University of Edinburgh, Scotland, UK" + }, + { + "type": "text", + "bbox": [ + 0.423, + 0.344, + 0.582, + 0.356 + ], + "angle": 0, + "content": "yuanchao.li@ed.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.234, + 0.356, + 0.768, + 0.37 + ], + "angle": 0, + "content": "4 Engineering Research Center of Next-Generation Search and Recommendation" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.439, + 0.744, + 0.745 + ], + "angle": 0, + "content": "Abstract. The Audio-Visual Segmentation (AVS) task aims to segment sounding objects in the visual space using audio cues. However, in this work, it is recognized that previous AVS methods show a heavy reliance on detrimental segmentation preferences related to audible objects, rather than precise audio guidance. We argue that the primary reason is that audio lacks robust semantics compared to vision, especially in multi-source sounding scenes, resulting in weak audio guidance over the visual space. Motivated by the fact that text modality is well explored and contains rich abstract semantics, we propose leveraging text cues from the visual scene to enhance audio guidance with the semantics inherent in text. Our approach begins by obtaining scene descriptions through an off-the-shelf image captioner and prompting a frozen large language model to deduce potential sounding objects as text cues. Subsequently, we introduce a novel semantics-driven audio modeling module with a dynamic mask to integrate audio features with text cues, leading to representative sounding object features. These features not only encompass audio cues but also possess vivid semantics, providing clearer guidance in the visual space. Experimental results on AVS benchmarks validate that our method exhibits enhanced sensitivity to audio when aided by text cues, achieving highly competitive performance on all three subsets. Project page: https://github.com/GeWu-Lab/Sounding-Object-Segmentation-Preference" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.761, + 0.741, + 0.79 + ], + "angle": 0, + "content": "Keywords: Audio-Visual Segmentation \\(\\cdot\\) Textual Guidance \\(\\cdot\\) Multi-modal learning" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.811, + 0.378, + 0.825 + ], + "angle": 0, + "content": "\\(\\dagger\\) Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.825, + 0.399, + 0.84 + ], + "angle": 0, + "content": "\\(\\mathbf{\\Pi}^{\\text{三}}\\) Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.145, + 0.756, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.496 + ], + "angle": 0, + "content": "Fig. 1: The previous methods (left) achieve satisfactory results in receiving normal audio input. However, even when the sound is completely silent, they still segment vast pixels to represent the guitar regarding the segmentation preference of audible objects built during the training. In contrast, our approach (right) utilizes a frozen LLM to reason from scene descriptions that the male is not playing the guitar, as his hands are off the strings. Guided by this semantic information, our method produces more precise segmentation with finer audio guidance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.527, + 0.376, + 0.543 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.718 + ], + "angle": 0, + "content": "In human perception, visual and auditory senses are closely related, and audio can provide vision with additional and dynamic scene information to enhance visual understanding. In the context of audio-visual understanding tasks, Audio-Visual Localization (AVL) enables the localization of sounding objects within visual scenes using audio references [15, 36]. However, the coarse localization of sounding objects is no longer sufficient to meet the practical demands of complex scene understanding, such as in autonomous driving [12, 49] and augmented reality [33]. To address this challenge, tasks are gradually shifting from bounding-box or rough heat-map localization to finer-grained pixel-level segmented localization, which is known as Audio-Visual Segmentation (AVS)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The existing works on AVS can be broadly categorized into fusion-based methods [8, 16, 20, 21, 29, 47] and prompt-based methods [30, 34, 40]. The former primarily focuses on localizing sounding objects by fusing audio and visual features, while the latter emphasizes constructing effective audio prompts for the visual foundation model. However, most of these methods fail to establish compact audio-visual correlations to achieve clear audio guidance in the visual space. As depicted in the left of Fig. 1, previous methods could achieve satisfactory results with normal audio input. Nevertheless, when the audio is silent," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Sounding Object Segmentation Preference" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "these methods still segment many pixels to represent potential sounding objects based on the segmentation preferences of audible objects1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.359 + ], + "angle": 0, + "content": "Furthermore, from the experiment with manually muted audio input, we observe that most of the behavior of AVS baselines depends on the segmentation preference established during the training process rather than relying on reliable audio-visual correlation. For instance, these models tend to take a shortcut by segmenting out the muted guitars because the audio guidance is weak, and they simply learned with the easy-to-learn visual feature alone (i.e., shortcut learning) that guitars are often associated with sound-emitting during the training phase. The unreliable audio-visual correlation can be attributed to two main factors. Firstly, the scarcity of training data for AVS is a significant challenge due to the demanding nature of pixel-level annotation. Secondly, the audio modality itself presents inherent complexity and ambiguity [16, 21, 29], especially in scenarios involving multiple audio sources that may be intertwined." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.361, + 0.788, + 0.692 + ], + "angle": 0, + "content": "Given the challenges discussed above, our approach aims to build a robust audio-visual correlation by integrating the text modality, which inherently provides meaningful and robust semantic information about objects and their interactions. Firstly, in the Scene Describing phase, we utilize an off-the-shelf image captioner (LLaVA-1.5 [28]) to generate detailed dense captions for scene understanding. Then, in the Sounding Objects Reasoning phase, we instruct the frozen large language models (LLMs, e.g., LLaMA2 [39]) as a text cues capturer to collect potential sounding objects from the scene description as text cues. Notably, we introduce an elaborate few-shot prompt template to guide the text cues capturer to reason with Chain-of-Thought (CoT) instructions. After that, in the Semantics-Driven Audio Modeling (SeDAM) module, we project the audio feature into latent audio features. However, instead of directly using these latent features to interact with visual features for mask decoding, we leverage these latent features to interact with text cues. Specifically, the latent features are attended to and collected to form sounding object features with semantics provided by the text cues through the crossmodal transformer. We further propose the Prompting Mask Queries with Semantics (PMQS) module to introduce the sounding object features into pre-trained mask queries for later segmentation. Finally, we use simple but effective bottleneck adapters in the Audio-Prompted Decoding phase for better segmentation quality. The experimental results demonstrate that our proposed method improves audio-visual grounding by incorporating guidance from both audio and text cues." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.693, + 0.788, + 0.754 + ], + "angle": 0, + "content": "Experiments in Sec. 4.4 reveal the advantage of our method, as verified by its increased sensitivity to the changes in audio input compared to networks without enhancing audio guidance with text cues. In summary, our contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.77, + 0.788, + 0.801 + ], + "angle": 0, + "content": "- We recognize the behavior of AVS model is influenced by the preference segmentation phenomenon, which can not be solved easily by previous audio-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.788, + 0.842 + ], + "angle": 0, + "content": "1 To be clear, we define audible objects as objects capable of producing sound, while sounding objects are defined as objects that are currently producing sound." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.147, + 0.785, + 0.177 + ], + "angle": 0, + "content": "visual based methods. A novel strategy introduces textual semantics as the bridge into the AVS task, establishing a better audio-visual correlation." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.178, + 0.784, + 0.237 + ], + "angle": 0, + "content": "- We introduce the SeDAM module to form sounding object features with audio and text cues. We also suggest a PMQS module to prompt the pretrained mask queries with these sounding object features for sounding object segmentation using a visual foundation model." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.239, + 0.784, + 0.268 + ], + "angle": 0, + "content": "- Our experiments show that our method can be more sensitive to the audio guidance, manifested in its high sensitivity to the changes in audio input." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.293, + 0.396, + 0.309 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.326, + 0.629, + 0.342 + ], + "angle": 0, + "content": "2.1 Audio-Visual Localization and Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.352, + 0.788, + 0.625 + ], + "angle": 0, + "content": "Traditional AVL tasks [15,35,36] predict the positions of sounding objects using bounding-box or coarse heat-maps through unsupervised learning of audio-visual correlation. In recent years, driven by the growing demand for more precise localization in industries such as autonomous driving [12,49] and augmented reality [33], the AVL task has evolved towards a finer-grained AVS task that localizes sounding objects at a pixel-level. The existing AVS works can be broadly categorized into fusion-based [8, 16, 20, 21, 29, 43, 47] and prompt-based [30, 32, 40] The pioneering AVS work [47] is fusion-based and adopts a multi-stage strategy to integrate audio with multi-scale visual features. Liu et al. [29] address the issues of inadequate fusion of audio-visual features with an audio-aware query-enhanced transformer (AuTR). In comparison, AVSegFormer [8] directly decodes the fused feature with the audio query. Wang et al. [40] innovatively prompt the visual foundation model with audio input, harnessing the abundant visual before achieving generalizable AVS in zero-shot and few-shot scenarios. However, these methods have not effectively established a reliable audio-visual correlation, preventing the models from perceiving robust audio guidance. On the contrary, our experiments in Sec. 4.4 also reveal that some AVS baselines' competitive performance may stem from segmentation preferences formed during the training." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.648, + 0.536, + 0.663 + ], + "angle": 0, + "content": "2.2 Text Aided Scene Understanding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.764 + ], + "angle": 0, + "content": "Before the explosion of LLM, textual semantic information had been widely used for visual understanding tasks. Zhan et al. [38] and Sharma et al. [37] improve Visual Question Answering by using the extra semantic information of image captions, while Hur and Park [17] utilize the image captioner to help zero-shot image classification. Wang et al. [41] incorporating text with audio and visual cues to solve reference AVS task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In recent years, the community has witnessed an emergent interest in strong off-the-shelf capabilities in open-world visual understanding [24-26, 45]. Notable examples, such as GPT4(V) [44] and LLaVA [27] have showcased remarkable linguistic and visual capabilities in zero-shot scene understanding [7, 42] without any training requirement. Innovatively, in the field of AVS, BAVS [23] replaces" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Sounding Object Segmentation Preference" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.144, + 0.768, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.341, + 0.788, + 0.455 + ], + "angle": 0, + "content": "Fig. 2: Overall pipeline of the proposed TeSO method. We utilize an off-the-shelf image captioner for dense scene describing and employ a frozen LLM to reason out potential sounding objects as text cues. These semantic text cues are then aggregated with audio features in the SeDAM module to form the sounding object features. Subsequently, we introduce the sounding object features into pre-trained mask queries in the PMQS module. Finally, we use adapters to tune the visual-only mask decoder for AVS in the Audio-prompted Decoding phase. \"MSDA\" is the multi-scale deformable attention proposed by Zhu et al. [48]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.788, + 0.604 + ], + "angle": 0, + "content": "audio input with textual audio tags from a large pre-trained vision (UniDiffuser [2]) and audio (BEATs [3]) foundation model, establishing a reliable AVS system with foundation knowledge. However, due to the significant gap in class distribution between the pre-training dataset and the AVS dataset, as well as the presence of multi-source scenarios in the AVS task, the extracted audio tags are not ideal to provide accurate semantics. In contrast, we utilize the text modality with robust semantic features to enhance rather than replace the audio guidance and address the segmentation preference issue." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.627, + 0.676, + 0.646 + ], + "angle": 0, + "content": "3 Text-guided Sounding Object Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As shown in Fig. 2, we introduce our Text-guided Sounding Object Segmentation (TeSO) method in this section. Our experiments about muted audio input imply that audio does not exert reasonable guidance over the visual space for some AVS baselines, resulting in the models relying more on the segmentation preferences established during the training process. In essence, they learn which objects are most likely audible during training and take shortcuts during inference. To address the above issue, our method aims to enhance the audio-visual correlation by leveraging the text modality, which inherently possesses robust semantic information, to obtain finer-grained audio guidance. We begin by acquiring detailed scene descriptions through an off-the-shelf image captioner. Subsequently, a frozen LLM works as the text cues capturer, collecting potential sounding objects as text cues from the scene descriptions with CoT instructions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.194 + ], + "angle": 0, + "content": "Finally, we introduce a novel SeDAM module with a dynamic mask to seamlessly integrate audio features and text semantics through a crossmodal transformer, providing finer audio guidance with text semantic cues." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.214, + 0.492, + 0.23 + ], + "angle": 0, + "content": "3.1 Multimodal Representation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.787, + 0.313 + ], + "angle": 0, + "content": "Visual: Following previous works [40, 47], we sample frames from the video at 1-second intervals. Mask2Former [4] serves as our visual foundation model, and we extract visual features \\( F_V \\in \\mathbb{R}^{d_V \\times H \\times W} \\) from a pre-trained Swin transformer [6]. We use simple two-layer MLP adapters to tune the visual encoder for better visual representations." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.314, + 0.787, + 0.404 + ], + "angle": 0, + "content": "Audio: Similar to the video processing, we split the audio into clips at 1-second intervals. We extract audio features \\( F_{As} = \\{F_{A_1}, F_{A_2}, \\dots, F_{A_T}\\} \\in \\mathbb{R}^{T \\times d_A} \\) using VGGish [9, 13], where \\( T \\) represents the length of the audio in seconds, matching the number of video frames. Therefore, each video frame corresponds to \\( F_{A} = F_{As}[i] \\). We freeze the pre-trained parameters for a better comparison with previous works." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.405, + 0.787, + 0.451 + ], + "angle": 0, + "content": "Text: For each video frame, we complement the captured text cues into a sentence using a template (\"This is a \\(\\{\\_ \\}\\).\") to extract text features \\(F_{T} \\in \\mathbb{R}^{N_{T} \\times d_{T}}\\) with ImageBind [10], where \\(N_{T}\\) is the number of text cues." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.471, + 0.446, + 0.485 + ], + "angle": 0, + "content": "3.2 Text Cues Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.787, + 0.556 + ], + "angle": 0, + "content": "To obtain more precise text cues, we use a two-stage inference flow. Firstly, we employ an off-the-shelf image captioner to generate dense scene descriptions. Secondly, we ask a text cues capturer to collect potential sounding objects from the generated descriptions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.577, + 0.788, + 0.7 + ], + "angle": 0, + "content": "Scene Describing. As shown in the top-left of Fig. 3, in this phase, we employ the commonly used and straightforward prompt: \"Please carefully understand the image content and provide as rich a description as possible.\" This instructs the image captioner to generate a detailed description of the visual scene, a task commonly referred to as dense captioning [18]. This process enables the image captioner to produce comprehensive and informative descriptions that not only describe the objects like \"a human and guitar in the image\" but also capture their interactions and relationships." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Sounding Object Collecting. To extract potential sounding objects from the generated dense descriptions, we design a task-specific CoT prompt template with few-shot demonstrations, as illustrated in Fig. 3. Specifically, we carefully craft examples that cover various scenarios and provide reasoning approaches for each example. The process begins by identifying the objects present in the dense scene descriptions, followed by step-by-step common-sense reasoning according to the CoT instructions, as depicted in the lower right of Fig. 3. During the CoT reasoning, the text cues capturer first categorizes the sample objects and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Sounding Object Segmentation Preference" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.144, + 0.455, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.143, + 0.685, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.361, + 0.788, + 0.418 + ], + "angle": 0, + "content": "Fig. 3: \"P.S.O\" stands for \"Potential Sounding Object\". A frozen LLM reasoner works as a text cues capturer by considering the interaction of audible objects. For instance, a person may sing a song while playing the guitar, but he would not sing along with a saxophone. In contrast, a noun parser simply captures any nouns that are present." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.57 + ], + "angle": 0, + "content": "assesses whether objects in that category are likely making sounds. Subsequently, the capturer analyzes which group of potential sounding objects can produce sounds simultaneously. Finally, the group with higher confidence is captured as the text cues of length \\( N_{T} \\) and encoded in \\( F_{T} \\) for further guidance. The right part of Fig. 3 also illustrates why we use the frozen LLM as a text cues capturer instead of using a naive noun parser. It is important to note that the use of LLM is solely for generating high-quality text cues, the LLM itself is not engaged in the training process for a fairness concern." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.596, + 0.54, + 0.612 + ], + "angle": 0, + "content": "3.3 Semantic-Driven Audio Modeling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.788, + 0.761 + ], + "angle": 0, + "content": "After obtaining potential sounding objects, we can enhance the semantics of audio features with these text cues in this SeDAM module. To allow semantic text cues to capture different aspects of audio features \\( F_{A} \\), we project \\( F_{A} \\) into distinct latent features \\( F_{A_{L}} = \\{a_{i}\\}_{i=1}^{N_{L}} \\in \\mathbb{R}^{N_{L} \\times d_{A_{L}}} \\), where \\( N_{L} \\) is the number of latent features. It is worth noting that, we incorporate an auxiliary loss term \\( \\mathcal{L}_{infoNCE} \\) aims at ensuring the distinction among latent audio features. This way, we project audio features into the latent space, where each unique \\( a_{i} \\) represents a latent audio feature to be captured by semantic text cues, forming clear and loud sounding object features with robust semantics for finer guidance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.762, + 0.788, + 0.807 + ], + "angle": 0, + "content": "To incorporate the semantics of text cues, we employ text cues as the query to attend and combine the relevant latent features through crossmodal multi-head attention, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.375, + 0.825, + 0.787, + 0.844 + ], + "angle": 0, + "content": "\\[\nF _ {\\widetilde {A} _ {L}} = \\operatorname {M u l t i H e a d} \\left(F _ {T}, F _ {A _ {L}}, F _ {A _ {L}}\\right), \\tag {1}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.178 + ], + "angle": 0, + "content": "where \\(F_{\\widetilde{A}_L} \\in \\mathbb{R}^{N_T \\times d_A}\\) is the attended latent audio features and will be combined with text cues \\(F_T\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.193, + 0.786, + 0.213 + ], + "angle": 0, + "content": "\\[\nF _ {A _ {T}} = F C \\left(L N \\left(F _ {T} + F _ {\\widetilde {A} _ {L}}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.219, + 0.785, + 0.339 + ], + "angle": 0, + "content": "where \\(F_{A_T} \\in \\mathbb{R}^{N_T \\times d_A}\\) is the combined feature with both audio information and semantic text cues, and \\(FC\\) and \\(LN\\) for linear layer and layer normalization. Note that the LLM reasoner may generate a different number of text cues, and we either pad with a zero matrix or truncate to \\(N_T\\). It is noteworthy that we propose task-specific CoT instructions to prompt the text cues capturer to infer more accurate potential sounding objects as the text cues. Additionally, we also introduce a dynamic mask based on attention score to mitigate the potential noisy information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.784, + 0.372 + ], + "angle": 0, + "content": "We first compute the average pooling attention scores \\(\\zeta \\in \\mathbb{R}^{N_T}\\) for text cues on the attention head and key dimension:" + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.383, + 0.786, + 0.425 + ], + "angle": 0, + "content": "\\[\n\\xi = \\operatorname {P o o l} \\left(\\frac {F _ {T} F _ {A _ {L}} ^ {\\mathrm {T}}}{\\sqrt {d _ {A _ {L}}}}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.437, + 0.784, + 0.469 + ], + "angle": 0, + "content": "where \\(\\sqrt{d_{A_L}}\\) is the scaling factor for dot production. Then we form the attention mask \\(\\mathcal{M} = \\{m_1, m_2, m_3, \\dots, m_{N_T}\\}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.481, + 0.786, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {i} = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} \\xi_ {i} > 0 \\\\ - \\infty & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.532, + 0.784, + 0.579 + ], + "angle": 0, + "content": "where \\(\\mathcal{M} \\in \\mathbb{R}^{N_T}\\), will be used as a key padding mask for the PMQS module in the next section to ensure that the attention weight of redundant text cues approaches zero." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.602, + 0.6, + 0.618 + ], + "angle": 0, + "content": "3.4 Prompting Mask Queries with Semantics" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.629, + 0.784, + 0.72 + ], + "angle": 0, + "content": "Current segmentation foundation models [4, 5, 19] adopt transformer-based decoder with mask queries as the mainstream approach. To prompt such foundation models, we propose the PMQS module, aiming to introduce semantic information to mask query features \\( F_{Q} = \\{q_{1}, q_{2}, \\dots, q_{N_{Q}}\\} \\), where \\( F_{Q} \\in \\mathbb{R}^{N_{Q} \\times d_{V}} \\), and \\( N_{Q} \\) is the number of mask queries in crossmodal attention. We formulate this updated expression as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.311, + 0.732, + 0.786, + 0.773 + ], + "angle": 0, + "content": "\\[\n\\phi_ {c a} \\left(F _ {Q}, F _ {A _ {T}}, F _ {A _ {T}}\\right) = S o f t m a x \\left(\\frac {F _ {Q} F _ {A _ {T}} ^ {\\mathrm {T}}}{\\sqrt {d _ {A _ {T}}}} + \\mathcal {M}\\right) F _ {A _ {T}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.779, + 0.784, + 0.81 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}\\) is added to the attention score to mask the redundant text cues for each query. Then we obtain the updated mask query features \\(\\widetilde{F}_Q\\in \\mathbb{R}^{N_Q\\times d_V}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.391, + 0.824, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\widetilde {F} _ {Q} = F _ {Q} + \\phi_ {c a} \\left(F _ {Q}, F _ {A _ {T}}, F _ {A _ {T}}\\right). \\tag {6}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.449, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Sounding Object Segmentation Preference" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.484, + 0.163 + ], + "angle": 0, + "content": "3.5 Audio-Prompted Decoding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.171, + 0.784, + 0.248 + ], + "angle": 0, + "content": "As shown on the right side of Fig. 2, the updated mask query features \\(\\widetilde{F}_Q\\) are forwarded into the transformer-based decoder, generating mask queries \\(M_Q \\in \\mathbb{R}^{N_Q \\times H \\times W}\\) and class queries \\(C_Q \\in \\mathbb{R}^{N_Q \\times N_C}\\), where \\(N_C\\) is the number of class queries pre-defined in the segmentation foundation model. Subsequently, we can obtain the final segmentation \\(M_{pred} \\in \\mathbb{R}^{N_C \\times H \\times W}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.261, + 0.786, + 0.279 + ], + "angle": 0, + "content": "\\[\nM _ {p r e d} = \\operatorname {E i n s u m} \\left(C _ {Q}, \\bar {F} _ {Q}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.289, + 0.63, + 0.304 + ], + "angle": 0, + "content": "where \\( \\text{Einsum}(\\cdot) \\) is the Einstein Summation Convention." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.305, + 0.784, + 0.366 + ], + "angle": 0, + "content": "Notably, to enhance the quality of the segmentation, we use two-layer bottle-neck MLP adapters [14] to lightly fine-tune the output projection of the following attentions: a) Swin Window Attention (W-MSA) [31], b) Deformable Attention in pixel decoder [48] and c) Masked Attention in transformer decoder [4]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.387, + 0.429, + 0.403 + ], + "angle": 0, + "content": "3.6 Learning Objectives" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.412, + 0.784, + 0.443 + ], + "angle": 0, + "content": "Segmentation Loss. During the model training process, we employ the binary cross-entropy loss and dice loss to optimize the mask quality:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.456, + 0.785, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {b c e} \\cdot \\mathcal {L} _ {b c e} + \\lambda_ {d i c e} \\cdot \\mathcal {L} _ {d i c e}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.482, + 0.784, + 0.51 + ], + "angle": 0, + "content": "Further, we adopt the mask classification loss [5] to compose the final segmentation loss:" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.514, + 0.785, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s e g} = \\mathcal {L} _ {m a s k} + \\lambda_ {c l s} \\cdot \\mathcal {L} _ {c l s}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.535, + 0.784, + 0.581 + ], + "angle": 0, + "content": "Latent Component Loss. This loss term aims to explicitly guide the model to learn distinct representations for each latent audio component, helping prevent redundancy and overlapping information across the latent components." + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.592, + 0.785, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {i n f o N C E}} = - \\log \\left(\\frac {\\exp \\left(\\operatorname {s i m} \\left(a , a ^ {+}\\right)\\right)}{\\sum_ {i = 1} ^ {N _ {L}} \\exp \\left(\\operatorname {s i m} \\left(a , a _ {i}\\right)\\right)}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.642, + 0.756, + 0.657 + ], + "angle": 0, + "content": "where \\(a^+\\) and \\(sim(\\cdot)\\) denote the positive example and the cosine similarity." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.658, + 0.784, + 0.686 + ], + "angle": 0, + "content": "Total Loss The final loss function is composed of the weighted sum of the two kinds of losses above:" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.701, + 0.785, + 0.716 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {s e g}} + \\lambda_ {\\text {i n f o N C E}} \\cdot \\mathcal {L} _ {\\text {i n f o N C E}}. \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.738, + 0.376, + 0.755 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.77, + 0.457, + 0.785 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.841 + ], + "angle": 0, + "content": "Dataset. Our proposed method is evaluated on the AVS Benchmarks [47], which contains three subsets. Firstly, the single-source subset (V1S) contains 4932 videos over 23 categories, covering various sounds, such as humans, animals," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.252 + ], + "angle": 0, + "content": "vehicles, and musical instruments. For videos in the training split of this subset, only the first sampled frame is annotated. Secondly, the multi-source subset (V1M) contains 424 videos that include two or more categories from V1S, and all sounding objects are visible in the frames. Finally, the AVSBench-semantic (AVSS) subset, an extension of V1S and V1M, contains 12,356 videos. These newly collected videos are trimmed to 10 seconds, differing from the 5 seconds in V1S and V1M." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.358 + ], + "angle": 0, + "content": "Setting. We conduct training and evaluation using the VGGish backbone pretrained on Youtube-8M [1] and Swin-base Transformer backbone pretrained on semantic-ADE20K [46] by Mask2Former [4]. The number \\(N\\) of the crossmodal transformer layers is set to 4, and the parameters \\(\\lambda_{bce},\\lambda_{dice},\\lambda_{cls},\\lambda_{infoNCE}\\) in the loss are set to 5,5,2,1. The AdamW optimizer is adopted with a learning rate of 1e-4 for the visual encoder adapters and 1e-3 for other learnable parameters, and the training epoch is roughly set to 60." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.406 + ], + "angle": 0, + "content": "Metrics. To conduct a comprehensive evaluation of our model, we carry out tests using mean Intersection over Union (mIoU) and F-score as the performance metrics, following previous works [8, 47]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.425, + 0.43, + 0.44 + ], + "angle": 0, + "content": "4.2 Comparison Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.569 + ], + "angle": 0, + "content": "When compared to methods that do not incorporate text as additional semantic information, our method, which leverages captioning and reasoning, demonstrates strong competitiveness. As illustrated in Sec. 4.2, our method achieves comparable results across all V1S, V1M, and AVSS-binary subsets, showcasing average performance improvements of up to \\(1.25\\%\\) in mIoU and \\(2.7\\%\\) in F-score, respectively. Furthermore, qualitative analysis indicates that our predicted masks exhibit superior quality, as demonstrated in sub-figure (1) of Fig. 4 (more examples in supplementary materials)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.788, + 0.659 + ], + "angle": 0, + "content": "It is necessary to mention that our proposed TeSO does not show significant improvements on V1S over certain existing methods, such as BAVS [23], which also utilizes text semantic information. Apart from marginal effects, this phenomenon can also be attributed to the singular sound source and simplistic nature of the data scenarios present in V1S. On V1M, however, our performance significantly surpasses that of other methods, including BAVS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.735 + ], + "angle": 0, + "content": "In addition to the AVS task, we also conduct experiments on the Audio-Visual Semantic Segmentation (AVSS) task, as shown in Tab. 2. Our model exceeds the best performance in previous models by \\(5.37\\%\\) on mIoU, which further demonstrates that our model is capable of establishing a better audio-visual correlation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.756, + 0.468, + 0.77 + ], + "angle": 0, + "content": "4.3 Bottlenecks and Fairness" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Our method (TeSO) collects text cues from dense image descriptions at \\(\\mathrm{REC}_{\\#1}\\) \\((80.41\\%)\\) and a filter with a dynamic mask at \\(\\mathrm{REC}_{\\#2}\\) \\((80.95\\%)\\), indicating a robust startup for the text-guided AVS framework, in comparison with another two-stage model, BAVS." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Sounding Object Segmentation Preference" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.224, + 0.144, + 0.775, + 0.19 + ], + "angle": 0, + "content": "Table 1: Performance on AVS-Benchmarks. The underscore is used to indicate suboptimal results. BAVS uses large pre-trained visual and audio foundation models to replace audio input with text tag. In comparison, TeSO (ours) enhances rather than replaces audio guidance based on the robust semantics of text. It is evident that our method has a good F-score result while ensuring mIoU." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.198, + 0.774, + 0.364 + ], + "angle": 0, + "content": "
MethodAudio-backboneVisual-backboneV1SV1MAVSS-binary
mIoU(%)F-scoremIoU(%)F-scoremIoU(%)F-score
AVSBench [47]VGGishPVT-v278.700.87954.000.64562.450.756
AVSegFormer [8]VGGishPVT-v282.060.89958.360.69364.340.759
AVSC [22]VGGishPVT-v281.290.88659.500.657--
AVS-BG [11]VGGishPVT-v281.710.90455.100.668--
AQFormer [16]VGGishPVT-v281.600.89461.100.721--
†CATR [20]VGGishPVT-v281.400.89659.000.700--
AV-SAM [34]ResNet18ViT-Base40.470.566----
SAMA [30]VGGishViT-Huge81.530.88663.140.691--
GAVS [40]VGGishViT-Base80.060.90263.700.77467.700.788
AuTR [29]VGGishSwin-Base80.400.89156.200.672--
MUTR [43]VGGishVideo-Swin-Base81.600.89764.000.735--
BAVS [23]‡BEATsSwin-Base82.680.89859.630.65955.450.640
TeSO (ours)VGGishSwin-base83.270.93366.020.80168.530.813
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.235, + 0.366, + 0.772, + 0.388 + ], + "angle": 0, + "content": "\\(\\dagger\\) : To make a fair comparison, the results of CATR here are without supplemented annotation of the training set." + }, + { + "type": "table_footnote", + "bbox": [ + 0.235, + 0.388, + 0.771, + 0.41 + ], + "angle": 0, + "content": ": BEATs is a stronger audio backbone, trained on AudioSet dataset with 90M parameters. In comparison, VGGish is trained on Youtube-8M with 70M parameters." + }, + { + "type": "list", + "bbox": [ + 0.235, + 0.366, + 0.772, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.425, + 0.487, + 0.495 + ], + "angle": 0, + "content": "Table 2: Performance on AVSS dataset. Our method showcases a substantial improvement in the semantic subset, achieved by leveraging robust text semantics to enhance audio guidance." + }, + { + "type": "table", + "bbox": [ + 0.228, + 0.506, + 0.472, + 0.583 + ], + "angle": 0, + "content": "
MethodBackboneAVSS
mIoU(%)F-Score
AVSBenchPVT-v229.800.352
CATRPVT-v232.800.385
BAVSSwin-base33.590.375
TeSO (Ours)Swin-base38.960.451
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.425, + 0.788, + 0.494 + ], + "angle": 0, + "content": "Table 3: Potential sounding objects recognition (REC) performance at different model stages and the corresponding segmentation results. Experiments are conducted on AVS-V1M test set." + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.506, + 0.774, + 0.583 + ], + "angle": 0, + "content": "
MethodREC(#1)AP(%)REC(#2)AP(%)SEGmIoU(%)
BAVS24.2279.6959.63
TeSOp(ours)77.9678.3363.87
TeSOn(ours)77.8278.2763.31
TeSO(ours)80.4180.9566.02
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.787, + 0.764 + ], + "angle": 0, + "content": "In addition, we develop several variants of TeSO to investigate its performance sources and bottlenecks, as well as its fairness in comparison with other methods. \\(\\mathrm{TeSO}_p\\) denotes the variant using one-shot prompts, while \\(\\mathrm{TeSO}_n\\) denotes the variant employing a noun parser instead of a frozen LLM for text cues collection. As shown in Tab. 3, the results demonstrate that both the number (shots) of demonstrations and the method of text cues collection significantly impact the model's ability to accurately identify audio semantics and perform segmentation (SEG). Thus, utilizing few-shot prompts and off-the-shelf foundation models are both essential for ensuring the acquisition of high-quality text cues (i.e., the bottleneck)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.842 + ], + "angle": 0, + "content": "However, we would like to emphasize that, even without using LLM to capture text cues, our method variant \\((\\mathrm{TeSO}_n)\\) still fairly achieves competitive performance (63.31% mIoU) compared with BAVS (59.63% mIoU) and state-of-the-art method GAVS (63.70% mIoU) in the complex multi-source scenario (AVS-V1M), and significantly outperforms other methods. This suggests that" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 4: Impact on mIoU and F-score by muted audio or noise-only audio on AVS-V1M test set. TeSO (ours) exhibits the most intensive attenuation up to \\(48.8\\%\\). It is noteworthy that the other models still generate \"satisfactory\" results even in silent or purely noisy environments. This observation suggests that the models might rely more on the segmentation preferences formed during the training rather than truly learning efficient audio guidance." + }, + { + "type": "table", + "bbox": [ + 0.24, + 0.24, + 0.761, + 0.306 + ], + "angle": 0, + "content": "
MethodMuteWGN-10dBWGN-40dB
mIoU% (Δm%)F-score (Δf%)mIoU% (Δm%)F-score (Δf%)mIoU% (Δm%)F-score (Δf%)mIoU% (Δm%)
AVS-Bench49.15 (6.6)0.622 (4.9)46.76 (11.1)0.618 (5.5)47.46 (9.8)0.610 (6.7)
GAVS42.53 (29.2)0.695 (8.4)41.51 (34.8)0.692 (10.6)41.87 (30.3)0.696 (8.4)
BAVS45.06 (26.1)0.644 (5.8)43.27 (27.4)0.601 (8.8)42.93 (31.6)0.595 (9.7)
TeSO (ours)37.89 (42.6)0.671 (16.2)34.20 (49.6)0.661 (17.5)34.74 (47.4)0.656 (18.1)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.331, + 0.786, + 0.361 + ], + "angle": 0, + "content": "leveraging text-enhanced audio for visual control with robust semantics is a viable and promising direction." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.381, + 0.383, + 0.396 + ], + "angle": 0, + "content": "4.4 Audio Control" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.787, + 0.525 + ], + "angle": 0, + "content": "Based on our observations, we notice a distinct shape of the predicted mask in sub-figure (2) of Fig. 4 when some popular AVS baselines are used for inference. Interestingly, this shape remains distinguishable even when the inference is conducted with muted audio. This suggests that the network can predict rough masks without relying on audio input, indicating that the audio-visual grounding is not effectively guided by audio cues. As a result, this undesirable phenomenon can be attributed to the segmentation preference, which is a result of these models lacking proper audio control." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.788, + 0.753 + ], + "angle": 0, + "content": "To evaluate the influence of audio control on segmentation, various models are subjected to inference using muted audio and noise-only audio (at 10 dB and 40 dB) instead of the corresponding audio signals. It should be noted that the visual features remained unchanged throughout the evaluation. As shown in Tab. 4, the disparity \\(\\Delta_{m}\\) (and \\(\\Delta_{f}\\)) refers to the difference in mIoU (and F-score) values before and after the audio substitution is applied as an evaluation metric for the sensitivity of audio changes. A larger drop in mIoU and F-score indicates a greater degree of audio control over the visual space. In addition to demonstrating strong segmentation capabilities in the AVS task, TeSO (ours) exhibits the highest level of attenuation, reaching up to \\(47.4\\%\\), when subjected to muted and noise-only audio. Additionally, qualitative analysis revealed that the predicted masks remained predominantly blank, as depicted in sub-figure (2) of Fig. 4 (more examples in supplementary materials). This observation indicates that our model has better audio guidance over the visual space, enabling it to play a crucial role in audio-visual grounding." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.772, + 0.391, + 0.788 + ], + "angle": 0, + "content": "4.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "As shown in Tab. 5, we conduct ablation experiments targeting the designed TeSO modules and the guidance from different modalities. To explore the guiding ability, we progressively remove either the audio or text guidance for TeSO." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.45, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Sounding Object Segmentation Preference" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.145, + 0.728, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.375, + 0.788, + 0.432 + ], + "angle": 0, + "content": "Fig. 4: Examples of the impact of normal audio input and all-mute audio on popular methods. In normal scenarios, our method shows better masks than previous methods. In all-mute scenarios, our approach exhibits strong sensitivity towards audio inputs, as it is capable of generating blank masks for silent audio clips." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.788, + 0.572 + ], + "angle": 0, + "content": "The experiments indicate that control from both audio and text inputs for visual space can establish a strong audio-visual correlation. Furthermore, by utilizing our proposed TeSO, we achieve enhanced audio guidance and reach a \\(66.02\\%\\) mIoU while incorporating precise semantic information from the text modality. Moreover, through the utilization of our proposed TeSO method that incorporates robust semantic information from the text modality, we achieved improved audio guidance and attained a mIoU of \\(66.02\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.574, + 0.788, + 0.635 + ], + "angle": 0, + "content": "As shown in Tab. 6, we also implement a baseline that directly filters the inferred semantic labels during the model evaluation process by intersecting them with the potential sounding object collected by the frozen LLM capturer. This implies that no text cues are utilized to enhance the audio guidance for better" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.665, + 0.473, + 0.722 + ], + "angle": 0, + "content": "Table 5: Ablation study on AVS-V1M. Our method benefits from using both audio and text as the guiding information for sounding objects." + }, + { + "type": "table", + "bbox": [ + 0.245, + 0.734, + 0.445, + 0.839 + ], + "angle": 0, + "content": "
ControlmIoU(%)F-score
TeSO66.020.801
(-) LinfoNCE65.340.793
(-) Dynamic Mask64.780.781
(-) PMQS62.900.774
(-) SeDAM62.220.760
(-) Text62.280.768
(-) Audio42.110.692
(-) Text & Audio37.350.656
" + }, + { + "type": "table_caption", + "bbox": [ + 0.497, + 0.665, + 0.788, + 0.748 + ], + "angle": 0, + "content": "Table 6: Comparison of our proposed baseline and TeSO. The baseline method employs LLM to filter the predicted results but does not enhance the audio guidance with text cues. Experiments are conducted on AVS-V1M test set." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.761, + 0.776, + 0.811 + ], + "angle": 0, + "content": "
Normal mIoU%Mute mIoU% (Δm%)WGN-40dB mIoU% (Δm%)
Baseline64.1543.37 (32.4)43.09 (32.8)
TeSO66.0237.89 (42.6)34.74 (47.4)
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.193 + ], + "angle": 0, + "content": "control over the visual space. As expected, the sensitivity \\((\\Delta_{m})\\) of model performance to changes in audio input has decreased compared with TeSO, indicating the effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.221, + 0.355, + 0.237 + ], + "angle": 0, + "content": "5 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.258, + 0.787, + 0.379 + ], + "angle": 0, + "content": "Noisy text cues. Besides dense image captions for collecting semantic text cues, we also constructed a dynamic mask in Eq. (4) to further select effective semantics based on audio correspondence. We use audio recognition results to determine the accuracy of text cues. Please refer to Tab. 3 to compare the audio recognition results based on text cues collection stage \\((\\mathrm{REC}_{\\#1})\\) and dynamic mask filtering stage \\((\\mathrm{REC}_{\\#2})\\). Our approach achieves higher audio recognition performance, compared with using the audio foundation model in BAVS. This indicates the reliability and accuracy of our strategy for obtaining text cues." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.381, + 0.787, + 0.47 + ], + "angle": 0, + "content": "Prompt with CoT. As shown in the Fig. 3, we utilize task-specific prompt templates with multiple scenario demonstrations and CoT instructions to assist the frozen LLM in collecting more accurate potential sounding objects as text cues. For the detailed prompt templates with CoT instructions and related experimental information, please refer to the Prompt Template part and Text Assistance part in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.472, + 0.787, + 0.563 + ], + "angle": 0, + "content": "Versatility. We explore an alternative visual foundation model like the Segment Anything Model [19] for segmentation. Conducting experiments using a ViT-based backbone with minimal adjustments, our model achieves an impressive \\(65.36\\%\\) mIoU on V1M, still getting comparable performance. This demonstrates the versatility and effectiveness of our model. Additional experiment details and analysis can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.591, + 0.358, + 0.607 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.629, + 0.787, + 0.84 + ], + "angle": 0, + "content": "This study represents a pioneering effort to incorporate text cues into the AVS task. Unlike previous approaches that solely rely on audio assistance, our TeSO method establishes a stronger audio-visual correlation by harnessing the robust semantics unique to the text modality By integrating CoT instructions and common-sense reasoning, we enhance the ability to capture potential sounding objects from detailed scene descriptions, thereby optimizing the utilization of text cues. Furthermore, we have developed a dynamic mask based on attention scores to effectively handle noisy information present in the text cues. Our comprehensive experiments highlight the effectiveness of our approach, demonstrating comparative performance across all AVS Benchmarks. In particular, more observation experiments validate that enhancing audio with text cues grants superior audio guidance. In summary, our work provides a new direction for building reliable audio-visual correlation with textual semantics, and we hope that it can help the community construct more robust audio-visual systems." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.449, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Sounding Object Segmentation Preference" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.179, + 0.785, + 0.21 + ], + "angle": 0, + "content": "This research was supported by National Natural Science Foundation of China (NO.62106272), and Public Computing Cloud, Renmin University of China." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.234, + 0.323, + 0.251 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.267, + 0.786, + 0.31 + ], + "angle": 0, + "content": "1. Abu-El-Haija, S., Kothari, N., Lee, J., Natsev, P., Toderici, G., Varadarajan, B., Vijayanarasimhan, S.: Youtube-8m: A large-scale video classification benchmark. arXiv preprint arXiv:1609.08675 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.787, + 0.352 + ], + "angle": 0, + "content": "2. Bao, F., Nie, S., Xue, K., Li, C., Pu, S., Wang, Y., Yue, G., Cao, Y., Su, H., Zhu, J.: One transformer fits all distributions in multi-modal diffusion at scale. arXiv preprint arXiv:2303.06555 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.352, + 0.787, + 0.381 + ], + "angle": 0, + "content": "3. Chen, S., Wu, Y., Wang, C., Liu, S., Tompkins, D., Chen, Z., Wei, F.: Beats: Audio pre-training with acoustic tokenizers. arXiv preprint arXiv:2212.09058 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.381, + 0.787, + 0.408 + ], + "angle": 0, + "content": "4. Cheng, B., Choudhuri, A., Misra, I., Kirillov, A., Girdhar, R., Schwing, A.G.: Mask2former for video instance segmentation (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.409, + 0.787, + 0.464 + ], + "angle": 0, + "content": "5. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290-1299 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.464, + 0.787, + 0.519 + ], + "angle": 0, + "content": "6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.519, + 0.787, + 0.575 + ], + "angle": 0, + "content": "7. Espejel, J.L., Ettifouri, E.H., Alassan, M.S.Y., Chouham, E.M., Dadhane, W.: Gpt-3.5, gpt-4, or bard? evaluating llms reasoning ability in zero-shot setting and performance boosting through prompts. Natural Language Processing Journal 5, 100032 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.576, + 0.787, + 0.603 + ], + "angle": 0, + "content": "8. Gao, S., Chen, Z., Chen, G., Wang, W., Lu, T.: Avsegformer: Audio-visual segmentation with transformer (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.603, + 0.787, + 0.659 + ], + "angle": 0, + "content": "9. Gemmeke, J.F., Ellis, D.P., Freedman, D., Jansen, A., Lawrence, W., Moore, R.C., Plakal, M., Ritter, M.: Audio set: An ontology and human-labeled dataset for audio events. In: 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP). pp. 776-780. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.787, + 0.714 + ], + "angle": 0, + "content": "0. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15180-15190 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.714, + 0.787, + 0.743 + ], + "angle": 0, + "content": "1. Hao, D., Mao, Y., He, B., Han, X., Dai, Y., Zhong, Y.: Improving audio-visual segmentation with bidirectional generation. arXiv preprint arXiv:2308.08288 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.743, + 0.787, + 0.785 + ], + "angle": 0, + "content": "2. Van der Heiden, R.M., Janssen, C.P., Donker, S.F., Hardeman, L.E., Mans, K., Kenemans, J.L.: Susceptibility to audio signals during autonomous driving. PloS one 13(8), e0201963 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.787, + 0.841 + ], + "angle": 0, + "content": "3. Hershey, S., Chaudhuri, S., Ellis, D.P., Gemmeke, J.F., Jansen, A., Moore, R.C., Plakal, M., Platt, D., Saurous, R.A., Seybold, B., et al.: Cnn architectures for large-scale audio classification. In: 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP). pp. 131-135. IEEE (2017)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.267, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "14. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.231 + ], + "angle": 0, + "content": "15. Hu, D., Wei, Y., Qian, R., Lin, W., Song, R., Wen, J.R.: Class-aware sounding objects localization via audiovisual correspondence. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9844-9859 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "16. Huang, S., Li, H., Wang, Y., Zhu, H., Dai, J., Han, J., Rong, W., Liu, S.: Discovering sounding objects by audio queries for audio visual segmentation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "17. Hur, C., Park, H.: Zero-shot image classification with rectified embedding vectors using a caption generator. Applied Sciences 13(12), 7071 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.288, + 0.785, + 0.327 + ], + "angle": 0, + "content": "18. Johnson, J., Karpathy, A., Fei-Fei, L.: Densecap: Fully convolutional localization networks for dense captioning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4565-4574 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.328, + 0.785, + 0.369 + ], + "angle": 0, + "content": "19. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.411 + ], + "angle": 0, + "content": "20. Li, K., Yang, Z., Chen, L., Yang, Y., Xun, J.: Catr: Combinatorial-dependence audio-queried transformer for audio-visual video segmentation. arXiv preprint arXiv:2309.09709 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.438 + ], + "angle": 0, + "content": "21. Ling, Y., Li, Y., Gan, Z., Zhang, J., Chi, M., Wang, Y.: Hear to segment: Unmixing the audio to guide the semantic segmentation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "22. Liu, C., Li, P., Qi, X., Zhang, H., Li, L., Wang, D., Yu, X.: Audio-visual segmentation by exploring cross-modal mutual semantics (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "23. Liu, C., Li, P., Zhang, H., Li, L., Huang, Z., Wang, D., Yu, X.: Bavs: Bootstrapping audio-visual segmentation by integrating foundation knowledge. arXiv preprint arXiv:2308.10175 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "24. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.577 + ], + "angle": 0, + "content": "25. Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: Llava-last: Improved reasoning,OCR,and world knowledge (January 2024),https://llava-v1.github.io/blog/2024-01-30-llava-last/" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.689, + 0.591 + ], + "angle": 0, + "content": "26. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.619 + ], + "angle": 0, + "content": "27. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.646 + ], + "angle": 0, + "content": "28. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "29. Liu, J., Ju, C., Ma, C., Wang, Y., Wang, Y., Zhang, Y.: Audio-aware query-enhanced transformer for audio-visual segmentation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "30. Liu, J., Wang, Y., Ju, C., Zhang, Y., Xie, W.: Annotation-free audio-visual segmentation. arXiv preprint arXiv:2305.11019 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.757 + ], + "angle": 0, + "content": "31. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "32. Ma, J., Sun, P., Wang, Y., Hu, D.: Stepping stones: A progressive training strategy for audio-visual semantic segmentation. IEEE European Conference on Computer Vision (ECCV) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "33. Majumder, S., Al-Halah, Z., Grauman, K.: Move2hear: Active audio-visual source separation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 275–285 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.449, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Sounding Object Segmentation Preference" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "34. Mo, S., Tian, Y.: Av-sam: Segment anything model meets audio-visual localization and segmentation. arXiv preprint arXiv:2305.01836 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.787, + 0.217 + ], + "angle": 0, + "content": "35. Park, S., Senocak, A., Chung, J.S.: Marginnce: Robust sound localization with a negative margin. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.785, + 0.257 + ], + "angle": 0, + "content": "36. Senocak, A., Oh, T.H., Kim, J., Yang, M.H., Kweon, I.S.: Learning to localize sound source in visual scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4358-4366 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.284 + ], + "angle": 0, + "content": "37. Sharma, H., Jalal, A.S.: Image captioning improved visual question answering. Multimedia tools and applications 81(24), 34775-34796 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.787, + 0.311 + ], + "angle": 0, + "content": "38. Shi, Z., Zhou, X., Qiu, X., Zhu, X.: Improving image captioning with better use of captions. arXiv preprint arXiv:2006.11807 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.462 + ], + "angle": 0, + "content": "39. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D., Blecher, L., Ferrer, C.C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I., Korenev, A., Koura, P.S., Lachaux, M.A., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A., Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X.E., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.463, + 0.785, + 0.503 + ], + "angle": 0, + "content": "40. Wang, Y., Liu, W., Li, G., Ding, J., Hu, D., Li, X.: Prompting segmentation with sound is generalizable audio-visual source localizer. arXiv preprint arXiv:2309.07929 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.504, + 0.785, + 0.544 + ], + "angle": 0, + "content": "41. Wang, Y., Sun, P., Zhou, D., Li, G., Zhang, H., Hu, D.: Ref-avs: Refer and segment objects in audio-visual scenes. IEEE European Conference on Computer Vision (ECCV) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.545, + 0.785, + 0.572 + ], + "angle": 0, + "content": "42. Wu, W., Yao, H., Zhang, M., Song, Y., Ouyang, W., Wang, J.: Gpt4vis: What can gpt-4 do for zero-shot visual recognition? arXiv preprint arXiv:2311.15732 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.572, + 0.785, + 0.612 + ], + "angle": 0, + "content": "43. Yan, S., Zhang, R., Guo, Z., Chen, W., Zhang, W., Li, H., Qiao, Y., He, Z., Gao, P.: Referred by multi-modality: A unified temporal transformer for video object segmentation. arXiv preprint arXiv:2305.16318 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.613, + 0.785, + 0.653 + ], + "angle": 0, + "content": "44. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of Imms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9(1), 1 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.654, + 0.785, + 0.681 + ], + "angle": 0, + "content": "45. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.681, + 0.785, + 0.722 + ], + "angle": 0, + "content": "46. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127(3), 302-321 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.722, + 0.785, + 0.762 + ], + "angle": 0, + "content": "47. Zhou, J., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., Zhong, Y.: Audio-visual segmentation. In: European Conference on Computer Vision (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.762, + 0.785, + 0.803 + ], + "angle": 0, + "content": "48. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.804, + 0.785, + 0.831 + ], + "angle": 0, + "content": "49. Zürn, J., Burgard, W.: Self-supervised moving vehicle detection from audio-visual cues (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.787, + 0.831 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Can Textual Semantics Mitigate Sounding Object Segmentation Preference_/88d8a449-254f-4c3c-bdf6-605afa985fb1_origin.pdf b/2024/Can Textual Semantics Mitigate Sounding Object Segmentation Preference_/88d8a449-254f-4c3c-bdf6-605afa985fb1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..81ff79cabe7d0d3d6de38684eda873d174e5c396 --- /dev/null +++ b/2024/Can Textual Semantics Mitigate Sounding Object Segmentation Preference_/88d8a449-254f-4c3c-bdf6-605afa985fb1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc1b2bd00b2632b22a1389b3e8f56328b867479f7554e00c30488cafdd2eb8c0 +size 3985125 diff --git a/2024/Can Textual Semantics Mitigate Sounding Object Segmentation Preference_/full.md b/2024/Can Textual Semantics Mitigate Sounding Object Segmentation Preference_/full.md new file mode 100644 index 0000000000000000000000000000000000000000..495de6ec832135139f130406fd11398075104069 --- /dev/null +++ b/2024/Can Textual Semantics Mitigate Sounding Object Segmentation Preference_/full.md @@ -0,0 +1,326 @@ +# Can Textual Semantics Mitigate Sounding Object Segmentation Preference? + +Yaoting Wang $^{1\dagger}$ , Peiwen Sun $^{2\dagger}$ , Yuanchao Li $^{3}$ , Honggang Zhang $^{2}$ , and Di Hu $^{\text{空}1,4}$ + +1 Gaoling School of Artificial Intelligence, Renmin University of China, China + +yaoting.wang@outlook.com + +dihu@ruc.edu.cn + +$^{2}$ Beijing University of Posts and Telecommunications, Beijing, China + +{sunpeiwen,zhhg}@bupt.edu.cn + +3 University of Edinburgh, Scotland, UK + +yuanchao.li@ed.ac.uk + +4 Engineering Research Center of Next-Generation Search and Recommendation + +Abstract. The Audio-Visual Segmentation (AVS) task aims to segment sounding objects in the visual space using audio cues. However, in this work, it is recognized that previous AVS methods show a heavy reliance on detrimental segmentation preferences related to audible objects, rather than precise audio guidance. We argue that the primary reason is that audio lacks robust semantics compared to vision, especially in multi-source sounding scenes, resulting in weak audio guidance over the visual space. Motivated by the fact that text modality is well explored and contains rich abstract semantics, we propose leveraging text cues from the visual scene to enhance audio guidance with the semantics inherent in text. Our approach begins by obtaining scene descriptions through an off-the-shelf image captioner and prompting a frozen large language model to deduce potential sounding objects as text cues. Subsequently, we introduce a novel semantics-driven audio modeling module with a dynamic mask to integrate audio features with text cues, leading to representative sounding object features. These features not only encompass audio cues but also possess vivid semantics, providing clearer guidance in the visual space. Experimental results on AVS benchmarks validate that our method exhibits enhanced sensitivity to audio when aided by text cues, achieving highly competitive performance on all three subsets. Project page: https://github.com/GeWu-Lab/Sounding-Object-Segmentation-Preference + +Keywords: Audio-Visual Segmentation $\cdot$ Textual Guidance $\cdot$ Multi-modal learning + +![](images/654b534457e0ff9452c91ebdf4e1d512b7c36414ac7bfb9b2ff7b9260d0c95d8.jpg) +Fig. 1: The previous methods (left) achieve satisfactory results in receiving normal audio input. However, even when the sound is completely silent, they still segment vast pixels to represent the guitar regarding the segmentation preference of audible objects built during the training. In contrast, our approach (right) utilizes a frozen LLM to reason from scene descriptions that the male is not playing the guitar, as his hands are off the strings. Guided by this semantic information, our method produces more precise segmentation with finer audio guidance. + +# 1 Introduction + +In human perception, visual and auditory senses are closely related, and audio can provide vision with additional and dynamic scene information to enhance visual understanding. In the context of audio-visual understanding tasks, Audio-Visual Localization (AVL) enables the localization of sounding objects within visual scenes using audio references [15, 36]. However, the coarse localization of sounding objects is no longer sufficient to meet the practical demands of complex scene understanding, such as in autonomous driving [12, 49] and augmented reality [33]. To address this challenge, tasks are gradually shifting from bounding-box or rough heat-map localization to finer-grained pixel-level segmented localization, which is known as Audio-Visual Segmentation (AVS). + +The existing works on AVS can be broadly categorized into fusion-based methods [8, 16, 20, 21, 29, 47] and prompt-based methods [30, 34, 40]. The former primarily focuses on localizing sounding objects by fusing audio and visual features, while the latter emphasizes constructing effective audio prompts for the visual foundation model. However, most of these methods fail to establish compact audio-visual correlations to achieve clear audio guidance in the visual space. As depicted in the left of Fig. 1, previous methods could achieve satisfactory results with normal audio input. Nevertheless, when the audio is silent, + +these methods still segment many pixels to represent potential sounding objects based on the segmentation preferences of audible objects1. + +Furthermore, from the experiment with manually muted audio input, we observe that most of the behavior of AVS baselines depends on the segmentation preference established during the training process rather than relying on reliable audio-visual correlation. For instance, these models tend to take a shortcut by segmenting out the muted guitars because the audio guidance is weak, and they simply learned with the easy-to-learn visual feature alone (i.e., shortcut learning) that guitars are often associated with sound-emitting during the training phase. The unreliable audio-visual correlation can be attributed to two main factors. Firstly, the scarcity of training data for AVS is a significant challenge due to the demanding nature of pixel-level annotation. Secondly, the audio modality itself presents inherent complexity and ambiguity [16, 21, 29], especially in scenarios involving multiple audio sources that may be intertwined. + +Given the challenges discussed above, our approach aims to build a robust audio-visual correlation by integrating the text modality, which inherently provides meaningful and robust semantic information about objects and their interactions. Firstly, in the Scene Describing phase, we utilize an off-the-shelf image captioner (LLaVA-1.5 [28]) to generate detailed dense captions for scene understanding. Then, in the Sounding Objects Reasoning phase, we instruct the frozen large language models (LLMs, e.g., LLaMA2 [39]) as a text cues capturer to collect potential sounding objects from the scene description as text cues. Notably, we introduce an elaborate few-shot prompt template to guide the text cues capturer to reason with Chain-of-Thought (CoT) instructions. After that, in the Semantics-Driven Audio Modeling (SeDAM) module, we project the audio feature into latent audio features. However, instead of directly using these latent features to interact with visual features for mask decoding, we leverage these latent features to interact with text cues. Specifically, the latent features are attended to and collected to form sounding object features with semantics provided by the text cues through the crossmodal transformer. We further propose the Prompting Mask Queries with Semantics (PMQS) module to introduce the sounding object features into pre-trained mask queries for later segmentation. Finally, we use simple but effective bottleneck adapters in the Audio-Prompted Decoding phase for better segmentation quality. The experimental results demonstrate that our proposed method improves audio-visual grounding by incorporating guidance from both audio and text cues. + +Experiments in Sec. 4.4 reveal the advantage of our method, as verified by its increased sensitivity to the changes in audio input compared to networks without enhancing audio guidance with text cues. In summary, our contributions are threefold: + +- We recognize the behavior of AVS model is influenced by the preference segmentation phenomenon, which can not be solved easily by previous audio- + +visual based methods. A novel strategy introduces textual semantics as the bridge into the AVS task, establishing a better audio-visual correlation. + +- We introduce the SeDAM module to form sounding object features with audio and text cues. We also suggest a PMQS module to prompt the pretrained mask queries with these sounding object features for sounding object segmentation using a visual foundation model. + +- Our experiments show that our method can be more sensitive to the audio guidance, manifested in its high sensitivity to the changes in audio input. + +# 2 Related Works + +# 2.1 Audio-Visual Localization and Segmentation + +Traditional AVL tasks [15,35,36] predict the positions of sounding objects using bounding-box or coarse heat-maps through unsupervised learning of audio-visual correlation. In recent years, driven by the growing demand for more precise localization in industries such as autonomous driving [12,49] and augmented reality [33], the AVL task has evolved towards a finer-grained AVS task that localizes sounding objects at a pixel-level. The existing AVS works can be broadly categorized into fusion-based [8, 16, 20, 21, 29, 43, 47] and prompt-based [30, 32, 40] The pioneering AVS work [47] is fusion-based and adopts a multi-stage strategy to integrate audio with multi-scale visual features. Liu et al. [29] address the issues of inadequate fusion of audio-visual features with an audio-aware query-enhanced transformer (AuTR). In comparison, AVSegFormer [8] directly decodes the fused feature with the audio query. Wang et al. [40] innovatively prompt the visual foundation model with audio input, harnessing the abundant visual before achieving generalizable AVS in zero-shot and few-shot scenarios. However, these methods have not effectively established a reliable audio-visual correlation, preventing the models from perceiving robust audio guidance. On the contrary, our experiments in Sec. 4.4 also reveal that some AVS baselines' competitive performance may stem from segmentation preferences formed during the training. + +# 2.2 Text Aided Scene Understanding + +Before the explosion of LLM, textual semantic information had been widely used for visual understanding tasks. Zhan et al. [38] and Sharma et al. [37] improve Visual Question Answering by using the extra semantic information of image captions, while Hur and Park [17] utilize the image captioner to help zero-shot image classification. Wang et al. [41] incorporating text with audio and visual cues to solve reference AVS task. + +In recent years, the community has witnessed an emergent interest in strong off-the-shelf capabilities in open-world visual understanding [24-26, 45]. Notable examples, such as GPT4(V) [44] and LLaVA [27] have showcased remarkable linguistic and visual capabilities in zero-shot scene understanding [7, 42] without any training requirement. Innovatively, in the field of AVS, BAVS [23] replaces + +![](images/3b7384491ed32f0e29fa98610f0ed5b3dcd66374660b90b82c87554356c724c4.jpg) +Fig. 2: Overall pipeline of the proposed TeSO method. We utilize an off-the-shelf image captioner for dense scene describing and employ a frozen LLM to reason out potential sounding objects as text cues. These semantic text cues are then aggregated with audio features in the SeDAM module to form the sounding object features. Subsequently, we introduce the sounding object features into pre-trained mask queries in the PMQS module. Finally, we use adapters to tune the visual-only mask decoder for AVS in the Audio-prompted Decoding phase. "MSDA" is the multi-scale deformable attention proposed by Zhu et al. [48]. + +audio input with textual audio tags from a large pre-trained vision (UniDiffuser [2]) and audio (BEATs [3]) foundation model, establishing a reliable AVS system with foundation knowledge. However, due to the significant gap in class distribution between the pre-training dataset and the AVS dataset, as well as the presence of multi-source scenarios in the AVS task, the extracted audio tags are not ideal to provide accurate semantics. In contrast, we utilize the text modality with robust semantic features to enhance rather than replace the audio guidance and address the segmentation preference issue. + +# 3 Text-guided Sounding Object Segmentation + +As shown in Fig. 2, we introduce our Text-guided Sounding Object Segmentation (TeSO) method in this section. Our experiments about muted audio input imply that audio does not exert reasonable guidance over the visual space for some AVS baselines, resulting in the models relying more on the segmentation preferences established during the training process. In essence, they learn which objects are most likely audible during training and take shortcuts during inference. To address the above issue, our method aims to enhance the audio-visual correlation by leveraging the text modality, which inherently possesses robust semantic information, to obtain finer-grained audio guidance. We begin by acquiring detailed scene descriptions through an off-the-shelf image captioner. Subsequently, a frozen LLM works as the text cues capturer, collecting potential sounding objects as text cues from the scene descriptions with CoT instructions. + +Finally, we introduce a novel SeDAM module with a dynamic mask to seamlessly integrate audio features and text semantics through a crossmodal transformer, providing finer audio guidance with text semantic cues. + +# 3.1 Multimodal Representation + +Visual: Following previous works [40, 47], we sample frames from the video at 1-second intervals. Mask2Former [4] serves as our visual foundation model, and we extract visual features $F_V \in \mathbb{R}^{d_V \times H \times W}$ from a pre-trained Swin transformer [6]. We use simple two-layer MLP adapters to tune the visual encoder for better visual representations. + +Audio: Similar to the video processing, we split the audio into clips at 1-second intervals. We extract audio features $F_{As} = \{F_{A_1}, F_{A_2}, \dots, F_{A_T}\} \in \mathbb{R}^{T \times d_A}$ using VGGish [9, 13], where $T$ represents the length of the audio in seconds, matching the number of video frames. Therefore, each video frame corresponds to $F_{A} = F_{As}[i]$ . We freeze the pre-trained parameters for a better comparison with previous works. + +Text: For each video frame, we complement the captured text cues into a sentence using a template ("This is a $\{\_ \}$ .") to extract text features $F_{T} \in \mathbb{R}^{N_{T} \times d_{T}}$ with ImageBind [10], where $N_{T}$ is the number of text cues. + +# 3.2 Text Cues Generation + +To obtain more precise text cues, we use a two-stage inference flow. Firstly, we employ an off-the-shelf image captioner to generate dense scene descriptions. Secondly, we ask a text cues capturer to collect potential sounding objects from the generated descriptions. + +Scene Describing. As shown in the top-left of Fig. 3, in this phase, we employ the commonly used and straightforward prompt: "Please carefully understand the image content and provide as rich a description as possible." This instructs the image captioner to generate a detailed description of the visual scene, a task commonly referred to as dense captioning [18]. This process enables the image captioner to produce comprehensive and informative descriptions that not only describe the objects like "a human and guitar in the image" but also capture their interactions and relationships. + +Sounding Object Collecting. To extract potential sounding objects from the generated dense descriptions, we design a task-specific CoT prompt template with few-shot demonstrations, as illustrated in Fig. 3. Specifically, we carefully craft examples that cover various scenarios and provide reasoning approaches for each example. The process begins by identifying the objects present in the dense scene descriptions, followed by step-by-step common-sense reasoning according to the CoT instructions, as depicted in the lower right of Fig. 3. During the CoT reasoning, the text cues capturer first categorizes the sample objects and + +![](images/9b78b8aefc443ced04e43ae181a90ca207178e1f8a63d5cdb0ba66a8549c3e3e.jpg) +Fig. 3: "P.S.O" stands for "Potential Sounding Object". A frozen LLM reasoner works as a text cues capturer by considering the interaction of audible objects. For instance, a person may sing a song while playing the guitar, but he would not sing along with a saxophone. In contrast, a noun parser simply captures any nouns that are present. + +![](images/9eeb3064ba0d82e908aa542331c861dda184f5a60a1efcf17b9011ca2e9e3a82.jpg) + +assesses whether objects in that category are likely making sounds. Subsequently, the capturer analyzes which group of potential sounding objects can produce sounds simultaneously. Finally, the group with higher confidence is captured as the text cues of length $N_{T}$ and encoded in $F_{T}$ for further guidance. The right part of Fig. 3 also illustrates why we use the frozen LLM as a text cues capturer instead of using a naive noun parser. It is important to note that the use of LLM is solely for generating high-quality text cues, the LLM itself is not engaged in the training process for a fairness concern. + +# 3.3 Semantic-Driven Audio Modeling + +After obtaining potential sounding objects, we can enhance the semantics of audio features with these text cues in this SeDAM module. To allow semantic text cues to capture different aspects of audio features $F_{A}$ , we project $F_{A}$ into distinct latent features $F_{A_{L}} = \{a_{i}\}_{i=1}^{N_{L}} \in \mathbb{R}^{N_{L} \times d_{A_{L}}}$ , where $N_{L}$ is the number of latent features. It is worth noting that, we incorporate an auxiliary loss term $\mathcal{L}_{infoNCE}$ aims at ensuring the distinction among latent audio features. This way, we project audio features into the latent space, where each unique $a_{i}$ represents a latent audio feature to be captured by semantic text cues, forming clear and loud sounding object features with robust semantics for finer guidance. + +To incorporate the semantics of text cues, we employ text cues as the query to attend and combine the relevant latent features through crossmodal multi-head attention, formulated as: + +$$ +F _ {\widetilde {A} _ {L}} = \operatorname {M u l t i H e a d} \left(F _ {T}, F _ {A _ {L}}, F _ {A _ {L}}\right), \tag {1} +$$ + +where $F_{\widetilde{A}_L} \in \mathbb{R}^{N_T \times d_A}$ is the attended latent audio features and will be combined with text cues $F_T$ as follows: + +$$ +F _ {A _ {T}} = F C \left(L N \left(F _ {T} + F _ {\widetilde {A} _ {L}}\right)\right), \tag {2} +$$ + +where $F_{A_T} \in \mathbb{R}^{N_T \times d_A}$ is the combined feature with both audio information and semantic text cues, and $FC$ and $LN$ for linear layer and layer normalization. Note that the LLM reasoner may generate a different number of text cues, and we either pad with a zero matrix or truncate to $N_T$ . It is noteworthy that we propose task-specific CoT instructions to prompt the text cues capturer to infer more accurate potential sounding objects as the text cues. Additionally, we also introduce a dynamic mask based on attention score to mitigate the potential noisy information. + +We first compute the average pooling attention scores $\zeta \in \mathbb{R}^{N_T}$ for text cues on the attention head and key dimension: + +$$ +\xi = \operatorname {P o o l} \left(\frac {F _ {T} F _ {A _ {L}} ^ {\mathrm {T}}}{\sqrt {d _ {A _ {L}}}}\right), \tag {3} +$$ + +where $\sqrt{d_{A_L}}$ is the scaling factor for dot production. Then we form the attention mask $\mathcal{M} = \{m_1, m_2, m_3, \dots, m_{N_T}\}$ as follows: + +$$ +\mathcal {M} _ {i} = \left\{ \begin{array}{l l} 0 & \text {i f} \xi_ {i} > 0 \\ - \infty & \text {o t h e r w i s e} \end{array} , \right. \tag {4} +$$ + +where $\mathcal{M} \in \mathbb{R}^{N_T}$ , will be used as a key padding mask for the PMQS module in the next section to ensure that the attention weight of redundant text cues approaches zero. + +# 3.4 Prompting Mask Queries with Semantics + +Current segmentation foundation models [4, 5, 19] adopt transformer-based decoder with mask queries as the mainstream approach. To prompt such foundation models, we propose the PMQS module, aiming to introduce semantic information to mask query features $F_{Q} = \{q_{1}, q_{2}, \dots, q_{N_{Q}}\}$ , where $F_{Q} \in \mathbb{R}^{N_{Q} \times d_{V}}$ , and $N_{Q}$ is the number of mask queries in crossmodal attention. We formulate this updated expression as follows: + +$$ +\phi_ {c a} \left(F _ {Q}, F _ {A _ {T}}, F _ {A _ {T}}\right) = S o f t m a x \left(\frac {F _ {Q} F _ {A _ {T}} ^ {\mathrm {T}}}{\sqrt {d _ {A _ {T}}}} + \mathcal {M}\right) F _ {A _ {T}}, \tag {5} +$$ + +where $\mathcal{M}$ is added to the attention score to mask the redundant text cues for each query. Then we obtain the updated mask query features $\widetilde{F}_Q\in \mathbb{R}^{N_Q\times d_V}$ as: + +$$ +\widetilde {F} _ {Q} = F _ {Q} + \phi_ {c a} \left(F _ {Q}, F _ {A _ {T}}, F _ {A _ {T}}\right). \tag {6} +$$ + +# 3.5 Audio-Prompted Decoding + +As shown on the right side of Fig. 2, the updated mask query features $\widetilde{F}_Q$ are forwarded into the transformer-based decoder, generating mask queries $M_Q \in \mathbb{R}^{N_Q \times H \times W}$ and class queries $C_Q \in \mathbb{R}^{N_Q \times N_C}$ , where $N_C$ is the number of class queries pre-defined in the segmentation foundation model. Subsequently, we can obtain the final segmentation $M_{pred} \in \mathbb{R}^{N_C \times H \times W}$ : + +$$ +M _ {p r e d} = \operatorname {E i n s u m} \left(C _ {Q}, \bar {F} _ {Q}\right), \tag {7} +$$ + +where $\text{Einsum}(\cdot)$ is the Einstein Summation Convention. + +Notably, to enhance the quality of the segmentation, we use two-layer bottle-neck MLP adapters [14] to lightly fine-tune the output projection of the following attentions: a) Swin Window Attention (W-MSA) [31], b) Deformable Attention in pixel decoder [48] and c) Masked Attention in transformer decoder [4]. + +# 3.6 Learning Objectives + +Segmentation Loss. During the model training process, we employ the binary cross-entropy loss and dice loss to optimize the mask quality: + +$$ +\mathcal {L} _ {\text {m a s k}} = \lambda_ {b c e} \cdot \mathcal {L} _ {b c e} + \lambda_ {d i c e} \cdot \mathcal {L} _ {d i c e}. \tag {8} +$$ + +Further, we adopt the mask classification loss [5] to compose the final segmentation loss: + +$$ +\mathcal {L} _ {s e g} = \mathcal {L} _ {m a s k} + \lambda_ {c l s} \cdot \mathcal {L} _ {c l s}. \tag {9} +$$ + +Latent Component Loss. This loss term aims to explicitly guide the model to learn distinct representations for each latent audio component, helping prevent redundancy and overlapping information across the latent components. + +$$ +\mathcal {L} _ {\text {i n f o N C E}} = - \log \left(\frac {\exp \left(\operatorname {s i m} \left(a , a ^ {+}\right)\right)}{\sum_ {i = 1} ^ {N _ {L}} \exp \left(\operatorname {s i m} \left(a , a _ {i}\right)\right)}\right), \tag {10} +$$ + +where $a^+$ and $sim(\cdot)$ denote the positive example and the cosine similarity. + +Total Loss The final loss function is composed of the weighted sum of the two kinds of losses above: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {s e g}} + \lambda_ {\text {i n f o N C E}} \cdot \mathcal {L} _ {\text {i n f o N C E}}. \tag {11} +$$ + +# 4 Experiments + +# 4.1 Implementation Details + +Dataset. Our proposed method is evaluated on the AVS Benchmarks [47], which contains three subsets. Firstly, the single-source subset (V1S) contains 4932 videos over 23 categories, covering various sounds, such as humans, animals, + +vehicles, and musical instruments. For videos in the training split of this subset, only the first sampled frame is annotated. Secondly, the multi-source subset (V1M) contains 424 videos that include two or more categories from V1S, and all sounding objects are visible in the frames. Finally, the AVSBench-semantic (AVSS) subset, an extension of V1S and V1M, contains 12,356 videos. These newly collected videos are trimmed to 10 seconds, differing from the 5 seconds in V1S and V1M. + +Setting. We conduct training and evaluation using the VGGish backbone pretrained on Youtube-8M [1] and Swin-base Transformer backbone pretrained on semantic-ADE20K [46] by Mask2Former [4]. The number $N$ of the crossmodal transformer layers is set to 4, and the parameters $\lambda_{bce},\lambda_{dice},\lambda_{cls},\lambda_{infoNCE}$ in the loss are set to 5,5,2,1. The AdamW optimizer is adopted with a learning rate of 1e-4 for the visual encoder adapters and 1e-3 for other learnable parameters, and the training epoch is roughly set to 60. + +Metrics. To conduct a comprehensive evaluation of our model, we carry out tests using mean Intersection over Union (mIoU) and F-score as the performance metrics, following previous works [8, 47]. + +# 4.2 Comparison Results + +When compared to methods that do not incorporate text as additional semantic information, our method, which leverages captioning and reasoning, demonstrates strong competitiveness. As illustrated in Sec. 4.2, our method achieves comparable results across all V1S, V1M, and AVSS-binary subsets, showcasing average performance improvements of up to $1.25\%$ in mIoU and $2.7\%$ in F-score, respectively. Furthermore, qualitative analysis indicates that our predicted masks exhibit superior quality, as demonstrated in sub-figure (1) of Fig. 4 (more examples in supplementary materials). + +It is necessary to mention that our proposed TeSO does not show significant improvements on V1S over certain existing methods, such as BAVS [23], which also utilizes text semantic information. Apart from marginal effects, this phenomenon can also be attributed to the singular sound source and simplistic nature of the data scenarios present in V1S. On V1M, however, our performance significantly surpasses that of other methods, including BAVS. + +In addition to the AVS task, we also conduct experiments on the Audio-Visual Semantic Segmentation (AVSS) task, as shown in Tab. 2. Our model exceeds the best performance in previous models by $5.37\%$ on mIoU, which further demonstrates that our model is capable of establishing a better audio-visual correlation. + +# 4.3 Bottlenecks and Fairness + +Our method (TeSO) collects text cues from dense image descriptions at $\mathrm{REC}_{\#1}$ $(80.41\%)$ and a filter with a dynamic mask at $\mathrm{REC}_{\#2}$ $(80.95\%)$ , indicating a robust startup for the text-guided AVS framework, in comparison with another two-stage model, BAVS. + +Table 1: Performance on AVS-Benchmarks. The underscore is used to indicate suboptimal results. BAVS uses large pre-trained visual and audio foundation models to replace audio input with text tag. In comparison, TeSO (ours) enhances rather than replaces audio guidance based on the robust semantics of text. It is evident that our method has a good F-score result while ensuring mIoU. + +
MethodAudio-backboneVisual-backboneV1SV1MAVSS-binary
mIoU(%)F-scoremIoU(%)F-scoremIoU(%)F-score
AVSBench [47]VGGishPVT-v278.700.87954.000.64562.450.756
AVSegFormer [8]VGGishPVT-v282.060.89958.360.69364.340.759
AVSC [22]VGGishPVT-v281.290.88659.500.657--
AVS-BG [11]VGGishPVT-v281.710.90455.100.668--
AQFormer [16]VGGishPVT-v281.600.89461.100.721--
†CATR [20]VGGishPVT-v281.400.89659.000.700--
AV-SAM [34]ResNet18ViT-Base40.470.566----
SAMA [30]VGGishViT-Huge81.530.88663.140.691--
GAVS [40]VGGishViT-Base80.060.90263.700.77467.700.788
AuTR [29]VGGishSwin-Base80.400.89156.200.672--
MUTR [43]VGGishVideo-Swin-Base81.600.89764.000.735--
BAVS [23]‡BEATsSwin-Base82.680.89859.630.65955.450.640
TeSO (ours)VGGishSwin-base83.270.93366.020.80168.530.813
+ +$\dagger$ : To make a fair comparison, the results of CATR here are without supplemented annotation of the training set. +: BEATs is a stronger audio backbone, trained on AudioSet dataset with 90M parameters. In comparison, VGGish is trained on Youtube-8M with 70M parameters. + +Table 2: Performance on AVSS dataset. Our method showcases a substantial improvement in the semantic subset, achieved by leveraging robust text semantics to enhance audio guidance. + +
MethodBackboneAVSS
mIoU(%)F-Score
AVSBenchPVT-v229.800.352
CATRPVT-v232.800.385
BAVSSwin-base33.590.375
TeSO (Ours)Swin-base38.960.451
+ +Table 3: Potential sounding objects recognition (REC) performance at different model stages and the corresponding segmentation results. Experiments are conducted on AVS-V1M test set. + +
MethodREC(#1)AP(%)REC(#2)AP(%)SEGmIoU(%)
BAVS24.2279.6959.63
TeSOp(ours)77.9678.3363.87
TeSOn(ours)77.8278.2763.31
TeSO(ours)80.4180.9566.02
+ +In addition, we develop several variants of TeSO to investigate its performance sources and bottlenecks, as well as its fairness in comparison with other methods. $\mathrm{TeSO}_p$ denotes the variant using one-shot prompts, while $\mathrm{TeSO}_n$ denotes the variant employing a noun parser instead of a frozen LLM for text cues collection. As shown in Tab. 3, the results demonstrate that both the number (shots) of demonstrations and the method of text cues collection significantly impact the model's ability to accurately identify audio semantics and perform segmentation (SEG). Thus, utilizing few-shot prompts and off-the-shelf foundation models are both essential for ensuring the acquisition of high-quality text cues (i.e., the bottleneck). + +However, we would like to emphasize that, even without using LLM to capture text cues, our method variant $(\mathrm{TeSO}_n)$ still fairly achieves competitive performance (63.31% mIoU) compared with BAVS (59.63% mIoU) and state-of-the-art method GAVS (63.70% mIoU) in the complex multi-source scenario (AVS-V1M), and significantly outperforms other methods. This suggests that + +Table 4: Impact on mIoU and F-score by muted audio or noise-only audio on AVS-V1M test set. TeSO (ours) exhibits the most intensive attenuation up to $48.8\%$ . It is noteworthy that the other models still generate "satisfactory" results even in silent or purely noisy environments. This observation suggests that the models might rely more on the segmentation preferences formed during the training rather than truly learning efficient audio guidance. + +
MethodMuteWGN-10dBWGN-40dB
mIoU% (Δm%)F-score (Δf%)mIoU% (Δm%)F-score (Δf%)mIoU% (Δm%)F-score (Δf%)mIoU% (Δm%)
AVS-Bench49.15 (6.6)0.622 (4.9)46.76 (11.1)0.618 (5.5)47.46 (9.8)0.610 (6.7)
GAVS42.53 (29.2)0.695 (8.4)41.51 (34.8)0.692 (10.6)41.87 (30.3)0.696 (8.4)
BAVS45.06 (26.1)0.644 (5.8)43.27 (27.4)0.601 (8.8)42.93 (31.6)0.595 (9.7)
TeSO (ours)37.89 (42.6)0.671 (16.2)34.20 (49.6)0.661 (17.5)34.74 (47.4)0.656 (18.1)
+ +leveraging text-enhanced audio for visual control with robust semantics is a viable and promising direction. + +# 4.4 Audio Control + +Based on our observations, we notice a distinct shape of the predicted mask in sub-figure (2) of Fig. 4 when some popular AVS baselines are used for inference. Interestingly, this shape remains distinguishable even when the inference is conducted with muted audio. This suggests that the network can predict rough masks without relying on audio input, indicating that the audio-visual grounding is not effectively guided by audio cues. As a result, this undesirable phenomenon can be attributed to the segmentation preference, which is a result of these models lacking proper audio control. + +To evaluate the influence of audio control on segmentation, various models are subjected to inference using muted audio and noise-only audio (at 10 dB and 40 dB) instead of the corresponding audio signals. It should be noted that the visual features remained unchanged throughout the evaluation. As shown in Tab. 4, the disparity $\Delta_{m}$ (and $\Delta_{f}$ ) refers to the difference in mIoU (and F-score) values before and after the audio substitution is applied as an evaluation metric for the sensitivity of audio changes. A larger drop in mIoU and F-score indicates a greater degree of audio control over the visual space. In addition to demonstrating strong segmentation capabilities in the AVS task, TeSO (ours) exhibits the highest level of attenuation, reaching up to $47.4\%$ , when subjected to muted and noise-only audio. Additionally, qualitative analysis revealed that the predicted masks remained predominantly blank, as depicted in sub-figure (2) of Fig. 4 (more examples in supplementary materials). This observation indicates that our model has better audio guidance over the visual space, enabling it to play a crucial role in audio-visual grounding. + +# 4.5 Ablation Study + +As shown in Tab. 5, we conduct ablation experiments targeting the designed TeSO modules and the guidance from different modalities. To explore the guiding ability, we progressively remove either the audio or text guidance for TeSO. + +![](images/bf36008771978a66a07336716644a2120701f1bf9601010fb270afffcc490b7a.jpg) +Fig. 4: Examples of the impact of normal audio input and all-mute audio on popular methods. In normal scenarios, our method shows better masks than previous methods. In all-mute scenarios, our approach exhibits strong sensitivity towards audio inputs, as it is capable of generating blank masks for silent audio clips. + +The experiments indicate that control from both audio and text inputs for visual space can establish a strong audio-visual correlation. Furthermore, by utilizing our proposed TeSO, we achieve enhanced audio guidance and reach a $66.02\%$ mIoU while incorporating precise semantic information from the text modality. Moreover, through the utilization of our proposed TeSO method that incorporates robust semantic information from the text modality, we achieved improved audio guidance and attained a mIoU of $66.02\%$ . + +As shown in Tab. 6, we also implement a baseline that directly filters the inferred semantic labels during the model evaluation process by intersecting them with the potential sounding object collected by the frozen LLM capturer. This implies that no text cues are utilized to enhance the audio guidance for better + +Table 5: Ablation study on AVS-V1M. Our method benefits from using both audio and text as the guiding information for sounding objects. + +
ControlmIoU(%)F-score
TeSO66.020.801
(-) LinfoNCE65.340.793
(-) Dynamic Mask64.780.781
(-) PMQS62.900.774
(-) SeDAM62.220.760
(-) Text62.280.768
(-) Audio42.110.692
(-) Text & Audio37.350.656
+ +Table 6: Comparison of our proposed baseline and TeSO. The baseline method employs LLM to filter the predicted results but does not enhance the audio guidance with text cues. Experiments are conducted on AVS-V1M test set. + +
Normal mIoU%Mute mIoU% (Δm%)WGN-40dB mIoU% (Δm%)
Baseline64.1543.37 (32.4)43.09 (32.8)
TeSO66.0237.89 (42.6)34.74 (47.4)
+ +control over the visual space. As expected, the sensitivity $(\Delta_{m})$ of model performance to changes in audio input has decreased compared with TeSO, indicating the effectiveness of our method. + +# 5 Discussion + +Noisy text cues. Besides dense image captions for collecting semantic text cues, we also constructed a dynamic mask in Eq. (4) to further select effective semantics based on audio correspondence. We use audio recognition results to determine the accuracy of text cues. Please refer to Tab. 3 to compare the audio recognition results based on text cues collection stage $(\mathrm{REC}_{\#1})$ and dynamic mask filtering stage $(\mathrm{REC}_{\#2})$ . Our approach achieves higher audio recognition performance, compared with using the audio foundation model in BAVS. This indicates the reliability and accuracy of our strategy for obtaining text cues. + +Prompt with CoT. As shown in the Fig. 3, we utilize task-specific prompt templates with multiple scenario demonstrations and CoT instructions to assist the frozen LLM in collecting more accurate potential sounding objects as text cues. For the detailed prompt templates with CoT instructions and related experimental information, please refer to the Prompt Template part and Text Assistance part in the supplementary materials. + +Versatility. We explore an alternative visual foundation model like the Segment Anything Model [19] for segmentation. Conducting experiments using a ViT-based backbone with minimal adjustments, our model achieves an impressive $65.36\%$ mIoU on V1M, still getting comparable performance. This demonstrates the versatility and effectiveness of our model. Additional experiment details and analysis can be found in the supplementary materials. + +# 6 Conclusion + +This study represents a pioneering effort to incorporate text cues into the AVS task. Unlike previous approaches that solely rely on audio assistance, our TeSO method establishes a stronger audio-visual correlation by harnessing the robust semantics unique to the text modality By integrating CoT instructions and common-sense reasoning, we enhance the ability to capture potential sounding objects from detailed scene descriptions, thereby optimizing the utilization of text cues. Furthermore, we have developed a dynamic mask based on attention scores to effectively handle noisy information present in the text cues. Our comprehensive experiments highlight the effectiveness of our approach, demonstrating comparative performance across all AVS Benchmarks. In particular, more observation experiments validate that enhancing audio with text cues grants superior audio guidance. In summary, our work provides a new direction for building reliable audio-visual correlation with textual semantics, and we hope that it can help the community construct more robust audio-visual systems. + +# Acknowledgements + +This research was supported by National Natural Science Foundation of China (NO.62106272), and Public Computing Cloud, Renmin University of China. + +# References + +1. Abu-El-Haija, S., Kothari, N., Lee, J., Natsev, P., Toderici, G., Varadarajan, B., Vijayanarasimhan, S.: Youtube-8m: A large-scale video classification benchmark. arXiv preprint arXiv:1609.08675 (2016) +2. Bao, F., Nie, S., Xue, K., Li, C., Pu, S., Wang, Y., Yue, G., Cao, Y., Su, H., Zhu, J.: One transformer fits all distributions in multi-modal diffusion at scale. arXiv preprint arXiv:2303.06555 (2023) +3. Chen, S., Wu, Y., Wang, C., Liu, S., Tompkins, D., Chen, Z., Wei, F.: Beats: Audio pre-training with acoustic tokenizers. arXiv preprint arXiv:2212.09058 (2022) +4. Cheng, B., Choudhuri, A., Misra, I., Kirillov, A., Girdhar, R., Schwing, A.G.: Mask2former for video instance segmentation (2021) +5. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290-1299 (2022) +6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +7. Espejel, J.L., Ettifouri, E.H., Alassan, M.S.Y., Chouham, E.M., Dadhane, W.: Gpt-3.5, gpt-4, or bard? evaluating llms reasoning ability in zero-shot setting and performance boosting through prompts. Natural Language Processing Journal 5, 100032 (2023) +8. Gao, S., Chen, Z., Chen, G., Wang, W., Lu, T.: Avsegformer: Audio-visual segmentation with transformer (2023) +9. Gemmeke, J.F., Ellis, D.P., Freedman, D., Jansen, A., Lawrence, W., Moore, R.C., Plakal, M., Ritter, M.: Audio set: An ontology and human-labeled dataset for audio events. In: 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP). pp. 776-780. IEEE (2017) +0. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15180-15190 (2023) +1. Hao, D., Mao, Y., He, B., Han, X., Dai, Y., Zhong, Y.: Improving audio-visual segmentation with bidirectional generation. arXiv preprint arXiv:2308.08288 (2023) +2. Van der Heiden, R.M., Janssen, C.P., Donker, S.F., Hardeman, L.E., Mans, K., Kenemans, J.L.: Susceptibility to audio signals during autonomous driving. PloS one 13(8), e0201963 (2018) +3. Hershey, S., Chaudhuri, S., Ellis, D.P., Gemmeke, J.F., Jansen, A., Moore, R.C., Plakal, M., Platt, D., Saurous, R.A., Seybold, B., et al.: Cnn architectures for large-scale audio classification. In: 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP). pp. 131-135. IEEE (2017) + +14. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019) +15. Hu, D., Wei, Y., Qian, R., Lin, W., Song, R., Wen, J.R.: Class-aware sounding objects localization via audiovisual correspondence. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9844-9859 (2021) +16. Huang, S., Li, H., Wang, Y., Zhu, H., Dai, J., Han, J., Rong, W., Liu, S.: Discovering sounding objects by audio queries for audio visual segmentation (2023) +17. Hur, C., Park, H.: Zero-shot image classification with rectified embedding vectors using a caption generator. Applied Sciences 13(12), 7071 (2023) +18. Johnson, J., Karpathy, A., Fei-Fei, L.: Densecap: Fully convolutional localization networks for dense captioning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4565-4574 (2016) +19. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023) +20. Li, K., Yang, Z., Chen, L., Yang, Y., Xun, J.: Catr: Combinatorial-dependence audio-queried transformer for audio-visual video segmentation. arXiv preprint arXiv:2309.09709 (2023) +21. Ling, Y., Li, Y., Gan, Z., Zhang, J., Chi, M., Wang, Y.: Hear to segment: Unmixing the audio to guide the semantic segmentation (2023) +22. Liu, C., Li, P., Qi, X., Zhang, H., Li, L., Wang, D., Yu, X.: Audio-visual segmentation by exploring cross-modal mutual semantics (2023) +23. Liu, C., Li, P., Zhang, H., Li, L., Huang, Z., Wang, D., Yu, X.: Bavs: Bootstrapping audio-visual segmentation by integrating foundation knowledge. arXiv preprint arXiv:2308.10175 (2023) +24. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023) +25. Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: Llava-last: Improved reasoning,OCR,and world knowledge (January 2024),https://llava-v1.github.io/blog/2024-01-30-llava-last/ +26. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023) +27. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023) +28. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024) +29. Liu, J., Ju, C., Ma, C., Wang, Y., Wang, Y., Zhang, Y.: Audio-aware query-enhanced transformer for audio-visual segmentation (2023) +30. Liu, J., Wang, Y., Ju, C., Zhang, Y., Xie, W.: Annotation-free audio-visual segmentation. arXiv preprint arXiv:2305.11019 (2023) +31. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) +32. Ma, J., Sun, P., Wang, Y., Hu, D.: Stepping stones: A progressive training strategy for audio-visual semantic segmentation. IEEE European Conference on Computer Vision (ECCV) (2024) +33. Majumder, S., Al-Halah, Z., Grauman, K.: Move2hear: Active audio-visual source separation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 275–285 (2021) + +34. Mo, S., Tian, Y.: Av-sam: Segment anything model meets audio-visual localization and segmentation. arXiv preprint arXiv:2305.01836 (2023) +35. Park, S., Senocak, A., Chung, J.S.: Marginnce: Robust sound localization with a negative margin. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023) +36. Senocak, A., Oh, T.H., Kim, J., Yang, M.H., Kweon, I.S.: Learning to localize sound source in visual scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4358-4366 (2018) +37. Sharma, H., Jalal, A.S.: Image captioning improved visual question answering. Multimedia tools and applications 81(24), 34775-34796 (2022) +38. Shi, Z., Zhou, X., Qiu, X., Zhu, X.: Improving image captioning with better use of captions. arXiv preprint arXiv:2006.11807 (2020) +39. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D., Blecher, L., Ferrer, C.C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I., Korenev, A., Koura, P.S., Lachaux, M.A., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A., Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X.E., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models (2023) +40. Wang, Y., Liu, W., Li, G., Ding, J., Hu, D., Li, X.: Prompting segmentation with sound is generalizable audio-visual source localizer. arXiv preprint arXiv:2309.07929 (2023) +41. Wang, Y., Sun, P., Zhou, D., Li, G., Zhang, H., Hu, D.: Ref-avs: Refer and segment objects in audio-visual scenes. IEEE European Conference on Computer Vision (ECCV) (2024) +42. Wu, W., Yao, H., Zhang, M., Song, Y., Ouyang, W., Wang, J.: Gpt4vis: What can gpt-4 do for zero-shot visual recognition? arXiv preprint arXiv:2311.15732 (2023) +43. Yan, S., Zhang, R., Guo, Z., Chen, W., Zhang, W., Li, H., Qiao, Y., He, Z., Gao, P.: Referred by multi-modality: A unified temporal transformer for video object segmentation. arXiv preprint arXiv:2305.16318 (2023) +44. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of Imms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9(1), 1 (2023) +45. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding (2023) +46. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127(3), 302-321 (2019) +47. Zhou, J., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., Zhong, Y.: Audio-visual segmentation. In: European Conference on Computer Vision (2022) +48. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020) +49. Zürn, J., Burgard, W.: Self-supervised moving vehicle detection from audio-visual cues (2022) \ No newline at end of file diff --git a/2024/Can Textual Semantics Mitigate Sounding Object Segmentation Preference_/images.zip b/2024/Can Textual Semantics Mitigate Sounding Object Segmentation Preference_/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..68776e44f033925332edd55a42a9ec1bc209f707 --- /dev/null +++ b/2024/Can Textual Semantics Mitigate Sounding Object Segmentation Preference_/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46fd1a37f9ebb534eaeb3566c33726dbc9d5961fefad45c993b105063150a1fe +size 473568 diff --git a/2024/Can Textual Semantics Mitigate Sounding Object Segmentation Preference_/layout.json b/2024/Can Textual Semantics Mitigate Sounding Object Segmentation Preference_/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1680a5a874bf761acb31ea5443617b58da56ca18 --- /dev/null +++ b/2024/Can Textual Semantics Mitigate Sounding Object Segmentation Preference_/layout.json @@ -0,0 +1,8763 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "type": "text", + "content": "Can Textual Semantics Mitigate Sounding Object Segmentation Preference?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 136, + 170, + 478, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 170, + 478, + 195 + ], + "spans": [ + { + "bbox": [ + 136, + 170, + 478, + 195 + ], + "type": "text", + "content": "Yaoting Wang" + }, + { + "bbox": [ + 136, + 170, + 478, + 195 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 136, + 170, + 478, + 195 + ], + "type": "text", + "content": ", Peiwen Sun" + }, + { + "bbox": [ + 136, + 170, + 478, + 195 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 136, + 170, + 478, + 195 + ], + "type": "text", + "content": ", Yuanchao Li" + }, + { + "bbox": [ + 136, + 170, + 478, + 195 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 136, + 170, + 478, + 195 + ], + "type": "text", + "content": ", Honggang Zhang" + }, + { + "bbox": [ + 136, + 170, + 478, + 195 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 136, + 170, + 478, + 195 + ], + "type": "text", + "content": ", and Di Hu" + }, + { + "bbox": [ + 136, + 170, + 478, + 195 + ], + "type": "inline_equation", + "content": "^{\\text{空}1,4}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 148, + 204, + 465, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 204, + 465, + 216 + ], + "spans": [ + { + "bbox": [ + 148, + 204, + 465, + 216 + ], + "type": "text", + "content": "1 Gaoling School of Artificial Intelligence, Renmin University of China, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 249, + 217, + 365, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 217, + 365, + 228 + ], + "spans": [ + { + "bbox": [ + 249, + 217, + 365, + 228 + ], + "type": "text", + "content": "yaoting.wang@outlook.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 270, + 228, + 343, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 228, + 343, + 237 + ], + "spans": [ + { + "bbox": [ + 270, + 228, + 343, + 237 + ], + "type": "text", + "content": "dihu@ruc.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 164, + 238, + 448, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 238, + 448, + 249 + ], + "spans": [ + { + "bbox": [ + 164, + 238, + 448, + 249 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 164, + 238, + 448, + 249 + ], + "type": "text", + "content": " Beijing University of Posts and Telecommunications, Beijing, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 241, + 250, + 373, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 250, + 373, + 260 + ], + "spans": [ + { + "bbox": [ + 241, + 250, + 373, + 260 + ], + "type": "text", + "content": "{sunpeiwen,zhhg}@bupt.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 223, + 260, + 389, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 260, + 389, + 270 + ], + "spans": [ + { + "bbox": [ + 223, + 260, + 389, + 270 + ], + "type": "text", + "content": "3 University of Edinburgh, Scotland, UK" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 258, + 272, + 356, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 272, + 356, + 281 + ], + "spans": [ + { + "bbox": [ + 258, + 272, + 356, + 281 + ], + "type": "text", + "content": "yuanchao.li@ed.ac.uk" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 281, + 470, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 281, + 470, + 293 + ], + "spans": [ + { + "bbox": [ + 143, + 281, + 470, + 293 + ], + "type": "text", + "content": "4 Engineering Research Center of Next-Generation Search and Recommendation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 159, + 347, + 455, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 347, + 455, + 590 + ], + "spans": [ + { + "bbox": [ + 159, + 347, + 455, + 590 + ], + "type": "text", + "content": "Abstract. The Audio-Visual Segmentation (AVS) task aims to segment sounding objects in the visual space using audio cues. However, in this work, it is recognized that previous AVS methods show a heavy reliance on detrimental segmentation preferences related to audible objects, rather than precise audio guidance. We argue that the primary reason is that audio lacks robust semantics compared to vision, especially in multi-source sounding scenes, resulting in weak audio guidance over the visual space. Motivated by the fact that text modality is well explored and contains rich abstract semantics, we propose leveraging text cues from the visual scene to enhance audio guidance with the semantics inherent in text. Our approach begins by obtaining scene descriptions through an off-the-shelf image captioner and prompting a frozen large language model to deduce potential sounding objects as text cues. Subsequently, we introduce a novel semantics-driven audio modeling module with a dynamic mask to integrate audio features with text cues, leading to representative sounding object features. These features not only encompass audio cues but also possess vivid semantics, providing clearer guidance in the visual space. Experimental results on AVS benchmarks validate that our method exhibits enhanced sensitivity to audio when aided by text cues, achieving highly competitive performance on all three subsets. Project page: https://github.com/GeWu-Lab/Sounding-Object-Segmentation-Preference" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 159, + 602, + 453, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 602, + 453, + 625 + ], + "spans": [ + { + "bbox": [ + 159, + 602, + 453, + 625 + ], + "type": "text", + "content": "Keywords: Audio-Visual Segmentation " + }, + { + "bbox": [ + 159, + 602, + 453, + 625 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 602, + 453, + 625 + ], + "type": "text", + "content": " Textual Guidance " + }, + { + "bbox": [ + 159, + 602, + 453, + 625 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 602, + 453, + 625 + ], + "type": "text", + "content": " Multi-modal learning" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 642, + 231, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 642, + 231, + 653 + ], + "spans": [ + { + "bbox": [ + 141, + 642, + 231, + 653 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 141, + 642, + 231, + 653 + ], + "type": "text", + "content": " Equal contribution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 653, + 244, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 653, + 244, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 653, + 244, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{\\Pi}^{\\text{三}}" + }, + { + "bbox": [ + 141, + 653, + 244, + 665 + ], + "type": "text", + "content": " Corresponding author." + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 114, + 462, + 306 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 462, + 306 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 462, + 306 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 462, + 306 + ], + "type": "image", + "image_path": "654b534457e0ff9452c91ebdf4e1d512b7c36414ac7bfb9b2ff7b9260d0c95d8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 315, + 482, + 392 + ], + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 392 + ], + "type": "text", + "content": "Fig. 1: The previous methods (left) achieve satisfactory results in receiving normal audio input. However, even when the sound is completely silent, they still segment vast pixels to represent the guitar regarding the segmentation preference of audible objects built during the training. In contrast, our approach (right) utilizes a frozen LLM to reason from scene descriptions that the male is not playing the guitar, as his hands are off the strings. Guided by this semantic information, our method produces more precise segmentation with finer audio guidance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 417, + 230, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 417, + 230, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 417, + 230, + 430 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 449, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 568 + ], + "type": "text", + "content": "In human perception, visual and auditory senses are closely related, and audio can provide vision with additional and dynamic scene information to enhance visual understanding. In the context of audio-visual understanding tasks, Audio-Visual Localization (AVL) enables the localization of sounding objects within visual scenes using audio references [15, 36]. However, the coarse localization of sounding objects is no longer sufficient to meet the practical demands of complex scene understanding, such as in autonomous driving [12, 49] and augmented reality [33]. To address this challenge, tasks are gradually shifting from bounding-box or rough heat-map localization to finer-grained pixel-level segmented localization, which is known as Audio-Visual Segmentation (AVS)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "The existing works on AVS can be broadly categorized into fusion-based methods [8, 16, 20, 21, 29, 47] and prompt-based methods [30, 34, 40]. The former primarily focuses on localizing sounding objects by fusing audio and visual features, while the latter emphasizes constructing effective audio prompts for the visual foundation model. However, most of these methods fail to establish compact audio-visual correlations to achieve clear audio guidance in the visual space. As depicted in the left of Fig. 1, previous methods could achieve satisfactory results with normal audio input. Nevertheless, when the audio is silent," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "these methods still segment many pixels to represent potential sounding objects based on the segmentation preferences of audible objects1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 284 + ], + "type": "text", + "content": "Furthermore, from the experiment with manually muted audio input, we observe that most of the behavior of AVS baselines depends on the segmentation preference established during the training process rather than relying on reliable audio-visual correlation. For instance, these models tend to take a shortcut by segmenting out the muted guitars because the audio guidance is weak, and they simply learned with the easy-to-learn visual feature alone (i.e., shortcut learning) that guitars are often associated with sound-emitting during the training phase. The unreliable audio-visual correlation can be attributed to two main factors. Firstly, the scarcity of training data for AVS is a significant challenge due to the demanding nature of pixel-level annotation. Secondly, the audio modality itself presents inherent complexity and ambiguity [16, 21, 29], especially in scenarios involving multiple audio sources that may be intertwined." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 285, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 548 + ], + "type": "text", + "content": "Given the challenges discussed above, our approach aims to build a robust audio-visual correlation by integrating the text modality, which inherently provides meaningful and robust semantic information about objects and their interactions. Firstly, in the Scene Describing phase, we utilize an off-the-shelf image captioner (LLaVA-1.5 [28]) to generate detailed dense captions for scene understanding. Then, in the Sounding Objects Reasoning phase, we instruct the frozen large language models (LLMs, e.g., LLaMA2 [39]) as a text cues capturer to collect potential sounding objects from the scene description as text cues. Notably, we introduce an elaborate few-shot prompt template to guide the text cues capturer to reason with Chain-of-Thought (CoT) instructions. After that, in the Semantics-Driven Audio Modeling (SeDAM) module, we project the audio feature into latent audio features. However, instead of directly using these latent features to interact with visual features for mask decoding, we leverage these latent features to interact with text cues. Specifically, the latent features are attended to and collected to form sounding object features with semantics provided by the text cues through the crossmodal transformer. We further propose the Prompting Mask Queries with Semantics (PMQS) module to introduce the sounding object features into pre-trained mask queries for later segmentation. Finally, we use simple but effective bottleneck adapters in the Audio-Prompted Decoding phase for better segmentation quality. The experimental results demonstrate that our proposed method improves audio-visual grounding by incorporating guidance from both audio and text cues." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 548, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 597 + ], + "type": "text", + "content": "Experiments in Sec. 4.4 reveal the advantage of our method, as verified by its increased sensitivity to the changes in audio input compared to networks without enhancing audio guidance with text cues. In summary, our contributions are threefold:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 609, + 482, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 609, + 482, + 634 + ], + "spans": [ + { + "bbox": [ + 138, + 609, + 482, + 634 + ], + "type": "text", + "content": "- We recognize the behavior of AVS model is influenced by the preference segmentation phenomenon, which can not be solved easily by previous audio-" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Sounding Object Segmentation Preference" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 482, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 482, + 666 + ], + "type": "text", + "content": "1 To be clear, we define audible objects as objects capable of producing sound, while sounding objects are defined as objects that are currently producing sound." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 147, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 480, + 140 + ], + "type": "text", + "content": "visual based methods. A novel strategy introduces textual semantics as the bridge into the AVS task, establishing a better audio-visual correlation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 479, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 479, + 187 + ], + "type": "text", + "content": "- We introduce the SeDAM module to form sounding object features with audio and text cues. We also suggest a PMQS module to prompt the pretrained mask queries with these sounding object features for sounding object segmentation using a visual foundation model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 189, + 479, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 189, + 479, + 212 + ], + "spans": [ + { + "bbox": [ + 138, + 189, + 479, + 212 + ], + "type": "text", + "content": "- Our experiments show that our method can be more sensitive to the audio guidance, manifested in its high sensitivity to the changes in audio input." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 232, + 242, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 232, + 242, + 244 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 242, + 244 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 258, + 384, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 258, + 384, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 258, + 384, + 270 + ], + "type": "text", + "content": "2.1 Audio-Visual Localization and Segmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 278, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 495 + ], + "type": "text", + "content": "Traditional AVL tasks [15,35,36] predict the positions of sounding objects using bounding-box or coarse heat-maps through unsupervised learning of audio-visual correlation. In recent years, driven by the growing demand for more precise localization in industries such as autonomous driving [12,49] and augmented reality [33], the AVL task has evolved towards a finer-grained AVS task that localizes sounding objects at a pixel-level. The existing AVS works can be broadly categorized into fusion-based [8, 16, 20, 21, 29, 43, 47] and prompt-based [30, 32, 40] The pioneering AVS work [47] is fusion-based and adopts a multi-stage strategy to integrate audio with multi-scale visual features. Liu et al. [29] address the issues of inadequate fusion of audio-visual features with an audio-aware query-enhanced transformer (AuTR). In comparison, AVSegFormer [8] directly decodes the fused feature with the audio query. Wang et al. [40] innovatively prompt the visual foundation model with audio input, harnessing the abundant visual before achieving generalizable AVS in zero-shot and few-shot scenarios. However, these methods have not effectively established a reliable audio-visual correlation, preventing the models from perceiving robust audio guidance. On the contrary, our experiments in Sec. 4.4 also reveal that some AVS baselines' competitive performance may stem from segmentation preferences formed during the training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 513, + 328, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 513, + 328, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 513, + 328, + 525 + ], + "type": "text", + "content": "2.2 Text Aided Scene Understanding" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 605 + ], + "type": "text", + "content": "Before the explosion of LLM, textual semantic information had been widely used for visual understanding tasks. Zhan et al. [38] and Sharma et al. [37] improve Visual Question Answering by using the extra semantic information of image captions, while Hur and Park [17] utilize the image captioner to help zero-shot image classification. Wang et al. [41] incorporating text with audio and visual cues to solve reference AVS task." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 606, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 481, + 666 + ], + "type": "text", + "content": "In recent years, the community has witnessed an emergent interest in strong off-the-shelf capabilities in open-world visual understanding [24-26, 45]. Notable examples, such as GPT4(V) [44] and LLaVA [27] have showcased remarkable linguistic and visual capabilities in zero-shot scene understanding [7, 42] without any training requirement. Innovatively, in the field of AVS, BAVS [23] replaces" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 144, + 114, + 470, + 259 + ], + "blocks": [ + { + "bbox": [ + 144, + 114, + 470, + 259 + ], + "lines": [ + { + "bbox": [ + 144, + 114, + 470, + 259 + ], + "spans": [ + { + "bbox": [ + 144, + 114, + 470, + 259 + ], + "type": "image", + "image_path": "3b7384491ed32f0e29fa98610f0ed5b3dcd66374660b90b82c87554356c724c4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 270, + 482, + 360 + ], + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 360 + ], + "type": "text", + "content": "Fig. 2: Overall pipeline of the proposed TeSO method. We utilize an off-the-shelf image captioner for dense scene describing and employ a frozen LLM to reason out potential sounding objects as text cues. These semantic text cues are then aggregated with audio features in the SeDAM module to form the sounding object features. Subsequently, we introduce the sounding object features into pre-trained mask queries in the PMQS module. Finally, we use adapters to tune the visual-only mask decoder for AVS in the Audio-prompted Decoding phase. \"MSDA\" is the multi-scale deformable attention proposed by Zhu et al. [48]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 381, + 482, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 482, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 482, + 478 + ], + "type": "text", + "content": "audio input with textual audio tags from a large pre-trained vision (UniDiffuser [2]) and audio (BEATs [3]) foundation model, establishing a reliable AVS system with foundation knowledge. However, due to the significant gap in class distribution between the pre-training dataset and the AVS dataset, as well as the presence of multi-source scenarios in the AVS task, the extracted audio tags are not ideal to provide accurate semantics. In contrast, we utilize the text modality with robust semantic features to enhance rather than replace the audio guidance and address the segmentation preference issue." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 496, + 413, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 496, + 413, + 511 + ], + "spans": [ + { + "bbox": [ + 131, + 496, + 413, + 511 + ], + "type": "text", + "content": "3 Text-guided Sounding Object Segmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "As shown in Fig. 2, we introduce our Text-guided Sounding Object Segmentation (TeSO) method in this section. Our experiments about muted audio input imply that audio does not exert reasonable guidance over the visual space for some AVS baselines, resulting in the models relying more on the segmentation preferences established during the training process. In essence, they learn which objects are most likely audible during training and take shortcuts during inference. To address the above issue, our method aims to enhance the audio-visual correlation by leveraging the text modality, which inherently possesses robust semantic information, to obtain finer-grained audio guidance. We begin by acquiring detailed scene descriptions through an off-the-shelf image captioner. Subsequently, a frozen LLM works as the text cues capturer, collecting potential sounding objects as text cues from the scene descriptions with CoT instructions." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Sounding Object Segmentation Preference" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": "Finally, we introduce a novel SeDAM module with a dynamic mask to seamlessly integrate audio features and text semantics through a crossmodal transformer, providing finer audio guidance with text semantic cues." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 169, + 301, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 169, + 301, + 182 + ], + "spans": [ + { + "bbox": [ + 131, + 169, + 301, + 182 + ], + "type": "text", + "content": "3.1 Multimodal Representation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 188, + 481, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 481, + 247 + ], + "type": "text", + "content": "Visual: Following previous works [40, 47], we sample frames from the video at 1-second intervals. Mask2Former [4] serves as our visual foundation model, and we extract visual features " + }, + { + "bbox": [ + 130, + 188, + 481, + 247 + ], + "type": "inline_equation", + "content": "F_V \\in \\mathbb{R}^{d_V \\times H \\times W}" + }, + { + "bbox": [ + 130, + 188, + 481, + 247 + ], + "type": "text", + "content": " from a pre-trained Swin transformer [6]. We use simple two-layer MLP adapters to tune the visual encoder for better visual representations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 248, + 481, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 248, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 131, + 248, + 481, + 319 + ], + "type": "text", + "content": "Audio: Similar to the video processing, we split the audio into clips at 1-second intervals. We extract audio features " + }, + { + "bbox": [ + 131, + 248, + 481, + 319 + ], + "type": "inline_equation", + "content": "F_{As} = \\{F_{A_1}, F_{A_2}, \\dots, F_{A_T}\\} \\in \\mathbb{R}^{T \\times d_A}" + }, + { + "bbox": [ + 131, + 248, + 481, + 319 + ], + "type": "text", + "content": " using VGGish [9, 13], where " + }, + { + "bbox": [ + 131, + 248, + 481, + 319 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 248, + 481, + 319 + ], + "type": "text", + "content": " represents the length of the audio in seconds, matching the number of video frames. Therefore, each video frame corresponds to " + }, + { + "bbox": [ + 131, + 248, + 481, + 319 + ], + "type": "inline_equation", + "content": "F_{A} = F_{As}[i]" + }, + { + "bbox": [ + 131, + 248, + 481, + 319 + ], + "type": "text", + "content": ". We freeze the pre-trained parameters for a better comparison with previous works." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 320, + 481, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 320, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 131, + 320, + 481, + 357 + ], + "type": "text", + "content": "Text: For each video frame, we complement the captured text cues into a sentence using a template (\"This is a " + }, + { + "bbox": [ + 131, + 320, + 481, + 357 + ], + "type": "inline_equation", + "content": "\\{\\_ \\}" + }, + { + "bbox": [ + 131, + 320, + 481, + 357 + ], + "type": "text", + "content": ".\") to extract text features " + }, + { + "bbox": [ + 131, + 320, + 481, + 357 + ], + "type": "inline_equation", + "content": "F_{T} \\in \\mathbb{R}^{N_{T} \\times d_{T}}" + }, + { + "bbox": [ + 131, + 320, + 481, + 357 + ], + "type": "text", + "content": " with ImageBind [10], where " + }, + { + "bbox": [ + 131, + 320, + 481, + 357 + ], + "type": "inline_equation", + "content": "N_{T}" + }, + { + "bbox": [ + 131, + 320, + 481, + 357 + ], + "type": "text", + "content": " is the number of text cues." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 373, + 272, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 373, + 272, + 384 + ], + "spans": [ + { + "bbox": [ + 131, + 373, + 272, + 384 + ], + "type": "text", + "content": "3.2 Text Cues Generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 392, + 481, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 481, + 440 + ], + "type": "text", + "content": "To obtain more precise text cues, we use a two-stage inference flow. Firstly, we employ an off-the-shelf image captioner to generate dense scene descriptions. Secondly, we ask a text cues capturer to collect potential sounding objects from the generated descriptions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 456, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 554 + ], + "type": "text", + "content": "Scene Describing. As shown in the top-left of Fig. 3, in this phase, we employ the commonly used and straightforward prompt: \"Please carefully understand the image content and provide as rich a description as possible.\" This instructs the image captioner to generate a detailed description of the visual scene, a task commonly referred to as dense captioning [18]. This process enables the image captioner to produce comprehensive and informative descriptions that not only describe the objects like \"a human and guitar in the image\" but also capture their interactions and relationships." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Sounding Object Collecting. To extract potential sounding objects from the generated dense descriptions, we design a task-specific CoT prompt template with few-shot demonstrations, as illustrated in Fig. 3. Specifically, we carefully craft examples that cover various scenarios and provide reasoning approaches for each example. The process begins by identifying the objects present in the dense scene descriptions, followed by step-by-step common-sense reasoning according to the CoT instructions, as depicted in the lower right of Fig. 3. During the CoT reasoning, the text cues capturer first categorizes the sample objects and" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 195, + 114, + 278, + 276 + ], + "blocks": [ + { + "bbox": [ + 195, + 114, + 278, + 276 + ], + "lines": [ + { + "bbox": [ + 195, + 114, + 278, + 276 + ], + "spans": [ + { + "bbox": [ + 195, + 114, + 278, + 276 + ], + "type": "image", + "image_path": "9b78b8aefc443ced04e43ae181a90ca207178e1f8a63d5cdb0ba66a8549c3e3e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 285, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 331 + ], + "type": "text", + "content": "Fig. 3: \"P.S.O\" stands for \"Potential Sounding Object\". A frozen LLM reasoner works as a text cues capturer by considering the interaction of audible objects. For instance, a person may sing a song while playing the guitar, but he would not sing along with a saxophone. In contrast, a noun parser simply captures any nouns that are present." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 278, + 113, + 419, + 277 + ], + "blocks": [ + { + "bbox": [ + 278, + 113, + 419, + 277 + ], + "lines": [ + { + "bbox": [ + 278, + 113, + 419, + 277 + ], + "spans": [ + { + "bbox": [ + 278, + 113, + 419, + 277 + ], + "type": "image", + "image_path": "9eeb3064ba0d82e908aa542331c861dda184f5a60a1efcf17b9011ca2e9e3a82.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 356, + 482, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 451 + ], + "type": "text", + "content": "assesses whether objects in that category are likely making sounds. Subsequently, the capturer analyzes which group of potential sounding objects can produce sounds simultaneously. Finally, the group with higher confidence is captured as the text cues of length " + }, + { + "bbox": [ + 130, + 356, + 482, + 451 + ], + "type": "inline_equation", + "content": "N_{T}" + }, + { + "bbox": [ + 130, + 356, + 482, + 451 + ], + "type": "text", + "content": " and encoded in " + }, + { + "bbox": [ + 130, + 356, + 482, + 451 + ], + "type": "inline_equation", + "content": "F_{T}" + }, + { + "bbox": [ + 130, + 356, + 482, + 451 + ], + "type": "text", + "content": " for further guidance. The right part of Fig. 3 also illustrates why we use the frozen LLM as a text cues capturer instead of using a naive noun parser. It is important to note that the use of LLM is solely for generating high-quality text cues, the LLM itself is not engaged in the training process for a fairness concern." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 472, + 330, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 330, + 484 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 330, + 484 + ], + "type": "text", + "content": "3.3 Semantic-Driven Audio Modeling" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 495, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 602 + ], + "type": "text", + "content": "After obtaining potential sounding objects, we can enhance the semantics of audio features with these text cues in this SeDAM module. To allow semantic text cues to capture different aspects of audio features " + }, + { + "bbox": [ + 130, + 495, + 482, + 602 + ], + "type": "inline_equation", + "content": "F_{A}" + }, + { + "bbox": [ + 130, + 495, + 482, + 602 + ], + "type": "text", + "content": ", we project " + }, + { + "bbox": [ + 130, + 495, + 482, + 602 + ], + "type": "inline_equation", + "content": "F_{A}" + }, + { + "bbox": [ + 130, + 495, + 482, + 602 + ], + "type": "text", + "content": " into distinct latent features " + }, + { + "bbox": [ + 130, + 495, + 482, + 602 + ], + "type": "inline_equation", + "content": "F_{A_{L}} = \\{a_{i}\\}_{i=1}^{N_{L}} \\in \\mathbb{R}^{N_{L} \\times d_{A_{L}}}" + }, + { + "bbox": [ + 130, + 495, + 482, + 602 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 495, + 482, + 602 + ], + "type": "inline_equation", + "content": "N_{L}" + }, + { + "bbox": [ + 130, + 495, + 482, + 602 + ], + "type": "text", + "content": " is the number of latent features. It is worth noting that, we incorporate an auxiliary loss term " + }, + { + "bbox": [ + 130, + 495, + 482, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{infoNCE}" + }, + { + "bbox": [ + 130, + 495, + 482, + 602 + ], + "type": "text", + "content": " aims at ensuring the distinction among latent audio features. This way, we project audio features into the latent space, where each unique " + }, + { + "bbox": [ + 130, + 495, + 482, + 602 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 130, + 495, + 482, + 602 + ], + "type": "text", + "content": " represents a latent audio feature to be captured by semantic text cues, forming clear and loud sounding object features with robust semantics for finer guidance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 603, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 603, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 603, + 482, + 639 + ], + "type": "text", + "content": "To incorporate the semantics of text cues, we employ text cues as the query to attend and combine the relevant latent features through crossmodal multi-head attention, formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 229, + 653, + 481, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 653, + 481, + 668 + ], + "spans": [ + { + "bbox": [ + 229, + 653, + 481, + 668 + ], + "type": "interline_equation", + "content": "F _ {\\widetilde {A} _ {L}} = \\operatorname {M u l t i H e a d} \\left(F _ {T}, F _ {A _ {L}}, F _ {A _ {L}}\\right), \\tag {1}", + "image_path": "ca12bedf993bafd595988de9aba3d0c227e019024b739e49c2b5a2cc5d3f4612.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Sounding Object Segmentation Preference" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "F_{\\widetilde{A}_L} \\in \\mathbb{R}^{N_T \\times d_A}" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": " is the attended latent audio features and will be combined with text cues " + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "F_T" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 244, + 152, + 481, + 168 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 152, + 481, + 168 + ], + "spans": [ + { + "bbox": [ + 244, + 152, + 481, + 168 + ], + "type": "interline_equation", + "content": "F _ {A _ {T}} = F C \\left(L N \\left(F _ {T} + F _ {\\widetilde {A} _ {L}}\\right)\\right), \\tag {2}", + "image_path": "a26b29364175e6dea8afc75901a1805754afc1cf1ae7c00c9acec127e8dc4788.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 173, + 480, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 173, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 173, + 480, + 268 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 173, + 480, + 268 + ], + "type": "inline_equation", + "content": "F_{A_T} \\in \\mathbb{R}^{N_T \\times d_A}" + }, + { + "bbox": [ + 130, + 173, + 480, + 268 + ], + "type": "text", + "content": " is the combined feature with both audio information and semantic text cues, and " + }, + { + "bbox": [ + 130, + 173, + 480, + 268 + ], + "type": "inline_equation", + "content": "FC" + }, + { + "bbox": [ + 130, + 173, + 480, + 268 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 173, + 480, + 268 + ], + "type": "inline_equation", + "content": "LN" + }, + { + "bbox": [ + 130, + 173, + 480, + 268 + ], + "type": "text", + "content": " for linear layer and layer normalization. Note that the LLM reasoner may generate a different number of text cues, and we either pad with a zero matrix or truncate to " + }, + { + "bbox": [ + 130, + 173, + 480, + 268 + ], + "type": "inline_equation", + "content": "N_T" + }, + { + "bbox": [ + 130, + 173, + 480, + 268 + ], + "type": "text", + "content": ". It is noteworthy that we propose task-specific CoT instructions to prompt the text cues capturer to infer more accurate potential sounding objects as the text cues. Additionally, we also introduce a dynamic mask based on attention score to mitigate the potential noisy information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 270, + 479, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 479, + 294 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 479, + 294 + ], + "type": "text", + "content": "We first compute the average pooling attention scores " + }, + { + "bbox": [ + 130, + 270, + 479, + 294 + ], + "type": "inline_equation", + "content": "\\zeta \\in \\mathbb{R}^{N_T}" + }, + { + "bbox": [ + 130, + 270, + 479, + 294 + ], + "type": "text", + "content": " for text cues on the attention head and key dimension:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 260, + 303, + 481, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 303, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 260, + 303, + 481, + 336 + ], + "type": "interline_equation", + "content": "\\xi = \\operatorname {P o o l} \\left(\\frac {F _ {T} F _ {A _ {L}} ^ {\\mathrm {T}}}{\\sqrt {d _ {A _ {L}}}}\\right), \\tag {3}", + "image_path": "c256ea34ce29dbac3054afd0db6dce2e95adc89fd99282bde34d561ec93aab7a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 346, + 479, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 479, + 371 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 479, + 371 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 346, + 479, + 371 + ], + "type": "inline_equation", + "content": "\\sqrt{d_{A_L}}" + }, + { + "bbox": [ + 130, + 346, + 479, + 371 + ], + "type": "text", + "content": " is the scaling factor for dot production. Then we form the attention mask " + }, + { + "bbox": [ + 130, + 346, + 479, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{m_1, m_2, m_3, \\dots, m_{N_T}\\}" + }, + { + "bbox": [ + 130, + 346, + 479, + 371 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 251, + 380, + 481, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 380, + 481, + 412 + ], + "spans": [ + { + "bbox": [ + 251, + 380, + 481, + 412 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {i} = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} \\xi_ {i} > 0 \\\\ - \\infty & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {4}", + "image_path": "78c70433b81ac59073410a89694ee4c950eb3100e05334d2ac651c7e7f92cd7d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 421, + 479, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 421, + 479, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 421, + 479, + 458 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 421, + 479, + 458 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\in \\mathbb{R}^{N_T}" + }, + { + "bbox": [ + 130, + 421, + 479, + 458 + ], + "type": "text", + "content": ", will be used as a key padding mask for the PMQS module in the next section to ensure that the attention weight of redundant text cues approaches zero." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 476, + 367, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 476, + 367, + 489 + ], + "spans": [ + { + "bbox": [ + 131, + 476, + 367, + 489 + ], + "type": "text", + "content": "3.4 Prompting Mask Queries with Semantics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 498, + 479, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 479, + 570 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 479, + 570 + ], + "type": "text", + "content": "Current segmentation foundation models [4, 5, 19] adopt transformer-based decoder with mask queries as the mainstream approach. To prompt such foundation models, we propose the PMQS module, aiming to introduce semantic information to mask query features " + }, + { + "bbox": [ + 131, + 498, + 479, + 570 + ], + "type": "inline_equation", + "content": "F_{Q} = \\{q_{1}, q_{2}, \\dots, q_{N_{Q}}\\}" + }, + { + "bbox": [ + 131, + 498, + 479, + 570 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 498, + 479, + 570 + ], + "type": "inline_equation", + "content": "F_{Q} \\in \\mathbb{R}^{N_{Q} \\times d_{V}}" + }, + { + "bbox": [ + 131, + 498, + 479, + 570 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 498, + 479, + 570 + ], + "type": "inline_equation", + "content": "N_{Q}" + }, + { + "bbox": [ + 131, + 498, + 479, + 570 + ], + "type": "text", + "content": " is the number of mask queries in crossmodal attention. We formulate this updated expression as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 190, + 579, + 481, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 579, + 481, + 612 + ], + "spans": [ + { + "bbox": [ + 190, + 579, + 481, + 612 + ], + "type": "interline_equation", + "content": "\\phi_ {c a} \\left(F _ {Q}, F _ {A _ {T}}, F _ {A _ {T}}\\right) = S o f t m a x \\left(\\frac {F _ {Q} F _ {A _ {T}} ^ {\\mathrm {T}}}{\\sqrt {d _ {A _ {T}}}} + \\mathcal {M}\\right) F _ {A _ {T}}, \\tag {5}", + "image_path": "84d51aa459920434b2a7f87af8cd3290f3d2f75a15c197b2cecdf20469800efc.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 616, + 479, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 616, + 479, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 616, + 479, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 616, + 479, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 616, + 479, + 641 + ], + "type": "text", + "content": " is added to the attention score to mask the redundant text cues for each query. Then we obtain the updated mask query features " + }, + { + "bbox": [ + 130, + 616, + 479, + 641 + ], + "type": "inline_equation", + "content": "\\widetilde{F}_Q\\in \\mathbb{R}^{N_Q\\times d_V}" + }, + { + "bbox": [ + 130, + 616, + 479, + 641 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 239, + 652, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 652, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 239, + 652, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\widetilde {F} _ {Q} = F _ {Q} + \\phi_ {c a} \\left(F _ {Q}, F _ {A _ {T}}, F _ {A _ {T}}\\right). \\tag {6}", + "image_path": "094f9fbaa96d0ae2407f06e73b30a2d91207b1002bba9d83b051ef7d1af4e2aa.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 296, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 296, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 296, + 129 + ], + "type": "text", + "content": "3.5 Audio-Prompted Decoding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 135, + 479, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 135, + 479, + 196 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 479, + 196 + ], + "type": "text", + "content": "As shown on the right side of Fig. 2, the updated mask query features " + }, + { + "bbox": [ + 130, + 135, + 479, + 196 + ], + "type": "inline_equation", + "content": "\\widetilde{F}_Q" + }, + { + "bbox": [ + 130, + 135, + 479, + 196 + ], + "type": "text", + "content": " are forwarded into the transformer-based decoder, generating mask queries " + }, + { + "bbox": [ + 130, + 135, + 479, + 196 + ], + "type": "inline_equation", + "content": "M_Q \\in \\mathbb{R}^{N_Q \\times H \\times W}" + }, + { + "bbox": [ + 130, + 135, + 479, + 196 + ], + "type": "text", + "content": " and class queries " + }, + { + "bbox": [ + 130, + 135, + 479, + 196 + ], + "type": "inline_equation", + "content": "C_Q \\in \\mathbb{R}^{N_Q \\times N_C}" + }, + { + "bbox": [ + 130, + 135, + 479, + 196 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 135, + 479, + 196 + ], + "type": "inline_equation", + "content": "N_C" + }, + { + "bbox": [ + 130, + 135, + 479, + 196 + ], + "type": "text", + "content": " is the number of class queries pre-defined in the segmentation foundation model. Subsequently, we can obtain the final segmentation " + }, + { + "bbox": [ + 130, + 135, + 479, + 196 + ], + "type": "inline_equation", + "content": "M_{pred} \\in \\mathbb{R}^{N_C \\times H \\times W}" + }, + { + "bbox": [ + 130, + 135, + 479, + 196 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 249, + 206, + 481, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 206, + 481, + 220 + ], + "spans": [ + { + "bbox": [ + 249, + 206, + 481, + 220 + ], + "type": "interline_equation", + "content": "M _ {p r e d} = \\operatorname {E i n s u m} \\left(C _ {Q}, \\bar {F} _ {Q}\\right), \\tag {7}", + "image_path": "e795bfa5ac5d8801169f741ba181d6e519a9f1e1c888f62abf99735dd0c50e16.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 228, + 385, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 228, + 385, + 240 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 385, + 240 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 228, + 385, + 240 + ], + "type": "inline_equation", + "content": "\\text{Einsum}(\\cdot)" + }, + { + "bbox": [ + 131, + 228, + 385, + 240 + ], + "type": "text", + "content": " is the Einstein Summation Convention." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 241, + 479, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 241, + 479, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 479, + 289 + ], + "type": "text", + "content": "Notably, to enhance the quality of the segmentation, we use two-layer bottle-neck MLP adapters [14] to lightly fine-tune the output projection of the following attentions: a) Swin Window Attention (W-MSA) [31], b) Deformable Attention in pixel decoder [48] and c) Masked Attention in transformer decoder [4]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 306, + 262, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 306, + 262, + 319 + ], + "spans": [ + { + "bbox": [ + 131, + 306, + 262, + 319 + ], + "type": "text", + "content": "3.6 Learning Objectives" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 326, + 479, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 326, + 479, + 350 + ], + "spans": [ + { + "bbox": [ + 131, + 326, + 479, + 350 + ], + "type": "text", + "content": "Segmentation Loss. During the model training process, we employ the binary cross-entropy loss and dice loss to optimize the mask quality:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 233, + 361, + 480, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 361, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 233, + 361, + 480, + 372 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {b c e} \\cdot \\mathcal {L} _ {b c e} + \\lambda_ {d i c e} \\cdot \\mathcal {L} _ {d i c e}. \\tag {8}", + "image_path": "2dd5b062aa48516b2420532cd02a25fa61a00116406a802f883e58a7355c0f09.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 381, + 479, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 381, + 479, + 403 + ], + "spans": [ + { + "bbox": [ + 131, + 381, + 479, + 403 + ], + "type": "text", + "content": "Further, we adopt the mask classification loss [5] to compose the final segmentation loss:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 249, + 407, + 480, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 407, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 249, + 407, + 480, + 418 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s e g} = \\mathcal {L} _ {m a s k} + \\lambda_ {c l s} \\cdot \\mathcal {L} _ {c l s}. \\tag {9}", + "image_path": "13a10e3b24a6627db70ca63929bbdaa25c64bdb985f8795e357708e4b7269fc9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 423, + 479, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 423, + 479, + 460 + ], + "spans": [ + { + "bbox": [ + 131, + 423, + 479, + 460 + ], + "type": "text", + "content": "Latent Component Loss. This loss term aims to explicitly guide the model to learn distinct representations for each latent audio component, helping prevent redundancy and overlapping information across the latent components." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 211, + 468, + 480, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 211, + 468, + 480, + 500 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {i n f o N C E}} = - \\log \\left(\\frac {\\exp \\left(\\operatorname {s i m} \\left(a , a ^ {+}\\right)\\right)}{\\sum_ {i = 1} ^ {N _ {L}} \\exp \\left(\\operatorname {s i m} \\left(a , a _ {i}\\right)\\right)}\\right), \\tag {10}", + "image_path": "6c42cade68d89e2ef64cab711c982d3dd3c3641f6d8af3331c173656310650f5.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 508, + 462, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 508, + 462, + 520 + ], + "spans": [ + { + "bbox": [ + 131, + 508, + 462, + 520 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 508, + 462, + 520 + ], + "type": "inline_equation", + "content": "a^+" + }, + { + "bbox": [ + 131, + 508, + 462, + 520 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 508, + 462, + 520 + ], + "type": "inline_equation", + "content": "sim(\\cdot)" + }, + { + "bbox": [ + 131, + 508, + 462, + 520 + ], + "type": "text", + "content": " denote the positive example and the cosine similarity." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 521, + 479, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 521, + 479, + 543 + ], + "spans": [ + { + "bbox": [ + 131, + 521, + 479, + 543 + ], + "type": "text", + "content": "Total Loss The final loss function is composed of the weighted sum of the two kinds of losses above:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 236, + 555, + 480, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 555, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 236, + 555, + 480, + 567 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {s e g}} + \\lambda_ {\\text {i n f o N C E}} \\cdot \\mathcal {L} _ {\\text {i n f o N C E}}. \\tag {11}", + "image_path": "918bde25b988d91ffad7f894cea9f03303277e7cd81cc8cb5a7d9f1bd0567b8e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 584, + 230, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 584, + 230, + 597 + ], + "spans": [ + { + "bbox": [ + 131, + 584, + 230, + 597 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 131, + 609, + 279, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 279, + 621 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 279, + 621 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "content": "Dataset. Our proposed method is evaluated on the AVS Benchmarks [47], which contains three subsets. Firstly, the single-source subset (V1S) contains 4932 videos over 23 categories, covering various sounds, such as humans, animals," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "text", + "content": "Sounding Object Segmentation Preference" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 199 + ], + "type": "text", + "content": "vehicles, and musical instruments. For videos in the training split of this subset, only the first sampled frame is annotated. Secondly, the multi-source subset (V1M) contains 424 videos that include two or more categories from V1S, and all sounding objects are visible in the frames. Finally, the AVSBench-semantic (AVSS) subset, an extension of V1S and V1M, contains 12,356 videos. These newly collected videos are trimmed to 10 seconds, differing from the 5 seconds in V1S and V1M." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 283 + ], + "type": "text", + "content": "Setting. We conduct training and evaluation using the VGGish backbone pretrained on Youtube-8M [1] and Swin-base Transformer backbone pretrained on semantic-ADE20K [46] by Mask2Former [4]. The number " + }, + { + "bbox": [ + 130, + 200, + 482, + 283 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 200, + 482, + 283 + ], + "type": "text", + "content": " of the crossmodal transformer layers is set to 4, and the parameters " + }, + { + "bbox": [ + 130, + 200, + 482, + 283 + ], + "type": "inline_equation", + "content": "\\lambda_{bce},\\lambda_{dice},\\lambda_{cls},\\lambda_{infoNCE}" + }, + { + "bbox": [ + 130, + 200, + 482, + 283 + ], + "type": "text", + "content": " in the loss are set to 5,5,2,1. The AdamW optimizer is adopted with a learning rate of 1e-4 for the visual encoder adapters and 1e-3 for other learnable parameters, and the training epoch is roughly set to 60." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 284, + 482, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 321 + ], + "type": "text", + "content": "Metrics. To conduct a comprehensive evaluation of our model, we carry out tests using mean Intersection over Union (mIoU) and F-score as the performance metrics, following previous works [8, 47]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 336, + 263, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 336, + 263, + 348 + ], + "spans": [ + { + "bbox": [ + 131, + 336, + 263, + 348 + ], + "type": "text", + "content": "4.2 Comparison Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 354, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 450 + ], + "type": "text", + "content": "When compared to methods that do not incorporate text as additional semantic information, our method, which leverages captioning and reasoning, demonstrates strong competitiveness. As illustrated in Sec. 4.2, our method achieves comparable results across all V1S, V1M, and AVSS-binary subsets, showcasing average performance improvements of up to " + }, + { + "bbox": [ + 130, + 354, + 482, + 450 + ], + "type": "inline_equation", + "content": "1.25\\%" + }, + { + "bbox": [ + 130, + 354, + 482, + 450 + ], + "type": "text", + "content": " in mIoU and " + }, + { + "bbox": [ + 130, + 354, + 482, + 450 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 130, + 354, + 482, + 450 + ], + "type": "text", + "content": " in F-score, respectively. Furthermore, qualitative analysis indicates that our predicted masks exhibit superior quality, as demonstrated in sub-figure (1) of Fig. 4 (more examples in supplementary materials)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 451, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 521 + ], + "type": "text", + "content": "It is necessary to mention that our proposed TeSO does not show significant improvements on V1S over certain existing methods, such as BAVS [23], which also utilizes text semantic information. Apart from marginal effects, this phenomenon can also be attributed to the singular sound source and simplistic nature of the data scenarios present in V1S. On V1M, however, our performance significantly surpasses that of other methods, including BAVS." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 522, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 582 + ], + "type": "text", + "content": "In addition to the AVS task, we also conduct experiments on the Audio-Visual Semantic Segmentation (AVSS) task, as shown in Tab. 2. Our model exceeds the best performance in previous models by " + }, + { + "bbox": [ + 130, + 522, + 482, + 582 + ], + "type": "inline_equation", + "content": "5.37\\%" + }, + { + "bbox": [ + 130, + 522, + 482, + 582 + ], + "type": "text", + "content": " on mIoU, which further demonstrates that our model is capable of establishing a better audio-visual correlation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 598, + 286, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 598, + 286, + 609 + ], + "spans": [ + { + "bbox": [ + 131, + 598, + 286, + 609 + ], + "type": "text", + "content": "4.3 Bottlenecks and Fairness" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "Our method (TeSO) collects text cues from dense image descriptions at " + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{REC}_{\\#1}" + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "inline_equation", + "content": "(80.41\\%)" + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": " and a filter with a dynamic mask at " + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{REC}_{\\#2}" + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "inline_equation", + "content": "(80.95\\%)" + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": ", indicating a robust startup for the text-guided AVS framework, in comparison with another two-stage model, BAVS." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 156, + 473, + 288 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 474, + 150 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 474, + 150 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 474, + 150 + ], + "type": "text", + "content": "Table 1: Performance on AVS-Benchmarks. The underscore is used to indicate suboptimal results. BAVS uses large pre-trained visual and audio foundation models to replace audio input with text tag. In comparison, TeSO (ours) enhances rather than replaces audio guidance based on the robust semantics of text. It is evident that our method has a good F-score result while ensuring mIoU." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 156, + 473, + 288 + ], + "lines": [ + { + "bbox": [ + 138, + 156, + 473, + 288 + ], + "spans": [ + { + "bbox": [ + 138, + 156, + 473, + 288 + ], + "type": "table", + "html": "
MethodAudio-backboneVisual-backboneV1SV1MAVSS-binary
mIoU(%)F-scoremIoU(%)F-scoremIoU(%)F-score
AVSBench [47]VGGishPVT-v278.700.87954.000.64562.450.756
AVSegFormer [8]VGGishPVT-v282.060.89958.360.69364.340.759
AVSC [22]VGGishPVT-v281.290.88659.500.657--
AVS-BG [11]VGGishPVT-v281.710.90455.100.668--
AQFormer [16]VGGishPVT-v281.600.89461.100.721--
†CATR [20]VGGishPVT-v281.400.89659.000.700--
AV-SAM [34]ResNet18ViT-Base40.470.566----
SAMA [30]VGGishViT-Huge81.530.88663.140.691--
GAVS [40]VGGishViT-Base80.060.90263.700.77467.700.788
AuTR [29]VGGishSwin-Base80.400.89156.200.672--
MUTR [43]VGGishVideo-Swin-Base81.600.89764.000.735--
BAVS [23]‡BEATsSwin-Base82.680.89859.630.65955.450.640
TeSO (ours)VGGishSwin-base83.270.93366.020.80168.530.813
", + "image_path": "fc507692e3d3bdf620ed78ac8dca7706b0dbbf5a718d386e804495a857c21969.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 143, + 289, + 472, + 307 + ], + "lines": [ + { + "bbox": [ + 143, + 289, + 472, + 307 + ], + "spans": [ + { + "bbox": [ + 143, + 289, + 472, + 307 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 143, + 289, + 472, + 307 + ], + "type": "text", + "content": " : To make a fair comparison, the results of CATR here are without supplemented annotation of the training set." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 143, + 307, + 471, + 324 + ], + "lines": [ + { + "bbox": [ + 143, + 307, + 471, + 324 + ], + "spans": [ + { + "bbox": [ + 143, + 307, + 471, + 324 + ], + "type": "text", + "content": ": BEATs is a stronger audio backbone, trained on AudioSet dataset with 90M parameters. In comparison, VGGish is trained on Youtube-8M with 70M parameters." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 139, + 400, + 288, + 461 + ], + "blocks": [ + { + "bbox": [ + 131, + 336, + 298, + 392 + ], + "lines": [ + { + "bbox": [ + 131, + 336, + 298, + 392 + ], + "spans": [ + { + "bbox": [ + 131, + 336, + 298, + 392 + ], + "type": "text", + "content": "Table 2: Performance on AVSS dataset. Our method showcases a substantial improvement in the semantic subset, achieved by leveraging robust text semantics to enhance audio guidance." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 139, + 400, + 288, + 461 + ], + "lines": [ + { + "bbox": [ + 139, + 400, + 288, + 461 + ], + "spans": [ + { + "bbox": [ + 139, + 400, + 288, + 461 + ], + "type": "table", + "html": "
MethodBackboneAVSS
mIoU(%)F-Score
AVSBenchPVT-v229.800.352
CATRPVT-v232.800.385
BAVSSwin-base33.590.375
TeSO (Ours)Swin-base38.960.451
", + "image_path": "0a21a4be6dabfc189cea795ffc70a0c44acd6ef62119d25db1c2cb6f230ec6e2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 321, + 400, + 473, + 461 + ], + "blocks": [ + { + "bbox": [ + 314, + 336, + 482, + 391 + ], + "lines": [ + { + "bbox": [ + 314, + 336, + 482, + 391 + ], + "spans": [ + { + "bbox": [ + 314, + 336, + 482, + 391 + ], + "type": "text", + "content": "Table 3: Potential sounding objects recognition (REC) performance at different model stages and the corresponding segmentation results. Experiments are conducted on AVS-V1M test set." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 400, + 473, + 461 + ], + "lines": [ + { + "bbox": [ + 321, + 400, + 473, + 461 + ], + "spans": [ + { + "bbox": [ + 321, + 400, + 473, + 461 + ], + "type": "table", + "html": "
MethodREC(#1)AP(%)REC(#2)AP(%)SEGmIoU(%)
BAVS24.2279.6959.63
TeSOp(ours)77.9678.3363.87
TeSOn(ours)77.8278.2763.31
TeSO(ours)80.4180.9566.02
", + "image_path": "d6459ea934d376eb43a55aea806421fa31f72c862f0b7f615b89540294d8c873.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 485, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 481, + 605 + ], + "type": "text", + "content": "In addition, we develop several variants of TeSO to investigate its performance sources and bottlenecks, as well as its fairness in comparison with other methods. " + }, + { + "bbox": [ + 130, + 485, + 481, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{TeSO}_p" + }, + { + "bbox": [ + 130, + 485, + 481, + 605 + ], + "type": "text", + "content": " denotes the variant using one-shot prompts, while " + }, + { + "bbox": [ + 130, + 485, + 481, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{TeSO}_n" + }, + { + "bbox": [ + 130, + 485, + 481, + 605 + ], + "type": "text", + "content": " denotes the variant employing a noun parser instead of a frozen LLM for text cues collection. As shown in Tab. 3, the results demonstrate that both the number (shots) of demonstrations and the method of text cues collection significantly impact the model's ability to accurately identify audio semantics and perform segmentation (SEG). Thus, utilizing few-shot prompts and off-the-shelf foundation models are both essential for ensuring the acquisition of high-quality text cues (i.e., the bottleneck)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "However, we would like to emphasize that, even without using LLM to capture text cues, our method variant " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "(\\mathrm{TeSO}_n)" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": " still fairly achieves competitive performance (63.31% mIoU) compared with BAVS (59.63% mIoU) and state-of-the-art method GAVS (63.70% mIoU) in the complex multi-source scenario (AVS-V1M), and significantly outperforms other methods. This suggests that" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Sounding Object Segmentation Preference" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 146, + 190, + 465, + 242 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 4: Impact on mIoU and F-score by muted audio or noise-only audio on AVS-V1M test set. TeSO (ours) exhibits the most intensive attenuation up to " + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "48.8\\%" + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": ". It is noteworthy that the other models still generate \"satisfactory\" results even in silent or purely noisy environments. This observation suggests that the models might rely more on the segmentation preferences formed during the training rather than truly learning efficient audio guidance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 146, + 190, + 465, + 242 + ], + "lines": [ + { + "bbox": [ + 146, + 190, + 465, + 242 + ], + "spans": [ + { + "bbox": [ + 146, + 190, + 465, + 242 + ], + "type": "table", + "html": "
MethodMuteWGN-10dBWGN-40dB
mIoU% (Δm%)F-score (Δf%)mIoU% (Δm%)F-score (Δf%)mIoU% (Δm%)F-score (Δf%)mIoU% (Δm%)
AVS-Bench49.15 (6.6)0.622 (4.9)46.76 (11.1)0.618 (5.5)47.46 (9.8)0.610 (6.7)
GAVS42.53 (29.2)0.695 (8.4)41.51 (34.8)0.692 (10.6)41.87 (30.3)0.696 (8.4)
BAVS45.06 (26.1)0.644 (5.8)43.27 (27.4)0.601 (8.8)42.93 (31.6)0.595 (9.7)
TeSO (ours)37.89 (42.6)0.671 (16.2)34.20 (49.6)0.661 (17.5)34.74 (47.4)0.656 (18.1)
", + "image_path": "73bdf3b3cef44a48d43bb36ffc69215e1df2f2266cc88f1ce07113188e464339.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 262, + 481, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 481, + 285 + ], + "type": "text", + "content": "leveraging text-enhanced audio for visual control with robust semantics is a viable and promising direction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 301, + 234, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 301, + 234, + 313 + ], + "spans": [ + { + "bbox": [ + 131, + 301, + 234, + 313 + ], + "type": "text", + "content": "4.4 Audio Control" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 319, + 481, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 481, + 415 + ], + "type": "text", + "content": "Based on our observations, we notice a distinct shape of the predicted mask in sub-figure (2) of Fig. 4 when some popular AVS baselines are used for inference. Interestingly, this shape remains distinguishable even when the inference is conducted with muted audio. This suggests that the network can predict rough masks without relying on audio input, indicating that the audio-visual grounding is not effectively guided by audio cues. As a result, this undesirable phenomenon can be attributed to the segmentation preference, which is a result of these models lacking proper audio control." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 416, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 482, + 596 + ], + "type": "text", + "content": "To evaluate the influence of audio control on segmentation, various models are subjected to inference using muted audio and noise-only audio (at 10 dB and 40 dB) instead of the corresponding audio signals. It should be noted that the visual features remained unchanged throughout the evaluation. As shown in Tab. 4, the disparity " + }, + { + "bbox": [ + 130, + 416, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\Delta_{m}" + }, + { + "bbox": [ + 130, + 416, + 482, + 596 + ], + "type": "text", + "content": " (and " + }, + { + "bbox": [ + 130, + 416, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\Delta_{f}" + }, + { + "bbox": [ + 130, + 416, + 482, + 596 + ], + "type": "text", + "content": ") refers to the difference in mIoU (and F-score) values before and after the audio substitution is applied as an evaluation metric for the sensitivity of audio changes. A larger drop in mIoU and F-score indicates a greater degree of audio control over the visual space. In addition to demonstrating strong segmentation capabilities in the AVS task, TeSO (ours) exhibits the highest level of attenuation, reaching up to " + }, + { + "bbox": [ + 130, + 416, + 482, + 596 + ], + "type": "inline_equation", + "content": "47.4\\%" + }, + { + "bbox": [ + 130, + 416, + 482, + 596 + ], + "type": "text", + "content": ", when subjected to muted and noise-only audio. Additionally, qualitative analysis revealed that the predicted masks remained predominantly blank, as depicted in sub-figure (2) of Fig. 4 (more examples in supplementary materials). This observation indicates that our model has better audio guidance over the visual space, enabling it to play a crucial role in audio-visual grounding." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 611, + 239, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 611, + 239, + 624 + ], + "spans": [ + { + "bbox": [ + 131, + 611, + 239, + 624 + ], + "type": "text", + "content": "4.5 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "As shown in Tab. 5, we conduct ablation experiments targeting the designed TeSO modules and the guidance from different modalities. To explore the guiding ability, we progressively remove either the audio or text guidance for TeSO." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 167, + 114, + 445, + 287 + ], + "blocks": [ + { + "bbox": [ + 167, + 114, + 445, + 287 + ], + "lines": [ + { + "bbox": [ + 167, + 114, + 445, + 287 + ], + "spans": [ + { + "bbox": [ + 167, + 114, + 445, + 287 + ], + "type": "image", + "image_path": "bf36008771978a66a07336716644a2120701f1bf9601010fb270afffcc490b7a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 297, + 482, + 342 + ], + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 342 + ], + "type": "text", + "content": "Fig. 4: Examples of the impact of normal audio input and all-mute audio on popular methods. In normal scenarios, our method shows better masks than previous methods. In all-mute scenarios, our approach exhibits strong sensitivity towards audio inputs, as it is capable of generating blank masks for silent audio clips." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 369, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 453 + ], + "type": "text", + "content": "The experiments indicate that control from both audio and text inputs for visual space can establish a strong audio-visual correlation. Furthermore, by utilizing our proposed TeSO, we achieve enhanced audio guidance and reach a " + }, + { + "bbox": [ + 130, + 369, + 482, + 453 + ], + "type": "inline_equation", + "content": "66.02\\%" + }, + { + "bbox": [ + 130, + 369, + 482, + 453 + ], + "type": "text", + "content": " mIoU while incorporating precise semantic information from the text modality. Moreover, through the utilization of our proposed TeSO method that incorporates robust semantic information from the text modality, we achieved improved audio guidance and attained a mIoU of " + }, + { + "bbox": [ + 130, + 369, + 482, + 453 + ], + "type": "inline_equation", + "content": "66.02\\%" + }, + { + "bbox": [ + 130, + 369, + 482, + 453 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 454, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 454, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 454, + 482, + 502 + ], + "type": "text", + "content": "As shown in Tab. 6, we also implement a baseline that directly filters the inferred semantic labels during the model evaluation process by intersecting them with the potential sounding object collected by the frozen LLM capturer. This implies that no text cues are utilized to enhance the audio guidance for better" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 149, + 581, + 272, + 664 + ], + "blocks": [ + { + "bbox": [ + 132, + 526, + 289, + 571 + ], + "lines": [ + { + "bbox": [ + 132, + 526, + 289, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 289, + 571 + ], + "type": "text", + "content": "Table 5: Ablation study on AVS-V1M. Our method benefits from using both audio and text as the guiding information for sounding objects." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 581, + 272, + 664 + ], + "lines": [ + { + "bbox": [ + 149, + 581, + 272, + 664 + ], + "spans": [ + { + "bbox": [ + 149, + 581, + 272, + 664 + ], + "type": "table", + "html": "
ControlmIoU(%)F-score
TeSO66.020.801
(-) LinfoNCE65.340.793
(-) Dynamic Mask64.780.781
(-) PMQS62.900.774
(-) SeDAM62.220.760
(-) Text62.280.768
(-) Audio42.110.692
(-) Text & Audio37.350.656
", + "image_path": "0be921ca0e094cf0aa9ead2e0d306f75dcf676a3f972b67f753205c4bf274e88.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 310, + 602, + 474, + 642 + ], + "blocks": [ + { + "bbox": [ + 304, + 526, + 482, + 592 + ], + "lines": [ + { + "bbox": [ + 304, + 526, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 526, + 482, + 592 + ], + "type": "text", + "content": "Table 6: Comparison of our proposed baseline and TeSO. The baseline method employs LLM to filter the predicted results but does not enhance the audio guidance with text cues. Experiments are conducted on AVS-V1M test set." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 602, + 474, + 642 + ], + "lines": [ + { + "bbox": [ + 310, + 602, + 474, + 642 + ], + "spans": [ + { + "bbox": [ + 310, + 602, + 474, + 642 + ], + "type": "table", + "html": "
Normal mIoU%Mute mIoU% (Δm%)WGN-40dB mIoU% (Δm%)
Baseline64.1543.37 (32.4)43.09 (32.8)
TeSO66.0237.89 (42.6)34.74 (47.4)
", + "image_path": "15822cf2b8f68ad588b581b5fe3fcf80ae8cf19d73a2780cb991ff503d55adaa.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 275, + 91, + 447, + 102 + ], + "type": "text", + "content": "Sounding Object Segmentation Preference" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "content": "control over the visual space. As expected, the sensitivity " + }, + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "inline_equation", + "content": "(\\Delta_{m})" + }, + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "content": " of model performance to changes in audio input has decreased compared with TeSO, indicating the effectiveness of our method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 175, + 217, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 175, + 217, + 187 + ], + "spans": [ + { + "bbox": [ + 132, + 175, + 217, + 187 + ], + "type": "text", + "content": "5 Discussion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 204, + 481, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 204, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 204, + 481, + 300 + ], + "type": "text", + "content": "Noisy text cues. Besides dense image captions for collecting semantic text cues, we also constructed a dynamic mask in Eq. (4) to further select effective semantics based on audio correspondence. We use audio recognition results to determine the accuracy of text cues. Please refer to Tab. 3 to compare the audio recognition results based on text cues collection stage " + }, + { + "bbox": [ + 132, + 204, + 481, + 300 + ], + "type": "inline_equation", + "content": "(\\mathrm{REC}_{\\#1})" + }, + { + "bbox": [ + 132, + 204, + 481, + 300 + ], + "type": "text", + "content": " and dynamic mask filtering stage " + }, + { + "bbox": [ + 132, + 204, + 481, + 300 + ], + "type": "inline_equation", + "content": "(\\mathrm{REC}_{\\#2})" + }, + { + "bbox": [ + 132, + 204, + 481, + 300 + ], + "type": "text", + "content": ". Our approach achieves higher audio recognition performance, compared with using the audio foundation model in BAVS. This indicates the reliability and accuracy of our strategy for obtaining text cues." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 301, + 481, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 301, + 481, + 372 + ], + "spans": [ + { + "bbox": [ + 132, + 301, + 481, + 372 + ], + "type": "text", + "content": "Prompt with CoT. As shown in the Fig. 3, we utilize task-specific prompt templates with multiple scenario demonstrations and CoT instructions to assist the frozen LLM in collecting more accurate potential sounding objects as text cues. For the detailed prompt templates with CoT instructions and related experimental information, please refer to the Prompt Template part and Text Assistance part in the supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 373, + 481, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 373, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 373, + 481, + 445 + ], + "type": "text", + "content": "Versatility. We explore an alternative visual foundation model like the Segment Anything Model [19] for segmentation. Conducting experiments using a ViT-based backbone with minimal adjustments, our model achieves an impressive " + }, + { + "bbox": [ + 132, + 373, + 481, + 445 + ], + "type": "inline_equation", + "content": "65.36\\%" + }, + { + "bbox": [ + 132, + 373, + 481, + 445 + ], + "type": "text", + "content": " mIoU on V1M, still getting comparable performance. This demonstrates the versatility and effectiveness of our model. Additional experiment details and analysis can be found in the supplementary materials." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 468, + 219, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 219, + 480 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 219, + 480 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 498, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 665 + ], + "type": "text", + "content": "This study represents a pioneering effort to incorporate text cues into the AVS task. Unlike previous approaches that solely rely on audio assistance, our TeSO method establishes a stronger audio-visual correlation by harnessing the robust semantics unique to the text modality By integrating CoT instructions and common-sense reasoning, we enhance the ability to capture potential sounding objects from detailed scene descriptions, thereby optimizing the utilization of text cues. Furthermore, we have developed a dynamic mask based on attention scores to effectively handle noisy information present in the text cues. Our comprehensive experiments highlight the effectiveness of our approach, demonstrating comparative performance across all AVS Benchmarks. In particular, more observation experiments validate that enhancing audio with text cues grants superior audio guidance. In summary, our work provides a new direction for building reliable audio-visual correlation with textual semantics, and we hope that it can help the community construct more robust audio-visual systems." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 480, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 480, + 166 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 480, + 166 + ], + "type": "text", + "content": "This research was supported by National Natural Science Foundation of China (NO.62106272), and Public Computing Cloud, Renmin University of China." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "spans": [ + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 211, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 211, + 481, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 211, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 211, + 481, + 245 + ], + "type": "text", + "content": "1. Abu-El-Haija, S., Kothari, N., Lee, J., Natsev, P., Toderici, G., Varadarajan, B., Vijayanarasimhan, S.: Youtube-8m: A large-scale video classification benchmark. arXiv preprint arXiv:1609.08675 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 245, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 481, + 278 + ], + "type": "text", + "content": "2. Bao, F., Nie, S., Xue, K., Li, C., Pu, S., Wang, Y., Yue, G., Cao, Y., Su, H., Zhu, J.: One transformer fits all distributions in multi-modal diffusion at scale. arXiv preprint arXiv:2303.06555 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 278, + 481, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 278, + 481, + 301 + ], + "spans": [ + { + "bbox": [ + 138, + 278, + 481, + 301 + ], + "type": "text", + "content": "3. Chen, S., Wu, Y., Wang, C., Liu, S., Tompkins, D., Chen, Z., Wei, F.: Beats: Audio pre-training with acoustic tokenizers. arXiv preprint arXiv:2212.09058 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 301, + 481, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 301, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 138, + 301, + 481, + 323 + ], + "type": "text", + "content": "4. Cheng, B., Choudhuri, A., Misra, I., Kirillov, A., Girdhar, R., Schwing, A.G.: Mask2former for video instance segmentation (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 323, + 481, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 481, + 367 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 481, + 367 + ], + "type": "text", + "content": "5. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290-1299 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 367, + 481, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 367, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 367, + 481, + 411 + ], + "type": "text", + "content": "6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 411, + 481, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 481, + 455 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 481, + 455 + ], + "type": "text", + "content": "7. Espejel, J.L., Ettifouri, E.H., Alassan, M.S.Y., Chouham, E.M., Dadhane, W.: Gpt-3.5, gpt-4, or bard? evaluating llms reasoning ability in zero-shot setting and performance boosting through prompts. Natural Language Processing Journal 5, 100032 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 456, + 481, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 481, + 477 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 481, + 477 + ], + "type": "text", + "content": "8. Gao, S., Chen, Z., Chen, G., Wang, W., Lu, T.: Avsegformer: Audio-visual segmentation with transformer (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 477, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 477, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 477, + 481, + 521 + ], + "type": "text", + "content": "9. Gemmeke, J.F., Ellis, D.P., Freedman, D., Jansen, A., Lawrence, W., Moore, R.C., Plakal, M., Ritter, M.: Audio set: An ontology and human-labeled dataset for audio events. In: 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP). pp. 776-780. IEEE (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 521, + 481, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 481, + 565 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 481, + 565 + ], + "type": "text", + "content": "0. Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: Imagebind: One embedding space to bind them all. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15180-15190 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 565, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 565, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 565, + 481, + 588 + ], + "type": "text", + "content": "1. Hao, D., Mao, Y., He, B., Han, X., Dai, Y., Zhong, Y.: Improving audio-visual segmentation with bidirectional generation. arXiv preprint arXiv:2308.08288 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 588, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 481, + 621 + ], + "type": "text", + "content": "2. Van der Heiden, R.M., Janssen, C.P., Donker, S.F., Hardeman, L.E., Mans, K., Kenemans, J.L.: Susceptibility to audio signals during autonomous driving. PloS one 13(8), e0201963 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 621, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 481, + 666 + ], + "type": "text", + "content": "3. Hershey, S., Chaudhuri, S., Ellis, D.P., Gemmeke, J.F., Jansen, A., Moore, R.C., Plakal, M., Platt, D., Saurous, R.A., Seybold, B., et al.: Cnn architectures for large-scale audio classification. In: 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP). pp. 131-135. IEEE (2017)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "text", + "content": "Sounding Object Segmentation Preference" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "14. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 182 + ], + "type": "text", + "content": "15. Hu, D., Wei, Y., Qian, R., Lin, W., Song, R., Wen, J.R.: Class-aware sounding objects localization via audiovisual correspondence. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12), 9844-9859 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "type": "text", + "content": "16. Huang, S., Li, H., Wang, Y., Zhu, H., Dai, J., Han, J., Rong, W., Liu, S.: Discovering sounding objects by audio queries for audio visual segmentation (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 480, + 226 + ], + "type": "text", + "content": "17. Hur, C., Park, H.: Zero-shot image classification with rectified embedding vectors using a caption generator. Applied Sciences 13(12), 7071 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 228, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 228, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 133, + 228, + 480, + 258 + ], + "type": "text", + "content": "18. Johnson, J., Karpathy, A., Fei-Fei, L.: Densecap: Fully convolutional localization networks for dense captioning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4565-4574 (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 259, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 259, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 133, + 259, + 480, + 292 + ], + "type": "text", + "content": "19. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 325 + ], + "type": "text", + "content": "20. Li, K., Yang, Z., Chen, L., Yang, Y., Xun, J.: Catr: Combinatorial-dependence audio-queried transformer for audio-visual video segmentation. arXiv preprint arXiv:2309.09709 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 346 + ], + "type": "text", + "content": "21. Ling, Y., Li, Y., Gan, Z., Zhang, J., Chi, M., Wang, Y.: Hear to segment: Unmixing the audio to guide the semantic segmentation (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 369 + ], + "type": "text", + "content": "22. Liu, C., Li, P., Qi, X., Zhang, H., Li, L., Wang, D., Yu, X.: Audio-visual segmentation by exploring cross-modal mutual semantics (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "23. Liu, C., Li, P., Zhang, H., Li, L., Huang, Z., Wang, D., Yu, X.: Bavs: Bootstrapping audio-visual segmentation by integrating foundation knowledge. arXiv preprint arXiv:2308.10175 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "text", + "content": "24. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "type": "text", + "content": "25. Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: Llava-last: Improved reasoning,OCR,and world knowledge (January 2024),https://llava-v1.github.io/blog/2024-01-30-llava-last/" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 457, + 421, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 421, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 421, + 468 + ], + "type": "text", + "content": "26. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "text", + "content": "27. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "text", + "content": "28. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "text", + "content": "29. Liu, J., Ju, C., Ma, C., Wang, Y., Wang, Y., Zhang, Y.: Audio-aware query-enhanced transformer for audio-visual segmentation (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "text", + "content": "30. Liu, J., Wang, Y., Ju, C., Zhang, Y., Xie, W.: Annotation-free audio-visual segmentation. arXiv preprint arXiv:2305.11019 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 555, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 599 + ], + "type": "text", + "content": "31. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "32. Ma, J., Sun, P., Wang, Y., Hu, D.: Stepping stones: A progressive training strategy for audio-visual semantic segmentation. IEEE European Conference on Computer Vision (ECCV) (2024)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "33. Majumder, S., Al-Halah, Z., Grauman, K.: Move2hear: Active audio-visual source separation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 275–285 (2021)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 658 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "34. Mo, S., Tian, Y.: Av-sam: Segment anything model meets audio-visual localization and segmentation. arXiv preprint arXiv:2305.01836 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 481, + 171 + ], + "type": "text", + "content": "35. Park, S., Senocak, A., Chung, J.S.: Marginnce: Robust sound localization with a negative margin. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 480, + 203 + ], + "type": "text", + "content": "36. Senocak, A., Oh, T.H., Kim, J., Yang, M.H., Kweon, I.S.: Learning to localize sound source in visual scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4358-4366 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "text", + "content": "37. Sharma, H., Jalal, A.S.: Image captioning improved visual question answering. Multimedia tools and applications 81(24), 34775-34796 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 481, + 246 + ], + "type": "text", + "content": "38. Shi, Z., Zhou, X., Qiu, X., Zhu, X.: Improving image captioning with better use of captions. arXiv preprint arXiv:2006.11807 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 246, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 365 + ], + "type": "text", + "content": "39. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D., Blecher, L., Ferrer, C.C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N., Hartshorn, A., Hosseini, S., Hou, R., Inan, H., Kardas, M., Kerkez, V., Khabsa, M., Kloumann, I., Korenev, A., Koura, P.S., Lachaux, M.A., Lavril, T., Lee, J., Liskovich, D., Lu, Y., Mao, Y., Martinet, X., Mihaylov, T., Mishra, P., Molybog, I., Nie, Y., Poulton, A., Reizenstein, J., Rungta, R., Saladi, K., Schelten, A., Silva, R., Smith, E.M., Subramanian, R., Tan, X.E., Tang, B., Taylor, R., Williams, A., Kuan, J.X., Xu, P., Yan, Z., Zarov, I., Zhang, Y., Fan, A., Kambadur, M., Narang, S., Rodriguez, A., Stojnic, R., Edunov, S., Scialom, T.: Llama 2: Open foundation and fine-tuned chat models (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 366, + 480, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 366, + 480, + 398 + ], + "spans": [ + { + "bbox": [ + 132, + 366, + 480, + 398 + ], + "type": "text", + "content": "40. Wang, Y., Liu, W., Li, G., Ding, J., Hu, D., Li, X.: Prompting segmentation with sound is generalizable audio-visual source localizer. arXiv preprint arXiv:2309.07929 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 399, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 399, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 399, + 480, + 430 + ], + "type": "text", + "content": "41. Wang, Y., Sun, P., Zhou, D., Li, G., Zhang, H., Hu, D.: Ref-avs: Refer and segment objects in audio-visual scenes. IEEE European Conference on Computer Vision (ECCV) (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 431, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 480, + 453 + ], + "type": "text", + "content": "42. Wu, W., Yao, H., Zhang, M., Song, Y., Ouyang, W., Wang, J.: Gpt4vis: What can gpt-4 do for zero-shot visual recognition? arXiv preprint arXiv:2311.15732 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 453, + 480, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 480, + 484 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 480, + 484 + ], + "type": "text", + "content": "43. Yan, S., Zhang, R., Guo, Z., Chen, W., Zhang, W., Li, H., Qiao, Y., He, Z., Gao, P.: Referred by multi-modality: A unified temporal transformer for video object segmentation. arXiv preprint arXiv:2305.16318 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 485, + 480, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 485, + 480, + 517 + ], + "spans": [ + { + "bbox": [ + 132, + 485, + 480, + 517 + ], + "type": "text", + "content": "44. Yang, Z., Li, L., Lin, K., Wang, J., Lin, C.C., Liu, Z., Wang, L.: The dawn of Imms: Preliminary explorations with gpt-4v (ision). arXiv preprint arXiv:2309.17421 9(1), 1 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 517, + 480, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 517, + 480, + 539 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 480, + 539 + ], + "type": "text", + "content": "45. Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual language model for video understanding (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 539, + 480, + 571 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 539, + 480, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 539, + 480, + 571 + ], + "type": "text", + "content": "46. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127(3), 302-321 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 571, + 480, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 571, + 480, + 603 + ], + "spans": [ + { + "bbox": [ + 132, + 571, + 480, + 603 + ], + "type": "text", + "content": "47. Zhou, J., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., Zhong, Y.: Audio-visual segmentation. In: European Conference on Computer Vision (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 603, + 480, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 603, + 480, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 603, + 480, + 635 + ], + "type": "text", + "content": "48. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 636, + 480, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 636, + 480, + 658 + ], + "spans": [ + { + "bbox": [ + 132, + 636, + 480, + 658 + ], + "type": "text", + "content": "49. Zürn, J., Burgard, W.: Self-supervised moving vehicle detection from audio-visual cues (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "text", + "content": "Sounding Object Segmentation Preference" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Canonical Shape Projection is All You Need for 3D Few-shot Class Incremental Learning/354b1d2e-f84c-4d0d-a2a6-9954d66577ef_content_list.json b/2024/Canonical Shape Projection is All You Need for 3D Few-shot Class Incremental Learning/354b1d2e-f84c-4d0d-a2a6-9954d66577ef_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c3a42af33ea6b6ef4c7b12a7d9e87fbb422faf78 --- /dev/null +++ b/2024/Canonical Shape Projection is All You Need for 3D Few-shot Class Incremental Learning/354b1d2e-f84c-4d0d-a2a6-9954d66577ef_content_list.json @@ -0,0 +1,1594 @@ +[ + { + "type": "text", + "text": "Canonical Shape Projection is All You Need for 3D Few-shot Class Incremental Learning", + "text_level": 1, + "bbox": [ + 228, + 140, + 776, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ali Cheraghian\\*,1,2, Zeeshan Hayden\\*,1,2, Sameera Ramasinghe\\*3, Shafin Rahman4, Javad Jafaryahya5, Lars Petersson2, and Mehrtash Harandi6", + "bbox": [ + 236, + 210, + 764, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Data61, CSIRO, Australia, 2 Australian National University, 3 Amazon, 4 North South University, 5 University of Technology Sydney, 6 Monash University {ali.cheraghian, zeeshan.hayder, lars.petersson}@data61.csiro.au, ramasisa@amazon.com, shafin.rahman@northsouth.edu, javad.jafaryahya@student.uts.edu.au, mehrtash.harandi@monash.edu", + "bbox": [ + 227, + 268, + 774, + 340 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In recent years, robust pre-trained foundation models have been successfully used in many downstream tasks. Here, we would like to use such powerful models to address the problem of few-shot class incremental learning (FSCIL) tasks on 3D point cloud objects. Our approach is to reprogram the well-known CLIP-based foundation model (trained on 2D images and text pairs) for this purpose. The CLIP model works by ingesting 2D images, so to leverage it in our context, we project the 3D object point cloud onto 2D image space to create proper depth maps. For this, prior works consider a fixed and non-trainable set of camera poses. In contrast, we propose to train the network to find a projection that best describes the object and is appropriate for extracting 2D image features from the CLIP vision encoder. Directly using the generated depth map is not suitable for the CLIP model, so we apply the model reprogramming paradigm to the depth map to augment the foreground and background to adapt it. This removes the need for modification or fine-tuning of the foundation model. In the setting we have investigated, we have limited access to data from novel classes, resulting in a problem with overfitting. Here, we address this problem via the use of a prompt engineering approach using multiple GPT-generated text descriptions. Our method, C3PR, successfully outperforms existing FSCIL methods on ModelNet, ShapeNet, ScanObjectNN, and CO3D datasets. The code is available at https://github.com/alichr/C3PR.", + "bbox": [ + 261, + 378, + 740, + 683 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D shape projection $\\cdot$ Model reprogramming $\\cdot$ Few-shot class incremental learning", + "bbox": [ + 259, + 696, + 740, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 752, + 375, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Maintaining a balance between acquiring new concepts and retaining existing knowledge presents a significant challenge for machine learning algorithms, a", + "bbox": [ + 212, + 785, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* denotes equal contribution.", + "bbox": [ + 217, + 824, + 415, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/362c13261a5cf4cefef74e2dd6ffe2218a4d9886d7caea7e9e702f19356f56dd.jpg", + "image_caption": [ + "Fig. 1: (a) Traditional methods using image-based models for point cloud processing often generate multiple depth maps from fixed camera poses, potentially causing feature alignment issues and difficulties in recognizing 3D objects from unsuitable camera angles. (b) In contrast, our proposed approach (C3PR) determines the most suitable camera pose for projecting a single depth map by leveraging feedback from the foundation model. (c) To improve the foundation model's performance in handling 3D data for the FSCIL task, we introduce a novel model reprogramming technique (without modifying or fine-tuning the foundation model) to further refine the depth map." + ], + "image_footnote": [], + "bbox": [ + 302, + 146, + 702, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "task that comes naturally to humans. This task becomes more difficult when only a few samples are available for new concepts. In the literature, this problem is referred to as few-shot class incremental learning (FSCIL) [7,9,10,24,37]. Substantial effort on this topic has been invested, particularly on image data [7,9,10, 24,37], whereas relatively few studies are tailored specifically towards point cloud data. For example, in the 2D domain, there are vision-language models such as CLIP [29], ALIGN [17], and FLAVA [33] models with an inherent capacity to generate highly versatile representations and demonstrate zero-shot capabilities, all without parameter fine-tuning [20, 25, 42]. Gathering training data for 3D point clouds is substantially more challenging than for image data, and here we explore whether the CLIP model can help alleviate this issue. In this paper, we attempt to leverage CLIP for FSCIL on 3D Point Cloud Objects.", + "bbox": [ + 212, + 609, + 787, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Popular approaches capture 2D images from a fixed set of camera poses and then pass the images to 2D foundation models for feature extraction [48, 51]. The main drawback of such a strategy is that camera poses are not adaptable", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "A. Cheraghian et al.", + "bbox": [ + 271, + 114, + 410, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to the target dataset, hindering capturing the objects in poses ideal for feature extraction (see Fig. 1 (a)). To address this problem, we propose to learn the camera projection pose based on the dataset that best describes the 3D objects (see Fig. 1 (b)). Our proposed model reprogramming approach is advantageous for representing real-world scanned 3D objects where point cloud data contains sensor noise (see Fig. 1 (c)). The unique camera pose can still capture the global shape of noisy 3D objects, even though the object belongs to novel classes. For language/text modality, many existing works [9-11] leverages vision-language learning to address the overfitting issue associated with few-shot novel classes. We argue that vital knowledge lies within the language domain, which may aid in the training process of the vision domain, particularly for few-shot classes, to alleviate overfitting. However, thus far, existing methods have only utilized class names from the language domain [9, 10, 12]. In contrast, we propose a method based on prompt engineering to better leverage the language domain.", + "bbox": [ + 212, + 146, + 787, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This paper presents a novel approach for extracting valuable insights from foundational models like CLIP, focusing on FSCIL on 3D point cloud data. As CLIP's training data comprises millions of image-text pairs, our target is to transfer this learned knowledge as a complete black box (without knowing or modifying pre-trained parameters) for 3D data. To this end, we propose reprogramming [38] the CLIP model, specifically for point cloud objects. The proposed approach is different from prior attempts [48,51] where they used a fixed set of camera poses for 3D to 2D image projection and did not augment the input data for better usage of the CLIP encoder. Our C3PR method emphasizes generating depth map images with consistent, informative, background pixels, diverging from conventional model reprogramming techniques that frequently produce random pixel information for the background. We observed that maintaining consistent background pixels contributes to improved generalization, particularly for few-shot novel classes. These steps indirectly add essential perturbation to the input depth map images so that the black box CLIP vision encoder can operate without fine-tuning. Furthermore, we propose a novel approach that leverages information from the language domain using a Large Language Model, GPT, to combat overfitting issues in few-shot novel classes. Specifically, we generate multiple descriptive prompts (instead of single class name-based descriptions as used in [11]) for base and novel classes. It also compensates for the issue with limited data during novel class learning. The generated prompts are then employed in the CLIP text encoder, reducing overfitting during incremental training sessions. Both the vision and text pipelines of our method can be trained end-to-end. We report state-of-the-art performances when evaluated on 3D synthetic datasets, ModelNet [45] and ShapeNet [6], and two 3D real world-scanned datasets, ScanObjectNN [40] and Common Objects in 3D (CO3D) [31].", + "bbox": [ + 212, + 359, + 787, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main contributions of our proposed method are:", + "bbox": [ + 238, + 756, + 620, + 771 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Learning the preferred object-based camera pose for 3D to 2D projections;", + "- A novel model reprogramming approach for 2D CLIP encoders targeting 3D point cloud data;" + ], + "bbox": [ + 225, + 791, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Canonical Shape Projection for 3D FSCIL", + "bbox": [ + 447, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Prompt engineering based on a GPT model for novel few-shot classes to mitigate overfitting.", + "bbox": [ + 223, + 146, + 787, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 200, + 382, + 217 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D point cloud processing: In recent years, there has been a surge in research focused on leveraging deep learning for the direct processing of 3D point cloud objects. Notable contributions include PointNet [27] by Qi et al., which introduced the use of multi-layer perceptron (MLP) networks for 3D point cloud processing. However, this approach overlooked the inherent local structures within the input data. Subsequent techniques were developed to address this limitation, including PointNet++ [28], which hierarchically extracts features to take advantage of local information. Building upon this, several works [21,23,26,30,44,47] have proposed convolution strategies for extracting local information. PointConv [44] introduced a novel convolution operation that locates a Monte Carlo estimation of the hidden continuous 3D convolution based on a powerful sampling technique. SFCNN [30] proposed graph convolution on spherical points. Other papers [41,43,49] treat each point cloud as a graph vertex to learn features in either spatial or spectral domains. DGCNN [43] constructs a graph in feature space and dynamically updates it using MLP for each edge, while PointGCN [49] suggests an approach that generates the graph using k-nearest neighbors from a point cloud to capture local information.", + "bbox": [ + 212, + 234, + 787, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Few-shot class-incremental learning: Tao et al. [37] were the pioneers in introducing the FSCIL setting for image data. They proposed the use of a neural gas network to address the issue of forgetting by preserving the topology of classes based on their feature vectors. Following this, in [10], a mixture of the subspaces are generated based on the training data distribution. In a related vein, Mazumder et al [24] devised a method that selectively adapts a subset of model parameters for learning novel classes, thereby curbing overfitting. They also combat forgetting by freezing critical parameters in the model. Cheraghian et al. [9] proposed a novel vision-language strategy, incorporating class semantic information from language space using a distillation technique to alleviate the impact of catastrophic forgetting. They also integrated an attention mechanism to counter overfitting on few-shot novel tasks. Introduced in [12], the Specific FSCIL tailored for 3D point cloud data presents the innovative concept of Microshape, which captures essential details from the primary task, effectively addressing both forgetting and overfitting challenges. Additionally, the work by [35] delves into cross-domain FSCIL applied to point-cloud recognition. This study dissects catastrophic forgetting into two distinct components: base class forgetting and incremental class forgetting, offering separate mitigation strategies.", + "bbox": [ + 212, + 492, + 787, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model reprogramming: Foundation model reprogramming has recently gained attention [8], primarily due to its reduced computational and time requirements compared to the fine-tuning approach. Prior to this, Elsayed et al. [15] introduced adversarial reprogramming. This technique involves mapping ImageNet labels to task labels, aiming to perplex the ImageNet classifier and induce be", + "bbox": [ + 212, + 765, + 789, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "A. Cheraghian et al.", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "havior akin to a CIFAR-10 or MNIST classifier. Dinh et al. [13] proposed that input reprogramming is more effective in situations with limited labeled data than training from scratch and fine-tuning. Tsai et al. [39] introduced a technique called black-box adversarial reprogramming. This method involves transforming the input and integrating label mapping mechanisms into the output. In this paper, we present a novel foundation model reprogramming approach tailored for the CLIP model.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 275, + 330, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first introduce the FSCIL problem settings. Let's consider a sequence of $T$ tasks denoted by $\\mathbf{Q} = \\{\\mathcal{Q}^1,\\mathcal{Q}^2,\\dots ,\\mathcal{Q}^T\\}$ , where $\\mathcal{V}^t$ represents the set of classes in task $\\mathcal{Q}^t$ , and $\\mathcal{V}^i\\cap \\mathcal{V}^j = \\emptyset$ . Additionally, each class in all tasks is associated with a set of prompt class descriptions generated by an LLM, e.g., GPT [4], denoted as $\\mathcal{P}^t$ . Thus, we can represent each task as a tuple $\\mathcal{Q}^t = \\{\\mathcal{X}_i^t,\\mathbf{y}_i^t,\\mathbf{p}_i^t\\}_{i = 1}^{n_t}$ , where $\\mathcal{X}_i^t = \\{\\mathbf{x}_{i,j}^t\\}_{j = 1}^l$ refers to a 3D point cloud object with coordinates $\\mathbf{x}_{i,j}^{t}\\in \\mathbb{R}^{3}$ . Moreover, $\\mathbf{y}_i^t\\in \\mathcal{V}^t$ and $\\mathbf{p}_i^t\\in \\mathcal{P}^t$ correspond to the label of the point cloud and its corresponding class prompt description, respectively. In the proposed FSCIL framework, $\\mathcal{Q}^1$ is the base task ( $t = 1$ ), where the model undergoes training on a large-scale synthetic 3D dataset. For $t > 1$ , the training data are drawn from real-world 3D point clouds consisting of only a few instances. The model is trained sequentially across the tasks $t = 1,\\ldots ,T$ . However, during the $t$ -th task, the model is exposed to $\\mathcal{X}^t$ , $\\mathbf{y}^t$ , and $\\{\\mathcal{P}^1,\\mathcal{P}^2,\\dots ,\\mathcal{P}^t\\}$ . During inference, the model trained on the current task $\\mathcal{Q}^t$ is expected to classify test samples from both the current and previous tasks, i.e., $\\{\\mathcal{Q}^1,\\mathcal{Q}^2,\\dots ,\\mathcal{Q}^t\\}$ .", + "bbox": [ + 212, + 306, + 789, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Model overview", + "text_level": 1, + "bbox": [ + 214, + 558, + 393, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present an overview of our proposed methodology, illustrated in Fig. 2. We begin with a point cloud sample denoted as $\\mathcal{X}_i^t$ , which is passed through the projection module $F$ . This module utilizes the rotation matrix $\\mathbf{R}_i^t$ derived from the transformation module $T_r$ to generate a depth map labelled as $\\mathbf{I}_{D_i}^t$ . For the projection module $F$ , we use the method proposed by [51], which is fixed during the training stage. The transformation module $T_r$ is crucial for determining the appropriate angle $\\theta$ for projection. Additionally, a soft constraint loss function $\\mathcal{L}_c$ is applied to the rotation matrix to ensure that the transformation module generates a valid rotation matrix. Following this, we extract a binary mask $\\mathbf{I}_{M_i}^t$ from the depth map. Within this mask $M$ , background pixels are assigned a value of zero, while non-zero depth map values are set to one. Subsequently, this mask is fed into a model reprogramming module $C$ to add pixel information to the background of the depth map. To ensure that the generated embedding does not converge to a noisy image and follows the pattern of the mask input, an image gradient consistency loss function $\\mathcal{L}_g$ is used between the input and output of the module $C$ . The resulting image embedding $\\mathbf{I}_{H_i}^t$ is then merged with the original depth map in the subsequent stage, resulting in", + "bbox": [ + 212, + 583, + 787, + 843 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Canonical Shape Projection for 3D FSCIL", + "bbox": [ + 447, + 113, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/14f2f243edd12565d6f9449f7e653a655f888504c9340c334d5c0d6bc19201b5.jpg", + "image_caption": [ + "Fig. 2: Beginning with a point cloud sample $\\mathcal{X}_i^t$ , we use the projection module $F$ to generate a depth map $\\mathbf{I}_{D_i}^t$ via the rotation matrix $\\mathbf{R}_i^t$ from the transformation module $T_r$ . A mask $\\mathbf{I}_{M_i}^t$ is then extracted from the depth map. This mask is input into module $C$ and the resulting image is merged with the depth map, yielding the image $\\mathbf{I}_{R_i}^t$ . Next, $\\mathbf{I}_{R_i}^t$ is input to the CLIP vision encoder model $V_e$ to generate a vision embedding $\\mathbf{v}_i^t \\in \\mathbb{R}^m$ . Simultaneously, $\\mathcal{X}_i^t$ is fed into $B$ to create a point cloud representation $\\mathbf{z}_i^t \\in \\mathbb{R}^m$ . Moreover, the GPT model generates multiple class name descriptions, which are then stored in a prompt bank. To this end, prompts $\\{\\mathbf{p}_1, \\mathbf{p}_2, \\dots, \\mathbf{p}_{|\\mathcal{Y}_{tl}|}\\}$ for all old classes and the current task $t$ are forwarded into the text encoder $T_e$ , producing corresponding text embeddings $\\{\\mathbf{e}_1, \\mathbf{e}_2, \\dots, \\mathbf{e}_{|\\mathcal{Y}_{tl}|}\\}$ , where $\\mathbf{e}_i^t \\in \\mathbb{R}^m$ . Finally, $\\mathbf{v}_i^t$ , $\\mathbf{z}_i^t$ , and $\\mathbf{e}_j^t$ are concatenated and passed through the relation module $R_l$ to generate a scalar value $r_i^t$ , serving as a similarity score." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the production of a reprogrammed depth map image denoted as $\\mathbf{I}_{R_i}^t$ . After that, $\\mathbf{I}_{R_i}^t$ is input into the vision encoder of the CLIP model $V_{e}$ to generate a vision embedding represented by $\\mathbf{v}_i^t\\in \\mathbb{R}^m$ . Simultaneously, the point cloud input $\\mathcal{X}_i^t$ is forwarded to a 3D point cloud backbone, for instance, Curvenet [46], to generate a feature description $\\mathbf{z}_i^t\\in \\mathbb{R}^m$ . At the same time, GPT is employed to generate multiple descriptions for each class using the command: describe a depth map of a [class], thereby creating several prompts, which are stored in the prompt bank module. Then, the prompts of all classes $\\{\\mathbf{p}_j\\}_{j = 1}^{|\\mathcal{Y}_{tl}|}$ , where $\\mathcal{Y}_{tl} = \\bigcup_{i = 1}^{t}\\mathcal{Y}^{i}$ , observed up to task $t$ are fed into the text encoder $T_{e}$ to generate the corresponding text embeddings denoted as $\\{\\mathbf{e}_j\\}_{j = 1}^{|\\mathcal{Y}_{tl}|}$ , where each $\\mathbf{e}_j^t\\in \\mathbb{R}^m$ . The prompts are generated using LLMs, such as GPT [4], given a command: Describe a depth map of a [class]. In the subsequent stage, all $\\mathbf{v}_i^t$ , $\\mathbf{z}_i^t$ , and $\\mathbf{e}_i^t$ are concatenated and passed through the relation module [34] $R_{l}$ to generate a scalar value $r_i^t$ ranging from 0 to 1. This value serves as a measure of the similarity between the visual and prompt feature embeddings. Finally, a binary cross-entropy loss function $\\mathcal{L}_r$ is used to train the entire proposed pipeline.", + "bbox": [ + 212, + 510, + 787, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Canonical 3D shape projection", + "text_level": 1, + "bbox": [ + 215, + 784, + 516, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A 3D-to-2D projection matrix can be designed to project any 3D point cloud on a 2D plane considering any camera angle. Nevertheless, the feature embedding", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "A. Cheraghian et al.", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d7a377aabb210e3f2f7785bbb9c27189deaf50fd77d8c0e1391a395fdda77b67.jpg", + "image_caption": [ + "Fig. 3: We demonstrate the sensitivity of the CLIP model to the angle of depth map projection. Here, we consider two random projections to generate depth maps. Additionally, we assume that there are three classes: car, guitar, and person. Even though the number of classes is limited to only three, the CLIP model can only correctly classify two of these depth maps. This highlights the importance of finding the optimum projection angle to fully leverage the CLIP model for 3D point cloud processing." + ], + "image_footnote": [], + "bbox": [ + 300, + 143, + 702, + 289 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "produced by the visual encoder of large image-based foundation models (e.g., CLIP) can be affected by the content that can be controlled by the projection angle. Figure 3 shows that an adjustment is pivotal due to the CLIP model's sensitivity to the projection angle, especially during depth map processing. We are interested in finding the optimal projection angle to enhance the performance of the CLIP model. When projecting from a 3D point cloud to 2D images, we leverage the mapping function introduced by [51]. We provide optimum projection angle $\\theta$ via a rotation matrix $\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}$ , where $\\mathbf{R} = T_r(\\mathcal{X})$ , generated by the transformation module $T_r$ to ascertain the most effective projection angle for converting a 3D point cloud into a depth map. The generated $\\mathbf{R}$ is then applied to the given 3D point to derive the appropriate angle for the projection module $\\theta$ . To ensure that $\\mathbf{R}$ is a valid rotation matrix, we apply the following loss function at the output of the transformation module $T_r$ ,", + "bbox": [ + 212, + 412, + 787, + 609 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(\\left\\| \\mathbf {R} _ {i} ^ {T} \\mathbf {R} _ {i} - \\mathbf {I} \\right\\| _ {\\mathcal {F}} ^ {2} + \\left| \\det (\\mathbf {R} _ {i}) - 1 \\right|\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 621, + 785, + 664 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{R}_i = T_r(\\mathcal{X}_i)$ , $n$ is the number of samples in the training set, $\\| \\cdot \\|_{\\mathcal{F}}^2$ denotes the Frobenius norm, $\\mathbf{I}$ is the identity matrix. The orthogonality condition $\\left\\| \\mathbf{R}_i^T\\mathbf{R}_i - \\mathbf{I}\\right\\|_F^2$ guarantee that the columns of $\\mathbf{R}_i$ form an orthonormal basis to preserve distances and angles for rotations. $\\left|\\operatorname*{det}(\\mathbf{R}_i) - 1\\right|$ affirms that $\\mathbf{R}_i$ is an orientation-preserving transformation matrix and doesn't involve any reflection.", + "bbox": [ + 212, + 667, + 787, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In practice, the $T_{r}$ matrix may not be guaranteed to be a rotation matrix on a per-batch basis as the above condition is only a soft constraint. This is overcome by continuous re-orthogonalization during the training process. Moreover, alternative methods like euler-angle or quaternion representations could be used. Nevertheless, the proposed approach of enforcing orthogonality through the $L_{c}$ is a known method in the literature [1]. In the end, when provided with a 3D", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Canonical Shape Projection for 3D FSCIL", + "bbox": [ + 447, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "point cloud sample $\\mathcal{X}_i$ , the transformation module $T_r$ computes the desired rotation matrix $\\mathbf{R}_i$ while taking into account all the aforementioned constraints to ensure its validity. This process results in the generation of a learned depth map image, $\\mathbf{I}_{D_i} = F(\\mathcal{X}_i,\\mathbf{R}_i)$", + "bbox": [ + 212, + 146, + 787, + 208 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Model reprogramming", + "text_level": 1, + "bbox": [ + 214, + 229, + 449, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Model reprogramming of computer vision models, especially foundational ones, is a crucial step in their effective deployment, enabling a seamless transition to new tasks and domains with minimal adjustments to existing architectures [38]. We present a novel model reprogramming approach specifically tailored for the CLIP model applied to the FSCIL task on 3D point cloud objects. It is worth noting that the CLIP model is primarily trained on RGB images rather than depth maps. Therefore, to enhance the CLIP model's effectiveness with depth maps, we incorporate pixel information into the background of the depth map images. More specifically, this modification assists the CLIP model in accurately discerning the boundary of the depth map shape information through a $C$ module, which takes the form of a U-Net [32]. To accomplish this, we initially compute the mask map $\\mathbf{I}_{M_i}$ corresponding to a given depth map $\\mathbf{I}_{D_i}$ , which isolates the shape boundary from the background. Next, we input $\\mathbf{I}_{M_i}$ into the $C$ module to extract the pixel information for the background, denoted as $\\mathbf{I}_{H_i}$ . To ensure the consistency of the generated pixels in $\\mathbf{I}_{M_i}$ from the $C$ module and to prevent convergence to arbitrary values, we employ the following loss function,", + "bbox": [ + 212, + 257, + 787, + 500 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {g} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(\\frac {\\partial \\mathbf {I} _ {M _ {i}}}{\\partial x} - \\frac {\\partial \\mathbf {I} _ {H _ {i}}}{\\partial x}\\right) ^ {2} + \\left(\\frac {\\partial \\mathbf {I} _ {M _ {i}}}{\\partial y} - \\frac {\\partial \\mathbf {I} _ {H _ {i}}}{\\partial y}\\right) ^ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 510, + 785, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\frac{\\partial}{\\partial x}$ and $\\frac{\\partial}{\\partial y}$ denote the derivatives of the images in the $x$ and $y$ directions, respectively. This loss enforces a smoothness constraint on the gradient to separate it from foreground objects explicitly.", + "bbox": [ + 212, + 561, + 787, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "It helps to learn novel classes with few examples. Subsequently, we combine the depth map $\\mathbf{I}_{D_i}$ from the projection module $F$ with the extracted values $\\mathbf{I}_{H_i}$ from $C$ to generate the final image, $\\mathbf{I}_{R_i} = \\mathbf{I}_{D_i} \\odot C(M(\\mathbf{I}_{D_i}))$ , where $\\odot$ represents the element-wise multiplication.", + "bbox": [ + 212, + 608, + 787, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Prompt engineering", + "text_level": 1, + "bbox": [ + 214, + 691, + 428, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Advancements in prompt engineering have unveiled significant potential, primarily driven by LLMs [18, 19]. LLMs can extract richer information from appropriately tailored prompts. With this motivation, we employ GPT to obtain more insights from plain class names and enhance the FSCIL task on 3D data. In contrast to prevailing methods [9, 10, 12] that rely solely on the class name for the FSCIL task, we extend our method by generating additional descriptions to extract more semantic information from the CLIP text encoder (denoted as $T_{e}$ ). This augmentation ensures a more comprehensive utilization of the LLM's", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "A. Cheraghian et al.", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "capabilities, thereby enhancing the overall efficacy of our approach. In the FS-CIL, a significant challenge arises due to the scarcity of training samples for novel classes, potentially leading to overfitting. To address this issue, we propose an innovative strategy leveraging the capabilities of LLM, specifically the GPT model, and utilizing this as novel prompts for the text encoder of the CLIP model. More precisely, we utilize the GPT model to generate multiple descriptions for each class by providing it with a specific command. These prompts are formally defined for each sample as follows:", + "bbox": [ + 212, + 146, + 782, + 267 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {i} = \\operatorname {G P T} (\\text {C o m m a n d s}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 276, + 784, + 292 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Here, the Commands are structured as Describe a depth of a [class]. For every class in the dataset, we substitute the respective class names in place of '[class]' in the command and input them into GPT-4. This results in the generation of depth map-specific descriptions with comprehensive category-wise semantics. Subsequently, we feed the descriptions of each category through CLIP's textual encoder $\\mathcal{T}_e$ . Finally, the generated prompts $\\mathbf{p}_i$ are used in the training of the few-shot classes.", + "bbox": [ + 212, + 297, + 784, + 402 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 Class incremental learning", + "text_level": 1, + "bbox": [ + 214, + 422, + 480, + 436 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training. For the base task, denoted as $\\mathcal{Q}^1$ , we conduct training for the transformation model $T_{r}$ , the 3D backbone $B$ , and the relation module $R_{l}$ . In the case of subsequent tasks, denoted as $\\mathcal{Q}^t$ , where $t > 1$ , only the relation module undergoes fine-tuning. Throughout both base and subsequent novel tasks, all other components in our proposed architecture (refer to Figure 2) remain frozen. Additionally, for each previously learned class, we randomly select a sample to be stored in a compact memory module $\\mathcal{M}$ . Given the $i^{th}$ point cloud $\\mathcal{X}_i^t$ of the $t^{th}$ task, the $B$ module generates the features $\\mathbf{z}_i^t$ and the $V_{e}$ module produces $\\mathbf{v}_i^t$ . Simultaneously, the prompts of all classes for tasks $\\mathcal{Q}^1,\\ldots ,\\mathcal{Q}^t$ are fed into the text encoder $T_{e}$ to generate the corresponding text embeddings denoted as $\\{\\mathbf{e}_j\\}_{j = 1}^{|\\mathcal{Y}_{tl}|}$ . Then, we concatenate $\\mathbf{z}_i^t$ , $\\mathbf{v}_i^t$ , and $\\mathbf{e}_j^t$ and forward them into the relation module $R_{l}$ , which provides a score between [0,1], representing the similarity between them. In other words, for each training sample, we generate a score against each of the classes in both the subsequent and previous tasks as $r_{ij}^{t} = \\gamma \\circ R\\circ (\\mathbf{z}_{i}^{t}\\oplus \\mathbf{v}_{i}\\oplus \\mathbf{e}_{j}), j\\in \\mathcal{Y}_{tl}$ , where $\\mathcal{Y}_{tl} = \\bigcup_{i = 1}^{t}\\mathcal{Y}^{i}$ , $\\oplus$ is the concatenation operator, $R$ is the relation module, and $\\gamma$ is the sigmoid function. Finally, for each feature $r_{ij}$ and the corresponding ground truth $\\mathbf{y}_i$ , we employ a binary cross-entropy cost function to train the model as follows,", + "bbox": [ + 212, + 443, + 787, + 719 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {r} = - \\frac {1}{| \\mathcal {Y} _ {t l} | | S |} \\sum_ {k \\in \\mathcal {Y} _ {t l}} \\sum_ {\\mathbf {y} _ {i} \\in \\mathcal {S}} \\Bigg (1 (\\boldsymbol {y} _ {i} ^ {t} = = k) \\log (r _ {i k}) + \\big (1 - 1 (\\boldsymbol {y} _ {i} ^ {t} = = k) \\big) \\log (1 - r _ {i k}) \\Bigg),\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 738, + 794, + 777 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathcal{S}$ denotes the set of true labels in the current task, along with the memory $\\mathcal{M}$ . Finally, the total loss to train the entire pipeline is as follows,", + "bbox": [ + 214, + 782, + 784, + 811 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t} = \\lambda_ {1} \\mathcal {L} _ {c} + \\lambda_ {2} \\mathcal {L} _ {g} + \\lambda_ {3} \\mathcal {L} _ {r} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 821, + 784, + 837 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Canonical Shape Projection for 3D FSCIL", + "bbox": [ + 447, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda_1, \\lambda_2$ , and $\\lambda_3$ are the weighting factors for the constraint loss $(\\mathcal{L}c)$ , gradient loss $(\\mathcal{L}g)$ , and the relation loss $(\\mathcal{L}_r)$ respectively.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Inference. During the inference process, when given an unlabeled sample $\\mathcal{X}^c$ , where $c \\in \\mathcal{Y}_{tl}$ , the label prediction $y^{*}$ is computed using the following equation:", + "bbox": [ + 212, + 176, + 784, + 207 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\ny ^ {*} = \\underset {j \\in \\mathcal {Y} _ {t l}} {\\arg \\max } (R \\circ (\\mathbf {z} _ {c} \\oplus \\mathbf {v} _ {c} \\oplus \\mathbf {s} _ {j})), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 220, + 785, + 247 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\mathbf{z}_c$ and $\\mathbf{v}_c$ represent the calculated features for the sample $\\mathcal{X}^c$ . We use a standardized prompt for all classes, specifically This is a depth map of a [class].", + "bbox": [ + 212, + 257, + 782, + 287 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 310, + 375, + 328 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. We employed two synthetic 3D datasets, Modelnet40 [45] and Shape-net [6], in addition to two real scanned 3D datasets, ScanobjectNN [40] and CO3D [31]. We follow the setup proposed by [11] for Few-Shot Classification with Incremental Learning (FSCIL) on point cloud objects. Our proposed method is assessed in two distinct scenarios. Firstly, we evaluate its performance on the FSCIL task within the same dataset, where both the base and novel tasks originate from identical datasets. Secondly, we subject our method to more demanding and realistic challenges by testing it in cross-dataset scenarios. In this setting, the base task involves synthetic datasets, while the novel tasks comprise classes from real scanned few-shot datasets.", + "bbox": [ + 212, + 340, + 784, + 491 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Prompt generation. GPT-4 was employed to produce diverse class descriptions for each class. The following command was utilized: Describe a depth of a [class]. Approximately 100 class descriptions were generated for each class and utilized during the training phase.", + "bbox": [ + 212, + 492, + 784, + 553 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation details. In our experimentation, we utilize Curvenet as the backbone for 3D point clouds $(B)$ [46]. The CLIP vision $(V_{e})$ and text encoder $(T_{e})$ are instantiated using the ViT-B-16 architecture [14]. The $C$ model is constructed based on the U-Net architecture [32]. The projection module $(F)$ follows the methodology proposed by Zhu et al. [51], and for the transformation module $(T_{r})$ , we employ the STN model [16]. Additionally, the relation module comprises several fully connected layers with sizes (1024, 2048, 1). We use the farthest 1024 points from 3D point cloud objects as input for all samples. For the base task, we train our proposed method (Figure 2) in an end-to-end manner for 250 epochs using the Adam optimizer with a learning rate of 0.001. For the novel task, we fine-tune our model for 200 epochs using the Adam optimizer with a learning rate of 0.001. It is important to note that we randomly shift and scale points in the input point cloud during both base and novel class training, with the additional step of randomly dropping out points. The feature vector dimensions for $B$ , $T_{e}$ , and $V_{e}$ are set to 512. Additionally, the weights of the loss functions $\\lambda_{1}$ , $\\lambda_{2}$ , and $\\lambda_{3}$ are configured as 0.01, 1, and 1 across the entire datasets. Our experiments are conducted using the PyTorch framework.", + "bbox": [ + 212, + 553, + 785, + 809 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation metrics. Following the training of each task, we compute the accuracy by consolidating both base and novel classes. Subsequently, following the", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "A. Cheraghian et al.", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "recommendation in [36], we determine the relative accuracy dropping rate, denoted as $\\varDelta$ , using the formula $\\varDelta=\\frac{|acc_T-acc_0|}{acc_0}\\times 100$ , where $acc_{T}$ and $acc_{0}$ denote the accuracy of the last and first tasks, respectively. The value of $\\varDelta$ serves as a comprehensive metric for method evaluation. A lower relative accuracy signifies superior performance.", + "bbox": [ + 212, + 146, + 787, + 224 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Main results", + "text_level": 1, + "bbox": [ + 215, + 258, + 366, + 272 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We conducted a comparative analysis of our approach against established methods in the field, including IL2M [2], ScaIL [3], EEIL [5], LwF [22], FACT [50], Sem-aware [9], and Microshape [12]. Notably, all these methods are tailored for image data, with the exception of Microshape [12], which was originally designed for point cloud data.", + "bbox": [ + 212, + 294, + 782, + 369 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Within dataset: In this scenario, both the base and novel tasks are derived from the same dataset. Roughly half of the classes in the dataset are assigned to the base task, while the remaining half are incrementally introduced as multiple novel tasks. It is presumed that these novel tasks encompass only a limited number of samples for training. Our proposed method has been evaluated on three datasets—ModelNet, CO3D, and Shapenet. As depicted in Table 1, our approach demonstrates superior performance across all methods in terms of the relative accuracy dropping rate, denoted as $\\Delta$ , indicating the effectiveness of our proposed method.", + "bbox": [ + 212, + 371, + 787, + 507 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Cross dataset: In this arrangement, the base task centers around synthetic point cloud objects, while the novel few-shot tasks incorporate real-scanned 3D objects. Notably, the latter often exhibit incompleteness or noise in contrast to the synthetic data, which typically boast cleanliness and completeness. Consequently, this configuration introduces a more realistic and challenging scenario compared to within-dataset experiments. Three datasets are considered in this setting: Shapenet $\\rightarrow$ CO3D, ModelNet40 $\\rightarrow$ ScanObjectNN, and Shapenet $\\rightarrow$ ScanObjectNN. As illustrated in Table 2, our method outperforms other approaches in all three setups, underscoring its efficacy in scenarios involving a domain gap between base and novel tasks.", + "bbox": [ + 212, + 510, + 787, + 660 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/4f1ece16acc37c7f000773b3db179f8c3f102890bb4036082a8d5a6efaf5e045.jpg", + "table_caption": [ + "Table 1: Summary of FSCIL results for within-dataset experiments." + ], + "table_footnote": [], + "table_body": "
ModelNetCO3DShapeNet
Method2025303540Δ↓253035404550Δ↓25303540455055 Δ↓
FT89.89.74.33.33.096.776.711.23.63.21.80.899.087.025.76.81.30.90.60.4 99.5
Joint89.888.287.083.580.510.476.769.464.862.760.759.822.087.085.284.383.082.582.281.3 6.6
LwF [22]89.836.09.13.63.196.076.714.74.73.52.31.098.787.060.833.515.93.83.11.8 97.9
IL2M [2]89.865.558.452.353.640.376.731.527.718.127.121.971.487.058.645.740.750.149.449.3 43.3
ScaIL [3]89.866.864.558.756.537.176.739.534.124.130.127.564.187.056.651.844.350.346.445.4 47.8
EEIL [5]89.875.467.260.155.638.176.761.452.442.839.532.857.287.077.773.269.366.465.965.8 22.4
FACT [50]90.481.377.173.565.028.177.967.159.754.850.246.740.087.575.371.469.967.565.762.5 28.6
Sem-aware [9]91.382.274.370.064.729.176.866.959.253.649.142.944.187.274.968.969.168.066.963.8 26.8
Microshape [11]93.683.178.275.867.128.378.567.360.156.151.447.239.987.683.281.579.076.873.5 72.617.1
C3PR (ours)91.682.375.872.270.922.581.569.466.563.054.253.834.088.081.677.876.776.776.976.2715.1
", + "bbox": [ + 220, + 708, + 795, + 820 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Canonical Shape Projection for 3D FSCIL", + "bbox": [ + 447, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/bdfcf39ada3895884b55569cfdf314cf1999d3872fcaeff6c0e4fa37d3548dde.jpg", + "table_caption": [ + "Table 2: Summary of FSCIL results for within-dataset and cross-dataset experiments." + ], + "table_footnote": [], + "table_body": "
ShapeNet → CO3DModelNet → ScanObjectNNShapeNet → ScanObjectNN
Method3944495459646974798489Δ↓26303437Δ↓44495459Δ↓
FT81.020.22.31.70.81.01.01.30.90.51.698.088.46.46.01.997.981.438.74.00.998.9
Joint81.079.578.375.275.174.872.371.370.068.867.316.988.479.774.071.219.581.482.579.878.73.3
LwF [22]81.057.419.32.31.00.90.81.31.10.81.997.788.435.85.82.597.281.447.914.05.992.8
IL2M [2]81.045.636.835.131.833.334.031.530.632.330.063.088.458.252.952.041.281.453.243.945.843.7
ScaIL [3]81.050.145.739.139.037.938.036.033.733.035.256.588.456.555.952.940.281.449.046.740.050.9
EEIL [5]81.075.269.363.260.557.953.051.951.347.847.641.288.470.261.056.835.781.474.569.863.422.1
FACT [50]81.476.070.368.165.863.563.060.158.257.555.931.389.172.568.363.528.782.374.669.966.818.8
Sem-aware [9]80.669.566.562.963.263.061.258.358.157.255.231.688.573.967.764.227.581.370.665.262.922.6
Microshape [11]82.677.973.972.767.766.265.463.460.658.157.130.989.373.268.465.127.182.574.871.267.118.7
C3PR (ours)83.680.077.875.472.872.370.367.964.964.163.224.488.375.770.667.823.284.577.875.571.914.9
", + "bbox": [ + 218, + 160, + 803, + 273 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 292, + 401, + 306 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Evaluating model reprogramming: In the conducted experiment (refer to Figure 4 (a)), we systematically omitted the model reprogramming module $C$ and directly fed the depth map image $\\mathbf{I}_D$ into the vision encoder $V_{e}$ of the CLIP model. The results clearly illustrate the significance of the model reprogramming in our proposed approach, as evidenced by notable differences between the outcomes with and without this module.", + "bbox": [ + 212, + 325, + 785, + 416 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Impact of the prompt engineering: In our ablation study (see Figure 4 (b)), we investigate the influence of prompt engineering in our methodology. Specifically, we streamline the training stage by substituting the original multiple prompts per class with a single fixed prompt. The outcomes underscore the effectiveness of our GPT-based prompt engineering approach.", + "bbox": [ + 212, + 419, + 785, + 494 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Impact of canonical shape projection: Here, we evaluate our proposed canonical shape projection methodology in comparison to PointCLIPv2 [51], which generates multiple fixed-depth map projections for each input point cloud. Specifically, we eliminate the transformation module $T_r$ . Subsequently, we adopt the depth map generation method employed in PointCLIP v2, involving the use of ten fixed camera pose angles for computing ten depth map images. These images are then fed into the vision encoder $V_e$ of the CLIP model, and the relation module is applied to merge the feature vector, mirroring our proposed pipeline. Notably, our method surpasses PointCLIPv2 by a significant margin (see Figure 4 (c)), highlighting the efficacy of our learned camera pose angle.", + "bbox": [ + 212, + 497, + 787, + 650 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f2735cff78476fd9eec68ebb24ea1f7d2fb4381219dff78ddc26c7952aea34be.jpg", + "image_caption": [ + "Fig. 4: This experiment is performed within the Shapenet to ScanobjectNN setting. In (a), we present the influence of our novel CLIP model reprogramming strategy. (b) Highlights the impact of prompt engineering as proposed in our method. (c) Our approach is compared with PointCLIP v2 [51]." + ], + "image_footnote": [], + "bbox": [ + 217, + 679, + 408, + 768 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/891842fb6002b36c8659e7ce176eb5e3ad87715498c7c2ae39b4daba34465bf7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 679, + 596, + 768 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/42a582476af33b9e1cbc6568663637754912fe8e31dcff1b891bf7ef7170887d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 680, + 782, + 767 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "A. Cheraghian et al.", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/eb73fc7a41ed88ab3af361dd4727f27eae7410e8e44844d1a11f8deb8b78f207.jpg", + "image_caption": [ + "Fig. 5: We visualize the quality of features (from 44 base classes of ShapeNet) generated by the CLIP vision encoder using (Right) our proposed method, C3PR, and (Left) PointCLIPv2 [51]. Our method produced a denser cluster than PointCLIP." + ], + "image_footnote": [], + "bbox": [ + 303, + 143, + 493, + 258 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/bcf4f8654d88c7b467beafb144fcc896a12b208d66957f5687610ce8c46ea02e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 143, + 699, + 258 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Feature quality: In Figure 5, we assess the feature quality produced by the CLIP vision encoder using our proposed method, C3PR, and compare it with PointCLIPv2 [51]. As illustrated, our approach yields more discriminative clusters, enhancing overall generalization capabilities.", + "bbox": [ + 212, + 325, + 782, + 386 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Influence of $\\mathcal{L}_c$ and $\\mathcal{L}_g$ : In Table 3a, we highlight the impact of both $\\mathcal{L}_c$ and $\\mathcal{L}_g$ within our proposed pipeline. It is evident that both loss functions contribute significantly, but $\\mathcal{L}_c$ holds greater importance, underscoring the significance of identifying the learned projection angle for depth map images on image foundation models.", + "bbox": [ + 212, + 387, + 784, + 462 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Model reprogramming: Conventional vs. our: In this section, we contrast our innovative model reprogramming approach with the conventional method introduced by [15]. Table 3b illustrates the superior effectiveness of our proposed model reprogramming strategy.", + "bbox": [ + 212, + 464, + 784, + 525 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/cfe8fd63775a6853ed7dad5d688ac528b8d2e70ea28cb111bad9d7ff5204c5ad.jpg", + "table_caption": [ + "Table 3: Ablation study on the influence of loss (a) and model reprogramming (b).", + "(a) Influence of $\\mathcal{L}_c$ and $\\mathcal{L}_g$ in our approach." + ], + "table_footnote": [], + "table_body": "
ShapeNet → ScanObjectNN
Method44495459Δ↓
Lr84.476.270.16522.9
Lr + Lg84.377.370.866.920.6
Lr + Lc85.780.274.570.817.3
Lr + Lg + Lc84.577.875.571.914.9
", + "bbox": [ + 217, + 578, + 491, + 667 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/5a1476d1d41162eb787e47e8ccf172f793ad09d4ebcbb2865f18b3f88cf3a116.jpg", + "table_caption": [ + "(b) Comparing our proposed model reprogramming strategy, C3PR, with the conventional model reprogramming technique employed in [15]." + ], + "table_footnote": [], + "table_body": "
ShapeNet → ScanObjectNN
Method44495459Δ↓
[15]85.780.274.570.817.3
C3PR (ours)84.577.875.571.914.9
", + "bbox": [ + 506, + 599, + 777, + 659 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Discussion", + "text_level": 1, + "bbox": [ + 215, + 688, + 346, + 700 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Point-cloud classification: Although our learning projection module $Tr$ and model reprogramming module $C$ have demonstrated effectiveness in the FSCIL context, they can also be applied in the point-cloud classification setting (shown in the table 4a. However, it is worth to note that our proposed prompt engineering strategy is specifically tailored to improve performance in FSCIL problems.", + "bbox": [ + 212, + 717, + 782, + 792 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Module $C$ role:", + "text_level": 1, + "bbox": [ + 215, + 792, + 343, + 806 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The role of Module $C$ is twofold: it learns background information for the depth maps and smoothens the foreground. Specifically, we combine the depth", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Canonical Shape Projection for 3D FSCIL", + "bbox": [ + 447, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "map $\\mathbf{I}_{D_i}$ from the projection module $F$ with the extracted values $\\mathbf{I}_{T_i}$ from Module $C$ to generate the final image, $\\mathbf{I}_{R_i} = \\mathbf{I}_{D_i} \\odot C(M(\\mathbf{I}_{D_i}))$ , where $\\odot$ represents element-wise multiplication. To this end, we conducted experiments to demonstrate the impact of Module $C$ in our proposed method (see table 4b).", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "(a) The performance of our proposed method on point-cloud classification in ModelNet40.", + "bbox": [ + 215, + 234, + 480, + 257 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "(b) The impact of module $C$ in our proposed method.", + "bbox": [ + 501, + 234, + 789, + 257 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/5b199210ac6a1651149622468d952902599504d0fda9044d6a33669865d56f1f.jpg", + "table_caption": [ + "Table 4: Ablation on point-cloud classification (a) and impact of $C$ module (b)." + ], + "table_footnote": [], + "table_body": "
MethodAccuracy
DGCNN [43]92.9
CurveNet [46]93.8
Ours94.6
", + "bbox": [ + 274, + 268, + 441, + 330 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/1c354097e1da4f62214c19a73ba56d57c186d0cf521338a3a1505f62b7655d5d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ShapeNet → ScanObjectNN
Method44495459Δ↓
Without C84.576.373.268.918.4
With C84.577.875.571.914.9
", + "bbox": [ + 501, + 268, + 761, + 330 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "**Limitation:** The proposed method exhibits several limitations that warrant discussion. Firstly, its reliance on storing exemplars in memory, with one sample per class, and addressing the forgetting issue raises the question of exploring exemplar-free alternatives, which could offer intriguing avenues for further research. Secondly, our method would perform worse on new concepts from significantly different domains, synthetic to real-scan datasets, as can be seen in Table 2. However, it works well when the new concepts come from similar domains, synthetic to synthetic datasets, as can be seen in Table 1.", + "bbox": [ + 212, + 338, + 787, + 459 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 479, + 359, + 494 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we leverage the CLIP architecture of image-to-text matching to address the FSCIL problem on 3D point cloud objects. Traditional approaches usually project a 3D object on a 2D plane to produce images from different camera angles. Among those 2D images, some have bad/poor camera angles, making them difficult to recognize. Therefore, we propose a learning mechanism to find one camera projection that best describes the 3D objects. In doing so, we adopt a model reprogramming approach that perturbs the input 2D images (from the best projection angle) with background-foreground colorization (segmentation) without modifying or fine-tuning the original CLIP architectures. Our reprogramming approach is different from conventional approaches because our custom loss produces a uniform background color instead of the noisy background of previous works. In the text pipeline, we use the CLIP text encoder and train the whole architecture end-to-end. Instead of using a single class name-based text input, we suggest using prompts based on GPT-based description. We propose a prompt engineering approach to train the networks with 3D object and text description pairs. Finally, we evaluate our proposed method using 3D synthetic datasets, ModelNet and ShapeNet, and two 3D real-world-scanned datasets, ScanObjectNN and CO3D, and report state-of-the-art performances. In future, we will investigate model reprogramming for other vision problems beyond FSCIL, like segmentation and domain adaptation on 3D point clouds.", + "bbox": [ + 212, + 507, + 787, + 809 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgment. Mehrtash Harandi thanks the Australian Research Council for support through the Discovery Program (DP230101176).", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "A. Cheraghian et al.", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bansal, N., Chen, X., Wang, Z.: Can we gain more from orthogonality regularizations in training deep networks? In: Bengio, S., Wallach, H., Larochelle, H., Grauman, K., Cesa-Bianchi, N., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 31. Curran Associates, Inc. (2018), https://proceedings.neurips.cc/paper_files/paper/2018/file/bf424cb7b0dea050a42b9739eb261a3a-Paper.pdf", + "2. Belouadah, E., Popescu, A.: II2m: Class incremental learning with dual memory. In: CVPR (2019)", + "3. Belouadah, E., Popescu, A.: Scail: Classifier weights scaling for class incremental learning. In: WACV (2020)", + "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 1877-1901. Curran Associates, Inc. (2020), https://proceedings.neurips.cc/paper_files/paper/2020/file/1457c0d6bcbd4967418bf8ac142f64a-Paper.pdf", + "5. Castro, F.M., Marín-Jiménez, M.J., Guil, N., Schmid, C., Alahari, K.: End-to-end incremental learning. In: ECCV (2018)", + "6. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015)", + "7. Chen, K., Lee, C.G.: Incremental few-shot learning via vector quantization in deep embedded space. In: ICLR (2021)", + "8. Chen, P.Y.: Model reprogramming: Resource-efficient cross-domain machine learning (2023)", + "9. Cheraghian, A., Rahman, S., Fang, P., Roy, S.K., Petersson, L., Harandi, M.: Semantic-aware knowledge distillation for few-shot class-incremental learning. In: CVPR (2021)", + "0. Cheraghian, A., Rahman, S., Ramasinghe, S., Fang, P., Simon, C., Petersson, L., Harandi, M.: Synthesized feature based few-shot class-incremental learning on a mixture of subspaces. In: ICCV (2021)", + "1. Chowdhury, T., Cheraghian, A., Ramasinghe, S., Ahmadi, S., Saberi, M., Rahman, S.: Few-shot class-incremental learning for 3d point cloud objects. In: Avidan, S., Brostow, G., Cissé, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 204-220. Springer Nature Switzerland, Cham (2022)", + "2. Chowdhury, T., Cheraghian, A., Ramasinghe, S., Ahmadi, S., Saberi, M., Rahman, S.: Few-shot class-incremental learning for 3d point cloud objects. In: ECCV (2022)", + "3. Dinh, T., Seo, D., Du, Z., Shang, L., Lee, K.: Improved input reprogramming for gan conditioning (2022)", + "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021)", + "5. Elsayed, G.F., Goodfellow, I., Sohl-Dickstein, J.: Adversarial reprogramming of neural networks. In: International Conference on Learning Representations (2019), https://openreview.net/forum?id=Syx_Ss05tm" + ], + "bbox": [ + 225, + 172, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Canonical Shape Projection for 3D FSCIL", + "bbox": [ + 447, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Jaderberg, M., Simonyan, K., Zisserman, A., kavukcuoglu, k.: Spatial transformer networks. In: Cortes, C., Lawrence, N., Lee, D., Sugiyama, M., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 28. Curran Associates, Inc. (2015), https://proceedings.neurips.cc/paper_files/paper/2015/file/33ceb07bf4eeb3da587e268d663aba1a-Paper.pdf", + "17. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 4904-4916. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/jia21b.html", + "18. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision (ECCV) (2022)", + "19. Lee, D.H., Pujara, J., Sewak, M., White, R.W., Jauhar, S.K.: Making large language models better data creators. In: The 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP) (2023), https://openreview.net/forum?id=2Rdfdri2oT", + "20. Lee, K.Y., Zhong, Y., Wang, Y.X.: Do pre-trained models benefit equally in continual learning? In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 6485-6493 (January 2023)", + "21. Li, Y., Bu, R., Sun, M., Wu, W., Di, X., Chen, B.: Pointcnn: Convolution on x-transformed points. In: NeurIPS (2018)", + "22. Li, Z., Hoiem, D.: Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence (2018)", + "23. Liu, Y., Fan, B., Xiang, S., Pan, C.: Relation-shape convolutional neural network for point cloud analysis. In: CVPR (2019)", + "24. Mazumder, P., Singh, P., Rai, P.: Few-shot lifelong learning. In: AAAI (2021)", + "25. Pei, Y., Qing, Z., CEN, J., Wang, X., Zhang, S., Wang, Y., Tang, M., Sang, N., Qian, X.: Learning a condensed frame for memory-efficient video class-incremental learning. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=1CGYC7pXWNQ", + "26. Poulenard, A., Rakotosaona, M.J., Ponty, Y., Ovsjanikov, M.: Effective rotation-invariant point cnn with spherical harmonics kernels. In: 3DV (2019)", + "27. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR (2017)", + "28. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: NeurIPS (2017)", + "29. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 8748-8763. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/radford21a.html", + "30. Rao, Y., Lu, J., Zhou, J.: Spherical fractal convolutional neural networks for point cloud recognition. In: CVPR (2019)", + "31. Reizenstein, J., Shapovalov, R., Henzler, P., Sbordone, L., Labatut, P., Novotny, D.: Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In: ICCV (2021)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "A. Cheraghian et al.", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F. (eds.) Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015. pp. 234-241. Springer International Publishing, Cham (2015)", + "33. Singh, A., Hu, R., Goswami, V., Couairon, G., Galuba, W., Rohrbach, M., Kiela, D.: FLAVA: A foundational language and vision alignment model. In: CVPR (2022)", + "34. Sung, F., Yang, Y., Zhang, L., Xiang, T., Torr, P.H., Hospedales, T.M.: Learning to compare: Relation network for few-shot learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2018)", + "35. Tan, Y., Xiang, X.: Cross-domain few-shot incremental learning for point-cloud recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 2307-2316 (January 2024)", + "36. Tan, Z., Ding, K., Guo, R., Liu, H.: Graph few-shot class-incremental learning. In: Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining (2022)", + "37. Tao, X., Hong, X., Chang, X., Dong, S., Wei, X., Gong, Y.: Few-shot class-incremental learning. In: CVPR (2020)", + "38. Tsai, Y.Y., Chen, P.Y., Ho, T.Y.: Transfer learning without knowing: Reprogramming black-box machine learning models with scarce data and limited resources. In: III, H.D., Singh, A. (eds.) Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 9614-9624. PMLR (13-18 Jul 2020), https://proceedings.mlr.press/v119/tsai20a.html", + "39. Tsai, Y.Y., Chen, P.Y., Ho, T.Y.: Transfer learning without knowing: Reprogramming black-box machine learning models with scarce data and limited resources. In: Proceedings of the 37th International Conference on Machine Learning. ICML'20, JMLR.org (2020)", + "40. Uy, M.A., Pham, Q.H., Hua, B.S., Nguyen, D.T., Yeung, S.K.: Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In: ICCV (2019)", + "41. Wang, C., Samari, B., Siddiqi, K.: Local spectral graph convolution for point set feature learning. In: ECCV (2018)", + "42. Wang, R., Duan, X., Kang, G., Liu, J., Lin, S., Xu, S., Lu, J., Zhang, B.: Attriclip: A non-incremental learner for incremental knowledge learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3654-3663 (June 2023)", + "43. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog) (2019)", + "44. Wu, W., Qi, Z., Fuxin, L.: PointCONV: Deep convolutional networks on 3D point clouds. In: CVPR (2019)", + "45. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3D ShapeNets: A deep representation for volumetric shapes. In: CVPR (2015)", + "46. Xiang, T., Zhang, C., Song, Y., Yu, J., Cai, W.: Walk in the cloud: Learning curves for point clouds shape analysis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 915-924 (October 2021)", + "47. Xu, Y., Fan, T., Xu, M., Zeng, L., Qiao, Y.: SpiderCNN: Deep learning on point sets with parameterized convolutional filters. In: ECCV (2018)", + "48. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. arXiv preprint arXiv:2112.02413 (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Canonical Shape Projection for 3D FSCIL", + "bbox": [ + 447, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. Zhang, Y., Rabbat, M.: A graph-cnn for 3d point cloud classification. In: ICASSP (2018)", + "50. Zhou, D.W., Wang, F.Y., Ye, H.J., Ma, L., Pu, S., Zhan, D.C.: Forward compatible few-shot class-incremental learning. In: CVPR (2022)", + "51. Zhu, X., Zhang, R., He, B., Guo, Z., Zeng, Z., Qin, Z., Zhang, S., Gao, P.: Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. arXiv preprint arXiv:2211.11682 (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 244 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "A. Cheraghian et al.", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Canonical Shape Projection is All You Need for 3D Few-shot Class Incremental Learning/354b1d2e-f84c-4d0d-a2a6-9954d66577ef_model.json b/2024/Canonical Shape Projection is All You Need for 3D Few-shot Class Incremental Learning/354b1d2e-f84c-4d0d-a2a6-9954d66577ef_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b9e3f60ab21240bef67cf27bc98d00d5cb9fa39f --- /dev/null +++ b/2024/Canonical Shape Projection is All You Need for 3D Few-shot Class Incremental Learning/354b1d2e-f84c-4d0d-a2a6-9954d66577ef_model.json @@ -0,0 +1,2194 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.229, + 0.141, + 0.777, + 0.187 + ], + "angle": 0, + "content": "Canonical Shape Projection is All You Need for 3D Few-shot Class Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.212, + 0.765, + 0.258 + ], + "angle": 0, + "content": "Ali Cheraghian\\*,1,2, Zeeshan Hayden\\*,1,2, Sameera Ramasinghe\\*3, Shafin Rahman4, Javad Jafaryahya5, Lars Petersson2, and Mehrtash Harandi6" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.269, + 0.776, + 0.342 + ], + "angle": 0, + "content": "1 Data61, CSIRO, Australia, 2 Australian National University, 3 Amazon, 4 North South University, 5 University of Technology Sydney, 6 Monash University {ali.cheraghian, zeeshan.hayder, lars.petersson}@data61.csiro.au, ramasisa@amazon.com, shafin.rahman@northsouth.edu, javad.jafaryahya@student.uts.edu.au, mehrtash.harandi@monash.edu" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.38, + 0.741, + 0.684 + ], + "angle": 0, + "content": "Abstract. In recent years, robust pre-trained foundation models have been successfully used in many downstream tasks. Here, we would like to use such powerful models to address the problem of few-shot class incremental learning (FSCIL) tasks on 3D point cloud objects. Our approach is to reprogram the well-known CLIP-based foundation model (trained on 2D images and text pairs) for this purpose. The CLIP model works by ingesting 2D images, so to leverage it in our context, we project the 3D object point cloud onto 2D image space to create proper depth maps. For this, prior works consider a fixed and non-trainable set of camera poses. In contrast, we propose to train the network to find a projection that best describes the object and is appropriate for extracting 2D image features from the CLIP vision encoder. Directly using the generated depth map is not suitable for the CLIP model, so we apply the model reprogramming paradigm to the depth map to augment the foreground and background to adapt it. This removes the need for modification or fine-tuning of the foundation model. In the setting we have investigated, we have limited access to data from novel classes, resulting in a problem with overfitting. Here, we address this problem via the use of a prompt engineering approach using multiple GPT-generated text descriptions. Our method, C3PR, successfully outperforms existing FSCIL methods on ModelNet, ShapeNet, ScanObjectNN, and CO3D datasets. The code is available at https://github.com/alichr/C3PR." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.698, + 0.741, + 0.726 + ], + "angle": 0, + "content": "Keywords: 3D shape projection \\(\\cdot\\) Model reprogramming \\(\\cdot\\) Few-shot class incremental learning" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.753, + 0.377, + 0.769 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.786, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Maintaining a balance between acquiring new concepts and retaining existing knowledge presents a significant challenge for machine learning algorithms, a" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.416, + 0.84 + ], + "angle": 0, + "content": "* denotes equal contribution." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.411, + 0.129 + ], + "angle": 0, + "content": "A. Cheraghian et al." + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.147, + 0.703, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.463, + 0.79, + 0.577 + ], + "angle": 0, + "content": "Fig. 1: (a) Traditional methods using image-based models for point cloud processing often generate multiple depth maps from fixed camera poses, potentially causing feature alignment issues and difficulties in recognizing 3D objects from unsuitable camera angles. (b) In contrast, our proposed approach (C3PR) determines the most suitable camera pose for projecting a single depth map by leveraging feedback from the foundation model. (c) To improve the foundation model's performance in handling 3D data for the FSCIL task, we introduce a novel model reprogramming technique (without modifying or fine-tuning the foundation model) to further refine the depth map." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.788, + 0.793 + ], + "angle": 0, + "content": "task that comes naturally to humans. This task becomes more difficult when only a few samples are available for new concepts. In the literature, this problem is referred to as few-shot class incremental learning (FSCIL) [7,9,10,24,37]. Substantial effort on this topic has been invested, particularly on image data [7,9,10, 24,37], whereas relatively few studies are tailored specifically towards point cloud data. For example, in the 2D domain, there are vision-language models such as CLIP [29], ALIGN [17], and FLAVA [33] models with an inherent capacity to generate highly versatile representations and demonstrate zero-shot capabilities, all without parameter fine-tuning [20, 25, 42]. Gathering training data for 3D point clouds is substantially more challenging than for image data, and here we explore whether the CLIP model can help alleviate this issue. In this paper, we attempt to leverage CLIP for FSCIL on 3D Point Cloud Objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Popular approaches capture 2D images from a fixed set of camera poses and then pass the images to 2D foundation models for feature extraction [48, 51]. The main drawback of such a strategy is that camera poses are not adaptable" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Canonical Shape Projection for 3D FSCIL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.359 + ], + "angle": 0, + "content": "to the target dataset, hindering capturing the objects in poses ideal for feature extraction (see Fig. 1 (a)). To address this problem, we propose to learn the camera projection pose based on the dataset that best describes the 3D objects (see Fig. 1 (b)). Our proposed model reprogramming approach is advantageous for representing real-world scanned 3D objects where point cloud data contains sensor noise (see Fig. 1 (c)). The unique camera pose can still capture the global shape of noisy 3D objects, even though the object belongs to novel classes. For language/text modality, many existing works [9-11] leverages vision-language learning to address the overfitting issue associated with few-shot novel classes. We argue that vital knowledge lies within the language domain, which may aid in the training process of the vision domain, particularly for few-shot classes, to alleviate overfitting. However, thus far, existing methods have only utilized class names from the language domain [9, 10, 12]. In contrast, we propose a method based on prompt engineering to better leverage the language domain." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.361, + 0.788, + 0.755 + ], + "angle": 0, + "content": "This paper presents a novel approach for extracting valuable insights from foundational models like CLIP, focusing on FSCIL on 3D point cloud data. As CLIP's training data comprises millions of image-text pairs, our target is to transfer this learned knowledge as a complete black box (without knowing or modifying pre-trained parameters) for 3D data. To this end, we propose reprogramming [38] the CLIP model, specifically for point cloud objects. The proposed approach is different from prior attempts [48,51] where they used a fixed set of camera poses for 3D to 2D image projection and did not augment the input data for better usage of the CLIP encoder. Our C3PR method emphasizes generating depth map images with consistent, informative, background pixels, diverging from conventional model reprogramming techniques that frequently produce random pixel information for the background. We observed that maintaining consistent background pixels contributes to improved generalization, particularly for few-shot novel classes. These steps indirectly add essential perturbation to the input depth map images so that the black box CLIP vision encoder can operate without fine-tuning. Furthermore, we propose a novel approach that leverages information from the language domain using a Large Language Model, GPT, to combat overfitting issues in few-shot novel classes. Specifically, we generate multiple descriptive prompts (instead of single class name-based descriptions as used in [11]) for base and novel classes. It also compensates for the issue with limited data during novel class learning. The generated prompts are then employed in the CLIP text encoder, reducing overfitting during incremental training sessions. Both the vision and text pipelines of our method can be trained end-to-end. We report state-of-the-art performances when evaluated on 3D synthetic datasets, ModelNet [45] and ShapeNet [6], and two 3D real world-scanned datasets, ScanObjectNN [40] and Common Objects in 3D (CO3D) [31]." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.757, + 0.621, + 0.772 + ], + "angle": 0, + "content": "The main contributions of our proposed method are:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.792, + 0.781, + 0.808 + ], + "angle": 0, + "content": "- Learning the preferred object-based camera pose for 3D to 2D projections;" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- A novel model reprogramming approach for 2D CLIP encoders targeting 3D point cloud data;" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.792, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.41, + 0.129 + ], + "angle": 0, + "content": "A. Cheraghian et al." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.147, + 0.788, + 0.178 + ], + "angle": 0, + "content": "- Prompt engineering based on a GPT model for novel few-shot classes to mitigate overfitting." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.202, + 0.383, + 0.218 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.236, + 0.789, + 0.492 + ], + "angle": 0, + "content": "3D point cloud processing: In recent years, there has been a surge in research focused on leveraging deep learning for the direct processing of 3D point cloud objects. Notable contributions include PointNet [27] by Qi et al., which introduced the use of multi-layer perceptron (MLP) networks for 3D point cloud processing. However, this approach overlooked the inherent local structures within the input data. Subsequent techniques were developed to address this limitation, including PointNet++ [28], which hierarchically extracts features to take advantage of local information. Building upon this, several works [21,23,26,30,44,47] have proposed convolution strategies for extracting local information. PointConv [44] introduced a novel convolution operation that locates a Monte Carlo estimation of the hidden continuous 3D convolution based on a powerful sampling technique. SFCNN [30] proposed graph convolution on spherical points. Other papers [41,43,49] treat each point cloud as a graph vertex to learn features in either spatial or spectral domains. DGCNN [43] constructs a graph in feature space and dynamically updates it using MLP for each edge, while PointGCN [49] suggests an approach that generates the graph using k-nearest neighbors from a point cloud to capture local information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.789, + 0.765 + ], + "angle": 0, + "content": "Few-shot class-incremental learning: Tao et al. [37] were the pioneers in introducing the FSCIL setting for image data. They proposed the use of a neural gas network to address the issue of forgetting by preserving the topology of classes based on their feature vectors. Following this, in [10], a mixture of the subspaces are generated based on the training data distribution. In a related vein, Mazumder et al [24] devised a method that selectively adapts a subset of model parameters for learning novel classes, thereby curbing overfitting. They also combat forgetting by freezing critical parameters in the model. Cheraghian et al. [9] proposed a novel vision-language strategy, incorporating class semantic information from language space using a distillation technique to alleviate the impact of catastrophic forgetting. They also integrated an attention mechanism to counter overfitting on few-shot novel tasks. Introduced in [12], the Specific FSCIL tailored for 3D point cloud data presents the innovative concept of Microshape, which captures essential details from the primary task, effectively addressing both forgetting and overfitting challenges. Additionally, the work by [35] delves into cross-domain FSCIL applied to point-cloud recognition. This study dissects catastrophic forgetting into two distinct components: base class forgetting and incremental class forgetting, offering separate mitigation strategies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.79, + 0.84 + ], + "angle": 0, + "content": "Model reprogramming: Foundation model reprogramming has recently gained attention [8], primarily due to its reduced computational and time requirements compared to the fine-tuning approach. Prior to this, Elsayed et al. [15] introduced adversarial reprogramming. This technique involves mapping ImageNet labels to task labels, aiming to perplex the ImageNet classifier and induce be" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Canonical Shape Projection for 3D FSCIL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "havior akin to a CIFAR-10 or MNIST classifier. Dinh et al. [13] proposed that input reprogramming is more effective in situations with limited labeled data than training from scratch and fine-tuning. Tsai et al. [39] introduced a technique called black-box adversarial reprogramming. This method involves transforming the input and integrating label mapping mechanisms into the output. In this paper, we present a novel foundation model reprogramming approach tailored for the CLIP model." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.276, + 0.331, + 0.291 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.308, + 0.79, + 0.538 + ], + "angle": 0, + "content": "In this section, we first introduce the FSCIL problem settings. Let's consider a sequence of \\(T\\) tasks denoted by \\(\\mathbf{Q} = \\{\\mathcal{Q}^1,\\mathcal{Q}^2,\\dots ,\\mathcal{Q}^T\\}\\), where \\(\\mathcal{V}^t\\) represents the set of classes in task \\(\\mathcal{Q}^t\\), and \\(\\mathcal{V}^i\\cap \\mathcal{V}^j = \\emptyset\\). Additionally, each class in all tasks is associated with a set of prompt class descriptions generated by an LLM, e.g., GPT [4], denoted as \\(\\mathcal{P}^t\\). Thus, we can represent each task as a tuple \\(\\mathcal{Q}^t = \\{\\mathcal{X}_i^t,\\mathbf{y}_i^t,\\mathbf{p}_i^t\\}_{i = 1}^{n_t}\\), where \\(\\mathcal{X}_i^t = \\{\\mathbf{x}_{i,j}^t\\}_{j = 1}^l\\) refers to a 3D point cloud object with coordinates \\(\\mathbf{x}_{i,j}^{t}\\in \\mathbb{R}^{3}\\). Moreover, \\(\\mathbf{y}_i^t\\in \\mathcal{V}^t\\) and \\(\\mathbf{p}_i^t\\in \\mathcal{P}^t\\) correspond to the label of the point cloud and its corresponding class prompt description, respectively. In the proposed FSCIL framework, \\(\\mathcal{Q}^1\\) is the base task (\\(t = 1\\)), where the model undergoes training on a large-scale synthetic 3D dataset. For \\(t > 1\\), the training data are drawn from real-world 3D point clouds consisting of only a few instances. The model is trained sequentially across the tasks \\(t = 1,\\ldots ,T\\). However, during the \\(t\\)-th task, the model is exposed to \\(\\mathcal{X}^t\\), \\(\\mathbf{y}^t\\), and \\(\\{\\mathcal{P}^1,\\mathcal{P}^2,\\dots ,\\mathcal{P}^t\\}\\). During inference, the model trained on the current task \\(\\mathcal{Q}^t\\) is expected to classify test samples from both the current and previous tasks, i.e., \\(\\{\\mathcal{Q}^1,\\mathcal{Q}^2,\\dots ,\\mathcal{Q}^t\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.559, + 0.395, + 0.573 + ], + "angle": 0, + "content": "3.1 Model overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.844 + ], + "angle": 0, + "content": "In this section, we present an overview of our proposed methodology, illustrated in Fig. 2. We begin with a point cloud sample denoted as \\(\\mathcal{X}_i^t\\), which is passed through the projection module \\(F\\). This module utilizes the rotation matrix \\(\\mathbf{R}_i^t\\) derived from the transformation module \\(T_r\\) to generate a depth map labelled as \\(\\mathbf{I}_{D_i}^t\\). For the projection module \\(F\\), we use the method proposed by [51], which is fixed during the training stage. The transformation module \\(T_r\\) is crucial for determining the appropriate angle \\(\\theta\\) for projection. Additionally, a soft constraint loss function \\(\\mathcal{L}_c\\) is applied to the rotation matrix to ensure that the transformation module generates a valid rotation matrix. Following this, we extract a binary mask \\(\\mathbf{I}_{M_i}^t\\) from the depth map. Within this mask \\(M\\), background pixels are assigned a value of zero, while non-zero depth map values are set to one. Subsequently, this mask is fed into a model reprogramming module \\(C\\) to add pixel information to the background of the depth map. To ensure that the generated embedding does not converge to a noisy image and follows the pattern of the mask input, an image gradient consistency loss function \\(\\mathcal{L}_g\\) is used between the input and output of the module \\(C\\). The resulting image embedding \\(\\mathbf{I}_{H_i}^t\\) is then merged with the original depth map in the subsequent stage, resulting in" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.129 + ], + "angle": 0, + "content": "A. Cheraghian et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.788, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.309, + 0.789, + 0.482 + ], + "angle": 0, + "content": "Fig. 2: Beginning with a point cloud sample \\(\\mathcal{X}_i^t\\), we use the projection module \\(F\\) to generate a depth map \\(\\mathbf{I}_{D_i}^t\\) via the rotation matrix \\(\\mathbf{R}_i^t\\) from the transformation module \\(T_r\\). A mask \\(\\mathbf{I}_{M_i}^t\\) is then extracted from the depth map. This mask is input into module \\(C\\) and the resulting image is merged with the depth map, yielding the image \\(\\mathbf{I}_{R_i}^t\\). Next, \\(\\mathbf{I}_{R_i}^t\\) is input to the CLIP vision encoder model \\(V_e\\) to generate a vision embedding \\(\\mathbf{v}_i^t \\in \\mathbb{R}^m\\). Simultaneously, \\(\\mathcal{X}_i^t\\) is fed into \\(B\\) to create a point cloud representation \\(\\mathbf{z}_i^t \\in \\mathbb{R}^m\\). Moreover, the GPT model generates multiple class name descriptions, which are then stored in a prompt bank. To this end, prompts \\(\\{\\mathbf{p}_1, \\mathbf{p}_2, \\dots, \\mathbf{p}_{|\\mathcal{Y}_{tl}|}\\}\\) for all old classes and the current task \\(t\\) are forwarded into the text encoder \\(T_e\\), producing corresponding text embeddings \\(\\{\\mathbf{e}_1, \\mathbf{e}_2, \\dots, \\mathbf{e}_{|\\mathcal{Y}_{tl}|}\\}\\), where \\(\\mathbf{e}_i^t \\in \\mathbb{R}^m\\). Finally, \\(\\mathbf{v}_i^t\\), \\(\\mathbf{z}_i^t\\), and \\(\\mathbf{e}_j^t\\) are concatenated and passed through the relation module \\(R_l\\) to generate a scalar value \\(r_i^t\\), serving as a similarity score." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.789, + 0.765 + ], + "angle": 0, + "content": "the production of a reprogrammed depth map image denoted as \\(\\mathbf{I}_{R_i}^t\\). After that, \\(\\mathbf{I}_{R_i}^t\\) is input into the vision encoder of the CLIP model \\(V_{e}\\) to generate a vision embedding represented by \\(\\mathbf{v}_i^t\\in \\mathbb{R}^m\\). Simultaneously, the point cloud input \\(\\mathcal{X}_i^t\\) is forwarded to a 3D point cloud backbone, for instance, Curvenet [46], to generate a feature description \\(\\mathbf{z}_i^t\\in \\mathbb{R}^m\\). At the same time, GPT is employed to generate multiple descriptions for each class using the command: describe a depth map of a [class], thereby creating several prompts, which are stored in the prompt bank module. Then, the prompts of all classes \\(\\{\\mathbf{p}_j\\}_{j = 1}^{|\\mathcal{Y}_{tl}|}\\), where \\(\\mathcal{Y}_{tl} = \\bigcup_{i = 1}^{t}\\mathcal{Y}^{i}\\), observed up to task \\(t\\) are fed into the text encoder \\(T_{e}\\) to generate the corresponding text embeddings denoted as \\(\\{\\mathbf{e}_j\\}_{j = 1}^{|\\mathcal{Y}_{tl}|}\\), where each \\(\\mathbf{e}_j^t\\in \\mathbb{R}^m\\). The prompts are generated using LLMs, such as GPT [4], given a command: Describe a depth map of a [class]. In the subsequent stage, all \\(\\mathbf{v}_i^t\\), \\(\\mathbf{z}_i^t\\), and \\(\\mathbf{e}_i^t\\) are concatenated and passed through the relation module [34] \\(R_{l}\\) to generate a scalar value \\(r_i^t\\) ranging from 0 to 1. This value serves as a measure of the similarity between the visual and prompt feature embeddings. Finally, a binary cross-entropy loss function \\(\\mathcal{L}_r\\) is used to train the entire proposed pipeline." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.785, + 0.517, + 0.801 + ], + "angle": 0, + "content": "3.2 Canonical 3D shape projection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "A 3D-to-2D projection matrix can be designed to project any 3D point cloud on a 2D plane considering any camera angle. Nevertheless, the feature embedding" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Canonical Shape Projection for 3D FSCIL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.145, + 0.703, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.3, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Fig. 3: We demonstrate the sensitivity of the CLIP model to the angle of depth map projection. Here, we consider two random projections to generate depth maps. Additionally, we assume that there are three classes: car, guitar, and person. Even though the number of classes is limited to only three, the CLIP model can only correctly classify two of these depth maps. This highlights the importance of finding the optimum projection angle to fully leverage the CLIP model for 3D point cloud processing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.789, + 0.611 + ], + "angle": 0, + "content": "produced by the visual encoder of large image-based foundation models (e.g., CLIP) can be affected by the content that can be controlled by the projection angle. Figure 3 shows that an adjustment is pivotal due to the CLIP model's sensitivity to the projection angle, especially during depth map processing. We are interested in finding the optimal projection angle to enhance the performance of the CLIP model. When projecting from a 3D point cloud to 2D images, we leverage the mapping function introduced by [51]. We provide optimum projection angle \\(\\theta\\) via a rotation matrix \\(\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}\\), where \\(\\mathbf{R} = T_r(\\mathcal{X})\\), generated by the transformation module \\(T_r\\) to ascertain the most effective projection angle for converting a 3D point cloud into a depth map. The generated \\(\\mathbf{R}\\) is then applied to the given 3D point to derive the appropriate angle for the projection module \\(\\theta\\). To ensure that \\(\\mathbf{R}\\) is a valid rotation matrix, we apply the following loss function at the output of the transformation module \\(T_r\\)," + }, + { + "type": "equation", + "bbox": [ + 0.338, + 0.622, + 0.787, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(\\left\\| \\mathbf {R} _ {i} ^ {T} \\mathbf {R} _ {i} - \\mathbf {I} \\right\\| _ {\\mathcal {F}} ^ {2} + \\left| \\det (\\mathbf {R} _ {i}) - 1 \\right|\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.669, + 0.788, + 0.749 + ], + "angle": 0, + "content": "where \\(\\mathbf{R}_i = T_r(\\mathcal{X}_i)\\), \\(n\\) is the number of samples in the training set, \\(\\| \\cdot \\|_{\\mathcal{F}}^2\\) denotes the Frobenius norm, \\(\\mathbf{I}\\) is the identity matrix. The orthogonality condition \\(\\left\\| \\mathbf{R}_i^T\\mathbf{R}_i - \\mathbf{I}\\right\\|_F^2\\) guarantee that the columns of \\(\\mathbf{R}_i\\) form an orthonormal basis to preserve distances and angles for rotations. \\(\\left|\\operatorname*{det}(\\mathbf{R}_i) - 1\\right|\\) affirms that \\(\\mathbf{R}_i\\) is an orientation-preserving transformation matrix and doesn't involve any reflection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In practice, the \\( T_{r} \\) matrix may not be guaranteed to be a rotation matrix on a per-batch basis as the above condition is only a soft constraint. This is overcome by continuous re-orthogonalization during the training process. Moreover, alternative methods like euler-angle or quaternion representations could be used. Nevertheless, the proposed approach of enforcing orthogonality through the \\( L_{c} \\) is a known method in the literature [1]. In the end, when provided with a 3D" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.129 + ], + "angle": 0, + "content": "A. Cheraghian et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.209 + ], + "angle": 0, + "content": "point cloud sample \\(\\mathcal{X}_i\\), the transformation module \\(T_r\\) computes the desired rotation matrix \\(\\mathbf{R}_i\\) while taking into account all the aforementioned constraints to ensure its validity. This process results in the generation of a learned depth map image, \\(\\mathbf{I}_{D_i} = F(\\mathcal{X}_i,\\mathbf{R}_i)\\)" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.231, + 0.45, + 0.248 + ], + "angle": 0, + "content": "3.3 Model reprogramming" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.258, + 0.789, + 0.501 + ], + "angle": 0, + "content": "Model reprogramming of computer vision models, especially foundational ones, is a crucial step in their effective deployment, enabling a seamless transition to new tasks and domains with minimal adjustments to existing architectures [38]. We present a novel model reprogramming approach specifically tailored for the CLIP model applied to the FSCIL task on 3D point cloud objects. It is worth noting that the CLIP model is primarily trained on RGB images rather than depth maps. Therefore, to enhance the CLIP model's effectiveness with depth maps, we incorporate pixel information into the background of the depth map images. More specifically, this modification assists the CLIP model in accurately discerning the boundary of the depth map shape information through a \\(C\\) module, which takes the form of a U-Net [32]. To accomplish this, we initially compute the mask map \\(\\mathbf{I}_{M_i}\\) corresponding to a given depth map \\(\\mathbf{I}_{D_i}\\), which isolates the shape boundary from the background. Next, we input \\(\\mathbf{I}_{M_i}\\) into the \\(C\\) module to extract the pixel information for the background, denoted as \\(\\mathbf{I}_{H_i}\\). To ensure the consistency of the generated pixels in \\(\\mathbf{I}_{M_i}\\) from the \\(C\\) module and to prevent convergence to arbitrary values, we employ the following loss function," + }, + { + "type": "equation", + "bbox": [ + 0.323, + 0.511, + 0.787, + 0.55 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {g} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(\\frac {\\partial \\mathbf {I} _ {M _ {i}}}{\\partial x} - \\frac {\\partial \\mathbf {I} _ {H _ {i}}}{\\partial x}\\right) ^ {2} + \\left(\\frac {\\partial \\mathbf {I} _ {M _ {i}}}{\\partial y} - \\frac {\\partial \\mathbf {I} _ {H _ {i}}}{\\partial y}\\right) ^ {2} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.562, + 0.788, + 0.608 + ], + "angle": 0, + "content": "where \\(\\frac{\\partial}{\\partial x}\\) and \\(\\frac{\\partial}{\\partial y}\\) denote the derivatives of the images in the \\(x\\) and \\(y\\) directions, respectively. This loss enforces a smoothness constraint on the gradient to separate it from foreground objects explicitly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.609, + 0.788, + 0.671 + ], + "angle": 0, + "content": "It helps to learn novel classes with few examples. Subsequently, we combine the depth map \\(\\mathbf{I}_{D_i}\\) from the projection module \\(F\\) with the extracted values \\(\\mathbf{I}_{H_i}\\) from \\(C\\) to generate the final image, \\(\\mathbf{I}_{R_i} = \\mathbf{I}_{D_i} \\odot C(M(\\mathbf{I}_{D_i}))\\), where \\(\\odot\\) represents the element-wise multiplication." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.693, + 0.429, + 0.71 + ], + "angle": 0, + "content": "3.4 Prompt engineering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Advancements in prompt engineering have unveiled significant potential, primarily driven by LLMs [18, 19]. LLMs can extract richer information from appropriately tailored prompts. With this motivation, we employ GPT to obtain more insights from plain class names and enhance the FSCIL task on 3D data. In contrast to prevailing methods [9, 10, 12] that rely solely on the class name for the FSCIL task, we extend our method by generating additional descriptions to extract more semantic information from the CLIP text encoder (denoted as \\( T_{e} \\)). This augmentation ensures a more comprehensive utilization of the LLM's" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Canonical Shape Projection for 3D FSCIL" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.268 + ], + "angle": 0, + "content": "capabilities, thereby enhancing the overall efficacy of our approach. In the FS-CIL, a significant challenge arises due to the scarcity of training samples for novel classes, potentially leading to overfitting. To address this issue, we propose an innovative strategy leveraging the capabilities of LLM, specifically the GPT model, and utilizing this as novel prompts for the text encoder of the CLIP model. More precisely, we utilize the GPT model to generate multiple descriptions for each class by providing it with a specific command. These prompts are formally defined for each sample as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.277, + 0.785, + 0.293 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {i} = \\operatorname {G P T} (\\text {C o m m a n d s}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.785, + 0.404 + ], + "angle": 0, + "content": "Here, the Commands are structured as Describe a depth of a [class]. For every class in the dataset, we substitute the respective class names in place of '[class]' in the command and input them into GPT-4. This results in the generation of depth map-specific descriptions with comprehensive category-wise semantics. Subsequently, we feed the descriptions of each category through CLIP's textual encoder \\(\\mathcal{T}_e\\). Finally, the generated prompts \\(\\mathbf{p}_i\\) are used in the training of the few-shot classes." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.423, + 0.482, + 0.438 + ], + "angle": 0, + "content": "3.5 Class incremental learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.788, + 0.72 + ], + "angle": 0, + "content": "Training. For the base task, denoted as \\(\\mathcal{Q}^1\\), we conduct training for the transformation model \\(T_{r}\\), the 3D backbone \\(B\\), and the relation module \\(R_{l}\\). In the case of subsequent tasks, denoted as \\(\\mathcal{Q}^t\\), where \\(t > 1\\), only the relation module undergoes fine-tuning. Throughout both base and subsequent novel tasks, all other components in our proposed architecture (refer to Figure 2) remain frozen. Additionally, for each previously learned class, we randomly select a sample to be stored in a compact memory module \\(\\mathcal{M}\\). Given the \\(i^{th}\\) point cloud \\(\\mathcal{X}_i^t\\) of the \\(t^{th}\\) task, the \\(B\\) module generates the features \\(\\mathbf{z}_i^t\\) and the \\(V_{e}\\) module produces \\(\\mathbf{v}_i^t\\). Simultaneously, the prompts of all classes for tasks \\(\\mathcal{Q}^1,\\ldots ,\\mathcal{Q}^t\\) are fed into the text encoder \\(T_{e}\\) to generate the corresponding text embeddings denoted as \\(\\{\\mathbf{e}_j\\}_{j = 1}^{|\\mathcal{Y}_{tl}|}\\). Then, we concatenate \\(\\mathbf{z}_i^t\\), \\(\\mathbf{v}_i^t\\), and \\(\\mathbf{e}_j^t\\) and forward them into the relation module \\(R_{l}\\), which provides a score between [0,1], representing the similarity between them. In other words, for each training sample, we generate a score against each of the classes in both the subsequent and previous tasks as \\(r_{ij}^{t} = \\gamma \\circ R\\circ (\\mathbf{z}_{i}^{t}\\oplus \\mathbf{v}_{i}\\oplus \\mathbf{e}_{j}), j\\in \\mathcal{Y}_{tl}\\), where \\(\\mathcal{Y}_{tl} = \\bigcup_{i = 1}^{t}\\mathcal{Y}^{i}\\), \\(\\oplus\\) is the concatenation operator, \\(R\\) is the relation module, and \\(\\gamma\\) is the sigmoid function. Finally, for each feature \\(r_{ij}\\) and the corresponding ground truth \\(\\mathbf{y}_i\\), we employ a binary cross-entropy cost function to train the model as follows," + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.739, + 0.795, + 0.779 + ], + "angle": 0, + "content": "\\[\nL _ {r} = - \\frac {1}{| \\mathcal {Y} _ {t l} | | S |} \\sum_ {k \\in \\mathcal {Y} _ {t l}} \\sum_ {\\mathbf {y} _ {i} \\in \\mathcal {S}} \\Bigg (1 (\\boldsymbol {y} _ {i} ^ {t} = = k) \\log (r _ {i k}) + \\big (1 - 1 (\\boldsymbol {y} _ {i} ^ {t} = = k) \\big) \\log (1 - r _ {i k}) \\Bigg),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.783, + 0.785, + 0.813 + ], + "angle": 0, + "content": "where \\(\\mathcal{S}\\) denotes the set of true labels in the current task, along with the memory \\(\\mathcal{M}\\). Finally, the total loss to train the entire pipeline is as follows," + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.822, + 0.785, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t} = \\lambda_ {1} \\mathcal {L} _ {c} + \\lambda_ {2} \\mathcal {L} _ {g} + \\lambda_ {3} \\mathcal {L} _ {r} \\tag {4}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.129 + ], + "angle": 0, + "content": "A. Cheraghian et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "where \\(\\lambda_1, \\lambda_2\\), and \\(\\lambda_3\\) are the weighting factors for the constraint loss \\((\\mathcal{L}c)\\), gradient loss \\((\\mathcal{L}g)\\), and the relation loss \\((\\mathcal{L}_r)\\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.208 + ], + "angle": 0, + "content": "Inference. During the inference process, when given an unlabeled sample \\(\\mathcal{X}^c\\), where \\(c \\in \\mathcal{Y}_{tl}\\), the label prediction \\(y^{*}\\) is computed using the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.222, + 0.786, + 0.248 + ], + "angle": 0, + "content": "\\[\ny ^ {*} = \\underset {j \\in \\mathcal {Y} _ {t l}} {\\arg \\max } (R \\circ (\\mathbf {z} _ {c} \\oplus \\mathbf {v} _ {c} \\oplus \\mathbf {s} _ {j})), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.258, + 0.784, + 0.289 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}_c\\) and \\(\\mathbf{v}_c\\) represent the calculated features for the sample \\(\\mathcal{X}^c\\). We use a standardized prompt for all classes, specifically This is a depth map of a [class]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.311, + 0.376, + 0.329 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.785, + 0.492 + ], + "angle": 0, + "content": "Datasets. We employed two synthetic 3D datasets, Modelnet40 [45] and Shape-net [6], in addition to two real scanned 3D datasets, ScanobjectNN [40] and CO3D [31]. We follow the setup proposed by [11] for Few-Shot Classification with Incremental Learning (FSCIL) on point cloud objects. Our proposed method is assessed in two distinct scenarios. Firstly, we evaluate its performance on the FSCIL task within the same dataset, where both the base and novel tasks originate from identical datasets. Secondly, we subject our method to more demanding and realistic challenges by testing it in cross-dataset scenarios. In this setting, the base task involves synthetic datasets, while the novel tasks comprise classes from real scanned few-shot datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.785, + 0.554 + ], + "angle": 0, + "content": "Prompt generation. GPT-4 was employed to produce diverse class descriptions for each class. The following command was utilized: Describe a depth of a [class]. Approximately 100 class descriptions were generated for each class and utilized during the training phase." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.786, + 0.81 + ], + "angle": 0, + "content": "Implementation details. In our experimentation, we utilize Curvenet as the backbone for 3D point clouds \\((B)\\) [46]. The CLIP vision \\((V_{e})\\) and text encoder \\((T_{e})\\) are instantiated using the ViT-B-16 architecture [14]. The \\(C\\) model is constructed based on the U-Net architecture [32]. The projection module \\((F)\\) follows the methodology proposed by Zhu et al. [51], and for the transformation module \\((T_{r})\\), we employ the STN model [16]. Additionally, the relation module comprises several fully connected layers with sizes (1024, 2048, 1). We use the farthest 1024 points from 3D point cloud objects as input for all samples. For the base task, we train our proposed method (Figure 2) in an end-to-end manner for 250 epochs using the Adam optimizer with a learning rate of 0.001. For the novel task, we fine-tune our model for 200 epochs using the Adam optimizer with a learning rate of 0.001. It is important to note that we randomly shift and scale points in the input point cloud during both base and novel class training, with the additional step of randomly dropping out points. The feature vector dimensions for \\(B\\), \\(T_{e}\\), and \\(V_{e}\\) are set to 512. Additionally, the weights of the loss functions \\(\\lambda_{1}\\), \\(\\lambda_{2}\\), and \\(\\lambda_{3}\\) are configured as 0.01, 1, and 1 across the entire datasets. Our experiments are conducted using the PyTorch framework." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Evaluation metrics. Following the training of each task, we compute the accuracy by consolidating both base and novel classes. Subsequently, following the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Canonical Shape Projection for 3D FSCIL" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.226 + ], + "angle": 0, + "content": "recommendation in [36], we determine the relative accuracy dropping rate, denoted as \\(\\varDelta\\), using the formula \\(\\varDelta=\\frac{|acc_T-acc_0|}{acc_0}\\times 100\\), where \\(acc_{T}\\) and \\(acc_{0}\\) denote the accuracy of the last and first tasks, respectively. The value of \\(\\varDelta\\) serves as a comprehensive metric for method evaluation. A lower relative accuracy signifies superior performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.259, + 0.367, + 0.273 + ], + "angle": 0, + "content": "4.1 Main results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.295, + 0.784, + 0.37 + ], + "angle": 0, + "content": "We conducted a comparative analysis of our approach against established methods in the field, including IL2M [2], ScaIL [3], EEIL [5], LwF [22], FACT [50], Sem-aware [9], and Microshape [12]. Notably, all these methods are tailored for image data, with the exception of Microshape [12], which was originally designed for point cloud data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.788, + 0.508 + ], + "angle": 0, + "content": "Within dataset: In this scenario, both the base and novel tasks are derived from the same dataset. Roughly half of the classes in the dataset are assigned to the base task, while the remaining half are incrementally introduced as multiple novel tasks. It is presumed that these novel tasks encompass only a limited number of samples for training. Our proposed method has been evaluated on three datasets—ModelNet, CO3D, and Shapenet. As depicted in Table 1, our approach demonstrates superior performance across all methods in terms of the relative accuracy dropping rate, denoted as \\(\\Delta\\), indicating the effectiveness of our proposed method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.788, + 0.661 + ], + "angle": 0, + "content": "Cross dataset: In this arrangement, the base task centers around synthetic point cloud objects, while the novel few-shot tasks incorporate real-scanned 3D objects. Notably, the latter often exhibit incompleteness or noise in contrast to the synthetic data, which typically boast cleanliness and completeness. Consequently, this configuration introduces a more realistic and challenging scenario compared to within-dataset experiments. Three datasets are considered in this setting: Shapenet \\(\\rightarrow\\) CO3D, ModelNet40 \\(\\rightarrow\\) ScanObjectNN, and Shapenet \\(\\rightarrow\\) ScanObjectNN. As illustrated in Table 2, our method outperforms other approaches in all three setups, underscoring its efficacy in scenarios involving a domain gap between base and novel tasks." + }, + { + "type": "table_caption", + "bbox": [ + 0.271, + 0.693, + 0.731, + 0.707 + ], + "angle": 0, + "content": "Table 1: Summary of FSCIL results for within-dataset experiments." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.709, + 0.797, + 0.821 + ], + "angle": 0, + "content": "
ModelNetCO3DShapeNet
Method2025303540Δ↓253035404550Δ↓25303540455055 Δ↓
FT89.89.74.33.33.096.776.711.23.63.21.80.899.087.025.76.81.30.90.60.4 99.5
Joint89.888.287.083.580.510.476.769.464.862.760.759.822.087.085.284.383.082.582.281.3 6.6
LwF [22]89.836.09.13.63.196.076.714.74.73.52.31.098.787.060.833.515.93.83.11.8 97.9
IL2M [2]89.865.558.452.353.640.376.731.527.718.127.121.971.487.058.645.740.750.149.449.3 43.3
ScaIL [3]89.866.864.558.756.537.176.739.534.124.130.127.564.187.056.651.844.350.346.445.4 47.8
EEIL [5]89.875.467.260.155.638.176.761.452.442.839.532.857.287.077.773.269.366.465.965.8 22.4
FACT [50]90.481.377.173.565.028.177.967.159.754.850.246.740.087.575.371.469.967.565.762.5 28.6
Sem-aware [9]91.382.274.370.064.729.176.866.959.253.649.142.944.187.274.968.969.168.066.963.8 26.8
Microshape [11]93.683.178.275.867.128.378.567.360.156.151.447.239.987.683.281.579.076.873.5 72.617.1
C3PR (ours)91.682.375.872.270.922.581.569.466.563.054.253.834.088.081.677.876.776.776.976.2715.1
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.129 + ], + "angle": 0, + "content": "A. Cheraghian et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.786, + 0.159 + ], + "angle": 0, + "content": "Table 2: Summary of FSCIL results for within-dataset and cross-dataset experiments." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.161, + 0.805, + 0.274 + ], + "angle": 0, + "content": "
ShapeNet → CO3DModelNet → ScanObjectNNShapeNet → ScanObjectNN
Method3944495459646974798489Δ↓26303437Δ↓44495459Δ↓
FT81.020.22.31.70.81.01.01.30.90.51.698.088.46.46.01.997.981.438.74.00.998.9
Joint81.079.578.375.275.174.872.371.370.068.867.316.988.479.774.071.219.581.482.579.878.73.3
LwF [22]81.057.419.32.31.00.90.81.31.10.81.997.788.435.85.82.597.281.447.914.05.992.8
IL2M [2]81.045.636.835.131.833.334.031.530.632.330.063.088.458.252.952.041.281.453.243.945.843.7
ScaIL [3]81.050.145.739.139.037.938.036.033.733.035.256.588.456.555.952.940.281.449.046.740.050.9
EEIL [5]81.075.269.363.260.557.953.051.951.347.847.641.288.470.261.056.835.781.474.569.863.422.1
FACT [50]81.476.070.368.165.863.563.060.158.257.555.931.389.172.568.363.528.782.374.669.966.818.8
Sem-aware [9]80.669.566.562.963.263.061.258.358.157.255.231.688.573.967.764.227.581.370.665.262.922.6
Microshape [11]82.677.973.972.767.766.265.463.460.658.157.130.989.373.268.465.127.182.574.871.267.118.7
C3PR (ours)83.680.077.875.472.872.370.367.964.964.163.224.488.375.770.667.823.284.577.875.571.914.9
" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.293, + 0.402, + 0.307 + ], + "angle": 0, + "content": "4.2 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.787, + 0.417 + ], + "angle": 0, + "content": "Evaluating model reprogramming: In the conducted experiment (refer to Figure 4 (a)), we systematically omitted the model reprogramming module \\( C \\) and directly fed the depth map image \\( \\mathbf{I}_D \\) into the vision encoder \\( V_{e} \\) of the CLIP model. The results clearly illustrate the significance of the model reprogramming in our proposed approach, as evidenced by notable differences between the outcomes with and without this module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.787, + 0.496 + ], + "angle": 0, + "content": "Impact of the prompt engineering: In our ablation study (see Figure 4 (b)), we investigate the influence of prompt engineering in our methodology. Specifically, we streamline the training stage by substituting the original multiple prompts per class with a single fixed prompt. The outcomes underscore the effectiveness of our GPT-based prompt engineering approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.789, + 0.651 + ], + "angle": 0, + "content": "Impact of canonical shape projection: Here, we evaluate our proposed canonical shape projection methodology in comparison to PointCLIPv2 [51], which generates multiple fixed-depth map projections for each input point cloud. Specifically, we eliminate the transformation module \\( T_r \\). Subsequently, we adopt the depth map generation method employed in PointCLIP v2, involving the use of ten fixed camera pose angles for computing ten depth map images. These images are then fed into the vision encoder \\( V_e \\) of the CLIP model, and the relation module is applied to merge the feature vector, mirroring our proposed pipeline. Notably, our method surpasses PointCLIPv2 by a significant margin (see Figure 4 (c)), highlighting the efficacy of our learned camera pose angle." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.68, + 0.409, + 0.769 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.68, + 0.597, + 0.769 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.681, + 0.784, + 0.768 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.837 + ], + "angle": 0, + "content": "Fig. 4: This experiment is performed within the Shapenet to ScanobjectNN setting. In (a), we present the influence of our novel CLIP model reprogramming strategy. (b) Highlights the impact of prompt engineering as proposed in our method. (c) Our approach is compared with PointCLIP v2 [51]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Canonical Shape Projection for 3D FSCIL" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.145, + 0.495, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.145, + 0.7, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.271, + 0.785, + 0.314 + ], + "angle": 0, + "content": "Fig. 5: We visualize the quality of features (from 44 base classes of ShapeNet) generated by the CLIP vision encoder using (Right) our proposed method, C3PR, and (Left) PointCLIPv2 [51]. Our method produced a denser cluster than PointCLIP." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.784, + 0.387 + ], + "angle": 0, + "content": "Feature quality: In Figure 5, we assess the feature quality produced by the CLIP vision encoder using our proposed method, C3PR, and compare it with PointCLIPv2 [51]. As illustrated, our approach yields more discriminative clusters, enhancing overall generalization capabilities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.785, + 0.463 + ], + "angle": 0, + "content": "Influence of \\(\\mathcal{L}_c\\) and \\(\\mathcal{L}_g\\): In Table 3a, we highlight the impact of both \\(\\mathcal{L}_c\\) and \\(\\mathcal{L}_g\\) within our proposed pipeline. It is evident that both loss functions contribute significantly, but \\(\\mathcal{L}_c\\) holds greater importance, underscoring the significance of identifying the learned projection angle for depth map images on image foundation models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.785, + 0.526 + ], + "angle": 0, + "content": "Model reprogramming: Conventional vs. our: In this section, we contrast our innovative model reprogramming approach with the conventional method introduced by [15]. Table 3b illustrates the superior effectiveness of our proposed model reprogramming strategy." + }, + { + "type": "table_caption", + "bbox": [ + 0.221, + 0.535, + 0.778, + 0.551 + ], + "angle": 0, + "content": "Table 3: Ablation study on the influence of loss (a) and model reprogramming (b)." + }, + { + "type": "table_caption", + "bbox": [ + 0.226, + 0.555, + 0.482, + 0.569 + ], + "angle": 0, + "content": "(a) Influence of \\(\\mathcal{L}_c\\) and \\(\\mathcal{L}_g\\) in our approach." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.579, + 0.492, + 0.668 + ], + "angle": 0, + "content": "
ShapeNet → ScanObjectNN
Method44495459Δ↓
Lr84.476.270.16522.9
Lr + Lg84.377.370.866.920.6
Lr + Lc85.780.274.570.817.3
Lr + Lg + Lc84.577.875.571.914.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.556, + 0.785, + 0.589 + ], + "angle": 0, + "content": "(b) Comparing our proposed model reprogramming strategy, C3PR, with the conventional model reprogramming technique employed in [15]." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.6, + 0.778, + 0.66 + ], + "angle": 0, + "content": "
ShapeNet → ScanObjectNN
Method44495459Δ↓
[15]85.780.274.570.817.3
C3PR (ours)84.577.875.571.914.9
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.689, + 0.348, + 0.702 + ], + "angle": 0, + "content": "4.3 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.718, + 0.784, + 0.793 + ], + "angle": 0, + "content": "Point-cloud classification: Although our learning projection module \\(Tr\\) and model reprogramming module \\(C\\) have demonstrated effectiveness in the FSCIL context, they can also be applied in the point-cloud classification setting (shown in the table 4a. However, it is worth to note that our proposed prompt engineering strategy is specifically tailored to improve performance in FSCIL problems." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.794, + 0.344, + 0.807 + ], + "angle": 0, + "content": "Module \\(C\\) role:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "The role of Module \\( C \\) is twofold: it learns background information for the depth maps and smoothens the foreground. Specifically, we combine the depth" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.129 + ], + "angle": 0, + "content": "A. Cheraghian et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "map \\(\\mathbf{I}_{D_i}\\) from the projection module \\(F\\) with the extracted values \\(\\mathbf{I}_{T_i}\\) from Module \\(C\\) to generate the final image, \\(\\mathbf{I}_{R_i} = \\mathbf{I}_{D_i} \\odot C(M(\\mathbf{I}_{D_i}))\\), where \\(\\odot\\) represents element-wise multiplication. To this end, we conducted experiments to demonstrate the impact of Module \\(C\\) in our proposed method (see table 4b)." + }, + { + "type": "table_caption", + "bbox": [ + 0.232, + 0.215, + 0.772, + 0.23 + ], + "angle": 0, + "content": "Table 4: Ablation on point-cloud classification (a) and impact of \\( C \\) module (b)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.236, + 0.482, + 0.258 + ], + "angle": 0, + "content": "(a) The performance of our proposed method on point-cloud classification in ModelNet40." + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.236, + 0.79, + 0.258 + ], + "angle": 0, + "content": "(b) The impact of module \\(C\\) in our proposed method." + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.269, + 0.442, + 0.331 + ], + "angle": 0, + "content": "
MethodAccuracy
DGCNN [43]92.9
CurveNet [46]93.8
Ours94.6
" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.269, + 0.763, + 0.331 + ], + "angle": 0, + "content": "
ShapeNet → ScanObjectNN
Method44495459Δ↓
Without C84.576.373.268.918.4
With C84.577.875.571.914.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.788, + 0.46 + ], + "angle": 0, + "content": "**Limitation:** The proposed method exhibits several limitations that warrant discussion. Firstly, its reliance on storing exemplars in memory, with one sample per class, and addressing the forgetting issue raises the question of exploring exemplar-free alternatives, which could offer intriguing avenues for further research. Secondly, our method would perform worse on new concepts from significantly different domains, synthetic to real-scan datasets, as can be seen in Table 2. However, it works well when the new concepts come from similar domains, synthetic to synthetic datasets, as can be seen in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.48, + 0.36, + 0.496 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.789, + 0.81 + ], + "angle": 0, + "content": "In this paper, we leverage the CLIP architecture of image-to-text matching to address the FSCIL problem on 3D point cloud objects. Traditional approaches usually project a 3D object on a 2D plane to produce images from different camera angles. Among those 2D images, some have bad/poor camera angles, making them difficult to recognize. Therefore, we propose a learning mechanism to find one camera projection that best describes the 3D objects. In doing so, we adopt a model reprogramming approach that perturbs the input 2D images (from the best projection angle) with background-foreground colorization (segmentation) without modifying or fine-tuning the original CLIP architectures. Our reprogramming approach is different from conventional approaches because our custom loss produces a uniform background color instead of the noisy background of previous works. In the text pipeline, we use the CLIP text encoder and train the whole architecture end-to-end. Instead of using a single class name-based text input, we suggest using prompts based on GPT-based description. We propose a prompt engineering approach to train the networks with 3D object and text description pairs. Finally, we evaluate our proposed method using 3D synthetic datasets, ModelNet and ShapeNet, and two 3D real-world-scanned datasets, ScanObjectNN and CO3D, and report state-of-the-art performances. In future, we will investigate model reprogramming for other vision problems beyond FSCIL, like segmentation and domain adaptation on 3D point clouds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Acknowledgment. Mehrtash Harandi thanks the Australian Research Council for support through the Discovery Program (DP230101176)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Canonical Shape Projection for 3D FSCIL" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.787, + 0.256 + ], + "angle": 0, + "content": "1. Bansal, N., Chen, X., Wang, Z.: Can we gain more from orthogonality regularizations in training deep networks? In: Bengio, S., Wallach, H., Larochelle, H., Grauman, K., Cesa-Bianchi, N., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 31. Curran Associates, Inc. (2018), https://proceedings.neurips.cc/paper_files/paper/2018/file/bf424cb7b0dea050a42b9739eb261a3a-Paper.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.257, + 0.786, + 0.283 + ], + "angle": 0, + "content": "2. Belouadah, E., Popescu, A.: II2m: Class incremental learning with dual memory. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.283, + 0.786, + 0.31 + ], + "angle": 0, + "content": "3. Belouadah, E., Popescu, A.: Scail: Classifier weights scaling for class incremental learning. In: WACV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.787, + 0.434 + ], + "angle": 0, + "content": "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 1877-1901. Curran Associates, Inc. (2020), https://proceedings.neurips.cc/paper_files/paper/2020/file/1457c0d6bcbd4967418bf8ac142f64a-Paper.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.786, + 0.46 + ], + "angle": 0, + "content": "5. Castro, F.M., Marín-Jiménez, M.J., Guil, N., Schmid, C., Alahari, K.: End-to-end incremental learning. In: ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.46, + 0.786, + 0.501 + ], + "angle": 0, + "content": "6. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.501, + 0.786, + 0.528 + ], + "angle": 0, + "content": "7. Chen, K., Lee, C.G.: Incremental few-shot learning via vector quantization in deep embedded space. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.528, + 0.786, + 0.554 + ], + "angle": 0, + "content": "8. Chen, P.Y.: Model reprogramming: Resource-efficient cross-domain machine learning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.554, + 0.786, + 0.595 + ], + "angle": 0, + "content": "9. Cheraghian, A., Rahman, S., Fang, P., Roy, S.K., Petersson, L., Harandi, M.: Semantic-aware knowledge distillation for few-shot class-incremental learning. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.595, + 0.786, + 0.636 + ], + "angle": 0, + "content": "0. Cheraghian, A., Rahman, S., Ramasinghe, S., Fang, P., Simon, C., Petersson, L., Harandi, M.: Synthesized feature based few-shot class-incremental learning on a mixture of subspaces. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.636, + 0.786, + 0.691 + ], + "angle": 0, + "content": "1. Chowdhury, T., Cheraghian, A., Ramasinghe, S., Ahmadi, S., Saberi, M., Rahman, S.: Few-shot class-incremental learning for 3d point cloud objects. In: Avidan, S., Brostow, G., Cissé, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 204-220. Springer Nature Switzerland, Cham (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.786, + 0.718 + ], + "angle": 0, + "content": "2. Chowdhury, T., Cheraghian, A., Ramasinghe, S., Ahmadi, S., Saberi, M., Rahman, S.: Few-shot class-incremental learning for 3d point cloud objects. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.786, + 0.744 + ], + "angle": 0, + "content": "3. Dinh, T., Seo, D., Du, Z., Shang, L., Lee, K.: Improved input reprogramming for gan conditioning (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.786, + 0.799 + ], + "angle": 0, + "content": "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "5. Elsayed, G.F., Goodfellow, I., Sohl-Dickstein, J.: Adversarial reprogramming of neural networks. In: International Conference on Learning Representations (2019), https://openreview.net/forum?id=Syx_Ss05tm" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.129 + ], + "angle": 0, + "content": "A. Cheraghian et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.218 + ], + "angle": 0, + "content": "16. Jaderberg, M., Simonyan, K., Zisserman, A., kavukcuoglu, k.: Spatial transformer networks. In: Cortes, C., Lawrence, N., Lee, D., Sugiyama, M., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 28. Curran Associates, Inc. (2015), https://proceedings.neurips.cc/paper_files/paper/2015/file/33ceb07bf4eeb3da587e268d663aba1a-Paper.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.787, + 0.302 + ], + "angle": 0, + "content": "17. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 4904-4916. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/jia21b.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.302, + 0.787, + 0.33 + ], + "angle": 0, + "content": "18. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.33, + 0.787, + 0.385 + ], + "angle": 0, + "content": "19. Lee, D.H., Pujara, J., Sewak, M., White, R.W., Jauhar, S.K.: Making large language models better data creators. In: The 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP) (2023), https://openreview.net/forum?id=2Rdfdri2oT" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.386, + 0.787, + 0.43 + ], + "angle": 0, + "content": "20. Lee, K.Y., Zhong, Y., Wang, Y.X.: Do pre-trained models benefit equally in continual learning? In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 6485-6493 (January 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.43, + 0.787, + 0.458 + ], + "angle": 0, + "content": "21. Li, Y., Bu, R., Sun, M., Wu, W., Di, X., Chen, B.: Pointcnn: Convolution on x-transformed points. In: NeurIPS (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.458, + 0.787, + 0.487 + ], + "angle": 0, + "content": "22. Li, Z., Hoiem, D.: Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.487, + 0.787, + 0.515 + ], + "angle": 0, + "content": "23. Liu, Y., Fan, B., Xiang, S., Pan, C.: Relation-shape convolutional neural network for point cloud analysis. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.787, + 0.53 + ], + "angle": 0, + "content": "24. Mazumder, P., Singh, P., Rai, P.: Few-shot lifelong learning. In: AAAI (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.53, + 0.787, + 0.599 + ], + "angle": 0, + "content": "25. Pei, Y., Qing, Z., CEN, J., Wang, X., Zhang, S., Wang, Y., Tang, M., Sang, N., Qian, X.: Learning a condensed frame for memory-efficient video class-incremental learning. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=1CGYC7pXWNQ" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.6, + 0.787, + 0.628 + ], + "angle": 0, + "content": "26. Poulenard, A., Rakotosaona, M.J., Ponty, Y., Ovsjanikov, M.: Effective rotation-invariant point cnn with spherical harmonics kernels. In: 3DV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.628, + 0.787, + 0.657 + ], + "angle": 0, + "content": "27. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.787, + 0.685 + ], + "angle": 0, + "content": "28. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: NeurIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.686, + 0.787, + 0.769 + ], + "angle": 0, + "content": "29. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 8748-8763. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/radford21a.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.77, + 0.787, + 0.798 + ], + "angle": 0, + "content": "30. Rao, Y., Lu, J., Zhou, J.: Spherical fractal convolutional neural networks for point cloud recognition. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "31. Reizenstein, J., Shapovalov, R., Henzler, P., Sbordone, L., Labatut, P., Novotny, D.: Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In: ICCV (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Canonical Shape Projection for 3D FSCIL" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.204 + ], + "angle": 0, + "content": "32. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F. (eds.) Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015. pp. 234-241. Springer International Publishing, Cham (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.205, + 0.787, + 0.231 + ], + "angle": 0, + "content": "33. Singh, A., Hu, R., Goswami, V., Couairon, G., Galuba, W., Rohrbach, M., Kiela, D.: FLAVA: A foundational language and vision alignment model. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.232, + 0.787, + 0.272 + ], + "angle": 0, + "content": "34. Sung, F., Yang, Y., Zhang, L., Xiang, T., Torr, P.H., Hospedales, T.M.: Learning to compare: Relation network for few-shot learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "35. Tan, Y., Xiang, X.: Cross-domain few-shot incremental learning for point-cloud recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 2307-2316 (January 2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "36. Tan, Z., Ding, K., Guo, R., Liu, H.: Graph few-shot class-incremental learning. In: Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.357, + 0.787, + 0.383 + ], + "angle": 0, + "content": "37. Tao, X., Hong, X., Chang, X., Dong, S., Wei, X., Gong, Y.: Few-shot class-incremental learning. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.384, + 0.787, + 0.464 + ], + "angle": 0, + "content": "38. Tsai, Y.Y., Chen, P.Y., Ho, T.Y.: Transfer learning without knowing: Reprogramming black-box machine learning models with scarce data and limited resources. In: III, H.D., Singh, A. (eds.) Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 9614-9624. PMLR (13-18 Jul 2020), https://proceedings.mlr.press/v119/tsai20a.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.467, + 0.787, + 0.522 + ], + "angle": 0, + "content": "39. Tsai, Y.Y., Chen, P.Y., Ho, T.Y.: Transfer learning without knowing: Reprogramming black-box machine learning models with scarce data and limited resources. In: Proceedings of the 37th International Conference on Machine Learning. ICML'20, JMLR.org (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.523, + 0.787, + 0.563 + ], + "angle": 0, + "content": "40. Uy, M.A., Pham, Q.H., Hua, B.S., Nguyen, D.T., Yeung, S.K.: Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.565, + 0.787, + 0.59 + ], + "angle": 0, + "content": "41. Wang, C., Samari, B., Siddiqi, K.: Local spectral graph convolution for point set feature learning. In: ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.592, + 0.787, + 0.646 + ], + "angle": 0, + "content": "42. Wang, R., Duan, X., Kang, G., Liu, J., Lin, S., Xu, S., Lu, J., Zhang, B.: Attriclip: A non-incremental learner for incremental knowledge learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3654-3663 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.647, + 0.787, + 0.674 + ], + "angle": 0, + "content": "43. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.675, + 0.787, + 0.701 + ], + "angle": 0, + "content": "44. Wu, W., Qi, Z., Fuxin, L.: PointCONV: Deep convolutional networks on 3D point clouds. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.702, + 0.787, + 0.729 + ], + "angle": 0, + "content": "45. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3D ShapeNets: A deep representation for volumetric shapes. In: CVPR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.73, + 0.787, + 0.77 + ], + "angle": 0, + "content": "46. Xiang, T., Zhang, C., Song, Y., Yu, J., Cai, W.: Walk in the cloud: Learning curves for point clouds shape analysis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 915-924 (October 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.772, + 0.787, + 0.798 + ], + "angle": 0, + "content": "47. Xu, Y., Fan, T., Xu, M., Zeng, L., Qiao, Y.: SpiderCNN: Deep learning on point sets with parameterized convolutional filters. In: ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "48. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. arXiv preprint arXiv:2112.02413 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.129 + ], + "angle": 0, + "content": "A. Cheraghian et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "49. Zhang, Y., Rabbat, M.: A graph-cnn for 3d point cloud classification. In: ICASSP (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.203 + ], + "angle": 0, + "content": "50. Zhou, D.W., Wang, F.Y., Ye, H.J., Ma, L., Pu, S., Zhan, D.C.: Forward compatible few-shot class-incremental learning. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.245 + ], + "angle": 0, + "content": "51. Zhu, X., Zhang, R., He, B., Guo, Z., Zeng, Z., Qin, Z., Zhang, S., Gao, P.: Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. arXiv preprint arXiv:2211.11682 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.245 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Canonical Shape Projection is All You Need for 3D Few-shot Class Incremental Learning/354b1d2e-f84c-4d0d-a2a6-9954d66577ef_origin.pdf b/2024/Canonical Shape Projection is All You Need for 3D Few-shot Class Incremental Learning/354b1d2e-f84c-4d0d-a2a6-9954d66577ef_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7e7646b0f5f39c29924e532a63afdb2732641834 --- /dev/null +++ b/2024/Canonical Shape Projection is All You Need for 3D Few-shot Class Incremental Learning/354b1d2e-f84c-4d0d-a2a6-9954d66577ef_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c7a3ab1d2cc7e1c37b70b6619d5e4d9b695029e0348e26de9931880660ad07e +size 2154466 diff --git a/2024/Canonical Shape Projection is All You Need for 3D Few-shot Class Incremental Learning/full.md b/2024/Canonical Shape Projection is All You Need for 3D Few-shot Class Incremental Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d2b400e046e2bd5a4969268469e725dc8edf0892 --- /dev/null +++ b/2024/Canonical Shape Projection is All You Need for 3D Few-shot Class Incremental Learning/full.md @@ -0,0 +1,269 @@ +# Canonical Shape Projection is All You Need for 3D Few-shot Class Incremental Learning + +Ali Cheraghian\*,1,2, Zeeshan Hayden\*,1,2, Sameera Ramasinghe\*3, Shafin Rahman4, Javad Jafaryahya5, Lars Petersson2, and Mehrtash Harandi6 + +1 Data61, CSIRO, Australia, 2 Australian National University, 3 Amazon, 4 North South University, 5 University of Technology Sydney, 6 Monash University {ali.cheraghian, zeeshan.hayder, lars.petersson}@data61.csiro.au, ramasisa@amazon.com, shafin.rahman@northsouth.edu, javad.jafaryahya@student.uts.edu.au, mehrtash.harandi@monash.edu + +Abstract. In recent years, robust pre-trained foundation models have been successfully used in many downstream tasks. Here, we would like to use such powerful models to address the problem of few-shot class incremental learning (FSCIL) tasks on 3D point cloud objects. Our approach is to reprogram the well-known CLIP-based foundation model (trained on 2D images and text pairs) for this purpose. The CLIP model works by ingesting 2D images, so to leverage it in our context, we project the 3D object point cloud onto 2D image space to create proper depth maps. For this, prior works consider a fixed and non-trainable set of camera poses. In contrast, we propose to train the network to find a projection that best describes the object and is appropriate for extracting 2D image features from the CLIP vision encoder. Directly using the generated depth map is not suitable for the CLIP model, so we apply the model reprogramming paradigm to the depth map to augment the foreground and background to adapt it. This removes the need for modification or fine-tuning of the foundation model. In the setting we have investigated, we have limited access to data from novel classes, resulting in a problem with overfitting. Here, we address this problem via the use of a prompt engineering approach using multiple GPT-generated text descriptions. Our method, C3PR, successfully outperforms existing FSCIL methods on ModelNet, ShapeNet, ScanObjectNN, and CO3D datasets. The code is available at https://github.com/alichr/C3PR. + +Keywords: 3D shape projection $\cdot$ Model reprogramming $\cdot$ Few-shot class incremental learning + +# 1 Introduction + +Maintaining a balance between acquiring new concepts and retaining existing knowledge presents a significant challenge for machine learning algorithms, a + +![](images/362c13261a5cf4cefef74e2dd6ffe2218a4d9886d7caea7e9e702f19356f56dd.jpg) +Fig. 1: (a) Traditional methods using image-based models for point cloud processing often generate multiple depth maps from fixed camera poses, potentially causing feature alignment issues and difficulties in recognizing 3D objects from unsuitable camera angles. (b) In contrast, our proposed approach (C3PR) determines the most suitable camera pose for projecting a single depth map by leveraging feedback from the foundation model. (c) To improve the foundation model's performance in handling 3D data for the FSCIL task, we introduce a novel model reprogramming technique (without modifying or fine-tuning the foundation model) to further refine the depth map. + +task that comes naturally to humans. This task becomes more difficult when only a few samples are available for new concepts. In the literature, this problem is referred to as few-shot class incremental learning (FSCIL) [7,9,10,24,37]. Substantial effort on this topic has been invested, particularly on image data [7,9,10, 24,37], whereas relatively few studies are tailored specifically towards point cloud data. For example, in the 2D domain, there are vision-language models such as CLIP [29], ALIGN [17], and FLAVA [33] models with an inherent capacity to generate highly versatile representations and demonstrate zero-shot capabilities, all without parameter fine-tuning [20, 25, 42]. Gathering training data for 3D point clouds is substantially more challenging than for image data, and here we explore whether the CLIP model can help alleviate this issue. In this paper, we attempt to leverage CLIP for FSCIL on 3D Point Cloud Objects. + +Popular approaches capture 2D images from a fixed set of camera poses and then pass the images to 2D foundation models for feature extraction [48, 51]. The main drawback of such a strategy is that camera poses are not adaptable + +to the target dataset, hindering capturing the objects in poses ideal for feature extraction (see Fig. 1 (a)). To address this problem, we propose to learn the camera projection pose based on the dataset that best describes the 3D objects (see Fig. 1 (b)). Our proposed model reprogramming approach is advantageous for representing real-world scanned 3D objects where point cloud data contains sensor noise (see Fig. 1 (c)). The unique camera pose can still capture the global shape of noisy 3D objects, even though the object belongs to novel classes. For language/text modality, many existing works [9-11] leverages vision-language learning to address the overfitting issue associated with few-shot novel classes. We argue that vital knowledge lies within the language domain, which may aid in the training process of the vision domain, particularly for few-shot classes, to alleviate overfitting. However, thus far, existing methods have only utilized class names from the language domain [9, 10, 12]. In contrast, we propose a method based on prompt engineering to better leverage the language domain. + +This paper presents a novel approach for extracting valuable insights from foundational models like CLIP, focusing on FSCIL on 3D point cloud data. As CLIP's training data comprises millions of image-text pairs, our target is to transfer this learned knowledge as a complete black box (without knowing or modifying pre-trained parameters) for 3D data. To this end, we propose reprogramming [38] the CLIP model, specifically for point cloud objects. The proposed approach is different from prior attempts [48,51] where they used a fixed set of camera poses for 3D to 2D image projection and did not augment the input data for better usage of the CLIP encoder. Our C3PR method emphasizes generating depth map images with consistent, informative, background pixels, diverging from conventional model reprogramming techniques that frequently produce random pixel information for the background. We observed that maintaining consistent background pixels contributes to improved generalization, particularly for few-shot novel classes. These steps indirectly add essential perturbation to the input depth map images so that the black box CLIP vision encoder can operate without fine-tuning. Furthermore, we propose a novel approach that leverages information from the language domain using a Large Language Model, GPT, to combat overfitting issues in few-shot novel classes. Specifically, we generate multiple descriptive prompts (instead of single class name-based descriptions as used in [11]) for base and novel classes. It also compensates for the issue with limited data during novel class learning. The generated prompts are then employed in the CLIP text encoder, reducing overfitting during incremental training sessions. Both the vision and text pipelines of our method can be trained end-to-end. We report state-of-the-art performances when evaluated on 3D synthetic datasets, ModelNet [45] and ShapeNet [6], and two 3D real world-scanned datasets, ScanObjectNN [40] and Common Objects in 3D (CO3D) [31]. + +The main contributions of our proposed method are: + +- Learning the preferred object-based camera pose for 3D to 2D projections; +- A novel model reprogramming approach for 2D CLIP encoders targeting 3D point cloud data; + +- Prompt engineering based on a GPT model for novel few-shot classes to mitigate overfitting. + +# 2 Related work + +3D point cloud processing: In recent years, there has been a surge in research focused on leveraging deep learning for the direct processing of 3D point cloud objects. Notable contributions include PointNet [27] by Qi et al., which introduced the use of multi-layer perceptron (MLP) networks for 3D point cloud processing. However, this approach overlooked the inherent local structures within the input data. Subsequent techniques were developed to address this limitation, including PointNet++ [28], which hierarchically extracts features to take advantage of local information. Building upon this, several works [21,23,26,30,44,47] have proposed convolution strategies for extracting local information. PointConv [44] introduced a novel convolution operation that locates a Monte Carlo estimation of the hidden continuous 3D convolution based on a powerful sampling technique. SFCNN [30] proposed graph convolution on spherical points. Other papers [41,43,49] treat each point cloud as a graph vertex to learn features in either spatial or spectral domains. DGCNN [43] constructs a graph in feature space and dynamically updates it using MLP for each edge, while PointGCN [49] suggests an approach that generates the graph using k-nearest neighbors from a point cloud to capture local information. + +Few-shot class-incremental learning: Tao et al. [37] were the pioneers in introducing the FSCIL setting for image data. They proposed the use of a neural gas network to address the issue of forgetting by preserving the topology of classes based on their feature vectors. Following this, in [10], a mixture of the subspaces are generated based on the training data distribution. In a related vein, Mazumder et al [24] devised a method that selectively adapts a subset of model parameters for learning novel classes, thereby curbing overfitting. They also combat forgetting by freezing critical parameters in the model. Cheraghian et al. [9] proposed a novel vision-language strategy, incorporating class semantic information from language space using a distillation technique to alleviate the impact of catastrophic forgetting. They also integrated an attention mechanism to counter overfitting on few-shot novel tasks. Introduced in [12], the Specific FSCIL tailored for 3D point cloud data presents the innovative concept of Microshape, which captures essential details from the primary task, effectively addressing both forgetting and overfitting challenges. Additionally, the work by [35] delves into cross-domain FSCIL applied to point-cloud recognition. This study dissects catastrophic forgetting into two distinct components: base class forgetting and incremental class forgetting, offering separate mitigation strategies. + +Model reprogramming: Foundation model reprogramming has recently gained attention [8], primarily due to its reduced computational and time requirements compared to the fine-tuning approach. Prior to this, Elsayed et al. [15] introduced adversarial reprogramming. This technique involves mapping ImageNet labels to task labels, aiming to perplex the ImageNet classifier and induce be + +havior akin to a CIFAR-10 or MNIST classifier. Dinh et al. [13] proposed that input reprogramming is more effective in situations with limited labeled data than training from scratch and fine-tuning. Tsai et al. [39] introduced a technique called black-box adversarial reprogramming. This method involves transforming the input and integrating label mapping mechanisms into the output. In this paper, we present a novel foundation model reprogramming approach tailored for the CLIP model. + +# 3 Method + +In this section, we first introduce the FSCIL problem settings. Let's consider a sequence of $T$ tasks denoted by $\mathbf{Q} = \{\mathcal{Q}^1,\mathcal{Q}^2,\dots ,\mathcal{Q}^T\}$ , where $\mathcal{V}^t$ represents the set of classes in task $\mathcal{Q}^t$ , and $\mathcal{V}^i\cap \mathcal{V}^j = \emptyset$ . Additionally, each class in all tasks is associated with a set of prompt class descriptions generated by an LLM, e.g., GPT [4], denoted as $\mathcal{P}^t$ . Thus, we can represent each task as a tuple $\mathcal{Q}^t = \{\mathcal{X}_i^t,\mathbf{y}_i^t,\mathbf{p}_i^t\}_{i = 1}^{n_t}$ , where $\mathcal{X}_i^t = \{\mathbf{x}_{i,j}^t\}_{j = 1}^l$ refers to a 3D point cloud object with coordinates $\mathbf{x}_{i,j}^{t}\in \mathbb{R}^{3}$ . Moreover, $\mathbf{y}_i^t\in \mathcal{V}^t$ and $\mathbf{p}_i^t\in \mathcal{P}^t$ correspond to the label of the point cloud and its corresponding class prompt description, respectively. In the proposed FSCIL framework, $\mathcal{Q}^1$ is the base task ( $t = 1$ ), where the model undergoes training on a large-scale synthetic 3D dataset. For $t > 1$ , the training data are drawn from real-world 3D point clouds consisting of only a few instances. The model is trained sequentially across the tasks $t = 1,\ldots ,T$ . However, during the $t$ -th task, the model is exposed to $\mathcal{X}^t$ , $\mathbf{y}^t$ , and $\{\mathcal{P}^1,\mathcal{P}^2,\dots ,\mathcal{P}^t\}$ . During inference, the model trained on the current task $\mathcal{Q}^t$ is expected to classify test samples from both the current and previous tasks, i.e., $\{\mathcal{Q}^1,\mathcal{Q}^2,\dots ,\mathcal{Q}^t\}$ . + +# 3.1 Model overview + +In this section, we present an overview of our proposed methodology, illustrated in Fig. 2. We begin with a point cloud sample denoted as $\mathcal{X}_i^t$ , which is passed through the projection module $F$ . This module utilizes the rotation matrix $\mathbf{R}_i^t$ derived from the transformation module $T_r$ to generate a depth map labelled as $\mathbf{I}_{D_i}^t$ . For the projection module $F$ , we use the method proposed by [51], which is fixed during the training stage. The transformation module $T_r$ is crucial for determining the appropriate angle $\theta$ for projection. Additionally, a soft constraint loss function $\mathcal{L}_c$ is applied to the rotation matrix to ensure that the transformation module generates a valid rotation matrix. Following this, we extract a binary mask $\mathbf{I}_{M_i}^t$ from the depth map. Within this mask $M$ , background pixels are assigned a value of zero, while non-zero depth map values are set to one. Subsequently, this mask is fed into a model reprogramming module $C$ to add pixel information to the background of the depth map. To ensure that the generated embedding does not converge to a noisy image and follows the pattern of the mask input, an image gradient consistency loss function $\mathcal{L}_g$ is used between the input and output of the module $C$ . The resulting image embedding $\mathbf{I}_{H_i}^t$ is then merged with the original depth map in the subsequent stage, resulting in + +![](images/14f2f243edd12565d6f9449f7e653a655f888504c9340c334d5c0d6bc19201b5.jpg) +Fig. 2: Beginning with a point cloud sample $\mathcal{X}_i^t$ , we use the projection module $F$ to generate a depth map $\mathbf{I}_{D_i}^t$ via the rotation matrix $\mathbf{R}_i^t$ from the transformation module $T_r$ . A mask $\mathbf{I}_{M_i}^t$ is then extracted from the depth map. This mask is input into module $C$ and the resulting image is merged with the depth map, yielding the image $\mathbf{I}_{R_i}^t$ . Next, $\mathbf{I}_{R_i}^t$ is input to the CLIP vision encoder model $V_e$ to generate a vision embedding $\mathbf{v}_i^t \in \mathbb{R}^m$ . Simultaneously, $\mathcal{X}_i^t$ is fed into $B$ to create a point cloud representation $\mathbf{z}_i^t \in \mathbb{R}^m$ . Moreover, the GPT model generates multiple class name descriptions, which are then stored in a prompt bank. To this end, prompts $\{\mathbf{p}_1, \mathbf{p}_2, \dots, \mathbf{p}_{|\mathcal{Y}_{tl}|}\}$ for all old classes and the current task $t$ are forwarded into the text encoder $T_e$ , producing corresponding text embeddings $\{\mathbf{e}_1, \mathbf{e}_2, \dots, \mathbf{e}_{|\mathcal{Y}_{tl}|}\}$ , where $\mathbf{e}_i^t \in \mathbb{R}^m$ . Finally, $\mathbf{v}_i^t$ , $\mathbf{z}_i^t$ , and $\mathbf{e}_j^t$ are concatenated and passed through the relation module $R_l$ to generate a scalar value $r_i^t$ , serving as a similarity score. + +the production of a reprogrammed depth map image denoted as $\mathbf{I}_{R_i}^t$ . After that, $\mathbf{I}_{R_i}^t$ is input into the vision encoder of the CLIP model $V_{e}$ to generate a vision embedding represented by $\mathbf{v}_i^t\in \mathbb{R}^m$ . Simultaneously, the point cloud input $\mathcal{X}_i^t$ is forwarded to a 3D point cloud backbone, for instance, Curvenet [46], to generate a feature description $\mathbf{z}_i^t\in \mathbb{R}^m$ . At the same time, GPT is employed to generate multiple descriptions for each class using the command: describe a depth map of a [class], thereby creating several prompts, which are stored in the prompt bank module. Then, the prompts of all classes $\{\mathbf{p}_j\}_{j = 1}^{|\mathcal{Y}_{tl}|}$ , where $\mathcal{Y}_{tl} = \bigcup_{i = 1}^{t}\mathcal{Y}^{i}$ , observed up to task $t$ are fed into the text encoder $T_{e}$ to generate the corresponding text embeddings denoted as $\{\mathbf{e}_j\}_{j = 1}^{|\mathcal{Y}_{tl}|}$ , where each $\mathbf{e}_j^t\in \mathbb{R}^m$ . The prompts are generated using LLMs, such as GPT [4], given a command: Describe a depth map of a [class]. In the subsequent stage, all $\mathbf{v}_i^t$ , $\mathbf{z}_i^t$ , and $\mathbf{e}_i^t$ are concatenated and passed through the relation module [34] $R_{l}$ to generate a scalar value $r_i^t$ ranging from 0 to 1. This value serves as a measure of the similarity between the visual and prompt feature embeddings. Finally, a binary cross-entropy loss function $\mathcal{L}_r$ is used to train the entire proposed pipeline. + +# 3.2 Canonical 3D shape projection + +A 3D-to-2D projection matrix can be designed to project any 3D point cloud on a 2D plane considering any camera angle. Nevertheless, the feature embedding + +![](images/d7a377aabb210e3f2f7785bbb9c27189deaf50fd77d8c0e1391a395fdda77b67.jpg) +Fig. 3: We demonstrate the sensitivity of the CLIP model to the angle of depth map projection. Here, we consider two random projections to generate depth maps. Additionally, we assume that there are three classes: car, guitar, and person. Even though the number of classes is limited to only three, the CLIP model can only correctly classify two of these depth maps. This highlights the importance of finding the optimum projection angle to fully leverage the CLIP model for 3D point cloud processing. + +produced by the visual encoder of large image-based foundation models (e.g., CLIP) can be affected by the content that can be controlled by the projection angle. Figure 3 shows that an adjustment is pivotal due to the CLIP model's sensitivity to the projection angle, especially during depth map processing. We are interested in finding the optimal projection angle to enhance the performance of the CLIP model. When projecting from a 3D point cloud to 2D images, we leverage the mapping function introduced by [51]. We provide optimum projection angle $\theta$ via a rotation matrix $\mathbf{R} \in \mathbb{R}^{3 \times 3}$ , where $\mathbf{R} = T_r(\mathcal{X})$ , generated by the transformation module $T_r$ to ascertain the most effective projection angle for converting a 3D point cloud into a depth map. The generated $\mathbf{R}$ is then applied to the given 3D point to derive the appropriate angle for the projection module $\theta$ . To ensure that $\mathbf{R}$ is a valid rotation matrix, we apply the following loss function at the output of the transformation module $T_r$ , + +$$ +\mathcal {L} _ {c} = \frac {1}{n} \sum_ {i = 1} ^ {n} \left(\left\| \mathbf {R} _ {i} ^ {T} \mathbf {R} _ {i} - \mathbf {I} \right\| _ {\mathcal {F}} ^ {2} + \left| \det (\mathbf {R} _ {i}) - 1 \right|\right) \tag {1} +$$ + +where $\mathbf{R}_i = T_r(\mathcal{X}_i)$ , $n$ is the number of samples in the training set, $\| \cdot \|_{\mathcal{F}}^2$ denotes the Frobenius norm, $\mathbf{I}$ is the identity matrix. The orthogonality condition $\left\| \mathbf{R}_i^T\mathbf{R}_i - \mathbf{I}\right\|_F^2$ guarantee that the columns of $\mathbf{R}_i$ form an orthonormal basis to preserve distances and angles for rotations. $\left|\operatorname*{det}(\mathbf{R}_i) - 1\right|$ affirms that $\mathbf{R}_i$ is an orientation-preserving transformation matrix and doesn't involve any reflection. + +In practice, the $T_{r}$ matrix may not be guaranteed to be a rotation matrix on a per-batch basis as the above condition is only a soft constraint. This is overcome by continuous re-orthogonalization during the training process. Moreover, alternative methods like euler-angle or quaternion representations could be used. Nevertheless, the proposed approach of enforcing orthogonality through the $L_{c}$ is a known method in the literature [1]. In the end, when provided with a 3D + +point cloud sample $\mathcal{X}_i$ , the transformation module $T_r$ computes the desired rotation matrix $\mathbf{R}_i$ while taking into account all the aforementioned constraints to ensure its validity. This process results in the generation of a learned depth map image, $\mathbf{I}_{D_i} = F(\mathcal{X}_i,\mathbf{R}_i)$ + +# 3.3 Model reprogramming + +Model reprogramming of computer vision models, especially foundational ones, is a crucial step in their effective deployment, enabling a seamless transition to new tasks and domains with minimal adjustments to existing architectures [38]. We present a novel model reprogramming approach specifically tailored for the CLIP model applied to the FSCIL task on 3D point cloud objects. It is worth noting that the CLIP model is primarily trained on RGB images rather than depth maps. Therefore, to enhance the CLIP model's effectiveness with depth maps, we incorporate pixel information into the background of the depth map images. More specifically, this modification assists the CLIP model in accurately discerning the boundary of the depth map shape information through a $C$ module, which takes the form of a U-Net [32]. To accomplish this, we initially compute the mask map $\mathbf{I}_{M_i}$ corresponding to a given depth map $\mathbf{I}_{D_i}$ , which isolates the shape boundary from the background. Next, we input $\mathbf{I}_{M_i}$ into the $C$ module to extract the pixel information for the background, denoted as $\mathbf{I}_{H_i}$ . To ensure the consistency of the generated pixels in $\mathbf{I}_{M_i}$ from the $C$ module and to prevent convergence to arbitrary values, we employ the following loss function, + +$$ +\mathcal {L} _ {g} = \frac {1}{n} \sum_ {i = 1} ^ {n} \left(\frac {\partial \mathbf {I} _ {M _ {i}}}{\partial x} - \frac {\partial \mathbf {I} _ {H _ {i}}}{\partial x}\right) ^ {2} + \left(\frac {\partial \mathbf {I} _ {M _ {i}}}{\partial y} - \frac {\partial \mathbf {I} _ {H _ {i}}}{\partial y}\right) ^ {2} \tag {2} +$$ + +where $\frac{\partial}{\partial x}$ and $\frac{\partial}{\partial y}$ denote the derivatives of the images in the $x$ and $y$ directions, respectively. This loss enforces a smoothness constraint on the gradient to separate it from foreground objects explicitly. + +It helps to learn novel classes with few examples. Subsequently, we combine the depth map $\mathbf{I}_{D_i}$ from the projection module $F$ with the extracted values $\mathbf{I}_{H_i}$ from $C$ to generate the final image, $\mathbf{I}_{R_i} = \mathbf{I}_{D_i} \odot C(M(\mathbf{I}_{D_i}))$ , where $\odot$ represents the element-wise multiplication. + +# 3.4 Prompt engineering + +Advancements in prompt engineering have unveiled significant potential, primarily driven by LLMs [18, 19]. LLMs can extract richer information from appropriately tailored prompts. With this motivation, we employ GPT to obtain more insights from plain class names and enhance the FSCIL task on 3D data. In contrast to prevailing methods [9, 10, 12] that rely solely on the class name for the FSCIL task, we extend our method by generating additional descriptions to extract more semantic information from the CLIP text encoder (denoted as $T_{e}$ ). This augmentation ensures a more comprehensive utilization of the LLM's + +capabilities, thereby enhancing the overall efficacy of our approach. In the FS-CIL, a significant challenge arises due to the scarcity of training samples for novel classes, potentially leading to overfitting. To address this issue, we propose an innovative strategy leveraging the capabilities of LLM, specifically the GPT model, and utilizing this as novel prompts for the text encoder of the CLIP model. More precisely, we utilize the GPT model to generate multiple descriptions for each class by providing it with a specific command. These prompts are formally defined for each sample as follows: + +$$ +\mathbf {p} _ {i} = \operatorname {G P T} (\text {C o m m a n d s}). \tag {3} +$$ + +Here, the Commands are structured as Describe a depth of a [class]. For every class in the dataset, we substitute the respective class names in place of '[class]' in the command and input them into GPT-4. This results in the generation of depth map-specific descriptions with comprehensive category-wise semantics. Subsequently, we feed the descriptions of each category through CLIP's textual encoder $\mathcal{T}_e$ . Finally, the generated prompts $\mathbf{p}_i$ are used in the training of the few-shot classes. + +# 3.5 Class incremental learning + +Training. For the base task, denoted as $\mathcal{Q}^1$ , we conduct training for the transformation model $T_{r}$ , the 3D backbone $B$ , and the relation module $R_{l}$ . In the case of subsequent tasks, denoted as $\mathcal{Q}^t$ , where $t > 1$ , only the relation module undergoes fine-tuning. Throughout both base and subsequent novel tasks, all other components in our proposed architecture (refer to Figure 2) remain frozen. Additionally, for each previously learned class, we randomly select a sample to be stored in a compact memory module $\mathcal{M}$ . Given the $i^{th}$ point cloud $\mathcal{X}_i^t$ of the $t^{th}$ task, the $B$ module generates the features $\mathbf{z}_i^t$ and the $V_{e}$ module produces $\mathbf{v}_i^t$ . Simultaneously, the prompts of all classes for tasks $\mathcal{Q}^1,\ldots ,\mathcal{Q}^t$ are fed into the text encoder $T_{e}$ to generate the corresponding text embeddings denoted as $\{\mathbf{e}_j\}_{j = 1}^{|\mathcal{Y}_{tl}|}$ . Then, we concatenate $\mathbf{z}_i^t$ , $\mathbf{v}_i^t$ , and $\mathbf{e}_j^t$ and forward them into the relation module $R_{l}$ , which provides a score between [0,1], representing the similarity between them. In other words, for each training sample, we generate a score against each of the classes in both the subsequent and previous tasks as $r_{ij}^{t} = \gamma \circ R\circ (\mathbf{z}_{i}^{t}\oplus \mathbf{v}_{i}\oplus \mathbf{e}_{j}), j\in \mathcal{Y}_{tl}$ , where $\mathcal{Y}_{tl} = \bigcup_{i = 1}^{t}\mathcal{Y}^{i}$ , $\oplus$ is the concatenation operator, $R$ is the relation module, and $\gamma$ is the sigmoid function. Finally, for each feature $r_{ij}$ and the corresponding ground truth $\mathbf{y}_i$ , we employ a binary cross-entropy cost function to train the model as follows, + +$$ +L _ {r} = - \frac {1}{| \mathcal {Y} _ {t l} | | S |} \sum_ {k \in \mathcal {Y} _ {t l}} \sum_ {\mathbf {y} _ {i} \in \mathcal {S}} \Bigg (1 (\boldsymbol {y} _ {i} ^ {t} = = k) \log (r _ {i k}) + \big (1 - 1 (\boldsymbol {y} _ {i} ^ {t} = = k) \big) \log (1 - r _ {i k}) \Bigg), +$$ + +where $\mathcal{S}$ denotes the set of true labels in the current task, along with the memory $\mathcal{M}$ . Finally, the total loss to train the entire pipeline is as follows, + +$$ +\mathcal {L} _ {t} = \lambda_ {1} \mathcal {L} _ {c} + \lambda_ {2} \mathcal {L} _ {g} + \lambda_ {3} \mathcal {L} _ {r} \tag {4} +$$ + +where $\lambda_1, \lambda_2$ , and $\lambda_3$ are the weighting factors for the constraint loss $(\mathcal{L}c)$ , gradient loss $(\mathcal{L}g)$ , and the relation loss $(\mathcal{L}_r)$ respectively. + +Inference. During the inference process, when given an unlabeled sample $\mathcal{X}^c$ , where $c \in \mathcal{Y}_{tl}$ , the label prediction $y^{*}$ is computed using the following equation: + +$$ +y ^ {*} = \underset {j \in \mathcal {Y} _ {t l}} {\arg \max } (R \circ (\mathbf {z} _ {c} \oplus \mathbf {v} _ {c} \oplus \mathbf {s} _ {j})), \tag {5} +$$ + +where $\mathbf{z}_c$ and $\mathbf{v}_c$ represent the calculated features for the sample $\mathcal{X}^c$ . We use a standardized prompt for all classes, specifically This is a depth map of a [class]. + +# 4 Experiments + +Datasets. We employed two synthetic 3D datasets, Modelnet40 [45] and Shape-net [6], in addition to two real scanned 3D datasets, ScanobjectNN [40] and CO3D [31]. We follow the setup proposed by [11] for Few-Shot Classification with Incremental Learning (FSCIL) on point cloud objects. Our proposed method is assessed in two distinct scenarios. Firstly, we evaluate its performance on the FSCIL task within the same dataset, where both the base and novel tasks originate from identical datasets. Secondly, we subject our method to more demanding and realistic challenges by testing it in cross-dataset scenarios. In this setting, the base task involves synthetic datasets, while the novel tasks comprise classes from real scanned few-shot datasets. + +Prompt generation. GPT-4 was employed to produce diverse class descriptions for each class. The following command was utilized: Describe a depth of a [class]. Approximately 100 class descriptions were generated for each class and utilized during the training phase. + +Implementation details. In our experimentation, we utilize Curvenet as the backbone for 3D point clouds $(B)$ [46]. The CLIP vision $(V_{e})$ and text encoder $(T_{e})$ are instantiated using the ViT-B-16 architecture [14]. The $C$ model is constructed based on the U-Net architecture [32]. The projection module $(F)$ follows the methodology proposed by Zhu et al. [51], and for the transformation module $(T_{r})$ , we employ the STN model [16]. Additionally, the relation module comprises several fully connected layers with sizes (1024, 2048, 1). We use the farthest 1024 points from 3D point cloud objects as input for all samples. For the base task, we train our proposed method (Figure 2) in an end-to-end manner for 250 epochs using the Adam optimizer with a learning rate of 0.001. For the novel task, we fine-tune our model for 200 epochs using the Adam optimizer with a learning rate of 0.001. It is important to note that we randomly shift and scale points in the input point cloud during both base and novel class training, with the additional step of randomly dropping out points. The feature vector dimensions for $B$ , $T_{e}$ , and $V_{e}$ are set to 512. Additionally, the weights of the loss functions $\lambda_{1}$ , $\lambda_{2}$ , and $\lambda_{3}$ are configured as 0.01, 1, and 1 across the entire datasets. Our experiments are conducted using the PyTorch framework. + +Evaluation metrics. Following the training of each task, we compute the accuracy by consolidating both base and novel classes. Subsequently, following the + +recommendation in [36], we determine the relative accuracy dropping rate, denoted as $\varDelta$ , using the formula $\varDelta=\frac{|acc_T-acc_0|}{acc_0}\times 100$ , where $acc_{T}$ and $acc_{0}$ denote the accuracy of the last and first tasks, respectively. The value of $\varDelta$ serves as a comprehensive metric for method evaluation. A lower relative accuracy signifies superior performance. + +# 4.1 Main results + +We conducted a comparative analysis of our approach against established methods in the field, including IL2M [2], ScaIL [3], EEIL [5], LwF [22], FACT [50], Sem-aware [9], and Microshape [12]. Notably, all these methods are tailored for image data, with the exception of Microshape [12], which was originally designed for point cloud data. + +Within dataset: In this scenario, both the base and novel tasks are derived from the same dataset. Roughly half of the classes in the dataset are assigned to the base task, while the remaining half are incrementally introduced as multiple novel tasks. It is presumed that these novel tasks encompass only a limited number of samples for training. Our proposed method has been evaluated on three datasets—ModelNet, CO3D, and Shapenet. As depicted in Table 1, our approach demonstrates superior performance across all methods in terms of the relative accuracy dropping rate, denoted as $\Delta$ , indicating the effectiveness of our proposed method. + +Cross dataset: In this arrangement, the base task centers around synthetic point cloud objects, while the novel few-shot tasks incorporate real-scanned 3D objects. Notably, the latter often exhibit incompleteness or noise in contrast to the synthetic data, which typically boast cleanliness and completeness. Consequently, this configuration introduces a more realistic and challenging scenario compared to within-dataset experiments. Three datasets are considered in this setting: Shapenet $\rightarrow$ CO3D, ModelNet40 $\rightarrow$ ScanObjectNN, and Shapenet $\rightarrow$ ScanObjectNN. As illustrated in Table 2, our method outperforms other approaches in all three setups, underscoring its efficacy in scenarios involving a domain gap between base and novel tasks. + +Table 1: Summary of FSCIL results for within-dataset experiments. + +
ModelNetCO3DShapeNet
Method2025303540Δ↓253035404550Δ↓25303540455055 Δ↓
FT89.89.74.33.33.096.776.711.23.63.21.80.899.087.025.76.81.30.90.60.4 99.5
Joint89.888.287.083.580.510.476.769.464.862.760.759.822.087.085.284.383.082.582.281.3 6.6
LwF [22]89.836.09.13.63.196.076.714.74.73.52.31.098.787.060.833.515.93.83.11.8 97.9
IL2M [2]89.865.558.452.353.640.376.731.527.718.127.121.971.487.058.645.740.750.149.449.3 43.3
ScaIL [3]89.866.864.558.756.537.176.739.534.124.130.127.564.187.056.651.844.350.346.445.4 47.8
EEIL [5]89.875.467.260.155.638.176.761.452.442.839.532.857.287.077.773.269.366.465.965.8 22.4
FACT [50]90.481.377.173.565.028.177.967.159.754.850.246.740.087.575.371.469.967.565.762.5 28.6
Sem-aware [9]91.382.274.370.064.729.176.866.959.253.649.142.944.187.274.968.969.168.066.963.8 26.8
Microshape [11]93.683.178.275.867.128.378.567.360.156.151.447.239.987.683.281.579.076.873.5 72.617.1
C3PR (ours)91.682.375.872.270.922.581.569.466.563.054.253.834.088.081.677.876.776.776.976.2715.1
+ +Table 2: Summary of FSCIL results for within-dataset and cross-dataset experiments. + +
ShapeNet → CO3DModelNet → ScanObjectNNShapeNet → ScanObjectNN
Method3944495459646974798489Δ↓26303437Δ↓44495459Δ↓
FT81.020.22.31.70.81.01.01.30.90.51.698.088.46.46.01.997.981.438.74.00.998.9
Joint81.079.578.375.275.174.872.371.370.068.867.316.988.479.774.071.219.581.482.579.878.73.3
LwF [22]81.057.419.32.31.00.90.81.31.10.81.997.788.435.85.82.597.281.447.914.05.992.8
IL2M [2]81.045.636.835.131.833.334.031.530.632.330.063.088.458.252.952.041.281.453.243.945.843.7
ScaIL [3]81.050.145.739.139.037.938.036.033.733.035.256.588.456.555.952.940.281.449.046.740.050.9
EEIL [5]81.075.269.363.260.557.953.051.951.347.847.641.288.470.261.056.835.781.474.569.863.422.1
FACT [50]81.476.070.368.165.863.563.060.158.257.555.931.389.172.568.363.528.782.374.669.966.818.8
Sem-aware [9]80.669.566.562.963.263.061.258.358.157.255.231.688.573.967.764.227.581.370.665.262.922.6
Microshape [11]82.677.973.972.767.766.265.463.460.658.157.130.989.373.268.465.127.182.574.871.267.118.7
C3PR (ours)83.680.077.875.472.872.370.367.964.964.163.224.488.375.770.667.823.284.577.875.571.914.9
+ +# 4.2 Ablation Studies + +Evaluating model reprogramming: In the conducted experiment (refer to Figure 4 (a)), we systematically omitted the model reprogramming module $C$ and directly fed the depth map image $\mathbf{I}_D$ into the vision encoder $V_{e}$ of the CLIP model. The results clearly illustrate the significance of the model reprogramming in our proposed approach, as evidenced by notable differences between the outcomes with and without this module. + +Impact of the prompt engineering: In our ablation study (see Figure 4 (b)), we investigate the influence of prompt engineering in our methodology. Specifically, we streamline the training stage by substituting the original multiple prompts per class with a single fixed prompt. The outcomes underscore the effectiveness of our GPT-based prompt engineering approach. + +Impact of canonical shape projection: Here, we evaluate our proposed canonical shape projection methodology in comparison to PointCLIPv2 [51], which generates multiple fixed-depth map projections for each input point cloud. Specifically, we eliminate the transformation module $T_r$ . Subsequently, we adopt the depth map generation method employed in PointCLIP v2, involving the use of ten fixed camera pose angles for computing ten depth map images. These images are then fed into the vision encoder $V_e$ of the CLIP model, and the relation module is applied to merge the feature vector, mirroring our proposed pipeline. Notably, our method surpasses PointCLIPv2 by a significant margin (see Figure 4 (c)), highlighting the efficacy of our learned camera pose angle. + +![](images/f2735cff78476fd9eec68ebb24ea1f7d2fb4381219dff78ddc26c7952aea34be.jpg) +Fig. 4: This experiment is performed within the Shapenet to ScanobjectNN setting. In (a), we present the influence of our novel CLIP model reprogramming strategy. (b) Highlights the impact of prompt engineering as proposed in our method. (c) Our approach is compared with PointCLIP v2 [51]. + +![](images/891842fb6002b36c8659e7ce176eb5e3ad87715498c7c2ae39b4daba34465bf7.jpg) + +![](images/42a582476af33b9e1cbc6568663637754912fe8e31dcff1b891bf7ef7170887d.jpg) + +![](images/eb73fc7a41ed88ab3af361dd4727f27eae7410e8e44844d1a11f8deb8b78f207.jpg) +Fig. 5: We visualize the quality of features (from 44 base classes of ShapeNet) generated by the CLIP vision encoder using (Right) our proposed method, C3PR, and (Left) PointCLIPv2 [51]. Our method produced a denser cluster than PointCLIP. + +![](images/bcf4f8654d88c7b467beafb144fcc896a12b208d66957f5687610ce8c46ea02e.jpg) + +Feature quality: In Figure 5, we assess the feature quality produced by the CLIP vision encoder using our proposed method, C3PR, and compare it with PointCLIPv2 [51]. As illustrated, our approach yields more discriminative clusters, enhancing overall generalization capabilities. + +Influence of $\mathcal{L}_c$ and $\mathcal{L}_g$ : In Table 3a, we highlight the impact of both $\mathcal{L}_c$ and $\mathcal{L}_g$ within our proposed pipeline. It is evident that both loss functions contribute significantly, but $\mathcal{L}_c$ holds greater importance, underscoring the significance of identifying the learned projection angle for depth map images on image foundation models. + +Model reprogramming: Conventional vs. our: In this section, we contrast our innovative model reprogramming approach with the conventional method introduced by [15]. Table 3b illustrates the superior effectiveness of our proposed model reprogramming strategy. + +Table 3: Ablation study on the influence of loss (a) and model reprogramming (b). +(a) Influence of $\mathcal{L}_c$ and $\mathcal{L}_g$ in our approach. + +
ShapeNet → ScanObjectNN
Method44495459Δ↓
Lr84.476.270.16522.9
Lr + Lg84.377.370.866.920.6
Lr + Lc85.780.274.570.817.3
Lr + Lg + Lc84.577.875.571.914.9
+ +(b) Comparing our proposed model reprogramming strategy, C3PR, with the conventional model reprogramming technique employed in [15]. + +
ShapeNet → ScanObjectNN
Method44495459Δ↓
[15]85.780.274.570.817.3
C3PR (ours)84.577.875.571.914.9
+ +# 4.3 Discussion + +Point-cloud classification: Although our learning projection module $Tr$ and model reprogramming module $C$ have demonstrated effectiveness in the FSCIL context, they can also be applied in the point-cloud classification setting (shown in the table 4a. However, it is worth to note that our proposed prompt engineering strategy is specifically tailored to improve performance in FSCIL problems. + +# Module $C$ role: + +The role of Module $C$ is twofold: it learns background information for the depth maps and smoothens the foreground. Specifically, we combine the depth + +map $\mathbf{I}_{D_i}$ from the projection module $F$ with the extracted values $\mathbf{I}_{T_i}$ from Module $C$ to generate the final image, $\mathbf{I}_{R_i} = \mathbf{I}_{D_i} \odot C(M(\mathbf{I}_{D_i}))$ , where $\odot$ represents element-wise multiplication. To this end, we conducted experiments to demonstrate the impact of Module $C$ in our proposed method (see table 4b). + +(a) The performance of our proposed method on point-cloud classification in ModelNet40. + +(b) The impact of module $C$ in our proposed method. + +Table 4: Ablation on point-cloud classification (a) and impact of $C$ module (b). + +
MethodAccuracy
DGCNN [43]92.9
CurveNet [46]93.8
Ours94.6
+ +
ShapeNet → ScanObjectNN
Method44495459Δ↓
Without C84.576.373.268.918.4
With C84.577.875.571.914.9
+ +**Limitation:** The proposed method exhibits several limitations that warrant discussion. Firstly, its reliance on storing exemplars in memory, with one sample per class, and addressing the forgetting issue raises the question of exploring exemplar-free alternatives, which could offer intriguing avenues for further research. Secondly, our method would perform worse on new concepts from significantly different domains, synthetic to real-scan datasets, as can be seen in Table 2. However, it works well when the new concepts come from similar domains, synthetic to synthetic datasets, as can be seen in Table 1. + +# 5 Conclusion + +In this paper, we leverage the CLIP architecture of image-to-text matching to address the FSCIL problem on 3D point cloud objects. Traditional approaches usually project a 3D object on a 2D plane to produce images from different camera angles. Among those 2D images, some have bad/poor camera angles, making them difficult to recognize. Therefore, we propose a learning mechanism to find one camera projection that best describes the 3D objects. In doing so, we adopt a model reprogramming approach that perturbs the input 2D images (from the best projection angle) with background-foreground colorization (segmentation) without modifying or fine-tuning the original CLIP architectures. Our reprogramming approach is different from conventional approaches because our custom loss produces a uniform background color instead of the noisy background of previous works. In the text pipeline, we use the CLIP text encoder and train the whole architecture end-to-end. Instead of using a single class name-based text input, we suggest using prompts based on GPT-based description. We propose a prompt engineering approach to train the networks with 3D object and text description pairs. Finally, we evaluate our proposed method using 3D synthetic datasets, ModelNet and ShapeNet, and two 3D real-world-scanned datasets, ScanObjectNN and CO3D, and report state-of-the-art performances. In future, we will investigate model reprogramming for other vision problems beyond FSCIL, like segmentation and domain adaptation on 3D point clouds. + +Acknowledgment. Mehrtash Harandi thanks the Australian Research Council for support through the Discovery Program (DP230101176). + +# References + +1. Bansal, N., Chen, X., Wang, Z.: Can we gain more from orthogonality regularizations in training deep networks? In: Bengio, S., Wallach, H., Larochelle, H., Grauman, K., Cesa-Bianchi, N., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 31. Curran Associates, Inc. (2018), https://proceedings.neurips.cc/paper_files/paper/2018/file/bf424cb7b0dea050a42b9739eb261a3a-Paper.pdf +2. Belouadah, E., Popescu, A.: II2m: Class incremental learning with dual memory. In: CVPR (2019) +3. Belouadah, E., Popescu, A.: Scail: Classifier weights scaling for class incremental learning. In: WACV (2020) +4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 1877-1901. Curran Associates, Inc. (2020), https://proceedings.neurips.cc/paper_files/paper/2020/file/1457c0d6bcbd4967418bf8ac142f64a-Paper.pdf +5. Castro, F.M., Marín-Jiménez, M.J., Guil, N., Schmid, C., Alahari, K.: End-to-end incremental learning. In: ECCV (2018) +6. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015) +7. Chen, K., Lee, C.G.: Incremental few-shot learning via vector quantization in deep embedded space. In: ICLR (2021) +8. Chen, P.Y.: Model reprogramming: Resource-efficient cross-domain machine learning (2023) +9. Cheraghian, A., Rahman, S., Fang, P., Roy, S.K., Petersson, L., Harandi, M.: Semantic-aware knowledge distillation for few-shot class-incremental learning. In: CVPR (2021) +0. Cheraghian, A., Rahman, S., Ramasinghe, S., Fang, P., Simon, C., Petersson, L., Harandi, M.: Synthesized feature based few-shot class-incremental learning on a mixture of subspaces. In: ICCV (2021) +1. Chowdhury, T., Cheraghian, A., Ramasinghe, S., Ahmadi, S., Saberi, M., Rahman, S.: Few-shot class-incremental learning for 3d point cloud objects. In: Avidan, S., Brostow, G., Cissé, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 204-220. Springer Nature Switzerland, Cham (2022) +2. Chowdhury, T., Cheraghian, A., Ramasinghe, S., Ahmadi, S., Saberi, M., Rahman, S.: Few-shot class-incremental learning for 3d point cloud objects. In: ECCV (2022) +3. Dinh, T., Seo, D., Du, Z., Shang, L., Lee, K.: Improved input reprogramming for gan conditioning (2022) +4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021) +5. Elsayed, G.F., Goodfellow, I., Sohl-Dickstein, J.: Adversarial reprogramming of neural networks. In: International Conference on Learning Representations (2019), https://openreview.net/forum?id=Syx_Ss05tm + +16. Jaderberg, M., Simonyan, K., Zisserman, A., kavukcuoglu, k.: Spatial transformer networks. In: Cortes, C., Lawrence, N., Lee, D., Sugiyama, M., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 28. Curran Associates, Inc. (2015), https://proceedings.neurips.cc/paper_files/paper/2015/file/33ceb07bf4eeb3da587e268d663aba1a-Paper.pdf +17. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 4904-4916. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/jia21b.html +18. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision (ECCV) (2022) +19. Lee, D.H., Pujara, J., Sewak, M., White, R.W., Jauhar, S.K.: Making large language models better data creators. In: The 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP) (2023), https://openreview.net/forum?id=2Rdfdri2oT +20. Lee, K.Y., Zhong, Y., Wang, Y.X.: Do pre-trained models benefit equally in continual learning? In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 6485-6493 (January 2023) +21. Li, Y., Bu, R., Sun, M., Wu, W., Di, X., Chen, B.: Pointcnn: Convolution on x-transformed points. In: NeurIPS (2018) +22. Li, Z., Hoiem, D.: Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence (2018) +23. Liu, Y., Fan, B., Xiang, S., Pan, C.: Relation-shape convolutional neural network for point cloud analysis. In: CVPR (2019) +24. Mazumder, P., Singh, P., Rai, P.: Few-shot lifelong learning. In: AAAI (2021) +25. Pei, Y., Qing, Z., CEN, J., Wang, X., Zhang, S., Wang, Y., Tang, M., Sang, N., Qian, X.: Learning a condensed frame for memory-efficient video class-incremental learning. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=1CGYC7pXWNQ +26. Poulenard, A., Rakotosaona, M.J., Ponty, Y., Ovsjanikov, M.: Effective rotation-invariant point cnn with spherical harmonics kernels. In: 3DV (2019) +27. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR (2017) +28. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: NeurIPS (2017) +29. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 8748-8763. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/radford21a.html +30. Rao, Y., Lu, J., Zhou, J.: Spherical fractal convolutional neural networks for point cloud recognition. In: CVPR (2019) +31. Reizenstein, J., Shapovalov, R., Henzler, P., Sbordone, L., Labatut, P., Novotny, D.: Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In: ICCV (2021) + +32. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F. (eds.) Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015. pp. 234-241. Springer International Publishing, Cham (2015) +33. Singh, A., Hu, R., Goswami, V., Couairon, G., Galuba, W., Rohrbach, M., Kiela, D.: FLAVA: A foundational language and vision alignment model. In: CVPR (2022) +34. Sung, F., Yang, Y., Zhang, L., Xiang, T., Torr, P.H., Hospedales, T.M.: Learning to compare: Relation network for few-shot learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2018) +35. Tan, Y., Xiang, X.: Cross-domain few-shot incremental learning for point-cloud recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 2307-2316 (January 2024) +36. Tan, Z., Ding, K., Guo, R., Liu, H.: Graph few-shot class-incremental learning. In: Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining (2022) +37. Tao, X., Hong, X., Chang, X., Dong, S., Wei, X., Gong, Y.: Few-shot class-incremental learning. In: CVPR (2020) +38. Tsai, Y.Y., Chen, P.Y., Ho, T.Y.: Transfer learning without knowing: Reprogramming black-box machine learning models with scarce data and limited resources. In: III, H.D., Singh, A. (eds.) Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 9614-9624. PMLR (13-18 Jul 2020), https://proceedings.mlr.press/v119/tsai20a.html +39. Tsai, Y.Y., Chen, P.Y., Ho, T.Y.: Transfer learning without knowing: Reprogramming black-box machine learning models with scarce data and limited resources. In: Proceedings of the 37th International Conference on Machine Learning. ICML'20, JMLR.org (2020) +40. Uy, M.A., Pham, Q.H., Hua, B.S., Nguyen, D.T., Yeung, S.K.: Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In: ICCV (2019) +41. Wang, C., Samari, B., Siddiqi, K.: Local spectral graph convolution for point set feature learning. In: ECCV (2018) +42. Wang, R., Duan, X., Kang, G., Liu, J., Lin, S., Xu, S., Lu, J., Zhang, B.: Attriclip: A non-incremental learner for incremental knowledge learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3654-3663 (June 2023) +43. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog) (2019) +44. Wu, W., Qi, Z., Fuxin, L.: PointCONV: Deep convolutional networks on 3D point clouds. In: CVPR (2019) +45. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3D ShapeNets: A deep representation for volumetric shapes. In: CVPR (2015) +46. Xiang, T., Zhang, C., Song, Y., Yu, J., Cai, W.: Walk in the cloud: Learning curves for point clouds shape analysis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 915-924 (October 2021) +47. Xu, Y., Fan, T., Xu, M., Zeng, L., Qiao, Y.: SpiderCNN: Deep learning on point sets with parameterized convolutional filters. In: ECCV (2018) +48. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. arXiv preprint arXiv:2112.02413 (2021) + +49. Zhang, Y., Rabbat, M.: A graph-cnn for 3d point cloud classification. In: ICASSP (2018) +50. Zhou, D.W., Wang, F.Y., Ye, H.J., Ma, L., Pu, S., Zhan, D.C.: Forward compatible few-shot class-incremental learning. In: CVPR (2022) +51. Zhu, X., Zhang, R., He, B., Guo, Z., Zeng, Z., Qin, Z., Zhang, S., Gao, P.: Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. arXiv preprint arXiv:2211.11682 (2022) \ No newline at end of file diff --git a/2024/Canonical Shape Projection is All You Need for 3D Few-shot Class Incremental Learning/images.zip b/2024/Canonical Shape Projection is All You Need for 3D Few-shot Class Incremental Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..28b083892876ac7a964bf5b2e93cc858b7d0a34d --- /dev/null +++ b/2024/Canonical Shape Projection is All You Need for 3D Few-shot Class Incremental Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff3e161c406c020637d9b03e0b14e2dcdb0de4287a19b4a6d2dd241bd9b56fc5 +size 434967 diff --git a/2024/Canonical Shape Projection is All You Need for 3D Few-shot Class Incremental Learning/layout.json b/2024/Canonical Shape Projection is All You Need for 3D Few-shot Class Incremental Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8830b0697ac99d44cd6e8cb4cbc54a5ceacffb1e --- /dev/null +++ b/2024/Canonical Shape Projection is All You Need for 3D Few-shot Class Incremental Learning/layout.json @@ -0,0 +1,10947 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 475, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 475, + 148 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 475, + 148 + ], + "type": "text", + "content": "Canonical Shape Projection is All You Need for 3D Few-shot Class Incremental Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 167, + 468, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 167, + 468, + 204 + ], + "spans": [ + { + "bbox": [ + 145, + 167, + 468, + 204 + ], + "type": "text", + "content": "Ali Cheraghian\\*,1,2, Zeeshan Hayden\\*,1,2, Sameera Ramasinghe\\*3, Shafin Rahman4, Javad Jafaryahya5, Lars Petersson2, and Mehrtash Harandi6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 139, + 213, + 474, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 213, + 474, + 270 + ], + "spans": [ + { + "bbox": [ + 139, + 213, + 474, + 270 + ], + "type": "text", + "content": "1 Data61, CSIRO, Australia, 2 Australian National University, 3 Amazon, 4 North South University, 5 University of Technology Sydney, 6 Monash University {ali.cheraghian, zeeshan.hayder, lars.petersson}@data61.csiro.au, ramasisa@amazon.com, shafin.rahman@northsouth.edu, javad.jafaryahya@student.uts.edu.au, mehrtash.harandi@monash.edu" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 300, + 453, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 300, + 453, + 541 + ], + "spans": [ + { + "bbox": [ + 160, + 300, + 453, + 541 + ], + "type": "text", + "content": "Abstract. In recent years, robust pre-trained foundation models have been successfully used in many downstream tasks. Here, we would like to use such powerful models to address the problem of few-shot class incremental learning (FSCIL) tasks on 3D point cloud objects. Our approach is to reprogram the well-known CLIP-based foundation model (trained on 2D images and text pairs) for this purpose. The CLIP model works by ingesting 2D images, so to leverage it in our context, we project the 3D object point cloud onto 2D image space to create proper depth maps. For this, prior works consider a fixed and non-trainable set of camera poses. In contrast, we propose to train the network to find a projection that best describes the object and is appropriate for extracting 2D image features from the CLIP vision encoder. Directly using the generated depth map is not suitable for the CLIP model, so we apply the model reprogramming paradigm to the depth map to augment the foreground and background to adapt it. This removes the need for modification or fine-tuning of the foundation model. In the setting we have investigated, we have limited access to data from novel classes, resulting in a problem with overfitting. Here, we address this problem via the use of a prompt engineering approach using multiple GPT-generated text descriptions. Our method, C3PR, successfully outperforms existing FSCIL methods on ModelNet, ShapeNet, ScanObjectNN, and CO3D datasets. The code is available at https://github.com/alichr/C3PR." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 552, + 453, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 552, + 453, + 574 + ], + "spans": [ + { + "bbox": [ + 159, + 552, + 453, + 574 + ], + "type": "text", + "content": "Keywords: 3D shape projection " + }, + { + "bbox": [ + 159, + 552, + 453, + 574 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 552, + 453, + 574 + ], + "type": "text", + "content": " Model reprogramming " + }, + { + "bbox": [ + 159, + 552, + 453, + 574 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 552, + 453, + 574 + ], + "type": "text", + "content": " Few-shot class incremental learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 596, + 230, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 596, + 230, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 596, + 230, + 609 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 622, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 622, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 622, + 482, + 647 + ], + "type": "text", + "content": "Maintaining a balance between acquiring new concepts and retaining existing knowledge presents a significant challenge for machine learning algorithms, a" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 254, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 254, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 254, + 665 + ], + "type": "text", + "content": "* denotes equal contribution." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 185, + 116, + 430, + 358 + ], + "blocks": [ + { + "bbox": [ + 185, + 116, + 430, + 358 + ], + "lines": [ + { + "bbox": [ + 185, + 116, + 430, + 358 + ], + "spans": [ + { + "bbox": [ + 185, + 116, + 430, + 358 + ], + "type": "image", + "image_path": "362c13261a5cf4cefef74e2dd6ffe2218a4d9886d7caea7e9e702f19356f56dd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 366, + 483, + 456 + ], + "lines": [ + { + "bbox": [ + 130, + 366, + 483, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 483, + 456 + ], + "type": "text", + "content": "Fig. 1: (a) Traditional methods using image-based models for point cloud processing often generate multiple depth maps from fixed camera poses, potentially causing feature alignment issues and difficulties in recognizing 3D objects from unsuitable camera angles. (b) In contrast, our proposed approach (C3PR) determines the most suitable camera pose for projecting a single depth map by leveraging feedback from the foundation model. (c) To improve the foundation model's performance in handling 3D data for the FSCIL task, we introduce a novel model reprogramming technique (without modifying or fine-tuning the foundation model) to further refine the depth map." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 483, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 628 + ], + "type": "text", + "content": "task that comes naturally to humans. This task becomes more difficult when only a few samples are available for new concepts. In the literature, this problem is referred to as few-shot class incremental learning (FSCIL) [7,9,10,24,37]. Substantial effort on this topic has been invested, particularly on image data [7,9,10, 24,37], whereas relatively few studies are tailored specifically towards point cloud data. For example, in the 2D domain, there are vision-language models such as CLIP [29], ALIGN [17], and FLAVA [33] models with an inherent capacity to generate highly versatile representations and demonstrate zero-shot capabilities, all without parameter fine-tuning [20, 25, 42]. Gathering training data for 3D point clouds is substantially more challenging than for image data, and here we explore whether the CLIP model can help alleviate this issue. In this paper, we attempt to leverage CLIP for FSCIL on 3D Point Cloud Objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Popular approaches capture 2D images from a fixed set of camera poses and then pass the images to 2D foundation models for feature extraction [48, 51]. The main drawback of such a strategy is that camera poses are not adaptable" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 251, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 251, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 251, + 102 + ], + "type": "text", + "content": "A. Cheraghian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": "to the target dataset, hindering capturing the objects in poses ideal for feature extraction (see Fig. 1 (a)). To address this problem, we propose to learn the camera projection pose based on the dataset that best describes the 3D objects (see Fig. 1 (b)). Our proposed model reprogramming approach is advantageous for representing real-world scanned 3D objects where point cloud data contains sensor noise (see Fig. 1 (c)). The unique camera pose can still capture the global shape of noisy 3D objects, even though the object belongs to novel classes. For language/text modality, many existing works [9-11] leverages vision-language learning to address the overfitting issue associated with few-shot novel classes. We argue that vital knowledge lies within the language domain, which may aid in the training process of the vision domain, particularly for few-shot classes, to alleviate overfitting. However, thus far, existing methods have only utilized class names from the language domain [9, 10, 12]. In contrast, we propose a method based on prompt engineering to better leverage the language domain." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 285, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 597 + ], + "type": "text", + "content": "This paper presents a novel approach for extracting valuable insights from foundational models like CLIP, focusing on FSCIL on 3D point cloud data. As CLIP's training data comprises millions of image-text pairs, our target is to transfer this learned knowledge as a complete black box (without knowing or modifying pre-trained parameters) for 3D data. To this end, we propose reprogramming [38] the CLIP model, specifically for point cloud objects. The proposed approach is different from prior attempts [48,51] where they used a fixed set of camera poses for 3D to 2D image projection and did not augment the input data for better usage of the CLIP encoder. Our C3PR method emphasizes generating depth map images with consistent, informative, background pixels, diverging from conventional model reprogramming techniques that frequently produce random pixel information for the background. We observed that maintaining consistent background pixels contributes to improved generalization, particularly for few-shot novel classes. These steps indirectly add essential perturbation to the input depth map images so that the black box CLIP vision encoder can operate without fine-tuning. Furthermore, we propose a novel approach that leverages information from the language domain using a Large Language Model, GPT, to combat overfitting issues in few-shot novel classes. Specifically, we generate multiple descriptive prompts (instead of single class name-based descriptions as used in [11]) for base and novel classes. It also compensates for the issue with limited data during novel class learning. The generated prompts are then employed in the CLIP text encoder, reducing overfitting during incremental training sessions. Both the vision and text pipelines of our method can be trained end-to-end. We report state-of-the-art performances when evaluated on 3D synthetic datasets, ModelNet [45] and ShapeNet [6], and two 3D real world-scanned datasets, ScanObjectNN [40] and Common Objects in 3D (CO3D) [31]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 599, + 380, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 599, + 380, + 611 + ], + "spans": [ + { + "bbox": [ + 146, + 599, + 380, + 611 + ], + "type": "text", + "content": "The main contributions of our proposed method are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 627, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 627, + 477, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 627, + 477, + 639 + ], + "spans": [ + { + "bbox": [ + 138, + 627, + 477, + 639 + ], + "type": "text", + "content": "- Learning the preferred object-based camera pose for 3D to 2D projections;" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 641, + 481, + 665 + ], + "type": "text", + "content": "- A novel model reprogramming approach for 2D CLIP encoders targeting 3D point cloud data;" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "text", + "content": "Canonical Shape Projection for 3D FSCIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 482, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 482, + 140 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 482, + 140 + ], + "type": "text", + "content": "- Prompt engineering based on a GPT model for novel few-shot classes to mitigate overfitting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 159, + 234, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 159, + 234, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 159, + 234, + 172 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 186, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 186, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 482, + 389 + ], + "type": "text", + "content": "3D point cloud processing: In recent years, there has been a surge in research focused on leveraging deep learning for the direct processing of 3D point cloud objects. Notable contributions include PointNet [27] by Qi et al., which introduced the use of multi-layer perceptron (MLP) networks for 3D point cloud processing. However, this approach overlooked the inherent local structures within the input data. Subsequent techniques were developed to address this limitation, including PointNet++ [28], which hierarchically extracts features to take advantage of local information. Building upon this, several works [21,23,26,30,44,47] have proposed convolution strategies for extracting local information. PointConv [44] introduced a novel convolution operation that locates a Monte Carlo estimation of the hidden continuous 3D convolution based on a powerful sampling technique. SFCNN [30] proposed graph convolution on spherical points. Other papers [41,43,49] treat each point cloud as a graph vertex to learn features in either spatial or spectral domains. DGCNN [43] constructs a graph in feature space and dynamically updates it using MLP for each edge, while PointGCN [49] suggests an approach that generates the graph using k-nearest neighbors from a point cloud to capture local information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 390, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 605 + ], + "type": "text", + "content": "Few-shot class-incremental learning: Tao et al. [37] were the pioneers in introducing the FSCIL setting for image data. They proposed the use of a neural gas network to address the issue of forgetting by preserving the topology of classes based on their feature vectors. Following this, in [10], a mixture of the subspaces are generated based on the training data distribution. In a related vein, Mazumder et al [24] devised a method that selectively adapts a subset of model parameters for learning novel classes, thereby curbing overfitting. They also combat forgetting by freezing critical parameters in the model. Cheraghian et al. [9] proposed a novel vision-language strategy, incorporating class semantic information from language space using a distillation technique to alleviate the impact of catastrophic forgetting. They also integrated an attention mechanism to counter overfitting on few-shot novel tasks. Introduced in [12], the Specific FSCIL tailored for 3D point cloud data presents the innovative concept of Microshape, which captures essential details from the primary task, effectively addressing both forgetting and overfitting challenges. Additionally, the work by [35] delves into cross-domain FSCIL applied to point-cloud recognition. This study dissects catastrophic forgetting into two distinct components: base class forgetting and incremental class forgetting, offering separate mitigation strategies." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 606, + 483, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 483, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 483, + 665 + ], + "type": "text", + "content": "Model reprogramming: Foundation model reprogramming has recently gained attention [8], primarily due to its reduced computational and time requirements compared to the fine-tuning approach. Prior to this, Elsayed et al. [15] introduced adversarial reprogramming. This technique involves mapping ImageNet labels to task labels, aiming to perplex the ImageNet classifier and induce be" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "A. Cheraghian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "havior akin to a CIFAR-10 or MNIST classifier. Dinh et al. [13] proposed that input reprogramming is more effective in situations with limited labeled data than training from scratch and fine-tuning. Tsai et al. [39] introduced a technique called black-box adversarial reprogramming. This method involves transforming the input and integrating label mapping mechanisms into the output. In this paper, we present a novel foundation model reprogramming approach tailored for the CLIP model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 218, + 202, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 218, + 202, + 230 + ], + "spans": [ + { + "bbox": [ + 131, + 218, + 202, + 230 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "text", + "content": "In this section, we first introduce the FSCIL problem settings. Let's consider a sequence of " + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "text", + "content": " tasks denoted by " + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} = \\{\\mathcal{Q}^1,\\mathcal{Q}^2,\\dots ,\\mathcal{Q}^T\\}" + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^t" + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "text", + "content": " represents the set of classes in task " + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}^t" + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^i\\cap \\mathcal{V}^j = \\emptyset" + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "text", + "content": ". Additionally, each class in all tasks is associated with a set of prompt class descriptions generated by an LLM, e.g., GPT [4], denoted as " + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^t" + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "text", + "content": ". Thus, we can represent each task as a tuple " + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}^t = \\{\\mathcal{X}_i^t,\\mathbf{y}_i^t,\\mathbf{p}_i^t\\}_{i = 1}^{n_t}" + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i^t = \\{\\mathbf{x}_{i,j}^t\\}_{j = 1}^l" + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "text", + "content": " refers to a 3D point cloud object with coordinates " + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i,j}^{t}\\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "text", + "content": ". Moreover, " + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i^t\\in \\mathcal{V}^t" + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i^t\\in \\mathcal{P}^t" + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "text", + "content": " correspond to the label of the point cloud and its corresponding class prompt description, respectively. In the proposed FSCIL framework, " + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}^1" + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "text", + "content": " is the base task (" + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "text", + "content": "), where the model undergoes training on a large-scale synthetic 3D dataset. For " + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "inline_equation", + "content": "t > 1" + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "text", + "content": ", the training data are drawn from real-world 3D point clouds consisting of only a few instances. The model is trained sequentially across the tasks " + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "inline_equation", + "content": "t = 1,\\ldots ,T" + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "text", + "content": ". However, during the " + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "text", + "content": "-th task, the model is exposed to " + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^t" + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^t" + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{P}^1,\\mathcal{P}^2,\\dots ,\\mathcal{P}^t\\}" + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "text", + "content": ". During inference, the model trained on the current task " + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}^t" + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "text", + "content": " is expected to classify test samples from both the current and previous tasks, i.e., " + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{Q}^1,\\mathcal{Q}^2,\\dots ,\\mathcal{Q}^t\\}" + }, + { + "bbox": [ + 130, + 243, + 483, + 426 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 442, + 241, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 442, + 241, + 453 + ], + "spans": [ + { + "bbox": [ + 131, + 442, + 241, + 453 + ], + "type": "text", + "content": "3.1 Model overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": "In this section, we present an overview of our proposed methodology, illustrated in Fig. 2. We begin with a point cloud sample denoted as " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i^t" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": ", which is passed through the projection module " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": ". This module utilizes the rotation matrix " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i^t" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " derived from the transformation module " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "T_r" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " to generate a depth map labelled as " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{D_i}^t" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": ". For the projection module " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": ", we use the method proposed by [51], which is fixed during the training stage. The transformation module " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "T_r" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " is crucial for determining the appropriate angle " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " for projection. Additionally, a soft constraint loss function " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " is applied to the rotation matrix to ensure that the transformation module generates a valid rotation matrix. Following this, we extract a binary mask " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{M_i}^t" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " from the depth map. Within this mask " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": ", background pixels are assigned a value of zero, while non-zero depth map values are set to one. Subsequently, this mask is fed into a model reprogramming module " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " to add pixel information to the background of the depth map. To ensure that the generated embedding does not converge to a noisy image and follows the pattern of the mask input, an image gradient consistency loss function " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_g" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " is used between the input and output of the module " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": ". The resulting image embedding " + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{H_i}^t" + }, + { + "bbox": [ + 130, + 462, + 482, + 668 + ], + "type": "text", + "content": " is then merged with the original depth map in the subsequent stage, resulting in" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 90, + 447, + 102 + ], + "type": "text", + "content": "Canonical Shape Projection for 3D FSCIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 228 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 228 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 228 + ], + "type": "image", + "image_path": "14f2f243edd12565d6f9449f7e653a655f888504c9340c334d5c0d6bc19201b5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "lines": [ + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "text", + "content": "Fig. 2: Beginning with a point cloud sample " + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i^t" + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "text", + "content": ", we use the projection module " + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "text", + "content": " to generate a depth map " + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{D_i}^t" + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "text", + "content": " via the rotation matrix " + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i^t" + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "text", + "content": " from the transformation module " + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "inline_equation", + "content": "T_r" + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "text", + "content": ". A mask " + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{M_i}^t" + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "text", + "content": " is then extracted from the depth map. This mask is input into module " + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "text", + "content": " and the resulting image is merged with the depth map, yielding the image " + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{R_i}^t" + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "text", + "content": ". Next, " + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{R_i}^t" + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "text", + "content": " is input to the CLIP vision encoder model " + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "inline_equation", + "content": "V_e" + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "text", + "content": " to generate a vision embedding " + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i^t \\in \\mathbb{R}^m" + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "text", + "content": ". Simultaneously, " + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i^t" + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "text", + "content": " is fed into " + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "text", + "content": " to create a point cloud representation " + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^t \\in \\mathbb{R}^m" + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "text", + "content": ". Moreover, the GPT model generates multiple class name descriptions, which are then stored in a prompt bank. To this end, prompts " + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}_1, \\mathbf{p}_2, \\dots, \\mathbf{p}_{|\\mathcal{Y}_{tl}|}\\}" + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "text", + "content": " for all old classes and the current task " + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "text", + "content": " are forwarded into the text encoder " + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "inline_equation", + "content": "T_e" + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "text", + "content": ", producing corresponding text embeddings " + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{e}_1, \\mathbf{e}_2, \\dots, \\mathbf{e}_{|\\mathcal{Y}_{tl}|}\\}" + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_i^t \\in \\mathbb{R}^m" + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "text", + "content": ". Finally, " + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i^t" + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^t" + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_j^t" + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "text", + "content": " are concatenated and passed through the relation module " + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "inline_equation", + "content": "R_l" + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "text", + "content": " to generate a scalar value " + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "inline_equation", + "content": "r_i^t" + }, + { + "bbox": [ + 130, + 244, + 482, + 381 + ], + "type": "text", + "content": ", serving as a similarity score." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "text", + "content": "the production of a reprogrammed depth map image denoted as " + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{R_i}^t" + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "text", + "content": ". After that, " + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{R_i}^t" + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "text", + "content": " is input into the vision encoder of the CLIP model " + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "inline_equation", + "content": "V_{e}" + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "text", + "content": " to generate a vision embedding represented by " + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i^t\\in \\mathbb{R}^m" + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "text", + "content": ". Simultaneously, the point cloud input " + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i^t" + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "text", + "content": " is forwarded to a 3D point cloud backbone, for instance, Curvenet [46], to generate a feature description " + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^t\\in \\mathbb{R}^m" + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "text", + "content": ". At the same time, GPT is employed to generate multiple descriptions for each class using the command: describe a depth map of a [class], thereby creating several prompts, which are stored in the prompt bank module. Then, the prompts of all classes " + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}_j\\}_{j = 1}^{|\\mathcal{Y}_{tl}|}" + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_{tl} = \\bigcup_{i = 1}^{t}\\mathcal{Y}^{i}" + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "text", + "content": ", observed up to task " + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "text", + "content": " are fed into the text encoder " + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "inline_equation", + "content": "T_{e}" + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "text", + "content": " to generate the corresponding text embeddings denoted as " + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{e}_j\\}_{j = 1}^{|\\mathcal{Y}_{tl}|}" + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_j^t\\in \\mathbb{R}^m" + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "text", + "content": ". The prompts are generated using LLMs, such as GPT [4], given a command: Describe a depth map of a [class]. In the subsequent stage, all " + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i^t" + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^t" + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_i^t" + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "text", + "content": " are concatenated and passed through the relation module [34] " + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "inline_equation", + "content": "R_{l}" + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "text", + "content": " to generate a scalar value " + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "inline_equation", + "content": "r_i^t" + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "text", + "content": " ranging from 0 to 1. This value serves as a measure of the similarity between the visual and prompt feature embeddings. Finally, a binary cross-entropy loss function " + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 130, + 404, + 482, + 605 + ], + "type": "text", + "content": " is used to train the entire proposed pipeline." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 621, + 316, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 316, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 316, + 634 + ], + "type": "text", + "content": "3.2 Canonical 3D shape projection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "A 3D-to-2D projection matrix can be designed to project any 3D point cloud on a 2D plane considering any camera angle. Nevertheless, the feature embedding" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "A. Cheraghian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 184, + 114, + 430, + 229 + ], + "blocks": [ + { + "bbox": [ + 184, + 114, + 430, + 229 + ], + "lines": [ + { + "bbox": [ + 184, + 114, + 430, + 229 + ], + "spans": [ + { + "bbox": [ + 184, + 114, + 430, + 229 + ], + "type": "image", + "image_path": "d7a377aabb210e3f2f7785bbb9c27189deaf50fd77d8c0e1391a395fdda77b67.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 237, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 131, + 237, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 131, + 237, + 482, + 304 + ], + "type": "text", + "content": "Fig. 3: We demonstrate the sensitivity of the CLIP model to the angle of depth map projection. Here, we consider two random projections to generate depth maps. Additionally, we assume that there are three classes: car, guitar, and person. Even though the number of classes is limited to only three, the CLIP model can only correctly classify two of these depth maps. This highlights the importance of finding the optimum projection angle to fully leverage the CLIP model for 3D point cloud processing." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "text", + "content": "produced by the visual encoder of large image-based foundation models (e.g., CLIP) can be affected by the content that can be controlled by the projection angle. Figure 3 shows that an adjustment is pivotal due to the CLIP model's sensitivity to the projection angle, especially during depth map processing. We are interested in finding the optimal projection angle to enhance the performance of the CLIP model. When projecting from a 3D point cloud to 2D images, we leverage the mapping function introduced by [51]. We provide optimum projection angle " + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "text", + "content": " via a rotation matrix " + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{R} = T_r(\\mathcal{X})" + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "text", + "content": ", generated by the transformation module " + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "inline_equation", + "content": "T_r" + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "text", + "content": " to ascertain the most effective projection angle for converting a 3D point cloud into a depth map. The generated " + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "text", + "content": " is then applied to the given 3D point to derive the appropriate angle for the projection module " + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "text", + "content": ". To ensure that " + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "text", + "content": " is a valid rotation matrix, we apply the following loss function at the output of the transformation module " + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "inline_equation", + "content": "T_r" + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 206, + 492, + 481, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 492, + 481, + 526 + ], + "spans": [ + { + "bbox": [ + 206, + 492, + 481, + 526 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(\\left\\| \\mathbf {R} _ {i} ^ {T} \\mathbf {R} _ {i} - \\mathbf {I} \\right\\| _ {\\mathcal {F}} ^ {2} + \\left| \\det (\\mathbf {R} _ {i}) - 1 \\right|\\right) \\tag {1}", + "image_path": "7b243510f47154400647e0c55d4a75ac597d7ce179fb631caaa9663c84e6b406.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 529, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 482, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 529, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i = T_r(\\mathcal{X}_i)" + }, + { + "bbox": [ + 130, + 529, + 482, + 593 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 529, + 482, + 593 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 529, + 482, + 593 + ], + "type": "text", + "content": " is the number of samples in the training set, " + }, + { + "bbox": [ + 130, + 529, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_{\\mathcal{F}}^2" + }, + { + "bbox": [ + 130, + 529, + 482, + 593 + ], + "type": "text", + "content": " denotes the Frobenius norm, " + }, + { + "bbox": [ + 130, + 529, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 529, + 482, + 593 + ], + "type": "text", + "content": " is the identity matrix. The orthogonality condition " + }, + { + "bbox": [ + 130, + 529, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\left\\| \\mathbf{R}_i^T\\mathbf{R}_i - \\mathbf{I}\\right\\|_F^2" + }, + { + "bbox": [ + 130, + 529, + 482, + 593 + ], + "type": "text", + "content": " guarantee that the columns of " + }, + { + "bbox": [ + 130, + 529, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i" + }, + { + "bbox": [ + 130, + 529, + 482, + 593 + ], + "type": "text", + "content": " form an orthonormal basis to preserve distances and angles for rotations. " + }, + { + "bbox": [ + 130, + 529, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\left|\\operatorname*{det}(\\mathbf{R}_i) - 1\\right|" + }, + { + "bbox": [ + 130, + 529, + 482, + 593 + ], + "type": "text", + "content": " affirms that " + }, + { + "bbox": [ + 130, + 529, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i" + }, + { + "bbox": [ + 130, + 529, + 482, + 593 + ], + "type": "text", + "content": " is an orientation-preserving transformation matrix and doesn't involve any reflection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "In practice, the " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "T_{r}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " matrix may not be guaranteed to be a rotation matrix on a per-batch basis as the above condition is only a soft constraint. This is overcome by continuous re-orthogonalization during the training process. Moreover, alternative methods like euler-angle or quaternion representations could be used. Nevertheless, the proposed approach of enforcing orthogonality through the " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "L_{c}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " is a known method in the literature [1]. In the end, when provided with a 3D" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 448, + 102 + ], + "type": "text", + "content": "Canonical Shape Projection for 3D FSCIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": "point cloud sample " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": ", the transformation module " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "T_r" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": " computes the desired rotation matrix " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": " while taking into account all the aforementioned constraints to ensure its validity. This process results in the generation of a learned depth map image, " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{D_i} = F(\\mathcal{X}_i,\\mathbf{R}_i)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 182, + 275, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 182, + 275, + 196 + ], + "spans": [ + { + "bbox": [ + 131, + 182, + 275, + 196 + ], + "type": "text", + "content": "3.3 Model reprogramming" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 204, + 482, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 204, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 482, + 396 + ], + "type": "text", + "content": "Model reprogramming of computer vision models, especially foundational ones, is a crucial step in their effective deployment, enabling a seamless transition to new tasks and domains with minimal adjustments to existing architectures [38]. We present a novel model reprogramming approach specifically tailored for the CLIP model applied to the FSCIL task on 3D point cloud objects. It is worth noting that the CLIP model is primarily trained on RGB images rather than depth maps. Therefore, to enhance the CLIP model's effectiveness with depth maps, we incorporate pixel information into the background of the depth map images. More specifically, this modification assists the CLIP model in accurately discerning the boundary of the depth map shape information through a " + }, + { + "bbox": [ + 130, + 204, + 482, + 396 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 204, + 482, + 396 + ], + "type": "text", + "content": " module, which takes the form of a U-Net [32]. To accomplish this, we initially compute the mask map " + }, + { + "bbox": [ + 130, + 204, + 482, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{M_i}" + }, + { + "bbox": [ + 130, + 204, + 482, + 396 + ], + "type": "text", + "content": " corresponding to a given depth map " + }, + { + "bbox": [ + 130, + 204, + 482, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{D_i}" + }, + { + "bbox": [ + 130, + 204, + 482, + 396 + ], + "type": "text", + "content": ", which isolates the shape boundary from the background. Next, we input " + }, + { + "bbox": [ + 130, + 204, + 482, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{M_i}" + }, + { + "bbox": [ + 130, + 204, + 482, + 396 + ], + "type": "text", + "content": " into the " + }, + { + "bbox": [ + 130, + 204, + 482, + 396 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 204, + 482, + 396 + ], + "type": "text", + "content": " module to extract the pixel information for the background, denoted as " + }, + { + "bbox": [ + 130, + 204, + 482, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{H_i}" + }, + { + "bbox": [ + 130, + 204, + 482, + 396 + ], + "type": "text", + "content": ". To ensure the consistency of the generated pixels in " + }, + { + "bbox": [ + 130, + 204, + 482, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{M_i}" + }, + { + "bbox": [ + 130, + 204, + 482, + 396 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 130, + 204, + 482, + 396 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 204, + 482, + 396 + ], + "type": "text", + "content": " module and to prevent convergence to arbitrary values, we employ the following loss function," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 197, + 404, + 481, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 404, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 197, + 404, + 481, + 435 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {g} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(\\frac {\\partial \\mathbf {I} _ {M _ {i}}}{\\partial x} - \\frac {\\partial \\mathbf {I} _ {H _ {i}}}{\\partial x}\\right) ^ {2} + \\left(\\frac {\\partial \\mathbf {I} _ {M _ {i}}}{\\partial y} - \\frac {\\partial \\mathbf {I} _ {H _ {i}}}{\\partial y}\\right) ^ {2} \\tag {2}", + "image_path": "521acd5f91dd00e1e5cd5d430edede7ac6d9077315d114aa82557ae6f4bd2d9a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 445, + 482, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 481 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 445, + 482, + 481 + ], + "type": "inline_equation", + "content": "\\frac{\\partial}{\\partial x}" + }, + { + "bbox": [ + 130, + 445, + 482, + 481 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 445, + 482, + 481 + ], + "type": "inline_equation", + "content": "\\frac{\\partial}{\\partial y}" + }, + { + "bbox": [ + 130, + 445, + 482, + 481 + ], + "type": "text", + "content": " denote the derivatives of the images in the " + }, + { + "bbox": [ + 130, + 445, + 482, + 481 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 445, + 482, + 481 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 445, + 482, + 481 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 445, + 482, + 481 + ], + "type": "text", + "content": " directions, respectively. This loss enforces a smoothness constraint on the gradient to separate it from foreground objects explicitly." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 482, + 482, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 482, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 482, + 531 + ], + "type": "text", + "content": "It helps to learn novel classes with few examples. Subsequently, we combine the depth map " + }, + { + "bbox": [ + 130, + 482, + 482, + 531 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{D_i}" + }, + { + "bbox": [ + 130, + 482, + 482, + 531 + ], + "type": "text", + "content": " from the projection module " + }, + { + "bbox": [ + 130, + 482, + 482, + 531 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 482, + 482, + 531 + ], + "type": "text", + "content": " with the extracted values " + }, + { + "bbox": [ + 130, + 482, + 482, + 531 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{H_i}" + }, + { + "bbox": [ + 130, + 482, + 482, + 531 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 482, + 482, + 531 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 482, + 482, + 531 + ], + "type": "text", + "content": " to generate the final image, " + }, + { + "bbox": [ + 130, + 482, + 482, + 531 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{R_i} = \\mathbf{I}_{D_i} \\odot C(M(\\mathbf{I}_{D_i}))" + }, + { + "bbox": [ + 130, + 482, + 482, + 531 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 482, + 482, + 531 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 482, + 482, + 531 + ], + "type": "text", + "content": " represents the element-wise multiplication." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 548, + 262, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 262, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 262, + 562 + ], + "type": "text", + "content": "3.4 Prompt engineering" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Advancements in prompt engineering have unveiled significant potential, primarily driven by LLMs [18, 19]. LLMs can extract richer information from appropriately tailored prompts. With this motivation, we employ GPT to obtain more insights from plain class names and enhance the FSCIL task on 3D data. In contrast to prevailing methods [9, 10, 12] that rely solely on the class name for the FSCIL task, we extend our method by generating additional descriptions to extract more semantic information from the CLIP text encoder (denoted as " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "T_{e}" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "). This augmentation ensures a more comprehensive utilization of the LLM's" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "A. Cheraghian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 212 + ], + "type": "text", + "content": "capabilities, thereby enhancing the overall efficacy of our approach. In the FS-CIL, a significant challenge arises due to the scarcity of training samples for novel classes, potentially leading to overfitting. To address this issue, we propose an innovative strategy leveraging the capabilities of LLM, specifically the GPT model, and utilizing this as novel prompts for the text encoder of the CLIP model. More precisely, we utilize the GPT model to generate multiple descriptions for each class by providing it with a specific command. These prompts are formally defined for each sample as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 255, + 219, + 480, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 219, + 480, + 232 + ], + "spans": [ + { + "bbox": [ + 255, + 219, + 480, + 232 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {i} = \\operatorname {G P T} (\\text {C o m m a n d s}). \\tag {3}", + "image_path": "62cfde791f2b7e7f90787777ffdca2a269a97a1588732cb8e262501bfc586ba3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 480, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 480, + 319 + ], + "type": "text", + "content": "Here, the Commands are structured as Describe a depth of a [class]. For every class in the dataset, we substitute the respective class names in place of '[class]' in the command and input them into GPT-4. This results in the generation of depth map-specific descriptions with comprehensive category-wise semantics. Subsequently, we feed the descriptions of each category through CLIP's textual encoder " + }, + { + "bbox": [ + 130, + 236, + 480, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_e" + }, + { + "bbox": [ + 130, + 236, + 480, + 319 + ], + "type": "text", + "content": ". Finally, the generated prompts " + }, + { + "bbox": [ + 130, + 236, + 480, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 130, + 236, + 480, + 319 + ], + "type": "text", + "content": " are used in the training of the few-shot classes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 335, + 294, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 335, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 131, + 335, + 294, + 346 + ], + "type": "text", + "content": "3.5 Class incremental learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": "Training. For the base task, denoted as " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}^1" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": ", we conduct training for the transformation model " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "T_{r}" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": ", the 3D backbone " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": ", and the relation module " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "R_{l}" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": ". In the case of subsequent tasks, denoted as " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}^t" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "t > 1" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": ", only the relation module undergoes fine-tuning. Throughout both base and subsequent novel tasks, all other components in our proposed architecture (refer to Figure 2) remain frozen. Additionally, for each previously learned class, we randomly select a sample to be stored in a compact memory module " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": ". Given the " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": " point cloud " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i^t" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "t^{th}" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": " task, the " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": " module generates the features " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^t" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "V_{e}" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": " module produces " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i^t" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": ". Simultaneously, the prompts of all classes for tasks " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}^1,\\ldots ,\\mathcal{Q}^t" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": " are fed into the text encoder " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "T_{e}" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": " to generate the corresponding text embeddings denoted as " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{e}_j\\}_{j = 1}^{|\\mathcal{Y}_{tl}|}" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": ". Then, we concatenate " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^t" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i^t" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_j^t" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": " and forward them into the relation module " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "R_{l}" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": ", which provides a score between [0,1], representing the similarity between them. In other words, for each training sample, we generate a score against each of the classes in both the subsequent and previous tasks as " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "r_{ij}^{t} = \\gamma \\circ R\\circ (\\mathbf{z}_{i}^{t}\\oplus \\mathbf{v}_{i}\\oplus \\mathbf{e}_{j}), j\\in \\mathcal{Y}_{tl}" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_{tl} = \\bigcup_{i = 1}^{t}\\mathcal{Y}^{i}" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": " is the concatenation operator, " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": " is the relation module, and " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": " is the sigmoid function. Finally, for each feature " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "r_{ij}" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": " and the corresponding ground truth " + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i" + }, + { + "bbox": [ + 130, + 351, + 482, + 570 + ], + "type": "text", + "content": ", we employ a binary cross-entropy cost function to train the model as follows," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 585, + 486, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 585, + 486, + 616 + ], + "spans": [ + { + "bbox": [ + 133, + 585, + 486, + 616 + ], + "type": "interline_equation", + "content": "L _ {r} = - \\frac {1}{| \\mathcal {Y} _ {t l} | | S |} \\sum_ {k \\in \\mathcal {Y} _ {t l}} \\sum_ {\\mathbf {y} _ {i} \\in \\mathcal {S}} \\Bigg (1 (\\boldsymbol {y} _ {i} ^ {t} = = k) \\log (r _ {i k}) + \\big (1 - 1 (\\boldsymbol {y} _ {i} ^ {t} = = k) \\big) \\log (1 - r _ {i k}) \\Bigg),", + "image_path": "10df96ce2fcf24ccf90696f2b8ea40550c684eabd2316b468b93f75e9b63f643.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 620, + 480, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 620, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 131, + 620, + 480, + 643 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 620, + 480, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 131, + 620, + 480, + 643 + ], + "type": "text", + "content": " denotes the set of true labels in the current task, along with the memory " + }, + { + "bbox": [ + 131, + 620, + 480, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 131, + 620, + 480, + 643 + ], + "type": "text", + "content": ". Finally, the total loss to train the entire pipeline is as follows," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 250, + 651, + 480, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 651, + 480, + 663 + ], + "spans": [ + { + "bbox": [ + 250, + 651, + 480, + 663 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t} = \\lambda_ {1} \\mathcal {L} _ {c} + \\lambda_ {2} \\mathcal {L} _ {g} + \\lambda_ {3} \\mathcal {L} _ {r} \\tag {4}", + "image_path": "5234bde2901b2b1612ead14174844958b9acc343281238cfdb05f6fa46d6118b.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 448, + 102 + ], + "type": "text", + "content": "Canonical Shape Projection for 3D FSCIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\lambda_3" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " are the weighting factors for the constraint loss " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}c)" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": ", gradient loss " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}g)" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": ", and the relation loss " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_r)" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 164 + ], + "type": "text", + "content": "Inference. During the inference process, when given an unlabeled sample " + }, + { + "bbox": [ + 130, + 140, + 480, + 164 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^c" + }, + { + "bbox": [ + 130, + 140, + 480, + 164 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 140, + 480, + 164 + ], + "type": "inline_equation", + "content": "c \\in \\mathcal{Y}_{tl}" + }, + { + "bbox": [ + 130, + 140, + 480, + 164 + ], + "type": "text", + "content": ", the label prediction " + }, + { + "bbox": [ + 130, + 140, + 480, + 164 + ], + "type": "inline_equation", + "content": "y^{*}" + }, + { + "bbox": [ + 130, + 140, + 480, + 164 + ], + "type": "text", + "content": " is computed using the following equation:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 233, + 175, + 481, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 175, + 481, + 196 + ], + "spans": [ + { + "bbox": [ + 233, + 175, + 481, + 196 + ], + "type": "interline_equation", + "content": "y ^ {*} = \\underset {j \\in \\mathcal {Y} _ {t l}} {\\arg \\max } (R \\circ (\\mathbf {z} _ {c} \\oplus \\mathbf {v} _ {c} \\oplus \\mathbf {s} _ {j})), \\tag {5}", + "image_path": "8feea8f0d9dd33d1c2f5570a7f5a169bcc39dc169659967ab1c4279f23283810.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 204, + 479, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 204, + 479, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 479, + 228 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 204, + 479, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_c" + }, + { + "bbox": [ + 130, + 204, + 479, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 204, + 479, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_c" + }, + { + "bbox": [ + 130, + 204, + 479, + 228 + ], + "type": "text", + "content": " represent the calculated features for the sample " + }, + { + "bbox": [ + 130, + 204, + 479, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^c" + }, + { + "bbox": [ + 130, + 204, + 479, + 228 + ], + "type": "text", + "content": ". We use a standardized prompt for all classes, specifically This is a depth map of a [class]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 246, + 230, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 246, + 230, + 260 + ], + "spans": [ + { + "bbox": [ + 131, + 246, + 230, + 260 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 270, + 480, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 480, + 389 + ], + "type": "text", + "content": "Datasets. We employed two synthetic 3D datasets, Modelnet40 [45] and Shape-net [6], in addition to two real scanned 3D datasets, ScanobjectNN [40] and CO3D [31]. We follow the setup proposed by [11] for Few-Shot Classification with Incremental Learning (FSCIL) on point cloud objects. Our proposed method is assessed in two distinct scenarios. Firstly, we evaluate its performance on the FSCIL task within the same dataset, where both the base and novel tasks originate from identical datasets. Secondly, we subject our method to more demanding and realistic challenges by testing it in cross-dataset scenarios. In this setting, the base task involves synthetic datasets, while the novel tasks comprise classes from real scanned few-shot datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 390, + 480, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 480, + 438 + ], + "type": "text", + "content": "Prompt generation. GPT-4 was employed to produce diverse class descriptions for each class. The following command was utilized: Describe a depth of a [class]. Approximately 100 class descriptions were generated for each class and utilized during the training phase." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "text", + "content": "Implementation details. In our experimentation, we utilize Curvenet as the backbone for 3D point clouds " + }, + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "inline_equation", + "content": "(B)" + }, + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "text", + "content": " [46]. The CLIP vision " + }, + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "inline_equation", + "content": "(V_{e})" + }, + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "text", + "content": " and text encoder " + }, + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "inline_equation", + "content": "(T_{e})" + }, + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "text", + "content": " are instantiated using the ViT-B-16 architecture [14]. The " + }, + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "text", + "content": " model is constructed based on the U-Net architecture [32]. The projection module " + }, + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "inline_equation", + "content": "(F)" + }, + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "text", + "content": " follows the methodology proposed by Zhu et al. [51], and for the transformation module " + }, + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "inline_equation", + "content": "(T_{r})" + }, + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "text", + "content": ", we employ the STN model [16]. Additionally, the relation module comprises several fully connected layers with sizes (1024, 2048, 1). We use the farthest 1024 points from 3D point cloud objects as input for all samples. For the base task, we train our proposed method (Figure 2) in an end-to-end manner for 250 epochs using the Adam optimizer with a learning rate of 0.001. For the novel task, we fine-tune our model for 200 epochs using the Adam optimizer with a learning rate of 0.001. It is important to note that we randomly shift and scale points in the input point cloud during both base and novel class training, with the additional step of randomly dropping out points. The feature vector dimensions for " + }, + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "inline_equation", + "content": "T_{e}" + }, + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "inline_equation", + "content": "V_{e}" + }, + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "text", + "content": " are set to 512. Additionally, the weights of the loss functions " + }, + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "inline_equation", + "content": "\\lambda_{3}" + }, + { + "bbox": [ + 130, + 438, + 481, + 641 + ], + "type": "text", + "content": " are configured as 0.01, 1, and 1 across the entire datasets. Our experiments are conducted using the PyTorch framework." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Evaluation metrics. Following the training of each task, we compute the accuracy by consolidating both base and novel classes. Subsequently, following the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "A. Cheraghian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "type": "text", + "content": "recommendation in [36], we determine the relative accuracy dropping rate, denoted as " + }, + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "type": "inline_equation", + "content": "\\varDelta" + }, + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "type": "text", + "content": ", using the formula " + }, + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "type": "inline_equation", + "content": "\\varDelta=\\frac{|acc_T-acc_0|}{acc_0}\\times 100" + }, + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "type": "inline_equation", + "content": "acc_{T}" + }, + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "type": "inline_equation", + "content": "acc_{0}" + }, + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "type": "text", + "content": " denote the accuracy of the last and first tasks, respectively. The value of " + }, + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "type": "inline_equation", + "content": "\\varDelta" + }, + { + "bbox": [ + 130, + 116, + 482, + 178 + ], + "type": "text", + "content": " serves as a comprehensive metric for method evaluation. A lower relative accuracy signifies superior performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 205, + 224, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 224, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 224, + 216 + ], + "type": "text", + "content": "4.1 Main results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 233, + 479, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 233, + 479, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 479, + 293 + ], + "type": "text", + "content": "We conducted a comparative analysis of our approach against established methods in the field, including IL2M [2], ScaIL [3], EEIL [5], LwF [22], FACT [50], Sem-aware [9], and Microshape [12]. Notably, all these methods are tailored for image data, with the exception of Microshape [12], which was originally designed for point cloud data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 294, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 402 + ], + "type": "text", + "content": "Within dataset: In this scenario, both the base and novel tasks are derived from the same dataset. Roughly half of the classes in the dataset are assigned to the base task, while the remaining half are incrementally introduced as multiple novel tasks. It is presumed that these novel tasks encompass only a limited number of samples for training. Our proposed method has been evaluated on three datasets—ModelNet, CO3D, and Shapenet. As depicted in Table 1, our approach demonstrates superior performance across all methods in terms of the relative accuracy dropping rate, denoted as " + }, + { + "bbox": [ + 130, + 294, + 482, + 402 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 294, + 482, + 402 + ], + "type": "text", + "content": ", indicating the effectiveness of our proposed method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 404, + 482, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 482, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 482, + 523 + ], + "type": "text", + "content": "Cross dataset: In this arrangement, the base task centers around synthetic point cloud objects, while the novel few-shot tasks incorporate real-scanned 3D objects. Notably, the latter often exhibit incompleteness or noise in contrast to the synthetic data, which typically boast cleanliness and completeness. Consequently, this configuration introduces a more realistic and challenging scenario compared to within-dataset experiments. Three datasets are considered in this setting: Shapenet " + }, + { + "bbox": [ + 130, + 404, + 482, + 523 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 404, + 482, + 523 + ], + "type": "text", + "content": " CO3D, ModelNet40 " + }, + { + "bbox": [ + 130, + 404, + 482, + 523 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 404, + 482, + 523 + ], + "type": "text", + "content": " ScanObjectNN, and Shapenet " + }, + { + "bbox": [ + 130, + 404, + 482, + 523 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 404, + 482, + 523 + ], + "type": "text", + "content": " ScanObjectNN. As illustrated in Table 2, our method outperforms other approaches in all three setups, underscoring its efficacy in scenarios involving a domain gap between base and novel tasks." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 135, + 561, + 487, + 650 + ], + "blocks": [ + { + "bbox": [ + 165, + 548, + 447, + 559 + ], + "lines": [ + { + "bbox": [ + 165, + 548, + 447, + 559 + ], + "spans": [ + { + "bbox": [ + 165, + 548, + 447, + 559 + ], + "type": "text", + "content": "Table 1: Summary of FSCIL results for within-dataset experiments." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 561, + 487, + 650 + ], + "lines": [ + { + "bbox": [ + 135, + 561, + 487, + 650 + ], + "spans": [ + { + "bbox": [ + 135, + 561, + 487, + 650 + ], + "type": "table", + "html": "
ModelNetCO3DShapeNet
Method2025303540Δ↓253035404550Δ↓25303540455055 Δ↓
FT89.89.74.33.33.096.776.711.23.63.21.80.899.087.025.76.81.30.90.60.4 99.5
Joint89.888.287.083.580.510.476.769.464.862.760.759.822.087.085.284.383.082.582.281.3 6.6
LwF [22]89.836.09.13.63.196.076.714.74.73.52.31.098.787.060.833.515.93.83.11.8 97.9
IL2M [2]89.865.558.452.353.640.376.731.527.718.127.121.971.487.058.645.740.750.149.449.3 43.3
ScaIL [3]89.866.864.558.756.537.176.739.534.124.130.127.564.187.056.651.844.350.346.445.4 47.8
EEIL [5]89.875.467.260.155.638.176.761.452.442.839.532.857.287.077.773.269.366.465.965.8 22.4
FACT [50]90.481.377.173.565.028.177.967.159.754.850.246.740.087.575.371.469.967.565.762.5 28.6
Sem-aware [9]91.382.274.370.064.729.176.866.959.253.649.142.944.187.274.968.969.168.066.963.8 26.8
Microshape [11]93.683.178.275.867.128.378.567.360.156.151.447.239.987.683.281.579.076.873.5 72.617.1
C3PR (ours)91.682.375.872.270.922.581.569.466.563.054.253.834.088.081.677.876.776.776.976.2715.1
", + "image_path": "4f1ece16acc37c7f000773b3db179f8c3f102890bb4036082a8d5a6efaf5e045.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "text", + "content": "Canonical Shape Projection for 3D FSCIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 127, + 492, + 217 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 125 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 125 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 125 + ], + "type": "text", + "content": "Table 2: Summary of FSCIL results for within-dataset and cross-dataset experiments." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 127, + 492, + 217 + ], + "lines": [ + { + "bbox": [ + 134, + 127, + 492, + 217 + ], + "spans": [ + { + "bbox": [ + 134, + 127, + 492, + 217 + ], + "type": "table", + "html": "
ShapeNet → CO3DModelNet → ScanObjectNNShapeNet → ScanObjectNN
Method3944495459646974798489Δ↓26303437Δ↓44495459Δ↓
FT81.020.22.31.70.81.01.01.30.90.51.698.088.46.46.01.997.981.438.74.00.998.9
Joint81.079.578.375.275.174.872.371.370.068.867.316.988.479.774.071.219.581.482.579.878.73.3
LwF [22]81.057.419.32.31.00.90.81.31.10.81.997.788.435.85.82.597.281.447.914.05.992.8
IL2M [2]81.045.636.835.131.833.334.031.530.632.330.063.088.458.252.952.041.281.453.243.945.843.7
ScaIL [3]81.050.145.739.139.037.938.036.033.733.035.256.588.456.555.952.940.281.449.046.740.050.9
EEIL [5]81.075.269.363.260.557.953.051.951.347.847.641.288.470.261.056.835.781.474.569.863.422.1
FACT [50]81.476.070.368.165.863.563.060.158.257.555.931.389.172.568.363.528.782.374.669.966.818.8
Sem-aware [9]80.669.566.562.963.263.061.258.358.157.255.231.688.573.967.764.227.581.370.665.262.922.6
Microshape [11]82.677.973.972.767.766.265.463.460.658.157.130.989.373.268.465.127.182.574.871.267.118.7
C3PR (ours)83.680.077.875.472.872.370.367.964.964.163.224.488.375.770.667.823.284.577.875.571.914.9
", + "image_path": "bdfcf39ada3895884b55569cfdf314cf1999d3872fcaeff6c0e4fa37d3548dde.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 232, + 246, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 232, + 246, + 243 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 246, + 243 + ], + "type": "text", + "content": "4.2 Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 258, + 481, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 481, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 481, + 330 + ], + "type": "text", + "content": "Evaluating model reprogramming: In the conducted experiment (refer to Figure 4 (a)), we systematically omitted the model reprogramming module " + }, + { + "bbox": [ + 130, + 258, + 481, + 330 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 258, + 481, + 330 + ], + "type": "text", + "content": " and directly fed the depth map image " + }, + { + "bbox": [ + 130, + 258, + 481, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_D" + }, + { + "bbox": [ + 130, + 258, + 481, + 330 + ], + "type": "text", + "content": " into the vision encoder " + }, + { + "bbox": [ + 130, + 258, + 481, + 330 + ], + "type": "inline_equation", + "content": "V_{e}" + }, + { + "bbox": [ + 130, + 258, + 481, + 330 + ], + "type": "text", + "content": " of the CLIP model. The results clearly illustrate the significance of the model reprogramming in our proposed approach, as evidenced by notable differences between the outcomes with and without this module." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 332, + 481, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 481, + 392 + ], + "type": "text", + "content": "Impact of the prompt engineering: In our ablation study (see Figure 4 (b)), we investigate the influence of prompt engineering in our methodology. Specifically, we streamline the training stage by substituting the original multiple prompts per class with a single fixed prompt. The outcomes underscore the effectiveness of our GPT-based prompt engineering approach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 394, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 482, + 515 + ], + "type": "text", + "content": "Impact of canonical shape projection: Here, we evaluate our proposed canonical shape projection methodology in comparison to PointCLIPv2 [51], which generates multiple fixed-depth map projections for each input point cloud. Specifically, we eliminate the transformation module " + }, + { + "bbox": [ + 130, + 394, + 482, + 515 + ], + "type": "inline_equation", + "content": "T_r" + }, + { + "bbox": [ + 130, + 394, + 482, + 515 + ], + "type": "text", + "content": ". Subsequently, we adopt the depth map generation method employed in PointCLIP v2, involving the use of ten fixed camera pose angles for computing ten depth map images. These images are then fed into the vision encoder " + }, + { + "bbox": [ + 130, + 394, + 482, + 515 + ], + "type": "inline_equation", + "content": "V_e" + }, + { + "bbox": [ + 130, + 394, + 482, + 515 + ], + "type": "text", + "content": " of the CLIP model, and the relation module is applied to merge the feature vector, mirroring our proposed pipeline. Notably, our method surpasses PointCLIPv2 by a significant margin (see Figure 4 (c)), highlighting the efficacy of our learned camera pose angle." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 133, + 538, + 250, + 609 + ], + "blocks": [ + { + "bbox": [ + 133, + 538, + 250, + 609 + ], + "lines": [ + { + "bbox": [ + 133, + 538, + 250, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 538, + 250, + 609 + ], + "type": "image", + "image_path": "f2735cff78476fd9eec68ebb24ea1f7d2fb4381219dff78ddc26c7952aea34be.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 617, + 482, + 662 + ], + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 662 + ], + "type": "text", + "content": "Fig. 4: This experiment is performed within the Shapenet to ScanobjectNN setting. In (a), we present the influence of our novel CLIP model reprogramming strategy. (b) Highlights the impact of prompt engineering as proposed in our method. (c) Our approach is compared with PointCLIP v2 [51]." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 250, + 538, + 365, + 609 + ], + "blocks": [ + { + "bbox": [ + 250, + 538, + 365, + 609 + ], + "lines": [ + { + "bbox": [ + 250, + 538, + 365, + 609 + ], + "spans": [ + { + "bbox": [ + 250, + 538, + 365, + 609 + ], + "type": "image", + "image_path": "891842fb6002b36c8659e7ce176eb5e3ad87715498c7c2ae39b4daba34465bf7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 365, + 539, + 479, + 608 + ], + "blocks": [ + { + "bbox": [ + 365, + 539, + 479, + 608 + ], + "lines": [ + { + "bbox": [ + 365, + 539, + 479, + 608 + ], + "spans": [ + { + "bbox": [ + 365, + 539, + 479, + 608 + ], + "type": "image", + "image_path": "42a582476af33b9e1cbc6568663637754912fe8e31dcff1b891bf7ef7170887d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "A. Cheraghian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 186, + 114, + 302, + 205 + ], + "blocks": [ + { + "bbox": [ + 186, + 114, + 302, + 205 + ], + "lines": [ + { + "bbox": [ + 186, + 114, + 302, + 205 + ], + "spans": [ + { + "bbox": [ + 186, + 114, + 302, + 205 + ], + "type": "image", + "image_path": "eb73fc7a41ed88ab3af361dd4727f27eae7410e8e44844d1a11f8deb8b78f207.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 214, + 480, + 248 + ], + "lines": [ + { + "bbox": [ + 131, + 214, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 131, + 214, + 480, + 248 + ], + "type": "text", + "content": "Fig. 5: We visualize the quality of features (from 44 base classes of ShapeNet) generated by the CLIP vision encoder using (Right) our proposed method, C3PR, and (Left) PointCLIPv2 [51]. Our method produced a denser cluster than PointCLIP." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 428, + 205 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 428, + 205 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 428, + 205 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 428, + 205 + ], + "type": "image", + "image_path": "bcf4f8654d88c7b467beafb144fcc896a12b208d66957f5687610ce8c46ea02e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 258, + 479, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 479, + 306 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 479, + 306 + ], + "type": "text", + "content": "Feature quality: In Figure 5, we assess the feature quality produced by the CLIP vision encoder using our proposed method, C3PR, and compare it with PointCLIPv2 [51]. As illustrated, our approach yields more discriminative clusters, enhancing overall generalization capabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 307, + 480, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 480, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 480, + 366 + ], + "type": "text", + "content": "Influence of " + }, + { + "bbox": [ + 130, + 307, + 480, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 130, + 307, + 480, + 366 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 307, + 480, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_g" + }, + { + "bbox": [ + 130, + 307, + 480, + 366 + ], + "type": "text", + "content": ": In Table 3a, we highlight the impact of both " + }, + { + "bbox": [ + 130, + 307, + 480, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 130, + 307, + 480, + 366 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 307, + 480, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_g" + }, + { + "bbox": [ + 130, + 307, + 480, + 366 + ], + "type": "text", + "content": " within our proposed pipeline. It is evident that both loss functions contribute significantly, but " + }, + { + "bbox": [ + 130, + 307, + 480, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 130, + 307, + 480, + 366 + ], + "type": "text", + "content": " holds greater importance, underscoring the significance of identifying the learned projection angle for depth map images on image foundation models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 368, + 480, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 480, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 480, + 416 + ], + "type": "text", + "content": "Model reprogramming: Conventional vs. our: In this section, we contrast our innovative model reprogramming approach with the conventional method introduced by [15]. Table 3b illustrates the superior effectiveness of our proposed model reprogramming strategy." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 133, + 458, + 301, + 529 + ], + "blocks": [ + { + "bbox": [ + 135, + 423, + 476, + 436 + ], + "lines": [ + { + "bbox": [ + 135, + 423, + 476, + 436 + ], + "spans": [ + { + "bbox": [ + 135, + 423, + 476, + 436 + ], + "type": "text", + "content": "Table 3: Ablation study on the influence of loss (a) and model reprogramming (b)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 439, + 294, + 450 + ], + "lines": [ + { + "bbox": [ + 138, + 439, + 294, + 450 + ], + "spans": [ + { + "bbox": [ + 138, + 439, + 294, + 450 + ], + "type": "text", + "content": "(a) Influence of " + }, + { + "bbox": [ + 138, + 439, + 294, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 138, + 439, + 294, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 138, + 439, + 294, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_g" + }, + { + "bbox": [ + 138, + 439, + 294, + 450 + ], + "type": "text", + "content": " in our approach." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 458, + 301, + 529 + ], + "lines": [ + { + "bbox": [ + 133, + 458, + 301, + 529 + ], + "spans": [ + { + "bbox": [ + 133, + 458, + 301, + 529 + ], + "type": "table", + "html": "
ShapeNet → ScanObjectNN
Method44495459Δ↓
Lr84.476.270.16522.9
Lr + Lg84.377.370.866.920.6
Lr + Lc85.780.274.570.817.3
Lr + Lg + Lc84.577.875.571.914.9
", + "image_path": "cfe8fd63775a6853ed7dad5d688ac528b8d2e70ea28cb111bad9d7ff5204c5ad.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 310, + 475, + 476, + 522 + ], + "blocks": [ + { + "bbox": [ + 304, + 440, + 480, + 466 + ], + "lines": [ + { + "bbox": [ + 304, + 440, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 480, + 466 + ], + "type": "text", + "content": "(b) Comparing our proposed model reprogramming strategy, C3PR, with the conventional model reprogramming technique employed in [15]." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 475, + 476, + 522 + ], + "lines": [ + { + "bbox": [ + 310, + 475, + 476, + 522 + ], + "spans": [ + { + "bbox": [ + 310, + 475, + 476, + 522 + ], + "type": "table", + "html": "
ShapeNet → ScanObjectNN
Method44495459Δ↓
[15]85.780.274.570.817.3
C3PR (ours)84.577.875.571.914.9
", + "image_path": "5a1476d1d41162eb787e47e8ccf172f793ad09d4ebcbb2865f18b3f88cf3a116.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 545, + 212, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 212, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 212, + 555 + ], + "type": "text", + "content": "4.3 Discussion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 568, + 479, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 479, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 479, + 628 + ], + "type": "text", + "content": "Point-cloud classification: Although our learning projection module " + }, + { + "bbox": [ + 130, + 568, + 479, + 628 + ], + "type": "inline_equation", + "content": "Tr" + }, + { + "bbox": [ + 130, + 568, + 479, + 628 + ], + "type": "text", + "content": " and model reprogramming module " + }, + { + "bbox": [ + 130, + 568, + 479, + 628 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 568, + 479, + 628 + ], + "type": "text", + "content": " have demonstrated effectiveness in the FSCIL context, they can also be applied in the point-cloud classification setting (shown in the table 4a. However, it is worth to note that our proposed prompt engineering strategy is specifically tailored to improve performance in FSCIL problems." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 628, + 210, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 628, + 210, + 639 + ], + "spans": [ + { + "bbox": [ + 132, + 628, + 210, + 639 + ], + "type": "text", + "content": "Module " + }, + { + "bbox": [ + 132, + 628, + 210, + 639 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 132, + 628, + 210, + 639 + ], + "type": "text", + "content": " role:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "The role of Module " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": " is twofold: it learns background information for the depth maps and smoothens the foreground. Specifically, we combine the depth" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 448, + 102 + ], + "type": "text", + "content": "Canonical Shape Projection for 3D FSCIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "map " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{D_i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": " from the projection module " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": " with the extracted values " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{T_i}" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": " from Module " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": " to generate the final image, " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{R_i} = \\mathbf{I}_{D_i} \\odot C(M(\\mathbf{I}_{D_i}))" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": " represents element-wise multiplication. To this end, we conducted experiments to demonstrate the impact of Module " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": " in our proposed method (see table 4b)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 186, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 186, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 186, + 294, + 204 + ], + "type": "text", + "content": "(a) The performance of our proposed method on point-cloud classification in ModelNet40." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 307, + 186, + 483, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 483, + 204 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 483, + 204 + ], + "type": "text", + "content": "(b) The impact of module " + }, + { + "bbox": [ + 307, + 186, + 483, + 204 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 307, + 186, + 483, + 204 + ], + "type": "text", + "content": " in our proposed method." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 168, + 213, + 270, + 262 + ], + "blocks": [ + { + "bbox": [ + 141, + 170, + 472, + 182 + ], + "lines": [ + { + "bbox": [ + 141, + 170, + 472, + 182 + ], + "spans": [ + { + "bbox": [ + 141, + 170, + 472, + 182 + ], + "type": "text", + "content": "Table 4: Ablation on point-cloud classification (a) and impact of " + }, + { + "bbox": [ + 141, + 170, + 472, + 182 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 141, + 170, + 472, + 182 + ], + "type": "text", + "content": " module (b)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 168, + 213, + 270, + 262 + ], + "lines": [ + { + "bbox": [ + 168, + 213, + 270, + 262 + ], + "spans": [ + { + "bbox": [ + 168, + 213, + 270, + 262 + ], + "type": "table", + "html": "
MethodAccuracy
DGCNN [43]92.9
CurveNet [46]93.8
Ours94.6
", + "image_path": "5b199210ac6a1651149622468d952902599504d0fda9044d6a33669865d56f1f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 213, + 466, + 262 + ], + "blocks": [ + { + "bbox": [ + 307, + 213, + 466, + 262 + ], + "lines": [ + { + "bbox": [ + 307, + 213, + 466, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 213, + 466, + 262 + ], + "type": "table", + "html": "
ShapeNet → ScanObjectNN
Method44495459Δ↓
Without C84.576.373.268.918.4
With C84.577.875.571.914.9
", + "image_path": "1c354097e1da4f62214c19a73ba56d57c186d0cf521338a3a1505f62b7655d5d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 268, + 482, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 482, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 482, + 364 + ], + "type": "text", + "content": "**Limitation:** The proposed method exhibits several limitations that warrant discussion. Firstly, its reliance on storing exemplars in memory, with one sample per class, and addressing the forgetting issue raises the question of exploring exemplar-free alternatives, which could offer intriguing avenues for further research. Secondly, our method would perform worse on new concepts from significantly different domains, synthetic to real-scan datasets, as can be seen in Table 2. However, it works well when the new concepts come from similar domains, synthetic to synthetic datasets, as can be seen in Table 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 220, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 220, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 220, + 392 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 402, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 641 + ], + "type": "text", + "content": "In this paper, we leverage the CLIP architecture of image-to-text matching to address the FSCIL problem on 3D point cloud objects. Traditional approaches usually project a 3D object on a 2D plane to produce images from different camera angles. Among those 2D images, some have bad/poor camera angles, making them difficult to recognize. Therefore, we propose a learning mechanism to find one camera projection that best describes the 3D objects. In doing so, we adopt a model reprogramming approach that perturbs the input 2D images (from the best projection angle) with background-foreground colorization (segmentation) without modifying or fine-tuning the original CLIP architectures. Our reprogramming approach is different from conventional approaches because our custom loss produces a uniform background color instead of the noisy background of previous works. In the text pipeline, we use the CLIP text encoder and train the whole architecture end-to-end. Instead of using a single class name-based text input, we suggest using prompts based on GPT-based description. We propose a prompt engineering approach to train the networks with 3D object and text description pairs. Finally, we evaluate our proposed method using 3D synthetic datasets, ModelNet and ShapeNet, and two 3D real-world-scanned datasets, ScanObjectNN and CO3D, and report state-of-the-art performances. In future, we will investigate model reprogramming for other vision problems beyond FSCIL, like segmentation and domain adaptation on 3D point clouds." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Acknowledgment. Mehrtash Harandi thanks the Australian Research Council for support through the Discovery Program (DP230101176)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "A. Cheraghian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 137, + 481, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 481, + 202 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 481, + 202 + ], + "type": "text", + "content": "1. Bansal, N., Chen, X., Wang, Z.: Can we gain more from orthogonality regularizations in training deep networks? In: Bengio, S., Wallach, H., Larochelle, H., Grauman, K., Cesa-Bianchi, N., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 31. Curran Associates, Inc. (2018), https://proceedings.neurips.cc/paper_files/paper/2018/file/bf424cb7b0dea050a42b9739eb261a3a-Paper.pdf" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 203, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 203, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 203, + 481, + 224 + ], + "type": "text", + "content": "2. Belouadah, E., Popescu, A.: II2m: Class incremental learning with dual memory. In: CVPR (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 224, + 481, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 481, + 245 + ], + "type": "text", + "content": "3. Belouadah, E., Popescu, A.: Scail: Classifier weights scaling for class incremental learning. In: WACV (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 245, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 481, + 343 + ], + "type": "text", + "content": "4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 1877-1901. Curran Associates, Inc. (2020), https://proceedings.neurips.cc/paper_files/paper/2020/file/1457c0d6bcbd4967418bf8ac142f64a-Paper.pdf" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 343, + 481, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 481, + 364 + ], + "type": "text", + "content": "5. Castro, F.M., Marín-Jiménez, M.J., Guil, N., Schmid, C., Alahari, K.: End-to-end incremental learning. In: ECCV (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 364, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 364, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 364, + 481, + 396 + ], + "type": "text", + "content": "6. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012 (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 396, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 481, + 418 + ], + "type": "text", + "content": "7. Chen, K., Lee, C.G.: Incremental few-shot learning via vector quantization in deep embedded space. In: ICLR (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 418, + 481, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 481, + 438 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 481, + 438 + ], + "type": "text", + "content": "8. Chen, P.Y.: Model reprogramming: Resource-efficient cross-domain machine learning (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 438, + 481, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 438, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 438, + 481, + 471 + ], + "type": "text", + "content": "9. Cheraghian, A., Rahman, S., Fang, P., Roy, S.K., Petersson, L., Harandi, M.: Semantic-aware knowledge distillation for few-shot class-incremental learning. In: CVPR (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 471, + 481, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 471, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 138, + 471, + 481, + 503 + ], + "type": "text", + "content": "0. Cheraghian, A., Rahman, S., Ramasinghe, S., Fang, P., Simon, C., Petersson, L., Harandi, M.: Synthesized feature based few-shot class-incremental learning on a mixture of subspaces. In: ICCV (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 503, + 481, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 503, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 138, + 503, + 481, + 547 + ], + "type": "text", + "content": "1. Chowdhury, T., Cheraghian, A., Ramasinghe, S., Ahmadi, S., Saberi, M., Rahman, S.: Few-shot class-incremental learning for 3d point cloud objects. In: Avidan, S., Brostow, G., Cissé, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 204-220. Springer Nature Switzerland, Cham (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 547, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 481, + 568 + ], + "type": "text", + "content": "2. Chowdhury, T., Cheraghian, A., Ramasinghe, S., Ahmadi, S., Saberi, M., Rahman, S.: Few-shot class-incremental learning for 3d point cloud objects. In: ECCV (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 568, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 481, + 589 + ], + "type": "text", + "content": "3. Dinh, T., Seo, D., Du, Z., Shang, L., Lee, K.: Improved input reprogramming for gan conditioning (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 589, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 481, + 632 + ], + "type": "text", + "content": "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "5. Elsayed, G.F., Goodfellow, I., Sohl-Dickstein, J.: Adversarial reprogramming of neural networks. In: International Conference on Learning Representations (2019), https://openreview.net/forum?id=Syx_Ss05tm" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "text", + "content": "Canonical Shape Projection for 3D FSCIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 172 + ], + "type": "text", + "content": "16. Jaderberg, M., Simonyan, K., Zisserman, A., kavukcuoglu, k.: Spatial transformer networks. In: Cortes, C., Lawrence, N., Lee, D., Sugiyama, M., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 28. Curran Associates, Inc. (2015), https://proceedings.neurips.cc/paper_files/paper/2015/file/33ceb07bf4eeb3da587e268d663aba1a-Paper.pdf" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 172, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 239 + ], + "type": "text", + "content": "17. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 4904-4916. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/jia21b.html" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 239, + 481, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 239, + 481, + 261 + ], + "spans": [ + { + "bbox": [ + 133, + 239, + 481, + 261 + ], + "type": "text", + "content": "18. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 261, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 261, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 133, + 261, + 481, + 304 + ], + "type": "text", + "content": "19. Lee, D.H., Pujara, J., Sewak, M., White, R.W., Jauhar, S.K.: Making large language models better data creators. In: The 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP) (2023), https://openreview.net/forum?id=2Rdfdri2oT" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 305, + 481, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 305, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 305, + 481, + 340 + ], + "type": "text", + "content": "20. Lee, K.Y., Zhong, Y., Wang, Y.X.: Do pre-trained models benefit equally in continual learning? In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 6485-6493 (January 2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 340, + 481, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 340, + 481, + 362 + ], + "spans": [ + { + "bbox": [ + 132, + 340, + 481, + 362 + ], + "type": "text", + "content": "21. Li, Y., Bu, R., Sun, M., Wu, W., Di, X., Chen, B.: Pointcnn: Convolution on x-transformed points. In: NeurIPS (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 362, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 362, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 362, + 481, + 385 + ], + "type": "text", + "content": "22. Li, Z., Hoiem, D.: Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 385, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 481, + 407 + ], + "type": "text", + "content": "23. Liu, Y., Fan, B., Xiang, S., Pan, C.: Relation-shape convolutional neural network for point cloud analysis. In: CVPR (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 407, + 481, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 419 + ], + "type": "text", + "content": "24. Mazumder, P., Singh, P., Rai, P.: Few-shot lifelong learning. In: AAAI (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 419, + 481, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 419, + 481, + 474 + ], + "spans": [ + { + "bbox": [ + 133, + 419, + 481, + 474 + ], + "type": "text", + "content": "25. Pei, Y., Qing, Z., CEN, J., Wang, X., Zhang, S., Wang, Y., Tang, M., Sang, N., Qian, X.: Learning a condensed frame for memory-efficient video class-incremental learning. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=1CGYC7pXWNQ" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 475, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 481, + 497 + ], + "type": "text", + "content": "26. Poulenard, A., Rakotosaona, M.J., Ponty, Y., Ovsjanikov, M.: Effective rotation-invariant point cnn with spherical harmonics kernels. In: 3DV (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 497, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 497, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 497, + 481, + 520 + ], + "type": "text", + "content": "27. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: CVPR (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 520, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 481, + 542 + ], + "type": "text", + "content": "28. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In: NeurIPS (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 543, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 543, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 543, + 481, + 609 + ], + "type": "text", + "content": "29. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 8748-8763. PMLR (18-24 Jul 2021), https://proceedings.mlr.press/v139/radford21a.html" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 609, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 481, + 632 + ], + "type": "text", + "content": "30. Rao, Y., Lu, J., Zhou, J.: Spherical fractal convolutional neural networks for point cloud recognition. In: CVPR (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "31. Reizenstein, J., Shapovalov, R., Henzler, P., Sbordone, L., Labatut, P., Novotny, D.: Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In: ICCV (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "A. Cheraghian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "type": "text", + "content": "32. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F. (eds.) Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015. pp. 234-241. Springer International Publishing, Cham (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 162, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 162, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 162, + 481, + 182 + ], + "type": "text", + "content": "33. Singh, A., Hu, R., Goswami, V., Couairon, G., Galuba, W., Rohrbach, M., Kiela, D.: FLAVA: A foundational language and vision alignment model. In: CVPR (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "text", + "content": "34. Sung, F., Yang, Y., Zhang, L., Xiang, T., Torr, P.H., Hospedales, T.M.: Learning to compare: Relation network for few-shot learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "35. Tan, Y., Xiang, X.: Cross-domain few-shot incremental learning for point-cloud recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 2307-2316 (January 2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "36. Tan, Z., Ding, K., Guo, R., Liu, H.: Graph few-shot class-incremental learning. In: Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "type": "text", + "content": "37. Tao, X., Hong, X., Chang, X., Dong, S., Wei, X., Gong, Y.: Few-shot class-incremental learning. In: CVPR (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 367 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 367 + ], + "type": "text", + "content": "38. Tsai, Y.Y., Chen, P.Y., Ho, T.Y.: Transfer learning without knowing: Reprogramming black-box machine learning models with scarce data and limited resources. In: III, H.D., Singh, A. (eds.) Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 9614-9624. PMLR (13-18 Jul 2020), https://proceedings.mlr.press/v119/tsai20a.html" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 369, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 413 + ], + "type": "text", + "content": "39. Tsai, Y.Y., Chen, P.Y., Ho, T.Y.: Transfer learning without knowing: Reprogramming black-box machine learning models with scarce data and limited resources. In: Proceedings of the 37th International Conference on Machine Learning. ICML'20, JMLR.org (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "text", + "content": "40. Uy, M.A., Pham, Q.H., Hua, B.S., Nguyen, D.T., Yeung, S.K.: Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In: ICCV (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 447, + 481, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 467 + ], + "type": "text", + "content": "41. Wang, C., Samari, B., Siddiqi, K.: Local spectral graph convolution for point set feature learning. In: ECCV (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 511 + ], + "type": "text", + "content": "42. Wang, R., Duan, X., Kang, G., Liu, J., Lin, S., Xu, S., Lu, J., Zhang, B.: Attriclip: A non-incremental learner for incremental knowledge learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3654-3663 (June 2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "text", + "content": "43. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog) (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 555 + ], + "type": "text", + "content": "44. Wu, W., Qi, Z., Fuxin, L.: PointCONV: Deep convolutional networks on 3D point clouds. In: CVPR (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 555, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 577 + ], + "type": "text", + "content": "45. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., Xiao, J.: 3D ShapeNets: A deep representation for volumetric shapes. In: CVPR (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 578, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 609 + ], + "type": "text", + "content": "46. Xiang, T., Zhang, C., Song, Y., Yu, J., Cai, W.: Walk in the cloud: Learning curves for point clouds shape analysis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 915-924 (October 2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "text", + "content": "47. Xu, Y., Fan, T., Xu, M., Zeng, L., Qiao, Y.: SpiderCNN: Deep learning on point sets with parameterized convolutional filters. In: ECCV (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "48. Zhang, R., Guo, Z., Zhang, W., Li, K., Miao, X., Cui, B., Qiao, Y., Gao, P., Li, H.: Pointclip: Point cloud understanding by clip. arXiv preprint arXiv:2112.02413 (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "text", + "content": "Canonical Shape Projection for 3D FSCIL" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 194 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "49. Zhang, Y., Rabbat, M.: A graph-cnn for 3d point cloud classification. In: ICASSP (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "type": "text", + "content": "50. Zhou, D.W., Wang, F.Y., Ye, H.J., Ma, L., Pu, S., Zhan, D.C.: Forward compatible few-shot class-incremental learning. In: CVPR (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 194 + ], + "type": "text", + "content": "51. Zhu, X., Zhang, R., He, B., Guo, Z., Zeng, Z., Qin, Z., Zhang, S., Gao, P.: Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. arXiv preprint arXiv:2211.11682 (2022)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "A. Cheraghian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file